{
 "count": 41103,
 "records": [
  {
   "backend": "ExLlamaV2",
   "batch": null,
   "ctx": 941,
   "date": "2023-09",
   "flags": [],
   "hardware": "A6000",
   "id": "exldisc-10-a6000-freewilly2-70b-v1-1",
   "model": "FreeWilly2",
   "notes": "v1 = ExLlamaV2 v1 GPTQ runtime, v2 = v2.0 EXL2 format, same 4bit 70b weights; line is a full-run average over the generated tokens from the starting context; config=v1",
   "params": "70B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "GPTQ 4bit",
   "quote": "TheBloke_FreeWilly2-GPTQ_gptq-4bit-32g-actorder_True [70b]; gpu - a6000; v1; Output generated in 398.96 seconds (7.91 tokens/s, 3156 tokens, context 941, seed 1913709604)",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "ExLlamaV2 discussion #10 (community report)",
   "source_url": "https://github.com/turboderp-org/exllamav2/discussions/10",
   "tg_tokens": 3156,
   "tps": 7.91,
   "ttft_s": null
  },
  {
   "backend": "ExLlamaV2",
   "batch": null,
   "ctx": 941,
   "date": "2023-09",
   "flags": [],
   "hardware": "A6000",
   "id": "exldisc-10-a6000-freewilly2-70b-v1-2",
   "model": "FreeWilly2",
   "notes": "v1 = ExLlamaV2 v1 GPTQ runtime, v2 = v2.0 EXL2 format, same 4bit 70b weights; line is a full-run average over the generated tokens from the starting context; config=v1",
   "params": "70B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "GPTQ 4bit",
   "quote": "TheBloke_FreeWilly2-GPTQ_gptq-4bit-32g-actorder_True [70b]; gpu - a6000; v1; Output generated in 13.17 seconds (7.52 tokens/s, 99 tokens, context 941, seed 632398902)",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "ExLlamaV2 discussion #10 (community report)",
   "source_url": "https://github.com/turboderp-org/exllamav2/discussions/10",
   "tg_tokens": 99,
   "tps": 7.52,
   "ttft_s": null
  },
  {
   "backend": "ExLlamaV2",
   "batch": null,
   "ctx": 1040,
   "date": "2023-09",
   "flags": [],
   "hardware": "A6000",
   "id": "exldisc-10-a6000-freewilly2-70b-v1-3",
   "model": "FreeWilly2",
   "notes": "v1 = ExLlamaV2 v1 GPTQ runtime, v2 = v2.0 EXL2 format, same 4bit 70b weights; line is a full-run average over the generated tokens from the starting context; config=v1",
   "params": "70B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "GPTQ 4bit",
   "quote": "TheBloke_FreeWilly2-GPTQ_gptq-4bit-32g-actorder_True [70b]; gpu - a6000; v1; Output generated in 50.28 seconds (8.81 tokens/s, 443 tokens, context 1040, seed 1079219730)",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "ExLlamaV2 discussion #10 (community report)",
   "source_url": "https://github.com/turboderp-org/exllamav2/discussions/10",
   "tg_tokens": 443,
   "tps": 8.81,
   "ttft_s": null
  },
  {
   "backend": "ExLlamaV2",
   "batch": null,
   "ctx": 1483,
   "date": "2023-09",
   "flags": [],
   "hardware": "A6000",
   "id": "exldisc-10-a6000-freewilly2-70b-v1-4",
   "model": "FreeWilly2",
   "notes": "v1 = ExLlamaV2 v1 GPTQ runtime, v2 = v2.0 EXL2 format, same 4bit 70b weights; line is a full-run average over the generated tokens from the starting context; config=v1",
   "params": "70B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "GPTQ 4bit",
   "quote": "TheBloke_FreeWilly2-GPTQ_gptq-4bit-32g-actorder_True [70b]; gpu - a6000; v1; Output generated in 57.31 seconds (8.45 tokens/s, 484 tokens, context 1483, seed 46450141)",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "ExLlamaV2 discussion #10 (community report)",
   "source_url": "https://github.com/turboderp-org/exllamav2/discussions/10",
   "tg_tokens": 484,
   "tps": 8.45,
   "ttft_s": null
  },
  {
   "backend": "ExLlamaV2",
   "batch": null,
   "ctx": 1967,
   "date": "2023-09",
   "flags": [],
   "hardware": "A6000",
   "id": "exldisc-10-a6000-freewilly2-70b-v1-5",
   "model": "FreeWilly2",
   "notes": "v1 = ExLlamaV2 v1 GPTQ runtime, v2 = v2.0 EXL2 format, same 4bit 70b weights; line is a full-run average over the generated tokens from the starting context; config=v1",
   "params": "70B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "GPTQ 4bit",
   "quote": "TheBloke_FreeWilly2-GPTQ_gptq-4bit-32g-actorder_True [70b]; gpu - a6000; v1; Output generated in 67.15 seconds (8.07 tokens/s, 542 tokens, context 1967, seed 633887777)",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "ExLlamaV2 discussion #10 (community report)",
   "source_url": "https://github.com/turboderp-org/exllamav2/discussions/10",
   "tg_tokens": 542,
   "tps": 8.07,
   "ttft_s": null
  },
  {
   "backend": "ExLlamaV2",
   "batch": null,
   "ctx": 2509,
   "date": "2023-09",
   "flags": [],
   "hardware": "A6000",
   "id": "exldisc-10-a6000-freewilly2-70b-v1-6",
   "model": "FreeWilly2",
   "notes": "v1 = ExLlamaV2 v1 GPTQ runtime, v2 = v2.0 EXL2 format, same 4bit 70b weights; line is a full-run average over the generated tokens from the starting context; config=v1",
   "params": "70B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "GPTQ 4bit",
   "quote": "TheBloke_FreeWilly2-GPTQ_gptq-4bit-32g-actorder_True [70b]; gpu - a6000; v1; Output generated in 45.11 seconds (7.78 tokens/s, 351 tokens, context 2509, seed 2075010672)",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "ExLlamaV2 discussion #10 (community report)",
   "source_url": "https://github.com/turboderp-org/exllamav2/discussions/10",
   "tg_tokens": 351,
   "tps": 7.78,
   "ttft_s": null
  },
  {
   "backend": "ExLlamaV2",
   "batch": null,
   "ctx": 2860,
   "date": "2023-09",
   "flags": [],
   "hardware": "A6000",
   "id": "exldisc-10-a6000-freewilly2-70b-v1-7",
   "model": "FreeWilly2",
   "notes": "v1 = ExLlamaV2 v1 GPTQ runtime, v2 = v2.0 EXL2 format, same 4bit 70b weights; line is a full-run average over the generated tokens from the starting context; config=v1",
   "params": "70B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "GPTQ 4bit",
   "quote": "TheBloke_FreeWilly2-GPTQ_gptq-4bit-32g-actorder_True [70b]; gpu - a6000; v1; Output generated in 42.57 seconds (7.56 tokens/s, 322 tokens, context 2860, seed 358840736)",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "ExLlamaV2 discussion #10 (community report)",
   "source_url": "https://github.com/turboderp-org/exllamav2/discussions/10",
   "tg_tokens": 322,
   "tps": 7.56,
   "ttft_s": null
  },
  {
   "backend": "ExLlamaV2",
   "batch": null,
   "ctx": 3182,
   "date": "2023-09",
   "flags": [],
   "hardware": "A6000",
   "id": "exldisc-10-a6000-freewilly2-70b-v1-8",
   "model": "FreeWilly2",
   "notes": "v1 = ExLlamaV2 v1 GPTQ runtime, v2 = v2.0 EXL2 format, same 4bit 70b weights; line is a full-run average over the generated tokens from the starting context; config=v1",
   "params": "70B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "GPTQ 4bit",
   "quote": "TheBloke_FreeWilly2-GPTQ_gptq-4bit-32g-actorder_True [70b]; gpu - a6000; v1; Output generated in 55.01 seconds (7.38 tokens/s, 406 tokens, context 3182, seed 12195232)",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "ExLlamaV2 discussion #10 (community report)",
   "source_url": "https://github.com/turboderp-org/exllamav2/discussions/10",
   "tg_tokens": 406,
   "tps": 7.38,
   "ttft_s": null
  },
  {
   "backend": "ExLlamaV2",
   "batch": null,
   "ctx": 3588,
   "date": "2023-09",
   "flags": [],
   "hardware": "A6000",
   "id": "exldisc-10-a6000-freewilly2-70b-v1-9",
   "model": "FreeWilly2",
   "notes": "v1 = ExLlamaV2 v1 GPTQ runtime, v2 = v2.0 EXL2 format, same 4bit 70b weights; line is a full-run average over the generated tokens from the starting context; config=v1",
   "params": "70B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "GPTQ 4bit",
   "quote": "TheBloke_FreeWilly2-GPTQ_gptq-4bit-32g-actorder_True [70b]; gpu - a6000; v1; Output generated in 71.95 seconds (7.07 tokens/s, 509 tokens, context 3588, seed 486408052)",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "ExLlamaV2 discussion #10 (community report)",
   "source_url": "https://github.com/turboderp-org/exllamav2/discussions/10",
   "tg_tokens": 509,
   "tps": 7.07,
   "ttft_s": null
  },
  {
   "backend": "ExLlamaV2",
   "batch": null,
   "ctx": 941,
   "date": "2023-09",
   "flags": [],
   "hardware": "A6000",
   "id": "exldisc-10-a6000-freewilly2-70b-v2-1",
   "model": "FreeWilly2",
   "notes": "v1 = ExLlamaV2 v1 GPTQ runtime, v2 = v2.0 EXL2 format, same 4bit 70b weights; line is a full-run average over the generated tokens from the starting context; config=v2; author: \"It's definitely faster. I applaud this work.\"",
   "params": "70B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL2 4bit",
   "quote": "TheBloke_FreeWilly2-GPTQ_gptq-4bit-32g-actorder_True [70b]; gpu - a6000; v2; Output generated in 316.86 seconds (9.96 tokens/s, 3156 tokens, context 941, seed 839325157)",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "ExLlamaV2 discussion #10 (community report)",
   "source_url": "https://github.com/turboderp-org/exllamav2/discussions/10",
   "tg_tokens": 3156,
   "tps": 9.96,
   "ttft_s": null
  },
  {
   "backend": "ExLlamaV2",
   "batch": null,
   "ctx": 941,
   "date": "2023-09",
   "flags": [],
   "hardware": "A6000",
   "id": "exldisc-10-a6000-freewilly2-70b-v2-2",
   "model": "FreeWilly2",
   "notes": "v1 = ExLlamaV2 v1 GPTQ runtime, v2 = v2.0 EXL2 format, same 4bit 70b weights; line is a full-run average over the generated tokens from the starting context; config=v2",
   "params": "70B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL2 4bit",
   "quote": "TheBloke_FreeWilly2-GPTQ_gptq-4bit-32g-actorder_True [70b]; gpu - a6000; v2; Output generated in 327.32 seconds (9.64 tokens/s, 3156 tokens, context 941, seed 964072975)",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "ExLlamaV2 discussion #10 (community report)",
   "source_url": "https://github.com/turboderp-org/exllamav2/discussions/10",
   "tg_tokens": 3156,
   "tps": 9.64,
   "ttft_s": null
  },
  {
   "backend": "ExLlamaV2",
   "batch": null,
   "ctx": 941,
   "date": "2023-09",
   "flags": [],
   "hardware": "A6000",
   "id": "exldisc-10-a6000-freewilly2-70b-v2-3",
   "model": "FreeWilly2",
   "notes": "v1 = ExLlamaV2 v1 GPTQ runtime, v2 = v2.0 EXL2 format, same 4bit 70b weights; line is a full-run average over the generated tokens from the starting context; config=v2",
   "params": "70B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL2 4bit",
   "quote": "TheBloke_FreeWilly2-GPTQ_gptq-4bit-32g-actorder_True [70b]; gpu - a6000; v2; Output generated in 11.19 seconds (9.38 tokens/s, 105 tokens, context 941, seed 21383890161)",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "ExLlamaV2 discussion #10 (community report)",
   "source_url": "https://github.com/turboderp-org/exllamav2/discussions/10",
   "tg_tokens": 105,
   "tps": 9.38,
   "ttft_s": null
  },
  {
   "backend": "ExLlamaV2",
   "batch": null,
   "ctx": 1046,
   "date": "2023-09",
   "flags": [],
   "hardware": "A6000",
   "id": "exldisc-10-a6000-freewilly2-70b-v2-4",
   "model": "FreeWilly2",
   "notes": "v1 = ExLlamaV2 v1 GPTQ runtime, v2 = v2.0 EXL2 format, same 4bit 70b weights; line is a full-run average over the generated tokens from the starting context; config=v2",
   "params": "70B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL2 4bit",
   "quote": "TheBloke_FreeWilly2-GPTQ_gptq-4bit-32g-actorder_True [70b]; gpu - a6000; v2; Output generated in 32.63 seconds (10.97 tokens/s, 358 tokens, context 1046, seed 698368887)",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "ExLlamaV2 discussion #10 (community report)",
   "source_url": "https://github.com/turboderp-org/exllamav2/discussions/10",
   "tg_tokens": 358,
   "tps": 10.97,
   "ttft_s": null
  },
  {
   "backend": "ExLlamaV2",
   "batch": null,
   "ctx": 1404,
   "date": "2023-09",
   "flags": [],
   "hardware": "A6000",
   "id": "exldisc-10-a6000-freewilly2-70b-v2-5",
   "model": "FreeWilly2",
   "notes": "v1 = ExLlamaV2 v1 GPTQ runtime, v2 = v2.0 EXL2 format, same 4bit 70b weights; line is a full-run average over the generated tokens from the starting context; config=v2",
   "params": "70B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL2 4bit",
   "quote": "TheBloke_FreeWilly2-GPTQ_gptq-4bit-32g-actorder_True [70b]; gpu - a6000; v2; Output generated in 45.70 seconds (10.42 tokens/s, 476 tokens, context 1404, seed 267215585)",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "ExLlamaV2 discussion #10 (community report)",
   "source_url": "https://github.com/turboderp-org/exllamav2/discussions/10",
   "tg_tokens": 476,
   "tps": 10.42,
   "ttft_s": null
  },
  {
   "backend": "ExLlamaV2",
   "batch": null,
   "ctx": 1880,
   "date": "2023-09",
   "flags": [],
   "hardware": "A6000",
   "id": "exldisc-10-a6000-freewilly2-70b-v2-6",
   "model": "FreeWilly2",
   "notes": "v1 = ExLlamaV2 v1 GPTQ runtime, v2 = v2.0 EXL2 format, same 4bit 70b weights; line is a full-run average over the generated tokens from the starting context; config=v2",
   "params": "70B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL2 4bit",
   "quote": "TheBloke_FreeWilly2-GPTQ_gptq-4bit-32g-actorder_True [70b]; gpu - a6000; v2; Output generated in 50.37 seconds (9.73 tokens/s, 490 tokens, context 1880, seed 1519311258)",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "ExLlamaV2 discussion #10 (community report)",
   "source_url": "https://github.com/turboderp-org/exllamav2/discussions/10",
   "tg_tokens": 490,
   "tps": 9.73,
   "ttft_s": null
  },
  {
   "backend": "ExLlamaV2",
   "batch": null,
   "ctx": 2370,
   "date": "2023-09",
   "flags": [],
   "hardware": "A6000",
   "id": "exldisc-10-a6000-freewilly2-70b-v2-7",
   "model": "FreeWilly2",
   "notes": "v1 = ExLlamaV2 v1 GPTQ runtime, v2 = v2.0 EXL2 format, same 4bit 70b weights; line is a full-run average over the generated tokens from the starting context; config=v2",
   "params": "70B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL2 4bit",
   "quote": "TheBloke_FreeWilly2-GPTQ_gptq-4bit-32g-actorder_True [70b]; gpu - a6000; v2; Output generated in 71.70 seconds (9.16 tokens/s, 657 tokens, context 2370, seed 1991901992)",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "ExLlamaV2 discussion #10 (community report)",
   "source_url": "https://github.com/turboderp-org/exllamav2/discussions/10",
   "tg_tokens": 657,
   "tps": 9.16,
   "ttft_s": null
  },
  {
   "backend": "ExLlamaV2",
   "batch": null,
   "ctx": 3027,
   "date": "2023-09",
   "flags": [],
   "hardware": "A6000",
   "id": "exldisc-10-a6000-freewilly2-70b-v2-8",
   "model": "FreeWilly2",
   "notes": "v1 = ExLlamaV2 v1 GPTQ runtime, v2 = v2.0 EXL2 format, same 4bit 70b weights; line is a full-run average over the generated tokens from the starting context; config=v2",
   "params": "70B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL2 4bit",
   "quote": "TheBloke_FreeWilly2-GPTQ_gptq-4bit-32g-actorder_True [70b]; gpu - a6000; v2; Output generated in 62.19 seconds (8.43 tokens/s, 524 tokens, context 3027, seed 1124908778)",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "ExLlamaV2 discussion #10 (community report)",
   "source_url": "https://github.com/turboderp-org/exllamav2/discussions/10",
   "tg_tokens": 524,
   "tps": 8.43,
   "ttft_s": null
  },
  {
   "backend": "ExLlamaV2",
   "batch": null,
   "ctx": 3551,
   "date": "2023-09",
   "flags": [],
   "hardware": "A6000",
   "id": "exldisc-10-a6000-freewilly2-70b-v2-9",
   "model": "FreeWilly2",
   "notes": "v1 = ExLlamaV2 v1 GPTQ runtime, v2 = v2.0 EXL2 format, same 4bit 70b weights; line is a full-run average over the generated tokens from the starting context; config=v2",
   "params": "70B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL2 4bit",
   "quote": "TheBloke_FreeWilly2-GPTQ_gptq-4bit-32g-actorder_True [70b]; gpu - a6000; v2; Output generated in 46.36 seconds (7.49 tokens/s, 347 tokens, context 3551, seed 1028076373)",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "ExLlamaV2 discussion #10 (community report)",
   "source_url": "https://github.com/turboderp-org/exllamav2/discussions/10",
   "tg_tokens": 347,
   "tps": 7.49,
   "ttft_s": null
  },
  {
   "backend": "ExLlamaV2",
   "batch": null,
   "ctx": null,
   "date": "2023-10",
   "flags": [],
   "hardware": "3090",
   "id": "exldisc-119-3090-llama2-7b-exl2-4bpw",
   "model": "Llama 2 7B",
   "notes": "test_inference on 3090 + Xeon E5 CPU, Windows; author reports one Xeon core pinned at 100% @ 3.8 GHz, most likely CPU bottlenecked; ~80 t/s as reported",
   "params": "7B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL2 4bpw",
   "quote": "I ran test_inference on a 3090 and a Xeon E5 CPU on Windows and the process was most likely CPU bottlenecked; I got ~80t/s with llama2 7b EXL2 4bpw",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "ExLlamaV2 discussion #119 (community report)",
   "source_url": "https://github.com/turboderp-org/exllamav2/discussions/119",
   "tg_tokens": null,
   "tps": 80,
   "ttft_s": null
  },
  {
   "backend": "ExLlamaV2",
   "batch": null,
   "ctx": null,
   "date": "2023-10",
   "flags": [],
   "hardware": "3090",
   "id": "exldisc-119-3090-70b-exl2-2-5bpw",
   "model": "70B",
   "notes": "test_inference on 3090 + Xeon E5 CPU, Windows; ~20 t/s as reported, GPU at 98% utilization / 400 W tdp per author; source names the model only as '70b', Llama 2 implied by the paired llama2 7b run",
   "params": "70B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL2 2.5bpw",
   "quote": "I got ~20t/s with 70b EXL2 2.5bpw, which is about 50% and 70% of the speed listed for a 3090Ti; generation apparently fully loads the GPU @ 98% utilization and 400w tdp; Any chance that the inference performance can still be improved on Xeon CPUs that are clocked lower?",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "ExLlamaV2 discussion #119 (community report)",
   "source_url": "https://github.com/turboderp-org/exllamav2/discussions/119",
   "tg_tokens": null,
   "tps": 20,
   "ttft_s": null
  },
  {
   "backend": "ExLlamaV2",
   "batch": null,
   "ctx": null,
   "date": "2023-12",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "exldisc-247-rtx3090-mistral-exl2-5bpw-batched",
   "model": "Mistral",
   "notes": "title-level claim: 200+ t/s aggregate with concurrent requests through the author's own batching/caching API (EricLLM), a drop-in vLLM replacement; body links the repo without per-request numbers",
   "params": "7B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL2 5.0bpw",
   "quote": "Here's my batching/caching API I made over the weekend. 200+ tk/s with Mistral 5.0bpw.exl2 on an RTX 3090 with concurrent requests; It was for a personal project, and it's not complete, but it's very fast.",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "ExLlamaV2 discussion #247 (community report)",
   "source_url": "https://github.com/turboderp-org/exllamav2/discussions/247",
   "tg_tokens": null,
   "tps": 200,
   "ttft_s": null
  },
  {
   "backend": "LM Studio",
   "batch": null,
   "ctx": 8192,
   "date": "2024-05",
   "flags": [],
   "hardware": "RTX 4070 Ti SUPER",
   "id": "exldisc-471-rtx4070tisuper-llama3-8b-q8-lmstudio",
   "model": "Llama 3 8B Instruct",
   "notes": "Llama 3 8B Instruct Q8.0 GGUF fully offloaded; single fixed prompt (periodic table), ~56 t/s as reported by LM Studio; cross-backend comparison with the EXUI row in this discussion",
   "params": "8B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "run a quick speed test using the Llama 3 8B Instruct Q8.0 quants in both LM Studio and EXUI; i7-14700K, 64GB DDR4 4300MHz of RAM, RTX 4070Ti Super 16GB VRAM, Windows 11 Pro; List the first 30 elements of the periodic table, stating their atomic masses in brackets. Do it as a numbered list.; flash attention on, context set to 8192, FP16 cache in Exui and no speculative decoding, gguf fully offloaded to the GPU; LM Studio reported ~56 t/s",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "ExLlamaV2 discussion #471 (community report)",
   "source_url": "https://github.com/turboderp-org/exllamav2/discussions/471",
   "tg_tokens": null,
   "tps": 56,
   "ttft_s": null
  },
  {
   "backend": "ExLlamaV2",
   "batch": null,
   "ctx": 8192,
   "date": "2024-05",
   "flags": [],
   "hardware": "RTX 4070 Ti SUPER",
   "id": "exldisc-471-rtx4070tisuper-llama3-8b-exl2",
   "model": "Llama 3 8B Instruct",
   "notes": "Llama 3 8B Instruct EXL2 8.0bpw (author calls both quants 'Q8.0'); FP16 cache, flash attention on, no speculative decoding; ~64 t/s in EXUI, >14% faster than the LM Studio GGUF run on the same prompt; i7-14700K system",
   "params": "8B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL2 8.0bpw",
   "quote": "run a quick speed test using the Llama 3 8B Instruct Q8.0 quants in both LM Studio and EXUI; i7-14700K, 64GB DDR4 4300MHz of RAM, RTX 4070Ti Super 16GB VRAM, Windows 11 Pro; List the first 30 elements of the periodic table, stating their atomic masses in brackets. Do it as a numbered list.; flash attention on, context set to 8192, FP16 cache in Exui and no speculative decoding, gguf fully offloaded to the GPU; EXUI ~64 t/s which makes exl2 >14% faster that gguf in this specific test",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "ExLlamaV2 discussion #471 (community report)",
   "source_url": "https://github.com/turboderp-org/exllamav2/discussions/471",
   "tg_tokens": null,
   "tps": 64,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2024-07",
   "flags": [],
   "hardware": "RX 6800",
   "id": "exldisc-562-rx6800-phi3mini-llamacpp-gguf",
   "model": "Phi-3 mini 4k instruct",
   "notes": "llama-bench on an unquantized GGUF (F16) of Phi-3-mini-4k-instruct, ~42 t/s as reported; ROCm 6.1.2, gfx1030",
   "params": "3.8B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "GGUF F16",
   "quote": "I tried a simple test for Phi3 (microsoft/Phi-3-mini-4k-instruct), no quantization; For llama.cpp I converted it to a .gguf and ran llama-bench on it, resulting in a speed of ~42 tokens/sec on my RX 6800 (aka gfx1030)",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "ExLlamaV2 discussion #562 (community report)",
   "source_url": "https://github.com/turboderp-org/exllamav2/discussions/562",
   "tg_tokens": null,
   "tps": 42,
   "ttft_s": null
  },
  {
   "backend": "ExLlamaV2",
   "batch": null,
   "ctx": null,
   "date": "2024-07",
   "flags": [],
   "hardware": "RX 6800",
   "id": "exldisc-562-rx6800-phi3mini-exllamav2",
   "model": "Phi-3 mini 4k instruct",
   "notes": "examples/inference.py on the unquantized Phi-3-mini-4k-instruct, 18 t/s as reported; ROCm 6.1.2 gfx1030 build; author notes the gap to llama.cpp and asks for a sanity check",
   "params": "3.8B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "F16 (unquantized)",
   "quote": "I tried a simple test for Phi3 (microsoft/Phi-3-mini-4k-instruct), no quantization; For llama.cpp I converted it to a .gguf and ran llama-bench on it, resulting in a speed of ~42 tokens/sec on my RX 6800 (aka gfx1030); However, a test with `python examples/inference.py` gives a token generation speed of only 18 tokens/sec",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "ExLlamaV2 discussion #562 (community report)",
   "source_url": "https://github.com/turboderp-org/exllamav2/discussions/562",
   "tg_tokens": null,
   "tps": 18,
   "ttft_s": null
  },
  {
   "backend": "ExLlamaV2",
   "batch": null,
   "ctx": 0,
   "date": "2024-07",
   "flags": [],
   "hardware": "4x A10G",
   "id": "exldisc-572-4xa10g-llama3ft-env017-ctx0",
   "model": "Llama3 fine-tuned",
   "notes": "environment A/B, not code: 0.1.7 vs 0.1.8 venvs (torch 2.3 vs 2.4, older vs newer CUDA libs), same model per author (BW6); Request 1 only, Request 2+ ranges not point values; at ctx 12,500 the 0.1.8 env degrades ~0.05 t/s per 100 output tokens; thread title: 70B llama3.1 journey; config=env-0.1.7",
   "params": "70B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL2 (bpw not stated)",
   "quote": "The hardware: 4x A10G GPUs (96GB VRAM), 48 CPUs, 196GB RAM, using auto-split during load; The model: Llama3 fine-tuned, max window size 256k, allocated length 100k, BW6 : (We are using the same exact model for these two tests); Context window 0: 0.1.7 performance: Request 1: 8.98 t/s, Request 2+: 9.25 - 9.85 t/s",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "ExLlamaV2 discussion #572 (community report)",
   "source_url": "https://github.com/turboderp-org/exllamav2/discussions/572",
   "tg_tokens": null,
   "tps": 8.98,
   "ttft_s": null
  },
  {
   "backend": "ExLlamaV2",
   "batch": null,
   "ctx": 0,
   "date": "2024-07",
   "flags": [],
   "hardware": "4x A10G",
   "id": "exldisc-572-4xa10g-llama3ft-env018-ctx0",
   "model": "Llama3 fine-tuned",
   "notes": "environment A/B, not code: 0.1.7 vs 0.1.8 venvs (torch 2.3 vs 2.4, older vs newer CUDA libs), same model per author (BW6); Request 1 only, Request 2+ ranges not point values; at ctx 12,500 the 0.1.8 env degrades ~0.05 t/s per 100 output tokens; thread title: 70B llama3.1 journey; config=env-0.1.8",
   "params": "70B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL2 (bpw not stated)",
   "quote": "The hardware: 4x A10G GPUs (96GB VRAM), 48 CPUs, 196GB RAM, using auto-split during load; The model: Llama3 fine-tuned, max window size 256k, allocated length 100k, BW6 : (We are using the same exact model for these two tests); Context window 0: 0.1.8 performance: Request 1: 8.76 t/s, Request 2+ shows progressive degradation",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "ExLlamaV2 discussion #572 (community report)",
   "source_url": "https://github.com/turboderp-org/exllamav2/discussions/572",
   "tg_tokens": null,
   "tps": 8.76,
   "ttft_s": null
  },
  {
   "backend": "ExLlamaV2",
   "batch": null,
   "ctx": 12500,
   "date": "2024-07",
   "flags": [],
   "hardware": "4x A10G",
   "id": "exldisc-572-4xa10g-llama3ft-env017-ctx12500",
   "model": "Llama3 fine-tuned",
   "notes": "environment A/B, not code: 0.1.7 vs 0.1.8 venvs (torch 2.3 vs 2.4, older vs newer CUDA libs), same model per author (BW6); Request 1 only, Request 2+ ranges not point values; at ctx 12,500 the 0.1.8 env degrades ~0.05 t/s per 100 output tokens; thread title: 70B llama3.1 journey; config=env-0.1.7",
   "params": "70B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL2 (bpw not stated)",
   "quote": "The hardware: 4x A10G GPUs (96GB VRAM), 48 CPUs, 196GB RAM, using auto-split during load; The model: Llama3 fine-tuned, max window size 256k, allocated length 100k, BW6 : (We are using the same exact model for these two tests); Context window 12,500: 0.1.7 performance: Request 1: 7.68 t/s, Request 2+: 8.09 - 8.25 t/s",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "ExLlamaV2 discussion #572 (community report)",
   "source_url": "https://github.com/turboderp-org/exllamav2/discussions/572",
   "tg_tokens": null,
   "tps": 7.68,
   "ttft_s": null
  },
  {
   "backend": "ExLlamaV2",
   "batch": null,
   "ctx": 12500,
   "date": "2024-07",
   "flags": [],
   "hardware": "4x A10G",
   "id": "exldisc-572-4xa10g-llama3ft-env018-ctx12500",
   "model": "Llama3 fine-tuned",
   "notes": "environment A/B, not code: 0.1.7 vs 0.1.8 venvs (torch 2.3 vs 2.4, older vs newer CUDA libs), same model per author (BW6); Request 1 only, Request 2+ ranges not point values; at ctx 12,500 the 0.1.8 env degrades ~0.05 t/s per 100 output tokens; thread title: 70B llama3.1 journey; config=env-0.1.8",
   "params": "70B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL2 (bpw not stated)",
   "quote": "The hardware: 4x A10G GPUs (96GB VRAM), 48 CPUs, 196GB RAM, using auto-split during load; The model: Llama3 fine-tuned, max window size 256k, allocated length 100k, BW6 : (We are using the same exact model for these two tests); Context window 12,500: 0.1.8 performance: Request 1: 3.76 t/s, Request 2+ shows progressive degradation",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "ExLlamaV2 discussion #572 (community report)",
   "source_url": "https://github.com/turboderp-org/exllamav2/discussions/572",
   "tg_tokens": null,
   "tps": 3.76,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Threadripper PRO 3975WX",
   "id": "lc29973-tr-msvc-b11160",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Threadripper PRO 3975WX (Zen 2, 32C/64T, AVX2), 128 GB DDR4-2667 (~46.6 GB/s), pure CPU (-ngl 0 -dev none), llama-bench -p 512 -n 128 -r 3; MinGW vs MSVC build regression on qwen4exp MoE decode path; unsloth b11160-mix-a6922cc release, MSVC 19.44; config=msvc-b11160",
   "params": "177B",
   "power_w": null,
   "pp_tokens": "512",
   "pp_tps": "55.86",
   "provenance": "community",
   "quant": "UD-Q3_K_XL",
   "quote": "Threadripper PRO 3975WX; Qwen3.8-Flash-Next; unsloth release | MSVC 19.44 | haswell | **55.86** | **6.83**",
   "ram_gb": "128",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #29973",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29973",
   "tg_tokens": "128",
   "tps": "6.83",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Threadripper PRO 3975WX",
   "id": "lc29973-tr-mingw-znver2",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Threadripper PRO 3975WX (Zen 2, 32C/64T, AVX2), 128 GB DDR4-2667 (~46.6 GB/s), pure CPU (-ngl 0 -dev none), llama-bench -p 512 -n 128 -r 3; MinGW vs MSVC build regression on qwen4exp MoE decode path; self-built b11160 MinGW gcc 16.2.0 -march=native (znver2), decode 5.4x slower than MSVC; config=mingw-znver2",
   "params": "177B",
   "power_w": null,
   "pp_tokens": "512",
   "pp_tps": "52.60",
   "provenance": "community",
   "quant": "UD-Q3_K_XL",
   "quote": "Threadripper PRO 3975WX; Qwen3.8-Flash-Next; self-built A | MinGW gcc 16.2.0 | `-march=native` (znver2) | 52.60 | **1.27**",
   "ram_gb": "128",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #29973",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29973",
   "tg_tokens": "128",
   "tps": "1.27",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Threadripper PRO 3975WX",
   "id": "lc29973-tr-mingw-haswell",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Threadripper PRO 3975WX (Zen 2, 32C/64T, AVX2), 128 GB DDR4-2667 (~46.6 GB/s), pure CPU (-ngl 0 -dev none), llama-bench -p 512 -n 128 -r 3; MinGW vs MSVC build regression on qwen4exp MoE decode path; self-built b11160 MinGW multi-variant build (haswell dll confirmed loaded), same decode collapse; config=mingw-haswell",
   "params": "177B",
   "power_w": null,
   "pp_tokens": "512",
   "pp_tps": "51.08",
   "provenance": "community",
   "quant": "UD-Q3_K_XL",
   "quote": "Threadripper PRO 3975WX; Qwen3.8-Flash-Next; self-built B | MinGW gcc 16.2.0 | haswell (multi-variant build) | 51.08 | **1.28**",
   "ram_gb": "128",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #29973",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29973",
   "tg_tokens": "128",
   "tps": "1.28",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Threadripper PRO 3975WX",
   "id": "lc29973-tr-t16-short",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Threadripper PRO 3975WX, pure CPU, llama-server prefill (short prompt 68 tok / long prompt 1012 tok); -tb (threads-batch) is the largest prefill lever, llama-bench has no -tb; threads t16, short prompt 68 tok: 28.79 tok/s; config=t16-short",
   "params": "177B",
   "power_w": null,
   "pp_tokens": "68",
   "pp_tps": "28.79",
   "provenance": "community",
   "quant": "UD-Q3_K_XL",
   "quote": "Threadripper PRO 3975WX; Qwen3.8-Flash-Next; `t16` | 28.79 | 41.50",
   "ram_gb": "128",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #29973",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29973",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Threadripper PRO 3975WX",
   "id": "lc29973-tr-t16-long",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Threadripper PRO 3975WX, pure CPU, llama-server prefill (short prompt 68 tok / long prompt 1012 tok); -tb (threads-batch) is the largest prefill lever, llama-bench has no -tb; threads t16, long prompt 1012 tok: 41.50 tok/s; config=t16-long",
   "params": "177B",
   "power_w": null,
   "pp_tokens": "1012",
   "pp_tps": "41.50",
   "provenance": "community",
   "quant": "UD-Q3_K_XL",
   "quote": "Threadripper PRO 3975WX; Qwen3.8-Flash-Next; `t16` | 28.79 | 41.50",
   "ram_gb": "128",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #29973",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29973",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Threadripper PRO 3975WX",
   "id": "lc29973-tr-t20-short",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Threadripper PRO 3975WX, pure CPU, llama-server prefill (short prompt 68 tok / long prompt 1012 tok); -tb (threads-batch) is the largest prefill lever, llama-bench has no -tb; threads t20, short prompt 68 tok: 32.74 tok/s; config=t20-short",
   "params": "177B",
   "power_w": null,
   "pp_tokens": "68",
   "pp_tps": "32.74",
   "provenance": "community",
   "quant": "UD-Q3_K_XL",
   "quote": "Threadripper PRO 3975WX; Qwen3.8-Flash-Next; `t20` | 32.74 | 48.80",
   "ram_gb": "128",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #29973",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29973",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Threadripper PRO 3975WX",
   "id": "lc29973-tr-t20-long",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Threadripper PRO 3975WX, pure CPU, llama-server prefill (short prompt 68 tok / long prompt 1012 tok); -tb (threads-batch) is the largest prefill lever, llama-bench has no -tb; threads t20, long prompt 1012 tok: 48.80 tok/s; config=t20-long",
   "params": "177B",
   "power_w": null,
   "pp_tokens": "1012",
   "pp_tps": "48.80",
   "provenance": "community",
   "quant": "UD-Q3_K_XL",
   "quote": "Threadripper PRO 3975WX; Qwen3.8-Flash-Next; `t20` | 32.74 | 48.80",
   "ram_gb": "128",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #29973",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29973",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Threadripper PRO 3975WX",
   "id": "lc29973-tr-t20-tb24-short",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Threadripper PRO 3975WX, pure CPU, llama-server prefill (short prompt 68 tok / long prompt 1012 tok); -tb (threads-batch) is the largest prefill lever, llama-bench has no -tb; threads t20-tb24, short prompt 68 tok: 35.46 tok/s; config=t20-tb24-short",
   "params": "177B",
   "power_w": null,
   "pp_tokens": "68",
   "pp_tps": "35.46",
   "provenance": "community",
   "quant": "UD-Q3_K_XL",
   "quote": "Threadripper PRO 3975WX; Qwen3.8-Flash-Next; `t20-tb24` | 35.46 | 54.46",
   "ram_gb": "128",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #29973",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29973",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Threadripper PRO 3975WX",
   "id": "lc29973-tr-t20-tb24-long",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Threadripper PRO 3975WX, pure CPU, llama-server prefill (short prompt 68 tok / long prompt 1012 tok); -tb (threads-batch) is the largest prefill lever, llama-bench has no -tb; threads t20-tb24, long prompt 1012 tok: 54.46 tok/s; config=t20-tb24-long",
   "params": "177B",
   "power_w": null,
   "pp_tokens": "1012",
   "pp_tps": "54.46",
   "provenance": "community",
   "quant": "UD-Q3_K_XL",
   "quote": "Threadripper PRO 3975WX; Qwen3.8-Flash-Next; `t20-tb24` | 35.46 | 54.46",
   "ram_gb": "128",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #29973",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29973",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Threadripper PRO 3975WX",
   "id": "lc29973-tr-spec-pmin0",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Threadripper PRO 3975WX, pure CPU, llama-server decode with MTP speculative decoding (-t 20 -tb 24), acceptance in notes; default --spec-draft-p-min 0, acceptance 46.9%; config=pmin0",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q3_K_XL",
   "quote": "Threadripper PRO 3975WX; Qwen3.8-Flash-Next; `p-min 0` (default) | 8.12 | 46.9%",
   "ram_gb": "128",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #29973",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29973",
   "tg_tokens": null,
   "tps": "8.12",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Threadripper PRO 3975WX",
   "id": "lc29973-tr-spec-pmin030",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Threadripper PRO 3975WX, pure CPU, llama-server decode with MTP speculative decoding (-t 20 -tb 24), acceptance in notes; p-min 0.30, the author's best, acceptance 54.9%; config=pmin030",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q3_K_XL",
   "quote": "Threadripper PRO 3975WX; Qwen3.8-Flash-Next; **`p-min 0.30`** | **8.94** | 54.9%",
   "ram_gb": "128",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #29973",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29973",
   "tg_tokens": null,
   "tps": "8.94",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Threadripper PRO 3975WX",
   "id": "lc29973-tr-spec-pmin050",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Threadripper PRO 3975WX, pure CPU, llama-server decode with MTP speculative decoding (-t 20 -tb 24), acceptance in notes; p-min 0.50, highest acceptance but slower than 0.30, acceptance 62.1%; config=pmin050",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q3_K_XL",
   "quote": "Threadripper PRO 3975WX; Qwen3.8-Flash-Next; `p-min 0.50` | 8.69 | **62.1%**",
   "ram_gb": "128",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #29973",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29973",
   "tg_tokens": null,
   "tps": "8.69",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Threadripper PRO 3975WX",
   "id": "lc29973-tr-spec-nmax5",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Threadripper PRO 3975WX, pure CPU, llama-server decode with MTP speculative decoding (-t 20 -tb 24), acceptance in notes; n-max 5 collapses acceptance, acceptance 33.8%; config=nmax5",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q3_K_XL",
   "quote": "Threadripper PRO 3975WX; Qwen3.8-Flash-Next; `n-max 5` | 7.10 | 33.8%",
   "ram_gb": "128",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #29973",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29973",
   "tg_tokens": null,
   "tps": "7.10",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Threadripper PRO 3975WX",
   "id": "lc29973-tr-spec-ngram-only",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Threadripper PRO 3975WX, pure CPU, llama-server decode with MTP speculative decoding (-t 20 -tb 24), acceptance in notes; ngram-mod only, no MTP draft, acceptance 0%; config=ngram-only",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q3_K_XL",
   "quote": "Threadripper PRO 3975WX; Qwen3.8-Flash-Next; ngram-mod only (no MTP) | 6.86 | 0%",
   "ram_gb": "128",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #29973",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29973",
   "tg_tokens": null,
   "tps": "6.86",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Threadripper PRO 3975WX",
   "id": "lc29973-tr-spec-mtp-ngram",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Threadripper PRO 3975WX, pure CPU, llama-server decode with MTP speculative decoding (-t 20 -tb 24), acceptance in notes; MTP + ngram-mod hybrid, loses to tuned MTP, acceptance 46.9%; config=mtp-ngram",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q3_K_XL",
   "quote": "Threadripper PRO 3975WX; Qwen3.8-Flash-Next; MTP + ngram-mod (hybrid) | 8.27 | 46.9%",
   "ram_gb": "128",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #29973",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29973",
   "tg_tokens": null,
   "tps": "8.27",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Threadripper PRO 3975WX",
   "id": "lc29973-tr-relcuda-cpu",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Threadripper PRO 3975WX, GPU fully disabled (-ngl 0 -dev none), release-build CPU comparison: the CUDA build's CPU code is 12% faster than the Vulkan build's; windows-x64-cuda12-legacy build used CPU-only; config=rel-cuda-build",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "35.52",
   "provenance": "community",
   "quant": "UD-Q3_K_XL",
   "quote": "Threadripper PRO 3975WX; Qwen3.8-Flash-Next; windows-x64-cuda12-legacy | **35.52** | **8.87**",
   "ram_gb": "128",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #29973",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29973",
   "tg_tokens": null,
   "tps": "8.87",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Threadripper PRO 3975WX",
   "id": "lc29973-tr-relvulkan-cpu",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Threadripper PRO 3975WX, GPU fully disabled (-ngl 0 -dev none), release-build CPU comparison: the CUDA build's CPU code is 12% faster than the Vulkan build's; windows-x64-vulkan build used CPU-only; config=rel-vulkan-build",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "31.46",
   "provenance": "community",
   "quant": "UD-Q3_K_XL",
   "quote": "Threadripper PRO 3975WX; Qwen3.8-Flash-Next; windows-x64-vulkan | 31.46 | 8.28",
   "ram_gb": "128",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #29973",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29973",
   "tg_tokens": null,
   "tps": "8.28",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Threadripper PRO 3975WX + Quadro P2000",
   "id": "lc29973-tr-p2000-offload",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Threadripper PRO 3975WX + Quadro P2000 5 GB (sm_61), unsloth cuda12-legacy build: full GPU offload is a big loss (per-layer activation transfer over PCIe swamps any gain on 5 GB Pascal); baseline CPU tg128 6.83",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "14.36",
   "provenance": "community",
   "quant": "UD-Q3_K_XL",
   "quote": "Threadripper PRO 3975WX; Quadro P2000; Qwen3.8-Flash-Next; GPU offload (`-ngl 999 -ncmoe 48 -nkvo 1`) | pp 14.36 (\u221274%), tg 2.86 (\u221258%)",
   "ram_gb": "128",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #29973",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29973",
   "tg_tokens": null,
   "tps": "2.86",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Threadripper PRO 3975WX",
   "id": "lc29973-tr-t32",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Threadripper PRO 3975WX, pure CPU, tg128 thread/affinity sweep (MSVC build, baseline 6.83 at -t 20); -t 32, SMT contention on Zen 2; config=t32",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q3_K_XL",
   "quote": "Threadripper PRO 3975WX; Qwen3.8-Flash-Next; regresses at `-t 32` (6.21)",
   "ram_gb": "128",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #29973",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29973",
   "tg_tokens": "128",
   "tps": "6.21",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Threadripper PRO 3975WX",
   "id": "lc29973-tr-mask-full",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Threadripper PRO 3975WX, pure CPU, tg128 thread/affinity sweep (MSVC build, baseline 6.83 at -t 20); all 32 logical cores via affinity mask, below unbound; config=mask-full",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q3_K_XL",
   "quote": "Threadripper PRO 3975WX; Qwen3.8-Flash-Next; full mask 5.85",
   "ram_gb": "128",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #29973",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29973",
   "tg_tokens": "128",
   "tps": "5.85",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Threadripper PRO 3975WX",
   "id": "lc29973-tr-mask-nosmt",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Threadripper PRO 3975WX, pure CPU, tg128 thread/affinity sweep (MSVC build, baseline 6.83 at -t 20); affinity masks 0x55555555 / 0xAAAAAAAA, 16 threads one per physical core, published range 6.22-6.24 (stored low end); config=mask-nosmt",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q3_K_XL",
   "quote": "Threadripper PRO 3975WX; Qwen3.8-Flash-Next; both reach 6.22\u20136.24",
   "ram_gb": "128",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #29973",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29973",
   "tg_tokens": "128",
   "tps": "6.22",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "RTX 5070 + Ryzen 5 5600GT",
   "id": "lc29930-5070-streaming-bench",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Qwen3.8-Flash-Next 177B UD-IQ3_XXS expert-streaming setup (hashyy, llama.cpp based) on RTX 5070 12 GB + Ryzen 5 5600GT + 32 GB DDR4-2400 (PCIe Gen3), Windows; gains from fixed Windows I/O queue depth, one file handle per worker, page-locked hot-expert tier; published benchmark run, about 11.5 tok/s; config=streaming-bench",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ3_XXS",
   "quote": "RTX 5070; Ryzen 5 5600GT; Qwen3.8-Flash-Next; Benchmark is about 11.5 tok/s, up from roughly 7 tok/s on the inherited setup",
   "ram_gb": "32",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #29930",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29930",
   "tg_tokens": null,
   "tps": "11.5",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "RTX 5070 + Ryzen 5 5600GT",
   "id": "lc29930-5070-inherited",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Qwen3.8-Flash-Next 177B UD-IQ3_XXS expert-streaming setup (hashyy, llama.cpp based) on RTX 5070 12 GB + Ryzen 5 5600GT + 32 GB DDR4-2400 (PCIe Gen3), Windows; gains from fixed Windows I/O queue depth, one file handle per worker, page-locked hot-expert tier; inherited setup baseline before the streaming rewrite, roughly 7 tok/s; config=inherited",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ3_XXS",
   "quote": "RTX 5070; Ryzen 5 5600GT; Qwen3.8-Flash-Next; Benchmark is about 11.5 tok/s, up from roughly 7 tok/s on the inherited setup",
   "ram_gb": "32",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #29930",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29930",
   "tg_tokens": null,
   "tps": "7",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "RTX 5070 + Ryzen 5 5600GT",
   "id": "lc29930-5070-coding",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Qwen3.8-Flash-Next 177B UD-IQ3_XXS expert-streaming setup (hashyy, llama.cpp based) on RTX 5070 12 GB + Ryzen 5 5600GT + 32 GB DDR4-2400 (PCIe Gen3), Windows; gains from fixed Windows I/O queue depth, one file handle per worker, page-locked hot-expert tier; long coding prompt, 4892 tokens, working single-file Snake game output; config=coding-prompt",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ3_XXS",
   "quote": "RTX 5070; Ryzen 5 5600GT; Qwen3.8-Flash-Next; a long coding prompt generated 4,892 tokens at 10.15 tok/s",
   "ram_gb": "32",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #29930",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29930",
   "tg_tokens": "4892",
   "tps": "10.15",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "RTX 5070 + Ryzen 5 5600GT",
   "id": "lc29930-5070-video-10k",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Qwen3.8-Flash-Next 177B UD-IQ3_XXS expert-streaming setup (hashyy, llama.cpp based) on RTX 5070 12 GB + Ryzen 5 5600GT + 32 GB DDR4-2400 (PCIe Gen3), Windows; gains from fixed Windows I/O queue depth, one file handle per worker, page-locked hot-expert tier; video-verified 10k-token run, about 16 minutes; config=video-10k",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ3_XXS",
   "quote": "RTX 5070; Ryzen 5 5600GT; Qwen3.8-Flash-Next; around 16 minutes for 10k tokens and 10.41 tok/s",
   "ram_gb": "32",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #29930",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29930",
   "tg_tokens": "10000",
   "tps": "10.41",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "RTX 5070 + Ryzen 5 5600GT",
   "id": "lc29930-5070-conversations",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Qwen3.8-Flash-Next 177B UD-IQ3_XXS expert-streaming setup (hashyy, llama.cpp based) on RTX 5070 12 GB + Ryzen 5 5600GT + 32 GB DDR4-2400 (PCIe Gen3), Windows; gains from fixed Windows I/O queue depth, one file handle per worker, page-locked hot-expert tier; normal conversation use, published range 14-15 tok/s (stored low end); config=conversations",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ3_XXS",
   "quote": "RTX 5070; Ryzen 5 5600GT; Qwen3.8-Flash-Next; In normal conversations I\u2019ve seen 14\u201315 tok/s",
   "ram_gb": "32",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #29930",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29930",
   "tg_tokens": null,
   "tps": "14",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc24528-3090-qwen-iq4xs-q80-q51-thecodacus",
   "model": "Qwen3.8-Flash-Next",
   "notes": "2x RTX 3090 (x8+x8 lanes) + Ryzen 9950X, 192 GB DDR5-3600, Windows; expert-caching fork comparison on Qwen3.8-Flash-Next, llama-server, all rows with MTP and per-branch optimized parameters; thecodacus branch = expert cache + async expert prefetch + CUDA-pinned host RAM; quant IQ4_XS, KV quant q8_0/q5_1, branch thecodacus; config=iq4xs-q80-q51-thecodacus",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "2x RTX 3090; Qwen3.8-Flash-Next; Qwen3.8-Flash-Next | IQ4_XS | q8_0/q5_1 | thecodacus | 41.2",
   "ram_gb": "192",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": null,
   "tps": "41.2",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc24528-3090-qwen-iq4xs-q80-q51-thetom",
   "model": "Qwen3.8-Flash-Next",
   "notes": "2x RTX 3090 (x8+x8 lanes) + Ryzen 9950X, 192 GB DDR5-3600, Windows; expert-caching fork comparison on Qwen3.8-Flash-Next, llama-server, all rows with MTP and per-branch optimized parameters; thecodacus branch = expert cache + async expert prefetch + CUDA-pinned host RAM; quant IQ4_XS, KV quant q8_0/q5_1, branch TheTom; config=iq4xs-q80-q51-thetom",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "2x RTX 3090; Qwen3.8-Flash-Next; Qwen3.8-Flash-Next | IQ4_XS | q8_0/q5_1 | TheTom | 24.3",
   "ram_gb": "192",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": null,
   "tps": "24.3",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc24528-3090-qwen-iq4xs-q80-q51-generelschwerz",
   "model": "Qwen3.8-Flash-Next",
   "notes": "2x RTX 3090 (x8+x8 lanes) + Ryzen 9950X, 192 GB DDR5-3600, Windows; expert-caching fork comparison on Qwen3.8-Flash-Next, llama-server, all rows with MTP and per-branch optimized parameters; thecodacus branch = expert cache + async expert prefetch + CUDA-pinned host RAM; quant IQ4_XS, KV quant q8_0/q5_1, branch GenerelSchwerz; config=iq4xs-q80-q51-generelschwerz",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "2x RTX 3090; Qwen3.8-Flash-Next; Qwen3.8-Flash-Next | IQ4_XS | q8_0/q5_1 | GenerelSchwerz | 11.9",
   "ram_gb": "192",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": null,
   "tps": "11.9",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc24528-3090-qwen-iq4xs-q80-q51-upstream",
   "model": "Qwen3.8-Flash-Next",
   "notes": "2x RTX 3090 (x8+x8 lanes) + Ryzen 9950X, 192 GB DDR5-3600, Windows; expert-caching fork comparison on Qwen3.8-Flash-Next, llama-server, all rows with MTP and per-branch optimized parameters; thecodacus branch = expert cache + async expert prefetch + CUDA-pinned host RAM; quant IQ4_XS, KV quant q8_0/q5_1, branch upstream; config=iq4xs-q80-q51-upstream",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "2x RTX 3090; Qwen3.8-Flash-Next; Qwen3.8-Flash-Next | IQ4_XS | q8_0/q5_1 | upstream | 26.8",
   "ram_gb": "192",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": null,
   "tps": "26.8",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc24528-3090-qwen-iq4xs-q80-q80-thecodacus",
   "model": "Qwen3.8-Flash-Next",
   "notes": "2x RTX 3090 (x8+x8 lanes) + Ryzen 9950X, 192 GB DDR5-3600, Windows; expert-caching fork comparison on Qwen3.8-Flash-Next, llama-server, all rows with MTP and per-branch optimized parameters; thecodacus branch = expert cache + async expert prefetch + CUDA-pinned host RAM; quant IQ4_XS, KV quant q8_0/q8_0, branch thecodacus; config=iq4xs-q80-q80-thecodacus",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "2x RTX 3090; Qwen3.8-Flash-Next; Qwen3.8-Flash-Next | IQ4_XS | q8_0/q8_0 | thecodacus | 40.7",
   "ram_gb": "192",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": null,
   "tps": "40.7",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc24528-3090-qwen-iq4xs-q80-q80-thetom",
   "model": "Qwen3.8-Flash-Next",
   "notes": "2x RTX 3090 (x8+x8 lanes) + Ryzen 9950X, 192 GB DDR5-3600, Windows; expert-caching fork comparison on Qwen3.8-Flash-Next, llama-server, all rows with MTP and per-branch optimized parameters; thecodacus branch = expert cache + async expert prefetch + CUDA-pinned host RAM; quant IQ4_XS, KV quant q8_0/q8_0, branch TheTom; config=iq4xs-q80-q80-thetom",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "2x RTX 3090; Qwen3.8-Flash-Next; Qwen3.8-Flash-Next | IQ4_XS | q8_0/q8_0 | TheTom | 28.6",
   "ram_gb": "192",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": null,
   "tps": "28.6",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc24528-3090-qwen-iq4xs-q80-q80-upstream",
   "model": "Qwen3.8-Flash-Next",
   "notes": "2x RTX 3090 (x8+x8 lanes) + Ryzen 9950X, 192 GB DDR5-3600, Windows; expert-caching fork comparison on Qwen3.8-Flash-Next, llama-server, all rows with MTP and per-branch optimized parameters; thecodacus branch = expert cache + async expert prefetch + CUDA-pinned host RAM; quant IQ4_XS, KV quant q8_0/q8_0, branch upstream; config=iq4xs-q80-q80-upstream",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "2x RTX 3090; Qwen3.8-Flash-Next; Qwen3.8-Flash-Next | IQ4_XS | q8_0/q8_0 | upstream | 27.3",
   "ram_gb": "192",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": null,
   "tps": "27.3",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc24528-3090-qwen-iq4xs-f16-f16-thecodacus",
   "model": "Qwen3.8-Flash-Next",
   "notes": "2x RTX 3090 (x8+x8 lanes) + Ryzen 9950X, 192 GB DDR5-3600, Windows; expert-caching fork comparison on Qwen3.8-Flash-Next, llama-server, all rows with MTP and per-branch optimized parameters; thecodacus branch = expert cache + async expert prefetch + CUDA-pinned host RAM; quant IQ4_XS, KV quant f16/f16, branch thecodacus; config=iq4xs-f16-f16-thecodacus",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "2x RTX 3090; Qwen3.8-Flash-Next; Qwen3.8-Flash-Next | IQ4_XS | f16/f16 | thecodacus | 33.3",
   "ram_gb": "192",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": null,
   "tps": "33.3",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc24528-3090-qwen-iq4xs-f16-f16-upstream",
   "model": "Qwen3.8-Flash-Next",
   "notes": "2x RTX 3090 (x8+x8 lanes) + Ryzen 9950X, 192 GB DDR5-3600, Windows; expert-caching fork comparison on Qwen3.8-Flash-Next, llama-server, all rows with MTP and per-branch optimized parameters; thecodacus branch = expert cache + async expert prefetch + CUDA-pinned host RAM; quant IQ4_XS, KV quant f16/f16, branch upstream; config=iq4xs-f16-f16-upstream",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "2x RTX 3090; Qwen3.8-Flash-Next; Qwen3.8-Flash-Next | IQ4_XS | f16/f16 | upstream | 26.8",
   "ram_gb": "192",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": null,
   "tps": "26.8",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc24528-3090-qwen-q4km-q80-q51-thecodacus",
   "model": "Qwen3.8-Flash-Next",
   "notes": "2x RTX 3090 (x8+x8 lanes) + Ryzen 9950X, 192 GB DDR5-3600, Windows; expert-caching fork comparison on Qwen3.8-Flash-Next, llama-server, all rows with MTP and per-branch optimized parameters; thecodacus branch = expert cache + async expert prefetch + CUDA-pinned host RAM; quant Q4_K_M, KV quant q8_0/q5_1, branch thecodacus; config=q4km-q80-q51-thecodacus",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "2x RTX 3090; Qwen3.8-Flash-Next; Qwen3.8-Flash-Next | Q4_K_M | q8_0/q5_1 | thecodacus | 36.6",
   "ram_gb": "192",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": null,
   "tps": "36.6",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc24528-3090-qwen-q4km-q80-q51-upstream",
   "model": "Qwen3.8-Flash-Next",
   "notes": "2x RTX 3090 (x8+x8 lanes) + Ryzen 9950X, 192 GB DDR5-3600, Windows; expert-caching fork comparison on Qwen3.8-Flash-Next, llama-server, all rows with MTP and per-branch optimized parameters; thecodacus branch = expert cache + async expert prefetch + CUDA-pinned host RAM; quant Q4_K_M, KV quant q8_0/q5_1, branch upstream; config=q4km-q80-q51-upstream",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "2x RTX 3090; Qwen3.8-Flash-Next; Qwen3.8-Flash-Next | Q4_K_M | q8_0/q5_1 | upstream | 21.5",
   "ram_gb": "192",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": null,
   "tps": "21.5",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc24528-3090-qwen-q4km-q80-q80-thecodacus",
   "model": "Qwen3.8-Flash-Next",
   "notes": "2x RTX 3090 (x8+x8 lanes) + Ryzen 9950X, 192 GB DDR5-3600, Windows; expert-caching fork comparison on Qwen3.8-Flash-Next, llama-server, all rows with MTP and per-branch optimized parameters; thecodacus branch = expert cache + async expert prefetch + CUDA-pinned host RAM; quant Q4_K_M, KV quant q8_0/q8_0, branch thecodacus; config=q4km-q80-q80-thecodacus",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "2x RTX 3090; Qwen3.8-Flash-Next; Qwen3.8-Flash-Next | Q4_K_M | q8_0/q8_0 | thecodacus | 35.5",
   "ram_gb": "192",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": null,
   "tps": "35.5",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc24528-3090-qwen-q4km-q80-q80-upstream",
   "model": "Qwen3.8-Flash-Next",
   "notes": "2x RTX 3090 (x8+x8 lanes) + Ryzen 9950X, 192 GB DDR5-3600, Windows; expert-caching fork comparison on Qwen3.8-Flash-Next, llama-server, all rows with MTP and per-branch optimized parameters; thecodacus branch = expert cache + async expert prefetch + CUDA-pinned host RAM; quant Q4_K_M, KV quant q8_0/q8_0, branch upstream; config=q4km-q80-q80-upstream",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "2x RTX 3090; Qwen3.8-Flash-Next; Qwen3.8-Flash-Next | Q4_K_M | q8_0/q8_0 | upstream | 23.0",
   "ram_gb": "192",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": null,
   "tps": "23.0",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc24528-3090-qwen-q4km-f16-f16-thecodacus",
   "model": "Qwen3.8-Flash-Next",
   "notes": "2x RTX 3090 (x8+x8 lanes) + Ryzen 9950X, 192 GB DDR5-3600, Windows; expert-caching fork comparison on Qwen3.8-Flash-Next, llama-server, all rows with MTP and per-branch optimized parameters; thecodacus branch = expert cache + async expert prefetch + CUDA-pinned host RAM; quant Q4_K_M, KV quant f16/f16, branch thecodacus; config=q4km-f16-f16-thecodacus",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "2x RTX 3090; Qwen3.8-Flash-Next; Qwen3.8-Flash-Next | Q4_K_M | f16/f16 | thecodacus | 30.4",
   "ram_gb": "192",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": null,
   "tps": "30.4",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc24528-3090-qwen-q4km-f16-f16-upstream",
   "model": "Qwen3.8-Flash-Next",
   "notes": "2x RTX 3090 (x8+x8 lanes) + Ryzen 9950X, 192 GB DDR5-3600, Windows; expert-caching fork comparison on Qwen3.8-Flash-Next, llama-server, all rows with MTP and per-branch optimized parameters; thecodacus branch = expert cache + async expert prefetch + CUDA-pinned host RAM; quant Q4_K_M, KV quant f16/f16, branch upstream; config=q4km-f16-f16-upstream",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "2x RTX 3090; Qwen3.8-Flash-Next; Qwen3.8-Flash-Next | Q4_K_M | f16/f16 | upstream | 20.0",
   "ram_gb": "192",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": null,
   "tps": "20.0",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc24528-3090-qwen-q5km-q80-q51-thecodacus",
   "model": "Qwen3.8-Flash-Next",
   "notes": "2x RTX 3090 (x8+x8 lanes) + Ryzen 9950X, 192 GB DDR5-3600, Windows; expert-caching fork comparison on Qwen3.8-Flash-Next, llama-server, all rows with MTP and per-branch optimized parameters; thecodacus branch = expert cache + async expert prefetch + CUDA-pinned host RAM; quant Q5_K_M, KV quant q8_0/q5_1, branch thecodacus; config=q5km-q80-q51-thecodacus",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q5_K_M",
   "quote": "2x RTX 3090; Qwen3.8-Flash-Next; Qwen3.8-Flash-Next | Q5_K_M | q8_0/q5_1 | thecodacus | 24.4",
   "ram_gb": "192",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": null,
   "tps": "24.4",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc24528-3090-qwen-q5km-q80-q51-upstream",
   "model": "Qwen3.8-Flash-Next",
   "notes": "2x RTX 3090 (x8+x8 lanes) + Ryzen 9950X, 192 GB DDR5-3600, Windows; expert-caching fork comparison on Qwen3.8-Flash-Next, llama-server, all rows with MTP and per-branch optimized parameters; thecodacus branch = expert cache + async expert prefetch + CUDA-pinned host RAM; quant Q5_K_M, KV quant q8_0/q5_1, branch upstream; config=q5km-q80-q51-upstream",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q5_K_M",
   "quote": "2x RTX 3090; Qwen3.8-Flash-Next; Qwen3.8-Flash-Next | Q5_K_M | q8_0/q5_1 | upstream | 14.2",
   "ram_gb": "192",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": null,
   "tps": "14.2",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc24528-3090-qwen-q5km-q80-q80-thecodacus",
   "model": "Qwen3.8-Flash-Next",
   "notes": "2x RTX 3090 (x8+x8 lanes) + Ryzen 9950X, 192 GB DDR5-3600, Windows; expert-caching fork comparison on Qwen3.8-Flash-Next, llama-server, all rows with MTP and per-branch optimized parameters; thecodacus branch = expert cache + async expert prefetch + CUDA-pinned host RAM; quant Q5_K_M, KV quant q8_0/q8_0, branch thecodacus; config=q5km-q80-q80-thecodacus",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q5_K_M",
   "quote": "2x RTX 3090; Qwen3.8-Flash-Next; Qwen3.8-Flash-Next | Q5_K_M | q8_0/q8_0 | thecodacus | 27.9",
   "ram_gb": "192",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": null,
   "tps": "27.9",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc24528-3090-qwen-q5km-q80-q80-upstream",
   "model": "Qwen3.8-Flash-Next",
   "notes": "2x RTX 3090 (x8+x8 lanes) + Ryzen 9950X, 192 GB DDR5-3600, Windows; expert-caching fork comparison on Qwen3.8-Flash-Next, llama-server, all rows with MTP and per-branch optimized parameters; thecodacus branch = expert cache + async expert prefetch + CUDA-pinned host RAM; quant Q5_K_M, KV quant q8_0/q8_0, branch upstream; config=q5km-q80-q80-upstream",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q5_K_M",
   "quote": "2x RTX 3090; Qwen3.8-Flash-Next; Qwen3.8-Flash-Next | Q5_K_M | q8_0/q8_0 | upstream | 16.8",
   "ram_gb": "192",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": null,
   "tps": "16.8",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc24528-3090-qwen-q5km-f16-f16-thecodacus",
   "model": "Qwen3.8-Flash-Next",
   "notes": "2x RTX 3090 (x8+x8 lanes) + Ryzen 9950X, 192 GB DDR5-3600, Windows; expert-caching fork comparison on Qwen3.8-Flash-Next, llama-server, all rows with MTP and per-branch optimized parameters; thecodacus branch = expert cache + async expert prefetch + CUDA-pinned host RAM; quant Q5_K_M, KV quant f16/f16, branch thecodacus; config=q5km-f16-f16-thecodacus",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q5_K_M",
   "quote": "2x RTX 3090; Qwen3.8-Flash-Next; Qwen3.8-Flash-Next | Q5_K_M | f16/f16 | thecodacus | 26.2",
   "ram_gb": "192",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": null,
   "tps": "26.2",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc24528-3090-qwen-q5km-f16-f16-upstream",
   "model": "Qwen3.8-Flash-Next",
   "notes": "2x RTX 3090 (x8+x8 lanes) + Ryzen 9950X, 192 GB DDR5-3600, Windows; expert-caching fork comparison on Qwen3.8-Flash-Next, llama-server, all rows with MTP and per-branch optimized parameters; thecodacus branch = expert cache + async expert prefetch + CUDA-pinned host RAM; quant Q5_K_M, KV quant f16/f16, branch upstream; config=q5km-f16-f16-upstream",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q5_K_M",
   "quote": "2x RTX 3090; Qwen3.8-Flash-Next; Qwen3.8-Flash-Next | Q5_K_M | f16/f16 | upstream | 15.5",
   "ram_gb": "192",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": null,
   "tps": "15.5",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc24528-3090-qwen-ud-q6kxl-q80-q51-thecodacus",
   "model": "Qwen3.8-Flash-Next",
   "notes": "2x RTX 3090 (x8+x8 lanes) + Ryzen 9950X, 192 GB DDR5-3600, Windows; expert-caching fork comparison on Qwen3.8-Flash-Next, llama-server, all rows with MTP and per-branch optimized parameters; thecodacus branch = expert cache + async expert prefetch + CUDA-pinned host RAM; quant UD-Q6_K_XL, KV quant q8_0/q5_1, branch thecodacus; config=ud-q6kxl-q80-q51-thecodacus",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q6_K_XL",
   "quote": "2x RTX 3090; Qwen3.8-Flash-Next; Qwen3.8-Flash-Next | UD-Q6_K_XL | q8_0/q5_1 | thecodacus | 23.6",
   "ram_gb": "192",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": null,
   "tps": "23.6",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc24528-3090-qwen-ud-q6kxl-q80-q51-upstream",
   "model": "Qwen3.8-Flash-Next",
   "notes": "2x RTX 3090 (x8+x8 lanes) + Ryzen 9950X, 192 GB DDR5-3600, Windows; expert-caching fork comparison on Qwen3.8-Flash-Next, llama-server, all rows with MTP and per-branch optimized parameters; thecodacus branch = expert cache + async expert prefetch + CUDA-pinned host RAM; quant UD-Q6_K_XL, KV quant q8_0/q5_1, branch upstream; config=ud-q6kxl-q80-q51-upstream",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q6_K_XL",
   "quote": "2x RTX 3090; Qwen3.8-Flash-Next; Qwen3.8-Flash-Next | UD-Q6_K_XL | q8_0/q5_1 | upstream | 13.9",
   "ram_gb": "192",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": null,
   "tps": "13.9",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc24528-3090-qwen-ud-q6kxl-q80-q80-thecodacus",
   "model": "Qwen3.8-Flash-Next",
   "notes": "2x RTX 3090 (x8+x8 lanes) + Ryzen 9950X, 192 GB DDR5-3600, Windows; expert-caching fork comparison on Qwen3.8-Flash-Next, llama-server, all rows with MTP and per-branch optimized parameters; thecodacus branch = expert cache + async expert prefetch + CUDA-pinned host RAM; quant UD-Q6_K_XL, KV quant q8_0/q8_0, branch thecodacus; config=ud-q6kxl-q80-q80-thecodacus",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q6_K_XL",
   "quote": "2x RTX 3090; Qwen3.8-Flash-Next; Qwen3.8-Flash-Next | UD-Q6_K_XL | q8_0/q8_0 | thecodacus | 25.5",
   "ram_gb": "192",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": null,
   "tps": "25.5",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc24528-3090-qwen-ud-q6kxl-q80-q80-upstream",
   "model": "Qwen3.8-Flash-Next",
   "notes": "2x RTX 3090 (x8+x8 lanes) + Ryzen 9950X, 192 GB DDR5-3600, Windows; expert-caching fork comparison on Qwen3.8-Flash-Next, llama-server, all rows with MTP and per-branch optimized parameters; thecodacus branch = expert cache + async expert prefetch + CUDA-pinned host RAM; quant UD-Q6_K_XL, KV quant q8_0/q8_0, branch upstream; config=ud-q6kxl-q80-q80-upstream",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q6_K_XL",
   "quote": "2x RTX 3090; Qwen3.8-Flash-Next; Qwen3.8-Flash-Next | UD-Q6_K_XL | q8_0/q8_0 | upstream | 16.2",
   "ram_gb": "192",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": null,
   "tps": "16.2",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc24528-3090-qwen-ud-q6kxl-f16-f16-thecodacus",
   "model": "Qwen3.8-Flash-Next",
   "notes": "2x RTX 3090 (x8+x8 lanes) + Ryzen 9950X, 192 GB DDR5-3600, Windows; expert-caching fork comparison on Qwen3.8-Flash-Next, llama-server, all rows with MTP and per-branch optimized parameters; thecodacus branch = expert cache + async expert prefetch + CUDA-pinned host RAM; quant UD-Q6_K_XL, KV quant f16/f16, branch thecodacus; config=ud-q6kxl-f16-f16-thecodacus",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q6_K_XL",
   "quote": "2x RTX 3090; Qwen3.8-Flash-Next; Qwen3.8-Flash-Next | UD-Q6_K_XL | f16/f16 | thecodacus | 23.6",
   "ram_gb": "192",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": null,
   "tps": "23.6",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc24528-3090-qwen-ud-q6kxl-f16-f16-upstream",
   "model": "Qwen3.8-Flash-Next",
   "notes": "2x RTX 3090 (x8+x8 lanes) + Ryzen 9950X, 192 GB DDR5-3600, Windows; expert-caching fork comparison on Qwen3.8-Flash-Next, llama-server, all rows with MTP and per-branch optimized parameters; thecodacus branch = expert cache + async expert prefetch + CUDA-pinned host RAM; quant UD-Q6_K_XL, KV quant f16/f16, branch upstream; config=ud-q6kxl-f16-f16-upstream",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q6_K_XL",
   "quote": "2x RTX 3090; Qwen3.8-Flash-Next; Qwen3.8-Flash-Next | UD-Q6_K_XL | f16/f16 | upstream | 12.8",
   "ram_gb": "192",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": null,
   "tps": "12.8",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "65535",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc24528-3090-dsv4-c65535-leloch",
   "model": "DeepSeek-V4-Flash",
   "notes": "2x RTX 3090 (x8+x8 lanes) + Ryzen 9950X, 192 GB DDR5-3600, Windows; DeepSeek V4 Flash UD-Q8_K_XL (161 GB incl DSpark), DSpark MTP, q8_0 KV quant, per-branch optimized parameters; TheTom branch pinned to the Sep 2 2026 state; context 65535, branch leloch; config=c65535-leloch",
   "params": "284B-A13B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q8_K_XL",
   "quote": "2x RTX 3090; DeepSeek-V4-Flash; 65535 | leloch| 13.57",
   "ram_gb": "192",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": null,
   "tps": "13.57",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "65535",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc24528-3090-dsv4-c65535-TheTom",
   "model": "DeepSeek-V4-Flash",
   "notes": "2x RTX 3090 (x8+x8 lanes) + Ryzen 9950X, 192 GB DDR5-3600, Windows; DeepSeek V4 Flash UD-Q8_K_XL (161 GB incl DSpark), DSpark MTP, q8_0 KV quant, per-branch optimized parameters; TheTom branch pinned to the Sep 2 2026 state; context 65535, branch TheTom; config=c65535-TheTom",
   "params": "284B-A13B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q8_K_XL",
   "quote": "2x RTX 3090; DeepSeek-V4-Flash; 65535 | TheTom | 12.7",
   "ram_gb": "192",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": null,
   "tps": "12.7",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "65535",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc24528-3090-dsv4-c65535-upstream",
   "model": "DeepSeek-V4-Flash",
   "notes": "2x RTX 3090 (x8+x8 lanes) + Ryzen 9950X, 192 GB DDR5-3600, Windows; DeepSeek V4 Flash UD-Q8_K_XL (161 GB incl DSpark), DSpark MTP, q8_0 KV quant, per-branch optimized parameters; TheTom branch pinned to the Sep 2 2026 state; context 65535, branch upstream; config=c65535-upstream",
   "params": "284B-A13B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q8_K_XL",
   "quote": "2x RTX 3090; DeepSeek-V4-Flash; 65535 | upstream | 6.88",
   "ram_gb": "192",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": null,
   "tps": "6.88",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "405000",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc24528-3090-dsv4-c405000-leloch",
   "model": "DeepSeek-V4-Flash",
   "notes": "2x RTX 3090 (x8+x8 lanes) + Ryzen 9950X, 192 GB DDR5-3600, Windows; DeepSeek V4 Flash UD-Q8_K_XL (161 GB incl DSpark), DSpark MTP, q8_0 KV quant, per-branch optimized parameters; TheTom branch pinned to the Sep 2 2026 state; context 405000, branch leloch; config=c405000-leloch",
   "params": "284B-A13B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q8_K_XL",
   "quote": "2x RTX 3090; DeepSeek-V4-Flash; 405000 | leloch| 13.03",
   "ram_gb": "192",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": null,
   "tps": "13.03",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "405000",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc24528-3090-dsv4-c405000-TheTom",
   "model": "DeepSeek-V4-Flash",
   "notes": "2x RTX 3090 (x8+x8 lanes) + Ryzen 9950X, 192 GB DDR5-3600, Windows; DeepSeek V4 Flash UD-Q8_K_XL (161 GB incl DSpark), DSpark MTP, q8_0 KV quant, per-branch optimized parameters; TheTom branch pinned to the Sep 2 2026 state; context 405000, branch TheTom; config=c405000-TheTom",
   "params": "284B-A13B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q8_K_XL",
   "quote": "2x RTX 3090; DeepSeek-V4-Flash; 405000 | TheTom | 12.57",
   "ram_gb": "192",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": null,
   "tps": "12.57",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "405000",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc24528-3090-dsv4-c405000-upstream",
   "model": "DeepSeek-V4-Flash",
   "notes": "2x RTX 3090 (x8+x8 lanes) + Ryzen 9950X, 192 GB DDR5-3600, Windows; DeepSeek V4 Flash UD-Q8_K_XL (161 GB incl DSpark), DSpark MTP, q8_0 KV quant, per-branch optimized parameters; TheTom branch pinned to the Sep 2 2026 state; context 405000, branch upstream; config=c405000-upstream",
   "params": "284B-A13B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q8_K_XL",
   "quote": "2x RTX 3090; DeepSeek-V4-Flash; 405000 | upstream | 7.08",
   "ram_gb": "192",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": null,
   "tps": "7.08",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "1048576",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc24528-3090-dsv4-c1048576-leloch",
   "model": "DeepSeek-V4-Flash",
   "notes": "2x RTX 3090 (x8+x8 lanes) + Ryzen 9950X, 192 GB DDR5-3600, Windows; DeepSeek V4 Flash UD-Q8_K_XL (161 GB incl DSpark), DSpark MTP, q8_0 KV quant, per-branch optimized parameters; TheTom branch pinned to the Sep 2 2026 state; context 1048576, branch leloch; config=c1048576-leloch",
   "params": "284B-A13B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q8_K_XL",
   "quote": "2x RTX 3090; DeepSeek-V4-Flash; 1048576 | leloch| 10.96",
   "ram_gb": "192",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": null,
   "tps": "10.96",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "1048576",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc24528-3090-dsv4-c1048576-TheTom",
   "model": "DeepSeek-V4-Flash",
   "notes": "2x RTX 3090 (x8+x8 lanes) + Ryzen 9950X, 192 GB DDR5-3600, Windows; DeepSeek V4 Flash UD-Q8_K_XL (161 GB incl DSpark), DSpark MTP, q8_0 KV quant, per-branch optimized parameters; TheTom branch pinned to the Sep 2 2026 state; context 1048576, branch TheTom; config=c1048576-TheTom",
   "params": "284B-A13B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q8_K_XL",
   "quote": "2x RTX 3090; DeepSeek-V4-Flash; 1048576 | TheTom | 9.71",
   "ram_gb": "192",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": null,
   "tps": "9.71",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "1048576",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc24528-3090-dsv4-c1048576-upstream",
   "model": "DeepSeek-V4-Flash",
   "notes": "2x RTX 3090 (x8+x8 lanes) + Ryzen 9950X, 192 GB DDR5-3600, Windows; DeepSeek V4 Flash UD-Q8_K_XL (161 GB incl DSpark), DSpark MTP, q8_0 KV quant, per-branch optimized parameters; TheTom branch pinned to the Sep 2 2026 state; context 1048576, branch upstream; config=c1048576-upstream",
   "params": "284B-A13B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q8_K_XL",
   "quote": "2x RTX 3090; DeepSeek-V4-Flash; 1048576 | upstream | 7.53",
   "ram_gb": "192",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": null,
   "tps": "7.53",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "M3 Pro 36GB",
   "id": "lc24528-m3pro-resident",
   "model": "Qwen3 30B A3B",
   "notes": "Metal on-demand expert loading directly from disk (exact, no cache approximation); M3 Pro 36 GB, fully resident baseline; config=resident",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "M3 Pro 36GB; Qwen3-30B-A3B; Qwen3-30B-A3B-Q6_K went from 38.1 tok/s fully resident to 27.5 tok/s while saving 11GB of memory, and 22.1 tok/s while saving 16.6GB",
   "ram_gb": "36",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528#discussioncomment-18312714",
   "tg_tokens": null,
   "tps": "38.1",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "M3 Pro 36GB",
   "id": "lc24528-m3pro-stream11",
   "model": "Qwen3 30B A3B",
   "notes": "Metal on-demand expert loading directly from disk (exact, no cache approximation); M3 Pro 36 GB, streaming experts from disk saving 11 GB; config=stream-11gb",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "M3 Pro 36GB; Qwen3-30B-A3B; Qwen3-30B-A3B-Q6_K went from 38.1 tok/s fully resident to 27.5 tok/s while saving 11GB of memory, and 22.1 tok/s while saving 16.6GB",
   "ram_gb": "36",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528#discussioncomment-18312714",
   "tg_tokens": null,
   "tps": "27.5",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "M3 Pro 36GB",
   "id": "lc24528-m3pro-stream16",
   "model": "Qwen3 30B A3B",
   "notes": "Metal on-demand expert loading directly from disk (exact, no cache approximation); M3 Pro 36 GB, streaming experts from disk saving 16.6 GB; config=stream-16-6gb",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "M3 Pro 36GB; Qwen3-30B-A3B; Qwen3-30B-A3B-Q6_K went from 38.1 tok/s fully resident to 27.5 tok/s while saving 11GB of memory, and 22.1 tok/s while saving 16.6GB",
   "ram_gb": "36",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528#discussioncomment-18312714",
   "tg_tokens": null,
   "tps": "22.1",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "M1 Pro 16GB",
   "id": "lc24528-m1pro-stream",
   "model": "Qwen3 30B A3B",
   "notes": "Metal on-demand expert loading directly from disk (exact, no cache approximation); M1 Pro 16 GB, running via on-demand expert loading; config=m1pro",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "M1 Pro 16GB; Qwen3-30B-A3B; I also got it running on an M1 Pro 16GB at 13 tok/s",
   "ram_gb": "16",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528#discussioncomment-18312714",
   "tg_tokens": null,
   "tps": "13",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "262144",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc24528-3090-simlu-dsv4-tg500",
   "model": "DeepSeek-V4-Flash",
   "notes": "simlu expert-cache branch (GGML_CUDA_MOE_CACHE_RESERVE_MB=0, --moe-cache 48000, DSpark/MTP draft, q8_0 KV+draft KV), llama-server on Volunteer-1's rig (Ryzen 9950X, 192 GB DDR5-3600, Windows), parameters optimized for decode; DeepSeek V4 Flash UD-Q8_K_XL with DSpark draft, tg500; config=simlu-dsv4-tg500",
   "params": "284B-A13B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q8_K_XL",
   "quote": "2x RTX 3090; DeepSeek-V4-Flash; DeepSeek V4 Flash UD-Q8_K_XL; tg500 13.20 t/s",
   "ram_gb": "192",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": "500",
   "tps": "13.20",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "262144",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc24528-3090-simlu-dsv4-tg1000",
   "model": "DeepSeek-V4-Flash",
   "notes": "simlu expert-cache branch (GGML_CUDA_MOE_CACHE_RESERVE_MB=0, --moe-cache 48000, DSpark/MTP draft, q8_0 KV+draft KV), llama-server on Volunteer-1's rig (Ryzen 9950X, 192 GB DDR5-3600, Windows), parameters optimized for decode; DeepSeek V4 Flash UD-Q8_K_XL with DSpark draft, tg1000; config=simlu-dsv4-tg1000",
   "params": "284B-A13B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q8_K_XL",
   "quote": "2x RTX 3090; DeepSeek-V4-Flash; DeepSeek V4 Flash UD-Q8_K_XL; tg1000 13.12 t/s",
   "ram_gb": "192",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": "1000",
   "tps": "13.12",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "1048576",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc24528-3090-simlu-qwen-tg500",
   "model": "Qwen3.8-Flash-Next",
   "notes": "simlu expert-cache branch (GGML_CUDA_MOE_CACHE_RESERVE_MB=0, --moe-cache 48000, DSpark/MTP draft, q8_0 KV+draft KV), llama-server on Volunteer-1's rig (Ryzen 9950X, 192 GB DDR5-3600, Windows), parameters optimized for decode; Qwen3.8-Flash-Next IQ4_XS-PLEQ4 (AesSedai) with MTP draft, tg500, max context 1M; config=simlu-qwen-tg500",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ4_XS-PLEQ4",
   "quote": "2x RTX 3090; Qwen3.8-Flash-Next; Qwen3.8-Flash-Next IQ4_XS-PLEQ4; tg500 37.7 t/s",
   "ram_gb": "192",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": "500",
   "tps": "37.7",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "1048576",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc24528-3090-simlu-qwen-tg1000",
   "model": "Qwen3.8-Flash-Next",
   "notes": "simlu expert-cache branch (GGML_CUDA_MOE_CACHE_RESERVE_MB=0, --moe-cache 48000, DSpark/MTP draft, q8_0 KV+draft KV), llama-server on Volunteer-1's rig (Ryzen 9950X, 192 GB DDR5-3600, Windows), parameters optimized for decode; Qwen3.8-Flash-Next IQ4_XS-PLEQ4 (AesSedai) with MTP draft, tg1000 (source writes ts/s), max context 1M; config=simlu-qwen-tg1000",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ4_XS-PLEQ4",
   "quote": "2x RTX 3090; Qwen3.8-Flash-Next; Qwen3.8-Flash-Next IQ4_XS-PLEQ4; tg1000 38.8 ts/s",
   "ram_gb": "192",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": "1000",
   "tps": "38.8",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "262144",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc24528-3090-simlu-glm-tg500",
   "model": "GLM-5.3-Flash",
   "notes": "simlu expert-cache branch (GGML_CUDA_MOE_CACHE_RESERVE_MB=0, --moe-cache 48000, DSpark/MTP draft, q8_0 KV+draft KV), llama-server on Volunteer-1's rig (Ryzen 9950X, 192 GB DDR5-3600, Windows), parameters optimized for decode; GLM-5.3-Flash Q4_K_M (AesSedai), no draft model, tg500; config=simlu-glm-tg500",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "2x RTX 3090; GLM 5.3 Flash Q4_K_M; tg500 7.72 t/s",
   "ram_gb": "192",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": "500",
   "tps": "7.72",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "262144",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc24528-3090-simlu-glm-tg1000",
   "model": "GLM-5.3-Flash",
   "notes": "simlu expert-cache branch (GGML_CUDA_MOE_CACHE_RESERVE_MB=0, --moe-cache 48000, DSpark/MTP draft, q8_0 KV+draft KV), llama-server on Volunteer-1's rig (Ryzen 9950X, 192 GB DDR5-3600, Windows), parameters optimized for decode; GLM-5.3-Flash Q4_K_M (AesSedai), no draft model, tg1000; config=simlu-glm-tg1000",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "2x RTX 3090; GLM 5.3 Flash Q4_K_M; tg1000 7.90 t/s",
   "ram_gb": "192",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": "1000",
   "tps": "7.90",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "32768",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "2x RTX 3090 + EPYC 7663",
   "id": "lc24528-epyc-csantiago-partial",
   "model": "Qwen3.8-Flash-Next",
   "notes": "csantiago78 expert-cache branch (PR #27861 reconstruction: danielhanchen MTP base 53b1389d0 + bccbacdb894 + cherry-pick 08930e7f3 + follow-up patches), rented server: 2x RTX 3090 sm_86 across two NUMA nodes no P2P, 2x EPYC 7663 (112C/224T), 640 GB DDR4 ECC, Linux, CUDA 12.8; Qwen3.8-Flash-Next UD-IQ4_XS, shared Q8_0 MTP draft GPU-resident, small-batch gate 3, partial count mid-generation; config=csantiago-partial",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "2x RTX 3090; EPYC 7663; Qwen3.8-Flash-Next; reached **39.55 t/s at 3,493 generated tokens**",
   "ram_gb": "640",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": "3493",
   "tps": "39.55",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "32768",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "2x RTX 3090 + EPYC 7663",
   "id": "lc24528-epyc-csantiago-full",
   "model": "Qwen3.8-Flash-Next",
   "notes": "csantiago78 expert-cache branch (PR #27861 reconstruction: danielhanchen MTP base 53b1389d0 + bccbacdb894 + cherry-pick 08930e7f3 + follow-up patches), rented server: 2x RTX 3090 sm_86 across two NUMA nodes no P2P, 2x EPYC 7663 (112C/224T), 640 GB DDR4 ECC, Linux, CUDA 12.8; same run, full 8513-token generation, prompt processing 30.83 t/s on a 380-token prompt; config=csantiago-full",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "30.83",
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "2x RTX 3090; EPYC 7663; Qwen3.8-Flash-Next; **41.01 t/s over the complete 8,513-token generation**; | Prompt processing | 30.83 t/s |",
   "ram_gb": "640",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": "8513",
   "tps": "41.01",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "262144",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "2x RTX 3090 + EPYC 7663",
   "id": "lc24528-epyc-gs-mtp2",
   "model": "Qwen3.8-Flash-Next",
   "notes": "GenerelSchwerz moe-cache branch 0031a507b on the same rented server (2x RTX 3090, 2x EPYC 7663, 640 GB), Qwen3.8-Flash-Next UD-IQ4_XS, MTP2 with shared Q8_0 draft pinned to CUDA1, 256k-capacity run (not 256k in KV), MTP2 acceptance 43.92%, mean accepted length 1.88; config=gs-mtp2-256k",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "73.97",
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "2x RTX 3090; EPYC 7663; Qwen3.8-Flash-Next; | Prompt processing | 73.97 tok/s |; | Sustained generation | **62.02 tok/s** |",
   "ram_gb": "640",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": "5100",
   "tps": "62.02",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc24528-3090-csantiago-493",
   "model": "Qwen3.8-Flash-Next",
   "notes": "csantiago78 branch PR #27861 with thread patches plus the author's own, Qwen3.8-Flash-Next IQ4_XS on Volunteer-1's 2x RTX 3090 rig, best generation speed at 3500 tokens; config=csantiago-493",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "2x RTX 3090; Qwen3.8-Flash-Next; **49.3 t/s @ 3500 tokens**",
   "ram_gb": "192",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": "3500",
   "tps": "49.3",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc24528-3090-csantiago-combined",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Volunteer-1's combined plug-and-play patch set for the csantiago78 moe-cache work (MTP not yet integrated), Qwen3.8-Flash-Next IQ4_XS at 3000 tokens; config=csantiago-combined",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "2x RTX 3090; Qwen3.8-Flash-Next; **42 t/s @ 3000 tokens**",
   "ram_gb": "192",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": "3000",
   "tps": "42",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "262000",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "RTX 5070 Ti 16 GB + RTX 4060 Ti 16 GB",
   "id": "lc28248-dualgpu-mec-0-baseline",
   "model": "DeepSeek-V4-Flash-Vision-Exp",
   "notes": "TmDagger fix-moe-pool-rail fork (persistent per-tensor VRAM expert pool -mec, per-device budget with compute-buffer rail), RTX 5070 Ti 16 GB on PCIe 5.0 x16 (CUDA0) + RTX 4060 Ti 16 GB on PCIe 4.0 x4 (CUDA1), Windows 11, 128 GB DDR4-3200, DeepSeek-V4-Flash-Vision-Exp UD-IQ3_XXS without mmproj, -c 262000 -ngl 99 --split-mode layer -ncmoe 99 --seed 42; baseline -mec 0 with -ncmoe 39 (author notes -ncmoe 99 baseline was slower); config=mec-0-baseline",
   "params": "284B-A13B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "43.7",
   "provenance": "community",
   "quant": "UD-IQ3_XXS",
   "quote": "2 GPUs: RTX 5070 Ti 16 GB; RTX 4060 Ti 16 GB; unsloth/DeepSeek-V4-Flash-Vision-Exp, UD-IQ3_XXS; | `-mec 0 -ncmoe 39` (baseline) | - | - | - | 5.36 | 43.7 |",
   "ram_gb": "128",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #28248",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28248",
   "tg_tokens": null,
   "tps": "5.36",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "262000",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "RTX 5070 Ti 16 GB + RTX 4060 Ti 16 GB",
   "id": "lc28248-dualgpu-mec-48",
   "model": "DeepSeek-V4-Flash-Vision-Exp",
   "notes": "TmDagger fix-moe-pool-rail fork (persistent per-tensor VRAM expert pool -mec, per-device budget with compute-buffer rail), RTX 5070 Ti 16 GB on PCIe 5.0 x16 (CUDA0) + RTX 4060 Ti 16 GB on PCIe 4.0 x4 (CUDA1), Windows 11, 128 GB DDR4-3200, DeepSeek-V4-Flash-Vision-Exp UD-IQ3_XXS without mmproj, -c 262000 -ngl 99 --split-mode layer -ncmoe 99 --seed 42; -mec 48 both devices, window hit 85%/93%; a first post-rebase run measured 8.97 before spill fixes; config=mec-48",
   "params": "284B-A13B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "44.2",
   "provenance": "community",
   "quant": "UD-IQ3_XXS",
   "quote": "2 GPUs: RTX 5070 Ti 16 GB; RTX 4060 Ti 16 GB; unsloth/DeepSeek-V4-Flash-Vision-Exp, UD-IQ3_XXS; | `-mec 48` | 48 | 48 | 85% / 93% | 9.09 | 44.2 |",
   "ram_gb": "128",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #28248",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28248",
   "tg_tokens": null,
   "tps": "9.09",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "262000",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "RTX 5070 Ti 16 GB + RTX 4060 Ti 16 GB",
   "id": "lc28248-dualgpu-mec-48-56",
   "model": "DeepSeek-V4-Flash-Vision-Exp",
   "notes": "TmDagger fix-moe-pool-rail fork (persistent per-tensor VRAM expert pool -mec, per-device budget with compute-buffer rail), RTX 5070 Ti 16 GB on PCIe 5.0 x16 (CUDA0) + RTX 4060 Ti 16 GB on PCIe 4.0 x4 (CUDA1), Windows 11, 128 GB DDR4-3200, DeepSeek-V4-Flash-Vision-Exp UD-IQ3_XXS without mmproj, -c 262000 -ngl 99 --split-mode layer -ncmoe 99 --seed 42; -mec 48,56 per-device slots, hit 85%/95%; config=mec-48-56",
   "params": "284B-A13B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "44.3",
   "provenance": "community",
   "quant": "UD-IQ3_XXS",
   "quote": "2 GPUs: RTX 5070 Ti 16 GB; RTX 4060 Ti 16 GB; unsloth/DeepSeek-V4-Flash-Vision-Exp, UD-IQ3_XXS; | `-mec 48,56` | 48 | 56 | 85% / 95% | 9.71 | 44.3 |",
   "ram_gb": "128",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #28248",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28248",
   "tg_tokens": null,
   "tps": "9.71",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "262000",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "RTX 5070 Ti 16 GB + RTX 4060 Ti 16 GB",
   "id": "lc28248-dualgpu-mec-48-0",
   "model": "DeepSeek-V4-Flash-Vision-Exp",
   "notes": "TmDagger fix-moe-pool-rail fork (persistent per-tensor VRAM expert pool -mec, per-device budget with compute-buffer rail), RTX 5070 Ti 16 GB on PCIe 5.0 x16 (CUDA0) + RTX 4060 Ti 16 GB on PCIe 4.0 x4 (CUDA1), Windows 11, 128 GB DDR4-3200, DeepSeek-V4-Flash-Vision-Exp UD-IQ3_XXS without mmproj, -c 262000 -ngl 99 --split-mode layer -ncmoe 99 --seed 42; -mec 48,0 pool on CUDA0 only: the PCIe 4.0 x4 second GPU gains little from a pool; config=mec-48-0",
   "params": "284B-A13B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "45.1",
   "provenance": "community",
   "quant": "UD-IQ3_XXS",
   "quote": "2 GPUs: RTX 5070 Ti 16 GB; RTX 4060 Ti 16 GB; unsloth/DeepSeek-V4-Flash-Vision-Exp, UD-IQ3_XXS; | `-mec 48,0` | 48 | 0 | 84% / - | 6.10 | 45.1 |",
   "ram_gb": "128",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #28248",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28248",
   "tg_tokens": null,
   "tps": "6.10",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "262000",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "RTX 5070 Ti 16 GB + RTX 4060 Ti 16 GB",
   "id": "lc28248-dualgpu-mec-48-128",
   "model": "DeepSeek-V4-Flash-Vision-Exp",
   "notes": "TmDagger fix-moe-pool-rail fork (persistent per-tensor VRAM expert pool -mec, per-device budget with compute-buffer rail), RTX 5070 Ti 16 GB on PCIe 5.0 x16 (CUDA0) + RTX 4060 Ti 16 GB on PCIe 4.0 x4 (CUDA1), Windows 11, 128 GB DDR4-3200, DeepSeek-V4-Flash-Vision-Exp UD-IQ3_XXS without mmproj, -c 262000 -ngl 99 --split-mode layer -ncmoe 99 --seed 42; -mec 48,128, CUDA1 auto-scaled to ~55 slots (44%); config=mec-48-128",
   "params": "284B-A13B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "45.7",
   "provenance": "community",
   "quant": "UD-IQ3_XXS",
   "quote": "2 GPUs: RTX 5070 Ti 16 GB; RTX 4060 Ti 16 GB; unsloth/DeepSeek-V4-Flash-Vision-Exp, UD-IQ3_XXS; | `-mec 48,128` (auto-scaled) | 48 | ~55 (44%) | 84% / 95% | 9.90 | 45.7 |",
   "ram_gb": "128",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #28248",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28248",
   "tg_tokens": null,
   "tps": "9.90",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "262000",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "RTX 5070 Ti 16 GB + RTX 4060 Ti 16 GB",
   "id": "lc28248-dualgpu-mec-128-128",
   "model": "DeepSeek-V4-Flash-Vision-Exp",
   "notes": "TmDagger fix-moe-pool-rail fork (persistent per-tensor VRAM expert pool -mec, per-device budget with compute-buffer rail), RTX 5070 Ti 16 GB on PCIe 5.0 x16 (CUDA0) + RTX 4060 Ti 16 GB on PCIe 4.0 x4 (CUDA1), Windows 11, 128 GB DDR4-3200, DeepSeek-V4-Flash-Vision-Exp UD-IQ3_XXS without mmproj, -c 262000 -ngl 99 --split-mode layer -ncmoe 99 --seed 42; -mec 128,128 auto-scaled to ~47 (37%) / ~58 (46%): best result, uniform scaling beats hand-picked 48 slots (memory spill costs ~5%); config=mec-128-128",
   "params": "284B-A13B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "49.5",
   "provenance": "community",
   "quant": "UD-IQ3_XXS",
   "quote": "2 GPUs: RTX 5070 Ti 16 GB; RTX 4060 Ti 16 GB; unsloth/DeepSeek-V4-Flash-Vision-Exp, UD-IQ3_XXS; | `-mec 128,128` (auto-scaled, final) | ~47 (37%) | ~58 (46%) | 83% / 96% | 10.32 | 49.5 |",
   "ram_gb": "128",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #28248",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28248",
   "tg_tokens": null,
   "tps": "10.32",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "8192",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "GTX 1080 Ti",
   "id": "lc24528-1080ti-cache-off",
   "model": "Qwen3.6-35B-A3B",
   "notes": "batot1's moe-cache budget sweep on leloch moe-cache-pr branch 8853f0535, single GTX 1080 Ti 11 GB (sm_61, build arch 61-real), Qwen3.6-35B-A3B UD-Q8_K_XL with MTP, -c 8192 -n 512 -t 6 -b 512 -ub 512 -ngl all -sm none -fa on -ctk q8_0 -ctv q8_0 --cpu-moe --cache-ram 0 GGML_CUDA_MOE_CACHE_HOTSET=0; cache path hard-disabled baseline; config=cache-off",
   "params": "35B-A3B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q8_K_XL",
   "quote": "NVIDIA GeForce GTX 1080 Ti; Qwen3.6-35B-A3B-UDT-Q8_K_XL_MTP.gguf; | hard OFF | 0 MB | 19.32 | 100.0% | baseline |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": "512",
   "tps": "19.32",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "8192",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "GTX 1080 Ti",
   "id": "lc24528-1080ti-cache-4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "batot1's moe-cache budget sweep on leloch moe-cache-pr branch 8853f0535, single GTX 1080 Ti 11 GB (sm_61, build arch 61-real), Qwen3.6-35B-A3B UD-Q8_K_XL with MTP, -c 8192 -n 512 -t 6 -b 512 -ub 512 -ngl all -sm none -fa on -ctk q8_0 -ctv q8_0 --cpu-moe --cache-ram 0 GGML_CUDA_MOE_CACHE_HOTSET=0; 4096 MB cache budget, worst regression (-31%); config=cache-4096",
   "params": "35B-A3B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q8_K_XL",
   "quote": "NVIDIA GeForce GTX 1080 Ti; Qwen3.6-35B-A3B-UDT-Q8_K_XL_MTP.gguf; | ON | 4096 MB | 13.25 | 68.6% | regression |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": "512",
   "tps": "13.25",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "8192",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "GTX 1080 Ti",
   "id": "lc24528-1080ti-cache-1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "batot1's moe-cache budget sweep on leloch moe-cache-pr branch 8853f0535, single GTX 1080 Ti 11 GB (sm_61, build arch 61-real), Qwen3.6-35B-A3B UD-Q8_K_XL with MTP, -c 8192 -n 512 -t 6 -b 512 -ub 512 -ngl all -sm none -fa on -ctk q8_0 -ctv q8_0 --cpu-moe --cache-ram 0 GGML_CUDA_MOE_CACHE_HOTSET=0; 1024 MB cache budget; config=cache-1024",
   "params": "35B-A3B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q8_K_XL",
   "quote": "NVIDIA GeForce GTX 1080 Ti; Qwen3.6-35B-A3B-UDT-Q8_K_XL_MTP.gguf; | ON | 1024 MB | 14.38 | 74.4% | regression |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": "512",
   "tps": "14.38",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "8192",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "GTX 1080 Ti",
   "id": "lc24528-1080ti-cache-512",
   "model": "Qwen3.6-35B-A3B",
   "notes": "batot1's moe-cache budget sweep on leloch moe-cache-pr branch 8853f0535, single GTX 1080 Ti 11 GB (sm_61, build arch 61-real), Qwen3.6-35B-A3B UD-Q8_K_XL with MTP, -c 8192 -n 512 -t 6 -b 512 -ub 512 -ngl all -sm none -fa on -ctk q8_0 -ctv q8_0 --cpu-moe --cache-ram 0 GGML_CUDA_MOE_CACHE_HOTSET=0; 512 MB cache budget; config=cache-512",
   "params": "35B-A3B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q8_K_XL",
   "quote": "NVIDIA GeForce GTX 1080 Ti; Qwen3.6-35B-A3B-UDT-Q8_K_XL_MTP.gguf; | ON | 512 MB | 17.26 | 89.3% | regression |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": "512",
   "tps": "17.26",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "8192",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "GTX 1080 Ti",
   "id": "lc24528-1080ti-cache-256",
   "model": "Qwen3.6-35B-A3B",
   "notes": "batot1's moe-cache budget sweep on leloch moe-cache-pr branch 8853f0535, single GTX 1080 Ti 11 GB (sm_61, build arch 61-real), Qwen3.6-35B-A3B UD-Q8_K_XL with MTP, -c 8192 -n 512 -t 6 -b 512 -ub 512 -ngl all -sm none -fa on -ctk q8_0 -ctv q8_0 --cpu-moe --cache-ram 0 GGML_CUDA_MOE_CACHE_HOTSET=0; 256 MB cache budget; config=cache-256",
   "params": "35B-A3B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q8_K_XL",
   "quote": "NVIDIA GeForce GTX 1080 Ti; Qwen3.6-35B-A3B-UDT-Q8_K_XL_MTP.gguf; | ON | 256 MB | 17.19 | 89.0% | regression |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": "512",
   "tps": "17.19",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "8192",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "GTX 1080 Ti",
   "id": "lc24528-1080ti-cache-128",
   "model": "Qwen3.6-35B-A3B",
   "notes": "batot1's moe-cache budget sweep on leloch moe-cache-pr branch 8853f0535, single GTX 1080 Ti 11 GB (sm_61, build arch 61-real), Qwen3.6-35B-A3B UD-Q8_K_XL with MTP, -c 8192 -n 512 -t 6 -b 512 -ub 512 -ngl all -sm none -fa on -ctk q8_0 -ctv q8_0 --cpu-moe --cache-ram 0 GGML_CUDA_MOE_CACHE_HOTSET=0; 128 MB cache budget; config=cache-128",
   "params": "35B-A3B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q8_K_XL",
   "quote": "NVIDIA GeForce GTX 1080 Ti; Qwen3.6-35B-A3B-UDT-Q8_K_XL_MTP.gguf; | ON | 128 MB | 18.64 | 96.5% | regression |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": "512",
   "tps": "18.64",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "8192",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "GTX 1080 Ti",
   "id": "lc24528-1080ti-cache-64",
   "model": "Qwen3.6-35B-A3B",
   "notes": "batot1's moe-cache budget sweep on leloch moe-cache-pr branch 8853f0535, single GTX 1080 Ti 11 GB (sm_61, build arch 61-real), Qwen3.6-35B-A3B UD-Q8_K_XL with MTP, -c 8192 -n 512 -t 6 -b 512 -ub 512 -ngl all -sm none -fa on -ctk q8_0 -ctv q8_0 --cpu-moe --cache-ram 0 GGML_CUDA_MOE_CACHE_HOTSET=0; 64 MB cache budget; config=cache-64",
   "params": "35B-A3B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q8_K_XL",
   "quote": "NVIDIA GeForce GTX 1080 Ti; Qwen3.6-35B-A3B-UDT-Q8_K_XL_MTP.gguf; | ON | 64 MB | 18.71 | 96.9% | regression |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": "512",
   "tps": "18.71",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "8192",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "GTX 1080 Ti",
   "id": "lc24528-1080ti-cache-32",
   "model": "Qwen3.6-35B-A3B",
   "notes": "batot1's moe-cache budget sweep on leloch moe-cache-pr branch 8853f0535, single GTX 1080 Ti 11 GB (sm_61, build arch 61-real), Qwen3.6-35B-A3B UD-Q8_K_XL with MTP, -c 8192 -n 512 -t 6 -b 512 -ub 512 -ngl all -sm none -fa on -ctk q8_0 -ctv q8_0 --cpu-moe --cache-ram 0 GGML_CUDA_MOE_CACHE_HOTSET=0; 32 MB cache budget, smallest regression; every budget regressed vs hard OFF; config=cache-32",
   "params": "35B-A3B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q8_K_XL",
   "quote": "NVIDIA GeForce GTX 1080 Ti; Qwen3.6-35B-A3B-UDT-Q8_K_XL_MTP.gguf; | ON | 32 MB | 18.72 | 96.9% | regression |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": "512",
   "tps": "18.72",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "81920",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "2x RTX 3090 + EPYC 7663",
   "id": "lc24528-epyc-gs-mtp2-80k",
   "model": "Qwen3.8-Flash-Next",
   "notes": "GenerelSchwerz moe-cache branch on the rented server (2x RTX 3090 sm_86 across two NUMA nodes, 2x EPYC 7663, 640 GB DDR4 ECC), Qwen3.8-Flash-Next UD-IQ4_XS with shared Q8_0 MTP draft, larger 21,000/18,000 MiB cache budgets; 80k-capacity reference run, 10781 tokens at 65.00 tok/s, MTP2 acceptance 55.89%; draft device not pinned so overlap fell back to normal decode (author does not treat it as a controlled A/B vs the 256k run); config=gs-mtp2-80k",
   "params": "177B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "2x RTX 3090; EPYC 7663; Qwen3.8-Flash-Next; an earlier 80k-capacity run using larger 21,000/18,000 MiB cache budgets generated 10,781 tokens at 65.00 tok/s with 55.89% MTP2 acceptance",
   "ram_gb": "640",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528",
   "tg_tokens": "10781",
   "tps": "65.00",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc24528-5060ti-matched-cache",
   "model": "Qwen3 30B A3B",
   "notes": "mfethe1's matched-budget A/B on a single RTX 5060 Ti 16 GB + Ryzen 7 3700X (DDR4), CUDA 12.9, Windows 11, Qwen3-30B-A3B Q4_K_M (17.28 GiB, 48 layers, 128 experts/layer), leloch moe-cache branch e3096b0 same build both arms, -mg 0 -sm none fa on t 8 KV f16, ABBA order 3 reps, both arms at ~12.5 GiB VRAM; cache arm -ngl 99 --n-cpu-moe 48 --moe-cache 12288: 85.3% steady-state expert hit rate (124860/146312), zero evictions, still loses decode by 1.80x and prefill by 2.34x; config=matched-cache",
   "params": "30B-A3B",
   "power_w": null,
   "pp_tokens": "512",
   "pp_tps": "64.0",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "single RTX 5060 Ti 16 GB; Qwen3-30B-A3B Q4_K_M; | cache | `-ngl 99 --n-cpu-moe 48 --moe-cache 12288` | ~12.4 GiB | 64.0 \u00b1 0.3 t/s | 19.7 \u00b1 0.1 t/s |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528#discussioncomment-18243371",
   "tg_tokens": "128",
   "tps": "19.7",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc24528-5060ti-matched-resident",
   "model": "Qwen3 30B A3B",
   "notes": "mfethe1's matched-budget A/B on a single RTX 5060 Ti 16 GB + Ryzen 7 3700X (DDR4), CUDA 12.9, Windows 11, Qwen3-30B-A3B Q4_K_M (17.28 GiB, 48 layers, 128 experts/layer), leloch moe-cache branch e3096b0 same build both arms, -mg 0 -sm none fa on t 8 KV f16, ABBA order 3 reps, both arms at ~12.5 GiB VRAM; resident arm -ngl 33 --moe-cache off: same VRAM budget spent on resident layers instead of the cache; resident weights serve 100% of tokens with zero misses; config=matched-resident",
   "params": "30B-A3B",
   "power_w": null,
   "pp_tokens": "512",
   "pp_tps": "149.6",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "single RTX 5060 Ti 16 GB; Qwen3-30B-A3B Q4_K_M; | resident | `-ngl 33 --moe-cache off` | ~12.5 GiB | 149.6 \u00b1 0.1 t/s | 35.4 \u00b1 0.4 t/s |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528#discussioncomment-18243371",
   "tg_tokens": "128",
   "tps": "35.4",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-04-01",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc20969-4090-st-kv-f16",
   "model": "Qwen3 30B A3B",
   "notes": "AmesianX TurboQuant v1.2.0 fork on RTX 4090 24GB (CUDA 12.8, Windows 11), Qwen3-30B-A3B Q4_K_M (17.28 GiB), llama-bench pp512/tg128 mean of 3 reps; llama-bench rejects tbq KV types in this build so only baseline KV caches measured; KV cache f16 (f16 baseline / q8_0 / q4_0 K and V); config=st-kv-f16",
   "power_w": null,
   "pp_tokens": "512",
   "pp_tps": "1130.8",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "RTX 4090 24GB, CUDA 12.8, Windows 11, AmesianX TurboQuant v1.2.0; Qwen3-30B-A3B \u2014 first RTX 4090 (SM89 Ada) PPL data + throughput; | f16/f16 | 1130.8 | 225.9 |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16403244",
   "tg_tokens": "128",
   "tps": "225.9",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-04-01",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc20969-4090-st-kv-q80",
   "model": "Qwen3 30B A3B",
   "notes": "AmesianX TurboQuant v1.2.0 fork on RTX 4090 24GB (CUDA 12.8, Windows 11), Qwen3-30B-A3B Q4_K_M (17.28 GiB), llama-bench pp512/tg128 mean of 3 reps; llama-bench rejects tbq KV types in this build so only baseline KV caches measured; KV cache q80 (f16 baseline / q8_0 / q4_0 K and V); config=st-kv-q80",
   "power_w": null,
   "pp_tokens": "512",
   "pp_tps": "1132.8",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "RTX 4090 24GB, CUDA 12.8, Windows 11, AmesianX TurboQuant v1.2.0; Qwen3-30B-A3B \u2014 first RTX 4090 (SM89 Ada) PPL data + throughput; | q8_0/q8_0 | 1132.8 | 223.2 |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16403244",
   "tg_tokens": "128",
   "tps": "223.2",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-04-01",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc20969-4090-st-kv-q40",
   "model": "Qwen3 30B A3B",
   "notes": "AmesianX TurboQuant v1.2.0 fork on RTX 4090 24GB (CUDA 12.8, Windows 11), Qwen3-30B-A3B Q4_K_M (17.28 GiB), llama-bench pp512/tg128 mean of 3 reps; llama-bench rejects tbq KV types in this build so only baseline KV caches measured; KV cache q40 (f16 baseline / q8_0 / q4_0 K and V); config=st-kv-q40",
   "power_w": null,
   "pp_tokens": "512",
   "pp_tps": "1087.8",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "RTX 4090 24GB, CUDA 12.8, Windows 11, AmesianX TurboQuant v1.2.0; Qwen3-30B-A3B \u2014 first RTX 4090 (SM89 Ada) PPL data + throughput; | q4_0/q4_0 | 1087.8 | 211.0 |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16403244",
   "tg_tokens": "128",
   "tps": "211.0",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-04-03",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc20969-4090-wsf1-nofa-f16",
   "model": "Llama 3.1 8B",
   "notes": "spiritbuun feature/turboquant-kv-cache @ 5047b3e9a on RTX 4090 24GB + i9-14900K (CUDA 12.4, Windows, VS 2022), Llama 3.1 8B Instruct Q4_K_M (4.58 GiB), llama-bench pp512/tg128; built without -DGGML_CUDA_FA flags: fused FA kernels missing, turbo KV types cannot run at runtime; config=wsf1-nofa-f16",
   "power_w": null,
   "pp_tokens": "512",
   "pp_tps": "9365",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "RTX 4090 24GB, i9-14900K, CUDA 12.4, Windows, VS 2022; Model:** Llama 3.1 8B Instruct Q4_K_M (4.58 GiB); | Without FA flags | 9,365 | 150.4 |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16436834",
   "tg_tokens": "128",
   "tps": "150.4",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-04-03",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc20969-4090-wsf1-fa-f16",
   "model": "Llama 3.1 8B",
   "notes": "spiritbuun feature/turboquant-kv-cache @ 5047b3e9a on RTX 4090 24GB + i9-14900K (CUDA 12.4, Windows, VS 2022), Llama 3.1 8B Instruct Q4_K_M (4.58 GiB), llama-bench pp512/tg128; same binary rebuilt with -DGGML_CUDA_FA=ON -DGGML_CUDA_FA_ALL_QUANTS=ON: +7.4% prefill, decode unchanged; config=wsf1-fa-f16",
   "power_w": null,
   "pp_tokens": "512",
   "pp_tps": "10060",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "RTX 4090 24GB, i9-14900K, CUDA 12.4, Windows, VS 2022; Model:** Llama 3.1 8B Instruct Q4_K_M (4.58 GiB); | With FA flags | 10,060 | 149.3 |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16436834",
   "tg_tokens": "128",
   "tps": "149.3",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-04-03",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc20969-4090-wsf1-q8-t4",
   "model": "Llama 3.1 8B",
   "notes": "spiritbuun feature/turboquant-kv-cache @ 5047b3e9a on RTX 4090 24GB + i9-14900K (CUDA 12.4, Windows, VS 2022), Llama 3.1 8B Instruct Q4_K_M (4.58 GiB), llama-bench pp512/tg128; q8_0 K + turbo4 V: fastest prefill of any config (+8.4% vs f16) with only 6.2% decode cost, author's recommended config; config=wsf1-q8-t4",
   "power_w": null,
   "pp_tokens": "512",
   "pp_tps": "10903",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "RTX 4090 24GB, i9-14900K, CUDA 12.4, Windows, VS 2022; Model:** Llama 3.1 8B Instruct Q4_K_M (4.58 GiB); | **q8_0 / turbo4** | **10,903 \u00b1 545** | **140.0 \u00b1 0.2** | +8.4% | \u22126.2% |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16436834",
   "tg_tokens": "128",
   "tps": "140.0",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-04-03",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc20969-4090-wsf1-q8-t3",
   "model": "Llama 3.1 8B",
   "notes": "spiritbuun feature/turboquant-kv-cache @ 5047b3e9a on RTX 4090 24GB + i9-14900K (CUDA 12.4, Windows, VS 2022), Llama 3.1 8B Instruct Q4_K_M (4.58 GiB), llama-bench pp512/tg128; q8_0 K + turbo3 V; config=wsf1-q8-t3",
   "power_w": null,
   "pp_tokens": "512",
   "pp_tps": "10925",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "RTX 4090 24GB, i9-14900K, CUDA 12.4, Windows, VS 2022; Model:** Llama 3.1 8B Instruct Q4_K_M (4.58 GiB); | q8_0 / turbo3 | 10,925 \u00b1 503 | 138.4 \u00b1 0.2 | +8.6% | \u22127.3% |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16436834",
   "tg_tokens": "128",
   "tps": "138.4",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-04-03",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc20969-4090-wsf1-t3-t3",
   "model": "Llama 3.1 8B",
   "notes": "spiritbuun feature/turboquant-kv-cache @ 5047b3e9a on RTX 4090 24GB + i9-14900K (CUDA 12.4, Windows, VS 2022), Llama 3.1 8B Instruct Q4_K_M (4.58 GiB), llama-bench pp512/tg128; symmetric turbo3 (adaptive=1); beats symmetric turbo4 at decode; config=wsf1-t3-t3",
   "power_w": null,
   "pp_tokens": "512",
   "pp_tps": "10476",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "RTX 4090 24GB, i9-14900K, CUDA 12.4, Windows, VS 2022; Model:** Llama 3.1 8B Instruct Q4_K_M (4.58 GiB); | turbo3 / turbo3 (adaptive=1) | 10,476 \u00b1 432 | 128.3 \u00b1 0.2 | +4.1% | \u221214.1% |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16436834",
   "tg_tokens": "128",
   "tps": "128.3",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-04-03",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc20969-4090-wsf1-t4-t4",
   "model": "Llama 3.1 8B",
   "notes": "spiritbuun feature/turboquant-kv-cache @ 5047b3e9a on RTX 4090 24GB + i9-14900K (CUDA 12.4, Windows, VS 2022), Llama 3.1 8B Instruct Q4_K_M (4.58 GiB), llama-bench pp512/tg128; symmetric turbo4: worst decode of the matrix; config=wsf1-t4-t4",
   "power_w": null,
   "pp_tokens": "512",
   "pp_tps": "10405",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "RTX 4090 24GB, i9-14900K, CUDA 12.4, Windows, VS 2022; Model:** Llama 3.1 8B Instruct Q4_K_M (4.58 GiB); | turbo4 / turbo4 | 10,405 \u00b1 496 | 124.3 \u00b1 0.1 | +3.4% | \u221216.8% |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16436834",
   "tg_tokens": "128",
   "tps": "124.3",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-04-04",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc20969-wsf2-4090-off-f16",
   "model": "Llama 3.1 8B",
   "notes": "spiritbuun llama-cpp-turboquant-cuda v2, CUDA 12.8 FA flags on, AMD 9950X3D + 64 GB DDR5, Meta-Llama-3.1-8B-Instruct-Q4_K_M (4.58 GiB), llama-bench pp512/tg128; CUDA 13.2 sm_120a identical to CUDA 12.8 sm_120 within noise; LA=off, f16 KV baseline; config=wsf2-4090-off-f16",
   "power_w": null,
   "pp_tokens": "512",
   "pp_tps": "10112.60",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "AMD 9950X3D \u00b7 64 GB DDR5 \u00b7 RTX 5090 (32 GB) + RTX 4090 (24 GB); Meta-Llama-3.1-8B-Instruct-Q4_K_M (4.58 GiB); | f16 (default)     | 10,112.60   | 150.59",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16447542",
   "tg_tokens": "128",
   "tps": "150.59",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-04-04",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc20969-wsf2-4090-off-q8t4",
   "model": "Llama 3.1 8B",
   "notes": "spiritbuun llama-cpp-turboquant-cuda v2, CUDA 12.8 FA flags on, AMD 9950X3D + 64 GB DDR5, Meta-Llama-3.1-8B-Instruct-Q4_K_M (4.58 GiB), llama-bench pp512/tg128; CUDA 13.2 sm_120a identical to CUDA 12.8 sm_120 within noise; LA=off, q8_0 K + turbo4 V; config=wsf2-4090-off-q8t4",
   "power_w": null,
   "pp_tokens": "512",
   "pp_tps": "10808.36",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "AMD 9950X3D \u00b7 64 GB DDR5 \u00b7 RTX 5090 (32 GB) + RTX 4090 (24 GB); Meta-Llama-3.1-8B-Instruct-Q4_K_M (4.58 GiB); | q8_0-K + turbo4-V | 10,808.36   | 136.37",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16447542",
   "tg_tokens": "128",
   "tps": "136.37",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-04-04",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc20969-wsf2-4090-off-t3",
   "model": "Llama 3.1 8B",
   "notes": "spiritbuun llama-cpp-turboquant-cuda v2, CUDA 12.8 FA flags on, AMD 9950X3D + 64 GB DDR5, Meta-Llama-3.1-8B-Instruct-Q4_K_M (4.58 GiB), llama-bench pp512/tg128; CUDA 13.2 sm_120a identical to CUDA 12.8 sm_120 within noise; LA=off, symmetric turbo3; config=wsf2-4090-off-t3",
   "power_w": null,
   "pp_tokens": "512",
   "pp_tps": "10154.79",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "AMD 9950X3D \u00b7 64 GB DDR5 \u00b7 RTX 5090 (32 GB) + RTX 4090 (24 GB); Meta-Llama-3.1-8B-Instruct-Q4_K_M (4.58 GiB); | turbo3 / turbo3   | 10,154.79   | 122.68",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16447542",
   "tg_tokens": "128",
   "tps": "122.68",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-04-04",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc20969-wsf2-4090-off-t4",
   "model": "Llama 3.1 8B",
   "notes": "spiritbuun llama-cpp-turboquant-cuda v2, CUDA 12.8 FA flags on, AMD 9950X3D + 64 GB DDR5, Meta-Llama-3.1-8B-Instruct-Q4_K_M (4.58 GiB), llama-bench pp512/tg128; CUDA 13.2 sm_120a identical to CUDA 12.8 sm_120 within noise; LA=off, symmetric turbo4; config=wsf2-4090-off-t4",
   "power_w": null,
   "pp_tokens": "512",
   "pp_tps": "10384.88",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "AMD 9950X3D \u00b7 64 GB DDR5 \u00b7 RTX 5090 (32 GB) + RTX 4090 (24 GB); Meta-Llama-3.1-8B-Instruct-Q4_K_M (4.58 GiB); | turbo4 / turbo4   | 10,384.88   | 123.95",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16447542",
   "tg_tokens": "128",
   "tps": "123.95",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-04-04",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc20969-wsf2-4090-off-q8t3",
   "model": "Llama 3.1 8B",
   "notes": "spiritbuun llama-cpp-turboquant-cuda v2, CUDA 12.8 FA flags on, AMD 9950X3D + 64 GB DDR5, Meta-Llama-3.1-8B-Instruct-Q4_K_M (4.58 GiB), llama-bench pp512/tg128; CUDA 13.2 sm_120a identical to CUDA 12.8 sm_120 within noise; LA=off, q8_0 K + turbo3 V; config=wsf2-4090-off-q8t3",
   "power_w": null,
   "pp_tokens": "512",
   "pp_tps": "10794.80",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "AMD 9950X3D \u00b7 64 GB DDR5 \u00b7 RTX 5090 (32 GB) + RTX 4090 (24 GB); Meta-Llama-3.1-8B-Instruct-Q4_K_M (4.58 GiB); | q8_0-K + turbo3-V | 10,794.80   | 134.73",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16447542",
   "tg_tokens": "128",
   "tps": "134.73",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-04-04",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc20969-wsf2-4090-la1-f16",
   "model": "Llama 3.1 8B",
   "notes": "spiritbuun llama-cpp-turboquant-cuda v2, CUDA 12.8 FA flags on, AMD 9950X3D + 64 GB DDR5, Meta-Llama-3.1-8B-Instruct-Q4_K_M (4.58 GiB), llama-bench pp512/tg128; CUDA 13.2 sm_120a identical to CUDA 12.8 sm_120 within noise; LA=1 (first 4 + last 4 layers at q8_0), f16 KV baseline; config=wsf2-4090-la1-f16",
   "power_w": null,
   "pp_tokens": "512",
   "pp_tps": "10076.76",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "AMD 9950X3D \u00b7 64 GB DDR5 \u00b7 RTX 5090 (32 GB) + RTX 4090 (24 GB); Meta-Llama-3.1-8B-Instruct-Q4_K_M (4.58 GiB); | f16 (default)     | 10,076.76   | 150.17",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16447542",
   "tg_tokens": "128",
   "tps": "150.17",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-04-04",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc20969-wsf2-4090-la1-q8t4",
   "model": "Llama 3.1 8B",
   "notes": "spiritbuun llama-cpp-turboquant-cuda v2, CUDA 12.8 FA flags on, AMD 9950X3D + 64 GB DDR5, Meta-Llama-3.1-8B-Instruct-Q4_K_M (4.58 GiB), llama-bench pp512/tg128; CUDA 13.2 sm_120a identical to CUDA 12.8 sm_120 within noise; LA=1, q8_0 K + turbo4 V: author's recommended config, decode recovers from 136.4 to 139.8; config=wsf2-4090-la1-q8t4",
   "power_w": null,
   "pp_tokens": "512",
   "pp_tps": "10957.29",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "AMD 9950X3D \u00b7 64 GB DDR5 \u00b7 RTX 5090 (32 GB) + RTX 4090 (24 GB); Meta-Llama-3.1-8B-Instruct-Q4_K_M (4.58 GiB); | q8_0-K + turbo4-V | 10,957.29   | 139.83",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16447542",
   "tg_tokens": "128",
   "tps": "139.83",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-04-04",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc20969-wsf2-4090-la1-t3",
   "model": "Llama 3.1 8B",
   "notes": "spiritbuun llama-cpp-turboquant-cuda v2, CUDA 12.8 FA flags on, AMD 9950X3D + 64 GB DDR5, Meta-Llama-3.1-8B-Instruct-Q4_K_M (4.58 GiB), llama-bench pp512/tg128; CUDA 13.2 sm_120a identical to CUDA 12.8 sm_120 within noise; LA=1, symmetric turbo3; config=wsf2-4090-la1-t3",
   "power_w": null,
   "pp_tokens": "512",
   "pp_tps": "10456.24",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "AMD 9950X3D \u00b7 64 GB DDR5 \u00b7 RTX 5090 (32 GB) + RTX 4090 (24 GB); Meta-Llama-3.1-8B-Instruct-Q4_K_M (4.58 GiB); | turbo3 / turbo3   | 10,456.24   | 127.79",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16447542",
   "tg_tokens": "128",
   "tps": "127.79",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-04-04",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc20969-wsf2-4090-la1-t4",
   "model": "Llama 3.1 8B",
   "notes": "spiritbuun llama-cpp-turboquant-cuda v2, CUDA 12.8 FA flags on, AMD 9950X3D + 64 GB DDR5, Meta-Llama-3.1-8B-Instruct-Q4_K_M (4.58 GiB), llama-bench pp512/tg128; CUDA 13.2 sm_120a identical to CUDA 12.8 sm_120 within noise; LA=1, symmetric turbo4; config=wsf2-4090-la1-t4",
   "power_w": null,
   "pp_tokens": "512",
   "pp_tps": "10571.92",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "AMD 9950X3D \u00b7 64 GB DDR5 \u00b7 RTX 5090 (32 GB) + RTX 4090 (24 GB); Meta-Llama-3.1-8B-Instruct-Q4_K_M (4.58 GiB); | turbo4 / turbo4   | 10,571.92   | 129.26",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16447542",
   "tg_tokens": "128",
   "tps": "129.26",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-04-04",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc20969-wsf2-4090-la1-q8t3",
   "model": "Llama 3.1 8B",
   "notes": "spiritbuun llama-cpp-turboquant-cuda v2, CUDA 12.8 FA flags on, AMD 9950X3D + 64 GB DDR5, Meta-Llama-3.1-8B-Instruct-Q4_K_M (4.58 GiB), llama-bench pp512/tg128; CUDA 13.2 sm_120a identical to CUDA 12.8 sm_120 within noise; LA=1, q8_0 K + turbo3 V; config=wsf2-4090-la1-q8t3",
   "power_w": null,
   "pp_tokens": "512",
   "pp_tps": "10936.64",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "AMD 9950X3D \u00b7 64 GB DDR5 \u00b7 RTX 5090 (32 GB) + RTX 4090 (24 GB); Meta-Llama-3.1-8B-Instruct-Q4_K_M (4.58 GiB); | q8_0-K + turbo3-V | 10,936.64   | 137.81",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16447542",
   "tg_tokens": "128",
   "tps": "137.81",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-04-04",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc20969-wsf2-5090-off-f16",
   "model": "Llama 3.1 8B",
   "notes": "spiritbuun llama-cpp-turboquant-cuda v2, CUDA 12.8 FA flags on, AMD 9950X3D + 64 GB DDR5, Meta-Llama-3.1-8B-Instruct-Q4_K_M (4.58 GiB), llama-bench pp512/tg128; CUDA 13.2 sm_120a identical to CUDA 12.8 sm_120 within noise; LA=off, f16 KV baseline; config=wsf2-5090-off-f16",
   "power_w": null,
   "pp_tokens": "512",
   "pp_tps": "12850.64",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "AMD 9950X3D \u00b7 64 GB DDR5 \u00b7 RTX 5090 (32 GB) + RTX 4090 (24 GB); Meta-Llama-3.1-8B-Instruct-Q4_K_M (4.58 GiB); | f16 (default)     | 12,850.64   | 230.70",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16447542",
   "tg_tokens": "128",
   "tps": "230.70",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-04-04",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc20969-wsf2-5090-off-q8t4",
   "model": "Llama 3.1 8B",
   "notes": "spiritbuun llama-cpp-turboquant-cuda v2, CUDA 12.8 FA flags on, AMD 9950X3D + 64 GB DDR5, Meta-Llama-3.1-8B-Instruct-Q4_K_M (4.58 GiB), llama-bench pp512/tg128; CUDA 13.2 sm_120a identical to CUDA 12.8 sm_120 within noise; LA=off, q8_0 K + turbo4 V: Blackwell loses 29% decode, vs 6-9% on Ada; author attributes it to TurboQuant's dequant path matching Ada int8 dp4a but not Blackwell fp8/fp4 tensor cores; config=wsf2-5090-off-q8t4",
   "power_w": null,
   "pp_tokens": "512",
   "pp_tps": "12388.47",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "AMD 9950X3D \u00b7 64 GB DDR5 \u00b7 RTX 5090 (32 GB) + RTX 4090 (24 GB); Meta-Llama-3.1-8B-Instruct-Q4_K_M (4.58 GiB); | q8_0-K + turbo4-V | 12,388.47   | 162.95",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16447542",
   "tg_tokens": "128",
   "tps": "162.95",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-04-04",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc20969-wsf2-5090-off-t3",
   "model": "Llama 3.1 8B",
   "notes": "spiritbuun llama-cpp-turboquant-cuda v2, CUDA 12.8 FA flags on, AMD 9950X3D + 64 GB DDR5, Meta-Llama-3.1-8B-Instruct-Q4_K_M (4.58 GiB), llama-bench pp512/tg128; CUDA 13.2 sm_120a identical to CUDA 12.8 sm_120 within noise; LA=off, symmetric turbo3: -45.6% decode vs f16; config=wsf2-5090-off-t3",
   "power_w": null,
   "pp_tokens": "512",
   "pp_tps": "11309.70",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "AMD 9950X3D \u00b7 64 GB DDR5 \u00b7 RTX 5090 (32 GB) + RTX 4090 (24 GB); Meta-Llama-3.1-8B-Instruct-Q4_K_M (4.58 GiB); | turbo3 / turbo3   | 11,309.70   | 125.44",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16447542",
   "tg_tokens": "128",
   "tps": "125.44",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-04-04",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc20969-wsf2-5090-off-t4",
   "model": "Llama 3.1 8B",
   "notes": "spiritbuun llama-cpp-turboquant-cuda v2, CUDA 12.8 FA flags on, AMD 9950X3D + 64 GB DDR5, Meta-Llama-3.1-8B-Instruct-Q4_K_M (4.58 GiB), llama-bench pp512/tg128; CUDA 13.2 sm_120a identical to CUDA 12.8 sm_120 within noise; LA=off, symmetric turbo4; config=wsf2-5090-off-t4",
   "power_w": null,
   "pp_tokens": "512",
   "pp_tps": "11914.27",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "AMD 9950X3D \u00b7 64 GB DDR5 \u00b7 RTX 5090 (32 GB) + RTX 4090 (24 GB); Meta-Llama-3.1-8B-Instruct-Q4_K_M (4.58 GiB); | turbo4 / turbo4   | 11,914.27   | 134.12",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16447542",
   "tg_tokens": "128",
   "tps": "134.12",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-04-04",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc20969-wsf2-5090-off-q8t3",
   "model": "Llama 3.1 8B",
   "notes": "spiritbuun llama-cpp-turboquant-cuda v2, CUDA 12.8 FA flags on, AMD 9950X3D + 64 GB DDR5, Meta-Llama-3.1-8B-Instruct-Q4_K_M (4.58 GiB), llama-bench pp512/tg128; CUDA 13.2 sm_120a identical to CUDA 12.8 sm_120 within noise; LA=off, q8_0 K + turbo3 V; config=wsf2-5090-off-q8t3",
   "power_w": null,
   "pp_tokens": "512",
   "pp_tps": "12369.44",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "AMD 9950X3D \u00b7 64 GB DDR5 \u00b7 RTX 5090 (32 GB) + RTX 4090 (24 GB); Meta-Llama-3.1-8B-Instruct-Q4_K_M (4.58 GiB); | q8_0-K + turbo3-V | 12,369.44   | 153.54",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16447542",
   "tg_tokens": "128",
   "tps": "153.54",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-04-04",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc20969-wsf2-5090-la1-f16",
   "model": "Llama 3.1 8B",
   "notes": "spiritbuun llama-cpp-turboquant-cuda v2, CUDA 12.8 FA flags on, AMD 9950X3D + 64 GB DDR5, Meta-Llama-3.1-8B-Instruct-Q4_K_M (4.58 GiB), llama-bench pp512/tg128; CUDA 13.2 sm_120a identical to CUDA 12.8 sm_120 within noise; LA=1, f16 KV baseline; config=wsf2-5090-la1-f16",
   "power_w": null,
   "pp_tokens": "512",
   "pp_tps": "12761.21",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "AMD 9950X3D \u00b7 64 GB DDR5 \u00b7 RTX 5090 (32 GB) + RTX 4090 (24 GB); Meta-Llama-3.1-8B-Instruct-Q4_K_M (4.58 GiB); | f16 (default)     | 12,761.21   | 231.93",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16447542",
   "tg_tokens": "128",
   "tps": "231.93",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-04-04",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc20969-wsf2-5090-la1-q8t4",
   "model": "Llama 3.1 8B",
   "notes": "spiritbuun llama-cpp-turboquant-cuda v2, CUDA 12.8 FA flags on, AMD 9950X3D + 64 GB DDR5, Meta-Llama-3.1-8B-Instruct-Q4_K_M (4.58 GiB), llama-bench pp512/tg128; CUDA 13.2 sm_120a identical to CUDA 12.8 sm_120 within noise; LA=1, q8_0 K + turbo4 V; config=wsf2-5090-la1-q8t4",
   "power_w": null,
   "pp_tokens": "512",
   "pp_tps": "12352.17",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "AMD 9950X3D \u00b7 64 GB DDR5 \u00b7 RTX 5090 (32 GB) + RTX 4090 (24 GB); Meta-Llama-3.1-8B-Instruct-Q4_K_M (4.58 GiB); | q8_0-K + turbo4-V | 12,352.17   | 173.43",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16447542",
   "tg_tokens": "128",
   "tps": "173.43",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-04-04",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc20969-wsf2-5090-la1-t3",
   "model": "Llama 3.1 8B",
   "notes": "spiritbuun llama-cpp-turboquant-cuda v2, CUDA 12.8 FA flags on, AMD 9950X3D + 64 GB DDR5, Meta-Llama-3.1-8B-Instruct-Q4_K_M (4.58 GiB), llama-bench pp512/tg128; CUDA 13.2 sm_120a identical to CUDA 12.8 sm_120 within noise; LA=1, symmetric turbo3; config=wsf2-5090-la1-t3",
   "power_w": null,
   "pp_tokens": "512",
   "pp_tps": "11799.41",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "AMD 9950X3D \u00b7 64 GB DDR5 \u00b7 RTX 5090 (32 GB) + RTX 4090 (24 GB); Meta-Llama-3.1-8B-Instruct-Q4_K_M (4.58 GiB); | turbo3 / turbo3   | 11,799.41   | 139.90",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16447542",
   "tg_tokens": "128",
   "tps": "139.90",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-04-04",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc20969-wsf2-5090-la1-t4",
   "model": "Llama 3.1 8B",
   "notes": "spiritbuun llama-cpp-turboquant-cuda v2, CUDA 12.8 FA flags on, AMD 9950X3D + 64 GB DDR5, Meta-Llama-3.1-8B-Instruct-Q4_K_M (4.58 GiB), llama-bench pp512/tg128; CUDA 13.2 sm_120a identical to CUDA 12.8 sm_120 within noise; LA=1, symmetric turbo4; config=wsf2-5090-la1-t4",
   "power_w": null,
   "pp_tokens": "512",
   "pp_tps": "11739.74",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "AMD 9950X3D \u00b7 64 GB DDR5 \u00b7 RTX 5090 (32 GB) + RTX 4090 (24 GB); Meta-Llama-3.1-8B-Instruct-Q4_K_M (4.58 GiB); | turbo4 / turbo4   | 11,739.74   | 144.22",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16447542",
   "tg_tokens": "128",
   "tps": "144.22",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-04-04",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc20969-wsf2-5090-la1-q8t3",
   "model": "Llama 3.1 8B",
   "notes": "spiritbuun llama-cpp-turboquant-cuda v2, CUDA 12.8 FA flags on, AMD 9950X3D + 64 GB DDR5, Meta-Llama-3.1-8B-Instruct-Q4_K_M (4.58 GiB), llama-bench pp512/tg128; CUDA 13.2 sm_120a identical to CUDA 12.8 sm_120 within noise; LA=1, q8_0 K + turbo3 V; config=wsf2-5090-la1-q8t3",
   "power_w": null,
   "pp_tokens": "512",
   "pp_tps": "12420.45",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "AMD 9950X3D \u00b7 64 GB DDR5 \u00b7 RTX 5090 (32 GB) + RTX 4090 (24 GB); Meta-Llama-3.1-8B-Instruct-Q4_K_M (4.58 GiB); | q8_0-K + turbo3-V | 12,420.45   | 161.79",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16447542",
   "tg_tokens": "128",
   "tps": "161.79",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "32768",
   "date": "2026-04-01",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc20969-5090-md59-q8",
   "model": "Qwen 3.5 27B",
   "notes": "Madreag/turbo3-cuda release/cuda-optimized fork, RTX 5090, Qwen 3.5 27B Q6_K, llama-bench tg128 at 32K depth (-d 32768), pp512 from the prefill scaling table; q8_0 KV baseline, short-context tg128 63.40, prefill 3512 pp512 falling to 2839 at pp32768; config=md59-q8",
   "power_w": null,
   "pp_tokens": "512",
   "pp_tps": "3512",
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "Key results (RTX 5090, Qwen 3.5 27B Q6_K); | q8_0 | 8.5 | 1.88\u00d7 | 63.40 | 55.60 | baseline |; | pp512 | 3,512 | 3,548 | 3,547 | 3,649 | 3,577 |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16404751",
   "tg_tokens": "128",
   "tps": "55.60",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "32768",
   "date": "2026-04-01",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc20969-5090-md59-t4",
   "model": "Qwen 3.5 27B",
   "notes": "Madreag/turbo3-cuda release/cuda-optimized fork, RTX 5090, Qwen 3.5 27B Q6_K, llama-bench tg128 at 32K depth (-d 32768), pp512 from the prefill scaling table; turbo4 KV 3.76x compression, +2.0% vs q8_0 at 32K, short-context 63.70; config=md59-t4",
   "power_w": null,
   "pp_tokens": "512",
   "pp_tps": "3548",
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "Key results (RTX 5090, Qwen 3.5 27B Q6_K); | turbo4 | 4.25 | 3.76\u00d7 | 63.70 | **56.73** | **+2.0%** |; | pp512 | 3,512 | 3,548 | 3,547 | 3,649 | 3,577 |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16404751",
   "tg_tokens": "128",
   "tps": "56.73",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "32768",
   "date": "2026-04-01",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc20969-5090-md59-t3",
   "model": "Qwen 3.5 27B",
   "notes": "Madreag/turbo3-cuda release/cuda-optimized fork, RTX 5090, Qwen 3.5 27B Q6_K, llama-bench tg128 at 32K depth (-d 32768), pp512 from the prefill scaling table; turbo3 KV 5.12x compression, +0.4% vs q8_0 at 32K, short-context 63.55; config=md59-t3",
   "power_w": null,
   "pp_tokens": "512",
   "pp_tps": "3547",
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "Key results (RTX 5090, Qwen 3.5 27B Q6_K); | turbo3 | 3.125 | 5.12\u00d7 | 63.55 | **55.84** | **+0.4%** |; | pp512 | 3,512 | 3,548 | 3,547 | 3,649 | 3,577 |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16404751",
   "tg_tokens": "128",
   "tps": "55.84",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "32768",
   "date": "2026-04-01",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc20969-5090-md59-t2",
   "model": "Qwen 3.5 27B",
   "notes": "Madreag/turbo3-cuda release/cuda-optimized fork, RTX 5090, Qwen 3.5 27B Q6_K, llama-bench tg128 at 32K depth (-d 32768), pp512 from the prefill scaling table; turbo2 KV 7.53x compression, the 32K champion at +5.4% vs q8_0 despite compression, short-context 65.50; config=md59-t2",
   "power_w": null,
   "pp_tokens": "512",
   "pp_tps": "3649",
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "Key results (RTX 5090, Qwen 3.5 27B Q6_K); | **turbo2** | **2.125** | **7.53\u00d7** | **65.50** | **58.61** | **+5.4%** |; | pp512 | 3,512 | 3,548 | 3,547 | 3,649 | 3,577 |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16404751",
   "tg_tokens": "128",
   "tps": "58.61",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "32768",
   "date": "2026-04-01",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc20969-5090-md59-t15",
   "model": "Qwen 3.5 27B",
   "notes": "Madreag/turbo3-cuda release/cuda-optimized fork, RTX 5090, Qwen 3.5 27B Q6_K, llama-bench tg128 at 32K depth (-d 32768), pp512 from the prefill scaling table; turbo1.5 KV 8x compression, -0.8% vs q8_0 at 32K, short-context 63.13; config=md59-t15",
   "power_w": null,
   "pp_tokens": "512",
   "pp_tps": "3577",
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "Key results (RTX 5090, Qwen 3.5 27B Q6_K); | turbo1.5 | 2.00 | 8.0\u00d7 | 63.13 | 55.16 | -0.8% |; | pp512 | 3,512 | 3,548 | 3,547 | 3,649 | 3,577 |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16404751",
   "tg_tokens": "128",
   "tps": "55.16",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "32768",
   "date": "2026-04-01",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc20969-5090-md59b-t4",
   "model": "Qwen 3.5 27B",
   "notes": "Madreag/turbo3-cuda release/cuda-optimized fork, RTX 5090, Qwen 3.5 27B Q6_K, llama-bench tg128 at 32K depth (-d 32768), pp512 from the prefill scaling table; base TurboQuant fork (TheTom feature/turboquant-kv-cache, unoptimized CUDA kernels) t4 at 32K, before the kernel rewrite; config=md59b-t4",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "Key results (RTX 5090, Qwen 3.5 27B Q6_K); | turbo4 32K | 38.88 | 56.73 | **+45.9%** |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16404751",
   "tg_tokens": "128",
   "tps": "38.88",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "32768",
   "date": "2026-04-01",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc20969-5090-md59b-t3",
   "model": "Qwen 3.5 27B",
   "notes": "Madreag/turbo3-cuda release/cuda-optimized fork, RTX 5090, Qwen 3.5 27B Q6_K, llama-bench tg128 at 32K depth (-d 32768), pp512 from the prefill scaling table; base TurboQuant fork (TheTom feature/turboquant-kv-cache, unoptimized CUDA kernels) t3 at 32K, before the kernel rewrite; config=md59b-t3",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "Key results (RTX 5090, Qwen 3.5 27B Q6_K); | turbo3 32K | 46.62 | 55.84 | **+19.8%** |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16404751",
   "tg_tokens": "128",
   "tps": "46.62",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "32768",
   "date": "2026-04-01",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc20969-5090-md59b-t2",
   "model": "Qwen 3.5 27B",
   "notes": "Madreag/turbo3-cuda release/cuda-optimized fork, RTX 5090, Qwen 3.5 27B Q6_K, llama-bench tg128 at 32K depth (-d 32768), pp512 from the prefill scaling table; base TurboQuant fork (TheTom feature/turboquant-kv-cache, unoptimized CUDA kernels) t2 at 32K, before the kernel rewrite; config=md59b-t2",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "Key results (RTX 5090, Qwen 3.5 27B Q6_K); | turbo2 32K | 51.69 | 58.61 | **+13.4%** |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16404751",
   "tg_tokens": "128",
   "tps": "51.69",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "32768",
   "date": "2026-04-01",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc20969-3090-md39-q8",
   "model": "Qwen 3.5 9B",
   "notes": "Madreag/turbo3-cuda optimized fork vs base TurboQuant fork, RTX 3090, Qwen 3.5 9B Q8_0, llama-bench tg128 at 32K/64K depth, back-to-back same GPU same model; q8_0 KV at 32K, optimized fork; config=md39-q8",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "### RTX 3090 (9B Q8_0); Qwen 3.5 9B Q8_0; | q8_0 32K | 56.91 | 61.0 | **+7.2%** |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16404751",
   "tg_tokens": "128",
   "tps": "61.0",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "32768",
   "date": "2026-04-01",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc20969-3090-md39-t4",
   "model": "Qwen 3.5 9B",
   "notes": "Madreag/turbo3-cuda optimized fork vs base TurboQuant fork, RTX 3090, Qwen 3.5 9B Q8_0, llama-bench tg128 at 32K/64K depth, back-to-back same GPU same model; turbo4 KV at 32K, optimized fork: biggest gain of the whole study, +69% over the base fork; config=md39-t4",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "### RTX 3090 (9B Q8_0); Qwen 3.5 9B Q8_0; | turbo4 32K | 35.63 | 60.28 | **+69%** |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16404751",
   "tg_tokens": "128",
   "tps": "60.28",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "32768",
   "date": "2026-04-01",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc20969-3090-md39-t3",
   "model": "Qwen 3.5 9B",
   "notes": "Madreag/turbo3-cuda optimized fork vs base TurboQuant fork, RTX 3090, Qwen 3.5 9B Q8_0, llama-bench tg128 at 32K/64K depth, back-to-back same GPU same model; turbo3 KV at 32K, optimized fork; config=md39-t3",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "### RTX 3090 (9B Q8_0); Qwen 3.5 9B Q8_0; | turbo3 32K | 44.79 | 56.82 | **+27%** |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16404751",
   "tg_tokens": "128",
   "tps": "56.82",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "32768",
   "date": "2026-04-01",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc20969-3090-md39-t2",
   "model": "Qwen 3.5 9B",
   "notes": "Madreag/turbo3-cuda optimized fork vs base TurboQuant fork, RTX 3090, Qwen 3.5 9B Q8_0, llama-bench tg128 at 32K/64K depth, back-to-back same GPU same model; turbo2 KV at 32K, optimized fork, fastest of this table; config=md39-t2",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "### RTX 3090 (9B Q8_0); Qwen 3.5 9B Q8_0; | turbo2 32K | 53.21 | 63.12 | **+19%** |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16404751",
   "tg_tokens": "128",
   "tps": "63.12",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "65536",
   "date": "2026-04-01",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc20969-3090-md39-t3-64",
   "model": "Qwen 3.5 9B",
   "notes": "Madreag/turbo3-cuda optimized fork vs base TurboQuant fork, RTX 3090, Qwen 3.5 9B Q8_0, llama-bench tg128 at 32K/64K depth, back-to-back same GPU same model; turbo3 KV at 64K, optimized fork; config=md39-t3-64",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "### RTX 3090 (9B Q8_0); Qwen 3.5 9B Q8_0; | turbo3 64K | 33.43 | 49.27 | **+47%** |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16404751",
   "tg_tokens": "128",
   "tps": "49.27",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "65536",
   "date": "2026-04-01",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc20969-3090-md39-t2-64",
   "model": "Qwen 3.5 9B",
   "notes": "Madreag/turbo3-cuda optimized fork vs base TurboQuant fork, RTX 3090, Qwen 3.5 9B Q8_0, llama-bench tg128 at 32K/64K depth, back-to-back same GPU same model; turbo2 KV at 64K, optimized fork; config=md39-t2-64",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "### RTX 3090 (9B Q8_0); Qwen 3.5 9B Q8_0; | turbo2 64K | 42.45 | 56.91 | **+34%** |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16404751",
   "tg_tokens": "128",
   "tps": "56.91",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "32768",
   "date": "2026-04-01",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc20969-3090-md39b-q8",
   "model": "Qwen 3.5 9B",
   "notes": "Madreag/turbo3-cuda optimized fork vs base TurboQuant fork, RTX 3090, Qwen 3.5 9B Q8_0, llama-bench tg128 at 32K/64K depth, back-to-back same GPU same model; base fork q8_0 KV at 32K; config=md39b-q8",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "### RTX 3090 (9B Q8_0); Qwen 3.5 9B Q8_0; | q8_0 32K | 56.91 | 61.0 | **+7.2%** |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16404751",
   "tg_tokens": "128",
   "tps": "56.91",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "32768",
   "date": "2026-04-01",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc20969-3090-md39b-t4",
   "model": "Qwen 3.5 9B",
   "notes": "Madreag/turbo3-cuda optimized fork vs base TurboQuant fork, RTX 3090, Qwen 3.5 9B Q8_0, llama-bench tg128 at 32K/64K depth, back-to-back same GPU same model; base fork turbo4 KV at 32K, 41% below q8_0 before optimization; config=md39b-t4",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "### RTX 3090 (9B Q8_0); Qwen 3.5 9B Q8_0; | turbo4 32K | 35.63 | 60.28 | **+69%** |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16404751",
   "tg_tokens": "128",
   "tps": "35.63",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "32768",
   "date": "2026-04-01",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc20969-3090-md39b-t3",
   "model": "Qwen 3.5 9B",
   "notes": "Madreag/turbo3-cuda optimized fork vs base TurboQuant fork, RTX 3090, Qwen 3.5 9B Q8_0, llama-bench tg128 at 32K/64K depth, back-to-back same GPU same model; base fork turbo3 KV at 32K; config=md39b-t3",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "### RTX 3090 (9B Q8_0); Qwen 3.5 9B Q8_0; | turbo3 32K | 44.79 | 56.82 | **+27%** |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16404751",
   "tg_tokens": "128",
   "tps": "44.79",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "32768",
   "date": "2026-04-01",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc20969-3090-md39b-t2",
   "model": "Qwen 3.5 9B",
   "notes": "Madreag/turbo3-cuda optimized fork vs base TurboQuant fork, RTX 3090, Qwen 3.5 9B Q8_0, llama-bench tg128 at 32K/64K depth, back-to-back same GPU same model; base fork turbo2 KV at 32K; config=md39b-t2",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "### RTX 3090 (9B Q8_0); Qwen 3.5 9B Q8_0; | turbo2 32K | 53.21 | 63.12 | **+19%** |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16404751",
   "tg_tokens": "128",
   "tps": "53.21",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "65536",
   "date": "2026-04-01",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc20969-3090-md39b-t3-64",
   "model": "Qwen 3.5 9B",
   "notes": "Madreag/turbo3-cuda optimized fork vs base TurboQuant fork, RTX 3090, Qwen 3.5 9B Q8_0, llama-bench tg128 at 32K/64K depth, back-to-back same GPU same model; base fork turbo3 KV at 64K; config=md39b-t3-64",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "### RTX 3090 (9B Q8_0); Qwen 3.5 9B Q8_0; | turbo3 64K | 33.43 | 49.27 | **+47%** |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16404751",
   "tg_tokens": "128",
   "tps": "33.43",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "65536",
   "date": "2026-04-01",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc20969-3090-md39b-t2-64",
   "model": "Qwen 3.5 9B",
   "notes": "Madreag/turbo3-cuda optimized fork vs base TurboQuant fork, RTX 3090, Qwen 3.5 9B Q8_0, llama-bench tg128 at 32K/64K depth, back-to-back same GPU same model; base fork turbo2 KV at 64K; config=md39b-t2-64",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "### RTX 3090 (9B Q8_0); Qwen 3.5 9B Q8_0; | turbo2 64K | 42.45 | 56.91 | **+34%** |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16404751",
   "tg_tokens": "128",
   "tps": "42.45",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "32768",
   "date": "2026-04-01",
   "flags": [],
   "hardware": "RTX 4090M",
   "id": "lc20969-4090m-md4m-q8",
   "model": "Qwen 3.5 9B",
   "notes": "Madreag/turbo3-cuda optimized fork vs base fork, RTX 4090M 16 GB (SM89), Qwen 3.5 9B Q8_0, llama-bench tg128 at 32K depth; optimized fork q8_0 KV at 32K; config=md4m-q8",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "### RTX 4090M (9B Q8_0); Qwen 3.5 9B Q8_0; | q8_0 32K | 48.2 | 52.0 | **+8%** |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16404751",
   "tg_tokens": "128",
   "tps": "52.0",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "32768",
   "date": "2026-04-01",
   "flags": [],
   "hardware": "RTX 4090M",
   "id": "lc20969-4090m-md4m-t4",
   "model": "Qwen 3.5 9B",
   "notes": "Madreag/turbo3-cuda optimized fork vs base fork, RTX 4090M 16 GB (SM89), Qwen 3.5 9B Q8_0, llama-bench tg128 at 32K depth; optimized fork turbo4 KV at 32K; config=md4m-t4",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "### RTX 4090M (9B Q8_0); Qwen 3.5 9B Q8_0; | turbo4 32K | 34.5 | 52.4 | **+52%** |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16404751",
   "tg_tokens": "128",
   "tps": "52.4",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "32768",
   "date": "2026-04-01",
   "flags": [],
   "hardware": "RTX 4090M",
   "id": "lc20969-4090m-md4m-t3",
   "model": "Qwen 3.5 9B",
   "notes": "Madreag/turbo3-cuda optimized fork vs base fork, RTX 4090M 16 GB (SM89), Qwen 3.5 9B Q8_0, llama-bench tg128 at 32K depth; optimized fork turbo3 KV at 32K; config=md4m-t3",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "### RTX 4090M (9B Q8_0); Qwen 3.5 9B Q8_0; | turbo3 32K | 40.3 | 49.0 | **+22%** |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16404751",
   "tg_tokens": "128",
   "tps": "49.0",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "32768",
   "date": "2026-04-01",
   "flags": [],
   "hardware": "RTX 4090M",
   "id": "lc20969-4090m-md4m-t2",
   "model": "Qwen 3.5 9B",
   "notes": "Madreag/turbo3-cuda optimized fork vs base fork, RTX 4090M 16 GB (SM89), Qwen 3.5 9B Q8_0, llama-bench tg128 at 32K depth; optimized fork turbo2 KV at 32K; config=md4m-t2",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "### RTX 4090M (9B Q8_0); Qwen 3.5 9B Q8_0; | turbo2 32K | 44.9 | 52.7 | **+17%** |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16404751",
   "tg_tokens": "128",
   "tps": "52.7",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "32768",
   "date": "2026-04-01",
   "flags": [],
   "hardware": "RTX 4090M",
   "id": "lc20969-4090m-md4mb-q8",
   "model": "Qwen 3.5 9B",
   "notes": "Madreag/turbo3-cuda optimized fork vs base fork, RTX 4090M 16 GB (SM89), Qwen 3.5 9B Q8_0, llama-bench tg128 at 32K depth; base fork q8_0 KV at 32K; config=md4mb-q8",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "### RTX 4090M (9B Q8_0); Qwen 3.5 9B Q8_0; | q8_0 32K | 48.2 | 52.0 | **+8%** |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16404751",
   "tg_tokens": "128",
   "tps": "48.2",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "32768",
   "date": "2026-04-01",
   "flags": [],
   "hardware": "RTX 4090M",
   "id": "lc20969-4090m-md4mb-t4",
   "model": "Qwen 3.5 9B",
   "notes": "Madreag/turbo3-cuda optimized fork vs base fork, RTX 4090M 16 GB (SM89), Qwen 3.5 9B Q8_0, llama-bench tg128 at 32K depth; base fork turbo4 KV at 32K; config=md4mb-t4",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "### RTX 4090M (9B Q8_0); Qwen 3.5 9B Q8_0; | turbo4 32K | 34.5 | 52.4 | **+52%** |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16404751",
   "tg_tokens": "128",
   "tps": "34.5",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "32768",
   "date": "2026-04-01",
   "flags": [],
   "hardware": "RTX 4090M",
   "id": "lc20969-4090m-md4mb-t3",
   "model": "Qwen 3.5 9B",
   "notes": "Madreag/turbo3-cuda optimized fork vs base fork, RTX 4090M 16 GB (SM89), Qwen 3.5 9B Q8_0, llama-bench tg128 at 32K depth; base fork turbo3 KV at 32K; config=md4mb-t3",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "### RTX 4090M (9B Q8_0); Qwen 3.5 9B Q8_0; | turbo3 32K | 40.3 | 49.0 | **+22%** |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16404751",
   "tg_tokens": "128",
   "tps": "40.3",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "32768",
   "date": "2026-04-01",
   "flags": [],
   "hardware": "RTX 4090M",
   "id": "lc20969-4090m-md4mb-t2",
   "model": "Qwen 3.5 9B",
   "notes": "Madreag/turbo3-cuda optimized fork vs base fork, RTX 4090M 16 GB (SM89), Qwen 3.5 9B Q8_0, llama-bench tg128 at 32K depth; base fork turbo2 KV at 32K; config=md4mb-t2",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "### RTX 4090M (9B Q8_0); Qwen 3.5 9B Q8_0; | turbo2 32K | 44.9 | 52.7 | **+17%** |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16404751",
   "tg_tokens": "128",
   "tps": "44.9",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "32768",
   "date": "2026-04-01",
   "flags": [],
   "hardware": "RTX 3090 Ti",
   "id": "lc20969-3090ti-mdti-q8",
   "model": "Qwen 3.5 9B",
   "notes": "Madreag/turbo3-cuda optimized fork, RTX 3090 Ti (SM86, memory OC +2200), Qwen 3.5 9B Q8_0, llama-bench tg128 at depth, +-0.3% variance; q8_0 KV at 32K, short-context 91.01, OOM at 64K; config=mdti-q8",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "### RTX 3090 Ti (SM86, OC +2200 mem, Qwen 3.5 9B Q8_0); | q8_0 | 91.01 | 77.44 | OOM |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16404751",
   "tg_tokens": "128",
   "tps": "77.44",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "32768",
   "date": "2026-04-01",
   "flags": [],
   "hardware": "RTX 3090 Ti",
   "id": "lc20969-3090ti-mdti-t4",
   "model": "Qwen 3.5 9B",
   "notes": "Madreag/turbo3-cuda optimized fork, RTX 3090 Ti (SM86, memory OC +2200), Qwen 3.5 9B Q8_0, llama-bench tg128 at depth, +-0.3% variance; turbo4 KV at 32K, short-context 90.03, OOM at 64K; config=mdti-t4",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "### RTX 3090 Ti (SM86, OC +2200 mem, Qwen 3.5 9B Q8_0); | turbo4 | 90.03 | 75.55 | OOM |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16404751",
   "tg_tokens": "128",
   "tps": "75.55",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "32768",
   "date": "2026-04-01",
   "flags": [],
   "hardware": "RTX 3090 Ti",
   "id": "lc20969-3090ti-mdti-t3",
   "model": "Qwen 3.5 9B",
   "notes": "Madreag/turbo3-cuda optimized fork, RTX 3090 Ti (SM86, memory OC +2200), Qwen 3.5 9B Q8_0, llama-bench tg128 at depth, +-0.3% variance; turbo3 KV at 32K, short-context 90.35; config=mdti-t3",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "### RTX 3090 Ti (SM86, OC +2200 mem, Qwen 3.5 9B Q8_0); | turbo3 | 90.35 | 75.01 | 61.47 |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16404751",
   "tg_tokens": "128",
   "tps": "75.01",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "32768",
   "date": "2026-04-01",
   "flags": [],
   "hardware": "RTX 3090 Ti",
   "id": "lc20969-3090ti-mdti-t2",
   "model": "Qwen 3.5 9B",
   "notes": "Madreag/turbo3-cuda optimized fork, RTX 3090 Ti (SM86, memory OC +2200), Qwen 3.5 9B Q8_0, llama-bench tg128 at depth, +-0.3% variance; turbo2 KV at 32K, beats q8_0 by 5.3%, short-context 90.75; config=mdti-t2",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "### RTX 3090 Ti (SM86, OC +2200 mem, Qwen 3.5 9B Q8_0); | **turbo2** | **90.75** | **81.58** | **72.79** |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16404751",
   "tg_tokens": "128",
   "tps": "81.58",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "32768",
   "date": "2026-04-01",
   "flags": [],
   "hardware": "RTX 3090 Ti",
   "id": "lc20969-3090ti-mdti-t15",
   "model": "Qwen 3.5 9B",
   "notes": "Madreag/turbo3-cuda optimized fork, RTX 3090 Ti (SM86, memory OC +2200), Qwen 3.5 9B Q8_0, llama-bench tg128 at depth, +-0.3% variance; turbo1.5 KV at 32K, short-context 90.13; config=mdti-t15",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "### RTX 3090 Ti (SM86, OC +2200 mem, Qwen 3.5 9B Q8_0); | turbo1.5 | 90.13 | 74.85 | 63.44 |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16404751",
   "tg_tokens": "128",
   "tps": "74.85",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "65536",
   "date": "2026-04-01",
   "flags": [],
   "hardware": "RTX 3090 Ti",
   "id": "lc20969-3090ti-mdti-t3-64",
   "model": "Qwen 3.5 9B",
   "notes": "Madreag/turbo3-cuda optimized fork, RTX 3090 Ti (SM86, memory OC +2200), Qwen 3.5 9B Q8_0, llama-bench tg128 at depth, +-0.3% variance; turbo3 KV at 64K; config=mdti-t3-64",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "### RTX 3090 Ti (SM86, OC +2200 mem, Qwen 3.5 9B Q8_0); | turbo3 | 90.35 | 75.01 | 61.47 |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16404751",
   "tg_tokens": "128",
   "tps": "61.47",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "65536",
   "date": "2026-04-01",
   "flags": [],
   "hardware": "RTX 3090 Ti",
   "id": "lc20969-3090ti-mdti-t2-64",
   "model": "Qwen 3.5 9B",
   "notes": "Madreag/turbo3-cuda optimized fork, RTX 3090 Ti (SM86, memory OC +2200), Qwen 3.5 9B Q8_0, llama-bench tg128 at depth, +-0.3% variance; turbo2 KV at 64K where q8_0 and turbo4 OOM; config=mdti-t2-64",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "### RTX 3090 Ti (SM86, OC +2200 mem, Qwen 3.5 9B Q8_0); | **turbo2** | **90.75** | **81.58** | **72.79** |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16404751",
   "tg_tokens": "128",
   "tps": "72.79",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "65536",
   "date": "2026-04-01",
   "flags": [],
   "hardware": "RTX 3090 Ti",
   "id": "lc20969-3090ti-mdti-t15-64",
   "model": "Qwen 3.5 9B",
   "notes": "Madreag/turbo3-cuda optimized fork, RTX 3090 Ti (SM86, memory OC +2200), Qwen 3.5 9B Q8_0, llama-bench tg128 at depth, +-0.3% variance; turbo1.5 KV at 64K; config=mdti-t15-64",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "### RTX 3090 Ti (SM86, OC +2200 mem, Qwen 3.5 9B Q8_0); | turbo1.5 | 90.13 | 74.85 | 63.44 |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16404751",
   "tg_tokens": "128",
   "tps": "63.44",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": "4",
   "ctx": "8192",
   "date": "2026-03-31",
   "flags": [],
   "hardware": "2x RTX 5060 Ti",
   "id": "lc20969-5060ti-df-l31-b",
   "model": "Llama 3.1 8B",
   "notes": "Defilan on 2x RTX 5060 Ti (SM 120 Blackwell, CUDA 13.0) with the animehacker TurboQuant fork vs stock llama.cpp, flash attention on for both, only difference tq3_0 K+V cache, 8K context, 4 concurrent requests, throughput per server; animehacker kernels were tuned for Ampere, author reports roughly 5x penalty across the board and asks whether spiritbuun or Madreag forks do better on SM 120; stock f16 KV baseline; config=df-l31-b",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q5_K_M",
   "quote": "Tested the animehacker fork on 2x RTX 5060 Ti (SM 120, CUDA 13.0, 16GB each); - Llama 3.1 8B Q5_K_M: 50 tok/s baseline, 8.4 tok/s TQ",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16391932",
   "tg_tokens": null,
   "tps": "50",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": "4",
   "ctx": "8192",
   "date": "2026-03-31",
   "flags": [],
   "hardware": "2x RTX 5060 Ti",
   "id": "lc20969-5060ti-df-l31-t",
   "model": "Llama 3.1 8B",
   "notes": "Defilan on 2x RTX 5060 Ti (SM 120 Blackwell, CUDA 13.0) with the animehacker TurboQuant fork vs stock llama.cpp, flash attention on for both, only difference tq3_0 K+V cache, 8K context, 4 concurrent requests, throughput per server; animehacker kernels were tuned for Ampere, author reports roughly 5x penalty across the board and asks whether spiritbuun or Madreag forks do better on SM 120; tq3_0 K and V cache; config=df-l31-t",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q5_K_M",
   "quote": "Tested the animehacker fork on 2x RTX 5060 Ti (SM 120, CUDA 13.0, 16GB each); - Llama 3.1 8B Q5_K_M: 50 tok/s baseline, 8.4 tok/s TQ",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16391932",
   "tg_tokens": null,
   "tps": "8.4",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": "4",
   "ctx": "8192",
   "date": "2026-03-31",
   "flags": [],
   "hardware": "2x RTX 5060 Ti",
   "id": "lc20969-5060ti-df-q14-b",
   "model": "Qwen 2.5 14B",
   "notes": "Defilan on 2x RTX 5060 Ti (SM 120 Blackwell, CUDA 13.0) with the animehacker TurboQuant fork vs stock llama.cpp, flash attention on for both, only difference tq3_0 K+V cache, 8K context, 4 concurrent requests, throughput per server; animehacker kernels were tuned for Ampere, author reports roughly 5x penalty across the board and asks whether spiritbuun or Madreag forks do better on SM 120; stock f16 KV baseline; config=df-q14-b",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q5_K_M",
   "quote": "Tested the animehacker fork on 2x RTX 5060 Ti (SM 120, CUDA 13.0, 16GB each); - Qwen 2.5 14B Q5_K_M: 28 tok/s baseline, 5.3 tok/s TQ",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16391932",
   "tg_tokens": null,
   "tps": "28",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": "4",
   "ctx": "8192",
   "date": "2026-03-31",
   "flags": [],
   "hardware": "2x RTX 5060 Ti",
   "id": "lc20969-5060ti-df-q14-t",
   "model": "Qwen 2.5 14B",
   "notes": "Defilan on 2x RTX 5060 Ti (SM 120 Blackwell, CUDA 13.0) with the animehacker TurboQuant fork vs stock llama.cpp, flash attention on for both, only difference tq3_0 K+V cache, 8K context, 4 concurrent requests, throughput per server; animehacker kernels were tuned for Ampere, author reports roughly 5x penalty across the board and asks whether spiritbuun or Madreag forks do better on SM 120; tq3_0 K and V cache; config=df-q14-t",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q5_K_M",
   "quote": "Tested the animehacker fork on 2x RTX 5060 Ti (SM 120, CUDA 13.0, 16GB each); - Qwen 2.5 14B Q5_K_M: 28 tok/s baseline, 5.3 tok/s TQ",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16391932",
   "tg_tokens": null,
   "tps": "5.3",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": "4",
   "ctx": "8192",
   "date": "2026-03-31",
   "flags": [],
   "hardware": "2x RTX 5060 Ti",
   "id": "lc20969-5060ti-df-q32-b",
   "model": "Qwen 2.5 32B",
   "notes": "Defilan on 2x RTX 5060 Ti (SM 120 Blackwell, CUDA 13.0) with the animehacker TurboQuant fork vs stock llama.cpp, flash attention on for both, only difference tq3_0 K+V cache, 8K context, 4 concurrent requests, throughput per server; animehacker kernels were tuned for Ampere, author reports roughly 5x penalty across the board and asks whether spiritbuun or Madreag forks do better on SM 120; stock f16 KV baseline; config=df-q32-b",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Tested the animehacker fork on 2x RTX 5060 Ti (SM 120, CUDA 13.0, 16GB each); - Qwen 2.5 32B Q4_K_M: 14.3 tok/s baseline, 5.5 tok/s TQ",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16391932",
   "tg_tokens": null,
   "tps": "14.3",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": "4",
   "ctx": "8192",
   "date": "2026-03-31",
   "flags": [],
   "hardware": "2x RTX 5060 Ti",
   "id": "lc20969-5060ti-df-q32-t",
   "model": "Qwen 2.5 32B",
   "notes": "Defilan on 2x RTX 5060 Ti (SM 120 Blackwell, CUDA 13.0) with the animehacker TurboQuant fork vs stock llama.cpp, flash attention on for both, only difference tq3_0 K+V cache, 8K context, 4 concurrent requests, throughput per server; animehacker kernels were tuned for Ampere, author reports roughly 5x penalty across the board and asks whether spiritbuun or Madreag forks do better on SM 120; tq3_0 K and V cache; config=df-q32-t",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Tested the animehacker fork on 2x RTX 5060 Ti (SM 120, CUDA 13.0, 16GB each); - Qwen 2.5 32B Q4_K_M: 14.3 tok/s baseline, 5.5 tok/s TQ",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16391932",
   "tg_tokens": null,
   "tps": "5.5",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "131072",
   "date": "2026-04-02",
   "flags": [],
   "hardware": "2x EVGA RTX 3080 Ti 12GB",
   "id": "lc20969-3080ti-sr-128k",
   "model": "Qwen3.5-9B",
   "notes": "seanrasch, dual RTX 3080 Ti 12 GB (24 GB total) with seanrasch/llama-cpp-turboquant master, Qwen3.5-9B Q4_K_M with turbo2 K+V cache, native 256K context push; decode stays flat at ~105 tok/s regardless of context depth while prefill falls with O(n^2) attention, 10 GB free across both cards at 256K so the model context limit is the ceiling, not VRAM; on a single 12 GB card 128K OOMs with every turbo config; full-context prefill 2027 tok/s, tg128 decode 104.7 tok/s at 131072 tokens; config=sr-128k",
   "power_w": null,
   "pp_tokens": "131072",
   "pp_tps": "2027",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "2x EVGA RTX 3080 Ti 12GB (SM 86, Ampere, PCIe Gen 3 x16); Qwen3.5 9B Q4_K_M (bartowski); | 128K | 2027 | 104.7 |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16424635",
   "tg_tokens": "128",
   "tps": "104.7",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "163840",
   "date": "2026-04-02",
   "flags": [],
   "hardware": "2x EVGA RTX 3080 Ti 12GB",
   "id": "lc20969-3080ti-sr-160k",
   "model": "Qwen3.5-9B",
   "notes": "seanrasch, dual RTX 3080 Ti 12 GB (24 GB total) with seanrasch/llama-cpp-turboquant master, Qwen3.5-9B Q4_K_M with turbo2 K+V cache, native 256K context push; decode stays flat at ~105 tok/s regardless of context depth while prefill falls with O(n^2) attention, 10 GB free across both cards at 256K so the model context limit is the ceiling, not VRAM; on a single 12 GB card 128K OOMs with every turbo config; full-context prefill 1827 tok/s, tg128 decode 104.4 tok/s at 163840 tokens; config=sr-160k",
   "power_w": null,
   "pp_tokens": "163840",
   "pp_tps": "1827",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "2x EVGA RTX 3080 Ti 12GB (SM 86, Ampere, PCIe Gen 3 x16); Qwen3.5 9B Q4_K_M (bartowski); | 160K | 1827 | 104.4 |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16424635",
   "tg_tokens": "128",
   "tps": "104.4",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "262144",
   "date": "2026-04-02",
   "flags": [],
   "hardware": "2x EVGA RTX 3080 Ti 12GB",
   "id": "lc20969-3080ti-sr-256k",
   "model": "Qwen3.5-9B",
   "notes": "seanrasch, dual RTX 3080 Ti 12 GB (24 GB total) with seanrasch/llama-cpp-turboquant master, Qwen3.5-9B Q4_K_M with turbo2 K+V cache, native 256K context push; decode stays flat at ~105 tok/s regardless of context depth while prefill falls with O(n^2) attention, 10 GB free across both cards at 256K so the model context limit is the ceiling, not VRAM; on a single 12 GB card 128K OOMs with every turbo config; full-context prefill 1423 tok/s, tg128 decode 104.8 tok/s at 262144 tokens; config=sr-256k",
   "power_w": null,
   "pp_tokens": "262144",
   "pp_tps": "1423",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "2x EVGA RTX 3080 Ti 12GB (SM 86, Ampere, PCIe Gen 3 x16); Qwen3.5 9B Q4_K_M (bartowski); | 256K | 1423 | 104.8 |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16424635",
   "tg_tokens": "128",
   "tps": "104.8",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-04-01",
   "flags": [],
   "hardware": "DGX Spark",
   "id": "lc20969-spark-ax-35b-q80",
   "model": "Qwen3.5 35B",
   "notes": "AmesianX TurboQuant v1.2/v1.3 fork on DGX Spark (SM 121 GB10), author-reported Blackwell decode comparison; q8_0 K+V cache decode on Qwen3.5-35B; config=ax-35b-q80",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "we only have DGX Spark (SM 121, GB10) numbers so far; we see turbo *faster* than q8_0 on Blackwell decode (63.5 vs 50.1 t/s on Qwen3.5-35B)",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16408848",
   "tg_tokens": "128",
   "tps": "50.1",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-04-01",
   "flags": [],
   "hardware": "DGX Spark",
   "id": "lc20969-spark-ax-35b-turbo",
   "model": "Qwen3.5 35B",
   "notes": "AmesianX TurboQuant v1.2/v1.3 fork on DGX Spark (SM 121 GB10), author-reported Blackwell decode comparison; turbo K+V cache decode, +26.7% over q8_0, the opposite of WaveboSF's Blackwell results with the spiritbuun fork; config=ax-35b-turbo",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "we only have DGX Spark (SM 121, GB10) numbers so far; we see turbo *faster* than q8_0 on Blackwell decode (63.5 vs 50.1 t/s on Qwen3.5-35B)",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16408848",
   "tg_tokens": "128",
   "tps": "63.5",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-04-02",
   "flags": [],
   "hardware": "DGX Spark",
   "id": "lc20969-spark-ax-27b-f16pp",
   "model": "Qwen3.5-27B",
   "notes": "AmesianX fork on DGX Spark (SM 121 native build), Qwen3.5-27B heretic-v3 i1-Q3_K_M, prompt processing with tbqp3-K/tbq3-V vs f16 KV, author reports only ~15% pp drop with the native build (release binaries had wrong SM targets and JIT-emulated far slower); f16 KV prompt processing 94 tok/s; config=ax-27b-f16pp",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "94",
   "provenance": "community",
   "quant": "i1-Q3_K_M",
   "quote": "Output is completely normal on our local build (SM 121 native, DGX Spark GB10); Qwen3.5-27B-heretic-v3.i1-Q3_K_M.gguf; On our SM 121 native build, tbqp3/tbq3 prompt processing drops only ~15% vs f16 (94\u219280 t/s)",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16419773",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-04-02",
   "flags": [],
   "hardware": "DGX Spark",
   "id": "lc20969-spark-ax-27b-tbqpp",
   "model": "Qwen3.5-27B",
   "notes": "AmesianX fork on DGX Spark (SM 121 native build), Qwen3.5-27B heretic-v3 i1-Q3_K_M, prompt processing with tbqp3-K/tbq3-V vs f16 KV, author reports only ~15% pp drop with the native build (release binaries had wrong SM targets and JIT-emulated far slower); tbqp3/tbq3 KV prompt processing 80 tok/s; config=ax-27b-tbqpp",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "80",
   "provenance": "community",
   "quant": "i1-Q3_K_M",
   "quote": "Output is completely normal on our local build (SM 121 native, DGX Spark GB10); Qwen3.5-27B-heretic-v3.i1-Q3_K_M.gguf; On our SM 121 native build, tbqp3/tbq3 prompt processing drops only ~15% vs f16 (94\u219280 t/s)",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp discussion #20969",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/20969#discussioncomment-16419773",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 92160,
   "date": "2026-10",
   "flags": [],
   "hardware": "GTX 1080 8GB",
   "id": "lc-disc-29885-gtx1080-qwen36-pp-gpu",
   "model": "Qwen3.6-35B-A3B",
   "notes": "40K-token agentic prompt, cold prefill 151s, TTFT ~155s; experts on GPU; config=experts-gpu",
   "params": null,
   "power_w": null,
   "pp_tps": 265,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "Measured on GTX 1080 8GB, Qwen3.6-35B-A3B, `-c 92160`; Cold: 151s prefill -> first token at ~155s; keep experts on GPU, let KV follow the layers - pushing 20 MoE layers to CPU collapsed prefill 265->30 tok/s",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29885 (qqkzlm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29885",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": 155
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 92160,
   "date": "2026-10",
   "flags": [],
   "hardware": "GTX 1080 8GB",
   "id": "lc-disc-29885-gtx1080-qwen36-pp-cpu20",
   "model": "Qwen3.6-35B-A3B",
   "notes": "same prompt/box with 20 MoE layers pushed to CPU; config=20-moe-cpu",
   "params": null,
   "power_w": null,
   "pp_tps": 30,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "Measured on GTX 1080 8GB, Qwen3.6-35B-A3B, `-c 92160`; Cold: 151s prefill -> first token at ~155s; keep experts on GPU, let KV follow the layers - pushing 20 MoE layers to CPU collapsed prefill 265->30 tok/s",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29885 (qqkzlm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29885",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "CUDA",
   "batch": null,
   "ctx": 262144,
   "date": "2026-10",
   "flags": [],
   "hardware": "3x RTX PRO 4000 Blackwell",
   "id": "lc-disc-29964-qwen38-master",
   "model": "Qwen3.8",
   "notes": "pipeline-parallel patch A/B, 52421-token RFC prompt, server; 2nd request 961 t/s; config=master",
   "params": null,
   "power_w": null,
   "pp_tps": 952,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "Benchmarks (3x RTX PRO 4000 blackwell, PCIe 5.0x16, EPYC 9274F, 384 GB DDR5 (12), CUDA 13.3); | Qwen3.8 (fit, c=262144) | 952 / 961 | 1756 / 1982 | 40.8 -> 41.6 |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29964 (drrros)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29964",
   "tg_tokens": null,
   "tps": 40.8,
   "ttft_s": null
  },
  {
   "backend": "CUDA",
   "batch": null,
   "ctx": 262144,
   "date": "2026-10",
   "flags": [],
   "hardware": "3x RTX PRO 4000 Blackwell",
   "id": "lc-disc-29964-qwen38-patched",
   "model": "Qwen3.8",
   "notes": "pipeline-parallel patch A/B, 52421-token RFC prompt, server; 2nd request 1982 t/s; config=patched",
   "params": null,
   "power_w": null,
   "pp_tps": 1756,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "Benchmarks (3x RTX PRO 4000 blackwell, PCIe 5.0x16, EPYC 9274F, 384 GB DDR5 (12), CUDA 13.3); | Qwen3.8 (fit, c=262144) | 952 / 961 | 1756 / 1982 | 40.8 -> 41.6 |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29964 (drrros)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29964",
   "tg_tokens": null,
   "tps": 41.6,
   "ttft_s": null
  },
  {
   "backend": "CUDA",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "3x RTX PRO 4000 Blackwell",
   "id": "lc-disc-29964-dsv4f-master",
   "model": "DeepSeek-V4-Flash",
   "notes": "pipeline-parallel patch A/B, 52421-token RFC prompt, server, fit; 2nd request 629 t/s; config=master",
   "params": null,
   "power_w": null,
   "pp_tps": 625,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "Benchmarks (3x RTX PRO 4000 blackwell, PCIe 5.0x16, EPYC 9274F, 384 GB DDR5 (12), CUDA 13.3); | DeepSeek-V4-Flash (fit) | 625 / 629 | 1128 / 1321 | 23.0 -> 23.1 |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29964 (drrros)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29964",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "CUDA",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "3x RTX PRO 4000 Blackwell",
   "id": "lc-disc-29964-dsv4f-patched",
   "model": "DeepSeek-V4-Flash",
   "notes": "pipeline-parallel patch A/B, 52421-token RFC prompt, server, fit; 2nd request 1321 t/s; config=patched",
   "params": null,
   "power_w": null,
   "pp_tps": 1128,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "Benchmarks (3x RTX PRO 4000 blackwell, PCIe 5.0x16, EPYC 9274F, 384 GB DDR5 (12), CUDA 13.3); | DeepSeek-V4-Flash (fit) | 625 / 629 | 1128 / 1321 | 23.0 -> 23.1 |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29964 (drrros)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29964",
   "tg_tokens": null,
   "tps": 23.1,
   "ttft_s": null
  },
  {
   "backend": "CUDA",
   "batch": null,
   "ctx": 131072,
   "date": "2026-10",
   "flags": [],
   "hardware": "3x RTX PRO 4000 Blackwell",
   "id": "lc-disc-29964-mimo-master",
   "model": "MiMo-V2.6-Flash",
   "notes": "pipeline-parallel patch A/B, 51341-token prompt, server, fit; 2nd request 886 t/s; config=master",
   "params": null,
   "power_w": null,
   "pp_tps": 883,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "Benchmarks (3x RTX PRO 4000 blackwell, PCIe 5.0x16, EPYC 9274F, 384 GB DDR5 (12), CUDA 13.3); | MiMo-V2.6-Flash (fit, c=131072) | 883 / 886 | 1741 / 1834 | 33.6 -> 33.3 |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29964 (drrros)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29964",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "CUDA",
   "batch": null,
   "ctx": 131072,
   "date": "2026-10",
   "flags": [],
   "hardware": "3x RTX PRO 4000 Blackwell",
   "id": "lc-disc-29964-mimo-patched",
   "model": "MiMo-V2.6-Flash",
   "notes": "pipeline-parallel patch A/B, 51341-token prompt, server, fit; 2nd request 1834 t/s; config=patched",
   "params": null,
   "power_w": null,
   "pp_tps": 1741,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "Benchmarks (3x RTX PRO 4000 blackwell, PCIe 5.0x16, EPYC 9274F, 384 GB DDR5 (12), CUDA 13.3); | MiMo-V2.6-Flash (fit, c=131072) | 883 / 886 | 1741 / 1834 | 33.6 -> 33.3 |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29964 (drrros)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29964",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "CUDA",
   "batch": null,
   "ctx": 65536,
   "date": "2026-10",
   "flags": [],
   "hardware": "3x RTX PRO 4000 Blackwell",
   "id": "lc-disc-29964-qwen3535b-master",
   "model": "Qwen3.5 35B",
   "notes": "bisect arm: master cb7934c52 (post-#29184), 52421-token prompt; 2nd request 3716 t/s; config=master-bisect",
   "params": null,
   "power_w": null,
   "pp_tps": 3701,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "Benchmarks (3x RTX PRO 4000 blackwell, PCIe 5.0x16, EPYC 9274F, 384 GB DDR5 (12), CUDA 13.3); Qwen3.5-35B fully in VRAM, 3 GPUs, server, 52421-token prompt, c=65536 - 7131 / 7632 t/s on the parent 4ebdf2c74, 3740 / 3747 on bed0a8566 (#29184), 3701 / 3716 on master cb7934c52, and 7148 / 7675 with only this fix on top of master",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29964 (drrros)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29964",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "CUDA",
   "batch": null,
   "ctx": 65536,
   "date": "2026-10",
   "flags": [],
   "hardware": "3x RTX PRO 4000 Blackwell",
   "id": "lc-disc-29964-qwen3535b-patched",
   "model": "Qwen3.5 35B",
   "notes": "bisect arm: master + topology fix only, 52421-token prompt; 2nd request 7675 t/s; pre-regression parent 4ebdf2c74 was 7131/7632; config=patched-bisect",
   "params": null,
   "power_w": null,
   "pp_tps": 7148,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "Benchmarks (3x RTX PRO 4000 blackwell, PCIe 5.0x16, EPYC 9274F, 384 GB DDR5 (12), CUDA 13.3); Qwen3.5-35B fully in VRAM, 3 GPUs, server, 52421-token prompt, c=65536 - 7131 / 7632 t/s on the parent 4ebdf2c74, 3740 / 3747 on bed0a8566 (#29184), 3701 / 3716 on master cb7934c52, and 7148 / 7675 with only this fix on top of master",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29964 (drrros)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29964",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "CUDA",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "3x RTX PRO 4000 Blackwell",
   "id": "lc-disc-29964-qwen3535ba3b-pp8192-master",
   "model": "Qwen3.5-35B-A3B",
   "notes": "all experts in RAM (-ncmoe 999, no fit), pp8192; config=master-pp8192",
   "params": null,
   "power_w": null,
   "pp_tps": 3078,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "Benchmarks (3x RTX PRO 4000 blackwell, PCIe 5.0x16, EPYC 9274F, 384 GB DDR5 (12), CUDA 13.3); All experts in RAM (Qwen3.5-35B-A3B, `-ncmoe 999`, no `fit`): pp8192 3078 -> 4393, pp32768 2814 -> 6274",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29964 (drrros)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29964",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "CUDA",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "3x RTX PRO 4000 Blackwell",
   "id": "lc-disc-29964-qwen3535ba3b-pp8192-patched",
   "model": "Qwen3.5-35B-A3B",
   "notes": "all experts in RAM (-ncmoe 999, no fit), pp8192; config=patched-pp8192",
   "params": null,
   "power_w": null,
   "pp_tps": 4393,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "Benchmarks (3x RTX PRO 4000 blackwell, PCIe 5.0x16, EPYC 9274F, 384 GB DDR5 (12), CUDA 13.3); All experts in RAM (Qwen3.5-35B-A3B, `-ncmoe 999`, no `fit`): pp8192 3078 -> 4393, pp32768 2814 -> 6274",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29964 (drrros)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29964",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "CUDA",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "3x RTX PRO 4000 Blackwell",
   "id": "lc-disc-29964-qwen3535ba3b-pp32768-master",
   "model": "Qwen3.5-35B-A3B",
   "notes": "all experts in RAM (-ncmoe 999, no fit), pp32768; config=master-pp32768",
   "params": null,
   "power_w": null,
   "pp_tps": 2814,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "Benchmarks (3x RTX PRO 4000 blackwell, PCIe 5.0x16, EPYC 9274F, 384 GB DDR5 (12), CUDA 13.3); All experts in RAM (Qwen3.5-35B-A3B, `-ncmoe 999`, no `fit`): pp8192 3078 -> 4393, pp32768 2814 -> 6274",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29964 (drrros)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29964",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "CUDA",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "3x RTX PRO 4000 Blackwell",
   "id": "lc-disc-29964-qwen3535ba3b-pp32768-patched",
   "model": "Qwen3.5-35B-A3B",
   "notes": "all experts in RAM (-ncmoe 999, no fit), pp32768; config=patched-pp32768",
   "params": null,
   "power_w": null,
   "pp_tps": 6274,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "Benchmarks (3x RTX PRO 4000 blackwell, PCIe 5.0x16, EPYC 9274F, 384 GB DDR5 (12), CUDA 13.3); All experts in RAM (Qwen3.5-35B-A3B, `-ncmoe 999`, no `fit`): pp8192 3078 -> 4393, pp32768 2814 -> 6274",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29964 (drrros)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29964",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "CUDA",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "3x RTX PRO 4000 Blackwell",
   "id": "lc-disc-29964-qwen38-2gpu-master",
   "model": "Qwen3.8",
   "notes": "pp16384 with fit, only 2 of the 3 GPUs used (model does not fit in VRAM either way); config=2gpu-master",
   "params": null,
   "power_w": null,
   "pp_tps": 1254,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "Benchmarks (3x RTX PRO 4000 blackwell, PCIe 5.0x16, EPYC 9274F, 384 GB DDR5 (12), CUDA 13.3); 2 GPUs with `fit` (Qwen3.8, which does not fit in VRAM either way): pp16384 1254 -> 1496",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29964 (drrros)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29964",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "CUDA",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "3x RTX PRO 4000 Blackwell",
   "id": "lc-disc-29964-qwen38-2gpu-patched",
   "model": "Qwen3.8",
   "notes": "pp16384 with fit, only 2 of the 3 GPUs used (model does not fit in VRAM either way); config=2gpu-patched",
   "params": null,
   "power_w": null,
   "pp_tps": 1496,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "Benchmarks (3x RTX PRO 4000 blackwell, PCIe 5.0x16, EPYC 9274F, 384 GB DDR5 (12), CUDA 13.3); 2 GPUs with `fit` (Qwen3.8, which does not fit in VRAM either way): pp16384 1254 -> 1496",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29964 (drrros)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29964",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "CUDA",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "3x RTX PRO 4000 Blackwell",
   "id": "lc-disc-29964-gptoss20b-master",
   "model": "gpt-oss 20B",
   "notes": "llama-bench pp8192 noise check, patch does not regress; config=master",
   "params": null,
   "power_w": null,
   "pp_tps": 17812,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "Benchmarks (3x RTX PRO 4000 blackwell, PCIe 5.0x16, EPYC 9274F, 384 GB DDR5 (12), CUDA 13.3); llama-bench pp8192 gpt-oss-20b 17812 -> 17746, Llama 3.3 8B 12615 -> 12583, both are noise",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29964 (drrros)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29964",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "CUDA",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "3x RTX PRO 4000 Blackwell",
   "id": "lc-disc-29964-gptoss20b-patched",
   "model": "gpt-oss 20B",
   "notes": "llama-bench pp8192 noise check, patch does not regress; config=patched",
   "params": null,
   "power_w": null,
   "pp_tps": 17746,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "Benchmarks (3x RTX PRO 4000 blackwell, PCIe 5.0x16, EPYC 9274F, 384 GB DDR5 (12), CUDA 13.3); llama-bench pp8192 gpt-oss-20b 17812 -> 17746, Llama 3.3 8B 12615 -> 12583, both are noise",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29964 (drrros)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29964",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "CUDA",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "3x RTX PRO 4000 Blackwell",
   "id": "lc-disc-29964-llama33-8b-master",
   "model": "Llama 3.3 8B",
   "notes": "llama-bench pp8192 noise check, patch does not regress; config=master",
   "params": null,
   "power_w": null,
   "pp_tps": 12615,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "Benchmarks (3x RTX PRO 4000 blackwell, PCIe 5.0x16, EPYC 9274F, 384 GB DDR5 (12), CUDA 13.3); llama-bench pp8192 gpt-oss-20b 17812 -> 17746, Llama 3.3 8B 12615 -> 12583, both are noise",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29964 (drrros)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29964",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "CUDA",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "3x RTX PRO 4000 Blackwell",
   "id": "lc-disc-29964-llama33-8b-patched",
   "model": "Llama 3.3 8B",
   "notes": "llama-bench pp8192 noise check, patch does not regress; config=patched",
   "params": null,
   "power_w": null,
   "pp_tps": 12583,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "Benchmarks (3x RTX PRO 4000 blackwell, PCIe 5.0x16, EPYC 9274F, 384 GB DDR5 (12), CUDA 13.3); llama-bench pp8192 gpt-oss-20b 17812 -> 17746, Llama 3.3 8B 12615 -> 12583, both are noise",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29964 (drrros)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29964",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "CUDA",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "3x RTX PRO 4000 Blackwell",
   "id": "lc-disc-29964-qwen38-q4x",
   "model": "Qwen3.8",
   "notes": "pp8192 llama-bench, patch with CUDA_SCALE_LAUNCH_QUEUES=4x; config=patch-queues4x",
   "params": null,
   "power_w": null,
   "pp_tps": 1685,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "Benchmarks (3x RTX PRO 4000 blackwell, PCIe 5.0x16, EPYC 9274F, 384 GB DDR5 (12), CUDA 13.3); both sides use `CUDA_SCALE_LAUNCH_QUEUES=4x`. Without it the patch loses 22..27% on Qwen3.8 (pp8192 1685 -> 1314)",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29964 (drrros)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29964",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "CUDA",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "3x RTX PRO 4000 Blackwell",
   "id": "lc-disc-29964-qwen38-qnone",
   "model": "Qwen3.8",
   "notes": "pp8192 llama-bench, patch without CUDA_SCALE_LAUNCH_QUEUES (default queues); config=patch-noqueues",
   "params": null,
   "power_w": null,
   "pp_tps": 1314,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "Benchmarks (3x RTX PRO 4000 blackwell, PCIe 5.0x16, EPYC 9274F, 384 GB DDR5 (12), CUDA 13.3); both sides use `CUDA_SCALE_LAUNCH_QUEUES=4x`. Without it the patch loses 22..27% on Qwen3.8 (pp8192 1685 -> 1314)",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29964 (drrros)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29964",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 131072,
   "date": "2026-09",
   "flags": [],
   "hardware": "Arc B580",
   "id": "lc-disc-29643-bonsai-branch",
   "model": "Bonsai 2 27B",
   "notes": "128K ctx, plain generation 42 t/s, prompt reading 895 t/s on the branch vs PrismML fork 21/207; config=branch",
   "params": null,
   "power_w": null,
   "pp_tps": 895,
   "provenance": "community",
   "quant": "PTQ1_0",
   "quote": "I've been working on the SYCL backend for Intel Arc B580 (Xe2), mostly for PrismML's ternary Bonsai 2 27B; Results on a B580 at 128K context: about 90 t/s writing new code, 250-370 t/s on code edits, 42 t/s plain, 44 t/s with 115K tokens of history; Same card, model and flags against the SYCL support in PrismML's fork: prompt reading 207 vs 895 t/s, plain generation 21 vs 42, generation at 32K context 8 vs 36",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29643 (Torchit1)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29643",
   "tg_tokens": null,
   "tps": 42,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 131072,
   "date": "2026-09",
   "flags": [],
   "hardware": "Arc B580",
   "id": "lc-disc-29643-bonsai-hist115k",
   "model": "Bonsai 2 27B",
   "notes": "generation with 115K tokens of conversation history, MTP plus n-gram drafts (author clarification in the follow-up); config=branch-115k-history",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "PTQ1_0",
   "quote": "I've been working on the SYCL backend for Intel Arc B580 (Xe2), mostly for PrismML's ternary Bonsai 2 27B; Results on a B580 at 128K context: about 90 t/s writing new code, 250-370 t/s on code edits, 42 t/s plain, 44 t/s with 115K tokens of history",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29643 (Torchit1)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29643",
   "tg_tokens": null,
   "tps": 44,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 131072,
   "date": "2026-09",
   "flags": [],
   "hardware": "Arc B580",
   "id": "lc-disc-29643-bonsai-mtp-codegen",
   "model": "Bonsai 2 27B",
   "notes": "MTP-assisted generation while writing new code (published as about 90 t/s); config=branch-mtp-codegen",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "PTQ1_0",
   "quote": "I've been working on the SYCL backend for Intel Arc B580 (Xe2), mostly for PrismML's ternary Bonsai 2 27B; Results on a B580 at 128K context: about 90 t/s writing new code, 250-370 t/s on code edits, 42 t/s plain, 44 t/s with 115K tokens of history",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29643 (Torchit1)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29643",
   "tg_tokens": null,
   "tps": 90,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 131072,
   "date": "2026-09",
   "flags": [],
   "hardware": "Arc B580",
   "id": "lc-disc-29643-bonsai-mtp-edits",
   "model": "Bonsai 2 27B",
   "notes": "MTP-assisted generation on code edits, published range 250-370 t/s stored at low end; config=branch-mtp-codeedits",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "PTQ1_0",
   "quote": "I've been working on the SYCL backend for Intel Arc B580 (Xe2), mostly for PrismML's ternary Bonsai 2 27B; Results on a B580 at 128K context: about 90 t/s writing new code, 250-370 t/s on code edits, 42 t/s plain, 44 t/s with 115K tokens of history",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29643 (Torchit1)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29643",
   "tg_tokens": null,
   "tps": 250,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 32768,
   "date": "2026-09",
   "flags": [],
   "hardware": "Arc B580",
   "id": "lc-disc-29643-bonsai-branch32k",
   "model": "Bonsai 2 27B",
   "notes": "generation at 32K context on the branch vs 8 t/s on the PrismML fork; config=branch-32k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "PTQ1_0",
   "quote": "I've been working on the SYCL backend for Intel Arc B580 (Xe2), mostly for PrismML's ternary Bonsai 2 27B; Same card, model and flags against the SYCL support in PrismML's fork: prompt reading 207 vs 895 t/s, plain generation 21 vs 42, generation at 32K context 8 vs 36",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29643 (Torchit1)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29643",
   "tg_tokens": null,
   "tps": 36,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (PrismML SYCL fork)",
   "batch": null,
   "ctx": 131072,
   "date": "2026-09",
   "flags": [],
   "hardware": "Arc B580",
   "id": "lc-disc-29643-bonsai-fork",
   "model": "Bonsai 2 27B",
   "notes": "PrismML fork SYCL baseline: plain generation 21 t/s, prompt reading 207 t/s; config=fork",
   "params": null,
   "power_w": null,
   "pp_tps": 207,
   "provenance": "community",
   "quant": "PTQ1_0",
   "quote": "I've been working on the SYCL backend for Intel Arc B580 (Xe2), mostly for PrismML's ternary Bonsai 2 27B; Same card, model and flags against the SYCL support in PrismML's fork: prompt reading 207 vs 895 t/s, plain generation 21 vs 42, generation at 32K context 8 vs 36",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29643 (Torchit1)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29643",
   "tg_tokens": null,
   "tps": 21,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (PrismML SYCL fork)",
   "batch": null,
   "ctx": 32768,
   "date": "2026-09",
   "flags": [],
   "hardware": "Arc B580",
   "id": "lc-disc-29643-bonsai-fork32k",
   "model": "Bonsai 2 27B",
   "notes": "PrismML fork SYCL generation at 32K context; config=fork-32k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "PTQ1_0",
   "quote": "I've been working on the SYCL backend for Intel Arc B580 (Xe2), mostly for PrismML's ternary Bonsai 2 27B; Same card, model and flags against the SYCL support in PrismML's fork: prompt reading 207 vs 895 t/s, plain generation 21 vs 42, generation at 32K context 8 vs 36",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29643 (Torchit1)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29643",
   "tg_tokens": null,
   "tps": 8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "Arc B580",
   "id": "lc-disc-29643-gemma412b-old",
   "model": "Gemma 4 12B",
   "notes": "with Google's assistant drafter, before the new XMX attention kernels (published as about 40 t/s); config=old-kernels",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "QAT",
   "quote": "I've been working on the SYCL backend for Intel Arc B580 (Xe2), mostly for PrismML's ternary Bonsai 2 27B; The same attention kernel generalises to Gemma 4 12B (head 512, GQA 16), plus a q4_0 small-batch GEMM on XMX: Gemma 4 12B QAT with Google's assistant drafter goes from about 40 to 100 t/s on new code",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29643 (Torchit1)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29643",
   "tg_tokens": null,
   "tps": 40,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "Arc B580",
   "id": "lc-disc-29643-gemma412b-new",
   "model": "Gemma 4 12B",
   "notes": "with Google's assistant drafter, on the new q4_0 KV attention + XMX q4_0 GEMM code; config=new-code",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "QAT",
   "quote": "I've been working on the SYCL backend for Intel Arc B580 (Xe2), mostly for PrismML's ternary Bonsai 2 27B; The same attention kernel generalises to Gemma 4 12B (head 512, GQA 16), plus a q4_0 small-batch GEMM on XMX: Gemma 4 12B QAT with Google's assistant drafter goes from about 40 to 100 t/s on new code",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29643 (Torchit1)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29643",
   "tg_tokens": null,
   "tps": 100,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (bonsai fork)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Mac Mini M4 16GB",
   "id": "lc-disc-29861-macmini-bonsai-before",
   "model": "Bonsai",
   "notes": "prompt processing before cross-backend split (GPU only); config=before-split",
   "params": null,
   "power_w": null,
   "pp_tps": 60,
   "provenance": "community",
   "quant": "PQ2_0",
   "quote": "On bonsai-llama.cpp on Apple Mac Mini M4 16GB, tensor+mode-split GPU + NPU made PQ2_0 go from 60tok/s pp to 115tok/s",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29861 (phhusson)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29861",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (bonsai fork)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Mac Mini M4 16GB",
   "id": "lc-disc-29861-macmini-bonsai-after",
   "model": "Bonsai",
   "notes": "prompt processing with GPU+NPU tensor+mode-split, decode unchanged; config=after-split",
   "params": null,
   "power_w": null,
   "pp_tps": 115,
   "provenance": "community",
   "quant": "PQ2_0",
   "quote": "On bonsai-llama.cpp on Apple Mac Mini M4 16GB, tensor+mode-split GPU + NPU made PQ2_0 go from 60tok/s pp to 115tok/s",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29861 (phhusson)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29861",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (cross-backend split)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Ryzen 8845HS",
   "id": "lc-disc-29861-8845hs-qwen38-before",
   "model": "Qwen3.8-Flash-Next",
   "notes": "prompt processing before split; config=before-split",
   "params": null,
   "power_w": null,
   "pp_tps": 150,
   "provenance": "community",
   "quant": "UD-IQ1_S",
   "quote": "On Ryzen 8845HS, Qwen3.8-Flash-Next-UD-IQ1_S went from 150tok/s to 330tok/s using GPU/NPU tensor+modesplit, without changing decode speed thanks to mode-split",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29861 (phhusson)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29861",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (cross-backend split)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Ryzen 8845HS",
   "id": "lc-disc-29861-8845hs-qwen38-after",
   "model": "Qwen3.8-Flash-Next",
   "notes": "prompt processing with GPU/NPU tensor+mode-split, decode unchanged; config=after-split",
   "params": null,
   "power_w": null,
   "pp_tps": 330,
   "provenance": "community",
   "quant": "UD-IQ1_S",
   "quote": "On Ryzen 8845HS, Qwen3.8-Flash-Next-UD-IQ1_S went from 150tok/s to 330tok/s using GPU/NPU tensor+modesplit, without changing decode speed thanks to mode-split",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29861 (phhusson)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29861",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-disc-29531-5090-qwen3827b-vanilla",
   "model": "Qwen3.8-27B",
   "notes": "1h coding session at 230K context, hot window 3-40k; config=vanilla",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "I've been running Qwen3.8-27B (Q6_K quant) on a single RTX 5090 for coding work; Benchmark (1hr identical coding session, RTX 5090, Q6_K, 230K context length): Avg TPS | 76.2 t/s | 102.7 t/s",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29531 (edwardyoon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29531",
   "tg_tokens": null,
   "tps": 76.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (focus-llama)",
   "batch": null,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-disc-29531-5090-qwen3827b-focus",
   "model": "Qwen3.8-27B",
   "notes": "Declarative Attention hot-window patch, same session (+35% published); config=focus-da",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "I've been running Qwen3.8-27B (Q6_K quant) on a single RTX 5090 for coding work; Benchmark (1hr identical coding session, RTX 5090, Q6_K, 230K context length): Avg TPS | 76.2 t/s | 102.7 t/s",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29531 (edwardyoon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29531",
   "tg_tokens": null,
   "tps": 102.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (RPC)",
   "batch": null,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "Radeon 860M",
   "id": "lc-disc-29662-860m-qwen2515b",
   "model": "Qwen2.5-1.5B",
   "notes": "LAN cluster with a 4-core Linux mini PC via ggml-rpc, published as over 50 tok/s stored at the bound",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "By clustering an AMD Radeon 860M and a 4-core Linux mini PC over a standard LAN; For smaller models like Qwen2.5-1.5B-Q4_K_M, the distributed cluster sustains over 50 tok/s",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29662 (rwilliamspbg-ops)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29662",
   "tg_tokens": null,
   "tps": 50,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "lc-disc-29037-qwen34bi-q80",
   "model": "Qwen3-4B-Instruct",
   "notes": "target 30 tok/s, 5% perplexity loss budget; Q8_0 default failed; config=q80-default",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Measured on an Apple M3 Pro, llama.cpp b9960; | Qwen3-4B-Instruct | 30 tok/s, 5% loss | 24.22 tok/s, failed | `Q4_K_M` @ ngl=33, 30.81 tok/s, 1.73% |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29037 (leeyunseokarchive)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29037",
   "tg_tokens": null,
   "tps": 24.22,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "lc-disc-29037-qwen34bi-pass",
   "model": "Qwen3-4B-Instruct",
   "notes": "target 30 tok/s, 5% perplexity loss budget; Q8_0 default failed; lightest config that passed; ngl=33; config=passed-config",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Measured on an Apple M3 Pro, llama.cpp b9960; | Qwen3-4B-Instruct | 30 tok/s, 5% loss | 24.22 tok/s, failed | `Q4_K_M` @ ngl=33, 30.81 tok/s, 1.73% |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29037 (leeyunseokarchive)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29037",
   "tg_tokens": null,
   "tps": 30.81,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "lc-disc-29037-smolm135-q80",
   "model": "SmolLM2-135M",
   "notes": "target 240 tok/s, 5% loss budget; Q8_0 failed; author notes Q4_K_M measured slower than Q6_K here; config=q80-default",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Measured on an Apple M3 Pro, llama.cpp b9960; | SmolLM2-135M | 240 tok/s, 5% loss | 205.91 tok/s, failed | `Q6_K`, 249.50 tok/s, 0.53% |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29037 (leeyunseokarchive)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29037",
   "tg_tokens": null,
   "tps": 205.91,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "lc-disc-29037-smolm135-pass",
   "model": "SmolLM2-135M",
   "notes": "target 240 tok/s, 5% loss budget; Q8_0 failed; author notes Q4_K_M measured slower than Q6_K here; lightest config that passed; config=passed-config",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "Measured on an Apple M3 Pro, llama.cpp b9960; | SmolLM2-135M | 240 tok/s, 5% loss | 205.91 tok/s, failed | `Q6_K`, 249.50 tok/s, 0.53% |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29037 (leeyunseokarchive)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29037",
   "tg_tokens": null,
   "tps": 249.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "lc-disc-29037-midm2mini-q80",
   "model": "Midm-2.0-Mini",
   "notes": "target 40 tok/s, 5% loss budget; Q8_0 failed; config=q80-default",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Measured on an Apple M3 Pro, llama.cpp b9960; | Midm-2.0-Mini (KR) | 40 tok/s, 5% loss | 34.26 tok/s, failed | `Q4_K_M` @ ngl=48, 44.62 tok/s, 2.58% |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29037 (leeyunseokarchive)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29037",
   "tg_tokens": null,
   "tps": 34.26,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "lc-disc-29037-midm2mini-pass",
   "model": "Midm-2.0-Mini",
   "notes": "target 40 tok/s, 5% loss budget; Q8_0 failed; lightest config that passed; ngl=48; config=passed-config",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Measured on an Apple M3 Pro, llama.cpp b9960; | Midm-2.0-Mini (KR) | 40 tok/s, 5% loss | 34.26 tok/s, failed | `Q4_K_M` @ ngl=48, 44.62 tok/s, 2.58% |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29037 (leeyunseokarchive)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29037",
   "tg_tokens": null,
   "tps": 44.62,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (PrismML fork)",
   "batch": null,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "L4",
   "id": "lc-disc-29387-bonsai-gsm8k-plain",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "plain decode on gsm8k, fastest plain config (PTQ1_0, no speculation); config=gsm8k-plain",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "PTQ1_0",
   "quote": "Setup: `llama-server`, NVIDIA L4 (sm_89, CUDA 12.8), greedy, `-np 1`, target `Ternary-Bonsai-2-27B-PQ2_0.gguf`, drafter Q4_K_M; | GSM8K | 100 | 31.4 | 67.6 | 2.15x | 0.94 -> 0.93 |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29387 (NakliTechie)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29387",
   "tg_tokens": null,
   "tps": 31.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (PrismML fork)",
   "batch": null,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "L4",
   "id": "lc-disc-29387-bonsai-gsm8k-dflash",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "DFlash2 speculative decode on gsm8k (drafter Q4_K_M, draft-n-max 7); config=gsm8k-dflash2",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "PQ2_0",
   "quote": "Setup: `llama-server`, NVIDIA L4 (sm_89, CUDA 12.8), greedy, `-np 1`, target `Ternary-Bonsai-2-27B-PQ2_0.gguf`, drafter Q4_K_M; | GSM8K | 100 | 31.4 | 67.6 | 2.15x | 0.94 -> 0.93 |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29387 (NakliTechie)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29387",
   "tg_tokens": null,
   "tps": 67.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (PrismML fork)",
   "batch": null,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "L4",
   "id": "lc-disc-29387-bonsai-mbpp-plain",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "plain decode on mbpp, fastest plain config (PTQ1_0, no speculation); config=mbpp-plain",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "PTQ1_0",
   "quote": "Setup: `llama-server`, NVIDIA L4 (sm_89, CUDA 12.8), greedy, `-np 1`, target `Ternary-Bonsai-2-27B-PQ2_0.gguf`, drafter Q4_K_M; | MBPP (sanitized) | 100 | 31.6 | 68.4 | 2.16x | 0.80 -> 0.80 |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29387 (NakliTechie)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29387",
   "tg_tokens": null,
   "tps": 31.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (PrismML fork)",
   "batch": null,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "L4",
   "id": "lc-disc-29387-bonsai-mbpp-dflash",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "DFlash2 speculative decode on mbpp (drafter Q4_K_M, draft-n-max 7); config=mbpp-dflash2",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "PQ2_0",
   "quote": "Setup: `llama-server`, NVIDIA L4 (sm_89, CUDA 12.8), greedy, `-np 1`, target `Ternary-Bonsai-2-27B-PQ2_0.gguf`, drafter Q4_K_M; | MBPP (sanitized) | 100 | 31.6 | 68.4 | 2.16x | 0.80 -> 0.80 |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29387 (NakliTechie)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29387",
   "tg_tokens": null,
   "tps": 68.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (PrismML fork)",
   "batch": null,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "L4",
   "id": "lc-disc-29387-bonsai-math500-plain",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "plain decode on math500, fastest plain config (PTQ1_0, no speculation); config=math500-plain",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "PTQ1_0",
   "quote": "Setup: `llama-server`, NVIDIA L4 (sm_89, CUDA 12.8), greedy, `-np 1`, target `Ternary-Bonsai-2-27B-PQ2_0.gguf`, drafter Q4_K_M; | MATH-500 | 100 | 30.6 | 67.8 | 2.22x | 0.76 -> 0.75 |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29387 (NakliTechie)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29387",
   "tg_tokens": null,
   "tps": 30.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (PrismML fork)",
   "batch": null,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "L4",
   "id": "lc-disc-29387-bonsai-math500-dflash",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "DFlash2 speculative decode on math500 (drafter Q4_K_M, draft-n-max 7); config=math500-dflash2",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "PQ2_0",
   "quote": "Setup: `llama-server`, NVIDIA L4 (sm_89, CUDA 12.8), greedy, `-np 1`, target `Ternary-Bonsai-2-27B-PQ2_0.gguf`, drafter Q4_K_M; | MATH-500 | 100 | 30.6 | 67.8 | 2.22x | 0.76 -> 0.75 |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29387 (NakliTechie)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29387",
   "tg_tokens": null,
   "tps": 67.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (PrismML fork)",
   "batch": null,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "L4",
   "id": "lc-disc-29387-bonsai-mtbench-plain",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "plain decode on mtbench, fastest plain config (PTQ1_0, no speculation); config=mtbench-plain",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "PTQ1_0",
   "quote": "Setup: `llama-server`, NVIDIA L4 (sm_89, CUDA 12.8), greedy, `-np 1`, target `Ternary-Bonsai-2-27B-PQ2_0.gguf`, drafter Q4_K_M; | MT-Bench turn 1 | 80 | 31.3 | 42.9 | 1.37x | - |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29387 (NakliTechie)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29387",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (PrismML fork)",
   "batch": null,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "L4",
   "id": "lc-disc-29387-bonsai-mtbench-dflash",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "DFlash2 speculative decode on mtbench (drafter Q4_K_M, draft-n-max 7); config=mtbench-dflash2",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "PQ2_0",
   "quote": "Setup: `llama-server`, NVIDIA L4 (sm_89, CUDA 12.8), greedy, `-np 1`, target `Ternary-Bonsai-2-27B-PQ2_0.gguf`, drafter Q4_K_M; | MT-Bench turn 1 | 80 | 31.3 | 42.9 | 1.37x | - |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29387 (NakliTechie)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29387",
   "tg_tokens": null,
   "tps": 42.9,
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": null,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "K12-1 Ryzen H255 64gb",
   "id": "lc-disc-29253-qwen34b-solo",
   "model": "Qwen3 4B",
   "notes": "pp64/tg32 solo baseline on K12-1 (no RPC); config=solo",
   "params": null,
   "power_w": null,
   "pp_tps": 393.674985,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "K12-1 Ryzen H255 64gb; | Qwen3 4B Q8_0, K12-1 solo | 393.674985 +/- 12.040010 | 18.040894 +/- 0.429658 |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29253 (Alexander2995)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29253",
   "tg_tokens": null,
   "tps": 18.040894,
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": null,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "K12-1 Ryzen H255 64gb",
   "id": "lc-disc-29253-gemma431b-solo",
   "model": "Gemma-4-31B",
   "notes": "pp64/tg32 solo baseline on K12-1 (no RPC); config=solo",
   "params": null,
   "power_w": null,
   "pp_tps": 73.446822,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "K12-1 Ryzen H255 64gb; | Gemma 4 31B Q4_0, K12-1 solo | 73.446822 +/- 0.900541 | 4.486488 +/- 0.019191 |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29253 (Alexander2995)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29253",
   "tg_tokens": null,
   "tps": 4.486488,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (RPC)",
   "batch": null,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "2 GMKtec K12",
   "id": "lc-disc-29253-qwen34b-upstream",
   "model": "Qwen3 4B",
   "notes": "pp64/tg32 over TCP RPC, upstream (no graph cache), 2.5G Ethernet; config=upstream",
   "params": null,
   "power_w": null,
   "pp_tps": 63.902555,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "2 GMKtec K12; transport - Ethernet 2.5G; API - Vulkan; | Qwen3 4B Q8_0, upstream TCP RPC | 63.902555 +/- 1.447661 | 5.411568 +/- 0.002232 |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29253 (Alexander2995)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29253",
   "tg_tokens": null,
   "tps": 5.411568,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (RPC)",
   "batch": null,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "2 GMKtec K12",
   "id": "lc-disc-29253-qwen34b-graphcache",
   "model": "Qwen3 4B",
   "notes": "pp64/tg32 over TCP RPC with per-UID graph caching; config=graph-cache",
   "params": null,
   "power_w": null,
   "pp_tps": 66.181528,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "2 GMKtec K12; transport - Ethernet 2.5G; API - Vulkan; | Qwen3 4B Q8_0, graph-cache TCP RPC | 66.181528 +/- 0.286268 | **13.211557 +/- 0.324566** |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29253 (Alexander2995)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29253",
   "tg_tokens": null,
   "tps": 13.211557,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (RPC)",
   "batch": null,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "2 GMKtec K12",
   "id": "lc-disc-29253-gemma431b-upstream",
   "model": "Gemma-4-31B",
   "notes": "pp64/tg32 over TCP RPC, upstream (no graph cache), 2.5G Ethernet; config=upstream",
   "params": null,
   "power_w": null,
   "pp_tps": 18.288838,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "2 GMKtec K12; transport - Ethernet 2.5G; API - Vulkan; | Gemma 4 31B Q4_0, upstream TCP RPC | 18.288838 +/- 0.026488 | 1.637130 +/- 0.001639 |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29253 (Alexander2995)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29253",
   "tg_tokens": null,
   "tps": 1.63713,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (RPC)",
   "batch": null,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "2 GMKtec K12",
   "id": "lc-disc-29253-gemma431b-graphcache",
   "model": "Gemma-4-31B",
   "notes": "pp64/tg32 over TCP RPC with per-UID graph caching; config=graph-cache",
   "params": null,
   "power_w": null,
   "pp_tps": 18.7754,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "2 GMKtec K12; transport - Ethernet 2.5G; API - Vulkan; | Gemma 4 31B Q4_0, graph-cache TCP RPC | 18.775400 +/- 0.012552 | 4.111661 +/- 0.139460 |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29253 (Alexander2995)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29253",
   "tg_tokens": null,
   "tps": 4.111661,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "4070 laptop card",
   "id": "lc-disc-29249-4070lp-gpu",
   "model": "Qwen3-4B-Instruct-2507",
   "notes": "code-writing throughput on the GPU; config=gpu",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "A small model (Qwen3-4B-Instruct-2507, Q4_K_M) writes code; Measured on a 4070 laptop card (8 GB), Q4_K_M: writing 72 tok/s on the GPU, 11.4 tok/s on 4 CPU cores with `-ngl 0`",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29249 (1picassoai)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29249",
   "tg_tokens": null,
   "tps": 72,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "4070 laptop card",
   "id": "lc-disc-29249-4070lp-cpu",
   "model": "Qwen3-4B-Instruct-2507",
   "notes": "code-writing on 4 CPU cores, -ngl 0; threads=4; ngl=0; config=cpu-ngl0",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "A small model (Qwen3-4B-Instruct-2507, Q4_K_M) writes code; Measured on a 4070 laptop card (8 GB), Q4_K_M: writing 72 tok/s on the GPU, 11.4 tok/s on 4 CPU cores with `-ngl 0`",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29249 (1picassoai)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29249",
   "tg_tokens": null,
   "tps": 11.4,
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": null,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "780M",
   "id": "lc-disc-29072-780m-gemma26a4b-before",
   "model": "Gemma 4 26B-A4B",
   "notes": "tg before narrow-MoE MMVQ patch, same-prompt speed test; config=before",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "QAT-Q4_0",
   "quote": "Question for Vulkan maintainers about gfx1103 / 780M (MMVQ related); gemma-4-26B-A4B-it-QAT-Q4_0.gguf; | TG throughput | 20.59 t/s | 23.11 t/s | +12.2% |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29072 (arch-btw)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29072",
   "tg_tokens": null,
   "tps": 20.59,
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": null,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "780M",
   "id": "lc-disc-29072-780m-gemma26a4b-after",
   "model": "Gemma 4 26B-A4B",
   "notes": "tg after narrow-MoE MMVQ patch (+12.2% published); config=after",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "QAT-Q4_0",
   "quote": "Question for Vulkan maintainers about gfx1103 / 780M (MMVQ related); gemma-4-26B-A4B-it-QAT-Q4_0.gguf; | TG throughput | 20.59 t/s | 23.11 t/s | +12.2% |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29072 (arch-btw)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29072",
   "tg_tokens": null,
   "tps": 23.11,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 50000,
   "date": "2026-09",
   "flags": [],
   "hardware": "5070/2x5060/4060",
   "id": "lc-disc-28767-flashnext-iq4xs-win",
   "model": "Qwen3.8-Flash-Next",
   "notes": "4-GPU layer split -sm layer -ts 1.45,0.80,0.85,0.90 with -n-cpu-moe 9, DDR4-3200, Windows; threads=8; config=win-4gpu-cpumoe",
   "params": null,
   "power_w": null,
   "pp_tps": 170,
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "E:\\llama\\llama-bin\\llama-server.exe -m Qwen3.8-Flash-Next-UD-IQ4_XS-00001-of-00003.gguf; My setup of 4gpus is quite retarded in terms of pci-lane bottleneck as its x8/x4/x1/x1 for 5070/2x5060/4060; My results of pp 170 and decode of 12tok/s at 50k ctx seems lower to what people state at their ddr5 multi-chan memories",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28767 (pepedombo)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28767",
   "tg_tokens": null,
   "tps": 12,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 50000,
   "date": "2026-09",
   "flags": [],
   "hardware": "5070/2x5060/4060",
   "id": "lc-disc-28767-flashnext-q3xxs-linux",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Ubuntu, full offload at 80k ctx, decode drops from 45 to 30-35 at 50k stored at low end, prefill 550 vs 350 on Windows; config=linux-q3xxs",
   "params": null,
   "power_w": null,
   "pp_tps": 550,
   "provenance": "community",
   "quant": "Q3_XXS",
   "quote": "E:\\llama\\llama-bin\\llama-server.exe -m Qwen3.8-Flash-Next-UD-IQ4_XS-00001-of-00003.gguf; its x8/x4/x1/x1 for 5070/2x5060/4060; On Linux Flash-Next behaves completely different: Q3_XXS (full offload 80k ctx): starts 45tok/s and drops to 30-35 at 50k ctx (no mtp, acceptable result), PP jumped from windows 350tok/s to 550tok/s",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28767 (pepedombo)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28767",
   "tg_tokens": null,
   "tps": 30,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "5070/2x5060/4060",
   "id": "lc-disc-28767-flashnext-q3xxs-win",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Windows prefill for the same Q3_XXS full-offload run (550 on Ubuntu); config=win-q3xxs",
   "params": null,
   "power_w": null,
   "pp_tps": 350,
   "provenance": "community",
   "quant": "Q3_XXS",
   "quote": "E:\\llama\\llama-bin\\llama-server.exe -m Qwen3.8-Flash-Next-UD-IQ4_XS-00001-of-00003.gguf; its x8/x4/x1/x1 for 5070/2x5060/4060; On Linux Flash-Next behaves completely different: Q3_XXS (full offload 80k ctx): starts 45tok/s and drops to 30-35 at 50k ctx (no mtp, acceptable result), PP jumped from windows 350tok/s to 550tok/s",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28767 (pepedombo)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28767",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 40000,
   "date": "2026-09",
   "flags": [],
   "hardware": "5070/2x5060/4060",
   "id": "lc-disc-28767-flashnext-q4xs-linux",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Ubuntu, partial offload with autofit on DDR4-3200 dual-channel, 20 tok/s vs 10 on Windows; config=linux-q4xs-autofit",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_XS",
   "quote": "E:\\llama\\llama-bin\\llama-server.exe -m Qwen3.8-Flash-Next-UD-IQ4_XS-00001-of-00003.gguf; its x8/x4/x1/x1 for 5070/2x5060/4060; Q4_XS (partial offload, autofit): stable 20tok/s at 40k ctx instead of 10, slower due to offload (ddr4 3200 dual-chan memory)",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28767 (pepedombo)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28767",
   "tg_tokens": null,
   "tps": 20,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 16384,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 4080 SUPER",
   "id": "lc-disc-28363-flashnext-control",
   "model": "Qwen3.8-Flash-Next",
   "notes": "expert-only CPU control at 16K, three 500-token runs; config=cpu-control",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "My test box: RTX 4080 Super 16 GiB, i9-14900KF, 96 GiB DDR5-4000, Windows, Qwen3.8-Flash-Next 176.94B IQ4_XS-family GGUF (QWEN4EXP arch), pinned llama.cpp revision `4e97ac8`; same-path Native expert-only CPU control: 15.135 tok/s",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28363 (zhaoyilun)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28363",
   "tg_tokens": null,
   "tps": 15.135,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 16384,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 4080 SUPER",
   "id": "lc-disc-28363-flashnext-pool4g",
   "model": "Qwen3.8-Flash-Next",
   "notes": "fixed GPU expert pool 4 GiB at 16K; config=pool-4gib",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "My test box: RTX 4080 Super 16 GiB, i9-14900KF, 96 GiB DDR5-4000, Windows, Qwen3.8-Flash-Next 176.94B IQ4_XS-family GGUF (QWEN4EXP arch), pinned llama.cpp revision `4e97ac8`; fixed cache 4 GiB: 16.096 tok/s",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28363 (zhaoyilun)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28363",
   "tg_tokens": null,
   "tps": 16.096,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 16384,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 4080 SUPER",
   "id": "lc-disc-28363-flashnext-pool6g",
   "model": "Qwen3.8-Flash-Next",
   "notes": "fixed GPU expert pool 6 GiB at 16K; config=pool-6gib",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "My test box: RTX 4080 Super 16 GiB, i9-14900KF, 96 GiB DDR5-4000, Windows, Qwen3.8-Flash-Next 176.94B IQ4_XS-family GGUF (QWEN4EXP arch), pinned llama.cpp revision `4e97ac8`; fixed cache 6 GiB: 17.275 tok/s",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28363 (zhaoyilun)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28363",
   "tg_tokens": null,
   "tps": 17.275,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 16384,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 4080 SUPER",
   "id": "lc-disc-28363-flashnext-pool8g",
   "model": "Qwen3.8-Flash-Next",
   "notes": "fixed GPU expert pool 8 GiB at 16K, +23.91% over control; config=pool-8gib",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "My test box: RTX 4080 Super 16 GiB, i9-14900KF, 96 GiB DDR5-4000, Windows, Qwen3.8-Flash-Next 176.94B IQ4_XS-family GGUF (QWEN4EXP arch), pinned llama.cpp revision `4e97ac8`; fixed cache 8 GiB: 18.753 tok/s",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28363 (zhaoyilun)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28363",
   "tg_tokens": null,
   "tps": 18.753,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 16384,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 4080 SUPER",
   "id": "lc-disc-28363-flashnext-pool8g-nograph",
   "model": "Qwen3.8-Flash-Next",
   "notes": "8 GiB pool with CUDA Graph disabled; config=pool-8gib-graphs-off",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "My test box: RTX 4080 Super 16 GiB, i9-14900KF, 96 GiB DDR5-4000, Windows, Qwen3.8-Flash-Next 176.94B IQ4_XS-family GGUF (QWEN4EXP arch), pinned llama.cpp revision `4e97ac8`; the same 8 GiB config with graphs off drops to 13.567 tok/s",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28363 (zhaoyilun)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28363",
   "tg_tokens": null,
   "tps": 13.567,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 2048,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 4080 SUPER",
   "id": "lc-disc-28363-flashnext-ctx2k",
   "model": "Qwen3.8-Flash-Next",
   "notes": "8 GiB pool at 2K context; config=ctx-2k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "My test box: RTX 4080 Super 16 GiB, i9-14900KF, 96 GiB DDR5-4000, Windows, Qwen3.8-Flash-Next 176.94B IQ4_XS-family GGUF (QWEN4EXP arch), pinned llama.cpp revision `4e97ac8`; Short-context repeats: 19.531 tok/s at 2K",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28363 (zhaoyilun)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28363",
   "tg_tokens": null,
   "tps": 19.531,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 8192,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 4080 SUPER",
   "id": "lc-disc-28363-flashnext-ctx8k",
   "model": "Qwen3.8-Flash-Next",
   "notes": "8 GiB pool at 8K context; config=ctx-8k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "My test box: RTX 4080 Super 16 GiB, i9-14900KF, 96 GiB DDR5-4000, Windows, Qwen3.8-Flash-Next 176.94B IQ4_XS-family GGUF (QWEN4EXP arch), pinned llama.cpp revision `4e97ac8`; 18.321 at 8K",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28363 (zhaoyilun)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28363",
   "tg_tokens": null,
   "tps": 18.321,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 262144,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 4080 SUPER",
   "id": "lc-disc-28363-flashnext-262k-q8q4",
   "model": "Qwen3.8-Flash-Next",
   "notes": "q8_0 K / q4_0 V KV with 4 GiB pool at 262K, p50 170.7 ms; config=262k-q8q4-pool4g",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "My test box: RTX 4080 Super 16 GiB, i9-14900KF, 96 GiB DDR5-4000, Windows, Qwen3.8-Flash-Next 176.94B IQ4_XS-family GGUF (QWEN4EXP arch), pinned llama.cpp revision `4e97ac8`; 262K, q8-q4 KV + 4 GiB pool: 5.776 tok/s (p50 170.7 ms / p95 197.2 ms), peak 13,611 MiB",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28363 (zhaoyilun)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28363",
   "tg_tokens": null,
   "tps": 5.776,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 262144,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 4080 SUPER",
   "id": "lc-disc-28363-flashnext-262k-q4q4",
   "model": "Qwen3.8-Flash-Next",
   "notes": "q4_0 KV with 6 GiB pool at 262K; config=262k-q4q4-pool6g",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "My test box: RTX 4080 Super 16 GiB, i9-14900KF, 96 GiB DDR5-4000, Windows, Qwen3.8-Flash-Next 176.94B IQ4_XS-family GGUF (QWEN4EXP arch), pinned llama.cpp revision `4e97ac8`; q4-q4 KV + 6 GiB pool: 5.649 tok/s \u2014 the q8-q4 + 4 GiB combo is slightly faster while using less VRAM",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28363 (zhaoyilun)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28363",
   "tg_tokens": null,
   "tps": 5.649,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 262144,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 4080 SUPER",
   "id": "lc-disc-28363-flashnext-262k-collapse",
   "model": "Qwen3.8-Flash-Next",
   "notes": "q8-q4 KV + 6 GiB pool at 262K, 250-token run collapsed via Windows shared-memory paging; config=262k-collapse-paging",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "My test box: RTX 4080 Super 16 GiB, i9-14900KF, 96 GiB DDR5-4000, Windows, Qwen3.8-Flash-Next 176.94B IQ4_XS-family GGUF (QWEN4EXP arch), pinned llama.cpp revision `4e97ac8`; a 250-token run of it collapsed to 0.71 tok/s mean with p95 at 7.5 s, because total demand crossed physical VRAM and Windows started paging through shared memory",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28363 (zhaoyilun)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28363",
   "tg_tokens": null,
   "tps": 0.71,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (KVMem fork)",
   "batch": null,
   "ctx": 262144,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 5060 Ti 16GB",
   "id": "lc-disc-28894-qwen3827b-iq3",
   "model": "Qwen3.8-27B",
   "notes": "ISTA Qwen3.8-27B IQ3_S with MTP, vision on GPU, Q8 KV, 36K history window on GPU + 16K reserved; overall prefill 243 tok/s, last-512 decode 29 tok/s; config=iq3-mtp",
   "params": null,
   "power_w": null,
   "pp_tps": 437,
   "provenance": "community",
   "quant": "IQ3_S",
   "quote": "RTX 5060 Ti 16GB, 32GB RAM (WSL2, ~19.5GB visible to Linux); **IQ3:** ISTA Qwen3.8-27B IQ3_S \u2014 MTP, vision on GPU, Q8 KV (36K history on GPU / 16K reserved for new tokens); 256K tool task (33 requests, ending at 262058 / 262144 tokens); | Prefill, first pass | 437 tok/s |; | Decode over the tool rounds | 30 tok/s |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28894 (wangleye)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28894",
   "tg_tokens": null,
   "tps": 30,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (KVMem fork)",
   "batch": null,
   "ctx": 262144,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 5060 Ti 16GB",
   "id": "lc-disc-28894-qwen3827b-iq4",
   "model": "Qwen3.8-27B",
   "notes": "Unsloth Qwen3.8-27B IQ4_XS with MTP, vision on CPU, Q5 KV, 32K window + 12K reserved; overall prefill 255 tok/s, last-512 decode 35 tok/s; config=iq4-mtp",
   "params": null,
   "power_w": null,
   "pp_tps": 466,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "RTX 5060 Ti 16GB, 32GB RAM (WSL2, ~19.5GB visible to Linux); **IQ4:** Unsloth Qwen3.8-27B IQ4_XS \u2014 MTP, vision on CPU, Q5 KV (32K / 12K); 256K tool task (33 requests, ending at 262058 / 262144 tokens); | Prefill, first pass | 466 tok/s |; | Decode over the tool rounds | 33 tok/s |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28894 (wangleye)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28894",
   "tg_tokens": null,
   "tps": 33,
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": null,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "Bosgame M5 (Ryzen AI Max+ 395, Radeon 8060S, 128 GB)",
   "id": "lc-disc-28512-flashnext-shortcode-before",
   "model": "Qwen3.8-Flash-Next",
   "notes": "apepojken build, stock UD-IQ4_XS, full Q8_0 draft head, draft-mtp + ngram-mod n-max 6; config=shortcode-before",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "I run Qwen3.8-Flash-Next as the model server for my agents on a Bosgame M5 (Ryzen AI Max+ 395, Radeon 8060S, 128 GB), on Vulkan with the RADV driver; Before and after, single stream, greedy, nothing cached; | short code, no context | 49.0 | 32.8 | 58.1 | |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28512 (drluoto)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28512",
   "tg_tokens": null,
   "tps": 49.0,
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": null,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "Bosgame M5 (Ryzen AI Max+ 395, Radeon 8060S, 128 GB)",
   "id": "lc-disc-28512-flashnext-shortcode-nospec",
   "model": "Qwen3.8-Flash-Next",
   "notes": "same build without speculation; config=shortcode-nospec",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "I run Qwen3.8-Flash-Next as the model server for my agents on a Bosgame M5 (Ryzen AI Max+ 395, Radeon 8060S, 128 GB), on Vulkan with the RADV driver; Before and after, single stream, greedy, nothing cached; | short code, no context | 49.0 | 32.8 | 58.1 | |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28512 (drluoto)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28512",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": null,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "Bosgame M5 (Ryzen AI Max+ 395, Radeon 8060S, 128 GB)",
   "id": "lc-disc-28512-flashnext-shortcode-now",
   "model": "Qwen3.8-Flash-Next",
   "notes": "after row-id hoisting (#28501), p-min 0 full 3-token drafts, trimmed-vocab FR-Spec draft head; config=shortcode-now",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "I run Qwen3.8-Flash-Next as the model server for my agents on a Bosgame M5 (Ryzen AI Max+ 395, Radeon 8060S, 128 GB), on Vulkan with the RADV driver; Before and after, single stream, greedy, nothing cached; | short code, no context | 49.0 | 32.8 | 58.1 | |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28512 (drluoto)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28512",
   "tg_tokens": null,
   "tps": 58.1,
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": null,
   "ctx": 8192,
   "date": "2026-09",
   "flags": [],
   "hardware": "Bosgame M5 (Ryzen AI Max+ 395, Radeon 8060S, 128 GB)",
   "id": "lc-disc-28512-flashnext-newcode8k-before",
   "model": "Qwen3.8-Flash-Next",
   "notes": "apepojken build, stock UD-IQ4_XS, full Q8_0 draft head, draft-mtp + ngram-mod n-max 6; config=newcode8k-before",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "I run Qwen3.8-Flash-Next as the model server for my agents on a Bosgame M5 (Ryzen AI Max+ 395, Radeon 8060S, 128 GB), on Vulkan with the RADV driver; Before and after, single stream, greedy, nothing cached; | new code @8k | 30.7 | 29.6 | 41.6 | |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28512 (drluoto)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28512",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": null,
   "ctx": 8192,
   "date": "2026-09",
   "flags": [],
   "hardware": "Bosgame M5 (Ryzen AI Max+ 395, Radeon 8060S, 128 GB)",
   "id": "lc-disc-28512-flashnext-newcode8k-nospec",
   "model": "Qwen3.8-Flash-Next",
   "notes": "same build without speculation; config=newcode8k-nospec",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "I run Qwen3.8-Flash-Next as the model server for my agents on a Bosgame M5 (Ryzen AI Max+ 395, Radeon 8060S, 128 GB), on Vulkan with the RADV driver; Before and after, single stream, greedy, nothing cached; | new code @8k | 30.7 | 29.6 | 41.6 | |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28512 (drluoto)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28512",
   "tg_tokens": null,
   "tps": 29.6,
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": null,
   "ctx": 8192,
   "date": "2026-09",
   "flags": [],
   "hardware": "Bosgame M5 (Ryzen AI Max+ 395, Radeon 8060S, 128 GB)",
   "id": "lc-disc-28512-flashnext-newcode8k-now",
   "model": "Qwen3.8-Flash-Next",
   "notes": "after row-id hoisting (#28501), p-min 0 full 3-token drafts, trimmed-vocab FR-Spec draft head; config=newcode8k-now",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "I run Qwen3.8-Flash-Next as the model server for my agents on a Bosgame M5 (Ryzen AI Max+ 395, Radeon 8060S, 128 GB), on Vulkan with the RADV driver; Before and after, single stream, greedy, nothing cached; | new code @8k | 30.7 | 29.6 | 41.6 | |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28512 (drluoto)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28512",
   "tg_tokens": null,
   "tps": 41.6,
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": null,
   "ctx": 8192,
   "date": "2026-09",
   "flags": [],
   "hardware": "Bosgame M5 (Ryzen AI Max+ 395, Radeon 8060S, 128 GB)",
   "id": "lc-disc-28512-flashnext-prose8k-before",
   "model": "Qwen3.8-Flash-Next",
   "notes": "apepojken build, stock UD-IQ4_XS, full Q8_0 draft head, draft-mtp + ngram-mod n-max 6; config=prose8k-before",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "I run Qwen3.8-Flash-Next as the model server for my agents on a Bosgame M5 (Ryzen AI Max+ 395, Radeon 8060S, 128 GB), on Vulkan with the RADV driver; Before and after, single stream, greedy, nothing cached; | prose @8k | 23.6 | 29.3 | 30.1 | |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28512 (drluoto)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28512",
   "tg_tokens": null,
   "tps": 23.6,
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": null,
   "ctx": 8192,
   "date": "2026-09",
   "flags": [],
   "hardware": "Bosgame M5 (Ryzen AI Max+ 395, Radeon 8060S, 128 GB)",
   "id": "lc-disc-28512-flashnext-prose8k-nospec",
   "model": "Qwen3.8-Flash-Next",
   "notes": "same build without speculation; config=prose8k-nospec",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "I run Qwen3.8-Flash-Next as the model server for my agents on a Bosgame M5 (Ryzen AI Max+ 395, Radeon 8060S, 128 GB), on Vulkan with the RADV driver; Before and after, single stream, greedy, nothing cached; | prose @8k | 23.6 | 29.3 | 30.1 | |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28512 (drluoto)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28512",
   "tg_tokens": null,
   "tps": 29.3,
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": null,
   "ctx": 8192,
   "date": "2026-09",
   "flags": [],
   "hardware": "Bosgame M5 (Ryzen AI Max+ 395, Radeon 8060S, 128 GB)",
   "id": "lc-disc-28512-flashnext-prose8k-now",
   "model": "Qwen3.8-Flash-Next",
   "notes": "after row-id hoisting (#28501), p-min 0 full 3-token drafts, trimmed-vocab FR-Spec draft head; config=prose8k-now",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "I run Qwen3.8-Flash-Next as the model server for my agents on a Bosgame M5 (Ryzen AI Max+ 395, Radeon 8060S, 128 GB), on Vulkan with the RADV driver; Before and after, single stream, greedy, nothing cached; | prose @8k | 23.6 | 29.3 | 30.1 | |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28512 (drluoto)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28512",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": null,
   "ctx": 8192,
   "date": "2026-09",
   "flags": [],
   "hardware": "Bosgame M5 (Ryzen AI Max+ 395, Radeon 8060S, 128 GB)",
   "id": "lc-disc-28512-flashnext-filerewrite8k-before",
   "model": "Qwen3.8-Flash-Next",
   "notes": "apepojken build, stock UD-IQ4_XS, full Q8_0 draft head, draft-mtp + ngram-mod n-max 6; config=filerewrite8k-before",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "I run Qwen3.8-Flash-Next as the model server for my agents on a Bosgame M5 (Ryzen AI Max+ 395, Radeon 8060S, 128 GB), on Vulkan with the RADV driver; Before and after, single stream, greedy, nothing cached; | file rewrite @8k | 30.9 | 29.0 | 55.4 | 63.2 |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28512 (drluoto)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28512",
   "tg_tokens": null,
   "tps": 30.9,
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": null,
   "ctx": 8192,
   "date": "2026-09",
   "flags": [],
   "hardware": "Bosgame M5 (Ryzen AI Max+ 395, Radeon 8060S, 128 GB)",
   "id": "lc-disc-28512-flashnext-filerewrite8k-nospec",
   "model": "Qwen3.8-Flash-Next",
   "notes": "same build without speculation; config=filerewrite8k-nospec",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "I run Qwen3.8-Flash-Next as the model server for my agents on a Bosgame M5 (Ryzen AI Max+ 395, Radeon 8060S, 128 GB), on Vulkan with the RADV driver; Before and after, single stream, greedy, nothing cached; | file rewrite @8k | 30.9 | 29.0 | 55.4 | 63.2 |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28512 (drluoto)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28512",
   "tg_tokens": null,
   "tps": 29.0,
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": null,
   "ctx": 8192,
   "date": "2026-09",
   "flags": [],
   "hardware": "Bosgame M5 (Ryzen AI Max+ 395, Radeon 8060S, 128 GB)",
   "id": "lc-disc-28512-flashnext-filerewrite8k-now",
   "model": "Qwen3.8-Flash-Next",
   "notes": "after row-id hoisting (#28501), p-min 0 full 3-token drafts, trimmed-vocab FR-Spec draft head; config=filerewrite8k-now",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "I run Qwen3.8-Flash-Next as the model server for my agents on a Bosgame M5 (Ryzen AI Max+ 395, Radeon 8060S, 128 GB), on Vulkan with the RADV driver; Before and after, single stream, greedy, nothing cached; | file rewrite @8k | 30.9 | 29.0 | 55.4 | 63.2 |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28512 (drluoto)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28512",
   "tg_tokens": null,
   "tps": 55.4,
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": null,
   "ctx": 8192,
   "date": "2026-09",
   "flags": [],
   "hardware": "Bosgame M5 (Ryzen AI Max+ 395, Radeon 8060S, 128 GB)",
   "id": "lc-disc-28512-flashnext-filerewrite8k-nmax6",
   "model": "Qwen3.8-Flash-Next",
   "notes": "tuned stack with n-max 6 drafts; config=filerewrite8k-nmax6",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "I run Qwen3.8-Flash-Next as the model server for my agents on a Bosgame M5 (Ryzen AI Max+ 395, Radeon 8060S, 128 GB), on Vulkan with the RADV driver; Before and after, single stream, greedy, nothing cached; | file rewrite @8k | 30.9 | 29.0 | 55.4 | 63.2 |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28512 (drluoto)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28512",
   "tg_tokens": null,
   "tps": 63.2,
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": null,
   "ctx": 32768,
   "date": "2026-09",
   "flags": [],
   "hardware": "Bosgame M5 (Ryzen AI Max+ 395, Radeon 8060S, 128 GB)",
   "id": "lc-disc-28512-flashnext-newcode32k-before",
   "model": "Qwen3.8-Flash-Next",
   "notes": "apepojken build, stock UD-IQ4_XS, full Q8_0 draft head, draft-mtp + ngram-mod n-max 6; config=newcode32k-before",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "I run Qwen3.8-Flash-Next as the model server for my agents on a Bosgame M5 (Ryzen AI Max+ 395, Radeon 8060S, 128 GB), on Vulkan with the RADV driver; Before and after, single stream, greedy, nothing cached; | new code @32k | 34.5 | 25.1 | 37.8 | |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28512 (drluoto)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28512",
   "tg_tokens": null,
   "tps": 34.5,
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": null,
   "ctx": 32768,
   "date": "2026-09",
   "flags": [],
   "hardware": "Bosgame M5 (Ryzen AI Max+ 395, Radeon 8060S, 128 GB)",
   "id": "lc-disc-28512-flashnext-newcode32k-nospec",
   "model": "Qwen3.8-Flash-Next",
   "notes": "same build without speculation; config=newcode32k-nospec",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "I run Qwen3.8-Flash-Next as the model server for my agents on a Bosgame M5 (Ryzen AI Max+ 395, Radeon 8060S, 128 GB), on Vulkan with the RADV driver; Before and after, single stream, greedy, nothing cached; | new code @32k | 34.5 | 25.1 | 37.8 | |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28512 (drluoto)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28512",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": null,
   "ctx": 32768,
   "date": "2026-09",
   "flags": [],
   "hardware": "Bosgame M5 (Ryzen AI Max+ 395, Radeon 8060S, 128 GB)",
   "id": "lc-disc-28512-flashnext-newcode32k-now",
   "model": "Qwen3.8-Flash-Next",
   "notes": "after row-id hoisting (#28501), p-min 0 full 3-token drafts, trimmed-vocab FR-Spec draft head; config=newcode32k-now",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "I run Qwen3.8-Flash-Next as the model server for my agents on a Bosgame M5 (Ryzen AI Max+ 395, Radeon 8060S, 128 GB), on Vulkan with the RADV driver; Before and after, single stream, greedy, nothing cached; | new code @32k | 34.5 | 25.1 | 37.8 | |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28512 (drluoto)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28512",
   "tg_tokens": null,
   "tps": 37.8,
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": null,
   "ctx": 32768,
   "date": "2026-09",
   "flags": [],
   "hardware": "Bosgame M5 (Ryzen AI Max+ 395, Radeon 8060S, 128 GB)",
   "id": "lc-disc-28512-flashnext-filerewrite32k-before",
   "model": "Qwen3.8-Flash-Next",
   "notes": "apepojken build, stock UD-IQ4_XS, full Q8_0 draft head, draft-mtp + ngram-mod n-max 6; config=filerewrite32k-before",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "I run Qwen3.8-Flash-Next as the model server for my agents on a Bosgame M5 (Ryzen AI Max+ 395, Radeon 8060S, 128 GB), on Vulkan with the RADV driver; Before and after, single stream, greedy, nothing cached; | file rewrite @32k | 37.4 | 25.1 | 48.5 | 55.2 |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28512 (drluoto)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28512",
   "tg_tokens": null,
   "tps": 37.4,
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": null,
   "ctx": 32768,
   "date": "2026-09",
   "flags": [],
   "hardware": "Bosgame M5 (Ryzen AI Max+ 395, Radeon 8060S, 128 GB)",
   "id": "lc-disc-28512-flashnext-filerewrite32k-nospec",
   "model": "Qwen3.8-Flash-Next",
   "notes": "same build without speculation; config=filerewrite32k-nospec",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "I run Qwen3.8-Flash-Next as the model server for my agents on a Bosgame M5 (Ryzen AI Max+ 395, Radeon 8060S, 128 GB), on Vulkan with the RADV driver; Before and after, single stream, greedy, nothing cached; | file rewrite @32k | 37.4 | 25.1 | 48.5 | 55.2 |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28512 (drluoto)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28512",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": null,
   "ctx": 32768,
   "date": "2026-09",
   "flags": [],
   "hardware": "Bosgame M5 (Ryzen AI Max+ 395, Radeon 8060S, 128 GB)",
   "id": "lc-disc-28512-flashnext-filerewrite32k-now",
   "model": "Qwen3.8-Flash-Next",
   "notes": "after row-id hoisting (#28501), p-min 0 full 3-token drafts, trimmed-vocab FR-Spec draft head; config=filerewrite32k-now",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "I run Qwen3.8-Flash-Next as the model server for my agents on a Bosgame M5 (Ryzen AI Max+ 395, Radeon 8060S, 128 GB), on Vulkan with the RADV driver; Before and after, single stream, greedy, nothing cached; | file rewrite @32k | 37.4 | 25.1 | 48.5 | 55.2 |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28512 (drluoto)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28512",
   "tg_tokens": null,
   "tps": 48.5,
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": null,
   "ctx": 32768,
   "date": "2026-09",
   "flags": [],
   "hardware": "Bosgame M5 (Ryzen AI Max+ 395, Radeon 8060S, 128 GB)",
   "id": "lc-disc-28512-flashnext-filerewrite32k-nmax6",
   "model": "Qwen3.8-Flash-Next",
   "notes": "tuned stack with n-max 6 drafts; config=filerewrite32k-nmax6",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "I run Qwen3.8-Flash-Next as the model server for my agents on a Bosgame M5 (Ryzen AI Max+ 395, Radeon 8060S, 128 GB), on Vulkan with the RADV driver; Before and after, single stream, greedy, nothing cached; | file rewrite @32k | 37.4 | 25.1 | 48.5 | 55.2 |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28512 (drluoto)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28512",
   "tg_tokens": null,
   "tps": 55.2,
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": null,
   "ctx": 8192,
   "date": "2026-09",
   "flags": [],
   "hardware": "Bosgame M5 (Ryzen AI Max+ 395, Radeon 8060S, 128 GB)",
   "id": "lc-disc-28512-flashnext-pp8k-before",
   "model": "Qwen3.8-Flash-Next",
   "notes": "prefill at 8k before the fixes; config=pp-8k-before",
   "params": null,
   "power_w": null,
   "pp_tps": 340,
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "I run Qwen3.8-Flash-Next as the model server for my agents on a Bosgame M5 (Ryzen AI Max+ 395, Radeon 8060S, 128 GB), on Vulkan with the RADV driver; Prefill went from 340 to 510 tok/s at 8k and 280 to 390 at 32k",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28512 (drluoto)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28512",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": null,
   "ctx": 8192,
   "date": "2026-09",
   "flags": [],
   "hardware": "Bosgame M5 (Ryzen AI Max+ 395, Radeon 8060S, 128 GB)",
   "id": "lc-disc-28512-flashnext-pp8k-now",
   "model": "Qwen3.8-Flash-Next",
   "notes": "prefill at 8k after row-id hoisting (+19% published); config=pp-8k-now",
   "params": null,
   "power_w": null,
   "pp_tps": 510,
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "I run Qwen3.8-Flash-Next as the model server for my agents on a Bosgame M5 (Ryzen AI Max+ 395, Radeon 8060S, 128 GB), on Vulkan with the RADV driver; Prefill went from 340 to 510 tok/s at 8k and 280 to 390 at 32k",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28512 (drluoto)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28512",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": null,
   "ctx": 32768,
   "date": "2026-09",
   "flags": [],
   "hardware": "Bosgame M5 (Ryzen AI Max+ 395, Radeon 8060S, 128 GB)",
   "id": "lc-disc-28512-flashnext-pp32k-before",
   "model": "Qwen3.8-Flash-Next",
   "notes": "prefill at 32k before the fixes; config=pp-32k-before",
   "params": null,
   "power_w": null,
   "pp_tps": 280,
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "I run Qwen3.8-Flash-Next as the model server for my agents on a Bosgame M5 (Ryzen AI Max+ 395, Radeon 8060S, 128 GB), on Vulkan with the RADV driver; Prefill went from 340 to 510 tok/s at 8k and 280 to 390 at 32k",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28512 (drluoto)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28512",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": null,
   "ctx": 32768,
   "date": "2026-09",
   "flags": [],
   "hardware": "Bosgame M5 (Ryzen AI Max+ 395, Radeon 8060S, 128 GB)",
   "id": "lc-disc-28512-flashnext-pp32k-now",
   "model": "Qwen3.8-Flash-Next",
   "notes": "prefill at 32k after the fixes; config=pp-32k-now",
   "params": null,
   "power_w": null,
   "pp_tps": 390,
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "I run Qwen3.8-Flash-Next as the model server for my agents on a Bosgame M5 (Ryzen AI Max+ 395, Radeon 8060S, 128 GB), on Vulkan with the RADV driver; Prefill went from 340 to 510 tok/s at 8k and 280 to 390 at 32k",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28512 (drluoto)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28512",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 353993,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 5060 Ti 16GB",
   "id": "lc-disc-28584-5060ti-qwen36-353k",
   "model": "Qwen3.6-35B-A3B",
   "notes": "353,993-token live coding session, all 99 layers in VRAM (15166/16311 MiB), decode 24.9-25.9 stored at low end, 716 tok/s prefill measured at the end of the 353k chunk; config=353k-live",
   "params": null,
   "power_w": null,
   "pp_tps": 716,
   "provenance": "community",
   "quant": "2.6763bpw",
   "quote": "push the **Qwen 3.6 MoE (35B-A3B)** model far beyond its native limits on a single consumer-grade **RTX 5060 Ti 16GB** GPU; **Model:** Qwen3.6-35B-A3B-12Gb-2.6763bpw.gguf; **Context processed:** 353,993 tokens (real usage inside a live coding session); **Prefill time:** 8m 14s (~716 tok/s at the very end of the 353k chunk); **Decode speed:** 24.9 - 25.9 tok/s (Prompt Caching works perfectly on follow-up turns!)",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28584 (0-KaM-0)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28584",
   "tg_tokens": null,
   "tps": 24.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 16384,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-disc-28514-5090-nvfp4-stock",
   "model": "Qwen3.8-27B",
   "notes": "pp16384 stock llama.cpp (6,122 +- 8); config=stock",
   "params": null,
   "power_w": null,
   "pp_tps": 6122,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "RTX 5090 (sm_120a, 600 W), CUDA 13.3, Qwen3.8-27B NVFP4, `llama-bench -ub 1024 -p 16384 -r 3`, no drafter; | stock llama.cpp | 6,122 \u00b1 8 |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28514 (jesdga95)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28514",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 16384,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-disc-28514-5090-nvfp4-poc",
   "model": "Qwen3.8-27B",
   "notes": "pp16384 with the PoC patch (chunked gated-delta-net prefill, GEMM fusions, TMA-fed NVFP4 MMQ), +44.7%; config=poc",
   "params": null,
   "power_w": null,
   "pp_tps": 8860,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "RTX 5090 (sm_120a, 600 W), CUDA 13.3, Qwen3.8-27B NVFP4, `llama-bench -ub 1024 -p 16384 -r 3`, no drafter; | stock + PoC patch | 8,860 \u00b1 16 (+44.7%) |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28514 (jesdga95)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28514",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "NInfer",
   "batch": null,
   "ctx": 16384,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-disc-28514-5090-nvfp4-ninfer",
   "model": "Qwen3.8-27B",
   "notes": "pp16384 NInfer hand-written W4A4, speculation off; config=ninfer",
   "params": null,
   "power_w": null,
   "pp_tps": 8466,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "RTX 5090 (sm_120a, 600 W), CUDA 13.3, Qwen3.8-27B NVFP4, `llama-bench -ub 1024 -p 16384 -r 3`, no drafter; | NInfer (hand-written W4A4), speculation off | 8,466 |",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28514 (jesdga95)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28514",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 16384,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-disc-28514-5090-nvfp4-poc-nomtp",
   "model": "Qwen3.8-27B",
   "notes": "engine-reported prefill on the PoC build without MTP; config=poc-mtp-off",
   "params": null,
   "power_w": null,
   "pp_tps": 8499,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "RTX 5090 (sm_120a, 600 W), CUDA 13.3, Qwen3.8-27B NVFP4, `llama-bench -ub 1024 -p 16384 -r 3`, no drafter; `--spec-type draft-mtp` cost 23% prefill (8,499 \u2192 6,518 engine-reported) vs NInfer's MTP at \u22129%",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28514 (jesdga95)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28514",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 16384,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-disc-28514-5090-nvfp4-poc-mtp",
   "model": "Qwen3.8-27B",
   "notes": "engine-reported prefill on the PoC build with draft-mtp (-23%, host hand-off per ubatch); config=poc-mtp-on",
   "params": null,
   "power_w": null,
   "pp_tps": 6518,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "RTX 5090 (sm_120a, 600 W), CUDA 13.3, Qwen3.8-27B NVFP4, `llama-bench -ub 1024 -p 16384 -r 3`, no drafter; `--spec-type draft-mtp` cost 23% prefill (8,499 \u2192 6,518 engine-reported) vs NInfer's MTP at \u22129%",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28514 (jesdga95)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28514",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 16384,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-disc-28514-5090-nvfp4-mtp-hosthandoff",
   "model": "Qwen3.8-27B",
   "notes": "MTP prefill with host hand-off, under profiler; config=mtp-host-handoff",
   "params": null,
   "power_w": null,
   "pp_tps": 7065,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "RTX 5090 (sm_120a, 600 W), CUDA 13.3, Qwen3.8-27B NVFP4, `llama-bench -ub 1024 -p 16384 -r 3`, no drafter; Three more commits keep the hand-off on device (device tensor + backend event); idle 15% \u2192 7.7%, 7,065 \u2192 7,550 under the profiler, output is identical",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28514 (jesdga95)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28514",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 16384,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-disc-28514-5090-nvfp4-mtp-device",
   "model": "Qwen3.8-27B",
   "notes": "MTP prefill with device-side hand-off fix, under profiler; config=mtp-device-handoff",
   "params": null,
   "power_w": null,
   "pp_tps": 7550,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "RTX 5090 (sm_120a, 600 W), CUDA 13.3, Qwen3.8-27B NVFP4, `llama-bench -ub 1024 -p 16384 -r 3`, no drafter; Three more commits keep the hand-off on device (device tensor + backend event); idle 15% \u2192 7.7%, 7,065 \u2192 7,550 under the profiler, output is identical",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28514 (jesdga95)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28514",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-disc-28514-5090-nvfp4-tg-stock",
   "model": "Qwen3.8-27B",
   "notes": "tg128 before the patch, not yet A/B tested; config=tg-before",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "RTX 5090 (sm_120a, 600 W), CUDA 13.3, Qwen3.8-27B NVFP4, `llama-bench -ub 1024 -p 16384 -r 3`, no drafter; tg128 82.8 \u2192 78.9, not yet A/B tested",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28514 (jesdga95)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28514",
   "tg_tokens": null,
   "tps": 82.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-disc-28514-5090-nvfp4-tg-poc",
   "model": "Qwen3.8-27B",
   "notes": "tg128 after the patch, not yet A/B tested; config=tg-after",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "RTX 5090 (sm_120a, 600 W), CUDA 13.3, Qwen3.8-27B NVFP4, `llama-bench -ub 1024 -p 16384 -r 3`, no drafter; tg128 82.8 \u2192 78.9, not yet A/B tested",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28514 (jesdga95)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28514",
   "tg_tokens": null,
   "tps": 78.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 131072,
   "date": "2026-09",
   "flags": [],
   "hardware": "Arc B580",
   "id": "lc-disc-29643-bonsai-plain115k",
   "model": "Bonsai 2 27B",
   "notes": "plain decode (no MTP) on the same 114.6K-token prompt as the history row, measured in the follow-up; MTP lifts it back to 43-44 tok/s; config=branch-plain-115k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "PTQ1_0",
   "quote": "I've been working on the SYCL backend for Intel Arc B580 (Xe2), mostly for PrismML's ternary Bonsai 2 27B; I measured plain decode on the same 114.6K-token prompt: 26.4 t/s",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29643 (Torchit1)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29643#discussioncomment-18675258",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "CUDA",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "4500 + 3060",
   "id": "lc-disc-29964-gemma426b-master",
   "model": "Gemma 4 26B-A4B",
   "notes": "1:1 layer split, -ncmoe 999 (all MoE experts on CPU), prompt processing on master: 4500 peaks at RX ~10GiB/s, 3060 mostly idle; config=master-1to1",
   "params": null,
   "power_w": null,
   "pp_tps": 1100,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "gemma-4-26B-A4B-it-UD-Q4_K_XL.gguf; On master the 4500; 3060 is mostly idle, and I get about 1100 t/s pp",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29964 (drrros)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29964#discussioncomment-18765235",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "CUDA",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "4500 + 3060",
   "id": "lc-disc-29964-gemma426b-pr29963",
   "model": "Gemma 4 26B-A4B",
   "notes": "1:1 layer split, -ncmoe 999, on PR #29963 branch: 3060 peaks at RX ~3GiB/s, 4500 mostly idle; disabling pipeline_parallel on the PR recovers full performance, so the penalty is the pipeline itself; 4:1-10:1 splits beat 1:1 but stay ~14% below the 4500-alone run; config=pr29963-1to1",
   "params": null,
   "power_w": null,
   "pp_tps": 600,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "gemma-4-26B-A4B-it-UD-Q4_K_XL.gguf; With https://github.com/ggml-org/llama.cpp/pull/29963 the 3060 peaks at RX of about 3GiB/s, the 4500; 3060 peaks at RX of about 3GiB/s, the 4500 is mostly idle, and I get only 600 t/s pp",
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29964 (drrros)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29964#discussioncomment-18765235",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (xyz-llama fork)",
   "batch": 1,
   "ctx": 163840,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 4070 Ti SUPER",
   "id": "lc29621-4070tis-xyzkv2",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "xyz-llama fork (xyznq1), xyzkv2 2-bit rotated KV cache, ~161k ctx, temp 1.0 top-k 20 top-p 0.95, 10 generations / 11964 tokens, 2.82 tokens per round, lossless block-verified draft (every token an exact sample); config=xyzkv2",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "PTQ1_0",
   "quote": "On an RTX 4070 Ti SUPER (16 GB, stock power limit) at around 161k context; we measured 139.5 tokens/s and 2.82 tokens per round, and 144.2 tokens/s with our optional CUDA runtime; PrismML's Ternary Bonsai 2 27B",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29621 (xyznq1)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29621",
   "tg_tokens": 11964,
   "tps": 139.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (xyz-llama fork)",
   "batch": 1,
   "ctx": 163840,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 4070 Ti SUPER",
   "id": "lc29621-4070tis-xyzengine",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "same run with XYZ_ENGINE=1 (optional CUDA runtime for the speculative rounds, Windows), same text; config=xyz-engine",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "PTQ1_0",
   "quote": "On an RTX 4070 Ti SUPER (16 GB, stock power limit) at around 161k context; we measured 139.5 tokens/s and 2.82 tokens per round, and 144.2 tokens/s with our optional CUDA runtime; PrismML's Ternary Bonsai 2 27B",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29621 (xyznq1)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29621",
   "tg_tokens": 11964,
   "tps": 144.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (xyz-llama fork)",
   "batch": 1,
   "ctx": 163840,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 4070 Ti SUPER",
   "id": "lc29621-4070tis-xyzkv2-r2",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "re-measured later, same card/build/prompts, 2.94 tokens per round, KV 7.5 GiB; config=xyzkv2-r2",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "PTQ1_0",
   "quote": "Measured on the same card at 161k, same build and prompts: xyzkv2 156.3 tokens/s and 7.5 GiB, q4_0 141.3 tokens/s and 9.0 GiB; On an RTX 4070 Ti SUPER (16 GB, stock power limit); PrismML's Ternary Bonsai 2 27B",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29621 (xyznq1)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29621",
   "tg_tokens": null,
   "tps": 156.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (xyz-llama fork)",
   "batch": 1,
   "ctx": 163840,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 4070 Ti SUPER",
   "id": "lc29621-4070tis-q40-r2",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "q4_0 KV cache control, same build and prompts, 2.94 tokens per round, KV 9.0 GiB; about 11% slower and 1.5 GiB larger than xyzkv2, gap is all cache reads; config=kv-q40-r2",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "PTQ1_0",
   "quote": "Measured on the same card at 161k, same build and prompts: xyzkv2 156.3 tokens/s and 7.5 GiB, q4_0 141.3 tokens/s and 9.0 GiB; On an RTX 4070 Ti SUPER (16 GB, stock power limit); PrismML's Ternary Bonsai 2 27B",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #29621 (xyznq1)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29621",
   "tg_tokens": null,
   "tps": 141.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (JigSawPT DSV4.1 port)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc28766-5090-dsv41-new",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "DSV4.1 port on b10269 (JigSawPT/llama.cpp, branch dsv41-porte, expert streaming is nibor1896's Crow patch series), routed experts streamed from NVMe through a VRAM cache and pinned host tier, 189 GiB engram tables on disk read by the host; decode on new content is disk-bound (20% compute / 26% PCIe / 54% NVMe per remap, ~3.8 requests in flight per layer); config=dsv41-stream-new",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4 (experts)",
   "quote": "I ported DeepSeek-V4.1-Flash to a llama.cpp fork and ran it on a single RTX 5090 (31.8 GiB of VRAM) with 125.7 GiB of RAM; at 5.1 tokens/s on new content and 21 tokens/s on content already resident in the cache",
   "ram_gb": 125.7,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28766 (JigSawPT)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28766",
   "tg_tokens": null,
   "tps": 5.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (JigSawPT DSV4.1 port)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc28766-5090-dsv41-cached",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "same setup, content already resident in the expert cache (105 GiB working set vs 31.8 GiB VRAM); config=dsv41-cached",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4 (experts)",
   "quote": "I ported DeepSeek-V4.1-Flash to a llama.cpp fork and ran it on a single RTX 5090 (31.8 GiB of VRAM) with 125.7 GiB of RAM; at 5.1 tokens/s on new content and 21 tokens/s on content already resident in the cache",
   "ram_gb": 125.7,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28766 (JigSawPT)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28766",
   "tg_tokens": null,
   "tps": 21,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (JigSawPT DSV4.1 port)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc28766-5090-dsv41-ceil-nomiss",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "measured ceiling with no disk miss at all (oracle routing, no disk traffic); config=ceil-nomiss",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4 (experts)",
   "quote": "Ceilings, measured: 6.2 tokens/s with no disk miss at all, 21.3 with everything resident; I ported DeepSeek-V4.1-Flash to a llama.cpp fork and ran it on a single RTX 5090 (31.8 GiB of VRAM)",
   "ram_gb": 125.7,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28766 (JigSawPT)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28766",
   "tg_tokens": null,
   "tps": 6.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (JigSawPT DSV4.1 port)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc28766-5090-dsv41-ceil-resident",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "measured ceiling with everything resident, which a 105 GiB working set does not allow on 31.8 GiB VRAM; config=ceil-resident",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4 (experts)",
   "quote": "Ceilings, measured: 6.2 tokens/s with no disk miss at all, 21.3 with everything resident; I ported DeepSeek-V4.1-Flash to a llama.cpp fork and ran it on a single RTX 5090 (31.8 GiB of VRAM)",
   "ram_gb": 125.7,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp discussion #28766 (JigSawPT)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/28766",
   "tg_tokens": null,
   "tps": 21.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 204800,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "Threadripper PRO 3975WX + 2x RTX 5090",
   "id": "lc24528-3975wx-2x5090-glm53flash-master-nocache",
   "model": "GLM-5.3-Flash",
   "notes": "Short-prompt code decode, ctx 204800; MoE expert pool on CPU (8x DDR4-3200, ~94 GB/s), attention on 2x RTX 5090 PCIe 4; downstream moe-cache-v2-pr (e3096b046) + GLM support from #27754; master build, no cache; A/B arm 1/4 from c18786889 (AI-drafted comment, author checked numbers vs raw data); per-token ms profile table parked; config=master-nocache",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ3_XXS",
   "quote": "GLM-5.3-Flash UD-IQ3_XXS, ctx 204800; Threadripper PRO 3975WX; 2x RTX 5090 on PCIe 4; master, experts on CPU, no cache | 18.5",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528#discussioncomment-18786889",
   "tg_tokens": null,
   "tps": "18.5",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 204800,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "Threadripper PRO 3975WX + 2x RTX 5090",
   "id": "lc24528-3975wx-2x5090-glm53flash-master-copycache",
   "model": "GLM-5.3-Flash",
   "notes": "Short-prompt code decode, ctx 204800; MoE expert pool on CPU (8x DDR4-3200, ~94 GB/s), attention on 2x RTX 5090 PCIe 4; downstream moe-cache-v2-pr (e3096b046) + GLM support from #27754; master + #29887 copy cache (one GPU, 10.7 GiB cache, 47% hits); slower than no cache: ~1.5 GiB uploaded per token, PCIe held at ~25 GB/s; A/B arm 2/4 from c18786889; config=master-copycache-29887",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ3_XXS",
   "quote": "GLM-5.3-Flash UD-IQ3_XXS, ctx 204800; Threadripper PRO 3975WX; 2x RTX 5090 on PCIe 4; master + #29887 copy cache (one GPU, 10.7 GiB cache, 47% hits) | 12.9",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528#discussioncomment-18786889",
   "tg_tokens": null,
   "tps": "12.9",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 204800,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "Threadripper PRO 3975WX + 2x RTX 5090",
   "id": "lc24528-3975wx-2x5090-glm53flash-v2-cpumiss",
   "model": "GLM-5.3-Flash",
   "notes": "Short-prompt code decode, ctx 204800; MoE expert pool on CPU (8x DDR4-3200, ~94 GB/s), attention on 2x RTX 5090 PCIe 4; downstream moe-cache-v2-pr (e3096b046) + GLM support from #27754; v2 CPU-on-miss cache: miss experts computed on host, no PCIe traffic; ~18-21 GiB expert pool per card, 61-69% decode hit rate, speculation off; A/B arm 3/4 from c18786889; config=v2-cpumiss",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ3_XXS",
   "quote": "GLM-5.3-Flash UD-IQ3_XXS, ctx 204800; Threadripper PRO 3975WX; 2x RTX 5090 on PCIe 4; v2 CPU-on-miss cache | 23.7",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528#discussioncomment-18786889",
   "tg_tokens": null,
   "tps": "23.7",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 204800,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "Threadripper PRO 3975WX + 2x RTX 5090",
   "id": "lc24528-3975wx-2x5090-glm53flash-v2-dflash2",
   "model": "GLM-5.3-Flash",
   "notes": "Short-prompt code decode, ctx 204800; MoE expert pool on CPU (8x DDR4-3200, ~94 GB/s), attention on 2x RTX 5090 PCIe 4; downstream moe-cache-v2-pr (e3096b046) + GLM support from #27754; v2 CPU-on-miss cache + DFlash2 n=2, ub 1024, best config (3 boots); A/B arm 4/4 from c18786889; config=v2-dflash2",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ3_XXS",
   "quote": "GLM-5.3-Flash UD-IQ3_XXS, ctx 204800; Threadripper PRO 3975WX; 2x RTX 5090 on PCIe 4; v2 CPU-on-miss cache + DFlash2 n=2, ub 1024 (best config) | 33.5 (3 boots)",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "llama.cpp discussion #24528",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/24528#discussioncomment-18786889",
   "tg_tokens": null,
   "tps": "33.5",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 51200,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "6x RTX 3090",
   "id": "lc30071-6xrtx3090-pf50k1s-pp-up",
   "model": "Qwen3.8-Flash-Next",
   "notes": "lukolszewski/llama.cpp-multigpu fork vs upstream df03399b8, same GGUF and command line, q8_0 KV, layer split over six GPUs, consumer PCIe (4 via USB4 eGPU docks), 5 slots x 262k; prefill, 1 session, 50k; config=upstream-df03399b8; fork arm: 2127 tok/s",
   "params": null,
   "power_w": null,
   "pp_tokens": 51200,
   "pp_tps": "568",
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Qwen3.8-Flash-Next; unsloth UD-Q4_K_XL; 6 x RTX 3090; prefill, 1 session | 50k | 568 -> 2127 (3.7x)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "llama.cpp discussion #30071",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/30071",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (multigpu fork)",
   "batch": 1,
   "ctx": 51200,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "6x RTX 3090",
   "id": "lc30071-6xrtx3090-pf50k1s-pp-fork",
   "model": "Qwen3.8-Flash-Next",
   "notes": "lukolszewski/llama.cpp-multigpu fork vs upstream df03399b8, same GGUF and command line, q8_0 KV, layer split over six GPUs, consumer PCIe (4 via USB4 eGPU docks), 5 slots x 262k; prefill, 1 session, 50k; config=fork-multigpu; upstream arm: 568 tok/s",
   "params": null,
   "power_w": null,
   "pp_tokens": 51200,
   "pp_tps": "2127",
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Qwen3.8-Flash-Next; unsloth UD-Q4_K_XL; 6 x RTX 3090; prefill, 1 session | 50k | 568 -> 2127 (3.7x)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "llama.cpp discussion #30071",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/30071",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 51200,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "6x RTX 4090",
   "id": "lc30071-6xrtx4090-pf50k1s-pp-up",
   "model": "Qwen3.8-Flash-Next",
   "notes": "lukolszewski/llama.cpp-multigpu fork vs upstream df03399b8, same GGUF and command line, q8_0 KV, layer split over six GPUs, consumer PCIe (4 via USB4 eGPU docks), 5 slots x 262k; prefill, 1 session, 50k; config=upstream-df03399b8; fork arm: 7050 tok/s",
   "params": null,
   "power_w": null,
   "pp_tokens": 51200,
   "pp_tps": "1728",
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Qwen3.8-Flash-Next; unsloth UD-Q4_K_XL; 6 x RTX 4090; 1728 -> 7050 (4.1x)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "llama.cpp discussion #30071",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/30071",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (multigpu fork)",
   "batch": 1,
   "ctx": 51200,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "6x RTX 4090",
   "id": "lc30071-6xrtx4090-pf50k1s-pp-fork",
   "model": "Qwen3.8-Flash-Next",
   "notes": "lukolszewski/llama.cpp-multigpu fork vs upstream df03399b8, same GGUF and command line, q8_0 KV, layer split over six GPUs, consumer PCIe (4 via USB4 eGPU docks), 5 slots x 262k; prefill, 1 session, 50k; config=fork-multigpu; upstream arm: 1728 tok/s",
   "params": null,
   "power_w": null,
   "pp_tokens": 51200,
   "pp_tps": "7050",
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Qwen3.8-Flash-Next; unsloth UD-Q4_K_XL; 6 x RTX 4090; 1728 -> 7050 (4.1x)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "llama.cpp discussion #30071",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/30071",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 256000,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "6x RTX 3090",
   "id": "lc30071-6xrtx3090-pf250k1s-pp-up",
   "model": "Qwen3.8-Flash-Next",
   "notes": "lukolszewski/llama.cpp-multigpu fork vs upstream df03399b8, same GGUF and command line, q8_0 KV, layer split over six GPUs, consumer PCIe (4 via USB4 eGPU docks), 5 slots x 262k; prefill, 1 session, 250k; config=upstream-df03399b8; fork arm: 2111 tok/s",
   "params": null,
   "power_w": null,
   "pp_tokens": 256000,
   "pp_tps": "263",
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Qwen3.8-Flash-Next; unsloth UD-Q4_K_XL; 6 x RTX 3090; prefill, 1 session | 250k | 263 -> 2111 (8.0x)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "llama.cpp discussion #30071",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/30071",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (multigpu fork)",
   "batch": 1,
   "ctx": 256000,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "6x RTX 3090",
   "id": "lc30071-6xrtx3090-pf250k1s-pp-fork",
   "model": "Qwen3.8-Flash-Next",
   "notes": "lukolszewski/llama.cpp-multigpu fork vs upstream df03399b8, same GGUF and command line, q8_0 KV, layer split over six GPUs, consumer PCIe (4 via USB4 eGPU docks), 5 slots x 262k; prefill, 1 session, 250k; config=fork-multigpu; upstream arm: 263 tok/s",
   "params": null,
   "power_w": null,
   "pp_tokens": 256000,
   "pp_tps": "2111",
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Qwen3.8-Flash-Next; unsloth UD-Q4_K_XL; 6 x RTX 3090; prefill, 1 session | 250k | 263 -> 2111 (8.0x)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "llama.cpp discussion #30071",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/30071",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 256000,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "6x RTX 4090",
   "id": "lc30071-6xrtx4090-pf250k1s-pp-up",
   "model": "Qwen3.8-Flash-Next",
   "notes": "lukolszewski/llama.cpp-multigpu fork vs upstream df03399b8, same GGUF and command line, q8_0 KV, layer split over six GPUs, consumer PCIe (4 via USB4 eGPU docks), 5 slots x 262k; prefill, 1 session, 250k; config=upstream-df03399b8; fork arm: 7403 tok/s",
   "params": null,
   "power_w": null,
   "pp_tokens": 256000,
   "pp_tps": "744",
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Qwen3.8-Flash-Next; unsloth UD-Q4_K_XL; 6 x RTX 4090; 744 -> 7403 (10.0x)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "llama.cpp discussion #30071",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/30071",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (multigpu fork)",
   "batch": 1,
   "ctx": 256000,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "6x RTX 4090",
   "id": "lc30071-6xrtx4090-pf250k1s-pp-fork",
   "model": "Qwen3.8-Flash-Next",
   "notes": "lukolszewski/llama.cpp-multigpu fork vs upstream df03399b8, same GGUF and command line, q8_0 KV, layer split over six GPUs, consumer PCIe (4 via USB4 eGPU docks), 5 slots x 262k; prefill, 1 session, 250k; config=fork-multigpu; upstream arm: 744 tok/s",
   "params": null,
   "power_w": null,
   "pp_tokens": 256000,
   "pp_tps": "7403",
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Qwen3.8-Flash-Next; unsloth UD-Q4_K_XL; 6 x RTX 4090; 744 -> 7403 (10.0x)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "llama.cpp discussion #30071",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/30071",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 5,
   "ctx": 256000,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "6x RTX 3090",
   "id": "lc30071-6xrtx3090-pf250k5s-pp-up",
   "model": "Qwen3.8-Flash-Next",
   "notes": "lukolszewski/llama.cpp-multigpu fork vs upstream df03399b8, same GGUF and command line, q8_0 KV, layer split over six GPUs, consumer PCIe (4 via USB4 eGPU docks), 5 slots x 262k; prefill, 5 sessions aggregate, 250k; config=upstream-df03399b8; fork arm: 2060 tok/s",
   "params": null,
   "power_w": null,
   "pp_tokens": 256000,
   "pp_tps": "265",
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Qwen3.8-Flash-Next; unsloth UD-Q4_K_XL; 6 x RTX 3090; prefill, 5 sessions, aggregate | 250k | 265 -> 2060 (7.8x)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "llama.cpp discussion #30071",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/30071",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (multigpu fork)",
   "batch": 5,
   "ctx": 256000,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "6x RTX 3090",
   "id": "lc30071-6xrtx3090-pf250k5s-pp-fork",
   "model": "Qwen3.8-Flash-Next",
   "notes": "lukolszewski/llama.cpp-multigpu fork vs upstream df03399b8, same GGUF and command line, q8_0 KV, layer split over six GPUs, consumer PCIe (4 via USB4 eGPU docks), 5 slots x 262k; prefill, 5 sessions aggregate, 250k; config=fork-multigpu; upstream arm: 265 tok/s",
   "params": null,
   "power_w": null,
   "pp_tokens": 256000,
   "pp_tps": "2060",
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Qwen3.8-Flash-Next; unsloth UD-Q4_K_XL; 6 x RTX 3090; prefill, 5 sessions, aggregate | 250k | 265 -> 2060 (7.8x)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "llama.cpp discussion #30071",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/30071",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 51200,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "6x RTX 3090",
   "id": "lc30071-6xrtx3090-gen50k1s-tg-up",
   "model": "Qwen3.8-Flash-Next",
   "notes": "lukolszewski/llama.cpp-multigpu fork vs upstream df03399b8, same GGUF and command line, q8_0 KV, layer split over six GPUs, consumer PCIe (4 via USB4 eGPU docks), 5 slots x 262k; generation, 1 session, 50k; config=upstream-df03399b8; fork arm: 41.9 tok/s",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Qwen3.8-Flash-Next; unsloth UD-Q4_K_XL; 6 x RTX 3090; generation, 1 session | 50k | 25.7 -> 41.9 (1.6x)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "llama.cpp discussion #30071",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/30071",
   "tg_tokens": null,
   "tps": "25.7",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (multigpu fork)",
   "batch": 1,
   "ctx": 51200,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "6x RTX 3090",
   "id": "lc30071-6xrtx3090-gen50k1s-tg-fork",
   "model": "Qwen3.8-Flash-Next",
   "notes": "lukolszewski/llama.cpp-multigpu fork vs upstream df03399b8, same GGUF and command line, q8_0 KV, layer split over six GPUs, consumer PCIe (4 via USB4 eGPU docks), 5 slots x 262k; generation, 1 session, 50k; config=fork-multigpu; upstream arm: 25.7 tok/s",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Qwen3.8-Flash-Next; unsloth UD-Q4_K_XL; 6 x RTX 3090; generation, 1 session | 50k | 25.7 -> 41.9 (1.6x)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "llama.cpp discussion #30071",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/30071",
   "tg_tokens": null,
   "tps": "41.9",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 51200,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "6x RTX 4090",
   "id": "lc30071-6xrtx4090-gen50k1s-tg-up",
   "model": "Qwen3.8-Flash-Next",
   "notes": "lukolszewski/llama.cpp-multigpu fork vs upstream df03399b8, same GGUF and command line, q8_0 KV, layer split over six GPUs, consumer PCIe (4 via USB4 eGPU docks), 5 slots x 262k; generation, 1 session, 50k; config=upstream-df03399b8; fork arm: 57.5 tok/s",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Qwen3.8-Flash-Next; unsloth UD-Q4_K_XL; 6 x RTX 4090; 46.2 -> 57.5 (1.2x)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "llama.cpp discussion #30071",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/30071",
   "tg_tokens": null,
   "tps": "46.2",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (multigpu fork)",
   "batch": 1,
   "ctx": 51200,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "6x RTX 4090",
   "id": "lc30071-6xrtx4090-gen50k1s-tg-fork",
   "model": "Qwen3.8-Flash-Next",
   "notes": "lukolszewski/llama.cpp-multigpu fork vs upstream df03399b8, same GGUF and command line, q8_0 KV, layer split over six GPUs, consumer PCIe (4 via USB4 eGPU docks), 5 slots x 262k; generation, 1 session, 50k; config=fork-multigpu; upstream arm: 46.2 tok/s",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Qwen3.8-Flash-Next; unsloth UD-Q4_K_XL; 6 x RTX 4090; 46.2 -> 57.5 (1.2x)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "llama.cpp discussion #30071",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/30071",
   "tg_tokens": null,
   "tps": "57.5",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 256000,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "6x RTX 3090",
   "id": "lc30071-6xrtx3090-gen250k1s-tg-up",
   "model": "Qwen3.8-Flash-Next",
   "notes": "lukolszewski/llama.cpp-multigpu fork vs upstream df03399b8, same GGUF and command line, q8_0 KV, layer split over six GPUs, consumer PCIe (4 via USB4 eGPU docks), 5 slots x 262k; generation, 1 session, 250k; config=upstream-df03399b8; fork arm: 33.7 tok/s",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Qwen3.8-Flash-Next; unsloth UD-Q4_K_XL; 6 x RTX 3090; generation, 1 session | 250k | 10.2 -> 33.7 (3.3x)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "llama.cpp discussion #30071",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/30071",
   "tg_tokens": null,
   "tps": "10.2",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (multigpu fork)",
   "batch": 1,
   "ctx": 256000,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "6x RTX 3090",
   "id": "lc30071-6xrtx3090-gen250k1s-tg-fork",
   "model": "Qwen3.8-Flash-Next",
   "notes": "lukolszewski/llama.cpp-multigpu fork vs upstream df03399b8, same GGUF and command line, q8_0 KV, layer split over six GPUs, consumer PCIe (4 via USB4 eGPU docks), 5 slots x 262k; generation, 1 session, 250k; config=fork-multigpu; upstream arm: 10.2 tok/s",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Qwen3.8-Flash-Next; unsloth UD-Q4_K_XL; 6 x RTX 3090; generation, 1 session | 250k | 10.2 -> 33.7 (3.3x)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "llama.cpp discussion #30071",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/30071",
   "tg_tokens": null,
   "tps": "33.7",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 256000,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "6x RTX 4090",
   "id": "lc30071-6xrtx4090-gen250k1s-tg-up",
   "model": "Qwen3.8-Flash-Next",
   "notes": "lukolszewski/llama.cpp-multigpu fork vs upstream df03399b8, same GGUF and command line, q8_0 KV, layer split over six GPUs, consumer PCIe (4 via USB4 eGPU docks), 5 slots x 262k; generation, 1 session, 250k; config=upstream-df03399b8; fork arm: 48.7 tok/s",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Qwen3.8-Flash-Next; unsloth UD-Q4_K_XL; 6 x RTX 4090; 21.0 -> 48.7 (2.3x)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "llama.cpp discussion #30071",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/30071",
   "tg_tokens": null,
   "tps": "21.0",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (multigpu fork)",
   "batch": 1,
   "ctx": 256000,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "6x RTX 4090",
   "id": "lc30071-6xrtx4090-gen250k1s-tg-fork",
   "model": "Qwen3.8-Flash-Next",
   "notes": "lukolszewski/llama.cpp-multigpu fork vs upstream df03399b8, same GGUF and command line, q8_0 KV, layer split over six GPUs, consumer PCIe (4 via USB4 eGPU docks), 5 slots x 262k; generation, 1 session, 250k; config=fork-multigpu; upstream arm: 21.0 tok/s",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Qwen3.8-Flash-Next; unsloth UD-Q4_K_XL; 6 x RTX 4090; 21.0 -> 48.7 (2.3x)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "llama.cpp discussion #30071",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/30071",
   "tg_tokens": null,
   "tps": "48.7",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 5,
   "ctx": 256000,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "6x RTX 3090",
   "id": "lc30071-6xrtx3090-gen250k5s-tg-up",
   "model": "Qwen3.8-Flash-Next",
   "notes": "lukolszewski/llama.cpp-multigpu fork vs upstream df03399b8, same GGUF and command line, q8_0 KV, layer split over six GPUs, consumer PCIe (4 via USB4 eGPU docks), 5 slots x 262k; generation, 5 sessions per session, 250k; config=upstream-df03399b8; fork arm: 27.3 tok/s",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Qwen3.8-Flash-Next; unsloth UD-Q4_K_XL; 6 x RTX 3090; generation, 5 sessions, per session | 250k | 2.3 -> 27.3 (10.8x)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "llama.cpp discussion #30071",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/30071",
   "tg_tokens": null,
   "tps": "2.3",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (multigpu fork)",
   "batch": 5,
   "ctx": 256000,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "6x RTX 3090",
   "id": "lc30071-6xrtx3090-gen250k5s-tg-fork",
   "model": "Qwen3.8-Flash-Next",
   "notes": "lukolszewski/llama.cpp-multigpu fork vs upstream df03399b8, same GGUF and command line, q8_0 KV, layer split over six GPUs, consumer PCIe (4 via USB4 eGPU docks), 5 slots x 262k; generation, 5 sessions per session, 250k; config=fork-multigpu; upstream arm: 2.3 tok/s",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Qwen3.8-Flash-Next; unsloth UD-Q4_K_XL; 6 x RTX 3090; generation, 5 sessions, per session | 250k | 2.3 -> 27.3 (10.8x)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "llama.cpp discussion #30071",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/30071",
   "tg_tokens": null,
   "tps": "27.3",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (multigpu fork)",
   "batch": 5,
   "ctx": 256000,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "6x RTX 4090",
   "id": "lc30071-6xrtx4090-pf250k5s-pp-fork",
   "model": "Qwen3.8-Flash-Next",
   "notes": "lukolszewski/llama.cpp-multigpu fork vs upstream df03399b8, same GGUF and command line, q8_0 KV, layer split over six GPUs, consumer PCIe (4 via USB4 eGPU docks), 5 slots x 262k; prefill, 5 sessions aggregate, 250k; config=fork-multigpu; upstream not run",
   "params": null,
   "power_w": null,
   "pp_tokens": 256000,
   "pp_tps": "7241",
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Qwen3.8-Flash-Next; unsloth UD-Q4_K_XL; 6 x RTX 4090; not run -> 7241",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "llama.cpp discussion #30071",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/30071",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (multigpu fork)",
   "batch": 5,
   "ctx": 256000,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "6x RTX 4090",
   "id": "lc30071-6xrtx4090-gen250k5s-tg-fork",
   "model": "Qwen3.8-Flash-Next",
   "notes": "lukolszewski/llama.cpp-multigpu fork vs upstream df03399b8, same GGUF and command line, q8_0 KV, layer split over six GPUs, consumer PCIe (4 via USB4 eGPU docks), 5 slots x 262k; generation, 5 sessions per session, 250k; config=fork-multigpu; upstream not run",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Qwen3.8-Flash-Next; unsloth UD-Q4_K_XL; 6 x RTX 4090; not run -> 30.8",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "llama.cpp discussion #30071",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/30071",
   "tg_tokens": null,
   "tps": "30.8",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "Raspberry Pi 5",
   "id": "lc30069-rpi5-qwen2505b-q40-t1-pp",
   "model": "Qwen2.5-0.5B",
   "notes": "llama-bench -p 512 -n 128 -r 5, KleidiAI off, build 836d571, CPU backend; BCM2712 4x Cortex-A76 @ 2.4 GHz, 2 GB LPDDR4X; table label qwen2 1B Q4_0 (llama-bench misdetect), GGUF qwen2.5-0.5b-instruct-q4_0; config=t1",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "93.49",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Qwen2.5-0.5B Q4_0; Raspberry Pi 5; qwen2 1B Q4_0 | 403.20 MiB | 630.17 M | CPU | 1 | pp512 | 93.49",
   "ram_gb": 2,
   "retrieved": "2026-10-08",
   "source_name": "llama.cpp discussion #30069",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/30069",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "Raspberry Pi 5",
   "id": "lc30069-rpi5-qwen2505b-q40-t1-tg",
   "model": "Qwen2.5-0.5B",
   "notes": "llama-bench -p 512 -n 128 -r 5, KleidiAI off, build 836d571, CPU backend; BCM2712 4x Cortex-A76 @ 2.4 GHz, 2 GB LPDDR4X; table label qwen2 1B Q4_0 (llama-bench misdetect), GGUF qwen2.5-0.5b-instruct-q4_0; config=t1",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Qwen2.5-0.5B Q4_0; Raspberry Pi 5; qwen2 1B Q4_0 | 403.20 MiB | 630.17 M | CPU | 1 | tg128 | 32.42",
   "ram_gb": 2,
   "retrieved": "2026-10-08",
   "source_name": "llama.cpp discussion #30069",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/30069",
   "tg_tokens": 128,
   "tps": "32.42",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "Raspberry Pi 5",
   "id": "lc30069-rpi5-qwen2505b-q40-t2-pp",
   "model": "Qwen2.5-0.5B",
   "notes": "llama-bench -p 512 -n 128 -r 5, KleidiAI off, build 836d571, CPU backend; BCM2712 4x Cortex-A76 @ 2.4 GHz, 2 GB LPDDR4X; table label qwen2 1B Q4_0 (llama-bench misdetect), GGUF qwen2.5-0.5b-instruct-q4_0; config=t2",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "173.68",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Qwen2.5-0.5B Q4_0; Raspberry Pi 5; qwen2 1B Q4_0 | 403.20 MiB | 630.17 M | CPU | 2 | pp512 | 173.68",
   "ram_gb": 2,
   "retrieved": "2026-10-08",
   "source_name": "llama.cpp discussion #30069",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/30069",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "Raspberry Pi 5",
   "id": "lc30069-rpi5-qwen2505b-q40-t2-tg",
   "model": "Qwen2.5-0.5B",
   "notes": "llama-bench -p 512 -n 128 -r 5, KleidiAI off, build 836d571, CPU backend; BCM2712 4x Cortex-A76 @ 2.4 GHz, 2 GB LPDDR4X; table label qwen2 1B Q4_0 (llama-bench misdetect), GGUF qwen2.5-0.5b-instruct-q4_0; config=t2",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Qwen2.5-0.5B Q4_0; Raspberry Pi 5; qwen2 1B Q4_0 | 403.20 MiB | 630.17 M | CPU | 2 | tg128 | 33.84",
   "ram_gb": 2,
   "retrieved": "2026-10-08",
   "source_name": "llama.cpp discussion #30069",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/30069",
   "tg_tokens": 128,
   "tps": "33.84",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "Raspberry Pi 5",
   "id": "lc30069-rpi5-qwen2505b-q40-t3-pp",
   "model": "Qwen2.5-0.5B",
   "notes": "llama-bench -p 512 -n 128 -r 5, KleidiAI off, build 836d571, CPU backend; BCM2712 4x Cortex-A76 @ 2.4 GHz, 2 GB LPDDR4X; table label qwen2 1B Q4_0 (llama-bench misdetect), GGUF qwen2.5-0.5b-instruct-q4_0; config=t3",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "233.22",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Qwen2.5-0.5B Q4_0; Raspberry Pi 5; qwen2 1B Q4_0 | 403.20 MiB | 630.17 M | CPU | 3 | pp512 | 233.22",
   "ram_gb": 2,
   "retrieved": "2026-10-08",
   "source_name": "llama.cpp discussion #30069",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/30069",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "Raspberry Pi 5",
   "id": "lc30069-rpi5-qwen2505b-q40-t3-tg",
   "model": "Qwen2.5-0.5B",
   "notes": "llama-bench -p 512 -n 128 -r 5, KleidiAI off, build 836d571, CPU backend; BCM2712 4x Cortex-A76 @ 2.4 GHz, 2 GB LPDDR4X; table label qwen2 1B Q4_0 (llama-bench misdetect), GGUF qwen2.5-0.5b-instruct-q4_0; config=t3",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Qwen2.5-0.5B Q4_0; Raspberry Pi 5; qwen2 1B Q4_0 | 403.20 MiB | 630.17 M | CPU | 3 | tg128 | 29.37",
   "ram_gb": 2,
   "retrieved": "2026-10-08",
   "source_name": "llama.cpp discussion #30069",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/30069",
   "tg_tokens": 128,
   "tps": "29.37",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "Raspberry Pi 5",
   "id": "lc30069-rpi5-qwen2505b-q40-t4-pp",
   "model": "Qwen2.5-0.5B",
   "notes": "llama-bench -p 512 -n 128 -r 5, KleidiAI off, build 836d571, CPU backend; BCM2712 4x Cortex-A76 @ 2.4 GHz, 2 GB LPDDR4X; table label qwen2 1B Q4_0 (llama-bench misdetect), GGUF qwen2.5-0.5b-instruct-q4_0; config=t4",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "289.92",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Qwen2.5-0.5B Q4_0; Raspberry Pi 5; qwen2 1B Q4_0 | 403.20 MiB | 630.17 M | CPU | 4 | pp512 | 289.92",
   "ram_gb": 2,
   "retrieved": "2026-10-08",
   "source_name": "llama.cpp discussion #30069",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/30069",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "Raspberry Pi 5",
   "id": "lc30069-rpi5-qwen2505b-q40-t4-tg",
   "model": "Qwen2.5-0.5B",
   "notes": "llama-bench -p 512 -n 128 -r 5, KleidiAI off, build 836d571, CPU backend; BCM2712 4x Cortex-A76 @ 2.4 GHz, 2 GB LPDDR4X; table label qwen2 1B Q4_0 (llama-bench misdetect), GGUF qwen2.5-0.5b-instruct-q4_0; config=t4",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Qwen2.5-0.5B Q4_0; Raspberry Pi 5; qwen2 1B Q4_0 | 403.20 MiB | 630.17 M | CPU | 4 | tg128 | 27.28",
   "ram_gb": 2,
   "retrieved": "2026-10-08",
   "source_name": "llama.cpp discussion #30069",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/30069",
   "tg_tokens": 128,
   "tps": "27.28",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 1024,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "githubissues-rtx5090-qwen35-27b-q4-k-m",
   "model": "Qwen3.5-27B",
   "notes": "llama-bench -ngl 99 -fa 1 -p 1024 -n 400, 5 reps; llama.cpp b10741 Windows binaries, Windows 11; reporter notes 76% of roofline on native Linux build, 28% on this Windows rig",
   "params": "27B",
   "power_w": null,
   "pp_tps": 1515.83,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "GPU: NVIDIA GeForce RTX 5090 (32 GB GDDR7, driver 616.56); Model: Qwen3.5-27B Q4_K_M GGUF, 15.65 GiB; qwen35 27B Q4_K_M | pp1024 | 1515.83 \u00b1 47.03; qwen35 27B Q4_K_M | tg400 | 30.09 \u00b1 0.44",
   "ram_gb": 32,
   "retrieved": "2026-09-26",
   "source_name": "llama.cpp issue #28196 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28196",
   "tps": 30.09,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 1024,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "githubissues-rtx5090-qwen3moe-30b-a3b-q4-k-m",
   "model": "qwen3moe 30B.A3B",
   "notes": "MoE control model on the same rig and build, per issue; same llama-bench command and reps",
   "params": "30B-A3B",
   "power_w": null,
   "pp_tps": 4081.2,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "GPU: NVIDIA GeForce RTX 5090 (32 GB GDDR7, driver 616.56); qwen3moe 30B.A3B Q4_K_M (control, same rig) | pp1024 | 4081.20 \u00b1 89.86; qwen3moe 30B.A3B Q4_K_M (control, same rig) | tg400 | 134.18 \u00b1 1.84",
   "ram_gb": 32,
   "retrieved": "2026-09-26",
   "source_name": "llama.cpp issue #28196 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28196",
   "tps": 134.18,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-05",
   "flags": [],
   "hardware": "GB10 (DGX Spark)",
   "id": "githubissues-dgxspark-gb10-qwen36-35b-a3b-q4-k-m",
   "model": "Qwen3.6-35B-A3B",
   "notes": "warm decode tps; unsloth UD-Q4_K_M build; llama.cpp server-cuda-b9014 container on DGX Spark (GB10, sm_120, aarch64)",
   "params": "35B-A3B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "NVIDIA GB10; DGX Spark (GB10 Grace Blackwell); I served Qwen3.6-35B-A3B-UD-Q4_K_M.gguf (unsloth, SHA ac0e2c1189e055faa36eff361580e79c5bd6f8e76bffb4ce547f167d53e31a61) on the same image, same GPU, same flags; Coherent output, 7 tok/s cold / 59 tok/s warm decode across multiple prompts",
   "ram_gb": null,
   "retrieved": "2026-09-26",
   "source_name": "llama.cpp issue #23010 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/23010",
   "tps": 59,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-05",
   "flags": [],
   "hardware": "GB10 (DGX Spark)",
   "id": "githubissues-dgxspark-gb10-qwen3-coder-next-q4-k-m",
   "model": "Qwen3-Coder-Next",
   "notes": "decode throughput reported by issue reporter; server configured with --ctx-size 131072 but measured with a minimal prompt; model output was corrupted (all-? tokens) though throughput was normal",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "NVIDIA GB10; DGX Spark (GB10 Grace Blackwell, compute capability 12.1, aarch64), llama-server from the ghcr.io/ggml-org/llama.cpp:server-cuda-b9014 image loads qwen3-coder-next-Q4_K_M.gguf successfully, accepts chat completions, but every output token is ? (Unicode U+003F). SHA256 of the GGUF matches upstream. Prompt processing log timings are normal (~220 tok/s prefill). Decode throughput is normal (~63 tok/s).",
   "ram_gb": null,
   "retrieved": "2026-09-26",
   "source_name": "llama.cpp issue #23010 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/23010",
   "tps": 63,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-05",
   "flags": [],
   "hardware": "RTX 5090 \u00d72",
   "id": "githubissues-rtx5090-x2-qwen3-coder-next-q4-k-m",
   "model": "Qwen3-Coder-Next",
   "notes": "dual RTX 5090, amd64; same server-cuda-b9014 image family; cross-platform check cited in the issue",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "The same qwen3-coder-next-Q4_K_M.gguf runs cleanly on RTX 5090 \u00d7 2 (Blackwell, sm_120, amd64) at ~74 tok/s decode using the same server-cuda-b9014 image family.",
   "ram_gb": null,
   "retrieved": "2026-09-26",
   "source_name": "llama.cpp issue #23010 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/23010",
   "tps": 74,
   "ttft_s": null
  },
  {
   "backend": "ExLlamaV2",
   "batch": null,
   "ctx": null,
   "date": "2025-09",
   "flags": [],
   "hardware": "RTX 5090 24GB (laptop)",
   "id": "githubissues-rtx5090-laptop-tinyllama-exl2",
   "model": "TinyLlama-1.1B-intermediate",
   "notes": "EXL2 4.0 bpw build; reporter's best sustained tps on a 24GB laptop 5090, i9 24-core CPU, 64GB RAM; CUDA 12.x, PyTorch 2.8.0",
   "params": "1.1B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL2 4.0bpw",
   "quote": "TinyLlama-1.1B-intermediate; GPU: RTX 5090 24GB laptop; The maximum result I achieved in terms of speed is around 350/370 tokens/s",
   "ram_gb": 24,
   "retrieved": "2026-09-26",
   "source_name": "ExLlamaV2 issue #806 (community report)",
   "source_url": "https://github.com/turboderp-org/exllamav2/issues/806",
   "tps": 350,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-07",
   "flags": [],
   "hardware": "GB10 (DGX Spark)",
   "id": "githubissues-gb10-dgxspark-qwen35-122b-a10b-int4-vllm",
   "model": "Qwen3.5-122B-A10B",
   "notes": "single-stream decode, static MTP k=2; the dynamic k schedule variant reports 50.0 t/s single-stream; 8 concurrent x 180-token aggregate ~232 tok/s; FLASH_ATTN, enforce_eager=false",
   "params": "122B-A10B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "INT4 (AutoRound)",
   "quote": "NVIDIA GB10 / DGX Spark, sm_121a, 128 GB unified memory (~273 GB/s); vLLM 0.23.1rc1.dev1302+ge765bbc97; Model: Qwen3.5-122B-A10B INT4 (AutoRound), native MTP speculator, num_speculative_tokens: 2; Single-stream tok/s | 59.6 | 50.0 (expected PIECEWISE downgrade)",
   "ram_gb": 128,
   "retrieved": "2026-09-26",
   "source_name": "vLLM issue #49548 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/49548",
   "tps": 59.6,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-07",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "vllm-49011-rtx5090-qwen36-27b-nvfp4",
   "model": "Qwen3.6-27B",
   "notes": "single-stream decode with MTP num_speculative_tokens=4 (acceptance 91-100%), effective output rate; measured on a prototype SM120 nvfp4-KV wiring (XQA decode + fa2 prefill) on top of vLLM 0.25.1",
   "params": "27B hybrid GDN + full attention (MTP head)",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "GPU: RTX 5090 (SM120, 32 GB) | driver 595.71.05, CUDA 13.2; vLLM 0.25.1 | flashinfer-python 0.6.13 | torch 2.11.0; Model: unsloth/Qwen3.6-27B-NVFP4 (hybrid GDN + full attention, 24 Q heads / 4 KV heads, head_dim 256, MTP head); 95.6 tok/s single-stream, MTP acceptance 91\u2013100%; KV ceiling 244,928 tokens at util 0.97 (fp8 KV tops out ~163K on the same card)",
   "ram_gb": 32,
   "retrieved": "2026-09-26",
   "source_name": "vLLM issue #49011 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/49011",
   "tps": 95.6,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-06",
   "flags": [],
   "hardware": "B300",
   "id": "vllm-45741-b300-llama33-70b-nvfp4",
   "model": "Llama 3.3 70B",
   "notes": "con=1, ISL/OSL 1000/1000, aiperf profile, vLLM nightly (CUDA 13.0.2, PyTorch 2.11.0+cu130); fully-reverted zeros build = pre-#42988 behavior; the minimal one-line fix build measured 135.40 tok/s (ITL 7.35 ms) in the same B300 isolation table",
   "params": "70B dense",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "NVIDIA Blackwell \u2014 verified on B300 (TP4, 1 node); Llama-3.3-70B-Instruct, NVFP4 checkpoint (dense, TP4); ISL=1000, OSL=1000; in the B300 isolation run the fully-reverted (zeros) build measured 162.29 tok/s, i.e. no regression relative to the historical baseline",
   "ram_gb": null,
   "retrieved": "2026-09-26",
   "source_name": "vLLM issue #45741 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/45741",
   "tps": 162.29,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-06",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-45741-gb300-llama33-70b-nvfp4",
   "model": "Llama 3.3 70B",
   "notes": "con=1, ISL/OSL 1000/1000, fixed (zeros) build after the #42988 NVFP4 decode regression; the regressed build measured 18.39 tok/s on the same tray",
   "params": "70B dense",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Llama-3.3-70B-Instruct, NVFP4 checkpoint (dense, TP4); GB300NVL (TP4, 1 node); con=1, ISL/OSL=1000/1000; Llama-3.3-70B (dense) zeros (fixed) 143.25 6.88 997.8",
   "ram_gb": null,
   "retrieved": "2026-09-26",
   "source_name": "vLLM issue #45741 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/45741",
   "tps": 143.25,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-06",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-45741-gb300-kimi-k25-nvfp4",
   "model": "Kimi-K2.5",
   "notes": "con=1, ISL/OSL 1000/1000, fixed (zeros) build after the #42988 NVFP4 decode regression; the regressed build measured 10.16 tok/s on the same tray",
   "params": "1T-A32B MoE",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Kimi-K2.5, NVFP4 checkpoint (1T-A32B MoE, TP4, --kv-cache-dtype fp8_e4m3); GB300NVL (TP4, 1 node); con=1, ISL/OSL=1000/1000; Kimi-K2.5 (1T MoE) zeros (fixed) 148.28 6.58 999.6",
   "ram_gb": null,
   "retrieved": "2026-09-26",
   "source_name": "vLLM issue #45741 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/45741",
   "tps": 148.28,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 128,
   "ctx": null,
   "date": "2026-06",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-45741-gb300-llama33-70b-nvfp4-c128",
   "model": "Llama 3.3 70B",
   "notes": "aggregate output tok/s at concurrency 128, ISL/OSL 1000/1000, fixed build; ITL 10.73 ms; throughput scales monotonically across 1/16/64/128 in the same table",
   "params": "70B dense",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Llama-3.3-70B-Instruct, NVFP4 checkpoint (dense, TP4); GB300NVL (TP4, 1 node); Result 3 \u2014 Fixed build scales cleanly across concurrency on GB300NVL; 128 11478.84 10.73 999.8 4913.39 25.48 1000.0",
   "ram_gb": null,
   "retrieved": "2026-09-26",
   "source_name": "vLLM issue #45741 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/45741",
   "tps": 11478.84,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 128,
   "ctx": null,
   "date": "2026-06",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-45741-gb300-kimi-k25-nvfp4-c128",
   "model": "Kimi-K2.5",
   "notes": "aggregate output tok/s at concurrency 128, ISL/OSL 1000/1000, fixed build; ITL 25.48 ms",
   "params": "1T-A32B MoE",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Kimi-K2.5, NVFP4 checkpoint (1T-A32B MoE, TP4, --kv-cache-dtype fp8_e4m3); GB300NVL (TP4, 1 node); Result 3 \u2014 Fixed build scales cleanly across concurrency on GB300NVL; 128 11478.84 10.73 999.8 4913.39 25.48 1000.0",
   "ram_gb": null,
   "retrieved": "2026-09-26",
   "source_name": "vLLM issue #45741 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/45741",
   "tps": 4913.39,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 200,
   "ctx": null,
   "date": "2026-07",
   "flags": [],
   "hardware": "B300",
   "id": "vllm-49369-b300-dsv4-flash-fp4-dspark",
   "model": "DeepSeek-V4-Flash",
   "notes": "DSpark speculative decoding k=7 (mean accepted length 3.87, draft acceptance ~41%); aggregate output tok/s; vLLM 0.25.0, fp8 KV cache, FP4 indexer cache, max_model_len=16384, max_num_seqs=256; same issue reports ~773 aggregate tok/s for the base model without DSpark (prefix caching on); config=dspark-k7",
   "params": "MoE",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP4",
   "quote": "Running DeepSeek-V4-Flash-DSpark offline (in-process LLM + llm.chat()) on a single B300 (SM100); Offline batch of ~200 short prompts, reasoning on (~310 output tokens each); DSpark on (speculative_config above); ~344; Aggregate output tok/s",
   "ram_gb": null,
   "retrieved": "2026-09-26",
   "source_name": "vLLM issue #49369 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/49369",
   "tps": "344",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 200,
   "ctx": null,
   "date": "2026-07",
   "flags": [],
   "hardware": "B300",
   "id": "vllm-49370-b300-dsv4-flash-fp4-base",
   "model": "DeepSeek-V4-Flash",
   "notes": "default startup (breakable cudagraph auto-enabled, no torch.compile); aggregate offline throughput; vLLM 0.25.0, fp8 KV cache, FP4 indexer cache, max_model_len=16384, max_num_seqs=256, offline batch ~200 requests; identical config reported as the No DSpark row (~773) in companion issue #49369; config=breakable-on",
   "params": "MoE",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP4",
   "quote": "Base DeepSeek-V4-Flash (no speculative decoding), offline LLM + llm.chat() on a single B300 (SM100); Setting VLLM_USE_BREAKABLE_CUDAGRAPH=0 raised aggregate offline throughput from ~773 to ~1216 tok/s (~1.57x)",
   "ram_gb": null,
   "retrieved": "2026-09-26",
   "source_name": "vLLM issue #49370 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/49370",
   "tps": "773",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 200,
   "ctx": null,
   "date": "2026-07",
   "flags": [],
   "hardware": "B300",
   "id": "vllm-49370-b300-dsv4-flash-fp4-nobreak",
   "model": "DeepSeek-V4-Flash",
   "notes": "VLLM_USE_BREAKABLE_CUDAGRAPH=0 moves cudagraph capture to the normal FULL_AND_PIECEWISE path; gpu_memory_utilization 0.95, PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True; ~1.57x vs the 773 default config, same workload; config=breakable-off",
   "params": "MoE",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP4",
   "quote": "Base DeepSeek-V4-Flash (no speculative decoding), offline LLM + llm.chat() on a single B300 (SM100); Setting VLLM_USE_BREAKABLE_CUDAGRAPH=0 raised aggregate offline throughput from ~773 to ~1216 tok/s (~1.57x)",
   "ram_gb": null,
   "retrieved": "2026-09-26",
   "source_name": "vLLM issue #49370 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/49370",
   "tps": "1216",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-58031-gb300-dsv41-flash-mxfp4",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "single-stream decode with DSpark speculative k=5; measured with VLLM_FLASHINFER_AUTOTUNE_SKIP_OPS set (default untuned kernel selection for trtllm_fp4_block_scale_moe); vLLM nightly 0.29.1rc1.dev452, FlashInfer 0.6.18.post1",
   "params": "MoE",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4 (experts) / MXFP8 (rest)",
   "quote": "DGX Station GB300 (SM103, compute capability 10.3); Model: DeepSeek-V4.1-Flash (MXFP4 experts / MXFP8 rest), TP1, --cpu-offload-gb 75, DSpark speculative k=5; Measured decode on this deployment is ~91 tok/s single-stream with the skip in place",
   "ram_gb": null,
   "retrieved": "2026-09-26",
   "source_name": "vLLM issue #58031 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58031",
   "tps": "91",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-07",
   "flags": [],
   "hardware": "H200",
   "id": "vllm-48071-h200-gemma4-31b-c1",
   "model": "Gemma-4-31B",
   "notes": "concurrency 1; --compilation-config.pass_config.fuse_allreduce_rms false (Workaround B); vLLM v0.23.1rc1.dev925 nightly 2026-07-08",
   "params": "31B dense",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "as-published",
   "quote": "GPUs: 8x NVIDIA H200-NVL, single node, bridged dual-island NVLink topology; vllm serve google/gemma-4-31b-it; ITL by concurrency with fusion disabled:; 1 | 13.1 ms | 73.9",
   "ram_gb": null,
   "retrieved": "2026-09-26",
   "source_name": "vLLM issue #48071 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/48071",
   "tps": "73.9",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": null,
   "date": "2026-07",
   "flags": [],
   "hardware": "H200",
   "id": "vllm-48071-h200-gemma4-31b-c4",
   "model": "Gemma-4-31B",
   "notes": "concurrency 4; --compilation-config.pass_config.fuse_allreduce_rms false (Workaround B); vLLM v0.23.1rc1.dev925 nightly 2026-07-08",
   "params": "31B dense",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "as-published",
   "quote": "GPUs: 8x NVIDIA H200-NVL, single node, bridged dual-island NVLink topology; vllm serve google/gemma-4-31b-it; ITL by concurrency with fusion disabled:; 4 | 13.4 ms | 280.4",
   "ram_gb": null,
   "retrieved": "2026-09-26",
   "source_name": "vLLM issue #48071 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/48071",
   "tps": "280.4",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": null,
   "date": "2026-07",
   "flags": [],
   "hardware": "H200",
   "id": "vllm-48071-h200-gemma4-31b-c8",
   "model": "Gemma-4-31B",
   "notes": "concurrency 8; --compilation-config.pass_config.fuse_allreduce_rms false (Workaround B); vLLM v0.23.1rc1.dev925 nightly 2026-07-08",
   "params": "31B dense",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "as-published",
   "quote": "GPUs: 8x NVIDIA H200-NVL, single node, bridged dual-island NVLink topology; vllm serve google/gemma-4-31b-it; ITL by concurrency with fusion disabled:; 8 | 14.1 ms | 525.4",
   "ram_gb": null,
   "retrieved": "2026-09-26",
   "source_name": "vLLM issue #48071 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/48071",
   "tps": "525.4",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": null,
   "date": "2026-07",
   "flags": [],
   "hardware": "H200",
   "id": "vllm-48071-h200-gemma4-31b-c16",
   "model": "Gemma-4-31B",
   "notes": "concurrency 16; --compilation-config.pass_config.fuse_allreduce_rms false (Workaround B); vLLM v0.23.1rc1.dev925 nightly 2026-07-08",
   "params": "31B dense",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "as-published",
   "quote": "GPUs: 8x NVIDIA H200-NVL, single node, bridged dual-island NVLink topology; vllm serve google/gemma-4-31b-it; ITL by concurrency with fusion disabled:; 16 | 15.1 ms | 945.3",
   "ram_gb": null,
   "retrieved": "2026-09-26",
   "source_name": "vLLM issue #48071 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/48071",
   "tps": "945.3",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 32,
   "ctx": null,
   "date": "2026-07",
   "flags": [],
   "hardware": "H200",
   "id": "vllm-48071-h200-gemma4-31b-c32",
   "model": "Gemma-4-31B",
   "notes": "concurrency 32; --compilation-config.pass_config.fuse_allreduce_rms false (Workaround B); vLLM v0.23.1rc1.dev925 nightly 2026-07-08",
   "params": "31B dense",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "as-published",
   "quote": "GPUs: 8x NVIDIA H200-NVL, single node, bridged dual-island NVLink topology; vllm serve google/gemma-4-31b-it; ITL by concurrency with fusion disabled:; 32 | 20.7 ms | 1364.4",
   "ram_gb": null,
   "retrieved": "2026-09-26",
   "source_name": "vLLM issue #48071 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/48071",
   "tps": "1364.4",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 32,
   "ctx": null,
   "date": "2026-07",
   "flags": [],
   "hardware": "H200",
   "id": "vllm-48071-h200-gemma4-31b-c32-trtllm",
   "model": "Gemma-4-31B",
   "notes": "trtllm allreduce fallback path, works but very slow (ITL p50 719.6 ms); bridged dual-island NVLink topology without NVSwitch multicast; vLLM v0.23.1rc1.dev925 nightly 2026-07-08; full NVSwitch H200x8 runs the same image at cc32 ITL ~9 ms; config=trtllm-allreduce",
   "params": "31B dense",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "as-published",
   "quote": "GPUs: 8x NVIDIA H200-NVL, single node, bridged dual-island NVLink topology; vllm serve google/gemma-4-31b-it; Workaround A: Force trtllm, Works But Very Slow; Output token throughput | 44.3 tok/s",
   "ram_gb": null,
   "retrieved": "2026-09-26",
   "source_name": "vLLM issue #48071 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/48071",
   "tps": "44.3",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 18,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "A100",
   "id": "vllm-55139-a100-gptoss20b-mrv1",
   "model": "gpt-oss 20B",
   "notes": "vllm bench serve, sonnet dataset, ISL/OSL 4096/1024, --max-concurrency 18, --request-rate inf; MRV1 allreduce; -pp 2 -tp 2 on 4x A100 80GB PCIe; config=MRV1",
   "params": "20B MoE",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "as-published",
   "quote": "GPU 0: NVIDIA A100 80GB PCIe; --model openai/gpt-oss-20b --enforce-eager; --max-model-len=6144 -pp 2 -tp=2; MRV1; Output token throughput (tok/s): 914.97",
   "ram_gb": null,
   "retrieved": "2026-09-26",
   "source_name": "vLLM issue #55139 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/55139",
   "tps": "914.97",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 18,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "A100",
   "id": "vllm-55139-a100-gptoss20b-mrv2",
   "model": "gpt-oss 20B",
   "notes": "vllm bench serve, sonnet dataset, ISL/OSL 4096/1024, --max-concurrency 18, --request-rate inf; MRV2 allreduce; -pp 2 -tp 2 on 4x A100 80GB PCIe; config=MRV2",
   "params": "20B MoE",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "as-published",
   "quote": "GPU 0: NVIDIA A100 80GB PCIe; --model openai/gpt-oss-20b --enforce-eager; --max-model-len=6144 -pp 2 -tp=2; MRV2; Output token throughput (tok/s): 501.49",
   "ram_gb": null,
   "retrieved": "2026-09-26",
   "source_name": "vLLM issue #55139 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/55139",
   "tps": "501.49",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": null,
   "date": "2026-04",
   "flags": [],
   "hardware": "H100",
   "id": "vllm-39323-h100-qwen35-35b-a3b-fp8-fa3",
   "model": "Qwen3.5-35B-A3B",
   "notes": "FlashAttn v3 (vLLM 0.19.0 default attention backend, buggy: 8.5x perf regression, fixed in nightly); vllm bench serve, random ISL/OSL 4096/256, 256 prompts, request-rate inf; same bench with --attention-backend FLASHINFER: 1,665 tok/s",
   "params": "35B MoE (A3B active)",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "On a 1xH100 machine, I launched vLLM with two configurations; Config 1) vllm serve Qwen/Qwen3.5-35B-A3B-FP8; | Output token throughput (tok/s) | 195 | 1,665 | +754% |",
   "ram_gb": null,
   "retrieved": "2026-09-26",
   "source_name": "vLLM issue #39323 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/39323",
   "tps": "195",
   "ttft_s": null
  },
  {
   "backend": "ExLlamaV2",
   "batch": 1,
   "ctx": null,
   "date": "2024-05",
   "flags": [],
   "hardware": "2080 Ti",
   "id": "githubissues-rtx2080ti-llama3-exl2-b1-batched",
   "model": "Llama 3",
   "notes": "generation t/s (only generated tokens, but time includes a bit of prefill), 16 prompts; issue body says llama3 while the linked gist script loads mistral-7b-instruct-exl2 4.0bpw; config=batched",
   "params": "8B (body says llama3; linked gist script loads mistral-7b-instruct-exl2)",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL2 4.0bpw",
   "quote": "All measurements done on llama3 exl2 4 bit on 2080 ti, CUDA 11.8, latest stable release of exllamav2; scaling with batched_inference.py; bs=1: 82 t/s (1x)",
   "ram_gb": 11,
   "retrieved": "2026-09-26",
   "source_name": "ExLlamaV2 issue #450 (community report)",
   "source_url": "https://github.com/turboderp-org/exllamav2/issues/450",
   "tps": 82,
   "ttft_s": null
  },
  {
   "backend": "ExLlamaV2",
   "batch": 2,
   "ctx": null,
   "date": "2024-05",
   "flags": [],
   "hardware": "2080 Ti",
   "id": "githubissues-rtx2080ti-llama3-exl2-b2-batched",
   "model": "Llama 3",
   "notes": "generation t/s (only generated tokens, but time includes a bit of prefill), 16 prompts; issue body says llama3 while the linked gist script loads mistral-7b-instruct-exl2 4.0bpw; config=batched",
   "params": "8B (body says llama3; linked gist script loads mistral-7b-instruct-exl2)",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL2 4.0bpw",
   "quote": "All measurements done on llama3 exl2 4 bit on 2080 ti, CUDA 11.8, latest stable release of exllamav2; scaling with batched_inference.py; bs=2: 121 t/s (1.48x of bs=1)",
   "ram_gb": 11,
   "retrieved": "2026-09-26",
   "source_name": "ExLlamaV2 issue #450 (community report)",
   "source_url": "https://github.com/turboderp-org/exllamav2/issues/450",
   "tps": 121,
   "ttft_s": null
  },
  {
   "backend": "ExLlamaV2",
   "batch": 4,
   "ctx": null,
   "date": "2024-05",
   "flags": [],
   "hardware": "2080 Ti",
   "id": "githubissues-rtx2080ti-llama3-exl2-b4-batched",
   "model": "Llama 3",
   "notes": "generation t/s (only generated tokens, but time includes a bit of prefill), 16 prompts; issue body says llama3 while the linked gist script loads mistral-7b-instruct-exl2 4.0bpw; config=batched",
   "params": "8B (body says llama3; linked gist script loads mistral-7b-instruct-exl2)",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL2 4.0bpw",
   "quote": "All measurements done on llama3 exl2 4 bit on 2080 ti, CUDA 11.8, latest stable release of exllamav2; scaling with batched_inference.py; bs=4: 177 t/s (2.16x)",
   "ram_gb": 11,
   "retrieved": "2026-09-26",
   "source_name": "ExLlamaV2 issue #450 (community report)",
   "source_url": "https://github.com/turboderp-org/exllamav2/issues/450",
   "tps": 177,
   "ttft_s": null
  },
  {
   "backend": "ExLlamaV2",
   "batch": 8,
   "ctx": null,
   "date": "2024-05",
   "flags": [],
   "hardware": "2080 Ti",
   "id": "githubissues-rtx2080ti-llama3-exl2-b8-batched",
   "model": "Llama 3",
   "notes": "generation t/s (only generated tokens, but time includes a bit of prefill), 16 prompts; issue body says llama3 while the linked gist script loads mistral-7b-instruct-exl2 4.0bpw; config=batched",
   "params": "8B (body says llama3; linked gist script loads mistral-7b-instruct-exl2)",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL2 4.0bpw",
   "quote": "All measurements done on llama3 exl2 4 bit on 2080 ti, CUDA 11.8, latest stable release of exllamav2; scaling with batched_inference.py; bs=8: 195 t/s (2.38x)",
   "ram_gb": 11,
   "retrieved": "2026-09-26",
   "source_name": "ExLlamaV2 issue #450 (community report)",
   "source_url": "https://github.com/turboderp-org/exllamav2/issues/450",
   "tps": 195,
   "ttft_s": null
  },
  {
   "backend": "ExLlamaV2",
   "batch": 1,
   "ctx": null,
   "date": "2024-05",
   "flags": [],
   "hardware": "2080 Ti",
   "id": "githubissues-rtx2080ti-llama3-exl2-b1-caches",
   "model": "Llama 3",
   "notes": "generation only (no prefill, guarded by torch.cuda.synchronize), max_parallel_seqs=1, 16 prompts; issue body says llama3 while the linked gist script loads mistral-7b-instruct-exl2 4.0bpw; config=multi-cache",
   "params": "8B (body says llama3; linked gist script loads mistral-7b-instruct-exl2)",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL2 4.0bpw",
   "quote": "All measurements done on llama3 exl2 4 bit on 2080 ti, CUDA 11.8, latest stable release of exllamav2; scaling with multiple_caches.py by varying max_parallel_seqs; bs=1: 80 t/s (1x)",
   "ram_gb": 11,
   "retrieved": "2026-09-26",
   "source_name": "ExLlamaV2 issue #450 (community report)",
   "source_url": "https://github.com/turboderp-org/exllamav2/issues/450",
   "tps": 80,
   "ttft_s": null
  },
  {
   "backend": "ExLlamaV2",
   "batch": 2,
   "ctx": null,
   "date": "2024-05",
   "flags": [],
   "hardware": "2080 Ti",
   "id": "githubissues-rtx2080ti-llama3-exl2-b2-caches",
   "model": "Llama 3",
   "notes": "generation only (no prefill, guarded by torch.cuda.synchronize), max_parallel_seqs=2, 16 prompts; issue body says llama3 while the linked gist script loads mistral-7b-instruct-exl2 4.0bpw; config=multi-cache",
   "params": "8B (body says llama3; linked gist script loads mistral-7b-instruct-exl2)",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL2 4.0bpw",
   "quote": "All measurements done on llama3 exl2 4 bit on 2080 ti, CUDA 11.8, latest stable release of exllamav2; scaling with multiple_caches.py by varying max_parallel_seqs; bs=2: 101 t/s (1.26x)",
   "ram_gb": 11,
   "retrieved": "2026-09-26",
   "source_name": "ExLlamaV2 issue #450 (community report)",
   "source_url": "https://github.com/turboderp-org/exllamav2/issues/450",
   "tps": 101,
   "ttft_s": null
  },
  {
   "backend": "ExLlamaV2",
   "batch": 4,
   "ctx": null,
   "date": "2024-05",
   "flags": [],
   "hardware": "2080 Ti",
   "id": "githubissues-rtx2080ti-llama3-exl2-b4-caches",
   "model": "Llama 3",
   "notes": "generation only (no prefill, guarded by torch.cuda.synchronize), max_parallel_seqs=4, 16 prompts; issue body says llama3 while the linked gist script loads mistral-7b-instruct-exl2 4.0bpw; config=multi-cache",
   "params": "8B (body says llama3; linked gist script loads mistral-7b-instruct-exl2)",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL2 4.0bpw",
   "quote": "All measurements done on llama3 exl2 4 bit on 2080 ti, CUDA 11.8, latest stable release of exllamav2; scaling with multiple_caches.py by varying max_parallel_seqs; bs=4: 113 t/s (1.41x)",
   "ram_gb": 11,
   "retrieved": "2026-09-26",
   "source_name": "ExLlamaV2 issue #450 (community report)",
   "source_url": "https://github.com/turboderp-org/exllamav2/issues/450",
   "tps": 113,
   "ttft_s": null
  },
  {
   "backend": "ExLlamaV2",
   "batch": 8,
   "ctx": null,
   "date": "2024-05",
   "flags": [],
   "hardware": "2080 Ti",
   "id": "githubissues-rtx2080ti-llama3-exl2-b8-caches",
   "model": "Llama 3",
   "notes": "generation only (no prefill, guarded by torch.cuda.synchronize), max_parallel_seqs=8, 16 prompts; issue body says llama3 while the linked gist script loads mistral-7b-instruct-exl2 4.0bpw; config=multi-cache",
   "params": "8B (body says llama3; linked gist script loads mistral-7b-instruct-exl2)",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL2 4.0bpw",
   "quote": "All measurements done on llama3 exl2 4 bit on 2080 ti, CUDA 11.8, latest stable release of exllamav2; scaling with multiple_caches.py by varying max_parallel_seqs; bs=8: 119 t/s (1.48x)",
   "ram_gb": 11,
   "retrieved": "2026-09-26",
   "source_name": "ExLlamaV2 issue #450 (community report)",
   "source_url": "https://github.com/turboderp-org/exllamav2/issues/450",
   "tps": 119,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": null,
   "date": "2026-07",
   "flags": [],
   "hardware": "H100",
   "id": "vllm-48518-h100-qwen3-8b-fp8-l2persist",
   "model": "Qwen3 8B",
   "notes": "512 ShareGPT prompts, no prefix caching, vLLM v0.25.1; run right after server start, high-priority L2 data still persisting from startup (4.83 ms TPOT, worse); config=l2-persist",
   "params": "8B dense",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "GPU 0: NVIDIA H100 80GB HBM3; --model RedHatAI/Qwen3-8B-FP8-dynamic --max-num-seqs 4 --tensor-parallel-size 1 --no-enable-prefix-caching; --max-concurrency 4; Output token throughput (tok/s): 809.40",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #48518 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/48518",
   "tps": "809.40",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": null,
   "date": "2026-07",
   "flags": [],
   "hardware": "H100",
   "id": "vllm-48518-h100-qwen3-8b-fp8-l2cleared",
   "model": "Qwen3 8B",
   "notes": "512 ShareGPT prompts, no prefix caching, vLLM v0.25.1; one-shot L2 cache clearing script run before benchmark (4.66 ms TPOT, better); config=l2-cleared",
   "params": "8B dense",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "GPU 0: NVIDIA H100 80GB HBM3; --model RedHatAI/Qwen3-8B-FP8-dynamic --max-num-seqs 4 --tensor-parallel-size 1 --no-enable-prefix-caching; --max-concurrency 4; Output token throughput (tok/s): 839.75",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #48518 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/48518",
   "tps": "839.75",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": null,
   "date": "2025-09",
   "flags": [],
   "hardware": "A100",
   "id": "vllm-24728-a100-qwen-img-inf",
   "model": "Qwen2.5-VL-7B",
   "notes": "VLM benchmark (multimodal, image input), vllm bench serve --request-rate inf max-QPS test, --max-concurrency 1, 50 prompts, ShareGPT4Video dataset; aggregate output tok/s; config=vlm-image",
   "params": "7B VLM",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "Hardware / Environment: single NVIDIA A100 40 GB, CUDA 12.x, vLLM latest main; Qwen/Qwen2.5-VL-7B-Instruct; Image Input Performance; Output Token Throughput (tok/s) | 1517.21 | 2320.44 | 2674.70 | 4734.68",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #24728 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/24728",
   "tps": "1517.21",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": null,
   "date": "2025-09",
   "flags": [],
   "hardware": "A100",
   "id": "vllm-24728-a100-minicpm-img-inf",
   "model": "MiniCPM-V-4",
   "notes": "VLM benchmark (multimodal, image input), vllm bench serve --request-rate inf max-QPS test, --max-concurrency 1, 50 prompts, ShareGPT4Video dataset; aggregate output tok/s; config=vlm-image",
   "params": "VLM",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "as-published",
   "quote": "Hardware / Environment: single NVIDIA A100 40 GB, CUDA 12.x, vLLM latest main; openbmb/MiniCPM-V-4; Image Input Performance; Output Token Throughput (tok/s) | 1517.21 | 2320.44 | 2674.70 | 4734.68",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #24728 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/24728",
   "tps": "2320.44",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": null,
   "date": "2025-09",
   "flags": [],
   "hardware": "A100",
   "id": "vllm-24728-a100-iv4-img-inf",
   "model": "InternVL3.5-4B",
   "notes": "VLM benchmark (multimodal, image input), vllm bench serve --request-rate inf max-QPS test, --max-concurrency 1, 50 prompts, ShareGPT4Video dataset; aggregate output tok/s; config=vlm-image",
   "params": "4B VLM",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "as-published",
   "quote": "Hardware / Environment: single NVIDIA A100 40 GB, CUDA 12.x, vLLM latest main; OpenGVLab/InternVL3_5-4B; Image Input Performance; Output Token Throughput (tok/s) | 1517.21 | 2320.44 | 2674.70 | 4734.68",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #24728 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/24728",
   "tps": "2674.70",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": null,
   "date": "2025-09",
   "flags": [],
   "hardware": "A100",
   "id": "vllm-24728-a100-iv2-img-inf",
   "model": "InternVL3.5-2B",
   "notes": "VLM benchmark (multimodal, image input), vllm bench serve --request-rate inf max-QPS test, --max-concurrency 1, 50 prompts, ShareGPT4Video dataset; aggregate output tok/s; config=vlm-image",
   "params": "2B VLM",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "as-published",
   "quote": "Hardware / Environment: single NVIDIA A100 40 GB, CUDA 12.x, vLLM latest main; OpenGVLab/InternVL3_5-2B; Image Input Performance; Output Token Throughput (tok/s) | 1517.21 | 2320.44 | 2674.70 | 4734.68",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #24728 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/24728",
   "tps": "4734.68",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": null,
   "date": "2025-09",
   "flags": [],
   "hardware": "A100",
   "id": "vllm-24728-a100-qwen-vid-inf",
   "model": "Qwen2.5-VL-7B",
   "notes": "VLM benchmark (multimodal, video input), vllm bench serve --request-rate inf max-QPS test, --max-concurrency 1, 50 prompts, ShareGPT4Video dataset; aggregate output tok/s; config=vlm-video",
   "params": "7B VLM",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "Hardware / Environment: single NVIDIA A100 40 GB, CUDA 12.x, vLLM latest main; Qwen/Qwen2.5-VL-7B-Instruct; Video Input Performance; Output Token Throughput (tok/s) | 1795.85 | 1272.71 | 2120.87 | 4143.84",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #24728 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/24728",
   "tps": "1795.85",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": null,
   "date": "2025-09",
   "flags": [],
   "hardware": "A100",
   "id": "vllm-24728-a100-minicpm-vid-inf",
   "model": "MiniCPM-V-4",
   "notes": "VLM benchmark (multimodal, video input), vllm bench serve --request-rate inf max-QPS test, --max-concurrency 1, 50 prompts, ShareGPT4Video dataset; aggregate output tok/s; config=vlm-video",
   "params": "VLM",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "as-published",
   "quote": "Hardware / Environment: single NVIDIA A100 40 GB, CUDA 12.x, vLLM latest main; openbmb/MiniCPM-V-4; Video Input Performance; Output Token Throughput (tok/s) | 1795.85 | 1272.71 | 2120.87 | 4143.84",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #24728 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/24728",
   "tps": "1272.71",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": null,
   "date": "2025-09",
   "flags": [],
   "hardware": "A100",
   "id": "vllm-24728-a100-iv4-vid-inf",
   "model": "InternVL3.5-4B",
   "notes": "VLM benchmark (multimodal, video input), vllm bench serve --request-rate inf max-QPS test, --max-concurrency 1, 50 prompts, ShareGPT4Video dataset; aggregate output tok/s; config=vlm-video",
   "params": "4B VLM",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "as-published",
   "quote": "Hardware / Environment: single NVIDIA A100 40 GB, CUDA 12.x, vLLM latest main; OpenGVLab/InternVL3_5-4B; Video Input Performance; Output Token Throughput (tok/s) | 1795.85 | 1272.71 | 2120.87 | 4143.84",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #24728 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/24728",
   "tps": "2120.87",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": null,
   "date": "2025-09",
   "flags": [],
   "hardware": "A100",
   "id": "vllm-24728-a100-iv2-vid-inf",
   "model": "InternVL3.5-2B",
   "notes": "VLM benchmark (multimodal, video input), vllm bench serve --request-rate inf max-QPS test, --max-concurrency 1, 50 prompts, ShareGPT4Video dataset; aggregate output tok/s; config=vlm-video",
   "params": "2B VLM",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "as-published",
   "quote": "Hardware / Environment: single NVIDIA A100 40 GB, CUDA 12.x, vLLM latest main; OpenGVLab/InternVL3_5-2B; Video Input Performance; Output Token Throughput (tok/s) | 1795.85 | 1272.71 | 2120.87 | 4143.84",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #24728 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/24728",
   "tps": "4143.84",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 10,
   "ctx": null,
   "date": "2025-09",
   "flags": [],
   "hardware": "A100",
   "id": "vllm-24728-a100-qwen-img-c10",
   "model": "Qwen2.5-VL-7B",
   "notes": "VLM benchmark (multimodal, image input), vllm bench serve fixed concurrency 10, ShareGPT4Video dataset; aggregate output tok/s; config=vlm-image",
   "params": "7B VLM",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "Hardware / Environment: single NVIDIA A100 40 GB, CUDA 12.x, vLLM latest main; Qwen/Qwen2.5-VL-7B; 10 | 6.08 | 713.45 | 28.28 | 13.69",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #24728 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/24728",
   "tps": "713.45",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 50,
   "ctx": null,
   "date": "2025-09",
   "flags": [],
   "hardware": "A100",
   "id": "vllm-24728-a100-qwen-img-c50",
   "model": "Qwen2.5-VL-7B",
   "notes": "VLM benchmark (multimodal, image input), vllm bench serve fixed concurrency 50, ShareGPT4Video dataset; aggregate output tok/s; config=vlm-image",
   "params": "7B VLM",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "Hardware / Environment: single NVIDIA A100 40 GB, CUDA 12.x, vLLM latest main; Qwen/Qwen2.5-VL-7B; 50 | 20.89 | 2451.80 | 58.31 | 16.36",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #24728 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/24728",
   "tps": "2451.80",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2025-09",
   "flags": [],
   "hardware": "A100",
   "id": "vllm-24728-a100-qwen-vid-c1",
   "model": "Qwen2.5-VL-7B",
   "notes": "VLM benchmark (multimodal, video input), vllm bench serve fixed concurrency 1, ShareGPT4Video dataset; aggregate output tok/s; config=vlm-video",
   "params": "7B VLM",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "Hardware / Environment: single NVIDIA A100 40 GB, CUDA 12.x, vLLM latest main; Qwen/Qwen2.5-VL-7B; 1 | 0.21 | 76.46 | 21.98 | 13.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #24728 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/24728",
   "tps": "76.46",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 10,
   "ctx": null,
   "date": "2025-09",
   "flags": [],
   "hardware": "A100",
   "id": "vllm-24728-a100-qwen-vid-c10",
   "model": "Qwen2.5-VL-7B",
   "notes": "VLM benchmark (multimodal, video input), vllm bench serve fixed concurrency 10, ShareGPT4Video dataset; aggregate output tok/s; config=vlm-video",
   "params": "7B VLM",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "Hardware / Environment: single NVIDIA A100 40 GB, CUDA 12.x, vLLM latest main; Qwen/Qwen2.5-VL-7B; 10 | 1.98 | 709.00 | 28.54 | 13.77",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #24728 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/24728",
   "tps": "709.00",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 50,
   "ctx": null,
   "date": "2025-09",
   "flags": [],
   "hardware": "A100",
   "id": "vllm-24728-a100-qwen-vid-c50",
   "model": "Qwen2.5-VL-7B",
   "notes": "VLM benchmark (multimodal, video input), vllm bench serve fixed concurrency 50, ShareGPT4Video dataset; aggregate output tok/s; config=vlm-video",
   "params": "7B VLM",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "Hardware / Environment: single NVIDIA A100 40 GB, CUDA 12.x, vLLM latest main; Qwen/Qwen2.5-VL-7B; 50 | 7.35 | 2632.88 | 63.79 | 16.25",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #24728 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/24728",
   "tps": "2632.88",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2025-09",
   "flags": [],
   "hardware": "A100",
   "id": "vllm-24728-a100-minicpm-img-c1",
   "model": "MiniCPM-V-4",
   "notes": "VLM benchmark (multimodal, image input), vllm bench serve fixed concurrency 1, ShareGPT4Video dataset; aggregate output tok/s; config=vlm-image",
   "params": "VLM",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "as-published",
   "quote": "Hardware / Environment: single NVIDIA A100 40 GB, CUDA 12.x, vLLM latest main; openbmb/MiniCPM-V-4; 1 | 1.07 | 126.22 | 15.16 | 7.88",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #24728 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/24728",
   "tps": "126.22",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 10,
   "ctx": null,
   "date": "2025-09",
   "flags": [],
   "hardware": "A100",
   "id": "vllm-24728-a100-minicpm-img-c10",
   "model": "MiniCPM-V-4",
   "notes": "VLM benchmark (multimodal, image input), vllm bench serve fixed concurrency 10, ShareGPT4Video dataset; aggregate output tok/s; config=vlm-image",
   "params": "VLM",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "as-published",
   "quote": "Hardware / Environment: single NVIDIA A100 40 GB, CUDA 12.x, vLLM latest main; openbmb/MiniCPM-V-4; 10 | 9.52 | 1114.85 | 19.11 | 8.64",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #24728 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/24728",
   "tps": "1114.85",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 50,
   "ctx": null,
   "date": "2025-09",
   "flags": [],
   "hardware": "A100",
   "id": "vllm-24728-a100-minicpm-img-c50",
   "model": "MiniCPM-V-4",
   "notes": "VLM benchmark (multimodal, image input), vllm bench serve fixed concurrency 50, ShareGPT4Video dataset; aggregate output tok/s; config=vlm-image",
   "params": "VLM",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "as-published",
   "quote": "Hardware / Environment: single NVIDIA A100 40 GB, CUDA 12.x, vLLM latest main; openbmb/MiniCPM-V-4; 50 | 32.73 | 3822.97 | 51.44 | 10.54",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #24728 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/24728",
   "tps": "3822.97",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2025-09",
   "flags": [],
   "hardware": "A100",
   "id": "vllm-24728-a100-minicpm-vid-c1",
   "model": "MiniCPM-V-4",
   "notes": "VLM benchmark (multimodal, video input), vllm bench serve fixed concurrency 1, ShareGPT4Video dataset; aggregate output tok/s; config=vlm-video",
   "params": "VLM",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "as-published",
   "quote": "Hardware / Environment: single NVIDIA A100 40 GB, CUDA 12.x, vLLM latest main; openbmb/MiniCPM-V-4; 1 | 0.69 | 126.05 | 14.95 | 7.89",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #24728 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/24728",
   "tps": "126.05",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 10,
   "ctx": null,
   "date": "2025-09",
   "flags": [],
   "hardware": "A100",
   "id": "vllm-24728-a100-minicpm-vid-c10",
   "model": "MiniCPM-V-4",
   "notes": "VLM benchmark (multimodal, video input), vllm bench serve fixed concurrency 10, ShareGPT4Video dataset; aggregate output tok/s; config=vlm-video",
   "params": "VLM",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "as-published",
   "quote": "Hardware / Environment: single NVIDIA A100 40 GB, CUDA 12.x, vLLM latest main; openbmb/MiniCPM-V-4; 10 | 6.04 | 1065.36 | 22.33 | 8.67",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #24728 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/24728",
   "tps": "1065.36",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 50,
   "ctx": null,
   "date": "2025-09",
   "flags": [],
   "hardware": "A100",
   "id": "vllm-24728-a100-minicpm-vid-c50",
   "model": "MiniCPM-V-4",
   "notes": "VLM benchmark (multimodal, video input), vllm bench serve fixed concurrency 50, ShareGPT4Video dataset; aggregate output tok/s; config=vlm-video",
   "params": "VLM",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "as-published",
   "quote": "Hardware / Environment: single NVIDIA A100 40 GB, CUDA 12.x, vLLM latest main; openbmb/MiniCPM-V-4; 50 | 18.95 | 3239.55 | 65.85 | 10.95",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #24728 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/24728",
   "tps": "3239.55",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2025-09",
   "flags": [],
   "hardware": "A100",
   "id": "vllm-24728-a100-iv2-img-c1",
   "model": "InternVL3.5-2B",
   "notes": "VLM benchmark (multimodal, image input), vllm bench serve fixed concurrency 1, ShareGPT4Video dataset; aggregate output tok/s; config=vlm-image",
   "params": "2B VLM",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "as-published",
   "quote": "Hardware / Environment: single NVIDIA A100 40 GB, CUDA 12.x, vLLM latest main; OpenGVLab/InternVL3.5-2B; 1 | 1.09 | 202.74 | 11.12 | 4.89",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #24728 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/24728",
   "tps": "202.74",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 10,
   "ctx": null,
   "date": "2025-09",
   "flags": [],
   "hardware": "A100",
   "id": "vllm-24728-a100-iv2-img-c10",
   "model": "InternVL3.5-2B",
   "notes": "VLM benchmark (multimodal, image input), vllm bench serve fixed concurrency 10, ShareGPT4Video dataset; aggregate output tok/s; config=vlm-image",
   "params": "2B VLM",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "as-published",
   "quote": "Hardware / Environment: single NVIDIA A100 40 GB, CUDA 12.x, vLLM latest main; OpenGVLab/InternVL3.5-2B; 10 | 9.13 | 1691.45 | 17.53 | 5.66",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #24728 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/24728",
   "tps": "1691.45",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 50,
   "ctx": null,
   "date": "2025-09",
   "flags": [],
   "hardware": "A100",
   "id": "vllm-24728-a100-iv2-img-c50",
   "model": "InternVL3.5-2B",
   "notes": "VLM benchmark (multimodal, image input), vllm bench serve fixed concurrency 50, ShareGPT4Video dataset; aggregate output tok/s; config=vlm-image",
   "params": "2B VLM",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "as-published",
   "quote": "Hardware / Environment: single NVIDIA A100 40 GB, CUDA 12.x, vLLM latest main; OpenGVLab/InternVL3.5-2B; 50 | 31.93 | 5912.53 | 48.34 | 6.65",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #24728 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/24728",
   "tps": "5912.53",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2025-09",
   "flags": [],
   "hardware": "A100",
   "id": "vllm-24728-a100-iv2-vid-c1",
   "model": "InternVL3.5-2B",
   "notes": "VLM benchmark (multimodal, video input), vllm bench serve fixed concurrency 1, ShareGPT4Video dataset; aggregate output tok/s; config=vlm-video",
   "params": "2B VLM",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "as-published",
   "quote": "Hardware / Environment: single NVIDIA A100 40 GB, CUDA 12.x, vLLM latest main; OpenGVLab/InternVL3.5-2B; 1 | 0.59 | 199.48 | 13.55 | 4.97",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #24728 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/24728",
   "tps": "199.48",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 10,
   "ctx": null,
   "date": "2025-09",
   "flags": [],
   "hardware": "A100",
   "id": "vllm-24728-a100-iv2-vid-c10",
   "model": "InternVL3.5-2B",
   "notes": "VLM benchmark (multimodal, video input), vllm bench serve fixed concurrency 10, ShareGPT4Video dataset; aggregate output tok/s; config=vlm-video",
   "params": "2B VLM",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "as-published",
   "quote": "Hardware / Environment: single NVIDIA A100 40 GB, CUDA 12.x, vLLM latest main; OpenGVLab/InternVL3.5-2B; 10 | 4.72 | 1683.27 | 18.37 | 5.80",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #24728 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/24728",
   "tps": "1683.27",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 50,
   "ctx": null,
   "date": "2025-09",
   "flags": [],
   "hardware": "A100",
   "id": "vllm-24728-a100-iv2-vid-c50",
   "model": "InternVL3.5-2B",
   "notes": "VLM benchmark (multimodal, video input), vllm bench serve fixed concurrency 50, ShareGPT4Video dataset; aggregate output tok/s; config=vlm-video",
   "params": "2B VLM",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "as-published",
   "quote": "Hardware / Environment: single NVIDIA A100 40 GB, CUDA 12.x, vLLM latest main; OpenGVLab/InternVL3.5-2B; 50 | 17.40 | 6016.16 | 48.43 | 7.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #24728 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/24728",
   "tps": "6016.16",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": null,
   "date": "2026-03",
   "flags": [],
   "hardware": "4090D",
   "id": "vllm-36629-4090d-qwen25-14b-fp8-eagle3",
   "model": "Qwen2.5-14B",
   "notes": "max-concurrency 16 (batch16), 500 prompts, 50 RPS cap; EAGLE3 speculative decoding, FP8; vllm 0.16.0; config=fp8-eagle3",
   "params": "14B dense",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "W4A16+eagle3 not better than fp8+eagle3 with Qwen2.5-14B; 4090d * 1; vllm 0.16.0; fp8+eagle3 batch16; Output token throughput (tok/s): 1131.19",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #36629 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/36629",
   "tps": "1131.19",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": null,
   "date": "2026-03",
   "flags": [],
   "hardware": "4090D",
   "id": "vllm-36629-4090d-qwen25-14b-w4a16-eagle3",
   "model": "Qwen2.5-14B",
   "notes": "max-concurrency 16 (batch16), 500 prompts, 50 RPS cap; EAGLE3 speculative decoding, W4A16; vllm 0.16.0; config=w4a16-eagle3",
   "params": "14B dense",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "W4A16",
   "quote": "W4A16+eagle3 not better than fp8+eagle3 with Qwen2.5-14B; 4090d * 1; vllm 0.16.0; w4a16+eagle3 batch16; Output token throughput (tok/s): 1079.15",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #36629 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/36629",
   "tps": "1079.15",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-03",
   "flags": [],
   "hardware": "H200",
   "id": "vllm-37441-h200-gptoss120b-v0160-triton35",
   "model": "gpt-oss 120B",
   "notes": "single 8x H200 node, --tensor-parallel-size 2, --max_num_seqs 2, eagle speculative decoding (3 draft tokens), max-model-len 131072, con=1; vLLM 0.16.0 + Triton 3.5 (baseline); config=v0160-triton35",
   "params": "120B MoE",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "as-published",
   "quote": "GPU 0: NVIDIA H200; Running GPT OSS 120B on H200 GPU; --tensor-parallel-size 2; --max-concurrency 1; Output token throughput (tok/s): 275.84",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #37441 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/37441",
   "tps": "275.84",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-03",
   "flags": [],
   "hardware": "H200",
   "id": "vllm-37441-h200-gptoss120b-v0171-triton36",
   "model": "gpt-oss 120B",
   "notes": "single 8x H200 node, --tensor-parallel-size 2, --max_num_seqs 2, eagle speculative decoding (3 draft tokens), max-model-len 131072, con=1; vLLM 0.17.1 + Triton 3.6 new _reduce kernel (regression); config=v0171-triton36",
   "params": "120B MoE",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "as-published",
   "quote": "GPU 0: NVIDIA H200; Running GPT OSS 120B on H200 GPU; --tensor-parallel-size 2; --max-concurrency 1; Output token throughput (tok/s): 231.22",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #37441 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/37441",
   "tps": "231.22",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-03",
   "flags": [],
   "hardware": "H200",
   "id": "vllm-37441-h200-gptoss120b-v0171-legacy-triton35",
   "model": "gpt-oss 120B",
   "notes": "single 8x H200 node, --tensor-parallel-size 2, --max_num_seqs 2, eagle speculative decoding (3 draft tokens), max-model-len 131072, con=1; vLLM 0.17.1 with reverted Triton 3.5 legacy kernels (full recovery); config=v0171-legacy-triton35",
   "params": "120B MoE",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "as-published",
   "quote": "GPU 0: NVIDIA H200; Running GPT OSS 120B on H200 GPU; --tensor-parallel-size 2; --max-concurrency 1; Output token throughput (tok/s): 275.84",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #37441 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/37441",
   "tps": "275.84",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 256,
   "ctx": null,
   "date": "2025-11",
   "flags": [],
   "hardware": "B200",
   "id": "vllm-29662-b200-dsr1-0528-fp8-async",
   "model": "DeepSeek-R1-0528",
   "notes": "8x B200, TP8, MTP speculative decoding (3 tokens), max-num-seqs 1024, con=256, 1280 random prompts (1024/1024), --async-scheduling; kv-cache-dtype fp8, FLASHINFER_MLA; config=fp8-async",
   "params": "DSV1 MoE (0528)",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "B200 machine.; --model deepseek-ai/DeepSeek-R1-0528; --max-concurrency 256; with  --async-scheduling, Output token throughput (tok/s): 1994",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #29662 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/29662",
   "tps": "1994",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 256,
   "ctx": null,
   "date": "2025-11",
   "flags": [],
   "hardware": "B200",
   "id": "vllm-29662-b200-dsr1-0528-fp8-sync",
   "model": "DeepSeek-R1-0528",
   "notes": "8x B200, TP8, MTP speculative decoding (3 tokens), max-num-seqs 1024, con=256, 1280 random prompts (1024/1024), no --async-scheduling; kv-cache-dtype fp8, FLASHINFER_MLA; config=fp8-sync",
   "params": "DSV1 MoE (0528)",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "B200 machine.; --model deepseek-ai/DeepSeek-R1-0528; --max-concurrency 256; without  --async-scheduling, Output token throughput (tok/s): 2421",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #29662 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/29662",
   "tps": "2421",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 256,
   "ctx": null,
   "date": "2025-11",
   "flags": [],
   "hardware": "B200",
   "id": "vllm-29662-b200-dsr1-0528-fp4-async",
   "model": "DeepSeek-R1-0528-FP4-v2",
   "notes": "8x B200, TP8, MTP speculative decoding (3 tokens), max-num-seqs 1024, con=256, 1280 random prompts (1024/1024), --async-scheduling; kv-cache-dtype fp8, FLASHINFER_MLA; config=fp4-async",
   "params": "DSV1 MoE (0528)",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP4",
   "quote": "B200 machine.; --model nvidia/DeepSeek-R1-0528-FP4-v2; --max-concurrency 256; with  --async-scheduling, Output token throughput (tok/s): 2375",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #29662 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/29662",
   "tps": "2375",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 256,
   "ctx": null,
   "date": "2025-11",
   "flags": [],
   "hardware": "B200",
   "id": "vllm-29662-b200-dsr1-0528-fp4-sync",
   "model": "DeepSeek-R1-0528-FP4-v2",
   "notes": "8x B200, TP8, MTP speculative decoding (3 tokens), max-num-seqs 1024, con=256, 1280 random prompts (1024/1024), no --async-scheduling; kv-cache-dtype fp8, FLASHINFER_MLA; config=fp4-sync",
   "params": "DSV1 MoE (0528)",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP4",
   "quote": "B200 machine.; --model nvidia/DeepSeek-R1-0528-FP4-v2; --max-concurrency 256; without  --async-scheduling, Output token throughput (tok/s): 2895",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #29662 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/29662",
   "tps": "2895",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 100,
   "ctx": null,
   "date": "2026-02",
   "flags": [],
   "hardware": "RTX PRO 6000 Blackwell Max-Q",
   "id": "vllm-35625-rtxpro6000mq-qwen35-35b-bf16-c100",
   "model": "Qwen3.5-35B-A3B",
   "notes": "concurrency-100 bench (1000 requests, 128k in / 128k out tokens), Qwen3-Next MTP k=2 speculative decoding (acceptance 71.28%); mean TTFT 611.98 ms, P99 4756.51 ms in the same run; config=qwen3next-mtp2",
   "params": "35B MoE A3B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "GPU 0: NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition; vllm serve Qwen/Qwen3.5-35B-A3B; --speculative-config '{\"method\":\"qwen3_next_mtp\",\"num_speculative_tokens\":2}'; Maximum request concurrency:             100; Output token throughput (tok/s):         3341.89",
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "https://github.com/vllm-project/vllm/issues/35625",
   "source_url": "https://github.com/vllm-project/vllm/issues/35625",
   "tps": "3341.89",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 100,
   "ctx": null,
   "date": "2026-02",
   "flags": [],
   "hardware": "DGX Spark",
   "id": "vllm-35625-dgxspark-qwen35-35b-bf16-c100",
   "model": "Qwen3.5-35B-A3B",
   "notes": "same concurrency-100 bench (1000 requests) on the DGX Spark (GB10 chip, same hardware as the GB10 rows in this dataset); Qwen3-Next MTP k=2 (acceptance 70.24%); mean TTFT 2242.31 ms in the same run; config=qwen3next-mtp2",
   "params": "35B MoE A3B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "running a concurrency benchmark on a Blackwell 6000 Pro and DGX Spark; vllm serve Qwen/Qwen3.5-35B-A3B; DGX Spark:; Output token throughput (tok/s):         431.42",
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "https://github.com/vllm-project/vllm/issues/35625",
   "source_url": "https://github.com/vllm-project/vllm/issues/35625",
   "tps": "431.42",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 80,
   "ctx": null,
   "date": "2026-04",
   "flags": [],
   "hardware": "RTX PRO 6000 Blackwell Workstation Edition",
   "id": "vllm-40551-rtxpro6000-qwen3-8b-bf16-eagle3-mrv1-temp0",
   "model": "Qwen3 8B",
   "notes": "MT-Bench 80 requests (all concurrent, peak 80), Temp=0, MRV1 table, EAGLE3 RedHatAI speculator k=7; config=eagle3-mrv1-temp0",
   "params": "8B dense",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "GPU 0: NVIDIA RTX PRO 6000 Blackwell Workstation Edition; serve Qwen/Qwen3-8B --port 9000 --speculative-config '{\"method\": \"eagle3\", \"model\": \"RedHatAI/Qwen3-8B-speculator.eagle3\", \"num_speculative_tokens\": 7}'; ### MRV1 Temp=0; Output token throughput (tok/s):         3936.20",
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "https://github.com/vllm-project/vllm/issues/40551",
   "source_url": "https://github.com/vllm-project/vllm/issues/40551",
   "tps": "3936.20",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 80,
   "ctx": null,
   "date": "2026-04",
   "flags": [],
   "hardware": "RTX PRO 6000 Blackwell Workstation Edition",
   "id": "vllm-40551-rtxpro6000-qwen3-8b-bf16-eagle3-mrv1-temp1",
   "model": "Qwen3 8B",
   "notes": "MT-Bench 80 requests (all concurrent, peak 80), Temp=1, MRV1 table, EAGLE3 RedHatAI speculator k=7; config=eagle3-mrv1-temp1",
   "params": "8B dense",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "GPU 0: NVIDIA RTX PRO 6000 Blackwell Workstation Edition; serve Qwen/Qwen3-8B --port 9000 --speculative-config '{\"method\": \"eagle3\", \"model\": \"RedHatAI/Qwen3-8B-speculator.eagle3\", \"num_speculative_tokens\": 7}'; ### MRV1 Temp=1; Output token throughput (tok/s):         3713.23",
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "https://github.com/vllm-project/vllm/issues/40551",
   "source_url": "https://github.com/vllm-project/vllm/issues/40551",
   "tps": "3713.23",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 80,
   "ctx": null,
   "date": "2026-04",
   "flags": [],
   "hardware": "RTX PRO 6000 Blackwell Workstation Edition",
   "id": "vllm-40551-rtxpro6000-qwen3-8b-bf16-eagle3-mrv2-temp0",
   "model": "Qwen3 8B",
   "notes": "MT-Bench 80 requests (all concurrent, peak 80), Temp=0, MRV2 table, EAGLE3 RedHatAI speculator k=7; config=eagle3-mrv2-temp0",
   "params": "8B dense",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "GPU 0: NVIDIA RTX PRO 6000 Blackwell Workstation Edition; serve Qwen/Qwen3-8B --port 9000 --speculative-config '{\"method\": \"eagle3\", \"model\": \"RedHatAI/Qwen3-8B-speculator.eagle3\", \"num_speculative_tokens\": 7}'; ### MRV2 Temp=0; Output token throughput (tok/s):         3841.50",
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "https://github.com/vllm-project/vllm/issues/40551",
   "source_url": "https://github.com/vllm-project/vllm/issues/40551",
   "tps": "3841.50",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 80,
   "ctx": null,
   "date": "2026-04",
   "flags": [],
   "hardware": "RTX PRO 6000 Blackwell Workstation Edition",
   "id": "vllm-40551-rtxpro6000-qwen3-8b-bf16-eagle3-mrv2-temp1",
   "model": "Qwen3 8B",
   "notes": "MT-Bench 80 requests (all concurrent, peak 80), Temp=1, MRV2 table, EAGLE3 RedHatAI speculator k=7; config=eagle3-mrv2-temp1",
   "params": "8B dense",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "GPU 0: NVIDIA RTX PRO 6000 Blackwell Workstation Edition; serve Qwen/Qwen3-8B --port 9000 --speculative-config '{\"method\": \"eagle3\", \"model\": \"RedHatAI/Qwen3-8B-speculator.eagle3\", \"num_speculative_tokens\": 7}'; ### MRV2 Temp=1; Output token throughput (tok/s):         2021.33",
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "https://github.com/vllm-project/vllm/issues/40551",
   "source_url": "https://github.com/vllm-project/vllm/issues/40551",
   "tps": "2021.33",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 32,
   "ctx": null,
   "date": "2026-03",
   "flags": [],
   "hardware": "Xeon(R) 6767P",
   "id": "vllm-38586-xeon6767p-whisper-medium-bf16-c32",
   "model": "whisper-medium",
   "notes": "CPU-only (torch 2.10.0+cpu), whisper-medium ASR on librispeech_asr test (32 prompts, 8.2 s average audio), TP2, OMP threads 32 per socket on a dual-socket box; RTX 2.67x real-time in the same run; config=cpu-tp2",
   "params": "768M ASR",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "Model name:                              Intel(R) Xeon(R) 6767P; vllm serve openai/whisper-medium; --max-concurrency 32; Output token throughput (tok/s):         9.08",
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "https://github.com/vllm-project/vllm/issues/38586",
   "source_url": "https://github.com/vllm-project/vllm/issues/38586",
   "tps": "9.08",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": null,
   "date": "2025-10",
   "flags": [],
   "hardware": "A100 PCIe",
   "id": "vllm-27021-a100-pcie-qwen3vl30b-fp8-pre",
   "model": "Qwen3-VL-30B-A3B-Instruct-FP8",
   "notes": "request-rate 10 RPS, 100 VisionArena prompts, prefill/decode position-computation benchmark vs PR #25337 (before that PR); config=pre-pr25337",
   "params": "30B MoE A3B VLM",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "GPU: NVIDIA A100 PCIe; vllm serve \"Qwen/Qwen3-VL-30B-A3B-Instruct-FP8\"; ### Before PR #25337; Output token throughput (tok/s):         681.42",
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "https://github.com/vllm-project/vllm/issues/27021",
   "source_url": "https://github.com/vllm-project/vllm/issues/27021",
   "tps": "681.42",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": null,
   "date": "2025-10",
   "flags": [],
   "hardware": "A100 PCIe",
   "id": "vllm-27021-a100-pcie-qwen3vl30b-fp8-post",
   "model": "Qwen3-VL-30B-A3B-Instruct-FP8",
   "notes": "same request-rate 10 RPS bench with PR #25337 merged; config=pr25337",
   "params": "30B MoE A3B VLM",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "GPU: NVIDIA A100 PCIe; vllm serve \"Qwen/Qwen3-VL-30B-A3B-Instruct-FP8\"; ### After PR #25337; Output token throughput (tok/s):         689.02",
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "https://github.com/vllm-project/vllm/issues/27021",
   "source_url": "https://github.com/vllm-project/vllm/issues/27021",
   "tps": "689.02",
   "ttft_s": null
  },
  {
   "backend": "ExLlamaV2",
   "batch": 1,
   "ctx": null,
   "date": "2023-09",
   "flags": [],
   "hardware": "Tesla P40",
   "id": "exllamav2-40-p40-codellama34b-exl2",
   "model": "CodeLlama-34B",
   "notes": "2023 report, P40 at 80 W underload (the bug); test_inference.py at length 1024, all seven positions (1 to 896) land between 1.17 and 1.19 t/s; the position-1 value is recorded; the same issue mentions a 3090 at 36 to 39 t/s under a driver change but does not name the model, so it is not a row",
   "params": "34B",
   "power_w": 80,
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL2 4.0bpw H6",
   "quote": "Tesla P40 performance is still very low, only using 80W underload.; python test_inference.py -m models/CodeLlama-34B-instruct-4.0bpw-h6-exl2/ -s -l 1024; ** Position     1 + 127 tokens:    1.1905 t/s",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "ExLlamaV2 issue #40 (community report)",
   "source_url": "https://github.com/turboderp-org/exllamav2/issues/40",
   "tps": "1.1905",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-08",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-27734-r65536",
   "model": "Qwen3.8-27B",
   "notes": "llama-server single-stream 1200-token completion, wall-clock throughput; config=kvf16-suballoc-default",
   "params": "27B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "ctx | KV dtype | suballoc block size | wall time | throughput; AMD Radeon RX 7900 XTX (24 GB, RDNA3, gfx1100); Qwen3.8-27B (hybrid linear-attention, 64 layers), `Qwen3.8-27B-UD-Q4_K_XL.gguf` (16.35 GiB); 65536 | f16 | default (1 GiB) | ~30 s | **40.4 tok/s**",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "llama.cpp issue #27734 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27734",
   "tps": "40.4",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": 98304,
   "date": "2026-08",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-27734-r98304",
   "model": "Qwen3.8-27B",
   "notes": "llama-server single-stream 1200-token completion, wall-clock throughput; config=kvq8-suballoc-default",
   "params": "27B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "ctx | KV dtype | suballoc block size | wall time | throughput; AMD Radeon RX 7900 XTX (24 GB, RDNA3, gfx1100); Qwen3.8-27B (hybrid linear-attention, 64 layers), `Qwen3.8-27B-UD-Q4_K_XL.gguf` (16.35 GiB); 98304 | q8_0 | default | ~29 s | **40.8 tok/s**",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "llama.cpp issue #27734 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27734",
   "tps": "40.8",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-08",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-27734-r131072",
   "model": "Qwen3.8-27B",
   "notes": "llama-server single-stream 1200-token completion, wall-clock throughput; 131072 default suballoc is the fragmentation cliff; config=kvq8-suballoc-default",
   "params": "27B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "ctx | KV dtype | suballoc block size | wall time | throughput; AMD Radeon RX 7900 XTX (24 GB, RDNA3, gfx1100); Qwen3.8-27B (hybrid linear-attention, 64 layers), `Qwen3.8-27B-UD-Q4_K_XL.gguf` (16.35 GiB); **131072** | q8_0 | default | **~145 s** | **8.9 tok/s**",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "llama.cpp issue #27734 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27734",
   "tps": "8.9",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-08",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-27734-r131072-4gib",
   "model": "Qwen3.8-27B",
   "notes": "llama-server single-stream 1200-token completion, wall-clock throughput; 131072 default suballoc is the fragmentation cliff; config=kvq8-suballoc-4gib",
   "params": "27B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "ctx | KV dtype | suballoc block size | wall time | throughput; AMD Radeon RX 7900 XTX (24 GB, RDNA3, gfx1100); Qwen3.8-27B (hybrid linear-attention, 64 layers), `Qwen3.8-27B-UD-Q4_K_XL.gguf` (16.35 GiB); **131072** | q8_0 | **4 GiB** | **~30 s** | **40.3 tok/s**",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "llama.cpp issue #27734 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27734",
   "tps": "40.3",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-29410-b1",
   "model": "Qwopus3.6-35B-A3B-Coder-MTP",
   "notes": "empty-context llama-bench, ngl 99, ncmoe 18, fa on; config=build-c77ae695c",
   "params": "35B-A3B",
   "power_w": null,
   "pp_tps": 1174,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "GPU: AMD Radeon RX 7800 XT 16 GB (RDNA3, discrete), AMD proprietary Vulkan driver; Model: Qwen3.5/3.6 35B-A3B MoE, Q4_K_M, native MTP (`Qwopus3.6-35B-A3B-Coder-MTP-Q4_K_M.gguf`); Test | c77ae695c (17 Sep) | 84e76d8a2 (24 Sep); llama-bench -m Qwopus3.6-35B-A3B-Coder-MTP-Q4_K_M.gguf -dev Vulkan0 -ngl 99 -ncmoe 18 -lm none -fa on -p 512 -n 32 -r 2; llama-bench tg32 (no MTP) | 42.4 t/s | 36.1 \u00b1 0.6 t/s; llama-bench pp512 | 1174 t/s | 1151 \u00b1 9 t/s",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "llama.cpp issue #29410 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29410",
   "tps": "42.4",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-29410-b2",
   "model": "Qwopus3.6-35B-A3B-Coder-MTP",
   "notes": "empty-context llama-bench, ngl 99, ncmoe 18, fa on; config=build-84e76d8a2",
   "params": "35B-A3B",
   "power_w": null,
   "pp_tps": 1151,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "GPU: AMD Radeon RX 7800 XT 16 GB (RDNA3, discrete), AMD proprietary Vulkan driver; Model: Qwen3.5/3.6 35B-A3B MoE, Q4_K_M, native MTP (`Qwopus3.6-35B-A3B-Coder-MTP-Q4_K_M.gguf`); Test | c77ae695c (17 Sep) | 84e76d8a2 (24 Sep); llama-bench -m Qwopus3.6-35B-A3B-Coder-MTP-Q4_K_M.gguf -dev Vulkan0 -ngl 99 -ncmoe 18 -lm none -fa on -p 512 -n 32 -r 2; llama-bench tg32 (no MTP) | 42.4 t/s | 36.1 \u00b1 0.6 t/s; llama-bench pp512 | 1174 t/s | 1151 \u00b1 9 t/s",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "llama.cpp issue #29410 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29410",
   "tps": "36.1",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": 48000,
   "date": "2026-09",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-29410-b3",
   "model": "Qwopus3.6-35B-A3B-Coder-MTP",
   "notes": "llama-server draft-mtp n_max 3, tg128 at KV depth 48000, cache q4_0, ctx-size 163840, threads=12; config=build-c77ae695c-mtp3",
   "params": "35B-A3B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "GPU: AMD Radeon RX 7800 XT 16 GB (RDNA3, discrete), AMD proprietary Vulkan driver; Model: Qwen3.5/3.6 35B-A3B MoE, Q4_K_M, native MTP (`Qwopus3.6-35B-A3B-Coder-MTP-Q4_K_M.gguf`); Test | c77ae695c (17 Sep) | 84e76d8a2 (24 Sep); MTP n=3, tg128 @ 48k | 55.79 t/s | 25.80 t/s",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "llama.cpp issue #29410 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29410",
   "tps": "55.79",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": 130000,
   "date": "2026-09",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-29410-b4",
   "model": "Qwopus3.6-35B-A3B-Coder-MTP",
   "notes": "llama-server draft-mtp n_max 3, tg128 at KV depth 130000, cache q4_0, ctx-size 163840, threads=12; config=build-c77ae695c-mtp3",
   "params": "35B-A3B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "GPU: AMD Radeon RX 7800 XT 16 GB (RDNA3, discrete), AMD proprietary Vulkan driver; Model: Qwen3.5/3.6 35B-A3B MoE, Q4_K_M, native MTP (`Qwopus3.6-35B-A3B-Coder-MTP-Q4_K_M.gguf`); Test | c77ae695c (17 Sep) | 84e76d8a2 (24 Sep); MTP n=3, tg128 @ 130k | 39.15 t/s | 27.14 t/s",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "llama.cpp issue #29410 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29410",
   "tps": "39.15",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": 48000,
   "date": "2026-09",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-29410-b5",
   "model": "Qwopus3.6-35B-A3B-Coder-MTP",
   "notes": "llama-server draft-mtp n_max 3, tg128 at KV depth 48000, cache q4_0, ctx-size 163840, threads=12; config=build-84e76d8a2-mtp3",
   "params": "35B-A3B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "GPU: AMD Radeon RX 7800 XT 16 GB (RDNA3, discrete), AMD proprietary Vulkan driver; Model: Qwen3.5/3.6 35B-A3B MoE, Q4_K_M, native MTP (`Qwopus3.6-35B-A3B-Coder-MTP-Q4_K_M.gguf`); Test | c77ae695c (17 Sep) | 84e76d8a2 (24 Sep); MTP n=3, tg128 @ 48k | 55.79 t/s | 25.80 t/s",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "llama.cpp issue #29410 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29410",
   "tps": "25.8",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": 130000,
   "date": "2026-09",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-29410-b6",
   "model": "Qwopus3.6-35B-A3B-Coder-MTP",
   "notes": "llama-server draft-mtp n_max 3, tg128 at KV depth 130000, cache q4_0, ctx-size 163840, threads=12; config=build-84e76d8a2-mtp3",
   "params": "35B-A3B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "GPU: AMD Radeon RX 7800 XT 16 GB (RDNA3, discrete), AMD proprietary Vulkan driver; Model: Qwen3.5/3.6 35B-A3B MoE, Q4_K_M, native MTP (`Qwopus3.6-35B-A3B-Coder-MTP-Q4_K_M.gguf`); Test | c77ae695c (17 Sep) | 84e76d8a2 (24 Sep); MTP n=3, tg128 @ 130k | 39.15 t/s | 27.14 t/s",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "llama.cpp issue #29410 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29410",
   "tps": "27.14",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "RTX 5070 Ti",
   "id": "lc-27171-b10283",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fit-target 1024, fa on, KV q4_0, threads=8; config=build-10283",
   "params": "35B-A3B",
   "power_w": null,
   "pp_tps": 1265.61,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "1 x NVIDIA GeForce RTX 5070 Ti; llama-bench -m /var/lib/machines/bartowski/Qwen_Qwen3.6-35B-A3B-Q4_K_M.gguf --fit-target 1024 -fa 1 -ctk q4_0 -ctv q4_0 -t 8 -p 2048 -n 256; qwen35moe 35B.A3B Q4_K - Medium | 20.74 GiB | 35.51 B | CUDA,BLAS | 8 | q4_0 | q4_0 | 1 | 1024 | tg256 | 101.00 \u00b1 0.39; qwen35moe 35B.A3B Q4_K - Medium | 20.74 GiB | 35.51 B | CUDA,BLAS | 8 | q4_0 | q4_0 | 1 | 1024 | pp2048 | 1265.61 \u00b1 29.25",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "llama.cpp issue #27171 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27171",
   "tps": "101.0",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "RTX 5070 Ti",
   "id": "lc-27171-b10284",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fit-target 1024, fa on, KV q4_0, threads=8; config=build-10284",
   "params": "35B-A3B",
   "power_w": null,
   "pp_tps": 923.45,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "1 x NVIDIA GeForce RTX 5070 Ti; llama-bench -m /var/lib/machines/bartowski/Qwen_Qwen3.6-35B-A3B-Q4_K_M.gguf --fit-target 1024 -fa 1 -ctk q4_0 -ctv q4_0 -t 8 -p 2048 -n 256; qwen35moe 35B.A3B Q4_K - Medium | 20.74 GiB | 35.51 B | CUDA,BLAS | 8 | q4_0 | q4_0 | 1 | 1024 | tg256 | 87.92 \u00b1 0.82; qwen35moe 35B.A3B Q4_K - Medium | 20.74 GiB | 35.51 B | CUDA,BLAS | 8 | q4_0 | q4_0 | 1 | 1024 | pp2048 | 923.45 \u00b1 14.56",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "llama.cpp issue #27171 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27171",
   "tps": "87.92",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "Radeon RX 6700 XT",
   "id": "lc-26702-rocm",
   "model": "gemma-4-12b-it",
   "notes": "avg of 3 runs, LACT power 182/223W, VRAM 1900MHz, fa on, KV q8_0",
   "params": "12B",
   "power_w": null,
   "pp_tps": 653.9,
   "provenance": "community",
   "quant": "IQ4_NL",
   "quote": "AMD Radeon RX 6700 XT (gfx1031, RDNA2, Navi22); Model: `gemma-4-12b-it-IQ4_NL.gguf` (6.24 GiB, 11.91B params); llama-bench --n-prompt 8192 --n-gen 512 --batch-size 1024 --ubatch-size 1024 --cache-type-k q8_0 --cache-type-v q8_0 --n-gpu-layers 99 --flash-attn on; Backend | pp8192 (t/s) | tg512 (t/s); ROCm (avg of 3) | 653.9 | 34.60",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "llama.cpp issue #26702 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/26702",
   "tps": "34.6",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "Radeon RX 6700 XT",
   "id": "lc-26702-vulkan",
   "model": "gemma-4-12b-it",
   "notes": "avg of 3 runs, LACT power 182/223W, VRAM 1900MHz, fa on, KV q8_0",
   "params": "12B",
   "power_w": null,
   "pp_tps": 354.4,
   "provenance": "community",
   "quant": "IQ4_NL",
   "quote": "AMD Radeon RX 6700 XT (gfx1031, RDNA2, Navi22); Model: `gemma-4-12b-it-IQ4_NL.gguf` (6.24 GiB, 11.91B params); llama-bench --n-prompt 8192 --n-gen 512 --batch-size 1024 --ubatch-size 1024 --cache-type-k q8_0 --cache-type-v q8_0 --n-gpu-layers 99 --flash-attn on; Backend | pp8192 (t/s) | tg512 (t/s); Vulkan (avg of 3) | 354.4 | 40.92",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "llama.cpp issue #26702 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/26702",
   "tps": "40.92",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-26674-gemma4-31b",
   "model": "gemma-4-31B-it",
   "notes": "n-gl 999, fa on; real-inference llama-server ~6 tok/s per reporter, CPU ~95-100%",
   "params": "31B",
   "power_w": null,
   "pp_tps": 41.84,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "GPU: NVIDIA GeForce RTX 5060 Ti 16GB; llama-bench.exe -m \"gemma-4-31B-it-Q6_K.gguf\" -ngl 999 -fa on; gemma4 31B Q6_K | CUDA | 999 | 1 | tg128 | 0.62 \u00b1 0.00; gemma4 31B Q6_K | CUDA | 999 | 1 | pp512 | 41.84 \u00b1 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "llama.cpp issue #26674 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/26674",
   "tps": "0.62",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-26674-qwen36-35b",
   "model": "Qwen3.6-35B-A3B",
   "notes": "ngl 999, fa on; Qwen fine on same hardware, Gemma 4 slow (reporter)",
   "params": "35B-A3B",
   "power_w": null,
   "pp_tps": 108.23,
   "provenance": "community",
   "quant": "UD-Q6_K",
   "quote": "GPU: NVIDIA GeForce RTX 5060 Ti 16GB; llama-bench.exe -m \"Qwen3.6-35B-A3B-UD-Q6_K.gguf\" -ngl 999 -fa on; qwen35moe 35B.A3B Q6_K | CUDA | 999 | 1 | tg128 | 5.49 \u00b1 0.01; qwen35moe 35B.A3B Q6_K | CUDA | 999 | 1 | pp512 | 108.23 \u00b1 0.67",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "llama.cpp issue #26674 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/26674",
   "tps": "5.49",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (OpenVINO)",
   "batch": 1,
   "ctx": 128,
   "date": "2026-09",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-29235-gpu1-d128",
   "model": "Qwen2.5-7B-Instruct",
   "notes": "llama-bench -r 3, OpenVINO device GPU.1 vs GPU name (other GPU not masked); d128 depth; config=openvino-gpu1",
   "params": "7B",
   "power_w": null,
   "pp_tps": 5437.3,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Intel Arc Pro B70; Qwen2.5-7B-Instruct-Q4_K_M.gguf; A=\"-m $HOME/models/Qwen2.5-7B-Instruct-Q4_K_M.gguf -ngl 999 -p 512 -n 128 -d 128,4096 -r 3\"; OpenVINO, Arc as GPU.1 | 44.1 | **9.6** (4.3x slow down) | 5437.3 | 2388.4",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "llama.cpp issue #29235 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29235",
   "tps": "44.1",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (OpenVINO)",
   "batch": 1,
   "ctx": 128,
   "date": "2026-09",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-29235-gpu-d128",
   "model": "Qwen2.5-7B-Instruct",
   "notes": "llama-bench -r 3, OpenVINO device GPU name with ZE_AFFINITY_MASK=0; d128 depth; config=openvino-gpu",
   "params": "7B",
   "power_w": null,
   "pp_tps": 5337.8,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Intel Arc Pro B70; Qwen2.5-7B-Instruct-Q4_K_M.gguf; A=\"-m $HOME/models/Qwen2.5-7B-Instruct-Q4_K_M.gguf -ngl 999 -p 512 -n 128 -d 128,4096 -r 3\"; OpenVINO, Arc as GPU | 59.7 | 41.1 | 5337.8 | 3043.0",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "llama.cpp issue #29235 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29235",
   "tps": "59.7",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (OpenVINO)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-09",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-29235-gpu1-d4096",
   "model": "Qwen2.5-7B-Instruct",
   "notes": "llama-bench -r 3, OpenVINO device GPU.1; d4096 depth, 4.3x slowdown vs d128; config=openvino-gpu1",
   "params": "7B",
   "power_w": null,
   "pp_tps": 2388.4,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Intel Arc Pro B70; Qwen2.5-7B-Instruct-Q4_K_M.gguf; A=\"-m $HOME/models/Qwen2.5-7B-Instruct-Q4_K_M.gguf -ngl 999 -p 512 -n 128 -d 128,4096 -r 3\"; OpenVINO, Arc as GPU.1 | 44.1 | **9.6** (4.3x slow down) | 5437.3 | 2388.4",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "llama.cpp issue #29235 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29235",
   "tps": "9.6",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (OpenVINO)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-09",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-29235-gpu-d4096",
   "model": "Qwen2.5-7B-Instruct",
   "notes": "llama-bench -r 3, OpenVINO device GPU name; d4096 depth; config=openvino-gpu",
   "params": "7B",
   "power_w": null,
   "pp_tps": 3043.0,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Intel Arc Pro B70; Qwen2.5-7B-Instruct-Q4_K_M.gguf; A=\"-m $HOME/models/Qwen2.5-7B-Instruct-Q4_K_M.gguf -ngl 999 -p 512 -n 128 -d 128,4096 -r 3\"; OpenVINO, Arc as GPU | 59.7 | 41.1 | 5337.8 | 3043.0",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "llama.cpp issue #29235 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29235",
   "tps": "41.1",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-29172-d0",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "ternary PTQ1_0 weights, qwen35 hybrid, ngl 99, fa on, engine-internal timings; config=depth0",
   "params": "27B",
   "power_w": null,
   "pp_tps": 464.81,
   "provenance": "community",
   "quant": "PTQ1_0",
   "quote": "GPU: NVIDIA GeForce RTX 5060 Ti 16GB (16283 MiB, compute capability 12.0 Blackwell, sm_120); prism-ml/Ternary-Bonsai-2-27B-gguf; llama-bench.exe -m Ternary-Bonsai-2-27B-PTQ1_0.gguf -ngl 99 -fa on -p 512 -n 32 -r 1; test | t/s; pp512 (d0) | 464.81; tg32 (d0) | 44.68",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "llama.cpp issue #29172 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29172",
   "tps": "44.68",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 154855,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-29172-d154855",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "ternary PTQ1_0 weights, qwen35 hybrid, ngl 99, fa on; ~21x prefill and ~5.5x decode collapse at KV depth 154855; config=depth154855",
   "params": "27B",
   "power_w": null,
   "pp_tps": 21.87,
   "provenance": "community",
   "quant": "PTQ1_0",
   "quote": "GPU: NVIDIA GeForce RTX 5060 Ti 16GB (16283 MiB, compute capability 12.0 Blackwell, sm_120); prism-ml/Ternary-Bonsai-2-27B-gguf; llama-bench.exe -m Ternary-Bonsai-2-27B-PTQ1_0.gguf -ngl 99 -fa on -p 512 -n 32 -r 1; test | t/s; **pp512 @ d154855** | **21.87**; **tg32 @ d154855** | **8.10**",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "llama.cpp issue #29172 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29172",
   "tps": "8.1",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "RTX 3060 + Intel Arc A770",
   "id": "lc-27137-b9006",
   "model": "Qwen3.6-27B",
   "notes": "model split across both GPUs (65/65 layers offloaded, 6577 MiB Vulkan0 + 9452 MiB Vulkan1), ngl 99; config=build-9006",
   "params": "27B",
   "power_w": null,
   "pp_tps": 97.37,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Hardware/configuration is an RTX 3060 + Intel Arc A770 using Vulkan, with a Qwen3.6 27B Q4_K_M model split across both GPUs; qwen35 27B Q4_K - Medium | 16.32 GiB | 26.90 B | Vulkan | 99 | tg128 | 7.88 \u00b1 0.36; qwen35 27B Q4_K - Medium | 16.32 GiB | 26.90 B | Vulkan | 99 | pp512 | 97.37 \u00b1 1.29",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "llama.cpp issue #27137 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27137",
   "tps": "7.88",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "RTX 3060 + Intel Arc A770",
   "id": "lc-27137-b10433",
   "model": "Qwen3.6-27B",
   "notes": "model split across both GPUs, ngl -1, auto flash-attn on; config=build-10433",
   "params": "27B",
   "power_w": null,
   "pp_tps": 98.26,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Hardware/configuration is an RTX 3060 + Intel Arc A770 using Vulkan, with a Qwen3.6 27B Q4_K_M model split across both GPUs; qwen35 27B Q4_K - Medium | 16.32 GiB | 26.90 B | Vulkan | -1 | tg128 | 2.93 \u00b1 0.01; qwen35 27B Q4_K - Medium | 16.32 GiB | 26.90 B | Vulkan | -1 | pp512 | 98.26 \u00b1 0.77",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "llama.cpp issue #27137 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27137",
   "tps": "2.93",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "RTX PRO 4000 Blackwell",
   "id": "lc-26750-cuda-mtp",
   "model": "Qwen3.5-9B",
   "notes": "config=mtp6; llama-server spec-type draft-mtp n-max 6, b10290 official image, 108 runs per cell, median; 61 with MTP vs 90 without",
   "params": "9B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "32% (61 vs 90 tok/s); unsloth/Qwen3.5-9B-MTP-GGUF; RTX PRO 4000 Blackwell 24GB",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #26750",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/26750",
   "tps": "61"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "RTX PRO 4000 Blackwell",
   "id": "lc-26750-cuda-base",
   "model": "Qwen3.5-9B",
   "notes": "config=baseline; same cell without speculation, b10290 official image; 90 vs 61 with MTP",
   "params": "9B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "32% (61 vs 90 tok/s); unsloth/Qwen3.5-9B-MTP-GGUF; RTX PRO 4000 Blackwell 24GB",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #26750",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/26750",
   "tps": "90"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "Radeon PRO W7900",
   "id": "lc-26750-w7900-mtp",
   "model": "Qwen3.5-9B",
   "notes": "config=mtp6; llama-server spec-type draft-mtp n-max 6, b10290 official image, 108 runs per cell, median; 218 with MTP vs 96 without",
   "params": "9B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "128% (218 vs 96 tok/s); unsloth/Qwen3.5-9B-MTP-GGUF; Radeon PRO W7900 48GB",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #26750",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/26750",
   "tps": "218"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "Radeon PRO W7900",
   "id": "lc-26750-w7900-base",
   "model": "Qwen3.5-9B",
   "notes": "config=baseline; same cell without speculation, b10290 official image; 96 vs 218 with MTP",
   "params": "9B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "128% (218 vs 96 tok/s); unsloth/Qwen3.5-9B-MTP-GGUF; Radeon PRO W7900 48GB",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #26750",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/26750",
   "tps": "96"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "RX 7800 XT",
   "id": "lc-26750-7800xt-mtp",
   "model": "Qwen3.5-9B",
   "notes": "config=mtp6; llama-server spec-type draft-mtp n-max 6, b10290 official image, 108 runs per cell, median; 124 with MTP vs 70 without",
   "params": "9B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "77% (124 vs 70 tok/s); unsloth/Qwen3.5-9B-MTP-GGUF; RX 7800 XT 16GB",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #26750",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/26750",
   "tps": "124"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "RX 7800 XT",
   "id": "lc-26750-7800xt-base",
   "model": "Qwen3.5-9B",
   "notes": "config=baseline; same cell without speculation, b10290 official image; 70 vs 124 with MTP",
   "params": "9B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "77% (124 vs 70 tok/s); unsloth/Qwen3.5-9B-MTP-GGUF; RX 7800 XT 16GB",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #26750",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/26750",
   "tps": "70"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-27050-b9660-1s",
   "model": "Qwen2.5-7B-Instruct",
   "notes": "config=build-9660; 512 prompt + 128 generated tokens per request, closed-loop 1 concurrent, mean of 2 repetitions",
   "params": "7B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "F16",
   "quote": "| `7dad2f1a1` (build 9660), default | 99.8 | 9.67 ms | 700.9 | 41.50 ms |; Model: Qwen2.5-7B-Instruct, F16 GGUF (14.19 GiB), converted from the official HF weights; NVIDIA GeForce RTX 5090 32GB (Blackwell, sm_120)",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27050",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27050",
   "tps": "99.8"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 32,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-27050-b9660-32s",
   "model": "Qwen2.5-7B-Instruct",
   "notes": "config=build-9660; 512 prompt + 128 generated tokens per request, closed-loop 32 concurrent, mean of 2 repetitions",
   "params": "7B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "F16",
   "quote": "| `7dad2f1a1` (build 9660), default | 99.8 | 9.67 ms | 700.9 | 41.50 ms |; Model: Qwen2.5-7B-Instruct, F16 GGUF (14.19 GiB), converted from the official HF weights; NVIDIA GeForce RTX 5090 32GB (Blackwell, sm_120)",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27050",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27050",
   "tps": "700.9"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-27050-b10423-1s",
   "model": "Qwen2.5-7B-Instruct",
   "notes": "config=build-10423; 512 prompt + 128 generated tokens per request, closed-loop 1 concurrent, mean of 2 repetitions",
   "params": "7B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "F16",
   "quote": "| `a94d563ed` (build 10423), default | 100.1 | 9.65 ms | 705.9 | 41.21 ms |; Model: Qwen2.5-7B-Instruct, F16 GGUF (14.19 GiB), converted from the official HF weights; NVIDIA GeForce RTX 5090 32GB (Blackwell, sm_120)",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27050",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27050",
   "tps": "100.1"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 32,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-27050-b10423-32s",
   "model": "Qwen2.5-7B-Instruct",
   "notes": "config=build-10423; 512 prompt + 128 generated tokens per request, closed-loop 32 concurrent, mean of 2 repetitions",
   "params": "7B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "F16",
   "quote": "| `a94d563ed` (build 10423), default | 100.1 | 9.65 ms | 705.9 | 41.21 ms |; Model: Qwen2.5-7B-Instruct, F16 GGUF (14.19 GiB), converted from the official HF weights; NVIDIA GeForce RTX 5090 32GB (Blackwell, sm_120)",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27050",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27050",
   "tps": "705.9"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-27050-b10423bs-1s",
   "model": "Qwen2.5-7B-Instruct",
   "notes": "config=build-10423-bs; 512 prompt + 128 generated tokens per request, closed-loop 1 concurrent, mean of 2 repetitions",
   "params": "7B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "F16",
   "quote": "| `a94d563ed` (build 10423), **`-bs`** | 101.4 | 9.43 ms | **1045.8** | **25.39 ms** |; Model: Qwen2.5-7B-Instruct, F16 GGUF (14.19 GiB), converted from the official HF weights; NVIDIA GeForce RTX 5090 32GB (Blackwell, sm_120)",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27050",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27050",
   "tps": "101.4"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 32,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-27050-b10423bs-32s",
   "model": "Qwen2.5-7B-Instruct",
   "notes": "config=build-10423-bs; 512 prompt + 128 generated tokens per request, closed-loop 32 concurrent, mean of 2 repetitions",
   "params": "7B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "F16",
   "quote": "| `a94d563ed` (build 10423), **`-bs`** | 101.4 | 9.43 ms | **1045.8** | **25.39 ms** |; Model: Qwen2.5-7B-Instruct, F16 GGUF (14.19 GiB), converted from the official HF weights; NVIDIA GeForce RTX 5090 32GB (Blackwell, sm_120)",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27050",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27050",
   "tps": "1045.8"
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-27050-vllm-1s",
   "model": "Qwen2.5-7B-Instruct",
   "notes": "config=vllm-bf16; 512 prompt + 128 generated tokens per request, closed-loop 1 concurrent, mean of 2 repetitions",
   "params": "7B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "| vLLM BF16 (reference) | 92.9 | 10.41 ms | 1404.7 | 17.68 ms |; Model: Qwen2.5-7B-Instruct, F16 GGUF (14.19 GiB), converted from the official HF weights; NVIDIA GeForce RTX 5090 32GB (Blackwell, sm_120)",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27050",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27050",
   "tps": "92.9"
  },
  {
   "backend": "vLLM",
   "batch": 32,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-27050-vllm-32s",
   "model": "Qwen2.5-7B-Instruct",
   "notes": "config=vllm-bf16; 512 prompt + 128 generated tokens per request, closed-loop 32 concurrent, mean of 2 repetitions",
   "params": "7B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "| vLLM BF16 (reference) | 92.9 | 10.41 ms | 1404.7 | 17.68 ms |; Model: Qwen2.5-7B-Instruct, F16 GGUF (14.19 GiB), converted from the official HF weights; NVIDIA GeForce RTX 5090 32GB (Blackwell, sm_120)",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27050",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27050",
   "tps": "1404.7"
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 16,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-27117-nospec16",
   "model": "muse-glimmer 30B",
   "notes": "config=nospec-np16; 16 identical concurrent requests, aggregate throughput, no speculation",
   "params": "30B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K",
   "quote": "measured **~80 tok/s aggregate vs 37 tok/s** no-spec baseline at `-np 16`; Models: Muse-Glimmer-30B target (Q4_K \"kquant\" GGUF, ~17 GB, all layers on GPU); AMD Ryzen AI MAX+ 395 / Radeon 8060S iGPU (gfx1151, 96 GB unified)",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27117",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27117",
   "tps": "37"
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 16,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-27117-dflash-nmax1",
   "model": "muse-glimmer 30B",
   "notes": "config=dflash-nmax1-np16; 16 concurrent requests, aggregate, spec-type draft-dflash n-max 1, acceptance 0.83-0.92, 2.2x the no-spec baseline",
   "params": "30B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K",
   "quote": "measured **~80 tok/s aggregate vs 37 tok/s** no-spec baseline at `-np 16`; Models: Muse-Glimmer-30B target (Q4_K \"kquant\" GGUF, ~17 GB, all layers on GPU); AMD Ryzen AI MAX+ 395 / Radeon 8060S iGPU (gfx1151, 96 GB unified)",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27117",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27117",
   "tps": "80"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "RTX 5080",
   "id": "lc-27327-t1",
   "model": "KAT-Coder-V2.5",
   "notes": "config=gdn-fused-fix; run 1 of 3; server log tg after the Gated DeltaNet fused-op fix (build 10154), model partially in 16 GB VRAM",
   "params": "35B",
   "pp_tps": "779.29",
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "[39225] 0.48.339.656 I slot print_timing: id  3 | task 1 | n_decoded =    215, tg =  71.63 t/s, tg_3s =  71.63 t/s; Ryzen 7950x3d, RAM 64gb, Nvidia RTX5080 16Gb; KAT-Coder-V2.5-Dev-APEX-I-Quality",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27327",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27327",
   "tps": "71.63"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "RTX 5080",
   "id": "lc-27327-t285",
   "model": "KAT-Coder-V2.5",
   "notes": "config=gdn-fused-fix; run 2 of 3; server log tg after the Gated DeltaNet fused-op fix (build 10154), model partially in 16 GB VRAM",
   "params": "35B",
   "pp_tps": "1024.01",
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "[39225] 1.00.110.632 I slot print_timing: id  3 | task 285 | n_decoded =    211, tg =  70.29 t/s, tg_3s =  70.29 t/s; Ryzen 7950x3d, RAM 64gb, Nvidia RTX5080 16Gb; KAT-Coder-V2.5-Dev-APEX-I-Quality",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27327",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27327",
   "tps": "70.29"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "RTX 5080",
   "id": "lc-27327-t556",
   "model": "KAT-Coder-V2.5",
   "notes": "config=gdn-fused-fix; run 3 of 3; server log tg after the Gated DeltaNet fused-op fix (build 10154), model partially in 16 GB VRAM",
   "params": "35B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "[39225] 1.04.337.789 I slot print_timing: id  3 | task 556 | n_decoded =    210, tg =  69.93 t/s, tg_3s =  69.93 t/s; Ryzen 7950x3d, RAM 64gb, Nvidia RTX5080 16Gb; KAT-Coder-V2.5-Dev-APEX-I-Quality",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27327",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27327",
   "tps": "69.93"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": 50000,
   "date": "2026-08",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "lc-27420-ub256-f16",
   "model": "qwen35 dense 27B",
   "notes": "config=ubub256-kvf16; ctx 50000, MTP draft-mtp n-max 3, 4096-token filler prompt, n_predict 512; values as published with ~ prefix",
   "params": "27B",
   "pp_tps": "36",
   "provenance": "community",
   "quant": "Q5_K_P",
   "quote": "| 256    | f16/f16   | ~36      | ~9.5     | BROKEN  |; GPU: AMD Radeon AI PRO R9700 (32 GB), Vulkan backend; Model: qwen35 dense 27B, Q5_K_P quantization (~18.8 GiB)",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27420",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27420",
   "tps": "9.5"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": 50000,
   "date": "2026-08",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "lc-27420-ub1024-f16",
   "model": "qwen35 dense 27B",
   "notes": "config=ubub1024-kvf16; ctx 50000, MTP draft-mtp n-max 3, 4096-token filler prompt, n_predict 512; values as published with ~ prefix",
   "params": "27B",
   "pp_tps": "641",
   "provenance": "community",
   "quant": "Q5_K_P",
   "quote": "| 1024   | f16/f16   | ~641     | ~50      | healthy |; GPU: AMD Radeon AI PRO R9700 (32 GB), Vulkan backend; Model: qwen35 dense 27B, Q5_K_P quantization (~18.8 GiB)",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27420",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27420",
   "tps": "50"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": 50000,
   "date": "2026-08",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "lc-27420-ub256-q8",
   "model": "qwen35 dense 27B",
   "notes": "config=ubub256-kvq8; ctx 50000, MTP draft-mtp n-max 3, 4096-token filler prompt, n_predict 512; values as published with ~ prefix",
   "params": "27B",
   "pp_tps": "594",
   "provenance": "community",
   "quant": "Q5_K_P",
   "quote": "| 256    | q8_0/q8_0 | ~594     | ~53      | healthy |; GPU: AMD Radeon AI PRO R9700 (32 GB), Vulkan backend; Model: qwen35 dense 27B, Q5_K_P quantization (~18.8 GiB)",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27420",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27420",
   "tps": "53"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": 50000,
   "date": "2026-08",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "lc-27420-ub1024-q8",
   "model": "qwen35 dense 27B",
   "notes": "config=ubub1024-kvq8; ctx 50000, MTP draft-mtp n-max 3, 4096-token filler prompt, n_predict 512; values as published with ~ prefix",
   "params": "27B",
   "pp_tps": "630",
   "provenance": "community",
   "quant": "Q5_K_P",
   "quote": "| 1024   | q8_0/q8_0 | ~630     | ~60      | healthy |; GPU: AMD Radeon AI PRO R9700 (32 GB), Vulkan backend; Model: qwen35 dense 27B, Q5_K_P quantization (~18.8 GiB)",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27420",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27420",
   "tps": "60"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "2xR9700",
   "id": "lc-27544-vk-1-s",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-nmax1; single-session TG, sweep summary raw table, spec-type draft-mtp",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "| 6a | 1 | 2 | 2 | 4 | 33.72 | **42.96** | 1.27\u00d7 | 0.644 | 0.661 \u2013 0.686 |; My setup has 2xR9700; hf                     = unsloth/Qwen3.8-27B-GGUF:UD-Q4_K_XL",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27544",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27544",
   "tps": "33.72"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 2,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "2xR9700",
   "id": "lc-27544-vk-1-p",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-nmax1; parallel TG (sum of 2 sessions), sweep summary raw table",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "| 6a | 1 | 2 | 2 | 4 | 33.72 | **42.96** | 1.27\u00d7 | 0.644 | 0.661 \u2013 0.686 |; My setup has 2xR9700; hf                     = unsloth/Qwen3.8-27B-GGUF:UD-Q4_K_XL",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27544",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27544",
   "tps": "42.96"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "2xR9700",
   "id": "lc-27544-vk-r1-s",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-nmax1; single-session TG, sweep summary raw table, spec-type draft-mtp",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "| 1 | 1 | 2 | 3 | 6 | 35.11 | **52.65** | 1.50\u00d7 | 0.699 | 0.656 \u2013 0.780 |; My setup has 2xR9700; hf                     = unsloth/Qwen3.8-27B-GGUF:UD-Q4_K_XL",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27544",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27544",
   "tps": "35.11"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 3,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "2xR9700",
   "id": "lc-27544-vk-r1-p",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-nmax1; parallel TG (sum of 3 sessions), sweep summary raw table",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "| 1 | 1 | 2 | 3 | 6 | 35.11 | **52.65** | 1.50\u00d7 | 0.699 | 0.656 \u2013 0.780 |; My setup has 2xR9700; hf                     = unsloth/Qwen3.8-27B-GGUF:UD-Q4_K_XL",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27544",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27544",
   "tps": "52.65"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "2xR9700",
   "id": "lc-27544-vk-r4-s",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-nmax1; single-session TG, sweep summary raw table, spec-type draft-mtp",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "| 4 | 1 | 2 | 3 | 6 | 34.07 | **55.78** | 1.64\u00d7 | 0.693 | 0.643 \u2013 0.675 |; My setup has 2xR9700; hf                     = unsloth/Qwen3.8-27B-GGUF:UD-Q4_K_XL",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27544",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27544",
   "tps": "34.07"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 3,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "2xR9700",
   "id": "lc-27544-vk-r4-p",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-nmax1; parallel TG (sum of 3 sessions), sweep summary raw table",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "| 4 | 1 | 2 | 3 | 6 | 34.07 | **55.78** | 1.64\u00d7 | 0.693 | 0.643 \u2013 0.675 |; My setup has 2xR9700; hf                     = unsloth/Qwen3.8-27B-GGUF:UD-Q4_K_XL",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27544",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27544",
   "tps": "55.78"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "2xR9700",
   "id": "lc-27544-vk-6b-s",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-nmax3; single-session TG, sweep summary raw table, spec-type draft-mtp",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "| **6b** | **3** | **4** | **2** | **8** | 35.42 | **41.12** | **1.16\u00d7** | 0.416 | 0.415 \u2013 0.425 |; My setup has 2xR9700; hf                     = unsloth/Qwen3.8-27B-GGUF:UD-Q4_K_XL",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27544",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27544",
   "tps": "35.42"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 2,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "2xR9700",
   "id": "lc-27544-vk-6b-p",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-nmax3; parallel TG (sum of 2 sessions), sweep summary raw table",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "| **6b** | **3** | **4** | **2** | **8** | 35.42 | **41.12** | **1.16\u00d7** | 0.416 | 0.415 \u2013 0.425 |; My setup has 2xR9700; hf                     = unsloth/Qwen3.8-27B-GGUF:UD-Q4_K_XL",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27544",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27544",
   "tps": "41.12"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "2xR9700",
   "id": "lc-27544-vk-r5-s",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-nmax1; single-session TG, sweep summary raw table, spec-type draft-mtp",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "| **5** | **1** | **2** | **4** | **8** | 33.87 | **55.58** | **1.64\u00d7** | 0.632 | 0.632 \u2013 0.683 |; My setup has 2xR9700; hf                     = unsloth/Qwen3.8-27B-GGUF:UD-Q4_K_XL",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27544",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27544",
   "tps": "33.87"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 4,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "2xR9700",
   "id": "lc-27544-vk-r5-p",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-nmax1; parallel TG (sum of 4 sessions), sweep summary raw table",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "| **5** | **1** | **2** | **4** | **8** | 33.87 | **55.58** | **1.64\u00d7** | 0.632 | 0.632 \u2013 0.683 |; My setup has 2xR9700; hf                     = unsloth/Qwen3.8-27B-GGUF:UD-Q4_K_XL",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27544",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27544",
   "tps": "55.58"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "2xR9700",
   "id": "lc-27544-vk-r3-s",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-nmax2; single-session TG, sweep summary raw table, spec-type draft-mtp",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "| 3 | 2 | 3 | 3 | 9 | 35.95 | **38.52** | 1.07\u00d7 | 0.529 | 0.530 \u2013 0.560 |; My setup has 2xR9700; hf                     = unsloth/Qwen3.8-27B-GGUF:UD-Q4_K_XL",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27544",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27544",
   "tps": "35.95"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 3,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "2xR9700",
   "id": "lc-27544-vk-r3-p",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-nmax2; parallel TG (sum of 3 sessions), sweep summary raw table",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "| 3 | 2 | 3 | 3 | 9 | 35.95 | **38.52** | 1.07\u00d7 | 0.529 | 0.530 \u2013 0.560 |; My setup has 2xR9700; hf                     = unsloth/Qwen3.8-27B-GGUF:UD-Q4_K_XL",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27544",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27544",
   "tps": "38.52"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "2xR9700",
   "id": "lc-27544-vk-r2-s",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-nmax3; single-session TG, sweep summary raw table, spec-type draft-mtp",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "| 2 | 3 | 4 | 3 | 12 | 37.52 | **40.78** | 1.09\u00d7 | 0.458 | 0.393 \u2013 0.500 |; My setup has 2xR9700; hf                     = unsloth/Qwen3.8-27B-GGUF:UD-Q4_K_XL",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27544",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27544",
   "tps": "37.52"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 3,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "2xR9700",
   "id": "lc-27544-vk-r2-p",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-nmax3; parallel TG (sum of 3 sessions), sweep summary raw table",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "| 2 | 3 | 4 | 3 | 12 | 37.52 | **40.78** | 1.09\u00d7 | 0.458 | 0.393 \u2013 0.500 |; My setup has 2xR9700; hf                     = unsloth/Qwen3.8-27B-GGUF:UD-Q4_K_XL",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27544",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27544",
   "tps": "40.78"
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "2xR9700",
   "id": "lc-27544-r1s",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-nmax1; aggregate TG, single session, ROCm 2xR9700",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Aggregate TG: single **32.40** \u2192 dual **44.56** (1.38\u00d7) \u2192 triple **50.49** (1.56\u00d7).; My setup has 2xR9700; hf                     = unsloth/Qwen3.8-27B-GGUF:UD-Q4_K_XL",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27544",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27544",
   "tps": "32.40"
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 2,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "2xR9700",
   "id": "lc-27544-r1d",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-nmax1; aggregate TG, 2 parallel sessions, ROCm 2xR9700",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Aggregate TG: single **32.40** \u2192 dual **44.56** (1.38\u00d7) \u2192 triple **50.49** (1.56\u00d7).; My setup has 2xR9700; hf                     = unsloth/Qwen3.8-27B-GGUF:UD-Q4_K_XL",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27544",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27544",
   "tps": "44.56"
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 3,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "2xR9700",
   "id": "lc-27544-r1t",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-nmax1; aggregate TG, 3 parallel sessions, ROCm 2xR9700",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Aggregate TG: single **32.40** \u2192 dual **44.56** (1.38\u00d7) \u2192 triple **50.49** (1.56\u00d7).; My setup has 2xR9700; hf                     = unsloth/Qwen3.8-27B-GGUF:UD-Q4_K_XL",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27544",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27544",
   "tps": "50.49"
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "2xR9700",
   "id": "lc-27544-r2s",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-nmax3; aggregate TG, single session, ROCm 2xR9700",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Aggregate TG: single **32.84** \u2192 dual **35.17** (1.07\u00d7) \u2192 triple **71.42** (2.18\u00d7).; My setup has 2xR9700; hf                     = unsloth/Qwen3.8-27B-GGUF:UD-Q4_K_XL",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27544",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27544",
   "tps": "32.84"
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 2,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "2xR9700",
   "id": "lc-27544-r2d",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-nmax3; aggregate TG, 2 parallel sessions, ROCm 2xR9700",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Aggregate TG: single **32.84** \u2192 dual **35.17** (1.07\u00d7) \u2192 triple **71.42** (2.18\u00d7).; My setup has 2xR9700; hf                     = unsloth/Qwen3.8-27B-GGUF:UD-Q4_K_XL",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27544",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27544",
   "tps": "35.17"
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 3,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "2xR9700",
   "id": "lc-27544-r2t",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-nmax3; aggregate TG, 3 parallel sessions, ROCm 2xR9700",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Aggregate TG: single **32.84** \u2192 dual **35.17** (1.07\u00d7) \u2192 triple **71.42** (2.18\u00d7).; My setup has 2xR9700; hf                     = unsloth/Qwen3.8-27B-GGUF:UD-Q4_K_XL",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27544",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27544",
   "tps": "71.42"
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-27572-d620",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-np4-pred620; single-stream tg before the sched-sync workaround (racing build), temp 0, 620-token prompt",
   "params": "27B",
   "pp_tps": "291.6",
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "| 620 | 291.6 \u2192 290.2 | 16.3 \u2192 15.5 |; AMD Radeon 8060S (Strix Halo, gfx1151), 256 GB unified; unsloth/Qwen3.8-27B-GGUF",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27572",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27572",
   "tps": "16.3"
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-27572-d4603",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-np4-pred4603; single-stream tg before the sched-sync workaround (racing build), temp 0, 4603-token prompt",
   "params": "27B",
   "pp_tps": "339.9",
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "| 4603 | 339.9 \u2192 338.4 | 17.7 \u2192 16.6 |; AMD Radeon 8060S (Strix Halo, gfx1151), 256 GB unified; unsloth/Qwen3.8-27B-GGUF",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27572",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27572",
   "tps": "17.7"
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-27572-d9154",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-np4-pred9154; single-stream tg before the sched-sync workaround (racing build), temp 0, 9154-token prompt",
   "params": "27B",
   "pp_tps": "552.1",
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "| 9154 | 552.1 \u2192 547.1 | 25.3 \u2192 ~15 |; AMD Radeon 8060S (Strix Halo, gfx1151), 256 GB unified; unsloth/Qwen3.8-27B-GGUF",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27572",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27572",
   "tps": "25.3"
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-27572-d18256",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-np4-pred18256; single-stream tg before the sched-sync workaround (racing build), temp 0, 18256-token prompt",
   "params": "27B",
   "pp_tps": "507.6",
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "| 18256 | 507.6 \u2192 503.9 | 24.9 \u2192 ~14 |; AMD Radeon 8060S (Strix Halo, gfx1151), 256 GB unified; unsloth/Qwen3.8-27B-GGUF",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27572",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27572",
   "tps": "24.9"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "RTX 4080 SUPER",
   "id": "lc-27623-p45574",
   "model": "Qwen3.8-27B",
   "notes": "config=position45574; position-sweep decode at KV position 45574, q8 KV, no flash attention, WSL2 llama-server, as published (last row: 1.4 or timeout)",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q3_K_XL",
   "quote": "| 45,574 | 35.6 |; GPU**: RTX 4080 SUPER 16GB, driver 591.86 (CUDA 13.1), nvcc 13.1, sm_89; unsloth/Qwen3.8-27B-GGUF",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27623",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27623",
   "tps": "35.6"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "RTX 4080 SUPER",
   "id": "lc-27623-p60309",
   "model": "Qwen3.8-27B",
   "notes": "config=position60309; position-sweep decode at KV position 60309, q8 KV, no flash attention, WSL2 llama-server, as published (last row: 1.4 or timeout)",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q3_K_XL",
   "quote": "| 60,309 | 33.9 |; GPU**: RTX 4080 SUPER 16GB, driver 591.86 (CUDA 13.1), nvcc 13.1, sm_89; unsloth/Qwen3.8-27B-GGUF",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27623",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27623",
   "tps": "33.9"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "RTX 4080 SUPER",
   "id": "lc-27623-p68642",
   "model": "Qwen3.8-27B",
   "notes": "config=position68642; position-sweep decode at KV position 68642, q8 KV, no flash attention, WSL2 llama-server, as published (last row: 1.4 or timeout)",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q3_K_XL",
   "quote": "| 68,642 | 33.1 |; GPU**: RTX 4080 SUPER 16GB, driver 591.86 (CUDA 13.1), nvcc 13.1, sm_89; unsloth/Qwen3.8-27B-GGUF",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27623",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27623",
   "tps": "33.1"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "RTX 4080 SUPER",
   "id": "lc-27623-p91077",
   "model": "Qwen3.8-27B",
   "notes": "config=position91077; position-sweep decode at KV position 91077, q8 KV, no flash attention, WSL2 llama-server, as published (last row: 1.4 or timeout)",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q3_K_XL",
   "quote": "| 91,077 | **1.4 (or timeout)** |; GPU**: RTX 4080 SUPER 16GB, driver 591.86 (CUDA 13.1), nvcc 13.1, sm_89; unsloth/Qwen3.8-27B-GGUF",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27623",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27623",
   "tps": "1.4"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "4\u00d7 Tesla P40",
   "id": "lc-27980-r1",
   "model": "Qwen3.8-Flash-Next",
   "notes": "config=mmap-embd-cpu; decode, 150 tokens, temperature 0.7, cache_prompt false, averaged over 3-4 runs",
   "params": "125B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q3_K_XL",
   "quote": "| 1 | `-ot \"per_layer_token_embd=CPU\"` (mmap) | **18.0 tok/s** |; 4\u00d7 Tesla P40 (sm_61, 24 GB each); unsloth/Qwen3.8-Flash-Next-GGUF",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27980",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27980",
   "tps": "18.0"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "4\u00d7 Tesla P40",
   "id": "lc-27980-r2",
   "model": "Qwen3.8-Flash-Next",
   "notes": "config=load-none-embd-cpu; decode, 150 tokens, temperature 0.7, cache_prompt false, averaged over 3-4 runs",
   "params": "125B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q3_K_XL",
   "quote": "| 2 | `-ot \"per_layer_token_embd=CPU\" --load-mode none` (+28 GB host RAM) | **18.0 tok/s** |; 4\u00d7 Tesla P40 (sm_61, 24 GB each); unsloth/Qwen3.8-Flash-Next-GGUF",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27980",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27980",
   "tps": "18.0"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "4\u00d7 Tesla P40",
   "id": "lc-27980-r3",
   "model": "Qwen3.8-Flash-Next",
   "notes": "config=all-vram; decode, 150 tokens, temperature 0.7, cache_prompt false, averaged over 3-4 runs",
   "params": "125B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q3_K_XL",
   "quote": "| 3 | no offload at all, all weights in VRAM, `-c 16384 --parallel 1` | **18.1 tok/s** |; 4\u00d7 Tesla P40 (sm_61, 24 GB each); unsloth/Qwen3.8-Flash-Next-GGUF",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27980",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27980",
   "tps": "18.1"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "4\u00d7 Tesla P40",
   "id": "lc-27980-r4",
   "model": "Qwen3.8-Flash-Next",
   "notes": "config=force-mmq; decode, 150 tokens, temperature 0.7, cache_prompt false, averaged over 3-4 runs",
   "params": "125B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q3_K_XL",
   "quote": "| 4 | `GGML_CUDA_FORCE_MMQ=1` | **18.1 tok/s** |; 4\u00d7 Tesla P40 (sm_61, 24 GB each); unsloth/Qwen3.8-Flash-Next-GGUF",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27980",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27980",
   "tps": "18.1"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "4\u00d7 Tesla P40",
   "id": "lc-27980-r5",
   "model": "qwen3.5moe 122B-A10B",
   "notes": "config=ref-moe-same-gpus; reference decode of the 10B-active MoE on the same 4 GPUs and build, 150 tokens, temperature 0.7",
   "params": "122B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q3_K_XL",
   "quote": "| \u2014 | **qwen3.5moe 122B-A10B, same GPUs/build** | **29.5 tok/s** |; 4\u00d7 Tesla P40 (sm_61, 24 GB each); unsloth/Qwen3.8-Flash-Next-GGUF",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27980",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27980",
   "tps": "29.5"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-28218-m1",
   "model": "Qwen3.8-27B",
   "notes": "config=build-b10734-mtp; official b10734, spec-type draft-mtp, no tensor split, baseline; ~60 greedy tokens, server timings API",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ3_S",
   "quote": "| 1 | official b10734 | MTP, no `-ts` | **38.8 tok/s** | baseline |; Windows 11 26200, RTX 5060 Ti 16 GB (sm_120), CUDA driver 13.2, 24-thread CPU; Model: Qwen3.8-27B-UD-IQ3_S",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28218",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28218",
   "tps": "38.8"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-28218-m2",
   "model": "Qwen3.8-27B",
   "notes": "config=build-b10734-mtp-ts1; official b10734, draft-mtp + tensor split 1, the 500 ms per token cliff; ~60 greedy tokens, server timings API",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ3_S",
   "quote": "| 2 | official b10734 | MTP + `-ts 1` | **1.67 tok/s** | cliff |; Windows 11 26200, RTX 5060 Ti 16 GB (sm_120), CUDA driver 13.2, 24-thread CPU; Model: Qwen3.8-27B-UD-IQ3_S",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28218",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28218",
   "tps": "1.67"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-28218-m3",
   "model": "Qwen3.8-27B",
   "notes": "config=build-msvc-mtp; local MSVC 19.44 build of the b10734 commit, draft-mtp, cliff; ~60 greedy tokens, server timings API",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ3_S",
   "quote": "| 3 | local MSVC build | MTP | **2.01 tok/s** | cliff |; Windows 11 26200, RTX 5060 Ti 16 GB (sm_120), CUDA driver 13.2, 24-thread CPU; Model: Qwen3.8-27B-UD-IQ3_S",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28218",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28218",
   "tps": "2.01"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-28218-m4",
   "model": "Qwen3.8-27B",
   "notes": "config=build-msvc-dflash; local MSVC build, draft-dflash (DFlash2), cliff; ~60 greedy tokens, server timings API",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ3_S",
   "quote": "| 4 | local MSVC build | draft-dflash (DFlash2) | **2.01 tok/s** | cliff |; Windows 11 26200, RTX 5060 Ti 16 GB (sm_120), CUDA driver 13.2, 24-thread CPU; Model: Qwen3.8-27B-UD-IQ3_S",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28218",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28218",
   "tps": "2.01"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-28218-m5",
   "model": "Qwen3.8-27B",
   "notes": "config=build-msvc-nospec; local MSVC build, no speculative decoding, normal speed; ~60 greedy tokens, server timings API",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ3_S",
   "quote": "| 5 | local MSVC build | no speculative decoding | **26.55 tok/s** | normal |; Windows 11 26200, RTX 5060 Ti 16 GB (sm_120), CUDA driver 13.2, 24-thread CPU; Model: Qwen3.8-27B-UD-IQ3_S",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28218",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28218",
   "tps": "26.55"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 5080",
   "id": "lc-28274-cuda",
   "model": "Qwen3.8-27B",
   "notes": "config=build-b10766; CUDA 13.3 build on driver 610.88, around 16 t/s as published",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "CUDA gets around 16t/s\nwhile Vulkan started with 40t/s and settled down on 35.59 t/s; -m \"Qwen3.8-27B-ZB4.00-MIN-v5-IQ4_XS.gguf\"; On NVIDIA driver 610.88 with RTX 5080",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28274",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28274",
   "tps": "16"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 5080",
   "id": "lc-28274-vulkan",
   "model": "Qwen3.8-27B",
   "notes": "config=build-b10766; Vulkan build, settled speed as published (started at 40 t/s)",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "CUDA gets around 16t/s\nwhile Vulkan started with 40t/s and settled down on 35.59 t/s; -m \"Qwen3.8-27B-ZB4.00-MIN-v5-IQ4_XS.gguf\"; On NVIDIA driver 610.88 with RTX 5080",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28274",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28274",
   "tps": "35.59"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-28721-d1k-vk",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp3-depthd1k; deep-context decode at depth 1k, q8_0 KV, b10883 + Mesa 26.2.2 ANV, MTP3 draft Q8_0 n-max 3",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "| ~1k   | 35.6 t/s | 33.4 t/s | 38.5 t/s | 1.8 s  | 1.8 s  |; GPU:** Intel Arc Pro B70 (32 GB, BMG-G31), single GPU, full offload; Model:** `Qwen3.8-27B-Q4_K_M` (dense, 16.4 GB weights), MTP3 draft (`Q8_0`, n-max 3, p-min 0.10)",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28721",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28721",
   "tps": "35.6"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-28721-d1k-vkf16",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp3-f16kv-depthd1k; same scan with f16 KV control, MTP3 draft",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "| ~1k   | 35.6 t/s | 33.4 t/s | 38.5 t/s | 1.8 s  | 1.8 s  |; GPU:** Intel Arc Pro B70 (32 GB, BMG-G31), single GPU, full offload; Model:** `Qwen3.8-27B-Q4_K_M` (dense, 16.4 GB weights), MTP3 draft (`Q8_0`, n-max 3, p-min 0.10)",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28721",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28721",
   "tps": "33.4"
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-28721-d1k-sycl",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp3-depthd1k; same scan on SYCL v0.4.0 / c26.31, MTP3 draft",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "| ~1k   | 35.6 t/s | 33.4 t/s | 38.5 t/s | 1.8 s  | 1.8 s  |; GPU:** Intel Arc Pro B70 (32 GB, BMG-G31), single GPU, full offload; Model:** `Qwen3.8-27B-Q4_K_M` (dense, 16.4 GB weights), MTP3 draft (`Q8_0`, n-max 3, p-min 0.10)",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28721",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28721",
   "tps": "38.5"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-28721-d15k-vk",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp3-depthd15k; deep-context decode at depth 15k, q8_0 KV, b10883 + Mesa 26.2.2 ANV, MTP3 draft Q8_0 n-max 3",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "| ~15k  | 14.9 t/s | 13.3 t/s | 31.7 t/s | 43.2 s | 23.2 s |; GPU:** Intel Arc Pro B70 (32 GB, BMG-G31), single GPU, full offload; Model:** `Qwen3.8-27B-Q4_K_M` (dense, 16.4 GB weights), MTP3 draft (`Q8_0`, n-max 3, p-min 0.10)",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28721",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28721",
   "tps": "14.9"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-28721-d15k-vkf16",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp3-f16kv-depthd15k; same scan with f16 KV control, MTP3 draft",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "| ~15k  | 14.9 t/s | 13.3 t/s | 31.7 t/s | 43.2 s | 23.2 s |; GPU:** Intel Arc Pro B70 (32 GB, BMG-G31), single GPU, full offload; Model:** `Qwen3.8-27B-Q4_K_M` (dense, 16.4 GB weights), MTP3 draft (`Q8_0`, n-max 3, p-min 0.10)",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28721",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28721",
   "tps": "13.3"
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-28721-d15k-sycl",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp3-depthd15k; same scan on SYCL v0.4.0 / c26.31, MTP3 draft",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "| ~15k  | 14.9 t/s | 13.3 t/s | 31.7 t/s | 43.2 s | 23.2 s |; GPU:** Intel Arc Pro B70 (32 GB, BMG-G31), single GPU, full offload; Model:** `Qwen3.8-27B-Q4_K_M` (dense, 16.4 GB weights), MTP3 draft (`Q8_0`, n-max 3, p-min 0.10)",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28721",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28721",
   "tps": "31.7"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-28721-d32k-vk",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp3-depthd32k; deep-context decode at depth 32k, q8_0 KV, b10883 + Mesa 26.2.2 ANV, MTP3 draft Q8_0 n-max 3",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "| ~32k  | 8.4 t/s  | 8.2 t/s  | 22.8 t/s | 94.1 s | 28.5 s |; GPU:** Intel Arc Pro B70 (32 GB, BMG-G31), single GPU, full offload; Model:** `Qwen3.8-27B-Q4_K_M` (dense, 16.4 GB weights), MTP3 draft (`Q8_0`, n-max 3, p-min 0.10)",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28721",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28721",
   "tps": "8.4"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-28721-d32k-vkf16",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp3-f16kv-depthd32k; same scan with f16 KV control, MTP3 draft",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "| ~32k  | 8.4 t/s  | 8.2 t/s  | 22.8 t/s | 94.1 s | 28.5 s |; GPU:** Intel Arc Pro B70 (32 GB, BMG-G31), single GPU, full offload; Model:** `Qwen3.8-27B-Q4_K_M` (dense, 16.4 GB weights), MTP3 draft (`Q8_0`, n-max 3, p-min 0.10)",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28721",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28721",
   "tps": "8.2"
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-28721-d32k-sycl",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp3-depthd32k; same scan on SYCL v0.4.0 / c26.31, MTP3 draft",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "| ~32k  | 8.4 t/s  | 8.2 t/s  | 22.8 t/s | 94.1 s | 28.5 s |; GPU:** Intel Arc Pro B70 (32 GB, BMG-G31), single GPU, full offload; Model:** `Qwen3.8-27B-Q4_K_M` (dense, 16.4 GB weights), MTP3 draft (`Q8_0`, n-max 3, p-min 0.10)",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28721",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28721",
   "tps": "22.8"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-28721-d64k-vk",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp3-depthd64k; deep-context decode at depth 64k, q8_0 KV, b10883 + Mesa 26.2.2 ANV, MTP3 draft Q8_0 n-max 3",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "| ~64k  | 4.5 t/s  | 4.3 t/s  | 15.8 t/s | 329.0 s| 64.9 s |; GPU:** Intel Arc Pro B70 (32 GB, BMG-G31), single GPU, full offload; Model:** `Qwen3.8-27B-Q4_K_M` (dense, 16.4 GB weights), MTP3 draft (`Q8_0`, n-max 3, p-min 0.10)",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28721",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28721",
   "tps": "4.5"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-28721-d64k-vkf16",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp3-f16kv-depthd64k; same scan with f16 KV control, MTP3 draft",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "| ~64k  | 4.5 t/s  | 4.3 t/s  | 15.8 t/s | 329.0 s| 64.9 s |; GPU:** Intel Arc Pro B70 (32 GB, BMG-G31), single GPU, full offload; Model:** `Qwen3.8-27B-Q4_K_M` (dense, 16.4 GB weights), MTP3 draft (`Q8_0`, n-max 3, p-min 0.10)",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28721",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28721",
   "tps": "4.3"
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-28721-d64k-sycl",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp3-depthd64k; same scan on SYCL v0.4.0 / c26.31, MTP3 draft",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "| ~64k  | 4.5 t/s  | 4.3 t/s  | 15.8 t/s | 329.0 s| 64.9 s |; GPU:** Intel Arc Pro B70 (32 GB, BMG-G31), single GPU, full offload; Model:** `Qwen3.8-27B-Q4_K_M` (dense, 16.4 GB weights), MTP3 draft (`Q8_0`, n-max 3, p-min 0.10)",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28721",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28721",
   "tps": "15.8"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-28721-d1k-vk-nomtp",
   "model": "Qwen3.8-27B",
   "notes": "config=nomtp-depthd1k; no-draft decode control at depth 1k, q8_0 KV",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "| ~1k  | 21.4 t/s | 35.6 t/s | **+66%** | 22.5 t/s | 38.5 t/s | **+71%** |; GPU:** Intel Arc Pro B70 (32 GB, BMG-G31), single GPU, full offload; Model:** `Qwen3.8-27B-Q4_K_M` (dense, 16.4 GB weights), MTP3 draft (`Q8_0`, n-max 3, p-min 0.10)",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28721",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28721",
   "tps": "21.4"
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-28721-d1k-sycl-nomtp",
   "model": "Qwen3.8-27B",
   "notes": "config=nomtp-depthd1k; no-draft decode control at depth 1k, SYCL v0.4.0",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "| ~1k  | 21.4 t/s | 35.6 t/s | **+66%** | 22.5 t/s | 38.5 t/s | **+71%** |; GPU:** Intel Arc Pro B70 (32 GB, BMG-G31), single GPU, full offload; Model:** `Qwen3.8-27B-Q4_K_M` (dense, 16.4 GB weights), MTP3 draft (`Q8_0`, n-max 3, p-min 0.10)",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28721",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28721",
   "tps": "22.5"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-28721-d16k-vk-nomtp",
   "model": "Qwen3.8-27B",
   "notes": "config=nomtp-depthd16k; no-draft decode control at depth 16k, q8_0 KV",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "| ~16k | 13.7 t/s | 14.9 t/s | +9%    | 18.8 t/s | 31.7 t/s | **+69%** |; GPU:** Intel Arc Pro B70 (32 GB, BMG-G31), single GPU, full offload; Model:** `Qwen3.8-27B-Q4_K_M` (dense, 16.4 GB weights), MTP3 draft (`Q8_0`, n-max 3, p-min 0.10)",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28721",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28721",
   "tps": "13.7"
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-28721-d16k-sycl-nomtp",
   "model": "Qwen3.8-27B",
   "notes": "config=nomtp-depthd16k; no-draft decode control at depth 16k, SYCL v0.4.0",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "| ~16k | 13.7 t/s | 14.9 t/s | +9%    | 18.8 t/s | 31.7 t/s | **+69%** |; GPU:** Intel Arc Pro B70 (32 GB, BMG-G31), single GPU, full offload; Model:** `Qwen3.8-27B-Q4_K_M` (dense, 16.4 GB weights), MTP3 draft (`Q8_0`, n-max 3, p-min 0.10)",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28721",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28721",
   "tps": "18.8"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-28721-d32k-vk-nomtp",
   "model": "Qwen3.8-27B",
   "notes": "config=nomtp-depthd32k; no-draft decode control at depth 32k, q8_0 KV",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "| ~32k | 9.8 t/s  | 8.4 t/s  | \u221214%   | 15.9 t/s | 22.8 t/s | **+43%** |; GPU:** Intel Arc Pro B70 (32 GB, BMG-G31), single GPU, full offload; Model:** `Qwen3.8-27B-Q4_K_M` (dense, 16.4 GB weights), MTP3 draft (`Q8_0`, n-max 3, p-min 0.10)",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28721",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28721",
   "tps": "9.8"
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-28721-d32k-sycl-nomtp",
   "model": "Qwen3.8-27B",
   "notes": "config=nomtp-depthd32k; no-draft decode control at depth 32k, SYCL v0.4.0",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "| ~32k | 9.8 t/s  | 8.4 t/s  | \u221214%   | 15.9 t/s | 22.8 t/s | **+43%** |; GPU:** Intel Arc Pro B70 (32 GB, BMG-G31), single GPU, full offload; Model:** `Qwen3.8-27B-Q4_K_M` (dense, 16.4 GB weights), MTP3 draft (`Q8_0`, n-max 3, p-min 0.10)",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28721",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28721",
   "tps": "15.9"
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "lc-28867-pre",
   "model": "Qwen3.8-27B",
   "notes": "config=fa-9113cc188; speculative decode (draft-mtp n-max 3), median of 5 x 256-token completions, temperature 0, 2 warm-ups discarded",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "| `9113cc188` (before #28102) | **51.70** (6 runs, 51.45\u201352.05) |; Radeon AI PRO R9700 (Navi 48, gfx1201, RDNA4), single GPU, 32 GiB; Qwen3.8-27B-UD-Q4_K_XL",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28867",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28867",
   "tps": "51.70"
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "lc-28867-master",
   "model": "Qwen3.8-27B",
   "notes": "config=fa-master-5f436dddb; speculative decode (draft-mtp n-max 3), median of 5 x 256-token completions, temperature 0, 2 warm-ups discarded",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "| master `5f436dddb` | **41.37** |; Radeon AI PRO R9700 (Navi 48, gfx1201, RDNA4), single GPU, 32 GiB; Qwen3.8-27B-UD-Q4_K_XL",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28867",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28867",
   "tps": "41.37"
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "lc-28867-t32",
   "model": "Qwen3.8-27B",
   "notes": "config=fa-thr32; speculative decode (draft-mtp n-max 3), median of 5 x 256-token completions, temperature 0, 2 warm-ups discarded",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "| master, threshold 32 | 40.96 |; Radeon AI PRO R9700 (Navi 48, gfx1201, RDNA4), single GPU, 32 GiB; Qwen3.8-27B-UD-Q4_K_XL",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28867",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28867",
   "tps": "40.96"
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "lc-28867-t64",
   "model": "Qwen3.8-27B",
   "notes": "config=fa-thr64; speculative decode (draft-mtp n-max 3), median of 5 x 256-token completions, temperature 0, 2 warm-ups discarded",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "| master, threshold **64** | **51.62** |; Radeon AI PRO R9700 (Navi 48, gfx1201, RDNA4), single GPU, 32 GiB; Qwen3.8-27B-UD-Q4_K_XL",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28867",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28867",
   "tps": "51.62"
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "lc-28867-t128",
   "model": "Qwen3.8-27B",
   "notes": "config=fa-thr128; speculative decode (draft-mtp n-max 3), median of 5 x 256-token completions, temperature 0, 2 warm-ups discarded",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "| master, threshold 128 | 51.69 |; Radeon AI PRO R9700 (Navi 48, gfx1201, RDNA4), single GPU, 32 GiB; Qwen3.8-27B-UD-Q4_K_XL",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28867",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28867",
   "tps": "51.69"
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "lc-28867-t256",
   "model": "Qwen3.8-27B",
   "notes": "config=fa-thr256; speculative decode (draft-mtp n-max 3), median of 5 x 256-token completions, temperature 0, 2 warm-ups discarded",
   "params": "27B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "| master, threshold 256 | 51.37 |; Radeon AI PRO R9700 (Navi 48, gfx1201, RDNA4), single GPU, 32 GiB; Qwen3.8-27B-UD-Q4_K_XL",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28867",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28867",
   "tps": "51.37"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 2070 Super",
   "id": "lc-29168-b10750-mtp",
   "model": "gemma-4-26B-A4B-it",
   "notes": "config=build-10750-mtp; build-10750, greedy, output byte-identical to no-draft, -ngl 31 --n-cpu-moe 21 partial offload, QAT UD-Q4_K_XL + Q4_0 MTP sidecar",
   "params": "26B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "tok/s **with** MTP draft | **46.35** | 35.36 \u2013 36.73 |; RTX 2070 SUPER (sm_75, 8 GB); Partial offload: `-ngl 31 --n-cpu-moe 21`; gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #29168",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29168",
   "tps": "46.35"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 2070 Super",
   "id": "lc-29168-b10750-plain",
   "model": "gemma-4-26B-A4B-it",
   "notes": "config=build-10750-plain; build-10750, no draft, -ngl 31 --n-cpu-moe 21 partial offload, QAT UD-Q4_K_XL + Q4_0 MTP sidecar",
   "params": "26B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "tok/s **without** draft | 38.10 | 37.02 \u2013 38.33 |; RTX 2070 SUPER (sm_75, 8 GB); Partial offload: `-ngl 31 --n-cpu-moe 21`; gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #29168",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29168",
   "tps": "38.10"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 2070 Super",
   "id": "lc-29168-b10964-mtp",
   "model": "gemma-4-26B-A4B-it",
   "notes": "config=build-10964-mtp; build-10964, the v0.4.1 designated build, -ngl 31 --n-cpu-moe 21 partial offload, QAT UD-Q4_K_XL + Q4_0 MTP sidecar",
   "params": "26B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "tok/s **with** MTP draft | **46.35** | 35.36 \u2013 36.73 |; RTX 2070 SUPER (sm_75, 8 GB); Partial offload: `-ngl 31 --n-cpu-moe 21`; gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #29168",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29168",
   "tps": "35.36"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 2070 Super",
   "id": "lc-29168-b10964-plain",
   "model": "gemma-4-26B-A4B-it",
   "notes": "config=build-10964-plain; build-10964, no draft, -ngl 31 --n-cpu-moe 21 partial offload, QAT UD-Q4_K_XL + Q4_0 MTP sidecar",
   "params": "26B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "tok/s **without** draft | 38.10 | 37.02 \u2013 38.33 |; RTX 2070 SUPER (sm_75, 8 GB); Partial offload: `-ngl 31 --n-cpu-moe 21`; gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #29168",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29168",
   "tps": "37.02"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 2070 Super",
   "id": "lc-29168-b11057-mtp",
   "model": "gemma-4-26B-A4B-it",
   "notes": "config=build-11057-mtp; build-11057, still present at head, -ngl 31 --n-cpu-moe 21 partial offload, QAT UD-Q4_K_XL + Q4_0 MTP sidecar",
   "params": "26B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "tok/s **with** MTP draft | **46.35** | 35.36 \u2013 36.73 |; RTX 2070 SUPER (sm_75, 8 GB); Partial offload: `-ngl 31 --n-cpu-moe 21`; gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #29168",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29168",
   "tps": "36.73"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 2070 Super",
   "id": "lc-29168-b11057-plain",
   "model": "gemma-4-26B-A4B-it",
   "notes": "config=build-11057-plain; build-11057, no draft, -ngl 31 --n-cpu-moe 21 partial offload, QAT UD-Q4_K_XL + Q4_0 MTP sidecar",
   "params": "26B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "tok/s **without** draft | 38.10 | 37.02 \u2013 38.33 |; RTX 2070 SUPER (sm_75, 8 GB); Partial offload: `-ngl 31 --n-cpu-moe 21`; gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #29168",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29168",
   "tps": "38.33"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-27464-gb10-nemo-npl1-cur",
   "model": "Nemotron 3 Nano 30B-A3B",
   "notes": "config=master-78ec4c3; TG t/s, llama-batched-bench -npp 64 -ntg 128, npl=1 concurrent streams, current master 78ec4c3, GEMV dispatch",
   "params": "30B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Nemotron 3 Nano 30B-A3B (NVFP4):; | 1   |  67.5 |  67.7 | 0% |; The numbers are on master `78ec4c3` with `llama-batched-bench -npp 64 -ntg 128 -ngl 1,8,32,256`.; on GB10 (sm121) and RTX PRO 6000 (sm120). The results are shown below as text-generation tokens/s.",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27464",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27464",
   "tps": "67.5"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-27464-gb10-nemo-npl1-pat",
   "model": "Nemotron 3 Nano 30B-A3B",
   "notes": "config=patched-mamba2-flat; TG t/s, llama-batched-bench -npp 64 -ntg 128, npl=1 concurrent streams, proposed mamba2 flat-2d fix, GEMM dispatch",
   "params": "30B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Nemotron 3 Nano 30B-A3B (NVFP4):; | 1   |  67.5 |  67.7 | 0% |; The numbers are on master `78ec4c3` with `llama-batched-bench -npp 64 -ntg 128 -ngl 1,8,32,256`.; on GB10 (sm121) and RTX PRO 6000 (sm120). The results are shown below as text-generation tokens/s.",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27464",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27464",
   "tps": "67.7"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 8,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-27464-gb10-nemo-npl8-cur",
   "model": "Nemotron 3 Nano 30B-A3B",
   "notes": "config=master-78ec4c3; TG t/s, llama-batched-bench -npp 64 -ntg 128, npl=8 concurrent streams, current master 78ec4c3, GEMV dispatch",
   "params": "30B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Nemotron 3 Nano 30B-A3B (NVFP4):; | 8   | 152.4 | 202.6 | +33% |; The numbers are on master `78ec4c3` with `llama-batched-bench -npp 64 -ntg 128 -ngl 1,8,32,256`.; on GB10 (sm121) and RTX PRO 6000 (sm120). The results are shown below as text-generation tokens/s.",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27464",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27464",
   "tps": "152.4"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 8,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-27464-gb10-nemo-npl8-pat",
   "model": "Nemotron 3 Nano 30B-A3B",
   "notes": "config=patched-mamba2-flat; TG t/s, llama-batched-bench -npp 64 -ntg 128, npl=8 concurrent streams, proposed mamba2 flat-2d fix, GEMM dispatch",
   "params": "30B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Nemotron 3 Nano 30B-A3B (NVFP4):; | 8   | 152.4 | 202.6 | +33% |; The numbers are on master `78ec4c3` with `llama-batched-bench -npp 64 -ntg 128 -ngl 1,8,32,256`.; on GB10 (sm121) and RTX PRO 6000 (sm120). The results are shown below as text-generation tokens/s.",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27464",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27464",
   "tps": "202.6"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 32,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-27464-gb10-nemo-npl32-cur",
   "model": "Nemotron 3 Nano 30B-A3B",
   "notes": "config=master-78ec4c3; TG t/s, llama-batched-bench -npp 64 -ntg 128, npl=32 concurrent streams, current master 78ec4c3, GEMV dispatch",
   "params": "30B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Nemotron 3 Nano 30B-A3B (NVFP4):; | 32  | 214.8 | 350.8 | +63% |; The numbers are on master `78ec4c3` with `llama-batched-bench -npp 64 -ntg 128 -ngl 1,8,32,256`.; on GB10 (sm121) and RTX PRO 6000 (sm120). The results are shown below as text-generation tokens/s.",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27464",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27464",
   "tps": "214.8"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 32,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-27464-gb10-nemo-npl32-pat",
   "model": "Nemotron 3 Nano 30B-A3B",
   "notes": "config=patched-mamba2-flat; TG t/s, llama-batched-bench -npp 64 -ntg 128, npl=32 concurrent streams, proposed mamba2 flat-2d fix, GEMM dispatch",
   "params": "30B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Nemotron 3 Nano 30B-A3B (NVFP4):; | 32  | 214.8 | 350.8 | +63% |; The numbers are on master `78ec4c3` with `llama-batched-bench -npp 64 -ntg 128 -ngl 1,8,32,256`.; on GB10 (sm121) and RTX PRO 6000 (sm120). The results are shown below as text-generation tokens/s.",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27464",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27464",
   "tps": "350.8"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 256,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-27464-gb10-nemo-npl256-cur",
   "model": "Nemotron 3 Nano 30B-A3B",
   "notes": "config=master-78ec4c3; TG t/s, llama-batched-bench -npp 64 -ntg 128, npl=256 concurrent streams, current master 78ec4c3, GEMV dispatch",
   "params": "30B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Nemotron 3 Nano 30B-A3B (NVFP4):; | 256 | 292.5 | 660.8 | +126% |; The numbers are on master `78ec4c3` with `llama-batched-bench -npp 64 -ntg 128 -ngl 1,8,32,256`.; on GB10 (sm121) and RTX PRO 6000 (sm120). The results are shown below as text-generation tokens/s.",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27464",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27464",
   "tps": "292.5"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 256,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-27464-gb10-nemo-npl256-pat",
   "model": "Nemotron 3 Nano 30B-A3B",
   "notes": "config=patched-mamba2-flat; TG t/s, llama-batched-bench -npp 64 -ntg 128, npl=256 concurrent streams, proposed mamba2 flat-2d fix, GEMM dispatch",
   "params": "30B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Nemotron 3 Nano 30B-A3B (NVFP4):; | 256 | 292.5 | 660.8 | +126% |; The numbers are on master `78ec4c3` with `llama-batched-bench -npp 64 -ntg 128 -ngl 1,8,32,256`.; on GB10 (sm121) and RTX PRO 6000 (sm120). The results are shown below as text-generation tokens/s.",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27464",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27464",
   "tps": "660.8"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 32,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-27464-gb10-mamba227b-qQ8_0-cur",
   "model": "mamba2-2.7b",
   "notes": "config=master-78ec4c3; TG t/s, llama-batched-bench -npp 64 -ntg 128, npl=32 concurrent streams, current master 78ec4c3, GEMV dispatch",
   "params": "2.7B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "| Mamba2 2.7B Q8_0 | 98.4 | 246.7 | +151% |; The numbers are on master `78ec4c3` with `llama-batched-bench -npp 64 -ntg 128 -ngl 1,8,32,256`.; on GB10 (sm121) and RTX PRO 6000 (sm120). The results are shown below as text-generation tokens/s.",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27464",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27464",
   "tps": "98.4"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 32,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-27464-gb10-mamba227b-qQ8_0-pat",
   "model": "mamba2-2.7b",
   "notes": "config=patched-mamba2-flat; TG t/s, llama-batched-bench -npp 64 -ntg 128, npl=32 concurrent streams, proposed mamba2 flat-2d fix, GEMM dispatch",
   "params": "2.7B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "| Mamba2 2.7B Q8_0 | 98.4 | 246.7 | +151% |; The numbers are on master `78ec4c3` with `llama-batched-bench -npp 64 -ntg 128 -ngl 1,8,32,256`.; on GB10 (sm121) and RTX PRO 6000 (sm120). The results are shown below as text-generation tokens/s.",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27464",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27464",
   "tps": "246.7"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 32,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-27464-gb10-falconh17b-qBF16-cur",
   "model": "Falcon-H1-7B",
   "notes": "config=master-78ec4c3; TG t/s, llama-batched-bench -npp 64 -ntg 128, npl=32 concurrent streams, current master 78ec4c3, GEMV dispatch",
   "params": "7B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "| Falcon-H1 7B BF16 | 80.0 | 199.0 | +149% |; The numbers are on master `78ec4c3` with `llama-batched-bench -npp 64 -ntg 128 -ngl 1,8,32,256`.; on GB10 (sm121) and RTX PRO 6000 (sm120). The results are shown below as text-generation tokens/s.",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27464",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27464",
   "tps": "80.0"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 32,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-27464-gb10-falconh17b-qBF16-pat",
   "model": "Falcon-H1-7B",
   "notes": "config=patched-mamba2-flat; TG t/s, llama-batched-bench -npp 64 -ntg 128, npl=32 concurrent streams, proposed mamba2 flat-2d fix, GEMM dispatch",
   "params": "7B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "| Falcon-H1 7B BF16 | 80.0 | 199.0 | +149% |; The numbers are on master `78ec4c3` with `llama-batched-bench -npp 64 -ntg 128 -ngl 1,8,32,256`.; on GB10 (sm121) and RTX PRO 6000 (sm120). The results are shown below as text-generation tokens/s.",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27464",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27464",
   "tps": "199.0"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 32,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-27464-gb10-falconh17b-qQ4_K_M-cur",
   "model": "Falcon-H1-7B",
   "notes": "config=master-78ec4c3; TG t/s, llama-batched-bench -npp 64 -ntg 128, npl=32 concurrent streams, current master 78ec4c3, GEMV dispatch",
   "params": "7B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "| Falcon-H1 7B Q4_K_M | 187.7 | 273.1 | +45% |; The numbers are on master `78ec4c3` with `llama-batched-bench -npp 64 -ntg 128 -ngl 1,8,32,256`.; on GB10 (sm121) and RTX PRO 6000 (sm120). The results are shown below as text-generation tokens/s.",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27464",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27464",
   "tps": "187.7"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 32,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-27464-gb10-falconh17b-qQ4_K_M-pat",
   "model": "Falcon-H1-7B",
   "notes": "config=patched-mamba2-flat; TG t/s, llama-batched-bench -npp 64 -ntg 128, npl=32 concurrent streams, proposed mamba2 flat-2d fix, GEMM dispatch",
   "params": "7B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "| Falcon-H1 7B Q4_K_M | 187.7 | 273.1 | +45% |; The numbers are on master `78ec4c3` with `llama-batched-bench -npp 64 -ntg 128 -ngl 1,8,32,256`.; on GB10 (sm121) and RTX PRO 6000 (sm120). The results are shown below as text-generation tokens/s.",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27464",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27464",
   "tps": "273.1"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 32,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-27464-gb10-granite40htiny-qBF16-cur",
   "model": "granite-4.0-h-tiny",
   "notes": "config=master-78ec4c3; TG t/s, llama-batched-bench -npp 64 -ntg 128, npl=32 concurrent streams, current master 78ec4c3, GEMV dispatch",
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "| Granite 4.0 H Tiny BF16 | 271.3 | 390.4 | +44% |; The numbers are on master `78ec4c3` with `llama-batched-bench -npp 64 -ntg 128 -ngl 1,8,32,256`.; on GB10 (sm121) and RTX PRO 6000 (sm120). The results are shown below as text-generation tokens/s.",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27464",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27464",
   "tps": "271.3"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 32,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-27464-gb10-granite40htiny-qBF16-pat",
   "model": "granite-4.0-h-tiny",
   "notes": "config=patched-mamba2-flat; TG t/s, llama-batched-bench -npp 64 -ntg 128, npl=32 concurrent streams, proposed mamba2 flat-2d fix, GEMM dispatch",
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "| Granite 4.0 H Tiny BF16 | 271.3 | 390.4 | +44% |; The numbers are on master `78ec4c3` with `llama-batched-bench -npp 64 -ntg 128 -ngl 1,8,32,256`.; on GB10 (sm121) and RTX PRO 6000 (sm120). The results are shown below as text-generation tokens/s.",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27464",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27464",
   "tps": "390.4"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 32,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-27464-gb10-granite40htiny-qQ4_K_M-cur",
   "model": "granite-4.0-h-tiny",
   "notes": "config=master-78ec4c3; TG t/s, llama-batched-bench -npp 64 -ntg 128, npl=32 concurrent streams, current master 78ec4c3, GEMV dispatch",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "| Granite 4.0 H Tiny Q4_K_M | 456.4 | 551.2 | +21% |; The numbers are on master `78ec4c3` with `llama-batched-bench -npp 64 -ntg 128 -ngl 1,8,32,256`.; on GB10 (sm121) and RTX PRO 6000 (sm120). The results are shown below as text-generation tokens/s.",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27464",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27464",
   "tps": "456.4"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 32,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-27464-gb10-granite40htiny-qQ4_K_M-pat",
   "model": "granite-4.0-h-tiny",
   "notes": "config=patched-mamba2-flat; TG t/s, llama-batched-bench -npp 64 -ntg 128, npl=32 concurrent streams, proposed mamba2 flat-2d fix, GEMM dispatch",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "| Granite 4.0 H Tiny Q4_K_M | 456.4 | 551.2 | +21% |; The numbers are on master `78ec4c3` with `llama-batched-bench -npp 64 -ntg 128 -ngl 1,8,32,256`.; on GB10 (sm121) and RTX PRO 6000 (sm120). The results are shown below as text-generation tokens/s.",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27464",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27464",
   "tps": "551.2"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "lc-27464-pro6000-nemo-npl1-cur",
   "model": "Nemotron 3 Nano 30B-A3B",
   "notes": "config=master-78ec4c3; TG t/s, llama-batched-bench -npp 64 -ntg 128, npl=1 concurrent streams, current master 78ec4c3, GEMV dispatch",
   "params": "30B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Nemotron 3 Nano 30B-A3B (NVFP4):; | 1   |  333.2 |  332.0 | 0% |; The numbers are on master `78ec4c3` with `llama-batched-bench -npp 64 -ntg 128 -ngl 1,8,32,256`.; on GB10 (sm121) and RTX PRO 6000 (sm120). The results are shown below as text-generation tokens/s.",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27464",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27464",
   "tps": "333.2"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "lc-27464-pro6000-nemo-npl1-pat",
   "model": "Nemotron 3 Nano 30B-A3B",
   "notes": "config=patched-mamba2-flat; TG t/s, llama-batched-bench -npp 64 -ntg 128, npl=1 concurrent streams, proposed mamba2 flat-2d fix, GEMM dispatch",
   "params": "30B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Nemotron 3 Nano 30B-A3B (NVFP4):; | 1   |  333.2 |  332.0 | 0% |; The numbers are on master `78ec4c3` with `llama-batched-bench -npp 64 -ntg 128 -ngl 1,8,32,256`.; on GB10 (sm121) and RTX PRO 6000 (sm120). The results are shown below as text-generation tokens/s.",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27464",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27464",
   "tps": "332.0"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 8,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "lc-27464-pro6000-nemo-npl8-cur",
   "model": "Nemotron 3 Nano 30B-A3B",
   "notes": "config=master-78ec4c3; TG t/s, llama-batched-bench -npp 64 -ntg 128, npl=8 concurrent streams, current master 78ec4c3, GEMV dispatch",
   "params": "30B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Nemotron 3 Nano 30B-A3B (NVFP4):; | 8   | 1035.5 | 1234.1 | +19% |; The numbers are on master `78ec4c3` with `llama-batched-bench -npp 64 -ntg 128 -ngl 1,8,32,256`.; on GB10 (sm121) and RTX PRO 6000 (sm120). The results are shown below as text-generation tokens/s.",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27464",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27464",
   "tps": "1035.5"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 8,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "lc-27464-pro6000-nemo-npl8-pat",
   "model": "Nemotron 3 Nano 30B-A3B",
   "notes": "config=patched-mamba2-flat; TG t/s, llama-batched-bench -npp 64 -ntg 128, npl=8 concurrent streams, proposed mamba2 flat-2d fix, GEMM dispatch",
   "params": "30B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Nemotron 3 Nano 30B-A3B (NVFP4):; | 8   | 1035.5 | 1234.1 | +19% |; The numbers are on master `78ec4c3` with `llama-batched-bench -npp 64 -ntg 128 -ngl 1,8,32,256`.; on GB10 (sm121) and RTX PRO 6000 (sm120). The results are shown below as text-generation tokens/s.",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27464",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27464",
   "tps": "1234.1"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 32,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "lc-27464-pro6000-nemo-npl32-cur",
   "model": "Nemotron 3 Nano 30B-A3B",
   "notes": "config=master-78ec4c3; TG t/s, llama-batched-bench -npp 64 -ntg 128, npl=32 concurrent streams, current master 78ec4c3, GEMV dispatch",
   "params": "30B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Nemotron 3 Nano 30B-A3B (NVFP4):; | 32  | 1542.0 | 2132.4 | +38% |; The numbers are on master `78ec4c3` with `llama-batched-bench -npp 64 -ntg 128 -ngl 1,8,32,256`.; on GB10 (sm121) and RTX PRO 6000 (sm120). The results are shown below as text-generation tokens/s.",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27464",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27464",
   "tps": "1542.0"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 32,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "lc-27464-pro6000-nemo-npl32-pat",
   "model": "Nemotron 3 Nano 30B-A3B",
   "notes": "config=patched-mamba2-flat; TG t/s, llama-batched-bench -npp 64 -ntg 128, npl=32 concurrent streams, proposed mamba2 flat-2d fix, GEMM dispatch",
   "params": "30B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Nemotron 3 Nano 30B-A3B (NVFP4):; | 32  | 1542.0 | 2132.4 | +38% |; The numbers are on master `78ec4c3` with `llama-batched-bench -npp 64 -ntg 128 -ngl 1,8,32,256`.; on GB10 (sm121) and RTX PRO 6000 (sm120). The results are shown below as text-generation tokens/s.",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27464",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27464",
   "tps": "2132.4"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 256,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "lc-27464-pro6000-nemo-npl256-cur",
   "model": "Nemotron 3 Nano 30B-A3B",
   "notes": "config=master-78ec4c3; TG t/s, llama-batched-bench -npp 64 -ntg 128, npl=256 concurrent streams, current master 78ec4c3, GEMV dispatch",
   "params": "30B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Nemotron 3 Nano 30B-A3B (NVFP4):; | 256 | 1959.8 | 3159.1 | +61% |; The numbers are on master `78ec4c3` with `llama-batched-bench -npp 64 -ntg 128 -ngl 1,8,32,256`.; on GB10 (sm121) and RTX PRO 6000 (sm120). The results are shown below as text-generation tokens/s.",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27464",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27464",
   "tps": "1959.8"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 256,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "lc-27464-pro6000-nemo-npl256-pat",
   "model": "Nemotron 3 Nano 30B-A3B",
   "notes": "config=patched-mamba2-flat; TG t/s, llama-batched-bench -npp 64 -ntg 128, npl=256 concurrent streams, proposed mamba2 flat-2d fix, GEMM dispatch",
   "params": "30B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Nemotron 3 Nano 30B-A3B (NVFP4):; | 256 | 1959.8 | 3159.1 | +61% |; The numbers are on master `78ec4c3` with `llama-batched-bench -npp 64 -ntg 128 -ngl 1,8,32,256`.; on GB10 (sm121) and RTX PRO 6000 (sm120). The results are shown below as text-generation tokens/s.",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27464",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27464",
   "tps": "3159.1"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 32,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "lc-27464-pro6000-mamba227b-qBF16-cur",
   "model": "mamba2-2.7b",
   "notes": "config=master-78ec4c3; TG t/s, llama-batched-bench -npp 64 -ntg 128, npl=32 concurrent streams, current master 78ec4c3, GEMV dispatch",
   "params": "2.7B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "| mamba2-2.7b BF16 | 630.9 | 1508.0 | +139% |; The numbers are on master `78ec4c3` with `llama-batched-bench -npp 64 -ntg 128 -ngl 1,8,32,256`.; on GB10 (sm121) and RTX PRO 6000 (sm120). The results are shown below as text-generation tokens/s.",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27464",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27464",
   "tps": "630.9"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 32,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "lc-27464-pro6000-mamba227b-qBF16-pat",
   "model": "mamba2-2.7b",
   "notes": "config=patched-mamba2-flat; TG t/s, llama-batched-bench -npp 64 -ntg 128, npl=32 concurrent streams, proposed mamba2 flat-2d fix, GEMM dispatch",
   "params": "2.7B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "| mamba2-2.7b BF16 | 630.9 | 1508.0 | +139% |; The numbers are on master `78ec4c3` with `llama-batched-bench -npp 64 -ntg 128 -ngl 1,8,32,256`.; on GB10 (sm121) and RTX PRO 6000 (sm120). The results are shown below as text-generation tokens/s.",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27464",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27464",
   "tps": "1508.0"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 32,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "lc-27464-pro6000-mamba227b-qQ4_K_M-cur",
   "model": "mamba2-2.7b",
   "notes": "config=master-78ec4c3; TG t/s, llama-batched-bench -npp 64 -ntg 128, npl=32 concurrent streams, current master 78ec4c3, GEMV dispatch",
   "params": "2.7B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "| mamba2-2.7b Q4_K_M | 953.0 | 1638.9 | +72% |; The numbers are on master `78ec4c3` with `llama-batched-bench -npp 64 -ntg 128 -ngl 1,8,32,256`.; on GB10 (sm121) and RTX PRO 6000 (sm120). The results are shown below as text-generation tokens/s.",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27464",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27464",
   "tps": "953.0"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 32,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "lc-27464-pro6000-mamba227b-qQ4_K_M-pat",
   "model": "mamba2-2.7b",
   "notes": "config=patched-mamba2-flat; TG t/s, llama-batched-bench -npp 64 -ntg 128, npl=32 concurrent streams, proposed mamba2 flat-2d fix, GEMM dispatch",
   "params": "2.7B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "| mamba2-2.7b Q4_K_M | 953.0 | 1638.9 | +72% |; The numbers are on master `78ec4c3` with `llama-batched-bench -npp 64 -ntg 128 -ngl 1,8,32,256`.; on GB10 (sm121) and RTX PRO 6000 (sm120). The results are shown below as text-generation tokens/s.",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27464",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27464",
   "tps": "1638.9"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 32,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "lc-27464-pro6000-falconh17b-qBF16-cur",
   "model": "Falcon-H1-7B",
   "notes": "config=master-78ec4c3; TG t/s, llama-batched-bench -npp 64 -ntg 128, npl=32 concurrent streams, current master 78ec4c3, GEMV dispatch",
   "params": "7B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "| Falcon-H1-7B BF16 | 771.5 | 1188.2 | +54% |; The numbers are on master `78ec4c3` with `llama-batched-bench -npp 64 -ntg 128 -ngl 1,8,32,256`.; on GB10 (sm121) and RTX PRO 6000 (sm120). The results are shown below as text-generation tokens/s.",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27464",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27464",
   "tps": "771.5"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 32,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "lc-27464-pro6000-falconh17b-qBF16-pat",
   "model": "Falcon-H1-7B",
   "notes": "config=patched-mamba2-flat; TG t/s, llama-batched-bench -npp 64 -ntg 128, npl=32 concurrent streams, proposed mamba2 flat-2d fix, GEMM dispatch",
   "params": "7B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "| Falcon-H1-7B BF16 | 771.5 | 1188.2 | +54% |; The numbers are on master `78ec4c3` with `llama-batched-bench -npp 64 -ntg 128 -ngl 1,8,32,256`.; on GB10 (sm121) and RTX PRO 6000 (sm120). The results are shown below as text-generation tokens/s.",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27464",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27464",
   "tps": "1188.2"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 32,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "lc-27464-pro6000-falconh17b-qQ8_0-cur",
   "model": "Falcon-H1-7B",
   "notes": "config=master-78ec4c3; TG t/s, llama-batched-bench -npp 64 -ntg 128, npl=32 concurrent streams, current master 78ec4c3, GEMV dispatch",
   "params": "7B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "| Falcon-H1-7B Q8_0 | 963.3 | 1318.3 | +37% |; The numbers are on master `78ec4c3` with `llama-batched-bench -npp 64 -ntg 128 -ngl 1,8,32,256`.; on GB10 (sm121) and RTX PRO 6000 (sm120). The results are shown below as text-generation tokens/s.",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27464",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27464",
   "tps": "963.3"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 32,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "lc-27464-pro6000-falconh17b-qQ8_0-pat",
   "model": "Falcon-H1-7B",
   "notes": "config=patched-mamba2-flat; TG t/s, llama-batched-bench -npp 64 -ntg 128, npl=32 concurrent streams, proposed mamba2 flat-2d fix, GEMM dispatch",
   "params": "7B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "| Falcon-H1-7B Q8_0 | 963.3 | 1318.3 | +37% |; The numbers are on master `78ec4c3` with `llama-batched-bench -npp 64 -ntg 128 -ngl 1,8,32,256`.; on GB10 (sm121) and RTX PRO 6000 (sm120). The results are shown below as text-generation tokens/s.",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27464",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27464",
   "tps": "1318.3"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 32,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "lc-27464-pro6000-falconh17b-qQ4_K_M-cur",
   "model": "Falcon-H1-7B",
   "notes": "config=master-78ec4c3; TG t/s, llama-batched-bench -npp 64 -ntg 128, npl=32 concurrent streams, current master 78ec4c3, GEMV dispatch",
   "params": "7B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "| Falcon-H1-7B Q4_K_M | 1067.9 | 1417.1 | +33% |; The numbers are on master `78ec4c3` with `llama-batched-bench -npp 64 -ntg 128 -ngl 1,8,32,256`.; on GB10 (sm121) and RTX PRO 6000 (sm120). The results are shown below as text-generation tokens/s.",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27464",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27464",
   "tps": "1067.9"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 32,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "lc-27464-pro6000-falconh17b-qQ4_K_M-pat",
   "model": "Falcon-H1-7B",
   "notes": "config=patched-mamba2-flat; TG t/s, llama-batched-bench -npp 64 -ntg 128, npl=32 concurrent streams, proposed mamba2 flat-2d fix, GEMM dispatch",
   "params": "7B",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "| Falcon-H1-7B Q4_K_M | 1067.9 | 1417.1 | +33% |; The numbers are on master `78ec4c3` with `llama-batched-bench -npp 64 -ntg 128 -ngl 1,8,32,256`.; on GB10 (sm121) and RTX PRO 6000 (sm120). The results are shown below as text-generation tokens/s.",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27464",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27464",
   "tps": "1417.1"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 32,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "lc-27464-pro6000-granite40htiny-qBF16-cur",
   "model": "granite-4.0-h-tiny",
   "notes": "config=master-78ec4c3; TG t/s, llama-batched-bench -npp 64 -ntg 128, npl=32 concurrent streams, current master 78ec4c3, GEMV dispatch",
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "| granite-4.0-h-tiny BF16 | 1427.0 | 2287.6 | +60% |; The numbers are on master `78ec4c3` with `llama-batched-bench -npp 64 -ntg 128 -ngl 1,8,32,256`.; on GB10 (sm121) and RTX PRO 6000 (sm120). The results are shown below as text-generation tokens/s.",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27464",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27464",
   "tps": "1427.0"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 32,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "lc-27464-pro6000-granite40htiny-qBF16-pat",
   "model": "granite-4.0-h-tiny",
   "notes": "config=patched-mamba2-flat; TG t/s, llama-batched-bench -npp 64 -ntg 128, npl=32 concurrent streams, proposed mamba2 flat-2d fix, GEMM dispatch",
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "| granite-4.0-h-tiny BF16 | 1427.0 | 2287.6 | +60% |; The numbers are on master `78ec4c3` with `llama-batched-bench -npp 64 -ntg 128 -ngl 1,8,32,256`.; on GB10 (sm121) and RTX PRO 6000 (sm120). The results are shown below as text-generation tokens/s.",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27464",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27464",
   "tps": "2287.6"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 32,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "lc-27464-pro6000-granite40htiny-qQ4_K_M-cur",
   "model": "granite-4.0-h-tiny",
   "notes": "config=master-78ec4c3; TG t/s, llama-batched-bench -npp 64 -ntg 128, npl=32 concurrent streams, current master 78ec4c3, GEMV dispatch",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "| granite-4.0-h-tiny Q4_K_M | 2192.4 | 2541.9 | +16% |; The numbers are on master `78ec4c3` with `llama-batched-bench -npp 64 -ntg 128 -ngl 1,8,32,256`.; on GB10 (sm121) and RTX PRO 6000 (sm120). The results are shown below as text-generation tokens/s.",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27464",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27464",
   "tps": "2192.4"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 32,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "lc-27464-pro6000-granite40htiny-qQ4_K_M-pat",
   "model": "granite-4.0-h-tiny",
   "notes": "config=patched-mamba2-flat; TG t/s, llama-batched-bench -npp 64 -ntg 128, npl=32 concurrent streams, proposed mamba2 flat-2d fix, GEMM dispatch",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "| granite-4.0-h-tiny Q4_K_M | 2192.4 | 2541.9 | +16% |; The numbers are on master `78ec4c3` with `llama-batched-bench -npp 64 -ntg 128 -ngl 1,8,32,256`.; on GB10 (sm121) and RTX PRO 6000 (sm120). The results are shown below as text-generation tokens/s.",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27464",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27464",
   "tps": "2541.9"
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "H100 NVL",
   "id": "vllm-57680-h100nvl-qwen36-35b-v0240-c1",
   "model": "Qwen3.6-35B-A3B-FP8",
   "notes": "H100 NVL SEV-SNP confidential computing, TP=1; 512-token output, ~122-token prompt, median of 5; image A vLLM 0.24.0/torch 2.11/triton 3.6/CUDA 12.9; ITL 7.08 ms; config=v0240",
   "pp_tokens": 122,
   "provenance": "community",
   "quant": "FP8",
   "quote": "- GPU: 1\u00d7 H100 NVL (SM90), no MIG, TP=1, PP=1; - Model: `Qwen/Qwen3.6-35B-A3B-FP8` (hybrid Gated DeltaNet + attention, MoE 35B total / 3B active); Output 512 tokens with `ignore_eos`, ~122-token prompt, **median of 5 repetitions** (single measurements are unusable, see below):; | A (0.24.0) | 138.8 tok/s | 1183.3 tok/s | 7.08 ms |",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #57680 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/57680",
   "tg_tokens": 512,
   "tps": "138.8"
  },
  {
   "backend": "vLLM",
   "batch": 12,
   "ctx": null,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "H100 NVL",
   "id": "vllm-57680-h100nvl-qwen36-35b-v0240-c12",
   "model": "Qwen3.6-35B-A3B-FP8",
   "notes": "H100 NVL SEV-SNP confidential computing, TP=1; 512-token output, ~122-token prompt, median of 5; image A vLLM 0.24.0, 12 concurrent; ITL 7.08 ms; config=v0240",
   "pp_tokens": 122,
   "provenance": "community",
   "quant": "FP8",
   "quote": "- GPU: 1\u00d7 H100 NVL (SM90), no MIG, TP=1, PP=1; - Model: `Qwen/Qwen3.6-35B-A3B-FP8` (hybrid Gated DeltaNet + attention, MoE 35B total / 3B active); Output 512 tokens with `ignore_eos`, ~122-token prompt, **median of 5 repetitions** (single measurements are unusable, see below):; | A (0.24.0) | 138.8 tok/s | 1183.3 tok/s | 7.08 ms |",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #57680 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/57680",
   "tg_tokens": 512,
   "tps": "1183.3"
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "H100 NVL",
   "id": "vllm-57680-h100nvl-qwen36-35b-v0260-c1",
   "model": "Qwen3.6-35B-A3B-FP8",
   "notes": "H100 NVL SEV-SNP confidential computing, TP=1; 512-token output, ~122-token prompt, median of 5; image B vLLM 0.26.0/torch 2.11/triton 3.6/CUDA 13.0; ITL 6.68 ms; config=v0260",
   "pp_tokens": 122,
   "provenance": "community",
   "quant": "FP8",
   "quote": "- GPU: 1\u00d7 H100 NVL (SM90), no MIG, TP=1, PP=1; - Model: `Qwen/Qwen3.6-35B-A3B-FP8` (hybrid Gated DeltaNet + attention, MoE 35B total / 3B active); Output 512 tokens with `ignore_eos`, ~122-token prompt, **median of 5 repetitions** (single measurements are unusable, see below):; | **B (0.26.0)** | **148.3 tok/s** | **1295.0 tok/s** | **6.68 ms** |",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #57680 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/57680",
   "tg_tokens": 512,
   "tps": "148.3"
  },
  {
   "backend": "vLLM",
   "batch": 12,
   "ctx": null,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "H100 NVL",
   "id": "vllm-57680-h100nvl-qwen36-35b-v0260-c12",
   "model": "Qwen3.6-35B-A3B-FP8",
   "notes": "H100 NVL SEV-SNP confidential computing, TP=1; 512-token output, ~122-token prompt, median of 5; image B vLLM 0.26.0, 12 concurrent; ITL 6.68 ms; config=v0260",
   "pp_tokens": 122,
   "provenance": "community",
   "quant": "FP8",
   "quote": "- GPU: 1\u00d7 H100 NVL (SM90), no MIG, TP=1, PP=1; - Model: `Qwen/Qwen3.6-35B-A3B-FP8` (hybrid Gated DeltaNet + attention, MoE 35B total / 3B active); Output 512 tokens with `ignore_eos`, ~122-token prompt, **median of 5 repetitions** (single measurements are unusable, see below):; | **B (0.26.0)** | **148.3 tok/s** | **1295.0 tok/s** | **6.68 ms** |",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #57680 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/57680",
   "tg_tokens": 512,
   "tps": "1295.0"
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "H100 NVL",
   "id": "vllm-57680-h100nvl-qwen36-35b-v0290-c1",
   "model": "Qwen3.6-35B-A3B-FP8",
   "notes": "H100 NVL SEV-SNP confidential computing, TP=1; 512-token output, ~122-token prompt, median of 5; image C vLLM 0.29.0/torch 2.13/triton 3.7.1; ITL 23.88 ms; c=1 alternates 41.4/60.8 per request (1.47x swing, not prefix caching); config=v0290",
   "pp_tokens": 122,
   "provenance": "community",
   "quant": "FP8",
   "quote": "- GPU: 1\u00d7 H100 NVL (SM90), no MIG, TP=1, PP=1; - Model: `Qwen/Qwen3.6-35B-A3B-FP8` (hybrid Gated DeltaNet + attention, MoE 35B total / 3B active); Output 512 tokens with `ignore_eos`, ~122-token prompt, **median of 5 repetitions** (single measurements are unusable, see below):; | **C (0.29.0)** | **41.4 tok/s** | **362.3 tok/s** | **23.88 ms** |",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #57680 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/57680",
   "tg_tokens": 512,
   "tps": "41.4"
  },
  {
   "backend": "vLLM",
   "batch": 12,
   "ctx": null,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "H100 NVL",
   "id": "vllm-57680-h100nvl-qwen36-35b-v0290-c12",
   "model": "Qwen3.6-35B-A3B-FP8",
   "notes": "H100 NVL SEV-SNP confidential computing, TP=1; 512-token output, ~122-token prompt, median of 5; image C vLLM 0.29.0, 12 concurrent; ITL 23.88 ms; config=v0290",
   "pp_tokens": 122,
   "provenance": "community",
   "quant": "FP8",
   "quote": "- GPU: 1\u00d7 H100 NVL (SM90), no MIG, TP=1, PP=1; - Model: `Qwen/Qwen3.6-35B-A3B-FP8` (hybrid Gated DeltaNet + attention, MoE 35B total / 3B active); Output 512 tokens with `ignore_eos`, ~122-token prompt, **median of 5 repetitions** (single measurements are unusable, see below):; | **C (0.29.0)** | **41.4 tok/s** | **362.3 tok/s** | **23.88 ms** |",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #57680 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/57680",
   "tg_tokens": 512,
   "tps": "362.3"
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "2\u00d7 NVIDIA RTX A5000 24 GB",
   "id": "vllm-40124-a5000x2-qwen3-35b-decode",
   "model": "Qwen3-35B-A3B",
   "notes": "after all 13 patches (Sandermage/genesis-vllm-patches), vLLM 0.19.1rc1 nightly 2026-04-16, TurboQuant k8v4 KV cache, max_model_len 163840; single request, short context; stability 10/10 runs sigma 0.11 tok/s; CUDAGraphs required (eager is 13x slower); config=tq-k8v4",
   "provenance": "community",
   "quant": "FP8",
   "quote": "- **GPUs**: 2\u00d7 NVIDIA RTX A5000 24 GB (SM86, Ampere), `tensor_parallel_size=2`; - **Model**: `Qwen/Qwen3-35B-A3B-FP8` \u2014 hybrid MoE, `block_shape=[128,128]`; Hardware: 2\u00d7 RTX A5000, TP=2, TurboQuant K8V4, max_model_len=160k; - **Decode speed**: 145.3 tok/s (single request, short context); - **Stability**: 10/10 runs, \u03c3=0.11 tok/s",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #40124 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/40124",
   "tps": "145.3"
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 163840,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "2\u00d7 NVIDIA RTX A5000 24 GB",
   "id": "vllm-40124-a5000x2-qwen3-35b-160k",
   "model": "Qwen3-35B-A3B",
   "notes": "after all 13 patches (Sandermage/genesis-vllm-patches), vLLM 0.19.1rc1 nightly 2026-04-16, TurboQuant k8v4 KV cache, max_model_len 163840; 160k context as published; config=tq-k8v4",
   "provenance": "community",
   "quant": "FP8",
   "quote": "- **GPUs**: 2\u00d7 NVIDIA RTX A5000 24 GB (SM86, Ampere), `tensor_parallel_size=2`; - **Model**: `Qwen/Qwen3-35B-A3B-FP8` \u2014 hybrid MoE, `block_shape=[128,128]`; Hardware: 2\u00d7 RTX A5000, TP=2, TurboQuant K8V4, max_model_len=160k; - **160k context**: 48.6 tok/s \u2705",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #40124 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/40124",
   "tps": "48.6"
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "2\u00d7 NVIDIA RTX A5000 24 GB",
   "id": "vllm-40124-a5000x2-qwen3-35b-eager",
   "model": "Qwen3-35B-A3B",
   "notes": "after all 13 patches (Sandermage/genesis-vllm-patches), vLLM 0.19.1rc1 nightly 2026-04-16, TurboQuant k8v4 KV cache, max_model_len 163840; enforce-eager (no CUDAGraphs): drops 143 to 11 tok/s, 13x slower; config=tq-k8v4-enforceeager",
   "provenance": "community",
   "quant": "FP8",
   "quote": "- **GPUs**: 2\u00d7 NVIDIA RTX A5000 24 GB (SM86, Ampere), `tensor_parallel_size=2`; - **Model**: `Qwen/Qwen3-35B-A3B-FP8` \u2014 hybrid MoE, `block_shape=[128,128]`; Hardware: 2\u00d7 RTX A5000, TP=2, TurboQuant K8V4, max_model_len=160k; - **CUDAGraphs mandatory**: `--enforce-eager` drops speed 143\u219211 tok/s (13\u00d7 slower)",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #40124 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/40124",
   "tps": "11"
  },
  {
   "backend": "vLLM",
   "batch": 3,
   "ctx": null,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "8x NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition (96 GB each)",
   "id": "vllm-56892-rtxpro6000x8-dsv41flash-eager",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM official deepseekv41-flash-0909 image, TP=8, kv fp8, enforce-eager (CUDA graphs unusable on SM120 for V4.1-Flash); engine-side avg over a 3-concurrent-request bench, 300 completion tokens each, ~1 tok/s per request; config=8xtp-eager",
   "provenance": "community",
   "quant": "FP8",
   "quote": "- **GPU**: 8x NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition (96 GB each), compute capability **12.0 (SM120)**; - **Model**: DeepSeek-V4.1-Flash (FP8 checkpoint), served with the community SM120 patch set; Engine 000: Avg generation throughput: 2.3\u20134.0 tokens/s, Running: 3 reqs",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #56892 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/56892",
   "tps": "4.0"
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "8x NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition (96 GB each)",
   "id": "vllm-56892-rtxpro6000x8-dsv4flash-published",
   "model": "DS-V4-Flash",
   "notes": "reference aggregate from the issue body: DS-V4-Flash (the V4-Flash checkpoint) on the same 8x RTX PRO 6000 per published recipes, not measured by this poster; config=published-reference",
   "provenance": "community",
   "quant": "FP8",
   "quote": "- **GPU**: 8x NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition (96 GB each), compute capability **12.0 (SM120)**; for reference, DS-V4-Flash on the same 8x RTX PRO 6000 reaches ~2452 tok/s aggregate in published recipes",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #56892 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/56892",
   "tps": "2452"
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "vllm-58639-r9700-pp3-step-sidestreams",
   "model": "MoE hybrid model",
   "notes": "PP=3 (V620/V620/R9700, decode rank R9700), MTP2 spec, FULL HIP graphs, ROCm 7.14; model name not published; pipeline decode throughput as published (~33 t/s); R9700 forward ~16.9-17.0 ms, step period ~56-58 ms; config=pp3-step-sidestreams",
   "provenance": "community",
   "quant": "as-published",
   "quote": "On AMD gfx1201 (Radeon AI PRO R9700), a stream of small kernels picks up about 17 \u00b5s per dispatch when two or more other HIP streams hold pending waits on it (default hardware-queue limit, graph mode).; The setup was PP=3 and TP=1 with ranks on V620, V620 and R9700 (last rank). It used an MoE hybrid model with MTP2 speculative decoding, FULL HIP graphs and ROCm 7.14.; | Side streams (default) | ~16.9\u201317.0 ms | ~56\u201358 ms | ~33 tok/s |",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #58639 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58639",
   "tps": "33"
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "vllm-58639-r9700-pp3-step-mainstream",
   "model": "MoE hybrid model",
   "notes": "PP=3 (V620/V620/R9700, decode rank R9700), MTP2 spec, FULL HIP graphs, ROCm 7.14; model name not published; pipeline decode throughput, main-stream fix, 45.3-46.6 range as published; R9700 forward 7.18-7.23 ms, step period 41.6-42.2 ms; config=pp3-step-mainstream",
   "provenance": "community",
   "quant": "as-published",
   "quote": "On AMD gfx1201 (Radeon AI PRO R9700), a stream of small kernels picks up about 17 \u00b5s per dispatch when two or more other HIP streams hold pending waits on it (default hardware-queue limit, graph mode).; The setup was PP=3 and TP=1 with ranks on V620, V620 and R9700 (last rank). It used an MoE hybrid model with MTP2 speculative decoding, FULL HIP graphs and ROCm 7.14.; | Last-rank sends and output copy on the main stream | 7.18\u20137.23 ms | 41.6\u201342.2 ms | 45.3\u201346.6 tok/s |",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #58639 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58639",
   "tps": "46.6"
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "vllm-58639-r9700-pp3-serve-sidestreams",
   "model": "MoE hybrid model",
   "notes": "PP=3 (V620/V620/R9700, decode rank R9700), MTP2 spec, FULL HIP graphs, ROCm 7.14; model name not published; separate serving measurement, median decode before the main-stream fix; config=pp3-serve-sidestreams",
   "provenance": "community",
   "quant": "as-published",
   "quote": "On AMD gfx1201 (Radeon AI PRO R9700), a stream of small kernels picks up about 17 \u00b5s per dispatch when two or more other HIP streams hold pending waits on it (default hardware-queue limit, graph mode).; The setup was PP=3 and TP=1 with ranks on V620, V620 and R9700 (last rank). It used an MoE hybrid model with MTP2 speculative decoding, FULL HIP graphs and ROCm 7.14.; In a separate serving measurement, median decode rose from 31.0 to 46.5 tok/s.",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #58639 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58639",
   "tps": "31.0"
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "vllm-58639-r9700-pp3-serve-mainstream",
   "model": "MoE hybrid model",
   "notes": "PP=3 (V620/V620/R9700, decode rank R9700), MTP2 spec, FULL HIP graphs, ROCm 7.14; model name not published; separate serving measurement, median decode after the main-stream fix; config=pp3-serve-mainstream",
   "provenance": "community",
   "quant": "as-published",
   "quote": "On AMD gfx1201 (Radeon AI PRO R9700), a stream of small kernels picks up about 17 \u00b5s per dispatch when two or more other HIP streams hold pending waits on it (default hardware-queue limit, graph mode).; The setup was PP=3 and TP=1 with ranks on V620, V620 and R9700 (last rank). It used an MoE hybrid model with MTP2 speculative decoding, FULL HIP graphs and ROCm 7.14.; In a separate serving measurement, median decode rose from 31.0 to 46.5 tok/s.",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #58639 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58639",
   "tps": "46.5"
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "vllm-58639-r9700-pp3-armA-r1",
   "model": "MoE hybrid model",
   "notes": "PP=3 (V620/V620/R9700, decode rank R9700), MTP2 spec, FULL HIP graphs, ROCm 7.14; model name not published; arm A: last-rank main stream on, GPU_MAX_HW_QUEUES unset; step period median 41.744 ms, 16K TTFT 9.900 s; three xhigh decode rates, run 1 of 3; config=armA",
   "provenance": "community",
   "quant": "as-published",
   "quote": "On AMD gfx1201 (Radeon AI PRO R9700), a stream of small kernels picks up about 17 \u00b5s per dispatch when two or more other HIP streams hold pending waits on it (default hardware-queue limit, graph mode).; The setup was PP=3 and TP=1 with ranks on V620, V620 and R9700 (last rank). It used an MoE hybrid model with MTP2 speculative decoding, FULL HIP graphs and ROCm 7.14.; | A | On | Unset | 7.216 / 7.683 ms | 41.744 ms | 48.2 / 48.9 / 42.2 tok/s | 9.900 s |",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #58639 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58639",
   "tps": "48.2"
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "vllm-58639-r9700-pp3-armA-r2",
   "model": "MoE hybrid model",
   "notes": "PP=3 (V620/V620/R9700, decode rank R9700), MTP2 spec, FULL HIP graphs, ROCm 7.14; model name not published; arm A: last-rank main stream on, GPU_MAX_HW_QUEUES unset; step period median 41.744 ms, 16K TTFT 9.900 s; three xhigh decode rates, run 2 of 3; config=armA",
   "provenance": "community",
   "quant": "as-published",
   "quote": "On AMD gfx1201 (Radeon AI PRO R9700), a stream of small kernels picks up about 17 \u00b5s per dispatch when two or more other HIP streams hold pending waits on it (default hardware-queue limit, graph mode).; The setup was PP=3 and TP=1 with ranks on V620, V620 and R9700 (last rank). It used an MoE hybrid model with MTP2 speculative decoding, FULL HIP graphs and ROCm 7.14.; | A | On | Unset | 7.216 / 7.683 ms | 41.744 ms | 48.2 / 48.9 / 42.2 tok/s | 9.900 s |",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #58639 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58639",
   "tps": "48.9"
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "vllm-58639-r9700-pp3-armA-r3",
   "model": "MoE hybrid model",
   "notes": "PP=3 (V620/V620/R9700, decode rank R9700), MTP2 spec, FULL HIP graphs, ROCm 7.14; model name not published; arm A: last-rank main stream on, GPU_MAX_HW_QUEUES unset; step period median 41.744 ms, 16K TTFT 9.900 s; three xhigh decode rates, run 3 of 3; config=armA",
   "provenance": "community",
   "quant": "as-published",
   "quote": "On AMD gfx1201 (Radeon AI PRO R9700), a stream of small kernels picks up about 17 \u00b5s per dispatch when two or more other HIP streams hold pending waits on it (default hardware-queue limit, graph mode).; The setup was PP=3 and TP=1 with ranks on V620, V620 and R9700 (last rank). It used an MoE hybrid model with MTP2 speculative decoding, FULL HIP graphs and ROCm 7.14.; | A | On | Unset | 7.216 / 7.683 ms | 41.744 ms | 48.2 / 48.9 / 42.2 tok/s | 9.900 s |",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #58639 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58639",
   "tps": "42.2"
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "vllm-58639-r9700-pp3-armB-r1",
   "model": "MoE hybrid model",
   "notes": "PP=3 (V620/V620/R9700, decode rank R9700), MTP2 spec, FULL HIP graphs, ROCm 7.14; model name not published; arm B: last-rank main stream off, GPU_MAX_HW_QUEUES=2; step period median 52.755 ms, 16K TTFT 9.915 s; three xhigh decode rates, run 1 of 3; config=armB",
   "provenance": "community",
   "quant": "as-published",
   "quote": "On AMD gfx1201 (Radeon AI PRO R9700), a stream of small kernels picks up about 17 \u00b5s per dispatch when two or more other HIP streams hold pending waits on it (default hardware-queue limit, graph mode).; The setup was PP=3 and TP=1 with ranks on V620, V620 and R9700 (last rank). It used an MoE hybrid model with MTP2 speculative decoding, FULL HIP graphs and ROCm 7.14.; | B | Off | 2 | 7.220 / 7.656 ms | 52.755 ms | 35.9 / 35.8 / 34.3 tok/s | 9.915 s |",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #58639 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58639",
   "tps": "35.9"
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "vllm-58639-r9700-pp3-armB-r2",
   "model": "MoE hybrid model",
   "notes": "PP=3 (V620/V620/R9700, decode rank R9700), MTP2 spec, FULL HIP graphs, ROCm 7.14; model name not published; arm B: last-rank main stream off, GPU_MAX_HW_QUEUES=2; step period median 52.755 ms, 16K TTFT 9.915 s; three xhigh decode rates, run 2 of 3; config=armB",
   "provenance": "community",
   "quant": "as-published",
   "quote": "On AMD gfx1201 (Radeon AI PRO R9700), a stream of small kernels picks up about 17 \u00b5s per dispatch when two or more other HIP streams hold pending waits on it (default hardware-queue limit, graph mode).; The setup was PP=3 and TP=1 with ranks on V620, V620 and R9700 (last rank). It used an MoE hybrid model with MTP2 speculative decoding, FULL HIP graphs and ROCm 7.14.; | B | Off | 2 | 7.220 / 7.656 ms | 52.755 ms | 35.9 / 35.8 / 34.3 tok/s | 9.915 s |",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #58639 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58639",
   "tps": "35.8"
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "vllm-58639-r9700-pp3-armB-r3",
   "model": "MoE hybrid model",
   "notes": "PP=3 (V620/V620/R9700, decode rank R9700), MTP2 spec, FULL HIP graphs, ROCm 7.14; model name not published; arm B: last-rank main stream off, GPU_MAX_HW_QUEUES=2; step period median 52.755 ms, 16K TTFT 9.915 s; three xhigh decode rates, run 3 of 3; config=armB",
   "provenance": "community",
   "quant": "as-published",
   "quote": "On AMD gfx1201 (Radeon AI PRO R9700), a stream of small kernels picks up about 17 \u00b5s per dispatch when two or more other HIP streams hold pending waits on it (default hardware-queue limit, graph mode).; The setup was PP=3 and TP=1 with ranks on V620, V620 and R9700 (last rank). It used an MoE hybrid model with MTP2 speculative decoding, FULL HIP graphs and ROCm 7.14.; | B | Off | 2 | 7.220 / 7.656 ms | 52.755 ms | 35.9 / 35.8 / 34.3 tok/s | 9.915 s |",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #58639 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58639",
   "tps": "34.3"
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "vllm-58639-r9700-pp3-armC-r1",
   "model": "MoE hybrid model",
   "notes": "PP=3 (V620/V620/R9700, decode rank R9700), MTP2 spec, FULL HIP graphs, ROCm 7.14; model name not published; arm C: last-rank main stream on + GPU_MAX_HW_QUEUES=2; step period median 52.239 ms, 16K TTFT 9.954 s; three xhigh decode rates, run 1 of 3; config=armC",
   "provenance": "community",
   "quant": "as-published",
   "quote": "On AMD gfx1201 (Radeon AI PRO R9700), a stream of small kernels picks up about 17 \u00b5s per dispatch when two or more other HIP streams hold pending waits on it (default hardware-queue limit, graph mode).; The setup was PP=3 and TP=1 with ranks on V620, V620 and R9700 (last rank). It used an MoE hybrid model with MTP2 speculative decoding, FULL HIP graphs and ROCm 7.14.; | C | On | 2 | 7.261 / 7.729 ms | 52.239 ms | 35.9 / 36.2 / 37.2 tok/s | 9.954 s |",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #58639 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58639",
   "tps": "35.9"
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "vllm-58639-r9700-pp3-armC-r2",
   "model": "MoE hybrid model",
   "notes": "PP=3 (V620/V620/R9700, decode rank R9700), MTP2 spec, FULL HIP graphs, ROCm 7.14; model name not published; arm C: last-rank main stream on + GPU_MAX_HW_QUEUES=2; step period median 52.239 ms, 16K TTFT 9.954 s; three xhigh decode rates, run 2 of 3; config=armC",
   "provenance": "community",
   "quant": "as-published",
   "quote": "On AMD gfx1201 (Radeon AI PRO R9700), a stream of small kernels picks up about 17 \u00b5s per dispatch when two or more other HIP streams hold pending waits on it (default hardware-queue limit, graph mode).; The setup was PP=3 and TP=1 with ranks on V620, V620 and R9700 (last rank). It used an MoE hybrid model with MTP2 speculative decoding, FULL HIP graphs and ROCm 7.14.; | C | On | 2 | 7.261 / 7.729 ms | 52.239 ms | 35.9 / 36.2 / 37.2 tok/s | 9.954 s |",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #58639 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58639",
   "tps": "36.2"
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "vllm-58639-r9700-pp3-armC-r3",
   "model": "MoE hybrid model",
   "notes": "PP=3 (V620/V620/R9700, decode rank R9700), MTP2 spec, FULL HIP graphs, ROCm 7.14; model name not published; arm C: last-rank main stream on + GPU_MAX_HW_QUEUES=2; step period median 52.239 ms, 16K TTFT 9.954 s; three xhigh decode rates, run 3 of 3; config=armC",
   "provenance": "community",
   "quant": "as-published",
   "quote": "On AMD gfx1201 (Radeon AI PRO R9700), a stream of small kernels picks up about 17 \u00b5s per dispatch when two or more other HIP streams hold pending waits on it (default hardware-queue limit, graph mode).; The setup was PP=3 and TP=1 with ranks on V620, V620 and R9700 (last rank). It used an MoE hybrid model with MTP2 speculative decoding, FULL HIP graphs and ROCm 7.14.; | C | On | 2 | 7.261 / 7.729 ms | 52.239 ms | 35.9 / 36.2 / 37.2 tok/s | 9.954 s |",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #58639 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58639",
   "tps": "37.2"
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "RX 7900 XTX",
   "id": "lc-28863-rocm-solo-c0-q4kxl",
   "model": "Qwen3.8-27B",
   "notes": "llama-bench, -r 5, spread 0.04 t/s; solo card 0 (Sapphire Nitro+ Vapor-X), -sm none; UMC 65% instant, 65.9% from 632.5 GB/s",
   "pp_tokens": 2048,
   "pp_tps": 1074.51,
   "provenance": "community",
   "quant": "Q4_K_XL",
   "quote": "Qwen3.8-27B-UD-Q4_K_XL (16.34 GiB = 17.545 GB) fits on one card; card 0 is a Sapphire Nitro+ Vapor-X; 2x RX 7900 XTX (gfx1100), PCIe 4.0 x8/x8 as configured in BIOS; | 1 card, `-sm none`, card 0 (Q4_K_XL) | 1074.51 \u00b1 1.00 | 36.05 \u00b1 0.04 | 27.74 | 632.5 | 65% | 65.9% |; -p 2048 -n 512 -d 0 -b 4096 -ub 512 -fa 1 --load-mode none -r 5; Stock `f9f09f02c` (10794), no third-party patches",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28863 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28863",
   "tg_tokens": 512,
   "tps": "36.05"
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "RX 7900 XTX",
   "id": "lc-28863-rocm-solo-c1-q4kxl",
   "model": "Qwen3.8-27B",
   "notes": "llama-bench, -r 5, spread 0.04 t/s; solo card 1 (PowerColor Red Devil), -sm none, HIP_VISIBLE_DEVICES=1; UMC 63% instant, 65.0% from 623.5 GB/s",
   "pp_tokens": 2048,
   "pp_tps": 1008.57,
   "provenance": "community",
   "quant": "Q4_K_XL",
   "quote": "Qwen3.8-27B-UD-Q4_K_XL (16.34 GiB = 17.545 GB) fits on one card; card 1 a PowerColor Red Devil; 2x RX 7900 XTX (gfx1100), PCIe 4.0 x8/x8 as configured in BIOS; | 1 card, `-sm none`, card 1 (Q4_K_XL) | 1008.57 \u00b1 1.85 | 35.54 \u00b1 0.04 | 28.14 | 623.5 | 63% | 65.0% |; -p 2048 -n 512 -d 0 -b 4096 -ub 512 -fa 1 --load-mode none -r 5; Stock `f9f09f02c` (10794), no third-party patches",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28863 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28863",
   "tg_tokens": 512,
   "tps": "35.54"
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "2x RX 7900 XTX",
   "id": "lc-28863-rocm-tensor-q4kxl",
   "model": "Qwen3.8-27B",
   "notes": "llama-bench, -r 5, spread 0.24 t/s; tensor split, default even split (8.951 / 9.041 GiB), RCCL P2P direct; UMC 42% instant, 43.0% from 825.7 GB/s aggregate; config=sm-tensor",
   "pp_tokens": 2048,
   "pp_tps": 1547.96,
   "provenance": "community",
   "quant": "Q4_K_XL",
   "quote": "Qwen3.8-27B-UD-Q4_K_XL (16.34 GiB = 17.545 GB) fits on one card; 2x RX 7900 XTX (gfx1100), PCIe 4.0 x8/x8 as configured in BIOS; | 2 cards, `-sm tensor` (Q4_K_XL) | 1547.96 \u00b1 5.25 | 47.06 \u00b1 0.24 | 21.25 | 825.7 agg / 412.8 per card | 42% | 43.0% |; -p 2048 -n 512 -d 0 -b 4096 -ub 512 -fa 1 --load-mode none -r 5; Stock `f9f09f02c` (10794), no third-party patches",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28863 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28863",
   "tg_tokens": 512,
   "tps": "47.06"
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "2x RX 7900 XTX",
   "id": "lc-28863-rocm-tensor-q8",
   "model": "Qwen3.8-27B",
   "notes": "llama-bench, -r 5, spread 0.09 t/s; tensor split; Q8_0 is 29.034 GB so no solo control; pp2048 not reported (spread 210.76 too wide); config=sm-tensor",
   "pp_tokens": 2048,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Qwen3.8-27B-UD-Q4_K_XL (16.34 GiB = 17.545 GB) fits on one card; 2x RX 7900 XTX (gfx1100), PCIe 4.0 x8/x8 as configured in BIOS; | 2 cards, `-sm tensor` (Q8_0) | not reported, see below | 36.26 \u00b1 0.09 | 27.58 | 1052.8 agg / 526.4 per card | 53% | 54.8% |; -p 2048 -n 512 -d 0 -b 4096 -ub 512 -fa 1 --load-mode none -r 5; Stock `f9f09f02c` (10794), no third-party patches",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28863 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28863",
   "tg_tokens": 512,
   "tps": "36.26"
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "RX 7900 XTX",
   "id": "lc-28863-rocm-solo-c0-q8-7b",
   "model": "Qwen2.5-7B",
   "notes": "llama-bench, -r 5, spread 0.09 t/s; solo card 0, -sm none; UMC 76% instant, 80.9% from 776.4 GB/s; 7.54 GiB checkpoint",
   "pp_tokens": 2048,
   "pp_tps": 4534.64,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "A dense Qwen2.5-7B at Q8_0 fits on one card; card 0 is a Sapphire Nitro+ Vapor-X; 2x RX 7900 XTX (gfx1100), PCIe 4.0 x8/x8 as configured in BIOS; | 1 card, `-sm none`, card 0 (7B Q8_0) | 4534.64 \u00b1 5.66 | 95.90 \u00b1 0.09 | 10.43 | 776.4 | 76% | 80.9% |; -p 2048 -n 512 -d 0 -b 4096 -ub 512 -fa 1 --load-mode none -r 5; Stock `f9f09f02c` (10794), no third-party patches",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28863 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28863",
   "tg_tokens": 512,
   "tps": "95.9"
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "RX 7900 XTX",
   "id": "lc-28863-rocm-solo-c1-q8-7b",
   "model": "Qwen2.5-7B",
   "notes": "llama-bench, -r 5, spread 0.06 t/s; solo card 1, -sm none, HIP_VISIBLE_DEVICES=1; 80.3% from 770.7 GB/s",
   "pp_tokens": 2048,
   "pp_tps": 4308.07,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "A dense Qwen2.5-7B at Q8_0 fits on one card; card 1 a PowerColor Red Devil; 2x RX 7900 XTX (gfx1100), PCIe 4.0 x8/x8 as configured in BIOS; | 1 card, `-sm none`, card 1 (7B Q8_0) | 4308.07 \u00b1 3.54 | 95.20 \u00b1 0.06 | 10.50 | 770.7 | \u2014 | 80.3% |; -p 2048 -n 512 -d 0 -b 4096 -ub 512 -fa 1 --load-mode none -r 5; Stock `f9f09f02c` (10794), no third-party patches",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28863 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28863",
   "tg_tokens": 512,
   "tps": "95.2"
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "2x RX 7900 XTX",
   "id": "lc-28863-rocm-tensor-q8-7b",
   "model": "Qwen2.5-7B",
   "notes": "llama-bench, -r 5, spread 5.09 t/s (widest decode row of the table); tensor split; pp2048 not reported, unstable (5513.20 \u00b1 1707.63)",
   "pp_tokens": 2048,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "A dense Qwen2.5-7B at Q8_0 fits on one card; 2x RX 7900 XTX (gfx1100), PCIe 4.0 x8/x8 as configured in BIOS; | 2 cards, `-sm tensor` (7B Q8_0) | not reported, unstable | 122.63 \u00b1 5.09 | 8.15 | 992.8 agg / 496.4 per card | 49% | 51.7% |; -p 2048 -n 512 -d 0 -b 4096 -ub 512 -fa 1 --load-mode none -r 5; Stock `f9f09f02c` (10794), no third-party patches",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28863 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28863",
   "tg_tokens": 512,
   "tps": "122.63"
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "2x RX 7900 XTX",
   "id": "lc-28863-rocm-layer-q4kxl",
   "model": "Qwen3.8-27B",
   "notes": "llama-bench, -r 5, spread 0.09 t/s; -sm layer, later session; worse than tensor split (34.03 vs 47.06); pp2048 spread too wide to use (1550.18 \u00b1 235.38); config=sm-layer",
   "pp_tokens": 2048,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_XL",
   "quote": "Qwen3.8-27B-UD-Q4_K_XL (16.34 GiB = 17.545 GB) fits on one card; 2x RX 7900 XTX (gfx1100), PCIe 4.0 x8/x8 as configured in BIOS; | 2 cards, `-sm layer` (Q4_K_XL) | not reported, unstable | 34.03 \u00b1 0.09 | 29.38 | 597.1 | \u2014 | 62.2% |; -p 2048 -n 512 -d 0 -b 4096 -ub 512 -fa 1 --load-mode none -r 5; Stock `f9f09f02c` (10794), no third-party patches",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28863 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28863",
   "tg_tokens": 512,
   "tps": "34.03"
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "2x RX 7900 XTX",
   "id": "lc-28863-rocm-layer-q8",
   "model": "Qwen3.8-27B",
   "notes": "llama-bench, -r 5, spread 0.02 t/s; -sm layer, later session; per active card 74.6% from 716.6 GB/s; config=sm-layer",
   "pp_tokens": 2048,
   "pp_tps": 1719.88,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Qwen3.8-27B-UD-Q4_K_XL (16.34 GiB = 17.545 GB) fits on one card; 2x RX 7900 XTX (gfx1100), PCIe 4.0 x8/x8 as configured in BIOS; | 2 cards, `-sm layer` (Q8_0) | 1719.88 \u00b1 1.50 | 24.68 \u00b1 0.02 | 40.52 | 716.6 | \u2014 | 74.6% |; -p 2048 -n 512 -d 0 -b 4096 -ub 512 -fa 1 --load-mode none -r 5; Stock `f9f09f02c` (10794), no third-party patches",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28863 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28863",
   "tg_tokens": 512,
   "tps": "24.68"
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "2x RX 7900 XTX",
   "id": "lc-28863-rocm-ub512",
   "model": "Qwen3.8-27B",
   "notes": "-ub 512 sweep row (2-card tensor split); decode column is a re-measure of the same config (run-to-run noise, not an effect); pp ratio 1.466 vs 1 card; config=sm-tensor",
   "pp_tokens": 2048,
   "pp_tps": 1569.15,
   "provenance": "community",
   "quant": "Q4_K_XL",
   "quote": "Qwen3.8-27B-UD-Q4_K_XL (16.34 GiB = 17.545 GB) fits on one card; 2x RX 7900 XTX (gfx1100), PCIe 4.0 x8/x8 as configured in BIOS; | 512 | ~1070 (\u00b1112) | 1569.15 \u00b1 2.25 | 1.466 | 36.04 | 46.28 | 1.284 |; -p 2048 -n 512 -d 0 -b 4096 -ub 512 -fa 1 --load-mode none -r 5; Stock `f9f09f02c` (10794), no third-party patches",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28863 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28863",
   "tg_tokens": 512,
   "tps": "46.28"
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "2x RX 7900 XTX",
   "id": "lc-28863-rocm-ub1024",
   "model": "Qwen3.8-27B",
   "notes": "-ub 1024 sweep row (2-card tensor split); decode re-measure of same config; pp ratio 1.518 vs 1 card; config=sm-tensor",
   "pp_tokens": 2048,
   "pp_tps": 1630.91,
   "provenance": "community",
   "quant": "Q4_K_XL",
   "quote": "Qwen3.8-27B-UD-Q4_K_XL (16.34 GiB = 17.545 GB) fits on one card; 2x RX 7900 XTX (gfx1100), PCIe 4.0 x8/x8 as configured in BIOS; | 1024 | 1074.37 \u00b1 0.45 | 1630.91 \u00b1 4.64 | 1.518 | 36.03 | 46.90 | 1.302 |; -p 2048 -n 512 -d 0 -b 4096 -ub 512 -fa 1 --load-mode none -r 5; Stock `f9f09f02c` (10794), no third-party patches",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28863 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28863",
   "tg_tokens": 512,
   "tps": "46.9"
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "2x RX 7900 XTX",
   "id": "lc-28863-rocm-ub2048",
   "model": "Qwen3.8-27B",
   "notes": "-ub 2048 sweep row (2-card tensor split); decode re-measure of same config; pair gains 7.9% prefill from ub 512 to 2048, ratio 1.579; config=sm-tensor",
   "pp_tokens": 2048,
   "pp_tps": 1692.4,
   "provenance": "community",
   "quant": "Q4_K_XL",
   "quote": "Qwen3.8-27B-UD-Q4_K_XL (16.34 GiB = 17.545 GB) fits on one card; 2x RX 7900 XTX (gfx1100), PCIe 4.0 x8/x8 as configured in BIOS; | 2048 | 1071.89 \u00b1 30.40 | 1692.40 \u00b1 5.91 | 1.579 | 35.81 | 46.48 | 1.298 |; -p 2048 -n 512 -d 0 -b 4096 -ub 512 -fa 1 --load-mode none -r 5; Stock `f9f09f02c` (10794), no third-party patches",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28863 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28863",
   "tg_tokens": 512,
   "tps": "46.48"
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "2x RX 7900 XTX",
   "id": "lc-28863-rocm-ub4096",
   "model": "Qwen3.8-27B",
   "notes": "-ub 4096 sweep row (2-card tensor split); decode re-measure of same config; ratio 1.582, no trend toward 2.0 over 8x ubatch; config=sm-tensor",
   "pp_tokens": 2048,
   "pp_tps": 1686.89,
   "provenance": "community",
   "quant": "Q4_K_XL",
   "quote": "Qwen3.8-27B-UD-Q4_K_XL (16.34 GiB = 17.545 GB) fits on one card; 2x RX 7900 XTX (gfx1100), PCIe 4.0 x8/x8 as configured in BIOS; | 4096 | 1066.34 \u00b1 30.10 | 1686.89 \u00b1 4.78 | 1.582 | 35.77 | 46.99 | 1.314 |; -p 2048 -n 512 -d 0 -b 4096 -ub 512 -fa 1 --load-mode none -r 5; Stock `f9f09f02c` (10794), no third-party patches",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28863 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28863",
   "tg_tokens": 512,
   "tps": "46.99"
  },
  {
   "backend": "Vulkan",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "RX 7900 XTX",
   "id": "lc-28863-vulkan-solo-tg128",
   "model": "Qwen3.8-27B",
   "notes": "Vulkan single-card, +8.5% over ROCm at tg128 (ruled-out section, same model and build); tg128 test",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_XL",
   "quote": "Qwen3.8-27B-UD-Q4_K_XL (16.34 GiB = 17.545 GB) fits on one card; card 0 is a Sapphire Nitro+ Vapor-X; 2x RX 7900 XTX (gfx1100), PCIe 4.0 x8/x8 as configured in BIOS; Vulkan single-card is +8.5% over ROCm (39.07 against 36.02 at tg128), but Vulkan's two-card tensor path is worse than its own single-card result (23.36 against 39.07) and appears to serialise (pp512 at 0.55x of single-card).; Stock `f9f09f02c` (10794), no third-party patches",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28863 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28863",
   "tg_tokens": 128,
   "tps": "39.07"
  },
  {
   "backend": "Vulkan",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "2x RX 7900 XTX",
   "id": "lc-28863-vulkan-tensor-tg128",
   "model": "Qwen3.8-27B",
   "notes": "Vulkan two-card tensor path, worse than its own single-card result (23.36 vs 39.07), appears to serialise; tg128 test",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_XL",
   "quote": "Qwen3.8-27B-UD-Q4_K_XL (16.34 GiB = 17.545 GB) fits on one card; 2x RX 7900 XTX (gfx1100), PCIe 4.0 x8/x8 as configured in BIOS; Vulkan single-card is +8.5% over ROCm (39.07 against 36.02 at tg128), but Vulkan's two-card tensor path is worse than its own single-card result (23.36 against 39.07) and appears to serialise (pp512 at 0.55x of single-card).; Stock `f9f09f02c` (10794), no third-party patches",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28863 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28863",
   "tg_tokens": 128,
   "tps": "23.36"
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "RX 7900 XTX",
   "id": "lc-28863-rocm-solo-tg128",
   "model": "Qwen3.8-27B",
   "notes": "ROCm single-card tg128 control for the Vulkan comparison (ruled-out section)",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_XL",
   "quote": "Qwen3.8-27B-UD-Q4_K_XL (16.34 GiB = 17.545 GB) fits on one card; card 0 is a Sapphire Nitro+ Vapor-X; 2x RX 7900 XTX (gfx1100), PCIe 4.0 x8/x8 as configured in BIOS; Vulkan single-card is +8.5% over ROCm (39.07 against 36.02 at tg128), but Vulkan's two-card tensor path is worse than its own single-card result (23.36 against 39.07) and appears to serialise (pp512 at 0.55x of single-card).; Stock `f9f09f02c` (10794), no third-party patches",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28863 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28863",
   "tg_tokens": 128,
   "tps": "36.02"
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": 42000,
   "date": "2026-09",
   "flags": [],
   "hardware": "5x RTX3090",
   "id": "lc-28734-master-42k",
   "model": "Qwen3.8-Flash-Next",
   "notes": "single-stream decode at 42k ctx on master with the linear depth decay (QSA top-k running as a dense scan); layer split across 5 GPUs; config=master",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Ryzen 7950x on x670 MB 192MB RAM 5x RTX3090 (3 of them via USB4, but this detail is not relevant to the bug); unsloth/Qwen3.8-Flash-Next-GGUF  UD-Q4_K_XL -ngl 99, layer split across 5 GPUs,; 37.2tok/s to 50.8tok/s at 42k ctx (+37%); Base: llama.cpp master `df03399b8`, branch `levers`",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28734 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28734",
   "tps": "37.2"
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": 42000,
   "date": "2026-09",
   "flags": [],
   "hardware": "5x RTX3090",
   "id": "lc-28734-patch-42k",
   "model": "Qwen3.8-Flash-Next",
   "notes": "same 42k ctx point after the 7-patch lever fix (sparse FA path, pooled block-key cache); +37% vs master; config=lever-patches",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Ryzen 7950x on x670 MB 192MB RAM 5x RTX3090 (3 of them via USB4, but this detail is not relevant to the bug); unsloth/Qwen3.8-Flash-Next-GGUF  UD-Q4_K_XL -ngl 99, layer split across 5 GPUs,; 37.2tok/s to 50.8tok/s at 42k ctx (+37%); Base: llama.cpp master `df03399b8`, branch `levers`",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28734 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28734",
   "tps": "50.8"
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": 111000,
   "date": "2026-09",
   "flags": [],
   "hardware": "5x RTX3090",
   "id": "lc-28734-master-111k",
   "model": "Qwen3.8-Flash-Next",
   "notes": "master at 111k ctx, decayed; prefill at 111k was 609 t/s on master before the patch; config=master",
   "pp_tps": 609.0,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Ryzen 7950x on x670 MB 192MB RAM 5x RTX3090 (3 of them via USB4, but this detail is not relevant to the bug); unsloth/Qwen3.8-Flash-Next-GGUF  UD-Q4_K_XL -ngl 99, layer split across 5 GPUs,; 22.5tok/s to 48.0tok/s at 111k (+113%); (shallow essentially unchanged: 54.4 \u2192 56.5. Prefill improves as a side effect: 609 \u2192 716 t/s @111k.); Base: llama.cpp master `df03399b8`, branch `levers`",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28734 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28734",
   "tps": "22.5"
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": 111000,
   "date": "2026-09",
   "flags": [],
   "hardware": "5x RTX3090",
   "id": "lc-28734-patch-111k",
   "model": "Qwen3.8-Flash-Next",
   "notes": "111k ctx after the lever patches; prefill side effect 609 -> 716 t/s at 111k; config=lever-patches",
   "pp_tps": 716.0,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Ryzen 7950x on x670 MB 192MB RAM 5x RTX3090 (3 of them via USB4, but this detail is not relevant to the bug); unsloth/Qwen3.8-Flash-Next-GGUF  UD-Q4_K_XL -ngl 99, layer split across 5 GPUs,; 22.5tok/s to 48.0tok/s at 111k (+113%); (shallow essentially unchanged: 54.4 \u2192 56.5. Prefill improves as a side effect: 609 \u2192 716 t/s @111k.); Base: llama.cpp master `df03399b8`, branch `levers`",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28734 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28734",
   "tps": "48.0"
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": 250000,
   "date": "2026-09",
   "flags": [],
   "hardware": "5x RTX3090",
   "id": "lc-28734-master-250k",
   "model": "Qwen3.8-Flash-Next",
   "notes": "master at 250k ctx, \"nighly speed\" at the bottom of the context window; config=master",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Ryzen 7950x on x670 MB 192MB RAM 5x RTX3090 (3 of them via USB4, but this detail is not relevant to the bug); unsloth/Qwen3.8-Flash-Next-GGUF  UD-Q4_K_XL -ngl 99, layer split across 5 GPUs,; 11tok/s to 41.8tok/s at 250k (+280%); Base: llama.cpp master `df03399b8`, branch `levers`",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28734 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28734",
   "tps": "11"
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": 250000,
   "date": "2026-09",
   "flags": [],
   "hardware": "5x RTX3090",
   "id": "lc-28734-patch-250k",
   "model": "Qwen3.8-Flash-Next",
   "notes": "250k ctx after the lever patches, +280% vs master at the same depth; config=lever-patches",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Ryzen 7950x on x670 MB 192MB RAM 5x RTX3090 (3 of them via USB4, but this detail is not relevant to the bug); unsloth/Qwen3.8-Flash-Next-GGUF  UD-Q4_K_XL -ngl 99, layer split across 5 GPUs,; 11tok/s to 41.8tok/s at 250k (+280%); Base: llama.cpp master `df03399b8`, branch `levers`",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28734 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28734",
   "tps": "41.8"
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "5x RTX3090",
   "id": "lc-28734-master-shallow",
   "model": "Qwen3.8-Flash-Next",
   "notes": "shallow-context decode on master, essentially unchanged by the decay (control point); config=master",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Ryzen 7950x on x670 MB 192MB RAM 5x RTX3090 (3 of them via USB4, but this detail is not relevant to the bug); unsloth/Qwen3.8-Flash-Next-GGUF  UD-Q4_K_XL -ngl 99, layer split across 5 GPUs,; (shallow essentially unchanged: 54.4 \u2192 56.5. Prefill improves as a side effect: 609 \u2192 716 t/s @111k.); Base: llama.cpp master `df03399b8`, branch `levers`",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28734 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28734",
   "tps": "54.4"
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "5x RTX3090",
   "id": "lc-28734-patch-shallow",
   "model": "Qwen3.8-Flash-Next",
   "notes": "shallow-context decode after the lever patches; \"shallow essentially unchanged: 54.4 -> 56.5\"; config=lever-patches",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Ryzen 7950x on x670 MB 192MB RAM 5x RTX3090 (3 of them via USB4, but this detail is not relevant to the bug); unsloth/Qwen3.8-Flash-Next-GGUF  UD-Q4_K_XL -ngl 99, layer split across 5 GPUs,; (shallow essentially unchanged: 54.4 \u2192 56.5. Prefill improves as a side effect: 609 \u2192 716 t/s @111k.); Base: llama.cpp master `df03399b8`, branch `levers`",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28734 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28734",
   "tps": "56.5"
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": 98304,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX A2000 12GB",
   "id": "lc-28484-mtp-depth2",
   "model": "Qwen3.8-27B",
   "notes": "draft-mtp depth 2 over RPC tensor split to a Windows RTX 3080 10GB via 2.5GbE, commit 6a1a922 (regressed from 27-32 at bb4caa7); 1209-token prompt / 1024-token generation; config=draft-mtp",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "Hardware: Linux = i5-12600H, RTX A2000 12GB, 96GB DDR5 5600Mt; windows = RTX 3080 10GB, AMD 5800x3d, 32gb DDR4, 2.5GbE LAN between both.; Model: unsloth Qwen3.8-27B UD-IQ4_XS GGUF (14.25 GB) with embedded MTP heads.; | draft-mtp depth 2 (flags above) | 27\u201332 tok/s | 21.3 tok/s |",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28484 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28484",
   "tps": "21.3"
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": 98304,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX A2000 12GB",
   "id": "lc-28484-dense",
   "model": "Qwen3.8-27B",
   "notes": "no MTP (raw dense decode) over the same RPC split at 6a1a922; dense decode unchanged by the regression; config=nomtp",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "Hardware: Linux = i5-12600H, RTX A2000 12GB, 96GB DDR5 5600Mt; windows = RTX 3080 10GB, AMD 5800x3d, 32gb DDR4, 2.5GbE LAN between both.; Model: unsloth Qwen3.8-27B UD-IQ4_XS GGUF (14.25 GB) with embedded MTP heads.; | no MTP (raw dense decode, same split) | ~17\u201318 tok/s | 17.5 tok/s |",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28484 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28484",
   "tps": "17.5"
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": 98304,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX A2000 12GB",
   "id": "lc-28484-glimmer-control",
   "model": "Muse Glimmer",
   "notes": "control point, 31B dense model without speculative decoding, same RPC split, identical on both commits",
   "pp_tps": null,
   "provenance": "community",
   "quant": "as-published",
   "quote": "Hardware: Linux = i5-12600H, RTX A2000 12GB, 96GB DDR5 5600Mt; windows = RTX 3080 10GB, AMD 5800x3d, 32gb DDR4, 2.5GbE LAN between both.; Model: unsloth Qwen3.8-27B UD-IQ4_XS GGUF (14.25 GB) with embedded MTP heads.; a 31B dense model (Muse Glimmer) (no speculative decoding) on the identical RPC split, flags, and both commits measures 17.66 tok/s before and after",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28484 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28484",
   "tps": "17.66"
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": 152000,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 5090 Laptop",
   "id": "lc-28790-msvc-mtp",
   "model": "Qwen3.8-27B-NVFP4-MTP-LOW",
   "notes": "self-compiled MSVC 19.44 + CUDA 12.8 build (master 43f3dda62, b10915) with draft-mtp; prefill collapses to ~decode speed on this build (57x slower); config=msvc-mtp",
   "pp_tokens": 4800,
   "pp_tps": 32.7,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "- **GPU**: RTX 5090 Laptop, 24GB GDDR7, sm_120; - **Model**: `Qwen3.8-27B-NVFP4-MTP-LOW.gguf` (15.5 GB, Qwen3.8 family, VLM + MTP head) + `mmproj-Q8_0.gguf`; | Build | prefill, 4800-token prompt | decode (256 tok) |; | **self-compiled + MTP** | **32.7 tok/s** | 77.1 tok/s |; -ngl 99 -fa on -fit off -c 152000",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28790 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28790",
   "tg_tokens": 256,
   "tps": "77.1"
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": 152000,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 5090 Laptop",
   "id": "lc-28790-msvc-nomtp",
   "model": "Qwen3.8-27B-NVFP4-MTP-LOW",
   "notes": "same self-compiled build with MTP disabled; the prefill path is healthy here; config=msvc-nomtp",
   "pp_tokens": 4800,
   "pp_tps": 1867.5,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "- **GPU**: RTX 5090 Laptop, 24GB GDDR7, sm_120; - **Model**: `Qwen3.8-27B-NVFP4-MTP-LOW.gguf` (15.5 GB, Qwen3.8 family, VLM + MTP head) + `mmproj-Q8_0.gguf`; | Build | prefill, 4800-token prompt | decode (256 tok) |; | **self-compiled, MTP disabled** | **1867.5 tok/s** | 39.3 tok/s |; -ngl 99 -fa on -fit off -c 152000",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28790 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28790",
   "tg_tokens": 256,
   "tps": "39.3"
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": 152000,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 5090 Laptop",
   "id": "lc-28790-b10917-mtp",
   "model": "Qwen3.8-27B-NVFP4-MTP-LOW",
   "notes": "official Windows prebuilt b10917 (Clang + CUDA 13.3) with MTP; no slowdown, isolates the fault to the MSVC + CUDA 12.8 toolchain; config=b10917-mtp",
   "pp_tokens": 4800,
   "pp_tps": 1344.9,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "- **GPU**: RTX 5090 Laptop, 24GB GDDR7, sm_120; - **Model**: `Qwen3.8-27B-NVFP4-MTP-LOW.gguf` (15.5 GB, Qwen3.8 family, VLM + MTP head) + `mmproj-Q8_0.gguf`; | Build | prefill, 4800-token prompt | decode (256 tok) |; | official b10917 (Clang + CUDA 13.3) + MTP | 1344.9 tok/s | 50.3 tok/s |; -ngl 99 -fa on -fit off -c 152000",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28790 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28790",
   "tg_tokens": 256,
   "tps": "50.3"
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": 152000,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 5090 Laptop",
   "id": "lc-28790-b10889-mtp",
   "model": "Qwen3.8-27B-NVFP4-MTP-LOW",
   "notes": "official Windows prebuilt b10889 (Clang + CUDA 13.x) with MTP; second official build confirms; config=b10889-mtp",
   "pp_tokens": 4800,
   "pp_tps": 1482.0,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "- **GPU**: RTX 5090 Laptop, 24GB GDDR7, sm_120; - **Model**: `Qwen3.8-27B-NVFP4-MTP-LOW.gguf` (15.5 GB, Qwen3.8 family, VLM + MTP head) + `mmproj-Q8_0.gguf`; | Build | prefill, 4800-token prompt | decode (256 tok) |; | official b10889 (Clang + CUDA 13.x) + MTP | 1482 tok/s | 51.7 tok/s |; -ngl 99 -fa on -fit off -c 152000",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28790 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28790",
   "tg_tokens": 256,
   "tps": "51.7"
  },
  {
   "backend": "Metal",
   "batch": 1,
   "ctx": 256,
   "date": "2026-09",
   "flags": [],
   "hardware": "Apple M3 Ultra",
   "id": "lc-29335-q4exp-c256-n1",
   "model": "Qwen3.8-Flash-Next",
   "notes": "aggregate decode tok/s, N=1, 256 ctx; b10830 Unsloth prebuilt, Metal, --kv-unified, unique prompts, 128 generated tokens each",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Apple M3 Ultra (Mac15,14), 512 GB unified memory, macOS 26.6.2 (25G83); `unsloth/Qwen3.8-Flash-Next-GGUF`, Q8_0 (arch `qwen4exp`, 512-expert MoE, 10 active). Also seen with GLM-5.3-Flash UD-Q4_K_XL (`glm5next`, see below).; | 256 | 32.8 | 45 | 60 | **78** | 33 \u2192 112 |; 128 generated tokens, thinking off",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #29335 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29335",
   "tg_tokens": 128,
   "tps": "32.8"
  },
  {
   "backend": "Metal",
   "batch": 2,
   "ctx": 256,
   "date": "2026-09",
   "flags": [],
   "hardware": "Apple M3 Ultra",
   "id": "lc-29335-q4exp-c256-n2",
   "model": "Qwen3.8-Flash-Next",
   "notes": "aggregate decode tok/s, N=2, 256 ctx; batching works as expected at short context",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Apple M3 Ultra (Mac15,14), 512 GB unified memory, macOS 26.6.2 (25G83); `unsloth/Qwen3.8-Flash-Next-GGUF`, Q8_0 (arch `qwen4exp`, 512-expert MoE, 10 active). Also seen with GLM-5.3-Flash UD-Q4_K_XL (`glm5next`, see below).; | 256 | 32.8 | 45 | 60 | **78** | 33 \u2192 112 |; 128 generated tokens, thinking off",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #29335 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29335",
   "tg_tokens": 128,
   "tps": "45"
  },
  {
   "backend": "Metal",
   "batch": 4,
   "ctx": 256,
   "date": "2026-09",
   "flags": [],
   "hardware": "Apple M3 Ultra",
   "id": "lc-29335-q4exp-c256-n4",
   "model": "Qwen3.8-Flash-Next",
   "notes": "aggregate decode tok/s, N=4, 256 ctx",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Apple M3 Ultra (Mac15,14), 512 GB unified memory, macOS 26.6.2 (25G83); `unsloth/Qwen3.8-Flash-Next-GGUF`, Q8_0 (arch `qwen4exp`, 512-expert MoE, 10 active). Also seen with GLM-5.3-Flash UD-Q4_K_XL (`glm5next`, see below).; | 256 | 32.8 | 45 | 60 | **78** | 33 \u2192 112 |; 128 generated tokens, thinking off",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #29335 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29335",
   "tg_tokens": 128,
   "tps": "60"
  },
  {
   "backend": "Metal",
   "batch": 8,
   "ctx": 256,
   "date": "2026-09",
   "flags": [],
   "hardware": "Apple M3 Ultra",
   "id": "lc-29335-q4exp-c256-n8",
   "model": "Qwen3.8-Flash-Next",
   "notes": "aggregate decode tok/s, N=8, 256 ctx; ms per decode pass climbs 33 -> 112 across N",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Apple M3 Ultra (Mac15,14), 512 GB unified memory, macOS 26.6.2 (25G83); `unsloth/Qwen3.8-Flash-Next-GGUF`, Q8_0 (arch `qwen4exp`, 512-expert MoE, 10 active). Also seen with GLM-5.3-Flash UD-Q4_K_XL (`glm5next`, see below).; | 256 | 32.8 | 45 | 60 | **78** | 33 \u2192 112 |; 128 generated tokens, thinking off",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #29335 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29335",
   "tg_tokens": 128,
   "tps": "78"
  },
  {
   "backend": "Metal",
   "batch": 8,
   "ctx": 4000,
   "date": "2026-09",
   "flags": [],
   "hardware": "Apple M3 Ultra",
   "id": "lc-29335-q4exp-c4k-n8",
   "model": "Qwen3.8-Flash-Next",
   "notes": "aggregate decode tok/s, N=8, 4k ctx",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Apple M3 Ultra (Mac15,14), 512 GB unified memory, macOS 26.6.2 (25G83); `unsloth/Qwen3.8-Flash-Next-GGUF`, Q8_0 (arch `qwen4exp`, 512-expert MoE, 10 active). Also seen with GLM-5.3-Flash UD-Q4_K_XL (`glm5next`, see below).; | 4 k | \u2014 | \u2014 | \u2014 | 23 | |; 128 generated tokens, thinking off",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #29335 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29335",
   "tg_tokens": 128,
   "tps": "23"
  },
  {
   "backend": "Metal",
   "batch": 4,
   "ctx": 8000,
   "date": "2026-09",
   "flags": [],
   "hardware": "Apple M3 Ultra",
   "id": "lc-29335-q4exp-c8k-n4",
   "model": "Qwen3.8-Flash-Next",
   "notes": "aggregate decode tok/s, N=4, 8k ctx",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Apple M3 Ultra (Mac15,14), 512 GB unified memory, macOS 26.6.2 (25G83); `unsloth/Qwen3.8-Flash-Next-GGUF`, Q8_0 (arch `qwen4exp`, 512-expert MoE, 10 active). Also seen with GLM-5.3-Flash UD-Q4_K_XL (`glm5next`, see below).; | 8 k | \u2014 | \u2014 | 23.5 | 11.6 | |; 128 generated tokens, thinking off",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #29335 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29335",
   "tg_tokens": 128,
   "tps": "23.5"
  },
  {
   "backend": "Metal",
   "batch": 8,
   "ctx": 8000,
   "date": "2026-09",
   "flags": [],
   "hardware": "Apple M3 Ultra",
   "id": "lc-29335-q4exp-c8k-n8",
   "model": "Qwen3.8-Flash-Next",
   "notes": "aggregate decode tok/s, N=8, 8k ctx; collapse worse than linear in sequence count",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Apple M3 Ultra (Mac15,14), 512 GB unified memory, macOS 26.6.2 (25G83); `unsloth/Qwen3.8-Flash-Next-GGUF`, Q8_0 (arch `qwen4exp`, 512-expert MoE, 10 active). Also seen with GLM-5.3-Flash UD-Q4_K_XL (`glm5next`, see below).; | 8 k | \u2014 | \u2014 | 23.5 | 11.6 | |; 128 generated tokens, thinking off",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #29335 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29335",
   "tg_tokens": 128,
   "tps": "11.6"
  },
  {
   "backend": "Metal",
   "batch": 1,
   "ctx": 16000,
   "date": "2026-09",
   "flags": [],
   "hardware": "Apple M3 Ultra",
   "id": "lc-29335-q4exp-c16k-n1",
   "model": "Qwen3.8-Flash-Next",
   "notes": "aggregate decode tok/s, N=1, 16k ctx; at 16k and above N=1 gives the highest aggregate",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Apple M3 Ultra (Mac15,14), 512 GB unified memory, macOS 26.6.2 (25G83); `unsloth/Qwen3.8-Flash-Next-GGUF`, Q8_0 (arch `qwen4exp`, 512-expert MoE, 10 active). Also seen with GLM-5.3-Flash UD-Q4_K_XL (`glm5next`, see below).; | 16 k | **25.9** | 10.3 | 6.9 | **5.2** | 39 \u2192 195 \u2192 580 \u2192 ~1500 |; 128 generated tokens, thinking off",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #29335 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29335",
   "tg_tokens": 128,
   "tps": "25.9"
  },
  {
   "backend": "Metal",
   "batch": 2,
   "ctx": 16000,
   "date": "2026-09",
   "flags": [],
   "hardware": "Apple M3 Ultra",
   "id": "lc-29335-q4exp-c16k-n2",
   "model": "Qwen3.8-Flash-Next",
   "notes": "aggregate decode tok/s, N=2, 16k ctx",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Apple M3 Ultra (Mac15,14), 512 GB unified memory, macOS 26.6.2 (25G83); `unsloth/Qwen3.8-Flash-Next-GGUF`, Q8_0 (arch `qwen4exp`, 512-expert MoE, 10 active). Also seen with GLM-5.3-Flash UD-Q4_K_XL (`glm5next`, see below).; | 16 k | **25.9** | 10.3 | 6.9 | **5.2** | 39 \u2192 195 \u2192 580 \u2192 ~1500 |; 128 generated tokens, thinking off",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #29335 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29335",
   "tg_tokens": 128,
   "tps": "10.3"
  },
  {
   "backend": "Metal",
   "batch": 4,
   "ctx": 16000,
   "date": "2026-09",
   "flags": [],
   "hardware": "Apple M3 Ultra",
   "id": "lc-29335-q4exp-c16k-n4",
   "model": "Qwen3.8-Flash-Next",
   "notes": "aggregate decode tok/s, N=4, 16k ctx",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Apple M3 Ultra (Mac15,14), 512 GB unified memory, macOS 26.6.2 (25G83); `unsloth/Qwen3.8-Flash-Next-GGUF`, Q8_0 (arch `qwen4exp`, 512-expert MoE, 10 active). Also seen with GLM-5.3-Flash UD-Q4_K_XL (`glm5next`, see below).; | 16 k | **25.9** | 10.3 | 6.9 | **5.2** | 39 \u2192 195 \u2192 580 \u2192 ~1500 |; 128 generated tokens, thinking off",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #29335 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29335",
   "tg_tokens": 128,
   "tps": "6.9"
  },
  {
   "backend": "Metal",
   "batch": 8,
   "ctx": 16000,
   "date": "2026-09",
   "flags": [],
   "hardware": "Apple M3 Ultra",
   "id": "lc-29335-q4exp-c16k-n8",
   "model": "Qwen3.8-Flash-Next",
   "notes": "aggregate decode tok/s, N=8, 16k ctx; ms per decode pass 39 -> ~1500 across N",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Apple M3 Ultra (Mac15,14), 512 GB unified memory, macOS 26.6.2 (25G83); `unsloth/Qwen3.8-Flash-Next-GGUF`, Q8_0 (arch `qwen4exp`, 512-expert MoE, 10 active). Also seen with GLM-5.3-Flash UD-Q4_K_XL (`glm5next`, see below).; | 16 k | **25.9** | 10.3 | 6.9 | **5.2** | 39 \u2192 195 \u2192 580 \u2192 ~1500 |; 128 generated tokens, thinking off",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #29335 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29335",
   "tg_tokens": 128,
   "tps": "5.2"
  },
  {
   "backend": "Metal",
   "batch": 1,
   "ctx": 32000,
   "date": "2026-09",
   "flags": [],
   "hardware": "Apple M3 Ultra",
   "id": "lc-29335-q4exp-c32k-n1",
   "model": "Qwen3.8-Flash-Next",
   "notes": "aggregate decode tok/s, N=1, 32k ctx; matches the 1-process row of the multi-process table",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Apple M3 Ultra (Mac15,14), 512 GB unified memory, macOS 26.6.2 (25G83); `unsloth/Qwen3.8-Flash-Next-GGUF`, Q8_0 (arch `qwen4exp`, 512-expert MoE, 10 active). Also seen with GLM-5.3-Flash UD-Q4_K_XL (`glm5next`, see below).; | 32 k | **25.7** | 8.6 | 5.5 | \u2014 | 39 \u2192 233 \u2192 730 |; 128 generated tokens, thinking off",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #29335 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29335",
   "tg_tokens": 128,
   "tps": "25.7"
  },
  {
   "backend": "Metal",
   "batch": 2,
   "ctx": 32000,
   "date": "2026-09",
   "flags": [],
   "hardware": "Apple M3 Ultra",
   "id": "lc-29335-q4exp-c32k-n2",
   "model": "Qwen3.8-Flash-Next",
   "notes": "aggregate decode tok/s, N=2, 32k ctx; per stream 4.3",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Apple M3 Ultra (Mac15,14), 512 GB unified memory, macOS 26.6.2 (25G83); `unsloth/Qwen3.8-Flash-Next-GGUF`, Q8_0 (arch `qwen4exp`, 512-expert MoE, 10 active). Also seen with GLM-5.3-Flash UD-Q4_K_XL (`glm5next`, see below).; | 32 k | **25.7** | 8.6 | 5.5 | \u2014 | 39 \u2192 233 \u2192 730 |; 128 generated tokens, thinking off",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #29335 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29335",
   "tg_tokens": 128,
   "tps": "8.6"
  },
  {
   "backend": "Metal",
   "batch": 4,
   "ctx": 32000,
   "date": "2026-09",
   "flags": [],
   "hardware": "Apple M3 Ultra",
   "id": "lc-29335-q4exp-c32k-n4",
   "model": "Qwen3.8-Flash-Next",
   "notes": "aggregate decode tok/s, N=4, 32k ctx; per stream 1.4",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Apple M3 Ultra (Mac15,14), 512 GB unified memory, macOS 26.6.2 (25G83); `unsloth/Qwen3.8-Flash-Next-GGUF`, Q8_0 (arch `qwen4exp`, 512-expert MoE, 10 active). Also seen with GLM-5.3-Flash UD-Q4_K_XL (`glm5next`, see below).; | 32 k | **25.7** | 8.6 | 5.5 | \u2014 | 39 \u2192 233 \u2192 730 |; 128 generated tokens, thinking off",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #29335 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29335",
   "tg_tokens": 128,
   "tps": "5.5"
  },
  {
   "backend": "Metal",
   "batch": 4,
   "ctx": 256,
   "date": "2026-09",
   "flags": [],
   "hardware": "Apple M3 Ultra",
   "id": "lc-29335-glm53-c256-n4",
   "model": "GLM-5.3-Flash",
   "notes": "aggregate decode tok/s, N=4, 256 ctx, GLM-5.3-Flash (glm5next); same cliff on a different architecture",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Apple M3 Ultra (Mac15,14), 512 GB unified memory, macOS 26.6.2 (25G83); `unsloth/Qwen3.8-Flash-Next-GGUF`, Q8_0 (arch `qwen4exp`, 512-expert MoE, 10 active). Also seen with GLM-5.3-Flash UD-Q4_K_XL (`glm5next`, see below).; GLM-5.3-Flash (`glm5next`, served from the same prebuilt; I know it's still an open PR upstream, #27752) goes from 40 to 10.6 tok/s aggregate at N=4 between 256 and 8 k context; 128 generated tokens, thinking off",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #29335 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29335",
   "tg_tokens": 128,
   "tps": "40"
  },
  {
   "backend": "Metal",
   "batch": 4,
   "ctx": 8000,
   "date": "2026-09",
   "flags": [],
   "hardware": "Apple M3 Ultra",
   "id": "lc-29335-glm53-c8k-n4",
   "model": "GLM-5.3-Flash",
   "notes": "aggregate decode tok/s, N=4, 8k ctx, GLM-5.3-Flash; goes from 40 to 10.6 between 256 and 8k context",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Apple M3 Ultra (Mac15,14), 512 GB unified memory, macOS 26.6.2 (25G83); `unsloth/Qwen3.8-Flash-Next-GGUF`, Q8_0 (arch `qwen4exp`, 512-expert MoE, 10 active). Also seen with GLM-5.3-Flash UD-Q4_K_XL (`glm5next`, see below).; GLM-5.3-Flash (`glm5next`, served from the same prebuilt; I know it's still an open PR upstream, #27752) goes from 40 to 10.6 tok/s aggregate at N=4 between 256 and 8 k context; 128 generated tokens, thinking off",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #29335 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29335",
   "tg_tokens": 128,
   "tps": "10.6"
  },
  {
   "backend": "Metal",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-09",
   "flags": [],
   "hardware": "2x Apple M1 Max",
   "id": "lc-28805-m1max-64k",
   "model": "Qwen3.8-Flash-Next",
   "notes": "retest on master df03399 (build 300) at 64K prompt, single slot, GPU FA; IQ4_XS 28-shard unsloth-style quant (3.84 bpw); reproduced on both M1 Max boards",
   "pp_tps": 153.8,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "2x Apple M1 Max (24-core and 32-core GPU), 64 GB unified memory each \u2014 reproduced on both; unsloth-style AD-3.84bpw-IQ4_XS-M64 (28 shards) \u2014 16K/64K OK, 128K FAIL; 64K x1 = 3/3 (prefill 153.8 tok/s, decode 11.1 \u2014 sparse-FA perf improvements visible); qwen4exp (Qwen3.8-Flash-Next)",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28805 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28805",
   "tps": "11.1"
  },
  {
   "backend": "Vulkan",
   "batch": 1,
   "ctx": 786432,
   "date": "2026-09",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-28433-ngram-c786k",
   "model": "Qwen3.8-27B",
   "notes": "Vulkan (RADV STRIX_HALO, gfx1151), --ctx-size 786432 --parallel 3, ngram-mod only (no draft-mtp); b10752 container",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "AMD Ryzen AI Max+ 395 \u2014 Radeon 8060S Graphics (RADV STRIX_HALO, gfx1151),; `Qwen3.8-27B-Q4_K_M.gguf` (dense, 27.3B params, `n_ctx_train = 262144`), which; Vulkan (Mesa RADV); | 786432 | 3 | off | `ngram-mod` | \u2014 | works (12.8 tok/s, 99.99% cache reuse) |",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28433 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28433",
   "tps": "12.8"
  },
  {
   "backend": "Vulkan",
   "batch": 1,
   "ctx": 393216,
   "date": "2026-09",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-28433-mtp-c393k",
   "model": "Qwen3.8-27B",
   "notes": "Vulkan, --ctx-size 393216 --parallel 3, ngram-mod + draft-mtp; the MTP draft-context sizing bug does not trigger at this size",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "AMD Ryzen AI Max+ 395 \u2014 Radeon 8060S Graphics (RADV STRIX_HALO, gfx1151),; `Qwen3.8-27B-Q4_K_M.gguf` (dense, 27.3B params, `n_ctx_train = 262144`), which; Vulkan (Mesa RADV); | 393216 | 3 | off | `ngram-mod,draft-mtp` | 393216 | works (24 tok/s) |",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28433 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28433",
   "tps": "24"
  },
  {
   "backend": "Vulkan",
   "batch": 1,
   "ctx": 524288,
   "date": "2026-09",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-28433-mtp-c524k-unified",
   "model": "Qwen3.8-27B",
   "notes": "Vulkan, --ctx-size 524288 --parallel 3, kv-unified with --kv-unified-per-slot 262144, ngram-mod + draft-mtp",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "AMD Ryzen AI Max+ 395 \u2014 Radeon 8060S Graphics (RADV STRIX_HALO, gfx1151),; `Qwen3.8-27B-Q4_K_M.gguf` (dense, 27.3B params, `n_ctx_train = 262144`), which; Vulkan (Mesa RADV); | 524288 | 3 | on + `--kv-unified-per-slot 262144` | `ngram-mod,draft-mtp` | 524288 | works (23.7 tok/s) |",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28433 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28433",
   "tps": "23.7"
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": 1048576,
   "date": "2026-09",
   "flags": [],
   "hardware": "4x RTX A6000",
   "id": "lc-28454-a6000-kvf16",
   "model": "DeepSeek-V4-Flash",
   "notes": "single-stream decode with f16 KV (default); ~ value as published; sparse-fa mma_f16 path reachable with non-quantized KV; 1M ctx; config=kv-f16",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_K_XL",
   "quote": "- GPU: 4x RTX A6000 (Ampere SM86), tensor parallel (`-sm tensor`); - Model: DeepSeek-V4-Flash-0731, GGUF `Q8_K_XL`; - Launch: `-c 1048576 -np 2 -sm tensor -fa on -ctk q8_0 -ctv q8_0`; | f16 (default) | ~33 tok/s (same setup, prior measurement) |",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28454 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28454",
   "tps": "33"
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": 1048576,
   "date": "2026-09",
   "flags": [],
   "hardware": "4x RTX A6000",
   "id": "lc-28454-a6000-kvq8",
   "model": "DeepSeek-V4-Flash",
   "notes": "single-stream decode with -ctk q8_0 -ctv q8_0; ~2.8x slowdown because quantized KV never reaches the sparse-fa fused kernel; ~ value as published; config=kv-q8",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_K_XL",
   "quote": "- GPU: 4x RTX A6000 (Ampere SM86), tensor parallel (`-sm tensor`); - Model: DeepSeek-V4-Flash-0731, GGUF `Q8_K_XL`; - Launch: `-c 1048576 -np 2 -sm tensor -fa on -ctk q8_0 -ctv q8_0`; | q8_0 / q8_0 | ~11.7 tok/s |",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28454 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28454",
   "tps": "11.7"
  },
  {
   "backend": "Vulkan",
   "batch": 1,
   "ctx": 98304,
   "date": "2026-09",
   "flags": [],
   "hardware": "AMD GPU via the Vulkan backend",
   "id": "lc-29175-vulkan-b4096",
   "model": "Qwen3.8-27B",
   "notes": "solo decode on the Vulkan host, -b 4096 --ubatch-size 2048 --parallel 8; KV q4_0, no flash-attn, b10724",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q5_K",
   "quote": "Model: Qwen3.8-27B, Q5_K quant, `--cache-type-k q4_0 --cache-type-v q4_0`; Host B: AMD GPU via the Vulkan backend, `--ctx-size 98304`; **Solo decode throughput is unchanged**: 24.24 \u2192 24.86 tok/s on the Vulkan host across the same change.; Raising `-b` from 4096 to 16384 on a `--parallel 8 --ubatch-size 2048` host",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #29175 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29175",
   "tps": "24.24"
  },
  {
   "backend": "Vulkan",
   "batch": 1,
   "ctx": 98304,
   "date": "2026-09",
   "flags": [],
   "hardware": "AMD GPU via the Vulkan backend",
   "id": "lc-29175-vulkan-b16384",
   "model": "Qwen3.8-27B",
   "notes": "solo decode on the Vulkan host after raising -b to 16384; throughput unchanged, only the stall count moves",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q5_K",
   "quote": "Model: Qwen3.8-27B, Q5_K quant, `--cache-type-k q4_0 --cache-type-v q4_0`; Host B: AMD GPU via the Vulkan backend, `--ctx-size 98304`; **Solo decode throughput is unchanged**: 24.24 \u2192 24.86 tok/s on the Vulkan host across the same change.; Raising `-b` from 4096 to 16384 on a `--parallel 8 --ubatch-size 2048` host",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #29175 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29175",
   "tps": "24.86"
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": 262144,
   "date": "2026-09",
   "flags": [],
   "hardware": "Quadro RTX 8000",
   "id": "lc-29175-quadro-before",
   "model": "Qwen3.8-27B",
   "notes": "solo decode on the SM 7.5 host before a single 159,437-token request; deep-KV collapse from #29172 tested and not reproduced",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q5_K",
   "quote": "Model: Qwen3.8-27B, Q5_K quant, `--cache-type-k q4_0 --cache-type-v q4_0`; Host A: Quadro RTX 8000 (Turing, SM 7.5), CUDA, `--ctx-size 262144`; **It did not reproduce.** Solo decode 20.25 tok/s \u2192 one **159,437-token** request (541.8 s) \u2192 solo decode **19.32 tok/s**, i.e. **0.95\u00d7**.",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #29175 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29175",
   "tps": "20.25"
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": 262144,
   "date": "2026-09",
   "flags": [],
   "hardware": "Quadro RTX 8000",
   "id": "lc-29175-quadro-after",
   "model": "Qwen3.8-27B",
   "notes": "solo decode after one 159,437-token request (541.8 s), 0.95x of the prior solo rate",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q5_K",
   "quote": "Model: Qwen3.8-27B, Q5_K quant, `--cache-type-k q4_0 --cache-type-v q4_0`; Host A: Quadro RTX 8000 (Turing, SM 7.5), CUDA, `--ctx-size 262144`; **It did not reproduce.** Solo decode 20.25 tok/s \u2192 one **159,437-token** request (541.8 s) \u2192 solo decode **19.32 tok/s**, i.e. **0.95\u00d7**.",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #29175 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29175",
   "tps": "19.32"
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": 262144,
   "date": "2026-09",
   "flags": [],
   "hardware": "8\u00d7 NVIDIA RTX 3080 20GB",
   "id": "lc-29345-mimo-8x3080",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "control decode without MTP, 40/48 expert layers offloaded across the 8 cards (-ts 1,1,1,1,1,1,1,1); ~ value as published; -c 262144 -fa on -b 4096 -ub 1024",
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "8\u00d7 NVIDIA RTX 3080 20GB (sm86), dual NUMA; main: MiMo-V2.6-Flash-RL-MXFP4; The main model loads and runs fine without the MTP flags (~18.6 tok/s decode; on this box with 40/48 expert layers offloaded to GPU",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #29345 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29345",
   "tps": "18.6"
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-28219-b10734-baseline",
   "model": "Qwen3.8-27B",
   "notes": "official b10734, MTP, no -ts: baseline, the fast configuration; config=official-mtp",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ3_S",
   "quote": "- Windows 11 26200, RTX 5060 Ti 16 GB (sm_120), CUDA driver 13.2, 24-thread CPU; - Model: Qwen3.8-27B-UD-IQ3_S (full GPU, no sysmem fallback; -fa on, -c 65536, -ctk q4_0 -ctv q4_0 (and -ctkd q4_0 -ctvd q4_0 with MTP); | 1 | official b10734 | MTP, no `-ts` | **38.8 tok/s** | baseline |; Speculative decoding (MTP and draft-dflash) collapses to a constant ~2.0 tok/s \u2014 a suspiciously exact 500 ms per token \u2014 in two independently reproducible configurations on a single-GPU Windows 11 machine.",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28219 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28219",
   "tps": "38.8"
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-28219-b10734-ts1",
   "model": "Qwen3.8-27B",
   "notes": "official b10734, MTP + -ts 1: the cliff, ~500 ms per token; config=official-ts1",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ3_S",
   "quote": "- Windows 11 26200, RTX 5060 Ti 16 GB (sm_120), CUDA driver 13.2, 24-thread CPU; - Model: Qwen3.8-27B-UD-IQ3_S (full GPU, no sysmem fallback; -fa on, -c 65536, -ctk q4_0 -ctv q4_0 (and -ctkd q4_0 -ctvd q4_0 with MTP); | 2 | official b10734 | MTP + `-ts 1` | **1.67 tok/s** | cliff |; Speculative decoding (MTP and draft-dflash) collapses to a constant ~2.0 tok/s \u2014 a suspiciously exact 500 ms per token \u2014 in two independently reproducible configurations on a single-GPU Windows 11 machine.",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28219 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28219",
   "tps": "1.67"
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-28219-msvc-mtp",
   "model": "Qwen3.8-27B",
   "notes": "local MSVC build, MTP: same ~500 ms/token cliff; exact 500 ms stepping in per-3s log lines; config=msvc-mtp",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ3_S",
   "quote": "- Windows 11 26200, RTX 5060 Ti 16 GB (sm_120), CUDA driver 13.2, 24-thread CPU; - Model: Qwen3.8-27B-UD-IQ3_S (full GPU, no sysmem fallback; -fa on, -c 65536, -ctk q4_0 -ctv q4_0 (and -ctkd q4_0 -ctvd q4_0 with MTP); | 3 | local MSVC build | MTP | **2.01 tok/s** | cliff |; Speculative decoding (MTP and draft-dflash) collapses to a constant ~2.0 tok/s \u2014 a suspiciously exact 500 ms per token \u2014 in two independently reproducible configurations on a single-GPU Windows 11 machine.",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28219 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28219",
   "tps": "2.01"
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-28219-msvc-dflash",
   "model": "Qwen3.8-27B",
   "notes": "local MSVC build, draft-dflash (DFlash2) instead of MTP: identical cliff, so it is not MTP-specific; config=msvc-dflash",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ3_S",
   "quote": "- Windows 11 26200, RTX 5060 Ti 16 GB (sm_120), CUDA driver 13.2, 24-thread CPU; - Model: Qwen3.8-27B-UD-IQ3_S (full GPU, no sysmem fallback; -fa on, -c 65536, -ctk q4_0 -ctv q4_0 (and -ctkd q4_0 -ctvd q4_0 with MTP); | 4 | local MSVC build | draft-dflash (DFlash2) | **2.01 tok/s** | cliff |; Speculative decoding (MTP and draft-dflash) collapses to a constant ~2.0 tok/s \u2014 a suspiciously exact 500 ms per token \u2014 in two independently reproducible configurations on a single-GPU Windows 11 machine.",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28219 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28219",
   "tps": "2.01"
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-28219-msvc-nospec",
   "model": "Qwen3.8-27B",
   "notes": "local MSVC build with no speculative decoding: normal speed restored on the same build; config=msvc-nospec",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ3_S",
   "quote": "- Windows 11 26200, RTX 5060 Ti 16 GB (sm_120), CUDA driver 13.2, 24-thread CPU; - Model: Qwen3.8-27B-UD-IQ3_S (full GPU, no sysmem fallback; -fa on, -c 65536, -ctk q4_0 -ctv q4_0 (and -ctkd q4_0 -ctvd q4_0 with MTP); | 5 | local MSVC build | no speculative decoding | **26.55 tok/s** | normal |; Speculative decoding (MTP and draft-dflash) collapses to a constant ~2.0 tok/s \u2014 a suspiciously exact 500 ms per token \u2014 in two independently reproducible configurations on a single-GPU Windows 11 machine.",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28219 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28219",
   "tps": "26.55"
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": 128000,
   "date": "2026-09",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-28219-b10734-128k",
   "model": "Qwen3.8-27B",
   "notes": "official b10734 at 128K ctx, no speculative decoding; range 19-28 tok/s as published, lower bound recorded; config=official-128k",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ3_S",
   "quote": "- Windows 11 26200, RTX 5060 Ti 16 GB (sm_120), CUDA driver 13.2, 24-thread CPU; - Model: Qwen3.8-27B-UD-IQ3_S (full GPU, no sysmem fallback; -fa on, -c 65536, -ctk q4_0 -ctv q4_0 (and -ctkd q4_0 -ctvd q4_0 with MTP); | 6 | official b10734 | 128K ctx, no speculative decoding | 19\u201328 tok/s | normal |; Speculative decoding (MTP and draft-dflash) collapses to a constant ~2.0 tok/s \u2014 a suspiciously exact 500 ms per token \u2014 in two independently reproducible configurations on a single-GPU Windows 11 machine.",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28219 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28219",
   "tps": "19"
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-26",
   "flags": [],
   "hardware": "Snapdragon 7 Gen 4 (SM7750)",
   "id": "lc-29473-1",
   "model": "Llama-3.2-1B",
   "notes": "CPU (-dev none, -ngl 0), correct output; -t 4; build 86a24a1; threads=4",
   "pp_tokens": 512,
   "pp_tps": 216.3,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Motorola Edge 70 (roadstr, XT2601-2), Qualcomm Snapdragon 7 Gen 4 (SM7750), Hexagon v73, 12 GB LPDDR5X; Llama-3.2-1B Q4_0, -t 4 pp512 t/s tg64 t/s output; CPU 216.3 34.4 correct",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #29473 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29473",
   "tg_tokens": 64,
   "tps": "34.4"
  },
  {
   "backend": "llama.cpp (Hexagon)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-26",
   "flags": [],
   "hardware": "Snapdragon 7 Gen 4 (SM7750)",
   "id": "lc-29473-2",
   "model": "Llama-3.2-1B",
   "notes": "HTP0 default HMX path, garbled output (MUL_MAT inf for n>=5); -t 4; build 86a24a1; config=hexagon-hmx; threads=4",
   "pp_tokens": 512,
   "pp_tps": 1749.4,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Motorola Edge 70 (roadstr, XT2601-2), Qualcomm Snapdragon 7 Gen 4 (SM7750), Hexagon v73, 12 GB LPDDR5X; Llama-3.2-1B Q4_0, -t 4 pp512 t/s tg64 t/s output; HTP0, default (HMX) 1749.4 29.9 garbled",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #29473 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29473",
   "tg_tokens": 64,
   "tps": "29.9"
  },
  {
   "backend": "llama.cpp (Hexagon)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-26",
   "flags": [],
   "hardware": "Snapdragon 7 Gen 4 (SM7750)",
   "id": "lc-29473-3",
   "model": "Llama-3.2-1B",
   "notes": "HTP0 with GGML_HEXAGON_MM_SELECT=1 (HVX only) + OPFILTER and -fa off, correct output, slower than CPU; -t 4; build 86a24a1; config=hexagon-workaround; threads=4",
   "pp_tokens": 512,
   "pp_tps": 78.7,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Motorola Edge 70 (roadstr, XT2601-2), Qualcomm Snapdragon 7 Gen 4 (SM7750), Hexagon v73, 12 GB LPDDR5X; Llama-3.2-1B Q4_0, -t 4 pp512 t/s tg64 t/s output; HTP0, workaround above 78.7 27.0 correct",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #29473 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29473",
   "tg_tokens": 64,
   "tps": "27.0"
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-26",
   "flags": [],
   "hardware": "Snapdragon 7 Gen 4 (SM7750)",
   "id": "lc-29473-4",
   "model": "Qwen3.5-4B",
   "notes": "CPU (-dev none, -ngl 0), correct output; -t 4; build 86a24a1; threads=4",
   "pp_tokens": 512,
   "pp_tps": 46.3,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Motorola Edge 70 (roadstr, XT2601-2), Qualcomm Snapdragon 7 Gen 4 (SM7750), Hexagon v73, 12 GB LPDDR5X; Qwen3.5-4B Q4_0, -t 4 pp512 t/s tg64 t/s output; CPU 46.3 9.1 correct",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #29473 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29473",
   "tg_tokens": 64,
   "tps": "9.1"
  },
  {
   "backend": "llama.cpp (OpenCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-26",
   "flags": [],
   "hardware": "Snapdragon 7 Gen 4 (SM7750)",
   "id": "lc-29473-5",
   "model": "Qwen3.5-4B",
   "notes": "Adreno 722 GPU via OpenCL, correct output; -t 4; build 86a24a1; threads=4",
   "pp_tokens": 512,
   "pp_tps": 77.5,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Motorola Edge 70 (roadstr, XT2601-2), Qualcomm Snapdragon 7 Gen 4 (SM7750), Hexagon v73, 12 GB LPDDR5X; Qwen3.5-4B Q4_0, -t 4 pp512 t/s tg64 t/s output; GPU (OpenCL, Adreno 722) 77.5 7.5 correct",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #29473 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29473",
   "tg_tokens": 64,
   "tps": "7.5"
  },
  {
   "backend": "llama.cpp (Hexagon)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-26",
   "flags": [],
   "hardware": "Snapdragon 7 Gen 4 (SM7750)",
   "id": "lc-29473-6",
   "model": "Qwen3.5-4B",
   "notes": "HTP0 default HMX path, garbled output; -t 4; build 86a24a1; config=hexagon-hmx; threads=4",
   "pp_tokens": 512,
   "pp_tps": 437.4,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Motorola Edge 70 (roadstr, XT2601-2), Qualcomm Snapdragon 7 Gen 4 (SM7750), Hexagon v73, 12 GB LPDDR5X; Qwen3.5-4B Q4_0, -t 4 pp512 t/s tg64 t/s output; HTP0, default (HMX) 437.4 8.1 garbled",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #29473 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29473",
   "tg_tokens": 64,
   "tps": "8.1"
  },
  {
   "backend": "llama.cpp (Hexagon)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-26",
   "flags": [],
   "hardware": "Snapdragon 7 Gen 4 (SM7750)",
   "id": "lc-29473-7",
   "model": "Qwen3.5-4B",
   "notes": "HTP0 with HVX-only workaround, correct output; -t 4; build 86a24a1; config=hexagon-workaround; threads=4",
   "pp_tokens": 512,
   "pp_tps": 25.6,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Motorola Edge 70 (roadstr, XT2601-2), Qualcomm Snapdragon 7 Gen 4 (SM7750), Hexagon v73, 12 GB LPDDR5X; Qwen3.5-4B Q4_0, -t 4 pp512 t/s tg64 t/s output; HTP0, workaround above 25.6 6.2 correct",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #29473 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29473",
   "tg_tokens": 64,
   "tps": "6.2"
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 65536,
   "date": "2026-09-25",
   "flags": [],
   "hardware": "Intel Arc B580 Limited Edition",
   "id": "lc-29419-1",
   "model": "gemma-4-31B-it",
   "notes": "2x B580 LE (Gen4x8 + Gen4x4), -sm tensor, MTP draft n-max 4, KV bf16, ctx 65536, fa on; timing sample at n_gen 103 before the SYCL FA abort; run 1 of 2; SYCL build cd74ef6; ngl=52; config=sycl-mtp",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Vulkan1: Intel Arc B580 Limited Edition 12GB Gen4x8; Vulkan2: Intel Arc B580 Limited Edition 12GB Gen4x4; unsloth/gemma-4-31B-it-qat-GGUF:UD-Q4_K_XL with MTP draft (mtp-gemma-4-31B-it.gguf); -ctk bf16 -ctv bf16 -c 65536 -b 256 -ub 256 -ngl 52 --no-mmproj-offload -sm tensor --spec-type draft-mtp --spec-draft-n-max 4; n_gen = 103, tg = 28.44 t/s, tg_3s = 28.72 t/s",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #29419 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29419",
   "tps": "28.44"
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 65536,
   "date": "2026-09-25",
   "flags": [],
   "hardware": "Intel Arc B580 Limited Edition",
   "id": "lc-29419-2",
   "model": "gemma-4-31B-it",
   "notes": "same run, timing sample at n_gen 159, SIGABRT in ggml_sycl_flash_attn_ext right after; run 2 of 2; ngl=52; config=sycl-mtp",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Vulkan1: Intel Arc B580 Limited Edition 12GB Gen4x8; Vulkan2: Intel Arc B580 Limited Edition 12GB Gen4x4; unsloth/gemma-4-31B-it-qat-GGUF:UD-Q4_K_XL with MTP draft (mtp-gemma-4-31B-it.gguf); n_gen = 159, tg = 23.79 t/s, tg_3s = 18.33 t/s",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #29419 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29419",
   "tps": "23.79"
  },
  {
   "backend": "Vulkan",
   "batch": null,
   "ctx": 65536,
   "date": "2026-09-25",
   "flags": [],
   "hardware": "Intel Arc B580 Limited Edition",
   "id": "lc-29418-1",
   "model": "gemma-4-31B-it",
   "notes": "2x B580 LE, -sm tensor -dev Vulkan1,Vulkan2, MTP draft n-max 4, KV bf16, ctx 65536; timing sample at n_gen 103 before the neq0==HSK assert; run 1 of 2; Vulkan build b1175; ngl=60; config=vulkan-mtp",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Vulkan1: Intel Arc B580 Limited Edition 12GB Gen4x8; Vulkan2: Intel Arc B580 Limited Edition 12GB Gen4x4; unsloth/gemma-4-31B-it-qat-GGUF:UD-Q4_K_XL with MTP draft (mtp-gemma-4-31B-it.gguf); -ctk bf16 -ctv bf16 -c 65536 -b 256 -ub 256 -ngl 60 --no-mmproj-offload -sm tensor --spec-type draft-mtp --spec-draft-n-max 4 -dev Vulkan1,Vulkan2; n_gen = 103, tg = 23.11 t/s, tg_3s = 23.33 t/s",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #29418 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29418",
   "tps": "23.11"
  },
  {
   "backend": "Vulkan",
   "batch": null,
   "ctx": 65536,
   "date": "2026-09-25",
   "flags": [],
   "hardware": "Intel Arc B580 Limited Edition",
   "id": "lc-29418-2",
   "model": "gemma-4-31B-it",
   "notes": "same run, timing sample at n_gen 156, assert GGML_ASSERT(neq0 == HSK) right after; run 2 of 2; ngl=60; config=vulkan-mtp",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Vulkan1: Intel Arc B580 Limited Edition 12GB Gen4x8; Vulkan2: Intel Arc B580 Limited Edition 12GB Gen4x4; unsloth/gemma-4-31B-it-qat-GGUF:UD-Q4_K_XL with MTP draft (mtp-gemma-4-31B-it.gguf); n_gen = 156, tg = 20.87 t/s, tg_3s = 17.60 t/s",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #29418 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29418",
   "tps": "20.87"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 8192,
   "date": "2026-09-26",
   "flags": [],
   "hardware": "P100",
   "id": "lc-29466-1",
   "model": "Qwen3.8-27B",
   "notes": "2x P100 (Pascal, sm_60), -sm tensor -ts 1/1, KV q8_0, fa on, ctx 8192; request 1 ran clean on the patched v0.4.0 tree (29 Pascal patches), second request crashes in the meta backend",
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q5_K_XL",
   "quote": "shinbunbun/llama-cpp-p100-patches (29 patches); CUDA_VISIBLE_DEVICES=0,1 /opt/llama.cpp-p100/build/bin/llama-server -m Qwen3.8-27B-UD-Q5_K_XL.gguf; request 1 completes successfully, graphs reused = 291, tg ~18.5 t/s",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #29466 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29466",
   "tps": "18.5"
  },
  {
   "backend": "ExLlamaV2",
   "batch": null,
   "ctx": null,
   "date": "2026-03-16",
   "flags": [],
   "hardware": "RTX PRO 4000 Blackwell",
   "id": "exl-813-1",
   "model": "Qwen2.5-7B-Instruct",
   "notes": "control run for the RTX 5060 Ti Thunderbolt 5 eGPU freeze report: the SFF card runs stable on direct PCIe; TabbyAPI, LoneStriker 6.0bpw h6 EXL2 quant",
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL2 6.0bpw H6",
   "quote": "LoneStriker / Qwen2.5-7B-Instruct-6.0bpw-h6-exl2; RTX PRO 4000 Blackwell SFF (PCIe direct) works perfectly; stable inference at ~47 t/s",
   "retrieved": "2026-09-27",
   "source_name": "ExLlamaV2 issue #813 (community report)",
   "source_url": "https://github.com/turboderp-org/exllamav2/issues/813",
   "tps": "47"
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-25",
   "flags": [],
   "hardware": "B300",
   "id": "vllm-58638-b300-qwen36-g46",
   "model": "Qwen3.6-35B-A3B",
   "notes": "RFC own measurement (section 4.1), DFlash drafter, 46 KV cache groups before fix; c=1; config=kvgroups-46",
   "pp_tps": null,
   "provenance": "community",
   "quant": "as-published",
   "quote": "Qwen3.6-35B-A3B + DFlash on B300 goes from 460 \u2192 733 tok/s at c=1 when the 46 groups drop to 17",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #58638 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58638",
   "tps": "460"
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-25",
   "flags": [],
   "hardware": "B300",
   "id": "vllm-58638-b300-qwen36-g17",
   "model": "Qwen3.6-35B-A3B",
   "notes": "RFC own measurement (section 4.1), DFlash drafter after the group-size fix; c=1; config=kvgroups-17",
   "pp_tps": null,
   "provenance": "community",
   "quant": "as-published",
   "quote": "Qwen3.6-35B-A3B + DFlash on B300 goes from 460 \u2192 733 tok/s at c=1 when the 46 groups drop to 17",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #58638 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58638",
   "tps": "733"
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-24",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "vllm-58578-b70-qwen38-dvfull",
   "model": "Qwen3.8-27B",
   "notes": "reduced draft vocabulary experiment, stock drafter with the full 248,320 id vocabulary; 3 drafts; XPU eager, fp8 KV; config=draftvocab-full",
   "pp_tps": null,
   "provenance": "community",
   "quant": "AutoRound int4",
   "quote": "What we measured on an Intel Arc Pro B70 (XPU, intel/llm-scaler-vllm:0.26.0-b2, vLLM 0.26.1.dev0), eager, fp8 KV, one stream on an idle server; Qwen3.8-27B at 3 drafts: acceptance did not move (2.99-3.04 tokens a step with the full vocabulary, 2.96-3.00 with the list), so one stream went from 59.5 to 74.6 tok/s",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #58578 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58578",
   "tps": "59.5"
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-24",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "vllm-58578-b70-qwen38-dvred",
   "model": "Qwen3.8-27B",
   "notes": "reduced draft vocabulary experiment, 50,521 id list at the same acceptance; 3 drafts; XPU eager, fp8 KV; config=draftvocab-reduced",
   "pp_tps": null,
   "provenance": "community",
   "quant": "AutoRound int4",
   "quote": "What we measured on an Intel Arc Pro B70 (XPU, intel/llm-scaler-vllm:0.26.0-b2, vLLM 0.26.1.dev0), eager, fp8 KV, one stream on an idle server; Qwen3.8-27B at 3 drafts: acceptance did not move (2.99-3.04 tokens a step with the full vocabulary, 2.96-3.00 with the list), so one stream went from 59.5 to 74.6 tok/s",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #58578 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58578",
   "tps": "74.6"
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-24",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "vllm-58578-b70-qwen36-dvfull",
   "model": "Qwen3.6-35B-A3B",
   "notes": "reduced draft vocabulary experiment, stock drafter (1 GB shared lm_head); XPU eager, fp8 KV; config=draftvocab-full",
   "pp_tps": null,
   "provenance": "community",
   "quant": "sym_int4",
   "quote": "What we measured on an Intel Arc Pro B70 (XPU, intel/llm-scaler-vllm:0.26.0-b2, vLLM 0.26.1.dev0), eager, fp8 KV, one stream on an idle server; Qwen3.6-35B-A3B (BF16 quantized to sym_int4 at load; its drafter reads a 1 GB head): a step went from 19.3 to 15.2 ms at the same acceptance, 147.4 to 189.8 tok/s",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #58578 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58578",
   "tps": "147.4"
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-24",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "vllm-58578-b70-qwen36-dvred",
   "model": "Qwen3.6-35B-A3B",
   "notes": "reduced draft vocabulary experiment, token list at the same acceptance, 19.3 to 15.2 ms per step; XPU eager, fp8 KV; config=draftvocab-reduced",
   "pp_tps": null,
   "provenance": "community",
   "quant": "sym_int4",
   "quote": "What we measured on an Intel Arc Pro B70 (XPU, intel/llm-scaler-vllm:0.26.0-b2, vLLM 0.26.1.dev0), eager, fp8 KV, one stream on an idle server; Qwen3.6-35B-A3B (BF16 quantized to sym_int4 at load; its drafter reads a 1 GB head): a step went from 19.3 to 15.2 ms at the same acceptance, 147.4 to 189.8 tok/s",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #58578 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58578",
   "tps": "189.8"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 4,
   "ctx": 400000,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "lc-29510-gemma-master",
   "model": "gemma-4-26B-A4B-it",
   "notes": "llama-batched-bench, unified KV cache (-kvu), -fa on, ctx 400000, -npl 4, -b 2048; master baseline with unified KV penalty; config=unified-kv-master",
   "pp_tokens": 99936,
   "pp_tps": 2982.14,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Tested on Qwen3.8-27B and Gemma-26BA4B. On RTX Pro 6000:; llama-batched-bench -m ~/workspace/gemma-4-26B-A4B-it-Q8_0.gguf -ngl 99 -fa on -c 400000 -b 2048 -ub 2048 -npp 99936 -ntg 64 -npl 4 -kvu; 99936 64 4 400000 134.046 2982.14 0.936 273.42 134.982 2963.35",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp PR #29510 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29510",
   "tg_tokens": 64,
   "tps": "273.42"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 4,
   "ctx": 400000,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "lc-29510-gemma-pr",
   "model": "gemma-4-26B-A4B-it",
   "notes": "llama-batched-bench, unified KV cache (-kvu), -fa on, ctx 400000, -npl 4, -b 2048; PR branch with flash_attn_ext_rows; config=flash-attn-ext-rows",
   "pp_tokens": 99936,
   "pp_tps": 8625.75,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Tested on Qwen3.8-27B and Gemma-26BA4B. On RTX Pro 6000:; llama-batched-bench -m ~/workspace/gemma-4-26B-A4B-it-Q8_0.gguf -ngl 99 -fa on -c 400000 -b 2048 -ub 2048 -npp 99936 -ntg 64 -npl 4 -kvu; 99936 64 4 400000 46.343 8625.75 0.889 288.11 47.232 8468.89",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp PR #29510 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29510",
   "tg_tokens": 64,
   "tps": "288.11"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 4,
   "ctx": 400000,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "lc-29510-qwen-master",
   "model": "Qwen3.8-27B",
   "notes": "llama-batched-bench, unified KV cache (-kvu), -fa on, ctx 400000, -npl 4, -b 2048; master baseline with unified KV penalty; config=unified-kv-master",
   "pp_tokens": 99936,
   "pp_tps": 1106.13,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Tested on Qwen3.8-27B and Gemma-26BA4B. On RTX Pro 6000:; llama-batched-bench -m ~/workspace/Qwen3.8-27B-Q8_0.gguf -ngl 99 -fa on -c 400000 -b 2048 -ub 2048 -npp 99936 -ntg 64 -npl 4 -kvu; 99936 64 4 400000 361.390 1106.13 2.835 90.29 364.226 1098.22",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp PR #29510 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29510",
   "tg_tokens": 64,
   "tps": "90.29"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 4,
   "ctx": 400000,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "lc-29510-qwen-pr",
   "model": "Qwen3.8-27B",
   "notes": "llama-batched-bench, unified KV cache (-kvu), -fa on, ctx 400000, -npl 4, -b 2048; PR branch with flash_attn_ext_rows; config=flash-attn-ext-rows",
   "pp_tokens": 99936,
   "pp_tps": 2778.22,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Tested on Qwen3.8-27B and Gemma-26BA4B. On RTX Pro 6000:; llama-batched-bench -m ~/workspace/Qwen3.8-27B-Q8_0.gguf -ngl 99 -fa on -c 400000 -b 2048 -ub 2048 -npp 99936 -ntg 64 -npl 4 -kvu; 99936 64 4 400000 143.885 2778.22 2.508 102.08 146.393 2732.37",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp PR #29510 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29510",
   "tg_tokens": 64,
   "tps": "102.08"
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": 32768,
   "date": "2026-07-31",
   "flags": [],
   "hardware": "M3 Max",
   "id": "hf1-dsv4flash-m3max",
   "model": "DeepSeek-V4-Flash",
   "notes": "Custom 2+4-bit mixed-precision ds4 GGUF with imatrix (ox-ox re-quant of the 0731 model); single-run Metal CLI, -n 256 short prompt; 128 GB unified memory. Upstream (DeepSeek) reference on the same machine class: 58.52/26.68 t/s.",
   "pp_tps": 45.08,
   "provenance": "community",
   "quant": "DS4 imatrix",
   "quote": "Performance Apple M3 Max, 128 GB. Single-run Metal CLI, --ctx 32768 --nothink --temp 0 -n 256 , short prompt; the same conditions upstream uses for its own table: imatrix build prefill 45.08 t/s generation 27.01 t/s; ox-ox/DeepSeek-V4-Flash-0731-gguf-ds4",
   "retrieved": "2026-09-27",
   "source_name": "HF model card: ox-ox/DeepSeek-V4-Flash-0731-gguf-ds4",
   "source_url": "https://huggingface.co/ox-ox/DeepSeek-V4-Flash-0731-gguf-ds4",
   "tps": "27.01"
  },
  {
   "backend": "LM Studio",
   "batch": null,
   "ctx": null,
   "date": "2026-07-19",
   "flags": [],
   "hardware": "5090",
   "id": "hf1-qwen359b-5090-q4ks",
   "model": "Qwen3.5-9B",
   "notes": "DavidAU 'The Defiant Fable' fine-tune, author's own rough LM Studio test on Windows 11; 'about 130 t/s' (approximate, uncontrolled).",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_S",
   "quote": "SPEED: On Q4_K_S (4bit) quant, regular GGUFs are about 130 t/s, whereas MTP GGUFs (acceptance at 60%, 2 tokens) can exceed 185 T/S. (5090, Windows 11, testing in LMStudio); DavidAU/Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-MTP-GGUF",
   "retrieved": "2026-09-27",
   "source_name": "HF model card: DavidAU Qwen3.5-9B MTP GGUF",
   "source_url": "https://huggingface.co/DavidAU/Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-MTP-GGUF",
   "tps": "130"
  },
  {
   "backend": "LM Studio",
   "batch": null,
   "ctx": null,
   "date": "2026-07-19",
   "flags": [],
   "hardware": "5090",
   "id": "hf1-qwen359b-5090-mtp",
   "model": "Qwen3.5-9B",
   "notes": "MTP (multi-token prediction) quant, acceptance ~60%, 2 tokens; 'can exceed 185 T/S' (upper-bound claim). DavidAU fine-tune, LM Studio, Windows 11.",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_S MTP",
   "quote": "SPEED: On Q4_K_S (4bit) quant, regular GGUFs are about 130 t/s, whereas MTP GGUFs (acceptance at 60%, 2 tokens) can exceed 185 T/S. (5090, Windows 11, testing in LMStudio); DavidAU/Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-MTP-GGUF",
   "retrieved": "2026-09-27",
   "source_name": "HF model card: DavidAU Qwen3.5-9B MTP GGUF",
   "source_url": "https://huggingface.co/DavidAU/Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-MTP-GGUF",
   "tps": "185"
  },
  {
   "backend": "LM Studio",
   "batch": null,
   "ctx": null,
   "date": "2026-07-17",
   "flags": [],
   "hardware": "5090",
   "id": "hf1-qwen3627b-5090-q4ks",
   "model": "Qwen3.6-27B",
   "notes": "DavidAU 'Fable Fusion 711' fine-tune, author's own rough LM Studio test on Windows 11; 'about 75 t/s' (approximate, uncontrolled).",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_S",
   "quote": "SPEED: On Q4_K_S (4bit) quant, regular GGUFs are about 75 t/s, whereas MTP GGUFs (acceptance at 60%, 2 tokens) can exceed 90 T/S. (5090, Windows 11, testing in LMStudio); DavidAU/Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-NEO-MAX-MTP-GGUF",
   "retrieved": "2026-09-27",
   "source_name": "HF model card: DavidAU Qwen3.6-27B MTP GGUF",
   "source_url": "https://huggingface.co/DavidAU/Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-NEO-MAX-MTP-GGUF",
   "tps": "75"
  },
  {
   "backend": "LM Studio",
   "batch": null,
   "ctx": null,
   "date": "2026-07-17",
   "flags": [],
   "hardware": "5090",
   "id": "hf1-qwen3627b-5090-mtp",
   "model": "Qwen3.6-27B",
   "notes": "MTP quant, acceptance ~60%, 2 tokens; 'can exceed 90 T/S' (upper-bound claim). DavidAU fine-tune, LM Studio, Windows 11.",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_S MTP",
   "quote": "SPEED: On Q4_K_S (4bit) quant, regular GGUFs are about 75 t/s, whereas MTP GGUFs (acceptance at 60%, 2 tokens) can exceed 90 T/S. (5090, Windows 11, testing in LMStudio); DavidAU/Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-NEO-MAX-MTP-GGUF",
   "retrieved": "2026-09-27",
   "source_name": "HF model card: DavidAU Qwen3.6-27B MTP GGUF",
   "source_url": "https://huggingface.co/DavidAU/Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-NEO-MAX-MTP-GGUF",
   "tps": "90"
  },
  {
   "backend": "CPU",
   "batch": null,
   "ctx": null,
   "date": "2026-07-28",
   "flags": [],
   "hardware": "M5 Max",
   "id": "hf1-lfsmall-m5max",
   "model": "LFM2.5-2.6B",
   "notes": "Vendor benchmark from the Liquid AI model card, CPU Inference section; ~2.5 GB footprint (Q4-class). Matches the independent llmcheck MLX Q4_K_M figure (220) for M5 Max.",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4 (<2.5 GB)",
   "quote": "Due to its efficient LFM2 architecture, LFM2.5-2.6B is the fastest model we tested, with decode speeds of 220 tokens/s on an M5 Max and 113 tokens/s on a Ryzen AI Max+ 395",
   "retrieved": "2026-09-27",
   "source_name": "HF model card: LiquidAI/LFM2.5-2.6B",
   "source_url": "https://huggingface.co/LiquidAI/LFM2.5-2.6B",
   "tps": "220"
  },
  {
   "backend": "CPU",
   "batch": null,
   "ctx": null,
   "date": "2026-07-28",
   "flags": [],
   "hardware": "Ryzen AI Max+ 395",
   "id": "hf1-lfsmall-ryzen395",
   "model": "LFM2.5-2.6B",
   "notes": "Vendor benchmark from the Liquid AI model card, CPU Inference section; ~2.5 GB footprint (Q4-class).",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4 (<2.5 GB)",
   "quote": "Due to its efficient LFM2 architecture, LFM2.5-2.6B is the fastest model we tested, with decode speeds of 220 tokens/s on an M5 Max and 113 tokens/s on a Ryzen AI Max+ 395",
   "retrieved": "2026-09-27",
   "source_name": "HF model card: LiquidAI/LFM2.5-2.6B",
   "source_url": "https://huggingface.co/LiquidAI/LFM2.5-2.6B",
   "tps": "113"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-16",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "hf1-bonsai-rtx5090-pq2",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "prism-ml llama-bench, custom ternary kernels, batch 1, depth 0, no vision tower; RTX 5090 (32 GB) 129.9 3893 1.95 120.5 1805 2.15 row; PQ2_0 energy 1.95 J/tok (board power incl. HBM/GDDR).",
   "pp_tokens": 512,
   "pp_tps": 3893,
   "provenance": "community",
   "quant": "Ternary PQ2_0",
   "quote": "tg128 is token-generation throughput over 128 generated tokens (the memory-bandwidth-bound, interactive phase); pp512 is prompt-processing throughput over 512 input tokens (the compute-bound phase). Both in tokens/s, measured with llama-bench on these GGUF packs (custom low-bit kernels), at batch size 1 and depth 0 with no vision tower; Platform PQ2_0 TG128 PQ2_0 PP512 PQ2_0 J/tok PTQ1_0 TG128 PTQ1_0 PP512 PTQ1_0 J/tok RTX 5090 (32 GB) 129.9 3893 1.95 120.5 1805 2.15 RTX PRO 6000 Blackwell 124.8 4020 2.49 117.9 1972 2.77 H100 SXM (80 GB) 113.9 2830 2.69 86.9 1237 3.18 RTX 6000 Ada (48 GB) 82.8 2431 2.51 90.4 1657 2.49 RTX 4090 (24 GB) 81.2 3124 2.99 91.1 1645 2.58 L40S (48 GB) 74.4 2868 3.24 81.8 1543 2.82 A100 SXM (80 GB) 73.9 1328 3.43 54.7 706 4.28 L4 (24 GB, 72 W) 29.8 777 2.42 32.1 467 2.25; hf download prism-ml/Ternary-Bonsai-2-27B-gguf Ternary-Bonsai-2-27B-PQ2_0.gguf --local-dir .",
   "retrieved": "2026-09-27",
   "source_name": "HF model card: prism-ml/Ternary-Bonsai-2-27B-gguf",
   "source_url": "https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-gguf",
   "tg_tokens": 128,
   "tps": "129.9"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-16",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "hf1-bonsai-rtx5090-ptq",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "prism-ml llama-bench, custom ternary kernels, batch 1, depth 0, no vision tower; RTX 5090 (32 GB) 129.9 3893 1.95 120.5 1805 2.15 row; PTQ1_0 energy 2.15 J/tok (board power incl. HBM/GDDR).",
   "pp_tokens": 512,
   "pp_tps": 1805,
   "provenance": "community",
   "quant": "Ternary PTQ1_0",
   "quote": "tg128 is token-generation throughput over 128 generated tokens (the memory-bandwidth-bound, interactive phase); pp512 is prompt-processing throughput over 512 input tokens (the compute-bound phase). Both in tokens/s, measured with llama-bench on these GGUF packs (custom low-bit kernels), at batch size 1 and depth 0 with no vision tower; Platform PQ2_0 TG128 PQ2_0 PP512 PQ2_0 J/tok PTQ1_0 TG128 PTQ1_0 PP512 PTQ1_0 J/tok RTX 5090 (32 GB) 129.9 3893 1.95 120.5 1805 2.15 RTX PRO 6000 Blackwell 124.8 4020 2.49 117.9 1972 2.77 H100 SXM (80 GB) 113.9 2830 2.69 86.9 1237 3.18 RTX 6000 Ada (48 GB) 82.8 2431 2.51 90.4 1657 2.49 RTX 4090 (24 GB) 81.2 3124 2.99 91.1 1645 2.58 L40S (48 GB) 74.4 2868 3.24 81.8 1543 2.82 A100 SXM (80 GB) 73.9 1328 3.43 54.7 706 4.28 L4 (24 GB, 72 W) 29.8 777 2.42 32.1 467 2.25; hf download prism-ml/Ternary-Bonsai-2-27B-gguf Ternary-Bonsai-2-27B-PQ2_0.gguf --local-dir .",
   "retrieved": "2026-09-27",
   "source_name": "HF model card: prism-ml/Ternary-Bonsai-2-27B-gguf",
   "source_url": "https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-gguf",
   "tg_tokens": 128,
   "tps": "120.5"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-16",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "hf1-bonsai-rtxpro6000-pq2",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "prism-ml llama-bench, custom ternary kernels, batch 1, depth 0, no vision tower; RTX PRO 6000 Blackwell 124.8 4020 2.49 117.9 1972 2.77 row; PQ2_0 energy 2.49 J/tok (board power incl. HBM/GDDR).",
   "pp_tokens": 512,
   "pp_tps": 4020,
   "provenance": "community",
   "quant": "Ternary PQ2_0",
   "quote": "tg128 is token-generation throughput over 128 generated tokens (the memory-bandwidth-bound, interactive phase); pp512 is prompt-processing throughput over 512 input tokens (the compute-bound phase). Both in tokens/s, measured with llama-bench on these GGUF packs (custom low-bit kernels), at batch size 1 and depth 0 with no vision tower; Platform PQ2_0 TG128 PQ2_0 PP512 PQ2_0 J/tok PTQ1_0 TG128 PTQ1_0 PP512 PTQ1_0 J/tok RTX 5090 (32 GB) 129.9 3893 1.95 120.5 1805 2.15 RTX PRO 6000 Blackwell 124.8 4020 2.49 117.9 1972 2.77 H100 SXM (80 GB) 113.9 2830 2.69 86.9 1237 3.18 RTX 6000 Ada (48 GB) 82.8 2431 2.51 90.4 1657 2.49 RTX 4090 (24 GB) 81.2 3124 2.99 91.1 1645 2.58 L40S (48 GB) 74.4 2868 3.24 81.8 1543 2.82 A100 SXM (80 GB) 73.9 1328 3.43 54.7 706 4.28 L4 (24 GB, 72 W) 29.8 777 2.42 32.1 467 2.25; hf download prism-ml/Ternary-Bonsai-2-27B-gguf Ternary-Bonsai-2-27B-PQ2_0.gguf --local-dir .",
   "retrieved": "2026-09-27",
   "source_name": "HF model card: prism-ml/Ternary-Bonsai-2-27B-gguf",
   "source_url": "https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-gguf",
   "tg_tokens": 128,
   "tps": "124.8"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-16",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "hf1-bonsai-rtxpro6000-ptq",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "prism-ml llama-bench, custom ternary kernels, batch 1, depth 0, no vision tower; RTX PRO 6000 Blackwell 124.8 4020 2.49 117.9 1972 2.77 row; PTQ1_0 energy 2.77 J/tok (board power incl. HBM/GDDR).",
   "pp_tokens": 512,
   "pp_tps": 1972,
   "provenance": "community",
   "quant": "Ternary PTQ1_0",
   "quote": "tg128 is token-generation throughput over 128 generated tokens (the memory-bandwidth-bound, interactive phase); pp512 is prompt-processing throughput over 512 input tokens (the compute-bound phase). Both in tokens/s, measured with llama-bench on these GGUF packs (custom low-bit kernels), at batch size 1 and depth 0 with no vision tower; Platform PQ2_0 TG128 PQ2_0 PP512 PQ2_0 J/tok PTQ1_0 TG128 PTQ1_0 PP512 PTQ1_0 J/tok RTX 5090 (32 GB) 129.9 3893 1.95 120.5 1805 2.15 RTX PRO 6000 Blackwell 124.8 4020 2.49 117.9 1972 2.77 H100 SXM (80 GB) 113.9 2830 2.69 86.9 1237 3.18 RTX 6000 Ada (48 GB) 82.8 2431 2.51 90.4 1657 2.49 RTX 4090 (24 GB) 81.2 3124 2.99 91.1 1645 2.58 L40S (48 GB) 74.4 2868 3.24 81.8 1543 2.82 A100 SXM (80 GB) 73.9 1328 3.43 54.7 706 4.28 L4 (24 GB, 72 W) 29.8 777 2.42 32.1 467 2.25; hf download prism-ml/Ternary-Bonsai-2-27B-gguf Ternary-Bonsai-2-27B-PQ2_0.gguf --local-dir .",
   "retrieved": "2026-09-27",
   "source_name": "HF model card: prism-ml/Ternary-Bonsai-2-27B-gguf",
   "source_url": "https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-gguf",
   "tg_tokens": 128,
   "tps": "117.9"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-16",
   "flags": [],
   "hardware": "H100",
   "id": "hf1-bonsai-h100-pq2",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "prism-ml llama-bench, custom ternary kernels, batch 1, depth 0, no vision tower; H100 SXM (80 GB) 113.9 2830 2.69 86.9 1237 3.18 row; PQ2_0 energy 2.69 J/tok (board power incl. HBM/GDDR).",
   "pp_tokens": 512,
   "pp_tps": 2830,
   "provenance": "community",
   "quant": "Ternary PQ2_0",
   "quote": "tg128 is token-generation throughput over 128 generated tokens (the memory-bandwidth-bound, interactive phase); pp512 is prompt-processing throughput over 512 input tokens (the compute-bound phase). Both in tokens/s, measured with llama-bench on these GGUF packs (custom low-bit kernels), at batch size 1 and depth 0 with no vision tower; Platform PQ2_0 TG128 PQ2_0 PP512 PQ2_0 J/tok PTQ1_0 TG128 PTQ1_0 PP512 PTQ1_0 J/tok RTX 5090 (32 GB) 129.9 3893 1.95 120.5 1805 2.15 RTX PRO 6000 Blackwell 124.8 4020 2.49 117.9 1972 2.77 H100 SXM (80 GB) 113.9 2830 2.69 86.9 1237 3.18 RTX 6000 Ada (48 GB) 82.8 2431 2.51 90.4 1657 2.49 RTX 4090 (24 GB) 81.2 3124 2.99 91.1 1645 2.58 L40S (48 GB) 74.4 2868 3.24 81.8 1543 2.82 A100 SXM (80 GB) 73.9 1328 3.43 54.7 706 4.28 L4 (24 GB, 72 W) 29.8 777 2.42 32.1 467 2.25; hf download prism-ml/Ternary-Bonsai-2-27B-gguf Ternary-Bonsai-2-27B-PQ2_0.gguf --local-dir .",
   "retrieved": "2026-09-27",
   "source_name": "HF model card: prism-ml/Ternary-Bonsai-2-27B-gguf",
   "source_url": "https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-gguf",
   "tg_tokens": 128,
   "tps": "113.9"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-16",
   "flags": [],
   "hardware": "H100",
   "id": "hf1-bonsai-h100-ptq",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "prism-ml llama-bench, custom ternary kernels, batch 1, depth 0, no vision tower; H100 SXM (80 GB) 113.9 2830 2.69 86.9 1237 3.18 row; PTQ1_0 energy 3.18 J/tok (board power incl. HBM/GDDR).",
   "pp_tokens": 512,
   "pp_tps": 1237,
   "provenance": "community",
   "quant": "Ternary PTQ1_0",
   "quote": "tg128 is token-generation throughput over 128 generated tokens (the memory-bandwidth-bound, interactive phase); pp512 is prompt-processing throughput over 512 input tokens (the compute-bound phase). Both in tokens/s, measured with llama-bench on these GGUF packs (custom low-bit kernels), at batch size 1 and depth 0 with no vision tower; Platform PQ2_0 TG128 PQ2_0 PP512 PQ2_0 J/tok PTQ1_0 TG128 PTQ1_0 PP512 PTQ1_0 J/tok RTX 5090 (32 GB) 129.9 3893 1.95 120.5 1805 2.15 RTX PRO 6000 Blackwell 124.8 4020 2.49 117.9 1972 2.77 H100 SXM (80 GB) 113.9 2830 2.69 86.9 1237 3.18 RTX 6000 Ada (48 GB) 82.8 2431 2.51 90.4 1657 2.49 RTX 4090 (24 GB) 81.2 3124 2.99 91.1 1645 2.58 L40S (48 GB) 74.4 2868 3.24 81.8 1543 2.82 A100 SXM (80 GB) 73.9 1328 3.43 54.7 706 4.28 L4 (24 GB, 72 W) 29.8 777 2.42 32.1 467 2.25; hf download prism-ml/Ternary-Bonsai-2-27B-gguf Ternary-Bonsai-2-27B-PQ2_0.gguf --local-dir .",
   "retrieved": "2026-09-27",
   "source_name": "HF model card: prism-ml/Ternary-Bonsai-2-27B-gguf",
   "source_url": "https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-gguf",
   "tg_tokens": 128,
   "tps": "86.9"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-16",
   "flags": [],
   "hardware": "RTX 6000 Ada",
   "id": "hf1-bonsai-rtx6000ada-pq2",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "prism-ml llama-bench, custom ternary kernels, batch 1, depth 0, no vision tower; RTX 6000 Ada (48 GB) 82.8 2431 2.51 90.4 1657 2.49 row; PQ2_0 energy 2.51 J/tok (board power incl. HBM/GDDR).",
   "pp_tokens": 512,
   "pp_tps": 2431,
   "provenance": "community",
   "quant": "Ternary PQ2_0",
   "quote": "tg128 is token-generation throughput over 128 generated tokens (the memory-bandwidth-bound, interactive phase); pp512 is prompt-processing throughput over 512 input tokens (the compute-bound phase). Both in tokens/s, measured with llama-bench on these GGUF packs (custom low-bit kernels), at batch size 1 and depth 0 with no vision tower; Platform PQ2_0 TG128 PQ2_0 PP512 PQ2_0 J/tok PTQ1_0 TG128 PTQ1_0 PP512 PTQ1_0 J/tok RTX 5090 (32 GB) 129.9 3893 1.95 120.5 1805 2.15 RTX PRO 6000 Blackwell 124.8 4020 2.49 117.9 1972 2.77 H100 SXM (80 GB) 113.9 2830 2.69 86.9 1237 3.18 RTX 6000 Ada (48 GB) 82.8 2431 2.51 90.4 1657 2.49 RTX 4090 (24 GB) 81.2 3124 2.99 91.1 1645 2.58 L40S (48 GB) 74.4 2868 3.24 81.8 1543 2.82 A100 SXM (80 GB) 73.9 1328 3.43 54.7 706 4.28 L4 (24 GB, 72 W) 29.8 777 2.42 32.1 467 2.25; hf download prism-ml/Ternary-Bonsai-2-27B-gguf Ternary-Bonsai-2-27B-PQ2_0.gguf --local-dir .",
   "retrieved": "2026-09-27",
   "source_name": "HF model card: prism-ml/Ternary-Bonsai-2-27B-gguf",
   "source_url": "https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-gguf",
   "tg_tokens": 128,
   "tps": "82.8"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-16",
   "flags": [],
   "hardware": "RTX 6000 Ada",
   "id": "hf1-bonsai-rtx6000ada-ptq",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "prism-ml llama-bench, custom ternary kernels, batch 1, depth 0, no vision tower; RTX 6000 Ada (48 GB) 82.8 2431 2.51 90.4 1657 2.49 row; PTQ1_0 energy 2.49 J/tok (board power incl. HBM/GDDR).",
   "pp_tokens": 512,
   "pp_tps": 1657,
   "provenance": "community",
   "quant": "Ternary PTQ1_0",
   "quote": "tg128 is token-generation throughput over 128 generated tokens (the memory-bandwidth-bound, interactive phase); pp512 is prompt-processing throughput over 512 input tokens (the compute-bound phase). Both in tokens/s, measured with llama-bench on these GGUF packs (custom low-bit kernels), at batch size 1 and depth 0 with no vision tower; Platform PQ2_0 TG128 PQ2_0 PP512 PQ2_0 J/tok PTQ1_0 TG128 PTQ1_0 PP512 PTQ1_0 J/tok RTX 5090 (32 GB) 129.9 3893 1.95 120.5 1805 2.15 RTX PRO 6000 Blackwell 124.8 4020 2.49 117.9 1972 2.77 H100 SXM (80 GB) 113.9 2830 2.69 86.9 1237 3.18 RTX 6000 Ada (48 GB) 82.8 2431 2.51 90.4 1657 2.49 RTX 4090 (24 GB) 81.2 3124 2.99 91.1 1645 2.58 L40S (48 GB) 74.4 2868 3.24 81.8 1543 2.82 A100 SXM (80 GB) 73.9 1328 3.43 54.7 706 4.28 L4 (24 GB, 72 W) 29.8 777 2.42 32.1 467 2.25; hf download prism-ml/Ternary-Bonsai-2-27B-gguf Ternary-Bonsai-2-27B-PQ2_0.gguf --local-dir .",
   "retrieved": "2026-09-27",
   "source_name": "HF model card: prism-ml/Ternary-Bonsai-2-27B-gguf",
   "source_url": "https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-gguf",
   "tg_tokens": 128,
   "tps": "90.4"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-16",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "hf1-bonsai-rtx4090-pq2",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "prism-ml llama-bench, custom ternary kernels, batch 1, depth 0, no vision tower; RTX 4090 (24 GB) 81.2 3124 2.99 91.1 1645 2.58 row; PQ2_0 energy 2.99 J/tok (board power incl. HBM/GDDR).",
   "pp_tokens": 512,
   "pp_tps": 3124,
   "provenance": "community",
   "quant": "Ternary PQ2_0",
   "quote": "tg128 is token-generation throughput over 128 generated tokens (the memory-bandwidth-bound, interactive phase); pp512 is prompt-processing throughput over 512 input tokens (the compute-bound phase). Both in tokens/s, measured with llama-bench on these GGUF packs (custom low-bit kernels), at batch size 1 and depth 0 with no vision tower; Platform PQ2_0 TG128 PQ2_0 PP512 PQ2_0 J/tok PTQ1_0 TG128 PTQ1_0 PP512 PTQ1_0 J/tok RTX 5090 (32 GB) 129.9 3893 1.95 120.5 1805 2.15 RTX PRO 6000 Blackwell 124.8 4020 2.49 117.9 1972 2.77 H100 SXM (80 GB) 113.9 2830 2.69 86.9 1237 3.18 RTX 6000 Ada (48 GB) 82.8 2431 2.51 90.4 1657 2.49 RTX 4090 (24 GB) 81.2 3124 2.99 91.1 1645 2.58 L40S (48 GB) 74.4 2868 3.24 81.8 1543 2.82 A100 SXM (80 GB) 73.9 1328 3.43 54.7 706 4.28 L4 (24 GB, 72 W) 29.8 777 2.42 32.1 467 2.25; hf download prism-ml/Ternary-Bonsai-2-27B-gguf Ternary-Bonsai-2-27B-PQ2_0.gguf --local-dir .",
   "retrieved": "2026-09-27",
   "source_name": "HF model card: prism-ml/Ternary-Bonsai-2-27B-gguf",
   "source_url": "https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-gguf",
   "tg_tokens": 128,
   "tps": "81.2"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-16",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "hf1-bonsai-rtx4090-ptq",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "prism-ml llama-bench, custom ternary kernels, batch 1, depth 0, no vision tower; RTX 4090 (24 GB) 81.2 3124 2.99 91.1 1645 2.58 row; PTQ1_0 energy 2.58 J/tok (board power incl. HBM/GDDR).",
   "pp_tokens": 512,
   "pp_tps": 1645,
   "provenance": "community",
   "quant": "Ternary PTQ1_0",
   "quote": "tg128 is token-generation throughput over 128 generated tokens (the memory-bandwidth-bound, interactive phase); pp512 is prompt-processing throughput over 512 input tokens (the compute-bound phase). Both in tokens/s, measured with llama-bench on these GGUF packs (custom low-bit kernels), at batch size 1 and depth 0 with no vision tower; Platform PQ2_0 TG128 PQ2_0 PP512 PQ2_0 J/tok PTQ1_0 TG128 PTQ1_0 PP512 PTQ1_0 J/tok RTX 5090 (32 GB) 129.9 3893 1.95 120.5 1805 2.15 RTX PRO 6000 Blackwell 124.8 4020 2.49 117.9 1972 2.77 H100 SXM (80 GB) 113.9 2830 2.69 86.9 1237 3.18 RTX 6000 Ada (48 GB) 82.8 2431 2.51 90.4 1657 2.49 RTX 4090 (24 GB) 81.2 3124 2.99 91.1 1645 2.58 L40S (48 GB) 74.4 2868 3.24 81.8 1543 2.82 A100 SXM (80 GB) 73.9 1328 3.43 54.7 706 4.28 L4 (24 GB, 72 W) 29.8 777 2.42 32.1 467 2.25; hf download prism-ml/Ternary-Bonsai-2-27B-gguf Ternary-Bonsai-2-27B-PQ2_0.gguf --local-dir .",
   "retrieved": "2026-09-27",
   "source_name": "HF model card: prism-ml/Ternary-Bonsai-2-27B-gguf",
   "source_url": "https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-gguf",
   "tg_tokens": 128,
   "tps": "91.1"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-16",
   "flags": [],
   "hardware": "L40S",
   "id": "hf1-bonsai-l40s-pq2",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "prism-ml llama-bench, custom ternary kernels, batch 1, depth 0, no vision tower; L40S (48 GB) 74.4 2868 3.24 81.8 1543 2.82 row; PQ2_0 energy 3.24 J/tok (board power incl. HBM/GDDR).",
   "pp_tokens": 512,
   "pp_tps": 2868,
   "provenance": "community",
   "quant": "Ternary PQ2_0",
   "quote": "tg128 is token-generation throughput over 128 generated tokens (the memory-bandwidth-bound, interactive phase); pp512 is prompt-processing throughput over 512 input tokens (the compute-bound phase). Both in tokens/s, measured with llama-bench on these GGUF packs (custom low-bit kernels), at batch size 1 and depth 0 with no vision tower; Platform PQ2_0 TG128 PQ2_0 PP512 PQ2_0 J/tok PTQ1_0 TG128 PTQ1_0 PP512 PTQ1_0 J/tok RTX 5090 (32 GB) 129.9 3893 1.95 120.5 1805 2.15 RTX PRO 6000 Blackwell 124.8 4020 2.49 117.9 1972 2.77 H100 SXM (80 GB) 113.9 2830 2.69 86.9 1237 3.18 RTX 6000 Ada (48 GB) 82.8 2431 2.51 90.4 1657 2.49 RTX 4090 (24 GB) 81.2 3124 2.99 91.1 1645 2.58 L40S (48 GB) 74.4 2868 3.24 81.8 1543 2.82 A100 SXM (80 GB) 73.9 1328 3.43 54.7 706 4.28 L4 (24 GB, 72 W) 29.8 777 2.42 32.1 467 2.25; hf download prism-ml/Ternary-Bonsai-2-27B-gguf Ternary-Bonsai-2-27B-PQ2_0.gguf --local-dir .",
   "retrieved": "2026-09-27",
   "source_name": "HF model card: prism-ml/Ternary-Bonsai-2-27B-gguf",
   "source_url": "https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-gguf",
   "tg_tokens": 128,
   "tps": "74.4"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-16",
   "flags": [],
   "hardware": "L40S",
   "id": "hf1-bonsai-l40s-ptq",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "prism-ml llama-bench, custom ternary kernels, batch 1, depth 0, no vision tower; L40S (48 GB) 74.4 2868 3.24 81.8 1543 2.82 row; PTQ1_0 energy 2.82 J/tok (board power incl. HBM/GDDR).",
   "pp_tokens": 512,
   "pp_tps": 1543,
   "provenance": "community",
   "quant": "Ternary PTQ1_0",
   "quote": "tg128 is token-generation throughput over 128 generated tokens (the memory-bandwidth-bound, interactive phase); pp512 is prompt-processing throughput over 512 input tokens (the compute-bound phase). Both in tokens/s, measured with llama-bench on these GGUF packs (custom low-bit kernels), at batch size 1 and depth 0 with no vision tower; Platform PQ2_0 TG128 PQ2_0 PP512 PQ2_0 J/tok PTQ1_0 TG128 PTQ1_0 PP512 PTQ1_0 J/tok RTX 5090 (32 GB) 129.9 3893 1.95 120.5 1805 2.15 RTX PRO 6000 Blackwell 124.8 4020 2.49 117.9 1972 2.77 H100 SXM (80 GB) 113.9 2830 2.69 86.9 1237 3.18 RTX 6000 Ada (48 GB) 82.8 2431 2.51 90.4 1657 2.49 RTX 4090 (24 GB) 81.2 3124 2.99 91.1 1645 2.58 L40S (48 GB) 74.4 2868 3.24 81.8 1543 2.82 A100 SXM (80 GB) 73.9 1328 3.43 54.7 706 4.28 L4 (24 GB, 72 W) 29.8 777 2.42 32.1 467 2.25; hf download prism-ml/Ternary-Bonsai-2-27B-gguf Ternary-Bonsai-2-27B-PQ2_0.gguf --local-dir .",
   "retrieved": "2026-09-27",
   "source_name": "HF model card: prism-ml/Ternary-Bonsai-2-27B-gguf",
   "source_url": "https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-gguf",
   "tg_tokens": 128,
   "tps": "81.8"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-16",
   "flags": [],
   "hardware": "A100",
   "id": "hf1-bonsai-a100-pq2",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "prism-ml llama-bench, custom ternary kernels, batch 1, depth 0, no vision tower; A100 SXM (80 GB) 73.9 1328 3.43 54.7 706 4.28 row; PQ2_0 energy 3.43 J/tok (board power incl. HBM/GDDR).",
   "pp_tokens": 512,
   "pp_tps": 1328,
   "provenance": "community",
   "quant": "Ternary PQ2_0",
   "quote": "tg128 is token-generation throughput over 128 generated tokens (the memory-bandwidth-bound, interactive phase); pp512 is prompt-processing throughput over 512 input tokens (the compute-bound phase). Both in tokens/s, measured with llama-bench on these GGUF packs (custom low-bit kernels), at batch size 1 and depth 0 with no vision tower; Platform PQ2_0 TG128 PQ2_0 PP512 PQ2_0 J/tok PTQ1_0 TG128 PTQ1_0 PP512 PTQ1_0 J/tok RTX 5090 (32 GB) 129.9 3893 1.95 120.5 1805 2.15 RTX PRO 6000 Blackwell 124.8 4020 2.49 117.9 1972 2.77 H100 SXM (80 GB) 113.9 2830 2.69 86.9 1237 3.18 RTX 6000 Ada (48 GB) 82.8 2431 2.51 90.4 1657 2.49 RTX 4090 (24 GB) 81.2 3124 2.99 91.1 1645 2.58 L40S (48 GB) 74.4 2868 3.24 81.8 1543 2.82 A100 SXM (80 GB) 73.9 1328 3.43 54.7 706 4.28 L4 (24 GB, 72 W) 29.8 777 2.42 32.1 467 2.25; hf download prism-ml/Ternary-Bonsai-2-27B-gguf Ternary-Bonsai-2-27B-PQ2_0.gguf --local-dir .",
   "retrieved": "2026-09-27",
   "source_name": "HF model card: prism-ml/Ternary-Bonsai-2-27B-gguf",
   "source_url": "https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-gguf",
   "tg_tokens": 128,
   "tps": "73.9"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-16",
   "flags": [],
   "hardware": "A100",
   "id": "hf1-bonsai-a100-ptq",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "prism-ml llama-bench, custom ternary kernels, batch 1, depth 0, no vision tower; A100 SXM (80 GB) 73.9 1328 3.43 54.7 706 4.28 row; PTQ1_0 energy 4.28 J/tok (board power incl. HBM/GDDR).",
   "pp_tokens": 512,
   "pp_tps": 706,
   "provenance": "community",
   "quant": "Ternary PTQ1_0",
   "quote": "tg128 is token-generation throughput over 128 generated tokens (the memory-bandwidth-bound, interactive phase); pp512 is prompt-processing throughput over 512 input tokens (the compute-bound phase). Both in tokens/s, measured with llama-bench on these GGUF packs (custom low-bit kernels), at batch size 1 and depth 0 with no vision tower; Platform PQ2_0 TG128 PQ2_0 PP512 PQ2_0 J/tok PTQ1_0 TG128 PTQ1_0 PP512 PTQ1_0 J/tok RTX 5090 (32 GB) 129.9 3893 1.95 120.5 1805 2.15 RTX PRO 6000 Blackwell 124.8 4020 2.49 117.9 1972 2.77 H100 SXM (80 GB) 113.9 2830 2.69 86.9 1237 3.18 RTX 6000 Ada (48 GB) 82.8 2431 2.51 90.4 1657 2.49 RTX 4090 (24 GB) 81.2 3124 2.99 91.1 1645 2.58 L40S (48 GB) 74.4 2868 3.24 81.8 1543 2.82 A100 SXM (80 GB) 73.9 1328 3.43 54.7 706 4.28 L4 (24 GB, 72 W) 29.8 777 2.42 32.1 467 2.25; hf download prism-ml/Ternary-Bonsai-2-27B-gguf Ternary-Bonsai-2-27B-PQ2_0.gguf --local-dir .",
   "retrieved": "2026-09-27",
   "source_name": "HF model card: prism-ml/Ternary-Bonsai-2-27B-gguf",
   "source_url": "https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-gguf",
   "tg_tokens": 128,
   "tps": "54.7"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-16",
   "flags": [],
   "hardware": "L4",
   "id": "hf1-bonsai-l4-pq2",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "prism-ml llama-bench, custom ternary kernels, batch 1, depth 0, no vision tower; L4 (24 GB, 72 W) 29.8 777 2.42 32.1 467 2.25 row; PQ2_0 energy 2.42 J/tok (board power incl. HBM/GDDR).",
   "pp_tokens": 512,
   "pp_tps": 777,
   "provenance": "community",
   "quant": "Ternary PQ2_0",
   "quote": "tg128 is token-generation throughput over 128 generated tokens (the memory-bandwidth-bound, interactive phase); pp512 is prompt-processing throughput over 512 input tokens (the compute-bound phase). Both in tokens/s, measured with llama-bench on these GGUF packs (custom low-bit kernels), at batch size 1 and depth 0 with no vision tower; Platform PQ2_0 TG128 PQ2_0 PP512 PQ2_0 J/tok PTQ1_0 TG128 PTQ1_0 PP512 PTQ1_0 J/tok RTX 5090 (32 GB) 129.9 3893 1.95 120.5 1805 2.15 RTX PRO 6000 Blackwell 124.8 4020 2.49 117.9 1972 2.77 H100 SXM (80 GB) 113.9 2830 2.69 86.9 1237 3.18 RTX 6000 Ada (48 GB) 82.8 2431 2.51 90.4 1657 2.49 RTX 4090 (24 GB) 81.2 3124 2.99 91.1 1645 2.58 L40S (48 GB) 74.4 2868 3.24 81.8 1543 2.82 A100 SXM (80 GB) 73.9 1328 3.43 54.7 706 4.28 L4 (24 GB, 72 W) 29.8 777 2.42 32.1 467 2.25; hf download prism-ml/Ternary-Bonsai-2-27B-gguf Ternary-Bonsai-2-27B-PQ2_0.gguf --local-dir .",
   "retrieved": "2026-09-27",
   "source_name": "HF model card: prism-ml/Ternary-Bonsai-2-27B-gguf",
   "source_url": "https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-gguf",
   "tg_tokens": 128,
   "tps": "29.8"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-16",
   "flags": [],
   "hardware": "L4",
   "id": "hf1-bonsai-l4-ptq",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "prism-ml llama-bench, custom ternary kernels, batch 1, depth 0, no vision tower; L4 (24 GB, 72 W) 29.8 777 2.42 32.1 467 2.25 row; PTQ1_0 energy 2.25 J/tok (board power incl. HBM/GDDR).",
   "pp_tokens": 512,
   "pp_tps": 467,
   "provenance": "community",
   "quant": "Ternary PTQ1_0",
   "quote": "tg128 is token-generation throughput over 128 generated tokens (the memory-bandwidth-bound, interactive phase); pp512 is prompt-processing throughput over 512 input tokens (the compute-bound phase). Both in tokens/s, measured with llama-bench on these GGUF packs (custom low-bit kernels), at batch size 1 and depth 0 with no vision tower; Platform PQ2_0 TG128 PQ2_0 PP512 PQ2_0 J/tok PTQ1_0 TG128 PTQ1_0 PP512 PTQ1_0 J/tok RTX 5090 (32 GB) 129.9 3893 1.95 120.5 1805 2.15 RTX PRO 6000 Blackwell 124.8 4020 2.49 117.9 1972 2.77 H100 SXM (80 GB) 113.9 2830 2.69 86.9 1237 3.18 RTX 6000 Ada (48 GB) 82.8 2431 2.51 90.4 1657 2.49 RTX 4090 (24 GB) 81.2 3124 2.99 91.1 1645 2.58 L40S (48 GB) 74.4 2868 3.24 81.8 1543 2.82 A100 SXM (80 GB) 73.9 1328 3.43 54.7 706 4.28 L4 (24 GB, 72 W) 29.8 777 2.42 32.1 467 2.25; hf download prism-ml/Ternary-Bonsai-2-27B-gguf Ternary-Bonsai-2-27B-PQ2_0.gguf --local-dir .",
   "retrieved": "2026-09-27",
   "source_name": "HF model card: prism-ml/Ternary-Bonsai-2-27B-gguf",
   "source_url": "https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-gguf",
   "tg_tokens": 128,
   "tps": "32.1"
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "hf1-bonsai-m5pro-cur",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "prism-ml llama-bench, custom ternary kernels, batch 1; PQ2_0 row; measured on a quiet machine (~4% swing with background load); 34.1 W across CPU+GPU via powermetrics (GPU rail 27.5 W); no J/tok published for Apple (different instrumentation).",
   "power_w": 34.1,
   "pp_tokens": 512,
   "pp_tps": 387,
   "provenance": "community",
   "quant": "Ternary PQ2_0",
   "quote": "tg128 is token-generation throughput over 128 generated tokens (the memory-bandwidth-bound, interactive phase); pp512 is prompt-processing throughput over 512 input tokens (the compute-bound phase). Both in tokens/s, measured with llama-bench on these GGUF packs (custom low-bit kernels), at batch size 1 and depth 0 with no vision tower; Laptop (Apple M5 Pro, Metal) 28.1 387; hf download prism-ml/Ternary-Bonsai-2-27B-gguf Ternary-Bonsai-2-27B-PQ2_0.gguf --local-dir .; the M5 Pro decodes at 27.5 W on the GPU rail and 34.1 W across CPU and GPU",
   "retrieved": "2026-09-27",
   "source_name": "HF model card: prism-ml/Ternary-Bonsai-2-27B-gguf",
   "source_url": "https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-gguf",
   "tg_tokens": 128,
   "tps": "28.1"
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "hf1-bonsai-m5max",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "prism-ml card, re-measurement table (earlier pre-rotation build, pending re-measurement on the current stack); 7.2 GB footprint; llama.cpp Metal, TG128/PP512 per llama-bench convention.",
   "pp_tokens": 512,
   "pp_tps": 765,
   "provenance": "community",
   "quant": "Ternary (pre-rotation build)",
   "quote": "Platform Footprint TG128 (tok/s) PP512 (tok/s) Laptop (Apple M5 Max, Metal) 7.2 GB 47.0 765 Laptop (Apple M5 Pro, Metal) 7.2 GB 28.7 393 Laptop (Apple M4 Pro, Metal) 7.2 GB 18.0 125; hf download prism-ml/Ternary-Bonsai-2-27B-gguf Ternary-Bonsai-2-27B-PQ2_0.gguf --local-dir .",
   "retrieved": "2026-09-27",
   "source_name": "HF model card: prism-ml/Ternary-Bonsai-2-27B-gguf",
   "source_url": "https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-gguf",
   "tg_tokens": 128,
   "tps": "47.0"
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "hf1-bonsai-m5pro-rep",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "prism-ml card, re-measurement table (earlier pre-rotation build, pending re-measurement on the current stack); 7.2 GB footprint; llama.cpp Metal.",
   "pp_tokens": 512,
   "pp_tps": 393,
   "provenance": "community",
   "quant": "Ternary (pre-rotation build)",
   "quote": "Platform Footprint TG128 (tok/s) PP512 (tok/s) Laptop (Apple M5 Max, Metal) 7.2 GB 47.0 765 Laptop (Apple M5 Pro, Metal) 7.2 GB 28.7 393 Laptop (Apple M4 Pro, Metal) 7.2 GB 18.0 125; hf download prism-ml/Ternary-Bonsai-2-27B-gguf Ternary-Bonsai-2-27B-PQ2_0.gguf --local-dir .",
   "retrieved": "2026-09-27",
   "source_name": "HF model card: prism-ml/Ternary-Bonsai-2-27B-gguf",
   "source_url": "https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-gguf",
   "tg_tokens": 128,
   "tps": "28.7"
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "hf1-bonsai-m4pro",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "prism-ml card, re-measurement table (earlier pre-rotation build, pending re-measurement on the current stack); 7.2 GB footprint; llama.cpp Metal; prefill is the practical limit on this machine per the card.",
   "pp_tokens": 512,
   "pp_tps": 125,
   "provenance": "community",
   "quant": "Ternary (pre-rotation build)",
   "quote": "Platform Footprint TG128 (tok/s) PP512 (tok/s) Laptop (Apple M5 Max, Metal) 7.2 GB 47.0 765 Laptop (Apple M5 Pro, Metal) 7.2 GB 28.7 393 Laptop (Apple M4 Pro, Metal) 7.2 GB 18.0 125; hf download prism-ml/Ternary-Bonsai-2-27B-gguf Ternary-Bonsai-2-27B-PQ2_0.gguf --local-dir .",
   "retrieved": "2026-09-27",
   "source_name": "HF model card: prism-ml/Ternary-Bonsai-2-27B-gguf",
   "source_url": "https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-gguf",
   "tg_tokens": 128,
   "tps": "18.0"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-09",
   "flags": [],
   "hardware": "RTX 3060 12GB",
   "id": "hf1-gemma12b-3060",
   "model": "gemma4 12B Q4_KM",
   "notes": "Community re-hosted card with a speed-benchmark table (Ryzen 5 5600x, 64 GB RAM, RTX 3060 12GB); runtime not stated on the card (GGUF benchmark, assumed llama.cpp CUDA path). Generation 33.34 / prompt processing 609.26 tok/s.",
   "pp_tps": 609.26,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Speed Benchmarks Tested on: Ryzen 5 5600x | 64GB RAM | Nvidia RTX 3060 12GB Model Size Generation Prompt Processing gemma-4-12B-Q4_K_M.gguf 7.4 GB 33.34 tok/s 609.26 tok/s",
   "retrieved": "2026-09-27",
   "source_name": "HF model card: superkuh/gemma-4-12b-Q4_K_M-GGUF",
   "source_url": "https://huggingface.co/superkuh/gemma-4-12b-Q4_K_M-GGUF",
   "tps": "33.34"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 16384,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "RTX 2060 6 GB",
   "id": "lc-29513-cuda-first",
   "model": "Qwen3.5-4B",
   "notes": "unsloth Qwen3.5-4B Q4_K_M GGUF; ngl=99, ctx 16384, ctk q8_0, ctv q8_0, reasoning-budget 0; Windows x64 release build b11206 (2b129ccfa), driver 610.88; first chat completion right after model load (Gated DeltaNet warm-up at startup defaults; /health was ok before it); requests from the third on run at normal speed; config=first-request",
   "pp_tps": 0.44,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "- GPU: RTX 2060 6 GB (sm_75), driver 610.88; - CPU i7-9750H; Windows 11 Pro; - Model: unsloth Qwen3.5-4B, Q4_K_M GGUF; llama-server.exe -m Qwen3.5-4B-Q4_K_M.gguf -ngl 99 -c 16384 -ctk q8_0 -ctv q8_0 --reasoning-budget 0 --no-webui; | CUDA | 1st | 0.44 tok/s | 3.09 tok/s |",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #29513",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29513",
   "tps": "3.09"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 16384,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "RTX 2060 6 GB",
   "id": "lc-29513-cuda-warm",
   "model": "Qwen3.5-4B",
   "notes": "unsloth Qwen3.5-4B Q4_K_M GGUF; ngl=99, ctx 16384, ctk q8_0, ctv q8_0, reasoning-budget 0; Windows x64 release build b11206 (2b129ccfa), driver 610.88; requests from the third on (warm; the first request after load is extremely slow on this hybrid Gated DeltaNet model, see companion row); prompt eval 117-131 tok/s range as published; config=warm-3rd",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "- GPU: RTX 2060 6 GB (sm_75), driver 610.88; - CPU i7-9750H; Windows 11 Pro; - Model: unsloth Qwen3.5-4B, Q4_K_M GGUF; llama-server.exe -m Qwen3.5-4B-Q4_K_M.gguf -ngl 99 -c 16384 -ctk q8_0 -ctv q8_0 --reasoning-budget 0 --no-webui; | CUDA | 3rd+ | 117-131 tok/s | 56.1 tok/s |",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #29513",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29513",
   "tps": "56.1"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": 16384,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "RTX 2060 6 GB",
   "id": "lc-29513-vulkan-first",
   "model": "Qwen3.5-4B",
   "notes": "unsloth Qwen3.5-4B Q4_K_M GGUF; ngl=99, ctx 16384, ctk q8_0, ctv q8_0, reasoning-budget 0; Windows x64 release build b11206 (2b129ccfa), driver 610.88; Vulkan release binary, first chat completion right after model load (Gated DeltaNet warm-up at startup defaults); requests from the third on run at normal speed; config=first-request",
   "pp_tps": 6.4,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "- GPU: RTX 2060 6 GB (sm_75), driver 610.88; - CPU i7-9750H; Windows 11 Pro; - Model: unsloth Qwen3.5-4B, Q4_K_M GGUF; llama-server.exe -m Qwen3.5-4B-Q4_K_M.gguf -ngl 99 -c 16384 -ctk q8_0 -ctv q8_0 --reasoning-budget 0 --no-webui; | Vulkan | 1st | 6.4 tok/s | 15.4 tok/s |",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #29513",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29513",
   "tps": "15.4"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": 16384,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "RTX 2060 6 GB",
   "id": "lc-29513-vulkan-warm",
   "model": "Qwen3.5-4B",
   "notes": "unsloth Qwen3.5-4B Q4_K_M GGUF; ngl=99, ctx 16384, ctk q8_0, ctv q8_0, reasoning-budget 0; Windows x64 release build b11206 (2b129ccfa), driver 610.88; Vulkan release binary, requests from the third on (warm; the first request after load is extremely slow on this hybrid Gated DeltaNet model, see companion row); prompt eval 84-90 tok/s range as published; config=warm-3rd",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "- GPU: RTX 2060 6 GB (sm_75), driver 610.88; - CPU i7-9750H; Windows 11 Pro; - Model: unsloth Qwen3.5-4B, Q4_K_M GGUF; llama-server.exe -m Qwen3.5-4B-Q4_K_M.gguf -ngl 99 -c 16384 -ctk q8_0 -ctv q8_0 --reasoning-budget 0 --no-webui; | Vulkan | 3rd+ | 84-90 tok/s | 47.0 tok/s |",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #29513",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29513",
   "tps": "47.0"
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "4x GB300",
   "id": "vllm-58880-gb300-main-c1",
   "model": "MiniMax-M3",
   "notes": "MiniMax-M3-MXFP8, TP4 on 4x GB300, vLLM main nightly vs this PR; bench: random 1024 in / 256 out, ignore-eos; main baseline (MoEPrepareAndFinalizeNoDPEPModular); config=main",
   "pp_tokens": 1024,
   "provenance": "community",
   "quant": "MXFP8",
   "quote": "**End to end, MiniMax-M3-MXFP8, TP4, 4x GB300:**; vllm serve MiniMaxAI/MiniMax-M3-MXFP8 -tp 4 --max-model-len 8192; vllm bench serve --dataset-name random --random-input-len 1024 --random-output-len 256; | Output throughput, concurrency 1 | 233.0 tok/s | 243.5 tok/s (+4.5%) |; | Output throughput, concurrency 8 | 942.9 tok/s | 966.2 tok/s (+2.5%) |; | Output throughput, concurrency 1 | 233.0 tok/s | 243.5 tok/s (+4.5%) |",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #58880 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58880",
   "tg_tokens": 256,
   "tps": "233.0"
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "4x GB300",
   "id": "vllm-58880-gb300-fused-c1",
   "model": "MiniMax-M3",
   "notes": "MiniMax-M3-MXFP8, TP4 on 4x GB300, vLLM main nightly vs this PR; bench: random 1024 in / 256 out, ignore-eos; PR branch (monolithic MXFP8 TRT-LLM with fused MiniMax2 routing, routed_scaling_factor 2.0); +4.5%; config=fused-routing",
   "pp_tokens": 1024,
   "provenance": "community",
   "quant": "MXFP8",
   "quote": "**End to end, MiniMax-M3-MXFP8, TP4, 4x GB300:**; vllm serve MiniMaxAI/MiniMax-M3-MXFP8 -tp 4 --max-model-len 8192; vllm bench serve --dataset-name random --random-input-len 1024 --random-output-len 256; | Output throughput, concurrency 1 | 233.0 tok/s | 243.5 tok/s (+4.5%) |; | Output throughput, concurrency 8 | 942.9 tok/s | 966.2 tok/s (+2.5%) |; | Output throughput, concurrency 1 | 233.0 tok/s | 243.5 tok/s (+4.5%) |",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #58880 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58880",
   "tg_tokens": 256,
   "tps": "243.5"
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": null,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "4x GB300",
   "id": "vllm-58880-gb300-main-c8",
   "model": "MiniMax-M3",
   "notes": "MiniMax-M3-MXFP8, TP4 on 4x GB300, vLLM main nightly vs this PR; bench: random 1024 in / 256 out, ignore-eos; main baseline, aggregate output throughput at concurrency 8; config=main",
   "pp_tokens": 1024,
   "provenance": "community",
   "quant": "MXFP8",
   "quote": "**End to end, MiniMax-M3-MXFP8, TP4, 4x GB300:**; vllm serve MiniMaxAI/MiniMax-M3-MXFP8 -tp 4 --max-model-len 8192; vllm bench serve --dataset-name random --random-input-len 1024 --random-output-len 256; | Output throughput, concurrency 1 | 233.0 tok/s | 243.5 tok/s (+4.5%) |; | Output throughput, concurrency 8 | 942.9 tok/s | 966.2 tok/s (+2.5%) |; | Output throughput, concurrency 8 | 942.9 tok/s | 966.2 tok/s (+2.5%) |",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #58880 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58880",
   "tg_tokens": 256,
   "tps": "942.9"
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": null,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "4x GB300",
   "id": "vllm-58880-gb300-fused-c8",
   "model": "MiniMax-M3",
   "notes": "MiniMax-M3-MXFP8, TP4 on 4x GB300, vLLM main nightly vs this PR; bench: random 1024 in / 256 out, ignore-eos; PR branch, aggregate output throughput at concurrency 8; +2.5%; config=fused-routing",
   "pp_tokens": 1024,
   "provenance": "community",
   "quant": "MXFP8",
   "quote": "**End to end, MiniMax-M3-MXFP8, TP4, 4x GB300:**; vllm serve MiniMaxAI/MiniMax-M3-MXFP8 -tp 4 --max-model-len 8192; vllm bench serve --dataset-name random --random-input-len 1024 --random-output-len 256; | Output throughput, concurrency 1 | 233.0 tok/s | 243.5 tok/s (+4.5%) |; | Output throughput, concurrency 8 | 942.9 tok/s | 966.2 tok/s (+2.5%) |; | Output throughput, concurrency 8 | 942.9 tok/s | 966.2 tok/s (+2.5%) |",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #58880 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58880",
   "tg_tokens": 256,
   "tps": "966.2"
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "8\u00d7 H20",
   "id": "vllm-58872-h20-main-1k-c1",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "DeepSeek-V4.1-Flash, TP8 on 8x H20 (SM90, 96 GB, NVLink), DSpark synthetic-acceptance speculation (mean acceptance length 3.51, identical in both arms); A1/A2 are main runs, B1/B2 are PR runs (this row cites the A1/B1 cells); 1K in / 1K out at C1, below the 2048-token split threshold, main run A1 (A2 335.2); the PR runs the same kernels here and scores +0.2% mean; config=main",
   "pp_tokens": 1024,
   "provenance": "community",
   "quant": "MXFP8",
   "quote": "Hardware: 8\u00d7 H20 (SM90, 96 GB, NVLink).; vllm serve deepseek-ai/DeepSeek-V4.1-Flash --revision dba1be0a40aa45a94ad051997016db3960a90277; vllm bench serve --backend openai --endpoint /v1/completions --dataset-name random; | 1K in / 1K out, C1 | output tok/s | 335.2 | 335.2 | 338.0 | 334.1 | +0.2% |; | 1K in / 1K out, C64 | output tok/s | 1983.8 | 1987.6 | 2014.9 | 2010.1 | +1.4% |; | 8K in / 1K out, C16 | output tok/s | 601.1 | 602.1 | 621.4 | 620.8 | +3.2% |; | 8K in / 1K out, C64 | output tok/s | 709.9 | 709.4 | 739.0 | 739.4 | +4.2% |",
   "retrieved": "2026-10-01",
   "source_name": "vLLM issue #58872 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58872",
   "tg_tokens": 1024,
   "tps": "335.2"
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "8\u00d7 H20",
   "id": "vllm-58872-h20-split-1k-c1",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "DeepSeek-V4.1-Flash, TP8 on 8x H20 (SM90, 96 GB, NVLink), DSpark synthetic-acceptance speculation (mean acceptance length 3.51, identical in both arms); A1/A2 are main runs, B1/B2 are PR runs (this row cites the A1/B1 cells); 1K in / 1K out at C1, below the 2048-token split threshold, PR run B1 (B2 334.1); the PR notes the two PR servers differ by 1.2% there, more than the gap between the arms; config=split-projections",
   "pp_tokens": 1024,
   "provenance": "community",
   "quant": "MXFP8",
   "quote": "Hardware: 8\u00d7 H20 (SM90, 96 GB, NVLink).; vllm serve deepseek-ai/DeepSeek-V4.1-Flash --revision dba1be0a40aa45a94ad051997016db3960a90277; vllm bench serve --backend openai --endpoint /v1/completions --dataset-name random; | 1K in / 1K out, C1 | output tok/s | 335.2 | 335.2 | 338.0 | 334.1 | +0.2% |; | 1K in / 1K out, C64 | output tok/s | 1983.8 | 1987.6 | 2014.9 | 2010.1 | +1.4% |; | 8K in / 1K out, C16 | output tok/s | 601.1 | 602.1 | 621.4 | 620.8 | +3.2% |; | 8K in / 1K out, C64 | output tok/s | 709.9 | 709.4 | 739.0 | 739.4 | +4.2% |",
   "retrieved": "2026-10-01",
   "source_name": "vLLM issue #58872 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58872",
   "tg_tokens": 1024,
   "tps": "338.0"
  },
  {
   "backend": "vLLM",
   "batch": 64,
   "ctx": null,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "8\u00d7 H20",
   "id": "vllm-58872-h20-main-1k-c64",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "DeepSeek-V4.1-Flash, TP8 on 8x H20 (SM90, 96 GB, NVLink), DSpark synthetic-acceptance speculation (mean acceptance length 3.51, identical in both arms); A1/A2 are main runs, B1/B2 are PR runs (this row cites the A1/B1 cells); 1K in / 1K out at C64, aggregate output throughput, main run A1 (A2 1987.6); config=main",
   "pp_tokens": 1024,
   "provenance": "community",
   "quant": "MXFP8",
   "quote": "Hardware: 8\u00d7 H20 (SM90, 96 GB, NVLink).; vllm serve deepseek-ai/DeepSeek-V4.1-Flash --revision dba1be0a40aa45a94ad051997016db3960a90277; vllm bench serve --backend openai --endpoint /v1/completions --dataset-name random; | 1K in / 1K out, C1 | output tok/s | 335.2 | 335.2 | 338.0 | 334.1 | +0.2% |; | 1K in / 1K out, C64 | output tok/s | 1983.8 | 1987.6 | 2014.9 | 2010.1 | +1.4% |; | 8K in / 1K out, C16 | output tok/s | 601.1 | 602.1 | 621.4 | 620.8 | +3.2% |; | 8K in / 1K out, C64 | output tok/s | 709.9 | 709.4 | 739.0 | 739.4 | +4.2% |",
   "retrieved": "2026-10-01",
   "source_name": "vLLM issue #58872 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58872",
   "tg_tokens": 1024,
   "tps": "1983.8"
  },
  {
   "backend": "vLLM",
   "batch": 64,
   "ctx": null,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "8\u00d7 H20",
   "id": "vllm-58872-h20-split-1k-c64",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "DeepSeek-V4.1-Flash, TP8 on 8x H20 (SM90, 96 GB, NVLink), DSpark synthetic-acceptance speculation (mean acceptance length 3.51, identical in both arms); A1/A2 are main runs, B1/B2 are PR runs (this row cites the A1/B1 cells); 1K in / 1K out at C64, aggregate output throughput, PR run B1 (B2 2010.1); +1.4% mean; config=split-projections",
   "pp_tokens": 1024,
   "provenance": "community",
   "quant": "MXFP8",
   "quote": "Hardware: 8\u00d7 H20 (SM90, 96 GB, NVLink).; vllm serve deepseek-ai/DeepSeek-V4.1-Flash --revision dba1be0a40aa45a94ad051997016db3960a90277; vllm bench serve --backend openai --endpoint /v1/completions --dataset-name random; | 1K in / 1K out, C1 | output tok/s | 335.2 | 335.2 | 338.0 | 334.1 | +0.2% |; | 1K in / 1K out, C64 | output tok/s | 1983.8 | 1987.6 | 2014.9 | 2010.1 | +1.4% |; | 8K in / 1K out, C16 | output tok/s | 601.1 | 602.1 | 621.4 | 620.8 | +3.2% |; | 8K in / 1K out, C64 | output tok/s | 709.9 | 709.4 | 739.0 | 739.4 | +4.2% |",
   "retrieved": "2026-10-01",
   "source_name": "vLLM issue #58872 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58872",
   "tg_tokens": 1024,
   "tps": "2014.9"
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": null,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "8\u00d7 H20",
   "id": "vllm-58872-h20-main-8k-c16",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "DeepSeek-V4.1-Flash, TP8 on 8x H20 (SM90, 96 GB, NVLink), DSpark synthetic-acceptance speculation (mean acceptance length 3.51, identical in both arms); A1/A2 are main runs, B1/B2 are PR runs (this row cites the A1/B1 cells); 8K in / 1K out at C16 (8192-token prefill crosses the 2048 threshold), aggregate output throughput, main run A1 (A2 602.1); config=main",
   "pp_tokens": 8192,
   "provenance": "community",
   "quant": "MXFP8",
   "quote": "Hardware: 8\u00d7 H20 (SM90, 96 GB, NVLink).; vllm serve deepseek-ai/DeepSeek-V4.1-Flash --revision dba1be0a40aa45a94ad051997016db3960a90277; vllm bench serve --backend openai --endpoint /v1/completions --dataset-name random; | 1K in / 1K out, C1 | output tok/s | 335.2 | 335.2 | 338.0 | 334.1 | +0.2% |; | 1K in / 1K out, C64 | output tok/s | 1983.8 | 1987.6 | 2014.9 | 2010.1 | +1.4% |; | 8K in / 1K out, C16 | output tok/s | 601.1 | 602.1 | 621.4 | 620.8 | +3.2% |; | 8K in / 1K out, C64 | output tok/s | 709.9 | 709.4 | 739.0 | 739.4 | +4.2% |",
   "retrieved": "2026-10-01",
   "source_name": "vLLM issue #58872 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58872",
   "tg_tokens": 1024,
   "tps": "601.1"
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": null,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "8\u00d7 H20",
   "id": "vllm-58872-h20-split-8k-c16",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "DeepSeek-V4.1-Flash, TP8 on 8x H20 (SM90, 96 GB, NVLink), DSpark synthetic-acceptance speculation (mean acceptance length 3.51, identical in both arms); A1/A2 are main runs, B1/B2 are PR runs (this row cites the A1/B1 cells); 8K in / 1K out at C16, aggregate output throughput, PR run B1 (B2 620.8); +3.2% mean; config=split-projections",
   "pp_tokens": 8192,
   "provenance": "community",
   "quant": "MXFP8",
   "quote": "Hardware: 8\u00d7 H20 (SM90, 96 GB, NVLink).; vllm serve deepseek-ai/DeepSeek-V4.1-Flash --revision dba1be0a40aa45a94ad051997016db3960a90277; vllm bench serve --backend openai --endpoint /v1/completions --dataset-name random; | 1K in / 1K out, C1 | output tok/s | 335.2 | 335.2 | 338.0 | 334.1 | +0.2% |; | 1K in / 1K out, C64 | output tok/s | 1983.8 | 1987.6 | 2014.9 | 2010.1 | +1.4% |; | 8K in / 1K out, C16 | output tok/s | 601.1 | 602.1 | 621.4 | 620.8 | +3.2% |; | 8K in / 1K out, C64 | output tok/s | 709.9 | 709.4 | 739.0 | 739.4 | +4.2% |",
   "retrieved": "2026-10-01",
   "source_name": "vLLM issue #58872 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58872",
   "tg_tokens": 1024,
   "tps": "621.4"
  },
  {
   "backend": "vLLM",
   "batch": 64,
   "ctx": null,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "8\u00d7 H20",
   "id": "vllm-58872-h20-main-8k-c64",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "DeepSeek-V4.1-Flash, TP8 on 8x H20 (SM90, 96 GB, NVLink), DSpark synthetic-acceptance speculation (mean acceptance length 3.51, identical in both arms); A1/A2 are main runs, B1/B2 are PR runs (this row cites the A1/B1 cells); 8K in / 1K out at C64, aggregate output throughput, main run A1 (A2 709.4); config=main",
   "pp_tokens": 8192,
   "provenance": "community",
   "quant": "MXFP8",
   "quote": "Hardware: 8\u00d7 H20 (SM90, 96 GB, NVLink).; vllm serve deepseek-ai/DeepSeek-V4.1-Flash --revision dba1be0a40aa45a94ad051997016db3960a90277; vllm bench serve --backend openai --endpoint /v1/completions --dataset-name random; | 1K in / 1K out, C1 | output tok/s | 335.2 | 335.2 | 338.0 | 334.1 | +0.2% |; | 1K in / 1K out, C64 | output tok/s | 1983.8 | 1987.6 | 2014.9 | 2010.1 | +1.4% |; | 8K in / 1K out, C16 | output tok/s | 601.1 | 602.1 | 621.4 | 620.8 | +3.2% |; | 8K in / 1K out, C64 | output tok/s | 709.9 | 709.4 | 739.0 | 739.4 | +4.2% |",
   "retrieved": "2026-10-01",
   "source_name": "vLLM issue #58872 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58872",
   "tg_tokens": 1024,
   "tps": "709.9"
  },
  {
   "backend": "vLLM",
   "batch": 64,
   "ctx": null,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "8\u00d7 H20",
   "id": "vllm-58872-h20-split-8k-c64",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "DeepSeek-V4.1-Flash, TP8 on 8x H20 (SM90, 96 GB, NVLink), DSpark synthetic-acceptance speculation (mean acceptance length 3.51, identical in both arms); A1/A2 are main runs, B1/B2 are PR runs (this row cites the A1/B1 cells); 8K in / 1K out at C64, aggregate output throughput, PR run B1 (B2 739.4); +4.2% mean; config=split-projections",
   "pp_tokens": 8192,
   "provenance": "community",
   "quant": "MXFP8",
   "quote": "Hardware: 8\u00d7 H20 (SM90, 96 GB, NVLink).; vllm serve deepseek-ai/DeepSeek-V4.1-Flash --revision dba1be0a40aa45a94ad051997016db3960a90277; vllm bench serve --backend openai --endpoint /v1/completions --dataset-name random; | 1K in / 1K out, C1 | output tok/s | 335.2 | 335.2 | 338.0 | 334.1 | +0.2% |; | 1K in / 1K out, C64 | output tok/s | 1983.8 | 1987.6 | 2014.9 | 2010.1 | +1.4% |; | 8K in / 1K out, C16 | output tok/s | 601.1 | 602.1 | 621.4 | 620.8 | +3.2% |; | 8K in / 1K out, C64 | output tok/s | 709.9 | 709.4 | 739.0 | 739.4 | +4.2% |",
   "retrieved": "2026-10-01",
   "source_name": "vLLM issue #58872 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58872",
   "tg_tokens": 1024,
   "tps": "739.0"
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-26",
   "flags": [],
   "hardware": "TP4",
   "id": "vllm-58845-glm53-main-c1",
   "model": "GLM-5.3",
   "notes": "zai-org/GLM-5.3, -tp 4, no explicit quantization flag (served at native checkpoint precision, BF16), MTP speculation with 5 draft tokens, guidellm synthetic 8000 prompt / 1000 output tokens, max-num-seqs 48; guidellm Throughput column (output tokens/s), main arm, streams 1 (175.4 vs 176.8 diff); avg latency 5.68 s, TTFT 360.5 ms; config=main",
   "pp_tokens": 8000,
   "provenance": "community",
   "quant": "BF16",
   "quote": "vllm serve zai-org/GLM-5.3 \\; -tp 4; method\":\"mtp\",\"num_speculative_tokens\":5; prompt_tokens=8000,output_tokens=1000; 1,4,16; | 1 | 5.68 s | 5.64 s | 175.4 tok/s | 176.8 tok/s\uff09 | 360.5 ms | **332.9 ms\uff08\u22127.7%\uff09** \ud83d\ude80 |; | 4 | 10.08 s | 10.57 s | 412.4 tok/s | 398.5 tok/s | 406.1 ms | **377.9 ms\uff08\u22127.0%\uff09** \ud83d\ude80 |; | 16 | 19.89 s | 18.99 s | 826.2 tok/s | 870.4 tok/s | 454.7 ms | **434.5 ms\uff08\u22124.4%\uff09** \ud83d\ude80 |",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #58845 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58845",
   "tg_tokens": 1000,
   "tps": "175.4"
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-26",
   "flags": [],
   "hardware": "TP4",
   "id": "vllm-58845-glm53-skip-c1",
   "model": "GLM-5.3",
   "notes": "zai-org/GLM-5.3, -tp 4, no explicit quantization flag (served at native checkpoint precision, BF16), MTP speculation with 5 draft tokens, guidellm synthetic 8000 prompt / 1000 output tokens, max-num-seqs 48; guidellm Throughput column (output tokens/s), PR arm (skip qlnorm for MHA), streams 1; avg latency 5.64 s, TTFT 332.9 ms (-7.7%); config=skip-qlnorm",
   "pp_tokens": 8000,
   "provenance": "community",
   "quant": "BF16",
   "quote": "vllm serve zai-org/GLM-5.3 \\; -tp 4; method\":\"mtp\",\"num_speculative_tokens\":5; prompt_tokens=8000,output_tokens=1000; 1,4,16; | 1 | 5.68 s | 5.64 s | 175.4 tok/s | 176.8 tok/s\uff09 | 360.5 ms | **332.9 ms\uff08\u22127.7%\uff09** \ud83d\ude80 |; | 4 | 10.08 s | 10.57 s | 412.4 tok/s | 398.5 tok/s | 406.1 ms | **377.9 ms\uff08\u22127.0%\uff09** \ud83d\ude80 |; | 16 | 19.89 s | 18.99 s | 826.2 tok/s | 870.4 tok/s | 454.7 ms | **434.5 ms\uff08\u22124.4%\uff09** \ud83d\ude80 |",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #58845 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58845",
   "tg_tokens": 1000,
   "tps": "176.8"
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": null,
   "date": "2026-09-26",
   "flags": [],
   "hardware": "TP4",
   "id": "vllm-58845-glm53-main-c4",
   "model": "GLM-5.3",
   "notes": "zai-org/GLM-5.3, -tp 4, no explicit quantization flag (served at native checkpoint precision, BF16), MTP speculation with 5 draft tokens, guidellm synthetic 8000 prompt / 1000 output tokens, max-num-seqs 48; guidellm Throughput column (output tokens/s), main arm, streams 4, aggregate (398.5 diff); avg latency 10.08 s; config=main",
   "pp_tokens": 8000,
   "provenance": "community",
   "quant": "BF16",
   "quote": "vllm serve zai-org/GLM-5.3 \\; -tp 4; method\":\"mtp\",\"num_speculative_tokens\":5; prompt_tokens=8000,output_tokens=1000; 1,4,16; | 1 | 5.68 s | 5.64 s | 175.4 tok/s | 176.8 tok/s\uff09 | 360.5 ms | **332.9 ms\uff08\u22127.7%\uff09** \ud83d\ude80 |; | 4 | 10.08 s | 10.57 s | 412.4 tok/s | 398.5 tok/s | 406.1 ms | **377.9 ms\uff08\u22127.0%\uff09** \ud83d\ude80 |; | 16 | 19.89 s | 18.99 s | 826.2 tok/s | 870.4 tok/s | 454.7 ms | **434.5 ms\uff08\u22124.4%\uff09** \ud83d\ude80 |",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #58845 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58845",
   "tg_tokens": 1000,
   "tps": "412.4"
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": null,
   "date": "2026-09-26",
   "flags": [],
   "hardware": "TP4",
   "id": "vllm-58845-glm53-skip-c4",
   "model": "GLM-5.3",
   "notes": "zai-org/GLM-5.3, -tp 4, no explicit quantization flag (served at native checkpoint precision, BF16), MTP speculation with 5 draft tokens, guidellm synthetic 8000 prompt / 1000 output tokens, max-num-seqs 48; guidellm Throughput column (output tokens/s), PR arm (skip qlnorm for MHA), streams 4, aggregate; avg latency 10.57 s, TTFT 377.9 ms (-7.0%); config=skip-qlnorm",
   "pp_tokens": 8000,
   "provenance": "community",
   "quant": "BF16",
   "quote": "vllm serve zai-org/GLM-5.3 \\; -tp 4; method\":\"mtp\",\"num_speculative_tokens\":5; prompt_tokens=8000,output_tokens=1000; 1,4,16; | 1 | 5.68 s | 5.64 s | 175.4 tok/s | 176.8 tok/s\uff09 | 360.5 ms | **332.9 ms\uff08\u22127.7%\uff09** \ud83d\ude80 |; | 4 | 10.08 s | 10.57 s | 412.4 tok/s | 398.5 tok/s | 406.1 ms | **377.9 ms\uff08\u22127.0%\uff09** \ud83d\ude80 |; | 16 | 19.89 s | 18.99 s | 826.2 tok/s | 870.4 tok/s | 454.7 ms | **434.5 ms\uff08\u22124.4%\uff09** \ud83d\ude80 |",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #58845 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58845",
   "tg_tokens": 1000,
   "tps": "398.5"
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": null,
   "date": "2026-09-26",
   "flags": [],
   "hardware": "TP4",
   "id": "vllm-58845-glm53-main-c16",
   "model": "GLM-5.3",
   "notes": "zai-org/GLM-5.3, -tp 4, no explicit quantization flag (served at native checkpoint precision, BF16), MTP speculation with 5 draft tokens, guidellm synthetic 8000 prompt / 1000 output tokens, max-num-seqs 48; guidellm Throughput column (output tokens/s), main arm, streams 16, aggregate (870.4 diff); avg latency 19.89 s; config=main",
   "pp_tokens": 8000,
   "provenance": "community",
   "quant": "BF16",
   "quote": "vllm serve zai-org/GLM-5.3 \\; -tp 4; method\":\"mtp\",\"num_speculative_tokens\":5; prompt_tokens=8000,output_tokens=1000; 1,4,16; | 1 | 5.68 s | 5.64 s | 175.4 tok/s | 176.8 tok/s\uff09 | 360.5 ms | **332.9 ms\uff08\u22127.7%\uff09** \ud83d\ude80 |; | 4 | 10.08 s | 10.57 s | 412.4 tok/s | 398.5 tok/s | 406.1 ms | **377.9 ms\uff08\u22127.0%\uff09** \ud83d\ude80 |; | 16 | 19.89 s | 18.99 s | 826.2 tok/s | 870.4 tok/s | 454.7 ms | **434.5 ms\uff08\u22124.4%\uff09** \ud83d\ude80 |",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #58845 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58845",
   "tg_tokens": 1000,
   "tps": "826.2"
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": null,
   "date": "2026-09-26",
   "flags": [],
   "hardware": "TP4",
   "id": "vllm-58845-glm53-skip-c16",
   "model": "GLM-5.3",
   "notes": "zai-org/GLM-5.3, -tp 4, no explicit quantization flag (served at native checkpoint precision, BF16), MTP speculation with 5 draft tokens, guidellm synthetic 8000 prompt / 1000 output tokens, max-num-seqs 48; guidellm Throughput column (output tokens/s), PR arm (skip qlnorm for MHA), streams 16, aggregate; avg latency 18.99 s, TTFT 434.5 ms (-4.4%); config=skip-qlnorm",
   "pp_tokens": 8000,
   "provenance": "community",
   "quant": "BF16",
   "quote": "vllm serve zai-org/GLM-5.3 \\; -tp 4; method\":\"mtp\",\"num_speculative_tokens\":5; prompt_tokens=8000,output_tokens=1000; 1,4,16; | 1 | 5.68 s | 5.64 s | 175.4 tok/s | 176.8 tok/s\uff09 | 360.5 ms | **332.9 ms\uff08\u22127.7%\uff09** \ud83d\ude80 |; | 4 | 10.08 s | 10.57 s | 412.4 tok/s | 398.5 tok/s | 406.1 ms | **377.9 ms\uff08\u22127.0%\uff09** \ud83d\ude80 |; | 16 | 19.89 s | 18.99 s | 826.2 tok/s | 870.4 tok/s | 454.7 ms | **434.5 ms\uff08\u22124.4%\uff09** \ud83d\ude80 |",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #58845 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58845",
   "tg_tokens": 1000,
   "tps": "870.4"
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": null,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-58887-mi355x-main-c4",
   "model": "Kimi-K3",
   "notes": "moonshotai/Kimi-K3, TP8 on MI355X (gfx950; 8x per the #58886 env table), vllm/vllm-openai-rocm nightly-rocm100 (29468dde), ROCm 10, VLLM_ROCM_USE_AITER_MLA, FP8 KV cache, async scheduling, dummy weights (throughput run, not a quality run), sglang bench_serving random 8k in / 1k out, aggregate output tok/s; A/B of the AITER MLA FP8 prefill metadata race, before vs after the pinned-host staging fix; no explicit weight quantization flag (BF16 assumed, KV cache fp8); bug reported in issue #58886; config=current-staging",
   "pp_tokens": 8192,
   "provenance": "community",
   "quant": "BF16",
   "quote": "Kimi-K3; MI355X, TP8, ROCm 10 nightly (`29468dde`), FP8 KV cache, async scheduling.; Performance, `sglang.bench_serving` random 8k/1k, on streams that do not crash before the fix:; | 4 | 236.05 / 236.05 | 15.49 / 15.50 ms |",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #58887 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58887",
   "tg_tokens": 1024,
   "tps": "236.05"
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": null,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-58887-mi355x-fix-c4",
   "model": "Kimi-K3",
   "notes": "moonshotai/Kimi-K3, TP8 on MI355X (gfx950; 8x per the #58886 env table), vllm/vllm-openai-rocm nightly-rocm100 (29468dde), ROCm 10, VLLM_ROCM_USE_AITER_MLA, FP8 KV cache, async scheduling, dummy weights (throughput run, not a quality run), sglang bench_serving random 8k in / 1k out, aggregate output tok/s; A/B of the AITER MLA FP8 prefill metadata race, before vs after the pinned-host staging fix; no explicit weight quantization flag (BF16 assumed, KV cache fp8); bug reported in issue #58886; config=pinned-host-staging",
   "pp_tokens": 8192,
   "provenance": "community",
   "quant": "BF16",
   "quote": "Kimi-K3; MI355X, TP8, ROCm 10 nightly (`29468dde`), FP8 KV cache, async scheduling.; Performance, `sglang.bench_serving` random 8k/1k, on streams that do not crash before the fix:; | 4 | 236.05 / 236.05 | 15.49 / 15.50 ms |",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #58887 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58887",
   "tg_tokens": 1024,
   "tps": "236.05"
  },
  {
   "backend": "vLLM",
   "batch": 32,
   "ctx": null,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-58887-mi355x-main-c32",
   "model": "Kimi-K3",
   "notes": "moonshotai/Kimi-K3, TP8 on MI355X (gfx950; 8x per the #58886 env table), vllm/vllm-openai-rocm nightly-rocm100 (29468dde), ROCm 10, VLLM_ROCM_USE_AITER_MLA, FP8 KV cache, async scheduling, dummy weights (throughput run, not a quality run), sglang bench_serving random 8k in / 1k out, aggregate output tok/s; A/B of the AITER MLA FP8 prefill metadata race, before vs after the pinned-host staging fix; no explicit weight quantization flag (BF16 assumed, KV cache fp8); bug reported in issue #58886; config=current-staging",
   "pp_tokens": 8192,
   "provenance": "community",
   "quant": "BF16",
   "quote": "Kimi-K3; MI355X, TP8, ROCm 10 nightly (`29468dde`), FP8 KV cache, async scheduling.; Performance, `sglang.bench_serving` random 8k/1k, on streams that do not crash before the fix:; | 32 | 1083.0 / 1082.6 | 18.14 / 18.14 ms |",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #58887 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58887",
   "tg_tokens": 1024,
   "tps": "1083.0"
  },
  {
   "backend": "vLLM",
   "batch": 32,
   "ctx": null,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-58887-mi355x-fix-c32",
   "model": "Kimi-K3",
   "notes": "moonshotai/Kimi-K3, TP8 on MI355X (gfx950; 8x per the #58886 env table), vllm/vllm-openai-rocm nightly-rocm100 (29468dde), ROCm 10, VLLM_ROCM_USE_AITER_MLA, FP8 KV cache, async scheduling, dummy weights (throughput run, not a quality run), sglang bench_serving random 8k in / 1k out, aggregate output tok/s; A/B of the AITER MLA FP8 prefill metadata race, before vs after the pinned-host staging fix; no explicit weight quantization flag (BF16 assumed, KV cache fp8); bug reported in issue #58886; config=pinned-host-staging",
   "pp_tokens": 8192,
   "provenance": "community",
   "quant": "BF16",
   "quote": "Kimi-K3; MI355X, TP8, ROCm 10 nightly (`29468dde`), FP8 KV cache, async scheduling.; Performance, `sglang.bench_serving` random 8k/1k, on streams that do not crash before the fix:; | 32 | 1083.0 / 1082.6 | 18.14 / 18.14 ms |",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #58887 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58887",
   "tg_tokens": 1024,
   "tps": "1082.6"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 262144,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc-29523-2x3090-released",
   "model": "Qwen3.8-27B",
   "notes": "build b29c606 (v0.4.1), 2x RTX 3090, --split-mode layer, --tensor-split 0.525,0.475, --spec-type draft-mtp, draft acceptance 71.4%, 10464-token prompt with cache_prompt false, 200 tokens greedy, 3 rounds per arm; quant not named in the PR (the #27428 report used unsloth Qwen3.8-27B GGUF Q6_K_XL / Q3_K_XL); the released build also OOMs the MTP draft context buffer on device 1 and silently drops pipeline parallelism; config=as-released",
   "pp_tokens": 10464,
   "pp_tps": "1132.0",
   "provenance": "community",
   "quant": "not stated",
   "quote": "Two builds of `b29c606` (v0.4.1) differing only by this branch, same machine (2x RTX 3090, `--split-mode layer`, `--tensor-split 0.525,0.475`), Qwen3.8-27B hybrid (arch `qwen35`), `--spec-type draft-mtp`, `--ctx-size 262144`.; Identical 10464-token prompt, greedy sampling, `cache_prompt: false` so every round pays the full prefill, 200 tokens generated, 3 rounds per arm.; Prompt processing is the warm rounds (2 and 3); including the cold first round the figures are 1111.5 and 1779.1 t/s, so +60% rather than +63%.; | | prompt processing | generation |; | `b29c606` as released | 1132.0 t/s | 57.98 t/s |; Generation is the mean of all three rounds, with a round-to-round spread of 2.7% and 2.2% respectively; and draft acceptance was 71.4% in both arms, so nothing about generation is being traded away.",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp PR #29523 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29523",
   "tg_tokens": 200,
   "tps": "57.98"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 262144,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc-29523-2x3090-mtp-lag",
   "model": "Qwen3.8-27B",
   "notes": "build b29c606 (v0.4.1), 2x RTX 3090, --split-mode layer, --tensor-split 0.525,0.475, --spec-type draft-mtp, draft acceptance 71.4%, 10464-token prompt with cache_prompt false, 200 tokens greedy, 3 rounds per arm; quant not named in the PR (the #27428 report used unsloth Qwen3.8-27B GGUF Q6_K_XL / Q3_K_XL); the released build also OOMs the MTP draft context buffer on device 1 and silently drops pipeline parallelism; config=mtp-catchup-lag",
   "pp_tokens": 10464,
   "pp_tps": "1848.8",
   "provenance": "community",
   "quant": "not stated",
   "quote": "Two builds of `b29c606` (v0.4.1) differing only by this branch, same machine (2x RTX 3090, `--split-mode layer`, `--tensor-split 0.525,0.475`), Qwen3.8-27B hybrid (arch `qwen35`), `--spec-type draft-mtp`, `--ctx-size 262144`.; Identical 10464-token prompt, greedy sampling, `cache_prompt: false` so every round pays the full prefill, 200 tokens generated, 3 rounds per arm.; Prompt processing is the warm rounds (2 and 3); including the cold first round the figures are 1111.5 and 1779.1 t/s, so +60% rather than +63%.; | | prompt processing | generation |; | this branch | **1848.8 t/s** | 59.98 t/s |; Generation is the mean of all three rounds, with a round-to-round spread of 2.7% and 2.2% respectively; and draft acceptance was 71.4% in both arms, so nothing about generation is being traded away.",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp PR #29523 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29523",
   "tg_tokens": 200,
   "tps": "59.98"
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-12",
   "flags": [],
   "hardware": "8x NVIDIA H100 80GB HBM3",
   "id": "vllm-56564-h100fa-c1",
   "model": "GLM-5.3-Flash",
   "notes": "zai-org/GLM-5.3-Flash, TP8 (single node, 8x H100 80GB HBM3), vLLM nightly 385dce36b with FlashInfer 0.6.18, vllm bench serve on the same container, sonnet dataset 300 in / 256 out (20 prompts, 2 warmups, ignore eos, temperature 0, seed 0), MTP 5-token speculation (accept len 4.89, rate 77.8% on this arm), aggregate output tok/s, no explicit weight quantization flag (BF16 assumed); backend forced with --attention-backend FLASH_ATTN_MLA_SPARSE; MTP draft acceptance identical across the two backends; random dataset worst case c1: 249.3 vs 162.9; same ratio with --enable-expert-parallel and the Hopper recipe override; config=flash-attn-sparse",
   "pp_tokens": 300,
   "provenance": "community",
   "quant": "BF16",
   "quote": "GPU                          : 8x NVIDIA H100 80GB HBM3 (single node, TP=8); **Summary:** For `zai-org/GLM-5.3-Flash` on H100 the auto-selected sparse-MLA attention backend is `FLASHINFER_MLA_SPARSE_SM90`;   --tensor-parallel-size 8 \\;   --speculative-config '{\"method\":\"mtp\",\"num_speculative_tokens\":5}' \\;   --dataset-name sonnet --dataset-path /vllm-workspace/benchmarks/sonnet.txt \\;   --sonnet-input-len 300 --sonnet-output-len 256 --sonnet-prefix-len 100 \\;   --num-prompts 20 --num-warmups 2 --max-concurrency {1,8} --ignore-eos --disable-shuffle --temperature 0 --seed 0; | sonnet | 1 | `FLASH_ATTN_MLA_SPARSE` (forced) | **368.8** | 2.28 ms | 10.9 ms | 698 ms | 4.89 / 77.8% |",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #56564 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/56564",
   "tg_tokens": 256,
   "tps": "368.8"
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-12",
   "flags": [],
   "hardware": "8x NVIDIA H100 80GB HBM3",
   "id": "vllm-56564-h100fi-c1",
   "model": "GLM-5.3-Flash",
   "notes": "zai-org/GLM-5.3-Flash, TP8 (single node, 8x H100 80GB HBM3), vLLM nightly 385dce36b with FlashInfer 0.6.18, vllm bench serve on the same container, sonnet dataset 300 in / 256 out (20 prompts, 2 warmups, ignore eos, temperature 0, seed 0), MTP 5-token speculation (accept len 4.89, rate 77.7% on this arm), aggregate output tok/s, no explicit weight quantization flag (BF16 assumed); backend auto-selected as the SM90 default; issue reports it 36% slower at concurrency 1 and 70% at concurrency 8 vs the forced FLASH_ATTN_MLA_SPARSE, same MTP acceptance so the gap is the per-step attention kernel; MTP draft acceptance identical across the two backends; random dataset worst case c1: 249.3 vs 162.9; same ratio with --enable-expert-parallel and the Hopper recipe override; config=flashinfer-sm90-sparse",
   "pp_tokens": 300,
   "provenance": "community",
   "quant": "BF16",
   "quote": "GPU                          : 8x NVIDIA H100 80GB HBM3 (single node, TP=8); **Summary:** For `zai-org/GLM-5.3-Flash` on H100 the auto-selected sparse-MLA attention backend is `FLASHINFER_MLA_SPARSE_SM90`;   --tensor-parallel-size 8 \\;   --speculative-config '{\"method\":\"mtp\",\"num_speculative_tokens\":5}' \\;   --dataset-name sonnet --dataset-path /vllm-workspace/benchmarks/sonnet.txt \\;   --sonnet-input-len 300 --sonnet-output-len 256 --sonnet-prefix-len 100 \\;   --num-prompts 20 --num-warmups 2 --max-concurrency {1,8} --ignore-eos --disable-shuffle --temperature 0 --seed 0; | sonnet | 1 | `FLASHINFER_MLA_SPARSE_SM90` (default) | 237.0 | 3.83 ms | 18.3 ms | 1064 ms | 4.89 / 77.7% |",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #56564 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/56564",
   "tg_tokens": 256,
   "tps": "237.0"
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": null,
   "date": "2026-09-12",
   "flags": [],
   "hardware": "8x NVIDIA H100 80GB HBM3",
   "id": "vllm-56564-h100fa-c8",
   "model": "GLM-5.3-Flash",
   "notes": "zai-org/GLM-5.3-Flash, TP8 (single node, 8x H100 80GB HBM3), vLLM nightly 385dce36b with FlashInfer 0.6.18, vllm bench serve on the same container, sonnet dataset 300 in / 256 out (20 prompts, 2 warmups, ignore eos, temperature 0, seed 0), MTP 5-token speculation (accept len 4.90, rate 77.9% on this arm), aggregate output tok/s, no explicit weight quantization flag (BF16 assumed); backend forced with --attention-backend FLASH_ATTN_MLA_SPARSE; MTP draft acceptance identical across the two backends; random dataset worst case c1: 249.3 vs 162.9; same ratio with --enable-expert-parallel and the Hopper recipe override; config=flash-attn-sparse",
   "pp_tokens": 300,
   "provenance": "community",
   "quant": "BF16",
   "quote": "GPU                          : 8x NVIDIA H100 80GB HBM3 (single node, TP=8); **Summary:** For `zai-org/GLM-5.3-Flash` on H100 the auto-selected sparse-MLA attention backend is `FLASHINFER_MLA_SPARSE_SM90`;   --tensor-parallel-size 8 \\;   --speculative-config '{\"method\":\"mtp\",\"num_speculative_tokens\":5}' \\;   --dataset-name sonnet --dataset-path /vllm-workspace/benchmarks/sonnet.txt \\;   --sonnet-input-len 300 --sonnet-output-len 256 --sonnet-prefix-len 100 \\;   --num-prompts 20 --num-warmups 2 --max-concurrency {1,8} --ignore-eos --disable-shuffle --temperature 0 --seed 0; | sonnet | 8 | `FLASH_ATTN_MLA_SPARSE` (forced) | **754.1** | 5.00 ms | 23.9 ms | 1914 ms | 4.90 / 77.9% |",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #56564 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/56564",
   "tg_tokens": 256,
   "tps": "754.1"
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": null,
   "date": "2026-09-12",
   "flags": [],
   "hardware": "8x NVIDIA H100 80GB HBM3",
   "id": "vllm-56564-h100fi-c8",
   "model": "GLM-5.3-Flash",
   "notes": "zai-org/GLM-5.3-Flash, TP8 (single node, 8x H100 80GB HBM3), vLLM nightly 385dce36b with FlashInfer 0.6.18, vllm bench serve on the same container, sonnet dataset 300 in / 256 out (20 prompts, 2 warmups, ignore eos, temperature 0, seed 0), MTP 5-token speculation (accept len 4.87, rate 77.3% on this arm), aggregate output tok/s, no explicit weight quantization flag (BF16 assumed); backend auto-selected as the SM90 default; issue reports it 36% slower at concurrency 1 and 70% at concurrency 8 vs the forced FLASH_ATTN_MLA_SPARSE, same MTP acceptance so the gap is the per-step attention kernel; MTP draft acceptance identical across the two backends; random dataset worst case c1: 249.3 vs 162.9; same ratio with --enable-expert-parallel and the Hopper recipe override; config=flashinfer-sm90-sparse",
   "pp_tokens": 300,
   "provenance": "community",
   "quant": "BF16",
   "quote": "GPU                          : 8x NVIDIA H100 80GB HBM3 (single node, TP=8); **Summary:** For `zai-org/GLM-5.3-Flash` on H100 the auto-selected sparse-MLA attention backend is `FLASHINFER_MLA_SPARSE_SM90`;   --tensor-parallel-size 8 \\;   --speculative-config '{\"method\":\"mtp\",\"num_speculative_tokens\":5}' \\;   --dataset-name sonnet --dataset-path /vllm-workspace/benchmarks/sonnet.txt \\;   --sonnet-input-len 300 --sonnet-output-len 256 --sonnet-prefix-len 100 \\;   --num-prompts 20 --num-warmups 2 --max-concurrency {1,8} --ignore-eos --disable-shuffle --temperature 0 --seed 0; | sonnet | 8 | `FLASHINFER_MLA_SPARSE_SM90` (default) | 229.8 | 17.4 ms | 82.9 ms | 5470 ms | 4.87 / 77.3% |",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #56564 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/56564",
   "tg_tokens": 256,
   "tps": "229.8"
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-12",
   "flags": [],
   "hardware": "8x NVIDIA H100 80GB HBM3",
   "id": "vllm-56564-h100fa-c1-random",
   "model": "GLM-5.3-Flash",
   "notes": "zai-org/GLM-5.3-Flash, TP8 (single node, 8x H100 80GB HBM3), vLLM nightly 385dce36b with FlashInfer 0.6.18, vllm bench serve on the same container, random dataset 300 in / 256 out (worst case for MTP on this model; ratio between backends the same as sonnet) (20 prompts, 2 warmups, ignore eos, temperature 0, seed 0), MTP 5-token speculation (accept len 3.08, rate 41.7% on this arm), aggregate output tok/s, no explicit weight quantization flag (BF16 assumed); backend forced with --attention-backend FLASH_ATTN_MLA_SPARSE; MTP draft acceptance identical across the two backends; same ratio with --enable-expert-parallel and the Hopper recipe override; config=flash-attn-sparse-random",
   "pp_tokens": 300,
   "provenance": "community",
   "quant": "BF16",
   "quote": "GPU                          : 8x NVIDIA H100 80GB HBM3 (single node, TP=8); **Summary:** For `zai-org/GLM-5.3-Flash` on H100 the auto-selected sparse-MLA attention backend is `FLASHINFER_MLA_SPARSE_SM90`;   --tensor-parallel-size 8 \\;   --speculative-config '{\"method\":\"mtp\",\"num_speculative_tokens\":5}' \\; # \"random\" rows: --dataset-name random --random-input-len 300 --random-output-len 256;   --num-prompts 20 --num-warmups 2 --max-concurrency {1,8} --ignore-eos --disable-shuffle --temperature 0 --seed 0; | random | 1 | `FLASH_ATTN_MLA_SPARSE` (forced) | **249.3** | 3.59 ms | 10.9 ms | 1006 ms | 3.08 / 41.7% |",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #56564 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/56564",
   "tg_tokens": 256,
   "tps": "249.3"
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-12",
   "flags": [],
   "hardware": "8x NVIDIA H100 80GB HBM3",
   "id": "vllm-56564-h100fi-c1-random",
   "model": "GLM-5.3-Flash",
   "notes": "zai-org/GLM-5.3-Flash, TP8 (single node, 8x H100 80GB HBM3), vLLM nightly 385dce36b with FlashInfer 0.6.18, vllm bench serve on the same container, random dataset 300 in / 256 out (worst case for MTP on this model; ratio between backends the same as sonnet) (20 prompts, 2 warmups, ignore eos, temperature 0, seed 0), MTP 5-token speculation (accept len 2.97, rate 39.5% on this arm), aggregate output tok/s, no explicit weight quantization flag (BF16 assumed); backend auto-selected as the SM90 default; issue reports it 36% slower at concurrency 1 and 70% at concurrency 8 vs the forced FLASH_ATTN_MLA_SPARSE, same MTP acceptance so the gap is the per-step attention kernel; MTP draft acceptance identical across the two backends; same ratio with --enable-expert-parallel and the Hopper recipe override; config=flashinfer-sm90-sparse-random",
   "pp_tokens": 300,
   "provenance": "community",
   "quant": "BF16",
   "quote": "GPU                          : 8x NVIDIA H100 80GB HBM3 (single node, TP=8); **Summary:** For `zai-org/GLM-5.3-Flash` on H100 the auto-selected sparse-MLA attention backend is `FLASHINFER_MLA_SPARSE_SM90`;   --tensor-parallel-size 8 \\;   --speculative-config '{\"method\":\"mtp\",\"num_speculative_tokens\":5}' \\; # \"random\" rows: --dataset-name random --random-input-len 300 --random-output-len 256;   --num-prompts 20 --num-warmups 2 --max-concurrency {1,8} --ignore-eos --disable-shuffle --temperature 0 --seed 0; | random | 1 | `FLASHINFER_MLA_SPARSE_SM90` (default) | 162.9 | 5.78 ms | 17.0 ms | 1570 ms | 2.97 / 39.5% |",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #56564 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/56564",
   "tg_tokens": 256,
   "tps": "162.9"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": 131072,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "Arc A770",
   "id": "lc-29526-a770-vk",
   "model": "Qwen3.5-9B",
   "notes": "llama-server 0.5.0-dev build 11160 (70c4e1582), ngl = 99, ctx 131072, fa on, q4_0 KV cache, np 2; decode tg from the 0-7h normal window of the long-run stability timeline (bug: empty EOS replies start after ~7h); mmproj BF16 loaded",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "- **GPU**: Intel Arc A770 16GB (DG2, PCI `8086:56a0` rev 08); - **Model**: Qwen3.5-9B (UD-Q4_K_XL), with multimodal mmproj (BF16); - **Backend**: Vulkan; - **llama.cpp**: `0.5.0-dev` (build 11160, commit `70c4e1582`), built with GNU 11.4.0 for Linux x86_64; -ngl 99 -c 131072 -fa on -ctk q4_0 -ctv q4_0 -t 4 -np 2 \\; | 0 \u2013 ~7h | Normal. `tg \u2248 26 t/s`, valid content. |",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #29526",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29526",
   "tps": "26",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 131072,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "Arc A770",
   "id": "lc-29527-a770-sycl",
   "model": "Qwen3.5-9B",
   "notes": "llama-server 0.4.1-dev build 11100 (7ab4ee7ba), oneAPI 2026.1, ngl = 99, ctx 131072 (reduced from 786432 on this build), fa on, q4_0 KV cache, np 2; decode tg from the 0-18h normal window of the long-run stability timeline (bug: hard decode fence deadlock after ~18h); companion of #29526 on the same host",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "- **GPU**: Intel Arc A770 16GB (DG2, PCI `8086:56a0` rev 08); - **Model**: Qwen3.5-9B (UD-Q4_K_XL), with multimodal mmproj (BF16); - **Backend**: SYCL / Level Zero; - **llama.cpp**: `0.4.1-dev` (build 11100, commit `7ab4ee7ba`), built with IntelLLVM 2026.1.1 for Linux x86_64; Start with a long context (note: SYCL Level Zero runtime has higher VRAM overhead than Vulkan, so `-c` had to be reduced from 786432 to 131072 to avoid `UR_RESULT_ERROR_OUT_OF_DEVICE_MEMORY`); -ngl 99 -c 131072 -fa on -ctk q4_0 -ctv q4_0 -t 4 -np 2 \\; | 0 \u2013 ~18h | Normal. `tg \u2248 21.19 t/s`, valid content, slots release normally. |",
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #29527",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29527",
   "tps": "21.19",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-05-26",
   "flags": [],
   "hardware": "RTX 6000 Pro (96GB)",
   "id": "hf1-dsv4flash-pro6000-q2",
   "model": "DeepSeek-V4-Flash",
   "notes": "Card's published figure, phrased as a lower bound ('can reach more than 35 tokens per second'). Custom fork Fringe210/llama.cpp-deepseek-v4-flash-cuda, WSL2 Ubuntu 24.04, CUDA 13.0, llama-cli -n 40960, Q2 GGUF.",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q2",
   "quote": "DS4 Test environment Windows, WSL2, Ubuntu 24.04, RTX 6000 Pro (96GB), CUDA 13.0; In this environment, inference can reach more than 35 tokens per second.; llama-cli -m huihui-ai/Huihui-DeepSeek-V4-Flash-BF16-abliterated-ds4-Q2.gguf -n 40960",
   "retrieved": "2026-09-27",
   "source_name": "HF model card: huihui-ai/Huihui-DeepSeek-V4-Flash-abliterated-ds4-GGUF",
   "source_url": "https://huggingface.co/huihui-ai/Huihui-DeepSeek-V4-Flash-abliterated-ds4-GGUF",
   "tg_tokens": null,
   "tps": "35"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-08-10",
   "flags": [],
   "hardware": "8\u00d7H100 (640 GB HBM)",
   "id": "hf1-kimik3-8xh100-iq1sx",
   "model": "Kimi-K3",
   "notes": "Card's published expectation ('Expect ~6.8 tok/s generation') on the author's 8xH100 + 2TB RAM node with mmap/page cache. MoE offloaded to CPU: -ngl 99 --n-cpu-moe 93; 34-shard IQ1_S-XS build.",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ1_S-XS",
   "quote": "8\u00d7H100 (640 GB HBM) cannot hold this alone and must be paired with large system RAM; our node is 8\u00d7H100 + 2TB RAM, where mmap keeps resident memory around 22 GB and leaves the rest in page cache; Expect ~6.8 tok/s generation and a ~4.5 min first (cold) load.; --model ./K3abl/Kimi-K3-abliterated-IQ1_S-XS-00001-of-00034.gguf \\ -ngl 99 --n-cpu-moe 93 -c 8192",
   "retrieved": "2026-09-27",
   "source_name": "HF model card: Ryanchen911/Kimi-K3-Uncensored-GGUF",
   "source_url": "https://huggingface.co/Ryanchen911/Kimi-K3-Uncensored-GGUF",
   "tg_tokens": null,
   "tps": "6.8"
  },
  {
   "backend": "LM Studio",
   "batch": null,
   "ctx": null,
   "date": "2026-09-17",
   "flags": [],
   "hardware": "5090",
   "id": "hf1-lfm2526b-5090-q6",
   "model": "LFM2.5-2.6B",
   "notes": "DavidAU quant-list entry (Q6, non-imatrix, 2 GB); the card does not name the harness, the author's other 5090 entries are LM Studio tests",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q6",
   "quote": "LFM2.5-2.6B-Qwen3.8-Turbo-Brilliance-Power-X12-NEO-MAX-GGUF; Q6, non imatrix, standard quant ; 380 t/s [5090], Quant size: 2 GB",
   "retrieved": "2026-09-27",
   "source_name": "HF model card: DavidAU LFM2.5-2.6B Q6 GGUF",
   "source_url": "https://huggingface.co/DavidAU/LFM2.5-2.6B-Qwen3.8-Turbo-Brilliance-Power-X12-NEO-MAX-GGUF",
   "tps": "380"
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-08-03",
   "flags": [],
   "hardware": "2\u00d7 RTX 3090 + Tesla V100",
   "id": "hf1-apertus-v15-70b-2x3090v100",
   "model": "Apertus v1.5 70B",
   "notes": "HF model card, 'What was actually tested' section; TEST quants, tensor split across three GPUs (2x RTX 3090 + Tesla V100, 64 GB VRAM total); ~17 tok/s generation and ~81 tok/s prompt at 65536 context with q8_0 KV cache; no imatrix, no perplexity measured",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "81",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Apertus v1.5 70B \u2014 text-only GGUF (TEST quants) GGUF conversions of swiss-ai/Apertus-v1.5-70B , made because there were no GGUF quants of v1.5 anywhere at the time of upload \u2014 the only Apertus GGUFs on the Hub are for the older 2509 (v1.0) release, and people who just want to try the model were stuck setting up the SwissAI vLLM fork. \u26a0\ufe0f Read this before downloading These are TEST quants. They were made to check whether the model answers at all , not as a polished release. No imatrix, no perplexity measurements, no benchmark runs. See What was actually tested . This is text-only. The vision and audio towers of v1.5 were stripped , not converted. You cannot feed images or audio to these files. See How this was made . Two launch flags are mandatory \u2014 --override-kv tokenizer.ggml.eos_token_id=int:68 (or the model never stops generating) and -sp (or reasoning appears to be missing). Neither is a quirk of these quants; both follow from the model itself. See Required flags for copy-paste commands. If someone produces proper imatrix quants of this model, prefer theirs over mine. Files File Size Notes apertus-70b-Q4_K_M.gguf 43.7 GB (40.7 GiB) The one that was actually run and tested apertus-70b-q8_0.gguf 75.0 GB (69.9 GiB) Converted directly from bf16; see testing notes Both were produced from the original bf16 weights, Q4_K_M by quantizing the q8_0 file. How this was made Apertus v1.5 is Apertus1p5ForConditionalGeneration \u2014 a multimodal model: a text decoder plus a VQ-VAE image tokenizer and a WavTokenizer audio tokenizer. llama.cpp does not support that architecture. What it does support is LLM_ARCH_APERTUS , added for Apertus 1.0. And the v1.5 decoder is architecturally the same as 1.0 \u2014 non-gated MLP with xIELU, mandatory q_norm/k_norm. So instead of teaching llama.cpp a new architecture, the model was reshaped into something the existing converter already understands: Both towers dropped \u2014 every tensor under model.vision_tokenizer. and model.audio_tokenizer. is excluded from the weight map. This is what makes the result text-only, and it is irreversible from these files. Vocabulary truncated from vocab_size 266752 down to output_vocab_size 131072. The tail of the embedding matrix holds image and audio codebook entries; the text model never emits them, and lm_head was already 131072 rows in the original. Multimodal entries were also removed from added_tokens in the tokenizer. Decoder prefix rewritten from model.language_model. to model. , which is where the Apertus 1.0 converter expects it. The result is an ordinary apertus -architecture GGUF. llama.cpp was not patched by a single byte \u2014 the conversion class is registered at runtime and then stock convert_hf_to_gguf.main() does the work. Both scripts are included in this repo: prep_apertus_text.py \u2014 builds a text-only HF folder. Weights are symlinked , not copied, so it costs ~20 MB of disk and a few seconds rather than a second copy of a 70B model. convert_apertus15.py \u2014 registers Apertus1p5ForConditionalGeneration and calls the stock converter. python3 prep_apertus_text.py ./Apertus-v1.5-70B ./Apertus-70B-text LLAMA_CPP_DIR=/path/to/llama.cpp python3 convert_apertus15.py ./Apertus-70B-text \\ --outfile apertus-70b-q8_0.gguf --outtype q8_0 Requirements No special build is needed. These load in stock upstream llama.cpp \u2014 the GGUF declares general.architecture = apertus , which has been supported since PR #15852 , released in b6671 (2 Oct 2025). Anything newer works; these files were run on a b10069 -era build. There is no vLLM path for these files \u2014 vLLM does not read GGUF for this architecture. If you want v1.5 in vLLM, you need the SwissAI fork and the original safetensors. Required flags \u2014 do not skip these Two flags are mandatory. Launch without them and the model looks broken when it is not: Flag Without it --override-kv tokenizer.ggml.eos_token_id=int:68 The model never stops . It runs past the end of its reply, answers its own question, starts a new turn, and only halts when it hits your token limit. -sp ( --special ) Reasoning appears to be missing . The model still thinks correctly, but the server strips the markers out of content , so you never see the block \u2014 and your frontend cannot split thinking from the answer. Why the EOS one is needed: the original config declares three EOS ids, eos_token_id: [2, 68, 72] . GGUF only stores one, and the converter writes 2 ( </s> ) \u2014 but the token that actually ends an assistant turn is 68 = <|assistant_end|> . So the correct stop token is present in the file, it is simply not the one marked as EOS. ( 72 is <|tools_suffix|> , for tool-calling mode.) -sp is available on both llama-cli and llama-server . Copy-paste: server llama-server \\ -m apertus-70b-Q4_K_M.gguf \\ --jinja -sp \\ --override-kv tokenizer.ggml.eos_token_id=int:68 \\ -ngl 999 --flash-attn on \\ -c 65536 -b 2048 -ub 512 \\ --cache-type-k q8_0 --cache-type-v q8_0 --cache-reuse 256 \\ --context-shift --parallel 1 \\ --host 127.0.0.1 --port 8084 Copy-paste: quick CLI check llama-cli \\ -m apertus-70b-Q4_K_M.gguf \\ -sp \\ --override-kv tokenizer.ggml.eos_token_id=int:68 \\ -ngl 999 -c 4096 \\ -p \"<|system_start|>You are a helpful assistant.<|system_end|><|user_start|>What is the capital of Switzerland?<|user_end|><|assistant_start|>\" That prompt is the literal format the model expects \u2014 <|system_start|> \u2026 <|system_end|> , <|user_start|> \u2026 <|user_end|> , then <|assistant_start|> to hand it the turn. It is exactly the call used to verify these files, and it answers The capital of Switzerland is Bern. To turn on reasoning, add the developer block after the system part \u2014 see Deliberation mode below. Deliberation (reasoning) mode Apertus v1.5 reasons between its own dedicated tokens, not <think> : id token meaning 32 <|inner_prefix|> start of reasoning 33 <|inner_suffix|> end of reasoning 68 <|assistant_end|> end of the whole reply Reasoning is enabled through a developer-role block, not a flag: <|developer_start|>Deliberation: enabled Tool Capabilities: disabled<|developer_end|> Quantization did not damage this. On a raw-token check of Q4_K_M , the model emitted token 32 at position 0, token 33 at position 131, and token 68 at position 262 \u2014 reasoning opened, closed, and the turn ended, exactly where they belong. But two things get in the way of seeing it: llama-server strips control tokens from content . Add -sp to get them through. llama.cpp's reasoning parser knows <think> and <|channel|> , not the Apertus markers, so /v1/chat/completions will not populate reasoning_content . Parse <|inner_prefix|> / <|inner_suffix|> on your side \u2014 or let your frontend do it (see below). SillyTavern Ready-made master-import presets are in this repo \u2014 import via Settings \u2192 Master Import : File System prompt sillytavern/Apertus-v1.5-EN.json English sillytavern/Apertus-v1.5-RU.json Russian The token layout is identical in both \u2014 only the system prompt language differs, so pick by the language you want the model instructed in, not by the language you intend to chat in (it follows the user either way). The system prompts are deliberately plain and general-purpose: swap in your own, the point of the presets is the plumbing below. No sampler settings are included, because none were tuned for this model \u2014 use your own. If you would rather set it up by hand, this is the same working configuration: Instruct Template System Sequence <|system_start|> System Suffix <|system_end|> Input Sequence <|user_start|> Input Suffix <|user_end|> Output Sequence <|assistant_start|> Output Suffix <|assistant_end|> Stop Sequence <|assistant_end|> Story String Prefix <|system_start|> Story String Suffix <|system_end|><|developer_start|>Deliberation: enabled Tool Capabilities: disabled<|developer_end|> Note where the deliberation block goes : on the Story String suffix, not the System suffix. Put it on System Suffix and it gets repeated after every system message instead of being stated once. Also enable System same as user , Sequences as stop strings , and set Names behavior to force ; leave wrap off. Reasoning settings \u2014 SillyTavern will then fold the thinking into a collapsible block by itself, no manual parsing: Prefix <|inner_prefix|> Suffix <|inner_suffix|> Separator \\n\\n What was actually tested Honest scope, on 2\u00d7RTX 3090 + Tesla V100 (64 GB VRAM total): Q4_K_M Loads and generates coherent multi-turn text (tested in Russian), ~17 tok/s generation, ~81 tok/s prompt at 65536 context with q8_0 KV cache across three GPUs.",
   "retrieved": "2026-09-27",
   "source_name": "HF model card: katya228/Apertus-v1.5-70B-text-GGUF",
   "source_url": "https://huggingface.co/katya228/Apertus-v1.5-70B-text-GGUF",
   "tg_tokens": null,
   "tps": "17",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": null,
   "date": "2026-02-13",
   "flags": [],
   "hardware": "M3 Max",
   "id": "hf1-minimax-m25-m3max",
   "model": "MiniMax M2.5",
   "notes": "HF model card quants table: Q3_K_L ~110 GB, 'Runs natively in RAM with high t/s (28 ON MAC M3 MAX)' on 128 GB Macs; card's perplexity-validation section cites 28.7 t/s for the same Q3_K_L quant; MoE 230B (8 active experts)",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q3_K_L",
   "quote": "File Name Method Size Use Case minimax-m2.5-Q4_K_M.gguf Q4_K_M 138 GB Highest logic preservation. Requires >128GB RAM or SSD swap. minimax-m2.5-Q3_K_L.gguf Q3_K_L ~110 GB Sweet spot for 128GB Macs . Runs natively in RAM with high t/s ( 28 ON MAC M3 MAX )",
   "retrieved": "2026-09-27",
   "source_name": "HF model card: ox-ox/MiniMax-M2.5-GGUF",
   "source_url": "https://huggingface.co/ox-ox/MiniMax-M2.5-GGUF",
   "tg_tokens": null,
   "tps": "28",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1024,
   "date": "2026-08",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "lc-27181-r9700-mmq-q2_k",
   "model": "Qwen3.8-27B",
   "notes": "identical weights: pure requant of one UD-Q6_K_XL source file; llama-bench -ngl 99 -fa on -b 2048 -ub 1024 -r 5; config=Q2_K",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "471.2",
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "AMD Radeon AI PRO R9700, gfx1201 (RDNA4, 40 CU, 32 GB); Source Qwen3.8-27B-UD-Q6_K_XL.gguf. llama-bench -ngl 99 -fa on -b 2048 -ub 1024 -r 5.; pp512 471.2 \u00b12.5",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27181 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27181",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1024,
   "date": "2026-08",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "lc-27181-r9700-mmq-q4_k",
   "model": "Qwen3.8-27B",
   "notes": "identical weights: pure requant of one UD-Q6_K_XL source file; llama-bench -ngl 99 -fa on -b 2048 -ub 1024 -r 5; config=Q4_K",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "1120.7",
   "provenance": "community",
   "quant": "Q4_K",
   "quote": "AMD Radeon AI PRO R9700, gfx1201 (RDNA4, 40 CU, 32 GB); Source Qwen3.8-27B-UD-Q6_K_XL.gguf. llama-bench -ngl 99 -fa on -b 2048 -ub 1024 -r 5.; 1120.7 \u00b119.2",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27181 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27181",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1024,
   "date": "2026-08",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "lc-27181-r9700-mmq-q8_0",
   "model": "Qwen3.8-27B",
   "notes": "identical weights: pure requant of one UD-Q6_K_XL source file; llama-bench -ngl 99 -fa on -b 2048 -ub 1024 -r 5; config=Q8_0",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "1178.3",
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "AMD Radeon AI PRO R9700, gfx1201 (RDNA4, 40 CU, 32 GB); Source Qwen3.8-27B-UD-Q6_K_XL.gguf. llama-bench -ngl 99 -fa on -b 2048 -ub 1024 -r 5.; 1178.3 \u00b115.5",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #27181 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/27181",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "Radeon RX 9070",
   "id": "lc-28135-rx9070-f16-scratch-off",
   "model": "Qwen3.8-27B",
   "notes": "hybrid-model (non-unified KV) f16-scratch fast path disabled; decode published as unchanged; config=fastpath-off",
   "params": null,
   "power_w": null,
   "pp_tokens": 21500,
   "pp_tps": "65",
   "provenance": "community",
   "quant": "as-published",
   "quote": "Measured impact (Qwen3.8-27B, 21.5K prompt); AMD Radeon RX 9070 (RDNA4), Windows; prefill 65 tok/s 119 tok/s (+83%)",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28135 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28135",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "Radeon RX 9070",
   "id": "lc-28135-rx9070-f16-scratch-on",
   "model": "Qwen3.8-27B",
   "notes": "hybrid-model (non-unified KV) f16-scratch fast path enabled; decode published as unchanged; config=fastpath-on",
   "params": null,
   "power_w": null,
   "pp_tokens": 21500,
   "pp_tps": "119",
   "provenance": "community",
   "quant": "as-published",
   "quote": "Measured impact (Qwen3.8-27B, 21.5K prompt); AMD Radeon RX 9070 (RDNA4), Windows; prefill 65 tok/s 119 tok/s (+83%)",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28135 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28135",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 8000,
   "date": "2026-09",
   "flags": [],
   "hardware": "2\u00d7 TU106",
   "id": "lc-28761-tu106-qwen-tile32-8k",
   "model": "Qwen3.8-27B",
   "notes": "stock sm_75 FA build (32-token Q-tile cap), prompt_per_second with FA on, KV q4_0, ubatch 2048, MTP; config=tile32",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "698",
   "provenance": "community",
   "quant": "IQ3_S",
   "quote": "head_dim=256 (Qwen3.8-27B IQ3_S, 2 \u00d7 TU106 tensor-split, ubatch 2048, KV q4_0, FA on, MTP); 8 K 698 tok/s 439 tok/s",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28761 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28761",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 8000,
   "date": "2026-09",
   "flags": [],
   "hardware": "2\u00d7 TU106",
   "id": "lc-28761-tu106-qwen-tile64-8k",
   "model": "Qwen3.8-27B",
   "notes": "forced tile 64 build (register-spill variant at head_dim=256), prompt_per_second with FA on, KV q4_0, ubatch 2048, MTP; config=tile64",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "439",
   "provenance": "community",
   "quant": "IQ3_S",
   "quote": "head_dim=256 (Qwen3.8-27B IQ3_S, 2 \u00d7 TU106 tensor-split, ubatch 2048, KV q4_0, FA on, MTP); 8 K 698 tok/s 439 tok/s",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28761 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28761",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 10000,
   "date": "2026-09",
   "flags": [],
   "hardware": "2\u00d7 TU106",
   "id": "lc-28761-tu106-qwen-tile32-10k",
   "model": "Qwen3.8-27B",
   "notes": "stock sm_75 FA build (32-token Q-tile cap), prompt_per_second with FA on, KV q4_0, ubatch 2048, MTP; config=tile32",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "657",
   "provenance": "community",
   "quant": "IQ3_S",
   "quote": "head_dim=256 (Qwen3.8-27B IQ3_S, 2 \u00d7 TU106 tensor-split, ubatch 2048, KV q4_0, FA on, MTP); 10 K 657 683",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28761 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28761",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 10000,
   "date": "2026-09",
   "flags": [],
   "hardware": "2\u00d7 TU106",
   "id": "lc-28761-tu106-qwen-tile64-10k",
   "model": "Qwen3.8-27B",
   "notes": "forced tile 64 build (register-spill variant at head_dim=256), prompt_per_second with FA on, KV q4_0, ubatch 2048, MTP; config=tile64",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "683",
   "provenance": "community",
   "quant": "IQ3_S",
   "quote": "head_dim=256 (Qwen3.8-27B IQ3_S, 2 \u00d7 TU106 tensor-split, ubatch 2048, KV q4_0, FA on, MTP); 10 K 657 683",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28761 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28761",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 16000,
   "date": "2026-09",
   "flags": [],
   "hardware": "2\u00d7 TU106",
   "id": "lc-28761-tu106-qwen-tile32-16k",
   "model": "Qwen3.8-27B",
   "notes": "stock sm_75 FA build (32-token Q-tile cap), prompt_per_second with FA on, KV q4_0, ubatch 2048, MTP; config=tile32",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "658",
   "provenance": "community",
   "quant": "IQ3_S",
   "quote": "head_dim=256 (Qwen3.8-27B IQ3_S, 2 \u00d7 TU106 tensor-split, ubatch 2048, KV q4_0, FA on, MTP); 16 K 658 664",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28761 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28761",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 16000,
   "date": "2026-09",
   "flags": [],
   "hardware": "2\u00d7 TU106",
   "id": "lc-28761-tu106-qwen-tile64-16k",
   "model": "Qwen3.8-27B",
   "notes": "forced tile 64 build (register-spill variant at head_dim=256), prompt_per_second with FA on, KV q4_0, ubatch 2048, MTP; config=tile64",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "664",
   "provenance": "community",
   "quant": "IQ3_S",
   "quote": "head_dim=256 (Qwen3.8-27B IQ3_S, 2 \u00d7 TU106 tensor-split, ubatch 2048, KV q4_0, FA on, MTP); 16 K 658 664",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28761 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28761",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 32000,
   "date": "2026-09",
   "flags": [],
   "hardware": "2\u00d7 TU106",
   "id": "lc-28761-tu106-qwen-tile32-32k",
   "model": "Qwen3.8-27B",
   "notes": "stock sm_75 FA build (32-token Q-tile cap), prompt_per_second with FA on, KV q4_0, ubatch 2048, MTP; config=tile32",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "544",
   "provenance": "community",
   "quant": "IQ3_S",
   "quote": "head_dim=256 (Qwen3.8-27B IQ3_S, 2 \u00d7 TU106 tensor-split, ubatch 2048, KV q4_0, FA on, MTP); 32 K 544 650",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28761 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28761",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 32000,
   "date": "2026-09",
   "flags": [],
   "hardware": "2\u00d7 TU106",
   "id": "lc-28761-tu106-qwen-tile64-32k",
   "model": "Qwen3.8-27B",
   "notes": "forced tile 64 build (register-spill variant at head_dim=256), prompt_per_second with FA on, KV q4_0, ubatch 2048, MTP; config=tile64",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "650",
   "provenance": "community",
   "quant": "IQ3_S",
   "quote": "head_dim=256 (Qwen3.8-27B IQ3_S, 2 \u00d7 TU106 tensor-split, ubatch 2048, KV q4_0, FA on, MTP); 32 K 544 650",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28761 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28761",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 230000,
   "date": "2026-09",
   "flags": [],
   "hardware": "2\u00d7 TU106",
   "id": "lc-28761-tu106-qwen-tile32-230k",
   "model": "Qwen3.8-27B",
   "notes": "stock sm_75 FA build (32-token Q-tile cap), prompt_per_second with FA on, KV q4_0, ubatch 2048, MTP; config=tile32",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "255",
   "provenance": "community",
   "quant": "IQ3_S",
   "quote": "head_dim=256 (Qwen3.8-27B IQ3_S, 2 \u00d7 TU106 tensor-split, ubatch 2048, KV q4_0, FA on, MTP); 230 K 255 384",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28761 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28761",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 230000,
   "date": "2026-09",
   "flags": [],
   "hardware": "2\u00d7 TU106",
   "id": "lc-28761-tu106-qwen-tile64-230k",
   "model": "Qwen3.8-27B",
   "notes": "forced tile 64 build (register-spill variant at head_dim=256), prompt_per_second with FA on, KV q4_0, ubatch 2048, MTP; config=tile64",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "384",
   "provenance": "community",
   "quant": "IQ3_S",
   "quote": "head_dim=256 (Qwen3.8-27B IQ3_S, 2 \u00d7 TU106 tensor-split, ubatch 2048, KV q4_0, FA on, MTP); 230 K 255 384",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28761 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28761",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 8000,
   "date": "2026-09",
   "flags": [],
   "hardware": "2\u00d7 TU106",
   "id": "lc-28761-tu106-qwen-dynamic-8k",
   "model": "Qwen3.8-27B",
   "notes": "dynamic tile build (tile64 iff KV > 8192), prompt_per_second, KV q4_0, ubatch 2048, MTP; config=dynamic",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "699",
   "provenance": "community",
   "quant": "IQ3_S",
   "quote": "head_dim=256 (Qwen3.8-27B IQ3_S, 2 \u00d7 TU106 tensor-split, ubatch 2048, KV q4_0, FA on, MTP); 8 K 698 439 699 (= tile 32)",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28761 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28761",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 16000,
   "date": "2026-09",
   "flags": [],
   "hardware": "2\u00d7 TU106",
   "id": "lc-28761-tu106-qwen-dynamic-16k",
   "model": "Qwen3.8-27B",
   "notes": "dynamic tile build (tile64 iff KV > 8192), prompt_per_second, KV q4_0, ubatch 2048, MTP; config=dynamic",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "664",
   "provenance": "community",
   "quant": "IQ3_S",
   "quote": "head_dim=256 (Qwen3.8-27B IQ3_S, 2 \u00d7 TU106 tensor-split, ubatch 2048, KV q4_0, FA on, MTP); 16 K 658 664 664 (= tile 64)",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28761 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28761",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 32000,
   "date": "2026-09",
   "flags": [],
   "hardware": "2\u00d7 TU106",
   "id": "lc-28761-tu106-qwen-dynamic-32k",
   "model": "Qwen3.8-27B",
   "notes": "dynamic tile build (tile64 iff KV > 8192), prompt_per_second, KV q4_0, ubatch 2048, MTP; config=dynamic",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "650",
   "provenance": "community",
   "quant": "IQ3_S",
   "quote": "head_dim=256 (Qwen3.8-27B IQ3_S, 2 \u00d7 TU106 tensor-split, ubatch 2048, KV q4_0, FA on, MTP); 32 K 544 650 650 (= tile 64)",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28761 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28761",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 230000,
   "date": "2026-09",
   "flags": [],
   "hardware": "2\u00d7 TU106",
   "id": "lc-28761-tu106-qwen-dynamic-230k",
   "model": "Qwen3.8-27B",
   "notes": "dynamic tile build (tile64 iff KV > 8192), prompt_per_second, KV q4_0, ubatch 2048, MTP; config=dynamic",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "387",
   "provenance": "community",
   "quant": "IQ3_S",
   "quote": "head_dim=256 (Qwen3.8-27B IQ3_S, 2 \u00d7 TU106 tensor-split, ubatch 2048, KV q4_0, FA on, MTP); 230 K 255 384 387 (= tile 64)",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28761 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28761",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 2,
   "ctx": 230000,
   "date": "2026-09",
   "flags": [],
   "hardware": "2\u00d7 TU106",
   "id": "lc-28761-tu106-qwen-dynamic-230k-concurrent",
   "model": "Qwen3.8-27B",
   "notes": "dynamic build, two concurrent slots (8 K + 230 K) on the 2 \u00d7 TU106 split; config=dynamic",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "385",
   "provenance": "community",
   "quant": "IQ3_S",
   "quote": "head_dim=256 (Qwen3.8-27B IQ3_S, 2 \u00d7 TU106 tensor-split, ubatch 2048, KV q4_0, FA on, MTP); the 230 K slot runs at its full solo speed (385 tok/s) while the 8 K slot uses tile32",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28761 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28761",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 8000,
   "date": "2026-09",
   "flags": [],
   "hardware": "TU106",
   "id": "lc-28761-tu106-llama-tile32-8k",
   "model": "Llama 3.1 8B",
   "notes": "single TU106, prompt_per_second, KV 8 K, tile32 build; config=tile32",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "2052",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "head_dim=128 (Llama-3.1-8B Q4_0, single TU106, GQA 4); 8 K 2052 tok/s 2127 tok/s",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28761 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28761",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 8000,
   "date": "2026-09",
   "flags": [],
   "hardware": "TU106",
   "id": "lc-28761-tu106-llama-tile64-8k",
   "model": "Llama 3.1 8B",
   "notes": "single TU106, prompt_per_second, KV 8 K, tile64 build; config=tile64",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "2127",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "head_dim=128 (Llama-3.1-8B Q4_0, single TU106, GQA 4); 8 K 2052 tok/s 2127 tok/s",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28761 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28761",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": 2048,
   "date": "2026-09",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-28828-7800xt-b10727-2048",
   "model": "Qwen3.8-27B",
   "notes": "official b10727 prebuilt, prompt_per_second (llama-bench -p 2048, -n 0), -ngl 99 -ctk q4_0 -ctv q4_0 -fa 1; config=b10727",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "468.35",
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "AMD Radeon RX 7800 XT (16 GB, gfx1101, RDNA3, Wave32); Qwen3.8-27B-UD IQ4_XS (13.26 GiB, 4.25 bpw), same family Q3_K_XL for control; 2048 468.35 502.91",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28828 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28828",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": 2048,
   "date": "2026-09",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-28828-7800xt-b10909-2048",
   "model": "Qwen3.8-27B",
   "notes": "self-built b10909, prompt_per_second (llama-bench -p 2048, -n 0), -ngl 99 -ctk q4_0 -ctv q4_0 -fa 1; config=b10909",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "502.91",
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "AMD Radeon RX 7800 XT (16 GB, gfx1101, RDNA3, Wave32); Qwen3.8-27B-UD IQ4_XS (13.26 GiB, 4.25 bpw), same family Q3_K_XL for control; 2048 468.35 502.91",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28828 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28828",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": 8192,
   "date": "2026-09",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-28828-7800xt-b10727-8192",
   "model": "Qwen3.8-27B",
   "notes": "official b10727 prebuilt, prompt_per_second (llama-bench -p 8192, -n 0), -ngl 99 -ctk q4_0 -ctv q4_0 -fa 1; config=b10727",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "535.11",
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "AMD Radeon RX 7800 XT (16 GB, gfx1101, RDNA3, Wave32); Qwen3.8-27B-UD IQ4_XS (13.26 GiB, 4.25 bpw), same family Q3_K_XL for control; 8192 535.11 251.74",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28828 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28828",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": 8192,
   "date": "2026-09",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-28828-7800xt-b10909-8192",
   "model": "Qwen3.8-27B",
   "notes": "self-built b10909, prompt_per_second (llama-bench -p 8192, -n 0), -ngl 99 -ctk q4_0 -ctv q4_0 -fa 1; config=b10909; single sample, source flags it as a likely artifact",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "251.74",
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "AMD Radeon RX 7800 XT (16 GB, gfx1101, RDNA3, Wave32); Qwen3.8-27B-UD IQ4_XS (13.26 GiB, 4.25 bpw), same family Q3_K_XL for control; 8192 535.11 251.74",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28828 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28828",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": 16384,
   "date": "2026-09",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-28828-7800xt-b10727-16384",
   "model": "Qwen3.8-27B",
   "notes": "official b10727 prebuilt, prompt_per_second (llama-bench -p 16384, -n 0), -ngl 99 -ctk q4_0 -ctv q4_0 -fa 1; config=b10727",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "458.78",
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "AMD Radeon RX 7800 XT (16 GB, gfx1101, RDNA3, Wave32); Qwen3.8-27B-UD IQ4_XS (13.26 GiB, 4.25 bpw), same family Q3_K_XL for control; 16384 458.78 499.40",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28828 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28828",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": 16384,
   "date": "2026-09",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-28828-7800xt-b10909-16384",
   "model": "Qwen3.8-27B",
   "notes": "self-built b10909, prompt_per_second (llama-bench -p 16384, -n 0), -ngl 99 -ctk q4_0 -ctv q4_0 -fa 1; config=b10909",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "499.40",
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "AMD Radeon RX 7800 XT (16 GB, gfx1101, RDNA3, Wave32); Qwen3.8-27B-UD IQ4_XS (13.26 GiB, 4.25 bpw), same family Q3_K_XL for control; 16384 458.78 499.40",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28828 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28828",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": 28672,
   "date": "2026-09",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-28828-7800xt-b10727-28672",
   "model": "Qwen3.8-27B",
   "notes": "official b10727 prebuilt, prompt_per_second (llama-bench -p 28672, -n 0), -ngl 99 -ctk q4_0 -ctv q4_0 -fa 1; config=b10727",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "425.16",
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "AMD Radeon RX 7800 XT (16 GB, gfx1101, RDNA3, Wave32); Qwen3.8-27B-UD IQ4_XS (13.26 GiB, 4.25 bpw), same family Q3_K_XL for control; 28672 425.16 495.57 481.75",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28828 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28828",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": 28672,
   "date": "2026-09",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-28828-7800xt-b10909-28672",
   "model": "Qwen3.8-27B",
   "notes": "self-built b10909, prompt_per_second (llama-bench -p 28672, -n 0), -ngl 99 -ctk q4_0 -ctv q4_0 -fa 1; config=b10909",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "495.57",
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "AMD Radeon RX 7800 XT (16 GB, gfx1101, RDNA3, Wave32); Qwen3.8-27B-UD IQ4_XS (13.26 GiB, 4.25 bpw), same family Q3_K_XL for control; 28672 425.16 495.57 481.75",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28828 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28828",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": 31744,
   "date": "2026-09",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-28828-7800xt-b10727-31744",
   "model": "Qwen3.8-27B",
   "notes": "official b10727 prebuilt, prompt_per_second (llama-bench -p 31744, -n 0), -ngl 99 -ctk q4_0 -ctv q4_0 -fa 1; config=b10727",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "417.64",
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "AMD Radeon RX 7800 XT (16 GB, gfx1101, RDNA3, Wave32); Qwen3.8-27B-UD IQ4_XS (13.26 GiB, 4.25 bpw), same family Q3_K_XL for control; 31744 417.64 488.03 478.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28828 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28828",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": 31744,
   "date": "2026-09",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-28828-7800xt-b10909-31744",
   "model": "Qwen3.8-27B",
   "notes": "self-built b10909, prompt_per_second (llama-bench -p 31744, -n 0), -ngl 99 -ctk q4_0 -ctv q4_0 -fa 1; config=b10909",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "488.03",
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "AMD Radeon RX 7800 XT (16 GB, gfx1101, RDNA3, Wave32); Qwen3.8-27B-UD IQ4_XS (13.26 GiB, 4.25 bpw), same family Q3_K_XL for control; 31744 417.64 488.03 478.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28828 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28828",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": 32768,
   "date": "2026-09",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-28828-7800xt-b10727-32768",
   "model": "Qwen3.8-27B",
   "notes": "official b10727 prebuilt, prompt_per_second (llama-bench -p 32768, -n 0), -ngl 99 -ctk q4_0 -ctv q4_0 -fa 1; config=b10727",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "412.17",
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "AMD Radeon RX 7800 XT (16 GB, gfx1101, RDNA3, Wave32); Qwen3.8-27B-UD IQ4_XS (13.26 GiB, 4.25 bpw), same family Q3_K_XL for control; 32768 412.17 199.17 480.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28828 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28828",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": 32768,
   "date": "2026-09",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-28828-7800xt-b10909-32768",
   "model": "Qwen3.8-27B",
   "notes": "self-built b10909, prompt_per_second (llama-bench -p 32768, -n 0), -ngl 99 -ctk q4_0 -ctv q4_0 -fa 1; config=b10909",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "199.17",
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "AMD Radeon RX 7800 XT (16 GB, gfx1101, RDNA3, Wave32); Qwen3.8-27B-UD IQ4_XS (13.26 GiB, 4.25 bpw), same family Q3_K_XL for control; 32768 412.17 199.17 480.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28828 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28828",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": 33792,
   "date": "2026-09",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-28828-7800xt-b10727-33792",
   "model": "Qwen3.8-27B",
   "notes": "official b10727 prebuilt, prompt_per_second (llama-bench -p 33792, -n 0), -ngl 99 -ctk q4_0 -ctv q4_0 -fa 1; config=b10727",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "168.49",
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "AMD Radeon RX 7800 XT (16 GB, gfx1101, RDNA3, Wave32); Qwen3.8-27B-UD IQ4_XS (13.26 GiB, 4.25 bpw), same family Q3_K_XL for control; 33792 168.49 202.21 474.16",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28828 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28828",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": 33792,
   "date": "2026-09",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-28828-7800xt-b10909-33792",
   "model": "Qwen3.8-27B",
   "notes": "self-built b10909, prompt_per_second (llama-bench -p 33792, -n 0), -ngl 99 -ctk q4_0 -ctv q4_0 -fa 1; config=b10909",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "202.21",
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "AMD Radeon RX 7800 XT (16 GB, gfx1101, RDNA3, Wave32); Qwen3.8-27B-UD IQ4_XS (13.26 GiB, 4.25 bpw), same family Q3_K_XL for control; 33792 168.49 202.21 474.16",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28828 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28828",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": 28672,
   "date": "2026-09",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-28828-7800xt-b10909-q3kxl-28672",
   "model": "Qwen3.8-27B",
   "notes": "Q3_K_XL control, self-built b10909, prompt_per_second (llama-bench -p 28672, -n 0); flat across the cliff region; config=b10909",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "481.75",
   "provenance": "community",
   "quant": "Q3_K_XL",
   "quote": "AMD Radeon RX 7800 XT (16 GB, gfx1101, RDNA3, Wave32); Qwen3.8-27B-UD IQ4_XS (13.26 GiB, 4.25 bpw), same family Q3_K_XL for control; 28672 425.16 495.57 481.75",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28828 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28828",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": 31744,
   "date": "2026-09",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-28828-7800xt-b10909-q3kxl-31744",
   "model": "Qwen3.8-27B",
   "notes": "Q3_K_XL control, self-built b10909, prompt_per_second (llama-bench -p 31744, -n 0); flat across the cliff region; config=b10909",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "478.00",
   "provenance": "community",
   "quant": "Q3_K_XL",
   "quote": "AMD Radeon RX 7800 XT (16 GB, gfx1101, RDNA3, Wave32); Qwen3.8-27B-UD IQ4_XS (13.26 GiB, 4.25 bpw), same family Q3_K_XL for control; 31744 417.64 488.03 478.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28828 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28828",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": 32768,
   "date": "2026-09",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-28828-7800xt-b10909-q3kxl-32768",
   "model": "Qwen3.8-27B",
   "notes": "Q3_K_XL control, self-built b10909, prompt_per_second (llama-bench -p 32768, -n 0); flat across the cliff region; config=b10909",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "480.00",
   "provenance": "community",
   "quant": "Q3_K_XL",
   "quote": "AMD Radeon RX 7800 XT (16 GB, gfx1101, RDNA3, Wave32); Qwen3.8-27B-UD IQ4_XS (13.26 GiB, 4.25 bpw), same family Q3_K_XL for control; 32768 412.17 199.17 480.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28828 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28828",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": 33792,
   "date": "2026-09",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-28828-7800xt-b10909-q3kxl-33792",
   "model": "Qwen3.8-27B",
   "notes": "Q3_K_XL control, self-built b10909, prompt_per_second (llama-bench -p 33792, -n 0); flat across the cliff region; config=b10909",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "474.16",
   "provenance": "community",
   "quant": "Q3_K_XL",
   "quote": "AMD Radeon RX 7800 XT (16 GB, gfx1101, RDNA3, Wave32); Qwen3.8-27B-UD IQ4_XS (13.26 GiB, 4.25 bpw), same family Q3_K_XL for control; 33792 168.49 202.21 474.16",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28828 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28828",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": 32100,
   "date": "2026-09",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-28828-7800xt-b10909-mtp-32100",
   "model": "Qwen3.8-27B",
   "notes": "llama-server, MTP spec decode (--spec-type draft-mtp), 32 100-token prompt inside the collapsed region; config=draft-mtp",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "93.7",
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "AMD Radeon RX 7800 XT (16 GB, gfx1101, RDNA3, Wave32); Qwen3.8-27B-UD IQ4_XS (13.26 GiB, 4.25 bpw), same family Q3_K_XL for control; Server measurement (self-built b10909, IQ4_XS, 32 100-tok prompt, --spec-type draft-mtp): prefill 93.7 t/s, TTFT 342.7 s. Same binary, Q3_K_XL: 461.8 t/s.",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28828 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28828",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": "342.7"
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": 32100,
   "date": "2026-09",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-28828-7800xt-b10909-mtp-q3kxl-32100",
   "model": "Qwen3.8-27B",
   "notes": "llama-server, MTP spec decode, 32 100-token prompt, same binary and conditions as the IQ4_XS row; config=draft-mtp",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "461.8",
   "provenance": "community",
   "quant": "Q3_K_XL",
   "quote": "AMD Radeon RX 7800 XT (16 GB, gfx1101, RDNA3, Wave32); Qwen3.8-27B-UD IQ4_XS (13.26 GiB, 4.25 bpw), same family Q3_K_XL for control; Server measurement (self-built b10909, IQ4_XS, 32 100-tok prompt, --spec-type draft-mtp): prefill 93.7 t/s, TTFT 342.7 s. Same binary, Q3_K_XL: 461.8 t/s.",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #28828 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/28828",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "R9700",
   "id": "lc-29536-r9700-stock",
   "model": "Qwen3.8-27B",
   "notes": "stock LLVM 23 ROCm/HIP build, MMQ VGPR spilling before the fix; figures published with ~ prefix; config=llvm23-stock",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "298",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Prefill on the R9700 goes from about 298 to about 1250 t/s on Qwen3.8-27B Q4_K_M with stock LLVM 23, the same as with AMD's compiler",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #29536 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29536",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "R9700",
   "id": "lc-29536-r9700-nospill",
   "model": "Qwen3.8-27B",
   "notes": "stock LLVM 23 ROCm/HIP build, MMQ VGPR spilling after the fix (no spill); figures published with ~ prefix; config=llvm23-nospill",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "1250",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Prefill on the R9700 goes from about 298 to about 1250 t/s on Qwen3.8-27B Q4_K_M with stock LLVM 23, the same as with AMD's compiler",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp issue #29536 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29536",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": null,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "2 nodes \u00d7 8\u00d7 H100 80GB",
   "id": "vllm-58920-2nodes-h100-none-c8",
   "model": "GLM 5.2",
   "notes": "2-node 16x H100 80GB, pipeline parallel 8 x tensor parallel 2, vLLM 0.30.0 official image, Model Runner V2 (default on 0.30) with async scheduling, 12k-token prompts, 1024 forced output tokens, distinct prompts per stream; no-KV-connector arm; decode TPOT table 18.6 / 23.4 / 32.1 ms at c1 / c4 / c8; V1 model runner shows no overhead (18.3 vs 17.0 ms at c1); ~250 is the published approximate aggregate at 8 running requests; config=connector-none",
   "pp_tokens": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "2 nodes \u00d7 8\u00d7 H100 80GB, pipeline parallel 8 \u00d7 tensor parallel 2 (PP0\u20133 on node A, PP4\u20137 on node B); Model: GLM-5.2 FP8 (`GlmMoeDsaForCausalLM`), `--max-num-seqs 8`; Throughput at 8 running requests drops from ~250 tok/s to ~133 tok/s.",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #58920 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58920",
   "tg_tokens": 1024,
   "tps": "250"
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": null,
   "date": "2026-09-27",
   "flags": [],
   "hardware": "2 nodes \u00d7 8\u00d7 H100 80GB",
   "id": "vllm-58920-2nodes-h100-noop-c8",
   "model": "GLM 5.2",
   "notes": "no-op KV connector (empty metadata, no loads or stores) measures the same as the team's real connector; decode TPOT 27.1 / 34.7 / 60.0 ms at c1 / c4 / c8 (+87% at c8 vs the no-connector arm); throughput ~133 t/s at 8 running; root cause per the report: all 16 PP x TP ranks reply to every execute_model / sample_tokens RPC once a KV aggregator exists (output_rank None), and the reply-ring writer spins holding the GIL, starving prepare_inputs on the worker main thread; config=connector-noop",
   "pp_tokens": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "2 nodes \u00d7 8\u00d7 H100 80GB, pipeline parallel 8 \u00d7 tensor parallel 2 (PP0\u20133 on node A, PP4\u20137 on node B); Model: GLM-5.2 FP8 (`GlmMoeDsaForCausalLM`), `--max-num-seqs 8`; Throughput at 8 running requests drops from ~250 tok/s to ~133 tok/s.",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #58920 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58920",
   "tg_tokens": 1024,
   "tps": "133"
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-18",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "hf1-katcoder25-5090-apex",
   "model": "KAT-Coder-V2.5",
   "notes": "IsValorum APEX-I-MiniPlus V2.1 GGUF card (KAT-Coder-V2.5-Dev, 256K hybrid linear-attention model): single-stream generation approx. 247-251 t/s on RTX 5090 (stored lower bound 247), prompt prefill 2800-3900+ t/s as published (stored lower bound 2800); card claims 65536 ctx runs 100% in 16 GB VRAM, empirically verified in Unsloth Studio & llama.cpp with full GPU offload (-ngl 99), RTX 5090 figure attributed to the zephel01 Occamy V2 reference",
   "pp_tps": 2800,
   "provenance": "community",
   "quant": "APEX-I-MiniPlus V2.1",
   "quote": "IsValorum/KAT-Coder-V2.5-Dev-APEX-I-MiniPlus-V2.1-GGUF; Measured Real-World Throughput: Sustained single-stream generation of approx. 247 \u2013 251 tok/s on NVIDIA RTX 5090.; NVIDIA RTX 5080 / 5090 (Blackwell) Full GPU ( -ngl 99 ) approx. 247 \u2013 251 tok/s 2,800 \u2013 3,900+ tok/s Empirically verified on RTX 5090 by zephel01 (Occamy V2 Reference)",
   "retrieved": "2026-09-27",
   "source_name": "HF model card: IsValorum/KAT-Coder-V2.5-Dev-APEX-I-MiniPlus-V2.1-GGUF",
   "source_url": "https://huggingface.co/IsValorum/KAT-Coder-V2.5-Dev-APEX-I-MiniPlus-V2.1-GGUF",
   "tps": "247"
  },
  {
   "backend": "vLLM",
   "batch": 64,
   "ctx": null,
   "date": "2026-09-25",
   "flags": [],
   "hardware": "H100",
   "id": "vllm-58804-h100-fs-tiering-ab35354",
   "model": "gpt-oss 120B",
   "notes": "Guidellm synthetic multi-turn: 20K initial prefix, 4096 tokens per turn, 5 turns, concurrency 64, 700 s; OffloadingConnector with TieringOffloadingSpec, 24 GB CPU tier plus NVMe FS secondary tier, block_size 1024, prefix caching enabled; tensor parallel 2 on a 4x H100 80GB node; baseline commit ab35354 (direct parent of the bisected commit); Total tokens/s is the Guidellm aggregate; config=fs-tiering-ab35354",
   "pp_tokens": 20000,
   "provenance": "community",
   "quant": "as-published",
   "quote": "Machine : H100; vllm serve openai/gpt-oss-120b; --tensor-parallel-size=2; \"spec_name\": \"TieringOffloadingSpec\"; \"cpu_bytes_to_use\": 25769803776; \"block_size\": 1024; Multi-turn conversations with initial prompt of size 20K, subsequent prompts 4K and 5 Turns. Run at concurrency 64 for 700s; BENCH_PREFIX_TOKENS=20000; BENCH_PROMPT_TOKENS=\"4096\"; commit `ab35354` (before f12fe1) : Total tokens / s = 119092; A git bisect, led me to the commit https://github.com/vllm-project/vllm/pull/51787",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #58804 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58804",
   "tg_tokens": null,
   "tps": "119092"
  },
  {
   "backend": "vLLM",
   "batch": 64,
   "ctx": null,
   "date": "2026-09-25",
   "flags": [],
   "hardware": "H100",
   "id": "vllm-58804-h100-fs-tiering-f12fe1",
   "model": "gpt-oss 120B",
   "notes": "Guidellm synthetic multi-turn: 20K initial prefix, 4096 tokens per turn, 5 turns, concurrency 64, 700 s; OffloadingConnector with TieringOffloadingSpec, 24 GB CPU tier plus NVMe FS secondary tier, block_size 1024, prefix caching enabled; tensor parallel 2 on a 4x H100 80GB node; regressed commit f12fe1 (lands PR #51787, the bisect hit); Total tokens/s is the Guidellm aggregate; config=fs-tiering-f12fe1",
   "pp_tokens": 20000,
   "provenance": "community",
   "quant": "as-published",
   "quote": "Machine : H100; vllm serve openai/gpt-oss-120b; --tensor-parallel-size=2; \"spec_name\": \"TieringOffloadingSpec\"; \"cpu_bytes_to_use\": 25769803776; \"block_size\": 1024; Multi-turn conversations with initial prompt of size 20K, subsequent prompts 4K and 5 Turns. Run at concurrency 64 for 700s; BENCH_PREFIX_TOKENS=20000; BENCH_PROMPT_TOKENS=\"4096\"; commit `f12fe1` : Total tokens / s = 102876; A git bisect, led me to the commit https://github.com/vllm-project/vllm/pull/51787",
   "retrieved": "2026-09-27",
   "source_name": "vLLM issue #58804 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58804",
   "tg_tokens": null,
   "tps": "102876"
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": null,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-58944-mi355x-addmm-c16",
   "model": "Kimi-K3",
   "notes": "ROCm Kimi-K3 TP8 on MI355X, vLLM with VLLM_ROCM_USE_AITER=1 and AITER_SITUV2_A8W4=1 (a8w4 routed experts), max_num_seqs 128, prefix caching, Kimi-K3 tool/reasoning parsers; vllm bench serve random 1k in / 1k out, aggregate output tok/s; A/B of the sharded latent-MoE tail addmm_ (hipBLASLt Cijk 8.57 us/layer) vs aiter gemm_a16w16_atomic accumulate=True in-place add (4.35 us/layer); gsm8k 5-shot exact match 0.9636 with the tuned kernel; no explicit weight quant flag (BF16 assumed); kernel-level before/after table in the Purpose section; config=addmm",
   "pp_tokens": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "On MI355X this is hipBLASLt `Cijk_..._MT16x16x512`, 8.57 us per layer at decode; vllm serve /data/Kimi-K3; --dataset-name random --random-input-len 1024 --random-output-len 1024; Maximum request concurrency:             16; Output token throughput (tok/s):         658.42",
   "retrieved": "2026-09-28",
   "source_name": "vLLM issue #58944 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58944",
   "tg_tokens": 1024,
   "tps": "658.42"
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": null,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-58944-mi355x-atomic-c16",
   "model": "Kimi-K3",
   "notes": "ROCm Kimi-K3 TP8 on MI355X, vLLM with VLLM_ROCM_USE_AITER=1 and AITER_SITUV2_A8W4=1 (a8w4 routed experts), max_num_seqs 128, prefix caching, Kimi-K3 tool/reasoning parsers; vllm bench serve random 1k in / 1k out, aggregate output tok/s; A/B of the sharded latent-MoE tail addmm_ (hipBLASLt Cijk 8.57 us/layer) vs aiter gemm_a16w16_atomic accumulate=True in-place add (4.35 us/layer); gsm8k 5-shot exact match 0.9636 with the tuned kernel; no explicit weight quant flag (BF16 assumed); kernel-level before/after table in the Purpose section; config=atomic",
   "pp_tokens": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "On MI355X this is hipBLASLt `Cijk_..._MT16x16x512`, 8.57 us per layer at decode; vllm serve /data/Kimi-K3; --dataset-name random --random-input-len 1024 --random-output-len 1024; Maximum request concurrency:             16; Output token throughput (tok/s):         668.97",
   "retrieved": "2026-09-28",
   "source_name": "vLLM issue #58944 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58944",
   "tg_tokens": 1024,
   "tps": "668.97"
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": 4096,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "RTX PRO 5000 72GB",
   "id": "vllm-58989-rtxpro5000-deepgemm",
   "model": "Qwen3-30B-A3B-Instruct-2507-FP8",
   "notes": "2x RTX PRO 5000 72GB Blackwell (SM 12.0) rig, TP1, driver 595.80, torch 2.13.0+cu130; vllm bench throughput random 1024 in / 256 out, 256 prompts, max model len 4096; block-FP8 MoE backend selection A/B: DEEPGEMM auto default vs TRITON fused MoE (PR #58989); real weights, bs1 latency 0.800 s, bs32 latency 3.739 s, GSM8K 5-shot 88.02%; config=deepgemm",
   "pp_tokens": 1024,
   "provenance": "community",
   "quant": "FP8",
   "quote": "2x RTX PRO 5000 72GB Blackwell (SM 12.0), driver 595.80, torch 2.13.0+cu130; Real weights, Qwen3-30B-A3B-Instruct-2507-FP8, TP=1; DeepGEMM (current) 0.800 3.739 19778 88.02%; --dataset-name random --input-len 1024 --output-len 256 --num-prompts 256; --max-model-len 4096",
   "retrieved": "2026-09-28",
   "source_name": "vLLM issue #58989 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58989",
   "tg_tokens": 256,
   "tps": "19778"
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": 4096,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "RTX PRO 5000 72GB",
   "id": "vllm-58989-rtxpro5000-triton",
   "model": "Qwen3-30B-A3B-Instruct-2507-FP8",
   "notes": "2x RTX PRO 5000 72GB Blackwell (SM 12.0) rig, TP1, driver 595.80, torch 2.13.0+cu130; vllm bench throughput random 1024 in / 256 out, 256 prompts, max model len 4096; block-FP8 MoE backend selection A/B: DEEPGEMM auto default vs TRITON fused MoE (PR #58989, the selected default after the PR); real weights, bs1 latency 0.678 s, bs32 latency 3.250 s, GSM8K 5-shot 88.55%; config=triton",
   "pp_tokens": 1024,
   "provenance": "community",
   "quant": "FP8",
   "quote": "2x RTX PRO 5000 72GB Blackwell (SM 12.0), driver 595.80, torch 2.13.0+cu130; Real weights, Qwen3-30B-A3B-Instruct-2507-FP8, TP=1; Triton (this PR) 0.678 3.250 19806 88.55%; --dataset-name random --input-len 1024 --output-len 256 --num-prompts 256; --max-model-len 4096",
   "retrieved": "2026-09-28",
   "source_name": "vLLM issue #58989 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58989",
   "tg_tokens": 256,
   "tps": "19806"
  },
  {
   "backend": "llama.cpp",
   "ctx": null,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "M5 Pro (20-core GPU)",
   "id": "lc-29570-m5-pp512-master",
   "model": "Qwen3.8-27B",
   "notes": "M5 Pro (20-core GPU), Apple silicon; llama-bench Qwen3.8-27B Q8_0 with -fa 1, master vs the tensor-API FLASH_ATTN_EXT kernel (32 queries per threadgroup, 64-row KV blocks, matmul2d); test pp512; config=master",
   "pp_tps": "383.3",
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "llama-bench, Qwen3.8-27B Q8_0, -fa 1; M5 Pro (20-core GPU), master vs PR; pp512 383.3 385.8",
   "retrieved": "2026-09-28",
   "source_name": "llama.cpp issue #29570 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29570",
   "tps": null
  },
  {
   "backend": "llama.cpp",
   "ctx": null,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "M5 Pro (20-core GPU)",
   "id": "lc-29570-m5-pp512-tensor",
   "model": "Qwen3.8-27B",
   "notes": "M5 Pro (20-core GPU), Apple silicon; llama-bench Qwen3.8-27B Q8_0 with -fa 1, master vs the tensor-API FLASH_ATTN_EXT kernel (32 queries per threadgroup, 64-row KV blocks, matmul2d); test pp512; config=fa-ext",
   "pp_tps": "385.8",
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "llama-bench, Qwen3.8-27B Q8_0, -fa 1; M5 Pro (20-core GPU), master vs PR; pp512 383.3 385.8",
   "retrieved": "2026-09-28",
   "source_name": "llama.cpp issue #29570 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29570",
   "tps": null
  },
  {
   "backend": "llama.cpp",
   "ctx": 8192,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "M5 Pro (20-core GPU)",
   "id": "lc-29570-m5-pp512-d8192-master",
   "model": "Qwen3.8-27B",
   "notes": "M5 Pro (20-core GPU), Apple silicon; llama-bench Qwen3.8-27B Q8_0 with -fa 1, master vs the tensor-API FLASH_ATTN_EXT kernel (32 queries per threadgroup, 64-row KV blocks, matmul2d); test pp512 @ d8192; config=master",
   "pp_tps": "295.7",
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "llama-bench, Qwen3.8-27B Q8_0, -fa 1; M5 Pro (20-core GPU), master vs PR; pp512 @ d8192 295.7 340.0",
   "retrieved": "2026-09-28",
   "source_name": "llama.cpp issue #29570 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29570",
   "tps": null
  },
  {
   "backend": "llama.cpp",
   "ctx": 8192,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "M5 Pro (20-core GPU)",
   "id": "lc-29570-m5-pp512-d8192-tensor",
   "model": "Qwen3.8-27B",
   "notes": "M5 Pro (20-core GPU), Apple silicon; llama-bench Qwen3.8-27B Q8_0 with -fa 1, master vs the tensor-API FLASH_ATTN_EXT kernel (32 queries per threadgroup, 64-row KV blocks, matmul2d); test pp512 @ d8192; config=fa-ext",
   "pp_tps": "340.0",
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "llama-bench, Qwen3.8-27B Q8_0, -fa 1; M5 Pro (20-core GPU), master vs PR; pp512 @ d8192 295.7 340.0",
   "retrieved": "2026-09-28",
   "source_name": "llama.cpp issue #29570 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29570",
   "tps": null
  },
  {
   "backend": "llama.cpp",
   "ctx": 32768,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "M5 Pro (20-core GPU)",
   "id": "lc-29570-m5-pp512-d32768-master",
   "model": "Qwen3.8-27B",
   "notes": "M5 Pro (20-core GPU), Apple silicon; llama-bench Qwen3.8-27B Q8_0 with -fa 1, master vs the tensor-API FLASH_ATTN_EXT kernel (32 queries per threadgroup, 64-row KV blocks, matmul2d); test pp512 @ d32768; config=master",
   "pp_tps": "174.3",
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "llama-bench, Qwen3.8-27B Q8_0, -fa 1; M5 Pro (20-core GPU), master vs PR; pp512 @ d32768 174.3 252.9",
   "retrieved": "2026-09-28",
   "source_name": "llama.cpp issue #29570 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29570",
   "tps": null
  },
  {
   "backend": "llama.cpp",
   "ctx": 32768,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "M5 Pro (20-core GPU)",
   "id": "lc-29570-m5-pp512-d32768-tensor",
   "model": "Qwen3.8-27B",
   "notes": "M5 Pro (20-core GPU), Apple silicon; llama-bench Qwen3.8-27B Q8_0 with -fa 1, master vs the tensor-API FLASH_ATTN_EXT kernel (32 queries per threadgroup, 64-row KV blocks, matmul2d); test pp512 @ d32768; config=fa-ext",
   "pp_tps": "252.9",
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "llama-bench, Qwen3.8-27B Q8_0, -fa 1; M5 Pro (20-core GPU), master vs PR; pp512 @ d32768 174.3 252.9",
   "retrieved": "2026-09-28",
   "source_name": "llama.cpp issue #29570 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29570",
   "tps": null
  },
  {
   "backend": "llama.cpp",
   "ctx": 8192,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "M5 Pro (20-core GPU)",
   "id": "lc-29570-m5-pp32-d8192-master",
   "model": "Qwen3.8-27B",
   "notes": "M5 Pro (20-core GPU), Apple silicon; llama-bench Qwen3.8-27B Q8_0 with -fa 1, master vs the tensor-API FLASH_ATTN_EXT kernel (32 queries per threadgroup, 64-row KV blocks, matmul2d); test pp32 @ d8192 (below the kernel threshold); config=master",
   "pp_tps": "107.2",
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "llama-bench, Qwen3.8-27B Q8_0, -fa 1; M5 Pro (20-core GPU), master vs PR; pp32 @ d8192 (below the threshold) 107.2 106.2",
   "retrieved": "2026-09-28",
   "source_name": "llama.cpp issue #29570 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29570",
   "tps": null
  },
  {
   "backend": "llama.cpp",
   "ctx": 8192,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "M5 Pro (20-core GPU)",
   "id": "lc-29570-m5-pp32-d8192-tensor",
   "model": "Qwen3.8-27B",
   "notes": "M5 Pro (20-core GPU), Apple silicon; llama-bench Qwen3.8-27B Q8_0 with -fa 1, master vs the tensor-API FLASH_ATTN_EXT kernel (32 queries per threadgroup, 64-row KV blocks, matmul2d); test pp32 @ d8192 (below the kernel threshold); config=fa-ext",
   "pp_tps": "106.2",
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "llama-bench, Qwen3.8-27B Q8_0, -fa 1; M5 Pro (20-core GPU), master vs PR; pp32 @ d8192 (below the threshold) 107.2 106.2",
   "retrieved": "2026-09-28",
   "source_name": "llama.cpp issue #29570 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29570",
   "tps": null
  },
  {
   "backend": "llama.cpp",
   "ctx": null,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "M5 Pro (20-core GPU)",
   "id": "lc-29570-m5-tg32-master",
   "model": "Qwen3.8-27B",
   "notes": "M5 Pro (20-core GPU), Apple silicon; llama-bench Qwen3.8-27B Q8_0 with -fa 1, master vs the tensor-API FLASH_ATTN_EXT kernel (32 queries per threadgroup, 64-row KV blocks, matmul2d); test tg32; config=master",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "llama-bench, Qwen3.8-27B Q8_0, -fa 1; M5 Pro (20-core GPU), master vs PR; tg32 9.18 9.02",
   "retrieved": "2026-09-28",
   "source_name": "llama.cpp issue #29570 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29570",
   "tps": "9.18"
  },
  {
   "backend": "llama.cpp",
   "ctx": null,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "M5 Pro (20-core GPU)",
   "id": "lc-29570-m5-tg32-tensor",
   "model": "Qwen3.8-27B",
   "notes": "M5 Pro (20-core GPU), Apple silicon; llama-bench Qwen3.8-27B Q8_0 with -fa 1, master vs the tensor-API FLASH_ATTN_EXT kernel (32 queries per threadgroup, 64-row KV blocks, matmul2d); test tg32; config=fa-ext",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "llama-bench, Qwen3.8-27B Q8_0, -fa 1; M5 Pro (20-core GPU), master vs PR; tg32 9.18 9.02",
   "retrieved": "2026-09-28",
   "source_name": "llama.cpp issue #29570 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29570",
   "tps": "9.02"
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": null,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "8x H20",
   "id": "vllm-59010-h20-triton",
   "model": "Qwen3.8-Flash-Next",
   "notes": "8x H20 (96 GB) rig, TP4, CUDA 13.3, torch 2.13.0, branch built with TORCH_CUDA_ARCH_LIST=9.0a; prefill-heavy serving A/B on Qwen/Qwen3.8-Flash-Next-FP8, vllm bench serve random 16000 in / 8 out, concurrency 16, 128 prompts, two seeds (native second seed 15,413 tok/s); stored value is the first published seed of the total token throughput; config=triton",
   "pp_tokens": 16000,
   "pp_tps": "15083",
   "provenance": "community",
   "quant": "FP8",
   "quote": "vllm serve Qwen/Qwen3.8-Flash-Next-FP8 --tensor-parallel-size 4; All runs were on 8x H20 (96 GB), CUDA 13.3, torch 2.13.0; Serving (TP4, H20), 16k-token prompts, 8 output tokens, concurrency 16, 128 prompts, two seeds; total token throughput (tok/s) 15,083 / 15,083 15,416 / 15,413 +2.2%; mean TTFT (s) 12.81 / 12.81 12.53 / 12.53 -2.2%",
   "retrieved": "2026-09-28",
   "source_name": "vLLM issue #59010 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/59010",
   "tg_tokens": 8,
   "ttft_s": "12.81"
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": null,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "8x H20",
   "id": "vllm-59010-h20-native",
   "model": "Qwen3.8-Flash-Next",
   "notes": "8x H20 (96 GB) rig, TP4, CUDA 13.3, torch 2.13.0, branch built with TORCH_CUDA_ARCH_LIST=9.0a; prefill-heavy serving A/B on Qwen/Qwen3.8-Flash-Next-FP8, vllm bench serve random 16000 in / 8 out, concurrency 16, 128 prompts, two seeds (native second seed 15,413 tok/s); stored value is the first published seed of the total token throughput; config=qsa-sm90-native",
   "pp_tokens": 16000,
   "pp_tps": "15416",
   "provenance": "community",
   "quant": "FP8",
   "quote": "vllm serve Qwen/Qwen3.8-Flash-Next-FP8 --tensor-parallel-size 4; All runs were on 8x H20 (96 GB), CUDA 13.3, torch 2.13.0; Serving (TP4, H20), 16k-token prompts, 8 output tokens, concurrency 16, 128 prompts, two seeds; total token throughput (tok/s) 15,083 / 15,083 15,416 / 15,413 +2.2%; mean TTFT (s) 12.81 / 12.81 12.53 / 12.53 -2.2%",
   "retrieved": "2026-09-28",
   "source_name": "vLLM issue #59010 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/59010",
   "tg_tokens": 8,
   "ttft_s": "12.53"
  },
  {
   "backend": "vLLM",
   "batch": 32,
   "ctx": 16384,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "4\u00d7 H20",
   "id": "vllm-58986-h20-control-2048-512-c32",
   "model": "GLM-5.3-Flash",
   "notes": "4x H20 rig, TP4, vLLM 0.30.0+cu129, PyTorch 2.13.0+cu129, driver 535.261.03; 62-shard block-FP8 checkpoint, MTP3, BF16 activations, max length 16384, max sequences 32, batch tokens 4096, memory utilization 0.92, OMP threads 1; vllm bench serve random with range ratio, 128 requests per repeat, 3 repeats, medians of complete paired repeats, fresh cache salt per run; primary comparison is c1-control to c1-return (both arms carry the mHC register-lifetime fix, only the MoE tile config differs); the superseded pre-fix +27.10% headline is historical; P99 TTFT 11271.37 to 9205.95 ms, P99 TPOT 38.02 to 33.33 ms in this comparison; config=moe-default",
   "pp_tokens": 2048,
   "provenance": "community",
   "quant": "FP8",
   "quote": "GLM-5.3-Flash at TP4 uses 288 FP8 experts with a per-rank intermediate size of 512 and 128\u00d7128 weight blocks; 4\u00d7H20, TP4, FP8 weights/BF16 activations, MTP3, max length 16384, max sequences 32, batch tokens 4096; 2048 / 512 / 32 962.39 1137.92",
   "retrieved": "2026-09-28",
   "source_name": "vLLM issue #58986 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58986",
   "tg_tokens": 512,
   "tps": "962.39"
  },
  {
   "backend": "vLLM",
   "batch": 32,
   "ctx": 16384,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "4\u00d7 H20",
   "id": "vllm-58986-h20-tuned-2048-512-c32",
   "model": "GLM-5.3-Flash",
   "notes": "4x H20 rig, TP4, vLLM 0.30.0+cu129, PyTorch 2.13.0+cu129, driver 535.261.03; 62-shard block-FP8 checkpoint, MTP3, BF16 activations, max length 16384, max sequences 32, batch tokens 4096, memory utilization 0.92, OMP threads 1; vllm bench serve random with range ratio, 128 requests per repeat, 3 repeats, medians of complete paired repeats, fresh cache salt per run; primary comparison is c1-control to c1-return (both arms carry the mHC register-lifetime fix, only the MoE tile config differs); the superseded pre-fix +27.10% headline is historical; P99 TTFT 11271.37 to 9205.95 ms, P99 TPOT 38.02 to 33.33 ms in this comparison; config=moe-tuned",
   "pp_tokens": 2048,
   "provenance": "community",
   "quant": "FP8",
   "quote": "GLM-5.3-Flash at TP4 uses 288 FP8 experts with a per-rank intermediate size of 512 and 128\u00d7128 weight blocks; 4\u00d7H20, TP4, FP8 weights/BF16 activations, MTP3, max length 16384, max sequences 32, batch tokens 4096; 2048 / 512 / 32 962.39 1137.92",
   "retrieved": "2026-09-28",
   "source_name": "vLLM issue #58986 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58986",
   "tg_tokens": 512,
   "tps": "1137.92"
  },
  {
   "backend": "vLLM",
   "batch": 32,
   "ctx": 16384,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "4\u00d7 H20",
   "id": "vllm-58986-h20-control-512-1024-c32",
   "model": "GLM-5.3-Flash",
   "notes": "4x H20 rig, TP4, vLLM 0.30.0+cu129, PyTorch 2.13.0+cu129, driver 535.261.03; 62-shard block-FP8 checkpoint, MTP3, BF16 activations, max length 16384, max sequences 32, batch tokens 4096, memory utilization 0.92, OMP threads 1; vllm bench serve random with range ratio, 128 requests per repeat, 3 repeats, medians of complete paired repeats, fresh cache salt per run; primary comparison is c1-control to c1-return (both arms carry the mHC register-lifetime fix, only the MoE tile config differs); the superseded pre-fix +27.10% headline is historical; P99 TTFT 9531.62 to 7236.48 ms, P99 TPOT 26.25 to 19.03 ms in this comparison; config=moe-default",
   "pp_tokens": 512,
   "provenance": "community",
   "quant": "FP8",
   "quote": "GLM-5.3-Flash at TP4 uses 288 FP8 experts with a per-rank intermediate size of 512 and 128\u00d7128 weight blocks; 4\u00d7H20, TP4, FP8 weights/BF16 activations, MTP3, max length 16384, max sequences 32, batch tokens 4096; 512 / 1024 / 32 1619.19 2045.35",
   "retrieved": "2026-09-28",
   "source_name": "vLLM issue #58986 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58986",
   "tg_tokens": 1024,
   "tps": "1619.19"
  },
  {
   "backend": "vLLM",
   "batch": 32,
   "ctx": 16384,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "4\u00d7 H20",
   "id": "vllm-58986-h20-tuned-512-1024-c32",
   "model": "GLM-5.3-Flash",
   "notes": "4x H20 rig, TP4, vLLM 0.30.0+cu129, PyTorch 2.13.0+cu129, driver 535.261.03; 62-shard block-FP8 checkpoint, MTP3, BF16 activations, max length 16384, max sequences 32, batch tokens 4096, memory utilization 0.92, OMP threads 1; vllm bench serve random with range ratio, 128 requests per repeat, 3 repeats, medians of complete paired repeats, fresh cache salt per run; primary comparison is c1-control to c1-return (both arms carry the mHC register-lifetime fix, only the MoE tile config differs); the superseded pre-fix +27.10% headline is historical; P99 TTFT 9531.62 to 7236.48 ms, P99 TPOT 26.25 to 19.03 ms in this comparison; config=moe-tuned",
   "pp_tokens": 512,
   "provenance": "community",
   "quant": "FP8",
   "quote": "GLM-5.3-Flash at TP4 uses 288 FP8 experts with a per-rank intermediate size of 512 and 128\u00d7128 weight blocks; 4\u00d7H20, TP4, FP8 weights/BF16 activations, MTP3, max length 16384, max sequences 32, batch tokens 4096; 512 / 1024 / 32 1619.19 2045.35",
   "retrieved": "2026-09-28",
   "source_name": "vLLM issue #58986 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58986",
   "tg_tokens": 1024,
   "tps": "2045.35"
  },
  {
   "backend": "vLLM",
   "batch": 32,
   "ctx": 16384,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "4\u00d7 H20",
   "id": "vllm-58986-h20-control-512-128-c32",
   "model": "GLM-5.3-Flash",
   "notes": "4x H20 rig, TP4, vLLM 0.30.0+cu129, PyTorch 2.13.0+cu129, driver 535.261.03; 62-shard block-FP8 checkpoint, MTP3, BF16 activations, max length 16384, max sequences 32, batch tokens 4096, memory utilization 0.92, OMP threads 1; vllm bench serve random with range ratio, 128 requests per repeat, 3 repeats, medians of complete paired repeats, fresh cache salt per run; primary comparison is c1-control to c1-return (both arms carry the mHC register-lifetime fix, only the MoE tile config differs); the superseded pre-fix +27.10% headline is historical; P99 TTFT 2697.35 to 2351.59 ms, P99 TPOT 57.26 to 49.48 ms in this comparison; config=moe-default",
   "pp_tokens": 512,
   "provenance": "community",
   "quant": "FP8",
   "quote": "GLM-5.3-Flash at TP4 uses 288 FP8 experts with a per-rank intermediate size of 512 and 128\u00d7128 weight blocks; 4\u00d7H20, TP4, FP8 weights/BF16 activations, MTP3, max length 16384, max sequences 32, batch tokens 4096; 512 / 128 / 32 820.00 903.60",
   "retrieved": "2026-09-28",
   "source_name": "vLLM issue #58986 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58986",
   "tg_tokens": 128,
   "tps": "820.00"
  },
  {
   "backend": "vLLM",
   "batch": 32,
   "ctx": 16384,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "4\u00d7 H20",
   "id": "vllm-58986-h20-tuned-512-128-c32",
   "model": "GLM-5.3-Flash",
   "notes": "4x H20 rig, TP4, vLLM 0.30.0+cu129, PyTorch 2.13.0+cu129, driver 535.261.03; 62-shard block-FP8 checkpoint, MTP3, BF16 activations, max length 16384, max sequences 32, batch tokens 4096, memory utilization 0.92, OMP threads 1; vllm bench serve random with range ratio, 128 requests per repeat, 3 repeats, medians of complete paired repeats, fresh cache salt per run; primary comparison is c1-control to c1-return (both arms carry the mHC register-lifetime fix, only the MoE tile config differs); the superseded pre-fix +27.10% headline is historical; P99 TTFT 2697.35 to 2351.59 ms, P99 TPOT 57.26 to 49.48 ms in this comparison; config=moe-tuned",
   "pp_tokens": 512,
   "provenance": "community",
   "quant": "FP8",
   "quote": "GLM-5.3-Flash at TP4 uses 288 FP8 experts with a per-rank intermediate size of 512 and 128\u00d7128 weight blocks; 4\u00d7H20, TP4, FP8 weights/BF16 activations, MTP3, max length 16384, max sequences 32, batch tokens 4096; 512 / 128 / 32 820.00 903.60",
   "retrieved": "2026-09-28",
   "source_name": "vLLM issue #58986 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/58986",
   "tg_tokens": 128,
   "tps": "903.60"
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": null,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "vllm-59040-radeon8060s-spec-none",
   "model": "Qwen3.5-27B",
   "notes": "Qwen3.5 27B int8 W8A8 via Quark on gfx1151 (Radeon 8060S), vLLM 0.28.0 on ROCm 7.2; single prompt, 400 output tokens, greedy, 3 runs (per-run wall times in the table row); only --speculative-config differs between rows; spec-mtp-before is the zero-initialized mtp.fc bug (0.0% acceptance, 9% slower than no speculation); spec-mtp-after is the one-line quark carve-out fix (22-45% acceptance); ngram speculation on the same box runs 59.62 s at 41% acceptance (no tok/s listed, parked); config=spec-none",
   "pp_tokens": null,
   "provenance": "community",
   "quant": "W8A8",
   "quote": "Qwen3.5 27B, int8 w8a8 via Quark, gfx1151 (Radeon 8060S), vLLM 0.28.0 on ROCm 7.2; no speculation 78.13 / 78.13 / 81.10 s 5.12",
   "retrieved": "2026-09-28",
   "source_name": "vLLM issue #59040 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/59040",
   "tg_tokens": 400,
   "tps": "5.12"
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": null,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "vllm-59040-radeon8060s-spec-mtp-before",
   "model": "Qwen3.5-27B",
   "notes": "Qwen3.5 27B int8 W8A8 via Quark on gfx1151 (Radeon 8060S), vLLM 0.28.0 on ROCm 7.2; single prompt, 400 output tokens, greedy, 3 runs (per-run wall times in the table row); only --speculative-config differs between rows; spec-mtp-before is the zero-initialized mtp.fc bug (0.0% acceptance, 9% slower than no speculation); spec-mtp-after is the one-line quark carve-out fix (22-45% acceptance); ngram speculation on the same box runs 59.62 s at 41% acceptance (no tok/s listed, parked); config=spec-mtp-before",
   "pp_tokens": null,
   "provenance": "community",
   "quant": "W8A8",
   "quote": "Qwen3.5 27B, int8 w8a8 via Quark, gfx1151 (Radeon 8060S), vLLM 0.28.0 on ROCm 7.2; mtp, before 85.63 / 85.64 / 86.27 s 4.67",
   "retrieved": "2026-09-28",
   "source_name": "vLLM issue #59040 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/59040",
   "tg_tokens": 400,
   "tps": "4.67"
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": null,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "vllm-59040-radeon8060s-spec-mtp-after",
   "model": "Qwen3.5-27B",
   "notes": "Qwen3.5 27B int8 W8A8 via Quark on gfx1151 (Radeon 8060S), vLLM 0.28.0 on ROCm 7.2; single prompt, 400 output tokens, greedy, 3 runs (per-run wall times in the table row); only --speculative-config differs between rows; spec-mtp-before is the zero-initialized mtp.fc bug (0.0% acceptance, 9% slower than no speculation); spec-mtp-after is the one-line quark carve-out fix (22-45% acceptance); ngram speculation on the same box runs 59.62 s at 41% acceptance (no tok/s listed, parked); config=spec-mtp-after",
   "pp_tokens": null,
   "provenance": "community",
   "quant": "W8A8",
   "quote": "Qwen3.5 27B, int8 w8a8 via Quark, gfx1151 (Radeon 8060S), vLLM 0.28.0 on ROCm 7.2; mtp, after 63.83 / 63.83 / 64.26 s 6.27",
   "retrieved": "2026-09-28",
   "source_name": "vLLM issue #59040 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/59040",
   "tg_tokens": 400,
   "tps": "6.27"
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": 67000,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "4x CMP 170HX",
   "id": "vllm-59054-cmp170hx-spec-2d-c67k",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "Triton attention backend: spec-decode verify batches (q_len>1) are forced onto the 2D varlen kernel by the max_seqlen_q>1 condition (4 CTAs on a 70-SM GPU at 67K context), MTP k=2, unique-content prompts, thinking off, warmup dropped, steady-state decode; kernel-level table (ms per call, q_len 1-4 at 67K/19K) parked; 67K context, spec decode on, stock 2D-forced path; config=verify-2d",
   "pp_tokens": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "4x CMP 170HX (sm80, PP=4), MiMo-V2.6-Flash-RL official fp8 checkpoint, MTP k=2, steady-state decode; 67K context: 16.5 tok/s (stock) vs 33.7 tok/s (3D verify)",
   "retrieved": "2026-09-28",
   "source_name": "vLLM issue #59054 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/59054",
   "tg_tokens": null,
   "tps": "16.5"
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": 67000,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "4x CMP 170HX",
   "id": "vllm-59054-cmp170hx-spec-3d-c67k",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "Triton attention backend: spec-decode verify batches (q_len>1) are forced onto the 2D varlen kernel by the max_seqlen_q>1 condition (4 CTAs on a 70-SM GPU at 67K context), MTP k=2, unique-content prompts, thinking off, warmup dropped, steady-state decode; kernel-level table (ms per call, q_len 1-4 at 67K/19K) parked; 67K context, spec decode on, 3D split-KV verify path (MiaAI-Lab DGX-Spark recipe, env-gated); config=verify-3d",
   "pp_tokens": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "4x CMP 170HX (sm80, PP=4), MiMo-V2.6-Flash-RL official fp8 checkpoint, MTP k=2, steady-state decode; 67K context: 16.5 tok/s (stock) vs 33.7 tok/s (3D verify)",
   "retrieved": "2026-09-28",
   "source_name": "vLLM issue #59054 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/59054",
   "tg_tokens": null,
   "tps": "33.7"
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": 67000,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "4x CMP 170HX",
   "id": "vllm-59054-cmp170hx-nospec-c67k",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "Triton attention backend: spec-decode verify batches (q_len>1) are forced onto the 2D varlen kernel by the max_seqlen_q>1 condition (4 CTAs on a 70-SM GPU at 67K context), MTP k=2, unique-content prompts, thinking off, warmup dropped, steady-state decode; kernel-level table (ms per call, q_len 1-4 at 67K/19K) parked; 67K context, no speculative decoding; config=no-spec",
   "pp_tokens": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "4x CMP 170HX (sm80, PP=4), MiMo-V2.6-Flash-RL official fp8 checkpoint, MTP k=2, steady-state decode; Without speculative decoding the same rig decodes 67K context at 56.1 tok/s",
   "retrieved": "2026-09-28",
   "source_name": "vLLM issue #59054 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/59054",
   "tg_tokens": null,
   "tps": "56.1"
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": 5120,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "4x CMP 170HX",
   "id": "vllm-59054-cmp170hx-spec-2d-c5k",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "Triton attention backend: spec-decode verify batches (q_len>1) are forced onto the 2D varlen kernel by the max_seqlen_q>1 condition (4 CTAs on a 70-SM GPU at 67K context), MTP k=2, unique-content prompts, thinking off, warmup dropped, steady-state decode; kernel-level table (ms per call, q_len 1-4 at 67K/19K) parked; around 5K context (cliff start point per body), spec decode on, stock 2D-forced path; config=verify-2d",
   "pp_tokens": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "4x CMP 170HX (sm80, PP=4), MiMo-V2.6-Flash-RL official fp8 checkpoint, MTP k=2, steady-state decode; The cliff starts around 5K context (42.5 tok/s with spec vs ~63 without)",
   "retrieved": "2026-09-28",
   "source_name": "vLLM issue #59054 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/59054",
   "tg_tokens": null,
   "tps": "42.5"
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": 5120,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "4x CMP 170HX",
   "id": "vllm-59054-cmp170hx-nospec-c5k",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "Triton attention backend: spec-decode verify batches (q_len>1) are forced onto the 2D varlen kernel by the max_seqlen_q>1 condition (4 CTAs on a 70-SM GPU at 67K context), MTP k=2, unique-content prompts, thinking off, warmup dropped, steady-state decode; kernel-level table (ms per call, q_len 1-4 at 67K/19K) parked; around 5K context (approximate, body says ~63), no speculative decoding; config=no-spec",
   "pp_tokens": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "4x CMP 170HX (sm80, PP=4), MiMo-V2.6-Flash-RL official fp8 checkpoint, MTP k=2, steady-state decode; The cliff starts around 5K context (42.5 tok/s with spec vs ~63 without)",
   "retrieved": "2026-09-28",
   "source_name": "vLLM issue #59054 (community report)",
   "source_url": "https://github.com/vllm-project/vllm/issues/59054",
   "tg_tokens": null,
   "tps": "63"
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "2x Arc Pro B70",
   "id": "lc-29604-master-tg128",
   "model": "Qwen3.8-27B",
   "notes": "tensor-split allreduce on 2 GPUs, master vs master; llama-bench, f16 KV, -fa on, Windows 11 oneAPI 2026.0; tg128; config=master-tg128",
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "On 2x Arc Pro B70 with Qwen3.8-27B in a Q8_0:; tensor tg128: 18.2 -> 20.2 tok/s (~10-11%)",
   "retrieved": "2026-09-28",
   "source_name": "llama.cpp PR #29604",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29604",
   "tg_tokens": 128,
   "tps": "18.2"
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "2x Arc Pro B70",
   "id": "lc-29604-master-vmm1k",
   "model": "Qwen3.8-27B",
   "notes": "tensor-split allreduce on 2 GPUs, master vs master; llama-bench, f16 KV, -fa on, Windows 11 oneAPI 2026.0; pp2048, ub 1024, GGML_SYCL_ENABLE_VMM=1; config=master-vmm1k",
   "pp_tps": 675,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "On 2x Arc Pro B70 with Qwen3.8-27B in a Q8_0:; tensor pp2048: 675 -> 918 tok/s (~36%) with ub 1024 with GGML_SYCL_ENABLE_VMM=1",
   "retrieved": "2026-09-28",
   "source_name": "llama.cpp PR #29604",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29604",
   "tg_tokens": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "2x Arc Pro B70",
   "id": "lc-29604-master-vmm256",
   "model": "Qwen3.8-27B",
   "notes": "tensor-split allreduce on 2 GPUs, master vs master; llama-bench, f16 KV, -fa on, Windows 11 oneAPI 2026.0; pp2048, ub 256, GGML_SYCL_ENABLE_VMM=1; config=master-vmm256",
   "pp_tps": 478,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "On 2x Arc Pro B70 with Qwen3.8-27B in a Q8_0:; tensor pp2048: 478 -> 661 tok/s (~38%) with ub 256 with GGML_SYCL_ENABLE_VMM=1",
   "retrieved": "2026-09-28",
   "source_name": "llama.cpp PR #29604",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29604",
   "tg_tokens": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "2x Arc Pro B70",
   "id": "lc-29604-master-novmm256",
   "model": "Qwen3.8-27B",
   "notes": "tensor-split allreduce on 2 GPUs, master vs master; llama-bench, f16 KV, -fa on, Windows 11 oneAPI 2026.0; pp2048, ub 256, VMM disabled; config=master-novmm256",
   "pp_tps": 477,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "On 2x Arc Pro B70 with Qwen3.8-27B in a Q8_0:; tensor pp2048: 477 -> 692 (+45%) at ub 256  without GGML_SYCL_ENABLE_VMM=0",
   "retrieved": "2026-09-28",
   "source_name": "llama.cpp PR #29604",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29604",
   "tg_tokens": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "2x Arc Pro B70",
   "id": "lc-29604-master-novmm1k",
   "model": "Qwen3.8-27B",
   "notes": "tensor-split allreduce on 2 GPUs, master vs master; llama-bench, f16 KV, -fa on, Windows 11 oneAPI 2026.0; pp2048, ub 1024, VMM disabled; config=master-novmm1k",
   "pp_tps": 665,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "On 2x Arc Pro B70 with Qwen3.8-27B in a Q8_0:; tensor pp2048: 665 -> 928 (+40%) at ub 1024 without GGML_SYCL_ENABLE_VMM=0",
   "retrieved": "2026-09-28",
   "source_name": "llama.cpp PR #29604",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29604",
   "tg_tokens": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "2x Arc Pro B70",
   "id": "lc-29604-patched-tg128",
   "model": "Qwen3.8-27B",
   "notes": "tensor-split allreduce on 2 GPUs, patched vs master; llama-bench, f16 KV, -fa on, Windows 11 oneAPI 2026.0; tg128; config=patched-tg128",
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "On 2x Arc Pro B70 with Qwen3.8-27B in a Q8_0:; tensor tg128: 18.2 -> 20.2 tok/s (~10-11%)",
   "retrieved": "2026-09-28",
   "source_name": "llama.cpp PR #29604",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29604",
   "tg_tokens": 128,
   "tps": "20.2"
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "2x Arc Pro B70",
   "id": "lc-29604-patched-vmm1k",
   "model": "Qwen3.8-27B",
   "notes": "tensor-split allreduce on 2 GPUs, patched vs master; llama-bench, f16 KV, -fa on, Windows 11 oneAPI 2026.0; pp2048, ub 1024, GGML_SYCL_ENABLE_VMM=1; config=patched-vmm1k",
   "pp_tps": 918,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "On 2x Arc Pro B70 with Qwen3.8-27B in a Q8_0:; tensor pp2048: 675 -> 918 tok/s (~36%) with ub 1024 with GGML_SYCL_ENABLE_VMM=1",
   "retrieved": "2026-09-28",
   "source_name": "llama.cpp PR #29604",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29604",
   "tg_tokens": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "2x Arc Pro B70",
   "id": "lc-29604-patched-vmm256",
   "model": "Qwen3.8-27B",
   "notes": "tensor-split allreduce on 2 GPUs, patched vs master; llama-bench, f16 KV, -fa on, Windows 11 oneAPI 2026.0; pp2048, ub 256, GGML_SYCL_ENABLE_VMM=1; config=patched-vmm256",
   "pp_tps": 661,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "On 2x Arc Pro B70 with Qwen3.8-27B in a Q8_0:; tensor pp2048: 478 -> 661 tok/s (~38%) with ub 256 with GGML_SYCL_ENABLE_VMM=1",
   "retrieved": "2026-09-28",
   "source_name": "llama.cpp PR #29604",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29604",
   "tg_tokens": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "2x Arc Pro B70",
   "id": "lc-29604-patched-novmm256",
   "model": "Qwen3.8-27B",
   "notes": "tensor-split allreduce on 2 GPUs, patched vs master; llama-bench, f16 KV, -fa on, Windows 11 oneAPI 2026.0; pp2048, ub 256, VMM disabled; config=patched-novmm256",
   "pp_tps": 692,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "On 2x Arc Pro B70 with Qwen3.8-27B in a Q8_0:; tensor pp2048: 477 -> 692 (+45%) at ub 256  without GGML_SYCL_ENABLE_VMM=0",
   "retrieved": "2026-09-28",
   "source_name": "llama.cpp PR #29604",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29604",
   "tg_tokens": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "2x Arc Pro B70",
   "id": "lc-29604-patched-novmm1k",
   "model": "Qwen3.8-27B",
   "notes": "tensor-split allreduce on 2 GPUs, patched vs master; llama-bench, f16 KV, -fa on, Windows 11 oneAPI 2026.0; pp2048, ub 1024, VMM disabled; config=patched-novmm1k",
   "pp_tps": 928,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "On 2x Arc Pro B70 with Qwen3.8-27B in a Q8_0:; tensor pp2048: 665 -> 928 (+40%) at ub 1024 without GGML_SYCL_ENABLE_VMM=0",
   "retrieved": "2026-09-28",
   "source_name": "llama.cpp PR #29604",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29604",
   "tg_tokens": null
  },
  {
   "backend": "llama.cpp (Hexagon)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "OnePlus 13T",
   "id": "lc-29606-htp-fixed-pp128",
   "model": "Qwen3.5-4B",
   "notes": "Hexagon HTP0, -ngl 99 -t 4 -p 128 -n 32 -r 2, NDK r28b / Hexagon SDK 6.6.0.0; fixed vs vanilla queue; pp128; config=htp-fixed",
   "pp_tps": "665.20",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Previously recorded upstream hardware reproduction: OnePlus 13T / v79, Qwen3.5-4B Q4_0; Fixed vs vanilla measured 665.20 vs 669.88 tok/s for pp128 and 14.38 vs 14.47 tok/s for tg32",
   "retrieved": "2026-09-28",
   "source_name": "llama.cpp PR #29606",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29606",
   "tg_tokens": null
  },
  {
   "backend": "llama.cpp (Hexagon)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "OnePlus 13T",
   "id": "lc-29606-htp-vanilla-pp128",
   "model": "Qwen3.5-4B",
   "notes": "Hexagon HTP0, -ngl 99 -t 4 -p 128 -n 32 -r 2, NDK r28b / Hexagon SDK 6.6.0.0; vanilla vs vanilla queue; pp128; config=htp-vanilla",
   "pp_tps": "669.88",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Previously recorded upstream hardware reproduction: OnePlus 13T / v79, Qwen3.5-4B Q4_0; Fixed vs vanilla measured 665.20 vs 669.88 tok/s for pp128 and 14.38 vs 14.47 tok/s for tg32",
   "retrieved": "2026-09-28",
   "source_name": "llama.cpp PR #29606",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29606",
   "tg_tokens": null
  },
  {
   "backend": "llama.cpp (Hexagon)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "OnePlus 13T",
   "id": "lc-29606-htp-fixed-tg32",
   "model": "Qwen3.5-4B",
   "notes": "Hexagon HTP0, -ngl 99 -t 4 -p 128 -n 32 -r 2, NDK r28b / Hexagon SDK 6.6.0.0; fixed vs vanilla queue; tg32; config=htp-fixed",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Previously recorded upstream hardware reproduction: OnePlus 13T / v79, Qwen3.5-4B Q4_0; Fixed vs vanilla measured 665.20 vs 669.88 tok/s for pp128 and 14.38 vs 14.47 tok/s for tg32",
   "retrieved": "2026-09-28",
   "source_name": "llama.cpp PR #29606",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29606",
   "tg_tokens": 32,
   "tps": "14.38"
  },
  {
   "backend": "llama.cpp (Hexagon)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "OnePlus 13T",
   "id": "lc-29606-htp-vanilla-tg32",
   "model": "Qwen3.5-4B",
   "notes": "Hexagon HTP0, -ngl 99 -t 4 -p 128 -n 32 -r 2, NDK r28b / Hexagon SDK 6.6.0.0; vanilla vs vanilla queue; tg32; config=htp-vanilla",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Previously recorded upstream hardware reproduction: OnePlus 13T / v79, Qwen3.5-4B Q4_0; Fixed vs vanilla measured 665.20 vs 669.88 tok/s for pp128 and 14.38 vs 14.47 tok/s for tg32",
   "retrieved": "2026-09-28",
   "source_name": "llama.cpp PR #29606",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29606",
   "tg_tokens": 32,
   "tps": "14.47"
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 100000,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "AMX",
   "id": "vllm-59082-amx-main-no-mtp-auto",
   "model": "Qwen3.5-4B",
   "notes": "vllm bench serve, random 100k input / 1k output, max-concurrency 1, 3 timed requests, no MTP, auto KV; main vs main; config=main-no-mtp-auto",
   "provenance": "community",
   "quant": "BF16",
   "quote": "Improve long-context MTP attention scheduling on AMX; Main `6b24bd8e75a3` values are medians of three timed requests; Serving used 100k input/1k output tokens with zero failures; No MTP, auto KV | 6.73 | 7.08 | 79.29 | 75.49; vllm serve Qwen/Qwen3.5-4B --dtype bfloat16",
   "retrieved": "2026-09-28",
   "source_name": "vLLM PR #59082",
   "source_url": "https://github.com/vllm-project/vllm/pull/59082",
   "tg_tokens": null,
   "tps": 6.73
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 100000,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "AMX",
   "id": "vllm-59082-amx-branch-no-mtp-auto",
   "model": "Qwen3.5-4B",
   "notes": "vllm bench serve, random 100k input / 1k output, max-concurrency 1, 3 timed requests, no MTP, auto KV; branch vs main; config=branch-no-mtp-auto",
   "provenance": "community",
   "quant": "BF16",
   "quote": "Improve long-context MTP attention scheduling on AMX; Main `6b24bd8e75a3` values are medians of three timed requests; Serving used 100k input/1k output tokens with zero failures; No MTP, auto KV | 6.73 | 7.08 | 79.29 | 75.49; vllm serve Qwen/Qwen3.5-4B --dtype bfloat16",
   "retrieved": "2026-09-28",
   "source_name": "vLLM PR #59082",
   "source_url": "https://github.com/vllm-project/vllm/pull/59082",
   "tg_tokens": null,
   "tps": 7.08
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 100000,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "AMX",
   "id": "vllm-59082-amx-main-mtp3-auto",
   "model": "Qwen3.5-4B",
   "notes": "vllm bench serve, random 100k input / 1k output, max-concurrency 1, 3 timed requests, MTP3 spec decode (3 tokens), auto KV; main vs main; config=main-mtp3-auto",
   "provenance": "community",
   "quant": "BF16",
   "quote": "Improve long-context MTP attention scheduling on AMX; Main `6b24bd8e75a3` values are medians of three timed requests; Serving used 100k input/1k output tokens with zero failures; MTP3, auto KV | 8.01 | 9.48 | 48.72 | 34.33; vllm serve Qwen/Qwen3.5-4B --dtype bfloat16",
   "retrieved": "2026-09-28",
   "source_name": "vLLM PR #59082",
   "source_url": "https://github.com/vllm-project/vllm/pull/59082",
   "tg_tokens": null,
   "tps": 8.01
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 100000,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "AMX",
   "id": "vllm-59082-amx-branch-mtp3-auto",
   "model": "Qwen3.5-4B",
   "notes": "vllm bench serve, random 100k input / 1k output, max-concurrency 1, 3 timed requests, MTP3 spec decode (3 tokens), auto KV; branch vs main; config=branch-mtp3-auto",
   "provenance": "community",
   "quant": "BF16",
   "quote": "Improve long-context MTP attention scheduling on AMX; Main `6b24bd8e75a3` values are medians of three timed requests; Serving used 100k input/1k output tokens with zero failures; MTP3, auto KV | 8.01 | 9.48 | 48.72 | 34.33; vllm serve Qwen/Qwen3.5-4B --dtype bfloat16",
   "retrieved": "2026-09-28",
   "source_name": "vLLM PR #59082",
   "source_url": "https://github.com/vllm-project/vllm/pull/59082",
   "tg_tokens": null,
   "tps": 9.48
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 100000,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "AMX",
   "id": "vllm-59082-amx-main-mtp3-fp8kv",
   "model": "Qwen3.5-4B",
   "notes": "vllm bench serve, random 100k input / 1k output, max-concurrency 1, 3 timed requests, MTP3 spec decode (3 tokens), FP8 KV cache; main vs main; config=main-mtp3-fp8kv",
   "provenance": "community",
   "quant": "BF16",
   "quote": "Improve long-context MTP attention scheduling on AMX; Main `6b24bd8e75a3` values are medians of three timed requests; Serving used 100k input/1k output tokens with zero failures; MTP3, FP8 KV, post-fix | 8.62 | 9.36 | 37.25 | 27.57; vllm serve Qwen/Qwen3.5-4B --dtype bfloat16",
   "retrieved": "2026-09-28",
   "source_name": "vLLM PR #59082",
   "source_url": "https://github.com/vllm-project/vllm/pull/59082",
   "tg_tokens": null,
   "tps": 8.62
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 100000,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "AMX",
   "id": "vllm-59082-amx-branch-mtp3-fp8kv",
   "model": "Qwen3.5-4B",
   "notes": "vllm bench serve, random 100k input / 1k output, max-concurrency 1, MTP3 spec decode (3 tokens), FP8 KV cache, post-fix run (one timed request after one warmup, legacy-materializer routing); branch vs main; config=branch-mtp3-fp8kv",
   "provenance": "community",
   "quant": "BF16",
   "quote": "Improve long-context MTP attention scheduling on AMX; Main `6b24bd8e75a3` values are medians of three timed requests; Serving used 100k input/1k output tokens with zero failures; MTP3, FP8 KV, post-fix | 8.62 | 9.36 | 37.25 | 27.57; vllm serve Qwen/Qwen3.5-4B --dtype bfloat16",
   "retrieved": "2026-09-28",
   "source_name": "vLLM PR #59082",
   "source_url": "https://github.com/vllm-project/vllm/pull/59082",
   "tg_tokens": null,
   "tps": 9.36
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "RTX 5070 Ti Laptop 12 GB",
   "id": "lc-29619-rtx5070ti-mtp-off-c16k",
   "model": "Qwen3-27B",
   "notes": "needle harness, 12 960-token prompt, --spec-draft-n-max 4, ctx 16 384, KV q4_0, MTP off baseline; VRAM peak 10 744 MiB; config=mtp-off",
   "params": "27B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "REAP192",
   "quote": "single RTX 5070 Ti Laptop 12 GB (Windows); Measured on Qwen3-27B REAP192 with the Qwen3-MTP drafter; MTP OFF | 10 744 MiB | 28.7 | \u2014 | \u2713 8271 / TANGO-CHARLIE",
   "ram_gb": 12,
   "retrieved": "2026-09-28",
   "source_name": "llama.cpp PR #29619",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29619",
   "tg_tokens": null,
   "tps": 28.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "RTX 5070 Ti Laptop 12 GB",
   "id": "lc-29619-rtx5070ti-cpu-mtp-c16k",
   "model": "Qwen3-27B",
   "notes": "pure --cpu-mtp, MTP block + RS snapshots on host, acceptance 69.2 pct; VRAM peak 10 944 MiB; config=cpu-mtp",
   "params": "27B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "REAP192",
   "quote": "single RTX 5070 Ti Laptop 12 GB (Windows); Measured on Qwen3-27B REAP192 with the Qwen3-MTP drafter; pure `--cpu-mtp` | 10 944 MiB | 31.1 | 69.2 % | \u2713 8271 / TANGO-CHARLIE",
   "ram_gb": 12,
   "retrieved": "2026-09-28",
   "source_name": "llama.cpp PR #29619",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29619",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "RTX 5070 Ti Laptop 12 GB",
   "id": "lc-29619-rtx5070ti-hybrid-c16k",
   "model": "Qwen3-27B",
   "notes": "hybrid, --cpu-mtp plus -ot pins MTP block to GPU, acceptance 72.4 pct; VRAM peak 11 128 MiB, saves ~751-755 MiB vs full-GPU MTP; config=hybrid",
   "params": "27B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "REAP192",
   "quote": "single RTX 5070 Ti Laptop 12 GB (Windows); Measured on Qwen3-27B REAP192 with the Qwen3-MTP drafter; **hybrid** (`--cpu-mtp -ot 'blk\\.64\\..*=CUDA0'`) | **11 128 MiB** | **51.2** | **72.4 %** | \u2713 8271 / TANGO-CHARLIE",
   "ram_gb": 12,
   "retrieved": "2026-09-28",
   "source_name": "llama.cpp PR #29619",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29619",
   "tg_tokens": null,
   "tps": 51.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "RTX 5070 Ti Laptop 12 GB",
   "id": "lc-29619-rtx5070ti-mtp-full-c16k",
   "model": "Qwen3-27B",
   "notes": "full-GPU MTP baseline, acceptance 79.6 pct; VRAM peak 11 766 MiB; config=mtp-full",
   "params": "27B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "REAP192",
   "quote": "single RTX 5070 Ti Laptop 12 GB (Windows); Measured on Qwen3-27B REAP192 with the Qwen3-MTP drafter; full-GPU MTP | 11 766 MiB | 62.5 | 79.6 % | \u2713 8271 / TANGO-CHARLIE",
   "ram_gb": 12,
   "retrieved": "2026-09-28",
   "source_name": "llama.cpp PR #29619",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29619",
   "tg_tokens": null,
   "tps": 62.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 60000,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "RTX 5070 Ti Laptop 12 GB",
   "id": "lc-29619-rtx5070ti-cpu-mtp-c60k",
   "model": "Qwen3-27B",
   "notes": "prior campaign at 60 k ctx, pure --cpu-mtp; config=cpu-mtp-c60k",
   "params": "27B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "REAP192",
   "quote": "single RTX 5070 Ti Laptop 12 GB (Windows); Measured on Qwen3-27B REAP192 with the Qwen3-MTP drafter; Longer context widens the hybrid advantage: prior campaign measured +36 % gen for hybrid over pure `--cpu-mtp` at 60 k ctx (40.8 vs. 30.0 t/s) on the same model.",
   "ram_gb": 12,
   "retrieved": "2026-09-28",
   "source_name": "llama.cpp PR #29619",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29619",
   "tg_tokens": null,
   "tps": 30.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 60000,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "RTX 5070 Ti Laptop 12 GB",
   "id": "lc-29619-rtx5070ti-hybrid-c60k",
   "model": "Qwen3-27B",
   "notes": "prior campaign at 60 k ctx, hybrid, +36 pct gen over pure --cpu-mtp; config=hybrid-c60k",
   "params": "27B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "REAP192",
   "quote": "single RTX 5070 Ti Laptop 12 GB (Windows); Measured on Qwen3-27B REAP192 with the Qwen3-MTP drafter; Longer context widens the hybrid advantage: prior campaign measured +36 % gen for hybrid over pure `--cpu-mtp` at 60 k ctx (40.8 vs. 30.0 t/s) on the same model.",
   "ram_gb": 12,
   "retrieved": "2026-09-28",
   "source_name": "llama.cpp PR #29619",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29619",
   "tg_tokens": null,
   "tps": 40.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "RTX 5070 Ti Laptop 12 GB",
   "id": "lc-29620-rtx5070ti-mtp-off-c16k",
   "model": "Qwen3-27B",
   "notes": "fresh sweep on the final PR commits, 12 960-token needle prompt, --spec-draft-n-max 4, ctx 16 384, KV q4_0, MTP off baseline; VRAM peak 10 744 MiB; config=mtp-off",
   "params": "27B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "REAP192",
   "quote": "single RTX 5070 Ti Laptop 12 GB (Windows); Measured on Qwen3-27B REAP192 with the Qwen3-MTP drafter; MTP OFF | 10 744 MiB | 28.9 | - | correct",
   "ram_gb": 12,
   "retrieved": "2026-09-29",
   "source_name": "llama.cpp PR #29620",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29620",
   "tg_tokens": null,
   "tps": 28.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "RTX 5070 Ti Laptop 12 GB",
   "id": "lc-29620-rtx5070ti-cpu-mtp-c16k",
   "model": "Qwen3-27B",
   "notes": "pure --cpu-mtp on final commits, acceptance 78.1 pct, +27 pct gen over MTP-off; VRAM peak 10 944 MiB; config=cpu-mtp-final",
   "params": "27B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "REAP192",
   "quote": "single RTX 5070 Ti Laptop 12 GB (Windows); Measured on Qwen3-27B REAP192 with the Qwen3-MTP drafter; pure `--cpu-mtp` | 10 944 MiB | 36.7 | 78.1 % | correct",
   "ram_gb": 12,
   "retrieved": "2026-09-29",
   "source_name": "llama.cpp PR #29620",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29620",
   "tg_tokens": null,
   "tps": 36.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "RTX 5070 Ti Laptop 12 GB",
   "id": "lc-29620-rtx5070ti-hybrid-c16k",
   "model": "Qwen3-27B",
   "notes": "hybrid, --cpu-mtp plus -ot pins MTP block to GPU, acceptance 78.1 pct, 94 pct of full-GPU throughput at 638 MiB less VRAM; config=hybrid",
   "params": "27B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "REAP192",
   "quote": "single RTX 5070 Ti Laptop 12 GB (Windows); Measured on Qwen3-27B REAP192 with the Qwen3-MTP drafter; hybrid (`--cpu-mtp -ot 'blk\\.64\\..*=CUDA0'`) | 11 128 MiB | 55.2 | 78.1 % | correct",
   "ram_gb": 12,
   "retrieved": "2026-09-29",
   "source_name": "llama.cpp PR #29620",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29620",
   "tg_tokens": null,
   "tps": 55.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "RTX 5070 Ti Laptop 12 GB",
   "id": "lc-29620-rtx5070ti-mtp-full-c16k",
   "model": "Qwen3-27B",
   "notes": "full-GPU MTP baseline on final commits, acceptance 70.2 pct; VRAM peak 11 766 MiB; config=mtp-full",
   "params": "27B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "REAP192",
   "quote": "single RTX 5070 Ti Laptop 12 GB (Windows); Measured on Qwen3-27B REAP192 with the Qwen3-MTP drafter; full-GPU MTP | 11 766 MiB | 58.7 | 70.2 % | correct",
   "ram_gb": 12,
   "retrieved": "2026-09-29",
   "source_name": "llama.cpp PR #29620",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29620",
   "tg_tokens": null,
   "tps": 58.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 60000,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "RTX 5070 Ti Laptop 12 GB",
   "id": "lc-29620-rtx5070ti-cpu-mtp-c60k",
   "model": "Qwen3-27B",
   "notes": "prior campaign at 60 k ctx, pure --cpu-mtp; config=cpu-mtp-c60k",
   "params": "27B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "REAP192",
   "quote": "single RTX 5070 Ti Laptop 12 GB (Windows); Measured on Qwen3-27B REAP192 with the Qwen3-MTP drafter; Longer context widens the hybrid advantage: prior campaign measured +36 % gen for hybrid over pure `--cpu-mtp` at 60 k ctx (40.8 vs. 30.0 t/s) on the same model.",
   "ram_gb": 12,
   "retrieved": "2026-09-29",
   "source_name": "llama.cpp PR #29620",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29620",
   "tg_tokens": null,
   "tps": 30.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 60000,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "RTX 5070 Ti Laptop 12 GB",
   "id": "lc-29620-rtx5070ti-hybrid-c60k",
   "model": "Qwen3-27B",
   "notes": "prior campaign at 60 k ctx, hybrid, +36 pct gen over pure --cpu-mtp; config=hybrid-c60k",
   "params": "27B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "REAP192",
   "quote": "single RTX 5070 Ti Laptop 12 GB (Windows); Measured on Qwen3-27B REAP192 with the Qwen3-MTP drafter; Longer context widens the hybrid advantage: prior campaign measured +36 % gen for hybrid over pure `--cpu-mtp` at 60 k ctx (40.8 vs. 30.0 t/s) on the same model.",
   "ram_gb": 12,
   "retrieved": "2026-09-29",
   "source_name": "llama.cpp PR #29620",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29620",
   "tg_tokens": null,
   "tps": 40.8,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": null,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "4\u00d7 NVIDIA CMP 170HX",
   "id": "vllm-59109-cmp170hx-before-c32k",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "single-stream decode before GLOO_SOCKET_IFNAME=lo fix (docker bridge control plane, ~535 ms per gloo send); config=before-c32k",
   "params": "764B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL3 2/3 bpw mixed",
   "quote": "4\u00d7 NVIDIA CMP 170HX (GA100, sm_80, 64 GiB, PCIe Gen2 x4, no P2P); DeepSeek-V4.1-Flash (764B, EXL3 2/3 bpw mixed), PP4 (10,10,11,9); Single-stream decode, warm, identical prompts:; | 32K | 57.0 | 60.7 | +6 % |",
   "ram_gb": 64,
   "retrieved": "2026-09-29",
   "source_name": "vLLM issue #59109",
   "source_url": "https://github.com/vllm-project/vllm/issues/59109",
   "tg_tokens": null,
   "tps": 57.0,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": null,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "4\u00d7 NVIDIA CMP 170HX",
   "id": "vllm-59109-cmp170hx-after-c32k",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "single-stream decode after GLOO_SOCKET_IFNAME=lo fix; config=after-c32k",
   "params": "764B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL3 2/3 bpw mixed",
   "quote": "4\u00d7 NVIDIA CMP 170HX (GA100, sm_80, 64 GiB, PCIe Gen2 x4, no P2P); DeepSeek-V4.1-Flash (764B, EXL3 2/3 bpw mixed), PP4 (10,10,11,9); Single-stream decode, warm, identical prompts:; | 32K | 57.0 | 60.7 | +6 % |",
   "ram_gb": 64,
   "retrieved": "2026-09-29",
   "source_name": "vLLM issue #59109",
   "source_url": "https://github.com/vllm-project/vllm/issues/59109",
   "tg_tokens": null,
   "tps": 60.7,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": null,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "4\u00d7 NVIDIA CMP 170HX",
   "id": "vllm-59109-cmp170hx-before-c128k",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "single-stream decode before fix; config=before-c128k",
   "params": "764B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL3 2/3 bpw mixed",
   "quote": "4\u00d7 NVIDIA CMP 170HX (GA100, sm_80, 64 GiB, PCIe Gen2 x4, no P2P); DeepSeek-V4.1-Flash (764B, EXL3 2/3 bpw mixed), PP4 (10,10,11,9); Single-stream decode, warm, identical prompts:; | 128K | 46.7 | 58.3 | +25 % |",
   "ram_gb": 64,
   "retrieved": "2026-09-29",
   "source_name": "vLLM issue #59109",
   "source_url": "https://github.com/vllm-project/vllm/issues/59109",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": null,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "4\u00d7 NVIDIA CMP 170HX",
   "id": "vllm-59109-cmp170hx-after-c128k",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "single-stream decode after fix; config=after-c128k",
   "params": "764B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL3 2/3 bpw mixed",
   "quote": "4\u00d7 NVIDIA CMP 170HX (GA100, sm_80, 64 GiB, PCIe Gen2 x4, no P2P); DeepSeek-V4.1-Flash (764B, EXL3 2/3 bpw mixed), PP4 (10,10,11,9); Single-stream decode, warm, identical prompts:; | 128K | 46.7 | 58.3 | +25 % |",
   "ram_gb": 64,
   "retrieved": "2026-09-29",
   "source_name": "vLLM issue #59109",
   "source_url": "https://github.com/vllm-project/vllm/issues/59109",
   "tg_tokens": null,
   "tps": 58.3,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": null,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "4\u00d7 NVIDIA CMP 170HX",
   "id": "vllm-59109-cmp170hx-before-c300k",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "single-stream decode before fix, approximate (~40) per the report; config=before-c300k",
   "params": "764B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL3 2/3 bpw mixed",
   "quote": "4\u00d7 NVIDIA CMP 170HX (GA100, sm_80, 64 GiB, PCIe Gen2 x4, no P2P); DeepSeek-V4.1-Flash (764B, EXL3 2/3 bpw mixed), PP4 (10,10,11,9); Single-stream decode, warm, identical prompts:; | 300K | ~40 | 56.7 | +40 % |",
   "ram_gb": 64,
   "retrieved": "2026-09-29",
   "source_name": "vLLM issue #59109",
   "source_url": "https://github.com/vllm-project/vllm/issues/59109",
   "tg_tokens": null,
   "tps": 40,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": null,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "4\u00d7 NVIDIA CMP 170HX",
   "id": "vllm-59109-cmp170hx-after-c300k",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "single-stream decode after fix; config=after-c300k",
   "params": "764B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL3 2/3 bpw mixed",
   "quote": "4\u00d7 NVIDIA CMP 170HX (GA100, sm_80, 64 GiB, PCIe Gen2 x4, no P2P); DeepSeek-V4.1-Flash (764B, EXL3 2/3 bpw mixed), PP4 (10,10,11,9); Single-stream decode, warm, identical prompts:; | 300K | ~40 | 56.7 | +40 % |",
   "ram_gb": 64,
   "retrieved": "2026-09-29",
   "source_name": "vLLM issue #59109",
   "source_url": "https://github.com/vllm-project/vllm/issues/59109",
   "tg_tokens": null,
   "tps": 56.7,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": null,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "4\u00d7 NVIDIA CMP 170HX",
   "id": "vllm-59109-cmp170hx-before-c524k",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "single-stream decode before fix; config=before-c524k",
   "params": "764B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL3 2/3 bpw mixed",
   "quote": "4\u00d7 NVIDIA CMP 170HX (GA100, sm_80, 64 GiB, PCIe Gen2 x4, no P2P); DeepSeek-V4.1-Flash (764B, EXL3 2/3 bpw mixed), PP4 (10,10,11,9); Single-stream decode, warm, identical prompts:; | 524K | 35.8 | **54.7** | **+53 %** |",
   "ram_gb": 64,
   "retrieved": "2026-09-29",
   "source_name": "vLLM issue #59109",
   "source_url": "https://github.com/vllm-project/vllm/issues/59109",
   "tg_tokens": null,
   "tps": 35.8,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": null,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "4\u00d7 NVIDIA CMP 170HX",
   "id": "vllm-59109-cmp170hx-after-c524k",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "single-stream decode after fix, +53 pct over pre-fix; config=after-c524k",
   "params": "764B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL3 2/3 bpw mixed",
   "quote": "4\u00d7 NVIDIA CMP 170HX (GA100, sm_80, 64 GiB, PCIe Gen2 x4, no P2P); DeepSeek-V4.1-Flash (764B, EXL3 2/3 bpw mixed), PP4 (10,10,11,9); Single-stream decode, warm, identical prompts:; | 524K | 35.8 | **54.7** | **+53 %** |",
   "ram_gb": 64,
   "retrieved": "2026-09-29",
   "source_name": "vLLM issue #59109",
   "source_url": "https://github.com/vllm-project/vllm/issues/59109",
   "tg_tokens": null,
   "tps": 54.7,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": null,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "4\u00d7 NVIDIA CMP 170HX",
   "id": "vllm-59109-cmp170hx-after-c8",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "mixed-scenario warm round after fix, concurrency C8; config=after-c8",
   "params": "764B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL3 2/3 bpw mixed",
   "quote": "4\u00d7 NVIDIA CMP 170HX (GA100, sm_80, 64 GiB, PCIe Gen2 x4, no P2P); DeepSeek-V4.1-Flash (764B, EXL3 2/3 bpw mixed), PP4 (10,10,11,9); Single-stream decode, warm, identical prompts:; concurrency C8 200.6 / C16 199.4 / C24 252.1 decode tok/s (mixed scenarios, warm round)",
   "ram_gb": 64,
   "retrieved": "2026-09-29",
   "source_name": "vLLM issue #59109",
   "source_url": "https://github.com/vllm-project/vllm/issues/59109",
   "tg_tokens": null,
   "tps": 200.6,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": null,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "4\u00d7 NVIDIA CMP 170HX",
   "id": "vllm-59109-cmp170hx-after-c16",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "mixed-scenario warm round after fix, concurrency C16; config=after-c16",
   "params": "764B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL3 2/3 bpw mixed",
   "quote": "4\u00d7 NVIDIA CMP 170HX (GA100, sm_80, 64 GiB, PCIe Gen2 x4, no P2P); DeepSeek-V4.1-Flash (764B, EXL3 2/3 bpw mixed), PP4 (10,10,11,9); Single-stream decode, warm, identical prompts:; concurrency C8 200.6 / C16 199.4 / C24 252.1 decode tok/s (mixed scenarios, warm round)",
   "ram_gb": 64,
   "retrieved": "2026-09-29",
   "source_name": "vLLM issue #59109",
   "source_url": "https://github.com/vllm-project/vllm/issues/59109",
   "tg_tokens": null,
   "tps": 199.4,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": null,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "4\u00d7 NVIDIA CMP 170HX",
   "id": "vllm-59109-cmp170hx-after-c24",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "mixed-scenario warm round after fix, concurrency C24; config=after-c24",
   "params": "764B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL3 2/3 bpw mixed",
   "quote": "4\u00d7 NVIDIA CMP 170HX (GA100, sm_80, 64 GiB, PCIe Gen2 x4, no P2P); DeepSeek-V4.1-Flash (764B, EXL3 2/3 bpw mixed), PP4 (10,10,11,9); Single-stream decode, warm, identical prompts:; concurrency C8 200.6 / C16 199.4 / C24 252.1 decode tok/s (mixed scenarios, warm round)",
   "ram_gb": 64,
   "retrieved": "2026-09-29",
   "source_name": "vLLM issue #59109",
   "source_url": "https://github.com/vllm-project/vllm/issues/59109",
   "tg_tokens": null,
   "tps": 252.1,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": null,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "B300",
   "id": "vllm-59112-b300-lbhnc-c1",
   "model": "Qwen3.6-35B-A3B-FP8",
   "notes": "LBHNC layout, 1024-in/1024-out c1 serve, MRv2 spec decode 7 tokens, BF16 KV, averaged over two runs; config=lbhnc-c1",
   "params": "35B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "1\u00d7 B300, PyTorch 2.13.0+cu130, FlashInfer 0.7.0; Qwen3.6-35B-A3B-FP8 + z-lab/Qwen3.6-35B-A3B-DFlash; | 1024 input / 1024 output, c1 | output tok/s | 286.56 | 427.01 | **+49.0%** |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "vLLM PR #59112",
   "source_url": "https://github.com/vllm-project/vllm/pull/59112",
   "tg_tokens": null,
   "tps": 286.56,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": null,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "B300",
   "id": "vllm-59112-b300-blhnc-c1",
   "model": "Qwen3.6-35B-A3B-FP8",
   "notes": "BLHNC packed layout, 1024-in/1024-out c1 serve, +49.0 pct over LBHNC, KV groups 46 to 5; config=blhnc-c1",
   "params": "35B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "1\u00d7 B300, PyTorch 2.13.0+cu130, FlashInfer 0.7.0; Qwen3.6-35B-A3B-FP8 + z-lab/Qwen3.6-35B-A3B-DFlash; | 1024 input / 1024 output, c1 | output tok/s | 286.56 | 427.01 | **+49.0%** |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "vLLM PR #59112",
   "source_url": "https://github.com/vllm-project/vllm/pull/59112",
   "tg_tokens": null,
   "tps": 427.01,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": null,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "B300",
   "id": "vllm-59112-b300-lbhnc-c32",
   "model": "Qwen3.6-35B-A3B-FP8",
   "notes": "LBHNC layout, 1024-in/1024-out c32 serve, averaged over two runs; config=lbhnc-c32",
   "params": "35B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "1\u00d7 B300, PyTorch 2.13.0+cu130, FlashInfer 0.7.0; Qwen3.6-35B-A3B-FP8 + z-lab/Qwen3.6-35B-A3B-DFlash; | 1024 input / 1024 output, c32 | output tok/s | 5115.53 | 7221.04 | **+41.2%** |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "vLLM PR #59112",
   "source_url": "https://github.com/vllm-project/vllm/pull/59112",
   "tg_tokens": null,
   "tps": 5115.53,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": null,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "B300",
   "id": "vllm-59112-b300-blhnc-c32",
   "model": "Qwen3.6-35B-A3B-FP8",
   "notes": "BLHNC packed layout, 1024-in/1024-out c32 serve, +41.2 pct over LBHNC; per-run 7258.52 and 7183.57 tok/s; config=blhnc-c32",
   "params": "35B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "1\u00d7 B300, PyTorch 2.13.0+cu130, FlashInfer 0.7.0; Qwen3.6-35B-A3B-FP8 + z-lab/Qwen3.6-35B-A3B-DFlash; | 1024 input / 1024 output, c32 | output tok/s | 5115.53 | 7221.04 | **+41.2%** |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "vLLM PR #59112",
   "source_url": "https://github.com/vllm-project/vllm/pull/59112",
   "tg_tokens": null,
   "tps": 7221.04,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": null,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "B300",
   "id": "vllm-59112-b300-lbhnc-8k1",
   "model": "Qwen3.6-35B-A3B-FP8",
   "notes": "LBHNC layout, 8192-in/1-out c1, req/s at one output token; config=lbhnc-8k1",
   "params": "35B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "1\u00d7 B300, PyTorch 2.13.0+cu130, FlashInfer 0.7.0; Qwen3.6-35B-A3B-FP8 + z-lab/Qwen3.6-35B-A3B-DFlash; | 8192 input / 1 output, c1 | req/s | 6.419 | 7.015 | **+9.3%** |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "vLLM PR #59112",
   "source_url": "https://github.com/vllm-project/vllm/pull/59112",
   "tg_tokens": null,
   "tps": 6.419,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": null,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "B300",
   "id": "vllm-59112-b300-blhnc-8k1",
   "model": "Qwen3.6-35B-A3B-FP8",
   "notes": "BLHNC packed layout, 8192-in/1-out c1, req/s at one output token, +9.3 pct; config=blhnc-8k1",
   "params": "35B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "1\u00d7 B300, PyTorch 2.13.0+cu130, FlashInfer 0.7.0; Qwen3.6-35B-A3B-FP8 + z-lab/Qwen3.6-35B-A3B-DFlash; | 8192 input / 1 output, c1 | req/s | 6.419 | 7.015 | **+9.3%** |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "vLLM PR #59112",
   "source_url": "https://github.com/vllm-project/vllm/pull/59112",
   "tg_tokens": null,
   "tps": 7.015,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": null,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "4x AMD MI210",
   "id": "vllm-59151-mi210-triton-unfused-c1",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "single-stream decode with DFlash drafter, explicit --moe-backend triton_unfused (the backend auto-selection was missing on CDNA2), TP=4, 262K context config; needle-in-a-haystack retrieval verified at 1,038,700 tokens; config=triton-unfused-c1",
   "params": "309B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "4x AMD MI210 (gfx90a / CDNA2), 64 GB HBM2e each, PCIe, no Infinity Fabric; XiaomiMiMo/MiMo-V2.6-Flash-RL (mxfp4), TP=4, 262K context; with `--moe-backend triton_unfused` on that setup: 31 tok/s single-stream decode with a DFlash drafter, 160 tok/s aggregate at 32 concurrent streams, needle-in-a-haystack retrieval verified at 1,038,700 tokens",
   "ram_gb": 64,
   "retrieved": "2026-09-29",
   "source_name": "vLLM issue #59151",
   "source_url": "https://github.com/vllm-project/vllm/issues/59151",
   "tg_tokens": null,
   "tps": 31,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": null,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "4x AMD MI210",
   "id": "vllm-59151-mi210-triton-unfused-c32",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "aggregate decode across 32 concurrent streams, --moe-backend triton_unfused, TP=4, DFlash drafter; config=triton-unfused-c32",
   "params": "309B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "4x AMD MI210 (gfx90a / CDNA2), 64 GB HBM2e each, PCIe, no Infinity Fabric; XiaomiMiMo/MiMo-V2.6-Flash-RL (mxfp4), TP=4, 262K context; with `--moe-backend triton_unfused` on that setup: 31 tok/s single-stream decode with a DFlash drafter, 160 tok/s aggregate at 32 concurrent streams, needle-in-a-haystack retrieval verified at 1,038,700 tokens",
   "ram_gb": 64,
   "retrieved": "2026-09-29",
   "source_name": "vLLM issue #59151",
   "source_url": "https://github.com/vllm-project/vllm/issues/59151",
   "tg_tokens": null,
   "tps": 160,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "GTX 1060",
   "id": "lc-29635-gtx1060-master",
   "model": "Llama-3.2-1B",
   "notes": "llama-bench tg128 on the master build, regression reference for the prefix-LM mask PR; pp512 2381 t/s; means of the last two of three alternating runs; config=master",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Llama-3.2-1B Q8_0 on a GTX 1060: perplexity on wikitext-2 (40 chunks) is 14.3649 on both builds; llama-bench pp512 2380 vs 2381 t/s and tg128 90.3 vs 90.7 t/s (PR vs master, means of the last two of three alternating runs)",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "llama.cpp PR #29635",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29635",
   "tg_tokens": null,
   "tps": 90.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "GTX 1060",
   "id": "lc-29635-gtx1060-pr",
   "model": "Llama-3.2-1B",
   "notes": "llama-bench tg128 on the PR build (prefix-LM attention mask, no perf change expected); pp512 2380 t/s; means of the last two of three alternating runs; config=pr",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Llama-3.2-1B Q8_0 on a GTX 1060: perplexity on wikitext-2 (40 chunks) is 14.3649 on both builds; llama-bench pp512 2380 vs 2381 t/s and tg128 90.3 vs 90.7 t/s (PR vs master, means of the last two of three alternating runs)",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "llama.cpp PR #29635",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29635",
   "tg_tokens": null,
   "tps": 90.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 0,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "R9700",
   "id": "lc-29639-r9700-master-c0",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Vulkan sparse FA for quantized K/V disabled (master + #29591 baseline), depth 0, q8_0 KV cache; interleaved round; config=master-c0",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "q8_0 KV",
   "quote": "Qwen3.8-Flash-Next QSA, DeepSeek-style; q8_0 costs 61 us with sparse vs 86 us with it forced off on the R9700, and 73 vs 131 us on the 7900 XT; Flash-Next decode, q8_0 KV, 4 interleaved rounds. Baseline is master with #29591 cherry-picked, at the same commit:; | 0          | 51.4 t/s        | 51.6    | 0                             |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "llama.cpp PR #29639",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29639",
   "tg_tokens": null,
   "tps": 51.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 0,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "R9700",
   "id": "lc-29639-r9700-pr-c0",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Vulkan sparse FA enabled for quantized K/V, depth 0, q8_0 KV cache; interleaved round; config=pr-c0",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "q8_0 KV",
   "quote": "Qwen3.8-Flash-Next QSA, DeepSeek-style; q8_0 costs 61 us with sparse vs 86 us with it forced off on the R9700, and 73 vs 131 us on the 7900 XT; Flash-Next decode, q8_0 KV, 4 interleaved rounds. Baseline is master with #29591 cherry-picked, at the same commit:; | 0          | 51.4 t/s        | 51.6    | 0                             |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "llama.cpp PR #29639",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29639",
   "tg_tokens": null,
   "tps": 51.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32000,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "R9700",
   "id": "lc-29639-r9700-master-c32k",
   "model": "Qwen3.8-Flash-Next",
   "notes": "baseline, KV depth 32k; per-session spread +2.1% / -0.6%; config=master-c32k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "q8_0 KV",
   "quote": "Qwen3.8-Flash-Next QSA, DeepSeek-style; q8_0 costs 61 us with sparse vs 86 us with it forced off on the R9700, and 73 vs 131 us on the 7900 XT; Flash-Next decode, q8_0 KV, 4 interleaved rounds. Baseline is master with #29591 cherry-picked, at the same commit:; | 32k        | 38.9            | 39.7    | ~0 (+2.1% / -0.6% per session)|",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "llama.cpp PR #29639",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29639",
   "tg_tokens": null,
   "tps": 38.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32000,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "R9700",
   "id": "lc-29639-r9700-pr-c32k",
   "model": "Qwen3.8-Flash-Next",
   "notes": "PR build, KV depth 32k; per-session spread +2.1% / -0.6%; config=pr-c32k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "q8_0 KV",
   "quote": "Qwen3.8-Flash-Next QSA, DeepSeek-style; q8_0 costs 61 us with sparse vs 86 us with it forced off on the R9700, and 73 vs 131 us on the 7900 XT; Flash-Next decode, q8_0 KV, 4 interleaved rounds. Baseline is master with #29591 cherry-picked, at the same commit:; | 32k        | 38.9            | 39.7    | ~0 (+2.1% / -0.6% per session)|",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "llama.cpp PR #29639",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29639",
   "tg_tokens": null,
   "tps": 39.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 64000,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "R9700",
   "id": "lc-29639-r9700-master-c64k",
   "model": "Qwen3.8-Flash-Next",
   "notes": "baseline, KV depth 64k; config=master-c64k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "q8_0 KV",
   "quote": "Qwen3.8-Flash-Next QSA, DeepSeek-style; q8_0 costs 61 us with sparse vs 86 us with it forced off on the R9700, and 73 vs 131 us on the 7900 XT; Flash-Next decode, q8_0 KV, 4 interleaved rounds. Baseline is master with #29591 cherry-picked, at the same commit:; | 64k        | 27.9            | 32.3    | +15.6%                        |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "llama.cpp PR #29639",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29639",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 64000,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "R9700",
   "id": "lc-29639-r9700-pr-c64k",
   "model": "Qwen3.8-Flash-Next",
   "notes": "PR build, KV depth 64k, +15.6 pct over baseline, the largest gain in the table; config=pr-c64k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "q8_0 KV",
   "quote": "Qwen3.8-Flash-Next QSA, DeepSeek-style; q8_0 costs 61 us with sparse vs 86 us with it forced off on the R9700, and 73 vs 131 us on the 7900 XT; Flash-Next decode, q8_0 KV, 4 interleaved rounds. Baseline is master with #29591 cherry-picked, at the same commit:; | 64k        | 27.9            | 32.3    | +15.6%                        |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "llama.cpp PR #29639",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29639",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 128000,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "R9700",
   "id": "lc-29639-r9700-master-c128k",
   "model": "Qwen3.8-Flash-Next",
   "notes": "baseline, KV depth 128k; config=master-c128k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "q8_0 KV",
   "quote": "Qwen3.8-Flash-Next QSA, DeepSeek-style; q8_0 costs 61 us with sparse vs 86 us with it forced off on the R9700, and 73 vs 131 us on the 7900 XT; Flash-Next decode, q8_0 KV, 4 interleaved rounds. Baseline is master with #29591 cherry-picked, at the same commit:; | 128k       | 20.6            | 21.4    | +4.1%                         |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "llama.cpp PR #29639",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29639",
   "tg_tokens": null,
   "tps": 20.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 128000,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "R9700",
   "id": "lc-29639-r9700-pr-c128k",
   "model": "Qwen3.8-Flash-Next",
   "notes": "PR build, KV depth 128k, +4.1 pct; author attributes the smaller 128k gain to master's single-workgroup index compaction growing with context; config=pr-c128k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "q8_0 KV",
   "quote": "Qwen3.8-Flash-Next QSA, DeepSeek-style; q8_0 costs 61 us with sparse vs 86 us with it forced off on the R9700, and 73 vs 131 us on the 7900 XT; Flash-Next decode, q8_0 KV, 4 interleaved rounds. Baseline is master with #29591 cherry-picked, at the same commit:; | 128k       | 20.6            | 21.4    | +4.1%                         |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "llama.cpp PR #29639",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29639",
   "tg_tokens": null,
   "tps": 21.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 120000,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "RTX 3090 + RX 7900 XT",
   "id": "lc-29654-3090-7900xt-pcie",
   "model": "Qwen3.8-27B",
   "notes": "PCIe bandwidth exhaustion on a two-GPU rig (RTX 3090 on PCIe4 x16, RX 7900 XT on PCIe4 x1): during the episode decode drops below 1 t/s (upper-bound value, not a benchmark), 39 GiB/s TX+RX on the x16 lanes; VK device lost before the bigger timeout; model + context fully in VRAM; 120K+ context scenario; first noticed since Qwen3.8-flash-next support, possibly coincidental; unified context, cache ram, mtp/dflash2 combinations tried without success",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Qwen3.8 27B Q8_0; RTX 3090  + RX 7900 XT; When the situation occurs the decode speed drops drastically down to less than 1 t/s.; The most common scenario is restoring big context (120K+) after reboot",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "llama.cpp issue #29654",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29654",
   "tg_tokens": null,
   "tps": 1,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 4,
   "ctx": 60000,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59280-mi355x-60k-gluon-c4",
   "model": "GLM 5.2",
   "notes": "FlyDSL prefill MQA-logits kernel for the sparse indexer (gfx950, opt-in VLLM_ROCM_USE_AITER_FLYDSL_MQA_LOGITS); TP4 MI355X, per-GPU throughput, ISL 60000 / OSL 600; Gluon baseline; TTFT 2866.7 ms; config=gluon-c4-60k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "GLM-5.2-MXFP4, TP4, MI355X.; ISL 60000 / OSL 600; | 4 | 3372.9 | 3419.9 | +1.4% | 2866.7 | 2708.7 | \u22125.5% |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "vLLM PR #59280",
   "source_url": "https://github.com/vllm-project/vllm/pull/59280",
   "tg_tokens": null,
   "tps": 3372.9,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 4,
   "ctx": 60000,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59280-mi355x-60k-flydsl-c4",
   "model": "GLM 5.2",
   "notes": "FlyDSL prefill MQA-logits kernel for the sparse indexer (gfx950, opt-in VLLM_ROCM_USE_AITER_FLYDSL_MQA_LOGITS); TP4 MI355X, per-GPU throughput, ISL 60000 / OSL 600; FlyDSL; TTFT 2708.7 ms; config=flydsl-c4-60k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "GLM-5.2-MXFP4, TP4, MI355X.; ISL 60000 / OSL 600; | 4 | 3372.9 | 3419.9 | +1.4% | 2866.7 | 2708.7 | \u22125.5% |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "vLLM PR #59280",
   "source_url": "https://github.com/vllm-project/vllm/pull/59280",
   "tg_tokens": null,
   "tps": 3419.9,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 8,
   "ctx": 60000,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59280-mi355x-60k-gluon-c8",
   "model": "GLM 5.2",
   "notes": "FlyDSL prefill MQA-logits kernel for the sparse indexer (gfx950, opt-in VLLM_ROCM_USE_AITER_FLYDSL_MQA_LOGITS); TP4 MI355X, per-GPU throughput, ISL 60000 / OSL 600; Gluon baseline; TTFT 3589.9 ms; config=gluon-c8-60k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "GLM-5.2-MXFP4, TP4, MI355X.; ISL 60000 / OSL 600; | 8 | 4250.0 | 4297.1 | +1.1% | 3589.9 | 3559.9 | \u22120.8% |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "vLLM PR #59280",
   "source_url": "https://github.com/vllm-project/vllm/pull/59280",
   "tg_tokens": null,
   "tps": 4250.0,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 8,
   "ctx": 60000,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59280-mi355x-60k-flydsl-c8",
   "model": "GLM 5.2",
   "notes": "FlyDSL prefill MQA-logits kernel for the sparse indexer (gfx950, opt-in VLLM_ROCM_USE_AITER_FLYDSL_MQA_LOGITS); TP4 MI355X, per-GPU throughput, ISL 60000 / OSL 600; FlyDSL; TTFT 3559.9 ms; config=flydsl-c8-60k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "GLM-5.2-MXFP4, TP4, MI355X.; ISL 60000 / OSL 600; | 8 | 4250.0 | 4297.1 | +1.1% | 3589.9 | 3559.9 | \u22120.8% |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "vLLM PR #59280",
   "source_url": "https://github.com/vllm-project/vllm/pull/59280",
   "tg_tokens": null,
   "tps": 4297.1,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 16,
   "ctx": 60000,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59280-mi355x-60k-gluon-c16",
   "model": "GLM 5.2",
   "notes": "FlyDSL prefill MQA-logits kernel for the sparse indexer (gfx950, opt-in VLLM_ROCM_USE_AITER_FLYDSL_MQA_LOGITS); TP4 MI355X, per-GPU throughput, ISL 60000 / OSL 600; Gluon baseline; TTFT 5697.8 ms; config=gluon-c16-60k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "GLM-5.2-MXFP4, TP4, MI355X.; ISL 60000 / OSL 600; | 16 | 4890.4 | 4964.6 | +1.5% | 5697.8 | 5505.0 | \u22123.4% |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "vLLM PR #59280",
   "source_url": "https://github.com/vllm-project/vllm/pull/59280",
   "tg_tokens": null,
   "tps": 4890.4,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 16,
   "ctx": 60000,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59280-mi355x-60k-flydsl-c16",
   "model": "GLM 5.2",
   "notes": "FlyDSL prefill MQA-logits kernel for the sparse indexer (gfx950, opt-in VLLM_ROCM_USE_AITER_FLYDSL_MQA_LOGITS); TP4 MI355X, per-GPU throughput, ISL 60000 / OSL 600; FlyDSL; TTFT 5505.0 ms; config=flydsl-c16-60k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "GLM-5.2-MXFP4, TP4, MI355X.; ISL 60000 / OSL 600; | 16 | 4890.4 | 4964.6 | +1.5% | 5697.8 | 5505.0 | \u22123.4% |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "vLLM PR #59280",
   "source_url": "https://github.com/vllm-project/vllm/pull/59280",
   "tg_tokens": null,
   "tps": 4964.6,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 32,
   "ctx": 60000,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59280-mi355x-60k-gluon-c32",
   "model": "GLM 5.2",
   "notes": "FlyDSL prefill MQA-logits kernel for the sparse indexer (gfx950, opt-in VLLM_ROCM_USE_AITER_FLYDSL_MQA_LOGITS); TP4 MI355X, per-GPU throughput, ISL 60000 / OSL 600; Gluon baseline; TTFT 13285.8 ms; config=gluon-c32-60k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "GLM-5.2-MXFP4, TP4, MI355X.; ISL 60000 / OSL 600; | 32 | 5315.0 | 5399.4 | +1.6% | 13285.8 | 12848.0 | \u22123.3% |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "vLLM PR #59280",
   "source_url": "https://github.com/vllm-project/vllm/pull/59280",
   "tg_tokens": null,
   "tps": 5315.0,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 32,
   "ctx": 60000,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59280-mi355x-60k-flydsl-c32",
   "model": "GLM 5.2",
   "notes": "FlyDSL prefill MQA-logits kernel for the sparse indexer (gfx950, opt-in VLLM_ROCM_USE_AITER_FLYDSL_MQA_LOGITS); TP4 MI355X, per-GPU throughput, ISL 60000 / OSL 600; FlyDSL; TTFT 12848.0 ms; config=flydsl-c32-60k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "GLM-5.2-MXFP4, TP4, MI355X.; ISL 60000 / OSL 600; | 32 | 5315.0 | 5399.4 | +1.6% | 13285.8 | 12848.0 | \u22123.3% |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "vLLM PR #59280",
   "source_url": "https://github.com/vllm-project/vllm/pull/59280",
   "tg_tokens": null,
   "tps": 5399.4,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 4,
   "ctx": 8192,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59280-mi355x-8k-gluon-c4",
   "model": "GLM 5.2",
   "notes": "FlyDSL prefill MQA-logits kernel for the sparse indexer (gfx950, opt-in VLLM_ROCM_USE_AITER_FLYDSL_MQA_LOGITS); TP4 MI355X, per-GPU throughput, ISL 8192 / OSL 1024; Gluon baseline; config=gluon-c4-8k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "GLM-5.2-MXFP4, TP4, MI355X.; ISL 8192 / OSL 1024; | 4 | 645.1 | 646.2 | +0.2% | 156.6 | 155.5 | \u22120.7% |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "vLLM PR #59280",
   "source_url": "https://github.com/vllm-project/vllm/pull/59280",
   "tg_tokens": null,
   "tps": 645.1,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 4,
   "ctx": 8192,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59280-mi355x-8k-flydsl-c4",
   "model": "GLM 5.2",
   "notes": "FlyDSL prefill MQA-logits kernel for the sparse indexer (gfx950, opt-in VLLM_ROCM_USE_AITER_FLYDSL_MQA_LOGITS); TP4 MI355X, per-GPU throughput, ISL 8192 / OSL 1024; FlyDSL; config=flydsl-c4-8k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "GLM-5.2-MXFP4, TP4, MI355X.; ISL 8192 / OSL 1024; | 4 | 645.1 | 646.2 | +0.2% | 156.6 | 155.5 | \u22120.7% |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "vLLM PR #59280",
   "source_url": "https://github.com/vllm-project/vllm/pull/59280",
   "tg_tokens": null,
   "tps": 646.2,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 4,
   "ctx": 1024,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59280-mi355x-1k-gluon-c4",
   "model": "GLM 5.2",
   "notes": "FlyDSL prefill MQA-logits kernel for the sparse indexer (gfx950, opt-in VLLM_ROCM_USE_AITER_FLYDSL_MQA_LOGITS); TP4 MI355X, per-GPU throughput, ISL 1024 / OSL 1024; Gluon baseline; config=gluon-c4-1k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "GLM-5.2-MXFP4, TP4, MI355X.; ISL 1024 / OSL 1024; | 4 | 645.1 | 646.2 | +0.2% | 156.6 | 155.5 | \u22120.7% |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "vLLM PR #59280",
   "source_url": "https://github.com/vllm-project/vllm/pull/59280",
   "tg_tokens": null,
   "tps": 156.6,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 4,
   "ctx": 1024,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59280-mi355x-1k-flydsl-c4",
   "model": "GLM 5.2",
   "notes": "FlyDSL prefill MQA-logits kernel for the sparse indexer (gfx950, opt-in VLLM_ROCM_USE_AITER_FLYDSL_MQA_LOGITS); TP4 MI355X, per-GPU throughput, ISL 1024 / OSL 1024; FlyDSL; config=flydsl-c4-1k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "GLM-5.2-MXFP4, TP4, MI355X.; ISL 1024 / OSL 1024; | 4 | 645.1 | 646.2 | +0.2% | 156.6 | 155.5 | \u22120.7% |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "vLLM PR #59280",
   "source_url": "https://github.com/vllm-project/vllm/pull/59280",
   "tg_tokens": null,
   "tps": 155.5,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 8,
   "ctx": 8192,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59280-mi355x-8k-gluon-c8",
   "model": "GLM 5.2",
   "notes": "FlyDSL prefill MQA-logits kernel for the sparse indexer (gfx950, opt-in VLLM_ROCM_USE_AITER_FLYDSL_MQA_LOGITS); TP4 MI355X, per-GPU throughput, ISL 8192 / OSL 1024; Gluon baseline; config=gluon-c8-8k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "GLM-5.2-MXFP4, TP4, MI355X.; ISL 8192 / OSL 1024; | 8 | 1058.7 | 1053.9 | \u22120.4% | 270.4 | 267.4 | \u22121.1% |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "vLLM PR #59280",
   "source_url": "https://github.com/vllm-project/vllm/pull/59280",
   "tg_tokens": null,
   "tps": 1058.7,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 8,
   "ctx": 8192,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59280-mi355x-8k-flydsl-c8",
   "model": "GLM 5.2",
   "notes": "FlyDSL prefill MQA-logits kernel for the sparse indexer (gfx950, opt-in VLLM_ROCM_USE_AITER_FLYDSL_MQA_LOGITS); TP4 MI355X, per-GPU throughput, ISL 8192 / OSL 1024; FlyDSL; config=flydsl-c8-8k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "GLM-5.2-MXFP4, TP4, MI355X.; ISL 8192 / OSL 1024; | 8 | 1058.7 | 1053.9 | \u22120.4% | 270.4 | 267.4 | \u22121.1% |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "vLLM PR #59280",
   "source_url": "https://github.com/vllm-project/vllm/pull/59280",
   "tg_tokens": null,
   "tps": 1053.9,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 8,
   "ctx": 1024,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59280-mi355x-1k-gluon-c8",
   "model": "GLM 5.2",
   "notes": "FlyDSL prefill MQA-logits kernel for the sparse indexer (gfx950, opt-in VLLM_ROCM_USE_AITER_FLYDSL_MQA_LOGITS); TP4 MI355X, per-GPU throughput, ISL 1024 / OSL 1024; Gluon baseline; config=gluon-c8-1k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "GLM-5.2-MXFP4, TP4, MI355X.; ISL 1024 / OSL 1024; | 8 | 1058.7 | 1053.9 | \u22120.4% | 270.4 | 267.4 | \u22121.1% |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "vLLM PR #59280",
   "source_url": "https://github.com/vllm-project/vllm/pull/59280",
   "tg_tokens": null,
   "tps": 270.4,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 8,
   "ctx": 1024,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59280-mi355x-1k-flydsl-c8",
   "model": "GLM 5.2",
   "notes": "FlyDSL prefill MQA-logits kernel for the sparse indexer (gfx950, opt-in VLLM_ROCM_USE_AITER_FLYDSL_MQA_LOGITS); TP4 MI355X, per-GPU throughput, ISL 1024 / OSL 1024; FlyDSL; config=flydsl-c8-1k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "GLM-5.2-MXFP4, TP4, MI355X.; ISL 1024 / OSL 1024; | 8 | 1058.7 | 1053.9 | \u22120.4% | 270.4 | 267.4 | \u22121.1% |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "vLLM PR #59280",
   "source_url": "https://github.com/vllm-project/vllm/pull/59280",
   "tg_tokens": null,
   "tps": 267.4,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 16,
   "ctx": 8192,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59280-mi355x-8k-gluon-c16",
   "model": "GLM 5.2",
   "notes": "FlyDSL prefill MQA-logits kernel for the sparse indexer (gfx950, opt-in VLLM_ROCM_USE_AITER_FLYDSL_MQA_LOGITS); TP4 MI355X, per-GPU throughput, ISL 8192 / OSL 1024; Gluon baseline; config=gluon-c16-8k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "GLM-5.2-MXFP4, TP4, MI355X.; ISL 8192 / OSL 1024; | 16 | 1593.2 | 1586.5 | \u22120.4% | 436.9 | 434.1 | \u22120.6% |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "vLLM PR #59280",
   "source_url": "https://github.com/vllm-project/vllm/pull/59280",
   "tg_tokens": null,
   "tps": 1593.2,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 16,
   "ctx": 8192,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59280-mi355x-8k-flydsl-c16",
   "model": "GLM 5.2",
   "notes": "FlyDSL prefill MQA-logits kernel for the sparse indexer (gfx950, opt-in VLLM_ROCM_USE_AITER_FLYDSL_MQA_LOGITS); TP4 MI355X, per-GPU throughput, ISL 8192 / OSL 1024; FlyDSL; config=flydsl-c16-8k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "GLM-5.2-MXFP4, TP4, MI355X.; ISL 8192 / OSL 1024; | 16 | 1593.2 | 1586.5 | \u22120.4% | 436.9 | 434.1 | \u22120.6% |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "vLLM PR #59280",
   "source_url": "https://github.com/vllm-project/vllm/pull/59280",
   "tg_tokens": null,
   "tps": 1586.5,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 16,
   "ctx": 1024,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59280-mi355x-1k-gluon-c16",
   "model": "GLM 5.2",
   "notes": "FlyDSL prefill MQA-logits kernel for the sparse indexer (gfx950, opt-in VLLM_ROCM_USE_AITER_FLYDSL_MQA_LOGITS); TP4 MI355X, per-GPU throughput, ISL 1024 / OSL 1024; Gluon baseline; config=gluon-c16-1k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "GLM-5.2-MXFP4, TP4, MI355X.; ISL 1024 / OSL 1024; | 16 | 1593.2 | 1586.5 | \u22120.4% | 436.9 | 434.1 | \u22120.6% |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "vLLM PR #59280",
   "source_url": "https://github.com/vllm-project/vllm/pull/59280",
   "tg_tokens": null,
   "tps": 436.9,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 16,
   "ctx": 1024,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59280-mi355x-1k-flydsl-c16",
   "model": "GLM 5.2",
   "notes": "FlyDSL prefill MQA-logits kernel for the sparse indexer (gfx950, opt-in VLLM_ROCM_USE_AITER_FLYDSL_MQA_LOGITS); TP4 MI355X, per-GPU throughput, ISL 1024 / OSL 1024; FlyDSL; config=flydsl-c16-1k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "GLM-5.2-MXFP4, TP4, MI355X.; ISL 1024 / OSL 1024; | 16 | 1593.2 | 1586.5 | \u22120.4% | 436.9 | 434.1 | \u22120.6% |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "vLLM PR #59280",
   "source_url": "https://github.com/vllm-project/vllm/pull/59280",
   "tg_tokens": null,
   "tps": 434.1,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 32,
   "ctx": 8192,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59280-mi355x-8k-gluon-c32",
   "model": "GLM 5.2",
   "notes": "FlyDSL prefill MQA-logits kernel for the sparse indexer (gfx950, opt-in VLLM_ROCM_USE_AITER_FLYDSL_MQA_LOGITS); TP4 MI355X, per-GPU throughput, ISL 8192 / OSL 1024; Gluon baseline; config=gluon-c32-8k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "GLM-5.2-MXFP4, TP4, MI355X.; ISL 8192 / OSL 1024; | 32 | 2248.1 | 2232.8 | \u22120.7% | 679.3 | 674.7 | \u22120.7% |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "vLLM PR #59280",
   "source_url": "https://github.com/vllm-project/vllm/pull/59280",
   "tg_tokens": null,
   "tps": 2248.1,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 32,
   "ctx": 8192,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59280-mi355x-8k-flydsl-c32",
   "model": "GLM 5.2",
   "notes": "FlyDSL prefill MQA-logits kernel for the sparse indexer (gfx950, opt-in VLLM_ROCM_USE_AITER_FLYDSL_MQA_LOGITS); TP4 MI355X, per-GPU throughput, ISL 8192 / OSL 1024; FlyDSL; config=flydsl-c32-8k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "GLM-5.2-MXFP4, TP4, MI355X.; ISL 8192 / OSL 1024; | 32 | 2248.1 | 2232.8 | \u22120.7% | 679.3 | 674.7 | \u22120.7% |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "vLLM PR #59280",
   "source_url": "https://github.com/vllm-project/vllm/pull/59280",
   "tg_tokens": null,
   "tps": 2232.8,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 32,
   "ctx": 1024,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59280-mi355x-1k-gluon-c32",
   "model": "GLM 5.2",
   "notes": "FlyDSL prefill MQA-logits kernel for the sparse indexer (gfx950, opt-in VLLM_ROCM_USE_AITER_FLYDSL_MQA_LOGITS); TP4 MI355X, per-GPU throughput, ISL 1024 / OSL 1024; Gluon baseline; config=gluon-c32-1k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "GLM-5.2-MXFP4, TP4, MI355X.; ISL 1024 / OSL 1024; | 32 | 2248.1 | 2232.8 | \u22120.7% | 679.3 | 674.7 | \u22120.7% |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "vLLM PR #59280",
   "source_url": "https://github.com/vllm-project/vllm/pull/59280",
   "tg_tokens": null,
   "tps": 679.3,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 32,
   "ctx": 1024,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59280-mi355x-1k-flydsl-c32",
   "model": "GLM 5.2",
   "notes": "FlyDSL prefill MQA-logits kernel for the sparse indexer (gfx950, opt-in VLLM_ROCM_USE_AITER_FLYDSL_MQA_LOGITS); TP4 MI355X, per-GPU throughput, ISL 1024 / OSL 1024; FlyDSL; config=flydsl-c32-1k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "GLM-5.2-MXFP4, TP4, MI355X.; ISL 1024 / OSL 1024; | 32 | 2248.1 | 2232.8 | \u22120.7% | 679.3 | 674.7 | \u22120.7% |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "vLLM PR #59280",
   "source_url": "https://github.com/vllm-project/vllm/pull/59280",
   "tg_tokens": null,
   "tps": 674.7,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 64,
   "ctx": 8192,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59280-mi355x-8k-gluon-c64",
   "model": "GLM 5.2",
   "notes": "FlyDSL prefill MQA-logits kernel for the sparse indexer (gfx950, opt-in VLLM_ROCM_USE_AITER_FLYDSL_MQA_LOGITS); TP4 MI355X, per-GPU throughput, ISL 8192 / OSL 1024; Gluon baseline; config=gluon-c64-8k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "GLM-5.2-MXFP4, TP4, MI355X.; ISL 8192 / OSL 1024; | 64 | 3118.3 | 3106.5 | \u22120.4% | 1060.7 | 1051.9 | \u22120.8% |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "vLLM PR #59280",
   "source_url": "https://github.com/vllm-project/vllm/pull/59280",
   "tg_tokens": null,
   "tps": 3118.3,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 64,
   "ctx": 8192,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59280-mi355x-8k-flydsl-c64",
   "model": "GLM 5.2",
   "notes": "FlyDSL prefill MQA-logits kernel for the sparse indexer (gfx950, opt-in VLLM_ROCM_USE_AITER_FLYDSL_MQA_LOGITS); TP4 MI355X, per-GPU throughput, ISL 8192 / OSL 1024; FlyDSL; config=flydsl-c64-8k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "GLM-5.2-MXFP4, TP4, MI355X.; ISL 8192 / OSL 1024; | 64 | 3118.3 | 3106.5 | \u22120.4% | 1060.7 | 1051.9 | \u22120.8% |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "vLLM PR #59280",
   "source_url": "https://github.com/vllm-project/vllm/pull/59280",
   "tg_tokens": null,
   "tps": 3106.5,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 64,
   "ctx": 1024,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59280-mi355x-1k-gluon-c64",
   "model": "GLM 5.2",
   "notes": "FlyDSL prefill MQA-logits kernel for the sparse indexer (gfx950, opt-in VLLM_ROCM_USE_AITER_FLYDSL_MQA_LOGITS); TP4 MI355X, per-GPU throughput, ISL 1024 / OSL 1024; Gluon baseline; config=gluon-c64-1k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "GLM-5.2-MXFP4, TP4, MI355X.; ISL 1024 / OSL 1024; | 64 | 3118.3 | 3106.5 | \u22120.4% | 1060.7 | 1051.9 | \u22120.8% |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "vLLM PR #59280",
   "source_url": "https://github.com/vllm-project/vllm/pull/59280",
   "tg_tokens": null,
   "tps": 1060.7,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 64,
   "ctx": 1024,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59280-mi355x-1k-flydsl-c64",
   "model": "GLM 5.2",
   "notes": "FlyDSL prefill MQA-logits kernel for the sparse indexer (gfx950, opt-in VLLM_ROCM_USE_AITER_FLYDSL_MQA_LOGITS); TP4 MI355X, per-GPU throughput, ISL 1024 / OSL 1024; FlyDSL; config=flydsl-c64-1k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "GLM-5.2-MXFP4, TP4, MI355X.; ISL 1024 / OSL 1024; | 64 | 3118.3 | 3106.5 | \u22120.4% | 1060.7 | 1051.9 | \u22120.8% |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "vLLM PR #59280",
   "source_url": "https://github.com/vllm-project/vllm/pull/59280",
   "tg_tokens": null,
   "tps": 1051.9,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 128,
   "ctx": 8192,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59280-mi355x-8k-gluon-c128",
   "model": "GLM 5.2",
   "notes": "FlyDSL prefill MQA-logits kernel for the sparse indexer (gfx950, opt-in VLLM_ROCM_USE_AITER_FLYDSL_MQA_LOGITS); TP4 MI355X, per-GPU throughput, ISL 8192 / OSL 1024; Gluon baseline; config=gluon-c128-8k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "GLM-5.2-MXFP4, TP4, MI355X.; ISL 8192 / OSL 1024; | 128 | 4060.3 | 4042.8 | \u22120.4% | 1637.7 | 1624.8 | \u22120.8% |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "vLLM PR #59280",
   "source_url": "https://github.com/vllm-project/vllm/pull/59280",
   "tg_tokens": null,
   "tps": 4060.3,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 128,
   "ctx": 8192,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59280-mi355x-8k-flydsl-c128",
   "model": "GLM 5.2",
   "notes": "FlyDSL prefill MQA-logits kernel for the sparse indexer (gfx950, opt-in VLLM_ROCM_USE_AITER_FLYDSL_MQA_LOGITS); TP4 MI355X, per-GPU throughput, ISL 8192 / OSL 1024; FlyDSL; config=flydsl-c128-8k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "GLM-5.2-MXFP4, TP4, MI355X.; ISL 8192 / OSL 1024; | 128 | 4060.3 | 4042.8 | \u22120.4% | 1637.7 | 1624.8 | \u22120.8% |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "vLLM PR #59280",
   "source_url": "https://github.com/vllm-project/vllm/pull/59280",
   "tg_tokens": null,
   "tps": 4042.8,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 128,
   "ctx": 1024,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59280-mi355x-1k-gluon-c128",
   "model": "GLM 5.2",
   "notes": "FlyDSL prefill MQA-logits kernel for the sparse indexer (gfx950, opt-in VLLM_ROCM_USE_AITER_FLYDSL_MQA_LOGITS); TP4 MI355X, per-GPU throughput, ISL 1024 / OSL 1024; Gluon baseline; config=gluon-c128-1k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "GLM-5.2-MXFP4, TP4, MI355X.; ISL 1024 / OSL 1024; | 128 | 4060.3 | 4042.8 | \u22120.4% | 1637.7 | 1624.8 | \u22120.8% |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "vLLM PR #59280",
   "source_url": "https://github.com/vllm-project/vllm/pull/59280",
   "tg_tokens": null,
   "tps": 1637.7,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 128,
   "ctx": 1024,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59280-mi355x-1k-flydsl-c128",
   "model": "GLM 5.2",
   "notes": "FlyDSL prefill MQA-logits kernel for the sparse indexer (gfx950, opt-in VLLM_ROCM_USE_AITER_FLYDSL_MQA_LOGITS); TP4 MI355X, per-GPU throughput, ISL 1024 / OSL 1024; FlyDSL; config=flydsl-c128-1k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "GLM-5.2-MXFP4, TP4, MI355X.; ISL 1024 / OSL 1024; | 128 | 4060.3 | 4042.8 | \u22120.4% | 1637.7 | 1624.8 | \u22120.8% |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "vLLM PR #59280",
   "source_url": "https://github.com/vllm-project/vllm/pull/59280",
   "tg_tokens": null,
   "tps": 1624.8,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 256,
   "ctx": 8192,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59280-mi355x-8k-gluon-c256",
   "model": "GLM 5.2",
   "notes": "FlyDSL prefill MQA-logits kernel for the sparse indexer (gfx950, opt-in VLLM_ROCM_USE_AITER_FLYDSL_MQA_LOGITS); TP4 MI355X, per-GPU throughput, ISL 8192 / OSL 1024; Gluon baseline; config=gluon-c256-8k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "GLM-5.2-MXFP4, TP4, MI355X.; ISL 8192 / OSL 1024; | 256 | 4925.5 | 4907.5 | \u22120.4% | 2424.7 | 2414.3 | \u22120.4% |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "vLLM PR #59280",
   "source_url": "https://github.com/vllm-project/vllm/pull/59280",
   "tg_tokens": null,
   "tps": 4925.5,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 256,
   "ctx": 8192,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59280-mi355x-8k-flydsl-c256",
   "model": "GLM 5.2",
   "notes": "FlyDSL prefill MQA-logits kernel for the sparse indexer (gfx950, opt-in VLLM_ROCM_USE_AITER_FLYDSL_MQA_LOGITS); TP4 MI355X, per-GPU throughput, ISL 8192 / OSL 1024; FlyDSL; config=flydsl-c256-8k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "GLM-5.2-MXFP4, TP4, MI355X.; ISL 8192 / OSL 1024; | 256 | 4925.5 | 4907.5 | \u22120.4% | 2424.7 | 2414.3 | \u22120.4% |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "vLLM PR #59280",
   "source_url": "https://github.com/vllm-project/vllm/pull/59280",
   "tg_tokens": null,
   "tps": 4907.5,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 256,
   "ctx": 1024,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59280-mi355x-1k-gluon-c256",
   "model": "GLM 5.2",
   "notes": "FlyDSL prefill MQA-logits kernel for the sparse indexer (gfx950, opt-in VLLM_ROCM_USE_AITER_FLYDSL_MQA_LOGITS); TP4 MI355X, per-GPU throughput, ISL 1024 / OSL 1024; Gluon baseline; config=gluon-c256-1k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "GLM-5.2-MXFP4, TP4, MI355X.; ISL 1024 / OSL 1024; | 256 | 4925.5 | 4907.5 | \u22120.4% | 2424.7 | 2414.3 | \u22120.4% |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "vLLM PR #59280",
   "source_url": "https://github.com/vllm-project/vllm/pull/59280",
   "tg_tokens": null,
   "tps": 2424.7,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 256,
   "ctx": 1024,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59280-mi355x-1k-flydsl-c256",
   "model": "GLM 5.2",
   "notes": "FlyDSL prefill MQA-logits kernel for the sparse indexer (gfx950, opt-in VLLM_ROCM_USE_AITER_FLYDSL_MQA_LOGITS); TP4 MI355X, per-GPU throughput, ISL 1024 / OSL 1024; FlyDSL; config=flydsl-c256-1k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "GLM-5.2-MXFP4, TP4, MI355X.; ISL 1024 / OSL 1024; | 256 | 4925.5 | 4907.5 | \u22120.4% | 2424.7 | 2414.3 | \u22120.4% |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "vLLM PR #59280",
   "source_url": "https://github.com/vllm-project/vllm/pull/59280",
   "tg_tokens": null,
   "tps": 2414.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "RX 7900 XTX",
   "id": "lc-29679-7900xtx-control-master",
   "model": "ThinkingCap-Qwen3.8-27B",
   "notes": "Vulkan MMVQ 4-row policy kept only at 8 columns on RDNA3 proprietary driver; Windows 11, AMD proprietary driver 26.8.1 (LLPC), master da89bb3; llama-server 64k ctx, KV q4_0, MTP draft, t/s per client, average of 4 runs; 1 slot, draft n_max 2 (control); arm=master (config=control",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_S",
   "quote": "Setup: RX 7900 XTX, Windows 11, AMD proprietary driver 26.8.1 (LLPC), master `da89bb3`.; `llama-bench -p N -n 0`, ms per batch (ThinkingCap-Qwen3.8-27B):; `llama-server`, Q4_K_S, 64k ctx, KV `q4_0`, MTP draft, t/s per client (average of 4 runs):; | 1 slot, draft `n_max` 2 (control) | 81.5 | 78.9 |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "llama.cpp PR #29679",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29679",
   "tg_tokens": null,
   "tps": 81.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "RX 7900 XTX",
   "id": "lc-29679-7900xtx-control-pr",
   "model": "ThinkingCap-Qwen3.8-27B",
   "notes": "Vulkan MMVQ 4-row policy kept only at 8 columns on RDNA3 proprietary driver; Windows 11, AMD proprietary driver 26.8.1 (LLPC), master da89bb3; llama-server 64k ctx, KV q4_0, MTP draft, t/s per client, average of 4 runs; 1 slot, draft n_max 2 (control); arm=pr (config=control",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_S",
   "quote": "Setup: RX 7900 XTX, Windows 11, AMD proprietary driver 26.8.1 (LLPC), master `da89bb3`.; `llama-bench -p N -n 0`, ms per batch (ThinkingCap-Qwen3.8-27B):; `llama-server`, Q4_K_S, 64k ctx, KV `q4_0`, MTP draft, t/s per client (average of 4 runs):; | 1 slot, draft `n_max` 2 (control) | 81.5 | 78.9 |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "llama.cpp PR #29679",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29679",
   "tg_tokens": null,
   "tps": 78.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 2,
   "ctx": 65536,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "RX 7900 XTX",
   "id": "lc-29679-7900xtx-2slot-master",
   "model": "ThinkingCap-Qwen3.8-27B",
   "notes": "Vulkan MMVQ 4-row policy kept only at 8 columns on RDNA3 proprietary driver; Windows 11, AMD proprietary driver 26.8.1 (LLPC), master da89bb3; llama-server 64k ctx, KV q4_0, MTP draft, t/s per client, average of 4 runs; 2 slots, n_max 2, 2 clients; arm=master (config=master)",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_S",
   "quote": "Setup: RX 7900 XTX, Windows 11, AMD proprietary driver 26.8.1 (LLPC), master `da89bb3`.; `llama-bench -p N -n 0`, ms per batch (ThinkingCap-Qwen3.8-27B):; `llama-server`, Q4_K_S, 64k ctx, KV `q4_0`, MTP draft, t/s per client (average of 4 runs):; | 2 slots, `n_max` 2, 2 clients | 15.5 | 45.6 |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "llama.cpp PR #29679",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29679",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 2,
   "ctx": 65536,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "RX 7900 XTX",
   "id": "lc-29679-7900xtx-2slot-pr",
   "model": "ThinkingCap-Qwen3.8-27B",
   "notes": "Vulkan MMVQ 4-row policy kept only at 8 columns on RDNA3 proprietary driver; Windows 11, AMD proprietary driver 26.8.1 (LLPC), master da89bb3; llama-server 64k ctx, KV q4_0, MTP draft, t/s per client, average of 4 runs; 2 slots, n_max 2, 2 clients; arm=pr (config=pr)",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_S",
   "quote": "Setup: RX 7900 XTX, Windows 11, AMD proprietary driver 26.8.1 (LLPC), master `da89bb3`.; `llama-bench -p N -n 0`, ms per batch (ThinkingCap-Qwen3.8-27B):; `llama-server`, Q4_K_S, 64k ctx, KV `q4_0`, MTP draft, t/s per client (average of 4 runs):; | 2 slots, `n_max` 2, 2 clients | 15.5 | 45.6 |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "llama.cpp PR #29679",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29679",
   "tg_tokens": null,
   "tps": 45.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 3,
   "ctx": 65536,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "RX 7900 XTX",
   "id": "lc-29679-7900xtx-3slot-master",
   "model": "ThinkingCap-Qwen3.8-27B",
   "notes": "Vulkan MMVQ 4-row policy kept only at 8 columns on RDNA3 proprietary driver; Windows 11, AMD proprietary driver 26.8.1 (LLPC), master da89bb3; llama-server 64k ctx, KV q4_0, MTP draft, t/s per client, average of 4 runs; 3 slots, n_max 1, 3 clients; arm=master (config=master)",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_S",
   "quote": "Setup: RX 7900 XTX, Windows 11, AMD proprietary driver 26.8.1 (LLPC), master `da89bb3`.; `llama-bench -p N -n 0`, ms per batch (ThinkingCap-Qwen3.8-27B):; `llama-server`, Q4_K_S, 64k ctx, KV `q4_0`, MTP draft, t/s per client (average of 4 runs):; | 3 slots, `n_max` 1, 3 clients | 10.6 | 34.2 |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "llama.cpp PR #29679",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29679",
   "tg_tokens": null,
   "tps": 10.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 3,
   "ctx": 65536,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "RX 7900 XTX",
   "id": "lc-29679-7900xtx-3slot-pr",
   "model": "ThinkingCap-Qwen3.8-27B",
   "notes": "Vulkan MMVQ 4-row policy kept only at 8 columns on RDNA3 proprietary driver; Windows 11, AMD proprietary driver 26.8.1 (LLPC), master da89bb3; llama-server 64k ctx, KV q4_0, MTP draft, t/s per client, average of 4 runs; 3 slots, n_max 1, 3 clients; arm=pr (config=pr)",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_S",
   "quote": "Setup: RX 7900 XTX, Windows 11, AMD proprietary driver 26.8.1 (LLPC), master `da89bb3`.; `llama-bench -p N -n 0`, ms per batch (ThinkingCap-Qwen3.8-27B):; `llama-server`, Q4_K_S, 64k ctx, KV `q4_0`, MTP draft, t/s per client (average of 4 runs):; | 3 slots, `n_max` 1, 3 clients | 10.6 | 34.2 |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "llama.cpp PR #29679",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29679",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-09-29",
   "flags": [
    "contradiction"
   ],
   "hardware": "RX 7900 XTX",
   "id": "lc-29679-7900xtx-nmax5-master",
   "model": "ThinkingCap-Qwen3.8-27B",
   "notes": "Vulkan MMVQ 4-row policy kept only at 8 columns on RDNA3 proprietary driver; Windows 11, AMD proprietary driver 26.8.1 (LLPC), master da89bb3; llama-server 64k ctx, KV q4_0, MTP draft, t/s per client, average of 4 runs; 1 slot, n_max 5; arm=master (config=nmax5",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_S",
   "quote": "Setup: RX 7900 XTX, Windows 11, AMD proprietary driver 26.8.1 (LLPC), master `da89bb3`.; `llama-bench -p N -n 0`, ms per batch (ThinkingCap-Qwen3.8-27B):; `llama-server`, Q4_K_S, 64k ctx, KV `q4_0`, MTP draft, t/s per client (average of 4 runs):; | 1 slot, `n_max` 5 | 26.7 | 92.4 |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "llama.cpp PR #29679",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29679",
   "tg_tokens": null,
   "tps": 26.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-09-29",
   "flags": [],
   "hardware": "RX 7900 XTX",
   "id": "lc-29679-7900xtx-nmax5-pr",
   "model": "ThinkingCap-Qwen3.8-27B",
   "notes": "Vulkan MMVQ 4-row policy kept only at 8 columns on RDNA3 proprietary driver; Windows 11, AMD proprietary driver 26.8.1 (LLPC), master da89bb3; llama-server 64k ctx, KV q4_0, MTP draft, t/s per client, average of 4 runs; 1 slot, n_max 5; arm=pr (config=nmax5",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_S",
   "quote": "Setup: RX 7900 XTX, Windows 11, AMD proprietary driver 26.8.1 (LLPC), master `da89bb3`.; `llama-bench -p N -n 0`, ms per batch (ThinkingCap-Qwen3.8-27B):; `llama-server`, Q4_K_S, 64k ctx, KV `q4_0`, MTP draft, t/s per client (average of 4 runs):; | 1 slot, `n_max` 5 | 26.7 | 92.4 |",
   "ram_gb": null,
   "retrieved": "2026-09-29",
   "source_name": "llama.cpp PR #29679",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29679",
   "tg_tokens": null,
   "tps": 92.4,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 128,
   "ctx": 10000,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "H200",
   "id": "vllm-59355-h200-oob-before",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "Spec-decode residual-mass OOB fix, before arm: unpatched master build crashes at conc=128 (2/2 runs, 131/320 requests completed, 189 Internal Server Errors, cudaErrorIllegalAddress); 3972 tok/s is the partial-run output throughput; dspark spec config, 5 speculative tokens, block verification, adaptive verification; TP2 x DP4 + EP, vLLM 0.30.0 container; config=oob-before",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "Repro environment: DeepSeek-V4.1-Flash, 8x H200 (SM90), vLLM 0.30.0,; random workload (10000 in / 5000 out), `--max-concurrency 128`.; | output throughput | 3972 (partial) | 4031.5 tok/s |",
   "ram_gb": null,
   "retrieved": "2026-09-30",
   "source_name": "vLLM PR #59355",
   "source_url": "https://github.com/vllm-project/vllm/pull/59355",
   "tg_tokens": null,
   "tps": 3972,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 128,
   "ctx": 10000,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "H200",
   "id": "vllm-59355-h200-oob-after",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "Spec-decode residual-mass OOB fix, after arm: 12-line guard (tl.num_programs(0) bound on the draft_sampled+1 read), 1/1 clean, 320/320 requests, 0 errors; output throughput 4031.5 tok/s vs 3972 partial before; mean spec acceptance length 2.399 (n=54) vs 2.145 (n=23, early death); same deployment and workload, only the patch differs; config=oob-after",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "Repro environment: DeepSeek-V4.1-Flash, 8x H200 (SM90), vLLM 0.30.0,; random workload (10000 in / 5000 out), `--max-concurrency 128`.; | output throughput | 3972 (partial) | 4031.5 tok/s |",
   "ram_gb": null,
   "retrieved": "2026-09-30",
   "source_name": "vLLM PR #59355",
   "source_url": "https://github.com/vllm-project/vllm/pull/59355",
   "tg_tokens": null,
   "tps": 4031.5,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 8,
   "ctx": 32768,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "4x H200 NVL",
   "id": "vllm-59367-h200nvl-bothfix-pp-c32k",
   "model": "GLM-5.3",
   "notes": "GLM-5.3-NVFP4, 4x H200 NVL, TP4 DCP4 EP, MTP 3, fp8_ds_mla, max-model-len 786432, nightly af7f9488c; DCP top-k merge scratch reserved through the workspace manager plus the CUDA-graph pool estimate fix; fresh prefill 8 sequences x 32768, patched (both fixes), 65.7 s; config=bothfix-pp-c32k",
   "params": null,
   "power_w": null,
   "pp_tps": 5092,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Serving: GLM-5.3-NVFP4, 4x H200 NVL, TP4 DCP4 EP, MTP 3, `fp8_ds_mla`, `max-model-len 786432`, `max-num-batched-tokens 8192`, `max-num-seqs 32`, `gpu-memory-utilization 0.945`, on nightly af7f948 with this PR and the fix for #59318.; Fresh prefill 8x32768: 65.7 s, 5092 tok/s (5129 tok/s on the unpatched production image the same night); 2x131072: 69.7 s, 4941 tok/s (4988 before).; Decode N=1 110.7, N=8 329.8, N=32 761.1 tok/s.; fresh prefill 8x32768 in 63.0 s (5103 tok/s), 2x131072 in 72.6 s, decode N=1 119.1, N=8 279.2, N=32 823.9 tok/s, no errors.",
   "ram_gb": null,
   "retrieved": "2026-09-30",
   "source_name": "vLLM PR #59367",
   "source_url": "https://github.com/vllm-project/vllm/pull/59367",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 8,
   "ctx": 32768,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "4x H200 NVL",
   "id": "vllm-59367-h200nvl-unpatched-pp-c32k",
   "model": "GLM-5.3",
   "notes": "GLM-5.3-NVFP4, 4x H200 NVL, TP4 DCP4 EP, MTP 3, fp8_ds_mla, max-model-len 786432, nightly af7f9488c; DCP top-k merge scratch reserved through the workspace manager plus the CUDA-graph pool estimate fix; fresh prefill 8 x 32768 on the unpatched production image the same night; config=unpatched-pp-c32k",
   "params": null,
   "power_w": null,
   "pp_tps": 5129,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Serving: GLM-5.3-NVFP4, 4x H200 NVL, TP4 DCP4 EP, MTP 3, `fp8_ds_mla`, `max-model-len 786432`, `max-num-batched-tokens 8192`, `max-num-seqs 32`, `gpu-memory-utilization 0.945`, on nightly af7f948 with this PR and the fix for #59318.; Fresh prefill 8x32768: 65.7 s, 5092 tok/s (5129 tok/s on the unpatched production image the same night); 2x131072: 69.7 s, 4941 tok/s (4988 before).; Decode N=1 110.7, N=8 329.8, N=32 761.1 tok/s.; fresh prefill 8x32768 in 63.0 s (5103 tok/s), 2x131072 in 72.6 s, decode N=1 119.1, N=8 279.2, N=32 823.9 tok/s, no errors.",
   "ram_gb": null,
   "retrieved": "2026-09-30",
   "source_name": "vLLM PR #59367",
   "source_url": "https://github.com/vllm-project/vllm/pull/59367",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 2,
   "ctx": 131072,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "4x H200 NVL",
   "id": "vllm-59367-h200nvl-bothfix-pp-c131k",
   "model": "GLM-5.3",
   "notes": "GLM-5.3-NVFP4, 4x H200 NVL, TP4 DCP4 EP, MTP 3, fp8_ds_mla, max-model-len 786432, nightly af7f9488c; DCP top-k merge scratch reserved through the workspace manager plus the CUDA-graph pool estimate fix; fresh prefill 2 x 131072, patched (both fixes), 69.7 s; config=bothfix-pp-c131k",
   "params": null,
   "power_w": null,
   "pp_tps": 4941,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Serving: GLM-5.3-NVFP4, 4x H200 NVL, TP4 DCP4 EP, MTP 3, `fp8_ds_mla`, `max-model-len 786432`, `max-num-batched-tokens 8192`, `max-num-seqs 32`, `gpu-memory-utilization 0.945`, on nightly af7f948 with this PR and the fix for #59318.; Fresh prefill 8x32768: 65.7 s, 5092 tok/s (5129 tok/s on the unpatched production image the same night); 2x131072: 69.7 s, 4941 tok/s (4988 before).; Decode N=1 110.7, N=8 329.8, N=32 761.1 tok/s.; fresh prefill 8x32768 in 63.0 s (5103 tok/s), 2x131072 in 72.6 s, decode N=1 119.1, N=8 279.2, N=32 823.9 tok/s, no errors.",
   "ram_gb": null,
   "retrieved": "2026-09-30",
   "source_name": "vLLM PR #59367",
   "source_url": "https://github.com/vllm-project/vllm/pull/59367",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 2,
   "ctx": 131072,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "4x H200 NVL",
   "id": "vllm-59367-h200nvl-unpatched-pp-c131k",
   "model": "GLM-5.3",
   "notes": "GLM-5.3-NVFP4, 4x H200 NVL, TP4 DCP4 EP, MTP 3, fp8_ds_mla, max-model-len 786432, nightly af7f9488c; DCP top-k merge scratch reserved through the workspace manager plus the CUDA-graph pool estimate fix; fresh prefill 2 x 131072 on the unpatched production image; config=unpatched-pp-c131k",
   "params": null,
   "power_w": null,
   "pp_tps": 4988,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Serving: GLM-5.3-NVFP4, 4x H200 NVL, TP4 DCP4 EP, MTP 3, `fp8_ds_mla`, `max-model-len 786432`, `max-num-batched-tokens 8192`, `max-num-seqs 32`, `gpu-memory-utilization 0.945`, on nightly af7f948 with this PR and the fix for #59318.; Fresh prefill 8x32768: 65.7 s, 5092 tok/s (5129 tok/s on the unpatched production image the same night); 2x131072: 69.7 s, 4941 tok/s (4988 before).; Decode N=1 110.7, N=8 329.8, N=32 761.1 tok/s.; fresh prefill 8x32768 in 63.0 s (5103 tok/s), 2x131072 in 72.6 s, decode N=1 119.1, N=8 279.2, N=32 823.9 tok/s, no errors.",
   "ram_gb": null,
   "retrieved": "2026-09-30",
   "source_name": "vLLM PR #59367",
   "source_url": "https://github.com/vllm-project/vllm/pull/59367",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "4x H200 NVL",
   "id": "vllm-59367-h200nvl-bothfix-decode-n1",
   "model": "GLM-5.3",
   "notes": "GLM-5.3-NVFP4, 4x H200 NVL, TP4 DCP4 EP, MTP 3, fp8_ds_mla, max-model-len 786432, nightly af7f9488c; DCP top-k merge scratch reserved through the workspace manager plus the CUDA-graph pool estimate fix; decode N=1, patched, single sample; config=bothfix-decode-n1",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Serving: GLM-5.3-NVFP4, 4x H200 NVL, TP4 DCP4 EP, MTP 3, `fp8_ds_mla`, `max-model-len 786432`, `max-num-batched-tokens 8192`, `max-num-seqs 32`, `gpu-memory-utilization 0.945`, on nightly af7f948 with this PR and the fix for #59318.; Fresh prefill 8x32768: 65.7 s, 5092 tok/s (5129 tok/s on the unpatched production image the same night); 2x131072: 69.7 s, 4941 tok/s (4988 before).; Decode N=1 110.7, N=8 329.8, N=32 761.1 tok/s.; fresh prefill 8x32768 in 63.0 s (5103 tok/s), 2x131072 in 72.6 s, decode N=1 119.1, N=8 279.2, N=32 823.9 tok/s, no errors.",
   "ram_gb": null,
   "retrieved": "2026-09-30",
   "source_name": "vLLM PR #59367",
   "source_url": "https://github.com/vllm-project/vllm/pull/59367",
   "tg_tokens": null,
   "tps": 110.7,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 8,
   "ctx": null,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "4x H200 NVL",
   "id": "vllm-59367-h200nvl-bothfix-decode-n8",
   "model": "GLM-5.3",
   "notes": "GLM-5.3-NVFP4, 4x H200 NVL, TP4 DCP4 EP, MTP 3, fp8_ds_mla, max-model-len 786432, nightly af7f9488c; DCP top-k merge scratch reserved through the workspace manager plus the CUDA-graph pool estimate fix; decode N=8, patched, single sample; unpatched image ranged 263 to 317 across seven boots, so no decode change is claimed; config=bothfix-decode-n8",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Serving: GLM-5.3-NVFP4, 4x H200 NVL, TP4 DCP4 EP, MTP 3, `fp8_ds_mla`, `max-model-len 786432`, `max-num-batched-tokens 8192`, `max-num-seqs 32`, `gpu-memory-utilization 0.945`, on nightly af7f948 with this PR and the fix for #59318.; Fresh prefill 8x32768: 65.7 s, 5092 tok/s (5129 tok/s on the unpatched production image the same night); 2x131072: 69.7 s, 4941 tok/s (4988 before).; Decode N=1 110.7, N=8 329.8, N=32 761.1 tok/s.; fresh prefill 8x32768 in 63.0 s (5103 tok/s), 2x131072 in 72.6 s, decode N=1 119.1, N=8 279.2, N=32 823.9 tok/s, no errors.",
   "ram_gb": null,
   "retrieved": "2026-09-30",
   "source_name": "vLLM PR #59367",
   "source_url": "https://github.com/vllm-project/vllm/pull/59367",
   "tg_tokens": null,
   "tps": 329.8,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 32,
   "ctx": null,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "4x H200 NVL",
   "id": "vllm-59367-h200nvl-bothfix-decode-n32",
   "model": "GLM-5.3",
   "notes": "GLM-5.3-NVFP4, 4x H200 NVL, TP4 DCP4 EP, MTP 3, fp8_ds_mla, max-model-len 786432, nightly af7f9488c; DCP top-k merge scratch reserved through the workspace manager plus the CUDA-graph pool estimate fix; decode N=32, patched, single sample; unpatched image ranged 701 to 838 across seven boots; config=bothfix-decode-n32",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Serving: GLM-5.3-NVFP4, 4x H200 NVL, TP4 DCP4 EP, MTP 3, `fp8_ds_mla`, `max-model-len 786432`, `max-num-batched-tokens 8192`, `max-num-seqs 32`, `gpu-memory-utilization 0.945`, on nightly af7f948 with this PR and the fix for #59318.; Fresh prefill 8x32768: 65.7 s, 5092 tok/s (5129 tok/s on the unpatched production image the same night); 2x131072: 69.7 s, 4941 tok/s (4988 before).; Decode N=1 110.7, N=8 329.8, N=32 761.1 tok/s.; fresh prefill 8x32768 in 63.0 s (5103 tok/s), 2x131072 in 72.6 s, decode N=1 119.1, N=8 279.2, N=32 823.9 tok/s, no errors.",
   "ram_gb": null,
   "retrieved": "2026-09-30",
   "source_name": "vLLM PR #59367",
   "source_url": "https://github.com/vllm-project/vllm/pull/59367",
   "tg_tokens": null,
   "tps": 761.1,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 8,
   "ctx": 32768,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "4x H200 NVL",
   "id": "vllm-59367-h200nvl-est0-pp-c32k",
   "model": "GLM-5.3",
   "notes": "GLM-5.3-NVFP4, 4x H200 NVL, TP4 DCP4 EP, MTP 3, fp8_ds_mla, max-model-len 786432, nightly af7f9488c; DCP top-k merge scratch reserved through the workspace manager plus the CUDA-graph pool estimate fix; second boot with VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS=0 and only the #59317 merge-scratch fix (the combination that OOMed before), fresh prefill 8 x 32768 in 63.0 s; config=est0-pp-c32k",
   "params": null,
   "power_w": null,
   "pp_tps": 5103,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Serving: GLM-5.3-NVFP4, 4x H200 NVL, TP4 DCP4 EP, MTP 3, `fp8_ds_mla`, `max-model-len 786432`, `max-num-batched-tokens 8192`, `max-num-seqs 32`, `gpu-memory-utilization 0.945`, on nightly af7f948 with this PR and the fix for #59318.; Fresh prefill 8x32768: 65.7 s, 5092 tok/s (5129 tok/s on the unpatched production image the same night); 2x131072: 69.7 s, 4941 tok/s (4988 before).; Decode N=1 110.7, N=8 329.8, N=32 761.1 tok/s.; fresh prefill 8x32768 in 63.0 s (5103 tok/s), 2x131072 in 72.6 s, decode N=1 119.1, N=8 279.2, N=32 823.9 tok/s, no errors.",
   "ram_gb": null,
   "retrieved": "2026-09-30",
   "source_name": "vLLM PR #59367",
   "source_url": "https://github.com/vllm-project/vllm/pull/59367",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 1,
   "ctx": null,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "4x H200 NVL",
   "id": "vllm-59367-h200nvl-est0-decode-n1",
   "model": "GLM-5.3",
   "notes": "GLM-5.3-NVFP4, 4x H200 NVL, TP4 DCP4 EP, MTP 3, fp8_ds_mla, max-model-len 786432, nightly af7f9488c; DCP top-k merge scratch reserved through the workspace manager plus the CUDA-graph pool estimate fix; second boot (ESTIMATE_CUDAGRAPHS=0, merge-scratch fix only), decode N=1; config=est0-decode-n1",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Serving: GLM-5.3-NVFP4, 4x H200 NVL, TP4 DCP4 EP, MTP 3, `fp8_ds_mla`, `max-model-len 786432`, `max-num-batched-tokens 8192`, `max-num-seqs 32`, `gpu-memory-utilization 0.945`, on nightly af7f948 with this PR and the fix for #59318.; Fresh prefill 8x32768: 65.7 s, 5092 tok/s (5129 tok/s on the unpatched production image the same night); 2x131072: 69.7 s, 4941 tok/s (4988 before).; Decode N=1 110.7, N=8 329.8, N=32 761.1 tok/s.; fresh prefill 8x32768 in 63.0 s (5103 tok/s), 2x131072 in 72.6 s, decode N=1 119.1, N=8 279.2, N=32 823.9 tok/s, no errors.",
   "ram_gb": null,
   "retrieved": "2026-09-30",
   "source_name": "vLLM PR #59367",
   "source_url": "https://github.com/vllm-project/vllm/pull/59367",
   "tg_tokens": null,
   "tps": 119.1,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 8,
   "ctx": null,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "4x H200 NVL",
   "id": "vllm-59367-h200nvl-est0-decode-n8",
   "model": "GLM-5.3",
   "notes": "GLM-5.3-NVFP4, 4x H200 NVL, TP4 DCP4 EP, MTP 3, fp8_ds_mla, max-model-len 786432, nightly af7f9488c; DCP top-k merge scratch reserved through the workspace manager plus the CUDA-graph pool estimate fix; second boot (ESTIMATE_CUDAGRAPHS=0, merge-scratch fix only), decode N=8; config=est0-decode-n8",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Serving: GLM-5.3-NVFP4, 4x H200 NVL, TP4 DCP4 EP, MTP 3, `fp8_ds_mla`, `max-model-len 786432`, `max-num-batched-tokens 8192`, `max-num-seqs 32`, `gpu-memory-utilization 0.945`, on nightly af7f948 with this PR and the fix for #59318.; Fresh prefill 8x32768: 65.7 s, 5092 tok/s (5129 tok/s on the unpatched production image the same night); 2x131072: 69.7 s, 4941 tok/s (4988 before).; Decode N=1 110.7, N=8 329.8, N=32 761.1 tok/s.; fresh prefill 8x32768 in 63.0 s (5103 tok/s), 2x131072 in 72.6 s, decode N=1 119.1, N=8 279.2, N=32 823.9 tok/s, no errors.",
   "ram_gb": null,
   "retrieved": "2026-09-30",
   "source_name": "vLLM PR #59367",
   "source_url": "https://github.com/vllm-project/vllm/pull/59367",
   "tg_tokens": null,
   "tps": 279.2,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 32,
   "ctx": null,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "4x H200 NVL",
   "id": "vllm-59367-h200nvl-est0-decode-n32",
   "model": "GLM-5.3",
   "notes": "GLM-5.3-NVFP4, 4x H200 NVL, TP4 DCP4 EP, MTP 3, fp8_ds_mla, max-model-len 786432, nightly af7f9488c; DCP top-k merge scratch reserved through the workspace manager plus the CUDA-graph pool estimate fix; second boot (ESTIMATE_CUDAGRAPHS=0, merge-scratch fix only), decode N=32; config=est0-decode-n32",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Serving: GLM-5.3-NVFP4, 4x H200 NVL, TP4 DCP4 EP, MTP 3, `fp8_ds_mla`, `max-model-len 786432`, `max-num-batched-tokens 8192`, `max-num-seqs 32`, `gpu-memory-utilization 0.945`, on nightly af7f948 with this PR and the fix for #59318.; Fresh prefill 8x32768: 65.7 s, 5092 tok/s (5129 tok/s on the unpatched production image the same night); 2x131072: 69.7 s, 4941 tok/s (4988 before).; Decode N=1 110.7, N=8 329.8, N=32 761.1 tok/s.; fresh prefill 8x32768 in 63.0 s (5103 tok/s), 2x131072 in 72.6 s, decode N=1 119.1, N=8 279.2, N=32 823.9 tok/s, no errors.",
   "ram_gb": null,
   "retrieved": "2026-09-30",
   "source_name": "vLLM PR #59367",
   "source_url": "https://github.com/vllm-project/vllm/pull/59367",
   "tg_tokens": null,
   "tps": 823.9,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 8,
   "ctx": 128,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "AMD Zen 5 (Turin, 32 cores pinned to one NUMA node)",
   "id": "vllm-59434-zen5-before-s8",
   "model": "gpt-oss 20B",
   "notes": "gpt-oss 20B BF16-w8a8 (amd/gpt-oss-20b-BF16-w8a8), experts fused for the in-tree loader, vllm bench throughput random 128/128 256 prompts, total tok/s; ZenCPUExpertsInt8 scale dtype fix, f32 scales reach the ZenDNN fast routed MoE path (bf16 fell back to the generic per-expert path); config=bf16scales-s8; bf16 scales (before)",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16-w8a8",
   "quote": "Throughput measured on AMD Zen 5 (Turin, 32 cores pinned to one NUMA node), `amd/gpt-oss-20b-BF16-w8a8` with the experts fused for the in-tree loader, `vllm bench throughput` with the random dataset at 128/128, 256 prompts:; | 8 | 154.23 total tok/s | 158.21 total tok/s |; | 32 | 369.79 total tok/s | 375.74 total tok/s |",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59434",
   "source_url": "https://github.com/vllm-project/vllm/pull/59434",
   "tg_tokens": null,
   "tps": 154.23,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 8,
   "ctx": 128,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "AMD Zen 5 (Turin, 32 cores pinned to one NUMA node)",
   "id": "vllm-59434-zen5-after-s8",
   "model": "gpt-oss 20B",
   "notes": "gpt-oss 20B BF16-w8a8 (amd/gpt-oss-20b-BF16-w8a8), experts fused for the in-tree loader, vllm bench throughput random 128/128 256 prompts, total tok/s; ZenCPUExpertsInt8 scale dtype fix, f32 scales reach the ZenDNN fast routed MoE path (bf16 fell back to the generic per-expert path); config=f32scales-s8; f32 scales (after)",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16-w8a8",
   "quote": "Throughput measured on AMD Zen 5 (Turin, 32 cores pinned to one NUMA node), `amd/gpt-oss-20b-BF16-w8a8` with the experts fused for the in-tree loader, `vllm bench throughput` with the random dataset at 128/128, 256 prompts:; | 8 | 154.23 total tok/s | 158.21 total tok/s |; | 32 | 369.79 total tok/s | 375.74 total tok/s |",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59434",
   "source_url": "https://github.com/vllm-project/vllm/pull/59434",
   "tg_tokens": null,
   "tps": 158.21,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 32,
   "ctx": 128,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "AMD Zen 5 (Turin, 32 cores pinned to one NUMA node)",
   "id": "vllm-59434-zen5-before-s32",
   "model": "gpt-oss 20B",
   "notes": "gpt-oss 20B BF16-w8a8 (amd/gpt-oss-20b-BF16-w8a8), experts fused for the in-tree loader, vllm bench throughput random 128/128 256 prompts, total tok/s; ZenCPUExpertsInt8 scale dtype fix, f32 scales reach the ZenDNN fast routed MoE path (bf16 fell back to the generic per-expert path); config=bf16scales-s32; bf16 scales (before)",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16-w8a8",
   "quote": "Throughput measured on AMD Zen 5 (Turin, 32 cores pinned to one NUMA node), `amd/gpt-oss-20b-BF16-w8a8` with the experts fused for the in-tree loader, `vllm bench throughput` with the random dataset at 128/128, 256 prompts:; | 8 | 154.23 total tok/s | 158.21 total tok/s |; | 32 | 369.79 total tok/s | 375.74 total tok/s |",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59434",
   "source_url": "https://github.com/vllm-project/vllm/pull/59434",
   "tg_tokens": null,
   "tps": 369.79,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 32,
   "ctx": 128,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "AMD Zen 5 (Turin, 32 cores pinned to one NUMA node)",
   "id": "vllm-59434-zen5-after-s32",
   "model": "gpt-oss 20B",
   "notes": "gpt-oss 20B BF16-w8a8 (amd/gpt-oss-20b-BF16-w8a8), experts fused for the in-tree loader, vllm bench throughput random 128/128 256 prompts, total tok/s; ZenCPUExpertsInt8 scale dtype fix, f32 scales reach the ZenDNN fast routed MoE path (bf16 fell back to the generic per-expert path); config=f32scales-s32; f32 scales (after)",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16-w8a8",
   "quote": "Throughput measured on AMD Zen 5 (Turin, 32 cores pinned to one NUMA node), `amd/gpt-oss-20b-BF16-w8a8` with the experts fused for the in-tree loader, `vllm bench throughput` with the random dataset at 128/128, 256 prompts:; | 8 | 154.23 total tok/s | 158.21 total tok/s |; | 32 | 369.79 total tok/s | 375.74 total tok/s |",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59434",
   "source_url": "https://github.com/vllm-project/vllm/pull/59434",
   "tg_tokens": null,
   "tps": 375.74,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 16,
   "ctx": 512,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "RTX 3070",
   "id": "vllm-59489-rtx3070-rust",
   "model": "Qwen3-0.6B",
   "notes": "Qwen3-0.6B (Qwen/Qwen3-0.6B, --reasoning-parser qwen3) on RTX 3070; vllm bench serve --backend openai-responses, random 512 in / 128 out, 200 prompts, concurrency 16, temperature 0, seed 42; output throughput over three runs; Rust vllm-bench (VLLM_USE_RUST_BENCH=1), 1676-1680 tok/s, mean TTFT 217-222 ms, mean TPOT 7.60-7.63 ms, median ITL 7.32-7.34 ms, 200/200; config=rust-responses",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "RTX 3070, `vllm serve Qwen/Qwen3-0.6B --reasoning-parser qwen3`. Python `vllm bench serve` and Rust `vllm-bench`, both with `--backend openai-responses`, alternating after one warmup, three runs each: random dataset, 512 in, 128 out, 200 prompts, concurrency 16, temperature 0, seed 42.; | Output throughput (tok/s) | 1676-1680 | 1671-1673 |; With this branch, the `VLLM_USE_RUST_BENCH=1` command above completes 200/200 at 1676 tok/s, 219 ms mean TTFT, 7.64 ms TPOT and 7.33 ms median ITL.",
   "ram_gb": null,
   "retrieved": "2026-09-30",
   "source_name": "vLLM PR #59489",
   "source_url": "https://github.com/vllm-project/vllm/pull/59489",
   "tg_tokens": null,
   "tps": 1676,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 16,
   "ctx": 512,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "RTX 3070",
   "id": "vllm-59489-rtx3070-python",
   "model": "Qwen3-0.6B",
   "notes": "Qwen3-0.6B (Qwen/Qwen3-0.6B, --reasoning-parser qwen3) on RTX 3070; vllm bench serve --backend openai-responses, random 512 in / 128 out, 200 prompts, concurrency 16, temperature 0, seed 42; output throughput over three runs; Python vllm bench serve, 1671-1673 tok/s, mean TTFT 219 ms, mean TPOT 7.65-7.66 ms, median ITL 7.34-7.35 ms, 200/200; config=python-responses",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "RTX 3070, `vllm serve Qwen/Qwen3-0.6B --reasoning-parser qwen3`. Python `vllm bench serve` and Rust `vllm-bench`, both with `--backend openai-responses`, alternating after one warmup, three runs each: random dataset, 512 in, 128 out, 200 prompts, concurrency 16, temperature 0, seed 42.; | Output throughput (tok/s) | 1676-1680 | 1671-1673 |; With this branch, the `VLLM_USE_RUST_BENCH=1` command above completes 200/200 at 1676 tok/s, 219 ms mean TTFT, 7.64 ms TPOT and 7.33 ms median ITL.",
   "ram_gb": null,
   "retrieved": "2026-09-30",
   "source_name": "vLLM PR #59489",
   "source_url": "https://github.com/vllm-project/vllm/pull/59489",
   "tg_tokens": null,
   "tps": 1671,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 64,
   "ctx": 512,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "H200",
   "id": "vllm-59488-h200-fp8-dec-marlin",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "8x H200 (SM90), DeepSeek-V4.1-Flash MXFP4 experts, per-token FP8 activations, TP8+EP, --moe-backend humming, offline LLM runs, 2 repeats, decode 64 prompts x 512 in / 256 out, total output tok/s; Marlin MoE GEMM, today's default on Hopper; config=fp8-marlin",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "For reference, with FP8 at decode, grouped is 16% slower than Marlin (2,222 vs 2,648 output tok/s) and indexed is 14% faster.; Throughput: offline `LLM` runs of DeepSeek-V4.1-Flash on 8x H200, TP8+EP, `--moe-backend humming`, 2 repeats.; Decode is 64 random 512-token prompts with 256 output tokens.; The rule applies with BF16 activations and with per-token FP8 activations",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59488 (draft)",
   "source_url": "https://github.com/vllm-project/vllm/pull/59488",
   "tg_tokens": 256,
   "tps": 2648,
   "ttft_s": null
  },
  {
   "backend": "vllm",
   "batch": 64,
   "ctx": 512,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "H200",
   "id": "vllm-59488-h200-fp8-dec-humming",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "8x H200 (SM90), DeepSeek-V4.1-Flash MXFP4 experts, per-token FP8 activations, TP8+EP, --moe-backend humming, offline LLM runs, 2 repeats, decode 64 prompts x 512 in / 256 out, total output tok/s; Humming grouped MoE GEMM, 16% behind Marlin; the PR's indexed GEMM is 14% faster than this (no absolute published); config=fp8-humming-grouped",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "For reference, with FP8 at decode, grouped is 16% slower than Marlin (2,222 vs 2,648 output tok/s) and indexed is 14% faster.; Throughput: offline `LLM` runs of DeepSeek-V4.1-Flash on 8x H200, TP8+EP, `--moe-backend humming`, 2 repeats.; Decode is 64 random 512-token prompts with 256 output tokens.; The rule applies with BF16 activations and with per-token FP8 activations",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59488 (draft)",
   "source_url": "https://github.com/vllm-project/vllm/pull/59488",
   "tg_tokens": 256,
   "tps": 2222,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "lc-29820-r9700-npl1-stock",
   "model": "Qwopus3.8-27B-Flash-V2",
   "notes": "Qwopus3.8-27B-Flash-V2 Q4_K_M (qwen35 arch), Radeon AI PRO R9700 (gfx1201), ROCm 7.2.4, llama-batched-bench -c 32768 -npp 512 -ntg 128 -npl 1 -dev ROCm0, -ngl 99 -fa on -ctk q8_0 -ctv q4_0, TG t/s; stock build, MMVQ_MAX_BATCH_SIZE=8; config=mmvq-max8",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "On the Radeon AI PRO R9700 (gfx1201), the HIP backend sends quantized matmuls with 5 to 8 columns to MMVQ, which seems to be up to 46% slower than MMQ at those sizes. Lowering `MMVQ_MAX_BATCH_SIZE` from 8 to 4 fixes it.; | GPU | AMD Radeon AI PRO R9700, gfx1201 |; | ROCm | 7.2.4 |; | Target model | `Jackrong/Qwopus3.8-27B-Flash-V2-GGUF:Q4_K_M`, qwen35 architecture |; | Common flags | `-ngl 99 -fa on -ctk q8_0 -ctv q4_0` |; llama-batched-bench -hf Jackrong/Qwopus3.8-27B-Flash-V2-GGUF:Q4_K_M -ngl 99 -fa on -ctk q8_0 -ctv q4_0 -c 32768 -npp 512 -ntg 128 -npl 1,2,4,5,6,8,16 -dev ROCm0; | Sequences | Stock TG (t/s) | Patched TG (t/s) | Change | Stock step (ms) | Patched step (ms) |; | 1 | 29.5 | 29.4 | \u2014 | 34 | 34 |",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp #29820",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29820",
   "tg_tokens": 128,
   "tps": "29.5",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "lc-29820-r9700-npl1-patched",
   "model": "Qwopus3.8-27B-Flash-V2",
   "notes": "Qwopus3.8-27B-Flash-V2 Q4_K_M (qwen35 arch), Radeon AI PRO R9700 (gfx1201), ROCm 7.2.4, llama-batched-bench -c 32768 -npp 512 -ntg 128 -npl 1 -dev ROCm0, -ngl 99 -fa on -ctk q8_0 -ctv q4_0, TG t/s; patched build, MMVQ_MAX_BATCH_SIZE=4; config=mmvq-max4",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "On the Radeon AI PRO R9700 (gfx1201), the HIP backend sends quantized matmuls with 5 to 8 columns to MMVQ, which seems to be up to 46% slower than MMQ at those sizes. Lowering `MMVQ_MAX_BATCH_SIZE` from 8 to 4 fixes it.; | GPU | AMD Radeon AI PRO R9700, gfx1201 |; | ROCm | 7.2.4 |; | Target model | `Jackrong/Qwopus3.8-27B-Flash-V2-GGUF:Q4_K_M`, qwen35 architecture |; | Common flags | `-ngl 99 -fa on -ctk q8_0 -ctv q4_0` |; llama-batched-bench -hf Jackrong/Qwopus3.8-27B-Flash-V2-GGUF:Q4_K_M -ngl 99 -fa on -ctk q8_0 -ctv q4_0 -c 32768 -npp 512 -ntg 128 -npl 1,2,4,5,6,8,16 -dev ROCm0; | Sequences | Stock TG (t/s) | Patched TG (t/s) | Change | Stock step (ms) | Patched step (ms) |; | 1 | 29.5 | 29.4 | \u2014 | 34 | 34 |",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp #29820",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29820",
   "tg_tokens": 128,
   "tps": "29.4",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 2,
   "ctx": 32768,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "lc-29820-r9700-npl2-stock",
   "model": "Qwopus3.8-27B-Flash-V2",
   "notes": "Qwopus3.8-27B-Flash-V2 Q4_K_M (qwen35 arch), Radeon AI PRO R9700 (gfx1201), ROCm 7.2.4, llama-batched-bench -c 32768 -npp 512 -ntg 128 -npl 2 -dev ROCm0, -ngl 99 -fa on -ctk q8_0 -ctv q4_0, TG t/s; stock build, MMVQ_MAX_BATCH_SIZE=8; config=mmvq-max8",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "On the Radeon AI PRO R9700 (gfx1201), the HIP backend sends quantized matmuls with 5 to 8 columns to MMVQ, which seems to be up to 46% slower than MMQ at those sizes. Lowering `MMVQ_MAX_BATCH_SIZE` from 8 to 4 fixes it.; | GPU | AMD Radeon AI PRO R9700, gfx1201 |; | ROCm | 7.2.4 |; | Target model | `Jackrong/Qwopus3.8-27B-Flash-V2-GGUF:Q4_K_M`, qwen35 architecture |; | Common flags | `-ngl 99 -fa on -ctk q8_0 -ctv q4_0` |; llama-batched-bench -hf Jackrong/Qwopus3.8-27B-Flash-V2-GGUF:Q4_K_M -ngl 99 -fa on -ctk q8_0 -ctv q4_0 -c 32768 -npp 512 -ntg 128 -npl 1,2,4,5,6,8,16 -dev ROCm0; | Sequences | Stock TG (t/s) | Patched TG (t/s) | Change | Stock step (ms) | Patched step (ms) |; | 2 | 50.9 | 50.6 | \u2014 | 39 | 40 |",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp #29820",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29820",
   "tg_tokens": 128,
   "tps": "50.9",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 2,
   "ctx": 32768,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "lc-29820-r9700-npl2-patched",
   "model": "Qwopus3.8-27B-Flash-V2",
   "notes": "Qwopus3.8-27B-Flash-V2 Q4_K_M (qwen35 arch), Radeon AI PRO R9700 (gfx1201), ROCm 7.2.4, llama-batched-bench -c 32768 -npp 512 -ntg 128 -npl 2 -dev ROCm0, -ngl 99 -fa on -ctk q8_0 -ctv q4_0, TG t/s; patched build, MMVQ_MAX_BATCH_SIZE=4; config=mmvq-max4",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "On the Radeon AI PRO R9700 (gfx1201), the HIP backend sends quantized matmuls with 5 to 8 columns to MMVQ, which seems to be up to 46% slower than MMQ at those sizes. Lowering `MMVQ_MAX_BATCH_SIZE` from 8 to 4 fixes it.; | GPU | AMD Radeon AI PRO R9700, gfx1201 |; | ROCm | 7.2.4 |; | Target model | `Jackrong/Qwopus3.8-27B-Flash-V2-GGUF:Q4_K_M`, qwen35 architecture |; | Common flags | `-ngl 99 -fa on -ctk q8_0 -ctv q4_0` |; llama-batched-bench -hf Jackrong/Qwopus3.8-27B-Flash-V2-GGUF:Q4_K_M -ngl 99 -fa on -ctk q8_0 -ctv q4_0 -c 32768 -npp 512 -ntg 128 -npl 1,2,4,5,6,8,16 -dev ROCm0; | Sequences | Stock TG (t/s) | Patched TG (t/s) | Change | Stock step (ms) | Patched step (ms) |; | 2 | 50.9 | 50.6 | \u2014 | 39 | 40 |",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp #29820",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29820",
   "tg_tokens": 128,
   "tps": "50.6",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 4,
   "ctx": 32768,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "lc-29820-r9700-npl4-stock",
   "model": "Qwopus3.8-27B-Flash-V2",
   "notes": "Qwopus3.8-27B-Flash-V2 Q4_K_M (qwen35 arch), Radeon AI PRO R9700 (gfx1201), ROCm 7.2.4, llama-batched-bench -c 32768 -npp 512 -ntg 128 -npl 4 -dev ROCm0, -ngl 99 -fa on -ctk q8_0 -ctv q4_0, TG t/s; stock build, MMVQ_MAX_BATCH_SIZE=8; config=mmvq-max8",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "On the Radeon AI PRO R9700 (gfx1201), the HIP backend sends quantized matmuls with 5 to 8 columns to MMVQ, which seems to be up to 46% slower than MMQ at those sizes. Lowering `MMVQ_MAX_BATCH_SIZE` from 8 to 4 fixes it.; | GPU | AMD Radeon AI PRO R9700, gfx1201 |; | ROCm | 7.2.4 |; | Target model | `Jackrong/Qwopus3.8-27B-Flash-V2-GGUF:Q4_K_M`, qwen35 architecture |; | Common flags | `-ngl 99 -fa on -ctk q8_0 -ctv q4_0` |; llama-batched-bench -hf Jackrong/Qwopus3.8-27B-Flash-V2-GGUF:Q4_K_M -ngl 99 -fa on -ctk q8_0 -ctv q4_0 -c 32768 -npp 512 -ntg 128 -npl 1,2,4,5,6,8,16 -dev ROCm0; | Sequences | Stock TG (t/s) | Patched TG (t/s) | Change | Stock step (ms) | Patched step (ms) |; | 4 | 77.0 | 75.9 | \u2014 | 52 | 53 |",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp #29820",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29820",
   "tg_tokens": 128,
   "tps": "77.0",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 4,
   "ctx": 32768,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "lc-29820-r9700-npl4-patched",
   "model": "Qwopus3.8-27B-Flash-V2",
   "notes": "Qwopus3.8-27B-Flash-V2 Q4_K_M (qwen35 arch), Radeon AI PRO R9700 (gfx1201), ROCm 7.2.4, llama-batched-bench -c 32768 -npp 512 -ntg 128 -npl 4 -dev ROCm0, -ngl 99 -fa on -ctk q8_0 -ctv q4_0, TG t/s; patched build, MMVQ_MAX_BATCH_SIZE=4; config=mmvq-max4",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "On the Radeon AI PRO R9700 (gfx1201), the HIP backend sends quantized matmuls with 5 to 8 columns to MMVQ, which seems to be up to 46% slower than MMQ at those sizes. Lowering `MMVQ_MAX_BATCH_SIZE` from 8 to 4 fixes it.; | GPU | AMD Radeon AI PRO R9700, gfx1201 |; | ROCm | 7.2.4 |; | Target model | `Jackrong/Qwopus3.8-27B-Flash-V2-GGUF:Q4_K_M`, qwen35 architecture |; | Common flags | `-ngl 99 -fa on -ctk q8_0 -ctv q4_0` |; llama-batched-bench -hf Jackrong/Qwopus3.8-27B-Flash-V2-GGUF:Q4_K_M -ngl 99 -fa on -ctk q8_0 -ctv q4_0 -c 32768 -npp 512 -ntg 128 -npl 1,2,4,5,6,8,16 -dev ROCm0; | Sequences | Stock TG (t/s) | Patched TG (t/s) | Change | Stock step (ms) | Patched step (ms) |; | 4 | 77.0 | 75.9 | \u2014 | 52 | 53 |",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp #29820",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29820",
   "tg_tokens": 128,
   "tps": "75.9",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 5,
   "ctx": 32768,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "lc-29820-r9700-npl5-stock",
   "model": "Qwopus3.8-27B-Flash-V2",
   "notes": "Qwopus3.8-27B-Flash-V2 Q4_K_M (qwen35 arch), Radeon AI PRO R9700 (gfx1201), ROCm 7.2.4, llama-batched-bench -c 32768 -npp 512 -ntg 128 -npl 5 -dev ROCm0, -ngl 99 -fa on -ctk q8_0 -ctv q4_0, TG t/s; stock build, MMVQ_MAX_BATCH_SIZE=8; config=mmvq-max8",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "On the Radeon AI PRO R9700 (gfx1201), the HIP backend sends quantized matmuls with 5 to 8 columns to MMVQ, which seems to be up to 46% slower than MMQ at those sizes. Lowering `MMVQ_MAX_BATCH_SIZE` from 8 to 4 fixes it.; | GPU | AMD Radeon AI PRO R9700, gfx1201 |; | ROCm | 7.2.4 |; | Target model | `Jackrong/Qwopus3.8-27B-Flash-V2-GGUF:Q4_K_M`, qwen35 architecture |; | Common flags | `-ngl 99 -fa on -ctk q8_0 -ctv q4_0` |; llama-batched-bench -hf Jackrong/Qwopus3.8-27B-Flash-V2-GGUF:Q4_K_M -ngl 99 -fa on -ctk q8_0 -ctv q4_0 -c 32768 -npp 512 -ntg 128 -npl 1,2,4,5,6,8,16 -dev ROCm0; | Sequences | Stock TG (t/s) | Patched TG (t/s) | Change | Stock step (ms) | Patched step (ms) |; | 5 | 81.1 | 91.8 | +13% | 62 | 54 |",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp #29820",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29820",
   "tg_tokens": 128,
   "tps": "81.1",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 5,
   "ctx": 32768,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "lc-29820-r9700-npl5-patched",
   "model": "Qwopus3.8-27B-Flash-V2",
   "notes": "Qwopus3.8-27B-Flash-V2 Q4_K_M (qwen35 arch), Radeon AI PRO R9700 (gfx1201), ROCm 7.2.4, llama-batched-bench -c 32768 -npp 512 -ntg 128 -npl 5 -dev ROCm0, -ngl 99 -fa on -ctk q8_0 -ctv q4_0, TG t/s; patched build, MMVQ_MAX_BATCH_SIZE=4; config=mmvq-max4",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "On the Radeon AI PRO R9700 (gfx1201), the HIP backend sends quantized matmuls with 5 to 8 columns to MMVQ, which seems to be up to 46% slower than MMQ at those sizes. Lowering `MMVQ_MAX_BATCH_SIZE` from 8 to 4 fixes it.; | GPU | AMD Radeon AI PRO R9700, gfx1201 |; | ROCm | 7.2.4 |; | Target model | `Jackrong/Qwopus3.8-27B-Flash-V2-GGUF:Q4_K_M`, qwen35 architecture |; | Common flags | `-ngl 99 -fa on -ctk q8_0 -ctv q4_0` |; llama-batched-bench -hf Jackrong/Qwopus3.8-27B-Flash-V2-GGUF:Q4_K_M -ngl 99 -fa on -ctk q8_0 -ctv q4_0 -c 32768 -npp 512 -ntg 128 -npl 1,2,4,5,6,8,16 -dev ROCm0; | Sequences | Stock TG (t/s) | Patched TG (t/s) | Change | Stock step (ms) | Patched step (ms) |; | 5 | 81.1 | 91.8 | +13% | 62 | 54 |",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp #29820",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29820",
   "tg_tokens": 128,
   "tps": "91.8",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 6,
   "ctx": 32768,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "lc-29820-r9700-npl6-stock",
   "model": "Qwopus3.8-27B-Flash-V2",
   "notes": "Qwopus3.8-27B-Flash-V2 Q4_K_M (qwen35 arch), Radeon AI PRO R9700 (gfx1201), ROCm 7.2.4, llama-batched-bench -c 32768 -npp 512 -ntg 128 -npl 6 -dev ROCm0, -ngl 99 -fa on -ctk q8_0 -ctv q4_0, TG t/s; stock build, MMVQ_MAX_BATCH_SIZE=8; config=mmvq-max8",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "On the Radeon AI PRO R9700 (gfx1201), the HIP backend sends quantized matmuls with 5 to 8 columns to MMVQ, which seems to be up to 46% slower than MMQ at those sizes. Lowering `MMVQ_MAX_BATCH_SIZE` from 8 to 4 fixes it.; | GPU | AMD Radeon AI PRO R9700, gfx1201 |; | ROCm | 7.2.4 |; | Target model | `Jackrong/Qwopus3.8-27B-Flash-V2-GGUF:Q4_K_M`, qwen35 architecture |; | Common flags | `-ngl 99 -fa on -ctk q8_0 -ctv q4_0` |; llama-batched-bench -hf Jackrong/Qwopus3.8-27B-Flash-V2-GGUF:Q4_K_M -ngl 99 -fa on -ctk q8_0 -ctv q4_0 -c 32768 -npp 512 -ntg 128 -npl 1,2,4,5,6,8,16 -dev ROCm0; | Sequences | Stock TG (t/s) | Patched TG (t/s) | Change | Stock step (ms) | Patched step (ms) |; | 6 | 84.4 | 104.9 | +24% | 71 | 57 |",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp #29820",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29820",
   "tg_tokens": 128,
   "tps": "84.4",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 6,
   "ctx": 32768,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "lc-29820-r9700-npl6-patched",
   "model": "Qwopus3.8-27B-Flash-V2",
   "notes": "Qwopus3.8-27B-Flash-V2 Q4_K_M (qwen35 arch), Radeon AI PRO R9700 (gfx1201), ROCm 7.2.4, llama-batched-bench -c 32768 -npp 512 -ntg 128 -npl 6 -dev ROCm0, -ngl 99 -fa on -ctk q8_0 -ctv q4_0, TG t/s; patched build, MMVQ_MAX_BATCH_SIZE=4; config=mmvq-max4",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "On the Radeon AI PRO R9700 (gfx1201), the HIP backend sends quantized matmuls with 5 to 8 columns to MMVQ, which seems to be up to 46% slower than MMQ at those sizes. Lowering `MMVQ_MAX_BATCH_SIZE` from 8 to 4 fixes it.; | GPU | AMD Radeon AI PRO R9700, gfx1201 |; | ROCm | 7.2.4 |; | Target model | `Jackrong/Qwopus3.8-27B-Flash-V2-GGUF:Q4_K_M`, qwen35 architecture |; | Common flags | `-ngl 99 -fa on -ctk q8_0 -ctv q4_0` |; llama-batched-bench -hf Jackrong/Qwopus3.8-27B-Flash-V2-GGUF:Q4_K_M -ngl 99 -fa on -ctk q8_0 -ctv q4_0 -c 32768 -npp 512 -ntg 128 -npl 1,2,4,5,6,8,16 -dev ROCm0; | Sequences | Stock TG (t/s) | Patched TG (t/s) | Change | Stock step (ms) | Patched step (ms) |; | 6 | 84.4 | 104.9 | +24% | 71 | 57 |",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp #29820",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29820",
   "tg_tokens": 128,
   "tps": "104.9",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 8,
   "ctx": 32768,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "lc-29820-r9700-npl8-stock",
   "model": "Qwopus3.8-27B-Flash-V2",
   "notes": "Qwopus3.8-27B-Flash-V2 Q4_K_M (qwen35 arch), Radeon AI PRO R9700 (gfx1201), ROCm 7.2.4, llama-batched-bench -c 32768 -npp 512 -ntg 128 -npl 8 -dev ROCm0, -ngl 99 -fa on -ctk q8_0 -ctv q4_0, TG t/s; stock build, MMVQ_MAX_BATCH_SIZE=8; config=mmvq-max8",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "On the Radeon AI PRO R9700 (gfx1201), the HIP backend sends quantized matmuls with 5 to 8 columns to MMVQ, which seems to be up to 46% slower than MMQ at those sizes. Lowering `MMVQ_MAX_BATCH_SIZE` from 8 to 4 fixes it.; | GPU | AMD Radeon AI PRO R9700, gfx1201 |; | ROCm | 7.2.4 |; | Target model | `Jackrong/Qwopus3.8-27B-Flash-V2-GGUF:Q4_K_M`, qwen35 architecture |; | Common flags | `-ngl 99 -fa on -ctk q8_0 -ctv q4_0` |; llama-batched-bench -hf Jackrong/Qwopus3.8-27B-Flash-V2-GGUF:Q4_K_M -ngl 99 -fa on -ctk q8_0 -ctv q4_0 -c 32768 -npp 512 -ntg 128 -npl 1,2,4,5,6,8,16 -dev ROCm0; | Sequences | Stock TG (t/s) | Patched TG (t/s) | Change | Stock step (ms) | Patched step (ms) |; | 8 | 88.6 | 127.3 | +44% | 90 | 63 |",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp #29820",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29820",
   "tg_tokens": 128,
   "tps": "88.6",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 8,
   "ctx": 32768,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "lc-29820-r9700-npl8-patched",
   "model": "Qwopus3.8-27B-Flash-V2",
   "notes": "Qwopus3.8-27B-Flash-V2 Q4_K_M (qwen35 arch), Radeon AI PRO R9700 (gfx1201), ROCm 7.2.4, llama-batched-bench -c 32768 -npp 512 -ntg 128 -npl 8 -dev ROCm0, -ngl 99 -fa on -ctk q8_0 -ctv q4_0, TG t/s; patched build, MMVQ_MAX_BATCH_SIZE=4; config=mmvq-max4",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "On the Radeon AI PRO R9700 (gfx1201), the HIP backend sends quantized matmuls with 5 to 8 columns to MMVQ, which seems to be up to 46% slower than MMQ at those sizes. Lowering `MMVQ_MAX_BATCH_SIZE` from 8 to 4 fixes it.; | GPU | AMD Radeon AI PRO R9700, gfx1201 |; | ROCm | 7.2.4 |; | Target model | `Jackrong/Qwopus3.8-27B-Flash-V2-GGUF:Q4_K_M`, qwen35 architecture |; | Common flags | `-ngl 99 -fa on -ctk q8_0 -ctv q4_0` |; llama-batched-bench -hf Jackrong/Qwopus3.8-27B-Flash-V2-GGUF:Q4_K_M -ngl 99 -fa on -ctk q8_0 -ctv q4_0 -c 32768 -npp 512 -ntg 128 -npl 1,2,4,5,6,8,16 -dev ROCm0; | Sequences | Stock TG (t/s) | Patched TG (t/s) | Change | Stock step (ms) | Patched step (ms) |; | 8 | 88.6 | 127.3 | +44% | 90 | 63 |",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp #29820",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29820",
   "tg_tokens": 128,
   "tps": "127.3",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 16,
   "ctx": 32768,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "lc-29820-r9700-npl16-stock",
   "model": "Qwopus3.8-27B-Flash-V2",
   "notes": "Qwopus3.8-27B-Flash-V2 Q4_K_M (qwen35 arch), Radeon AI PRO R9700 (gfx1201), ROCm 7.2.4, llama-batched-bench -c 32768 -npp 512 -ntg 128 -npl 16 -dev ROCm0, -ngl 99 -fa on -ctk q8_0 -ctv q4_0, TG t/s; stock build, MMVQ_MAX_BATCH_SIZE=8; config=mmvq-max8",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "On the Radeon AI PRO R9700 (gfx1201), the HIP backend sends quantized matmuls with 5 to 8 columns to MMVQ, which seems to be up to 46% slower than MMQ at those sizes. Lowering `MMVQ_MAX_BATCH_SIZE` from 8 to 4 fixes it.; | GPU | AMD Radeon AI PRO R9700, gfx1201 |; | ROCm | 7.2.4 |; | Target model | `Jackrong/Qwopus3.8-27B-Flash-V2-GGUF:Q4_K_M`, qwen35 architecture |; | Common flags | `-ngl 99 -fa on -ctk q8_0 -ctv q4_0` |; llama-batched-bench -hf Jackrong/Qwopus3.8-27B-Flash-V2-GGUF:Q4_K_M -ngl 99 -fa on -ctk q8_0 -ctv q4_0 -c 32768 -npp 512 -ntg 128 -npl 1,2,4,5,6,8,16 -dev ROCm0; | Sequences | Stock TG (t/s) | Patched TG (t/s) | Change | Stock step (ms) | Patched step (ms) |; | 16 | 174.5 | 174.6 | \u2014 | 92 | 92 |",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp #29820",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29820",
   "tg_tokens": 128,
   "tps": "174.5",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 16,
   "ctx": 32768,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "lc-29820-r9700-npl16-patched",
   "model": "Qwopus3.8-27B-Flash-V2",
   "notes": "Qwopus3.8-27B-Flash-V2 Q4_K_M (qwen35 arch), Radeon AI PRO R9700 (gfx1201), ROCm 7.2.4, llama-batched-bench -c 32768 -npp 512 -ntg 128 -npl 16 -dev ROCm0, -ngl 99 -fa on -ctk q8_0 -ctv q4_0, TG t/s; patched build, MMVQ_MAX_BATCH_SIZE=4; config=mmvq-max4",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "On the Radeon AI PRO R9700 (gfx1201), the HIP backend sends quantized matmuls with 5 to 8 columns to MMVQ, which seems to be up to 46% slower than MMQ at those sizes. Lowering `MMVQ_MAX_BATCH_SIZE` from 8 to 4 fixes it.; | GPU | AMD Radeon AI PRO R9700, gfx1201 |; | ROCm | 7.2.4 |; | Target model | `Jackrong/Qwopus3.8-27B-Flash-V2-GGUF:Q4_K_M`, qwen35 architecture |; | Common flags | `-ngl 99 -fa on -ctk q8_0 -ctv q4_0` |; llama-batched-bench -hf Jackrong/Qwopus3.8-27B-Flash-V2-GGUF:Q4_K_M -ngl 99 -fa on -ctk q8_0 -ctv q4_0 -c 32768 -npp 512 -ntg 128 -npl 1,2,4,5,6,8,16 -dev ROCm0; | Sequences | Stock TG (t/s) | Patched TG (t/s) | Change | Stock step (ms) | Patched step (ms) |; | 16 | 174.5 | 174.6 | \u2014 | 92 | 92 |",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp #29820",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29820",
   "tg_tokens": 128,
   "tps": "174.6",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-29809-b70-20b-dec-before",
   "model": "gpt-oss 20B",
   "notes": "gpt-oss 20B MXFP4 MoE, Arc Pro B70, SYCL backend, Windows 11, oneAPI 2026.0, AMD EPYC 7402P host, flash attention on, decode with 5 repetitions (20B) or 2 (120B); before the patch (table-based MXFP4 decode); config=mxfp4-table-decode",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Speeds up MXFP4 MoE inference in the SYCL backend:; gpt-oss-20b: 55.84 \u2192 106.64 tok/s (1.91\u00d7) decode, 5 repetitions; gpt-oss-120b on two cards: 33.27 \u2192 67.14 tok/s (2.02\u00d7) decode, 2 repetitions; gpt-oss-120b prompt processing: 594.5 \u2192 600.1 tok/s; This was tested and benchmarked on Intel Arc Pro B70 GPUs running Windows 11 and oneAPI 2026.0 and AMD EPYC 7402P, with flash attention enabled.; Gpt-oss-20b ran on one card and gpt-oss-120b on two cards.",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29809",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29809",
   "tg_tokens": null,
   "tps": "55.84",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-29809-b70-20b-dec-after",
   "model": "gpt-oss 20B",
   "notes": "gpt-oss 20B MXFP4 MoE, Arc Pro B70, SYCL backend, Windows 11, oneAPI 2026.0, AMD EPYC 7402P host, flash attention on, decode with 5 repetitions (20B) or 2 (120B); after the patch (arithmetic decode + weight reordering), 1.91x; config=mxfp4-arithmetic-decode",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Speeds up MXFP4 MoE inference in the SYCL backend:; gpt-oss-20b: 55.84 \u2192 106.64 tok/s (1.91\u00d7) decode, 5 repetitions; gpt-oss-120b on two cards: 33.27 \u2192 67.14 tok/s (2.02\u00d7) decode, 2 repetitions; gpt-oss-120b prompt processing: 594.5 \u2192 600.1 tok/s; This was tested and benchmarked on Intel Arc Pro B70 GPUs running Windows 11 and oneAPI 2026.0 and AMD EPYC 7402P, with flash attention enabled.; Gpt-oss-20b ran on one card and gpt-oss-120b on two cards.",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29809",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29809",
   "tg_tokens": null,
   "tps": "106.64",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-29809-2xb70-120b-dec-before",
   "model": "gpt-oss 120B",
   "notes": "gpt-oss 120B MXFP4 MoE, 2x Arc Pro B70 x2 (two cards), SYCL backend, Windows 11, oneAPI 2026.0, AMD EPYC 7402P host, flash attention on, decode with 5 repetitions (20B) or 2 (120B); before the patch (table-based MXFP4 decode); config=mxfp4-table-decode",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Speeds up MXFP4 MoE inference in the SYCL backend:; gpt-oss-20b: 55.84 \u2192 106.64 tok/s (1.91\u00d7) decode, 5 repetitions; gpt-oss-120b on two cards: 33.27 \u2192 67.14 tok/s (2.02\u00d7) decode, 2 repetitions; gpt-oss-120b prompt processing: 594.5 \u2192 600.1 tok/s; This was tested and benchmarked on Intel Arc Pro B70 GPUs running Windows 11 and oneAPI 2026.0 and AMD EPYC 7402P, with flash attention enabled.; Gpt-oss-20b ran on one card and gpt-oss-120b on two cards.",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29809",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29809",
   "tg_tokens": null,
   "tps": "33.27",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-29809-2xb70-120b-dec-after",
   "model": "gpt-oss 120B",
   "notes": "gpt-oss 120B MXFP4 MoE, 2x Arc Pro B70 x2 (two cards), SYCL backend, Windows 11, oneAPI 2026.0, AMD EPYC 7402P host, flash attention on, decode with 5 repetitions (20B) or 2 (120B); after the patch (arithmetic decode + weight reordering), 2.02x; config=mxfp4-arithmetic-decode",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Speeds up MXFP4 MoE inference in the SYCL backend:; gpt-oss-20b: 55.84 \u2192 106.64 tok/s (1.91\u00d7) decode, 5 repetitions; gpt-oss-120b on two cards: 33.27 \u2192 67.14 tok/s (2.02\u00d7) decode, 2 repetitions; gpt-oss-120b prompt processing: 594.5 \u2192 600.1 tok/s; This was tested and benchmarked on Intel Arc Pro B70 GPUs running Windows 11 and oneAPI 2026.0 and AMD EPYC 7402P, with flash attention enabled.; Gpt-oss-20b ran on one card and gpt-oss-120b on two cards.",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29809",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29809",
   "tg_tokens": null,
   "tps": "67.14",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-29809-2xb70-120b-pp-before",
   "model": "gpt-oss 120B",
   "notes": "gpt-oss 120B MXFP4 MoE, 2x Arc Pro B70 x2 (two cards), SYCL backend, Windows 11, oneAPI 2026.0, AMD EPYC 7402P host, flash attention on, decode with 5 repetitions (20B) or 2 (120B); before the patch, prompt processing; config=mxfp4-table-decode",
   "params": null,
   "power_w": null,
   "pp_tps": "594.5",
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Speeds up MXFP4 MoE inference in the SYCL backend:; gpt-oss-20b: 55.84 \u2192 106.64 tok/s (1.91\u00d7) decode, 5 repetitions; gpt-oss-120b on two cards: 33.27 \u2192 67.14 tok/s (2.02\u00d7) decode, 2 repetitions; gpt-oss-120b prompt processing: 594.5 \u2192 600.1 tok/s; This was tested and benchmarked on Intel Arc Pro B70 GPUs running Windows 11 and oneAPI 2026.0 and AMD EPYC 7402P, with flash attention enabled.; Gpt-oss-20b ran on one card and gpt-oss-120b on two cards.",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29809",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29809",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-29809-2xb70-120b-pp-after",
   "model": "gpt-oss 120B",
   "notes": "gpt-oss 120B MXFP4 MoE, 2x Arc Pro B70 x2 (two cards), SYCL backend, Windows 11, oneAPI 2026.0, AMD EPYC 7402P host, flash attention on, decode with 5 repetitions (20B) or 2 (120B); after the patch, prompt processing; config=mxfp4-arithmetic-decode",
   "params": null,
   "power_w": null,
   "pp_tps": "600.1",
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Speeds up MXFP4 MoE inference in the SYCL backend:; gpt-oss-20b: 55.84 \u2192 106.64 tok/s (1.91\u00d7) decode, 5 repetitions; gpt-oss-120b on two cards: 33.27 \u2192 67.14 tok/s (2.02\u00d7) decode, 2 repetitions; gpt-oss-120b prompt processing: 594.5 \u2192 600.1 tok/s; This was tested and benchmarked on Intel Arc Pro B70 GPUs running Windows 11 and oneAPI 2026.0 and AMD EPYC 7402P, with flash attention enabled.; Gpt-oss-20b ran on one card and gpt-oss-120b on two cards.",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29809",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29809",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "lc-29807-pro6000b-pp512-master",
   "model": "nemotron_h_moe 31B.A3.5B",
   "notes": "nemotron_h_moe 31B.A3.5B Q4_K_M (NVIDIA-Nemotron-3.5-Lightning-30B-A3B-UD), RTX PRO 6000 Blackwell (edition not stated), Windows, llama-bench MTP off, pp512, mean \u00b1 stddev; master; config=ssmcopy-master",
   "params": null,
   "power_w": null,
   "pp_tps": "11485.61",
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "Model: NVIDIA-Nemotron-3.5-Lightning-30B-A3B-UD-Q4_K_M.gguf; System: Windows with NVIDIA RTX Pro 6000 Blackwell GPU; | model | test | Master t/s | PR t/s | Speed-up |; | nemotron_h_moe 31B.A3.5B Q4_K - Medium | pp512 | 11485.61 \u00b1 643.02 | 11474.42 \u00b1 646.61 | 1.00x |",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29807",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29807",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "lc-29807-pro6000b-pp512-pr",
   "model": "nemotron_h_moe 31B.A3.5B",
   "notes": "nemotron_h_moe 31B.A3.5B Q4_K_M (NVIDIA-Nemotron-3.5-Lightning-30B-A3B-UD), RTX PRO 6000 Blackwell (edition not stated), Windows, llama-bench MTP off, pp512, mean \u00b1 stddev; PR; config=ssmcopy-pr",
   "params": null,
   "power_w": null,
   "pp_tps": "11474.42",
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "Model: NVIDIA-Nemotron-3.5-Lightning-30B-A3B-UD-Q4_K_M.gguf; System: Windows with NVIDIA RTX Pro 6000 Blackwell GPU; | model | test | Master t/s | PR t/s | Speed-up |; | nemotron_h_moe 31B.A3.5B Q4_K - Medium | pp512 | 11485.61 \u00b1 643.02 | 11474.42 \u00b1 646.61 | 1.00x |",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29807",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29807",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "lc-29807-pro6000b-tg128-master",
   "model": "nemotron_h_moe 31B.A3.5B",
   "notes": "nemotron_h_moe 31B.A3.5B Q4_K_M (NVIDIA-Nemotron-3.5-Lightning-30B-A3B-UD), RTX PRO 6000 Blackwell (edition not stated), Windows, llama-bench MTP off, tg128, mean \u00b1 stddev; master; config=ssmcopy-master",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "Model: NVIDIA-Nemotron-3.5-Lightning-30B-A3B-UD-Q4_K_M.gguf; System: Windows with NVIDIA RTX Pro 6000 Blackwell GPU; | model | test | Master t/s | PR t/s | Speed-up |; | nemotron_h_moe 31B.A3.5B Q4_K - Medium | tg128 | 313.28 \u00b1 1.27 | 314.14 \u00b1 0.60 | 1.00x |",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29807",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29807",
   "tg_tokens": null,
   "tps": "313.28",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "lc-29807-pro6000b-tg128-pr",
   "model": "nemotron_h_moe 31B.A3.5B",
   "notes": "nemotron_h_moe 31B.A3.5B Q4_K_M (NVIDIA-Nemotron-3.5-Lightning-30B-A3B-UD), RTX PRO 6000 Blackwell (edition not stated), Windows, llama-bench MTP off, tg128, mean \u00b1 stddev; PR; config=ssmcopy-pr",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "Model: NVIDIA-Nemotron-3.5-Lightning-30B-A3B-UD-Q4_K_M.gguf; System: Windows with NVIDIA RTX Pro 6000 Blackwell GPU; | model | test | Master t/s | PR t/s | Speed-up |; | nemotron_h_moe 31B.A3.5B Q4_K - Medium | tg128 | 313.28 \u00b1 1.27 | 314.14 \u00b1 0.60 | 1.00x |",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29807",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29807",
   "tg_tokens": null,
   "tps": "314.14",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 8192,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "lc-29807-pro6000b-mtp-master",
   "model": "nemotron_h_moe 31B.A3.5B",
   "notes": "nemotron_h_moe 31B.A3.5B Q4_K_M (NVIDIA-Nemotron-3.5-Lightning-30B-A3B-UD), RTX PRO 6000 Blackwell (edition not stated), Windows, llama-server -c 8192 --spec-type draft-mtp, SPEED-Bench coding 80 prompts x 1024 tokens greedy, mean of 2 runs, decode t/s; master; config=ssmcopy-master",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "Model: NVIDIA-Nemotron-3.5-Lightning-30B-A3B-UD-Q4_K_M.gguf; System: Windows with NVIDIA RTX Pro 6000 Blackwell GPU; | model | test | Master t/s | PR t/s | Speed-up |; | nemotron_h_moe 31B.A3.5B Q4_K - Medium | pp512 | 11485.61 \u00b1 643.02 | 11474.42 \u00b1 646.61 | 1.00x |; | nemotron_h_moe 31B.A3.5B Q4_K - Medium | tg128 | 313.28 \u00b1 1.27 | 314.14 \u00b1 0.60 | 1.00x |; MTP on: Command- `llama-server -m NVIDIA-Nemotron-3.5-Lightning-30B-A3B-UD-Q4_K_M.gguf -ngl 99 -fa on -c 8192 --spec-type draft-mtp --reasoning off` (SPEED-Bench coding prompts, greedy, 1024 tokens per prompt, tok/s is the mean of 2 runs per build); | test | Master t/s | PR t/s | Speed-up |; | SPEED-Bench coding, 80 prompts x 1024 tokens | 432.58 | 447.63 | 1.035x |; Acceptance is identical on master and PR: acceptance rate 0.777 (57244 of 73648 drafted tokens accepted)",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29807",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29807",
   "tg_tokens": 1024,
   "tps": "432.58",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 8192,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "lc-29807-pro6000b-mtp-pr",
   "model": "nemotron_h_moe 31B.A3.5B",
   "notes": "nemotron_h_moe 31B.A3.5B Q4_K_M (NVIDIA-Nemotron-3.5-Lightning-30B-A3B-UD), RTX PRO 6000 Blackwell (edition not stated), Windows, llama-server -c 8192 --spec-type draft-mtp, SPEED-Bench coding 80 prompts x 1024 tokens greedy, mean of 2 runs, decode t/s; PR; config=ssmcopy-pr",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "Model: NVIDIA-Nemotron-3.5-Lightning-30B-A3B-UD-Q4_K_M.gguf; System: Windows with NVIDIA RTX Pro 6000 Blackwell GPU; | model | test | Master t/s | PR t/s | Speed-up |; | nemotron_h_moe 31B.A3.5B Q4_K - Medium | pp512 | 11485.61 \u00b1 643.02 | 11474.42 \u00b1 646.61 | 1.00x |; | nemotron_h_moe 31B.A3.5B Q4_K - Medium | tg128 | 313.28 \u00b1 1.27 | 314.14 \u00b1 0.60 | 1.00x |; MTP on: Command- `llama-server -m NVIDIA-Nemotron-3.5-Lightning-30B-A3B-UD-Q4_K_M.gguf -ngl 99 -fa on -c 8192 --spec-type draft-mtp --reasoning off` (SPEED-Bench coding prompts, greedy, 1024 tokens per prompt, tok/s is the mean of 2 runs per build); | test | Master t/s | PR t/s | Speed-up |; | SPEED-Bench coding, 80 prompts x 1024 tokens | 432.58 | 447.63 | 1.035x |; Acceptance is identical on master and PR: acceptance rate 0.777 (57244 of 73648 drafted tokens accepted)",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29807",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29807",
   "tg_tokens": 1024,
   "tps": "447.63",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 131072,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-29769-rtx5090-npl8-before",
   "model": "Qwen3.8-27B",
   "notes": "Qwen3.8-27B Q4_K_M, F16 KV cache, RTX 5090 (CC 12.0, Windows x64, driver 617.14, CUDA 12.9, static build), llama-batched-bench -c 131072 -npp 15872 -ntg 512 -npl 8, decode median of 10 paired runs; without the unified decode mask patch; config=decodemask-off",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Measured on RTX 5090 (compute capability 12.0; Windows x64, driver 617.14, CUDA 12.9, static build) with Qwen3.8-27B Q4_K_M and an F16 KV cache.; llama-batched-bench -c 131072 -b 2048 -ub 512 -npp 15872 -ntg 512 -npl 8 -ngl 99 -fa on -ctk f16 -ctv f16 -kvu; Medians without -> with; the change is the mean of the 10 paired per-repetition changes:; 274.54 -> 291.21 tok/s, +5.77%; (95% CI +5.11% to +6.43%)",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29769",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29769",
   "tg_tokens": 512,
   "tps": "274.54",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 131072,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-29769-rtx5090-npl8-after",
   "model": "Qwen3.8-27B",
   "notes": "Qwen3.8-27B Q4_K_M, F16 KV cache, RTX 5090 (CC 12.0, Windows x64, driver 617.14, CUDA 12.9, static build), llama-batched-bench -c 131072 -npp 15872 -ntg 512 -npl 8, decode median of 10 paired runs, +5.77% (95% CI +5.11% to +6.43%); with the unified decode mask patch; config=decodemask-on",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Measured on RTX 5090 (compute capability 12.0; Windows x64, driver 617.14, CUDA 12.9, static build) with Qwen3.8-27B Q4_K_M and an F16 KV cache.; llama-batched-bench -c 131072 -b 2048 -ub 512 -npp 15872 -ntg 512 -npl 8 -ngl 99 -fa on -ctk f16 -ctv f16 -kvu; Medians without -> with; the change is the mean of the 10 paired per-repetition changes:; 274.54 -> 291.21 tok/s, +5.77%; (95% CI +5.11% to +6.43%)",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29769",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29769",
   "tg_tokens": 512,
   "tps": "291.21",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-29769-rtx5090-pp512-before",
   "model": "Qwen3.8-27B",
   "notes": "Qwen3.8-27B Q4_K_M, F16 KV cache, RTX 5090 (CC 12.0, Windows x64, driver 617.14, CUDA 12.9, static build), llama-bench pp512, 2 rounds x 5 reps mirrored, sanity run (fast path not triggered); without the unified decode mask patch; config=decodemask-off",
   "params": null,
   "power_w": null,
   "pp_tps": "3731.18",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Measured on RTX 5090 (compute capability 12.0; Windows x64, driver 617.14, CUDA 12.9, static build) with Qwen3.8-27B Q4_K_M and an F16 KV cache.; For completeness, ran the following as well on 5090, these don't trigger the fast path so were only sanity runs for merge requirements; llama-bench -ngl 99 -fa 1, two rounds of 5 repetitions in mirrored order: pp512 3731.18 -> 3718.85 tok/s (-0.33%), tg128 78.67 -> 78.49 (-0.23%)",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29769",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29769",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-29769-rtx5090-pp512-after",
   "model": "Qwen3.8-27B",
   "notes": "Qwen3.8-27B Q4_K_M, F16 KV cache, RTX 5090 (CC 12.0, Windows x64, driver 617.14, CUDA 12.9, static build), llama-bench pp512, 2 rounds x 5 reps mirrored, sanity run (fast path not triggered), -0.33%; with the unified decode mask patch; config=decodemask-on",
   "params": null,
   "power_w": null,
   "pp_tps": "3718.85",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Measured on RTX 5090 (compute capability 12.0; Windows x64, driver 617.14, CUDA 12.9, static build) with Qwen3.8-27B Q4_K_M and an F16 KV cache.; For completeness, ran the following as well on 5090, these don't trigger the fast path so were only sanity runs for merge requirements; llama-bench -ngl 99 -fa 1, two rounds of 5 repetitions in mirrored order: pp512 3731.18 -> 3718.85 tok/s (-0.33%), tg128 78.67 -> 78.49 (-0.23%)",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29769",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29769",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-29769-rtx5090-tg128-before",
   "model": "Qwen3.8-27B",
   "notes": "Qwen3.8-27B Q4_K_M, F16 KV cache, RTX 5090 (CC 12.0, Windows x64, driver 617.14, CUDA 12.9, static build), llama-bench tg128, 2 rounds x 5 reps mirrored, sanity run (fast path not triggered); without the unified decode mask patch; config=decodemask-off",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Measured on RTX 5090 (compute capability 12.0; Windows x64, driver 617.14, CUDA 12.9, static build) with Qwen3.8-27B Q4_K_M and an F16 KV cache.; For completeness, ran the following as well on 5090, these don't trigger the fast path so were only sanity runs for merge requirements; llama-bench -ngl 99 -fa 1, two rounds of 5 repetitions in mirrored order: pp512 3731.18 -> 3718.85 tok/s (-0.33%), tg128 78.67 -> 78.49 (-0.23%)",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29769",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29769",
   "tg_tokens": null,
   "tps": "78.67",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-29769-rtx5090-tg128-after",
   "model": "Qwen3.8-27B",
   "notes": "Qwen3.8-27B Q4_K_M, F16 KV cache, RTX 5090 (CC 12.0, Windows x64, driver 617.14, CUDA 12.9, static build), llama-bench tg128, 2 rounds x 5 reps mirrored, sanity run (fast path not triggered), -0.23%; with the unified decode mask patch; config=decodemask-on",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Measured on RTX 5090 (compute capability 12.0; Windows x64, driver 617.14, CUDA 12.9, static build) with Qwen3.8-27B Q4_K_M and an F16 KV cache.; For completeness, ran the following as well on 5090, these don't trigger the fast path so were only sanity runs for merge requirements; llama-bench -ngl 99 -fa 1, two rounds of 5 repetitions in mirrored order: pp512 3731.18 -> 3718.85 tok/s (-0.33%), tg128 78.67 -> 78.49 (-0.23%)",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29769",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29769",
   "tg_tokens": null,
   "tps": "78.49",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "M5 Max",
   "id": "lc-29784-m5max-pp512-master",
   "model": "muse-glimmer 30B",
   "notes": "muse-glimmer 30B Q4_K_XL, M5 Max (6 Super + 12 Performance cores), llama-bench -ngl 0 -r 3 (CPU-only, MTL,BLAS), mean of 3 runs; master (thread-count default fix, threads 6 -> 18); pp512; config=thrdefault-master",
   "params": null,
   "power_w": null,
   "pp_tps": "26.98",
   "provenance": "community",
   "quant": "Q4_K_XL",
   "quote": "On Apple chips with both a Super and a Performance cluster (M5 Max here: perflevel0 \"Super\" = 6 cores, perflevel1 \"Performance\" = 12), common_cpu_get_num_physical_cores() read only hw.perflevel0.physicalcpu , so the default thread count was 6 of 18 cores.; Muse-Glimmer-30B Q4_K_XL, -ngl 0 , -r 3 , M5 Max, master 0c1e570 -> this branch:; Only the M5 Max was measured.; pp512: 26.98 -> 58.76 t/s (2.18x)",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29784",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29784",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "M5 Max",
   "id": "lc-29784-m5max-pp512-fix",
   "model": "muse-glimmer 30B",
   "notes": "muse-glimmer 30B Q4_K_XL, M5 Max (6 Super + 12 Performance cores), llama-bench -ngl 0 -r 3 (CPU-only, MTL,BLAS), mean of 3 runs; with the fix (thread-count default fix, threads 6 -> 18); pp512, 2.18x; config=thrdefault-fix",
   "params": null,
   "power_w": null,
   "pp_tps": "58.76",
   "provenance": "community",
   "quant": "Q4_K_XL",
   "quote": "On Apple chips with both a Super and a Performance cluster (M5 Max here: perflevel0 \"Super\" = 6 cores, perflevel1 \"Performance\" = 12), common_cpu_get_num_physical_cores() read only hw.perflevel0.physicalcpu , so the default thread count was 6 of 18 cores.; Muse-Glimmer-30B Q4_K_XL, -ngl 0 , -r 3 , M5 Max, master 0c1e570 -> this branch:; Only the M5 Max was measured.; pp512: 26.98 -> 58.76 t/s (2.18x)",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29784",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29784",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "M5 Max",
   "id": "lc-29784-m5max-tg128-master",
   "model": "muse-glimmer 30B",
   "notes": "muse-glimmer 30B Q4_K_XL, M5 Max (6 Super + 12 Performance cores), llama-bench -ngl 0 -r 3 (CPU-only, MTL,BLAS), mean of 3 runs; master (thread-count default fix, threads 6 -> 18); tg128; config=thrdefault-master",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_XL",
   "quote": "On Apple chips with both a Super and a Performance cluster (M5 Max here: perflevel0 \"Super\" = 6 cores, perflevel1 \"Performance\" = 12), common_cpu_get_num_physical_cores() read only hw.perflevel0.physicalcpu , so the default thread count was 6 of 18 cores.; Muse-Glimmer-30B Q4_K_XL, -ngl 0 , -r 3 , M5 Max, master 0c1e570 -> this branch:; Only the M5 Max was measured.; tg128: 9.24 -> 13.64 t/s (1.48x)",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29784",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29784",
   "tg_tokens": null,
   "tps": "9.24",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "M5 Max",
   "id": "lc-29784-m5max-tg128-fix",
   "model": "muse-glimmer 30B",
   "notes": "muse-glimmer 30B Q4_K_XL, M5 Max (6 Super + 12 Performance cores), llama-bench -ngl 0 -r 3 (CPU-only, MTL,BLAS), mean of 3 runs; with the fix (thread-count default fix, threads 6 -> 18); tg128, 1.48x; config=thrdefault-fix",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_XL",
   "quote": "On Apple chips with both a Super and a Performance cluster (M5 Max here: perflevel0 \"Super\" = 6 cores, perflevel1 \"Performance\" = 12), common_cpu_get_num_physical_cores() read only hw.perflevel0.physicalcpu , so the default thread count was 6 of 18 cores.; Muse-Glimmer-30B Q4_K_XL, -ngl 0 , -r 3 , M5 Max, master 0c1e570 -> this branch:; Only the M5 Max was measured.; tg128: 9.24 -> 13.64 t/s (1.48x)",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29784",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29784",
   "tg_tokens": null,
   "tps": "13.64",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": 2048,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "lc-29824-m2ultra-2048-master",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Qwen3.8-Flash-Next GGUF (ggml-org), M2 Ultra, Metal, llama-batched-bench B=1, npp=ctx, ntg=32, S_PP and S_TG columns; master (mask graph from 1-element seed repeats, #29751); config=maskbuild-master",
   "params": null,
   "power_w": null,
   "pp_tps": "832.23",
   "provenance": "community",
   "quant": "unspecified",
   "quote": "M2 Ultra perf using https://huggingface.co/ggml-org/Qwen3.8-Flash-Next-GGUF; The graph for constructing the masks by repeating 1-element seeds is very inefficient in some backends such as Metal. Optimize it by repeating rows instead.; master PP TG B N_KV T_PP s S_PP t/s T_TG s S_TG t/s T s S t/s; 2048 32 1 2080 2.461 832.23 0.839 38.13 3.300 630.28",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29824",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29824",
   "tg_tokens": 32,
   "tps": "38.13",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "lc-29824-m2ultra-4096-master",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Qwen3.8-Flash-Next GGUF (ggml-org), M2 Ultra, Metal, llama-batched-bench B=1, npp=ctx, ntg=32, S_PP and S_TG columns; master (mask graph from 1-element seed repeats, #29751); config=maskbuild-master",
   "params": null,
   "power_w": null,
   "pp_tps": "802.72",
   "provenance": "community",
   "quant": "unspecified",
   "quote": "M2 Ultra perf using https://huggingface.co/ggml-org/Qwen3.8-Flash-Next-GGUF; The graph for constructing the masks by repeating 1-element seeds is very inefficient in some backends such as Metal. Optimize it by repeating rows instead.; master PP TG B N_KV T_PP s S_PP t/s T_TG s S_TG t/s T s S t/s; 4096 32 1 4128 5.103 802.72 0.842 38.00 5.945 694.38",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29824",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29824",
   "tg_tokens": 32,
   "tps": "38.00",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "lc-29824-m2ultra-8192-master",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Qwen3.8-Flash-Next GGUF (ggml-org), M2 Ultra, Metal, llama-batched-bench B=1, npp=ctx, ntg=32, S_PP and S_TG columns; master (mask graph from 1-element seed repeats, #29751); config=maskbuild-master",
   "params": null,
   "power_w": null,
   "pp_tps": "768.48",
   "provenance": "community",
   "quant": "unspecified",
   "quote": "M2 Ultra perf using https://huggingface.co/ggml-org/Qwen3.8-Flash-Next-GGUF; The graph for constructing the masks by repeating 1-element seeds is very inefficient in some backends such as Metal. Optimize it by repeating rows instead.; master PP TG B N_KV T_PP s S_PP t/s T_TG s S_TG t/s T s S t/s; 8192 32 1 8224 10.660 768.48 0.853 37.51 11.513 714.31",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29824",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29824",
   "tg_tokens": 32,
   "tps": "37.51",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "lc-29824-m2ultra-16384-master",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Qwen3.8-Flash-Next GGUF (ggml-org), M2 Ultra, Metal, llama-batched-bench B=1, npp=ctx, ntg=32, S_PP and S_TG columns; master (mask graph from 1-element seed repeats, #29751); config=maskbuild-master",
   "params": null,
   "power_w": null,
   "pp_tps": "717.95",
   "provenance": "community",
   "quant": "unspecified",
   "quote": "M2 Ultra perf using https://huggingface.co/ggml-org/Qwen3.8-Flash-Next-GGUF; The graph for constructing the masks by repeating 1-element seeds is very inefficient in some backends such as Metal. Optimize it by repeating rows instead.; master PP TG B N_KV T_PP s S_PP t/s T_TG s S_TG t/s T s S t/s; 16384 32 1 16416 22.820 717.95 0.875 36.57 23.695 692.79",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29824",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29824",
   "tg_tokens": 32,
   "tps": "36.57",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "lc-29824-m2ultra-32768-master",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Qwen3.8-Flash-Next GGUF (ggml-org), M2 Ultra, Metal, llama-batched-bench B=1, npp=ctx, ntg=32, S_PP and S_TG columns; master (mask graph from 1-element seed repeats, #29751); config=maskbuild-master",
   "params": null,
   "power_w": null,
   "pp_tps": "637.29",
   "provenance": "community",
   "quant": "unspecified",
   "quote": "M2 Ultra perf using https://huggingface.co/ggml-org/Qwen3.8-Flash-Next-GGUF; The graph for constructing the masks by repeating 1-element seeds is very inefficient in some backends such as Metal. Optimize it by repeating rows instead.; master PP TG B N_KV T_PP s S_PP t/s T_TG s S_TG t/s T s S t/s; 32768 32 1 32800 51.418 637.29 0.918 34.85 52.336 626.72",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29824",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29824",
   "tg_tokens": 32,
   "tps": "34.85",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "lc-29824-m2ultra-65536-master",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Qwen3.8-Flash-Next GGUF (ggml-org), M2 Ultra, Metal, llama-batched-bench B=1, npp=ctx, ntg=32, S_PP and S_TG columns; master (mask graph from 1-element seed repeats, #29751); config=maskbuild-master",
   "params": null,
   "power_w": null,
   "pp_tps": "528.55",
   "provenance": "community",
   "quant": "unspecified",
   "quote": "M2 Ultra perf using https://huggingface.co/ggml-org/Qwen3.8-Flash-Next-GGUF; The graph for constructing the masks by repeating 1-element seeds is very inefficient in some backends such as Metal. Optimize it by repeating rows instead.; master PP TG B N_KV T_PP s S_PP t/s T_TG s S_TG t/s T s S t/s; 65536 32 1 65568 123.992 528.55 0.997 32.08 124.990 524.59",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29824",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29824",
   "tg_tokens": 32,
   "tps": "32.08",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": 2048,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "lc-29824-m2ultra-2048-pr",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Qwen3.8-Flash-Next GGUF (ggml-org), M2 Ultra, Metal, llama-batched-bench B=1, npp=ctx, ntg=32, S_PP and S_TG columns; PR (mask construction repeats rows); config=maskbuild-pr",
   "params": null,
   "power_w": null,
   "pp_tps": "851.40",
   "provenance": "community",
   "quant": "unspecified",
   "quote": "M2 Ultra perf using https://huggingface.co/ggml-org/Qwen3.8-Flash-Next-GGUF; The graph for constructing the masks by repeating 1-element seeds is very inefficient in some backends such as Metal. Optimize it by repeating rows instead.; PR PP TG B N_KV T_PP s S_PP t/s T_TG s S_TG t/s T s S t/s; 2048 32 1 2080 2.405 851.40 0.859 37.25 3.264 637.18",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29824",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29824",
   "tg_tokens": 32,
   "tps": "37.25",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "lc-29824-m2ultra-4096-pr",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Qwen3.8-Flash-Next GGUF (ggml-org), M2 Ultra, Metal, llama-batched-bench B=1, npp=ctx, ntg=32, S_PP and S_TG columns; PR (mask construction repeats rows); config=maskbuild-pr",
   "params": null,
   "power_w": null,
   "pp_tps": "830.32",
   "provenance": "community",
   "quant": "unspecified",
   "quote": "M2 Ultra perf using https://huggingface.co/ggml-org/Qwen3.8-Flash-Next-GGUF; The graph for constructing the masks by repeating 1-element seeds is very inefficient in some backends such as Metal. Optimize it by repeating rows instead.; PR PP TG B N_KV T_PP s S_PP t/s T_TG s S_TG t/s T s S t/s; 4096 32 1 4128 4.933 830.32 0.870 36.78 5.803 711.35",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29824",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29824",
   "tg_tokens": 32,
   "tps": "36.78",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "lc-29824-m2ultra-8192-pr",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Qwen3.8-Flash-Next GGUF (ggml-org), M2 Ultra, Metal, llama-batched-bench B=1, npp=ctx, ntg=32, S_PP and S_TG columns; PR (mask construction repeats rows); config=maskbuild-pr",
   "params": null,
   "power_w": null,
   "pp_tps": "817.21",
   "provenance": "community",
   "quant": "unspecified",
   "quote": "M2 Ultra perf using https://huggingface.co/ggml-org/Qwen3.8-Flash-Next-GGUF; The graph for constructing the masks by repeating 1-element seeds is very inefficient in some backends such as Metal. Optimize it by repeating rows instead.; PR PP TG B N_KV T_PP s S_PP t/s T_TG s S_TG t/s T s S t/s; 8192 32 1 8224 10.024 817.21 0.881 36.31 10.906 754.10",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29824",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29824",
   "tg_tokens": 32,
   "tps": "36.31",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "lc-29824-m2ultra-16384-pr",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Qwen3.8-Flash-Next GGUF (ggml-org), M2 Ultra, Metal, llama-batched-bench B=1, npp=ctx, ntg=32, S_PP and S_TG columns; PR (mask construction repeats rows); config=maskbuild-pr",
   "params": null,
   "power_w": null,
   "pp_tps": "801.99",
   "provenance": "community",
   "quant": "unspecified",
   "quote": "M2 Ultra perf using https://huggingface.co/ggml-org/Qwen3.8-Flash-Next-GGUF; The graph for constructing the masks by repeating 1-element seeds is very inefficient in some backends such as Metal. Optimize it by repeating rows instead.; PR PP TG B N_KV T_PP s S_PP t/s T_TG s S_TG t/s T s S t/s; 16384 32 1 16416 20.429 801.99 0.907 35.30 21.336 769.41",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29824",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29824",
   "tg_tokens": 32,
   "tps": "35.30",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "lc-29824-m2ultra-32768-pr",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Qwen3.8-Flash-Next GGUF (ggml-org), M2 Ultra, Metal, llama-batched-bench B=1, npp=ctx, ntg=32, S_PP and S_TG columns; PR (mask construction repeats rows); config=maskbuild-pr",
   "params": null,
   "power_w": null,
   "pp_tps": "781.91",
   "provenance": "community",
   "quant": "unspecified",
   "quote": "M2 Ultra perf using https://huggingface.co/ggml-org/Qwen3.8-Flash-Next-GGUF; The graph for constructing the masks by repeating 1-element seeds is very inefficient in some backends such as Metal. Optimize it by repeating rows instead.; PR PP TG B N_KV T_PP s S_PP t/s T_TG s S_TG t/s T s S t/s; 32768 32 1 32800 41.908 781.91 0.932 34.35 42.839 765.65",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29824",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29824",
   "tg_tokens": 32,
   "tps": "34.35",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "lc-29824-m2ultra-65536-pr",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Qwen3.8-Flash-Next GGUF (ggml-org), M2 Ultra, Metal, llama-batched-bench B=1, npp=ctx, ntg=32, S_PP and S_TG columns; PR (mask construction repeats rows); config=maskbuild-pr",
   "params": null,
   "power_w": null,
   "pp_tps": "754.63",
   "provenance": "community",
   "quant": "unspecified",
   "quote": "M2 Ultra perf using https://huggingface.co/ggml-org/Qwen3.8-Flash-Next-GGUF; The graph for constructing the masks by repeating 1-element seeds is very inefficient in some backends such as Metal. Optimize it by repeating rows instead.; PR PP TG B N_KV T_PP s S_PP t/s T_TG s S_TG t/s T s S t/s; 65536 32 1 65568 86.845 754.63 1.004 31.86 87.850 746.36",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29824",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29824",
   "tg_tokens": 32,
   "tps": "31.86",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Hexagon)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "IQ-9075",
   "id": "lc-29779-qwen35-1pp-master",
   "model": "Qwen3.5-2B",
   "notes": "Qwen3.5-2B Q4_0, IQ-9075 Hexagon NPU (HTP0:0, -t 6 --cpu-mask 0xfc), llama-batched-bench -c 4096 -npp 128 -ntg 32 -npl 1, prompt t/s at 1 sequence; master the HMX flatten-matmul patch; config=hmxflat-master",
   "params": null,
   "power_w": null,
   "pp_tps": "673.17",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Test on IQ-9075, run two models (that cover the path) with llama-batched-bench --device HTP0:0 -ngl 99 -fa on -t 6 --cpu-mask 0xfc -c 4096 -b 512 -ub 256 -npp 128 -ntg 32 -npl 1,2,4; into 2d to use HMX, which is faster on n_seqs > 1; Qwen3.5-2B Q4_0 1 pp128 673.17 673.00 -0.0%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29779",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29779",
   "tg_tokens": 128,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Hexagon)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "IQ-9075",
   "id": "lc-29779-qwen35-1pp-patch",
   "model": "Qwen3.5-2B",
   "notes": "Qwen3.5-2B Q4_0, IQ-9075 Hexagon NPU (HTP0:0, -t 6 --cpu-mask 0xfc), llama-batched-bench -c 4096 -npp 128 -ntg 32 -npl 1, prompt t/s at 1 sequence; patch the HMX flatten-matmul patch; config=hmxflat-patch",
   "params": null,
   "power_w": null,
   "pp_tps": "673.00",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Test on IQ-9075, run two models (that cover the path) with llama-batched-bench --device HTP0:0 -ngl 99 -fa on -t 6 --cpu-mask 0xfc -c 4096 -b 512 -ub 256 -npp 128 -ntg 32 -npl 1,2,4; into 2d to use HMX, which is faster on n_seqs > 1; Qwen3.5-2B Q4_0 1 pp128 673.17 673.00 -0.0%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29779",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29779",
   "tg_tokens": 128,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Hexagon)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "IQ-9075",
   "id": "lc-29779-qwen35-1tg-master",
   "model": "Qwen3.5-2B",
   "notes": "Qwen3.5-2B Q4_0, IQ-9075 Hexagon NPU (HTP0:0, -t 6 --cpu-mask 0xfc), llama-batched-bench -c 4096 -npp 128 -ntg 32 -npl 1, decode t/s at 1 sequence; master the HMX flatten-matmul patch; config=hmxflat-master",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Test on IQ-9075, run two models (that cover the path) with llama-batched-bench --device HTP0:0 -ngl 99 -fa on -t 6 --cpu-mask 0xfc -c 4096 -b 512 -ub 256 -npp 128 -ntg 32 -npl 1,2,4; into 2d to use HMX, which is faster on n_seqs > 1; Qwen3.5-2B Q4_0 1 tg32 20.68 20.66 -0.1%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29779",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29779",
   "tg_tokens": 32,
   "tps": "20.68",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Hexagon)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "IQ-9075",
   "id": "lc-29779-qwen35-1tg-patch",
   "model": "Qwen3.5-2B",
   "notes": "Qwen3.5-2B Q4_0, IQ-9075 Hexagon NPU (HTP0:0, -t 6 --cpu-mask 0xfc), llama-batched-bench -c 4096 -npp 128 -ntg 32 -npl 1, decode t/s at 1 sequence; patch the HMX flatten-matmul patch; config=hmxflat-patch",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Test on IQ-9075, run two models (that cover the path) with llama-batched-bench --device HTP0:0 -ngl 99 -fa on -t 6 --cpu-mask 0xfc -c 4096 -b 512 -ub 256 -npp 128 -ntg 32 -npl 1,2,4; into 2d to use HMX, which is faster on n_seqs > 1; Qwen3.5-2B Q4_0 1 tg32 20.68 20.66 -0.1%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29779",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29779",
   "tg_tokens": 32,
   "tps": "20.66",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Hexagon)",
   "batch": 2,
   "ctx": 4096,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "IQ-9075",
   "id": "lc-29779-qwen35-2pp-master",
   "model": "Qwen3.5-2B",
   "notes": "Qwen3.5-2B Q4_0, IQ-9075 Hexagon NPU (HTP0:0, -t 6 --cpu-mask 0xfc), llama-batched-bench -c 4096 -npp 128 -ntg 32 -npl 2, prompt t/s at 2 sequences; master the HMX flatten-matmul patch; config=hmxflat-master",
   "params": null,
   "power_w": null,
   "pp_tps": "452.70",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Test on IQ-9075, run two models (that cover the path) with llama-batched-bench --device HTP0:0 -ngl 99 -fa on -t 6 --cpu-mask 0xfc -c 4096 -b 512 -ub 256 -npp 128 -ntg 32 -npl 1,2,4; into 2d to use HMX, which is faster on n_seqs > 1; Qwen3.5-2B Q4_0 2 pp128 x 2 452.70 719.49 +58.9%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29779",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29779",
   "tg_tokens": 128,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Hexagon)",
   "batch": 2,
   "ctx": 4096,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "IQ-9075",
   "id": "lc-29779-qwen35-2pp-patch",
   "model": "Qwen3.5-2B",
   "notes": "Qwen3.5-2B Q4_0, IQ-9075 Hexagon NPU (HTP0:0, -t 6 --cpu-mask 0xfc), llama-batched-bench -c 4096 -npp 128 -ntg 32 -npl 2, prompt t/s at 2 sequences; patch the HMX flatten-matmul patch; config=hmxflat-patch",
   "params": null,
   "power_w": null,
   "pp_tps": "719.49",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Test on IQ-9075, run two models (that cover the path) with llama-batched-bench --device HTP0:0 -ngl 99 -fa on -t 6 --cpu-mask 0xfc -c 4096 -b 512 -ub 256 -npp 128 -ntg 32 -npl 1,2,4; into 2d to use HMX, which is faster on n_seqs > 1; Qwen3.5-2B Q4_0 2 pp128 x 2 452.70 719.49 +58.9%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29779",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29779",
   "tg_tokens": 128,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Hexagon)",
   "batch": 2,
   "ctx": 4096,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "IQ-9075",
   "id": "lc-29779-qwen35-2tg-master",
   "model": "Qwen3.5-2B",
   "notes": "Qwen3.5-2B Q4_0, IQ-9075 Hexagon NPU (HTP0:0, -t 6 --cpu-mask 0xfc), llama-batched-bench -c 4096 -npp 128 -ntg 32 -npl 2, decode t/s at 2 sequences; master the HMX flatten-matmul patch; config=hmxflat-master",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Test on IQ-9075, run two models (that cover the path) with llama-batched-bench --device HTP0:0 -ngl 99 -fa on -t 6 --cpu-mask 0xfc -c 4096 -b 512 -ub 256 -npp 128 -ntg 32 -npl 1,2,4; into 2d to use HMX, which is faster on n_seqs > 1; Qwen3.5-2B Q4_0 2 tg32 x 2 29.53 30.19 +2.2%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29779",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29779",
   "tg_tokens": 32,
   "tps": "29.53",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Hexagon)",
   "batch": 2,
   "ctx": 4096,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "IQ-9075",
   "id": "lc-29779-qwen35-2tg-patch",
   "model": "Qwen3.5-2B",
   "notes": "Qwen3.5-2B Q4_0, IQ-9075 Hexagon NPU (HTP0:0, -t 6 --cpu-mask 0xfc), llama-batched-bench -c 4096 -npp 128 -ntg 32 -npl 2, decode t/s at 2 sequences; patch the HMX flatten-matmul patch; config=hmxflat-patch",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Test on IQ-9075, run two models (that cover the path) with llama-batched-bench --device HTP0:0 -ngl 99 -fa on -t 6 --cpu-mask 0xfc -c 4096 -b 512 -ub 256 -npp 128 -ntg 32 -npl 1,2,4; into 2d to use HMX, which is faster on n_seqs > 1; Qwen3.5-2B Q4_0 2 tg32 x 2 29.53 30.19 +2.2%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29779",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29779",
   "tg_tokens": 32,
   "tps": "30.19",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Hexagon)",
   "batch": 4,
   "ctx": 4096,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "IQ-9075",
   "id": "lc-29779-qwen35-4pp-master",
   "model": "Qwen3.5-2B",
   "notes": "Qwen3.5-2B Q4_0, IQ-9075 Hexagon NPU (HTP0:0, -t 6 --cpu-mask 0xfc), llama-batched-bench -c 4096 -npp 128 -ntg 32 -npl 4, prompt t/s at 4 sequences; master the HMX flatten-matmul patch; config=hmxflat-master",
   "params": null,
   "power_w": null,
   "pp_tps": "445.93",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Test on IQ-9075, run two models (that cover the path) with llama-batched-bench --device HTP0:0 -ngl 99 -fa on -t 6 --cpu-mask 0xfc -c 4096 -b 512 -ub 256 -npp 128 -ntg 32 -npl 1,2,4; into 2d to use HMX, which is faster on n_seqs > 1; Qwen3.5-2B Q4_0 4 pp128 x 4 445.93 703.51 +57.8%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29779",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29779",
   "tg_tokens": 128,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Hexagon)",
   "batch": 4,
   "ctx": 4096,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "IQ-9075",
   "id": "lc-29779-qwen35-4pp-patch",
   "model": "Qwen3.5-2B",
   "notes": "Qwen3.5-2B Q4_0, IQ-9075 Hexagon NPU (HTP0:0, -t 6 --cpu-mask 0xfc), llama-batched-bench -c 4096 -npp 128 -ntg 32 -npl 4, prompt t/s at 4 sequences; patch the HMX flatten-matmul patch; config=hmxflat-patch",
   "params": null,
   "power_w": null,
   "pp_tps": "703.51",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Test on IQ-9075, run two models (that cover the path) with llama-batched-bench --device HTP0:0 -ngl 99 -fa on -t 6 --cpu-mask 0xfc -c 4096 -b 512 -ub 256 -npp 128 -ntg 32 -npl 1,2,4; into 2d to use HMX, which is faster on n_seqs > 1; Qwen3.5-2B Q4_0 4 pp128 x 4 445.93 703.51 +57.8%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29779",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29779",
   "tg_tokens": 128,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Hexagon)",
   "batch": 4,
   "ctx": 4096,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "IQ-9075",
   "id": "lc-29779-qwen35-4tg-master",
   "model": "Qwen3.5-2B",
   "notes": "Qwen3.5-2B Q4_0, IQ-9075 Hexagon NPU (HTP0:0, -t 6 --cpu-mask 0xfc), llama-batched-bench -c 4096 -npp 128 -ntg 32 -npl 4, decode t/s at 4 sequences; master the HMX flatten-matmul patch; config=hmxflat-master",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Test on IQ-9075, run two models (that cover the path) with llama-batched-bench --device HTP0:0 -ngl 99 -fa on -t 6 --cpu-mask 0xfc -c 4096 -b 512 -ub 256 -npp 128 -ntg 32 -npl 1,2,4; into 2d to use HMX, which is faster on n_seqs > 1; Qwen3.5-2B Q4_0 4 tg32 x 4 33.58 34.18 +1.8%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29779",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29779",
   "tg_tokens": 32,
   "tps": "33.58",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Hexagon)",
   "batch": 4,
   "ctx": 4096,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "IQ-9075",
   "id": "lc-29779-qwen35-4tg-patch",
   "model": "Qwen3.5-2B",
   "notes": "Qwen3.5-2B Q4_0, IQ-9075 Hexagon NPU (HTP0:0, -t 6 --cpu-mask 0xfc), llama-batched-bench -c 4096 -npp 128 -ntg 32 -npl 4, decode t/s at 4 sequences; patch the HMX flatten-matmul patch; config=hmxflat-patch",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Test on IQ-9075, run two models (that cover the path) with llama-batched-bench --device HTP0:0 -ngl 99 -fa on -t 6 --cpu-mask 0xfc -c 4096 -b 512 -ub 256 -npp 128 -ntg 32 -npl 1,2,4; into 2d to use HMX, which is faster on n_seqs > 1; Qwen3.5-2B Q4_0 4 tg32 x 4 33.58 34.18 +1.8%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29779",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29779",
   "tg_tokens": 32,
   "tps": "34.18",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Hexagon)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "IQ-9075",
   "id": "lc-29779-lfm2-1pp-master",
   "model": "LFM2-2.6B",
   "notes": "LFM2-2.6B Q4_0, IQ-9075 Hexagon NPU (HTP0:0, -t 6 --cpu-mask 0xfc), llama-batched-bench -c 4096 -npp 128 -ntg 32 -npl 1, prompt t/s at 1 sequence; master the HMX flatten-matmul patch; config=hmxflat-master",
   "params": null,
   "power_w": null,
   "pp_tps": "685.60",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Test on IQ-9075, run two models (that cover the path) with llama-batched-bench --device HTP0:0 -ngl 99 -fa on -t 6 --cpu-mask 0xfc -c 4096 -b 512 -ub 256 -npp 128 -ntg 32 -npl 1,2,4; into 2d to use HMX, which is faster on n_seqs > 1; LFM2-2.6B Q4_0 1 pp128 685.60 683.94 -0.2%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29779",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29779",
   "tg_tokens": 128,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Hexagon)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "IQ-9075",
   "id": "lc-29779-lfm2-1pp-patch",
   "model": "LFM2-2.6B",
   "notes": "LFM2-2.6B Q4_0, IQ-9075 Hexagon NPU (HTP0:0, -t 6 --cpu-mask 0xfc), llama-batched-bench -c 4096 -npp 128 -ntg 32 -npl 1, prompt t/s at 1 sequence; patch the HMX flatten-matmul patch; config=hmxflat-patch",
   "params": null,
   "power_w": null,
   "pp_tps": "683.94",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Test on IQ-9075, run two models (that cover the path) with llama-batched-bench --device HTP0:0 -ngl 99 -fa on -t 6 --cpu-mask 0xfc -c 4096 -b 512 -ub 256 -npp 128 -ntg 32 -npl 1,2,4; into 2d to use HMX, which is faster on n_seqs > 1; LFM2-2.6B Q4_0 1 pp128 685.60 683.94 -0.2%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29779",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29779",
   "tg_tokens": 128,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Hexagon)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "IQ-9075",
   "id": "lc-29779-lfm2-1tg-master",
   "model": "LFM2-2.6B",
   "notes": "LFM2-2.6B Q4_0, IQ-9075 Hexagon NPU (HTP0:0, -t 6 --cpu-mask 0xfc), llama-batched-bench -c 4096 -npp 128 -ntg 32 -npl 1, decode t/s at 1 sequence; master the HMX flatten-matmul patch; config=hmxflat-master",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Test on IQ-9075, run two models (that cover the path) with llama-batched-bench --device HTP0:0 -ngl 99 -fa on -t 6 --cpu-mask 0xfc -c 4096 -b 512 -ub 256 -npp 128 -ntg 32 -npl 1,2,4; into 2d to use HMX, which is faster on n_seqs > 1; LFM2-2.6B Q4_0 1 tg32 22.32 22.64 +1.4%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29779",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29779",
   "tg_tokens": 32,
   "tps": "22.32",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Hexagon)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "IQ-9075",
   "id": "lc-29779-lfm2-1tg-patch",
   "model": "LFM2-2.6B",
   "notes": "LFM2-2.6B Q4_0, IQ-9075 Hexagon NPU (HTP0:0, -t 6 --cpu-mask 0xfc), llama-batched-bench -c 4096 -npp 128 -ntg 32 -npl 1, decode t/s at 1 sequence; patch the HMX flatten-matmul patch; config=hmxflat-patch",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Test on IQ-9075, run two models (that cover the path) with llama-batched-bench --device HTP0:0 -ngl 99 -fa on -t 6 --cpu-mask 0xfc -c 4096 -b 512 -ub 256 -npp 128 -ntg 32 -npl 1,2,4; into 2d to use HMX, which is faster on n_seqs > 1; LFM2-2.6B Q4_0 1 tg32 22.32 22.64 +1.4%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29779",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29779",
   "tg_tokens": 32,
   "tps": "22.64",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Hexagon)",
   "batch": 2,
   "ctx": 4096,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "IQ-9075",
   "id": "lc-29779-lfm2-2pp-master",
   "model": "LFM2-2.6B",
   "notes": "LFM2-2.6B Q4_0, IQ-9075 Hexagon NPU (HTP0:0, -t 6 --cpu-mask 0xfc), llama-batched-bench -c 4096 -npp 128 -ntg 32 -npl 2, prompt t/s at 2 sequences; master the HMX flatten-matmul patch; config=hmxflat-master",
   "params": null,
   "power_w": null,
   "pp_tps": "209.92",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Test on IQ-9075, run two models (that cover the path) with llama-batched-bench --device HTP0:0 -ngl 99 -fa on -t 6 --cpu-mask 0xfc -c 4096 -b 512 -ub 256 -npp 128 -ntg 32 -npl 1,2,4; into 2d to use HMX, which is faster on n_seqs > 1; LFM2-2.6B Q4_0 2 pp128 x 2 209.92 687.22 +227.4%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29779",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29779",
   "tg_tokens": 128,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Hexagon)",
   "batch": 2,
   "ctx": 4096,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "IQ-9075",
   "id": "lc-29779-lfm2-2pp-patch",
   "model": "LFM2-2.6B",
   "notes": "LFM2-2.6B Q4_0, IQ-9075 Hexagon NPU (HTP0:0, -t 6 --cpu-mask 0xfc), llama-batched-bench -c 4096 -npp 128 -ntg 32 -npl 2, prompt t/s at 2 sequences; patch the HMX flatten-matmul patch; config=hmxflat-patch",
   "params": null,
   "power_w": null,
   "pp_tps": "687.22",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Test on IQ-9075, run two models (that cover the path) with llama-batched-bench --device HTP0:0 -ngl 99 -fa on -t 6 --cpu-mask 0xfc -c 4096 -b 512 -ub 256 -npp 128 -ntg 32 -npl 1,2,4; into 2d to use HMX, which is faster on n_seqs > 1; LFM2-2.6B Q4_0 2 pp128 x 2 209.92 687.22 +227.4%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29779",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29779",
   "tg_tokens": 128,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Hexagon)",
   "batch": 2,
   "ctx": 4096,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "IQ-9075",
   "id": "lc-29779-lfm2-2tg-master",
   "model": "LFM2-2.6B",
   "notes": "LFM2-2.6B Q4_0, IQ-9075 Hexagon NPU (HTP0:0, -t 6 --cpu-mask 0xfc), llama-batched-bench -c 4096 -npp 128 -ntg 32 -npl 2, decode t/s at 2 sequences; master the HMX flatten-matmul patch; config=hmxflat-master",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Test on IQ-9075, run two models (that cover the path) with llama-batched-bench --device HTP0:0 -ngl 99 -fa on -t 6 --cpu-mask 0xfc -c 4096 -b 512 -ub 256 -npp 128 -ntg 32 -npl 1,2,4; into 2d to use HMX, which is faster on n_seqs > 1; LFM2-2.6B Q4_0 2 tg32 x 2 27.62 28.13 +1.9%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29779",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29779",
   "tg_tokens": 32,
   "tps": "27.62",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Hexagon)",
   "batch": 2,
   "ctx": 4096,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "IQ-9075",
   "id": "lc-29779-lfm2-2tg-patch",
   "model": "LFM2-2.6B",
   "notes": "LFM2-2.6B Q4_0, IQ-9075 Hexagon NPU (HTP0:0, -t 6 --cpu-mask 0xfc), llama-batched-bench -c 4096 -npp 128 -ntg 32 -npl 2, decode t/s at 2 sequences; patch the HMX flatten-matmul patch; config=hmxflat-patch",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Test on IQ-9075, run two models (that cover the path) with llama-batched-bench --device HTP0:0 -ngl 99 -fa on -t 6 --cpu-mask 0xfc -c 4096 -b 512 -ub 256 -npp 128 -ntg 32 -npl 1,2,4; into 2d to use HMX, which is faster on n_seqs > 1; LFM2-2.6B Q4_0 2 tg32 x 2 27.62 28.13 +1.9%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29779",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29779",
   "tg_tokens": 32,
   "tps": "28.13",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Hexagon)",
   "batch": 4,
   "ctx": 4096,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "IQ-9075",
   "id": "lc-29779-lfm2-4pp-master",
   "model": "LFM2-2.6B",
   "notes": "LFM2-2.6B Q4_0, IQ-9075 Hexagon NPU (HTP0:0, -t 6 --cpu-mask 0xfc), llama-batched-bench -c 4096 -npp 128 -ntg 32 -npl 4, prompt t/s at 4 sequences; master the HMX flatten-matmul patch; config=hmxflat-master",
   "params": null,
   "power_w": null,
   "pp_tps": "209.26",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Test on IQ-9075, run two models (that cover the path) with llama-batched-bench --device HTP0:0 -ngl 99 -fa on -t 6 --cpu-mask 0xfc -c 4096 -b 512 -ub 256 -npp 128 -ntg 32 -npl 1,2,4; into 2d to use HMX, which is faster on n_seqs > 1; LFM2-2.6B Q4_0 4 pp128 x 4 209.26 679.47 +224.7%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29779",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29779",
   "tg_tokens": 128,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Hexagon)",
   "batch": 4,
   "ctx": 4096,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "IQ-9075",
   "id": "lc-29779-lfm2-4pp-patch",
   "model": "LFM2-2.6B",
   "notes": "LFM2-2.6B Q4_0, IQ-9075 Hexagon NPU (HTP0:0, -t 6 --cpu-mask 0xfc), llama-batched-bench -c 4096 -npp 128 -ntg 32 -npl 4, prompt t/s at 4 sequences; patch the HMX flatten-matmul patch; config=hmxflat-patch",
   "params": null,
   "power_w": null,
   "pp_tps": "679.47",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Test on IQ-9075, run two models (that cover the path) with llama-batched-bench --device HTP0:0 -ngl 99 -fa on -t 6 --cpu-mask 0xfc -c 4096 -b 512 -ub 256 -npp 128 -ntg 32 -npl 1,2,4; into 2d to use HMX, which is faster on n_seqs > 1; LFM2-2.6B Q4_0 4 pp128 x 4 209.26 679.47 +224.7%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29779",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29779",
   "tg_tokens": 128,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Hexagon)",
   "batch": 4,
   "ctx": 4096,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "IQ-9075",
   "id": "lc-29779-lfm2-4tg-master",
   "model": "LFM2-2.6B",
   "notes": "LFM2-2.6B Q4_0, IQ-9075 Hexagon NPU (HTP0:0, -t 6 --cpu-mask 0xfc), llama-batched-bench -c 4096 -npp 128 -ntg 32 -npl 4, decode t/s at 4 sequences; master the HMX flatten-matmul patch; config=hmxflat-master",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Test on IQ-9075, run two models (that cover the path) with llama-batched-bench --device HTP0:0 -ngl 99 -fa on -t 6 --cpu-mask 0xfc -c 4096 -b 512 -ub 256 -npp 128 -ntg 32 -npl 1,2,4; into 2d to use HMX, which is faster on n_seqs > 1; LFM2-2.6B Q4_0 4 tg32 x 4 30.74 31.47 +2.4%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29779",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29779",
   "tg_tokens": 32,
   "tps": "30.74",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Hexagon)",
   "batch": 4,
   "ctx": 4096,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "IQ-9075",
   "id": "lc-29779-lfm2-4tg-patch",
   "model": "LFM2-2.6B",
   "notes": "LFM2-2.6B Q4_0, IQ-9075 Hexagon NPU (HTP0:0, -t 6 --cpu-mask 0xfc), llama-batched-bench -c 4096 -npp 128 -ntg 32 -npl 4, decode t/s at 4 sequences; patch the HMX flatten-matmul patch; config=hmxflat-patch",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Test on IQ-9075, run two models (that cover the path) with llama-batched-bench --device HTP0:0 -ngl 99 -fa on -t 6 --cpu-mask 0xfc -c 4096 -b 512 -ub 256 -npp 128 -ntg 32 -npl 1,2,4; into 2d to use HMX, which is faster on n_seqs > 1; LFM2-2.6B Q4_0 4 tg32 x 4 30.74 31.47 +2.4%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29779",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29779",
   "tg_tokens": 32,
   "tps": "31.47",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 8192,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59668-mi355x-8k1k-c4-default",
   "model": "DeepSeek V4.1",
   "notes": "DeepSeek V4.1, MI355X TP4 + EP, FP8, fused single-launch DSA decode candidate mask, vllm bench serve output throughput at concurrency 4, mean TPOT 8.05 ms; stock four-kernel mask path; config=dsamask-default",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "On ROCm, the DeepSeek V4.1 decode candidate mask goes from four kernels per consumer indexer layer to one.; vLLM on MI355X, TP4 + EP, FP8, fused against default. Mean TPOT in ms and output throughput in tok/s:; input / output concurrency TPOT default TPOT fused; 8192 / 1024 4 8.05 7.99 -0.7% 470.57 472.20 +0.3%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59668",
   "source_url": "https://github.com/vllm-project/vllm/pull/59668",
   "tg_tokens": 1024,
   "tps": "470.57",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 8192,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59668-mi355x-8k1k-c4-fused",
   "model": "DeepSeek V4.1",
   "notes": "DeepSeek V4.1, MI355X TP4 + EP, FP8, fused single-launch DSA decode candidate mask, vllm bench serve output throughput at concurrency 4, mean TPOT 7.99 ms; fused single-kernel mask path; config=dsamask-fused",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "On ROCm, the DeepSeek V4.1 decode candidate mask goes from four kernels per consumer indexer layer to one.; vLLM on MI355X, TP4 + EP, FP8, fused against default. Mean TPOT in ms and output throughput in tok/s:; input / output concurrency TPOT default TPOT fused; 8192 / 1024 4 8.05 7.99 -0.7% 470.57 472.20 +0.3%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59668",
   "source_url": "https://github.com/vllm-project/vllm/pull/59668",
   "tg_tokens": 1024,
   "tps": "472.20",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 8192,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59668-mi355x-8k1k-c8-default",
   "model": "DeepSeek V4.1",
   "notes": "DeepSeek V4.1, MI355X TP4 + EP, FP8, fused single-launch DSA decode candidate mask, vllm bench serve output throughput at concurrency 8, mean TPOT 9.81 ms; stock four-kernel mask path; config=dsamask-default",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "On ROCm, the DeepSeek V4.1 decode candidate mask goes from four kernels per consumer indexer layer to one.; vLLM on MI355X, TP4 + EP, FP8, fused against default. Mean TPOT in ms and output throughput in tok/s:; input / output concurrency TPOT default TPOT fused; 8192 / 1024 8 9.81 9.73 -0.9% 772.13 778.70 +0.9%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59668",
   "source_url": "https://github.com/vllm-project/vllm/pull/59668",
   "tg_tokens": 1024,
   "tps": "772.13",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 8192,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59668-mi355x-8k1k-c8-fused",
   "model": "DeepSeek V4.1",
   "notes": "DeepSeek V4.1, MI355X TP4 + EP, FP8, fused single-launch DSA decode candidate mask, vllm bench serve output throughput at concurrency 8, mean TPOT 9.73 ms; fused single-kernel mask path; config=dsamask-fused",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "On ROCm, the DeepSeek V4.1 decode candidate mask goes from four kernels per consumer indexer layer to one.; vLLM on MI355X, TP4 + EP, FP8, fused against default. Mean TPOT in ms and output throughput in tok/s:; input / output concurrency TPOT default TPOT fused; 8192 / 1024 8 9.81 9.73 -0.9% 772.13 778.70 +0.9%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59668",
   "source_url": "https://github.com/vllm-project/vllm/pull/59668",
   "tg_tokens": 1024,
   "tps": "778.70",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": 8192,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59668-mi355x-8k1k-c16-default",
   "model": "DeepSeek V4.1",
   "notes": "DeepSeek V4.1, MI355X TP4 + EP, FP8, fused single-launch DSA decode candidate mask, vllm bench serve output throughput at concurrency 16, mean TPOT 11.85 ms; stock four-kernel mask path; config=dsamask-default",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "On ROCm, the DeepSeek V4.1 decode candidate mask goes from four kernels per consumer indexer layer to one.; vLLM on MI355X, TP4 + EP, FP8, fused against default. Mean TPOT in ms and output throughput in tok/s:; input / output concurrency TPOT default TPOT fused; 8192 / 1024 16 11.85 11.74 -0.9% 1262.90 1273.77 +0.9%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59668",
   "source_url": "https://github.com/vllm-project/vllm/pull/59668",
   "tg_tokens": 1024,
   "tps": "1262.90",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": 8192,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59668-mi355x-8k1k-c16-fused",
   "model": "DeepSeek V4.1",
   "notes": "DeepSeek V4.1, MI355X TP4 + EP, FP8, fused single-launch DSA decode candidate mask, vllm bench serve output throughput at concurrency 16, mean TPOT 11.74 ms; fused single-kernel mask path; config=dsamask-fused",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "On ROCm, the DeepSeek V4.1 decode candidate mask goes from four kernels per consumer indexer layer to one.; vLLM on MI355X, TP4 + EP, FP8, fused against default. Mean TPOT in ms and output throughput in tok/s:; input / output concurrency TPOT default TPOT fused; 8192 / 1024 16 11.85 11.74 -0.9% 1262.90 1273.77 +0.9%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59668",
   "source_url": "https://github.com/vllm-project/vllm/pull/59668",
   "tg_tokens": 1024,
   "tps": "1273.77",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 32,
   "ctx": 8192,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59668-mi355x-8k1k-c32-default",
   "model": "DeepSeek V4.1",
   "notes": "DeepSeek V4.1, MI355X TP4 + EP, FP8, fused single-launch DSA decode candidate mask, vllm bench serve output throughput at concurrency 32, mean TPOT 16.16 ms; stock four-kernel mask path; config=dsamask-default",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "On ROCm, the DeepSeek V4.1 decode candidate mask goes from four kernels per consumer indexer layer to one.; vLLM on MI355X, TP4 + EP, FP8, fused against default. Mean TPOT in ms and output throughput in tok/s:; input / output concurrency TPOT default TPOT fused; 8192 / 1024 32 16.16 16.07 -0.6% 1783.04 1793.30 +0.6%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59668",
   "source_url": "https://github.com/vllm-project/vllm/pull/59668",
   "tg_tokens": 1024,
   "tps": "1783.04",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 32,
   "ctx": 8192,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59668-mi355x-8k1k-c32-fused",
   "model": "DeepSeek V4.1",
   "notes": "DeepSeek V4.1, MI355X TP4 + EP, FP8, fused single-launch DSA decode candidate mask, vllm bench serve output throughput at concurrency 32, mean TPOT 16.07 ms; fused single-kernel mask path; config=dsamask-fused",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "On ROCm, the DeepSeek V4.1 decode candidate mask goes from four kernels per consumer indexer layer to one.; vLLM on MI355X, TP4 + EP, FP8, fused against default. Mean TPOT in ms and output throughput in tok/s:; input / output concurrency TPOT default TPOT fused; 8192 / 1024 32 16.16 16.07 -0.6% 1783.04 1793.30 +0.6%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59668",
   "source_url": "https://github.com/vllm-project/vllm/pull/59668",
   "tg_tokens": 1024,
   "tps": "1793.30",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 64,
   "ctx": 8192,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59668-mi355x-8k1k-c64-default",
   "model": "DeepSeek V4.1",
   "notes": "DeepSeek V4.1, MI355X TP4 + EP, FP8, fused single-launch DSA decode candidate mask, vllm bench serve output throughput at concurrency 64, mean TPOT 23.07 ms; stock four-kernel mask path; config=dsamask-default",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "On ROCm, the DeepSeek V4.1 decode candidate mask goes from four kernels per consumer indexer layer to one.; vLLM on MI355X, TP4 + EP, FP8, fused against default. Mean TPOT in ms and output throughput in tok/s:; input / output concurrency TPOT default TPOT fused; 8192 / 1024 64 23.07 22.84 -1.0% 2523.76 2545.75 +0.9%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59668",
   "source_url": "https://github.com/vllm-project/vllm/pull/59668",
   "tg_tokens": 1024,
   "tps": "2523.76",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 64,
   "ctx": 8192,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59668-mi355x-8k1k-c64-fused",
   "model": "DeepSeek V4.1",
   "notes": "DeepSeek V4.1, MI355X TP4 + EP, FP8, fused single-launch DSA decode candidate mask, vllm bench serve output throughput at concurrency 64, mean TPOT 22.84 ms; fused single-kernel mask path; config=dsamask-fused",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "On ROCm, the DeepSeek V4.1 decode candidate mask goes from four kernels per consumer indexer layer to one.; vLLM on MI355X, TP4 + EP, FP8, fused against default. Mean TPOT in ms and output throughput in tok/s:; input / output concurrency TPOT default TPOT fused; 8192 / 1024 64 23.07 22.84 -1.0% 2523.76 2545.75 +0.9%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59668",
   "source_url": "https://github.com/vllm-project/vllm/pull/59668",
   "tg_tokens": 1024,
   "tps": "2545.75",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 128,
   "ctx": 8192,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59668-mi355x-8k1k-c128-default",
   "model": "DeepSeek V4.1",
   "notes": "DeepSeek V4.1, MI355X TP4 + EP, FP8, fused single-launch DSA decode candidate mask, vllm bench serve output throughput at concurrency 128, mean TPOT 35.20 ms; stock four-kernel mask path; config=dsamask-default",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "On ROCm, the DeepSeek V4.1 decode candidate mask goes from four kernels per consumer indexer layer to one.; vLLM on MI355X, TP4 + EP, FP8, fused against default. Mean TPOT in ms and output throughput in tok/s:; input / output concurrency TPOT default TPOT fused; 8192 / 1024 128 35.20 35.15 -0.1% 3176.89 3182.50 +0.2%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59668",
   "source_url": "https://github.com/vllm-project/vllm/pull/59668",
   "tg_tokens": 1024,
   "tps": "3176.89",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 128,
   "ctx": 8192,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59668-mi355x-8k1k-c128-fused",
   "model": "DeepSeek V4.1",
   "notes": "DeepSeek V4.1, MI355X TP4 + EP, FP8, fused single-launch DSA decode candidate mask, vllm bench serve output throughput at concurrency 128, mean TPOT 35.15 ms; fused single-kernel mask path; config=dsamask-fused",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "On ROCm, the DeepSeek V4.1 decode candidate mask goes from four kernels per consumer indexer layer to one.; vLLM on MI355X, TP4 + EP, FP8, fused against default. Mean TPOT in ms and output throughput in tok/s:; input / output concurrency TPOT default TPOT fused; 8192 / 1024 128 35.20 35.15 -0.1% 3176.89 3182.50 +0.2%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59668",
   "source_url": "https://github.com/vllm-project/vllm/pull/59668",
   "tg_tokens": 1024,
   "tps": "3182.50",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 256,
   "ctx": 8192,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59668-mi355x-8k1k-c256-default",
   "model": "DeepSeek V4.1",
   "notes": "DeepSeek V4.1, MI355X TP4 + EP, FP8, fused single-launch DSA decode candidate mask, vllm bench serve output throughput at concurrency 256, mean TPOT 57.90 ms; stock four-kernel mask path; config=dsamask-default",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "On ROCm, the DeepSeek V4.1 decode candidate mask goes from four kernels per consumer indexer layer to one.; vLLM on MI355X, TP4 + EP, FP8, fused against default. Mean TPOT in ms and output throughput in tok/s:; input / output concurrency TPOT default TPOT fused; 8192 / 1024 256 57.90 57.72 -0.3% 3814.25 3827.03 +0.3%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59668",
   "source_url": "https://github.com/vllm-project/vllm/pull/59668",
   "tg_tokens": 1024,
   "tps": "3814.25",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 256,
   "ctx": 8192,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59668-mi355x-8k1k-c256-fused",
   "model": "DeepSeek V4.1",
   "notes": "DeepSeek V4.1, MI355X TP4 + EP, FP8, fused single-launch DSA decode candidate mask, vllm bench serve output throughput at concurrency 256, mean TPOT 57.72 ms; fused single-kernel mask path; config=dsamask-fused",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "On ROCm, the DeepSeek V4.1 decode candidate mask goes from four kernels per consumer indexer layer to one.; vLLM on MI355X, TP4 + EP, FP8, fused against default. Mean TPOT in ms and output throughput in tok/s:; input / output concurrency TPOT default TPOT fused; 8192 / 1024 256 57.90 57.72 -0.3% 3814.25 3827.03 +0.3%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59668",
   "source_url": "https://github.com/vllm-project/vllm/pull/59668",
   "tg_tokens": 1024,
   "tps": "3827.03",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 60000,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59668-mi355x-60k600-c4-default",
   "model": "DeepSeek V4.1",
   "notes": "DeepSeek V4.1, MI355X TP4 + EP, FP8, fused single-launch DSA decode candidate mask, vllm bench serve output throughput at concurrency 4, mean TPOT 14.05 ms; stock four-kernel mask path; config=dsamask-default",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "On ROCm, the DeepSeek V4.1 decode candidate mask goes from four kernels per consumer indexer layer to one.; vLLM on MI355X, TP4 + EP, FP8, fused against default. Mean TPOT in ms and output throughput in tok/s:; input / output concurrency TPOT default TPOT fused; 60000 / 600 4 14.05 13.97 -0.6% 227.76 229.03 +0.6%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59668",
   "source_url": "https://github.com/vllm-project/vllm/pull/59668",
   "tg_tokens": 600,
   "tps": "227.76",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 60000,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59668-mi355x-60k600-c4-fused",
   "model": "DeepSeek V4.1",
   "notes": "DeepSeek V4.1, MI355X TP4 + EP, FP8, fused single-launch DSA decode candidate mask, vllm bench serve output throughput at concurrency 4, mean TPOT 13.97 ms; fused single-kernel mask path; config=dsamask-fused",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "On ROCm, the DeepSeek V4.1 decode candidate mask goes from four kernels per consumer indexer layer to one.; vLLM on MI355X, TP4 + EP, FP8, fused against default. Mean TPOT in ms and output throughput in tok/s:; input / output concurrency TPOT default TPOT fused; 60000 / 600 4 14.05 13.97 -0.6% 227.76 229.03 +0.6%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59668",
   "source_url": "https://github.com/vllm-project/vllm/pull/59668",
   "tg_tokens": 600,
   "tps": "229.03",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": 60000,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59668-mi355x-60k600-c16-default",
   "model": "DeepSeek V4.1",
   "notes": "DeepSeek V4.1, MI355X TP4 + EP, FP8, fused single-launch DSA decode candidate mask, vllm bench serve output throughput at concurrency 16, mean TPOT 40.27 ms; stock four-kernel mask path; config=dsamask-default",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "On ROCm, the DeepSeek V4.1 decode candidate mask goes from four kernels per consumer indexer layer to one.; vLLM on MI355X, TP4 + EP, FP8, fused against default. Mean TPOT in ms and output throughput in tok/s:; input / output concurrency TPOT default TPOT fused; 60000 / 600 16 40.27 40.02 -0.6% 335.86 337.69 +0.5%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59668",
   "source_url": "https://github.com/vllm-project/vllm/pull/59668",
   "tg_tokens": 600,
   "tps": "335.86",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": 60000,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59668-mi355x-60k600-c16-fused",
   "model": "DeepSeek V4.1",
   "notes": "DeepSeek V4.1, MI355X TP4 + EP, FP8, fused single-launch DSA decode candidate mask, vllm bench serve output throughput at concurrency 16, mean TPOT 40.02 ms; fused single-kernel mask path; config=dsamask-fused",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "On ROCm, the DeepSeek V4.1 decode candidate mask goes from four kernels per consumer indexer layer to one.; vLLM on MI355X, TP4 + EP, FP8, fused against default. Mean TPOT in ms and output throughput in tok/s:; input / output concurrency TPOT default TPOT fused; 60000 / 600 16 40.27 40.02 -0.6% 335.86 337.69 +0.5%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59668",
   "source_url": "https://github.com/vllm-project/vllm/pull/59668",
   "tg_tokens": 600,
   "tps": "337.69",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 64,
   "ctx": 60000,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59668-mi355x-60k600-c64-default",
   "model": "DeepSeek V4.1",
   "notes": "DeepSeek V4.1, MI355X TP4 + EP, FP8, fused single-launch DSA decode candidate mask, vllm bench serve output throughput at concurrency 64, mean TPOT 137.05 ms; stock four-kernel mask path; config=dsamask-default",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "On ROCm, the DeepSeek V4.1 decode candidate mask goes from four kernels per consumer indexer layer to one.; vLLM on MI355X, TP4 + EP, FP8, fused against default. Mean TPOT in ms and output throughput in tok/s:; input / output concurrency TPOT default TPOT fused; 60000 / 600 64 137.05 137.15 +0.1% 391.59 391.60 +0.0%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59668",
   "source_url": "https://github.com/vllm-project/vllm/pull/59668",
   "tg_tokens": 600,
   "tps": "391.59",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 64,
   "ctx": 60000,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59668-mi355x-60k600-c64-fused",
   "model": "DeepSeek V4.1",
   "notes": "DeepSeek V4.1, MI355X TP4 + EP, FP8, fused single-launch DSA decode candidate mask, vllm bench serve output throughput at concurrency 64, mean TPOT 137.15 ms; fused single-kernel mask path; config=dsamask-fused",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "On ROCm, the DeepSeek V4.1 decode candidate mask goes from four kernels per consumer indexer layer to one.; vLLM on MI355X, TP4 + EP, FP8, fused against default. Mean TPOT in ms and output throughput in tok/s:; input / output concurrency TPOT default TPOT fused; 60000 / 600 64 137.05 137.15 +0.1% 391.59 391.60 +0.0%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59668",
   "source_url": "https://github.com/vllm-project/vllm/pull/59668",
   "tg_tokens": 600,
   "tps": "391.60",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 128000,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59668-mi355x-128k1k-c4-default",
   "model": "DeepSeek V4.1",
   "notes": "DeepSeek V4.1, MI355X TP4 + EP, FP8, fused single-launch DSA decode candidate mask, vllm bench serve output throughput at concurrency 4, mean TPOT 16.74 ms; stock four-kernel mask path; config=dsamask-default",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "On ROCm, the DeepSeek V4.1 decode candidate mask goes from four kernels per consumer indexer layer to one.; vLLM on MI355X, TP4 + EP, FP8, fused against default. Mean TPOT in ms and output throughput in tok/s:; input / output concurrency TPOT default TPOT fused; 128000 / 1024 4 16.74 16.68 -0.4% 185.73 186.21 +0.3%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59668",
   "source_url": "https://github.com/vllm-project/vllm/pull/59668",
   "tg_tokens": 1024,
   "tps": "185.73",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 128000,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59668-mi355x-128k1k-c4-fused",
   "model": "DeepSeek V4.1",
   "notes": "DeepSeek V4.1, MI355X TP4 + EP, FP8, fused single-launch DSA decode candidate mask, vllm bench serve output throughput at concurrency 4, mean TPOT 16.68 ms; fused single-kernel mask path; config=dsamask-fused",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "On ROCm, the DeepSeek V4.1 decode candidate mask goes from four kernels per consumer indexer layer to one.; vLLM on MI355X, TP4 + EP, FP8, fused against default. Mean TPOT in ms and output throughput in tok/s:; input / output concurrency TPOT default TPOT fused; 128000 / 1024 4 16.74 16.68 -0.4% 185.73 186.21 +0.3%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59668",
   "source_url": "https://github.com/vllm-project/vllm/pull/59668",
   "tg_tokens": 1024,
   "tps": "186.21",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": 128000,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59668-mi355x-128k1k-c16-default",
   "model": "DeepSeek V4.1",
   "notes": "DeepSeek V4.1, MI355X TP4 + EP, FP8, fused single-launch DSA decode candidate mask, vllm bench serve output throughput at concurrency 16, mean TPOT 53.98 ms; stock four-kernel mask path; config=dsamask-default",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "On ROCm, the DeepSeek V4.1 decode candidate mask goes from four kernels per consumer indexer layer to one.; vLLM on MI355X, TP4 + EP, FP8, fused against default. Mean TPOT in ms and output throughput in tok/s:; input / output concurrency TPOT default TPOT fused; 128000 / 1024 16 53.98 54.17 +0.3% 248.75 248.60 -0.1%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59668",
   "source_url": "https://github.com/vllm-project/vllm/pull/59668",
   "tg_tokens": 1024,
   "tps": "248.75",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": 128000,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59668-mi355x-128k1k-c16-fused",
   "model": "DeepSeek V4.1",
   "notes": "DeepSeek V4.1, MI355X TP4 + EP, FP8, fused single-launch DSA decode candidate mask, vllm bench serve output throughput at concurrency 16, mean TPOT 54.17 ms; fused single-kernel mask path; config=dsamask-fused",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "On ROCm, the DeepSeek V4.1 decode candidate mask goes from four kernels per consumer indexer layer to one.; vLLM on MI355X, TP4 + EP, FP8, fused against default. Mean TPOT in ms and output throughput in tok/s:; input / output concurrency TPOT default TPOT fused; 128000 / 1024 16 53.98 54.17 +0.3% 248.75 248.60 -0.1%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59668",
   "source_url": "https://github.com/vllm-project/vllm/pull/59668",
   "tg_tokens": 1024,
   "tps": "248.60",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "MI350X",
   "id": "vllm-59653-mi350x-dec-c1-off",
   "model": "MiniMax-M3",
   "notes": "MiniMax-M3 MXFP4, 4x MI350X TP4, FP8 KV, fused sparse MoE decode via AITER, decode 8K in / 256 out, output tok/s at concurrency 1, mean TPOT 6.14 ms; unfused baseline layers; config=aiterfused-off",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Adds an opt-in decode path for MiniMax-M3 on ROCm that runs each sparse MoE layer as one AITER kernel launch per GPU; 4x MI350X, amd/MiniMax-M3-MXFP4, TP=4:; Decode, 8K input / 256 output:; 1 6.14 ms 3.15 ms 1.95x 141 254",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59653",
   "source_url": "https://github.com/vllm-project/vllm/pull/59653",
   "tg_tokens": 256,
   "tps": "141",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "MI350X",
   "id": "vllm-59653-mi350x-dec-c1-on",
   "model": "MiniMax-M3",
   "notes": "MiniMax-M3 MXFP4, 4x MI350X TP4, FP8 KV, fused sparse MoE decode via AITER, decode 8K in / 256 out, output tok/s at concurrency 1, mean TPOT 3.15 ms; fused AITER path; config=aiterfused-on",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Adds an opt-in decode path for MiniMax-M3 on ROCm that runs each sparse MoE layer as one AITER kernel launch per GPU; 4x MI350X, amd/MiniMax-M3-MXFP4, TP=4:; Decode, 8K input / 256 output:; 1 6.14 ms 3.15 ms 1.95x 141 254",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59653",
   "source_url": "https://github.com/vllm-project/vllm/pull/59653",
   "tg_tokens": 256,
   "tps": "254",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 2,
   "ctx": 8192,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "MI350X",
   "id": "vllm-59653-mi350x-dec-c2-off",
   "model": "MiniMax-M3",
   "notes": "MiniMax-M3 MXFP4, 4x MI350X TP4, FP8 KV, fused sparse MoE decode via AITER, decode 8K in / 256 out, output tok/s at concurrency 2, mean TPOT 6.57 ms; unfused baseline layers; config=aiterfused-off",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Adds an opt-in decode path for MiniMax-M3 on ROCm that runs each sparse MoE layer as one AITER kernel launch per GPU; 4x MI350X, amd/MiniMax-M3-MXFP4, TP=4:; Decode, 8K input / 256 output:; 2 6.57 ms 3.80 ms 1.73x 254 405",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59653",
   "source_url": "https://github.com/vllm-project/vllm/pull/59653",
   "tg_tokens": 256,
   "tps": "254",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 2,
   "ctx": 8192,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "MI350X",
   "id": "vllm-59653-mi350x-dec-c2-on",
   "model": "MiniMax-M3",
   "notes": "MiniMax-M3 MXFP4, 4x MI350X TP4, FP8 KV, fused sparse MoE decode via AITER, decode 8K in / 256 out, output tok/s at concurrency 2, mean TPOT 3.80 ms; fused AITER path; config=aiterfused-on",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Adds an opt-in decode path for MiniMax-M3 on ROCm that runs each sparse MoE layer as one AITER kernel launch per GPU; 4x MI350X, amd/MiniMax-M3-MXFP4, TP=4:; Decode, 8K input / 256 output:; 2 6.57 ms 3.80 ms 1.73x 254 405",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59653",
   "source_url": "https://github.com/vllm-project/vllm/pull/59653",
   "tg_tokens": 256,
   "tps": "405",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 8192,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "MI350X",
   "id": "vllm-59653-mi350x-dec-c4-off",
   "model": "MiniMax-M3",
   "notes": "MiniMax-M3 MXFP4, 4x MI350X TP4, FP8 KV, fused sparse MoE decode via AITER, decode 8K in / 256 out, output tok/s at concurrency 4, mean TPOT 7.24 ms; unfused baseline layers; config=aiterfused-off",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Adds an opt-in decode path for MiniMax-M3 on ROCm that runs each sparse MoE layer as one AITER kernel launch per GPU; 4x MI350X, amd/MiniMax-M3-MXFP4, TP=4:; Decode, 8K input / 256 output:; 4 7.24 ms 4.61 ms 1.57x 415 574",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59653",
   "source_url": "https://github.com/vllm-project/vllm/pull/59653",
   "tg_tokens": 256,
   "tps": "415",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 8192,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "MI350X",
   "id": "vllm-59653-mi350x-dec-c4-on",
   "model": "MiniMax-M3",
   "notes": "MiniMax-M3 MXFP4, 4x MI350X TP4, FP8 KV, fused sparse MoE decode via AITER, decode 8K in / 256 out, output tok/s at concurrency 4, mean TPOT 4.61 ms; fused AITER path; config=aiterfused-on",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Adds an opt-in decode path for MiniMax-M3 on ROCm that runs each sparse MoE layer as one AITER kernel launch per GPU; 4x MI350X, amd/MiniMax-M3-MXFP4, TP=4:; Decode, 8K input / 256 output:; 4 7.24 ms 4.61 ms 1.57x 415 574",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59653",
   "source_url": "https://github.com/vllm-project/vllm/pull/59653",
   "tg_tokens": 256,
   "tps": "574",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 8192,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "MI350X",
   "id": "vllm-59653-mi350x-dec-c8-off",
   "model": "MiniMax-M3",
   "notes": "MiniMax-M3 MXFP4, 4x MI350X TP4, FP8 KV, fused sparse MoE decode via AITER, decode 8K in / 256 out, output tok/s at concurrency 8, mean TPOT 9.03 ms; unfused baseline layers; config=aiterfused-off",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Adds an opt-in decode path for MiniMax-M3 on ROCm that runs each sparse MoE layer as one AITER kernel launch per GPU; 4x MI350X, amd/MiniMax-M3-MXFP4, TP=4:; Decode, 8K input / 256 output:; 8 9.03 ms 6.90 ms 1.31x 599 682",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59653",
   "source_url": "https://github.com/vllm-project/vllm/pull/59653",
   "tg_tokens": 256,
   "tps": "599",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 8192,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "MI350X",
   "id": "vllm-59653-mi350x-dec-c8-on",
   "model": "MiniMax-M3",
   "notes": "MiniMax-M3 MXFP4, 4x MI350X TP4, FP8 KV, fused sparse MoE decode via AITER, decode 8K in / 256 out, output tok/s at concurrency 8, mean TPOT 6.90 ms; fused AITER path; config=aiterfused-on",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Adds an opt-in decode path for MiniMax-M3 on ROCm that runs each sparse MoE layer as one AITER kernel launch per GPU; 4x MI350X, amd/MiniMax-M3-MXFP4, TP=4:; Decode, 8K input / 256 output:; 8 9.03 ms 6.90 ms 1.31x 599 682",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59653",
   "source_url": "https://github.com/vllm-project/vllm/pull/59653",
   "tg_tokens": 256,
   "tps": "682",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": 8192,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "MI350X",
   "id": "vllm-59653-mi350x-dec-c16-off",
   "model": "MiniMax-M3",
   "notes": "MiniMax-M3 MXFP4, 4x MI350X TP4, FP8 KV, fused sparse MoE decode via AITER, decode 8K in / 256 out, output tok/s at concurrency 16, mean TPOT 12.47 ms; unfused baseline layers; config=aiterfused-off",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Adds an opt-in decode path for MiniMax-M3 on ROCm that runs each sparse MoE layer as one AITER kernel launch per GPU; 4x MI350X, amd/MiniMax-M3-MXFP4, TP=4:; Decode, 8K input / 256 output:; 16 12.47 ms 11.10 ms 1.12x 825 910",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59653",
   "source_url": "https://github.com/vllm-project/vllm/pull/59653",
   "tg_tokens": 256,
   "tps": "825",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": 8192,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "MI350X",
   "id": "vllm-59653-mi350x-dec-c16-on",
   "model": "MiniMax-M3",
   "notes": "MiniMax-M3 MXFP4, 4x MI350X TP4, FP8 KV, fused sparse MoE decode via AITER, decode 8K in / 256 out, output tok/s at concurrency 16, mean TPOT 11.10 ms; fused AITER path; config=aiterfused-on",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Adds an opt-in decode path for MiniMax-M3 on ROCm that runs each sparse MoE layer as one AITER kernel launch per GPU; 4x MI350X, amd/MiniMax-M3-MXFP4, TP=4:; Decode, 8K input / 256 output:; 16 12.47 ms 11.10 ms 1.12x 825 910",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59653",
   "source_url": "https://github.com/vllm-project/vllm/pull/59653",
   "tg_tokens": 256,
   "tps": "910",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 2,
   "ctx": 131072,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "MI350X",
   "id": "vllm-59653-mi350x-srv-c2-off",
   "model": "MiniMax-M3",
   "notes": "MiniMax-M3 MXFP4, 4x MI350X TP4, FP8 KV, fused sparse MoE decode via AITER, serving 128K in / 1K out, output tok/s at concurrency 2, mean TPOT 8.08 ms; unfused baseline layers; config=aiterfused-off",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Adds an opt-in decode path for MiniMax-M3 on ROCm that runs each sparse MoE layer as one AITER kernel launch per GPU; 4x MI350X, amd/MiniMax-M3-MXFP4, TP=4:; Serving, 128K input / 1K output:; 2 8.08 ms 5.72 ms 157 195 +24%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59653",
   "source_url": "https://github.com/vllm-project/vllm/pull/59653",
   "tg_tokens": 1024,
   "tps": "157",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 2,
   "ctx": 131072,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "MI350X",
   "id": "vllm-59653-mi350x-srv-c2-on",
   "model": "MiniMax-M3",
   "notes": "MiniMax-M3 MXFP4, 4x MI350X TP4, FP8 KV, fused sparse MoE decode via AITER, serving 128K in / 1K out, output tok/s at concurrency 2, mean TPOT 5.72 ms; fused AITER path, +24%; config=aiterfused-on",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Adds an opt-in decode path for MiniMax-M3 on ROCm that runs each sparse MoE layer as one AITER kernel launch per GPU; 4x MI350X, amd/MiniMax-M3-MXFP4, TP=4:; Serving, 128K input / 1K output:; 2 8.08 ms 5.72 ms 157 195 +24%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59653",
   "source_url": "https://github.com/vllm-project/vllm/pull/59653",
   "tg_tokens": 1024,
   "tps": "195",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 131072,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "MI350X",
   "id": "vllm-59653-mi350x-srv-c4-off",
   "model": "MiniMax-M3",
   "notes": "MiniMax-M3 MXFP4, 4x MI350X TP4, FP8 KV, fused sparse MoE decode via AITER, serving 128K in / 1K out, output tok/s at concurrency 4, mean TPOT 11.53 ms; unfused baseline layers; config=aiterfused-off",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Adds an opt-in decode path for MiniMax-M3 on ROCm that runs each sparse MoE layer as one AITER kernel launch per GPU; 4x MI350X, amd/MiniMax-M3-MXFP4, TP=4:; Serving, 128K input / 1K output:; 4 11.53 ms 9.04 ms 206 243 +17%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59653",
   "source_url": "https://github.com/vllm-project/vllm/pull/59653",
   "tg_tokens": 1024,
   "tps": "206",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 131072,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "MI350X",
   "id": "vllm-59653-mi350x-srv-c4-on",
   "model": "MiniMax-M3",
   "notes": "MiniMax-M3 MXFP4, 4x MI350X TP4, FP8 KV, fused sparse MoE decode via AITER, serving 128K in / 1K out, output tok/s at concurrency 4, mean TPOT 9.04 ms; fused AITER path, +17%; config=aiterfused-on",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Adds an opt-in decode path for MiniMax-M3 on ROCm that runs each sparse MoE layer as one AITER kernel launch per GPU; 4x MI350X, amd/MiniMax-M3-MXFP4, TP=4:; Serving, 128K input / 1K output:; 4 11.53 ms 9.04 ms 206 243 +17%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59653",
   "source_url": "https://github.com/vllm-project/vllm/pull/59653",
   "tg_tokens": 1024,
   "tps": "243",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 131072,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "MI350X",
   "id": "vllm-59653-mi350x-srv-c8-off",
   "model": "MiniMax-M3",
   "notes": "MiniMax-M3 MXFP4, 4x MI350X TP4, FP8 KV, fused sparse MoE decode via AITER, serving 128K in / 1K out, output tok/s at concurrency 8, mean TPOT 18.88 ms; unfused baseline layers; config=aiterfused-off",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Adds an opt-in decode path for MiniMax-M3 on ROCm that runs each sparse MoE layer as one AITER kernel launch per GPU; 4x MI350X, amd/MiniMax-M3-MXFP4, TP=4:; Serving, 128K input / 1K output:; 8 18.88 ms 16.35 ms 246 273 +11%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59653",
   "source_url": "https://github.com/vllm-project/vllm/pull/59653",
   "tg_tokens": 1024,
   "tps": "246",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 131072,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "MI350X",
   "id": "vllm-59653-mi350x-srv-c8-on",
   "model": "MiniMax-M3",
   "notes": "MiniMax-M3 MXFP4, 4x MI350X TP4, FP8 KV, fused sparse MoE decode via AITER, serving 128K in / 1K out, output tok/s at concurrency 8, mean TPOT 16.35 ms; fused AITER path, +11%; config=aiterfused-on",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Adds an opt-in decode path for MiniMax-M3 on ROCm that runs each sparse MoE layer as one AITER kernel launch per GPU; 4x MI350X, amd/MiniMax-M3-MXFP4, TP=4:; Serving, 128K input / 1K output:; 8 18.88 ms 16.35 ms 246 273 +11%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59653",
   "source_url": "https://github.com/vllm-project/vllm/pull/59653",
   "tg_tokens": 1024,
   "tps": "273",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 256,
   "ctx": 1024,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-59567-gb300-p1024-base",
   "model": "Qwen3 8B",
   "notes": "Qwen3-8B on one GB300, batch sampling-mask transport PR, non-streaming /inference/v1/generate, c=256, median of 2 measured passes, output tok/s with --return-sampling-mask at 1024/1024; no sampling mask requested (baseline); config=mask-base",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "Qwen3-8B, one GB300 GPU per server. temperature 1.0, top_k 50, top_p 0.9 (wide: top_p 1.0); base = this commit, no --return-sampling-mask; stock = the unmodified nightly image; branch = this commit with --return-sampling-mask; 1024/1024 c=256 1.7 16515 15705 -4.9% 16364 -0.9%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59567",
   "source_url": "https://github.com/vllm-project/vllm/pull/59567",
   "tg_tokens": 1024,
   "tps": "16515",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 256,
   "ctx": 1024,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-59567-gb300-p1024-stock",
   "model": "Qwen3 8B",
   "notes": "Qwen3-8B on one GB300, batch sampling-mask transport PR, non-streaming /inference/v1/generate, c=256, median of 2 measured passes, output tok/s with --return-sampling-mask at 1024/1024; stock nightly mask path, -4.9%; config=mask-stock",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "Qwen3-8B, one GB300 GPU per server. temperature 1.0, top_k 50, top_p 0.9 (wide: top_p 1.0); base = this commit, no --return-sampling-mask; stock = the unmodified nightly image; branch = this commit with --return-sampling-mask; 1024/1024 c=256 1.7 16515 15705 -4.9% 16364 -0.9%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59567",
   "source_url": "https://github.com/vllm-project/vllm/pull/59567",
   "tg_tokens": 1024,
   "tps": "15705",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 256,
   "ctx": 1024,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-59567-gb300-p1024-branch",
   "model": "Qwen3 8B",
   "notes": "Qwen3-8B on one GB300, batch sampling-mask transport PR, non-streaming /inference/v1/generate, c=256, median of 2 measured passes, output tok/s with --return-sampling-mask at 1024/1024; batched mask transport, -0.9%; config=mask-branch",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "Qwen3-8B, one GB300 GPU per server. temperature 1.0, top_k 50, top_p 0.9 (wide: top_p 1.0); base = this commit, no --return-sampling-mask; stock = the unmodified nightly image; branch = this commit with --return-sampling-mask; 1024/1024 c=256 1.7 16515 15705 -4.9% 16364 -0.9%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59567",
   "source_url": "https://github.com/vllm-project/vllm/pull/59567",
   "tg_tokens": 1024,
   "tps": "16364",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 256,
   "ctx": 256,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-59567-gb300-p256-base",
   "model": "Qwen3 8B",
   "notes": "Qwen3-8B on one GB300, batch sampling-mask transport PR, non-streaming /inference/v1/generate, c=256, median of 2 measured passes, output tok/s with --return-sampling-mask at 256/256; no sampling mask requested (baseline); config=mask-base",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "Qwen3-8B, one GB300 GPU per server. temperature 1.0, top_k 50, top_p 0.9 (wide: top_p 1.0); base = this commit, no --return-sampling-mask; stock = the unmodified nightly image; branch = this commit with --return-sampling-mask; 256/256 c=256 3.0 25515 23488 -7.9% 24811 -2.8%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59567",
   "source_url": "https://github.com/vllm-project/vllm/pull/59567",
   "tg_tokens": 256,
   "tps": "25515",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 256,
   "ctx": 256,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-59567-gb300-p256-stock",
   "model": "Qwen3 8B",
   "notes": "Qwen3-8B on one GB300, batch sampling-mask transport PR, non-streaming /inference/v1/generate, c=256, median of 2 measured passes, output tok/s with --return-sampling-mask at 256/256; stock nightly mask path, -7.9%; config=mask-stock",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "Qwen3-8B, one GB300 GPU per server. temperature 1.0, top_k 50, top_p 0.9 (wide: top_p 1.0); base = this commit, no --return-sampling-mask; stock = the unmodified nightly image; branch = this commit with --return-sampling-mask; 256/256 c=256 3.0 25515 23488 -7.9% 24811 -2.8%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59567",
   "source_url": "https://github.com/vllm-project/vllm/pull/59567",
   "tg_tokens": 256,
   "tps": "23488",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 256,
   "ctx": 256,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-59567-gb300-p256-branch",
   "model": "Qwen3 8B",
   "notes": "Qwen3-8B on one GB300, batch sampling-mask transport PR, non-streaming /inference/v1/generate, c=256, median of 2 measured passes, output tok/s with --return-sampling-mask at 256/256; batched mask transport, -2.8%; config=mask-branch",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "Qwen3-8B, one GB300 GPU per server. temperature 1.0, top_k 50, top_p 0.9 (wide: top_p 1.0); base = this commit, no --return-sampling-mask; stock = the unmodified nightly image; branch = this commit with --return-sampling-mask; 256/256 c=256 3.0 25515 23488 -7.9% 24811 -2.8%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59567",
   "source_url": "https://github.com/vllm-project/vllm/pull/59567",
   "tg_tokens": 256,
   "tps": "24811",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 256,
   "ctx": 256,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-59567-gb300-p256wide-base",
   "model": "Qwen3 8B",
   "notes": "Qwen3-8B on one GB300, batch sampling-mask transport PR, non-streaming /inference/v1/generate, c=256, median of 2 measured passes, output tok/s with --return-sampling-mask at 256/256 (wide top_p=1.0, 50 ids/token); no sampling mask requested (baseline); config=mask-base",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "Qwen3-8B, one GB300 GPU per server. temperature 1.0, top_k 50, top_p 0.9 (wide: top_p 1.0); base = this commit, no --return-sampling-mask; stock = the unmodified nightly image; branch = this commit with --return-sampling-mask; 256/256 c=256, wide (top_p=1.0) 50 25781 21249 -17.6% 22873 -11.3%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59567",
   "source_url": "https://github.com/vllm-project/vllm/pull/59567",
   "tg_tokens": 256,
   "tps": "25781",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 256,
   "ctx": 256,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-59567-gb300-p256wide-stock",
   "model": "Qwen3 8B",
   "notes": "Qwen3-8B on one GB300, batch sampling-mask transport PR, non-streaming /inference/v1/generate, c=256, median of 2 measured passes, output tok/s with --return-sampling-mask at 256/256 (wide top_p=1.0, 50 ids/token); stock nightly mask path, -17.6%; config=mask-stock",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "Qwen3-8B, one GB300 GPU per server. temperature 1.0, top_k 50, top_p 0.9 (wide: top_p 1.0); base = this commit, no --return-sampling-mask; stock = the unmodified nightly image; branch = this commit with --return-sampling-mask; 256/256 c=256, wide (top_p=1.0) 50 25781 21249 -17.6% 22873 -11.3%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59567",
   "source_url": "https://github.com/vllm-project/vllm/pull/59567",
   "tg_tokens": 256,
   "tps": "21249",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 256,
   "ctx": 256,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-59567-gb300-p256wide-branch",
   "model": "Qwen3 8B",
   "notes": "Qwen3-8B on one GB300, batch sampling-mask transport PR, non-streaming /inference/v1/generate, c=256, median of 2 measured passes, output tok/s with --return-sampling-mask at 256/256 (wide top_p=1.0, 50 ids/token); batched mask transport, -11.3%; config=mask-branch",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "Qwen3-8B, one GB300 GPU per server. temperature 1.0, top_k 50, top_p 0.9 (wide: top_p 1.0); base = this commit, no --return-sampling-mask; stock = the unmodified nightly image; branch = this commit with --return-sampling-mask; 256/256 c=256, wide (top_p=1.0) 50 25781 21249 -17.6% 22873 -11.3%",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM PR #59567",
   "source_url": "https://github.com/vllm-project/vllm/pull/59567",
   "tg_tokens": 256,
   "tps": "22873",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "H200",
   "id": "vllm-59520-h200-sparse-cuda",
   "model": "Qwen3.5-9B",
   "notes": "Qwen3.5-9B BF16 on 2x H200 TP2, non-spec decode throughput bench, 200 prefix-repetition requests, 204,800 output tokens per arm, prefix cache retention Sparse (retention 0); default CUDA GDN wrapper (regressed path); config=gdn-cuda",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "The original two-H200 sequential comparison found that selecting the standard Triton path; Every arm used Qwen/Qwen3.5-9B revision; BF16, Model Runner V2, seed 42; Sparse (`0`) | 1,426.638 | 1,859.655 | **+30.35%**",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM issue #59520",
   "source_url": "https://github.com/vllm-project/vllm/issues/59520",
   "tg_tokens": null,
   "tps": "1426.638",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "H200",
   "id": "vllm-59520-h200-sparse-triton",
   "model": "Qwen3.5-9B",
   "notes": "Qwen3.5-9B BF16 on 2x H200 TP2, non-spec decode throughput bench, 200 prefix-repetition requests, 204,800 output tokens per arm, prefix cache retention Sparse (retention 0); Triton GDN decode path, +30.35% vs CUDA; config=sparse",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "The original two-H200 sequential comparison found that selecting the standard Triton path; Every arm used Qwen/Qwen3.5-9B revision; BF16, Model Runner V2, seed 42; Sparse (`0`) | 1,426.638 | 1,859.655 | **+30.35%**",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM issue #59520",
   "source_url": "https://github.com/vllm-project/vllm/issues/59520",
   "tg_tokens": null,
   "tps": "1859.655",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "H200",
   "id": "vllm-59520-h200-dense-cuda",
   "model": "Qwen3.5-9B",
   "notes": "Qwen3.5-9B BF16 on 2x H200 TP2, non-spec decode throughput bench, 200 prefix-repetition requests, 204,800 output tokens per arm, prefix cache retention Dense (retention None); default CUDA GDN wrapper (regressed path); config=gdn-cuda",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "The original two-H200 sequential comparison found that selecting the standard Triton path; Every arm used Qwen/Qwen3.5-9B revision; BF16, Model Runner V2, seed 42; Dense (`None`) | 1,536.536 | 2,069.343 | **+34.68%**",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM issue #59520",
   "source_url": "https://github.com/vllm-project/vllm/issues/59520",
   "tg_tokens": null,
   "tps": "1536.536",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "H200",
   "id": "vllm-59520-h200-dense-triton",
   "model": "Qwen3.5-9B",
   "notes": "Qwen3.5-9B BF16 on 2x H200 TP2, non-spec decode throughput bench, 200 prefix-repetition requests, 204,800 output tokens per arm, prefix cache retention Dense (retention None); Triton GDN decode path, +34.68% vs CUDA; config=dense",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "The original two-H200 sequential comparison found that selecting the standard Triton path; Every arm used Qwen/Qwen3.5-9B revision; BF16, Model Runner V2, seed 42; Dense (`None`) | 1,536.536 | 2,069.343 | **+34.68%**",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM issue #59520",
   "source_url": "https://github.com/vllm-project/vllm/issues/59520",
   "tg_tokens": null,
   "tps": "2069.343",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [
    "contradiction"
   ],
   "hardware": "L4",
   "id": "vllm-59548-l4-a-default",
   "model": "Qwen3 4B",
   "notes": "Qwen3-4B + DFlash-b16 drafter, 5 spec tokens, GSM8K prompt set replayed, concurrency 1, vLLM 0.29.0, container session A, mean of 12 boots; CUDA-graph default path, boot-to-boot CV 13.92%, eager ratio 0.487; config=0290-default",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "Qwen3-4B target, `z-lab/Qwen3-4B-DFlash-b16` drafter, 5 speculative tokens, bf16 weights; 12 boots per arm inside a single container, one physical GPU per session; Throughput at concurrency 1, output tokens/s:; | L4 (SM89) | 0.29.0 | A | 39.2 | 80.5 | **0.487** | **13.92%** |",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM issue #59548",
   "source_url": "https://github.com/vllm-project/vllm/issues/59548",
   "tg_tokens": null,
   "tps": "39.2",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "L4",
   "id": "vllm-59548-l4-a-eager",
   "model": "Qwen3 4B",
   "notes": "Qwen3-4B + DFlash-b16 drafter, 5 spec tokens, GSM8K prompt set replayed, concurrency 1, vLLM 0.29.0, container session A, mean of 12 boots; --enforce-eager control arm (stable across boots); config=0290-eager",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "Qwen3-4B target, `z-lab/Qwen3-4B-DFlash-b16` drafter, 5 speculative tokens, bf16 weights; 12 boots per arm inside a single container, one physical GPU per session; Throughput at concurrency 1, output tokens/s:; | L4 (SM89) | 0.29.0 | A | 39.2 | 80.5 | **0.487** | **13.92%** |",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM issue #59548",
   "source_url": "https://github.com/vllm-project/vllm/issues/59548",
   "tg_tokens": null,
   "tps": "80.5",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "L4",
   "id": "vllm-59548-l4-b-default",
   "model": "Qwen3 4B",
   "notes": "Qwen3-4B + DFlash-b16 drafter, 5 spec tokens, GSM8K prompt set replayed, concurrency 1, vLLM 0.29.0, container session B, mean of 12 boots; CUDA-graph default path, boot-to-boot CV 3.73%, eager ratio 0.768; config=0290-default",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "Qwen3-4B target, `z-lab/Qwen3-4B-DFlash-b16` drafter, 5 speculative tokens, bf16 weights; 12 boots per arm inside a single container, one physical GPU per session; Throughput at concurrency 1, output tokens/s:; | L4 (SM89) | 0.29.0 | B | 61.2 | 79.7 | **0.768** | **3.73%** |",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM issue #59548",
   "source_url": "https://github.com/vllm-project/vllm/issues/59548",
   "tg_tokens": null,
   "tps": "61.2",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "L4",
   "id": "vllm-59548-l4-b-eager",
   "model": "Qwen3 4B",
   "notes": "Qwen3-4B + DFlash-b16 drafter, 5 spec tokens, GSM8K prompt set replayed, concurrency 1, vLLM 0.29.0, container session B, mean of 12 boots; --enforce-eager control arm (stable across boots); config=0290-eager",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "Qwen3-4B target, `z-lab/Qwen3-4B-DFlash-b16` drafter, 5 speculative tokens, bf16 weights; 12 boots per arm inside a single container, one physical GPU per session; Throughput at concurrency 1, output tokens/s:; | L4 (SM89) | 0.29.0 | B | 61.2 | 79.7 | **0.768** | **3.73%** |",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM issue #59548",
   "source_url": "https://github.com/vllm-project/vllm/issues/59548",
   "tg_tokens": null,
   "tps": "79.7",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "L4",
   "id": "vllm-59548-l4-c-default",
   "model": "Qwen3 4B",
   "notes": "Qwen3-4B + DFlash-b16 drafter, 5 spec tokens, GSM8K prompt set replayed, concurrency 1, vLLM 0.30.0, container session C, mean of 12 boots; CUDA-graph default path, boot-to-boot CV 0.99%, eager ratio 1.014; config=0300-default",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "Qwen3-4B target, `z-lab/Qwen3-4B-DFlash-b16` drafter, 5 speculative tokens, bf16 weights; 12 boots per arm inside a single container, one physical GPU per session; Throughput at concurrency 1, output tokens/s:; | L4 (SM89) | 0.30.0 | C | 82.8 | 81.6 | **1.014** | 0.99% |",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM issue #59548",
   "source_url": "https://github.com/vllm-project/vllm/issues/59548",
   "tg_tokens": null,
   "tps": "82.8",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "L4",
   "id": "vllm-59548-l4-c-eager",
   "model": "Qwen3 4B",
   "notes": "Qwen3-4B + DFlash-b16 drafter, 5 spec tokens, GSM8K prompt set replayed, concurrency 1, vLLM 0.30.0, container session C, mean of 12 boots; --enforce-eager control arm (stable across boots); config=0300-eager",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "Qwen3-4B target, `z-lab/Qwen3-4B-DFlash-b16` drafter, 5 speculative tokens, bf16 weights; 12 boots per arm inside a single container, one physical GPU per session; Throughput at concurrency 1, output tokens/s:; | L4 (SM89) | 0.30.0 | C | 82.8 | 81.6 | **1.014** | 0.99% |",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM issue #59548",
   "source_url": "https://github.com/vllm-project/vllm/issues/59548",
   "tg_tokens": null,
   "tps": "81.6",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "L4",
   "id": "vllm-59548-l4-d-default",
   "model": "Qwen3 4B",
   "notes": "Qwen3-4B + DFlash-b16 drafter, 5 spec tokens, GSM8K prompt set replayed, concurrency 1, vLLM 0.30.0, container session D, mean of 12 boots; CUDA-graph default path, boot-to-boot CV 1.61%, eager ratio 1.017; config=0300-default",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "Qwen3-4B target, `z-lab/Qwen3-4B-DFlash-b16` drafter, 5 speculative tokens, bf16 weights; 12 boots per arm inside a single container, one physical GPU per session; Throughput at concurrency 1, output tokens/s:; | L4 (SM89) | 0.30.0 | D | 78.2 | 77.0 | **1.017** | 1.61% |",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM issue #59548",
   "source_url": "https://github.com/vllm-project/vllm/issues/59548",
   "tg_tokens": null,
   "tps": "78.2",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "L4",
   "id": "vllm-59548-l4-d-eager",
   "model": "Qwen3 4B",
   "notes": "Qwen3-4B + DFlash-b16 drafter, 5 spec tokens, GSM8K prompt set replayed, concurrency 1, vLLM 0.30.0, container session D, mean of 12 boots; --enforce-eager control arm (stable across boots); config=0300-eager",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "Qwen3-4B target, `z-lab/Qwen3-4B-DFlash-b16` drafter, 5 speculative tokens, bf16 weights; 12 boots per arm inside a single container, one physical GPU per session; Throughput at concurrency 1, output tokens/s:; | L4 (SM89) | 0.30.0 | D | 78.2 | 77.0 | **1.017** | 1.61% |",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM issue #59548",
   "source_url": "https://github.com/vllm-project/vllm/issues/59548",
   "tg_tokens": null,
   "tps": "77.0",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "A10",
   "id": "vllm-59548-a10-e-default",
   "model": "Qwen3 4B",
   "notes": "Qwen3-4B + DFlash-b16 drafter, 5 spec tokens, GSM8K prompt set replayed, concurrency 1, vLLM 0.29.0, container session E, mean of 12 boots; CUDA-graph default path, boot-to-boot CV 2.08%, eager ratio 0.959; config=0290-default",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "Qwen3-4B target, `z-lab/Qwen3-4B-DFlash-b16` drafter, 5 speculative tokens, bf16 weights; 12 boots per arm inside a single container, one physical GPU per session; Throughput at concurrency 1, output tokens/s:; | A10 (SM86) | 0.29.0 | E | 132.6 | 138.3 | 0.959 | 2.08% |",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM issue #59548",
   "source_url": "https://github.com/vllm-project/vllm/issues/59548",
   "tg_tokens": null,
   "tps": "132.6",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "A10",
   "id": "vllm-59548-a10-e-eager",
   "model": "Qwen3 4B",
   "notes": "Qwen3-4B + DFlash-b16 drafter, 5 spec tokens, GSM8K prompt set replayed, concurrency 1, vLLM 0.29.0, container session E, mean of 12 boots; --enforce-eager control arm (stable across boots); config=0290-eager",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "Qwen3-4B target, `z-lab/Qwen3-4B-DFlash-b16` drafter, 5 speculative tokens, bf16 weights; 12 boots per arm inside a single container, one physical GPU per session; Throughput at concurrency 1, output tokens/s:; | A10 (SM86) | 0.29.0 | E | 132.6 | 138.3 | 0.959 | 2.08% |",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "vLLM issue #59548",
   "source_url": "https://github.com/vllm-project/vllm/issues/59548",
   "tg_tokens": null,
   "tps": "138.3",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-59655-gb300-a-c1-v030",
   "model": "DeepSeek V4.1",
   "notes": "DeepSeek V4.1 Flash on GB300 TP4, DSpark k=5 speculative decoding, MT-Bench output tok/s, node A, concurrency 1, acceptance length 1.00 / 2.87 / 2.81; unmodified v0.30.0 serves the NVFP4 checkpoint with the draft experts mis-quantized (acceptance length 1.00, every draft token rejected); PR vs v0.30.0 2.19x, PR vs native -0.9%; config=A-c1-v030",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "All runs used vLLM v0.30.0 on GB300 (TP4), with this PR's quant_config.py mounted over the installed one; MT-Bench serving (80 prompts, greedy, TP4; 0 failed requests); A 1 1.00 / 2.87 / 2.81 114 / 250 / 252 2.19x -0.9%; \"v0.30.0\" and \"PR\" both serve nvidia/DeepSeek-V4.1-Flash-NVFP4 ; \"native\" is deepseek-ai/DeepSeek-V4.1-Flash on v0.30.0.",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59655",
   "source_url": "https://github.com/vllm-project/vllm/pull/59655",
   "tg_tokens": null,
   "tps": "114",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-59655-gb300-a-c1-pr",
   "model": "DeepSeek V4.1",
   "notes": "DeepSeek V4.1 Flash on GB300 TP4, DSpark k=5 speculative decoding, MT-Bench output tok/s, node A, concurrency 1, acceptance length 1.00 / 2.87 / 2.81; v0.30.0 with the fix: ignored mtp.* experts keep MXFP4, draft works; PR vs v0.30.0 2.19x, PR vs native -0.9%; config=A-c1-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "All runs used vLLM v0.30.0 on GB300 (TP4), with this PR's quant_config.py mounted over the installed one; MT-Bench serving (80 prompts, greedy, TP4; 0 failed requests); A 1 1.00 / 2.87 / 2.81 114 / 250 / 252 2.19x -0.9%; \"v0.30.0\" and \"PR\" both serve nvidia/DeepSeek-V4.1-Flash-NVFP4 ; \"native\" is deepseek-ai/DeepSeek-V4.1-Flash on v0.30.0.",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59655",
   "source_url": "https://github.com/vllm-project/vllm/pull/59655",
   "tg_tokens": null,
   "tps": "250",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-59655-gb300-a-c1-native",
   "model": "DeepSeek V4.1",
   "notes": "DeepSeek V4.1 Flash on GB300 TP4, DSpark k=5 speculative decoding, MT-Bench output tok/s, node A, concurrency 1, acceptance length 1.00 / 2.87 / 2.81; native deepseek-ai/DeepSeek-V4.1-Flash checkpoint on v0.30.0; PR vs v0.30.0 2.19x, PR vs native -0.9%; config=A-c1-native",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "All runs used vLLM v0.30.0 on GB300 (TP4), with this PR's quant_config.py mounted over the installed one; MT-Bench serving (80 prompts, greedy, TP4; 0 failed requests); A 1 1.00 / 2.87 / 2.81 114 / 250 / 252 2.19x -0.9%; \"v0.30.0\" and \"PR\" both serve nvidia/DeepSeek-V4.1-Flash-NVFP4 ; \"native\" is deepseek-ai/DeepSeek-V4.1-Flash on v0.30.0.",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59655",
   "source_url": "https://github.com/vllm-project/vllm/pull/59655",
   "tg_tokens": null,
   "tps": "252",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 32768,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-59655-gb300-a-c8-v030",
   "model": "DeepSeek V4.1",
   "notes": "DeepSeek V4.1 Flash on GB300 TP4, DSpark k=5 speculative decoding, MT-Bench output tok/s, node A, concurrency 8, acceptance length 1.00 / 2.85 / 2.83; unmodified v0.30.0 serves the NVFP4 checkpoint with the draft experts mis-quantized (acceptance length 1.00, every draft token rejected); PR vs v0.30.0 1.90x, PR vs native +0.4%; config=A-c8-v030",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "All runs used vLLM v0.30.0 on GB300 (TP4), with this PR's quant_config.py mounted over the installed one; MT-Bench serving (80 prompts, greedy, TP4; 0 failed requests); A 8 1.00 / 2.85 / 2.83 773 / 1471 / 1465 1.90x +0.4%; \"v0.30.0\" and \"PR\" both serve nvidia/DeepSeek-V4.1-Flash-NVFP4 ; \"native\" is deepseek-ai/DeepSeek-V4.1-Flash on v0.30.0.",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59655",
   "source_url": "https://github.com/vllm-project/vllm/pull/59655",
   "tg_tokens": null,
   "tps": "773",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 32768,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-59655-gb300-a-c8-pr",
   "model": "DeepSeek V4.1",
   "notes": "DeepSeek V4.1 Flash on GB300 TP4, DSpark k=5 speculative decoding, MT-Bench output tok/s, node A, concurrency 8, acceptance length 1.00 / 2.85 / 2.83; v0.30.0 with the fix: ignored mtp.* experts keep MXFP4, draft works; PR vs v0.30.0 1.90x, PR vs native +0.4%; config=A-c8-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "All runs used vLLM v0.30.0 on GB300 (TP4), with this PR's quant_config.py mounted over the installed one; MT-Bench serving (80 prompts, greedy, TP4; 0 failed requests); A 8 1.00 / 2.85 / 2.83 773 / 1471 / 1465 1.90x +0.4%; \"v0.30.0\" and \"PR\" both serve nvidia/DeepSeek-V4.1-Flash-NVFP4 ; \"native\" is deepseek-ai/DeepSeek-V4.1-Flash on v0.30.0.",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59655",
   "source_url": "https://github.com/vllm-project/vllm/pull/59655",
   "tg_tokens": null,
   "tps": "1471",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 32768,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-59655-gb300-a-c8-native",
   "model": "DeepSeek V4.1",
   "notes": "DeepSeek V4.1 Flash on GB300 TP4, DSpark k=5 speculative decoding, MT-Bench output tok/s, node A, concurrency 8, acceptance length 1.00 / 2.85 / 2.83; native deepseek-ai/DeepSeek-V4.1-Flash checkpoint on v0.30.0; PR vs v0.30.0 1.90x, PR vs native +0.4%; config=A-c8-native",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "All runs used vLLM v0.30.0 on GB300 (TP4), with this PR's quant_config.py mounted over the installed one; MT-Bench serving (80 prompts, greedy, TP4; 0 failed requests); A 8 1.00 / 2.85 / 2.83 773 / 1471 / 1465 1.90x +0.4%; \"v0.30.0\" and \"PR\" both serve nvidia/DeepSeek-V4.1-Flash-NVFP4 ; \"native\" is deepseek-ai/DeepSeek-V4.1-Flash on v0.30.0.",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59655",
   "source_url": "https://github.com/vllm-project/vllm/pull/59655",
   "tg_tokens": null,
   "tps": "1465",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-59655-gb300-b-c1-v030",
   "model": "DeepSeek V4.1",
   "notes": "DeepSeek V4.1 Flash on GB300 TP4, DSpark k=5 speculative decoding, MT-Bench output tok/s, node B, concurrency 1, acceptance length 1.00 / 2.86 / 2.81; unmodified v0.30.0 serves the NVFP4 checkpoint with the draft experts mis-quantized (acceptance length 1.00, every draft token rejected); PR vs v0.30.0 2.39x, PR vs native -13.1%; config=B-c1-v030",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "All runs used vLLM v0.30.0 on GB300 (TP4), with this PR's quant_config.py mounted over the installed one; MT-Bench serving (80 prompts, greedy, TP4; 0 failed requests); B 1 1.00 / 2.86 / 2.81 114 / 273 / 314 2.39x -13.1%; \"v0.30.0\" and \"PR\" both serve nvidia/DeepSeek-V4.1-Flash-NVFP4 ; \"native\" is deepseek-ai/DeepSeek-V4.1-Flash on v0.30.0.",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59655",
   "source_url": "https://github.com/vllm-project/vllm/pull/59655",
   "tg_tokens": null,
   "tps": "114",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-59655-gb300-b-c1-pr",
   "model": "DeepSeek V4.1",
   "notes": "DeepSeek V4.1 Flash on GB300 TP4, DSpark k=5 speculative decoding, MT-Bench output tok/s, node B, concurrency 1, acceptance length 1.00 / 2.86 / 2.81; v0.30.0 with the fix: ignored mtp.* experts keep MXFP4, draft works; PR vs v0.30.0 2.39x, PR vs native -13.1%; config=B-c1-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "All runs used vLLM v0.30.0 on GB300 (TP4), with this PR's quant_config.py mounted over the installed one; MT-Bench serving (80 prompts, greedy, TP4; 0 failed requests); B 1 1.00 / 2.86 / 2.81 114 / 273 / 314 2.39x -13.1%; \"v0.30.0\" and \"PR\" both serve nvidia/DeepSeek-V4.1-Flash-NVFP4 ; \"native\" is deepseek-ai/DeepSeek-V4.1-Flash on v0.30.0.",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59655",
   "source_url": "https://github.com/vllm-project/vllm/pull/59655",
   "tg_tokens": null,
   "tps": "273",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-59655-gb300-b-c1-native",
   "model": "DeepSeek V4.1",
   "notes": "DeepSeek V4.1 Flash on GB300 TP4, DSpark k=5 speculative decoding, MT-Bench output tok/s, node B, concurrency 1, acceptance length 1.00 / 2.86 / 2.81; native deepseek-ai/DeepSeek-V4.1-Flash checkpoint on v0.30.0; PR vs v0.30.0 2.39x, PR vs native -13.1%; config=B-c1-native",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "All runs used vLLM v0.30.0 on GB300 (TP4), with this PR's quant_config.py mounted over the installed one; MT-Bench serving (80 prompts, greedy, TP4; 0 failed requests); B 1 1.00 / 2.86 / 2.81 114 / 273 / 314 2.39x -13.1%; \"v0.30.0\" and \"PR\" both serve nvidia/DeepSeek-V4.1-Flash-NVFP4 ; \"native\" is deepseek-ai/DeepSeek-V4.1-Flash on v0.30.0.",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59655",
   "source_url": "https://github.com/vllm-project/vllm/pull/59655",
   "tg_tokens": null,
   "tps": "314",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 32768,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-59655-gb300-b-c8-v030",
   "model": "DeepSeek V4.1",
   "notes": "DeepSeek V4.1 Flash on GB300 TP4, DSpark k=5 speculative decoding, MT-Bench output tok/s, node B, concurrency 8, acceptance length 1.00 / 2.84 / 2.83; unmodified v0.30.0 serves the NVFP4 checkpoint with the draft experts mis-quantized (acceptance length 1.00, every draft token rejected); PR vs v0.30.0 1.95x, PR vs native -7.4%; config=B-c8-v030",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "All runs used vLLM v0.30.0 on GB300 (TP4), with this PR's quant_config.py mounted over the installed one; MT-Bench serving (80 prompts, greedy, TP4; 0 failed requests); B 8 1.00 / 2.84 / 2.83 796 / 1554 / 1678 1.95x -7.4%; \"v0.30.0\" and \"PR\" both serve nvidia/DeepSeek-V4.1-Flash-NVFP4 ; \"native\" is deepseek-ai/DeepSeek-V4.1-Flash on v0.30.0.",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59655",
   "source_url": "https://github.com/vllm-project/vllm/pull/59655",
   "tg_tokens": null,
   "tps": "796",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 32768,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-59655-gb300-b-c8-pr",
   "model": "DeepSeek V4.1",
   "notes": "DeepSeek V4.1 Flash on GB300 TP4, DSpark k=5 speculative decoding, MT-Bench output tok/s, node B, concurrency 8, acceptance length 1.00 / 2.84 / 2.83; v0.30.0 with the fix: ignored mtp.* experts keep MXFP4, draft works; PR vs v0.30.0 1.95x, PR vs native -7.4%; config=B-c8-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "All runs used vLLM v0.30.0 on GB300 (TP4), with this PR's quant_config.py mounted over the installed one; MT-Bench serving (80 prompts, greedy, TP4; 0 failed requests); B 8 1.00 / 2.84 / 2.83 796 / 1554 / 1678 1.95x -7.4%; \"v0.30.0\" and \"PR\" both serve nvidia/DeepSeek-V4.1-Flash-NVFP4 ; \"native\" is deepseek-ai/DeepSeek-V4.1-Flash on v0.30.0.",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59655",
   "source_url": "https://github.com/vllm-project/vllm/pull/59655",
   "tg_tokens": null,
   "tps": "1554",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 32768,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-59655-gb300-b-c8-native",
   "model": "DeepSeek V4.1",
   "notes": "DeepSeek V4.1 Flash on GB300 TP4, DSpark k=5 speculative decoding, MT-Bench output tok/s, node B, concurrency 8, acceptance length 1.00 / 2.84 / 2.83; native deepseek-ai/DeepSeek-V4.1-Flash checkpoint on v0.30.0; PR vs v0.30.0 1.95x, PR vs native -7.4%; config=B-c8-native",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "All runs used vLLM v0.30.0 on GB300 (TP4), with this PR's quant_config.py mounted over the installed one; MT-Bench serving (80 prompts, greedy, TP4; 0 failed requests); B 8 1.00 / 2.84 / 2.83 796 / 1554 / 1678 1.95x -7.4%; \"v0.30.0\" and \"PR\" both serve nvidia/DeepSeek-V4.1-Flash-NVFP4 ; \"native\" is deepseek-ai/DeepSeek-V4.1-Flash on v0.30.0.",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59655",
   "source_url": "https://github.com/vllm-project/vllm/pull/59655",
   "tg_tokens": null,
   "tps": "1678",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-59655-gb300-c-c1-v030",
   "model": "DeepSeek V4.1",
   "notes": "DeepSeek V4.1 Flash on GB300 TP4, DSpark k=5 speculative decoding, MT-Bench output tok/s, node C, concurrency 1, acceptance length 1.00 / 2.86 / 2.81; unmodified v0.30.0 serves the NVFP4 checkpoint with the draft experts mis-quantized (acceptance length 1.00, every draft token rejected); PR vs v0.30.0 2.45x, PR vs native -0.6%; config=C-c1-v030",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "All runs used vLLM v0.30.0 on GB300 (TP4), with this PR's quant_config.py mounted over the installed one; MT-Bench serving (80 prompts, greedy, TP4; 0 failed requests); C 1 1.00 / 2.86 / 2.81 114 / 280 / 282 2.45x -0.6%; \"v0.30.0\" and \"PR\" both serve nvidia/DeepSeek-V4.1-Flash-NVFP4 ; \"native\" is deepseek-ai/DeepSeek-V4.1-Flash on v0.30.0.",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59655",
   "source_url": "https://github.com/vllm-project/vllm/pull/59655",
   "tg_tokens": null,
   "tps": "114",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-59655-gb300-c-c1-pr",
   "model": "DeepSeek V4.1",
   "notes": "DeepSeek V4.1 Flash on GB300 TP4, DSpark k=5 speculative decoding, MT-Bench output tok/s, node C, concurrency 1, acceptance length 1.00 / 2.86 / 2.81; v0.30.0 with the fix: ignored mtp.* experts keep MXFP4, draft works; PR vs v0.30.0 2.45x, PR vs native -0.6%; config=C-c1-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "All runs used vLLM v0.30.0 on GB300 (TP4), with this PR's quant_config.py mounted over the installed one; MT-Bench serving (80 prompts, greedy, TP4; 0 failed requests); C 1 1.00 / 2.86 / 2.81 114 / 280 / 282 2.45x -0.6%; \"v0.30.0\" and \"PR\" both serve nvidia/DeepSeek-V4.1-Flash-NVFP4 ; \"native\" is deepseek-ai/DeepSeek-V4.1-Flash on v0.30.0.",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59655",
   "source_url": "https://github.com/vllm-project/vllm/pull/59655",
   "tg_tokens": null,
   "tps": "280",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-59655-gb300-c-c1-native",
   "model": "DeepSeek V4.1",
   "notes": "DeepSeek V4.1 Flash on GB300 TP4, DSpark k=5 speculative decoding, MT-Bench output tok/s, node C, concurrency 1, acceptance length 1.00 / 2.86 / 2.81; native deepseek-ai/DeepSeek-V4.1-Flash checkpoint on v0.30.0; PR vs v0.30.0 2.45x, PR vs native -0.6%; config=C-c1-native",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "All runs used vLLM v0.30.0 on GB300 (TP4), with this PR's quant_config.py mounted over the installed one; MT-Bench serving (80 prompts, greedy, TP4; 0 failed requests); C 1 1.00 / 2.86 / 2.81 114 / 280 / 282 2.45x -0.6%; \"v0.30.0\" and \"PR\" both serve nvidia/DeepSeek-V4.1-Flash-NVFP4 ; \"native\" is deepseek-ai/DeepSeek-V4.1-Flash on v0.30.0.",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59655",
   "source_url": "https://github.com/vllm-project/vllm/pull/59655",
   "tg_tokens": null,
   "tps": "282",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 32768,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-59655-gb300-c-c8-v030",
   "model": "DeepSeek V4.1",
   "notes": "DeepSeek V4.1 Flash on GB300 TP4, DSpark k=5 speculative decoding, MT-Bench output tok/s, node C, concurrency 8, acceptance length 1.00 / 2.86 / 2.80; unmodified v0.30.0 serves the NVFP4 checkpoint with the draft experts mis-quantized (acceptance length 1.00, every draft token rejected); PR vs v0.30.0 2.08x, PR vs native -0.7%; config=C-c8-v030",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "All runs used vLLM v0.30.0 on GB300 (TP4), with this PR's quant_config.py mounted over the installed one; MT-Bench serving (80 prompts, greedy, TP4; 0 failed requests); C 8 1.00 / 2.86 / 2.80 766 / 1595 / 1605 2.08x -0.7%; \"v0.30.0\" and \"PR\" both serve nvidia/DeepSeek-V4.1-Flash-NVFP4 ; \"native\" is deepseek-ai/DeepSeek-V4.1-Flash on v0.30.0.",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59655",
   "source_url": "https://github.com/vllm-project/vllm/pull/59655",
   "tg_tokens": null,
   "tps": "766",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 32768,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-59655-gb300-c-c8-pr",
   "model": "DeepSeek V4.1",
   "notes": "DeepSeek V4.1 Flash on GB300 TP4, DSpark k=5 speculative decoding, MT-Bench output tok/s, node C, concurrency 8, acceptance length 1.00 / 2.86 / 2.80; v0.30.0 with the fix: ignored mtp.* experts keep MXFP4, draft works; PR vs v0.30.0 2.08x, PR vs native -0.7%; config=C-c8-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "All runs used vLLM v0.30.0 on GB300 (TP4), with this PR's quant_config.py mounted over the installed one; MT-Bench serving (80 prompts, greedy, TP4; 0 failed requests); C 8 1.00 / 2.86 / 2.80 766 / 1595 / 1605 2.08x -0.7%; \"v0.30.0\" and \"PR\" both serve nvidia/DeepSeek-V4.1-Flash-NVFP4 ; \"native\" is deepseek-ai/DeepSeek-V4.1-Flash on v0.30.0.",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59655",
   "source_url": "https://github.com/vllm-project/vllm/pull/59655",
   "tg_tokens": null,
   "tps": "1595",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 32768,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-59655-gb300-c-c8-native",
   "model": "DeepSeek V4.1",
   "notes": "DeepSeek V4.1 Flash on GB300 TP4, DSpark k=5 speculative decoding, MT-Bench output tok/s, node C, concurrency 8, acceptance length 1.00 / 2.86 / 2.80; native deepseek-ai/DeepSeek-V4.1-Flash checkpoint on v0.30.0; PR vs v0.30.0 2.08x, PR vs native -0.7%; config=C-c8-native",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "All runs used vLLM v0.30.0 on GB300 (TP4), with this PR's quant_config.py mounted over the installed one; MT-Bench serving (80 prompts, greedy, TP4; 0 failed requests); C 8 1.00 / 2.86 / 2.80 766 / 1595 / 1605 2.08x -0.7%; \"v0.30.0\" and \"PR\" both serve nvidia/DeepSeek-V4.1-Flash-NVFP4 ; \"native\" is deepseek-ai/DeepSeek-V4.1-Flash on v0.30.0.",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59655",
   "source_url": "https://github.com/vllm-project/vllm/pull/59655",
   "tg_tokens": null,
   "tps": "1605",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "Radeon 860M",
   "id": "lc-29772-radeon860m-fwht-master",
   "model": "Bonsai 2 27B",
   "notes": "Vulkan FWHT extended to block widths above 512; master falls back to dense f32 matmul for wide Hadamard blocks. Radeon 860M (RDNA 3.5), Bonsai 2 27B Q2_0: tg64 7.26, pp512 59.2, pp2048 55.8 t/s; master arm; config=fwht-master",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "59.2",
   "provenance": "community",
   "quant": "Q2_0",
   "quote": "Also tested by @Tough-Respawn on an AMD Radeon 860M (RDNA 3.5, subgroup 64, 32 KB shared memory); Bonsai 2 27B Q2_0 end to end pp512 59.2 to 66.4, pp2048 55.8 to 62.2, tg64 7.26 to 8.17 t/s (+12 to 13%), KLD unchanged",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp PR #29772",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29772",
   "tg_tokens": 64,
   "tps": "7.26",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "Radeon 860M",
   "id": "lc-29772-radeon860m-fwht-pr",
   "model": "Bonsai 2 27B",
   "notes": "Vulkan FWHT extended to block widths above 512; wide blocks run the shared-memory FWHT shader instead of the dense fallback. Radeon 860M (RDNA 3.5), Bonsai 2 27B Q2_0: tg64 8.17, pp512 66.4, pp2048 62.2 t/s (+12 to 13%), KLD unchanged; PR arm; config=fwht-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "66.4",
   "provenance": "community",
   "quant": "Q2_0",
   "quote": "Also tested by @Tough-Respawn on an AMD Radeon 860M (RDNA 3.5, subgroup 64, 32 KB shared memory); Bonsai 2 27B Q2_0 end to end pp512 59.2 to 66.4, pp2048 55.8 to 62.2, tg64 7.26 to 8.17 t/s (+12 to 13%), KLD unchanged",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp PR #29772",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29772",
   "tg_tokens": 64,
   "tps": "8.17",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 2048,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "vllm-59600-8060s-target-only",
   "model": "LFM2.5-1.2B",
   "notes": "ShortConv drafter state restore fix (MRV2), GSM8K first 32 questions offline output tok/s, target only, 16/32 correct, 2.424 questions/s; config=spec-off",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "Retested commit 359db8bc6143 on one Radeon 8060S ( gfx1151 ), ROCm SDK 10.0.0 / PyTorch 2.13.0, MRV2, TP=1, BF16; Models: LiquidAI/LFM2.5-1.2B-Instruct target and LiquidAI/LFM2.5-350M drafter; Engine: max sequences 8, max model length 2048, eager execution, prefix caching disabled; Target only 16/32 496.40 2.424",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59600",
   "source_url": "https://github.com/vllm-project/vllm/pull/59600",
   "tg_tokens": 256,
   "tps": "496.40",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 2048,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "vllm-59600-8060s-drafter-k3",
   "model": "LFM2.5-1.2B",
   "notes": "ShortConv drafter state restore fix (MRV2), GSM8K first 32 questions offline output tok/s, target + fixed LFM2.5-350M drafter K=3: 16/32 correct, 3.236 questions/s, accepted 4,491/6,252 draft tokens (71.83%); before the fix the drafter state was not restored and speculation did not help; config=spec-k3",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "Retested commit 359db8bc6143 on one Radeon 8060S ( gfx1151 ), ROCm SDK 10.0.0 / PyTorch 2.13.0, MRV2, TP=1, BF16; Models: LiquidAI/LFM2.5-1.2B-Instruct target and LiquidAI/LFM2.5-350M drafter; Engine: max sequences 8, max model length 2048, eager execution, prefix caching disabled; Target + fixed drafter, K=3 16/32 663.49 3.236",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59600",
   "source_url": "https://github.com/vllm-project/vllm/pull/59600",
   "tg_tokens": 256,
   "tps": "663.49",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "DGX Spark",
   "id": "vllm-59606-spark-decode-cutlass",
   "model": "CYBER-FROST-3.8-NVFP4",
   "notes": "MoE backend A/B on DGX Spark TP2 (vLLM 0.30.0, b12x 1.3.0, FlashInfer 0.6.18.post1), 8-stream aggregate decode: flashinfer_cutlass 155.6 tok/s (43 ms per step at 1 stream); b12x crashes with illegal memory access under padded CUDA graphs and kills the worker on a 6,941-token prefill; NLL identical; config=fi-cutlass",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "DGX Spark (GB10, SM121), TP2 over RoCE; Blackfrost-AI/CYBER-FROST-3.8-NVFP4: ModelOpt NVFP4 W4A4 experts, 512 experts, top-10; MTP with 3 speculative tokens; 8-stream aggregate | 155.6 tok/s | 137.6 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM issue #59606",
   "source_url": "https://github.com/vllm-project/vllm/issues/59606",
   "tg_tokens": null,
   "tps": "155.6",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "DGX Spark",
   "id": "vllm-59606-spark-decode-b12x",
   "model": "CYBER-FROST-3.8-NVFP4",
   "notes": "MoE backend A/B on DGX Spark TP2, native b12x NVFP4 MoE backend with VLLM_MOE_SKIP_PADDING=0 (boots only with padding rows routed to real experts): 137.6 tok/s at 8 streams (48 ms per step at 1 stream), -11.6% vs flashinfer_cutlass, and prefill still crashes; config=b12x",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "DGX Spark (GB10, SM121), TP2 over RoCE; Blackfrost-AI/CYBER-FROST-3.8-NVFP4: ModelOpt NVFP4 W4A4 experts, 512 experts, top-10; MTP with 3 speculative tokens; 8-stream aggregate | 155.6 tok/s | 137.6 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM issue #59606",
   "source_url": "https://github.com/vllm-project/vllm/issues/59606",
   "tg_tokens": null,
   "tps": "137.6",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": null,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "H100",
   "id": "vllm-59514-h100-main-kernel",
   "model": "Llama 3.1 8B",
   "notes": "flat-grid silu_and_mul_quant fix (dead vector loop, fp32 fast-math chain), H100, vllm bench serve output tok/s, medians of three interleaved rounds; main kernel arm; per-round req/s 37.67/37.59/37.73; config=main-kernel",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "E2e, Llama-3.1-8B-FP8, H100; vllm bench serve, three interleaved rounds per arm; Output token throughput (tok/s) 4821 5005 4977 +3.8% +0.6%",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59514",
   "source_url": "https://github.com/vllm-project/vllm/pull/59514",
   "tg_tokens": null,
   "tps": "4821",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": null,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "H100",
   "id": "vllm-59514-h100-pr",
   "model": "Llama 3.1 8B",
   "notes": "flat-grid silu_and_mul_quant fix, H100, vllm bench serve output tok/s, three interleaved rounds; PR kernel arm, +3.8% vs main kernel, every main round below every PR round; fused op falls 69.87 -> 27.63 ms of GPU time per profile iteration; mean TPOT 73.31 -> 70.66 ms; config=pr",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "E2e, Llama-3.1-8B-FP8, H100; vllm bench serve, three interleaved rounds per arm; Output token throughput (tok/s) 4821 5005 4977 +3.8% +0.6%",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59514",
   "source_url": "https://github.com/vllm-project/vllm/pull/59514",
   "tg_tokens": null,
   "tps": "5005",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": null,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "H100",
   "id": "vllm-59514-h100-inductor",
   "model": "Llama 3.1 8B",
   "notes": "flat-grid silu_and_mul_quant fix, H100, vllm bench serve output tok/s, three interleaved rounds; torch Inductor path arm (bit-identical bf16 outputs to the PR kernel, +0.6% apart = noise); config=inductor",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "E2e, Llama-3.1-8B-FP8, H100; vllm bench serve, three interleaved rounds per arm; Output token throughput (tok/s) 4821 5005 4977 +3.8% +0.6%",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59514",
   "source_url": "https://github.com/vllm-project/vllm/pull/59514",
   "tg_tokens": null,
   "tps": "4977",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": 8192,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "B200",
   "id": "vllm-59679-b200-glm-4-9b-chat-main",
   "model": "GLM-4-9B-Chat",
   "notes": "migration of GLM-4-9B-Chat to the Transformers modeling backend, B200, vllm bench throughput 1024/256 x 1000 prompts, one run per cell; main arm, gsm8k 0.637; author states the +-1.5% spread is run-to-run noise; config=main",
   "params": null,
   "power_w": null,
   "pp_tokens": 1024,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "1x B200 per model, main ( d848c4ed3d ) vs this branch, transformers 5.17.0; vllm bench throughput --model <model> --max-model-len 8192 --gpu-memory-utilization 0.6 --input-len 1024 --output-len 256 --num-prompts 1000; zai-org/glm-4-9b-chat-hf 0.637 0.650 7,481 7,591",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59679",
   "source_url": "https://github.com/vllm-project/vllm/pull/59679",
   "tg_tokens": 256,
   "tps": "7481",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": 8192,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "B200",
   "id": "vllm-59679-b200-glm-4-9b-chat-branch",
   "model": "GLM-4-9B-Chat",
   "notes": "migration of GLM-4-9B-Chat to the Transformers modeling backend, B200, vllm bench throughput 1024/256 x 1000 prompts, one run per cell; branch arm, gsm8k 0.650; author states the +-1.5% spread is run-to-run noise; config=branch",
   "params": null,
   "power_w": null,
   "pp_tokens": 1024,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "1x B200 per model, main ( d848c4ed3d ) vs this branch, transformers 5.17.0; vllm bench throughput --model <model> --max-model-len 8192 --gpu-memory-utilization 0.6 --input-len 1024 --output-len 256 --num-prompts 1000; zai-org/glm-4-9b-chat-hf 0.637 0.650 7,481 7,591",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59679",
   "source_url": "https://github.com/vllm-project/vllm/pull/59679",
   "tg_tokens": 256,
   "tps": "7591",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": 8192,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "B200",
   "id": "vllm-59679-b200-afm-4-5b-base-main",
   "model": "AFM-4.5B-Base",
   "notes": "migration of AFM-4.5B-Base to the Transformers modeling backend, B200, vllm bench throughput 1024/256 x 1000 prompts, one run per cell; main arm, gsm8k 0.518; author states the +-1.5% spread is run-to-run noise; config=main",
   "params": null,
   "power_w": null,
   "pp_tokens": 1024,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "1x B200 per model, main ( d848c4ed3d ) vs this branch, transformers 5.17.0; vllm bench throughput --model <model> --max-model-len 8192 --gpu-memory-utilization 0.6 --input-len 1024 --output-len 256 --num-prompts 1000; arcee-ai/AFM-4.5B-Base 0.518 0.521 13,657 13,820",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59679",
   "source_url": "https://github.com/vllm-project/vllm/pull/59679",
   "tg_tokens": 256,
   "tps": "13657",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": 8192,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "B200",
   "id": "vllm-59679-b200-afm-4-5b-base-branch",
   "model": "AFM-4.5B-Base",
   "notes": "migration of AFM-4.5B-Base to the Transformers modeling backend, B200, vllm bench throughput 1024/256 x 1000 prompts, one run per cell; branch arm, gsm8k 0.521; author states the +-1.5% spread is run-to-run noise; config=branch",
   "params": null,
   "power_w": null,
   "pp_tokens": 1024,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "1x B200 per model, main ( d848c4ed3d ) vs this branch, transformers 5.17.0; vllm bench throughput --model <model> --max-model-len 8192 --gpu-memory-utilization 0.6 --input-len 1024 --output-len 256 --num-prompts 1000; arcee-ai/AFM-4.5B-Base 0.518 0.521 13,657 13,820",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59679",
   "source_url": "https://github.com/vllm-project/vllm/pull/59679",
   "tg_tokens": 256,
   "tps": "13820",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": 8192,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "B200",
   "id": "vllm-59679-b200-cwm-main",
   "model": "facebook/cwm",
   "notes": "migration of facebook/cwm to the Transformers modeling backend, B200, vllm bench throughput 1024/256 x 1000 prompts, one run per cell; main arm, gsm8k 0.850; author states the +-1.5% spread is run-to-run noise; config=main",
   "params": null,
   "power_w": null,
   "pp_tokens": 1024,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "1x B200 per model, main ( d848c4ed3d ) vs this branch, transformers 5.17.0; vllm bench throughput --model <model> --max-model-len 8192 --gpu-memory-utilization 0.6 --input-len 1024 --output-len 256 --num-prompts 1000; facebook/cwm 0.850 0.852 1,761 1,766",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59679",
   "source_url": "https://github.com/vllm-project/vllm/pull/59679",
   "tg_tokens": 256,
   "tps": "1761",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": 8192,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "B200",
   "id": "vllm-59679-b200-cwm-branch",
   "model": "facebook/cwm",
   "notes": "migration of facebook/cwm to the Transformers modeling backend, B200, vllm bench throughput 1024/256 x 1000 prompts, one run per cell; branch arm, gsm8k 0.852; author states the +-1.5% spread is run-to-run noise; config=branch",
   "params": null,
   "power_w": null,
   "pp_tokens": 1024,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "1x B200 per model, main ( d848c4ed3d ) vs this branch, transformers 5.17.0; vllm bench throughput --model <model> --max-model-len 8192 --gpu-memory-utilization 0.6 --input-len 1024 --output-len 256 --num-prompts 1000; facebook/cwm 0.850 0.852 1,761 1,766",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59679",
   "source_url": "https://github.com/vllm-project/vllm/pull/59679",
   "tg_tokens": 256,
   "tps": "1766",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": 8192,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "B200",
   "id": "vllm-59679-b200-mellum2-12b-a2-5b-main",
   "model": "Mellum2-12B-A2.5B-Base",
   "notes": "migration of Mellum2-12B-A2.5B-Base to the Transformers modeling backend, B200, vllm bench throughput 1024/256 x 1000 prompts, one run per cell; main arm, gsm8k 0.803; author states the +-1.5% spread is run-to-run noise; config=main",
   "params": null,
   "power_w": null,
   "pp_tokens": 1024,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "1x B200 per model, main ( d848c4ed3d ) vs this branch, transformers 5.17.0; vllm bench throughput --model <model> --max-model-len 8192 --gpu-memory-utilization 0.6 --input-len 1024 --output-len 256 --num-prompts 1000; JetBrains/Mellum2-12B-A2.5B-Base 0.803 0.809 16,604 16,498",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59679",
   "source_url": "https://github.com/vllm-project/vllm/pull/59679",
   "tg_tokens": 256,
   "tps": "16604",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": 8192,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "B200",
   "id": "vllm-59679-b200-mellum2-12b-a2-5b-branch",
   "model": "Mellum2-12B-A2.5B-Base",
   "notes": "migration of Mellum2-12B-A2.5B-Base to the Transformers modeling backend, B200, vllm bench throughput 1024/256 x 1000 prompts, one run per cell; branch arm, gsm8k 0.809; author states the +-1.5% spread is run-to-run noise; config=branch",
   "params": null,
   "power_w": null,
   "pp_tokens": 1024,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "1x B200 per model, main ( d848c4ed3d ) vs this branch, transformers 5.17.0; vllm bench throughput --model <model> --max-model-len 8192 --gpu-memory-utilization 0.6 --input-len 1024 --output-len 256 --num-prompts 1000; JetBrains/Mellum2-12B-A2.5B-Base 0.803 0.809 16,604 16,498",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59679",
   "source_url": "https://github.com/vllm-project/vllm/pull/59679",
   "tg_tokens": 256,
   "tps": "16498",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "vllm-59594-pro6000-mtp-blnhc-c1",
   "model": "Qwen3.8-Flash-Next",
   "notes": "KV-cache layout fallback fix, mtp arm at concurrency 1, output tok/s; config=mtp-blnhc-c1",
   "params": null,
   "power_w": null,
   "pp_tokens": 8192,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "End to end, on RTX PRO 6000 with Mooncake over RDMA, Qwen3.8-Flash-Next (QSA attention + GDN, 2 KV heads); Performance at equal TP2 (single instance, no connector, VLLM_KV_CACHE_LAYOUT set, 8K in / 1K out random prompts, output tok/s); TP2, MTP3 speculative decoding: C BLNHC BLHNC; 1 236.1 235.6",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59594",
   "source_url": "https://github.com/vllm-project/vllm/pull/59594",
   "tg_tokens": 1024,
   "tps": "236.1",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "vllm-59594-pro6000-mtp-blhnc-c1",
   "model": "Qwen3.8-Flash-Next",
   "notes": "KV-cache layout fallback fix, mtp arm at concurrency 1, output tok/s; config=mtp-blhnc-c1",
   "params": null,
   "power_w": null,
   "pp_tokens": 8192,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "End to end, on RTX PRO 6000 with Mooncake over RDMA, Qwen3.8-Flash-Next (QSA attention + GDN, 2 KV heads); Performance at equal TP2 (single instance, no connector, VLLM_KV_CACHE_LAYOUT set, 8K in / 1K out random prompts, output tok/s); TP2, MTP3 speculative decoding: C BLNHC BLHNC; 1 236.1 235.6",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59594",
   "source_url": "https://github.com/vllm-project/vllm/pull/59594",
   "tg_tokens": 1024,
   "tps": "235.6",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "vllm-59594-pro6000-mtp-blnhc-c4",
   "model": "Qwen3.8-Flash-Next",
   "notes": "KV-cache layout fallback fix, mtp arm at concurrency 4, output tok/s; config=mtp-blnhc-c4",
   "params": null,
   "power_w": null,
   "pp_tokens": 8192,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "End to end, on RTX PRO 6000 with Mooncake over RDMA, Qwen3.8-Flash-Next (QSA attention + GDN, 2 KV heads); Performance at equal TP2 (single instance, no connector, VLLM_KV_CACHE_LAYOUT set, 8K in / 1K out random prompts, output tok/s); TP2, MTP3 speculative decoding: C BLNHC BLHNC; 4 653.9 643.8",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59594",
   "source_url": "https://github.com/vllm-project/vllm/pull/59594",
   "tg_tokens": 1024,
   "tps": "653.9",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "vllm-59594-pro6000-mtp-blhnc-c4",
   "model": "Qwen3.8-Flash-Next",
   "notes": "KV-cache layout fallback fix, mtp arm at concurrency 4, output tok/s; config=mtp-blhnc-c4",
   "params": null,
   "power_w": null,
   "pp_tokens": 8192,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "End to end, on RTX PRO 6000 with Mooncake over RDMA, Qwen3.8-Flash-Next (QSA attention + GDN, 2 KV heads); Performance at equal TP2 (single instance, no connector, VLLM_KV_CACHE_LAYOUT set, 8K in / 1K out random prompts, output tok/s); TP2, MTP3 speculative decoding: C BLNHC BLHNC; 4 653.9 643.8",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59594",
   "source_url": "https://github.com/vllm-project/vllm/pull/59594",
   "tg_tokens": 1024,
   "tps": "643.8",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "vllm-59594-pro6000-mtp-blnhc-c16",
   "model": "Qwen3.8-Flash-Next",
   "notes": "KV-cache layout fallback fix, mtp arm at concurrency 16, output tok/s; config=mtp-blnhc-c16",
   "params": null,
   "power_w": null,
   "pp_tokens": 8192,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "End to end, on RTX PRO 6000 with Mooncake over RDMA, Qwen3.8-Flash-Next (QSA attention + GDN, 2 KV heads); Performance at equal TP2 (single instance, no connector, VLLM_KV_CACHE_LAYOUT set, 8K in / 1K out random prompts, output tok/s); TP2, MTP3 speculative decoding: C BLNHC BLHNC; 16 1282.8 1293.8",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59594",
   "source_url": "https://github.com/vllm-project/vllm/pull/59594",
   "tg_tokens": 1024,
   "tps": "1282.8",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "vllm-59594-pro6000-mtp-blhnc-c16",
   "model": "Qwen3.8-Flash-Next",
   "notes": "KV-cache layout fallback fix, mtp arm at concurrency 16, output tok/s; config=mtp-blhnc-c16",
   "params": null,
   "power_w": null,
   "pp_tokens": 8192,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "End to end, on RTX PRO 6000 with Mooncake over RDMA, Qwen3.8-Flash-Next (QSA attention + GDN, 2 KV heads); Performance at equal TP2 (single instance, no connector, VLLM_KV_CACHE_LAYOUT set, 8K in / 1K out random prompts, output tok/s); TP2, MTP3 speculative decoding: C BLNHC BLHNC; 16 1282.8 1293.8",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59594",
   "source_url": "https://github.com/vllm-project/vllm/pull/59594",
   "tg_tokens": 1024,
   "tps": "1293.8",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 32,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "vllm-59594-pro6000-mtp-blnhc-c32",
   "model": "Qwen3.8-Flash-Next",
   "notes": "KV-cache layout fallback fix, mtp arm at concurrency 32, output tok/s; config=mtp-blnhc-c32",
   "params": null,
   "power_w": null,
   "pp_tokens": 8192,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "End to end, on RTX PRO 6000 with Mooncake over RDMA, Qwen3.8-Flash-Next (QSA attention + GDN, 2 KV heads); Performance at equal TP2 (single instance, no connector, VLLM_KV_CACHE_LAYOUT set, 8K in / 1K out random prompts, output tok/s); TP2, MTP3 speculative decoding: C BLNHC BLHNC; 32 1628.3 1640.9",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59594",
   "source_url": "https://github.com/vllm-project/vllm/pull/59594",
   "tg_tokens": 1024,
   "tps": "1628.3",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 32,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "vllm-59594-pro6000-mtp-blhnc-c32",
   "model": "Qwen3.8-Flash-Next",
   "notes": "KV-cache layout fallback fix, mtp arm at concurrency 32, output tok/s; config=mtp-blhnc-c32",
   "params": null,
   "power_w": null,
   "pp_tokens": 8192,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "End to end, on RTX PRO 6000 with Mooncake over RDMA, Qwen3.8-Flash-Next (QSA attention + GDN, 2 KV heads); Performance at equal TP2 (single instance, no connector, VLLM_KV_CACHE_LAYOUT set, 8K in / 1K out random prompts, output tok/s); TP2, MTP3 speculative decoding: C BLNHC BLHNC; 32 1628.3 1640.9",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59594",
   "source_url": "https://github.com/vllm-project/vllm/pull/59594",
   "tg_tokens": 1024,
   "tps": "1640.9",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 64,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "vllm-59594-pro6000-mtp-blnhc-c64",
   "model": "Qwen3.8-Flash-Next",
   "notes": "KV-cache layout fallback fix, mtp arm at concurrency 64, output tok/s; config=mtp-blnhc-c64",
   "params": null,
   "power_w": null,
   "pp_tokens": 8192,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "End to end, on RTX PRO 6000 with Mooncake over RDMA, Qwen3.8-Flash-Next (QSA attention + GDN, 2 KV heads); Performance at equal TP2 (single instance, no connector, VLLM_KV_CACHE_LAYOUT set, 8K in / 1K out random prompts, output tok/s); TP2, MTP3 speculative decoding: C BLNHC BLHNC; 64 1493.9 1491.0",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59594",
   "source_url": "https://github.com/vllm-project/vllm/pull/59594",
   "tg_tokens": 1024,
   "tps": "1493.9",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 64,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "vllm-59594-pro6000-mtp-blhnc-c64",
   "model": "Qwen3.8-Flash-Next",
   "notes": "KV-cache layout fallback fix, mtp arm at concurrency 64, output tok/s; config=mtp-blhnc-c64",
   "params": null,
   "power_w": null,
   "pp_tokens": 8192,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "End to end, on RTX PRO 6000 with Mooncake over RDMA, Qwen3.8-Flash-Next (QSA attention + GDN, 2 KV heads); Performance at equal TP2 (single instance, no connector, VLLM_KV_CACHE_LAYOUT set, 8K in / 1K out random prompts, output tok/s); TP2, MTP3 speculative decoding: C BLNHC BLHNC; 64 1493.9 1491.0",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59594",
   "source_url": "https://github.com/vllm-project/vllm/pull/59594",
   "tg_tokens": 1024,
   "tps": "1491.0",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "vllm-59594-pro6000-nospec-blnhc-c1",
   "model": "Qwen3.8-Flash-Next",
   "notes": "KV-cache layout fallback fix, nospec arm at concurrency 1, output tok/s; config=nospec-blnhc-c1",
   "params": null,
   "power_w": null,
   "pp_tokens": 8192,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "End to end, on RTX PRO 6000 with Mooncake over RDMA, Qwen3.8-Flash-Next (QSA attention + GDN, 2 KV heads); Performance at equal TP2 (single instance, no connector, VLLM_KV_CACHE_LAYOUT set, 8K in / 1K out random prompts, output tok/s); TP2, no speculative decoding: C BLNHC BLHNC; 1 110.2 110.2",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59594",
   "source_url": "https://github.com/vllm-project/vllm/pull/59594",
   "tg_tokens": 1024,
   "tps": "110.2",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "vllm-59594-pro6000-nospec-blhnc-c1",
   "model": "Qwen3.8-Flash-Next",
   "notes": "KV-cache layout fallback fix, nospec arm at concurrency 1, output tok/s; config=nospec-blhnc-c1",
   "params": null,
   "power_w": null,
   "pp_tokens": 8192,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "End to end, on RTX PRO 6000 with Mooncake over RDMA, Qwen3.8-Flash-Next (QSA attention + GDN, 2 KV heads); Performance at equal TP2 (single instance, no connector, VLLM_KV_CACHE_LAYOUT set, 8K in / 1K out random prompts, output tok/s); TP2, no speculative decoding: C BLNHC BLHNC; 1 110.2 110.2",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59594",
   "source_url": "https://github.com/vllm-project/vllm/pull/59594",
   "tg_tokens": 1024,
   "tps": "110.2",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "vllm-59594-pro6000-nospec-blnhc-c4",
   "model": "Qwen3.8-Flash-Next",
   "notes": "KV-cache layout fallback fix, nospec arm at concurrency 4, output tok/s; config=nospec-blnhc-c4",
   "params": null,
   "power_w": null,
   "pp_tokens": 8192,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "End to end, on RTX PRO 6000 with Mooncake over RDMA, Qwen3.8-Flash-Next (QSA attention + GDN, 2 KV heads); Performance at equal TP2 (single instance, no connector, VLLM_KV_CACHE_LAYOUT set, 8K in / 1K out random prompts, output tok/s); TP2, no speculative decoding: C BLNHC BLHNC; 4 348.5 348.6",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59594",
   "source_url": "https://github.com/vllm-project/vllm/pull/59594",
   "tg_tokens": 1024,
   "tps": "348.5",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "vllm-59594-pro6000-nospec-blhnc-c4",
   "model": "Qwen3.8-Flash-Next",
   "notes": "KV-cache layout fallback fix, nospec arm at concurrency 4, output tok/s; config=nospec-blhnc-c4",
   "params": null,
   "power_w": null,
   "pp_tokens": 8192,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "End to end, on RTX PRO 6000 with Mooncake over RDMA, Qwen3.8-Flash-Next (QSA attention + GDN, 2 KV heads); Performance at equal TP2 (single instance, no connector, VLLM_KV_CACHE_LAYOUT set, 8K in / 1K out random prompts, output tok/s); TP2, no speculative decoding: C BLNHC BLHNC; 4 348.5 348.6",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59594",
   "source_url": "https://github.com/vllm-project/vllm/pull/59594",
   "tg_tokens": 1024,
   "tps": "348.6",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "vllm-59594-pro6000-nospec-blnhc-c16",
   "model": "Qwen3.8-Flash-Next",
   "notes": "KV-cache layout fallback fix, nospec arm at concurrency 16, output tok/s; config=nospec-blnhc-c16",
   "params": null,
   "power_w": null,
   "pp_tokens": 8192,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "End to end, on RTX PRO 6000 with Mooncake over RDMA, Qwen3.8-Flash-Next (QSA attention + GDN, 2 KV heads); Performance at equal TP2 (single instance, no connector, VLLM_KV_CACHE_LAYOUT set, 8K in / 1K out random prompts, output tok/s); TP2, no speculative decoding: C BLNHC BLHNC; 16 831.4 836.4",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59594",
   "source_url": "https://github.com/vllm-project/vllm/pull/59594",
   "tg_tokens": 1024,
   "tps": "831.4",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "vllm-59594-pro6000-nospec-blhnc-c16",
   "model": "Qwen3.8-Flash-Next",
   "notes": "KV-cache layout fallback fix, nospec arm at concurrency 16, output tok/s; config=nospec-blhnc-c16",
   "params": null,
   "power_w": null,
   "pp_tokens": 8192,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "End to end, on RTX PRO 6000 with Mooncake over RDMA, Qwen3.8-Flash-Next (QSA attention + GDN, 2 KV heads); Performance at equal TP2 (single instance, no connector, VLLM_KV_CACHE_LAYOUT set, 8K in / 1K out random prompts, output tok/s); TP2, no speculative decoding: C BLNHC BLHNC; 16 831.4 836.4",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59594",
   "source_url": "https://github.com/vllm-project/vllm/pull/59594",
   "tg_tokens": 1024,
   "tps": "836.4",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 32,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "vllm-59594-pro6000-nospec-blnhc-c32",
   "model": "Qwen3.8-Flash-Next",
   "notes": "KV-cache layout fallback fix, nospec arm at concurrency 32, output tok/s; config=nospec-blnhc-c32",
   "params": null,
   "power_w": null,
   "pp_tokens": 8192,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "End to end, on RTX PRO 6000 with Mooncake over RDMA, Qwen3.8-Flash-Next (QSA attention + GDN, 2 KV heads); Performance at equal TP2 (single instance, no connector, VLLM_KV_CACHE_LAYOUT set, 8K in / 1K out random prompts, output tok/s); TP2, no speculative decoding: C BLNHC BLHNC; 32 1176.1 1175.5",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59594",
   "source_url": "https://github.com/vllm-project/vllm/pull/59594",
   "tg_tokens": 1024,
   "tps": "1176.1",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 32,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "vllm-59594-pro6000-nospec-blhnc-c32",
   "model": "Qwen3.8-Flash-Next",
   "notes": "KV-cache layout fallback fix, nospec arm at concurrency 32, output tok/s; config=nospec-blhnc-c32",
   "params": null,
   "power_w": null,
   "pp_tokens": 8192,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "End to end, on RTX PRO 6000 with Mooncake over RDMA, Qwen3.8-Flash-Next (QSA attention + GDN, 2 KV heads); Performance at equal TP2 (single instance, no connector, VLLM_KV_CACHE_LAYOUT set, 8K in / 1K out random prompts, output tok/s); TP2, no speculative decoding: C BLNHC BLHNC; 32 1176.1 1175.5",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59594",
   "source_url": "https://github.com/vllm-project/vllm/pull/59594",
   "tg_tokens": 1024,
   "tps": "1175.5",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 64,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "vllm-59594-pro6000-nospec-blnhc-c64",
   "model": "Qwen3.8-Flash-Next",
   "notes": "KV-cache layout fallback fix, nospec arm at concurrency 64, output tok/s; config=nospec-blnhc-c64",
   "params": null,
   "power_w": null,
   "pp_tokens": 8192,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "End to end, on RTX PRO 6000 with Mooncake over RDMA, Qwen3.8-Flash-Next (QSA attention + GDN, 2 KV heads); Performance at equal TP2 (single instance, no connector, VLLM_KV_CACHE_LAYOUT set, 8K in / 1K out random prompts, output tok/s); TP2, no speculative decoding: C BLNHC BLHNC; 64 1214.0 1215.2",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59594",
   "source_url": "https://github.com/vllm-project/vllm/pull/59594",
   "tg_tokens": 1024,
   "tps": "1214.0",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 64,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "vllm-59594-pro6000-nospec-blhnc-c64",
   "model": "Qwen3.8-Flash-Next",
   "notes": "KV-cache layout fallback fix, nospec arm at concurrency 64, output tok/s; config=nospec-blhnc-c64",
   "params": null,
   "power_w": null,
   "pp_tokens": 8192,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "End to end, on RTX PRO 6000 with Mooncake over RDMA, Qwen3.8-Flash-Next (QSA attention + GDN, 2 KV heads); Performance at equal TP2 (single instance, no connector, VLLM_KV_CACHE_LAYOUT set, 8K in / 1K out random prompts, output tok/s); TP2, no speculative decoding: C BLNHC BLHNC; 64 1214.0 1215.2",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59594",
   "source_url": "https://github.com/vllm-project/vllm/pull/59594",
   "tg_tokens": 1024,
   "tps": "1215.2",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "H200",
   "id": "vllm-59607-h200-coder480b-eager",
   "model": "Qwen3-Coder-480B-A35B-Instruct-FP8",
   "notes": "Xid 31 MMU fault makes CUDA-graph replay unusable in the EP16 FP8 MoE path across 2x8 H200 over RoCE; the only working mode is --enforce-eager, measured here at ~583 tok/s aggregate at concurrency 16 with TPOT ~80 ms, far above the 30-50 ms target; cudagraph single-node reference is TPOT ~16 ms (no tok/s published); no A/B arm, single measurement",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "Model: Qwen3-Coder-480B-A35B-Instruct-FP8 (MoE, FP8 block quantization, block_n=128); GPUs: 2 nodes \u00d7 8\u00d7 NVIDIA H200 (16 total); NVLink/NVSwitch intra-node, RoCE inter-node; Parallelism: tensor_parallel_size=8, data_parallel_size=2, enable_expert_parallel=True (EP16); eager: **TPOT ~80 ms, ITL ~80 ms, TTFT ~272 ms, ~583 tok/s** @ conc 16 (400/400 clean)",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM issue #59607",
   "source_url": "https://github.com/vllm-project/vllm/issues/59607",
   "tg_tokens": null,
   "tps": "583",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": 8192,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "B200",
   "id": "vllm-59701-b200-pythia-12b-main",
   "model": "pythia-12b",
   "notes": "migration of pythia-12b to the Transformers modeling backend, B200, vllm bench throughput 1024/256 x 1000 prompts, one run per cell; main arm, gsm8k 0.032; author states the +-1% spread is run-to-run noise; config=main",
   "params": null,
   "power_w": null,
   "pp_tokens": 1024,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "1x B200 per model, main ( ccda9098cc ) vs this branch, transformers 5.17.0; vllm bench throughput --model <model> --gpu-memory-utilization 0.6 [--max-model-len 8192] --input-len 1024 --output-len 256 --num-prompts 1000; EleutherAI/pythia-12b 0.032 0.032 2,912 2,911",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59701",
   "source_url": "https://github.com/vllm-project/vllm/pull/59701",
   "tg_tokens": 256,
   "tps": "2912",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": 8192,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "B200",
   "id": "vllm-59701-b200-pythia-12b-branch",
   "model": "pythia-12b",
   "notes": "migration of pythia-12b to the Transformers modeling backend, B200, vllm bench throughput 1024/256 x 1000 prompts, one run per cell; branch arm, gsm8k 0.032; author states the +-1% spread is run-to-run noise; config=branch",
   "params": null,
   "power_w": null,
   "pp_tokens": 1024,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "1x B200 per model, main ( ccda9098cc ) vs this branch, transformers 5.17.0; vllm bench throughput --model <model> --gpu-memory-utilization 0.6 [--max-model-len 8192] --input-len 1024 --output-len 256 --num-prompts 1000; EleutherAI/pythia-12b 0.032 0.032 2,912 2,911",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59701",
   "source_url": "https://github.com/vllm-project/vllm/pull/59701",
   "tg_tokens": 256,
   "tps": "2911",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": 8192,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "B200",
   "id": "vllm-59701-b200-phi-2-main",
   "model": "phi-2",
   "notes": "migration of phi-2 to the Transformers modeling backend, B200, vllm bench throughput 1024/256 x 1000 prompts, one run per cell; main arm, gsm8k 0.565; author states the +-1% spread is run-to-run noise; config=main",
   "params": null,
   "power_w": null,
   "pp_tokens": 1024,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "1x B200 per model, main ( ccda9098cc ) vs this branch, transformers 5.17.0; vllm bench throughput --model <model> --gpu-memory-utilization 0.6 [--max-model-len 8192] --input-len 1024 --output-len 256 --num-prompts 1000; microsoft/phi-2 0.565 0.565 5,144 5,155",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59701",
   "source_url": "https://github.com/vllm-project/vllm/pull/59701",
   "tg_tokens": 256,
   "tps": "5144",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": 8192,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "B200",
   "id": "vllm-59701-b200-phi-2-branch",
   "model": "phi-2",
   "notes": "migration of phi-2 to the Transformers modeling backend, B200, vllm bench throughput 1024/256 x 1000 prompts, one run per cell; branch arm, gsm8k 0.565; author states the +-1% spread is run-to-run noise; config=branch",
   "params": null,
   "power_w": null,
   "pp_tokens": 1024,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "1x B200 per model, main ( ccda9098cc ) vs this branch, transformers 5.17.0; vllm bench throughput --model <model> --gpu-memory-utilization 0.6 [--max-model-len 8192] --input-len 1024 --output-len 256 --num-prompts 1000; microsoft/phi-2 0.565 0.565 5,144 5,155",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59701",
   "source_url": "https://github.com/vllm-project/vllm/pull/59701",
   "tg_tokens": 256,
   "tps": "5155",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": 8192,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "B200",
   "id": "vllm-59701-b200-seed-oss-36b-base-main",
   "model": "Seed-OSS-36B-Base",
   "notes": "migration of Seed-OSS-36B-Base to the Transformers modeling backend, B200, vllm bench throughput 1024/256 x 1000 prompts, one run per cell; main arm, gsm8k 0.876; author states the +-1% spread is run-to-run noise; config=main",
   "params": null,
   "power_w": null,
   "pp_tokens": 1024,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "1x B200 per model, main ( ccda9098cc ) vs this branch, transformers 5.17.0; vllm bench throughput --model <model> --gpu-memory-utilization 0.6 [--max-model-len 8192] --input-len 1024 --output-len 256 --num-prompts 1000; ByteDance-Seed/Seed-OSS-36B-Base 0.876 0.876 1,540 1,542",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59701",
   "source_url": "https://github.com/vllm-project/vllm/pull/59701",
   "tg_tokens": 256,
   "tps": "1540",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": 8192,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "B200",
   "id": "vllm-59701-b200-seed-oss-36b-base-branch",
   "model": "Seed-OSS-36B-Base",
   "notes": "migration of Seed-OSS-36B-Base to the Transformers modeling backend, B200, vllm bench throughput 1024/256 x 1000 prompts, one run per cell; branch arm, gsm8k 0.876; author states the +-1% spread is run-to-run noise; config=branch",
   "params": null,
   "power_w": null,
   "pp_tokens": 1024,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "1x B200 per model, main ( ccda9098cc ) vs this branch, transformers 5.17.0; vllm bench throughput --model <model> --gpu-memory-utilization 0.6 [--max-model-len 8192] --input-len 1024 --output-len 256 --num-prompts 1000; ByteDance-Seed/Seed-OSS-36B-Base 0.876 0.876 1,540 1,542",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59701",
   "source_url": "https://github.com/vllm-project/vllm/pull/59701",
   "tg_tokens": 256,
   "tps": "1542",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": 8192,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "B200",
   "id": "vllm-59701-b200-jais-2-8b-chat-main",
   "model": "Jais-2-8B-Chat",
   "notes": "migration of Jais-2-8B-Chat to the Transformers modeling backend, B200, vllm bench throughput 1024/256 x 1000 prompts, one run per cell; main arm, gsm8k 0.643; author states the +-1% spread is run-to-run noise; config=main",
   "params": null,
   "power_w": null,
   "pp_tokens": 1024,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "1x B200 per model, main ( ccda9098cc ) vs this branch, transformers 5.17.0; vllm bench throughput --model <model> --gpu-memory-utilization 0.6 [--max-model-len 8192] --input-len 1024 --output-len 256 --num-prompts 1000; inceptionai/Jais-2-8B-Chat 0.643 0.636 5,324 5,359",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59701",
   "source_url": "https://github.com/vllm-project/vllm/pull/59701",
   "tg_tokens": 256,
   "tps": "5324",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": 8192,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "B200",
   "id": "vllm-59701-b200-jais-2-8b-chat-branch",
   "model": "Jais-2-8B-Chat",
   "notes": "migration of Jais-2-8B-Chat to the Transformers modeling backend, B200, vllm bench throughput 1024/256 x 1000 prompts, one run per cell; branch arm, gsm8k 0.636; author states the +-1% spread is run-to-run noise; config=branch",
   "params": null,
   "power_w": null,
   "pp_tokens": 1024,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "1x B200 per model, main ( ccda9098cc ) vs this branch, transformers 5.17.0; vllm bench throughput --model <model> --gpu-memory-utilization 0.6 [--max-model-len 8192] --input-len 1024 --output-len 256 --num-prompts 1000; inceptionai/Jais-2-8B-Chat 0.643 0.636 5,324 5,359",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59701",
   "source_url": "https://github.com/vllm-project/vllm/pull/59701",
   "tg_tokens": 256,
   "tps": "5359",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "Radeon 8065S",
   "id": "hf3-glm53flash-8065s-pp2048",
   "model": "GLM-5.3-Flash",
   "notes": "config=pp2048 | ROCmFPX main prefill (with #35/#36 prefill work); plain llama.cpp master is slower on prefill, same decode.",
   "params": "llama-bench -fa 1 -b 2048 -ub 2048, 3 runs; ROCmFPX main (GLM-5.3 support), Mesa RADV",
   "pp_tokens": 2048,
   "pp_tps": "256.4",
   "provenance": "community",
   "quant": "Q5K-IQ3S mixed (3.84 bpw)",
   "quote": "Built from Unsloth's GLM-5.3-Flash UD-IQ4_XS GGUF and its importance matrix ( glm-imatrix.gguf ), requantizing the attention and dense tensors to Q5_K and the routed experts as above. The file was then converted from the pre-merge glm5next layout to the merged glm5-next layout: keys renamed, indexer.kpool_select_tail = true and indexer.types (all full ) taken from the official text_config , and the gate/up experts stored as separate tensors. The conversion is a byte copy of the weights. Quality KL divergence against Unsloth's UD-IQ4_XS GGUF (the file this mix was made from): mean KLD 0.040, same top-1 token 93.1% ( llama-perplexity --kl-divergence , 16 x 1024-token chunks, Vulkan). Vulkan vs CPU on this exact file (12K tokens, ctx 2048): mean KLD 0.018, same top-1 96.3%. This measures backend numerics, not quantization loss. Speed on the Framework Desktop (Gorgon Halo) AMD Ryzen AI Max+ PRO 495 / Radeon 8065S, 192 GB LPDDR5X, Ubuntu 26.04, kernel 7.0, Mesa 26.0.8 RADV, Vulkan backend. ROCmFPX main (GLM-5.3 support from #34 , prefill work from #35 and #36 ). llama-bench -fa 1 -b 2048 -ub 2048 , 3 runs: test t/s prefill, 2048 tokens 256.4 decode 17.4 prefill, 2048 tokens at 32K context 95.9 decode at 32K context 15.0",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "HF model card: cafonez GLM-5.3-Flash Gorgon GGUF",
   "source_url": "https://huggingface.co/cafonez/GLM-5.3-Flash-Gorgon-GGUF",
   "tg_tokens": null,
   "tps": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "Radeon 8065S",
   "id": "hf3-glm53flash-8065s-tg",
   "model": "GLM-5.3-Flash",
   "notes": "config=tg | Plain llama.cpp master runs the file at the same decode speed.",
   "params": "llama-bench -fa 1 -b 2048 -ub 2048, 3 runs; ROCmFPX main (GLM-5.3 support), Mesa RADV",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q5K-IQ3S mixed (3.84 bpw)",
   "quote": "Built from Unsloth's GLM-5.3-Flash UD-IQ4_XS GGUF and its importance matrix ( glm-imatrix.gguf ), requantizing the attention and dense tensors to Q5_K and the routed experts as above. The file was then converted from the pre-merge glm5next layout to the merged glm5-next layout: keys renamed, indexer.kpool_select_tail = true and indexer.types (all full ) taken from the official text_config , and the gate/up experts stored as separate tensors. The conversion is a byte copy of the weights. Quality KL divergence against Unsloth's UD-IQ4_XS GGUF (the file this mix was made from): mean KLD 0.040, same top-1 token 93.1% ( llama-perplexity --kl-divergence , 16 x 1024-token chunks, Vulkan). Vulkan vs CPU on this exact file (12K tokens, ctx 2048): mean KLD 0.018, same top-1 96.3%. This measures backend numerics, not quantization loss. Speed on the Framework Desktop (Gorgon Halo) AMD Ryzen AI Max+ PRO 495 / Radeon 8065S, 192 GB LPDDR5X, Ubuntu 26.04, kernel 7.0, Mesa 26.0.8 RADV, Vulkan backend. ROCmFPX main (GLM-5.3 support from #34 , prefill work from #35 and #36 ). llama-bench -fa 1 -b 2048 -ub 2048 , 3 runs: test t/s prefill, 2048 tokens 256.4 decode 17.4 prefill, 2048 tokens at 32K context 95.9 decode at 32K context 15.0",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "HF model card: cafonez GLM-5.3-Flash Gorgon GGUF",
   "source_url": "https://huggingface.co/cafonez/GLM-5.3-Flash-Gorgon-GGUF",
   "tg_tokens": null,
   "tps": "17.4"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": 32768,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "Radeon 8065S",
   "id": "hf3-glm53flash-8065s-pp2048-d32k",
   "model": "GLM-5.3-Flash",
   "notes": "config=pp2048-d32k | 32K-context prefill decay from 256.4.",
   "params": "llama-bench -fa 1 -b 2048 -ub 2048, 3 runs; ROCmFPX main (GLM-5.3 support), Mesa RADV",
   "pp_tokens": 2048,
   "pp_tps": "95.9",
   "provenance": "community",
   "quant": "Q5K-IQ3S mixed (3.84 bpw)",
   "quote": "Built from Unsloth's GLM-5.3-Flash UD-IQ4_XS GGUF and its importance matrix ( glm-imatrix.gguf ), requantizing the attention and dense tensors to Q5_K and the routed experts as above. The file was then converted from the pre-merge glm5next layout to the merged glm5-next layout: keys renamed, indexer.kpool_select_tail = true and indexer.types (all full ) taken from the official text_config , and the gate/up experts stored as separate tensors. The conversion is a byte copy of the weights. Quality KL divergence against Unsloth's UD-IQ4_XS GGUF (the file this mix was made from): mean KLD 0.040, same top-1 token 93.1% ( llama-perplexity --kl-divergence , 16 x 1024-token chunks, Vulkan). Vulkan vs CPU on this exact file (12K tokens, ctx 2048): mean KLD 0.018, same top-1 96.3%. This measures backend numerics, not quantization loss. Speed on the Framework Desktop (Gorgon Halo) AMD Ryzen AI Max+ PRO 495 / Radeon 8065S, 192 GB LPDDR5X, Ubuntu 26.04, kernel 7.0, Mesa 26.0.8 RADV, Vulkan backend. ROCmFPX main (GLM-5.3 support from #34 , prefill work from #35 and #36 ). llama-bench -fa 1 -b 2048 -ub 2048 , 3 runs: test t/s prefill, 2048 tokens 256.4 decode 17.4 prefill, 2048 tokens at 32K context 95.9 decode at 32K context 15.0",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "HF model card: cafonez GLM-5.3-Flash Gorgon GGUF",
   "source_url": "https://huggingface.co/cafonez/GLM-5.3-Flash-Gorgon-GGUF",
   "tg_tokens": null,
   "tps": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": 32768,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "Radeon 8065S",
   "id": "hf3-glm53flash-8065s-tg-d32k",
   "model": "GLM-5.3-Flash",
   "notes": "config=tg-d32k | 32K-context decode from 17.4.",
   "params": "llama-bench -fa 1 -b 2048 -ub 2048, 3 runs; ROCmFPX main (GLM-5.3 support), Mesa RADV",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q5K-IQ3S mixed (3.84 bpw)",
   "quote": "Built from Unsloth's GLM-5.3-Flash UD-IQ4_XS GGUF and its importance matrix ( glm-imatrix.gguf ), requantizing the attention and dense tensors to Q5_K and the routed experts as above. The file was then converted from the pre-merge glm5next layout to the merged glm5-next layout: keys renamed, indexer.kpool_select_tail = true and indexer.types (all full ) taken from the official text_config , and the gate/up experts stored as separate tensors. The conversion is a byte copy of the weights. Quality KL divergence against Unsloth's UD-IQ4_XS GGUF (the file this mix was made from): mean KLD 0.040, same top-1 token 93.1% ( llama-perplexity --kl-divergence , 16 x 1024-token chunks, Vulkan). Vulkan vs CPU on this exact file (12K tokens, ctx 2048): mean KLD 0.018, same top-1 96.3%. This measures backend numerics, not quantization loss. Speed on the Framework Desktop (Gorgon Halo) AMD Ryzen AI Max+ PRO 495 / Radeon 8065S, 192 GB LPDDR5X, Ubuntu 26.04, kernel 7.0, Mesa 26.0.8 RADV, Vulkan backend. ROCmFPX main (GLM-5.3 support from #34 , prefill work from #35 and #36 ). llama-bench -fa 1 -b 2048 -ub 2048 , 3 runs: test t/s prefill, 2048 tokens 256.4 decode 17.4 prefill, 2048 tokens at 32K context 95.9 decode at 32K context 15.0",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "HF model card: cafonez GLM-5.3-Flash Gorgon GGUF",
   "source_url": "https://huggingface.co/cafonez/GLM-5.3-Flash-Gorgon-GGUF",
   "tg_tokens": null,
   "tps": "15.0"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 262144,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hf3-qwen3827b-paw-3090-coding-draft",
   "model": "Qwen3.8-27B",
   "notes": "config=coding-draft | DFlash2 drafter on; fastest workload (copying-style code).",
   "params": "llama-paw (llama.cpp CUDA fork, trellis codec); -fa on -ctk q4_0 -ctv q4_0 -c 262144; DFlash2 drafter where noted",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "PAW 3.5 bpw",
   "quote": "Run Qwen/Qwen3.8-27B with good quality on a single RTX 3090; generation, coding (drafter on) 129.09",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "HF model card: lackonendes PAW-27B-X3.1 GGUF",
   "source_url": "https://huggingface.co/lackonendes/PAW-27B-X3.1-GGUF",
   "tg_tokens": null,
   "tps": "129.09"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 262144,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hf3-qwen3827b-paw-3090-newcode-draft",
   "model": "Qwen3.8-27B",
   "notes": "config=newcode-draft | DFlash2 drafter on; novel code slower than copying.",
   "params": "llama-paw (llama.cpp CUDA fork, trellis codec); -fa on -ctk q4_0 -ctv q4_0 -c 262144; DFlash2 drafter where noted",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "PAW 3.5 bpw",
   "quote": "Run Qwen/Qwen3.8-27B with good quality on a single RTX 3090; generation, new code from a long prompt (drafter on) 73.74",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "HF model card: lackonendes PAW-27B-X3.1 GGUF",
   "source_url": "https://huggingface.co/lackonendes/PAW-27B-X3.1-GGUF",
   "tg_tokens": null,
   "tps": "73.74"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 262144,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hf3-qwen3827b-paw-3090-nodraft",
   "model": "Qwen3.8-27B",
   "notes": "config=nodraft | The card's throughput table also lists 31.17 tok/s no-drafter; drafter roughly doubles speed.",
   "params": "llama-paw (llama.cpp CUDA fork, trellis codec); -fa on -ctk q4_0 -ctv q4_0 -c 262144; DFlash2 drafter where noted",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "PAW 3.5 bpw",
   "quote": "Run Qwen/Qwen3.8-27B with good quality on a single RTX 3090; generation, no drafter 30.16",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "HF model card: lackonendes PAW-27B-X3.1 GGUF",
   "source_url": "https://huggingface.co/lackonendes/PAW-27B-X3.1-GGUF",
   "tg_tokens": null,
   "tps": "30.16"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 262144,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hf3-qwen3827b-paw-3090-pp-short",
   "model": "Qwen3.8-27B",
   "notes": "config=pp-short | Prompt processing throughput table.",
   "params": "llama-paw (llama.cpp CUDA fork, trellis codec); -fa on -ctk q4_0 -ctv q4_0 -c 262144; DFlash2 drafter where noted",
   "pp_tokens": null,
   "pp_tps": "982.87",
   "provenance": "community",
   "quant": "PAW 3.5 bpw",
   "quote": "Run Qwen/Qwen3.8-27B with good quality on a single RTX 3090; Prompt processing, short prompt 982.87",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "HF model card: lackonendes PAW-27B-X3.1 GGUF",
   "source_url": "https://huggingface.co/lackonendes/PAW-27B-X3.1-GGUF",
   "tg_tokens": null,
   "tps": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 262144,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hf3-qwen3827b-paw-3090-pp-8k",
   "model": "Qwen3.8-27B",
   "notes": "config=pp-8k | Peak over an 8k-token prompt.",
   "params": "llama-paw (llama.cpp CUDA fork, trellis codec); -fa on -ctk q4_0 -ctv q4_0 -c 262144; DFlash2 drafter where noted",
   "pp_tokens": 8192,
   "pp_tps": "1268.67",
   "provenance": "community",
   "quant": "PAW 3.5 bpw",
   "quote": "Run Qwen/Qwen3.8-27B with good quality on a single RTX 3090; Prompt processing, 8k-token prompt up to 1268.67 (peak)",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "HF model card: lackonendes PAW-27B-X3.1 GGUF",
   "source_url": "https://huggingface.co/lackonendes/PAW-27B-X3.1-GGUF",
   "tg_tokens": null,
   "tps": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "hf3-qwen38flashnext-gyro-r9700-1x",
   "model": "Qwen3.8-Flash-Next",
   "notes": "config=pp512-tg128 | llama-bench batch 1, no speculative decoding; measured by a tester with the agentionai Vulkan build.",
   "params": "agentionai/llama.cpp Vulkan build (rotor formats), llama-bench batch 1, no drafting",
   "pp_tokens": 512,
   "pp_tps": 1246,
   "provenance": "community",
   "quant": "Gyro-S",
   "quote": "Qwen3.8-Flash-Next; Radeon AI PRO R9700, 32 GB llama.cpp fork, Vulkan 57.8 (code with MTP: 70.6) 1,246 (pp512)",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "HF model card: agentionai Qwen3.8-Flash-Next Gyro GGUF",
   "source_url": "https://huggingface.co/agentionai/Qwen3.8-Flash-Next-Gyro-GGUF",
   "tg_tokens": 128,
   "tps": 57.8
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "hf3-qwen38flashnext-gyro-r9700-mtp-code-on",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP draft, agentionai draft-mtp flags, n-max 6; code generation with MTP; config=mtp-code-on",
   "params": "agentionai/llama.cpp Vulkan build (rotor formats), llama-bench batch 1, no drafting",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Gyro-S",
   "quote": "Qwen3.8-Flash-Next; Radeon AI PRO R9700, 32 GB llama.cpp fork, Vulkan 57.8 (code with MTP: 70.6) 1,246 (pp512)",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "HF model card: agentionai Qwen3.8-Flash-Next Gyro GGUF",
   "source_url": "https://huggingface.co/agentionai/Qwen3.8-Flash-Next-Gyro-GGUF",
   "tg_tokens": null,
   "tps": 70.6
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "AMD Strix Halo",
   "id": "hf3-qwen38flashnext-gyro-8060s-gyros",
   "model": "Qwen3.8-Flash-Next",
   "notes": "llama-bench batch 1, no drafting; Gyro-S decode; config=pp512-tg128-gyros",
   "params": "agentionai/llama.cpp Vulkan build (rotor formats), llama-bench batch 1, no drafting",
   "pp_tokens": 512,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Gyro-S",
   "quote": "Qwen3.8-Flash-Next; AMD Strix Halo llama.cpp fork, Vulkan 31.7 (MTP: code 50, JSON 55)",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "HF model card: agentionai Qwen3.8-Flash-Next Gyro GGUF",
   "source_url": "https://huggingface.co/agentionai/Qwen3.8-Flash-Next-Gyro-GGUF",
   "tg_tokens": 128,
   "tps": 31.7
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "AMD Strix Halo",
   "id": "hf3-qwen38flashnext-gyro-8060s-gyrom",
   "model": "Qwen3.8-Flash-Next",
   "notes": "llama-bench batch 1, no drafting; Gyro-M decode; config=pp512-tg128-gyrom",
   "params": "agentionai/llama.cpp Vulkan build (rotor formats), llama-bench batch 1, no drafting",
   "pp_tokens": 512,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Gyro-M",
   "quote": "Qwen3.8-Flash-Next; AMD Strix Halo llama.cpp fork, Vulkan 31.7 (MTP: code 50, JSON 55) 26.4",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "HF model card: agentionai Qwen3.8-Flash-Next Gyro GGUF",
   "source_url": "https://huggingface.co/agentionai/Qwen3.8-Flash-Next-Gyro-GGUF",
   "tg_tokens": 128,
   "tps": 26.4
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "AMD Strix Halo",
   "id": "hf3-qwen38flashnext-gyro-8060s-mtp-code-on",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP draft from the mtp- file; code generation with MTP; config=mtp-code-on",
   "params": "agentionai/llama.cpp Vulkan build, MTP draft where noted",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Gyro-S",
   "quote": "Qwen3.8-Flash-Next; AMD Strix Halo llama.cpp fork, Vulkan 31.7 (MTP: code 50, JSON 55)",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "HF model card: agentionai Qwen3.8-Flash-Next Gyro GGUF",
   "source_url": "https://huggingface.co/agentionai/Qwen3.8-Flash-Next-Gyro-GGUF",
   "tg_tokens": null,
   "tps": 50
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "AMD Strix Halo",
   "id": "hf3-qwen38flashnext-gyro-8060s-mtp-json-on",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP draft from the mtp- file; JSON generation with MTP; config=mtp-json-on",
   "params": "agentionai/llama.cpp Vulkan build, MTP draft where noted",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Gyro-S",
   "quote": "Qwen3.8-Flash-Next; AMD Strix Halo llama.cpp fork, Vulkan 31.7 (MTP: code 50, JSON 55)",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "HF model card: agentionai Qwen3.8-Flash-Next Gyro GGUF",
   "source_url": "https://huggingface.co/agentionai/Qwen3.8-Flash-Next-Gyro-GGUF",
   "tg_tokens": null,
   "tps": 55
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": "2026-09-10",
   "flags": [],
   "hardware": "M4 Pro (48 GB)",
   "id": "hf3-qwen3827b-aeon-mlx-coding",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-b4-coding | AEON abliterated build, vision tower preserved; mlx-vlm with native MTP drafter.",
   "params": "mlx-vlm 0.6.13 / mlx 0.32.0; MTP drafter (block size noted); Mac mini M4 Pro 48 GB (~273 GB/s)",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "6-bit MLX",
   "quote": "--mlx-path Qwen3.8-27B-AEON-Ultimate-mlx-6bit Affine mode, group size 64, no calibration (RTN is near-lossless at 6-bit). Language model and vision tower are both quantized at the global setting; mtp.* tensors are excluded by design \u2014 mlx-vlm loads the MTP drafter as a separate model (see the drafter repo for the split recipe). Measured performance Test machine: Mac mini M4 Pro, 48 GB unified memory (~273 GB/s). Higher-bandwidth chips (M-series Max/Ultra) scale decode proportionally \u2014 serial decode on this quantization measures at ~100% of the device's streaming-bandwidth roofline. Workload Config Decode Coding (temp 0.2) MTP block 4 21.5 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "HF model card: VisualInference Qwen3.8-27B AEON MLX 6-bit",
   "source_url": "https://huggingface.co/VisualInference/Qwen3.8-27B-AEON-Ultimate-Multimodal-MLX-6bit",
   "tg_tokens": null,
   "tps": "21.5"
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": "2026-09-10",
   "flags": [],
   "hardware": "M4 Pro (48 GB)",
   "id": "hf3-qwen3827b-aeon-mlx-docqa",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-b3-docqa-13k | 13k-token context; draft acceptance workload-dependent.",
   "params": "mlx-vlm 0.6.13 / mlx 0.32.0; MTP drafter (block size noted); Mac mini M4 Pro 48 GB (~273 GB/s)",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "6-bit MLX",
   "quote": "--mlx-path Qwen3.8-27B-AEON-Ultimate-mlx-6bit Affine mode, group size 64, no calibration (RTN is near-lossless at 6-bit). Language model and vision tower are both quantized at the global setting; mtp.* tensors are excluded by design \u2014 mlx-vlm loads the MTP drafter as a separate model (see the drafter repo for the split recipe). Measured performance Test machine: Mac mini M4 Pro, 48 GB unified memory (~273 GB/s). Higher-bandwidth chips (M-series Max/Ultra) scale decode proportionally \u2014 serial decode on this quantization measures at ~100% of the device's streaming-bandwidth roofline. Workload Config Decode Coding (temp 0.2) MTP block 4 21.5 tok/s Document QA @ 13k ctx (temp 0.7) MTP block 3 16.9 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "HF model card: VisualInference Qwen3.8-27B AEON MLX 6-bit",
   "source_url": "https://huggingface.co/VisualInference/Qwen3.8-27B-AEON-Ultimate-Multimodal-MLX-6bit",
   "tg_tokens": null,
   "tps": "16.9"
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": "2026-09-10",
   "flags": [],
   "hardware": "M4 Pro (48 GB)",
   "id": "hf3-qwen3827b-aeon-mlx-prose",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-b3-prose | Open-ended prose, lowest draft acceptance.",
   "params": "mlx-vlm 0.6.13 / mlx 0.32.0; MTP drafter (block size noted); Mac mini M4 Pro 48 GB (~273 GB/s)",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "6-bit MLX",
   "quote": "--mlx-path Qwen3.8-27B-AEON-Ultimate-mlx-6bit Affine mode, group size 64, no calibration (RTN is near-lossless at 6-bit). Language model and vision tower are both quantized at the global setting; mtp.* tensors are excluded by design \u2014 mlx-vlm loads the MTP drafter as a separate model (see the drafter repo for the split recipe). Measured performance Test machine: Mac mini M4 Pro, 48 GB unified memory (~273 GB/s). Higher-bandwidth chips (M-series Max/Ultra) scale decode proportionally \u2014 serial decode on this quantization measures at ~100% of the device's streaming-bandwidth roofline. Workload Config Decode Coding (temp 0.2) MTP block 4 21.5 tok/s Document QA @ 13k ctx (temp 0.7) MTP block 3 16.9 tok/s Creative prose (temp 0.7) MTP block 3 15.9 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "HF model card: VisualInference Qwen3.8-27B AEON MLX 6-bit",
   "source_url": "https://huggingface.co/VisualInference/Qwen3.8-27B-AEON-Ultimate-Multimodal-MLX-6bit",
   "tg_tokens": null,
   "tps": "15.9"
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": "2026-09-10",
   "flags": [],
   "hardware": "M4 Pro (48 GB)",
   "id": "hf3-qwen3827b-aeon-mlx-serial",
   "model": "Qwen3.8-27B",
   "notes": "config=serial | Serial decode measures ~100% of the M4 Pro streaming-bandwidth roofline.",
   "params": "mlx-vlm 0.6.13 / mlx 0.32.0; MTP drafter (block size noted); Mac mini M4 Pro 48 GB (~273 GB/s)",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "6-bit MLX",
   "quote": "--mlx-path Qwen3.8-27B-AEON-Ultimate-mlx-6bit Affine mode, group size 64, no calibration (RTN is near-lossless at 6-bit). Language model and vision tower are both quantized at the global setting; mtp.* tensors are excluded by design \u2014 mlx-vlm loads the MTP drafter as a separate model (see the drafter repo for the split recipe). Measured performance Test machine: Mac mini M4 Pro, 48 GB unified memory (~273 GB/s). Higher-bandwidth chips (M-series Max/Ultra) scale decode proportionally \u2014 serial decode on this quantization measures at ~100% of the device's streaming-bandwidth roofline. Workload Config Decode Coding (temp 0.2) MTP block 4 21.5 tok/s Document QA @ 13k ctx (temp 0.7) MTP block 3 16.9 tok/s Creative prose (temp 0.7) MTP block 3 15.9 tok/s Serial (no drafter) \u2014 11.4 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "HF model card: VisualInference Qwen3.8-27B AEON MLX 6-bit",
   "source_url": "https://huggingface.co/VisualInference/Qwen3.8-27B-AEON-Ultimate-Multimodal-MLX-6bit",
   "tg_tokens": null,
   "tps": "11.4"
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": "2026-09-10",
   "flags": [],
   "hardware": "M4 Pro (48 GB)",
   "id": "hf3-qwen3827b-aeon-mlx-prefill",
   "model": "Qwen3.8-27B",
   "notes": "config=prefill | Published as ~105-110; lower bound stored.",
   "params": "mlx-vlm 0.6.13 / mlx 0.32.0; MTP drafter (block size noted); Mac mini M4 Pro 48 GB (~273 GB/s)",
   "pp_tokens": null,
   "pp_tps": "105",
   "provenance": "community",
   "quant": "6-bit MLX",
   "quote": "--mlx-path Qwen3.8-27B-AEON-Ultimate-mlx-6bit Affine mode, group size 64, no calibration (RTN is near-lossless at 6-bit). Language model and vision tower are both quantized at the global setting; mtp.* tensors are excluded by design \u2014 mlx-vlm loads the MTP drafter as a separate model (see the drafter repo for the split recipe). Measured performance Test machine: Mac mini M4 Pro, 48 GB unified memory (~273 GB/s). Higher-bandwidth chips (M-series Max/Ultra) scale decode proportionally \u2014 serial decode on this quantization measures at ~100% of the device's streaming-bandwidth roofline. Workload Config Decode Coding (temp 0.2) MTP block 4 21.5 tok/s Document QA @ 13k ctx (temp 0.7) MTP block 3 16.9 tok/s Creative prose (temp 0.7) MTP block 3 15.9 tok/s Serial (no drafter) \u2014 11.4 tok/s Prefill measured ~105\u2013110 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "HF model card: VisualInference Qwen3.8-27B AEON MLX 6-bit",
   "source_url": "https://huggingface.co/VisualInference/Qwen3.8-27B-AEON-Ultimate-Multimodal-MLX-6bit",
   "tg_tokens": null,
   "tps": null
  },
  {
   "backend": "llama.cpp (zDNN)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "IBM Z Neural Network Processing Assist (NNPA)",
   "id": "lc-29836-nnpa-pp512-base",
   "model": "llama 1B",
   "notes": "llama 1B Q8_0, IBM Z NNPA (zDNN backend), llama-bench pp512, baseline build b11284; config=zdnn-b11284",
   "params": null,
   "power_w": null,
   "pp_tps": "25.81",
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "IBM Z Neural Network Processing Assist (NNPA), CPU; llama 1B Q8_0; | CPU | Model | Test | t/s b11284 | t/s feat/zdnn-i8-upscale | Speedup |; | IBM Z Neural Network Processing Assist (NNPA), CPU | llama 1B Q8_0 | pp512 | 25.81 | 249.62 | 9.67 |",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp PR #29836",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29836",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (zDNN)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "IBM Z Neural Network Processing Assist (NNPA)",
   "id": "lc-29836-nnpa-pp512-pr",
   "model": "llama 1B",
   "notes": "llama 1B Q8_0, IBM Z NNPA (zDNN backend), llama-bench pp512, feat/zdnn-i8-upscale build (Q8_0 matmul added); config=zdnn-i8-upscale",
   "params": null,
   "power_w": null,
   "pp_tps": "249.62",
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "IBM Z Neural Network Processing Assist (NNPA), CPU; llama 1B Q8_0; | CPU | Model | Test | t/s b11284 | t/s feat/zdnn-i8-upscale | Speedup |; | IBM Z Neural Network Processing Assist (NNPA), CPU | llama 1B Q8_0 | pp512 | 25.81 | 249.62 | 9.67 |",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp PR #29836",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29836",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (zDNN)",
   "batch": null,
   "ctx": 1024,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "IBM Z Neural Network Processing Assist (NNPA)",
   "id": "lc-29836-nnpa-pp512d1024-base",
   "model": "llama 1B",
   "notes": "llama 1B Q8_0, IBM Z NNPA (zDNN backend), llama-bench pp512@d1024, baseline build b11284; config=zdnn-b11284",
   "params": null,
   "power_w": null,
   "pp_tps": "22.85",
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "IBM Z Neural Network Processing Assist (NNPA), CPU; llama 1B Q8_0; | CPU | Model | Test | t/s b11284 | t/s feat/zdnn-i8-upscale | Speedup |; | IBM Z Neural Network Processing Assist (NNPA), CPU | llama 1B Q8_0 | pp512@d1024 | 22.85 | 107.69 | 4.71 |",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp PR #29836",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29836",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (zDNN)",
   "batch": null,
   "ctx": 1024,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "IBM Z Neural Network Processing Assist (NNPA)",
   "id": "lc-29836-nnpa-pp512d1024-pr",
   "model": "llama 1B",
   "notes": "llama 1B Q8_0, IBM Z NNPA (zDNN backend), llama-bench pp512@d1024, feat/zdnn-i8-upscale build (Q8_0 matmul added); config=zdnn-i8-upscale",
   "params": null,
   "power_w": null,
   "pp_tps": "107.69",
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "IBM Z Neural Network Processing Assist (NNPA), CPU; llama 1B Q8_0; | CPU | Model | Test | t/s b11284 | t/s feat/zdnn-i8-upscale | Speedup |; | IBM Z Neural Network Processing Assist (NNPA), CPU | llama 1B Q8_0 | pp512@d1024 | 22.85 | 107.69 | 4.71 |",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp PR #29836",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29836",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (zDNN)",
   "batch": null,
   "ctx": 2048,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "IBM Z Neural Network Processing Assist (NNPA)",
   "id": "lc-29836-nnpa-pp512d2048-base",
   "model": "llama 1B",
   "notes": "llama 1B Q8_0, IBM Z NNPA (zDNN backend), llama-bench pp512@d2048, baseline build b11284; config=zdnn-b11284",
   "params": null,
   "power_w": null,
   "pp_tps": "20.41",
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "IBM Z Neural Network Processing Assist (NNPA), CPU; llama 1B Q8_0; | CPU | Model | Test | t/s b11284 | t/s feat/zdnn-i8-upscale | Speedup |; | IBM Z Neural Network Processing Assist (NNPA), CPU | llama 1B Q8_0 | pp512@d2048 | 20.41 | 68.02 | 3.33 |",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp PR #29836",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29836",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (zDNN)",
   "batch": null,
   "ctx": 2048,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "IBM Z Neural Network Processing Assist (NNPA)",
   "id": "lc-29836-nnpa-pp512d2048-pr",
   "model": "llama 1B",
   "notes": "llama 1B Q8_0, IBM Z NNPA (zDNN backend), llama-bench pp512@d2048, feat/zdnn-i8-upscale build (Q8_0 matmul added); config=zdnn-i8-upscale",
   "params": null,
   "power_w": null,
   "pp_tps": "68.02",
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "IBM Z Neural Network Processing Assist (NNPA), CPU; llama 1B Q8_0; | CPU | Model | Test | t/s b11284 | t/s feat/zdnn-i8-upscale | Speedup |; | IBM Z Neural Network Processing Assist (NNPA), CPU | llama 1B Q8_0 | pp512@d2048 | 20.41 | 68.02 | 3.33 |",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp PR #29836",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29836",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (zDNN)",
   "batch": null,
   "ctx": 4096,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "IBM Z Neural Network Processing Assist (NNPA)",
   "id": "lc-29836-nnpa-pp512d4096-base",
   "model": "llama 1B",
   "notes": "llama 1B Q8_0, IBM Z NNPA (zDNN backend), llama-bench pp512@d4096, baseline build b11284; config=zdnn-b11284",
   "params": null,
   "power_w": null,
   "pp_tps": "16.89",
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "IBM Z Neural Network Processing Assist (NNPA), CPU; llama 1B Q8_0; | CPU | Model | Test | t/s b11284 | t/s feat/zdnn-i8-upscale | Speedup |; | IBM Z Neural Network Processing Assist (NNPA), CPU | llama 1B Q8_0 | pp512@d4096 | 16.89 | 40.41 | 2.39 |",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp PR #29836",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29836",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (zDNN)",
   "batch": null,
   "ctx": 4096,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "IBM Z Neural Network Processing Assist (NNPA)",
   "id": "lc-29836-nnpa-pp512d4096-pr",
   "model": "llama 1B",
   "notes": "llama 1B Q8_0, IBM Z NNPA (zDNN backend), llama-bench pp512@d4096, feat/zdnn-i8-upscale build (Q8_0 matmul added); config=zdnn-i8-upscale",
   "params": null,
   "power_w": null,
   "pp_tps": "40.41",
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "IBM Z Neural Network Processing Assist (NNPA), CPU; llama 1B Q8_0; | CPU | Model | Test | t/s b11284 | t/s feat/zdnn-i8-upscale | Speedup |; | IBM Z Neural Network Processing Assist (NNPA), CPU | llama 1B Q8_0 | pp512@d4096 | 16.89 | 40.41 | 2.39 |",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp PR #29836",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29836",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (zDNN)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "IBM Z Neural Network Processing Assist (NNPA)",
   "id": "lc-29836-nnpa-tg128-base",
   "model": "llama 1B",
   "notes": "llama 1B Q8_0, IBM Z NNPA (zDNN backend), llama-bench tg128, baseline build b11284; config=zdnn-b11284",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "IBM Z Neural Network Processing Assist (NNPA), CPU; llama 1B Q8_0; | CPU | Model | Test | t/s b11284 | t/s feat/zdnn-i8-upscale | Speedup |; | IBM Z Neural Network Processing Assist (NNPA), CPU | llama 1B Q8_0 | tg128 | 9.02 | 12.53 | 1.39 |",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp PR #29836",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29836",
   "tg_tokens": null,
   "tps": "9.02",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (zDNN)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "IBM Z Neural Network Processing Assist (NNPA)",
   "id": "lc-29836-nnpa-tg128-pr",
   "model": "llama 1B",
   "notes": "llama 1B Q8_0, IBM Z NNPA (zDNN backend), llama-bench tg128, feat/zdnn-i8-upscale build (Q8_0 matmul added); config=zdnn-i8-upscale",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "IBM Z Neural Network Processing Assist (NNPA), CPU; llama 1B Q8_0; | CPU | Model | Test | t/s b11284 | t/s feat/zdnn-i8-upscale | Speedup |; | IBM Z Neural Network Processing Assist (NNPA), CPU | llama 1B Q8_0 | tg128 | 9.02 | 12.53 | 1.39 |",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp PR #29836",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29836",
   "tg_tokens": null,
   "tps": "12.53",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (zDNN)",
   "batch": null,
   "ctx": 1024,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "IBM Z Neural Network Processing Assist (NNPA)",
   "id": "lc-29836-nnpa-tg128d1024-base",
   "model": "llama 1B",
   "notes": "llama 1B Q8_0, IBM Z NNPA (zDNN backend), llama-bench tg128@d1024, baseline build b11284; config=zdnn-b11284",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "IBM Z Neural Network Processing Assist (NNPA), CPU; llama 1B Q8_0; | CPU | Model | Test | t/s b11284 | t/s feat/zdnn-i8-upscale | Speedup |; | IBM Z Neural Network Processing Assist (NNPA), CPU | llama 1B Q8_0 | tg128@d1024 | 4.61 | 5.39 | 1.17 |",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp PR #29836",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29836",
   "tg_tokens": null,
   "tps": "4.61",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (zDNN)",
   "batch": null,
   "ctx": 1024,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "IBM Z Neural Network Processing Assist (NNPA)",
   "id": "lc-29836-nnpa-tg128d1024-pr",
   "model": "llama 1B",
   "notes": "llama 1B Q8_0, IBM Z NNPA (zDNN backend), llama-bench tg128@d1024, feat/zdnn-i8-upscale build (Q8_0 matmul added); config=zdnn-i8-upscale",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "IBM Z Neural Network Processing Assist (NNPA), CPU; llama 1B Q8_0; | CPU | Model | Test | t/s b11284 | t/s feat/zdnn-i8-upscale | Speedup |; | IBM Z Neural Network Processing Assist (NNPA), CPU | llama 1B Q8_0 | tg128@d1024 | 4.61 | 5.39 | 1.17 |",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp PR #29836",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29836",
   "tg_tokens": null,
   "tps": "5.39",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (zDNN)",
   "batch": null,
   "ctx": 2048,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "IBM Z Neural Network Processing Assist (NNPA)",
   "id": "lc-29836-nnpa-tg128d2048-base",
   "model": "llama 1B",
   "notes": "llama 1B Q8_0, IBM Z NNPA (zDNN backend), llama-bench tg128@d2048, baseline build b11284; config=zdnn-b11284",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "IBM Z Neural Network Processing Assist (NNPA), CPU; llama 1B Q8_0; | CPU | Model | Test | t/s b11284 | t/s feat/zdnn-i8-upscale | Speedup |; | IBM Z Neural Network Processing Assist (NNPA), CPU | llama 1B Q8_0 | tg128@d2048 | 3.34 | 3.79 | 1.14 |",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp PR #29836",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29836",
   "tg_tokens": null,
   "tps": "3.34",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (zDNN)",
   "batch": null,
   "ctx": 2048,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "IBM Z Neural Network Processing Assist (NNPA)",
   "id": "lc-29836-nnpa-tg128d2048-pr",
   "model": "llama 1B",
   "notes": "llama 1B Q8_0, IBM Z NNPA (zDNN backend), llama-bench tg128@d2048, feat/zdnn-i8-upscale build (Q8_0 matmul added); config=zdnn-i8-upscale",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "IBM Z Neural Network Processing Assist (NNPA), CPU; llama 1B Q8_0; | CPU | Model | Test | t/s b11284 | t/s feat/zdnn-i8-upscale | Speedup |; | IBM Z Neural Network Processing Assist (NNPA), CPU | llama 1B Q8_0 | tg128@d2048 | 3.34 | 3.79 | 1.14 |",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp PR #29836",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29836",
   "tg_tokens": null,
   "tps": "3.79",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (zDNN)",
   "batch": null,
   "ctx": 4096,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "IBM Z Neural Network Processing Assist (NNPA)",
   "id": "lc-29836-nnpa-tg128d4096-base",
   "model": "llama 1B",
   "notes": "llama 1B Q8_0, IBM Z NNPA (zDNN backend), llama-bench tg128@d4096, baseline build b11284; config=zdnn-b11284",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "IBM Z Neural Network Processing Assist (NNPA), CPU; llama 1B Q8_0; | CPU | Model | Test | t/s b11284 | t/s feat/zdnn-i8-upscale | Speedup |; | IBM Z Neural Network Processing Assist (NNPA), CPU | llama 1B Q8_0 | tg128@d4096 | 2.18 | 2.32 | 1.06 |",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp PR #29836",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29836",
   "tg_tokens": null,
   "tps": "2.18",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (zDNN)",
   "batch": null,
   "ctx": 4096,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "IBM Z Neural Network Processing Assist (NNPA)",
   "id": "lc-29836-nnpa-tg128d4096-pr",
   "model": "llama 1B",
   "notes": "llama 1B Q8_0, IBM Z NNPA (zDNN backend), llama-bench tg128@d4096, feat/zdnn-i8-upscale build (Q8_0 matmul added); config=zdnn-i8-upscale",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "IBM Z Neural Network Processing Assist (NNPA), CPU; llama 1B Q8_0; | CPU | Model | Test | t/s b11284 | t/s feat/zdnn-i8-upscale | Speedup |; | IBM Z Neural Network Processing Assist (NNPA), CPU | llama 1B Q8_0 | tg128@d4096 | 2.18 | 2.32 | 1.06 |",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp PR #29836",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29836",
   "tg_tokens": null,
   "tps": "2.32",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 8192,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59732-mi355x-8k1k-c4-ctrl",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Qwen3.8-Flash-Next-Quark-MXFP4, MI355X TP2, MXFP4, vllm bench serve random, aggregate output tok/s at c4, 8192/1024; fused QSA prepare launch (main-attention QK-norm, RoPE, gate split and K/V cache write folded into the prepare launch, as on NVIDIA) vs unfused control; config=qsa-unfused",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "MI355X TP2 amd/Qwen3.8-Flash-Next-Quark-MXFP4; ISL OSL Conc Output tok/s Mean TPOT (ms) Per repeat, tok/s; 8,192 1,024 4 462.0 \u2192 475.5 (+2.9%) 7.39 \u2192 7.12 (\u22123.7%)",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59732",
   "source_url": "https://github.com/vllm-project/vllm/pull/59732",
   "tg_tokens": 1024,
   "tps": 462.0,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 8192,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59732-mi355x-8k1k-c4-fused",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Qwen3.8-Flash-Next-Quark-MXFP4, MI355X TP2, MXFP4, vllm bench serve random, aggregate output tok/s at c4, 8192/1024; fused QSA prepare launch (main-attention QK-norm, RoPE, gate split and K/V cache write folded into the prepare launch, as on NVIDIA) vs unfused control; config=qsa-fused",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "MI355X TP2 amd/Qwen3.8-Flash-Next-Quark-MXFP4; ISL OSL Conc Output tok/s Mean TPOT (ms) Per repeat, tok/s; 8,192 1,024 4 462.0 \u2192 475.5 (+2.9%) 7.39 \u2192 7.12 (\u22123.7%)",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59732",
   "source_url": "https://github.com/vllm-project/vllm/pull/59732",
   "tg_tokens": 1024,
   "tps": 475.5,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": 8192,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59732-mi355x-8k1k-c16-ctrl",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Qwen3.8-Flash-Next-Quark-MXFP4, MI355X TP2, MXFP4, vllm bench serve random, aggregate output tok/s at c16, 8192/1024; fused QSA prepare launch (main-attention QK-norm, RoPE, gate split and K/V cache write folded into the prepare launch, as on NVIDIA) vs unfused control; config=qsa-unfused",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "MI355X TP2 amd/Qwen3.8-Flash-Next-Quark-MXFP4; ISL OSL Conc Output tok/s Mean TPOT (ms) Per repeat, tok/s; 8,192 1,024 16 964.1 \u2192 975.3 (+1.2%) 13.70 \u2192 13.51 (\u22121.4%)",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59732",
   "source_url": "https://github.com/vllm-project/vllm/pull/59732",
   "tg_tokens": 1024,
   "tps": 964.1,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": 8192,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59732-mi355x-8k1k-c16-fused",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Qwen3.8-Flash-Next-Quark-MXFP4, MI355X TP2, MXFP4, vllm bench serve random, aggregate output tok/s at c16, 8192/1024; fused QSA prepare launch (main-attention QK-norm, RoPE, gate split and K/V cache write folded into the prepare launch, as on NVIDIA) vs unfused control; config=qsa-fused",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "MI355X TP2 amd/Qwen3.8-Flash-Next-Quark-MXFP4; ISL OSL Conc Output tok/s Mean TPOT (ms) Per repeat, tok/s; 8,192 1,024 16 964.1 \u2192 975.3 (+1.2%) 13.70 \u2192 13.51 (\u22121.4%)",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59732",
   "source_url": "https://github.com/vllm-project/vllm/pull/59732",
   "tg_tokens": 1024,
   "tps": 975.3,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 64,
   "ctx": 8192,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59732-mi355x-8k1k-c64-ctrl",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Qwen3.8-Flash-Next-Quark-MXFP4, MI355X TP2, MXFP4, vllm bench serve random, aggregate output tok/s at c64, 8192/1024; fused QSA prepare launch (main-attention QK-norm, RoPE, gate split and K/V cache write folded into the prepare launch, as on NVIDIA) vs unfused control; config=qsa-unfused",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "MI355X TP2 amd/Qwen3.8-Flash-Next-Quark-MXFP4; ISL OSL Conc Output tok/s Mean TPOT (ms) Per repeat, tok/s; 8,192 1,024 64 1457.2 \u2192 1471.0 (+0.9%) 35.44 \u2192 35.17 (\u22120.8%)",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59732",
   "source_url": "https://github.com/vllm-project/vllm/pull/59732",
   "tg_tokens": 1024,
   "tps": 1457.2,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 64,
   "ctx": 8192,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59732-mi355x-8k1k-c64-fused",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Qwen3.8-Flash-Next-Quark-MXFP4, MI355X TP2, MXFP4, vllm bench serve random, aggregate output tok/s at c64, 8192/1024; fused QSA prepare launch (main-attention QK-norm, RoPE, gate split and K/V cache write folded into the prepare launch, as on NVIDIA) vs unfused control; config=qsa-fused",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "MI355X TP2 amd/Qwen3.8-Flash-Next-Quark-MXFP4; ISL OSL Conc Output tok/s Mean TPOT (ms) Per repeat, tok/s; 8,192 1,024 64 1457.2 \u2192 1471.0 (+0.9%) 35.44 \u2192 35.17 (\u22120.8%)",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59732",
   "source_url": "https://github.com/vllm-project/vllm/pull/59732",
   "tg_tokens": 1024,
   "tps": 1471.0,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 60000,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59732-mi355x-60k600-c4-ctrl",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Qwen3.8-Flash-Next-Quark-MXFP4, MI355X TP2, MXFP4, vllm bench serve random, aggregate output tok/s at c4, 60000/600; fused QSA prepare launch (main-attention QK-norm, RoPE, gate split and K/V cache write folded into the prepare launch, as on NVIDIA) vs unfused control; config=qsa-unfused",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "MI355X TP2 amd/Qwen3.8-Flash-Next-Quark-MXFP4; ISL OSL Conc Output tok/s Mean TPOT (ms) Per repeat, tok/s; 60,000 600 4 144.8 \u2192 147.4 (+1.8%) 14.23 \u2192 13.96 (\u22121.9%)",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59732",
   "source_url": "https://github.com/vllm-project/vllm/pull/59732",
   "tg_tokens": 600,
   "tps": 144.8,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 60000,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59732-mi355x-60k600-c4-fused",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Qwen3.8-Flash-Next-Quark-MXFP4, MI355X TP2, MXFP4, vllm bench serve random, aggregate output tok/s at c4, 60000/600; fused QSA prepare launch (main-attention QK-norm, RoPE, gate split and K/V cache write folded into the prepare launch, as on NVIDIA) vs unfused control; config=qsa-fused",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "MI355X TP2 amd/Qwen3.8-Flash-Next-Quark-MXFP4; ISL OSL Conc Output tok/s Mean TPOT (ms) Per repeat, tok/s; 60,000 600 4 144.8 \u2192 147.4 (+1.8%) 14.23 \u2192 13.96 (\u22121.9%)",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59732",
   "source_url": "https://github.com/vllm-project/vllm/pull/59732",
   "tg_tokens": 600,
   "tps": 147.4,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 8192,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59733-mi355x-k3-36768d1",
   "model": "Kimi-K3",
   "notes": "Kimi-K3 (MXFP4) + DSpark draft, MI355X TP8, MXFP4, vllm bench serve random 8K/1K c8, output tok/s; nightly 36768d1, before #57652 (4 KV cache groups); config=k3-36768d1",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Kimi-K3 (MXFP4) + Kimi-K3-DSpark, MI355X \u00d78, TP8; image KV cache groups total tput/GPU (tok/s) output tput (tok/s) median TPOT (ms) median ITL (ms); nightly 36768d1 (before #57652) 4 729 648 11.13 33.4",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59733",
   "source_url": "https://github.com/vllm-project/vllm/pull/59733",
   "tg_tokens": 1024,
   "tps": 648,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 8192,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59733-mi355x-k3-ac9126e",
   "model": "Kimi-K3",
   "notes": "Kimi-K3 (MXFP4) + DSpark draft, MI355X TP8, MXFP4, vllm bench serve random 8K/1K c8, output tok/s; nightly ac9126e, with #57652 (20 KV cache groups, the regression); config=k3-ac9126e",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Kimi-K3 (MXFP4) + Kimi-K3-DSpark, MI355X \u00d78, TP8; image KV cache groups total tput/GPU (tok/s) output tput (tok/s) median TPOT (ms) median ITL (ms); nightly ac9126e (includes #57652) 20 701 623 11.47 34.4",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59733",
   "source_url": "https://github.com/vllm-project/vllm/pull/59733",
   "tg_tokens": 1024,
   "tps": 623,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 8192,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59733-mi355x-k3-revert57652",
   "model": "Kimi-K3",
   "notes": "Kimi-K3 (MXFP4) + DSpark draft, MI355X TP8, MXFP4, vllm bench serve random 8K/1K c8, output tok/s; ac9126e + full revert of #57652 (4 KV cache groups); config=k3-revert57652",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Kimi-K3 (MXFP4) + Kimi-K3-DSpark, MI355X \u00d78, TP8; image KV cache groups total tput/GPU (tok/s) output tput (tok/s) median TPOT (ms) median ITL (ms); ac9126e + revert #57652 4 726 646 11.15 33.2",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59733",
   "source_url": "https://github.com/vllm-project/vllm/pull/59733",
   "tg_tokens": 1024,
   "tps": 646,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 8192,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59733-mi355x-k3-59733",
   "model": "Kimi-K3",
   "notes": "Kimi-K3 (MXFP4) + DSpark draft, MI355X TP8, MXFP4, vllm bench serve random 8K/1K c8, output tok/s; ac9126e + this PR (max_tp_shards=1, 4 KV cache groups); config=k3-59733",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Kimi-K3 (MXFP4) + Kimi-K3-DSpark, MI355X \u00d78, TP8; image KV cache groups total tput/GPU (tok/s) output tput (tok/s) median TPOT (ms) median ITL (ms); ac9126e + this PR 4 727 647 11.09 33.1",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59733",
   "source_url": "https://github.com/vllm-project/vllm/pull/59733",
   "tg_tokens": 1024,
   "tps": 647,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": null,
   "ctx": 8192,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "OPPO Reno7 (Dimensity 900)",
   "id": "luigi-gemma4e2b-ren7",
   "model": "Gemma 4 E2B",
   "notes": "Real-speed replay of a 2 h 08 zh-TW meeting on the phone (no drift; lag 67 s median, 115 s max per ~4 min window). Effective speed 16 tok/s prefill, 4.5 tok/s decode. llama.cpp, 8 threads, CPU only (Dimensity 900). QAT Q4_0 meeting-agent fine-tune, 8k context restart. Numbers are the effective replay speed, not a clean llama-bench run.",
   "pp_tps": 16.0,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Luigi/gemma-4-E2B-meeting-agent-zh; Live on an OPPO Reno7 (Dimensity 900, CPU only, llama.cpp, 8 threads). A 2 h 08 meeting was replayed at real speed: the lag after each ~4 min window was 67 s median and 115 s max, with no drift. Effective speed was 16 tok/s of prefill and 4.5 tok/s of decode.",
   "retrieved": "2026-10-02",
   "source_name": "HF model card: Luigi/gemma-4-E2B-meeting-agent-zh-GGUF",
   "source_url": "https://huggingface.co/Luigi/gemma-4-E2B-meeting-agent-zh-GGUF",
   "tps": 4.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "M5",
   "id": "takenoko-qwen3508b-m5-cpu-q80",
   "model": "Qwen3.5-0.8B-Japanese-SFT",
   "notes": "llama-bench, generation 128 tokens, on Apple M5. CPU 4 threads (\u30b9\u30ec\u30c3\u30c9 4). The GPU figure in the same card row is 123 tok/s (Metal). GPU \u7d04 = approximate.",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Qwen3.5-0.8B-Japanese-SFT-Q8_0.gguf 0.8 GB \u7d04 78 / 123 \u30c8\u30fc\u30af\u30f3/\u79d2; \u901f\u3055\u306f Apple M5\uff08llama-bench\u3001\u751f\u6210 128 \u30c8\u30fc\u30af\u30f3\uff09\u3067\u6e21\u3063\u305f\u3082\u306e\u3067\u3059\u3002",
   "retrieved": "2026-10-02",
   "source_name": "HF model card: Takenoko12345678/Qwen3.5-0.8B-Japanese-SFT-GGUF",
   "source_url": "https://huggingface.co/Takenoko12345678/Qwen3.5-0.8B-Japanese-SFT-GGUF",
   "tg_tokens": 128,
   "tps": 78,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "M5",
   "id": "takenoko-qwen3508b-m5-cpu-q4km",
   "model": "Qwen3.5-0.8B-Japanese-SFT",
   "notes": "llama-bench, generation 128 tokens, on Apple M5. CPU 4 threads (\u30b9\u30ec\u30c3\u30c9 4). The GPU figure in the same card row is 169 tok/s (Metal). GPU \u7d04 = approximate.",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Qwen3.5-0.8B-Japanese-SFT-Q4_K_M.gguf 0.5 GB \u7d04 92 / 169 \u30c8\u30fc\u30af\u30f3/\u79d2; \u901f\u3055\u306f Apple M5\uff08llama-bench\u3001\u751f\u6210 128 \u30c8\u30fc\u30af\u30f3\uff09\u3067\u6e21\u3063\u305f\u3082\u306e\u3067\u3059\u3002",
   "retrieved": "2026-10-02",
   "source_name": "HF model card: Takenoko12345678/Qwen3.5-0.8B-Japanese-SFT-GGUF",
   "source_url": "https://huggingface.co/Takenoko12345678/Qwen3.5-0.8B-Japanese-SFT-GGUF",
   "tg_tokens": 128,
   "tps": 92,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "M5",
   "id": "takenoko-qwen3508b-m5-metal-q80",
   "model": "Qwen3.5-0.8B-Japanese-SFT",
   "notes": "llama-bench, generation 128 tokens, on Apple M5. GPU (Metal). The CPU figure in the same card row is 78 tok/s (4 threads). CPU \u7d04 = approximate.",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Qwen3.5-0.8B-Japanese-SFT-Q8_0.gguf 0.8 GB \u7d04 78 / 123 \u30c8\u30fc\u30af\u30f3/\u79d2; \u901f\u3055\u306f Apple M5\uff08llama-bench\u3001\u751f\u6210 128 \u30c8\u30fc\u30af\u30f3\uff09\u3067\u6e21\u3063\u305f\u3082\u306e\u3067\u3059\u3002",
   "retrieved": "2026-10-02",
   "source_name": "HF model card: Takenoko12345678/Qwen3.5-0.8B-Japanese-SFT-GGUF",
   "source_url": "https://huggingface.co/Takenoko12345678/Qwen3.5-0.8B-Japanese-SFT-GGUF",
   "tg_tokens": 128,
   "tps": 123,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "M5",
   "id": "takenoko-qwen3508b-m5-metal-q4km",
   "model": "Qwen3.5-0.8B-Japanese-SFT",
   "notes": "llama-bench, generation 128 tokens, on Apple M5. GPU (Metal). The CPU figure in the same card row is 92 tok/s (4 threads). CPU \u7d04 = approximate.",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Qwen3.5-0.8B-Japanese-SFT-Q4_K_M.gguf 0.5 GB \u7d04 92 / 169 \u30c8\u30fc\u30af\u30f3/\u79d2; \u901f\u3055\u306f Apple M5\uff08llama-bench\u3001\u751f\u6210 128 \u30c8\u30fc\u30af\u30f3\uff09\u3067\u6e21\u3063\u305f\u3082\u306e\u3067\u3059\u3002",
   "retrieved": "2026-10-02",
   "source_name": "HF model card: Takenoko12345678/Qwen3.5-0.8B-Japanese-SFT-GGUF",
   "source_url": "https://huggingface.co/Takenoko12345678/Qwen3.5-0.8B-Japanese-SFT-GGUF",
   "tg_tokens": 128,
   "tps": 169,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "Radeon 8065S",
   "id": "hf3-mimov26flashrl-8065s-pp2048",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "config=pp2048 | ROCmFPX main prefill; MiMo-V2.6-Flash-RL-UNCENSORED 309.8B MoE, full GPU offload.",
   "params": "llama-bench -fa 1 -b 2048 -ub 2048, 3 runs; ROCmFPX main (fused gate/up expert tensors, from PR #33), Mesa RADV",
   "pp_tokens": 2048,
   "pp_tps": 484.9,
   "provenance": "community",
   "quant": "MXFP4-Q5K mixed (4.29 bpw)",
   "quote": "MiMo-V2.6-Flash-RL UNCENSORED GGUF for the Framework Desktop (Gorgon Halo) A 4.29 bpw GGUF of dealignai/MiMo-V2.6-Flash-RL-UNCENSORED, the refusal-removed edition of XiaomiMiMo/MiMo-V2.6-Flash-RL (309.8B total parameters, MoE with 256 experts, top-8). It is sized to run fully on the GPU of the 192 GB AMD Ryzen AI Max 400 (\"Gorgon Halo\") Framework Desktop, and was built and tested on that machine; Speed on the Framework Desktop (Gorgon Halo) AMD Ryzen AI Max+ PRO 495 / Radeon 8065S, 192 GB LPDDR5X, Ubuntu 26.04, kernel 7.0, Mesa 26.0.8 RADV, Vulkan backend, ROCmFPX main. llama-bench -fa 1 -b 2048 -ub 2048, 3 runs: test t/s prefill, 2048 tokens 484.9 decode 23.6 prefill, 2048 tokens at 32K context 258.7 decode at 32K context 22.1",
   "retrieved": "2026-10-02",
   "source_name": "HF model card: cafonez MiMo-V2.6-Flash-RL Gorgon GGUF",
   "source_url": "https://huggingface.co/cafonez/MiMo-V2.6-Flash-RL-UNCENSORED-Gorgon-GGUF",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "Radeon 8065S",
   "id": "hf3-mimov26flashrl-8065s-tg",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "config=tg | decode (llama-bench tg128 default test); 309.8B MoE at 4.29 bpw (MXFP4 experts, Q5_K attention/dense/MTP).",
   "params": "llama-bench -fa 1 -b 2048 -ub 2048, 3 runs; ROCmFPX main (fused gate/up expert tensors, from PR #33), Mesa RADV",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4-Q5K mixed (4.29 bpw)",
   "quote": "MiMo-V2.6-Flash-RL UNCENSORED GGUF for the Framework Desktop (Gorgon Halo) A 4.29 bpw GGUF of dealignai/MiMo-V2.6-Flash-RL-UNCENSORED, the refusal-removed edition of XiaomiMiMo/MiMo-V2.6-Flash-RL (309.8B total parameters, MoE with 256 experts, top-8). It is sized to run fully on the GPU of the 192 GB AMD Ryzen AI Max 400 (\"Gorgon Halo\") Framework Desktop, and was built and tested on that machine; Speed on the Framework Desktop (Gorgon Halo) AMD Ryzen AI Max+ PRO 495 / Radeon 8065S, 192 GB LPDDR5X, Ubuntu 26.04, kernel 7.0, Mesa 26.0.8 RADV, Vulkan backend, ROCmFPX main. llama-bench -fa 1 -b 2048 -ub 2048, 3 runs: test t/s prefill, 2048 tokens 484.9 decode 23.6 prefill, 2048 tokens at 32K context 258.7 decode at 32K context 22.1",
   "retrieved": "2026-10-02",
   "source_name": "HF model card: cafonez MiMo-V2.6-Flash-RL Gorgon GGUF",
   "source_url": "https://huggingface.co/cafonez/MiMo-V2.6-Flash-RL-UNCENSORED-Gorgon-GGUF",
   "tg_tokens": 128,
   "tps": 23.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": 32768,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "Radeon 8065S",
   "id": "hf3-mimov26flashrl-8065s-pp2048-d32k",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "config=pp2048-d32k | 32K-context prefill decay from 484.9.",
   "params": "llama-bench -fa 1 -b 2048 -ub 2048, 3 runs; ROCmFPX main (fused gate/up expert tensors, from PR #33), Mesa RADV",
   "pp_tokens": 2048,
   "pp_tps": 258.7,
   "provenance": "community",
   "quant": "MXFP4-Q5K mixed (4.29 bpw)",
   "quote": "MiMo-V2.6-Flash-RL UNCENSORED GGUF for the Framework Desktop (Gorgon Halo) A 4.29 bpw GGUF of dealignai/MiMo-V2.6-Flash-RL-UNCENSORED, the refusal-removed edition of XiaomiMiMo/MiMo-V2.6-Flash-RL (309.8B total parameters, MoE with 256 experts, top-8). It is sized to run fully on the GPU of the 192 GB AMD Ryzen AI Max 400 (\"Gorgon Halo\") Framework Desktop, and was built and tested on that machine; Speed on the Framework Desktop (Gorgon Halo) AMD Ryzen AI Max+ PRO 495 / Radeon 8065S, 192 GB LPDDR5X, Ubuntu 26.04, kernel 7.0, Mesa 26.0.8 RADV, Vulkan backend, ROCmFPX main. llama-bench -fa 1 -b 2048 -ub 2048, 3 runs: test t/s prefill, 2048 tokens 484.9 decode 23.6 prefill, 2048 tokens at 32K context 258.7 decode at 32K context 22.1",
   "retrieved": "2026-10-02",
   "source_name": "HF model card: cafonez MiMo-V2.6-Flash-RL Gorgon GGUF",
   "source_url": "https://huggingface.co/cafonez/MiMo-V2.6-Flash-RL-UNCENSORED-Gorgon-GGUF",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": 32768,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "Radeon 8065S",
   "id": "hf3-mimov26flashrl-8065s-tg-d32k",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "config=tg-d32k | 32K-context decode from 23.6.",
   "params": "llama-bench -fa 1 -b 2048 -ub 2048, 3 runs; ROCmFPX main (fused gate/up expert tensors, from PR #33), Mesa RADV",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4-Q5K mixed (4.29 bpw)",
   "quote": "MiMo-V2.6-Flash-RL UNCENSORED GGUF for the Framework Desktop (Gorgon Halo) A 4.29 bpw GGUF of dealignai/MiMo-V2.6-Flash-RL-UNCENSORED, the refusal-removed edition of XiaomiMiMo/MiMo-V2.6-Flash-RL (309.8B total parameters, MoE with 256 experts, top-8). It is sized to run fully on the GPU of the 192 GB AMD Ryzen AI Max 400 (\"Gorgon Halo\") Framework Desktop, and was built and tested on that machine; Speed on the Framework Desktop (Gorgon Halo) AMD Ryzen AI Max+ PRO 495 / Radeon 8065S, 192 GB LPDDR5X, Ubuntu 26.04, kernel 7.0, Mesa 26.0.8 RADV, Vulkan backend, ROCmFPX main. llama-bench -fa 1 -b 2048 -ub 2048, 3 runs: test t/s prefill, 2048 tokens 484.9 decode 23.6 prefill, 2048 tokens at 32K context 258.7 decode at 32K context 22.1",
   "retrieved": "2026-10-02",
   "source_name": "HF model card: cafonez MiMo-V2.6-Flash-RL Gorgon GGUF",
   "source_url": "https://huggingface.co/cafonez/MiMo-V2.6-Flash-RL-UNCENSORED-Gorgon-GGUF",
   "tg_tokens": 128,
   "tps": 22.1,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "GB10",
   "id": "vllm-59778-gb10-b1k3-base",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Qwen3.8-Flash-Next-NVFP4, GB10 TP=1, MTP k=3; aiperf TPOT, ms/token (lower is better), n=3 per arm, ISL 2048/OSL 256; baseline, no CuTe skinny-GEMM table; config=skinny-b1k3-base",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Qwen3.8-Flash-Next-NVFP4, single GB10, TP=1, ISL 2048 / OSL 256; bs=1, k=3 48.14 \u00b1 0.90 ms 44.86 \u00b1 1.59 ms -6.8% \u00b1 2.2% (3.1\u03c3) aiperf TPOT, n=3 per arm",
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59778",
   "source_url": "https://github.com/vllm-project/vllm/pull/59778",
   "tg_tokens": 256,
   "tps": "48.14",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "GB10",
   "id": "vllm-59778-gb10-b1k3-tuned",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Qwen3.8-Flash-Next-NVFP4, GB10 TP=1, MTP k=3; aiperf TPOT, ms/token (lower is better), n=3 per arm, ISL 2048/OSL 256; with the sm_121 CuTe skinny-GEMM table enabled; config=skinny-b1k3-tuned",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Qwen3.8-Flash-Next-NVFP4, single GB10, TP=1, ISL 2048 / OSL 256; bs=1, k=3 48.14 \u00b1 0.90 ms 44.86 \u00b1 1.59 ms -6.8% \u00b1 2.2% (3.1\u03c3) aiperf TPOT, n=3 per arm",
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59778",
   "source_url": "https://github.com/vllm-project/vllm/pull/59778",
   "tg_tokens": 256,
   "tps": "44.86",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "GB10",
   "id": "vllm-59778-gb10-b1k1-base",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Qwen3.8-Flash-Next-NVFP4, GB10 TP=1, MTP k=1; aiperf TPOT, ms/token (lower is better), n=3 per arm, ISL 2048/OSL 256; baseline, no CuTe skinny-GEMM table; config=skinny-b1k1-base",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Qwen3.8-Flash-Next-NVFP4, single GB10, TP=1, ISL 2048 / OSL 256; bs=1, k=1 47.70 \u00b1 0.34 ms 46.32 \u00b1 0.15 ms -2.9% (6.4\u03c3) aiperf TPOT, n=3 per arm",
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59778",
   "source_url": "https://github.com/vllm-project/vllm/pull/59778",
   "tg_tokens": 256,
   "tps": "47.70",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "GB10",
   "id": "vllm-59778-gb10-b1k1-tuned",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Qwen3.8-Flash-Next-NVFP4, GB10 TP=1, MTP k=1; aiperf TPOT, ms/token (lower is better), n=3 per arm, ISL 2048/OSL 256; with the sm_121 CuTe skinny-GEMM table enabled; config=skinny-b1k1-tuned",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Qwen3.8-Flash-Next-NVFP4, single GB10, TP=1, ISL 2048 / OSL 256; bs=1, k=1 47.70 \u00b1 0.34 ms 46.32 \u00b1 0.15 ms -2.9% (6.4\u03c3) aiperf TPOT, n=3 per arm",
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59778",
   "source_url": "https://github.com/vllm-project/vllm/pull/59778",
   "tg_tokens": 256,
   "tps": "46.32",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 2,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "GB10",
   "id": "vllm-59778-gb10-b2k3-base",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Qwen3.8-Flash-Next-NVFP4, GB10 TP=1, MTP k=3; aiperf TPOT, ms/token (lower is better), n=3 per arm, ISL 2048/OSL 256; baseline, no CuTe skinny-GEMM table; config=skinny-b2k3-base",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Qwen3.8-Flash-Next-NVFP4, single GB10, TP=1, ISL 2048 / OSL 256; bs=2, k=3 55.28 \u00b1 0.50 ms 54.28 \u00b1 0.22 ms -1.8% (3.2\u03c3) aiperf TPOT, n=3 per arm",
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59778",
   "source_url": "https://github.com/vllm-project/vllm/pull/59778",
   "tg_tokens": 256,
   "tps": "55.28",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 2,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "GB10",
   "id": "vllm-59778-gb10-b2k3-tuned",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Qwen3.8-Flash-Next-NVFP4, GB10 TP=1, MTP k=3; aiperf TPOT, ms/token (lower is better), n=3 per arm, ISL 2048/OSL 256; with the sm_121 CuTe skinny-GEMM table enabled; config=skinny-b2k3-tuned",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Qwen3.8-Flash-Next-NVFP4, single GB10, TP=1, ISL 2048 / OSL 256; bs=2, k=3 55.28 \u00b1 0.50 ms 54.28 \u00b1 0.22 ms -1.8% (3.2\u03c3) aiperf TPOT, n=3 per arm",
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59778",
   "source_url": "https://github.com/vllm-project/vllm/pull/59778",
   "tg_tokens": 256,
   "tps": "54.28",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "GB10",
   "id": "vllm-59778-gb10-b4k3-base",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Qwen3.8-Flash-Next-NVFP4, GB10 TP=1, MTP k=3; aiperf TPOT, ms/token (lower is better), n=3 per arm, ISL 2048/OSL 256; baseline, no CuTe skinny-GEMM table; config=skinny-b4k3-base",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Qwen3.8-Flash-Next-NVFP4, single GB10, TP=1, ISL 2048 / OSL 256; bs=4, k=3 73.43 \u00b1 0.05 ms 72.64 \u00b1 0.92 ms -1.1% (1.5\u03c3) aiperf TPOT, n=3 per arm",
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59778",
   "source_url": "https://github.com/vllm-project/vllm/pull/59778",
   "tg_tokens": 256,
   "tps": "73.43",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "GB10",
   "id": "vllm-59778-gb10-b4k3-tuned",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Qwen3.8-Flash-Next-NVFP4, GB10 TP=1, MTP k=3; aiperf TPOT, ms/token (lower is better), n=3 per arm, ISL 2048/OSL 256; with the sm_121 CuTe skinny-GEMM table enabled; config=skinny-b4k3-tuned",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Qwen3.8-Flash-Next-NVFP4, single GB10, TP=1, ISL 2048 / OSL 256; bs=4, k=3 73.43 \u00b1 0.05 ms 72.64 \u00b1 0.92 ms -1.1% (1.5\u03c3) aiperf TPOT, n=3 per arm",
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59778",
   "source_url": "https://github.com/vllm-project/vllm/pull/59778",
   "tg_tokens": 256,
   "tps": "72.64",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "B300",
   "id": "vllm-59779-b300-c1-main",
   "model": "Qwen3 4B",
   "notes": "B300, Qwen3-4B + EAGLE3 (AngelSlim/Qwen3-4B_eagle3, K=3, probabilistic), BF16; watermarking fix for stale prompt lengths under CUDA graphs; aggregate output tok/s, random 1024 in / 256 out, --ignore-eos, c=1; main; config=wm-c1-main",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "B300, Qwen3-4B with EAGLE3; 1 464.7 \u2192 462.3 \u22120.5% \u00b1 0.6 +0.5% \u00b1 0.6 +1.8% \u00b1 2.0 1.632 \u2192 1.632 \u22120.4% \u00b1 0.6 +0.6% \u00b1 3.4",
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59779",
   "source_url": "https://github.com/vllm-project/vllm/pull/59779",
   "tg_tokens": 256,
   "tps": "464.7",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "B300",
   "id": "vllm-59779-b300-c1-fix",
   "model": "Qwen3 4B",
   "notes": "B300, Qwen3-4B + EAGLE3 (AngelSlim/Qwen3-4B_eagle3, K=3, probabilistic), BF16; watermarking fix for stale prompt lengths under CUDA graphs; aggregate output tok/s, random 1024 in / 256 out, --ignore-eos, c=1; fix; config=wm-c1-fix",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "B300, Qwen3-4B with EAGLE3; 1 464.7 \u2192 462.3 \u22120.5% \u00b1 0.6 +0.5% \u00b1 0.6 +1.8% \u00b1 2.0 1.632 \u2192 1.632 \u22120.4% \u00b1 0.6 +0.6% \u00b1 3.4",
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59779",
   "source_url": "https://github.com/vllm-project/vllm/pull/59779",
   "tg_tokens": 256,
   "tps": "462.3",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "B300",
   "id": "vllm-59779-b300-c8-main",
   "model": "Qwen3 4B",
   "notes": "B300, Qwen3-4B + EAGLE3 (AngelSlim/Qwen3-4B_eagle3, K=3, probabilistic), BF16; watermarking fix for stale prompt lengths under CUDA graphs; aggregate output tok/s, random 1024 in / 256 out, --ignore-eos, c=8; main; config=wm-c8-main",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "B300, Qwen3-4B with EAGLE3; 8 3032 \u2192 3012 \u22120.6% \u00b1 0.8 +0.4% \u00b1 0.9 +2.3% \u00b1 1.7 1.658 \u2192 1.662 \u22120.3% \u00b1 3.3 \u22121.2% \u00b1 5.6",
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59779",
   "source_url": "https://github.com/vllm-project/vllm/pull/59779",
   "tg_tokens": 256,
   "tps": "3032",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "B300",
   "id": "vllm-59779-b300-c8-fix",
   "model": "Qwen3 4B",
   "notes": "B300, Qwen3-4B + EAGLE3 (AngelSlim/Qwen3-4B_eagle3, K=3, probabilistic), BF16; watermarking fix for stale prompt lengths under CUDA graphs; aggregate output tok/s, random 1024 in / 256 out, --ignore-eos, c=8; fix; config=wm-c8-fix",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "B300, Qwen3-4B with EAGLE3; 8 3032 \u2192 3012 \u22120.6% \u00b1 0.8 +0.4% \u00b1 0.9 +2.3% \u00b1 1.7 1.658 \u2192 1.662 \u22120.3% \u00b1 3.3 \u22121.2% \u00b1 5.6",
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59779",
   "source_url": "https://github.com/vllm-project/vllm/pull/59779",
   "tg_tokens": 256,
   "tps": "3012",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 64,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "B300",
   "id": "vllm-59779-b300-c64-main",
   "model": "Qwen3 4B",
   "notes": "B300, Qwen3-4B + EAGLE3 (AngelSlim/Qwen3-4B_eagle3, K=3, probabilistic), BF16; watermarking fix for stale prompt lengths under CUDA graphs; aggregate output tok/s, random 1024 in / 256 out, --ignore-eos, c=64; main; config=wm-c64-main",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "B300, Qwen3-4B with EAGLE3; 64 10543 \u2192 10495 \u22120.4% \u00b1 0.8 +0.3% \u00b1 0.9 +0.8% \u00b1 1.5 1.533 \u2192 1.533 +0.3% \u00b1 1.2 \u22120.7% \u00b1 4.1",
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59779",
   "source_url": "https://github.com/vllm-project/vllm/pull/59779",
   "tg_tokens": 256,
   "tps": "10543",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 64,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "B300",
   "id": "vllm-59779-b300-c64-fix",
   "model": "Qwen3 4B",
   "notes": "B300, Qwen3-4B + EAGLE3 (AngelSlim/Qwen3-4B_eagle3, K=3, probabilistic), BF16; watermarking fix for stale prompt lengths under CUDA graphs; aggregate output tok/s, random 1024 in / 256 out, --ignore-eos, c=64; fix; config=wm-c64-fix",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "B300, Qwen3-4B with EAGLE3; 64 10543 \u2192 10495 \u22120.4% \u00b1 0.8 +0.3% \u00b1 0.9 +0.8% \u00b1 1.5 1.533 \u2192 1.533 +0.3% \u00b1 1.2 \u22120.7% \u00b1 4.1",
   "retrieved": "2026-10-02",
   "source_name": "vLLM PR #59779",
   "source_url": "https://github.com/vllm-project/vllm/pull/59779",
   "tg_tokens": 256,
   "tps": "10495",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "GB10",
   "id": "cdt-ornith397-gb10-nmax0",
   "model": "Ornith-1.5-397B-A17B",
   "notes": "ik_llama.cpp DFlash drafter, n_max sweep (spec-type draft-dflash); target Ornith-1.5-397B-A17B IQ2_XXS (finetune of z-lab/Qwen3.5-397B-A17B-DFlash); ~32k prompts, 131072 KV q8_0, 256-token generations, one boot two reps; config=dflash-nmax0",
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ2_XXS",
   "quote": "DFlash Drafter for Ornith 1.5 397B A17B IQ2_XXS (GGUF); Measured on an NVIDIA GB10 (DGX Spark), 121.6 GiB unified memory shared between CPU and GPU; n_max tok/s vs spec-off acceptance 0 15.00 \u2014 \u2014 2 19.07 +27.3% 0.562 3 19.18 +28.0% 0.462 4 19.82 +32.4% 0.431 5 17.72 +18.5% 0.334",
   "retrieved": "2026-10-02",
   "source_name": "HF model card: cdtio33/Ornith-1.5-397B-IQ2_XXS-DFlash-Drafter-GGUF",
   "source_url": "https://huggingface.co/cdtio33/Ornith-1.5-397B-IQ2_XXS-DFlash-Drafter-GGUF",
   "tps": "15.00",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "GB10",
   "id": "cdt-ornith397-gb10-nmax2",
   "model": "Ornith-1.5-397B-A17B",
   "notes": "ik_llama.cpp DFlash drafter, n_max sweep (spec-type draft-dflash); target Ornith-1.5-397B-A17B IQ2_XXS (finetune of z-lab/Qwen3.5-397B-A17B-DFlash); ~32k prompts, 131072 KV q8_0, 256-token generations, one boot two reps, acceptance 0.562; config=dflash-nmax2",
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ2_XXS",
   "quote": "DFlash Drafter for Ornith 1.5 397B A17B IQ2_XXS (GGUF); Measured on an NVIDIA GB10 (DGX Spark), 121.6 GiB unified memory shared between CPU and GPU; n_max tok/s vs spec-off acceptance 0 15.00 \u2014 \u2014 2 19.07 +27.3% 0.562 3 19.18 +28.0% 0.462 4 19.82 +32.4% 0.431 5 17.72 +18.5% 0.334",
   "retrieved": "2026-10-02",
   "source_name": "HF model card: cdtio33/Ornith-1.5-397B-IQ2_XXS-DFlash-Drafter-GGUF",
   "source_url": "https://huggingface.co/cdtio33/Ornith-1.5-397B-IQ2_XXS-DFlash-Drafter-GGUF",
   "tps": "19.07",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "GB10",
   "id": "cdt-ornith397-gb10-nmax3",
   "model": "Ornith-1.5-397B-A17B",
   "notes": "ik_llama.cpp DFlash drafter, n_max sweep (spec-type draft-dflash); target Ornith-1.5-397B-A17B IQ2_XXS (finetune of z-lab/Qwen3.5-397B-A17B-DFlash); ~32k prompts, 131072 KV q8_0, 256-token generations, one boot two reps, acceptance 0.462; config=dflash-nmax3",
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ2_XXS",
   "quote": "DFlash Drafter for Ornith 1.5 397B A17B IQ2_XXS (GGUF); Measured on an NVIDIA GB10 (DGX Spark), 121.6 GiB unified memory shared between CPU and GPU; n_max tok/s vs spec-off acceptance 0 15.00 \u2014 \u2014 2 19.07 +27.3% 0.562 3 19.18 +28.0% 0.462 4 19.82 +32.4% 0.431 5 17.72 +18.5% 0.334",
   "retrieved": "2026-10-02",
   "source_name": "HF model card: cdtio33/Ornith-1.5-397B-IQ2_XXS-DFlash-Drafter-GGUF",
   "source_url": "https://huggingface.co/cdtio33/Ornith-1.5-397B-IQ2_XXS-DFlash-Drafter-GGUF",
   "tps": "19.18",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "GB10",
   "id": "cdt-ornith397-gb10-nmax4",
   "model": "Ornith-1.5-397B-A17B",
   "notes": "ik_llama.cpp DFlash drafter, n_max sweep (spec-type draft-dflash); target Ornith-1.5-397B-A17B IQ2_XXS (finetune of z-lab/Qwen3.5-397B-A17B-DFlash); ~32k prompts, 131072 KV q8_0, 256-token generations, one boot two reps, acceptance 0.431; config=dflash-nmax4",
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ2_XXS",
   "quote": "DFlash Drafter for Ornith 1.5 397B A17B IQ2_XXS (GGUF); Measured on an NVIDIA GB10 (DGX Spark), 121.6 GiB unified memory shared between CPU and GPU; n_max tok/s vs spec-off acceptance 0 15.00 \u2014 \u2014 2 19.07 +27.3% 0.562 3 19.18 +28.0% 0.462 4 19.82 +32.4% 0.431 5 17.72 +18.5% 0.334",
   "retrieved": "2026-10-02",
   "source_name": "HF model card: cdtio33/Ornith-1.5-397B-IQ2_XXS-DFlash-Drafter-GGUF",
   "source_url": "https://huggingface.co/cdtio33/Ornith-1.5-397B-IQ2_XXS-DFlash-Drafter-GGUF",
   "tps": "19.82",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "GB10",
   "id": "cdt-ornith397-gb10-nmax5",
   "model": "Ornith-1.5-397B-A17B",
   "notes": "ik_llama.cpp DFlash drafter, n_max sweep (spec-type draft-dflash); target Ornith-1.5-397B-A17B IQ2_XXS (finetune of z-lab/Qwen3.5-397B-A17B-DFlash); ~32k prompts, 131072 KV q8_0, 256-token generations, one boot two reps, acceptance 0.334; config=dflash-nmax5",
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ2_XXS",
   "quote": "DFlash Drafter for Ornith 1.5 397B A17B IQ2_XXS (GGUF); Measured on an NVIDIA GB10 (DGX Spark), 121.6 GiB unified memory shared between CPU and GPU; n_max tok/s vs spec-off acceptance 0 15.00 \u2014 \u2014 2 19.07 +27.3% 0.562 3 19.18 +28.0% 0.462 4 19.82 +32.4% 0.431 5 17.72 +18.5% 0.334",
   "retrieved": "2026-10-02",
   "source_name": "HF model card: cdtio33/Ornith-1.5-397B-IQ2_XXS-DFlash-Drafter-GGUF",
   "source_url": "https://huggingface.co/cdtio33/Ornith-1.5-397B-IQ2_XXS-DFlash-Drafter-GGUF",
   "tps": "17.72",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "2 x RTX 3060",
   "id": "lemesoft-ornith35-2x3060",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "2 x RTX 3060 full-GPU (-ngl 999, tensor split), IQ4_XS imatrix; MTP --spec-type draft-mtp n-max 3, mmproj, max ctx 180224, 10 CPU threads, flash attention on; peak fill ~1800 tok/s, peak generation ~120 tok/s (author's approximations, one run per cell, not a clean benchmark); agentic heavy-KV use 50-90 tok/s",
   "pp_tps": "1800",
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "Llamacpp imatrix IQ4_XS Quantization of Ornith-1.5-35B-A3B by ornith-ai; With 2 x RTX 3060 the peak fill rate is about 1800 tok/s and peak generation rate is about 120 tok/s",
   "retrieved": "2026-10-02",
   "source_name": "HF model card: LemesoftNostalgic/Ornith-1.5-35B-A3B-GGUF",
   "source_url": "https://huggingface.co/LemesoftNostalgic/Ornith-1.5-35B-A3B-GGUF",
   "tps": "120",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "dargoth-qwen36-5090-q4km",
   "model": "Qwen3.6-35B-A3B",
   "notes": "llama-bench pp512/tg128, single RTX 5090 (32 GB), CUDA; custom per-tensor requantizer of the Hermes3.6-35B-A3B fine-tune of Qwen3.6-35B-A3B, dequantized from Q8_0 source in memory (no F16/BF16 intermediate); blends calibrated on diverse imatrix (n_ctx 6144); run-to-run variance ~5%; config=Q4_K_M",
   "pp_tokens": 512,
   "pp_tps": "8316",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-Final-Experimental-GGUF; Throughput: llama-bench, single RTX 5090 (32 GB), llama.cpp CUDA build, all models measured back-to-back in one session; Q4_K_M (reference recipe) 19.7 GiB 7.078 8316 281",
   "retrieved": "2026-10-02",
   "source_name": "HF model card: Dargoth/Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-Final-Experimental-GGUF",
   "source_url": "https://huggingface.co/Dargoth/Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-Final-Experimental-GGUF",
   "tg_tokens": 128,
   "tps": "281",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "dargoth-qwen36-5090-nvfp4",
   "model": "Qwen3.6-35B-A3B",
   "notes": "llama-bench pp512/tg128, single RTX 5090 (32 GB), CUDA; custom per-tensor requantizer of the Hermes3.6-35B-A3B fine-tune of Qwen3.6-35B-A3B, dequantized from Q8_0 source in memory (no F16/BF16 intermediate); blends calibrated on diverse imatrix (n_ctx 6144); run-to-run variance ~5%; config=NVFP4",
   "pp_tokens": 512,
   "pp_tps": "8855",
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-Final-Experimental-GGUF; Throughput: llama-bench, single RTX 5090 (32 GB), llama.cpp CUDA build, all models measured back-to-back in one session; NVFP4 (tuned) 28.4 GiB 7.265 8855 250",
   "retrieved": "2026-10-02",
   "source_name": "HF model card: Dargoth/Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-Final-Experimental-GGUF",
   "source_url": "https://huggingface.co/Dargoth/Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-Final-Experimental-GGUF",
   "tg_tokens": 128,
   "tps": "250",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "dargoth-qwen36-5090-blend2",
   "model": "Qwen3.6-35B-A3B",
   "notes": "llama-bench pp512/tg128, single RTX 5090 (32 GB), CUDA; custom per-tensor requantizer of the Hermes3.6-35B-A3B fine-tune of Qwen3.6-35B-A3B, dequantized from Q8_0 source in memory (no F16/BF16 intermediate); blends calibrated on diverse imatrix (n_ctx 6144); run-to-run variance ~5%; config=Q4_blend2",
   "pp_tokens": 512,
   "pp_tps": "8566",
   "provenance": "community",
   "quant": "Q4_blend2",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-Final-Experimental-GGUF; Throughput: llama-bench, single RTX 5090 (32 GB), llama.cpp CUDA build, all models measured back-to-back in one session; Q4_blend2 19.5 GiB 7.050 8566 247",
   "retrieved": "2026-10-02",
   "source_name": "HF model card: Dargoth/Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-Final-Experimental-GGUF",
   "source_url": "https://huggingface.co/Dargoth/Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-Final-Experimental-GGUF",
   "tg_tokens": 128,
   "tps": "247",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "dargoth-qwen36-5090-blend4",
   "model": "Qwen3.6-35B-A3B",
   "notes": "llama-bench pp512/tg128, single RTX 5090 (32 GB), CUDA; custom per-tensor requantizer of the Hermes3.6-35B-A3B fine-tune of Qwen3.6-35B-A3B, dequantized from Q8_0 source in memory (no F16/BF16 intermediate); blends calibrated on diverse imatrix (n_ctx 6144); run-to-run variance ~5%; config=Q4_blend4",
   "pp_tokens": 512,
   "pp_tps": "8137",
   "provenance": "community",
   "quant": "Q4_blend4",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-Final-Experimental-GGUF; Throughput: llama-bench, single RTX 5090 (32 GB), llama.cpp CUDA build, all models measured back-to-back in one session; Q4_blend4 19.1 GiB 7.058 8137 258",
   "retrieved": "2026-10-02",
   "source_name": "HF model card: Dargoth/Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-Final-Experimental-GGUF",
   "source_url": "https://huggingface.co/Dargoth/Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-Final-Experimental-GGUF",
   "tg_tokens": 128,
   "tps": "258",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "dargoth-qwen36-5090-blend5",
   "model": "Qwen3.6-35B-A3B",
   "notes": "llama-bench pp512/tg128, single RTX 5090 (32 GB), CUDA; custom per-tensor requantizer of the Hermes3.6-35B-A3B fine-tune of Qwen3.6-35B-A3B, dequantized from Q8_0 source in memory (no F16/BF16 intermediate); blends calibrated on diverse imatrix (n_ctx 6144); run-to-run variance ~5%; config=Q4_blend5",
   "pp_tokens": 512,
   "pp_tps": "8123",
   "provenance": "community",
   "quant": "Q4_blend5",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-Final-Experimental-GGUF; Throughput: llama-bench, single RTX 5090 (32 GB), llama.cpp CUDA build, all models measured back-to-back in one session; Q4_blend5 18.9 GiB 7.058 8123 258",
   "retrieved": "2026-10-02",
   "source_name": "HF model card: Dargoth/Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-Final-Experimental-GGUF",
   "source_url": "https://huggingface.co/Dargoth/Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-Final-Experimental-GGUF",
   "tg_tokens": 128,
   "tps": "258",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "M3 Ultra (60-core GPU)",
   "id": "lc-29869-m3u-dflash-code-master",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #29869, M3 Ultra (60-core GPU), Qwen3.8-27B Q4_0 target + DFlash2 Q8_0 drafter, -ngl 99 -fa on -c 8192 -np 1, 64 generated tokens, median of 5 requests, mean of 2 server runs; DFlash2 code T=0, master build; config=dflash-code-master",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Qwen3.8-27B DFlash2 Q8_0 drafter; M3 Ultra (60-core GPU); DFlash2 code 0 30.2 110.0",
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp PR #29869 (Metal few-row MMA for spec decode)",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29869",
   "tg_tokens": 64,
   "tps": "30.2",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "M3 Ultra (60-core GPU)",
   "id": "lc-29869-m3u-dflash-code-pr",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #29869, M3 Ultra (60-core GPU), Qwen3.8-27B Q4_0 target + DFlash2 Q8_0 drafter, -ngl 99 -fa on -c 8192 -np 1, 64 generated tokens, median of 5 requests, mean of 2 server runs; DFlash2 code T=0, this PR build; config=dflash-code-pr",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Qwen3.8-27B DFlash2 Q8_0 drafter; M3 Ultra (60-core GPU); DFlash2 code 0 30.2 110.0",
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp PR #29869 (Metal few-row MMA for spec decode)",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29869",
   "tg_tokens": 64,
   "tps": "110.0",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "M3 Ultra (60-core GPU)",
   "id": "lc-29869-m3u-dflash-prose-master",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #29869, M3 Ultra (60-core GPU), Qwen3.8-27B Q4_0 target + DFlash2 Q8_0 drafter, -ngl 99 -fa on -c 8192 -np 1, 64 generated tokens, median of 5 requests, mean of 2 server runs; DFlash2 prose T=0, master build; config=dflash-prose-master",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Qwen3.8-27B DFlash2 Q8_0 drafter; M3 Ultra (60-core GPU); DFlash2 prose 0 16.8 62.6",
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp PR #29869 (Metal few-row MMA for spec decode)",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29869",
   "tg_tokens": 64,
   "tps": "16.8",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "M3 Ultra (60-core GPU)",
   "id": "lc-29869-m3u-dflash-prose-pr",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #29869, M3 Ultra (60-core GPU), Qwen3.8-27B Q4_0 target + DFlash2 Q8_0 drafter, -ngl 99 -fa on -c 8192 -np 1, 64 generated tokens, median of 5 requests, mean of 2 server runs; DFlash2 prose T=0, this PR build; config=dflash-prose-pr",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Qwen3.8-27B DFlash2 Q8_0 drafter; M3 Ultra (60-core GPU); DFlash2 prose 0 16.8 62.6",
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp PR #29869 (Metal few-row MMA for spec decode)",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29869",
   "tg_tokens": 64,
   "tps": "62.6",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "M3 Ultra (60-core GPU)",
   "id": "lc-29869-m3u-serial-code-master",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #29869, M3 Ultra (60-core GPU), Qwen3.8-27B Q4_0 target + DFlash2 Q8_0 drafter, -ngl 99 -fa on -c 8192 -np 1, 64 generated tokens, median of 5 requests, mean of 2 server runs; serial code T=0 baseline, master build; config=serial-code-master",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Qwen3.8-27B DFlash2 Q8_0 drafter; M3 Ultra (60-core GPU); serial code 0 32.1 32.0",
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp PR #29869 (Metal few-row MMA for spec decode)",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29869",
   "tg_tokens": 64,
   "tps": "32.1",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "M3 Ultra (60-core GPU)",
   "id": "lc-29869-m3u-serial-code-pr",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #29869, M3 Ultra (60-core GPU), Qwen3.8-27B Q4_0 target + DFlash2 Q8_0 drafter, -ngl 99 -fa on -c 8192 -np 1, 64 generated tokens, median of 5 requests, mean of 2 server runs; serial code T=0 baseline, this PR build; config=serial-code-pr",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Qwen3.8-27B DFlash2 Q8_0 drafter; M3 Ultra (60-core GPU); serial code 0 32.1 32.0",
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp PR #29869 (Metal few-row MMA for spec decode)",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29869",
   "tg_tokens": 64,
   "tps": "32.0",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "M3 Ultra (60-core GPU)",
   "id": "lc-29869-m3u-serial-prose-master",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #29869, M3 Ultra (60-core GPU), Qwen3.8-27B Q4_0 target + DFlash2 Q8_0 drafter, -ngl 99 -fa on -c 8192 -np 1, 64 generated tokens, median of 5 requests, mean of 2 server runs; serial prose T=0 baseline, master build; config=serial-prose-master",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Qwen3.8-27B DFlash2 Q8_0 drafter; M3 Ultra (60-core GPU); serial prose 0 32.1 32.0",
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp PR #29869 (Metal few-row MMA for spec decode)",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29869",
   "tg_tokens": 64,
   "tps": "32.1",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "M3 Ultra (60-core GPU)",
   "id": "lc-29869-m3u-serial-prose-pr",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #29869, M3 Ultra (60-core GPU), Qwen3.8-27B Q4_0 target + DFlash2 Q8_0 drafter, -ngl 99 -fa on -c 8192 -np 1, 64 generated tokens, median of 5 requests, mean of 2 server runs; serial prose T=0 baseline, this PR build; config=serial-prose-pr",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Qwen3.8-27B DFlash2 Q8_0 drafter; M3 Ultra (60-core GPU); serial prose 0 32.1 32.0",
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp PR #29869 (Metal few-row MMA for spec decode)",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29869",
   "tg_tokens": 64,
   "tps": "32.0",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "M3 Ultra (60-core GPU)",
   "id": "lc-29869-m3u-bench-master",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #29869, M3 Ultra (60-core GPU), Qwen3.8-27B Q4_0 target + DFlash2 Q8_0 drafter, -ngl 99 -fa on -c 8192 -np 1, 64 generated tokens, median of 5 requests, mean of 2 server runs; llama-bench -fa 1 pp512/tg128 (two interleaved runs, first shown), master build; config=bench-master",
   "pp_tokens": 512,
   "pp_tps": "317.2",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Qwen3.8-27B; M3 Ultra (60-core GPU); pp512 317.2 t/s on both builds, and tg128 32.45 t/s on master and 32.36 t/s with this PR",
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp PR #29869 (Metal few-row MMA for spec decode)",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29869",
   "tg_tokens": 128,
   "tps": "32.45",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "M3 Ultra (60-core GPU)",
   "id": "lc-29869-m3u-bench-pr",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #29869, M3 Ultra (60-core GPU), Qwen3.8-27B Q4_0 target + DFlash2 Q8_0 drafter, -ngl 99 -fa on -c 8192 -np 1, 64 generated tokens, median of 5 requests, mean of 2 server runs; llama-bench -fa 1 pp512/tg128 (two interleaved runs, first shown), this PR build; config=bench-pr",
   "pp_tokens": 512,
   "pp_tps": "317.2",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Qwen3.8-27B; M3 Ultra (60-core GPU); pp512 317.2 t/s on both builds, and tg128 32.45 t/s on master and 32.36 t/s with this PR",
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp PR #29869 (Metal few-row MMA for spec decode)",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29869",
   "tg_tokens": 128,
   "tps": "32.36",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": 1,
   "ctx": 262144,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "cb-swift15-m5pro-weighted",
   "model": "Swift 1.5",
   "notes": "Splash (Metal) with DFlash2, 48 GB M5 Pro MacBook Pro (20 GPU cores); Q4/Q6/Q8 mixed-precision build of the Swift 1.5 fine-tune of Qwen3.8 27B (base ukisai/Swift-1.5-Qwen3.8-27b), 262144-token window; request-level throughput (prompt+gen mixed) unless noted, one request; weighted per-request throughput; config=weighted; median TTFT 1.93 s",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "Q4/Q6/Q8 mixed",
   "quote": "Badger Mixed is a Q4/Q6/Q8 mixed-precision build of Swift 1.5 designed for fast local agents on 48 GB Apple Silicon; Measured on a 48 GB MacBook Pro with an M5 Pro and 20 GPU cores; Weighted per-request throughput 32.58 tok/s",
   "retrieved": "2026-10-02",
   "source_name": "HF model card: CleverBadger/Swift-1.5-Qwen3.8-27B-Badger-Mixed-GGUF",
   "source_url": "https://huggingface.co/CleverBadger/Swift-1.5-Qwen3.8-27B-Badger-Mixed-GGUF",
   "tg_tokens": null,
   "tps": "32.58",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": 1,
   "ctx": 262144,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "cb-swift15-m5pro-median",
   "model": "Swift 1.5",
   "notes": "Splash (Metal) with DFlash2, 48 GB M5 Pro MacBook Pro (20 GPU cores); Q4/Q6/Q8 mixed-precision build of the Swift 1.5 fine-tune of Qwen3.8 27B (base ukisai/Swift-1.5-Qwen3.8-27b), 262144-token window; request-level throughput (prompt+gen mixed) unless noted, one request; median per-request throughput; config=median",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "Q4/Q6/Q8 mixed",
   "quote": "Badger Mixed is a Q4/Q6/Q8 mixed-precision build of Swift 1.5 designed for fast local agents on 48 GB Apple Silicon; Measured on a 48 GB MacBook Pro with an M5 Pro and 20 GPU cores; Median per-request throughput 34.42 tok/s",
   "retrieved": "2026-10-02",
   "source_name": "HF model card: CleverBadger/Swift-1.5-Qwen3.8-27B-Badger-Mixed-GGUF",
   "source_url": "https://huggingface.co/CleverBadger/Swift-1.5-Qwen3.8-27B-Badger-Mixed-GGUF",
   "tg_tokens": null,
   "tps": "34.42",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": 1,
   "ctx": 262144,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "cb-swift15-m5pro-thinking",
   "model": "Swift 1.5",
   "notes": "Splash (Metal) with DFlash2, 48 GB M5 Pro MacBook Pro (20 GPU cores); Q4/Q6/Q8 mixed-precision build of the Swift 1.5 fine-tune of Qwen3.8 27B (base ukisai/Swift-1.5-Qwen3.8-27b), 262144-token window; request-level throughput (prompt+gen mixed) unless noted, one request; thinking throughput; config=thinking",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "Q4/Q6/Q8 mixed",
   "quote": "Badger Mixed is a Q4/Q6/Q8 mixed-precision build of Swift 1.5 designed for fast local agents on 48 GB Apple Silicon; Measured on a 48 GB MacBook Pro with an M5 Pro and 20 GPU cores; Thinking throughput 32.43 tok/s",
   "retrieved": "2026-10-02",
   "source_name": "HF model card: CleverBadger/Swift-1.5-Qwen3.8-27B-Badger-Mixed-GGUF",
   "source_url": "https://huggingface.co/CleverBadger/Swift-1.5-Qwen3.8-27B-Badger-Mixed-GGUF",
   "tg_tokens": null,
   "tps": "32.43",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": 1,
   "ctx": 262144,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "cb-swift15-m5pro-code",
   "model": "Swift 1.5",
   "notes": "Splash (Metal) with DFlash2, 48 GB M5 Pro MacBook Pro (20 GPU cores); Q4/Q6/Q8 mixed-precision build of the Swift 1.5 fine-tune of Qwen3.8 27B (base ukisai/Swift-1.5-Qwen3.8-27b), 262144-token window; request-level throughput (prompt+gen mixed) unless noted, one request; code throughput; config=code",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "Q4/Q6/Q8 mixed",
   "quote": "Badger Mixed is a Q4/Q6/Q8 mixed-precision build of Swift 1.5 designed for fast local agents on 48 GB Apple Silicon; Measured on a 48 GB MacBook Pro with an M5 Pro and 20 GPU cores; Code throughput 67.52 tok/s",
   "retrieved": "2026-10-02",
   "source_name": "HF model card: CleverBadger/Swift-1.5-Qwen3.8-27B-Badger-Mixed-GGUF",
   "source_url": "https://huggingface.co/CleverBadger/Swift-1.5-Qwen3.8-27B-Badger-Mixed-GGUF",
   "tg_tokens": null,
   "tps": "67.52",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "AMD Ryzen 7 7735HS",
   "id": "lc-29875-ryzen-code-static",
   "model": "Qwen2.5 3B",
   "notes": "llama.cpp issue #29875 feature request (open, 0 comments): `--spec-draft-entropy-max` Shannon-entropy-gated early-exit speculative drafting; AMD Ryzen 7 7735HS, 8 CPU threads, Qwen2.5-3B target + Qwen2.5-0.5B draft, K=8 draft budget, 128+ tokens; gating strategies compared: Static Speculation (no gate), --spec-draft-p-min 0.6, Shannon-entropy gate (entropy 1.5); unmerged working branch feat/entropy-gated-speculation; spec-decode speed (draft+target); code task, Static Speculation; config=code-static",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "I benchmarked Shannon entropy gating head-to-head against `--spec-draft-p-min 0.6` on an AMD Ryzen 7 7735HS (8 CPU threads) using Qwen2.5-3B (target) and Qwen2.5-0.5B (draft) with K=8 draft budget over 128+ tokens; #### 1. Code Generation (`longest common subsequence`, 134 tokens); | Static Speculation (no gate) | 11.71 t/s | 192 | 110 | 57.29% |",
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp issue #29875 (feature request)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29875",
   "tg_tokens": null,
   "tps": "11.71",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "AMD Ryzen 7 7735HS",
   "id": "lc-29875-ryzen-code-pmin06",
   "model": "Qwen2.5 3B",
   "notes": "llama.cpp issue #29875 feature request (open, 0 comments): `--spec-draft-entropy-max` Shannon-entropy-gated early-exit speculative drafting; AMD Ryzen 7 7735HS, 8 CPU threads, Qwen2.5-3B target + Qwen2.5-0.5B draft, K=8 draft budget, 128+ tokens; gating strategies compared: Static Speculation (no gate), --spec-draft-p-min 0.6, Shannon-entropy gate (entropy 1.5); unmerged working branch feat/entropy-gated-speculation; spec-decode speed (draft+target); code task, p_min 0.6; config=code-pmin06",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "I benchmarked Shannon entropy gating head-to-head against `--spec-draft-p-min 0.6` on an AMD Ryzen 7 7735HS (8 CPU threads) using Qwen2.5-3B (target) and Qwen2.5-0.5B (draft) with K=8 draft budget over 128+ tokens; #### 1. Code Generation (`longest common subsequence`, 134 tokens); | `--spec-draft-p-min 0.6` | 14.83 t/s | 138 | 109 | 78.99% |",
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp issue #29875 (feature request)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29875",
   "tg_tokens": null,
   "tps": "14.83",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "AMD Ryzen 7 7735HS",
   "id": "lc-29875-ryzen-code-entropy15",
   "model": "Qwen2.5 3B",
   "notes": "llama.cpp issue #29875 feature request (open, 0 comments): `--spec-draft-entropy-max` Shannon-entropy-gated early-exit speculative drafting; AMD Ryzen 7 7735HS, 8 CPU threads, Qwen2.5-3B target + Qwen2.5-0.5B draft, K=8 draft budget, 128+ tokens; gating strategies compared: Static Speculation (no gate), --spec-draft-p-min 0.6, Shannon-entropy gate (entropy 1.5); unmerged working branch feat/entropy-gated-speculation; spec-decode speed (draft+target); code task, entropy gate; config=code-entropy15",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "I benchmarked Shannon entropy gating head-to-head against `--spec-draft-p-min 0.6` on an AMD Ryzen 7 7735HS (8 CPU threads) using Qwen2.5-3B (target) and Qwen2.5-0.5B (draft) with K=8 draft budget over 128+ tokens; #### 1. Code Generation (`longest common subsequence`, 134 tokens); | **Entropy Gate (`entropy 1.5`)** | **16.50 t/s** | **130** | **115** | **88.46%** |",
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp issue #29875 (feature request)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29875",
   "tg_tokens": null,
   "tps": "16.50",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "AMD Ryzen 7 7735HS",
   "id": "lc-29875-ryzen-riddle-static",
   "model": "Qwen2.5 3B",
   "notes": "llama.cpp issue #29875 feature request (open, 0 comments): `--spec-draft-entropy-max` Shannon-entropy-gated early-exit speculative drafting; AMD Ryzen 7 7735HS, 8 CPU threads, Qwen2.5-3B target + Qwen2.5-0.5B draft, K=8 draft budget, 128+ tokens; gating strategies compared: Static Speculation (no gate), --spec-draft-p-min 0.6, Shannon-entropy gate (entropy 1.5); unmerged working branch feat/entropy-gated-speculation; spec-decode speed (draft+target); riddle/divergent reasoning task, Static Speculation; config=riddle-static",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "I benchmarked Shannon entropy gating head-to-head against `--spec-draft-p-min 0.6` on an AMD Ryzen 7 7735HS (8 CPU threads) using Qwen2.5-3B (target) and Qwen2.5-0.5B (draft) with K=8 draft budget over 128+ tokens; #### 2. Divergent Reasoning / Riddle (130 tokens); | Static Speculation (no gate) | 6.70 t/s | 324 | 90 | 27.78% |",
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp issue #29875 (feature request)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29875",
   "tg_tokens": null,
   "tps": "6.70",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "AMD Ryzen 7 7735HS",
   "id": "lc-29875-ryzen-riddle-pmin06",
   "model": "Qwen2.5 3B",
   "notes": "llama.cpp issue #29875 feature request (open, 0 comments): `--spec-draft-entropy-max` Shannon-entropy-gated early-exit speculative drafting; AMD Ryzen 7 7735HS, 8 CPU threads, Qwen2.5-3B target + Qwen2.5-0.5B draft, K=8 draft budget, 128+ tokens; gating strategies compared: Static Speculation (no gate), --spec-draft-p-min 0.6, Shannon-entropy gate (entropy 1.5); unmerged working branch feat/entropy-gated-speculation; spec-decode speed (draft+target); riddle/divergent reasoning task, p_min 0.6; config=riddle-pmin06",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "I benchmarked Shannon entropy gating head-to-head against `--spec-draft-p-min 0.6` on an AMD Ryzen 7 7735HS (8 CPU threads) using Qwen2.5-3B (target) and Qwen2.5-0.5B (draft) with K=8 draft budget over 128+ tokens; #### 2. Divergent Reasoning / Riddle (130 tokens); | `--spec-draft-p-min 0.6` | 12.91 t/s | 102 | 73 | 71.57% |",
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp issue #29875 (feature request)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29875",
   "tg_tokens": null,
   "tps": "12.91",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "AMD Ryzen 7 7735HS",
   "id": "lc-29875-ryzen-riddle-entropy15",
   "model": "Qwen2.5 3B",
   "notes": "llama.cpp issue #29875 feature request (open, 0 comments): `--spec-draft-entropy-max` Shannon-entropy-gated early-exit speculative drafting; AMD Ryzen 7 7735HS, 8 CPU threads, Qwen2.5-3B target + Qwen2.5-0.5B draft, K=8 draft budget, 128+ tokens; gating strategies compared: Static Speculation (no gate), --spec-draft-p-min 0.6, Shannon-entropy gate (entropy 1.5); unmerged working branch feat/entropy-gated-speculation; spec-decode speed (draft+target); riddle/divergent reasoning task, entropy gate; config=riddle-entropy15",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "I benchmarked Shannon entropy gating head-to-head against `--spec-draft-p-min 0.6` on an AMD Ryzen 7 7735HS (8 CPU threads) using Qwen2.5-3B (target) and Qwen2.5-0.5B (draft) with K=8 draft budget over 128+ tokens; #### 2. Divergent Reasoning / Riddle (130 tokens); | **Entropy Gate (`entropy 1.5`)** | **13.19 t/s** | **78** | **66** | **84.62%** |",
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp issue #29875 (feature request)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29875",
   "tg_tokens": null,
   "tps": "13.19",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "8 threads",
   "id": "davidmg-qwen38flash-mtp-nospec",
   "model": "Qwen3.8-Flash-Next",
   "notes": "HF card field note: MTP draft-mtp speculative decoding (--spec-type draft-mtp) on the Qwen3.8-Flash-Next 176B MoE pruned 512 -> 256 experts (GSQ-RCO), main model IQ1_M + separate Q8_0 MTP head; CPU, 8 threads, ctx 4096, one code prompt, temperature 0, thinking off, 160 tokens; author: 'treat as indications, not benchmarks'; a separate MTP head loaded as its own model via -md; serial decode baseline (no speculation), 3.49 tok/s; config=nospec",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "IQ1_M main / Q8_0 head",
   "quote": "Qwen3.8-Flash-Next \u2014 separate MTP head on the expert-pruned (256-expert) GSQ-RCO model; CPU only (8 threads, ctx 4096, one code prompt, 160 tokens); llama-server -m Qwen3.8-Flash-Next-GSQ-RCO-IQ1_M-00001-of-00002.gguf; -md mtp-Qwen3.8-Flash-Next-256exp-Q8_0.gguf; no speculation 3.49",
   "retrieved": "2026-10-02",
   "source_name": "HF model card: Davidmg0815/Qwen3.8-Flash-Next-MTP-256exp-GGUF",
   "source_url": "https://huggingface.co/Davidmg0815/Qwen3.8-Flash-Next-MTP-256exp-GGUF",
   "tg_tokens": null,
   "tps": "3.49",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "8 threads",
   "id": "davidmg-qwen38flash-mtp-h512",
   "model": "Qwen3.8-Flash-Next",
   "notes": "HF card field note: MTP draft-mtp speculative decoding (--spec-type draft-mtp) on the Qwen3.8-Flash-Next 176B MoE pruned 512 -> 256 experts (GSQ-RCO), main model IQ1_M + separate Q8_0 MTP head; CPU, 8 threads, ctx 4096, one code prompt, temperature 0, thinking off, 160 tokens; author: 'treat as indications, not benchmarks'; a separate MTP head loaded as its own model via -md; unsloth full MTP head (512 experts, 4.1 GB), 5.61 tok/s (+61%), draft acceptance 81.7% (98/120); config=h512",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "IQ1_M main / Q8_0 head",
   "quote": "Qwen3.8-Flash-Next \u2014 separate MTP head on the expert-pruned (256-expert) GSQ-RCO model; CPU only (8 threads, ctx 4096, one code prompt, 160 tokens); llama-server -m Qwen3.8-Flash-Next-GSQ-RCO-IQ1_M-00001-of-00002.gguf; -md mtp-Qwen3.8-Flash-Next-256exp-Q8_0.gguf; original head, 512 experts 5.61 (+61 %) 81.7 % (98/120)",
   "retrieved": "2026-10-02",
   "source_name": "HF model card: Davidmg0815/Qwen3.8-Flash-Next-MTP-256exp-GGUF",
   "source_url": "https://huggingface.co/Davidmg0815/Qwen3.8-Flash-Next-MTP-256exp-GGUF",
   "tg_tokens": null,
   "tps": "5.61",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "8 threads",
   "id": "davidmg-qwen38flash-mtp-h256",
   "model": "Qwen3.8-Flash-Next",
   "notes": "HF card field note: MTP draft-mtp speculative decoding (--spec-type draft-mtp) on the Qwen3.8-Flash-Next 176B MoE pruned 512 -> 256 experts (GSQ-RCO), main model IQ1_M + separate Q8_0 MTP head; CPU, 8 threads, ctx 4096, one code prompt, temperature 0, thinking off, 160 tokens; author: 'treat as indications, not benchmarks'; a separate MTP head loaded as its own model via -md; pruned MTP head (256 experts, 2.8 GB), 5.13 tok/s (+47%), draft acceptance 74.8% (95/127); config=h256",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "IQ1_M main / Q8_0 head",
   "quote": "Qwen3.8-Flash-Next \u2014 separate MTP head on the expert-pruned (256-expert) GSQ-RCO model; CPU only (8 threads, ctx 4096, one code prompt, 160 tokens); llama-server -m Qwen3.8-Flash-Next-GSQ-RCO-IQ1_M-00001-of-00002.gguf; -md mtp-Qwen3.8-Flash-Next-256exp-Q8_0.gguf; this head, 256 experts 5.13 (+47 %) 74.8 % (95/127)",
   "retrieved": "2026-10-02",
   "source_name": "HF model card: Davidmg0815/Qwen3.8-Flash-Next-MTP-256exp-GGUF",
   "source_url": "https://huggingface.co/Davidmg0815/Qwen3.8-Flash-Next-MTP-256exp-GGUF",
   "tg_tokens": null,
   "tps": "5.13",
   "ttft_s": null
  },
  {
   "backend": "training (TPU)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "A100-SXM4-40GB",
   "id": "resonatex-t1-tpuv6e-training",
   "model": "ResonateX T1",
   "notes": "HF card: training (not inference) telemetry for the 125.3M-parameter ResonateX T1 talking model, 1 TPU v6e worker, 22,100 steps, 724.2M tokens (5.78 tok/param), single run; tps is the average training throughput over the run (throughput stabilized ~53K tok/s after warm-up); average training throughput 53,225 tok/s over 22,100 steps (3.83 h), 1 TPU v6e worker",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "ResonatexIntegratedTechnologies/ResonateX-T1-125M-Talking-GGUF; A100-SXM4-40GB; Throughput 162,214 tok/s",
   "retrieved": "2026-10-02",
   "source_name": "HF model card: ResonatexIntegratedTechnologies/ResonateX-T1-125M-Talking-GGUF",
   "source_url": "https://huggingface.co/ResonatexIntegratedTechnologies/ResonateX-T1-125M-Talking-GGUF",
   "tg_tokens": null,
   "tps": "162214",
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "B70 Arc Pro",
   "id": "lc-29882-b70-gemma4-fa0-b11369",
   "model": "Gemma4 26B",
   "notes": "llama.cpp PR #29882, Vulkan rms-norm subgroup-reduction opt before/after; config=fa0-b11369; ngl -1; r 5; Windows",
   "pp_tokens": 8192,
   "pp_tps": "1525.36",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; B70 Arc Pro (Windows); gemma4 26B.A4B Q4_K - Medium 15.70 GiB 25.23 B Vulkan -1 0 pp8192 1525.36 \u00b1 6.33",
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp PR #29882 (Vulkan rms-norm subgroup opt, before/after build)",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29882",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "B70 Arc Pro",
   "id": "lc-29882-b70-gemma4-fa1-b11369",
   "model": "Gemma4 26B",
   "notes": "llama.cpp PR #29882, Vulkan rms-norm subgroup-reduction opt before/after; config=fa1-b11369; ngl -1; r 5; Windows",
   "pp_tokens": 8192,
   "pp_tps": "874.89",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; B70 Arc Pro (Windows); gemma4 26B.A4B Q4_K - Medium 15.70 GiB 25.23 B Vulkan -1 1 pp8192 874.89 \u00b1 2.02",
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp PR #29882 (Vulkan rms-norm subgroup opt, before/after build)",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29882",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "B70 Arc Pro",
   "id": "lc-29882-b70-gemma4-fa0-b11371",
   "model": "Gemma4 26B",
   "notes": "llama.cpp PR #29882, Vulkan rms-norm subgroup-reduction opt before/after; config=fa0-b11371; ngl -1; r 5; Windows",
   "pp_tokens": 8192,
   "pp_tps": "1544.14",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; B70 Arc Pro (Windows); gemma4 26B.A4B Q4_K - Medium 15.70 GiB 25.23 B Vulkan -1 0 pp8192 1544.14 \u00b1 6.48",
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp PR #29882 (Vulkan rms-norm subgroup opt, before/after build)",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29882",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "B70 Arc Pro",
   "id": "lc-29882-b70-gemma4-fa1-b11371",
   "model": "Gemma4 26B",
   "notes": "llama.cpp PR #29882, Vulkan rms-norm subgroup-reduction opt before/after; config=fa1-b11371; ngl -1; r 5; Windows",
   "pp_tokens": 8192,
   "pp_tps": "880.44",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "gemma-4-26B-A4B-it-UD-Q4_K_M.gguf; B70 Arc Pro (Windows); gemma4 26B.A4B Q4_K - Medium 15.70 GiB 25.23 B Vulkan -1 1 pp8192 880.44 \u00b1 2.10",
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp PR #29882 (Vulkan rms-norm subgroup opt, before/after build)",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29882",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": 1,
   "ctx": 512,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "B70 Arc Pro",
   "id": "lc-29882-b70-qwen36-fa0-b11369",
   "model": "Qwen3.6-35B-A3B",
   "notes": "llama.cpp PR #29882, Vulkan rms-norm subgroup-reduction opt before/after; config=fa0-b11369; ngl -1; r 5; Windows",
   "pp_tokens": 512,
   "pp_tps": "1589.22",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Qwen3.6-35B-A3B-UD-Q4_K_M.gguf; B70 Arc Pro (Windows); qwen35moe 35B.A3B Q4_K - Medium 20.60 GiB 34.66 B Vulkan -1 0 pp512 1589.22 \u00b1 8.57; qwen35moe 35B.A3B Q4_K - Medium 20.60 GiB 34.66 B Vulkan -1 0 tg128 105.92 \u00b1 0.17",
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp PR #29882 (Vulkan rms-norm subgroup opt, before/after build)",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29882",
   "tg_tokens": 128,
   "tps": "105.92",
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": 1,
   "ctx": 512,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "B70 Arc Pro",
   "id": "lc-29882-b70-qwen36-fa1-b11369",
   "model": "Qwen3.6-35B-A3B",
   "notes": "llama.cpp PR #29882, Vulkan rms-norm subgroup-reduction opt before/after; config=fa1-b11369; ngl -1; r 5; Windows",
   "pp_tokens": 512,
   "pp_tps": "1555.32",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Qwen3.6-35B-A3B-UD-Q4_K_M.gguf; B70 Arc Pro (Windows); qwen35moe 35B.A3B Q4_K - Medium 20.60 GiB 34.66 B Vulkan -1 1 pp512 1555.32 \u00b1 11.68; qwen35moe 35B.A3B Q4_K - Medium 20.60 GiB 34.66 B Vulkan -1 1 tg128 106.39 \u00b1 0.18",
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp PR #29882 (Vulkan rms-norm subgroup opt, before/after build)",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29882",
   "tg_tokens": 128,
   "tps": "106.39",
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": 1,
   "ctx": 512,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "B70 Arc Pro",
   "id": "lc-29882-b70-qwen36-fa0-b11371",
   "model": "Qwen3.6-35B-A3B",
   "notes": "llama.cpp PR #29882, Vulkan rms-norm subgroup-reduction opt before/after; config=fa0-b11371; ngl -1; r 5; Windows",
   "pp_tokens": 512,
   "pp_tps": "1620.24",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Qwen3.6-35B-A3B-UD-Q4_K_M.gguf; B70 Arc Pro (Windows); qwen35moe 35B.A3B Q4_K - Medium 20.60 GiB 34.66 B Vulkan -1 0 pp512 1620.24 \u00b1 14.11; qwen35moe 35B.A3B Q4_K - Medium 20.60 GiB 34.66 B Vulkan -1 0 tg128 106.28 \u00b1 0.15",
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp PR #29882 (Vulkan rms-norm subgroup opt, before/after build)",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29882",
   "tg_tokens": 128,
   "tps": "106.28",
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": 1,
   "ctx": 512,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "B70 Arc Pro",
   "id": "lc-29882-b70-qwen36-fa1-b11371",
   "model": "Qwen3.6-35B-A3B",
   "notes": "llama.cpp PR #29882, Vulkan rms-norm subgroup-reduction opt before/after; config=fa1-b11371; ngl -1; r 5; Windows",
   "pp_tokens": 512,
   "pp_tps": "1585.87",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Qwen3.6-35B-A3B-UD-Q4_K_M.gguf; B70 Arc Pro (Windows); qwen35moe 35B.A3B Q4_K - Medium 20.60 GiB 34.66 B Vulkan -1 1 pp512 1585.87 \u00b1 7.19; qwen35moe 35B.A3B Q4_K - Medium 20.60 GiB 34.66 B Vulkan -1 1 tg128 106.77 \u00b1 0.18",
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp PR #29882 (Vulkan rms-norm subgroup opt, before/after build)",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29882",
   "tg_tokens": 128,
   "tps": "106.77",
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": 1,
   "ctx": 512,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "B70 Arc Pro",
   "id": "lc-29882-b70-qwen06-fa0-b11369",
   "model": "Qwen3-0.6B",
   "notes": "llama.cpp PR #29882, Vulkan rms-norm subgroup-reduction opt before/after; config=fa0-b11369; ngl -1; r 5; Windows",
   "pp_tokens": 512,
   "pp_tps": "21235.41",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Qwen3-0.6B-Q4_K_M.gguf; B70 Arc Pro (Windows); qwen3 0.6B Q4_K - Medium 456.11 MiB 751.63 M Vulkan -1 0 pp512 21235.41 \u00b1 672.52; qwen3 0.6B Q4_K - Medium 456.11 MiB 751.63 M Vulkan -1 0 tg128 452.70 \u00b1 4.64",
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp PR #29882 (Vulkan rms-norm subgroup opt, before/after build)",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29882",
   "tg_tokens": 128,
   "tps": "452.70",
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": 1,
   "ctx": 512,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "B70 Arc Pro",
   "id": "lc-29882-b70-qwen06-fa1-b11369",
   "model": "Qwen3-0.6B",
   "notes": "llama.cpp PR #29882, Vulkan rms-norm subgroup-reduction opt before/after; config=fa1-b11369; ngl -1; r 5; Windows",
   "pp_tokens": 512,
   "pp_tps": "17508.58",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Qwen3-0.6B-Q4_K_M.gguf; B70 Arc Pro (Windows); qwen3 0.6B Q4_K - Medium 456.11 MiB 751.63 M Vulkan -1 1 pp512 17508.58 \u00b1 135.80; qwen3 0.6B Q4_K - Medium 456.11 MiB 751.63 M Vulkan -1 1 tg128 467.96 \u00b1 0.64",
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp PR #29882 (Vulkan rms-norm subgroup opt, before/after build)",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29882",
   "tg_tokens": 128,
   "tps": "467.96",
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": 1,
   "ctx": 512,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "B70 Arc Pro",
   "id": "lc-29882-b70-qwen06-fa0-b11371",
   "model": "Qwen3-0.6B",
   "notes": "llama.cpp PR #29882, Vulkan rms-norm subgroup-reduction opt before/after; config=fa0-b11371; ngl -1; r 5; Windows",
   "pp_tokens": 512,
   "pp_tps": "25346.93",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Qwen3-0.6B-Q4_K_M.gguf; B70 Arc Pro (Windows); qwen3 0.6B Q4_K - Medium 456.11 MiB 751.63 M Vulkan -1 0 pp512 25346.93 \u00b1 137.76; qwen3 0.6B Q4_K - Medium 456.11 MiB 751.63 M Vulkan -1 0 tg128 456.94 \u00b1 9.38",
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp PR #29882 (Vulkan rms-norm subgroup opt, before/after build)",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29882",
   "tg_tokens": 128,
   "tps": "456.94",
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": 1,
   "ctx": 512,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "B70 Arc Pro",
   "id": "lc-29882-b70-qwen06-fa1-b11371",
   "model": "Qwen3-0.6B",
   "notes": "llama.cpp PR #29882, Vulkan rms-norm subgroup-reduction opt before/after; config=fa1-b11371; ngl -1; r 5; Windows",
   "pp_tokens": 512,
   "pp_tps": "20217.98",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Qwen3-0.6B-Q4_K_M.gguf; B70 Arc Pro (Windows); qwen3 0.6B Q4_K - Medium 456.11 MiB 751.63 M Vulkan -1 1 pp512 20217.98 \u00b1 89.04; qwen3 0.6B Q4_K - Medium 456.11 MiB 751.63 M Vulkan -1 1 tg128 475.76 \u00b1 0.77",
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp PR #29882 (Vulkan rms-norm subgroup opt, before/after build)",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29882",
   "tg_tokens": 128,
   "tps": "475.76",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 2048,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "2\u00d7 Quadro RTX 4000",
   "id": "livesport-rune26b-rtx4000-pp2048",
   "model": "rune-26b-a4b-v3",
   "notes": "HF card (re-mined 2026-10-03 after the card dropped the -ub batch sweep): prompt processing (prefill) of a ~2,000-token prompt on 2x Quadro RTX 4000 8 GB, IQ3_S-dyn 26B MoE (gemma4 arch) + q8_0 KV. llama-server 2,098 tok/s; llama-bench pp2048 cross-check 2,049-2,065 tok/s.",
   "pp_tokens": 2048,
   "pp_tps": "2098",
   "provenance": "community",
   "quant": "IQ3_S-dyn",
   "quote": "Livesport/rune-26b-a4b-v3-GGUF; Running on 2\u00d7 Quadro RTX 4000 8 GB (measured); 2,098 tok/s (llama-server)",
   "retrieved": "2026-10-03",
   "source_name": "Livesport/rune-26b-a4b-v3-GGUF",
   "source_url": "https://huggingface.co/Livesport/rune-26b-a4b-v3-GGUF",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 14478,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "2\u00d7 Quadro RTX 4000",
   "id": "livesport-rune26b-rtx4000-pp14478",
   "model": "rune-26b-a4b-v3",
   "notes": "HF card (re-mined): prefill of a 14,478-token prompt on 2x Quadro RTX 4000 8 GB, IQ3_S-dyn. 1,904 tok/s, essentially flat vs the ~2,000-token rate (no long-context prefill cliff).",
   "pp_tokens": 14478,
   "pp_tps": "1904",
   "provenance": "community",
   "quant": "IQ3_S-dyn",
   "quote": "Livesport/rune-26b-a4b-v3-GGUF; Running on 2\u00d7 Quadro RTX 4000 8 GB (measured); prompt processing, 14,478 tokens 1,904 tok/s",
   "retrieved": "2026-10-03",
   "source_name": "Livesport/rune-26b-a4b-v3-GGUF",
   "source_url": "https://huggingface.co/Livesport/rune-26b-a4b-v3-GGUF",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59824-c10-off",
   "model": "Kimi-K3",
   "notes": "Kimi-K3 FP4 TP8 MI355X ROCm, DSpark K=7, 16k-token code prompts, 1024 output tokens, fresh servers. adaptive verification off (static K=7). config=adv-c10-off",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP4",
   "quote": "Kimi-K3 FP4, TP8, MI355X; conc 10, out tok/s 453.4 485.5",
   "retrieved": "2026-10-03",
   "source_name": "vllm-project/vllm#59824",
   "source_url": "https://github.com/vllm-project/vllm/pull/59824",
   "tg_tokens": null,
   "tps": "453.4",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59824-c10-on",
   "model": "Kimi-K3",
   "notes": "Kimi-K3 FP4 TP8 MI355X ROCm, DSpark K=7, 16k-token code prompts, 1024 output tokens, fresh servers. adaptive verification on. config=adv-c10-on",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP4",
   "quote": "Kimi-K3 FP4, TP8, MI355X; conc 10, out tok/s 453.4 485.5",
   "retrieved": "2026-10-03",
   "source_name": "vllm-project/vllm#59824",
   "source_url": "https://github.com/vllm-project/vllm/pull/59824",
   "tg_tokens": null,
   "tps": "485.5",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59824-c12-off",
   "model": "Kimi-K3",
   "notes": "Kimi-K3 FP4 TP8 MI355X ROCm, DSpark K=7, 16k-token code prompts, 1024 output tokens, fresh servers. adaptive verification off (static K=7). config=adv-c12-off",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP4",
   "quote": "Kimi-K3 FP4, TP8, MI355X; conc 12, out tok/s 499.3 510.9",
   "retrieved": "2026-10-03",
   "source_name": "vllm-project/vllm#59824",
   "source_url": "https://github.com/vllm-project/vllm/pull/59824",
   "tg_tokens": null,
   "tps": "499.3",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59824-c12-on",
   "model": "Kimi-K3",
   "notes": "Kimi-K3 FP4 TP8 MI355X ROCm, DSpark K=7, 16k-token code prompts, 1024 output tokens, fresh servers. adaptive verification on. config=adv-c12-on",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP4",
   "quote": "Kimi-K3 FP4, TP8, MI355X; conc 12, out tok/s 499.3 510.9",
   "retrieved": "2026-10-03",
   "source_name": "vllm-project/vllm#59824",
   "source_url": "https://github.com/vllm-project/vllm/pull/59824",
   "tg_tokens": null,
   "tps": "510.9",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59824-c14-off",
   "model": "Kimi-K3",
   "notes": "Kimi-K3 FP4 TP8 MI355X ROCm, DSpark K=7, 16k-token code prompts, 1024 output tokens, fresh servers. adaptive verification off (static K=7). config=adv-c14-off",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP4",
   "quote": "Kimi-K3 FP4, TP8, MI355X; conc 14, out tok/s (3 runs) 515.1 529.7",
   "retrieved": "2026-10-03",
   "source_name": "vllm-project/vllm#59824",
   "source_url": "https://github.com/vllm-project/vllm/pull/59824",
   "tg_tokens": null,
   "tps": "515.1",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-59824-c14-on",
   "model": "Kimi-K3",
   "notes": "Kimi-K3 FP4 TP8 MI355X ROCm, DSpark K=7, 16k-token code prompts, 1024 output tokens, fresh servers. adaptive verification on. config=adv-c14-on",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP4",
   "quote": "Kimi-K3 FP4, TP8, MI355X; conc 14, out tok/s (3 runs) 515.1 529.7",
   "retrieved": "2026-10-03",
   "source_name": "vllm-project/vllm#59824",
   "source_url": "https://github.com/vllm-project/vllm/pull/59824",
   "tg_tokens": null,
   "tps": "529.7",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 512,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "Dimensity 9400",
   "id": "lc-29884-d9400-armv8.0_1",
   "model": "Qwen3-Embedding-0.6B",
   "notes": "llama.cpp #29884 CPU backend variant sweep, MediaTek Dimensity 9400 (MT6991), Android 16, Qwen3-Embedding-0.6B Q8_0, llama-bench -p 512 -r 3 (pp512), each variant run in isolation. The ggml_backend_score function selects armv9.0_1 (123.45) but armv8.6_1 (259.77) is the fastest; armv9.2_1/2 are not loaded (no SME). config=armv8.0_1; baseline, no dotprod/i8mm",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "102.33",
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Model: Qwen3-Embedding-0.6B-Q8_0.gguf; MediaTek Dimensity 9400 (MT6991), Android 16; android_armv8.0_1 102.33",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp issue #29884 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29884",
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 512,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "Dimensity 9400",
   "id": "lc-29884-d9400-armv8.2_1",
   "model": "Qwen3-Embedding-0.6B",
   "notes": "llama.cpp #29884 CPU backend variant sweep, MediaTek Dimensity 9400 (MT6991), Android 16, Qwen3-Embedding-0.6B Q8_0, llama-bench -p 512 -r 3 (pp512), each variant run in isolation. The ggml_backend_score function selects armv9.0_1 (123.45) but armv8.6_1 (259.77) is the fastest; armv9.2_1/2 are not loaded (no SME). config=armv8.2_1",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "240.05",
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Model: Qwen3-Embedding-0.6B-Q8_0.gguf; MediaTek Dimensity 9400 (MT6991), Android 16; android_armv8.2_1 240.05",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp issue #29884 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29884",
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 512,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "Dimensity 9400",
   "id": "lc-29884-d9400-armv8.2_2",
   "model": "Qwen3-Embedding-0.6B",
   "notes": "llama.cpp #29884 CPU backend variant sweep, MediaTek Dimensity 9400 (MT6991), Android 16, Qwen3-Embedding-0.6B Q8_0, llama-bench -p 512 -r 3 (pp512), each variant run in isolation. The ggml_backend_score function selects armv9.0_1 (123.45) but armv8.6_1 (259.77) is the fastest; armv9.2_1/2 are not loaded (no SME). config=armv8.2_2",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "229.56",
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Model: Qwen3-Embedding-0.6B-Q8_0.gguf; MediaTek Dimensity 9400 (MT6991), Android 16; android_armv8.2_2 229.56",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp issue #29884 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29884",
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 512,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "Dimensity 9400",
   "id": "lc-29884-d9400-armv8.6_1",
   "model": "Qwen3-Embedding-0.6B",
   "notes": "llama.cpp #29884 CPU backend variant sweep, MediaTek Dimensity 9400 (MT6991), Android 16, Qwen3-Embedding-0.6B Q8_0, llama-bench -p 512 -r 3 (pp512), each variant run in isolation. The ggml_backend_score function selects armv9.0_1 (123.45) but armv8.6_1 (259.77) is the fastest; armv9.2_1/2 are not loaded (no SME). config=armv8.6_1; fastest, not selected",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "259.77",
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Model: Qwen3-Embedding-0.6B-Q8_0.gguf; MediaTek Dimensity 9400 (MT6991), Android 16; android_armv8.6_1 259.77",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp issue #29884 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29884",
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 512,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "Dimensity 9400",
   "id": "lc-29884-d9400-armv9.0_1",
   "model": "Qwen3-Embedding-0.6B",
   "notes": "llama.cpp #29884 CPU backend variant sweep, MediaTek Dimensity 9400 (MT6991), Android 16, Qwen3-Embedding-0.6B Q8_0, llama-bench -p 512 -r 3 (pp512), each variant run in isolation. The ggml_backend_score function selects armv9.0_1 (123.45) but armv8.6_1 (259.77) is the fastest; armv9.2_1/2 are not loaded (no SME). config=armv9.0_1; selected by default",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "123.45",
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Model: Qwen3-Embedding-0.6B-Q8_0.gguf; MediaTek Dimensity 9400 (MT6991), Android 16; android_armv9.0_1 123.45",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp issue #29884 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29884",
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hf3-qwen3635b-paw3090-mixed",
   "issue": null,
   "model": "Qwen3.6-35B-A3B",
   "notes": "config=mixed | Decode throughput, mixed general workload. PAW 1.54 bpw trellis-coded quant, one RTX 3090 (24 GB).",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "PAW 1.54 bpw",
   "quote": "Qwen/Qwen3.6-35B-A3B; RTX 3090; 133.7 tok/s mixed",
   "retrieved": "2026-10-03",
   "source_name": "HF model card: lackonendes PAW-35B-A3B GGUF",
   "source_url": "https://huggingface.co/lackonendes/PAW-35B-A3B-GGUF",
   "tg_tokens": null,
   "tps": "133.7",
   "ttft_s": null,
   "venue": "HF"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hf3-qwen3635b-paw3090-code",
   "issue": null,
   "model": "Qwen3.6-35B-A3B",
   "notes": "config=code | Decode throughput, code workload. PAW 1.54 bpw trellis-coded quant, one RTX 3090 (24 GB).",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "PAW 1.54 bpw",
   "quote": "Qwen/Qwen3.6-35B-A3B; RTX 3090; 182.7 tok/s code",
   "retrieved": "2026-10-03",
   "source_name": "HF model card: lackonendes PAW-35B-A3B GGUF",
   "source_url": "https://huggingface.co/lackonendes/PAW-35B-A3B-GGUF",
   "tg_tokens": null,
   "tps": "182.7",
   "ttft_s": null,
   "venue": "HF"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 16384,
   "flags": [],
   "hardware": "Quadro RTX 4000",
   "id": "hf3-clefflash-quadro4000-tp",
   "issue": null,
   "model": "clef-flash",
   "notes": "Serving throughput for 16,384-token requests on one Quadro RTX 4000 (Turing, 8 GB). Approximate; llama-bench flat over ubatch 256-2048, compute-bound. Same as earlier Q4_K_M build.",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q5_K_M-dyn",
   "quote": "Cloudflare/clef-flash; Quadro RTX 4000; 1,100 tokens/s",
   "retrieved": "2026-10-03",
   "source_name": "HF model card: Livesport clef-flash GGUF",
   "source_url": "https://huggingface.co/Livesport/clef-flash-GGUF",
   "tg_tokens": null,
   "tps": "1100",
   "ttft_s": null,
   "venue": "HF"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hf3-qwen3827b-paw27b-nodraft",
   "issue": null,
   "model": "Qwen3.8-27B",
   "notes": "config=nodraft | Baseline decode without the MTP drafter, short context. PAW ~2.17 bpw trellis, one RTX 3090, 256k context.",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "PAW 2.17 bpw",
   "quote": "Qwen/Qwen3.8-27B; single RTX 3090; 22.7 -> 78.4 tok/s",
   "retrieved": "2026-10-03",
   "source_name": "HF model card: lackonendes PAW-27B GGUF",
   "source_url": "https://huggingface.co/lackonendes/PAW-27B-GGUF",
   "tg_tokens": null,
   "tps": "22.7",
   "ttft_s": null,
   "venue": "HF"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hf3-qwen3827b-paw27b-mtpshort",
   "issue": null,
   "model": "Qwen3.8-27B",
   "notes": "config=mtpshort | MTP drafter (--spec-type draft-mtp) on, short context. Card: 22.7 -> 78.4 lossless.",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "PAW 2.17 bpw",
   "quote": "Qwen/Qwen3.8-27B; single RTX 3090; 78.4 tok/s short",
   "retrieved": "2026-10-03",
   "source_name": "HF model card: lackonendes PAW-27B GGUF",
   "source_url": "https://huggingface.co/lackonendes/PAW-27B-GGUF",
   "tg_tokens": null,
   "tps": "78.4",
   "ttft_s": null,
   "venue": "HF"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 191000,
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hf3-qwen3827b-paw27b-mtp191k",
   "issue": null,
   "model": "Qwen3.8-27B",
   "notes": "config=mtp191k | MTP drafter on at 191k code context (context-depth decay).",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "PAW 2.17 bpw",
   "quote": "Qwen/Qwen3.8-27B; single RTX 3090; 47.8 tok/s at 191k",
   "retrieved": "2026-10-03",
   "source_name": "HF model card: lackonendes PAW-27B GGUF",
   "source_url": "https://huggingface.co/lackonendes/PAW-27B-GGUF",
   "tg_tokens": null,
   "tps": "47.8",
   "ttft_s": null,
   "venue": "HF"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 262144,
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hf3-qwen3827b-pawx3-pp512",
   "issue": null,
   "model": "Qwen3.8-27B",
   "notes": "config=pp512 | Prompt processing, 512-token prompt. PAW X3 3.5 bpw, one RTX 3090.",
   "pp_tokens": 512,
   "pp_tps": "1076.5",
   "provenance": "community",
   "quant": "PAW 3.5 bpw",
   "quote": "Qwen/Qwen3.8-27B; one RTX 3090; PP512 (short prompt) 1076.5",
   "retrieved": "2026-10-03",
   "source_name": "HF model card: lackonendes PAW-27B-X3 GGUF",
   "source_url": "https://huggingface.co/lackonendes/PAW-27B-X3-GGUF",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null,
   "venue": "HF"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 262144,
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hf3-qwen3827b-pawx3-pp8192",
   "issue": null,
   "model": "Qwen3.8-27B",
   "notes": "config=pp8192 | Prompt processing, 8192-token prompt (chat length).",
   "pp_tokens": 8192,
   "pp_tps": "1215.4",
   "provenance": "community",
   "quant": "PAW 3.5 bpw",
   "quote": "Qwen/Qwen3.8-27B; one RTX 3090; PP8192 (chat length) 1215.4",
   "retrieved": "2026-10-03",
   "source_name": "HF model card: lackonendes PAW-27B-X3 GGUF",
   "source_url": "https://huggingface.co/lackonendes/PAW-27B-X3-GGUF",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null,
   "venue": "HF"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 262144,
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hf3-qwen3827b-pawx3-pp211k",
   "issue": null,
   "model": "Qwen3.8-27B",
   "notes": "config=pp211k | Prompt processing, 211k-token prompt at 262144 ctx (-ub 2048).",
   "pp_tokens": 211000,
   "pp_tps": "563.0",
   "provenance": "community",
   "quant": "PAW 3.5 bpw",
   "quote": "Qwen/Qwen3.8-27B; one RTX 3090; PP, 211k-token prompt at 262144 ctx (-ub 2048) 563.0",
   "retrieved": "2026-10-03",
   "source_name": "HF model card: lackonendes PAW-27B-X3 GGUF",
   "source_url": "https://huggingface.co/lackonendes/PAW-27B-X3-GGUF",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null,
   "venue": "HF"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 262144,
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hf3-qwen3827b-pawx3-tg128",
   "issue": null,
   "model": "Qwen3.8-27B",
   "notes": "config=tg128 | Decode at tg128 (generation, chat scale).",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "PAW 3.5 bpw",
   "quote": "Qwen/Qwen3.8-27B; one RTX 3090; TG128 (generation, chat scale) 33.5",
   "retrieved": "2026-10-03",
   "source_name": "HF model card: lackonendes PAW-27B-X3 GGUF",
   "source_url": "https://huggingface.co/lackonendes/PAW-27B-X3-GGUF",
   "tg_tokens": 128,
   "tps": "33.5",
   "ttft_s": null,
   "venue": "HF"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 211000,
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hf3-qwen3827b-pawx3-tg211k",
   "issue": null,
   "model": "Qwen3.8-27B",
   "notes": "config=tg211k | Decode at 211k context depth (context-fill decay).",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "PAW 3.5 bpw",
   "quote": "Qwen/Qwen3.8-27B; one RTX 3090; TG at 211k depth 19.9",
   "retrieved": "2026-10-03",
   "source_name": "HF model card: lackonendes PAW-27B-X3 GGUF",
   "source_url": "https://huggingface.co/lackonendes/PAW-27B-X3-GGUF",
   "tg_tokens": null,
   "tps": "19.9",
   "ttft_s": null,
   "venue": "HF"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 8192,
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hf3-qwen3827b-pawx3-spec8k",
   "issue": null,
   "model": "Qwen3.8-27B",
   "notes": "config=spec8k | Verified spec decode (3.5bit + DFlash2 drafter) at 8k code context, output hash-identical.",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "PAW 3.5 bpw",
   "quote": "Qwen/Qwen3.8-27B; one RTX 3090; speculative decode, 8k code context 100.45",
   "retrieved": "2026-10-03",
   "source_name": "HF model card: lackonendes PAW-27B-X3 GGUF",
   "source_url": "https://huggingface.co/lackonendes/PAW-27B-X3-GGUF",
   "tg_tokens": null,
   "tps": "100.45",
   "ttft_s": null,
   "venue": "HF"
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "flags": [],
   "hardware": "10\u00d7 MI100",
   "id": "hf3-glm53flash-mi100-q4kxl",
   "issue": null,
   "model": "GLM-5.3-Flash",
   "notes": "config=q4kxl | llama-bench -p 2048 -n 128, no spec decode, -fa off, fork of llama.cpp (glm5next arch). 10x MI100 (gfx908); ROCm (gfx908 default, card does not name the GPU backend explicitly).",
   "pp_tokens": 2048,
   "pp_tps": "780",
   "provenance": "community",
   "quant": "Q4_K_XL",
   "quote": "GLM-5.3-Flash; 10 \u00d7 MI100, gfx908; Q4_K_XL 780 t/s prefill / 33.8 t/s decode",
   "retrieved": "2026-10-03",
   "source_name": "HF model card: SixVolts GLM-5.3-Flash-ewaste-edition GGUF",
   "source_url": "https://huggingface.co/SixVolts/GLM-5.3-Flash-ewaste-edition-GGUF",
   "tg_tokens": 128,
   "tps": "33.8",
   "ttft_s": null,
   "venue": "HF"
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "flags": [],
   "hardware": "10\u00d7 MI100",
   "id": "hf3-glm53flash-mi100-q3km",
   "issue": null,
   "model": "GLM-5.3-Flash",
   "notes": "config=q3km | llama-bench -p 2048 -n 128, no spec decode, -fa off, fork of llama.cpp (glm5next arch). 10x MI100 (gfx908); ROCm. Card: 3-bit not faster than 4-bit here (K-quant dequant cost on this GPU class, why the e-waste editions drop i-quants).",
   "pp_tokens": 2048,
   "pp_tps": "730",
   "provenance": "community",
   "quant": "Q3_K_M",
   "quote": "GLM-5.3-Flash; 10 \u00d7 MI100, gfx908; Q3_K_M 730 / 32.2",
   "retrieved": "2026-10-03",
   "source_name": "HF model card: SixVolts GLM-5.3-Flash-ewaste-edition GGUF",
   "source_url": "https://huggingface.co/SixVolts/GLM-5.3-Flash-ewaste-edition-GGUF",
   "tg_tokens": 128,
   "tps": "32.2",
   "ttft_s": null,
   "venue": "HF"
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "4\u00d7 RTX PRO 6000 Blackwell",
   "id": "vllm-59846-sm120-v4rec-c1",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "4x RTX PRO 6000 Blackwell (SM120), TP4, DeepSeek-V4.1-Flash (FP8 model); KV-cache dtype A/B: V4 record before this PR (fp8) vs nvfp4_ds_mla (new on SM120); aggregate output tok/s, random 8k in / 1k out, c=1; V4 record (before); config=v4rec-c1",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "DeepSeek-V4.1-Flash; 4\u00d7 RTX PRO 6000 Blackwell (SM120); | 1 | 105.7 | 105.6 | | 106.9 | 104.7 |",
   "retrieved": "2026-10-03",
   "source_name": "vLLM PR #59846",
   "source_url": "https://github.com/vllm-project/vllm/pull/59846",
   "tg_tokens": null,
   "tps": "105.7",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "4\u00d7 RTX PRO 6000 Blackwell",
   "id": "vllm-59846-sm120-nvfp4-c1",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "4x RTX PRO 6000 Blackwell (SM120), TP4, DeepSeek-V4.1-Flash (FP8 model); KV-cache dtype A/B: V4 record before this PR (fp8) vs nvfp4_ds_mla (new on SM120); aggregate output tok/s, random 8k in / 1k out, c=1; nvfp4_ds_mla; config=nvfp4-c1",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "DeepSeek-V4.1-Flash; 4\u00d7 RTX PRO 6000 Blackwell (SM120); | 1 | 105.7 | 105.6 | | 106.9 | 104.7 |",
   "retrieved": "2026-10-03",
   "source_name": "vLLM PR #59846",
   "source_url": "https://github.com/vllm-project/vllm/pull/59846",
   "tg_tokens": null,
   "tps": "104.7",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 2,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "4\u00d7 RTX PRO 6000 Blackwell",
   "id": "vllm-59846-sm120-v4rec-c2",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "4x RTX PRO 6000 Blackwell (SM120), TP4, DeepSeek-V4.1-Flash (FP8 model); KV-cache dtype A/B: V4 record before this PR (fp8) vs nvfp4_ds_mla (new on SM120); aggregate output tok/s, random 8k in / 1k out, c=2; V4 record (before); config=v4rec-c2",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "DeepSeek-V4.1-Flash; 4\u00d7 RTX PRO 6000 Blackwell (SM120); | 2 | 180.8 | 178.5 | | 183.3 | 179.8 |",
   "retrieved": "2026-10-03",
   "source_name": "vLLM PR #59846",
   "source_url": "https://github.com/vllm-project/vllm/pull/59846",
   "tg_tokens": null,
   "tps": "180.8",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 2,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "4\u00d7 RTX PRO 6000 Blackwell",
   "id": "vllm-59846-sm120-nvfp4-c2",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "4x RTX PRO 6000 Blackwell (SM120), TP4, DeepSeek-V4.1-Flash (FP8 model); KV-cache dtype A/B: V4 record before this PR (fp8) vs nvfp4_ds_mla (new on SM120); aggregate output tok/s, random 8k in / 1k out, c=2; nvfp4_ds_mla; config=nvfp4-c2",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "DeepSeek-V4.1-Flash; 4\u00d7 RTX PRO 6000 Blackwell (SM120); | 2 | 180.8 | 178.5 | | 183.3 | 179.8 |",
   "retrieved": "2026-10-03",
   "source_name": "vLLM PR #59846",
   "source_url": "https://github.com/vllm-project/vllm/pull/59846",
   "tg_tokens": null,
   "tps": "179.8",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "4\u00d7 RTX PRO 6000 Blackwell",
   "id": "vllm-59846-sm120-v4rec-c4",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "4x RTX PRO 6000 Blackwell (SM120), TP4, DeepSeek-V4.1-Flash (FP8 model); KV-cache dtype A/B: V4 record before this PR (fp8) vs nvfp4_ds_mla (new on SM120); aggregate output tok/s, random 8k in / 1k out, c=4; V4 record (before); config=v4rec-c4",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "DeepSeek-V4.1-Flash; 4\u00d7 RTX PRO 6000 Blackwell (SM120); | 4 | 314.9 | 315.7 | | 313.4 | 314.3 |",
   "retrieved": "2026-10-03",
   "source_name": "vLLM PR #59846",
   "source_url": "https://github.com/vllm-project/vllm/pull/59846",
   "tg_tokens": null,
   "tps": "314.9",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "4\u00d7 RTX PRO 6000 Blackwell",
   "id": "vllm-59846-sm120-nvfp4-c4",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "4x RTX PRO 6000 Blackwell (SM120), TP4, DeepSeek-V4.1-Flash (FP8 model); KV-cache dtype A/B: V4 record before this PR (fp8) vs nvfp4_ds_mla (new on SM120); aggregate output tok/s, random 8k in / 1k out, c=4; nvfp4_ds_mla; config=nvfp4-c4",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "DeepSeek-V4.1-Flash; 4\u00d7 RTX PRO 6000 Blackwell (SM120); | 4 | 314.9 | 315.7 | | 313.4 | 314.3 |",
   "retrieved": "2026-10-03",
   "source_name": "vLLM PR #59846",
   "source_url": "https://github.com/vllm-project/vllm/pull/59846",
   "tg_tokens": null,
   "tps": "314.3",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "4\u00d7 RTX PRO 6000 Blackwell",
   "id": "vllm-59846-sm120-v4rec-c8",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "4x RTX PRO 6000 Blackwell (SM120), TP4, DeepSeek-V4.1-Flash (FP8 model); KV-cache dtype A/B: V4 record before this PR (fp8) vs nvfp4_ds_mla (new on SM120); aggregate output tok/s, random 8k in / 1k out, c=8; V4 record (before); config=v4rec-c8",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "DeepSeek-V4.1-Flash; 4\u00d7 RTX PRO 6000 Blackwell (SM120); | 8 | 502.4 | 504.6 | | 505.2 | 502.0 |",
   "retrieved": "2026-10-03",
   "source_name": "vLLM PR #59846",
   "source_url": "https://github.com/vllm-project/vllm/pull/59846",
   "tg_tokens": null,
   "tps": "502.4",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "4\u00d7 RTX PRO 6000 Blackwell",
   "id": "vllm-59846-sm120-nvfp4-c8",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "4x RTX PRO 6000 Blackwell (SM120), TP4, DeepSeek-V4.1-Flash (FP8 model); KV-cache dtype A/B: V4 record before this PR (fp8) vs nvfp4_ds_mla (new on SM120); aggregate output tok/s, random 8k in / 1k out, c=8; nvfp4_ds_mla; config=nvfp4-c8",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "DeepSeek-V4.1-Flash; 4\u00d7 RTX PRO 6000 Blackwell (SM120); | 8 | 502.4 | 504.6 | | 505.2 | 502.0 |",
   "retrieved": "2026-10-03",
   "source_name": "vLLM PR #59846",
   "source_url": "https://github.com/vllm-project/vllm/pull/59846",
   "tg_tokens": null,
   "tps": "502.0",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "4\u00d7 RTX PRO 6000 Blackwell",
   "id": "vllm-59846-sm120-v4rec-c16",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "4x RTX PRO 6000 Blackwell (SM120), TP4, DeepSeek-V4.1-Flash (FP8 model); KV-cache dtype A/B: V4 record before this PR (fp8) vs nvfp4_ds_mla (new on SM120); aggregate output tok/s, random 8k in / 1k out, c=16; V4 record (before); config=v4rec-c16",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "DeepSeek-V4.1-Flash; 4\u00d7 RTX PRO 6000 Blackwell (SM120); | 16 | 732.1 | 734.2 | | 734.3 | 732.9 |",
   "retrieved": "2026-10-03",
   "source_name": "vLLM PR #59846",
   "source_url": "https://github.com/vllm-project/vllm/pull/59846",
   "tg_tokens": null,
   "tps": "732.1",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "4\u00d7 RTX PRO 6000 Blackwell",
   "id": "vllm-59846-sm120-nvfp4-c16",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "4x RTX PRO 6000 Blackwell (SM120), TP4, DeepSeek-V4.1-Flash (FP8 model); KV-cache dtype A/B: V4 record before this PR (fp8) vs nvfp4_ds_mla (new on SM120); aggregate output tok/s, random 8k in / 1k out, c=16; nvfp4_ds_mla; config=nvfp4-c16",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "DeepSeek-V4.1-Flash; 4\u00d7 RTX PRO 6000 Blackwell (SM120); | 16 | 732.1 | 734.2 | | 734.3 | 732.9 |",
   "retrieved": "2026-10-03",
   "source_name": "vLLM PR #59846",
   "source_url": "https://github.com/vllm-project/vllm/pull/59846",
   "tg_tokens": null,
   "tps": "732.9",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 32,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "4\u00d7 RTX PRO 6000 Blackwell",
   "id": "vllm-59846-sm120-v4rec-c32",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "4x RTX PRO 6000 Blackwell (SM120), TP4, DeepSeek-V4.1-Flash (FP8 model); KV-cache dtype A/B: V4 record before this PR (fp8) vs nvfp4_ds_mla (new on SM120); aggregate output tok/s, random 8k in / 1k out, c=32; V4 record (before); config=v4rec-c32",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "DeepSeek-V4.1-Flash; 4\u00d7 RTX PRO 6000 Blackwell (SM120); | 32 | 857.2 | 856.4 | | 859.0 | 1003.1 |",
   "retrieved": "2026-10-03",
   "source_name": "vLLM PR #59846",
   "source_url": "https://github.com/vllm-project/vllm/pull/59846",
   "tg_tokens": null,
   "tps": "857.2",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 32,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "4\u00d7 RTX PRO 6000 Blackwell",
   "id": "vllm-59846-sm120-nvfp4-c32",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "4x RTX PRO 6000 Blackwell (SM120), TP4, DeepSeek-V4.1-Flash (FP8 model); KV-cache dtype A/B: V4 record before this PR (fp8) vs nvfp4_ds_mla (new on SM120); aggregate output tok/s, random 8k in / 1k out, c=32; nvfp4_ds_mla; config=nvfp4-c32",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "DeepSeek-V4.1-Flash; 4\u00d7 RTX PRO 6000 Blackwell (SM120); | 32 | 857.2 | 856.4 | | 859.0 | 1003.1 |",
   "retrieved": "2026-10-03",
   "source_name": "vLLM PR #59846",
   "source_url": "https://github.com/vllm-project/vllm/pull/59846",
   "tg_tokens": null,
   "tps": "1003.1",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 64,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "4\u00d7 RTX PRO 6000 Blackwell",
   "id": "vllm-59846-sm120-v4rec-c64",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "4x RTX PRO 6000 Blackwell (SM120), TP4, DeepSeek-V4.1-Flash (FP8 model); KV-cache dtype A/B: V4 record before this PR (fp8) vs nvfp4_ds_mla (new on SM120); aggregate output tok/s, random 8k in / 1k out, c=64; V4 record (before); config=v4rec-c64",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "DeepSeek-V4.1-Flash; 4\u00d7 RTX PRO 6000 Blackwell (SM120); | 64 | 1039.7 | 1041.2 | | 1042.2 | 1036.4 |",
   "retrieved": "2026-10-03",
   "source_name": "vLLM PR #59846",
   "source_url": "https://github.com/vllm-project/vllm/pull/59846",
   "tg_tokens": null,
   "tps": "1039.7",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 64,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "4\u00d7 RTX PRO 6000 Blackwell",
   "id": "vllm-59846-sm120-nvfp4-c64",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "4x RTX PRO 6000 Blackwell (SM120), TP4, DeepSeek-V4.1-Flash (FP8 model); KV-cache dtype A/B: V4 record before this PR (fp8) vs nvfp4_ds_mla (new on SM120); aggregate output tok/s, random 8k in / 1k out, c=64; nvfp4_ds_mla; config=nvfp4-c64",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "DeepSeek-V4.1-Flash; 4\u00d7 RTX PRO 6000 Blackwell (SM120); | 64 | 1039.7 | 1041.2 | | 1042.2 | 1036.4 |",
   "retrieved": "2026-10-03",
   "source_name": "vLLM PR #59846",
   "source_url": "https://github.com/vllm-project/vllm/pull/59846",
   "tg_tokens": null,
   "tps": "1036.4",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 3,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "2\u00d7 20 GB",
   "id": "davidmg-qwen38flash-gpu-nospec",
   "model": "Qwen3.8-Flash-Next",
   "notes": "HF card field note: MTP draft-mtp speculative decoding (--spec-type draft-mtp) on the Qwen3.8-Flash-Next 176B MoE pruned 512 -> 256 experts (GSQ-RCO), main model IQ1_M + separate Q8_0 MTP head; GPU 2\u00d7 20 GB (CUDA, tensor-split 45,55), 3-slot llama-server, KV q5_1, temperature 0, thinking off, 400-token runs, one run per cell, author: 'treat as indications, not benchmarks'; code vs German prose (54k each); no-MTP baseline, 61.5 tok/s for code and prose alike; context not explicitly stated (listed before the per-context table); config=nospec",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "IQ1_M main / Q8_0 head",
   "quote": "Qwen3.8-Flash-Next \u2014 separate MTP head on the expert-pruned (256-expert) GSQ-RCO model; GPU (2\u00d7 20 GB, 3 slots, KV q5_1, 3 runs of 400 tokens each; Baseline without a head: **61.5 tok/s** for code and prose alike",
   "retrieved": "2026-10-03",
   "source_name": "HF model card: Davidmg0815/Qwen3.8-Flash-Next-MTP-256exp-GGUF",
   "source_url": "https://huggingface.co/Davidmg0815/Qwen3.8-Flash-Next-MTP-256exp-GGUF",
   "tg_tokens": null,
   "tps": "61.5",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 3,
   "ctx": 65536,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "2\u00d7 20 GB",
   "id": "davidmg-qwen38flash-gpu-nmax1-code",
   "model": "Qwen3.8-Flash-Next",
   "notes": "HF card field note: MTP draft-mtp speculative decoding (--spec-type draft-mtp) on the Qwen3.8-Flash-Next 176B MoE pruned 512 -> 256 experts (GSQ-RCO), main model IQ1_M + separate Q8_0 MTP head; GPU 2\u00d7 20 GB (CUDA, tensor-split 45,55), 3-slot llama-server, KV q5_1, temperature 0, thinking off, 400-token runs, one run per cell, author: 'treat as indications, not benchmarks'; code vs German prose (54k each); n-max 1, 256 head, 3\u00d764K, code: 55.4 tok/s (prose 49.4); config=nmax1-code",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "IQ1_M main / Q8_0 head",
   "quote": "Qwen3.8-Flash-Next \u2014 separate MTP head on the expert-pruned (256-expert) GSQ-RCO model; GPU (2\u00d7 20 GB, 3 slots, KV q5_1, 3 runs of 400 tokens each; 256 head, 3\u00d764K, n-max 1 55.4 49.4",
   "retrieved": "2026-10-03",
   "source_name": "HF model card: Davidmg0815/Qwen3.8-Flash-Next-MTP-256exp-GGUF",
   "source_url": "https://huggingface.co/Davidmg0815/Qwen3.8-Flash-Next-MTP-256exp-GGUF",
   "tg_tokens": null,
   "tps": "55.4",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 3,
   "ctx": 65536,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "2\u00d7 20 GB",
   "id": "davidmg-qwen38flash-gpu-nmax1-prose",
   "model": "Qwen3.8-Flash-Next",
   "notes": "HF card field note: MTP draft-mtp speculative decoding (--spec-type draft-mtp) on the Qwen3.8-Flash-Next 176B MoE pruned 512 -> 256 experts (GSQ-RCO), main model IQ1_M + separate Q8_0 MTP head; GPU 2\u00d7 20 GB (CUDA, tensor-split 45,55), 3-slot llama-server, KV q5_1, temperature 0, thinking off, 400-token runs, one run per cell, author: 'treat as indications, not benchmarks'; code vs German prose (54k each); n-max 1, 256 head, 3\u00d764K, German prose: 49.4 tok/s (code 55.4); config=nmax1-prose",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "IQ1_M main / Q8_0 head",
   "quote": "Qwen3.8-Flash-Next \u2014 separate MTP head on the expert-pruned (256-expert) GSQ-RCO model; GPU (2\u00d7 20 GB, 3 slots, KV q5_1, 3 runs of 400 tokens each; 256 head, 3\u00d764K, n-max 1 55.4 49.4",
   "retrieved": "2026-10-03",
   "source_name": "HF model card: Davidmg0815/Qwen3.8-Flash-Next-MTP-256exp-GGUF",
   "source_url": "https://huggingface.co/Davidmg0815/Qwen3.8-Flash-Next-MTP-256exp-GGUF",
   "tg_tokens": null,
   "tps": "49.4",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 3,
   "ctx": 65536,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "2\u00d7 20 GB",
   "id": "davidmg-qwen38flash-gpu-nmax2-code",
   "model": "Qwen3.8-Flash-Next",
   "notes": "HF card field note: MTP draft-mtp speculative decoding (--spec-type draft-mtp) on the Qwen3.8-Flash-Next 176B MoE pruned 512 -> 256 experts (GSQ-RCO), main model IQ1_M + separate Q8_0 MTP head; GPU 2\u00d7 20 GB (CUDA, tensor-split 45,55), 3-slot llama-server, KV q5_1, temperature 0, thinking off, 400-token runs, one run per cell, author: 'treat as indications, not benchmarks'; code vs German prose (54k each); n-max 2, 256 head, 3\u00d764K, code: 67.1 tok/s (prose 50.3); config=nmax2-code",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "IQ1_M main / Q8_0 head",
   "quote": "Qwen3.8-Flash-Next \u2014 separate MTP head on the expert-pruned (256-expert) GSQ-RCO model; GPU (2\u00d7 20 GB, 3 slots, KV q5_1, 3 runs of 400 tokens each; 256 head, 3\u00d764K, n-max 2 67.1 50.3",
   "retrieved": "2026-10-03",
   "source_name": "HF model card: Davidmg0815/Qwen3.8-Flash-Next-MTP-256exp-GGUF",
   "source_url": "https://huggingface.co/Davidmg0815/Qwen3.8-Flash-Next-MTP-256exp-GGUF",
   "tg_tokens": null,
   "tps": "67.1",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 3,
   "ctx": 65536,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "2\u00d7 20 GB",
   "id": "davidmg-qwen38flash-gpu-nmax2-prose",
   "model": "Qwen3.8-Flash-Next",
   "notes": "HF card field note: MTP draft-mtp speculative decoding (--spec-type draft-mtp) on the Qwen3.8-Flash-Next 176B MoE pruned 512 -> 256 experts (GSQ-RCO), main model IQ1_M + separate Q8_0 MTP head; GPU 2\u00d7 20 GB (CUDA, tensor-split 45,55), 3-slot llama-server, KV q5_1, temperature 0, thinking off, 400-token runs, one run per cell, author: 'treat as indications, not benchmarks'; code vs German prose (54k each); n-max 2, 256 head, 3\u00d764K, German prose: 50.3 tok/s (code 67.1); config=nmax2-prose",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "IQ1_M main / Q8_0 head",
   "quote": "Qwen3.8-Flash-Next \u2014 separate MTP head on the expert-pruned (256-expert) GSQ-RCO model; GPU (2\u00d7 20 GB, 3 slots, KV q5_1, 3 runs of 400 tokens each; 256 head, 3\u00d764K, n-max 2 67.1 50.3",
   "retrieved": "2026-10-03",
   "source_name": "HF model card: Davidmg0815/Qwen3.8-Flash-Next-MTP-256exp-GGUF",
   "source_url": "https://huggingface.co/Davidmg0815/Qwen3.8-Flash-Next-MTP-256exp-GGUF",
   "tg_tokens": null,
   "tps": "50.3",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 3,
   "ctx": 65536,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "2\u00d7 20 GB",
   "id": "davidmg-qwen38flash-gpu-nmax3-code",
   "model": "Qwen3.8-Flash-Next",
   "notes": "HF card field note: MTP draft-mtp speculative decoding (--spec-type draft-mtp) on the Qwen3.8-Flash-Next 176B MoE pruned 512 -> 256 experts (GSQ-RCO), main model IQ1_M + separate Q8_0 MTP head; GPU 2\u00d7 20 GB (CUDA, tensor-split 45,55), 3-slot llama-server, KV q5_1, temperature 0, thinking off, 400-token runs, one run per cell, author: 'treat as indications, not benchmarks'; code vs German prose (54k each); n-max 3, 256 head, 3\u00d764K, code: 70.5 tok/s (prose 59.6, range 45-70; author: unstable, CUDA OOM under three parallel 54k prompts); config=nmax3-code",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "IQ1_M main / Q8_0 head",
   "quote": "Qwen3.8-Flash-Next \u2014 separate MTP head on the expert-pruned (256-expert) GSQ-RCO model; GPU (2\u00d7 20 GB, 3 slots, KV q5_1, 3 runs of 400 tokens each; 256 head, 3\u00d764K, n-max 3 70.5 59.6",
   "retrieved": "2026-10-03",
   "source_name": "HF model card: Davidmg0815/Qwen3.8-Flash-Next-MTP-256exp-GGUF",
   "source_url": "https://huggingface.co/Davidmg0815/Qwen3.8-Flash-Next-MTP-256exp-GGUF",
   "tg_tokens": null,
   "tps": "70.5",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 3,
   "ctx": 65536,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "2\u00d7 20 GB",
   "id": "davidmg-qwen38flash-gpu-nmax3-prose",
   "model": "Qwen3.8-Flash-Next",
   "notes": "HF card field note: MTP draft-mtp speculative decoding (--spec-type draft-mtp) on the Qwen3.8-Flash-Next 176B MoE pruned 512 -> 256 experts (GSQ-RCO), main model IQ1_M + separate Q8_0 MTP head; GPU 2\u00d7 20 GB (CUDA, tensor-split 45,55), 3-slot llama-server, KV q5_1, temperature 0, thinking off, 400-token runs, one run per cell, author: 'treat as indications, not benchmarks'; code vs German prose (54k each); n-max 3, 256 head, 3\u00d764K, German prose: 59.6 tok/s (range 45-70, code 70.5; unstable per author); config=nmax3-prose",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "IQ1_M main / Q8_0 head",
   "quote": "Qwen3.8-Flash-Next \u2014 separate MTP head on the expert-pruned (256-expert) GSQ-RCO model; GPU (2\u00d7 20 GB, 3 slots, KV q5_1, 3 runs of 400 tokens each; 256 head, 3\u00d764K, n-max 3 70.5 59.6",
   "retrieved": "2026-10-03",
   "source_name": "HF model card: Davidmg0815/Qwen3.8-Flash-Next-MTP-256exp-GGUF",
   "source_url": "https://huggingface.co/Davidmg0815/Qwen3.8-Flash-Next-MTP-256exp-GGUF",
   "tg_tokens": null,
   "tps": "59.6",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 3,
   "ctx": 65536,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "2\u00d7 20 GB",
   "id": "davidmg-qwen38flash-gpu-nmax3pmin-code",
   "model": "Qwen3.8-Flash-Next",
   "notes": "HF card field note: MTP draft-mtp speculative decoding (--spec-type draft-mtp) on the Qwen3.8-Flash-Next 176B MoE pruned 512 -> 256 experts (GSQ-RCO), main model IQ1_M + separate Q8_0 MTP head; GPU 2\u00d7 20 GB (CUDA, tensor-split 45,55), 3-slot llama-server, KV q5_1, temperature 0, thinking off, 400-token runs, one run per cell, author: 'treat as indications, not benchmarks'; code vs German prose (54k each); n-max 3, p-min 0.6, 256 head, 3\u00d764K, code: 58.9 tok/s (prose 41.2); config=nmax3pmin-code",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "IQ1_M main / Q8_0 head",
   "quote": "Qwen3.8-Flash-Next \u2014 separate MTP head on the expert-pruned (256-expert) GSQ-RCO model; GPU (2\u00d7 20 GB, 3 slots, KV q5_1, 3 runs of 400 tokens each; 256 head, 3\u00d764K, n-max 3, p-min 0.6 58.9 41.2",
   "retrieved": "2026-10-03",
   "source_name": "HF model card: Davidmg0815/Qwen3.8-Flash-Next-MTP-256exp-GGUF",
   "source_url": "https://huggingface.co/Davidmg0815/Qwen3.8-Flash-Next-MTP-256exp-GGUF",
   "tg_tokens": null,
   "tps": "58.9",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 3,
   "ctx": 65536,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "2\u00d7 20 GB",
   "id": "davidmg-qwen38flash-gpu-nmax3pmin-prose",
   "model": "Qwen3.8-Flash-Next",
   "notes": "HF card field note: MTP draft-mtp speculative decoding (--spec-type draft-mtp) on the Qwen3.8-Flash-Next 176B MoE pruned 512 -> 256 experts (GSQ-RCO), main model IQ1_M + separate Q8_0 MTP head; GPU 2\u00d7 20 GB (CUDA, tensor-split 45,55), 3-slot llama-server, KV q5_1, temperature 0, thinking off, 400-token runs, one run per cell, author: 'treat as indications, not benchmarks'; code vs German prose (54k each); n-max 3, p-min 0.6, 256 head, 3\u00d764K, German prose: 41.2 tok/s (code 58.9); config=nmax3pmin-prose",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "IQ1_M main / Q8_0 head",
   "quote": "Qwen3.8-Flash-Next \u2014 separate MTP head on the expert-pruned (256-expert) GSQ-RCO model; GPU (2\u00d7 20 GB, 3 slots, KV q5_1, 3 runs of 400 tokens each; 256 head, 3\u00d764K, n-max 3, p-min 0.6 58.9 41.2",
   "retrieved": "2026-10-03",
   "source_name": "HF model card: Davidmg0815/Qwen3.8-Flash-Next-MTP-256exp-GGUF",
   "source_url": "https://huggingface.co/Davidmg0815/Qwen3.8-Flash-Next-MTP-256exp-GGUF",
   "tg_tokens": null,
   "tps": "41.2",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "hf3-swift15-r9700-a-plain",
   "model": "Swift 1.5",
   "notes": "HF card field note: MXFP4 GGUF of the Swift-1.5 fine-tune of Qwen3.8 27b (MTP layer kept for self-spec decode); base MXFP4 with 3 output-head variants (A Q6_K / B Q8_0 / C Q4_K); llama.cpp b11214 ROCm, R9700 32 GB USB4 eGPU, single request, short prompts, 2k prefill; variant A (Q6_K head), plain decode 33.3 / 2k prefill 1215 tok/s; config=a-plain",
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": "1215",
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Swift-1.5-Qwen3.8-27b MXFP4 GGUF; Measured performance (WHIRL 0.2.0: Radeon 8060S and AMD Radeon AI PRO R9700; llama.cpp-only rows: R9700, 32 GB); Plain decode, tok/s (mean of 4 prompts) | 33.3 | 32.8 | 33.8",
   "retrieved": "2026-10-03",
   "source_name": "HF model card: tsaipifong Swift-1.5-Qwen3.8-27b-MXFP4",
   "source_url": "https://huggingface.co/tsaipifong/Swift-1.5-Qwen3.8-27b-MXFP4-GGUF",
   "tg_tokens": null,
   "tps": "33.3",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "hf3-swift15-r9700-a-mtp",
   "model": "Swift 1.5",
   "notes": "HF card field note: MXFP4 GGUF of the Swift-1.5 fine-tune of Qwen3.8 27b (MTP layer kept for self-spec decode); base MXFP4 with 3 output-head variants (A Q6_K / B Q8_0 / C Q4_K); llama.cpp b11214 ROCm, R9700 32 GB USB4 eGPU, single request, short prompts, 2k prefill; variant A (Q6_K head), MTP n-max 3 decode 58.1 tok/s (acceptance 73.6%); config=a-mtp",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Swift-1.5-Qwen3.8-27b MXFP4 GGUF; Measured performance (WHIRL 0.2.0: Radeon 8060S and AMD Radeon AI PRO R9700; llama.cpp-only rows: R9700, 32 GB); MTP decode, tok/s (--spec-draft-n-max 3) | 58.1 | 55.9 | 60.7",
   "retrieved": "2026-10-03",
   "source_name": "HF model card: tsaipifong Swift-1.5-Qwen3.8-27b-MXFP4",
   "source_url": "https://huggingface.co/tsaipifong/Swift-1.5-Qwen3.8-27b-MXFP4-GGUF",
   "tg_tokens": null,
   "tps": "58.1",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "hf3-swift15-r9700-b-plain",
   "model": "Swift 1.5",
   "notes": "HF card field note: MXFP4 GGUF of the Swift-1.5 fine-tune of Qwen3.8 27b (MTP layer kept for self-spec decode); base MXFP4 with 3 output-head variants (A Q6_K / B Q8_0 / C Q4_K); llama.cpp b11214 ROCm, R9700 32 GB USB4 eGPU, single request, short prompts, 2k prefill; variant B (Q8_0 head), plain decode 32.8 / 2k prefill 1212 tok/s; config=b-plain",
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": "1212",
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Swift-1.5-Qwen3.8-27b MXFP4 GGUF; Measured performance (WHIRL 0.2.0: Radeon 8060S and AMD Radeon AI PRO R9700; llama.cpp-only rows: R9700, 32 GB); Plain decode, tok/s (mean of 4 prompts) | 33.3 | 32.8 | 33.8",
   "retrieved": "2026-10-03",
   "source_name": "HF model card: tsaipifong Swift-1.5-Qwen3.8-27b-MXFP4",
   "source_url": "https://huggingface.co/tsaipifong/Swift-1.5-Qwen3.8-27b-MXFP4-GGUF",
   "tg_tokens": null,
   "tps": "32.8",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "hf3-swift15-r9700-b-mtp",
   "model": "Swift 1.5",
   "notes": "HF card field note: MXFP4 GGUF of the Swift-1.5 fine-tune of Qwen3.8 27b (MTP layer kept for self-spec decode); base MXFP4 with 3 output-head variants (A Q6_K / B Q8_0 / C Q4_K); llama.cpp b11214 ROCm, R9700 32 GB USB4 eGPU, single request, short prompts, 2k prefill; variant B (Q8_0 head), MTP n-max 3 decode 55.9 tok/s (acceptance 72.2%); config=b-mtp",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Swift-1.5-Qwen3.8-27b MXFP4 GGUF; Measured performance (WHIRL 0.2.0: Radeon 8060S and AMD Radeon AI PRO R9700; llama.cpp-only rows: R9700, 32 GB); MTP decode, tok/s (--spec-draft-n-max 3) | 58.1 | 55.9 | 60.7",
   "retrieved": "2026-10-03",
   "source_name": "HF model card: tsaipifong Swift-1.5-Qwen3.8-27b-MXFP4",
   "source_url": "https://huggingface.co/tsaipifong/Swift-1.5-Qwen3.8-27b-MXFP4-GGUF",
   "tg_tokens": null,
   "tps": "55.9",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "hf3-swift15-r9700-c-plain",
   "model": "Swift 1.5",
   "notes": "HF card field note: MXFP4 GGUF of the Swift-1.5 fine-tune of Qwen3.8 27b (MTP layer kept for self-spec decode); base MXFP4 with 3 output-head variants (A Q6_K / B Q8_0 / C Q4_K); llama.cpp b11214 ROCm, R9700 32 GB USB4 eGPU, single request, short prompts, 2k prefill; variant C (Q4_K head), plain decode 33.8 / 2k prefill 1206 tok/s; config=c-plain",
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": "1206",
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Swift-1.5-Qwen3.8-27b MXFP4 GGUF; Measured performance (WHIRL 0.2.0: Radeon 8060S and AMD Radeon AI PRO R9700; llama.cpp-only rows: R9700, 32 GB); Plain decode, tok/s (mean of 4 prompts) | 33.3 | 32.8 | 33.8",
   "retrieved": "2026-10-03",
   "source_name": "HF model card: tsaipifong Swift-1.5-Qwen3.8-27b-MXFP4",
   "source_url": "https://huggingface.co/tsaipifong/Swift-1.5-Qwen3.8-27b-MXFP4-GGUF",
   "tg_tokens": null,
   "tps": "33.8",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "hf3-swift15-r9700-c-mtp",
   "model": "Swift 1.5",
   "notes": "HF card field note: MXFP4 GGUF of the Swift-1.5 fine-tune of Qwen3.8 27b (MTP layer kept for self-spec decode); base MXFP4 with 3 output-head variants (A Q6_K / B Q8_0 / C Q4_K); llama.cpp b11214 ROCm, R9700 32 GB USB4 eGPU, single request, short prompts, 2k prefill; variant C (Q4_K head), MTP n-max 3 decode 60.7 tok/s (acceptance 75.0%); config=c-mtp",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Swift-1.5-Qwen3.8-27b MXFP4 GGUF; Measured performance (WHIRL 0.2.0: Radeon 8060S and AMD Radeon AI PRO R9700; llama.cpp-only rows: R9700, 32 GB); MTP decode, tok/s (--spec-draft-n-max 3) | 58.1 | 55.9 | 60.7",
   "retrieved": "2026-10-03",
   "source_name": "HF model card: tsaipifong Swift-1.5-Qwen3.8-27b-MXFP4",
   "source_url": "https://huggingface.co/tsaipifong/Swift-1.5-Qwen3.8-27b-MXFP4-GGUF",
   "tg_tokens": null,
   "tps": "60.7",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc-29887-4090-master",
   "model": "Qwen3.8-Flash-Next",
   "notes": "llama.cpp PR #29887, MoE expert GPU cache (port of qvac-fabric); Qwen3.8-Flash-Next Q4_0 (93.7 GiB, 65.4 GiB experts); SPEED-Bench qualitative (--osl 512, temp 0, no MTP); llama-server -fa on -c 32768 -b 2048 -ub 2048 -t 16, CPU on the GPU NUMA node (EPYC 7742, 16 cores), PCIe 4.0 x16; baseline is master with the default --fit; master --fit baseline (no cache), 16.5+0.0 GB whole layers, tg 25.0; pp2048 1388; config=master",
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Qwen3.8-Flash-Next Q4_0; RTX 4090 | master --fit | 16.5 + 0.0 GB | 25.0 | 1.00x",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp PR #29887 (MoE expert GPU cache)",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29887",
   "tg_tokens": 512,
   "tps": "25.0",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc-29887-4090-fitcache",
   "model": "Qwen3.8-Flash-Next",
   "notes": "llama.cpp PR #29887, MoE expert GPU cache (port of qvac-fabric); Qwen3.8-Flash-Next Q4_0 (93.7 GiB, 65.4 GiB experts); SPEED-Bench qualitative (--osl 512, temp 0, no MTP); llama-server -fa on -c 32768 -b 2048 -ub 2048 -t 16, CPU on the GPU NUMA node (EPYC 7742, 16 cores), PCIe 4.0 x16; baseline is master with the default --fit; --fit --moe-cache-mib 6544, 6.4 GB cache (10.4+6.4 GB), hit rate 72%, tg 39.4 (1.57x); pp2048 1292; config=fitcache",
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Qwen3.8-Flash-Next Q4_0; RTX 4090 | --fit --moe-cache-mib 6544 | 10.4 + 6.4 GB | 39.4 | 1.57x",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp PR #29887 (MoE expert GPU cache)",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29887",
   "tg_tokens": 512,
   "tps": "39.4",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc-29887-4090-cmoecache",
   "model": "Qwen3.8-Flash-Next",
   "notes": "llama.cpp PR #29887, MoE expert GPU cache (port of qvac-fabric); Qwen3.8-Flash-Next Q4_0 (93.7 GiB, 65.4 GiB experts); SPEED-Bench qualitative (--osl 512, temp 0, no MTP); llama-server -fa on -c 32768 -b 2048 -ub 2048 -t 16, CPU on the GPU NUMA node (EPYC 7742, 16 cores), PCIe 4.0 x16; baseline is master with the default --fit; -cmoe full cache (11000 mib, 10.7 GB, 0 whole layers), hit rate 77%, tg 40.7 (1.62x); pp2048 1080; config=cmoecache",
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Qwen3.8-Flash-Next Q4_0; RTX 4090 | -cmoe --moe-cache-mib 11000 | 0.0 + 10.7 GB | 40.7 | 1.62x",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp PR #29887 (MoE expert GPU cache)",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29887",
   "tg_tokens": 512,
   "tps": "40.7",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-29887-5090-master",
   "model": "Qwen3.8-Flash-Next",
   "notes": "llama.cpp PR #29887, MoE expert GPU cache (port of qvac-fabric); Qwen3.8-Flash-Next Q4_0 (93.7 GiB, 65.4 GiB experts); SPEED-Bench qualitative (--osl 512, temp 0, no MTP); llama-server -fa on -c 32768 -b 2048 -ub 2048 -t 16, CPU on the GPU NUMA node (EPYC 7742, 16 cores), PCIe 4.0 x16; baseline is master with the default --fit; master --fit baseline (no cache), 24.5+0.0 GB whole layers, tg 30.8; config=master",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Qwen3.8-Flash-Next Q4_0; RTX 5090 | master --fit | 24.5 + 0.0 GB | 30.8 | 1.00x",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp PR #29887 (MoE expert GPU cache)",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29887",
   "tg_tokens": 512,
   "tps": "30.8",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-29887-5090-fitcache",
   "model": "Qwen3.8-Flash-Next",
   "notes": "llama.cpp PR #29887, MoE expert GPU cache (port of qvac-fabric); Qwen3.8-Flash-Next Q4_0 (93.7 GiB, 65.4 GiB experts); SPEED-Bench qualitative (--osl 512, temp 0, no MTP); llama-server -fa on -c 32768 -b 2048 -ub 2048 -t 16, CPU on the GPU NUMA node (EPYC 7742, 16 cores), PCIe 4.0 x16; baseline is master with the default --fit; --fit --moe-cache-mib 6544, 6.4 GB cache (17.9+6.4 GB), hit rate 77%, tg 54.5 (1.77x); config=fitcache",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Qwen3.8-Flash-Next Q4_0; RTX 5090 | --fit --moe-cache-mib 6544 | 17.9 + 6.4 GB | 54.5 | 1.77x",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp PR #29887 (MoE expert GPU cache)",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29887",
   "tg_tokens": 512,
   "tps": "54.5",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-29887-5090-cmoecache",
   "model": "Qwen3.8-Flash-Next",
   "notes": "llama.cpp PR #29887, MoE expert GPU cache (port of qvac-fabric); Qwen3.8-Flash-Next Q4_0 (93.7 GiB, 65.4 GiB experts); SPEED-Bench qualitative (--osl 512, temp 0, no MTP); llama-server -fa on -c 32768 -b 2048 -ub 2048 -t 16, CPU on the GPU NUMA node (EPYC 7742, 16 cores), PCIe 4.0 x16; baseline is master with the default --fit; -cmoe full cache (19000 mib, 18.6 GB, 0 whole layers), hit rate 89%, tg 67.8 (2.20x); config=cmoecache",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Qwen3.8-Flash-Next Q4_0; RTX 5090 | -cmoe --moe-cache-mib 19000 | 0.0 + 18.6 GB | 67.8 | 2.20x",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp PR #29887 (MoE expert GPU cache)",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29887",
   "tg_tokens": 512,
   "tps": "67.8",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "RTX PRO 6000 Blackwell Max-Q",
   "id": "deluxetiky-qwen38turbo-vllm-c1",
   "model": "Qwen3.8-27B-TURBO",
   "notes": "HF card field note: Qwen3.8-27B TURBO NVFP4-W4A16 (modelopt FP4, native MTP head n=1), vLLM 0.27.1 + GDN decode backport (PR #41966), KV fp8, 262144 max ctx, single RTX PRO 6000 Blackwell Max-Q 96 GB; concurrency-sweep aggregate tok/s (scale_and_context.py); c1 = 106.3 tok/s; config=c1",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Benchmark: Agnes-3.0-Flash (BF16) vs Qwen3.8-27B TURBO (NVFP4) on one 96 GB GPU; 1\u00d7 NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 97 887 MiB (96 GB), single GPU (TP1); Qwen3.8 NVFP4 | 106.3 | 209.8 | 274.5 | 765.6 | 581.0",
   "retrieved": "2026-10-03",
   "source_name": "HF model card: deluxetiky agnes-vs-qwen38-benchmark",
   "source_url": "https://huggingface.co/deluxetiky/agnes-vs-qwen38-benchmark",
   "tg_tokens": null,
   "tps": "106.3",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 2,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "RTX PRO 6000 Blackwell Max-Q",
   "id": "deluxetiky-qwen38turbo-vllm-c2",
   "model": "Qwen3.8-27B-TURBO",
   "notes": "HF card field note: Qwen3.8-27B TURBO NVFP4-W4A16 (modelopt FP4, native MTP head n=1), vLLM 0.27.1 + GDN decode backport (PR #41966), KV fp8, 262144 max ctx, single RTX PRO 6000 Blackwell Max-Q 96 GB; concurrency-sweep aggregate tok/s (scale_and_context.py); c2 = 209.8 tok/s; config=c2",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Benchmark: Agnes-3.0-Flash (BF16) vs Qwen3.8-27B TURBO (NVFP4) on one 96 GB GPU; 1\u00d7 NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 97 887 MiB (96 GB), single GPU (TP1); Qwen3.8 NVFP4 | 106.3 | 209.8 | 274.5 | 765.6 | 581.0",
   "retrieved": "2026-10-03",
   "source_name": "HF model card: deluxetiky agnes-vs-qwen38-benchmark",
   "source_url": "https://huggingface.co/deluxetiky/agnes-vs-qwen38-benchmark",
   "tg_tokens": null,
   "tps": "209.8",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "RTX PRO 6000 Blackwell Max-Q",
   "id": "deluxetiky-qwen38turbo-vllm-c4",
   "model": "Qwen3.8-27B-TURBO",
   "notes": "HF card field note: Qwen3.8-27B TURBO NVFP4-W4A16 (modelopt FP4, native MTP head n=1), vLLM 0.27.1 + GDN decode backport (PR #41966), KV fp8, 262144 max ctx, single RTX PRO 6000 Blackwell Max-Q 96 GB; concurrency-sweep aggregate tok/s (scale_and_context.py); c4 = 274.5 tok/s; config=c4",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Benchmark: Agnes-3.0-Flash (BF16) vs Qwen3.8-27B TURBO (NVFP4) on one 96 GB GPU; 1\u00d7 NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 97 887 MiB (96 GB), single GPU (TP1); Qwen3.8 NVFP4 | 106.3 | 209.8 | 274.5 | 765.6 | 581.0",
   "retrieved": "2026-10-03",
   "source_name": "HF model card: deluxetiky agnes-vs-qwen38-benchmark",
   "source_url": "https://huggingface.co/deluxetiky/agnes-vs-qwen38-benchmark",
   "tg_tokens": null,
   "tps": "274.5",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "RTX PRO 6000 Blackwell Max-Q",
   "id": "deluxetiky-qwen38turbo-vllm-c8",
   "model": "Qwen3.8-27B-TURBO",
   "notes": "HF card field note: Qwen3.8-27B TURBO NVFP4-W4A16 (modelopt FP4, native MTP head n=1), vLLM 0.27.1 + GDN decode backport (PR #41966), KV fp8, 262144 max ctx, single RTX PRO 6000 Blackwell Max-Q 96 GB; concurrency-sweep aggregate tok/s (scale_and_context.py); c8 = 765.6 tok/s (sweep peak); config=c8",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Benchmark: Agnes-3.0-Flash (BF16) vs Qwen3.8-27B TURBO (NVFP4) on one 96 GB GPU; 1\u00d7 NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 97 887 MiB (96 GB), single GPU (TP1); Qwen3.8 NVFP4 | 106.3 | 209.8 | 274.5 | 765.6 | 581.0",
   "retrieved": "2026-10-03",
   "source_name": "HF model card: deluxetiky agnes-vs-qwen38-benchmark",
   "source_url": "https://huggingface.co/deluxetiky/agnes-vs-qwen38-benchmark",
   "tg_tokens": null,
   "tps": "765.6",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "RTX PRO 6000 Blackwell Max-Q",
   "id": "deluxetiky-qwen38turbo-vllm-c16",
   "model": "Qwen3.8-27B-TURBO",
   "notes": "HF card field note: Qwen3.8-27B TURBO NVFP4-W4A16 (modelopt FP4, native MTP head n=1), vLLM 0.27.1 + GDN decode backport (PR #41966), KV fp8, 262144 max ctx, single RTX PRO 6000 Blackwell Max-Q 96 GB; concurrency-sweep aggregate tok/s (scale_and_context.py); c16 = 581.0 tok/s; config=c16",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Benchmark: Agnes-3.0-Flash (BF16) vs Qwen3.8-27B TURBO (NVFP4) on one 96 GB GPU; 1\u00d7 NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 97 887 MiB (96 GB), single GPU (TP1); Qwen3.8 NVFP4 | 106.3 | 209.8 | 274.5 | 765.6 | 581.0",
   "retrieved": "2026-10-03",
   "source_name": "HF model card: deluxetiky agnes-vs-qwen38-benchmark",
   "source_url": "https://huggingface.co/deluxetiky/agnes-vs-qwen38-benchmark",
   "tg_tokens": null,
   "tps": "581.0",
   "ttft_s": null
  },
  {
   "backend": "SGLang",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "RTX PRO 6000 Blackwell Max-Q",
   "id": "deluxetiky-agnes-sglang-c1",
   "model": "Agnes-3.0-Flash",
   "notes": "HF card field note: Agnes-3.0-Flash BF16 (66.2 GB), SGLang 0.5.19 + checkpoint sglang_patch, NEXTN spec-decode (steps=3 draft=4), KV fp8, 262144 max ctx, single RTX PRO 6000 Blackwell Max-Q 96 GB; concurrency-sweep aggregate tok/s (scale_and_context.py); c1 = 35.2 tok/s; config=c1",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "Benchmark: Agnes-3.0-Flash (BF16) vs Qwen3.8-27B TURBO (NVFP4) on one 96 GB GPU; 1\u00d7 NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 97 887 MiB (96 GB), single GPU (TP1); Agnes BF16 | 35.2 | 59.0 | 123.7 | 125.6 | 71.6",
   "retrieved": "2026-10-03",
   "source_name": "HF model card: deluxetiky agnes-vs-qwen38-benchmark",
   "source_url": "https://huggingface.co/deluxetiky/agnes-vs-qwen38-benchmark",
   "tg_tokens": null,
   "tps": "35.2",
   "ttft_s": null
  },
  {
   "backend": "SGLang",
   "batch": 2,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "RTX PRO 6000 Blackwell Max-Q",
   "id": "deluxetiky-agnes-sglang-c2",
   "model": "Agnes-3.0-Flash",
   "notes": "HF card field note: Agnes-3.0-Flash BF16 (66.2 GB), SGLang 0.5.19 + checkpoint sglang_patch, NEXTN spec-decode (steps=3 draft=4), KV fp8, 262144 max ctx, single RTX PRO 6000 Blackwell Max-Q 96 GB; concurrency-sweep aggregate tok/s (scale_and_context.py); c2 = 59.0 tok/s; config=c2",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "Benchmark: Agnes-3.0-Flash (BF16) vs Qwen3.8-27B TURBO (NVFP4) on one 96 GB GPU; 1\u00d7 NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 97 887 MiB (96 GB), single GPU (TP1); Agnes BF16 | 35.2 | 59.0 | 123.7 | 125.6 | 71.6",
   "retrieved": "2026-10-03",
   "source_name": "HF model card: deluxetiky agnes-vs-qwen38-benchmark",
   "source_url": "https://huggingface.co/deluxetiky/agnes-vs-qwen38-benchmark",
   "tg_tokens": null,
   "tps": "59.0",
   "ttft_s": null
  },
  {
   "backend": "SGLang",
   "batch": 4,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "RTX PRO 6000 Blackwell Max-Q",
   "id": "deluxetiky-agnes-sglang-c4",
   "model": "Agnes-3.0-Flash",
   "notes": "HF card field note: Agnes-3.0-Flash BF16 (66.2 GB), SGLang 0.5.19 + checkpoint sglang_patch, NEXTN spec-decode (steps=3 draft=4), KV fp8, 262144 max ctx, single RTX PRO 6000 Blackwell Max-Q 96 GB; concurrency-sweep aggregate tok/s (scale_and_context.py); c4 = 123.7 tok/s; config=c4",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "Benchmark: Agnes-3.0-Flash (BF16) vs Qwen3.8-27B TURBO (NVFP4) on one 96 GB GPU; 1\u00d7 NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 97 887 MiB (96 GB), single GPU (TP1); Agnes BF16 | 35.2 | 59.0 | 123.7 | 125.6 | 71.6",
   "retrieved": "2026-10-03",
   "source_name": "HF model card: deluxetiky agnes-vs-qwen38-benchmark",
   "source_url": "https://huggingface.co/deluxetiky/agnes-vs-qwen38-benchmark",
   "tg_tokens": null,
   "tps": "123.7",
   "ttft_s": null
  },
  {
   "backend": "SGLang",
   "batch": 8,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "RTX PRO 6000 Blackwell Max-Q",
   "id": "deluxetiky-agnes-sglang-c8",
   "model": "Agnes-3.0-Flash",
   "notes": "HF card field note: Agnes-3.0-Flash BF16 (66.2 GB), SGLang 0.5.19 + checkpoint sglang_patch, NEXTN spec-decode (steps=3 draft=4), KV fp8, 262144 max ctx, single RTX PRO 6000 Blackwell Max-Q 96 GB; concurrency-sweep aggregate tok/s (scale_and_context.py); c8 = 125.6 tok/s; config=c8",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "Benchmark: Agnes-3.0-Flash (BF16) vs Qwen3.8-27B TURBO (NVFP4) on one 96 GB GPU; 1\u00d7 NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 97 887 MiB (96 GB), single GPU (TP1); Agnes BF16 | 35.2 | 59.0 | 123.7 | 125.6 | 71.6",
   "retrieved": "2026-10-03",
   "source_name": "HF model card: deluxetiky agnes-vs-qwen38-benchmark",
   "source_url": "https://huggingface.co/deluxetiky/agnes-vs-qwen38-benchmark",
   "tg_tokens": null,
   "tps": "125.6",
   "ttft_s": null
  },
  {
   "backend": "SGLang",
   "batch": 16,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "RTX PRO 6000 Blackwell Max-Q",
   "id": "deluxetiky-agnes-sglang-c16",
   "model": "Agnes-3.0-Flash",
   "notes": "HF card field note: Agnes-3.0-Flash BF16 (66.2 GB), SGLang 0.5.19 + checkpoint sglang_patch, NEXTN spec-decode (steps=3 draft=4), KV fp8, 262144 max ctx, single RTX PRO 6000 Blackwell Max-Q 96 GB; concurrency-sweep aggregate tok/s (scale_and_context.py); c16 = 71.6 tok/s; config=c16",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "Benchmark: Agnes-3.0-Flash (BF16) vs Qwen3.8-27B TURBO (NVFP4) on one 96 GB GPU; 1\u00d7 NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 97 887 MiB (96 GB), single GPU (TP1); Agnes BF16 | 35.2 | 59.0 | 123.7 | 125.6 | 71.6",
   "retrieved": "2026-10-03",
   "source_name": "HF model card: deluxetiky agnes-vs-qwen38-benchmark",
   "source_url": "https://huggingface.co/deluxetiky/agnes-vs-qwen38-benchmark",
   "tg_tokens": null,
   "tps": "71.6",
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "M4 Max (64 GB)",
   "id": "rariruluis-elyza-m4max-medium",
   "model": "ELYZA-Thinking-1.0-llm-jp-4-32b-a3b",
   "notes": "HF card field note: ELYZA-Thinking-1.0 (Japanese 4th-gen 32B MoE, 3B active), 4-bit MLX (group 64, 4.501 bpw), mlx-lm 0.32.0, Apple M4 Max 64 GB, 4096 ctx, greedy; math 17x23 prompt (reasoning_effort=medium) 114 tok/s; author notes these are small sanity checks, not a full benchmark; config=medium",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "4-bit",
   "quote": "ELYZA-Thinking-1.0-llm-jp-4-32b-a3b-mlx-4bit; Greedy generation with mlx-lm 0.32.0 on an Apple M4 Max (64 GB); 123 | 114 tok/s",
   "retrieved": "2026-10-03",
   "source_name": "HF model card: rariruluis ELYZA-Thinking-1.0-llm-jp-4-32b-a3b-mlx-4bit",
   "source_url": "https://huggingface.co/rariruluis/ELYZA-Thinking-1.0-llm-jp-4-32b-a3b-mlx-4bit",
   "tg_tokens": null,
   "tps": "114",
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "M4 Max (64 GB)",
   "id": "rariruluis-elyza-m4max-low",
   "model": "ELYZA-Thinking-1.0-llm-jp-4-32b-a3b",
   "notes": "HF card field note: ELYZA-Thinking-1.0 (Japanese 4th-gen 32B MoE, 3B active), 4-bit MLX (group 64, 4.501 bpw), mlx-lm 0.32.0, Apple M4 Max 64 GB, 4096 ctx, greedy; code fib(n) prompt (reasoning_effort=low) 129 tok/s; author notes these are small sanity checks, not a full benchmark; config=low",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "4-bit",
   "quote": "ELYZA-Thinking-1.0-llm-jp-4-32b-a3b-mlx-4bit; Greedy generation with mlx-lm 0.32.0 on an Apple M4 Max (64 GB); 583 | 129 tok/s",
   "retrieved": "2026-10-03",
   "source_name": "HF model card: rariruluis ELYZA-Thinking-1.0-llm-jp-4-32b-a3b-mlx-4bit",
   "source_url": "https://huggingface.co/rariruluis/ELYZA-Thinking-1.0-llm-jp-4-32b-a3b-mlx-4bit",
   "tg_tokens": null,
   "tps": "129",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": 48000,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-29892-ub1024-parent",
   "model": "Qwen3.6-35B-A3B",
   "notes": "llama.cpp #29892, Vulkan RDNA4 prefill regression since #29182 (MOE-aware mat_mul_id tile selection); Qwen3.6-35B-A3B UD-IQ3_S (256 experts top-8); llama-bench pp2048 @ d48000, -b 4096 -fa 1, all-VRAM; parent=da89bb3cc (no #29182), master=bed0a8566 (with #29182); config=ub1024-parent",
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": "2154.41",
   "provenance": "community",
   "quant": "IQ3_S",
   "quote": "Qwen3.6-35B-A3B (unsloth UD-IQ3_S, 256 experts, top-8); AMD Radeon RX 9070 XT (RADV GFX1201) (radv); | qwen35moe 35B.A3B IQ3_S - 3.4375 bpw | 12.73 GiB | 34.66 B | Vulkan | 99 | 4096 | 1024 | 1 | pp2048 @ d48000 | 2154.41 \u00b1 6.50 |",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp #29892 (Vulkan RDNA4 prefill regression since #29182)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29892",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": 48000,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-29892-ub1024-master",
   "model": "Qwen3.6-35B-A3B",
   "notes": "llama.cpp #29892, Vulkan RDNA4 prefill regression since #29182 (MOE-aware mat_mul_id tile selection); Qwen3.6-35B-A3B UD-IQ3_S (256 experts top-8); llama-bench pp2048 @ d48000, -b 4096 -fa 1, all-VRAM; parent=da89bb3cc (no #29182), master=bed0a8566 (with #29182); config=ub1024-master",
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": "1875.05",
   "provenance": "community",
   "quant": "IQ3_S",
   "quote": "Qwen3.6-35B-A3B (unsloth UD-IQ3_S, 256 experts, top-8); AMD Radeon RX 9070 XT (RADV GFX1201) (radv); | qwen35moe 35B.A3B IQ3_S - 3.4375 bpw | 12.73 GiB | 34.66 B | Vulkan | 99 | 4096 | 1024 | 1 | pp2048 @ d48000 | 1875.05 \u00b1 5.03 |",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp #29892 (Vulkan RDNA4 prefill regression since #29182)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29892",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": 48000,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-29892-ub2048-parent",
   "model": "Qwen3.6-35B-A3B",
   "notes": "llama.cpp #29892, Vulkan RDNA4 prefill regression since #29182 (MOE-aware mat_mul_id tile selection); Qwen3.6-35B-A3B UD-IQ3_S (256 experts top-8); llama-bench pp2048 @ d48000, -b 4096 -fa 1, all-VRAM; parent=da89bb3cc (no #29182), master=bed0a8566 (with #29182); config=ub2048-parent",
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": "2075.07",
   "provenance": "community",
   "quant": "IQ3_S",
   "quote": "Qwen3.6-35B-A3B (unsloth UD-IQ3_S, 256 experts, top-8); AMD Radeon RX 9070 XT (RADV GFX1201) (radv); | qwen35moe 35B.A3B IQ3_S - 3.4375 bpw | 12.73 GiB | 34.66 B | Vulkan | 99 | 4096 | 2048 | 1 | pp2048 @ d48000 | 2075.07 \u00b1 119.82 |",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp #29892 (Vulkan RDNA4 prefill regression since #29182)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29892",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": 48000,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-29892-ub2048-master",
   "model": "Qwen3.6-35B-A3B",
   "notes": "llama.cpp #29892, Vulkan RDNA4 prefill regression since #29182 (MOE-aware mat_mul_id tile selection); Qwen3.6-35B-A3B UD-IQ3_S (256 experts top-8); llama-bench pp2048 @ d48000, -b 4096 -fa 1, all-VRAM; parent=da89bb3cc (no #29182), master=bed0a8566 (with #29182); config=ub2048-master",
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": "2073.69",
   "provenance": "community",
   "quant": "IQ3_S",
   "quote": "Qwen3.6-35B-A3B (unsloth UD-IQ3_S, 256 experts, top-8); AMD Radeon RX 9070 XT (RADV GFX1201) (radv); | qwen35moe 35B.A3B IQ3_S - 3.4375 bpw | 12.73 GiB | 34.66 B | Vulkan | 99 | 4096 | 2048 | 1 | pp2048 @ d48000 | 2073.69 \u00b1 123.86 |",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp #29892 (Vulkan RDNA4 prefill regression since #29182)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29892",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "lc-29901-prod6000-base",
   "model": "Qwen4-exp",
   "notes": "llama.cpp PR #29901, cuda: tile the lightning indexer over keys/tokens for 4 heads (port of the Vulkan shader); qwen4exp prod config; prefill @ 128k ctx, TG unchanged, pp length unstated; base=before tiling, pr=after; config=base",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "3867",
   "provenance": "community",
   "quant": "unspecified",
   "quote": "On an RTX PRO 6000 the kernel is 2.5x faster in prefill (6.4 vs 16.4 ms at kv 65536, nb 2048), and qwen4exp with my prod config goes from PP 3867 -> 4038 t/s @ 128k with TG unchanged, the indexer dropping from 9.6% to 4.0% of the GPU time in nsys.",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp PR #29901 (CUDA lightning-indexer tiling for 4 heads)",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29901",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "lc-29901-prod6000-pr",
   "model": "Qwen4-exp",
   "notes": "llama.cpp PR #29901, cuda: tile the lightning indexer over keys/tokens for 4 heads (port of the Vulkan shader); qwen4exp prod config; prefill @ 128k ctx, TG unchanged, pp length unstated; base=before tiling, pr=after; config=pr",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "4038",
   "provenance": "community",
   "quant": "unspecified",
   "quote": "On an RTX PRO 6000 the kernel is 2.5x faster in prefill (6.4 vs 16.4 ms at kv 65536, nb 2048), and qwen4exp with my prod config goes from PP 3867 -> 4038 t/s @ 128k with TG unchanged, the indexer dropping from 9.6% to 4.0% of the GPU time in nsys.",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp PR #29901 (CUDA lightning-indexer tiling for 4 heads)",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29901",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "MacBook Pro 128GB",
   "id": "nightmedia-qwen38-27b-mac128-mxfp8",
   "model": "Qwen3.8-27B",
   "notes": "config=mxfp8; MLX quant sweep of the MindMeld-AREX merge (Qwen3.8-27B) on a 128 GB MacBook Pro (M-chip not stated in the card); the card also benchmarks the parent Qwen3.6-27B-MindMeld and AREX-2 as components, not mined here",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "MXFP8",
   "quote": "Qwen3.8-27B-MindMeld-AREX; Our lab is one Macbook Pro 128GB; Quant Perplexity Peak Memory Tokens/sec mxfp8 3.731 \u00b1 0.023 34.74 GB 209 q8-hi 3.721 \u00b1 0.023 37.26 GB 213 qx86-hi 3.721 \u00b1 0.023 33.25 GB 210 qx64-hi 3.727 \u00b1 0.023 27.03 GB 204 mxfp4 3.790 \u00b1 0.023 21.30 GB 206",
   "retrieved": "2026-10-03",
   "source_name": "HF model card: nightmedia Qwen3.8-27B-MindMeld-AREX-mxfp8-mlx",
   "source_url": "https://huggingface.co/nightmedia/Qwen3.8-27B-MindMeld-AREX-mxfp8-mlx",
   "tg_tokens": null,
   "tps": "209",
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "MacBook Pro 128GB",
   "id": "nightmedia-qwen38-27b-mac128-q8hi",
   "model": "Qwen3.8-27B",
   "notes": "config=q8-hi; MLX quant sweep of the MindMeld-AREX merge (Qwen3.8-27B) on a 128 GB MacBook Pro (M-chip not stated in the card); the card also benchmarks the parent Qwen3.6-27B-MindMeld and AREX-2 as components, not mined here",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "q8-hi",
   "quote": "Qwen3.8-27B-MindMeld-AREX; Our lab is one Macbook Pro 128GB; Quant Perplexity Peak Memory Tokens/sec mxfp8 3.731 \u00b1 0.023 34.74 GB 209 q8-hi 3.721 \u00b1 0.023 37.26 GB 213 qx86-hi 3.721 \u00b1 0.023 33.25 GB 210 qx64-hi 3.727 \u00b1 0.023 27.03 GB 204 mxfp4 3.790 \u00b1 0.023 21.30 GB 206",
   "retrieved": "2026-10-03",
   "source_name": "HF model card: nightmedia Qwen3.8-27B-MindMeld-AREX-mxfp8-mlx",
   "source_url": "https://huggingface.co/nightmedia/Qwen3.8-27B-MindMeld-AREX-mxfp8-mlx",
   "tg_tokens": null,
   "tps": "213",
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "MacBook Pro 128GB",
   "id": "nightmedia-qwen38-27b-mac128-qx86hi",
   "model": "Qwen3.8-27B",
   "notes": "config=qx86-hi; MLX quant sweep of the MindMeld-AREX merge (Qwen3.8-27B) on a 128 GB MacBook Pro (M-chip not stated in the card); the card also benchmarks the parent Qwen3.6-27B-MindMeld and AREX-2 as components, not mined here",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "qx86-hi",
   "quote": "Qwen3.8-27B-MindMeld-AREX; Our lab is one Macbook Pro 128GB; Quant Perplexity Peak Memory Tokens/sec mxfp8 3.731 \u00b1 0.023 34.74 GB 209 q8-hi 3.721 \u00b1 0.023 37.26 GB 213 qx86-hi 3.721 \u00b1 0.023 33.25 GB 210 qx64-hi 3.727 \u00b1 0.023 27.03 GB 204 mxfp4 3.790 \u00b1 0.023 21.30 GB 206",
   "retrieved": "2026-10-03",
   "source_name": "HF model card: nightmedia Qwen3.8-27B-MindMeld-AREX-mxfp8-mlx",
   "source_url": "https://huggingface.co/nightmedia/Qwen3.8-27B-MindMeld-AREX-mxfp8-mlx",
   "tg_tokens": null,
   "tps": "210",
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "MacBook Pro 128GB",
   "id": "nightmedia-qwen38-27b-mac128-qx64hi",
   "model": "Qwen3.8-27B",
   "notes": "config=qx64-hi; MLX quant sweep of the MindMeld-AREX merge (Qwen3.8-27B) on a 128 GB MacBook Pro (M-chip not stated in the card); the card also benchmarks the parent Qwen3.6-27B-MindMeld and AREX-2 as components, not mined here",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "qx64-hi",
   "quote": "Qwen3.8-27B-MindMeld-AREX; Our lab is one Macbook Pro 128GB; Quant Perplexity Peak Memory Tokens/sec mxfp8 3.731 \u00b1 0.023 34.74 GB 209 q8-hi 3.721 \u00b1 0.023 37.26 GB 213 qx86-hi 3.721 \u00b1 0.023 33.25 GB 210 qx64-hi 3.727 \u00b1 0.023 27.03 GB 204 mxfp4 3.790 \u00b1 0.023 21.30 GB 206",
   "retrieved": "2026-10-03",
   "source_name": "HF model card: nightmedia Qwen3.8-27B-MindMeld-AREX-mxfp8-mlx",
   "source_url": "https://huggingface.co/nightmedia/Qwen3.8-27B-MindMeld-AREX-mxfp8-mlx",
   "tg_tokens": null,
   "tps": "204",
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "MacBook Pro 128GB",
   "id": "nightmedia-qwen38-27b-mac128-mxfp4",
   "model": "Qwen3.8-27B",
   "notes": "config=mxfp4; MLX quant sweep of the MindMeld-AREX merge (Qwen3.8-27B) on a 128 GB MacBook Pro (M-chip not stated in the card); the card also benchmarks the parent Qwen3.6-27B-MindMeld and AREX-2 as components, not mined here",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Qwen3.8-27B-MindMeld-AREX; Our lab is one Macbook Pro 128GB; Quant Perplexity Peak Memory Tokens/sec mxfp8 3.731 \u00b1 0.023 34.74 GB 209 q8-hi 3.721 \u00b1 0.023 37.26 GB 213 qx86-hi 3.721 \u00b1 0.023 33.25 GB 210 qx64-hi 3.727 \u00b1 0.023 27.03 GB 204 mxfp4 3.790 \u00b1 0.023 21.30 GB 206",
   "retrieved": "2026-10-03",
   "source_name": "HF model card: nightmedia Qwen3.8-27B-MindMeld-AREX-mxfp8-mlx",
   "source_url": "https://huggingface.co/nightmedia/Qwen3.8-27B-MindMeld-AREX-mxfp8-mlx",
   "tg_tokens": null,
   "tps": "206",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "sm_80",
   "id": "vllm-59878-sm80-1k256-triton",
   "model": "gpt-oss 20B",
   "notes": "vLLM PR #59878 (open): routes attention sinks through FlashInfer fa2 wrappers on SM8x, so gpt-oss-20b with a BF16 KV cache selects FLASHINFER instead of TRITON_ATTN by backend priority. vllm bench throughput, 200 prompts, 1024 in / 256 out, BF16 KV cache, averaged over two sm_80 GPUs back to back in opposite orders; tps is total tok/s (input+output); batch-1 decode kernel 9.07 us/layer vs Triton 4.73, so the batch-1 latency gap is in decode and prefill matches; config=triton",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "gpt-oss-20b; All runs are on sm_80; Throughput, 200 prompts, 1024 in / 256 out | 11,283 tok/s | 11,546 tok/s | 2.34% higher",
   "retrieved": "2026-10-03",
   "source_name": "vLLM PR #59878",
   "source_url": "https://github.com/vllm-project/vllm/pull/59878",
   "tg_tokens": 256,
   "tps": "11283",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "sm_80",
   "id": "vllm-59878-sm80-1k256-flashinfer",
   "model": "gpt-oss 20B",
   "notes": "vLLM PR #59878 (open): routes attention sinks through FlashInfer fa2 wrappers on SM8x, so gpt-oss-20b with a BF16 KV cache selects FLASHINFER instead of TRITON_ATTN by backend priority. vllm bench throughput, 200 prompts, 1024 in / 256 out, BF16 KV cache, averaged over two sm_80 GPUs back to back in opposite orders; tps is total tok/s (input+output); batch-1 decode kernel 9.07 us/layer vs Triton 4.73, so the batch-1 latency gap is in decode and prefill matches; config=flashinfer",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "gpt-oss-20b; All runs are on sm_80; Throughput, 200 prompts, 1024 in / 256 out | 11,283 tok/s | 11,546 tok/s | 2.34% higher",
   "retrieved": "2026-10-03",
   "source_name": "vLLM PR #59878",
   "source_url": "https://github.com/vllm-project/vllm/pull/59878",
   "tg_tokens": 256,
   "tps": "11546",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "Arc Pro B50",
   "id": "lc-29908-arcpro-b50-iq3xxs",
   "model": "Qwen3.8-27B",
   "notes": "Whole-model decode tg128 on Arc Pro B50 (Vulkan, -ngl 99). The issue is a bug report on in-model vs isolated MUL_MAT_VEC cost; this is the whole-model decode baseline.",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "GSQ-RCO IQ3_XXS",
   "quote": "Qwen3.8-27B; Arc Pro B50; 10.1 t/s",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp #29908",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29908",
   "tg_tokens": 128,
   "tps": "10.1",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "Arc Pro B50",
   "id": "lc-29908-arcpro-b50-q3km",
   "model": "Qwen3.8-27B",
   "notes": "Full-model Q3_K_M decode on Arc Pro B50 (Vulkan, -ngl 99), used to rule out quant type as the cause of the in-model slowdown.",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "Q3_K_M",
   "quote": "Qwen3.8-27B; Arc Pro B50; Q3_K_M (pure K-quants) decodes at the same 10.2 t/s",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp #29908",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29908",
   "tg_tokens": 128,
   "tps": "10.2",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "RTX 3060 12 GB",
   "id": "lc-29911-3060-before-r1",
   "model": "Mistral Small 4 119B",
   "notes": "llama.cpp PR #29911 (open): MUL_MAT_ID with GGML_PREC_F32 on quantized weights was rejected by CUDA/Vulkan, offloading every ffn_down_exps mul_mat_id to the CPU and copying the full expert matrix VRAM->host per token. Interleaved A/B swapping only libggml-cuda.so, 200 prompt tokens, taskset -c 0-5 -t 6, -ncmoe 33; D2H copies 232->0; arm=before-r1; config=before-r1",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "mistral-small-4-119b-a6b-heretic; RTX 3060 12 GB; before 4.83 t/s 5.40 t/s",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp #29911",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29911",
   "tg_tokens": null,
   "tps": "4.83",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "RTX 3060 12 GB",
   "id": "lc-29911-3060-before-r2",
   "model": "Mistral Small 4 119B",
   "notes": "llama.cpp PR #29911 (open): MUL_MAT_ID with GGML_PREC_F32 on quantized weights was rejected by CUDA/Vulkan, offloading every ffn_down_exps mul_mat_id to the CPU and copying the full expert matrix VRAM->host per token. Interleaved A/B swapping only libggml-cuda.so, 200 prompt tokens, taskset -c 0-5 -t 6, -ncmoe 33; D2H copies 232->0; arm=before-r2; config=before-r2",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "mistral-small-4-119b-a6b-heretic; RTX 3060 12 GB; before 4.83 t/s 5.40 t/s",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp #29911",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29911",
   "tg_tokens": null,
   "tps": "5.40",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "RTX 3060 12 GB",
   "id": "lc-29911-3060-after-r1",
   "model": "Mistral Small 4 119B",
   "notes": "llama.cpp PR #29911 (open): MUL_MAT_ID with GGML_PREC_F32 on quantized weights was rejected by CUDA/Vulkan, offloading every ffn_down_exps mul_mat_id to the CPU and copying the full expert matrix VRAM->host per token. Interleaved A/B swapping only libggml-cuda.so, 200 prompt tokens, taskset -c 0-5 -t 6, -ncmoe 33; D2H copies 232->0; arm=after-r1; config=after-r1",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "mistral-small-4-119b-a6b-heretic; RTX 3060 12 GB; after 18.68 t/s 19.36 t/s",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp #29911",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29911",
   "tg_tokens": null,
   "tps": "18.68",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "RTX 3060 12 GB",
   "id": "lc-29911-3060-after-r2",
   "model": "Mistral Small 4 119B",
   "notes": "llama.cpp PR #29911 (open): MUL_MAT_ID with GGML_PREC_F32 on quantized weights was rejected by CUDA/Vulkan, offloading every ffn_down_exps mul_mat_id to the CPU and copying the full expert matrix VRAM->host per token. Interleaved A/B swapping only libggml-cuda.so, 200 prompt tokens, taskset -c 0-5 -t 6, -ncmoe 33; D2H copies 232->0; arm=after-r2; config=after-r2",
   "power_w": null,
   "pp_tokens": null,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "mistral-small-4-119b-a6b-heretic; RTX 3060 12 GB; after 18.68 t/s 19.36 t/s",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp #29911",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29911",
   "tg_tokens": null,
   "tps": "19.36",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 2048,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "MI50",
   "id": "lc-29910-mi50-ub16-master",
   "model": "Meta-Llama-3-8B-Instruct",
   "notes": "llama.cpp PR #29910 (open): Q2_K mmq VGPR spills fixed with a gentler unroll and removing an unneeded temporary loop; pp2048 -b 2048 -r 10; -ub sweep, Meta-Llama-3-8B-Instruct-Q2_K.gguf; -ub 16, pp_tps 310.24; config=ub16-master",
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": "310.24",
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Meta-Llama-3-8B-Instruct-Q2_K.gguf; MI50; 16 310.24 318.84 +2.77%",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp #29910",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29910",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 2048,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "MI50",
   "id": "lc-29910-mi50-ub16-optimized",
   "model": "Meta-Llama-3-8B-Instruct",
   "notes": "llama.cpp PR #29910 (open): Q2_K mmq VGPR spills fixed with a gentler unroll and removing an unneeded temporary loop; pp2048 -b 2048 -r 10; -ub sweep, Meta-Llama-3-8B-Instruct-Q2_K.gguf; -ub 16, pp_tps 318.84; config=ub16-optimized",
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": "318.84",
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Meta-Llama-3-8B-Instruct-Q2_K.gguf; MI50; 16 310.24 318.84 +2.77%",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp #29910",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29910",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 2048,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "MI50",
   "id": "lc-29910-mi50-ub32-master",
   "model": "Meta-Llama-3-8B-Instruct",
   "notes": "llama.cpp PR #29910 (open): Q2_K mmq VGPR spills fixed with a gentler unroll and removing an unneeded temporary loop; pp2048 -b 2048 -r 10; -ub sweep, Meta-Llama-3-8B-Instruct-Q2_K.gguf; -ub 32, pp_tps 362.25; config=ub32-master",
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": "362.25",
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Meta-Llama-3-8B-Instruct-Q2_K.gguf; MI50; 32 362.25 426.41 +17.71%",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp #29910",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29910",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 2048,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "MI50",
   "id": "lc-29910-mi50-ub32-optimized",
   "model": "Meta-Llama-3-8B-Instruct",
   "notes": "llama.cpp PR #29910 (open): Q2_K mmq VGPR spills fixed with a gentler unroll and removing an unneeded temporary loop; pp2048 -b 2048 -r 10; -ub sweep, Meta-Llama-3-8B-Instruct-Q2_K.gguf; -ub 32, pp_tps 426.41; config=ub32-optimized",
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": "426.41",
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Meta-Llama-3-8B-Instruct-Q2_K.gguf; MI50; 32 362.25 426.41 +17.71%",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp #29910",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29910",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 2048,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "MI50",
   "id": "lc-29910-mi50-ub64-master",
   "model": "Meta-Llama-3-8B-Instruct",
   "notes": "llama.cpp PR #29910 (open): Q2_K mmq VGPR spills fixed with a gentler unroll and removing an unneeded temporary loop; pp2048 -b 2048 -r 10; -ub sweep, Meta-Llama-3-8B-Instruct-Q2_K.gguf; -ub 64, pp_tps 169.92; config=ub64-master",
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": "169.92",
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Meta-Llama-3-8B-Instruct-Q2_K.gguf; MI50; 64 169.92 406.99 +139.52%",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp #29910",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29910",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 2048,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "MI50",
   "id": "lc-29910-mi50-ub64-optimized",
   "model": "Meta-Llama-3-8B-Instruct",
   "notes": "llama.cpp PR #29910 (open): Q2_K mmq VGPR spills fixed with a gentler unroll and removing an unneeded temporary loop; pp2048 -b 2048 -r 10; -ub sweep, Meta-Llama-3-8B-Instruct-Q2_K.gguf; -ub 64, pp_tps 406.99; config=ub64-optimized",
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": "406.99",
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Meta-Llama-3-8B-Instruct-Q2_K.gguf; MI50; 64 169.92 406.99 +139.52%",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp #29910",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29910",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 2048,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "MI50",
   "id": "lc-29910-mi50-ub128-master",
   "model": "Meta-Llama-3-8B-Instruct",
   "notes": "llama.cpp PR #29910 (open): Q2_K mmq VGPR spills fixed with a gentler unroll and removing an unneeded temporary loop; pp2048 -b 2048 -r 10; -ub sweep, Meta-Llama-3-8B-Instruct-Q2_K.gguf; -ub 128, pp_tps 181.72; config=ub128-master",
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": "181.72",
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Meta-Llama-3-8B-Instruct-Q2_K.gguf; MI50; 128 181.72 454.36 +150.03%",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp #29910",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29910",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 2048,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "MI50",
   "id": "lc-29910-mi50-ub128-optimized",
   "model": "Meta-Llama-3-8B-Instruct",
   "notes": "llama.cpp PR #29910 (open): Q2_K mmq VGPR spills fixed with a gentler unroll and removing an unneeded temporary loop; pp2048 -b 2048 -r 10; -ub sweep, Meta-Llama-3-8B-Instruct-Q2_K.gguf; -ub 128, pp_tps 454.36; config=ub128-optimized",
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": "454.36",
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Meta-Llama-3-8B-Instruct-Q2_K.gguf; MI50; 128 181.72 454.36 +150.03%",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp #29910",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29910",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 2048,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "MI50",
   "id": "lc-29910-mi50-ub256-master",
   "model": "Meta-Llama-3-8B-Instruct",
   "notes": "llama.cpp PR #29910 (open): Q2_K mmq VGPR spills fixed with a gentler unroll and removing an unneeded temporary loop; pp2048 -b 2048 -r 10; -ub sweep, Meta-Llama-3-8B-Instruct-Q2_K.gguf; -ub 256, pp_tps 197.86; config=ub256-master",
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": "197.86",
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Meta-Llama-3-8B-Instruct-Q2_K.gguf; MI50; 256 197.86 531.68 +168.72%",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp #29910",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29910",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 2048,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "MI50",
   "id": "lc-29910-mi50-ub256-optimized",
   "model": "Meta-Llama-3-8B-Instruct",
   "notes": "llama.cpp PR #29910 (open): Q2_K mmq VGPR spills fixed with a gentler unroll and removing an unneeded temporary loop; pp2048 -b 2048 -r 10; -ub sweep, Meta-Llama-3-8B-Instruct-Q2_K.gguf; -ub 256, pp_tps 531.68; config=ub256-optimized",
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": "531.68",
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Meta-Llama-3-8B-Instruct-Q2_K.gguf; MI50; 256 197.86 531.68 +168.72%",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp #29910",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29910",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 2048,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "MI50",
   "id": "lc-29910-mi50-ub512-master",
   "model": "Meta-Llama-3-8B-Instruct",
   "notes": "llama.cpp PR #29910 (open): Q2_K mmq VGPR spills fixed with a gentler unroll and removing an unneeded temporary loop; pp2048 -b 2048 -r 10; -ub sweep, Meta-Llama-3-8B-Instruct-Q2_K.gguf; -ub 512, pp_tps 204.92; config=ub512-master",
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": "204.92",
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Meta-Llama-3-8B-Instruct-Q2_K.gguf; MI50; 512 204.92 571.23 +178.76%",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp #29910",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29910",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 2048,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "MI50",
   "id": "lc-29910-mi50-ub512-optimized",
   "model": "Meta-Llama-3-8B-Instruct",
   "notes": "llama.cpp PR #29910 (open): Q2_K mmq VGPR spills fixed with a gentler unroll and removing an unneeded temporary loop; pp2048 -b 2048 -r 10; -ub sweep, Meta-Llama-3-8B-Instruct-Q2_K.gguf; -ub 512, pp_tps 571.23; config=ub512-optimized",
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": "571.23",
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Meta-Llama-3-8B-Instruct-Q2_K.gguf; MI50; 512 204.92 571.23 +178.76%",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp #29910",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29910",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 2048,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "gfx1152",
   "id": "lc-29910-gfx1152-ub16-master",
   "model": "Meta-Llama-3-8B-Instruct",
   "notes": "llama.cpp PR #29910 (open): Q2_K mmq VGPR spills fixed with a gentler unroll and removing an unneeded temporary loop; pp2048 -b 2048 -r 10; -ub sweep, Meta-Llama-3-8B-Instruct-Q2_K.gguf; -ub 16, pp_tps 69.09; config=ub16-master",
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": "69.09",
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Meta-Llama-3-8B-Instruct-Q2_K.gguf; gfx1152; 16 69.09 92.57 +33.98%",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp #29910",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29910",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 2048,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "gfx1152",
   "id": "lc-29910-gfx1152-ub16-optimized",
   "model": "Meta-Llama-3-8B-Instruct",
   "notes": "llama.cpp PR #29910 (open): Q2_K mmq VGPR spills fixed with a gentler unroll and removing an unneeded temporary loop; pp2048 -b 2048 -r 10; -ub sweep, Meta-Llama-3-8B-Instruct-Q2_K.gguf; -ub 16, pp_tps 92.57; config=ub16-optimized",
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": "92.57",
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Meta-Llama-3-8B-Instruct-Q2_K.gguf; gfx1152; 16 69.09 92.57 +33.98%",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp #29910",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29910",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 2048,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "gfx1152",
   "id": "lc-29910-gfx1152-ub32-master",
   "model": "Meta-Llama-3-8B-Instruct",
   "notes": "llama.cpp PR #29910 (open): Q2_K mmq VGPR spills fixed with a gentler unroll and removing an unneeded temporary loop; pp2048 -b 2048 -r 10; -ub sweep, Meta-Llama-3-8B-Instruct-Q2_K.gguf; -ub 32, pp_tps 98.26; config=ub32-master",
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": "98.26",
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Meta-Llama-3-8B-Instruct-Q2_K.gguf; gfx1152; 32 98.26 117.49 +19.57%",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp #29910",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29910",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 2048,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "gfx1152",
   "id": "lc-29910-gfx1152-ub32-optimized",
   "model": "Meta-Llama-3-8B-Instruct",
   "notes": "llama.cpp PR #29910 (open): Q2_K mmq VGPR spills fixed with a gentler unroll and removing an unneeded temporary loop; pp2048 -b 2048 -r 10; -ub sweep, Meta-Llama-3-8B-Instruct-Q2_K.gguf; -ub 32, pp_tps 117.49; config=ub32-optimized",
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": "117.49",
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Meta-Llama-3-8B-Instruct-Q2_K.gguf; gfx1152; 32 98.26 117.49 +19.57%",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp #29910",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29910",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 2048,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "gfx1152",
   "id": "lc-29910-gfx1152-ub64-master",
   "model": "Meta-Llama-3-8B-Instruct",
   "notes": "llama.cpp PR #29910 (open): Q2_K mmq VGPR spills fixed with a gentler unroll and removing an unneeded temporary loop; pp2048 -b 2048 -r 10; -ub sweep, Meta-Llama-3-8B-Instruct-Q2_K.gguf; -ub 64, pp_tps 128.82; config=ub64-master",
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": "128.82",
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Meta-Llama-3-8B-Instruct-Q2_K.gguf; gfx1152; 64 128.82 128.20 -0.48%",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp #29910",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29910",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 2048,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "gfx1152",
   "id": "lc-29910-gfx1152-ub64-optimized",
   "model": "Meta-Llama-3-8B-Instruct",
   "notes": "llama.cpp PR #29910 (open): Q2_K mmq VGPR spills fixed with a gentler unroll and removing an unneeded temporary loop; pp2048 -b 2048 -r 10; -ub sweep, Meta-Llama-3-8B-Instruct-Q2_K.gguf; -ub 64, pp_tps 128.20; config=ub64-optimized",
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": "128.20",
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Meta-Llama-3-8B-Instruct-Q2_K.gguf; gfx1152; 64 128.82 128.20 -0.48%",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp #29910",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29910",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 2048,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "gfx1152",
   "id": "lc-29910-gfx1152-ub128-master",
   "model": "Meta-Llama-3-8B-Instruct",
   "notes": "llama.cpp PR #29910 (open): Q2_K mmq VGPR spills fixed with a gentler unroll and removing an unneeded temporary loop; pp2048 -b 2048 -r 10; -ub sweep, Meta-Llama-3-8B-Instruct-Q2_K.gguf; -ub 128, pp_tps 135.69; config=ub128-master",
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": "135.69",
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Meta-Llama-3-8B-Instruct-Q2_K.gguf; gfx1152; 128 135.69 135.13 -0.41%",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp #29910",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29910",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 2048,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "gfx1152",
   "id": "lc-29910-gfx1152-ub128-optimized",
   "model": "Meta-Llama-3-8B-Instruct",
   "notes": "llama.cpp PR #29910 (open): Q2_K mmq VGPR spills fixed with a gentler unroll and removing an unneeded temporary loop; pp2048 -b 2048 -r 10; -ub sweep, Meta-Llama-3-8B-Instruct-Q2_K.gguf; -ub 128, pp_tps 135.13; config=ub128-optimized",
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": "135.13",
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Meta-Llama-3-8B-Instruct-Q2_K.gguf; gfx1152; 128 135.69 135.13 -0.41%",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp #29910",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29910",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 2048,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "gfx1152",
   "id": "lc-29910-gfx1152-ub256-master",
   "model": "Meta-Llama-3-8B-Instruct",
   "notes": "llama.cpp PR #29910 (open): Q2_K mmq VGPR spills fixed with a gentler unroll and removing an unneeded temporary loop; pp2048 -b 2048 -r 10; -ub sweep, Meta-Llama-3-8B-Instruct-Q2_K.gguf; -ub 256, pp_tps 191.79; config=ub256-master",
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": "191.79",
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Meta-Llama-3-8B-Instruct-Q2_K.gguf; gfx1152; 256 191.79 188.99 -1.46%",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp #29910",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29910",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 2048,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "gfx1152",
   "id": "lc-29910-gfx1152-ub256-optimized",
   "model": "Meta-Llama-3-8B-Instruct",
   "notes": "llama.cpp PR #29910 (open): Q2_K mmq VGPR spills fixed with a gentler unroll and removing an unneeded temporary loop; pp2048 -b 2048 -r 10; -ub sweep, Meta-Llama-3-8B-Instruct-Q2_K.gguf; -ub 256, pp_tps 188.99; config=ub256-optimized",
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": "188.99",
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Meta-Llama-3-8B-Instruct-Q2_K.gguf; gfx1152; 256 191.79 188.99 -1.46%",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp #29910",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29910",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 2048,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "gfx1152",
   "id": "lc-29910-gfx1152-ub512-master",
   "model": "Meta-Llama-3-8B-Instruct",
   "notes": "llama.cpp PR #29910 (open): Q2_K mmq VGPR spills fixed with a gentler unroll and removing an unneeded temporary loop; pp2048 -b 2048 -r 10; -ub sweep, Meta-Llama-3-8B-Instruct-Q2_K.gguf; -ub 512, pp_tps 204.13; config=ub512-master",
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": "204.13",
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Meta-Llama-3-8B-Instruct-Q2_K.gguf; gfx1152; 512 204.13 205.85 +0.84%",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp #29910",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29910",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 2048,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "gfx1152",
   "id": "lc-29910-gfx1152-ub512-optimized",
   "model": "Meta-Llama-3-8B-Instruct",
   "notes": "llama.cpp PR #29910 (open): Q2_K mmq VGPR spills fixed with a gentler unroll and removing an unneeded temporary loop; pp2048 -b 2048 -r 10; -ub sweep, Meta-Llama-3-8B-Instruct-Q2_K.gguf; -ub 512, pp_tps 205.85; config=ub512-optimized",
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": "205.85",
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Meta-Llama-3-8B-Instruct-Q2_K.gguf; gfx1152; 512 204.13 205.85 +0.84%",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp #29910",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29910",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-29924-l318b-ngram-prePR",
   "model": "Llama 3.1 8B",
   "notes": "llama.cpp PR #29924 (open): fix n-gram drafts rejected at temp>0 after truncation; ngram-mod, temp 0.7, 8 seeds, pre-PR build; median t/s (worst seed 419.1), mean drafted 1048, acceptance 0.493; config=ngram-mod-prePR",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Llama-3.1-8B-Instruct Q4_K_M; Tested on RTX 5090; ngram-mod pre-PR 509.6 419.1 1048 0.493",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp #29924",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29924",
   "tg_tokens": null,
   "tps": "509.6",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-29924-l318b-ngram-tot",
   "model": "Llama 3.1 8B",
   "notes": "llama.cpp PR #29924 (open): fix n-gram drafts rejected at temp>0 after truncation; ngram-mod, temp 0.7, 8 seeds, ToT build; median t/s (worst seed 334.7), mean drafted 1368, acceptance 0.365; config=ngram-mod-tot",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Llama-3.1-8B-Instruct Q4_K_M; Tested on RTX 5090; ToT 375.1 334.7 1368 0.365",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp #29924",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29924",
   "tg_tokens": null,
   "tps": "375.1",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-29924-l318b-ngram-fix",
   "model": "Llama 3.1 8B",
   "notes": "llama.cpp PR #29924 (open): fix n-gram drafts rejected at temp>0 after truncation; ngram-mod, temp 0.7, 8 seeds, fix build; median t/s (worst seed 419.0), mean drafted 1048, acceptance 0.493; config=ngram-mod-fix",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Llama-3.1-8B-Instruct Q4_K_M; Tested on RTX 5090; fix 500.7 419.0 1048 0.493",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp #29924",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29924",
   "tg_tokens": null,
   "tps": "500.7",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-29924-q36-ngram-mtp1-tot",
   "model": "Qwen3.6-35B-A3B",
   "notes": "llama.cpp PR #29924 (open): fix n-gram drafts rejected at temp>0 after truncation; ngram-mod,draft-mtp (np=1), temp 0.7, 8 seeds, ToT build; median t/s (worst seed 203.1), mean drafted 1306, acceptance 0.496; config=ngram-mtp-np1-tot",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Qwen3.6-35B-A3B Q4_K_M; Tested on RTX 5090; ngram-mod,draft-mtp (np=1) ToT 301.7 203.1 1306 0.496",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp #29924",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29924",
   "tg_tokens": null,
   "tps": "301.7",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-29924-q36-ngram-mtp1-fix",
   "model": "Qwen3.6-35B-A3B",
   "notes": "llama.cpp PR #29924 (open): fix n-gram drafts rejected at temp>0 after truncation; ngram-mod,draft-mtp (np=1), temp 0.7, 8 seeds, fix build; median t/s (worst seed 304.9), mean drafted 1168, acceptance 0.553; config=ngram-mtp-np1-fix",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Qwen3.6-35B-A3B Q4_K_M; Tested on RTX 5090; fix 333.9 304.9 1168 0.553",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp #29924",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29924",
   "tg_tokens": null,
   "tps": "333.9",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-29924-q36-ngram-tot",
   "model": "Qwen3.6-35B-A3B",
   "notes": "llama.cpp PR #29924 (open): fix n-gram drafts rejected at temp>0 after truncation; ngram-mod, temp 0.7, 8 seeds, ToT build; median t/s (worst seed 184.9), mean drafted 849, acceptance 0.377; config=ngram-tot",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Qwen3.6-35B-A3B Q4_K_M; Tested on RTX 5090; ngram-mod ToT 227.8 184.9 849 0.377",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp #29924",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29924",
   "tg_tokens": null,
   "tps": "227.8",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-29924-q36-ngram-fix",
   "model": "Qwen3.6-35B-A3B",
   "notes": "llama.cpp PR #29924 (open): fix n-gram drafts rejected at temp>0 after truncation; ngram-mod, temp 0.7, 8 seeds, fix build; median t/s (worst seed 262.0), mean drafted 780, acceptance 0.514; config=ngram-fix",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Qwen3.6-35B-A3B Q4_K_M; Tested on RTX 5090; fix 292.8 262.0 780 0.514",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp #29924",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29924",
   "tg_tokens": null,
   "tps": "292.8",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-29924-q36-ngram-mtp2-tot",
   "model": "Qwen3.6-35B-A3B",
   "notes": "llama.cpp PR #29924 (open): fix n-gram drafts rejected at temp>0 after truncation; ngram-mod,draft-mtp (np=2), temp 0.7, 8 seeds, ToT build; median t/s (worst seed 145.9), mean drafted 1254, acceptance 0.538; config=ngram-mtp-np2-tot",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Qwen3.6-35B-A3B Q4_K_M; Tested on RTX 5090; ngram-mod,draft-mtp (np=2) ToT 187.0 145.9 1254 0.538",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp #29924",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29924",
   "tg_tokens": null,
   "tps": "187.0",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-29924-q36-ngram-mtp2-fix",
   "model": "Qwen3.6-35B-A3B",
   "notes": "llama.cpp PR #29924 (open): fix n-gram drafts rejected at temp>0 after truncation; ngram-mod,draft-mtp (np=2), temp 0.7, 8 seeds, fix build; median t/s (worst seed 225.4), mean drafted 1041, acceptance 0.670; config=ngram-mtp-np2-fix",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Qwen3.6-35B-A3B Q4_K_M; Tested on RTX 5090; fix 329.4 225.4 1041 0.670",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp #29924",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29924",
   "tg_tokens": null,
   "tps": "329.4",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-29924-q36-ngram-mtpprob-tot",
   "model": "Qwen3.6-35B-A3B",
   "notes": "llama.cpp PR #29924 (open): fix n-gram drafts rejected at temp>0 after truncation; ngram-mod,draft-mtp (probabilistic), temp 0.7, 8 seeds, ToT build; median t/s (worst seed 218.4), mean drafted 1233, acceptance 0.522; config=ngram-mtp-prob-tot",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Qwen3.6-35B-A3B Q4_K_M; Tested on RTX 5090; ngram-mod,draft-mtp (--spec-draft-sampling probabilistic) ToT 307.6 218.4 1233 0.522",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp #29924",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29924",
   "tg_tokens": null,
   "tps": "307.6",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-29924-q36-ngram-mtpprob-fix",
   "model": "Qwen3.6-35B-A3B",
   "notes": "llama.cpp PR #29924 (open): fix n-gram drafts rejected at temp>0 after truncation; ngram-mod,draft-mtp (probabilistic), temp 0.7, 8 seeds, fix build; median t/s (worst seed 314.2), mean drafted 1214, acceptance 0.529; config=ngram-mtp-prob-fix",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Qwen3.6-35B-A3B Q4_K_M; Tested on RTX 5090; fix 344.1 314.2 1214 0.529",
   "retrieved": "2026-10-03",
   "source_name": "llama.cpp #29924",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29924",
   "tg_tokens": null,
   "tps": "344.1",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "GB10",
   "id": "hf-promzeus-glm46-gb10-nospec",
   "model": "GLM-4.6",
   "notes": "config=no-spec",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ2_XXS+Q5_K",
   "quote": "GLM-4.6 (355B-A32B MoE); GB10; no speculation 32k q8_0 11.45 MTP, n-max 1 113k q4_0 17.14 0.80-0.94 MTP, n-max 2 32k q8_0 14.36 0.41-0.47 MTP, n-max 3 32k q8_0 12.17 0.26-0.33; 57.6k 197 4.10, with MTP 6.28 111.9k 125 3.61 with MTP",
   "retrieved": "2026-10-03",
   "source_name": "promzeus/gh0stx-glm46-gb10-GGUF",
   "source_url": "https://huggingface.co/promzeus/gh0stx-glm46-gb10-GGUF",
   "tg_tokens": 128,
   "tps": "11.45",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 113664,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "GB10",
   "id": "hf-promzeus-glm46-gb10-mtp1",
   "model": "GLM-4.6",
   "notes": "config=mtp-nmax1",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ2_XXS+Q5_K",
   "quote": "GLM-4.6 (355B-A32B MoE); GB10; no speculation 32k q8_0 11.45 MTP, n-max 1 113k q4_0 17.14 0.80-0.94 MTP, n-max 2 32k q8_0 14.36 0.41-0.47 MTP, n-max 3 32k q8_0 12.17 0.26-0.33; 57.6k 197 4.10, with MTP 6.28 111.9k 125 3.61 with MTP",
   "retrieved": "2026-10-03",
   "source_name": "promzeus/gh0stx-glm46-gb10-GGUF",
   "source_url": "https://huggingface.co/promzeus/gh0stx-glm46-gb10-GGUF",
   "tg_tokens": 128,
   "tps": "17.14",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "GB10",
   "id": "hf-promzeus-glm46-gb10-mtp2",
   "model": "GLM-4.6",
   "notes": "config=mtp-nmax2",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ2_XXS+Q5_K",
   "quote": "GLM-4.6 (355B-A32B MoE); GB10; no speculation 32k q8_0 11.45 MTP, n-max 1 113k q4_0 17.14 0.80-0.94 MTP, n-max 2 32k q8_0 14.36 0.41-0.47 MTP, n-max 3 32k q8_0 12.17 0.26-0.33; 57.6k 197 4.10, with MTP 6.28 111.9k 125 3.61 with MTP",
   "retrieved": "2026-10-03",
   "source_name": "promzeus/gh0stx-glm46-gb10-GGUF",
   "source_url": "https://huggingface.co/promzeus/gh0stx-glm46-gb10-GGUF",
   "tg_tokens": 128,
   "tps": "14.36",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "GB10",
   "id": "hf-promzeus-glm46-gb10-mtp3",
   "model": "GLM-4.6",
   "notes": "config=mtp-nmax3",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ2_XXS+Q5_K",
   "quote": "GLM-4.6 (355B-A32B MoE); GB10; no speculation 32k q8_0 11.45 MTP, n-max 1 113k q4_0 17.14 0.80-0.94 MTP, n-max 2 32k q8_0 14.36 0.41-0.47 MTP, n-max 3 32k q8_0 12.17 0.26-0.33; 57.6k 197 4.10, with MTP 6.28 111.9k 125 3.61 with MTP",
   "retrieved": "2026-10-03",
   "source_name": "promzeus/gh0stx-glm46-gb10-GGUF",
   "source_url": "https://huggingface.co/promzeus/gh0stx-glm46-gb10-GGUF",
   "tg_tokens": 128,
   "tps": "12.17",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 57600,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "GB10",
   "id": "hf-promzeus-glm46-gb10-57k",
   "model": "GLM-4.6",
   "notes": "config=depth-57k",
   "pp_tokens": null,
   "pp_tps": "197",
   "provenance": "community",
   "quant": "IQ2_XXS+Q5_K",
   "quote": "GLM-4.6 (355B-A32B MoE); GB10; no speculation 32k q8_0 11.45 MTP, n-max 1 113k q4_0 17.14 0.80-0.94 MTP, n-max 2 32k q8_0 14.36 0.41-0.47 MTP, n-max 3 32k q8_0 12.17 0.26-0.33; 57.6k 197 4.10, with MTP 6.28 111.9k 125 3.61 with MTP",
   "retrieved": "2026-10-03",
   "source_name": "promzeus/gh0stx-glm46-gb10-GGUF",
   "source_url": "https://huggingface.co/promzeus/gh0stx-glm46-gb10-GGUF",
   "tg_tokens": 128,
   "tps": "4.10",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 111900,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "GB10",
   "id": "hf-promzeus-glm46-gb10-112k",
   "model": "GLM-4.6",
   "notes": "config=depth-112k",
   "pp_tokens": null,
   "pp_tps": "125",
   "provenance": "community",
   "quant": "IQ2_XXS+Q5_K",
   "quote": "GLM-4.6 (355B-A32B MoE); GB10; no speculation 32k q8_0 11.45 MTP, n-max 1 113k q4_0 17.14 0.80-0.94 MTP, n-max 2 32k q8_0 14.36 0.41-0.47 MTP, n-max 3 32k q8_0 12.17 0.26-0.33; 57.6k 197 4.10, with MTP 6.28 111.9k 125 3.61 with MTP",
   "retrieved": "2026-10-03",
   "source_name": "promzeus/gh0stx-glm46-gb10-GGUF",
   "source_url": "https://huggingface.co/promzeus/gh0stx-glm46-gb10-GGUF",
   "tg_tokens": 128,
   "tps": "3.61",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 16384,
   "date": "2026-10-04",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hf1-qwen359b-3090-fp16",
   "model": "Qwen3.5-9B",
   "notes": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "Qwen3.5-9B; RTX 3090; 78.2 t/s",
   "retrieved": "2026-10-04",
   "source_name": "HF model card: DuoNeural Qwen3.5-9B TAP-DPQ v7",
   "source_url": "https://huggingface.co/DuoNeural/Qwen3.5-9B-TAP-DPQ-v7-GGUF",
   "tg_tokens": null,
   "tps": "78.2",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 16384,
   "date": "2026-10-04",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hf1-qwen359b-3090-q4km",
   "model": "Qwen3.5-9B",
   "notes": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Qwen3.5-9B; RTX 3090; 112.5 t/s",
   "retrieved": "2026-10-04",
   "source_name": "HF model card: DuoNeural Qwen3.5-9B TAP-DPQ v7",
   "source_url": "https://huggingface.co/DuoNeural/Qwen3.5-9B-TAP-DPQ-v7-GGUF",
   "tg_tokens": null,
   "tps": "112.5",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 16384,
   "date": "2026-10-04",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hf1-qwen359b-3090-iq3xxs",
   "model": "Qwen3.5-9B",
   "notes": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ3_XXS",
   "quote": "Qwen3.5-9B; RTX 3090; 138.7 t/s",
   "retrieved": "2026-10-04",
   "source_name": "HF model card: DuoNeural Qwen3.5-9B TAP-DPQ v7",
   "source_url": "https://huggingface.co/DuoNeural/Qwen3.5-9B-TAP-DPQ-v7-GGUF",
   "tg_tokens": null,
   "tps": "138.7",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 16384,
   "date": "2026-10-04",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hf1-qwen359b-3090-iq2m",
   "model": "Qwen3.5-9B",
   "notes": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ2_M",
   "quote": "Qwen3.5-9B; RTX 3090; 145.2 t/s",
   "retrieved": "2026-10-04",
   "source_name": "HF model card: DuoNeural Qwen3.5-9B TAP-DPQ v7",
   "source_url": "https://huggingface.co/DuoNeural/Qwen3.5-9B-TAP-DPQ-v7-GGUF",
   "tg_tokens": null,
   "tps": "145.2",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 16384,
   "date": "2026-10-04",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hf1-qwen359b-3090-iq2xxs",
   "model": "Qwen3.5-9B",
   "notes": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ2_XXS",
   "quote": "Qwen3.5-9B; RTX 3090; 156.4 t/s",
   "retrieved": "2026-10-04",
   "source_name": "HF model card: DuoNeural Qwen3.5-9B TAP-DPQ v7",
   "source_url": "https://huggingface.co/DuoNeural/Qwen3.5-9B-TAP-DPQ-v7-GGUF",
   "tg_tokens": null,
   "tps": "156.4",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 16384,
   "date": "2026-10-04",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hf1-qwen359b-3090-iq1s",
   "model": "Qwen3.5-9B",
   "notes": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ1_S",
   "quote": "Qwen3.5-9B; RTX 3090; 168.1 t/s",
   "retrieved": "2026-10-04",
   "source_name": "HF model card: DuoNeural Qwen3.5-9B TAP-DPQ v7",
   "source_url": "https://huggingface.co/DuoNeural/Qwen3.5-9B-TAP-DPQ-v7-GGUF",
   "tg_tokens": null,
   "tps": "168.1",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-04",
   "flags": [],
   "hardware": "RTX 4080 SUPER",
   "id": "hf1-lfm258b-4080s-fp16",
   "model": "LFM2.5-8B-A1B",
   "notes": "",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "LFM2.5-8B-A1B; RTX 4080 Super; 118.5 t/s",
   "retrieved": "2026-10-04",
   "source_name": "DuoNeural/LFM2.5-8B-A1B-TAP-DPQ-v7-GGUF",
   "source_url": "https://huggingface.co/DuoNeural/LFM2.5-8B-A1B-TAP-DPQ-v7-GGUF",
   "tg_tokens": null,
   "tps": "118.5",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-04",
   "flags": [],
   "hardware": "RTX 4080 SUPER",
   "id": "hf1-lfm258b-4080s-q4km",
   "model": "LFM2.5-8B-A1B",
   "notes": "",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "LFM2.5-8B-A1B; RTX 4080 Super; 172.4 t/s",
   "retrieved": "2026-10-04",
   "source_name": "DuoNeural/LFM2.5-8B-A1B-TAP-DPQ-v7-GGUF",
   "source_url": "https://huggingface.co/DuoNeural/LFM2.5-8B-A1B-TAP-DPQ-v7-GGUF",
   "tg_tokens": null,
   "tps": "172.4",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-04",
   "flags": [],
   "hardware": "RTX 4080 SUPER",
   "id": "hf1-lfm258b-4080s-iq3xxs",
   "model": "LFM2.5-8B-A1B",
   "notes": "",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ3_XXS",
   "quote": "LFM2.5-8B-A1B; RTX 4080 Super; 196.2 t/s",
   "retrieved": "2026-10-04",
   "source_name": "DuoNeural/LFM2.5-8B-A1B-TAP-DPQ-v7-GGUF",
   "source_url": "https://huggingface.co/DuoNeural/LFM2.5-8B-A1B-TAP-DPQ-v7-GGUF",
   "tg_tokens": null,
   "tps": "196.2",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-04",
   "flags": [],
   "hardware": "RTX 4080 SUPER",
   "id": "hf1-lfm258b-4080s-iq2m",
   "model": "LFM2.5-8B-A1B",
   "notes": "",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ2_M",
   "quote": "LFM2.5-8B-A1B; RTX 4080 Super; 208.5 t/s",
   "retrieved": "2026-10-04",
   "source_name": "DuoNeural/LFM2.5-8B-A1B-TAP-DPQ-v7-GGUF",
   "source_url": "https://huggingface.co/DuoNeural/LFM2.5-8B-A1B-TAP-DPQ-v7-GGUF",
   "tg_tokens": null,
   "tps": "208.5",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-04",
   "flags": [],
   "hardware": "RTX 4080 SUPER",
   "id": "hf1-lfm258b-4080s-iq2xxs",
   "model": "LFM2.5-8B-A1B",
   "notes": "",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ2_XXS",
   "quote": "LFM2.5-8B-A1B; RTX 4080 Super; 224.1 t/s",
   "retrieved": "2026-10-04",
   "source_name": "DuoNeural/LFM2.5-8B-A1B-TAP-DPQ-v7-GGUF",
   "source_url": "https://huggingface.co/DuoNeural/LFM2.5-8B-A1B-TAP-DPQ-v7-GGUF",
   "tg_tokens": null,
   "tps": "224.1",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-04",
   "flags": [],
   "hardware": "RTX 4080 SUPER",
   "id": "hf1-lfm258b-4080s-iq1s",
   "model": "LFM2.5-8B-A1B",
   "notes": "",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ1_S",
   "quote": "LFM2.5-8B-A1B; RTX 4080 Super; 241.8 t/s",
   "retrieved": "2026-10-04",
   "source_name": "DuoNeural/LFM2.5-8B-A1B-TAP-DPQ-v7-GGUF",
   "source_url": "https://huggingface.co/DuoNeural/LFM2.5-8B-A1B-TAP-DPQ-v7-GGUF",
   "tg_tokens": null,
   "tps": "241.8",
   "ttft_s": null
  },
  {
   "backend": "TensorFold (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-04",
   "flags": [],
   "hardware": "RTX 4090 Laptop GPU",
   "id": "capyctl-frognano-tf-single",
   "model": "FrogNano-4B-2609",
   "notes": "HF card field note: 4-bit MLX affine (group 64), TensorFold 0.6.3 CUDA engine, greedy, thinking on, median of 3 prompts; dedicated single-stream decode. config=single",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "4-bit",
   "quote": "FrogNano-4B-2609; RTX 4090 Laptop GPU; 52.4 tok/s",
   "retrieved": "2026-10-04",
   "source_name": "HF model card: capyctl FrogNano-4B-2609 MLX 4bit",
   "source_url": "https://huggingface.co/capyctl/FrogNano-4B-2609-MLX-4bit",
   "tg_tokens": null,
   "tps": "52.4",
   "ttft_s": null
  },
  {
   "backend": "TensorFold (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-04",
   "flags": [],
   "hardware": "RTX 4090 Laptop GPU",
   "id": "capyctl-frognano-tf-c1",
   "model": "FrogNano-4B-2609",
   "notes": "HF card field note: 4-bit MLX affine (group 64), TensorFold 0.6.3 --parallel 8, 512 tokens a request, greedy, thinking on, median of 5 rounds; concurrency-sweep aggregate tok/s. config=c1",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "4-bit",
   "quote": "FrogNano-4B-2609; RTX 4090 Laptop GPU; 50.4 tok/s",
   "retrieved": "2026-10-04",
   "source_name": "HF model card: capyctl FrogNano-4B-2609 MLX 4bit",
   "source_url": "https://huggingface.co/capyctl/FrogNano-4B-2609-MLX-4bit",
   "tg_tokens": null,
   "tps": "50.4",
   "ttft_s": null
  },
  {
   "backend": "TensorFold (CUDA)",
   "batch": 2,
   "ctx": null,
   "date": "2026-10-04",
   "flags": [],
   "hardware": "RTX 4090 Laptop GPU",
   "id": "capyctl-frognano-tf-c2",
   "model": "FrogNano-4B-2609",
   "notes": "HF card field note: 4-bit MLX affine (group 64), TensorFold 0.6.3 --parallel 8, 512 tokens a request, greedy, thinking on, median of 5 rounds; concurrency-sweep aggregate tok/s. config=c2",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "4-bit",
   "quote": "FrogNano-4B-2609; RTX 4090 Laptop GPU; 93.4 tok/s",
   "retrieved": "2026-10-04",
   "source_name": "HF model card: capyctl FrogNano-4B-2609 MLX 4bit",
   "source_url": "https://huggingface.co/capyctl/FrogNano-4B-2609-MLX-4bit",
   "tg_tokens": null,
   "tps": "93.4",
   "ttft_s": null
  },
  {
   "backend": "TensorFold (CUDA)",
   "batch": 4,
   "ctx": null,
   "date": "2026-10-04",
   "flags": [],
   "hardware": "RTX 4090 Laptop GPU",
   "id": "capyctl-frognano-tf-c4",
   "model": "FrogNano-4B-2609",
   "notes": "HF card field note: 4-bit MLX affine (group 64), TensorFold 0.6.3 --parallel 8, 512 tokens a request, greedy, thinking on, median of 5 rounds; concurrency-sweep aggregate tok/s. config=c4",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "4-bit",
   "quote": "FrogNano-4B-2609; RTX 4090 Laptop GPU; 176.5 tok/s",
   "retrieved": "2026-10-04",
   "source_name": "HF model card: capyctl FrogNano-4B-2609 MLX 4bit",
   "source_url": "https://huggingface.co/capyctl/FrogNano-4B-2609-MLX-4bit",
   "tg_tokens": null,
   "tps": "176.5",
   "ttft_s": null
  },
  {
   "backend": "TensorFold (CUDA)",
   "batch": 8,
   "ctx": null,
   "date": "2026-10-04",
   "flags": [],
   "hardware": "RTX 4090 Laptop GPU",
   "id": "capyctl-frognano-tf-c8",
   "model": "FrogNano-4B-2609",
   "notes": "HF card field note: 4-bit MLX affine (group 64), TensorFold 0.6.3 --parallel 8, 512 tokens a request, greedy, thinking on, median of 5 rounds; concurrency-sweep aggregate tok/s. config=c8",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "4-bit",
   "quote": "FrogNano-4B-2609; RTX 4090 Laptop GPU; 319.1 tok/s",
   "retrieved": "2026-10-04",
   "source_name": "HF model card: capyctl FrogNano-4B-2609 MLX 4bit",
   "source_url": "https://huggingface.co/capyctl/FrogNano-4B-2609-MLX-4bit",
   "tg_tokens": null,
   "tps": "319.1",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "4x B300",
   "id": "vllm-59916-b300x4-glm53-tp1pcp4-replicated",
   "model": "GLM-5.3",
   "notes": "prefix-pooled prefill 60K shared + 6K unique, OSL 1, concurrency 8; pp_tps is prefill rate in K tok/s (thousand tokens/s); config=TP1\u00d7PCP4, replicated (no DCP/KVPP); baseline without KV sharding; KV 637K tokens; 36-prefix n/a; source body rewritten 2026-10-06 with new measurements, record tracks current body",
   "params": null,
   "power_w": null,
   "pp_tps": "51.5",
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Prefill throughput on GLM-5.3-NVFP4 (4x B300, 32 GiB/GPU KV budget); TP1\u00d7PCP4 | No DCP | 51,526 | 636,928; That is 2.85x no DCP (51.5K)",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #59916 (DCP gather benchmark)",
   "source_url": "https://github.com/vllm-project/vllm/pull/59916",
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "4x B300",
   "id": "vllm-59916-b300x4-glm53-tp1pcp4-kvpp",
   "model": "GLM-5.3",
   "notes": "prefix-pooled prefill 60K shared + 6K unique, OSL 1, concurrency 8; pp_tps is prefill rate in K tok/s (thousand tokens/s); config=TP1\u00d7PCP4 + KVPP (#59059, layer-sharded KV); KV 2.18M tokens; 36-prefix rate 98.3K; source body rewritten 2026-10-06 with new measurements, record tracks current body",
   "params": null,
   "power_w": null,
   "pp_tps": "142.7",
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Prefill throughput on GLM-5.3-NVFP4 (4x B300, 32 GiB/GPU KV budget); TP1\u00d7PCP4 | KVPP (#59059) | 142,717 | 98,273 | 2,182,528; against 144.1K for native DCP and 142.7K for KVPP",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #59916 (DCP gather benchmark)",
   "source_url": "https://github.com/vllm-project/vllm/pull/59916",
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "4x B300",
   "id": "vllm-59916-b300x4-glm53-tp1pcp4-nativedcp4",
   "model": "GLM-5.3",
   "notes": "prefix-pooled prefill 60K shared + 6K unique, OSL 1, concurrency 8; pp_tps is prefill rate in K tok/s (thousand tokens/s); config=TP1\u00d7PCP4 + native DCP4; KV 2.54M tokens; 36-prefix rate 127.6K; source body rewritten 2026-10-06 with new measurements, record tracks current body",
   "params": null,
   "power_w": null,
   "pp_tps": "144.1",
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Prefill throughput on GLM-5.3-NVFP4 (4x B300, 32 GiB/GPU KV budget); TP1\u00d7PCP4 | Native DCP4 (#59951) | 144,131 | 127,610 | 2,540,936; Gather-DCP runs at 146.0K tok/s, against 144.1K for native DCP",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #59916 (DCP gather benchmark)",
   "source_url": "https://github.com/vllm-project/vllm/pull/59916",
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "4x B300",
   "id": "vllm-59916-b300x4-glm53-tp1pcp4-gatherdcp4",
   "model": "GLM-5.3",
   "notes": "prefix-pooled prefill 60K shared + 6K unique, OSL 1, concurrency 8; pp_tps is prefill rate in K tok/s (thousand tokens/s); config=TP1\u00d7PCP4 + gather DCP4 (this PR, -dcp 4 auto gather); KV 2.40M tokens; 36-prefix rate 122.3K; source body rewritten 2026-10-06 with new measurements, record tracks current body",
   "params": null,
   "power_w": null,
   "pp_tps": "146.0",
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Prefill throughput on GLM-5.3-NVFP4 (4x B300, 32 GiB/GPU KV budget); TP1\u00d7PCP4 | --dcp-gather, DCP4 | 146,035 | 122,330 | 2,400,510; Gather-DCP runs at 146.0K tok/s, against 144.1K for native DCP and 142.7K for KVPP",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #59916 (DCP gather benchmark)",
   "source_url": "https://github.com/vllm-project/vllm/pull/59916",
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "4x B300",
   "id": "vllm-59916-b300x4-glm53-tp4-kvpp",
   "model": "GLM-5.3",
   "notes": "prefix-pooled prefill 60K shared + 6K unique, OSL 1, concurrency 8; pp_tps is prefill rate in K tok/s (thousand tokens/s); config=TP4 + KVPP (#59059, layer-sharded KV); KV 2.18M tokens; source body rewritten 2026-10-06 with new measurements, record tracks current body",
   "params": null,
   "power_w": null,
   "pp_tps": "62.2",
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Prefill throughput on GLM-5.3-NVFP4 (4x B300, 32 GiB/GPU KV budget); TP4 | KVPP (#59059) | 62,247 | 2,182,528; TP4: 62.2K tok/s, 1.57x native DCP (39.7K) and level with KVPP (62.2K)",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #59916 (DCP gather benchmark)",
   "source_url": "https://github.com/vllm-project/vllm/pull/59916",
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "4x B300",
   "id": "vllm-59916-b300x4-glm53-tp4-nativedcp4",
   "model": "GLM-5.3",
   "notes": "prefix-pooled prefill 60K shared + 6K unique, OSL 1, concurrency 8; pp_tps is prefill rate in K tok/s (thousand tokens/s); config=TP4 + native DCP4; KV 2.54M tokens; source body rewritten 2026-10-06 with new measurements, record tracks current body",
   "params": null,
   "power_w": null,
   "pp_tps": "39.7",
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Prefill throughput on GLM-5.3-NVFP4 (4x B300, 32 GiB/GPU KV budget); TP4 | Native DCP4 | 39,716 | 2,540,936; 1.57x native DCP (39.7K)",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #59916 (DCP gather benchmark)",
   "source_url": "https://github.com/vllm-project/vllm/pull/59916",
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "4x B300",
   "id": "vllm-59916-b300x4-glm53-tp4-gatherdcp4",
   "model": "GLM-5.3",
   "notes": "prefix-pooled prefill 60K shared + 6K unique, OSL 1, concurrency 8; pp_tps is prefill rate in K tok/s (thousand tokens/s); config=TP4 + gather DCP4 (this PR, -dcp 4 auto gather); KV 2.40M tokens; source body rewritten 2026-10-06 with new measurements, record tracks current body",
   "params": null,
   "power_w": null,
   "pp_tps": "62.2",
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Prefill throughput on GLM-5.3-NVFP4 (4x B300, 32 GiB/GPU KV budget); TP4 | --dcp-gather, DCP4 | 62,204 | 2,400,510; TP4: 62.2K tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #59916 (DCP gather benchmark)",
   "source_url": "https://github.com/vllm-project/vllm/pull/59916",
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "4x B300",
   "id": "vllm-59916-b300x4-glm53-pp2tp2-replicated",
   "model": "GLM-5.3",
   "notes": "prefix-pooled prefill 60K shared + 6K unique, OSL 1, concurrency 8; pp_tps is prefill rate in K tok/s (thousand tokens/s); config=PP2\u00d7TP2, replicated (no DCP/KVPP); baseline without KV sharding; source body rewritten 2026-10-06 with new measurements, record tracks current body",
   "params": null,
   "power_w": null,
   "pp_tps": "59.0",
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Prefill throughput on GLM-5.3-NVFP4 (4x B300, 32 GiB/GPU KV budget); PP2\u00d7TP2 | No DCP | 58,961 | 1,264,448; 1.73x no DCP (59.0K)",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #59916 (DCP gather benchmark)",
   "source_url": "https://github.com/vllm-project/vllm/pull/59916",
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "4x B300",
   "id": "vllm-59916-b300x4-glm53-pp2tp2-kvpp",
   "model": "GLM-5.3",
   "notes": "prefix-pooled prefill 60K shared + 6K unique, OSL 1, concurrency 8; pp_tps is prefill rate in K tok/s (thousand tokens/s); config=PP2\u00d7TP2 + KVPP (#59059, layer-sharded KV); source body rewritten 2026-10-06 with new measurements, record tracks current body",
   "params": null,
   "power_w": null,
   "pp_tps": "102.1",
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Prefill throughput on GLM-5.3-NVFP4 (4x B300, 32 GiB/GPU KV budget); PP2\u00d7TP2 | KVPP (#59059) | 102,097 | 2,182,528; PP2\u00d7TP2: 102.0K tok/s, 1.73x no DCP (59.0K) and level with KVPP (102.1K)",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #59916 (DCP gather benchmark)",
   "source_url": "https://github.com/vllm-project/vllm/pull/59916",
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "4x B300",
   "id": "vllm-59916-b300x4-glm53-pp2tp2-gatherdcp2",
   "model": "GLM-5.3",
   "notes": "prefix-pooled prefill 60K shared + 6K unique, OSL 1, concurrency 8; pp_tps is prefill rate in K tok/s (thousand tokens/s); config=PP2\u00d7TP2 + gather DCP2 (this PR, -dcp 2 auto gather); source body rewritten 2026-10-06 with new measurements, record tracks current body",
   "params": null,
   "power_w": null,
   "pp_tps": "102.0",
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Prefill throughput on GLM-5.3-NVFP4 (4x B300, 32 GiB/GPU KV budget); PP2\u00d7TP2 | --dcp-gather, DCP2 | 102,020 | 2,388,149; PP2\u00d7TP2: 102.0K tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #59916 (DCP gather benchmark)",
   "source_url": "https://github.com/vllm-project/vllm/pull/59916",
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 30000,
   "date": "2026-10-04",
   "flags": [],
   "hardware": "2x RTX 3080 20GB",
   "id": "lc-29935-2x3080-qwen35arch27b-f16-30k",
   "model": "qwen3.5-arch 27B",
   "notes": "perf issue #29935 (CUDA fattn KV streaming on sm86, context-decay report); local build 836d571 (= b11379 + 2 commits, fattn identical); 2x RTX 3080 20GB, qwen3.5-arch 27B NVFP4, -fa on (tensor split), MTP3 spec (verify batch 4), single request; official b11379 f16 48.2 @30K for reference; decode tok/s at 30K context, f16 KV; f16 KV decays slower than q8_0 (no conversion pass); config=f16-30k",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4 / f16 KV",
   "quote": "2x RTX 3080 20GB (sm86, `-sm tensor`), Windows, model = qwen3.5-arch 27B NVFP4; | local 836d571 | f16 | 50.6 | 44.0 | ~0.075 ms / 1K tok |",
   "ram_gb": null,
   "retrieved": "2026-10-04",
   "source_name": "llama.cpp issue #29935 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29935",
   "tg_tokens": null,
   "tps": 50.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 90000,
   "date": "2026-10-04",
   "flags": [],
   "hardware": "2x RTX 3080 20GB",
   "id": "lc-29935-2x3080-qwen35arch27b-f16-90k",
   "model": "qwen3.5-arch 27B",
   "notes": "perf issue #29935 (CUDA fattn KV streaming on sm86, context-decay report); local build 836d571 (= b11379 + 2 commits, fattn identical); 2x RTX 3080 20GB, qwen3.5-arch 27B NVFP4, -fa on (tensor split), MTP3 spec (verify batch 4), single request; official b11379 f16 48.2 @30K for reference; decode tok/s at 90K context, f16 KV; step-time slope ~0.075 ms/1K tok; config=f16-90k",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4 / f16 KV",
   "quote": "2x RTX 3080 20GB (sm86, `-sm tensor`), Windows, model = qwen3.5-arch 27B NVFP4; | local 836d571 | f16 | 50.6 | 44.0 | ~0.075 ms / 1K tok |",
   "ram_gb": null,
   "retrieved": "2026-10-04",
   "source_name": "llama.cpp issue #29935 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29935",
   "tg_tokens": null,
   "tps": 44.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 30000,
   "date": "2026-10-04",
   "flags": [],
   "hardware": "2x RTX 3080 20GB",
   "id": "lc-29935-2x3080-qwen35arch27b-q8-30k",
   "model": "qwen3.5-arch 27B",
   "notes": "perf issue #29935 (CUDA fattn KV streaming on sm86, context-decay report); local build 836d571 (= b11379 + 2 commits, fattn identical); 2x RTX 3080 20GB, qwen3.5-arch 27B NVFP4, -fa on (tensor split), MTP3 spec (verify batch 4), single request; official b11379 f16 48.2 @30K for reference; decode tok/s at 30K context, q8_0 KV; q8_0 KV takes the MMA path + whole-cache f16 conversion at MTP3 verify batch 4; config=q8-30k",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4 / q8_0 KV",
   "quote": "2x RTX 3080 20GB (sm86, `-sm tensor`), Windows, model = qwen3.5-arch 27B NVFP4; | local 836d571 | q8_0 | 48.7 | 39.4 | ~0.081 ms / 1K tok (conversion pass visible) |",
   "ram_gb": null,
   "retrieved": "2026-10-04",
   "source_name": "llama.cpp issue #29935 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29935",
   "tg_tokens": null,
   "tps": 48.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 90000,
   "date": "2026-10-04",
   "flags": [],
   "hardware": "2x RTX 3080 20GB",
   "id": "lc-29935-2x3080-qwen35arch27b-q8-90k",
   "model": "qwen3.5-arch 27B",
   "notes": "perf issue #29935 (CUDA fattn KV streaming on sm86, context-decay report); local build 836d571 (= b11379 + 2 commits, fattn identical); 2x RTX 3080 20GB, qwen3.5-arch 27B NVFP4, -fa on (tensor split), MTP3 spec (verify batch 4), single request; official b11379 f16 48.2 @30K for reference; decode tok/s at 90K context, q8_0 KV; step-time slope ~0.081 ms/1K tok (conversion pass visible), decays faster than f16; config=q8-90k",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4 / q8_0 KV",
   "quote": "2x RTX 3080 20GB (sm86, `-sm tensor`), Windows, model = qwen3.5-arch 27B NVFP4; | local 836d571 | q8_0 | 48.7 | 39.4 | ~0.081 ms / 1K tok (conversion pass visible) |",
   "ram_gb": null,
   "retrieved": "2026-10-04",
   "source_name": "llama.cpp issue #29935 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29935",
   "tg_tokens": null,
   "tps": 39.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10-04",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "hf5-qwen3827b-nvfp4-5090-default",
   "model": "Qwen3.8-27B",
   "notes": "akopytko Blackwell-native NVFP4 (no per-tensor scales, quant-time MSE scale search), runs on current llama.cpp, Blackwell GPUs (RTX 50 series/B200/B300); default flags -b 2048 -ub 512, tg128 decode / pp2048 prefill; unsloth NVFP4 reference on the same card 87.69 tg128 / 6018.71 pp2048; config=akopytko-nvfp4-5090-default",
   "params": "27B",
   "power_w": 575,
   "pp_tokens": 2048,
   "pp_tps": 6258.74,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Qwen3.8-27B NVFP4 GGUF Fast; For benchmarks, we used RTX 5090 running at stock 575 W; this NVFP4  6258.74 +/- 15.37  6432.25 +/- 143.73  87.45 +/- 0.27",
   "ram_gb": 32,
   "retrieved": "2026-10-04",
   "source_name": "HF model card: akopytko/Qwen3.8-27B-NVFP4-GGUF",
   "source_url": "https://huggingface.co/akopytko/Qwen3.8-27B-NVFP4-GGUF",
   "tg_tokens": 128,
   "tps": 87.45,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-04",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "hf5-qwen3827b-nvfp4-5090-mtp",
   "model": "Qwen3.8-27B",
   "notes": "akopytko NVFP4 with embedded MTP draft head, --spec-type draft-mtp --spec-draft-n-max 8 --spec-draft-p-min 0.8; 152 t/s decode at 0.75 draft acceptance on a typical prompt; config=akopytko-nvfp4-5090-mtp",
   "params": "27B",
   "power_w": 575,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Qwen3.8-27B NVFP4 GGUF Fast; For benchmarks, we used RTX 5090 running at stock 575 W; we observed 152 t/s decode at 0.75 draft acceptance on a typical prompt",
   "ram_gb": 32,
   "retrieved": "2026-10-04",
   "source_name": "HF model card: akopytko/Qwen3.8-27B-NVFP4-GGUF",
   "source_url": "https://huggingface.co/akopytko/Qwen3.8-27B-NVFP4-GGUF",
   "tg_tokens": null,
   "tps": 152,
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-04",
   "flags": [],
   "hardware": "B70 Arc Pro",
   "id": "lc-29936-b70arc-b11017-fa0",
   "model": "gemma4 26B.A4B",
   "notes": "PR #29936 (Vulkan Intel prefill regression fix, git-bisect shows #29182 halved Intel MoE prefill); B70 Arc Pro (Windows), gemma4 26B.A4B Q4_K_M, llama-bench -p 8192 -n 0; build b11017 (7490357f2), fa 0 (flash attn off); pp8192 prefill; config=b11017 (7490357f2)-fa0",
   "params": "25.23B",
   "power_w": null,
   "pp_tokens": 8192,
   "pp_tps": 3147.38,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "B70 Arc Pro on Windows; gemma4 26B.A4B Q4_K - Medium  15.70 GiB  25.23 B  Vulkan  -1  0  pp8192  3147.38 \u00b1 12.10",
   "ram_gb": null,
   "retrieved": "2026-10-04",
   "source_name": "llama.cpp #29936",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29936",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-04",
   "flags": [],
   "hardware": "B70 Arc Pro",
   "id": "lc-29936-b70arc-b11017-fa1",
   "model": "gemma4 26B.A4B",
   "notes": "PR #29936 (Vulkan Intel prefill regression fix, git-bisect shows #29182 halved Intel MoE prefill); B70 Arc Pro (Windows), gemma4 26B.A4B Q4_K_M, llama-bench -p 8192 -n 0; build b11017 (7490357f2), fa 1 (flash attn on); pp8192 prefill; config=b11017 (7490357f2)-fa1",
   "params": "25.23B",
   "power_w": null,
   "pp_tokens": 8192,
   "pp_tps": 1230.41,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "B70 Arc Pro on Windows; gemma4 26B.A4B Q4_K - Medium  15.70 GiB  25.23 B  Vulkan  -1  1  pp8192  1230.41 \u00b1 2.42",
   "ram_gb": null,
   "retrieved": "2026-10-04",
   "source_name": "llama.cpp #29936",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29936",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-04",
   "flags": [],
   "hardware": "B70 Arc Pro",
   "id": "lc-29936-b70arc-b11352-fa0",
   "model": "gemma4 26B.A4B",
   "notes": "PR #29936 (Vulkan Intel prefill regression fix, git-bisect shows #29182 halved Intel MoE prefill); B70 Arc Pro (Windows), gemma4 26B.A4B Q4_K_M, llama-bench -p 8192 -n 0; build b11352 (4e2713c16), fa 0 (flash attn off); pp8192 prefill; config=b11352 (4e2713c16)-fa0",
   "params": "25.23B",
   "power_w": null,
   "pp_tokens": 8192,
   "pp_tps": 1526.29,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "B70 Arc Pro on Windows; gemma4 26B.A4B Q4_K - Medium  15.70 GiB  25.23 B  Vulkan  -1  0  pp8192  1526.29 \u00b1 3.82",
   "ram_gb": null,
   "retrieved": "2026-10-04",
   "source_name": "llama.cpp #29936",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29936",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-04",
   "flags": [],
   "hardware": "B70 Arc Pro",
   "id": "lc-29936-b70arc-b11352-fa1",
   "model": "gemma4 26B.A4B",
   "notes": "PR #29936 (Vulkan Intel prefill regression fix, git-bisect shows #29182 halved Intel MoE prefill); B70 Arc Pro (Windows), gemma4 26B.A4B Q4_K_M, llama-bench -p 8192 -n 0; build b11352 (4e2713c16), fa 1 (flash attn on); pp8192 prefill; config=b11352 (4e2713c16)-fa1",
   "params": "25.23B",
   "power_w": null,
   "pp_tokens": 8192,
   "pp_tps": 879.92,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "B70 Arc Pro on Windows; gemma4 26B.A4B Q4_K - Medium  15.70 GiB  25.23 B  Vulkan  -1  1  pp8192  879.92 \u00b1 2.33",
   "ram_gb": null,
   "retrieved": "2026-10-04",
   "source_name": "llama.cpp #29936",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29936",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-04",
   "flags": [],
   "hardware": "7900 XTX",
   "id": "lc-29949-7900xtx-iq3xxs-nopool",
   "model": "Qwen3.8-Flash-Next",
   "notes": "config=iq3xxs-nopool; MoE expert cache GPU LRU, feature request with working impl",
   "params": "125B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ3_XXS",
   "quote": "7900 XTX, 24GB, single GPU; Qwen3.8-Flash-Next; no pool, plain 9.7",
   "ram_gb": 24,
   "retrieved": "2026-10-04",
   "source_name": "llama.cpp issue #29949 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29949",
   "tg_tokens": null,
   "tps": 9.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-04",
   "flags": [],
   "hardware": "7900 XTX",
   "id": "lc-29949-7900xtx-iq3xxs-nopool-mtp",
   "model": "Qwen3.8-Flash-Next",
   "notes": "config=iq3xxs-nopool-mtp; MoE expert cache GPU LRU, feature request with working impl",
   "params": "125B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ3_XXS",
   "quote": "7900 XTX, 24GB, single GPU; Qwen3.8-Flash-Next; no pool + MTP-Q8 11.3",
   "ram_gb": 24,
   "retrieved": "2026-10-04",
   "source_name": "llama.cpp issue #29949 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29949",
   "tg_tokens": null,
   "tps": 11.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-04",
   "flags": [],
   "hardware": "7900 XTX",
   "id": "lc-29949-7900xtx-iq3xxs-160s",
   "model": "Qwen3.8-Flash-Next",
   "notes": "config=iq3xxs-160s; MoE expert cache GPU LRU, feature request with working impl",
   "params": "125B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ3_XXS",
   "quote": "7900 XTX, 24GB, single GPU; Qwen3.8-Flash-Next; 160 slots, plain 37.2",
   "ram_gb": 24,
   "retrieved": "2026-10-04",
   "source_name": "llama.cpp issue #29949 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29949",
   "tg_tokens": null,
   "tps": 37.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-04",
   "flags": [],
   "hardware": "7900 XTX",
   "id": "lc-29949-7900xtx-iq3xxs-160s-mtp",
   "model": "Qwen3.8-Flash-Next",
   "notes": "config=iq3xxs-160s-mtp; MoE expert cache GPU LRU, feature request with working impl",
   "params": "125B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ3_XXS",
   "quote": "7900 XTX, 24GB, single GPU; Qwen3.8-Flash-Next; 160 slots + MTP-Q8 44.8",
   "ram_gb": 24,
   "retrieved": "2026-10-04",
   "source_name": "llama.cpp issue #29949 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29949",
   "tg_tokens": null,
   "tps": 44.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-04",
   "flags": [],
   "hardware": "7900 XTX",
   "id": "lc-29949-7900xtx-iq4xs-72s",
   "model": "Qwen3.8-Flash-Next",
   "notes": "config=iq4xs-72s; MoE expert cache GPU LRU, feature request with working impl",
   "params": "125B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "7900 XTX, 24GB, single GPU; Qwen3.8-Flash-Next; 72 slots, plain (6.4 GiB pool) 31.2",
   "ram_gb": 24,
   "retrieved": "2026-10-04",
   "source_name": "llama.cpp issue #29949 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29949",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-04",
   "flags": [],
   "hardware": "7900 XTX",
   "id": "lc-29949-7900xtx-iq4xs-40s",
   "model": "Qwen3.8-Flash-Next",
   "notes": "config=iq4xs-40s; MoE expert cache GPU LRU, feature request with working impl",
   "params": "125B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "7900 XTX, 24GB, single GPU; Qwen3.8-Flash-Next; 40 slots, plain (3.6 GiB pool) 28.8",
   "ram_gb": 24,
   "retrieved": "2026-10-04",
   "source_name": "llama.cpp issue #29949 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29949",
   "tg_tokens": null,
   "tps": 28.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-04",
   "flags": [],
   "hardware": "7900 XTX",
   "id": "lc-29949-7900xtx-iq4xs-40s-mtp",
   "model": "Qwen3.8-Flash-Next",
   "notes": "config=iq4xs-40s-mtp; MoE expert cache GPU LRU, feature request with working impl",
   "params": "125B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "7900 XTX, 24GB, single GPU; Qwen3.8-Flash-Next; 40 slots + MTP-Q4 30.5",
   "ram_gb": 24,
   "retrieved": "2026-10-04",
   "source_name": "llama.cpp issue #29949 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29949",
   "tg_tokens": null,
   "tps": 30.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-04",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc-29949-2x3090-iq4xs-plain",
   "model": "Qwen3.8-Flash-Next",
   "notes": "config=iq4xs-plain; MoE expert cache GPU LRU, feature request with working impl",
   "params": "125B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "2x RTX 3090, 48GB total; Qwen3.8-Flash-Next; EN, plain 47.6",
   "ram_gb": 48,
   "retrieved": "2026-10-04",
   "source_name": "llama.cpp issue #29949 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29949",
   "tg_tokens": null,
   "tps": 47.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-04",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc-29949-2x3090-iq4xs-mtp",
   "model": "Qwen3.8-Flash-Next",
   "notes": "config=iq4xs-mtp; MoE expert cache GPU LRU, feature request with working impl",
   "params": "125B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "2x RTX 3090, 48GB total; Qwen3.8-Flash-Next; EN + MTP-Q8 59.6",
   "ram_gb": 48,
   "retrieved": "2026-10-04",
   "source_name": "llama.cpp issue #29949 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29949",
   "tg_tokens": null,
   "tps": 59.6,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-04",
   "flags": [],
   "hardware": "GB10",
   "id": "vllm-59973-gb10-bf16-c1",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Qwen3.8-Flash-Next NVFP4 checkpoint + MTP k=3 drafter, one DGX Spark (GB10), measured on three nodes with SM121 skinny-GEMM plans from #59753; opt-in quantized draft lm_head (draft_lm_head_quantization); stock BF16 draft head, c=1 SPEED-Bench decode 39.4 tok/s; nvfp4 draft head 46.5 tok/s (+18%), average AL unchanged (-0.4%); config=bf16-c1",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Decode at c=1 on SPEED-Bench goes from 39.4 to 46.5 tok/s; Qwen3.8-Flash-Next-NVFP4 + MTP (k=3), one DGX Spark; Decode on GB10",
   "ram_gb": null,
   "retrieved": "2026-10-13",
   "source_name": "vLLM PR #59973",
   "source_url": "https://github.com/vllm-project/vllm/pull/59973",
   "tg_tokens": null,
   "tps": 39.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "H200 NVL",
   "id": "hf6-qui-linta13-iq2m-baseline",
   "model": "Qwen3.8-27B",
   "notes": "IQ2_M fused MTP, H200 NVL, -ngl 99, flash attention, F16 KV, 256 gen tokens, median of 3; spec-decode n_max swept 1-3; no spec-decode, prose prompt; config=iq2m-baseline",
   "params": "27B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ2_M",
   "quote": "H200 NVL; Qwen3.8-27B; prose none - 75.4 1.00x",
   "ram_gb": 141,
   "retrieved": "2026-10-05",
   "source_name": "HF model card: Qui-Linta13/Qwen3.8-27B-Uncensored-GGUF",
   "source_url": "https://huggingface.co/Qui-Linta13/Qwen3.8-27B-Uncensored-GGUF",
   "tg_tokens": 256,
   "tps": 75.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "H200 NVL",
   "id": "hf6-qui-linta13-iq2m-mtp-n2",
   "model": "Qwen3.8-27B",
   "notes": "IQ2_M fused MTP, H200 NVL, -ngl 99, flash attention, F16 KV, 256 gen tokens, median of 3; spec-decode n_max swept 1-3; MTP n_max 2, code prompt (best); config=iq2m-mtp-n2",
   "params": "27B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ2_M",
   "quote": "H200 NVL; Qwen3.8-27B; code draft-mtp 2 99.8 1.32x",
   "ram_gb": 141,
   "retrieved": "2026-10-05",
   "source_name": "HF model card: Qui-Linta13/Qwen3.8-27B-Uncensored-GGUF",
   "source_url": "https://huggingface.co/Qui-Linta13/Qwen3.8-27B-Uncensored-GGUF",
   "tg_tokens": 256,
   "tps": 99.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "H200 NVL",
   "id": "hf6-qui-linta13-iq2m-split-draft",
   "model": "Qwen3.8-27B",
   "notes": "IQ2_M fused MTP, H200 NVL, -ngl 99, flash attention, F16 KV, 256 gen tokens, median of 3; spec-decode n_max swept 1-3; split noMTP-IQ2_M + draft-Q8_0, n_max 2, prose; config=iq2m-split-draft",
   "params": "27B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ2_M",
   "quote": "H200 NVL; Qwen3.8-27B; noMTP-IQ2_M with the published draft-Q8_0 reaches 97.7 tok/s on the prose prompt",
   "ram_gb": 141,
   "retrieved": "2026-10-05",
   "source_name": "HF model card: Qui-Linta13/Qwen3.8-27B-Uncensored-GGUF",
   "source_url": "https://huggingface.co/Qui-Linta13/Qwen3.8-27B-Uncensored-GGUF",
   "tg_tokens": 256,
   "tps": 97.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "RTX A6000",
   "id": "hf6-gyro-a6000-gyro-s",
   "model": "Qwen3.8-Flash-Next",
   "notes": "llama-bench, batch size 1, no speculative decoding; config=a6000-gyro-s",
   "params": "126B",
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": 740,
   "provenance": "community",
   "quant": "Gyro-S",
   "quote": "Qwen3.8-Flash-Next; RTX A6000, 48 GB llama.cpp fork main, CUDA 57.4 740 (pp2048) 52.3",
   "ram_gb": 48,
   "retrieved": "2026-10-05",
   "source_name": "HF model card: agentionai/Qwen3.8-Flash-Next-Gyro-GGUF",
   "source_url": "https://huggingface.co/agentionai/Qwen3.8-Flash-Next-Gyro-GGUF",
   "tg_tokens": 128,
   "tps": 57.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "RTX A6000",
   "id": "hf6-gyro-a6000-gyro-m",
   "model": "Qwen3.8-Flash-Next",
   "notes": "llama-bench, batch size 1, no speculative decoding; Gyro-M decode; config=a6000-gyro-m",
   "params": "126B",
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Gyro-M",
   "quote": "Qwen3.8-Flash-Next; RTX A6000, 48 GB llama.cpp fork main, CUDA 57.4 740 (pp2048) 52.3",
   "ram_gb": 48,
   "retrieved": "2026-10-05",
   "source_name": "HF model card: agentionai/Qwen3.8-Flash-Next-Gyro-GGUF",
   "source_url": "https://huggingface.co/agentionai/Qwen3.8-Flash-Next-Gyro-GGUF",
   "tg_tokens": 128,
   "tps": 52.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-04",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "hf6-gyro-5090-cuda-decode",
   "model": "Qwen3.8-Flash-Next",
   "notes": "agentionai/llama.cpp main 2026-10-04 CUDA kernel update; config=cuda-kernel-decode",
   "params": "126B",
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": 3000,
   "provenance": "community",
   "quant": "Gyro-S",
   "quote": "Qwen3.8-Flash-Next; RTX 5090, 32 GB llama.cpp fork main, CUDA 113 (MTP: JSON 210, code 180, copy 218) ~3,000 (pp2048)",
   "ram_gb": 32,
   "retrieved": "2026-10-05",
   "source_name": "HF model card: agentionai/Qwen3.8-Flash-Next-Gyro-GGUF",
   "source_url": "https://huggingface.co/agentionai/Qwen3.8-Flash-Next-Gyro-GGUF",
   "tg_tokens": 128,
   "tps": 113,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-04",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "hf6-gyro-5090-cuda-mtp-json",
   "model": "Qwen3.8-Flash-Next",
   "notes": "agentionai/llama.cpp main 2026-10-04 CUDA kernel update, MTP JSON; config=cuda-kernel-mtp-json",
   "params": "126B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Gyro-S",
   "quote": "Qwen3.8-Flash-Next; RTX 5090, 32 GB llama.cpp fork main, CUDA 113 (MTP: JSON 210, code 180, copy 218) ~3,000 (pp2048)",
   "ram_gb": 32,
   "retrieved": "2026-10-05",
   "source_name": "HF model card: agentionai/Qwen3.8-Flash-Next-Gyro-GGUF",
   "source_url": "https://huggingface.co/agentionai/Qwen3.8-Flash-Next-Gyro-GGUF",
   "tg_tokens": 128,
   "tps": 210,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-04",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "hf6-gyro-5090-cuda-mtp-code",
   "model": "Qwen3.8-Flash-Next",
   "notes": "agentionai/llama.cpp main 2026-10-04 CUDA kernel update, MTP code; config=cuda-kernel-mtp-code",
   "params": "126B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Gyro-S",
   "quote": "Qwen3.8-Flash-Next; RTX 5090, 32 GB llama.cpp fork main, CUDA 113 (MTP: JSON 210, code 180, copy 218) ~3,000 (pp2048)",
   "ram_gb": 32,
   "retrieved": "2026-10-05",
   "source_name": "HF model card: agentionai/Qwen3.8-Flash-Next-Gyro-GGUF",
   "source_url": "https://huggingface.co/agentionai/Qwen3.8-Flash-Next-Gyro-GGUF",
   "tg_tokens": 128,
   "tps": 180,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-04",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "hf6-gyro-5090-cuda-mtp-copy",
   "model": "Qwen3.8-Flash-Next",
   "notes": "agentionai/llama.cpp main 2026-10-04 CUDA kernel update, MTP copied text; config=cuda-kernel-mtp-copy",
   "params": "126B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Gyro-S",
   "quote": "Qwen3.8-Flash-Next; RTX 5090, 32 GB llama.cpp fork main, CUDA 113 (MTP: JSON 210, code 180, copy 218) ~3,000 (pp2048)",
   "ram_gb": 32,
   "retrieved": "2026-10-05",
   "source_name": "HF model card: agentionai/Qwen3.8-Flash-Next-Gyro-GGUF",
   "source_url": "https://huggingface.co/agentionai/Qwen3.8-Flash-Next-Gyro-GGUF",
   "tg_tokens": 128,
   "tps": 218,
   "ttft_s": null
  },
  {
   "backend": "Strata",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "hf6-gyro-5090-strata-prose",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Strata rc1, all experts cached, greedy, MTP draft always on; config=strata-prose",
   "params": "126B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Gyro-S",
   "quote": "Qwen3.8-Flash-Next; RTX 5090, 32 GB Strata rc1 (release candidate) prose 140 , code 209, JSON 221 4,883 (16k prompt) prose 114",
   "ram_gb": 32,
   "retrieved": "2026-10-05",
   "source_name": "HF model card: agentionai/Qwen3.8-Flash-Next-Gyro-GGUF",
   "source_url": "https://huggingface.co/agentionai/Qwen3.8-Flash-Next-Gyro-GGUF",
   "tg_tokens": 128,
   "tps": 140,
   "ttft_s": null
  },
  {
   "backend": "Strata",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "hf6-gyro-5090-strata-json",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Strata rc1, all experts cached, greedy, MTP draft always on; config=strata-json",
   "params": "126B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Gyro-S",
   "quote": "Qwen3.8-Flash-Next; RTX 5090, 32 GB Strata rc1 (release candidate) prose 140 , code 209, JSON 221 4,883 (16k prompt) prose 114",
   "ram_gb": 32,
   "retrieved": "2026-10-05",
   "source_name": "HF model card: agentionai/Qwen3.8-Flash-Next-Gyro-GGUF",
   "source_url": "https://huggingface.co/agentionai/Qwen3.8-Flash-Next-Gyro-GGUF",
   "tg_tokens": 128,
   "tps": 221,
   "ttft_s": null
  },
  {
   "backend": "Strata",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "hf6-gyro-5090-strata-code",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Strata rc1, all experts cached, greedy, MTP draft always on; config=strata-code",
   "params": "126B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Gyro-S",
   "quote": "Qwen3.8-Flash-Next; RTX 5090, 32 GB Strata rc1 (release candidate) prose 140 , code 209, JSON 221 4,883 (16k prompt) prose 114",
   "ram_gb": 32,
   "retrieved": "2026-10-05",
   "source_name": "HF model card: agentionai/Qwen3.8-Flash-Next-Gyro-GGUF",
   "source_url": "https://huggingface.co/agentionai/Qwen3.8-Flash-Next-Gyro-GGUF",
   "tg_tokens": 128,
   "tps": 209,
   "ttft_s": null
  },
  {
   "backend": "Strata",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "hf6-gyro-5090-strata-prefill",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Strata rc1, all experts cached, greedy, MTP draft always on; prefill 16k prompt; config=strata-prefill",
   "params": "126B",
   "power_w": null,
   "pp_tokens": 16384,
   "pp_tps": 4883,
   "provenance": "community",
   "quant": "Gyro-S",
   "quote": "Qwen3.8-Flash-Next; RTX 5090, 32 GB Strata rc1 (release candidate) prose 140 , code 209, JSON 221 4,883 (16k prompt) prose 114",
   "ram_gb": 32,
   "retrieved": "2026-10-05",
   "source_name": "HF model card: agentionai/Qwen3.8-Flash-Next-Gyro-GGUF",
   "source_url": "https://huggingface.co/agentionai/Qwen3.8-Flash-Next-Gyro-GGUF",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "4x GB200",
   "id": "vllm-60008-4xgb200-pcoff-c1",
   "model": "Nemotron-3.5-Lightning",
   "notes": "Hybrid Mamba prefix caching align mode, Nemotron-3.5-Lightning NVFP4, 4x GB200 DP4/EP4, 8K-in/1K-out, median of 6 paired rounds; PC off, c=1, total throughput; config=pcoff-c1",
   "params": "30B-A3B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "4x GB200; Nemotron-3.5-Lightning; 317 [",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "vLLM issue #60008",
   "source_url": "https://github.com/vllm-project/vllm/issues/60008",
   "tg_tokens": 1024,
   "tps": 317,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "4x GB200",
   "id": "vllm-60008-4xgb200-pcoff-c8",
   "model": "Nemotron-3.5-Lightning",
   "notes": "Hybrid Mamba prefix caching align mode, Nemotron-3.5-Lightning NVFP4, 4x GB200 DP4/EP4, 8K-in/1K-out, median of 6 paired rounds; PC off, c=8, total throughput; config=pcoff-c8",
   "params": "30B-A3B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "4x GB200; Nemotron-3.5-Lightning; 2178 [",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "vLLM issue #60008",
   "source_url": "https://github.com/vllm-project/vllm/issues/60008",
   "tg_tokens": 1024,
   "tps": 2178,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 32,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "4x GB200",
   "id": "vllm-60008-4xgb200-pcoff-c32",
   "model": "Nemotron-3.5-Lightning",
   "notes": "Hybrid Mamba prefix caching align mode, Nemotron-3.5-Lightning NVFP4, 4x GB200 DP4/EP4, 8K-in/1K-out, median of 6 paired rounds; PC off, c=32, total throughput; config=pcoff-c32",
   "params": "30B-A3B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "4x GB200; Nemotron-3.5-Lightning; 7174 [",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "vLLM issue #60008",
   "source_url": "https://github.com/vllm-project/vllm/issues/60008",
   "tg_tokens": 1024,
   "tps": 7174,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 64,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "4x GB200",
   "id": "vllm-60008-4xgb200-pcoff-c64",
   "model": "Nemotron-3.5-Lightning",
   "notes": "Hybrid Mamba prefix caching align mode, Nemotron-3.5-Lightning NVFP4, 4x GB200 DP4/EP4, 8K-in/1K-out, median of 6 paired rounds; PC off, c=64, total throughput; config=pcoff-c64",
   "params": "30B-A3B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "4x GB200; Nemotron-3.5-Lightning; 10371 [",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "vLLM issue #60008",
   "source_url": "https://github.com/vllm-project/vllm/issues/60008",
   "tg_tokens": 1024,
   "tps": 10371,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "4x GB200",
   "id": "vllm-60008-4xgb200-pcon-c1",
   "model": "Nemotron-3.5-Lightning",
   "notes": "Hybrid Mamba prefix caching align mode, Nemotron-3.5-Lightning NVFP4, 4x GB200 DP4/EP4, 8K-in/1K-out, median of 6 paired rounds; PC on, c=1, total throughput; config=pcon-c1",
   "params": "30B-A3B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "4x GB200; Nemotron-3.5-Lightning; 277 [",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "vLLM issue #60008",
   "source_url": "https://github.com/vllm-project/vllm/issues/60008",
   "tg_tokens": 1024,
   "tps": 277,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "4x GB200",
   "id": "vllm-60008-4xgb200-pcon-c8",
   "model": "Nemotron-3.5-Lightning",
   "notes": "Hybrid Mamba prefix caching align mode, Nemotron-3.5-Lightning NVFP4, 4x GB200 DP4/EP4, 8K-in/1K-out, median of 6 paired rounds; PC on, c=8, total throughput; config=pcon-c8",
   "params": "30B-A3B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "4x GB200; Nemotron-3.5-Lightning; 1879 [",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "vLLM issue #60008",
   "source_url": "https://github.com/vllm-project/vllm/issues/60008",
   "tg_tokens": 1024,
   "tps": 1879,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 32,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "4x GB200",
   "id": "vllm-60008-4xgb200-pcon-c32",
   "model": "Nemotron-3.5-Lightning",
   "notes": "Hybrid Mamba prefix caching align mode, Nemotron-3.5-Lightning NVFP4, 4x GB200 DP4/EP4, 8K-in/1K-out, median of 6 paired rounds; PC on, c=32, total throughput; config=pcon-c32",
   "params": "30B-A3B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "4x GB200; Nemotron-3.5-Lightning; 5476 [",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "vLLM issue #60008",
   "source_url": "https://github.com/vllm-project/vllm/issues/60008",
   "tg_tokens": 1024,
   "tps": 5476,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 64,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "4x GB200",
   "id": "vllm-60008-4xgb200-pcon-c64",
   "model": "Nemotron-3.5-Lightning",
   "notes": "Hybrid Mamba prefix caching align mode, Nemotron-3.5-Lightning NVFP4, 4x GB200 DP4/EP4, 8K-in/1K-out, median of 6 paired rounds; PC on, c=64, total throughput; config=pcon-c64",
   "params": "30B-A3B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "4x GB200; Nemotron-3.5-Lightning; 9109 [",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "vLLM issue #60008",
   "source_url": "https://github.com/vllm-project/vllm/issues/60008",
   "tg_tokens": 1024,
   "tps": 9109,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Dimensity 9400",
   "id": "lc-30006-d9400-cpu",
   "model": "Qwen3.5-0.8B",
   "notes": "Mali-G925 coopmat prefill pessimization (Dimensity 9400, Mali-G925-Immortalis MC12, Android 16 Termux, llama.cpp 6c59c40), Qwen3.5-0.8B Q4_K_M, llama-bench -ngl 99 -t 4 -p 512 -n 32 -r 2; CPU arm (-ngl 0, 4 threads); prefill flat vs GPU, decode 12.6x slower",
   "params": "752M",
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 51.19,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Dimensity 9400; Qwen3.5-0.8B; -ngl 0` (CPU, 4 threads) | 51.19 \u00b1 0.09 | 2.95 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30006",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30006",
   "tg_tokens": 32,
   "tps": 2.95,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Dimensity 9400",
   "id": "lc-30006-d9400-coopmat-on",
   "model": "Qwen3.5-0.8B",
   "notes": "Mali-G925 coopmat prefill pessimization (Dimensity 9400, Mali-G925-Immortalis MC12, Android 16 Termux, llama.cpp 6c59c40), Qwen3.5-0.8B Q4_K_M, llama-bench -ngl 99 -t 4 -p 512 -n 32 -r 2; Vulkan default (coopmat on); config=coopmat-on",
   "params": "752M",
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 53.3,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Dimensity 9400; Qwen3.5-0.8B; none) | **53.30** | 36.40 | KHR_coopmat",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30006",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30006",
   "tg_tokens": 32,
   "tps": 36.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Dimensity 9400",
   "id": "lc-30006-d9400-coopmat-off",
   "model": "Qwen3.5-0.8B",
   "notes": "Mali-G925 coopmat prefill pessimization (Dimensity 9400, Mali-G925-Immortalis MC12, Android 16 Termux, llama.cpp 6c59c40), Qwen3.5-0.8B Q4_K_M, llama-bench -ngl 99 -t 4 -p 512 -n 32 -r 2; GGML_VK_DISABLE_COOPMAT=1; prefill 2.36x faster, decode unchanged; config=coopmat-off",
   "params": "752M",
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 125.78,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Dimensity 9400; Qwen3.5-0.8B; GGML_VK_DISABLE_COOPMAT=1** | **125.78** | 36.62",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30006",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30006",
   "tg_tokens": 32,
   "tps": 36.62,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Dimensity 9400",
   "id": "lc-30006-d9400-coopmat-off-nof16",
   "model": "Qwen3.5-0.8B",
   "notes": "Mali-G925 coopmat prefill pessimization (Dimensity 9400, Mali-G925-Immortalis MC12, Android 16 Termux, llama.cpp 6c59c40), Qwen3.5-0.8B Q4_K_M, llama-bench -ngl 99 -t 4 -p 512 -n 32 -r 2; coopmat off + f16 off, slightly worse than coopmat off alone; config=coopmat-off-f16-off",
   "params": "752M",
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 117.27,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Dimensity 9400; Qwen3.5-0.8B; GGML_VK_DISABLE_COOPMAT=1` + `GGML_VK_DISABLE_F16=1** | 117.27 | 35.80",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30006",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30006",
   "tg_tokens": 32,
   "tps": 35.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Dimensity 9400",
   "id": "lc-30006-d9400-intdot-off",
   "model": "Qwen3.5-0.8B",
   "notes": "Mali-G925 coopmat prefill pessimization (Dimensity 9400, Mali-G925-Immortalis MC12, Android 16 Termux, llama.cpp 6c59c40), Qwen3.5-0.8B Q4_K_M, llama-bench -ngl 99 -t 4 -p 512 -n 32 -r 2; integer dot product disabled, no effect; config=intdot-off",
   "params": "752M",
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 54.53,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Dimensity 9400; Qwen3.5-0.8B; GGML_VK_DISABLE_INTEGER_DOT_PRODUCT=1** | 54.53 | 36.03",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30006",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30006",
   "tg_tokens": 32,
   "tps": 36.03,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Dimensity 9400",
   "id": "lc-30006-d9400-hostmem",
   "model": "Qwen3.5-0.8B",
   "notes": "Mali-G925 coopmat prefill pessimization (Dimensity 9400, Mali-G925-Immortalis MC12, Android 16 Termux, llama.cpp 6c59c40), Qwen3.5-0.8B Q4_K_M, llama-bench -ngl 99 -t 4 -p 512 -n 32 -r 2; prefer host memory, no effect; config=hostmem",
   "params": "752M",
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 54.04,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Dimensity 9400; Qwen3.5-0.8B; GGML_VK_PREFER_HOST_MEMORY=1** | 54.04 | 36.88",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30006",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30006",
   "tg_tokens": 32,
   "tps": 36.88,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-30000-5060ti-qwen3emb8b-b6d9c82",
   "model": "Qwen3-Embedding-8B",
   "notes": "Vulkan Q8_0 prefill regression from #25773 (91f6a6cf3), RTX 5060 Ti 16 GB (driver 617.14), i9-14900K, Windows, llama-bench -dev Vulkan0 -sm none -ngl 99 -fa 1 -ub 512 -b 512 -r 5 -p 512; bisect pair (6d9c82ea2 vs 91f6a6cf3); build 6d9c82ea2 (before #25773); config=6d9c82ea2",
   "params": "8B",
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2798,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "RTX 5060 Ti; Qwen3-Embedding-8B; Qwen3-Embedding-8B Q8_0 | 2,798 | 2,348",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30000",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30000",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-30000-5060ti-qwen3emb8b-b91f6a6",
   "model": "Qwen3-Embedding-8B",
   "notes": "Vulkan Q8_0 prefill regression from #25773 (91f6a6cf3), RTX 5060 Ti 16 GB (driver 617.14), i9-14900K, Windows, llama-bench -dev Vulkan0 -sm none -ngl 99 -fa 1 -ub 512 -b 512 -r 5 -p 512; bisect pair (6d9c82ea2 vs 91f6a6cf3); build 91f6a6cf3 (#25773, first bad commit); config=91f6a6cf3",
   "params": "8B",
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2348,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "RTX 5060 Ti; Qwen3-Embedding-8B; Qwen3-Embedding-8B Q8_0 | 2,798 | 2,348",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30000",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30000",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-30000-5060ti-qwen3rerank8b-b6d9c82",
   "model": "Qwen3-Reranker-8B",
   "notes": "Vulkan Q8_0 prefill regression from #25773 (91f6a6cf3), RTX 5060 Ti 16 GB (driver 617.14), i9-14900K, Windows, llama-bench -dev Vulkan0 -sm none -ngl 99 -fa 1 -ub 512 -b 512 -r 5 -p 512; bisect pair (6d9c82ea2 vs 91f6a6cf3); build 6d9c82ea2 (before #25773); config=6d9c82ea2",
   "params": "8B",
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2790,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "RTX 5060 Ti; Qwen3-Reranker-8B; Qwen3-Reranker-8B Q8_0 | 2,790 | 2,337",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30000",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30000",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-30000-5060ti-qwen3rerank8b-b91f6a6",
   "model": "Qwen3-Reranker-8B",
   "notes": "Vulkan Q8_0 prefill regression from #25773 (91f6a6cf3), RTX 5060 Ti 16 GB (driver 617.14), i9-14900K, Windows, llama-bench -dev Vulkan0 -sm none -ngl 99 -fa 1 -ub 512 -b 512 -r 5 -p 512; bisect pair (6d9c82ea2 vs 91f6a6cf3); build 91f6a6cf3 (#25773, first bad commit); config=91f6a6cf3",
   "params": "8B",
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2337,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "RTX 5060 Ti; Qwen3-Reranker-8B; Qwen3-Reranker-8B Q8_0 | 2,790 | 2,337",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30000",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30000",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-30000-5060ti-granite428b-b6d9c82",
   "model": "granite-4.2-8b",
   "notes": "Vulkan Q8_0 prefill regression from #25773 (91f6a6cf3), RTX 5060 Ti 16 GB (driver 617.14), i9-14900K, Windows, llama-bench -dev Vulkan0 -sm none -ngl 99 -fa 1 -ub 512 -b 512 -r 5 -p 512; bisect pair (6d9c82ea2 vs 91f6a6cf3); build 6d9c82ea2 (before #25773); config=6d9c82ea2",
   "params": "8B",
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2955,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "RTX 5060 Ti; granite-4.2-8b; granite-4.2-8b Q8_0 | 2,955 | 2,437",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30000",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30000",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-30000-5060ti-granite428b-b91f6a6",
   "model": "granite-4.2-8b",
   "notes": "Vulkan Q8_0 prefill regression from #25773 (91f6a6cf3), RTX 5060 Ti 16 GB (driver 617.14), i9-14900K, Windows, llama-bench -dev Vulkan0 -sm none -ngl 99 -fa 1 -ub 512 -b 512 -r 5 -p 512; bisect pair (6d9c82ea2 vs 91f6a6cf3); build 91f6a6cf3 (#25773, first bad commit); config=91f6a6cf3",
   "params": "8B",
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2437,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "RTX 5060 Ti; granite-4.2-8b; granite-4.2-8b Q8_0 | 2,955 | 2,437",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30000",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30000",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-30000-5060ti-granite428b-tg-pre",
   "model": "granite-4.2-8b",
   "notes": "Vulkan Q8_0 prefill regression from #25773 (91f6a6cf3), RTX 5060 Ti 16 GB (driver 617.14), i9-14900K, Windows, llama-bench -dev Vulkan0 -sm none -ngl 99 -fa 1 -ub 512 -b 512 -r 5 -p 512; bisect pair (6d9c82ea2 vs 91f6a6cf3); tg128 decode, build 6d9c82ea2, unchanged by #25773; config=tg-6d9c82ea2",
   "params": "8B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "RTX 5060 Ti; granite-4.2-8b; granite tg128: 41.9 \u2192 42.1 t/s",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30000",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30000",
   "tg_tokens": 128,
   "tps": 41.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-30000-5060ti-granite428b-tg-post",
   "model": "granite-4.2-8b",
   "notes": "Vulkan Q8_0 prefill regression from #25773 (91f6a6cf3), RTX 5060 Ti 16 GB (driver 617.14), i9-14900K, Windows, llama-bench -dev Vulkan0 -sm none -ngl 99 -fa 1 -ub 512 -b 512 -r 5 -p 512; bisect pair (6d9c82ea2 vs 91f6a6cf3); tg128 decode, build 91f6a6cf3, unchanged by #25773; config=tg-91f6a6cf3",
   "params": "8B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "RTX 5060 Ti; granite-4.2-8b; granite tg128: 41.9 \u2192 42.1 t/s",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30000",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30000",
   "tg_tokens": 128,
   "tps": 42.1,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "4x B200",
   "id": "vllm-60070-4xb200-baseline",
   "model": "GLM-5.2-NVFP4",
   "notes": "vLLM PR #60070 persistent FlashInfer multi-CTA-KV counter buffer in sparse MLA decode, GLM-5.2-NVFP4 on 4x B200 TP4, BS=1, ISL=4, OSL=2048, fp8_e4m3 KV cache, like-with-like on the same FlashInfer build; baseline, VLLM_DISABLE_PERSISTENT_MLA_COUNTER=1; counterpart patched 132.5 tok/s; config=baseline",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "4x B200; GLM-5.2-NVFP4; Output throughput (tok/s) 130.8 132.5",
   "ram_gb": null,
   "retrieved": "2026-10-12",
   "source_name": "vLLM PR #60070",
   "source_url": "https://github.com/vllm-project/vllm/pull/60070",
   "tg_tokens": 2048,
   "tps": 130.8,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "4x B200",
   "id": "vllm-60070-4xb200-patched",
   "model": "GLM-5.2-NVFP4",
   "notes": "vLLM PR #60070 persistent FlashInfer multi-CTA-KV counter buffer in sparse MLA decode, GLM-5.2-NVFP4 on 4x B200 TP4, BS=1, ISL=4, OSL=2048, fp8_e4m3 KV cache, like-with-like on the same FlashInfer build; patched, persistent counter buffer, +1.35%; counterpart baseline 130.8 tok/s; config=patched",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "4x B200; GLM-5.2-NVFP4; Output throughput (tok/s) 130.8 132.5",
   "ram_gb": null,
   "retrieved": "2026-10-12",
   "source_name": "vLLM PR #60070",
   "source_url": "https://github.com/vllm-project/vllm/pull/60070",
   "tg_tokens": 2048,
   "tps": 132.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "hf-tielcoder-8060s-stock-dflash",
   "model": "Tiel-Coder",
   "notes": "Tiel-Coder 35B-A3B on Radeon 8060S (gfx1151, 96 GiB unified, Ryzen AI Max+ 395), Debian 13 LXC, ROCm 10.0, temperature 0; effective generation speed over a 17-task code battery (tokens/seconds, prompt time included, passing tasks only); stock llama.cpp + DFlash draft, UD-Q4_K_XL; config=stock-dflash",
   "params": "35B-A3B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Radeon 8060S; Tiel-Coder; passing tasks only) | 69.0 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "HF card ramon-balaguer/Tiel-Coder-35B-A3B-ROCmFP4-GGUF",
   "source_url": "https://huggingface.co/ramon-balaguer/Tiel-Coder-35B-A3B-ROCmFP4-GGUF",
   "tg_tokens": null,
   "tps": 69.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "hf-tielcoder-8060s-fpx-dflash",
   "model": "Tiel-Coder",
   "notes": "Tiel-Coder 35B-A3B on Radeon 8060S (gfx1151, 96 GiB unified, Ryzen AI Max+ 395), Debian 13 LXC, ROCm 10.0, temperature 0; effective generation speed over a 17-task code battery (tokens/seconds, prompt time included, passing tasks only); ROCmFPX engine + DFlash n4 draft, ROCmFP4 requant (17.3 GiB vs 20.8 GiB); config=fpx-dflash-n4",
   "params": "35B-A3B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "ROCmFP4",
   "quote": "Radeon 8060S; Tiel-Coder; **86.6 tok/s** (+25 %) with DFlash",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "HF card ramon-balaguer/Tiel-Coder-35B-A3B-ROCmFP4-GGUF",
   "source_url": "https://huggingface.co/ramon-balaguer/Tiel-Coder-35B-A3B-ROCmFP4-GGUF",
   "tg_tokens": null,
   "tps": 86.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "hf-tielcoder-8060s-fpx-mtp",
   "model": "Tiel-Coder",
   "notes": "Tiel-Coder 35B-A3B on Radeon 8060S (gfx1151, 96 GiB unified, Ryzen AI Max+ 395), Debian 13 LXC, ROCm 10.0, temperature 0; effective generation speed over a 17-task code battery (tokens/seconds, prompt time included, passing tasks only); ROCmFPX engine + native MTP head n2, draft acceptance 71-87 percent; config=fpx-mtp-n2",
   "params": "35B-A3B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "ROCmFP4",
   "quote": "Radeon 8060S; Tiel-Coder; 77.4 tok/s with the Tiel MTP head",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "HF card ramon-balaguer/Tiel-Coder-35B-A3B-ROCmFP4-GGUF",
   "source_url": "https://huggingface.co/ramon-balaguer/Tiel-Coder-35B-A3B-ROCmFP4-GGUF",
   "tg_tokens": null,
   "tps": 77.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "hf-tielcoder-8060s-fpx-decode",
   "model": "Tiel-Coder",
   "notes": "Tiel-Coder 35B-A3B on Radeon 8060S (gfx1151, 96 GiB unified, Ryzen AI Max+ 395), Debian 13 LXC, ROCm 10.0, temperature 0; effective generation speed over a 17-task code battery (tokens/seconds, prompt time included, passing tasks only); plain decode on a short code prompt via llama-swap, published as a range 117-125 tok/s (stored low end); config=fpx-decode",
   "params": "35B-A3B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "ROCmFP4",
   "quote": "Radeon 8060S; Tiel-Coder; 117\u2013125 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "HF card ramon-balaguer/Tiel-Coder-35B-A3B-ROCmFP4-GGUF",
   "source_url": "https://huggingface.co/ramon-balaguer/Tiel-Coder-35B-A3B-ROCmFP4-GGUF",
   "tg_tokens": null,
   "tps": 117,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "48 GB GPU",
   "id": "hf-volundr-48gb-ctx1k",
   "model": "Agens-Volundr-32B-Preview",
   "notes": "Agens-Volundr-32B-Preview Q4_K_M on one 48 GB GPU, llama.cpp -ngl 99, flash attention, F16 KV; 1K context decode 32.3 tok/s, 1K-prompt prefill 2218 tok/s; config=ctx1k",
   "params": "32B",
   "power_w": null,
   "pp_tokens": 1024,
   "pp_tps": 2218,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "48 GB GPU; Agens-Volundr-32B-Preview; decode | 32.3 tok/s; Prompt processing of a 1K-token prompt: 2,218 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "HF card Blockway/Agens-Volundr-32B-Preview-GGUF",
   "source_url": "https://huggingface.co/Blockway/Agens-Volundr-32B-Preview-GGUF",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "48 GB GPU",
   "id": "hf-volundr-48gb-ctx8k",
   "model": "Agens-Volundr-32B-Preview",
   "notes": "Agens-Volundr-32B-Preview Q4_K_M on one 48 GB GPU, llama.cpp -ngl 99, flash attention, F16 KV; 8K context decode; config=ctx8k",
   "params": "32B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "48 GB GPU; Agens-Volundr-32B-Preview; 27.4 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "HF card Blockway/Agens-Volundr-32B-Preview-GGUF",
   "source_url": "https://huggingface.co/Blockway/Agens-Volundr-32B-Preview-GGUF",
   "tg_tokens": null,
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "48 GB GPU",
   "id": "hf-volundr-48gb-ctx32k",
   "model": "Agens-Volundr-32B-Preview",
   "notes": "Agens-Volundr-32B-Preview Q4_K_M on one 48 GB GPU, llama.cpp -ngl 99, flash attention, F16 KV; 32K context decode 25.0 tok/s, 32K-prompt prefill about 620 tok/s (block selection slows with prompt length); config=ctx32k",
   "params": "32B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 620,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "48 GB GPU; Agens-Volundr-32B-Preview; 25.0 tok/s; about 620 tok/s for a 32K-token prompt",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "HF card Blockway/Agens-Volundr-32B-Preview-GGUF",
   "source_url": "https://huggingface.co/Blockway/Agens-Volundr-32B-Preview-GGUF",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "ranma.cpp (ROCm)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "hf-ranma-r9700-128gb-bench-305",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Qwen3.8-Flash-Next RANMA EXL3 (ranma.cpp ranma_20261005, MoE experts in host RAM over PCIe), llama-bench TG128 at depth 0, warm expert cache, exclusive mode, measured once; 128 GB host RAM tier; config=r128-bench-305",
   "params": "80B-A3B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL3 3.05 bpw",
   "quote": "Radeon AI PRO R9700; Qwen3.8-Flash-Next; R9700 + 128 GB RAM | 53.5 | 49.6",
   "ram_gb": "128",
   "retrieved": "2026-10-05",
   "source_name": "HF card SaturnHeaven/Qwen3.8-Flash-Next-RANMA-EXL3-GGUF",
   "source_url": "https://huggingface.co/SaturnHeaven/Qwen3.8-Flash-Next-RANMA-EXL3-GGUF",
   "tg_tokens": "128",
   "tps": 53.5,
   "ttft_s": null
  },
  {
   "backend": "ranma.cpp (ROCm)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "hf-ranma-r9700-128gb-bench-405",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Qwen3.8-Flash-Next RANMA EXL3 (ranma.cpp ranma_20261005, MoE experts in host RAM over PCIe), llama-bench TG128 at depth 0, warm expert cache, exclusive mode, measured once; 128 GB host RAM tier; config=r128-bench-405",
   "params": "80B-A3B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL3 4.05 bpw",
   "quote": "Radeon AI PRO R9700; Qwen3.8-Flash-Next; R9700 + 128 GB RAM | 53.5 | 49.6",
   "ram_gb": "128",
   "retrieved": "2026-10-05",
   "source_name": "HF card SaturnHeaven/Qwen3.8-Flash-Next-RANMA-EXL3-GGUF",
   "source_url": "https://huggingface.co/SaturnHeaven/Qwen3.8-Flash-Next-RANMA-EXL3-GGUF",
   "tg_tokens": "128",
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "ranma.cpp (ROCm)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "hf-ranma-r9700-64gb-bench-305",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Qwen3.8-Flash-Next RANMA EXL3 (ranma.cpp ranma_20261005, MoE experts in host RAM over PCIe), llama-bench TG128 at depth 0, warm expert cache, exclusive mode, measured once; 64 GB host RAM tier; config=r64-bench-305",
   "params": "80B-A3B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL3 3.05 bpw",
   "quote": "Radeon AI PRO R9700; Qwen3.8-Flash-Next; R9700 + 64 GB RAM | 52.4 | 47.7",
   "ram_gb": "64",
   "retrieved": "2026-10-05",
   "source_name": "HF card SaturnHeaven/Qwen3.8-Flash-Next-RANMA-EXL3-GGUF",
   "source_url": "https://huggingface.co/SaturnHeaven/Qwen3.8-Flash-Next-RANMA-EXL3-GGUF",
   "tg_tokens": "128",
   "tps": 52.4,
   "ttft_s": null
  },
  {
   "backend": "ranma.cpp (ROCm)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "hf-ranma-r9700-64gb-bench-405",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Qwen3.8-Flash-Next RANMA EXL3 (ranma.cpp ranma_20261005, MoE experts in host RAM over PCIe), llama-bench TG128 at depth 0, warm expert cache, exclusive mode, measured once; 64 GB host RAM tier; config=r64-bench-405",
   "params": "80B-A3B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL3 4.05 bpw",
   "quote": "Radeon AI PRO R9700; Qwen3.8-Flash-Next; R9700 + 64 GB RAM | 52.4 | 47.7",
   "ram_gb": "64",
   "retrieved": "2026-10-05",
   "source_name": "HF card SaturnHeaven/Qwen3.8-Flash-Next-RANMA-EXL3-GGUF",
   "source_url": "https://huggingface.co/SaturnHeaven/Qwen3.8-Flash-Next-RANMA-EXL3-GGUF",
   "tg_tokens": "128",
   "tps": 47.7,
   "ttft_s": null
  },
  {
   "backend": "ranma.cpp (ROCm)",
   "batch": 1,
   "ctx": "65536",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "hf-ranma-r9700-128gb-mtp-305",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Qwen3.8-Flash-Next RANMA EXL3 (ranma.cpp ranma_20261005), MTP smart decoding via server (draft head mtp-Qwen3.8-Flash-Next-Q8_0.gguf), context 65536, greedy, at most 384 generated tokens, published as a range over coding and four roleplay presets (stored low end); 128 GB host RAM tier, range 60.8-66.3 tok/s; config=r128-mtp-305",
   "params": "80B-A3B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL3 3.05 bpw",
   "quote": "Radeon AI PRO R9700; Qwen3.8-Flash-Next; 60.8 - 66.3",
   "ram_gb": "128",
   "retrieved": "2026-10-05",
   "source_name": "HF card SaturnHeaven/Qwen3.8-Flash-Next-RANMA-EXL3-GGUF",
   "source_url": "https://huggingface.co/SaturnHeaven/Qwen3.8-Flash-Next-RANMA-EXL3-GGUF",
   "tg_tokens": "384",
   "tps": 60.8,
   "ttft_s": null
  },
  {
   "backend": "ranma.cpp (ROCm)",
   "batch": 1,
   "ctx": "65536",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "hf-ranma-r9700-128gb-mtp-405",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Qwen3.8-Flash-Next RANMA EXL3 (ranma.cpp ranma_20261005), MTP smart decoding via server (draft head mtp-Qwen3.8-Flash-Next-Q8_0.gguf), context 65536, greedy, at most 384 generated tokens, published as a range over coding and four roleplay presets (stored low end); 128 GB host RAM tier, range 53.9-57.0 tok/s; config=r128-mtp-405",
   "params": "80B-A3B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL3 4.05 bpw",
   "quote": "Radeon AI PRO R9700; Qwen3.8-Flash-Next; 53.9 - 57.0",
   "ram_gb": "128",
   "retrieved": "2026-10-05",
   "source_name": "HF card SaturnHeaven/Qwen3.8-Flash-Next-RANMA-EXL3-GGUF",
   "source_url": "https://huggingface.co/SaturnHeaven/Qwen3.8-Flash-Next-RANMA-EXL3-GGUF",
   "tg_tokens": "384",
   "tps": 53.9,
   "ttft_s": null
  },
  {
   "backend": "ranma.cpp (ROCm)",
   "batch": 1,
   "ctx": "65536",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "hf-ranma-r9700-64gb-mtp-305",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Qwen3.8-Flash-Next RANMA EXL3 (ranma.cpp ranma_20261005), MTP smart decoding via server (draft head mtp-Qwen3.8-Flash-Next-Q8_0.gguf), context 65536, greedy, at most 384 generated tokens, published as a range over coding and four roleplay presets (stored low end); 64 GB host RAM tier, range 58.8-66.4 tok/s; config=r64-mtp-305",
   "params": "80B-A3B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL3 3.05 bpw",
   "quote": "Radeon AI PRO R9700; Qwen3.8-Flash-Next; 58.8 - 66.4",
   "ram_gb": "64",
   "retrieved": "2026-10-05",
   "source_name": "HF card SaturnHeaven/Qwen3.8-Flash-Next-RANMA-EXL3-GGUF",
   "source_url": "https://huggingface.co/SaturnHeaven/Qwen3.8-Flash-Next-RANMA-EXL3-GGUF",
   "tg_tokens": "384",
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "ranma.cpp (ROCm)",
   "batch": 1,
   "ctx": "65536",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "hf-ranma-r9700-64gb-mtp-405",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Qwen3.8-Flash-Next RANMA EXL3 (ranma.cpp ranma_20261005), MTP smart decoding via server (draft head mtp-Qwen3.8-Flash-Next-Q8_0.gguf), context 65536, greedy, at most 384 generated tokens, published as a range over coding and four roleplay presets (stored low end); 64 GB host RAM tier, range 53.3-57.3 tok/s; config=r64-mtp-405",
   "params": "80B-A3B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL3 4.05 bpw",
   "quote": "Radeon AI PRO R9700; Qwen3.8-Flash-Next; 53.3 - 57.3",
   "ram_gb": "64",
   "retrieved": "2026-10-05",
   "source_name": "HF card SaturnHeaven/Qwen3.8-Flash-Next-RANMA-EXL3-GGUF",
   "source_url": "https://huggingface.co/SaturnHeaven/Qwen3.8-Flash-Next-RANMA-EXL3-GGUF",
   "tg_tokens": "384",
   "tps": 53.3,
   "ttft_s": null
  },
  {
   "backend": "ranma.cpp (ROCm)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "RX 9070 XT (emulated)",
   "id": "hf-ranma-9070xt-emul-64gb-bench-305",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Qwen3.8-Flash-Next RANMA EXL3 (ranma.cpp ranma_20261005, MoE experts in host RAM over PCIe), llama-bench TG128 at depth 0, warm expert cache, exclusive mode, measured once; emulated: runs on the R9700 with the expert cache limited to a 16 GiB card budget; config=r64-bench-305",
   "params": "80B-A3B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL3 3.05 bpw",
   "quote": "RX 9070 XT (emulated); Qwen3.8-Flash-Next; RX 9070 XT (emulated) + 64 GB RAM | 43.2 | 37.0",
   "ram_gb": "64",
   "retrieved": "2026-10-05",
   "source_name": "HF card SaturnHeaven/Qwen3.8-Flash-Next-RANMA-EXL3-GGUF",
   "source_url": "https://huggingface.co/SaturnHeaven/Qwen3.8-Flash-Next-RANMA-EXL3-GGUF",
   "tg_tokens": "128",
   "tps": 43.2,
   "ttft_s": null
  },
  {
   "backend": "ranma.cpp (ROCm)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "RX 9070 XT (emulated)",
   "id": "hf-ranma-9070xt-emul-64gb-bench-405",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Qwen3.8-Flash-Next RANMA EXL3 (ranma.cpp ranma_20261005, MoE experts in host RAM over PCIe), llama-bench TG128 at depth 0, warm expert cache, exclusive mode, measured once; emulated: runs on the R9700 with the expert cache limited to a 16 GiB card budget; config=r64-bench-405",
   "params": "80B-A3B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL3 4.05 bpw",
   "quote": "RX 9070 XT (emulated); Qwen3.8-Flash-Next; RX 9070 XT (emulated) + 64 GB RAM | 43.2 | 37.0",
   "ram_gb": "64",
   "retrieved": "2026-10-05",
   "source_name": "HF card SaturnHeaven/Qwen3.8-Flash-Next-RANMA-EXL3-GGUF",
   "source_url": "https://huggingface.co/SaturnHeaven/Qwen3.8-Flash-Next-RANMA-EXL3-GGUF",
   "tg_tokens": "128",
   "tps": 37.0,
   "ttft_s": null
  },
  {
   "backend": "ranma.cpp (ROCm)",
   "batch": 1,
   "ctx": "65536",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "RX 9070 XT (emulated)",
   "id": "hf-ranma-9070xt-emul-64gb-mtp-305",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Qwen3.8-Flash-Next RANMA EXL3 (ranma.cpp ranma_20261005), MTP smart decoding via server (draft head mtp-Qwen3.8-Flash-Next-Q8_0.gguf), context 65536, greedy, at most 384 generated tokens, published as a range over coding and four roleplay presets (stored low end); emulated 16 GiB card budget, range 39.6-42.2 tok/s; config=r64-mtp-305",
   "params": "80B-A3B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL3 3.05 bpw",
   "quote": "RX 9070 XT (emulated); Qwen3.8-Flash-Next; 39.6 - 42.2",
   "ram_gb": "64",
   "retrieved": "2026-10-05",
   "source_name": "HF card SaturnHeaven/Qwen3.8-Flash-Next-RANMA-EXL3-GGUF",
   "source_url": "https://huggingface.co/SaturnHeaven/Qwen3.8-Flash-Next-RANMA-EXL3-GGUF",
   "tg_tokens": "384",
   "tps": 39.6,
   "ttft_s": null
  },
  {
   "backend": "ranma.cpp (ROCm)",
   "batch": 1,
   "ctx": "65536",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "RX 9070 XT (emulated)",
   "id": "hf-ranma-9070xt-emul-64gb-mtp-405",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Qwen3.8-Flash-Next RANMA EXL3 (ranma.cpp ranma_20261005), MTP smart decoding via server (draft head mtp-Qwen3.8-Flash-Next-Q8_0.gguf), context 65536, greedy, at most 384 generated tokens, published as a range over coding and four roleplay presets (stored low end); emulated 16 GiB card budget, range 29.1-30.8 tok/s; config=r64-mtp-405",
   "params": "80B-A3B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "EXL3 4.05 bpw",
   "quote": "RX 9070 XT (emulated); Qwen3.8-Flash-Next; 29.1 - 30.8",
   "ram_gb": "64",
   "retrieved": "2026-10-05",
   "source_name": "HF card SaturnHeaven/Qwen3.8-Flash-Next-RANMA-EXL3-GGUF",
   "source_url": "https://huggingface.co/SaturnHeaven/Qwen3.8-Flash-Next-RANMA-EXL3-GGUF",
   "tg_tokens": "384",
   "tps": 29.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "4096",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc30001-4090-kv4096-default",
   "model": "Qwen2.5-7B",
   "notes": "RTX 4090 24 GB, Qwen2.5-7B Q4_0 weights with q8_0 KV cache, same git commit A/B, decode vs KV depth; default build falls back to f16 K/V conversion for q8_0-q4_0, native build compiles only the q8_0-q4_0 FlashAttention kernel (+0.47 MiB libggml-cuda.so); KV depth 4096, default build (f16 fallback), published 149.39 +/- std; config=kv4096-default",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 4090; Qwen2.5-7B; | 4096 | 149.39 +/- 0.45 | 163.51 +/- 0.57 |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30001",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30001",
   "tg_tokens": null,
   "tps": "149.39",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "4096",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc30001-4090-kv4096-native",
   "model": "Qwen2.5-7B",
   "notes": "RTX 4090 24 GB, Qwen2.5-7B Q4_0 weights with q8_0 KV cache, same git commit A/B, decode vs KV depth; default build falls back to f16 K/V conversion for q8_0-q4_0, native build compiles only the q8_0-q4_0 FlashAttention kernel (+0.47 MiB libggml-cuda.so); KV depth 4096, native q8_0-q4_0 FA build, published 163.51 +/- std, speedup +9.5%; config=kv4096-native",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 4090; Qwen2.5-7B; | 4096 | 149.39 +/- 0.45 | 163.51 +/- 0.57 |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30001",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30001",
   "tg_tokens": null,
   "tps": "163.51",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "16384",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc30001-4090-kv16384-default",
   "model": "Qwen2.5-7B",
   "notes": "RTX 4090 24 GB, Qwen2.5-7B Q4_0 weights with q8_0 KV cache, same git commit A/B, decode vs KV depth; default build falls back to f16 K/V conversion for q8_0-q4_0, native build compiles only the q8_0-q4_0 FlashAttention kernel (+0.47 MiB libggml-cuda.so); KV depth 16384, default build (f16 fallback), published 110.87 +/- std; config=kv16384-default",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 4090; Qwen2.5-7B; | 16384 | 110.87 +/- 0.34 | 136.26 +/- 0.41 |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30001",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30001",
   "tg_tokens": null,
   "tps": "110.87",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "16384",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc30001-4090-kv16384-native",
   "model": "Qwen2.5-7B",
   "notes": "RTX 4090 24 GB, Qwen2.5-7B Q4_0 weights with q8_0 KV cache, same git commit A/B, decode vs KV depth; default build falls back to f16 K/V conversion for q8_0-q4_0, native build compiles only the q8_0-q4_0 FlashAttention kernel (+0.47 MiB libggml-cuda.so); KV depth 16384, native q8_0-q4_0 FA build, published 136.26 +/- std, speedup +22.9%; config=kv16384-native",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 4090; Qwen2.5-7B; | 16384 | 110.87 +/- 0.34 | 136.26 +/- 0.41 |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30001",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30001",
   "tg_tokens": null,
   "tps": "136.26",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "32768",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc30001-4090-kv32768-default",
   "model": "Qwen2.5-7B",
   "notes": "RTX 4090 24 GB, Qwen2.5-7B Q4_0 weights with q8_0 KV cache, same git commit A/B, decode vs KV depth; default build falls back to f16 K/V conversion for q8_0-q4_0, native build compiles only the q8_0-q4_0 FlashAttention kernel (+0.47 MiB libggml-cuda.so); KV depth 32768, default build (f16 fallback), published 88.07 +/- std; config=kv32768-default",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 4090; Qwen2.5-7B; | 32768 | 88.07 +/- 0.11 | 110.98 +/- 0.60 |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30001",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30001",
   "tg_tokens": null,
   "tps": "88.07",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "32768",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc30001-4090-kv32768-native",
   "model": "Qwen2.5-7B",
   "notes": "RTX 4090 24 GB, Qwen2.5-7B Q4_0 weights with q8_0 KV cache, same git commit A/B, decode vs KV depth; default build falls back to f16 K/V conversion for q8_0-q4_0, native build compiles only the q8_0-q4_0 FlashAttention kernel (+0.47 MiB libggml-cuda.so); KV depth 32768, native q8_0-q4_0 FA build, published 110.98 +/- std, speedup +26.0%; config=kv32768-native",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 4090; Qwen2.5-7B; | 32768 | 88.07 +/- 0.11 | 110.98 +/- 0.60 |",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30001",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30001",
   "tg_tokens": null,
   "tps": "110.98",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": "32768",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Radeon 780M",
   "id": "lc29965-780m-groups-off",
   "model": "GLM-4.7-Flash",
   "notes": "Radeon 780M (gfx1103, RADV Mesa 26.2.4), CachyOS, llama.cpp Vulkan coopmat1 path, GLM-4.7-Flash Q4_K_M at 32K context; GLM-4.7-Flash has qk_ratio 20 > max_gqa 16 so GQA mode is off by default and every head reads the whole KV cache; baseline (branch cb7934c52 vs baseline 1fb7ef3e3), heads not grouped; config=fa-gqa-groups-off",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Radeon 780M; GLM-4.7-Flash; On a Radeon 780M with GLM-4.7-Flash Q4_K_M, token generation improves from 8.83 to 12.91 t/s at 32K context (+46%) and by 15-19% at 8K. Prompt processing is unchanged.",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #29965",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29965",
   "tg_tokens": null,
   "tps": "8.83",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": "32768",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Radeon 780M",
   "id": "lc29965-780m-groups-on",
   "model": "GLM-4.7-Flash",
   "notes": "Radeon 780M (gfx1103, RADV Mesa 26.2.4), CachyOS, llama.cpp Vulkan coopmat1 path, GLM-4.7-Flash Q4_K_M at 32K context; GLM-4.7-Flash has qk_ratio 20 > max_gqa 16 so GQA mode is off by default and every head reads the whole KV cache; heads grouped to the largest divisor of qk_ratio at or below max_gqa (group floor 4 heads); the author also measured +15-19% at 8K and unchanged prefill; config=fa-gqa-groups-on",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Radeon 780M; GLM-4.7-Flash; On a Radeon 780M with GLM-4.7-Flash Q4_K_M, token generation improves from 8.83 to 12.91 t/s at 32K context (+46%) and by 15-19% at 8K. Prompt processing is unchanged.",
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #29965",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/29965",
   "tg_tokens": null,
   "tps": "12.91",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "256000",
   "date": "2026-09-03",
   "flags": [],
   "hardware": "RTX 5080 (PCIe Gen 4) + Ryzen 3900X",
   "id": "lc-lid-qwen36-cache",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Lidenburg moe-expert-caching fork (three-tier GPU/RAM/disk expert cache, LFU with aging, Linux, io_uring + O_DIRECT disk tier), RTX 5080 (PCIe Gen 4) + Ryzen 3900X, 32 GB DDR4, decade-old SATA SSD ~500 MB/s, llama-cli -fa on -ctk q8_0 -ctv q8_0 -c 256000 --jinja -ngl 99 -t 11 -n 1000, short prompt, published as approximate t/s; expert cache on, 120 experts/tensor in VRAM + 136 in pinned RAM, no disk tier, 15253 MiB VRAM, 12 GB RAM; published ~80 t/s, +60% vs the -ncmoe arm; config=lid-qwen36-cache",
   "params": "35B-A3B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_M",
   "quote": "RTX 5080 (PCIe Gen 4), Ryzen 3900X, 32 GB DDR4 RAM; Qwen3.6-35B-A3B-UD-Q4_K_M.gguf; 120/136/0 15,253 MiB 12 GB ~80 t/s +60%",
   "ram_gb": "32",
   "retrieved": "2026-10-05",
   "source_name": "Lidenburg/llama.cpp README (moe-expert-caching)",
   "source_url": "https://github.com/Lidenburg/llama.cpp/blob/moe-expert-caching/README.md",
   "tg_tokens": "1000",
   "tps": "80",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "256000",
   "date": "2026-09-03",
   "flags": [],
   "hardware": "RTX 5080 (PCIe Gen 4) + Ryzen 3900X",
   "id": "lc-lid-qwen36-ncmoe",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Lidenburg moe-expert-caching fork (three-tier GPU/RAM/disk expert cache, LFU with aging, Linux, io_uring + O_DIRECT disk tier), RTX 5080 (PCIe Gen 4) + Ryzen 3900X, 32 GB DDR4, decade-old SATA SSD ~500 MB/s, llama-cli -fa on -ctk q8_0 -ctv q8_0 -c 256000 --jinja -ngl 99 -t 11 -n 1000, short prompt, published as approximate t/s; stock --n-cpu-moe 21 at matched VRAM, 15766 MiB, 12.5 GB RAM; published ~50 t/s; config=lid-qwen36-ncmoe",
   "params": "35B-A3B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_M",
   "quote": "RTX 5080 (PCIe Gen 4), Ryzen 3900X, 32 GB DDR4 RAM; Qwen3.6-35B-A3B-UD-Q4_K_M.gguf; -ncmoe 21 15,766 MiB 12.5 GB ~50 t/s",
   "ram_gb": "32",
   "retrieved": "2026-10-05",
   "source_name": "Lidenburg/llama.cpp README (moe-expert-caching)",
   "source_url": "https://github.com/Lidenburg/llama.cpp/blob/moe-expert-caching/README.md",
   "tg_tokens": "1000",
   "tps": "50",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "256000",
   "date": "2026-09-03",
   "flags": [],
   "hardware": "RTX 5080 (PCIe Gen 4) + Ryzen 3900X",
   "id": "lc-lid-codernext-cache",
   "model": "Qwen3-Coder-Next",
   "notes": "Lidenburg moe-expert-caching fork (three-tier GPU/RAM/disk expert cache, LFU with aging, Linux, io_uring + O_DIRECT disk tier), RTX 5080 (PCIe Gen 4) + Ryzen 3900X, 32 GB DDR4, decade-old SATA SSD ~500 MB/s, llama-cli -fa on -ctk q8_0 -ctv q8_0 -c 256000 --jinja -ngl 99 -t 11 -n 1000, short prompt, published as approximate t/s; expert cache on, 110 in VRAM + 402 in pinned RAM, 14732 MiB VRAM, 28.7 GB RAM; published ~50 t/s, +66% vs the -ncmoe arm; config=lid-codernext-cache",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "RTX 5080 (PCIe Gen 4), Ryzen 3900X, 32 GB DDR4 RAM; Qwen3-Coder-Next-UD-IQ4_XS.gguf; 110/402/0 14,732 MiB 28.7 GB ~50 t/s +66%",
   "ram_gb": "32",
   "retrieved": "2026-10-05",
   "source_name": "Lidenburg/llama.cpp README (moe-expert-caching)",
   "source_url": "https://github.com/Lidenburg/llama.cpp/blob/moe-expert-caching/README.md",
   "tg_tokens": "1000",
   "tps": "50",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "256000",
   "date": "2026-09-03",
   "flags": [],
   "hardware": "RTX 5080 (PCIe Gen 4) + Ryzen 3900X",
   "id": "lc-lid-codernext-ncmoe",
   "model": "Qwen3-Coder-Next",
   "notes": "Lidenburg moe-expert-caching fork (three-tier GPU/RAM/disk expert cache, LFU with aging, Linux, io_uring + O_DIRECT disk tier), RTX 5080 (PCIe Gen 4) + Ryzen 3900X, 32 GB DDR4, decade-old SATA SSD ~500 MB/s, llama-cli -fa on -ctk q8_0 -ctv q8_0 -c 256000 --jinja -ngl 99 -t 11 -n 1000, short prompt, published as approximate t/s; stock --n-cpu-moe 37 at matched VRAM, 15091 MiB, 28.4 GB RAM; published ~30 t/s; config=lid-codernext-ncmoe",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "RTX 5080 (PCIe Gen 4), Ryzen 3900X, 32 GB DDR4 RAM; Qwen3-Coder-Next-UD-IQ4_XS.gguf; -ncmoe 37 15,091 MiB 28.4 GB ~30 t/s",
   "ram_gb": "32",
   "retrieved": "2026-10-05",
   "source_name": "Lidenburg/llama.cpp README (moe-expert-caching)",
   "source_url": "https://github.com/Lidenburg/llama.cpp/blob/moe-expert-caching/README.md",
   "tg_tokens": "1000",
   "tps": "30",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": "64",
   "ctx": "16384",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "MI350X",
   "id": "vllm-60091-mi350x-dsv41-main",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #60091 on MI350X (gfx950), DeepSeek-V4.1-Flash TP=4 with expert parallelism, VLLM_ROCM_USE_AITER=1, max_num_batched_tokens 16384, mean of 3-5 reps per arm, main and PR swapped in place and interleaved in rotated order, ranges disjoint in every row; main: prompt 31440 tok/s at rate inf, aggregate output 1717 tok/s at concurrency 64 (8K/1K), median TTFT 340.7 ms at conc 1 / 8192 in, 608.7 ms at conc 64, median TPOT 33.34 ms; config=60091-main",
   "params": "?",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "31440",
   "provenance": "sourced",
   "quant": "unspecified",
   "quote": "MI350X (gfx950), DeepSeek-V4.1-Flash, TP=4 with expert parallelism; 8192 \u00b1 25 % in, rate inf prompt tok/s 31,440 32,443 +3.2 %; 8K/1K \u00b1 25 %, concurrency 64 output tok/s 1,717 1,763 +2.7 %",
   "retrieved": "2026-10-05",
   "source_name": "vLLM PR #60091",
   "source_url": "https://github.com/vllm-project/vllm/pull/60091",
   "tg_tokens": null,
   "tps": "1717",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": "64",
   "ctx": "16384",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "MI350X",
   "id": "vllm-60091-mi350x-dsv41-pr",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #60091 on MI350X (gfx950), DeepSeek-V4.1-Flash TP=4 with expert parallelism, VLLM_ROCM_USE_AITER=1, max_num_batched_tokens 16384, mean of 3-5 reps per arm, main and PR swapped in place and interleaved in rotated order, ranges disjoint in every row; PR (pass output_size to repeat_interleave, removes 160 host syncs per prefill step): prompt 32443 tok/s (+3.2%), aggregate output 1763 tok/s (+2.7%), median TTFT 321.1 ms at conc 1 (-5.7%), 572.2 ms at conc 64 (-6.0%), median TPOT 32.47 ms (-2.6%); about 20 ms saved per prefill step; GSM8K 0.9014 vs 0.9083; config=60091-pr",
   "params": "?",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "32443",
   "provenance": "sourced",
   "quant": "unspecified",
   "quote": "MI350X (gfx950), DeepSeek-V4.1-Flash, TP=4 with expert parallelism; 8192 \u00b1 25 % in, rate inf prompt tok/s 31,440 32,443 +3.2 %; 8K/1K \u00b1 25 %, concurrency 64 output tok/s 1,717 1,763 +2.7 %",
   "retrieved": "2026-10-05",
   "source_name": "vLLM PR #60091",
   "source_url": "https://github.com/vllm-project/vllm/pull/60091",
   "tg_tokens": null,
   "tps": "1763",
   "ttft_s": null
  },
  {
   "backend": "Ollama",
   "batch": 1,
   "ctx": "16384",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "tq-m1pro-ollama-q4km-fp16kv",
   "model": "Qwen2.5 3B",
   "notes": "devYRPauli turboquant-m1pro-evaluation, MacBook Pro Apple M1 Pro 16 GB unified memory, Qwen2.5-3B-Instruct at 16K context (15947 prompt tokens, up to 80 generated), corrected 2026-10-05 table from benchmarks/phase3_results.json; Ollama Q4_K_M GGUF with its default f16 KV cache, 37.5 tok/s is Ollama's own decode rate excluding the prompt, wall time for the request 49.3 s including the HTTP call and any model load",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Apple M1 Pro chip, 16GB unified memory; Qwen2.5-3B-Instruct (Q4_K_M GGUF via Ollama, and 4-bit MLX via Hugging Face); Ollama (FP16 KV) | 37.5 (decode only) | 49.3 s",
   "ram_gb": "16",
   "retrieved": "2026-10-05",
   "source_name": "devYRPauli/turboquant-m1pro-evaluation README",
   "source_url": "https://github.com/devYRPauli/turboquant-m1pro-evaluation/blob/main/README.md",
   "tg_tokens": "80",
   "tps": "37.5",
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": 1,
   "ctx": "16384",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "tq-m1pro-mlx-fp16kv",
   "model": "Qwen2.5 3B",
   "notes": "devYRPauli turboquant-m1pro-evaluation, MacBook Pro Apple M1 Pro 16 GB unified memory, Qwen2.5-3B-Instruct at 16K context (15947 prompt tokens, up to 80 generated), corrected 2026-10-05 table from benchmarks/phase3_results.json; 4-bit MLX model, FP16 KV cache: 2.0 tok/s divides generated tokens by total request time including the 16K prefill, wall time 39.9 s, measured KV cache 563.4 MB, MLX peak memory 3076.8 MB; config=tq-fp16kv",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "4-bit (MLX)",
   "quote": "Apple M1 Pro chip, 16GB unified memory; Qwen2.5-3B-Instruct (Q4_K_M GGUF via Ollama, and 4-bit MLX via Hugging Face); MLX baseline FP16 | 2.0 (prefill included) | 39.9 s",
   "ram_gb": "16",
   "retrieved": "2026-10-05",
   "source_name": "devYRPauli/turboquant-m1pro-evaluation README",
   "source_url": "https://github.com/devYRPauli/turboquant-m1pro-evaluation/blob/main/README.md",
   "tg_tokens": "80",
   "tps": "2.0",
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": 1,
   "ctx": "16384",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "tq-m1pro-mlx-hybrid-kv",
   "model": "Qwen2.5 3B",
   "notes": "devYRPauli turboquant-m1pro-evaluation, MacBook Pro Apple M1 Pro 16 GB unified memory, Qwen2.5-3B-Instruct at 16K context (15947 prompt tokens, up to 80 generated), corrected 2026-10-05 table from benchmarks/phase3_results.json; same run with the Hybrid K5/V4 quantized KV cache (5-bit keys 4-bit MSE + 1-bit QJL, 4-bit values): 1.1 tok/s prefill included, wall time 71.4 s, stored cache 435.6 MB (1.29x smaller than FP16, not the 4x formula figure), MLX peak memory 3245.4 MB, higher than FP16 because dequantized K/V returns float32; slower because it dequantizes the full cache every step in unfused MLX ops; config=tq-hybrid-kv",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "4-bit (MLX)",
   "quote": "Apple M1 Pro chip, 16GB unified memory; Qwen2.5-3B-Instruct (Q4_K_M GGUF via Ollama, and 4-bit MLX via Hugging Face); MLX Hybrid K5/V4 | 1.1 (prefill included) | 71.4 s",
   "ram_gb": "16",
   "retrieved": "2026-10-05",
   "source_name": "devYRPauli/turboquant-m1pro-evaluation README",
   "source_url": "https://github.com/devYRPauli/turboquant-m1pro-evaluation/blob/main/README.md",
   "tg_tokens": "80",
   "tps": "1.1",
   "ttft_s": null
  },
  {
   "backend": "Strata",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "hf6-gyro-5090-strata-gyrom-prose",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Strata rc1, all experts cached, greedy, MTP draft always on; Gyro-M prose; config=strata-gyrom-prose",
   "params": "126B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Gyro-M",
   "quote": "Qwen3.8-Flash-Next; RTX 5090, 32 GB Strata rc1 (release candidate) prose 140 , code 209, JSON 221 4,883 (16k prompt) prose 114",
   "ram_gb": 32,
   "retrieved": "2026-10-05",
   "source_name": "HF model card: agentionai/Qwen3.8-Flash-Next-Gyro-GGUF",
   "source_url": "https://huggingface.co/agentionai/Qwen3.8-Flash-Next-Gyro-GGUF",
   "tg_tokens": 128,
   "tps": 114,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 16,
   "ctx": 2048,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "A800M",
   "id": "lc30021-a800m-q8_0-ub16-master",
   "model": "Granite 3.0 3B",
   "notes": "pp2048 prefill, Q8_0, ub 16, master; config=q8_0-ub16-master",
   "params": null,
   "power_w": null,
   "pp_tps": 345.34,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Granite 3.0 3B; Granite 3.0 3B A800M; 16 345.34 \u00b1 0.43 t/s 613.21 \u00b1 2.03 t/s **+77.57%**",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30021",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30021",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 16,
   "ctx": 2048,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "A800M",
   "id": "lc30021-a800m-q8_0-ub16-pr",
   "model": "Granite 3.0 3B",
   "notes": "pp2048 prefill, Q8_0, ub 16, PR; config=q8_0-ub16-pr",
   "params": null,
   "power_w": null,
   "pp_tps": 613.21,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Granite 3.0 3B; Granite 3.0 3B A800M; 16 345.34 \u00b1 0.43 t/s 613.21 \u00b1 2.03 t/s **+77.57%**",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30021",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30021",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 32,
   "ctx": 2048,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "A800M",
   "id": "lc30021-a800m-q8_0-ub32-master",
   "model": "Granite 3.0 3B",
   "notes": "pp2048 prefill, Q8_0, ub 32, master; config=q8_0-ub32-master",
   "params": null,
   "power_w": null,
   "pp_tps": 866.6,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Granite 3.0 3B; Granite 3.0 3B A800M; 32 866.60 \u00b1 4.54 t/s 866.49 \u00b1 3.71 t/s -0.01%",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30021",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30021",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 32,
   "ctx": 2048,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "A800M",
   "id": "lc30021-a800m-q8_0-ub32-pr",
   "model": "Granite 3.0 3B",
   "notes": "pp2048 prefill, Q8_0, ub 32, PR; config=q8_0-ub32-pr",
   "params": null,
   "power_w": null,
   "pp_tps": 866.49,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Granite 3.0 3B; Granite 3.0 3B A800M; 32 866.60 \u00b1 4.54 t/s 866.49 \u00b1 3.71 t/s -0.01%",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30021",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30021",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 64,
   "ctx": 2048,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "A800M",
   "id": "lc30021-a800m-q8_0-ub64-master",
   "model": "Granite 3.0 3B",
   "notes": "pp2048 prefill, Q8_0, ub 64, master; config=q8_0-ub64-master",
   "params": null,
   "power_w": null,
   "pp_tps": 1093.52,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Granite 3.0 3B; Granite 3.0 3B A800M; 64 1093.52 \u00b1 9.49 t/s 1092.78 \u00b1 8.57 t/s -0.07%",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30021",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30021",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 64,
   "ctx": 2048,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "A800M",
   "id": "lc30021-a800m-q8_0-ub64-pr",
   "model": "Granite 3.0 3B",
   "notes": "pp2048 prefill, Q8_0, ub 64, PR; config=q8_0-ub64-pr",
   "params": null,
   "power_w": null,
   "pp_tps": 1092.78,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Granite 3.0 3B; Granite 3.0 3B A800M; 64 1093.52 \u00b1 9.49 t/s 1092.78 \u00b1 8.57 t/s -0.07%",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30021",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30021",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 80,
   "ctx": 2048,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "A800M",
   "id": "lc30021-a800m-q8_0-ub80-master",
   "model": "Granite 3.0 3B",
   "notes": "pp2048 prefill, Q8_0, ub 80, master; config=q8_0-ub80-master",
   "params": null,
   "power_w": null,
   "pp_tps": 1170.83,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Granite 3.0 3B; Granite 3.0 3B A800M; 80 1170.83 \u00b1 7.22 t/s 1171.68 \u00b1 9.34 t/s +0.07%",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30021",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30021",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 80,
   "ctx": 2048,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "A800M",
   "id": "lc30021-a800m-q8_0-ub80-pr",
   "model": "Granite 3.0 3B",
   "notes": "pp2048 prefill, Q8_0, ub 80, PR; config=q8_0-ub80-pr",
   "params": null,
   "power_w": null,
   "pp_tps": 1171.68,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Granite 3.0 3B; Granite 3.0 3B A800M; 80 1170.83 \u00b1 7.22 t/s 1171.68 \u00b1 9.34 t/s +0.07%",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30021",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30021",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 96,
   "ctx": 2048,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "A800M",
   "id": "lc30021-a800m-q8_0-ub96-master",
   "model": "Granite 3.0 3B",
   "notes": "pp2048 prefill, Q8_0, ub 96, master; config=q8_0-ub96-master",
   "params": null,
   "power_w": null,
   "pp_tps": 1164.62,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Granite 3.0 3B; Granite 3.0 3B A800M; 96 1164.62 \u00b1 7.94 t/s 1169.14 \u00b1 9.26 t/s +0.39%",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30021",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30021",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 96,
   "ctx": 2048,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "A800M",
   "id": "lc30021-a800m-q8_0-ub96-pr",
   "model": "Granite 3.0 3B",
   "notes": "pp2048 prefill, Q8_0, ub 96, PR; config=q8_0-ub96-pr",
   "params": null,
   "power_w": null,
   "pp_tps": 1169.14,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Granite 3.0 3B; Granite 3.0 3B A800M; 96 1164.62 \u00b1 7.94 t/s 1169.14 \u00b1 9.26 t/s +0.39%",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30021",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30021",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 128,
   "ctx": 2048,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "A800M",
   "id": "lc30021-a800m-q8_0-ub128-master",
   "model": "Granite 3.0 3B",
   "notes": "pp2048 prefill, Q8_0, ub 128, master; config=q8_0-ub128-master",
   "params": null,
   "power_w": null,
   "pp_tps": 1314.74,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Granite 3.0 3B; Granite 3.0 3B A800M; 128 1314.74 \u00b1 10.86 t/s 1312.95 \u00b1 9.19 t/s -0.14%",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30021",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30021",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 128,
   "ctx": 2048,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "A800M",
   "id": "lc30021-a800m-q8_0-ub128-pr",
   "model": "Granite 3.0 3B",
   "notes": "pp2048 prefill, Q8_0, ub 128, PR; config=q8_0-ub128-pr",
   "params": null,
   "power_w": null,
   "pp_tps": 1312.95,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Granite 3.0 3B; Granite 3.0 3B A800M; 128 1314.74 \u00b1 10.86 t/s 1312.95 \u00b1 9.19 t/s -0.14%",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30021",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30021",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 256,
   "ctx": 2048,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "A800M",
   "id": "lc30021-a800m-q8_0-ub256-master",
   "model": "Granite 3.0 3B",
   "notes": "pp2048 prefill, Q8_0, ub 256, master; config=q8_0-ub256-master",
   "params": null,
   "power_w": null,
   "pp_tps": 2363.34,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Granite 3.0 3B; Granite 3.0 3B A800M; 256 2363.34 \u00b1 43.50 t/s 2364.72 \u00b1 44.69 t/s +0.06%",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30021",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30021",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 256,
   "ctx": 2048,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "A800M",
   "id": "lc30021-a800m-q8_0-ub256-pr",
   "model": "Granite 3.0 3B",
   "notes": "pp2048 prefill, Q8_0, ub 256, PR; config=q8_0-ub256-pr",
   "params": null,
   "power_w": null,
   "pp_tps": 2364.72,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Granite 3.0 3B; Granite 3.0 3B A800M; 256 2363.34 \u00b1 43.50 t/s 2364.72 \u00b1 44.69 t/s +0.06%",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30021",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30021",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 512,
   "ctx": 2048,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "A800M",
   "id": "lc30021-a800m-q8_0-ub512-master",
   "model": "Granite 3.0 3B",
   "notes": "pp2048 prefill, Q8_0, ub 512, master; config=q8_0-ub512-master",
   "params": null,
   "power_w": null,
   "pp_tps": 3445.3,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Granite 3.0 3B; Granite 3.0 3B A800M; 512 3445.30 \u00b1 98.02 t/s 3450.53 \u00b1 98.40 t/s +0.15%",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30021",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30021",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 512,
   "ctx": 2048,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "A800M",
   "id": "lc30021-a800m-q8_0-ub512-pr",
   "model": "Granite 3.0 3B",
   "notes": "pp2048 prefill, Q8_0, ub 512, PR; config=q8_0-ub512-pr",
   "params": null,
   "power_w": null,
   "pp_tps": 3450.53,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Granite 3.0 3B; Granite 3.0 3B A800M; 512 3445.30 \u00b1 98.02 t/s 3450.53 \u00b1 98.40 t/s +0.15%",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30021",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30021",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 16,
   "ctx": 2048,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "A800M",
   "id": "lc30021-a800m-iq4_xs-ub16-master",
   "model": "Granite 3.0 3B",
   "notes": "pp2048 prefill, IQ4_XS, ub 16, master; config=iq4_xs-ub16-master",
   "params": null,
   "power_w": null,
   "pp_tps": 594.71,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "Granite 3.0 3B; Granite 3.0 3B A800M; 16 594.71 \u00b1 1.80 t/s 647.84 \u00b1 1.76 t/s **+8.93%**",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30021",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30021",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 16,
   "ctx": 2048,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "A800M",
   "id": "lc30021-a800m-iq4_xs-ub16-pr",
   "model": "Granite 3.0 3B",
   "notes": "pp2048 prefill, IQ4_XS, ub 16, PR; config=iq4_xs-ub16-pr",
   "params": null,
   "power_w": null,
   "pp_tps": 647.84,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "Granite 3.0 3B; Granite 3.0 3B A800M; 16 594.71 \u00b1 1.80 t/s 647.84 \u00b1 1.76 t/s **+8.93%**",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30021",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30021",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 32,
   "ctx": 2048,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "A800M",
   "id": "lc30021-a800m-iq4_xs-ub32-master",
   "model": "Granite 3.0 3B",
   "notes": "pp2048 prefill, IQ4_XS, ub 32, master; config=iq4_xs-ub32-master",
   "params": null,
   "power_w": null,
   "pp_tps": 932.02,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "Granite 3.0 3B; Granite 3.0 3B A800M; 32 932.02 \u00b1 6.08 t/s 930.00 \u00b1 2.46 t/s -0.22%",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30021",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30021",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 32,
   "ctx": 2048,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "A800M",
   "id": "lc30021-a800m-iq4_xs-ub32-pr",
   "model": "Granite 3.0 3B",
   "notes": "pp2048 prefill, IQ4_XS, ub 32, PR; config=iq4_xs-ub32-pr",
   "params": null,
   "power_w": null,
   "pp_tps": 930.0,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "Granite 3.0 3B; Granite 3.0 3B A800M; 32 932.02 \u00b1 6.08 t/s 930.00 \u00b1 2.46 t/s -0.22%",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30021",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30021",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 48,
   "ctx": 2048,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "A800M",
   "id": "lc30021-a800m-iq4_xs-ub48-master",
   "model": "Granite 3.0 3B",
   "notes": "pp2048 prefill, IQ4_XS, ub 48, master; config=iq4_xs-ub48-master",
   "params": null,
   "power_w": null,
   "pp_tps": 973.78,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "Granite 3.0 3B; Granite 3.0 3B A800M; 48 973.78 \u00b1 3.68 t/s 1168.95 \u00b1 10.42 t/s **+20.04%**",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30021",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30021",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 48,
   "ctx": 2048,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "A800M",
   "id": "lc30021-a800m-iq4_xs-ub48-pr",
   "model": "Granite 3.0 3B",
   "notes": "pp2048 prefill, IQ4_XS, ub 48, PR; config=iq4_xs-ub48-pr",
   "params": null,
   "power_w": null,
   "pp_tps": 1168.95,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "Granite 3.0 3B; Granite 3.0 3B A800M; 48 973.78 \u00b1 3.68 t/s 1168.95 \u00b1 10.42 t/s **+20.04%**",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30021",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30021",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 64,
   "ctx": 2048,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "A800M",
   "id": "lc30021-a800m-iq4_xs-ub64-master",
   "model": "Granite 3.0 3B",
   "notes": "pp2048 prefill, IQ4_XS, ub 64, master; config=iq4_xs-ub64-master",
   "params": null,
   "power_w": null,
   "pp_tps": 1135.67,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "Granite 3.0 3B; Granite 3.0 3B A800M; 64 1135.67 \u00b1 5.38 t/s 1129.44 \u00b1 5.87 t/s -0.55%",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30021",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30021",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 64,
   "ctx": 2048,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "A800M",
   "id": "lc30021-a800m-iq4_xs-ub64-pr",
   "model": "Granite 3.0 3B",
   "notes": "pp2048 prefill, IQ4_XS, ub 64, PR; config=iq4_xs-ub64-pr",
   "params": null,
   "power_w": null,
   "pp_tps": 1129.44,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "Granite 3.0 3B; Granite 3.0 3B A800M; 64 1135.67 \u00b1 5.38 t/s 1129.44 \u00b1 5.87 t/s -0.55%",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30021",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30021",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 128,
   "ctx": 2048,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "A800M",
   "id": "lc30021-a800m-iq4_xs-ub128-master",
   "model": "Granite 3.0 3B",
   "notes": "pp2048 prefill, IQ4_XS, ub 128, master; config=iq4_xs-ub128-master",
   "params": null,
   "power_w": null,
   "pp_tps": 1970.72,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "Granite 3.0 3B; Granite 3.0 3B A800M; 128 1970.72 \u00b1 19.55 t/s 1974.53 \u00b1 18.91 t/s +0.19%",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30021",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30021",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 128,
   "ctx": 2048,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "A800M",
   "id": "lc30021-a800m-iq4_xs-ub128-pr",
   "model": "Granite 3.0 3B",
   "notes": "pp2048 prefill, IQ4_XS, ub 128, PR; config=iq4_xs-ub128-pr",
   "params": null,
   "power_w": null,
   "pp_tps": 1974.53,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "Granite 3.0 3B; Granite 3.0 3B A800M; 128 1970.72 \u00b1 19.55 t/s 1974.53 \u00b1 18.91 t/s +0.19%",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30021",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30021",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 256,
   "ctx": 2048,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "A800M",
   "id": "lc30021-a800m-iq4_xs-ub256-master",
   "model": "Granite 3.0 3B",
   "notes": "pp2048 prefill, IQ4_XS, ub 256, master; config=iq4_xs-ub256-master",
   "params": null,
   "power_w": null,
   "pp_tps": 3230.53,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "Granite 3.0 3B; Granite 3.0 3B A800M; 256 3230.53 \u00b1 69.99 t/s 3236.54 \u00b1 65.08 t/s +0.19%",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30021",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30021",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 256,
   "ctx": 2048,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "A800M",
   "id": "lc30021-a800m-iq4_xs-ub256-pr",
   "model": "Granite 3.0 3B",
   "notes": "pp2048 prefill, IQ4_XS, ub 256, PR; config=iq4_xs-ub256-pr",
   "params": null,
   "power_w": null,
   "pp_tps": 3236.54,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "Granite 3.0 3B; Granite 3.0 3B A800M; 256 3230.53 \u00b1 69.99 t/s 3236.54 \u00b1 65.08 t/s +0.19%",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30021",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30021",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 512,
   "ctx": 2048,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "A800M",
   "id": "lc30021-a800m-iq4_xs-ub512-master",
   "model": "Granite 3.0 3B",
   "notes": "pp2048 prefill, IQ4_XS, ub 512, master; config=iq4_xs-ub512-master",
   "params": null,
   "power_w": null,
   "pp_tps": 4070.07,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "Granite 3.0 3B; Granite 3.0 3B A800M; 512 4070.07 \u00b1 103.33 t/s 4073.25 \u00b1 101.33 t/s +0.08%",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30021",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30021",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 512,
   "ctx": 2048,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "A800M",
   "id": "lc30021-a800m-iq4_xs-ub512-pr",
   "model": "Granite 3.0 3B",
   "notes": "pp2048 prefill, IQ4_XS, ub 512, PR; config=iq4_xs-ub512-pr",
   "params": null,
   "power_w": null,
   "pp_tps": 4073.25,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "Granite 3.0 3B; Granite 3.0 3B A800M; 512 4070.07 \u00b1 103.33 t/s 4073.25 \u00b1 101.33 t/s +0.08%",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30021",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30021",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 16,
   "ctx": 2048,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "A800M",
   "id": "lc30021-a800m-iq4_nl-ub16-master",
   "model": "Meta Llama 3 8B",
   "notes": "pp2048 prefill, IQ4_NL, ub 16, master; config=iq4_nl-ub16-master",
   "params": null,
   "power_w": null,
   "pp_tps": 310.76,
   "provenance": "community",
   "quant": "IQ4_NL",
   "quote": "Meta Llama 3 8B; Granite 3.0 3B A800M; 16 310.76 \u00b1 0.21 t/s 348.75 \u00b1 0.29 t/s **+12.22%**",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30021",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30021",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 16,
   "ctx": 2048,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "A800M",
   "id": "lc30021-a800m-iq4_nl-ub16-pr",
   "model": "Meta Llama 3 8B",
   "notes": "pp2048 prefill, IQ4_NL, ub 16, PR; config=iq4_nl-ub16-pr",
   "params": null,
   "power_w": null,
   "pp_tps": 348.75,
   "provenance": "community",
   "quant": "IQ4_NL",
   "quote": "Meta Llama 3 8B; Granite 3.0 3B A800M; 16 310.76 \u00b1 0.21 t/s 348.75 \u00b1 0.29 t/s **+12.22%**",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30021",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30021",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 24,
   "ctx": 2048,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "A800M",
   "id": "lc30021-a800m-iq4_nl-ub24-master",
   "model": "Meta Llama 3 8B",
   "notes": "pp2048 prefill, IQ4_NL, ub 24, master; config=iq4_nl-ub24-master",
   "params": null,
   "power_w": null,
   "pp_tps": 404.1,
   "provenance": "community",
   "quant": "IQ4_NL",
   "quote": "Meta Llama 3 8B; Granite 3.0 3B A800M; 24 404.10 \u00b1 0.38 t/s 404.85 \u00b1 0.38 t/s +0.19%",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30021",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30021",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 24,
   "ctx": 2048,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "A800M",
   "id": "lc30021-a800m-iq4_nl-ub24-pr",
   "model": "Meta Llama 3 8B",
   "notes": "pp2048 prefill, IQ4_NL, ub 24, PR; config=iq4_nl-ub24-pr",
   "params": null,
   "power_w": null,
   "pp_tps": 404.85,
   "provenance": "community",
   "quant": "IQ4_NL",
   "quote": "Meta Llama 3 8B; Granite 3.0 3B A800M; 24 404.10 \u00b1 0.38 t/s 404.85 \u00b1 0.38 t/s +0.19%",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30021",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30021",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 32,
   "ctx": 2048,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "A800M",
   "id": "lc30021-a800m-iq4_nl-ub32-master",
   "model": "Meta Llama 3 8B",
   "notes": "pp2048 prefill, IQ4_NL, ub 32, master; config=iq4_nl-ub32-master",
   "params": null,
   "power_w": null,
   "pp_tps": 499.45,
   "provenance": "community",
   "quant": "IQ4_NL",
   "quote": "Meta Llama 3 8B; Granite 3.0 3B A800M; 32 499.45 \u00b1 0.61 t/s 498.73 \u00b1 0.68 t/s -0.14%",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30021",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30021",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 32,
   "ctx": 2048,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "A800M",
   "id": "lc30021-a800m-iq4_nl-ub32-pr",
   "model": "Meta Llama 3 8B",
   "notes": "pp2048 prefill, IQ4_NL, ub 32, PR; config=iq4_nl-ub32-pr",
   "params": null,
   "power_w": null,
   "pp_tps": 498.73,
   "provenance": "community",
   "quant": "IQ4_NL",
   "quote": "Meta Llama 3 8B; Granite 3.0 3B A800M; 32 499.45 \u00b1 0.61 t/s 498.73 \u00b1 0.68 t/s -0.14%",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30021",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30021",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 40,
   "ctx": 2048,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "A800M",
   "id": "lc30021-a800m-iq4_nl-ub40-master",
   "model": "Meta Llama 3 8B",
   "notes": "pp2048 prefill, IQ4_NL, ub 40, master; config=iq4_nl-ub40-master",
   "params": null,
   "power_w": null,
   "pp_tps": 461.94,
   "provenance": "community",
   "quant": "IQ4_NL",
   "quote": "Meta Llama 3 8B; Granite 3.0 3B A800M; 40 461.94 \u00b1 0.48 t/s 461.13 \u00b1 0.48 t/s -0.18%",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30021",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30021",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 40,
   "ctx": 2048,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "A800M",
   "id": "lc30021-a800m-iq4_nl-ub40-pr",
   "model": "Meta Llama 3 8B",
   "notes": "pp2048 prefill, IQ4_NL, ub 40, PR; config=iq4_nl-ub40-pr",
   "params": null,
   "power_w": null,
   "pp_tps": 461.13,
   "provenance": "community",
   "quant": "IQ4_NL",
   "quote": "Meta Llama 3 8B; Granite 3.0 3B A800M; 40 461.94 \u00b1 0.48 t/s 461.13 \u00b1 0.48 t/s -0.18%",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30021",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30021",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 48,
   "ctx": 2048,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "A800M",
   "id": "lc30021-a800m-iq4_nl-ub48-master",
   "model": "Meta Llama 3 8B",
   "notes": "pp2048 prefill, IQ4_NL, ub 48, master; config=iq4_nl-ub48-master",
   "params": null,
   "power_w": null,
   "pp_tps": 481.94,
   "provenance": "community",
   "quant": "IQ4_NL",
   "quote": "Meta Llama 3 8B; Granite 3.0 3B A800M; 48 481.94 \u00b1 0.56 t/s 537.13 \u00b1 0.68 t/s **+11.45%**",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30021",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30021",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 48,
   "ctx": 2048,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "A800M",
   "id": "lc30021-a800m-iq4_nl-ub48-pr",
   "model": "Meta Llama 3 8B",
   "notes": "pp2048 prefill, IQ4_NL, ub 48, PR; config=iq4_nl-ub48-pr",
   "params": null,
   "power_w": null,
   "pp_tps": 537.13,
   "provenance": "community",
   "quant": "IQ4_NL",
   "quote": "Meta Llama 3 8B; Granite 3.0 3B A800M; 48 481.94 \u00b1 0.56 t/s 537.13 \u00b1 0.68 t/s **+11.45%**",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30021",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30021",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 64,
   "ctx": 2048,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "A800M",
   "id": "lc30021-a800m-iq4_nl-ub64-master",
   "model": "Meta Llama 3 8B",
   "notes": "pp2048 prefill, IQ4_NL, ub 64, master; config=iq4_nl-ub64-master",
   "params": null,
   "power_w": null,
   "pp_tps": 593.97,
   "provenance": "community",
   "quant": "IQ4_NL",
   "quote": "Meta Llama 3 8B; Granite 3.0 3B A800M; 64 593.97 \u00b1 0.80 t/s 594.73 \u00b1 0.83 t/s +0.13%",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30021",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30021",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 64,
   "ctx": 2048,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "A800M",
   "id": "lc30021-a800m-iq4_nl-ub64-pr",
   "model": "Meta Llama 3 8B",
   "notes": "pp2048 prefill, IQ4_NL, ub 64, PR; config=iq4_nl-ub64-pr",
   "params": null,
   "power_w": null,
   "pp_tps": 594.73,
   "provenance": "community",
   "quant": "IQ4_NL",
   "quote": "Meta Llama 3 8B; Granite 3.0 3B A800M; 64 593.97 \u00b1 0.80 t/s 594.73 \u00b1 0.83 t/s +0.13%",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30021",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30021",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 128,
   "ctx": 2048,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "A800M",
   "id": "lc30021-a800m-iq4_nl-ub128-master",
   "model": "Meta Llama 3 8B",
   "notes": "pp2048 prefill, IQ4_NL, ub 128, master; config=iq4_nl-ub128-master",
   "params": null,
   "power_w": null,
   "pp_tps": 660.72,
   "provenance": "community",
   "quant": "IQ4_NL",
   "quote": "Meta Llama 3 8B; Granite 3.0 3B A800M; 128 660.72 \u00b1 0.96 t/s 661.46 \u00b1 1.03 t/s +0.11%",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30021",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30021",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 128,
   "ctx": 2048,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "A800M",
   "id": "lc30021-a800m-iq4_nl-ub128-pr",
   "model": "Meta Llama 3 8B",
   "notes": "pp2048 prefill, IQ4_NL, ub 128, PR; config=iq4_nl-ub128-pr",
   "params": null,
   "power_w": null,
   "pp_tps": 661.46,
   "provenance": "community",
   "quant": "IQ4_NL",
   "quote": "Meta Llama 3 8B; Granite 3.0 3B A800M; 128 660.72 \u00b1 0.96 t/s 661.46 \u00b1 1.03 t/s +0.11%",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30021",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30021",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 256,
   "ctx": 2048,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "A800M",
   "id": "lc30021-a800m-iq4_nl-ub256-master",
   "model": "Meta Llama 3 8B",
   "notes": "pp2048 prefill, IQ4_NL, ub 256, master; config=iq4_nl-ub256-master",
   "params": null,
   "power_w": null,
   "pp_tps": 794.69,
   "provenance": "community",
   "quant": "IQ4_NL",
   "quote": "Meta Llama 3 8B; Granite 3.0 3B A800M; 256 794.69 \u00b1 2.42 t/s 795.33 \u00b1 2.44 t/s +0.08%",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30021",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30021",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 256,
   "ctx": 2048,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "A800M",
   "id": "lc30021-a800m-iq4_nl-ub256-pr",
   "model": "Meta Llama 3 8B",
   "notes": "pp2048 prefill, IQ4_NL, ub 256, PR; config=iq4_nl-ub256-pr",
   "params": null,
   "power_w": null,
   "pp_tps": 795.33,
   "provenance": "community",
   "quant": "IQ4_NL",
   "quote": "Meta Llama 3 8B; Granite 3.0 3B A800M; 256 794.69 \u00b1 2.42 t/s 795.33 \u00b1 2.44 t/s +0.08%",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30021",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30021",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 512,
   "ctx": 2048,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "A800M",
   "id": "lc30021-a800m-iq4_nl-ub512-master",
   "model": "Meta Llama 3 8B",
   "notes": "pp2048 prefill, IQ4_NL, ub 512, master; config=iq4_nl-ub512-master",
   "params": null,
   "power_w": null,
   "pp_tps": 867.15,
   "provenance": "community",
   "quant": "IQ4_NL",
   "quote": "Meta Llama 3 8B; Granite 3.0 3B A800M; 512 867.15 \u00b1 2.91 t/s 867.05 \u00b1 2.94 t/s -0.01%",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30021",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30021",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 512,
   "ctx": 2048,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "A800M",
   "id": "lc30021-a800m-iq4_nl-ub512-pr",
   "model": "Meta Llama 3 8B",
   "notes": "pp2048 prefill, IQ4_NL, ub 512, PR; config=iq4_nl-ub512-pr",
   "params": null,
   "power_w": null,
   "pp_tps": 867.05,
   "provenance": "community",
   "quant": "IQ4_NL",
   "quote": "Meta Llama 3 8B; Granite 3.0 3B A800M; 512 867.15 \u00b1 2.91 t/s 867.05 \u00b1 2.94 t/s -0.01%",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30021",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30021",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 512,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Arc B580",
   "id": "lc30018-b580-gemma4-before",
   "model": "Gemma 4 E4B",
   "notes": "parent of the regressing commit (2ca15f54), tg128 / pp512; config=before",
   "params": null,
   "power_w": null,
   "pp_tps": 3323,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Gemma-4-E4B; Intel Arc B580; 2ca15f54 (CUDA swizzling refactor, no llama-side changes) 90.7 t/s 3323 t/s",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30018",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30018",
   "tg_tokens": null,
   "tps": 90.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 512,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Arc B580",
   "id": "lc30018-b580-gemma4-after",
   "model": "Gemma 4 E4B",
   "notes": "regressing commit 0bb496db (#29622), tg128 / pp512; config=after",
   "params": null,
   "power_w": null,
   "pp_tps": 3449,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Gemma-4-E4B; Intel Arc B580; 0bb496db (#29622) 86.2 t/s 3449 t/s",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30018",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30018",
   "tg_tokens": null,
   "tps": 86.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 512,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Arc B580",
   "id": "lc30018-b580-gemma4-bf79dbb",
   "model": "Gemma 4 E4B",
   "notes": "bisect point bf79dbb, tg128; config=bisect-bf79dbb",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Gemma-4-E4B; Intel Arc B580; 6716df6=90.7, bf79dbb=90.8, a7fb71f=87.0, 0bb496db=86.2, 2ca15f54=90.7",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30018",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30018",
   "tg_tokens": null,
   "tps": 90.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 512,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Arc B580",
   "id": "lc30018-b580-gemma4-a7fb71f",
   "model": "Gemma 4 E4B",
   "notes": "bisect point a7fb71f, tg128; config=bisect-a7fb71f",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Gemma-4-E4B; Intel Arc B580; 6716df6=90.7, bf79dbb=90.8, a7fb71f=87.0, 0bb496db=86.2, 2ca15f54=90.7",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30018",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30018",
   "tg_tokens": null,
   "tps": 87.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 512,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Arc B580",
   "id": "lc30018-b580-qwen314b-before",
   "model": "Qwen3 14B",
   "notes": "same boundary, unaffected; config=before",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "Qwen3-14B; Intel Arc B580; Qwen3-14B is unaffected across the same boundary (40.0 t/s on both sides)",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30018",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30018",
   "tg_tokens": null,
   "tps": 40.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 512,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Arc B580",
   "id": "lc30018-b580-qwen314b-after",
   "model": "Qwen3 14B",
   "notes": "same boundary, unaffected; config=after",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "Qwen3-14B; Intel Arc B580; Qwen3-14B is unaffected across the same boundary (40.0 t/s on both sides)",
   "ram_gb": null,
   "retrieved": "2026-10-05",
   "source_name": "llama.cpp #30018",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30018",
   "tg_tokens": null,
   "tps": 40.0,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 14,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "8x MI355X",
   "id": "vllm-60110-mi355x-kimik3-before",
   "model": "Kimi-K3",
   "notes": "InferenceX agentic c14, before folding q_b_proj + W_UK into one GEMM; after arm 693.13 tok/s (+1.1%); config=before",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unspecified",
   "quote": "Kimi-K3; 8x MI355X; Output throughput (tok/s) | 685.62 | 693.13 | +1.1%; Mean TTFT (ms) | 340.08 | 336.53 | -1.0%; c14 point",
   "ram_gb": null,
   "retrieved": "2026-10-12",
   "source_name": "vLLM PR #60110",
   "source_url": "https://github.com/vllm-project/vllm/pull/60110",
   "tg_tokens": null,
   "tps": 685.62,
   "ttft_s": 340.08
  },
  {
   "backend": "vLLM",
   "batch": 14,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "8x MI355X",
   "id": "vllm-60110-mi355x-kimik3-after",
   "model": "Kimi-K3",
   "notes": "InferenceX agentic c14, after the folded GEMM; before arm 685.62 tok/s; config=after",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unspecified",
   "quote": "Kimi-K3; 8x MI355X; Output throughput (tok/s) | 685.62 | 693.13 | +1.1%; Mean TTFT (ms) | 340.08 | 336.53 | -1.0%; c14 point",
   "ram_gb": null,
   "retrieved": "2026-10-12",
   "source_name": "vLLM PR #60110",
   "source_url": "https://github.com/vllm-project/vllm/pull/60110",
   "tg_tokens": null,
   "tps": 693.13,
   "ttft_s": 336.53
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 2048,
   "ctx": 32768,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "hf-sakura-swift3-iq3xxs-mtp2-b2048",
   "model": "Sakura-Qwen3.8-Flash-Next-Swift",
   "notes": "32k ctx, KV q8_0, ~1700-token prompt; MTP n-max 2 sidecar, fork build (danielhanchen qwen4exp/mtp 6fcaa16); config=iq3xxs-mtp2-b2048",
   "params": null,
   "power_w": null,
   "pp_tokens": 1700,
   "pp_tps": 314.5,
   "provenance": "community",
   "quant": "IQ3_XXS",
   "quote": "Sakura-Qwen3.8-Flash-Next-Swift; Sakura-Qwen3.8-Flash-Next-Swift-K352-IQ3_XXS-00001-of-00002.gguf; Speed (AMD Strix Halo, 64 GB, Radeon 8060S, Windows, Vulkan); | 2048 | MTP n-max 2 | 314.5 tok/s | **34.5 tok/s** |",
   "ram_gb": 64,
   "retrieved": "2026-10-05",
   "source_name": "HF card webmp3/Sakura-Qwen3.8-Flash-Next-Swift-3bit-352E-GGUF",
   "source_url": "https://huggingface.co/webmp3/Sakura-Qwen3.8-Flash-Next-Swift-3bit-352E-GGUF",
   "tg_tokens": null,
   "tps": 34.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 2048,
   "ctx": 32768,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "hf-sakura-swift3-iq3xxs-mtp3-b2048",
   "model": "Sakura-Qwen3.8-Flash-Next-Swift",
   "notes": "32k ctx, KV q8_0, ~1700-token prompt; MTP n-max 3, fork build; config=iq3xxs-mtp3-b2048",
   "params": null,
   "power_w": null,
   "pp_tokens": 1700,
   "pp_tps": 316.7,
   "provenance": "community",
   "quant": "IQ3_XXS",
   "quote": "Sakura-Qwen3.8-Flash-Next-Swift; Sakura-Qwen3.8-Flash-Next-Swift-K352-IQ3_XXS-00001-of-00002.gguf; Speed (AMD Strix Halo, 64 GB, Radeon 8060S, Windows, Vulkan); | 2048 | MTP n-max 3 | 316.7 tok/s | 33.6 tok/s |",
   "ram_gb": 64,
   "retrieved": "2026-10-05",
   "source_name": "HF card webmp3/Sakura-Qwen3.8-Flash-Next-Swift-3bit-352E-GGUF",
   "source_url": "https://huggingface.co/webmp3/Sakura-Qwen3.8-Flash-Next-Swift-3bit-352E-GGUF",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 512,
   "ctx": 32768,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "hf-sakura-swift3-iq3xxs-mtp3-b512",
   "model": "Sakura-Qwen3.8-Flash-Next-Swift",
   "notes": "32k ctx, KV q8_0, ~1700-token prompt; MTP n-max 3, fork build; config=iq3xxs-mtp3-b512",
   "params": null,
   "power_w": null,
   "pp_tokens": 1700,
   "pp_tps": 257.5,
   "provenance": "community",
   "quant": "IQ3_XXS",
   "quote": "Sakura-Qwen3.8-Flash-Next-Swift; Sakura-Qwen3.8-Flash-Next-Swift-K352-IQ3_XXS-00001-of-00002.gguf; Speed (AMD Strix Halo, 64 GB, Radeon 8060S, Windows, Vulkan); | 512 | MTP n-max 3 | 257.5 tok/s | 33.4 tok/s |",
   "ram_gb": 64,
   "retrieved": "2026-10-05",
   "source_name": "HF card webmp3/Sakura-Qwen3.8-Flash-Next-Swift-3bit-352E-GGUF",
   "source_url": "https://huggingface.co/webmp3/Sakura-Qwen3.8-Flash-Next-Swift-3bit-352E-GGUF",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 512,
   "ctx": 32768,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "hf-sakura-swift3-iq3xxs-mtp2-b512",
   "model": "Sakura-Qwen3.8-Flash-Next-Swift",
   "notes": "32k ctx, KV q8_0, ~1700-token prompt; MTP n-max 2, fork build; config=iq3xxs-mtp2-b512",
   "params": null,
   "power_w": null,
   "pp_tokens": 1700,
   "pp_tps": 253.4,
   "provenance": "community",
   "quant": "IQ3_XXS",
   "quote": "Sakura-Qwen3.8-Flash-Next-Swift; Sakura-Qwen3.8-Flash-Next-Swift-K352-IQ3_XXS-00001-of-00002.gguf; Speed (AMD Strix Halo, 64 GB, Radeon 8060S, Windows, Vulkan); | 512 | MTP n-max 2 | 253.4 tok/s | 32.8 tok/s |",
   "ram_gb": 64,
   "retrieved": "2026-10-05",
   "source_name": "HF card webmp3/Sakura-Qwen3.8-Flash-Next-Swift-3bit-352E-GGUF",
   "source_url": "https://huggingface.co/webmp3/Sakura-Qwen3.8-Flash-Next-Swift-3bit-352E-GGUF",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": 32768,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "hf-sakura-swift3-iq3xxs-official",
   "model": "Sakura-Qwen3.8-Flash-Next-Swift",
   "notes": "official llama.cpp b11259, no speculation, approximate value measured during a coding task; config=iq3xxs-official",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ3_XXS",
   "quote": "Sakura-Qwen3.8-Flash-Next-Swift; Sakura-Qwen3.8-Flash-Next-Swift-K352-IQ3_XXS-00001-of-00002.gguf; Speed (AMD Strix Halo, 64 GB, Radeon 8060S, Windows, Vulkan); Without speculation in the official llama.cpp b11259 (Vulkan) the same file decodes at about 21 tok/s",
   "ram_gb": 64,
   "retrieved": "2026-10-05",
   "source_name": "HF card webmp3/Sakura-Qwen3.8-Flash-Next-Swift-3bit-352E-GGUF",
   "source_url": "https://huggingface.co/webmp3/Sakura-Qwen3.8-Flash-Next-Swift-3bit-352E-GGUF",
   "tg_tokens": null,
   "tps": 21,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "hf-sakura-swift3-iq3xxs-rocm-collapse",
   "model": "Sakura-Qwen3.8-Flash-Next-Swift",
   "notes": "ROCm on Windows, approximate; resident weights (~31 GiB) exceed dedicated GPU memory and decode collapses",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ3_XXS",
   "quote": "Sakura-Qwen3.8-Flash-Next-Swift; Sakura-Qwen3.8-Flash-Next-Swift-K352-IQ3_XXS-00001-of-00002.gguf; on Windows the decode collapses to about 5 tok/s as soon as the resident weights; Speed (AMD Strix Halo, 64 GB, Radeon 8060S, Windows, Vulkan)",
   "ram_gb": 64,
   "retrieved": "2026-10-05",
   "source_name": "HF card webmp3/Sakura-Qwen3.8-Flash-Next-Swift-3bit-352E-GGUF",
   "source_url": "https://huggingface.co/webmp3/Sakura-Qwen3.8-Flash-Next-Swift-3bit-352E-GGUF",
   "tg_tokens": null,
   "tps": 5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 2048,
   "ctx": 32768,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "hf-sakura-swift25-iq2xs-mtp2-b2048",
   "model": "Sakura-Qwen3.8-Flash-Next-Swift",
   "notes": "32k ctx, KV q8_0, ~1700-token prompt; MTP sidecar, fork build (danielhanchen qwen4exp/mtp 6fcaa16); config=iq2xs-mtp2-b2048",
   "params": null,
   "power_w": null,
   "pp_tokens": 1700,
   "pp_tps": 238.0,
   "provenance": "community",
   "quant": "IQ2_XS",
   "quote": "Sakura-Qwen3.8-Flash-Next-Swift; Sakura-Qwen3.8-Flash-Next-Swift-K352-IQ2_XS-00001-of-00002.gguf; Speed (AMD Strix Halo, 64 GB, Radeon 8060S, Windows, Vulkan); | 2048 | MTP n-max 2 | 238.0 tok/s | **37.6 tok/s** |",
   "ram_gb": 64,
   "retrieved": "2026-10-05",
   "source_name": "HF card webmp3/Sakura-Qwen3.8-Flash-Next-Swift-2.5bit-352E-32GB-GGUF",
   "source_url": "https://huggingface.co/webmp3/Sakura-Qwen3.8-Flash-Next-Swift-2.5bit-352E-32GB-GGUF",
   "tg_tokens": null,
   "tps": 37.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 2048,
   "ctx": 32768,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "hf-sakura-swift25-iq2xs-mtp3-b2048",
   "model": "Sakura-Qwen3.8-Flash-Next-Swift",
   "notes": "32k ctx, KV q8_0, ~1700-token prompt; MTP sidecar, fork build (danielhanchen qwen4exp/mtp 6fcaa16); config=iq2xs-mtp3-b2048",
   "params": null,
   "power_w": null,
   "pp_tokens": 1700,
   "pp_tps": 253.3,
   "provenance": "community",
   "quant": "IQ2_XS",
   "quote": "Sakura-Qwen3.8-Flash-Next-Swift; Sakura-Qwen3.8-Flash-Next-Swift-K352-IQ2_XS-00001-of-00002.gguf; Speed (AMD Strix Halo, 64 GB, Radeon 8060S, Windows, Vulkan); | 2048 | MTP n-max 3 | 253.3 tok/s | **35.2 tok/s** |",
   "ram_gb": 64,
   "retrieved": "2026-10-05",
   "source_name": "HF card webmp3/Sakura-Qwen3.8-Flash-Next-Swift-2.5bit-352E-32GB-GGUF",
   "source_url": "https://huggingface.co/webmp3/Sakura-Qwen3.8-Flash-Next-Swift-2.5bit-352E-32GB-GGUF",
   "tg_tokens": null,
   "tps": 35.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": 32768,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "hf-sakura-swift25-iq2xs-official",
   "model": "Sakura-Qwen3.8-Flash-Next-Swift",
   "notes": "official llama.cpp, no speculation, approximate; config=iq2xs-official",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ2_XS",
   "quote": "Sakura-Qwen3.8-Flash-Next-Swift; Sakura-Qwen3.8-Flash-Next-Swift-K352-IQ2_XS-00001-of-00002.gguf; Without speculation the same file decodes at about 24 tok/s in official llama.cpp (Vulkan, b11259 and b11297); Speed (AMD Strix Halo, 64 GB, Radeon 8060S, Windows, Vulkan)",
   "ram_gb": 64,
   "retrieved": "2026-10-05",
   "source_name": "HF card webmp3/Sakura-Qwen3.8-Flash-Next-Swift-2.5bit-352E-32GB-GGUF",
   "source_url": "https://huggingface.co/webmp3/Sakura-Qwen3.8-Flash-Next-Swift-2.5bit-352E-32GB-GGUF",
   "tg_tokens": null,
   "tps": 24,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": 32768,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "hf-sakura-swift25-iq2xs-rocm",
   "model": "Sakura-Qwen3.8-Flash-Next-Swift",
   "notes": "ROCm halo-box build, no speculation; resident weights (~26 GiB) fit dedicated GPU memory so no collapse",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 322.0,
   "provenance": "community",
   "quant": "IQ2_XS",
   "quote": "Sakura-Qwen3.8-Flash-Next-Swift; Sakura-Qwen3.8-Flash-Next-Swift-K352-IQ2_XS-00001-of-00002.gguf; On ROCm (community halo-box build, no speculation) we measured 322 tok/s prefill and 21.6 tok/s decode for this file; Speed (AMD Strix Halo, 64 GB, Radeon 8060S, Windows, Vulkan)",
   "ram_gb": 64,
   "retrieved": "2026-10-05",
   "source_name": "HF card webmp3/Sakura-Qwen3.8-Flash-Next-Swift-2.5bit-352E-32GB-GGUF",
   "source_url": "https://huggingface.co/webmp3/Sakura-Qwen3.8-Flash-Next-Swift-2.5bit-352E-32GB-GGUF",
   "tg_tokens": null,
   "tps": 21.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": 32768,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "hf-sakura-darwinr3-iq3xxs-mtp",
   "model": "Sakura-Qwen3.8-Flash-Next-ISTA-Darwin-R3",
   "notes": "fork build (danielhanchen qwen4exp/mtp 6fcaa16) with shared MTP head, 32k ctx, KV q8_0",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 236.0,
   "provenance": "community",
   "quant": "IQ3_XXS",
   "quote": "Sakura-Qwen3.8-Flash-Next-ISTA-Darwin-R3-K352-IQ3_XXS-00001-of-00002.gguf; | Decode / prefill speed (AMD Strix Halo, 64 GB, Radeon 8060S, Windows, Vulkan, MTP head) | - | 31.4 tok/s decode, 236 tok/s prefill |",
   "ram_gb": 64,
   "retrieved": "2026-10-05",
   "source_name": "HF card webmp3/Sakura-Qwen3.8-Flash-Next-ISTA-Darwin-R3-3bit-352E-GGUF",
   "source_url": "https://huggingface.co/webmp3/Sakura-Qwen3.8-Flash-Next-ISTA-Darwin-R3-3bit-352E-GGUF",
   "tg_tokens": null,
   "tps": 31.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 512,
   "ctx": 32768,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "hf-sakura-ista3-iq3xxs-vulkan",
   "model": "Sakura-Qwen3.8-Flash-Next",
   "notes": "official llama.cpp b11259, no speculation, 32k ctx, KV q8_0, ~1700-token prompt, -b 512 -ub 512",
   "params": null,
   "power_w": null,
   "pp_tokens": 1700,
   "pp_tps": 186.0,
   "provenance": "community",
   "quant": "IQ3_XXS",
   "quote": "Sakura-Qwen3.8-Flash-Next-K352-IQ3_XXS-00001-of-00002.gguf; Speed (AMD Strix Halo, 64 GB, Radeon 8060S, Windows, Vulkan); official llama.cpp b11259 (Vulkan): **22.7 tok/s decode, 186 tok/s prefill**",
   "ram_gb": 64,
   "retrieved": "2026-10-05",
   "source_name": "HF card webmp3/Sakura-Qwen3.8-Flash-Next-ISTA-3bit-352E-GGUF",
   "source_url": "https://huggingface.co/webmp3/Sakura-Qwen3.8-Flash-Next-ISTA-3bit-352E-GGUF",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "hf-sakura-ista3-iq3xxs-rocm",
   "model": "Sakura-Qwen3.8-Flash-Next",
   "notes": "ROCm on Windows, approximate; resident weights (~31 GiB) exceed dedicated GPU memory",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ3_XXS",
   "quote": "Sakura-Qwen3.8-Flash-Next-K352-IQ3_XXS-00001-of-00002.gguf; on Windows the decode collapses to about 4.9 tok/s decode as soon as the resident weights; Speed (AMD Strix Halo, 64 GB, Radeon 8060S, Windows, Vulkan)",
   "ram_gb": 64,
   "retrieved": "2026-10-05",
   "source_name": "HF card webmp3/Sakura-Qwen3.8-Flash-Next-ISTA-3bit-352E-GGUF",
   "source_url": "https://huggingface.co/webmp3/Sakura-Qwen3.8-Flash-Next-ISTA-3bit-352E-GGUF",
   "tg_tokens": null,
   "tps": 4.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": 16384,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "hf-mimo-mopd160-iq2xs-ctx16k",
   "model": "MiMo-V2.6-Flash-MOPD-P160",
   "notes": "published range 9-10.5 tok/s decode, prefill range 13-25 on a 4k prompt, stored at the low end; desktop in parallel use; config=ctx16k",
   "params": null,
   "power_w": null,
   "pp_tokens": 4096,
   "pp_tps": 13.0,
   "provenance": "community",
   "quant": "IQ2_XS",
   "quote": "MiMo-V2.6-Flash-MOPD-P160-OnPolicy-v3la-IQ2_XS.gguf; Speed: AMD Ryzen AI Max+ 395 / Radeon 8060S, 64 GB, Vulkan, MiMo SSD Streaming setup below; | **fast (default)** | 16k | none | **9\u201310.5** | 13\u201325 |",
   "ram_gb": 64,
   "retrieved": "2026-10-05",
   "source_name": "HF card webmp3/Sakura-MiMo-V2.6-Flash-MOPD-P160-GGUF",
   "source_url": "https://huggingface.co/webmp3/Sakura-MiMo-V2.6-Flash-MOPD-P160-GGUF",
   "tg_tokens": null,
   "tps": 9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": 32768,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "hf-mimo-mopd160-iq2xs-ctx32k-lock",
   "model": "MiMo-V2.6-Flash-MOPD-P160",
   "notes": "published range 5.6-7.7 decode, 14-18 prefill, stored at the low end; expert lock on hottest 30%; config=ctx32k-lock",
   "params": null,
   "power_w": null,
   "pp_tokens": 4096,
   "pp_tps": 14.0,
   "provenance": "community",
   "quant": "IQ2_XS",
   "quote": "MiMo-V2.6-Flash-MOPD-P160-OnPolicy-v3la-IQ2_XS.gguf; Speed: AMD Ryzen AI Max+ 395 / Radeon 8060S, 64 GB, Vulkan, MiMo SSD Streaming setup below; | long | 32k | hottest 30 % (~6 GiB) | 5.6\u20137.7 | 14\u201318 |",
   "ram_gb": 64,
   "retrieved": "2026-10-05",
   "source_name": "HF card webmp3/Sakura-MiMo-V2.6-Flash-MOPD-P160-GGUF",
   "source_url": "https://huggingface.co/webmp3/Sakura-MiMo-V2.6-Flash-MOPD-P160-GGUF",
   "tg_tokens": null,
   "tps": 5.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": 24576,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "hf-mimo-mopd160-iq2xs-ctx24k",
   "model": "MiMo-V2.6-Flash-MOPD-P160",
   "notes": "published range 6.6-7.4 decode, stored at the low end; config=ctx24k",
   "params": null,
   "power_w": null,
   "pp_tokens": 4096,
   "pp_tps": 32.0,
   "provenance": "community",
   "quant": "IQ2_XS",
   "quote": "MiMo-V2.6-Flash-MOPD-P160-OnPolicy-v3la-IQ2_XS.gguf; Speed: AMD Ryzen AI Max+ 395 / Radeon 8060S, 64 GB, Vulkan, MiMo SSD Streaming setup below; | (24k, for reference) | 24k | none | 6.6\u20137.4 | 32 |",
   "ram_gb": 64,
   "retrieved": "2026-10-05",
   "source_name": "HF card webmp3/Sakura-MiMo-V2.6-Flash-MOPD-P160-GGUF",
   "source_url": "https://huggingface.co/webmp3/Sakura-MiMo-V2.6-Flash-MOPD-P160-GGUF",
   "tg_tokens": null,
   "tps": 6.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 512,
   "ctx": 16384,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "hf-mimo-mopd160-iq2xs-ub512",
   "model": "MiMo-V2.6-Flash-MOPD-P160",
   "notes": "ubatch sweep at the default 16k fast mode, prefill on a 2k prompt; config=ub512",
   "params": null,
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": 12.7,
   "provenance": "community",
   "quant": "IQ2_XS",
   "quote": "MiMo-V2.6-Flash-MOPD-P160-OnPolicy-v3la-IQ2_XS.gguf; Speed: AMD Ryzen AI Max+ 395 / Radeon 8060S, 64 GB, Vulkan, MiMo SSD Streaming setup below; | **512 (default)** | 12.7 | **9.6** |",
   "ram_gb": 64,
   "retrieved": "2026-10-05",
   "source_name": "HF card webmp3/Sakura-MiMo-V2.6-Flash-MOPD-P160-GGUF",
   "source_url": "https://huggingface.co/webmp3/Sakura-MiMo-V2.6-Flash-MOPD-P160-GGUF",
   "tg_tokens": null,
   "tps": 9.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1024,
   "ctx": 16384,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "hf-mimo-mopd160-iq2xs-ub1024",
   "model": "MiMo-V2.6-Flash-MOPD-P160",
   "notes": "ubatch sweep, prefill on a 2k prompt; GPU buffers spill into shared RAM and nearly halve decode; config=ub1024",
   "params": null,
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": 22.3,
   "provenance": "community",
   "quant": "IQ2_XS",
   "quote": "MiMo-V2.6-Flash-MOPD-P160-OnPolicy-v3la-IQ2_XS.gguf; Speed: AMD Ryzen AI Max+ 395 / Radeon 8060S, 64 GB, Vulkan, MiMo SSD Streaming setup below; | 1024 | 22.3 | 5.9 |",
   "ram_gb": 64,
   "retrieved": "2026-10-05",
   "source_name": "HF card webmp3/Sakura-MiMo-V2.6-Flash-MOPD-P160-GGUF",
   "source_url": "https://huggingface.co/webmp3/Sakura-MiMo-V2.6-Flash-MOPD-P160-GGUF",
   "tg_tokens": null,
   "tps": 5.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 2048,
   "ctx": 16384,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "hf-mimo-mopd160-iq2xs-ub2048",
   "model": "MiMo-V2.6-Flash-MOPD-P160",
   "notes": "ubatch sweep, prefill range 34-39 on a 4k prompt stored at the low end, decode approximate; config=ub2048",
   "params": null,
   "power_w": null,
   "pp_tokens": 4096,
   "pp_tps": 34.0,
   "provenance": "community",
   "quant": "IQ2_XS",
   "quote": "MiMo-V2.6-Flash-MOPD-P160-OnPolicy-v3la-IQ2_XS.gguf; Speed: AMD Ryzen AI Max+ 395 / Radeon 8060S, 64 GB, Vulkan, MiMo SSD Streaming setup below; | 2048 | 34\u201339 (4k prompt) | ~5.3 |",
   "ram_gb": 64,
   "retrieved": "2026-10-05",
   "source_name": "HF card webmp3/Sakura-MiMo-V2.6-Flash-MOPD-P160-GGUF",
   "source_url": "https://huggingface.co/webmp3/Sakura-MiMo-V2.6-Flash-MOPD-P160-GGUF",
   "tg_tokens": null,
   "tps": 5.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": 32768,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "hf-mimo-flashrl-iq2-full-nvme",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "full 512-expert model (91.9 GiB) on the same 64 GB machine, experts mostly paged from NVMe",
   "params": "309B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ2",
   "quote": "MiMo-V2.6-Flash-RL (TrevorJS IQ2, reference); 256 (full model, reference) | 91.9 GiB | 5.04 | 4.23 | \u2014 (2.6 tok/s here, mostly paged from NVMe); Speed: AMD Ryzen AI Max+ 395 / Radeon 8060S, 64 GB, Vulkan, MiMo SSD Streaming setup below",
   "ram_gb": 64,
   "retrieved": "2026-10-05",
   "source_name": "HF card webmp3/Sakura-MiMo-V2.6-Flash-MOPD-P160-GGUF",
   "source_url": "https://huggingface.co/webmp3/Sakura-MiMo-V2.6-Flash-MOPD-P160-GGUF",
   "tg_tokens": null,
   "tps": 2.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "256000",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "RTX 5080 (PCIe Gen 4) + Ryzen 3900X",
   "id": "lc-lid-qwen36-cache-12gb",
   "model": "Qwen3.6-35B-A3B",
   "notes": "expert cache on at the 16 GB RAM + 12 GB VRAM target: 70 experts in VRAM, 186 in pinned RAM, no disk tier; published approximate; config=lid-qwen36-cache-12gb",
   "params": "35B-A3B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_M",
   "quote": "RTX 5080 (PCIe Gen 4), Ryzen 3900X, 32 GB DDR4 RAM; Qwen3.6-35B-A3B-UD-Q4_K_M.gguf; Qwen3.6-35B-A3B-UD-Q4_K_M.gguf; 70/186/0 11,643 MiB ~16 GB ~66 t/s +65%",
   "ram_gb": "32",
   "retrieved": "2026-10-05",
   "source_name": "Lidenburg/llama.cpp README (moe-expert-caching)",
   "source_url": "https://github.com/Lidenburg/llama.cpp/blob/moe-expert-caching/README.md",
   "tg_tokens": "1000",
   "tps": 66,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "256000",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "RTX 5080 (PCIe Gen 4) + Ryzen 3900X",
   "id": "lc-lid-qwen36-ncmoe-12gb",
   "model": "Qwen3.6-35B-A3B",
   "notes": "stock --n-cpu-moe 30 at the 16 GB RAM + 12 GB VRAM target; published approximate; config=lid-qwen36-ncmoe-12gb",
   "params": "35B-A3B",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_M",
   "quote": "RTX 5080 (PCIe Gen 4), Ryzen 3900X, 32 GB DDR4 RAM; Qwen3.6-35B-A3B-UD-Q4_K_M.gguf; Qwen3.6-35B-A3B-UD-Q4_K_M.gguf; -ncmoe 30 11,580 MiB 16.5 GB ~40 t/s",
   "ram_gb": "32",
   "retrieved": "2026-10-05",
   "source_name": "Lidenburg/llama.cpp README (moe-expert-caching)",
   "source_url": "https://github.com/Lidenburg/llama.cpp/blob/moe-expert-caching/README.md",
   "tg_tokens": "1000",
   "tps": 40,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "256000",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "RTX 5080 (PCIe Gen 4) + Ryzen 3900X",
   "id": "lc-lid-codernext-cache-12gb",
   "model": "Qwen3-Coder-Next",
   "notes": "expert cache on at the 32 GB RAM + 12 GB VRAM target: 55 VRAM / 402 RAM / 55 disk experts, disk tier active on the SATA SSD; published approximate; config=lid-codernext-cache-12gb",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "RTX 5080 (PCIe Gen 4), Ryzen 3900X, 32 GB DDR4 RAM; Qwen3-Coder-Next-UD-IQ4_XS.gguf; Qwen3-Coder-Next-UD-IQ4_XS.gguf; 55/402/55 10,988 MiB 28.7 GB ~27 t/s +58%",
   "ram_gb": "32",
   "retrieved": "2026-10-05",
   "source_name": "Lidenburg/llama.cpp README (moe-expert-caching)",
   "source_url": "https://github.com/Lidenburg/llama.cpp/blob/moe-expert-caching/README.md",
   "tg_tokens": "1000",
   "tps": 27,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": "256000",
   "date": "2026-10-05",
   "flags": [],
   "hardware": "RTX 5080 (PCIe Gen 4) + Ryzen 3900X",
   "id": "lc-lid-codernext-ncmoe-12gb",
   "model": "Qwen3-Coder-Next",
   "notes": "stock --n-cpu-moe 41 at the 32 GB RAM + 12 GB VRAM target; published approximate; config=lid-codernext-ncmoe-12gb",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "RTX 5080 (PCIe Gen 4), Ryzen 3900X, 32 GB DDR4 RAM; Qwen3-Coder-Next-UD-IQ4_XS.gguf; Qwen3-Coder-Next-UD-IQ4_XS.gguf; -ncmoe 41 12,162 MiB ~2 GB ~17 t/s",
   "ram_gb": "32",
   "retrieved": "2026-10-05",
   "source_name": "Lidenburg/llama.cpp README (moe-expert-caching)",
   "source_url": "https://github.com/Lidenburg/llama.cpp/blob/moe-expert-caching/README.md",
   "tg_tokens": "1000",
   "tps": 17,
   "ttft_s": null
  },
  {
   "backend": "Splash",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "M6",
   "id": "m6-qwen36-splash",
   "model": "Qwen3.6-35B-A3B-Splash",
   "notes": "Splash engine (DFlash2 speculative drafts); decode by prompt type: prose DE 51.5/59.0, prose EN 69.6, code 135.9, JSON 234.1 tok/s; stored at the prose EN column",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Splash",
   "quote": "Apple M6: 12 CPU cores (2 Super, 4 Performance, 6 Efficiency), 12 GPU cores, Metal 4; | qwen3.6-35b-a3b-splash | MoE, 3B active | Splash | 51.5 | 59.0 | 69.6 | 135.9 | **234.1** |",
   "ram_gb": 32,
   "retrieved": "2026-10-05",
   "source_name": "HDZucht/Mac-Mini-M6-LLM-benchmarks README",
   "source_url": "https://github.com/HDZucht/Mac-Mini-M6-LLM-benchmarks",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "Splash",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "M6",
   "id": "m6-ornith-splash",
   "model": "Ornith-1.5-35B-A3B-Splash",
   "notes": "Splash engine; prose DE 52.6/59.7, prose EN 61.5, code 119.9, JSON 142.6 tok/s; stored at the prose EN column",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Splash",
   "quote": "Apple M6: 12 CPU cores (2 Super, 4 Performance, 6 Efficiency), 12 GPU cores, Metal 4; | ornith-1.5-35b-a3b-splash | MoE, 3B active | Splash | 52.6 | 59.7 | 61.5 | 119.9 | 142.6 |",
   "ram_gb": 32,
   "retrieved": "2026-10-05",
   "source_name": "HDZucht/Mac-Mini-M6-LLM-benchmarks README",
   "source_url": "https://github.com/HDZucht/Mac-Mini-M6-LLM-benchmarks",
   "tg_tokens": null,
   "tps": 61.5,
   "ttft_s": null
  },
  {
   "backend": "LM Studio",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "M6",
   "id": "m6-ornith-q4km",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "LM Studio GGUF baseline vs its Splash arm (+41% prose EN with drafts); prose DE 37.6/41.2, prose EN 43.7, code 59.3, JSON 53.8 tok/s; stored at the prose EN column",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Apple M6: 12 CPU cores (2 Super, 4 Performance, 6 Efficiency), 12 GPU cores, Metal 4; | ornith-1.5-35b-a3b | MoE, 3B active | GGUF Q4_K_M | 37.6 | 41.2 | 43.7 | 59.3 | 53.8 |",
   "ram_gb": 32,
   "retrieved": "2026-10-05",
   "source_name": "HDZucht/Mac-Mini-M6-LLM-benchmarks README",
   "source_url": "https://github.com/HDZucht/Mac-Mini-M6-LLM-benchmarks",
   "tg_tokens": null,
   "tps": 43.7,
   "ttft_s": null
  },
  {
   "backend": "LM Studio",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "M6",
   "id": "m6-nemotron-nano-omni",
   "model": "Nemotron-3-Nano-Omni",
   "notes": "flat 45-46 tok/s across all five prompt types (bandwidth-bound MoE); stored at the prose EN column",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "Apple M6: 12 CPU cores (2 Super, 4 Performance, 6 Efficiency), 12 GPU cores, Metal 4; | nvidia/nemotron-3-nano-omni | MoE, 3B active | | 45.0 | 45.1 | 45.0 | 45.0 | 45.9 |",
   "ram_gb": 32,
   "retrieved": "2026-10-05",
   "source_name": "HDZucht/Mac-Mini-M6-LLM-benchmarks README",
   "source_url": "https://github.com/HDZucht/Mac-Mini-M6-LLM-benchmarks",
   "tg_tokens": null,
   "tps": 45.0,
   "ttft_s": null
  },
  {
   "backend": "LM Studio",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "M6",
   "id": "m6-nemotron35-udq3kxl",
   "model": "Nemotron-3.5-Lightning",
   "notes": "unsloth UD-Q3_K_XL arm, ~7% slower than the Q4_K_M build; flat 42.2-42.5 tok/s; stored at the prose EN column",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q3_K_XL",
   "quote": "Apple M6: 12 CPU cores (2 Super, 4 Performance, 6 Efficiency), 12 GPU cores, Metal 4; | nemotron-3.5-lightning (unsloth) | MoE, 3B active | GGUF UD-Q3_K_XL | 42.3 | 42.2 | 42.3 | 42.4 | 42.5 |",
   "ram_gb": 32,
   "retrieved": "2026-10-05",
   "source_name": "HDZucht/Mac-Mini-M6-LLM-benchmarks README",
   "source_url": "https://github.com/HDZucht/Mac-Mini-M6-LLM-benchmarks",
   "tg_tokens": null,
   "tps": 42.3,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "M6",
   "id": "m6-qwen35-9b-mlx",
   "model": "Qwen 3.5 9B",
   "notes": "dense 9B, flat 30.1-30.8 tok/s across prompt types; stored at the prose EN column",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "Apple M6: 12 CPU cores (2 Super, 4 Performance, 6 Efficiency), 12 GPU cores, Metal 4; | qwen3.5-9b | dense 9B | MLX | 30.6 | 30.8 | 30.3 | 30.1 | 30.3 |",
   "ram_gb": 32,
   "retrieved": "2026-10-05",
   "source_name": "HDZucht/Mac-Mini-M6-LLM-benchmarks README",
   "source_url": "https://github.com/HDZucht/Mac-Mini-M6-LLM-benchmarks",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "Splash",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "M6",
   "id": "m6-qwen38-27b-splash",
   "model": "Qwen3.8-27B",
   "notes": "Splash drafts lift code/JSON 2.6-3.5x over the GGUF arm (17.5/19.3); prose EN 20.2 stored",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Splash",
   "quote": "Apple M6: 12 CPU cores (2 Super, 4 Performance, 6 Efficiency), 12 GPU cores, Metal 4; | qwen3.8-27b-splash | dense 27B | Splash | 15.3 | 17.2 | 20.2 | 45.2 | 67.7 |",
   "ram_gb": 32,
   "retrieved": "2026-10-05",
   "source_name": "HDZucht/Mac-Mini-M6-LLM-benchmarks README",
   "source_url": "https://github.com/HDZucht/Mac-Mini-M6-LLM-benchmarks",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "Splash",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "M6",
   "id": "m6-swift-qwen38-27b-splash",
   "model": "Swift-Qwen3.8-27B",
   "notes": "Swift fine-tune under the same Splash setup, within 4% of the base Splash arm; stored at the prose EN column",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Splash",
   "quote": "Apple M6: 12 CPU cores (2 Super, 4 Performance, 6 Efficiency), 12 GPU cores, Metal 4; | swift-qwen3.8-27b-splash | dense 27B | Splash | 15.4 | 15.9 | 19.5 | 42.9 | 66.5 |",
   "ram_gb": 32,
   "retrieved": "2026-10-05",
   "source_name": "HDZucht/Mac-Mini-M6-LLM-benchmarks README",
   "source_url": "https://github.com/HDZucht/Mac-Mini-M6-LLM-benchmarks",
   "tg_tokens": null,
   "tps": 19.5,
   "ttft_s": null
  },
  {
   "backend": "LM Studio",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "M6",
   "id": "m6-qwen38-27b-q4km",
   "model": "Qwen3.8-27B",
   "notes": "dense 27B GGUF baseline on 32 GB: 11.2-19.3 tok/s; the Splash arm reaches 20.2-67.7; stored at the prose EN column",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Apple M6: 12 CPU cores (2 Super, 4 Performance, 6 Efficiency), 12 GPU cores, Metal 4; | qwen/qwen3.8-27b | dense 27B | GGUF Q4_K_M | 11.2 | 11.4 | 11.8 | 17.5 | 19.3 |",
   "ram_gb": 32,
   "retrieved": "2026-10-05",
   "source_name": "HDZucht/Mac-Mini-M6-LLM-benchmarks README",
   "source_url": "https://github.com/HDZucht/Mac-Mini-M6-LLM-benchmarks",
   "tg_tokens": null,
   "tps": 11.8,
   "ttft_s": null
  },
  {
   "backend": "LM Studio",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-04",
   "flags": [],
   "hardware": "M6",
   "id": "m6-arex2-lmstudio",
   "model": "AREX-2",
   "notes": "agent-style runs with a PubMed search tool, thinking on; published range 6.8-8.0 tok/s stored at the low end",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Apple M6: 12 CPU cores (2 Super, 4 Performance, 6 Efficiency), 12 GPU cores, Metal 4; BAAI AREX-2 (dense 27B, Qwen3.8 fine-tune); | LM Studio, llama.cpp, GGUF Q4_K_M | 6.8\u20138.0 |",
   "ram_gb": 32,
   "retrieved": "2026-10-05",
   "source_name": "HDZucht/Mac-Mini-M6-LLM-benchmarks README",
   "source_url": "https://github.com/HDZucht/Mac-Mini-M6-LLM-benchmarks",
   "tg_tokens": null,
   "tps": 6.8,
   "ttft_s": null
  },
  {
   "backend": "Splash",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-04",
   "flags": [],
   "hardware": "M6",
   "id": "m6-arex2-splash",
   "model": "AREX-2",
   "notes": "same GGUF under Splash with the base-model DFlash2 draft: 3-5x speedup on the fine-tune; published range 22-37 tok/s stored at the low end",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Apple M6: 12 CPU cores (2 Super, 4 Performance, 6 Efficiency), 12 GPU cores, Metal 4; BAAI AREX-2 (dense 27B, Qwen3.8 fine-tune); | Splash 1.2.0 standalone, same GGUF + Qwen3.8-27B DFlash2 draft (selected automatically) | **22\u201337** |",
   "ram_gb": 32,
   "retrieved": "2026-10-05",
   "source_name": "HDZucht/Mac-Mini-M6-LLM-benchmarks README",
   "source_url": "https://github.com/HDZucht/Mac-Mini-M6-LLM-benchmarks",
   "tg_tokens": null,
   "tps": 22,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (PrismML fork)",
   "batch": null,
   "ctx": 32768,
   "date": "2026-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "hf-schiltmans-ft5-none",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "DFlash2 spec-dec on Metal, where the fork's path is slower than no drafter (7.1-8.0 vs 20.7 tok/s); no drafter, plain decode; config=no-drafter",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "PQ2_0",
   "quote": "One Apple M4 Pro, PrismML's fork at `prism-b10754-2459f68` (Metal), the target `Ternary-Bonsai-2-27B-PQ2_0.gguf`; 32,768-token window, and the repository's bench5.py (five prompts, three repetitions, greedy); | none | | 20.70 | 5,706 |",
   "retrieved": "2026-10-06",
   "source_name": "Schiltmans/Ternary-Bonsai-2-27B-DFlash2-ft5-GGUF",
   "source_url": "https://huggingface.co/Schiltmans/Ternary-Bonsai-2-27B-DFlash2-ft5-GGUF",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (PrismML fork)",
   "batch": null,
   "ctx": 32768,
   "date": "2026-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "hf-schiltmans-ft5-ft5-q80",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "DFlash2 spec-dec on Metal, where the fork's path is slower than no drafter (7.1-8.0 vs 20.7 tok/s); ft5 drafter Q8_0, acceptance 0.432; config=ft5-q80",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "PQ2_0",
   "quote": "One Apple M4 Pro, PrismML's fork at `prism-b10754-2459f68` (Metal), the target `Ternary-Bonsai-2-27B-PQ2_0.gguf`; 32,768-token window, and the repository's bench5.py (five prompts, three repetitions, greedy); | ft5 Q8_0 | 0.432 | 7.92 | 5,706 |",
   "retrieved": "2026-10-06",
   "source_name": "Schiltmans/Ternary-Bonsai-2-27B-DFlash2-ft5-GGUF",
   "source_url": "https://huggingface.co/Schiltmans/Ternary-Bonsai-2-27B-DFlash2-ft5-GGUF",
   "tg_tokens": null,
   "tps": 7.92,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (PrismML fork)",
   "batch": null,
   "ctx": 32768,
   "date": "2026-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "hf-schiltmans-ft5-ft5-q4km",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "DFlash2 spec-dec on Metal, where the fork's path is slower than no drafter (7.1-8.0 vs 20.7 tok/s); ft5 drafter Q4_K_M, acceptance 0.427; config=ft5-q4km",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "PQ2_0",
   "quote": "One Apple M4 Pro, PrismML's fork at `prism-b10754-2459f68` (Metal), the target `Ternary-Bonsai-2-27B-PQ2_0.gguf`; 32,768-token window, and the repository's bench5.py (five prompts, three repetitions, greedy); | ft5 Q4_K_M | 0.427 | 7.88 | 5,706 |",
   "retrieved": "2026-10-06",
   "source_name": "Schiltmans/Ternary-Bonsai-2-27B-DFlash2-ft5-GGUF",
   "source_url": "https://huggingface.co/Schiltmans/Ternary-Bonsai-2-27B-DFlash2-ft5-GGUF",
   "tg_tokens": null,
   "tps": 7.88,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (PrismML fork)",
   "batch": null,
   "ctx": 32768,
   "date": "2026-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "hf-schiltmans-ft5-ft5-q2k",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "DFlash2 spec-dec on Metal, where the fork's path is slower than no drafter (7.1-8.0 vs 20.7 tok/s); ft5 drafter Q2_K, acceptance 0.398; config=ft5-q2k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "PQ2_0",
   "quote": "One Apple M4 Pro, PrismML's fork at `prism-b10754-2459f68` (Metal), the target `Ternary-Bonsai-2-27B-PQ2_0.gguf`; 32,768-token window, and the repository's bench5.py (five prompts, three repetitions, greedy); | ft5 Q2_K | 0.398 | 7.36 | 5,706 |",
   "retrieved": "2026-10-06",
   "source_name": "Schiltmans/Ternary-Bonsai-2-27B-DFlash2-ft5-GGUF",
   "source_url": "https://huggingface.co/Schiltmans/Ternary-Bonsai-2-27B-DFlash2-ft5-GGUF",
   "tg_tokens": null,
   "tps": 7.36,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (PrismML fork)",
   "batch": null,
   "ctx": 32768,
   "date": "2026-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "hf-schiltmans-ft5-stock-q80",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "DFlash2 spec-dec on Metal, where the fork's path is slower than no drafter (7.1-8.0 vs 20.7 tok/s); stock z-lab drafter Q8_0, acceptance 0.395; config=stock-q80",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "PQ2_0",
   "quote": "One Apple M4 Pro, PrismML's fork at `prism-b10754-2459f68` (Metal), the target `Ternary-Bonsai-2-27B-PQ2_0.gguf`; 32,768-token window, and the repository's bench5.py (five prompts, three repetitions, greedy); | stock Q8_0 | 0.395 | 7.10 | 5,706 |",
   "retrieved": "2026-10-06",
   "source_name": "Schiltmans/Ternary-Bonsai-2-27B-DFlash2-ft5-GGUF",
   "source_url": "https://huggingface.co/Schiltmans/Ternary-Bonsai-2-27B-DFlash2-ft5-GGUF",
   "tg_tokens": null,
   "tps": 7.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (PrismML fork)",
   "batch": null,
   "ctx": 32768,
   "date": "2026-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "hf-schiltmans-ft5-ft5-q80-repeat",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "DFlash2 spec-dec on Metal, where the fork's path is slower than no drafter (7.1-8.0 vs 20.7 tok/s); ft5 drafter Q8_0, repeated-last variant, acceptance 0.432; config=ft5-q80-repeated",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "PQ2_0",
   "quote": "One Apple M4 Pro, PrismML's fork at `prism-b10754-2459f68` (Metal), the target `Ternary-Bonsai-2-27B-PQ2_0.gguf`; 32,768-token window, and the repository's bench5.py (five prompts, three repetitions, greedy); | ft5 Q8_0, repeated last | 0.432 | 7.96 | 5,706 |",
   "retrieved": "2026-10-06",
   "source_name": "Schiltmans/Ternary-Bonsai-2-27B-DFlash2-ft5-GGUF",
   "source_url": "https://huggingface.co/Schiltmans/Ternary-Bonsai-2-27B-DFlash2-ft5-GGUF",
   "tg_tokens": null,
   "tps": 7.96,
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "hf-webmp3-istadarwin3b-mtp",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Sakura ISTA-Darwin-R3 K352 3-bit build with shared MTP head on the danielhanchen fork; config=mtp-fork",
   "params": null,
   "power_w": null,
   "pp_tps": 236.0,
   "provenance": "community",
   "quant": "IQ3_XXS",
   "quote": "# Sakura \u2014 Qwen3.8-Flash-Next, ISTA 3-bit cut with Darwin-R3 tensors, 352 of 512 experts (experimental); | Decode / prefill speed (AMD Strix Halo, 64 GB, Radeon 8060S, Windows, Vulkan, MTP head) | - | 31.4 tok/s decode, 236 tok/s prefill |",
   "retrieved": "2026-10-06",
   "source_name": "webmp3/Sakura-Qwen3.8-Flash-Next-Swift-2.5bit-352E-32GB-GGUF",
   "source_url": "https://huggingface.co/webmp3/Sakura-Qwen3.8-Flash-Next-ISTA-Darwin-R3-3bit-352E-GGUF",
   "tg_tokens": null,
   "tps": 31.4,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "M4 Pro (48 GB)",
   "id": "hf-aeon7-fp4-text",
   "model": "Qwen3.6-27B-AEON-Ultimate-Uncensored",
   "notes": "greedy single stream, mlx-vlm git main, post-warmup; 16 GB on disk, 17.07 GB peak",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "79",
   "provenance": "community",
   "quant": "mxfp4 (mixed)",
   "quote": "AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored-BF16; MacBook Pro M4 Pro \u00b7 48 GB; Text \u00b7 single stream 15.2 79 604 ms 17.07 GB 16 GB",
   "retrieved": "2026-10-06",
   "source_name": "HF model card: AEON-7 Qwen3.6-27B-AEON FP4",
   "source_url": "https://huggingface.co/AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored-Multimodal-MLX-FP4",
   "tg_tokens": null,
   "tps": "15.2",
   "ttft_s": "604"
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "M4 Pro (48 GB)",
   "id": "hf-aeon7-fp4-img",
   "model": "Qwen3.6-27B-AEON-Ultimate-Uncensored",
   "notes": "image + text multimodal, 274 image tokens, 17.7 GB peak",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "78.5",
   "provenance": "community",
   "quant": "mxfp4 (mixed)",
   "quote": "AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored-BF16; MacBook Pro M4 Pro \u00b7 48 GB; Image + text 15.0 78.5 (274 img tok) 17.7 GB",
   "retrieved": "2026-10-06",
   "source_name": "HF model card: AEON-7 Qwen3.6-27B-AEON FP4",
   "source_url": "https://huggingface.co/AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored-Multimodal-MLX-FP4",
   "tg_tokens": null,
   "tps": "15.0",
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "M4 Pro (48 GB)",
   "id": "hf-aeon7-fp4-mtp-bs2",
   "model": "Qwen3.6-27B-AEON-Ultimate-Uncensored",
   "notes": "native qwen3_5_mtp drafter, draft block size 2, 97.3% draft accept rate; config=mtp-bs2",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "mxfp4 (mixed)",
   "quote": "AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored-BF16; MacBook Pro M4 Pro \u00b7 48 GB; + MTP bs=2 23.5 1.58\u00d7 97.3% 1.97",
   "retrieved": "2026-10-06",
   "source_name": "HF model card: AEON-7 Qwen3.6-27B-AEON FP4",
   "source_url": "https://huggingface.co/AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored-Multimodal-MLX-FP4",
   "tg_tokens": null,
   "tps": "23.5",
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "M4 Pro (48 GB)",
   "id": "hf-aeon7-fp4-mtp-bs3",
   "model": "Qwen3.6-27B-AEON-Ultimate-Uncensored",
   "notes": "native qwen3_5_mtp drafter, draft block size 3 (card's sweet spot), 94.7% draft accept rate, lossless (verified); config=mtp-bs3",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "mxfp4 (mixed)",
   "quote": "AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored-BF16; MacBook Pro M4 Pro \u00b7 48 GB; + MTP bs=3 (sweet spot) 26.5 1.78\u00d7 94.7% 2.89",
   "retrieved": "2026-10-06",
   "source_name": "HF model card: AEON-7 Qwen3.6-27B-AEON FP4",
   "source_url": "https://huggingface.co/AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored-Multimodal-MLX-FP4",
   "tg_tokens": null,
   "tps": "26.5",
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "M4 Pro (48 GB)",
   "id": "hf-aeon7-fp4-mtp-bs4",
   "model": "Qwen3.6-27B-AEON-Ultimate-Uncensored",
   "notes": "native qwen3_5_mtp drafter, draft block size 4, 86.9% draft accept rate drags net throughput back; config=mtp-bs4",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "mxfp4 (mixed)",
   "quote": "AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored-BF16; MacBook Pro M4 Pro \u00b7 48 GB; + MTP bs=4 25.4 1.70\u00d7 86.9% 3.61",
   "retrieved": "2026-10-06",
   "source_name": "HF model card: AEON-7 Qwen3.6-27B-AEON FP4",
   "source_url": "https://huggingface.co/AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored-Multimodal-MLX-FP4",
   "tg_tokens": null,
   "tps": "25.4",
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "M4 Pro (48 GB)",
   "id": "hf-aeon7-fp4-mtp3-math",
   "model": "Qwen3.6-27B-AEON-Ultimate-Uncensored",
   "notes": "draft block size 3 on math prompts, 90.1% draft accept rate; per-category baseline is flat 14.7-15.3 tok/s (bandwidth bound); config=mtp3-math",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "mxfp4 (mixed)",
   "quote": "AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored-BF16; MacBook Pro M4 Pro \u00b7 48 GB; Math 613 ms 14.8 67.3 ms 25.7 38.9 ms 1.73\u00d7 90.1%",
   "retrieved": "2026-10-06",
   "source_name": "HF model card: AEON-7 Qwen3.6-27B-AEON FP4",
   "source_url": "https://huggingface.co/AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored-Multimodal-MLX-FP4",
   "tg_tokens": null,
   "tps": "25.7",
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "M4 Pro (48 GB)",
   "id": "hf-aeon7-fp4-mtp3-code",
   "model": "Qwen3.6-27B-AEON-Ultimate-Uncensored",
   "notes": "draft block size 3 on code prompts, 85.1% draft accept rate; per-category baseline is flat 14.7-15.3 tok/s (bandwidth bound); config=mtp3-code",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "mxfp4 (mixed)",
   "quote": "AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored-BF16; MacBook Pro M4 Pro \u00b7 48 GB; Code 610 ms 15.1 66.2 ms 25.2 39.7 ms 1.67\u00d7 85.1%",
   "retrieved": "2026-10-06",
   "source_name": "HF model card: AEON-7 Qwen3.6-27B-AEON FP4",
   "source_url": "https://huggingface.co/AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored-Multimodal-MLX-FP4",
   "tg_tokens": null,
   "tps": "25.2",
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "M4 Pro (48 GB)",
   "id": "hf-aeon7-fp4-mtp3-knowledge",
   "model": "Qwen3.6-27B-AEON-Ultimate-Uncensored",
   "notes": "draft block size 3 on knowledge prompts, 83.9% draft accept rate; per-category baseline is flat 14.7-15.3 tok/s (bandwidth bound); config=mtp3-knowledge",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "mxfp4 (mixed)",
   "quote": "AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored-BF16; MacBook Pro M4 Pro \u00b7 48 GB; Knowledge 612 ms 15.0 66.7 ms 25.0 40.0 ms 1.67\u00d7 83.9%",
   "retrieved": "2026-10-06",
   "source_name": "HF model card: AEON-7 Qwen3.6-27B-AEON FP4",
   "source_url": "https://huggingface.co/AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored-Multimodal-MLX-FP4",
   "tg_tokens": null,
   "tps": "25.0",
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "M4 Pro (48 GB)",
   "id": "hf-aeon7-fp4-mtp3-reasoning",
   "model": "Qwen3.6-27B-AEON-Ultimate-Uncensored",
   "notes": "draft block size 3 on reasoning prompts, 82.1% draft accept rate; per-category baseline is flat 14.7-15.3 tok/s (bandwidth bound); config=mtp3-reasoning",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "mxfp4 (mixed)",
   "quote": "AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored-BF16; MacBook Pro M4 Pro \u00b7 48 GB; Reasoning 906 ms 14.8 67.8 ms 23.5 42.6 ms 1.59\u00d7 82.1%",
   "retrieved": "2026-10-06",
   "source_name": "HF model card: AEON-7 Qwen3.6-27B-AEON FP4",
   "source_url": "https://huggingface.co/AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored-Multimodal-MLX-FP4",
   "tg_tokens": null,
   "tps": "23.5",
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "M4 Pro (48 GB)",
   "id": "hf-aeon7-fp4-mtp3-creative",
   "model": "Qwen3.6-27B-AEON-Ultimate-Uncensored",
   "notes": "draft block size 3 on creative prompts, 73.9% draft accept rate; per-category baseline is flat 14.7-15.3 tok/s (bandwidth bound); config=mtp3-creative",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "mxfp4 (mixed)",
   "quote": "AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored-BF16; MacBook Pro M4 Pro \u00b7 48 GB; Creative 625 ms 14.7 67.9 ms 23.2 43.1 ms 1.57\u00d7 73.9%",
   "retrieved": "2026-10-06",
   "source_name": "HF model card: AEON-7 Qwen3.6-27B-AEON FP4",
   "source_url": "https://huggingface.co/AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored-Multimodal-MLX-FP4",
   "tg_tokens": null,
   "tps": "23.2",
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "M4 Pro (48 GB)",
   "id": "hf-aeon7-fp4-mtp3-chat",
   "model": "Qwen3.6-27B-AEON-Ultimate-Uncensored",
   "notes": "draft block size 3 on chat prompts, 67.1% draft accept rate; per-category baseline is flat 14.7-15.3 tok/s (bandwidth bound); config=mtp3-chat",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "mxfp4 (mixed)",
   "quote": "AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored-BF16; MacBook Pro M4 Pro \u00b7 48 GB; Chat 603 ms 15.3 65.6 ms 21.8 45.9 ms 1.43\u00d7 67.1%",
   "retrieved": "2026-10-06",
   "source_name": "HF model card: AEON-7 Qwen3.6-27B-AEON FP4",
   "source_url": "https://huggingface.co/AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored-Multimodal-MLX-FP4",
   "tg_tokens": null,
   "tps": "21.8",
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "M4 Pro (48 GB)",
   "id": "hf-aeon7-8bit-text",
   "model": "Qwen3.6-27B-AEON-Ultimate-Uncensored",
   "notes": "8-bit affine build (bf16 state dynamics, vision tower, MTP head), 29.5 GB on disk, 29.85 GB peak; the FP4 sibling decodes 15.2 tok/s at 17 GB on the same card",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "73",
   "provenance": "community",
   "quant": "8bit",
   "quote": "AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored-BF16; MacBook Pro M4 Pro \u00b7 48 GB; MLX-8bit (this build) 8.2 73 659 ms 29.85 GB 29.5 GB",
   "retrieved": "2026-10-06",
   "source_name": "HF model card: AEON-7 Qwen3.6-27B-AEON 8bit",
   "source_url": "https://huggingface.co/AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored-Multimodal-MLX-8bit",
   "tg_tokens": null,
   "tps": "8.2",
   "ttft_s": "659"
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": 179200,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "hf-dankpaws-swift15-mlx-pp25k",
   "model": "Swift 1.5 Qwen3.8-Flash-Next",
   "notes": "mlx-serve 26.10.1, calibrated mixed-precision pack with pooled n-gram prefetch, 32 GB n-gram table memory-mapped, ctx 179200; right column is the Swift llama.cpp IQ3_XXS reference; config=mlx-serve-pp25k",
   "power_w": null,
   "pp_tokens": 25000,
   "pp_tps": "3191",
   "provenance": "community",
   "quant": "4.7bpw (4-bit experts, 8-bit spine)",
   "quote": "UkisAI's Swift 1.5 Qwen3.8-Flash-Next; M5 Ultra 96 GB; Prefill \u00b7 25k prompt 3,191 tok/s 1,427 tok/s",
   "retrieved": "2026-10-06",
   "source_name": "HF model card: Dankpaws Swift1.5-Qwen3.8-Flash-Next-MLX-4.7bpw",
   "source_url": "https://huggingface.co/Dankpaws/Swift1.5-Qwen3.8-Flash-Next-MLX-4.7bpw",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": 179200,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "hf-dankpaws-swift15-mlx-pp95k",
   "model": "Swift 1.5 Qwen3.8-Flash-Next",
   "notes": "mlx-serve 26.10.1, 95k prompt prefill; right column is the Swift llama.cpp IQ3_XXS reference; config=mlx-serve-pp95k",
   "power_w": null,
   "pp_tokens": 95000,
   "pp_tps": "2928",
   "provenance": "community",
   "quant": "4.7bpw (4-bit experts, 8-bit spine)",
   "quote": "UkisAI's Swift 1.5 Qwen3.8-Flash-Next; M5 Ultra 96 GB; Prefill \u00b7 95k prompt 2,928 tok/s 1,307 tok/s",
   "retrieved": "2026-10-06",
   "source_name": "HF model card: Dankpaws Swift1.5-Qwen3.8-Flash-Next-MLX-4.7bpw",
   "source_url": "https://huggingface.co/Dankpaws/Swift1.5-Qwen3.8-Flash-Next-MLX-4.7bpw",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": 179200,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "hf-dankpaws-swift15-mlx-d4k",
   "model": "Swift 1.5 Qwen3.8-Flash-Next",
   "notes": "mlx-serve 26.10.1, decode after 4k prompt; right column is the Swift llama.cpp IQ3_XXS reference; config=mlx-serve-d4k",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "4.7bpw (4-bit experts, 8-bit spine)",
   "quote": "UkisAI's Swift 1.5 Qwen3.8-Flash-Next; M5 Ultra 96 GB; Decode \u00b7 after 4k prompt 113.7 tok/s 62.7 tok/s",
   "retrieved": "2026-10-06",
   "source_name": "HF model card: Dankpaws Swift1.5-Qwen3.8-Flash-Next-MLX-4.7bpw",
   "source_url": "https://huggingface.co/Dankpaws/Swift1.5-Qwen3.8-Flash-Next-MLX-4.7bpw",
   "tg_tokens": null,
   "tps": "113.7",
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": 179200,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "hf-dankpaws-swift15-mlx-d95k",
   "model": "Swift 1.5 Qwen3.8-Flash-Next",
   "notes": "mlx-serve 26.10.1, decode after 95k prompt; right column is the Swift llama.cpp IQ3_XXS reference; config=mlx-serve-d95k",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "4.7bpw (4-bit experts, 8-bit spine)",
   "quote": "UkisAI's Swift 1.5 Qwen3.8-Flash-Next; M5 Ultra 96 GB; Decode \u00b7 after 95k prompt 81.1 tok/s 44.7 tok/s",
   "retrieved": "2026-10-06",
   "source_name": "HF model card: Dankpaws Swift1.5-Qwen3.8-Flash-Next-MLX-4.7bpw",
   "source_url": "https://huggingface.co/Dankpaws/Swift1.5-Qwen3.8-Flash-Next-MLX-4.7bpw",
   "tg_tokens": null,
   "tps": "81.1",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": 179200,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "hf-dankpaws-swift15-ic3-pp25k",
   "model": "Swift 1.5 Qwen3.8-Flash-Next",
   "notes": "Swift llama.cpp IQ3_XXS reference column, 25k prompt prefill, same M5 Ultra 96 GB Mac Studio; top-1 agreement with Swift BF16 84.1% vs 91.0% for the MLX pack; config=llamacpp-ic3xxs-pp25k",
   "power_w": null,
   "pp_tokens": 25000,
   "pp_tps": "1427",
   "provenance": "community",
   "quant": "IQ3_XXS",
   "quote": "UkisAI's Swift 1.5 Qwen3.8-Flash-Next; M5 Ultra 96 GB; Prefill \u00b7 25k prompt 3,191 tok/s 1,427 tok/s",
   "retrieved": "2026-10-06",
   "source_name": "HF model card: Dankpaws Swift1.5-Qwen3.8-Flash-Next-MLX-4.7bpw",
   "source_url": "https://huggingface.co/Dankpaws/Swift1.5-Qwen3.8-Flash-Next-MLX-4.7bpw",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": 179200,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "hf-dankpaws-swift15-ic3-pp95k",
   "model": "Swift 1.5 Qwen3.8-Flash-Next",
   "notes": "Swift llama.cpp IQ3_XXS reference column, 95k prompt prefill; config=llamacpp-ic3xxs-pp95k",
   "power_w": null,
   "pp_tokens": 95000,
   "pp_tps": "1307",
   "provenance": "community",
   "quant": "IQ3_XXS",
   "quote": "UkisAI's Swift 1.5 Qwen3.8-Flash-Next; M5 Ultra 96 GB; Prefill \u00b7 95k prompt 2,928 tok/s 1,307 tok/s",
   "retrieved": "2026-10-06",
   "source_name": "HF model card: Dankpaws Swift1.5-Qwen3.8-Flash-Next-MLX-4.7bpw",
   "source_url": "https://huggingface.co/Dankpaws/Swift1.5-Qwen3.8-Flash-Next-MLX-4.7bpw",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": 179200,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "hf-dankpaws-swift15-ic3-d4k",
   "model": "Swift 1.5 Qwen3.8-Flash-Next",
   "notes": "Swift llama.cpp IQ3_XXS reference column, decode after 4k prompt; config=llamacpp-ic3xxs-d4k",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ3_XXS",
   "quote": "UkisAI's Swift 1.5 Qwen3.8-Flash-Next; M5 Ultra 96 GB; Decode \u00b7 after 4k prompt 113.7 tok/s 62.7 tok/s",
   "retrieved": "2026-10-06",
   "source_name": "HF model card: Dankpaws Swift1.5-Qwen3.8-Flash-Next-MLX-4.7bpw",
   "source_url": "https://huggingface.co/Dankpaws/Swift1.5-Qwen3.8-Flash-Next-MLX-4.7bpw",
   "tg_tokens": null,
   "tps": "62.7",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": 179200,
   "date": "2026-10-05",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "hf-dankpaws-swift15-ic3-d95k",
   "model": "Swift 1.5 Qwen3.8-Flash-Next",
   "notes": "Swift llama.cpp IQ3_XXS reference column, decode after 95k prompt; config=llamacpp-ic3xxs-d95k",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ3_XXS",
   "quote": "UkisAI's Swift 1.5 Qwen3.8-Flash-Next; M5 Ultra 96 GB; Decode \u00b7 after 95k prompt 81.1 tok/s 44.7 tok/s",
   "retrieved": "2026-10-06",
   "source_name": "HF model card: Dankpaws Swift1.5-Qwen3.8-Flash-Next-MLX-4.7bpw",
   "source_url": "https://huggingface.co/Dankpaws/Swift1.5-Qwen3.8-Flash-Next-MLX-4.7bpw",
   "tg_tokens": null,
   "tps": "44.7",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 2,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "4\u00d7 MI325X",
   "id": "vllm-60153-mi325x4-off-262k",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "gfx942 mHC seam kernel routing (off arm), 262144-in / 1024-out, vllm bench serve c=2, seed 7, language model only; config=seam-off-262k",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "4\u00d7 MI325X (gfx942), TP4; DeepSeek-V4.1-Flash; Concurrency 2; 262144\u21921024 Off 32.84 37310.48 24.48 25.14",
   "ram_gb": null,
   "retrieved": "2026-10-12",
   "source_name": "vLLM PR #60153",
   "source_url": "https://github.com/vllm-project/vllm/pull/60153",
   "tg_tokens": null,
   "tps": 32.84,
   "ttft_s": 37310.48
  },
  {
   "backend": "vLLM",
   "batch": 2,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "4\u00d7 MI325X",
   "id": "vllm-60153-mi325x4-on-262k",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "gfx942 mHC seam kernel routing (on arm, mhc_seam_v3_b2_stnt for steps >=40 tokens), 262144-in / 1024-out, vllm bench serve c=2; config=seam-on-262k",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "4\u00d7 MI325X (gfx942), TP4; DeepSeek-V4.1-Flash; Concurrency 2; On 33.58 36063.46 24.35 26.51",
   "ram_gb": null,
   "retrieved": "2026-10-12",
   "source_name": "vLLM PR #60153",
   "source_url": "https://github.com/vllm-project/vllm/pull/60153",
   "tg_tokens": null,
   "tps": 33.58,
   "ttft_s": 36063.46
  },
  {
   "backend": "vLLM",
   "batch": 2,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "4\u00d7 MI325X",
   "id": "vllm-60153-mi325x4-off-8k",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "gfx942 mHC seam kernel routing (off arm), 8192-in / 1024-out, vllm bench serve c=2; config=seam-off-8k",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "4\u00d7 MI325X (gfx942), TP4; DeepSeek-V4.1-Flash; Concurrency 2; 8192\u21921024 Off 155.45 1101.50 11.80 11.80",
   "ram_gb": null,
   "retrieved": "2026-10-12",
   "source_name": "vLLM PR #60153",
   "source_url": "https://github.com/vllm-project/vllm/pull/60153",
   "tg_tokens": null,
   "tps": 155.45,
   "ttft_s": 1101.5
  },
  {
   "backend": "vLLM",
   "batch": 2,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "4\u00d7 MI325X",
   "id": "vllm-60153-mi325x4-on-8k",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "gfx942 mHC seam kernel routing (on arm), 8192-in / 1024-out, vllm bench serve c=2; config=seam-on-8k",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "4\u00d7 MI325X (gfx942), TP4; DeepSeek-V4.1-Flash; Concurrency 2; On 155.50 1070.71 11.83 12.03",
   "ram_gb": null,
   "retrieved": "2026-10-12",
   "source_name": "vLLM PR #60153",
   "source_url": "https://github.com/vllm-project/vllm/pull/60153",
   "tg_tokens": null,
   "tps": 155.5,
   "ttft_s": 1070.71
  },
  {
   "backend": "MLX",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max (128 GB)",
   "id": "hf-ddalcu-q38fn-mlx-serial",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mlx-serve 26.8.11, ~75 GB resident, prefix cache on, serial decode; config=mlx-serve-serial",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "4-bit experts, 8-bit rest",
   "quote": "Qwen3.8-Flash-Next for mlx-serve (4-bit experts, 8-bit rest); Measured on an M4 Max 128 GB with mlx-serve 26.8.11: ~75 GB resident, decode ~60 tok/s serial and 78 tok/s with MTP (--mtp; +41% on code, a few percent slower on prose), prefill ~730 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "HF model card: ddalcu Qwen3.8-Flash-Next-MLX-Serve-mixed-4-8bit",
   "source_url": "https://huggingface.co/ddalcu/Qwen3.8-Flash-Next-MLX-Serve-mixed-4-8bit",
   "tg_tokens": null,
   "tps": 60,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max (128 GB)",
   "id": "hf-ddalcu-q38fn-mlx-mtp",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mlx-serve 26.8.11, MTP head via --mtp (opt-in speculative path), +41% on code, a few percent slower on prose; config=mlx-serve-mtp",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "4-bit experts, 8-bit rest",
   "quote": "Qwen3.8-Flash-Next for mlx-serve (4-bit experts, 8-bit rest); Measured on an M4 Max 128 GB with mlx-serve 26.8.11: ~75 GB resident, decode ~60 tok/s serial and 78 tok/s with MTP (--mtp; +41% on code, a few percent slower on prose), prefill ~730 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "HF model card: ddalcu Qwen3.8-Flash-Next-MLX-Serve-mixed-4-8bit",
   "source_url": "https://huggingface.co/ddalcu/Qwen3.8-Flash-Next-MLX-Serve-mixed-4-8bit",
   "tg_tokens": null,
   "tps": 78,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max (128 GB)",
   "id": "hf-ddalcu-q38fn-mlx-prefill",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mlx-serve 26.8.11, prefill; needle at 24.8k tokens recovered with sparse attention engaged; config=mlx-serve-prefill",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 730,
   "provenance": "community",
   "quant": "4-bit experts, 8-bit rest",
   "quote": "Qwen3.8-Flash-Next for mlx-serve (4-bit experts, 8-bit rest); Measured on an M4 Max 128 GB with mlx-serve 26.8.11: ~75 GB resident, decode ~60 tok/s serial and 78 tok/s with MTP (--mtp; +41% on code, a few percent slower on prose), prefill ~730 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "HF model card: ddalcu Qwen3.8-Flash-Next-MLX-Serve-mixed-4-8bit",
   "source_url": "https://huggingface.co/ddalcu/Qwen3.8-Flash-Next-MLX-Serve-mixed-4-8bit",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": 1,
   "ctx": 196608,
   "date": "2026-10",
   "flags": [],
   "hardware": "2x Intel B70",
   "id": "lc30033-b70x2-29612",
   "model": "Qwen3.8-Flash-Next",
   "notes": "b11400 0bb496d, llama-server via llama-swap, -dev SYCL0,SYCL1 -fa on -fit off -ts 1,1 -sm layer -ctxcp 3 -ngl 999 -c 196608, VM with 32 Ryzen 9 5950X threads; measured at PR #29612, one commit before the regression; config=pre-29622",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 420,
   "provenance": "community",
   "quant": "UD-IQ3_XXS",
   "quote": "2x Intel B70; unsloth/Qwen3.8-Flash-Next-GGUF:UD-IQ3_XXS; PR #29612 (one previous to this): pp: ~420t/s tg: ~32t/s; PR #29622 (this one): pp: ~365t/s tg: ~24t/s",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp issue #30033",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30033",
   "tg_tokens": null,
   "tps": 32,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": 1,
   "ctx": 196608,
   "date": "2026-10",
   "flags": [],
   "hardware": "2x Intel B70",
   "id": "lc30033-b70x2-29622",
   "model": "Qwen3.8-Flash-Next",
   "notes": "same machine and flags at PR #29622 (first bad commit 0bb496d): ~25% tg and ~13% pp regression, observed through PR #29971; config=post-29622",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 365,
   "provenance": "community",
   "quant": "UD-IQ3_XXS",
   "quote": "2x Intel B70; unsloth/Qwen3.8-Flash-Next-GGUF:UD-IQ3_XXS; PR #29612 (one previous to this): pp: ~420t/s tg: ~32t/s; PR #29622 (this one): pp: ~365t/s tg: ~24t/s",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp issue #30033",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30033",
   "tg_tokens": null,
   "tps": 24,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "B300",
   "id": "vllm-60158-b300-diffgemma-triton",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "Triton fallback (main), diffusion-generative workload (8k input, 1024 output positions, 96 requests), block tok/s as published, ~20 denoising steps per canvas, GSM8K 126/128; config=triton-main",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "BF16",
   "quote": "DiffusionGemma-26B-A4B, BF16, TP1, one B300, concurrency 1; Triton (main) ~947 1.36 s 0.55 s 126/128; FA4 hd512 (this PR) 1152\u20131206 (3 runs) 1.03\u20131.07 s 0.39\u20130.42 s 125/128",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "vLLM PR #60158",
   "source_url": "https://github.com/vllm-project/vllm/pull/60158",
   "tg_tokens": null,
   "tps": 947,
   "ttft_s": 0.55
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "B300",
   "id": "vllm-60158-b300-diffgemma-fa4",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "FA4 hd512 kernel for the Blackwell global-attention layers (d=dv=512, SM100/SM110); published range 1152-1206 block tok/s over 3 runs, stored at low end; e2e 1.03-1.07 s, first block 0.39-0.42 s, GSM8K 125/128 (one-question flip, within noise); config=fa4-hd512",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "BF16",
   "quote": "DiffusionGemma-26B-A4B, BF16, TP1, one B300, concurrency 1; Triton (main) ~947 1.36 s 0.55 s 126/128; FA4 hd512 (this PR) 1152\u20131206 (3 runs) 1.03\u20131.07 s 0.39\u20130.42 s 125/128",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "vLLM PR #60158",
   "source_url": "https://github.com/vllm-project/vllm/pull/60158",
   "tg_tokens": null,
   "tps": 1152,
   "ttft_s": 0.39
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "8x MI355X",
   "id": "vllm-60159-mi355x8-triton-60k",
   "model": "Kimi-K3",
   "notes": "Kimi-K3 on 8x MI355X TP=8, FP8 KV cache, --load-format dummy, concurrency 16; output tok/s and mean TTFT from the profiled serving run; merge_attn_states native ROCm kernel vs Triton fallback is the A/B (merge kernel 4.1x faster at 60k, 4.3x at 128k, but 0.26-0.42% of total kernel time so serving metrics move within noise); 60k-in/600-out; config=triton-60k",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "Kimi-K3, 8x MI355X, TP=8, FP8 KV cache, --load-format dummy, concurrency 16; 60k/600 Triton 1176 118.90 ms 7280.80 3337.37 89.97 152.49; 60k/600 native 1176 29.30 ms 7291.12 3231.90 89.54 153.09; 128k/1k Triton 3744 424.50 ms 17272.42 9884.53 111.25 122.19; 128k/1k native 3792 98.31 ms 17312.48 10211.05 110.82 122.54",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "vLLM PR #60159",
   "source_url": "https://github.com/vllm-project/vllm/pull/60159",
   "tg_tokens": null,
   "tps": 152.49,
   "ttft_s": 7280.8
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "8x MI355X",
   "id": "vllm-60159-mi355x8-native-60k",
   "model": "Kimi-K3",
   "notes": "Kimi-K3 on 8x MI355X TP=8, FP8 KV cache, --load-format dummy, concurrency 16; output tok/s and mean TTFT from the profiled serving run; merge_attn_states native ROCm kernel vs Triton fallback is the A/B (merge kernel 4.1x faster at 60k, 4.3x at 128k, but 0.26-0.42% of total kernel time so serving metrics move within noise); 60k-in/600-out; config=native-60k",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "Kimi-K3, 8x MI355X, TP=8, FP8 KV cache, --load-format dummy, concurrency 16; 60k/600 Triton 1176 118.90 ms 7280.80 3337.37 89.97 152.49; 60k/600 native 1176 29.30 ms 7291.12 3231.90 89.54 153.09; 128k/1k Triton 3744 424.50 ms 17272.42 9884.53 111.25 122.19; 128k/1k native 3792 98.31 ms 17312.48 10211.05 110.82 122.54",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "vLLM PR #60159",
   "source_url": "https://github.com/vllm-project/vllm/pull/60159",
   "tg_tokens": null,
   "tps": 153.09,
   "ttft_s": 7291.12
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "8x MI355X",
   "id": "vllm-60159-mi355x8-triton-128k",
   "model": "Kimi-K3",
   "notes": "Kimi-K3 on 8x MI355X TP=8, FP8 KV cache, --load-format dummy, concurrency 16; output tok/s and mean TTFT from the profiled serving run; merge_attn_states native ROCm kernel vs Triton fallback is the A/B (merge kernel 4.1x faster at 60k, 4.3x at 128k, but 0.26-0.42% of total kernel time so serving metrics move within noise); 128k-in/1k-out; config=triton-128k",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "Kimi-K3, 8x MI355X, TP=8, FP8 KV cache, --load-format dummy, concurrency 16; 60k/600 Triton 1176 118.90 ms 7280.80 3337.37 89.97 152.49; 60k/600 native 1176 29.30 ms 7291.12 3231.90 89.54 153.09; 128k/1k Triton 3744 424.50 ms 17272.42 9884.53 111.25 122.19; 128k/1k native 3792 98.31 ms 17312.48 10211.05 110.82 122.54",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "vLLM PR #60159",
   "source_url": "https://github.com/vllm-project/vllm/pull/60159",
   "tg_tokens": null,
   "tps": 122.19,
   "ttft_s": 17272.42
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "8x MI355X",
   "id": "vllm-60159-mi355x8-native-128k",
   "model": "Kimi-K3",
   "notes": "Kimi-K3 on 8x MI355X TP=8, FP8 KV cache, --load-format dummy, concurrency 16; output tok/s and mean TTFT from the profiled serving run; merge_attn_states native ROCm kernel vs Triton fallback is the A/B (merge kernel 4.1x faster at 60k, 4.3x at 128k, but 0.26-0.42% of total kernel time so serving metrics move within noise); 128k-in/1k-out; config=native-128k",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "Kimi-K3, 8x MI355X, TP=8, FP8 KV cache, --load-format dummy, concurrency 16; 60k/600 Triton 1176 118.90 ms 7280.80 3337.37 89.97 152.49; 60k/600 native 1176 29.30 ms 7291.12 3231.90 89.54 153.09; 128k/1k Triton 3744 424.50 ms 17272.42 9884.53 111.25 122.19; 128k/1k native 3792 98.31 ms 17312.48 10211.05 110.82 122.54",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "vLLM PR #60159",
   "source_url": "https://github.com/vllm-project/vllm/pull/60159",
   "tg_tokens": null,
   "tps": 122.54,
   "ttft_s": 17312.48
  },
  {
   "backend": "vLLM",
   "batch": 256,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "4x NVIDIA GB300",
   "id": "vllm-60122-gb300x4-dsv31-main-mixed",
   "model": "DeepSeek-V3.1",
   "notes": "DeepSeek-V3.1 on 4x GB300 TP=4, TRITON_MLA, vllm bench serve --request-rate inf; A/B is main at 182247257f versus the batch-invariant TRITON_MLA PR (per-row prefill cost on this 128-head model; gsm8k strict 0.9553 -> 0.9591, unchanged within noise); 1024-in/128-out, 256 prompts at 256 concurrency; output tok/s and mean TTFT; config=main-mixed",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "Blackwell results, as promised in the description: 4x NVIDIA GB300 (SM 10.3), tensor-parallel 4, deepseek-ai/DeepSeek-V3.1; Output token throughput (tok/s): 1168.47; Mean TTFT (ms): 4434.90; Output token throughput (tok/s): 982.28; Mean TTFT (ms): 5609.71; Total token throughput (tok/s): 10575.95; Mean TTFT (ms): 50052.37; Total token throughput (tok/s): 3083.34; Mean TTFT (ms): 171527.46",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "vLLM PR #60122",
   "source_url": "https://github.com/vllm-project/vllm/pull/60122",
   "tg_tokens": null,
   "tps": 1168.47,
   "ttft_s": 4434.9
  },
  {
   "backend": "vLLM",
   "batch": 256,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "4x NVIDIA GB300",
   "id": "vllm-60122-gb300x4-dsv31-pr-mixed",
   "model": "DeepSeek-V3.1",
   "notes": "DeepSeek-V3.1 on 4x GB300 TP=4, TRITON_MLA, vllm bench serve --request-rate inf; A/B is main at 182247257f versus the batch-invariant TRITON_MLA PR (per-row prefill cost on this 128-head model; gsm8k strict 0.9553 -> 0.9591, unchanged within noise); 1024-in/128-out, 256 prompts at 256 concurrency; output tok/s and mean TTFT; config=pr-mixed",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "Blackwell results, as promised in the description: 4x NVIDIA GB300 (SM 10.3), tensor-parallel 4, deepseek-ai/DeepSeek-V3.1; Output token throughput (tok/s): 1168.47; Mean TTFT (ms): 4434.90; Output token throughput (tok/s): 982.28; Mean TTFT (ms): 5609.71; Total token throughput (tok/s): 10575.95; Mean TTFT (ms): 50052.37; Total token throughput (tok/s): 3083.34; Mean TTFT (ms): 171527.46",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "vLLM PR #60122",
   "source_url": "https://github.com/vllm-project/vllm/pull/60122",
   "tg_tokens": null,
   "tps": 982.28,
   "ttft_s": 5609.71
  },
  {
   "backend": "vLLM",
   "batch": 256,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "4x NVIDIA GB300",
   "id": "vllm-60122-gb300x4-dsv31-main-prefill",
   "model": "DeepSeek-V3.1",
   "notes": "DeepSeek-V3.1 on 4x GB300 TP=4, TRITON_MLA, vllm bench serve --request-rate inf; A/B is main at 182247257f versus the batch-invariant TRITON_MLA PR (per-row prefill cost on this 128-head model; gsm8k strict 0.9553 -> 0.9591, unchanged within noise); 8192-in/1-out, 128 prompts; total token throughput as the prefill rate; config=main-prefill",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 10575.95,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "Blackwell results, as promised in the description: 4x NVIDIA GB300 (SM 10.3), tensor-parallel 4, deepseek-ai/DeepSeek-V3.1; Output token throughput (tok/s): 1168.47; Mean TTFT (ms): 4434.90; Output token throughput (tok/s): 982.28; Mean TTFT (ms): 5609.71; Total token throughput (tok/s): 10575.95; Mean TTFT (ms): 50052.37; Total token throughput (tok/s): 3083.34; Mean TTFT (ms): 171527.46",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "vLLM PR #60122",
   "source_url": "https://github.com/vllm-project/vllm/pull/60122",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": 50052.37
  },
  {
   "backend": "vLLM",
   "batch": 256,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "4x NVIDIA GB300",
   "id": "vllm-60122-gb300x4-dsv31-pr-prefill",
   "model": "DeepSeek-V3.1",
   "notes": "DeepSeek-V3.1 on 4x GB300 TP=4, TRITON_MLA, vllm bench serve --request-rate inf; A/B is main at 182247257f versus the batch-invariant TRITON_MLA PR (per-row prefill cost on this 128-head model; gsm8k strict 0.9553 -> 0.9591, unchanged within noise); 8192-in/1-out, 128 prompts; total token throughput as the prefill rate; config=pr-prefill",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3083.34,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "Blackwell results, as promised in the description: 4x NVIDIA GB300 (SM 10.3), tensor-parallel 4, deepseek-ai/DeepSeek-V3.1; Output token throughput (tok/s): 1168.47; Mean TTFT (ms): 4434.90; Output token throughput (tok/s): 982.28; Mean TTFT (ms): 5609.71; Total token throughput (tok/s): 10575.95; Mean TTFT (ms): 50052.37; Total token throughput (tok/s): 3083.34; Mean TTFT (ms): 171527.46",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "vLLM PR #60122",
   "source_url": "https://github.com/vllm-project/vllm/pull/60122",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": 171527.46
  },
  {
   "backend": "vLLM",
   "batch": 32,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "64 H100",
   "id": "vllm-60161-h100x64-c32-pr",
   "model": "GLM-5.3",
   "notes": "GLM-5.3 HiSparse decode on 64 H100 (4P/4D, DP4 x TP8, 288 GiB shared host pool, DeepEP LL, cc_traces 256k, 900 s/level); A/B is VLLM_NIXL_HISPARSE_DRAM_RANK0_ONLY (this PR) versus stock; out tok/s and TTFT p99 as published; concurrency 32; config=pr-c32",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "End-to-end benchmark, 4P/4D, 64 H100; Results (llmd-roce cluster, GLM-5.3, DeepEP LL decode, 8 NICs); 32 602 vs 503 (+19.7%) 21.6 vs 89.0 s; 64 547 vs 563 98.2 vs 87.5 s; 128 482 vs 520 194 vs 183 s",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "vLLM PR #60161",
   "source_url": "https://github.com/vllm-project/vllm/pull/60161",
   "tg_tokens": null,
   "tps": 602,
   "ttft_s": 21.6
  },
  {
   "backend": "vLLM",
   "batch": 32,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "64 H100",
   "id": "vllm-60161-h100x64-c32-stock",
   "model": "GLM-5.3",
   "notes": "GLM-5.3 HiSparse decode on 64 H100 (4P/4D, DP4 x TP8, 288 GiB shared host pool, DeepEP LL, cc_traces 256k, 900 s/level); A/B is VLLM_NIXL_HISPARSE_DRAM_RANK0_ONLY (this PR) versus stock; out tok/s and TTFT p99 as published; concurrency 32; config=stock-c32",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "End-to-end benchmark, 4P/4D, 64 H100; Results (llmd-roce cluster, GLM-5.3, DeepEP LL decode, 8 NICs); 32 602 vs 503 (+19.7%) 21.6 vs 89.0 s; 64 547 vs 563 98.2 vs 87.5 s; 128 482 vs 520 194 vs 183 s",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "vLLM PR #60161",
   "source_url": "https://github.com/vllm-project/vllm/pull/60161",
   "tg_tokens": null,
   "tps": 503,
   "ttft_s": 89.0
  },
  {
   "backend": "vLLM",
   "batch": 64,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "64 H100",
   "id": "vllm-60161-h100x64-c64-pr",
   "model": "GLM-5.3",
   "notes": "GLM-5.3 HiSparse decode on 64 H100 (4P/4D, DP4 x TP8, 288 GiB shared host pool, DeepEP LL, cc_traces 256k, 900 s/level); A/B is VLLM_NIXL_HISPARSE_DRAM_RANK0_ONLY (this PR) versus stock; out tok/s and TTFT p99 as published; concurrency 64; config=pr-c64",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "End-to-end benchmark, 4P/4D, 64 H100; Results (llmd-roce cluster, GLM-5.3, DeepEP LL decode, 8 NICs); 32 602 vs 503 (+19.7%) 21.6 vs 89.0 s; 64 547 vs 563 98.2 vs 87.5 s; 128 482 vs 520 194 vs 183 s",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "vLLM PR #60161",
   "source_url": "https://github.com/vllm-project/vllm/pull/60161",
   "tg_tokens": null,
   "tps": 547,
   "ttft_s": 98.2
  },
  {
   "backend": "vLLM",
   "batch": 64,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "64 H100",
   "id": "vllm-60161-h100x64-c64-stock",
   "model": "GLM-5.3",
   "notes": "GLM-5.3 HiSparse decode on 64 H100 (4P/4D, DP4 x TP8, 288 GiB shared host pool, DeepEP LL, cc_traces 256k, 900 s/level); A/B is VLLM_NIXL_HISPARSE_DRAM_RANK0_ONLY (this PR) versus stock; out tok/s and TTFT p99 as published; concurrency 64; config=stock-c64",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "End-to-end benchmark, 4P/4D, 64 H100; Results (llmd-roce cluster, GLM-5.3, DeepEP LL decode, 8 NICs); 32 602 vs 503 (+19.7%) 21.6 vs 89.0 s; 64 547 vs 563 98.2 vs 87.5 s; 128 482 vs 520 194 vs 183 s",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "vLLM PR #60161",
   "source_url": "https://github.com/vllm-project/vllm/pull/60161",
   "tg_tokens": null,
   "tps": 563,
   "ttft_s": 87.5
  },
  {
   "backend": "vLLM",
   "batch": 128,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "64 H100",
   "id": "vllm-60161-h100x64-c128-pr",
   "model": "GLM-5.3",
   "notes": "GLM-5.3 HiSparse decode on 64 H100 (4P/4D, DP4 x TP8, 288 GiB shared host pool, DeepEP LL, cc_traces 256k, 900 s/level); A/B is VLLM_NIXL_HISPARSE_DRAM_RANK0_ONLY (this PR) versus stock; out tok/s and TTFT p99 as published; concurrency 128; config=pr-c128",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "End-to-end benchmark, 4P/4D, 64 H100; Results (llmd-roce cluster, GLM-5.3, DeepEP LL decode, 8 NICs); 32 602 vs 503 (+19.7%) 21.6 vs 89.0 s; 64 547 vs 563 98.2 vs 87.5 s; 128 482 vs 520 194 vs 183 s",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "vLLM PR #60161",
   "source_url": "https://github.com/vllm-project/vllm/pull/60161",
   "tg_tokens": null,
   "tps": 482,
   "ttft_s": 194
  },
  {
   "backend": "vLLM",
   "batch": 128,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "64 H100",
   "id": "vllm-60161-h100x64-c128-stock",
   "model": "GLM-5.3",
   "notes": "GLM-5.3 HiSparse decode on 64 H100 (4P/4D, DP4 x TP8, 288 GiB shared host pool, DeepEP LL, cc_traces 256k, 900 s/level); A/B is VLLM_NIXL_HISPARSE_DRAM_RANK0_ONLY (this PR) versus stock; out tok/s and TTFT p99 as published; concurrency 128; config=stock-c128",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "End-to-end benchmark, 4P/4D, 64 H100; Results (llmd-roce cluster, GLM-5.3, DeepEP LL decode, 8 NICs); 32 602 vs 503 (+19.7%) 21.6 vs 89.0 s; 64 547 vs 563 98.2 vs 87.5 s; 128 482 vs 520 194 vs 183 s",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "vLLM PR #60161",
   "source_url": "https://github.com/vllm-project/vllm/pull/60161",
   "tg_tokens": null,
   "tps": 520,
   "ttft_s": 183
  },
  {
   "backend": "llama.cpp (Hexagon)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "IQ-9075",
   "id": "lc30035-llama321b-refq40-master",
   "model": "Llama-3.2-1B",
   "notes": "IQ-9075 Hexagon NPU (HTP0:0), llama-bench -p 512 -n 64 -fa 1 -t 6 --cpu-mask 0xfc -ub 256 -b 512; tg64 t/s; A/B is master (c479922) versus the q2_k tile-reduction PR (bpw 4.0 to 2.625, activation scale 1/32 to 1/256); master arm; config=ref-q40-master",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Setup: IQ-9075 (Hexagon v73), single NPU HTP0:0; Llama-3.2-1B --pure Q4_0 (reference) all Q4_0 1555.7 / 50.33 1557.5 / 50.28 +0.1% / -0.1%",
   "ram_gb": null,
   "retrieved": "2026-10-12",
   "source_name": "llama.cpp PR #30035",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30035",
   "tg_tokens": null,
   "tps": 50.33,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Hexagon)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "IQ-9075",
   "id": "lc30035-llama321b-refq40-pr",
   "model": "Llama-3.2-1B",
   "notes": "IQ-9075 Hexagon NPU (HTP0:0), llama-bench -p 512 -n 64 -fa 1 -t 6 --cpu-mask 0xfc -ub 256 -b 512; tg64 t/s; A/B is master (c479922) versus the q2_k tile-reduction PR (bpw 4.0 to 2.625, activation scale 1/32 to 1/256); pr arm; config=ref-q40-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Setup: IQ-9075 (Hexagon v73), single NPU HTP0:0; Llama-3.2-1B --pure Q4_0 (reference) all Q4_0 1555.7 / 50.33 1557.5 / 50.28 +0.1% / -0.1%",
   "ram_gb": null,
   "retrieved": "2026-10-12",
   "source_name": "llama.cpp PR #30035",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30035",
   "tg_tokens": null,
   "tps": 50.28,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Hexagon)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "IQ-9075",
   "id": "lc30035-llama321b-pureq2k-master",
   "model": "Llama-3.2-1B",
   "notes": "IQ-9075 Hexagon NPU (HTP0:0), llama-bench -p 512 -n 64 -fa 1 -t 6 --cpu-mask 0xfc -ub 256 -b 512; tg64 t/s; A/B is master (c479922) versus the q2_k tile-reduction PR (bpw 4.0 to 2.625, activation scale 1/32 to 1/256); master arm; config=pure-q2k-master",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Setup: IQ-9075 (Hexagon v73), single NPU HTP0:0; Llama-3.2-1B --pure Q2_K all Q2_K 1561.6 / 62.66 1536.8 / 66.69 -1.6% / +6.4%",
   "ram_gb": null,
   "retrieved": "2026-10-12",
   "source_name": "llama.cpp PR #30035",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30035",
   "tg_tokens": null,
   "tps": 62.66,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Hexagon)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "IQ-9075",
   "id": "lc30035-llama321b-pureq2k-pr",
   "model": "Llama-3.2-1B",
   "notes": "IQ-9075 Hexagon NPU (HTP0:0), llama-bench -p 512 -n 64 -fa 1 -t 6 --cpu-mask 0xfc -ub 256 -b 512; tg64 t/s; A/B is master (c479922) versus the q2_k tile-reduction PR (bpw 4.0 to 2.625, activation scale 1/32 to 1/256); pr arm; config=pure-q2k-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Setup: IQ-9075 (Hexagon v73), single NPU HTP0:0; Llama-3.2-1B --pure Q2_K all Q2_K 1561.6 / 62.66 1536.8 / 66.69 -1.6% / +6.4%",
   "ram_gb": null,
   "retrieved": "2026-10-12",
   "source_name": "llama.cpp PR #30035",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30035",
   "tg_tokens": null,
   "tps": 66.69,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Hexagon)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "IQ-9075",
   "id": "lc30035-llama321b-mixedq2k-master",
   "model": "Llama-3.2-1B",
   "notes": "IQ-9075 Hexagon NPU (HTP0:0), llama-bench -p 512 -n 64 -fa 1 -t 6 --cpu-mask 0xfc -ub 256 -b 512; tg64 t/s; A/B is master (c479922) versus the q2_k tile-reduction PR (bpw 4.0 to 2.625, activation scale 1/32 to 1/256); master arm; config=mixed-q2k-master",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Setup: IQ-9075 (Hexagon v73), single NPU HTP0:0; Llama-3.2-1B Q2_K Q2_K + Q3_K + Q4_K + Q6_K 1540.6 / 54.98 1526.8 / 56.46 -0.9% / +2.7%",
   "ram_gb": null,
   "retrieved": "2026-10-12",
   "source_name": "llama.cpp PR #30035",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30035",
   "tg_tokens": null,
   "tps": 54.98,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Hexagon)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "IQ-9075",
   "id": "lc30035-llama321b-mixedq2k-pr",
   "model": "Llama-3.2-1B",
   "notes": "IQ-9075 Hexagon NPU (HTP0:0), llama-bench -p 512 -n 64 -fa 1 -t 6 --cpu-mask 0xfc -ub 256 -b 512; tg64 t/s; A/B is master (c479922) versus the q2_k tile-reduction PR (bpw 4.0 to 2.625, activation scale 1/32 to 1/256); pr arm; config=mixed-q2k-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Setup: IQ-9075 (Hexagon v73), single NPU HTP0:0; Llama-3.2-1B Q2_K Q2_K + Q3_K + Q4_K + Q6_K 1540.6 / 54.98 1526.8 / 56.46 -0.9% / +2.7%",
   "ram_gb": null,
   "retrieved": "2026-10-12",
   "source_name": "llama.cpp PR #30035",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30035",
   "tg_tokens": null,
   "tps": 56.46,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Hexagon)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "IQ-9075",
   "id": "lc30035-llama323b-pureq2k-master",
   "model": "Llama-3.2-3B",
   "notes": "IQ-9075 Hexagon NPU (HTP0:0), llama-bench -p 512 -n 64 -fa 1 -t 6 --cpu-mask 0xfc -ub 256 -b 512; tg64 t/s; A/B is master (c479922) versus the q2_k tile-reduction PR (bpw 4.0 to 2.625, activation scale 1/32 to 1/256); master arm; config=pure-q2k-master",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Setup: IQ-9075 (Hexagon v73), single NPU HTP0:0; Llama-3.2-3B --pure Q2_K all Q2_K 634.9 / 26.18 626.2 / 27.33 -1.4% / +4.4%",
   "ram_gb": null,
   "retrieved": "2026-10-12",
   "source_name": "llama.cpp PR #30035",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30035",
   "tg_tokens": null,
   "tps": 26.18,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Hexagon)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "IQ-9075",
   "id": "lc30035-llama323b-pureq2k-pr",
   "model": "Llama-3.2-3B",
   "notes": "IQ-9075 Hexagon NPU (HTP0:0), llama-bench -p 512 -n 64 -fa 1 -t 6 --cpu-mask 0xfc -ub 256 -b 512; tg64 t/s; A/B is master (c479922) versus the q2_k tile-reduction PR (bpw 4.0 to 2.625, activation scale 1/32 to 1/256); pr arm; config=pure-q2k-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Setup: IQ-9075 (Hexagon v73), single NPU HTP0:0; Llama-3.2-3B --pure Q2_K all Q2_K 634.9 / 26.18 626.2 / 27.33 -1.4% / +4.4%",
   "ram_gb": null,
   "retrieved": "2026-10-12",
   "source_name": "llama.cpp PR #30035",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30035",
   "tg_tokens": null,
   "tps": 27.33,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Hexagon)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "IQ-9075",
   "id": "lc30035-llama323b-mixedq2k-master",
   "model": "Llama-3.2-3B",
   "notes": "IQ-9075 Hexagon NPU (HTP0:0), llama-bench -p 512 -n 64 -fa 1 -t 6 --cpu-mask 0xfc -ub 256 -b 512; tg64 t/s; A/B is master (c479922) versus the q2_k tile-reduction PR (bpw 4.0 to 2.625, activation scale 1/32 to 1/256); master arm; config=mixed-q2k-master",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Setup: IQ-9075 (Hexagon v73), single NPU HTP0:0; Llama-3.2-3B Q2_K Q2_K + Q3_K + Q6_K 627.4 / 24.47 621.1 / 24.92 -1.0% / +1.8%",
   "ram_gb": null,
   "retrieved": "2026-10-12",
   "source_name": "llama.cpp PR #30035",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30035",
   "tg_tokens": null,
   "tps": 24.47,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Hexagon)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "IQ-9075",
   "id": "lc30035-llama323b-mixedq2k-pr",
   "model": "Llama-3.2-3B",
   "notes": "IQ-9075 Hexagon NPU (HTP0:0), llama-bench -p 512 -n 64 -fa 1 -t 6 --cpu-mask 0xfc -ub 256 -b 512; tg64 t/s; A/B is master (c479922) versus the q2_k tile-reduction PR (bpw 4.0 to 2.625, activation scale 1/32 to 1/256); pr arm; config=mixed-q2k-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Setup: IQ-9075 (Hexagon v73), single NPU HTP0:0; Llama-3.2-3B Q2_K Q2_K + Q3_K + Q6_K 627.4 / 24.47 621.1 / 24.92 -1.0% / +1.8%",
   "ram_gb": null,
   "retrieved": "2026-10-12",
   "source_name": "llama.cpp PR #30035",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30035",
   "tg_tokens": null,
   "tps": 24.92,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Hexagon)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "IQ-9075",
   "id": "lc30035-qwen352b-pureq2k-master",
   "model": "Qwen3.5-2B",
   "notes": "IQ-9075 Hexagon NPU (HTP0:0), llama-bench -p 512 -n 64 -fa 1 -t 6 --cpu-mask 0xfc -ub 256 -b 512; tg64 t/s; A/B is master (c479922) versus the q2_k tile-reduction PR (bpw 4.0 to 2.625, activation scale 1/32 to 1/256); master arm; config=pure-q2k-master",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Setup: IQ-9075 (Hexagon v73), single NPU HTP0:0; Qwen3.5-2B --pure Q2_K (imatrix) all Q2_K 891.9 / 30.84 883.0 / 31.86 -1.0% / +3.3%",
   "ram_gb": null,
   "retrieved": "2026-10-12",
   "source_name": "llama.cpp PR #30035",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30035",
   "tg_tokens": null,
   "tps": 30.84,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Hexagon)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "IQ-9075",
   "id": "lc30035-qwen352b-pureq2k-pr",
   "model": "Qwen3.5-2B",
   "notes": "IQ-9075 Hexagon NPU (HTP0:0), llama-bench -p 512 -n 64 -fa 1 -t 6 --cpu-mask 0xfc -ub 256 -b 512; tg64 t/s; A/B is master (c479922) versus the q2_k tile-reduction PR (bpw 4.0 to 2.625, activation scale 1/32 to 1/256); pr arm; config=pure-q2k-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Setup: IQ-9075 (Hexagon v73), single NPU HTP0:0; Qwen3.5-2B --pure Q2_K (imatrix) all Q2_K 891.9 / 30.84 883.0 / 31.86 -1.0% / +3.3%",
   "ram_gb": null,
   "retrieved": "2026-10-12",
   "source_name": "llama.cpp PR #30035",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30035",
   "tg_tokens": null,
   "tps": 31.86,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Hexagon)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "IQ-9075",
   "id": "lc30035-qwen352b-mixedq2k-master",
   "model": "Qwen3.5-2B",
   "notes": "IQ-9075 Hexagon NPU (HTP0:0), llama-bench -p 512 -n 64 -fa 1 -t 6 --cpu-mask 0xfc -ub 256 -b 512; tg64 t/s; A/B is master (c479922) versus the q2_k tile-reduction PR (bpw 4.0 to 2.625, activation scale 1/32 to 1/256); master arm; config=mixed-q2k-master",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Setup: IQ-9075 (Hexagon v73), single NPU HTP0:0; Qwen3.5-2B Q2_K (bartowski) Q2_K + Q3_K + Q4_K + Q8_0 + Q6_K 886.1 / 25.86 881.8 / 26.17 -0.5% / +1.2%",
   "ram_gb": null,
   "retrieved": "2026-10-12",
   "source_name": "llama.cpp PR #30035",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30035",
   "tg_tokens": null,
   "tps": 25.86,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Hexagon)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "IQ-9075",
   "id": "lc30035-qwen352b-mixedq2k-pr",
   "model": "Qwen3.5-2B",
   "notes": "IQ-9075 Hexagon NPU (HTP0:0), llama-bench -p 512 -n 64 -fa 1 -t 6 --cpu-mask 0xfc -ub 256 -b 512; tg64 t/s; A/B is master (c479922) versus the q2_k tile-reduction PR (bpw 4.0 to 2.625, activation scale 1/32 to 1/256); pr arm; config=mixed-q2k-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Setup: IQ-9075 (Hexagon v73), single NPU HTP0:0; Qwen3.5-2B Q2_K (bartowski) Q2_K + Q3_K + Q4_K + Q8_0 + Q6_K 886.1 / 25.86 881.8 / 26.17 -0.5% / +1.2%",
   "ram_gb": null,
   "retrieved": "2026-10-12",
   "source_name": "llama.cpp PR #30035",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30035",
   "tg_tokens": null,
   "tps": 26.17,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 2048,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-qwen34b-q80-pp2048-t4-master",
   "model": "Qwen3 4B",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; pp2048 at 4 threads, master arm; config=pp2048-t4-master",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 95.95,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Qwen3-4B Q8_0 pp2048 4 95.95 310.88 3.24",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 2048,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-qwen34b-q80-pp2048-t4-pr",
   "model": "Qwen3 4B",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; pp2048 at 4 threads, pr arm; config=pp2048-t4-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 310.88,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Qwen3-4B Q8_0 pp2048 4 95.95 310.88 3.24",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 2048,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-qwen34b-q80-pp2048-t8-master",
   "model": "Qwen3 4B",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; pp2048 at 8 threads, master arm; config=pp2048-t8-master",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 176.81,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Qwen3-4B Q8_0 pp2048 8 176.81 552.12 3.12",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 2048,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-qwen34b-q80-pp2048-t8-pr",
   "model": "Qwen3 4B",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; pp2048 at 8 threads, pr arm; config=pp2048-t8-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 552.12,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Qwen3-4B Q8_0 pp2048 8 176.81 552.12 3.12",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 2048,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-qwen34b-q80-pp2048-t16-master",
   "model": "Qwen3 4B",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; pp2048 at 16 threads, master arm; config=pp2048-t16-master",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 307.15,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Qwen3-4B Q8_0 pp2048 16 307.15 810.96 2.64",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 2048,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-qwen34b-q80-pp2048-t16-pr",
   "model": "Qwen3 4B",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; pp2048 at 16 threads, pr arm; config=pp2048-t16-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 810.96,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Qwen3-4B Q8_0 pp2048 16 307.15 810.96 2.64",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-qwen34b-q80-pp4096-t4-master",
   "model": "Qwen3 4B",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; pp4096 at 4 threads, master arm; config=pp4096-t4-master",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 87.88,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Qwen3-4B Q8_0 pp4096 4 87.88 237.88 2.71",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-qwen34b-q80-pp4096-t4-pr",
   "model": "Qwen3 4B",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; pp4096 at 4 threads, pr arm; config=pp4096-t4-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 237.88,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Qwen3-4B Q8_0 pp4096 4 87.88 237.88 2.71",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-qwen34b-q80-pp4096-t8-master",
   "model": "Qwen3 4B",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; pp4096 at 8 threads, master arm; config=pp4096-t8-master",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 162.5,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Qwen3-4B Q8_0 pp4096 8 162.50 430.49 2.65",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-qwen34b-q80-pp4096-t8-pr",
   "model": "Qwen3 4B",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; pp4096 at 8 threads, pr arm; config=pp4096-t8-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 430.49,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Qwen3-4B Q8_0 pp4096 8 162.50 430.49 2.65",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-qwen34b-q80-pp4096-t16-master",
   "model": "Qwen3 4B",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; pp4096 at 16 threads, master arm; config=pp4096-t16-master",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 284.02,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Qwen3-4B Q8_0 pp4096 16 284.02 657.98 2.32",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-qwen34b-q80-pp4096-t16-pr",
   "model": "Qwen3 4B",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; pp4096 at 16 threads, pr arm; config=pp4096-t16-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 657.98,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Qwen3-4B Q8_0 pp4096 16 284.02 657.98 2.32",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-qwen34b-q80-tg128-t4-master",
   "model": "Qwen3 4B",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; tg128 at 4 threads, master arm; config=tg128-t4-master",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Qwen3-4B Q8_0 tg128 4 11.48 11.48 1.00",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": 11.48,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-qwen34b-q80-tg128-t4-pr",
   "model": "Qwen3 4B",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; tg128 at 4 threads, pr arm; config=tg128-t4-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Qwen3-4B Q8_0 tg128 4 11.48 11.48 1.00",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": 11.48,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-qwen34b-q80-tg128-t8-master",
   "model": "Qwen3 4B",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; tg128 at 8 threads, master arm; config=tg128-t8-master",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Qwen3-4B Q8_0 tg128 8 11.32 11.22 0.99",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": 11.32,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-qwen34b-q80-tg128-t8-pr",
   "model": "Qwen3 4B",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; tg128 at 8 threads, pr arm; config=tg128-t8-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Qwen3-4B Q8_0 tg128 8 11.32 11.22 0.99",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": 11.22,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-qwen34b-q80-tg128-t16-master",
   "model": "Qwen3 4B",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; tg128 at 16 threads, master arm; config=tg128-t16-master",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Qwen3-4B Q8_0 tg128 16 10.92 10.78 0.99",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": 10.92,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-qwen34b-q80-tg128-t16-pr",
   "model": "Qwen3 4B",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; tg128 at 16 threads, pr arm; config=tg128-t16-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Qwen3-4B Q8_0 tg128 16 10.92 10.78 0.99",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": 10.78,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 2048,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-qwen34binstruct2507-q40-pp2048-t4-master",
   "model": "Qwen3-4B-Instruct-2507",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; pp2048 at 4 threads, master arm; config=pp2048-t4-master",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 156.0,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Qwen3-4B-Instruct-2507 Q4_0 pp2048 4 156.00 262.06 1.68",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 2048,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-qwen34binstruct2507-q40-pp2048-t4-pr",
   "model": "Qwen3-4B-Instruct-2507",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; pp2048 at 4 threads, pr arm; config=pp2048-t4-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 262.06,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Qwen3-4B-Instruct-2507 Q4_0 pp2048 4 156.00 262.06 1.68",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 2048,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-qwen34binstruct2507-q40-pp2048-t8-master",
   "model": "Qwen3-4B-Instruct-2507",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; pp2048 at 8 threads, master arm; config=pp2048-t8-master",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 281.31,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Qwen3-4B-Instruct-2507 Q4_0 pp2048 8 281.31 474.37 1.69",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 2048,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-qwen34binstruct2507-q40-pp2048-t8-pr",
   "model": "Qwen3-4B-Instruct-2507",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; pp2048 at 8 threads, pr arm; config=pp2048-t8-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 474.37,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Qwen3-4B-Instruct-2507 Q4_0 pp2048 8 281.31 474.37 1.69",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 2048,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-qwen34binstruct2507-q40-pp2048-t16-master",
   "model": "Qwen3-4B-Instruct-2507",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; pp2048 at 16 threads, master arm; config=pp2048-t16-master",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 451.55,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Qwen3-4B-Instruct-2507 Q4_0 pp2048 16 451.55 741.54 1.64",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 2048,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-qwen34binstruct2507-q40-pp2048-t16-pr",
   "model": "Qwen3-4B-Instruct-2507",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; pp2048 at 16 threads, pr arm; config=pp2048-t16-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 741.54,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Qwen3-4B-Instruct-2507 Q4_0 pp2048 16 451.55 741.54 1.64",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-qwen34binstruct2507-q40-pp4096-t4-master",
   "model": "Qwen3-4B-Instruct-2507",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; pp4096 at 4 threads, master arm; config=pp4096-t4-master",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 135.97,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Qwen3-4B-Instruct-2507 Q4_0 pp4096 4 135.97 208.94 1.54",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-qwen34binstruct2507-q40-pp4096-t4-pr",
   "model": "Qwen3-4B-Instruct-2507",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; pp4096 at 4 threads, pr arm; config=pp4096-t4-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 208.94,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Qwen3-4B-Instruct-2507 Q4_0 pp4096 4 135.97 208.94 1.54",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-qwen34binstruct2507-q40-pp4096-t8-master",
   "model": "Qwen3-4B-Instruct-2507",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; pp4096 at 8 threads, master arm; config=pp4096-t8-master",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 246.41,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Qwen3-4B-Instruct-2507 Q4_0 pp4096 8 246.41 381.71 1.55",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-qwen34binstruct2507-q40-pp4096-t8-pr",
   "model": "Qwen3-4B-Instruct-2507",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; pp4096 at 8 threads, pr arm; config=pp4096-t8-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 381.71,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Qwen3-4B-Instruct-2507 Q4_0 pp4096 8 246.41 381.71 1.55",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-qwen34binstruct2507-q40-pp4096-t16-master",
   "model": "Qwen3-4B-Instruct-2507",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; pp4096 at 16 threads, master arm; config=pp4096-t16-master",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 401.77,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Qwen3-4B-Instruct-2507 Q4_0 pp4096 16 401.77 610.93 1.52",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-qwen34binstruct2507-q40-pp4096-t16-pr",
   "model": "Qwen3-4B-Instruct-2507",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; pp4096 at 16 threads, pr arm; config=pp4096-t16-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 610.93,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Qwen3-4B-Instruct-2507 Q4_0 pp4096 16 401.77 610.93 1.52",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-qwen34binstruct2507-q40-tg128-t4-master",
   "model": "Qwen3-4B-Instruct-2507",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; tg128 at 4 threads, master arm; config=tg128-t4-master",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Qwen3-4B-Instruct-2507 Q4_0 tg128 4 20.22 20.30 1.00",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": 20.22,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-qwen34binstruct2507-q40-tg128-t4-pr",
   "model": "Qwen3-4B-Instruct-2507",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; tg128 at 4 threads, pr arm; config=tg128-t4-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Qwen3-4B-Instruct-2507 Q4_0 tg128 4 20.22 20.30 1.00",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": 20.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-qwen34binstruct2507-q40-tg128-t8-master",
   "model": "Qwen3-4B-Instruct-2507",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; tg128 at 8 threads, master arm; config=tg128-t8-master",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Qwen3-4B-Instruct-2507 Q4_0 tg128 8 19.97 20.05 1.00",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": 19.97,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-qwen34binstruct2507-q40-tg128-t8-pr",
   "model": "Qwen3-4B-Instruct-2507",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; tg128 at 8 threads, pr arm; config=tg128-t8-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Qwen3-4B-Instruct-2507 Q4_0 tg128 8 19.97 20.05 1.00",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": 20.05,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-qwen34binstruct2507-q40-tg128-t16-master",
   "model": "Qwen3-4B-Instruct-2507",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; tg128 at 16 threads, master arm; config=tg128-t16-master",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Qwen3-4B-Instruct-2507 Q4_0 tg128 16 19.21 19.14 1.00",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": 19.21,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-qwen34binstruct2507-q40-tg128-t16-pr",
   "model": "Qwen3-4B-Instruct-2507",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; tg128 at 16 threads, pr arm; config=tg128-t16-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Qwen3-4B-Instruct-2507 Q4_0 tg128 16 19.21 19.14 1.00",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": 19.14,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 2048,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-llama318b-q80-pp2048-t4-master",
   "model": "llama-3-1-8b-instruct",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; pp2048 at 4 threads, master arm; config=pp2048-t4-master",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 53.84,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Llama-3.1-8B-Instruct Q8_0 pp2048 4 53.84 209.10 3.88",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 2048,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-llama318b-q80-pp2048-t4-pr",
   "model": "llama-3-1-8b-instruct",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; pp2048 at 4 threads, pr arm; config=pp2048-t4-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 209.1,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Llama-3.1-8B-Instruct Q8_0 pp2048 4 53.84 209.10 3.88",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 2048,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-llama318b-q80-pp2048-t8-master",
   "model": "llama-3-1-8b-instruct",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; pp2048 at 8 threads, master arm; config=pp2048-t8-master",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 99.37,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Llama-3.1-8B-Instruct Q8_0 pp2048 8 99.37 366.70 3.69",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 2048,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-llama318b-q80-pp2048-t8-pr",
   "model": "llama-3-1-8b-instruct",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; pp2048 at 8 threads, pr arm; config=pp2048-t8-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 366.7,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Llama-3.1-8B-Instruct Q8_0 pp2048 8 99.37 366.70 3.69",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 2048,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-llama318b-q80-pp2048-t16-master",
   "model": "llama-3-1-8b-instruct",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; pp2048 at 16 threads, master arm; config=pp2048-t16-master",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 176.76,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Llama-3.1-8B-Instruct Q8_0 pp2048 16 176.76 571.95 3.24",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 2048,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-llama318b-q80-pp2048-t16-pr",
   "model": "llama-3-1-8b-instruct",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; pp2048 at 16 threads, pr arm; config=pp2048-t16-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 571.95,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Llama-3.1-8B-Instruct Q8_0 pp2048 16 176.76 571.95 3.24",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-llama318b-q80-pp4096-t4-master",
   "model": "llama-3-1-8b-instruct",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; pp4096 at 4 threads, master arm; config=pp4096-t4-master",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 51.5,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Llama-3.1-8B-Instruct Q8_0 pp4096 4 51.50 177.31 3.44",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-llama318b-q80-pp4096-t4-pr",
   "model": "llama-3-1-8b-instruct",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; pp4096 at 4 threads, pr arm; config=pp4096-t4-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 177.31,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Llama-3.1-8B-Instruct Q8_0 pp4096 4 51.50 177.31 3.44",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-llama318b-q80-pp4096-t8-master",
   "model": "llama-3-1-8b-instruct",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; pp4096 at 8 threads, master arm; config=pp4096-t8-master",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 95.36,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Llama-3.1-8B-Instruct Q8_0 pp4096 8 95.36 315.43 3.31",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-llama318b-q80-pp4096-t8-pr",
   "model": "llama-3-1-8b-instruct",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; pp4096 at 8 threads, pr arm; config=pp4096-t8-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 315.43,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Llama-3.1-8B-Instruct Q8_0 pp4096 8 95.36 315.43 3.31",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-llama318b-q80-pp4096-t16-master",
   "model": "llama-3-1-8b-instruct",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; pp4096 at 16 threads, master arm; config=pp4096-t16-master",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 168.85,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Llama-3.1-8B-Instruct Q8_0 pp4096 16 168.85 496.83 2.94",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-llama318b-q80-pp4096-t16-pr",
   "model": "llama-3-1-8b-instruct",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; pp4096 at 16 threads, pr arm; config=pp4096-t16-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 496.83,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Llama-3.1-8B-Instruct Q8_0 pp4096 16 168.85 496.83 2.94",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-llama318b-q80-tg128-t4-master",
   "model": "llama-3-1-8b-instruct",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; tg128 at 4 threads, master arm; config=tg128-t4-master",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Llama-3.1-8B-Instruct Q8_0 tg128 4 6.25 6.21 0.99",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": 6.25,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-llama318b-q80-tg128-t4-pr",
   "model": "llama-3-1-8b-instruct",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; tg128 at 4 threads, pr arm; config=tg128-t4-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Llama-3.1-8B-Instruct Q8_0 tg128 4 6.25 6.21 0.99",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": 6.21,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-llama318b-q80-tg128-t8-master",
   "model": "llama-3-1-8b-instruct",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; tg128 at 8 threads, master arm; config=tg128-t8-master",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Llama-3.1-8B-Instruct Q8_0 tg128 8 6.09 6.07 1.00",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": 6.09,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-llama318b-q80-tg128-t8-pr",
   "model": "llama-3-1-8b-instruct",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; tg128 at 8 threads, pr arm; config=tg128-t8-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Llama-3.1-8B-Instruct Q8_0 tg128 8 6.09 6.07 1.00",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": 6.07,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-llama318b-q80-tg128-t16-master",
   "model": "llama-3-1-8b-instruct",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; tg128 at 16 threads, master arm; config=tg128-t16-master",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Llama-3.1-8B-Instruct Q8_0 tg128 16 5.91 5.86 0.99",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": 5.91,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-llama318b-q80-tg128-t16-pr",
   "model": "llama-3-1-8b-instruct",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; tg128 at 16 threads, pr arm; config=tg128-t16-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Llama-3.1-8B-Instruct Q8_0 tg128 16 5.91 5.86 0.99",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": 5.86,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 2048,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-llama318b-q40-pp2048-t4-master",
   "model": "llama-3-1-8b-instruct",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; pp2048 at 4 threads, master arm; config=pp2048-t4-master",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 99.07,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Llama-3.1-8B-Instruct Q4_0 pp2048 4 99.07 205.03 2.07",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 2048,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-llama318b-q40-pp2048-t4-pr",
   "model": "llama-3-1-8b-instruct",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; pp2048 at 4 threads, pr arm; config=pp2048-t4-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.03,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Llama-3.1-8B-Instruct Q4_0 pp2048 4 99.07 205.03 2.07",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 2048,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-llama318b-q40-pp2048-t8-master",
   "model": "llama-3-1-8b-instruct",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; pp2048 at 8 threads, master arm; config=pp2048-t8-master",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 177.18,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Llama-3.1-8B-Instruct Q4_0 pp2048 8 177.18 365.14 2.06",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 2048,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-llama318b-q40-pp2048-t8-pr",
   "model": "llama-3-1-8b-instruct",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; pp2048 at 8 threads, pr arm; config=pp2048-t8-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 365.14,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Llama-3.1-8B-Instruct Q4_0 pp2048 8 177.18 365.14 2.06",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 2048,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-llama318b-q40-pp2048-t16-master",
   "model": "llama-3-1-8b-instruct",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; pp2048 at 16 threads, master arm; config=pp2048-t16-master",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 305.82,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Llama-3.1-8B-Instruct Q4_0 pp2048 16 305.82 580.20 1.90",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 2048,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-llama318b-q40-pp2048-t16-pr",
   "model": "llama-3-1-8b-instruct",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; pp2048 at 16 threads, pr arm; config=pp2048-t16-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 580.2,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Llama-3.1-8B-Instruct Q4_0 pp2048 16 305.82 580.20 1.90",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-llama318b-q40-pp4096-t4-master",
   "model": "llama-3-1-8b-instruct",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; pp4096 at 4 threads, master arm; config=pp4096-t4-master",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 91.36,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Llama-3.1-8B-Instruct Q4_0 pp4096 4 91.36 174.46 1.91",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-llama318b-q40-pp4096-t4-pr",
   "model": "llama-3-1-8b-instruct",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; pp4096 at 4 threads, pr arm; config=pp4096-t4-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 174.46,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Llama-3.1-8B-Instruct Q4_0 pp4096 4 91.36 174.46 1.91",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-llama318b-q40-pp4096-t8-master",
   "model": "llama-3-1-8b-instruct",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; pp4096 at 8 threads, master arm; config=pp4096-t8-master",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 164.27,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Llama-3.1-8B-Instruct Q4_0 pp4096 8 164.27 312.12 1.90",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-llama318b-q40-pp4096-t8-pr",
   "model": "llama-3-1-8b-instruct",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; pp4096 at 8 threads, pr arm; config=pp4096-t8-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 312.12,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Llama-3.1-8B-Instruct Q4_0 pp4096 8 164.27 312.12 1.90",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-llama318b-q40-pp4096-t16-master",
   "model": "llama-3-1-8b-instruct",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; pp4096 at 16 threads, master arm; config=pp4096-t16-master",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 283.17,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Llama-3.1-8B-Instruct Q4_0 pp4096 16 283.17 504.99 1.78",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-llama318b-q40-pp4096-t16-pr",
   "model": "llama-3-1-8b-instruct",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; pp4096 at 16 threads, pr arm; config=pp4096-t16-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 504.99,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Llama-3.1-8B-Instruct Q4_0 pp4096 16 283.17 504.99 1.78",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-llama318b-q40-tg128-t4-master",
   "model": "llama-3-1-8b-instruct",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; tg128 at 4 threads, master arm; config=tg128-t4-master",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Llama-3.1-8B-Instruct Q4_0 tg128 4 11.27 11.29 1.00",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": 11.27,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-llama318b-q40-tg128-t4-pr",
   "model": "llama-3-1-8b-instruct",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; tg128 at 4 threads, pr arm; config=tg128-t4-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Llama-3.1-8B-Instruct Q4_0 tg128 4 11.27 11.29 1.00",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": 11.29,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-llama318b-q40-tg128-t8-master",
   "model": "llama-3-1-8b-instruct",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; tg128 at 8 threads, master arm; config=tg128-t8-master",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Llama-3.1-8B-Instruct Q4_0 tg128 8 11.06 11.03 1.00",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": 11.06,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-llama318b-q40-tg128-t8-pr",
   "model": "llama-3-1-8b-instruct",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; tg128 at 8 threads, pr arm; config=tg128-t8-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Llama-3.1-8B-Instruct Q4_0 tg128 8 11.06 11.03 1.00",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": 11.03,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-llama318b-q40-tg128-t16-master",
   "model": "llama-3-1-8b-instruct",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; tg128 at 16 threads, master arm; config=tg128-t16-master",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Llama-3.1-8B-Instruct Q4_0 tg128 16 10.69 10.64 1.00",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": 10.69,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD Ryzen 9 9950X",
   "id": "lc30036-llama318b-q40-tg128-t16-pr",
   "model": "llama-3-1-8b-instruct",
   "notes": "AMD Ryzen 9 9950X (AVX512-VNNI), llama.cpp CPU; A/B is master versus the Q8_0/Q4_0 VNNI repack-kernel PR (vpdpbusd 8x4/16x4 layouts); llama-bench pp/tg as named; tg128 at 16 threads, pr arm; config=tg128-t16-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Ryzen 9 9950X (AVX512-VNNI); Llama-3.1-8B-Instruct Q4_0 tg128 16 10.69 10.64 1.00",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30036",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30036",
   "tg_tokens": null,
   "tps": 10.64,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": 10890,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD RX 5700 XT 8GB (gfx1010, RDNA1)",
   "id": "lc30039-rx5700xt-b10455",
   "model": "Qwen3.5-9B",
   "notes": "AMD RX 5700 XT 8GB (gfx1010, RDNA1), llama.cpp Vulkan; A/B is the b10455 build (2026-08-17) versus b11429 (v0.6.0) on a Qwen3.5-9B hybrid GDN/full-attention fine-tune Q4_K_M, 10890-token prompt, -fa on -ctk q4_0 -ctv q4_0 -ngl 99, single cold-start request; b10455 arm; config=b10455",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 269.7,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "AMD RX 5700 XT 8GB (gfx1010, RDNA1); Qwen3.5-9B; b10455-37a215c9e (2026-08-17) 269.7 44.5",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp issue #30039",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30039",
   "tg_tokens": null,
   "tps": 44.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": 10890,
   "date": "2026-10",
   "flags": [],
   "hardware": "AMD RX 5700 XT 8GB (gfx1010, RDNA1)",
   "id": "lc30039-rx5700xt-b11429",
   "model": "Qwen3.5-9B",
   "notes": "AMD RX 5700 XT 8GB (gfx1010, RDNA1), llama.cpp Vulkan; A/B is the b10455 build (2026-08-17) versus b11429 (v0.6.0) on a Qwen3.5-9B hybrid GDN/full-attention fine-tune Q4_K_M, 10890-token prompt, -fa on -ctk q4_0 -ctv q4_0 -ngl 99, single cold-start request; b11429 arm; config=b11429",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 151.6,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "AMD RX 5700 XT 8GB (gfx1010, RDNA1); Qwen3.5-9B; b11429-d81235049 (v0.6.0, 2026-10-05) 151.6 44.3",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp issue #30039",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30039",
   "tg_tokens": null,
   "tps": 44.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hf1-marcomini-reap05",
   "model": "Marco-Mini-Instruct",
   "notes": "kueizen REAP expert-pruned GGUF, llama-bench batch 1, all layers on GPU; config=reap05; REAP05 (244 experts, 9.33 GB, PPL 8.59).",
   "pp_tokens": 512,
   "pp_tps": 5160,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Nine files from 9.3 GB down to 3.1 GB, all Q4_K_M, all measured on one RTX 3090; File Experts kept Size (GB) PPL pp512 (t/s) tg128 (t/s) unpruned base (not in this repo) 256 10.00 8.17 6,221 \u2014 Marco-Mini-Instruct-REAP05-Q4_K_M.gguf 244 9.33 8.59 5,160 155\u00b9 Marco-Mini-Instruct-REAP10-Q4_K_M.gguf 231 8.85 9.30 6,482 281 Marco-Mini-Instruct-REAP15-Q4_K_M.gguf 218 8.36 10.28 6,806 256 Marco-Mini-Instruct-REAP20-Q4_K_M.gguf 205 7.88 11.40 6,005 277 Marco-Mini-Instruct-REAP30-Q4_K_M.gguf 179 6.94 14.64 7,661 278 Marco-Mini-Instruct-REAP40-Q4_K_M.gguf 154 5.97 18.74 6,970 278 Marco-Mini-Instruct-REAP50-Q4_K_M.gguf 128 5.00 26.42 6,966 316 Marco-Mini-Instruct-REAP60-Q4_K_M.gguf 102 4.07 38.62 9,335 287 Marco-Mini-Instruct-REAP70-Q4_K_M.gguf 77 3.10 64.43 9,713 278",
   "retrieved": "2026-10-06",
   "source_name": "HF model card: kueizen/Marco-Mini-Instruct-REAP-GGUF",
   "source_url": "https://huggingface.co/kueizen/Marco-Mini-Instruct-REAP-GGUF",
   "tg_tokens": 128,
   "tps": 155
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hf1-marcomini-reap10",
   "model": "Marco-Mini-Instruct",
   "notes": "kueizen REAP expert-pruned GGUF, llama-bench batch 1, all layers on GPU; config=reap10; REAP10 (231 experts, 8.85 GB, PPL 9.30).",
   "pp_tokens": 512,
   "pp_tps": 6482,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Nine files from 9.3 GB down to 3.1 GB, all Q4_K_M, all measured on one RTX 3090; File Experts kept Size (GB) PPL pp512 (t/s) tg128 (t/s) unpruned base (not in this repo) 256 10.00 8.17 6,221 \u2014 Marco-Mini-Instruct-REAP05-Q4_K_M.gguf 244 9.33 8.59 5,160 155\u00b9 Marco-Mini-Instruct-REAP10-Q4_K_M.gguf 231 8.85 9.30 6,482 281 Marco-Mini-Instruct-REAP15-Q4_K_M.gguf 218 8.36 10.28 6,806 256 Marco-Mini-Instruct-REAP20-Q4_K_M.gguf 205 7.88 11.40 6,005 277 Marco-Mini-Instruct-REAP30-Q4_K_M.gguf 179 6.94 14.64 7,661 278 Marco-Mini-Instruct-REAP40-Q4_K_M.gguf 154 5.97 18.74 6,970 278 Marco-Mini-Instruct-REAP50-Q4_K_M.gguf 128 5.00 26.42 6,966 316 Marco-Mini-Instruct-REAP60-Q4_K_M.gguf 102 4.07 38.62 9,335 287 Marco-Mini-Instruct-REAP70-Q4_K_M.gguf 77 3.10 64.43 9,713 278",
   "retrieved": "2026-10-06",
   "source_name": "HF model card: kueizen/Marco-Mini-Instruct-REAP-GGUF",
   "source_url": "https://huggingface.co/kueizen/Marco-Mini-Instruct-REAP-GGUF",
   "tg_tokens": 128,
   "tps": 281
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hf1-marcomini-reap15",
   "model": "Marco-Mini-Instruct",
   "notes": "kueizen REAP expert-pruned GGUF, llama-bench batch 1, all layers on GPU; config=reap15; REAP15 (218 experts, 8.36 GB, PPL 10.28).",
   "pp_tokens": 512,
   "pp_tps": 6806,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Nine files from 9.3 GB down to 3.1 GB, all Q4_K_M, all measured on one RTX 3090; File Experts kept Size (GB) PPL pp512 (t/s) tg128 (t/s) unpruned base (not in this repo) 256 10.00 8.17 6,221 \u2014 Marco-Mini-Instruct-REAP05-Q4_K_M.gguf 244 9.33 8.59 5,160 155\u00b9 Marco-Mini-Instruct-REAP10-Q4_K_M.gguf 231 8.85 9.30 6,482 281 Marco-Mini-Instruct-REAP15-Q4_K_M.gguf 218 8.36 10.28 6,806 256 Marco-Mini-Instruct-REAP20-Q4_K_M.gguf 205 7.88 11.40 6,005 277 Marco-Mini-Instruct-REAP30-Q4_K_M.gguf 179 6.94 14.64 7,661 278 Marco-Mini-Instruct-REAP40-Q4_K_M.gguf 154 5.97 18.74 6,970 278 Marco-Mini-Instruct-REAP50-Q4_K_M.gguf 128 5.00 26.42 6,966 316 Marco-Mini-Instruct-REAP60-Q4_K_M.gguf 102 4.07 38.62 9,335 287 Marco-Mini-Instruct-REAP70-Q4_K_M.gguf 77 3.10 64.43 9,713 278",
   "retrieved": "2026-10-06",
   "source_name": "HF model card: kueizen/Marco-Mini-Instruct-REAP-GGUF",
   "source_url": "https://huggingface.co/kueizen/Marco-Mini-Instruct-REAP-GGUF",
   "tg_tokens": 128,
   "tps": 256
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hf1-marcomini-reap20",
   "model": "Marco-Mini-Instruct",
   "notes": "kueizen REAP expert-pruned GGUF, llama-bench batch 1, all layers on GPU; config=reap20; REAP20 (205 experts, 7.88 GB, PPL 11.40).",
   "pp_tokens": 512,
   "pp_tps": 6005,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Nine files from 9.3 GB down to 3.1 GB, all Q4_K_M, all measured on one RTX 3090; File Experts kept Size (GB) PPL pp512 (t/s) tg128 (t/s) unpruned base (not in this repo) 256 10.00 8.17 6,221 \u2014 Marco-Mini-Instruct-REAP05-Q4_K_M.gguf 244 9.33 8.59 5,160 155\u00b9 Marco-Mini-Instruct-REAP10-Q4_K_M.gguf 231 8.85 9.30 6,482 281 Marco-Mini-Instruct-REAP15-Q4_K_M.gguf 218 8.36 10.28 6,806 256 Marco-Mini-Instruct-REAP20-Q4_K_M.gguf 205 7.88 11.40 6,005 277 Marco-Mini-Instruct-REAP30-Q4_K_M.gguf 179 6.94 14.64 7,661 278 Marco-Mini-Instruct-REAP40-Q4_K_M.gguf 154 5.97 18.74 6,970 278 Marco-Mini-Instruct-REAP50-Q4_K_M.gguf 128 5.00 26.42 6,966 316 Marco-Mini-Instruct-REAP60-Q4_K_M.gguf 102 4.07 38.62 9,335 287 Marco-Mini-Instruct-REAP70-Q4_K_M.gguf 77 3.10 64.43 9,713 278",
   "retrieved": "2026-10-06",
   "source_name": "HF model card: kueizen/Marco-Mini-Instruct-REAP-GGUF",
   "source_url": "https://huggingface.co/kueizen/Marco-Mini-Instruct-REAP-GGUF",
   "tg_tokens": 128,
   "tps": 277
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hf1-marcomini-reap30",
   "model": "Marco-Mini-Instruct",
   "notes": "kueizen REAP expert-pruned GGUF, llama-bench batch 1, all layers on GPU; config=reap30; REAP30 (179 experts, 6.94 GB, PPL 14.64).",
   "pp_tokens": 512,
   "pp_tps": 7661,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Nine files from 9.3 GB down to 3.1 GB, all Q4_K_M, all measured on one RTX 3090; File Experts kept Size (GB) PPL pp512 (t/s) tg128 (t/s) unpruned base (not in this repo) 256 10.00 8.17 6,221 \u2014 Marco-Mini-Instruct-REAP05-Q4_K_M.gguf 244 9.33 8.59 5,160 155\u00b9 Marco-Mini-Instruct-REAP10-Q4_K_M.gguf 231 8.85 9.30 6,482 281 Marco-Mini-Instruct-REAP15-Q4_K_M.gguf 218 8.36 10.28 6,806 256 Marco-Mini-Instruct-REAP20-Q4_K_M.gguf 205 7.88 11.40 6,005 277 Marco-Mini-Instruct-REAP30-Q4_K_M.gguf 179 6.94 14.64 7,661 278 Marco-Mini-Instruct-REAP40-Q4_K_M.gguf 154 5.97 18.74 6,970 278 Marco-Mini-Instruct-REAP50-Q4_K_M.gguf 128 5.00 26.42 6,966 316 Marco-Mini-Instruct-REAP60-Q4_K_M.gguf 102 4.07 38.62 9,335 287 Marco-Mini-Instruct-REAP70-Q4_K_M.gguf 77 3.10 64.43 9,713 278",
   "retrieved": "2026-10-06",
   "source_name": "HF model card: kueizen/Marco-Mini-Instruct-REAP-GGUF",
   "source_url": "https://huggingface.co/kueizen/Marco-Mini-Instruct-REAP-GGUF",
   "tg_tokens": 128,
   "tps": 278
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hf1-marcomini-reap40",
   "model": "Marco-Mini-Instruct",
   "notes": "kueizen REAP expert-pruned GGUF, llama-bench batch 1, all layers on GPU; config=reap40; REAP40 (154 experts, 5.97 GB, PPL 18.74).",
   "pp_tokens": 512,
   "pp_tps": 6970,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Nine files from 9.3 GB down to 3.1 GB, all Q4_K_M, all measured on one RTX 3090; File Experts kept Size (GB) PPL pp512 (t/s) tg128 (t/s) unpruned base (not in this repo) 256 10.00 8.17 6,221 \u2014 Marco-Mini-Instruct-REAP05-Q4_K_M.gguf 244 9.33 8.59 5,160 155\u00b9 Marco-Mini-Instruct-REAP10-Q4_K_M.gguf 231 8.85 9.30 6,482 281 Marco-Mini-Instruct-REAP15-Q4_K_M.gguf 218 8.36 10.28 6,806 256 Marco-Mini-Instruct-REAP20-Q4_K_M.gguf 205 7.88 11.40 6,005 277 Marco-Mini-Instruct-REAP30-Q4_K_M.gguf 179 6.94 14.64 7,661 278 Marco-Mini-Instruct-REAP40-Q4_K_M.gguf 154 5.97 18.74 6,970 278 Marco-Mini-Instruct-REAP50-Q4_K_M.gguf 128 5.00 26.42 6,966 316 Marco-Mini-Instruct-REAP60-Q4_K_M.gguf 102 4.07 38.62 9,335 287 Marco-Mini-Instruct-REAP70-Q4_K_M.gguf 77 3.10 64.43 9,713 278",
   "retrieved": "2026-10-06",
   "source_name": "HF model card: kueizen/Marco-Mini-Instruct-REAP-GGUF",
   "source_url": "https://huggingface.co/kueizen/Marco-Mini-Instruct-REAP-GGUF",
   "tg_tokens": 128,
   "tps": 278
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hf1-marcomini-reap50",
   "model": "Marco-Mini-Instruct",
   "notes": "kueizen REAP expert-pruned GGUF, llama-bench batch 1, all layers on GPU; config=reap50; REAP50 (128 experts, 5.00 GB, PPL 26.42).",
   "pp_tokens": 512,
   "pp_tps": 6966,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Nine files from 9.3 GB down to 3.1 GB, all Q4_K_M, all measured on one RTX 3090; File Experts kept Size (GB) PPL pp512 (t/s) tg128 (t/s) unpruned base (not in this repo) 256 10.00 8.17 6,221 \u2014 Marco-Mini-Instruct-REAP05-Q4_K_M.gguf 244 9.33 8.59 5,160 155\u00b9 Marco-Mini-Instruct-REAP10-Q4_K_M.gguf 231 8.85 9.30 6,482 281 Marco-Mini-Instruct-REAP15-Q4_K_M.gguf 218 8.36 10.28 6,806 256 Marco-Mini-Instruct-REAP20-Q4_K_M.gguf 205 7.88 11.40 6,005 277 Marco-Mini-Instruct-REAP30-Q4_K_M.gguf 179 6.94 14.64 7,661 278 Marco-Mini-Instruct-REAP40-Q4_K_M.gguf 154 5.97 18.74 6,970 278 Marco-Mini-Instruct-REAP50-Q4_K_M.gguf 128 5.00 26.42 6,966 316 Marco-Mini-Instruct-REAP60-Q4_K_M.gguf 102 4.07 38.62 9,335 287 Marco-Mini-Instruct-REAP70-Q4_K_M.gguf 77 3.10 64.43 9,713 278",
   "retrieved": "2026-10-06",
   "source_name": "HF model card: kueizen/Marco-Mini-Instruct-REAP-GGUF",
   "source_url": "https://huggingface.co/kueizen/Marco-Mini-Instruct-REAP-GGUF",
   "tg_tokens": 128,
   "tps": 316
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hf1-marcomini-reap60",
   "model": "Marco-Mini-Instruct",
   "notes": "kueizen REAP expert-pruned GGUF, llama-bench batch 1, all layers on GPU; config=reap60; REAP60 (102 experts, 4.07 GB, PPL 38.62).",
   "pp_tokens": 512,
   "pp_tps": 9335,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Nine files from 9.3 GB down to 3.1 GB, all Q4_K_M, all measured on one RTX 3090; File Experts kept Size (GB) PPL pp512 (t/s) tg128 (t/s) unpruned base (not in this repo) 256 10.00 8.17 6,221 \u2014 Marco-Mini-Instruct-REAP05-Q4_K_M.gguf 244 9.33 8.59 5,160 155\u00b9 Marco-Mini-Instruct-REAP10-Q4_K_M.gguf 231 8.85 9.30 6,482 281 Marco-Mini-Instruct-REAP15-Q4_K_M.gguf 218 8.36 10.28 6,806 256 Marco-Mini-Instruct-REAP20-Q4_K_M.gguf 205 7.88 11.40 6,005 277 Marco-Mini-Instruct-REAP30-Q4_K_M.gguf 179 6.94 14.64 7,661 278 Marco-Mini-Instruct-REAP40-Q4_K_M.gguf 154 5.97 18.74 6,970 278 Marco-Mini-Instruct-REAP50-Q4_K_M.gguf 128 5.00 26.42 6,966 316 Marco-Mini-Instruct-REAP60-Q4_K_M.gguf 102 4.07 38.62 9,335 287 Marco-Mini-Instruct-REAP70-Q4_K_M.gguf 77 3.10 64.43 9,713 278",
   "retrieved": "2026-10-06",
   "source_name": "HF model card: kueizen/Marco-Mini-Instruct-REAP-GGUF",
   "source_url": "https://huggingface.co/kueizen/Marco-Mini-Instruct-REAP-GGUF",
   "tg_tokens": 128,
   "tps": 287
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hf1-marcomini-reap70",
   "model": "Marco-Mini-Instruct",
   "notes": "kueizen REAP expert-pruned GGUF, llama-bench batch 1, all layers on GPU; config=reap70; REAP70 (77 experts, 3.10 GB, PPL 64.43).",
   "pp_tokens": 512,
   "pp_tps": 9713,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Nine files from 9.3 GB down to 3.1 GB, all Q4_K_M, all measured on one RTX 3090; File Experts kept Size (GB) PPL pp512 (t/s) tg128 (t/s) unpruned base (not in this repo) 256 10.00 8.17 6,221 \u2014 Marco-Mini-Instruct-REAP05-Q4_K_M.gguf 244 9.33 8.59 5,160 155\u00b9 Marco-Mini-Instruct-REAP10-Q4_K_M.gguf 231 8.85 9.30 6,482 281 Marco-Mini-Instruct-REAP15-Q4_K_M.gguf 218 8.36 10.28 6,806 256 Marco-Mini-Instruct-REAP20-Q4_K_M.gguf 205 7.88 11.40 6,005 277 Marco-Mini-Instruct-REAP30-Q4_K_M.gguf 179 6.94 14.64 7,661 278 Marco-Mini-Instruct-REAP40-Q4_K_M.gguf 154 5.97 18.74 6,970 278 Marco-Mini-Instruct-REAP50-Q4_K_M.gguf 128 5.00 26.42 6,966 316 Marco-Mini-Instruct-REAP60-Q4_K_M.gguf 102 4.07 38.62 9,335 287 Marco-Mini-Instruct-REAP70-Q4_K_M.gguf 77 3.10 64.43 9,713 278",
   "retrieved": "2026-10-06",
   "source_name": "HF model card: kueizen/Marco-Mini-Instruct-REAP-GGUF",
   "source_url": "https://huggingface.co/kueizen/Marco-Mini-Instruct-REAP-GGUF",
   "tg_tokens": 128,
   "tps": 278
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hf1-marcnano-reap05",
   "model": "Marco-Nano-Instruct",
   "notes": "kueizen REAP expert-pruned GGUF, llama-bench batch 1, all layers on GPU; config=reap05; REAP05 (220 experts, 4.77 GB, PPL 12.07).",
   "pp_tokens": 512,
   "pp_tps": 9142,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Nine files from 4.8 GB down to 1.7 GB, all Q4_K_M, all measured on one RTX 3090; File Experts kept Size (GB) PPL KLD (nats) pp512 (t/s) tg128 (t/s) unpruned base (not in this repo) 232 5.00 11.48 \u2014 8,460 277 Marco-Nano-Instruct-REAP05-Q4_K_M.gguf 220 4.77 12.07 0.065 9,142 282 Marco-Nano-Instruct-REAP10-Q4_K_M.gguf 209 4.52 13.09 0.151 8,578 282 Marco-Nano-Instruct-REAP15-Q4_K_M.gguf 197 4.30 14.03 0.214 9,378 241 Marco-Nano-Instruct-REAP20-Q4_K_M.gguf 186 4.05 15.36 0.291 9,402 271 Marco-Nano-Instruct-REAP30-Q4_K_M.gguf 162 3.58 18.03 0.458 9,681 268 Marco-Nano-Instruct-REAP40-Q4_K_M.gguf 139 3.11 22.71 0.671 10,185 279 Marco-Nano-Instruct-REAP50-Q4_K_M.gguf 116 2.61 30.38 0.941 10,157 296 Marco-Nano-Instruct-REAP60-Q4_K_M.gguf 93 2.14 45.04 1.289 11,146 296 Marco-Nano-Instruct-REAP70-Q4_K_M.gguf 70 1.66 74.79 1.718 11,416 302",
   "retrieved": "2026-10-06",
   "source_name": "HF model card: kueizen/Marco-Nano-Instruct-REAP-GGUF",
   "source_url": "https://huggingface.co/kueizen/Marco-Nano-Instruct-REAP-GGUF",
   "tg_tokens": 128,
   "tps": 282
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hf1-marcnano-reap10",
   "model": "Marco-Nano-Instruct",
   "notes": "kueizen REAP expert-pruned GGUF, llama-bench batch 1, all layers on GPU; config=reap10; REAP10 (209 experts, 4.52 GB, PPL 13.09).",
   "pp_tokens": 512,
   "pp_tps": 8578,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Nine files from 4.8 GB down to 1.7 GB, all Q4_K_M, all measured on one RTX 3090; File Experts kept Size (GB) PPL KLD (nats) pp512 (t/s) tg128 (t/s) unpruned base (not in this repo) 232 5.00 11.48 \u2014 8,460 277 Marco-Nano-Instruct-REAP05-Q4_K_M.gguf 220 4.77 12.07 0.065 9,142 282 Marco-Nano-Instruct-REAP10-Q4_K_M.gguf 209 4.52 13.09 0.151 8,578 282 Marco-Nano-Instruct-REAP15-Q4_K_M.gguf 197 4.30 14.03 0.214 9,378 241 Marco-Nano-Instruct-REAP20-Q4_K_M.gguf 186 4.05 15.36 0.291 9,402 271 Marco-Nano-Instruct-REAP30-Q4_K_M.gguf 162 3.58 18.03 0.458 9,681 268 Marco-Nano-Instruct-REAP40-Q4_K_M.gguf 139 3.11 22.71 0.671 10,185 279 Marco-Nano-Instruct-REAP50-Q4_K_M.gguf 116 2.61 30.38 0.941 10,157 296 Marco-Nano-Instruct-REAP60-Q4_K_M.gguf 93 2.14 45.04 1.289 11,146 296 Marco-Nano-Instruct-REAP70-Q4_K_M.gguf 70 1.66 74.79 1.718 11,416 302",
   "retrieved": "2026-10-06",
   "source_name": "HF model card: kueizen/Marco-Nano-Instruct-REAP-GGUF",
   "source_url": "https://huggingface.co/kueizen/Marco-Nano-Instruct-REAP-GGUF",
   "tg_tokens": 128,
   "tps": 282
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hf1-marcnano-reap15",
   "model": "Marco-Nano-Instruct",
   "notes": "kueizen REAP expert-pruned GGUF, llama-bench batch 1, all layers on GPU; config=reap15; REAP15 (197 experts, 4.30 GB, PPL 14.03).",
   "pp_tokens": 512,
   "pp_tps": 9378,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Nine files from 4.8 GB down to 1.7 GB, all Q4_K_M, all measured on one RTX 3090; File Experts kept Size (GB) PPL KLD (nats) pp512 (t/s) tg128 (t/s) unpruned base (not in this repo) 232 5.00 11.48 \u2014 8,460 277 Marco-Nano-Instruct-REAP05-Q4_K_M.gguf 220 4.77 12.07 0.065 9,142 282 Marco-Nano-Instruct-REAP10-Q4_K_M.gguf 209 4.52 13.09 0.151 8,578 282 Marco-Nano-Instruct-REAP15-Q4_K_M.gguf 197 4.30 14.03 0.214 9,378 241 Marco-Nano-Instruct-REAP20-Q4_K_M.gguf 186 4.05 15.36 0.291 9,402 271 Marco-Nano-Instruct-REAP30-Q4_K_M.gguf 162 3.58 18.03 0.458 9,681 268 Marco-Nano-Instruct-REAP40-Q4_K_M.gguf 139 3.11 22.71 0.671 10,185 279 Marco-Nano-Instruct-REAP50-Q4_K_M.gguf 116 2.61 30.38 0.941 10,157 296 Marco-Nano-Instruct-REAP60-Q4_K_M.gguf 93 2.14 45.04 1.289 11,146 296 Marco-Nano-Instruct-REAP70-Q4_K_M.gguf 70 1.66 74.79 1.718 11,416 302",
   "retrieved": "2026-10-06",
   "source_name": "HF model card: kueizen/Marco-Nano-Instruct-REAP-GGUF",
   "source_url": "https://huggingface.co/kueizen/Marco-Nano-Instruct-REAP-GGUF",
   "tg_tokens": 128,
   "tps": 241
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hf1-marcnano-reap20",
   "model": "Marco-Nano-Instruct",
   "notes": "kueizen REAP expert-pruned GGUF, llama-bench batch 1, all layers on GPU; config=reap20; REAP20 (186 experts, 4.05 GB, PPL 15.36).",
   "pp_tokens": 512,
   "pp_tps": 9402,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Nine files from 4.8 GB down to 1.7 GB, all Q4_K_M, all measured on one RTX 3090; File Experts kept Size (GB) PPL KLD (nats) pp512 (t/s) tg128 (t/s) unpruned base (not in this repo) 232 5.00 11.48 \u2014 8,460 277 Marco-Nano-Instruct-REAP05-Q4_K_M.gguf 220 4.77 12.07 0.065 9,142 282 Marco-Nano-Instruct-REAP10-Q4_K_M.gguf 209 4.52 13.09 0.151 8,578 282 Marco-Nano-Instruct-REAP15-Q4_K_M.gguf 197 4.30 14.03 0.214 9,378 241 Marco-Nano-Instruct-REAP20-Q4_K_M.gguf 186 4.05 15.36 0.291 9,402 271 Marco-Nano-Instruct-REAP30-Q4_K_M.gguf 162 3.58 18.03 0.458 9,681 268 Marco-Nano-Instruct-REAP40-Q4_K_M.gguf 139 3.11 22.71 0.671 10,185 279 Marco-Nano-Instruct-REAP50-Q4_K_M.gguf 116 2.61 30.38 0.941 10,157 296 Marco-Nano-Instruct-REAP60-Q4_K_M.gguf 93 2.14 45.04 1.289 11,146 296 Marco-Nano-Instruct-REAP70-Q4_K_M.gguf 70 1.66 74.79 1.718 11,416 302",
   "retrieved": "2026-10-06",
   "source_name": "HF model card: kueizen/Marco-Nano-Instruct-REAP-GGUF",
   "source_url": "https://huggingface.co/kueizen/Marco-Nano-Instruct-REAP-GGUF",
   "tg_tokens": 128,
   "tps": 271
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hf1-marcnano-reap30",
   "model": "Marco-Nano-Instruct",
   "notes": "kueizen REAP expert-pruned GGUF, llama-bench batch 1, all layers on GPU; config=reap30; REAP30 (162 experts, 3.58 GB, PPL 18.03).",
   "pp_tokens": 512,
   "pp_tps": 9681,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Nine files from 4.8 GB down to 1.7 GB, all Q4_K_M, all measured on one RTX 3090; File Experts kept Size (GB) PPL KLD (nats) pp512 (t/s) tg128 (t/s) unpruned base (not in this repo) 232 5.00 11.48 \u2014 8,460 277 Marco-Nano-Instruct-REAP05-Q4_K_M.gguf 220 4.77 12.07 0.065 9,142 282 Marco-Nano-Instruct-REAP10-Q4_K_M.gguf 209 4.52 13.09 0.151 8,578 282 Marco-Nano-Instruct-REAP15-Q4_K_M.gguf 197 4.30 14.03 0.214 9,378 241 Marco-Nano-Instruct-REAP20-Q4_K_M.gguf 186 4.05 15.36 0.291 9,402 271 Marco-Nano-Instruct-REAP30-Q4_K_M.gguf 162 3.58 18.03 0.458 9,681 268 Marco-Nano-Instruct-REAP40-Q4_K_M.gguf 139 3.11 22.71 0.671 10,185 279 Marco-Nano-Instruct-REAP50-Q4_K_M.gguf 116 2.61 30.38 0.941 10,157 296 Marco-Nano-Instruct-REAP60-Q4_K_M.gguf 93 2.14 45.04 1.289 11,146 296 Marco-Nano-Instruct-REAP70-Q4_K_M.gguf 70 1.66 74.79 1.718 11,416 302",
   "retrieved": "2026-10-06",
   "source_name": "HF model card: kueizen/Marco-Nano-Instruct-REAP-GGUF",
   "source_url": "https://huggingface.co/kueizen/Marco-Nano-Instruct-REAP-GGUF",
   "tg_tokens": 128,
   "tps": 268
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hf1-marcnano-reap40",
   "model": "Marco-Nano-Instruct",
   "notes": "kueizen REAP expert-pruned GGUF, llama-bench batch 1, all layers on GPU; config=reap40; REAP40 (139 experts, 3.11 GB, PPL 22.71).",
   "pp_tokens": 512,
   "pp_tps": 10185,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Nine files from 4.8 GB down to 1.7 GB, all Q4_K_M, all measured on one RTX 3090; File Experts kept Size (GB) PPL KLD (nats) pp512 (t/s) tg128 (t/s) unpruned base (not in this repo) 232 5.00 11.48 \u2014 8,460 277 Marco-Nano-Instruct-REAP05-Q4_K_M.gguf 220 4.77 12.07 0.065 9,142 282 Marco-Nano-Instruct-REAP10-Q4_K_M.gguf 209 4.52 13.09 0.151 8,578 282 Marco-Nano-Instruct-REAP15-Q4_K_M.gguf 197 4.30 14.03 0.214 9,378 241 Marco-Nano-Instruct-REAP20-Q4_K_M.gguf 186 4.05 15.36 0.291 9,402 271 Marco-Nano-Instruct-REAP30-Q4_K_M.gguf 162 3.58 18.03 0.458 9,681 268 Marco-Nano-Instruct-REAP40-Q4_K_M.gguf 139 3.11 22.71 0.671 10,185 279 Marco-Nano-Instruct-REAP50-Q4_K_M.gguf 116 2.61 30.38 0.941 10,157 296 Marco-Nano-Instruct-REAP60-Q4_K_M.gguf 93 2.14 45.04 1.289 11,146 296 Marco-Nano-Instruct-REAP70-Q4_K_M.gguf 70 1.66 74.79 1.718 11,416 302",
   "retrieved": "2026-10-06",
   "source_name": "HF model card: kueizen/Marco-Nano-Instruct-REAP-GGUF",
   "source_url": "https://huggingface.co/kueizen/Marco-Nano-Instruct-REAP-GGUF",
   "tg_tokens": 128,
   "tps": 279
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hf1-marcnano-reap50",
   "model": "Marco-Nano-Instruct",
   "notes": "kueizen REAP expert-pruned GGUF, llama-bench batch 1, all layers on GPU; config=reap50; REAP50 (116 experts, 2.61 GB, PPL 30.38).",
   "pp_tokens": 512,
   "pp_tps": 10157,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Nine files from 4.8 GB down to 1.7 GB, all Q4_K_M, all measured on one RTX 3090; File Experts kept Size (GB) PPL KLD (nats) pp512 (t/s) tg128 (t/s) unpruned base (not in this repo) 232 5.00 11.48 \u2014 8,460 277 Marco-Nano-Instruct-REAP05-Q4_K_M.gguf 220 4.77 12.07 0.065 9,142 282 Marco-Nano-Instruct-REAP10-Q4_K_M.gguf 209 4.52 13.09 0.151 8,578 282 Marco-Nano-Instruct-REAP15-Q4_K_M.gguf 197 4.30 14.03 0.214 9,378 241 Marco-Nano-Instruct-REAP20-Q4_K_M.gguf 186 4.05 15.36 0.291 9,402 271 Marco-Nano-Instruct-REAP30-Q4_K_M.gguf 162 3.58 18.03 0.458 9,681 268 Marco-Nano-Instruct-REAP40-Q4_K_M.gguf 139 3.11 22.71 0.671 10,185 279 Marco-Nano-Instruct-REAP50-Q4_K_M.gguf 116 2.61 30.38 0.941 10,157 296 Marco-Nano-Instruct-REAP60-Q4_K_M.gguf 93 2.14 45.04 1.289 11,146 296 Marco-Nano-Instruct-REAP70-Q4_K_M.gguf 70 1.66 74.79 1.718 11,416 302",
   "retrieved": "2026-10-06",
   "source_name": "HF model card: kueizen/Marco-Nano-Instruct-REAP-GGUF",
   "source_url": "https://huggingface.co/kueizen/Marco-Nano-Instruct-REAP-GGUF",
   "tg_tokens": 128,
   "tps": 296
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hf1-marcnano-reap60",
   "model": "Marco-Nano-Instruct",
   "notes": "kueizen REAP expert-pruned GGUF, llama-bench batch 1, all layers on GPU; config=reap60; REAP60 (93 experts, 2.14 GB, PPL 45.04).",
   "pp_tokens": 512,
   "pp_tps": 11146,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Nine files from 4.8 GB down to 1.7 GB, all Q4_K_M, all measured on one RTX 3090; File Experts kept Size (GB) PPL KLD (nats) pp512 (t/s) tg128 (t/s) unpruned base (not in this repo) 232 5.00 11.48 \u2014 8,460 277 Marco-Nano-Instruct-REAP05-Q4_K_M.gguf 220 4.77 12.07 0.065 9,142 282 Marco-Nano-Instruct-REAP10-Q4_K_M.gguf 209 4.52 13.09 0.151 8,578 282 Marco-Nano-Instruct-REAP15-Q4_K_M.gguf 197 4.30 14.03 0.214 9,378 241 Marco-Nano-Instruct-REAP20-Q4_K_M.gguf 186 4.05 15.36 0.291 9,402 271 Marco-Nano-Instruct-REAP30-Q4_K_M.gguf 162 3.58 18.03 0.458 9,681 268 Marco-Nano-Instruct-REAP40-Q4_K_M.gguf 139 3.11 22.71 0.671 10,185 279 Marco-Nano-Instruct-REAP50-Q4_K_M.gguf 116 2.61 30.38 0.941 10,157 296 Marco-Nano-Instruct-REAP60-Q4_K_M.gguf 93 2.14 45.04 1.289 11,146 296 Marco-Nano-Instruct-REAP70-Q4_K_M.gguf 70 1.66 74.79 1.718 11,416 302",
   "retrieved": "2026-10-06",
   "source_name": "HF model card: kueizen/Marco-Nano-Instruct-REAP-GGUF",
   "source_url": "https://huggingface.co/kueizen/Marco-Nano-Instruct-REAP-GGUF",
   "tg_tokens": 128,
   "tps": 296
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hf1-marcnano-reap70",
   "model": "Marco-Nano-Instruct",
   "notes": "kueizen REAP expert-pruned GGUF, llama-bench batch 1, all layers on GPU; config=reap70; REAP70 (70 experts, 1.66 GB, PPL 74.79).",
   "pp_tokens": 512,
   "pp_tps": 11416,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Nine files from 4.8 GB down to 1.7 GB, all Q4_K_M, all measured on one RTX 3090; File Experts kept Size (GB) PPL KLD (nats) pp512 (t/s) tg128 (t/s) unpruned base (not in this repo) 232 5.00 11.48 \u2014 8,460 277 Marco-Nano-Instruct-REAP05-Q4_K_M.gguf 220 4.77 12.07 0.065 9,142 282 Marco-Nano-Instruct-REAP10-Q4_K_M.gguf 209 4.52 13.09 0.151 8,578 282 Marco-Nano-Instruct-REAP15-Q4_K_M.gguf 197 4.30 14.03 0.214 9,378 241 Marco-Nano-Instruct-REAP20-Q4_K_M.gguf 186 4.05 15.36 0.291 9,402 271 Marco-Nano-Instruct-REAP30-Q4_K_M.gguf 162 3.58 18.03 0.458 9,681 268 Marco-Nano-Instruct-REAP40-Q4_K_M.gguf 139 3.11 22.71 0.671 10,185 279 Marco-Nano-Instruct-REAP50-Q4_K_M.gguf 116 2.61 30.38 0.941 10,157 296 Marco-Nano-Instruct-REAP60-Q4_K_M.gguf 93 2.14 45.04 1.289 11,146 296 Marco-Nano-Instruct-REAP70-Q4_K_M.gguf 70 1.66 74.79 1.718 11,416 302",
   "retrieved": "2026-10-06",
   "source_name": "HF model card: kueizen/Marco-Nano-Instruct-REAP-GGUF",
   "source_url": "https://huggingface.co/kueizen/Marco-Nano-Instruct-REAP-GGUF",
   "tg_tokens": 128,
   "tps": 302
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "A100 PCIe 80 GB",
   "id": "hf1-firetussin-pi-rco",
   "model": "Qwen3.8-27B-pi",
   "notes": "firetussin RCO allocation-transfer quant (experimental, 3.00 BPW, 10.094 GB); llama.cpp CUDA, FA on, q8_0 K/V, 8K ctx, single-run, Vast A100 ($2.46/2.09h); config=rco; prefill +15% vs matched IQ2_M control, decode flat.",
   "pp_tokens": 512,
   "pp_tps": 1088,
   "provenance": "community",
   "quant": "IQ3_XXS",
   "quote": "One on-demand NVIDIA A100 PCIe 80 GB was used for about 2.092 hours; Qwen3.8-27B-pi RCO allocation transfer; the transfer model processed a 512-token prompt at 1,088 tok/s and generated at 48.8 tok/s",
   "retrieved": "2026-10-06",
   "source_name": "HF model card: firetussin/Qwen3.8-27B-pi-RCO-allocation-transfer-GGUF",
   "source_url": "https://huggingface.co/firetussin/Qwen3.8-27B-pi-RCO-allocation-transfer-GGUF",
   "tg_tokens": null,
   "tps": 48.8
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "A100 PCIe 80 GB",
   "id": "hf1-firetussin-pi-iq2m",
   "model": "Qwen3.8-27B-pi",
   "notes": "firetussin same-source/same-imatrix conventional IQ2_M control (10.004 GB); llama.cpp CUDA, FA on, q8_0 K/V, 8K ctx, single-run, Vast A100; config=iq2m; baseline for the RCO transfer comparison.",
   "pp_tokens": 512,
   "pp_tps": 945,
   "provenance": "community",
   "quant": "IQ2_M",
   "quote": "One on-demand NVIDIA A100 PCIe 80 GB was used for about 2.092 hours; Qwen3.8-27B-pi RCO allocation transfer; The matched IQ2_M control measured 945 tok/s and 48.9 tok/s respectively",
   "retrieved": "2026-10-06",
   "source_name": "HF model card: firetussin/Qwen3.8-27B-pi-RCO-allocation-transfer-GGUF",
   "source_url": "https://huggingface.co/firetussin/Qwen3.8-27B-pi-RCO-allocation-transfer-GGUF",
   "tg_tokens": null,
   "tps": 48.9
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "16\u00d7 A100-SXM4-40GB",
   "id": "vllm-60187-a100x16-gsm8k-main",
   "model": "K-EXAONE-236B-A23B",
   "notes": "K-EXAONE RoPE bugfix, official HF config, GSM8K, c=16; main (MTP NoPE); MTP num_speculative_tokens=1, RoPE on MTP layer (the fix); config=c16-gsm8k-main",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "FP8",
   "quote": "16\u00d7 A100-SXM4-40GB; K-EXAONE-236B-A23B FP8 (compressed-tensors), TP=8; official, GSM8K  85.01% (1062 tok/s)  88.28% (1081 tok/s)",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "vLLM PR #60187",
   "source_url": "https://github.com/vllm-project/vllm/pull/60187",
   "tg_tokens": null,
   "tps": 1062,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "16\u00d7 A100-SXM4-40GB",
   "id": "vllm-60187-a100x16-gsm8k-pr",
   "model": "K-EXAONE-236B-A23B",
   "notes": "K-EXAONE RoPE bugfix, official HF config, GSM8K, c=16; this PR (MTP RoPE); MTP num_speculative_tokens=1, RoPE on MTP layer (the fix); config=c16-gsm8k-pr",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "FP8",
   "quote": "16\u00d7 A100-SXM4-40GB; K-EXAONE-236B-A23B FP8 (compressed-tensors), TP=8; official, GSM8K  85.01% (1062 tok/s)  88.28% (1081 tok/s)",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "vLLM PR #60187",
   "source_url": "https://github.com/vllm-project/vllm/pull/60187",
   "tg_tokens": null,
   "tps": 1081,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "16\u00d7 A100-SXM4-40GB",
   "id": "vllm-60187-a100x16-mtb-main",
   "model": "K-EXAONE-236B-A23B",
   "notes": "K-EXAONE RoPE bugfix, official HF config, MT-Bench, c=16; main (MTP NoPE); MTP num_speculative_tokens=1, RoPE on MTP layer (the fix); config=c16-mtb-main",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "FP8",
   "quote": "16\u00d7 A100-SXM4-40GB; K-EXAONE-236B-A23B FP8 (compressed-tensors), TP=8; official, MT-Bench  67.86% (940 tok/s)  73.07% (970 tok/s)",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "vLLM PR #60187",
   "source_url": "https://github.com/vllm-project/vllm/pull/60187",
   "tg_tokens": null,
   "tps": 940,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "16\u00d7 A100-SXM4-40GB",
   "id": "vllm-60187-a100x16-mtb-pr",
   "model": "K-EXAONE-236B-A23B",
   "notes": "K-EXAONE RoPE bugfix, official HF config, MT-Bench, c=16; this PR (MTP RoPE); MTP num_speculative_tokens=1, RoPE on MTP layer (the fix); config=c16-mtb-pr",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "FP8",
   "quote": "16\u00d7 A100-SXM4-40GB; K-EXAONE-236B-A23B FP8 (compressed-tensors), TP=8; official, MT-Bench  67.86% (940 tok/s)  73.07% (970 tok/s)",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "vLLM PR #60187",
   "source_url": "https://github.com/vllm-project/vllm/pull/60187",
   "tg_tokens": null,
   "tps": 970,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 16,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "lc30042-rtxpro6000-faon-8192-before",
   "model": "GLM-5.3-Flash",
   "notes": "RTX PRO 6000, llama.cpp; removes the obsolete gather path in the glm5-next sparse attention (gather_mla_rows); -fa on, -ub 16, ctx 8192; before (gather present); config=before-faon-8192",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ1_S",
   "quote": "RTX PRO 6000; GLM-5.3-Flash UD-IQ1_S; 8192 on 75.3 t/s 84.3 t/s",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30042",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30042",
   "tg_tokens": null,
   "tps": 75.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 16,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "lc30042-rtxpro6000-faon-8192-after",
   "model": "GLM-5.3-Flash",
   "notes": "RTX PRO 6000, llama.cpp; removes the obsolete gather path in the glm5-next sparse attention (gather_mla_rows); -fa on, -ub 16, ctx 8192; after (gather removed); config=after-faon-8192",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ1_S",
   "quote": "RTX PRO 6000; GLM-5.3-Flash UD-IQ1_S; 8192 on 75.3 t/s 84.3 t/s",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30042",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30042",
   "tg_tokens": null,
   "tps": 84.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 16,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "lc30042-rtxpro6000-faon-32768-before",
   "model": "GLM-5.3-Flash",
   "notes": "RTX PRO 6000, llama.cpp; removes the obsolete gather path in the glm5-next sparse attention (gather_mla_rows); -fa on, -ub 16, ctx 32768; before (gather present); config=before-faon-32768",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ1_S",
   "quote": "RTX PRO 6000; GLM-5.3-Flash UD-IQ1_S; 32768 on 71.9 t/s 80.1 t/s",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30042",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30042",
   "tg_tokens": null,
   "tps": 71.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 16,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "lc30042-rtxpro6000-faon-32768-after",
   "model": "GLM-5.3-Flash",
   "notes": "RTX PRO 6000, llama.cpp; removes the obsolete gather path in the glm5-next sparse attention (gather_mla_rows); -fa on, -ub 16, ctx 32768; after (gather removed); config=after-faon-32768",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ1_S",
   "quote": "RTX PRO 6000; GLM-5.3-Flash UD-IQ1_S; 32768 on 71.9 t/s 80.1 t/s",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30042",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30042",
   "tg_tokens": null,
   "tps": 80.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 16,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "lc30042-rtxpro6000-faoff-32768-before",
   "model": "GLM-5.3-Flash",
   "notes": "RTX PRO 6000, llama.cpp; removes the obsolete gather path in the glm5-next sparse attention (gather_mla_rows); -fa off, -ub 16, ctx 32768; before (gather present); config=before-faoff-32768",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ1_S",
   "quote": "RTX PRO 6000; GLM-5.3-Flash UD-IQ1_S; 32768 off 72.3 t/s 24.6 t/s",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30042",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30042",
   "tg_tokens": null,
   "tps": 72.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 16,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "lc30042-rtxpro6000-faoff-32768-after",
   "model": "GLM-5.3-Flash",
   "notes": "RTX PRO 6000, llama.cpp; removes the obsolete gather path in the glm5-next sparse attention (gather_mla_rows); -fa off, -ub 16, ctx 32768; after (gather removed); config=after-faoff-32768",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ1_S",
   "quote": "RTX PRO 6000; GLM-5.3-Flash UD-IQ1_S; 32768 off 72.3 t/s 24.6 t/s",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30042",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30042",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hf1-qwen3635ba3b-reap05",
   "model": "Qwen3.6-35B-A3B",
   "notes": "kueizen REAP expert-pruned GGUF, llama-bench batch 1, all layers on GPU; config=reap05; REAP05 (244 experts, 18.86 GB, PPL 6.63).",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2900,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Nine files from 18.9 GB down to 7.0 GB, all Q4_K_M, all measured on one RTX 3090; Qwen3.6-35B-A3B-REAP05-Q4_K_M.gguf 244 18.86 6.63 2,900 117",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "HF model card: kueizen/Qwen3.6-35B-A3B-REAP-GGUF",
   "source_url": "https://huggingface.co/kueizen/Qwen3.6-35B-A3B-REAP-GGUF",
   "tg_tokens": 128,
   "tps": 117,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hf1-qwen3635ba3b-reap10",
   "model": "Qwen3.6-35B-A3B",
   "notes": "kueizen REAP expert-pruned GGUF, llama-bench batch 1, all layers on GPU; config=reap10; REAP10 (230 experts, 17.93 GB, PPL 7.03).",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2977,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Nine files from 18.9 GB down to 7.0 GB, all Q4_K_M, all measured on one RTX 3090; Qwen3.6-35B-A3B-REAP10-Q4_K_M.gguf 230 17.93 7.03 2,977 118",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "HF model card: kueizen/Qwen3.6-35B-A3B-REAP-GGUF",
   "source_url": "https://huggingface.co/kueizen/Qwen3.6-35B-A3B-REAP-GGUF",
   "tg_tokens": 128,
   "tps": 118,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hf1-qwen3635ba3b-reap15",
   "model": "Qwen3.6-35B-A3B",
   "notes": "kueizen REAP expert-pruned GGUF, llama-bench batch 1, all layers on GPU; config=reap15; REAP15 (218 experts, 17.00 GB, PPL 7.55).",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2997,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Nine files from 18.9 GB down to 7.0 GB, all Q4_K_M, all measured on one RTX 3090; Qwen3.6-35B-A3B-REAP15-Q4_K_M.gguf 218 17.00 7.55 2,997 118",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "HF model card: kueizen/Qwen3.6-35B-A3B-REAP-GGUF",
   "source_url": "https://huggingface.co/kueizen/Qwen3.6-35B-A3B-REAP-GGUF",
   "tg_tokens": 128,
   "tps": 118,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hf1-qwen3635ba3b-reap20",
   "model": "Qwen3.6-35B-A3B",
   "notes": "kueizen REAP expert-pruned GGUF, llama-bench batch 1, all layers on GPU; config=reap20; REAP20 (205 experts, 16.08 GB, PPL 8.14).",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3092,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Nine files from 18.9 GB down to 7.0 GB, all Q4_K_M, all measured on one RTX 3090; Qwen3.6-35B-A3B-REAP20-Q4_K_M.gguf 205 16.08 8.14 3,092 118",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "HF model card: kueizen/Qwen3.6-35B-A3B-REAP-GGUF",
   "source_url": "https://huggingface.co/kueizen/Qwen3.6-35B-A3B-REAP-GGUF",
   "tg_tokens": 128,
   "tps": 118,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hf1-qwen3635ba3b-reap30",
   "model": "Qwen3.6-35B-A3B",
   "notes": "kueizen REAP expert-pruned GGUF, llama-bench batch 1, all layers on GPU; config=reap30; REAP30 (179 experts, 14.30 GB, PPL 9.41).",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3192,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Nine files from 18.9 GB down to 7.0 GB, all Q4_K_M, all measured on one RTX 3090; Qwen3.6-35B-A3B-REAP30-Q4_K_M.gguf 179 14.30 9.41 3,192 118",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "HF model card: kueizen/Qwen3.6-35B-A3B-REAP-GGUF",
   "source_url": "https://huggingface.co/kueizen/Qwen3.6-35B-A3B-REAP-GGUF",
   "tg_tokens": 128,
   "tps": 118,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hf1-qwen3635ba3b-reap40",
   "model": "Qwen3.6-35B-A3B",
   "notes": "kueizen REAP expert-pruned GGUF, llama-bench batch 1, all layers on GPU; config=reap40; REAP40 (154 experts, 12.44 GB, PPL 11.55).",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3320,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Nine files from 18.9 GB down to 7.0 GB, all Q4_K_M, all measured on one RTX 3090; Qwen3.6-35B-A3B-REAP40-Q4_K_M.gguf 154 12.44 11.55 3,320 119",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "HF model card: kueizen/Qwen3.6-35B-A3B-REAP-GGUF",
   "source_url": "https://huggingface.co/kueizen/Qwen3.6-35B-A3B-REAP-GGUF",
   "tg_tokens": 128,
   "tps": 119,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hf1-qwen3635ba3b-reap50",
   "model": "Qwen3.6-35B-A3B",
   "notes": "kueizen REAP expert-pruned GGUF, llama-bench batch 1, all layers on GPU; config=reap50; REAP50 (128 experts, 10.59 GB, PPL 14.55).",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3384,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Nine files from 18.9 GB down to 7.0 GB, all Q4_K_M, all measured on one RTX 3090; Qwen3.6-35B-A3B-REAP50-Q4_K_M.gguf 128 10.59 14.55 3,384 130",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "HF model card: kueizen/Qwen3.6-35B-A3B-REAP-GGUF",
   "source_url": "https://huggingface.co/kueizen/Qwen3.6-35B-A3B-REAP-GGUF",
   "tg_tokens": 128,
   "tps": 130,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hf1-qwen3635ba3b-reap60",
   "model": "Qwen3.6-35B-A3B",
   "notes": "kueizen REAP expert-pruned GGUF, llama-bench batch 1, all layers on GPU; config=reap60; REAP60 (102 experts, 8.81 GB, PPL 17.70).",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3597,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Nine files from 18.9 GB down to 7.0 GB, all Q4_K_M, all measured on one RTX 3090; Qwen3.6-35B-A3B-REAP60-Q4_K_M.gguf 102 8.81 17.70 3,597 120",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "HF model card: kueizen/Qwen3.6-35B-A3B-REAP-GGUF",
   "source_url": "https://huggingface.co/kueizen/Qwen3.6-35B-A3B-REAP-GGUF",
   "tg_tokens": 128,
   "tps": 120,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hf1-qwen3635ba3b-reap70",
   "model": "Qwen3.6-35B-A3B",
   "notes": "kueizen REAP expert-pruned GGUF, llama-bench batch 1, all layers on GPU; config=reap70; REAP70 (77 experts, 6.96 GB, PPL 24.39).",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3735,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Nine files from 18.9 GB down to 7.0 GB, all Q4_K_M, all measured on one RTX 3090; Qwen3.6-35B-A3B-REAP70-Q4_K_M.gguf 77 6.96 24.39 3,735 121",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "HF model card: kueizen/Qwen3.6-35B-A3B-REAP-GGUF",
   "source_url": "https://huggingface.co/kueizen/Qwen3.6-35B-A3B-REAP-GGUF",
   "tg_tokens": 128,
   "tps": 121,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Apple M3 Ultra",
   "id": "lc30047-q36a3b-seq1-master",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Apple M3 Ultra (60-core GPU), llama.cpp (Metal); llama-batched-bench -npp 512 -ntg 32, Qwen3.6-35B-A3B Q4_K_M text generation, seq 1; A/B master vs PR (few-row MMA MUL_MAT_ID kernels); master arm; config=seq1-master",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Apple M3 Ultra (60-core GPU), macOS 15.7.9; Qwen3.6-35B-A3B Q4_K_M, text generation; master (t/s) 95.5 145.9 185.0 213.6 297.3 346.8",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30047",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30047",
   "tg_tokens": 32,
   "tps": 95.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Apple M3 Ultra",
   "id": "lc30047-q36a3b-seq1-pr",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Apple M3 Ultra (60-core GPU), llama.cpp (Metal); llama-batched-bench -npp 512 -ntg 32, Qwen3.6-35B-A3B Q4_K_M text generation, seq 1; A/B master vs PR (few-row MMA MUL_MAT_ID kernels); pr arm; config=seq1-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Apple M3 Ultra (60-core GPU), macOS 15.7.9; Qwen3.6-35B-A3B Q4_K_M, text generation; this PR (t/s) 95.1 145.4 184.7 213.3 295.9 365.5",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30047",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30047",
   "tg_tokens": 32,
   "tps": 95.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 2,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Apple M3 Ultra",
   "id": "lc30047-q36a3b-seq2-master",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Apple M3 Ultra (60-core GPU), llama.cpp (Metal); llama-batched-bench -npp 512 -ntg 32, Qwen3.6-35B-A3B Q4_K_M text generation, seq 2; A/B master vs PR (few-row MMA MUL_MAT_ID kernels); master arm; config=seq2-master",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Apple M3 Ultra (60-core GPU), macOS 15.7.9; Qwen3.6-35B-A3B Q4_K_M, text generation; master (t/s) 95.5 145.9 185.0 213.6 297.3 346.8",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30047",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30047",
   "tg_tokens": 32,
   "tps": 145.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 2,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Apple M3 Ultra",
   "id": "lc30047-q36a3b-seq2-pr",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Apple M3 Ultra (60-core GPU), llama.cpp (Metal); llama-batched-bench -npp 512 -ntg 32, Qwen3.6-35B-A3B Q4_K_M text generation, seq 2; A/B master vs PR (few-row MMA MUL_MAT_ID kernels); pr arm; config=seq2-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Apple M3 Ultra (60-core GPU), macOS 15.7.9; Qwen3.6-35B-A3B Q4_K_M, text generation; this PR (t/s) 95.1 145.4 184.7 213.3 295.9 365.5",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30047",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30047",
   "tg_tokens": 32,
   "tps": 145.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 3,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Apple M3 Ultra",
   "id": "lc30047-q36a3b-seq3-master",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Apple M3 Ultra (60-core GPU), llama.cpp (Metal); llama-batched-bench -npp 512 -ntg 32, Qwen3.6-35B-A3B Q4_K_M text generation, seq 3; A/B master vs PR (few-row MMA MUL_MAT_ID kernels); master arm; config=seq3-master",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Apple M3 Ultra (60-core GPU), macOS 15.7.9; Qwen3.6-35B-A3B Q4_K_M, text generation; master (t/s) 95.5 145.9 185.0 213.6 297.3 346.8",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30047",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30047",
   "tg_tokens": 32,
   "tps": 185.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 3,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Apple M3 Ultra",
   "id": "lc30047-q36a3b-seq3-pr",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Apple M3 Ultra (60-core GPU), llama.cpp (Metal); llama-batched-bench -npp 512 -ntg 32, Qwen3.6-35B-A3B Q4_K_M text generation, seq 3; A/B master vs PR (few-row MMA MUL_MAT_ID kernels); pr arm; config=seq3-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Apple M3 Ultra (60-core GPU), macOS 15.7.9; Qwen3.6-35B-A3B Q4_K_M, text generation; this PR (t/s) 95.1 145.4 184.7 213.3 295.9 365.5",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30047",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30047",
   "tg_tokens": 32,
   "tps": 184.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 4,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Apple M3 Ultra",
   "id": "lc30047-q36a3b-seq4-master",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Apple M3 Ultra (60-core GPU), llama.cpp (Metal); llama-batched-bench -npp 512 -ntg 32, Qwen3.6-35B-A3B Q4_K_M text generation, seq 4; A/B master vs PR (few-row MMA MUL_MAT_ID kernels); master arm; config=seq4-master",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Apple M3 Ultra (60-core GPU), macOS 15.7.9; Qwen3.6-35B-A3B Q4_K_M, text generation; master (t/s) 95.5 145.9 185.0 213.6 297.3 346.8",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30047",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30047",
   "tg_tokens": 32,
   "tps": 213.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 4,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Apple M3 Ultra",
   "id": "lc30047-q36a3b-seq4-pr",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Apple M3 Ultra (60-core GPU), llama.cpp (Metal); llama-batched-bench -npp 512 -ntg 32, Qwen3.6-35B-A3B Q4_K_M text generation, seq 4; A/B master vs PR (few-row MMA MUL_MAT_ID kernels); pr arm; config=seq4-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Apple M3 Ultra (60-core GPU), macOS 15.7.9; Qwen3.6-35B-A3B Q4_K_M, text generation; this PR (t/s) 95.1 145.4 184.7 213.3 295.9 365.5",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30047",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30047",
   "tg_tokens": 32,
   "tps": 213.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 8,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Apple M3 Ultra",
   "id": "lc30047-q36a3b-seq8-master",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Apple M3 Ultra (60-core GPU), llama.cpp (Metal); llama-batched-bench -npp 512 -ntg 32, Qwen3.6-35B-A3B Q4_K_M text generation, seq 8; A/B master vs PR (few-row MMA MUL_MAT_ID kernels); master arm; config=seq8-master",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Apple M3 Ultra (60-core GPU), macOS 15.7.9; Qwen3.6-35B-A3B Q4_K_M, text generation; master (t/s) 95.5 145.9 185.0 213.6 297.3 346.8",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30047",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30047",
   "tg_tokens": 32,
   "tps": 297.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 8,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Apple M3 Ultra",
   "id": "lc30047-q36a3b-seq8-pr",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Apple M3 Ultra (60-core GPU), llama.cpp (Metal); llama-batched-bench -npp 512 -ntg 32, Qwen3.6-35B-A3B Q4_K_M text generation, seq 8; A/B master vs PR (few-row MMA MUL_MAT_ID kernels); pr arm; config=seq8-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Apple M3 Ultra (60-core GPU), macOS 15.7.9; Qwen3.6-35B-A3B Q4_K_M, text generation; this PR (t/s) 95.1 145.4 184.7 213.3 295.9 365.5",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30047",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30047",
   "tg_tokens": 32,
   "tps": 295.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 16,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Apple M3 Ultra",
   "id": "lc30047-q36a3b-seq16-master",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Apple M3 Ultra (60-core GPU), llama.cpp (Metal); llama-batched-bench -npp 512 -ntg 32, Qwen3.6-35B-A3B Q4_K_M text generation, seq 16; A/B master vs PR (few-row MMA MUL_MAT_ID kernels); master arm; config=seq16-master",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Apple M3 Ultra (60-core GPU), macOS 15.7.9; Qwen3.6-35B-A3B Q4_K_M, text generation; master (t/s) 95.5 145.9 185.0 213.6 297.3 346.8",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30047",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30047",
   "tg_tokens": 32,
   "tps": 346.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 16,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Apple M3 Ultra",
   "id": "lc30047-q36a3b-seq16-pr",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Apple M3 Ultra (60-core GPU), llama.cpp (Metal); llama-batched-bench -npp 512 -ntg 32, Qwen3.6-35B-A3B Q4_K_M text generation, seq 16; A/B master vs PR (few-row MMA MUL_MAT_ID kernels); pr arm; config=seq16-pr",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Apple M3 Ultra (60-core GPU), macOS 15.7.9; Qwen3.6-35B-A3B Q4_K_M, text generation; this PR (t/s) 95.1 145.4 184.7 213.3 295.9 365.5",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp PR #30047",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30047",
   "tg_tokens": 32,
   "tps": 365.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": 1048576,
   "date": "2026-10",
   "flags": [],
   "hardware": "Ryzen AI Max+ 395 (Strix Halo)",
   "id": "lc30048-q35b-strixhalo",
   "model": "Qwen3.5-2B",
   "notes": "AMD Strix Halo (Ryzen AI Max+ 395, Radeon 8060S iGPU), llama.cpp Vulkan, Qwen3.5-2B UD-Q4_K_XL; long-context prefill from a 2M-ctx FA quantized-KV OOM repro (bug #30048): -fa 1 -ctk q8_0 -ctv q8_0 -c 2097152, ~1.89M-token prompt, -ub 512; pp_tps is the cumulative prompt-processing rate over the 1M tokens (n=1048576) processed before the 2 GiB combined K+V scratch allocation failed; -n 1, generation not run",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 213.69,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "AMD Radeon 8060S / Ryzen AI Max+ 395 (Strix Halo); Qwen3.5-2B; prompt processing, n_tokens = 1048576, progress = 0.56, t = 4906.91 s / 213.69 tokens per second",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp issue #30048",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30048",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "8x MI355X",
   "id": "vllm-60216-mi355x8-c4-8k-off",
   "model": "GLM-5.3",
   "notes": "GLM-5.3 MXFP4, fp8 KV, 8x MI355X TP8, vLLM bench serve random dataset (random-range-ratio 0.8), 8192 in / out, c=4; stock path (flag off); out tok/s/GPU; config=c4-8k-off",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "8x MI355X, GLM-5.3 MXFP4, TP8, fp8 KV, serving benchmark:; 8192/1024 4 13.10 ms 7.74 ms -40.9% 36.67 61.02 +66.4%",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "vLLM PR #60216",
   "source_url": "https://github.com/vllm-project/vllm/pull/60216",
   "tg_tokens": null,
   "tps": 36.67,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "8x MI355X",
   "id": "vllm-60216-mi355x8-c4-8k-on",
   "model": "GLM-5.3",
   "notes": "GLM-5.3 MXFP4, fp8 KV, 8x MI355X TP8, vLLM bench serve random dataset (random-range-ratio 0.8), 8192 in / out, c=4; fused AITER MonoKernel path (VLLM_ROCM_GLM_MONO_DECODE=1); out tok/s/GPU; config=c4-8k-on",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "8x MI355X, GLM-5.3 MXFP4, TP8, fp8 KV, serving benchmark:; 8192/1024 4 13.10 ms 7.74 ms -40.9% 36.67 61.02 +66.4%",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "vLLM PR #60216",
   "source_url": "https://github.com/vllm-project/vllm/pull/60216",
   "tg_tokens": null,
   "tps": 61.02,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "8x MI355X",
   "id": "vllm-60216-mi355x8-c8-8k-off",
   "model": "GLM-5.3",
   "notes": "GLM-5.3 MXFP4, fp8 KV, 8x MI355X TP8, vLLM bench serve random dataset (random-range-ratio 0.8), 8192 in / out, c=8; stock path (flag off); out tok/s/GPU; config=c8-8k-off",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "8x MI355X, GLM-5.3 MXFP4, TP8, fp8 KV, serving benchmark:; 8192/1024 8 15.05 ms 10.92 ms -27.4% 62.31 85.72 +37.6%",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "vLLM PR #60216",
   "source_url": "https://github.com/vllm-project/vllm/pull/60216",
   "tg_tokens": null,
   "tps": 62.31,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "8x MI355X",
   "id": "vllm-60216-mi355x8-c8-8k-on",
   "model": "GLM-5.3",
   "notes": "GLM-5.3 MXFP4, fp8 KV, 8x MI355X TP8, vLLM bench serve random dataset (random-range-ratio 0.8), 8192 in / out, c=8; fused AITER MonoKernel path (VLLM_ROCM_GLM_MONO_DECODE=1); out tok/s/GPU; config=c8-8k-on",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "8x MI355X, GLM-5.3 MXFP4, TP8, fp8 KV, serving benchmark:; 8192/1024 8 15.05 ms 10.92 ms -27.4% 62.31 85.72 +37.6%",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "vLLM PR #60216",
   "source_url": "https://github.com/vllm-project/vllm/pull/60216",
   "tg_tokens": null,
   "tps": 85.72,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "8x MI355X",
   "id": "vllm-60216-mi355x8-c16-8k-off",
   "model": "GLM-5.3",
   "notes": "GLM-5.3 MXFP4, fp8 KV, 8x MI355X TP8, vLLM bench serve random dataset (random-range-ratio 0.8), 8192 in / out, c=16; stock path (flag off); out tok/s/GPU (fused path inactive at c16: decode steps above 8 tokens keep regular layers); config=c16-8k-off",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "8x MI355X, GLM-5.3 MXFP4, TP8, fp8 KV, serving benchmark:; 8192/1024 16 (path off) 17.89 ms 17.84 ms -0.3% 105.16 106.11 +0.9%",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "vLLM PR #60216",
   "source_url": "https://github.com/vllm-project/vllm/pull/60216",
   "tg_tokens": null,
   "tps": 105.16,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "8x MI355X",
   "id": "vllm-60216-mi355x8-c16-8k-on",
   "model": "GLM-5.3",
   "notes": "GLM-5.3 MXFP4, fp8 KV, 8x MI355X TP8, vLLM bench serve random dataset (random-range-ratio 0.8), 8192 in / out, c=16; fused AITER MonoKernel path (VLLM_ROCM_GLM_MONO_DECODE=1); out tok/s/GPU (fused path inactive at c16: decode steps above 8 tokens keep regular layers); config=c16-8k-on",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "8x MI355X, GLM-5.3 MXFP4, TP8, fp8 KV, serving benchmark:; 8192/1024 16 (path off) 17.89 ms 17.84 ms -0.3% 105.16 106.11 +0.9%",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "vLLM PR #60216",
   "source_url": "https://github.com/vllm-project/vllm/pull/60216",
   "tg_tokens": null,
   "tps": 106.11,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 60000,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "8x MI355X",
   "id": "vllm-60216-mi355x8-c4-60k-off",
   "model": "GLM-5.3",
   "notes": "GLM-5.3 MXFP4, fp8 KV, 8x MI355X TP8, vLLM bench serve random dataset (random-range-ratio 0.8), 60000 in / out, c=4; stock path (flag off); out tok/s/GPU; config=c4-60k-off",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "8x MI355X, GLM-5.3 MXFP4, TP8, fp8 KV, serving benchmark:; 60000/600 4 24.59 ms 20.07 ms -18.4% 16.62 19.62 +18.1%",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "vLLM PR #60216",
   "source_url": "https://github.com/vllm-project/vllm/pull/60216",
   "tg_tokens": null,
   "tps": 16.62,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 60000,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "8x MI355X",
   "id": "vllm-60216-mi355x8-c4-60k-on",
   "model": "GLM-5.3",
   "notes": "GLM-5.3 MXFP4, fp8 KV, 8x MI355X TP8, vLLM bench serve random dataset (random-range-ratio 0.8), 60000 in / out, c=4; fused AITER MonoKernel path (VLLM_ROCM_GLM_MONO_DECODE=1); out tok/s/GPU; config=c4-60k-on",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "8x MI355X, GLM-5.3 MXFP4, TP8, fp8 KV, serving benchmark:; 60000/600 4 24.59 ms 20.07 ms -18.4% 16.62 19.62 +18.1%",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "vLLM PR #60216",
   "source_url": "https://github.com/vllm-project/vllm/pull/60216",
   "tg_tokens": null,
   "tps": 19.62,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 60000,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "8x MI355X",
   "id": "vllm-60216-mi355x8-c8-60k-off",
   "model": "GLM-5.3",
   "notes": "GLM-5.3 MXFP4, fp8 KV, 8x MI355X TP8, vLLM bench serve random dataset (random-range-ratio 0.8), 60000 in / out, c=8; stock path (flag off); out tok/s/GPU; config=c8-60k-off",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "8x MI355X, GLM-5.3 MXFP4, TP8, fp8 KV, serving benchmark:; 60000/600 8 36.41 ms 33.46 ms -8.1% 23.13 24.88 +7.6%",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "vLLM PR #60216",
   "source_url": "https://github.com/vllm-project/vllm/pull/60216",
   "tg_tokens": null,
   "tps": 23.13,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 60000,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "8x MI355X",
   "id": "vllm-60216-mi355x8-c8-60k-on",
   "model": "GLM-5.3",
   "notes": "GLM-5.3 MXFP4, fp8 KV, 8x MI355X TP8, vLLM bench serve random dataset (random-range-ratio 0.8), 60000 in / out, c=8; fused AITER MonoKernel path (VLLM_ROCM_GLM_MONO_DECODE=1); out tok/s/GPU; config=c8-60k-on",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "8x MI355X, GLM-5.3 MXFP4, TP8, fp8 KV, serving benchmark:; 60000/600 8 36.41 ms 33.46 ms -8.1% 23.13 24.88 +7.6%",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "vLLM PR #60216",
   "source_url": "https://github.com/vllm-project/vllm/pull/60216",
   "tg_tokens": null,
   "tps": 24.88,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": 60000,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "8x MI355X",
   "id": "vllm-60216-mi355x8-c16-60k-off",
   "model": "GLM-5.3",
   "notes": "GLM-5.3 MXFP4, fp8 KV, 8x MI355X TP8, vLLM bench serve random dataset (random-range-ratio 0.8), 60000 in / out, c=16; stock path (flag off); out tok/s/GPU (fused path inactive at c16: decode steps above 8 tokens keep regular layers); config=c16-60k-off",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "8x MI355X, GLM-5.3 MXFP4, TP8, fp8 KV, serving benchmark:; 60000/600 16 (path off) 62.53 ms 62.51 ms 0.0% 27.22 27.34 +0.4%",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "vLLM PR #60216",
   "source_url": "https://github.com/vllm-project/vllm/pull/60216",
   "tg_tokens": null,
   "tps": 27.22,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": 60000,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "8x MI355X",
   "id": "vllm-60216-mi355x8-c16-60k-on",
   "model": "GLM-5.3",
   "notes": "GLM-5.3 MXFP4, fp8 KV, 8x MI355X TP8, vLLM bench serve random dataset (random-range-ratio 0.8), 60000 in / out, c=16; fused AITER MonoKernel path (VLLM_ROCM_GLM_MONO_DECODE=1); out tok/s/GPU (fused path inactive at c16: decode steps above 8 tokens keep regular layers); config=c16-60k-on",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "8x MI355X, GLM-5.3 MXFP4, TP8, fp8 KV, serving benchmark:; 60000/600 16 (path off) 62.53 ms 62.51 ms 0.0% 27.22 27.34 +0.4%",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "vLLM PR #60216",
   "source_url": "https://github.com/vllm-project/vllm/pull/60216",
   "tg_tokens": null,
   "tps": 27.34,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 64,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "8xH100",
   "id": "vllm-60220-h100-c64-8k-mnbt32k",
   "model": "GLM-5.3-Flash",
   "notes": "GLM-5.3-Flash FP8, 8xH100 TP8, max_num_batched_tokens=32768 (previously OOMed at warmup), vllm bench serve random dataset (--ignore-eos), 8192 in / 1024 out, c64; total tok/s; config=c64-8k-mnbt32k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "FP8",
   "quote": "GLM-5.3-Flash FP8 on 8xH100 (TP8, max_num_batched_tokens=32768; 8192/1024, c64 192/192 16717 16424",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "vLLM PR #60220",
   "source_url": "https://github.com/vllm-project/vllm/pull/60220",
   "tg_tokens": null,
   "tps": 16717,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 256,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "8xH100",
   "id": "vllm-60220-h100-c256-8k-mnbt32k",
   "model": "GLM-5.3-Flash",
   "notes": "GLM-5.3-Flash FP8, 8xH100 TP8, max_num_batched_tokens=32768 (previously OOMed at warmup), vllm bench serve random dataset (--ignore-eos), 8192 in / 1024 out, c256; total tok/s; config=c256-8k-mnbt32k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "FP8",
   "quote": "GLM-5.3-Flash FP8 on 8xH100 (TP8, max_num_batched_tokens=32768; 8192/1024, c256 768/768 20189 21522 (KV cache reached 100% at 32K)",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "vLLM PR #60220",
   "source_url": "https://github.com/vllm-project/vllm/pull/60220",
   "tg_tokens": null,
   "tps": 20189,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "8xH100",
   "id": "vllm-60220-h100-c16-64k-mnbt32k",
   "model": "GLM-5.3-Flash",
   "notes": "GLM-5.3-Flash FP8, 8xH100 TP8, max_num_batched_tokens=32768 (previously OOMed at warmup), vllm bench serve random dataset (--ignore-eos), 65536 in / 1024 out, c16; total tok/s; config=c16-64k-mnbt32k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "FP8",
   "quote": "GLM-5.3-Flash FP8 on 8xH100 (TP8, max_num_batched_tokens=32768; 65536/1024, c16 48/48 23357 23141",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "vLLM PR #60220",
   "source_url": "https://github.com/vllm-project/vllm/pull/60220",
   "tg_tokens": null,
   "tps": 23357,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 64,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "8xH100",
   "id": "vllm-60220-h100-c64-64k-mnbt32k",
   "model": "GLM-5.3-Flash",
   "notes": "GLM-5.3-Flash FP8, 8xH100 TP8, max_num_batched_tokens=32768 (previously OOMed at warmup), vllm bench serve random dataset (--ignore-eos), 65536 in / 1024 out, c64; total tok/s; config=c64-64k-mnbt32k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "FP8",
   "quote": "GLM-5.3-Flash FP8 on 8xH100 (TP8, max_num_batched_tokens=32768; 65536/1024, c64 192/192 24263 25101",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "vLLM PR #60220",
   "source_url": "https://github.com/vllm-project/vllm/pull/60220",
   "tg_tokens": null,
   "tps": 24263,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 64,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "8xH100",
   "id": "vllm-60220-h100-c64-8k-mnbt16k",
   "model": "GLM-5.3-Flash",
   "notes": "GLM-5.3-Flash FP8, 8xH100 TP8, max_num_batched_tokens=16384 (existing baseline config, larger KV pool 2.66M vs 1.88M tokens), vllm bench serve random dataset (--ignore-eos), 8192 in / 1024 out, c64; total tok/s; config=c64-8k-mnbt16k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "FP8",
   "quote": "GLM-5.3-Flash FP8 on 8xH100 (TP8, max_num_batched_tokens=32768; 8192/1024, c64 192/192 16717 16424",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "vLLM PR #60220",
   "source_url": "https://github.com/vllm-project/vllm/pull/60220",
   "tg_tokens": null,
   "tps": 16424,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 256,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "8xH100",
   "id": "vllm-60220-h100-c256-8k-mnbt16k",
   "model": "GLM-5.3-Flash",
   "notes": "GLM-5.3-Flash FP8, 8xH100 TP8, max_num_batched_tokens=16384 (existing baseline config, larger KV pool 2.66M vs 1.88M tokens), vllm bench serve random dataset (--ignore-eos), 8192 in / 1024 out, c256; total tok/s; config=c256-8k-mnbt16k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "FP8",
   "quote": "GLM-5.3-Flash FP8 on 8xH100 (TP8, max_num_batched_tokens=32768; 8192/1024, c256 768/768 20189 21522 (KV cache reached 100% at 32K)",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "vLLM PR #60220",
   "source_url": "https://github.com/vllm-project/vllm/pull/60220",
   "tg_tokens": null,
   "tps": 21522,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "8xH100",
   "id": "vllm-60220-h100-c16-64k-mnbt16k",
   "model": "GLM-5.3-Flash",
   "notes": "GLM-5.3-Flash FP8, 8xH100 TP8, max_num_batched_tokens=16384 (existing baseline config, larger KV pool 2.66M vs 1.88M tokens), vllm bench serve random dataset (--ignore-eos), 65536 in / 1024 out, c16; total tok/s; config=c16-64k-mnbt16k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "FP8",
   "quote": "GLM-5.3-Flash FP8 on 8xH100 (TP8, max_num_batched_tokens=32768; 65536/1024, c16 48/48 23357 23141",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "vLLM PR #60220",
   "source_url": "https://github.com/vllm-project/vllm/pull/60220",
   "tg_tokens": null,
   "tps": 23141,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 64,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "8xH100",
   "id": "vllm-60220-h100-c64-64k-mnbt16k",
   "model": "GLM-5.3-Flash",
   "notes": "GLM-5.3-Flash FP8, 8xH100 TP8, max_num_batched_tokens=16384 (existing baseline config, larger KV pool 2.66M vs 1.88M tokens), vllm bench serve random dataset (--ignore-eos), 65536 in / 1024 out, c64; total tok/s; config=c64-64k-mnbt16k",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "FP8",
   "quote": "GLM-5.3-Flash FP8 on 8xH100 (TP8, max_num_batched_tokens=32768; 65536/1024, c64 192/192 24263 25101",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "vLLM PR #60220",
   "source_url": "https://github.com/vllm-project/vllm/pull/60220",
   "tg_tokens": null,
   "tps": 25101,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "GB10",
   "id": "vllm-60226-gb10-stock-c1",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Context-aware sparse LM head on a quantized NVFP4 lm_head, MTP k=3, SPEED-Bench 11 categories, greedy, 256 output tokens, FP8 KV; stock NVFP4 head (baseline), c=1; config=stock-c1",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "GB10; Qwen3.6-35B-A3B; stock 2.935 3.060 118.1 tok/s 72.9 tok/s 97.0%",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "vLLM PR #60226",
   "source_url": "https://github.com/vllm-project/vllm/pull/60226",
   "tg_tokens": 256,
   "tps": 118.1,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "GB10",
   "id": "vllm-60226-gb10-stock-c4",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Context-aware sparse LM head on a quantized NVFP4 lm_head, MTP k=3, SPEED-Bench 11 categories, greedy, 256 output tokens, FP8 KV; stock NVFP4 head (baseline), c=4; config=stock-c4",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "GB10; Qwen3.6-35B-A3B; stock 2.935 3.060 118.1 tok/s 72.9 tok/s 97.0%",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "vLLM PR #60226",
   "source_url": "https://github.com/vllm-project/vllm/pull/60226",
   "tg_tokens": 256,
   "tps": 72.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Google Axion, Neoverse V2 with SVE2",
   "id": "lc30051-axion-fa1-t8-pp63",
   "model": "Gemma 4 E4B",
   "notes": "Google Axion (Neoverse V2 with SVE2) on GCP c4a-standard-16, llama.cpp b11443, Gemma 4 E4B Q4_0, llama-bench pp-only (-p 63 -n 0); FA-on cliff: from 64 queries the tiled flash-attention path runs the scalar simd_gemm fallback because SVE builds are excluded from the generic kernel; threads=8 fa=1 config=fa1-t8-pp63",
   "params": null,
   "power_w": null,
   "pp_tokens": 63,
   "pp_tps": 126.69,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Google Axion, Neoverse V2 with SVE2; gemma4 E4B Q4_0 | 4.26 GiB | 7.46 B | CPU | 8 | 1 | pp63 | 126.69 \u00b1 0.26",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp issue #30051",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30051",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Google Axion, Neoverse V2 with SVE2",
   "id": "lc30051-axion-fa1-t8-pp64",
   "model": "Gemma 4 E4B",
   "notes": "Google Axion (Neoverse V2 with SVE2) on GCP c4a-standard-16, llama.cpp b11443, Gemma 4 E4B Q4_0, llama-bench pp-only (-p 64 -n 0); FA-on cliff: from 64 queries the tiled flash-attention path runs the scalar simd_gemm fallback because SVE builds are excluded from the generic kernel; threads=8 fa=1 config=fa1-t8-pp64",
   "params": null,
   "power_w": null,
   "pp_tokens": 64,
   "pp_tps": 91.37,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Google Axion, Neoverse V2 with SVE2; gemma4 E4B Q4_0 | 4.26 GiB | 7.46 B | CPU | 8 | 1 | pp64 | 91.37 \u00b1 0.11",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp issue #30051",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30051",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Google Axion, Neoverse V2 with SVE2",
   "id": "lc30051-axion-fa1-t16-pp63",
   "model": "Gemma 4 E4B",
   "notes": "Google Axion (Neoverse V2 with SVE2) on GCP c4a-standard-16, llama.cpp b11443, Gemma 4 E4B Q4_0, llama-bench pp-only (-p 63 -n 0); FA-on cliff: from 64 queries the tiled flash-attention path runs the scalar simd_gemm fallback because SVE builds are excluded from the generic kernel; threads=16 fa=1 config=fa1-t16-pp63",
   "params": null,
   "power_w": null,
   "pp_tokens": 63,
   "pp_tps": 247.07,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Google Axion, Neoverse V2 with SVE2; gemma4 E4B Q4_0 | 4.26 GiB | 7.46 B | CPU | 16 | 1 | pp63 | 247.07 \u00b1 0.47",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp issue #30051",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30051",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Google Axion, Neoverse V2 with SVE2",
   "id": "lc30051-axion-fa1-t16-pp64",
   "model": "Gemma 4 E4B",
   "notes": "Google Axion (Neoverse V2 with SVE2) on GCP c4a-standard-16, llama.cpp b11443, Gemma 4 E4B Q4_0, llama-bench pp-only (-p 64 -n 0); FA-on cliff: from 64 queries the tiled flash-attention path runs the scalar simd_gemm fallback because SVE builds are excluded from the generic kernel; threads=16 fa=1 config=fa1-t16-pp64",
   "params": null,
   "power_w": null,
   "pp_tokens": 64,
   "pp_tps": 135.1,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Google Axion, Neoverse V2 with SVE2; gemma4 E4B Q4_0 | 4.26 GiB | 7.46 B | CPU | 16 | 1 | pp64 | 135.10 \u00b1 0.12",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp issue #30051",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30051",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Google Axion, Neoverse V2 with SVE2",
   "id": "lc30051-axion-fa0-t8-pp63",
   "model": "Gemma 4 E4B",
   "notes": "Google Axion (Neoverse V2 with SVE2) on GCP c4a-standard-16, llama.cpp b11443, Gemma 4 E4B Q4_0, llama-bench pp-only (-p 63 -n 0); FA-on cliff: from 64 queries the tiled flash-attention path runs the scalar simd_gemm fallback because SVE builds are excluded from the generic kernel; threads=8 fa=0 config=fa0-t8-pp63",
   "params": null,
   "power_w": null,
   "pp_tokens": 63,
   "pp_tps": 121.03,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Google Axion, Neoverse V2 with SVE2; gemma4 E4B Q4_0 | 4.26 GiB | 7.46 B | CPU | 8 | 0 | pp63 | 121.03 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp issue #30051",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30051",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Google Axion, Neoverse V2 with SVE2",
   "id": "lc30051-axion-fa0-t8-pp64",
   "model": "Gemma 4 E4B",
   "notes": "Google Axion (Neoverse V2 with SVE2) on GCP c4a-standard-16, llama.cpp b11443, Gemma 4 E4B Q4_0, llama-bench pp-only (-p 64 -n 0); FA-on cliff: from 64 queries the tiled flash-attention path runs the scalar simd_gemm fallback because SVE builds are excluded from the generic kernel; threads=8 fa=0 config=fa0-t8-pp64",
   "params": null,
   "power_w": null,
   "pp_tokens": 64,
   "pp_tps": 129.01,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Google Axion, Neoverse V2 with SVE2; gemma4 E4B Q4_0 | 4.26 GiB | 7.46 B | CPU | 8 | 0 | pp64 | 129.01 \u00b1 0.24",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "llama.cpp issue #30051",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30051",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 64,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-60250-mi355x-qwen3827b-c64-main",
   "model": "Qwen3.8-27B",
   "notes": "Qwen3.8-27B BF16 (unquantized), single MI355X TP1, vllm bench serve random 1024 in / 1024 out, c=64, 640 prompts, greedy (temperature 0), no prefix caching, async scheduling, max-model-len 262144; nightly image 2026-09-22 (vLLM 0961bbae28, torch 2.12, amd-aiter 0.1.22.post1); aggregate output token throughput; median TTFT 1074.2 ms, mean TPOT 22.23 ms; stock path (rms_norm_gated decomposed into two Inductor kernels); config=main",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "BF16",
   "quote": "Run on MI355X against the vllm/vllm-openai-rocm:nightly image of 2026-09-22; vllm serve Qwen/Qwen3.8-27B --tensor-parallel-size 1 --max-model-len 262144; 64 Main 2752.5 \u2014 1074.2 \u2014 22.23 \u2014",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "vLLM PR #60250",
   "source_url": "https://github.com/vllm-project/vllm/pull/60250",
   "tg_tokens": 1024,
   "tps": 2752.5,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 64,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-60250-mi355x-qwen3827b-c64-pr",
   "model": "Qwen3.8-27B",
   "notes": "Qwen3.8-27B BF16 (unquantized), single MI355X TP1, vllm bench serve random 1024 in / 1024 out, c=64, 640 prompts, greedy (temperature 0), no prefix caching, async scheduling, max-model-len 262144; nightly image 2026-09-22 (vLLM 0961bbae28, torch 2.12, amd-aiter 0.1.22.post1); aggregate output token throughput; median TTFT 1084.0 ms, mean TPOT 22.13 ms; rms_norm_gated custom op (vllm.fla_rms_norm_gated) enabled by default on ROCm, one Triton launch replaces two; config=pr",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "BF16",
   "quote": "Run on MI355X against the vllm/vllm-openai-rocm:nightly image of 2026-09-22; vllm serve Qwen/Qwen3.8-27B --tensor-parallel-size 1 --max-model-len 262144; 64 This PR 2762.8 +0.37% 1084.0 +0.91% 22.13 \u22120.45%",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "vLLM PR #60250",
   "source_url": "https://github.com/vllm-project/vllm/pull/60250",
   "tg_tokens": 1024,
   "tps": 2762.8,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-60248-mi355x-kimik3-p90-off",
   "model": "Kimi-K3",
   "notes": "Kimi-K3, single MI355X (gfx950), current ROCm vLLM nightly; p90 interactivity tok/s from the latest same-node/image/day full 60-minute scored comparison; both Kimi-K3 AMD stream overlaps off; weight quantization not stated; config=overlap-off",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "[ROCm][Perf][Kimi-K3] Overlap MoE routing and MLA gate projections; MI355X (gfx950), current ROCm vLLM nightly; overlap off: p90 interactivity 163.94 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "vLLM PR #60248",
   "source_url": "https://github.com/vllm-project/vllm/pull/60248",
   "tg_tokens": null,
   "tps": 163.94,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-60248-mi355x-kimik3-p90-v4",
   "model": "Kimi-K3",
   "notes": "Kimi-K3, single MI355X (gfx950), current ROCm vLLM nightly; p90 interactivity tok/s from the latest same-node/image/day full 60-minute scored comparison; both Kimi-K3 AMD stream overlaps enabled (VLLM_KIMI_K3_AMD_MOE_ROUTER_DOWN_PROJ_STREAM=1, VLLM_KIMI_K3_AMD_MLA_GATE_STREAM=1, decode-gated to at most 128 tokens); gain +8.95 tok/s over overlap off; weight quantization not stated; config=overlap-v4",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "[ROCm][Perf][Kimi-K3] Overlap MoE routing and MLA gate projections; MI355X (gfx950), current ROCm vLLM nightly; overlap v4: p90 interactivity 172.89 tok/s; gain: +8.95 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "vLLM PR #60248",
   "source_url": "https://github.com/vllm-project/vllm/pull/60248",
   "tg_tokens": null,
   "tps": 172.89,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-04",
   "flags": [],
   "hardware": "GB10",
   "id": "vllm-59973-gb10-nvfp4-c1",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Qwen3.8-Flash-Next NVFP4 checkpoint + MTP k=3 drafter, one DGX Spark (GB10), measured on three nodes with SM121 skinny-GEMM plans from #59753; opt-in quantized draft lm_head (draft_lm_head_quantization=nvfp4); NVFP4 draft head, c=1 SPEED-Bench decode 46.5 tok/s (+18% over stock), average AL unchanged (-0.4%); stock BF16 draft head 39.4 tok/s; config=nvfp4-c1",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Decode at c=1 on SPEED-Bench goes from 39.4 to 46.5 tok/s; Qwen3.8-Flash-Next-NVFP4 + MTP (k=3), one DGX Spark; Decode on GB10",
   "ram_gb": null,
   "retrieved": "2026-10-13",
   "source_name": "vLLM PR #59973",
   "source_url": "https://github.com/vllm-project/vllm/pull/59973",
   "tg_tokens": null,
   "tps": 46.5,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "hf-rolandoris-qwen389bdistill-m1max-oq4e",
   "model": "Qwen3.8-9B-Distill",
   "notes": "community distill of Qwen3.8-2.4T-A95B into the Qwen3.5-9B architecture (base empero-ai/Qwen3.8-9B-Distill), oMLX Universal Dynamic Quantization (oQ4e), MLX safetensors, native ctx 262144; generation tok/s measured by the card author on M1 Max 64 GB, model size 4.9 GiB, peak memory measured with a 37-token prompt; config=oQ4e",
   "params": "9B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-9B-Distill; M1 Max; 9B-oQ4e 4.9GB 101.5 43.0 5.471",
   "ram_gb": 64,
   "retrieved": "2026-10-13",
   "source_name": "HF model card: RolanDorisTech/Qwen3.8-9B-Distill-MLX-8bit",
   "source_url": "https://huggingface.co/RolanDorisTech/Qwen3.8-9B-Distill-MLX-8bit",
   "tg_tokens": null,
   "tps": 43.0,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "hf-rolandoris-qwen389bdistill-m1max-oq5e",
   "model": "Qwen3.8-9B-Distill",
   "notes": "community distill of Qwen3.8-2.4T-A95B into the Qwen3.5-9B architecture (base empero-ai/Qwen3.8-9B-Distill), oMLX Universal Dynamic Quantization (oQ5e), MLX safetensors, native ctx 262144; generation tok/s measured by the card author on M1 Max 64 GB, model size 6.0 GiB, peak memory measured with a 37-token prompt; config=oQ5e",
   "params": "9B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-9B-Distill; M1 Max; 9B-oQ5e 6.0GB 91.3 37.1 6.555",
   "ram_gb": 64,
   "retrieved": "2026-10-13",
   "source_name": "HF model card: RolanDorisTech/Qwen3.8-9B-Distill-MLX-8bit",
   "source_url": "https://huggingface.co/RolanDorisTech/Qwen3.8-9B-Distill-MLX-8bit",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "hf-rolandoris-qwen389bdistill-m1max-oq6e",
   "model": "Qwen3.8-9B-Distill",
   "notes": "community distill of Qwen3.8-2.4T-A95B into the Qwen3.5-9B architecture (base empero-ai/Qwen3.8-9B-Distill), oMLX Universal Dynamic Quantization (oQ6e), MLX safetensors, native ctx 262144; generation tok/s measured by the card author on M1 Max 64 GB, model size 7.0 GiB, peak memory measured with a 37-token prompt; config=oQ6e",
   "params": "9B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-9B-Distill; M1 Max; 9B-oQ6e 7.0GB 84.3 33.3 7.664",
   "ram_gb": 64,
   "retrieved": "2026-10-13",
   "source_name": "HF model card: RolanDorisTech/Qwen3.8-9B-Distill-MLX-8bit",
   "source_url": "https://huggingface.co/RolanDorisTech/Qwen3.8-9B-Distill-MLX-8bit",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "hf-rolandoris-qwen389bdistill-m1max-oq8e",
   "model": "Qwen3.8-9B-Distill",
   "notes": "community distill of Qwen3.8-2.4T-A95B into the Qwen3.5-9B architecture (base empero-ai/Qwen3.8-9B-Distill), oMLX Universal Dynamic Quantization (oQ8e), MLX safetensors, native ctx 262144; generation tok/s measured by the card author on M1 Max 64 GB, model size 8.9 GiB, peak memory measured with a 37-token prompt; config=oQ8e",
   "params": "9B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-9B-Distill; M1 Max; 9B-oQ8e 8.9GB 91.8 28.7 9.679",
   "ram_gb": 64,
   "retrieved": "2026-10-13",
   "source_name": "HF model card: RolanDorisTech/Qwen3.8-9B-Distill-MLX-8bit",
   "source_url": "https://huggingface.co/RolanDorisTech/Qwen3.8-9B-Distill-MLX-8bit",
   "tg_tokens": null,
   "tps": 28.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": 1,
   "ctx": 175000,
   "date": "2026-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "lc-30074-175k-none",
   "model": "Qwen3.8-Flash-Next",
   "notes": "decode, llama-server chat endpoint wall-clock, 179K-token book prompt, build b11429, depth-decay measurement; config=none",
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ3_S",
   "quote": "qwen4exp: Metal decode depth-decay on M5 Max \u2014 free-form 26.9 t/s at 175K vs 73.9 t/s with n-gram spec vs 53.9 t/s MLX; Mac, macOS 26.5, Apple M5 Max, 128 GB unified memory; SC117/Qwen3.8-Flash-Next-GSQ-RCO-abliterated IQ3_S (2 shards, 83.8 GB): 512 experts / top-10, 51.2B per_layer_token_embd (IQ4_NL), QSA indexer (top_k = 2048); llama-server -m <IQ3_S>-00001-of-00002.gguf -c 262144 -b 2048 -ub 512 -t 12 -fa on -ngl 99 -lzm auto -ctk f16 -ctv f16; | 175K, summarize (free-form) | none | 26.9 (b11429) | - |; | 175K, quote the last six paragraphs verbatim | ngram-map-k4v | 73.9 (b11451) | 67% |; | 32K, quote verbatim | ngram-map-k4v | 78.5 (b11451) | 68% |; | 32K, quote verbatim | ngram-simple | 89.4 (b11451) | 71% |; mlx-serve (same model family, 5bit quant) holds 53.9 t/s at 252K.",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "llama.cpp issue #30074 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30074",
   "tps": "26.9"
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": 1,
   "ctx": 175000,
   "date": "2026-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "lc-30074-175k-ngramk4v",
   "model": "Qwen3.8-Flash-Next",
   "notes": "decode with ngram-map-k4v spec, 67% draft accept, build b11451; config=ngram-map-k4v",
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ3_S",
   "quote": "qwen4exp: Metal decode depth-decay on M5 Max \u2014 free-form 26.9 t/s at 175K vs 73.9 t/s with n-gram spec vs 53.9 t/s MLX; Mac, macOS 26.5, Apple M5 Max, 128 GB unified memory; SC117/Qwen3.8-Flash-Next-GSQ-RCO-abliterated IQ3_S (2 shards, 83.8 GB): 512 experts / top-10, 51.2B per_layer_token_embd (IQ4_NL), QSA indexer (top_k = 2048); llama-server -m <IQ3_S>-00001-of-00002.gguf -c 262144 -b 2048 -ub 512 -t 12 -fa on -ngl 99 -lzm auto -ctk f16 -ctv f16; | 175K, summarize (free-form) | none | 26.9 (b11429) | - |; | 175K, quote the last six paragraphs verbatim | ngram-map-k4v | 73.9 (b11451) | 67% |; | 32K, quote verbatim | ngram-map-k4v | 78.5 (b11451) | 68% |; | 32K, quote verbatim | ngram-simple | 89.4 (b11451) | 71% |; mlx-serve (same model family, 5bit quant) holds 53.9 t/s at 252K.",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "llama.cpp issue #30074 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30074",
   "tps": "73.9"
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": 1,
   "ctx": 32000,
   "date": "2026-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "lc-30074-32k-ngramk4v",
   "model": "Qwen3.8-Flash-Next",
   "notes": "decode with ngram-map-k4v spec, 68% draft accept, build b11451; config=ngram-map-k4v",
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ3_S",
   "quote": "qwen4exp: Metal decode depth-decay on M5 Max \u2014 free-form 26.9 t/s at 175K vs 73.9 t/s with n-gram spec vs 53.9 t/s MLX; Mac, macOS 26.5, Apple M5 Max, 128 GB unified memory; SC117/Qwen3.8-Flash-Next-GSQ-RCO-abliterated IQ3_S (2 shards, 83.8 GB): 512 experts / top-10, 51.2B per_layer_token_embd (IQ4_NL), QSA indexer (top_k = 2048); llama-server -m <IQ3_S>-00001-of-00002.gguf -c 262144 -b 2048 -ub 512 -t 12 -fa on -ngl 99 -lzm auto -ctk f16 -ctv f16; | 175K, summarize (free-form) | none | 26.9 (b11429) | - |; | 175K, quote the last six paragraphs verbatim | ngram-map-k4v | 73.9 (b11451) | 67% |; | 32K, quote verbatim | ngram-map-k4v | 78.5 (b11451) | 68% |; | 32K, quote verbatim | ngram-simple | 89.4 (b11451) | 71% |; mlx-serve (same model family, 5bit quant) holds 53.9 t/s at 252K.",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "llama.cpp issue #30074 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30074",
   "tps": "78.5"
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": 1,
   "ctx": 32000,
   "date": "2026-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "lc-30074-32k-ngramsimple",
   "model": "Qwen3.8-Flash-Next",
   "notes": "decode with ngram-simple spec, 71% draft accept, build b11451; config=ngram-simple",
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ3_S",
   "quote": "qwen4exp: Metal decode depth-decay on M5 Max \u2014 free-form 26.9 t/s at 175K vs 73.9 t/s with n-gram spec vs 53.9 t/s MLX; Mac, macOS 26.5, Apple M5 Max, 128 GB unified memory; SC117/Qwen3.8-Flash-Next-GSQ-RCO-abliterated IQ3_S (2 shards, 83.8 GB): 512 experts / top-10, 51.2B per_layer_token_embd (IQ4_NL), QSA indexer (top_k = 2048); llama-server -m <IQ3_S>-00001-of-00002.gguf -c 262144 -b 2048 -ub 512 -t 12 -fa on -ngl 99 -lzm auto -ctk f16 -ctv f16; | 175K, summarize (free-form) | none | 26.9 (b11429) | - |; | 175K, quote the last six paragraphs verbatim | ngram-map-k4v | 73.9 (b11451) | 67% |; | 32K, quote verbatim | ngram-map-k4v | 78.5 (b11451) | 68% |; | 32K, quote verbatim | ngram-simple | 89.4 (b11451) | 71% |; mlx-serve (same model family, 5bit quant) holds 53.9 t/s at 252K.",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "llama.cpp issue #30074 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30074",
   "tps": "89.4"
  },
  {
   "backend": "MLX",
   "batch": 1,
   "ctx": 252000,
   "date": "2026-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "lc-30074-mlx-252k",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mlx-serve, same model family 5bit quant, same machine, 252K context; reference point against the llama.cpp depth decay",
   "pp_tps": null,
   "provenance": "community",
   "quant": "5bit",
   "quote": "qwen4exp: Metal decode depth-decay on M5 Max \u2014 free-form 26.9 t/s at 175K vs 73.9 t/s with n-gram spec vs 53.9 t/s MLX; Mac, macOS 26.5, Apple M5 Max, 128 GB unified memory; SC117/Qwen3.8-Flash-Next-GSQ-RCO-abliterated IQ3_S (2 shards, 83.8 GB): 512 experts / top-10, 51.2B per_layer_token_embd (IQ4_NL), QSA indexer (top_k = 2048); llama-server -m <IQ3_S>-00001-of-00002.gguf -c 262144 -b 2048 -ub 512 -t 12 -fa on -ngl 99 -lzm auto -ctk f16 -ctv f16; | 175K, summarize (free-form) | none | 26.9 (b11429) | - |; | 175K, quote the last six paragraphs verbatim | ngram-map-k4v | 73.9 (b11451) | 67% |; | 32K, quote verbatim | ngram-map-k4v | 78.5 (b11451) | 68% |; | 32K, quote verbatim | ngram-simple | 89.4 (b11451) | 71% |; mlx-serve (same model family, 5bit quant) holds 53.9 t/s at 252K.",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "llama.cpp issue #30074 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30074",
   "tps": "53.9"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-30077-q40-shallow-master",
   "model": "Qwen3 8B",
   "notes": "llama-bench tg128, q4_0 K/V cache, CUDA 12.8.1; stack-frame dequant bug present (master); config=q40-master",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Performance (RTX 5090, CUDA 12.8.1, llama-bench -fa 1 -ctk/-ctv q8_0,q4_0 -p 0 -n 128 -d 0,4096,16384,32768); | RTX 5090 | qwen3 8B Q4_K_M | q4_0 | tg128 | 215.63 | 235.00 | 1.09 |",
   "retrieved": "2026-10-07",
   "source_name": "llama.cpp issue #30077 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30077",
   "tg_tokens": 128,
   "tps": "215.63"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-30077-q40-shallow-branch",
   "model": "Qwen3 8B",
   "notes": "llama-bench tg128, q4_0 K/V cache, CUDA 12.8.1; stack-frame fix (branch); config=q40-branch",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Performance (RTX 5090, CUDA 12.8.1, llama-bench -fa 1 -ctk/-ctv q8_0,q4_0 -p 0 -n 128 -d 0,4096,16384,32768); | RTX 5090 | qwen3 8B Q4_K_M | q4_0 | tg128 | 215.63 | 235.00 | 1.09 |",
   "retrieved": "2026-10-07",
   "source_name": "llama.cpp issue #30077 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30077",
   "tg_tokens": 128,
   "tps": "235.00"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-30077-q40-d4096-master",
   "model": "Qwen3 8B",
   "notes": "llama-bench tg128, q4_0 K/V cache, CUDA 12.8.1; stack-frame dequant bug present (master); config=q40-master",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Performance (RTX 5090, CUDA 12.8.1, llama-bench -fa 1 -ctk/-ctv q8_0,q4_0 -p 0 -n 128 -d 0,4096,16384,32768); | RTX 5090 | qwen3 8B Q4_K_M | q4_0 | tg128@d4096 | 164.41 | 212.66 | 1.29 |",
   "retrieved": "2026-10-07",
   "source_name": "llama.cpp issue #30077 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30077",
   "tg_tokens": 128,
   "tps": "164.41"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-30077-q40-d4096-branch",
   "model": "Qwen3 8B",
   "notes": "llama-bench tg128, q4_0 K/V cache, CUDA 12.8.1; stack-frame fix (branch); config=q40-branch",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Performance (RTX 5090, CUDA 12.8.1, llama-bench -fa 1 -ctk/-ctv q8_0,q4_0 -p 0 -n 128 -d 0,4096,16384,32768); | RTX 5090 | qwen3 8B Q4_K_M | q4_0 | tg128@d4096 | 164.41 | 212.66 | 1.29 |",
   "retrieved": "2026-10-07",
   "source_name": "llama.cpp issue #30077 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30077",
   "tg_tokens": 128,
   "tps": "212.66"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-30077-q40-d16384-master",
   "model": "Qwen3 8B",
   "notes": "llama-bench tg128, q4_0 K/V cache, CUDA 12.8.1; stack-frame dequant bug present (master); config=q40-master",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Performance (RTX 5090, CUDA 12.8.1, llama-bench -fa 1 -ctk/-ctv q8_0,q4_0 -p 0 -n 128 -d 0,4096,16384,32768); | RTX 5090 | qwen3 8B Q4_K_M | q4_0 | tg128@d16384 | 100.54 | 171.85 | 1.71 |",
   "retrieved": "2026-10-07",
   "source_name": "llama.cpp issue #30077 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30077",
   "tg_tokens": 128,
   "tps": "100.54"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-30077-q40-d16384-branch",
   "model": "Qwen3 8B",
   "notes": "llama-bench tg128, q4_0 K/V cache, CUDA 12.8.1; stack-frame fix (branch); config=q40-branch",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Performance (RTX 5090, CUDA 12.8.1, llama-bench -fa 1 -ctk/-ctv q8_0,q4_0 -p 0 -n 128 -d 0,4096,16384,32768); | RTX 5090 | qwen3 8B Q4_K_M | q4_0 | tg128@d16384 | 100.54 | 171.85 | 1.71 |",
   "retrieved": "2026-10-07",
   "source_name": "llama.cpp issue #30077 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30077",
   "tg_tokens": 128,
   "tps": "171.85"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-30077-q40-d32768-master",
   "model": "Qwen3 8B",
   "notes": "llama-bench tg128, q4_0 K/V cache, CUDA 12.8.1; stack-frame dequant bug present (master); config=q40-master",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Performance (RTX 5090, CUDA 12.8.1, llama-bench -fa 1 -ctk/-ctv q8_0,q4_0 -p 0 -n 128 -d 0,4096,16384,32768); | RTX 5090 | qwen3 8B Q4_K_M | q4_0 | tg128@d32768 | 65.07 | 134.39 | 2.07 |",
   "retrieved": "2026-10-07",
   "source_name": "llama.cpp issue #30077 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30077",
   "tg_tokens": 128,
   "tps": "65.07"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-30077-q40-d32768-branch",
   "model": "Qwen3 8B",
   "notes": "llama-bench tg128, q4_0 K/V cache, CUDA 12.8.1; stack-frame fix (branch); config=q40-branch",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Performance (RTX 5090, CUDA 12.8.1, llama-bench -fa 1 -ctk/-ctv q8_0,q4_0 -p 0 -n 128 -d 0,4096,16384,32768); | RTX 5090 | qwen3 8B Q4_K_M | q4_0 | tg128@d32768 | 65.07 | 134.39 | 2.07 |",
   "retrieved": "2026-10-07",
   "source_name": "llama.cpp issue #30077 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30077",
   "tg_tokens": 128,
   "tps": "134.39"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-30077-q80-shallow-master",
   "model": "Qwen3 8B",
   "notes": "llama-bench tg128, q8_0 K/V cache, CUDA 12.8.1; stack-frame dequant bug present (master); config=q80-master",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Performance (RTX 5090, CUDA 12.8.1, llama-bench -fa 1 -ctk/-ctv q8_0,q4_0 -p 0 -n 128 -d 0,4096,16384,32768); | RTX 5090 | qwen3 8B Q4_K_M | q8_0 | tg128 | 234.31 | 234.35 | 1.00 |",
   "retrieved": "2026-10-07",
   "source_name": "llama.cpp issue #30077 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30077",
   "tg_tokens": 128,
   "tps": "234.31"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-30077-q80-shallow-branch",
   "model": "Qwen3 8B",
   "notes": "llama-bench tg128, q8_0 K/V cache, CUDA 12.8.1; stack-frame fix (branch); config=q80-branch",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Performance (RTX 5090, CUDA 12.8.1, llama-bench -fa 1 -ctk/-ctv q8_0,q4_0 -p 0 -n 128 -d 0,4096,16384,32768); | RTX 5090 | qwen3 8B Q4_K_M | q8_0 | tg128 | 234.31 | 234.35 | 1.00 |",
   "retrieved": "2026-10-07",
   "source_name": "llama.cpp issue #30077 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30077",
   "tg_tokens": 128,
   "tps": "234.35"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-30077-q80-d4096-master",
   "model": "Qwen3 8B",
   "notes": "llama-bench tg128, q8_0 K/V cache, CUDA 12.8.1; stack-frame dequant bug present (master); config=q80-master",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Performance (RTX 5090, CUDA 12.8.1, llama-bench -fa 1 -ctk/-ctv q8_0,q4_0 -p 0 -n 128 -d 0,4096,16384,32768); | RTX 5090 | qwen3 8B Q4_K_M | q8_0 | tg128@d4096 | 211.85 | 211.93 | 1.00 |",
   "retrieved": "2026-10-07",
   "source_name": "llama.cpp issue #30077 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30077",
   "tg_tokens": 128,
   "tps": "211.85"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-30077-q80-d4096-branch",
   "model": "Qwen3 8B",
   "notes": "llama-bench tg128, q8_0 K/V cache, CUDA 12.8.1; stack-frame fix (branch); config=q80-branch",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Performance (RTX 5090, CUDA 12.8.1, llama-bench -fa 1 -ctk/-ctv q8_0,q4_0 -p 0 -n 128 -d 0,4096,16384,32768); | RTX 5090 | qwen3 8B Q4_K_M | q8_0 | tg128@d4096 | 211.85 | 211.93 | 1.00 |",
   "retrieved": "2026-10-07",
   "source_name": "llama.cpp issue #30077 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30077",
   "tg_tokens": 128,
   "tps": "211.93"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-30077-q80-d16384-master",
   "model": "Qwen3 8B",
   "notes": "llama-bench tg128, q8_0 K/V cache, CUDA 12.8.1; stack-frame dequant bug present (master); config=q80-master",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Performance (RTX 5090, CUDA 12.8.1, llama-bench -fa 1 -ctk/-ctv q8_0,q4_0 -p 0 -n 128 -d 0,4096,16384,32768); | RTX 5090 | qwen3 8B Q4_K_M | q8_0 | tg128@d16384 | 169.06 | 168.94 | 1.00 |",
   "retrieved": "2026-10-07",
   "source_name": "llama.cpp issue #30077 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30077",
   "tg_tokens": 128,
   "tps": "169.06"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-30077-q80-d16384-branch",
   "model": "Qwen3 8B",
   "notes": "llama-bench tg128, q8_0 K/V cache, CUDA 12.8.1; stack-frame fix (branch); config=q80-branch",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Performance (RTX 5090, CUDA 12.8.1, llama-bench -fa 1 -ctk/-ctv q8_0,q4_0 -p 0 -n 128 -d 0,4096,16384,32768); | RTX 5090 | qwen3 8B Q4_K_M | q8_0 | tg128@d16384 | 169.06 | 168.94 | 1.00 |",
   "retrieved": "2026-10-07",
   "source_name": "llama.cpp issue #30077 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30077",
   "tg_tokens": 128,
   "tps": "168.94"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-30077-q80-d32768-master",
   "model": "Qwen3 8B",
   "notes": "llama-bench tg128, q8_0 K/V cache, CUDA 12.8.1; stack-frame dequant bug present (master); config=q80-master",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Performance (RTX 5090, CUDA 12.8.1, llama-bench -fa 1 -ctk/-ctv q8_0,q4_0 -p 0 -n 128 -d 0,4096,16384,32768); | RTX 5090 | qwen3 8B Q4_K_M | q8_0 | tg128@d32768 | 131.24 | 131.09 | 1.00 |",
   "retrieved": "2026-10-07",
   "source_name": "llama.cpp issue #30077 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30077",
   "tg_tokens": 128,
   "tps": "131.24"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-30077-q80-d32768-branch",
   "model": "Qwen3 8B",
   "notes": "llama-bench tg128, q8_0 K/V cache, CUDA 12.8.1; stack-frame fix (branch); config=q80-branch",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Performance (RTX 5090, CUDA 12.8.1, llama-bench -fa 1 -ctk/-ctv q8_0,q4_0 -p 0 -n 128 -d 0,4096,16384,32768); | RTX 5090 | qwen3 8B Q4_K_M | q8_0 | tg128@d32768 | 131.24 | 131.09 | 1.00 |",
   "retrieved": "2026-10-07",
   "source_name": "llama.cpp issue #30077 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30077",
   "tg_tokens": 128,
   "tps": "131.09"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "2\u00d7 Tesla P100",
   "id": "lc-30075-plain",
   "model": "Qwen3.6-35B-A3B",
   "notes": "decode wall-clock, --split-mode tensor, first 1-2 requests before output degeneration; approximate figure from the report; config=split-tensor",
   "pp_tps": null,
   "provenance": "community",
   "quant": "unknown",
   "quote": "On 2\u00d7 Tesla P100 (SM 60) with --split-mode tensor, the first 1\u20132 requests work normally (~130 t/s decode on Qwen3.6-35B-A3B), then the model degenerates; With MTP (--spec-type draft-mtp) the draft acceptance drops to 0 and throughput falls to ~36 t/s; the same degradation also happens without speculative decoding (plain greedy decode, no draft)",
   "retrieved": "2026-10-07",
   "source_name": "llama.cpp issue #30075 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30075",
   "tps": "130"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "2\u00d7 Tesla P100",
   "id": "lc-30075-mtp",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP draft mode, draft acceptance dropped to 0, degenerate state; approximate figure from the report; config=split-tensor-mtp",
   "pp_tps": null,
   "provenance": "community",
   "quant": "unknown",
   "quote": "On 2\u00d7 Tesla P100 (SM 60) with --split-mode tensor, the first 1\u20132 requests work normally (~130 t/s decode on Qwen3.6-35B-A3B), then the model degenerates; With MTP (--spec-type draft-mtp) the draft acceptance drops to 0 and throughput falls to ~36 t/s; the same degradation also happens without speculative decoding (plain greedy decode, no draft)",
   "retrieved": "2026-10-07",
   "source_name": "llama.cpp issue #30075 (community report)",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30075",
   "tps": "36"
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 2048,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60317-e2b-2k-b1",
   "model": "Gemma 4 E2B",
   "notes": "Triton unified attention head_dim=512/SWA optimization, vLLM engine prefill A/B, main ab905a885d baseline vs PR branch; A100-SXM4-80GB; prefill_2k (in=2048, out=1), bs=1; baseline 30314.2 vs PR 32396.1 tok/s (1.07x); prefill-only (out=1)",
   "params": null,
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": 30314.2,
   "provenance": "community",
   "quant": "BF16",
   "quote": "| gemma-4-E2B-it | prefill_2k | 1 | 67.56 | 63.22 | 1.07x | 30314.2 | 32396.1 | 1.07x |; A100-SXM4-80GB Measured",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #60317",
   "source_url": "https://github.com/vllm-project/vllm/pull/60317",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 2048,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60317-e2b-2k-b4",
   "model": "Gemma 4 E2B",
   "notes": "Triton unified attention head_dim=512/SWA optimization, vLLM engine prefill A/B, main ab905a885d baseline vs PR branch; A100-SXM4-80GB; prefill_2k (in=2048, out=1), bs=4; baseline 38015.2 vs PR 40284.1 tok/s (1.06x); prefill-only (out=1)",
   "params": null,
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": 38015.2,
   "provenance": "community",
   "quant": "BF16",
   "quote": "| gemma-4-E2B-it | prefill_2k | 4 | 215.49 | 203.36 | 1.06x | 38015.2 | 40284.1 | 1.06x |; A100-SXM4-80GB Measured",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #60317",
   "source_url": "https://github.com/vllm-project/vllm/pull/60317",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 2048,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60317-e2b-2k-b8",
   "model": "Gemma 4 E2B",
   "notes": "Triton unified attention head_dim=512/SWA optimization, vLLM engine prefill A/B, main ab905a885d baseline vs PR branch; A100-SXM4-80GB; prefill_2k (in=2048, out=1), bs=8; baseline 40572.3 vs PR 43659.4 tok/s (1.08x); prefill-only (out=1)",
   "params": null,
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": 40572.3,
   "provenance": "community",
   "quant": "BF16",
   "quote": "| gemma-4-E2B-it | prefill_2k | 8 | 403.82 | 375.27 | 1.08x | 40572.3 | 43659.4 | 1.08x |; A100-SXM4-80GB Measured",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #60317",
   "source_url": "https://github.com/vllm-project/vllm/pull/60317",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60317-e2b-4k-b1",
   "model": "Gemma 4 E2B",
   "notes": "Triton unified attention head_dim=512/SWA optimization, vLLM engine prefill A/B, main ab905a885d baseline vs PR branch; A100-SXM4-80GB; long_context_4k (in=4096, out=1), bs=1; baseline 30697.3 vs PR 33435.0 tok/s (1.09x); prefill-only (out=1)",
   "params": null,
   "power_w": null,
   "pp_tokens": 4096,
   "pp_tps": 30697.3,
   "provenance": "community",
   "quant": "BF16",
   "quote": "| gemma-4-E2B-it | long_context_4k | 1 | 133.43 | 122.51 | 1.09x | 30697.3 | 33435.0 | 1.09x |; A100-SXM4-80GB Measured",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #60317",
   "source_url": "https://github.com/vllm-project/vllm/pull/60317",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60317-e2b-4k-b4",
   "model": "Gemma 4 E2B",
   "notes": "Triton unified attention head_dim=512/SWA optimization, vLLM engine prefill A/B, main ab905a885d baseline vs PR branch; A100-SXM4-80GB; long_context_4k (in=4096, out=1), bs=4; baseline 35937.9 vs PR 39979.0 tok/s (1.11x); prefill-only (out=1)",
   "params": null,
   "power_w": null,
   "pp_tokens": 4096,
   "pp_tps": 35937.9,
   "provenance": "community",
   "quant": "BF16",
   "quote": "| gemma-4-E2B-it | long_context_4k | 4 | 455.90 | 409.81 | 1.11x | 35937.9 | 39979.0 | 1.11x |; A100-SXM4-80GB Measured",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #60317",
   "source_url": "https://github.com/vllm-project/vllm/pull/60317",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 2048,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60317-e4b-2k-b1",
   "model": "Gemma 4 E4B",
   "notes": "Triton unified attention head_dim=512/SWA optimization, vLLM engine prefill A/B, main ab905a885d baseline vs PR branch; A100-SXM4-80GB; prefill_2k (in=2048, out=1), bs=1; baseline 19742.0 vs PR 20139.2 tok/s (1.02x); prefill-only (out=1)",
   "params": null,
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": 19742.0,
   "provenance": "community",
   "quant": "BF16",
   "quote": "| gemma-4-E4B-it | prefill_2k | 1 | 103.74 | 101.69 | 1.02x | 19742.0 | 20139.2 | 1.02x |; A100-SXM4-80GB Measured",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #60317",
   "source_url": "https://github.com/vllm-project/vllm/pull/60317",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 2048,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60317-e4b-2k-b4",
   "model": "Gemma 4 E4B",
   "notes": "Triton unified attention head_dim=512/SWA optimization, vLLM engine prefill A/B, main ab905a885d baseline vs PR branch; A100-SXM4-80GB; prefill_2k (in=2048, out=1), bs=4; baseline 22494.2 vs PR 23404.3 tok/s (1.04x); prefill-only (out=1)",
   "params": null,
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": 22494.2,
   "provenance": "community",
   "quant": "BF16",
   "quote": "| gemma-4-E4B-it | prefill_2k | 4 | 364.18 | 350.02 | 1.04x | 22494.2 | 23404.3 | 1.04x |; A100-SXM4-80GB Measured",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #60317",
   "source_url": "https://github.com/vllm-project/vllm/pull/60317",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 2048,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60317-e4b-2k-b8",
   "model": "Gemma 4 E4B",
   "notes": "Triton unified attention head_dim=512/SWA optimization, vLLM engine prefill A/B, main ab905a885d baseline vs PR branch; A100-SXM4-80GB; prefill_2k (in=2048, out=1), bs=8; baseline 23855.8 vs PR 24966.0 tok/s (1.05x); prefill-only (out=1)",
   "params": null,
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": 23855.8,
   "provenance": "community",
   "quant": "BF16",
   "quote": "| gemma-4-E4B-it | prefill_2k | 8 | 686.79 | 656.25 | 1.05x | 23855.8 | 24966.0 | 1.05x |; A100-SXM4-80GB Measured",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #60317",
   "source_url": "https://github.com/vllm-project/vllm/pull/60317",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60317-e4b-4k-b1",
   "model": "Gemma 4 E4B",
   "notes": "Triton unified attention head_dim=512/SWA optimization, vLLM engine prefill A/B, main ab905a885d baseline vs PR branch; A100-SXM4-80GB; long_context_4k (in=4096, out=1), bs=1; baseline 20097.0 vs PR 21227.5 tok/s (1.06x); prefill-only (out=1)",
   "params": null,
   "power_w": null,
   "pp_tokens": 4096,
   "pp_tps": 20097.0,
   "provenance": "community",
   "quant": "BF16",
   "quote": "| gemma-4-E4B-it | long_context_4k | 1 | 203.81 | 192.96 | 1.06x | 20097.0 | 21227.5 | 1.06x |; A100-SXM4-80GB Measured",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #60317",
   "source_url": "https://github.com/vllm-project/vllm/pull/60317",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60317-e4b-4k-b4",
   "model": "Gemma 4 E4B",
   "notes": "Triton unified attention head_dim=512/SWA optimization, vLLM engine prefill A/B, main ab905a885d baseline vs PR branch; A100-SXM4-80GB; long_context_4k (in=4096, out=1), bs=4; baseline 22062.9 vs PR 23574.1 tok/s (1.07x); prefill-only (out=1)",
   "params": null,
   "power_w": null,
   "pp_tokens": 4096,
   "pp_tps": 22062.9,
   "provenance": "community",
   "quant": "BF16",
   "quote": "| gemma-4-E4B-it | long_context_4k | 4 | 742.60 | 695.00 | 1.07x | 22062.9 | 23574.1 | 1.07x |; A100-SXM4-80GB Measured",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #60317",
   "source_url": "https://github.com/vllm-project/vllm/pull/60317",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 2048,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60317-12b-2k-b1",
   "model": "gemma-4-12b-it",
   "notes": "Triton unified attention head_dim=512/SWA optimization, vLLM engine prefill A/B, main ab905a885d baseline vs PR branch; A100-SXM4-80GB; prefill_2k (in=2048, out=1), bs=1; baseline 8268.4 vs PR 8781.2 tok/s (1.06x); prefill-only (out=1)",
   "params": null,
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": 8268.4,
   "provenance": "community",
   "quant": "BF16",
   "quote": "| gemma-4-12B-it | prefill_2k | 1 | 247.69 | 233.23 | 1.06x | 8268.4 | 8781.2 | 1.06x |; A100-SXM4-80GB Measured",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #60317",
   "source_url": "https://github.com/vllm-project/vllm/pull/60317",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 2048,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60317-12b-2k-b4",
   "model": "gemma-4-12b-it",
   "notes": "Triton unified attention head_dim=512/SWA optimization, vLLM engine prefill A/B, main ab905a885d baseline vs PR branch; A100-SXM4-80GB; prefill_2k (in=2048, out=1), bs=4; baseline 8906.6 vs PR 9517.6 tok/s (1.07x); prefill-only (out=1)",
   "params": null,
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": 8906.6,
   "provenance": "community",
   "quant": "BF16",
   "quote": "| gemma-4-12B-it | prefill_2k | 4 | 919.76 | 860.72 | 1.07x | 8906.6 | 9517.6 | 1.07x |; A100-SXM4-80GB Measured",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #60317",
   "source_url": "https://github.com/vllm-project/vllm/pull/60317",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 2048,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60317-12b-2k-b8",
   "model": "gemma-4-12b-it",
   "notes": "Triton unified attention head_dim=512/SWA optimization, vLLM engine prefill A/B, main ab905a885d baseline vs PR branch; A100-SXM4-80GB; prefill_2k (in=2048, out=1), bs=8; baseline 9141.9 vs PR 9761.8 tok/s (1.07x); prefill-only (out=1)",
   "params": null,
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": 9141.9,
   "provenance": "community",
   "quant": "BF16",
   "quote": "| gemma-4-12B-it | prefill_2k | 8 | 1792.19 | 1678.39 | 1.07x | 9141.9 | 9761.8 | 1.07x |; A100-SXM4-80GB Measured",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #60317",
   "source_url": "https://github.com/vllm-project/vllm/pull/60317",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60317-12b-4k-b1",
   "model": "gemma-4-12b-it",
   "notes": "Triton unified attention head_dim=512/SWA optimization, vLLM engine prefill A/B, main ab905a885d baseline vs PR branch; A100-SXM4-80GB; long_context_4k (in=4096, out=1), bs=1; baseline 7900.1 vs PR 8642.6 tok/s (1.09x); prefill-only (out=1)",
   "params": null,
   "power_w": null,
   "pp_tokens": 4096,
   "pp_tps": 7900.1,
   "provenance": "community",
   "quant": "BF16",
   "quote": "| gemma-4-12B-it | long_context_4k | 1 | 518.47 | 473.93 | 1.09x | 7900.1 | 8642.6 | 1.09x |; A100-SXM4-80GB Measured",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #60317",
   "source_url": "https://github.com/vllm-project/vllm/pull/60317",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60317-12b-4k-b4",
   "model": "gemma-4-12b-it",
   "notes": "Triton unified attention head_dim=512/SWA optimization, vLLM engine prefill A/B, main ab905a885d baseline vs PR branch; A100-SXM4-80GB; long_context_4k (in=4096, out=1), bs=4; baseline 8435.7 vs PR 9285.1 tok/s (1.10x); prefill-only (out=1)",
   "params": null,
   "power_w": null,
   "pp_tokens": 4096,
   "pp_tps": 8435.7,
   "provenance": "community",
   "quant": "BF16",
   "quote": "| gemma-4-12B-it | long_context_4k | 4 | 1942.21 | 1764.55 | 1.10x | 8435.7 | 9285.1 | 1.10x |; A100-SXM4-80GB Measured",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #60317",
   "source_url": "https://github.com/vllm-project/vllm/pull/60317",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "vllm-60296-r9700-pp",
   "model": "gemma-4-12b-it",
   "notes": "fp8 query attention for fp8 KV cache, TRITON_ATTN backend, gfx1201, fp8 KV, prefix caching on; prefill 7.7K prompt cache-missing; baseline (before) 1102 tok/s vs PR 4574 tok/s (4.15x)",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1102,
   "provenance": "community",
   "quant": "FP8",
   "quote": "before 1102 tok/s (1102-1103) 33.38 (33.35-33.42); Single Radeon AI PRO R9700 (gfx1201). Model gemma-4-12B-it-FP8-Dynamic",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60296",
   "source_url": "https://github.com/vllm-project/vllm/pull/60296",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "vllm-60296-r9700-tg",
   "model": "gemma-4-12b-it",
   "notes": "fp8 query attention for fp8 KV cache, TRITON_ATTN backend, gfx1201, fp8 KV, prefix caching on; decode; baseline (before) 33.38 t/s vs PR 36.08 t/s (+8.1%)",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "before 1102 tok/s (1102-1103) 33.38 (33.35-33.42); Single Radeon AI PRO R9700 (gfx1201). Model gemma-4-12B-it-FP8-Dynamic",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60296",
   "source_url": "https://github.com/vllm-project/vllm/pull/60296",
   "tg_tokens": null,
   "tps": 33.38,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-60397-mi355x-dsv41flash-tp2c1",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM A/B mono decode fast path (VLLM_ROCM_DSV41_MONO_DECODE=1), DeepSeek-V4.1-Flash on MI355X (gfx950), AgentX e2e a8w4 MoE INT4, p90 interactivity tok/s/user, TP2, stored arm = mono-off baseline, mono-on arm: 429.1 (1.35x)",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "INT4 (a8w4)",
   "quote": "fast path for DeepSeek-V4.1-Flash decode on MI355X (gfx950); | c1 | 318.1 \u2192 429.1 (1.35\u00d7) | 21,573 \u2192 25,039 (1.16\u00d7) | 1.18 \u2192 1.28 |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM #60397",
   "source_url": "https://github.com/vllm-project/vllm/issues/60397",
   "tg_tokens": null,
   "tps": 318.1,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 2,
   "ctx": null,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-60397-mi355x-dsv41flash-tp2c2",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM A/B mono decode fast path (VLLM_ROCM_DSV41_MONO_DECODE=1), DeepSeek-V4.1-Flash on MI355X (gfx950), AgentX e2e a8w4 MoE INT4, p90 interactivity tok/s/user, TP2, stored arm = mono-off baseline, mono-on arm: 381.7 (1.28x)",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "INT4 (a8w4)",
   "quote": "fast path for DeepSeek-V4.1-Flash decode on MI355X (gfx950); | c2 | 299.3 \u2192 381.7 (1.28\u00d7) | 22,735 \u2192 25,695 (1.13\u00d7) | 1.10 \u2192 1.08 |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM #60397",
   "source_url": "https://github.com/vllm-project/vllm/issues/60397",
   "tg_tokens": null,
   "tps": 299.3,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": null,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-60397-mi355x-dsv41flash-tp2c4",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM A/B mono decode fast path (VLLM_ROCM_DSV41_MONO_DECODE=1), DeepSeek-V4.1-Flash on MI355X (gfx950), AgentX e2e a8w4 MoE INT4, p90 interactivity tok/s/user, TP2, stored arm = mono-off baseline, mono-on arm: 332.4 (1.15x)",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "INT4 (a8w4)",
   "quote": "fast path for DeepSeek-V4.1-Flash decode on MI355X (gfx950); | c4 | 288.0 \u2192 332.4 (1.15\u00d7) | 33,043 \u2192 35,885 (1.09\u00d7) | 1.27 \u2192 1.26 |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM #60397",
   "source_url": "https://github.com/vllm-project/vllm/issues/60397",
   "tg_tokens": null,
   "tps": 288.0,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": null,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "MI355X",
   "id": "vllm-60397-mi355x-dsv41flash-tp2c8",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM A/B mono decode fast path (VLLM_ROCM_DSV41_MONO_DECODE=1), DeepSeek-V4.1-Flash on MI355X (gfx950), AgentX e2e a8w4 MoE INT4, p90 interactivity tok/s/user, TP2, stored arm = mono-off baseline, mono-on arm: 245.3 (1.14x)",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "INT4 (a8w4)",
   "quote": "fast path for DeepSeek-V4.1-Flash decode on MI355X (gfx950); | c8 | 214.6 \u2192 245.3 (1.14\u00d7) | 58,294 \u2192 60,501 (1.04\u00d7) | 1.01 \u2192 1.08 |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM #60397",
   "source_url": "https://github.com/vllm-project/vllm/issues/60397",
   "tg_tokens": null,
   "tps": 214.6,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-60397-mi355x-dsv41flash-tp4c1",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM A/B mono decode fast path (VLLM_ROCM_DSV41_MONO_DECODE=1), DeepSeek-V4.1-Flash on MI355X (gfx950), AgentX e2e a8w4 MoE INT4, p90 interactivity tok/s/user, TP4, stored arm = InferenceX reference (mono-off), mono-on arm: 523.7 (1.52x)",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "INT4 (a8w4)",
   "quote": "fast path for DeepSeek-V4.1-Flash decode on MI355X (gfx950); | c1 | 343.7 \u2192 523.7 (1.52\u00d7) | 23,691 \u2192 26,796 (1.13\u00d7) |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM #60397",
   "source_url": "https://github.com/vllm-project/vllm/issues/60397",
   "tg_tokens": null,
   "tps": 343.7,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 2,
   "ctx": null,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-60397-mi355x-dsv41flash-tp4c2",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM A/B mono decode fast path (VLLM_ROCM_DSV41_MONO_DECODE=1), DeepSeek-V4.1-Flash on MI355X (gfx950), AgentX e2e a8w4 MoE INT4, p90 interactivity tok/s/user, TP4, stored arm = InferenceX reference (mono-off), mono-on arm: 471.1 (1.44x)",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "INT4 (a8w4)",
   "quote": "fast path for DeepSeek-V4.1-Flash decode on MI355X (gfx950); | c2 | 327.4 \u2192 471.1 (1.44\u00d7) | 23,579 \u2192 27,217 (1.15\u00d7) |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM #60397",
   "source_url": "https://github.com/vllm-project/vllm/issues/60397",
   "tg_tokens": null,
   "tps": 327.4,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": null,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-60397-mi355x-dsv41flash-tp4c4",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM A/B mono decode fast path (VLLM_ROCM_DSV41_MONO_DECODE=1), DeepSeek-V4.1-Flash on MI355X (gfx950), AgentX e2e a8w4 MoE INT4, p90 interactivity tok/s/user, TP4, stored arm = InferenceX reference (mono-off), mono-on arm: 445.1 (1.45x)",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "INT4 (a8w4)",
   "quote": "fast path for DeepSeek-V4.1-Flash decode on MI355X (gfx950); | c4 | 307.2 \u2192 445.1 (1.45\u00d7) | 34,831 \u2192 36,791 (1.06\u00d7) |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM #60397",
   "source_url": "https://github.com/vllm-project/vllm/issues/60397",
   "tg_tokens": null,
   "tps": 307.2,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": null,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-60397-mi355x-dsv41flash-tp4c8",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM A/B mono decode fast path (VLLM_ROCM_DSV41_MONO_DECODE=1), DeepSeek-V4.1-Flash on MI355X (gfx950), AgentX e2e a8w4 MoE INT4, p90 interactivity tok/s/user, TP4, stored arm = InferenceX reference (mono-off), mono-on arm: 358.6 (1.35x)",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "INT4 (a8w4)",
   "quote": "fast path for DeepSeek-V4.1-Flash decode on MI355X (gfx950); | c8 | 265.8 \u2192 358.6 (1.35\u00d7) | 60,266 \u2192 62,621 (1.04\u00d7) |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM #60397",
   "source_url": "https://github.com/vllm-project/vllm/issues/60397",
   "tg_tokens": null,
   "tps": 265.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 512,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc-30087-rtx4090-qwen3827b-pp512-master",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #30087, GDN half-warp reduce kernel, RTX 4090 FE, Qwen3.8-27B UD-Q4_K_XL, -ctk q8_0 -ctv q4_0, library-swap A/B; config=master-c479922ac",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3075.24,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Qwen3.8-27B (UD-Q4_K_XL), RTX 4090 FE; | pp512 |    3075.24 | 3243.60 |  +5.5% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "llama.cpp #30087",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30087",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 512,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc-30087-rtx4090-qwen3827b-pp512-pr",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #30087, GDN half-warp reduce kernel, RTX 4090 FE, Qwen3.8-27B UD-Q4_K_XL, -ctk q8_0 -ctv q4_0, library-swap A/B; config=pr-30087",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3243.6,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Qwen3.8-27B (UD-Q4_K_XL), RTX 4090 FE; | pp512 |    3075.24 | 3243.60 |  +5.5% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "llama.cpp #30087",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30087",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc-30087-rtx4090-qwen3827b-pp4096-master",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #30087, GDN half-warp reduce kernel, RTX 4090 FE, Qwen3.8-27B UD-Q4_K_XL, -ctk q8_0 -ctv q4_0, library-swap A/B; config=master-c479922ac",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3059.87,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Qwen3.8-27B (UD-Q4_K_XL), RTX 4090 FE; | pp4096 |   3059.87 | 3221.08 |  +5.3% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "llama.cpp #30087",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30087",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc-30087-rtx4090-qwen3827b-pp4096-pr",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #30087, GDN half-warp reduce kernel, RTX 4090 FE, Qwen3.8-27B UD-Q4_K_XL, -ctk q8_0 -ctv q4_0, library-swap A/B; config=pr-30087",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3221.08,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Qwen3.8-27B (UD-Q4_K_XL), RTX 4090 FE; | pp4096 |   3059.87 | 3221.08 |  +5.3% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "llama.cpp #30087",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30087",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 128,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc-30087-rtx4090-qwen3827b-tg128-master",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #30087, GDN half-warp reduce kernel, RTX 4090 FE, Qwen3.8-27B UD-Q4_K_XL, -ctk q8_0 -ctv q4_0, library-swap A/B; config=master-c479922ac",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Qwen3.8-27B (UD-Q4_K_XL), RTX 4090 FE; | tg128 |      47.62 |   47.67 |  +0.1% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "llama.cpp #30087",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30087",
   "tg_tokens": null,
   "tps": 47.62,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 128,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc-30087-rtx4090-qwen3827b-tg128-pr",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #30087, GDN half-warp reduce kernel, RTX 4090 FE, Qwen3.8-27B UD-Q4_K_XL, -ctk q8_0 -ctv q4_0, library-swap A/B; config=pr-30087",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Qwen3.8-27B (UD-Q4_K_XL), RTX 4090 FE; | tg128 |      47.62 |   47.67 |  +0.1% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "llama.cpp #30087",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30087",
   "tg_tokens": null,
   "tps": 47.67,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "2 \u00d7 GB200",
   "id": "vllm-60355-gb200-dsv41flash-nodespark-c1",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #60355, fused CuTe DSL all-reduce TP2 (supports_mhc_all_reduce), 2 \u00d7 GB200, DeepSeek-V4.1-Flash, vllm bench serve ISL 1024/OSL 512 mean of 3 seeds, stored arm = main baseline, PR arm: 213.0 (+15.4%); config=nodespark",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "2 \u00d7 GB200, DeepSeek-V4.1-Flash; | 1 | 5.19 \u2192 4.48 (\u221213.8%) | 184.6 \u2192 213.0 (+15.4%) |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM #60355",
   "source_url": "https://github.com/vllm-project/vllm/issues/60355",
   "tg_tokens": null,
   "tps": 184.6,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 2,
   "ctx": null,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "2 \u00d7 GB200",
   "id": "vllm-60355-gb200-dsv41flash-nodespark-c2",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #60355, fused CuTe DSL all-reduce TP2 (supports_mhc_all_reduce), 2 \u00d7 GB200, DeepSeek-V4.1-Flash, vllm bench serve ISL 1024/OSL 512 mean of 3 seeds, stored arm = main baseline, PR arm: 390.7 (+12.4%); config=nodespark",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "2 \u00d7 GB200, DeepSeek-V4.1-Flash; | 2 | 5.58 \u2192 4.93 (\u221211.7%) | 347.7 \u2192 390.7 (+12.4%) |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM #60355",
   "source_url": "https://github.com/vllm-project/vllm/issues/60355",
   "tg_tokens": null,
   "tps": 347.7,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": null,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "2 \u00d7 GB200",
   "id": "vllm-60355-gb200-dsv41flash-nodespark-c4",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #60355, fused CuTe DSL all-reduce TP2 (supports_mhc_all_reduce), 2 \u00d7 GB200, DeepSeek-V4.1-Flash, vllm bench serve ISL 1024/OSL 512 mean of 3 seeds, stored arm = main baseline, PR arm: 701.5 (+15.5%); config=nodespark",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "2 \u00d7 GB200, DeepSeek-V4.1-Flash; | 4 | 6.10 \u2192 5.25 (\u221213.9%) | 607.2 \u2192 701.5 (+15.5%) |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM #60355",
   "source_url": "https://github.com/vllm-project/vllm/issues/60355",
   "tg_tokens": null,
   "tps": 607.2,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": null,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "2 \u00d7 GB200",
   "id": "vllm-60355-gb200-dsv41flash-nodespark-c8",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #60355, fused CuTe DSL all-reduce TP2 (supports_mhc_all_reduce), 2 \u00d7 GB200, DeepSeek-V4.1-Flash, vllm bench serve ISL 1024/OSL 512 mean of 3 seeds, stored arm = main baseline, PR arm: 1212.9 (+11.1%); config=nodespark",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "2 \u00d7 GB200, DeepSeek-V4.1-Flash; | 8 | 6.86 \u2192 6.08 (\u221211.3%) | 1092.2 \u2192 1212.9 (+11.1%) |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM #60355",
   "source_url": "https://github.com/vllm-project/vllm/issues/60355",
   "tg_tokens": null,
   "tps": 1092.2,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": null,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "2 \u00d7 GB200",
   "id": "vllm-60355-gb200-dsv41flash-nodespark-c16",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #60355, fused CuTe DSL all-reduce TP2 (supports_mhc_all_reduce), 2 \u00d7 GB200, DeepSeek-V4.1-Flash, vllm bench serve ISL 1024/OSL 512 mean of 3 seeds, stored arm = main baseline, PR arm: 1919.2 (+8.6%); config=nodespark",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "2 \u00d7 GB200, DeepSeek-V4.1-Flash; | 16 | 8.48 \u2192 7.64 (\u22129.9%) | 1767.1 \u2192 1919.2 (+8.6%) |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM #60355",
   "source_url": "https://github.com/vllm-project/vllm/issues/60355",
   "tg_tokens": null,
   "tps": 1767.1,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 32,
   "ctx": null,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "2 \u00d7 GB200",
   "id": "vllm-60355-gb200-dsv41flash-nodespark-c32",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #60355, fused CuTe DSL all-reduce TP2 (supports_mhc_all_reduce), 2 \u00d7 GB200, DeepSeek-V4.1-Flash, vllm bench serve ISL 1024/OSL 512 mean of 3 seeds, stored arm = main baseline, PR arm: 2658.8 (\u22124.6%); config=nodespark",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "2 \u00d7 GB200, DeepSeek-V4.1-Flash; | 32 | 10.59 \u2192 10.55 (\u22120.4%) | 2786.3 \u2192 2658.8 (\u22124.6%, one outlier run) |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM #60355",
   "source_url": "https://github.com/vllm-project/vllm/issues/60355",
   "tg_tokens": null,
   "tps": 2786.3,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 64,
   "ctx": null,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "2 \u00d7 GB200",
   "id": "vllm-60355-gb200-dsv41flash-nodespark-c64",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #60355, fused CuTe DSL all-reduce TP2 (supports_mhc_all_reduce), 2 \u00d7 GB200, DeepSeek-V4.1-Flash, vllm bench serve ISL 1024/OSL 512 mean of 3 seeds, stored arm = main baseline, PR arm: 4346.2 (\u22120.4%); config=nodespark",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "2 \u00d7 GB200, DeepSeek-V4.1-Flash; | 64 | 13.44 \u2192 13.72 (+2.1%) | 4364.5 \u2192 4346.2 (\u22120.4%) |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM #60355",
   "source_url": "https://github.com/vllm-project/vllm/issues/60355",
   "tg_tokens": null,
   "tps": 4364.5,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "2 \u00d7 GB200",
   "id": "vllm-60355-gb200-dsv41flash-dspark-k5-c1",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #60355, fused CuTe DSL all-reduce TP2 (supports_mhc_all_reduce), 2 \u00d7 GB200, DeepSeek-V4.1-Flash, vllm bench serve ISL 1024/OSL 512 mean of 3 seeds, stored arm = main baseline, PR arm: 680.6 (+10.4%); config=dspark-k5",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "2 \u00d7 GB200, DeepSeek-V4.1-Flash; | 1 | 7.12 \u2192 6.49 (\u22128.8%) | 616.7 \u2192 680.6 (+10.4%) |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM #60355",
   "source_url": "https://github.com/vllm-project/vllm/issues/60355",
   "tg_tokens": null,
   "tps": 616.7,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 2,
   "ctx": null,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "2 \u00d7 GB200",
   "id": "vllm-60355-gb200-dsv41flash-dspark-k5-c2",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #60355, fused CuTe DSL all-reduce TP2 (supports_mhc_all_reduce), 2 \u00d7 GB200, DeepSeek-V4.1-Flash, vllm bench serve ISL 1024/OSL 512 mean of 3 seeds, stored arm = main baseline, PR arm: 1149.0 (+16.4%); config=dspark-k5",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "2 \u00d7 GB200, DeepSeek-V4.1-Flash; | 2 | 8.37 \u2192 7.27 (\u221213.2%) | 986.9 \u2192 1149.0 (+16.4%) |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM #60355",
   "source_url": "https://github.com/vllm-project/vllm/issues/60355",
   "tg_tokens": null,
   "tps": 986.9,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": null,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "2 \u00d7 GB200",
   "id": "vllm-60355-gb200-dsv41flash-dspark-k5-c4",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #60355, fused CuTe DSL all-reduce TP2 (supports_mhc_all_reduce), 2 \u00d7 GB200, DeepSeek-V4.1-Flash, vllm bench serve ISL 1024/OSL 512 mean of 3 seeds, stored arm = main baseline, PR arm: 1497.0 (+5.2%); config=dspark-k5",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "2 \u00d7 GB200, DeepSeek-V4.1-Flash; | 4 | 9.63 \u2192 8.33 (\u221213.6%) | 1423.3 \u2192 1497.0 (+5.2%) |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM #60355",
   "source_url": "https://github.com/vllm-project/vllm/issues/60355",
   "tg_tokens": null,
   "tps": 1423.3,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": null,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "DGX Spark",
   "id": "vllm-60387-dgxspark-llama318b-fullvocab",
   "model": "Llama 3.1 8B",
   "notes": "vLLM PR #60387, VLLM_SPEC_DRAFT_VOCAB reduced-vocab A/B, spec decode k=3, EAGLE-LLaMA3.1-Instruct-8B drafter, greedy, 10 Korean prompts; stored arm = full vocab (128,256) baseline, reduced-vocab arm: 17.95 (+13.4%); config=full-vocab",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "DGX Spark (GB10); | EAGLE-LLaMA3.1-Instruct-8B / Llama-3.1-8B-Instruct | full (128,256) | 15.83 | 1.55 | \u2014 |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM #60387",
   "source_url": "https://github.com/vllm-project/vllm/issues/60387",
   "tg_tokens": null,
   "tps": 15.83,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": null,
   "ctx": null,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "DGX Spark",
   "id": "vllm-60387-dgxspark-exaone45-33b-fullvocab",
   "model": "EXAONE-4.5-33B",
   "notes": "vLLM PR #60387, VLLM_SPEC_DRAFT_VOCAB reduced-vocab A/B, spec decode k=3, native MTP head (Exaone4_5_MTP), greedy, 10 Korean prompts; stored arm = full vocab (153,600) baseline, reduced-vocab arm: 5.96 (+7.0%); config=full-vocab",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "FP8",
   "quote": "DGX Spark (GB10); | EXAONE-4.5-33B-FP8 (native MTP head, Exaone4_5_MTP) | full (153,600) | 5.57 | 1.07 | \u2014 |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM #60387",
   "source_url": "https://github.com/vllm-project/vllm/issues/60387",
   "tg_tokens": null,
   "tps": 5.57,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": null,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "vllm-60390-rtxpro6000-qwen38flashnext-unpatched-1card-c16",
   "model": "Qwen3.8-Flash-Next",
   "notes": "vLLM PR #60390, shared CPU-offloaded Qwen4Exp PLE table, RTX PRO 6000, Qwen3.8-Flash-Next NVFP4, one card, unpatched reference, vllm bench serve 8K-in/512-out, request-rate inf; config=unpatched-1card",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "two RTX PRO 6000 (96 GB each); RadixArk/Qwen3.8-Flash-Next-NVFP4; | one card, unpatched (reference) | ~150 s | 85 GiB | 16 conc | 395 | 5.1 / 5.3 s | 0 |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM #60390",
   "source_url": "https://github.com/vllm-project/vllm/issues/60390",
   "tg_tokens": null,
   "tps": 395,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 32,
   "ctx": null,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "vllm-60390-rtxpro6000-qwen38flashnext-dp-shared-mem-c32",
   "model": "Qwen3.8-Flash-Next",
   "notes": "vLLM PR #60390, shared CPU-offloaded Qwen4Exp PLE table, RTX PRO 6000, Qwen3.8-Flash-Next NVFP4, two cards, dp_shared_memory DP2, vllm bench serve 8K-in/512-out, request-rate inf; config=dp-shared-mem",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "two RTX PRO 6000 (96 GB each); RadixArk/Qwen3.8-Flash-Next-NVFP4; | `--data-parallel-size 2`, `dp_shared_memory` (this PR) | 195 s | 101 GiB | 32 conc | 718 | 5.2 / 9.9 s | 0 |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM #60390",
   "source_url": "https://github.com/vllm-project/vllm/issues/60390",
   "tg_tokens": null,
   "tps": 718,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 32,
   "ctx": null,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "vllm-60390-rtxpro6000-qwen38flashnext-file-shared-c32",
   "model": "Qwen3.8-Flash-Next",
   "notes": "vLLM PR #60390, shared CPU-offloaded Qwen4Exp PLE table, RTX PRO 6000, Qwen3.8-Flash-Next NVFP4, two single-card processes, shared_host_table_dir, aggregate of two 1-card servers (406+411), 16 conc each at once; config=file-shared",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "two RTX PRO 6000 (96 GB each); RadixArk/Qwen3.8-Flash-Next-NVFP4; | two single-card processes, `shared_host_table_dir` (this PR) | both 150 s | 97 GiB | 16 conc each, at once | 406 + 411 = 817 | 4.9 / 9.4 s and 4.8 / 9.1 s | 0 |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM #60390",
   "source_url": "https://github.com/vllm-project/vllm/issues/60390",
   "tg_tokens": null,
   "tps": 817,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": 512,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "lc-30105-r9700-mistral4-pp512-before",
   "model": "Mistral Small 4",
   "notes": "llama.cpp PR #30105, Vulkan FP32 vector kernel selection (Mistral Small 4 perf fix), Radeon AI PRO R9700, Mistral Small 4 IQ4_XS, f16/f32, fitt 1024/fitc 4096; config=before",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 242.75,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "Mistral Small 4; AMD Radeon AI PRO R9700; | mistral4 ?B IQ4_XS - 4.25 bpw  |  54.13 GiB |   118.97 B | Vulkan     |  -1 |   1 |       1024 |        4096 |           pp512 |        242.75 \u00b1 2.00 |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "llama.cpp #30105",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30105",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": 512,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "lc-30105-r9700-mistral4-pp512-after",
   "model": "Mistral Small 4",
   "notes": "llama.cpp PR #30105, Vulkan FP32 vector kernel selection (Mistral Small 4 perf fix), Radeon AI PRO R9700, Mistral Small 4 IQ4_XS, f16/f32, fitt 1024/fitc 4096; config=after",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 240.0,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "Mistral Small 4; AMD Radeon AI PRO R9700; | mistral4 ?B IQ4_XS - 4.25 bpw  |  54.13 GiB |   118.97 B | Vulkan     |  -1 |   1 |       1024 |        4096 |           pp512 |        240.00 \u00b1 2.57 |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "llama.cpp #30105",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30105",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": 128,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "lc-30105-r9700-mistral4-tg128-before",
   "model": "Mistral Small 4",
   "notes": "llama.cpp PR #30105, Vulkan FP32 vector kernel selection (Mistral Small 4 perf fix), Radeon AI PRO R9700, Mistral Small 4 IQ4_XS, f16/f32, fitt 1024/fitc 4096; config=before",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "Mistral Small 4; AMD Radeon AI PRO R9700; | mistral4 ?B IQ4_XS - 4.25 bpw  |  54.13 GiB |   118.97 B | Vulkan     |  -1 |   1 |       1024 |        4096 |           tg128 |          2.99 \u00b1 0.02 |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "llama.cpp #30105",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30105",
   "tg_tokens": null,
   "tps": 2.99,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": 128,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "lc-30105-r9700-mistral4-tg128-after",
   "model": "Mistral Small 4",
   "notes": "llama.cpp PR #30105, Vulkan FP32 vector kernel selection (Mistral Small 4 perf fix), Radeon AI PRO R9700, Mistral Small 4 IQ4_XS, f16/f32, fitt 1024/fitc 4096; config=after",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "Mistral Small 4; AMD Radeon AI PRO R9700; | mistral4 ?B IQ4_XS - 4.25 bpw  |  54.13 GiB |   118.97 B | Vulkan     |  -1 |   1 |       1024 |        4096 |           tg128 |         31.56 \u00b1 0.10 |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "llama.cpp #30105",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30105",
   "tg_tokens": null,
   "tps": 31.56,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 100000,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-60426-mi355x-kimik3-c1",
   "model": "Kimi-K3",
   "notes": "vLLM #60426, Kimi-K3, MI355X (gfx950) TP8 DCP1, fp8 KV cache, FULL_DECODE_ONLY cudagraphs, vllm bench ISL 100000 OSL 512 c1; interactivity tok/s. Baseline arm (stock per-group metadata build); PR group-metadata reuse 184.1 (+3.8%), p90 ITL 5.640 to 5.431 ms. config=baseline",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "Kimi-K3's 93 layers are split by the hybrid allocator into 14 KDA and 5 MLA KV cache groups; Measured on MI355X at concurrency 1, ISL 100k; interactivity 177.3 to 184.1 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM #60426",
   "source_url": "https://github.com/vllm-project/vllm/issues/60426",
   "tg_tokens": null,
   "tps": 177.3,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-e2b-prefill-b1",
   "model": "Gemma 4 E2B",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve prefill_heavy (in_len=1024 out_len=1), total tokens/s (in+out); PR arm 34069.2 tok/s (1.34x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 25433.3,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-E2B-it` | prefill_heavy | 1 | 40.30 | 30.09 | **1.34x** | 25433.3 | 34069.2 | **1.34x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 1024,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-e2b-prefill-b4",
   "model": "Gemma 4 E2B",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve prefill_heavy (in_len=1024 out_len=1), total tokens/s (in+out); PR arm 67488.9 tok/s (1.90x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 35480.8,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-E2B-it` | prefill_heavy | 4 | 115.56 | 60.75 | **1.90x** | 35480.8 | 67488.9 | **1.90x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 1024,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-e2b-prefill-b8",
   "model": "Gemma 4 E2B",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve prefill_heavy (in_len=1024 out_len=1), total tokens/s (in+out); PR arm 75137.3 tok/s (1.99x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 37836.3,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-E2B-it` | prefill_heavy | 8 | 216.72 | 109.13 | **1.99x** | 37836.3 | 75137.3 | **1.99x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": 1024,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-e2b-prefill-b16",
   "model": "Gemma 4 E2B",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve prefill_heavy (in_len=1024 out_len=1), total tokens/s (in+out); PR arm 81074.3 tok/s (2.06x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 39380.0,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-E2B-it` | prefill_heavy | 16 | 416.45 | 202.28 | **2.06x** | 39380.0 | 81074.3 | **2.06x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 512,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-e2b-balanced-b1",
   "model": "Gemma 4 E2B",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve balanced (in_len=512 out_len=4), total tokens/s (in+out); PR arm 12839.4 tok/s (1.01x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 12762.8,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-E2B-it` | balanced | 1 | 40.43 | 40.19 | **1.01x** | 12762.8 | 12839.4 | **1.01x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 512,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-e2b-balanced-b4",
   "model": "Gemma 4 E2B",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve balanced (in_len=512 out_len=4), total tokens/s (in+out); PR arm 37118.4 tok/s (1.37x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 27009.7,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-E2B-it` | balanced | 4 | 76.42 | 55.61 | **1.37x** | 27009.7 | 37118.4 | **1.37x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 512,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-e2b-balanced-b8",
   "model": "Gemma 4 E2B",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve balanced (in_len=512 out_len=4), total tokens/s (in+out); PR arm 53667.5 tok/s (1.68x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 31879.1,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-E2B-it` | balanced | 8 | 129.49 | 76.92 | **1.68x** | 31879.1 | 53667.5 | **1.68x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": 512,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-e2b-balanced-b16",
   "model": "Gemma 4 E2B",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve balanced (in_len=512 out_len=4), total tokens/s (in+out); PR arm 62936.3 tok/s (1.75x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 36008.9,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-E2B-it` | balanced | 16 | 229.28 | 131.18 | **1.75x** | 36008.9 | 62936.3 | **1.75x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-e2b-decode-b1",
   "model": "Gemma 4 E2B",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve decode_heavy (in_len=128 out_len=8), total tokens/s (in+out); PR arm 2348.2 tok/s (0.94x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2489.8,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-E2B-it` | decode_heavy | 1 | 54.62 | 57.92 | **0.94x** | 2489.8 | 2348.2 | **0.94x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 128,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-e2b-decode-b4",
   "model": "Gemma 4 E2B",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve decode_heavy (in_len=128 out_len=8), total tokens/s (in+out); PR arm 8723.7 tok/s (1.00x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 8692.8,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-E2B-it` | decode_heavy | 4 | 62.58 | 62.36 | **1.00x** | 8692.8 | 8723.7 | **1.00x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 128,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-e2b-decode-b8",
   "model": "Gemma 4 E2B",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve decode_heavy (in_len=128 out_len=8), total tokens/s (in+out); PR arm 14349.1 tok/s (1.04x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 13812.5,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-E2B-it` | decode_heavy | 8 | 78.77 | 75.82 | **1.04x** | 13812.5 | 14349.1 | **1.04x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": 128,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-e2b-decode-b16",
   "model": "Gemma 4 E2B",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve decode_heavy (in_len=128 out_len=8), total tokens/s (in+out); PR arm 25580.3 tok/s (1.21x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 21129.3,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-E2B-it` | decode_heavy | 16 | 102.98 | 85.07 | **1.21x** | 21129.3 | 25580.3 | **1.21x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-e2b-longctx-b1",
   "model": "Gemma 4 E2B",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve long_context (in_len=4096 out_len=1), total tokens/s (in+out); PR arm 59297.5 tok/s (1.95x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 30383.4,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-E2B-it` | long_context | 1 | 134.84 | 69.09 | **1.95x** | 30383.4 | 59297.5 | **1.95x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 4096,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-e2b-longctx-b4",
   "model": "Gemma 4 E2B",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve long_context (in_len=4096 out_len=1), total tokens/s (in+out); PR arm 68749.8 tok/s (2.10x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 32728.3,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-E2B-it` | long_context | 4 | 500.73 | 238.37 | **2.10x** | 32728.3 | 68749.8 | **2.10x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-e4b-prefill-b1",
   "model": "Gemma 4 E4B",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve prefill_heavy (in_len=1024 out_len=1), total tokens/s (in+out); PR arm 25822.9 tok/s (1.41x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 18361.2,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-E4B-it` | prefill_heavy | 1 | 55.82 | 39.69 | **1.41x** | 18361.2 | 25822.9 | **1.41x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 1024,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-e4b-prefill-b4",
   "model": "Gemma 4 E4B",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve prefill_heavy (in_len=1024 out_len=1), total tokens/s (in+out); PR arm 34290.8 tok/s (1.50x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 22787.7,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-E4B-it` | prefill_heavy | 4 | 179.92 | 119.57 | **1.50x** | 22787.7 | 34290.8 | **1.50x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 1024,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-e4b-prefill-b8",
   "model": "Gemma 4 E4B",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve prefill_heavy (in_len=1024 out_len=1), total tokens/s (in+out); PR arm 36005.0 tok/s (1.53x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 23480.9,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-E4B-it` | prefill_heavy | 8 | 349.22 | 227.75 | **1.53x** | 23480.9 | 36005.0 | **1.53x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": 1024,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-e4b-prefill-b16",
   "model": "Gemma 4 E4B",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve prefill_heavy (in_len=1024 out_len=1), total tokens/s (in+out); PR arm 36874.4 tok/s (1.54x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 23924.9,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-E4B-it` | prefill_heavy | 16 | 685.48 | 444.75 | **1.54x** | 23924.9 | 36874.4 | **1.54x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 512,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-e4b-balanced-b1",
   "model": "Gemma 4 E4B",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve balanced (in_len=512 out_len=4), total tokens/s (in+out); PR arm 9908.0 tok/s (1.11x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 8918.4,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-E4B-it` | balanced | 1 | 57.86 | 52.08 | **1.11x** | 8918.4 | 9908.0 | **1.11x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 512,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-e4b-balanced-b4",
   "model": "Gemma 4 E4B",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve balanced (in_len=512 out_len=4), total tokens/s (in+out); PR arm 22388.8 tok/s (1.33x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 16890.1,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-E4B-it` | balanced | 4 | 122.20 | 92.19 | **1.33x** | 16890.1 | 22388.8 | **1.33x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 512,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-e4b-balanced-b8",
   "model": "Gemma 4 E4B",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve balanced (in_len=512 out_len=4), total tokens/s (in+out); PR arm 28708.2 tok/s (1.40x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 20515.7,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-E4B-it` | balanced | 8 | 201.21 | 143.79 | **1.40x** | 20515.7 | 28708.2 | **1.40x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": 512,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-e4b-balanced-b16",
   "model": "Gemma 4 E4B",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve balanced (in_len=512 out_len=4), total tokens/s (in+out); PR arm 32595.0 tok/s (1.45x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 22436.7,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-E4B-it` | balanced | 16 | 367.97 | 253.29 | **1.45x** | 22436.7 | 32595.0 | **1.45x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-e4b-decode-b1",
   "model": "Gemma 4 E4B",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve decode_heavy (in_len=128 out_len=8), total tokens/s (in+out); PR arm 1600.5 tok/s (0.96x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1666.5,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-E4B-it` | decode_heavy | 1 | 81.61 | 84.97 | **0.96x** | 1666.5 | 1600.5 | **0.96x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 128,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-e4b-decode-b4",
   "model": "Gemma 4 E4B",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve decode_heavy (in_len=128 out_len=8), total tokens/s (in+out); PR arm 6049.2 tok/s (1.02x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5943.5,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-E4B-it` | decode_heavy | 4 | 91.53 | 89.93 | **1.02x** | 5943.5 | 6049.2 | **1.02x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 128,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-e4b-decode-b8",
   "model": "Gemma 4 E4B",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve decode_heavy (in_len=128 out_len=8), total tokens/s (in+out); PR arm 10052.6 tok/s (1.08x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 9302.1,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-E4B-it` | decode_heavy | 8 | 116.96 | 108.23 | **1.08x** | 9302.1 | 10052.6 | **1.08x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": 128,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-e4b-decode-b16",
   "model": "Gemma 4 E4B",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve decode_heavy (in_len=128 out_len=8), total tokens/s (in+out); PR arm 16105.5 tok/s (1.20x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 13435.8,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-E4B-it` | decode_heavy | 16 | 161.96 | 135.11 | **1.20x** | 13435.8 | 16105.5 | **1.20x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-e4b-longctx-b1",
   "model": "Gemma 4 E4B",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve long_context (in_len=4096 out_len=1), total tokens/s (in+out); PR arm 31084.1 tok/s (1.52x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 20474.2,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-E4B-it` | long_context | 1 | 200.11 | 131.80 | **1.52x** | 20474.2 | 31084.1 | **1.52x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 4096,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-e4b-longctx-b4",
   "model": "Gemma 4 E4B",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve long_context (in_len=4096 out_len=1), total tokens/s (in+out); PR arm 33474.3 tok/s (1.57x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 21364.6,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-E4B-it` | long_context | 4 | 767.06 | 489.57 | **1.57x** | 21364.6 | 33474.3 | **1.57x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-12b-prefill-b1",
   "model": "gemma-4-12b-it",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve prefill_heavy (in_len=1024 out_len=1), total tokens/s (in+out); PR arm 8196.8 tok/s (0.99x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 8271.5,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-12B-it` | prefill_heavy | 1 | 123.92 | 125.05 | **0.99x** | 8271.5 | 8196.8 | **0.99x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 1024,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-12b-prefill-b4",
   "model": "gemma-4-12b-it",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve prefill_heavy (in_len=1024 out_len=1), total tokens/s (in+out); PR arm 9417.1 tok/s (1.00x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 9456.8,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-12B-it` | prefill_heavy | 4 | 433.55 | 435.38 | **1.00x** | 9456.8 | 9417.1 | **1.00x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 1024,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-12b-prefill-b8",
   "model": "gemma-4-12b-it",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve prefill_heavy (in_len=1024 out_len=1), total tokens/s (in+out); PR arm 9900.5 tok/s (1.00x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 9922.0,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-12B-it` | prefill_heavy | 8 | 826.45 | 828.25 | **1.00x** | 9922.0 | 9900.5 | **1.00x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": 1024,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-12b-prefill-b16",
   "model": "gemma-4-12b-it",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve prefill_heavy (in_len=1024 out_len=1), total tokens/s (in+out); PR arm 10009.9 tok/s (1.00x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 10004.8,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-12B-it` | prefill_heavy | 16 | 1639.21 | 1638.39 | **1.00x** | 10004.8 | 10009.9 | **1.00x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 512,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-12b-balanced-b1",
   "model": "gemma-4-12b-it",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve balanced (in_len=512 out_len=4), total tokens/s (in+out); PR arm 4331.7 tok/s (1.00x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4339.0,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-12B-it` | balanced | 1 | 118.92 | 119.12 | **1.00x** | 4339.0 | 4331.7 | **1.00x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 512,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-12b-balanced-b4",
   "model": "gemma-4-12b-it",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve balanced (in_len=512 out_len=4), total tokens/s (in+out); PR arm 7732.2 tok/s (1.00x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 7754.7,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-12B-it` | balanced | 4 | 266.16 | 266.93 | **1.00x** | 7754.7 | 7732.2 | **1.00x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 512,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-12b-balanced-b8",
   "model": "gemma-4-12b-it",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve balanced (in_len=512 out_len=4), total tokens/s (in+out); PR arm 8861.7 tok/s (1.00x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 8859.6,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-12B-it` | balanced | 8 | 465.94 | 465.83 | **1.00x** | 8859.6 | 8861.7 | **1.00x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": 512,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-12b-balanced-b16",
   "model": "gemma-4-12b-it",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve balanced (in_len=512 out_len=4), total tokens/s (in+out); PR arm 9823.6 tok/s (1.00x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 9818.5,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-12B-it` | balanced | 16 | 840.86 | 840.42 | **1.00x** | 9818.5 | 9823.6 | **1.00x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-12b-decode-b1",
   "model": "gemma-4-12b-it",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve decode_heavy (in_len=128 out_len=8), total tokens/s (in+out); PR arm 901.1 tok/s (1.01x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 893.6,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-12B-it` | decode_heavy | 1 | 152.19 | 150.92 | **1.01x** | 893.6 | 901.1 | **1.01x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 128,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-12b-decode-b4",
   "model": "gemma-4-12b-it",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve decode_heavy (in_len=128 out_len=8), total tokens/s (in+out); PR arm 2912.3 tok/s (1.00x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2914.4,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-12B-it` | decode_heavy | 4 | 186.66 | 186.79 | **1.00x** | 2914.4 | 2912.3 | **1.00x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 128,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-12b-decode-b8",
   "model": "gemma-4-12b-it",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve decode_heavy (in_len=128 out_len=8), total tokens/s (in+out); PR arm 4544.2 tok/s (1.00x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4529.7,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-12B-it` | decode_heavy | 8 | 240.19 | 239.43 | **1.00x** | 4529.7 | 4544.2 | **1.00x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": 128,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-12b-decode-b16",
   "model": "gemma-4-12b-it",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve decode_heavy (in_len=128 out_len=8), total tokens/s (in+out); PR arm 6427.7 tok/s (1.00x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6409.1,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-12B-it` | decode_heavy | 16 | 339.52 | 338.53 | **1.00x** | 6409.1 | 6427.7 | **1.00x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-12b-longctx-b1",
   "model": "gemma-4-12b-it",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve long_context (in_len=4096 out_len=1), total tokens/s (in+out); PR arm 8137.2 tok/s (1.00x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 8137.0,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-12B-it` | long_context | 1 | 503.50 | 503.49 | **1.00x** | 8137.0 | 8137.2 | **1.00x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 4096,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-12b-longctx-b4",
   "model": "gemma-4-12b-it",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve long_context (in_len=4096 out_len=1), total tokens/s (in+out); PR arm 8545.5 tok/s (1.00x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 8549.3,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-12B-it` | long_context | 4 | 1916.88 | 1917.73 | **1.00x** | 8549.3 | 8545.5 | **1.00x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-26b-prefill-b1",
   "model": "Gemma 4 26B-A4B",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve prefill_heavy (in_len=1024 out_len=1), total tokens/s (in+out); PR arm 15587.3 tok/s (0.99x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 15804.1,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-26B-A4B-it` | prefill_heavy | 1 | 64.86 | 65.76 | **0.99x** | 15804.1 | 15587.3 | **0.99x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 1024,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-26b-prefill-b4",
   "model": "Gemma 4 26B-A4B",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve prefill_heavy (in_len=1024 out_len=1), total tokens/s (in+out); PR arm 19968.5 tok/s (0.99x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 20223.1,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-26B-A4B-it` | prefill_heavy | 4 | 202.74 | 205.32 | **0.99x** | 20223.1 | 19968.5 | **0.99x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 1024,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-26b-prefill-b8",
   "model": "Gemma 4 26B-A4B",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve prefill_heavy (in_len=1024 out_len=1), total tokens/s (in+out); PR arm 21241.1 tok/s (0.99x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 21393.1,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-26B-A4B-it` | prefill_heavy | 8 | 383.30 | 386.04 | **0.99x** | 21393.1 | 21241.1 | **0.99x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": 1024,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-26b-prefill-b16",
   "model": "Gemma 4 26B-A4B",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve prefill_heavy (in_len=1024 out_len=1), total tokens/s (in+out); PR arm 21812.3 tok/s (1.00x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 21907.9,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-26B-A4B-it` | prefill_heavy | 16 | 748.59 | 751.87 | **1.00x** | 21907.9 | 21812.3 | **1.00x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 512,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-26b-balanced-b1",
   "model": "Gemma 4 26B-A4B",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve balanced (in_len=512 out_len=4), total tokens/s (in+out); PR arm 8601.6 tok/s (1.00x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 8592.1,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-26B-A4B-it` | balanced | 1 | 60.06 | 59.99 | **1.00x** | 8592.1 | 8601.6 | **1.00x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 512,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-26b-balanced-b4",
   "model": "Gemma 4 26B-A4B",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve balanced (in_len=512 out_len=4), total tokens/s (in+out); PR arm 14853.9 tok/s (1.00x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 14870.2,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-26B-A4B-it` | balanced | 4 | 138.80 | 138.95 | **1.00x** | 14870.2 | 14853.9 | **1.00x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 512,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-26b-balanced-b8",
   "model": "Gemma 4 26B-A4B",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve balanced (in_len=512 out_len=4), total tokens/s (in+out); PR arm 19228.8 tok/s (0.99x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 19341.7,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-26B-A4B-it` | balanced | 8 | 213.42 | 214.68 | **0.99x** | 19341.7 | 19228.8 | **0.99x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": 512,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-26b-balanced-b16",
   "model": "Gemma 4 26B-A4B",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve balanced (in_len=512 out_len=4), total tokens/s (in+out); PR arm 21197.1 tok/s (0.99x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 21462.7,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-26B-A4B-it` | balanced | 16 | 384.67 | 389.49 | **0.99x** | 21462.7 | 21197.1 | **0.99x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-26b-decode-b1",
   "model": "Gemma 4 26B-A4B",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve decode_heavy (in_len=128 out_len=8), total tokens/s (in+out); PR arm 1847.8 tok/s (1.00x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1853.8,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-26B-A4B-it` | decode_heavy | 1 | 73.36 | 73.60 | **1.00x** | 1853.8 | 1847.8 | **1.00x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 128,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-26b-decode-b4",
   "model": "Gemma 4 26B-A4B",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve decode_heavy (in_len=128 out_len=8), total tokens/s (in+out); PR arm 4605.6 tok/s (1.00x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4607.7,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-26B-A4B-it` | decode_heavy | 4 | 118.06 | 118.12 | **1.00x** | 4607.7 | 4605.6 | **1.00x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 128,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-26b-decode-b8",
   "model": "Gemma 4 26B-A4B",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve decode_heavy (in_len=128 out_len=8), total tokens/s (in+out); PR arm 9071.3 tok/s (0.98x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 9212.9,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-26B-A4B-it` | decode_heavy | 8 | 118.09 | 119.94 | **0.98x** | 9212.9 | 9071.3 | **0.98x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": 128,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-26b-decode-b16",
   "model": "Gemma 4 26B-A4B",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve decode_heavy (in_len=128 out_len=8), total tokens/s (in+out); PR arm 13260.1 tok/s (0.99x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 13345.6,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-26B-A4B-it` | decode_heavy | 16 | 163.05 | 164.10 | **0.99x** | 13345.6 | 13260.1 | **0.99x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-26b-longctx-b1",
   "model": "Gemma 4 26B-A4B",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve long_context (in_len=4096 out_len=1), total tokens/s (in+out); PR arm 16426.9 tok/s (1.00x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 16495.5,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-26B-A4B-it` | long_context | 1 | 248.37 | 249.41 | **1.00x** | 16495.5 | 16426.9 | **1.00x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 4096,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-26b-longctx-b4",
   "model": "Gemma 4 26B-A4B",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve long_context (in_len=4096 out_len=1), total tokens/s (in+out); PR arm 17636.1 tok/s (1.00x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 17685.1,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-26B-A4B-it` | long_context | 4 | 926.65 | 929.23 | **1.00x** | 17685.1 | 17636.1 | **1.00x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-31b-prefill-b1",
   "model": "gemma-4-31B-it",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve prefill_heavy (in_len=1024 out_len=1), total tokens/s (in+out); PR arm 3559.9 tok/s (1.00x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3566.7,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-31B-it` | prefill_heavy | 1 | 287.38 | 287.93 | **1.00x** | 3566.7 | 3559.9 | **1.00x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 1024,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-31b-prefill-b4",
   "model": "gemma-4-31B-it",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve prefill_heavy (in_len=1024 out_len=1), total tokens/s (in+out); PR arm 3830.3 tok/s (1.00x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3835.8,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-31B-it` | prefill_heavy | 4 | 1068.88 | 1070.41 | **1.00x** | 3835.8 | 3830.3 | **1.00x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 512,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-31b-balanced-b1",
   "model": "gemma-4-31B-it",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve balanced (in_len=512 out_len=4), total tokens/s (in+out); PR arm 1904.2 tok/s (1.00x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1906.1,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-31B-it` | balanced | 1 | 270.70 | 270.99 | **1.00x** | 1906.1 | 1904.2 | **1.00x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 512,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-31b-balanced-b4",
   "model": "gemma-4-31B-it",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve balanced (in_len=512 out_len=4), total tokens/s (in+out); PR arm 3172.4 tok/s (1.00x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3173.0,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-31B-it` | balanced | 4 | 650.49 | 650.61 | **1.00x** | 3173.0 | 3172.4 | **1.00x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-31b-decode-b1",
   "model": "gemma-4-31B-it",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve decode_heavy (in_len=128 out_len=8), total tokens/s (in+out); PR arm 398.0 tok/s (1.00x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 398.2,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-31B-it` | decode_heavy | 1 | 341.51 | 341.73 | **1.00x** | 398.2 | 398.0 | **1.00x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 128,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-31b-decode-b4",
   "model": "gemma-4-31B-it",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve decode_heavy (in_len=128 out_len=8), total tokens/s (in+out); PR arm 1265.7 tok/s (1.00x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1266.7,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-31B-it` | decode_heavy | 4 | 429.46 | 429.81 | **1.00x** | 1266.7 | 1265.7 | **1.00x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-31b-longctx-b1",
   "model": "gemma-4-31B-it",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve long_context (in_len=4096 out_len=1), total tokens/s (in+out); PR arm 3288.1 tok/s (0.99x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3306.2,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-31B-it` | long_context | 1 | 1239.19 | 1246.02 | **0.99x** | 3306.2 | 3288.1 | **0.99x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 4096,
   "date": "2026-10",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60445-31b-longctx-b4",
   "model": "gemma-4-31B-it",
   "notes": "vLLM PR #60445 KV-sharing fast prefill (YOCO) auto-enable + piecewise CUDAGraph/metadata optimization; vllm bench serve long_context (in_len=4096 out_len=1), total tokens/s (in+out); PR arm 3385.3 tok/s (1.00x); enforce_eager=False, torch.compile + CUDAGraphs",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3390.3,
   "provenance": "community",
   "quant": "BF16",
   "quote": "measured on NVIDIA A100-SXM4-80GB (CUDA 12.6, PyTorch 2.13.0, warmup=20, reps=50); | `gemma-4-31B-it` | long_context | 4 | 4833.84 | 4840.89 | **1.00x** | 3390.3 | 3385.3 | **1.00x** | 100.0% |",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM PR #60445 (gemma-4 KV-sharing fast prefill)",
   "source_url": "https://github.com/vllm-project/vllm/issues/60445",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ada-surgery fork)",
   "batch": null,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "RTX 4070",
   "id": "hf1-ternary4070-4k",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "PrismML Ternary Bonsai 2 27B (1.58-bit ternary PTQ1_0, 5.9 GB) re-served unchanged with the MTP draft head grafted back, via the ada-surgery fork (professorpalmer/bonsai-ada-surgery): tiered KV cache (first ~113k positions in VRAM, remainder in pinned system RAM) at q8_0 KV, one slot on a 12 GB card. 2026-10-08 depth table, fresh prompt at each depth, 400-token greedy answers with default drafting (lookup + MTP head); decode = new-text tok/s at this context depth; prefill = cumulative prefill tok/s; card also publishes a code-decode column.",
   "pp_tps": 1173,
   "provenance": "community",
   "quant": "PTQ1_0",
   "quote": "RTX 4070 12 GB, served, one slot; Ternary Bonsai 2 27B; 4k tokens of context 102 102 1,173",
   "ram_gb": 12,
   "retrieved": "2026-10-14",
   "source_name": "HF model card: CaryPalmer/Ternary-Bonsai-2-27B-262k-GGUF",
   "source_url": "https://huggingface.co/CaryPalmer/Ternary-Bonsai-2-27B-262k-GGUF",
   "tps": "102"
  },
  {
   "backend": "llama.cpp (ada-surgery fork)",
   "batch": null,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "RTX 4070",
   "id": "hf1-ternary4070-16k",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "PrismML Ternary Bonsai 2 27B (1.58-bit ternary PTQ1_0, 5.9 GB) re-served unchanged with the MTP draft head grafted back, via the ada-surgery fork (professorpalmer/bonsai-ada-surgery): tiered KV cache (first ~113k positions in VRAM, remainder in pinned system RAM) at q8_0 KV, one slot on a 12 GB card. 2026-10-08 depth table, fresh prompt at each depth, 400-token greedy answers with default drafting (lookup + MTP head); decode = new-text tok/s at this context depth; prefill = cumulative prefill tok/s; card also publishes a code-decode column.",
   "pp_tps": 1137,
   "provenance": "community",
   "quant": "PTQ1_0",
   "quote": "RTX 4070 12 GB, served, one slot; Ternary Bonsai 2 27B; 16k 83 242 1,137",
   "ram_gb": 12,
   "retrieved": "2026-10-14",
   "source_name": "HF model card: CaryPalmer/Ternary-Bonsai-2-27B-262k-GGUF",
   "source_url": "https://huggingface.co/CaryPalmer/Ternary-Bonsai-2-27B-262k-GGUF",
   "tps": "83"
  },
  {
   "backend": "llama.cpp (ada-surgery fork)",
   "batch": null,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "RTX 4070",
   "id": "hf1-ternary4070-32k",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "PrismML Ternary Bonsai 2 27B (1.58-bit ternary PTQ1_0, 5.9 GB) re-served unchanged with the MTP draft head grafted back, via the ada-surgery fork (professorpalmer/bonsai-ada-surgery): tiered KV cache (first ~113k positions in VRAM, remainder in pinned system RAM) at q8_0 KV, one slot on a 12 GB card. 2026-10-08 depth table, fresh prompt at each depth, 400-token greedy answers with default drafting (lookup + MTP head); decode = new-text tok/s at this context depth; prefill = cumulative prefill tok/s; card also publishes a code-decode column.",
   "pp_tps": 1031,
   "provenance": "community",
   "quant": "PTQ1_0",
   "quote": "RTX 4070 12 GB, served, one slot; Ternary Bonsai 2 27B; 32k 78 175 1,031",
   "ram_gb": 12,
   "retrieved": "2026-10-14",
   "source_name": "HF model card: CaryPalmer/Ternary-Bonsai-2-27B-262k-GGUF",
   "source_url": "https://huggingface.co/CaryPalmer/Ternary-Bonsai-2-27B-262k-GGUF",
   "tps": "78"
  },
  {
   "backend": "llama.cpp (ada-surgery fork)",
   "batch": null,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "RTX 4070",
   "id": "hf1-ternary4070-64k",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "PrismML Ternary Bonsai 2 27B (1.58-bit ternary PTQ1_0, 5.9 GB) re-served unchanged with the MTP draft head grafted back, via the ada-surgery fork (professorpalmer/bonsai-ada-surgery): tiered KV cache (first ~113k positions in VRAM, remainder in pinned system RAM) at q8_0 KV, one slot on a 12 GB card. 2026-10-08 depth table, fresh prompt at each depth, 400-token greedy answers with default drafting (lookup + MTP head); decode = new-text tok/s at this context depth; prefill = cumulative prefill tok/s; card also publishes a code-decode column.",
   "pp_tps": 867,
   "provenance": "community",
   "quant": "PTQ1_0",
   "quote": "RTX 4070 12 GB, served, one slot; Ternary Bonsai 2 27B; 64k 69 109 867",
   "ram_gb": 12,
   "retrieved": "2026-10-14",
   "source_name": "HF model card: CaryPalmer/Ternary-Bonsai-2-27B-262k-GGUF",
   "source_url": "https://huggingface.co/CaryPalmer/Ternary-Bonsai-2-27B-262k-GGUF",
   "tps": "69"
  },
  {
   "backend": "llama.cpp (ada-surgery fork)",
   "batch": null,
   "ctx": 114688,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "RTX 4070",
   "id": "hf1-ternary4070-112k",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "PrismML Ternary Bonsai 2 27B (1.58-bit ternary PTQ1_0, 5.9 GB) re-served unchanged with the MTP draft head grafted back, via the ada-surgery fork (professorpalmer/bonsai-ada-surgery): tiered KV cache (first ~113k positions in VRAM, remainder in pinned system RAM) at q8_0 KV, one slot on a 12 GB card. 2026-10-08 depth table, fresh prompt at each depth, 400-token greedy answers with default drafting (lookup + MTP head); decode = new-text tok/s at this context depth; prefill = cumulative prefill tok/s; card also publishes a code-decode column.",
   "pp_tps": 698,
   "provenance": "community",
   "quant": "PTQ1_0",
   "quote": "RTX 4070 12 GB, served, one slot; Ternary Bonsai 2 27B; 112k (in VRAM) 61 91 698",
   "ram_gb": 12,
   "retrieved": "2026-10-14",
   "source_name": "HF model card: CaryPalmer/Ternary-Bonsai-2-27B-262k-GGUF",
   "source_url": "https://huggingface.co/CaryPalmer/Ternary-Bonsai-2-27B-262k-GGUF",
   "tps": "61"
  },
  {
   "backend": "llama.cpp (ada-surgery fork)",
   "batch": null,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "RTX 4070",
   "id": "hf1-ternary4070-131k",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "PrismML Ternary Bonsai 2 27B (1.58-bit ternary PTQ1_0, 5.9 GB) re-served unchanged with the MTP draft head grafted back, via the ada-surgery fork (professorpalmer/bonsai-ada-surgery): tiered KV cache (first ~113k positions in VRAM, remainder in pinned system RAM) at q8_0 KV, one slot on a 12 GB card. 2026-10-08 depth table, fresh prompt at each depth, 400-token greedy answers with default drafting (lookup + MTP head); decode = new-text tok/s at this context depth; prefill = cumulative prefill tok/s; card also publishes a code-decode column.",
   "pp_tps": 649,
   "provenance": "community",
   "quant": "PTQ1_0",
   "quote": "RTX 4070 12 GB, served, one slot; Ternary Bonsai 2 27B; 131k 30 63 649",
   "ram_gb": 12,
   "retrieved": "2026-10-14",
   "source_name": "HF model card: CaryPalmer/Ternary-Bonsai-2-27B-262k-GGUF",
   "source_url": "https://huggingface.co/CaryPalmer/Ternary-Bonsai-2-27B-262k-GGUF",
   "tps": "30"
  },
  {
   "backend": "llama.cpp (ada-surgery fork)",
   "batch": null,
   "ctx": 184320,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "RTX 4070",
   "id": "hf1-ternary4070-180k",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "PrismML Ternary Bonsai 2 27B (1.58-bit ternary PTQ1_0, 5.9 GB) re-served unchanged with the MTP draft head grafted back, via the ada-surgery fork (professorpalmer/bonsai-ada-surgery): tiered KV cache (first ~113k positions in VRAM, remainder in pinned system RAM) at q8_0 KV, one slot on a 12 GB card. 2026-10-08 depth table, fresh prompt at each depth, 400-token greedy answers with default drafting (lookup + MTP head); decode = new-text tok/s at this context depth; prefill = cumulative prefill tok/s; card also publishes a code-decode column.",
   "pp_tps": 547,
   "provenance": "community",
   "quant": "PTQ1_0",
   "quote": "RTX 4070 12 GB, served, one slot; Ternary Bonsai 2 27B; 180k 17.9 28 547",
   "ram_gb": 12,
   "retrieved": "2026-10-14",
   "source_name": "HF model card: CaryPalmer/Ternary-Bonsai-2-27B-262k-GGUF",
   "source_url": "https://huggingface.co/CaryPalmer/Ternary-Bonsai-2-27B-262k-GGUF",
   "tps": "17.9"
  },
  {
   "backend": "llama.cpp (ada-surgery fork)",
   "batch": null,
   "ctx": 262144,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "RTX 4070",
   "id": "hf1-ternary4070-258k",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "PrismML Ternary Bonsai 2 27B (1.58-bit ternary PTQ1_0, 5.9 GB) re-served unchanged with the MTP draft head grafted back, via the ada-surgery fork (professorpalmer/bonsai-ada-surgery): tiered KV cache (first ~113k positions in VRAM, remainder in pinned system RAM) at q8_0 KV, one slot on a 12 GB card. 2026-10-08 depth table, fresh prompt at each depth, 400-token greedy answers with default drafting (lookup + MTP head); decode = new-text tok/s at this context depth; prefill = cumulative prefill tok/s; card also publishes a code-decode column.",
   "pp_tps": 436,
   "provenance": "community",
   "quant": "PTQ1_0",
   "quote": "RTX 4070 12 GB, served, one slot; Ternary Bonsai 2 27B; 258k (the window's end) 8.6 12.1 436",
   "ram_gb": 12,
   "retrieved": "2026-10-14",
   "source_name": "HF model card: CaryPalmer/Ternary-Bonsai-2-27B-262k-GGUF",
   "source_url": "https://huggingface.co/CaryPalmer/Ternary-Bonsai-2-27B-262k-GGUF",
   "tps": "8.6"
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Max",
   "id": "hf2-d1-3b-m3max",
   "model": "d1-3B",
   "notes": "LiquidAI d1-3B (3B decision model, MLX 8-bit) on M3 Max 36 GB, mlx-vlm; 246-token image prompt, greedy, warm; decode 79.0 / prefill 746 tok/s (4-bit 134.0, BF16 50.4).",
   "pp_tps": 746,
   "provenance": "community",
   "quant": "8bit",
   "quote": "DJLougen / d1-3B-MLX-8bit; Apple M3 Max with 36 GB unified memory; Decode tok/s (246-token image prompt) 79.0 Prefill tok/s 746",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "HF model card: DJLougen/d1-3B-MLX-8bit",
   "source_url": "https://huggingface.co/DJLougen/d1-3B-MLX-8bit",
   "tps": "79.0"
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": 26235,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "hf2-cyberfrost-m3ultra-26k",
   "model": "CYBER-FROST-3.8",
   "notes": "CYBER-FROST 3.8 (MLX 4-bit) on M3 Ultra 256 GB, TensorFold; 26,235-token agentic workload, 1,024-token replies, xhigh thinking, MTP drafting on; warm decode 96.3 (cold prefill 1,012).",
   "pp_tps": 1012,
   "provenance": "community",
   "quant": "4bit",
   "quote": "grant-ai / CYBER-FROST-3.8-MLX-4bit; Measured on an M3 Ultra Mac Studio (256 GB); xhigh thinking, TensorFold 96.3 tok/s",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "HF model card: grant-ai/CYBER-FROST-3.8-MLX-4bit",
   "source_url": "https://huggingface.co/grant-ai/CYBER-FROST-3.8-MLX-4bit",
   "tps": "96.3"
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "hf2-cyberfrost-m3ultra-short",
   "model": "CYBER-FROST-3.8",
   "notes": "CYBER-FROST 3.8 (MLX 4-bit) on M3 Ultra 256 GB, TensorFold short-reply bench (64-token replies, thinking off); 136.6 greedy / 109.5 sampled tok/s.",
   "pp_tps": null,
   "provenance": "community",
   "quant": "4bit",
   "quote": "grant-ai / CYBER-FROST-3.8-MLX-4bit; Measured on an M3 Ultra Mac Studio (256 GB); 136.6 tok/s greedy \u00b7 109.5 tok/s sampled",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "HF model card: grant-ai/CYBER-FROST-3.8-MLX-4bit",
   "source_url": "https://huggingface.co/grant-ai/CYBER-FROST-3.8-MLX-4bit",
   "tps": "136.6"
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Ultra (64 GB)",
   "id": "hf2-minicpmm46-m1ultra",
   "model": "MiniCPM-V-4.6",
   "notes": "MiniCPM-V 4.6 (Huihui abliterated, 8-bit MLX) on M1 Ultra 64 GB, oMLX; ~140-180 tok/s decode; 144 tok/s on 30-line large-page OCR (2.3 GB).",
   "pp_tps": null,
   "provenance": "community",
   "quant": "8bit",
   "quote": "YCF-AI / MiniCPM-V-4.6-Huihui-abliterated-8bit-MLX; Measured on Mac Studio M1 Ultra 64 GB; 144 tok/s (previous 145)",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "HF model card: YCF-AI/MiniCPM-V-4.6-Huihui-abliterated-8bit-MLX",
   "source_url": "https://huggingface.co/YCF-AI/MiniCPM-V-4.6-Huihui-abliterated-8bit-MLX",
   "tps": "144"
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Ultra (64 GB)",
   "id": "hf2-qwenh-m1ultra-chat",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "Qwen3.8-27B Huihui abliterated (oQ4e, MTP fixed depth 2, FP16) on M1 Ultra 64 GB, oMLX; chat decode 48-49 / prefill ~280 tok/s; dense 27B 4-bit ceiling on M1.",
   "pp_tps": 280,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "YCF-AI / Qwen3.8-27B-Huihui-abliterated-oQ4e-MTP-FP16-MLX; Measured on Mac Studio M1 Ultra 64 GB; chat decode 48\u201349 tok/s (MTP fixed depth 2); prefill \u2248 280 tok/s",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "HF model card: YCF-AI/Qwen3.8-27B-Huihui-abliterated-oQ4e-MTP-FP16-MLX",
   "source_url": "https://huggingface.co/YCF-AI/Qwen3.8-27B-Huihui-abliterated-oQ4e-MTP-FP16-MLX",
   "tps": "49"
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Ultra (64 GB)",
   "id": "hf2-qwenh-m1ultra-16k",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "Qwen3.8-27B Huihui abliterated (oQ4e, MTP) on M1 Ultra 64 GB, oMLX; 16K-context decode 37-38 tok/s.",
   "pp_tps": null,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "YCF-AI / Qwen3.8-27B-Huihui-abliterated-oQ4e-MTP-FP16-MLX; Measured on Mac Studio M1 Ultra 64 GB; 16K decode 37\u201338",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "HF model card: YCF-AI/Qwen3.8-27B-Huihui-abliterated-oQ4e-MTP-FP16-MLX",
   "source_url": "https://huggingface.co/YCF-AI/Qwen3.8-27B-Huihui-abliterated-oQ4e-MTP-FP16-MLX",
   "tps": "38"
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Ultra (64 GB)",
   "id": "hf2-cybertiel-m1ultra",
   "model": "CyberTiel-Coder-35B",
   "notes": "Cyber-Tiel-Coder 35B (35B-A3B MoE, oQ6e 6-bit, MTP) on Mac Studio M1 Ultra 64 GB, oMLX 0.7.0; chat decode 107-116, prefill 4K ~1890 / 16K ~1640 tok/s.",
   "pp_tps": 1890,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "YCF-AI / CyberTiel-Coder-35B-MLX; every day on a Mac Studio M1 Ultra 64 GB with oMLX 0.7.0; Speed (M1 Ultra, chat) 107\u2013116 tok/s decode; prefill 4K \u2248 1890, 16K \u2248 1640 tok/s",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "HF model card: YCF-AI/CyberTiel-Coder-35B-MLX",
   "source_url": "https://huggingface.co/YCF-AI/CyberTiel-Coder-35B-MLX",
   "tps": "116"
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Max",
   "id": "hf2-laguna-m3max",
   "model": "Laguna-M.1",
   "notes": "Laguna M.1 (Q3 MLX, 256-expert MoE top-16) on M3 Max 128 GB; ~26.5 tok/s, peak RAM ~100 GB (approximate, single value in card).",
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q3",
   "quote": "ox-ox / Laguna-M.1-MLX-Q3; Apple M3 Max 128 GB ~26.5 tok/s ~100 GB",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "HF model card: ox-ox/Laguna-M.1-MLX-Q3",
   "source_url": "https://huggingface.co/ox-ox/Laguna-M.1-MLX-Q3",
   "tps": "26.5"
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": 128,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra (512 GB)",
   "id": "hf2-altar1-m3ultra-128",
   "model": "altar-1",
   "notes": "altar-1 (4-bit MLX, ~330+ GB unified) on M3 Ultra 512 GB; ctx 128: decode 13.7 / prefill 86 tok/s (card performance table).",
   "pp_tps": 86,
   "provenance": "community",
   "quant": "4bit",
   "quote": "openvuln / altar-1-4bit-mlx; Performance (Mac Studio M3 Ultra, 512 GB); 128 86 1.5 s 13.7 315 GB",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "HF model card: openvuln/altar-1-4bit-mlx",
   "source_url": "https://huggingface.co/openvuln/altar-1-4bit-mlx",
   "tps": "13.7"
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": 2048,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra (512 GB)",
   "id": "hf2-altar1-m3ultra-2k",
   "model": "altar-1",
   "notes": "altar-1 (4-bit MLX) on M3 Ultra 512 GB; ctx 2k dense: decode 12.6 / prefill 205 tok/s (card performance table).",
   "pp_tps": 205,
   "provenance": "community",
   "quant": "4bit",
   "quote": "openvuln / altar-1-4bit-mlx; Performance (Mac Studio M3 Ultra, 512 GB); 2k (dense) 205 10.0 s 12.6 319 GB",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "HF model card: openvuln/altar-1-4bit-mlx",
   "source_url": "https://huggingface.co/openvuln/altar-1-4bit-mlx",
   "tps": "12.6"
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra (512 GB)",
   "id": "hf2-altar1-m3ultra-4k",
   "model": "altar-1",
   "notes": "altar-1 (4-bit MLX) on M3 Ultra 512 GB; ctx 4k: decode 13.0 / prefill 194 tok/s (card performance table).",
   "pp_tps": 194,
   "provenance": "community",
   "quant": "4bit",
   "quote": "openvuln / altar-1-4bit-mlx; Performance (Mac Studio M3 Ultra, 512 GB); 4k 194 21.1 s 13.0 320 GB",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "HF model card: openvuln/altar-1-4bit-mlx",
   "source_url": "https://huggingface.co/openvuln/altar-1-4bit-mlx",
   "tps": "13.0"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": 160000,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "hf2-ornith35b-4090-160k",
   "model": "Ornith-1.5-35B",
   "notes": "Ornith-1.5-35B (35B-A3B MoE, Q4_K_M) on RTX 4090 24 GB, llama.cpp; end-to-end decode 113.8 tok/s at 160K context with DFlash2 draft speculation (acceptance 0.471, 1.88 accepted/draft); prefill ~2.2k tok/s.",
   "pp_tps": 2210,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "3ntr0py-t4m3r / Ornith-1.5-DFlash2-Draft-GGUF; Target model Ornith-1.5-35B-Q4_K_M; 2026-10-07 on a single RTX 4090 24 GB, q4_0 against q4_0; 0.471 1.88 113.8",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "HF model card: 3ntr0py-t4m3r/Ornith-1.5-DFlash2-Draft-GGUF",
   "source_url": "https://huggingface.co/3ntr0py-t4m3r/Ornith-1.5-DFlash2-Draft-GGUF",
   "tps": "113.8"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "2x RTX 4090",
   "id": "llama30112-2x4090-master",
   "model": "Qwen3.8-Flash-Next",
   "notes": "llama.cpp PR #30112 multi-GPU MoE cache; SPEED-Bench qualitative, osl 512, concurrency 1, temp 0; master --fit baseline; config=master",
   "pp_tps": 440.7,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "2x RTX 4090 (layer split), Qwen3.8-Flash-Next Q4_0; master `--fit` | 34.78 | 1.00x | 440.7",
   "retrieved": "2026-10-07",
   "source_name": "llama.cpp issue #30112",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30112",
   "tps": "34.78"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "2x RTX 4090",
   "id": "llama30112-2x4090-moe6544",
   "model": "Qwen3.8-Flash-Next",
   "notes": "llama.cpp PR #30112 multi-GPU MoE cache; SPEED-Bench qualitative, osl 512, concurrency 1; --fit --moe-cache-mib 6544 (85.9% cache hit), 1.74x decode vs master; prompt t/s 402.0; config=moe-cache-6544",
   "pp_tps": 402.0,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "2x RTX 4090 (layer split), Qwen3.8-Flash-Next Q4_0; `--fit --moe-cache-mib 6544` | 60.38 | 1.74x | 402.0",
   "retrieved": "2026-10-07",
   "source_name": "llama.cpp issue #30112",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30112",
   "tps": "60.38"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "2x RTX 4090",
   "id": "llama30112-2x4090-cmoe15000",
   "model": "Qwen3.8-Flash-Next",
   "notes": "llama.cpp PR #30112 multi-GPU MoE cache; SPEED-Bench qualitative, osl 512, concurrency 1; -cmoe --moe-cache-mib 15000 (95.8% cache hit), 1.85x decode vs master; prompt t/s 321.5; config=cmoe-15000",
   "pp_tps": 321.5,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "2x RTX 4090 (layer split), Qwen3.8-Flash-Next Q4_0; `-cmoe --moe-cache-mib 15000` | 64.39 | 1.85x | 321.5",
   "retrieved": "2026-10-07",
   "source_name": "llama.cpp issue #30112",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30112",
   "tps": "64.39"
  },
  {
   "backend": "llama.cpp (Hexagon)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "Hexagon",
   "id": "llama30115-hexagon-lfm25",
   "model": "LFM2.5-2.6B",
   "notes": "llama.cpp PR #30115 tiled Q4_K/Q6_K GET_ROWS on Hexagon (HTP0); LFM2.5-2.6B Q4_K_M with LFM2.5-DSpark draft speculation, -n 256, c 4096; decode 43.88 / prompt 57.49 tok/s",
   "pp_tps": 57.49,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "LFM2.5-2.6B-Q4_K_M.gguf; LFM2.5-DSpark draft model onto hexagon; Decode: 43.88 tok/s; Prompt: 57.49 tok/s",
   "retrieved": "2026-10-07",
   "source_name": "llama.cpp issue #30115",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30115",
   "tps": "43.88"
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": null,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "H100 NVL",
   "id": "vllm60464-h100nvl-chat16",
   "model": "Qwen3 8B",
   "notes": "vLLM PR #60464 SM90 block-FP8 CUTLASS kernel dispatch by M size (swap-AB to M=64); vllm bench serve, output tok/s at concurrency; stock CUTLASS baseline (this PR 1,816, +51%); Qwen3-8B-FP8 H100 NVL",
   "pp_tps": 1202,
   "provenance": "community",
   "quant": "FP8",
   "quote": "Serving Qwen3-8B-FP8 on an H100 NVL with the CUTLASS path; chat, 16 concurrent | 1,202 | 1,816 | +51%",
   "retrieved": "2026-10-07",
   "source_name": "vLLM issue #60464",
   "source_url": "https://github.com/vllm-project/vllm/issues/60464",
   "tps": null
  },
  {
   "backend": "vLLM",
   "batch": 64,
   "ctx": null,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "H100 NVL",
   "id": "vllm60464-h100nvl-chat64",
   "model": "Qwen3 8B",
   "notes": "vLLM PR #60464 SM90 block-FP8 CUTLASS kernel dispatch by M size; vllm bench serve, output tok/s at concurrency; stock CUTLASS baseline (this PR 4,863, +17%); Qwen3-8B-FP8 H100 NVL",
   "pp_tps": 4160,
   "provenance": "community",
   "quant": "FP8",
   "quote": "Serving Qwen3-8B-FP8 on an H100 NVL with the CUTLASS path; chat, 64 concurrent | 4,160 | 4,863 | +17%",
   "retrieved": "2026-10-07",
   "source_name": "vLLM issue #60464",
   "source_url": "https://github.com/vllm-project/vllm/issues/60464",
   "tps": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": null,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "H100 NVL",
   "id": "vllm60464-h100nvl-long16",
   "model": "Qwen3 8B",
   "notes": "vLLM PR #60464 SM90 block-FP8 CUTLASS kernel dispatch by M size; vllm bench serve, output tok/s at concurrency; stock CUTLASS baseline (this PR 667, +26%); long prompts; Qwen3-8B-FP8 H100 NVL",
   "pp_tps": 529,
   "provenance": "community",
   "quant": "FP8",
   "quote": "Serving Qwen3-8B-FP8 on an H100 NVL with the CUTLASS path; long prompts, 16 concurrent | 529 | 667 | +26%",
   "retrieved": "2026-10-07",
   "source_name": "vLLM issue #60464",
   "source_url": "https://github.com/vllm-project/vllm/issues/60464",
   "tps": null
  },
  {
   "backend": "vLLM",
   "batch": 64,
   "ctx": null,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "H100 NVL",
   "id": "vllm60464-h100nvl-long64",
   "model": "Qwen3 8B",
   "notes": "vLLM PR #60464 SM90 block-FP8 CUTLASS kernel dispatch by M size; vllm bench serve, output tok/s at concurrency; stock CUTLASS baseline (this PR 965, +52%); long prompts; Qwen3-8B-FP8 H100 NVL",
   "pp_tps": 634,
   "provenance": "community",
   "quant": "FP8",
   "quote": "Serving Qwen3-8B-FP8 on an H100 NVL with the CUTLASS path; long prompts, 64 concurrent | 634 | 965 | +52%",
   "retrieved": "2026-10-07",
   "source_name": "vLLM issue #60464",
   "source_url": "https://github.com/vllm-project/vllm/issues/60464",
   "tps": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "Phytium D3000M",
   "id": "lc-30150-phytium-d3000m-qwen3508b-q80-base",
   "model": "Qwen3.5-0.8B",
   "notes": "llama.cpp PR #30150 fixed-width concat threaded over all rows, ARM NEON armv8.2a, 8 threads, 12 interleaved A/B pairs medians; config=master; PR arm: pp512 295.0 (+33%), tg128 26.79 (+9.8%)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 221.4,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Qwen3.5-0.8B Q8_0, Phytium D3000M(armv8.2a)/8t, 12 interleaved A/B pairs, medians: pp512 221.4 -> 295.0 t/s, tg128 24.40 -> 26.79 t/s",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "llama.cpp issue #30150",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30150",
   "tg_tokens": 128,
   "tps": 24.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "Phytium D3000M",
   "id": "lc-30150-phytium-d3000m-qwen3508b-q80-pr",
   "model": "Qwen3.5-0.8B",
   "notes": "llama.cpp PR #30150 fixed-width concat threaded over all rows, ARM NEON armv8.2a, 8 threads, 12 interleaved A/B pairs medians; config=pr-30150; master arm: pp512 221.4, tg128 24.40",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 295.0,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Qwen3.5-0.8B Q8_0, Phytium D3000M(armv8.2a)/8t, 12 interleaved A/B pairs, medians: pp512 221.4 -> 295.0 t/s, tg128 24.40 -> 26.79 t/s",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "llama.cpp issue #30150",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30150",
   "tg_tokens": 128,
   "tps": 26.79,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "Intel(R) Core(TM) Ultra 9 285K",
   "id": "lc-30150-core-ultra9-285k-qwen3508b-q40-base",
   "model": "Qwen3.5-0.8B",
   "notes": "llama.cpp PR #30150 fixed-width concat threaded over all rows, 8 threads cpu_mask 0xff cpu_strict 1, medians; config=master; PR arm: pp512 889.15 (+17.2%), tg128 86.21 (+1.4%)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 758.97,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Qwen3.5-0.8B Q4_0, Intel(R) Core(TM) Ultra 9 285K/24t, 8 threads, cpu_mask 0xff, cpu_strict 1, medians: pp512 758.97 -> 889.15 t/s (+17.2%), tg128 85.03 -> 86.21 t/s (+1.4%)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "llama.cpp issue #30150",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30150",
   "tg_tokens": 128,
   "tps": 85.03,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "Intel(R) Core(TM) Ultra 9 285K",
   "id": "lc-30150-core-ultra9-285k-qwen3508b-q40-pr",
   "model": "Qwen3.5-0.8B",
   "notes": "llama.cpp PR #30150 fixed-width concat threaded over all rows, 8 threads cpu_mask 0xff cpu_strict 1, medians; config=pr-30150; master arm: pp512 758.97, tg128 85.03",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 889.15,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Qwen3.5-0.8B Q4_0, Intel(R) Core(TM) Ultra 9 285K/24t, 8 threads, cpu_mask 0xff, cpu_strict 1, medians: pp512 758.97 -> 889.15 t/s (+17.2%), tg128 85.03 -> 86.21 t/s (+1.4%)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "llama.cpp issue #30150",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30150",
   "tg_tokens": 128,
   "tps": 86.21,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": null,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60674-a100-gemma4e2b-base",
   "model": "Gemma 4 E2B",
   "notes": "vLLM PR #60674 fused Triton kernels (QKV norm/RoPE, pre/post-FFN norms, PLE, MoE routing); 1k prefill batch 4; config=main; PR arm: 67488.9 tok/s (1.90x)",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 35480.8,
   "provenance": "community",
   "quant": "BF16",
   "quote": "A100-SXM4-80GB; gemma-4-E2B-it: Up to 1.90x prefill throughput speedup (67,488.9 tok/s vs 35,480.8 tok/s at BS=4, 1k prefill)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "vLLM issue #60674",
   "source_url": "https://github.com/vllm-project/vllm/pull/60674",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": null,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60674-a100-gemma4e2b-pr",
   "model": "Gemma 4 E2B",
   "notes": "vLLM PR #60674 fused Triton kernels (QKV norm/RoPE, pre/post-FFN norms, PLE, MoE routing); 1k prefill batch 4; config=pr-60674; main arm: 35480.8 tok/s",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 67488.9,
   "provenance": "community",
   "quant": "BF16",
   "quote": "A100-SXM4-80GB; gemma-4-E2B-it: Up to 1.90x prefill throughput speedup (67,488.9 tok/s vs 35,480.8 tok/s at BS=4, 1k prefill)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "vLLM issue #60674",
   "source_url": "https://github.com/vllm-project/vllm/pull/60674",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": null,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60674-a100-gemma4e4b-base",
   "model": "Gemma 4 E4B",
   "notes": "vLLM PR #60674 fused Triton kernels (QKV norm/RoPE, pre/post-FFN norms, PLE, MoE routing); 1k prefill batch 4; config=main; PR arm: 34290.8 tok/s (1.50x)",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 22787.7,
   "provenance": "community",
   "quant": "BF16",
   "quote": "A100-SXM4-80GB; gemma-4-E4B-it: Up to 1.50x prefill throughput speedup (34,290.8 tok/s vs 22,787.7 tok/s at BS=4, 1k prefill)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "vLLM issue #60674",
   "source_url": "https://github.com/vllm-project/vllm/pull/60674",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": null,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60674-a100-gemma4e4b-pr",
   "model": "Gemma 4 E4B",
   "notes": "vLLM PR #60674 fused Triton kernels (QKV norm/RoPE, pre/post-FFN norms, PLE, MoE routing); 1k prefill batch 4; config=pr-60674; main arm: 22787.7 tok/s",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 34290.8,
   "provenance": "community",
   "quant": "BF16",
   "quote": "A100-SXM4-80GB; gemma-4-E4B-it: Up to 1.50x prefill throughput speedup (34,290.8 tok/s vs 22,787.7 tok/s at BS=4, 1k prefill)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "vLLM issue #60674",
   "source_url": "https://github.com/vllm-project/vllm/pull/60674",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "4x RX 7900 XTX",
   "id": "vllm-60657-4xrx7900xtx-flashnext-unfused",
   "model": "Qwen3.8-Flash-Next",
   "notes": "vLLM PR #60657 fused HyperConnection combine+mix HIP kernels gfx1100, TP4, fp16 KV, no spec decode, 512 output tokens, 2 launches x 5 requests medians; config=unfused; fused arm: 76.6 tok/s (+5.8%)",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "GPTQ W4A16",
   "quote": "Decode, Qwen3.8-Flash-Next, TP4 on 4x RX 7900 XTX, fp16 KV cache, no spec decode, one user: 72.5 -> 76.6 tok/s (+5.8%)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "vLLM issue #60657",
   "source_url": "https://github.com/vllm-project/vllm/pull/60657",
   "tg_tokens": null,
   "tps": 72.5,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "4x RX 7900 XTX",
   "id": "vllm-60657-4xrx7900xtx-flashnext-fused",
   "model": "Qwen3.8-Flash-Next",
   "notes": "vLLM PR #60657 fused HyperConnection combine+mix HIP kernels gfx1100, TP4, fp16 KV, no spec decode, 512 output tokens, 2 launches x 5 requests medians; config=fused; unfused arm: 72.5 tok/s",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "GPTQ W4A16",
   "quote": "Decode, Qwen3.8-Flash-Next, TP4 on 4x RX 7900 XTX, fp16 KV cache, no spec decode, one user: 72.5 -> 76.6 tok/s (+5.8%)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "vLLM issue #60657",
   "source_url": "https://github.com/vllm-project/vllm/pull/60657",
   "tg_tokens": null,
   "tps": 76.6,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "4x GB200",
   "id": "vllm-60627-4xgb200-glm52nvfp4-c1-base",
   "model": "GLM-5.2-NVFP4",
   "notes": "vLLM PR #60627 vectorized SM100 direct DCP Output/LSE consumer, TP4/DCP4/EP4, FlashInfer sparse MLA, FP8 KV, 3072 in / 128 out, 1800 MHz SM clocks; config=base; PR arm: 47.538 tok/s (+2.07%)",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "4x GB200; GLM-5.2-NVFP4, TP4/DCP4/EP4, FlashInfer sparse MLA, FP8 KV, 3072 input tokens, 128 output tokens; c1 | 19.552 ms | 19.109 ms | 46.576 tok/s | 47.538 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "vLLM issue #60627",
   "source_url": "https://github.com/vllm-project/vllm/pull/60627",
   "tg_tokens": null,
   "tps": 46.576,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "4x GB200",
   "id": "vllm-60627-4xgb200-glm52nvfp4-c1-pr",
   "model": "GLM-5.2-NVFP4",
   "notes": "vLLM PR #60627 vectorized SM100 direct DCP Output/LSE consumer, TP4/DCP4/EP4, FlashInfer sparse MLA, FP8 KV, 3072 in / 128 out, 1800 MHz SM clocks; config=pr-60627; base arm: 46.576 tok/s",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "4x GB200; GLM-5.2-NVFP4, TP4/DCP4/EP4, FlashInfer sparse MLA, FP8 KV, 3072 input tokens, 128 output tokens; c1 | 19.552 ms | 19.109 ms | 46.576 tok/s | 47.538 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "vLLM issue #60627",
   "source_url": "https://github.com/vllm-project/vllm/pull/60627",
   "tg_tokens": null,
   "tps": 47.538,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 32,
   "ctx": null,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "4x GB200",
   "id": "vllm-60627-4xgb200-glm52nvfp4-c32-base",
   "model": "GLM-5.2-NVFP4",
   "notes": "vLLM PR #60627 vectorized SM100 direct DCP Output/LSE consumer, TP4/DCP4/EP4, FlashInfer sparse MLA, FP8 KV, 3072 in / 128 out, 1800 MHz SM clocks; config=base; PR arm: 340.987 tok/s (neutral)",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "4x GB200; GLM-5.2-NVFP4, TP4/DCP4/EP4, FlashInfer sparse MLA, FP8 KV, 3072 input tokens, 128 output tokens; c32 | 81.386 ms | 81.386 ms | 340.973 tok/s | 340.987 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "vLLM issue #60627",
   "source_url": "https://github.com/vllm-project/vllm/pull/60627",
   "tg_tokens": null,
   "tps": 340.973,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 32,
   "ctx": null,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "4x GB200",
   "id": "vllm-60627-4xgb200-glm52nvfp4-c32-pr",
   "model": "GLM-5.2-NVFP4",
   "notes": "vLLM PR #60627 vectorized SM100 direct DCP Output/LSE consumer, TP4/DCP4/EP4, FlashInfer sparse MLA, FP8 KV, 3072 in / 128 out, 1800 MHz SM clocks; config=pr-60627; base arm: 340.973 tok/s",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "4x GB200; GLM-5.2-NVFP4, TP4/DCP4/EP4, FlashInfer sparse MLA, FP8 KV, 3072 input tokens, 128 output tokens; c32 | 81.386 ms | 81.386 ms | 340.973 tok/s | 340.987 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "vLLM issue #60627",
   "source_url": "https://github.com/vllm-project/vllm/pull/60627",
   "tg_tokens": null,
   "tps": 340.987,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "GB10",
   "id": "vllm-60615-gb10-flashnext-t0-copy-mtp",
   "model": "Qwen3.8-Flash-Next",
   "notes": "vLLM PR #60615 n-gram lookup in front of MTP drafting, GB10, MTP k=3 stock BF16 head, thinking off, c=1, nightly 81198e97ba + PR + #56273; T=0 greedy copy probe, MTP baseline (2 runs 52.5/52.4); config=mtp-t0-copy; ngram_lookup arm: 67.4 (+28.4%)",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "GB10; Qwen3.8-Flash-Next-NVFP4 + MTP k=3 (stock BF16 head), one DGX Spark (GB10); Single request, T=0: copy probe 52.5 -> 67.4 tok/s (+28%), Blazedit 51.8 -> 60.9 (+17%)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "vLLM issue #60615",
   "source_url": "https://github.com/vllm-project/vllm/pull/60615",
   "tg_tokens": null,
   "tps": 52.5,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "GB10",
   "id": "vllm-60615-gb10-flashnext-t0-copy-ngram",
   "model": "Qwen3.8-Flash-Next",
   "notes": "vLLM PR #60615 n-gram lookup in front of MTP drafting, GB10, MTP k=3 stock BF16 head, thinking off, c=1, nightly 81198e97ba + PR + #56273; T=0 greedy copy probe with ngram_lookup (accept 3.92); config=ngram-t0-copy; MTP baseline: 52.5",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "GB10; Qwen3.8-Flash-Next-NVFP4 + MTP k=3 (stock BF16 head), one DGX Spark (GB10); Single request, T=0: copy probe 52.5 -> 67.4 tok/s (+28%), Blazedit 51.8 -> 60.9 (+17%)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "vLLM issue #60615",
   "source_url": "https://github.com/vllm-project/vllm/pull/60615",
   "tg_tokens": null,
   "tps": 67.4,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "GB10",
   "id": "vllm-60615-gb10-flashnext-t1-copy-mtp",
   "model": "Qwen3.8-Flash-Next",
   "notes": "vLLM PR #60615 n-gram lookup in front of MTP drafting, GB10, MTP k=3 stock BF16 head, thinking off, c=1, nightly 81198e97ba + PR + #56273; T=1 probabilistic copy probe, MTP baseline (2 runs 50.9/52.1); config=mtp-t1-copy; ngram_lookup arm: 66.2 (+28.6%)",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "GB10; Qwen3.8-Flash-Next-NVFP4 + MTP k=3 (stock BF16 head), one DGX Spark (GB10); Single request, T=1, probabilistic drafting: copy probe 51.5 -> 66.2 (+29%), Blazedit 51.1 -> 58.5 (+15%)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "vLLM issue #60615",
   "source_url": "https://github.com/vllm-project/vllm/pull/60615",
   "tg_tokens": null,
   "tps": 51.5,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "GB10",
   "id": "vllm-60615-gb10-flashnext-t1-copy-ngram",
   "model": "Qwen3.8-Flash-Next",
   "notes": "vLLM PR #60615 n-gram lookup in front of MTP drafting, GB10, MTP k=3 stock BF16 head, thinking off, c=1, nightly 81198e97ba + PR + #56273; T=1 probabilistic copy probe with ngram_lookup (accept 3.91); config=ngram-t1-copy; MTP baseline: 51.5",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "GB10; Qwen3.8-Flash-Next-NVFP4 + MTP k=3 (stock BF16 head), one DGX Spark (GB10); Single request, T=1, probabilistic drafting: copy probe 51.5 -> 66.2 (+29%), Blazedit 51.1 -> 58.5 (+15%)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "vLLM issue #60615",
   "source_url": "https://github.com/vllm-project/vllm/pull/60615",
   "tg_tokens": null,
   "tps": 66.2,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "GB10",
   "id": "vllm-60615-gb10-flashnext-t0-blazedit-mtp",
   "model": "Qwen3.8-Flash-Next",
   "notes": "vLLM PR #60615 n-gram lookup in front of MTP drafting, GB10, MTP k=3 stock BF16 head, thinking off, c=1, nightly 81198e97ba + PR + #56273; T=0 greedy Blazedit code editing (16 prompts, 1536 tokens), MTP baseline (2 runs 51.3/52.4); config=mtp-t0-blazedit; ngram_lookup arm: 60.9 (+17.4%)",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "GB10; Qwen3.8-Flash-Next-NVFP4 + MTP k=3 (stock BF16 head), one DGX Spark (GB10); Single request, T=0: copy probe 52.5 -> 67.4 tok/s (+28%), Blazedit 51.8 -> 60.9 (+17%)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "vLLM issue #60615",
   "source_url": "https://github.com/vllm-project/vllm/pull/60615",
   "tg_tokens": null,
   "tps": 51.8,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "GB10",
   "id": "vllm-60615-gb10-flashnext-t0-blazedit-ngram",
   "model": "Qwen3.8-Flash-Next",
   "notes": "vLLM PR #60615 n-gram lookup in front of MTP drafting, GB10, MTP k=3 stock BF16 head, thinking off, c=1, nightly 81198e97ba + PR + #56273; T=0 greedy Blazedit code editing with ngram_lookup (accept 3.73); config=ngram-t0-blazedit; MTP baseline: 51.8",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "GB10; Qwen3.8-Flash-Next-NVFP4 + MTP k=3 (stock BF16 head), one DGX Spark (GB10); Single request, T=0: copy probe 52.5 -> 67.4 tok/s (+28%), Blazedit 51.8 -> 60.9 (+17%)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "vLLM issue #60615",
   "source_url": "https://github.com/vllm-project/vllm/pull/60615",
   "tg_tokens": null,
   "tps": 60.9,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "GB10",
   "id": "vllm-60615-gb10-flashnext-t1-blazedit-mtp",
   "model": "Qwen3.8-Flash-Next",
   "notes": "vLLM PR #60615 n-gram lookup in front of MTP drafting, GB10, MTP k=3 stock BF16 head, thinking off, c=1, nightly 81198e97ba + PR + #56273; T=1 probabilistic Blazedit code editing, MTP baseline (2 runs 50.6/51.6); config=mtp-t1-blazedit; ngram_lookup arm: 58.5 (+14.6%)",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "GB10; Qwen3.8-Flash-Next-NVFP4 + MTP k=3 (stock BF16 head), one DGX Spark (GB10); Single request, T=1, probabilistic drafting: copy probe 51.5 -> 66.2 (+29%), Blazedit 51.1 -> 58.5 (+15%)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "vLLM issue #60615",
   "source_url": "https://github.com/vllm-project/vllm/pull/60615",
   "tg_tokens": null,
   "tps": 51.1,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "GB10",
   "id": "vllm-60615-gb10-flashnext-t1-blazedit-ngram",
   "model": "Qwen3.8-Flash-Next",
   "notes": "vLLM PR #60615 n-gram lookup in front of MTP drafting, GB10, MTP k=3 stock BF16 head, thinking off, c=1, nightly 81198e97ba + PR + #56273; T=1 probabilistic Blazedit code editing with ngram_lookup (accept 3.65); config=ngram-t1-blazedit; MTP baseline: 51.1",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "GB10; Qwen3.8-Flash-Next-NVFP4 + MTP k=3 (stock BF16 head), one DGX Spark (GB10); Single request, T=1, probabilistic drafting: copy probe 51.5 -> 66.2 (+29%), Blazedit 51.1 -> 58.5 (+15%)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "vLLM issue #60615",
   "source_url": "https://github.com/vllm-project/vllm/pull/60615",
   "tg_tokens": null,
   "tps": 58.5,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "GB10",
   "id": "vllm-60615-gb10-flashnext-t0-speed-mtp",
   "model": "Qwen3.8-Flash-Next",
   "notes": "vLLM PR #60615 n-gram lookup in front of MTP drafting, GB10, MTP k=3 stock BF16 head, thinking off, c=1, nightly 81198e97ba + PR + #56273; T=0 greedy SPEED-Bench 5 categories (160 prompts, 256 tokens), MTP baseline (2 runs 41.06/41.29); config=mtp-t0-speed; ngram_lookup arm: 42.84 (+4.0%)",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "GB10; Qwen3.8-Flash-Next-NVFP4 + MTP k=3 (stock BF16 head), one DGX Spark (GB10); SPEED-Bench, 5 categories | 41.06 / 41.29 (2.77) | 42.84 (2.79), +4.0% | 40.76 / 41.07 (2.78) | 42.07 (2.79), +2.8%",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "vLLM issue #60615",
   "source_url": "https://github.com/vllm-project/vllm/pull/60615",
   "tg_tokens": null,
   "tps": 41.06,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "GB10",
   "id": "vllm-60615-gb10-flashnext-t0-speed-ngram",
   "model": "Qwen3.8-Flash-Next",
   "notes": "vLLM PR #60615 n-gram lookup in front of MTP drafting, GB10, MTP k=3 stock BF16 head, thinking off, c=1, nightly 81198e97ba + PR + #56273; T=0 greedy SPEED-Bench 5 categories with ngram_lookup; config=ngram-t0-speed; MTP baseline: 41.06",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "GB10; Qwen3.8-Flash-Next-NVFP4 + MTP k=3 (stock BF16 head), one DGX Spark (GB10); SPEED-Bench, 5 categories | 41.06 / 41.29 (2.77) | 42.84 (2.79), +4.0% | 40.76 / 41.07 (2.78) | 42.07 (2.79), +2.8%",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "vLLM issue #60615",
   "source_url": "https://github.com/vllm-project/vllm/pull/60615",
   "tg_tokens": null,
   "tps": 42.84,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "GB10",
   "id": "vllm-60615-gb10-flashnext-t1-speed-mtp",
   "model": "Qwen3.8-Flash-Next",
   "notes": "vLLM PR #60615 n-gram lookup in front of MTP drafting, GB10, MTP k=3 stock BF16 head, thinking off, c=1, nightly 81198e97ba + PR + #56273; T=1 probabilistic SPEED-Bench 5 categories, MTP baseline (2 runs 40.76/41.07); config=mtp-t1-speed; ngram_lookup arm: 42.07 (+2.8%)",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "GB10; Qwen3.8-Flash-Next-NVFP4 + MTP k=3 (stock BF16 head), one DGX Spark (GB10); SPEED-Bench, 5 categories | 41.06 / 41.29 (2.77) | 42.84 (2.79), +4.0% | 40.76 / 41.07 (2.78) | 42.07 (2.79), +2.8%",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "vLLM issue #60615",
   "source_url": "https://github.com/vllm-project/vllm/pull/60615",
   "tg_tokens": null,
   "tps": 40.76,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "GB10",
   "id": "vllm-60615-gb10-flashnext-t1-speed-ngram",
   "model": "Qwen3.8-Flash-Next",
   "notes": "vLLM PR #60615 n-gram lookup in front of MTP drafting, GB10, MTP k=3 stock BF16 head, thinking off, c=1, nightly 81198e97ba + PR + #56273; T=1 probabilistic SPEED-Bench 5 categories with ngram_lookup; config=ngram-t1-speed; MTP baseline: 40.76",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "GB10; Qwen3.8-Flash-Next-NVFP4 + MTP k=3 (stock BF16 head), one DGX Spark (GB10); SPEED-Bench, 5 categories | 41.06 / 41.29 (2.77) | 42.84 (2.79), +4.0% | 40.76 / 41.07 (2.78) | 42.07 (2.79), +2.8%",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "vLLM issue #60615",
   "source_url": "https://github.com/vllm-project/vllm/pull/60615",
   "tg_tokens": null,
   "tps": 42.07,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "8\u00d7 B200",
   "id": "vllm-60570-8xb200-nemotron3super-c1-control",
   "model": "Nemotron-3-Super-120B-A12B",
   "notes": "vLLM PR #60570 persistent FlashInfer CUTLASS MoE workspace, 8x B200 178 GB node, TP1/DP8/EP, AIPerf 8000 in / 1000 out; concurrency 1, unpatched control (patched runs 128.07/128.39); config=control",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "8\u00d7 B200; nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4; Output tok/s, concurrency 1 | 127.98 | 128.07 | 128.39",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "vLLM issue #60570",
   "source_url": "https://github.com/vllm-project/vllm/pull/60570",
   "tg_tokens": null,
   "tps": 127.98,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "8\u00d7 B200",
   "id": "vllm-60570-8xb200-nemotron3super-c1-patched",
   "model": "Nemotron-3-Super-120B-A12B",
   "notes": "vLLM PR #60570 persistent FlashInfer CUTLASS MoE workspace, 8x B200 178 GB node, TP1/DP8/EP, AIPerf 8000 in / 1000 out; concurrency 1, patched run 1 (run 2: 128.39; control 127.98, neutral); config=patched",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "8\u00d7 B200; nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4; Output tok/s, concurrency 1 | 127.98 | 128.07 | 128.39",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "vLLM issue #60570",
   "source_url": "https://github.com/vllm-project/vllm/pull/60570",
   "tg_tokens": null,
   "tps": 128.07,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 64,
   "ctx": null,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "8\u00d7 B200",
   "id": "vllm-60570-8xb200-nemotron3super-c64-patched",
   "model": "Nemotron-3-Super-120B-A12B",
   "notes": "vLLM PR #60570 persistent FlashInfer CUTLASS MoE workspace, 8x B200 178 GB node, TP1/DP8/EP, AIPerf 8000 in / 1000 out; concurrency 64, patched run 1 (run 2: 5127.56; unpatched control OOMed at c64, no baseline); config=patched",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "8\u00d7 B200; nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4; Output tok/s, concurrency 64 | Not completed | 5,274.97 | 5,127.56",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "vLLM issue #60570",
   "source_url": "https://github.com/vllm-project/vllm/pull/60570",
   "tg_tokens": null,
   "tps": 5274.97,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 256,
   "ctx": null,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "8\u00d7 B200",
   "id": "vllm-60570-8xb200-nemotron3super-c256-patched",
   "model": "Nemotron-3-Super-120B-A12B",
   "notes": "vLLM PR #60570 persistent FlashInfer CUTLASS MoE workspace, 8x B200 178 GB node, TP1/DP8/EP, AIPerf 8000 in / 1000 out; concurrency 256, patched run 1 (run 2: 12638.27; control not run); config=patched",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "8\u00d7 B200; nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4; Output tok/s, concurrency 256 | Not run | 12,686.38 | 12,638.27",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "vLLM issue #60570",
   "source_url": "https://github.com/vllm-project/vllm/pull/60570",
   "tg_tokens": null,
   "tps": 12686.38,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": null,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "H100",
   "id": "vllm-60500-h100-nemotronparse-c8-on",
   "model": "Nemotron-Parse",
   "notes": "vLLM PR #60500 image rescale/normalize on device (FusedMMInputNorm), one H100, bfloat16, TRITON_ATTN, document page parsing, c=8 aggregate output tok/s, 3 interleaved reps fresh server each; device normalize on; config=on; off arm: 772.7 tok/s (+21%)",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "One H100; Nemotron-Parse; on | 132.5 ms | 131.8 ms | 937.9 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "vLLM issue #60500",
   "source_url": "https://github.com/vllm-project/vllm/pull/60500",
   "tg_tokens": null,
   "tps": 937.9,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": null,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "H100",
   "id": "vllm-60500-h100-nemotronparse-c8-off",
   "model": "Nemotron-Parse",
   "notes": "vLLM PR #60500 image rescale/normalize on device (FusedMMInputNorm), one H100, bfloat16, TRITON_ATTN, document page parsing, c=8 aggregate output tok/s, 3 interleaved reps fresh server each; device normalize off (CPU path); config=off; on arm: 937.9 tok/s",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "One H100; Nemotron-Parse; off | 263.1 ms | 261.5 ms | 772.7 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "vLLM issue #60500",
   "source_url": "https://github.com/vllm-project/vllm/pull/60500",
   "tg_tokens": null,
   "tps": 772.7,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": null,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60504-a100-gemma426ba4b-bs4-base",
   "model": "gemma-4-26B-A4B-it",
   "notes": "vLLM PR #60504 tuned Triton FusedMoE configs E=128 top-8, A100-SXM4-80GB, enforce_eager=False, max_model_len=8192, decode-heavy campaign batch 4; untuned default configs; config=main; tuned arm: 5493.9 tok/s (1.20x)",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "A100-SXM4-80GB; Gemma 4 26B-A4B-it; decode-heavy (up to 1.20x speedup, 118.7 ms -> 99.0 ms, 5493.9 vs 4582.2 tok/s at batch size 4)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "vLLM issue #60504",
   "source_url": "https://github.com/vllm-project/vllm/pull/60504",
   "tg_tokens": null,
   "tps": 4582.2,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": null,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "A100-SXM4-80GB",
   "id": "vllm-60504-a100-gemma426ba4b-bs4-tuned",
   "model": "gemma-4-26B-A4B-it",
   "notes": "vLLM PR #60504 tuned Triton FusedMoE configs E=128 top-8, A100-SXM4-80GB, enforce_eager=False, max_model_len=8192, decode-heavy campaign batch 4; tuned FusedMoE configs; config=pr-60504; main arm: 4582.2 tok/s",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "A100-SXM4-80GB; Gemma 4 26B-A4B-it; decode-heavy (up to 1.20x speedup, 118.7 ms -> 99.0 ms, 5493.9 vs 4582.2 tok/s at batch size 4)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "vLLM issue #60504",
   "source_url": "https://github.com/vllm-project/vllm/pull/60504",
   "tg_tokens": null,
   "tps": 5493.9,
   "ttft_s": null
  },
  {
   "backend": "NInfer",
   "batch": 8,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "hf-ninfer-coder390-q2lynn-ninfer-c8",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO",
   "notes": "Coder390 GGUF-NInfer card bench, one RTX PRO 6000 Blackwell, 1-minute quick benchmark, 512 generation cap, thinking on (xhigh), temp 1.0 top_p 0.95 top_k 20, aggregate tok/s; NInfer server --max-context 32768 --kv-capacity 32768 --kv-dtype rk8v4; Q2 LynnStyle on NInfer, C8 aggregate, built-in Q4 MTP, 512-token cap short run; llama.cpp arm: 174.8 tok/s (protocol differs); config=ninfer-c8",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q2 LynnStyle",
   "quote": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer; RTX PRO 6000 Blackwell; GGUF Q2 LynnStyle speed: NInfer with built-in Q4 MTP, C8 aggregate 521.0 tok/s (512-token cap, short run); llama.cpp with built-in Q4 MTP, C4 aggregate 174.8 tok/s (4096-token cap, long run)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "HF model card: nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer",
   "source_url": "https://huggingface.co/nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer",
   "tg_tokens": null,
   "tps": "521.0",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 4,
   "ctx": null,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "hf-ninfer-coder390-q2lynn-llamacpp-c4",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO",
   "notes": "Coder390 GGUF-NInfer card bench, one RTX PRO 6000 Blackwell, 1-minute quick benchmark, 512 generation cap, thinking on (xhigh), temp 1.0 top_p 0.95 top_k 20, aggregate tok/s; NInfer server --max-context 32768 --kv-capacity 32768 --kv-dtype rk8v4; Q2 LynnStyle on llama.cpp, C4 aggregate, built-in Q4 MTP, 4096-token cap long run; NInfer arm: 521.0 tok/s (protocol differs); config=llamacpp-c4",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q2 LynnStyle",
   "quote": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer; RTX PRO 6000 Blackwell; GGUF Q2 LynnStyle speed: NInfer with built-in Q4 MTP, C8 aggregate 521.0 tok/s (512-token cap, short run); llama.cpp with built-in Q4 MTP, C4 aggregate 174.8 tok/s (4096-token cap, long run)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "HF model card: nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer",
   "source_url": "https://huggingface.co/nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer",
   "tg_tokens": null,
   "tps": "174.8",
   "ttft_s": null
  },
  {
   "backend": "NInfer",
   "batch": 8,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "hf-ninfer-coder390-q3lynn-ninfer-c8",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO",
   "notes": "Coder390 GGUF-NInfer card bench, one RTX PRO 6000 Blackwell, 1-minute quick benchmark, 512 generation cap, thinking on (xhigh), temp 1.0 top_p 0.95 top_k 20, aggregate tok/s; NInfer server --max-context 32768 --kv-capacity 32768 --kv-dtype rk8v4; Q3 LynnStyle on NInfer, C8 aggregate, built-in Q4 MTP, 512-token cap short run; llama.cpp arm: 191.8 tok/s (protocol differs); config=ninfer-c8",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q3 LynnStyle",
   "quote": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer; RTX PRO 6000 Blackwell; GGUF Q3 LynnStyle speed: NInfer with built-in Q4 MTP, C8 aggregate 512.3 tok/s (512-token cap, short run); llama.cpp with built-in Q4 MTP, C4 aggregate 191.8 tok/s (4096-token cap, long run)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "HF model card: nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer",
   "source_url": "https://huggingface.co/nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer",
   "tg_tokens": null,
   "tps": "512.3",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 4,
   "ctx": null,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "hf-ninfer-coder390-q3lynn-llamacpp-c4",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO",
   "notes": "Coder390 GGUF-NInfer card bench, one RTX PRO 6000 Blackwell, 1-minute quick benchmark, 512 generation cap, thinking on (xhigh), temp 1.0 top_p 0.95 top_k 20, aggregate tok/s; NInfer server --max-context 32768 --kv-capacity 32768 --kv-dtype rk8v4; Q3 LynnStyle on llama.cpp, C4 aggregate, built-in Q4 MTP, 4096-token cap long run; NInfer arm: 512.3 tok/s (protocol differs); config=llamacpp-c4",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q3 LynnStyle",
   "quote": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer; RTX PRO 6000 Blackwell; GGUF Q3 LynnStyle speed: NInfer with built-in Q4 MTP, C8 aggregate 512.3 tok/s (512-token cap, short run); llama.cpp with built-in Q4 MTP, C4 aggregate 191.8 tok/s (4096-token cap, long run)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "HF model card: nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer",
   "source_url": "https://huggingface.co/nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer",
   "tg_tokens": null,
   "tps": "191.8",
   "ttft_s": null
  },
  {
   "backend": "NInfer",
   "batch": 8,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "hf-ninfer-coder390-q4lynn-ninfer-c8",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO",
   "notes": "Coder390 GGUF-NInfer card bench, one RTX PRO 6000 Blackwell, 1-minute quick benchmark, 512 generation cap, thinking on (xhigh), temp 1.0 top_p 0.95 top_k 20, aggregate tok/s; NInfer server --max-context 32768 --kv-capacity 32768 --kv-dtype rk8v4; Q4 LynnStyle on NInfer, C8 aggregate, built-in Q4 MTP, 512-token cap short run; llama.cpp arm: 81.1 tok/s (protocol differs); config=ninfer-c8",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4 LynnStyle",
   "quote": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer; RTX PRO 6000 Blackwell; GGUF Q4 LynnStyle speed: NInfer with built-in Q4 MTP, C8 aggregate 436.4 tok/s (512-token cap, short run); llama.cpp with built-in Q4 MTP, C4 aggregate 81.1 tok/s (4096-token cap, long run)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "HF model card: nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer",
   "source_url": "https://huggingface.co/nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer",
   "tg_tokens": null,
   "tps": "436.4",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 4,
   "ctx": null,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "hf-ninfer-coder390-q4lynn-llamacpp-c4",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO",
   "notes": "Coder390 GGUF-NInfer card bench, one RTX PRO 6000 Blackwell, 1-minute quick benchmark, 512 generation cap, thinking on (xhigh), temp 1.0 top_p 0.95 top_k 20, aggregate tok/s; NInfer server --max-context 32768 --kv-capacity 32768 --kv-dtype rk8v4; Q4 LynnStyle on llama.cpp, C4 aggregate, built-in Q4 MTP, 4096-token cap long run; NInfer arm: 436.4 tok/s (protocol differs); config=llamacpp-c4",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4 LynnStyle",
   "quote": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer; RTX PRO 6000 Blackwell; GGUF Q4 LynnStyle speed: NInfer with built-in Q4 MTP, C8 aggregate 436.4 tok/s (512-token cap, short run); llama.cpp with built-in Q4 MTP, C4 aggregate 81.1 tok/s (4096-token cap, long run)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "HF model card: nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer",
   "source_url": "https://huggingface.co/nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer",
   "tg_tokens": null,
   "tps": "81.1",
   "ttft_s": null
  },
  {
   "backend": "NInfer",
   "batch": 4,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "hf-ninfer-coder390-q6k-ninfer-c4",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO",
   "notes": "Coder390 GGUF-NInfer card bench, one RTX PRO 6000 Blackwell, 1-minute quick benchmark, 512 generation cap, thinking on (xhigh), temp 1.0 top_p 0.95 top_k 20, aggregate tok/s; NInfer server --max-context 32768 --kv-capacity 32768 --kv-dtype rk8v4; Q6_K on NInfer, C4 aggregate, MTP, 512-token cap quick run; llama.cpp arm: 189.0 tok/s (protocol differs); config=ninfer-c4",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer; RTX PRO 6000 Blackwell; GGUF Q6_K speed: NInfer with MTP, C4 aggregate 235.5 tok/s; llama.cpp with MTP, C8 aggregate 189.0 tok/s (both 512-token-cap quick runs)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "HF model card: nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer",
   "source_url": "https://huggingface.co/nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer",
   "tg_tokens": null,
   "tps": "235.5",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 8,
   "ctx": null,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "hf-ninfer-coder390-q6k-llamacpp-c8",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO",
   "notes": "Coder390 GGUF-NInfer card bench, one RTX PRO 6000 Blackwell, 1-minute quick benchmark, 512 generation cap, thinking on (xhigh), temp 1.0 top_p 0.95 top_k 20, aggregate tok/s; NInfer server --max-context 32768 --kv-capacity 32768 --kv-dtype rk8v4; Q6_K on llama.cpp, C8 aggregate, MTP, 512-token cap quick run; NInfer arm: 235.5 tok/s (protocol differs); config=llamacpp-c8",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer; RTX PRO 6000 Blackwell; GGUF Q6_K speed: NInfer with MTP, C4 aggregate 235.5 tok/s; llama.cpp with MTP, C8 aggregate 189.0 tok/s (both 512-token-cap quick runs)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "HF model card: nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer",
   "source_url": "https://huggingface.co/nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer",
   "tg_tokens": null,
   "tps": "189.0",
   "ttft_s": null
  },
  {
   "backend": "NInfer",
   "batch": 8,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "hf-ninfer-coder390-q80-ninfer-c8",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO",
   "notes": "Coder390 GGUF-NInfer card bench, one RTX PRO 6000 Blackwell, 1-minute quick benchmark, 512 generation cap, thinking on (xhigh), temp 1.0 top_p 0.95 top_k 20, aggregate tok/s; NInfer server --max-context 32768 --kv-capacity 32768 --kv-dtype rk8v4; Q8_0 on NInfer, C8 aggregate, MTP, same-caliber bench; llama.cpp arm: 190.8 tok/s (protocol differs); config=ninfer-c8",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer; RTX PRO 6000 Blackwell; GGUF Q8_0 speed: NInfer with MTP, C8 aggregate 486.3 tok/s (same-caliber bench); llama.cpp with MTP, C4 aggregate 190.8 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "HF model card: nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer",
   "source_url": "https://huggingface.co/nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer",
   "tg_tokens": null,
   "tps": "486.3",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 4,
   "ctx": null,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "hf-ninfer-coder390-q80-llamacpp-c4",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO",
   "notes": "Coder390 GGUF-NInfer card bench, one RTX PRO 6000 Blackwell, 1-minute quick benchmark, 512 generation cap, thinking on (xhigh), temp 1.0 top_p 0.95 top_k 20, aggregate tok/s; NInfer server --max-context 32768 --kv-capacity 32768 --kv-dtype rk8v4; Q8_0 on llama.cpp, C4 aggregate, MTP; NInfer arm: 486.3 tok/s (protocol differs); config=llamacpp-c4",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer; RTX PRO 6000 Blackwell; GGUF Q8_0 speed: NInfer with MTP, C8 aggregate 486.3 tok/s (same-caliber bench); llama.cpp with MTP, C4 aggregate 190.8 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "HF model card: nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer",
   "source_url": "https://huggingface.co/nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer",
   "tg_tokens": null,
   "tps": "190.8",
   "ttft_s": null
  },
  {
   "backend": "NInfer",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "hf-ninfer-coder390-q6k-ninfer-c1",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO",
   "notes": "Coder390 GGUF-NInfer card bench, one RTX PRO 6000 Blackwell, 1-minute quick benchmark, 512 generation cap, thinking on (xhigh), temp 1.0 top_p 0.95 top_k 20, aggregate tok/s; NInfer server --max-context 32768 --kv-capacity 32768 --kv-dtype rk8v4; single request with MTP (card context: Q6_K recommended config); config=ninfer-c1",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer; RTX PRO 6000 Blackwell; For Q6_K at a single request, NInfer with MTP reaches 131.9 tok/s, about 2.5\u00d7 llama.cpp without speculation (53.4 tok/s); llama.cpp with MTP reaches 97.2 tok/s aggregate at C1 (119.0 tok/s single-request decode)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "HF model card: nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer",
   "source_url": "https://huggingface.co/nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer",
   "tg_tokens": null,
   "tps": "131.9",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "hf-ninfer-coder390-q6k-llamacpp-nospec-c1",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO",
   "notes": "Coder390 GGUF-NInfer card bench, one RTX PRO 6000 Blackwell, 1-minute quick benchmark, 512 generation cap, thinking on (xhigh), temp 1.0 top_p 0.95 top_k 20, aggregate tok/s; NInfer server --max-context 32768 --kv-capacity 32768 --kv-dtype rk8v4; single request without speculation; config=llamacpp-nospec-c1",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer; RTX PRO 6000 Blackwell; For Q6_K at a single request, NInfer with MTP reaches 131.9 tok/s, about 2.5\u00d7 llama.cpp without speculation (53.4 tok/s); llama.cpp with MTP reaches 97.2 tok/s aggregate at C1 (119.0 tok/s single-request decode)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "HF model card: nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer",
   "source_url": "https://huggingface.co/nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer",
   "tg_tokens": null,
   "tps": "53.4",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "hf-ninfer-coder390-q6k-llamacpp-mtp-c1agg",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO",
   "notes": "Coder390 GGUF-NInfer card bench, one RTX PRO 6000 Blackwell, 1-minute quick benchmark, 512 generation cap, thinking on (xhigh), temp 1.0 top_p 0.95 top_k 20, aggregate tok/s; NInfer server --max-context 32768 --kv-capacity 32768 --kv-dtype rk8v4; C1 aggregate with MTP (single-request decode 119.0); config=llamacpp-mtp-c1-agg",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer; RTX PRO 6000 Blackwell; For Q6_K at a single request, NInfer with MTP reaches 131.9 tok/s, about 2.5\u00d7 llama.cpp without speculation (53.4 tok/s); llama.cpp with MTP reaches 97.2 tok/s aggregate at C1 (119.0 tok/s single-request decode)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "HF model card: nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer",
   "source_url": "https://huggingface.co/nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer",
   "tg_tokens": null,
   "tps": "97.2",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "hf-ninfer-coder390-q6k-llamacpp-mtp-c1single",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO",
   "notes": "Coder390 GGUF-NInfer card bench, one RTX PRO 6000 Blackwell, 1-minute quick benchmark, 512 generation cap, thinking on (xhigh), temp 1.0 top_p 0.95 top_k 20, aggregate tok/s; NInfer server --max-context 32768 --kv-capacity 32768 --kv-dtype rk8v4; single-request decode with MTP at C1; config=llamacpp-mtp-c1-single",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer; RTX PRO 6000 Blackwell; For Q6_K at a single request, NInfer with MTP reaches 131.9 tok/s, about 2.5\u00d7 llama.cpp without speculation (53.4 tok/s); llama.cpp with MTP reaches 97.2 tok/s aggregate at C1 (119.0 tok/s single-request decode)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "HF model card: nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer",
   "source_url": "https://huggingface.co/nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer",
   "tg_tokens": null,
   "tps": "119.0",
   "ttft_s": null
  },
  {
   "backend": "NInfer",
   "batch": 8,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "hf-ninfer-coder390-q6k-ninfer-c8",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO",
   "notes": "Coder390 GGUF-NInfer card bench, one RTX PRO 6000 Blackwell, 1-minute quick benchmark, 512 generation cap, thinking on (xhigh), temp 1.0 top_p 0.95 top_k 20, aggregate tok/s; NInfer server --max-context 32768 --kv-capacity 32768 --kv-dtype rk8v4; Q6_K NInfer C8 aggregate (recommended single-request config is C4 at 235.5); config=ninfer-c8",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer; RTX PRO 6000 Blackwell; for aggregate throughput alone, Q6_K at C8 is higher (388.1 tok/s)",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "HF model card: nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer",
   "source_url": "https://huggingface.co/nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer",
   "tg_tokens": null,
   "tps": "388.1",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "hf-ninfer-coder390-q3lynn-md-q8head-greedy",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO",
   "notes": "Coder390 card, llama.cpp Q3 LynnStyle target with external Q8_0 MTP draft (-md) vs built-in Q4 MTP head, --spec-draft-n-max 4, one RTX PRO 6000 Blackwell, 12 prompts, card states checked 2026-10-09; greedy, external Q8_0 MTP draft, acceptance 66.1%; built-in Q4 head: 112.2 tok/s (66.0%); config=md-q8-greedy",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q3 LynnStyle",
   "quote": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer; RTX PRO 6000 Blackwell; Greedy acceptance is 66.1% at 110.2 tok/s, against 66.0% at 112.2 tok/s for the built-in Q4 MTP head with the same command and no -md; At temperature 1.0 the two are 52.4% at 94.4 tok/s and 51.9% at 95.6 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "HF model card: nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer",
   "source_url": "https://huggingface.co/nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer",
   "tg_tokens": null,
   "tps": "110.2",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "hf-ninfer-coder390-q3lynn-builtin-q4-greedy",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO",
   "notes": "Coder390 card, llama.cpp Q3 LynnStyle target with external Q8_0 MTP draft (-md) vs built-in Q4 MTP head, --spec-draft-n-max 4, one RTX PRO 6000 Blackwell, 12 prompts, card states checked 2026-10-09; greedy, built-in Q4 MTP head, acceptance 66.0%; external Q8_0 draft: 110.2 tok/s (66.1%); config=builtin-q4-greedy",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q3 LynnStyle",
   "quote": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer; RTX PRO 6000 Blackwell; Greedy acceptance is 66.1% at 110.2 tok/s, against 66.0% at 112.2 tok/s for the built-in Q4 MTP head with the same command and no -md; At temperature 1.0 the two are 52.4% at 94.4 tok/s and 51.9% at 95.6 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "HF model card: nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer",
   "source_url": "https://huggingface.co/nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer",
   "tg_tokens": null,
   "tps": "112.2",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "hf-ninfer-coder390-q3lynn-md-q8head-t1",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO",
   "notes": "Coder390 card, llama.cpp Q3 LynnStyle target with external Q8_0 MTP draft (-md) vs built-in Q4 MTP head, --spec-draft-n-max 4, one RTX PRO 6000 Blackwell, 12 prompts, card states checked 2026-10-09; temperature 1.0, external Q8_0 MTP draft, acceptance 52.4%; built-in Q4 head: 95.6 tok/s (51.9%); config=md-q8-t1",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q3 LynnStyle",
   "quote": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer; RTX PRO 6000 Blackwell; Greedy acceptance is 66.1% at 110.2 tok/s, against 66.0% at 112.2 tok/s for the built-in Q4 MTP head with the same command and no -md; At temperature 1.0 the two are 52.4% at 94.4 tok/s and 51.9% at 95.6 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "HF model card: nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer",
   "source_url": "https://huggingface.co/nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer",
   "tg_tokens": null,
   "tps": "94.4",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "hf-ninfer-coder390-q3lynn-builtin-q4-t1",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO",
   "notes": "Coder390 card, llama.cpp Q3 LynnStyle target with external Q8_0 MTP draft (-md) vs built-in Q4 MTP head, --spec-draft-n-max 4, one RTX PRO 6000 Blackwell, 12 prompts, card states checked 2026-10-09; temperature 1.0, built-in Q4 MTP head, acceptance 51.9%; external Q8_0 draft: 94.4 tok/s (52.4%); config=builtin-q4-t1",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q3 LynnStyle",
   "quote": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer; RTX PRO 6000 Blackwell; Greedy acceptance is 66.1% at 110.2 tok/s, against 66.0% at 112.2 tok/s for the built-in Q4 MTP head with the same command and no -md; At temperature 1.0 the two are 52.4% at 94.4 tok/s and 51.9% at 95.6 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "HF model card: nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer",
   "source_url": "https://huggingface.co/nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer",
   "tg_tokens": null,
   "tps": "95.6",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 262144,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "RTX 4080 SUPER",
   "id": "hf-welp-welp35b-tg128",
   "model": "Welp-35B-A3B",
   "notes": "Welp-35B-A3B card, one RTX 4080 Super 16 GB, same harness single runs, llama-server -ngl 99 -fa on -c 262144 -ctk q4_0 -ctv q4_0 -ub 256; decode llama-bench tg128, prefill llama-bench pp4096; Welp-35B-A3B 13.21 GB decode tg128 at 262144 ctx (largest that fits with q4_0 KV); config=tg128",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "13.21 GB",
   "quote": "Welp-35B-A3B; RTX 4080 Super 16 GB; Welp-35B-A3B | 13.21 GB | 95.7% | 23/37 | 262k | 164 | 5,111",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "HF model card: quaedra/Welp-35B-A3B-GGUF",
   "source_url": "https://huggingface.co/quaedra/Welp-35B-A3B-GGUF",
   "tg_tokens": 128,
   "tps": "164",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 262144,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "RTX 4080 SUPER",
   "id": "hf-welp-welp35b-pp4096",
   "model": "Welp-35B-A3B",
   "notes": "Welp-35B-A3B card, one RTX 4080 Super 16 GB, same harness single runs, llama-server -ngl 99 -fa on -c 262144 -ctk q4_0 -ctv q4_0 -ub 256; decode llama-bench tg128, prefill llama-bench pp4096; Welp-35B-A3B 13.21 GB prefill pp4096 at 262144 ctx; config=pp4096",
   "params": null,
   "power_w": null,
   "pp_tokens": 4096,
   "pp_tps": "5111",
   "provenance": "community",
   "quant": "13.21 GB",
   "quote": "Welp-35B-A3B; RTX 4080 Super 16 GB; Welp-35B-A3B | 13.21 GB | 95.7% | 23/37 | 262k | 164 | 5,111",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "HF model card: quaedra/Welp-35B-A3B-GGUF",
   "source_url": "https://huggingface.co/quaedra/Welp-35B-A3B-GGUF",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 262144,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "RTX 4080 SUPER",
   "id": "hf-welp-welp35b-udiq3xxs-tg128",
   "model": "Welp-35B-A3B",
   "notes": "Welp-35B-A3B card, one RTX 4080 Super 16 GB, same harness single runs, llama-server -ngl 99 -fa on -c 262144 -ctk q4_0 -ctv q4_0 -ub 256; decode llama-bench tg128, prefill llama-bench pp4096; Welp-35B-A3B UD-IQ3_XXS decode tg128 at 262144 ctx (largest that fits with q4_0 KV); config=tg128",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ3_XXS",
   "quote": "Welp-35B-A3B; RTX 4080 Super 16 GB; Unsloth UD-IQ3_XXS | 13.21 GB | 93.9% | 21/37 | 262k | 165 | 5,165",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "HF model card: quaedra/Welp-35B-A3B-GGUF",
   "source_url": "https://huggingface.co/quaedra/Welp-35B-A3B-GGUF",
   "tg_tokens": 128,
   "tps": "165",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 262144,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "RTX 4080 SUPER",
   "id": "hf-welp-welp35b-udiq3xxs-pp4096",
   "model": "Welp-35B-A3B",
   "notes": "Welp-35B-A3B card, one RTX 4080 Super 16 GB, same harness single runs, llama-server -ngl 99 -fa on -c 262144 -ctk q4_0 -ctv q4_0 -ub 256; decode llama-bench tg128, prefill llama-bench pp4096; Welp-35B-A3B UD-IQ3_XXS prefill pp4096 at 262144 ctx; config=pp4096",
   "params": null,
   "power_w": null,
   "pp_tokens": 4096,
   "pp_tps": "5165",
   "provenance": "community",
   "quant": "UD-IQ3_XXS",
   "quote": "Welp-35B-A3B; RTX 4080 Super 16 GB; Unsloth UD-IQ3_XXS | 13.21 GB | 93.9% | 21/37 | 262k | 165 | 5,165",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "HF model card: quaedra/Welp-35B-A3B-GGUF",
   "source_url": "https://huggingface.co/quaedra/Welp-35B-A3B-GGUF",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "RTX 4080 SUPER",
   "id": "hf-welp-qwen3827b-udq3kxl-tg128",
   "model": "Qwen3.8-27B",
   "notes": "Welp-35B-A3B card, one RTX 4080 Super 16 GB, same harness single runs, llama-server -ngl 99 -fa on -c 262144 -ctk q4_0 -ctv q4_0 -ub 256; decode llama-bench tg128, prefill llama-bench pp4096; Qwen3.8-27B UD-Q3_K_XL decode tg128 at 131072 ctx (largest that fits with q4_0 KV); config=tg128",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-Q3_K_XL",
   "quote": "Welp-35B-A3B; RTX 4080 Super 16 GB; Qwen3.8-27B (Unsloth UD-Q3_K_XL) | 13.15 GB | 96.3% | 26/37 | 131k | 47 | 2,155",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "HF model card: quaedra/Welp-35B-A3B-GGUF",
   "source_url": "https://huggingface.co/quaedra/Welp-35B-A3B-GGUF",
   "tg_tokens": 128,
   "tps": "47",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "RTX 4080 SUPER",
   "id": "hf-welp-qwen3827b-udq3kxl-pp4096",
   "model": "Qwen3.8-27B",
   "notes": "Welp-35B-A3B card, one RTX 4080 Super 16 GB, same harness single runs, llama-server -ngl 99 -fa on -c 262144 -ctk q4_0 -ctv q4_0 -ub 256; decode llama-bench tg128, prefill llama-bench pp4096; Qwen3.8-27B UD-Q3_K_XL prefill pp4096 at 131072 ctx; config=pp4096",
   "params": null,
   "power_w": null,
   "pp_tokens": 4096,
   "pp_tps": "2155",
   "provenance": "community",
   "quant": "UD-Q3_K_XL",
   "quote": "Welp-35B-A3B; RTX 4080 Super 16 GB; Qwen3.8-27B (Unsloth UD-Q3_K_XL) | 13.15 GB | 96.3% | 26/37 | 131k | 47 | 2,155",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "HF model card: quaedra/Welp-35B-A3B-GGUF",
   "source_url": "https://huggingface.co/quaedra/Welp-35B-A3B-GGUF",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 262144,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "RTX 4080 SUPER",
   "id": "hf-welp-bonsai227b-tg128",
   "model": "Bonsai 2 27B",
   "notes": "Welp-35B-A3B card, one RTX 4080 Super 16 GB, same harness single runs, llama-server -ngl 99 -fa on -c 262144 -ctk q4_0 -ctv q4_0 -ub 256; decode llama-bench tg128, prefill llama-bench pp4096; Bonsai 2 27B 5.95 GB decode tg128 at 262144 ctx (largest that fits with q4_0 KV); config=tg128",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "5.95 GB",
   "quote": "Welp-35B-A3B; RTX 4080 Super 16 GB; Bonsai 2 27B | 5.95 GB | 91.5% | 15/37 | 262k | 85 | 2,278",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "HF model card: quaedra/Welp-35B-A3B-GGUF",
   "source_url": "https://huggingface.co/quaedra/Welp-35B-A3B-GGUF",
   "tg_tokens": 128,
   "tps": "85",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 262144,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "RTX 4080 SUPER",
   "id": "hf-welp-bonsai227b-pp4096",
   "model": "Bonsai 2 27B",
   "notes": "Welp-35B-A3B card, one RTX 4080 Super 16 GB, same harness single runs, llama-server -ngl 99 -fa on -c 262144 -ctk q4_0 -ctv q4_0 -ub 256; decode llama-bench tg128, prefill llama-bench pp4096; Bonsai 2 27B 5.95 GB prefill pp4096 at 262144 ctx; config=pp4096",
   "params": null,
   "power_w": null,
   "pp_tokens": 4096,
   "pp_tps": "2278",
   "provenance": "community",
   "quant": "5.95 GB",
   "quote": "Welp-35B-A3B; RTX 4080 Super 16 GB; Bonsai 2 27B | 5.95 GB | 91.5% | 15/37 | 262k | 85 | 2,278",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "HF model card: quaedra/Welp-35B-A3B-GGUF",
   "source_url": "https://huggingface.co/quaedra/Welp-35B-A3B-GGUF",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 262144,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "RTX 4080 SUPER",
   "id": "hf-welp-welp35b-262k-deep-tg",
   "model": "Welp-35B-A3B",
   "notes": "Welp-35B-A3B card, one RTX 4080 Super 16 GB, same harness single runs, llama-server -ngl 99 -fa on -c 262144 -ctk q4_0 -ctv q4_0 -ub 256; decode llama-bench tg128, prefill llama-bench pp4096; decode at full 262k ctx after a 214k-token prompt, q4_0 KV, passcode retrieved 3/3; config=262k-deep-tg",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "13.21 GB",
   "quote": "Welp-35B-A3B; RTX 4080 Super 16 GB; At the full 262k context (q4_0 KV), after a 214k-token prompt: decode 107 tok/s, prefill 2,043 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "HF model card: quaedra/Welp-35B-A3B-GGUF",
   "source_url": "https://huggingface.co/quaedra/Welp-35B-A3B-GGUF",
   "tg_tokens": null,
   "tps": "107",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 262144,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "RTX 4080 SUPER",
   "id": "hf-welp-welp35b-262k-deep-pp",
   "model": "Welp-35B-A3B",
   "notes": "Welp-35B-A3B card, one RTX 4080 Super 16 GB, same harness single runs, llama-server -ngl 99 -fa on -c 262144 -ctk q4_0 -ctv q4_0 -ub 256; decode llama-bench tg128, prefill llama-bench pp4096; prefill at full 262k ctx after a 214k-token prompt, q4_0 KV; config=262k-deep-pp",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "2043",
   "provenance": "community",
   "quant": "13.21 GB",
   "quote": "Welp-35B-A3B; RTX 4080 Super 16 GB; At the full 262k context (q4_0 KV), after a 214k-token prompt: decode 107 tok/s, prefill 2,043 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "HF model card: quaedra/Welp-35B-A3B-GGUF",
   "source_url": "https://huggingface.co/quaedra/Welp-35B-A3B-GGUF",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro (48 GB)",
   "id": "hf-phntmwvs-lfm258ba1bhermes-4bit-tg",
   "model": "LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated-v3",
   "notes": "MLX card bench on M4 Pro (48 GB), 4bit quant, peak memory 5.4 GB; generation; config=tg",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated-v3; M4 Pro (48 GB); Generation: 160.9 tok/s; prompt: 146.2 tok/s; peak memory: 5.4 GB",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "HF model card: phntmwvs/LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated-v3-4bit-MLX",
   "source_url": "https://huggingface.co/phntmwvs/LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated-v3-4bit-MLX",
   "tg_tokens": null,
   "tps": "160.9",
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro (48 GB)",
   "id": "hf-phntmwvs-lfm258ba1bhermes-4bit-pp",
   "model": "LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated-v3",
   "notes": "MLX card bench on M4 Pro (48 GB), 4bit quant, peak memory 5.4 GB; prompt; config=pp",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "146.2",
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated-v3; M4 Pro (48 GB); Generation: 160.9 tok/s; prompt: 146.2 tok/s; peak memory: 5.4 GB",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "HF model card: phntmwvs/LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated-v3-4bit-MLX",
   "source_url": "https://huggingface.co/phntmwvs/LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated-v3-4bit-MLX",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro (48 GB)",
   "id": "hf-phntmwvs-lfm258ba1bhermes-8bit-tg",
   "model": "LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated-v3",
   "notes": "MLX card bench on M4 Pro (48 GB), 8bit quant, peak memory 9.6 GB; generation; config=tg",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated-v3; M4 Pro (48 GB); Generation: 99.5 tok/s; prompt: 281.4 tok/s; peak memory: 9.6 GB",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "HF model card: phntmwvs/LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated-v3-8bit-MLX",
   "source_url": "https://huggingface.co/phntmwvs/LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated-v3-8bit-MLX",
   "tg_tokens": null,
   "tps": "99.5",
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro (48 GB)",
   "id": "hf-phntmwvs-lfm258ba1bhermes-8bit-pp",
   "model": "LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated-v3",
   "notes": "MLX card bench on M4 Pro (48 GB), 8bit quant, peak memory 9.6 GB; prompt; config=pp",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "281.4",
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated-v3; M4 Pro (48 GB); Generation: 99.5 tok/s; prompt: 281.4 tok/s; peak memory: 9.6 GB",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "HF model card: phntmwvs/LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated-v3-8bit-MLX",
   "source_url": "https://huggingface.co/phntmwvs/LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated-v3-8bit-MLX",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro (48 GB)",
   "id": "hf-phntmwvs-lfm258ba1bhermes-bf16-tg",
   "model": "LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated-v3",
   "notes": "MLX card bench on M4 Pro (48 GB), BF16 quant, peak memory 17.0 GB; generation; config=tg",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "BF16",
   "quote": "LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated-v3; M4 Pro (48 GB); Generation: 61.6 tok/s; prompt: 142.1 tok/s; peak memory: 17.0 GB",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "HF model card: phntmwvs/LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated-v3-BF16-MLX",
   "source_url": "https://huggingface.co/phntmwvs/LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated-v3-BF16-MLX",
   "tg_tokens": null,
   "tps": "61.6",
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro (48 GB)",
   "id": "hf-phntmwvs-lfm258ba1bhermes-bf16-pp",
   "model": "LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated-v3",
   "notes": "MLX card bench on M4 Pro (48 GB), BF16 quant, peak memory 17.0 GB; prompt; config=pp",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "142.1",
   "provenance": "community",
   "quant": "BF16",
   "quote": "LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated-v3; M4 Pro (48 GB); Generation: 61.6 tok/s; prompt: 142.1 tok/s; peak memory: 17.0 GB",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "HF model card: phntmwvs/LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated-v3-BF16-MLX",
   "source_url": "https://huggingface.co/phntmwvs/LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated-v3-BF16-MLX",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max (128 GB)",
   "id": "hf-pplxdecider27b-4bit-m5max-prefill-cool",
   "model": "pplx-decider-v1.1-27b",
   "notes": "pplx-decider-v1.1-27b 4bit MLX card bench, M5 Max (128 GB) High Power mode, over HTTP, sustained; 1.4-1.6x llama.cpp Q8_0/Q4_K_M same machine; prefill with cool GPU; config=prefill-cool",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "975",
   "provenance": "community",
   "quant": "4bit",
   "quote": "pplx-decider-v1.1-27b; M5 Max (128 GB); Speed depends on prefill, which is compute-bound: about 975 tok/s with a cool GPU, about 730 tok/s sustained",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "HF model card: mlx-community/pplx-decider-v1.1-27b-4bit",
   "source_url": "https://huggingface.co/mlx-community/pplx-decider-v1.1-27b-4bit",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max (128 GB)",
   "id": "hf-pplxdecider27b-4bit-m5max-prefill-sustained",
   "model": "pplx-decider-v1.1-27b",
   "notes": "pplx-decider-v1.1-27b 4bit MLX card bench, M5 Max (128 GB) High Power mode, over HTTP, sustained; 1.4-1.6x llama.cpp Q8_0/Q4_K_M same machine; prefill sustained; config=prefill-sustained",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": "730",
   "provenance": "community",
   "quant": "4bit",
   "quote": "pplx-decider-v1.1-27b; M5 Max (128 GB); Speed depends on prefill, which is compute-bound: about 975 tok/s with a cool GPU, about 730 tok/s sustained",
   "ram_gb": null,
   "retrieved": "2026-10-08",
   "source_name": "HF model card: mlx-community/pplx-decider-v1.1-27b-4bit",
   "source_url": "https://huggingface.co/mlx-community/pplx-decider-v1.1-27b-4bit",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 8,
   "ctx": 131072,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-29768-rtx5090-npl8-on",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #29768, CUDA graph warmup fix; RTX 5090 Qwen3.8-27B Q4_K_M; with patch arm (mean); -c 131072 -npl 8 -fa on; ngl 99",
   "params": null,
   "power_w": null,
   "pp_tokens": 15872,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Qwen3.8-27B Q4_K_M; RTX 5090 llama-batched-bench; 299.21 tok/s mean",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29768",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29768",
   "tg_tokens": 512,
   "tps": "299.21",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 8,
   "ctx": 131072,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-29768-rtx5090-npl8-off",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #29768, CUDA graph warmup fix; RTX 5090 Qwen3.8-27B Q4_K_M; without patch arm (mean); -c 131072 -npl 8 -fa on; ngl 99",
   "params": null,
   "power_w": null,
   "pp_tokens": 15872,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Qwen3.8-27B Q4_K_M; RTX 5090 llama-batched-bench; 295.85 tok/s mean",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp PR #29768",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29768",
   "tg_tokens": 512,
   "tps": "295.85",
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "RTX 4060 Ti",
   "id": "lc-29882-4060ti-qwen06-fa0-b11369",
   "model": "Qwen3-0.6B",
   "notes": "llama.cpp PR #29882, Vulkan rms-norm subgroup-reduction opt before/after; config=fa0-b11369; ngl -1; r 5; Windows",
   "pp_tokens": 1024,
   "pp_tps": "16094.75",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Qwen3-0.6B-Q4_K_M.gguf; Nvidia RTX 4060 TI; qwen3 0.6B Q4_K - Medium 372.65 MiB 596.05 M Vulkan -1 0 pp1024 16094.75 \u00b1 338.16; qwen3 0.6B Q4_K - Medium 372.65 MiB 596.05 M Vulkan -1 0 tg128 378.44 \u00b1 1.06",
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp PR #29882 (Vulkan rms-norm subgroup opt, before/after build)",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29882",
   "tg_tokens": 128,
   "tps": "378.44",
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "RTX 4060 Ti",
   "id": "lc-29882-4060ti-qwen06-fa1-b11369",
   "model": "Qwen3-0.6B",
   "notes": "llama.cpp PR #29882, Vulkan rms-norm subgroup-reduction opt before/after; config=fa1-b11369; ngl -1; r 5; Windows",
   "pp_tokens": 1024,
   "pp_tps": "24067.88",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Qwen3-0.6B-Q4_K_M.gguf; Nvidia RTX 4060 TI; qwen3 0.6B Q4_K - Medium 372.65 MiB 596.05 M Vulkan -1 1 pp1024 24067.88 \u00b1 136.88; qwen3 0.6B Q4_K - Medium 372.65 MiB 596.05 M Vulkan -1 1 tg128 390.33 \u00b1 9.78",
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp PR #29882 (Vulkan rms-norm subgroup opt, before/after build)",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29882",
   "tg_tokens": 128,
   "tps": "390.33",
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "RTX 4060 Ti",
   "id": "lc-29882-4060ti-qwen06-fa0-b11371",
   "model": "Qwen3-0.6B",
   "notes": "llama.cpp PR #29882, Vulkan rms-norm subgroup-reduction opt before/after; config=fa0-b11371; ngl -1; r 5; Windows",
   "pp_tokens": 1024,
   "pp_tps": "17654.48",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Qwen3-0.6B-Q4_K_M.gguf; Nvidia RTX 4060 TI; qwen3 0.6B Q4_K - Medium 372.65 MiB 596.05 M Vulkan -1 0 pp1024 17654.48 \u00b1 226.08; qwen3 0.6B Q4_K - Medium 372.65 MiB 596.05 M Vulkan -1 0 tg128 376.92 \u00b1 1.47",
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp PR #29882 (Vulkan rms-norm subgroup opt, before/after build)",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29882",
   "tg_tokens": 128,
   "tps": "376.92",
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "RTX 4060 Ti",
   "id": "lc-29882-4060ti-qwen06-fa1-b11371",
   "model": "Qwen3-0.6B",
   "notes": "llama.cpp PR #29882, Vulkan rms-norm subgroup-reduction opt before/after; config=fa1-b11371; ngl -1; r 5; Windows",
   "pp_tokens": 1024,
   "pp_tps": "28280.84",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Qwen3-0.6B-Q4_K_M.gguf; Nvidia RTX 4060 TI; qwen3 0.6B Q4_K - Medium 372.65 MiB 596.05 M Vulkan -1 1 pp1024 28280.84 \u00b1 312.71; qwen3 0.6B Q4_K - Medium 372.65 MiB 596.05 M Vulkan -1 1 tg128 396.13 \u00b1 8.27",
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp PR #29882 (Vulkan rms-norm subgroup opt, before/after build)",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29882",
   "tg_tokens": 128,
   "tps": "396.13",
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": 1,
   "ctx": 512,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "RTX 4060 Ti",
   "id": "lc-29882-4060ti-qwen35-fa0-b11369",
   "model": "Qwen3.5-9B",
   "notes": "llama.cpp PR #29882, Vulkan rms-norm subgroup-reduction opt before/after; config=fa0-b11369; ngl -1; r 5; Windows",
   "pp_tokens": 512,
   "pp_tps": "2314.17",
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "Qwen3.5-9B-Q6_K.gguf; Nvidia RTX 4060 TI; qwen35 9B Q6_K 6.94 GiB 8.95 B Vulkan -1 0 pp512 2314.17 \u00b1 8.11; qwen35 9B Q6_K 6.94 GiB 8.95 B Vulkan -1 0 tg128 37.54 \u00b1 0.03",
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp PR #29882 (Vulkan rms-norm subgroup opt, before/after build)",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29882",
   "tg_tokens": 128,
   "tps": "37.54",
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": 1,
   "ctx": 512,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "RTX 4060 Ti",
   "id": "lc-29882-4060ti-qwen35-fa1-b11369",
   "model": "Qwen3.5-9B",
   "notes": "llama.cpp PR #29882, Vulkan rms-norm subgroup-reduction opt before/after; config=fa1-b11369; ngl -1; r 5; Windows",
   "pp_tokens": 512,
   "pp_tps": "2320.02",
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "Qwen3.5-9B-Q6_K.gguf; Nvidia RTX 4060 TI; qwen35 9B Q6_K 6.94 GiB 8.95 B Vulkan -1 1 pp512 2320.02 \u00b1 37.29; qwen35 9B Q6_K 6.94 GiB 8.95 B Vulkan -1 1 tg128 37.46 \u00b1 0.14",
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp PR #29882 (Vulkan rms-norm subgroup opt, before/after build)",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29882",
   "tg_tokens": 128,
   "tps": "37.46",
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": 1,
   "ctx": 512,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "RTX 4060 Ti",
   "id": "lc-29882-4060ti-qwen35-fa0-b11371",
   "model": "Qwen3.5-9B",
   "notes": "llama.cpp PR #29882, Vulkan rms-norm subgroup-reduction opt before/after; config=fa0-b11371; ngl -1; r 5; Windows",
   "pp_tokens": 512,
   "pp_tps": "2407.05",
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "Qwen3.5-9B-Q6_K.gguf; Nvidia RTX 4060 TI; qwen35 9B Q6_K 6.94 GiB 8.95 B Vulkan -1 0 pp512 2407.05 \u00b1 8.79; qwen35 9B Q6_K 6.94 GiB 8.95 B Vulkan -1 0 tg128 37.55 \u00b1 0.05",
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp PR #29882 (Vulkan rms-norm subgroup opt, before/after build)",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29882",
   "tg_tokens": 128,
   "tps": "37.55",
   "ttft_s": null
  },
  {
   "backend": "Vulkan",
   "batch": 1,
   "ctx": 512,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "RTX 4060 Ti",
   "id": "lc-29882-4060ti-qwen35-fa1-b11371",
   "model": "Qwen3.5-9B",
   "notes": "llama.cpp PR #29882, Vulkan rms-norm subgroup-reduction opt before/after; config=fa1-b11371; ngl -1; r 5; Windows",
   "pp_tokens": 512,
   "pp_tps": "2430.82",
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "Qwen3.5-9B-Q6_K.gguf; Nvidia RTX 4060 TI; qwen35 9B Q6_K 6.94 GiB 8.95 B Vulkan -1 1 pp512 2430.82 \u00b1 6.47; qwen35 9B Q6_K 6.94 GiB 8.95 B Vulkan -1 1 tg128 37.64 \u00b1 0.06",
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp PR #29882 (Vulkan rms-norm subgroup opt, before/after build)",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29882",
   "tg_tokens": 128,
   "tps": "37.64",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": 8192,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "B200",
   "id": "vllm-59894-b200x4-8k1c16-main",
   "model": "DSV4.1-Flash",
   "notes": "#59894 decoder-replay cut-layer split, b200x4 point, main arm; vllm bench serve random, total tok/s",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "DEP4",
   "quote": "DSV4.1-Flash DEP4 on B200; 8k in / 1 out, C16, total tok/s 236,952 245,556; baseline",
   "retrieved": "2026-10-03",
   "source_name": "vLLM PR #59894",
   "source_url": "https://github.com/vllm-project/vllm/pull/59894",
   "tg_tokens": null,
   "tps": "236952",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": 8192,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "B200",
   "id": "vllm-59894-b200x4-8k1c16-pr",
   "model": "DSV4.1-Flash",
   "notes": "#59894 decoder-replay cut-layer split, b200x4 point, pr arm; vllm bench serve random, total tok/s",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "DEP4",
   "quote": "DSV4.1-Flash DEP4 on B200; 8k in / 1 out, C16, total tok/s 236,952 245,556; this PR",
   "retrieved": "2026-10-03",
   "source_name": "vLLM PR #59894",
   "source_url": "https://github.com/vllm-project/vllm/pull/59894",
   "tg_tokens": null,
   "tps": "245556",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": 32768,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "B200",
   "id": "vllm-59894-b200x4-32k1c16-main",
   "model": "DSV4.1-Flash",
   "notes": "#59894 decoder-replay cut-layer split, b200x4 point, main arm; vllm bench serve random, total tok/s",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "DEP4",
   "quote": "DSV4.1-Flash DEP4 on B200; 32k in / 1 out, C16 235,791 245,996; baseline",
   "retrieved": "2026-10-03",
   "source_name": "vLLM PR #59894",
   "source_url": "https://github.com/vllm-project/vllm/pull/59894",
   "tg_tokens": null,
   "tps": "235791",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": 32768,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "B200",
   "id": "vllm-59894-b200x4-32k1c16-pr",
   "model": "DSV4.1-Flash",
   "notes": "#59894 decoder-replay cut-layer split, b200x4 point, pr arm; vllm bench serve random, total tok/s",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "DEP4",
   "quote": "DSV4.1-Flash DEP4 on B200; 32k in / 1 out, C16 235,791 245,996; this PR",
   "retrieved": "2026-10-03",
   "source_name": "vLLM PR #59894",
   "source_url": "https://github.com/vllm-project/vllm/pull/59894",
   "tg_tokens": null,
   "tps": "245996",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 131072,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "B200",
   "id": "vllm-59894-b200x4-128k1c4-main",
   "model": "DSV4.1-Flash",
   "notes": "#59894 decoder-replay cut-layer split, b200x4 point, main arm; vllm bench serve random, total tok/s",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "DEP4",
   "quote": "DSV4.1-Flash DEP4 on B200; 128k in / 1 out, C4 194,012 204,764; baseline",
   "retrieved": "2026-10-03",
   "source_name": "vLLM PR #59894",
   "source_url": "https://github.com/vllm-project/vllm/pull/59894",
   "tg_tokens": null,
   "tps": "194012",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 131072,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "B200",
   "id": "vllm-59894-b200x4-128k1c4-pr",
   "model": "DSV4.1-Flash",
   "notes": "#59894 decoder-replay cut-layer split, b200x4 point, pr arm; vllm bench serve random, total tok/s",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "DEP4",
   "quote": "DSV4.1-Flash DEP4 on B200; 128k in / 1 out, C4 194,012 204,764; this PR",
   "retrieved": "2026-10-03",
   "source_name": "vLLM PR #59894",
   "source_url": "https://github.com/vllm-project/vllm/pull/59894",
   "tg_tokens": null,
   "tps": "204764",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 32,
   "ctx": 32768,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "B200",
   "id": "vllm-59894-b200x4-32k256c32-main",
   "model": "DSV4.1-Flash",
   "notes": "#59894 decoder-replay cut-layer split, b200x4 point, main arm; vllm bench serve random, total tok/s",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "DEP4",
   "quote": "DSV4.1-Flash DEP4 on B200; 32k in / 256 out, C32, output tok/s 1,281 1,303; baseline",
   "retrieved": "2026-10-03",
   "source_name": "vLLM PR #59894",
   "source_url": "https://github.com/vllm-project/vllm/pull/59894",
   "tg_tokens": null,
   "tps": "1281",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 32,
   "ctx": 32768,
   "date": "2026-10-03",
   "flags": [],
   "hardware": "B200",
   "id": "vllm-59894-b200x4-32k256c32-pr",
   "model": "DSV4.1-Flash",
   "notes": "#59894 decoder-replay cut-layer split, b200x4 point, pr arm; vllm bench serve random, total tok/s",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "DEP4",
   "quote": "DSV4.1-Flash DEP4 on B200; 32k in / 256 out, C32, output tok/s 1,281 1,303; this PR",
   "retrieved": "2026-10-03",
   "source_name": "vLLM PR #59894",
   "source_url": "https://github.com/vllm-project/vllm/pull/59894",
   "tg_tokens": null,
   "tps": "1303",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "GB10",
   "id": "vllm-60068-gb10-fixed3-c1",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MRV2 opt-in confidence stop for autoregressive drafting, MTP, SPEED-Bench qualitative 256 tokens, one run per arm, NVFP4 draft head from #59973, BF16 KV, measured on final commit 7c4775ff31; fixed 3 drafts, c=1; counterpart confidence stop 50.5 tok/s; config=fixed3-c1",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "GB10; Qwen3.8-Flash-Next; fixed 3 drafts; decode c=1, 11 categories; 49.1 tok/s (AL 2.80)",
   "ram_gb": null,
   "retrieved": "2026-10-12",
   "source_name": "vLLM PR #60068",
   "source_url": "https://github.com/vllm-project/vllm/pull/60068",
   "tg_tokens": 256,
   "tps": 49.1,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "GB10",
   "id": "vllm-60068-gb10-stop-c1",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MRV2 opt-in confidence stop for autoregressive drafting, MTP, SPEED-Bench qualitative 256 tokens, one run per arm, NVFP4 draft head from #59973, BF16 KV, measured on final commit 7c4775ff31; confidence stop max 6 fallback 3, c=1; counterpart fixed 3 drafts 49.1 tok/s; config=stop-c1",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "GB10; Qwen3.8-Flash-Next; confidence stop (max 6, fallback 3); decode c=1, 11 categories; 50.5 tok/s (AL 3.07)",
   "ram_gb": null,
   "retrieved": "2026-10-12",
   "source_name": "vLLM PR #60068",
   "source_url": "https://github.com/vllm-project/vllm/pull/60068",
   "tg_tokens": 256,
   "tps": 50.5,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": null,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "GB10",
   "id": "vllm-60068-gb10-fixed3-c4",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MRV2 opt-in confidence stop for autoregressive drafting, MTP, SPEED-Bench qualitative 256 tokens, one run per arm, NVFP4 draft head from #59973, BF16 KV, measured on final commit 7c4775ff31; fixed 3 drafts, c=4; counterpart confidence stop 32.6 tok/s; config=fixed3-c4",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "GB10; Qwen3.8-Flash-Next; fixed 3 drafts; decode c=4, 6 categories, per request; 33.0 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-12",
   "source_name": "vLLM PR #60068",
   "source_url": "https://github.com/vllm-project/vllm/pull/60068",
   "tg_tokens": 256,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": null,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "GB10",
   "id": "vllm-60068-gb10-stop-c4",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MRV2 opt-in confidence stop for autoregressive drafting, MTP, SPEED-Bench qualitative 256 tokens, one run per arm, NVFP4 draft head from #59973, BF16 KV, measured on final commit 7c4775ff31; confidence stop max 6 fallback 3, c=4; counterpart fixed 3 drafts 33.0 tok/s; config=stop-c4",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "GB10; Qwen3.8-Flash-Next; confidence stop (max 6, fallback 3); decode c=4, 6 categories, per request; 32.6 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-12",
   "source_name": "vLLM PR #60068",
   "source_url": "https://github.com/vllm-project/vllm/pull/60068",
   "tg_tokens": 256,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc-30493-qw3827b-master-pp512",
   "model": "Qwen3.8-27B",
   "notes": "CUDA, SM clock locked 2235 MHz; llama-bench -ngl 99 -r 5; MMQ activation tile copy int2; master vs PR pp512",
   "pp_tokens": 512,
   "pp_tps": "2984.8",
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Performance (RTX 4090, SM clock locked to 2235 MHz, llama-bench -ngl 99 -r 5, t/s):; Qwen3.8-27B UD-Q4_K_XL | 2984.8 | 3057.3 | 2954.5 | 3030.0 | +2.4% / +2.6%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30493",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30493"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc-30493-qw3827b-pr-pp512",
   "model": "Qwen3.8-27B",
   "notes": "CUDA, SM clock locked 2235 MHz; llama-bench -ngl 99 -r 5; MMQ activation tile copy int2; master vs PR pp512",
   "pp_tokens": 512,
   "pp_tps": "3057.3",
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Performance (RTX 4090, SM clock locked to 2235 MHz, llama-bench -ngl 99 -r 5, t/s):; Qwen3.8-27B UD-Q4_K_XL | 2984.8 | 3057.3 | 2954.5 | 3030.0 | +2.4% / +2.6%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30493",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30493"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc-30493-qw3827b-master-pp4096",
   "model": "Qwen3.8-27B",
   "notes": "CUDA, SM clock locked 2235 MHz; llama-bench -ngl 99 -r 5; MMQ activation tile copy int2; master vs PR pp4096",
   "pp_tokens": 4096,
   "pp_tps": "2954.5",
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Performance (RTX 4090, SM clock locked to 2235 MHz, llama-bench -ngl 99 -r 5, t/s):; Qwen3.8-27B UD-Q4_K_XL | 2984.8 | 3057.3 | 2954.5 | 3030.0 | +2.4% / +2.6%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30493",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30493"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc-30493-qw3827b-pr-pp4096",
   "model": "Qwen3.8-27B",
   "notes": "CUDA, SM clock locked 2235 MHz; llama-bench -ngl 99 -r 5; MMQ activation tile copy int2; master vs PR pp4096",
   "pp_tokens": 4096,
   "pp_tps": "3030.0",
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Performance (RTX 4090, SM clock locked to 2235 MHz, llama-bench -ngl 99 -r 5, t/s):; Qwen3.8-27B UD-Q4_K_XL | 2984.8 | 3057.3 | 2954.5 | 3030.0 | +2.4% / +2.6%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30493",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30493"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc-30493-glm47-master-pp512",
   "model": "GLM-4.7-Flash",
   "notes": "CUDA, SM clock locked 2235 MHz; llama-bench -ngl 99 -r 5; MMQ activation tile copy int2; master vs PR pp512",
   "pp_tokens": 512,
   "pp_tps": "6750.5",
   "provenance": "community",
   "quant": "Q4_K",
   "quote": "Performance (RTX 4090, SM clock locked to 2235 MHz, llama-bench -ngl 99 -r 5, t/s):; GLM-4.7-Flash Q4_K | 6750.5 | 6843.1 | 5732.7 | 5794.6 | +1.4% / +1.1%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30493",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30493"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc-30493-glm47-pr-pp512",
   "model": "GLM-4.7-Flash",
   "notes": "CUDA, SM clock locked 2235 MHz; llama-bench -ngl 99 -r 5; MMQ activation tile copy int2; master vs PR pp512",
   "pp_tokens": 512,
   "pp_tps": "6843.1",
   "provenance": "community",
   "quant": "Q4_K",
   "quote": "Performance (RTX 4090, SM clock locked to 2235 MHz, llama-bench -ngl 99 -r 5, t/s):; GLM-4.7-Flash Q4_K | 6750.5 | 6843.1 | 5732.7 | 5794.6 | +1.4% / +1.1%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30493",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30493"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc-30493-glm47-master-pp4096",
   "model": "GLM-4.7-Flash",
   "notes": "CUDA, SM clock locked 2235 MHz; llama-bench -ngl 99 -r 5; MMQ activation tile copy int2; master vs PR pp4096",
   "pp_tokens": 4096,
   "pp_tps": "5732.7",
   "provenance": "community",
   "quant": "Q4_K",
   "quote": "Performance (RTX 4090, SM clock locked to 2235 MHz, llama-bench -ngl 99 -r 5, t/s):; GLM-4.7-Flash Q4_K | 6750.5 | 6843.1 | 5732.7 | 5794.6 | +1.4% / +1.1%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30493",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30493"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc-30493-glm47-pr-pp4096",
   "model": "GLM-4.7-Flash",
   "notes": "CUDA, SM clock locked 2235 MHz; llama-bench -ngl 99 -r 5; MMQ activation tile copy int2; master vs PR pp4096",
   "pp_tokens": 4096,
   "pp_tps": "5794.6",
   "provenance": "community",
   "quant": "Q4_K",
   "quote": "Performance (RTX 4090, SM clock locked to 2235 MHz, llama-bench -ngl 99 -r 5, t/s):; GLM-4.7-Flash Q4_K | 6750.5 | 6843.1 | 5732.7 | 5794.6 | +1.4% / +1.1%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30493",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30493"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc-30493-qw3535b-master-pp512",
   "model": "Qwen3.5-35B-A3B",
   "notes": "CUDA, SM clock locked 2235 MHz; llama-bench -ngl 99 -r 5; MMQ activation tile copy int2; master vs PR pp512",
   "pp_tokens": 512,
   "pp_tps": "7473.5",
   "provenance": "community",
   "quant": "Q4_K_S",
   "quote": "Performance (RTX 4090, SM clock locked to 2235 MHz, llama-bench -ngl 99 -r 5, t/s):; Qwen3.5-35B-A3B Q4_K_S | 7473.5 | 7570.8 | 7283.5 | 7370.4 | +1.3% / +1.2%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30493",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30493"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc-30493-qw3535b-pr-pp512",
   "model": "Qwen3.5-35B-A3B",
   "notes": "CUDA, SM clock locked 2235 MHz; llama-bench -ngl 99 -r 5; MMQ activation tile copy int2; master vs PR pp512",
   "pp_tokens": 512,
   "pp_tps": "7570.8",
   "provenance": "community",
   "quant": "Q4_K_S",
   "quote": "Performance (RTX 4090, SM clock locked to 2235 MHz, llama-bench -ngl 99 -r 5, t/s):; Qwen3.5-35B-A3B Q4_K_S | 7473.5 | 7570.8 | 7283.5 | 7370.4 | +1.3% / +1.2%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30493",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30493"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc-30493-qw3535b-master-pp4096",
   "model": "Qwen3.5-35B-A3B",
   "notes": "CUDA, SM clock locked 2235 MHz; llama-bench -ngl 99 -r 5; MMQ activation tile copy int2; master vs PR pp4096",
   "pp_tokens": 4096,
   "pp_tps": "7283.5",
   "provenance": "community",
   "quant": "Q4_K_S",
   "quote": "Performance (RTX 4090, SM clock locked to 2235 MHz, llama-bench -ngl 99 -r 5, t/s):; Qwen3.5-35B-A3B Q4_K_S | 7473.5 | 7570.8 | 7283.5 | 7370.4 | +1.3% / +1.2%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30493",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30493"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc-30493-qw3535b-pr-pp4096",
   "model": "Qwen3.5-35B-A3B",
   "notes": "CUDA, SM clock locked 2235 MHz; llama-bench -ngl 99 -r 5; MMQ activation tile copy int2; master vs PR pp4096",
   "pp_tokens": 4096,
   "pp_tps": "7370.4",
   "provenance": "community",
   "quant": "Q4_K_S",
   "quote": "Performance (RTX 4090, SM clock locked to 2235 MHz, llama-bench -ngl 99 -r 5, t/s):; Qwen3.5-35B-A3B Q4_K_S | 7473.5 | 7570.8 | 7283.5 | 7370.4 | +1.3% / +1.2%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30493",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30493"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc-30493-qw359b-master-pp512",
   "model": "Qwen3.5-9B",
   "notes": "CUDA, SM clock locked 2235 MHz; llama-bench -ngl 99 -r 5; MMQ activation tile copy int2; master vs PR pp512",
   "pp_tokens": 512,
   "pp_tps": "9343.1",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Performance (RTX 4090, SM clock locked to 2235 MHz, llama-bench -ngl 99 -r 5, t/s):; Qwen3.5-9B Q4_K_M | 9343.1 | 9515.0 | 9472.7 | 9619.7 | +1.8% / +1.6%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30493",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30493"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc-30493-qw359b-pr-pp512",
   "model": "Qwen3.5-9B",
   "notes": "CUDA, SM clock locked 2235 MHz; llama-bench -ngl 99 -r 5; MMQ activation tile copy int2; master vs PR pp512",
   "pp_tokens": 512,
   "pp_tps": "9515.0",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Performance (RTX 4090, SM clock locked to 2235 MHz, llama-bench -ngl 99 -r 5, t/s):; Qwen3.5-9B Q4_K_M | 9343.1 | 9515.0 | 9472.7 | 9619.7 | +1.8% / +1.6%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30493",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30493"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc-30493-qw359b-master-pp4096",
   "model": "Qwen3.5-9B",
   "notes": "CUDA, SM clock locked 2235 MHz; llama-bench -ngl 99 -r 5; MMQ activation tile copy int2; master vs PR pp4096",
   "pp_tokens": 4096,
   "pp_tps": "9472.7",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Performance (RTX 4090, SM clock locked to 2235 MHz, llama-bench -ngl 99 -r 5, t/s):; Qwen3.5-9B Q4_K_M | 9343.1 | 9515.0 | 9472.7 | 9619.7 | +1.8% / +1.6%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30493",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30493"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc-30493-qw359b-pr-pp4096",
   "model": "Qwen3.5-9B",
   "notes": "CUDA, SM clock locked 2235 MHz; llama-bench -ngl 99 -r 5; MMQ activation tile copy int2; master vs PR pp4096",
   "pp_tokens": 4096,
   "pp_tps": "9619.7",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Performance (RTX 4090, SM clock locked to 2235 MHz, llama-bench -ngl 99 -r 5, t/s):; Qwen3.5-9B Q4_K_M | 9343.1 | 9515.0 | 9472.7 | 9619.7 | +1.8% / +1.6%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30493",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30493"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc-30493-kimi48b-master-pp512",
   "model": "Kimi-Linear-48B-A3B",
   "notes": "CUDA, SM clock locked 2235 MHz; llama-bench -ngl 99 -r 5; MMQ activation tile copy int2; master vs PR pp512",
   "pp_tokens": 512,
   "pp_tps": "5861.0",
   "provenance": "community",
   "quant": "IQ3_XS",
   "quote": "Performance (RTX 4090, SM clock locked to 2235 MHz, llama-bench -ngl 99 -r 5, t/s):; Kimi-Linear-48B-A3B IQ3_XS | 5861.0 | 6041.0 | 5854.8 | 6022.7 | +3.1% / +2.9%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30493",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30493"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc-30493-kimi48b-pr-pp512",
   "model": "Kimi-Linear-48B-A3B",
   "notes": "CUDA, SM clock locked 2235 MHz; llama-bench -ngl 99 -r 5; MMQ activation tile copy int2; master vs PR pp512",
   "pp_tokens": 512,
   "pp_tps": "6041.0",
   "provenance": "community",
   "quant": "IQ3_XS",
   "quote": "Performance (RTX 4090, SM clock locked to 2235 MHz, llama-bench -ngl 99 -r 5, t/s):; Kimi-Linear-48B-A3B IQ3_XS | 5861.0 | 6041.0 | 5854.8 | 6022.7 | +3.1% / +2.9%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30493",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30493"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc-30493-kimi48b-master-pp4096",
   "model": "Kimi-Linear-48B-A3B",
   "notes": "CUDA, SM clock locked 2235 MHz; llama-bench -ngl 99 -r 5; MMQ activation tile copy int2; master vs PR pp4096",
   "pp_tokens": 4096,
   "pp_tps": "5854.8",
   "provenance": "community",
   "quant": "IQ3_XS",
   "quote": "Performance (RTX 4090, SM clock locked to 2235 MHz, llama-bench -ngl 99 -r 5, t/s):; Kimi-Linear-48B-A3B IQ3_XS | 5861.0 | 6041.0 | 5854.8 | 6022.7 | +3.1% / +2.9%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30493",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30493"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc-30493-kimi48b-pr-pp4096",
   "model": "Kimi-Linear-48B-A3B",
   "notes": "CUDA, SM clock locked 2235 MHz; llama-bench -ngl 99 -r 5; MMQ activation tile copy int2; master vs PR pp4096",
   "pp_tokens": 4096,
   "pp_tps": "6022.7",
   "provenance": "community",
   "quant": "IQ3_XS",
   "quote": "Performance (RTX 4090, SM clock locked to 2235 MHz, llama-bench -ngl 99 -r 5, t/s):; Kimi-Linear-48B-A3B IQ3_XS | 5861.0 | 6041.0 | 5854.8 | 6022.7 | +3.1% / +2.9%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30493",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30493"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc-30489-glm47-master-pp2048",
   "model": "GLM-4.7-Flash",
   "notes": "CUDA, SM clock locked 2235 MHz; llama-bench -ngl 99 -r 5; GLU fused into MMQ Q8_1 activation quant; master vs PR",
   "pp_tokens": 2048,
   "pp_tps": "8437.7",
   "provenance": "community",
   "quant": "Q4_K",
   "quote": "Performance (RTX 4090, SM clock locked to 2235 MHz, llama-bench -ngl 99 -r 5):; GLM-4.7-Flash Q4_K, pp2048 | 8437.7 | 8587.5 | +1.8%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30489",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30489"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc-30489-glm47-pr-pp2048",
   "model": "GLM-4.7-Flash",
   "notes": "CUDA, SM clock locked 2235 MHz; llama-bench -ngl 99 -r 5; GLU fused into MMQ Q8_1 activation quant; master vs PR",
   "pp_tokens": 2048,
   "pp_tps": "8587.5",
   "provenance": "community",
   "quant": "Q4_K",
   "quote": "Performance (RTX 4090, SM clock locked to 2235 MHz, llama-bench -ngl 99 -r 5):; GLM-4.7-Flash Q4_K, pp2048 | 8437.7 | 8587.5 | +1.8%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30489",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30489"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc-30489-qw3535b-master-pp2048",
   "model": "Qwen3.5-35B-A3B",
   "notes": "CUDA, SM clock locked 2235 MHz; llama-bench -ngl 99 -r 5; GLU fused into MMQ Q8_1 activation quant; master vs PR",
   "pp_tokens": 2048,
   "pp_tps": "11554.2",
   "provenance": "community",
   "quant": "Q4_K_S",
   "quote": "Performance (RTX 4090, SM clock locked to 2235 MHz, llama-bench -ngl 99 -r 5):; Qwen3.5-35B-A3B Q4_K_S, pp2048 | 11554.2 | 11665.8 | +1.0%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30489",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30489"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc-30489-qw3535b-pr-pp2048",
   "model": "Qwen3.5-35B-A3B",
   "notes": "CUDA, SM clock locked 2235 MHz; llama-bench -ngl 99 -r 5; GLU fused into MMQ Q8_1 activation quant; master vs PR",
   "pp_tokens": 2048,
   "pp_tps": "11665.8",
   "provenance": "community",
   "quant": "Q4_K_S",
   "quote": "Performance (RTX 4090, SM clock locked to 2235 MHz, llama-bench -ngl 99 -r 5):; Qwen3.5-35B-A3B Q4_K_S, pp2048 | 11554.2 | 11665.8 | +1.0%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30489",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30489"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc-30489-qw3827b-master-pp4096",
   "model": "Qwen3.8-27B",
   "notes": "CUDA, SM clock locked 2235 MHz; llama-bench -ngl 99 -r 5; GLU fused into MMQ Q8_1 activation quant; master vs PR",
   "pp_tokens": 4096,
   "pp_tps": "2950.6",
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Performance (RTX 4090, SM clock locked to 2235 MHz, llama-bench -ngl 99 -r 5):; Qwen3.8-27B UD-Q4_K_XL, pp4096 | 2950.6 | 3000.7 | +1.7%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30489",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30489"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc-30489-qw3827b-pr-pp4096",
   "model": "Qwen3.8-27B",
   "notes": "CUDA, SM clock locked 2235 MHz; llama-bench -ngl 99 -r 5; GLU fused into MMQ Q8_1 activation quant; master vs PR",
   "pp_tokens": 4096,
   "pp_tps": "3000.7",
   "provenance": "community",
   "quant": "UD-Q4_K_XL",
   "quote": "Performance (RTX 4090, SM clock locked to 2235 MHz, llama-bench -ngl 99 -r 5):; Qwen3.8-27B UD-Q4_K_XL, pp4096 | 2950.6 | 3000.7 | +1.7%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30489",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30489"
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": 8192,
   "date": "2026-10",
   "flags": [],
   "hardware": "M5 Pro (20-core GPU)",
   "id": "lc-30491-qw3827b-master-pp512d8192",
   "model": "Qwen3.8-27B",
   "notes": "Metal, BF16 tensor-API FA kernel; -fa 1, KV depth 8192; master(BF16 KV) vs PR(BF16 KV) vs PR(F16 KV)",
   "pp_tokens": 512,
   "pp_tps": "323.8",
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M5 Pro (20-core GPU), master vs PR:; llama-bench, Qwen3.8-27B Q8_0, -fa 1; pp512 @ d8192 323.8 355.3 354.2",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30491",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30491"
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": 8192,
   "date": "2026-10",
   "flags": [],
   "hardware": "M5 Pro (20-core GPU)",
   "id": "lc-30491-qw3827b-pr-pp512d8192",
   "model": "Qwen3.8-27B",
   "notes": "Metal, BF16 tensor-API FA kernel; -fa 1, KV depth 8192; master(BF16 KV) vs PR(BF16 KV) vs PR(F16 KV)",
   "pp_tokens": 512,
   "pp_tps": "355.3",
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M5 Pro (20-core GPU), master vs PR:; llama-bench, Qwen3.8-27B Q8_0, -fa 1; pp512 @ d8192 323.8 355.3 354.2",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30491",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30491"
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": 32768,
   "date": "2026-10",
   "flags": [],
   "hardware": "M5 Pro (20-core GPU)",
   "id": "lc-30491-qw3827b-master-pp512d32768",
   "model": "Qwen3.8-27B",
   "notes": "Metal, BF16 tensor-API FA kernel; -fa 1, KV depth 32768",
   "pp_tokens": 512,
   "pp_tps": "184.2",
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M5 Pro (20-core GPU), master vs PR:; llama-bench, Qwen3.8-27B Q8_0, -fa 1; pp512 @ d32768 184.2 265.6 268.5",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30491",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30491"
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": 32768,
   "date": "2026-10",
   "flags": [],
   "hardware": "M5 Pro (20-core GPU)",
   "id": "lc-30491-qw3827b-pr-pp512d32768",
   "model": "Qwen3.8-27B",
   "notes": "Metal, BF16 tensor-API FA kernel; -fa 1, KV depth 32768",
   "pp_tokens": 512,
   "pp_tps": "265.6",
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M5 Pro (20-core GPU), master vs PR:; llama-bench, Qwen3.8-27B Q8_0, -fa 1; pp512 @ d32768 184.2 265.6 268.5",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30491",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30491"
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": 8192,
   "date": "2026-10",
   "flags": [],
   "hardware": "M5 Pro (20-core GPU)",
   "id": "lc-30491-qw3827b-master-tg32d8192",
   "model": "Qwen3.8-27B",
   "notes": "Metal, BF16 tensor-API FA kernel; -fa 1, KV depth 8192",
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M5 Pro (20-core GPU), master vs PR:; llama-bench, Qwen3.8-27B Q8_0, -fa 1; tg32 @ d8192 9.09 9.25 9.32",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30491",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30491",
   "tg_tokens": 32,
   "tps": "9.09"
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": 8192,
   "date": "2026-10",
   "flags": [],
   "hardware": "M5 Pro (20-core GPU)",
   "id": "lc-30491-qw3827b-pr-tg32d8192",
   "model": "Qwen3.8-27B",
   "notes": "Metal, BF16 tensor-API FA kernel; -fa 1, KV depth 8192",
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M5 Pro (20-core GPU), master vs PR:; llama-bench, Qwen3.8-27B Q8_0, -fa 1; tg32 @ d8192 9.09 9.25 9.32",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30491",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30491",
   "tg_tokens": 32,
   "tps": "9.25"
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Arc Pro B50",
   "id": "lc-30374-qw359b-master-pp512d0",
   "model": "Qwen3.5-9B",
   "notes": "SYCL, Arc Pro B50 (Battlemage); llama-bench -ngl 99 -fa 1 -ctk q8_0 -ctv q8_0 -ub 1024 -p 512 -n 32 -d 0,1; Q4_0/Q8_0 weight-reorder PP regression fix; master vs PR",
   "pp_tokens": 512,
   "pp_tps": "1029",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B50 (Battlemage), Qwen3.5-9B Q4_0:; pp512 (before any decode) 1029 1028",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30374",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30374"
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Arc Pro B50",
   "id": "lc-30374-qw359b-pr-pp512d0",
   "model": "Qwen3.5-9B",
   "notes": "SYCL, Arc Pro B50 (Battlemage); llama-bench -ngl 99 -fa 1 -ctk q8_0 -ctv q8_0 -ub 1024 -p 512 -n 32 -d 0,1; Q4_0/Q8_0 weight-reorder PP regression fix; master vs PR",
   "pp_tokens": 512,
   "pp_tps": "1028",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B50 (Battlemage), Qwen3.5-9B Q4_0:; pp512 (before any decode) 1029 1028",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30374",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30374"
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Arc Pro B50",
   "id": "lc-30374-qw359b-master-pp512d1",
   "model": "Qwen3.5-9B",
   "notes": "SYCL, Arc Pro B50 (Battlemage); llama-bench -ngl 99 -fa 1 -ctk q8_0 -ctv q8_0 -ub 1024 -p 512 -n 32 -d 0,1; Q4_0/Q8_0 weight-reorder PP regression fix; master vs PR",
   "pp_tokens": 512,
   "pp_tps": "607",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B50 (Battlemage), Qwen3.5-9B Q4_0:; pp512 @ d1 (after decode, reordered) 607 1005",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30374",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30374"
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Arc Pro B50",
   "id": "lc-30374-qw359b-pr-pp512d1",
   "model": "Qwen3.5-9B",
   "notes": "SYCL, Arc Pro B50 (Battlemage); llama-bench -ngl 99 -fa 1 -ctk q8_0 -ctv q8_0 -ub 1024 -p 512 -n 32 -d 0,1; Q4_0/Q8_0 weight-reorder PP regression fix; master vs PR",
   "pp_tokens": 512,
   "pp_tps": "1005",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B50 (Battlemage), Qwen3.5-9B Q4_0:; pp512 @ d1 (after decode, reordered) 607 1005",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30374",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30374"
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Arc Pro B50",
   "id": "lc-30374-qw359b-master-tg32",
   "model": "Qwen3.5-9B",
   "notes": "SYCL, Arc Pro B50 (Battlemage); llama-bench -ngl 99 -fa 1 -ctk q8_0 -ctv q8_0 -ub 1024 -p 512 -n 32 -d 0,1; Q4_0/Q8_0 weight-reorder PP regression fix; master vs PR",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B50 (Battlemage), Qwen3.5-9B Q4_0:; tg32 33 33",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30374",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30374",
   "tg_tokens": 32,
   "tps": "33"
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Arc Pro B50",
   "id": "lc-30374-qw359b-pr-tg32",
   "model": "Qwen3.5-9B",
   "notes": "SYCL, Arc Pro B50 (Battlemage); llama-bench -ngl 99 -fa 1 -ctk q8_0 -ctv q8_0 -ub 1024 -p 512 -n 32 -d 0,1; Q4_0/Q8_0 weight-reorder PP regression fix; master vs PR",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B50 (Battlemage), Qwen3.5-9B Q4_0:; tg32 33 33",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30374",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30374",
   "tg_tokens": 32,
   "tps": "33"
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Arc Pro B50",
   "id": "lc-30374-qw359b-master-tg32d1",
   "model": "Qwen3.5-9B",
   "notes": "SYCL, Arc Pro B50 (Battlemage); llama-bench -ngl 99 -fa 1 -ctk q8_0 -ctv q8_0 -ub 1024 -p 512 -n 32 -d 0,1; Q4_0/Q8_0 weight-reorder PP regression fix; master vs PR",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B50 (Battlemage), Qwen3.5-9B Q4_0:; tg32 @ d1 33 33",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30374",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30374",
   "tg_tokens": 32,
   "tps": "33"
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Arc Pro B50",
   "id": "lc-30374-qw359b-pr-tg32d1",
   "model": "Qwen3.5-9B",
   "notes": "SYCL, Arc Pro B50 (Battlemage); llama-bench -ngl 99 -fa 1 -ctk q8_0 -ctv q8_0 -ub 1024 -p 512 -n 32 -d 0,1; Q4_0/Q8_0 weight-reorder PP regression fix; master vs PR",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B50 (Battlemage), Qwen3.5-9B Q4_0:; tg32 @ d1 33 33",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30374",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30374",
   "tg_tokens": 32,
   "tps": "33"
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Arc Pro B50",
   "id": "lc-30374-qw359b-master-pp1500",
   "model": "Qwen3.5-9B",
   "notes": "SYCL, Arc Pro B50 (Battlemage); llama-bench -ngl 99 -fa 1 -ctk q8_0 -ctv q8_0 -ub 1024 -p 512 -n 32 -d 0,1; Q4_0/Q8_0 weight-reorder PP regression fix; master vs PR",
   "pp_tokens": 1500,
   "pp_tps": "767",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B50 (Battlemage), Qwen3.5-9B Q4_0:; server prompt, 1.5k tokens, after decode 767 1150",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30374",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30374"
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Arc Pro B50",
   "id": "lc-30374-qw359b-pr-pp1500",
   "model": "Qwen3.5-9B",
   "notes": "SYCL, Arc Pro B50 (Battlemage); llama-bench -ngl 99 -fa 1 -ctk q8_0 -ctv q8_0 -ub 1024 -p 512 -n 32 -d 0,1; Q4_0/Q8_0 weight-reorder PP regression fix; master vs PR",
   "pp_tokens": 1500,
   "pp_tps": "1150",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B50 (Battlemage), Qwen3.5-9B Q4_0:; server prompt, 1.5k tokens, after decode 767 1150",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30374",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30374"
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Arc Pro B50",
   "id": "lc-30374-qw352b-master-pp512d0",
   "model": "Qwen3.5-2B",
   "notes": "SYCL, Arc Pro B50 (Battlemage); llama-bench -ngl 99 -fa 1 -ctk q8_0 -ctv q8_0 -ub 1024 -p 512 -n 32 -d 0,1; Q4_0/Q8_0 weight-reorder PP regression fix; master vs PR",
   "pp_tokens": 512,
   "pp_tps": "4084",
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B50, Qwen3.5-2B Q8_0:; pp512 (before any decode) 4084 4092",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30374",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30374"
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Arc Pro B50",
   "id": "lc-30374-qw352b-pr-pp512d0",
   "model": "Qwen3.5-2B",
   "notes": "SYCL, Arc Pro B50 (Battlemage); llama-bench -ngl 99 -fa 1 -ctk q8_0 -ctv q8_0 -ub 1024 -p 512 -n 32 -d 0,1; Q4_0/Q8_0 weight-reorder PP regression fix; master vs PR",
   "pp_tokens": 512,
   "pp_tps": "4092",
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B50, Qwen3.5-2B Q8_0:; pp512 (before any decode) 4084 4092",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30374",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30374"
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Arc Pro B50",
   "id": "lc-30374-qw352b-master-pp512d1",
   "model": "Qwen3.5-2B",
   "notes": "SYCL, Arc Pro B50 (Battlemage); llama-bench -ngl 99 -fa 1 -ctk q8_0 -ctv q8_0 -ub 1024 -p 512 -n 32 -d 0,1; Q4_0/Q8_0 weight-reorder PP regression fix; master vs PR",
   "pp_tokens": 512,
   "pp_tps": "3117",
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B50, Qwen3.5-2B Q8_0:; pp512 @ d1 (after decode, reordered) 3117 4161",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30374",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30374"
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Arc Pro B50",
   "id": "lc-30374-qw352b-pr-pp512d1",
   "model": "Qwen3.5-2B",
   "notes": "SYCL, Arc Pro B50 (Battlemage); llama-bench -ngl 99 -fa 1 -ctk q8_0 -ctv q8_0 -ub 1024 -p 512 -n 32 -d 0,1; Q4_0/Q8_0 weight-reorder PP regression fix; master vs PR",
   "pp_tokens": 512,
   "pp_tps": "4161",
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B50, Qwen3.5-2B Q8_0:; pp512 @ d1 (after decode, reordered) 3117 4161",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30374",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30374"
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Arc Pro B50",
   "id": "lc-30374-qw352b-master-tg32",
   "model": "Qwen3.5-2B",
   "notes": "SYCL, Arc Pro B50 (Battlemage); llama-bench -ngl 99 -fa 1 -ctk q8_0 -ctv q8_0 -ub 1024 -p 512 -n 32 -d 0,1; Q4_0/Q8_0 weight-reorder PP regression fix; master vs PR",
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B50, Qwen3.5-2B Q8_0:; tg32 81 81",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30374",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30374",
   "tg_tokens": 32,
   "tps": "81"
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Arc Pro B50",
   "id": "lc-30374-qw352b-pr-tg32",
   "model": "Qwen3.5-2B",
   "notes": "SYCL, Arc Pro B50 (Battlemage); llama-bench -ngl 99 -fa 1 -ctk q8_0 -ctv q8_0 -ub 1024 -p 512 -n 32 -d 0,1; Q4_0/Q8_0 weight-reorder PP regression fix; master vs PR",
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B50, Qwen3.5-2B Q8_0:; tg32 81 81",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30374",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30374",
   "tg_tokens": 32,
   "tps": "81"
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Arc Pro B50",
   "id": "lc-30374-qw352b-master-tg32d1",
   "model": "Qwen3.5-2B",
   "notes": "SYCL, Arc Pro B50 (Battlemage); llama-bench -ngl 99 -fa 1 -ctk q8_0 -ctv q8_0 -ub 1024 -p 512 -n 32 -d 0,1; Q4_0/Q8_0 weight-reorder PP regression fix; master vs PR",
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B50, Qwen3.5-2B Q8_0:; tg32 @ d1 81 81",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30374",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30374",
   "tg_tokens": 32,
   "tps": "81"
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Arc Pro B50",
   "id": "lc-30374-qw352b-pr-tg32d1",
   "model": "Qwen3.5-2B",
   "notes": "SYCL, Arc Pro B50 (Battlemage); llama-bench -ngl 99 -fa 1 -ctk q8_0 -ctv q8_0 -ub 1024 -p 512 -n 32 -d 0,1; Q4_0/Q8_0 weight-reorder PP regression fix; master vs PR",
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B50, Qwen3.5-2B Q8_0:; tg32 @ d1 81 81",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30374",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30374",
   "tg_tokens": 32,
   "tps": "81"
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Arc Pro B50",
   "id": "lc-30374-qw352b-master-pp1500",
   "model": "Qwen3.5-2B",
   "notes": "SYCL, Arc Pro B50 (Battlemage); llama-bench -ngl 99 -fa 1 -ctk q8_0 -ctv q8_0 -ub 1024 -p 512 -n 32 -d 0,1; Q4_0/Q8_0 weight-reorder PP regression fix; master vs PR",
   "pp_tokens": 1500,
   "pp_tps": "3506",
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B50, Qwen3.5-2B Q8_0:; server prompt, 1.5k tokens, after decode 3506 4335",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30374",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30374"
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Arc Pro B50",
   "id": "lc-30374-qw352b-pr-pp1500",
   "model": "Qwen3.5-2B",
   "notes": "SYCL, Arc Pro B50 (Battlemage); llama-bench -ngl 99 -fa 1 -ctk q8_0 -ctv q8_0 -ub 1024 -p 512 -n 32 -d 0,1; Q4_0/Q8_0 weight-reorder PP regression fix; master vs PR",
   "pp_tokens": 1500,
   "pp_tps": "4335",
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B50, Qwen3.5-2B Q8_0:; server prompt, 1.5k tokens, after decode 3506 4335",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30374",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30374"
  },
  {
   "backend": "llama.cpp (OpenCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Adreno X2-90",
   "id": "lc-30249-gemma412b-before-pp2048d0",
   "model": "Gemma 4 12B",
   "notes": "OpenCL, Adreno X2-90; fused int8 KQ head-size-512; GGML_OPENCL_FA_KQ_INT8=1; before vs after",
   "pp_tokens": 2048,
   "pp_tps": "261.5",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Adreno X2-90, llama-bench -fa 1 -n 0, GGML_OPENCL_FA_KQ_INT8=1:; gemma-4-12B QAT Q4_0, pp2048 depth 0 261.5 263.2",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30249",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30249"
  },
  {
   "backend": "llama.cpp (OpenCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Adreno X2-90",
   "id": "lc-30249-gemma412b-after-pp2048d0",
   "model": "Gemma 4 12B",
   "notes": "OpenCL, Adreno X2-90; fused int8 KQ head-size-512; GGML_OPENCL_FA_KQ_INT8=1; before vs after",
   "pp_tokens": 2048,
   "pp_tps": "263.2",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Adreno X2-90, llama-bench -fa 1 -n 0, GGML_OPENCL_FA_KQ_INT8=1:; gemma-4-12B QAT Q4_0, pp2048 depth 0 261.5 263.2",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30249",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30249"
  },
  {
   "backend": "llama.cpp (OpenCL)",
   "batch": null,
   "ctx": 16384,
   "date": "2026-10",
   "flags": [],
   "hardware": "Adreno X2-90",
   "id": "lc-30249-gemma412b-before-pp2048d16384",
   "model": "Gemma 4 12B",
   "notes": "OpenCL, Adreno X2-90; fused int8 KQ head-size-512; GGML_OPENCL_FA_KQ_INT8=1; before vs after",
   "pp_tokens": 2048,
   "pp_tps": "186.7",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Adreno X2-90, llama-bench -fa 1 -n 0, GGML_OPENCL_FA_KQ_INT8=1:; gemma-4-12B QAT Q4_0, pp2048 depth 16384 186.7 209.3",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30249",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30249"
  },
  {
   "backend": "llama.cpp (OpenCL)",
   "batch": null,
   "ctx": 16384,
   "date": "2026-10",
   "flags": [],
   "hardware": "Adreno X2-90",
   "id": "lc-30249-gemma412b-after-pp2048d16384",
   "model": "Gemma 4 12B",
   "notes": "OpenCL, Adreno X2-90; fused int8 KQ head-size-512; GGML_OPENCL_FA_KQ_INT8=1; before vs after",
   "pp_tokens": 2048,
   "pp_tps": "209.3",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Adreno X2-90, llama-bench -fa 1 -n 0, GGML_OPENCL_FA_KQ_INT8=1:; gemma-4-12B QAT Q4_0, pp2048 depth 16384 186.7 209.3",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30249",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30249"
  },
  {
   "backend": "llama.cpp (OpenCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Adreno X2-90",
   "id": "lc-30249-gemma4e4b-before-pp2048d0",
   "model": "Gemma 4 E4B",
   "notes": "OpenCL, Adreno X2-90; fused int8 KQ head-size-512; GGML_OPENCL_FA_KQ_INT8=1; before vs after",
   "pp_tokens": 2048,
   "pp_tps": "565.2",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Adreno X2-90, llama-bench -fa 1 -n 0, GGML_OPENCL_FA_KQ_INT8=1:; gemma-4-E4B Q4_K_M, pp2048 depth 0 565.2 567.4",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30249",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30249"
  },
  {
   "backend": "llama.cpp (OpenCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Adreno X2-90",
   "id": "lc-30249-gemma4e4b-after-pp2048d0",
   "model": "Gemma 4 E4B",
   "notes": "OpenCL, Adreno X2-90; fused int8 KQ head-size-512; GGML_OPENCL_FA_KQ_INT8=1; before vs after",
   "pp_tokens": 2048,
   "pp_tps": "567.4",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Adreno X2-90, llama-bench -fa 1 -n 0, GGML_OPENCL_FA_KQ_INT8=1:; gemma-4-E4B Q4_K_M, pp2048 depth 0 565.2 567.4",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30249",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30249"
  },
  {
   "backend": "llama.cpp (OpenCL)",
   "batch": null,
   "ctx": 4096,
   "date": "2026-10",
   "flags": [],
   "hardware": "Adreno X2-90",
   "id": "lc-30247-gemma412b-before-tg64d4096",
   "model": "Gemma 4 12B",
   "notes": "OpenCL, Adreno X2-90; gemma-4 global-attention decode via FA decomposition; -fa 1 -p 0 -n 64; before vs after (measured pre-rebase on #30266)",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "gemma-4-12B QAT Q4_0, Adreno X2-90, llama-bench -fa 1 -p 0 -n 64; 4096 13.77 13.93",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30247",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30247",
   "tg_tokens": 64,
   "tps": "13.77"
  },
  {
   "backend": "llama.cpp (OpenCL)",
   "batch": null,
   "ctx": 4096,
   "date": "2026-10",
   "flags": [],
   "hardware": "Adreno X2-90",
   "id": "lc-30247-gemma412b-after-tg64d4096",
   "model": "Gemma 4 12B",
   "notes": "OpenCL, Adreno X2-90; gemma-4 global-attention decode via FA decomposition; -fa 1 -p 0 -n 64; before vs after (measured pre-rebase on #30266)",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "gemma-4-12B QAT Q4_0, Adreno X2-90, llama-bench -fa 1 -p 0 -n 64; 4096 13.77 13.93",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30247",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30247",
   "tg_tokens": 64,
   "tps": "13.93"
  },
  {
   "backend": "llama.cpp (OpenCL)",
   "batch": null,
   "ctx": 16384,
   "date": "2026-10",
   "flags": [],
   "hardware": "Adreno X2-90",
   "id": "lc-30247-gemma412b-before-tg64d16384",
   "model": "Gemma 4 12B",
   "notes": "OpenCL, Adreno X2-90; gemma-4 global-attention decode via FA decomposition; -fa 1 -p 0 -n 64; before vs after (measured pre-rebase on #30266)",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "gemma-4-12B QAT Q4_0, Adreno X2-90, llama-bench -fa 1 -p 0 -n 64; 16384 12.15 13.33",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30247",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30247",
   "tg_tokens": 64,
   "tps": "12.15"
  },
  {
   "backend": "llama.cpp (OpenCL)",
   "batch": null,
   "ctx": 16384,
   "date": "2026-10",
   "flags": [],
   "hardware": "Adreno X2-90",
   "id": "lc-30247-gemma412b-after-tg64d16384",
   "model": "Gemma 4 12B",
   "notes": "OpenCL, Adreno X2-90; gemma-4 global-attention decode via FA decomposition; -fa 1 -p 0 -n 64; before vs after (measured pre-rebase on #30266)",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "gemma-4-12B QAT Q4_0, Adreno X2-90, llama-bench -fa 1 -p 0 -n 64; 16384 12.15 13.33",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30247",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30247",
   "tg_tokens": 64,
   "tps": "13.33"
  },
  {
   "backend": "llama.cpp (OpenCL)",
   "batch": null,
   "ctx": 32768,
   "date": "2026-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "Adreno X2-90",
   "id": "lc-30247-gemma412b-before-tg64d32768",
   "model": "Gemma 4 12B",
   "notes": "OpenCL, Adreno X2-90; gemma-4 global-attention decode via FA decomposition; -fa 1 -p 0 -n 64; before vs after (measured pre-rebase on #30266)",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "gemma-4-12B QAT Q4_0, Adreno X2-90, llama-bench -fa 1 -p 0 -n 64; 32768 10.66 12.58",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30247",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30247",
   "tg_tokens": 64,
   "tps": "10.66"
  },
  {
   "backend": "llama.cpp (OpenCL)",
   "batch": null,
   "ctx": 32768,
   "date": "2026-10",
   "flags": [],
   "hardware": "Adreno X2-90",
   "id": "lc-30247-gemma412b-after-tg64d32768",
   "model": "Gemma 4 12B",
   "notes": "OpenCL, Adreno X2-90; gemma-4 global-attention decode via FA decomposition; -fa 1 -p 0 -n 64; before vs after (measured pre-rebase on #30266)",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "gemma-4-12B QAT Q4_0, Adreno X2-90, llama-bench -fa 1 -p 0 -n 64; 32768 10.66 12.58",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30247",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30247",
   "tg_tokens": 64,
   "tps": "12.58"
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Xeon 6985P-C",
   "id": "lc-30231-gemma4e4b-avx512-master-pp512",
   "model": "Gemma 4 E4B",
   "notes": "CPU, Xeon 6985P-C (Granite Rapids); vectorized GEGLU/GELU table lookup; -t 8; first of two runs shown",
   "pp_tokens": 512,
   "pp_tps": "152.06",
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Intel Xeon 6985P-C (Granite Rapids), GCP c4-highcpu-16 with all-core max turbo, GCC 14.2, master 08246a28f.; llama-bench, Gemma 4 E4B Q8_0, -t 8 -p 512 -n 128 -r 5, t/s, two runs each; AVX-512 (-DGGML_NATIVE=ON) pp512 152.06 / 152.27 155.77 / 156.07",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30231",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30231"
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Xeon 6985P-C",
   "id": "lc-30231-gemma4e4b-avx512-branch-pp512",
   "model": "Gemma 4 E4B",
   "notes": "CPU, Xeon 6985P-C (Granite Rapids); vectorized GEGLU/GELU table lookup; -t 8; first of two runs shown",
   "pp_tokens": 512,
   "pp_tps": "155.77",
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Intel Xeon 6985P-C (Granite Rapids), GCP c4-highcpu-16 with all-core max turbo, GCC 14.2, master 08246a28f.; llama-bench, Gemma 4 E4B Q8_0, -t 8 -p 512 -n 128 -r 5, t/s, two runs each; AVX-512 (-DGGML_NATIVE=ON) pp512 152.06 / 152.27 155.77 / 156.07",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30231",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30231"
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Xeon 6985P-C",
   "id": "lc-30231-gemma4e4b-avx512-master-tg128",
   "model": "Gemma 4 E4B",
   "notes": "CPU, Xeon 6985P-C (Granite Rapids); vectorized GEGLU/GELU table lookup; -t 8; first of two runs shown",
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Intel Xeon 6985P-C (Granite Rapids), GCP c4-highcpu-16 with all-core max turbo, GCC 14.2, master 08246a28f.; llama-bench, Gemma 4 E4B Q8_0, -t 8 -p 512 -n 128 -r 5, t/s, two runs each; AVX-512 (-DGGML_NATIVE=ON) tg128 10.43 / 10.55 10.60 / 10.75",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30231",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30231",
   "tg_tokens": 128,
   "tps": "10.43"
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Xeon 6985P-C",
   "id": "lc-30231-gemma4e4b-avx512-branch-tg128",
   "model": "Gemma 4 E4B",
   "notes": "CPU, Xeon 6985P-C (Granite Rapids); vectorized GEGLU/GELU table lookup; -t 8; first of two runs shown",
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Intel Xeon 6985P-C (Granite Rapids), GCP c4-highcpu-16 with all-core max turbo, GCC 14.2, master 08246a28f.; llama-bench, Gemma 4 E4B Q8_0, -t 8 -p 512 -n 128 -r 5, t/s, two runs each; AVX-512 (-DGGML_NATIVE=ON) tg128 10.43 / 10.55 10.60 / 10.75",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30231",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30231",
   "tg_tokens": 128,
   "tps": "10.60"
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Xeon 6985P-C",
   "id": "lc-30231-gemma4e4b-avx2-master-pp512",
   "model": "Gemma 4 E4B",
   "notes": "CPU, Xeon 6985P-C (Granite Rapids); vectorized GEGLU/GELU table lookup; -t 8; first of two runs shown",
   "pp_tokens": 512,
   "pp_tps": "39.75",
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Intel Xeon 6985P-C (Granite Rapids), GCP c4-highcpu-16 with all-core max turbo, GCC 14.2, master 08246a28f.; llama-bench, Gemma 4 E4B Q8_0, -t 8 -p 512 -n 128 -r 5, t/s, two runs each; AVX2 only pp512 39.75 / 39.88 39.96 / 40.12",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30231",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30231"
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Xeon 6985P-C",
   "id": "lc-30231-gemma4e4b-avx2-branch-pp512",
   "model": "Gemma 4 E4B",
   "notes": "CPU, Xeon 6985P-C (Granite Rapids); vectorized GEGLU/GELU table lookup; -t 8; first of two runs shown",
   "pp_tokens": 512,
   "pp_tps": "39.96",
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Intel Xeon 6985P-C (Granite Rapids), GCP c4-highcpu-16 with all-core max turbo, GCC 14.2, master 08246a28f.; llama-bench, Gemma 4 E4B Q8_0, -t 8 -p 512 -n 128 -r 5, t/s, two runs each; AVX2 only pp512 39.75 / 39.88 39.96 / 40.12",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30231",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30231"
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "Xeon 6985P-C",
   "id": "lc-30231-gemma4e4b-avx2-master-tg128",
   "model": "Gemma 4 E4B",
   "notes": "CPU, Xeon 6985P-C (Granite Rapids); vectorized GEGLU/GELU table lookup; -t 8; first of two runs shown",
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Intel Xeon 6985P-C (Granite Rapids), GCP c4-highcpu-16 with all-core max turbo, GCC 14.2, master 08246a28f.; llama-bench, Gemma 4 E4B Q8_0, -t 8 -p 512 -n 128 -r 5, t/s, two runs each; AVX2 only tg128 9.16 / 9.26 9.27 / 9.34",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30231",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30231",
   "tg_tokens": 128,
   "tps": "9.16"
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "Xeon 6985P-C",
   "id": "lc-30231-gemma4e4b-avx2-branch-tg128",
   "model": "Gemma 4 E4B",
   "notes": "CPU, Xeon 6985P-C (Granite Rapids); vectorized GEGLU/GELU table lookup; -t 8; first of two runs shown",
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Intel Xeon 6985P-C (Granite Rapids), GCP c4-highcpu-16 with all-core max turbo, GCC 14.2, master 08246a28f.; llama-bench, Gemma 4 E4B Q8_0, -t 8 -p 512 -n 128 -r 5, t/s, two runs each; AVX2 only tg128 9.16 / 9.26 9.27 / 9.34",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30231",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30231",
   "tg_tokens": 128,
   "tps": "9.27"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 2048,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "DGX Spark",
   "id": "lc-30215-qw3827b-base-pp512",
   "model": "Qwen3.8-27B",
   "notes": "Vulkan, DGX Spark GB10; L2-friendly workgroup scheduling for cm2 mulmat; llama-bench -fa 1 -n 0 -b 2048 -ub 2048 -r 3; base(master) vs remap(PR)",
   "pp_tokens": 512,
   "pp_tps": "605.0",
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "DGX Spark GB10 tests; Qwen3.8-27B Q6_K pp512 605.0 693.8",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30215",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30215"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 2048,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "DGX Spark",
   "id": "lc-30215-qw3827b-remap-pp512",
   "model": "Qwen3.8-27B",
   "notes": "Vulkan, DGX Spark GB10; L2-friendly workgroup scheduling for cm2 mulmat; llama-bench -fa 1 -n 0 -b 2048 -ub 2048 -r 3; base(master) vs remap(PR)",
   "pp_tokens": 512,
   "pp_tps": "693.8",
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "DGX Spark GB10 tests; Qwen3.8-27B Q6_K pp512 605.0 693.8",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30215",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30215"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 2048,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "DGX Spark",
   "id": "lc-30215-qw3827b-base-pp2048",
   "model": "Qwen3.8-27B",
   "notes": "Vulkan, DGX Spark GB10; L2-friendly workgroup scheduling for cm2 mulmat; llama-bench -fa 1 -n 0 -b 2048 -ub 2048 -r 3; base(master) vs remap(PR)",
   "pp_tokens": 2048,
   "pp_tps": "602.8",
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "DGX Spark GB10 tests; Qwen3.8-27B Q6_K pp2048 602.8 731.5",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30215",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30215"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 2048,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "DGX Spark",
   "id": "lc-30215-qw3827b-remap-pp2048",
   "model": "Qwen3.8-27B",
   "notes": "Vulkan, DGX Spark GB10; L2-friendly workgroup scheduling for cm2 mulmat; llama-bench -fa 1 -n 0 -b 2048 -ub 2048 -r 3; base(master) vs remap(PR)",
   "pp_tokens": 2048,
   "pp_tps": "731.5",
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "DGX Spark GB10 tests; Qwen3.8-27B Q6_K pp2048 602.8 731.5",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30215",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30215"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 2048,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "DGX Spark",
   "id": "lc-30215-llama318b-base-pp512",
   "model": "Llama 3.1 8B",
   "notes": "Vulkan, DGX Spark GB10; L2-friendly workgroup scheduling for cm2 mulmat; llama-bench -fa 1 -n 0 -b 2048 -ub 2048 -r 3; base(master) vs remap(PR)",
   "pp_tokens": 512,
   "pp_tps": "2238.9",
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "DGX Spark GB10 tests; Llama-3.1-8B Q6_K pp512 2238.9 2419.3",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30215",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30215"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 2048,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "DGX Spark",
   "id": "lc-30215-llama318b-remap-pp512",
   "model": "Llama 3.1 8B",
   "notes": "Vulkan, DGX Spark GB10; L2-friendly workgroup scheduling for cm2 mulmat; llama-bench -fa 1 -n 0 -b 2048 -ub 2048 -r 3; base(master) vs remap(PR)",
   "pp_tokens": 512,
   "pp_tps": "2419.3",
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "DGX Spark GB10 tests; Llama-3.1-8B Q6_K pp512 2238.9 2419.3",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30215",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30215"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 2048,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "DGX Spark",
   "id": "lc-30215-llama318b-base-pp2048",
   "model": "Llama 3.1 8B",
   "notes": "Vulkan, DGX Spark GB10; L2-friendly workgroup scheduling for cm2 mulmat; llama-bench -fa 1 -n 0 -b 2048 -ub 2048 -r 3; base(master) vs remap(PR)",
   "pp_tokens": 2048,
   "pp_tps": "2316.6",
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "DGX Spark GB10 tests; Llama-3.1-8B Q6_K pp2048 2316.6 2688.4",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30215",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30215"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 2048,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "DGX Spark",
   "id": "lc-30215-llama318b-remap-pp2048",
   "model": "Llama 3.1 8B",
   "notes": "Vulkan, DGX Spark GB10; L2-friendly workgroup scheduling for cm2 mulmat; llama-bench -fa 1 -n 0 -b 2048 -ub 2048 -r 3; base(master) vs remap(PR)",
   "pp_tokens": 2048,
   "pp_tps": "2688.4",
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "DGX Spark GB10 tests; Llama-3.1-8B Q6_K pp2048 2316.6 2688.4",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30215",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30215"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 2048,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "DGX Spark",
   "id": "lc-30215-gemma31b-base-pp512",
   "model": "Gemma-31B",
   "notes": "Vulkan, DGX Spark GB10; L2-friendly workgroup scheduling for cm2 mulmat; llama-bench -fa 1 -n 0 -b 2048 -ub 2048 -r 3; base(master) vs remap(PR)",
   "pp_tokens": 512,
   "pp_tps": "646.7",
   "provenance": "community",
   "quant": "Q5_K_M",
   "quote": "DGX Spark GB10 tests; Gemma-31B Q5_K_M pp512 646.7 726.2",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30215",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30215"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 2048,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "DGX Spark",
   "id": "lc-30215-gemma31b-remap-pp512",
   "model": "Gemma-31B",
   "notes": "Vulkan, DGX Spark GB10; L2-friendly workgroup scheduling for cm2 mulmat; llama-bench -fa 1 -n 0 -b 2048 -ub 2048 -r 3; base(master) vs remap(PR)",
   "pp_tokens": 512,
   "pp_tps": "726.2",
   "provenance": "community",
   "quant": "Q5_K_M",
   "quote": "DGX Spark GB10 tests; Gemma-31B Q5_K_M pp512 646.7 726.2",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30215",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30215"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 2048,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "DGX Spark",
   "id": "lc-30215-gemma31b-base-pp2048",
   "model": "Gemma-31B",
   "notes": "Vulkan, DGX Spark GB10; L2-friendly workgroup scheduling for cm2 mulmat; llama-bench -fa 1 -n 0 -b 2048 -ub 2048 -r 3; base(master) vs remap(PR)",
   "pp_tokens": 2048,
   "pp_tps": "636.9",
   "provenance": "community",
   "quant": "Q5_K_M",
   "quote": "DGX Spark GB10 tests; Gemma-31B Q5_K_M pp2048 636.9 745.1",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30215",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30215"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 2048,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "DGX Spark",
   "id": "lc-30215-gemma31b-remap-pp2048",
   "model": "Gemma-31B",
   "notes": "Vulkan, DGX Spark GB10; L2-friendly workgroup scheduling for cm2 mulmat; llama-bench -fa 1 -n 0 -b 2048 -ub 2048 -r 3; base(master) vs remap(PR)",
   "pp_tokens": 2048,
   "pp_tps": "745.1",
   "provenance": "community",
   "quant": "Q5_K_M",
   "quote": "DGX Spark GB10 tests; Gemma-31B Q5_K_M pp2048 636.9 745.1",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30215",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30215"
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "B70",
   "id": "lc-30226-museglimmer-before-pp4",
   "model": "Muse Glimmer",
   "notes": "SYCL, B70 (Arc Pro B70); Q4_K MMVQ wide row pairing; target only, -fa 1 -r 5; before vs after",
   "pp_tokens": 4,
   "pp_tps": "75.71",
   "provenance": "community",
   "quant": "Q4_K_L",
   "quote": "llama-bench B70, Q4_K_L, target only, fusion at default, -fa 1 -r 5, tok/s; on the B70 with Muse-Glimmer Q4_K_L with DFlash2; before 75.71 77.23 80.85 82.74 87.53",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30226",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30226"
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "B70",
   "id": "lc-30226-museglimmer-after-pp4",
   "model": "Muse Glimmer",
   "notes": "SYCL, B70 (Arc Pro B70); Q4_K MMVQ wide row pairing; target only, -fa 1 -r 5; before vs after",
   "pp_tokens": 4,
   "pp_tps": "75.61",
   "provenance": "community",
   "quant": "Q4_K_L",
   "quote": "llama-bench B70, Q4_K_L, target only, fusion at default, -fa 1 -r 5, tok/s; on the B70 with Muse-Glimmer Q4_K_L with DFlash2; after 75.61 83.72 86.17 88.41 95.43",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30226",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30226"
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "B70",
   "id": "lc-30226-museglimmer-before-pp5",
   "model": "Muse Glimmer",
   "notes": "SYCL, B70 (Arc Pro B70); Q4_K MMVQ wide row pairing; target only, -fa 1 -r 5; before vs after",
   "pp_tokens": 5,
   "pp_tps": "77.23",
   "provenance": "community",
   "quant": "Q4_K_L",
   "quote": "llama-bench B70, Q4_K_L, target only, fusion at default, -fa 1 -r 5, tok/s; on the B70 with Muse-Glimmer Q4_K_L with DFlash2; before 75.71 77.23 80.85 82.74 87.53",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30226",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30226"
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "B70",
   "id": "lc-30226-museglimmer-after-pp5",
   "model": "Muse Glimmer",
   "notes": "SYCL, B70 (Arc Pro B70); Q4_K MMVQ wide row pairing; target only, -fa 1 -r 5; before vs after",
   "pp_tokens": 5,
   "pp_tps": "83.72",
   "provenance": "community",
   "quant": "Q4_K_L",
   "quote": "llama-bench B70, Q4_K_L, target only, fusion at default, -fa 1 -r 5, tok/s; on the B70 with Muse-Glimmer Q4_K_L with DFlash2; after 75.61 83.72 86.17 88.41 95.43",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30226",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30226"
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "B70",
   "id": "lc-30226-museglimmer-before-pp6",
   "model": "Muse Glimmer",
   "notes": "SYCL, B70 (Arc Pro B70); Q4_K MMVQ wide row pairing; target only, -fa 1 -r 5; before vs after",
   "pp_tokens": 6,
   "pp_tps": "80.85",
   "provenance": "community",
   "quant": "Q4_K_L",
   "quote": "llama-bench B70, Q4_K_L, target only, fusion at default, -fa 1 -r 5, tok/s; on the B70 with Muse-Glimmer Q4_K_L with DFlash2; before 75.71 77.23 80.85 82.74 87.53",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30226",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30226"
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "B70",
   "id": "lc-30226-museglimmer-after-pp6",
   "model": "Muse Glimmer",
   "notes": "SYCL, B70 (Arc Pro B70); Q4_K MMVQ wide row pairing; target only, -fa 1 -r 5; before vs after",
   "pp_tokens": 6,
   "pp_tps": "86.17",
   "provenance": "community",
   "quant": "Q4_K_L",
   "quote": "llama-bench B70, Q4_K_L, target only, fusion at default, -fa 1 -r 5, tok/s; on the B70 with Muse-Glimmer Q4_K_L with DFlash2; after 75.61 83.72 86.17 88.41 95.43",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30226",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30226"
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "B70",
   "id": "lc-30226-museglimmer-before-pp7",
   "model": "Muse Glimmer",
   "notes": "SYCL, B70 (Arc Pro B70); Q4_K MMVQ wide row pairing; target only, -fa 1 -r 5; before vs after",
   "pp_tokens": 7,
   "pp_tps": "82.74",
   "provenance": "community",
   "quant": "Q4_K_L",
   "quote": "llama-bench B70, Q4_K_L, target only, fusion at default, -fa 1 -r 5, tok/s; on the B70 with Muse-Glimmer Q4_K_L with DFlash2; before 75.71 77.23 80.85 82.74 87.53",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30226",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30226"
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "B70",
   "id": "lc-30226-museglimmer-after-pp7",
   "model": "Muse Glimmer",
   "notes": "SYCL, B70 (Arc Pro B70); Q4_K MMVQ wide row pairing; target only, -fa 1 -r 5; before vs after",
   "pp_tokens": 7,
   "pp_tps": "88.41",
   "provenance": "community",
   "quant": "Q4_K_L",
   "quote": "llama-bench B70, Q4_K_L, target only, fusion at default, -fa 1 -r 5, tok/s; on the B70 with Muse-Glimmer Q4_K_L with DFlash2; after 75.61 83.72 86.17 88.41 95.43",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30226",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30226"
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "B70",
   "id": "lc-30226-museglimmer-before-pp8",
   "model": "Muse Glimmer",
   "notes": "SYCL, B70 (Arc Pro B70); Q4_K MMVQ wide row pairing; target only, -fa 1 -r 5; before vs after",
   "pp_tokens": 8,
   "pp_tps": "87.53",
   "provenance": "community",
   "quant": "Q4_K_L",
   "quote": "llama-bench B70, Q4_K_L, target only, fusion at default, -fa 1 -r 5, tok/s; on the B70 with Muse-Glimmer Q4_K_L with DFlash2; before 75.71 77.23 80.85 82.74 87.53",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30226",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30226"
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "B70",
   "id": "lc-30226-museglimmer-after-pp8",
   "model": "Muse Glimmer",
   "notes": "SYCL, B70 (Arc Pro B70); Q4_K MMVQ wide row pairing; target only, -fa 1 -r 5; before vs after",
   "pp_tokens": 8,
   "pp_tps": "95.43",
   "provenance": "community",
   "quant": "Q4_K_L",
   "quote": "llama-bench B70, Q4_K_L, target only, fusion at default, -fa 1 -r 5, tok/s; on the B70 with Muse-Glimmer Q4_K_L with DFlash2; after 75.61 83.72 86.17 88.41 95.43",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30226",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30226"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Arc B580",
   "id": "lc-30503-master-pp16384",
   "model": "Gemma 4 E4B",
   "notes": "Vulkan, Intel Arc B580 (driver 32.0.101.9033); coopmat FA cm1 shader never selected (driver exposes 8x16x16, shape_ok requires 16x16x16); pp16384; fa=off ceiling 3523 t/s stock master",
   "pp_tokens": "16384",
   "pp_tps": "896",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Observed on Intel Arc B580; gemma-4-E4B Q4_0 -t 14 -ub 512 -r 3 llama-bench; stock master 896 t/s",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30503",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30503"
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Arc B580",
   "id": "lc-30503-gaterelaxed-pp16384",
   "model": "Gemma 4 E4B",
   "notes": "Vulkan, Intel Arc B580 (driver 32.0.101.9033); coopmat FA cm1 shader never selected (driver exposes 8x16x16, shape_ok requires 16x16x16); pp16384; fa=off ceiling 3523 t/s gate relaxed (accept 8x16x16)",
   "pp_tokens": "16384",
   "pp_tps": "2731",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Observed on Intel Arc B580; gemma-4-E4B Q4_0 -t 14 -ub 512 -r 3 llama-bench; gate relaxed (accept 8x16x16) 2731 t/s",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30503",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30503"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 0,
   "date": "2026-10",
   "flags": [],
   "hardware": "Tesla P40",
   "id": "lc-30334-mx87bq2k-stock-tg128-d0",
   "model": "Mixtral 8x7B",
   "notes": "P40 (cc 6.1) MMVQ fusion A/B; stock(fusion on) vs MMVQ_TEST_NO_FUSE; tg128 median of 6, -ngl 99 -fa 1",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Tesla P40, compute capability 6.1; Mixtral 8x7B; Fusion off against stock, tg128; 32.42 -> 30.65, -5.46%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30334",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30334",
   "tg_tokens": 128,
   "tps": "32.42"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 0,
   "date": "2026-10",
   "flags": [],
   "hardware": "Tesla P40",
   "id": "lc-30334-mx87bq2k-nofuse-tg128-d0",
   "model": "Mixtral 8x7B",
   "notes": "P40 (cc 6.1) MMVQ fusion A/B; stock(fusion on) vs MMVQ_TEST_NO_FUSE; tg128 median of 6, -ngl 99 -fa 1",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Tesla P40, compute capability 6.1; Mixtral 8x7B; Fusion off against stock, tg128; 32.42 -> 30.65, -5.46%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30334",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30334",
   "tg_tokens": 128,
   "tps": "30.65"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 496,
   "date": "2026-10",
   "flags": [],
   "hardware": "Tesla P40",
   "id": "lc-30334-mx87bq2k-stock-tg128-d496",
   "model": "Mixtral 8x7B",
   "notes": "P40 (cc 6.1) MMVQ fusion A/B; stock(fusion on) vs MMVQ_TEST_NO_FUSE; tg128 median of 6, -ngl 99 -fa 1",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Tesla P40, compute capability 6.1; Mixtral 8x7B; Fusion off against stock, tg128; 32.03 -> 30.30, -5.41%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30334",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30334",
   "tg_tokens": 128,
   "tps": "32.03"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 496,
   "date": "2026-10",
   "flags": [],
   "hardware": "Tesla P40",
   "id": "lc-30334-mx87bq2k-nofuse-tg128-d496",
   "model": "Mixtral 8x7B",
   "notes": "P40 (cc 6.1) MMVQ fusion A/B; stock(fusion on) vs MMVQ_TEST_NO_FUSE; tg128 median of 6, -ngl 99 -fa 1",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Tesla P40, compute capability 6.1; Mixtral 8x7B; Fusion off against stock, tg128; 32.03 -> 30.30, -5.41%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30334",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30334",
   "tg_tokens": 128,
   "tps": "30.30"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 2016,
   "date": "2026-10",
   "flags": [],
   "hardware": "Tesla P40",
   "id": "lc-30334-mx87bq2k-stock-tg128-d2016",
   "model": "Mixtral 8x7B",
   "notes": "P40 (cc 6.1) MMVQ fusion A/B; stock(fusion on) vs MMVQ_TEST_NO_FUSE; tg128 median of 6, -ngl 99 -fa 1",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Tesla P40, compute capability 6.1; Mixtral 8x7B; Fusion off against stock, tg128; 31.37 -> 29.69, -5.35%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30334",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30334",
   "tg_tokens": 128,
   "tps": "31.37"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 2016,
   "date": "2026-10",
   "flags": [],
   "hardware": "Tesla P40",
   "id": "lc-30334-mx87bq2k-nofuse-tg128-d2016",
   "model": "Mixtral 8x7B",
   "notes": "P40 (cc 6.1) MMVQ fusion A/B; stock(fusion on) vs MMVQ_TEST_NO_FUSE; tg128 median of 6, -ngl 99 -fa 1",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Tesla P40, compute capability 6.1; Mixtral 8x7B; Fusion off against stock, tg128; 31.37 -> 29.69, -5.35%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30334",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30334",
   "tg_tokens": 128,
   "tps": "29.69"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 0,
   "date": "2026-10",
   "flags": [],
   "hardware": "Tesla P40",
   "id": "lc-30334-mx87bq3km-stock-tg128-d0",
   "model": "Mixtral 8x7B",
   "notes": "P40 (cc 6.1) MMVQ fusion A/B; stock(fusion on) vs MMVQ_TEST_NO_FUSE; tg128 median of 6, -ngl 99 -fa 1",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q3_K_M",
   "quote": "Tesla P40, compute capability 6.1; Mixtral 8x7B; Fusion off against stock, tg128; 25.76 -> 27.12, +5.29%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30334",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30334",
   "tg_tokens": 128,
   "tps": "25.76"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 0,
   "date": "2026-10",
   "flags": [],
   "hardware": "Tesla P40",
   "id": "lc-30334-mx87bq3km-nofuse-tg128-d0",
   "model": "Mixtral 8x7B",
   "notes": "P40 (cc 6.1) MMVQ fusion A/B; stock(fusion on) vs MMVQ_TEST_NO_FUSE; tg128 median of 6, -ngl 99 -fa 1",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q3_K_M",
   "quote": "Tesla P40, compute capability 6.1; Mixtral 8x7B; Fusion off against stock, tg128; 25.76 -> 27.12, +5.29%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30334",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30334",
   "tg_tokens": 128,
   "tps": "27.12"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 496,
   "date": "2026-10",
   "flags": [],
   "hardware": "Tesla P40",
   "id": "lc-30334-mx87bq3km-stock-tg128-d496",
   "model": "Mixtral 8x7B",
   "notes": "P40 (cc 6.1) MMVQ fusion A/B; stock(fusion on) vs MMVQ_TEST_NO_FUSE; tg128 median of 6, -ngl 99 -fa 1",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q3_K_M",
   "quote": "Tesla P40, compute capability 6.1; Mixtral 8x7B; Fusion off against stock, tg128; 25.54 -> 26.83, +5.06%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30334",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30334",
   "tg_tokens": 128,
   "tps": "25.54"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 496,
   "date": "2026-10",
   "flags": [],
   "hardware": "Tesla P40",
   "id": "lc-30334-mx87bq3km-nofuse-tg128-d496",
   "model": "Mixtral 8x7B",
   "notes": "P40 (cc 6.1) MMVQ fusion A/B; stock(fusion on) vs MMVQ_TEST_NO_FUSE; tg128 median of 6, -ngl 99 -fa 1",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q3_K_M",
   "quote": "Tesla P40, compute capability 6.1; Mixtral 8x7B; Fusion off against stock, tg128; 25.54 -> 26.83, +5.06%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30334",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30334",
   "tg_tokens": 128,
   "tps": "26.83"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 2016,
   "date": "2026-10",
   "flags": [],
   "hardware": "Tesla P40",
   "id": "lc-30334-mx87bq3km-stock-tg128-d2016",
   "model": "Mixtral 8x7B",
   "notes": "P40 (cc 6.1) MMVQ fusion A/B; stock(fusion on) vs MMVQ_TEST_NO_FUSE; tg128 median of 6, -ngl 99 -fa 1",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q3_K_M",
   "quote": "Tesla P40, compute capability 6.1; Mixtral 8x7B; Fusion off against stock, tg128; 25.06 -> 26.35, +5.15%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30334",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30334",
   "tg_tokens": 128,
   "tps": "25.06"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 2016,
   "date": "2026-10",
   "flags": [],
   "hardware": "Tesla P40",
   "id": "lc-30334-mx87bq3km-nofuse-tg128-d2016",
   "model": "Mixtral 8x7B",
   "notes": "P40 (cc 6.1) MMVQ fusion A/B; stock(fusion on) vs MMVQ_TEST_NO_FUSE; tg128 median of 6, -ngl 99 -fa 1",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q3_K_M",
   "quote": "Tesla P40, compute capability 6.1; Mixtral 8x7B; Fusion off against stock, tg128; 25.06 -> 26.35, +5.15%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30334",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30334",
   "tg_tokens": 128,
   "tps": "26.35"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 0,
   "date": "2026-10",
   "flags": [],
   "hardware": "Tesla P40",
   "id": "lc-30334-ll318bq40-stock-tg128-d0",
   "model": "Llama 3.1 8B",
   "notes": "P40 (cc 6.1) MMVQ fusion A/B; stock(fusion on) vs MMVQ_TEST_NO_FUSE; tg128 median of 6, -ngl 99 -fa 1",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Tesla P40, compute capability 6.1; Llama 3.1 8B; Fusion off against stock, tg128; 52.28 -> 54.75, +4.72%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30334",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30334",
   "tg_tokens": 128,
   "tps": "52.28"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 0,
   "date": "2026-10",
   "flags": [],
   "hardware": "Tesla P40",
   "id": "lc-30334-ll318bq40-nofuse-tg128-d0",
   "model": "Llama 3.1 8B",
   "notes": "P40 (cc 6.1) MMVQ fusion A/B; stock(fusion on) vs MMVQ_TEST_NO_FUSE; tg128 median of 6, -ngl 99 -fa 1",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Tesla P40, compute capability 6.1; Llama 3.1 8B; Fusion off against stock, tg128; 52.28 -> 54.75, +4.72%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30334",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30334",
   "tg_tokens": 128,
   "tps": "54.75"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 496,
   "date": "2026-10",
   "flags": [],
   "hardware": "Tesla P40",
   "id": "lc-30334-ll318bq40-stock-tg128-d496",
   "model": "Llama 3.1 8B",
   "notes": "P40 (cc 6.1) MMVQ fusion A/B; stock(fusion on) vs MMVQ_TEST_NO_FUSE; tg128 median of 6, -ngl 99 -fa 1",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Tesla P40, compute capability 6.1; Llama 3.1 8B; Fusion off against stock, tg128; 51.39 -> 53.69, +4.47%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30334",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30334",
   "tg_tokens": 128,
   "tps": "51.39"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 496,
   "date": "2026-10",
   "flags": [],
   "hardware": "Tesla P40",
   "id": "lc-30334-ll318bq40-nofuse-tg128-d496",
   "model": "Llama 3.1 8B",
   "notes": "P40 (cc 6.1) MMVQ fusion A/B; stock(fusion on) vs MMVQ_TEST_NO_FUSE; tg128 median of 6, -ngl 99 -fa 1",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Tesla P40, compute capability 6.1; Llama 3.1 8B; Fusion off against stock, tg128; 51.39 -> 53.69, +4.47%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30334",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30334",
   "tg_tokens": 128,
   "tps": "53.69"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 2016,
   "date": "2026-10",
   "flags": [],
   "hardware": "Tesla P40",
   "id": "lc-30334-ll318bq40-stock-tg128-d2016",
   "model": "Llama 3.1 8B",
   "notes": "P40 (cc 6.1) MMVQ fusion A/B; stock(fusion on) vs MMVQ_TEST_NO_FUSE; tg128 median of 6, -ngl 99 -fa 1",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Tesla P40, compute capability 6.1; Llama 3.1 8B; Fusion off against stock, tg128; 49.72 -> 51.87, +4.33%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30334",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30334",
   "tg_tokens": 128,
   "tps": "49.72"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 2016,
   "date": "2026-10",
   "flags": [],
   "hardware": "Tesla P40",
   "id": "lc-30334-ll318bq40-nofuse-tg128-d2016",
   "model": "Llama 3.1 8B",
   "notes": "P40 (cc 6.1) MMVQ fusion A/B; stock(fusion on) vs MMVQ_TEST_NO_FUSE; tg128 median of 6, -ngl 99 -fa 1",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Tesla P40, compute capability 6.1; Llama 3.1 8B; Fusion off against stock, tg128; 49.72 -> 51.87, +4.33%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30334",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30334",
   "tg_tokens": 128,
   "tps": "51.87"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 0,
   "date": "2026-10",
   "flags": [],
   "hardware": "Tesla P40",
   "id": "lc-30334-ll318bq4km-stock-tg128-d0",
   "model": "Llama 3.1 8B",
   "notes": "P40 (cc 6.1) MMVQ fusion A/B; stock(fusion on) vs MMVQ_TEST_NO_FUSE; tg128 median of 6, -ngl 99 -fa 1",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Tesla P40, compute capability 6.1; Llama 3.1 8B; Fusion off against stock, tg128; 48.94 -> 48.34, -1.21%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30334",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30334",
   "tg_tokens": 128,
   "tps": "48.94"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 0,
   "date": "2026-10",
   "flags": [],
   "hardware": "Tesla P40",
   "id": "lc-30334-ll318bq4km-nofuse-tg128-d0",
   "model": "Llama 3.1 8B",
   "notes": "P40 (cc 6.1) MMVQ fusion A/B; stock(fusion on) vs MMVQ_TEST_NO_FUSE; tg128 median of 6, -ngl 99 -fa 1",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Tesla P40, compute capability 6.1; Llama 3.1 8B; Fusion off against stock, tg128; 48.94 -> 48.34, -1.21%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30334",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30334",
   "tg_tokens": 128,
   "tps": "48.34"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 496,
   "date": "2026-10",
   "flags": [],
   "hardware": "Tesla P40",
   "id": "lc-30334-ll318bq4km-stock-tg128-d496",
   "model": "Llama 3.1 8B",
   "notes": "P40 (cc 6.1) MMVQ fusion A/B; stock(fusion on) vs MMVQ_TEST_NO_FUSE; tg128 median of 6, -ngl 99 -fa 1",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Tesla P40, compute capability 6.1; Llama 3.1 8B; Fusion off against stock, tg128; 48.04 -> 47.47, -1.19%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30334",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30334",
   "tg_tokens": 128,
   "tps": "48.04"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 496,
   "date": "2026-10",
   "flags": [],
   "hardware": "Tesla P40",
   "id": "lc-30334-ll318bq4km-nofuse-tg128-d496",
   "model": "Llama 3.1 8B",
   "notes": "P40 (cc 6.1) MMVQ fusion A/B; stock(fusion on) vs MMVQ_TEST_NO_FUSE; tg128 median of 6, -ngl 99 -fa 1",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Tesla P40, compute capability 6.1; Llama 3.1 8B; Fusion off against stock, tg128; 48.04 -> 47.47, -1.19%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30334",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30334",
   "tg_tokens": 128,
   "tps": "47.47"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 2016,
   "date": "2026-10",
   "flags": [],
   "hardware": "Tesla P40",
   "id": "lc-30334-ll318bq4km-stock-tg128-d2016",
   "model": "Llama 3.1 8B",
   "notes": "P40 (cc 6.1) MMVQ fusion A/B; stock(fusion on) vs MMVQ_TEST_NO_FUSE; tg128 median of 6, -ngl 99 -fa 1",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Tesla P40, compute capability 6.1; Llama 3.1 8B; Fusion off against stock, tg128; 46.53 -> 46.05, -1.04%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30334",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30334",
   "tg_tokens": 128,
   "tps": "46.53"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": 2016,
   "date": "2026-10",
   "flags": [],
   "hardware": "Tesla P40",
   "id": "lc-30334-ll318bq4km-nofuse-tg128-d2016",
   "model": "Llama 3.1 8B",
   "notes": "P40 (cc 6.1) MMVQ fusion A/B; stock(fusion on) vs MMVQ_TEST_NO_FUSE; tg128 median of 6, -ngl 99 -fa 1",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Tesla P40, compute capability 6.1; Llama 3.1 8B; Fusion off against stock, tg128; 46.53 -> 46.05, -1.04%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30334",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30334",
   "tg_tokens": 128,
   "tps": "46.05"
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Tesla P40",
   "id": "lc-30334-mx87bq2k-stock-pp512-ub1",
   "model": "Mixtral 8x7B",
   "notes": "P40 (cc 6.1) MMVQ fusion A/B; stock(fusion on) vs MMVQ_TEST_NO_FUSE; pp512 mean of 6, -ngl 99 -fa 1",
   "pp_tokens": 512,
   "pp_tps": "33.71",
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Tesla P40, compute capability 6.1; Mixtral 8x7B; fusion off at -ub 1, 2, 4, 8, pp512; 33.71 -> 31.82, -5.62%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30334",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30334",
   "tg_tokens": null,
   "tps": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Tesla P40",
   "id": "lc-30334-mx87bq2k-nofuse-pp512-ub1",
   "model": "Mixtral 8x7B",
   "notes": "P40 (cc 6.1) MMVQ fusion A/B; stock(fusion on) vs MMVQ_TEST_NO_FUSE; pp512 mean of 6, -ngl 99 -fa 1",
   "pp_tokens": 512,
   "pp_tps": "31.82",
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Tesla P40, compute capability 6.1; Mixtral 8x7B; fusion off at -ub 1, 2, 4, 8, pp512; 33.71 -> 31.82, -5.62%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30334",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30334",
   "tg_tokens": null,
   "tps": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 2,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Tesla P40",
   "id": "lc-30334-mx87bq2k-stock-pp512-ub2",
   "model": "Mixtral 8x7B",
   "notes": "P40 (cc 6.1) MMVQ fusion A/B; stock(fusion on) vs MMVQ_TEST_NO_FUSE; pp512 mean of 6, -ngl 99 -fa 1",
   "pp_tokens": 512,
   "pp_tps": "46.65",
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Tesla P40, compute capability 6.1; Mixtral 8x7B; fusion off at -ub 1, 2, 4, 8, pp512; 46.65 -> 45.35, -2.78%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30334",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30334",
   "tg_tokens": null,
   "tps": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 2,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Tesla P40",
   "id": "lc-30334-mx87bq2k-nofuse-pp512-ub2",
   "model": "Mixtral 8x7B",
   "notes": "P40 (cc 6.1) MMVQ fusion A/B; stock(fusion on) vs MMVQ_TEST_NO_FUSE; pp512 mean of 6, -ngl 99 -fa 1",
   "pp_tokens": 512,
   "pp_tps": "45.35",
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Tesla P40, compute capability 6.1; Mixtral 8x7B; fusion off at -ub 1, 2, 4, 8, pp512; 46.65 -> 45.35, -2.78%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30334",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30334",
   "tg_tokens": null,
   "tps": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 4,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Tesla P40",
   "id": "lc-30334-mx87bq2k-stock-pp512-ub4",
   "model": "Mixtral 8x7B",
   "notes": "P40 (cc 6.1) MMVQ fusion A/B; stock(fusion on) vs MMVQ_TEST_NO_FUSE; pp512 mean of 6, -ngl 99 -fa 1",
   "pp_tokens": 512,
   "pp_tps": "51.35",
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Tesla P40, compute capability 6.1; Mixtral 8x7B; fusion off at -ub 1, 2, 4, 8, pp512; 51.35 -> 50.64, -1.38%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30334",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30334",
   "tg_tokens": null,
   "tps": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 4,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Tesla P40",
   "id": "lc-30334-mx87bq2k-nofuse-pp512-ub4",
   "model": "Mixtral 8x7B",
   "notes": "P40 (cc 6.1) MMVQ fusion A/B; stock(fusion on) vs MMVQ_TEST_NO_FUSE; pp512 mean of 6, -ngl 99 -fa 1",
   "pp_tokens": 512,
   "pp_tps": "50.64",
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Tesla P40, compute capability 6.1; Mixtral 8x7B; fusion off at -ub 1, 2, 4, 8, pp512; 51.35 -> 50.64, -1.38%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30334",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30334",
   "tg_tokens": null,
   "tps": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 8,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Tesla P40",
   "id": "lc-30334-mx87bq2k-stock-pp512-ub8",
   "model": "Mixtral 8x7B",
   "notes": "P40 (cc 6.1) MMVQ fusion A/B; stock(fusion on) vs MMVQ_TEST_NO_FUSE; pp512 mean of 6, -ngl 99 -fa 1",
   "pp_tokens": 512,
   "pp_tps": "54.41",
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Tesla P40, compute capability 6.1; Mixtral 8x7B; fusion off at -ub 1, 2, 4, 8, pp512; 54.41 -> 54.33, -0.15%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30334",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30334",
   "tg_tokens": null,
   "tps": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 8,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Tesla P40",
   "id": "lc-30334-mx87bq2k-nofuse-pp512-ub8",
   "model": "Mixtral 8x7B",
   "notes": "P40 (cc 6.1) MMVQ fusion A/B; stock(fusion on) vs MMVQ_TEST_NO_FUSE; pp512 mean of 6, -ngl 99 -fa 1",
   "pp_tokens": 512,
   "pp_tps": "54.33",
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Tesla P40, compute capability 6.1; Mixtral 8x7B; fusion off at -ub 1, 2, 4, 8, pp512; 54.41 -> 54.33, -0.15%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30334",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30334",
   "tg_tokens": null,
   "tps": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Tesla P40",
   "id": "lc-30334-mx87bq3km-stock-pp512-ub1",
   "model": "Mixtral 8x7B",
   "notes": "P40 (cc 6.1) MMVQ fusion A/B; stock(fusion on) vs MMVQ_TEST_NO_FUSE; pp512 mean of 6, -ngl 99 -fa 1",
   "pp_tokens": 512,
   "pp_tps": "26.74",
   "provenance": "community",
   "quant": "Q3_K_M",
   "quote": "Tesla P40, compute capability 6.1; Mixtral 8x7B; fusion off at -ub 1, 2, 4, 8, pp512; 26.74 -> 28.14, +5.22%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30334",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30334",
   "tg_tokens": null,
   "tps": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Tesla P40",
   "id": "lc-30334-mx87bq3km-nofuse-pp512-ub1",
   "model": "Mixtral 8x7B",
   "notes": "P40 (cc 6.1) MMVQ fusion A/B; stock(fusion on) vs MMVQ_TEST_NO_FUSE; pp512 mean of 6, -ngl 99 -fa 1",
   "pp_tokens": 512,
   "pp_tps": "28.14",
   "provenance": "community",
   "quant": "Q3_K_M",
   "quote": "Tesla P40, compute capability 6.1; Mixtral 8x7B; fusion off at -ub 1, 2, 4, 8, pp512; 26.74 -> 28.14, +5.22%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30334",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30334",
   "tg_tokens": null,
   "tps": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 2,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Tesla P40",
   "id": "lc-30334-mx87bq3km-stock-pp512-ub2",
   "model": "Mixtral 8x7B",
   "notes": "P40 (cc 6.1) MMVQ fusion A/B; stock(fusion on) vs MMVQ_TEST_NO_FUSE; pp512 mean of 6, -ngl 99 -fa 1",
   "pp_tokens": 512,
   "pp_tps": "32.71",
   "provenance": "community",
   "quant": "Q3_K_M",
   "quote": "Tesla P40, compute capability 6.1; Mixtral 8x7B; fusion off at -ub 1, 2, 4, 8, pp512; 32.71 -> 36.54, +11.70%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30334",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30334",
   "tg_tokens": null,
   "tps": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 2,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Tesla P40",
   "id": "lc-30334-mx87bq3km-nofuse-pp512-ub2",
   "model": "Mixtral 8x7B",
   "notes": "P40 (cc 6.1) MMVQ fusion A/B; stock(fusion on) vs MMVQ_TEST_NO_FUSE; pp512 mean of 6, -ngl 99 -fa 1",
   "pp_tokens": 512,
   "pp_tps": "36.54",
   "provenance": "community",
   "quant": "Q3_K_M",
   "quote": "Tesla P40, compute capability 6.1; Mixtral 8x7B; fusion off at -ub 1, 2, 4, 8, pp512; 32.71 -> 36.54, +11.70%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30334",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30334",
   "tg_tokens": null,
   "tps": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 4,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Tesla P40",
   "id": "lc-30334-mx87bq3km-stock-pp512-ub4",
   "model": "Mixtral 8x7B",
   "notes": "P40 (cc 6.1) MMVQ fusion A/B; stock(fusion on) vs MMVQ_TEST_NO_FUSE; pp512 mean of 6, -ngl 99 -fa 1",
   "pp_tokens": 512,
   "pp_tps": "35.99",
   "provenance": "community",
   "quant": "Q3_K_M",
   "quote": "Tesla P40, compute capability 6.1; Mixtral 8x7B; fusion off at -ub 1, 2, 4, 8, pp512; 35.99 -> 40.58, +12.76%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30334",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30334",
   "tg_tokens": null,
   "tps": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 4,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Tesla P40",
   "id": "lc-30334-mx87bq3km-nofuse-pp512-ub4",
   "model": "Mixtral 8x7B",
   "notes": "P40 (cc 6.1) MMVQ fusion A/B; stock(fusion on) vs MMVQ_TEST_NO_FUSE; pp512 mean of 6, -ngl 99 -fa 1",
   "pp_tokens": 512,
   "pp_tps": "40.58",
   "provenance": "community",
   "quant": "Q3_K_M",
   "quote": "Tesla P40, compute capability 6.1; Mixtral 8x7B; fusion off at -ub 1, 2, 4, 8, pp512; 35.99 -> 40.58, +12.76%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30334",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30334",
   "tg_tokens": null,
   "tps": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 8,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Tesla P40",
   "id": "lc-30334-mx87bq3km-stock-pp512-ub8",
   "model": "Mixtral 8x7B",
   "notes": "P40 (cc 6.1) MMVQ fusion A/B; stock(fusion on) vs MMVQ_TEST_NO_FUSE; pp512 mean of 6, -ngl 99 -fa 1",
   "pp_tokens": 512,
   "pp_tps": "57.55",
   "provenance": "community",
   "quant": "Q3_K_M",
   "quote": "Tesla P40, compute capability 6.1; Mixtral 8x7B; fusion off at -ub 1, 2, 4, 8, pp512; 57.55 -> 57.53, -0.03%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30334",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30334",
   "tg_tokens": null,
   "tps": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": 8,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Tesla P40",
   "id": "lc-30334-mx87bq3km-nofuse-pp512-ub8",
   "model": "Mixtral 8x7B",
   "notes": "P40 (cc 6.1) MMVQ fusion A/B; stock(fusion on) vs MMVQ_TEST_NO_FUSE; pp512 mean of 6, -ngl 99 -fa 1",
   "pp_tokens": 512,
   "pp_tps": "57.53",
   "provenance": "community",
   "quant": "Q3_K_M",
   "quote": "Tesla P40, compute capability 6.1; Mixtral 8x7B; fusion off at -ub 1, 2, 4, 8, pp512; 57.55 -> 57.53, -0.03%",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30334",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30334",
   "tg_tokens": null,
   "tps": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 32,
   "ctx": 0,
   "date": "2026-10",
   "flags": [],
   "hardware": "Radeon 8065S",
   "id": "lc-30364-mx27-vk-pp512-ub32-d0",
   "model": "MiniMax-M2.7",
   "notes": "Strix Halo gfx1151; Vulkan pp512 ubatch x depth sweep; depth 32768 collapses above ubatch 128; -p 512 -n 0 -d 0,32768 -ub <n> -r 1",
   "pp_tokens": 512,
   "pp_tps": "54.6",
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "Radeon 8065S (RADV STRIX_HALO, gfx1151); MiniMax-M2.7 UD-IQ4_XS; llama-bench, pp512, Vulkan (t/s); 32 | 54.6 | 42.6",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30364",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30364",
   "tg_tokens": null,
   "tps": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 32,
   "ctx": 32768,
   "date": "2026-10",
   "flags": [],
   "hardware": "Radeon 8065S",
   "id": "lc-30364-mx27-vk-pp512-ub32-d32768",
   "model": "MiniMax-M2.7",
   "notes": "Strix Halo gfx1151; Vulkan pp512 ubatch x depth sweep; depth 32768 collapses above ubatch 128; -p 512 -n 0 -d 0,32768 -ub <n> -r 1",
   "pp_tokens": 512,
   "pp_tps": "42.6",
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "Radeon 8065S (RADV STRIX_HALO, gfx1151); MiniMax-M2.7 UD-IQ4_XS; llama-bench, pp512, Vulkan (t/s); 32 | 54.6 | 42.6",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30364",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30364",
   "tg_tokens": null,
   "tps": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 64,
   "ctx": 0,
   "date": "2026-10",
   "flags": [],
   "hardware": "Radeon 8065S",
   "id": "lc-30364-mx27-vk-pp512-ub64-d0",
   "model": "MiniMax-M2.7",
   "notes": "Strix Halo gfx1151; Vulkan pp512 ubatch x depth sweep; depth 32768 collapses above ubatch 128; -p 512 -n 0 -d 0,32768 -ub <n> -r 1",
   "pp_tokens": 512,
   "pp_tps": "78.9",
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "Radeon 8065S (RADV STRIX_HALO, gfx1151); MiniMax-M2.7 UD-IQ4_XS; llama-bench, pp512, Vulkan (t/s); 64 | 78.9 | 55.7",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30364",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30364",
   "tg_tokens": null,
   "tps": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 64,
   "ctx": 32768,
   "date": "2026-10",
   "flags": [],
   "hardware": "Radeon 8065S",
   "id": "lc-30364-mx27-vk-pp512-ub64-d32768",
   "model": "MiniMax-M2.7",
   "notes": "Strix Halo gfx1151; Vulkan pp512 ubatch x depth sweep; depth 32768 collapses above ubatch 128; -p 512 -n 0 -d 0,32768 -ub <n> -r 1",
   "pp_tokens": 512,
   "pp_tps": "55.7",
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "Radeon 8065S (RADV STRIX_HALO, gfx1151); MiniMax-M2.7 UD-IQ4_XS; llama-bench, pp512, Vulkan (t/s); 64 | 78.9 | 55.7",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30364",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30364",
   "tg_tokens": null,
   "tps": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 128,
   "ctx": 0,
   "date": "2026-10",
   "flags": [],
   "hardware": "Radeon 8065S",
   "id": "lc-30364-mx27-vk-pp512-ub128-d0",
   "model": "MiniMax-M2.7",
   "notes": "Strix Halo gfx1151; Vulkan pp512 ubatch x depth sweep; depth 32768 collapses above ubatch 128; -p 512 -n 0 -d 0,32768 -ub <n> -r 1",
   "pp_tokens": 512,
   "pp_tps": "124.0",
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "Radeon 8065S (RADV STRIX_HALO, gfx1151); MiniMax-M2.7 UD-IQ4_XS; llama-bench, pp512, Vulkan (t/s); 128 | 124.0 | 77.5",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30364",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30364",
   "tg_tokens": null,
   "tps": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 128,
   "ctx": 32768,
   "date": "2026-10",
   "flags": [],
   "hardware": "Radeon 8065S",
   "id": "lc-30364-mx27-vk-pp512-ub128-d32768",
   "model": "MiniMax-M2.7",
   "notes": "Strix Halo gfx1151; Vulkan pp512 ubatch x depth sweep; depth 32768 collapses above ubatch 128; -p 512 -n 0 -d 0,32768 -ub <n> -r 1",
   "pp_tokens": 512,
   "pp_tps": "77.5",
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "Radeon 8065S (RADV STRIX_HALO, gfx1151); MiniMax-M2.7 UD-IQ4_XS; llama-bench, pp512, Vulkan (t/s); 128 | 124.0 | 77.5",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30364",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30364",
   "tg_tokens": null,
   "tps": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 256,
   "ctx": 32768,
   "date": "2026-10",
   "flags": [],
   "hardware": "Radeon 8065S",
   "id": "lc-30364-mx27-vk-pp512-ub256-d32768",
   "model": "MiniMax-M2.7",
   "notes": "Strix Halo gfx1151; Vulkan pp512 ubatch x depth sweep; depth 32768 collapses above ubatch 128; -p 512 -n 0 -d 0,32768 -ub <n> -r 1",
   "pp_tokens": 512,
   "pp_tps": "51.5",
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "Radeon 8065S (RADV STRIX_HALO, gfx1151); MiniMax-M2.7 UD-IQ4_XS; llama-bench, pp512, Vulkan (t/s); 256 | not tested | 51.5",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30364",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30364",
   "tg_tokens": null,
   "tps": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 512,
   "ctx": 0,
   "date": "2026-10",
   "flags": [],
   "hardware": "Radeon 8065S",
   "id": "lc-30364-mx27-vk-pp512-ub512-d0",
   "model": "MiniMax-M2.7",
   "notes": "Strix Halo gfx1151; Vulkan pp512 ubatch x depth sweep; depth 32768 collapses above ubatch 128; -p 512 -n 0 -d 0,32768 -ub <n> -r 1",
   "pp_tokens": 512,
   "pp_tps": "309.3",
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "Radeon 8065S (RADV STRIX_HALO, gfx1151); MiniMax-M2.7 UD-IQ4_XS; llama-bench, pp512, Vulkan (t/s); 512 (default) | 309.3 | 13.3",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30364",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30364",
   "tg_tokens": null,
   "tps": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 512,
   "ctx": 32768,
   "date": "2026-10",
   "flags": [],
   "hardware": "Radeon 8065S",
   "id": "lc-30364-mx27-vk-pp512-ub512-d32768",
   "model": "MiniMax-M2.7",
   "notes": "Strix Halo gfx1151; Vulkan pp512 ubatch x depth sweep; depth 32768 collapses above ubatch 128; -p 512 -n 0 -d 0,32768 -ub <n> -r 1",
   "pp_tokens": 512,
   "pp_tps": "13.3",
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "Radeon 8065S (RADV STRIX_HALO, gfx1151); MiniMax-M2.7 UD-IQ4_XS; llama-bench, pp512, Vulkan (t/s); 512 (default) | 309.3 | 13.3",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30364",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30364",
   "tg_tokens": null,
   "tps": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 128,
   "ctx": 0,
   "date": "2026-10",
   "flags": [],
   "hardware": "Radeon 8065S",
   "id": "lc-30364-mx27-rocm-pp512-d0",
   "model": "MiniMax-M2.7",
   "notes": "Strix Halo gfx1151; ROCm build (c83f3058) ubatch 128 reference; no depth-32768 collapse on ROCm",
   "pp_tokens": 512,
   "pp_tps": "246.4",
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "Radeon 8065S (RADV STRIX_HALO, gfx1151); MiniMax-M2.7 UD-IQ4_XS; ubatch 128: pp512 is 246.4 at depth 0 and 151.0 at depth 32768",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30364",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30364",
   "tg_tokens": null,
   "tps": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 128,
   "ctx": 32768,
   "date": "2026-10",
   "flags": [],
   "hardware": "Radeon 8065S",
   "id": "lc-30364-mx27-rocm-pp512-d32768",
   "model": "MiniMax-M2.7",
   "notes": "Strix Halo gfx1151; ROCm build (c83f3058) ubatch 128 reference; no depth-32768 collapse on ROCm",
   "pp_tokens": 512,
   "pp_tps": "151.0",
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "Radeon 8065S (RADV STRIX_HALO, gfx1151); MiniMax-M2.7 UD-IQ4_XS; ubatch 128: pp512 is 246.4 at depth 0 and 151.0 at depth 32768",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30364",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30364",
   "tg_tokens": null,
   "tps": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Arc A770",
   "id": "lc-30478-g22b-q4km-fused-pp512",
   "model": "gemma-2-2b-it",
   "notes": "Arc A770 SYCL; MMVQ GLU fusion on(default) vs off; unfused ~1.08x faster on tg128 (Q4_K_M)",
   "pp_tokens": 512,
   "pp_tps": "2666.46",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Arc A770 using the SYCL backend; gemma-2-2b-it-Q4_K_M.gguf; gemma2 2B Q4_K - Medium | 1.59 GiB | 2.61 B | SYCL | -1 | pp512 | 2666.46 \u00b1 23.11",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30478",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30478",
   "tg_tokens": null,
   "tps": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Arc A770",
   "id": "lc-30478-g22b-q4km-fused-tg128",
   "model": "gemma-2-2b-it",
   "notes": "Arc A770 SYCL; MMVQ GLU fusion on(default) vs off; unfused ~1.08x faster on tg128 (Q4_K_M)",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Arc A770 using the SYCL backend; gemma-2-2b-it-Q4_K_M.gguf; gemma2 2B Q4_K - Medium | 1.59 GiB | 2.61 B | SYCL | -1 | tg128 | 81.26 \u00b1 0.10",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30478",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30478",
   "tg_tokens": 128,
   "tps": "81.26"
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Arc A770",
   "id": "lc-30478-g22b-q4km-unfused-pp512",
   "model": "gemma-2-2b-it",
   "notes": "Arc A770 SYCL; MMVQ GLU fusion on(default) vs off; unfused ~1.08x faster on tg128 (Q4_K_M)",
   "pp_tokens": 512,
   "pp_tps": "2664.11",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Arc A770 using the SYCL backend; gemma-2-2b-it-Q4_K_M.gguf; gemma2 2B Q4_K - Medium | 1.59 GiB | 2.61 B | SYCL | -1 | pp512 | 2664.11 \u00b1 33.71",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30478",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30478",
   "tg_tokens": null,
   "tps": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Arc A770",
   "id": "lc-30478-g22b-q4km-unfused-tg128",
   "model": "gemma-2-2b-it",
   "notes": "Arc A770 SYCL; MMVQ GLU fusion on(default) vs off; unfused ~1.08x faster on tg128 (Q4_K_M)",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Arc A770 using the SYCL backend; gemma-2-2b-it-Q4_K_M.gguf; gemma2 2B Q4_K - Medium | 1.59 GiB | 2.61 B | SYCL | -1 | tg128 | 87.94 \u00b1 0.07",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30478",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30478",
   "tg_tokens": 128,
   "tps": "87.94"
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Arc A770",
   "id": "lc-30478-g22b-q2k-fused-pp512",
   "model": "gemma-2-2b-it",
   "notes": "Arc A770 SYCL; MMVQ GLU fusion on(default) vs off; unfused ~1.08x faster on tg128 (Q4_K_M)",
   "pp_tokens": 512,
   "pp_tps": "2613.42",
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Arc A770 using the SYCL backend; gemma-2-2b-it-Q4_K_M.gguf; gemma2 2B Q2_K - Medium | 1.14 GiB | 2.61 B | SYCL | -1 | pp512 | 2613.42 \u00b1 26.24",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30478",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30478",
   "tg_tokens": null,
   "tps": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Arc A770",
   "id": "lc-30478-g22b-q2k-fused-tg128",
   "model": "gemma-2-2b-it",
   "notes": "Arc A770 SYCL; MMVQ GLU fusion on(default) vs off; unfused ~1.08x faster on tg128 (Q4_K_M)",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Arc A770 using the SYCL backend; gemma-2-2b-it-Q4_K_M.gguf; gemma2 2B Q2_K - Medium | 1.14 GiB | 2.61 B | SYCL | -1 | tg128 | 102.18 \u00b1 0.06",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30478",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30478",
   "tg_tokens": 128,
   "tps": "102.18"
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Arc A770",
   "id": "lc-30478-g22b-q2k-unfused-pp512",
   "model": "gemma-2-2b-it",
   "notes": "Arc A770 SYCL; MMVQ GLU fusion on(default) vs off; unfused ~1.08x faster on tg128 (Q4_K_M)",
   "pp_tokens": 512,
   "pp_tps": "2608.42",
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Arc A770 using the SYCL backend; gemma-2-2b-it-Q4_K_M.gguf; gemma2 2B Q2_K - Medium | 1.14 GiB | 2.61 B | SYCL | -1 | pp512 | 2608.42 \u00b1 21.18",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30478",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30478",
   "tg_tokens": null,
   "tps": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": 1,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "Arc A770",
   "id": "lc-30478-g22b-q2k-unfused-tg128",
   "model": "gemma-2-2b-it",
   "notes": "Arc A770 SYCL; MMVQ GLU fusion on(default) vs off; unfused ~1.08x faster on tg128 (Q4_K_M)",
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "Arc A770 using the SYCL backend; gemma-2-2b-it-Q4_K_M.gguf; gemma2 2B Q2_K - Medium | 1.14 GiB | 2.61 B | SYCL | -1 | tg128 | 102.03 \u00b1 0.24",
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp #30478",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30478",
   "tg_tokens": 128,
   "tps": "102.03"
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "AMD Strix Halo",
   "id": "vllm-60413-strixhalo-qwen3827b-w4a8",
   "model": "Qwen3.8-27B",
   "notes": "vLLM #60413, Qwen3.8-27B-Quark-AWQ-MXFP4, AMD Strix Halo (gfx1151 RDNA 3.5), W4A8 int8-dot GEMV decode kernel (this PR), MXFP4 lm_head, MTP=3, single stream max-num-seqs 1, vllm bench serve random 128 in/256 out, decode tok/s, GSM8K 96.51%. config=w4a8-mtp3",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "AMD Strix Halo (gfx1151, RDNA 3.5), amd/Qwen3.8-27B-Quark-AWQ-MXFP4, single stream; This PR + MXFP4 lm_head W4A8 int8-dot GEMV 28.89",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM #60413",
   "source_url": "https://github.com/vllm-project/vllm/issues/60413",
   "tg_tokens": null,
   "tps": 28.89,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "AMD Strix Halo",
   "id": "vllm-60413-strixhalo-qwen3827b-w4a16",
   "model": "Qwen3.8-27B",
   "notes": "vLLM #60413, Qwen3.8-27B-Quark-AWQ-MXFP4, AMD Strix Halo (gfx1151 RDNA 3.5), W4A16 scalar FMA decode kernel (#46676 baseline, WMMA for M>8), MXFP4 lm_head, MTP=3, single stream max-num-seqs 1, vllm bench serve random 128 in/256 out, decode tok/s, GSM8K 96.51%. config=w4a16-mtp3",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "AMD Strix Halo (gfx1151, RDNA 3.5), amd/Qwen3.8-27B-Quark-AWQ-MXFP4, single stream; #46676 + MXFP4 lm_head W4A16 scalar FMA (WMMA for M > 8) 26.19",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM #60413",
   "source_url": "https://github.com/vllm-project/vllm/issues/60413",
   "tg_tokens": null,
   "tps": 26.19,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "AMD Strix Halo",
   "id": "vllm-60413-strixhalo-qwen3827b-emulation",
   "model": "Qwen3.8-27B",
   "notes": "vLLM #60413, Qwen3.8-27B-Quark-AWQ-MXFP4, AMD Strix Halo (gfx1151 RDNA 3.5), dequant + F.linear emulation path, MXFP4 lm_head, MTP=3, single stream max-num-seqs 1, vllm bench serve random 128 in/256 out, decode tok/s, GSM8K 96.75%. config=emulation-mtp3",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "AMD Strix Halo (gfx1151, RDNA 3.5), amd/Qwen3.8-27B-Quark-AWQ-MXFP4, single stream; Emulation + MXFP4 lm_head dequant + F.linear 4.94",
   "ram_gb": null,
   "retrieved": "2026-10-07",
   "source_name": "vLLM #60413",
   "source_url": "https://github.com/vllm-project/vllm/issues/60413",
   "tg_tokens": null,
   "tps": 4.94,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-61482-mi355x8-dsv4pro-monooff-c1",
   "model": "DeepSeek-V4-Pro",
   "notes": "InferenceX AgentX closed-loop replay (DSpark 6 drafts, synthetic acceptance 3.77), TP8 on MI355X (gfx950), FP8 preshuffled weights, mono decode off; P50 tok/s/user from one 900 s run per point; mono on arm 299 tok/s; aggregate output 87.8 tok/s; P90 227 \u2192 278 tok/s; median ITL 4.16 \u2192 3.34 ms; config=mono-off",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "FP8",
   "quote": "DeepSeek-V4-Pro; 8\u00d7 MI355X; FP8 weights; Interactivity P50 (tok/s/user) | 241 \u2192 **299** (+24%); c1 off \u2192 on",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61482",
   "source_url": "https://github.com/vllm-project/vllm/pull/61482",
   "tg_tokens": null,
   "tps": 241,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-61482-mi355x8-dsv4pro-monoon-c1",
   "model": "DeepSeek-V4-Pro",
   "notes": "InferenceX AgentX closed-loop replay (DSpark 6 drafts, synthetic acceptance 3.77), TP8 on MI355X (gfx950), FP8 preshuffled weights, mono decode on (--monokernel, shared FlyDSL kernels); P50 tok/s/user from one 900 s run per point; mono off arm 241 tok/s; aggregate output 93.8 tok/s; P90 227 \u2192 278 tok/s; median ITL 4.16 \u2192 3.34 ms; config=mono-on",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "FP8",
   "quote": "DeepSeek-V4-Pro; 8\u00d7 MI355X; FP8 weights; Interactivity P50 (tok/s/user) | 241 \u2192 **299** (+24%); c1 off \u2192 on",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61482",
   "source_url": "https://github.com/vllm-project/vllm/pull/61482",
   "tg_tokens": null,
   "tps": 299,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": null,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-61482-mi355x8-dsv4pro-monooff-c4",
   "model": "DeepSeek-V4-Pro",
   "notes": "InferenceX AgentX closed-loop replay (DSpark 6 drafts, synthetic acceptance 3.77), TP8 on MI355X (gfx950), FP8 preshuffled weights, mono decode off; P50 tok/s/user from one 900 s run per point; mono on arm 247 tok/s; aggregate output 144 tok/s; P90 169 \u2192 176 tok/s; median ITL 4.59 \u2192 4.04 ms; config=mono-off",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "FP8",
   "quote": "DeepSeek-V4-Pro; 8\u00d7 MI355X; FP8 weights; Interactivity P50 (tok/s/user); 218 \u2192 **247** (+13%); c4 off \u2192 on",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61482",
   "source_url": "https://github.com/vllm-project/vllm/pull/61482",
   "tg_tokens": null,
   "tps": 218,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": null,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-61482-mi355x8-dsv4pro-monoon-c4",
   "model": "DeepSeek-V4-Pro",
   "notes": "InferenceX AgentX closed-loop replay (DSpark 6 drafts, synthetic acceptance 3.77), TP8 on MI355X (gfx950), FP8 preshuffled weights, mono decode on (--monokernel, shared FlyDSL kernels); P50 tok/s/user from one 900 s run per point; mono off arm 218 tok/s; aggregate output 194 tok/s; P90 169 \u2192 176 tok/s; median ITL 4.59 \u2192 4.04 ms; config=mono-on",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "FP8",
   "quote": "DeepSeek-V4-Pro; 8\u00d7 MI355X; FP8 weights; Interactivity P50 (tok/s/user); 218 \u2192 **247** (+13%); c4 off \u2192 on",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61482",
   "source_url": "https://github.com/vllm-project/vllm/pull/61482",
   "tg_tokens": null,
   "tps": 247,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": null,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-61482-mi355x8-dsv4pro-monooff-c8",
   "model": "DeepSeek-V4-Pro",
   "notes": "InferenceX AgentX closed-loop replay (DSpark 6 drafts, synthetic acceptance 3.77), TP8 on MI355X (gfx950), FP8 preshuffled weights, mono decode off; P50 tok/s/user from one 900 s run per point; mono on arm 201 tok/s; aggregate output 292 tok/s; P90 119 \u2192 147 tok/s; median ITL 5.20 \u2192 4.96 ms; config=mono-off",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "FP8",
   "quote": "DeepSeek-V4-Pro; 8\u00d7 MI355X; FP8 weights; Interactivity P50 (tok/s/user); 192 \u2192 **201** (+5%); c8 off \u2192 on",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61482",
   "source_url": "https://github.com/vllm-project/vllm/pull/61482",
   "tg_tokens": null,
   "tps": 192,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": null,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-61482-mi355x8-dsv4pro-monoon-c8",
   "model": "DeepSeek-V4-Pro",
   "notes": "InferenceX AgentX closed-loop replay (DSpark 6 drafts, synthetic acceptance 3.77), TP8 on MI355X (gfx950), FP8 preshuffled weights, mono decode on (--monokernel, shared FlyDSL kernels); P50 tok/s/user from one 900 s run per point; mono off arm 192 tok/s; aggregate output 302 tok/s; P90 119 \u2192 147 tok/s; median ITL 5.20 \u2192 4.96 ms; config=mono-on",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "FP8",
   "quote": "DeepSeek-V4-Pro; 8\u00d7 MI355X; FP8 weights; Interactivity P50 (tok/s/user); 192 \u2192 **201** (+5%); c8 off \u2192 on",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61482",
   "source_url": "https://github.com/vllm-project/vllm/pull/61482",
   "tg_tokens": null,
   "tps": 201,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "MI355X",
   "id": "vllm-61573-mi355x-qwen3827b-tp1-nonmono-c4",
   "model": "Qwen3.8-27B",
   "notes": "FlyDSL dense mono decode opt-in VLLM_ROCM_MONO_DECODE=1, TP1 (single MI355X gfx950), non-mono arm (stock CUDA graph), aggregate output tok/s at concurrency 4, 8k input / 1k output sglang.bench_serving random dataset; mono arm 576.4 tok/s (+34.0%); median ITL 8.33 \u2192 5.94 ms; median TPOT 8.89 \u2192 6.51 ms; median TTFT 202.4 \u2192 201.0 ms; config=non-mono",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "Qwen3.8-27B; amd/Qwen3.8-27B-Quark-AWQ-MXFP4; MI355X; MXFP4; Output tok/s, non-mono \u2192 mono; 1 | 4 | 430.1 \u2192 576.4 (+34.0%)",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61573",
   "source_url": "https://github.com/vllm-project/vllm/pull/61573",
   "tg_tokens": null,
   "tps": 430.1,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "MI355X",
   "id": "vllm-61573-mi355x-qwen3827b-tp1-mono-c4",
   "model": "Qwen3.8-27B",
   "notes": "FlyDSL dense mono decode opt-in VLLM_ROCM_MONO_DECODE=1, TP1 (single MI355X gfx950), mono arm (K1+K2 inside row caps), aggregate output tok/s at concurrency 4, 8k input / 1k output sglang.bench_serving random dataset; non-mono arm 430.1 tok/s; median ITL 8.33 \u2192 5.94 ms; median TPOT 8.89 \u2192 6.51 ms; median TTFT 202.4 \u2192 201.0 ms; config=mono",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "Qwen3.8-27B; amd/Qwen3.8-27B-Quark-AWQ-MXFP4; MI355X; MXFP4; Output tok/s, non-mono \u2192 mono; 1 | 4 | 430.1 \u2192 576.4 (+34.0%)",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61573",
   "source_url": "https://github.com/vllm-project/vllm/pull/61573",
   "tg_tokens": null,
   "tps": 576.4,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "MI355X",
   "id": "vllm-61573-mi355x-qwen3827b-tp1-nonmono-c8",
   "model": "Qwen3.8-27B",
   "notes": "FlyDSL dense mono decode opt-in VLLM_ROCM_MONO_DECODE=1, TP1 (single MI355X gfx950), non-mono arm (stock CUDA graph), aggregate output tok/s at concurrency 8, 8k input / 1k output sglang.bench_serving random dataset; mono arm 911.4 tok/s (+21.2%); median ITL 8.91 \u2192 7.10 ms; median TPOT 10.09 \u2192 8.30 ms; median TTFT 206.7 \u2192 208.4 ms; config=non-mono",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "Qwen3.8-27B; amd/Qwen3.8-27B-Quark-AWQ-MXFP4; MI355X; MXFP4; Output tok/s, non-mono \u2192 mono; 1 | 8 | 751.7 \u2192 911.4 (+21.2%)",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61573",
   "source_url": "https://github.com/vllm-project/vllm/pull/61573",
   "tg_tokens": null,
   "tps": 751.7,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "MI355X",
   "id": "vllm-61573-mi355x-qwen3827b-tp1-mono-c8",
   "model": "Qwen3.8-27B",
   "notes": "FlyDSL dense mono decode opt-in VLLM_ROCM_MONO_DECODE=1, TP1 (single MI355X gfx950), mono arm (K1+K2 inside row caps), aggregate output tok/s at concurrency 8, 8k input / 1k output sglang.bench_serving random dataset; non-mono arm 751.7 tok/s; median ITL 8.91 \u2192 7.10 ms; median TPOT 10.09 \u2192 8.30 ms; median TTFT 206.7 \u2192 208.4 ms; config=mono",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "Qwen3.8-27B; amd/Qwen3.8-27B-Quark-AWQ-MXFP4; MI355X; MXFP4; Output tok/s, non-mono \u2192 mono; 1 | 8 | 751.7 \u2192 911.4 (+21.2%)",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61573",
   "source_url": "https://github.com/vllm-project/vllm/pull/61573",
   "tg_tokens": null,
   "tps": 911.4,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-61573-mi355x-qwen3827b-tp2-nonmono-c4",
   "model": "Qwen3.8-27B",
   "notes": "FlyDSL dense mono decode opt-in VLLM_ROCM_MONO_DECODE=1, TP2 (2-GPU tensor parallel MI355X gfx950), non-mono arm (stock CUDA graph), aggregate output tok/s at concurrency 4, 8k input / 1k output sglang.bench_serving random dataset; mono arm 655.8 tok/s (+13.1%); median ITL 6.05 \u2192 5.15 ms; median TPOT 6.57 \u2192 5.66 ms; median TTFT 180.2 \u2192 183.3 ms; config=non-mono",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "Qwen3.8-27B; amd/Qwen3.8-27B-Quark-AWQ-MXFP4; MI355X; MXFP4; Output tok/s, non-mono \u2192 mono; 2 | 4 | 579.6 \u2192 655.8 (+13.1%)",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61573",
   "source_url": "https://github.com/vllm-project/vllm/pull/61573",
   "tg_tokens": null,
   "tps": 579.6,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-61573-mi355x-qwen3827b-tp2-mono-c4",
   "model": "Qwen3.8-27B",
   "notes": "FlyDSL dense mono decode opt-in VLLM_ROCM_MONO_DECODE=1, TP2 (2-GPU tensor parallel MI355X gfx950), mono arm (K1+K2 inside row caps), aggregate output tok/s at concurrency 4, 8k input / 1k output sglang.bench_serving random dataset; non-mono arm 579.6 tok/s; median ITL 6.05 \u2192 5.15 ms; median TPOT 6.57 \u2192 5.66 ms; median TTFT 180.2 \u2192 183.3 ms; config=mono",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "Qwen3.8-27B; amd/Qwen3.8-27B-Quark-AWQ-MXFP4; MI355X; MXFP4; Output tok/s, non-mono \u2192 mono; 2 | 4 | 579.6 \u2192 655.8 (+13.1%)",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61573",
   "source_url": "https://github.com/vllm-project/vllm/pull/61573",
   "tg_tokens": null,
   "tps": 655.8,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-61573-mi355x-qwen3827b-tp2-nonmono-c8",
   "model": "Qwen3.8-27B",
   "notes": "FlyDSL dense mono decode opt-in VLLM_ROCM_MONO_DECODE=1, TP2 (2-GPU tensor parallel MI355X gfx950), non-mono arm (stock CUDA graph), aggregate output tok/s at concurrency 8, 8k input / 1k output sglang.bench_serving random dataset; mono arm 1080.7 tok/s (+9.5%); median ITL 6.56 \u2192 5.79 ms; median TPOT 7.66 \u2192 6.99 ms; median TTFT 186.3 \u2192 205.6 ms; config=non-mono",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "Qwen3.8-27B; amd/Qwen3.8-27B-Quark-AWQ-MXFP4; MI355X; MXFP4; Output tok/s, non-mono \u2192 mono; 2 | 8 | 987.1 \u2192 1080.7 (+9.5%)",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61573",
   "source_url": "https://github.com/vllm-project/vllm/pull/61573",
   "tg_tokens": null,
   "tps": 987.1,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-61573-mi355x-qwen3827b-tp2-mono-c8",
   "model": "Qwen3.8-27B",
   "notes": "FlyDSL dense mono decode opt-in VLLM_ROCM_MONO_DECODE=1, TP2 (2-GPU tensor parallel MI355X gfx950), mono arm (K1+K2 inside row caps), aggregate output tok/s at concurrency 8, 8k input / 1k output sglang.bench_serving random dataset; non-mono arm 987.1 tok/s; median ITL 6.56 \u2192 5.79 ms; median TPOT 7.66 \u2192 6.99 ms; median TTFT 186.3 \u2192 205.6 ms; config=mono",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "Qwen3.8-27B; amd/Qwen3.8-27B-Quark-AWQ-MXFP4; MI355X; MXFP4; Output tok/s, non-mono \u2192 mono; 2 | 8 | 987.1 \u2192 1080.7 (+9.5%)",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61573",
   "source_url": "https://github.com/vllm-project/vllm/pull/61573",
   "tg_tokens": null,
   "tps": 1080.7,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 2,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "4\u00d7 MI325X",
   "id": "vllm-61682-mi325x4-dsv41flash-baseline-8k-mc2-dspark",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "baseline arm (previous default capture sizes start at the 128-token window); ROCm decoder-replay kept-length rule (unfinished prefill chunk keeps 1 replay row, layers 20-39 SWA), TP4 on 4\u00d7 MI325X (gfx942), MXFP8 emulation kernel, QuickReduce INT4 all-reduce, vLLM 0.32.0rc1 nightly rocm723; aggregate tok/s, vllm bench serve random 20 prompts output 1024, one run per cell; median TTFT 336.2 \u2192 337.1 ms; median TPOT 3.96 \u2192 4.03 ms; cell=8k-mc2-dspark; config=baseline",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP8",
   "quote": "DeepSeek-V4.1-Flash; 4\u00d7 MI325X; 8k\u21921k MC2, DSpark, seed 7; Output (tok/s) 431.1 430.2",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61682",
   "source_url": "https://github.com/vllm-project/vllm/pull/61682",
   "tg_tokens": null,
   "tps": 431.1,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 2,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "4\u00d7 MI325X",
   "id": "vllm-61682-mi325x4-dsv41flash-onrow-8k-mc2-dspark",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "one-row arm (explicit capture-size list including 1); ROCm decoder-replay kept-length rule (unfinished prefill chunk keeps 1 replay row, layers 20-39 SWA), TP4 on 4\u00d7 MI325X (gfx942), MXFP8 emulation kernel, QuickReduce INT4 all-reduce, vLLM 0.32.0rc1 nightly rocm723; aggregate tok/s, vllm bench serve random 20 prompts output 1024, one run per cell; median TTFT 336.2 \u2192 337.1 ms; median TPOT 3.96 \u2192 4.03 ms; cell=8k-mc2-dspark; config=one-row",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP8",
   "quote": "DeepSeek-V4.1-Flash; 4\u00d7 MI325X; 8k\u21921k MC2, DSpark, seed 7; Output (tok/s) 431.1 430.2",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61682",
   "source_url": "https://github.com/vllm-project/vllm/pull/61682",
   "tg_tokens": null,
   "tps": 430.2,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 2,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "4\u00d7 MI325X",
   "id": "vllm-61682-mi325x4-dsv41flash-baseline-8k-mc2-nodspark",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "baseline arm (previous default capture sizes start at the 128-token window); ROCm decoder-replay kept-length rule (unfinished prefill chunk keeps 1 replay row, layers 20-39 SWA), TP4 on 4\u00d7 MI325X (gfx942), MXFP8 emulation kernel, QuickReduce INT4 all-reduce, vLLM 0.32.0rc1 nightly rocm723; aggregate tok/s, vllm bench serve random 20 prompts output 1024, one run per cell; median TTFT 463.2 \u2192 463.7 ms; median TPOT 11.67 \u2192 11.63 ms; cell=8k-mc2-nodspark; config=baseline",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP8",
   "quote": "DeepSeek-V4.1-Flash; 4\u00d7 MI325X; 8k\u21921k MC2, no DSpark, seed 7; Output (tok/s) 165.8 165.7",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61682",
   "source_url": "https://github.com/vllm-project/vllm/pull/61682",
   "tg_tokens": null,
   "tps": 165.8,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 2,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "4\u00d7 MI325X",
   "id": "vllm-61682-mi325x4-dsv41flash-onrow-8k-mc2-nodspark",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "one-row arm (explicit capture-size list including 1); ROCm decoder-replay kept-length rule (unfinished prefill chunk keeps 1 replay row, layers 20-39 SWA), TP4 on 4\u00d7 MI325X (gfx942), MXFP8 emulation kernel, QuickReduce INT4 all-reduce, vLLM 0.32.0rc1 nightly rocm723; aggregate tok/s, vllm bench serve random 20 prompts output 1024, one run per cell; median TTFT 463.2 \u2192 463.7 ms; median TPOT 11.67 \u2192 11.63 ms; cell=8k-mc2-nodspark; config=one-row",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP8",
   "quote": "DeepSeek-V4.1-Flash; 4\u00d7 MI325X; 8k\u21921k MC2, no DSpark, seed 7; Output (tok/s) 165.8 165.7",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61682",
   "source_url": "https://github.com/vllm-project/vllm/pull/61682",
   "tg_tokens": null,
   "tps": 165.7,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 6,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "4\u00d7 MI325X",
   "id": "vllm-61682-mi325x4-dsv41flash-baseline-8k-mc6-dspark",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "baseline arm (previous default capture sizes start at the 128-token window); ROCm decoder-replay kept-length rule (unfinished prefill chunk keeps 1 replay row, layers 20-39 SWA), TP4 on 4\u00d7 MI325X (gfx942), MXFP8 emulation kernel, QuickReduce INT4 all-reduce, vLLM 0.32.0rc1 nightly rocm723; aggregate tok/s, vllm bench serve random 20 prompts output 1024, one run per cell; median TTFT 1075.7 \u2192 1068.3 ms; median TPOT 6.22 \u2192 6.29 ms; cell=8k-mc6-dspark; config=baseline",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP8",
   "quote": "DeepSeek-V4.1-Flash; 4\u00d7 MI325X; 8k\u21921k MC6, DSpark, seed 9; Output (tok/s) 699.8 714.8",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61682",
   "source_url": "https://github.com/vllm-project/vllm/pull/61682",
   "tg_tokens": null,
   "tps": 699.8,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 6,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "4\u00d7 MI325X",
   "id": "vllm-61682-mi325x4-dsv41flash-onrow-8k-mc6-dspark",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "one-row arm (explicit capture-size list including 1); ROCm decoder-replay kept-length rule (unfinished prefill chunk keeps 1 replay row, layers 20-39 SWA), TP4 on 4\u00d7 MI325X (gfx942), MXFP8 emulation kernel, QuickReduce INT4 all-reduce, vLLM 0.32.0rc1 nightly rocm723; aggregate tok/s, vllm bench serve random 20 prompts output 1024, one run per cell; median TTFT 1075.7 \u2192 1068.3 ms; median TPOT 6.22 \u2192 6.29 ms; cell=8k-mc6-dspark; config=one-row",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP8",
   "quote": "DeepSeek-V4.1-Flash; 4\u00d7 MI325X; 8k\u21921k MC6, DSpark, seed 9; Output (tok/s) 699.8 714.8",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61682",
   "source_url": "https://github.com/vllm-project/vllm/pull/61682",
   "tg_tokens": null,
   "tps": 714.8,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 2,
   "ctx": 262144,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "4\u00d7 MI325X",
   "id": "vllm-61682-mi325x4-dsv41flash-baseline-256k-mc2-dspark",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "baseline arm (previous default capture sizes start at the 128-token window); ROCm decoder-replay kept-length rule (unfinished prefill chunk keeps 1 replay row, layers 20-39 SWA), TP4 on 4\u00d7 MI325X (gfx942), MXFP8 emulation kernel, QuickReduce INT4 all-reduce, vLLM 0.32.0rc1 nightly rocm723; aggregate tok/s, vllm bench serve random 20 prompts output 1024, one run per cell; median TTFT 13109.6 \u2192 12181.2 ms; median TPOT 7.94 \u2192 8.27 ms; cell=256k-mc2-dspark; config=baseline",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP8",
   "quote": "DeepSeek-V4.1-Flash; 4\u00d7 MI325X; 256k\u21921k MC2, DSpark, seed 7; Output (tok/s) 95.5 98.2",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61682",
   "source_url": "https://github.com/vllm-project/vllm/pull/61682",
   "tg_tokens": null,
   "tps": 95.5,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 2,
   "ctx": 262144,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "4\u00d7 MI325X",
   "id": "vllm-61682-mi325x4-dsv41flash-onrow-256k-mc2-dspark",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "one-row arm (explicit capture-size list including 1); ROCm decoder-replay kept-length rule (unfinished prefill chunk keeps 1 replay row, layers 20-39 SWA), TP4 on 4\u00d7 MI325X (gfx942), MXFP8 emulation kernel, QuickReduce INT4 all-reduce, vLLM 0.32.0rc1 nightly rocm723; aggregate tok/s, vllm bench serve random 20 prompts output 1024, one run per cell; median TTFT 13109.6 \u2192 12181.2 ms; median TPOT 7.94 \u2192 8.27 ms; cell=256k-mc2-dspark; config=one-row",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP8",
   "quote": "DeepSeek-V4.1-Flash; 4\u00d7 MI325X; 256k\u21921k MC2, DSpark, seed 7; Output (tok/s) 95.5 98.2",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61682",
   "source_url": "https://github.com/vllm-project/vllm/pull/61682",
   "tg_tokens": null,
   "tps": 98.2,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 262144,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "4\u00d7 MI325X",
   "id": "vllm-61682-mi325x4-dsv41flash-baseline-256k-mc8-dspark",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "baseline arm (previous default capture sizes start at the 128-token window); ROCm decoder-replay kept-length rule (unfinished prefill chunk keeps 1 replay row, layers 20-39 SWA), TP4 on 4\u00d7 MI325X (gfx942), MXFP8 emulation kernel, QuickReduce INT4 all-reduce, vLLM 0.32.0rc1 nightly rocm723; aggregate tok/s, vllm bench serve random 20 prompts output 1024, one run per cell; median TTFT 23984.5 \u2192 22806.5 ms; median TPOT 47.31 \u2192 45.71 ms; cell=256k-mc8-dspark; config=baseline",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP8",
   "quote": "DeepSeek-V4.1-Flash; 4\u00d7 MI325X; 256k\u21921k MC8, DSpark, seed 10; Output (tok/s) 109.3 113.5",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61682",
   "source_url": "https://github.com/vllm-project/vllm/pull/61682",
   "tg_tokens": null,
   "tps": 109.3,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 262144,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "4\u00d7 MI325X",
   "id": "vllm-61682-mi325x4-dsv41flash-onrow-256k-mc8-dspark",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "one-row arm (explicit capture-size list including 1); ROCm decoder-replay kept-length rule (unfinished prefill chunk keeps 1 replay row, layers 20-39 SWA), TP4 on 4\u00d7 MI325X (gfx942), MXFP8 emulation kernel, QuickReduce INT4 all-reduce, vLLM 0.32.0rc1 nightly rocm723; aggregate tok/s, vllm bench serve random 20 prompts output 1024, one run per cell; median TTFT 23984.5 \u2192 22806.5 ms; median TPOT 47.31 \u2192 45.71 ms; cell=256k-mc8-dspark; config=one-row",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP8",
   "quote": "DeepSeek-V4.1-Flash; 4\u00d7 MI325X; 256k\u21921k MC8, DSpark, seed 10; Output (tok/s) 109.3 113.5",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61682",
   "source_url": "https://github.com/vllm-project/vllm/pull/61682",
   "tg_tokens": null,
   "tps": 113.5,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 32,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-60470-gb300-kimik3-native-c32",
   "model": "Kimi-K3",
   "notes": "vLLM PR #60470 FlashInfer backend=auto KDA decode vs native arm; Kimi-K3 NVFP4, TP8 over two GB300 nodes (SM103), vllm 0.31.1rc1 nightly cu134, FlashInfer from main; vllm bench serve random 1024-token prompts 128 output unless stated, ignore-eos, one run per arm, within +/-1.5% no regression signal; native arm; scenario=200 prompts, concurrency 32; median 797 ms TTFT, 27.30 ms TPOT; config=native",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "NVFP4",
   "quote": "Kimi-K3 NVFP4; GB300; 200 prompts, concurrency 32; native 7.10 909 797 / 1337 27.30",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #60470",
   "source_url": "https://github.com/vllm-project/vllm/issues/60470",
   "tg_tokens": null,
   "tps": 909,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 32,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-60470-gb300-kimik3-fiauto-c32",
   "model": "Kimi-K3",
   "notes": "vLLM PR #60470 FlashInfer backend=auto KDA decode vs native arm; Kimi-K3 NVFP4, TP8 over two GB300 nodes (SM103), vllm 0.31.1rc1 nightly cu134, FlashInfer from main; vllm bench serve random 1024-token prompts 128 output unless stated, ignore-eos, one run per arm, within +/-1.5% no regression signal; fiauto arm (backend=auto); scenario=200 prompts, concurrency 32; median 796 ms TTFT, 27.45 ms TPOT; config=fiauto",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "NVFP4",
   "quote": "Kimi-K3 NVFP4; GB300; 200 prompts, concurrency 32; fiauto 7.05 902 796 / 1330 27.45",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #60470",
   "source_url": "https://github.com/vllm-project/vllm/issues/60470",
   "tg_tokens": null,
   "tps": 902,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 32,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-60470-gb300-kimik3-native-c32w",
   "model": "Kimi-K3",
   "notes": "vLLM PR #60470 FlashInfer backend=auto KDA decode vs native arm; Kimi-K3 NVFP4, TP8 over two GB300 nodes (SM103), vllm 0.31.1rc1 nightly cu134, FlashInfer from main; vllm bench serve random 1024-token prompts 128 output unless stated, ignore-eos, one run per arm, within +/-1.5% no regression signal; native arm; scenario=32 prompts, concurrency 32 (warm-up pass, run first); median 1069 ms TTFT, 25.51 ms TPOT; config=native",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "NVFP4",
   "quote": "Kimi-K3 NVFP4; GB300; 32 prompts, concurrency 32 (warm-up pass, run first); native 7.39 946 1069 / 1496 25.51",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #60470",
   "source_url": "https://github.com/vllm-project/vllm/issues/60470",
   "tg_tokens": null,
   "tps": 946,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 32,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-60470-gb300-kimik3-fiauto-c32w",
   "model": "Kimi-K3",
   "notes": "vLLM PR #60470 FlashInfer backend=auto KDA decode vs native arm; Kimi-K3 NVFP4, TP8 over two GB300 nodes (SM103), vllm 0.31.1rc1 nightly cu134, FlashInfer from main; vllm bench serve random 1024-token prompts 128 output unless stated, ignore-eos, one run per arm, within +/-1.5% no regression signal; fiauto arm (backend=auto); scenario=32 prompts, concurrency 32 (warm-up pass, run first); median 1066 ms TTFT, 25.66 ms TPOT; config=fiauto",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "NVFP4",
   "quote": "Kimi-K3 NVFP4; GB300; 32 prompts, concurrency 32 (warm-up pass, run first); fiauto 7.36 942 1066 / 1494 25.66",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #60470",
   "source_url": "https://github.com/vllm-project/vllm/issues/60470",
   "tg_tokens": null,
   "tps": 942,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-60470-gb300-kimik3-native-c8",
   "model": "Kimi-K3",
   "notes": "vLLM PR #60470 FlashInfer backend=auto KDA decode vs native arm; Kimi-K3 NVFP4, TP8 over two GB300 nodes (SM103), vllm 0.31.1rc1 nightly cu134, FlashInfer from main; vllm bench serve random 1024-token prompts 128 output unless stated, ignore-eos, one run per arm, within +/-1.5% no regression signal; native arm; scenario=32 prompts, concurrency 8; median 506 ms TTFT, 15.47 ms TPOT; config=native",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "NVFP4",
   "quote": "Kimi-K3 NVFP4; GB300; 32 prompts, concurrency 8; native 3.22 413 506 / 553 15.47",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #60470",
   "source_url": "https://github.com/vllm-project/vllm/issues/60470",
   "tg_tokens": null,
   "tps": 413,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-60470-gb300-kimik3-fiauto-c8",
   "model": "Kimi-K3",
   "notes": "vLLM PR #60470 FlashInfer backend=auto KDA decode vs native arm; Kimi-K3 NVFP4, TP8 over two GB300 nodes (SM103), vllm 0.31.1rc1 nightly cu134, FlashInfer from main; vllm bench serve random 1024-token prompts 128 output unless stated, ignore-eos, one run per arm, within +/-1.5% no regression signal; fiauto arm (backend=auto); scenario=32 prompts, concurrency 8; median 505 ms TTFT, 15.40 ms TPOT; config=fiauto",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "NVFP4",
   "quote": "Kimi-K3 NVFP4; GB300; 32 prompts, concurrency 8; fiauto 3.24 415 505 / 535 15.40",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #60470",
   "source_url": "https://github.com/vllm-project/vllm/issues/60470",
   "tg_tokens": null,
   "tps": 415,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-60470-gb300-kimik3-native-c1",
   "model": "Kimi-K3",
   "notes": "vLLM PR #60470 FlashInfer backend=auto KDA decode vs native arm; Kimi-K3 NVFP4, TP8 over two GB300 nodes (SM103), vllm 0.31.1rc1 nightly cu134, FlashInfer from main; vllm bench serve random 1024-token prompts 128 output unless stated, ignore-eos, one run per arm, within +/-1.5% no regression signal; native arm; scenario=8 prompts, concurrency 1; median 225 ms TTFT, 11.06 ms TPOT; config=native",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "NVFP4",
   "quote": "Kimi-K3 NVFP4; GB300; 8 prompts, concurrency 1; native 0.611 78.2 225 / 256 11.06",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #60470",
   "source_url": "https://github.com/vllm-project/vllm/issues/60470",
   "tg_tokens": null,
   "tps": 78.2,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-60470-gb300-kimik3-fiauto-c1",
   "model": "Kimi-K3",
   "notes": "vLLM PR #60470 FlashInfer backend=auto KDA decode vs native arm; Kimi-K3 NVFP4, TP8 over two GB300 nodes (SM103), vllm 0.31.1rc1 nightly cu134, FlashInfer from main; vllm bench serve random 1024-token prompts 128 output unless stated, ignore-eos, one run per arm, within +/-1.5% no regression signal; fiauto arm (backend=auto); scenario=8 prompts, concurrency 1; median 223 ms TTFT, 11.05 ms TPOT; config=fiauto",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "NVFP4",
   "quote": "Kimi-K3 NVFP4; GB300; 8 prompts, concurrency 1; fiauto 0.615 78.7 223 / 231 11.05",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #60470",
   "source_url": "https://github.com/vllm-project/vllm/issues/60470",
   "tg_tokens": null,
   "tps": 78.7,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-60470-gb300-kimik3-native-c1-8k",
   "model": "Kimi-K3",
   "notes": "vLLM PR #60470 FlashInfer backend=auto KDA decode vs native arm; Kimi-K3 NVFP4, TP8 over two GB300 nodes (SM103), vllm 0.31.1rc1 nightly cu134, FlashInfer from main; vllm bench serve random 1024-token prompts 128 output unless stated, ignore-eos, one run per arm, within +/-1.5% no regression signal; native arm; scenario=4 prompts of 8192 tokens, 16 output, concurrency 1; median 351 ms TTFT, 11.40 ms TPOT; config=native",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "NVFP4",
   "quote": "Kimi-K3 NVFP4; GB300; 4 prompts of 8192 tokens, 16 output, concurrency 1; native 1.85 29.6 351 / 419 11.40",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #60470",
   "source_url": "https://github.com/vllm-project/vllm/issues/60470",
   "tg_tokens": null,
   "tps": 29.6,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-60470-gb300-kimik3-fiauto-c1-8k",
   "model": "Kimi-K3",
   "notes": "vLLM PR #60470 FlashInfer backend=auto KDA decode vs native arm; Kimi-K3 NVFP4, TP8 over two GB300 nodes (SM103), vllm 0.31.1rc1 nightly cu134, FlashInfer from main; vllm bench serve random 1024-token prompts 128 output unless stated, ignore-eos, one run per arm, within +/-1.5% no regression signal; fiauto arm (backend=auto); scenario=4 prompts of 8192 tokens, 16 output, concurrency 1; median 11.57 ms TPOT, 352 ms TTFT; config=fiauto",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "NVFP4",
   "quote": "Kimi-K3 NVFP4; GB300; 4 prompts of 8192 tokens, 16 output, concurrency 1; fiauto 1.84 29.4 352 / 428 11.57",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #60470",
   "source_url": "https://github.com/vllm-project/vllm/issues/60470",
   "tg_tokens": null,
   "tps": 29.4,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-61325-mi355x2-qwen38flashnext-stock-8k-c1",
   "model": "Qwen3.8-Flash-Next",
   "notes": "vLLM PR #61325 mono decode layer opt-in VLLM_ROCM_MONO_DECODE=1, TP2 on 2x MI355X (gfx950), checkpoint amd/Qwen3.8-Flash-Next-Quark-MXFP4, nightly-2eb491e1 + branch, arms differ only in the flag, one run per cell; aggregate output tok/s (sglang.bench_serving style ISL/OSL); stock arm (flag off), #59575 uplifts on main; ISL/OSL 8k/1k; mean TTFT 494 \u2192 491 ms; mean TPOT 6.34 \u2192 3.46 ms (-45%); config=stock",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "Qwen3.8-Flash-Next; MI355X; MXFP4; 8k / 1k | 1 (8) | 146.8 \u2192 253.8",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61325",
   "source_url": "https://github.com/vllm-project/vllm/pull/61325",
   "tg_tokens": null,
   "tps": 146.8,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-61325-mi355x2-qwen38flashnext-mono-8k-c1",
   "model": "Qwen3.8-Flash-Next",
   "notes": "vLLM PR #61325 mono decode layer opt-in VLLM_ROCM_MONO_DECODE=1, TP2 on 2x MI355X (gfx950), checkpoint amd/Qwen3.8-Flash-Next-Quark-MXFP4, nightly-2eb491e1 + branch, arms differ only in the flag, one run per cell; aggregate output tok/s (sglang.bench_serving style ISL/OSL); mono arm (flag on), #59575 uplifts on main; ISL/OSL 8k/1k; mean TTFT 494 \u2192 491 ms; mean TPOT 6.34 \u2192 3.46 ms (-45%); config=mono",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "Qwen3.8-Flash-Next; MI355X; MXFP4; 8k / 1k | 1 (8) | 146.8 \u2192 253.8",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61325",
   "source_url": "https://github.com/vllm-project/vllm/pull/61325",
   "tg_tokens": null,
   "tps": 253.8,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 2,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-61325-mi355x2-qwen38flashnext-stock-8k-c2",
   "model": "Qwen3.8-Flash-Next",
   "notes": "vLLM PR #61325 mono decode layer opt-in VLLM_ROCM_MONO_DECODE=1, TP2 on 2x MI355X (gfx950), checkpoint amd/Qwen3.8-Flash-Next-Quark-MXFP4, nightly-2eb491e1 + branch, arms differ only in the flag, one run per cell; aggregate output tok/s (sglang.bench_serving style ISL/OSL); stock arm (flag off), #59575 uplifts on main; ISL/OSL 8k/1k; mean TTFT 644 \u2192 643 ms; mean TPOT 6.86 \u2192 4.17 ms (-39%); config=stock",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "Qwen3.8-Flash-Next; MI355X; MXFP4; 8k / 1k | 2 (12) | 267.4 \u2192 417.0",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61325",
   "source_url": "https://github.com/vllm-project/vllm/pull/61325",
   "tg_tokens": null,
   "tps": 267.4,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 2,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-61325-mi355x2-qwen38flashnext-mono-8k-c2",
   "model": "Qwen3.8-Flash-Next",
   "notes": "vLLM PR #61325 mono decode layer opt-in VLLM_ROCM_MONO_DECODE=1, TP2 on 2x MI355X (gfx950), checkpoint amd/Qwen3.8-Flash-Next-Quark-MXFP4, nightly-2eb491e1 + branch, arms differ only in the flag, one run per cell; aggregate output tok/s (sglang.bench_serving style ISL/OSL); mono arm (flag on), #59575 uplifts on main; ISL/OSL 8k/1k; mean TTFT 644 \u2192 643 ms; mean TPOT 6.86 \u2192 4.17 ms (-39%); config=mono",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "Qwen3.8-Flash-Next; MI355X; MXFP4; 8k / 1k | 2 (12) | 267.4 \u2192 417.0",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61325",
   "source_url": "https://github.com/vllm-project/vllm/pull/61325",
   "tg_tokens": null,
   "tps": 417.0,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-61325-mi355x2-qwen38flashnext-stock-8k-c4",
   "model": "Qwen3.8-Flash-Next",
   "notes": "vLLM PR #61325 mono decode layer opt-in VLLM_ROCM_MONO_DECODE=1, TP2 on 2x MI355X (gfx950), checkpoint amd/Qwen3.8-Flash-Next-Quark-MXFP4, nightly-2eb491e1 + branch, arms differ only in the flag, one run per cell; aggregate output tok/s (sglang.bench_serving style ISL/OSL); stock arm (flag off), #59575 uplifts on main; ISL/OSL 8k/1k; mean TTFT 1,209 \u2192 1,233 ms; mean TPOT 7.84 \u2192 5.13 ms (-35%); config=stock",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "Qwen3.8-Flash-Next; MI355X; MXFP4; 8k / 1k | 4 (16) | 443.5 \u2192 632.2",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61325",
   "source_url": "https://github.com/vllm-project/vllm/pull/61325",
   "tg_tokens": null,
   "tps": 443.5,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-61325-mi355x2-qwen38flashnext-mono-8k-c4",
   "model": "Qwen3.8-Flash-Next",
   "notes": "vLLM PR #61325 mono decode layer opt-in VLLM_ROCM_MONO_DECODE=1, TP2 on 2x MI355X (gfx950), checkpoint amd/Qwen3.8-Flash-Next-Quark-MXFP4, nightly-2eb491e1 + branch, arms differ only in the flag, one run per cell; aggregate output tok/s (sglang.bench_serving style ISL/OSL); mono arm (flag on), #59575 uplifts on main; ISL/OSL 8k/1k; mean TTFT 1,209 \u2192 1,233 ms; mean TPOT 7.84 \u2192 5.13 ms (-35%); config=mono",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "Qwen3.8-Flash-Next; MI355X; MXFP4; 8k / 1k | 4 (16) | 443.5 \u2192 632.2",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61325",
   "source_url": "https://github.com/vllm-project/vllm/pull/61325",
   "tg_tokens": null,
   "tps": 632.2,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-61325-mi355x2-qwen38flashnext-stock-8k-c8",
   "model": "Qwen3.8-Flash-Next",
   "notes": "vLLM PR #61325 mono decode layer opt-in VLLM_ROCM_MONO_DECODE=1, TP2 on 2x MI355X (gfx950), checkpoint amd/Qwen3.8-Flash-Next-Quark-MXFP4, nightly-2eb491e1 + branch, arms differ only in the flag, one run per cell; aggregate output tok/s (sglang.bench_serving style ISL/OSL); stock arm (flag off), #59575 uplifts on main; ISL/OSL 8k/1k; mean TTFT 2,114 \u2192 2,032 ms; mean TPOT 10.39 \u2192 7.38 ms (-29%); config=stock",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "Qwen3.8-Flash-Next; MI355X; MXFP4; 8k / 1k | 8 (32) | 642.5 \u2192 854.7",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61325",
   "source_url": "https://github.com/vllm-project/vllm/pull/61325",
   "tg_tokens": null,
   "tps": 642.5,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-61325-mi355x2-qwen38flashnext-mono-8k-c8",
   "model": "Qwen3.8-Flash-Next",
   "notes": "vLLM PR #61325 mono decode layer opt-in VLLM_ROCM_MONO_DECODE=1, TP2 on 2x MI355X (gfx950), checkpoint amd/Qwen3.8-Flash-Next-Quark-MXFP4, nightly-2eb491e1 + branch, arms differ only in the flag, one run per cell; aggregate output tok/s (sglang.bench_serving style ISL/OSL); mono arm (flag on), #59575 uplifts on main; ISL/OSL 8k/1k; mean TTFT 2,114 \u2192 2,032 ms; mean TPOT 10.39 \u2192 7.38 ms (-29%); config=mono",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "Qwen3.8-Flash-Next; MI355X; MXFP4; 8k / 1k | 8 (32) | 642.5 \u2192 854.7",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61325",
   "source_url": "https://github.com/vllm-project/vllm/pull/61325",
   "tg_tokens": null,
   "tps": 854.7,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 61440,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-61325-mi355x2-qwen38flashnext-stock-60k-c1",
   "model": "Qwen3.8-Flash-Next",
   "notes": "vLLM PR #61325 mono decode layer opt-in VLLM_ROCM_MONO_DECODE=1, TP2 on 2x MI355X (gfx950), checkpoint amd/Qwen3.8-Flash-Next-Quark-MXFP4, nightly-2eb491e1 + branch, arms differ only in the flag, one run per cell; aggregate output tok/s (sglang.bench_serving style ISL/OSL); stock arm (flag off), #59575 uplifts on main; ISL/OSL 60k/600; mean TTFT 3,175 \u2192 3,078 ms; mean TPOT 6.55 \u2192 4.90 ms (-25%); config=stock",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "Qwen3.8-Flash-Next; MI355X; MXFP4; 60k / 600 | 1 (6) | 84.5 \u2192 99.8",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61325",
   "source_url": "https://github.com/vllm-project/vllm/pull/61325",
   "tg_tokens": null,
   "tps": 84.5,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 61440,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-61325-mi355x2-qwen38flashnext-mono-60k-c1",
   "model": "Qwen3.8-Flash-Next",
   "notes": "vLLM PR #61325 mono decode layer opt-in VLLM_ROCM_MONO_DECODE=1, TP2 on 2x MI355X (gfx950), checkpoint amd/Qwen3.8-Flash-Next-Quark-MXFP4, nightly-2eb491e1 + branch, arms differ only in the flag, one run per cell; aggregate output tok/s (sglang.bench_serving style ISL/OSL); mono arm (flag on), #59575 uplifts on main; ISL/OSL 60k/600; mean TTFT 3,175 \u2192 3,078 ms; mean TPOT 6.55 \u2192 4.90 ms (-25%); config=mono",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "Qwen3.8-Flash-Next; MI355X; MXFP4; 60k / 600 | 1 (6) | 84.5 \u2192 99.8",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61325",
   "source_url": "https://github.com/vllm-project/vllm/pull/61325",
   "tg_tokens": null,
   "tps": 99.8,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 61440,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-61325-mi355x2-qwen38flashnext-stock-60k-c4",
   "model": "Qwen3.8-Flash-Next",
   "notes": "vLLM PR #61325 mono decode layer opt-in VLLM_ROCM_MONO_DECODE=1, TP2 on 2x MI355X (gfx950), checkpoint amd/Qwen3.8-Flash-Next-Quark-MXFP4, nightly-2eb491e1 + branch, arms differ only in the flag, one run per cell; aggregate output tok/s (sglang.bench_serving style ISL/OSL); stock arm (flag off), #59575 uplifts on main; ISL/OSL 60k/600; mean TTFT 7,616 \u2192 7,917 ms; mean TPOT 15.22 \u2192 12.67 ms (-17%); config=stock",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "Qwen3.8-Flash-Next; MI355X; MXFP4; 60k / 600 | 4 (16) | 143.3 \u2192 154.7",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61325",
   "source_url": "https://github.com/vllm-project/vllm/pull/61325",
   "tg_tokens": null,
   "tps": 143.3,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 61440,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-61325-mi355x2-qwen38flashnext-mono-60k-c4",
   "model": "Qwen3.8-Flash-Next",
   "notes": "vLLM PR #61325 mono decode layer opt-in VLLM_ROCM_MONO_DECODE=1, TP2 on 2x MI355X (gfx950), checkpoint amd/Qwen3.8-Flash-Next-Quark-MXFP4, nightly-2eb491e1 + branch, arms differ only in the flag, one run per cell; aggregate output tok/s (sglang.bench_serving style ISL/OSL); mono arm (flag on), #59575 uplifts on main; ISL/OSL 60k/600; mean TTFT 7,616 \u2192 7,917 ms; mean TPOT 15.22 \u2192 12.67 ms (-17%); config=mono",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "Qwen3.8-Flash-Next; MI355X; MXFP4; 60k / 600 | 4 (16) | 143.3 \u2192 154.7",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61325",
   "source_url": "https://github.com/vllm-project/vllm/pull/61325",
   "tg_tokens": null,
   "tps": 154.7,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-61325-mi355x2-qwen38flashnext-stock-mtp-8k-c1",
   "model": "Qwen3.8-Flash-Next",
   "notes": "vLLM PR #61325 mono decode layer opt-in VLLM_ROCM_MONO_DECODE=1, TP2 on 2x MI355X (gfx950), checkpoint amd/Qwen3.8-Flash-Next-Quark-MXFP4, nightly-2eb491e1 + branch, arms differ only in the flag, one run per cell; aggregate output tok/s (sglang.bench_serving style ISL/OSL); stock arm with MTP-3 (num_speculative_tokens 3, max-num-batched-tokens 8192, re:mtp\\..* exclude fix applied), verify steps up to 8 rows run on the kernels; ; ISL/OSL 8k/1k; mean TTFT 494 \u2192 489 ms; mean TPOT 3.04 \u2192 2.16 ms (-29%); config=stock-mtp",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "Qwen3.8-Flash-Next; MI355X; MXFP4; MTP-3; 8k / 1k | 1 (8) | 284.4 \u2192 379.1",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61325",
   "source_url": "https://github.com/vllm-project/vllm/pull/61325",
   "tg_tokens": null,
   "tps": 284.4,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-61325-mi355x2-qwen38flashnext-mono-mtp-8k-c1",
   "model": "Qwen3.8-Flash-Next",
   "notes": "vLLM PR #61325 mono decode layer opt-in VLLM_ROCM_MONO_DECODE=1, TP2 on 2x MI355X (gfx950), checkpoint amd/Qwen3.8-Flash-Next-Quark-MXFP4, nightly-2eb491e1 + branch, arms differ only in the flag, one run per cell; aggregate output tok/s (sglang.bench_serving style ISL/OSL); mono arm (flag on) with MTP-3 (num_speculative_tokens 3, max-num-batched-tokens 8192, re:mtp\\..* exclude fix applied), verify steps up to 8 rows run on the kernels; ; ISL/OSL 8k/1k; mean TTFT 494 \u2192 489 ms; mean TPOT 3.04 \u2192 2.16 ms (-29%); config=mono-mtp",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "Qwen3.8-Flash-Next; MI355X; MXFP4; MTP-3; 8k / 1k | 1 (8) | 284.4 \u2192 379.1",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61325",
   "source_url": "https://github.com/vllm-project/vllm/pull/61325",
   "tg_tokens": null,
   "tps": 379.1,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 2,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-61325-mi355x2-qwen38flashnext-stock-mtp-8k-c2",
   "model": "Qwen3.8-Flash-Next",
   "notes": "vLLM PR #61325 mono decode layer opt-in VLLM_ROCM_MONO_DECODE=1, TP2 on 2x MI355X (gfx950), checkpoint amd/Qwen3.8-Flash-Next-Quark-MXFP4, nightly-2eb491e1 + branch, arms differ only in the flag, one run per cell; aggregate output tok/s (sglang.bench_serving style ISL/OSL); stock arm with MTP-3 (num_speculative_tokens 3, max-num-batched-tokens 8192, re:mtp\\..* exclude fix applied), verify steps up to 8 rows run on the kernels; ; ISL/OSL 8k/1k; mean TTFT 576 \u2192 563 ms; mean TPOT 4.39 \u2192 3.47 ms (-21%); config=stock-mtp",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "Qwen3.8-Flash-Next; MI355X; MXFP4; MTP-3; 8k / 1k | 2 (12) | 397.2 \u2192 491.2",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61325",
   "source_url": "https://github.com/vllm-project/vllm/pull/61325",
   "tg_tokens": null,
   "tps": 397.2,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 2,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-61325-mi355x2-qwen38flashnext-mono-mtp-8k-c2",
   "model": "Qwen3.8-Flash-Next",
   "notes": "vLLM PR #61325 mono decode layer opt-in VLLM_ROCM_MONO_DECODE=1, TP2 on 2x MI355X (gfx950), checkpoint amd/Qwen3.8-Flash-Next-Quark-MXFP4, nightly-2eb491e1 + branch, arms differ only in the flag, one run per cell; aggregate output tok/s (sglang.bench_serving style ISL/OSL); mono arm (flag on) with MTP-3 (num_speculative_tokens 3, max-num-batched-tokens 8192, re:mtp\\..* exclude fix applied), verify steps up to 8 rows run on the kernels; ; ISL/OSL 8k/1k; mean TTFT 576 \u2192 563 ms; mean TPOT 4.39 \u2192 3.47 ms (-21%); config=mono-mtp",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "Qwen3.8-Flash-Next; MI355X; MXFP4; MTP-3; 8k / 1k | 2 (12) | 397.2 \u2192 491.2",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61325",
   "source_url": "https://github.com/vllm-project/vllm/pull/61325",
   "tg_tokens": null,
   "tps": 491.2,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-61325-mi355x2-qwen38flashnext-stock-mtp-8k-c4",
   "model": "Qwen3.8-Flash-Next",
   "notes": "vLLM PR #61325 mono decode layer opt-in VLLM_ROCM_MONO_DECODE=1, TP2 on 2x MI355X (gfx950), checkpoint amd/Qwen3.8-Flash-Next-Quark-MXFP4, nightly-2eb491e1 + branch, arms differ only in the flag, one run per cell; aggregate output tok/s (sglang.bench_serving style ISL/OSL); c4 verify step has 16 rows past the 8-row cap, both arms run the stock path (noise floor); stock arm with MTP-3 (num_speculative_tokens 3, max-num-batched-tokens 8192, re:mtp\\..* exclude fix applied), verify steps up to 8 rows run on the kernels; ; ISL/OSL 8k/1k; mean TTFT 718 \u2192 742 ms; mean TPOT 5.07 \u2192 4.88 ms; config=stock-mtp",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "Qwen3.8-Flash-Next; MI355X; MXFP4; MTP-3; 8k / 1k | 4 (16), stock in both | 654.0 \u2192 684.5",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61325",
   "source_url": "https://github.com/vllm-project/vllm/pull/61325",
   "tg_tokens": null,
   "tps": 654.0,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-61325-mi355x2-qwen38flashnext-mono-mtp-8k-c4",
   "model": "Qwen3.8-Flash-Next",
   "notes": "vLLM PR #61325 mono decode layer opt-in VLLM_ROCM_MONO_DECODE=1, TP2 on 2x MI355X (gfx950), checkpoint amd/Qwen3.8-Flash-Next-Quark-MXFP4, nightly-2eb491e1 + branch, arms differ only in the flag, one run per cell; aggregate output tok/s (sglang.bench_serving style ISL/OSL); c4 verify step has 16 rows past the 8-row cap, both arms run the stock path (noise floor); mono arm (flag on) with MTP-3 (num_speculative_tokens 3, max-num-batched-tokens 8192, re:mtp\\..* exclude fix applied), verify steps up to 8 rows run on the kernels; ; ISL/OSL 8k/1k; mean TTFT 718 \u2192 742 ms; mean TPOT 5.07 \u2192 4.88 ms; config=mono-mtp",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "Qwen3.8-Flash-Next; MI355X; MXFP4; MTP-3; 8k / 1k | 4 (16), stock in both | 654.0 \u2192 684.5",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61325",
   "source_url": "https://github.com/vllm-project/vllm/pull/61325",
   "tg_tokens": null,
   "tps": 684.5,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 61440,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-61325-mi355x2-qwen38flashnext-stock-mtp-60k-c1",
   "model": "Qwen3.8-Flash-Next",
   "notes": "vLLM PR #61325 mono decode layer opt-in VLLM_ROCM_MONO_DECODE=1, TP2 on 2x MI355X (gfx950), checkpoint amd/Qwen3.8-Flash-Next-Quark-MXFP4, nightly-2eb491e1 + branch, arms differ only in the flag, one run per cell; aggregate output tok/s (sglang.bench_serving style ISL/OSL); stock arm with MTP-3 (num_speculative_tokens 3, max-num-batched-tokens 8192, re:mtp\\..* exclude fix applied), verify steps up to 8 rows run on the kernels; ; ISL/OSL 60k/600; mean TTFT 3,319 \u2192 3,300 ms; mean TPOT 2.56 \u2192 2.02 ms (-21%); config=stock-mtp",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "Qwen3.8-Flash-Next; MI355X; MXFP4; MTP-3; 60k / 600 | 1 (6) | 123.7 \u2192 133.0",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61325",
   "source_url": "https://github.com/vllm-project/vllm/pull/61325",
   "tg_tokens": null,
   "tps": 123.7,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 61440,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-61325-mi355x2-qwen38flashnext-mono-mtp-60k-c1",
   "model": "Qwen3.8-Flash-Next",
   "notes": "vLLM PR #61325 mono decode layer opt-in VLLM_ROCM_MONO_DECODE=1, TP2 on 2x MI355X (gfx950), checkpoint amd/Qwen3.8-Flash-Next-Quark-MXFP4, nightly-2eb491e1 + branch, arms differ only in the flag, one run per cell; aggregate output tok/s (sglang.bench_serving style ISL/OSL); mono arm (flag on) with MTP-3 (num_speculative_tokens 3, max-num-batched-tokens 8192, re:mtp\\..* exclude fix applied), verify steps up to 8 rows run on the kernels; ; ISL/OSL 60k/600; mean TTFT 3,319 \u2192 3,300 ms; mean TPOT 2.56 \u2192 2.02 ms (-21%); config=mono-mtp",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "Qwen3.8-Flash-Next; MI355X; MXFP4; MTP-3; 60k / 600 | 1 (6) | 123.7 \u2192 133.0",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61325",
   "source_url": "https://github.com/vllm-project/vllm/pull/61325",
   "tg_tokens": null,
   "tps": 133.0,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 96,
   "ctx": null,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "4x B200",
   "id": "vllm-60895-4xb200-glm53flash-tp4",
   "model": "GLM-5.3-Flash",
   "notes": "vLLM PR #60895 DeepGEMM FP8 MegaMoE (routed+shared in one kernel, EP dispatch/combine), official FP8 checkpoint (block FP8 E4M3 requantized to UE8M0 at load), 4x B200, nightly 8cbd5d03, MTP-5 synthetic acceptance 3.61, KV cache fp8; AgentX replay at concurrency 96, one run per arm; aggregate output tok/s; TP4 (--tensor-parallel-size 4, intra-node DP1 on the router); TTFT avg 3.93 s, ITL avg 18.3 ms, output tok/s per user 64.0; config=tp4",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "FP8",
   "quote": "GLM-5.3-Flash; 4x B200; FP8; output tok/s | 3,152 | 2,063 | 3,332",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #60895",
   "source_url": "https://github.com/vllm-project/vllm/pull/60895",
   "tg_tokens": null,
   "tps": 3152,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 96,
   "ctx": null,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "4x B200",
   "id": "vllm-60895-4xb200-glm53flash-dep4-generic",
   "model": "GLM-5.3-Flash",
   "notes": "vLLM PR #60895 DeepGEMM FP8 MegaMoE (routed+shared in one kernel, EP dispatch/combine), official FP8 checkpoint (block FP8 E4M3 requantized to UE8M0 at load), 4x B200, nightly 8cbd5d03, MTP-5 synthetic acceptance 3.61, KV cache fp8; AgentX replay at concurrency 96, one run per arm; aggregate output tok/s; DEP4 generic MoE path (picked FLASHINFER_TRTLLM for the FP8 experts); TTFT avg 2.36 s, ITL avg 45.6 ms, output tok/s per user 26.2; config=dep4-generic",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "FP8",
   "quote": "GLM-5.3-Flash; 4x B200; FP8; output tok/s | 3,152 | 2,063 | 3,332",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #60895",
   "source_url": "https://github.com/vllm-project/vllm/pull/60895",
   "tg_tokens": null,
   "tps": 2063,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 96,
   "ctx": null,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "4x B200",
   "id": "vllm-60895-4xb200-glm53flash-dep4-megamoe",
   "model": "GLM-5.3-Flash",
   "notes": "vLLM PR #60895 DeepGEMM FP8 MegaMoE (routed+shared in one kernel, EP dispatch/combine), official FP8 checkpoint (block FP8 E4M3 requantized to UE8M0 at load), 4x B200, nightly 8cbd5d03, MTP-5 synthetic acceptance 3.61, KV cache fp8; AgentX replay at concurrency 96, one run per arm; aggregate output tok/s; DEP4 with deep_gemm_mega_moe backend (TP1 DP4 expert-parallel); TTFT avg 0.81 s, ITL avg 19.5 ms, output tok/s per user 64.1; config=dep4-megamoe",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "FP8",
   "quote": "GLM-5.3-Flash; 4x B200; FP8; output tok/s | 3,152 | 2,063 | 3,332",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #60895",
   "source_url": "https://github.com/vllm-project/vllm/pull/60895",
   "tg_tokens": null,
   "tps": 3332,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "GB10",
   "id": "llama-30380-gb10-qwen38flashnext-full",
   "model": "Qwen3.8-Flash-Next",
   "notes": "llama.cpp PR #30380 --spec-draft-vocab-dynamic 16384 (trimmed draft LM head, low-rank scorer picks K best above id 32768); SPEED-Bench qualitative 165 samples 256 output tokens temp 0, one request at a time, --spec-type draft-mtp draft max 3; decode tok/s mean of 2 rounds; full draft head arm; target UD-IQ4_XS, drafter MTP Q4_K_M; gain with trimmed vocab +5.9%; acceptance 58.0% / 57.8%; config=full-head",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "UD-IQ4_XS",
   "quote": "Qwen3.8-Flash-Next; GB10; UD-IQ4_XS; DGX Spark GB10; Qwen3.8-Flash-Next UD-IQ4_XS / MTP Q4_K_M | 41.51 | 43.97",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30380",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30380",
   "tg_tokens": null,
   "tps": 41.51,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "GB10",
   "id": "llama-30380-gb10-qwen38flashnext-trim",
   "model": "Qwen3.8-Flash-Next",
   "notes": "llama.cpp PR #30380 --spec-draft-vocab-dynamic 16384 (trimmed draft LM head, low-rank scorer picks K best above id 32768); SPEED-Bench qualitative 165 samples 256 output tokens temp 0, one request at a time, --spec-type draft-mtp draft max 3; decode tok/s mean of 2 rounds; trimmed draft vocab arm (K=16384); target UD-IQ4_XS, drafter MTP Q4_K_M; gain vs full head +5.9%; acceptance 58.0% / 57.8%; config=trim-16384",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "UD-IQ4_XS",
   "quote": "Qwen3.8-Flash-Next; GB10; UD-IQ4_XS; DGX Spark GB10; Qwen3.8-Flash-Next UD-IQ4_XS / MTP Q4_K_M | 41.51 | 43.97",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30380",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30380",
   "tg_tokens": null,
   "tps": 43.97,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "GB10",
   "id": "llama-30380-gb10-gemma412b-full",
   "model": "Gemma 4 12B",
   "notes": "llama.cpp PR #30380 --spec-draft-vocab-dynamic 16384 (trimmed draft LM head, low-rank scorer picks K best above id 32768); SPEED-Bench qualitative 165 samples 256 output tokens temp 0, one request at a time, --spec-type draft-mtp draft max 3; decode tok/s mean of 2 rounds; full draft head arm; target Gemma-4-12B-it Q4_K_M, drafter assistant Q8_0; gain with trimmed vocab +1.6%; acceptance 66.8% / 66.7%; config=full-head",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "Gemma 4 12B; GB10; Q4_K_M; DGX Spark GB10; Gemma-4-12B-it Q4_K_M / assistant Q8_0 | 67.31 | 68.42",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30380",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30380",
   "tg_tokens": null,
   "tps": 67.31,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "GB10",
   "id": "llama-30380-gb10-gemma412b-trim",
   "model": "Gemma 4 12B",
   "notes": "llama.cpp PR #30380 --spec-draft-vocab-dynamic 16384 (trimmed draft LM head, low-rank scorer picks K best above id 32768); SPEED-Bench qualitative 165 samples 256 output tokens temp 0, one request at a time, --spec-type draft-mtp draft max 3; decode tok/s mean of 2 rounds; trimmed draft vocab arm (K=16384); target Gemma-4-12B-it Q4_K_M, drafter assistant Q8_0; gain vs full head +1.6%; acceptance 66.8% / 66.7%; config=trim-16384",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "Gemma 4 12B; GB10; Q4_K_M; DGX Spark GB10; Gemma-4-12B-it Q4_K_M / assistant Q8_0 | 67.31 | 68.42",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30380",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30380",
   "tg_tokens": null,
   "tps": 68.42,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "GB10",
   "id": "llama-30380-gb10-qwen3627b-full",
   "model": "Qwen3.6-27B",
   "notes": "llama.cpp PR #30380 --spec-draft-vocab-dynamic 16384 (trimmed draft LM head, low-rank scorer picks K best above id 32768); SPEED-Bench qualitative 165 samples 256 output tokens temp 0, one request at a time, --spec-type draft-mtp draft max 3; decode tok/s mean of 2 rounds; full draft head arm; target Q4_K_M, drafter MTP Q8_0; gain with trimmed vocab +10.2%; acceptance 65.2% / 64.4%; config=full-head",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "Qwen3.6-27B; GB10; Q4_K_M; DGX Spark GB10; Qwen3.6-27B Q4_K_M / MTP Q8_0 | 24.29 | 26.77",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30380",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30380",
   "tg_tokens": null,
   "tps": 24.29,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "GB10",
   "id": "llama-30380-gb10-qwen3627b-trim",
   "model": "Qwen3.6-27B",
   "notes": "llama.cpp PR #30380 --spec-draft-vocab-dynamic 16384 (trimmed draft LM head, low-rank scorer picks K best above id 32768); SPEED-Bench qualitative 165 samples 256 output tokens temp 0, one request at a time, --spec-type draft-mtp draft max 3; decode tok/s mean of 2 rounds; trimmed draft vocab arm (K=16384); target Q4_K_M, drafter MTP Q8_0; gain vs full head +10.2%; acceptance 65.2% / 64.4%; config=trim-16384",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "Qwen3.6-27B; GB10; Q4_K_M; DGX Spark GB10; Qwen3.6-27B Q4_K_M / MTP Q8_0 | 24.29 | 26.77",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30380",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30380",
   "tg_tokens": null,
   "tps": 26.77,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "RTX PRO 6000 Blackwell Workstation Edition",
   "id": "llama-30380-rtxpro6000-qwen38flashnext-full",
   "model": "Qwen3.8-Flash-Next",
   "notes": "llama.cpp PR #30380 --spec-draft-vocab-dynamic 16384 (trimmed draft LM head, low-rank scorer picks K best above id 32768); SPEED-Bench qualitative 165 samples 256 output tokens temp 0, one request at a time, --spec-type draft-mtp draft max 3; decode tok/s mean of 2 rounds; full draft head arm; target UD-IQ4_XS, drafter MTP Q4_K_M; gain with trimmed vocab +0.9%; acceptance 62.7% / 62.2%; config=full-head",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "UD-IQ4_XS",
   "quote": "Qwen3.8-Flash-Next; RTX PRO 6000 Blackwell Workstation Edition; UD-IQ4_XS; RTX PRO 6000 Blackwell Workstation Edition; Qwen3.8-Flash-Next UD-IQ4_XS / MTP Q4_K_M | 155.87 | 157.33",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30380",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30380",
   "tg_tokens": null,
   "tps": 155.87,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "RTX PRO 6000 Blackwell Workstation Edition",
   "id": "llama-30380-rtxpro6000-qwen38flashnext-trim",
   "model": "Qwen3.8-Flash-Next",
   "notes": "llama.cpp PR #30380 --spec-draft-vocab-dynamic 16384 (trimmed draft LM head, low-rank scorer picks K best above id 32768); SPEED-Bench qualitative 165 samples 256 output tokens temp 0, one request at a time, --spec-type draft-mtp draft max 3; decode tok/s mean of 2 rounds; trimmed draft vocab arm (K=16384); target UD-IQ4_XS, drafter MTP Q4_K_M; gain vs full head +0.9%; acceptance 62.7% / 62.2%; config=trim-16384",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "UD-IQ4_XS",
   "quote": "Qwen3.8-Flash-Next; RTX PRO 6000 Blackwell Workstation Edition; UD-IQ4_XS; RTX PRO 6000 Blackwell Workstation Edition; Qwen3.8-Flash-Next UD-IQ4_XS / MTP Q4_K_M | 155.87 | 157.33",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30380",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30380",
   "tg_tokens": null,
   "tps": 157.33,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "RTX PRO 6000 Blackwell Workstation Edition",
   "id": "llama-30380-rtxpro6000-gemma412b-full",
   "model": "Gemma 4 12B",
   "notes": "llama.cpp PR #30380 --spec-draft-vocab-dynamic 16384 (trimmed draft LM head, low-rank scorer picks K best above id 32768); SPEED-Bench qualitative 165 samples 256 output tokens temp 0, one request at a time, --spec-type draft-mtp draft max 3; decode tok/s mean of 2 rounds; full draft head arm; target Gemma-4-12B-it Q4_K_M, drafter assistant Q8_0; gain with trimmed vocab +0.5%; acceptance 68.9% / 68.9%; config=full-head",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "Gemma 4 12B; RTX PRO 6000 Blackwell Workstation Edition; Q4_K_M; RTX PRO 6000 Blackwell Workstation Edition; Gemma-4-12B-it Q4_K_M / assistant Q8_0 | 285.98 | 287.33",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30380",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30380",
   "tg_tokens": null,
   "tps": 285.98,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "RTX PRO 6000 Blackwell Workstation Edition",
   "id": "llama-30380-rtxpro6000-gemma412b-trim",
   "model": "Gemma 4 12B",
   "notes": "llama.cpp PR #30380 --spec-draft-vocab-dynamic 16384 (trimmed draft LM head, low-rank scorer picks K best above id 32768); SPEED-Bench qualitative 165 samples 256 output tokens temp 0, one request at a time, --spec-type draft-mtp draft max 3; decode tok/s mean of 2 rounds; trimmed draft vocab arm (K=16384); target Gemma-4-12B-it Q4_K_M, drafter assistant Q8_0; gain vs full head +0.5%; acceptance 68.9% / 68.9%; config=trim-16384",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "Gemma 4 12B; RTX PRO 6000 Blackwell Workstation Edition; Q4_K_M; RTX PRO 6000 Blackwell Workstation Edition; Gemma-4-12B-it Q4_K_M / assistant Q8_0 | 285.98 | 287.33",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30380",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30380",
   "tg_tokens": null,
   "tps": 287.33,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "RTX PRO 6000 Blackwell Workstation Edition",
   "id": "llama-30380-rtxpro6000-qwen3627b-full",
   "model": "Qwen3.6-27B",
   "notes": "llama.cpp PR #30380 --spec-draft-vocab-dynamic 16384 (trimmed draft LM head, low-rank scorer picks K best above id 32768); SPEED-Bench qualitative 165 samples 256 output tokens temp 0, one request at a time, --spec-type draft-mtp draft max 3; decode tok/s mean of 2 rounds; full draft head arm; target Q4_K_M, drafter MTP Q8_0; gain with trimmed vocab +6.7%; acceptance 65.6% / 64.8%; config=full-head",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "Qwen3.6-27B; RTX PRO 6000 Blackwell Workstation Edition; Q4_K_M; RTX PRO 6000 Blackwell Workstation Edition; Qwen3.6-27B Q4_K_M / MTP Q8_0 | 135.98 | 145.14",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30380",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30380",
   "tg_tokens": null,
   "tps": 135.98,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "RTX PRO 6000 Blackwell Workstation Edition",
   "id": "llama-30380-rtxpro6000-qwen3627b-trim",
   "model": "Qwen3.6-27B",
   "notes": "llama.cpp PR #30380 --spec-draft-vocab-dynamic 16384 (trimmed draft LM head, low-rank scorer picks K best above id 32768); SPEED-Bench qualitative 165 samples 256 output tokens temp 0, one request at a time, --spec-type draft-mtp draft max 3; decode tok/s mean of 2 rounds; trimmed draft vocab arm (K=16384); target Q4_K_M, drafter MTP Q8_0; gain vs full head +6.7%; acceptance 65.6% / 64.8%; config=trim-16384",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "Qwen3.6-27B; RTX PRO 6000 Blackwell Workstation Edition; Q4_K_M; RTX PRO 6000 Blackwell Workstation Edition; Qwen3.6-27B Q4_K_M / MTP Q8_0 | 135.98 | 145.14",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30380",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30380",
   "tg_tokens": null,
   "tps": 145.14,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "llama-30380-rtx5090-gemma412b-full",
   "model": "Gemma 4 12B",
   "notes": "llama.cpp PR #30380 --spec-draft-vocab-dynamic 16384 (trimmed draft LM head, low-rank scorer picks K best above id 32768); SPEED-Bench qualitative 165 samples 256 output tokens temp 0, one request at a time, --spec-type draft-mtp draft max 3; decode tok/s mean of 2 rounds; full draft head arm; target Gemma-4-12B-it Q4_K_M, drafter assistant Q8_0; gain with trimmed vocab -0.3% within noise; acceptance 67.5% / 67.5%; config=full-head",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "Gemma 4 12B; RTX 5090; Q4_K_M; RTX 5090; Gemma-4-12B-it Q4_K_M / assistant Q8_0 | 279.09 | 278.37",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30380",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30380",
   "tg_tokens": null,
   "tps": 279.09,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "llama-30380-rtx5090-gemma412b-trim",
   "model": "Gemma 4 12B",
   "notes": "llama.cpp PR #30380 --spec-draft-vocab-dynamic 16384 (trimmed draft LM head, low-rank scorer picks K best above id 32768); SPEED-Bench qualitative 165 samples 256 output tokens temp 0, one request at a time, --spec-type draft-mtp draft max 3; decode tok/s mean of 2 rounds; trimmed draft vocab arm (K=16384); target Gemma-4-12B-it Q4_K_M, drafter assistant Q8_0; gain vs full head -0.3% within noise; acceptance 67.5% / 67.5%; config=trim-16384",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "Gemma 4 12B; RTX 5090; Q4_K_M; RTX 5090; Gemma-4-12B-it Q4_K_M / assistant Q8_0 | 279.09 | 278.37",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30380",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30380",
   "tg_tokens": null,
   "tps": 278.37,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "llama-30380-rtx5090-qwen3627b-full",
   "model": "Qwen3.6-27B",
   "notes": "llama.cpp PR #30380 --spec-draft-vocab-dynamic 16384 (trimmed draft LM head, low-rank scorer picks K best above id 32768); SPEED-Bench qualitative 165 samples 256 output tokens temp 0, one request at a time, --spec-type draft-mtp draft max 3; decode tok/s mean of 2 rounds; full draft head arm; target Q4_K_M, drafter MTP Q8_0; gain with trimmed vocab +5.0%; acceptance 64.1% / 62.7%; config=full-head",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "Qwen3.6-27B; RTX 5090; Q4_K_M; RTX 5090; Qwen3.6-27B Q4_K_M / MTP Q8_0 | 130.11 | 136.59",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30380",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30380",
   "tg_tokens": null,
   "tps": 130.11,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "llama-30380-rtx5090-qwen3627b-trim",
   "model": "Qwen3.6-27B",
   "notes": "llama.cpp PR #30380 --spec-draft-vocab-dynamic 16384 (trimmed draft LM head, low-rank scorer picks K best above id 32768); SPEED-Bench qualitative 165 samples 256 output tokens temp 0, one request at a time, --spec-type draft-mtp draft max 3; decode tok/s mean of 2 rounds; trimmed draft vocab arm (K=16384); target Q4_K_M, drafter MTP Q8_0; gain vs full head +5.0%; acceptance 64.1% / 62.7%; config=trim-16384",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "Qwen3.6-27B; RTX 5090; Q4_K_M; RTX 5090; Qwen3.6-27B Q4_K_M / MTP Q8_0 | 130.11 | 136.59",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30380",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30380",
   "tg_tokens": null,
   "tps": 136.59,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 8,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "llama-30276-rtx5060ti-bonsai227b-c32768-np8",
   "model": "Bonsai 2 27B",
   "notes": "llama.cpp issue #30276 decode-throughput collapse at large total context with --parallel >= 4 (bug report); llama-server -ngl 99 -fa on -ctk q8_0 -ctv q8_0, Bonsai-2 27B PQ2_0 GGUF 7.2 GB on RTX 5060 Ti 16 GB (sm_120, CUDA 13.1); aggregate decode tok/s at --parallel N; n_ctx 32768, --parallel 8, n_ctx_slot 4096; 15.97 t/s per request; the c=262144 np=4 config collapses to 3.6-3.9 t/s per request (range, not mined)",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "PQ2_0",
   "quote": "Bonsai 2 27B; RTX 5060 Ti; PQ2_0; 32768 | 8 | 4096 | 15.97 t/s | 127.8 t/s",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp issue #30276",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30276",
   "tg_tokens": null,
   "tps": 127.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 8,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "llama-30276-rtx5060ti-bonsai227b-c131072-np8",
   "model": "Bonsai 2 27B",
   "notes": "llama.cpp issue #30276 decode-throughput collapse at large total context with --parallel >= 4 (bug report); llama-server -ngl 99 -fa on -ctk q8_0 -ctv q8_0, Bonsai-2 27B PQ2_0 GGUF 7.2 GB on RTX 5060 Ti 16 GB (sm_120, CUDA 13.1); aggregate decode tok/s at --parallel N; n_ctx 131072, --parallel 8, n_ctx_slot 16384; 15.91 t/s per request; the c=262144 np=4 config collapses to 3.6-3.9 t/s per request (range, not mined)",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "PQ2_0",
   "quote": "Bonsai 2 27B; RTX 5060 Ti; PQ2_0; 131072 | 8 | 16384 | 15.91 | 127.3",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp issue #30276",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30276",
   "tg_tokens": null,
   "tps": 127.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 4,
   "ctx": 196608,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "llama-30276-rtx5060ti-bonsai227b-c196608-np4",
   "model": "Bonsai 2 27B",
   "notes": "llama.cpp issue #30276 decode-throughput collapse at large total context with --parallel >= 4 (bug report); llama-server -ngl 99 -fa on -ctk q8_0 -ctv q8_0, Bonsai-2 27B PQ2_0 GGUF 7.2 GB on RTX 5060 Ti 16 GB (sm_120, CUDA 13.1); aggregate decode tok/s at --parallel N; n_ctx 196608, --parallel 4, n_ctx_slot 49152; 28.11 t/s per request; the c=262144 np=4 config collapses to 3.6-3.9 t/s per request (range, not mined)",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "PQ2_0",
   "quote": "Bonsai 2 27B; RTX 5060 Ti; PQ2_0; 196608 | 4 | 49152 | 28.11 | 112.4",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp issue #30276",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30276",
   "tg_tokens": null,
   "tps": 112.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 4,
   "ctx": 229376,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "llama-30276-rtx5060ti-bonsai227b-c229376-np4",
   "model": "Bonsai 2 27B",
   "notes": "llama.cpp issue #30276 decode-throughput collapse at large total context with --parallel >= 4 (bug report); llama-server -ngl 99 -fa on -ctk q8_0 -ctv q8_0, Bonsai-2 27B PQ2_0 GGUF 7.2 GB on RTX 5060 Ti 16 GB (sm_120, CUDA 13.1); aggregate decode tok/s at --parallel N; n_ctx 229376, --parallel 4, n_ctx_slot 57344; 27.94 t/s per request; the c=262144 np=4 config collapses to 3.6-3.9 t/s per request (range, not mined)",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "PQ2_0",
   "quote": "Bonsai 2 27B; RTX 5060 Ti; PQ2_0; 229376 | 4 | 57344 | 27.94 | 111.8",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp issue #30276",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30276",
   "tg_tokens": null,
   "tps": 111.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 4,
   "ctx": 245760,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "llama-30276-rtx5060ti-bonsai227b-c245760-np4",
   "model": "Bonsai 2 27B",
   "notes": "llama.cpp issue #30276 decode-throughput collapse at large total context with --parallel >= 4 (bug report); llama-server -ngl 99 -fa on -ctk q8_0 -ctv q8_0, Bonsai-2 27B PQ2_0 GGUF 7.2 GB on RTX 5060 Ti 16 GB (sm_120, CUDA 13.1); aggregate decode tok/s at --parallel N; n_ctx 245760, --parallel 4, n_ctx_slot 61440; 28.70 t/s per request; the c=262144 np=4 config collapses to 3.6-3.9 t/s per request (range, not mined)",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "PQ2_0",
   "quote": "Bonsai 2 27B; RTX 5060 Ti; PQ2_0; 245760 | 4 | 61440 | 28.70 | 114.8",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp issue #30276",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30276",
   "tg_tokens": null,
   "tps": 114.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 2,
   "ctx": 262144,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "llama-30276-rtx5060ti-bonsai227b-c262144-np2",
   "model": "Bonsai 2 27B",
   "notes": "llama.cpp issue #30276 decode-throughput collapse at large total context with --parallel >= 4 (bug report); llama-server -ngl 99 -fa on -ctk q8_0 -ctv q8_0, Bonsai-2 27B PQ2_0 GGUF 7.2 GB on RTX 5060 Ti 16 GB (sm_120, CUDA 13.1); aggregate decode tok/s at --parallel N; n_ctx 262144, --parallel 2, n_ctx_slot 131072; 39.83 t/s per request; the c=262144 np=4 config collapses to 3.6-3.9 t/s per request (range, not mined)",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "PQ2_0",
   "quote": "Bonsai 2 27B; RTX 5060 Ti; PQ2_0; 262144 | 2 | 131072 | 39.83 | 79.7",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp issue #30276",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30276",
   "tg_tokens": null,
   "tps": 79.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 1,
   "ctx": 262144,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "llama-30276-rtx5060ti-bonsai227b-c262144-np1",
   "model": "Bonsai 2 27B",
   "notes": "llama.cpp issue #30276 decode-throughput collapse at large total context with --parallel >= 4 (bug report); llama-server -ngl 99 -fa on -ctk q8_0 -ctv q8_0, Bonsai-2 27B PQ2_0 GGUF 7.2 GB on RTX 5060 Ti 16 GB (sm_120, CUDA 13.1); aggregate decode tok/s at --parallel N; n_ctx 262144, --parallel 1, n_ctx_slot 262144; 47.9 t/s per request; the c=262144 np=4 config collapses to 3.6-3.9 t/s per request (range, not mined)",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "PQ2_0",
   "quote": "Bonsai 2 27B; RTX 5060 Ti; PQ2_0; 262144 | 1 | 262144 | 47.9 | 47.9",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp issue #30276",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30276",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 8,
   "ctx": 262144,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "llama-30276-rtx5060ti-bonsai227b-c262144-np8",
   "model": "Bonsai 2 27B",
   "notes": "llama.cpp issue #30276 decode-throughput collapse at large total context with --parallel >= 4 (bug report); llama-server -ngl 99 -fa on -ctk q8_0 -ctv q8_0, Bonsai-2 27B PQ2_0 GGUF 7.2 GB on RTX 5060 Ti 16 GB (sm_120, CUDA 13.1); aggregate decode tok/s at --parallel N; n_ctx 262144, --parallel 8, n_ctx_slot 32768; 1.65 t/s per request, collapsed config; the c=262144 np=4 config collapses to 3.6-3.9 t/s per request (range, not mined)",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "PQ2_0",
   "quote": "Bonsai 2 27B; RTX 5060 Ti; PQ2_0; 262144 | 8 | 32768 | 1.65 | 13.3",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp issue #30276",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30276",
   "tg_tokens": null,
   "tps": 13.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 2,
   "ctx": 524288,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "llama-30276-rtx5060ti-bonsai227b-c524288-np2",
   "model": "Bonsai 2 27B",
   "notes": "llama.cpp issue #30276 decode-throughput collapse at large total context with --parallel >= 4 (bug report); llama-server -ngl 99 -fa on -ctk q8_0 -ctv q8_0, Bonsai-2 27B PQ2_0 GGUF 7.2 GB on RTX 5060 Ti 16 GB (sm_120, CUDA 13.1); aggregate decode tok/s at --parallel N; n_ctx 524288, --parallel 2, n_ctx_slot 262144; 5.48 t/s per request; the c=262144 np=4 config collapses to 3.6-3.9 t/s per request (range, not mined)",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "PQ2_0",
   "quote": "Bonsai 2 27B; RTX 5060 Ti; PQ2_0; 524288 | 2 | 262144 | 5.48 | 11.0",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp issue #30276",
   "source_url": "https://github.com/ggml-org/llama.cpp/issues/30276",
   "tg_tokens": null,
   "tps": 11.0,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-61264-mi355x-qwen253b-eager-i1024c16-base",
   "model": "Qwen2.5-3B-Instruct",
   "notes": "vLLM PR #61264 batched short TurboQuant continuation/verification requests with split-KV attention; one MI355X, Qwen/Qwen2.5-3B-Instruct BF16 weights, turboquant_4bit_nc KV cache, ngram speculation 3 draft tokens lookup 2..5, prefix caching and chunked prefill on, greedy 512-token cap; synthetic repetition workload, three alternating server pairs, arm medians; output tok/s includes prefill and decode; eager execution campaign, ISL/OSL 1024/128 at concurrency 16, base arm; median paired speedup patched/base 2.712x; config=base-eager",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "BF16",
   "quote": "Qwen2.5-3B-Instruct; MI355X; BF16; 1024 / 16 / 128 | 347.68 | 942.92",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61264",
   "source_url": "https://github.com/vllm-project/vllm/pull/61264",
   "tg_tokens": null,
   "tps": 347.68,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-61264-mi355x-qwen253b-eager-i1024c16-patched",
   "model": "Qwen2.5-3B-Instruct",
   "notes": "vLLM PR #61264 batched short TurboQuant continuation/verification requests with split-KV attention; one MI355X, Qwen/Qwen2.5-3B-Instruct BF16 weights, turboquant_4bit_nc KV cache, ngram speculation 3 draft tokens lookup 2..5, prefix caching and chunked prefill on, greedy 512-token cap; synthetic repetition workload, three alternating server pairs, arm medians; output tok/s includes prefill and decode; eager execution campaign, ISL/OSL 1024/128 at concurrency 16, patched arm; median paired speedup vs base 2.712x; config=patched-eager",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "BF16",
   "quote": "Qwen2.5-3B-Instruct; MI355X; BF16; 1024 / 16 / 128 | 347.68 | 942.92",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61264",
   "source_url": "https://github.com/vllm-project/vllm/pull/61264",
   "tg_tokens": null,
   "tps": 942.92,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-61264-mi355x-qwen253b-eager-i8192c16-base",
   "model": "Qwen2.5-3B-Instruct",
   "notes": "vLLM PR #61264 batched short TurboQuant continuation/verification requests with split-KV attention; one MI355X, Qwen/Qwen2.5-3B-Instruct BF16 weights, turboquant_4bit_nc KV cache, ngram speculation 3 draft tokens lookup 2..5, prefix caching and chunked prefill on, greedy 512-token cap; synthetic repetition workload, three alternating server pairs, arm medians; output tok/s includes prefill and decode; eager execution campaign, ISL/OSL 8192/128 at concurrency 16, base arm; median paired speedup patched/base 1.660x; config=base-eager",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "BF16",
   "quote": "Qwen2.5-3B-Instruct; MI355X; BF16; 8192 / 16 / 128 | 443.52 | 736.38",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61264",
   "source_url": "https://github.com/vllm-project/vllm/pull/61264",
   "tg_tokens": null,
   "tps": 443.52,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-61264-mi355x-qwen253b-eager-i8192c16-patched",
   "model": "Qwen2.5-3B-Instruct",
   "notes": "vLLM PR #61264 batched short TurboQuant continuation/verification requests with split-KV attention; one MI355X, Qwen/Qwen2.5-3B-Instruct BF16 weights, turboquant_4bit_nc KV cache, ngram speculation 3 draft tokens lookup 2..5, prefix caching and chunked prefill on, greedy 512-token cap; synthetic repetition workload, three alternating server pairs, arm medians; output tok/s includes prefill and decode; eager execution campaign, ISL/OSL 8192/128 at concurrency 16, patched arm; median paired speedup vs base 1.660x; config=patched-eager",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "BF16",
   "quote": "Qwen2.5-3B-Instruct; MI355X; BF16; 8192 / 16 / 128 | 443.52 | 736.38",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61264",
   "source_url": "https://github.com/vllm-project/vllm/pull/61264",
   "tg_tokens": null,
   "tps": 736.38,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 64,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-61264-mi355x-qwen253b-eager-i8192c64-base",
   "model": "Qwen2.5-3B-Instruct",
   "notes": "vLLM PR #61264 batched short TurboQuant continuation/verification requests with split-KV attention; one MI355X, Qwen/Qwen2.5-3B-Instruct BF16 weights, turboquant_4bit_nc KV cache, ngram speculation 3 draft tokens lookup 2..5, prefix caching and chunked prefill on, greedy 512-token cap; synthetic repetition workload, three alternating server pairs, arm medians; output tok/s includes prefill and decode; eager execution campaign, ISL/OSL 8192/128 at concurrency 64, base arm; median paired speedup patched/base 1.139x; config=base-eager",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "BF16",
   "quote": "Qwen2.5-3B-Instruct; MI355X; BF16; 8192 / 64 / 128 | 499.36 | 568.81",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61264",
   "source_url": "https://github.com/vllm-project/vllm/pull/61264",
   "tg_tokens": null,
   "tps": 499.36,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 64,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-61264-mi355x-qwen253b-eager-i8192c64-patched",
   "model": "Qwen2.5-3B-Instruct",
   "notes": "vLLM PR #61264 batched short TurboQuant continuation/verification requests with split-KV attention; one MI355X, Qwen/Qwen2.5-3B-Instruct BF16 weights, turboquant_4bit_nc KV cache, ngram speculation 3 draft tokens lookup 2..5, prefix caching and chunked prefill on, greedy 512-token cap; synthetic repetition workload, three alternating server pairs, arm medians; output tok/s includes prefill and decode; eager execution campaign, ISL/OSL 8192/128 at concurrency 64, patched arm; median paired speedup vs base 1.139x; config=patched-eager",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "BF16",
   "quote": "Qwen2.5-3B-Instruct; MI355X; BF16; 8192 / 64 / 128 | 499.36 | 568.81",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61264",
   "source_url": "https://github.com/vllm-project/vllm/pull/61264",
   "tg_tokens": null,
   "tps": 568.81,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-61264-mi355x-qwen253b-graph-i1024c16-base",
   "model": "Qwen2.5-3B-Instruct",
   "notes": "vLLM PR #61264 batched short TurboQuant continuation/verification requests with split-KV attention; one MI355X, Qwen/Qwen2.5-3B-Instruct BF16 weights, turboquant_4bit_nc KV cache, ngram speculation 3 draft tokens lookup 2..5, prefix caching and chunked prefill on, greedy 512-token cap; synthetic repetition workload, three alternating server pairs, arm medians; output tok/s includes prefill and decode; default graph-enabled serving scope check, ISL/OSL 1024/128 at concurrency 16, base arm; median paired speedup patched/base 0.997x; config=base-graph",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "BF16",
   "quote": "Qwen2.5-3B-Instruct; MI355X; BF16; 1024 / 16 / 128 | 2688.94 | 2681.27",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61264",
   "source_url": "https://github.com/vllm-project/vllm/pull/61264",
   "tg_tokens": null,
   "tps": 2688.94,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-61264-mi355x-qwen253b-graph-i1024c16-patched",
   "model": "Qwen2.5-3B-Instruct",
   "notes": "vLLM PR #61264 batched short TurboQuant continuation/verification requests with split-KV attention; one MI355X, Qwen/Qwen2.5-3B-Instruct BF16 weights, turboquant_4bit_nc KV cache, ngram speculation 3 draft tokens lookup 2..5, prefix caching and chunked prefill on, greedy 512-token cap; synthetic repetition workload, three alternating server pairs, arm medians; output tok/s includes prefill and decode; default graph-enabled serving scope check, ISL/OSL 1024/128 at concurrency 16, patched arm; median paired speedup vs base 0.997x; config=patched-graph",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "BF16",
   "quote": "Qwen2.5-3B-Instruct; MI355X; BF16; 1024 / 16 / 128 | 2688.94 | 2681.27",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61264",
   "source_url": "https://github.com/vllm-project/vllm/pull/61264",
   "tg_tokens": null,
   "tps": 2681.27,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-61264-mi355x-qwen253b-graph-i8192c16-base",
   "model": "Qwen2.5-3B-Instruct",
   "notes": "vLLM PR #61264 batched short TurboQuant continuation/verification requests with split-KV attention; one MI355X, Qwen/Qwen2.5-3B-Instruct BF16 weights, turboquant_4bit_nc KV cache, ngram speculation 3 draft tokens lookup 2..5, prefix caching and chunked prefill on, greedy 512-token cap; synthetic repetition workload, three alternating server pairs, arm medians; output tok/s includes prefill and decode; default graph-enabled serving scope check, ISL/OSL 8192/128 at concurrency 16, base arm; median paired speedup patched/base 1.038x; config=base-graph",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "BF16",
   "quote": "Qwen2.5-3B-Instruct; MI355X; BF16; 8192 / 16 / 128 | 918.01 | 953.94",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61264",
   "source_url": "https://github.com/vllm-project/vllm/pull/61264",
   "tg_tokens": null,
   "tps": 918.01,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-61264-mi355x-qwen253b-graph-i8192c16-patched",
   "model": "Qwen2.5-3B-Instruct",
   "notes": "vLLM PR #61264 batched short TurboQuant continuation/verification requests with split-KV attention; one MI355X, Qwen/Qwen2.5-3B-Instruct BF16 weights, turboquant_4bit_nc KV cache, ngram speculation 3 draft tokens lookup 2..5, prefix caching and chunked prefill on, greedy 512-token cap; synthetic repetition workload, three alternating server pairs, arm medians; output tok/s includes prefill and decode; default graph-enabled serving scope check, ISL/OSL 8192/128 at concurrency 16, patched arm; median paired speedup vs base 1.038x; config=patched-graph",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "BF16",
   "quote": "Qwen2.5-3B-Instruct; MI355X; BF16; 8192 / 16 / 128 | 918.01 | 953.94",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61264",
   "source_url": "https://github.com/vllm-project/vllm/pull/61264",
   "tg_tokens": null,
   "tps": 953.94,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 64,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-61264-mi355x-qwen253b-graph-i8192c64-base",
   "model": "Qwen2.5-3B-Instruct",
   "notes": "vLLM PR #61264 batched short TurboQuant continuation/verification requests with split-KV attention; one MI355X, Qwen/Qwen2.5-3B-Instruct BF16 weights, turboquant_4bit_nc KV cache, ngram speculation 3 draft tokens lookup 2..5, prefix caching and chunked prefill on, greedy 512-token cap; synthetic repetition workload, three alternating server pairs, arm medians; output tok/s includes prefill and decode; default graph-enabled serving scope check, ISL/OSL 8192/128 at concurrency 64, base arm; median paired speedup patched/base 1.003x; config=base-graph",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "BF16",
   "quote": "Qwen2.5-3B-Instruct; MI355X; BF16; 8192 / 64 / 128 | 570.16 | 570.92",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61264",
   "source_url": "https://github.com/vllm-project/vllm/pull/61264",
   "tg_tokens": null,
   "tps": 570.16,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 64,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-61264-mi355x-qwen253b-graph-i8192c64-patched",
   "model": "Qwen2.5-3B-Instruct",
   "notes": "vLLM PR #61264 batched short TurboQuant continuation/verification requests with split-KV attention; one MI355X, Qwen/Qwen2.5-3B-Instruct BF16 weights, turboquant_4bit_nc KV cache, ngram speculation 3 draft tokens lookup 2..5, prefix caching and chunked prefill on, greedy 512-token cap; synthetic repetition workload, three alternating server pairs, arm medians; output tok/s includes prefill and decode; default graph-enabled serving scope check, ISL/OSL 8192/128 at concurrency 64, patched arm; median paired speedup vs base 1.003x; config=patched-graph",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "BF16",
   "quote": "Qwen2.5-3B-Instruct; MI355X; BF16; 8192 / 64 / 128 | 570.16 | 570.92",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61264",
   "source_url": "https://github.com/vllm-project/vllm/pull/61264",
   "tg_tokens": null,
   "tps": 570.92,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "H20",
   "id": "vllm-61555-h20-dsv41-c1-eager",
   "model": "DeepSeek V4.1",
   "notes": "vLLM PR #61555 NVIDIA DeepSeek-V4.1 PCP decode-only CUDA graphs (stabilized graph-read metadata, graphs sized by rank-local requests), four H20 GPUs TP1/PCP4/EP4, fixed DSpark5, Humming MoE FP8 activations dense A16, fp8_ds_mla, shared CPU Engram, bounded replay on, APC off, chunk8192, max-len131072; GSM8K natural questions temp0 thinking off natural EOS real speculative acceptance; concurrency 1, eager decode arm (parent 9516299b), median of 3 repetitions; TPOT 29.293 \u2192 5.023 ms, TTFT 143.575 \u2192 138.941 ms; config=eager",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "FP8",
   "quote": "DeepSeek-V4.1; H20; FP8; c1 output tok/s, median [min-max] | 32.494 [31.889-32.985] | 169.756 [167.900-171.100]",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61555",
   "source_url": "https://github.com/vllm-project/vllm/pull/61555",
   "tg_tokens": null,
   "tps": 32.494,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "H20",
   "id": "vllm-61555-h20-dsv41-c1-graph",
   "model": "DeepSeek V4.1",
   "notes": "vLLM PR #61555 NVIDIA DeepSeek-V4.1 PCP decode-only CUDA graphs (stabilized graph-read metadata, graphs sized by rank-local requests), four H20 GPUs TP1/PCP4/EP4, fixed DSpark5, Humming MoE FP8 activations dense A16, fp8_ds_mla, shared CPU Engram, bounded replay on, APC off, chunk8192, max-len131072; GSM8K natural questions temp0 thinking off natural EOS real speculative acceptance; concurrency 1, decode CUDA graph arm (b2e6404), median of 3 repetitions; TPOT 29.293 \u2192 5.023 ms, TTFT 143.575 \u2192 138.941 ms; config=graph",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "FP8",
   "quote": "DeepSeek-V4.1; H20; FP8; c1 output tok/s, median [min-max] | 32.494 [31.889-32.985] | 169.756 [167.900-171.100]",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61555",
   "source_url": "https://github.com/vllm-project/vllm/pull/61555",
   "tg_tokens": null,
   "tps": 169.756,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": null,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "H20",
   "id": "vllm-61555-h20-dsv41-c4-eager",
   "model": "DeepSeek V4.1",
   "notes": "vLLM PR #61555 NVIDIA DeepSeek-V4.1 PCP decode-only CUDA graphs (stabilized graph-read metadata, graphs sized by rank-local requests), four H20 GPUs TP1/PCP4/EP4, fixed DSpark5, Humming MoE FP8 activations dense A16, fp8_ds_mla, shared CPU Engram, bounded replay on, APC off, chunk8192, max-len131072; GSM8K natural questions temp0 thinking off natural EOS real speculative acceptance; concurrency 4, eager decode arm (parent 9516299b), median of 3 repetitions; TPOT 29.936 \u2192 5.631 ms, TTFT 294.792 \u2192 284.813 ms; config=eager",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "FP8",
   "quote": "DeepSeek-V4.1; H20; FP8; c4 output tok/s, median [min-max] | 82.069 [81.581-97.824] | 369.892 [349.078-380.643]",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61555",
   "source_url": "https://github.com/vllm-project/vllm/pull/61555",
   "tg_tokens": null,
   "tps": 82.069,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": null,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "H20",
   "id": "vllm-61555-h20-dsv41-c4-graph",
   "model": "DeepSeek V4.1",
   "notes": "vLLM PR #61555 NVIDIA DeepSeek-V4.1 PCP decode-only CUDA graphs (stabilized graph-read metadata, graphs sized by rank-local requests), four H20 GPUs TP1/PCP4/EP4, fixed DSpark5, Humming MoE FP8 activations dense A16, fp8_ds_mla, shared CPU Engram, bounded replay on, APC off, chunk8192, max-len131072; GSM8K natural questions temp0 thinking off natural EOS real speculative acceptance; concurrency 4, decode CUDA graph arm (b2e6404), median of 3 repetitions; TPOT 29.936 \u2192 5.631 ms, TTFT 294.792 \u2192 284.813 ms; config=graph",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "FP8",
   "quote": "DeepSeek-V4.1; H20; FP8; c4 output tok/s, median [min-max] | 82.069 [81.581-97.824] | 369.892 [349.078-380.643]",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61555",
   "source_url": "https://github.com/vllm-project/vllm/pull/61555",
   "tg_tokens": null,
   "tps": 369.892,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "H20",
   "id": "vllm-61555-h20-dsv41-prefill64k-eager",
   "model": "DeepSeek V4.1",
   "notes": "vLLM PR #61555 NVIDIA DeepSeek-V4.1 PCP decode-only CUDA graphs (stabilized graph-read metadata, graphs sized by rank-local requests), four H20 GPUs TP1/PCP4/EP4, fixed DSpark5, Humming MoE FP8 activations dense A16, fp8_ds_mla, shared CPU Engram, bounded replay on, APC off, chunk8192, max-len131072; GSM8K natural questions temp0 thinking off natural EOS real speculative acceptance; cold 64K/c4/output1 prefill (8 requests per round, exact input hashes, cached_tokens=0), eager arm, median of 3 formal samples; effectively unchanged vs graph arm; config=eager-prefill",
   "params": null,
   "power_w": null,
   "pp_tps": 15255.437,
   "provenance": "sourced",
   "quant": "FP8",
   "quote": "DeepSeek-V4.1; H20; FP8; Cold 64K/c4/output1 prefill, input tok/s | 15,255.437 | 15,265.495",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61555",
   "source_url": "https://github.com/vllm-project/vllm/pull/61555",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "H20",
   "id": "vllm-61555-h20-dsv41-prefill64k-graph",
   "model": "DeepSeek V4.1",
   "notes": "vLLM PR #61555 NVIDIA DeepSeek-V4.1 PCP decode-only CUDA graphs (stabilized graph-read metadata, graphs sized by rank-local requests), four H20 GPUs TP1/PCP4/EP4, fixed DSpark5, Humming MoE FP8 activations dense A16, fp8_ds_mla, shared CPU Engram, bounded replay on, APC off, chunk8192, max-len131072; GSM8K natural questions temp0 thinking off natural EOS real speculative acceptance; cold 64K/c4/output1 prefill (8 requests per round, exact input hashes, cached_tokens=0), decode-graph arm (prefill stays eager in both arms), median of 3 formal samples; effectively unchanged vs eager arm; config=graph-prefill",
   "params": null,
   "power_w": null,
   "pp_tps": 15265.495,
   "provenance": "sourced",
   "quant": "FP8",
   "quote": "DeepSeek-V4.1; H20; FP8; Cold 64K/c4/output1 prefill, input tok/s | 15,255.437 | 15,265.495",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61555",
   "source_url": "https://github.com/vllm-project/vllm/pull/61555",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-60932-gb300-dsv41flash-c1-uva",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #60932 peer-device expert tier RFC measurement, DGX Station GB300 with RTX PRO 6000 Blackwell sidecar GPU, DeepSeek-V4.1-Flash native MXFP4 experts, 285 hot / 99 cold experts per layer, sidecar computes cold experts over NVLink-C2C; decode at concurrency 1, cold experts in host memory (UVA offload) arm; sidecar ratio 1.97x; GB300 board power decode C1/C16/C32 396/449/481 W; config=uva",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "DeepSeek-V4.1-Flash; GB300; MXFP4; decode C1, tok/s per user (mean of 3) | 268.7 | 528.7",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #60932",
   "source_url": "https://github.com/vllm-project/vllm/pull/60932",
   "tg_tokens": null,
   "tps": 268.7,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-60932-gb300-dsv41flash-c1-sidecar",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #60932 peer-device expert tier RFC measurement, DGX Station GB300 with RTX PRO 6000 Blackwell sidecar GPU, DeepSeek-V4.1-Flash native MXFP4 experts, 285 hot / 99 cold experts per layer, sidecar computes cold experts over NVLink-C2C; decode at concurrency 1, cold experts on the RTX PRO 6000 sidecar arm; ratio vs UVA 1.97x; combined GB300+RTX board power decode C1/C16/C32 665/990/1006 W; config=sidecar",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "DeepSeek-V4.1-Flash; GB300; MXFP4; decode C1, tok/s per user (mean of 3) | 268.7 | 528.7",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #60932",
   "source_url": "https://github.com/vllm-project/vllm/pull/60932",
   "tg_tokens": null,
   "tps": 528.7,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": null,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-60932-gb300-dsv41flash-c16-uva",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #60932 peer-device expert tier RFC measurement, DGX Station GB300 with RTX PRO 6000 Blackwell sidecar GPU, DeepSeek-V4.1-Flash native MXFP4 experts, 285 hot / 99 cold experts per layer, sidecar computes cold experts over NVLink-C2C; decode at concurrency 16, cold experts in host memory (UVA offload) arm; sidecar ratio 2.83x; GB300 board power decode C1/C16/C32 396/449/481 W; unlabeled parenthetical sub-metric 646.0 (UVA) and 1,739.0 (sidecar) on the page; config=uva",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "DeepSeek-V4.1-Flash; GB300; MXFP4; decode C16, aggregate (output tok/s) | 722.7 (646.0) | 2,048.7 (1,739.0)",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #60932",
   "source_url": "https://github.com/vllm-project/vllm/pull/60932",
   "tg_tokens": null,
   "tps": 722.7,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": null,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-60932-gb300-dsv41flash-c16-sidecar",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #60932 peer-device expert tier RFC measurement, DGX Station GB300 with RTX PRO 6000 Blackwell sidecar GPU, DeepSeek-V4.1-Flash native MXFP4 experts, 285 hot / 99 cold experts per layer, sidecar computes cold experts over NVLink-C2C; decode at concurrency 16, cold experts on the RTX PRO 6000 sidecar arm; ratio vs UVA 2.83x; combined GB300+RTX board power decode C1/C16/C32 665/990/1006 W; unlabeled parenthetical sub-metric 646.0 (UVA) and 1,739.0 (sidecar) on the page; config=sidecar",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "DeepSeek-V4.1-Flash; GB300; MXFP4; decode C16, aggregate (output tok/s) | 722.7 (646.0) | 2,048.7 (1,739.0)",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #60932",
   "source_url": "https://github.com/vllm-project/vllm/pull/60932",
   "tg_tokens": null,
   "tps": 2048.7,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 32,
   "ctx": null,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-60932-gb300-dsv41flash-c32-uva",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #60932 peer-device expert tier RFC measurement, DGX Station GB300 with RTX PRO 6000 Blackwell sidecar GPU, DeepSeek-V4.1-Flash native MXFP4 experts, 285 hot / 99 cold experts per layer, sidecar computes cold experts over NVLink-C2C; decode at concurrency 32, cold experts in host memory (UVA offload) arm; sidecar ratio 2.82x; GB300 board power decode C1/C16/C32 396/449/481 W; unlabeled parenthetical sub-metric 849.0 (UVA) and 2,311.4 (sidecar) on the page; config=uva",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "DeepSeek-V4.1-Flash; GB300; MXFP4; decode C32, aggregate (output tok/s) | 927.5 (849.0) | 2,616.5 (2,311.4)",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #60932",
   "source_url": "https://github.com/vllm-project/vllm/pull/60932",
   "tg_tokens": null,
   "tps": 927.5,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 32,
   "ctx": null,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-60932-gb300-dsv41flash-c32-sidecar",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #60932 peer-device expert tier RFC measurement, DGX Station GB300 with RTX PRO 6000 Blackwell sidecar GPU, DeepSeek-V4.1-Flash native MXFP4 experts, 285 hot / 99 cold experts per layer, sidecar computes cold experts over NVLink-C2C; decode at concurrency 32, cold experts on the RTX PRO 6000 sidecar arm; ratio vs UVA 2.82x; combined GB300+RTX board power decode C1/C16/C32 665/990/1006 W; unlabeled parenthetical sub-metric 849.0 (UVA) and 2,311.4 (sidecar) on the page; config=sidecar",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "DeepSeek-V4.1-Flash; GB300; MXFP4; decode C32, aggregate (output tok/s) | 927.5 (849.0) | 2,616.5 (2,311.4)",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #60932",
   "source_url": "https://github.com/vllm-project/vllm/pull/60932",
   "tg_tokens": null,
   "tps": 2616.5,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "vllm-61091-r9700-gemma412b-p1-s1-rowwise",
   "model": "Gemma 4 12B",
   "notes": "vLLM PR #61091 native Triton per-token FP8 scaled-MM linear backend (--linear-backend triton) tuned for Radeon AI PRO R9700 / gfx1201, TP1 serving of RedHatAI/gemma-4-12B-it-FP8-Dynamic, GuideLLM 0.7.3 workload 8 turns 2048 prompt tokens per turn 512 output tokens, streams 1/2/4/8 with 8/16/32/64 requests, two fresh-server pairs in reversed order, empty compilation caches; mean per-request output rate (output tokens / end-to-end latency); pair 1 at 1 stream(s), Torch RowWise FP8 linear arm; TTFT 915 \u2192 1,153 ms, E2E 18.24 \u2192 15.74 s; config=rowwise-p1",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "FP8",
   "quote": "Gemma 4 12B; Radeon AI PRO R9700; FP8; 1 | 1 | 28.09 \u2192 32.62",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61091",
   "source_url": "https://github.com/vllm-project/vllm/pull/61091",
   "tg_tokens": null,
   "tps": 28.09,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "vllm-61091-r9700-gemma412b-p1-s1-triton",
   "model": "Gemma 4 12B",
   "notes": "vLLM PR #61091 native Triton per-token FP8 scaled-MM linear backend (--linear-backend triton) tuned for Radeon AI PRO R9700 / gfx1201, TP1 serving of RedHatAI/gemma-4-12B-it-FP8-Dynamic, GuideLLM 0.7.3 workload 8 turns 2048 prompt tokens per turn 512 output tokens, streams 1/2/4/8 with 8/16/32/64 requests, two fresh-server pairs in reversed order, empty compilation caches; mean per-request output rate (output tokens / end-to-end latency); pair 1 at 1 stream(s), native Triton FP8 linear arm; TTFT 915 \u2192 1,153 ms, E2E 18.24 \u2192 15.74 s; config=triton-p1",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "FP8",
   "quote": "Gemma 4 12B; Radeon AI PRO R9700; FP8; 1 | 1 | 28.09 \u2192 32.62",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61091",
   "source_url": "https://github.com/vllm-project/vllm/pull/61091",
   "tg_tokens": null,
   "tps": 32.62,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 2,
   "ctx": null,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "vllm-61091-r9700-gemma412b-p1-s2-rowwise",
   "model": "Gemma 4 12B",
   "notes": "vLLM PR #61091 native Triton per-token FP8 scaled-MM linear backend (--linear-backend triton) tuned for Radeon AI PRO R9700 / gfx1201, TP1 serving of RedHatAI/gemma-4-12B-it-FP8-Dynamic, GuideLLM 0.7.3 workload 8 turns 2048 prompt tokens per turn 512 output tokens, streams 1/2/4/8 with 8/16/32/64 requests, two fresh-server pairs in reversed order, empty compilation caches; mean per-request output rate (output tokens / end-to-end latency); pair 1 at 2 stream(s), Torch RowWise FP8 linear arm; TTFT 1,149 \u2192 1,139 ms, E2E 19.10 \u2192 16.32 s; config=rowwise-p1",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "FP8",
   "quote": "Gemma 4 12B; Radeon AI PRO R9700; FP8; 1 | 2 | 26.82 \u2192 31.41",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61091",
   "source_url": "https://github.com/vllm-project/vllm/pull/61091",
   "tg_tokens": null,
   "tps": 26.82,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 2,
   "ctx": null,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "vllm-61091-r9700-gemma412b-p1-s2-triton",
   "model": "Gemma 4 12B",
   "notes": "vLLM PR #61091 native Triton per-token FP8 scaled-MM linear backend (--linear-backend triton) tuned for Radeon AI PRO R9700 / gfx1201, TP1 serving of RedHatAI/gemma-4-12B-it-FP8-Dynamic, GuideLLM 0.7.3 workload 8 turns 2048 prompt tokens per turn 512 output tokens, streams 1/2/4/8 with 8/16/32/64 requests, two fresh-server pairs in reversed order, empty compilation caches; mean per-request output rate (output tokens / end-to-end latency); pair 1 at 2 stream(s), native Triton FP8 linear arm; TTFT 1,149 \u2192 1,139 ms, E2E 19.10 \u2192 16.32 s; config=triton-p1",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "FP8",
   "quote": "Gemma 4 12B; Radeon AI PRO R9700; FP8; 1 | 2 | 26.82 \u2192 31.41",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61091",
   "source_url": "https://github.com/vllm-project/vllm/pull/61091",
   "tg_tokens": null,
   "tps": 31.41,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": null,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "vllm-61091-r9700-gemma412b-p1-s4-rowwise",
   "model": "Gemma 4 12B",
   "notes": "vLLM PR #61091 native Triton per-token FP8 scaled-MM linear backend (--linear-backend triton) tuned for Radeon AI PRO R9700 / gfx1201, TP1 serving of RedHatAI/gemma-4-12B-it-FP8-Dynamic, GuideLLM 0.7.3 workload 8 turns 2048 prompt tokens per turn 512 output tokens, streams 1/2/4/8 with 8/16/32/64 requests, two fresh-server pairs in reversed order, empty compilation caches; mean per-request output rate (output tokens / end-to-end latency); pair 1 at 4 stream(s), Torch RowWise FP8 linear arm; TTFT 1,850 \u2192 1,800 ms, E2E 20.70 \u2192 17.84 s; config=rowwise-p1",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "FP8",
   "quote": "Gemma 4 12B; Radeon AI PRO R9700; FP8; 1 | 4 | 24.79 \u2192 28.80",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61091",
   "source_url": "https://github.com/vllm-project/vllm/pull/61091",
   "tg_tokens": null,
   "tps": 24.79,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": null,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "vllm-61091-r9700-gemma412b-p1-s4-triton",
   "model": "Gemma 4 12B",
   "notes": "vLLM PR #61091 native Triton per-token FP8 scaled-MM linear backend (--linear-backend triton) tuned for Radeon AI PRO R9700 / gfx1201, TP1 serving of RedHatAI/gemma-4-12B-it-FP8-Dynamic, GuideLLM 0.7.3 workload 8 turns 2048 prompt tokens per turn 512 output tokens, streams 1/2/4/8 with 8/16/32/64 requests, two fresh-server pairs in reversed order, empty compilation caches; mean per-request output rate (output tokens / end-to-end latency); pair 1 at 4 stream(s), native Triton FP8 linear arm; TTFT 1,850 \u2192 1,800 ms, E2E 20.70 \u2192 17.84 s; config=triton-p1",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "FP8",
   "quote": "Gemma 4 12B; Radeon AI PRO R9700; FP8; 1 | 4 | 24.79 \u2192 28.80",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61091",
   "source_url": "https://github.com/vllm-project/vllm/pull/61091",
   "tg_tokens": null,
   "tps": 28.8,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": null,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "vllm-61091-r9700-gemma412b-p1-s8-rowwise",
   "model": "Gemma 4 12B",
   "notes": "vLLM PR #61091 native Triton per-token FP8 scaled-MM linear backend (--linear-backend triton) tuned for Radeon AI PRO R9700 / gfx1201, TP1 serving of RedHatAI/gemma-4-12B-it-FP8-Dynamic, GuideLLM 0.7.3 workload 8 turns 2048 prompt tokens per turn 512 output tokens, streams 1/2/4/8 with 8/16/32/64 requests, two fresh-server pairs in reversed order, empty compilation caches; mean per-request output rate (output tokens / end-to-end latency); pair 1 at 8 stream(s), Torch RowWise FP8 linear arm; TTFT 2,911 \u2192 2,823 ms, E2E 24.65 \u2192 21.56 s; config=rowwise-p1",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "FP8",
   "quote": "Gemma 4 12B; Radeon AI PRO R9700; FP8; 1 | 8 | 20.88 \u2192 23.91",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61091",
   "source_url": "https://github.com/vllm-project/vllm/pull/61091",
   "tg_tokens": null,
   "tps": 20.88,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": null,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "vllm-61091-r9700-gemma412b-p1-s8-triton",
   "model": "Gemma 4 12B",
   "notes": "vLLM PR #61091 native Triton per-token FP8 scaled-MM linear backend (--linear-backend triton) tuned for Radeon AI PRO R9700 / gfx1201, TP1 serving of RedHatAI/gemma-4-12B-it-FP8-Dynamic, GuideLLM 0.7.3 workload 8 turns 2048 prompt tokens per turn 512 output tokens, streams 1/2/4/8 with 8/16/32/64 requests, two fresh-server pairs in reversed order, empty compilation caches; mean per-request output rate (output tokens / end-to-end latency); pair 1 at 8 stream(s), native Triton FP8 linear arm; TTFT 2,911 \u2192 2,823 ms, E2E 24.65 \u2192 21.56 s; config=triton-p1",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "FP8",
   "quote": "Gemma 4 12B; Radeon AI PRO R9700; FP8; 1 | 8 | 20.88 \u2192 23.91",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61091",
   "source_url": "https://github.com/vllm-project/vllm/pull/61091",
   "tg_tokens": null,
   "tps": 23.91,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "vllm-61091-r9700-gemma412b-p2-s1-rowwise",
   "model": "Gemma 4 12B",
   "notes": "vLLM PR #61091 native Triton per-token FP8 scaled-MM linear backend (--linear-backend triton) tuned for Radeon AI PRO R9700 / gfx1201, TP1 serving of RedHatAI/gemma-4-12B-it-FP8-Dynamic, GuideLLM 0.7.3 workload 8 turns 2048 prompt tokens per turn 512 output tokens, streams 1/2/4/8 with 8/16/32/64 requests, two fresh-server pairs in reversed order, empty compilation caches; mean per-request output rate (output tokens / end-to-end latency); pair 2 at 1 stream(s), Torch RowWise FP8 linear arm; TTFT 925 \u2192 1,135 ms, E2E 18.24 \u2192 15.71 s; config=rowwise-p2",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "FP8",
   "quote": "Gemma 4 12B; Radeon AI PRO R9700; FP8; 2 | 1 | 28.09 \u2192 32.70",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61091",
   "source_url": "https://github.com/vllm-project/vllm/pull/61091",
   "tg_tokens": null,
   "tps": 28.09,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "vllm-61091-r9700-gemma412b-p2-s1-triton",
   "model": "Gemma 4 12B",
   "notes": "vLLM PR #61091 native Triton per-token FP8 scaled-MM linear backend (--linear-backend triton) tuned for Radeon AI PRO R9700 / gfx1201, TP1 serving of RedHatAI/gemma-4-12B-it-FP8-Dynamic, GuideLLM 0.7.3 workload 8 turns 2048 prompt tokens per turn 512 output tokens, streams 1/2/4/8 with 8/16/32/64 requests, two fresh-server pairs in reversed order, empty compilation caches; mean per-request output rate (output tokens / end-to-end latency); pair 2 at 1 stream(s), native Triton FP8 linear arm; TTFT 925 \u2192 1,135 ms, E2E 18.24 \u2192 15.71 s; config=triton-p2",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "FP8",
   "quote": "Gemma 4 12B; Radeon AI PRO R9700; FP8; 2 | 1 | 28.09 \u2192 32.70",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61091",
   "source_url": "https://github.com/vllm-project/vllm/pull/61091",
   "tg_tokens": null,
   "tps": 32.7,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 2,
   "ctx": null,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "vllm-61091-r9700-gemma412b-p2-s2-rowwise",
   "model": "Gemma 4 12B",
   "notes": "vLLM PR #61091 native Triton per-token FP8 scaled-MM linear backend (--linear-backend triton) tuned for Radeon AI PRO R9700 / gfx1201, TP1 serving of RedHatAI/gemma-4-12B-it-FP8-Dynamic, GuideLLM 0.7.3 workload 8 turns 2048 prompt tokens per turn 512 output tokens, streams 1/2/4/8 with 8/16/32/64 requests, two fresh-server pairs in reversed order, empty compilation caches; mean per-request output rate (output tokens / end-to-end latency); pair 2 at 2 stream(s), Torch RowWise FP8 linear arm; TTFT 1,171 \u2192 1,123 ms, E2E 19.23 \u2192 16.33 s; config=rowwise-p2",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "FP8",
   "quote": "Gemma 4 12B; Radeon AI PRO R9700; FP8; 2 | 2 | 26.64 \u2192 31.38",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61091",
   "source_url": "https://github.com/vllm-project/vllm/pull/61091",
   "tg_tokens": null,
   "tps": 26.64,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 2,
   "ctx": null,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "vllm-61091-r9700-gemma412b-p2-s2-triton",
   "model": "Gemma 4 12B",
   "notes": "vLLM PR #61091 native Triton per-token FP8 scaled-MM linear backend (--linear-backend triton) tuned for Radeon AI PRO R9700 / gfx1201, TP1 serving of RedHatAI/gemma-4-12B-it-FP8-Dynamic, GuideLLM 0.7.3 workload 8 turns 2048 prompt tokens per turn 512 output tokens, streams 1/2/4/8 with 8/16/32/64 requests, two fresh-server pairs in reversed order, empty compilation caches; mean per-request output rate (output tokens / end-to-end latency); pair 2 at 2 stream(s), native Triton FP8 linear arm; TTFT 1,171 \u2192 1,123 ms, E2E 19.23 \u2192 16.33 s; config=triton-p2",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "FP8",
   "quote": "Gemma 4 12B; Radeon AI PRO R9700; FP8; 2 | 2 | 26.64 \u2192 31.38",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61091",
   "source_url": "https://github.com/vllm-project/vllm/pull/61091",
   "tg_tokens": null,
   "tps": 31.38,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": null,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "vllm-61091-r9700-gemma412b-p2-s4-rowwise",
   "model": "Gemma 4 12B",
   "notes": "vLLM PR #61091 native Triton per-token FP8 scaled-MM linear backend (--linear-backend triton) tuned for Radeon AI PRO R9700 / gfx1201, TP1 serving of RedHatAI/gemma-4-12B-it-FP8-Dynamic, GuideLLM 0.7.3 workload 8 turns 2048 prompt tokens per turn 512 output tokens, streams 1/2/4/8 with 8/16/32/64 requests, two fresh-server pairs in reversed order, empty compilation caches; mean per-request output rate (output tokens / end-to-end latency); pair 2 at 4 stream(s), Torch RowWise FP8 linear arm; TTFT 1,776 \u2192 1,673 ms, E2E 20.88 \u2192 17.97 s; config=rowwise-p2",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "FP8",
   "quote": "Gemma 4 12B; Radeon AI PRO R9700; FP8; 2 | 4 | 24.57 \u2192 28.59",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61091",
   "source_url": "https://github.com/vllm-project/vllm/pull/61091",
   "tg_tokens": null,
   "tps": 24.57,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": null,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "vllm-61091-r9700-gemma412b-p2-s4-triton",
   "model": "Gemma 4 12B",
   "notes": "vLLM PR #61091 native Triton per-token FP8 scaled-MM linear backend (--linear-backend triton) tuned for Radeon AI PRO R9700 / gfx1201, TP1 serving of RedHatAI/gemma-4-12B-it-FP8-Dynamic, GuideLLM 0.7.3 workload 8 turns 2048 prompt tokens per turn 512 output tokens, streams 1/2/4/8 with 8/16/32/64 requests, two fresh-server pairs in reversed order, empty compilation caches; mean per-request output rate (output tokens / end-to-end latency); pair 2 at 4 stream(s), native Triton FP8 linear arm; TTFT 1,776 \u2192 1,673 ms, E2E 20.88 \u2192 17.97 s; config=triton-p2",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "FP8",
   "quote": "Gemma 4 12B; Radeon AI PRO R9700; FP8; 2 | 4 | 24.57 \u2192 28.59",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61091",
   "source_url": "https://github.com/vllm-project/vllm/pull/61091",
   "tg_tokens": null,
   "tps": 28.59,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": null,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "vllm-61091-r9700-gemma412b-p2-s8-rowwise",
   "model": "Gemma 4 12B",
   "notes": "vLLM PR #61091 native Triton per-token FP8 scaled-MM linear backend (--linear-backend triton) tuned for Radeon AI PRO R9700 / gfx1201, TP1 serving of RedHatAI/gemma-4-12B-it-FP8-Dynamic, GuideLLM 0.7.3 workload 8 turns 2048 prompt tokens per turn 512 output tokens, streams 1/2/4/8 with 8/16/32/64 requests, two fresh-server pairs in reversed order, empty compilation caches; mean per-request output rate (output tokens / end-to-end latency); pair 2 at 8 stream(s), Torch RowWise FP8 linear arm; TTFT 3,011 \u2192 2,872 ms, E2E 24.93 \u2192 21.88 s; config=rowwise-p2",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "FP8",
   "quote": "Gemma 4 12B; Radeon AI PRO R9700; FP8; 2 | 8 | 20.65 \u2192 23.60",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61091",
   "source_url": "https://github.com/vllm-project/vllm/pull/61091",
   "tg_tokens": null,
   "tps": 20.65,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": null,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "vllm-61091-r9700-gemma412b-p2-s8-triton",
   "model": "Gemma 4 12B",
   "notes": "vLLM PR #61091 native Triton per-token FP8 scaled-MM linear backend (--linear-backend triton) tuned for Radeon AI PRO R9700 / gfx1201, TP1 serving of RedHatAI/gemma-4-12B-it-FP8-Dynamic, GuideLLM 0.7.3 workload 8 turns 2048 prompt tokens per turn 512 output tokens, streams 1/2/4/8 with 8/16/32/64 requests, two fresh-server pairs in reversed order, empty compilation caches; mean per-request output rate (output tokens / end-to-end latency); pair 2 at 8 stream(s), native Triton FP8 linear arm; TTFT 3,011 \u2192 2,872 ms, E2E 24.93 \u2192 21.88 s; config=triton-p2",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "FP8",
   "quote": "Gemma 4 12B; Radeon AI PRO R9700; FP8; 2 | 8 | 20.65 \u2192 23.60",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61091",
   "source_url": "https://github.com/vllm-project/vllm/pull/61091",
   "tg_tokens": null,
   "tps": 23.6,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-llama318b-128-128-rdna",
   "model": "Llama 3.1 8B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Llama-3.1-8B-AWQ-INT4 EAGLE3 at in/out 128/128 (OSL 128), ours, ROCM_RDNA_SINGLESEQ_ATTN; TTFT 221 ms; EAGLE3 speculative decoding enabled; config=rdna",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-INT4",
   "quote": "Llama 3.1 8B; gfx1151; AWQ-INT4; Llama-3.1-8B-AWQ-INT4 EAGLE3 | 128/128 | 221 / 221 / 218 | 8.6 / 8.4 / 8.6",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 8.6,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-llama318b-128-128-rocm-attn",
   "model": "Llama 3.1 8B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Llama-3.1-8B-AWQ-INT4 EAGLE3 at in/out 128/128 (OSL 128), ROCM_ATTN baseline; TTFT 221 ms; EAGLE3 speculative decoding enabled; config=rocm-attn",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-INT4",
   "quote": "Llama 3.1 8B; gfx1151; AWQ-INT4; Llama-3.1-8B-AWQ-INT4 EAGLE3 | 128/128 | 221 / 221 / 218 | 8.6 / 8.4 / 8.6",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 8.4,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-llama318b-128-128-segmented",
   "model": "Llama 3.1 8B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Llama-3.1-8B-AWQ-INT4 EAGLE3 at in/out 128/128 (OSL 128), ROCM_SEGMENTED_ATTN baseline; TTFT 218 ms; EAGLE3 speculative decoding enabled; config=segmented",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-INT4",
   "quote": "Llama 3.1 8B; gfx1151; AWQ-INT4; Llama-3.1-8B-AWQ-INT4 EAGLE3 | 128/128 | 221 / 221 / 218 | 8.6 / 8.4 / 8.6",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 8.6,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-llama318b-128-4k-rdna",
   "model": "Llama 3.1 8B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Llama-3.1-8B-AWQ-INT4 EAGLE3 at in/out 128/4k (OSL 4096), ours, ROCM_RDNA_SINGLESEQ_ATTN; TTFT 242 ms; EAGLE3 speculative decoding enabled; config=rdna",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-INT4",
   "quote": "Llama 3.1 8B; gfx1151; AWQ-INT4; Llama-3.1-8B-AWQ-INT4 EAGLE3 | 128/4k | 242 / 255 / 265 | 9.2 / 8.4 / 9.1",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 9.2,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-llama318b-128-4k-rocm-attn",
   "model": "Llama 3.1 8B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Llama-3.1-8B-AWQ-INT4 EAGLE3 at in/out 128/4k (OSL 4096), ROCM_ATTN baseline; TTFT 255 ms; EAGLE3 speculative decoding enabled; config=rocm-attn",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-INT4",
   "quote": "Llama 3.1 8B; gfx1151; AWQ-INT4; Llama-3.1-8B-AWQ-INT4 EAGLE3 | 128/4k | 242 / 255 / 265 | 9.2 / 8.4 / 9.1",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 8.4,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-llama318b-128-4k-segmented",
   "model": "Llama 3.1 8B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Llama-3.1-8B-AWQ-INT4 EAGLE3 at in/out 128/4k (OSL 4096), ROCM_SEGMENTED_ATTN baseline; TTFT 265 ms; EAGLE3 speculative decoding enabled; config=segmented",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-INT4",
   "quote": "Llama 3.1 8B; gfx1151; AWQ-INT4; Llama-3.1-8B-AWQ-INT4 EAGLE3 | 128/4k | 242 / 255 / 265 | 9.2 / 8.4 / 9.1",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 9.1,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-llama318b-128-16k-rdna",
   "model": "Llama 3.1 8B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Llama-3.1-8B-AWQ-INT4 EAGLE3 at in/out 128/16k (OSL 16384), ours, ROCM_RDNA_SINGLESEQ_ATTN; TTFT 216 ms; EAGLE3 speculative decoding enabled; config=rdna",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-INT4",
   "quote": "Llama 3.1 8B; gfx1151; AWQ-INT4; Llama-3.1-8B-AWQ-INT4 EAGLE3 | 128/16k | 216 / \u2013 / 214 | 7.7 / \u2013 / 7.7",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 7.7,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-llama318b-128-16k-segmented",
   "model": "Llama 3.1 8B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Llama-3.1-8B-AWQ-INT4 EAGLE3 at in/out 128/16k (OSL 16384), ROCM_SEGMENTED_ATTN baseline; TTFT 214 ms; EAGLE3 speculative decoding enabled; config=segmented",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-INT4",
   "quote": "Llama 3.1 8B; gfx1151; AWQ-INT4; Llama-3.1-8B-AWQ-INT4 EAGLE3 | 128/16k | 216 / \u2013 / 214 | 7.7 / \u2013 / 7.7",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 7.7,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-llama318b-8k-128-rdna",
   "model": "Llama 3.1 8B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Llama-3.1-8B-AWQ-INT4 EAGLE3 at in/out 8k/128 (OSL 128), ours, ROCM_RDNA_SINGLESEQ_ATTN; TTFT 4226 ms; EAGLE3 speculative decoding enabled; config=rdna",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-INT4",
   "quote": "Llama 3.1 8B; gfx1151; AWQ-INT4; Llama-3.1-8B-AWQ-INT4 EAGLE3 | 8k/128 | 4226 / 5142 / 4370 | 6.4 / 5.1 / 6.4",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 6.4,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-llama318b-8k-128-rocm-attn",
   "model": "Llama 3.1 8B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Llama-3.1-8B-AWQ-INT4 EAGLE3 at in/out 8k/128 (OSL 128), ROCM_ATTN baseline; TTFT 5142 ms; EAGLE3 speculative decoding enabled; config=rocm-attn",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-INT4",
   "quote": "Llama 3.1 8B; gfx1151; AWQ-INT4; Llama-3.1-8B-AWQ-INT4 EAGLE3 | 8k/128 | 4226 / 5142 / 4370 | 6.4 / 5.1 / 6.4",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 5.1,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-llama318b-8k-128-segmented",
   "model": "Llama 3.1 8B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Llama-3.1-8B-AWQ-INT4 EAGLE3 at in/out 8k/128 (OSL 128), ROCM_SEGMENTED_ATTN baseline; TTFT 4370 ms; EAGLE3 speculative decoding enabled; config=segmented",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-INT4",
   "quote": "Llama 3.1 8B; gfx1151; AWQ-INT4; Llama-3.1-8B-AWQ-INT4 EAGLE3 | 8k/128 | 4226 / 5142 / 4370 | 6.4 / 5.1 / 6.4",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 6.4,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-llama318b-8k-4k-rdna",
   "model": "Llama 3.1 8B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Llama-3.1-8B-AWQ-INT4 EAGLE3 at in/out 8k/4k (OSL 4096), ours, ROCM_RDNA_SINGLESEQ_ATTN; TTFT 4599 ms; EAGLE3 speculative decoding enabled; config=rdna",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-INT4",
   "quote": "Llama 3.1 8B; gfx1151; AWQ-INT4; Llama-3.1-8B-AWQ-INT4 EAGLE3 | 8k/4k | 4599 / 5686 / 4774 | 8.2 / 6.0 / 8.1",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 8.2,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-llama318b-8k-4k-rocm-attn",
   "model": "Llama 3.1 8B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Llama-3.1-8B-AWQ-INT4 EAGLE3 at in/out 8k/4k (OSL 4096), ROCM_ATTN baseline; TTFT 5686 ms; EAGLE3 speculative decoding enabled; config=rocm-attn",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-INT4",
   "quote": "Llama 3.1 8B; gfx1151; AWQ-INT4; Llama-3.1-8B-AWQ-INT4 EAGLE3 | 8k/4k | 4599 / 5686 / 4774 | 8.2 / 6.0 / 8.1",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 6.0,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-llama318b-8k-4k-segmented",
   "model": "Llama 3.1 8B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Llama-3.1-8B-AWQ-INT4 EAGLE3 at in/out 8k/4k (OSL 4096), ROCM_SEGMENTED_ATTN baseline; TTFT 4774 ms; EAGLE3 speculative decoding enabled; config=segmented",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-INT4",
   "quote": "Llama 3.1 8B; gfx1151; AWQ-INT4; Llama-3.1-8B-AWQ-INT4 EAGLE3 | 8k/4k | 4599 / 5686 / 4774 | 8.2 / 6.0 / 8.1",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 8.1,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-llama318b-8k-16k-rdna",
   "model": "Llama 3.1 8B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Llama-3.1-8B-AWQ-INT4 EAGLE3 at in/out 8k/16k (OSL 16384), ours, ROCM_RDNA_SINGLESEQ_ATTN; TTFT 4302 ms; EAGLE3 speculative decoding enabled; config=rdna",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-INT4",
   "quote": "Llama 3.1 8B; gfx1151; AWQ-INT4; Llama-3.1-8B-AWQ-INT4 EAGLE3 | 8k/16k | 4302 / 5182 / 4412 | 7.2 / 4.6 / 7.1",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 7.2,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-llama318b-8k-16k-rocm-attn",
   "model": "Llama 3.1 8B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Llama-3.1-8B-AWQ-INT4 EAGLE3 at in/out 8k/16k (OSL 16384), ROCM_ATTN baseline; TTFT 5182 ms; EAGLE3 speculative decoding enabled; config=rocm-attn",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-INT4",
   "quote": "Llama 3.1 8B; gfx1151; AWQ-INT4; Llama-3.1-8B-AWQ-INT4 EAGLE3 | 8k/16k | 4302 / 5182 / 4412 | 7.2 / 4.6 / 7.1",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 4.6,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-llama318b-8k-16k-segmented",
   "model": "Llama 3.1 8B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Llama-3.1-8B-AWQ-INT4 EAGLE3 at in/out 8k/16k (OSL 16384), ROCM_SEGMENTED_ATTN baseline; TTFT 4412 ms; EAGLE3 speculative decoding enabled; config=segmented",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-INT4",
   "quote": "Llama 3.1 8B; gfx1151; AWQ-INT4; Llama-3.1-8B-AWQ-INT4 EAGLE3 | 8k/16k | 4302 / 5182 / 4412 | 7.2 / 4.6 / 7.1",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 7.1,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-qwen359b-128-128-rdna",
   "model": "Qwen3.5-9B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Qwen3.5-9B at in/out 128/128 (OSL 128), ours, ROCM_RDNA_SINGLESEQ_ATTN; TTFT 127 ms; config=rdna",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "Qwen3.5-9B; gfx1151; Qwen3.5-9B | 128/128 | 127 / 129 / 128 | 13.6 / 13.3 / 13.5",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-qwen359b-128-128-rocm-attn",
   "model": "Qwen3.5-9B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Qwen3.5-9B at in/out 128/128 (OSL 128), ROCM_ATTN baseline; TTFT 129 ms; config=rocm-attn",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "Qwen3.5-9B; gfx1151; Qwen3.5-9B | 128/128 | 127 / 129 / 128 | 13.6 / 13.3 / 13.5",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 13.3,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-qwen359b-128-128-segmented",
   "model": "Qwen3.5-9B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Qwen3.5-9B at in/out 128/128 (OSL 128), ROCM_SEGMENTED_ATTN baseline; TTFT 128 ms; config=segmented",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "Qwen3.5-9B; gfx1151; Qwen3.5-9B | 128/128 | 127 / 129 / 128 | 13.6 / 13.3 / 13.5",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 13.5,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-qwen359b-128-4k-rdna",
   "model": "Qwen3.5-9B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Qwen3.5-9B at in/out 128/4k (OSL 4096), ours, ROCM_RDNA_SINGLESEQ_ATTN; TTFT 126 ms; config=rdna",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "Qwen3.5-9B; gfx1151; Qwen3.5-9B | 128/4k | 126 / 127 / 126 | 13.6 / 11.9 / 13.6",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-qwen359b-128-4k-rocm-attn",
   "model": "Qwen3.5-9B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Qwen3.5-9B at in/out 128/4k (OSL 4096), ROCM_ATTN baseline; TTFT 127 ms; config=rocm-attn",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "Qwen3.5-9B; gfx1151; Qwen3.5-9B | 128/4k | 126 / 127 / 126 | 13.6 / 11.9 / 13.6",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 11.9,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-qwen359b-128-4k-segmented",
   "model": "Qwen3.5-9B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Qwen3.5-9B at in/out 128/4k (OSL 4096), ROCM_SEGMENTED_ATTN baseline; TTFT 126 ms; config=segmented",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "Qwen3.5-9B; gfx1151; Qwen3.5-9B | 128/4k | 126 / 127 / 126 | 13.6 / 11.9 / 13.6",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-qwen359b-128-16k-rdna",
   "model": "Qwen3.5-9B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Qwen3.5-9B at in/out 128/16k (OSL 16384), ours, ROCM_RDNA_SINGLESEQ_ATTN; TTFT 134 ms; config=rdna",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "Qwen3.5-9B; gfx1151; Qwen3.5-9B | 128/16k | 134 / 135 / 132 | 13.2 / 8.6 / 13.2",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 13.2,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-qwen359b-128-16k-rocm-attn",
   "model": "Qwen3.5-9B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Qwen3.5-9B at in/out 128/16k (OSL 16384), ROCM_ATTN baseline; TTFT 135 ms; config=rocm-attn",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "Qwen3.5-9B; gfx1151; Qwen3.5-9B | 128/16k | 134 / 135 / 132 | 13.2 / 8.6 / 13.2",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 8.6,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-qwen359b-128-16k-segmented",
   "model": "Qwen3.5-9B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Qwen3.5-9B at in/out 128/16k (OSL 16384), ROCM_SEGMENTED_ATTN baseline; TTFT 132 ms; config=segmented",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "Qwen3.5-9B; gfx1151; Qwen3.5-9B | 128/16k | 134 / 135 / 132 | 13.2 / 8.6 / 13.2",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 13.2,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-qwen359b-8k-128-rdna",
   "model": "Qwen3.5-9B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Qwen3.5-9B at in/out 8k/128 (OSL 128), ours, ROCM_RDNA_SINGLESEQ_ATTN; TTFT 4595 ms; config=rdna",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "Qwen3.5-9B; gfx1151; Qwen3.5-9B | 8k/128 | 4595 / 5462 / 4698 | 9.1 / 6.4 / 9.0",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 9.1,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-qwen359b-8k-128-rocm-attn",
   "model": "Qwen3.5-9B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Qwen3.5-9B at in/out 8k/128 (OSL 128), ROCM_ATTN baseline; TTFT 5462 ms; config=rocm-attn",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "Qwen3.5-9B; gfx1151; Qwen3.5-9B | 8k/128 | 4595 / 5462 / 4698 | 9.1 / 6.4 / 9.0",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 6.4,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-qwen359b-8k-128-segmented",
   "model": "Qwen3.5-9B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Qwen3.5-9B at in/out 8k/128 (OSL 128), ROCM_SEGMENTED_ATTN baseline; TTFT 4698 ms; config=segmented",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "Qwen3.5-9B; gfx1151; Qwen3.5-9B | 8k/128 | 4595 / 5462 / 4698 | 9.1 / 6.4 / 9.0",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 9.0,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-qwen359b-8k-4k-rdna",
   "model": "Qwen3.5-9B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Qwen3.5-9B at in/out 8k/4k (OSL 4096), ours, ROCM_RDNA_SINGLESEQ_ATTN; TTFT 4800 ms; config=rdna",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "Qwen3.5-9B; gfx1151; Qwen3.5-9B | 8k/4k | 4800 / 5646 / 4910 | 13.2 / 8.1 / 13.1",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 13.2,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-qwen359b-8k-4k-rocm-attn",
   "model": "Qwen3.5-9B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Qwen3.5-9B at in/out 8k/4k (OSL 4096), ROCM_ATTN baseline; TTFT 5646 ms; config=rocm-attn",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "Qwen3.5-9B; gfx1151; Qwen3.5-9B | 8k/4k | 4800 / 5646 / 4910 | 13.2 / 8.1 / 13.1",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 8.1,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-qwen359b-8k-4k-segmented",
   "model": "Qwen3.5-9B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Qwen3.5-9B at in/out 8k/4k (OSL 4096), ROCM_SEGMENTED_ATTN baseline; TTFT 4910 ms; config=segmented",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "Qwen3.5-9B; gfx1151; Qwen3.5-9B | 8k/4k | 4800 / 5646 / 4910 | 13.2 / 8.1 / 13.1",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 13.1,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-qwen359b-8k-16k-rdna",
   "model": "Qwen3.5-9B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Qwen3.5-9B at in/out 8k/16k (OSL 16384), ours, ROCM_RDNA_SINGLESEQ_ATTN; TTFT 4825 ms; config=rdna",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "Qwen3.5-9B; gfx1151; Qwen3.5-9B | 8k/16k | 4825 / 5679 / 4941 | 13.0 / 6.5 / 13.0",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-qwen359b-8k-16k-rocm-attn",
   "model": "Qwen3.5-9B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Qwen3.5-9B at in/out 8k/16k (OSL 16384), ROCM_ATTN baseline; TTFT 5679 ms; config=rocm-attn",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "Qwen3.5-9B; gfx1151; Qwen3.5-9B | 8k/16k | 4825 / 5679 / 4941 | 13.0 / 6.5 / 13.0",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 6.5,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-qwen359b-8k-16k-segmented",
   "model": "Qwen3.5-9B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Qwen3.5-9B at in/out 8k/16k (OSL 16384), ROCM_SEGMENTED_ATTN baseline; TTFT 4941 ms; config=segmented",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "Qwen3.5-9B; gfx1151; Qwen3.5-9B | 8k/16k | 4825 / 5679 / 4941 | 13.0 / 6.5 / 13.0",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-qwen3627b-128-128-rdna",
   "model": "Qwen3.6-27B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Qwen3.6-27B-AWQ-INT4 at in/out 128/128 (OSL 128), ours, ROCM_RDNA_SINGLESEQ_ATTN; TTFT 494 ms; config=rdna",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-INT4",
   "quote": "Qwen3.6-27B; gfx1151; AWQ-INT4; Qwen3.6-27B-AWQ-INT4 | 128/128 | 494 / 497 / 488 | 10.7 / 10.4 / 10.7",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 10.7,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-qwen3627b-128-128-rocm-attn",
   "model": "Qwen3.6-27B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Qwen3.6-27B-AWQ-INT4 at in/out 128/128 (OSL 128), ROCM_ATTN baseline; TTFT 497 ms; config=rocm-attn",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-INT4",
   "quote": "Qwen3.6-27B; gfx1151; AWQ-INT4; Qwen3.6-27B-AWQ-INT4 | 128/128 | 494 / 497 / 488 | 10.7 / 10.4 / 10.7",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 10.4,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-qwen3627b-128-128-segmented",
   "model": "Qwen3.6-27B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Qwen3.6-27B-AWQ-INT4 at in/out 128/128 (OSL 128), ROCM_SEGMENTED_ATTN baseline; TTFT 488 ms; config=segmented",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-INT4",
   "quote": "Qwen3.6-27B; gfx1151; AWQ-INT4; Qwen3.6-27B-AWQ-INT4 | 128/128 | 494 / 497 / 488 | 10.7 / 10.4 / 10.7",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 10.7,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-qwen3627b-128-4k-rdna",
   "model": "Qwen3.6-27B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Qwen3.6-27B-AWQ-INT4 at in/out 128/4k (OSL 4096), ours, ROCM_RDNA_SINGLESEQ_ATTN; TTFT 434 ms; config=rdna",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-INT4",
   "quote": "Qwen3.6-27B; gfx1151; AWQ-INT4; Qwen3.6-27B-AWQ-INT4 | 128/4k | 434 / 485 / 438 | 10.8 / 8.5 / 10.7",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 10.8,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-qwen3627b-128-4k-rocm-attn",
   "model": "Qwen3.6-27B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Qwen3.6-27B-AWQ-INT4 at in/out 128/4k (OSL 4096), ROCM_ATTN baseline; TTFT 485 ms; config=rocm-attn",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-INT4",
   "quote": "Qwen3.6-27B; gfx1151; AWQ-INT4; Qwen3.6-27B-AWQ-INT4 | 128/4k | 434 / 485 / 438 | 10.8 / 8.5 / 10.7",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 8.5,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-qwen3627b-128-4k-segmented",
   "model": "Qwen3.6-27B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Qwen3.6-27B-AWQ-INT4 at in/out 128/4k (OSL 4096), ROCM_SEGMENTED_ATTN baseline; TTFT 438 ms; config=segmented",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-INT4",
   "quote": "Qwen3.6-27B; gfx1151; AWQ-INT4; Qwen3.6-27B-AWQ-INT4 | 128/4k | 434 / 485 / 438 | 10.8 / 8.5 / 10.7",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 10.7,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-qwen3627b-128-16k-rdna",
   "model": "Qwen3.6-27B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Qwen3.6-27B-AWQ-INT4 at in/out 128/16k (OSL 16384), ours, ROCM_RDNA_SINGLESEQ_ATTN; TTFT 425 ms; config=rdna",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-INT4",
   "quote": "Qwen3.6-27B; gfx1151; AWQ-INT4; Qwen3.6-27B-AWQ-INT4 | 128/16k | 425 / 435 / 417 | 11.0 / 5.5 / 10.9",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 11.0,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-qwen3627b-128-16k-rocm-attn",
   "model": "Qwen3.6-27B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Qwen3.6-27B-AWQ-INT4 at in/out 128/16k (OSL 16384), ROCM_ATTN baseline; TTFT 435 ms; config=rocm-attn",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-INT4",
   "quote": "Qwen3.6-27B; gfx1151; AWQ-INT4; Qwen3.6-27B-AWQ-INT4 | 128/16k | 425 / 435 / 417 | 11.0 / 5.5 / 10.9",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 5.5,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-qwen3627b-128-16k-segmented",
   "model": "Qwen3.6-27B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Qwen3.6-27B-AWQ-INT4 at in/out 128/16k (OSL 16384), ROCM_SEGMENTED_ATTN baseline; TTFT 417 ms; config=segmented",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-INT4",
   "quote": "Qwen3.6-27B; gfx1151; AWQ-INT4; Qwen3.6-27B-AWQ-INT4 | 128/16k | 425 / 435 / 417 | 11.0 / 5.5 / 10.9",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 10.9,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-qwen3627b-8k-128-rdna",
   "model": "Qwen3.6-27B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Qwen3.6-27B-AWQ-INT4 at in/out 8k/128 (OSL 128), ours, ROCM_RDNA_SINGLESEQ_ATTN; TTFT 20065 ms; config=rdna",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-INT4",
   "quote": "Qwen3.6-27B; gfx1151; AWQ-INT4; Qwen3.6-27B-AWQ-INT4 | 8k/128 | 20065 / 22391 / 20484 | 4.0 / 2.8 / 4.0",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 4.0,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-qwen3627b-8k-128-rocm-attn",
   "model": "Qwen3.6-27B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Qwen3.6-27B-AWQ-INT4 at in/out 8k/128 (OSL 128), ROCM_ATTN baseline; TTFT 22391 ms; config=rocm-attn",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-INT4",
   "quote": "Qwen3.6-27B; gfx1151; AWQ-INT4; Qwen3.6-27B-AWQ-INT4 | 8k/128 | 20065 / 22391 / 20484 | 4.0 / 2.8 / 4.0",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 2.8,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-qwen3627b-8k-128-segmented",
   "model": "Qwen3.6-27B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Qwen3.6-27B-AWQ-INT4 at in/out 8k/128 (OSL 128), ROCM_SEGMENTED_ATTN baseline; TTFT 20484 ms; config=segmented",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-INT4",
   "quote": "Qwen3.6-27B; gfx1151; AWQ-INT4; Qwen3.6-27B-AWQ-INT4 | 8k/128 | 20065 / 22391 / 20484 | 4.0 / 2.8 / 4.0",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 4.0,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-qwen3627b-8k-4k-rdna",
   "model": "Qwen3.6-27B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Qwen3.6-27B-AWQ-INT4 at in/out 8k/4k (OSL 4096), ours, ROCM_RDNA_SINGLESEQ_ATTN; TTFT 19995 ms; config=rdna",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-INT4",
   "quote": "Qwen3.6-27B; gfx1151; AWQ-INT4; Qwen3.6-27B-AWQ-INT4 | 8k/4k | 19995 / 22335 / 20256 | 10.4 / 4.8 / 10.3",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 10.4,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-qwen3627b-8k-4k-rocm-attn",
   "model": "Qwen3.6-27B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Qwen3.6-27B-AWQ-INT4 at in/out 8k/4k (OSL 4096), ROCM_ATTN baseline; TTFT 22335 ms; config=rocm-attn",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-INT4",
   "quote": "Qwen3.6-27B; gfx1151; AWQ-INT4; Qwen3.6-27B-AWQ-INT4 | 8k/4k | 19995 / 22335 / 20256 | 10.4 / 4.8 / 10.3",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 4.8,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-qwen3627b-8k-4k-segmented",
   "model": "Qwen3.6-27B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Qwen3.6-27B-AWQ-INT4 at in/out 8k/4k (OSL 4096), ROCM_SEGMENTED_ATTN baseline; TTFT 20256 ms; config=segmented",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-INT4",
   "quote": "Qwen3.6-27B; gfx1151; AWQ-INT4; Qwen3.6-27B-AWQ-INT4 | 8k/4k | 19995 / 22335 / 20256 | 10.4 / 4.8 / 10.3",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 10.3,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-qwen3627b-8k-16k-rdna",
   "model": "Qwen3.6-27B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Qwen3.6-27B-AWQ-INT4 at in/out 8k/16k (OSL 16384), ours, ROCM_RDNA_SINGLESEQ_ATTN; TTFT 21616 ms; config=rdna",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-INT4",
   "quote": "Qwen3.6-27B; gfx1151; AWQ-INT4; Qwen3.6-27B-AWQ-INT4 | 8k/16k | 21616 / 22627 / 22485 | 10.6 / 3.7 / 10.5",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 10.6,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-qwen3627b-8k-16k-rocm-attn",
   "model": "Qwen3.6-27B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Qwen3.6-27B-AWQ-INT4 at in/out 8k/16k (OSL 16384), ROCM_ATTN baseline; TTFT 22627 ms; config=rocm-attn",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-INT4",
   "quote": "Qwen3.6-27B; gfx1151; AWQ-INT4; Qwen3.6-27B-AWQ-INT4 | 8k/16k | 21616 / 22627 / 22485 | 10.6 / 3.7 / 10.5",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 3.7,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-qwen3627b-8k-16k-segmented",
   "model": "Qwen3.6-27B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend (two HIP kernels, WMMA, GQA-packed, split-KV, CUDA graphs) for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, single sequence per step; decode tok/s; Qwen3.6-27B-AWQ-INT4 at in/out 8k/16k (OSL 16384), ROCM_SEGMENTED_ATTN baseline; TTFT 22485 ms; config=segmented",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-INT4",
   "quote": "Qwen3.6-27B; gfx1151; AWQ-INT4; Qwen3.6-27B-AWQ-INT4 | 8k/16k | 21616 / 22627 / 22485 | 10.6 / 3.7 / 10.5",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 10.5,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-gemma426ba4b-128-128-rdna",
   "model": "Gemma 4 26B-A4B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, Gemma 4 D=512 head dim where ROCM_ATTN and ROCM_SEGMENTED_ATTN do not start, compared against TRITON_ATTN; decode tok/s; Gemma-4-26B-A4B-AWQ-4bit at in/out 128/128 (OSL 128), ours, ROCM_RDNA_SINGLESEQ_ATTN; TTFT 117 ms; config=rdna",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-4bit",
   "quote": "Gemma 4 26B-A4B; gfx1151; AWQ-4bit; Gemma-4-26B-A4B-AWQ-4bit | 128/128 | 117 / 128 | 37.9 / 36.3",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 37.9,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-gemma426ba4b-128-128-triton-attn",
   "model": "Gemma 4 26B-A4B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, Gemma 4 D=512 head dim where ROCM_ATTN and ROCM_SEGMENTED_ATTN do not start, compared against TRITON_ATTN; decode tok/s; Gemma-4-26B-A4B-AWQ-4bit at in/out 128/128 (OSL 128), TRITON_ATTN baseline; TTFT 128 ms; config=triton-attn",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-4bit",
   "quote": "Gemma 4 26B-A4B; gfx1151; AWQ-4bit; Gemma-4-26B-A4B-AWQ-4bit | 128/128 | 117 / 128 | 37.9 / 36.3",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-gemma426ba4b-128-4k-rdna",
   "model": "Gemma 4 26B-A4B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, Gemma 4 D=512 head dim where ROCM_ATTN and ROCM_SEGMENTED_ATTN do not start, compared against TRITON_ATTN; decode tok/s; Gemma-4-26B-A4B-AWQ-4bit at in/out 128/4k (OSL 4096), ours, ROCM_RDNA_SINGLESEQ_ATTN; TTFT 118 ms; config=rdna",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-4bit",
   "quote": "Gemma 4 26B-A4B; gfx1151; AWQ-4bit; Gemma-4-26B-A4B-AWQ-4bit | 128/4k | 118 / 128 | 37.7 / 32.0",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 37.7,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-gemma426ba4b-128-4k-triton-attn",
   "model": "Gemma 4 26B-A4B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, Gemma 4 D=512 head dim where ROCM_ATTN and ROCM_SEGMENTED_ATTN do not start, compared against TRITON_ATTN; decode tok/s; Gemma-4-26B-A4B-AWQ-4bit at in/out 128/4k (OSL 4096), TRITON_ATTN baseline; TTFT 128 ms; config=triton-attn",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-4bit",
   "quote": "Gemma 4 26B-A4B; gfx1151; AWQ-4bit; Gemma-4-26B-A4B-AWQ-4bit | 128/4k | 118 / 128 | 37.7 / 32.0",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-gemma426ba4b-128-16k-rdna",
   "model": "Gemma 4 26B-A4B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, Gemma 4 D=512 head dim where ROCM_ATTN and ROCM_SEGMENTED_ATTN do not start, compared against TRITON_ATTN; decode tok/s; Gemma-4-26B-A4B-AWQ-4bit at in/out 128/16k (OSL 16384), ours, ROCM_RDNA_SINGLESEQ_ATTN; TTFT 125 ms; config=rdna",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-4bit",
   "quote": "Gemma 4 26B-A4B; gfx1151; AWQ-4bit; Gemma-4-26B-A4B-AWQ-4bit | 128/16k | 125 / 135 | 34.8 / 24.3",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 34.8,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-gemma426ba4b-128-16k-triton-attn",
   "model": "Gemma 4 26B-A4B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, Gemma 4 D=512 head dim where ROCM_ATTN and ROCM_SEGMENTED_ATTN do not start, compared against TRITON_ATTN; decode tok/s; Gemma-4-26B-A4B-AWQ-4bit at in/out 128/16k (OSL 16384), TRITON_ATTN baseline; TTFT 135 ms; config=triton-attn",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-4bit",
   "quote": "Gemma 4 26B-A4B; gfx1151; AWQ-4bit; Gemma-4-26B-A4B-AWQ-4bit | 128/16k | 125 / 135 | 34.8 / 24.3",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 24.3,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-gemma426ba4b-8k-128-rdna",
   "model": "Gemma 4 26B-A4B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, Gemma 4 D=512 head dim where ROCM_ATTN and ROCM_SEGMENTED_ATTN do not start, compared against TRITON_ATTN; decode tok/s; Gemma-4-26B-A4B-AWQ-4bit at in/out 8k/128 (OSL 128), ours, ROCM_RDNA_SINGLESEQ_ATTN; TTFT 3024 ms; config=rdna",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-4bit",
   "quote": "Gemma 4 26B-A4B; gfx1151; AWQ-4bit; Gemma-4-26B-A4B-AWQ-4bit | 8k/128 | 3024 / 10288 | 19.8 / 8.4",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 19.8,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-gemma426ba4b-8k-128-triton-attn",
   "model": "Gemma 4 26B-A4B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, Gemma 4 D=512 head dim where ROCM_ATTN and ROCM_SEGMENTED_ATTN do not start, compared against TRITON_ATTN; decode tok/s; Gemma-4-26B-A4B-AWQ-4bit at in/out 8k/128 (OSL 128), TRITON_ATTN baseline; TTFT 10288 ms; config=triton-attn",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-4bit",
   "quote": "Gemma 4 26B-A4B; gfx1151; AWQ-4bit; Gemma-4-26B-A4B-AWQ-4bit | 8k/128 | 3024 / 10288 | 19.8 / 8.4",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 8.4,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-gemma426ba4b-8k-4k-rdna",
   "model": "Gemma 4 26B-A4B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, Gemma 4 D=512 head dim where ROCM_ATTN and ROCM_SEGMENTED_ATTN do not start, compared against TRITON_ATTN; decode tok/s; Gemma-4-26B-A4B-AWQ-4bit at in/out 8k/4k (OSL 4096), ours, ROCM_RDNA_SINGLESEQ_ATTN; TTFT 3013 ms; config=rdna",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-4bit",
   "quote": "Gemma 4 26B-A4B; gfx1151; AWQ-4bit; Gemma-4-26B-A4B-AWQ-4bit | 8k/4k | 3013 / 10258 | 35.7 / 22.9",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 35.7,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-gemma426ba4b-8k-4k-triton-attn",
   "model": "Gemma 4 26B-A4B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, Gemma 4 D=512 head dim where ROCM_ATTN and ROCM_SEGMENTED_ATTN do not start, compared against TRITON_ATTN; decode tok/s; Gemma-4-26B-A4B-AWQ-4bit at in/out 8k/4k (OSL 4096), TRITON_ATTN baseline; TTFT 10258 ms; config=triton-attn",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-4bit",
   "quote": "Gemma 4 26B-A4B; gfx1151; AWQ-4bit; Gemma-4-26B-A4B-AWQ-4bit | 8k/4k | 3013 / 10258 | 35.7 / 22.9",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 22.9,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-gemma426ba4b-8k-16k-rdna",
   "model": "Gemma 4 26B-A4B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, Gemma 4 D=512 head dim where ROCM_ATTN and ROCM_SEGMENTED_ATTN do not start, compared against TRITON_ATTN; decode tok/s; Gemma-4-26B-A4B-AWQ-4bit at in/out 8k/16k (OSL 16384), ours, ROCM_RDNA_SINGLESEQ_ATTN; TTFT 3113 ms; config=rdna",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-4bit",
   "quote": "Gemma 4 26B-A4B; gfx1151; AWQ-4bit; Gemma-4-26B-A4B-AWQ-4bit | 8k/16k | 3113 / 10877 | 35.1 / 21.1",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 35.1,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-gemma426ba4b-8k-16k-triton-attn",
   "model": "Gemma 4 26B-A4B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, Gemma 4 D=512 head dim where ROCM_ATTN and ROCM_SEGMENTED_ATTN do not start, compared against TRITON_ATTN; decode tok/s; Gemma-4-26B-A4B-AWQ-4bit at in/out 8k/16k (OSL 16384), TRITON_ATTN baseline; TTFT 10877 ms; config=triton-attn",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-4bit",
   "quote": "Gemma 4 26B-A4B; gfx1151; AWQ-4bit; Gemma-4-26B-A4B-AWQ-4bit | 8k/16k | 3113 / 10877 | 35.1 / 21.1",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-gemma431b-128-128-rdna",
   "model": "Gemma-4-31B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, Gemma 4 D=512 head dim where ROCM_ATTN and ROCM_SEGMENTED_ATTN do not start, compared against TRITON_ATTN; decode tok/s; Gemma-4-31B-AWQ-4bit MTP at in/out 128/128 (OSL 128), ours, ROCM_RDNA_SINGLESEQ_ATTN; TTFT 518 ms; MTP speculative decoding enabled; config=rdna",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-4bit",
   "quote": "Gemma-4-31B; gfx1151; AWQ-4bit; Gemma-4-31B-AWQ-4bit MTP | 128/128 | 518 / 551 | 9.5 / 8.5",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 9.5,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-gemma431b-128-128-triton-attn",
   "model": "Gemma-4-31B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, Gemma 4 D=512 head dim where ROCM_ATTN and ROCM_SEGMENTED_ATTN do not start, compared against TRITON_ATTN; decode tok/s; Gemma-4-31B-AWQ-4bit MTP at in/out 128/128 (OSL 128), TRITON_ATTN baseline; TTFT 551 ms; MTP speculative decoding enabled; config=triton-attn",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-4bit",
   "quote": "Gemma-4-31B; gfx1151; AWQ-4bit; Gemma-4-31B-AWQ-4bit MTP | 128/128 | 518 / 551 | 9.5 / 8.5",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 8.5,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-gemma431b-128-4k-rdna",
   "model": "Gemma-4-31B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, Gemma 4 D=512 head dim where ROCM_ATTN and ROCM_SEGMENTED_ATTN do not start, compared against TRITON_ATTN; decode tok/s; Gemma-4-31B-AWQ-4bit MTP at in/out 128/4k (OSL 4096), ours, ROCM_RDNA_SINGLESEQ_ATTN; TTFT 594 ms; MTP speculative decoding enabled; config=rdna",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-4bit",
   "quote": "Gemma-4-31B; gfx1151; AWQ-4bit; Gemma-4-31B-AWQ-4bit MTP | 128/4k | 594 / 612 | 6.6 / 4.6",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 6.6,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-gemma431b-128-4k-triton-attn",
   "model": "Gemma-4-31B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, Gemma 4 D=512 head dim where ROCM_ATTN and ROCM_SEGMENTED_ATTN do not start, compared against TRITON_ATTN; decode tok/s; Gemma-4-31B-AWQ-4bit MTP at in/out 128/4k (OSL 4096), TRITON_ATTN baseline; TTFT 612 ms; MTP speculative decoding enabled; config=triton-attn",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-4bit",
   "quote": "Gemma-4-31B; gfx1151; AWQ-4bit; Gemma-4-31B-AWQ-4bit MTP | 128/4k | 594 / 612 | 6.6 / 4.6",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 4.6,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-gemma431b-128-16k-rdna",
   "model": "Gemma-4-31B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, Gemma 4 D=512 head dim where ROCM_ATTN and ROCM_SEGMENTED_ATTN do not start, compared against TRITON_ATTN; decode tok/s; Gemma-4-31B-AWQ-4bit MTP at in/out 128/16k (OSL 16384), ours, ROCM_RDNA_SINGLESEQ_ATTN; TTFT 542 ms; MTP speculative decoding enabled; config=rdna",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-4bit",
   "quote": "Gemma-4-31B; gfx1151; AWQ-4bit; Gemma-4-31B-AWQ-4bit MTP | 128/16k | 542 / 556 | 6.6 / 3.4",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 6.6,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 128,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-gemma431b-128-16k-triton-attn",
   "model": "Gemma-4-31B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, Gemma 4 D=512 head dim where ROCM_ATTN and ROCM_SEGMENTED_ATTN do not start, compared against TRITON_ATTN; decode tok/s; Gemma-4-31B-AWQ-4bit MTP at in/out 128/16k (OSL 16384), TRITON_ATTN baseline; TTFT 556 ms; MTP speculative decoding enabled; config=triton-attn",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-4bit",
   "quote": "Gemma-4-31B; gfx1151; AWQ-4bit; Gemma-4-31B-AWQ-4bit MTP | 128/16k | 542 / 556 | 6.6 / 3.4",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 3.4,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-gemma431b-8k-128-rdna",
   "model": "Gemma-4-31B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, Gemma 4 D=512 head dim where ROCM_ATTN and ROCM_SEGMENTED_ATTN do not start, compared against TRITON_ATTN; decode tok/s; Gemma-4-31B-AWQ-4bit MTP at in/out 8k/128 (OSL 128), ours, ROCM_RDNA_SINGLESEQ_ATTN; TTFT 22083 ms; MTP speculative decoding enabled; config=rdna",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-4bit",
   "quote": "Gemma-4-31B; gfx1151; AWQ-4bit; Gemma-4-31B-AWQ-4bit MTP | 8k/128 | 22083 / 51655 | 4.6 / 2.0",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 4.6,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-gemma431b-8k-128-triton-attn",
   "model": "Gemma-4-31B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, Gemma 4 D=512 head dim where ROCM_ATTN and ROCM_SEGMENTED_ATTN do not start, compared against TRITON_ATTN; decode tok/s; Gemma-4-31B-AWQ-4bit MTP at in/out 8k/128 (OSL 128), TRITON_ATTN baseline; TTFT 51655 ms; MTP speculative decoding enabled; config=triton-attn",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-4bit",
   "quote": "Gemma-4-31B; gfx1151; AWQ-4bit; Gemma-4-31B-AWQ-4bit MTP | 8k/128 | 22083 / 51655 | 4.6 / 2.0",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 2.0,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-gemma431b-8k-4k-rdna",
   "model": "Gemma-4-31B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, Gemma 4 D=512 head dim where ROCM_ATTN and ROCM_SEGMENTED_ATTN do not start, compared against TRITON_ATTN; decode tok/s; Gemma-4-31B-AWQ-4bit MTP at in/out 8k/4k (OSL 4096), ours, ROCM_RDNA_SINGLESEQ_ATTN; TTFT 25605 ms; MTP speculative decoding enabled; config=rdna",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-4bit",
   "quote": "Gemma-4-31B; gfx1151; AWQ-4bit; Gemma-4-31B-AWQ-4bit MTP | 8k/4k | 25605 / 57609 | 22.5 / 10.5",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 22.5,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-gemma431b-8k-4k-triton-attn",
   "model": "Gemma-4-31B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, Gemma 4 D=512 head dim where ROCM_ATTN and ROCM_SEGMENTED_ATTN do not start, compared against TRITON_ATTN; decode tok/s; Gemma-4-31B-AWQ-4bit MTP at in/out 8k/4k (OSL 4096), TRITON_ATTN baseline; TTFT 57609 ms; MTP speculative decoding enabled; config=triton-attn",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-4bit",
   "quote": "Gemma-4-31B; gfx1151; AWQ-4bit; Gemma-4-31B-AWQ-4bit MTP | 8k/4k | 25605 / 57609 | 22.5 / 10.5",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 10.5,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-gemma431b-8k-16k-rdna",
   "model": "Gemma-4-31B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, Gemma 4 D=512 head dim where ROCM_ATTN and ROCM_SEGMENTED_ATTN do not start, compared against TRITON_ATTN; decode tok/s; Gemma-4-31B-AWQ-4bit MTP at in/out 8k/16k (OSL 16384), ours, ROCM_RDNA_SINGLESEQ_ATTN; TTFT 25572 ms; MTP speculative decoding enabled; config=rdna",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-4bit",
   "quote": "Gemma-4-31B; gfx1151; AWQ-4bit; Gemma-4-31B-AWQ-4bit MTP | 8k/16k | 25572 / 54678 | 24.2 / 9.8",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 24.2,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "gfx1151",
   "id": "vllm-61421-gfx1151-gemma431b-8k-16k-triton-attn",
   "model": "Gemma-4-31B",
   "notes": "vLLM PR #61421 opt-in ROCM_RDNA_SINGLESEQ_ATTN backend for gfx1151 RDNA3.5, vllm bench serve --max-num-seqs 1, Gemma 4 D=512 head dim where ROCM_ATTN and ROCM_SEGMENTED_ATTN do not start, compared against TRITON_ATTN; decode tok/s; Gemma-4-31B-AWQ-4bit MTP at in/out 8k/16k (OSL 16384), TRITON_ATTN baseline; TTFT 54678 ms; MTP speculative decoding enabled; config=triton-attn",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "AWQ-4bit",
   "quote": "Gemma-4-31B; gfx1151; AWQ-4bit; Gemma-4-31B-AWQ-4bit MTP | 8k/16k | 25572 / 54678 | 24.2 / 9.8",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61421",
   "source_url": "https://github.com/vllm-project/vllm/pull/61421",
   "tg_tokens": null,
   "tps": 9.8,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "H20-3e",
   "id": "vllm-61088-h203e-dsv41flash-cold16k-c16-base",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #61088 fused DeepSeek-V4.1 C1/C2 compression with RoPE, FP8 quantization and 584-byte main-cache write, two TP4 instances crossover on H20-3e SM90 (disjoint GPU groups, swapped placement phase 1), DeepSeek-V4.1-Flash TP4/EP DP1 MRV2, Humming dense/MoE, Engram CPU offload, APC on, 16K chunk, real DSpark 5 acceptance, precision unstated on page; 1,024 actual output tokens per request from fixed 4K/16K ledger prompts, throughput includes complete HTTP batch and tail drain, six batches per arm/cell; aggregate output tok/s; Cold 16K / c16 \u2014 primary, baseline arm; pooled median change -0.05%, phase 0 -0.09%, swapped phase 1 +0.61%, CV 1.92% / 1.70%; cold requests report exactly zero cache hits; config=baseline",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "DeepSeek-V4.1-Flash; H20-3e; Cold 16K / c16 \u2014 primary | 355.06 | 354.86",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61088",
   "source_url": "https://github.com/vllm-project/vllm/pull/61088",
   "tg_tokens": null,
   "tps": 355.06,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "H20-3e",
   "id": "vllm-61088-h203e-dsv41flash-cold16k-c16-fused",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #61088 fused DeepSeek-V4.1 C1/C2 compression with RoPE, FP8 quantization and 584-byte main-cache write, two TP4 instances crossover on H20-3e SM90 (disjoint GPU groups, swapped placement phase 1), DeepSeek-V4.1-Flash TP4/EP DP1 MRV2, Humming dense/MoE, Engram CPU offload, APC on, 16K chunk, real DSpark 5 acceptance, precision unstated on page; 1,024 actual output tokens per request from fixed 4K/16K ledger prompts, throughput includes complete HTTP batch and tail drain, six batches per arm/cell; aggregate output tok/s; Cold 16K / c16 \u2014 primary, fused arm; pooled median change -0.05%, phase 0 -0.09%, swapped phase 1 +0.61%, CV 1.92% / 1.70%; cold requests report exactly zero cache hits; config=fused",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "DeepSeek-V4.1-Flash; H20-3e; Cold 16K / c16 \u2014 primary | 355.06 | 354.86",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61088",
   "source_url": "https://github.com/vllm-project/vllm/pull/61088",
   "tg_tokens": null,
   "tps": 354.86,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "H20-3e",
   "id": "vllm-61088-h203e-dsv41flash-warm16k-c16-base",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #61088 fused DeepSeek-V4.1 C1/C2 compression with RoPE, FP8 quantization and 584-byte main-cache write, two TP4 instances crossover on H20-3e SM90 (disjoint GPU groups, swapped placement phase 1), DeepSeek-V4.1-Flash TP4/EP DP1 MRV2, Humming dense/MoE, Engram CPU offload, APC on, 16K chunk, real DSpark 5 acceptance, precision unstated on page; 1,024 actual output tokens per request from fixed 4K/16K ledger prompts, throughput includes complete HTTP batch and tail drain, six batches per arm/cell; aggregate output tok/s; Warm 16K / c16, baseline arm; pooled median change -0.77%, phase 0 -2.09%, swapped phase 1 -0.15%; warm requests retain 64 residual tokens, observed hits 99.6% at 16K and 98.4% at 4K; config=baseline",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "DeepSeek-V4.1-Flash; H20-3e; Warm 16K / c16 | 846.13 | 839.59",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61088",
   "source_url": "https://github.com/vllm-project/vllm/pull/61088",
   "tg_tokens": null,
   "tps": 846.13,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "H20-3e",
   "id": "vllm-61088-h203e-dsv41flash-warm16k-c16-fused",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #61088 fused DeepSeek-V4.1 C1/C2 compression with RoPE, FP8 quantization and 584-byte main-cache write, two TP4 instances crossover on H20-3e SM90 (disjoint GPU groups, swapped placement phase 1), DeepSeek-V4.1-Flash TP4/EP DP1 MRV2, Humming dense/MoE, Engram CPU offload, APC on, 16K chunk, real DSpark 5 acceptance, precision unstated on page; 1,024 actual output tokens per request from fixed 4K/16K ledger prompts, throughput includes complete HTTP batch and tail drain, six batches per arm/cell; aggregate output tok/s; Warm 16K / c16, fused arm; pooled median change -0.77%, phase 0 -2.09%, swapped phase 1 -0.15%; warm requests retain 64 residual tokens, observed hits 99.6% at 16K and 98.4% at 4K; config=fused",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "DeepSeek-V4.1-Flash; H20-3e; Warm 16K / c16 | 846.13 | 839.59",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61088",
   "source_url": "https://github.com/vllm-project/vllm/pull/61088",
   "tg_tokens": null,
   "tps": 839.59,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "H20-3e",
   "id": "vllm-61088-h203e-dsv41flash-cold4k-c4-base",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #61088 fused DeepSeek-V4.1 C1/C2 compression with RoPE, FP8 quantization and 584-byte main-cache write, two TP4 instances crossover on H20-3e SM90 (disjoint GPU groups, swapped placement phase 1), DeepSeek-V4.1-Flash TP4/EP DP1 MRV2, Humming dense/MoE, Engram CPU offload, APC on, 16K chunk, real DSpark 5 acceptance, precision unstated on page; 1,024 actual output tokens per request from fixed 4K/16K ledger prompts, throughput includes complete HTTP batch and tail drain, six batches per arm/cell; aggregate output tok/s; Cold 4K / c4, baseline arm; pooled median change -1.92%, phase 0 -11.39%, swapped phase 1 +19.68%; cold requests report exactly zero cache hits; config=baseline",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "DeepSeek-V4.1-Flash; H20-3e; Cold 4K / c4 | 479.74 | 470.55",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61088",
   "source_url": "https://github.com/vllm-project/vllm/pull/61088",
   "tg_tokens": null,
   "tps": 479.74,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "H20-3e",
   "id": "vllm-61088-h203e-dsv41flash-cold4k-c4-fused",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #61088 fused DeepSeek-V4.1 C1/C2 compression with RoPE, FP8 quantization and 584-byte main-cache write, two TP4 instances crossover on H20-3e SM90 (disjoint GPU groups, swapped placement phase 1), DeepSeek-V4.1-Flash TP4/EP DP1 MRV2, Humming dense/MoE, Engram CPU offload, APC on, 16K chunk, real DSpark 5 acceptance, precision unstated on page; 1,024 actual output tokens per request from fixed 4K/16K ledger prompts, throughput includes complete HTTP batch and tail drain, six batches per arm/cell; aggregate output tok/s; Cold 4K / c4, fused arm; pooled median change -1.92%, phase 0 -11.39%, swapped phase 1 +19.68%; cold requests report exactly zero cache hits; config=fused",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "DeepSeek-V4.1-Flash; H20-3e; Cold 4K / c4 | 479.74 | 470.55",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61088",
   "source_url": "https://github.com/vllm-project/vllm/pull/61088",
   "tg_tokens": null,
   "tps": 470.55,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "H20-3e",
   "id": "vllm-61088-h203e-dsv41flash-warm4k-c4-base",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #61088 fused DeepSeek-V4.1 C1/C2 compression with RoPE, FP8 quantization and 584-byte main-cache write, two TP4 instances crossover on H20-3e SM90 (disjoint GPU groups, swapped placement phase 1), DeepSeek-V4.1-Flash TP4/EP DP1 MRV2, Humming dense/MoE, Engram CPU offload, APC on, 16K chunk, real DSpark 5 acceptance, precision unstated on page; 1,024 actual output tokens per request from fixed 4K/16K ledger prompts, throughput includes complete HTTP batch and tail drain, six batches per arm/cell; aggregate output tok/s; Warm 4K / c4, baseline arm; pooled median change -7.49%, phase 0 -15.63%, swapped phase 1 +9.45%; warm requests retain 64 residual tokens, observed hits 99.6% at 16K and 98.4% at 4K; config=baseline",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "DeepSeek-V4.1-Flash; H20-3e; Warm 4K / c4 | 562.92 | 520.78",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61088",
   "source_url": "https://github.com/vllm-project/vllm/pull/61088",
   "tg_tokens": null,
   "tps": 562.92,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "H20-3e",
   "id": "vllm-61088-h203e-dsv41flash-warm4k-c4-fused",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #61088 fused DeepSeek-V4.1 C1/C2 compression with RoPE, FP8 quantization and 584-byte main-cache write, two TP4 instances crossover on H20-3e SM90 (disjoint GPU groups, swapped placement phase 1), DeepSeek-V4.1-Flash TP4/EP DP1 MRV2, Humming dense/MoE, Engram CPU offload, APC on, 16K chunk, real DSpark 5 acceptance, precision unstated on page; 1,024 actual output tokens per request from fixed 4K/16K ledger prompts, throughput includes complete HTTP batch and tail drain, six batches per arm/cell; aggregate output tok/s; Warm 4K / c4, fused arm; pooled median change -7.49%, phase 0 -15.63%, swapped phase 1 +9.45%; warm requests retain 64 residual tokens, observed hits 99.6% at 16K and 98.4% at 4K; config=fused",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "DeepSeek-V4.1-Flash; H20-3e; Warm 4K / c4 | 562.92 | 520.78",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61088",
   "source_url": "https://github.com/vllm-project/vllm/pull/61088",
   "tg_tokens": null,
   "tps": 520.78,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "H20-3e",
   "id": "vllm-61088-h203e-dsv41flash-cold16k-c4-base",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #61088 fused DeepSeek-V4.1 C1/C2 compression with RoPE, FP8 quantization and 584-byte main-cache write, two TP4 instances crossover on H20-3e SM90 (disjoint GPU groups, swapped placement phase 1), DeepSeek-V4.1-Flash TP4/EP DP1 MRV2, Humming dense/MoE, Engram CPU offload, APC on, 16K chunk, real DSpark 5 acceptance, precision unstated on page; 1,024 actual output tokens per request from fixed 4K/16K ledger prompts, throughput includes complete HTTP batch and tail drain, six batches per arm/cell; aggregate output tok/s; Cold 16K / c4, baseline arm; pooled median change +2.29%, phase 0 +3.81%, swapped phase 1 +2.45%, CV 7.04% / 5.12%, positive in both placements; cold requests report exactly zero cache hits; config=baseline",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "DeepSeek-V4.1-Flash; H20-3e; Cold 16K / c4 | 270.43 | 276.63",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61088",
   "source_url": "https://github.com/vllm-project/vllm/pull/61088",
   "tg_tokens": null,
   "tps": 270.43,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "H20-3e",
   "id": "vllm-61088-h203e-dsv41flash-cold16k-c4-fused",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #61088 fused DeepSeek-V4.1 C1/C2 compression with RoPE, FP8 quantization and 584-byte main-cache write, two TP4 instances crossover on H20-3e SM90 (disjoint GPU groups, swapped placement phase 1), DeepSeek-V4.1-Flash TP4/EP DP1 MRV2, Humming dense/MoE, Engram CPU offload, APC on, 16K chunk, real DSpark 5 acceptance, precision unstated on page; 1,024 actual output tokens per request from fixed 4K/16K ledger prompts, throughput includes complete HTTP batch and tail drain, six batches per arm/cell; aggregate output tok/s; Cold 16K / c4, fused arm; pooled median change +2.29%, phase 0 +3.81%, swapped phase 1 +2.45%, CV 7.04% / 5.12%, positive in both placements; cold requests report exactly zero cache hits; config=fused",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "DeepSeek-V4.1-Flash; H20-3e; Cold 16K / c4 | 270.43 | 276.63",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61088",
   "source_url": "https://github.com/vllm-project/vllm/pull/61088",
   "tg_tokens": null,
   "tps": 276.63,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "H20-3e",
   "id": "vllm-61088-h203e-dsv41flash-warm16k-c4-base",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #61088 fused DeepSeek-V4.1 C1/C2 compression with RoPE, FP8 quantization and 584-byte main-cache write, two TP4 instances crossover on H20-3e SM90 (disjoint GPU groups, swapped placement phase 1), DeepSeek-V4.1-Flash TP4/EP DP1 MRV2, Humming dense/MoE, Engram CPU offload, APC on, 16K chunk, real DSpark 5 acceptance, precision unstated on page; 1,024 actual output tokens per request from fixed 4K/16K ledger prompts, throughput includes complete HTTP batch and tail drain, six batches per arm/cell; aggregate output tok/s; Warm 16K / c4, baseline arm; pooled median change +0.95%, phase 0 +8.60%, swapped phase 1 -14.68%; warm requests retain 64 residual tokens, observed hits 99.6% at 16K and 98.4% at 4K; config=baseline",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "DeepSeek-V4.1-Flash; H20-3e; Warm 16K / c4 | 526.67 | 531.65",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61088",
   "source_url": "https://github.com/vllm-project/vllm/pull/61088",
   "tg_tokens": null,
   "tps": 526.67,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "H20-3e",
   "id": "vllm-61088-h203e-dsv41flash-warm16k-c4-fused",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #61088 fused DeepSeek-V4.1 C1/C2 compression with RoPE, FP8 quantization and 584-byte main-cache write, two TP4 instances crossover on H20-3e SM90 (disjoint GPU groups, swapped placement phase 1), DeepSeek-V4.1-Flash TP4/EP DP1 MRV2, Humming dense/MoE, Engram CPU offload, APC on, 16K chunk, real DSpark 5 acceptance, precision unstated on page; 1,024 actual output tokens per request from fixed 4K/16K ledger prompts, throughput includes complete HTTP batch and tail drain, six batches per arm/cell; aggregate output tok/s; Warm 16K / c4, fused arm; pooled median change +0.95%, phase 0 +8.60%, swapped phase 1 -14.68%; warm requests retain 64 residual tokens, observed hits 99.6% at 16K and 98.4% at 4K; config=fused",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "DeepSeek-V4.1-Flash; H20-3e; Warm 16K / c4 | 526.67 | 531.65",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61088",
   "source_url": "https://github.com/vllm-project/vllm/pull/61088",
   "tg_tokens": null,
   "tps": 531.65,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 512,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "Adreno 840",
   "id": "lc-30535-Adreno840-llama-32-1b-pp512-before",
   "model": "Llama-3.2-1B",
   "notes": "llama.cpp PR #30535 use dp4a q4_0 GEMM by default for batched matmul on OpenCL, pp512 prefill tok/s before (f16 GEMM) vs after (dp4a); Llama-3.2-1B Q4_0 on Adreno 840, before arm (f16 GEMM); gain after/before +18%; config=before",
   "params": null,
   "power_w": null,
   "pp_tps": 849,
   "provenance": "sourced",
   "quant": "Q4_0",
   "quote": "Llama-3.2-1B; Q4_0; Adreno 840; Adreno 840 | Llama-3.2-1B Q4_0 | 849 | 998",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30535",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30535",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 512,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "Adreno 840",
   "id": "lc-30535-Adreno840-llama-32-1b-pp512-after",
   "model": "Llama-3.2-1B",
   "notes": "llama.cpp PR #30535 use dp4a q4_0 GEMM by default for batched matmul on OpenCL, pp512 prefill tok/s before (f16 GEMM) vs after (dp4a); Llama-3.2-1B Q4_0 on Adreno 840, after arm (dp4a q4_0 GEMM default); gain after/before +18%; config=after",
   "params": null,
   "power_w": null,
   "pp_tps": 998,
   "provenance": "sourced",
   "quant": "Q4_0",
   "quote": "Llama-3.2-1B; Q4_0; Adreno 840; Adreno 840 | Llama-3.2-1B Q4_0 | 849 | 998",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30535",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30535",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 512,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "Adreno 840",
   "id": "lc-30535-Adreno840-gemma-4-e4b-pp512-before",
   "model": "Gemma 4 E4B",
   "notes": "llama.cpp PR #30535 use dp4a q4_0 GEMM by default for batched matmul on OpenCL, pp512 prefill tok/s before (f16 GEMM) vs after (dp4a); gemma-4-E4B Q4_0 on Adreno 840, before arm (f16 GEMM); gain after/before +15%; config=before",
   "params": null,
   "power_w": null,
   "pp_tps": 132,
   "provenance": "sourced",
   "quant": "Q4_0",
   "quote": "Gemma 4 E4B; Q4_0; Adreno 840; Adreno 840 | gemma-4-E4B Q4_0 | 132 | 152",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30535",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30535",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 512,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "Adreno 840",
   "id": "lc-30535-Adreno840-gemma-4-e4b-pp512-after",
   "model": "Gemma 4 E4B",
   "notes": "llama.cpp PR #30535 use dp4a q4_0 GEMM by default for batched matmul on OpenCL, pp512 prefill tok/s before (f16 GEMM) vs after (dp4a); gemma-4-E4B Q4_0 on Adreno 840, after arm (dp4a q4_0 GEMM default); gain after/before +15%; config=after",
   "params": null,
   "power_w": null,
   "pp_tps": 152,
   "provenance": "sourced",
   "quant": "Q4_0",
   "quote": "Gemma 4 E4B; Q4_0; Adreno 840; Adreno 840 | gemma-4-E4B Q4_0 | 132 | 152",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30535",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30535",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 512,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "Adreno 850",
   "id": "lc-30535-Adreno850-llama-32-1b-pp512-before",
   "model": "Llama-3.2-1B",
   "notes": "llama.cpp PR #30535 use dp4a q4_0 GEMM by default for batched matmul on OpenCL, pp512 prefill tok/s before (f16 GEMM) vs after (dp4a); Llama-3.2-1B Q4_0 on Adreno 850, before arm (f16 GEMM); gain after/before +14%; config=before",
   "params": null,
   "power_w": null,
   "pp_tps": 1202,
   "provenance": "sourced",
   "quant": "Q4_0",
   "quote": "Llama-3.2-1B; Q4_0; Adreno 850; Adreno 850 | Llama-3.2-1B Q4_0 | 1202 | 1371",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30535",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30535",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 512,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "Adreno 850",
   "id": "lc-30535-Adreno850-llama-32-1b-pp512-after",
   "model": "Llama-3.2-1B",
   "notes": "llama.cpp PR #30535 use dp4a q4_0 GEMM by default for batched matmul on OpenCL, pp512 prefill tok/s before (f16 GEMM) vs after (dp4a); Llama-3.2-1B Q4_0 on Adreno 850, after arm (dp4a q4_0 GEMM default); gain after/before +14%; config=after",
   "params": null,
   "power_w": null,
   "pp_tps": 1371,
   "provenance": "sourced",
   "quant": "Q4_0",
   "quote": "Llama-3.2-1B; Q4_0; Adreno 850; Adreno 850 | Llama-3.2-1B Q4_0 | 1202 | 1371",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30535",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30535",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 512,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "X2-90",
   "id": "lc-30535-X2-90-gemma-4-e2b-pp512-before",
   "model": "Gemma 4 E2B",
   "notes": "llama.cpp PR #30535 use dp4a q4_0 GEMM by default for batched matmul on OpenCL, pp512 prefill tok/s before (f16 GEMM) vs after (dp4a); gemma-4-E2B Q4_0 on X2-90, before arm (f16 GEMM); gain after/before +17%; config=before",
   "params": null,
   "power_w": null,
   "pp_tps": 881,
   "provenance": "sourced",
   "quant": "Q4_0",
   "quote": "Gemma 4 E2B; Q4_0; X2-90; X2-90 | gemma-4-E2B Q4_0 | 881 | 1028",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30535",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30535",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 512,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "X2-90",
   "id": "lc-30535-X2-90-gemma-4-e2b-pp512-after",
   "model": "Gemma 4 E2B",
   "notes": "llama.cpp PR #30535 use dp4a q4_0 GEMM by default for batched matmul on OpenCL, pp512 prefill tok/s before (f16 GEMM) vs after (dp4a); gemma-4-E2B Q4_0 on X2-90, after arm (dp4a q4_0 GEMM default); gain after/before +17%; config=after",
   "params": null,
   "power_w": null,
   "pp_tps": 1028,
   "provenance": "sourced",
   "quant": "Q4_0",
   "quote": "Gemma 4 E2B; Q4_0; X2-90; X2-90 | gemma-4-E2B Q4_0 | 881 | 1028",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30535",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30535",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 512,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "X2-90",
   "id": "lc-30535-X2-90-qwen38-27b-pp512-before",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #30535 use dp4a q4_0 GEMM by default for batched matmul on OpenCL, pp512 prefill tok/s before (f16 GEMM) vs after (dp4a); Qwen3.8-27B Q4_0 on X2-90, before arm (f16 GEMM); gain after/before +15%; config=before",
   "params": null,
   "power_w": null,
   "pp_tps": 95.8,
   "provenance": "sourced",
   "quant": "Q4_0",
   "quote": "Qwen3.8-27B; Q4_0; X2-90; X2-90 | Qwen3.8-27B Q4_0 | 95.8 | 110.3",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30535",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30535",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 512,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "X2-90",
   "id": "lc-30535-X2-90-qwen38-27b-pp512-after",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #30535 use dp4a q4_0 GEMM by default for batched matmul on OpenCL, pp512 prefill tok/s before (f16 GEMM) vs after (dp4a); Qwen3.8-27B Q4_0 on X2-90, after arm (dp4a q4_0 GEMM default); gain after/before +15%; config=after",
   "params": null,
   "power_w": null,
   "pp_tps": 110.3,
   "provenance": "sourced",
   "quant": "Q4_0",
   "quote": "Qwen3.8-27B; Q4_0; X2-90; X2-90 | Qwen3.8-27B Q4_0 | 95.8 | 110.3",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30535",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30535",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 8,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "Intel(R) Xeon(R) Platinum 8488C",
   "id": "lc-30526-xeon8488c-gemmaemb03b-pp8-old",
   "model": "gemma-embedding2 0.3B",
   "notes": "llama.cpp PR #30526 ggml-cpu optimize pad_f32, llama-bench prefill tok/s old vs new on Intel(R) Xeon(R) Platinum 8488C, gemma-embedding2 0.3B Q4_K_M; pp8 test, old arm; speedup new/old 2.24; config=old",
   "params": null,
   "power_w": null,
   "pp_tps": 652.25,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "gemma-embedding2 0.3B; Q4_K_M; Intel(R) Xeon(R) Platinum 8488C; gemma-embedding2 0.3B Q4_K_M | pp8 | 652.25 | 1460.06",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30526",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30526",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 8,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "Intel(R) Xeon(R) Platinum 8488C",
   "id": "lc-30526-xeon8488c-gemmaemb03b-pp8-new",
   "model": "gemma-embedding2 0.3B",
   "notes": "llama.cpp PR #30526 ggml-cpu optimize pad_f32, llama-bench prefill tok/s old vs new on Intel(R) Xeon(R) Platinum 8488C, gemma-embedding2 0.3B Q4_K_M; pp8 test, new arm (optimized pad_f32); speedup new/old 2.24; config=new",
   "params": null,
   "power_w": null,
   "pp_tps": 1460.06,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "gemma-embedding2 0.3B; Q4_K_M; Intel(R) Xeon(R) Platinum 8488C; gemma-embedding2 0.3B Q4_K_M | pp8 | 652.25 | 1460.06",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30526",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30526",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "Intel(R) Xeon(R) Platinum 8488C",
   "id": "lc-30526-xeon8488c-gemmaemb03b-pp32-old",
   "model": "gemma-embedding2 0.3B",
   "notes": "llama.cpp PR #30526 ggml-cpu optimize pad_f32, llama-bench prefill tok/s old vs new on Intel(R) Xeon(R) Platinum 8488C, gemma-embedding2 0.3B Q4_K_M; pp32 test, old arm; speedup new/old 1.60; config=old",
   "params": null,
   "power_w": null,
   "pp_tps": 1782.69,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "gemma-embedding2 0.3B; Q4_K_M; Intel(R) Xeon(R) Platinum 8488C; gemma-embedding2 0.3B Q4_K_M | pp32 | 1782.69 | 2845.38",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30526",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30526",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "Intel(R) Xeon(R) Platinum 8488C",
   "id": "lc-30526-xeon8488c-gemmaemb03b-pp32-new",
   "model": "gemma-embedding2 0.3B",
   "notes": "llama.cpp PR #30526 ggml-cpu optimize pad_f32, llama-bench prefill tok/s old vs new on Intel(R) Xeon(R) Platinum 8488C, gemma-embedding2 0.3B Q4_K_M; pp32 test, new arm (optimized pad_f32); speedup new/old 1.60; config=new",
   "params": null,
   "power_w": null,
   "pp_tps": 2845.38,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "gemma-embedding2 0.3B; Q4_K_M; Intel(R) Xeon(R) Platinum 8488C; gemma-embedding2 0.3B Q4_K_M | pp32 | 1782.69 | 2845.38",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30526",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30526",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 128,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "Intel(R) Xeon(R) Platinum 8488C",
   "id": "lc-30526-xeon8488c-gemmaemb03b-pp128-old",
   "model": "gemma-embedding2 0.3B",
   "notes": "llama.cpp PR #30526 ggml-cpu optimize pad_f32, llama-bench prefill tok/s old vs new on Intel(R) Xeon(R) Platinum 8488C, gemma-embedding2 0.3B Q4_K_M; pp128 test, old arm; speedup new/old 1.13; config=old",
   "params": null,
   "power_w": null,
   "pp_tps": 2264.76,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "gemma-embedding2 0.3B; Q4_K_M; Intel(R) Xeon(R) Platinum 8488C; gemma-embedding2 0.3B Q4_K_M | pp128 | 2264.76 | 2552.93",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30526",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30526",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 128,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "Intel(R) Xeon(R) Platinum 8488C",
   "id": "lc-30526-xeon8488c-gemmaemb03b-pp128-new",
   "model": "gemma-embedding2 0.3B",
   "notes": "llama.cpp PR #30526 ggml-cpu optimize pad_f32, llama-bench prefill tok/s old vs new on Intel(R) Xeon(R) Platinum 8488C, gemma-embedding2 0.3B Q4_K_M; pp128 test, new arm (optimized pad_f32); speedup new/old 1.13; config=new",
   "params": null,
   "power_w": null,
   "pp_tps": 2552.93,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "gemma-embedding2 0.3B; Q4_K_M; Intel(R) Xeon(R) Platinum 8488C; gemma-embedding2 0.3B Q4_K_M | pp128 | 2264.76 | 2552.93",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30526",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30526",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 256,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "Intel(R) Xeon(R) Platinum 8488C",
   "id": "lc-30526-xeon8488c-gemmaemb03b-pp256-old",
   "model": "gemma-embedding2 0.3B",
   "notes": "llama.cpp PR #30526 ggml-cpu optimize pad_f32, llama-bench prefill tok/s old vs new on Intel(R) Xeon(R) Platinum 8488C, gemma-embedding2 0.3B Q4_K_M; pp256 test, old arm; speedup new/old 1.08; config=old",
   "params": null,
   "power_w": null,
   "pp_tps": 2576.44,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "gemma-embedding2 0.3B; Q4_K_M; Intel(R) Xeon(R) Platinum 8488C; gemma-embedding2 0.3B Q4_K_M | pp256 | 2576.44 | 2777.91",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30526",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30526",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 256,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "Intel(R) Xeon(R) Platinum 8488C",
   "id": "lc-30526-xeon8488c-gemmaemb03b-pp256-new",
   "model": "gemma-embedding2 0.3B",
   "notes": "llama.cpp PR #30526 ggml-cpu optimize pad_f32, llama-bench prefill tok/s old vs new on Intel(R) Xeon(R) Platinum 8488C, gemma-embedding2 0.3B Q4_K_M; pp256 test, new arm (optimized pad_f32); speedup new/old 1.08; config=new",
   "params": null,
   "power_w": null,
   "pp_tps": 2777.91,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "gemma-embedding2 0.3B; Q4_K_M; Intel(R) Xeon(R) Platinum 8488C; gemma-embedding2 0.3B Q4_K_M | pp256 | 2576.44 | 2777.91",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30526",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30526",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 512,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "Intel(R) Xeon(R) Platinum 8488C",
   "id": "lc-30526-xeon8488c-gemmaemb03b-pp512-old",
   "model": "gemma-embedding2 0.3B",
   "notes": "llama.cpp PR #30526 ggml-cpu optimize pad_f32, llama-bench prefill tok/s old vs new on Intel(R) Xeon(R) Platinum 8488C, gemma-embedding2 0.3B Q4_K_M; pp512 test, old arm; speedup new/old 1.10; config=old",
   "params": null,
   "power_w": null,
   "pp_tps": 2893.72,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "gemma-embedding2 0.3B; Q4_K_M; Intel(R) Xeon(R) Platinum 8488C; gemma-embedding2 0.3B Q4_K_M | pp512 | 2893.72 | 3177.44",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30526",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30526",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 512,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "Intel(R) Xeon(R) Platinum 8488C",
   "id": "lc-30526-xeon8488c-gemmaemb03b-pp512-new",
   "model": "gemma-embedding2 0.3B",
   "notes": "llama.cpp PR #30526 ggml-cpu optimize pad_f32, llama-bench prefill tok/s old vs new on Intel(R) Xeon(R) Platinum 8488C, gemma-embedding2 0.3B Q4_K_M; pp512 test, new arm (optimized pad_f32); speedup new/old 1.10; config=new",
   "params": null,
   "power_w": null,
   "pp_tps": 3177.44,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "gemma-embedding2 0.3B; Q4_K_M; Intel(R) Xeon(R) Platinum 8488C; gemma-embedding2 0.3B Q4_K_M | pp512 | 2893.72 | 3177.44",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30526",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30526",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "Intel(R) Xeon(R) Platinum 8488C",
   "id": "lc-30526-xeon8488c-gemmaemb03b-pp1024-old",
   "model": "gemma-embedding2 0.3B",
   "notes": "llama.cpp PR #30526 ggml-cpu optimize pad_f32, llama-bench prefill tok/s old vs new on Intel(R) Xeon(R) Platinum 8488C, gemma-embedding2 0.3B Q4_K_M; pp1024 test, old arm; speedup new/old 1.07; config=old",
   "params": null,
   "power_w": null,
   "pp_tps": 2366.24,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "gemma-embedding2 0.3B; Q4_K_M; Intel(R) Xeon(R) Platinum 8488C; gemma-embedding2 0.3B Q4_K_M | pp1024 | 2366.24 | 2535.64",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30526",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30526",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "Intel(R) Xeon(R) Platinum 8488C",
   "id": "lc-30526-xeon8488c-gemmaemb03b-pp1024-new",
   "model": "gemma-embedding2 0.3B",
   "notes": "llama.cpp PR #30526 ggml-cpu optimize pad_f32, llama-bench prefill tok/s old vs new on Intel(R) Xeon(R) Platinum 8488C, gemma-embedding2 0.3B Q4_K_M; pp1024 test, new arm (optimized pad_f32); speedup new/old 1.07; config=new",
   "params": null,
   "power_w": null,
   "pp_tps": 2535.64,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "gemma-embedding2 0.3B; Q4_K_M; Intel(R) Xeon(R) Platinum 8488C; gemma-embedding2 0.3B Q4_K_M | pp1024 | 2366.24 | 2535.64",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp PR #30526",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30526",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 40960,
   "date": "2026-09-10",
   "flags": [],
   "hardware": "GB10",
   "id": "vllm-56273-gb10-qwen38flashnext-b12x-c1",
   "model": "Qwen3.8-Flash-Next",
   "notes": "vLLM PR #56273 packed NVFP4 Qwen4Exp PLE embeddings, single-DGX-Spark reference measurements of Qwen3.8-Flash-Next NVFP4, one GB10 TP1/PP1 40960-token context, prefix caching, chunked prefill, MTP with four draft tokens; workload 32768 shared prefix + 2048 suffix + 256 output tokens, streaming chat temp0, three warmups per cell; aggregate active decode throughput; local-inference-lab / b12x at concurrency 1 (10 measured requests), mean latency 11.27 s, acceptance length 2.4; local checkpoint packed PLE payload 26.822 GiB resident in b12x managed CUDA weight storage, KV cache 8 GiB FP8, 27216 reused tokens per request; config=b12x",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "NVFP4",
   "quote": "Qwen3.8-Flash-Next; NVFP4; GB10; local-inference-lab / b12x | 1 | 10 | 11.27 | 3.40 | 32.41 | 2.40",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #56273",
   "source_url": "https://github.com/vllm-project/vllm/pull/56273",
   "tg_tokens": null,
   "tps": 32.41,
   "ttft_s": 3.4
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 40960,
   "date": "2026-09-10",
   "flags": [],
   "hardware": "GB10",
   "id": "vllm-56273-gb10-qwen38flashnext-b12x-c4",
   "model": "Qwen3.8-Flash-Next",
   "notes": "vLLM PR #56273 packed NVFP4 Qwen4Exp PLE embeddings, single-DGX-Spark reference measurements of Qwen3.8-Flash-Next NVFP4, one GB10 TP1/PP1 40960-token context, prefix caching, chunked prefill, MTP with four draft tokens; workload 32768 shared prefix + 2048 suffix + 256 output tokens, streaming chat temp0, three warmups per cell; aggregate active decode throughput; local-inference-lab / b12x at concurrency 4 (40 measured requests), mean latency 28.19 s, acceptance length 2.13; local checkpoint packed PLE payload 26.822 GiB resident in b12x managed CUDA weight storage, KV cache 8 GiB FP8; config=b12x",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "NVFP4",
   "quote": "Qwen3.8-Flash-Next; NVFP4; GB10; local-inference-lab / b12x | 4 | 40 | 28.19 | 5.44 | 36.31 | 2.13",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #56273",
   "source_url": "https://github.com/vllm-project/vllm/pull/56273",
   "tg_tokens": null,
   "tps": 36.31,
   "ttft_s": 5.44
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 40960,
   "date": "2026-09-10",
   "flags": [],
   "hardware": "GB10",
   "id": "vllm-56273-gb10-qwen38flashnext-b12x-c8",
   "model": "Qwen3.8-Flash-Next",
   "notes": "vLLM PR #56273 packed NVFP4 Qwen4Exp PLE embeddings, single-DGX-Spark reference measurements of Qwen3.8-Flash-Next NVFP4, one GB10 TP1/PP1 40960-token context, prefix caching, chunked prefill, MTP with four draft tokens; workload 32768 shared prefix + 2048 suffix + 256 output tokens, streaming chat temp0, three warmups per cell; aggregate active decode throughput; local-inference-lab / b12x at concurrency 8 (80 measured requests), mean latency 45.99 s, acceptance length 2.2; local checkpoint packed PLE payload 26.822 GiB resident in b12x managed CUDA weight storage, KV cache 8 GiB FP8; config=b12x",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "NVFP4",
   "quote": "Qwen3.8-Flash-Next; NVFP4; GB10; local-inference-lab / b12x | 8 | 80 | 45.99 | 6.75 | 43.44 | 2.2",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #56273",
   "source_url": "https://github.com/vllm-project/vllm/pull/56273",
   "tg_tokens": null,
   "tps": 43.44,
   "ttft_s": 6.75
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 40960,
   "date": "2026-09-10",
   "flags": [],
   "hardware": "GB10",
   "id": "vllm-56273-gb10-qwen38flashnext-nvdisk-c1",
   "model": "Qwen3.8-Flash-Next",
   "notes": "vLLM PR #56273 packed NVFP4 Qwen4Exp PLE embeddings, single-DGX-Spark reference measurements of Qwen3.8-Flash-Next NVFP4, one GB10 TP1/PP1 40960-token context, prefix caching, chunked prefill, MTP with four draft tokens; workload 32768 shared prefix + 2048 suffix + 256 output tokens, streaming chat temp0, three warmups per cell; aggregate active decode throughput; NVIDIA / vLLM with disk PLE at concurrency 1 (10 measured requests), mean latency 14.3 s, acceptance length 2.64; NVIDIA checkpoint FP8 PLE payload 47.684 GiB through disk-backed mmap and OS page cache, KV cache 10 GiB BF16, 30704 reused tokens per request, host swap activity observed; config=nvdisk",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "NVFP4",
   "quote": "Qwen3.8-Flash-Next; NVFP4; GB10; NVIDIA / vLLM with disk PLE | 1 | 10 | 14.30 | 3.45 | 23.51 | 2.64",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #56273",
   "source_url": "https://github.com/vllm-project/vllm/pull/56273",
   "tg_tokens": null,
   "tps": 23.51,
   "ttft_s": 3.45
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 40960,
   "date": "2026-09-10",
   "flags": [],
   "hardware": "GB10",
   "id": "vllm-56273-gb10-qwen38flashnext-nvdisk-c4",
   "model": "Qwen3.8-Flash-Next",
   "notes": "vLLM PR #56273 packed NVFP4 Qwen4Exp PLE embeddings, single-DGX-Spark reference measurements of Qwen3.8-Flash-Next NVFP4, one GB10 TP1/PP1 40960-token context, prefix caching, chunked prefill, MTP with four draft tokens; workload 32768 shared prefix + 2048 suffix + 256 output tokens, streaming chat temp0, three warmups per cell; aggregate active decode throughput; NVIDIA / vLLM with disk PLE at concurrency 4 (40 measured requests), mean latency 29.6 s, acceptance length 2.39; NVIDIA checkpoint FP8 PLE payload 47.684 GiB through disk-backed mmap and OS page cache, KV cache 10 GiB BF16; config=nvdisk",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "NVFP4",
   "quote": "Qwen3.8-Flash-Next; NVFP4; GB10; NVIDIA / vLLM with disk PLE | 4 | 40 | 29.60 | 5.09 | 34.25 | 2.39",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #56273",
   "source_url": "https://github.com/vllm-project/vllm/pull/56273",
   "tg_tokens": null,
   "tps": 34.25,
   "ttft_s": 5.09
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": 40960,
   "date": "2026-09-10",
   "flags": [],
   "hardware": "GB10",
   "id": "vllm-56273-gb10-qwen38flashnext-nvdisk-c8",
   "model": "Qwen3.8-Flash-Next",
   "notes": "vLLM PR #56273 packed NVFP4 Qwen4Exp PLE embeddings, single-DGX-Spark reference measurements of Qwen3.8-Flash-Next NVFP4, one GB10 TP1/PP1 40960-token context, prefix caching, chunked prefill, MTP with four draft tokens; workload 32768 shared prefix + 2048 suffix + 256 output tokens, streaming chat temp0, three warmups per cell; aggregate active decode throughput; NVIDIA / vLLM with disk PLE at concurrency 8 (80 measured requests), mean latency 44.34 s, acceptance length 2.59; NVIDIA checkpoint FP8 PLE payload 47.684 GiB through disk-backed mmap and OS page cache, KV cache 10 GiB BF16; config=nvdisk",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "NVFP4",
   "quote": "Qwen3.8-Flash-Next; NVFP4; GB10; NVIDIA / vLLM with disk PLE | 8 | 80 | 44.34 | 7.12 | 45.92 | 2.59",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #56273",
   "source_url": "https://github.com/vllm-project/vllm/pull/56273",
   "tg_tokens": null,
   "tps": 45.92,
   "ttft_s": 7.12
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 2048,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "lc-29353-rtxpro6000-qwen3535ba3b-q80-2048-master",
   "model": "Qwen3.5-35B-A3B",
   "notes": "llama.cpp PR #29353 cuda/hip GDN chunked kernel, llama-bench prompt-processing tok/s master vs PR, all rows -b 2048 -ub 2048 -fa 1 full GPU offload F16 KV cache, retested 2026-10-13 against master 296538ade; Qwen3.5-35B-A3B Q8_0 at 2048 prompt tokens on RTX PRO 6000 Blackwell, change +20.6%; config=master",
   "params": null,
   "power_w": null,
   "pp_tps": 13653.58,
   "provenance": "sourced",
   "quant": "Q8_0",
   "quote": "Qwen3.5-35B-A3B; Q8_0; RTX PRO 6000; RTX PRO 6000 Blackwell | Qwen3.5-35B-A3B-Q8_0 | 2048 | 13,653.58 | 16,459.56",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "llama.cpp PR #29353",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29353",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 2048,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "lc-29353-rtxpro6000-qwen3535ba3b-q80-2048-pr",
   "model": "Qwen3.5-35B-A3B",
   "notes": "llama.cpp PR #29353 cuda/hip GDN chunked kernel, llama-bench prompt-processing tok/s master vs PR, all rows -b 2048 -ub 2048 -fa 1 full GPU offload F16 KV cache, retested 2026-10-13 against master 296538ade; Qwen3.5-35B-A3B Q8_0 at 2048 prompt tokens on RTX PRO 6000 Blackwell, change +20.6%; config=pr",
   "params": null,
   "power_w": null,
   "pp_tps": 16459.56,
   "provenance": "sourced",
   "quant": "Q8_0",
   "quote": "Qwen3.5-35B-A3B; Q8_0; RTX PRO 6000; RTX PRO 6000 Blackwell | Qwen3.5-35B-A3B-Q8_0 | 2048 | 13,653.58 | 16,459.56",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "llama.cpp PR #29353",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29353",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "lc-29353-rtxpro6000-qwen3535ba3b-q80-4096-master",
   "model": "Qwen3.5-35B-A3B",
   "notes": "llama.cpp PR #29353 cuda/hip GDN chunked kernel, llama-bench prompt-processing tok/s master vs PR, all rows -b 2048 -ub 2048 -fa 1 full GPU offload F16 KV cache, retested 2026-10-13 against master 296538ade; Qwen3.5-35B-A3B Q8_0 at 4096 prompt tokens on RTX PRO 6000 Blackwell, change +17.8%; config=master",
   "params": null,
   "power_w": null,
   "pp_tps": 14090.26,
   "provenance": "sourced",
   "quant": "Q8_0",
   "quote": "Qwen3.5-35B-A3B; Q8_0; RTX PRO 6000; RTX PRO 6000 Blackwell | Qwen3.5-35B-A3B-Q8_0 | 4096 | 14,090.26 | 16,595.47",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "llama.cpp PR #29353",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29353",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "lc-29353-rtxpro6000-qwen3535ba3b-q80-4096-pr",
   "model": "Qwen3.5-35B-A3B",
   "notes": "llama.cpp PR #29353 cuda/hip GDN chunked kernel, llama-bench prompt-processing tok/s master vs PR, all rows -b 2048 -ub 2048 -fa 1 full GPU offload F16 KV cache, retested 2026-10-13 against master 296538ade; Qwen3.5-35B-A3B Q8_0 at 4096 prompt tokens on RTX PRO 6000 Blackwell, change +17.8%; config=pr",
   "params": null,
   "power_w": null,
   "pp_tps": 16595.47,
   "provenance": "sourced",
   "quant": "Q8_0",
   "quote": "Qwen3.5-35B-A3B; Q8_0; RTX PRO 6000; RTX PRO 6000 Blackwell | Qwen3.5-35B-A3B-Q8_0 | 4096 | 14,090.26 | 16,595.47",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "llama.cpp PR #29353",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29353",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 2048,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "lc-29353-rtxpro6000-qwen3827b-q80-2048-master",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #29353 cuda/hip GDN chunked kernel, llama-bench prompt-processing tok/s master vs PR, all rows -b 2048 -ub 2048 -fa 1 full GPU offload F16 KV cache, retested 2026-10-13 against master 296538ade; Qwen3.8-27B Q8_0 at 2048 prompt tokens on RTX PRO 6000 Blackwell, change +6.4%; config=master",
   "params": null,
   "power_w": null,
   "pp_tps": 4470.39,
   "provenance": "sourced",
   "quant": "Q8_0",
   "quote": "Qwen3.8-27B; Q8_0; RTX PRO 6000; RTX PRO 6000 Blackwell | Qwen3.8-27B-Q8_0 | 2048 | 4,470.39 | 4,757.20",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "llama.cpp PR #29353",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29353",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 2048,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "lc-29353-rtxpro6000-qwen3827b-q80-2048-pr",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #29353 cuda/hip GDN chunked kernel, llama-bench prompt-processing tok/s master vs PR, all rows -b 2048 -ub 2048 -fa 1 full GPU offload F16 KV cache, retested 2026-10-13 against master 296538ade; Qwen3.8-27B Q8_0 at 2048 prompt tokens on RTX PRO 6000 Blackwell, change +6.4%; config=pr",
   "params": null,
   "power_w": null,
   "pp_tps": 4757.2,
   "provenance": "sourced",
   "quant": "Q8_0",
   "quote": "Qwen3.8-27B; Q8_0; RTX PRO 6000; RTX PRO 6000 Blackwell | Qwen3.8-27B-Q8_0 | 2048 | 4,470.39 | 4,757.20",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "llama.cpp PR #29353",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29353",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "lc-29353-rtxpro6000-qwen3827b-q80-4096-master",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #29353 cuda/hip GDN chunked kernel, llama-bench prompt-processing tok/s master vs PR, all rows -b 2048 -ub 2048 -fa 1 full GPU offload F16 KV cache, retested 2026-10-13 against master 296538ade; Qwen3.8-27B Q8_0 at 4096 prompt tokens on RTX PRO 6000 Blackwell, change +5.8%; config=master",
   "params": null,
   "power_w": null,
   "pp_tps": 4442.45,
   "provenance": "sourced",
   "quant": "Q8_0",
   "quote": "Qwen3.8-27B; Q8_0; RTX PRO 6000; RTX PRO 6000 Blackwell | Qwen3.8-27B-Q8_0 | 4096 | 4,442.45 | 4,699.08",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "llama.cpp PR #29353",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29353",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "lc-29353-rtxpro6000-qwen3827b-q80-4096-pr",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #29353 cuda/hip GDN chunked kernel, llama-bench prompt-processing tok/s master vs PR, all rows -b 2048 -ub 2048 -fa 1 full GPU offload F16 KV cache, retested 2026-10-13 against master 296538ade; Qwen3.8-27B Q8_0 at 4096 prompt tokens on RTX PRO 6000 Blackwell, change +5.8%; config=pr",
   "params": null,
   "power_w": null,
   "pp_tps": 4699.08,
   "provenance": "sourced",
   "quant": "Q8_0",
   "quote": "Qwen3.8-27B; Q8_0; RTX PRO 6000; RTX PRO 6000 Blackwell | Qwen3.8-27B-Q8_0 | 4096 | 4,442.45 | 4,699.08",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "llama.cpp PR #29353",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29353",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 2048,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-29353-gb10-qwen3635ba3b-udq4km-2048-master",
   "model": "Qwen3.6-35B-A3B",
   "notes": "llama.cpp PR #29353 cuda/hip GDN chunked kernel, llama-bench prompt-processing tok/s master vs PR, all rows -b 2048 -ub 2048 -fa 1 full GPU offload F16 KV cache, retested 2026-10-13 against master 296538ade; Qwen3.6-35B-A3B UD-Q4_K_M at 2048 prompt tokens on Spark GB10, change +4.1%; config=master",
   "params": null,
   "power_w": null,
   "pp_tps": 3271.18,
   "provenance": "sourced",
   "quant": "UD-Q4_K_M",
   "quote": "Qwen3.6-35B-A3B; UD-Q4_K_M; GB10; Spark GB10 | Qwen3.6-35B-A3B-UD-Q4_K_M | 2048 | 3,271.18 | 3,406.68",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "llama.cpp PR #29353",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29353",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 2048,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-29353-gb10-qwen3635ba3b-udq4km-2048-pr",
   "model": "Qwen3.6-35B-A3B",
   "notes": "llama.cpp PR #29353 cuda/hip GDN chunked kernel, llama-bench prompt-processing tok/s master vs PR, all rows -b 2048 -ub 2048 -fa 1 full GPU offload F16 KV cache, retested 2026-10-13 against master 296538ade; Qwen3.6-35B-A3B UD-Q4_K_M at 2048 prompt tokens on Spark GB10, change +4.1%; config=pr",
   "params": null,
   "power_w": null,
   "pp_tps": 3406.68,
   "provenance": "sourced",
   "quant": "UD-Q4_K_M",
   "quote": "Qwen3.6-35B-A3B; UD-Q4_K_M; GB10; Spark GB10 | Qwen3.6-35B-A3B-UD-Q4_K_M | 2048 | 3,271.18 | 3,406.68",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "llama.cpp PR #29353",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29353",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-29353-gb10-qwen3635ba3b-udq4km-4096-master",
   "model": "Qwen3.6-35B-A3B",
   "notes": "llama.cpp PR #29353 cuda/hip GDN chunked kernel, llama-bench prompt-processing tok/s master vs PR, all rows -b 2048 -ub 2048 -fa 1 full GPU offload F16 KV cache, retested 2026-10-13 against master 296538ade; Qwen3.6-35B-A3B UD-Q4_K_M at 4096 prompt tokens on Spark GB10, change +3.4%; config=master",
   "params": null,
   "power_w": null,
   "pp_tps": 3254.79,
   "provenance": "sourced",
   "quant": "UD-Q4_K_M",
   "quote": "Qwen3.6-35B-A3B; UD-Q4_K_M; GB10; Spark GB10 | Qwen3.6-35B-A3B-UD-Q4_K_M | 4096 | 3,254.79 | 3,366.88",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "llama.cpp PR #29353",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29353",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-29353-gb10-qwen3635ba3b-udq4km-4096-pr",
   "model": "Qwen3.6-35B-A3B",
   "notes": "llama.cpp PR #29353 cuda/hip GDN chunked kernel, llama-bench prompt-processing tok/s master vs PR, all rows -b 2048 -ub 2048 -fa 1 full GPU offload F16 KV cache, retested 2026-10-13 against master 296538ade; Qwen3.6-35B-A3B UD-Q4_K_M at 4096 prompt tokens on Spark GB10, change +3.4%; config=pr",
   "params": null,
   "power_w": null,
   "pp_tps": 3366.88,
   "provenance": "sourced",
   "quant": "UD-Q4_K_M",
   "quote": "Qwen3.6-35B-A3B; UD-Q4_K_M; GB10; Spark GB10 | Qwen3.6-35B-A3B-UD-Q4_K_M | 4096 | 3,254.79 | 3,366.88",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "llama.cpp PR #29353",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29353",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 2048,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-29353-gb10-qwen3827b-udq8kl-2048-master",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #29353 cuda/hip GDN chunked kernel, llama-bench prompt-processing tok/s master vs PR, all rows -b 2048 -ub 2048 -fa 1 full GPU offload F16 KV cache, retested 2026-10-13 against master 296538ade; Qwen3.8-27B UD-Q8_K_L at 2048 prompt tokens on Spark GB10, change +4.5%; config=master",
   "params": null,
   "power_w": null,
   "pp_tps": 942.84,
   "provenance": "sourced",
   "quant": "UD-Q8_K_L",
   "quote": "Qwen3.8-27B; UD-Q8_K_L; GB10; Spark GB10 | Qwen3.8-27B-UD-Q8_K_L | 2048 | 942.84 | 984.86",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "llama.cpp PR #29353",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29353",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 2048,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-29353-gb10-qwen3827b-udq8kl-2048-pr",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #29353 cuda/hip GDN chunked kernel, llama-bench prompt-processing tok/s master vs PR, all rows -b 2048 -ub 2048 -fa 1 full GPU offload F16 KV cache, retested 2026-10-13 against master 296538ade; Qwen3.8-27B UD-Q8_K_L at 2048 prompt tokens on Spark GB10, change +4.5%; config=pr",
   "params": null,
   "power_w": null,
   "pp_tps": 984.86,
   "provenance": "sourced",
   "quant": "UD-Q8_K_L",
   "quote": "Qwen3.8-27B; UD-Q8_K_L; GB10; Spark GB10 | Qwen3.8-27B-UD-Q8_K_L | 2048 | 942.84 | 984.86",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "llama.cpp PR #29353",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29353",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-29353-gb10-qwen3827b-udq8kl-4096-master",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #29353 cuda/hip GDN chunked kernel, llama-bench prompt-processing tok/s master vs PR, all rows -b 2048 -ub 2048 -fa 1 full GPU offload F16 KV cache, retested 2026-10-13 against master 296538ade; Qwen3.8-27B UD-Q8_K_L at 4096 prompt tokens on Spark GB10, change +4.9%; config=master",
   "params": null,
   "power_w": null,
   "pp_tps": 928.16,
   "provenance": "sourced",
   "quant": "UD-Q8_K_L",
   "quote": "Qwen3.8-27B; UD-Q8_K_L; GB10; Spark GB10 | Qwen3.8-27B-UD-Q8_K_L | 4096 | 928.16 | 973.34",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "llama.cpp PR #29353",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29353",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-29353-gb10-qwen3827b-udq8kl-4096-pr",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #29353 cuda/hip GDN chunked kernel, llama-bench prompt-processing tok/s master vs PR, all rows -b 2048 -ub 2048 -fa 1 full GPU offload F16 KV cache, retested 2026-10-13 against master 296538ade; Qwen3.8-27B UD-Q8_K_L at 4096 prompt tokens on Spark GB10, change +4.9%; config=pr",
   "params": null,
   "power_w": null,
   "pp_tps": 973.34,
   "provenance": "sourced",
   "quant": "UD-Q8_K_L",
   "quote": "Qwen3.8-27B; UD-Q8_K_L; GB10; Spark GB10 | Qwen3.8-27B-UD-Q8_K_L | 4096 | 928.16 | 973.34",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "llama.cpp PR #29353",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29353",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 2048,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-29353-gb10-qwen38flashnext-udiq4xs-2048-master",
   "model": "Qwen3.8-Flash-Next",
   "notes": "llama.cpp PR #29353 cuda/hip GDN chunked kernel, llama-bench prompt-processing tok/s master vs PR, all rows -b 2048 -ub 2048 -fa 1 full GPU offload F16 KV cache, retested 2026-10-13 against master 296538ade; Qwen3.8-Flash-Next UD-IQ4_XS at 2048 prompt tokens on Spark GB10, change +5.7%; config=master",
   "params": null,
   "power_w": null,
   "pp_tps": 1044.71,
   "provenance": "sourced",
   "quant": "UD-IQ4_XS",
   "quote": "Qwen3.8-Flash-Next; UD-IQ4_XS; GB10; Spark GB10 | Qwen3.8-Flash-Next-UD-IQ4_XS (`-lzm on`) | 2048 | 1,044.71 | 1,104.31",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "llama.cpp PR #29353",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29353",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 2048,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-29353-gb10-qwen38flashnext-udiq4xs-2048-pr",
   "model": "Qwen3.8-Flash-Next",
   "notes": "llama.cpp PR #29353 cuda/hip GDN chunked kernel, llama-bench prompt-processing tok/s master vs PR, all rows -b 2048 -ub 2048 -fa 1 full GPU offload F16 KV cache, retested 2026-10-13 against master 296538ade; Qwen3.8-Flash-Next UD-IQ4_XS at 2048 prompt tokens on Spark GB10, change +5.7%; config=pr",
   "params": null,
   "power_w": null,
   "pp_tps": 1104.31,
   "provenance": "sourced",
   "quant": "UD-IQ4_XS",
   "quote": "Qwen3.8-Flash-Next; UD-IQ4_XS; GB10; Spark GB10 | Qwen3.8-Flash-Next-UD-IQ4_XS (`-lzm on`) | 2048 | 1,044.71 | 1,104.31",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "llama.cpp PR #29353",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29353",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-29353-gb10-qwen38flashnext-udiq4xs-4096-master",
   "model": "Qwen3.8-Flash-Next",
   "notes": "llama.cpp PR #29353 cuda/hip GDN chunked kernel, llama-bench prompt-processing tok/s master vs PR, all rows -b 2048 -ub 2048 -fa 1 full GPU offload F16 KV cache, retested 2026-10-13 against master 296538ade; Qwen3.8-Flash-Next UD-IQ4_XS at 4096 prompt tokens on Spark GB10, change +5.2%; config=master",
   "params": null,
   "power_w": null,
   "pp_tps": 1068.7,
   "provenance": "sourced",
   "quant": "UD-IQ4_XS",
   "quote": "Qwen3.8-Flash-Next; UD-IQ4_XS; GB10; Spark GB10 | Qwen3.8-Flash-Next-UD-IQ4_XS (`-lzm on`) | 4096 | 1,068.70 | 1,124.01",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "llama.cpp PR #29353",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29353",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-29353-gb10-qwen38flashnext-udiq4xs-4096-pr",
   "model": "Qwen3.8-Flash-Next",
   "notes": "llama.cpp PR #29353 cuda/hip GDN chunked kernel, llama-bench prompt-processing tok/s master vs PR, all rows -b 2048 -ub 2048 -fa 1 full GPU offload F16 KV cache, retested 2026-10-13 against master 296538ade; Qwen3.8-Flash-Next UD-IQ4_XS at 4096 prompt tokens on Spark GB10, change +5.2%; config=pr",
   "params": null,
   "power_w": null,
   "pp_tps": 1124.01,
   "provenance": "sourced",
   "quant": "UD-IQ4_XS",
   "quote": "Qwen3.8-Flash-Next; UD-IQ4_XS; GB10; Spark GB10 | Qwen3.8-Flash-Next-UD-IQ4_XS (`-lzm on`) | 4096 | 1,068.70 | 1,124.01",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "llama.cpp PR #29353",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29353",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 2048,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-29353-rtx5090-qwen3535ba3b-q4ks-2048-master",
   "model": "Qwen3.5-35B-A3B",
   "notes": "llama.cpp PR #29353 cuda/hip GDN chunked kernel, llama-bench prompt-processing tok/s master vs PR, all rows -b 2048 -ub 2048 -fa 1 full GPU offload F16 KV cache, retested 2026-10-13 against master 296538ade; Qwen3.5-35B-A3B Q4_K_S at 2048 prompt tokens on RTX 5090, change +16.4%; config=master",
   "params": null,
   "power_w": null,
   "pp_tps": 12477.32,
   "provenance": "sourced",
   "quant": "Q4_K_S",
   "quote": "Qwen3.5-35B-A3B; Q4_K_S; RTX 5090; RTX 5090 | Qwen3.5-35B-A3B-Q4_K_S | 2048 | 12,477.32 | 14,524.81",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "llama.cpp PR #29353",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29353",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 2048,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-29353-rtx5090-qwen3535ba3b-q4ks-2048-pr",
   "model": "Qwen3.5-35B-A3B",
   "notes": "llama.cpp PR #29353 cuda/hip GDN chunked kernel, llama-bench prompt-processing tok/s master vs PR, all rows -b 2048 -ub 2048 -fa 1 full GPU offload F16 KV cache, retested 2026-10-13 against master 296538ade; Qwen3.5-35B-A3B Q4_K_S at 2048 prompt tokens on RTX 5090, change +16.4%; config=pr",
   "params": null,
   "power_w": null,
   "pp_tps": 14524.81,
   "provenance": "sourced",
   "quant": "Q4_K_S",
   "quote": "Qwen3.5-35B-A3B; Q4_K_S; RTX 5090; RTX 5090 | Qwen3.5-35B-A3B-Q4_K_S | 2048 | 12,477.32 | 14,524.81",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "llama.cpp PR #29353",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29353",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-29353-rtx5090-qwen3535ba3b-q4ks-4096-master",
   "model": "Qwen3.5-35B-A3B",
   "notes": "llama.cpp PR #29353 cuda/hip GDN chunked kernel, llama-bench prompt-processing tok/s master vs PR, all rows -b 2048 -ub 2048 -fa 1 full GPU offload F16 KV cache, retested 2026-10-13 against master 296538ade; Qwen3.5-35B-A3B Q4_K_S at 4096 prompt tokens on RTX 5090, change +17.9%; config=master",
   "params": null,
   "power_w": null,
   "pp_tps": 12978.57,
   "provenance": "sourced",
   "quant": "Q4_K_S",
   "quote": "Qwen3.5-35B-A3B; Q4_K_S; RTX 5090; RTX 5090 | Qwen3.5-35B-A3B-Q4_K_S | 4096 | 12,978.57 | 15,298.00",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "llama.cpp PR #29353",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29353",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-29353-rtx5090-qwen3535ba3b-q4ks-4096-pr",
   "model": "Qwen3.5-35B-A3B",
   "notes": "llama.cpp PR #29353 cuda/hip GDN chunked kernel, llama-bench prompt-processing tok/s master vs PR, all rows -b 2048 -ub 2048 -fa 1 full GPU offload F16 KV cache, retested 2026-10-13 against master 296538ade; Qwen3.5-35B-A3B Q4_K_S at 4096 prompt tokens on RTX 5090, change +17.9%; config=pr",
   "params": null,
   "power_w": null,
   "pp_tps": 15298.0,
   "provenance": "sourced",
   "quant": "Q4_K_S",
   "quote": "Qwen3.5-35B-A3B; Q4_K_S; RTX 5090; RTX 5090 | Qwen3.5-35B-A3B-Q4_K_S | 4096 | 12,978.57 | 15,298.00",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "llama.cpp PR #29353",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29353",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 2048,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-29353-rtx5090-qwen3827b-q4km-2048-master",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #29353 cuda/hip GDN chunked kernel, llama-bench prompt-processing tok/s master vs PR, all rows -b 2048 -ub 2048 -fa 1 full GPU offload F16 KV cache, retested 2026-10-13 against master 296538ade; Qwen3.8-27B Q4_K_M at 2048 prompt tokens on RTX 5090, change +5.5%; config=master",
   "params": null,
   "power_w": null,
   "pp_tps": 4095.61,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "Qwen3.8-27B; Q4_K_M; RTX 5090; RTX 5090 | Qwen3.8-27B-Q4_K_M | 2048 | 4,095.61 | 4,320.75",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "llama.cpp PR #29353",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29353",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 2048,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-29353-rtx5090-qwen3827b-q4km-2048-pr",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #29353 cuda/hip GDN chunked kernel, llama-bench prompt-processing tok/s master vs PR, all rows -b 2048 -ub 2048 -fa 1 full GPU offload F16 KV cache, retested 2026-10-13 against master 296538ade; Qwen3.8-27B Q4_K_M at 2048 prompt tokens on RTX 5090, change +5.5%; config=pr",
   "params": null,
   "power_w": null,
   "pp_tps": 4320.75,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "Qwen3.8-27B; Q4_K_M; RTX 5090; RTX 5090 | Qwen3.8-27B-Q4_K_M | 2048 | 4,095.61 | 4,320.75",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "llama.cpp PR #29353",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29353",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-29353-rtx5090-qwen3827b-q4km-4096-master",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #29353 cuda/hip GDN chunked kernel, llama-bench prompt-processing tok/s master vs PR, all rows -b 2048 -ub 2048 -fa 1 full GPU offload F16 KV cache, retested 2026-10-13 against master 296538ade; Qwen3.8-27B Q4_K_M at 4096 prompt tokens on RTX 5090, change +5.8%; config=master",
   "params": null,
   "power_w": null,
   "pp_tps": 4114.11,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "Qwen3.8-27B; Q4_K_M; RTX 5090; RTX 5090 | Qwen3.8-27B-Q4_K_M | 4096 | 4,114.11 | 4,353.09",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "llama.cpp PR #29353",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29353",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-29353-rtx5090-qwen3827b-q4km-4096-pr",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #29353 cuda/hip GDN chunked kernel, llama-bench prompt-processing tok/s master vs PR, all rows -b 2048 -ub 2048 -fa 1 full GPU offload F16 KV cache, retested 2026-10-13 against master 296538ade; Qwen3.8-27B Q4_K_M at 4096 prompt tokens on RTX 5090, change +5.8%; config=pr",
   "params": null,
   "power_w": null,
   "pp_tps": 4353.09,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "Qwen3.8-27B; Q4_K_M; RTX 5090; RTX 5090 | Qwen3.8-27B-Q4_K_M | 4096 | 4,114.11 | 4,353.09",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "llama.cpp PR #29353",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29353",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 2048,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc-29353-rtx4090-qwen3535ba3b-q4ks-2048-master",
   "model": "Qwen3.5-35B-A3B",
   "notes": "llama.cpp PR #29353 cuda/hip GDN chunked kernel, llama-bench prompt-processing tok/s master vs PR, all rows -b 2048 -ub 2048 -fa 1 full GPU offload F16 KV cache, retested 2026-10-13 against master 296538ade; Qwen3.5-35B-A3B Q4_K_S at 2048 prompt tokens on RTX 4090, change +10.8%; config=master",
   "params": null,
   "power_w": null,
   "pp_tps": 10872.34,
   "provenance": "sourced",
   "quant": "Q4_K_S",
   "quote": "Qwen3.5-35B-A3B; Q4_K_S; RTX 4090; RTX 4090 | Qwen3.5-35B-A3B-Q4_K_S | 2048 | 10,872.34 | 12,044.25",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "llama.cpp PR #29353",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29353",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 2048,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc-29353-rtx4090-qwen3535ba3b-q4ks-2048-pr",
   "model": "Qwen3.5-35B-A3B",
   "notes": "llama.cpp PR #29353 cuda/hip GDN chunked kernel, llama-bench prompt-processing tok/s master vs PR, all rows -b 2048 -ub 2048 -fa 1 full GPU offload F16 KV cache, retested 2026-10-13 against master 296538ade; Qwen3.5-35B-A3B Q4_K_S at 2048 prompt tokens on RTX 4090, change +10.8%; config=pr",
   "params": null,
   "power_w": null,
   "pp_tps": 12044.25,
   "provenance": "sourced",
   "quant": "Q4_K_S",
   "quote": "Qwen3.5-35B-A3B; Q4_K_S; RTX 4090; RTX 4090 | Qwen3.5-35B-A3B-Q4_K_S | 2048 | 10,872.34 | 12,044.25",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "llama.cpp PR #29353",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29353",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc-29353-rtx4090-qwen3535ba3b-q4ks-4096-master",
   "model": "Qwen3.5-35B-A3B",
   "notes": "llama.cpp PR #29353 cuda/hip GDN chunked kernel, llama-bench prompt-processing tok/s master vs PR, all rows -b 2048 -ub 2048 -fa 1 full GPU offload F16 KV cache, retested 2026-10-13 against master 296538ade; Qwen3.5-35B-A3B Q4_K_S at 4096 prompt tokens on RTX 4090, change +11.1%; config=master",
   "params": null,
   "power_w": null,
   "pp_tps": 11187.34,
   "provenance": "sourced",
   "quant": "Q4_K_S",
   "quote": "Qwen3.5-35B-A3B; Q4_K_S; RTX 4090; RTX 4090 | Qwen3.5-35B-A3B-Q4_K_S | 4096 | 11,187.34 | 12,428.05",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "llama.cpp PR #29353",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29353",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc-29353-rtx4090-qwen3535ba3b-q4ks-4096-pr",
   "model": "Qwen3.5-35B-A3B",
   "notes": "llama.cpp PR #29353 cuda/hip GDN chunked kernel, llama-bench prompt-processing tok/s master vs PR, all rows -b 2048 -ub 2048 -fa 1 full GPU offload F16 KV cache, retested 2026-10-13 against master 296538ade; Qwen3.5-35B-A3B Q4_K_S at 4096 prompt tokens on RTX 4090, change +11.1%; config=pr",
   "params": null,
   "power_w": null,
   "pp_tps": 12428.05,
   "provenance": "sourced",
   "quant": "Q4_K_S",
   "quote": "Qwen3.5-35B-A3B; Q4_K_S; RTX 4090; RTX 4090 | Qwen3.5-35B-A3B-Q4_K_S | 4096 | 11,187.34 | 12,428.05",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "llama.cpp PR #29353",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29353",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 2048,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc-29353-rtx4090-qwen3827b-q4km-2048-master",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #29353 cuda/hip GDN chunked kernel, llama-bench prompt-processing tok/s master vs PR, all rows -b 2048 -ub 2048 -fa 1 full GPU offload F16 KV cache, retested 2026-10-13 against master 296538ade; Qwen3.8-27B Q4_K_M at 2048 prompt tokens on RTX 4090, change +3.1%; config=master",
   "params": null,
   "power_w": null,
   "pp_tps": 3336.69,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "Qwen3.8-27B; Q4_K_M; RTX 4090; RTX 4090 | Qwen3.8-27B-Q4_K_M | 2048 | 3,336.69 | 3,438.65",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "llama.cpp PR #29353",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29353",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 2048,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc-29353-rtx4090-qwen3827b-q4km-2048-pr",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #29353 cuda/hip GDN chunked kernel, llama-bench prompt-processing tok/s master vs PR, all rows -b 2048 -ub 2048 -fa 1 full GPU offload F16 KV cache, retested 2026-10-13 against master 296538ade; Qwen3.8-27B Q4_K_M at 2048 prompt tokens on RTX 4090, change +3.1%; config=pr",
   "params": null,
   "power_w": null,
   "pp_tps": 3438.65,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "Qwen3.8-27B; Q4_K_M; RTX 4090; RTX 4090 | Qwen3.8-27B-Q4_K_M | 2048 | 3,336.69 | 3,438.65",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "llama.cpp PR #29353",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29353",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc-29353-rtx4090-qwen3827b-q4km-4096-master",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #29353 cuda/hip GDN chunked kernel, llama-bench prompt-processing tok/s master vs PR, all rows -b 2048 -ub 2048 -fa 1 full GPU offload F16 KV cache, retested 2026-10-13 against master 296538ade; Qwen3.8-27B Q4_K_M at 4096 prompt tokens on RTX 4090, change +3.2%; config=master",
   "params": null,
   "power_w": null,
   "pp_tps": 3340.9,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "Qwen3.8-27B; Q4_K_M; RTX 4090; RTX 4090 | Qwen3.8-27B-Q4_K_M | 4096 | 3,340.90 | 3,446.88",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "llama.cpp PR #29353",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29353",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc-29353-rtx4090-qwen3827b-q4km-4096-pr",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #29353 cuda/hip GDN chunked kernel, llama-bench prompt-processing tok/s master vs PR, all rows -b 2048 -ub 2048 -fa 1 full GPU offload F16 KV cache, retested 2026-10-13 against master 296538ade; Qwen3.8-27B Q4_K_M at 4096 prompt tokens on RTX 4090, change +3.2%; config=pr",
   "params": null,
   "power_w": null,
   "pp_tps": 3446.88,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "Qwen3.8-27B; Q4_K_M; RTX 4090; RTX 4090 | Qwen3.8-27B-Q4_K_M | 4096 | 3,340.90 | 3,446.88",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "llama.cpp PR #29353",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29353",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 2048,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc-29353-rtx3090-qwen3535ba3b-q4ks-2048-master",
   "model": "Qwen3.5-35B-A3B",
   "notes": "llama.cpp PR #29353 cuda/hip GDN chunked kernel, llama-bench prompt-processing tok/s master vs PR, all rows -b 2048 -ub 2048 -fa 1 full GPU offload F16 KV cache, retested 2026-10-13 against master 296538ade; Qwen3.5-35B-A3B Q4_K_S at 2048 prompt tokens on RTX 3090, change +6.9%; config=master",
   "params": null,
   "power_w": null,
   "pp_tps": 5672.07,
   "provenance": "sourced",
   "quant": "Q4_K_S",
   "quote": "Qwen3.5-35B-A3B; Q4_K_S; RTX 3090; RTX 3090 | Qwen3.5-35B-A3B-Q4_K_S | 2048 | 5,672.07 | 6,061.59",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "llama.cpp PR #29353",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29353",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 2048,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc-29353-rtx3090-qwen3535ba3b-q4ks-2048-pr",
   "model": "Qwen3.5-35B-A3B",
   "notes": "llama.cpp PR #29353 cuda/hip GDN chunked kernel, llama-bench prompt-processing tok/s master vs PR, all rows -b 2048 -ub 2048 -fa 1 full GPU offload F16 KV cache, retested 2026-10-13 against master 296538ade; Qwen3.5-35B-A3B Q4_K_S at 2048 prompt tokens on RTX 3090, change +6.9%; config=pr",
   "params": null,
   "power_w": null,
   "pp_tps": 6061.59,
   "provenance": "sourced",
   "quant": "Q4_K_S",
   "quote": "Qwen3.5-35B-A3B; Q4_K_S; RTX 3090; RTX 3090 | Qwen3.5-35B-A3B-Q4_K_S | 2048 | 5,672.07 | 6,061.59",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "llama.cpp PR #29353",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29353",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc-29353-rtx3090-qwen3535ba3b-q4ks-4096-master",
   "model": "Qwen3.5-35B-A3B",
   "notes": "llama.cpp PR #29353 cuda/hip GDN chunked kernel, llama-bench prompt-processing tok/s master vs PR, all rows -b 2048 -ub 2048 -fa 1 full GPU offload F16 KV cache, retested 2026-10-13 against master 296538ade; Qwen3.5-35B-A3B Q4_K_S at 4096 prompt tokens on RTX 3090, change +3.2%; config=master",
   "params": null,
   "power_w": null,
   "pp_tps": 5505.49,
   "provenance": "sourced",
   "quant": "Q4_K_S",
   "quote": "Qwen3.5-35B-A3B; Q4_K_S; RTX 3090; RTX 3090 | Qwen3.5-35B-A3B-Q4_K_S | 4096 | 5,505.49 | 5,683.79",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "llama.cpp PR #29353",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29353",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc-29353-rtx3090-qwen3535ba3b-q4ks-4096-pr",
   "model": "Qwen3.5-35B-A3B",
   "notes": "llama.cpp PR #29353 cuda/hip GDN chunked kernel, llama-bench prompt-processing tok/s master vs PR, all rows -b 2048 -ub 2048 -fa 1 full GPU offload F16 KV cache, retested 2026-10-13 against master 296538ade; Qwen3.5-35B-A3B Q4_K_S at 4096 prompt tokens on RTX 3090, change +3.2%; config=pr",
   "params": null,
   "power_w": null,
   "pp_tps": 5683.79,
   "provenance": "sourced",
   "quant": "Q4_K_S",
   "quote": "Qwen3.5-35B-A3B; Q4_K_S; RTX 3090; RTX 3090 | Qwen3.5-35B-A3B-Q4_K_S | 4096 | 5,505.49 | 5,683.79",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "llama.cpp PR #29353",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29353",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 2048,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc-29353-rtx3090-qwen3827b-q4km-2048-master",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #29353 cuda/hip GDN chunked kernel, llama-bench prompt-processing tok/s master vs PR, all rows -b 2048 -ub 2048 -fa 1 full GPU offload F16 KV cache, retested 2026-10-13 against master 296538ade; Qwen3.8-27B Q4_K_M at 2048 prompt tokens on RTX 3090, change +0.9%; config=master",
   "params": null,
   "power_w": null,
   "pp_tps": 1400.46,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "Qwen3.8-27B; Q4_K_M; RTX 3090; RTX 3090 | Qwen3.8-27B-Q4_K_M | 2048 | 1,400.46 | 1,412.40",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "llama.cpp PR #29353",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29353",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 2048,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc-29353-rtx3090-qwen3827b-q4km-2048-pr",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #29353 cuda/hip GDN chunked kernel, llama-bench prompt-processing tok/s master vs PR, all rows -b 2048 -ub 2048 -fa 1 full GPU offload F16 KV cache, retested 2026-10-13 against master 296538ade; Qwen3.8-27B Q4_K_M at 2048 prompt tokens on RTX 3090, change +0.9%; config=pr",
   "params": null,
   "power_w": null,
   "pp_tps": 1412.4,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "Qwen3.8-27B; Q4_K_M; RTX 3090; RTX 3090 | Qwen3.8-27B-Q4_K_M | 2048 | 1,400.46 | 1,412.40",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "llama.cpp PR #29353",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29353",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc-29353-rtx3090-qwen3827b-q4km-4096-master",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #29353 cuda/hip GDN chunked kernel, llama-bench prompt-processing tok/s master vs PR, all rows -b 2048 -ub 2048 -fa 1 full GPU offload F16 KV cache, retested 2026-10-13 against master 296538ade; Qwen3.8-27B Q4_K_M at 4096 prompt tokens on RTX 3090, change +2.2%; config=master",
   "params": null,
   "power_w": null,
   "pp_tps": 1337.41,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "Qwen3.8-27B; Q4_K_M; RTX 3090; RTX 3090 | Qwen3.8-27B-Q4_K_M | 4096 | 1,337.41 | 1,367.46",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "llama.cpp PR #29353",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29353",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc-29353-rtx3090-qwen3827b-q4km-4096-pr",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #29353 cuda/hip GDN chunked kernel, llama-bench prompt-processing tok/s master vs PR, all rows -b 2048 -ub 2048 -fa 1 full GPU offload F16 KV cache, retested 2026-10-13 against master 296538ade; Qwen3.8-27B Q4_K_M at 4096 prompt tokens on RTX 3090, change +2.2%; config=pr",
   "params": null,
   "power_w": null,
   "pp_tps": 1367.46,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "Qwen3.8-27B; Q4_K_M; RTX 3090; RTX 3090 | Qwen3.8-27B-Q4_K_M | 4096 | 1,337.41 | 1,367.46",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "llama.cpp PR #29353",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/29353",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-30544-radeon8060s-mistral119b-pp512-pre29029",
   "model": "Mistral-Small-4-119B-2603",
   "notes": "llama.cpp PR #30544 vulkan keep F32 MUL_MAT_ID on GPU for mat-vec skip MMVQ, llama-bench -ngl 99 -fa 1 -p 512 -n 32 on AMD Ryzen AI Max Strix Halo Radeon 8060S gfx1151 RADV Mesa 26.0.8 Linux 7.0 uma 1 int dot 1, unsloth/Mistral-Small-4-119B-2603 UD-Q4_K_XL; pp512 on build fb92d8f18 before #29029; config=pre-29029",
   "params": null,
   "power_w": null,
   "pp_tps": 334,
   "provenance": "sourced",
   "quant": "UD-Q4_K_XL",
   "quote": "Mistral-Small-4-119B-2603; UD-Q4_K_XL; Radeon 8060S; fb92d8f (before #29029) | 334 | 41.2",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "llama.cpp PR #30544",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30544",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 32,
   "ctx": null,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-30544-radeon8060s-mistral119b-tg32-pre29029",
   "model": "Mistral-Small-4-119B-2603",
   "notes": "llama.cpp PR #30544 vulkan keep F32 MUL_MAT_ID on GPU for mat-vec skip MMVQ, llama-bench -ngl 99 -fa 1 -p 512 -n 32 on AMD Ryzen AI Max Strix Halo Radeon 8060S gfx1151 RADV Mesa 26.0.8 Linux 7.0 uma 1 int dot 1, unsloth/Mistral-Small-4-119B-2603 UD-Q4_K_XL; tg32 on build fb92d8f18 before #29029; config=pre-29029",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "UD-Q4_K_XL",
   "quote": "Mistral-Small-4-119B-2603; UD-Q4_K_XL; Radeon 8060S; fb92d8f (before #29029) | 334 | 41.2",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "llama.cpp PR #30544",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30544",
   "tg_tokens": null,
   "tps": 41.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-30544-radeon8060s-mistral119b-pp512-master",
   "model": "Mistral-Small-4-119B-2603",
   "notes": "llama.cpp PR #30544 vulkan keep F32 MUL_MAT_ID on GPU for mat-vec skip MMVQ, llama-bench -ngl 99 -fa 1 -p 512 -n 32 on AMD Ryzen AI Max Strix Halo Radeon 8060S gfx1151 RADV Mesa 26.0.8 Linux 7.0 uma 1 int dot 1, unsloth/Mistral-Small-4-119B-2603 UD-Q4_K_XL; pp512 on master 5cdadfc32, degraded by #29029 MMVQ path; config=master",
   "params": null,
   "power_w": null,
   "pp_tps": 319,
   "provenance": "sourced",
   "quant": "UD-Q4_K_XL",
   "quote": "Mistral-Small-4-119B-2603; UD-Q4_K_XL; Radeon 8060S; 5cdadfc (master) | 319 | 3.4",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "llama.cpp PR #30544",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30544",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 32,
   "ctx": null,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-30544-radeon8060s-mistral119b-tg32-master",
   "model": "Mistral-Small-4-119B-2603",
   "notes": "llama.cpp PR #30544 vulkan keep F32 MUL_MAT_ID on GPU for mat-vec skip MMVQ, llama-bench -ngl 99 -fa 1 -p 512 -n 32 on AMD Ryzen AI Max Strix Halo Radeon 8060S gfx1151 RADV Mesa 26.0.8 Linux 7.0 uma 1 int dot 1, unsloth/Mistral-Small-4-119B-2603 UD-Q4_K_XL; tg32 on master 5cdadfc32, degraded by #29029 MMVQ path; config=master",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "UD-Q4_K_XL",
   "quote": "Mistral-Small-4-119B-2603; UD-Q4_K_XL; Radeon 8060S; 5cdadfc (master) | 319 | 3.4",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "llama.cpp PR #30544",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30544",
   "tg_tokens": null,
   "tps": 3.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": 32,
   "ctx": null,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-30544-radeon8060s-mistral119b-tg32-pr",
   "model": "Mistral-Small-4-119B-2603",
   "notes": "llama.cpp PR #30544 vulkan keep F32 MUL_MAT_ID on GPU for mat-vec skip MMVQ, llama-bench -ngl 99 -fa 1 -p 512 -n 32 on AMD Ryzen AI Max Strix Halo Radeon 8060S gfx1151 RADV Mesa 26.0.8 Linux 7.0 uma 1 int dot 1, unsloth/Mistral-Small-4-119B-2603 UD-Q4_K_XL; tg32 with this PR applied via llama-server, same model; pp512 not measured for this arm; config=pr",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "UD-Q4_K_XL",
   "quote": "Mistral-Small-4-119B-2603; UD-Q4_K_XL; Radeon 8060S; this PR (llama-server, same model) | - | 39.5",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "llama.cpp PR #30544",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30544",
   "tg_tokens": null,
   "tps": 39.5,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 32,
   "ctx": null,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "A100-SXM4-40GB",
   "id": "vllm-61728-a100sxm4-qwen4exp-qsa-dcp-59279",
   "model": "Qwen4Exp",
   "notes": "vLLM PR #61728 Qwen4Exp QSA DCP follow-ups to #59279 (copy-free LSE combine fused gate, MTP verify rows on decode path), 8x A100-SXM4-40GB TP8+EP DCP4 FP8 KV MTP3 max-num-batched-tokens 2048, MLPerf agentic multi-turn workload 32 requests in flight 1x60 min per arm 0 failed; output tok/s per user; #59279 plus its block-table fix arm, 2.72 req/s, 140.8 output tok/s per GPU, TTFT P50/P95 990/2137 ms, TPOT P50/P95 18.8/29.3 ms, SLO pass 68.8%; config=59279",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "Qwen4Exp; A100-SXM4-40GB; #59279 (+ its block-table fix) | 2.72 | 140.8 | 42.5 | 990 / 2,137 ms | 18.8 / 29.3 ms | 68.8 %",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "vLLM PR #61728",
   "source_url": "https://github.com/vllm-project/vllm/pull/61728",
   "tg_tokens": null,
   "tps": 42.5,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 32,
   "ctx": null,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "A100-SXM4-40GB",
   "id": "vllm-61728-a100sxm4-qwen4exp-qsa-dcp-pr",
   "model": "Qwen4Exp",
   "notes": "vLLM PR #61728 Qwen4Exp QSA DCP follow-ups to #59279 (copy-free LSE combine fused gate, MTP verify rows on decode path), 8x A100-SXM4-40GB TP8+EP DCP4 FP8 KV MTP3 max-num-batched-tokens 2048, MLPerf agentic multi-turn workload 32 requests in flight 1x60 min per arm 0 failed; output tok/s per user; this PR arm, 2.80 req/s, 149.8 output tok/s per GPU, TTFT P50/P95 967/2233 ms, TPOT P50/P95 17.4/27.8 ms, SLO pass 76.2%; config=pr",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "Qwen4Exp; A100-SXM4-40GB; This PR | 2.80 | 149.8 | 45.3 | 967 / 2,233 ms | 17.4 / 27.8 ms | 76.2 %",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "vLLM PR #61728",
   "source_url": "https://github.com/vllm-project/vllm/pull/61728",
   "tg_tokens": null,
   "tps": 45.3,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-15",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "GB300",
   "id": "vllm-61868-int-c1-off",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #61868 low-latency decode paths; upstream main arm (low-latency decode switch off); P90 interactivity tok/s per user at concurrency 1; TP4 one GB300 tray; DSpark speculative decoding 5 draft tokens; Model Runner V2, FULL CUDA graphs decode; AgentX agentic-coding traces 3600 s per concurrency, median 350K token prefix-cached conversations",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "DeepSeek-V4.1-Flash; GB300; 1 655.2 \u2192 811.3",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "vLLM PR #61868",
   "source_url": "https://github.com/vllm-project/vllm/pull/61868",
   "tg_tokens": null,
   "tps": 655.2,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-15",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "GB300",
   "id": "vllm-61868-int-c1-on",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #61868 low-latency decode paths; this PR arm (low-latency decode switch on); P90 interactivity tok/s per user at concurrency 1; TP4 one GB300 tray; DSpark speculative decoding 5 draft tokens; Model Runner V2, FULL CUDA graphs decode; AgentX agentic-coding traces 3600 s per concurrency, median 350K token prefix-cached conversations",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "DeepSeek-V4.1-Flash; GB300; 1 655.2 \u2192 811.3",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "vLLM PR #61868",
   "source_url": "https://github.com/vllm-project/vllm/pull/61868",
   "tg_tokens": null,
   "tps": 811.3,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-61868-gpu-c1-off",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #61868 low-latency decode paths; upstream main arm (low-latency decode switch off); total tok/s per GPU at concurrency 1 (+6.2% change off to on); TP4 one GB300 tray; DSpark speculative decoding 5 draft tokens; Model Runner V2, FULL CUDA graphs decode; AgentX agentic-coding traces 3600 s per concurrency, median 350K token prefix-cached conversations",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "DeepSeek-V4.1-Flash; GB300; 7,402 \u2192 7,861 (+6.2%)",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "vLLM PR #61868",
   "source_url": "https://github.com/vllm-project/vllm/pull/61868",
   "tg_tokens": null,
   "tps": 7402,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": null,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-61868-gpu-c1-on",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #61868 low-latency decode paths; this PR arm (low-latency decode switch on); total tok/s per GPU at concurrency 1 (+6.2% change off to on); TP4 one GB300 tray; DSpark speculative decoding 5 draft tokens; Model Runner V2, FULL CUDA graphs decode; AgentX agentic-coding traces 3600 s per concurrency, median 350K token prefix-cached conversations",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "DeepSeek-V4.1-Flash; GB300; 7,402 \u2192 7,861 (+6.2%)",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "vLLM PR #61868",
   "source_url": "https://github.com/vllm-project/vllm/pull/61868",
   "tg_tokens": null,
   "tps": 7861,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": null,
   "date": "2026-10-15",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "GB300",
   "id": "vllm-61868-int-c4-off",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #61868 low-latency decode paths; upstream main arm (low-latency decode switch off); P90 interactivity tok/s per user at concurrency 4; TP4 one GB300 tray; DSpark speculative decoding 5 draft tokens; Model Runner V2, FULL CUDA graphs decode; AgentX agentic-coding traces 3600 s per concurrency, median 350K token prefix-cached conversations",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "DeepSeek-V4.1-Flash; GB300; 4 557.7 \u2192 681.7",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "vLLM PR #61868",
   "source_url": "https://github.com/vllm-project/vllm/pull/61868",
   "tg_tokens": null,
   "tps": 557.7,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": null,
   "date": "2026-10-15",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "GB300",
   "id": "vllm-61868-int-c4-on",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #61868 low-latency decode paths; this PR arm (low-latency decode switch on); P90 interactivity tok/s per user at concurrency 4; TP4 one GB300 tray; DSpark speculative decoding 5 draft tokens; Model Runner V2, FULL CUDA graphs decode; AgentX agentic-coding traces 3600 s per concurrency, median 350K token prefix-cached conversations",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "DeepSeek-V4.1-Flash; GB300; 4 557.7 \u2192 681.7",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "vLLM PR #61868",
   "source_url": "https://github.com/vllm-project/vllm/pull/61868",
   "tg_tokens": null,
   "tps": 681.7,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": null,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-61868-gpu-c4-off",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #61868 low-latency decode paths; upstream main arm (low-latency decode switch off); total tok/s per GPU at concurrency 4 (+4.9% change off to on); TP4 one GB300 tray; DSpark speculative decoding 5 draft tokens; Model Runner V2, FULL CUDA graphs decode; AgentX agentic-coding traces 3600 s per concurrency, median 350K token prefix-cached conversations",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "DeepSeek-V4.1-Flash; GB300; 9,590 \u2192 10,055 (+4.9%)",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "vLLM PR #61868",
   "source_url": "https://github.com/vllm-project/vllm/pull/61868",
   "tg_tokens": null,
   "tps": 9590,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": null,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-61868-gpu-c4-on",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #61868 low-latency decode paths; this PR arm (low-latency decode switch on); total tok/s per GPU at concurrency 4 (+4.9% change off to on); TP4 one GB300 tray; DSpark speculative decoding 5 draft tokens; Model Runner V2, FULL CUDA graphs decode; AgentX agentic-coding traces 3600 s per concurrency, median 350K token prefix-cached conversations",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "DeepSeek-V4.1-Flash; GB300; 9,590 \u2192 10,055 (+4.9%)",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "vLLM PR #61868",
   "source_url": "https://github.com/vllm-project/vllm/pull/61868",
   "tg_tokens": null,
   "tps": 10055,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": null,
   "date": "2026-10-15",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "GB300",
   "id": "vllm-61868-int-c8-off",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #61868 low-latency decode paths; upstream main arm (low-latency decode switch off); P90 interactivity tok/s per user at concurrency 8; TP4 one GB300 tray; DSpark speculative decoding 5 draft tokens; Model Runner V2, FULL CUDA graphs decode; AgentX agentic-coding traces 3600 s per concurrency, median 350K token prefix-cached conversations",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "DeepSeek-V4.1-Flash; GB300; 8 478.3 \u2192 560.9",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "vLLM PR #61868",
   "source_url": "https://github.com/vllm-project/vllm/pull/61868",
   "tg_tokens": null,
   "tps": 478.3,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": null,
   "date": "2026-10-15",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "GB300",
   "id": "vllm-61868-int-c8-on",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #61868 low-latency decode paths; this PR arm (low-latency decode switch on); P90 interactivity tok/s per user at concurrency 8; TP4 one GB300 tray; DSpark speculative decoding 5 draft tokens; Model Runner V2, FULL CUDA graphs decode; AgentX agentic-coding traces 3600 s per concurrency, median 350K token prefix-cached conversations",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "DeepSeek-V4.1-Flash; GB300; 8 478.3 \u2192 560.9",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "vLLM PR #61868",
   "source_url": "https://github.com/vllm-project/vllm/pull/61868",
   "tg_tokens": null,
   "tps": 560.9,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": null,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-61868-gpu-c8-off",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #61868 low-latency decode paths; upstream main arm (low-latency decode switch off); total tok/s per GPU at concurrency 8 (+3.3% change off to on); TP4 one GB300 tray; DSpark speculative decoding 5 draft tokens; Model Runner V2, FULL CUDA graphs decode; AgentX agentic-coding traces 3600 s per concurrency, median 350K token prefix-cached conversations",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "DeepSeek-V4.1-Flash; GB300; 16,443 \u2192 16,985 (+3.3%)",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "vLLM PR #61868",
   "source_url": "https://github.com/vllm-project/vllm/pull/61868",
   "tg_tokens": null,
   "tps": 16443,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 8,
   "ctx": null,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-61868-gpu-c8-on",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #61868 low-latency decode paths; this PR arm (low-latency decode switch on); total tok/s per GPU at concurrency 8 (+3.3% change off to on); TP4 one GB300 tray; DSpark speculative decoding 5 draft tokens; Model Runner V2, FULL CUDA graphs decode; AgentX agentic-coding traces 3600 s per concurrency, median 350K token prefix-cached conversations",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "DeepSeek-V4.1-Flash; GB300; 16,443 \u2192 16,985 (+3.3%)",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "vLLM PR #61868",
   "source_url": "https://github.com/vllm-project/vllm/pull/61868",
   "tg_tokens": null,
   "tps": 16985,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": null,
   "date": "2026-10-15",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "GB300",
   "id": "vllm-61868-int-c16-off",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #61868 low-latency decode paths; upstream main arm (low-latency decode switch off); P90 interactivity tok/s per user at concurrency 16; TP4 one GB300 tray; DSpark speculative decoding 5 draft tokens; Model Runner V2, FULL CUDA graphs decode; AgentX agentic-coding traces 3600 s per concurrency, median 350K token prefix-cached conversations",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "DeepSeek-V4.1-Flash; GB300; 16 365.5 \u2192 452.6",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "vLLM PR #61868",
   "source_url": "https://github.com/vllm-project/vllm/pull/61868",
   "tg_tokens": null,
   "tps": 365.5,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": null,
   "date": "2026-10-15",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "GB300",
   "id": "vllm-61868-int-c16-on",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #61868 low-latency decode paths; this PR arm (low-latency decode switch on); P90 interactivity tok/s per user at concurrency 16; TP4 one GB300 tray; DSpark speculative decoding 5 draft tokens; Model Runner V2, FULL CUDA graphs decode; AgentX agentic-coding traces 3600 s per concurrency, median 350K token prefix-cached conversations",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "DeepSeek-V4.1-Flash; GB300; 16 365.5 \u2192 452.6",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "vLLM PR #61868",
   "source_url": "https://github.com/vllm-project/vllm/pull/61868",
   "tg_tokens": null,
   "tps": 452.6,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": null,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-61868-gpu-c16-off",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #61868 low-latency decode paths; upstream main arm (low-latency decode switch off); total tok/s per GPU at concurrency 16 (+1.0% change off to on); TP4 one GB300 tray; DSpark speculative decoding 5 draft tokens; Model Runner V2, FULL CUDA graphs decode; AgentX agentic-coding traces 3600 s per concurrency, median 350K token prefix-cached conversations",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "DeepSeek-V4.1-Flash; GB300; 29,444 \u2192 29,727 (+1.0%)",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "vLLM PR #61868",
   "source_url": "https://github.com/vllm-project/vllm/pull/61868",
   "tg_tokens": null,
   "tps": 29444,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 16,
   "ctx": null,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-61868-gpu-c16-on",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #61868 low-latency decode paths; this PR arm (low-latency decode switch on); total tok/s per GPU at concurrency 16 (+1.0% change off to on); TP4 one GB300 tray; DSpark speculative decoding 5 draft tokens; Model Runner V2, FULL CUDA graphs decode; AgentX agentic-coding traces 3600 s per concurrency, median 350K token prefix-cached conversations",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "DeepSeek-V4.1-Flash; GB300; 29,444 \u2192 29,727 (+1.0%)",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "vLLM PR #61868",
   "source_url": "https://github.com/vllm-project/vllm/pull/61868",
   "tg_tokens": null,
   "tps": 29727,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 32,
   "ctx": null,
   "date": "2026-10-15",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "GB300",
   "id": "vllm-61868-int-c32-off",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #61868 low-latency decode paths; upstream main arm (low-latency decode switch off); P90 interactivity tok/s per user at concurrency 32; TP4 one GB300 tray; DSpark speculative decoding 5 draft tokens; Model Runner V2, FULL CUDA graphs decode; AgentX agentic-coding traces 3600 s per concurrency, median 350K token prefix-cached conversations",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "DeepSeek-V4.1-Flash; GB300; 32 256.8 \u2192 297.0",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "vLLM PR #61868",
   "source_url": "https://github.com/vllm-project/vllm/pull/61868",
   "tg_tokens": null,
   "tps": 256.8,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 32,
   "ctx": null,
   "date": "2026-10-15",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "GB300",
   "id": "vllm-61868-int-c32-on",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #61868 low-latency decode paths; this PR arm (low-latency decode switch on); P90 interactivity tok/s per user at concurrency 32; TP4 one GB300 tray; DSpark speculative decoding 5 draft tokens; Model Runner V2, FULL CUDA graphs decode; AgentX agentic-coding traces 3600 s per concurrency, median 350K token prefix-cached conversations",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "DeepSeek-V4.1-Flash; GB300; 32 256.8 \u2192 297.0",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "vLLM PR #61868",
   "source_url": "https://github.com/vllm-project/vllm/pull/61868",
   "tg_tokens": null,
   "tps": 297.0,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 32,
   "ctx": null,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-61868-gpu-c32-off",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #61868 low-latency decode paths; upstream main arm (low-latency decode switch off); total tok/s per GPU at concurrency 32 (+2.0% change off to on); TP4 one GB300 tray; DSpark speculative decoding 5 draft tokens; Model Runner V2, FULL CUDA graphs decode; AgentX agentic-coding traces 3600 s per concurrency, median 350K token prefix-cached conversations",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "DeepSeek-V4.1-Flash; GB300; 61,732 \u2192 62,969 (+2.0%)",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "vLLM PR #61868",
   "source_url": "https://github.com/vllm-project/vllm/pull/61868",
   "tg_tokens": null,
   "tps": 61732,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 32,
   "ctx": null,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-61868-gpu-c32-on",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #61868 low-latency decode paths; this PR arm (low-latency decode switch on); total tok/s per GPU at concurrency 32 (+2.0% change off to on); TP4 one GB300 tray; DSpark speculative decoding 5 draft tokens; Model Runner V2, FULL CUDA graphs decode; AgentX agentic-coding traces 3600 s per concurrency, median 350K token prefix-cached conversations",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "DeepSeek-V4.1-Flash; GB300; 61,732 \u2192 62,969 (+2.0%)",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "vLLM PR #61868",
   "source_url": "https://github.com/vllm-project/vllm/pull/61868",
   "tg_tokens": null,
   "tps": 62969,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 64,
   "ctx": null,
   "date": "2026-10-15",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "GB300",
   "id": "vllm-61868-int-c64-off",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #61868 low-latency decode paths; upstream main arm (low-latency decode switch off); P90 interactivity tok/s per user at concurrency 64; TP4 one GB300 tray; DSpark speculative decoding 5 draft tokens; Model Runner V2, FULL CUDA graphs decode; AgentX agentic-coding traces 3600 s per concurrency, median 350K token prefix-cached conversations",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "DeepSeek-V4.1-Flash; GB300; 64 175.8 \u2192 211.5",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "vLLM PR #61868",
   "source_url": "https://github.com/vllm-project/vllm/pull/61868",
   "tg_tokens": null,
   "tps": 175.8,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 64,
   "ctx": null,
   "date": "2026-10-15",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "GB300",
   "id": "vllm-61868-int-c64-on",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #61868 low-latency decode paths; this PR arm (low-latency decode switch on); P90 interactivity tok/s per user at concurrency 64; TP4 one GB300 tray; DSpark speculative decoding 5 draft tokens; Model Runner V2, FULL CUDA graphs decode; AgentX agentic-coding traces 3600 s per concurrency, median 350K token prefix-cached conversations",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "DeepSeek-V4.1-Flash; GB300; 64 175.8 \u2192 211.5",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "vLLM PR #61868",
   "source_url": "https://github.com/vllm-project/vllm/pull/61868",
   "tg_tokens": null,
   "tps": 211.5,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 64,
   "ctx": null,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-61868-gpu-c64-off",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #61868 low-latency decode paths; upstream main arm (low-latency decode switch off); total tok/s per GPU at concurrency 64 (+1.2% change off to on); TP4 one GB300 tray; DSpark speculative decoding 5 draft tokens; Model Runner V2, FULL CUDA graphs decode; AgentX agentic-coding traces 3600 s per concurrency, median 350K token prefix-cached conversations",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "DeepSeek-V4.1-Flash; GB300; 100,638 \u2192 101,811 (+1.2%)",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "vLLM PR #61868",
   "source_url": "https://github.com/vllm-project/vllm/pull/61868",
   "tg_tokens": null,
   "tps": 100638,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 64,
   "ctx": null,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "GB300",
   "id": "vllm-61868-gpu-c64-on",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "vLLM PR #61868 low-latency decode paths; this PR arm (low-latency decode switch on); total tok/s per GPU at concurrency 64 (+1.2% change off to on); TP4 one GB300 tray; DSpark speculative decoding 5 draft tokens; Model Runner V2, FULL CUDA graphs decode; AgentX agentic-coding traces 3600 s per concurrency, median 350K token prefix-cached conversations",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "DeepSeek-V4.1-Flash; GB300; 100,638 \u2192 101,811 (+1.2%)",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "vLLM PR #61868",
   "source_url": "https://github.com/vllm-project/vllm/pull/61868",
   "tg_tokens": null,
   "tps": 101811,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 128,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "4x H200",
   "id": "vllm-61875-ep-decode-bf16",
   "model": "DeepSeek-V4-Flash",
   "notes": "vLLM PR #61875 Humming MoE backend on H200 SM90; BF16 MoE activations arm (main); ep workload decode, max-model-len 16384; median of 3 runs",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "DeepSeek-V4-Flash; 4x H200; TP4 + EP decode: 512 in / 1024 out, conc 128 (3 runs) 4757 / 4911 / 4906 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "vLLM PR #61875",
   "source_url": "https://github.com/vllm-project/vllm/pull/61875",
   "tg_tokens": null,
   "tps": 4906,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 128,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "4x H200",
   "id": "vllm-61875-ep-decode-fp8",
   "model": "DeepSeek-V4-Flash",
   "notes": "vLLM PR #61875 Humming MoE backend on H200 SM90; block FP8 GS128 MoE activations arm (VLLM_HUMMING_INPUT_QUANT_CONFIG float8e4m3 group 128, resolves to this PR default); ep workload decode, max-model-len 16384; median of 3 runs",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "DeepSeek-V4-Flash; 4x H200; 4757 / 4911 / 4906 tok/s 6131 / 6135 / 6128 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "vLLM PR #61875",
   "source_url": "https://github.com/vllm-project/vllm/pull/61875",
   "tg_tokens": null,
   "tps": 6131,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 32,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "4x H200",
   "id": "vllm-61875-ep-prefill-bf16",
   "model": "DeepSeek-V4-Flash",
   "notes": "vLLM PR #61875 Humming MoE backend on H200 SM90; BF16 MoE activations arm (main); ep workload prefill, max-model-len 16384; total tok/s",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "DeepSeek-V4-Flash; 4x H200; TP4 + EP prefill: 8192 in / 64 out, conc 32 28939 tok/s total",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "vLLM PR #61875",
   "source_url": "https://github.com/vllm-project/vllm/pull/61875",
   "tg_tokens": null,
   "tps": 28939,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 32,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "4x H200",
   "id": "vllm-61875-ep-prefill-fp8",
   "model": "DeepSeek-V4-Flash",
   "notes": "vLLM PR #61875 Humming MoE backend on H200 SM90; block FP8 GS128 MoE activations arm (VLLM_HUMMING_INPUT_QUANT_CONFIG float8e4m3 group 128, resolves to this PR default); ep workload prefill, max-model-len 16384; total tok/s",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "DeepSeek-V4-Flash; 4x H200; 28939 tok/s total 30780 tok/s total",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "vLLM PR #61875",
   "source_url": "https://github.com/vllm-project/vllm/pull/61875",
   "tg_tokens": null,
   "tps": 30780,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "4x H200",
   "id": "vllm-61875-ep-low-bf16",
   "model": "DeepSeek-V4-Flash",
   "notes": "vLLM PR #61875 Humming MoE backend on H200 SM90; BF16 MoE activations arm (main); ep workload low, max-model-len 16384",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "DeepSeek-V4-Flash; 4x H200; TP4 + EP low concurrency: 2048 in / 512 out, conc 4 422.6 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "vLLM PR #61875",
   "source_url": "https://github.com/vllm-project/vllm/pull/61875",
   "tg_tokens": null,
   "tps": 422.6,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "4x H200",
   "id": "vllm-61875-ep-low-fp8",
   "model": "DeepSeek-V4-Flash",
   "notes": "vLLM PR #61875 Humming MoE backend on H200 SM90; block FP8 GS128 MoE activations arm (VLLM_HUMMING_INPUT_QUANT_CONFIG float8e4m3 group 128, resolves to this PR default); ep workload low, max-model-len 16384",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "DeepSeek-V4-Flash; 4x H200; 422.6 tok/s 421.0 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "vLLM PR #61875",
   "source_url": "https://github.com/vllm-project/vllm/pull/61875",
   "tg_tokens": null,
   "tps": 421.0,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 128,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "4x H200",
   "id": "vllm-61875-noep-decode-bf16",
   "model": "DeepSeek-V4-Flash",
   "notes": "vLLM PR #61875 Humming MoE backend on H200 SM90; BF16 MoE activations arm (main); noep workload decode, max-model-len 16384; 1 run",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "DeepSeek-V4-Flash; 4x H200; TP4 (no EP) decode (1 run) 5249 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "vLLM PR #61875",
   "source_url": "https://github.com/vllm-project/vllm/pull/61875",
   "tg_tokens": null,
   "tps": 5249,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 128,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "4x H200",
   "id": "vllm-61875-noep-decode-fp8",
   "model": "DeepSeek-V4-Flash",
   "notes": "vLLM PR #61875 Humming MoE backend on H200 SM90; block FP8 GS128 MoE activations arm (VLLM_HUMMING_INPUT_QUANT_CONFIG float8e4m3 group 128, resolves to this PR default); noep workload decode, max-model-len 16384; 1 run",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "DeepSeek-V4-Flash; 4x H200; 5249 tok/s 4473 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "vLLM PR #61875",
   "source_url": "https://github.com/vllm-project/vllm/pull/61875",
   "tg_tokens": null,
   "tps": 4473,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 32,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "4x H200",
   "id": "vllm-61875-noep-prefill-bf16",
   "model": "DeepSeek-V4-Flash",
   "notes": "vLLM PR #61875 Humming MoE backend on H200 SM90; BF16 MoE activations arm (main); noep workload prefill, max-model-len 16384; total tok/s",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "DeepSeek-V4-Flash; 4x H200; TP4 (no EP) prefill 29873 tok/s total",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "vLLM PR #61875",
   "source_url": "https://github.com/vllm-project/vllm/pull/61875",
   "tg_tokens": null,
   "tps": 29873,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 32,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "4x H200",
   "id": "vllm-61875-noep-prefill-fp8",
   "model": "DeepSeek-V4-Flash",
   "notes": "vLLM PR #61875 Humming MoE backend on H200 SM90; block FP8 GS128 MoE activations arm (VLLM_HUMMING_INPUT_QUANT_CONFIG float8e4m3 group 128, resolves to this PR default); noep workload prefill, max-model-len 16384; total tok/s",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "DeepSeek-V4-Flash; 4x H200; 29873 tok/s total 30309 tok/s total",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "vLLM PR #61875",
   "source_url": "https://github.com/vllm-project/vllm/pull/61875",
   "tg_tokens": null,
   "tps": 30309,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "4x H200",
   "id": "vllm-61875-noep-low-bf16",
   "model": "DeepSeek-V4-Flash",
   "notes": "vLLM PR #61875 Humming MoE backend on H200 SM90; BF16 MoE activations arm (main); noep workload low, max-model-len 16384",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "DeepSeek-V4-Flash; 4x H200; TP4 (no EP) low concurrency 485.9 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "vLLM PR #61875",
   "source_url": "https://github.com/vllm-project/vllm/pull/61875",
   "tg_tokens": null,
   "tps": 485.9,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "4x H200",
   "id": "vllm-61875-noep-low-fp8",
   "model": "DeepSeek-V4-Flash",
   "notes": "vLLM PR #61875 Humming MoE backend on H200 SM90; block FP8 GS128 MoE activations arm (VLLM_HUMMING_INPUT_QUANT_CONFIG float8e4m3 group 128, resolves to this PR default); noep workload low, max-model-len 16384",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "DeepSeek-V4-Flash; 4x H200; 485.9 tok/s 455.9 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "vLLM PR #61875",
   "source_url": "https://github.com/vllm-project/vllm/pull/61875",
   "tg_tokens": null,
   "tps": 455.9,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 32,
   "ctx": null,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-61942-tt-c32-main",
   "model": "GLM-5.1-FP8",
   "notes": "vLLM PR #61942 AITER top-k sparse indexer skip; main arm; GLM-5.1-FP8 MI355X TP4; 1K in, c32 total throughput, +2.6% within noise per PR",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "FP8",
   "quote": "GLM-5.1-FP8; MI355X; gsm8k flexible: main 0.9462 / 0.9409, PR 0.9439 / 0.9484. 1K c32 total throughput 16616 \u2192 17042 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-16",
   "source_name": "vLLM PR #61942",
   "source_url": "https://github.com/vllm-project/vllm/pull/61942",
   "tg_tokens": null,
   "tps": 16616,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 32,
   "ctx": null,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-61942-tt-c32-pr",
   "model": "GLM-5.1-FP8",
   "notes": "vLLM PR #61942 AITER top-k sparse indexer skip; pr arm; GLM-5.1-FP8 MI355X TP4; 1K in, c32 total throughput, +2.6% within noise per PR",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "FP8",
   "quote": "GLM-5.1-FP8; MI355X; gsm8k flexible: main 0.9462 / 0.9409, PR 0.9439 / 0.9484. 1K c32 total throughput 16616 \u2192 17042 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-16",
   "source_name": "vLLM PR #61942",
   "source_url": "https://github.com/vllm-project/vllm/pull/61942",
   "tg_tokens": null,
   "tps": 17042,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 512,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-30701-b70-qwen38-27b-q80-pp-stock",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #30701 SYCL fast q8_0 reorder dequant, llama-perplexity -v -f ggml-sycl.cpp --chunks 12 -n 64 -b 512 -c 512 on Intel Arc Pro B70 oneAPI 2026.1 Debian Trixie, 6144-token prompt, clean A/B stock b11539 vs patch; Qwen3.8-27B Q8_0; stock kernel; prompt eval pp_tps only, tg not measured; stock kernel arm",
   "params": null,
   "power_w": null,
   "pp_tps": 516.36,
   "provenance": "sourced",
   "quant": "Q8_0",
   "quote": "Measured on Intel Arc Pro B70; Qwen3.8-27B-Q8_0 Prompt eval (t/s) 516.36",
   "ram_gb": null,
   "retrieved": "2026-10-17",
   "source_name": "llama.cpp PR #30701",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30701",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 512,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-30701-b70-qwen38-27b-q80-pp-patch",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #30701 SYCL fast q8_0 reorder dequant, llama-perplexity -v -f ggml-sycl.cpp --chunks 12 -n 64 -b 512 -c 512 on Intel Arc Pro B70 oneAPI 2026.1 Debian Trixie, 6144-token prompt, clean A/B stock b11539 vs patch; Qwen3.8-27B Q8_0; patch kernel; prompt eval pp_tps only, tg not measured; stock kernel arm",
   "params": null,
   "power_w": null,
   "pp_tps": 906.28,
   "provenance": "sourced",
   "quant": "Q8_0",
   "quote": "Measured on Intel Arc Pro B70; Qwen3.8-27B-Q8_0 Prompt eval (t/s) 516.36 906.28",
   "ram_gb": null,
   "retrieved": "2026-10-17",
   "source_name": "llama.cpp PR #30701",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30701",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 512,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-30701-b70-qwen38-27b-q8kxl-pp-stock",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #30701 SYCL fast q8_0 reorder dequant, llama-perplexity -v -f ggml-sycl.cpp --chunks 12 -n 64 -b 512 -c 512 on Intel Arc Pro B70 oneAPI 2026.1 Debian Trixie, 6144-token prompt, clean A/B stock b11539 vs patch; Qwen3.8-27B UD-Q8_K_XL; stock kernel; prompt eval pp_tps only, tg not measured; stock kernel arm",
   "params": null,
   "power_w": null,
   "pp_tps": 547.57,
   "provenance": "sourced",
   "quant": "UD-Q8_K_XL",
   "quote": "Measured on Intel Arc Pro B70; Qwen3.8-27B-UD-Q8_K_XL Prompt eval (t/s) 547.57",
   "ram_gb": null,
   "retrieved": "2026-10-17",
   "source_name": "llama.cpp PR #30701",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30701",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 512,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-30701-b70-qwen38-27b-q8kxl-pp-patch",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #30701 SYCL fast q8_0 reorder dequant, llama-perplexity -v -f ggml-sycl.cpp --chunks 12 -n 64 -b 512 -c 512 on Intel Arc Pro B70 oneAPI 2026.1 Debian Trixie, 6144-token prompt, clean A/B stock b11539 vs patch; Qwen3.8-27B UD-Q8_K_XL; patch kernel; prompt eval pp_tps only, tg not measured; stock kernel arm",
   "params": null,
   "power_w": null,
   "pp_tps": 938.73,
   "provenance": "sourced",
   "quant": "UD-Q8_K_XL",
   "quote": "Measured on Intel Arc Pro B70; Qwen3.8-27B-UD-Q8_K_XL Prompt eval (t/s) 547.57 938.73",
   "ram_gb": null,
   "retrieved": "2026-10-17",
   "source_name": "llama.cpp PR #30701",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30701",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 512,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-30701-b70-qwen38-27b-q8k4xl-pp-stock",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #30701 SYCL fast q8_0 reorder dequant, llama-perplexity -v -f ggml-sycl.cpp --chunks 12 -n 64 -b 512 -c 512 on Intel Arc Pro B70 oneAPI 2026.1 Debian Trixie, 6144-token prompt, clean A/B stock b11539 vs patch; Qwen3.8-27B UD-Q8_K_4XL; stock kernel; prompt eval pp_tps only, tg not measured; stock kernel arm; dual GPU run -dev SYCL0,SYCL1",
   "params": null,
   "power_w": null,
   "pp_tps": 630.0,
   "provenance": "sourced",
   "quant": "UD-Q8_K_4XL",
   "quote": "Measured on Intel Arc Pro B70; Qwen3.8-27B-UD-Q8_K_4XL Prompt eval (t/s) 630.00",
   "ram_gb": null,
   "retrieved": "2026-10-17",
   "source_name": "llama.cpp PR #30701",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30701",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 512,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-30701-b70-qwen38-27b-q8k4xl-pp-patch",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #30701 SYCL fast q8_0 reorder dequant, llama-perplexity -v -f ggml-sycl.cpp --chunks 12 -n 64 -b 512 -c 512 on Intel Arc Pro B70 oneAPI 2026.1 Debian Trixie, 6144-token prompt, clean A/B stock b11539 vs patch; Qwen3.8-27B UD-Q8_K_4XL; patch kernel; prompt eval pp_tps only, tg not measured; stock kernel arm; dual GPU run -dev SYCL0,SYCL1",
   "params": null,
   "power_w": null,
   "pp_tps": 975.48,
   "provenance": "sourced",
   "quant": "UD-Q8_K_4XL",
   "quote": "Measured on Intel Arc Pro B70; Qwen3.8-27B-UD-Q8_K_4XL Prompt eval (t/s) 630.00 975.48",
   "ram_gb": null,
   "retrieved": "2026-10-17",
   "source_name": "llama.cpp PR #30701",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30701",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 512,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-30701-b70-mistral-small-32-24b-q80-pp-stock",
   "model": "Mistral-Small-3.2-24B-Instruct",
   "notes": "llama.cpp PR #30701 SYCL fast q8_0 reorder dequant, llama-perplexity -v -f ggml-sycl.cpp --chunks 12 -n 64 -b 512 -c 512 on Intel Arc Pro B70 oneAPI 2026.1 Debian Trixie, 6144-token prompt, clean A/B stock b11539 vs patch; Mistral-Small-3.2-24B-Instruct Q8_0; stock kernel; prompt eval pp_tps only, tg not measured; stock kernel arm",
   "params": null,
   "power_w": null,
   "pp_tps": 635.88,
   "provenance": "sourced",
   "quant": "Q8_0",
   "quote": "Measured on Intel Arc Pro B70; Mistral-Small-3.2-24B-Instruct-Q8_0 Prompt eval (t/s) 635.88",
   "ram_gb": null,
   "retrieved": "2026-10-17",
   "source_name": "llama.cpp PR #30701",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30701",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 512,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-30701-b70-mistral-small-32-24b-q80-pp-patch",
   "model": "Mistral-Small-3.2-24B-Instruct",
   "notes": "llama.cpp PR #30701 SYCL fast q8_0 reorder dequant, llama-perplexity -v -f ggml-sycl.cpp --chunks 12 -n 64 -b 512 -c 512 on Intel Arc Pro B70 oneAPI 2026.1 Debian Trixie, 6144-token prompt, clean A/B stock b11539 vs patch; Mistral-Small-3.2-24B-Instruct Q8_0; patch kernel; prompt eval pp_tps only, tg not measured; stock kernel arm",
   "params": null,
   "power_w": null,
   "pp_tps": 1164.05,
   "provenance": "sourced",
   "quant": "Q8_0",
   "quote": "Measured on Intel Arc Pro B70; Mistral-Small-3.2-24B-Instruct-Q8_0 Prompt eval (t/s) 635.88 1,164.05",
   "ram_gb": null,
   "retrieved": "2026-10-17",
   "source_name": "llama.cpp PR #30701",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30701",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 512,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-30701-b70-qwen3-4b-q80-pp-stock",
   "model": "Qwen3 4B",
   "notes": "llama.cpp PR #30701 SYCL fast q8_0 reorder dequant, llama-perplexity -v -f ggml-sycl.cpp --chunks 12 -n 64 -b 512 -c 512 on Intel Arc Pro B70 oneAPI 2026.1 Debian Trixie, 6144-token prompt, clean A/B stock b11539 vs patch; Qwen3-4B Q8_0; stock kernel; prompt eval pp_tps only, tg not measured; stock kernel arm",
   "params": null,
   "power_w": null,
   "pp_tps": 2985.5,
   "provenance": "sourced",
   "quant": "Q8_0",
   "quote": "Measured on Intel Arc Pro B70; Qwen3-4B-Q8_0 Prompt eval (t/s) 2,985.50",
   "ram_gb": null,
   "retrieved": "2026-10-17",
   "source_name": "llama.cpp PR #30701",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30701",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 512,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-30701-b70-qwen3-4b-q80-pp-patch",
   "model": "Qwen3 4B",
   "notes": "llama.cpp PR #30701 SYCL fast q8_0 reorder dequant, llama-perplexity -v -f ggml-sycl.cpp --chunks 12 -n 64 -b 512 -c 512 on Intel Arc Pro B70 oneAPI 2026.1 Debian Trixie, 6144-token prompt, clean A/B stock b11539 vs patch; Qwen3-4B Q8_0; patch kernel; prompt eval pp_tps only, tg not measured; stock kernel arm",
   "params": null,
   "power_w": null,
   "pp_tps": 4831.58,
   "provenance": "sourced",
   "quant": "Q8_0",
   "quote": "Measured on Intel Arc Pro B70; Qwen3-4B-Q8_0 Prompt eval (t/s) 2,985.50 4,831.58",
   "ram_gb": null,
   "retrieved": "2026-10-17",
   "source_name": "llama.cpp PR #30701",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30701",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max (128 GB)",
   "id": "lc-30728-sparkx254b-ctx4096",
   "model": "Spark-X2.5-4B",
   "notes": "llama.cpp PR #30728 Metal GQA vec decode kernel, master arm baseline decode tok/s, -fa on -ctk f16 -ctv f16 -b 512 -ub 128, median of 8-10 rounds of 128-token greedy decode with cache_prompt; Spark-X2.5-4B Q8_0 at 4096 context",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q8_0",
   "quote": "Spark-X2.5-4B Q8_0 (4, 256) | 4k / 32k / 100k | 27.9 / 18.2 / 9.6; Results (M4 Max 128 GB, macOS 27, AC power)",
   "ram_gb": null,
   "retrieved": "2026-10-17",
   "source_name": "llama.cpp PR #30728",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30728",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max (128 GB)",
   "id": "lc-30728-sparkx254b-ctx32768",
   "model": "Spark-X2.5-4B",
   "notes": "llama.cpp PR #30728 Metal GQA vec decode kernel, master arm baseline decode tok/s, -fa on -ctk f16 -ctv f16 -b 512 -ub 128, median of 8-10 rounds of 128-token greedy decode with cache_prompt; Spark-X2.5-4B Q8_0 at 32768 context",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q8_0",
   "quote": "Spark-X2.5-4B Q8_0 (4, 256) | 4k / 32k / 100k | 27.9 / 18.2 / 9.6; Results (M4 Max 128 GB, macOS 27, AC power)",
   "ram_gb": null,
   "retrieved": "2026-10-17",
   "source_name": "llama.cpp PR #30728",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30728",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 102400,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max (128 GB)",
   "id": "lc-30728-sparkx254b-ctx102400",
   "model": "Spark-X2.5-4B",
   "notes": "llama.cpp PR #30728 Metal GQA vec decode kernel, master arm baseline decode tok/s, -fa on -ctk f16 -ctv f16 -b 512 -ub 128, median of 8-10 rounds of 128-token greedy decode with cache_prompt; Spark-X2.5-4B Q8_0 at 102400 context",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q8_0",
   "quote": "Spark-X2.5-4B Q8_0 (4, 256) | 4k / 32k / 100k | 27.9 / 18.2 / 9.6; Results (M4 Max 128 GB, macOS 27, AC power)",
   "ram_gb": null,
   "retrieved": "2026-10-17",
   "source_name": "llama.cpp PR #30728",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30728",
   "tg_tokens": null,
   "tps": 9.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max (128 GB)",
   "id": "lc-30728-minicpm52b-ctx4096",
   "model": "MiniCPM5-2B",
   "notes": "llama.cpp PR #30728 Metal GQA vec decode kernel, master arm baseline decode tok/s, -fa on -ctk f16 -ctv f16 -b 512 -ub 128, median of 8-10 rounds of 128-token greedy decode with cache_prompt; MiniCPM5-2B Q8_0 at 4096 context",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q8_0",
   "quote": "MiniCPM5-2B Q8_0 (8, 128) | 4k / 32k / 100k | 35.5 / 20.6 / 11.4; Results (M4 Max 128 GB, macOS 27, AC power)",
   "ram_gb": null,
   "retrieved": "2026-10-17",
   "source_name": "llama.cpp PR #30728",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30728",
   "tg_tokens": null,
   "tps": 35.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max (128 GB)",
   "id": "lc-30728-minicpm52b-ctx32768",
   "model": "MiniCPM5-2B",
   "notes": "llama.cpp PR #30728 Metal GQA vec decode kernel, master arm baseline decode tok/s, -fa on -ctk f16 -ctv f16 -b 512 -ub 128, median of 8-10 rounds of 128-token greedy decode with cache_prompt; MiniCPM5-2B Q8_0 at 32768 context",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q8_0",
   "quote": "MiniCPM5-2B Q8_0 (8, 128) | 4k / 32k / 100k | 35.5 / 20.6 / 11.4; Results (M4 Max 128 GB, macOS 27, AC power)",
   "ram_gb": null,
   "retrieved": "2026-10-17",
   "source_name": "llama.cpp PR #30728",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30728",
   "tg_tokens": null,
   "tps": 20.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 102400,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max (128 GB)",
   "id": "lc-30728-minicpm52b-ctx102400",
   "model": "MiniCPM5-2B",
   "notes": "llama.cpp PR #30728 Metal GQA vec decode kernel, master arm baseline decode tok/s, -fa on -ctk f16 -ctv f16 -b 512 -ub 128, median of 8-10 rounds of 128-token greedy decode with cache_prompt; MiniCPM5-2B Q8_0 at 102400 context",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q8_0",
   "quote": "MiniCPM5-2B Q8_0 (8, 128) | 4k / 32k / 100k | 35.5 / 20.6 / 11.4; Results (M4 Max 128 GB, macOS 27, AC power)",
   "ram_gb": null,
   "retrieved": "2026-10-17",
   "source_name": "llama.cpp PR #30728",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30728",
   "tg_tokens": null,
   "tps": 11.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max (128 GB)",
   "id": "lc-30728-llama323b-ctx4096",
   "model": "Llama-3.2-3B",
   "notes": "llama.cpp PR #30728 Metal GQA vec decode kernel, master arm baseline decode tok/s, -fa on -ctk f16 -ctv f16 -b 512 -ub 128, median of 8-10 rounds of 128-token greedy decode with cache_prompt; Llama-3.2-3B Q8_0 at 4096 context",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q8_0",
   "quote": "Llama-3.2-3B Q8_0 (3, 128) | 4k / 32k / 100k | 34.0 / 12.8 / 5.9; Results (M4 Max 128 GB, macOS 27, AC power)",
   "ram_gb": null,
   "retrieved": "2026-10-17",
   "source_name": "llama.cpp PR #30728",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30728",
   "tg_tokens": null,
   "tps": 34.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max (128 GB)",
   "id": "lc-30728-llama323b-ctx32768",
   "model": "Llama-3.2-3B",
   "notes": "llama.cpp PR #30728 Metal GQA vec decode kernel, master arm baseline decode tok/s, -fa on -ctk f16 -ctv f16 -b 512 -ub 128, median of 8-10 rounds of 128-token greedy decode with cache_prompt; Llama-3.2-3B Q8_0 at 32768 context",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q8_0",
   "quote": "Llama-3.2-3B Q8_0 (3, 128) | 4k / 32k / 100k | 34.0 / 12.8 / 5.9; Results (M4 Max 128 GB, macOS 27, AC power)",
   "ram_gb": null,
   "retrieved": "2026-10-17",
   "source_name": "llama.cpp PR #30728",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30728",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 102400,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max (128 GB)",
   "id": "lc-30728-llama323b-ctx102400",
   "model": "Llama-3.2-3B",
   "notes": "llama.cpp PR #30728 Metal GQA vec decode kernel, master arm baseline decode tok/s, -fa on -ctk f16 -ctv f16 -b 512 -ub 128, median of 8-10 rounds of 128-token greedy decode with cache_prompt; Llama-3.2-3B Q8_0 at 102400 context",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q8_0",
   "quote": "Llama-3.2-3B Q8_0 (3, 128) | 4k / 32k / 100k | 34.0 / 12.8 / 5.9; Results (M4 Max 128 GB, macOS 27, AC power)",
   "ram_gb": null,
   "retrieved": "2026-10-17",
   "source_name": "llama.cpp PR #30728",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30728",
   "tg_tokens": null,
   "tps": 5.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max (128 GB)",
   "id": "lc-30728-llama318b-ctx32768",
   "model": "Llama 3.1 8B",
   "notes": "llama.cpp PR #30728 Metal GQA vec decode kernel, master arm baseline decode tok/s, -fa on -ctk f16 -ctv f16 -b 512 -ub 128, median of 8-10 rounds of 128-token greedy decode with cache_prompt; Llama-3.1-8B Q4_K_M at 32768 context",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "Llama-3.1-8B Q4_K_M (4, 128) | 32k / 65k | 12.2 / 7.9; Results (M4 Max 128 GB, macOS 27, AC power)",
   "ram_gb": null,
   "retrieved": "2026-10-17",
   "source_name": "llama.cpp PR #30728",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30728",
   "tg_tokens": null,
   "tps": 12.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 66560,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max (128 GB)",
   "id": "lc-30728-llama318b-ctx66560",
   "model": "Llama 3.1 8B",
   "notes": "llama.cpp PR #30728 Metal GQA vec decode kernel, master arm baseline decode tok/s, -fa on -ctk f16 -ctv f16 -b 512 -ub 128, median of 8-10 rounds of 128-token greedy decode with cache_prompt; Llama-3.1-8B Q4_K_M at 66560 context",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "Llama-3.1-8B Q4_K_M (4, 128) | 32k / 65k | 12.2 / 7.9; Results (M4 Max 128 GB, macOS 27, AC power)",
   "ram_gb": null,
   "retrieved": "2026-10-17",
   "source_name": "llama.cpp PR #30728",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30728",
   "tg_tokens": null,
   "tps": 7.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max (128 GB)",
   "id": "lc-30728-qwen330ba3b-ctx32768",
   "model": "Qwen3 30B A3B",
   "notes": "llama.cpp PR #30728 Metal GQA vec decode kernel, master arm baseline decode tok/s, -fa on -ctk f16 -ctv f16 -b 512 -ub 128, median of 8-10 rounds of 128-token greedy decode with cache_prompt; Qwen3 30B A3B Q4_K_M at 32768 context",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "Qwen3-30B-A3B Q4_K_M (8, 128) | 32k | 16.2; Results (M4 Max 128 GB, macOS 27, AC power)",
   "ram_gb": null,
   "retrieved": "2026-10-17",
   "source_name": "llama.cpp PR #30728",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30728",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max (128 GB)",
   "id": "lc-30728-gemma4e4b-ctx32768",
   "model": "Gemma 4 E4B",
   "notes": "llama.cpp PR #30728 Metal GQA vec decode kernel, master arm baseline decode tok/s, -fa on -ctk f16 -ctv f16 -b 512 -ub 128, median of 8-10 rounds of 128-token greedy decode with cache_prompt; Gemma 4 E4B Q8_0 at 32768 context",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q8_0",
   "quote": "gemma-4-E4B Q8_0 (4, 512, SWA) | 32k / 100k | 18.1 / 14.3; Results (M4 Max 128 GB, macOS 27, AC power)",
   "ram_gb": null,
   "retrieved": "2026-10-17",
   "source_name": "llama.cpp PR #30728",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30728",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 102400,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max (128 GB)",
   "id": "lc-30728-gemma4e4b-ctx102400",
   "model": "Gemma 4 E4B",
   "notes": "llama.cpp PR #30728 Metal GQA vec decode kernel, master arm baseline decode tok/s, -fa on -ctk f16 -ctv f16 -b 512 -ub 128, median of 8-10 rounds of 128-token greedy decode with cache_prompt; Gemma 4 E4B Q8_0 at 102400 context",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q8_0",
   "quote": "gemma-4-E4B Q8_0 (4, 512, SWA) | 32k / 100k | 18.1 / 14.3; Results (M4 Max 128 GB, macOS 27, AC power)",
   "ram_gb": null,
   "retrieved": "2026-10-17",
   "source_name": "llama.cpp PR #30728",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30728",
   "tg_tokens": null,
   "tps": 14.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max (128 GB)",
   "id": "lc-30728-gptoss20b-ctx32768",
   "model": "gpt-oss 20B",
   "notes": "llama.cpp PR #30728 Metal GQA vec decode kernel, master arm baseline decode tok/s, -fa on -ctk f16 -ctv f16 -b 512 -ub 128, median of 8-10 rounds of 128-token greedy decode with cache_prompt; gpt-oss 20B MXFP4 at 32768 context",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "gpt-oss-20b MXFP4 (8, 64, sinks) | 32k / 100k | 21.8 / 15.6; Results (M4 Max 128 GB, macOS 27, AC power)",
   "ram_gb": null,
   "retrieved": "2026-10-17",
   "source_name": "llama.cpp PR #30728",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30728",
   "tg_tokens": null,
   "tps": 21.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 102400,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max (128 GB)",
   "id": "lc-30728-gptoss20b-ctx102400",
   "model": "gpt-oss 20B",
   "notes": "llama.cpp PR #30728 Metal GQA vec decode kernel, master arm baseline decode tok/s, -fa on -ctk f16 -ctv f16 -b 512 -ub 128, median of 8-10 rounds of 128-token greedy decode with cache_prompt; gpt-oss 20B MXFP4 at 102400 context",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "gpt-oss-20b MXFP4 (8, 64, sinks) | 32k / 100k | 21.8 / 15.6; Results (M4 Max 128 GB, macOS 27, AC power)",
   "ram_gb": null,
   "retrieved": "2026-10-17",
   "source_name": "llama.cpp PR #30728",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30728",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 30720,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max (128 GB)",
   "id": "lc-30728-glm49b0414-ctx30720",
   "model": "GLM-4-9B-0414",
   "notes": "llama.cpp PR #30728 Metal GQA vec decode kernel, master arm baseline decode tok/s, -fa on -ctk f16 -ctv f16 -b 512 -ub 128, median of 8-10 rounds of 128-token greedy decode with cache_prompt; GLM-4-9B-0414 Q4_K_M at 30720 context",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "GLM-4-9B-0414 Q4_K_M (16, 128) | 30k | 9.3; Results (M4 Max 128 GB, macOS 27, AC power)",
   "ram_gb": null,
   "retrieved": "2026-10-17",
   "source_name": "llama.cpp PR #30728",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30728",
   "tg_tokens": null,
   "tps": 9.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 2048,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max (128 GB)",
   "id": "lc-30728-smollm217b-ctx2048",
   "model": "SmolLM2-1.7B",
   "notes": "llama.cpp PR #30728 Metal GQA vec decode kernel, master arm baseline decode tok/s, -fa on -ctk f16 -ctv f16 -b 512 -ub 128, median of 8-10 rounds of 128-token greedy decode with cache_prompt; SmolLM2-1.7B unspecified at 2048 context",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unspecified",
   "quote": "SmolLM2-1.7B (MHA, control) | 2k / 8k | 58.7 / 42.0; Results (M4 Max 128 GB, macOS 27, AC power)",
   "ram_gb": null,
   "retrieved": "2026-10-17",
   "source_name": "llama.cpp PR #30728",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30728",
   "tg_tokens": null,
   "tps": 58.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max (128 GB)",
   "id": "lc-30728-smollm217b-ctx8192",
   "model": "SmolLM2-1.7B",
   "notes": "llama.cpp PR #30728 Metal GQA vec decode kernel, master arm baseline decode tok/s, -fa on -ctk f16 -ctv f16 -b 512 -ub 128, median of 8-10 rounds of 128-token greedy decode with cache_prompt; SmolLM2-1.7B unspecified at 8192 context",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unspecified",
   "quote": "SmolLM2-1.7B (MHA, control) | 2k / 8k | 58.7 / 42.0; Results (M4 Max 128 GB, macOS 27, AC power)",
   "ram_gb": null,
   "retrieved": "2026-10-17",
   "source_name": "llama.cpp PR #30728",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30728",
   "tg_tokens": null,
   "tps": 42.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "AMD Radeon HD 7900",
   "id": "lc-30738-q40q40-pp512-before",
   "model": "Qwen3.5-4B",
   "notes": "llama.cpp PR #30738 vulkan Q4_0 KV dequant+transpose fast path in flash attention, llama-bench -d 32768 -fa 1, Qwen3.5-4B Q4_K_M on AMD Radeon HD 7900 (GCN1/radv), mean of runs; q4_0/q4_0 KV cache arm, before, llama-bench pp",
   "params": null,
   "power_w": null,
   "pp_tps": 82.46,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "| q4_0/q4_0 | 82.46 \u00b1 0.16 | **116.44 \u00b1 0.09** | 13.19 \u00b1 0.02 | **16.60 \u00b1 0.01** |; Qwen3.5-4B Q4_K_M, fa=1; AMD Radeon HD 7900 (RADV TAHITI, GCN1, Vulkan, no hardware int8 dot product)",
   "ram_gb": null,
   "retrieved": "2026-10-17",
   "source_name": "llama.cpp PR #30738",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30738",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "AMD Radeon HD 7900",
   "id": "lc-30738-q40q40-pp512-after",
   "model": "Qwen3.5-4B",
   "notes": "llama.cpp PR #30738 vulkan Q4_0 KV dequant+transpose fast path in flash attention, llama-bench -d 32768 -fa 1, Qwen3.5-4B Q4_K_M on AMD Radeon HD 7900 (GCN1/radv), mean of runs; q4_0/q4_0 KV cache arm, after PR, llama-bench pp",
   "params": null,
   "power_w": null,
   "pp_tps": 116.44,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "| q4_0/q4_0 | 82.46 \u00b1 0.16 | **116.44 \u00b1 0.09** | 13.19 \u00b1 0.02 | **16.60 \u00b1 0.01** |; Qwen3.5-4B Q4_K_M, fa=1; AMD Radeon HD 7900 (RADV TAHITI, GCN1, Vulkan, no hardware int8 dot product)",
   "ram_gb": null,
   "retrieved": "2026-10-17",
   "source_name": "llama.cpp PR #30738",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30738",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "AMD Radeon HD 7900",
   "id": "lc-30738-q40q40-tg128-before",
   "model": "Qwen3.5-4B",
   "notes": "llama.cpp PR #30738 vulkan Q4_0 KV dequant+transpose fast path in flash attention, llama-bench -d 32768 -fa 1, Qwen3.5-4B Q4_K_M on AMD Radeon HD 7900 (GCN1/radv), mean of runs; q4_0/q4_0 KV cache arm, before, llama-bench tg",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "| q4_0/q4_0 | 82.46 \u00b1 0.16 | **116.44 \u00b1 0.09** | 13.19 \u00b1 0.02 | **16.60 \u00b1 0.01** |; Qwen3.5-4B Q4_K_M, fa=1; AMD Radeon HD 7900 (RADV TAHITI, GCN1, Vulkan, no hardware int8 dot product)",
   "ram_gb": null,
   "retrieved": "2026-10-17",
   "source_name": "llama.cpp PR #30738",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30738",
   "tg_tokens": null,
   "tps": 13.19,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "AMD Radeon HD 7900",
   "id": "lc-30738-q40q40-tg128-after",
   "model": "Qwen3.5-4B",
   "notes": "llama.cpp PR #30738 vulkan Q4_0 KV dequant+transpose fast path in flash attention, llama-bench -d 32768 -fa 1, Qwen3.5-4B Q4_K_M on AMD Radeon HD 7900 (GCN1/radv), mean of runs; q4_0/q4_0 KV cache arm, after PR, llama-bench tg",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "| q4_0/q4_0 | 82.46 \u00b1 0.16 | **116.44 \u00b1 0.09** | 13.19 \u00b1 0.02 | **16.60 \u00b1 0.01** |; Qwen3.5-4B Q4_K_M, fa=1; AMD Radeon HD 7900 (RADV TAHITI, GCN1, Vulkan, no hardware int8 dot product)",
   "ram_gb": null,
   "retrieved": "2026-10-17",
   "source_name": "llama.cpp PR #30738",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30738",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "AMD Radeon HD 7900",
   "id": "lc-30738-q80q40-pp512-before",
   "model": "Qwen3.5-4B",
   "notes": "llama.cpp PR #30738 vulkan Q4_0 KV dequant+transpose fast path in flash attention, llama-bench -d 32768 -fa 1, Qwen3.5-4B Q4_K_M on AMD Radeon HD 7900 (GCN1/radv), mean of runs; q8_0/q4_0 mixed KV cache arm, before, llama-bench pp",
   "params": null,
   "power_w": null,
   "pp_tps": 73.78,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "| q8_0/q4_0 | 73.78 \u00b1 14.18 | **116.24 \u00b1 0.17** | 12.36 \u00b1 0.03 | **16.83 \u00b1 0.03** |; Qwen3.5-4B Q4_K_M, fa=1; AMD Radeon HD 7900 (RADV TAHITI, GCN1, Vulkan, no hardware int8 dot product)",
   "ram_gb": null,
   "retrieved": "2026-10-17",
   "source_name": "llama.cpp PR #30738",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30738",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "AMD Radeon HD 7900",
   "id": "lc-30738-q80q40-pp512-after",
   "model": "Qwen3.5-4B",
   "notes": "llama.cpp PR #30738 vulkan Q4_0 KV dequant+transpose fast path in flash attention, llama-bench -d 32768 -fa 1, Qwen3.5-4B Q4_K_M on AMD Radeon HD 7900 (GCN1/radv), mean of runs; q8_0/q4_0 mixed KV cache arm, after PR, llama-bench pp",
   "params": null,
   "power_w": null,
   "pp_tps": 116.24,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "| q8_0/q4_0 | 73.78 \u00b1 14.18 | **116.24 \u00b1 0.17** | 12.36 \u00b1 0.03 | **16.83 \u00b1 0.03** |; Qwen3.5-4B Q4_K_M, fa=1; AMD Radeon HD 7900 (RADV TAHITI, GCN1, Vulkan, no hardware int8 dot product)",
   "ram_gb": null,
   "retrieved": "2026-10-17",
   "source_name": "llama.cpp PR #30738",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30738",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "AMD Radeon HD 7900",
   "id": "lc-30738-q80q40-tg128-before",
   "model": "Qwen3.5-4B",
   "notes": "llama.cpp PR #30738 vulkan Q4_0 KV dequant+transpose fast path in flash attention, llama-bench -d 32768 -fa 1, Qwen3.5-4B Q4_K_M on AMD Radeon HD 7900 (GCN1/radv), mean of runs; q8_0/q4_0 mixed KV cache arm, before, llama-bench tg",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "| q8_0/q4_0 | 73.78 \u00b1 14.18 | **116.24 \u00b1 0.17** | 12.36 \u00b1 0.03 | **16.83 \u00b1 0.03** |; Qwen3.5-4B Q4_K_M, fa=1; AMD Radeon HD 7900 (RADV TAHITI, GCN1, Vulkan, no hardware int8 dot product)",
   "ram_gb": null,
   "retrieved": "2026-10-17",
   "source_name": "llama.cpp PR #30738",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30738",
   "tg_tokens": null,
   "tps": 12.36,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "AMD Radeon HD 7900",
   "id": "lc-30738-q80q40-tg128-after",
   "model": "Qwen3.5-4B",
   "notes": "llama.cpp PR #30738 vulkan Q4_0 KV dequant+transpose fast path in flash attention, llama-bench -d 32768 -fa 1, Qwen3.5-4B Q4_K_M on AMD Radeon HD 7900 (GCN1/radv), mean of runs; q8_0/q4_0 mixed KV cache arm, after PR, llama-bench tg",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "| q8_0/q4_0 | 73.78 \u00b1 14.18 | **116.24 \u00b1 0.17** | 12.36 \u00b1 0.03 | **16.83 \u00b1 0.03** |; Qwen3.5-4B Q4_K_M, fa=1; AMD Radeon HD 7900 (RADV TAHITI, GCN1, Vulkan, no hardware int8 dot product)",
   "ram_gb": null,
   "retrieved": "2026-10-17",
   "source_name": "llama.cpp PR #30738",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30738",
   "tg_tokens": null,
   "tps": 16.83,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1,
   "ctx": 2048,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "hf3-ornith15-35b-r9700-decode-2k",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "HF card field note: MXFP4 GGUF of the ornith-ai Ornith-1.5-35B-A3B MoE (~35B total / ~3B active, MTP layer kept); WHIRL 0.2.0 vs llama.cpp b11214 ROCm/Vulkan comparison table, numbers are the llama.cpp ROCm column; R9700 32 GB USB4 eGPU; re-measured 2026-10-17; decode after a 2k-token document, 256-token answer, plain greedy",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Ornith-1.5-35B-A3B MXFP4 (MoE, ~3B active); AMD Radeon AI PRO R9700 (32 GB), single run per cell; 2k 181.8 238.4 208.3 115.4 (plain) 115.9 (plain) +107% +106%",
   "retrieved": "2026-10-17",
   "source_name": "HF model card: tsaipifong Ornith-1.5-35B-A3B-MXFP4",
   "source_url": "https://huggingface.co/tsaipifong/Ornith-1.5-35B-A3B-MXFP4-GGUF",
   "tg_tokens": 256,
   "tps": "115.4",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "hf3-ornith15-35b-r9700-decode-8k",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "HF card field note: MXFP4 GGUF of the ornith-ai Ornith-1.5-35B-A3B MoE (~35B total / ~3B active, MTP layer kept); WHIRL 0.2.0 vs llama.cpp b11214 ROCm/Vulkan comparison table, numbers are the llama.cpp ROCm column; R9700 32 GB USB4 eGPU; re-measured 2026-10-17; decode after a 8k-token document, 256-token answer, plain greedy",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Ornith-1.5-35B-A3B MXFP4 (MoE, ~3B active); AMD Radeon AI PRO R9700 (32 GB), single run per cell; 8k 174.6 212.9 191.3 112.3 (plain) 111.9 (plain) +90% +90%",
   "retrieved": "2026-10-17",
   "source_name": "HF model card: tsaipifong Ornith-1.5-35B-A3B-MXFP4",
   "source_url": "https://huggingface.co/tsaipifong/Ornith-1.5-35B-A3B-MXFP4-GGUF",
   "tg_tokens": 256,
   "tps": "112.3",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "hf3-ornith15-35b-r9700-decode-32k",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "HF card field note: MXFP4 GGUF of the ornith-ai Ornith-1.5-35B-A3B MoE (~35B total / ~3B active, MTP layer kept); WHIRL 0.2.0 vs llama.cpp b11214 ROCm/Vulkan comparison table, numbers are the llama.cpp ROCm column; R9700 32 GB USB4 eGPU; re-measured 2026-10-17; decode after a 32k-token document, 256-token answer, plain greedy",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Ornith-1.5-35B-A3B MXFP4 (MoE, ~3B active); AMD Radeon AI PRO R9700 (32 GB), single run per cell; 32k 154.2 199.1 183.5 101.4 (plain) 100.6 (plain) +96% +98%",
   "retrieved": "2026-10-17",
   "source_name": "HF model card: tsaipifong Ornith-1.5-35B-A3B-MXFP4",
   "source_url": "https://huggingface.co/tsaipifong/Ornith-1.5-35B-A3B-MXFP4-GGUF",
   "tg_tokens": 256,
   "tps": "101.4",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "hf3-ornith15-35b-r9700-decode-64k",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "HF card field note: MXFP4 GGUF of the ornith-ai Ornith-1.5-35B-A3B MoE (~35B total / ~3B active, MTP layer kept); WHIRL 0.2.0 vs llama.cpp b11214 ROCm/Vulkan comparison table, numbers are the llama.cpp ROCm column; R9700 32 GB USB4 eGPU; re-measured 2026-10-17; decode after a 64k-token document, 256-token answer, plain greedy",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Ornith-1.5-35B-A3B MXFP4 (MoE, ~3B active); AMD Radeon AI PRO R9700 (32 GB), single run per cell; 64k 138.6 169.1 174.9 90.7 (plain) 89.6 (plain) +86% +89%",
   "retrieved": "2026-10-17",
   "source_name": "HF model card: tsaipifong Ornith-1.5-35B-A3B-MXFP4",
   "source_url": "https://huggingface.co/tsaipifong/Ornith-1.5-35B-A3B-MXFP4-GGUF",
   "tg_tokens": 256,
   "tps": "90.7",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1,
   "ctx": 98304,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "hf3-ornith15-35b-r9700-decode-96k",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "HF card field note: MXFP4 GGUF of the ornith-ai Ornith-1.5-35B-A3B MoE (~35B total / ~3B active, MTP layer kept); WHIRL 0.2.0 vs llama.cpp b11214 ROCm/Vulkan comparison table, numbers are the llama.cpp ROCm column; R9700 32 GB USB4 eGPU; re-measured 2026-10-17; decode after a 96k-token document, 256-token answer, plain greedy",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Ornith-1.5-35B-A3B MXFP4 (MoE, ~3B active); AMD Radeon AI PRO R9700 (32 GB), single run per cell; 96k 126.4 148.1 156.2 81.9 (plain) 81.1 (plain) +81% +83%",
   "retrieved": "2026-10-17",
   "source_name": "HF model card: tsaipifong Ornith-1.5-35B-A3B-MXFP4",
   "source_url": "https://huggingface.co/tsaipifong/Ornith-1.5-35B-A3B-MXFP4-GGUF",
   "tg_tokens": 256,
   "tps": "81.9",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "hf3-ornith15-35b-r9700-decode-128k",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "HF card field note: MXFP4 GGUF of the ornith-ai Ornith-1.5-35B-A3B MoE (~35B total / ~3B active, MTP layer kept); WHIRL 0.2.0 vs llama.cpp b11214 ROCm/Vulkan comparison table, numbers are the llama.cpp ROCm column; R9700 32 GB USB4 eGPU; re-measured 2026-10-17; decode after a 128k-token document, 256-token answer, plain greedy",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Ornith-1.5-35B-A3B MXFP4 (MoE, ~3B active); AMD Radeon AI PRO R9700 (32 GB), single run per cell; 128k 121.3 131.3 152.4 75.3 (plain) 73.8 (plain) +74% +78%",
   "retrieved": "2026-10-17",
   "source_name": "HF model card: tsaipifong Ornith-1.5-35B-A3B-MXFP4",
   "source_url": "https://huggingface.co/tsaipifong/Ornith-1.5-35B-A3B-MXFP4-GGUF",
   "tg_tokens": 256,
   "tps": "75.3",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1,
   "ctx": 2048,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "hf3-ornith15-35b-r9700-pp-2k",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "HF card field note: MXFP4 GGUF of the ornith-ai Ornith-1.5-35B-A3B MoE (~35B total / ~3B active, MTP layer kept); WHIRL 0.2.0 vs llama.cpp b11214 ROCm/Vulkan comparison table, numbers are the llama.cpp ROCm column; R9700 32 GB USB4 eGPU; re-measured 2026-10-17; prefill at 2k prompt",
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": "4014",
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Ornith-1.5-35B-A3B MXFP4 (MoE, ~3B active); AMD Radeon AI PRO R9700 (32 GB), single run per cell; 2k 7,575 11,550 11,801 4,014 3,551 +188% +225%",
   "retrieved": "2026-10-17",
   "source_name": "HF model card: tsaipifong Ornith-1.5-35B-A3B-MXFP4",
   "source_url": "https://huggingface.co/tsaipifong/Ornith-1.5-35B-A3B-MXFP4-GGUF",
   "tg_tokens": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "hf3-ornith15-35b-r9700-pp-8k",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "HF card field note: MXFP4 GGUF of the ornith-ai Ornith-1.5-35B-A3B MoE (~35B total / ~3B active, MTP layer kept); WHIRL 0.2.0 vs llama.cpp b11214 ROCm/Vulkan comparison table, numbers are the llama.cpp ROCm column; R9700 32 GB USB4 eGPU; re-measured 2026-10-17; prefill at 8k prompt",
   "power_w": null,
   "pp_tokens": 8192,
   "pp_tps": "4487",
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Ornith-1.5-35B-A3B MXFP4 (MoE, ~3B active); AMD Radeon AI PRO R9700 (32 GB), single run per cell; 8k 6,846 10,601 10,501 4,487 3,912 +136% +171%",
   "retrieved": "2026-10-17",
   "source_name": "HF model card: tsaipifong Ornith-1.5-35B-A3B-MXFP4",
   "source_url": "https://huggingface.co/tsaipifong/Ornith-1.5-35B-A3B-MXFP4-GGUF",
   "tg_tokens": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "hf3-ornith15-35b-r9700-pp-32k",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "HF card field note: MXFP4 GGUF of the ornith-ai Ornith-1.5-35B-A3B MoE (~35B total / ~3B active, MTP layer kept); WHIRL 0.2.0 vs llama.cpp b11214 ROCm/Vulkan comparison table, numbers are the llama.cpp ROCm column; R9700 32 GB USB4 eGPU; re-measured 2026-10-17; prefill at 32k prompt",
   "power_w": null,
   "pp_tokens": 32768,
   "pp_tps": "3872",
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Ornith-1.5-35B-A3B MXFP4 (MoE, ~3B active); AMD Radeon AI PRO R9700 (32 GB), single run per cell; 32k 5,959 8,469 8,450 3,872 3,395 +119% +149%",
   "retrieved": "2026-10-17",
   "source_name": "HF model card: tsaipifong Ornith-1.5-35B-A3B-MXFP4",
   "source_url": "https://huggingface.co/tsaipifong/Ornith-1.5-35B-A3B-MXFP4-GGUF",
   "tg_tokens": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "hf3-ornith15-35b-r9700-pp-64k",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "HF card field note: MXFP4 GGUF of the ornith-ai Ornith-1.5-35B-A3B MoE (~35B total / ~3B active, MTP layer kept); WHIRL 0.2.0 vs llama.cpp b11214 ROCm/Vulkan comparison table, numbers are the llama.cpp ROCm column; R9700 32 GB USB4 eGPU; re-measured 2026-10-17; prefill at 64k prompt",
   "power_w": null,
   "pp_tokens": 65536,
   "pp_tps": "3173",
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Ornith-1.5-35B-A3B MXFP4 (MoE, ~3B active); AMD Radeon AI PRO R9700 (32 GB), single run per cell; 64k 4,873 6,451 6,420 3,173 2,735 +103% +136%",
   "retrieved": "2026-10-17",
   "source_name": "HF model card: tsaipifong Ornith-1.5-35B-A3B-MXFP4",
   "source_url": "https://huggingface.co/tsaipifong/Ornith-1.5-35B-A3B-MXFP4-GGUF",
   "tg_tokens": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1,
   "ctx": 98304,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "hf3-ornith15-35b-r9700-pp-96k",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "HF card field note: MXFP4 GGUF of the ornith-ai Ornith-1.5-35B-A3B MoE (~35B total / ~3B active, MTP layer kept); WHIRL 0.2.0 vs llama.cpp b11214 ROCm/Vulkan comparison table, numbers are the llama.cpp ROCm column; R9700 32 GB USB4 eGPU; re-measured 2026-10-17; prefill at 96k prompt",
   "power_w": null,
   "pp_tokens": 98304,
   "pp_tps": "2667",
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Ornith-1.5-35B-A3B MXFP4 (MoE, ~3B active); AMD Radeon AI PRO R9700 (32 GB), single run per cell; 96k 4,098 5,156 5,143 2,667 2,280 +93% +126%",
   "retrieved": "2026-10-17",
   "source_name": "HF model card: tsaipifong Ornith-1.5-35B-A3B-MXFP4",
   "source_url": "https://huggingface.co/tsaipifong/Ornith-1.5-35B-A3B-MXFP4-GGUF",
   "tg_tokens": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "hf3-ornith15-35b-r9700-pp-128k",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "HF card field note: MXFP4 GGUF of the ornith-ai Ornith-1.5-35B-A3B MoE (~35B total / ~3B active, MTP layer kept); WHIRL 0.2.0 vs llama.cpp b11214 ROCm/Vulkan comparison table, numbers are the llama.cpp ROCm column; R9700 32 GB USB4 eGPU; re-measured 2026-10-17; prefill at 128k prompt",
   "power_w": null,
   "pp_tokens": 131072,
   "pp_tps": "2307",
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Ornith-1.5-35B-A3B MXFP4 (MoE, ~3B active); AMD Radeon AI PRO R9700 (32 GB), single run per cell; 128k 3,520 4,275 4,270 2,307 1,922 +85% +122%",
   "retrieved": "2026-10-17",
   "source_name": "HF model card: tsaipifong Ornith-1.5-35B-A3B-MXFP4",
   "source_url": "https://huggingface.co/tsaipifong/Ornith-1.5-35B-A3B-MXFP4-GGUF",
   "tg_tokens": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 512,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-30752-b70-qwen38-27b-q80-run1-pp-mainline",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #30752 SYCL q8 fast dequant (memory-structure change, wider stores), llama-perplexity -f ggml-sycl.cpp --chunks 12 -n 64 -b 512 -c 512 on Intel Arc Pro B70, 6144-token prompt, 12 pp512 passes, clean A/B mainline vs this PR; prompt eval pp_tps only, tg not measured; Qwen3.8-27B Q8_0, run 1 of a 2-run A/B; mainline arm",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1050.4,
   "provenance": "sourced",
   "quant": "Q8_0",
   "quote": "Intel Arc Pro B70; Qwen3.8-27B-Q8_0 (run 1) 1050.4",
   "retrieved": "2026-10-18",
   "source_name": "llama.cpp PR #30752",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30752",
   "tg_tokens": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 512,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-30752-b70-qwen38-27b-q80-run1-pp-pr",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #30752 SYCL q8 fast dequant (memory-structure change, wider stores), llama-perplexity -f ggml-sycl.cpp --chunks 12 -n 64 -b 512 -c 512 on Intel Arc Pro B70, 6144-token prompt, 12 pp512 passes, clean A/B mainline vs this PR; prompt eval pp_tps only, tg not measured; Qwen3.8-27B Q8_0, run 1 of a 2-run A/B; pr arm",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1095.1,
   "provenance": "sourced",
   "quant": "Q8_0",
   "quote": "Intel Arc Pro B70; Qwen3.8-27B-Q8_0 (run 1) 1050.4 1095.1",
   "retrieved": "2026-10-18",
   "source_name": "llama.cpp PR #30752",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30752",
   "tg_tokens": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 512,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-30752-b70-qwen38-27b-q80-repeat-pp-mainline",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #30752 SYCL q8 fast dequant (memory-structure change, wider stores), llama-perplexity -f ggml-sycl.cpp --chunks 12 -n 64 -b 512 -c 512 on Intel Arc Pro B70, 6144-token prompt, 12 pp512 passes, clean A/B mainline vs this PR; prompt eval pp_tps only, tg not measured; Qwen3.8-27B Q8_0, repeat run of the same A/B; mainline arm",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1043.4,
   "provenance": "sourced",
   "quant": "Q8_0",
   "quote": "Intel Arc Pro B70; Qwen3.8-27B-Q8_0 (repeat) 1043.4",
   "retrieved": "2026-10-18",
   "source_name": "llama.cpp PR #30752",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30752",
   "tg_tokens": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 512,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-30752-b70-qwen38-27b-q80-repeat-pp-pr",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #30752 SYCL q8 fast dequant (memory-structure change, wider stores), llama-perplexity -f ggml-sycl.cpp --chunks 12 -n 64 -b 512 -c 512 on Intel Arc Pro B70, 6144-token prompt, 12 pp512 passes, clean A/B mainline vs this PR; prompt eval pp_tps only, tg not measured; Qwen3.8-27B Q8_0, repeat run of the same A/B; pr arm",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1086.4,
   "provenance": "sourced",
   "quant": "Q8_0",
   "quote": "Intel Arc Pro B70; Qwen3.8-27B-Q8_0 (repeat) 1043.4 1086.4",
   "retrieved": "2026-10-18",
   "source_name": "llama.cpp PR #30752",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30752",
   "tg_tokens": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 512,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-30752-b70-qwen38-27b-q8kxl-pp-mainline",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #30752 SYCL q8 fast dequant (memory-structure change, wider stores), llama-perplexity -f ggml-sycl.cpp --chunks 12 -n 64 -b 512 -c 512 on Intel Arc Pro B70, 6144-token prompt, 12 pp512 passes, clean A/B mainline vs this PR; prompt eval pp_tps only, tg not measured; Qwen3.8-27B UD-Q8_K_XL; mainline arm",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1086.0,
   "provenance": "sourced",
   "quant": "UD-Q8_K_XL",
   "quote": "Intel Arc Pro B70; Qwen3.8-27B-UD-Q8_K_XL 1086.0",
   "retrieved": "2026-10-18",
   "source_name": "llama.cpp PR #30752",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30752",
   "tg_tokens": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 512,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-30752-b70-qwen38-27b-q8kxl-pp-pr",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #30752 SYCL q8 fast dequant (memory-structure change, wider stores), llama-perplexity -f ggml-sycl.cpp --chunks 12 -n 64 -b 512 -c 512 on Intel Arc Pro B70, 6144-token prompt, 12 pp512 passes, clean A/B mainline vs this PR; prompt eval pp_tps only, tg not measured; Qwen3.8-27B UD-Q8_K_XL; pr arm",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1130.2,
   "provenance": "sourced",
   "quant": "UD-Q8_K_XL",
   "quote": "Intel Arc Pro B70; Qwen3.8-27B-UD-Q8_K_XL 1086.0 1130.2",
   "retrieved": "2026-10-18",
   "source_name": "llama.cpp PR #30752",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30752",
   "tg_tokens": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 512,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-30752-b70-qwen38-27b-q8k4xl-pp-mainline",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #30752 SYCL q8 fast dequant (memory-structure change, wider stores), llama-perplexity -f ggml-sycl.cpp --chunks 12 -n 64 -b 512 -c 512 on Intel Arc Pro B70, 6144-token prompt, 12 pp512 passes, clean A/B mainline vs this PR; prompt eval pp_tps only, tg not measured; Qwen3.8-27B UD-Q8_K_4XL split across SYCL0/SYCL1 dual device; mainline arm",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1082.7,
   "provenance": "sourced",
   "quant": "UD-Q8_K_4XL",
   "quote": "Intel Arc Pro B70; Qwen3.8-27B-UD-Q8_K_4XL (SYCL0,SYCL1) 1082.7",
   "retrieved": "2026-10-18",
   "source_name": "llama.cpp PR #30752",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30752",
   "tg_tokens": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 512,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-30752-b70-qwen38-27b-q8k4xl-pp-pr",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp PR #30752 SYCL q8 fast dequant (memory-structure change, wider stores), llama-perplexity -f ggml-sycl.cpp --chunks 12 -n 64 -b 512 -c 512 on Intel Arc Pro B70, 6144-token prompt, 12 pp512 passes, clean A/B mainline vs this PR; prompt eval pp_tps only, tg not measured; Qwen3.8-27B UD-Q8_K_4XL split across SYCL0/SYCL1 dual device; pr arm",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1105.5,
   "provenance": "sourced",
   "quant": "UD-Q8_K_4XL",
   "quote": "Intel Arc Pro B70; Qwen3.8-27B-UD-Q8_K_4XL (SYCL0,SYCL1) 1082.7 1105.5",
   "retrieved": "2026-10-18",
   "source_name": "llama.cpp PR #30752",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30752",
   "tg_tokens": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 512,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-30752-b70-qwen3-4b-pp-mainline",
   "model": "Qwen3 4B",
   "notes": "llama.cpp PR #30752 SYCL q8 fast dequant (memory-structure change, wider stores), llama-perplexity -f ggml-sycl.cpp --chunks 12 -n 64 -b 512 -c 512 on Intel Arc Pro B70, 6144-token prompt, 12 pp512 passes, clean A/B mainline vs this PR; prompt eval pp_tps only, tg not measured; Qwen3 4B Q8_0; mainline arm",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5489.1,
   "provenance": "sourced",
   "quant": "Q8_0",
   "quote": "Intel Arc Pro B70; Qwen3-4B-Q8_0 5489.1",
   "retrieved": "2026-10-18",
   "source_name": "llama.cpp PR #30752",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30752",
   "tg_tokens": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 512,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-30752-b70-qwen3-4b-pp-pr",
   "model": "Qwen3 4B",
   "notes": "llama.cpp PR #30752 SYCL q8 fast dequant (memory-structure change, wider stores), llama-perplexity -f ggml-sycl.cpp --chunks 12 -n 64 -b 512 -c 512 on Intel Arc Pro B70, 6144-token prompt, 12 pp512 passes, clean A/B mainline vs this PR; prompt eval pp_tps only, tg not measured; Qwen3 4B Q8_0; pr arm",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5652.1,
   "provenance": "sourced",
   "quant": "Q8_0",
   "quote": "Intel Arc Pro B70; Qwen3-4B-Q8_0 5489.1 5652.1",
   "retrieved": "2026-10-18",
   "source_name": "llama.cpp PR #30752",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30752",
   "tg_tokens": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 512,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-30752-b70-mistral-small-32-24b-pp-mainline",
   "model": "Mistral-Small-3.2-24B-Instruct",
   "notes": "llama.cpp PR #30752 SYCL q8 fast dequant (memory-structure change, wider stores), llama-perplexity -f ggml-sycl.cpp --chunks 12 -n 64 -b 512 -c 512 on Intel Arc Pro B70, 6144-token prompt, 12 pp512 passes, clean A/B mainline vs this PR; prompt eval pp_tps only, tg not measured; Mistral-Small-3.2-24B-Instruct-2506 Q8_0; mainline arm",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1416.7,
   "provenance": "sourced",
   "quant": "Q8_0",
   "quote": "Intel Arc Pro B70; Mistral-Small-3.2-24B-Instruct-2506-Q8_0 1416.7",
   "retrieved": "2026-10-18",
   "source_name": "llama.cpp PR #30752",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30752",
   "tg_tokens": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 512,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-30752-b70-mistral-small-32-24b-pp-pr",
   "model": "Mistral-Small-3.2-24B-Instruct",
   "notes": "llama.cpp PR #30752 SYCL q8 fast dequant (memory-structure change, wider stores), llama-perplexity -f ggml-sycl.cpp --chunks 12 -n 64 -b 512 -c 512 on Intel Arc Pro B70, 6144-token prompt, 12 pp512 passes, clean A/B mainline vs this PR; prompt eval pp_tps only, tg not measured; Mistral-Small-3.2-24B-Instruct-2506 Q8_0; pr arm",
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1494.7,
   "provenance": "sourced",
   "quant": "Q8_0",
   "quote": "Intel Arc Pro B70; Mistral-Small-3.2-24B-Instruct-2506-Q8_0 1416.7 1494.7",
   "retrieved": "2026-10-18",
   "source_name": "llama.cpp PR #30752",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30752",
   "tg_tokens": null,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": null,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-61694-mi355x8-kimik3-main-recipe-c4",
   "model": "Kimi-K3",
   "notes": "Kimi-K3 serving on MI355X (gfx950), TP8/DCP1; quant unstated on the page; AgentX conc4 InferenceX recipe (kimik3/vllm/mi355x-fp4-mtp/agentic.yaml, DSpark k=7, synthetic acceptance 3.84, GPU-resident FP8 KV, prefix caching); main recipe as is (FP8 in_proj, stock decode); Output 174.25 tok/s, Total 23531 tok/s; one 3600 s run",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "Kimi-K3; 8\u00d7 MI355X; AgentX conc4; main , recipe as is (FP8 in_proj, stock decode) 174.25 23531",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61694",
   "source_url": "https://github.com/vllm-project/vllm/pull/61694",
   "tg_tokens": null,
   "tps": 174.25,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": null,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-61694-mi355x8-kimik3-main-bf16-c4",
   "model": "Kimi-K3",
   "notes": "Kimi-K3 serving on MI355X (gfx950), TP8/DCP1; quant unstated on the page; AgentX conc4 InferenceX recipe; main with BF16 in_proj, stock decode; Output 175.70 tok/s, Total 23214 tok/s; one 3600 s run",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "Kimi-K3; 8\u00d7 MI355X; AgentX conc4; main , BF16 in_proj, stock decode 175.70 23214",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61694",
   "source_url": "https://github.com/vllm-project/vllm/pull/61694",
   "tg_tokens": null,
   "tps": 175.7,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": null,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-61694-mi355x8-kimik3-main-latency-c4",
   "model": "Kimi-K3",
   "notes": "Kimi-K3 serving on MI355X (gfx950), TP8/DCP1; quant unstated on the page; AgentX conc4 InferenceX recipe; main + latency_specialized (#60683, 8-row steps only); Output 180.98 tok/s, Total 23899 tok/s; one 3600 s run",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "Kimi-K3; 8\u00d7 MI355X; AgentX conc4; main + latency_specialized ( #60683 , 8-row steps only) 180.98 23899",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61694",
   "source_url": "https://github.com/vllm-project/vllm/pull/61694",
   "tg_tokens": null,
   "tps": 180.98,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 4,
   "ctx": null,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-61694-mi355x8-kimik3-pr-c4",
   "model": "Kimi-K3",
   "notes": "Kimi-K3 serving on MI355X (gfx950), TP8/DCP1; quant unstated on the page; AgentX conc4 InferenceX recipe; this PR + latency_specialized (8/16/24/32-row steps); Output 190.90 tok/s, Total 25346 tok/s; one 3600 s run",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "Kimi-K3; 8\u00d7 MI355X; AgentX conc4; This PR + latency_specialized (8/16/24/32-row steps) 190.90 25346",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "vLLM PR #61694",
   "source_url": "https://github.com/vllm-project/vllm/pull/61694",
   "tg_tokens": null,
   "tps": 190.9,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 64,
   "ctx": null,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "2xH200",
   "id": "vllm-61903-out-c64-main",
   "model": "MiniMax M2.5",
   "notes": "MiniMax-M2.5 FP8 router GEMM on 2xH200, TP2; main arm, mixed 1K-in/128-out workload, concurrency 64; four runs 1422 / 1472 / 1461 / 1467 output tok/s (first stored); main and PR alternating in one container",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "FP8",
   "quote": "MiniMax-M2.5 FP8, TP2 on 2xH200; 1K in / 128 out, concurrency 64: output tok/s 1422 / 1472 / 1461 / 1467",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "vLLM PR #61903",
   "source_url": "https://github.com/vllm-project/vllm/pull/61903",
   "tg_tokens": null,
   "tps": 1422,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 64,
   "ctx": null,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "2xH200",
   "id": "vllm-61903-out-c64-pr",
   "model": "MiniMax M2.5",
   "notes": "MiniMax-M2.5 FP8 router GEMM on 2xH200, TP2; this PR arm, mixed 1K-in/128-out workload, concurrency 64; four runs 1518 / 1525 / 1523 / 1522 output tok/s (first stored), +4.6% vs main",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "FP8",
   "quote": "MiniMax-M2.5 FP8, TP2 on 2xH200; output tok/s 1422 / 1472 / 1461 / 1467 1518 / 1525 / 1523 / 1522",
   "ram_gb": null,
   "retrieved": "2026-10-15",
   "source_name": "vLLM PR #61903",
   "source_url": "https://github.com/vllm-project/vllm/pull/61903",
   "tg_tokens": null,
   "tps": 1518,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "MI355X",
   "id": "vllm-61435-mi355x8-kimik3-inferencex-dcp1-auto",
   "model": "Kimi-K3",
   "notes": "vLLM PR #61435 InferenceX recipe 8k/1k, c1, k=7, TP8 on 8x MI355X, decode_backend auto vs specialized; DCP1/DCP8 = decode context parallel degree; output tok/s; DCP1 auto (base)",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "Kimi-K3; MI355X; Output tok/s 179.77",
   "ram_gb": null,
   "retrieved": "2026-10-22",
   "source_name": "vLLM PR #61435",
   "source_url": "https://github.com/vllm-project/vllm/pull/61435",
   "tg_tokens": null,
   "tps": 179.77,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-61435-mi355x8-kimik3-inferencex-dcp1-spec",
   "model": "Kimi-K3",
   "notes": "vLLM PR #61435 InferenceX recipe 8k/1k, c1, k=7, TP8 on 8x MI355X, decode_backend auto vs specialized; DCP1/DCP8 = decode context parallel degree; output tok/s; DCP1 specialized, +29.5% over auto",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "Kimi-K3; MI355X; 179.77 232.87",
   "ram_gb": null,
   "retrieved": "2026-10-22",
   "source_name": "vLLM PR #61435",
   "source_url": "https://github.com/vllm-project/vllm/pull/61435",
   "tg_tokens": null,
   "tps": 232.87,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "MI355X",
   "id": "vllm-61435-mi355x8-kimik3-inferencex-dcp8-auto",
   "model": "Kimi-K3",
   "notes": "vLLM PR #61435 InferenceX recipe 8k/1k, c1, k=7, TP8 on 8x MI355X, decode_backend auto vs specialized; DCP1/DCP8 = decode context parallel degree; output tok/s; DCP8 auto (base)",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "Kimi-K3; MI355X; 164.79 233.11",
   "ram_gb": null,
   "retrieved": "2026-10-22",
   "source_name": "vLLM PR #61435",
   "source_url": "https://github.com/vllm-project/vllm/pull/61435",
   "tg_tokens": null,
   "tps": 164.79,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "MI355X",
   "id": "vllm-61435-mi355x8-kimik3-inferencex-dcp8-spec",
   "model": "Kimi-K3",
   "notes": "vLLM PR #61435 InferenceX recipe 8k/1k, c1, k=7, TP8 on 8x MI355X, decode_backend auto vs specialized; DCP1/DCP8 = decode context parallel degree; output tok/s; DCP8 specialized, +41.5% over auto",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "Kimi-K3; MI355X; 233.11 (+41.5%)",
   "ram_gb": null,
   "retrieved": "2026-10-22",
   "source_name": "vLLM PR #61435",
   "source_url": "https://github.com/vllm-project/vllm/pull/61435",
   "tg_tokens": null,
   "tps": 233.11,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-22",
   "flags": [],
   "hardware": "B300",
   "id": "vllm-63101-b300-glm53-mxfp4-c1-baseline",
   "model": "GLM-5.3",
   "notes": "vLLM PR #63101 MXFP4 MoE activation quant (sorted_expert_ids direct index) E2E, TP4 on B300, random 256 in / 1024 out, max-concurrency 1, fp8 KV cache, no spec decode; output tok/s; baseline arm (main)",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "GLM-5.3; B300; MXFP4; Baseline 86.66",
   "ram_gb": null,
   "retrieved": "2026-10-22",
   "source_name": "vLLM PR #63101",
   "source_url": "https://github.com/vllm-project/vllm/pull/63101",
   "tg_tokens": null,
   "tps": 86.66,
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-22",
   "flags": [],
   "hardware": "B300",
   "id": "vllm-63101-b300-glm53-mxfp4-c1-pr",
   "model": "GLM-5.3",
   "notes": "vLLM PR #63101 MXFP4 MoE activation quant (sorted_expert_ids direct index) E2E, TP4 on B300, random 256 in / 1024 out, max-concurrency 1, fp8 KV cache, no spec decode; output tok/s; this PR arm, +5.7% over baseline",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "GLM-5.3; B300; MXFP4; This PR 91.60",
   "ram_gb": null,
   "retrieved": "2026-10-22",
   "source_name": "vLLM PR #63101",
   "source_url": "https://github.com/vllm-project/vllm/pull/63101",
   "tg_tokens": null,
   "tps": 91.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": "2026-10-21",
   "flags": [
    "contradiction"
   ],
   "hardware": "AMD Ryzen 7 8845H",
   "id": "lc-30923-ryzen8845h-lfm2524ba2b-q4km-32768-master",
   "model": "LFM2.5-24B-A2B",
   "notes": "llama.cpp PR #30923 CPU softmax parallelism across head/batch dims; LFM2.5-24B-A2B Q4_K_M, CPU-only on AMD Ryzen 7 8845H (8 threads), 32K context, Q8_0 K-cache, F16 V-cache, flash-attn off; decode tok/s; master arm",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "LFM2.5-24B-A2B; Q4_K_M; AMD Ryzen 7 8845H; Decode throughput 15.156",
   "ram_gb": null,
   "retrieved": "2026-10-22",
   "source_name": "llama.cpp PR #30923",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30923",
   "tg_tokens": null,
   "tps": 15.156,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": "2026-10-21",
   "flags": [],
   "hardware": "AMD Ryzen 7 8845H",
   "id": "lc-30923-ryzen8845h-lfm2524ba2b-q4km-32768-pr",
   "model": "LFM2.5-24B-A2B",
   "notes": "llama.cpp PR #30923 CPU softmax parallelism across head/batch dims; LFM2.5-24B-A2B Q4_K_M, CPU-only on AMD Ryzen 7 8845H (8 threads), 32K context, Q8_0 K-cache, F16 V-cache, flash-attn off; decode tok/s; this PR arm, +16.4% over master",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "LFM2.5-24B-A2B; Q4_K_M; AMD Ryzen 7 8845H; 15.156 tok/s 17.640",
   "ram_gb": null,
   "retrieved": "2026-10-22",
   "source_name": "llama.cpp PR #30923",
   "source_url": "https://github.com/ggml-org/llama.cpp/pull/30923",
   "tg_tokens": null,
   "tps": 17.64,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 3060 12GB",
   "id": "hardwarecorner-rtx-3060-12gb-qwen3-8b-q4-k-ctx4096",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 3060 12GB hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 1696.8,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 3060 12GB, Qwen3 8B (Q4_K), 4k Ctx: generation 55.2 t/s, prompt processing 1,696.8 t/s",
   "ram_gb": 12.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3060-12gb/",
   "tps": 55.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 3060 12GB",
   "id": "hardwarecorner-rtx-3060-12gb-qwen3-8b-q4-k-ctx16384",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 3060 12GB hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 1119.2,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 3060 12GB, Qwen3 8B (Q4_K), 16k Ctx: generation 42.0 t/s, prompt processing 1,119.2 t/s",
   "ram_gb": 12.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3060-12gb/",
   "tps": 42.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 3060 12GB",
   "id": "hardwarecorner-rtx-3060-12gb-qwen3-8b-q4-k-ctx32768",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 3060 12GB hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 764.7,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 3060 12GB, Qwen3 8B (Q4_K), 32k Ctx: generation 31.9 t/s, prompt processing 764.7 t/s",
   "ram_gb": 12.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3060-12gb/",
   "tps": 31.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 3060 12GB",
   "id": "hardwarecorner-rtx-3060-12gb-qwen3-14b-q4-k-ctx4096",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 3060 12GB hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 972.6,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 3060 12GB, Qwen3 14B (Q4_K), 4k Ctx: generation 31.2 t/s, prompt processing 972.6 t/s",
   "ram_gb": 12.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3060-12gb/",
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 3060 12GB",
   "id": "hardwarecorner-rtx-3060-12gb-qwen3-14b-q4-k-ctx16384",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 3060 12GB hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 678.2,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 3060 12GB, Qwen3 14B (Q4_K), 16k Ctx: generation 22.7 t/s, prompt processing 678.2 t/s",
   "ram_gb": 12.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3060-12gb/",
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 3080 Ti",
   "id": "hardwarecorner-rtx-3080-ti-qwen3-8b-q4-k-ctx4096",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 3080 Ti hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 4211.7,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 3080 Ti, Qwen3 8B (Q4_K), 4k Ctx: generation 115.2 t/s, prompt processing 4,211.7 t/s",
   "ram_gb": 12.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3080-ti/",
   "tps": 115.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 3080 Ti",
   "id": "hardwarecorner-rtx-3080-ti-qwen3-8b-q4-k-ctx16384",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 3080 Ti hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 2658.5,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 3080 Ti, Qwen3 8B (Q4_K), 16k Ctx: generation 87.9 t/s, prompt processing 2,658.5 t/s",
   "ram_gb": 12.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3080-ti/",
   "tps": 87.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 3080 Ti",
   "id": "hardwarecorner-rtx-3080-ti-qwen3-8b-q4-k-ctx32768",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 3080 Ti hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 1761.2,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 3080 Ti, Qwen3 8B (Q4_K), 32k Ctx: generation 68.1 t/s, prompt processing 1,761.2 t/s",
   "ram_gb": 12.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3080-ti/",
   "tps": 68.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 3080 Ti",
   "id": "hardwarecorner-rtx-3080-ti-qwen3-14b-q4-k-ctx4096",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 3080 Ti hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 2600.1,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 3080 Ti, Qwen3 14B (Q4_K), 4k Ctx: generation 69.9 t/s, prompt processing 2,600.1 t/s",
   "ram_gb": 12.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3080-ti/",
   "tps": 69.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 3080 Ti",
   "id": "hardwarecorner-rtx-3080-ti-qwen3-14b-q4-k-ctx16384",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 3080 Ti hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 1773.9,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 3080 Ti, Qwen3 14B (Q4_K), 16k Ctx: generation 52.3 t/s, prompt processing 1,773.9 t/s",
   "ram_gb": 12.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3080-ti/",
   "tps": 52.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hardwarecorner-rtx-3090-gemma4-26b-q4-k-ctx4096",
   "model": "Gemma4 26B",
   "notes": "Hardware Corner RTX 3090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "26B",
   "power_w": null,
   "pp_tps": 3625.6,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 3090, Gemma4 26B (Q4_K), 4k Ctx: generation 119.4 t/s, prompt processing 3,625.6 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3090/",
   "tps": 119.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hardwarecorner-rtx-3090-gemma4-26b-q4-k-ctx16384",
   "model": "Gemma4 26B",
   "notes": "Hardware Corner RTX 3090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "26B",
   "power_w": null,
   "pp_tps": 3068.9,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 3090, Gemma4 26B (Q4_K), 16k Ctx: generation 115.0 t/s, prompt processing 3,068.9 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3090/",
   "tps": 115.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hardwarecorner-rtx-3090-gemma4-26b-q4-k-ctx32768",
   "model": "Gemma4 26B",
   "notes": "Hardware Corner RTX 3090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "26B",
   "power_w": null,
   "pp_tps": 2453.4,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 3090, Gemma4 26B (Q4_K), 32k Ctx: generation 107.5 t/s, prompt processing 2,453.4 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3090/",
   "tps": 107.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 65536,
   "date": null,
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hardwarecorner-rtx-3090-gemma4-26b-q4-k-ctx65536",
   "model": "Gemma4 26B",
   "notes": "Hardware Corner RTX 3090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "26B",
   "power_w": null,
   "pp_tps": 1765.1,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 3090, Gemma4 26B (Q4_K), 64k Ctx: generation 98.9 t/s, prompt processing 1,765.1 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3090/",
   "tps": 98.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 131072,
   "date": null,
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hardwarecorner-rtx-3090-gemma4-26b-q4-k-ctx131072",
   "model": "Gemma4 26B",
   "notes": "Hardware Corner RTX 3090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "26B",
   "power_w": null,
   "pp_tps": 1147.1,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 3090, Gemma4 26B (Q4_K), 128k Ctx: generation 83.0 t/s, prompt processing 1,147.1 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3090/",
   "tps": 83.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 262144,
   "date": null,
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hardwarecorner-rtx-3090-gemma4-26b-q4-k-ctx262144",
   "model": "Gemma4 26B",
   "notes": "Hardware Corner RTX 3090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "26B",
   "power_w": null,
   "pp_tps": 671.4,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 3090, Gemma4 26B (Q4_K), 256k Ctx: generation 64.4 t/s, prompt processing 671.4 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3090/",
   "tps": 64.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hardwarecorner-rtx-3090-qwen3-5-27b-q4-k-ctx4096",
   "model": "Qwen3.5-27B",
   "notes": "Hardware Corner RTX 3090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "27B",
   "power_w": null,
   "pp_tps": 1104.2,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 3090, Qwen3.5 27B (Q4_K), 4k Ctx: generation 33.5 t/s, prompt processing 1,104.2 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3090/",
   "tps": 33.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hardwarecorner-rtx-3090-qwen3-5-27b-q4-k-ctx16384",
   "model": "Qwen3.5-27B",
   "notes": "Hardware Corner RTX 3090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "27B",
   "power_w": null,
   "pp_tps": 977.4,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 3090, Qwen3.5 27B (Q4_K), 16k Ctx: generation 32.3 t/s, prompt processing 977.4 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3090/",
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hardwarecorner-rtx-3090-qwen3-5-27b-q4-k-ctx32768",
   "model": "Qwen3.5-27B",
   "notes": "Hardware Corner RTX 3090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "27B",
   "power_w": null,
   "pp_tps": 848.2,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 3090, Qwen3.5 27B (Q4_K), 32k Ctx: generation 31.0 t/s, prompt processing 848.2 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3090/",
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 65536,
   "date": null,
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hardwarecorner-rtx-3090-qwen3-5-27b-q4-k-ctx65536",
   "model": "Qwen3.5-27B",
   "notes": "Hardware Corner RTX 3090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "27B",
   "power_w": null,
   "pp_tps": 678.9,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 3090, Qwen3.5 27B (Q4_K), 64k Ctx: generation 28.8 t/s, prompt processing 678.9 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3090/",
   "tps": 28.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hardwarecorner-rtx-3090-gemma4-31b-q4-k-ctx4096",
   "model": "Gemma4 31B",
   "notes": "Hardware Corner RTX 3090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "31B",
   "power_w": null,
   "pp_tps": 1155.8,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 3090, Gemma4 31B (Q4_K), 4k Ctx: generation 34.7 t/s, prompt processing 1,155.8 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3090/",
   "tps": 34.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hardwarecorner-rtx-3090-gemma4-31b-q4-k-ctx16384",
   "model": "Gemma4 31B",
   "notes": "Hardware Corner RTX 3090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "31B",
   "power_w": null,
   "pp_tps": 913.2,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 3090, Gemma4 31B (Q4_K), 16k Ctx: generation 33.5 t/s, prompt processing 913.2 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3090/",
   "tps": 33.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hardwarecorner-rtx-3090-gemma4-31b-q4-k-ctx32768",
   "model": "Gemma4 31B",
   "notes": "Hardware Corner RTX 3090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "31B",
   "power_w": null,
   "pp_tps": 723.7,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 3090, Gemma4 31B (Q4_K), 32k Ctx: generation 31.4 t/s, prompt processing 723.7 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3090/",
   "tps": 31.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hardwarecorner-rtx-3090-qwen3-5-35b-mxfp4-ctx4096",
   "model": "Qwen3.5 35B",
   "notes": "Hardware Corner RTX 3090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "35B",
   "power_w": null,
   "pp_tps": 2622.1,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 3090, Qwen3.5 35B (MXFP4), 4k Ctx: generation 111.2 t/s, prompt processing 2,622.1 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3090/",
   "tps": 111.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hardwarecorner-rtx-3090-qwen3-5-35b-mxfp4-ctx16384",
   "model": "Qwen3.5 35B",
   "notes": "Hardware Corner RTX 3090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "35B",
   "power_w": null,
   "pp_tps": 2381.3,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 3090, Qwen3.5 35B (MXFP4), 16k Ctx: generation 107.1 t/s, prompt processing 2,381.3 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3090/",
   "tps": 107.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hardwarecorner-rtx-3090-qwen3-5-35b-mxfp4-ctx32768",
   "model": "Qwen3.5 35B",
   "notes": "Hardware Corner RTX 3090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "35B",
   "power_w": null,
   "pp_tps": 2121.6,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 3090, Qwen3.5 35B (MXFP4), 32k Ctx: generation 101.2 t/s, prompt processing 2,121.6 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3090/",
   "tps": 101.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 65536,
   "date": null,
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hardwarecorner-rtx-3090-qwen3-5-35b-mxfp4-ctx65536",
   "model": "Qwen3.5 35B",
   "notes": "Hardware Corner RTX 3090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "35B",
   "power_w": null,
   "pp_tps": 1749.8,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 3090, Qwen3.5 35B (MXFP4), 64k Ctx: generation 93.1 t/s, prompt processing 1,749.8 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3090/",
   "tps": 93.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 131072,
   "date": null,
   "flags": [],
   "hardware": "RTX 3090",
   "id": "hardwarecorner-rtx-3090-qwen3-5-35b-mxfp4-ctx131072",
   "model": "Qwen3.5 35B",
   "notes": "Hardware Corner RTX 3090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "35B",
   "power_w": null,
   "pp_tps": 1288.9,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 3090, Qwen3.5 35B (MXFP4), 128k Ctx: generation 79.4 t/s, prompt processing 1,288.9 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3090/",
   "tps": 79.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 3090 Ti",
   "id": "hardwarecorner-rtx-3090-ti-qwen3-8b-q4-k-ctx4096",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 3090 Ti hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 4467.9,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 3090 Ti, Qwen3 8B (Q4_K), 4k Ctx: generation 123.7 t/s, prompt processing 4,467.9 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3090-ti/",
   "tps": 123.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 3090 Ti",
   "id": "hardwarecorner-rtx-3090-ti-qwen3-8b-q4-k-ctx16384",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 3090 Ti hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 2834.2,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 3090 Ti, Qwen3 8B (Q4_K), 16k Ctx: generation 93.6 t/s, prompt processing 2,834.2 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3090-ti/",
   "tps": 93.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 3090 Ti",
   "id": "hardwarecorner-rtx-3090-ti-qwen3-8b-q4-k-ctx32768",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 3090 Ti hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 1868.0,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 3090 Ti, Qwen3 8B (Q4_K), 32k Ctx: generation 71.8 t/s, prompt processing 1,868.0 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3090-ti/",
   "tps": 71.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 65536,
   "date": null,
   "flags": [],
   "hardware": "RTX 3090 Ti",
   "id": "hardwarecorner-rtx-3090-ti-qwen3-8b-q4-k-ctx65536",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 3090 Ti hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 1111.1,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 3090 Ti, Qwen3 8B (Q4_K), 64k Ctx: generation 49.0 t/s, prompt processing 1,111.1 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3090-ti/",
   "tps": 49.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 131072,
   "date": null,
   "flags": [],
   "hardware": "RTX 3090 Ti",
   "id": "hardwarecorner-rtx-3090-ti-qwen3-8b-q4-k-ctx131072",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 3090 Ti hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 612.4,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 3090 Ti, Qwen3 8B (Q4_K), 128k Ctx: generation 29.0 t/s, prompt processing 612.4 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3090-ti/",
   "tps": 29.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 3090 Ti",
   "id": "hardwarecorner-rtx-3090-ti-qwen3-14b-q4-k-ctx4096",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 3090 Ti hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 2817.1,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 3090 Ti, Qwen3 14B (Q4_K), 4k Ctx: generation 76.2 t/s, prompt processing 2,817.1 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3090-ti/",
   "tps": 76.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 3090 Ti",
   "id": "hardwarecorner-rtx-3090-ti-qwen3-14b-q4-k-ctx16384",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 3090 Ti hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 1914.8,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 3090 Ti, Qwen3 14B (Q4_K), 16k Ctx: generation 56.9 t/s, prompt processing 1,914.8 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3090-ti/",
   "tps": 56.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 3090 Ti",
   "id": "hardwarecorner-rtx-3090-ti-qwen3-14b-q4-k-ctx32768",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 3090 Ti hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 1232.1,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 3090 Ti, Qwen3 14B (Q4_K), 32k Ctx: generation 42.1 t/s, prompt processing 1,232.1 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3090-ti/",
   "tps": 42.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 65536,
   "date": null,
   "flags": [],
   "hardware": "RTX 3090 Ti",
   "id": "hardwarecorner-rtx-3090-ti-qwen3-14b-q4-k-ctx65536",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 3090 Ti hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 828.5,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 3090 Ti, Qwen3 14B (Q4_K), 64k Ctx: generation 28.2 t/s, prompt processing 828.5 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3090-ti/",
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 3090 Ti",
   "id": "hardwarecorner-rtx-3090-ti-gpt-oss-20b-mxfp4-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 3090 Ti hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 4805.4,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 3090 Ti, gpt-oss 20B (MXFP4), 4k Ctx: generation 160.3 t/s, prompt processing 4,805.4 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3090-ti/",
   "tps": 160.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 3090 Ti",
   "id": "hardwarecorner-rtx-3090-ti-gpt-oss-20b-mxfp4-ctx16384",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 3090 Ti hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 3440.1,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 3090 Ti, gpt-oss 20B (MXFP4), 16k Ctx: generation 137.6 t/s, prompt processing 3,440.1 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3090-ti/",
   "tps": 137.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 3090 Ti",
   "id": "hardwarecorner-rtx-3090-ti-gpt-oss-20b-mxfp4-ctx32768",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 3090 Ti hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 2521.3,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 3090 Ti, gpt-oss 20B (MXFP4), 32k Ctx: generation 119.7 t/s, prompt processing 2,521.3 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3090-ti/",
   "tps": 119.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 65536,
   "date": null,
   "flags": [],
   "hardware": "RTX 3090 Ti",
   "id": "hardwarecorner-rtx-3090-ti-gpt-oss-20b-mxfp4-ctx65536",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 3090 Ti hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 1666.5,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 3090 Ti, gpt-oss 20B (MXFP4), 64k Ctx: generation 94.9 t/s, prompt processing 1,666.5 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3090-ti/",
   "tps": 94.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 131072,
   "date": null,
   "flags": [],
   "hardware": "RTX 3090 Ti",
   "id": "hardwarecorner-rtx-3090-ti-gpt-oss-20b-mxfp4-ctx131072",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 3090 Ti hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 936.4,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 3090 Ti, gpt-oss 20B (MXFP4), 128k Ctx: generation 61.5 t/s, prompt processing 936.4 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3090-ti/",
   "tps": 61.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 3090 Ti",
   "id": "hardwarecorner-rtx-3090-ti-qwen3-30b-a3b-q4-k-ctx4096",
   "model": "Qwen3 30B A3B",
   "notes": "Hardware Corner RTX 3090 Ti hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "30B",
   "power_w": null,
   "pp_tps": 3441.0,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 3090 Ti, Qwen3 30B A3B (Q4_K), 4k Ctx: generation 166.9 t/s, prompt processing 3,441.0 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3090-ti/",
   "tps": 166.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 3090 Ti",
   "id": "hardwarecorner-rtx-3090-ti-qwen3-30b-a3b-q4-k-ctx16384",
   "model": "Qwen3 30B A3B",
   "notes": "Hardware Corner RTX 3090 Ti hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "30B",
   "power_w": null,
   "pp_tps": 2205.6,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 3090 Ti, Qwen3 30B A3B (Q4_K), 16k Ctx: generation 121.9 t/s, prompt processing 2,205.6 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3090-ti/",
   "tps": 121.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 3090 Ti",
   "id": "hardwarecorner-rtx-3090-ti-qwen3-30b-a3b-q4-k-ctx32768",
   "model": "Qwen3 30B A3B",
   "notes": "Hardware Corner RTX 3090 Ti hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "30B",
   "power_w": null,
   "pp_tps": 1483.9,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 3090 Ti, Qwen3 30B A3B (Q4_K), 32k Ctx: generation 92.0 t/s, prompt processing 1,483.9 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3090-ti/",
   "tps": 92.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 3090 Ti",
   "id": "hardwarecorner-rtx-3090-ti-qwen3-32b-q4-k-ctx4096",
   "model": "Qwen3 32B",
   "notes": "Hardware Corner RTX 3090 Ti hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "32B",
   "power_w": null,
   "pp_tps": 1238.8,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 3090 Ti, Qwen3 32B (Q4_K), 4k Ctx: generation 38.0 t/s, prompt processing 1,238.8 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3090-ti/",
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 3090 Ti",
   "id": "hardwarecorner-rtx-3090-ti-qwen3-32b-q4-k-ctx16384",
   "model": "Qwen3 32B",
   "notes": "Hardware Corner RTX 3090 Ti hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "32B",
   "power_w": null,
   "pp_tps": 862.7,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 3090 Ti, Qwen3 32B (Q4_K), 16k Ctx: generation 32.6 t/s, prompt processing 862.7 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3090-ti/",
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 4060 Ti 16GB",
   "id": "hardwarecorner-rtx-4060-ti-16gb-qwen3-8b-q4-k-ctx4096",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 4060 Ti 16GB hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 2675.2,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4060 Ti 16GB, Qwen3 8B (Q4_K), 4k Ctx: generation 45.8 t/s, prompt processing 2,675.2 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4060-ti-16gb/",
   "tps": 45.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 4060 Ti 16GB",
   "id": "hardwarecorner-rtx-4060-ti-16gb-qwen3-8b-q4-k-ctx16384",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 4060 Ti 16GB hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 1480.8,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4060 Ti 16GB, Qwen3 8B (Q4_K), 16k Ctx: generation 34.3 t/s, prompt processing 1,480.8 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4060-ti-16gb/",
   "tps": 34.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 4060 Ti 16GB",
   "id": "hardwarecorner-rtx-4060-ti-16gb-qwen3-8b-q4-k-ctx32768",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 4060 Ti 16GB hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 760.3,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4060 Ti 16GB, Qwen3 8B (Q4_K), 32k Ctx: generation 25.5 t/s, prompt processing 760.3 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4060-ti-16gb/",
   "tps": 25.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 65536,
   "date": null,
   "flags": [],
   "hardware": "RTX 4060 Ti 16GB",
   "id": "hardwarecorner-rtx-4060-ti-16gb-qwen3-8b-q4-k-ctx65536",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 4060 Ti 16GB hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 392.1,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4060 Ti 16GB, Qwen3 8B (Q4_K), 64k Ctx: generation 13.0 t/s, prompt processing 392.1 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4060-ti-16gb/",
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 4060 Ti 16GB",
   "id": "hardwarecorner-rtx-4060-ti-16gb-qwen3-14b-q4-k-ctx4096",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 4060 Ti 16GB hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 1645.7,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4060 Ti 16GB, Qwen3 14B (Q4_K), 4k Ctx: generation 27.4 t/s, prompt processing 1,645.7 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4060-ti-16gb/",
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 4060 Ti 16GB",
   "id": "hardwarecorner-rtx-4060-ti-16gb-qwen3-14b-q4-k-ctx16384",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 4060 Ti 16GB hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 917.6,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4060 Ti 16GB, Qwen3 14B (Q4_K), 16k Ctx: generation 22.4 t/s, prompt processing 917.6 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4060-ti-16gb/",
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 4060 Ti 16GB",
   "id": "hardwarecorner-rtx-4060-ti-16gb-qwen3-14b-q4-k-ctx32768",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 4060 Ti 16GB hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 541.4,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4060 Ti 16GB, Qwen3 14B (Q4_K), 32k Ctx: generation 17.9 t/s, prompt processing 541.4 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4060-ti-16gb/",
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 4060 Ti 16GB",
   "id": "hardwarecorner-rtx-4060-ti-16gb-gpt-oss-20b-mxfp4-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 4060 Ti 16GB hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 3274.2,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 4060 Ti 16GB, gpt-oss 20B (MXFP4), 4k Ctx: generation 63.2 t/s, prompt processing 3,274.2 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4060-ti-16gb/",
   "tps": 63.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 4060 Ti 16GB",
   "id": "hardwarecorner-rtx-4060-ti-16gb-gpt-oss-20b-mxfp4-ctx16384",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 4060 Ti 16GB hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 2552.9,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 4060 Ti 16GB, gpt-oss 20B (MXFP4), 16k Ctx: generation 57.8 t/s, prompt processing 2,552.9 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4060-ti-16gb/",
   "tps": 57.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 4060 Ti 16GB",
   "id": "hardwarecorner-rtx-4060-ti-16gb-gpt-oss-20b-mxfp4-ctx32768",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 4060 Ti 16GB hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 1964.9,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 4060 Ti 16GB, gpt-oss 20B (MXFP4), 32k Ctx: generation 51.5 t/s, prompt processing 1,964.9 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4060-ti-16gb/",
   "tps": 51.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 65536,
   "date": null,
   "flags": [],
   "hardware": "RTX 4060 Ti 16GB",
   "id": "hardwarecorner-rtx-4060-ti-16gb-gpt-oss-20b-mxfp4-ctx65536",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 4060 Ti 16GB hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 1332.2,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 4060 Ti 16GB, gpt-oss 20B (MXFP4), 64k Ctx: generation 41.1 t/s, prompt processing 1,332.2 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4060-ti-16gb/",
   "tps": 41.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 131072,
   "date": null,
   "flags": [],
   "hardware": "RTX 4060 Ti 16GB",
   "id": "hardwarecorner-rtx-4060-ti-16gb-gpt-oss-20b-mxfp4-ctx131072",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 4060 Ti 16GB hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 780.2,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 4060 Ti 16GB, gpt-oss 20B (MXFP4), 128k Ctx: generation 31.1 t/s, prompt processing 780.2 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4060-ti-16gb/",
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 4070",
   "id": "hardwarecorner-rtx-4070-qwen3-8b-q4-k-ctx4096",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 4070 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 3564.1,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4070, Qwen3 8B (Q4_K), 4k Ctx: generation 71.2 t/s, prompt processing 3,564.1 t/s",
   "ram_gb": 12.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4070/",
   "tps": 71.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 4070",
   "id": "hardwarecorner-rtx-4070-qwen3-8b-q4-k-ctx16384",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 4070 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 2064.3,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4070, Qwen3 8B (Q4_K), 16k Ctx: generation 52.1 t/s, prompt processing 2,064.3 t/s",
   "ram_gb": 12.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4070/",
   "tps": 52.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 4070",
   "id": "hardwarecorner-rtx-4070-qwen3-8b-q4-k-ctx32768",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 4070 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 1116.7,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4070, Qwen3 8B (Q4_K), 32k Ctx: generation 38.1 t/s, prompt processing 1,116.7 t/s",
   "ram_gb": 12.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4070/",
   "tps": 38.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 4070",
   "id": "hardwarecorner-rtx-4070-qwen3-14b-q4-k-ctx4096",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 4070 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 2099.9,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4070, Qwen3 14B (Q4_K), 4k Ctx: generation 42.5 t/s, prompt processing 2,099.9 t/s",
   "ram_gb": 12.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4070/",
   "tps": 42.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 4070",
   "id": "hardwarecorner-rtx-4070-qwen3-14b-q4-k-ctx16384",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 4070 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 1355.8,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4070, Qwen3 14B (Q4_K), 16k Ctx: generation 32.7 t/s, prompt processing 1,355.8 t/s",
   "ram_gb": 12.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4070/",
   "tps": 32.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 4070 SUPER",
   "id": "hardwarecorner-rtx-4070-super-qwen3-8b-q4-k-ctx4096",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 4070 SUPER hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 4321.7,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4070 SUPER, Qwen3 8B (Q4_K), 4k Ctx: generation 75.4 t/s, prompt processing 4,321.7 t/s",
   "ram_gb": 12.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4070-super/",
   "tps": 75.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 4070 SUPER",
   "id": "hardwarecorner-rtx-4070-super-qwen3-8b-q4-k-ctx16384",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 4070 SUPER hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 2525.8,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4070 SUPER, Qwen3 8B (Q4_K), 16k Ctx: generation 56.2 t/s, prompt processing 2,525.8 t/s",
   "ram_gb": 12.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4070-super/",
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 4070 SUPER",
   "id": "hardwarecorner-rtx-4070-super-qwen3-8b-q4-k-ctx32768",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 4070 SUPER hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 1595.7,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4070 SUPER, Qwen3 8B (Q4_K), 32k Ctx: generation 42.2 t/s, prompt processing 1,595.7 t/s",
   "ram_gb": 12.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4070-super/",
   "tps": 42.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 4070 SUPER",
   "id": "hardwarecorner-rtx-4070-super-qwen3-14b-q4-k-ctx4096",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 4070 SUPER hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 2522.2,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4070 SUPER, Qwen3 14B (Q4_K), 4k Ctx: generation 45.5 t/s, prompt processing 2,522.2 t/s",
   "ram_gb": 12.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4070-super/",
   "tps": 45.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 4070 SUPER",
   "id": "hardwarecorner-rtx-4070-super-qwen3-14b-q4-k-ctx16384",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 4070 SUPER hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 1578.1,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4070 SUPER, Qwen3 14B (Q4_K), 16k Ctx: generation 37.2 t/s, prompt processing 1,578.1 t/s",
   "ram_gb": 12.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4070-super/",
   "tps": 37.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 4070 Ti",
   "id": "hardwarecorner-rtx-4070-ti-qwen3-8b-q4-k-ctx4096",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 4070 Ti hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 4877.7,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4070 Ti, Qwen3 8B (Q4_K), 4k Ctx: generation 75.8 t/s, prompt processing 4,877.7 t/s",
   "ram_gb": 12.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4070-ti/",
   "tps": 75.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 4070 Ti",
   "id": "hardwarecorner-rtx-4070-ti-qwen3-8b-q4-k-ctx16384",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 4070 Ti hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 2274.8,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4070 Ti, Qwen3 8B (Q4_K), 16k Ctx: generation 57.6 t/s, prompt processing 2,274.8 t/s",
   "ram_gb": 12.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4070-ti/",
   "tps": 57.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 4070 Ti",
   "id": "hardwarecorner-rtx-4070-ti-qwen3-8b-q4-k-ctx32768",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 4070 Ti hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 1099.7,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4070 Ti, Qwen3 8B (Q4_K), 32k Ctx: generation 42.1 t/s, prompt processing 1,099.7 t/s",
   "ram_gb": 12.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4070-ti/",
   "tps": 42.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 4070 Ti",
   "id": "hardwarecorner-rtx-4070-ti-qwen3-14b-q4-k-ctx4096",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 4070 Ti hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 2866.6,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4070 Ti, Qwen3 14B (Q4_K), 4k Ctx: generation 45.8 t/s, prompt processing 2,866.6 t/s",
   "ram_gb": 12.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4070-ti/",
   "tps": 45.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 4070 Ti",
   "id": "hardwarecorner-rtx-4070-ti-qwen3-14b-q4-k-ctx16384",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 4070 Ti hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 1951.4,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4070 Ti, Qwen3 14B (Q4_K), 16k Ctx: generation 37.8 t/s, prompt processing 1,951.4 t/s",
   "ram_gb": 12.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4070-ti/",
   "tps": 37.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 4070 Ti SUPER",
   "id": "hardwarecorner-rtx-4070-ti-super-qwen3-8b-q4-k-ctx4096",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 4070 Ti SUPER hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 5220.1,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4070 Ti SUPER, Qwen3 8B (Q4_K), 4k Ctx: generation 96.3 t/s, prompt processing 5,220.1 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4070-ti-super/",
   "tps": 96.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 4070 Ti SUPER",
   "id": "hardwarecorner-rtx-4070-ti-super-qwen3-8b-q4-k-ctx16384",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 4070 Ti SUPER hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 3050.7,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4070 Ti SUPER, Qwen3 8B (Q4_K), 16k Ctx: generation 72.2 t/s, prompt processing 3,050.7 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4070-ti-super/",
   "tps": 72.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 4070 Ti SUPER",
   "id": "hardwarecorner-rtx-4070-ti-super-qwen3-8b-q4-k-ctx32768",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 4070 Ti SUPER hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 1616.6,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4070 Ti SUPER, Qwen3 8B (Q4_K), 32k Ctx: generation 54.5 t/s, prompt processing 1,616.6 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4070-ti-super/",
   "tps": 54.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 65536,
   "date": null,
   "flags": [],
   "hardware": "RTX 4070 Ti SUPER",
   "id": "hardwarecorner-rtx-4070-ti-super-qwen3-8b-q4-k-ctx65536",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 4070 Ti SUPER hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 829.4,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4070 Ti SUPER, Qwen3 8B (Q4_K), 64k Ctx: generation 36.6 t/s, prompt processing 829.4 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4070-ti-super/",
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 4070 Ti SUPER",
   "id": "hardwarecorner-rtx-4070-ti-super-qwen3-14b-q4-k-ctx4096",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 4070 Ti SUPER hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 3048.0,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4070 Ti SUPER, Qwen3 14B (Q4_K), 4k Ctx: generation 58.1 t/s, prompt processing 3,048.0 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4070-ti-super/",
   "tps": 58.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 4070 Ti SUPER",
   "id": "hardwarecorner-rtx-4070-ti-super-qwen3-14b-q4-k-ctx16384",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 4070 Ti SUPER hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 2003.4,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4070 Ti SUPER, Qwen3 14B (Q4_K), 16k Ctx: generation 47.2 t/s, prompt processing 2,003.4 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4070-ti-super/",
   "tps": 47.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 4070 Ti SUPER",
   "id": "hardwarecorner-rtx-4070-ti-super-qwen3-14b-q4-k-ctx32768",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 4070 Ti SUPER hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 1236.9,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4070 Ti SUPER, Qwen3 14B (Q4_K), 32k Ctx: generation 37.7 t/s, prompt processing 1,236.9 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4070-ti-super/",
   "tps": 37.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 4070 Ti SUPER",
   "id": "hardwarecorner-rtx-4070-ti-super-gpt-oss-20b-mxfp4-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 4070 Ti SUPER hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 5628.1,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 4070 Ti SUPER, gpt-oss 20B (MXFP4), 4k Ctx: generation 128.9 t/s, prompt processing 5,628.1 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4070-ti-super/",
   "tps": 128.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 4070 Ti SUPER",
   "id": "hardwarecorner-rtx-4070-ti-super-gpt-oss-20b-mxfp4-ctx16384",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 4070 Ti SUPER hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 4182.7,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 4070 Ti SUPER, gpt-oss 20B (MXFP4), 16k Ctx: generation 113.6 t/s, prompt processing 4,182.7 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4070-ti-super/",
   "tps": 113.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 4070 Ti SUPER",
   "id": "hardwarecorner-rtx-4070-ti-super-gpt-oss-20b-mxfp4-ctx32768",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 4070 Ti SUPER hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 3020.0,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 4070 Ti SUPER, gpt-oss 20B (MXFP4), 32k Ctx: generation 98.7 t/s, prompt processing 3,020.0 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4070-ti-super/",
   "tps": 98.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 65536,
   "date": null,
   "flags": [],
   "hardware": "RTX 4070 Ti SUPER",
   "id": "hardwarecorner-rtx-4070-ti-super-gpt-oss-20b-mxfp4-ctx65536",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 4070 Ti SUPER hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 1716.2,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 4070 Ti SUPER, gpt-oss 20B (MXFP4), 64k Ctx: generation 78.7 t/s, prompt processing 1,716.2 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4070-ti-super/",
   "tps": 78.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 131072,
   "date": null,
   "flags": [],
   "hardware": "RTX 4070 Ti SUPER",
   "id": "hardwarecorner-rtx-4070-ti-super-gpt-oss-20b-mxfp4-ctx131072",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 4070 Ti SUPER hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 993.5,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 4070 Ti SUPER, gpt-oss 20B (MXFP4), 128k Ctx: generation 57.5 t/s, prompt processing 993.5 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4070-ti-super/",
   "tps": 57.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 4080",
   "id": "hardwarecorner-rtx-4080-qwen3-8b-q4-k-ctx4096",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 4080 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 6177.9,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4080, Qwen3 8B (Q4_K), 4k Ctx: generation 102.7 t/s, prompt processing 6,177.9 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4080/",
   "tps": 102.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 4080",
   "id": "hardwarecorner-rtx-4080-qwen3-8b-q4-k-ctx16384",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 4080 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 3809.8,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4080, Qwen3 8B (Q4_K), 16k Ctx: generation 77.9 t/s, prompt processing 3,809.8 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4080/",
   "tps": 77.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 4080",
   "id": "hardwarecorner-rtx-4080-qwen3-8b-q4-k-ctx32768",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 4080 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 1968.4,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4080, Qwen3 8B (Q4_K), 32k Ctx: generation 59.0 t/s, prompt processing 1,968.4 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4080/",
   "tps": 59.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 65536,
   "date": null,
   "flags": [],
   "hardware": "RTX 4080",
   "id": "hardwarecorner-rtx-4080-qwen3-8b-q4-k-ctx65536",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 4080 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 937.9,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4080, Qwen3 8B (Q4_K), 64k Ctx: generation 39.0 t/s, prompt processing 937.9 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4080/",
   "tps": 39.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 4080",
   "id": "hardwarecorner-rtx-4080-qwen3-14b-q4-k-ctx4096",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 4080 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 3574.7,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4080, Qwen3 14B (Q4_K), 4k Ctx: generation 62.0 t/s, prompt processing 3,574.7 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4080/",
   "tps": 62.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 4080",
   "id": "hardwarecorner-rtx-4080-qwen3-14b-q4-k-ctx16384",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 4080 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 2295.1,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4080, Qwen3 14B (Q4_K), 16k Ctx: generation 51.2 t/s, prompt processing 2,295.1 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4080/",
   "tps": 51.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 4080",
   "id": "hardwarecorner-rtx-4080-qwen3-14b-q4-k-ctx32768",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 4080 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 1395.7,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4080, Qwen3 14B (Q4_K), 32k Ctx: generation 40.6 t/s, prompt processing 1,395.7 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4080/",
   "tps": 40.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 4080",
   "id": "hardwarecorner-rtx-4080-gpt-oss-20b-mxfp4-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 4080 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 6218.5,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 4080, gpt-oss 20B (MXFP4), 4k Ctx: generation 136.5 t/s, prompt processing 6,218.5 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4080/",
   "tps": 136.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 4080",
   "id": "hardwarecorner-rtx-4080-gpt-oss-20b-mxfp4-ctx16384",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 4080 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 4329.2,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 4080, gpt-oss 20B (MXFP4), 16k Ctx: generation 120.2 t/s, prompt processing 4,329.2 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4080/",
   "tps": 120.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 4080",
   "id": "hardwarecorner-rtx-4080-gpt-oss-20b-mxfp4-ctx32768",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 4080 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 2831.1,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 4080, gpt-oss 20B (MXFP4), 32k Ctx: generation 106.0 t/s, prompt processing 2,831.1 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4080/",
   "tps": 106.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 65536,
   "date": null,
   "flags": [],
   "hardware": "RTX 4080",
   "id": "hardwarecorner-rtx-4080-gpt-oss-20b-mxfp4-ctx65536",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 4080 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 1559.0,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 4080, gpt-oss 20B (MXFP4), 64k Ctx: generation 81.3 t/s, prompt processing 1,559.0 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4080/",
   "tps": 81.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 131072,
   "date": null,
   "flags": [],
   "hardware": "RTX 4080",
   "id": "hardwarecorner-rtx-4080-gpt-oss-20b-mxfp4-ctx131072",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 4080 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 898.5,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 4080, gpt-oss 20B (MXFP4), 128k Ctx: generation 60.0 t/s, prompt processing 898.5 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4080/",
   "tps": 60.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 4080 SUPER",
   "id": "hardwarecorner-rtx-4080-super-qwen3-8b-q4-k-ctx4096",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 4080 SUPER hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 6137.0,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4080 SUPER, Qwen3 8B (Q4_K), 4k Ctx: generation 104.2 t/s, prompt processing 6,137.0 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4080-super/",
   "tps": 104.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 4080 SUPER",
   "id": "hardwarecorner-rtx-4080-super-qwen3-8b-q4-k-ctx16384",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 4080 SUPER hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 3858.1,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4080 SUPER, Qwen3 8B (Q4_K), 16k Ctx: generation 79.4 t/s, prompt processing 3,858.1 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4080-super/",
   "tps": 79.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 4080 SUPER",
   "id": "hardwarecorner-rtx-4080-super-qwen3-8b-q4-k-ctx32768",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 4080 SUPER hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 2537.1,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4080 SUPER, Qwen3 8B (Q4_K), 32k Ctx: generation 59.5 t/s, prompt processing 2,537.1 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4080-super/",
   "tps": 59.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 65536,
   "date": null,
   "flags": [],
   "hardware": "RTX 4080 SUPER",
   "id": "hardwarecorner-rtx-4080-super-qwen3-8b-q4-k-ctx65536",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 4080 SUPER hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 1501.5,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4080 SUPER, Qwen3 8B (Q4_K), 64k Ctx: generation 39.1 t/s, prompt processing 1,501.5 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4080-super/",
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 4080 SUPER",
   "id": "hardwarecorner-rtx-4080-super-qwen3-14b-q4-k-ctx4096",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 4080 SUPER hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 3745.0,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4080 SUPER, Qwen3 14B (Q4_K), 4k Ctx: generation 64.2 t/s, prompt processing 3,745.0 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4080-super/",
   "tps": 64.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 4080 SUPER",
   "id": "hardwarecorner-rtx-4080-super-qwen3-14b-q4-k-ctx16384",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 4080 SUPER hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 2526.7,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4080 SUPER, Qwen3 14B (Q4_K), 16k Ctx: generation 52.8 t/s, prompt processing 2,526.7 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4080-super/",
   "tps": 52.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 4080 SUPER",
   "id": "hardwarecorner-rtx-4080-super-qwen3-14b-q4-k-ctx32768",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 4080 SUPER hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 1769.3,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4080 SUPER, Qwen3 14B (Q4_K), 32k Ctx: generation 42.6 t/s, prompt processing 1,769.3 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4080-super/",
   "tps": 42.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 4080 SUPER",
   "id": "hardwarecorner-rtx-4080-super-gpt-oss-20b-mxfp4-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 4080 SUPER hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 6364.0,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 4080 SUPER, gpt-oss 20B (MXFP4), 4k Ctx: generation 139.1 t/s, prompt processing 6,364.0 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4080-super/",
   "tps": 139.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 4080 SUPER",
   "id": "hardwarecorner-rtx-4080-super-gpt-oss-20b-mxfp4-ctx16384",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 4080 SUPER hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 4708.7,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 4080 SUPER, gpt-oss 20B (MXFP4), 16k Ctx: generation 123.0 t/s, prompt processing 4,708.7 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4080-super/",
   "tps": 123.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 4080 SUPER",
   "id": "hardwarecorner-rtx-4080-super-gpt-oss-20b-mxfp4-ctx32768",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 4080 SUPER hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 3328.1,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 4080 SUPER, gpt-oss 20B (MXFP4), 32k Ctx: generation 107.2 t/s, prompt processing 3,328.1 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4080-super/",
   "tps": 107.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 65536,
   "date": null,
   "flags": [],
   "hardware": "RTX 4080 SUPER",
   "id": "hardwarecorner-rtx-4080-super-gpt-oss-20b-mxfp4-ctx65536",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 4080 SUPER hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 1961.5,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 4080 SUPER, gpt-oss 20B (MXFP4), 64k Ctx: generation 81.5 t/s, prompt processing 1,961.5 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4080-super/",
   "tps": 81.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 131072,
   "date": null,
   "flags": [],
   "hardware": "RTX 4080 SUPER",
   "id": "hardwarecorner-rtx-4080-super-gpt-oss-20b-mxfp4-ctx131072",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 4080 SUPER hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 1145.0,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 4080 SUPER, gpt-oss 20B (MXFP4), 128k Ctx: generation 60.0 t/s, prompt processing 1,145.0 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4080-super/",
   "tps": 60.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 4090",
   "id": "hardwarecorner-rtx-4090-qwen3-8b-q4-k-ctx4096",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 4090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 9250.5,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4090, Qwen3 8B (Q4_K), 4k Ctx: generation 141.3 t/s, prompt processing 9,250.5 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4090/",
   "tps": 141.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 4090",
   "id": "hardwarecorner-rtx-4090-qwen3-8b-q4-k-ctx16384",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 4090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 5530.5,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4090, Qwen3 8B (Q4_K), 16k Ctx: generation 108.0 t/s, prompt processing 5,530.5 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4090/",
   "tps": 108.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 4090",
   "id": "hardwarecorner-rtx-4090-qwen3-8b-q4-k-ctx32768",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 4090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 3560.1,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4090, Qwen3 8B (Q4_K), 32k Ctx: generation 82.3 t/s, prompt processing 3,560.1 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4090/",
   "tps": 82.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 65536,
   "date": null,
   "flags": [],
   "hardware": "RTX 4090",
   "id": "hardwarecorner-rtx-4090-qwen3-8b-q4-k-ctx65536",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 4090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 2028.6,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4090, Qwen3 8B (Q4_K), 64k Ctx: generation 56.1 t/s, prompt processing 2,028.6 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4090/",
   "tps": 56.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 131072,
   "date": null,
   "flags": [],
   "hardware": "RTX 4090",
   "id": "hardwarecorner-rtx-4090-qwen3-8b-q4-k-ctx131072",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 4090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 1059.9,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4090, Qwen3 8B (Q4_K), 128k Ctx: generation 33.8 t/s, prompt processing 1,059.9 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4090/",
   "tps": 33.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 4090",
   "id": "hardwarecorner-rtx-4090-qwen3-14b-q4-k-ctx4096",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 4090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 5265.4,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4090, Qwen3 14B (Q4_K), 4k Ctx: generation 84.4 t/s, prompt processing 5,265.4 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4090/",
   "tps": 84.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 4090",
   "id": "hardwarecorner-rtx-4090-qwen3-14b-q4-k-ctx16384",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 4090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 3452.3,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4090, Qwen3 14B (Q4_K), 16k Ctx: generation 69.8 t/s, prompt processing 3,452.3 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4090/",
   "tps": 69.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 4090",
   "id": "hardwarecorner-rtx-4090-qwen3-14b-q4-k-ctx32768",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 4090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 2324.2,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4090, Qwen3 14B (Q4_K), 32k Ctx: generation 55.4 t/s, prompt processing 2,324.2 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4090/",
   "tps": 55.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 65536,
   "date": null,
   "flags": [],
   "hardware": "RTX 4090",
   "id": "hardwarecorner-rtx-4090-qwen3-14b-q4-k-ctx65536",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 4090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 1398.3,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4090, Qwen3 14B (Q4_K), 64k Ctx: generation 38.8 t/s, prompt processing 1,398.3 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4090/",
   "tps": 38.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 4090",
   "id": "hardwarecorner-rtx-4090-gpt-oss-20b-mxfp4-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 4090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 8369.3,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 4090, gpt-oss 20B (MXFP4), 4k Ctx: generation 190.6 t/s, prompt processing 8,369.3 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4090/",
   "tps": 190.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 4090",
   "id": "hardwarecorner-rtx-4090-gpt-oss-20b-mxfp4-ctx16384",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 4090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 5856.2,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 4090, gpt-oss 20B (MXFP4), 16k Ctx: generation 163.3 t/s, prompt processing 5,856.2 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4090/",
   "tps": 163.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 4090",
   "id": "hardwarecorner-rtx-4090-gpt-oss-20b-mxfp4-ctx32768",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 4090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 3705.6,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 4090, gpt-oss 20B (MXFP4), 32k Ctx: generation 140.7 t/s, prompt processing 3,705.6 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4090/",
   "tps": 140.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 65536,
   "date": null,
   "flags": [],
   "hardware": "RTX 4090",
   "id": "hardwarecorner-rtx-4090-gpt-oss-20b-mxfp4-ctx65536",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 4090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 2092.2,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 4090, gpt-oss 20B (MXFP4), 64k Ctx: generation 104.0 t/s, prompt processing 2,092.2 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4090/",
   "tps": 104.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 131072,
   "date": null,
   "flags": [],
   "hardware": "RTX 4090",
   "id": "hardwarecorner-rtx-4090-gpt-oss-20b-mxfp4-ctx131072",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 4090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 1114.7,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 4090, gpt-oss 20B (MXFP4), 128k Ctx: generation 70.3 t/s, prompt processing 1,114.7 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4090/",
   "tps": 70.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 4090",
   "id": "hardwarecorner-rtx-4090-qwen3-30b-a3b-q4-k-ctx4096",
   "model": "Qwen3 30B A3B",
   "notes": "Hardware Corner RTX 4090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "30B",
   "power_w": null,
   "pp_tps": 6159.4,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4090, Qwen3 30B A3B (Q4_K), 4k Ctx: generation 207.3 t/s, prompt processing 6,159.4 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4090/",
   "tps": 207.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 4090",
   "id": "hardwarecorner-rtx-4090-qwen3-30b-a3b-q4-k-ctx16384",
   "model": "Qwen3 30B A3B",
   "notes": "Hardware Corner RTX 4090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "30B",
   "power_w": null,
   "pp_tps": 4027.3,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4090, Qwen3 30B A3B (Q4_K), 16k Ctx: generation 139.2 t/s, prompt processing 4,027.3 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4090/",
   "tps": 139.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 4090",
   "id": "hardwarecorner-rtx-4090-qwen3-30b-a3b-q4-k-ctx32768",
   "model": "Qwen3 30B A3B",
   "notes": "Hardware Corner RTX 4090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "30B",
   "power_w": null,
   "pp_tps": 2627.7,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4090, Qwen3 30B A3B (Q4_K), 32k Ctx: generation 105.1 t/s, prompt processing 2,627.7 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4090/",
   "tps": 105.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 65536,
   "date": null,
   "flags": [],
   "hardware": "RTX 4090",
   "id": "hardwarecorner-rtx-4090-qwen3-30b-a3b-q4-k-ctx65536",
   "model": "Qwen3 30B A3B",
   "notes": "Hardware Corner RTX 4090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "30B",
   "power_w": null,
   "pp_tps": 1502.5,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4090, Qwen3 30B A3B (Q4_K), 64k Ctx: generation 68.2 t/s, prompt processing 1,502.5 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4090/",
   "tps": 68.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 4090",
   "id": "hardwarecorner-rtx-4090-qwen3-32b-q4-k-ctx4096",
   "model": "Qwen3 32B",
   "notes": "Hardware Corner RTX 4090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "32B",
   "power_w": null,
   "pp_tps": 2448.0,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4090, Qwen3 32B (Q4_K), 4k Ctx: generation 39.6 t/s, prompt processing 2,448.0 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4090/",
   "tps": 39.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 4090",
   "id": "hardwarecorner-rtx-4090-qwen3-32b-q4-k-ctx16384",
   "model": "Qwen3 32B",
   "notes": "Hardware Corner RTX 4090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "32B",
   "power_w": null,
   "pp_tps": 1707.4,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 4090, Qwen3 32B (Q4_K), 16k Ctx: generation 34.4 t/s, prompt processing 1,707.4 t/s",
   "ram_gb": 24.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4090/",
   "tps": 34.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 5060 Ti 16GB",
   "id": "hardwarecorner-rtx-5060-ti-16gb-qwen3-8b-q4-k-ctx4096",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 5060 Ti 16GB hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 2965.1,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5060 Ti 16GB, Qwen3 8B (Q4_K), 4k Ctx: generation 69.2 t/s, prompt processing 2,965.1 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5060-ti-16gb/",
   "tps": 69.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 5060 Ti 16GB",
   "id": "hardwarecorner-rtx-5060-ti-16gb-qwen3-8b-q4-k-ctx16384",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 5060 Ti 16GB hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 1447.9,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5060 Ti 16GB, Qwen3 8B (Q4_K), 16k Ctx: generation 51.4 t/s, prompt processing 1,447.9 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5060-ti-16gb/",
   "tps": 51.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 5060 Ti 16GB",
   "id": "hardwarecorner-rtx-5060-ti-16gb-qwen3-8b-q4-k-ctx32768",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 5060 Ti 16GB hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 915.1,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5060 Ti 16GB, Qwen3 8B (Q4_K), 32k Ctx: generation 38.9 t/s, prompt processing 915.1 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5060-ti-16gb/",
   "tps": 38.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 65536,
   "date": null,
   "flags": [],
   "hardware": "RTX 5060 Ti 16GB",
   "id": "hardwarecorner-rtx-5060-ti-16gb-qwen3-8b-q4-k-ctx65536",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 5060 Ti 16GB hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 529.8,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5060 Ti 16GB, Qwen3 8B (Q4_K), 64k Ctx: generation 25.8 t/s, prompt processing 529.8 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5060-ti-16gb/",
   "tps": 25.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 5060 Ti 16GB",
   "id": "hardwarecorner-rtx-5060-ti-16gb-qwen3-14b-q4-k-ctx4096",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 5060 Ti 16GB hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 1743.0,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5060 Ti 16GB, Qwen3 14B (Q4_K), 4k Ctx: generation 41.1 t/s, prompt processing 1,743.0 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5060-ti-16gb/",
   "tps": 41.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 5060 Ti 16GB",
   "id": "hardwarecorner-rtx-5060-ti-16gb-qwen3-14b-q4-k-ctx16384",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 5060 Ti 16GB hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 942.6,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5060 Ti 16GB, Qwen3 14B (Q4_K), 16k Ctx: generation 32.9 t/s, prompt processing 942.6 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5060-ti-16gb/",
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 5060 Ti 16GB",
   "id": "hardwarecorner-rtx-5060-ti-16gb-qwen3-14b-q4-k-ctx32768",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 5060 Ti 16GB hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 621.0,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5060 Ti 16GB, Qwen3 14B (Q4_K), 32k Ctx: generation 25.9 t/s, prompt processing 621.0 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5060-ti-16gb/",
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 5060 Ti 16GB",
   "id": "hardwarecorner-rtx-5060-ti-16gb-gpt-oss-20b-mxfp4-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 5060 Ti 16GB hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 3585.2,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 5060 Ti 16GB, gpt-oss 20B (MXFP4), 4k Ctx: generation 92.1 t/s, prompt processing 3,585.2 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5060-ti-16gb/",
   "tps": 92.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 5060 Ti 16GB",
   "id": "hardwarecorner-rtx-5060-ti-16gb-gpt-oss-20b-mxfp4-ctx16384",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 5060 Ti 16GB hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 2753.3,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 5060 Ti 16GB, gpt-oss 20B (MXFP4), 16k Ctx: generation 82.4 t/s, prompt processing 2,753.3 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5060-ti-16gb/",
   "tps": 82.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 5060 Ti 16GB",
   "id": "hardwarecorner-rtx-5060-ti-16gb-gpt-oss-20b-mxfp4-ctx32768",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 5060 Ti 16GB hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 1737.7,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 5060 Ti 16GB, gpt-oss 20B (MXFP4), 32k Ctx: generation 73.2 t/s, prompt processing 1,737.7 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5060-ti-16gb/",
   "tps": 73.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 65536,
   "date": null,
   "flags": [],
   "hardware": "RTX 5060 Ti 16GB",
   "id": "hardwarecorner-rtx-5060-ti-16gb-gpt-oss-20b-mxfp4-ctx65536",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 5060 Ti 16GB hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 1102.3,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 5060 Ti 16GB, gpt-oss 20B (MXFP4), 64k Ctx: generation 58.1 t/s, prompt processing 1,102.3 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5060-ti-16gb/",
   "tps": 58.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 131072,
   "date": null,
   "flags": [],
   "hardware": "RTX 5060 Ti 16GB",
   "id": "hardwarecorner-rtx-5060-ti-16gb-gpt-oss-20b-mxfp4-ctx131072",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 5060 Ti 16GB hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 685.3,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 5060 Ti 16GB, gpt-oss 20B (MXFP4), 128k Ctx: generation 43.8 t/s, prompt processing 685.3 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5060-ti-16gb/",
   "tps": 43.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 5070",
   "id": "hardwarecorner-rtx-5070-qwen3-8b-q4-k-ctx4096",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 5070 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 3487.7,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5070, Qwen3 8B (Q4_K), 4k Ctx: generation 85.8 t/s, prompt processing 3,487.7 t/s",
   "ram_gb": 12.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5070/",
   "tps": 85.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 5070",
   "id": "hardwarecorner-rtx-5070-qwen3-8b-q4-k-ctx16384",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 5070 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 1600.8,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5070, Qwen3 8B (Q4_K), 16k Ctx: generation 59.1 t/s, prompt processing 1,600.8 t/s",
   "ram_gb": 12.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5070/",
   "tps": 59.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 5070",
   "id": "hardwarecorner-rtx-5070-qwen3-8b-q4-k-ctx32768",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 5070 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 898.8,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5070, Qwen3 8B (Q4_K), 32k Ctx: generation 43.6 t/s, prompt processing 898.8 t/s",
   "ram_gb": 12.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5070/",
   "tps": 43.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 5070",
   "id": "hardwarecorner-rtx-5070-qwen3-14b-q4-k-ctx4096",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 5070 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 2144.2,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5070, Qwen3 14B (Q4_K), 4k Ctx: generation 54.2 t/s, prompt processing 2,144.2 t/s",
   "ram_gb": 12.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5070/",
   "tps": 54.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 5070",
   "id": "hardwarecorner-rtx-5070-qwen3-14b-q4-k-ctx16384",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 5070 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 1315.2,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5070, Qwen3 14B (Q4_K), 16k Ctx: generation 40.6 t/s, prompt processing 1,315.2 t/s",
   "ram_gb": 12.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5070/",
   "tps": 40.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 5070 Ti",
   "id": "hardwarecorner-rtx-5070-ti-qwen3-8b-q4-k-ctx4096",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 5070 Ti hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 5557.3,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5070 Ti, Qwen3 8B (Q4_K), 4k Ctx: generation 120.5 t/s, prompt processing 5,557.3 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5070-ti/",
   "tps": 120.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 5070 Ti",
   "id": "hardwarecorner-rtx-5070-ti-qwen3-8b-q4-k-ctx16384",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 5070 Ti hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 3653.8,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5070 Ti, Qwen3 8B (Q4_K), 16k Ctx: generation 87.5 t/s, prompt processing 3,653.8 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5070-ti/",
   "tps": 87.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 5070 Ti",
   "id": "hardwarecorner-rtx-5070-ti-qwen3-8b-q4-k-ctx32768",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 5070 Ti hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 2269.0,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5070 Ti, Qwen3 8B (Q4_K), 32k Ctx: generation 63.3 t/s, prompt processing 2,269.0 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5070-ti/",
   "tps": 63.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 65536,
   "date": null,
   "flags": [],
   "hardware": "RTX 5070 Ti",
   "id": "hardwarecorner-rtx-5070-ti-qwen3-8b-q4-k-ctx65536",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 5070 Ti hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 1078.7,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5070 Ti, Qwen3 8B (Q4_K), 64k Ctx: generation 40.4 t/s, prompt processing 1,078.7 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5070-ti/",
   "tps": 40.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 5070 Ti",
   "id": "hardwarecorner-rtx-5070-ti-qwen3-14b-q4-k-ctx4096",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 5070 Ti hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 3264.7,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5070 Ti, Qwen3 14B (Q4_K), 4k Ctx: generation 74.3 t/s, prompt processing 3,264.7 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5070-ti/",
   "tps": 74.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 5070 Ti",
   "id": "hardwarecorner-rtx-5070-ti-qwen3-14b-q4-k-ctx16384",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 5070 Ti hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 2303.9,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5070 Ti, Qwen3 14B (Q4_K), 16k Ctx: generation 58.0 t/s, prompt processing 2,303.9 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5070-ti/",
   "tps": 58.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 5070 Ti",
   "id": "hardwarecorner-rtx-5070-ti-qwen3-14b-q4-k-ctx32768",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 5070 Ti hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 1658.2,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5070 Ti, Qwen3 14B (Q4_K), 32k Ctx: generation 45.5 t/s, prompt processing 1,658.2 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5070-ti/",
   "tps": 45.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 5070 Ti",
   "id": "hardwarecorner-rtx-5070-ti-gpt-oss-20b-mxfp4-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 5070 Ti hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 6178.3,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 5070 Ti, gpt-oss 20B (MXFP4), 4k Ctx: generation 156.0 t/s, prompt processing 6,178.3 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5070-ti/",
   "tps": 156.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 5070 Ti",
   "id": "hardwarecorner-rtx-5070-ti-gpt-oss-20b-mxfp4-ctx16384",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 5070 Ti hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 4940.7,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 5070 Ti, gpt-oss 20B (MXFP4), 16k Ctx: generation 133.1 t/s, prompt processing 4,940.7 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5070-ti/",
   "tps": 133.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 5070 Ti",
   "id": "hardwarecorner-rtx-5070-ti-gpt-oss-20b-mxfp4-ctx32768",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 5070 Ti hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 3903.5,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 5070 Ti, gpt-oss 20B (MXFP4), 32k Ctx: generation 115.9 t/s, prompt processing 3,903.5 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5070-ti/",
   "tps": 115.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 65536,
   "date": null,
   "flags": [],
   "hardware": "RTX 5070 Ti",
   "id": "hardwarecorner-rtx-5070-ti-gpt-oss-20b-mxfp4-ctx65536",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 5070 Ti hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 2644.9,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 5070 Ti, gpt-oss 20B (MXFP4), 64k Ctx: generation 91.3 t/s, prompt processing 2,644.9 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5070-ti/",
   "tps": 91.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 131072,
   "date": null,
   "flags": [],
   "hardware": "RTX 5070 Ti",
   "id": "hardwarecorner-rtx-5070-ti-gpt-oss-20b-mxfp4-ctx131072",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 5070 Ti hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 1363.7,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 5070 Ti, gpt-oss 20B (MXFP4), 128k Ctx: generation 66.6 t/s, prompt processing 1,363.7 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5070-ti/",
   "tps": 66.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 5080",
   "id": "hardwarecorner-rtx-5080-qwen3-8b-q4-k-ctx4096",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 5080 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 6410.1,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5080, Qwen3 8B (Q4_K), 4k Ctx: generation 129.1 t/s, prompt processing 6,410.1 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5080/",
   "tps": 129.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 5080",
   "id": "hardwarecorner-rtx-5080-qwen3-8b-q4-k-ctx16384",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 5080 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 4024.4,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5080, Qwen3 8B (Q4_K), 16k Ctx: generation 94.1 t/s, prompt processing 4,024.4 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5080/",
   "tps": 94.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 5080",
   "id": "hardwarecorner-rtx-5080-qwen3-8b-q4-k-ctx32768",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 5080 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 1943.3,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5080, Qwen3 8B (Q4_K), 32k Ctx: generation 72.5 t/s, prompt processing 1,943.3 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5080/",
   "tps": 72.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 65536,
   "date": null,
   "flags": [],
   "hardware": "RTX 5080",
   "id": "hardwarecorner-rtx-5080-qwen3-8b-q4-k-ctx65536",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 5080 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 1124.6,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5080, Qwen3 8B (Q4_K), 64k Ctx: generation 49.0 t/s, prompt processing 1,124.6 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5080/",
   "tps": 49.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 5080",
   "id": "hardwarecorner-rtx-5080-qwen3-14b-q4-k-ctx4096",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 5080 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 3820.5,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5080, Qwen3 14B (Q4_K), 4k Ctx: generation 80.6 t/s, prompt processing 3,820.5 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5080/",
   "tps": 80.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 5080",
   "id": "hardwarecorner-rtx-5080-qwen3-14b-q4-k-ctx16384",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 5080 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 2542.0,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5080, Qwen3 14B (Q4_K), 16k Ctx: generation 64.0 t/s, prompt processing 2,542.0 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5080/",
   "tps": 64.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 5080",
   "id": "hardwarecorner-rtx-5080-qwen3-14b-q4-k-ctx32768",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 5080 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 1326.1,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5080, Qwen3 14B (Q4_K), 32k Ctx: generation 51.9 t/s, prompt processing 1,326.1 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5080/",
   "tps": 51.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 5080",
   "id": "hardwarecorner-rtx-5080-gpt-oss-20b-mxfp4-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 5080 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 5080, gpt-oss 20B (MXFP4), 4k Ctx: generation 172.4 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5080/",
   "tps": 172.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 5080",
   "id": "hardwarecorner-rtx-5080-gpt-oss-20b-mxfp4-ctx16384",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 5080 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 5080, gpt-oss 20B (MXFP4), 16k Ctx: generation 161.3 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5080/",
   "tps": 161.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 5080",
   "id": "hardwarecorner-rtx-5080-gpt-oss-20b-mxfp4-ctx32768",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 5080 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 5080, gpt-oss 20B (MXFP4), 32k Ctx: generation 149.3 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5080/",
   "tps": 149.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 65536,
   "date": null,
   "flags": [],
   "hardware": "RTX 5080",
   "id": "hardwarecorner-rtx-5080-gpt-oss-20b-mxfp4-ctx65536",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 5080 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 5080, gpt-oss 20B (MXFP4), 64k Ctx: generation 130.9 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5080/",
   "tps": 130.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 131072,
   "date": null,
   "flags": [],
   "hardware": "RTX 5080",
   "id": "hardwarecorner-rtx-5080-gpt-oss-20b-mxfp4-ctx131072",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 5080 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 5080, gpt-oss 20B (MXFP4), 128k Ctx: generation 105.6 t/s",
   "ram_gb": 16.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5080/",
   "tps": 105.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 5090",
   "id": "hardwarecorner-rtx-5090-qwen3-14b-q4-k-ctx4096",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 5090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 6497.6,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5090, Qwen3 14B (Q4_K), 4k Ctx: generation 123.8 t/s, prompt processing 6,497.6 t/s",
   "ram_gb": 32.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5090/",
   "tps": 123.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 5090",
   "id": "hardwarecorner-rtx-5090-qwen3-14b-q4-k-ctx16384",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 5090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 4473.1,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5090, Qwen3 14B (Q4_K), 16k Ctx: generation 102.7 t/s, prompt processing 4,473.1 t/s",
   "ram_gb": 32.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5090/",
   "tps": 102.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 5090",
   "id": "hardwarecorner-rtx-5090-qwen3-14b-q4-k-ctx32768",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 5090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 2908.4,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5090, Qwen3 14B (Q4_K), 32k Ctx: generation 82.4 t/s, prompt processing 2,908.4 t/s",
   "ram_gb": 32.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5090/",
   "tps": 82.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 65536,
   "date": null,
   "flags": [],
   "hardware": "RTX 5090",
   "id": "hardwarecorner-rtx-5090-qwen3-14b-q4-k-ctx65536",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 5090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 1707.2,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5090, Qwen3 14B (Q4_K), 64k Ctx: generation 57.6 t/s, prompt processing 1,707.2 t/s",
   "ram_gb": 32.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5090/",
   "tps": 57.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 131072,
   "date": null,
   "flags": [],
   "hardware": "RTX 5090",
   "id": "hardwarecorner-rtx-5090-qwen3-14b-q4-k-ctx131072",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 5090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 908.4,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5090, Qwen3 14B (Q4_K), 128k Ctx: generation 37.2 t/s, prompt processing 908.4 t/s",
   "ram_gb": 32.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5090/",
   "tps": 37.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 5090",
   "id": "hardwarecorner-rtx-5090-gpt-oss-20b-q4-k-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 5090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 9443.8,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5090, gpt-oss 20B (Q4_K), 4k Ctx: generation 298.2 t/s, prompt processing 9,443.8 t/s",
   "ram_gb": 32.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5090/",
   "tps": 298.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 5090",
   "id": "hardwarecorner-rtx-5090-gpt-oss-20b-q4-k-ctx16384",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 5090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 7168.1,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5090, gpt-oss 20B (Q4_K), 16k Ctx: generation 249.2 t/s, prompt processing 7,168.1 t/s",
   "ram_gb": 32.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5090/",
   "tps": 249.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 5090",
   "id": "hardwarecorner-rtx-5090-gpt-oss-20b-q4-k-ctx32768",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 5090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 5183.1,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5090, gpt-oss 20B (Q4_K), 32k Ctx: generation 215.1 t/s, prompt processing 5,183.1 t/s",
   "ram_gb": 32.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5090/",
   "tps": 215.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 65536,
   "date": null,
   "flags": [],
   "hardware": "RTX 5090",
   "id": "hardwarecorner-rtx-5090-gpt-oss-20b-q4-k-ctx65536",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 5090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 3019.7,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5090, gpt-oss 20B (Q4_K), 64k Ctx: generation 169.0 t/s, prompt processing 3,019.7 t/s",
   "ram_gb": 32.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5090/",
   "tps": 169.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 131072,
   "date": null,
   "flags": [],
   "hardware": "RTX 5090",
   "id": "hardwarecorner-rtx-5090-gpt-oss-20b-q4-k-ctx131072",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 5090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 1636.4,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5090, gpt-oss 20B (Q4_K), 128k Ctx: generation 112.0 t/s, prompt processing 1,636.4 t/s",
   "ram_gb": 32.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5090/",
   "tps": 112.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 5090",
   "id": "hardwarecorner-rtx-5090-gemma4-26b-q4-k-ctx4096",
   "model": "Gemma4 26B",
   "notes": "Hardware Corner RTX 5090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "26B",
   "power_w": null,
   "pp_tps": 8799.2,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5090, Gemma4 26B (Q4_K), 4k Ctx: generation 180.3 t/s, prompt processing 8,799.2 t/s",
   "ram_gb": 32.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5090/",
   "tps": 180.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 5090",
   "id": "hardwarecorner-rtx-5090-gemma4-26b-q4-k-ctx16384",
   "model": "Gemma4 26B",
   "notes": "Hardware Corner RTX 5090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "26B",
   "power_w": null,
   "pp_tps": 7733.2,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5090, Gemma4 26B (Q4_K), 16k Ctx: generation 167.2 t/s, prompt processing 7,733.2 t/s",
   "ram_gb": 32.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5090/",
   "tps": 167.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 5090",
   "id": "hardwarecorner-rtx-5090-gemma4-26b-q4-k-ctx32768",
   "model": "Gemma4 26B",
   "notes": "Hardware Corner RTX 5090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "26B",
   "power_w": null,
   "pp_tps": 6292.6,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5090, Gemma4 26B (Q4_K), 32k Ctx: generation 159.4 t/s, prompt processing 6,292.6 t/s",
   "ram_gb": 32.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5090/",
   "tps": 159.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 65536,
   "date": null,
   "flags": [],
   "hardware": "RTX 5090",
   "id": "hardwarecorner-rtx-5090-gemma4-26b-q4-k-ctx65536",
   "model": "Gemma4 26B",
   "notes": "Hardware Corner RTX 5090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "26B",
   "power_w": null,
   "pp_tps": 4360.6,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5090, Gemma4 26B (Q4_K), 64k Ctx: generation 149.4 t/s, prompt processing 4,360.6 t/s",
   "ram_gb": 32.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5090/",
   "tps": 149.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 131072,
   "date": null,
   "flags": [],
   "hardware": "RTX 5090",
   "id": "hardwarecorner-rtx-5090-gemma4-26b-q4-k-ctx131072",
   "model": "Gemma4 26B",
   "notes": "Hardware Corner RTX 5090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "26B",
   "power_w": null,
   "pp_tps": 2839.1,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5090, Gemma4 26B (Q4_K), 128k Ctx: generation 130.2 t/s, prompt processing 2,839.1 t/s",
   "ram_gb": 32.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5090/",
   "tps": 130.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 262144,
   "date": null,
   "flags": [],
   "hardware": "RTX 5090",
   "id": "hardwarecorner-rtx-5090-gemma4-26b-q4-k-ctx262144",
   "model": "Gemma4 26B",
   "notes": "Hardware Corner RTX 5090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "26B",
   "power_w": null,
   "pp_tps": 1707.2,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5090, Gemma4 26B (Q4_K), 256k Ctx: generation 106.0 t/s, prompt processing 1,707.2 t/s",
   "ram_gb": 32.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5090/",
   "tps": 106.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 5090",
   "id": "hardwarecorner-rtx-5090-qwen3-5-27b-q4-k-ctx4096",
   "model": "Qwen3.5-27B",
   "notes": "Hardware Corner RTX 5090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "27B",
   "power_w": null,
   "pp_tps": 3004.1,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5090, Qwen3.5 27B (Q4_K), 4k Ctx: generation 58.8 t/s, prompt processing 3,004.1 t/s",
   "ram_gb": 32.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5090/",
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 5090",
   "id": "hardwarecorner-rtx-5090-qwen3-5-27b-q4-k-ctx16384",
   "model": "Qwen3.5-27B",
   "notes": "Hardware Corner RTX 5090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "27B",
   "power_w": null,
   "pp_tps": 2721.5,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5090, Qwen3.5 27B (Q4_K), 16k Ctx: generation 55.8 t/s, prompt processing 2,721.5 t/s",
   "ram_gb": 32.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5090/",
   "tps": 55.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 5090",
   "id": "hardwarecorner-rtx-5090-qwen3-5-27b-q4-k-ctx32768",
   "model": "Qwen3.5-27B",
   "notes": "Hardware Corner RTX 5090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "27B",
   "power_w": null,
   "pp_tps": 2341.8,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5090, Qwen3.5 27B (Q4_K), 32k Ctx: generation 53.8 t/s, prompt processing 2,341.8 t/s",
   "ram_gb": 32.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5090/",
   "tps": 53.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 65536,
   "date": null,
   "flags": [],
   "hardware": "RTX 5090",
   "id": "hardwarecorner-rtx-5090-qwen3-5-27b-q4-k-ctx65536",
   "model": "Qwen3.5-27B",
   "notes": "Hardware Corner RTX 5090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "27B",
   "power_w": null,
   "pp_tps": 1606.4,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5090, Qwen3.5 27B (Q4_K), 64k Ctx: generation 50.1 t/s, prompt processing 1,606.4 t/s",
   "ram_gb": 32.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5090/",
   "tps": 50.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 131072,
   "date": null,
   "flags": [],
   "hardware": "RTX 5090",
   "id": "hardwarecorner-rtx-5090-qwen3-5-27b-q4-k-ctx131072",
   "model": "Qwen3.5-27B",
   "notes": "Hardware Corner RTX 5090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "27B",
   "power_w": null,
   "pp_tps": 1019.8,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5090, Qwen3.5 27B (Q4_K), 128k Ctx: generation 44.1 t/s, prompt processing 1,019.8 t/s",
   "ram_gb": 32.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5090/",
   "tps": 44.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 5090",
   "id": "hardwarecorner-rtx-5090-gemma4-31b-q4-k-ctx4096",
   "model": "Gemma4 31B",
   "notes": "Hardware Corner RTX 5090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "31B",
   "power_w": null,
   "pp_tps": 3395.0,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5090, Gemma4 31B (Q4_K), 4k Ctx: generation 61.1 t/s, prompt processing 3,395.0 t/s",
   "ram_gb": 32.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5090/",
   "tps": 61.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 5090",
   "id": "hardwarecorner-rtx-5090-gemma4-31b-q4-k-ctx16384",
   "model": "Gemma4 31B",
   "notes": "Hardware Corner RTX 5090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "31B",
   "power_w": null,
   "pp_tps": 2794.3,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5090, Gemma4 31B (Q4_K), 16k Ctx: generation 59.2 t/s, prompt processing 2,794.3 t/s",
   "ram_gb": 32.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5090/",
   "tps": 59.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 5090",
   "id": "hardwarecorner-rtx-5090-gemma4-31b-q4-k-ctx32768",
   "model": "Gemma4 31B",
   "notes": "Hardware Corner RTX 5090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "31B",
   "power_w": null,
   "pp_tps": 2229.1,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5090, Gemma4 31B (Q4_K), 32k Ctx: generation 55.4 t/s, prompt processing 2,229.1 t/s",
   "ram_gb": 32.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5090/",
   "tps": 55.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 65536,
   "date": null,
   "flags": [],
   "hardware": "RTX 5090",
   "id": "hardwarecorner-rtx-5090-gemma4-31b-q4-k-ctx65536",
   "model": "Gemma4 31B",
   "notes": "Hardware Corner RTX 5090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "31B",
   "power_w": null,
   "pp_tps": 1459.3,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5090, Gemma4 31B (Q4_K), 64k Ctx: generation 51.6 t/s, prompt processing 1,459.3 t/s",
   "ram_gb": 32.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5090/",
   "tps": 51.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 131072,
   "date": null,
   "flags": [],
   "hardware": "RTX 5090",
   "id": "hardwarecorner-rtx-5090-gemma4-31b-q4-k-ctx131072",
   "model": "Gemma4 31B",
   "notes": "Hardware Corner RTX 5090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "31B",
   "power_w": null,
   "pp_tps": 900.2,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5090, Gemma4 31B (Q4_K), 128k Ctx: generation 43.4 t/s, prompt processing 900.2 t/s",
   "ram_gb": 32.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5090/",
   "tps": 43.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 5090",
   "id": "hardwarecorner-rtx-5090-qwen3-32b-q4-k-ctx4096",
   "model": "Qwen3 32B",
   "notes": "Hardware Corner RTX 5090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "32B",
   "power_w": null,
   "pp_tps": 2931.3,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5090, Qwen3 32B (Q4_K), 4k Ctx: generation 61.4 t/s, prompt processing 2,931.3 t/s",
   "ram_gb": 32.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5090/",
   "tps": 61.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 5090",
   "id": "hardwarecorner-rtx-5090-qwen3-32b-q4-k-ctx16384",
   "model": "Qwen3 32B",
   "notes": "Hardware Corner RTX 5090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "32B",
   "power_w": null,
   "pp_tps": 2077.2,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5090, Qwen3 32B (Q4_K), 16k Ctx: generation 50.9 t/s, prompt processing 2,077.2 t/s",
   "ram_gb": 32.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5090/",
   "tps": 50.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 5090",
   "id": "hardwarecorner-rtx-5090-qwen3-32b-q4-k-ctx32768",
   "model": "Qwen3 32B",
   "notes": "Hardware Corner RTX 5090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "32B",
   "power_w": null,
   "pp_tps": 1451.1,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 5090, Qwen3 32B (Q4_K), 32k Ctx: generation 43.8 t/s, prompt processing 1,451.1 t/s",
   "ram_gb": 32.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5090/",
   "tps": 43.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 5090",
   "id": "hardwarecorner-rtx-5090-qwen3-5-35b-mxfp4-ctx4096",
   "model": "Qwen3.5 35B",
   "notes": "Hardware Corner RTX 5090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "35B",
   "power_w": null,
   "pp_tps": 6605.2,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 5090, Qwen3.5 35B (MXFP4), 4k Ctx: generation 165.2 t/s, prompt processing 6,605.2 t/s",
   "ram_gb": 32.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5090/",
   "tps": 165.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 5090",
   "id": "hardwarecorner-rtx-5090-qwen3-5-35b-mxfp4-ctx16384",
   "model": "Qwen3.5 35B",
   "notes": "Hardware Corner RTX 5090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "35B",
   "power_w": null,
   "pp_tps": 6142.3,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 5090, Qwen3.5 35B (MXFP4), 16k Ctx: generation 148.3 t/s, prompt processing 6,142.3 t/s",
   "ram_gb": 32.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5090/",
   "tps": 148.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 5090",
   "id": "hardwarecorner-rtx-5090-qwen3-5-35b-mxfp4-ctx32768",
   "model": "Qwen3.5 35B",
   "notes": "Hardware Corner RTX 5090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "35B",
   "power_w": null,
   "pp_tps": 5611.4,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 5090, Qwen3.5 35B (MXFP4), 32k Ctx: generation 143.2 t/s, prompt processing 5,611.4 t/s",
   "ram_gb": 32.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5090/",
   "tps": 143.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 65536,
   "date": null,
   "flags": [],
   "hardware": "RTX 5090",
   "id": "hardwarecorner-rtx-5090-qwen3-5-35b-mxfp4-ctx65536",
   "model": "Qwen3.5 35B",
   "notes": "Hardware Corner RTX 5090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "35B",
   "power_w": null,
   "pp_tps": 4624.5,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 5090, Qwen3.5 35B (MXFP4), 64k Ctx: generation 133.5 t/s, prompt processing 4,624.5 t/s",
   "ram_gb": 32.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5090/",
   "tps": 133.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 131072,
   "date": null,
   "flags": [],
   "hardware": "RTX 5090",
   "id": "hardwarecorner-rtx-5090-qwen3-5-35b-mxfp4-ctx131072",
   "model": "Qwen3.5 35B",
   "notes": "Hardware Corner RTX 5090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "35B",
   "power_w": null,
   "pp_tps": 3242.6,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 5090, Qwen3.5 35B (MXFP4), 128k Ctx: generation 118.2 t/s, prompt processing 3,242.6 t/s",
   "ram_gb": 32.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5090/",
   "tps": 118.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 262144,
   "date": null,
   "flags": [],
   "hardware": "RTX 5090",
   "id": "hardwarecorner-rtx-5090-qwen3-5-35b-mxfp4-ctx262144",
   "model": "Qwen3.5 35B",
   "notes": "Hardware Corner RTX 5090 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "35B",
   "power_w": null,
   "pp_tps": 2003.7,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 5090, Qwen3.5 35B (MXFP4), 256k Ctx: generation 97.3 t/s, prompt processing 2,003.7 t/s",
   "ram_gb": 32.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-5090/",
   "tps": 97.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 6000 Ada",
   "id": "hardwarecorner-rtx-6000-ada-qwen3-8b-q4-k-ctx4096",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 6000 Ada hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 4932.5,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 6000 Ada, Qwen3 8B (Q4_K), 4k Ctx: generation 133.9 t/s, prompt processing 4,932.5 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-6000-ada/",
   "tps": 133.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 6000 Ada",
   "id": "hardwarecorner-rtx-6000-ada-qwen3-8b-q4-k-ctx16384",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 6000 Ada hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 4096.4,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 6000 Ada, Qwen3 8B (Q4_K), 16k Ctx: generation 98.7 t/s, prompt processing 4,096.4 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-6000-ada/",
   "tps": 98.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 6000 Ada",
   "id": "hardwarecorner-rtx-6000-ada-qwen3-8b-q4-k-ctx32768",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 6000 Ada hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 2218.5,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 6000 Ada, Qwen3 8B (Q4_K), 32k Ctx: generation 68.9 t/s, prompt processing 2,218.5 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-6000-ada/",
   "tps": 68.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 65536,
   "date": null,
   "flags": [],
   "hardware": "RTX 6000 Ada",
   "id": "hardwarecorner-rtx-6000-ada-qwen3-8b-q4-k-ctx65536",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 6000 Ada hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 1152.6,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 6000 Ada, Qwen3 8B (Q4_K), 64k Ctx: generation 39.9 t/s, prompt processing 1,152.6 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-6000-ada/",
   "tps": 39.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 131072,
   "date": null,
   "flags": [],
   "hardware": "RTX 6000 Ada",
   "id": "hardwarecorner-rtx-6000-ada-qwen3-8b-q4-k-ctx131072",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX 6000 Ada hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 541.3,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 6000 Ada, Qwen3 8B (Q4_K), 128k Ctx: generation 20.8 t/s, prompt processing 541.3 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-6000-ada/",
   "tps": 20.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 6000 Ada",
   "id": "hardwarecorner-rtx-6000-ada-qwen3-14b-q4-k-ctx4096",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 6000 Ada hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 2912.1,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 6000 Ada, Qwen3 14B (Q4_K), 4k Ctx: generation 81.2 t/s, prompt processing 2,912.1 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-6000-ada/",
   "tps": 81.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 6000 Ada",
   "id": "hardwarecorner-rtx-6000-ada-qwen3-14b-q4-k-ctx16384",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 6000 Ada hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 2367.2,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 6000 Ada, Qwen3 14B (Q4_K), 16k Ctx: generation 58.5 t/s, prompt processing 2,367.2 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-6000-ada/",
   "tps": 58.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 6000 Ada",
   "id": "hardwarecorner-rtx-6000-ada-qwen3-14b-q4-k-ctx32768",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 6000 Ada hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 1310.2,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 6000 Ada, Qwen3 14B (Q4_K), 32k Ctx: generation 42.3 t/s, prompt processing 1,310.2 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-6000-ada/",
   "tps": 42.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 65536,
   "date": null,
   "flags": [],
   "hardware": "RTX 6000 Ada",
   "id": "hardwarecorner-rtx-6000-ada-qwen3-14b-q4-k-ctx65536",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 6000 Ada hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 727.4,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 6000 Ada, Qwen3 14B (Q4_K), 64k Ctx: generation 27.0 t/s, prompt processing 727.4 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-6000-ada/",
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 131072,
   "date": null,
   "flags": [],
   "hardware": "RTX 6000 Ada",
   "id": "hardwarecorner-rtx-6000-ada-qwen3-14b-q4-k-ctx131072",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX 6000 Ada hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 358.6,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 6000 Ada, Qwen3 14B (Q4_K), 128k Ctx: generation 15.2 t/s, prompt processing 358.6 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-6000-ada/",
   "tps": 15.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 6000 Ada",
   "id": "hardwarecorner-rtx-6000-ada-gpt-oss-20b-mxfp4-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 6000 Ada hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 6495.8,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 6000 Ada, gpt-oss 20B (MXFP4), 4k Ctx: generation 160.8 t/s, prompt processing 6,495.8 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-6000-ada/",
   "tps": 160.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 6000 Ada",
   "id": "hardwarecorner-rtx-6000-ada-gpt-oss-20b-mxfp4-ctx16384",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 6000 Ada hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 5350.5,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 6000 Ada, gpt-oss 20B (MXFP4), 16k Ctx: generation 137.1 t/s, prompt processing 5,350.5 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-6000-ada/",
   "tps": 137.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 6000 Ada",
   "id": "hardwarecorner-rtx-6000-ada-gpt-oss-20b-mxfp4-ctx32768",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 6000 Ada hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 3818.7,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 6000 Ada, gpt-oss 20B (MXFP4), 32k Ctx: generation 116.3 t/s, prompt processing 3,818.7 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-6000-ada/",
   "tps": 116.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 65536,
   "date": null,
   "flags": [],
   "hardware": "RTX 6000 Ada",
   "id": "hardwarecorner-rtx-6000-ada-gpt-oss-20b-mxfp4-ctx65536",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 6000 Ada hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 2270.3,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 6000 Ada, gpt-oss 20B (MXFP4), 64k Ctx: generation 88.4 t/s, prompt processing 2,270.3 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-6000-ada/",
   "tps": 88.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 131072,
   "date": null,
   "flags": [],
   "hardware": "RTX 6000 Ada",
   "id": "hardwarecorner-rtx-6000-ada-gpt-oss-20b-mxfp4-ctx131072",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX 6000 Ada hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 1177.7,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX 6000 Ada, gpt-oss 20B (MXFP4), 128k Ctx: generation 57.5 t/s, prompt processing 1,177.7 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-6000-ada/",
   "tps": 57.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 6000 Ada",
   "id": "hardwarecorner-rtx-6000-ada-qwen3-30b-a3b-q4-k-ctx4096",
   "model": "Qwen3 30B A3B",
   "notes": "Hardware Corner RTX 6000 Ada hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "30B",
   "power_w": null,
   "pp_tps": 3888.0,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 6000 Ada, Qwen3 30B A3B (Q4_K), 4k Ctx: generation 199.6 t/s, prompt processing 3,888.0 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-6000-ada/",
   "tps": 199.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 6000 Ada",
   "id": "hardwarecorner-rtx-6000-ada-qwen3-30b-a3b-q4-k-ctx16384",
   "model": "Qwen3 30B A3B",
   "notes": "Hardware Corner RTX 6000 Ada hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "30B",
   "power_w": null,
   "pp_tps": 3156.5,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 6000 Ada, Qwen3 30B A3B (Q4_K), 16k Ctx: generation 120.1 t/s, prompt processing 3,156.5 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-6000-ada/",
   "tps": 120.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 6000 Ada",
   "id": "hardwarecorner-rtx-6000-ada-qwen3-30b-a3b-q4-k-ctx32768",
   "model": "Qwen3 30B A3B",
   "notes": "Hardware Corner RTX 6000 Ada hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "30B",
   "power_w": null,
   "pp_tps": 2010.6,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 6000 Ada, Qwen3 30B A3B (Q4_K), 32k Ctx: generation 74.3 t/s, prompt processing 2,010.6 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-6000-ada/",
   "tps": 74.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 65536,
   "date": null,
   "flags": [],
   "hardware": "RTX 6000 Ada",
   "id": "hardwarecorner-rtx-6000-ada-qwen3-30b-a3b-q4-k-ctx65536",
   "model": "Qwen3 30B A3B",
   "notes": "Hardware Corner RTX 6000 Ada hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "30B",
   "power_w": null,
   "pp_tps": 913.2,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 6000 Ada, Qwen3 30B A3B (Q4_K), 64k Ctx: generation 40.8 t/s, prompt processing 913.2 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-6000-ada/",
   "tps": 40.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 131072,
   "date": null,
   "flags": [],
   "hardware": "RTX 6000 Ada",
   "id": "hardwarecorner-rtx-6000-ada-qwen3-30b-a3b-q4-k-ctx131072",
   "model": "Qwen3 30B A3B",
   "notes": "Hardware Corner RTX 6000 Ada hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "30B",
   "power_w": null,
   "pp_tps": 448.6,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 6000 Ada, Qwen3 30B A3B (Q4_K), 128k Ctx: generation 20.7 t/s, prompt processing 448.6 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-6000-ada/",
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 6000 Ada",
   "id": "hardwarecorner-rtx-6000-ada-qwen3-32b-q4-k-ctx4096",
   "model": "Qwen3 32B",
   "notes": "Hardware Corner RTX 6000 Ada hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "32B",
   "power_w": null,
   "pp_tps": 1203.9,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 6000 Ada, Qwen3 32B (Q4_K), 4k Ctx: generation 37.4 t/s, prompt processing 1,203.9 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-6000-ada/",
   "tps": 37.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 6000 Ada",
   "id": "hardwarecorner-rtx-6000-ada-qwen3-32b-q4-k-ctx16384",
   "model": "Qwen3 32B",
   "notes": "Hardware Corner RTX 6000 Ada hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "32B",
   "power_w": null,
   "pp_tps": 977.5,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 6000 Ada, Qwen3 32B (Q4_K), 16k Ctx: generation 25.1 t/s, prompt processing 977.5 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-6000-ada/",
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX 6000 Ada",
   "id": "hardwarecorner-rtx-6000-ada-qwen3-32b-q4-k-ctx32768",
   "model": "Qwen3 32B",
   "notes": "Hardware Corner RTX 6000 Ada hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "32B",
   "power_w": null,
   "pp_tps": 595.5,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 6000 Ada, Qwen3 32B (Q4_K), 32k Ctx: generation 18.2 t/s, prompt processing 595.5 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-6000-ada/",
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 65536,
   "date": null,
   "flags": [],
   "hardware": "RTX 6000 Ada",
   "id": "hardwarecorner-rtx-6000-ada-qwen3-32b-q4-k-ctx65536",
   "model": "Qwen3 32B",
   "notes": "Hardware Corner RTX 6000 Ada hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "32B",
   "power_w": null,
   "pp_tps": 348.6,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 6000 Ada, Qwen3 32B (Q4_K), 64k Ctx: generation 11.8 t/s, prompt processing 348.6 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-6000-ada/",
   "tps": 11.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX 6000 Ada",
   "id": "hardwarecorner-rtx-6000-ada-llama-3-3-70b-q4-k-ctx4096",
   "model": "Llama 3.3 70B",
   "notes": "Hardware Corner RTX 6000 Ada hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "70B",
   "power_w": null,
   "pp_tps": 678.1,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 6000 Ada, Llama 3.3 70B (Q4_K), 4k Ctx: generation 18.1 t/s, prompt processing 678.1 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-6000-ada/",
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX 6000 Ada",
   "id": "hardwarecorner-rtx-6000-ada-llama-3-3-70b-q4-k-ctx16384",
   "model": "Llama 3.3 70B",
   "notes": "Hardware Corner RTX 6000 Ada hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "70B",
   "power_w": null,
   "pp_tps": 526.0,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX 6000 Ada, Llama 3.3 70B (Q4_K), 16k Ctx: generation 13.7 t/s, prompt processing 526.0 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-6000-ada/",
   "tps": 13.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX A6000",
   "id": "hardwarecorner-rtx-a6000-qwen3-8b-q4-k-ctx4096",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX A6000 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 2756.0,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX A6000, Qwen3 8B (Q4_K), 4k Ctx: generation 90.5 t/s, prompt processing 2,756.0 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-a6000/",
   "tps": 90.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX A6000",
   "id": "hardwarecorner-rtx-a6000-qwen3-8b-q4-k-ctx16384",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX A6000 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 2428.0,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX A6000, Qwen3 8B (Q4_K), 16k Ctx: generation 64.3 t/s, prompt processing 2,428.0 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-a6000/",
   "tps": 64.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX A6000",
   "id": "hardwarecorner-rtx-a6000-qwen3-8b-q4-k-ctx32768",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX A6000 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 1593.3,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX A6000, Qwen3 8B (Q4_K), 32k Ctx: generation 44.6 t/s, prompt processing 1,593.3 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-a6000/",
   "tps": 44.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 65536,
   "date": null,
   "flags": [],
   "hardware": "RTX A6000",
   "id": "hardwarecorner-rtx-a6000-qwen3-8b-q4-k-ctx65536",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX A6000 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 951.0,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX A6000, Qwen3 8B (Q4_K), 64k Ctx: generation 27.9 t/s, prompt processing 951.0 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-a6000/",
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 131072,
   "date": null,
   "flags": [],
   "hardware": "RTX A6000",
   "id": "hardwarecorner-rtx-a6000-qwen3-8b-q4-k-ctx131072",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX A6000 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 527.9,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX A6000, Qwen3 8B (Q4_K), 128k Ctx: generation 16.0 t/s, prompt processing 527.9 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-a6000/",
   "tps": 16.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX A6000",
   "id": "hardwarecorner-rtx-a6000-qwen3-14b-q4-k-ctx4096",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX A6000 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 1713.3,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX A6000, Qwen3 14B (Q4_K), 4k Ctx: generation 53.8 t/s, prompt processing 1,713.3 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-a6000/",
   "tps": 53.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX A6000",
   "id": "hardwarecorner-rtx-a6000-qwen3-14b-q4-k-ctx16384",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX A6000 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 1529.0,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX A6000, Qwen3 14B (Q4_K), 16k Ctx: generation 40.7 t/s, prompt processing 1,529.0 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-a6000/",
   "tps": 40.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX A6000",
   "id": "hardwarecorner-rtx-a6000-qwen3-14b-q4-k-ctx32768",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX A6000 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 1063.4,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX A6000, Qwen3 14B (Q4_K), 32k Ctx: generation 30.8 t/s, prompt processing 1,063.4 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-a6000/",
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 65536,
   "date": null,
   "flags": [],
   "hardware": "RTX A6000",
   "id": "hardwarecorner-rtx-a6000-qwen3-14b-q4-k-ctx65536",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX A6000 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 668.9,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX A6000, Qwen3 14B (Q4_K), 64k Ctx: generation 20.5 t/s, prompt processing 668.9 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-a6000/",
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 131072,
   "date": null,
   "flags": [],
   "hardware": "RTX A6000",
   "id": "hardwarecorner-rtx-a6000-qwen3-14b-q4-k-ctx131072",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX A6000 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 382.4,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX A6000, Qwen3 14B (Q4_K), 128k Ctx: generation 12.1 t/s, prompt processing 382.4 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-a6000/",
   "tps": 12.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX A6000",
   "id": "hardwarecorner-rtx-a6000-gpt-oss-20b-mxfp4-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX A6000 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 3490.6,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX A6000, gpt-oss 20B (MXFP4), 4k Ctx: generation 108.4 t/s, prompt processing 3,490.6 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-a6000/",
   "tps": 108.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX A6000",
   "id": "hardwarecorner-rtx-a6000-gpt-oss-20b-mxfp4-ctx16384",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX A6000 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 2649.2,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX A6000, gpt-oss 20B (MXFP4), 16k Ctx: generation 91.2 t/s, prompt processing 2,649.2 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-a6000/",
   "tps": 91.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX A6000",
   "id": "hardwarecorner-rtx-a6000-gpt-oss-20b-mxfp4-ctx32768",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX A6000 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 1891.2,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX A6000, gpt-oss 20B (MXFP4), 32k Ctx: generation 74.4 t/s, prompt processing 1,891.2 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-a6000/",
   "tps": 74.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 65536,
   "date": null,
   "flags": [],
   "hardware": "RTX A6000",
   "id": "hardwarecorner-rtx-a6000-gpt-oss-20b-mxfp4-ctx65536",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX A6000 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 1207.2,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX A6000, gpt-oss 20B (MXFP4), 64k Ctx: generation 53.3 t/s, prompt processing 1,207.2 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-a6000/",
   "tps": 53.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 131072,
   "date": null,
   "flags": [],
   "hardware": "RTX A6000",
   "id": "hardwarecorner-rtx-a6000-gpt-oss-20b-mxfp4-ctx131072",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX A6000 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 702.8,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX A6000, gpt-oss 20B (MXFP4), 128k Ctx: generation 33.4 t/s, prompt processing 702.8 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-a6000/",
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX A6000",
   "id": "hardwarecorner-rtx-a6000-qwen3-30b-a3b-q4-k-ctx4096",
   "model": "Qwen3 30B A3B",
   "notes": "Hardware Corner RTX A6000 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "30B",
   "power_w": null,
   "pp_tps": 2142.4,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX A6000, Qwen3 30B A3B (Q4_K), 4k Ctx: generation 131.2 t/s, prompt processing 2,142.4 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-a6000/",
   "tps": 131.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX A6000",
   "id": "hardwarecorner-rtx-a6000-qwen3-30b-a3b-q4-k-ctx16384",
   "model": "Qwen3 30B A3B",
   "notes": "Hardware Corner RTX A6000 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "30B",
   "power_w": null,
   "pp_tps": 1901.1,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX A6000, Qwen3 30B A3B (Q4_K), 16k Ctx: generation 75.4 t/s, prompt processing 1,901.1 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-a6000/",
   "tps": 75.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX A6000",
   "id": "hardwarecorner-rtx-a6000-qwen3-30b-a3b-q4-k-ctx32768",
   "model": "Qwen3 30B A3B",
   "notes": "Hardware Corner RTX A6000 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "30B",
   "power_w": null,
   "pp_tps": 1275.6,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX A6000, Qwen3 30B A3B (Q4_K), 32k Ctx: generation 48.1 t/s, prompt processing 1,275.6 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-a6000/",
   "tps": 48.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 65536,
   "date": null,
   "flags": [],
   "hardware": "RTX A6000",
   "id": "hardwarecorner-rtx-a6000-qwen3-30b-a3b-q4-k-ctx65536",
   "model": "Qwen3 30B A3B",
   "notes": "Hardware Corner RTX A6000 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "30B",
   "power_w": null,
   "pp_tps": 764.2,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX A6000, Qwen3 30B A3B (Q4_K), 64k Ctx: generation 27.6 t/s, prompt processing 764.2 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-a6000/",
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 131072,
   "date": null,
   "flags": [],
   "hardware": "RTX A6000",
   "id": "hardwarecorner-rtx-a6000-qwen3-30b-a3b-q4-k-ctx131072",
   "model": "Qwen3 30B A3B",
   "notes": "Hardware Corner RTX A6000 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "30B",
   "power_w": null,
   "pp_tps": 429.2,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX A6000, Qwen3 30B A3B (Q4_K), 128k Ctx: generation 14.8 t/s, prompt processing 429.2 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-a6000/",
   "tps": 14.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX A6000",
   "id": "hardwarecorner-rtx-a6000-qwen3-32b-q4-k-ctx4096",
   "model": "Qwen3 32B",
   "notes": "Hardware Corner RTX A6000 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "32B",
   "power_w": null,
   "pp_tps": 736.2,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX A6000, Qwen3 32B (Q4_K), 4k Ctx: generation 24.6 t/s, prompt processing 736.2 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-a6000/",
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX A6000",
   "id": "hardwarecorner-rtx-a6000-qwen3-32b-q4-k-ctx16384",
   "model": "Qwen3 32B",
   "notes": "Hardware Corner RTX A6000 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "32B",
   "power_w": null,
   "pp_tps": 704.8,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX A6000, Qwen3 32B (Q4_K), 16k Ctx: generation 18.3 t/s, prompt processing 704.8 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-a6000/",
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX A6000",
   "id": "hardwarecorner-rtx-a6000-qwen3-32b-q4-k-ctx32768",
   "model": "Qwen3 32B",
   "notes": "Hardware Corner RTX A6000 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "32B",
   "power_w": null,
   "pp_tps": 502.7,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX A6000, Qwen3 32B (Q4_K), 32k Ctx: generation 13.5 t/s, prompt processing 502.7 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-a6000/",
   "tps": 13.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 65536,
   "date": null,
   "flags": [],
   "hardware": "RTX A6000",
   "id": "hardwarecorner-rtx-a6000-qwen3-32b-q4-k-ctx65536",
   "model": "Qwen3 32B",
   "notes": "Hardware Corner RTX A6000 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "32B",
   "power_w": null,
   "pp_tps": 319.0,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX A6000, Qwen3 32B (Q4_K), 64k Ctx: generation 8.7 t/s, prompt processing 319.0 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-a6000/",
   "tps": 8.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX A6000",
   "id": "hardwarecorner-rtx-a6000-llama-3-3-70b-q4-k-ctx4096",
   "model": "Llama 3.3 70B",
   "notes": "Hardware Corner RTX A6000 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "70B",
   "power_w": null,
   "pp_tps": 423.0,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX A6000, Llama 3.3 70B (Q4_K), 4k Ctx: generation 12.6 t/s, prompt processing 423.0 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-a6000/",
   "tps": 12.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX A6000",
   "id": "hardwarecorner-rtx-a6000-llama-3-3-70b-q4-k-ctx16384",
   "model": "Llama 3.3 70B",
   "notes": "Hardware Corner RTX A6000 hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "70B",
   "power_w": null,
   "pp_tps": 280.3,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX A6000, Llama 3.3 70B (Q4_K), 16k Ctx: generation 9.1 t/s, prompt processing 280.3 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-a6000/",
   "tps": 9.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX Pro 5000 Blackwell",
   "id": "hardwarecorner-rtx-pro-5000-blackwell-qwen3-8b-q4-k-ctx4096",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX Pro 5000 Blackwell hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 8175.3,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX Pro 5000 Blackwell, Qwen3 8B (Q4_K), 4k Ctx: generation 159.8 t/s, prompt processing 8,175.3 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-pro-5000-blackwell/",
   "tps": 159.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX Pro 5000 Blackwell",
   "id": "hardwarecorner-rtx-pro-5000-blackwell-qwen3-8b-q4-k-ctx16384",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX Pro 5000 Blackwell hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 5846.4,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX Pro 5000 Blackwell, Qwen3 8B (Q4_K), 16k Ctx: generation 126.3 t/s, prompt processing 5,846.4 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-pro-5000-blackwell/",
   "tps": 126.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX Pro 5000 Blackwell",
   "id": "hardwarecorner-rtx-pro-5000-blackwell-qwen3-8b-q4-k-ctx32768",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX Pro 5000 Blackwell hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 4006.2,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX Pro 5000 Blackwell, Qwen3 8B (Q4_K), 32k Ctx: generation 98.3 t/s, prompt processing 4,006.2 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-pro-5000-blackwell/",
   "tps": 98.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 65536,
   "date": null,
   "flags": [],
   "hardware": "RTX Pro 5000 Blackwell",
   "id": "hardwarecorner-rtx-pro-5000-blackwell-qwen3-8b-q4-k-ctx65536",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX Pro 5000 Blackwell hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 2120.7,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX Pro 5000 Blackwell, Qwen3 8B (Q4_K), 64k Ctx: generation 64.1 t/s, prompt processing 2,120.7 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-pro-5000-blackwell/",
   "tps": 64.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 131072,
   "date": null,
   "flags": [],
   "hardware": "RTX Pro 5000 Blackwell",
   "id": "hardwarecorner-rtx-pro-5000-blackwell-qwen3-8b-q4-k-ctx131072",
   "model": "Qwen3 8B",
   "notes": "Hardware Corner RTX Pro 5000 Blackwell hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "8B",
   "power_w": null,
   "pp_tps": 899.1,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX Pro 5000 Blackwell, Qwen3 8B (Q4_K), 128k Ctx: generation 38.1 t/s, prompt processing 899.1 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-pro-5000-blackwell/",
   "tps": 38.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX Pro 5000 Blackwell",
   "id": "hardwarecorner-rtx-pro-5000-blackwell-qwen3-14b-q4-k-ctx4096",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX Pro 5000 Blackwell hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 4515.7,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX Pro 5000 Blackwell, Qwen3 14B (Q4_K), 4k Ctx: generation 98.9 t/s, prompt processing 4,515.7 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-pro-5000-blackwell/",
   "tps": 98.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX Pro 5000 Blackwell",
   "id": "hardwarecorner-rtx-pro-5000-blackwell-qwen3-14b-q4-k-ctx16384",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX Pro 5000 Blackwell hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 3064.2,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX Pro 5000 Blackwell, Qwen3 14B (Q4_K), 16k Ctx: generation 84.3 t/s, prompt processing 3,064.2 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-pro-5000-blackwell/",
   "tps": 84.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX Pro 5000 Blackwell",
   "id": "hardwarecorner-rtx-pro-5000-blackwell-qwen3-14b-q4-k-ctx32768",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX Pro 5000 Blackwell hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 2069.8,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX Pro 5000 Blackwell, Qwen3 14B (Q4_K), 32k Ctx: generation 70.8 t/s, prompt processing 2,069.8 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-pro-5000-blackwell/",
   "tps": 70.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 65536,
   "date": null,
   "flags": [],
   "hardware": "RTX Pro 5000 Blackwell",
   "id": "hardwarecorner-rtx-pro-5000-blackwell-qwen3-14b-q4-k-ctx65536",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX Pro 5000 Blackwell hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 1125.1,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX Pro 5000 Blackwell, Qwen3 14B (Q4_K), 64k Ctx: generation 53.8 t/s, prompt processing 1,125.1 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-pro-5000-blackwell/",
   "tps": 53.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 131072,
   "date": null,
   "flags": [],
   "hardware": "RTX Pro 5000 Blackwell",
   "id": "hardwarecorner-rtx-pro-5000-blackwell-qwen3-14b-q4-k-ctx131072",
   "model": "Qwen3 14B",
   "notes": "Hardware Corner RTX Pro 5000 Blackwell hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "14B",
   "power_w": null,
   "pp_tps": 434.2,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX Pro 5000 Blackwell, Qwen3 14B (Q4_K), 128k Ctx: generation 36.5 t/s, prompt processing 434.2 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-pro-5000-blackwell/",
   "tps": 36.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX Pro 5000 Blackwell",
   "id": "hardwarecorner-rtx-pro-5000-blackwell-gpt-oss-20b-mxfp4-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX Pro 5000 Blackwell hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 10339.6,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX Pro 5000 Blackwell, gpt-oss 20B (MXFP4), 4k Ctx: generation 213.4 t/s, prompt processing 10,339.6 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-pro-5000-blackwell/",
   "tps": 213.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX Pro 5000 Blackwell",
   "id": "hardwarecorner-rtx-pro-5000-blackwell-gpt-oss-20b-mxfp4-ctx16384",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX Pro 5000 Blackwell hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 8747.7,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX Pro 5000 Blackwell, gpt-oss 20B (MXFP4), 16k Ctx: generation 196.2 t/s, prompt processing 8,747.7 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-pro-5000-blackwell/",
   "tps": 196.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX Pro 5000 Blackwell",
   "id": "hardwarecorner-rtx-pro-5000-blackwell-gpt-oss-20b-mxfp4-ctx32768",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX Pro 5000 Blackwell hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 7257.7,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX Pro 5000 Blackwell, gpt-oss 20B (MXFP4), 32k Ctx: generation 182.8 t/s, prompt processing 7,257.7 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-pro-5000-blackwell/",
   "tps": 182.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 65536,
   "date": null,
   "flags": [],
   "hardware": "RTX Pro 5000 Blackwell",
   "id": "hardwarecorner-rtx-pro-5000-blackwell-gpt-oss-20b-mxfp4-ctx65536",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX Pro 5000 Blackwell hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 5273.9,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX Pro 5000 Blackwell, gpt-oss 20B (MXFP4), 64k Ctx: generation 161.4 t/s, prompt processing 5,273.9 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-pro-5000-blackwell/",
   "tps": 161.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 131072,
   "date": null,
   "flags": [],
   "hardware": "RTX Pro 5000 Blackwell",
   "id": "hardwarecorner-rtx-pro-5000-blackwell-gpt-oss-20b-mxfp4-ctx131072",
   "model": "gpt-oss 20B",
   "notes": "Hardware Corner RTX Pro 5000 Blackwell hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "20B",
   "power_w": null,
   "pp_tps": 3163.3,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX Pro 5000 Blackwell, gpt-oss 20B (MXFP4), 128k Ctx: generation 130.9 t/s, prompt processing 3,163.3 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-pro-5000-blackwell/",
   "tps": 130.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX Pro 5000 Blackwell",
   "id": "hardwarecorner-rtx-pro-5000-blackwell-qwen3-30b-a3b-q4-k-ctx4096",
   "model": "Qwen3 30B A3B",
   "notes": "Hardware Corner RTX Pro 5000 Blackwell hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "30B",
   "power_w": null,
   "pp_tps": 5266.1,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX Pro 5000 Blackwell, Qwen3 30B A3B (Q4_K), 4k Ctx: generation 188.8 t/s, prompt processing 5,266.1 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-pro-5000-blackwell/",
   "tps": 188.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX Pro 5000 Blackwell",
   "id": "hardwarecorner-rtx-pro-5000-blackwell-qwen3-30b-a3b-q4-k-ctx16384",
   "model": "Qwen3 30B A3B",
   "notes": "Hardware Corner RTX Pro 5000 Blackwell hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "30B",
   "power_w": null,
   "pp_tps": 4043.8,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX Pro 5000 Blackwell, Qwen3 30B A3B (Q4_K), 16k Ctx: generation 148.5 t/s, prompt processing 4,043.8 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-pro-5000-blackwell/",
   "tps": 148.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX Pro 5000 Blackwell",
   "id": "hardwarecorner-rtx-pro-5000-blackwell-qwen3-30b-a3b-q4-k-ctx32768",
   "model": "Qwen3 30B A3B",
   "notes": "Hardware Corner RTX Pro 5000 Blackwell hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "30B",
   "power_w": null,
   "pp_tps": 3032.7,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX Pro 5000 Blackwell, Qwen3 30B A3B (Q4_K), 32k Ctx: generation 123.2 t/s, prompt processing 3,032.7 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-pro-5000-blackwell/",
   "tps": 123.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 65536,
   "date": null,
   "flags": [],
   "hardware": "RTX Pro 5000 Blackwell",
   "id": "hardwarecorner-rtx-pro-5000-blackwell-qwen3-30b-a3b-q4-k-ctx65536",
   "model": "Qwen3 30B A3B",
   "notes": "Hardware Corner RTX Pro 5000 Blackwell hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "30B",
   "power_w": null,
   "pp_tps": 1708.9,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX Pro 5000 Blackwell, Qwen3 30B A3B (Q4_K), 64k Ctx: generation 91.9 t/s, prompt processing 1,708.9 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-pro-5000-blackwell/",
   "tps": 91.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 131072,
   "date": null,
   "flags": [],
   "hardware": "RTX Pro 5000 Blackwell",
   "id": "hardwarecorner-rtx-pro-5000-blackwell-qwen3-30b-a3b-q4-k-ctx131072",
   "model": "Qwen3 30B A3B",
   "notes": "Hardware Corner RTX Pro 5000 Blackwell hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "30B",
   "power_w": null,
   "pp_tps": 702.7,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX Pro 5000 Blackwell, Qwen3 30B A3B (Q4_K), 128k Ctx: generation 61.4 t/s, prompt processing 702.7 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-pro-5000-blackwell/",
   "tps": 61.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 262144,
   "date": null,
   "flags": [],
   "hardware": "RTX Pro 5000 Blackwell",
   "id": "hardwarecorner-rtx-pro-5000-blackwell-qwen3-30b-a3b-q4-k-ctx262144",
   "model": "Qwen3 30B A3B",
   "notes": "Hardware Corner RTX Pro 5000 Blackwell hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "30B",
   "power_w": null,
   "pp_tps": 364.1,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX Pro 5000 Blackwell, Qwen3 30B A3B (Q4_K), 256k Ctx: generation 37.0 t/s, prompt processing 364.1 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-pro-5000-blackwell/",
   "tps": 37.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX Pro 5000 Blackwell",
   "id": "hardwarecorner-rtx-pro-5000-blackwell-qwen3-32b-q4-k-ctx4096",
   "model": "Qwen3 32B",
   "notes": "Hardware Corner RTX Pro 5000 Blackwell hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "32B",
   "power_w": null,
   "pp_tps": 2030.3,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX Pro 5000 Blackwell, Qwen3 32B (Q4_K), 4k Ctx: generation 46.6 t/s, prompt processing 2,030.3 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-pro-5000-blackwell/",
   "tps": 46.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX Pro 5000 Blackwell",
   "id": "hardwarecorner-rtx-pro-5000-blackwell-qwen3-32b-q4-k-ctx16384",
   "model": "Qwen3 32B",
   "notes": "Hardware Corner RTX Pro 5000 Blackwell hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "32B",
   "power_w": null,
   "pp_tps": 1497.7,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX Pro 5000 Blackwell, Qwen3 32B (Q4_K), 16k Ctx: generation 42.0 t/s, prompt processing 1,497.7 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-pro-5000-blackwell/",
   "tps": 42.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX Pro 5000 Blackwell",
   "id": "hardwarecorner-rtx-pro-5000-blackwell-qwen3-32b-q4-k-ctx32768",
   "model": "Qwen3 32B",
   "notes": "Hardware Corner RTX Pro 5000 Blackwell hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "32B",
   "power_w": null,
   "pp_tps": 1079.8,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX Pro 5000 Blackwell, Qwen3 32B (Q4_K), 32k Ctx: generation 36.4 t/s, prompt processing 1,079.8 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-pro-5000-blackwell/",
   "tps": 36.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 65536,
   "date": null,
   "flags": [],
   "hardware": "RTX Pro 5000 Blackwell",
   "id": "hardwarecorner-rtx-pro-5000-blackwell-qwen3-32b-q4-k-ctx65536",
   "model": "Qwen3 32B",
   "notes": "Hardware Corner RTX Pro 5000 Blackwell hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "32B",
   "power_w": null,
   "pp_tps": 629.6,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX Pro 5000 Blackwell, Qwen3 32B (Q4_K), 64k Ctx: generation 28.9 t/s, prompt processing 629.6 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-pro-5000-blackwell/",
   "tps": 28.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX Pro 5000 Blackwell",
   "id": "hardwarecorner-rtx-pro-5000-blackwell-llama-3-3-70b-q4-k-ctx4096",
   "model": "Llama 3.3 70B",
   "notes": "Hardware Corner RTX Pro 5000 Blackwell hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "70B",
   "power_w": null,
   "pp_tps": 1022.9,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX Pro 5000 Blackwell, Llama 3.3 70B (Q4_K), 4k Ctx: generation 24.0 t/s, prompt processing 1,022.9 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-pro-5000-blackwell/",
   "tps": 24.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX Pro 5000 Blackwell",
   "id": "hardwarecorner-rtx-pro-5000-blackwell-llama-3-3-70b-q4-k-ctx16384",
   "model": "Llama 3.3 70B",
   "notes": "Hardware Corner RTX Pro 5000 Blackwell hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "70B",
   "power_w": null,
   "pp_tps": 826.3,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX Pro 5000 Blackwell, Llama 3.3 70B (Q4_K), 16k Ctx: generation 22.4 t/s, prompt processing 826.3 t/s",
   "ram_gb": 48.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-pro-5000-blackwell/",
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "hardwarecorner-rtx-pro-6000-blackwell-gemma4-26b-q4-k-ctx4096",
   "model": "Gemma4 26B",
   "notes": "Hardware Corner RTX Pro 6000 Blackwell hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "26B",
   "power_w": null,
   "pp_tps": 9437.2,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX Pro 6000 Blackwell, Gemma4 26B (Q4_K), 4k Ctx: generation 196.9 t/s, prompt processing 9,437.2 t/s",
   "ram_gb": 96.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-pro-6000-blackwell/",
   "tps": 196.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "hardwarecorner-rtx-pro-6000-blackwell-gemma4-26b-q4-k-ctx16384",
   "model": "Gemma4 26B",
   "notes": "Hardware Corner RTX Pro 6000 Blackwell hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "26B",
   "power_w": null,
   "pp_tps": 8453.8,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX Pro 6000 Blackwell, Gemma4 26B (Q4_K), 16k Ctx: generation 172.7 t/s, prompt processing 8,453.8 t/s",
   "ram_gb": 96.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-pro-6000-blackwell/",
   "tps": 172.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "hardwarecorner-rtx-pro-6000-blackwell-gemma4-26b-q4-k-ctx32768",
   "model": "Gemma4 26B",
   "notes": "Hardware Corner RTX Pro 6000 Blackwell hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "26B",
   "power_w": null,
   "pp_tps": 7107.6,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX Pro 6000 Blackwell, Gemma4 26B (Q4_K), 32k Ctx: generation 170.3 t/s, prompt processing 7,107.6 t/s",
   "ram_gb": 96.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-pro-6000-blackwell/",
   "tps": 170.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 65536,
   "date": null,
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "hardwarecorner-rtx-pro-6000-blackwell-gemma4-26b-q4-k-ctx65536",
   "model": "Gemma4 26B",
   "notes": "Hardware Corner RTX Pro 6000 Blackwell hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "26B",
   "power_w": null,
   "pp_tps": 5379.6,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX Pro 6000 Blackwell, Gemma4 26B (Q4_K), 64k Ctx: generation 161.0 t/s, prompt processing 5,379.6 t/s",
   "ram_gb": 96.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-pro-6000-blackwell/",
   "tps": 161.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 131072,
   "date": null,
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "hardwarecorner-rtx-pro-6000-blackwell-gemma4-26b-q4-k-ctx131072",
   "model": "Gemma4 26B",
   "notes": "Hardware Corner RTX Pro 6000 Blackwell hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "26B",
   "power_w": null,
   "pp_tps": 3667.8,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX Pro 6000 Blackwell, Gemma4 26B (Q4_K), 128k Ctx: generation 133.2 t/s, prompt processing 3,667.8 t/s",
   "ram_gb": 96.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-pro-6000-blackwell/",
   "tps": 133.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 262144,
   "date": null,
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "hardwarecorner-rtx-pro-6000-blackwell-gemma4-26b-q4-k-ctx262144",
   "model": "Gemma4 26B",
   "notes": "Hardware Corner RTX Pro 6000 Blackwell hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "26B",
   "power_w": null,
   "pp_tps": 2245.7,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX Pro 6000 Blackwell, Gemma4 26B (Q4_K), 256k Ctx: generation 112.5 t/s, prompt processing 2,245.7 t/s",
   "ram_gb": 96.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-pro-6000-blackwell/",
   "tps": 112.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "hardwarecorner-rtx-pro-6000-blackwell-qwen3-5-27b-q4-k-ctx4096",
   "model": "Qwen3.5-27B",
   "notes": "Hardware Corner RTX Pro 6000 Blackwell hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "27B",
   "power_w": null,
   "pp_tps": 3338.5,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX Pro 6000 Blackwell, Qwen3.5 27B (Q4_K), 4k Ctx: generation 60.9 t/s, prompt processing 3,338.5 t/s",
   "ram_gb": 96.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-pro-6000-blackwell/",
   "tps": 60.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "hardwarecorner-rtx-pro-6000-blackwell-qwen3-5-27b-q4-k-ctx16384",
   "model": "Qwen3.5-27B",
   "notes": "Hardware Corner RTX Pro 6000 Blackwell hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "27B",
   "power_w": null,
   "pp_tps": 2980.5,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX Pro 6000 Blackwell, Qwen3.5 27B (Q4_K), 16k Ctx: generation 57.5 t/s, prompt processing 2,980.5 t/s",
   "ram_gb": 96.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-pro-6000-blackwell/",
   "tps": 57.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "hardwarecorner-rtx-pro-6000-blackwell-qwen3-5-27b-q4-k-ctx32768",
   "model": "Qwen3.5-27B",
   "notes": "Hardware Corner RTX Pro 6000 Blackwell hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "27B",
   "power_w": null,
   "pp_tps": 2526.7,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX Pro 6000 Blackwell, Qwen3.5 27B (Q4_K), 32k Ctx: generation 55.1 t/s, prompt processing 2,526.7 t/s",
   "ram_gb": 96.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-pro-6000-blackwell/",
   "tps": 55.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 65536,
   "date": null,
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "hardwarecorner-rtx-pro-6000-blackwell-qwen3-5-27b-q4-k-ctx65536",
   "model": "Qwen3.5-27B",
   "notes": "Hardware Corner RTX Pro 6000 Blackwell hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "27B",
   "power_w": null,
   "pp_tps": 1856.7,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX Pro 6000 Blackwell, Qwen3.5 27B (Q4_K), 64k Ctx: generation 51.3 t/s, prompt processing 1,856.7 t/s",
   "ram_gb": 96.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-pro-6000-blackwell/",
   "tps": 51.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 131072,
   "date": null,
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "hardwarecorner-rtx-pro-6000-blackwell-qwen3-5-27b-q4-k-ctx131072",
   "model": "Qwen3.5-27B",
   "notes": "Hardware Corner RTX Pro 6000 Blackwell hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "27B",
   "power_w": null,
   "pp_tps": 1404.9,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX Pro 6000 Blackwell, Qwen3.5 27B (Q4_K), 128k Ctx: generation 45.1 t/s, prompt processing 1,404.9 t/s",
   "ram_gb": 96.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-pro-6000-blackwell/",
   "tps": 45.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 262144,
   "date": null,
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "hardwarecorner-rtx-pro-6000-blackwell-qwen3-5-27b-q4-k-ctx262144",
   "model": "Qwen3.5-27B",
   "notes": "Hardware Corner RTX Pro 6000 Blackwell hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "27B",
   "power_w": null,
   "pp_tps": 903.1,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX Pro 6000 Blackwell, Qwen3.5 27B (Q4_K), 256k Ctx: generation 36.3 t/s, prompt processing 903.1 t/s",
   "ram_gb": 96.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-pro-6000-blackwell/",
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "hardwarecorner-rtx-pro-6000-blackwell-gemma4-31b-q4-k-ctx4096",
   "model": "Gemma4 31B",
   "notes": "Hardware Corner RTX Pro 6000 Blackwell hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "31B",
   "power_w": null,
   "pp_tps": 3749.8,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX Pro 6000 Blackwell, Gemma4 31B (Q4_K), 4k Ctx: generation 61.6 t/s, prompt processing 3,749.8 t/s",
   "ram_gb": 96.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-pro-6000-blackwell/",
   "tps": 61.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "hardwarecorner-rtx-pro-6000-blackwell-gemma4-31b-q4-k-ctx16384",
   "model": "Gemma4 31B",
   "notes": "Hardware Corner RTX Pro 6000 Blackwell hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "31B",
   "power_w": null,
   "pp_tps": 3061.9,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX Pro 6000 Blackwell, Gemma4 31B (Q4_K), 16k Ctx: generation 59.6 t/s, prompt processing 3,061.9 t/s",
   "ram_gb": 96.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-pro-6000-blackwell/",
   "tps": 59.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "hardwarecorner-rtx-pro-6000-blackwell-gemma4-31b-q4-k-ctx32768",
   "model": "Gemma4 31B",
   "notes": "Hardware Corner RTX Pro 6000 Blackwell hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "31B",
   "power_w": null,
   "pp_tps": 2086.7,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX Pro 6000 Blackwell, Gemma4 31B (Q4_K), 32k Ctx: generation 55.8 t/s, prompt processing 2,086.7 t/s",
   "ram_gb": 96.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-pro-6000-blackwell/",
   "tps": 55.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 65536,
   "date": null,
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "hardwarecorner-rtx-pro-6000-blackwell-gemma4-31b-q4-k-ctx65536",
   "model": "Gemma4 31B",
   "notes": "Hardware Corner RTX Pro 6000 Blackwell hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "31B",
   "power_w": null,
   "pp_tps": 1423.0,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX Pro 6000 Blackwell, Gemma4 31B (Q4_K), 64k Ctx: generation 52.0 t/s, prompt processing 1,423.0 t/s",
   "ram_gb": 96.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-pro-6000-blackwell/",
   "tps": 52.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 131072,
   "date": null,
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "hardwarecorner-rtx-pro-6000-blackwell-gemma4-31b-q4-k-ctx131072",
   "model": "Gemma4 31B",
   "notes": "Hardware Corner RTX Pro 6000 Blackwell hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "31B",
   "power_w": null,
   "pp_tps": 876.8,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX Pro 6000 Blackwell, Gemma4 31B (Q4_K), 128k Ctx: generation 43.6 t/s, prompt processing 876.8 t/s",
   "ram_gb": 96.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-pro-6000-blackwell/",
   "tps": 43.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 262144,
   "date": null,
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "hardwarecorner-rtx-pro-6000-blackwell-gemma4-31b-q4-k-ctx262144",
   "model": "Gemma4 31B",
   "notes": "Hardware Corner RTX Pro 6000 Blackwell hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "31B",
   "power_w": null,
   "pp_tps": 506.8,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "RTX Pro 6000 Blackwell, Gemma4 31B (Q4_K), 256k Ctx: generation 34.4 t/s, prompt processing 506.8 t/s",
   "ram_gb": 96.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-pro-6000-blackwell/",
   "tps": 34.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "hardwarecorner-rtx-pro-6000-blackwell-gpt-oss-120b-mxfp4-ctx4096",
   "model": "gpt-oss 120B",
   "notes": "Hardware Corner RTX Pro 6000 Blackwell hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "120B",
   "power_w": null,
   "pp_tps": 4578.6,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX Pro 6000 Blackwell, gpt-oss 120B (MXFP4), 4k Ctx: generation 210.1 t/s, prompt processing 4,578.6 t/s",
   "ram_gb": 96.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-pro-6000-blackwell/",
   "tps": 210.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "hardwarecorner-rtx-pro-6000-blackwell-gpt-oss-120b-mxfp4-ctx16384",
   "model": "gpt-oss 120B",
   "notes": "Hardware Corner RTX Pro 6000 Blackwell hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "120B",
   "power_w": null,
   "pp_tps": 4060.7,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX Pro 6000 Blackwell, gpt-oss 120B (MXFP4), 16k Ctx: generation 182.2 t/s, prompt processing 4,060.7 t/s",
   "ram_gb": 96.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-pro-6000-blackwell/",
   "tps": 182.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "hardwarecorner-rtx-pro-6000-blackwell-gpt-oss-120b-mxfp4-ctx32768",
   "model": "gpt-oss 120B",
   "notes": "Hardware Corner RTX Pro 6000 Blackwell hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "120B",
   "power_w": null,
   "pp_tps": 3368.3,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX Pro 6000 Blackwell, gpt-oss 120B (MXFP4), 32k Ctx: generation 161.5 t/s, prompt processing 3,368.3 t/s",
   "ram_gb": 96.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-pro-6000-blackwell/",
   "tps": 161.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 65536,
   "date": null,
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "hardwarecorner-rtx-pro-6000-blackwell-gpt-oss-120b-mxfp4-ctx65536",
   "model": "gpt-oss 120B",
   "notes": "Hardware Corner RTX Pro 6000 Blackwell hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "120B",
   "power_w": null,
   "pp_tps": 2360.3,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX Pro 6000 Blackwell, gpt-oss 120B (MXFP4), 64k Ctx: generation 133.1 t/s, prompt processing 2,360.3 t/s",
   "ram_gb": 96.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-pro-6000-blackwell/",
   "tps": 133.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 131072,
   "date": null,
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "hardwarecorner-rtx-pro-6000-blackwell-gpt-oss-120b-mxfp4-ctx131072",
   "model": "gpt-oss 120B",
   "notes": "Hardware Corner RTX Pro 6000 Blackwell hub page; llama.cpp runtime per the site's benchmark methodology",
   "params": "120B",
   "power_w": null,
   "pp_tps": 1289.8,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "RTX Pro 6000 Blackwell, gpt-oss 120B (MXFP4), 128k Ctx: generation 99.8 t/s, prompt processing 1,289.8 t/s",
   "ram_gb": 96.0,
   "retrieved": "2026-10-03",
   "source_name": "Hardware Corner GPU LLM benchmarks (per-row attribution)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-pro-6000-blackwell/",
   "tps": 99.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M1 (7 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q8_0",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 108.21,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M1 (7 GPU), LLaMA 7B, Q8_0 PP   [t/s]: 108.21; Q8_0 TG   [t/s]: 7.92",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 7.92,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M1 (7 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 107.81,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M1 (7 GPU), LLaMA 7B, Q4_0 PP   [t/s]: 107.81; Q4_0 TG   [t/s]: 14.19",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 14.19,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M1 (8 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q8_0-2",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 117.25,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M1 (8 GPU), LLaMA 7B, Q8_0 PP   [t/s]: 117.25; Q8_0 TG   [t/s]: 7.91",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 7.91,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M1 (8 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 117.96,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M1 (8 GPU), LLaMA 7B, Q4_0 PP   [t/s]: 117.96; Q4_0 TG   [t/s]: 14.15",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 14.15,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M1 Pro (14 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-F16",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 262.65,
   "provenance": "community",
   "quant": "F16",
   "quote": "M1 Pro (14 GPU), LLaMA 7B, F16 PP   [t/s]: 262.65; F16 TG   [t/s]: 12.75",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 12.75,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M1 Pro (14 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q8_0-3",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 235.16,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M1 Pro (14 GPU), LLaMA 7B, Q8_0 PP   [t/s]: 235.16; Q8_0 TG   [t/s]: 21.95",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 21.95,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M1 Pro (14 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 232.55,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M1 Pro (14 GPU), LLaMA 7B, Q4_0 PP   [t/s]: 232.55; Q4_0 TG   [t/s]: 35.52",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 35.52,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M1 Pro (16 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-F16-2",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 302.14,
   "provenance": "community",
   "quant": "F16",
   "quote": "M1 Pro (16 GPU), LLaMA 7B, F16 PP   [t/s]: 302.14; F16 TG   [t/s]: 12.75",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 12.75,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M1 Pro (16 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q8_0-4",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 270.37,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M1 Pro (16 GPU), LLaMA 7B, Q8_0 PP   [t/s]: 270.37; Q8_0 TG   [t/s]: 22.34",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 22.34,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M1 Pro (16 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 266.25,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M1 Pro (16 GPU), LLaMA 7B, Q4_0 PP   [t/s]: 266.25; Q4_0 TG   [t/s]: 36.41",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 36.41,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M1 Max (24 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-F16-3",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 453.03,
   "provenance": "community",
   "quant": "F16",
   "quote": "M1 Max (24 GPU), LLaMA 7B, F16 PP   [t/s]: 453.03; F16 TG   [t/s]: 22.55",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 22.55,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M1 Max (24 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q8_0-5",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 405.87,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M1 Max (24 GPU), LLaMA 7B, Q8_0 PP   [t/s]: 405.87; Q8_0 TG   [t/s]: 37.81",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 37.81,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M1 Max (24 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q4_0-5",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 400.26,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M1 Max (24 GPU), LLaMA 7B, Q4_0 PP   [t/s]: 400.26; Q4_0 TG   [t/s]: 54.61",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 54.61,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M1 Max (32 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-F16-4",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 599.53,
   "provenance": "community",
   "quant": "F16",
   "quote": "M1 Max (32 GPU), LLaMA 7B, F16 PP   [t/s]: 599.53; F16 TG   [t/s]: 23.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 23.03,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M1 Max (32 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q8_0-6",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 537.37,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M1 Max (32 GPU), LLaMA 7B, Q8_0 PP   [t/s]: 537.37; Q8_0 TG   [t/s]: 40.2",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 40.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M1 Max (32 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q4_0-6",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 530.06,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M1 Max (32 GPU), LLaMA 7B, Q4_0 PP   [t/s]: 530.06; Q4_0 TG   [t/s]: 61.19",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 61.19,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M1 Ultra (48 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-F16-5",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 875.81,
   "provenance": "community",
   "quant": "F16",
   "quote": "M1 Ultra (48 GPU), LLaMA 7B, F16 PP   [t/s]: 875.81; F16 TG   [t/s]: 33.92",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 33.92,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M1 Ultra (48 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q8_0-7",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 783.45,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M1 Ultra (48 GPU), LLaMA 7B, Q8_0 PP   [t/s]: 783.45; Q8_0 TG   [t/s]: 55.69",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 55.69,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M1 Ultra (48 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q4_0-7",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 772.24,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M1 Ultra (48 GPU), LLaMA 7B, Q4_0 PP   [t/s]: 772.24; Q4_0 TG   [t/s]: 74.93",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 74.93,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M1 Ultra (64 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-F16-6",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1168.89,
   "provenance": "community",
   "quant": "F16",
   "quote": "M1 Ultra (64 GPU), LLaMA 7B, F16 PP   [t/s]: 1168.89; F16 TG   [t/s]: 37.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 37.01,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M1 Ultra (64 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q8_0-8",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1042.95,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M1 Ultra (64 GPU), LLaMA 7B, Q8_0 PP   [t/s]: 1042.95; Q8_0 TG   [t/s]: 59.87",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 59.87,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M1 Ultra (64 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q4_0-8",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1030.04,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M1 Ultra (64 GPU), LLaMA 7B, Q4_0 PP   [t/s]: 1030.04; Q4_0 TG   [t/s]: 83.73",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 83.73,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M2 (8 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q8_0-9",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 147.27,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M2 (8 GPU), LLaMA 7B, Q8_0 PP   [t/s]: 147.27; Q8_0 TG   [t/s]: 12.18",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 12.18,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M2 (8 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q4_0-9",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 145.91,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M2 (8 GPU), LLaMA 7B, Q4_0 PP   [t/s]: 145.91; Q4_0 TG   [t/s]: 21.7",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 21.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M2 (10 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-F16-7",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 201.34,
   "provenance": "community",
   "quant": "F16",
   "quote": "M2 (10 GPU), LLaMA 7B, F16 PP   [t/s]: 201.34; F16 TG   [t/s]: 6.72",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 6.72,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M2 (10 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q8_0-10",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 181.4,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M2 (10 GPU), LLaMA 7B, Q8_0 PP   [t/s]: 181.4; Q8_0 TG   [t/s]: 12.21",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 12.21,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M2 (10 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q4_0-10",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 179.57,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M2 (10 GPU), LLaMA 7B, Q4_0 PP   [t/s]: 179.57; Q4_0 TG   [t/s]: 21.91",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 21.91,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M2 Pro (16 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-F16-8",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 312.65,
   "provenance": "community",
   "quant": "F16",
   "quote": "M2 Pro (16 GPU), LLaMA 7B, F16 PP   [t/s]: 312.65; F16 TG   [t/s]: 12.47",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 12.47,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M2 Pro (16 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q8_0-11",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 288.46,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M2 Pro (16 GPU), LLaMA 7B, Q8_0 PP   [t/s]: 288.46; Q8_0 TG   [t/s]: 22.7",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M2 Pro (16 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q4_0-11",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 294.24,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M2 Pro (16 GPU), LLaMA 7B, Q4_0 PP   [t/s]: 294.24; Q4_0 TG   [t/s]: 37.87",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 37.87,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M2 Pro (19 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-F16-9",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 384.38,
   "provenance": "community",
   "quant": "F16",
   "quote": "M2 Pro (19 GPU), LLaMA 7B, F16 PP   [t/s]: 384.38; F16 TG   [t/s]: 13.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 13.06,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M2 Pro (19 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q8_0-12",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 344.5,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M2 Pro (19 GPU), LLaMA 7B, Q8_0 PP   [t/s]: 344.5; Q8_0 TG   [t/s]: 23.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 23.01,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M2 Pro (19 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q4_0-12",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 341.19,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M2 Pro (19 GPU), LLaMA 7B, Q4_0 PP   [t/s]: 341.19; Q4_0 TG   [t/s]: 38.86",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 38.86,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M2 Max (30 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-F16-10",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 600.46,
   "provenance": "community",
   "quant": "F16",
   "quote": "M2 Max (30 GPU), LLaMA 7B, F16 PP   [t/s]: 600.46; F16 TG   [t/s]: 24.16",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 24.16,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M2 Max (30 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q8_0-13",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 540.15,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M2 Max (30 GPU), LLaMA 7B, Q8_0 PP   [t/s]: 540.15; Q8_0 TG   [t/s]: 39.97",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 39.97,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M2 Max (30 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q4_0-13",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 537.6,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M2 Max (30 GPU), LLaMA 7B, Q4_0 PP   [t/s]: 537.6; Q4_0 TG   [t/s]: 60.99",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 60.99,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M2 Max (38 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-F16-11",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 755.67,
   "provenance": "community",
   "quant": "F16",
   "quote": "M2 Max (38 GPU), LLaMA 7B, F16 PP   [t/s]: 755.67; F16 TG   [t/s]: 24.65",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 24.65,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M2 Max (38 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q8_0-14",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 677.91,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M2 Max (38 GPU), LLaMA 7B, Q8_0 PP   [t/s]: 677.91; Q8_0 TG   [t/s]: 41.83",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 41.83,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M2 Max (38 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q4_0-14",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 671.31,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M2 Max (38 GPU), LLaMA 7B, Q4_0 PP   [t/s]: 671.31; Q4_0 TG   [t/s]: 65.95",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 65.95,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M2 Ultra (60 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-F16-12",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1128.59,
   "provenance": "community",
   "quant": "F16",
   "quote": "M2 Ultra (60 GPU), LLaMA 7B, F16 PP   [t/s]: 1128.59; F16 TG   [t/s]: 39.86",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 39.86,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M2 Ultra (60 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q8_0-15",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1003.16,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M2 Ultra (60 GPU), LLaMA 7B, Q8_0 PP   [t/s]: 1003.16; Q8_0 TG   [t/s]: 62.14",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 62.14,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M2 Ultra (60 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q4_0-15",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1013.81,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M2 Ultra (60 GPU), LLaMA 7B, Q4_0 PP   [t/s]: 1013.81; Q4_0 TG   [t/s]: 88.64",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 88.64,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M2 Ultra (76 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-F16-13",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1401.85,
   "provenance": "community",
   "quant": "F16",
   "quote": "M2 Ultra (76 GPU), LLaMA 7B, F16 PP   [t/s]: 1401.85; F16 TG   [t/s]: 41.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 41.02,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M2 Ultra (76 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q8_0-16",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1248.59,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M2 Ultra (76 GPU), LLaMA 7B, Q8_0 PP   [t/s]: 1248.59; Q8_0 TG   [t/s]: 66.64",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 66.64,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M2 Ultra (76 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q4_0-16",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1238.48,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M2 Ultra (76 GPU), LLaMA 7B, Q4_0 PP   [t/s]: 1238.48; Q4_0 TG   [t/s]: 94.27",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 94.27,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M3 (10 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q8_0-17",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 187.52,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M3 (10 GPU), LLaMA 7B, Q8_0 PP   [t/s]: 187.52; Q8_0 TG   [t/s]: 12.27",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 12.27,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M3 (10 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q4_0-17",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 186.75,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M3 (10 GPU), LLaMA 7B, Q4_0 PP   [t/s]: 186.75; Q4_0 TG   [t/s]: 21.34",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 21.34,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M3 Pro (14 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q8_0-18",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 272.11,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M3 Pro (14 GPU), LLaMA 7B, Q8_0 PP   [t/s]: 272.11; Q8_0 TG   [t/s]: 17.44",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 17.44,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M3 Pro (14 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q4_0-18",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 269.49,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M3 Pro (14 GPU), LLaMA 7B, Q4_0 PP   [t/s]: 269.49; Q4_0 TG   [t/s]: 30.65",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 30.65,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M3 Pro (18 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-F16-14",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 357.45,
   "provenance": "community",
   "quant": "F16",
   "quote": "M3 Pro (18 GPU), LLaMA 7B, F16 PP   [t/s]: 357.45; F16 TG   [t/s]: 9.89",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 9.89,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M3 Pro (18 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q8_0-19",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 344.66,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M3 Pro (18 GPU), LLaMA 7B, Q8_0 PP   [t/s]: 344.66; Q8_0 TG   [t/s]: 17.53",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 17.53,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M3 Pro (18 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q4_0-19",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 341.67,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M3 Pro (18 GPU), LLaMA 7B, Q4_0 PP   [t/s]: 341.67; Q4_0 TG   [t/s]: 30.74",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 30.74,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M3 Max (30 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-F16-15",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 589.41,
   "provenance": "community",
   "quant": "F16",
   "quote": "M3 Max (30 GPU), LLaMA 7B, F16 PP   [t/s]: 589.41; F16 TG   [t/s]: 19.54",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 19.54,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M3 Max (30 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q8_0-20",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 566.4,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M3 Max (30 GPU), LLaMA 7B, Q8_0 PP   [t/s]: 566.4; Q8_0 TG   [t/s]: 34.3",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 34.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M3 Max (30 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q4_0-20",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 567.59,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M3 Max (30 GPU), LLaMA 7B, Q4_0 PP   [t/s]: 567.59; Q4_0 TG   [t/s]: 56.58",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 56.58,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M3 Max (40 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-F16-16",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 779.17,
   "provenance": "community",
   "quant": "F16",
   "quote": "M3 Max (40 GPU), LLaMA 7B, F16 PP   [t/s]: 779.17; F16 TG   [t/s]: 25.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 25.09,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M3 Max (40 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q8_0-21",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 757.64,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M3 Max (40 GPU), LLaMA 7B, Q8_0 PP   [t/s]: 757.64; Q8_0 TG   [t/s]: 42.75",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 42.75,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M3 Max (40 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q4_0-21",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 759.7,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M3 Max (40 GPU), LLaMA 7B, Q4_0 PP   [t/s]: 759.7; Q4_0 TG   [t/s]: 66.31",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 66.31,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M3 Ultra (60 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-F16-17",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1121.8,
   "provenance": "community",
   "quant": "F16",
   "quote": "M3 Ultra (60 GPU), LLaMA 7B, F16 PP   [t/s]: 1121.80; F16 TG   [t/s]: 42.24",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 42.24,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M3 Ultra (60 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q8_0-22",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1085.76,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M3 Ultra (60 GPU), LLaMA 7B, Q8_0 PP   [t/s]: 1085.76; Q8_0 TG   [t/s]: 63.55",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 63.55,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M3 Ultra (60 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q4_0-22",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1073.09,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M3 Ultra (60 GPU), LLaMA 7B, Q4_0 PP   [t/s]: 1073.09; Q4_0 TG   [t/s]: 88.40",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 88.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M3 Ultra (80 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-F16-18",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1538.34,
   "provenance": "community",
   "quant": "F16",
   "quote": "M3 Ultra (80 GPU), LLaMA 7B, F16 PP   [t/s]: 1538.34; F16 TG   [t/s]: 39.78",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 39.78,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M3 Ultra (80 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q8_0-23",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1487.51,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M3 Ultra (80 GPU), LLaMA 7B, Q8_0 PP   [t/s]: 1487.51; Q8_0 TG   [t/s]: 63.93",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 63.93,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M3 Ultra (80 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q4_0-23",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1471.24,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M3 Ultra (80 GPU), LLaMA 7B, Q4_0 PP   [t/s]: 1471.24; Q4_0 TG   [t/s]: 92.14",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 92.14,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M4 (10 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-F16-19",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 230.18,
   "provenance": "community",
   "quant": "F16",
   "quote": "M4 (10 GPU), LLaMA 7B, F16 PP   [t/s]: 230.18; F16 TG   [t/s]: 7.43",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 7.43,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M4 (10 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q8_0-24",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 223.64,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M4 (10 GPU), LLaMA 7B, Q8_0 PP   [t/s]: 223.64; Q8_0 TG   [t/s]: 13.54",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 13.54,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M4 (10 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q4_0-24",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 221.29,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M4 (10 GPU), LLaMA 7B, Q4_0 PP   [t/s]: 221.29; Q4_0 TG   [t/s]: 24.11",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 24.11,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M4 Pro (16 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-F16-20",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 381.14,
   "provenance": "community",
   "quant": "F16",
   "quote": "M4 Pro (16 GPU), LLaMA 7B, F16 PP   [t/s]: 381.14; F16 TG   [t/s]: 17.19",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 17.19,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M4 Pro (16 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q8_0-25",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 367.13,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M4 Pro (16 GPU), LLaMA 7B, Q8_0 PP   [t/s]: 367.13; Q8_0 TG   [t/s]: 30.54",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 30.54,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M4 Pro (16 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q4_0-25",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 364.06,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M4 Pro (16 GPU), LLaMA 7B, Q4_0 PP   [t/s]: 364.06; Q4_0 TG   [t/s]: 49.64",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 49.64,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M4 Pro (20 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-F16-21",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 464.48,
   "provenance": "community",
   "quant": "F16",
   "quote": "M4 Pro (20 GPU), LLaMA 7B, F16 PP   [t/s]: 464.48; F16 TG   [t/s]: 17.18",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 17.18,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M4 Pro (20 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q8_0-26",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 449.62,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M4 Pro (20 GPU), LLaMA 7B, Q8_0 PP   [t/s]: 449.62; Q8_0 TG   [t/s]: 30.69",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 30.69,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M4 Pro (20 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q4_0-26",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 439.78,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M4 Pro (20 GPU), LLaMA 7B, Q4_0 PP   [t/s]: 439.78; Q4_0 TG   [t/s]: 50.74",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 50.74,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M4 Max (32 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-F16-22",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 736.25,
   "provenance": "community",
   "quant": "F16",
   "quote": "M4 Max (32 GPU), LLaMA 7B, F16 PP   [t/s]: 736.25; F16 TG   [t/s]: 24.29",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 24.29,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M4 Max (32 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q8_0-27",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 718.56,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M4 Max (32 GPU), LLaMA 7B, Q8_0 PP   [t/s]: 718.56; Q8_0 TG   [t/s]: 43.87",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 43.87,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M4 Max (32 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q4_0-27",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 713.93,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M4 Max (32 GPU), LLaMA 7B, Q4_0 PP   [t/s]: 713.93; Q4_0 TG   [t/s]: 69.95",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 69.95,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M4 Max (40 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-F16-23",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 922.83,
   "provenance": "community",
   "quant": "F16",
   "quote": "M4 Max (40 GPU), LLaMA 7B, F16 PP   [t/s]: 922.83; F16 TG   [t/s]: 31.64",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 31.64,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M4 Max (40 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q8_0-28",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 891.94,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M4 Max (40 GPU), LLaMA 7B, Q8_0 PP   [t/s]: 891.94; Q8_0 TG   [t/s]: 54.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 54.05,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M4 Max (40 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q4_0-28",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 885.68,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M4 Max (40 GPU), LLaMA 7B, Q4_0 PP   [t/s]: 885.68; Q4_0 TG   [t/s]: 83.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 83.06,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M5 (10 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q8_0-29",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 715.42,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M5 (10 GPU), LLaMA 7B, Q8_0 PP   [t/s]: 715.42; Q8_0 TG   [t/s]: 18.42",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 18.42,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M5 (10 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q4_0-29",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 722.79,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M5 (10 GPU), LLaMA 7B, Q4_0 PP   [t/s]: 722.79; Q4_0 TG   [t/s]: 31.88",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 31.88,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M5 Pro (20 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-F16-24",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1588.78,
   "provenance": "community",
   "quant": "F16",
   "quote": "M5 Pro (20 GPU), LLaMA 7B, F16 PP   [t/s]: 1588.78; F16 TG   [t/s]: 21.55",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 21.55,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M5 Pro (20 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q8_0-30",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1553.86,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M5 Pro (20 GPU), LLaMA 7B, Q8_0 PP   [t/s]: 1553.86; Q8_0 TG   [t/s]: 38.92",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 38.92,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M5 Pro (20 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q4_0-30",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1620.64,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M5 Pro (20 GPU), LLaMA 7B, Q4_0 PP   [t/s]: 1620.64; Q4_0 TG   [t/s]: 66.33",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 66.33,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M5 Max (40 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-F16-25",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3158.49,
   "provenance": "community",
   "quant": "F16",
   "quote": "M5 Max (40 GPU), LLaMA 7B, F16 PP   [t/s]: 3158.49; F16 TG   [t/s]: 37.11",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 37.11,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M5 Max (40 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q8_0-31",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3143.81,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M5 Max (40 GPU), LLaMA 7B, Q8_0 PP   [t/s]: 3143.81; Q8_0 TG   [t/s]: 72.42",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 72.42,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M5 Max (40 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7b-Q4_0-31",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3219.99,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M5 Max (40 GPU), LLaMA 7B, Q4_0 PP   [t/s]: 3219.99; Q4_0 TG   [t/s]: 119.92",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 119.92,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M2 Ultra (76 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-aspublished-F16",
   "model": "Llama 2 7B",
   "notes": "fa=0 config=8e672ef",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1401.85,
   "provenance": "community",
   "quant": "F16",
   "quote": "Llama 2 7B; M2 Ultra [^2] | 800 | 76 | 1401.85 | 41.02 | 1248.59 | 66.64 | 1238.48 | 94.27; M2 Ultra (76 GPU)",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 41.02,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M2 Ultra (76 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-aspublished-Q8_0",
   "model": "Llama 2 7B",
   "notes": "fa=0 config=8e672ef",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1248.59,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Llama 2 7B; M2 Ultra [^2] | 800 | 76 | 1401.85 | 41.02 | 1248.59 | 66.64 | 1238.48 | 94.27; M2 Ultra (76 GPU)",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 66.64,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M2 Ultra (76 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-aspublished-Q4_0",
   "model": "Llama 2 7B",
   "notes": "fa=0 config=8e672ef",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1238.48,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Llama 2 7B; M2 Ultra [^2] | 800 | 76 | 1401.85 | 41.02 | 1248.59 | 66.64 | 1238.48 | 94.27; M2 Ultra (76 GPU)",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 94.27,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M2 Ultra (76 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-aspublished-F16-2",
   "model": "Llama 2 7B",
   "notes": "fa=1 config=86ed72d2",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1525.95,
   "provenance": "community",
   "quant": "F16",
   "quote": "Llama 2 7B; M2 Ultra 86ed72d2 + FA | 800 | 76 | 1525.95 | 43.15 | 1368.18 | 73.11 | 1391.78 | 108.80; M2 Ultra (76 GPU)",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 43.15,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M2 Ultra (76 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-aspublished-Q8_0-2",
   "model": "Llama 2 7B",
   "notes": "fa=1 config=86ed72d2",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1368.18,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Llama 2 7B; M2 Ultra 86ed72d2 + FA | 800 | 76 | 1525.95 | 43.15 | 1368.18 | 73.11 | 1391.78 | 108.80; M2 Ultra (76 GPU)",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 73.11,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M2 Ultra (76 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-aspublished-Q4_0-2",
   "model": "Llama 2 7B",
   "notes": "fa=1 config=86ed72d2",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1391.78,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Llama 2 7B; M2 Ultra 86ed72d2 + FA | 800 | 76 | 1525.95 | 43.15 | 1368.18 | 73.11 | 1391.78 | 108.80; M2 Ultra (76 GPU)",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 108.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M2 Ultra (76 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-aspublished-F16-3",
   "model": "Llama 2 7B",
   "notes": "fa=1 config=5c0eb5ef5",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1561.35,
   "provenance": "community",
   "quant": "F16",
   "quote": "Llama 2 7B; M2 Ultra 5c0eb5ef5 + FA | 800 | 76 | 1561.35 | 43.24 | 1386.97 | 73.35 | 1412.42 | 109.41; M2 Ultra (76 GPU)",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 43.24,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M2 Ultra (76 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-aspublished-Q8_0-3",
   "model": "Llama 2 7B",
   "notes": "fa=1 config=5c0eb5ef5",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1386.97,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Llama 2 7B; M2 Ultra 5c0eb5ef5 + FA | 800 | 76 | 1561.35 | 43.24 | 1386.97 | 73.35 | 1412.42 | 109.41; M2 Ultra (76 GPU)",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 73.35,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M2 Ultra (76 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-aspublished-Q4_0-3",
   "model": "Llama 2 7B",
   "notes": "fa=1 config=5c0eb5ef5",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1412.42,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Llama 2 7B; M2 Ultra 5c0eb5ef5 + FA | 800 | 76 | 1561.35 | 43.24 | 1386.97 | 73.35 | 1412.42 | 109.41; M2 Ultra (76 GPU)",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 109.41,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M2 Ultra (76 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-aspublished-F16-4",
   "model": "Llama 2 7B",
   "notes": "fa=1 config=c1d0e7a00",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1627.82,
   "provenance": "community",
   "quant": "F16",
   "quote": "Llama 2 7B; M2 Ultra c1d0e7a00 + FA | 800 | 76 | 1627.82 | 49.55 | 1486.51 | 82.83 | 1488.81 | 125.21; M2 Ultra (76 GPU)",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 49.55,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M2 Ultra (76 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-aspublished-Q8_0-4",
   "model": "Llama 2 7B",
   "notes": "fa=1 config=c1d0e7a00",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1486.51,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Llama 2 7B; M2 Ultra c1d0e7a00 + FA | 800 | 76 | 1627.82 | 49.55 | 1486.51 | 82.83 | 1488.81 | 125.21; M2 Ultra (76 GPU)",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 82.83,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M2 Ultra (76 GPU)",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-aspublished-Q4_0-4",
   "model": "Llama 2 7B",
   "notes": "fa=1 config=c1d0e7a00",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1488.81,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Llama 2 7B; M2 Ultra c1d0e7a00 + FA | 800 | 76 | 1627.82 | 49.55 | 1486.51 | 82.83 | 1488.81 | 125.21; M2 Ultra (76 GPU)",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": null,
   "tps": 125.21,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7bmostly-F16",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 302.14,
   "provenance": "community",
   "quant": "F16",
   "quote": "M1 Pro; llama 7B mostly F16 | pp512 | 302.14 \u00b1 0.07; llama 7B mostly F16 | tg128 | 12.75 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": 128,
   "tps": 12.75,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7bmostly-Q8_0",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 270.37,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M1 Pro; llama 7B mostly Q8_0 | pp512 | 270.37 \u00b1 0.02; llama 7B mostly Q8_0 | tg128 | 22.34 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": 128,
   "tps": 22.34,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7bmostly-Q4_0",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 266.25,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M1 Pro; llama 7B mostly Q4_0 | pp512 | 266.25 \u00b1 0.07; llama 7B mostly Q4_0 | tg128 | 36.41 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": 128,
   "tps": 36.41,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7bmostly-F16-2",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1401.85,
   "provenance": "community",
   "quant": "F16",
   "quote": "M2 Ultra; llama 7B mostly F16 | pp512 | 1401.85 \u00b1 1.75; llama 7B mostly F16 | tg128 | 41.02 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": 128,
   "tps": 41.02,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7bmostly-Q8_0-2",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1248.59,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M2 Ultra; llama 7B mostly Q8_0 | pp512 | 1248.59 \u00b1 0.73; llama 7B mostly Q8_0 | tg128 | 66.64 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": 128,
   "tps": 66.64,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7bmostly-Q4_0-2",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1238.48,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M2 Ultra; llama 7B mostly Q4_0 | pp512 | 1238.48 \u00b1 0.76; llama 7B mostly Q4_0 | tg128 | 94.27 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": 128,
   "tps": 94.27,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M3 Max",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7bmostly-F16-3",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 794.26,
   "provenance": "community",
   "quant": "F16",
   "quote": "M3 Max; llama 7B mostly F16 | pp512 | 794.26 \u00b1 3.16; llama 7B mostly F16 | tg128 | 25.27 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": 128,
   "tps": 25.27,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M3 Max",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7bmostly-Q8_0-3",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 749.37,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M3 Max; llama 7B mostly Q8_0 | pp512 | 749.37 \u00b1 8.35; llama 7B mostly Q8_0 | tg128 | 43.00 \u00b1 0.12",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": 128,
   "tps": 43.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M3 Max",
   "id": "lc-disc-4167-chttps://github.com/ggml-org/llama.cpp/discussions/4167-llama7bmostly-Q4_0-3",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 690.99,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M3 Max; llama 7B mostly Q4_0 | pp512 | 690.99 \u00b1 33.76; llama 7B mostly Q4_0 | tg128 | 65.85 \u00b1 0.22",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tg_tokens": 128,
   "tps": 65.85,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M2",
   "id": "lc-disc-4167-c7644774-llama7bmostly-F16",
   "model": "llama 7B mostly",
   "notes": "threads=99; ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 201.34,
   "provenance": "community",
   "quant": "F16",
   "quote": "M2; llama 7B mostly F16 | pp512 | 201.34 \u00b1 0.21; llama 7B mostly F16 | tg128 | 6.72 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7644774",
   "tg_tokens": 128,
   "tps": 6.72,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M2",
   "id": "lc-disc-4167-c7644774-llama7bmostly-Q8_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 181.4,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M2; llama 7B mostly Q8_0 | pp512 | 181.40 \u00b1 0.05; llama 7B mostly Q8_0 | tg128 | 12.21 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7644774",
   "tg_tokens": 128,
   "tps": 12.21,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M2",
   "id": "lc-disc-4167-c7644774-llama7bmostly-Q4_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 179.57,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M2; llama 7B mostly Q4_0 | pp512 | 179.57 \u00b1 0.04; llama 7B mostly Q4_0 | tg128 | 21.91 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7644774",
   "tg_tokens": 128,
   "tps": 21.91,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M2 Max",
   "id": "lc-disc-4167-c7644774-llama7bmostly-F16-2",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 755.67,
   "provenance": "community",
   "quant": "F16",
   "quote": "M2 Max; llama 7B mostly F16 | pp512 | 755.67 \u00b1 0.11; llama 7B mostly F16 | tg128 | 24.65 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7644774",
   "tg_tokens": 128,
   "tps": 24.65,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M2 Max",
   "id": "lc-disc-4167-c7644774-llama7bmostly-Q8_0-2",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 677.91,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M2 Max; llama 7B mostly Q8_0 | pp512 | 677.91 \u00b1 0.26; llama 7B mostly Q8_0 | tg128 | 41.83 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7644774",
   "tg_tokens": 128,
   "tps": 41.83,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-22",
   "flags": [],
   "hardware": "M2 Max",
   "id": "lc-disc-4167-c7644774-llama7bmostly-Q4_0-2",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 671.31,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M2 Max; llama 7B mostly Q4_0 | pp512 | 671.31 \u00b1 0.20; llama 7B mostly Q4_0 | tg128 | 65.95 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7644774",
   "tg_tokens": 128,
   "tps": 65.95,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-23",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "lc-disc-4167-c7654488-llama7bmostly-F16",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1128.59,
   "provenance": "community",
   "quant": "F16",
   "quote": "M2 Ultra; llama 7B mostly F16 | pp512 | 1128.59 \u00b1 0.82; llama 7B mostly F16 | tg128 | 39.86 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7654488",
   "tg_tokens": 128,
   "tps": 39.86,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-23",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "lc-disc-4167-c7654488-llama7bmostly-Q8_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1003.16,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M2 Ultra; llama 7B mostly Q8_0 | pp512 | 1003.16 \u00b1 0.39; llama 7B mostly Q8_0 | tg128 | 62.14 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7654488",
   "tg_tokens": 128,
   "tps": 62.14,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-23",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "lc-disc-4167-c7654488-llama7bmostly-Q4_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1013.81,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M2 Ultra; llama 7B mostly Q4_0 | pp512 | 1013.81 \u00b1 0.92; llama 7B mostly Q4_0 | tg128 | 88.64 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7654488",
   "tg_tokens": 128,
   "tps": 88.64,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-24",
   "flags": [],
   "hardware": "M3 Max",
   "id": "lc-disc-4167-c7656623-llama7bmostly-F16",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01555",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 779.17,
   "provenance": "community",
   "quant": "F16",
   "quote": "M3 Max; llama 7B mostly F16 | pp512 | 779.17 \u00b1 0.49; llama 7B mostly F16 | tg128 | 25.09 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7656623",
   "tg_tokens": 128,
   "tps": 25.09,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-24",
   "flags": [],
   "hardware": "M3 Max",
   "id": "lc-disc-4167-c7656623-llama7bmostly-Q8_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01555",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 757.64,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M3 Max; llama 7B mostly Q8_0 | pp512 | 757.64 \u00b1 1.03; llama 7B mostly Q8_0 | tg128 | 42.75 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7656623",
   "tg_tokens": 128,
   "tps": 42.75,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-24",
   "flags": [],
   "hardware": "M3 Max",
   "id": "lc-disc-4167-c7656623-llama7bmostly-Q4_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01555",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 759.7,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M3 Max; llama 7B mostly Q4_0 | pp512 | 759.70 \u00b1 2.26; llama 7B mostly Q4_0 | tg128 | 66.31 \u00b1 0.12",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7656623",
   "tg_tokens": 128,
   "tps": 66.31,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-24",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "lc-disc-4167-c7663473-llama7bmostly-F16",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01560",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 312.65,
   "provenance": "community",
   "quant": "F16",
   "quote": "M2 Pro; llama 7B mostly F16 | pp512 | 312.65 \u00b1 15.75; llama 7B mostly F16 | tg128 | 12.47 \u00b1 0.71",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7663473",
   "tg_tokens": 128,
   "tps": 12.47,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-24",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "lc-disc-4167-c7663473-llama7bmostly-Q8_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01560",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 288.46,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M2 Pro; llama 7B mostly Q8_0 | pp512 | 288.46 \u00b1 0.06; llama 7B mostly Q8_0 | tg128 | 22.70 \u00b1 0.12",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7663473",
   "tg_tokens": 128,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-24",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "lc-disc-4167-c7663473-llama7bmostly-Q4_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01560",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 294.24,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M2 Pro; llama 7B mostly Q4_0 | pp512 | 294.24 \u00b1 0.10; llama 7B mostly Q4_0 | tg128 | 37.87 \u00b1 0.10",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7663473",
   "tg_tokens": 128,
   "tps": 37.87,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-25",
   "flags": [],
   "hardware": "M2",
   "id": "lc-disc-4167-c7664186-llama7bmostly-Q8_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01560",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 674.5,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M2; llama 7B mostly Q8_0 | pp512 | 674.50 \u00b1 0.58; llama 7B mostly Q8_0 | tg128 | 41.79 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7664186",
   "tg_tokens": 128,
   "tps": 41.79,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-25",
   "flags": [],
   "hardware": "M2",
   "id": "lc-disc-4167-c7664186-llama7bmostly-Q4_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01560",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 669.51,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M2; llama 7B mostly Q4_0 | pp512 | 669.51 \u00b1 1.17; llama 7B mostly Q4_0 | tg128 | 64.55 \u00b1 1.36",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7664186",
   "tg_tokens": 128,
   "tps": 64.55,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-26",
   "flags": [],
   "hardware": "M2",
   "id": "lc-disc-4167-c7670696-llama7bmostly-F16",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01566",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 747.99,
   "provenance": "community",
   "quant": "F16",
   "quote": "M2; llama 7B mostly F16 | pp512 | 747.99 \u00b1 0.28; llama 7B mostly F16 | tg128 | 24.54 \u00b1 0.22",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7670696",
   "tg_tokens": 128,
   "tps": 24.54,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-26",
   "flags": [],
   "hardware": "M2",
   "id": "lc-disc-4167-c7670696-llama7bmostly-Q8_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01566",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 674.37,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M2; llama 7B mostly Q8_0 | pp512 | 674.37 \u00b1 0.63; llama 7B mostly Q8_0 | tg128 | 40.67 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7670696",
   "tg_tokens": 128,
   "tps": 40.67,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-26",
   "flags": [],
   "hardware": "M2",
   "id": "lc-disc-4167-c7670696-llama7bmostly-Q4_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01566",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 668.28,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M2; llama 7B mostly Q4_0 | pp512 | 668.28 \u00b1 0.24; llama 7B mostly Q4_0 | tg128 | 62.98 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7670696",
   "tg_tokens": 128,
   "tps": 62.98,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-25",
   "flags": [],
   "hardware": "M1 Max",
   "id": "lc-disc-4167-c7664196-llama7bmostly-F16",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01560",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 599.53,
   "provenance": "community",
   "quant": "F16",
   "quote": "M1 Max; llama 7B mostly F16 | pp512 | 599.53 \u00b1 0.86; llama 7B mostly F16 | tg128 | 23.03 \u00b1 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7664196",
   "tg_tokens": 128,
   "tps": 23.03,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-25",
   "flags": [],
   "hardware": "M1 Max",
   "id": "lc-disc-4167-c7664196-llama7bmostly-Q8_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01560",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 537.37,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M1 Max; llama 7B mostly Q8_0 | pp512 | 537.37 \u00b1 0.19; llama 7B mostly Q8_0 | tg128 | 40.20 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7664196",
   "tg_tokens": 128,
   "tps": 40.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-25",
   "flags": [],
   "hardware": "M1 Max",
   "id": "lc-disc-4167-c7664196-llama7bmostly-Q4_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01560",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 530.06,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M1 Max; llama 7B mostly Q4_0 | pp512 | 530.06 \u00b1 0.17; llama 7B mostly Q4_0 | tg128 | 61.19 \u00b1 0.15",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7664196",
   "tg_tokens": 128,
   "tps": 61.19,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-25",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "lc-disc-4167-c7664636-llama7bmostly-Q8_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01560",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 272.11,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M3 Pro; llama 7B mostly Q8_0 | pp512 | 272.11 \u00b1 1.40; llama 7B mostly Q8_0 | tg128 | 17.44 \u00b1 0.42",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7664636",
   "tg_tokens": 128,
   "tps": 17.44,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-25",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "lc-disc-4167-c7664636-llama7bmostly-Q4_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01560",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 269.49,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M3 Pro; llama 7B mostly Q4_0 | pp512 | 269.49 \u00b1 1.14; llama 7B mostly Q4_0 | tg128 | 30.65 \u00b1 0.20",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7664636",
   "tg_tokens": 128,
   "tps": 30.65,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-25",
   "flags": [],
   "hardware": "M2",
   "id": "lc-disc-4167-c7667693-llama7bmostly-F16",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01493",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 754.39,
   "provenance": "community",
   "quant": "F16",
   "quote": "M2; llama 7B mostly F16 | pp512 | 754.39 \u00b1 0.36; llama 7B mostly F16 | tg128 | 24.31 \u00b1 0.38",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7667693",
   "tg_tokens": 128,
   "tps": 24.31,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-25",
   "flags": [],
   "hardware": "M2",
   "id": "lc-disc-4167-c7667693-llama7bmostly-Q8_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01493",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 671.33,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M2; llama 7B mostly Q8_0 | pp512 | 671.33 \u00b1 2.65; llama 7B mostly Q8_0 | tg128 | 40.85 \u00b1 0.32",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7667693",
   "tg_tokens": 128,
   "tps": 40.85,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-25",
   "flags": [],
   "hardware": "M2",
   "id": "lc-disc-4167-c7667693-llama7bmostly-Q4_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01493",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 664.07,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M2; llama 7B mostly Q4_0 | pp512 | 664.07 \u00b1 9.11; llama 7B mostly Q4_0 | tg128 | 63.29 \u00b1 0.15",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7667693",
   "tg_tokens": 128,
   "tps": 63.29,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "lc-disc-4167-c16049299-llama7bmostly-F16",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1018.3,
   "provenance": "community",
   "quant": "F16",
   "quote": "M5 Max; llama 7B mostly F16 | pp512 | 1018.30 \u00b1 0.89; llama 7B mostly F16 | tg128 | 37.58 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-16049299",
   "tg_tokens": 128,
   "tps": 37.58,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "lc-disc-4167-c16049299-llama7bmostly-Q8_0",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1051.59,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M5 Max; llama 7B mostly Q8_0 | pp512 | 1051.59 \u00b1 0.67; llama 7B mostly Q8_0 | tg128 | 64.61 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-16049299",
   "tg_tokens": 128,
   "tps": 64.61,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "lc-disc-4167-c16049299-llama7bmostly-Q4_0",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 987.1,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M5 Max; llama 7B mostly Q4_0 | pp512 | 987.10 \u00b1 1.05; llama 7B mostly Q4_0 | tg128 | 102.93 \u00b1 0.17",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-16049299",
   "tg_tokens": 128,
   "tps": 102.93,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-31",
   "flags": [],
   "hardware": "M5 Max",
   "id": "lc-disc-4167-c16400323-llama7bmostly-F16",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1032.18,
   "provenance": "community",
   "quant": "F16",
   "quote": "M5 Max; llama 7B mostly F16 | pp512 | 1032.18 \u00b1 1.34; llama 7B mostly F16 | tg128 | 38.06 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-16400323",
   "tg_tokens": 128,
   "tps": 38.06,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-31",
   "flags": [],
   "hardware": "M5 Max",
   "id": "lc-disc-4167-c16400323-llama7bmostly-Q8_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1054.83,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M5 Max; llama 7B mostly Q8_0 | pp512 | 1054.83 \u00b1 0.70; llama 7B mostly Q8_0 | tg128 | 65.11 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-16400323",
   "tg_tokens": 128,
   "tps": 65.11,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-31",
   "flags": [],
   "hardware": "M5 Max",
   "id": "lc-disc-4167-c16400323-llama7bmostly-Q4_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 990.53,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M5 Max; llama 7B mostly Q4_0 | pp512 | 990.53 \u00b1 0.44; llama 7B mostly Q4_0 | tg128 | 104.03 \u00b1 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-16400323",
   "tg_tokens": 128,
   "tps": 104.03,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2026-04-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "lc-disc-4167-c16556242-llama7b-F16",
   "model": "llama 7B",
   "notes": "threads=5; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1310.28,
   "provenance": "community",
   "quant": "F16",
   "quote": "M5 Pro; llama 7B F16 | pp512 | 1310.28 \u00b1 8.21; llama 7B F16 | tg128 | 21.11 \u00b1 0.18",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-16556242",
   "tg_tokens": 128,
   "tps": 21.11,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2026-04-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "lc-disc-4167-c16556242-llama7b-Q8_0",
   "model": "llama 7B",
   "notes": "threads=5; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1304.66,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M5 Pro; llama 7B Q8_0 | pp512 | 1304.66 \u00b1 1.03; llama 7B Q8_0 | tg128 | 37.78 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-16556242",
   "tg_tokens": 128,
   "tps": 37.78,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2026-04-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "lc-disc-4167-c16556242-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=5; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1350.43,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M5 Pro; llama 7B Q4_0 | pp512 | 1350.43 \u00b1 1.97; llama 7B Q4_0 | tg128 | 63.86 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-16556242",
   "tg_tokens": 128,
   "tps": 63.86,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2026-04-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "lc-disc-4167-c16556242-llama7b-F16-2",
   "model": "llama 7B",
   "notes": "threads=5; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 695.51,
   "provenance": "community",
   "quant": "F16",
   "quote": "M5 Pro; llama 7B F16 | pp512 | 695.51 \u00b1 0.74; llama 7B F16 | tg128 | 21.37 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-16556242",
   "tg_tokens": 128,
   "tps": 21.37,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2026-04-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "lc-disc-4167-c16556242-llama7b-Q8_0-2",
   "model": "llama 7B",
   "notes": "threads=5; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 964.18,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M5 Pro; llama 7B Q8_0 | pp512 | 964.18 \u00b1 0.87; llama 7B Q8_0 | tg128 | 37.70 \u00b1 0.14",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-16556242",
   "tg_tokens": 128,
   "tps": 37.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2026-04-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "lc-disc-4167-c16556242-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=5; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1135.71,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M5 Pro; llama 7B Q4_0 | pp512 | 1135.71 \u00b1 9.26; llama 7B Q4_0 | tg128 | 63.74 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-16556242",
   "tg_tokens": 128,
   "tps": 63.74,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2026-04-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "lc-disc-4167-c16556242-llama7bmostly-F16",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 423.3,
   "provenance": "community",
   "quant": "F16",
   "quote": "M5 Pro; llama 7B mostly F16 | pp512 | 423.30 \u00b1 0.14; llama 7B mostly F16 | tg128 | 20.22 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-16556242",
   "tg_tokens": 128,
   "tps": 20.22,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2026-04-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "lc-disc-4167-c16556242-llama7bmostly-Q8_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 431.14,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M5 Pro; llama 7B mostly Q8_0 | pp512 | 431.14 \u00b1 0.10; llama 7B mostly Q8_0 | tg128 | 35.86 \u00b1 0.24",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-16556242",
   "tg_tokens": 128,
   "tps": 35.86,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2026-04-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "lc-disc-4167-c16556242-llama7bmostly-Q4_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 403.19,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M5 Pro; llama 7B mostly Q4_0 | pp512 | 403.19 \u00b1 0.22; llama 7B mostly Q4_0 | tg128 | 60.04 \u00b1 0.21",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-16556242",
   "tg_tokens": 128,
   "tps": 60.04,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-12",
   "flags": [],
   "hardware": "M4",
   "id": "lc-disc-4167-c16886673-llama3b-Q4_K - Medium",
   "model": "llama 3B",
   "notes": "threads=4; config=b09100",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 530.01,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "M4; llama 3B Q4_K - Medium | pp512 | 530.01 \u00b1 0.86; llama 3B Q4_K - Medium | tg128 | 45.95 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-16886673",
   "tg_tokens": 128,
   "tps": 45.95,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-12",
   "flags": [],
   "hardware": "M4",
   "id": "lc-disc-4167-c16886673-qwen27b-Q4_K - Medium",
   "model": "qwen2 7B",
   "notes": "threads=4; config=b09100",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 239.73,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "M4; qwen2 7B Q4_K - Medium | pp512 | 239.73 \u00b1 2.07; qwen2 7B Q4_K - Medium | tg128 | 22.60 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-16886673",
   "tg_tokens": 128,
   "tps": 22.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-22",
   "flags": [],
   "hardware": "M5 Max",
   "id": "lc-disc-4167-c17386908-llama7b-F16",
   "model": "llama 7B",
   "notes": "threads=6; config=b09754",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3189.9,
   "provenance": "community",
   "quant": "F16",
   "quote": "M5 Max; llama 7B F16 | pp512 | 3189.90 \u00b1 8.62; llama 7B F16 | tg128 | 35.39 \u00b1 0.10",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-17386908",
   "tg_tokens": 128,
   "tps": 35.39,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-22",
   "flags": [],
   "hardware": "M5 Max",
   "id": "lc-disc-4167-c17386908-llama7b-Q8_0",
   "model": "llama 7B",
   "notes": "threads=6; config=b09754",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3219.94,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M5 Max; llama 7B Q8_0 | pp512 | 3219.94 \u00b1 2.06; llama 7B Q8_0 | tg128 | 72.52 \u00b1 0.11",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-17386908",
   "tg_tokens": 128,
   "tps": 72.52,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-22",
   "flags": [],
   "hardware": "M5 Max",
   "id": "lc-disc-4167-c17386908-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=6; config=b09754",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3347.44,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M5 Max; llama 7B Q4_0 | pp512 | 3347.44 \u00b1 1.53; llama 7B Q4_0 | tg128 | 119.08 \u00b1 0.69",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-17386908",
   "tg_tokens": 128,
   "tps": 119.08,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-25",
   "flags": [],
   "hardware": "M5 Max",
   "id": "lc-disc-4167-c18152504-llama7b-F16",
   "model": "llama 7B",
   "notes": "threads=6; config=b10621",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3158.49,
   "provenance": "community",
   "quant": "F16",
   "quote": "M5 Max; llama 7B F16 | pp512 | 3158.49 \u00b1 41.07; llama 7B F16 | tg128 | 37.11 \u00b1 0.17",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-18152504",
   "tg_tokens": 128,
   "tps": 37.11,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-25",
   "flags": [],
   "hardware": "M5 Max",
   "id": "lc-disc-4167-c18152504-llama7b-Q8_0",
   "model": "llama 7B",
   "notes": "threads=6; config=b10621",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3143.81,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M5 Max; llama 7B Q8_0 | pp512 | 3143.81 \u00b1 29.60; llama 7B Q8_0 | tg128 | 72.42 \u00b1 0.10",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-18152504",
   "tg_tokens": 128,
   "tps": 72.42,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-25",
   "flags": [],
   "hardware": "M5 Max",
   "id": "lc-disc-4167-c18152504-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=6; config=b10621",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3219.99,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M5 Max; llama 7B Q4_0 | pp512 | 3219.99 \u00b1 36.79; llama 7B Q4_0 | tg128 | 119.92 \u00b1 0.17",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-18152504",
   "tg_tokens": 128,
   "tps": 119.92,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-28",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "lc-disc-4167-c18191340-llama7b-F16",
   "model": "llama 7B",
   "notes": "threads=16; config=b10621",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1294.31,
   "provenance": "community",
   "quant": "F16",
   "quote": "M1 Ultra; llama 7B F16 | pp512 | 1294.31 \u00b1 0.51; llama 7B F16 | tg128 | 47.82 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-18191340",
   "tg_tokens": 128,
   "tps": 47.82,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-28",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "lc-disc-4167-c18191340-llama7b-Q8_0",
   "model": "llama 7B",
   "notes": "threads=16; config=b10621",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1167.67,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M1 Ultra; llama 7B Q8_0 | pp512 | 1167.67 \u00b1 2.97; llama 7B Q8_0 | tg128 | 79.13 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-18191340",
   "tg_tokens": 128,
   "tps": 79.13,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-28",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "lc-disc-4167-c18191340-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=16; config=b10621",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1179.89,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M1 Ultra; llama 7B Q4_0 | pp512 | 1179.89 \u00b1 1.71; llama 7B Q4_0 | tg128 | 113.79 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-18191340",
   "tg_tokens": 128,
   "tps": 113.79,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-30",
   "flags": [],
   "hardware": "M5",
   "id": "lc-disc-4167-c18210351-llama7b-Q8_0",
   "model": "llama 7B",
   "notes": "threads=4; config=b10621",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 715.42,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M5; llama 7B Q8_0 | pp512 | 715.42 \u00b1 4.90; llama 7B Q8_0 | tg128 | 18.42 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-18210351",
   "tg_tokens": 128,
   "tps": 18.42,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-30",
   "flags": [],
   "hardware": "M5",
   "id": "lc-disc-4167-c18210351-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=4; config=b10621",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 722.79,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M5; llama 7B Q4_0 | pp512 | 722.79 \u00b1 3.30; llama 7B Q4_0 | tg128 | 31.88 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-18210351",
   "tg_tokens": 128,
   "tps": 31.88,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-31",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "lc-disc-4167-c18213344-llama7b-F16",
   "model": "llama 7B",
   "notes": "threads=6",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1588.78,
   "provenance": "community",
   "quant": "F16",
   "quote": "M5 Pro; llama 7B F16 | pp512 | 1588.78 \u00b1 3.16; llama 7B F16 | tg128 | 21.55 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-18213344",
   "tg_tokens": 128,
   "tps": 21.55,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-31",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "lc-disc-4167-c18213344-llama7b-Q8_0",
   "model": "llama 7B",
   "notes": "threads=6",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1553.86,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M5 Pro; llama 7B Q8_0 | pp512 | 1553.86 \u00b1 2.18; llama 7B Q8_0 | tg128 | 38.92 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-18213344",
   "tg_tokens": 128,
   "tps": 38.92,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-31",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "lc-disc-4167-c18213344-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=6",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1620.64,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M5 Pro; llama 7B Q4_0 | pp512 | 1620.64 \u00b1 1.16; llama 7B Q4_0 | tg128 | 66.33 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-18213344",
   "tg_tokens": 128,
   "tps": 66.33,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-24",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "lc-disc-4167-c18583406-llama7b-F16",
   "model": "llama 7B",
   "notes": "threads=5",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1310.78,
   "provenance": "community",
   "quant": "F16",
   "quote": "M5 Pro; llama 7B F16 | pp512 | 1310.78 \u00b1 1.33; llama 7B F16 | tg128 | 21.39 \u00b1 0.53",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-18583406",
   "tg_tokens": 128,
   "tps": 21.39,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-24",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "lc-disc-4167-c18583406-llama7b-Q8_0",
   "model": "llama 7B",
   "notes": "threads=5",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1302.59,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M5 Pro; llama 7B Q8_0 | pp512 | 1302.59 \u00b1 2.05; llama 7B Q8_0 | tg128 | 38.94 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-18583406",
   "tg_tokens": 128,
   "tps": 38.94,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-24",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "lc-disc-4167-c18583406-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=5",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1340.44,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M5 Pro; llama 7B Q4_0 | pp512 | 1340.44 \u00b1 1.69; llama 7B Q4_0 | tg128 | 67.59 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-18583406",
   "tg_tokens": 128,
   "tps": 67.59,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-27",
   "flags": [],
   "hardware": "M3",
   "id": "lc-disc-4167-c7677748-llama7bmostly-Q8_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; ngl=99; config=b01570",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 187.52,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M3; llama 7B mostly Q8_0 | pp512 | 187.52 \u00b1 0.15; llama 7B mostly Q8_0 | tg128 | 12.27 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7677748",
   "tg_tokens": 128,
   "tps": 12.27,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-27",
   "flags": [],
   "hardware": "M3",
   "id": "lc-disc-4167-c7677748-llama7bmostly-Q4_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; ngl=99; config=b01570",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 186.75,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M3; llama 7B mostly Q4_0 | pp512 | 186.75 \u00b1 0.08; llama 7B mostly Q4_0 | tg128 | 21.34 \u00b1 0.59",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7677748",
   "tg_tokens": 128,
   "tps": 21.34,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-29",
   "flags": [],
   "hardware": "M2 Max",
   "id": "lc-disc-4167-c7711079-llama7bmostly-F16",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01575",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 600.46,
   "provenance": "community",
   "quant": "F16",
   "quote": "M2 Max; llama 7B mostly F16 | pp512 | 600.46 \u00b1 4.30; llama 7B mostly F16 | tg128 | 24.16 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7711079",
   "tg_tokens": 128,
   "tps": 24.16,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-29",
   "flags": [],
   "hardware": "M2 Max",
   "id": "lc-disc-4167-c7711079-llama7bmostly-Q8_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01575",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 540.15,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M2 Max; llama 7B mostly Q8_0 | pp512 | 540.15 \u00b1 0.37; llama 7B mostly Q8_0 | tg128 | 39.97 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7711079",
   "tg_tokens": 128,
   "tps": 39.97,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-11-29",
   "flags": [],
   "hardware": "M2 Max",
   "id": "lc-disc-4167-c7711079-llama7bmostly-Q4_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01575",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 537.6,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M2 Max; llama 7B mostly Q4_0 | pp512 | 537.60 \u00b1 0.31; llama 7B mostly Q4_0 | tg128 | 60.99 \u00b1 0.19",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7711079",
   "tg_tokens": 128,
   "tps": 60.99,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-01",
   "flags": [],
   "hardware": "M3",
   "id": "lc-disc-4167-c7727765-llama7bmostly-Q8_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 339.31,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M3; llama 7B mostly Q8_0 | pp512 | 339.31 \u00b1 0.40; llama 7B mostly Q8_0 | tg128 | 16.46 \u00b1 0.26",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7727765",
   "tg_tokens": 128,
   "tps": 16.46,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-01",
   "flags": [],
   "hardware": "M3",
   "id": "lc-disc-4167-c7727765-llama7bmostly-Q4_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 338.37,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M3; llama 7B mostly Q4_0 | pp512 | 338.37 \u00b1 1.88; llama 7B mostly Q4_0 | tg128 | 30.36 \u00b1 0.70",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7727765",
   "tg_tokens": 128,
   "tps": 30.36,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-01",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "lc-disc-4167-c7730376-llama7bmostly-F16",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 875.81,
   "provenance": "community",
   "quant": "F16",
   "quote": "M1 Ultra; llama 7B mostly F16 | pp512 | 875.81 \u00b1 0.28; llama 7B mostly F16 | tg128 | 33.92 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7730376",
   "tg_tokens": 128,
   "tps": 33.92,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-01",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "lc-disc-4167-c7730376-llama7bmostly-Q8_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 783.45,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M1 Ultra; llama 7B mostly Q8_0 | pp512 | 783.45 \u00b1 0.08; llama 7B mostly Q8_0 | tg128 | 55.69 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7730376",
   "tg_tokens": 128,
   "tps": 55.69,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-01",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "lc-disc-4167-c7730376-llama7bmostly-Q4_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 772.24,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M1 Ultra; llama 7B mostly Q4_0 | pp512 | 772.24 \u00b1 0.15; llama 7B mostly Q4_0 | tg128 | 74.93 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7730376",
   "tg_tokens": 128,
   "tps": 74.93,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-02",
   "flags": [],
   "hardware": "M1",
   "id": "lc-disc-4167-c7737353-llama7bmostly-Q4_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; ngl=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 117.96,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M1; llama 7B mostly Q4_0 | pp512 | 117.96 \u00b1 0.51; llama 7B mostly Q4_0 | tg128 | 14.15 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7737353",
   "tg_tokens": 128,
   "tps": 14.15,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-02",
   "flags": [],
   "hardware": "M1",
   "id": "lc-disc-4167-c7737353-llama7bmostly-Q8_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; ngl=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 117.25,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M1; llama 7B mostly Q8_0 | pp512 | 117.25 \u00b1 0.04; llama 7B mostly Q8_0 | tg128 | 7.91 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7737353",
   "tg_tokens": 128,
   "tps": 7.91,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-02",
   "flags": [],
   "hardware": "M1 Max",
   "id": "lc-disc-4167-c7737353-llama7bmostly-F16",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 453.03,
   "provenance": "community",
   "quant": "F16",
   "quote": "M1 Max; llama 7B mostly F16 | pp512 | 453.03 \u00b1 0.23; llama 7B mostly F16 | tg128 | 22.55 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7737353",
   "tg_tokens": 128,
   "tps": 22.55,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-02",
   "flags": [],
   "hardware": "M1 Max",
   "id": "lc-disc-4167-c7737353-llama7bmostly-Q8_0-2",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 405.87,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M1 Max; llama 7B mostly Q8_0 | pp512 | 405.87 \u00b1 0.09; llama 7B mostly Q8_0 | tg128 | 37.81 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7737353",
   "tg_tokens": 128,
   "tps": 37.81,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-02",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "lc-disc-4167-c7737353-llama7bmostly-Q4_0-2",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 400.26,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M1 Max; llama 7B mostly Q4_0 | pp512 | 400.26 \u00b1 0.12; llama 7B mostly Q4_0 | tg128 | 54.61 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7737353",
   "tg_tokens": 128,
   "tps": 54.61,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-04",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "lc-disc-4167-c7750625-llama7bmostly-F16",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1408.01,
   "provenance": "community",
   "quant": "F16",
   "quote": "M2 Ultra; llama 7B mostly F16 | pp512 | 1408.01 \u00b1 1.05; llama 7B mostly F16 | tg128 | 41.34 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7750625",
   "tg_tokens": 128,
   "tps": 41.34,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-04",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "lc-disc-4167-c7750625-llama7bmostly-Q8_0",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1250.86,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M2 Ultra; llama 7B mostly Q8_0 | pp512 | 1250.86 \u00b1 1.02; llama 7B mostly Q8_0 | tg128 | 67.36 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7750625",
   "tg_tokens": 128,
   "tps": 67.36,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-04",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "lc-disc-4167-c7750625-llama7bmostly-Q4_0",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1240.61,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M2 Ultra; llama 7B mostly Q4_0 | pp512 | 1240.61 \u00b1 0.76; llama 7B mostly Q4_0 | tg128 | 95.34 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7750625",
   "tg_tokens": 128,
   "tps": 95.34,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-05",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2",
   "id": "lc-disc-4167-c7764006-llama7bmostly-Q8_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 76.9,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M2; llama 7B mostly Q8_0 | pp512 | 76.90 \u00b1 0.80; llama 7B mostly Q8_0 | tg128 | 5.31 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7764006",
   "tg_tokens": 128,
   "tps": 5.31,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-05",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2",
   "id": "lc-disc-4167-c7764006-llama7bmostly-Q4_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 85.24,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M2; llama 7B mostly Q4_0 | pp512 | 85.24 \u00b1 0.74; llama 7B mostly Q4_0 | tg128 | 8.61 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7764006",
   "tg_tokens": 128,
   "tps": 8.61,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-09",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "lc-disc-4167-c7804807-llama7bmostly-F16",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 357.45,
   "provenance": "community",
   "quant": "F16",
   "quote": "M3 Pro; llama 7B mostly F16 | pp512 | 357.45 \u00b1 0.46; llama 7B mostly F16 | tg128 | 9.89 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7804807",
   "tg_tokens": 128,
   "tps": 9.89,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-09",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "lc-disc-4167-c7804807-llama7bmostly-Q8_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 344.66,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M3 Pro; llama 7B mostly Q8_0 | pp512 | 344.66 \u00b1 0.29; llama 7B mostly Q8_0 | tg128 | 17.53 \u00b1 0.17",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7804807",
   "tg_tokens": 128,
   "tps": 17.53,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-09",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "lc-disc-4167-c7804807-llama7bmostly-Q4_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 341.67,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M3 Pro; llama 7B mostly Q4_0 | pp512 | 341.67 \u00b1 4.07; llama 7B mostly Q4_0 | tg128 | 30.74 \u00b1 0.21",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7804807",
   "tg_tokens": 128,
   "tps": 30.74,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-11",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "lc-disc-4167-c7824473-llama7bmostly-F16",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 384.38,
   "provenance": "community",
   "quant": "F16",
   "quote": "M2 Pro; llama 7B mostly F16 | pp512 | 384.38 \u00b1 0.07; llama 7B mostly F16 | tg128 | 13.06 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7824473",
   "tg_tokens": 128,
   "tps": 13.06,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-11",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "lc-disc-4167-c7824473-llama7bmostly-Q8_0",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 344.5,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M2 Pro; llama 7B mostly Q8_0 | pp512 | 344.50 \u00b1 0.06; llama 7B mostly Q8_0 | tg128 | 23.01 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7824473",
   "tg_tokens": 128,
   "tps": 23.01,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-11",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "lc-disc-4167-c7824473-llama7bmostly-Q4_0",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 341.19,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M2 Pro; llama 7B mostly Q4_0 | pp512 | 341.19 \u00b1 0.10; llama 7B mostly Q4_0 | tg128 | 38.86 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7824473",
   "tg_tokens": 128,
   "tps": 38.86,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2025-03-14",
   "flags": [],
   "hardware": "M3 Ultra 20+8",
   "id": "lc-disc-4167-c12495617-llama7bmostly-F16",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1121.8,
   "provenance": "community",
   "quant": "F16",
   "quote": "M3 Ultra 20+8; llama 7B mostly F16 | pp512 | 1121.80 \u00b1 2.33; llama 7B mostly F16 | tg128 | 42.24 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-12495617",
   "tg_tokens": 128,
   "tps": 42.24,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2025-03-14",
   "flags": [],
   "hardware": "M3 Ultra 20+8",
   "id": "lc-disc-4167-c12495617-llama7bmostly-Q8_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1085.76,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M3 Ultra 20+8; llama 7B mostly Q8_0 | pp512 | 1085.76 \u00b1 0.90; llama 7B mostly Q8_0 | tg128 | 63.55 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-12495617",
   "tg_tokens": 128,
   "tps": 63.55,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2025-03-14",
   "flags": [],
   "hardware": "M3 Ultra 20+8",
   "id": "lc-disc-4167-c12495617-llama7bmostly-Q4_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1073.09,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M3 Ultra 20+8; llama 7B mostly Q4_0 | pp512 | 1073.09 \u00b1 1.29; llama 7B mostly Q4_0 | tg128 | 88.40 \u00b1 0.44",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-12495617",
   "tg_tokens": 128,
   "tps": 88.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2025-03-23",
   "flags": [],
   "hardware": "M3 Ultra 24+8",
   "id": "lc-disc-4167-c12594707-llama7bmostly-F16",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1538.34,
   "provenance": "community",
   "quant": "F16",
   "quote": "M3 Ultra 24+8; llama 7B mostly F16 | pp512 | 1538.34 \u00b1 2.14; llama 7B mostly F16 | tg128 | 39.78 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-12594707",
   "tg_tokens": 128,
   "tps": 39.78,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2025-03-23",
   "flags": [],
   "hardware": "M3 Ultra 24+8",
   "id": "lc-disc-4167-c12594707-llama7bmostly-Q8_0",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1487.51,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M3 Ultra 24+8; llama 7B mostly Q8_0 | pp512 | 1487.51 \u00b1 1.57; llama 7B mostly Q8_0 | tg128 | 63.93 \u00b1 0.24",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-12594707",
   "tg_tokens": 128,
   "tps": 63.93,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2025-03-23",
   "flags": [],
   "hardware": "M3 Ultra 24+8",
   "id": "lc-disc-4167-c12594707-llama7bmostly-Q4_0",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1471.24,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M3 Ultra 24+8; llama 7B mostly Q4_0 | pp512 | 1471.24 \u00b1 1.05; llama 7B mostly Q4_0 | tg128 | 92.14 \u00b1 0.66",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-12594707",
   "tg_tokens": 128,
   "tps": 92.14,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2025-04-22",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "lc-disc-4167-c12913344-llama7bmostly-F16",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1527.74,
   "provenance": "community",
   "quant": "F16",
   "quote": "M3 Ultra; llama 7B mostly F16 | pp512 | 1527.74 \u00b1 2.02; llama 7B mostly F16 | tg128 | 40.10 \u00b1 0.10",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-12913344",
   "tg_tokens": 128,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2025-04-22",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "lc-disc-4167-c12913344-llama7bmostly-Q8_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1488.84,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M3 Ultra; llama 7B mostly Q8_0 | pp512 | 1488.84 \u00b1 2.52; llama 7B mostly Q8_0 | tg128 | 64.16 \u00b1 0.38",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-12913344",
   "tg_tokens": 128,
   "tps": 64.16,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2025-04-22",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "lc-disc-4167-c12913344-llama7bmostly-Q4_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1473.76,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M3 Ultra; llama 7B mostly Q4_0 | pp512 | 1473.76 \u00b1 1.09; llama 7B mostly Q4_0 | tg128 | 91.93 \u00b1 0.48",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-12913344",
   "tg_tokens": 128,
   "tps": 91.93,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-20",
   "flags": [],
   "hardware": "M1",
   "id": "lc-disc-4167-c13208048-llama7bmostly-Q4_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 115.67,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M1; llama 7B mostly Q4_0 | pp512 | 115.67 \u00b1 0.88; llama 7B mostly Q4_0 | tg128 | 14.13 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-13208048",
   "tg_tokens": 128,
   "tps": 14.13,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-20",
   "flags": [],
   "hardware": "M1",
   "id": "lc-disc-4167-c13208048-llama7bmostly-Q8_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 121.73,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M1; llama 7B mostly Q8_0 | pp512 | 121.73 \u00b1 1.43; llama 7B mostly Q8_0 | tg128 | 7.69 \u00b1 0.12",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-13208048",
   "tg_tokens": 128,
   "tps": 7.69,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-20",
   "flags": [],
   "hardware": "M1",
   "id": "lc-disc-4167-c13208048-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=4",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 131.46,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M1; llama 7B Q4_0 | pp512 | 131.46 \u00b1 6.71; llama 7B Q4_0 | tg128 | 13.99 \u00b1 0.14",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-13208048",
   "tg_tokens": 128,
   "tps": 13.99,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-20",
   "flags": [],
   "hardware": "M1",
   "id": "lc-disc-4167-c13208048-llama7b-Q8_0",
   "model": "llama 7B",
   "notes": "threads=4",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 133.34,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M1; llama 7B Q8_0 | pp512 | 133.34 \u00b1 1.17; llama 7B Q8_0 | tg128 | 7.67 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-13208048",
   "tg_tokens": 128,
   "tps": 7.67,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-20",
   "flags": [],
   "hardware": "M4",
   "id": "lc-disc-4167-c13208048-llama7bmostly-F16",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 747.59,
   "provenance": "community",
   "quant": "F16",
   "quote": "M4; llama 7B mostly F16 | pp512 | 747.59 \u00b1 0.92; llama 7B mostly F16 | tg128 | 25.58 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-13208048",
   "tg_tokens": 128,
   "tps": 25.58,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-20",
   "flags": [],
   "hardware": "M4",
   "id": "lc-disc-4167-c13208048-llama7bmostly-Q8_0-2",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 720.38,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M4; llama 7B mostly Q8_0 | pp512 | 720.38 \u00b1 0.04; llama 7B mostly Q8_0 | tg128 | 43.80 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-13208048",
   "tg_tokens": 128,
   "tps": 43.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-20",
   "flags": [],
   "hardware": "M4",
   "id": "lc-disc-4167-c13208048-llama7bmostly-Q4_0-2",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 715.74,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M4; llama 7B mostly Q4_0 | pp512 | 715.74 \u00b1 0.52; llama 7B mostly Q4_0 | tg128 | 69.24 \u00b1 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-13208048",
   "tg_tokens": 128,
   "tps": 69.24,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-20",
   "flags": [],
   "hardware": "M4",
   "id": "lc-disc-4167-c13208048-llama7b-F16",
   "model": "llama 7B",
   "notes": "threads=10",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 790.33,
   "provenance": "community",
   "quant": "F16",
   "quote": "M4; llama 7B F16 | pp512 | 790.33 \u00b1 0.49; llama 7B F16 | tg128 | 26.05 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-13208048",
   "tg_tokens": 128,
   "tps": 26.05,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-20",
   "flags": [],
   "hardware": "M4",
   "id": "lc-disc-4167-c13208048-llama7b-Q8_0-2",
   "model": "llama 7B",
   "notes": "threads=10",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 702.39,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M4; llama 7B Q8_0 | pp512 | 702.39 \u00b1 11.36; llama 7B Q8_0 | tg128 | 44.89 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-13208048",
   "tg_tokens": 128,
   "tps": 44.89,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-20",
   "flags": [],
   "hardware": "M4",
   "id": "lc-disc-4167-c13208048-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=10",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 762.81,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M4; llama 7B Q4_0 | pp512 | 762.81 \u00b1 1.03; llama 7B Q4_0 | tg128 | 72.26 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-13208048",
   "tg_tokens": 128,
   "tps": 72.26,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-30",
   "flags": [],
   "hardware": "M5",
   "id": "lc-disc-4167-c14821709-llama7bmostly-F16",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 257.79,
   "provenance": "community",
   "quant": "F16",
   "quote": "M5; llama 7B mostly F16 | pp512 | 257.79 \u00b1 0.94; llama 7B mostly F16 | tg128 | 9.43 \u00b1 0.12",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-14821709",
   "tg_tokens": 128,
   "tps": 9.43,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-30",
   "flags": [],
   "hardware": "M5",
   "id": "lc-disc-4167-c14821709-llama7bmostly-Q8_0",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 264.15,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M5; llama 7B mostly Q8_0 | pp512 | 264.15 \u00b1 2.26; llama 7B mostly Q8_0 | tg128 | 16.62 \u00b1 0.15",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-14821709",
   "tg_tokens": 128,
   "tps": 16.62,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-30",
   "flags": [],
   "hardware": "M5",
   "id": "lc-disc-4167-c14821709-llama7bmostly-Q4_0",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 247.68,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M5; llama 7B mostly Q4_0 | pp512 | 247.68 \u00b1 2.63; llama 7B mostly Q4_0 | tg128 | 29.62 \u00b1 0.33",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-14821709",
   "tg_tokens": 128,
   "tps": 29.62,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-30",
   "flags": [],
   "hardware": "M5",
   "id": "lc-disc-4167-c14821709-llama7b-F16",
   "model": "llama 7B",
   "notes": "threads=4",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 374.35,
   "provenance": "community",
   "quant": "F16",
   "quote": "M5; llama 7B F16 | pp512 | 374.35 \u00b1 1.95; llama 7B F16 | tg128 | 9.67 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-14821709",
   "tg_tokens": 128,
   "tps": 9.67,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-30",
   "flags": [],
   "hardware": "M5",
   "id": "lc-disc-4167-c14821709-llama7b-Q8_0",
   "model": "llama 7B",
   "notes": "threads=4",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 489.78,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M5; llama 7B Q8_0 | pp512 | 489.78 \u00b1 3.90; llama 7B Q8_0 | tg128 | 17.50 \u00b1 0.38",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-14821709",
   "tg_tokens": 128,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-30",
   "flags": [],
   "hardware": "M5",
   "id": "lc-disc-4167-c14821709-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=4",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 636.36,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M5; llama 7B Q4_0 | pp512 | 636.36 \u00b1 1.92; llama 7B Q4_0 | tg128 | 31.02 \u00b1 0.37",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-14821709",
   "tg_tokens": 128,
   "tps": 31.02,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-30",
   "flags": [],
   "hardware": "M5",
   "id": "lc-disc-4167-c14821709-llama7b-F16-2",
   "model": "llama 7B",
   "notes": "threads=4",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 378.36,
   "provenance": "community",
   "quant": "F16",
   "quote": "M5; llama 7B F16 | pp512 | 378.36 \u00b1 4.69; llama 7B F16 | tg128 | 9.56 \u00b1 0.11",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-14821709",
   "tg_tokens": 128,
   "tps": 9.56,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-30",
   "flags": [],
   "hardware": "M5",
   "id": "lc-disc-4167-c14821709-llama7b-Q8_0-2",
   "model": "llama 7B",
   "notes": "threads=4",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 512.15,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M5; llama 7B Q8_0 | pp512 | 512.15 \u00b1 2.73; llama 7B Q8_0 | tg128 | 17.55 \u00b1 0.24",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-14821709",
   "tg_tokens": 128,
   "tps": 17.55,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-30",
   "flags": [],
   "hardware": "M5",
   "id": "lc-disc-4167-c14821709-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=4",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 650.65,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M5; llama 7B Q4_0 | pp512 | 650.65 \u00b1 5.06; llama 7B Q4_0 | tg128 | 30.67 \u00b1 0.25",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-14821709",
   "tg_tokens": 128,
   "tps": 30.67,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-15",
   "flags": [],
   "hardware": "M1",
   "id": "lc-disc-4167-c7861106-llama7bmostly-Q4_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; ngl=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 107.81,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M1; llama 7B mostly Q4_0 | pp512 | 107.81 \u00b1 0.10; llama 7B mostly Q4_0 | tg128 | 14.19 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7861106",
   "tg_tokens": 128,
   "tps": 14.19,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-15",
   "flags": [],
   "hardware": "M1",
   "id": "lc-disc-4167-c7861106-llama7bmostly-Q8_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; ngl=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 108.21,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M1; llama 7B mostly Q8_0 | pp512 | 108.21 \u00b1 0.27; llama 7B mostly Q8_0 | tg128 | 7.92 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7861106",
   "tg_tokens": 128,
   "tps": 7.92,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-17",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "lc-disc-4167-c7876137-llama7bmostly-Q8_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 236.72,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M1 Pro; llama 7B mostly Q8_0 | pp512 | 236.72 \u00b1 0.08; llama 7B mostly Q8_0 | tg128 | 22.16 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7876137",
   "tg_tokens": 128,
   "tps": 22.16,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-17",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "lc-disc-4167-c7876137-llama7bmostly-Q4_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 233.01,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M1 Pro; llama 7B mostly Q4_0 | pp512 | 233.01 \u00b1 0.08; llama 7B mostly Q4_0 | tg128 | 35.45 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7876137",
   "tg_tokens": 128,
   "tps": 35.45,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-17",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "lc-disc-4167-c7876137-llama7bmostly-Q8_0-2",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 201.58,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M1 Pro; llama 7B mostly Q8_0 | pp512 | 201.58 \u00b1 2.62; llama 7B mostly Q8_0 | tg128 | 20.55 \u00b1 0.14",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7876137",
   "tg_tokens": 128,
   "tps": 20.55,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-17",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "lc-disc-4167-c7876137-llama7bmostly-Q4_0-2",
   "model": "llama 7B mostly",
   "notes": "threads=99; ngl=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 205.41,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M1 Pro; llama 7B mostly Q4_0 | pp512 | 205.41 \u00b1 0.83; llama 7B mostly Q4_0 | tg128 | 30.60 \u00b1 0.29",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7876137",
   "tg_tokens": 128,
   "tps": 30.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "lc-disc-4167-c7877105-llama7bmostly-F16",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 596.85,
   "provenance": "community",
   "quant": "F16",
   "quote": "M1 Max; llama 7B mostly F16 | pp512 | 596.85 \u00b1 0.54; llama 7B mostly F16 | tg128 | 22.96 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7877105",
   "tg_tokens": 128,
   "tps": 22.96,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "lc-disc-4167-c7877105-llama7bmostly-Q8_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 537.23,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M1 Max; llama 7B mostly Q8_0 | pp512 | 537.23 \u00b1 0.23; llama 7B mostly Q8_0 | tg128 | 40.17 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7877105",
   "tg_tokens": 128,
   "tps": 40.17,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "lc-disc-4167-c7877105-llama7bmostly-Q4_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 529.4,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M1 Max; llama 7B mostly Q4_0 | pp512 | 529.40 \u00b1 0.94; llama 7B mostly Q4_0 | tg128 | 60.96 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7877105",
   "tg_tokens": 128,
   "tps": 60.96,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "lc-disc-4167-c7877105-llama7bmostly-F16-2",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2757.29,
   "provenance": "community",
   "quant": "F16",
   "quote": "M1 Max; llama 7B mostly F16 | pp512 | 2757.29 \u00b1 77.35; llama 7B mostly F16 | tg128 | 19.45 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7877105",
   "tg_tokens": 128,
   "tps": 19.45,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "lc-disc-4167-c7877105-llama7bmostly-Q8_0-2",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1873.64,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M1 Max; llama 7B mostly Q8_0 | pp512 | 1873.64 \u00b1 10.51; llama 7B mostly Q8_0 | tg128 | 35.02 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7877105",
   "tg_tokens": 128,
   "tps": 35.02,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "lc-disc-4167-c7877105-llama7bmostly-Q4_0-2",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1945.87,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M1 Max; llama 7B mostly Q4_0 | pp512 | 1945.87 \u00b1 1.90; llama 7B mostly Q4_0 | tg128 | 61.39 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7877105",
   "tg_tokens": 128,
   "tps": 61.39,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "lc-disc-4167-c7877105-llama7bmostly-F16-3",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 4096,
   "pp_tps": 516.68,
   "provenance": "community",
   "quant": "F16",
   "quote": "M1 Max; llama 7B mostly F16 | pp4096 | 516.68 \u00b1 0.11; llama 7B mostly F16 | tg1024 | 22.41 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7877105",
   "tg_tokens": 1024,
   "tps": 22.41,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "lc-disc-4167-c7877105-llama7bmostly-Q8_0-3",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 4096,
   "pp_tps": 471.64,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M1 Max; llama 7B mostly Q8_0 | pp4096 | 471.64 \u00b1 0.24; llama 7B mostly Q8_0 | tg1024 | 38.50 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7877105",
   "tg_tokens": 1024,
   "tps": 38.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "lc-disc-4167-c7877105-llama7bmostly-Q4_0-3",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 4096,
   "pp_tps": 466.05,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M1 Max; llama 7B mostly Q4_0 | pp4096 | 466.05 \u00b1 0.10; llama 7B mostly Q4_0 | tg1024 | 57.39 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7877105",
   "tg_tokens": 1024,
   "tps": 57.39,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "lc-disc-4167-c7877105-llama7bmostly-F16-4",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 4096,
   "pp_tps": 1453.27,
   "provenance": "community",
   "quant": "F16",
   "quote": "M1 Max; llama 7B mostly F16 | pp4096 | 1453.27 \u00b1 14.35; llama 7B mostly F16 | tg1024 | 19.10 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7877105",
   "tg_tokens": 1024,
   "tps": 19.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "lc-disc-4167-c7877105-llama7bmostly-Q8_0-4",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 4096,
   "pp_tps": 1157.46,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M1 Max; llama 7B mostly Q8_0 | pp4096 | 1157.46 \u00b1 7.28; llama 7B mostly Q8_0 | tg1024 | 33.86 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7877105",
   "tg_tokens": 1024,
   "tps": 33.86,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "lc-disc-4167-c7877105-llama7bmostly-Q4_0-4",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 4096,
   "pp_tps": 1183.28,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M1 Max; llama 7B mostly Q4_0 | pp4096 | 1183.28 \u00b1 8.18; llama 7B mostly Q4_0 | tg1024 | 57.87 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7877105",
   "tg_tokens": 1024,
   "tps": 57.87,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-21",
   "flags": [],
   "hardware": "M1",
   "id": "lc-disc-4167-c7921945-llama7bmostly-F16",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 598.96,
   "provenance": "community",
   "quant": "F16",
   "quote": "M1; llama 7B mostly F16 | pp512 | 598.96 \u00b1 1.05; llama 7B mostly F16 | tg128 | 22.83 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7921945",
   "tg_tokens": 128,
   "tps": 22.83,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-21",
   "flags": [],
   "hardware": "M1",
   "id": "lc-disc-4167-c7921945-llama7bmostly-Q8_0",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 536.5,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M1; llama 7B mostly Q8_0 | pp512 | 536.50 \u00b1 1.08; llama 7B mostly Q8_0 | tg128 | 39.60 \u00b1 0.20",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7921945",
   "tg_tokens": 128,
   "tps": 39.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-21",
   "flags": [],
   "hardware": "M1",
   "id": "lc-disc-4167-c7921945-llama7bmostly-Q4_0",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 527.76,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M1; llama 7B mostly Q4_0 | pp512 | 527.76 \u00b1 3.68; llama 7B mostly Q4_0 | tg128 | 60.40 \u00b1 0.11",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7921945",
   "tg_tokens": 128,
   "tps": 60.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-22",
   "flags": [],
   "hardware": "M3 Max 10+4",
   "id": "lc-disc-4167-c7931329-llama7bmostly-F16",
   "model": "llama 7B mostly",
   "notes": "threads=99; ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 589.41,
   "provenance": "community",
   "quant": "F16",
   "quote": "M3 Max 10+4; llama 7B mostly F16 | pp512 | 589.41 \u00b1 0.31; llama 7B mostly F16 | tg128 | 19.54 \u00b1 0.12",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7931329",
   "tg_tokens": 128,
   "tps": 19.54,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-22",
   "flags": [],
   "hardware": "M3 Max 10+4",
   "id": "lc-disc-4167-c7931329-llama7bmostly-Q8_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 566.4,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M3 Max 10+4; llama 7B mostly Q8_0 | pp512 | 566.40 \u00b1 3.32; llama 7B mostly Q8_0 | tg128 | 34.30 \u00b1 0.24",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7931329",
   "tg_tokens": 128,
   "tps": 34.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-22",
   "flags": [],
   "hardware": "M3 Max 10+4",
   "id": "lc-disc-4167-c7931329-llama7bmostly-Q4_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 567.59,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M3 Max 10+4; llama 7B mostly Q4_0 | pp512 | 567.59 \u00b1 0.53; llama 7B mostly Q4_0 | tg128 | 56.85 \u00b1 0.37",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7931329",
   "tg_tokens": 128,
   "tps": 56.85,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-28",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "lc-disc-4167-c7965958-llama7bmostly-F16",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 373.97,
   "provenance": "community",
   "quant": "F16",
   "quote": "M2 Pro; llama 7B mostly F16 | pp512 | 373.97 \u00b1 0.15; llama 7B mostly F16 | tg128 | 12.97 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7965958",
   "tg_tokens": 128,
   "tps": 12.97,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-28",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "lc-disc-4167-c7965958-llama7bmostly-Q8_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 335.9,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M2 Pro; llama 7B mostly Q8_0 | pp512 | 335.90 \u00b1 0.09; llama 7B mostly Q8_0 | tg128 | 23.24 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7965958",
   "tg_tokens": 128,
   "tps": 23.24,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-28",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "lc-disc-4167-c7965958-llama7bmostly-Q4_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 342.61,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M2 Pro; llama 7B mostly Q4_0 | pp512 | 342.61 \u00b1 0.06; llama 7B mostly Q4_0 | tg128 | 39.71 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7965958",
   "tg_tokens": 128,
   "tps": 39.71,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-28",
   "flags": [],
   "hardware": "M3 Max",
   "id": "lc-disc-4167-c7968183-llama7ball-F32",
   "model": "llama 7B all",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 644.56,
   "provenance": "community",
   "quant": "F32",
   "quote": "M3 Max; llama 7B all F32 | pp512 | 644.56 \u00b1 1.70; llama 7B all F32 | tg128 | 13.43 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7968183",
   "tg_tokens": 128,
   "tps": 13.43,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-28",
   "flags": [],
   "hardware": "M3 Max",
   "id": "lc-disc-4167-c7968183-llama7bmostly-F16",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 784.83,
   "provenance": "community",
   "quant": "F16",
   "quote": "M3 Max; llama 7B mostly F16 | pp512 | 784.83 \u00b1 2.11; llama 7B mostly F16 | tg128 | 25.14 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7968183",
   "tg_tokens": 128,
   "tps": 25.14,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2023-12-28",
   "flags": [],
   "hardware": "M3 Max",
   "id": "lc-disc-4167-c7968183-llama7bmostly-Q8_0",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 707.07,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M3 Max; llama 7B mostly Q8_0 | pp512 | 707.07 \u00b1 27.03; llama 7B mostly Q8_0 | tg128 | 42.37 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7968183",
   "tg_tokens": 128,
   "tps": 42.37,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2024-01-02",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "lc-disc-4167-c7997936-llama7bmostly-F16",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 262.65,
   "provenance": "community",
   "quant": "F16",
   "quote": "M1 Pro; llama 7B mostly F16 | pp512 | 262.65 \u00b1 0.16; llama 7B mostly F16 | tg128 | 12.75 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7997936",
   "tg_tokens": 128,
   "tps": 12.75,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2024-01-02",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "lc-disc-4167-c7997936-llama7bmostly-Q8_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 237.0,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M1 Pro; llama 7B mostly Q8_0 | pp512 | 237.00 \u00b1 0.05; llama 7B mostly Q8_0 | tg128 | 22.51 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7997936",
   "tg_tokens": 128,
   "tps": 22.51,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2024-01-02",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "lc-disc-4167-c7997936-llama7bmostly-Q4_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 233.49,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M1 Pro; llama 7B mostly Q4_0 | pp512 | 233.49 \u00b1 0.04; llama 7B mostly Q4_0 | tg128 | 36.34 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-7997936",
   "tg_tokens": 128,
   "tps": 36.34,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2024-01-07",
   "flags": [],
   "hardware": "M2",
   "id": "lc-disc-4167-c8034638-llama7bmostly-Q8_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; ngl=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 147.27,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M2; llama 7B mostly Q8_0 | pp512 | 147.27 \u00b1 0.53; llama 7B mostly Q8_0 | tg128 | 12.18 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-8034638",
   "tg_tokens": 128,
   "tps": 12.18,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2024-01-07",
   "flags": [],
   "hardware": "M2",
   "id": "lc-disc-4167-c8034638-llama7bmostly-Q4_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; ngl=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 145.91,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M2; llama 7B mostly Q4_0 | pp512 | 145.91 \u00b1 0.05; llama 7B mostly Q4_0 | tg128 | 21.70 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-8034638",
   "tg_tokens": 128,
   "tps": 21.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2024-01-10",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "lc-disc-4167-c8086260-llama7bmostly-F16",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1158.95,
   "provenance": "community",
   "quant": "F16",
   "quote": "M1 Ultra; llama 7B mostly F16 | pp512 | 1158.95 \u00b1 1.67; llama 7B mostly F16 | tg128 | 36.31 \u00b1 0.18",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-8086260",
   "tg_tokens": 128,
   "tps": 36.31,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2024-01-11",
   "flags": [],
   "hardware": "M3 Max 10+4",
   "id": "lc-disc-4167-c8089048-llama7bmostly-F16",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1168.89,
   "provenance": "community",
   "quant": "F16",
   "quote": "M3 Max 10+4; llama 7B mostly F16 | pp512 | 1168.89 \u00b1 0.49; llama 7B mostly F16 | tg128 | 37.01 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-8089048",
   "tg_tokens": 128,
   "tps": 37.01,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2024-01-11",
   "flags": [],
   "hardware": "M3 Max 10+4",
   "id": "lc-disc-4167-c8089048-llama7bmostly-Q8_0",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1042.95,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M3 Max 10+4; llama 7B mostly Q8_0 | pp512 | 1042.95 \u00b1 1.01; llama 7B mostly Q8_0 | tg128 | 59.87 \u00b1 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-8089048",
   "tg_tokens": 128,
   "tps": 59.87,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2024-01-11",
   "flags": [],
   "hardware": "M3 Max 10+4",
   "id": "lc-disc-4167-c8089048-llama7bmostly-Q4_0",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1030.04,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M3 Max 10+4; llama 7B mostly Q4_0 | pp512 | 1030.04 \u00b1 0.72; llama 7B mostly Q4_0 | tg128 | 83.73 \u00b1 0.13",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-8089048",
   "tg_tokens": 128,
   "tps": 83.73,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2024-01-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max 10+4",
   "id": "lc-disc-4167-c8089048-llama7bmostly-Q8_0-2",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 566.4,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M3 Max 10+4; llama 7B mostly Q8_0 | pp512 | 566.40 \u00b1 3.32; llama 7B mostly Q8_0 | tg128 | 34.30 \u00b1 0.24",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-8089048",
   "tg_tokens": 128,
   "tps": 34.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2024-01-11",
   "flags": [],
   "hardware": "M3",
   "id": "lc-disc-4167-c8089048-llama7bmostly-F16-2",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 962.94,
   "provenance": "community",
   "quant": "F16",
   "quote": "M3; llama 7B mostly F16 | pp512 | 962.94 \u00b1 10.27; llama 7B mostly F16 | tg128 | 35.28 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-8089048",
   "tg_tokens": 128,
   "tps": 35.28,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2024-01-11",
   "flags": [],
   "hardware": "M3",
   "id": "lc-disc-4167-c8089048-llama7bmostly-Q8_0-3",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1049.96,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M3; llama 7B mostly Q8_0 | pp512 | 1049.96 \u00b1 26.63; llama 7B mostly Q8_0 | tg128 | 50.47 \u00b1 0.16",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-8089048",
   "tg_tokens": 128,
   "tps": 50.47,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2024-01-11",
   "flags": [],
   "hardware": "M3",
   "id": "lc-disc-4167-c8089048-llama7bmostly-Q4_0-2",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1087.59,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M3; llama 7B mostly Q4_0 | pp512 | 1087.59 \u00b1 32.54; llama 7B mostly Q4_0 | tg128 | 69.49 \u00b1 0.29",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-8089048",
   "tg_tokens": 128,
   "tps": 69.49,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2024-01-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3",
   "id": "lc-disc-4167-c8089048-llama7bmostly-F16-3",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 648.07,
   "provenance": "community",
   "quant": "F16",
   "quote": "M3; llama 7B mostly F16 | pp512 | 648.07 \u00b1 2.41; llama 7B mostly F16 | tg128 | 18.46 \u00b1 0.67",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-8089048",
   "tg_tokens": 128,
   "tps": 18.46,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2024-01-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3",
   "id": "lc-disc-4167-c8089048-llama7bmostly-Q8_0-4",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 637.51,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M3; llama 7B mostly Q8_0 | pp512 | 637.51 \u00b1 102.02; llama 7B mostly Q8_0 | tg128 | 19.11 \u00b1 0.53",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-8089048",
   "tg_tokens": 128,
   "tps": 19.11,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2024-01-11",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3",
   "id": "lc-disc-4167-c8089048-llama7bmostly-Q4_0-3",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 671.1,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M3; llama 7B mostly Q4_0 | pp512 | 671.10 \u00b1 94.51; llama 7B mostly Q4_0 | tg128 | 19.23 \u00b1 1.21",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-8089048",
   "tg_tokens": 128,
   "tps": 19.23,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2024-01-21",
   "flags": [],
   "hardware": "M3",
   "id": "lc-disc-4167-c8196421-llama7bmostly-F16",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 925.62,
   "provenance": "community",
   "quant": "F16",
   "quote": "M3; llama 7B mostly F16 | pp512 | 925.62 \u00b1 23.50; llama 7B mostly F16 | tg128 | 35.42 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-8196421",
   "tg_tokens": 128,
   "tps": 35.42,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2024-01-21",
   "flags": [],
   "hardware": "M3",
   "id": "lc-disc-4167-c8196421-llama7bmostly-Q8_0",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1937.77,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M3; llama 7B mostly Q8_0 | pp512 | 1937.77  \u00b1 109.23; llama 7B mostly Q8_0 | tg128 | 50.75 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-8196421",
   "tg_tokens": 128,
   "tps": 50.75,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2024-01-21",
   "flags": [],
   "hardware": "M3",
   "id": "lc-disc-4167-c8196421-llama7bmostly-Q4_0",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1956.45,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M3; llama 7B mostly Q4_0 | pp512 | 1956.45  \u00b1 134.41; llama 7B mostly Q4_0 | tg128 | 69.83 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-8196421",
   "tg_tokens": 128,
   "tps": 69.83,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2024-01-21",
   "flags": [],
   "hardware": "M3",
   "id": "lc-disc-4167-c8196421-llama7bmostly-F16-2",
   "model": "llama 7B mostly",
   "notes": "threads=99; ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3022.37,
   "provenance": "community",
   "quant": "F16",
   "quote": "M3; llama 7B mostly F16 | pp512 | 3022.37  \u00b1 94.50; llama 7B mostly F16 | tg128 | 23.93 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-8196421",
   "tg_tokens": 128,
   "tps": 23.93,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2024-01-21",
   "flags": [],
   "hardware": "M3",
   "id": "lc-disc-4167-c8196421-llama7bmostly-Q8_0-2",
   "model": "llama 7B mostly",
   "notes": "threads=99; ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1688.37,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M3; llama 7B mostly Q8_0 | pp512 | 1688.37 \u00b1 51.25; llama 7B mostly Q8_0 | tg128 | 42.43 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-8196421",
   "tg_tokens": 128,
   "tps": 42.43,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2024-01-21",
   "flags": [],
   "hardware": "M3",
   "id": "lc-disc-4167-c8196421-llama7bmostly-Q4_0-2",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1816.45,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M3; llama 7B mostly Q4_0 | pp512 | 1816.45 \u00b1 68.24; llama 7B mostly Q4_0 | tg128 | 73.38 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-8196421",
   "tg_tokens": 128,
   "tps": 73.38,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2024-01-21",
   "flags": [],
   "hardware": "M3",
   "id": "lc-disc-4167-c8196421-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 310.95,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M3; llama 7B Q4_0 | pp512 | 310.95 \u00b1 2.10; llama 7B Q4_0 | tg128 | 21.47 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-8196421",
   "tg_tokens": 128,
   "tps": 21.47,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2024-05-20",
   "flags": [],
   "hardware": "M4",
   "id": "lc-disc-4167-c9491068-llama1b-F16",
   "model": "llama 1B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1375.87,
   "provenance": "community",
   "quant": "F16",
   "quote": "M4; llama 1B F16 | pp512 | 1375.87 \u00b1 2.49; llama 1B F16 | tg128 | 46.21 \u00b1 0.22",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-9491068",
   "tg_tokens": 128,
   "tps": 46.21,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2024-05-20",
   "flags": [],
   "hardware": "M4",
   "id": "lc-disc-4167-c9491068-llama1b-Q8_0",
   "model": "llama 1B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1333.82,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M4; llama 1B Q8_0 | pp512 | 1333.82 \u00b1 1.85; llama 1B Q8_0 | tg128 | 75.33 \u00b1 0.19",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-9491068",
   "tg_tokens": 128,
   "tps": 75.33,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2024-05-20",
   "flags": [],
   "hardware": "M4",
   "id": "lc-disc-4167-c9491068-llama1b-Q4_0",
   "model": "llama 1B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1334.54,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M4; llama 1B Q4_0 | pp512 | 1334.54 \u00b1 6.87; llama 1B Q4_0 | tg128 | 121.20 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-9491068",
   "tg_tokens": 128,
   "tps": 121.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2024-05-20",
   "flags": [],
   "hardware": "M4",
   "id": "lc-disc-4167-c9491068-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 212.13,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M4; llama 7B Q4_0 | pp512 | 212.13 \u00b1 0.11; llama 7B Q4_0 | tg128 | 23.29 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-9491068",
   "tg_tokens": 128,
   "tps": 23.29,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2024-11-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "lc-disc-4167-c11185614-llama7bmostly-F16",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 464.48,
   "provenance": "community",
   "quant": "F16",
   "quote": "M4 Pro; llama 7B mostly F16 | pp512 | 464.48 \u00b1 2.62; llama 7B mostly F16 | tg128 | 17.18 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-11185614",
   "tg_tokens": 128,
   "tps": 17.18,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2024-11-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "lc-disc-4167-c11185614-llama7bmostly-Q8_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 449.62,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M4 Pro; llama 7B mostly Q8_0 | pp512 | 449.62 \u00b1 0.12; llama 7B mostly Q8_0 | tg128 | 30.69 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-11185614",
   "tg_tokens": 128,
   "tps": 30.69,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2024-11-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "lc-disc-4167-c11185614-llama7bmostly-Q4_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 439.78,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M4 Pro; llama 7B mostly Q4_0 | pp512 | 439.78 \u00b1 6.51; llama 7B mostly Q4_0 | tg128 | 50.74 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-11185614",
   "tg_tokens": 128,
   "tps": 50.74,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2024-11-10",
   "flags": [],
   "hardware": "M4",
   "id": "lc-disc-4167-c11202046-llama7bmostly-F16",
   "model": "llama 7B mostly",
   "notes": "threads=99; ngl=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 230.18,
   "provenance": "community",
   "quant": "F16",
   "quote": "M4; llama 7B mostly F16 | pp512 | 230.18 \u00b1 0.16; llama 7B mostly F16 | tg128 | 7.43 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-11202046",
   "tg_tokens": 128,
   "tps": 7.43,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2024-11-10",
   "flags": [],
   "hardware": "M4",
   "id": "lc-disc-4167-c11202046-llama7bmostly-Q8_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; ngl=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 223.64,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M4; llama 7B mostly Q8_0 | pp512 | 223.64 \u00b1 0.05; llama 7B mostly Q8_0 | tg128 | 13.54 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-11202046",
   "tg_tokens": 128,
   "tps": 13.54,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2024-11-10",
   "flags": [],
   "hardware": "M4",
   "id": "lc-disc-4167-c11202046-llama7bmostly-Q4_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; ngl=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 221.29,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M4; llama 7B mostly Q4_0 | pp512 | 221.29 \u00b1 0.04; llama 7B mostly Q4_0 | tg128 | 24.11 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-11202046",
   "tg_tokens": 128,
   "tps": 24.11,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2024-12-04",
   "flags": [],
   "hardware": "M4 Max",
   "id": "lc-disc-4167-c11462237-llama7bmostly-F16",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 922.83,
   "provenance": "community",
   "quant": "F16",
   "quote": "M4 Max; llama 7B mostly F16 | pp512 | 922.83 \u00b1 1.12; llama 7B mostly F16 | tg128 | 31.64 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-11462237",
   "tg_tokens": 128,
   "tps": 31.64,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2024-12-04",
   "flags": [],
   "hardware": "M4 Max",
   "id": "lc-disc-4167-c11462237-llama7bmostly-Q8_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 891.94,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M4 Max; llama 7B mostly Q8_0 | pp512 | 891.94 \u00b1 0.28; llama 7B mostly Q8_0 | tg128 | 54.05 \u00b1 0.16",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-11462237",
   "tg_tokens": 128,
   "tps": 54.05,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2024-12-04",
   "flags": [],
   "hardware": "M4 Max",
   "id": "lc-disc-4167-c11462237-llama7bmostly-Q4_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 885.68,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M4 Max; llama 7B mostly Q4_0 | pp512 | 885.68 \u00b1 1.11; llama 7B mostly Q4_0 | tg128 | 83.06 \u00b1 0.22",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-11462237",
   "tg_tokens": 128,
   "tps": 83.06,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2024-12-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "lc-disc-4167-c11493483-llama7bmostly-F16",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 381.14,
   "provenance": "community",
   "quant": "F16",
   "quote": "M4 Pro; llama 7B mostly F16 | pp512 | 381.14 \u00b1 0.06; llama 7B mostly F16 | tg128 | 17.19 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-11493483",
   "tg_tokens": 128,
   "tps": 17.19,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2024-12-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "lc-disc-4167-c11493483-llama7bmostly-Q8_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 367.13,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M4 Pro; llama 7B mostly Q8_0 | pp512 | 367.13 \u00b1 0.06; llama 7B mostly Q8_0 | tg128 | 30.54 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-11493483",
   "tg_tokens": 128,
   "tps": 30.54,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2024-12-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "lc-disc-4167-c11493483-llama7bmostly-Q4_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 364.06,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M4 Pro; llama 7B mostly Q4_0 | pp512 | 364.06 \u00b1 0.11; llama 7B mostly Q4_0 | tg128 | 49.64 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-11493483",
   "tg_tokens": 128,
   "tps": 49.64,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2024-12-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "lc-disc-4167-c11595505-llama7bmostly-F16",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 923.55,
   "provenance": "community",
   "quant": "F16",
   "quote": "M4 Max; llama 7B mostly F16 | pp512 | 923.55 \u00b1 0.12; llama 7B mostly F16 | tg128 | 31.61 \u00b1 0.10",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-11595505",
   "tg_tokens": 128,
   "tps": 31.61,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2024-12-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "lc-disc-4167-c11595505-llama7bmostly-Q8_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 852.47,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M4 Max; llama 7B mostly Q8_0 | pp512 | 852.47 \u00b1 48.37; llama 7B mostly Q8_0 | tg128 | 53.06 \u00b1 0.48",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-11595505",
   "tg_tokens": 128,
   "tps": 53.06,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2024-12-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "lc-disc-4167-c11595505-llama7bmostly-Q4_0",
   "model": "llama 7B mostly",
   "notes": "threads=99; config=b01550",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 746.09,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M4 Max; llama 7B mostly Q4_0 | pp512 | 746.09 \u00b1 29.30; llama 7B mostly Q4_0 | tg128 | 82.52 \u00b1 0.13",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-11595505",
   "tg_tokens": 128,
   "tps": 82.52,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2025-02-19",
   "flags": [],
   "hardware": "M4 Max",
   "id": "lc-disc-4167-c12244870-llama7bmostly-F16",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 920.48,
   "provenance": "community",
   "quant": "F16",
   "quote": "M4 Max; llama 7B mostly F16 | pp512 | 920.48 \u00b1 3.25; llama 7B mostly F16 | tg128 | 31.56 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-12244870",
   "tg_tokens": 128,
   "tps": 31.56,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2025-02-19",
   "flags": [],
   "hardware": "M4 Max",
   "id": "lc-disc-4167-c12244870-llama7bmostly-Q8_0",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 891.32,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M4 Max; llama 7B mostly Q8_0 | pp512 | 891.32 \u00b1 0.95; llama 7B mostly Q8_0 | tg128 | 53.75 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-12244870",
   "tg_tokens": 128,
   "tps": 53.75,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2025-02-19",
   "flags": [],
   "hardware": "M4 Max",
   "id": "lc-disc-4167-c12244870-llama7bmostly-Q4_0",
   "model": "llama 7B mostly",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 884.59,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M4 Max; llama 7B mostly Q4_0 | pp512 | 884.59 \u00b1 0.78; llama 7B mostly Q4_0 | tg128 | 82.36 \u00b1 0.22",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-12244870",
   "tg_tokens": 128,
   "tps": 82.36,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": null,
   "ctx": null,
   "date": "2024-06-21",
   "flags": [],
   "hardware": "M2 Max",
   "id": "lc-disc-4167-c9839020-llama7b-F16",
   "model": "llama 7B",
   "notes": "threads=8",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 26.18,
   "provenance": "community",
   "quant": "F16",
   "quote": "M2 Max; llama 7B F16 | pp512 | 26.18 \u00b1 0.06; llama 7B F16 | tg128 | 8.19 \u00b1 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-9839020",
   "tg_tokens": 128,
   "tps": 8.19,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": null,
   "ctx": null,
   "date": "2024-06-21",
   "flags": [],
   "hardware": "M2 Max",
   "id": "lc-disc-4167-c9839020-llama7b-Q8_0",
   "model": "llama 7B",
   "notes": "threads=8",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 52.83,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M2 Max; llama 7B Q8_0 | pp512 | 52.83 \u00b1 0.36; llama 7B Q8_0 | tg128 | 13.79 \u00b1 0.49",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-9839020",
   "tg_tokens": 128,
   "tps": 13.79,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": null,
   "ctx": null,
   "date": "2024-06-21",
   "flags": [],
   "hardware": "M2 Max",
   "id": "lc-disc-4167-c9839020-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=8",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 49.04,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M2 Max; llama 7B Q4_0 | pp512 | 49.04 \u00b1 0.06; llama 7B Q4_0 | tg128 | 22.86 \u00b1 0.16",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-9839020",
   "tg_tokens": 128,
   "tps": 22.86,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2024-06-21",
   "flags": [],
   "hardware": "M2 Max",
   "id": "lc-disc-4167-c9839020-llama7b-F16-2",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 80.4,
   "provenance": "community",
   "quant": "F16",
   "quote": "M2 Max; llama 7B F16 | pp512 | 80.40 \u00b1 0.43; llama 7B F16 | tg128 | 2.02 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-9839020",
   "tg_tokens": 128,
   "tps": 2.02,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2024-06-21",
   "flags": [],
   "hardware": "M2 Max",
   "id": "lc-disc-4167-c9839020-llama7b-Q8_0-2",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 79.79,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "M2 Max; llama 7B Q8_0 | pp512 | 79.79 \u00b1 0.07; llama 7B Q8_0 | tg128 | 19.79 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-9839020",
   "tg_tokens": 128,
   "tps": 19.79,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2024-06-21",
   "flags": [],
   "hardware": "M2 Max",
   "id": "lc-disc-4167-c9839020-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 79.94,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M2 Max; llama 7B Q4_0 | pp512 | 79.94 \u00b1 0.13; llama 7B Q4_0 | tg128 | 24.07 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-9839020",
   "tg_tokens": 128,
   "tps": 24.07,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-22",
   "flags": [],
   "hardware": "Radeon RX 470 G",
   "id": "lc-disc-10879-c14193645-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=8; fa=0 ngl=99 sm=none; config=b06249",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 155.22,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 470 G; llama 7B Q4_0 | pp512 | 155.22 \u00b1 0.17; llama 7B Q4_0 | tg128 | 29.52 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14193645",
   "tg_tokens": 128,
   "tps": 29.52,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-22",
   "flags": [],
   "hardware": "Radeon RX 470 G",
   "id": "lc-disc-10879-c14193645-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=8; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99 sm=none; config=b06249",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 140.52,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 470 G; llama 7B Q4_0 | pp512 | 140.52 \u00b1 0.34; llama 7B Q4_0 | tg128 | 29.28 \u00b1 0.14",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14193645",
   "tg_tokens": 128,
   "tps": 29.28,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-22",
   "flags": [],
   "hardware": "Radeon RX 470 G",
   "id": "lc-disc-10879-c14193645-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=8; config=b06249",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 161.47,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 470 G; llama 7B Q4_0 | pp512 | 161.47 \u00b1 0.43; llama 7B Q4_0 | tg128 | 33.45 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14193645",
   "tg_tokens": 128,
   "tps": 33.45,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-22",
   "flags": [],
   "hardware": "Radeon RX 470 G",
   "id": "lc-disc-10879-c14193654-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=8; config=b06249",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 185.73,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 470 G; llama 7B Q4_0 | pp512 | 185.73 \u00b1 0.69; llama 7B Q4_0 | tg128 | 34.89 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14193654",
   "tg_tokens": 128,
   "tps": 34.89,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-22",
   "flags": [],
   "hardware": "Radeon RX 470 G",
   "id": "lc-disc-10879-c14193654-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=8; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06249",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 179.01,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 470 G; llama 7B Q4_0 | pp512 | 179.01 \u00b1 0.65; llama 7B Q4_0 | tg128 | 34.65 \u00b1 0.17",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14193654",
   "tg_tokens": 128,
   "tps": 34.65,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-23",
   "flags": [],
   "hardware": "Radeon RX 470 G",
   "id": "lc-disc-10879-c14195584-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b06242",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 218.07,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 470 G; llama 7B Q4_0 | pp512 | 218.07 \u00b1 0.56; llama 7B Q4_0 | tg128 | 38.63 \u00b1 0.21",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14195584",
   "tg_tokens": 128,
   "tps": 38.63,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-23",
   "flags": [],
   "hardware": "Radeon RX 470 G",
   "id": "lc-disc-10879-c14195584-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06242",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 197.94,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 470 G; llama 7B Q4_0 | pp512 | 197.94 \u00b1 2.94; llama 7B Q4_0 | tg128 | 35.14 \u00b1 1.93",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14195584",
   "tg_tokens": 128,
   "tps": 35.14,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-02-25",
   "flags": [],
   "hardware": "Radeon RX 470 G",
   "id": "lc-disc-10879-c15925877-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100; config=b08156",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 197.19,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 470 G; llama 7B Q4_0 | pp512 | 197.19 \u00b1 0.63; llama 7B Q4_0 | tg128 | 34.42 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15925877",
   "tg_tokens": 128,
   "tps": 34.42,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-02-25",
   "flags": [],
   "hardware": "Radeon RX 470 G",
   "id": "lc-disc-10879-c15925877-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b08156",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 197.26,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 470 G; llama 7B Q4_0 | pp512 | 197.26 \u00b1 0.27; llama 7B Q4_0 | tg128 | 37.28 \u00b1 0.11",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15925877",
   "tg_tokens": 128,
   "tps": 37.28,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-07-09",
   "flags": [],
   "hardware": "Radeon RX 470 G",
   "id": "lc-disc-10879-c17578819-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=-1; config=b09934",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 207.16,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 470 G; llama 7B Q4_0 | pp512 | 207.16 \u00b1 0.27; llama 7B Q4_0 | tg128 | 40.11 \u00b1 0.11",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-17578819",
   "tg_tokens": 128,
   "tps": 40.11,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-07-09",
   "flags": [],
   "hardware": "Radeon RX 470 G",
   "id": "lc-disc-10879-c17578819-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b09934",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 211.25,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 470 G; llama 7B Q4_0 | pp512 | 211.25 \u00b1 0.88; llama 7B Q4_0 | tg128 | 43.35 \u00b1 0.22",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-17578819",
   "tg_tokens": 128,
   "tps": 43.35,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2024-12-18",
   "flags": [],
   "hardware": "RTX 3080",
   "id": "lc-disc-10879-c11600977-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; ngl=99; config=b04351",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1706.07,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3080; llama 7B Q4_0 | pp512 | 1706.07 \u00b1 139.33; llama 7B Q4_0 | tg128 | 62.16 \u00b1 1.98",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-11600977",
   "tg_tokens": 128,
   "tps": 62.16,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2024-12-18",
   "flags": [],
   "hardware": "RTX 3080",
   "id": "lc-disc-10879-c11600977-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; config=b04351",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4499.47,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3080; llama 7B Q4_0 | pp512 | 4499.47 \u00b1 60.66; llama 7B Q4_0 | tg128 | 131.01 \u00b1 0.43",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-11600977",
   "tg_tokens": 128,
   "tps": 131.01,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-01-08",
   "flags": [],
   "hardware": "M2",
   "id": "lc-disc-10879-c11768021-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; ngl=100",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 38.67,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M2; llama 7B Q4_0 | pp512 | 38.67 \u00b1 0.03; llama 7B Q4_0 | tg128 | 11.07 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-11768021",
   "tg_tokens": 128,
   "tps": 11.07,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-01-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c11771840-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; ngl=100",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1545.39,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 1545.39 \u00b1 6.58; llama 7B Q4_0 | tg128 | 88.12 \u00b1 1.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-11771840",
   "tg_tokens": 128,
   "tps": 88.12,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-01-08",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c11772341-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; ngl=100",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2022.59,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2022.59 \u00b1 10.08; llama 7B Q4_0 | tg128 | 136.24 \u00b1 0.30",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-11772341",
   "tg_tokens": 128,
   "tps": 136.24,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-01-08",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c11772341-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; ngl=100",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2039.24,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2039.24 \u00b1 18.08; llama 7B Q4_0 | tg128 | 140.68 \u00b1 2.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-11772341",
   "tg_tokens": 128,
   "tps": 140.68,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-01-08",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c11772341-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100; ngl=100",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2062.17,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2062.17 \u00b1 5.36; llama 7B Q4_0 | tg128 | 143.99 \u00b1 0.23",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-11772341",
   "tg_tokens": 128,
   "tps": 143.99,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-01-08",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c11772341-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=100; ngl=100",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1997.04,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 1997.04 \u00b1 5.78; llama 7B Q4_0 | tg128 | 136.98 \u00b1 1.73",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-11772341",
   "tg_tokens": 128,
   "tps": 136.98,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-01-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c11772341-llama7b-Q4_0-5",
   "model": "llama 7B",
   "notes": "threads=100; ngl=100",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1668.19,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 1668.19 \u00b1 12.78; llama 7B Q4_0 | tg128 | 100.62 \u00b1 0.66",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-11772341",
   "tg_tokens": 128,
   "tps": 100.62,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-01-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c11772341-llama7b-Q4_0-6",
   "model": "llama 7B",
   "notes": "threads=100; ngl=100",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1566.38,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 1566.38 \u00b1 8.01; llama 7B Q4_0 | tg128 | 97.96 \u00b1 1.13",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-11772341",
   "tg_tokens": 128,
   "tps": 97.96,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-01-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c11772341-llama7b-Q4_0-7",
   "model": "llama 7B",
   "notes": "threads=100; ngl=100",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1484.04,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 1484.04 \u00b1 6.01; llama 7B Q4_0 | tg128 | 91.48 \u00b1 0.63",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-11772341",
   "tg_tokens": 128,
   "tps": 91.48,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-01-08",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc-disc-10879-c11772082-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; ngl=100; config=b04439",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3301.47,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3090; llama 7B Q4_0 | pp512 | 3301.47 \u00b1 33.76; llama 7B Q4_0 | tg128 | 123.72 \u00b1 0.14",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-11772082",
   "tg_tokens": 128,
   "tps": 123.72,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-01-08",
   "flags": [],
   "hardware": "Radeon RX 6800 XT",
   "id": "lc-disc-10879-c11772082-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; config=b04439",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 863.03,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6800 XT; llama 7B Q4_0 | pp512 | 863.03 \u00b1 0.70; llama 7B Q4_0 | tg128 | 91.59 \u00b1 0.40",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-11772082",
   "tg_tokens": 128,
   "tps": 91.59,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-01-08",
   "flags": [],
   "hardware": "Radeon RX 6800 XT",
   "id": "lc-disc-10879-c11772082-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100; ngl=100; config=b04439",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 312.02,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6800 XT; llama 7B Q4_0 | pp512 | 312.02 \u00b1 0.97; llama 7B Q4_0 | tg128 | 70.17 \u00b1 0.25",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-11772082",
   "tg_tokens": 128,
   "tps": 70.17,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-01-08",
   "flags": [],
   "hardware": "Arc A770",
   "id": "lc-disc-10879-c11772082-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=100; config=b04439",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 95.52,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A770; llama 7B Q4_0 | pp512 | 95.52 \u00b1 0.12; llama 7B Q4_0 | tg128 | 44.49 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-11772082",
   "tg_tokens": 128,
   "tps": 44.49,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-01-08",
   "flags": [],
   "hardware": "Arc A750",
   "id": "lc-disc-10879-c11778792-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 88.86,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A750; llama 7B Q4_0 | pp512 | 88.86 \u00b1 0.14; llama 7B Q4_0 | tg128 | 27.57 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-11778792",
   "tg_tokens": 128,
   "tps": 27.57,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2025-01-08",
   "flags": [],
   "hardware": "Arc A750",
   "id": "lc-disc-10879-c11778792-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1616.11,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A750; llama 7B Q4_0 | pp512 | 1616.11 \u00b1 5.28; llama 7B Q4_0 | tg128 | 36.64 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-11778792",
   "tg_tokens": 128,
   "tps": 36.64,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-01-09",
   "flags": [],
   "hardware": "Radeon RX 470 G",
   "id": "lc-disc-10879-c11791379-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=8; main_gpu=1 ngl=100; config=b04452",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 147.84,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 470 G; llama 7B Q4_0 | pp512 | 147.84 \u00b1 0.38; llama 7B Q4_0 | tg128 | 30.77 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-11791379",
   "tg_tokens": 128,
   "tps": 30.77,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-01-10",
   "flags": [],
   "hardware": "A770",
   "id": "lc-disc-10879-c11802247-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b04460",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 314.24,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A770; llama 7B Q4_0 | pp512 | 314.24 \u00b1 1.04; llama 7B Q4_0 | tg128 | 45.22 \u00b1 0.25",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-11802247",
   "tg_tokens": 128,
   "tps": 45.22,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-01-12",
   "flags": [],
   "hardware": "Radeon RX 5700 XT",
   "id": "lc-disc-10879-c11809587-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b04462",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 439.42,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 5700 XT; llama 7B Q4_0 | pp512 | 439.42 \u00b1 0.28; llama 7B Q4_0 | tg128 | 70.13 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-11809587",
   "tg_tokens": 128,
   "tps": 70.13,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-01-12",
   "flags": [],
   "hardware": "Radeon RX 5700 XT",
   "id": "lc-disc-10879-c11809587-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; config=b04462",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 354.17,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 5700 XT; llama 7B Q4_0 | pp512 | 354.17 \u00b1 0.18; llama 7B Q4_0 | tg128 | 67.55 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-11809587",
   "tg_tokens": 128,
   "tps": 67.55,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-01-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon RX 5700 XT",
   "id": "lc-disc-10879-c11809587-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100; config=b04462",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 214.48,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 5700 XT; llama 7B Q4_0 | pp512 | 214.48 \u00b1 2.31; llama 7B Q4_0 | tg128 | 23.21 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-11809587",
   "tg_tokens": 128,
   "tps": 23.21,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-01-12",
   "flags": [],
   "hardware": "Radeon RX 5700 XT",
   "id": "lc-disc-10879-c11809587-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=100; config=b04462",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 314.17,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 5700 XT; llama 7B Q4_0 | pp512 | 314.17 \u00b1 0.29; llama 7B Q4_0 | tg128 | 62.02 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-11809587",
   "tg_tokens": 128,
   "tps": 62.02,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-07-20",
   "flags": [],
   "hardware": "Radeon RX 6750 XT",
   "id": "lc-disc-10879-c17704220-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100 sm=none; config=b10073",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 525.04,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6750 XT; llama 7B Q4_0 | pp512 | 525.04 \u00b1 8.94; llama 7B Q4_0 | tg128 | 61.08 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-17704220",
   "tg_tokens": 128,
   "tps": 61.08,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-07-20",
   "flags": [],
   "hardware": "Radeon RX 6750 XT",
   "id": "lc-disc-10879-c17704220-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10073",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 920.41,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6750 XT; llama 7B Q4_0 | pp512 | 920.41 \u00b1 5.34; llama 7B Q4_0 | tg128 | 78.19 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-17704220",
   "tg_tokens": 128,
   "tps": 78.19,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-07-30",
   "flags": [],
   "hardware": "Radeon RX 7600",
   "id": "lc-disc-10879-c17845940-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b10156",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1167.2,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7600; llama 7B Q4_0 | pp512 | 1167.20 \u00b1 20.39; llama 7B Q4_0 | tg128 | 58.03 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-17845940",
   "tg_tokens": 128,
   "tps": 58.03,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-07-30",
   "flags": [],
   "hardware": "Radeon RX 7600",
   "id": "lc-disc-10879-c17845940-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10156",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1353.47,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7600; llama 7B Q4_0 | pp512 | 1353.47 \u00b1 26.13; llama 7B Q4_0 | tg128 | 60.88 \u00b1 0.25",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-17845940",
   "tg_tokens": 128,
   "tps": 60.88,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-07-30",
   "flags": [],
   "hardware": "Radeon RX 7600",
   "id": "lc-disc-10879-c17845940-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100; config=b10156",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1076.93,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7600; llama 7B Q4_0 | pp512 | 1076.93 \u00b1 43.17; llama 7B Q4_0 | tg128 | 53.48 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-17845940",
   "tg_tokens": 128,
   "tps": 53.48,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-07-30",
   "flags": [],
   "hardware": "Radeon RX 7600",
   "id": "lc-disc-10879-c17845940-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10156",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1312.19,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7600; llama 7B Q4_0 | pp512 | 1312.19 \u00b1 12.80; llama 7B Q4_0 | tg128 | 58.00 \u00b1 0.18",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-17845940",
   "tg_tokens": 128,
   "tps": 58.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-07",
   "flags": [],
   "hardware": "Radeon 840M",
   "id": "lc-disc-10879-c17937054-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=6; config=b10310",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 60.33,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 840M; llama 7B Q4_0 | pp512 | 60.33 \u00b1 0.24; llama 7B Q4_0 | tg128 | 12.95 \u00b1 0.11",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-17937054",
   "tg_tokens": 128,
   "tps": 12.95,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-07",
   "flags": [],
   "hardware": "Radeon 840M",
   "id": "lc-disc-10879-c17937054-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=6; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10310",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 61.42,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 840M; llama 7B Q4_0 | pp512 | 61.42 \u00b1 0.17; llama 7B Q4_0 | tg128 | 13.18 \u00b1 0.13",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-17937054",
   "tg_tokens": 128,
   "tps": 13.18,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-07",
   "flags": [],
   "hardware": "Radeon 840M",
   "id": "lc-disc-10879-c17937054-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100; config=b10310",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 134.09,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 840M; llama 7B Q4_0 | pp512 | 134.09 \u00b1 6.19; llama 7B Q4_0 | tg128 | 14.33 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-17937054",
   "tg_tokens": 128,
   "tps": 14.33,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-07",
   "flags": [],
   "hardware": "Radeon 840M",
   "id": "lc-disc-10879-c17937054-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10310",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 172.58,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 840M; llama 7B Q4_0 | pp512 | 172.58 \u00b1 5.49; llama 7B Q4_0 | tg128 | 16.12 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-17937054",
   "tg_tokens": 128,
   "tps": 16.12,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-07",
   "flags": [],
   "hardware": "Radeon 840M",
   "id": "lc-disc-10879-c17937054-llama7b-Q4_0-5",
   "model": "llama 7B",
   "notes": "threads=100; config=b10310",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 131.33,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 840M; llama 7B Q4_0 | pp512 | 131.33 \u00b1 1.20; llama 7B Q4_0 | tg128 | 17.66 \u00b1 0.15",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-17937054",
   "tg_tokens": 128,
   "tps": 17.66,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-07",
   "flags": [],
   "hardware": "Radeon 840M",
   "id": "lc-disc-10879-c17937054-llama7b-Q4_0-6",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10310",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 150.12,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 840M; llama 7B Q4_0 | pp512 | 150.12 \u00b1 1.11; llama 7B Q4_0 | tg128 | 19.03 \u00b1 0.23",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-17937054",
   "tg_tokens": 128,
   "tps": 19.03,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-14",
   "flags": [],
   "hardware": "A100",
   "id": "lc-disc-10879-c18017987-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b10430",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3222.01,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A100; llama 7B Q4_0 | pp512 | 3222.01 \u00b1 173.55; llama 7B Q4_0 | pp1024 | 3208.43 \u00b1 1.07; llama 7B Q4_0 | pp2048 | 2988.57 \u00b1 1.15; llama 7B Q4_0 | pp4096 | 2706.29 \u00b1 0.24; llama 7B Q4_0 | pp8192 | 2266.24 \u00b1 0.86; llama 7B Q4_0 | pp16384 | 1626.71 \u00b1 0.45; llama 7B Q4_0 | pp32768 | 868.27 \u00b1 16.06; llama 7B Q4_0 | tg128 | 102.46 \u00b1 1.46; llama 7B Q4_0 | tg256 | 101.76 \u00b1 0.17; llama 7B Q4_0 | tg512 | 96.93 \u00b1 0.17; llama 7B Q4_0 | tg1024 | 92.33 \u00b1 0.19; llama 7B Q4_0 | tg2048 | 86.28 \u00b1 0.39",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18017987",
   "tg_tokens": 128,
   "tps": 102.46,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-14",
   "flags": [],
   "hardware": "A100",
   "id": "lc-disc-10879-c18017987-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b10430",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4040.77,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A100; llama 7B Q4_0 | pp512 | 4040.77 \u00b1 16.50; llama 7B Q4_0 | pp1024 | 3979.22 \u00b1 48.47; llama 7B Q4_0 | pp2048 | 3840.83 \u00b1 11.61; llama 7B Q4_0 | pp4096 | 3659.82 \u00b1 12.29; llama 7B Q4_0 | pp8192 | 3398.63 \u00b1 3.48; llama 7B Q4_0 | pp16384 | 2955.32 \u00b1 6.90; llama 7B Q4_0 | pp32768 | 2353.98 \u00b1 2.81; llama 7B Q4_0 | tg128 | 108.39 \u00b1 0.89; llama 7B Q4_0 | tg256 | 106.71 \u00b1 0.84; llama 7B Q4_0 | tg512 | 100.83 \u00b1 0.10; llama 7B Q4_0 | tg1024 | 97.85 \u00b1 0.54; llama 7B Q4_0 | tg2048 | 91.99 \u00b1 0.59",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18017987",
   "tg_tokens": 128,
   "tps": 108.39,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-23",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c18122886-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3692.87,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; dev=VULKAN1; llama 7B Q4_0 | pp512 | 3692.87 \u00b1 84.82; llama 7B Q4_0 | tg128 | 170.69 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18122886",
   "tg_tokens": 128,
   "tps": 170.69,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-23",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c18122886-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3932.57,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; dev=VULKAN1; llama 7B Q4_0 | pp512 | 3932.57 \u00b1 33.28; llama 7B Q4_0 | tg128 | 179.66 \u00b1 0.10",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18122886",
   "tg_tokens": 128,
   "tps": 179.66,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-23",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c18122886-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; fa=layer ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3252.09,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3252.09 \u00b1 14.34; llama 7B Q4_0 | tg128 | 149.04 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18122886",
   "tg_tokens": 128,
   "tps": 149.04,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-23",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c18122886-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=layer ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3422.83,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3422.83 \u00b1 10.35; llama 7B Q4_0 | tg128 | 157.41 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18122886",
   "tg_tokens": 128,
   "tps": 157.41,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-23",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c18122886-llama7b-Q4_0-5",
   "model": "llama 7B",
   "notes": "threads=99; fa=tensor ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1727.1,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 1727.10 \u00b1 1.78; llama 7B Q4_0 | tg128 | 56.35 \u00b1 0.62",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18122886",
   "tg_tokens": 128,
   "tps": 56.35,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-23",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c18122886-gemma431b-Q8_0",
   "model": "Gemma4 31B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 851.56,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Radeon RX 7900 XTX; gemma4 31B Q8_0 | pp512 | 851.56 \u00b1 1.38; gemma4 31B Q8_0 | tg128 | 24.42 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18122886",
   "tg_tokens": 128,
   "tps": 24.42,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-23",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c18122886-gemma431b-Q8_0-2",
   "model": "Gemma4 31B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 609.25,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Radeon RX 7900 XTX; gemma4 31B Q8_0 | pp512 | 609.25 \u00b1 0.25; gemma4 31B Q8_0 | tg128 | 21.03 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18122886",
   "tg_tokens": 128,
   "tps": 21.03,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-10",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c18386832-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b10868",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3726.99,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3726.99 \u00b1 22.09; llama 7B Q4_0 | tg128 | 182.63 \u00b1 0.22",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18386832",
   "tg_tokens": 128,
   "tps": 182.63,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-10",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c18386832-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10868",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3889.54,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3889.54 \u00b1 16.52; llama 7B Q4_0 | tg128 | 190.92 \u00b1 0.19",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18386832",
   "tg_tokens": 128,
   "tps": 190.92,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-17",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-disc-10879-c18485429-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 5546.39,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 5090; dev=VULKAN0; llama 7B Q4_0 | pp512 | 5546.39 \u00b1 17.00; llama 7B Q4_0 | tg128 | 149.41 \u00b1 2.13",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18485429",
   "tg_tokens": 128,
   "tps": 149.41,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-17",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-disc-10879-c18485429-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; 160.0 W measured GPU power, peaked at ~160 W during the run (approximate); fa=1 ngl=99",
   "params": null,
   "power_w": 160.0,
   "pp_tokens": 512,
   "pp_tps": 6011.18,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 5090; dev=VULKAN0; llama 7B Q4_0 | pp512 | 6011.18 \u00b1 20.90; llama 7B Q4_0 | tg128 | 154.31 \u00b1 1.61",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18485429",
   "tg_tokens": 128,
   "tps": 154.31,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-01-13",
   "flags": [],
   "hardware": "RTX 4050",
   "id": "lc-disc-10879-c11825412-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b04393",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1154.28,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 4050; llama 7B Q4_0 | pp512 | 1154.28 + 15.76; llama 7B Q4_0 | tg128 | 41.89 + 0.10",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-11825412",
   "tg_tokens": 128,
   "tps": 41.89,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-01-13",
   "flags": [],
   "hardware": "RTX 4050",
   "id": "lc-disc-10879-c11825412-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; config=b04393",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1725.85,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 4050; llama 7B Q4_0 | pp512 | 1725.85 + 17.85; llama 7B Q4_0 | tg128 | 43.72 + 0.41",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-11825412",
   "tg_tokens": 128,
   "tps": 43.72,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-01-14",
   "flags": [],
   "hardware": "Radeon RX 6600 XT",
   "id": "lc-disc-10879-c11836134-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b04481",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 574.65,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6600 XT; llama 7B Q4_0 | pp512 | 574.65 \u00b1 0.86; llama 7B Q4_0 | tg128 | 53.92 \u00b1 0.11",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-11836134",
   "tg_tokens": 128,
   "tps": 53.92,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-01-16",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc-disc-10879-c11859632-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; ngl=99; config=b04465",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4368.74,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3090; llama 7B Q4_0 | pp512 | 4368.74 \u00b1 48.23; llama 7B Q4_0 | tg128 | 118.50 \u00b1 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-11859632",
   "tg_tokens": 128,
   "tps": 118.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-01-16",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc-disc-10879-c11859632-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; config=b04465",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4761.91,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3090; llama 7B Q4_0 | pp512 | 4761.91 \u00b1 5.84; llama 7B Q4_0 | tg128 | 107.32 \u00b1 0.34",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-11859632",
   "tg_tokens": 128,
   "tps": 107.32,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-01-15",
   "flags": [],
   "hardware": "Radeon RX 5700 XT",
   "id": "lc-disc-10879-c11846756-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b04488",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 609.57,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 5700 XT; llama 7B Q4_0 | pp512 | 609.57 \u00b1 0.35; llama 7B Q4_0 | tg128 | 67.53 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-11846756",
   "tg_tokens": 128,
   "tps": 67.53,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-01-15",
   "flags": [],
   "hardware": "Radeon RX 5700 XT",
   "id": "lc-disc-10879-c11846756-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; config=b04488",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 605.4,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 5700 XT; llama 7B Q4_0 | pp512 | 605.40 \u00b1 1.16; llama 7B Q4_0 | tg128 | 69.14 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-11846756",
   "tg_tokens": 128,
   "tps": 69.14,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-01-18",
   "flags": [],
   "hardware": "Radeon RX 6900 XT",
   "id": "lc-disc-10879-c11875823-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b04503",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1257.98,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6900 XT; llama 7B Q4_0 | pp512 | 1257.98 \u00b1 1.55; llama 7B Q4_0 | tg128 | 101.42 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-11875823",
   "tg_tokens": 128,
   "tps": 101.42,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-01-18",
   "flags": [],
   "hardware": "Radeon RX 6900 XT",
   "id": "lc-disc-10879-c11875823-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; config=b04503",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1018.94,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6900 XT; llama 7B Q4_0 | pp512 | 1018.94 \u00b1 2.02; llama 7B Q4_0 | tg128 | 95.70 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-11875823",
   "tg_tokens": 128,
   "tps": 95.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-01-18",
   "flags": [],
   "hardware": "Radeon RX 6900 XT",
   "id": "lc-disc-10879-c11875823-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100; config=b04503",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1963.96,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6900 XT; llama 7B Q4_0 | pp512 | 1963.96 \u00b1 2.14; llama 7B Q4_0 | tg128 | 85.48 \u00b1 0.17",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-11875823",
   "tg_tokens": 128,
   "tps": 85.48,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-02-05",
   "flags": [],
   "hardware": "Radeon RX 6900 XT",
   "id": "lc-disc-10879-c12074955-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b04644",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1121.36,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6900 XT; llama 7B Q4_0 | pp512 | 1121.36 \u00b1 0.91; llama 7B Q4_0 | tg128 | 98.89 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-12074955",
   "tg_tokens": 128,
   "tps": 98.89,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-01-21",
   "flags": [],
   "hardware": "RTX 3060",
   "id": "lc-disc-10879-c11908439-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; ngl=100",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1298.03,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3060; llama 7B Q4_0 | pp512 | 1298.03 \u00b1 23.40; llama 7B Q4_0 | tg128 | 54.28 \u00b1 1.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-11908439",
   "tg_tokens": 128,
   "tps": 54.28,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-02-06",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c12077156-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; ngl=100",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3236.63,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3236.63 \u00b1 71.63; llama 7B Q4_0 | tg128 | 148.21 \u00b1 0.94",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-12077156",
   "tg_tokens": 128,
   "tps": 148.21,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-02-06",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c12077156-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3266.07,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3266.07 \u00b1 29.65; llama 7B Q4_0 | tg128 | 157.91 \u00b1 0.13",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-12077156",
   "tg_tokens": 128,
   "tps": 157.91,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-02-07",
   "flags": [],
   "hardware": "RTX 2070 Super",
   "id": "lc-disc-10879-c12092476-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1199.13,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 2070 Super; llama 7B Q4_0 | pp512 | 1199.13 \u00b1 7.70; llama 7B Q4_0 | tg128 | 64.64 \u00b1 0.20",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-12092476",
   "tg_tokens": 128,
   "tps": 64.64,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-02-08",
   "flags": [],
   "hardware": "Radeon RX 6600 XT",
   "id": "lc-disc-10879-c12101309-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; ngl=100; config=b04650",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 428.92,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6600 XT; llama 7B Q4_0 | pp512 | 428.92 \u00b1 0.17; llama 7B Q4_0 | tg128 | 47.94 \u00b1 0.70",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-12101309",
   "tg_tokens": 128,
   "tps": 47.94,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-02-08",
   "flags": [],
   "hardware": "Radeon RX 6900 XT",
   "id": "lc-disc-10879-c12101309-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=10; config=b04650",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 713.77,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6900 XT; llama 7B Q4_0 | pp512 | 713.77 \u00b1 0.62; llama 7B Q4_0 | tg128 | 83.54 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-12101309",
   "tg_tokens": 128,
   "tps": 83.54,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-03",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-10879-c16795994-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b09010",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4666.15,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; dev=VULKAN0; llama 7B Q4_0 | pp512 | 4666.15 \u00b1 12.89; llama 7B Q4_0 | tg128 | 164.05 \u00b1 0.89",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-16795994",
   "tg_tokens": 128,
   "tps": 164.05,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-03",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-10879-c16795994-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b09010",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 5200.68,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; dev=VULKAN0; llama 7B Q4_0 | pp512 | 5200.68 \u00b1 16.44; llama 7B Q4_0 | tg128 | 171.61 \u00b1 2.69",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-16795994",
   "tg_tokens": 128,
   "tps": 171.61,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-27",
   "flags": [],
   "hardware": "Radeon RX 470 G",
   "id": "lc-disc-10879-c17069649-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b08966",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 198.66,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 470 G; llama 7B Q4_0 | pp512 | 198.66 \u00b1 0.83; llama 7B Q4_0 | tg128 | 38.96 \u00b1 0.12",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-17069649",
   "tg_tokens": 128,
   "tps": 38.96,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-27",
   "flags": [],
   "hardware": "Radeon RX 470 G",
   "id": "lc-disc-10879-c17069649-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; config=b08966",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 198.3,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 470 G; llama 7B Q4_0 | pp512 | 198.30 \u00b1 0.20; llama 7B Q4_0 | tg128 | 38.69 \u00b1 0.15",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-17069649",
   "tg_tokens": 128,
   "tps": 38.69,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-27",
   "flags": [],
   "hardware": "Radeon RX 470 G",
   "id": "lc-disc-10879-c17069649-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100; config=b08966",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 207.79,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 470 G; llama 7B Q4_0 | pp512 | 207.79 \u00b1 0.90; llama 7B Q4_0 | tg128 | 39.75 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-17069649",
   "tg_tokens": 128,
   "tps": 39.75,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-27",
   "flags": [],
   "hardware": "Radeon RX 470 G",
   "id": "lc-disc-10879-c17069649-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=100; config=b08966",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 206.61,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 470 G; dev=VULKAN0; llama 7B Q4_0 | pp512 | 206.61 \u00b1 0.77; llama 7B Q4_0 | tg128 | 39.85 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-17069649",
   "tg_tokens": 128,
   "tps": 39.85,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-30",
   "flags": [],
   "hardware": "Arc Pro B50",
   "id": "lc-disc-10879-c17118759-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b09437",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 590.01,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B50; llama 7B Q4_0 | pp512 | 590.01 \u00b1 0.93; llama 7B Q4_0 | tg128 | 40.13 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-17118759",
   "tg_tokens": 128,
   "tps": 40.13,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-30",
   "flags": [],
   "hardware": "Arc Pro B50",
   "id": "lc-disc-10879-c17118759-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b09437",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 581.78,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B50; llama 7B Q4_0 | pp512 | 581.78 \u00b1 1.61; llama 7B Q4_0 | tg128 | 41.82 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-17118759",
   "tg_tokens": 128,
   "tps": 41.82,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-30",
   "flags": [],
   "hardware": "Arc Pro B50",
   "id": "lc-disc-10879-c17118759-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=-1; config=b09437",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1186.45,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B50; llama 7B Q4_0 | pp512 | 1186.45 \u00b1 4.43; llama 7B Q4_0 | tg128 | 46.30 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-17118759",
   "tg_tokens": 128,
   "tps": 46.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-30",
   "flags": [],
   "hardware": "Arc Pro B50",
   "id": "lc-disc-10879-c17118759-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b09437",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1122.53,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B50; llama 7B Q4_0 | pp512 | 1122.53 \u00b1 3.73; llama 7B Q4_0 | tg128 | 47.69 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-17118759",
   "tg_tokens": 128,
   "tps": 47.69,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-31",
   "flags": [],
   "hardware": "Arc A770",
   "id": "lc-disc-10879-c17121334-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b09439",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1209.43,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A770; llama 7B Q4_0 | pp512 | 1209.43 \u00b1 1.62; llama 7B Q4_0 | tg128 | 39.13 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-17121334",
   "tg_tokens": 128,
   "tps": 39.13,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-31",
   "flags": [],
   "hardware": "Arc A770",
   "id": "lc-disc-10879-c17121334-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b09439",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1243.46,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A770; llama 7B Q4_0 | pp512 | 1243.46 \u00b1 2.55; llama 7B Q4_0 | tg128 | 40.62 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-17121334",
   "tg_tokens": 128,
   "tps": 40.62,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-21",
   "flags": [],
   "hardware": "Radeon RX 580",
   "id": "lc-disc-10879-c17377824-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100; config=b09222",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 96.83,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 580; llama 7B Q4_0 | pp512 | 96.83 \u00b1 0.78; llama 7B Q4_0 | tg128 | 17.00 \u00b1 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-17377824",
   "tg_tokens": 128,
   "tps": 17.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-21",
   "flags": [],
   "hardware": "Radeon RX 580",
   "id": "lc-disc-10879-c17377824-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b09222",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 102.86,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 580; llama 7B Q4_0 | pp512 | 102.86 \u00b1 0.22; llama 7B Q4_0 | tg128 | 18.92 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-17377824",
   "tg_tokens": 128,
   "tps": 18.92,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-07-14",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-disc-10879-c17635924-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=-1; config=b10001",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2789.51,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GB10; dev=VULKAN0; llama 7B Q4_0 | pp512 | 2789.51 \u00b1 18.33; llama 7B Q4_0 | tg128 | 58.51 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-17635924",
   "tg_tokens": 128,
   "tps": 58.51,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-07-14",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-disc-10879-c17635924-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10001",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3321.91,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GB10; dev=VULKAN0; llama 7B Q4_0 | pp512 | 3321.91 \u00b1 5.86; llama 7B Q4_0 | tg128 | 60.39 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-17635924",
   "tg_tokens": 128,
   "tps": 60.39,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-07-17",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-10879-c17675617-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b09739",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2531.36,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B70; llama 7B Q4_0 | pp512 | 2531.36 \u00b1 41.02; llama 7B Q4_0 | tg128 | 80.39 \u00b1 0.32",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-17675617",
   "tg_tokens": 128,
   "tps": 80.39,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-07-17",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-10879-c17675617-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b09739",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2461.92,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B70; llama 7B Q4_0 | pp512 | 2461.92 \u00b1 9.12; llama 7B Q4_0 | tg128 | 82.19 \u00b1 0.15",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-17675617",
   "tg_tokens": 128,
   "tps": 82.19,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-07-18",
   "flags": [],
   "hardware": "Radeon RX 6600",
   "id": "lc-disc-10879-c17680588-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b10066",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 758.5,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6600; llama 7B Q4_0 | pp512 | 758.50 \u00b1 0.25; llama 7B Q4_0 | tg128 | 50.57 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-17680588",
   "tg_tokens": 128,
   "tps": 50.57,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-07-18",
   "flags": [],
   "hardware": "Radeon RX 6600",
   "id": "lc-disc-10879-c17680588-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10066",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 804.64,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6600; llama 7B Q4_0 | pp512 | 804.64 \u00b1 0.21; llama 7B Q4_0 | tg128 | 53.17 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-17680588",
   "tg_tokens": 128,
   "tps": 53.17,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-02-09",
   "flags": [],
   "hardware": "Radeon 780M",
   "id": "lc-disc-10879-c12109521-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b04677",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 245.79,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 780M; llama 7B Q4_0 | pp512 | 245.79 \u00b1 2.97; llama 7B Q4_0 | tg128 | 20.10 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-12109521",
   "tg_tokens": 128,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-02-11",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "lc-disc-10879-c12139359-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b04687",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 259.11,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M3 Pro; llama 7B Q4_0 | pp512 | 259.11 \u00b1 1.46; llama 7B Q4_0 | tg128 | 26.23 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-12139359",
   "tg_tokens": 128,
   "tps": 26.23,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-02-11",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "lc-disc-10879-c12139359-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=6; config=b04687",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 263.7,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M3 Pro; llama 7B Q4_0 | pp512 | 263.70 \u00b1 0.02; llama 7B Q4_0 | tg128 | 26.39 \u00b1 0.14",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-12139359",
   "tg_tokens": 128,
   "tps": 26.39,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2025-02-11",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "lc-disc-10879-c12139359-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100; config=b04687",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 371.72,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M3 Pro; llama 7B Q4_0 | pp512 | 371.72 \u00b1 0.19; llama 7B Q4_0 | tg128 | 31.54 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-12139359",
   "tg_tokens": 128,
   "tps": 31.54,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2025-02-11",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "lc-disc-10879-c12139359-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=6; config=b04687",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 368.92,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M3 Pro; llama 7B Q4_0 | pp512 | 368.92 \u00b1 4.92; llama 7B Q4_0 | tg128 | 31.52 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-12139359",
   "tg_tokens": 128,
   "tps": 31.52,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-02-21",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-disc-10879-c12273095-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; ngl=100; config=b04749",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 883.2,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7800 XT; llama 7B Q4_0 | pp512 | 883.20 \u00b1 0.80; llama 7B Q4_0 | tg128 | 103.98 \u00b1 0.23",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-12273095",
   "tg_tokens": 128,
   "tps": 103.98,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-02-21",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-disc-10879-c12273655-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; ngl=100; config=b04749",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1260.54,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7800 XT; llama 7B Q4_0 | pp512 | 1260.54 \u00b1 10.51; llama 7B Q4_0 | tg128 | 107.53 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-12273655",
   "tg_tokens": 128,
   "tps": 107.53,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-03-01",
   "flags": [],
   "hardware": "Radeon 780M",
   "id": "lc-disc-10879-c12357861-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=8; config=b04793",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 237.73,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 780M; llama 7B Q4_0 | pp512 | 237.73 \u00b1 13.98; llama 7B Q4_0 | tg128 | 18.22 \u00b1 0.62",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-12357861",
   "tg_tokens": 128,
   "tps": 18.22,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-03-10",
   "flags": [],
   "hardware": "Radeon RX 6600",
   "id": "lc-disc-10879-c12443489-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 380.87,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6600; llama 7B Q4_0 | pp512 | 380.87 \u00b1 0.21; llama 7B Q4_0 | tg128 | 47.47 \u00b1 0.18",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-12443489",
   "tg_tokens": 128,
   "tps": 47.47,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-03-10",
   "flags": [],
   "hardware": "Radeon RX 6600",
   "id": "lc-disc-10879-c12443489-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 331.58,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6600; llama 7B Q4_0 | pp512 | 331.58 \u00b1 0.06; llama 7B Q4_0 | tg128 | 49.76 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-12443489",
   "tg_tokens": 128,
   "tps": 49.76,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-03-10",
   "flags": [],
   "hardware": "Radeon RX 6600",
   "id": "lc-disc-10879-c12443489-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=0",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 26.05,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6600; llama 7B Q4_0 | pp512 | 26.05 \u00b1 0.02; llama 7B Q4_0 | tg128 | 12.16 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-12443489",
   "tg_tokens": 128,
   "tps": 12.16,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": null,
   "ctx": null,
   "date": "2025-03-10",
   "flags": [],
   "hardware": "Radeon RX 6600",
   "id": "lc-disc-10879-c12443489-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=6",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 26.26,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6600; llama 7B Q4_0 | pp512 | 26.26 \u00b1 0.01; llama 7B Q4_0 | tg128 | 12.08 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-12443489",
   "tg_tokens": 128,
   "tps": 12.08,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-03-17",
   "flags": [],
   "hardware": "Radeon 780M",
   "id": "lc-disc-10879-c12531197-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b04635",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 281.62,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 780M; llama 7B Q4_0 | pp512 | 281.62 \u00b1 1.56; llama 7B Q4_0 | tg128 | 19.91 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-12531197",
   "tg_tokens": 128,
   "tps": 19.91,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-03-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon 780M",
   "id": "lc-disc-10879-c12531197-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; config=b04635",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 205.15,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 780M; llama 7B Q4_0 | pp512 | 205.15 \u00b1 3.01; llama 7B Q4_0 | tg128 | 16.60 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-12531197",
   "tg_tokens": 128,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-21",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c16241493-glm4moe106ba12b-Q3_K - Medium",
   "model": "glm4moe 106B.A12B",
   "notes": "threads=48; config=b08463",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 263.73,
   "provenance": "community",
   "quant": "Q3_K - Medium",
   "quote": "Radeon RX 7900 XTX; glm4moe 106B.A12B Q3_K - Medium | pp512 | 263.73 \u00b1 1.08; glm4moe 106B.A12B Q3_K - Medium | tg128 | 13.61 \u00b1 0.19",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-16241493",
   "tg_tokens": 128,
   "tps": 13.61,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-21",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c16241493-glm4moe106ba12b-Q3_K - Medium-2",
   "model": "glm4moe 106B.A12B",
   "notes": "threads=48; config=b08463",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 252.83,
   "provenance": "community",
   "quant": "Q3_K - Medium",
   "quote": "Radeon RX 7900 XTX; dev=VULKAN0; glm4moe 106B.A12B Q3_K - Medium | pp512 | 252.83 \u00b1 1.62; glm4moe 106B.A12B Q3_K - Medium | tg128 | 10.35 \u00b1 0.17",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-16241493",
   "tg_tokens": 128,
   "tps": 10.35,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-24",
   "flags": [],
   "hardware": "A770",
   "id": "lc-disc-10879-c16298977-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b08401",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1073.85,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A770; llama 7B Q4_0 | pp512 | 1073.85 + 29.68; llama 7B Q4_0 | tg128 | 52.56 + 0.11",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-16298977",
   "tg_tokens": 128,
   "tps": 52.56,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-24",
   "flags": [],
   "hardware": "A770",
   "id": "lc-disc-10879-c16298977-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b08401",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1119.68,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A770; llama 7B Q4_0 | pp512 | 1119.68 + 30.25; llama 7B Q4_0 | tg128 | 53.07 + 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-16298977",
   "tg_tokens": 128,
   "tps": 53.07,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-27",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-10879-c16350652-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4945.07,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 4945.07 \u00b1 439.22; llama 7B Q4_0 | tg128 | 125.64 \u00b1 0.51",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-16350652",
   "tg_tokens": 128,
   "tps": 125.64,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-27",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-10879-c16350652-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 5505.31,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 5505.31 \u00b1 6.05; llama 7B Q4_0 | tg128 | 131.01 \u00b1 0.59",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-16350652",
   "tg_tokens": 128,
   "tps": 131.01,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-04-03",
   "flags": [],
   "hardware": "RTX 3080",
   "id": "lc-disc-10879-c16432568-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=1 ngl=99; config=b08642",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 230.79,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3080; dev=VULKAN0; llama 7B Q4_0 | pp512 | 230.79 \u00b1 2.19; llama 7B Q4_0 | tg128 | 18.32 \u00b1 0.14",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-16432568",
   "tg_tokens": 128,
   "tps": 18.32,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-04-03",
   "flags": [],
   "hardware": "RTX 3080",
   "id": "lc-disc-10879-c16432568-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; config=b08642",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4889.13,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3080; dev=VULKAN1; llama 7B Q4_0 | pp512 | 4889.13 \u00b1 58.58; llama 7B Q4_0 | tg128 | 143.44 \u00b1 0.62",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-16432568",
   "tg_tokens": 128,
   "tps": 143.44,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-04-24",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-10879-c16706548-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b08892",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 7567.94,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; dev=VULKAN0; llama 7B Q4_0 | pp512 | 7567.94 \u00b1 41.96; llama 7B Q4_0 | tg128 | 176.18 \u00b1 1.12",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-16706548",
   "tg_tokens": 128,
   "tps": 176.18,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-04-25",
   "flags": [],
   "hardware": "Radeon 760M",
   "id": "lc-disc-10879-c16710467-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b08910",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 187.02,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 760M; llama 7B Q4_0 | pp512 | 187.02 \u00b1 1.12; llama 7B Q4_0 | tg128 | 17.69 \u00b1 0.65",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-16710467",
   "tg_tokens": 128,
   "tps": 17.69,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-04-25",
   "flags": [],
   "hardware": "Radeon 760M",
   "id": "lc-disc-10879-c16710467-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b08910",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 219.24,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 760M; llama 7B Q4_0 | pp512 | 219.24 \u00b1 3.57; llama 7B Q4_0 | tg128 | 18.89 \u00b1 0.35",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-16710467",
   "tg_tokens": 128,
   "tps": 18.89,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-04-25",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon RX 570 S",
   "id": "lc-disc-10879-c16711256-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b08910",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 52.66,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 570 S; llama 7B Q4_0 | pp512 | 52.66 \u00b1 0.01; llama 7B Q4_0 | tg128 | 24.11 \u00b1 0.11",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-16711256",
   "tg_tokens": 128,
   "tps": 24.11,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-04-25",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon RX 570 S",
   "id": "lc-disc-10879-c16711256-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b08910",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 53.68,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 570 S; llama 7B Q4_0 | pp512 | 53.68 \u00b1 0.00; llama 7B Q4_0 | tg128 | 24.63 \u00b1 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-16711256",
   "tg_tokens": 128,
   "tps": 24.63,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-04-25",
   "flags": [],
   "hardware": "Radeon RX 570 S",
   "id": "lc-disc-10879-c16711256-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100; config=b08910",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 226.11,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 570 S; llama 7B Q4_0 | pp512 | 226.11 \u00b1 0.45; llama 7B Q4_0 | tg128 | 37.44 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-16711256",
   "tg_tokens": 128,
   "tps": 37.44,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-04-25",
   "flags": [],
   "hardware": "Radeon RX 570 S",
   "id": "lc-disc-10879-c16711256-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b08910",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 229.23,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 570 S; llama 7B Q4_0 | pp512 | 229.23 \u00b1 0.45; llama 7B Q4_0 | tg128 | 40.18 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-16711256",
   "tg_tokens": 128,
   "tps": 40.18,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-04-30",
   "flags": [],
   "hardware": "RTX 3060 Ti",
   "id": "lc-disc-10879-c16769718-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b08982",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2361.43,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3060 Ti; llama 7B Q4_0 | pp512 | 2361.43 \u00b1 4.32; llama 7B Q4_0 | tg128 | 86.74 \u00b1 0.34",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-16769718",
   "tg_tokens": 128,
   "tps": 86.74,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-04-30",
   "flags": [],
   "hardware": "RTX 3060 Ti",
   "id": "lc-disc-10879-c16769718-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; config=b08982",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2644.95,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3060 Ti; llama 7B Q4_0 | pp512 | 2644.95 \u00b1 26.82; llama 7B Q4_0 | tg128 | 90.35 \u00b1 1.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-16769718",
   "tg_tokens": 128,
   "tps": 90.35,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-01",
   "flags": [],
   "hardware": "Radeon 780M",
   "id": "lc-disc-10879-c16781874-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b08768",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 267.27,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 780M; llama 7B Q4_0 | pp512 | 267.27 \u00b1 7.61; llama 7B Q4_0 | tg128 | 19.81 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-16781874",
   "tg_tokens": 128,
   "tps": 19.81,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-01",
   "flags": [],
   "hardware": "Radeon 780M",
   "id": "lc-disc-10879-c16781874-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b08768",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 311.02,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 780M; llama 7B Q4_0 | pp512 | 311.02 \u00b1 0.12; llama 7B Q4_0 | tg128 | 21.28 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-16781874",
   "tg_tokens": 128,
   "tps": 21.28,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-03-18",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-10879-c12544053-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b04920",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2206.24,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 2206.24 \u00b1 266.06; llama 7B Q4_0 | tg128 | 117.22 \u00b1 0.33",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-12544053",
   "tg_tokens": 128,
   "tps": 117.22,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-03-20",
   "flags": [
    "contradiction"
   ],
   "hardware": "RTX 3060",
   "id": "lc-disc-10879-c12568713-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; ngl=100; config=b04930",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 90.15,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3060; llama 7B Q4_0 | pp512 | 90.15 \u00b1 1.45; llama 7B Q4_0 | tg128 | 10.81 \u00b1 0.14",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-12568713",
   "tg_tokens": 128,
   "tps": 10.81,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-03-20",
   "flags": [],
   "hardware": "RTX 3060",
   "id": "lc-disc-10879-c12568713-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; ngl=100; config=b04930",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1059.76,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3060; llama 7B Q4_0 | pp512 | 1059.76 \u00b1 3.54; llama 7B Q4_0 | tg128 | 49.03 \u00b1 0.13",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-12568713",
   "tg_tokens": 128,
   "tps": 49.03,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-03-20",
   "flags": [],
   "hardware": "RTX 3060",
   "id": "lc-disc-10879-c12568713-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100; config=b04930",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 254.09,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3060; llama 7B Q4_0 | pp512 | 254.09 \u00b1 14.64; llama 7B Q4_0 | tg128 | 20.18 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-12568713",
   "tg_tokens": 128,
   "tps": 20.18,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-04-01",
   "flags": [],
   "hardware": "A770",
   "id": "lc-disc-10879-c12685855-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b05010",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1116.8,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A770; llama 7B Q4_0 | pp512 | 1116.80 \u00b1 1.84; llama 7B Q4_0 | tg128 | 48.45 \u00b1 1.32",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-12685855",
   "tg_tokens": 128,
   "tps": 48.45,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2025-04-01",
   "flags": [],
   "hardware": "A770",
   "id": "lc-disc-10879-c12685855-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; ngl=100; config=b05010",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2478.2,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A770; llama 7B Q4_0 | pp512 | 2478.20 \u00b1 10.42; llama 7B Q4_0 | tg128 | 70.18 \u00b1 0.57",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-12685855",
   "tg_tokens": 128,
   "tps": 70.18,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-04-03",
   "flags": [],
   "hardware": "A770",
   "id": "lc-disc-10879-c12716702-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b05039",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 603.46,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A770; llama 7B Q4_0 | pp512 | 603.46 \u00b1 0.92; llama 7B Q4_0 | tg128 | 49.02 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-12716702",
   "tg_tokens": 128,
   "tps": 49.02,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-04-10",
   "flags": [],
   "hardware": "Radeon RX 6600M",
   "id": "lc-disc-10879-c12795174-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b05097",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 605.59,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6600M; llama 7B Q4_0 | pp512 | 605.59 \u00b1 0.65; llama 7B Q4_0 | tg128 | 48.21 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-12795174",
   "tg_tokens": 128,
   "tps": 48.21,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-04-12",
   "flags": [],
   "hardware": "Radeon RX 6800 XT",
   "id": "lc-disc-10879-c12811083-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1533.6,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6800 XT; llama 7B Q4_0 | pp512 | 1533.60 \u00b1 2.47; llama 7B Q4_0 | tg128 | 95.56 \u00b1 0.72",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-12811083",
   "tg_tokens": 128,
   "tps": 95.56,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-04-12",
   "flags": [],
   "hardware": "Radeon RX 6800 XT",
   "id": "lc-disc-10879-c12811083-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 612.47,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6800 XT; llama 7B Q4_0 | pp512 | 612.47 \u00b1 0.87; llama 7B Q4_0 | tg128 | 71.37 \u00b1 0.98",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-12811083",
   "tg_tokens": 128,
   "tps": 71.37,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-04-12",
   "flags": [],
   "hardware": "Arc A770",
   "id": "lc-disc-10879-c12811083-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 661.92,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A770; llama 7B Q4_0 | pp512 | 661.92 \u00b1 0.47; llama 7B Q4_0 | tg128 | 39.41 \u00b1 0.41",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-12811083",
   "tg_tokens": 128,
   "tps": 39.41,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-04-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "RTX 3090",
   "id": "lc-disc-10879-c12811083-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4241.77,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3090; llama 7B Q4_0 | pp512 | 4241.77 \u00b1 88.02; llama 7B Q4_0 | tg128 | 122.61 \u00b1 0.93",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-12811083",
   "tg_tokens": 128,
   "tps": 122.61,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-04-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "RTX 3090",
   "id": "lc-disc-10879-c12811083-llama7b-Q4_0-5",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4516.92,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3090; llama 7B Q4_0 | pp512 | 4516.92 \u00b1 9.55; llama 7B Q4_0 | tg128 | 120.44 \u00b1 2.58",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-12811083",
   "tg_tokens": 128,
   "tps": 120.44,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-04-18",
   "flags": [],
   "hardware": "A770",
   "id": "lc-disc-10879-c12878591-qwen214b-Q4_0",
   "model": "qwen2 14B",
   "notes": "threads=99; config=b05083",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 647.35,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A770; qwen2 14B Q4_0 | pp512 | 647.35 \u00b1 1.85; qwen2 14B Q4_0 | tg128 | 31.09 \u00b1 0.50; qwen2 14B Q4_0 | tg256 | 30.72 \u00b1 0.45; qwen2 14B Q4_0 | tg512 | 31.04 \u00b1 0.31",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-12878591",
   "tg_tokens": 128,
   "tps": 31.09,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-04-18",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-10879-c12878591-qwen214b-Q4_0-2",
   "model": "qwen2 14B",
   "notes": "threads=99; config=b05083",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2039.45,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; qwen2 14B Q4_0 | pp512 | 2039.45 \u00b1 1.97; qwen2 14B Q4_0 | tg128 | 58.78 \u00b1 0.55; qwen2 14B Q4_0 | tg256 | 57.86 \u00b1 0.48; qwen2 14B Q4_0 | tg512 | 58.16 \u00b1 0.68",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-12878591",
   "tg_tokens": 128,
   "tps": 58.78,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-04-22",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-disc-10879-c12916199-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; ngl=99; config=b05170",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3211.73,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 5060 Ti; llama 7B Q4_0 | pp512 | 3211.73 \u00b1 24.44; llama 7B Q4_0 | tg128 | 81.48 \u00b1 3.50",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-12916199",
   "tg_tokens": 128,
   "tps": 81.48,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-04-22",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-disc-10879-c12916199-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; config=b05170",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3492.22,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 5060 Ti; llama 7B Q4_0 | pp512 | 3492.22 \u00b1 15.73; llama 7B Q4_0 | tg128 | 83.26 \u00b1 2.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-12916199",
   "tg_tokens": 128,
   "tps": 83.26,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-04-24",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-disc-10879-c12938830-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b05170",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4034.65,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 5060 Ti; llama 7B Q4_0 | pp512 | 4034.65 \u00b1 2.41; llama 7B Q4_0 | tg128 | 94.30 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-12938830",
   "tg_tokens": 128,
   "tps": 94.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-04-27",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "lc-disc-10879-c12958536-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=24; config=b05195",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1116.48,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M3 Ultra; llama 7B Q4_0 | pp512 | 1116.48 \u00b1 0.20; llama 7B Q4_0 | tg128 | 114.55 \u00b1 1.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-12958536",
   "tg_tokens": 128,
   "tps": 114.55,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-04-27",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "lc-disc-10879-c12958536-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; config=b05195",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1116.83,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M3 Ultra; llama 7B Q4_0 | pp512 | 1116.83 \u00b1 0.55; llama 7B Q4_0 | tg128 | 115.54 \u00b1 0.78",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-12958536",
   "tg_tokens": 128,
   "tps": 115.54,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2025-04-27",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "lc-disc-10879-c12958536-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100; config=b05195",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1574.09,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M3 Ultra; llama 7B Q4_0 | pp512 | 1574.09 \u00b1 1.42; llama 7B Q4_0 | tg128 | 102.28 \u00b1 0.78",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-12958536",
   "tg_tokens": 128,
   "tps": 102.28,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-01-29",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-disc-10879-c15642964-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b07875",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2017.33,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7800 XT; llama 7B Q4_0 | pp512 | 2017.33 \u00b1 19.30; llama 7B Q4_0 | tg128 | 118.27 \u00b1 0.27",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15642964",
   "tg_tokens": 128,
   "tps": 118.27,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-01-29",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-disc-10879-c15642964-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07875",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2197.05,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7800 XT; llama 7B Q4_0 | pp512 | 2197.05 \u00b1 6.03; llama 7B Q4_0 | tg128 | 124.86 \u00b1 0.10",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15642964",
   "tg_tokens": 128,
   "tps": 124.86,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-01-29",
   "flags": [],
   "hardware": "Radeon RX 5700 XT",
   "id": "lc-disc-10879-c15642964-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100; config=b07875",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 529.69,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 5700 XT; llama 7B Q4_0 | pp512 | 529.69 \u00b1 0.26; llama 7B Q4_0 | tg128 | 70.73 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15642964",
   "tg_tokens": 128,
   "tps": 70.73,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-01-29",
   "flags": [],
   "hardware": "Radeon RX 5700 XT",
   "id": "lc-disc-10879-c15642964-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07875",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 538.31,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 5700 XT; llama 7B Q4_0 | pp512 | 538.31 \u00b1 0.35; llama 7B Q4_0 | tg128 | 74.43 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15642964",
   "tg_tokens": 128,
   "tps": 74.43,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-02-03",
   "flags": [],
   "hardware": "Radeon RX 9060 XT",
   "id": "lc-disc-10879-c15686767-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b07898",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3460.92,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9060 XT; llama 7B Q4_0 | pp512 | 3460.92 \u00b1 7.16; llama 7B Q4_0 | tg128 | 93.51 \u00b1 0.15",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15686767",
   "tg_tokens": 128,
   "tps": 93.51,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-02-03",
   "flags": [],
   "hardware": "Radeon RX 9060 XT",
   "id": "lc-disc-10879-c15686767-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07898",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3912.65,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9060 XT; llama 7B Q4_0 | pp512 | 3912.65 \u00b1 5.86; llama 7B Q4_0 | tg128 | 97.01 \u00b1 0.14",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15686767",
   "tg_tokens": 128,
   "tps": 97.01,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-02-03",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon RX 9060 XT",
   "id": "lc-disc-10879-c15686767-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; config=b07898",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 103.01,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Radeon RX 9060 XT (RADV GFX1200) (radv) | uma: 0 | fp16: 1 | bf16: 1 | warp; llama 7B Q4_0 | pp512 | 103.01 \u00b1 0.24; llama 7B Q4_0 | tg128 | 15.07 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15686767",
   "tg_tokens": 128,
   "tps": 15.07,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-02-03",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon RX 9060 XT",
   "id": "lc-disc-10879-c15686767-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07898",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 142.04,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Radeon RX 9060 XT (RADV GFX1200) (radv) | uma: 0 | fp16: 1 | bf16: 1 | warp; llama 7B Q4_0 | pp512 | 142.04 \u00b1 0.29; llama 7B Q4_0 | tg128 | 28.37 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15686767",
   "tg_tokens": 128,
   "tps": 28.37,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-02-03",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-disc-10879-c15686845-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99 sm=none; config=b07898",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 511.59,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 5060 Ti; llama 7B Q4_0 | pp512 | 511.59 \u00b1 6.07; llama 7B Q4_0 | tg128 | 47.35 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15686845",
   "tg_tokens": 128,
   "tps": 47.35,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-02-03",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-disc-10879-c15686845-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99 sm=none; config=b07898",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 689.59,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 5060 Ti; llama 7B Q4_0 | pp512 | 689.59 \u00b1 1.20; llama 7B Q4_0 | tg128 | 60.02 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15686845",
   "tg_tokens": 128,
   "tps": 60.02,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-02-03",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-disc-10879-c15686845-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; config=b07898",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3342.62,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "NVIDIA GeForce RTX 5060 Ti (NVIDIA) | uma: 0 | fp16: 1 | bf16: 1 | warp size: 32; llama 7B Q4_0 | pp512 | 3342.62 \u00b1 6.17; llama 7B Q4_0 | tg128 | 92.19 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15686845",
   "tg_tokens": 128,
   "tps": 92.19,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-02-03",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-disc-10879-c15686845-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07898",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3848.92,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "NVIDIA GeForce RTX 5060 Ti (NVIDIA) | uma: 0 | fp16: 1 | bf16: 1 | warp size: 32; llama 7B Q4_0 | pp512 | 3848.92 \u00b1 4.29; llama 7B Q4_0 | tg128 | 95.98 \u00b1 0.14",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15686845",
   "tg_tokens": 128,
   "tps": 95.98,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-02-03",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-disc-10879-c15686874-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99 sm=none; config=b07898",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2021.05,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 5060 Ti; llama 7B Q4_0 | pp512 | 2021.05 \u00b1 3.73; llama 7B Q4_0 | tg128 | 49.44 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15686874",
   "tg_tokens": 128,
   "tps": 49.44,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-02-03",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-disc-10879-c15686874-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99 sm=none; config=b07898",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2199.15,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 5060 Ti; llama 7B Q4_0 | pp512 | 2199.15 \u00b1 4.57; llama 7B Q4_0 | tg128 | 62.05 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15686874",
   "tg_tokens": 128,
   "tps": 62.05,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-02-28",
   "flags": [],
   "hardware": "Radeon RX 9060 XT",
   "id": "lc-disc-10879-c15953889-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99 sm=none; config=b08157",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2093.56,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9060 XT; llama 7B Q4_0 | pp512 | 2093.56 \u00b1 17.98; llama 7B Q4_0 | tg128 | 71.45 \u00b1 0.66",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15953889",
   "tg_tokens": 128,
   "tps": 71.45,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-02-28",
   "flags": [],
   "hardware": "Radeon RX 9060 XT",
   "id": "lc-disc-10879-c15953889-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99 sm=none; config=b08157",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2311.69,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9060 XT; llama 7B Q4_0 | pp512 | 2311.69 \u00b1 11.02; llama 7B Q4_0 | tg128 | 73.86 \u00b1 0.51",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15953889",
   "tg_tokens": 128,
   "tps": 73.86,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-02-28",
   "flags": [],
   "hardware": "Radeon RX 9060 XT",
   "id": "lc-disc-10879-c15953889-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; config=b08157",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 446.42,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9060 XT; llama 7B Q4_0 | pp512 | 446.42 \u00b1 0.53; llama 7B Q4_0 | tg128 | 66.67 \u00b1 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15953889",
   "tg_tokens": 128,
   "tps": 66.67,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-02-28",
   "flags": [],
   "hardware": "Radeon RX 9060 XT",
   "id": "lc-disc-10879-c15953889-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b08157",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 411.98,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9060 XT; llama 7B Q4_0 | pp512 | 411.98 \u00b1 0.40; llama 7B Q4_0 | tg128 | 72.00 \u00b1 0.22",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15953889",
   "tg_tokens": 128,
   "tps": 72.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-02-12",
   "flags": [],
   "hardware": "Radeon RX 6800",
   "id": "lc-disc-10879-c15786878-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 933.71,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6800; llama 7B Q4_0 | pp512 | 933.71 \u00b1 0.71; llama 7B Q4_0 | tg128 | 97.30 \u00b1 0.26",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15786878",
   "tg_tokens": 128,
   "tps": 97.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-02-12",
   "flags": [],
   "hardware": "Radeon RX 6800",
   "id": "lc-disc-10879-c15786878-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 949.45,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6800; llama 7B Q4_0 | pp512 | 949.45 \u00b1 0.56; llama 7B Q4_0 | tg128 | 98.11 \u00b1 1.27",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15786878",
   "tg_tokens": 128,
   "tps": 98.11,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-02-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "Radeon RX 6800",
   "id": "lc-disc-10879-c15786878-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 103.87,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6800; dev=VULKAN1; llama 7B Q4_0 | pp512 | 103.87 \u00b1 0.02; llama 7B Q4_0 | tg128 | 9.63 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15786878",
   "tg_tokens": 128,
   "tps": 9.63,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-02-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "Radeon RX 6800",
   "id": "lc-disc-10879-c15786878-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=100",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 113.32,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6800; dev=VULKAN1; llama 7B Q4_0 | pp512 | 113.32 \u00b1 0.01; llama 7B Q4_0 | tg128 | 9.87 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15786878",
   "tg_tokens": 128,
   "tps": 9.87,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-02-13",
   "flags": [],
   "hardware": "Radeon RX 6800",
   "id": "lc-disc-10879-c15791261-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b08018",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1698.69,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6800; llama 7B Q4_0 | pp512 | 1698.69 \u00b1 0.80; llama 7B Q4_0 | tg128 | 95.61 \u00b1 0.19",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15791261",
   "tg_tokens": 128,
   "tps": 95.61,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-02-13",
   "flags": [],
   "hardware": "Radeon RX 6800",
   "id": "lc-disc-10879-c15791261-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; config=b08018",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1749.46,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6800; llama 7B Q4_0 | pp512 | 1749.46 \u00b1 3.36; llama 7B Q4_0 | tg128 | 96.65 \u00b1 0.48",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15791261",
   "tg_tokens": 128,
   "tps": 96.65,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-02-15",
   "flags": [],
   "hardware": "Radeon RX 7600 XT",
   "id": "lc-disc-10879-c15810129-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b08054",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 840.85,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7600 XT; llama 7B Q4_0 | pp512 | 840.85 \u00b1 3.02; llama 7B Q4_0 | tg128 | 53.02 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15810129",
   "tg_tokens": 128,
   "tps": 53.02,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-02-15",
   "flags": [],
   "hardware": "Radeon RX 7600 XT",
   "id": "lc-disc-10879-c15810129-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b08054",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1024.38,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7600 XT; llama 7B Q4_0 | pp512 | 1024.38 \u00b1 7.56; llama 7B Q4_0 | tg128 | 56.11 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15810129",
   "tg_tokens": 128,
   "tps": 56.11,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-02-15",
   "flags": [],
   "hardware": "Radeon RX 7600 XT",
   "id": "lc-disc-10879-c15810129-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; config=b08054",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2037.24,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7600 XT; llama 7B Q4_0 | pp512 | 2037.24 \u00b1 17.08; llama 7B Q4_0 | tg128 | 65.36 \u00b1 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15810129",
   "tg_tokens": 128,
   "tps": 65.36,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-02-15",
   "flags": [],
   "hardware": "Radeon RX 7600 XT",
   "id": "lc-disc-10879-c15810129-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b08054",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2251.85,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7600 XT; llama 7B Q4_0 | pp512 | 2251.85 \u00b1 3.84; llama 7B Q4_0 | tg128 | 67.98 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15810129",
   "tg_tokens": 128,
   "tps": 67.98,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-02-15",
   "flags": [],
   "hardware": "Radeon RX 7600 XT",
   "id": "lc-disc-10879-c15810129-llama7b-Q4_0-5",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b08054",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1039.52,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7600 XT; llama 7B Q4_0 | pp512 | 1039.52 \u00b1 1.38; llama 7B Q4_0 | tg128 | 52.69 \u00b1 0.13",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15810129",
   "tg_tokens": 128,
   "tps": 52.69,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-02-15",
   "flags": [],
   "hardware": "Radeon RX 7600 XT",
   "id": "lc-disc-10879-c15810129-llama7b-Q4_0-6",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b08054",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1028.16,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7600 XT; llama 7B Q4_0 | pp512 | 1028.16 \u00b1 10.02; llama 7B Q4_0 | tg128 | 55.20 \u00b1 0.28",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15810129",
   "tg_tokens": 128,
   "tps": 55.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-02-15",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "Radeon RX 7600 XT",
   "id": "lc-disc-10879-c15810129-llama7b-Q4_0-7",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b08054",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 34.95,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7600 XT; llama 7B Q4_0 | pp512 | 34.95 \u00b1 0.82; llama 7B Q4_0 | tg128 | 8.17 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15810129",
   "tg_tokens": 128,
   "tps": 8.17,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-02-15",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "Radeon RX 7600 XT",
   "id": "lc-disc-10879-c15810129-llama7b-Q4_0-8",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b08054",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 49.49,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7600 XT; llama 7B Q4_0 | pp512 | 49.49 \u00b1 2.23; llama 7B Q4_0 | tg128 | 16.06 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15810129",
   "tg_tokens": 128,
   "tps": 16.06,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-02-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon RX 7600 XT",
   "id": "lc-disc-10879-c15810129-llama7b-Q4_0-9",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b08054",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 401.03,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7600 XT; llama 7B Q4_0 | pp512 | 401.03 \u00b1 0.90; llama 7B Q4_0 | tg128 | 37.94 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15810129",
   "tg_tokens": 128,
   "tps": 37.94,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-02-15",
   "flags": [],
   "hardware": "Radeon RX 7600 XT",
   "id": "lc-disc-10879-c15810129-llama7b-Q4_0-10",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b08054",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 544.83,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7600 XT; llama 7B Q4_0 | pp512 | 544.83 \u00b1 2.40; llama 7B Q4_0 | tg128 | 52.80 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15810129",
   "tg_tokens": 128,
   "tps": 52.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-02-19",
   "flags": [],
   "hardware": "Radeon 680M",
   "id": "lc-disc-10879-c15860788-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b08099",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 670.41,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 680M; llama 7B Q4_0 | pp512 | 670.41 \u00b1 9.44; llama 7B Q4_0 | tg128 | 13.86 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15860788",
   "tg_tokens": 128,
   "tps": 13.86,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-02-19",
   "flags": [],
   "hardware": "Radeon 680M",
   "id": "lc-disc-10879-c15860788-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b08099",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1494.78,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 680M; llama 7B Q4_0 | pp512 | 1494.78 \u00b1 7.83; llama 7B Q4_0 | tg128 | 14.70 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15860788",
   "tg_tokens": 128,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-01",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15958734-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100; config=b08182",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3338.82,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; dev=VULKAN0; llama 7B Q4_0 | pp512 | 3338.82 \u00b1 69.83; llama 7B Q4_0 | tg128 | 161.07 \u00b1 0.15",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15958734",
   "tg_tokens": 128,
   "tps": 161.07,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-01",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15958734-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b08182",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3479.41,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; dev=VULKAN0; llama 7B Q4_0 | pp512 | 3479.41 \u00b1 25.85; llama 7B Q4_0 | tg128 | 168.62 \u00b1 0.45",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15958734",
   "tg_tokens": 128,
   "tps": 168.62,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-14",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-disc-10879-c16131738-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2731.93,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GB10; dev=VULKAN0; llama 7B Q4_0 | pp512 | 2731.93 \u00b1 62.70; llama 7B Q4_0 | tg128 | 47.55 \u00b1 2.16",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-16131738",
   "tg_tokens": 128,
   "tps": 47.55,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-14",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-disc-10879-c16131738-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3171.16,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GB10; dev=VULKAN0; llama 7B Q4_0 | pp512 | 3171.16 \u00b1 83.09; llama 7B Q4_0 | tg128 | 46.61 \u00b1 0.36",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-16131738",
   "tg_tokens": 128,
   "tps": 46.61,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-15",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-disc-10879-c16143155-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b08356",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2737.79,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GB10; dev=VULKAN0; llama 7B Q4_0 | pp512 | 2737.79 \u00b1 19.56; llama 7B Q4_0 | tg128 | 52.28 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-16143155",
   "tg_tokens": 128,
   "tps": 52.28,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-15",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-disc-10879-c16143155-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b08356",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3279.89,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GB10; dev=VULKAN0; llama 7B Q4_0 | pp512 | 3279.89 \u00b1 26.78; llama 7B Q4_0 | tg128 | 53.64 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-16143155",
   "tg_tokens": 128,
   "tps": 53.64,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-15",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-disc-10879-c16143310-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b08356",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2710.58,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GB10; dev=VULKAN0; llama 7B Q4_0 | pp512 | 2710.58 \u00b1 15.24; llama 7B Q4_0 | tg128 | 51.70 \u00b1 0.19",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-16143310",
   "tg_tokens": 128,
   "tps": 51.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-15",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-disc-10879-c16143310-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b08356",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3219.57,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GB10; dev=VULKAN0; llama 7B Q4_0 | pp512 | 3219.57 \u00b1 84.09; llama 7B Q4_0 | tg128 | 53.21 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-16143310",
   "tg_tokens": 128,
   "tps": 53.21,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-15",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-disc-10879-c16143310-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; config=b08356",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2692.06,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GB10; dev=VULKAN0; llama 7B Q4_0 | pp512 | 2692.06 \u00b1 32.45; llama 7B Q4_0 | tg128 | 48.42 \u00b1 2.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-16143310",
   "tg_tokens": 128,
   "tps": 48.42,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-15",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-disc-10879-c16143310-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b08356",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3137.97,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GB10; dev=VULKAN0; llama 7B Q4_0 | pp512 | 3137.97 \u00b1 30.23; llama 7B Q4_0 | tg128 | 46.67 \u00b1 0.16",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-16143310",
   "tg_tokens": 128,
   "tps": 46.67,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-21",
   "flags": [],
   "hardware": "Radeon RX 6600",
   "id": "lc-disc-10879-c16236948-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b08460",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 761.89,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6600; llama 7B Q4_0 | pp512 | 761.89 \u00b1 1.76; llama 7B Q4_0 | tg128 | 50.63 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-16236948",
   "tg_tokens": 128,
   "tps": 50.63,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-21",
   "flags": [],
   "hardware": "Radeon RX 6600",
   "id": "lc-disc-10879-c16236948-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b08460",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 808.76,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6600; llama 7B Q4_0 | pp512 | 808.76 \u00b1 0.15; llama 7B Q4_0 | tg128 | 53.24 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-16236948",
   "tg_tokens": 128,
   "tps": 53.24,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-07",
   "flags": [],
   "hardware": "Radeon RX 6800 XT",
   "id": "lc-disc-10879-c13064870-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=12; config=b05180",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 625.53,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Radeon RX 6800 XT (MoltenVK) | uma: 0 | fp16: 1 | warp size: 64 | shared mem; llama 7B Q4_0 | pp512 | 625.53 \u00b1 1.41; llama 7B Q4_0 | tg128 | 94.81 \u00b1 0.29",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13064870",
   "tg_tokens": 128,
   "tps": 94.81,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-07",
   "flags": [],
   "hardware": "Radeon RX 6800 XT",
   "id": "lc-disc-10879-c13064870-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=12; config=b05180",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 519.14,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6800 XT; llama 7B Q4_0 | pp512 | 519.14 \u00b1 0.13; llama 7B Q4_0 | tg128 | 87.56 \u00b1 0.19",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13064870",
   "tg_tokens": 128,
   "tps": 87.56,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-07",
   "flags": [],
   "hardware": "Radeon RX 6800 XT",
   "id": "lc-disc-10879-c13064870-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=12; config=b05180",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 522.39,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6800 XT; llama 7B Q4_0 | pp512 | 522.39 \u00b1 0.16; llama 7B Q4_0 | tg128 | 86.17 \u00b1 0.86",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13064870",
   "tg_tokens": 128,
   "tps": 86.17,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-07",
   "flags": [],
   "hardware": "Radeon RX 6800 XT",
   "id": "lc-disc-10879-c13064870-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=12; config=b05180",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 510.8,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6800 XT; llama 7B Q4_0 | pp512 | 510.80 \u00b1 0.13; llama 7B Q4_0 | tg128 | 86.47 \u00b1 0.46",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13064870",
   "tg_tokens": 128,
   "tps": 86.47,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-07",
   "flags": [],
   "hardware": "Radeon RX 6900 XT",
   "id": "lc-disc-10879-c13064870-llama7b-Q4_0-5",
   "model": "llama 7B",
   "notes": "threads=12; config=b05180",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 659.07,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Radeon RX 6900 XT (MoltenVK) | uma: 0 | fp16: 1 | warp size: 64 | shared mem; llama 7B Q4_0 | pp512 | 659.07 \u00b1 0.64; llama 7B Q4_0 | tg128 | 66.81 \u00b1 0.24",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13064870",
   "tg_tokens": 128,
   "tps": 66.81,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon RX 6800 XT",
   "id": "lc-disc-10879-c13064870-llama7b-Q4_0-6",
   "model": "llama 7B",
   "notes": "threads=12; config=b05180",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 570.06,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Radeon RX 6800 XT (MoltenVK) | uma: 0 | fp16: 1 | warp size: 64 | shared mem; llama 7B Q4_0 | pp512 | 570.06 \u00b1 0.64; llama 7B Q4_0 | tg128 | 78.30 \u00b1 0.41",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13064870",
   "tg_tokens": 128,
   "tps": 78.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon RX 6800 XT",
   "id": "lc-disc-10879-c13064870-llama7b-Q4_0-7",
   "model": "llama 7B",
   "notes": "threads=12; config=b05180",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 441.9,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6800 XT; llama 7B Q4_0 | pp512 | 441.90 \u00b1 0.34; llama 7B Q4_0 | tg128 | 79.35 \u00b1 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13064870",
   "tg_tokens": 128,
   "tps": 79.35,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon RX 6800 XT",
   "id": "lc-disc-10879-c13064870-llama7b-Q4_0-8",
   "model": "llama 7B",
   "notes": "threads=12; config=b05180",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 444.07,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6800 XT; llama 7B Q4_0 | pp512 | 444.07 \u00b1 0.63; llama 7B Q4_0 | tg128 | 79.60 \u00b1 0.35",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13064870",
   "tg_tokens": 128,
   "tps": 79.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-12",
   "flags": [],
   "hardware": "Radeon RX 6900 XT",
   "id": "lc-disc-10879-c13115818-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b04940",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1843.15,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6900 XT; llama 7B Q4_0 | pp512 | 1843.15 \u00b1 2.09; llama 7B Q4_0 | tg128 | 85.60 \u00b1 0.21",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13115818",
   "tg_tokens": 128,
   "tps": 85.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-12",
   "flags": [],
   "hardware": "Radeon RX 6900 XT",
   "id": "lc-disc-10879-c13115818-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; config=b04940",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1869.09,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6900 XT; llama 7B Q4_0 | pp512 | 1869.09 \u00b1 1.48; llama 7B Q4_0 | tg128 | 92.96 \u00b1 0.47",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13115818",
   "tg_tokens": 128,
   "tps": 92.96,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-12",
   "flags": [],
   "hardware": "Radeon RX 6900 XT",
   "id": "lc-disc-10879-c13115818-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; config=b04940",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1843.15,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Radeon RX 6900 XT, gfx1030 (0x1030), VMM: no, Wave Size: 32 \nDevice; llama 7B Q4_0 | pp512 | 1843.15 \u00b1 2.09; llama 7B Q4_0 | tg128 | 85.60 \u00b1 0.21",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13115818",
   "tg_tokens": 128,
   "tps": 85.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-12",
   "flags": [],
   "hardware": "Radeon RX 6900 XT",
   "id": "lc-disc-10879-c13115818-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 5 in this comment (separate llama-bench run; build/flags differ); config=b04940",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1615.12,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Radeon RX 6900 XT, gfx1030 (0x1030), VMM: no, Wave Size: 32 \nDevice; llama 7B Q4_0 | pp512 | 1615.12 \u00b1 0.86; llama 7B Q4_0 | tg128 | 77.84 \u00b1 0.29",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13115818",
   "tg_tokens": 128,
   "tps": 77.84,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-12",
   "flags": [],
   "hardware": "Radeon RX 6900 XT",
   "id": "lc-disc-10879-c13115818-llama7b-Q4_0-5",
   "model": "llama 7B",
   "notes": "threads=99; run 3 of 5 in this comment (separate llama-bench run; build/flags differ); config=b04940",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1962.01,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Radeon RX 6900 XT, gfx1030 (0x1030), VMM: no, Wave Size: 32 \nDevice; llama 7B Q4_0 | pp512 | 1962.01 \u00b1 3.56; llama 7B Q4_0 | tg128 | 85.42 \u00b1 0.32",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13115818",
   "tg_tokens": 128,
   "tps": 85.42,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-12",
   "flags": [],
   "hardware": "Radeon RX 6900 XT",
   "id": "lc-disc-10879-c13115818-llama7b-Q4_0-6",
   "model": "llama 7B",
   "notes": "threads=99; run 4 of 5 in this comment (separate llama-bench run; build/flags differ); config=b04940",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1736.24,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Radeon RX 6900 XT, gfx1030 (0x1030), VMM: no, Wave Size: 32 \nDevice; llama 7B Q4_0 | pp512 | 1736.24 \u00b1 1.09; llama 7B Q4_0 | tg128 | 75.24 \u00b1 0.14",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13115818",
   "tg_tokens": 128,
   "tps": 75.24,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-12",
   "flags": [],
   "hardware": "Radeon RX 6950 XT",
   "id": "lc-disc-10879-c13115818-llama7b-Q4_0-7",
   "model": "llama 7B",
   "notes": "threads=99; config=b04940",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1869.09,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Radeon RX 6950 XT, gfx1030 (0x1030), VMM: no, Wave Size: 32; llama 7B Q4_0 | pp512 | 1869.09 \u00b1 1.48; llama 7B Q4_0 | tg128 | 92.96 \u00b1 0.47",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13115818",
   "tg_tokens": 128,
   "tps": 92.96,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-12",
   "flags": [],
   "hardware": "Radeon RX 6950 XT",
   "id": "lc-disc-10879-c13115818-llama7b-Q4_0-8",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 4 in this comment (separate llama-bench run; build/flags differ); config=b04940",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1666.05,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Radeon RX 6950 XT, gfx1030 (0x1030), VMM: no, Wave Size: 32; llama 7B Q4_0 | pp512 | 1666.05 \u00b1 0.85; llama 7B Q4_0 | tg128 | 82.73 \u00b1 0.81",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13115818",
   "tg_tokens": 128,
   "tps": 82.73,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-12",
   "flags": [],
   "hardware": "Radeon RX 6950 XT",
   "id": "lc-disc-10879-c13115818-llama7b-Q4_0-9",
   "model": "llama 7B",
   "notes": "threads=99; run 3 of 4 in this comment (separate llama-bench run; build/flags differ); config=b04940",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1956.82,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Radeon RX 6950 XT, gfx1030 (0x1030), VMM: no, Wave Size: 32; llama 7B Q4_0 | pp512 | 1956.82 \u00b1 1.95; llama 7B Q4_0 | tg128 | 85.49 \u00b1 0.25",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13115818",
   "tg_tokens": 128,
   "tps": 85.49,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-12",
   "flags": [],
   "hardware": "Radeon RX 6950 XT",
   "id": "lc-disc-10879-c13115818-llama7b-Q4_0-10",
   "model": "llama 7B",
   "notes": "threads=99; run 4 of 4 in this comment (separate llama-bench run; build/flags differ); config=b04940",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1730.22,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Radeon RX 6950 XT, gfx1030 (0x1030), VMM: no, Wave Size: 32; llama 7B Q4_0 | pp512 | 1730.22 \u00b1 3.33; llama 7B Q4_0 | tg128 | 75.80 \u00b1 0.23",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13115818",
   "tg_tokens": 128,
   "tps": 75.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-04",
   "flags": [],
   "hardware": "Radeon RX 6500 XT",
   "id": "lc-disc-10879-c13033279-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 255.25,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6500 XT; llama 7B Q4_0 | pp512 | 255.25 \u00b1 0.35; llama 7B Q4_0 | tg128 | 27.81 \u00b1 0.10",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13033279",
   "tg_tokens": 128,
   "tps": 27.81,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-06",
   "flags": [],
   "hardware": "Radeon RX 7600 XT",
   "id": "lc-disc-10879-c13051834-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; ngl=100; config=b05273",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 494.8,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7600 XT; llama 7B Q4_0 | pp512 | 494.80 \u00b1 0.43; llama 7B Q4_0 | tg128 | 51.42 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13051834",
   "tg_tokens": 128,
   "tps": 51.42,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-10",
   "flags": [],
   "hardware": "Radeon RX 7600 XT",
   "id": "lc-disc-10879-c13101844-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b05339",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 632.88,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7600 XT; llama 7B Q4_0 | pp512 | 632.88 \u00b1 0.70; llama 7B Q4_0 | tg128 | 58.44 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13101844",
   "tg_tokens": 128,
   "tps": 58.44,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-10",
   "flags": [],
   "hardware": "Radeon RX 7600 XT",
   "id": "lc-disc-10879-c13101844-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; config=b05339",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 586.16,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7600 XT; llama 7B Q4_0 | pp512 | 586.16 \u00b1 2.43; llama 7B Q4_0 | tg128 | 59.02 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13101844",
   "tg_tokens": 128,
   "tps": 59.02,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-10",
   "flags": [],
   "hardware": "RTX 5070 Ti",
   "id": "lc-disc-10879-c13101529-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 5164.6,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 5070 Ti; llama 7B Q4_0 | pp512 | 5164.60 \u00b1 24.58; llama 7B Q4_0 | tg128 | 126.67 \u00b1 0.44",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13101529",
   "tg_tokens": 128,
   "tps": 126.67,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-10",
   "flags": [],
   "hardware": "RTX 5070 Ti",
   "id": "lc-disc-10879-c13101529-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 7802.37,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 5070 Ti; llama 7B Q4_0 | pp512 | 7802.37 \u00b1 26.52; llama 7B Q4_0 | tg128 | 168.20 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13101529",
   "tg_tokens": 128,
   "tps": 168.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-12",
   "flags": [],
   "hardware": "RTX 5070 Ti",
   "id": "lc-disc-10879-c13114398-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 6801.97,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 5070 Ti; llama 7B Q4_0 | pp512 | 6801.97 \u00b1 15.29; llama 7B Q4_0 | tg128 | 125.89 \u00b1 1.28",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13114398",
   "tg_tokens": 128,
   "tps": 125.89,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-12",
   "flags": [],
   "hardware": "Radeon RX 580 S",
   "id": "lc-disc-10879-c13121100-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b05359",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 258.03,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 580 S; llama 7B Q4_0 | pp512 | 258.03 \u00b1 0.71; llama 7B Q4_0 | tg128 | 39.32 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13121100",
   "tg_tokens": 128,
   "tps": 39.32,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-26",
   "flags": [],
   "hardware": "Radeon RX 580",
   "id": "lc-disc-10879-c15089097-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b05359",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 182.18,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 580; llama 7B Q4_0 | pp512 | 182.18 \u00b1 0.30; llama 7B Q4_0 | tg128 | 33.19 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15089097",
   "tg_tokens": 128,
   "tps": 33.19,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-14",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-disc-10879-c13139787-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1244.85,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7800 XT; llama 7B Q4_0 | pp512 | 1244.85 \u00b1 21.29; llama 7B Q4_0 | tg128 | 112.01 \u00b1 0.54",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13139787",
   "tg_tokens": 128,
   "tps": 112.01,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-14",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-disc-10879-c13139787-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1258.58,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7800 XT; llama 7B Q4_0 | pp512 | 1258.58 \u00b1 1.49; llama 7B Q4_0 | tg128 | 114.18 \u00b1 0.26",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13139787",
   "tg_tokens": 128,
   "tps": 114.18,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-disc-10879-c13139787-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2091.34,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7800 XT; llama 7B Q4_0 | pp512 | 2091.34 \u00b1 8.75; llama 7B Q4_0 | tg128 | 98.15 \u00b1 0.23",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13139787",
   "tg_tokens": 128,
   "tps": 98.15,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-disc-10879-c13139787-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1955.91,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7800 XT; llama 7B Q4_0 | pp512 | 1955.91 \u00b1 6.36; llama 7B Q4_0 | tg128 | 95.17 \u00b1 0.23",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13139787",
   "tg_tokens": 128,
   "tps": 95.17,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-14",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-disc-10879-c13139787-llama7b-Q4_0-5",
   "model": "llama 7B",
   "notes": "threads=100",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1531.85,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7800 XT; llama 7B Q4_0 | pp512 | 1531.85 \u00b1 1.46; llama 7B Q4_0 | tg128 | 111.97 \u00b1 0.37",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13139787",
   "tg_tokens": 128,
   "tps": 111.97,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-14",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-disc-10879-c13139787-llama7b-Q4_0-6",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1432.7,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7800 XT; llama 7B Q4_0 | pp512 | 1432.70 \u00b1 8.81; llama 7B Q4_0 | tg128 | 114.12 \u00b1 0.31",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13139787",
   "tg_tokens": 128,
   "tps": 114.12,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-disc-10879-c13139787-llama7b-Q4_0-7",
   "model": "llama 7B",
   "notes": "threads=100",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1487.1,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7800 XT; llama 7B Q4_0 | pp512 | 1487.10 \u00b1 1.74; llama 7B Q4_0 | tg128 | 98.13 \u00b1 0.32",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13139787",
   "tg_tokens": 128,
   "tps": 98.13,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-disc-10879-c13139787-llama7b-Q4_0-8",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 949.34,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7800 XT; llama 7B Q4_0 | pp512 | 949.34 \u00b1 3.92; llama 7B Q4_0 | tg128 | 95.20 \u00b1 0.32",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13139787",
   "tg_tokens": 128,
   "tps": 95.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-21",
   "flags": [],
   "hardware": "Radeon RX 7600 XT",
   "id": "lc-disc-10879-c13226854-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b05443",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 420.83,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7600 XT; llama 7B Q4_0 | pp512 | 420.83 \u00b1 57.28; llama 7B Q4_0 | tg128 | 48.31 \u00b1 0.57",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13226854",
   "tg_tokens": 128,
   "tps": 48.31,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-21",
   "flags": [],
   "hardware": "Radeon RX 7600 XT",
   "id": "lc-disc-10879-c13226854-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b05443",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 483.22,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7600 XT; llama 7B Q4_0 | pp512 | 483.22 \u00b1 16.43; llama 7B Q4_0 | tg128 | 50.63 \u00b1 0.25",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13226854",
   "tg_tokens": 128,
   "tps": 50.63,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-21",
   "flags": [],
   "hardware": "Radeon RX 7600 XT",
   "id": "lc-disc-10879-c13226854-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b05443",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 694.27,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7600 XT; llama 7B Q4_0 | pp512 | 694.27 \u00b1 0.63; llama 7B Q4_0 | tg128 | 46.98 \u00b1 0.61",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13226854",
   "tg_tokens": 128,
   "tps": 46.98,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-21",
   "flags": [],
   "hardware": "Radeon RX 7600 XT",
   "id": "lc-disc-10879-c13226854-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b05443",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 717.53,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7600 XT; llama 7B Q4_0 | pp512 | 717.53 \u00b1 12.97; llama 7B Q4_0 | tg128 | 47.94 \u00b1 0.52",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13226854",
   "tg_tokens": 128,
   "tps": 47.94,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-20",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c13206458-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; ngl=100; config=b05431",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2001.05,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2001.05 \u00b1 24.18; llama 7B Q4_0 | tg128 | 140.76 \u00b1 1.30",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13206458",
   "tg_tokens": 128,
   "tps": 140.76,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-20",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c13206458-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; config=b05431",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1986.43,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 1986.43 \u00b1 27.04; llama 7B Q4_0 | tg128 | 146.66 \u00b1 0.34",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13206458",
   "tg_tokens": 128,
   "tps": 146.66,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-20",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c13206458-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100; ngl=100; config=b05431",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3044.46,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3044.46 \u00b1 57.71; llama 7B Q4_0 | tg128 | 107.02 \u00b1 0.34",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13206458",
   "tg_tokens": 128,
   "tps": 107.02,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-21",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c13222968-gemma327b-Q6_K",
   "model": "gemma3 27B",
   "notes": "threads=48; config=b05431",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 628.07,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "Radeon RX 7900 XTX; gemma3 27B Q6_K | pp512 | 628.07 \u00b1 16.15; gemma3 27B Q6_K | tg128 | 6.81 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13222968",
   "tg_tokens": 128,
   "tps": 6.81,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-21",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c13222968-gemma327b-Q6_K-2",
   "model": "gemma3 27B",
   "notes": "threads=48; config=b05431",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 213.89,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "Radeon RX 7900 XTX; gemma3 27B Q6_K | pp512 | 213.89 \u00b1 12.86; gemma3 27B Q6_K | tg128 | 7.05 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13222968",
   "tg_tokens": 128,
   "tps": 7.05,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-21",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c13223671-gemma327b-Q4_0",
   "model": "gemma3 27B",
   "notes": "threads=48; config=b05431",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 500.3,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; gemma3 27B Q4_0 | pp512 | 500.30 \u00b1 1.23; gemma3 27B Q4_0 | tg128 | 35.25 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13223671",
   "tg_tokens": 128,
   "tps": 35.25,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-21",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c13223671-gemma327b-Q4_0-2",
   "model": "gemma3 27B",
   "notes": "threads=48; config=b05431",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1008.91,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; gemma3 27B Q4_0 | pp512 | 1008.91 \u00b1 3.17; gemma3 27B Q4_0 | tg128 | 32.73 \u00b1 0.53",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13223671",
   "tg_tokens": 128,
   "tps": 32.73,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-22",
   "flags": [],
   "hardware": "Arc B570",
   "id": "lc-disc-10879-c13235880-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; ngl=100; config=b05449",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 651.09,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc B570; llama 7B Q4_0 | pp512 | 651.09 \u00b1 0.16; llama 7B Q4_0 | tg128 | 31.44 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13235880",
   "tg_tokens": 128,
   "tps": 31.44,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-22",
   "flags": [],
   "hardware": "Arc B570",
   "id": "lc-disc-10879-c13235880-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; config=b05449",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 342.71,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc B570; llama 7B Q4_0 | pp512 | 342.71 \u00b1 0.07; llama 7B Q4_0 | tg128 | 30.88 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13235880",
   "tg_tokens": 128,
   "tps": 30.88,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-22",
   "flags": [],
   "hardware": "Arc B570",
   "id": "lc-disc-10879-c13235880-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100; ngl=100; config=b05449",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 356.09,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc B570; llama 7B Q4_0 | pp512 | 356.09 \u00b1 0.90; llama 7B Q4_0 | tg128 | 31.39 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13235880",
   "tg_tokens": 128,
   "tps": 31.39,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-22",
   "flags": [],
   "hardware": "Arc B570",
   "id": "lc-disc-10879-c13235880-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=100; config=b05449",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 226.82,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc B570; llama 7B Q4_0 | pp512 | 226.82 \u00b1 0.50; llama 7B Q4_0 | tg128 | 30.89 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13235880",
   "tg_tokens": 128,
   "tps": 30.89,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-18",
   "flags": [],
   "hardware": "Radeon RX 7900 GRE",
   "id": "lc-disc-10879-c15291338-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2336.31,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 GRE; llama 7B Q4_0 | pp512 | 2336.31 \u00b1 7.52; llama 7B Q4_0 | tg128 | 116.11 \u00b1 0.26",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15291338",
   "tg_tokens": 128,
   "tps": 116.11,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-18",
   "flags": [],
   "hardware": "Radeon RX 7900 GRE",
   "id": "lc-disc-10879-c15291338-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2251.37,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 GRE; llama 7B Q4_0 | pp512 | 2251.37 \u00b1 2.91; llama 7B Q4_0 | tg128 | 110.50 \u00b1 0.23",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15291338",
   "tg_tokens": 128,
   "tps": 110.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-18",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-10879-c15291338-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3493.01,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 3493.01 \u00b1 11.30; llama 7B Q4_0 | tg128 | 126.81 \u00b1 0.39",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15291338",
   "tg_tokens": 128,
   "tps": 126.81,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-18",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-10879-c15291338-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3465.31,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 3465.31 \u00b1 3.93; llama 7B Q4_0 | tg128 | 132.27 \u00b1 0.29",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15291338",
   "tg_tokens": 128,
   "tps": 132.27,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-21",
   "flags": [],
   "hardware": "Radeon RX 6600",
   "id": "lc-disc-10879-c15311225-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b07499",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 404.04,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6600; llama 7B Q4_0 | pp512 | 404.04 \u00b1 0.59; llama 7B Q4_0 | tg128 | 50.11 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15311225",
   "tg_tokens": 128,
   "tps": 50.11,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-21",
   "flags": [],
   "hardware": "Radeon RX 6600",
   "id": "lc-disc-10879-c15311225-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07499",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 410.19,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6600; llama 7B Q4_0 | pp512 | 410.19 \u00b1 0.33; llama 7B Q4_0 | tg128 | 52.47 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15311225",
   "tg_tokens": 128,
   "tps": 52.47,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-30",
   "flags": [],
   "hardware": "Radeon RX 9060 XT",
   "id": "lc-disc-10879-c15375022-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b07411",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 317.42,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9060 XT; llama 7B Q4_0 | pp512 | 317.42 \u00b1 3.12; llama 7B Q4_0 | tg128 | 38.31 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15375022",
   "tg_tokens": 128,
   "tps": 38.31,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-30",
   "flags": [],
   "hardware": "Radeon RX 9060 XT",
   "id": "lc-disc-10879-c15375022-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b07411",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 418.46,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9060 XT; llama 7B Q4_0 | pp512 | 418.46 \u00b1 1.36; llama 7B Q4_0 | tg128 | 54.84 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15375022",
   "tg_tokens": 128,
   "tps": 54.84,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-30",
   "flags": [],
   "hardware": "Radeon RX 9060 XT",
   "id": "lc-disc-10879-c15375022-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b07411",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1649.97,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9060 XT; llama 7B Q4_0 | pp512 | 1649.97 \u00b1 3.95; llama 7B Q4_0 | tg128 | 38.37 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15375022",
   "tg_tokens": 128,
   "tps": 38.37,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-30",
   "flags": [],
   "hardware": "Radeon RX 9060 XT",
   "id": "lc-disc-10879-c15375022-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b07411",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1651.6,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9060 XT; llama 7B Q4_0 | pp512 | 1651.60 \u00b1 2.00; llama 7B Q4_0 | tg128 | 54.96 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15375022",
   "tg_tokens": 128,
   "tps": 54.96,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-01-08",
   "flags": [],
   "hardware": "A4000",
   "id": "lc-disc-10879-c15441098-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b07664",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2523.51,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A4000; llama 7B Q4_0 | pp512 | 2523.51 \u00b1 28.10; llama 7B Q4_0 | tg128 | 78.12 \u00b1 0.11",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15441098",
   "tg_tokens": 128,
   "tps": 78.12,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-01-08",
   "flags": [],
   "hardware": "A4000",
   "id": "lc-disc-10879-c15441098-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07664",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2835.78,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A4000; llama 7B Q4_0 | pp512 | 2835.78 \u00b1 18.53; llama 7B Q4_0 | tg128 | 80.39 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15441098",
   "tg_tokens": 128,
   "tps": 80.39,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-01-08",
   "flags": [],
   "hardware": "A4000",
   "id": "lc-disc-10879-c15441098-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100; config=b07664",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2248.14,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A4000; llama 7B Q4_0 | pp512 | 2248.14 \u00b1 7.59; llama 7B Q4_0 | tg128 | 73.74 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15441098",
   "tg_tokens": 128,
   "tps": 73.74,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-01-08",
   "flags": [],
   "hardware": "A4000",
   "id": "lc-disc-10879-c15441098-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07664",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2482.74,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A4000; llama 7B Q4_0 | pp512 | 2482.74 \u00b1 26.05; llama 7B Q4_0 | tg128 | 76.07 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15441098",
   "tg_tokens": 128,
   "tps": 76.07,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-01-16",
   "flags": [],
   "hardware": "RTX 4070 SUPER",
   "id": "lc-disc-10879-c15520807-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b07757",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4608.2,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 4070 Super; llama 7B Q4_0 | pp512 | 4608.20 \u00b1 31.66; llama 7B Q4_0 | tg128 | 108.74 \u00b1 0.18",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15520807",
   "tg_tokens": 128,
   "tps": 108.74,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-01-16",
   "flags": [],
   "hardware": "RTX 4070 SUPER",
   "id": "lc-disc-10879-c15520807-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07757",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 5198.91,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 4070 Super; llama 7B Q4_0 | pp512 | 5198.91 \u00b1 29.02; llama 7B Q4_0 | tg128 | 112.00 \u00b1 0.17",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15520807",
   "tg_tokens": 128,
   "tps": 112.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-01-17",
   "flags": [],
   "hardware": "Radeon 890M",
   "id": "lc-disc-10879-c15523099-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; 34 W measured nvtop power, 1.75 W per t/s (published range 30-35W); fa=0 ngl=100",
   "params": null,
   "power_w": 34,
   "pp_tokens": 512,
   "pp_tps": 278.49,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 890M; llama 7B Q4_0 | pp512 | 278.49 \u00b1 7.53; llama 7B Q4_0 | tg128 | 19.16 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15523099",
   "tg_tokens": 128,
   "tps": 19.16,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-01-17",
   "flags": [],
   "hardware": "Radeon 890M",
   "id": "lc-disc-10879-c15523099-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 291.42,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 890M; llama 7B Q4_0 | pp512 | 291.42 \u00b1 1.43; llama 7B Q4_0 | tg128 | 20.00 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15523099",
   "tg_tokens": 128,
   "tps": 20.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-01-17",
   "flags": [],
   "hardware": "RTX 4070 Ti",
   "id": "lc-disc-10879-c15523099-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100; 221 W measured nvtop power, 2 W per t/s (published range 220-230W)",
   "params": null,
   "power_w": 221,
   "pp_tokens": 512,
   "pp_tps": 4981.44,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "NVIDIA GeForce RTX 4070 Ti (NVIDIA) | uma: 0 | fp16: 1 | bf16: 1 | warp size: 32; llama 7B Q4_0 | pp512 | 4981.44 \u00b1 102.35; llama 7B Q4_0 | tg128 | 110.53 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15523099",
   "tg_tokens": 128,
   "tps": 110.53,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-01-17",
   "flags": [],
   "hardware": "RTX 4070 Ti",
   "id": "lc-disc-10879-c15523099-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 5515.26,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "NVIDIA GeForce RTX 4070 Ti (NVIDIA) | uma: 0 | fp16: 1 | bf16: 1 | warp size: 32; llama 7B Q4_0 | pp512 | 5515.26 \u00b1 7.92; llama 7B Q4_0 | tg128 | 113.52 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15523099",
   "tg_tokens": 128,
   "tps": 113.52,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-01-17",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-10879-c15523099-llama7b-Q4_0-5",
   "model": "llama 7B",
   "notes": "threads=100; 126 W measured nvtop power, 1.84 W per t/s (published range 120-130W)",
   "params": null,
   "power_w": 126,
   "pp_tokens": 512,
   "pp_tps": 522.36,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B60; Intel(R) Arc(tm) Pro B60 Graphics (BMG G21) (Intel open-source Mesa driver) | um; llama 7B Q4_0 | pp512 | 522.36 \u00b1 3.60; llama 7B Q4_0 | tg128 | 68.55 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15523099",
   "tg_tokens": 128,
   "tps": 68.55,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-01-17",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-10879-c15523099-llama7b-Q4_0-6",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 274.76,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B60; Intel(R) Arc(tm) Pro B60 Graphics (BMG G21) (Intel open-source Mesa driver) | um; llama 7B Q4_0 | pp512 | 274.76 \u00b1 0.27; llama 7B Q4_0 | tg128 | 70.54 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15523099",
   "tg_tokens": 128,
   "tps": 70.54,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-01-17",
   "flags": [],
   "hardware": "Radeon 890M",
   "id": "lc-disc-10879-c15523182-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 437.1,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 890M; llama 7B Q4_0 | pp512 | 437.10 \u00b1 4.52; llama 7B Q4_0 | tg128 | 18.02 \u00b1 0.18",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15523182",
   "tg_tokens": 128,
   "tps": 18.02,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-01-17",
   "flags": [],
   "hardware": "Radeon 890M",
   "id": "lc-disc-10879-c15523182-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 531.03,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 890M; llama 7B Q4_0 | pp512 | 531.03 \u00b1 3.34; llama 7B Q4_0 | tg128 | 19.06 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15523182",
   "tg_tokens": 128,
   "tps": 19.06,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-01-17",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-10879-c15523182-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100; 152.0 W measured power draw",
   "params": null,
   "power_w": 152.0,
   "pp_tokens": 512,
   "pp_tps": 652.15,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B60; llama 7B Q4_0 | pp512 | 652.15 \u00b1 1.20; llama 7B Q4_0 | tg128 | 77.65 \u00b1 0.11",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15523182",
   "tg_tokens": 128,
   "tps": 77.65,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-01-17",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-10879-c15523182-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 399.3,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B60; llama 7B Q4_0 | pp512 | 399.30 \u00b1 1.18; llama 7B Q4_0 | tg128 | 43.93 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15523182",
   "tg_tokens": 128,
   "tps": 43.93,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-01-28",
   "flags": [
    "contradiction"
   ],
   "hardware": "A770",
   "id": "lc-disc-10879-c15631342-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100; config=b06529",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 916.88,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A770; llama 7B Q4_0 | pp512 | 916.88 + 1.21; llama 7B Q4_0 | tg128 | 52.76 + 0.78",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15631342",
   "tg_tokens": 128,
   "tps": 52.76,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-01-28",
   "flags": [
    "contradiction"
   ],
   "hardware": "A770",
   "id": "lc-disc-10879-c15631342-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b06529",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 437.75,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A770; llama 7B Q4_0 | pp512 | 437.75 + 3.16; llama 7B Q4_0 | tg128 | 16.57 + 0.10",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15631342",
   "tg_tokens": 128,
   "tps": 16.57,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-01-28",
   "flags": [
    "contradiction"
   ],
   "hardware": "A770",
   "id": "lc-disc-10879-c15631342-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100; config=b06529",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 845.83,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A770; llama 7B Q4_0 | pp512 | 845.83 + 1.87; llama 7B Q4_0 | tg128 | 38.10 + 1.46",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15631342",
   "tg_tokens": 128,
   "tps": 38.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-01-28",
   "flags": [
    "contradiction"
   ],
   "hardware": "A770",
   "id": "lc-disc-10879-c15631342-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b06529",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 497.1,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A770; llama 7B Q4_0 | pp512 | 497.10 + 1.11; llama 7B Q4_0 | tg128 | 11.14 + 0.28",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15631342",
   "tg_tokens": 128,
   "tps": 11.14,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-01-28",
   "flags": [],
   "hardware": "A770",
   "id": "lc-disc-10879-c15631342-llama7b-Q4_0-5",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100; config=b06529",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2498.81,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A770; llama 7B Q4_0 | pp512 | 2498.81 + 45.71; llama 7B Q4_0 | tg128 | 70.54 + 0.28",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15631342",
   "tg_tokens": 128,
   "tps": 70.54,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-01-28",
   "flags": [
    "outlier"
   ],
   "hardware": "A770",
   "id": "lc-disc-10879-c15631342-llama7b-Q4_0-6",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b06529",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2512.42,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A770; llama 7B Q4_0 | pp512 | 2512.42 + 15.24; llama 7B Q4_0 | tg128 | 69.73 + 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15631342",
   "tg_tokens": 128,
   "tps": 69.73,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-01-17",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-disc-10879-c15524098-nemotronhmoe31ba35b-Q4_K - Medium",
   "model": "nemotron_h_moe 31B.A3.5B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1661.48,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "GB10; nemotron_h_moe 31B.A3.5B Q4_K - Medium | pp512 | 1661.48 \u00b1 12.92; nemotron_h_moe 31B.A3.5B Q4_K - Medium | tg128 | 64.50 \u00b1 1.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15524098",
   "tg_tokens": 128,
   "tps": 64.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-01-17",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-disc-10879-c15524098-nemotronhmoe31ba35b-Q4_K - Medium-2",
   "model": "nemotron_h_moe 31B.A3.5B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2135.98,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "GB10; nemotron_h_moe 31B.A3.5B Q4_K - Medium | pp512 | 2135.98 \u00b1 9.54; nemotron_h_moe 31B.A3.5B Q4_K - Medium | tg128 | 61.76 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15524098",
   "tg_tokens": 128,
   "tps": 61.76,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-01-20",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-disc-10879-c15544966-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2665.61,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GB10; llama 7B Q4_0 | pp512 | 2665.61 \u00b1 19.61; llama 7B Q4_0 | tg128 | 50.80 \u00b1 1.27",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15544966",
   "tg_tokens": 128,
   "tps": 50.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-01-20",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-disc-10879-c15544966-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3310.48,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GB10; llama 7B Q4_0 | pp512 | 3310.48 \u00b1 23.73; llama 7B Q4_0 | tg128 | 60.24 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15544966",
   "tg_tokens": 128,
   "tps": 60.24,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": null,
   "ctx": null,
   "date": "2026-01-20",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-disc-10879-c15544966-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=20",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 313.01,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GB10; llama 7B Q4_0 | pp512 | 313.01 \u00b1 0.64; llama 7B Q4_0 | tg128 | 24.08 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15544966",
   "tg_tokens": 128,
   "tps": 24.08,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-01-24",
   "flags": [],
   "hardware": "RTX 5080",
   "id": "lc-disc-10879-c15594232-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b07823",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 6840.43,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 5080; llama 7B Q4_0 | pp512 | 6840.43 \u00b1 15.77; llama 7B Q4_0 | tg128 | 183.10 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15594232",
   "tg_tokens": 128,
   "tps": 183.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-01-24",
   "flags": [],
   "hardware": "RTX 5080",
   "id": "lc-disc-10879-c15594232-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07823",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 7022.52,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 5080; llama 7B Q4_0 | pp512 | 7022.52 \u00b1 28.53; llama 7B Q4_0 | tg128 | 190.78 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15594232",
   "tg_tokens": 128,
   "tps": 190.78,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-01-24",
   "flags": [],
   "hardware": "RTX 5080",
   "id": "lc-disc-10879-c15594232-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100; config=b07823",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 6840.43,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 5080; llama 7B Q4_0 | pp512 | 6840.43 \u00b1 15.77; llama 7B Q4_0 | tg128 | 183.10 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15594232",
   "tg_tokens": 128,
   "tps": 183.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-01-24",
   "flags": [],
   "hardware": "RTX 5080",
   "id": "lc-disc-10879-c15594232-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07823",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 7022.52,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 5080; llama 7B Q4_0 | pp512 | 7022.52 \u00b1 28.53; llama 7B Q4_0 | tg128 | 190.78 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15594232",
   "tg_tokens": 128,
   "tps": 190.78,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-01-31",
   "flags": [],
   "hardware": "RTX 5080",
   "id": "lc-disc-10879-c15659601-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b07865",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 7444.99,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 5080; llama 7B Q4_0 | pp512 | 7444.99 \u00b1 20.11; llama 7B Q4_0 | tg128 | 185.10 \u00b1 0.54",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15659601",
   "tg_tokens": 128,
   "tps": 185.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-01-31",
   "flags": [],
   "hardware": "RTX 5080",
   "id": "lc-disc-10879-c15659601-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07865",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 8054.59,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 5080; llama 7B Q4_0 | pp512 | 8054.59 \u00b1 35.68; llama 7B Q4_0 | tg128 | 192.17 \u00b1 0.21",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15659601",
   "tg_tokens": 128,
   "tps": 192.17,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-01-31",
   "flags": [],
   "hardware": "RTX 5080",
   "id": "lc-disc-10879-c15659601-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100; config=b07865",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 7444.99,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 5080; llama 7B Q4_0 | pp512 | 7444.99 \u00b1 20.11; llama 7B Q4_0 | tg128 | 185.10 \u00b1 0.54",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15659601",
   "tg_tokens": 128,
   "tps": 185.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-01-31",
   "flags": [],
   "hardware": "RTX 5080",
   "id": "lc-disc-10879-c15659601-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07865",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 8054.59,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 5080; llama 7B Q4_0 | pp512 | 8054.59 \u00b1 35.68; llama 7B Q4_0 | tg128 | 192.17 \u00b1 0.21",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15659601",
   "tg_tokens": 128,
   "tps": 192.17,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-01-24",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-10879-c15594467-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b07825",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 5036.04,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 5036.04 \u00b1 88.16; llama 7B Q4_0 | tg128 | 137.11 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15594467",
   "tg_tokens": 128,
   "tps": 137.11,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-01-24",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-10879-c15594467-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07825",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 5048.07,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 5048.07 \u00b1 2.12; llama 7B Q4_0 | tg128 | 131.54 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15594467",
   "tg_tokens": 128,
   "tps": 131.54,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-22",
   "flags": [],
   "hardware": "A100",
   "id": "lc-disc-10879-c13236852-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b05454",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4865.98,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A100; llama 7B Q4_0 | pp512 | 4865.98 \u00b1 11.74; llama 7B Q4_0 | tg128 | 168.23 \u00b1 0.25",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13236852",
   "tg_tokens": 128,
   "tps": 168.23,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-22",
   "flags": [],
   "hardware": "A100",
   "id": "lc-disc-10879-c13236852-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b05454",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 5141.45,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A100; llama 7B Q4_0 | pp512 | 5141.45 \u00b1 8.01; llama 7B Q4_0 | tg128 | 176.88 \u00b1 0.11",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13236852",
   "tg_tokens": 128,
   "tps": 176.88,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-22",
   "flags": [],
   "hardware": "A100",
   "id": "lc-disc-10879-c13236852-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; config=b05454",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3103.32,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A100; llama 7B Q4_0 | pp512 | 3103.32 \u00b1 4.21; llama 7B Q4_0 | tg128 | 121.83 \u00b1 0.54",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13236852",
   "tg_tokens": 128,
   "tps": 121.83,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-22",
   "flags": [],
   "hardware": "A100",
   "id": "lc-disc-10879-c13236852-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b05454",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3164.55,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A100; llama 7B Q4_0 | pp512 | 3164.55 \u00b1 5.00; llama 7B Q4_0 | tg128 | 120.53 \u00b1 0.41",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13236852",
   "tg_tokens": 128,
   "tps": 120.53,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-23",
   "flags": [],
   "hardware": "RTX 5070 Ti",
   "id": "lc-disc-10879-c13250309-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b05468",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 6213.63,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 5070 Ti; llama 7B Q4_0 | pp512 | 6213.63 \u00b1 27.72; llama 7B Q4_0 | tg128 | 135.63 \u00b1 0.18",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13250309",
   "tg_tokens": 128,
   "tps": 135.63,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-23",
   "flags": [],
   "hardware": "RTX 5070 Ti",
   "id": "lc-disc-10879-c13250309-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b05468",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 6614.86,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 5070 Ti; llama 7B Q4_0 | pp512 | 6614.86 \u00b1 8.32; llama 7B Q4_0 | tg128 | 133.94 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13250309",
   "tg_tokens": 128,
   "tps": 133.94,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-23",
   "flags": [],
   "hardware": "RTX 5070 Ti",
   "id": "lc-disc-10879-c13250309-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; config=b05468",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 6870.08,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 5070 Ti; llama 7B Q4_0 | pp512 | 6870.08 \u00b1 29.31; llama 7B Q4_0 | tg128 | 149.28 \u00b1 0.16",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13250309",
   "tg_tokens": 128,
   "tps": 149.28,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-23",
   "flags": [],
   "hardware": "RTX 5070 Ti",
   "id": "lc-disc-10879-c13250309-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b05468",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 7623.55,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 5070 Ti; llama 7B Q4_0 | pp512 | 7623.55 \u00b1 21.43; llama 7B Q4_0 | tg128 | 153.55 \u00b1 0.11",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13250309",
   "tg_tokens": 128,
   "tps": 153.55,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-26",
   "flags": [],
   "hardware": "A770",
   "id": "lc-disc-10879-c13268629-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b05474",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 725.31,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A770; llama 7B Q4_0 | pp512 | 725.31 + 0.98; llama 7B Q4_0 | tg128 | 49.43 + 1.45",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13268629",
   "tg_tokens": 128,
   "tps": 49.43,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-26",
   "flags": [],
   "hardware": "A770",
   "id": "lc-disc-10879-c13268629-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b05474",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 327.58,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A770; llama 7B Q4_0 | pp512 | 327.58 + 0.19; llama 7B Q4_0 | tg128 | 48.17 + 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13268629",
   "tg_tokens": 128,
   "tps": 48.17,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-26",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-10879-c13274730-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b05498",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2336.07,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 2336.07 \u00b1 5.92; llama 7B Q4_0 | tg128 | 117.72 \u00b1 0.34",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13274730",
   "tg_tokens": 128,
   "tps": 117.72,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-26",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-10879-c13274730-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b05498",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2296.83,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 2296.83 \u00b1 5.22; llama 7B Q4_0 | tg128 | 120.46 \u00b1 0.29",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13274730",
   "tg_tokens": 128,
   "tps": 120.46,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-26",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-10879-c13274730-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100; config=b05498",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2728.17,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 2728.17 \u00b1 68.71; llama 7B Q4_0 | tg128 | 84.34 \u00b1 1.18",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13274730",
   "tg_tokens": 128,
   "tps": 84.34,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-26",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-10879-c13274730-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b05498",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1841.07,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 1841.07 \u00b1 35.88; llama 7B Q4_0 | tg128 | 80.81 \u00b1 0.20",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13274730",
   "tg_tokens": 128,
   "tps": 80.81,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-26",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-10879-c13274730-llama7b-Q4_0-5",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100; config=b05498",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2738.43,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 2738.43 \u00b1 74.42; llama 7B Q4_0 | tg128 | 85.12 \u00b1 0.20",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13274730",
   "tg_tokens": 128,
   "tps": 85.12,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-26",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-10879-c13274730-llama7b-Q4_0-6",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b05498",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2909.86,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 2909.86 \u00b1 4.60; llama 7B Q4_0 | tg128 | 86.10 \u00b1 0.29",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13274730",
   "tg_tokens": 128,
   "tps": 86.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-26",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c13274882-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; ngl=100",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3419.13,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3419.13 \u00b1 29.51; llama 7B Q4_0 | tg128 | 144.90 \u00b1 0.92",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13274882",
   "tg_tokens": 128,
   "tps": 144.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-26",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c13274882-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3599.57,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3599.57 \u00b1 25.62; llama 7B Q4_0 | tg128 | 153.72 \u00b1 0.23",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13274882",
   "tg_tokens": 128,
   "tps": 153.72,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-26",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c13274882-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3743.15,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3743.15 \u00b1 45.40; llama 7B Q4_0 | tg128 | 143.98 \u00b1 0.55",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13274882",
   "tg_tokens": 128,
   "tps": 143.98,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-13",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c13462814-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; fa=1 ngl=100",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3187.61,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3187.61 \u00b1 38.34; llama 7B Q4_0 | tg128 | 141.52 \u00b1 0.10",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13462814",
   "tg_tokens": 128,
   "tps": 141.52,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-23",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c13552822-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=1; config=b05735",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1952.7,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 1952.70 \u00b1 18.25; llama 7B Q4_0 | tg128 | 145.78 \u00b1 0.35",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13552822",
   "tg_tokens": 128,
   "tps": 145.78,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-23",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c13552822-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b05735",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1943.62,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 1943.62 \u00b1 12.29; llama 7B Q4_0 | tg128 | 147.25 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13552822",
   "tg_tokens": 128,
   "tps": 147.25,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-23",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c13554210-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=1; config=b05735",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3110.71,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3110.71 \u00b1 56.52; llama 7B Q4_0 | tg128 | 133.42 \u00b1 5.22",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13554210",
   "tg_tokens": 128,
   "tps": 133.42,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-23",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c13554210-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b05735",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2925.56,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2925.56 \u00b1 31.93; llama 7B Q4_0 | tg128 | 130.90 \u00b1 0.17",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13554210",
   "tg_tokens": 128,
   "tps": 130.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-23",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c13554210-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=1; config=b05735",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3236.05,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3236.05 \u00b1 50.51; llama 7B Q4_0 | tg128 | 121.87 \u00b1 0.10",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13554210",
   "tg_tokens": 128,
   "tps": 121.87,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-23",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c13554210-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b05735",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3409.8,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3409.80 \u00b1 31.75; llama 7B Q4_0 | tg128 | 117.61 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13554210",
   "tg_tokens": 128,
   "tps": 117.61,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-29",
   "flags": [],
   "hardware": "RTX 3070",
   "id": "lc-disc-10879-c13306463-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b05529",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2113.02,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3070; llama 7B Q4_0 | pp512 | 2113.02 \u00b1 7.38; llama 7B Q4_0 | tg128 | 78.71 \u00b1 0.13",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13306463",
   "tg_tokens": 128,
   "tps": 78.71,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-05-29",
   "flags": [],
   "hardware": "RTX 3070",
   "id": "lc-disc-10879-c13306463-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; config=b05529",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2993.73,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3070; llama 7B Q4_0 | pp512 | 2993.73 \u00b1 10.90; llama 7B Q4_0 | tg128 | 84.62 \u00b1 0.10",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13306463",
   "tg_tokens": 128,
   "tps": 84.62,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-02",
   "flags": [],
   "hardware": "Radeon RX 580 S",
   "id": "lc-disc-10879-c13344355-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b05536",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 222.98,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 580 S; llama 7B Q4_0 | pp512 | 222.98 \u00b1 2.54; llama 7B Q4_0 | tg128 | 38.20 \u00b1 0.14",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13344355",
   "tg_tokens": 128,
   "tps": 38.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-02",
   "flags": [],
   "hardware": "Radeon RX 580 S",
   "id": "lc-disc-10879-c13344355-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b05536",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 212.32,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 580 S; llama 7B Q4_0 | pp512 | 212.32 \u00b1 1.64; llama 7B Q4_0 | tg128 | 27.51 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13344355",
   "tg_tokens": 128,
   "tps": 27.51,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-04",
   "flags": [],
   "hardware": "Radeon RX 7900 XT",
   "id": "lc-disc-10879-c13363850-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b05581",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2941.58,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XT; llama 7B Q4_0 | pp512 | 2941.58 \u00b1 17.17; llama 7B Q4_0 | tg128 | 123.18 \u00b1 0.40",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13363850",
   "tg_tokens": 128,
   "tps": 123.18,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-04",
   "flags": [],
   "hardware": "Radeon RX 7900 XT",
   "id": "lc-disc-10879-c13363850-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b05581",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2701.13,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XT; llama 7B Q4_0 | pp512 | 2701.13 \u00b1 8.75; llama 7B Q4_0 | tg128 | 120.62 \u00b1 0.36",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13363850",
   "tg_tokens": 128,
   "tps": 120.62,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-04",
   "flags": [],
   "hardware": "Radeon RX 7900 XT",
   "id": "lc-disc-10879-c13363850-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100; config=b05581",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3219.16,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XT; llama 7B Q4_0 | pp512 | 3219.16 \u00b1 21.01; llama 7B Q4_0 | tg128 | 122.86 \u00b1 0.58",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13363850",
   "tg_tokens": 128,
   "tps": 122.86,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-04",
   "flags": [],
   "hardware": "Radeon RX 7900 XT",
   "id": "lc-disc-10879-c13363850-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b05581",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3267.35,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XT; llama 7B Q4_0 | pp512 | 3267.35 \u00b1 12.45; llama 7B Q4_0 | tg128 | 117.11 \u00b1 0.61",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13363850",
   "tg_tokens": 128,
   "tps": 117.11,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-10",
   "flags": [],
   "hardware": "Radeon RX 6650 XT",
   "id": "lc-disc-10879-c13420099-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b05604",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 735.64,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6650 XT; llama 7B Q4_0 | pp512 | 735.64 \u00b1 3.12; llama 7B Q4_0 | tg128 | 59.22 \u00b1 0.11",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13420099",
   "tg_tokens": 128,
   "tps": 59.22,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-10",
   "flags": [],
   "hardware": "Radeon RX 6650 XT",
   "id": "lc-disc-10879-c13420099-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b05604",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 730.64,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6650 XT; llama 7B Q4_0 | pp512 | 730.64 \u00b1 0.25; llama 7B Q4_0 | tg128 | 59.18 \u00b1 0.44",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13420099",
   "tg_tokens": 128,
   "tps": 59.18,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-10",
   "flags": [],
   "hardware": "Radeon RX 6750 XT",
   "id": "lc-disc-10879-c13420099-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100; config=b05604",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1040.58,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6750 XT; llama 7B Q4_0 | pp512 | 1040.58 \u00b1 0.35; llama 7B Q4_0 | tg128 | 81.98 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13420099",
   "tg_tokens": 128,
   "tps": 81.98,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-10",
   "flags": [],
   "hardware": "Radeon RX 6750 XT",
   "id": "lc-disc-10879-c13420099-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b05604",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 997.05,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6750 XT; llama 7B Q4_0 | pp512 | 997.05 \u00b1 0.45; llama 7B Q4_0 | tg128 | 82.29 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13420099",
   "tg_tokens": 128,
   "tps": 82.29,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-26",
   "flags": [],
   "hardware": "RTX 3060",
   "id": "lc-disc-10879-c15089502-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 main_gpu=1 ngl=99 sm=none; config=b07118",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 439.87,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3060; llama 7B Q4_0 | pp512 | 439.87 \u00b1 0.61; llama 7B Q4_0 | tg128 | 50.23 \u00b1 0.14",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15089502",
   "tg_tokens": 128,
   "tps": 50.23,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-26",
   "flags": [],
   "hardware": "RTX 3060",
   "id": "lc-disc-10879-c15089502-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 main_gpu=1 ngl=99 sm=none; config=b07118",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 428.54,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3060; llama 7B Q4_0 | pp512 | 428.54 \u00b1 0.50; llama 7B Q4_0 | tg128 | 52.66 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15089502",
   "tg_tokens": 128,
   "tps": 52.66,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-26",
   "flags": [],
   "hardware": "M2",
   "id": "lc-disc-10879-c15089707-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b07166",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 53.1,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M2; llama 7B Q4_0 | pp512 | 53.10 \u00b1 0.02; llama 7B Q4_0 | tg128 | 58.47 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15089707",
   "tg_tokens": 128,
   "tps": 58.47,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-26",
   "flags": [],
   "hardware": "M2",
   "id": "lc-disc-10879-c15089707-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b07166",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 53.16,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M2; llama 7B Q4_0 | pp512 | 53.16 \u00b1 0.02; llama 7B Q4_0 | tg128 | 62.75 \u00b1 0.10",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15089707",
   "tg_tokens": 128,
   "tps": 62.75,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-26",
   "flags": [],
   "hardware": "M2",
   "id": "lc-disc-10879-c15089707-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b07166",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 678.14,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M2; llama 7B Q4_0 | pp512 | 678.14 \u00b1 1.40; llama 7B Q4_0 | tg128 | 63.16 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15089707",
   "tg_tokens": 128,
   "tps": 63.16,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-26",
   "flags": [],
   "hardware": "M2",
   "id": "lc-disc-10879-c15089707-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b07166",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 685.51,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M2; llama 7B Q4_0 | pp512 | 685.51 \u00b1 0.88; llama 7B Q4_0 | tg128 | 66.48 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15089707",
   "tg_tokens": 128,
   "tps": 66.48,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-28",
   "flags": [],
   "hardware": "M2",
   "id": "lc-disc-10879-c15107399-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b07166",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 675.02,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M2; llama 7B Q4_0 | pp512 | 675.02 \u00b1 1.28; llama 7B Q4_0 | tg128 | 63.10 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15107399",
   "tg_tokens": 128,
   "tps": 63.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-28",
   "flags": [],
   "hardware": "M2",
   "id": "lc-disc-10879-c15107399-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b07166",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 683.61,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M2; llama 7B Q4_0 | pp512 | 683.61 \u00b1 0.32; llama 7B Q4_0 | tg128 | 66.32 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15107399",
   "tg_tokens": 128,
   "tps": 66.32,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-29",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-10879-c15110576-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b07193",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4262.85,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 4262.85 \u00b1 63.19; llama 7B Q4_0 | tg128 | 137.31 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15110576",
   "tg_tokens": 128,
   "tps": 137.31,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-29",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-10879-c15110576-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07193",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4273.68,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 4273.68 \u00b1 9.55; llama 7B Q4_0 | tg128 | 131.99 \u00b1 0.14",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15110576",
   "tg_tokens": 128,
   "tps": 131.99,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-03",
   "flags": [],
   "hardware": "RTX 3060",
   "id": "lc-disc-10879-c15148900-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=8; config=b07243",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 74.29,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3060; llama 7B Q4_0 | pp512 | 74.29 \u00b1 0.42; llama 7B Q4_0 | tg128 | 9.48 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15148900",
   "tg_tokens": 128,
   "tps": 9.48,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-03",
   "flags": [],
   "hardware": "RTX 3060",
   "id": "lc-disc-10879-c15148900-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=8; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07243",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 72.76,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3060; llama 7B Q4_0 | pp512 | 72.76 \u00b1 0.37; llama 7B Q4_0 | tg128 | 9.79 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15148900",
   "tg_tokens": 128,
   "tps": 9.79,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-03",
   "flags": [],
   "hardware": "RTX 3060",
   "id": "lc-disc-10879-c15148900-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100; config=b07243",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 90.13,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3060; llama 7B Q4_0 | pp512 | 90.13 \u00b1 0.04; llama 7B Q4_0 | tg128 | 7.33 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15148900",
   "tg_tokens": 128,
   "tps": 7.33,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-03",
   "flags": [],
   "hardware": "RTX 3060",
   "id": "lc-disc-10879-c15148900-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b07243",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 91.04,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3060; llama 7B Q4_0 | pp512 | 91.04 \u00b1 0.04; llama 7B Q4_0 | tg128 | 7.69 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15148900",
   "tg_tokens": 128,
   "tps": 7.69,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-03",
   "flags": [],
   "hardware": "RTX 3060",
   "id": "lc-disc-10879-c15148900-llama7b-Q4_0-5",
   "model": "llama 7B",
   "notes": "threads=100; config=b07243",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1122.04,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3060; llama 7B Q4_0 | pp512 | 1122.04 \u00b1 138.44; llama 7B Q4_0 | tg128 | 68.34 \u00b1 3.53",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15148900",
   "tg_tokens": 128,
   "tps": 68.34,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-03",
   "flags": [],
   "hardware": "RTX 3060",
   "id": "lc-disc-10879-c15148900-llama7b-Q4_0-6",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07243",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1685.14,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3060; llama 7B Q4_0 | pp512 | 1685.14 \u00b1 6.90; llama 7B Q4_0 | tg128 | 73.71 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15148900",
   "tg_tokens": 128,
   "tps": 73.71,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-06",
   "flags": [],
   "hardware": "Arc B50",
   "id": "lc-disc-10879-c15180826-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=999; config=b07166",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 193.5,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc B50; llama 7B Q4_0 | pp512 | 193.50 \u00b1 0.24; llama 7B Q4_0 | tg128 | 39.99 \u00b1 0.10",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15180826",
   "tg_tokens": 128,
   "tps": 39.99,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-06",
   "flags": [],
   "hardware": "Arc B50",
   "id": "lc-disc-10879-c15180826-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=999; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07166",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 132.48,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc B50; llama 7B Q4_0 | pp512 | 132.48 \u00b1 0.04; llama 7B Q4_0 | tg128 | 41.02 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15180826",
   "tg_tokens": 128,
   "tps": 41.02,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-13",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c16118289-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b07302",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3290.92,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3290.92 \u00b1 37.67; llama 7B Q4_0 | tg128 | 172.86 \u00b1 0.23",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-16118289",
   "tg_tokens": 128,
   "tps": 172.86,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c16118289-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; config=b07302",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3229.07,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3229.07 \u00b1 53.81; llama 7B Q4_0 | tg128 | 150.97 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-16118289",
   "tg_tokens": 128,
   "tps": 150.97,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c16118289-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100; config=b07302",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2546.13,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2546.13 \u00b1 18.25; llama 7B Q4_0 | tg128 | 129.80 \u00b1 0.34",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-16118289",
   "tg_tokens": 128,
   "tps": 129.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15234456-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b07361",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2973.55,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2973.55 \u00b1 75.06; llama 7B Q4_0 | tg128 | 161.48 \u00b1 0.11",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234456",
   "tg_tokens": 128,
   "tps": 161.48,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15234456-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 8 in this comment (separate llama-bench run; build/flags differ); config=b07361",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2994.51,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2994.51 \u00b1 19.13; llama 7B Q4_0 | tg128 | 169.38 \u00b1 0.19",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234456",
   "tg_tokens": 128,
   "tps": 169.38,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15234456-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; run 3 of 8 in this comment (separate llama-bench run; build/flags differ); config=b07361",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3069.84,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3069.84 \u00b1 33.22; llama 7B Q4_0 | tg128 | 161.39 \u00b1 0.10",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234456",
   "tg_tokens": 128,
   "tps": 161.39,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15234456-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 4 of 8 in this comment (separate llama-bench run; build/flags differ); config=b07361",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3138.42,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3138.42 \u00b1 26.38; llama 7B Q4_0 | tg128 | 169.06 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234456",
   "tg_tokens": 128,
   "tps": 169.06,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15234456-llama7b-Q4_0-5",
   "model": "llama 7B",
   "notes": "threads=99; run 5 of 8 in this comment (separate llama-bench run; build/flags differ); config=b07361",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3056.15,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3056.15 \u00b1 35.64; llama 7B Q4_0 | tg128 | 160.77 \u00b1 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234456",
   "tg_tokens": 128,
   "tps": 160.77,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15234456-llama7b-Q4_0-6",
   "model": "llama 7B",
   "notes": "threads=99; run 6 of 8 in this comment (separate llama-bench run; build/flags differ); config=b07361",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3128.43,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3128.43 \u00b1 33.55; llama 7B Q4_0 | tg128 | 168.63 \u00b1 0.19",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234456",
   "tg_tokens": 128,
   "tps": 168.63,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15234456-llama7b-Q4_0-7",
   "model": "llama 7B",
   "notes": "threads=99; run 7 of 8 in this comment (separate llama-bench run; build/flags differ); config=b07361",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3051.47,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3051.47 \u00b1 16.96; llama 7B Q4_0 | tg128 | 160.78 \u00b1 0.19",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234456",
   "tg_tokens": 128,
   "tps": 160.78,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15234456-llama7b-Q4_0-8",
   "model": "llama 7B",
   "notes": "threads=99; run 8 of 8 in this comment (separate llama-bench run; build/flags differ); config=b07361",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3121.91,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3121.91 \u00b1 30.95; llama 7B Q4_0 | tg128 | 168.77 \u00b1 0.12",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234456",
   "tg_tokens": 128,
   "tps": 168.77,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15234456-llama7b-Q4_0-9",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b07361",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3143.31,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3143.31 \u00b1 76.22; llama 7B Q4_0 | tg128 | 165.94 \u00b1 0.18",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234456",
   "tg_tokens": 128,
   "tps": 165.94,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15234456-llama7b-Q4_0-10",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 8 in this comment (separate llama-bench run; build/flags differ); config=b07361",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3153.9,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3153.90 \u00b1 14.02; llama 7B Q4_0 | tg128 | 173.81 \u00b1 0.12",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234456",
   "tg_tokens": 128,
   "tps": 173.81,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15234456-llama7b-Q4_0-11",
   "model": "llama 7B",
   "notes": "threads=99; run 3 of 8 in this comment (separate llama-bench run; build/flags differ); config=b07361",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3170.06,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3170.06 \u00b1 37.51; llama 7B Q4_0 | tg128 | 165.57 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234456",
   "tg_tokens": 128,
   "tps": 165.57,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15234456-llama7b-Q4_0-12",
   "model": "llama 7B",
   "notes": "threads=99; run 4 of 8 in this comment (separate llama-bench run; build/flags differ); config=b07361",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3259.34,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3259.34 \u00b1 33.04; llama 7B Q4_0 | tg128 | 173.73 \u00b1 0.10",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234456",
   "tg_tokens": 128,
   "tps": 173.73,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15234456-llama7b-Q4_0-13",
   "model": "llama 7B",
   "notes": "threads=99; run 5 of 8 in this comment (separate llama-bench run; build/flags differ); config=b07361",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3170.2,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3170.20 \u00b1 34.63; llama 7B Q4_0 | tg128 | 165.38 \u00b1 0.13",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234456",
   "tg_tokens": 128,
   "tps": 165.38,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15234456-llama7b-Q4_0-14",
   "model": "llama 7B",
   "notes": "threads=99; run 6 of 8 in this comment (separate llama-bench run; build/flags differ); config=b07361",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3253.68,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3253.68 \u00b1 44.61; llama 7B Q4_0 | tg128 | 173.36 \u00b1 0.12",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234456",
   "tg_tokens": 128,
   "tps": 173.36,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15234456-llama7b-Q4_0-15",
   "model": "llama 7B",
   "notes": "threads=99; run 7 of 8 in this comment (separate llama-bench run; build/flags differ); config=b07361",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3165.77,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3165.77 \u00b1 32.11; llama 7B Q4_0 | tg128 | 165.23 \u00b1 0.22",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234456",
   "tg_tokens": 128,
   "tps": 165.23,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15234456-llama7b-Q4_0-16",
   "model": "llama 7B",
   "notes": "threads=99; run 8 of 8 in this comment (separate llama-bench run; build/flags differ); config=b07361",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3250.86,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3250.86 \u00b1 48.41; llama 7B Q4_0 | tg128 | 173.13 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234456",
   "tg_tokens": 128,
   "tps": 173.13,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15234456-llama7b-Q4_0-17",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b07361",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2078.19,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2078.19 \u00b1 4.95; llama 7B Q4_0 | tg128 | 121.57 \u00b1 0.25",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234456",
   "tg_tokens": 128,
   "tps": 121.57,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15234456-llama7b-Q4_0-18",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 8 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b07361",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2130.63,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2130.63 \u00b1 9.31; llama 7B Q4_0 | tg128 | 129.07 \u00b1 0.24",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234456",
   "tg_tokens": 128,
   "tps": 129.07,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15234456-llama7b-Q4_0-19",
   "model": "llama 7B",
   "notes": "threads=99; run 3 of 8 in this comment (separate llama-bench run; build/flags differ); fa=0 ngl=99; config=b07361",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2521.5,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2521.50 \u00b1 10.56; llama 7B Q4_0 | tg128 | 121.33 \u00b1 0.10",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234456",
   "tg_tokens": 128,
   "tps": 121.33,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15234456-llama7b-Q4_0-20",
   "model": "llama 7B",
   "notes": "threads=99; run 4 of 8 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b07361",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2590.43,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2590.43 \u00b1 5.81; llama 7B Q4_0 | tg128 | 129.07 \u00b1 0.11",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234456",
   "tg_tokens": 128,
   "tps": 129.07,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15234456-llama7b-Q4_0-21",
   "model": "llama 7B",
   "notes": "threads=99; run 5 of 8 in this comment (separate llama-bench run; build/flags differ); fa=0 ngl=99; config=b07361",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2524.02,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2524.02 \u00b1 13.16; llama 7B Q4_0 | tg128 | 121.61 \u00b1 0.12",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234456",
   "tg_tokens": 128,
   "tps": 121.61,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15234456-llama7b-Q4_0-22",
   "model": "llama 7B",
   "notes": "threads=99; run 6 of 8 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b07361",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2581.45,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2581.45 \u00b1 8.67; llama 7B Q4_0 | tg128 | 128.94 \u00b1 0.15",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234456",
   "tg_tokens": 128,
   "tps": 128.94,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15234456-llama7b-Q4_0-23",
   "model": "llama 7B",
   "notes": "threads=99; run 7 of 8 in this comment (separate llama-bench run; build/flags differ); fa=0 ngl=99; config=b07361",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2512.95,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2512.95 \u00b1 11.18; llama 7B Q4_0 | tg128 | 121.50 \u00b1 0.16",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234456",
   "tg_tokens": 128,
   "tps": 121.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15234456-llama7b-Q4_0-24",
   "model": "llama 7B",
   "notes": "threads=99; run 8 of 8 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b07361",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2580.6,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2580.60 \u00b1 13.18; llama 7B Q4_0 | tg128 | 128.95 \u00b1 0.20",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234456",
   "tg_tokens": 128,
   "tps": 128.95,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15234456-llama7b-Q4_0-25",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b07361",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2653.13,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2653.13 \u00b1 35.39; llama 7B Q4_0 | tg128 | 160.77 \u00b1 0.35",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234456",
   "tg_tokens": 128,
   "tps": 160.77,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15234456-llama7b-Q4_0-26",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 8 in this comment (separate llama-bench run; build/flags differ); config=b07361",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2622.48,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2622.48 \u00b1 24.42; llama 7B Q4_0 | tg128 | 168.74 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234456",
   "tg_tokens": 128,
   "tps": 168.74,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15234456-llama7b-Q4_0-27",
   "model": "llama 7B",
   "notes": "threads=99; run 3 of 8 in this comment (separate llama-bench run; build/flags differ); config=b07361",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2739.28,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2739.28 \u00b1 17.91; llama 7B Q4_0 | tg128 | 160.20 \u00b1 0.11",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234456",
   "tg_tokens": 128,
   "tps": 160.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15234456-llama7b-Q4_0-28",
   "model": "llama 7B",
   "notes": "threads=99; run 4 of 8 in this comment (separate llama-bench run; build/flags differ); config=b07361",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2727.7,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2727.70 \u00b1 11.74; llama 7B Q4_0 | tg128 | 167.78 \u00b1 0.44",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234456",
   "tg_tokens": 128,
   "tps": 167.78,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15234456-llama7b-Q4_0-29",
   "model": "llama 7B",
   "notes": "threads=99; run 5 of 8 in this comment (separate llama-bench run; build/flags differ); config=b07361",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2726.36,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2726.36 \u00b1 23.92; llama 7B Q4_0 | tg128 | 159.78 \u00b1 0.34",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234456",
   "tg_tokens": 128,
   "tps": 159.78,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15234456-llama7b-Q4_0-30",
   "model": "llama 7B",
   "notes": "threads=99; run 6 of 8 in this comment (separate llama-bench run; build/flags differ); config=b07361",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2716.65,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2716.65 \u00b1 12.42; llama 7B Q4_0 | tg128 | 167.79 \u00b1 0.52",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234456",
   "tg_tokens": 128,
   "tps": 167.79,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15234456-llama7b-Q4_0-31",
   "model": "llama 7B",
   "notes": "threads=99; run 7 of 8 in this comment (separate llama-bench run; build/flags differ); config=b07361",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2727.9,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2727.90 \u00b1 9.12; llama 7B Q4_0 | tg128 | 159.86 \u00b1 0.12",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234456",
   "tg_tokens": 128,
   "tps": 159.86,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15234456-llama7b-Q4_0-32",
   "model": "llama 7B",
   "notes": "threads=99; run 8 of 8 in this comment (separate llama-bench run; build/flags differ); config=b07361",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2714.98,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2714.98 \u00b1 18.80; llama 7B Q4_0 | tg128 | 168.04 \u00b1 0.23",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234456",
   "tg_tokens": 128,
   "tps": 168.04,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15234456-llama7b-Q4_0-33",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b07361",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2750.74,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2750.74 \u00b1 41.21; llama 7B Q4_0 | tg128 | 164.84 \u00b1 0.13",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234456",
   "tg_tokens": 128,
   "tps": 164.84,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15234456-llama7b-Q4_0-34",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 8 in this comment (separate llama-bench run; build/flags differ); config=b07361",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2706.82,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2706.82 \u00b1 26.85; llama 7B Q4_0 | tg128 | 172.84 \u00b1 0.13",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234456",
   "tg_tokens": 128,
   "tps": 172.84,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15234456-llama7b-Q4_0-35",
   "model": "llama 7B",
   "notes": "threads=99; run 3 of 8 in this comment (separate llama-bench run; build/flags differ); config=b07361",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2805.52,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2805.52 \u00b1 27.45; llama 7B Q4_0 | tg128 | 164.32 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234456",
   "tg_tokens": 128,
   "tps": 164.32,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15234456-llama7b-Q4_0-36",
   "model": "llama 7B",
   "notes": "threads=99; run 4 of 8 in this comment (separate llama-bench run; build/flags differ); config=b07361",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2806.12,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2806.12 \u00b1 29.99; llama 7B Q4_0 | tg128 | 172.79 \u00b1 0.14",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234456",
   "tg_tokens": 128,
   "tps": 172.79,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15234456-llama7b-Q4_0-37",
   "model": "llama 7B",
   "notes": "threads=99; run 5 of 8 in this comment (separate llama-bench run; build/flags differ); config=b07361",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2814.04,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2814.04 \u00b1 25.41; llama 7B Q4_0 | tg128 | 164.22 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234456",
   "tg_tokens": 128,
   "tps": 164.22,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15234456-llama7b-Q4_0-38",
   "model": "llama 7B",
   "notes": "threads=99; run 6 of 8 in this comment (separate llama-bench run; build/flags differ); config=b07361",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2801.46,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2801.46 \u00b1 31.62; llama 7B Q4_0 | tg128 | 172.76 \u00b1 0.27",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234456",
   "tg_tokens": 128,
   "tps": 172.76,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15234456-llama7b-Q4_0-39",
   "model": "llama 7B",
   "notes": "threads=99; run 7 of 8 in this comment (separate llama-bench run; build/flags differ); config=b07361",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2807.29,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2807.29 \u00b1 28.35; llama 7B Q4_0 | tg128 | 164.21 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234456",
   "tg_tokens": 128,
   "tps": 164.21,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15234456-llama7b-Q4_0-40",
   "model": "llama 7B",
   "notes": "threads=99; run 8 of 8 in this comment (separate llama-bench run; build/flags differ); config=b07361",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2805.99,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2805.99 \u00b1 21.82; llama 7B Q4_0 | tg128 | 172.71 \u00b1 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234456",
   "tg_tokens": 128,
   "tps": 172.71,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15234456-llama7b-Q4_0-41",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b07361",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1869.34,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 1869.34 \u00b1 15.52; llama 7B Q4_0 | tg128 | 121.55 \u00b1 0.13",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234456",
   "tg_tokens": 128,
   "tps": 121.55,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15234456-llama7b-Q4_0-42",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 8 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b07361",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1878.23,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 1878.23 \u00b1 6.35; llama 7B Q4_0 | tg128 | 129.20 \u00b1 0.14",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234456",
   "tg_tokens": 128,
   "tps": 129.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15234456-llama7b-Q4_0-43",
   "model": "llama 7B",
   "notes": "threads=99; run 3 of 8 in this comment (separate llama-bench run; build/flags differ); fa=0 ngl=99; config=b07361",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2273.79,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2273.79 \u00b1 9.07; llama 7B Q4_0 | tg128 | 121.62 \u00b1 0.15",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234456",
   "tg_tokens": 128,
   "tps": 121.62,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15234456-llama7b-Q4_0-44",
   "model": "llama 7B",
   "notes": "threads=99; run 4 of 8 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b07361",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2278.5,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2278.50 \u00b1 6.08; llama 7B Q4_0 | tg128 | 129.15 \u00b1 0.28",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234456",
   "tg_tokens": 128,
   "tps": 129.15,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15234456-llama7b-Q4_0-45",
   "model": "llama 7B",
   "notes": "threads=99; run 5 of 8 in this comment (separate llama-bench run; build/flags differ); fa=0 ngl=99; config=b07361",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2276.59,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2276.59 \u00b1 8.25; llama 7B Q4_0 | tg128 | 121.56 \u00b1 0.32",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234456",
   "tg_tokens": 128,
   "tps": 121.56,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15234456-llama7b-Q4_0-46",
   "model": "llama 7B",
   "notes": "threads=99; run 6 of 8 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b07361",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2285.41,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2285.41 \u00b1 6.25; llama 7B Q4_0 | tg128 | 129.30 \u00b1 0.13",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234456",
   "tg_tokens": 128,
   "tps": 129.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15234456-llama7b-Q4_0-47",
   "model": "llama 7B",
   "notes": "threads=99; run 7 of 8 in this comment (separate llama-bench run; build/flags differ); fa=0 ngl=99; config=b07361",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2272.76,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2272.76 \u00b1 17.46; llama 7B Q4_0 | tg128 | 122.02 \u00b1 0.18",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234456",
   "tg_tokens": 128,
   "tps": 122.02,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15234456-llama7b-Q4_0-48",
   "model": "llama 7B",
   "notes": "threads=99; run 8 of 8 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b07361",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2281.29,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2281.29 \u00b1 2.43; llama 7B Q4_0 | tg128 | 129.48 \u00b1 0.11",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234456",
   "tg_tokens": 128,
   "tps": 129.48,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-10879-c15234450-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b07358",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1044.51,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; llama 7B Q4_0 | pp512 | 1044.51 \u00b1 6.43; llama 7B Q4_0 | tg128 | 52.23 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234450",
   "tg_tokens": 128,
   "tps": 52.23,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-10879-c15234450-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b07358",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1147.07,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; llama 7B Q4_0 | pp512 | 1147.07 \u00b1 4.02; llama 7B Q4_0 | tg128 | 55.32 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234450",
   "tg_tokens": 128,
   "tps": 55.32,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc-disc-10879-c15234836-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b07358",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3877.12,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3090; dev=VULKAN0; llama 7B Q4_0 | pp512 | 3877.12 \u00b1 561.58; llama 7B Q4_0 | tg128 | 143.46 \u00b1 0.31",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234836",
   "tg_tokens": 128,
   "tps": 143.46,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc-disc-10879-c15234836-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07358",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4356.94,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3090; dev=VULKAN0; llama 7B Q4_0 | pp512 | 4356.94 \u00b1 5.59; llama 7B Q4_0 | tg128 | 151.65 \u00b1 0.25",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234836",
   "tg_tokens": 128,
   "tps": 151.65,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc-disc-10879-c15234836-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; config=b07358",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4033.29,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3090; dev=VULKAN1; llama 7B Q4_0 | pp512 | 4033.29 \u00b1 340.56; llama 7B Q4_0 | tg128 | 142.76 \u00b1 0.27",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234836",
   "tg_tokens": 128,
   "tps": 142.76,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc-disc-10879-c15234836-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07358",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4435.33,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3090; dev=VULKAN1; llama 7B Q4_0 | pp512 | 4435.33 \u00b1 5.14; llama 7B Q4_0 | tg128 | 151.65 \u00b1 0.23",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15234836",
   "tg_tokens": 128,
   "tps": 151.65,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-14",
   "flags": [],
   "hardware": "Radeon RX 7900 GRE",
   "id": "lc-disc-10879-c15249350-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1711.33,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 GRE; dev=VULKAN0; llama 7B Q4_0 | pp512 | 1711.33 \u00b1 5.64; llama 7B Q4_0 | tg128 | 104.75 \u00b1 0.46",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15249350",
   "tg_tokens": 128,
   "tps": 104.75,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-14",
   "flags": [],
   "hardware": "Radeon RX 7900 GRE",
   "id": "lc-disc-10879-c15249350-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1760.15,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 GRE; dev=VULKAN0; llama 7B Q4_0 | pp512 | 1760.15 \u00b1 3.42; llama 7B Q4_0 | tg128 | 110.80 \u00b1 0.32",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15249350",
   "tg_tokens": 128,
   "tps": 110.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-14",
   "flags": [],
   "hardware": "Radeon RX 7900 GRE",
   "id": "lc-disc-10879-c15249350-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2411.47,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 GRE; dev=VULKAN1; llama 7B Q4_0 | pp512 | 2411.47 \u00b1 14.04; llama 7B Q4_0 | tg128 | 105.91 \u00b1 0.25",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15249350",
   "tg_tokens": 128,
   "tps": 105.91,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-14",
   "flags": [],
   "hardware": "Radeon RX 7900 GRE",
   "id": "lc-disc-10879-c15249350-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2372.49,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 GRE; dev=VULKAN1; llama 7B Q4_0 | pp512 | 2372.49 \u00b1 3.79; llama 7B Q4_0 | tg128 | 110.73 \u00b1 0.13",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15249350",
   "tg_tokens": 128,
   "tps": 110.73,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-14",
   "flags": [],
   "hardware": "Radeon 890M",
   "id": "lc-disc-10879-c15252134-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b07386",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 318.6,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 890M; llama 7B Q4_0 | pp512 | 318.60 \u00b1 14.15; llama 7B Q4_0 | tg128 | 22.03 \u00b1 0.11",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15252134",
   "tg_tokens": 128,
   "tps": 22.03,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-14",
   "flags": [],
   "hardware": "Radeon 890M",
   "id": "lc-disc-10879-c15252134-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07386",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 373.78,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 890M; llama 7B Q4_0 | pp512 | 373.78 \u00b1 2.87; llama 7B Q4_0 | tg128 | 23.10 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15252134",
   "tg_tokens": 128,
   "tps": 23.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-14",
   "flags": [],
   "hardware": "Radeon 890M",
   "id": "lc-disc-10879-c15252153-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b07386",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 329.76,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 890M; llama 7B Q4_0 | pp512 | 329.76 \u00b1 1.86; llama 7B Q4_0 | tg128 | 21.99 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15252153",
   "tg_tokens": 128,
   "tps": 21.99,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-14",
   "flags": [],
   "hardware": "Radeon 890M",
   "id": "lc-disc-10879-c15252153-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07386",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 374.88,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 890M; llama 7B Q4_0 | pp512 | 374.88 \u00b1 1.30; llama 7B Q4_0 | tg128 | 23.06 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15252153",
   "tg_tokens": 128,
   "tps": 23.06,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-14",
   "flags": [],
   "hardware": "Radeon 890M",
   "id": "lc-disc-10879-c15252153-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100; config=b07386",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 424.24,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 890M; llama 7B Q4_0 | pp512 | 424.24 \u00b1 2.08; llama 7B Q4_0 | tg128 | 21.18 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15252153",
   "tg_tokens": 128,
   "tps": 21.18,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-14",
   "flags": [],
   "hardware": "Radeon 890M",
   "id": "lc-disc-10879-c15252153-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07386",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 545.1,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 890M; llama 7B Q4_0 | pp512 | 545.10 \u00b1 5.25; llama 7B Q4_0 | tg128 | 22.37 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15252153",
   "tg_tokens": 128,
   "tps": 22.37,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-17",
   "flags": [],
   "hardware": "M4",
   "id": "lc-disc-10879-c15278221-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b07312",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 6389.86,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M4; llama 7B Q4_0 | pp512 | 6389.86 \u00b1 4.83; llama 7B Q4_0 | tg128 | 160.78 \u00b1 0.16",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15278221",
   "tg_tokens": 128,
   "tps": 160.78,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-17",
   "flags": [],
   "hardware": "M4",
   "id": "lc-disc-10879-c15278221-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07312",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 7064.4,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M4; llama 7B Q4_0 | pp512 | 7064.40 \u00b1 1.63; llama 7B Q4_0 | tg128 | 170.56 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15278221",
   "tg_tokens": 128,
   "tps": 170.56,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-13",
   "flags": [],
   "hardware": "Radeon RX 6600",
   "id": "lc-disc-10879-c13460098-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b05648",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 422.67,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6600; llama 7B Q4_0 | pp512 | 422.67 \u00b1 1.27; llama 7B Q4_0 | tg128 | 34.34 \u00b1 0.11",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13460098",
   "tg_tokens": 128,
   "tps": 34.34,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-13",
   "flags": [],
   "hardware": "Radeon RX 6600",
   "id": "lc-disc-10879-c13460098-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b05648",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 277.2,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6600; llama 7B Q4_0 | pp512 | 277.20 \u00b1 1.68; llama 7B Q4_0 | tg128 | 25.61 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13460098",
   "tg_tokens": 128,
   "tps": 25.61,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-13",
   "flags": [],
   "hardware": "Radeon RX 9060 XT",
   "id": "lc-disc-10879-c13460098-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b05648",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 658.95,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9060 XT; llama 7B Q4_0 | pp512 | 658.95 \u00b1 7.24; llama 7B Q4_0 | tg128 | 44.10 \u00b1 0.20",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13460098",
   "tg_tokens": 128,
   "tps": 44.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-13",
   "flags": [],
   "hardware": "Radeon RX 9060 XT",
   "id": "lc-disc-10879-c13460098-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b05648",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 389.52,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9060 XT; llama 7B Q4_0 | pp512 | 389.52 \u00b1 12.75; llama 7B Q4_0 | tg128 | 34.72 \u00b1 0.27",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13460098",
   "tg_tokens": 128,
   "tps": 34.72,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-13",
   "flags": [],
   "hardware": "Radeon RX 9060 XT",
   "id": "lc-disc-10879-c13460098-llama7b-Q4_0-5",
   "model": "llama 7B",
   "notes": "threads=99; config=b05648",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2141.67,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9060 XT; llama 7B Q4_0 | pp512 | 2141.67 \u00b1 6.87; llama 7B Q4_0 | tg128 | 70.54 \u00b1 0.74",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13460098",
   "tg_tokens": 128,
   "tps": 70.54,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-13",
   "flags": [],
   "hardware": "Radeon RX 9060 XT",
   "id": "lc-disc-10879-c13460098-llama7b-Q4_0-6",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b05648",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1915.41,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9060 XT; llama 7B Q4_0 | pp512 | 1915.41 \u00b1 7.90; llama 7B Q4_0 | tg128 | 70.52 \u00b1 0.16",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13460098",
   "tg_tokens": 128,
   "tps": 70.52,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-17",
   "flags": [],
   "hardware": "RTX 3060",
   "id": "lc-disc-10879-c13499036-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b05688",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1681.81,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3060; llama 7B Q4_0 | pp512 | 1681.81 \u00b1 7.45; llama 7B Q4_0 | tg128 | 64.76 \u00b1 3.20",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13499036",
   "tg_tokens": 128,
   "tps": 64.76,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-17",
   "flags": [],
   "hardware": "RTX 3060",
   "id": "lc-disc-10879-c13499036-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b05688",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1826.49,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3060; llama 7B Q4_0 | pp512 | 1826.49 \u00b1 3.02; llama 7B Q4_0 | tg128 | 67.11 \u00b1 0.12",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13499036",
   "tg_tokens": 128,
   "tps": 67.11,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-17",
   "flags": [],
   "hardware": "RTX 3060",
   "id": "lc-disc-10879-c13499036-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100; config=b05688",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2020.4,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3060; llama 7B Q4_0 | pp512 | 2020.40 \u00b1 3.75; llama 7B Q4_0 | tg128 | 68.38 \u00b1 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13499036",
   "tg_tokens": 128,
   "tps": 68.38,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-17",
   "flags": [],
   "hardware": "RTX 3060",
   "id": "lc-disc-10879-c13499036-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b05688",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2229.42,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3060; llama 7B Q4_0 | pp512 | 2229.42 \u00b1 2.98; llama 7B Q4_0 | tg128 | 70.78 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13499036",
   "tg_tokens": 128,
   "tps": 70.78,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-23",
   "flags": [],
   "hardware": "RTX 3060",
   "id": "lc-disc-10879-c14195645-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b06242",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1730.33,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3060; llama 7B Q4_0 | pp512 | 1730.33 \u00b1 1.68; llama 7B Q4_0 | tg128 | 65.63 \u00b1 2.47",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14195645",
   "tg_tokens": 128,
   "tps": 65.63,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-23",
   "flags": [],
   "hardware": "RTX 3060",
   "id": "lc-disc-10879-c14195645-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b06242",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1715.17,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3060; llama 7B Q4_0 | pp512 | 1715.17 \u00b1 21.13; llama 7B Q4_0 | tg128 | 66.12 \u00b1 1.74",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14195645",
   "tg_tokens": 128,
   "tps": 66.12,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-18",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc-disc-10879-c13503953-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 main_gpu=1 ngl=100 sm=none; config=b05688",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3336.74,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "NVIDIA GeForce RTX 3090 (NVIDIA) | uma: 0 | fp16: 1 | warp size: 32 | shared mem; llama 7B Q4_0 | pp512 | 3336.74 \u00b1 18.64; llama 7B Q4_0 | tg128 | 125.60 \u00b1 0.30",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13503953",
   "tg_tokens": 128,
   "tps": 125.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-18",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc-disc-10879-c13503953-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 main_gpu=1 ngl=100 sm=none; config=b05688",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3384.6,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "NVIDIA GeForce RTX 3090 (NVIDIA) | uma: 0 | fp16: 1 | warp size: 32 | shared mem; llama 7B Q4_0 | pp512 | 3384.60 \u00b1 8.18; llama 7B Q4_0 | tg128 | 126.06 \u00b1 0.46",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13503953",
   "tg_tokens": 128,
   "tps": 126.06,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-18",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc-disc-10879-c13503953-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100; config=b05688",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 186.24,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3090; llama 7B Q4_0 | pp512 | 186.24 \u00b1 0.09; llama 7B Q4_0 | tg128 | 33.90 \u00b1 0.51",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13503953",
   "tg_tokens": 128,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-18",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc-disc-10879-c13503953-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b05688",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 180.97,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3090; llama 7B Q4_0 | pp512 | 180.97 \u00b1 0.74; llama 7B Q4_0 | tg128 | 34.16 \u00b1 0.10",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13503953",
   "tg_tokens": 128,
   "tps": 34.16,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-18",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc-disc-10879-c13503953-llama7b-Q4_0-5",
   "model": "llama 7B",
   "notes": "threads=30; config=b05688",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 311.56,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "NVIDIA GeForce RTX 3090 (NVIDIA) | uma: 0 | fp16: 1 | warp size: 32 | shared mem; llama 7B Q4_0 | pp512 | 311.56 \u00b1 0.29; llama 7B Q4_0 | tg128 | 22.41 \u00b1 0.12",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13503953",
   "tg_tokens": 128,
   "tps": 22.41,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-18",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc-disc-10879-c13503953-llama7b-Q4_0-6",
   "model": "llama 7B",
   "notes": "threads=30; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b05688",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 307.71,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "NVIDIA GeForce RTX 3090 (NVIDIA) | uma: 0 | fp16: 1 | warp size: 32 | shared mem; llama 7B Q4_0 | pp512 | 307.71 \u00b1 0.25; llama 7B Q4_0 | tg128 | 22.80 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13503953",
   "tg_tokens": 128,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-18",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc-disc-10879-c13503953-llama7b-Q4_0-7",
   "model": "llama 7B",
   "notes": "threads=15; config=b05688",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 42.12,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3090; llama 7B Q4_0 | pp512 | 42.12 \u00b1 0.05; llama 7B Q4_0 | tg128 | 5.61 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13503953",
   "tg_tokens": 128,
   "tps": 5.61,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-18",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc-disc-10879-c13503953-llama7b-Q4_0-8",
   "model": "llama 7B",
   "notes": "threads=15; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b05688",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 41.26,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3090; llama 7B Q4_0 | pp512 | 41.26 \u00b1 0.10; llama 7B Q4_0 | tg128 | 5.54 \u00b1 0.10",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13503953",
   "tg_tokens": 128,
   "tps": 5.54,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-18",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc-disc-10879-c13503953-llama7b-Q4_0-9",
   "model": "llama 7B",
   "notes": "threads=100; config=b05688",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 315.39,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3090; llama 7B Q4_0 | pp512 | 315.39 \u00b1 3.76; llama 7B Q4_0 | tg128 | 36.82 \u00b1 0.38",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13503953",
   "tg_tokens": 128,
   "tps": 36.82,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "RTX 3090",
   "id": "lc-disc-10879-c13503953-llama7b-Q4_0-10",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b05688",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 117.48,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3090; llama 7B Q4_0 | pp512 | 117.48 \u00b1 0.50; llama 7B Q4_0 | tg128 | 8.55 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13503953",
   "tg_tokens": 128,
   "tps": 8.55,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "RTX 3090",
   "id": "lc-disc-10879-c13503953-llama7b-Q4_0-11",
   "model": "llama 7B",
   "notes": "threads=100; config=b05688",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 78.79,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3090; llama 7B Q4_0 | pp512 | 78.79 \u00b1 0.10; llama 7B Q4_0 | tg128 | 16.05 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13503953",
   "tg_tokens": 128,
   "tps": 16.05,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "RTX 3090",
   "id": "lc-disc-10879-c13503953-llama7b-Q4_0-12",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b05688",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 75.59,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3090; llama 7B Q4_0 | pp512 | 75.59 \u00b1 0.19; llama 7B Q4_0 | tg128 | 16.56 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13503953",
   "tg_tokens": 128,
   "tps": 16.56,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-18",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc-disc-10879-c13503953-llama7b-Q4_0-13",
   "model": "llama 7B",
   "notes": "threads=7; config=b05688",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 17.7,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3090; llama 7B Q4_0 | pp512 | 17.70 \u00b1 0.08; llama 7B Q4_0 | tg128 | 5.71 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13503953",
   "tg_tokens": 128,
   "tps": 5.71,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-18",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc-disc-10879-c13503953-llama7b-Q4_0-14",
   "model": "llama 7B",
   "notes": "threads=7; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b05688",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 16.62,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3090; llama 7B Q4_0 | pp512 | 16.62 \u00b1 0.07; llama 7B Q4_0 | tg128 | 5.82 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13503953",
   "tg_tokens": 128,
   "tps": 5.82,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-18",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc-disc-10879-c13503953-llama7b-Q4_0-15",
   "model": "llama 7B",
   "notes": "threads=7; config=b05688",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 19.18,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3090; llama 7B Q4_0 | pp512 | 19.18 \u00b1 0.01; llama 7B Q4_0 | tg128 | 5.35 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13503953",
   "tg_tokens": 128,
   "tps": 5.35,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-18",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc-disc-10879-c13503953-llama7b-Q4_0-16",
   "model": "llama 7B",
   "notes": "threads=7; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b05688",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 15.22,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3090; llama 7B Q4_0 | pp512 | 15.22 \u00b1 0.01; llama 7B Q4_0 | tg128 | 5.42 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13503953",
   "tg_tokens": 128,
   "tps": 5.42,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-18",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "RTX 3090",
   "id": "lc-disc-10879-c13503953-llama7b-Q4_0-17",
   "model": "llama 7B",
   "notes": "threads=100; config=b05688",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 81.59,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3090; llama 7B Q4_0 | pp512 | 81.59 \u00b1 0.30; llama 7B Q4_0 | tg128 | 10.06 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13503953",
   "tg_tokens": 128,
   "tps": 10.06,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "RTX 3090",
   "id": "lc-disc-10879-c13503953-llama7b-Q4_0-18",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b05688",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 76.53,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3090; llama 7B Q4_0 | pp512 | 76.53 \u00b1 0.12; llama 7B Q4_0 | tg128 | 10.09 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13503953",
   "tg_tokens": 128,
   "tps": 10.09,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "RTX 3090",
   "id": "lc-disc-10879-c13503953-llama7b-Q4_0-19",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 main_gpu=1 ngl=100 sm=none; config=b05688",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 107.8,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "NVIDIA GeForce RTX 3090 (NVIDIA) | uma: 0 | fp16: 1 | warp size: 32 | shared mem; llama 7B Q4_0 | pp512 | 107.80 \u00b1 0.36; llama 7B Q4_0 | tg128 | 2.85 \u00b1 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13503953",
   "tg_tokens": 128,
   "tps": 2.85,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "RTX 3090",
   "id": "lc-disc-10879-c13503953-llama7b-Q4_0-20",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 main_gpu=1 ngl=100 sm=none; config=b05688",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 84.19,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "NVIDIA GeForce RTX 3090 (NVIDIA) | uma: 0 | fp16: 1 | warp size: 32 | shared mem; llama 7B Q4_0 | pp512 | 84.19 \u00b1 3.31; llama 7B Q4_0 | tg128 | 2.87 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13503953",
   "tg_tokens": 128,
   "tps": 2.87,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-18",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc-disc-10879-c13503953-llama7b-Q4_0-21",
   "model": "llama 7B",
   "notes": "threads=31; config=b05688",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 23.19,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3090; llama 7B Q4_0 | pp512 | 23.19 \u00b1 0.20; llama 7B Q4_0 | tg128 | 2.63 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13503953",
   "tg_tokens": 128,
   "tps": 2.63,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-18",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc-disc-10879-c13503953-llama7b-Q4_0-22",
   "model": "llama 7B",
   "notes": "threads=31; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b05688",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 23.19,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3090; llama 7B Q4_0 | pp512 | 23.19 \u00b1 0.05; llama 7B Q4_0 | tg128 | 2.67 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13503953",
   "tg_tokens": 128,
   "tps": 2.67,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-23",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-10879-c13546617-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b05740",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2313.09,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 2313.09 \u00b1 62.74; llama 7B Q4_0 | tg128 | 121.25 \u00b1 0.30",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13546617",
   "tg_tokens": 128,
   "tps": 121.25,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-23",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-10879-c13546617-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b05740",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2307.03,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 2307.03 \u00b1 6.20; llama 7B Q4_0 | tg128 | 123.70 \u00b1 0.63",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13546617",
   "tg_tokens": 128,
   "tps": 123.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-27",
   "flags": [],
   "hardware": "Radeon RX 6600",
   "id": "lc-disc-10879-c13602256-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b05767",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 615.03,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6600; llama 7B Q4_0 | pp512 | 615.03 \u00b1 0.61; llama 7B Q4_0 | tg128 | 47.49 \u00b1 0.18",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13602256",
   "tg_tokens": 128,
   "tps": 47.49,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-28",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c13606581-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=1; config=b05774",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3085.75,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3085.75 \u00b1 39.20; llama 7B Q4_0 | tg128 | 132.75 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13606581",
   "tg_tokens": 128,
   "tps": 132.75,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-28",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c13606581-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b05774",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2871.79,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2871.79 \u00b1 8.96; llama 7B Q4_0 | tg128 | 128.79 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13606581",
   "tg_tokens": 128,
   "tps": 128.79,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-28",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c13606581-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=1; config=b05774",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3169.7,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3169.70 \u00b1 59.71; llama 7B Q4_0 | tg128 | 122.56 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13606581",
   "tg_tokens": 128,
   "tps": 122.56,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-28",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c13606581-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b05774",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3323.43,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3323.43 \u00b1 10.17; llama 7B Q4_0 | tg128 | 118.23 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13606581",
   "tg_tokens": 128,
   "tps": 118.23,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-29",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-10879-c13611122-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b05774",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3787.28,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 3787.28 \u00b1 38.03; llama 7B Q4_0 | tg128 | 124.75 \u00b1 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13611122",
   "tg_tokens": 128,
   "tps": 124.75,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-29",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-10879-c13611122-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b05774",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3399.14,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 3399.14 \u00b1 5.12; llama 7B Q4_0 | tg128 | 122.54 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13611122",
   "tg_tokens": 128,
   "tps": 122.54,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-29",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-10879-c13611122-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100; config=b05774",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1689.15,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 1689.15 \u00b1 12.26; llama 7B Q4_0 | tg128 | 116.82 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13611122",
   "tg_tokens": 128,
   "tps": 116.82,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-06-29",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-10879-c13611122-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b05774",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1707.58,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 1707.58 \u00b1 6.44; llama 7B Q4_0 | tg128 | 120.52 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13611122",
   "tg_tokens": 128,
   "tps": 120.52,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-08",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc-disc-10879-c13699168-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; ngl=99; config=b05846",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4543.96,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3090; llama 7B Q4_0 | pp512 | 4543.96 \u00b1 73.80; llama 7B Q4_0 | tg128 | 136.81 \u00b1 3.63",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13699168",
   "tg_tokens": 128,
   "tps": 136.81,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-08",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc-disc-10879-c13699168-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; config=b05846",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4946.47,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3090; llama 7B Q4_0 | pp512 | 4946.47 \u00b1 77.89; llama 7B Q4_0 | tg128 | 133.63 \u00b1 4.50",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13699168",
   "tg_tokens": 128,
   "tps": 133.63,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-08",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc-disc-10879-c13699168-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; config=b05846",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 8534.56,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 4090; llama 7B Q4_0 | pp512 | 8534.56 \u00b1 200.32; llama 7B Q4_0 | tg128 | 170.95 \u00b1 0.32",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13699168",
   "tg_tokens": 128,
   "tps": 170.95,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-08",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc-disc-10879-c13699168-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; config=b05846",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 9032.44,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 4090; llama 7B Q4_0 | pp512 | 9032.44 \u00b1 323.38; llama 7B Q4_0 | tg128 | 166.45 \u00b1 0.18",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13699168",
   "tg_tokens": 128,
   "tps": 166.45,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-08",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc-disc-10879-c13699168-llama7b-Q4_0-5",
   "model": "llama 7B",
   "notes": "threads=99; config=b05846",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 14472.84,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 4090; llama 7B Q4_0 | pp512 | 14472.84 \u00b1 57.16; llama 7B Q4_0 | tg128 | 185.04 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13699168",
   "tg_tokens": 128,
   "tps": 185.04,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-03",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc-disc-10879-c13973420-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b06075",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 14753.2,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 4090; llama 7B Q4_0 | pp512 | 14753.20 \u00b1 92.92; llama 7B Q4_0 | tg128 | 189.07 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13973420",
   "tg_tokens": 128,
   "tps": 189.07,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-03",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc-disc-10879-c13973420-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; config=b06075",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 9481.99,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 4090; llama 7B Q4_0 | pp512 | 9481.99 \u00b1 780.92; llama 7B Q4_0 | tg128 | 179.62 \u00b1 0.27",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13973420",
   "tg_tokens": 128,
   "tps": 179.62,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-12",
   "flags": [],
   "hardware": "T4",
   "id": "lc-disc-10879-c13738518-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b05873",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 734.28,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "T4; llama 7B Q4_0 | pp512 | 734.28 \u00b1 2.66; llama 7B Q4_0 | tg128 | 34.66 \u00b1 0.51",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13738518",
   "tg_tokens": 128,
   "tps": 34.66,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-12",
   "flags": [],
   "hardware": "T4",
   "id": "lc-disc-10879-c13738518-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; config=b05873",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 708.19,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "T4; llama 7B Q4_0 | pp512 | 708.19 \u00b1 3.22; llama 7B Q4_0 | tg128 | 33.36 \u00b1 0.67",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13738518",
   "tg_tokens": 128,
   "tps": 33.36,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-16",
   "flags": [],
   "hardware": "Radeon RX 550",
   "id": "lc-disc-10879-c14984005-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 52.66,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 550; llama 7B Q4_0 | pp512 | 52.66 \u00b1 0.49; llama 7B Q4_0 | tg128 | 10.20 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14984005",
   "tg_tokens": 128,
   "tps": 10.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-18",
   "flags": [],
   "hardware": "RTX 3060",
   "id": "lc-disc-10879-c14995287-llama8b-Q4_K - Medium",
   "model": "llama 8B",
   "notes": "threads=99; ngl=99; config=b07087",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1061.32,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "RTX 3060; dev=VULKAN0; llama 8B Q4_K - Medium | pp512 | 1061.32 \u00b1 2.07; llama 8B Q4_K - Medium | tg128 | 9.06 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14995287",
   "tg_tokens": 128,
   "tps": 9.06,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-18",
   "flags": [],
   "hardware": "RTX 3060",
   "id": "lc-disc-10879-c14995287-llama8b-Q4_K - Medium-2",
   "model": "llama 8B",
   "notes": "threads=99; config=b07087",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 230.26,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "RTX 3060; dev=VULKAN1; llama 8B Q4_K - Medium | pp512 | 230.26 \u00b1 1.56; llama 8B Q4_K - Medium | tg128 | 12.11 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14995287",
   "tg_tokens": 128,
   "tps": 12.11,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-18",
   "flags": [],
   "hardware": "RTX 3060",
   "id": "lc-disc-10879-c14995287-llama7b-Q4_K - Medium",
   "model": "llama 7B",
   "notes": "threads=99; ngl=99; config=b07087",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1081.37,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "RTX 3060; dev=VULKAN0; llama 7B Q4_K - Medium | pp512 | 1081.37 \u00b1 2.29; llama 7B Q4_K - Medium | tg128 | 9.40 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14995287",
   "tg_tokens": 128,
   "tps": 9.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-18",
   "flags": [],
   "hardware": "RTX 3060",
   "id": "lc-disc-10879-c14995287-llama7b-Q4_K - Medium-2",
   "model": "llama 7B",
   "notes": "threads=99; config=b07087",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 218.01,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "RTX 3060; dev=VULKAN1; llama 7B Q4_K - Medium | pp512 | 218.01 \u00b1 1.39; llama 7B Q4_K - Medium | tg128 | 11.80 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14995287",
   "tg_tokens": 128,
   "tps": 11.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-18",
   "flags": [],
   "hardware": "RTX 3060",
   "id": "lc-disc-10879-c14995287-gptoss20bmxfp4moe-unknown",
   "model": "gpt-oss 20B MXFP4 MoE",
   "notes": "threads=10; config=b07087",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 184.9,
   "provenance": "community",
   "quant": "as-published",
   "quote": "RTX 3060; dev=VULKAN0; gpt-oss 20B MXFP4 MoE | pp512 | 184.90 \u00b1 2.73; gpt-oss 20B MXFP4 MoE | tg128 | 16.18 \u00b1 0.33",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14995287",
   "tg_tokens": 128,
   "tps": 16.18,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-18",
   "flags": [],
   "hardware": "RTX 3060",
   "id": "lc-disc-10879-c14995287-gptoss20bmxfp4moe-unknown-2",
   "model": "gpt-oss 20B MXFP4 MoE",
   "notes": "threads=99; config=b07087",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 324.21,
   "provenance": "community",
   "quant": "as-published",
   "quote": "RTX 3060; dev=VULKAN1; gpt-oss 20B MXFP4 MoE | pp512 | 324.21 \u00b1 2.15; gpt-oss 20B MXFP4 MoE | tg128 | 18.65 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14995287",
   "tg_tokens": 128,
   "tps": 18.65,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-18",
   "flags": [],
   "hardware": "RTX 3060",
   "id": "lc-disc-10879-c14995287-qwen3moe30ba3b-Q8_0",
   "model": "qwen3moe 30B.A3B",
   "notes": "threads=7; config=b07087",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 86.32,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "RTX 3060; dev=VULKAN0; qwen3moe 30B.A3B Q8_0 | pp512 | 86.32 \u00b1 4.69; qwen3moe 30B.A3B Q8_0 | tg128 | 8.69 \u00b1 0.15",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14995287",
   "tg_tokens": 128,
   "tps": 8.69,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-18",
   "flags": [],
   "hardware": "RTX 3060",
   "id": "lc-disc-10879-c14995287-qwen3moe30ba3b-Q8_0-2",
   "model": "qwen3moe 30B.A3B",
   "notes": "threads=99; config=b07087",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 171.43,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "RTX 3060; dev=VULKAN1; qwen3moe 30B.A3B Q8_0 | pp512 | 171.43 \u00b1 32.20; qwen3moe 30B.A3B Q8_0 | tg128 | 15.52 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14995287",
   "tg_tokens": 128,
   "tps": 15.52,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-18",
   "flags": [],
   "hardware": "RTX 3060",
   "id": "lc-disc-10879-c14995287-qwen3moe30ba3b-Q4_K - Medium",
   "model": "qwen3moe 30B.A3B",
   "notes": "threads=13; config=b07087",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 113.21,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "RTX 3060; dev=VULKAN0; qwen3moe 30B.A3B Q4_K - Medium | pp512 | 113.21 \u00b1 2.87; qwen3moe 30B.A3B Q4_K - Medium | tg128 | 15.47 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14995287",
   "tg_tokens": 128,
   "tps": 15.47,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-18",
   "flags": [],
   "hardware": "RTX 3060",
   "id": "lc-disc-10879-c14995287-qwen3moe30ba3b-Q4_K - Medium-2",
   "model": "qwen3moe 30B.A3B",
   "notes": "threads=99; config=b07087",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 212.78,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "RTX 3060; dev=VULKAN1; qwen3moe 30B.A3B Q4_K - Medium | pp512 | 212.78 \u00b1 38.20; qwen3moe 30B.A3B Q4_K - Medium | tg128 | 22.79 \u00b1 0.11",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14995287",
   "tg_tokens": 128,
   "tps": 22.79,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-19",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-10879-c15006740-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b07097",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1041.42,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; llama 7B Q4_0 | pp512 | 1041.42 \u00b1 3.60; llama 7B Q4_0 | tg128 | 51.68 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15006740",
   "tg_tokens": 128,
   "tps": 51.68,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-19",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-10879-c15006740-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b07097",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1135.17,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; llama 7B Q4_0 | pp512 | 1135.17 \u00b1 6.29; llama 7B Q4_0 | tg128 | 54.76 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15006740",
   "tg_tokens": 128,
   "tps": 54.76,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-19",
   "flags": [],
   "hardware": "M2",
   "id": "lc-disc-10879-c15017064-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b07108",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1391.39,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M2; llama 7B Q4_0 | pp512 | 1391.39 \u00b1 1.19; llama 7B Q4_0 | tg128 | 129.58 \u00b1 0.58",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15017064",
   "tg_tokens": 128,
   "tps": 129.58,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-19",
   "flags": [],
   "hardware": "M2",
   "id": "lc-disc-10879-c15017064-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07108",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1411.25,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M2; llama 7B Q4_0 | pp512 | 1411.25 \u00b1 2.12; llama 7B Q4_0 | tg128 | 142.13 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15017064",
   "tg_tokens": 128,
   "tps": 142.13,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-20",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15019368-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100; config=b07062",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3226.16,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3226.16 \u00b1 8.13; llama 7B Q4_0 | tg128 | 109.92 \u00b1 0.22",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15019368",
   "tg_tokens": 128,
   "tps": 109.92,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-20",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15019368-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b07062",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3447.07,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3447.07 \u00b1 23.95; llama 7B Q4_0 | tg128 | 116.83 \u00b1 0.40",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15019368",
   "tg_tokens": 128,
   "tps": 116.83,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-20",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15019368-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100; config=b07062",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3309.54,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3309.54 \u00b1 34.19; llama 7B Q4_0 | tg128 | 134.07 \u00b1 0.12",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15019368",
   "tg_tokens": 128,
   "tps": 134.07,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-20",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15019368-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b07062",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3516.41,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3516.41 \u00b1 10.27; llama 7B Q4_0 | tg128 | 140.94 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15019368",
   "tg_tokens": 128,
   "tps": 140.94,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-20",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15019368-llama7b-Q4_0-5",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100; config=b07062",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3365.22,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3365.22 \u00b1 38.17; llama 7B Q4_0 | tg128 | 133.59 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15019368",
   "tg_tokens": 128,
   "tps": 133.59,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-20",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15019368-llama7b-Q4_0-6",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b07062",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3588.15,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3588.15 \u00b1 18.48; llama 7B Q4_0 | tg128 | 140.27 \u00b1 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15019368",
   "tg_tokens": 128,
   "tps": 140.27,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-20",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15019368-llama7b-Q4_0-7",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100; config=b07062",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1873.97,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 1873.97 \u00b1 7.61; llama 7B Q4_0 | tg128 | 119.98 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15019368",
   "tg_tokens": 128,
   "tps": 119.98,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-20",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15019368-llama7b-Q4_0-8",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b07062",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1888.77,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 1888.77 \u00b1 7.22; llama 7B Q4_0 | tg128 | 126.06 \u00b1 0.11",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15019368",
   "tg_tokens": 128,
   "tps": 126.06,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-20",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15019368-llama7b-Q4_0-9",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100; config=b07062",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2170.71,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2170.71 \u00b1 7.67; llama 7B Q4_0 | tg128 | 162.02 \u00b1 0.15",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15019368",
   "tg_tokens": 128,
   "tps": 162.02,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-20",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15019368-llama7b-Q4_0-10",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b07062",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2175.84,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2175.84 \u00b1 7.68; llama 7B Q4_0 | tg128 | 167.31 \u00b1 0.14",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15019368",
   "tg_tokens": 128,
   "tps": 167.31,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-20",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15019368-llama7b-Q4_0-11",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100; config=b07062",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2284.71,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2284.71 \u00b1 27.48; llama 7B Q4_0 | tg128 | 162.78 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15019368",
   "tg_tokens": 128,
   "tps": 162.78,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-20",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15019368-llama7b-Q4_0-12",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b07062",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2333.75,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2333.75 \u00b1 10.54; llama 7B Q4_0 | tg128 | 168.53 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15019368",
   "tg_tokens": 128,
   "tps": 168.53,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-20",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15019368-llama7b-Q4_0-13",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100; config=b07062",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3226.16,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3226.16 \u00b1 8.13; llama 7B Q4_0 | tg128 | 109.92 \u00b1 0.22",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15019368",
   "tg_tokens": 128,
   "tps": 109.92,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-20",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15019368-llama7b-Q4_0-14",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b07062",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3447.07,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3447.07 \u00b1 23.95; llama 7B Q4_0 | tg128 | 116.83 \u00b1 0.40",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15019368",
   "tg_tokens": 128,
   "tps": 116.83,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-20",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15019368-llama7b-Q4_0-15",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100; config=b07062",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3309.54,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3309.54 \u00b1 34.19; llama 7B Q4_0 | tg128 | 134.07 \u00b1 0.12",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15019368",
   "tg_tokens": 128,
   "tps": 134.07,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-20",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15019368-llama7b-Q4_0-16",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b07062",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3516.41,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3516.41 \u00b1 10.27; llama 7B Q4_0 | tg128 | 140.94 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15019368",
   "tg_tokens": 128,
   "tps": 140.94,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-20",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15019368-llama7b-Q4_0-17",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100; config=b07062",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3365.22,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3365.22 \u00b1 38.17; llama 7B Q4_0 | tg128 | 133.59 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15019368",
   "tg_tokens": 128,
   "tps": 133.59,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-20",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15019368-llama7b-Q4_0-18",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b07062",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3588.15,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3588.15 \u00b1 18.48; llama 7B Q4_0 | tg128 | 140.27 \u00b1 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15019368",
   "tg_tokens": 128,
   "tps": 140.27,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-20",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15019368-llama7b-Q4_0-19",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100; config=b07062",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1873.97,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 1873.97 \u00b1 7.61; llama 7B Q4_0 | tg128 | 119.98 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15019368",
   "tg_tokens": 128,
   "tps": 119.98,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-20",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15019368-llama7b-Q4_0-20",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b07062",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1888.77,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 1888.77 \u00b1 7.22; llama 7B Q4_0 | tg128 | 126.06 \u00b1 0.11",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15019368",
   "tg_tokens": 128,
   "tps": 126.06,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-20",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15019368-llama7b-Q4_0-21",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100; config=b07062",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2170.71,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2170.71 \u00b1 7.67; llama 7B Q4_0 | tg128 | 162.02 \u00b1 0.15",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15019368",
   "tg_tokens": 128,
   "tps": 162.02,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-20",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15019368-llama7b-Q4_0-22",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b07062",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2175.84,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2175.84 \u00b1 7.68; llama 7B Q4_0 | tg128 | 167.31 \u00b1 0.14",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15019368",
   "tg_tokens": 128,
   "tps": 167.31,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-20",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15019368-llama7b-Q4_0-23",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100; config=b07062",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2284.71,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2284.71 \u00b1 27.48; llama 7B Q4_0 | tg128 | 162.78 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15019368",
   "tg_tokens": 128,
   "tps": 162.78,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-20",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15019368-llama7b-Q4_0-24",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b07062",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2333.75,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2333.75 \u00b1 10.54; llama 7B Q4_0 | tg128 | 168.53 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15019368",
   "tg_tokens": 128,
   "tps": 168.53,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-20",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15020331-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100; config=b07062",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3190.19,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3190.19 \u00b1 15.15; llama 7B Q4_0 | tg128 | 114.02 \u00b1 0.42",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15020331",
   "tg_tokens": 128,
   "tps": 114.02,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-20",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15020331-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b07062",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3408.06,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3408.06 \u00b1 16.21; llama 7B Q4_0 | tg128 | 119.97 \u00b1 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15020331",
   "tg_tokens": 128,
   "tps": 119.97,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-20",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15020331-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100; config=b07062",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3278.06,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3278.06 \u00b1 67.85; llama 7B Q4_0 | tg128 | 132.78 \u00b1 0.18",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15020331",
   "tg_tokens": 128,
   "tps": 132.78,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-20",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15020331-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b07062",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3482.15,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3482.15 \u00b1 20.03; llama 7B Q4_0 | tg128 | 139.68 \u00b1 0.15",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15020331",
   "tg_tokens": 128,
   "tps": 139.68,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-20",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15020331-llama7b-Q4_0-5",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100; config=b07062",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3364.06,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3364.06 \u00b1 47.17; llama 7B Q4_0 | tg128 | 134.01 \u00b1 0.15",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15020331",
   "tg_tokens": 128,
   "tps": 134.01,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-20",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15020331-llama7b-Q4_0-6",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b07062",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3577.58,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3577.58 \u00b1 11.35; llama 7B Q4_0 | tg128 | 140.83 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15020331",
   "tg_tokens": 128,
   "tps": 140.83,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-20",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15020331-llama7b-Q4_0-7",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100; config=b07062",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2530.61,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2530.61 \u00b1 15.73; llama 7B Q4_0 | tg128 | 122.90 \u00b1 0.13",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15020331",
   "tg_tokens": 128,
   "tps": 122.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-20",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15020331-llama7b-Q4_0-8",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b07062",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2561.41,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2561.41 \u00b1 12.97; llama 7B Q4_0 | tg128 | 129.22 \u00b1 0.20",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15020331",
   "tg_tokens": 128,
   "tps": 129.22,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-20",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15020331-llama7b-Q4_0-9",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100; config=b07062",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3044.63,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3044.63 \u00b1 56.27; llama 7B Q4_0 | tg128 | 161.63 \u00b1 0.21",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15020331",
   "tg_tokens": 128,
   "tps": 161.63,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-20",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15020331-llama7b-Q4_0-10",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b07062",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3042.23,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3042.23 \u00b1 36.07; llama 7B Q4_0 | tg128 | 167.42 \u00b1 0.17",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15020331",
   "tg_tokens": 128,
   "tps": 167.42,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-20",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15020331-llama7b-Q4_0-11",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100; config=b07062",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3181.81,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3181.81 \u00b1 71.86; llama 7B Q4_0 | tg128 | 164.28 \u00b1 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15020331",
   "tg_tokens": 128,
   "tps": 164.28,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-20",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c15020331-llama7b-Q4_0-12",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b07062",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3178.79,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3178.79 \u00b1 33.58; llama 7B Q4_0 | tg128 | 170.13 \u00b1 0.18",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15020331",
   "tg_tokens": 128,
   "tps": 170.13,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-21",
   "flags": [],
   "hardware": "M1",
   "id": "lc-disc-10879-c15038551-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b07123",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 38.29,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M1; llama 7B Q4_0 | pp512 | 38.29 \u00b1 0.00; llama 7B Q4_0 | tg128 | 12.47 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15038551",
   "tg_tokens": 128,
   "tps": 12.47,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-21",
   "flags": [],
   "hardware": "M1",
   "id": "lc-disc-10879-c15038551-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07123",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 35.93,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M1; llama 7B Q4_0 | pp512 | 35.93 \u00b1 0.00; llama 7B Q4_0 | tg128 | 12.85 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15038551",
   "tg_tokens": 128,
   "tps": 12.85,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-23",
   "flags": [],
   "hardware": "RTX 3060",
   "id": "lc-disc-10879-c15053908-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b07118",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1815.7,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3060; llama 7B Q4_0 | pp512 | 1815.70 \u00b1 5.85; llama 7B Q4_0 | tg128 | 75.94 \u00b1 0.80",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15053908",
   "tg_tokens": 128,
   "tps": 75.94,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-23",
   "flags": [],
   "hardware": "RTX 3060",
   "id": "lc-disc-10879-c15053908-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07118",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2012.88,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3060; llama 7B Q4_0 | pp512 | 2012.88 \u00b1 10.12; llama 7B Q4_0 | tg128 | 80.59 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15053908",
   "tg_tokens": 128,
   "tps": 80.59,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-25",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "lc-disc-10879-c15069272-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=24; config=b07142",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 205.98,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M2 Ultra; llama 7B Q4_0 | pp512 | 205.98 \u00b1 0.02; llama 7B Q4_0 | tg128 | 34.34 \u00b1 0.12",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15069272",
   "tg_tokens": 128,
   "tps": 34.34,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-25",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "lc-disc-10879-c15069272-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=24; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07142",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 198.83,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M2 Ultra; llama 7B Q4_0 | pp512 | 198.83 \u00b1 0.85; llama 7B Q4_0 | tg128 | 36.07 \u00b1 0.11",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15069272",
   "tg_tokens": 128,
   "tps": 36.07,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-25",
   "flags": [],
   "hardware": "Radeon RX 6650 XT",
   "id": "lc-disc-10879-c15076033-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=8; config=b07142",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1029.52,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6650 XT; llama 7B Q4_0 | pp512 | 1029.52 \u00b1 1.21; llama 7B Q4_0 | tg128 | 62.14 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15076033",
   "tg_tokens": 128,
   "tps": 62.14,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-25",
   "flags": [],
   "hardware": "Radeon RX 6650 XT",
   "id": "lc-disc-10879-c15076033-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=8; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07142",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1088.9,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6650 XT; llama 7B Q4_0 | pp512 | 1088.90 \u00b1 0.40; llama 7B Q4_0 | tg128 | 64.53 \u00b1 0.75",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15076033",
   "tg_tokens": 128,
   "tps": 64.53,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-26",
   "flags": [],
   "hardware": "Radeon RX 480 G",
   "id": "lc-disc-10879-c15089098-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100; config=b07166",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 321.57,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 480 G; dev=VULKAN0; llama 7B Q4_0 | pp512 | 321.57 \u00b1 0.93; llama 7B Q4_0 | tg128 | 41.48 \u00b1 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15089098",
   "tg_tokens": 128,
   "tps": 41.48,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-26",
   "flags": [],
   "hardware": "Radeon RX 480 G",
   "id": "lc-disc-10879-c15089098-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b07166",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 330.84,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 480 G; dev=VULKAN0; llama 7B Q4_0 | pp512 | 330.84 \u00b1 1.02; llama 7B Q4_0 | tg128 | 43.33 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15089098",
   "tg_tokens": 128,
   "tps": 43.33,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-26",
   "flags": [],
   "hardware": "Radeon RX 480 G",
   "id": "lc-disc-10879-c15089098-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100; config=b07166",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 311.9,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 480 G; dev=VULKAN1; llama 7B Q4_0 | pp512 | 311.90 \u00b1 0.22; llama 7B Q4_0 | tg128 | 46.18 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15089098",
   "tg_tokens": 128,
   "tps": 46.18,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-26",
   "flags": [],
   "hardware": "Radeon RX 480 G",
   "id": "lc-disc-10879-c15089098-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07166",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 325.3,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 480 G; dev=VULKAN1; llama 7B Q4_0 | pp512 | 325.30 \u00b1 0.25; llama 7B Q4_0 | tg128 | 48.64 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15089098",
   "tg_tokens": 128,
   "tps": 48.64,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-26",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon RX 480 G",
   "id": "lc-disc-10879-c15089098-llama7b-Q4_0-5",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100; config=b07166",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 214.61,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 480 G; dev=VULKAN2; llama 7B Q4_0 | pp512 | 214.61 \u00b1 0.50; llama 7B Q4_0 | tg128 | 37.29 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15089098",
   "tg_tokens": 128,
   "tps": 37.29,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-26",
   "flags": [],
   "hardware": "Radeon RX 480 G",
   "id": "lc-disc-10879-c15089098-llama7b-Q4_0-6",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b07166",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 217.77,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 480 G; dev=VULKAN2; llama 7B Q4_0 | pp512 | 217.77 \u00b1 1.05; llama 7B Q4_0 | tg128 | 40.37 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15089098",
   "tg_tokens": 128,
   "tps": 40.37,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "lc-disc-10879-c13767244-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b05904",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 724.77,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M4 Max; llama 7B Q4_0 | pp512 | 724.77 \u00b1 20.93; llama 7B Q4_0 | tg128 | 75.02 \u00b1 0.14",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13767244",
   "tg_tokens": 128,
   "tps": 75.02,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "lc-disc-10879-c13767244-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b05904",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 557.46,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M4 Max; llama 7B Q4_0 | pp512 | 557.46 \u00b1 26.87; llama 7B Q4_0 | tg128 | 71.79 \u00b1 4.16",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13767244",
   "tg_tokens": 128,
   "tps": 71.79,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "lc-disc-10879-c13767244-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=12; config=b05904",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 949.34,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M4 Max; llama 7B Q4_0 | pp512 | 949.34 \u00b1 2.54; llama 7B Q4_0 | tg128 | 95.19 \u00b1 0.59",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13767244",
   "tg_tokens": 128,
   "tps": 95.19,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "lc-disc-10879-c13767244-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=12; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b05904",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 823.99,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M4 Max; llama 7B Q4_0 | pp512 | 823.99 \u00b1 48.77; llama 7B Q4_0 | tg128 | 96.62 \u00b1 2.14",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13767244",
   "tg_tokens": 128,
   "tps": 96.62,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-17",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c13790394-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b05904",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1937.34,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 1937.34 \u00b1 14.08; llama 7B Q4_0 | tg128 | 145.35 \u00b1 1.20",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13790394",
   "tg_tokens": 128,
   "tps": 145.35,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-17",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c13790394-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b05904",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1876.28,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 1876.28 \u00b1 3.99; llama 7B Q4_0 | tg128 | 133.02 \u00b1 0.69",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13790394",
   "tg_tokens": 128,
   "tps": 133.02,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-17",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c13790394-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; config=b05904",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3486.59,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3486.59 \u00b1 96.22; llama 7B Q4_0 | tg128 | 127.32 \u00b1 4.27",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13790394",
   "tg_tokens": 128,
   "tps": 127.32,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-17",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c13790394-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b05904",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3224.31,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3224.31 \u00b1 83.04; llama 7B Q4_0 | tg128 | 131.95 \u00b1 0.31",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13790394",
   "tg_tokens": 128,
   "tps": 131.95,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-17",
   "flags": [],
   "hardware": "Radeon RX 6800M",
   "id": "lc-disc-10879-c13790394-llama7b-Q4_0-5",
   "model": "llama 7B",
   "notes": "threads=100; config=b05904",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 480.52,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6800M; llama 7B Q4_0 | pp512 | 480.52 \u00b1 6.73; llama 7B Q4_0 | tg128 | 24.56 \u00b1 0.30",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13790394",
   "tg_tokens": 128,
   "tps": 24.56,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-17",
   "flags": [],
   "hardware": "Radeon RX 6800M",
   "id": "lc-disc-10879-c13790394-llama7b-Q4_0-6",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b05904",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 435.39,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6800M; llama 7B Q4_0 | pp512 | 435.39 \u00b1 5.00; llama 7B Q4_0 | tg128 | 24.76 \u00b1 0.36",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13790394",
   "tg_tokens": 128,
   "tps": 24.76,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-17",
   "flags": [],
   "hardware": "Radeon RX 6800M",
   "id": "lc-disc-10879-c13790394-llama7b-Q4_0-7",
   "model": "llama 7B",
   "notes": "threads=99; config=b05904",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 861.99,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6800M; llama 7B Q4_0 | pp512 | 861.99 \u00b1 7.67; llama 7B Q4_0 | tg128 | 48.71 \u00b1 0.71",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13790394",
   "tg_tokens": 128,
   "tps": 48.71,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-17",
   "flags": [],
   "hardware": "Radeon RX 6800M",
   "id": "lc-disc-10879-c13790394-llama7b-Q4_0-8",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b05904",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 784.16,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6800M; llama 7B Q4_0 | pp512 | 784.16 \u00b1 2.76; llama 7B Q4_0 | tg128 | 49.06 \u00b1 0.34",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13790394",
   "tg_tokens": 128,
   "tps": 49.06,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-27",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c13900110-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b05958",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3401.14,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3401.14 \u00b1 11.25; llama 7B Q4_0 | tg128 | 121.14 \u00b1 0.12",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13900110",
   "tg_tokens": 128,
   "tps": 121.14,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-27",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c13900110-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b05958",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1934.05,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 1934.05 \u00b1 2.46; llama 7B Q4_0 | tg128 | 107.16 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13900110",
   "tg_tokens": 128,
   "tps": 107.16,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-27",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c13900110-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b05958",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3379.59,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3379.59 \u00b1 8.91; llama 7B Q4_0 | tg128 | 120.88 \u00b1 0.11",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13900110",
   "tg_tokens": 128,
   "tps": 120.88,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-27",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c13900110-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b05958",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3605.23,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3605.23 \u00b1 6.72; llama 7B Q4_0 | tg128 | 117.12 \u00b1 0.11",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13900110",
   "tg_tokens": 128,
   "tps": 117.12,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-27",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c13900110-llama7b-Q4_0-5",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b05958",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3405.36,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3405.36 \u00b1 14.06; llama 7B Q4_0 | tg128 | 123.69 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13900110",
   "tg_tokens": 128,
   "tps": 123.69,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-27",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c13900110-llama7b-Q4_0-6",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b05958",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3638.45,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3638.45 \u00b1 5.49; llama 7B Q4_0 | tg128 | 120.73 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13900110",
   "tg_tokens": 128,
   "tps": 120.73,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-22",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c13851206-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3489.67,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3489.67 \u00b1 82.17; llama 7B Q4_0 | tg128 | 145.00 \u00b1 0.89",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13851206",
   "tg_tokens": 128,
   "tps": 145.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-22",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c13851206-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3225.19,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3225.19 \u00b1 10.11; llama 7B Q4_0 | tg128 | 145.70 \u00b1 0.73",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13851206",
   "tg_tokens": 128,
   "tps": 145.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-22",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c13851206-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3529.1,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3529.10 \u00b1 56.47; llama 7B Q4_0 | tg128 | 150.86 \u00b1 0.29",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13851206",
   "tg_tokens": 128,
   "tps": 150.86,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-22",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c13851206-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3598.42,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3598.42 \u00b1 10.65; llama 7B Q4_0 | tg128 | 140.96 \u00b1 0.23",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13851206",
   "tg_tokens": 128,
   "tps": 140.96,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-22",
   "flags": [],
   "hardware": "Radeon RX 6800M",
   "id": "lc-disc-10879-c13851674-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b05958",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 861.99,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6800M; llama 7B Q4_0 | pp512 | 861.99 \u00b1 7.67; llama 7B Q4_0 | tg128 | 48.71 \u00b1 0.71",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13851674",
   "tg_tokens": 128,
   "tps": 48.71,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-22",
   "flags": [],
   "hardware": "Radeon RX 6800M",
   "id": "lc-disc-10879-c13851674-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b05958",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 784.16,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6800M; llama 7B Q4_0 | pp512 | 784.16 \u00b1 2.76; llama 7B Q4_0 | tg128 | 49.06 \u00b1 0.34",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13851674",
   "tg_tokens": 128,
   "tps": 49.06,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-22",
   "flags": [],
   "hardware": "Radeon RX 6800M",
   "id": "lc-disc-10879-c13851674-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; config=b05958",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 663.96,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6800M; llama 7B Q4_0 | pp512 | 663.96 \u00b1 4.82; llama 7B Q4_0 | tg128 | 49.58 \u00b1 1.58",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13851674",
   "tg_tokens": 128,
   "tps": 49.58,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-22",
   "flags": [],
   "hardware": "Radeon RX 6800M",
   "id": "lc-disc-10879-c13851674-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b05958",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 442.71,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6800M; llama 7B Q4_0 | pp512 | 442.71 \u00b1 4.26; llama 7B Q4_0 | tg128 | 51.36 \u00b1 0.25",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13851674",
   "tg_tokens": 128,
   "tps": 51.36,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-30",
   "flags": [],
   "hardware": "Radeon RX 7900 XT",
   "id": "lc-disc-10879-c13934683-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b06030",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3098.38,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XT; llama 7B Q4_0 | pp512 | 3098.38 \u00b1 24.02; llama 7B Q4_0 | tg128 | 116.15 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13934683",
   "tg_tokens": 128,
   "tps": 116.15,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-30",
   "flags": [],
   "hardware": "Radeon RX 7900 XT",
   "id": "lc-disc-10879-c13934683-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06030",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3261.75,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XT; llama 7B Q4_0 | pp512 | 3261.75 \u00b1 9.09; llama 7B Q4_0 | tg128 | 112.30 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13934683",
   "tg_tokens": 128,
   "tps": 112.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-30",
   "flags": [],
   "hardware": "Radeon RX 7900 XT",
   "id": "lc-disc-10879-c13934683-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100; config=b06030",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2823.35,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XT; llama 7B Q4_0 | pp512 | 2823.35 \u00b1 35.08; llama 7B Q4_0 | tg128 | 117.45 \u00b1 0.13",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13934683",
   "tg_tokens": 128,
   "tps": 117.45,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-30",
   "flags": [],
   "hardware": "Radeon RX 7900 XT",
   "id": "lc-disc-10879-c13934683-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b06030",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2609.51,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XT; llama 7B Q4_0 | pp512 | 2609.51 \u00b1 8.07; llama 7B Q4_0 | tg128 | 117.65 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13934683",
   "tg_tokens": 128,
   "tps": 117.65,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-30",
   "flags": [],
   "hardware": "Arc A770",
   "id": "lc-disc-10879-c13937285-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b06026",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1543.17,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A770; llama 7B Q4_0 | pp512 | 1543.17 \u00b1 1.75; llama 7B Q4_0 | tg128 | 74.24 \u00b1 0.58",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13937285",
   "tg_tokens": 128,
   "tps": 74.24,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-30",
   "flags": [],
   "hardware": "Arc A770",
   "id": "lc-disc-10879-c13937285-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06026",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1479.07,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A770; llama 7B Q4_0 | pp512 | 1479.07 \u00b1 0.33; llama 7B Q4_0 | tg128 | 75.75 \u00b1 2.46",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13937285",
   "tg_tokens": 128,
   "tps": 75.75,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-30",
   "flags": [],
   "hardware": "Arc A770",
   "id": "lc-disc-10879-c13937285-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; config=b06026",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1357.35,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A770; llama 7B Q4_0 | pp512 | 1357.35 \u00b1 1.41; llama 7B Q4_0 | tg128 | 71.93 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13937285",
   "tg_tokens": 128,
   "tps": 71.93,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-30",
   "flags": [],
   "hardware": "Arc A770",
   "id": "lc-disc-10879-c13937285-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06026",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 987.41,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A770; llama 7B Q4_0 | pp512 | 987.41 \u00b1 0.49; llama 7B Q4_0 | tg128 | 69.27 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13937285",
   "tg_tokens": 128,
   "tps": 69.27,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-30",
   "flags": [
    "contradiction"
   ],
   "hardware": "Arc A770",
   "id": "lc-disc-10879-c13937285-llama7b-Q4_0-5",
   "model": "llama 7B",
   "notes": "threads=99; config=b06026",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 706.67,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A770; llama 7B Q4_0 | pp512 | 706.67 \u00b1 0.39; llama 7B Q4_0 | tg128 | 38.07 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13937285",
   "tg_tokens": 128,
   "tps": 38.07,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-30",
   "flags": [
    "contradiction"
   ],
   "hardware": "Arc A770",
   "id": "lc-disc-10879-c13937285-llama7b-Q4_0-6",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06026",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 255.59,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A770; llama 7B Q4_0 | pp512 | 255.59 \u00b1 0.10; llama 7B Q4_0 | tg128 | 38.80 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13937285",
   "tg_tokens": 128,
   "tps": 38.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-30",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-disc-10879-c13937285-llama7b-Q4_0-7",
   "model": "llama 7B",
   "notes": "threads=99; config=b06026",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3226.8,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 5060 Ti; llama 7B Q4_0 | pp512 | 3226.80 \u00b1 4.88; llama 7B Q4_0 | tg128 | 74.47 \u00b1 6.42",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13937285",
   "tg_tokens": 128,
   "tps": 74.47,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-30",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-disc-10879-c13937285-llama7b-Q4_0-8",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 6 in this comment (separate llama-bench run; build/flags differ); config=b06026",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3504.54,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 5060 Ti; llama 7B Q4_0 | pp512 | 3504.54 \u00b1 3.82; llama 7B Q4_0 | tg128 | 78.10 \u00b1 0.23",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13937285",
   "tg_tokens": 128,
   "tps": 78.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-30",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-disc-10879-c13937285-llama7b-Q4_0-9",
   "model": "llama 7B",
   "notes": "threads=99; run 3 of 6 in this comment (separate llama-bench run; build/flags differ); config=b06026",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3060.78,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 5060 Ti; llama 7B Q4_0 | pp512 | 3060.78 \u00b1 154.26; llama 7B Q4_0 | tg128 | 74.52 \u00b1 0.17",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13937285",
   "tg_tokens": 128,
   "tps": 74.52,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-30",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-disc-10879-c13937285-llama7b-Q4_0-10",
   "model": "llama 7B",
   "notes": "threads=99; run 4 of 6 in this comment (separate llama-bench run; build/flags differ); config=b06026",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3386.08,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 5060 Ti; llama 7B Q4_0 | pp512 | 3386.08 \u00b1 3.45; llama 7B Q4_0 | tg128 | 74.80 \u00b1 0.25",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13937285",
   "tg_tokens": 128,
   "tps": 74.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-30",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-disc-10879-c13937285-llama7b-Q4_0-11",
   "model": "llama 7B",
   "notes": "threads=99; run 5 of 6 in this comment (separate llama-bench run; build/flags differ); config=b06026",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2784.46,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 5060 Ti; llama 7B Q4_0 | pp512 | 2784.46 \u00b1 2.02; llama 7B Q4_0 | tg128 | 43.79 \u00b1 1.20",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13937285",
   "tg_tokens": 128,
   "tps": 43.79,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-30",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-disc-10879-c13937285-llama7b-Q4_0-12",
   "model": "llama 7B",
   "notes": "threads=99; run 6 of 6 in this comment (separate llama-bench run; build/flags differ); config=b06026",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2718.32,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 5060 Ti; llama 7B Q4_0 | pp512 | 2718.32 \u00b1 0.79; llama 7B Q4_0 | tg128 | 42.47 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13937285",
   "tg_tokens": 128,
   "tps": 42.47,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-30",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-disc-10879-c13937285-llama7b-Q4_0-13",
   "model": "llama 7B",
   "notes": "threads=99; config=b06026",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3728.09,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 5060 Ti; llama 7B Q4_0 | pp512 | 3728.09 \u00b1 5.24; llama 7B Q4_0 | tg128 | 90.95 \u00b1 0.14",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13937285",
   "tg_tokens": 128,
   "tps": 90.95,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-30",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-disc-10879-c13937285-llama7b-Q4_0-14",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 6 in this comment (separate llama-bench run; build/flags differ); config=b06026",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4163.39,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 5060 Ti; llama 7B Q4_0 | pp512 | 4163.39 \u00b1 5.25; llama 7B Q4_0 | tg128 | 93.30 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13937285",
   "tg_tokens": 128,
   "tps": 93.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-30",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-disc-10879-c13937285-llama7b-Q4_0-15",
   "model": "llama 7B",
   "notes": "threads=99; run 3 of 6 in this comment (separate llama-bench run; build/flags differ); config=b06026",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3648.71,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 5060 Ti; llama 7B Q4_0 | pp512 | 3648.71 \u00b1 5.03; llama 7B Q4_0 | tg128 | 89.57 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13937285",
   "tg_tokens": 128,
   "tps": 89.57,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-30",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-disc-10879-c13937285-llama7b-Q4_0-16",
   "model": "llama 7B",
   "notes": "threads=99; run 4 of 6 in this comment (separate llama-bench run; build/flags differ); config=b06026",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4073.37,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 5060 Ti; llama 7B Q4_0 | pp512 | 4073.37 \u00b1 5.93; llama 7B Q4_0 | tg128 | 91.56 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13937285",
   "tg_tokens": 128,
   "tps": 91.56,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-30",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-disc-10879-c13937285-llama7b-Q4_0-17",
   "model": "llama 7B",
   "notes": "threads=99; run 5 of 6 in this comment (separate llama-bench run; build/flags differ); config=b06026",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3526.38,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 5060 Ti; llama 7B Q4_0 | pp512 | 3526.38 \u00b1 6.80; llama 7B Q4_0 | tg128 | 89.75 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13937285",
   "tg_tokens": 128,
   "tps": 89.75,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-30",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-disc-10879-c13937285-llama7b-Q4_0-18",
   "model": "llama 7B",
   "notes": "threads=99; run 6 of 6 in this comment (separate llama-bench run; build/flags differ); config=b06026",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3917.29,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 5060 Ti; llama 7B Q4_0 | pp512 | 3917.29 \u00b1 1.69; llama 7B Q4_0 | tg128 | 89.87 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13937285",
   "tg_tokens": 128,
   "tps": 89.87,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-30",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-disc-10879-c13937285-deci70b-IQ4_XS - 4.25 bpw",
   "model": "deci 70B",
   "notes": "threads=99; config=b06026",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 112.31,
   "provenance": "community",
   "quant": "IQ4_XS - 4.25 bpw",
   "quote": "NVIDIA GeForce RTX 5060 Ti, compute capability 12.0, VMM: yes\n  Device; deci 70B IQ4_XS - 4.25 bpw | pp512 | 112.31 \u00b1 0.79; deci 70B IQ4_XS - 4.25 bpw | tg128 | 10.61 \u00b1 1.54",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13937285",
   "tg_tokens": 128,
   "tps": 10.61,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-30",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-disc-10879-c13937285-deci70b-IQ4_XS - 4.25 bpw-2",
   "model": "deci 70B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06026",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 55.4,
   "provenance": "community",
   "quant": "IQ4_XS - 4.25 bpw",
   "quote": "NVIDIA GeForce RTX 5060 Ti, compute capability 12.0, VMM: yes\n  Device; deci 70B IQ4_XS - 4.25 bpw | pp512 | 55.40 \u00b1 0.15; deci 70B IQ4_XS - 4.25 bpw | tg128 | 6.73 \u00b1 0.46",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13937285",
   "tg_tokens": 128,
   "tps": 6.73,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-30",
   "flags": [],
   "hardware": "Arc A770",
   "id": "lc-disc-10879-c13937285-deci70b-IQ4_XS - 4.25 bpw-3",
   "model": "deci 70B",
   "notes": "threads=99; config=b06026",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 132.43,
   "provenance": "community",
   "quant": "IQ4_XS - 4.25 bpw",
   "quote": "Arc A770; deci 70B IQ4_XS - 4.25 bpw | pp512 | 132.43 \u00b1 0.30; deci 70B IQ4_XS - 4.25 bpw | tg128 | 13.33 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13937285",
   "tg_tokens": 128,
   "tps": 13.33,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-30",
   "flags": [],
   "hardware": "Arc A770",
   "id": "lc-disc-10879-c13937285-deci70b-IQ4_XS - 4.25 bpw-4",
   "model": "deci 70B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06026",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 109.97,
   "provenance": "community",
   "quant": "IQ4_XS - 4.25 bpw",
   "quote": "Arc A770; deci 70B IQ4_XS - 4.25 bpw | pp512 | 109.97 \u00b1 3.55; deci 70B IQ4_XS - 4.25 bpw | tg128 | 12.89 \u00b1 0.14",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13937285",
   "tg_tokens": 128,
   "tps": 12.89,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-30",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-disc-10879-c13937285-deci70b-IQ4_XS - 4.25 bpw-5",
   "model": "deci 70B",
   "notes": "threads=99; config=b06026",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 423.66,
   "provenance": "community",
   "quant": "IQ4_XS - 4.25 bpw",
   "quote": "RTX 5060 Ti; deci 70B IQ4_XS - 4.25 bpw | pp512 | 423.66 \u00b1 0.32; deci 70B IQ4_XS - 4.25 bpw | tg128 | 9.73 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13937285",
   "tg_tokens": 128,
   "tps": 9.73,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-30",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-disc-10879-c13937285-deci70b-IQ4_XS - 4.25 bpw-6",
   "model": "deci 70B",
   "notes": "threads=99; run 2 of 4 in this comment (separate llama-bench run; build/flags differ); config=b06026",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 438.01,
   "provenance": "community",
   "quant": "IQ4_XS - 4.25 bpw",
   "quote": "RTX 5060 Ti; deci 70B IQ4_XS - 4.25 bpw | pp512 | 438.01 \u00b1 0.04; deci 70B IQ4_XS - 4.25 bpw | tg128 | 9.76 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13937285",
   "tg_tokens": 128,
   "tps": 9.76,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-30",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-disc-10879-c13937285-deci70b-IQ4_XS - 4.25 bpw-7",
   "model": "deci 70B",
   "notes": "threads=99; run 3 of 4 in this comment (separate llama-bench run; build/flags differ); config=b06026",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 412.37,
   "provenance": "community",
   "quant": "IQ4_XS - 4.25 bpw",
   "quote": "RTX 5060 Ti; deci 70B IQ4_XS - 4.25 bpw | pp512 | 412.37 \u00b1 0.42; deci 70B IQ4_XS - 4.25 bpw | tg128 | 8.45 \u00b1 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13937285",
   "tg_tokens": 128,
   "tps": 8.45,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-30",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-disc-10879-c13937285-deci70b-IQ4_XS - 4.25 bpw-8",
   "model": "deci 70B",
   "notes": "threads=99; run 4 of 4 in this comment (separate llama-bench run; build/flags differ); config=b06026",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 426.57,
   "provenance": "community",
   "quant": "IQ4_XS - 4.25 bpw",
   "quote": "RTX 5060 Ti; deci 70B IQ4_XS - 4.25 bpw | pp512 | 426.57 \u00b1 0.06; deci 70B IQ4_XS - 4.25 bpw | tg128 | 8.48 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13937285",
   "tg_tokens": 128,
   "tps": 8.48,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-30",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-disc-10879-c13937285-deci70b-IQ4_XS - 4.25 bpw-9",
   "model": "deci 70B",
   "notes": "threads=99; config=b06026",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 549.87,
   "provenance": "community",
   "quant": "IQ4_XS - 4.25 bpw",
   "quote": "RTX 5060 Ti; deci 70B IQ4_XS - 4.25 bpw | pp512 | 549.87 \u00b1 0.33; deci 70B IQ4_XS - 4.25 bpw | tg128 | 14.41 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13937285",
   "tg_tokens": 128,
   "tps": 14.41,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-30",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-disc-10879-c13937285-deci70b-IQ4_XS - 4.25 bpw-10",
   "model": "deci 70B",
   "notes": "threads=99; run 2 of 4 in this comment (separate llama-bench run; build/flags differ); config=b06026",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 577.31,
   "provenance": "community",
   "quant": "IQ4_XS - 4.25 bpw",
   "quote": "RTX 5060 Ti; deci 70B IQ4_XS - 4.25 bpw | pp512 | 577.31 \u00b1 0.37; deci 70B IQ4_XS - 4.25 bpw | tg128 | 14.60 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13937285",
   "tg_tokens": 128,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-30",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-disc-10879-c13937285-deci70b-IQ4_XS - 4.25 bpw-11",
   "model": "deci 70B",
   "notes": "threads=99; run 3 of 4 in this comment (separate llama-bench run; build/flags differ); config=b06026",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 542.01,
   "provenance": "community",
   "quant": "IQ4_XS - 4.25 bpw",
   "quote": "RTX 5060 Ti; deci 70B IQ4_XS - 4.25 bpw | pp512 | 542.01 \u00b1 0.14; deci 70B IQ4_XS - 4.25 bpw | tg128 | 14.35 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13937285",
   "tg_tokens": 128,
   "tps": 14.35,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-30",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-disc-10879-c13937285-deci70b-IQ4_XS - 4.25 bpw-12",
   "model": "deci 70B",
   "notes": "threads=99; run 4 of 4 in this comment (separate llama-bench run; build/flags differ); config=b06026",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 573.75,
   "provenance": "community",
   "quant": "IQ4_XS - 4.25 bpw",
   "quote": "RTX 5060 Ti; deci 70B IQ4_XS - 4.25 bpw | pp512 | 573.75 \u00b1 0.22; deci 70B IQ4_XS - 4.25 bpw | tg128 | 14.59 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13937285",
   "tg_tokens": 128,
   "tps": 14.59,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-30",
   "flags": [],
   "hardware": "Arc A770",
   "id": "lc-disc-10879-c13937968-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b06026",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 737.29,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A770; llama 7B Q4_0 | pp512 | 737.29 \u00b1 10.12; llama 7B Q4_0 | tg128 | 40.56 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13937968",
   "tg_tokens": 128,
   "tps": 40.56,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-30",
   "flags": [],
   "hardware": "Arc A770",
   "id": "lc-disc-10879-c13937968-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b06026",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 257.53,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A770; llama 7B Q4_0 | pp512 | 257.53 \u00b1 0.96; llama 7B Q4_0 | tg128 | 41.19 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13937968",
   "tg_tokens": 128,
   "tps": 41.19,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-30",
   "flags": [],
   "hardware": "Arc A770",
   "id": "lc-disc-10879-c13937968-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b06026",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 934.47,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A770; llama 7B Q4_0 | pp512 | 934.47 \u00b1 6.84; llama 7B Q4_0 | tg128 | 31.45 \u00b1 0.16",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13937968",
   "tg_tokens": 128,
   "tps": 31.45,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-30",
   "flags": [],
   "hardware": "Arc A770",
   "id": "lc-disc-10879-c13937968-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b06026",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 288.17,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A770; llama 7B Q4_0 | pp512 | 288.17 \u00b1 0.25; llama 7B Q4_0 | tg128 | 18.22 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13937968",
   "tg_tokens": 128,
   "tps": 18.22,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-30",
   "flags": [],
   "hardware": "A770",
   "id": "lc-disc-10879-c13937968-llama7b-Q4_0-5",
   "model": "llama 7B",
   "notes": "threads=99; config=b06026",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2173.52,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A770; llama 7B Q4_0 | pp512 | 2173.52 \u00b1 24.68; llama 7B Q4_0 | tg128 | 52.98 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13937968",
   "tg_tokens": 128,
   "tps": 52.98,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2025-07-30",
   "flags": [],
   "hardware": "A770",
   "id": "lc-disc-10879-c13937968-llama7b-Q4_0-6",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06026",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2185.36,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A770; llama 7B Q4_0 | pp512 | 2185.36 \u00b1 2.39; llama 7B Q4_0 | tg128 | 52.26 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13937968",
   "tg_tokens": 128,
   "tps": 52.26,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-04",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-disc-10879-c14874993-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b06753",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1741.01,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 5060 Ti; llama 7B Q4_0 | pp512 | 1741.01 \u00b1 1.07; llama 7B Q4_0 | tg128 | 102.31 \u00b1 0.51",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14874993",
   "tg_tokens": 128,
   "tps": 102.31,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-04",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-disc-10879-c14874993-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06753",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1678.98,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 5060 Ti; llama 7B Q4_0 | pp512 | 1678.98 \u00b1 1.22; llama 7B Q4_0 | tg128 | 101.35 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14874993",
   "tg_tokens": 128,
   "tps": 101.35,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-04",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-disc-10879-c14874993-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; config=b06753",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1986.47,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 5060 Ti; llama 7B Q4_0 | pp512 | 1986.47 \u00b1 1.41; llama 7B Q4_0 | tg128 | 102.66 \u00b1 0.38",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14874993",
   "tg_tokens": 128,
   "tps": 102.66,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-04",
   "flags": [],
   "hardware": "RTX 5060 Ti",
   "id": "lc-disc-10879-c14874993-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06753",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1907.81,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 5060 Ti; llama 7B Q4_0 | pp512 | 1907.81 \u00b1 2.31; llama 7B Q4_0 | tg128 | 104.42 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14874993",
   "tg_tokens": 128,
   "tps": 104.42,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-04",
   "flags": [],
   "hardware": "Arc A770",
   "id": "lc-disc-10879-c14874993-llama7b-Q4_0-5",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b06753",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 926.02,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A770; llama 7B Q4_0 | pp512 | 926.02 \u00b1 17.99; llama 7B Q4_0 | tg128 | 50.47 \u00b1 0.32",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14874993",
   "tg_tokens": 128,
   "tps": 50.47,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-04",
   "flags": [],
   "hardware": "Arc A770",
   "id": "lc-disc-10879-c14874993-llama7b-Q4_0-6",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b06753",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 251.11,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A770; llama 7B Q4_0 | pp512 | 251.11 \u00b1 1.15; llama 7B Q4_0 | tg128 | 51.36 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14874993",
   "tg_tokens": 128,
   "tps": 51.36,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-04",
   "flags": [],
   "hardware": "Arc A770",
   "id": "lc-disc-10879-c14874993-llama7b-Q4_0-7",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b06753",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1185.04,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A770; llama 7B Q4_0 | pp512 | 1185.04 \u00b1 27.79; llama 7B Q4_0 | tg128 | 51.43 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14874993",
   "tg_tokens": 128,
   "tps": 51.43,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-04",
   "flags": [],
   "hardware": "Arc A770",
   "id": "lc-disc-10879-c14874993-llama7b-Q4_0-8",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b06753",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 296.62,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A770; llama 7B Q4_0 | pp512 | 296.62 \u00b1 1.50; llama 7B Q4_0 | tg128 | 52.48 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14874993",
   "tg_tokens": 128,
   "tps": 52.48,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-04",
   "flags": [
    "contradiction"
   ],
   "hardware": "RTX 5060 Ti",
   "id": "lc-disc-10879-c14874993-llama7b-Q4_0-9",
   "model": "llama 7B",
   "notes": "threads=99; config=b06753",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3248.01,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 5060 Ti; llama 7B Q4_0 | pp512 | 3248.01 \u00b1 3.09; llama 7B Q4_0 | tg128 | 87.05 \u00b1 7.36",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14874993",
   "tg_tokens": 128,
   "tps": 87.05,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-04",
   "flags": [
    "contradiction"
   ],
   "hardware": "RTX 5060 Ti",
   "id": "lc-disc-10879-c14874993-llama7b-Q4_0-10",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06753",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3720.85,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 5060 Ti; llama 7B Q4_0 | pp512 | 3720.85 \u00b1 7.03; llama 7B Q4_0 | tg128 | 91.85 \u00b1 0.25",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14874993",
   "tg_tokens": 128,
   "tps": 91.85,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-04",
   "flags": [
    "contradiction"
   ],
   "hardware": "RTX 5060 Ti",
   "id": "lc-disc-10879-c14874993-llama7b-Q4_0-11",
   "model": "llama 7B",
   "notes": "threads=99; config=b06753",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3240.12,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 5060 Ti; llama 7B Q4_0 | pp512 | 3240.12 \u00b1 2.67; llama 7B Q4_0 | tg128 | 91.41 \u00b1 0.14",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14874993",
   "tg_tokens": 128,
   "tps": 91.41,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-04",
   "flags": [
    "contradiction"
   ],
   "hardware": "RTX 5060 Ti",
   "id": "lc-disc-10879-c14874993-llama7b-Q4_0-12",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06753",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3713.27,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 5060 Ti; llama 7B Q4_0 | pp512 | 3713.27 \u00b1 5.43; llama 7B Q4_0 | tg128 | 93.76 \u00b1 0.18",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14874993",
   "tg_tokens": 128,
   "tps": 93.76,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-01",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-disc-10879-c13963727-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b06056",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2145.6,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7800 XT; llama 7B Q4_0 | pp512 | 2145.60 + 23.14; llama 7B Q4_0 | tg128 | 96.89 + 0.22",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13963727",
   "tg_tokens": 128,
   "tps": 96.89,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-01",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-disc-10879-c13963727-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06056",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2063.66,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7800 XT; llama 7B Q4_0 | pp512 | 2063.66 + 2.92; llama 7B Q4_0 | tg128 | 96.03 + 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13963727",
   "tg_tokens": 128,
   "tps": 96.03,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-01",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-disc-10879-c13963727-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; config=b06056",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2151.81,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7800 XT; llama 7B Q4_0 | pp512 | 2151.81 + 17.94; llama 7B Q4_0 | tg128 | 100.94 + 0.10",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13963727",
   "tg_tokens": 128,
   "tps": 100.94,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-01",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-disc-10879-c13963727-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06056",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2304.63,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7800 XT; llama 7B Q4_0 | pp512 | 2304.63 + 2.85; llama 7B Q4_0 | tg128 | 95.99 + 0.21",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-13963727",
   "tg_tokens": 128,
   "tps": 95.99,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-08",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-disc-10879-c14048792-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b06118",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2055.58,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7800 XT; llama 7B Q4_0 | pp512 | 2055.58 + 10.12; llama 7B Q4_0 | pp1024 | 1805.50 + 5.85; llama 7B Q4_0 | pp2048 | 1317.89 + 1.30; llama 7B Q4_0 | pp4096 | 1173.49 + 0.98; llama 7B Q4_0 | pp8192 | 980.86 + 0.47; llama 7B Q4_0 | pp16384 | 762.15 + 0.23; llama 7B Q4_0 | tg128 | 96.61 + 2.62; llama 7B Q4_0 | tg256 | 96.36 + 0.05; llama 7B Q4_0 | tg512 | 92.81 + 0.03; llama 7B Q4_0 | tg1024 | 83.30 + 0.03; llama 7B Q4_0 | tg2048 | 63.57 + 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14048792",
   "tg_tokens": 128,
   "tps": 96.61,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-08",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-disc-10879-c14048792-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06118",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2026.84,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7800 XT; llama 7B Q4_0 | pp512 | 2026.84 + 4.16; llama 7B Q4_0 | pp1024 | 1870.10 + 2.98; llama 7B Q4_0 | pp2048 | 1607.70 + 2.27; llama 7B Q4_0 | pp4096 | 1258.16 + 0.41; llama 7B Q4_0 | pp8192 | 878.79 + 0.51; llama 7B Q4_0 | pp16384 | 545.52 + 0.55; llama 7B Q4_0 | tg128 | 97.07 + 0.14; llama 7B Q4_0 | tg256 | 97.07 + 0.04; llama 7B Q4_0 | tg512 | 95.99 + 0.02; llama 7B Q4_0 | tg1024 | 93.71 + 0.01; llama 7B Q4_0 | tg2048 | 89.15 + 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14048792",
   "tg_tokens": 128,
   "tps": 97.07,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-06",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c14019775-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100; config=b06099",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3288.72,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3288.72 \u00b1 54.00; llama 7B Q4_0 | tg128 | 148.01 \u00b1 0.97",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14019775",
   "tg_tokens": 128,
   "tps": 148.01,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-06",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c14019775-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b06099",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3134.2,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3134.20 \u00b1 9.65; llama 7B Q4_0 | tg128 | 148.34 \u00b1 0.17",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14019775",
   "tg_tokens": 128,
   "tps": 148.34,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-07",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-10879-c14038254-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b06096",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3461.64,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 3461.64 \u00b1 30.11; llama 7B Q4_0 | tg128 | 130.56 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14038254",
   "tg_tokens": 128,
   "tps": 130.56,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-07",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-10879-c14038254-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06096",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3831.64,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 3831.64 \u00b1 1.82; llama 7B Q4_0 | tg128 | 130.57 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14038254",
   "tg_tokens": 128,
   "tps": 130.57,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-07",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-10879-c14038254-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; config=b06096",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3530.68,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 3530.68 \u00b1 18.14; llama 7B Q4_0 | tg128 | 125.45 \u00b1 0.28",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14038254",
   "tg_tokens": 128,
   "tps": 125.45,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-07",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-10879-c14038254-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=0 ngl=99; config=b06096",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3558.71,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 3558.71 \u00b1 2.87; llama 7B Q4_0 | tg128 | 124.33 \u00b1 0.56",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14038254",
   "tg_tokens": 128,
   "tps": 124.33,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-07",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-10879-c14038665-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=1 ngl=99; config=b06096",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 589.16,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 589.16 \u00b1 0.22; llama 7B Q4_0 | tg128 | 91.86 \u00b1 0.29",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14038665",
   "tg_tokens": 128,
   "tps": 91.86,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-07",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-10879-c14038665-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=0 ngl=99; config=b06096",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 723.41,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 723.41 \u00b1 0.15; llama 7B Q4_0 | tg128 | 90.08 \u00b1 0.39",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14038665",
   "tg_tokens": 128,
   "tps": 90.08,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-26",
   "flags": [],
   "hardware": "Radeon RX 6900 XT",
   "id": "lc-disc-10879-c14784210-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=10",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 116.28,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6900 XT; llama 7B Q4_0 | pp512 | 116.28 \u00b1 0.36; llama 7B Q4_0 | tg128 | 71.96 \u00b1 0.29",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14784210",
   "tg_tokens": 128,
   "tps": 71.96,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-26",
   "flags": [],
   "hardware": "Radeon RX 6900 XT",
   "id": "lc-disc-10879-c14784210-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=10; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 111.82,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6900 XT; llama 7B Q4_0 | pp512 | 111.82 \u00b1 0.08; llama 7B Q4_0 | tg128 | 72.39 \u00b1 0.20",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14784210",
   "tg_tokens": 128,
   "tps": 72.39,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-26",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon RX 6900 XT",
   "id": "lc-disc-10879-c14784210-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=10",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 23.9,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6900 XT; llama 7B Q4_0 | pp512 | 23.90 \u00b1 0.32; llama 7B Q4_0 | tg128 | 36.72 \u00b1 2.79",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14784210",
   "tg_tokens": 128,
   "tps": 36.72,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-26",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon RX 6900 XT",
   "id": "lc-disc-10879-c14784210-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=10; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 21.66,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6900 XT; llama 7B Q4_0 | pp512 | 21.66 \u00b1 0.21; llama 7B Q4_0 | tg128 | 37.33 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14784210",
   "tg_tokens": 128,
   "tps": 37.33,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-26",
   "flags": [],
   "hardware": "Radeon RX 6900 XT",
   "id": "lc-disc-10879-c14784210-llama7b-Q4_0-5",
   "model": "llama 7B",
   "notes": "threads=10",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 118.09,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6900 XT; llama 7B Q4_0 | pp512 | 118.09 \u00b1 0.39; llama 7B Q4_0 | tg128 | 72.31 \u00b1 0.17",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14784210",
   "tg_tokens": 128,
   "tps": 72.31,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-26",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon RX 6900 XT",
   "id": "lc-disc-10879-c14784210-llama7b-Q4_0-6",
   "model": "llama 7B",
   "notes": "threads=10",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 23.61,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6900 XT; llama 7B Q4_0 | pp512 | 23.61 \u00b1 0.61; llama 7B Q4_0 | tg128 | 37.59 \u00b1 0.41",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14784210",
   "tg_tokens": 128,
   "tps": 37.59,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-27",
   "flags": [],
   "hardware": "Radeon 890M",
   "id": "lc-disc-10879-c14798753-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b06854",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 293.41,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 890M; llama 7B Q4_0 | pp512 | 293.41 \u00b1 3.16; llama 7B Q4_0 | tg128 | 8.31 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14798753",
   "tg_tokens": 128,
   "tps": 8.31,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-27",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon 890M",
   "id": "lc-disc-10879-c14798753-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06854",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 315.24,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 890M; llama 7B Q4_0 | pp512 | 315.24 \u00b1 3.80; llama 7B Q4_0 | tg128 | 8.30 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14798753",
   "tg_tokens": 128,
   "tps": 8.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-27",
   "flags": [],
   "hardware": "Radeon 890M",
   "id": "lc-disc-10879-c14798753-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; config=b06854",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 201.59,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 890M; llama 7B Q4_0 | pp512 | 201.59 \u00b1 6.84; llama 7B Q4_0 | tg128 | 14.94 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14798753",
   "tg_tokens": 128,
   "tps": 14.94,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-27",
   "flags": [],
   "hardware": "Radeon 890M",
   "id": "lc-disc-10879-c14798753-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06854",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 246.65,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 890M; llama 7B Q4_0 | pp512 | 246.65 \u00b1 2.12; llama 7B Q4_0 | tg128 | 15.11 \u00b1 0.46",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14798753",
   "tg_tokens": 128,
   "tps": 15.11,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-27",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon RX 480 G",
   "id": "lc-disc-10879-c14798753-llama7b-Q4_0-5",
   "model": "llama 7B",
   "notes": "threads=100; config=b06854",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 103.24,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 480 G; dev=VULKAN0; llama 7B Q4_0 | pp512 | 103.24 \u00b1 0.17; llama 7B Q4_0 | tg128 | 31.23 \u00b1 0.19",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14798753",
   "tg_tokens": 128,
   "tps": 31.23,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-27",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon RX 480 G",
   "id": "lc-disc-10879-c14798753-llama7b-Q4_0-6",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06854",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 102.55,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 480 G; dev=VULKAN0; llama 7B Q4_0 | pp512 | 102.55 \u00b1 0.13; llama 7B Q4_0 | tg128 | 31.53 \u00b1 0.18",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14798753",
   "tg_tokens": 128,
   "tps": 31.53,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-27",
   "flags": [],
   "hardware": "Radeon RX 480 G",
   "id": "lc-disc-10879-c14798753-llama7b-Q4_0-7",
   "model": "llama 7B",
   "notes": "threads=100; config=b06854",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 107.79,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 480 G; dev=VULKAN1; llama 7B Q4_0 | pp512 | 107.79 \u00b1 0.01; llama 7B Q4_0 | tg128 | 39.93 \u00b1 0.10",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14798753",
   "tg_tokens": 128,
   "tps": 39.93,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-27",
   "flags": [],
   "hardware": "Radeon RX 480 G",
   "id": "lc-disc-10879-c14798753-llama7b-Q4_0-8",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06854",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 106.39,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 480 G; dev=VULKAN1; llama 7B Q4_0 | pp512 | 106.39 \u00b1 0.01; llama 7B Q4_0 | tg128 | 40.40 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14798753",
   "tg_tokens": 128,
   "tps": 40.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-04",
   "flags": [],
   "hardware": "Radeon 780M",
   "id": "lc-disc-10879-c14874902-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b06949",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 266.19,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 780M; llama 7B Q4_0 | pp512 | 266.19 \u00b1 1.36; llama 7B Q4_0 | tg128 | 20.53 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14874902",
   "tg_tokens": 128,
   "tps": 20.53,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-04",
   "flags": [],
   "hardware": "Radeon 780M",
   "id": "lc-disc-10879-c14874902-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06949",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 297.39,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 780M; llama 7B Q4_0 | pp512 | 297.39 \u00b1 1.22; llama 7B Q4_0 | tg128 | 20.59 \u00b1 0.38",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14874902",
   "tg_tokens": 128,
   "tps": 20.59,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-04",
   "flags": [],
   "hardware": "Radeon 780M",
   "id": "lc-disc-10879-c14874902-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; config=b06949",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 326.84,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 780M; llama 7B Q4_0 | pp512 | 326.84 \u00b1 1.76; llama 7B Q4_0 | tg128 | 19.66 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14874902",
   "tg_tokens": 128,
   "tps": 19.66,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-04",
   "flags": [],
   "hardware": "Radeon 780M",
   "id": "lc-disc-10879-c14874902-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06949",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 349.57,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 780M; llama 7B Q4_0 | pp512 | 349.57 \u00b1 0.69; llama 7B Q4_0 | tg128 | 19.81 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14874902",
   "tg_tokens": 128,
   "tps": 19.81,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-04",
   "flags": [],
   "hardware": "RTX 4060",
   "id": "lc-disc-10879-c14874924-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b06949",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2135.66,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 4060; llama 7B Q4_0 | pp512 | 2135.66 \u00b1 23.18; llama 7B Q4_0 | tg128 | 59.53 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14874924",
   "tg_tokens": 128,
   "tps": 59.53,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-04",
   "flags": [],
   "hardware": "RTX 4060",
   "id": "lc-disc-10879-c14874924-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06949",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2358.03,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 4060; llama 7B Q4_0 | pp512 | 2358.03 \u00b1 12.17; llama 7B Q4_0 | tg128 | 60.01 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14874924",
   "tg_tokens": 128,
   "tps": 60.01,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-05",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c14882060-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b06957",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3531.93,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3531.93 \u00b1 31.74; llama 7B Q4_0 | tg128 | 191.28 \u00b1 0.20",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14882060",
   "tg_tokens": 128,
   "tps": 191.28,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-05",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c14882060-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06957",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3332.9,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3332.90 \u00b1 11.47; llama 7B Q4_0 | tg128 | 195.30 \u00b1 0.23",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14882060",
   "tg_tokens": 128,
   "tps": 195.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-07",
   "flags": [],
   "hardware": "M2",
   "id": "lc-disc-10879-c14901609-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b00279",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 50.79,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M2; llama 7B Q4_0 | pp512 | 50.79 \u00b1 0.16; llama 7B Q4_0 | tg128 | 13.50 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14901609",
   "tg_tokens": 128,
   "tps": 13.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-07",
   "flags": [],
   "hardware": "M2",
   "id": "lc-disc-10879-c14901609-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b00279",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 46.81,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M2; llama 7B Q4_0 | pp512 | 46.81 \u00b1 0.08; llama 7B Q4_0 | tg128 | 12.25 \u00b1 2.30",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14901609",
   "tg_tokens": 128,
   "tps": 12.25,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-07",
   "flags": [],
   "hardware": "Arc A750",
   "id": "lc-disc-10879-c14904580-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b00279",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1075.94,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A750; llama 7B Q4_0 | pp512 | 1075.94 \u00b1 13.89; llama 7B Q4_0 | tg128 | 42.66 \u00b1 0.18",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14904580",
   "tg_tokens": 128,
   "tps": 42.66,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-07",
   "flags": [],
   "hardware": "Arc A750",
   "id": "lc-disc-10879-c14904580-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b00279",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 303.37,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A750; llama 7B Q4_0 | pp512 | 303.37 \u00b1 1.44; llama 7B Q4_0 | tg128 | 43.96 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14904580",
   "tg_tokens": 128,
   "tps": 43.96,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-07",
   "flags": [],
   "hardware": "Arc A750",
   "id": "lc-disc-10879-c14904580-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100; config=b00279",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1631.21,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A750; llama 7B Q4_0 | pp512 | 1631.21 \u00b1 5.81; llama 7B Q4_0 | tg128 | 36.20 \u00b1 0.18",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14904580",
   "tg_tokens": 128,
   "tps": 36.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-07",
   "flags": [],
   "hardware": "Arc A750",
   "id": "lc-disc-10879-c14904580-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b00279",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 506.22,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A750; llama 7B Q4_0 | pp512 | 506.22 \u00b1 3.80; llama 7B Q4_0 | tg128 | 23.61 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14904580",
   "tg_tokens": 128,
   "tps": 23.61,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (OpenCL)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-07",
   "flags": [],
   "hardware": "A750",
   "id": "lc-disc-10879-c14904580-llama7b-Q4_0-5",
   "model": "llama 7B",
   "notes": "threads=100; config=b00279",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 234.63,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A750; llama 7B Q4_0 | pp512 | 234.63 \u00b1 0.59; llama 7B Q4_0 | tg128 | 40.70 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14904580",
   "tg_tokens": 128,
   "tps": 40.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (OpenCL)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-07",
   "flags": [],
   "hardware": "A750",
   "id": "lc-disc-10879-c14904580-llama7b-Q4_0-6",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b00279",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 162.56,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A750; llama 7B Q4_0 | pp512 | 162.56 \u00b1 0.18; llama 7B Q4_0 | tg128 | 37.42 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14904580",
   "tg_tokens": 128,
   "tps": 37.42,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-12",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-10879-c14950915-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=16",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 308.17,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; llama 7B Q4_0 | pp512 | 308.17 \u00b1 1.49; llama 7B Q4_0 | tg128 | 27.58 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14950915",
   "tg_tokens": 128,
   "tps": 27.58,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-12",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-10879-c14950915-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=16; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 304.53,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; llama 7B Q4_0 | pp512 | 304.53 \u00b1 5.85; llama 7B Q4_0 | tg128 | 29.28 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14950915",
   "tg_tokens": 128,
   "tps": 29.28,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-12",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-10879-c14950915-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1023.92,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; llama 7B Q4_0 | pp512 | 1023.92 \u00b1 3.47; llama 7B Q4_0 | tg128 | 51.67 \u00b1 0.14",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14950915",
   "tg_tokens": 128,
   "tps": 51.67,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-12",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-10879-c14950915-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1083.51,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; llama 7B Q4_0 | pp512 | 1083.51 \u00b1 5.14; llama 7B Q4_0 | tg128 | 53.37 \u00b1 0.15",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14950915",
   "tg_tokens": 128,
   "tps": 53.37,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-15",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-10879-c14977811-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b07070",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1009.02,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; llama 7B Q4_0 | pp512 | 1009.02 \u00b1 2.03; llama 7B Q4_0 | tg128 | 46.68 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14977811",
   "tg_tokens": 128,
   "tps": 46.68,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-15",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-10879-c14977811-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07070",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1118.89,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; llama 7B Q4_0 | pp512 | 1118.89 \u00b1 4.92; llama 7B Q4_0 | tg128 | 50.69 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14977811",
   "tg_tokens": 128,
   "tps": 50.69,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-15",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-10879-c14977811-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100; config=b07070",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1004.84,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; llama 7B Q4_0 | pp512 | 1004.84 \u00b1 1.90; llama 7B Q4_0 | tg128 | 46.65 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14977811",
   "tg_tokens": 128,
   "tps": 46.65,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-15",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-10879-c14977811-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07070",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1117.55,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; llama 7B Q4_0 | pp512 | 1117.55 \u00b1 3.14; llama 7B Q4_0 | tg128 | 50.75 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14977811",
   "tg_tokens": 128,
   "tps": 50.75,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c14975049-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100; config=b07062",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1892.58,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 1892.58 \u00b1 15.25; llama 7B Q4_0 | tg128 | 119.69 \u00b1 0.61",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14975049",
   "tg_tokens": 128,
   "tps": 119.69,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c14975049-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b07062",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1941.47,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 1941.47 \u00b1 9.99; llama 7B Q4_0 | tg128 | 123.70 \u00b1 0.22",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14975049",
   "tg_tokens": 128,
   "tps": 123.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-15",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c14975049-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100; config=b07062",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2162.71,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2162.71 \u00b1 18.29; llama 7B Q4_0 | tg128 | 159.46 \u00b1 0.36",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14975049",
   "tg_tokens": 128,
   "tps": 159.46,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-15",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c14975049-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b07062",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2148.72,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2148.72 \u00b1 14.88; llama 7B Q4_0 | tg128 | 165.38 \u00b1 0.33",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14975049",
   "tg_tokens": 128,
   "tps": 165.38,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c14975049-llama7b-Q4_0-5",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100; config=b07062",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2222.31,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2222.31 \u00b1 22.29; llama 7B Q4_0 | tg128 | 143.75 \u00b1 0.10",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14975049",
   "tg_tokens": 128,
   "tps": 143.75,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c14975049-llama7b-Q4_0-6",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b07062",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2369.54,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2369.54 \u00b1 28.15; llama 7B Q4_0 | tg128 | 150.34 \u00b1 0.12",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14975049",
   "tg_tokens": 128,
   "tps": 150.34,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-11",
   "flags": [],
   "hardware": "Radeon RX 6900 XT",
   "id": "lc-disc-10879-c14068774-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=12; config=b06123",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 326.29,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6900 XT; llama 7B Q4_0 | pp512 | 326.29 \u00b1 0.64; llama 7B Q4_0 | tg128 | 98.27 \u00b1 0.80",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14068774",
   "tg_tokens": 128,
   "tps": 98.27,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-11",
   "flags": [],
   "hardware": "Radeon RX 6900 XT",
   "id": "lc-disc-10879-c14068774-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=12; config=b06123",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 643.37,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6900 XT; llama 7B Q4_0 | pp512 | 643.37 \u00b1 0.65; llama 7B Q4_0 | tg128 | 82.26 \u00b1 0.20",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14068774",
   "tg_tokens": 128,
   "tps": 82.26,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-14",
   "flags": [],
   "hardware": "Radeon 890M",
   "id": "lc-disc-10879-c14102867-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b06156",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 257.01,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 890M; llama 7B Q4_0 | pp512 | 257.01 \u00b1 0.39; llama 7B Q4_0 | tg128 | 19.93 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14102867",
   "tg_tokens": 128,
   "tps": 19.93,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-14",
   "flags": [],
   "hardware": "Radeon 890M",
   "id": "lc-disc-10879-c14102867-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b06156",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 264.67,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 890M; llama 7B Q4_0 | pp512 | 264.67 \u00b1 3.26; llama 7B Q4_0 | tg128 | 19.81 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14102867",
   "tg_tokens": 128,
   "tps": 19.81,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-15",
   "flags": [],
   "hardware": "Radeon RX 6600",
   "id": "lc-disc-10879-c14111848-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b06170",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 617.85,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6600; llama 7B Q4_0 | pp512 | 617.85 \u00b1 0.28; llama 7B Q4_0 | tg128 | 48.52 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14111848",
   "tg_tokens": 128,
   "tps": 48.52,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-15",
   "flags": [],
   "hardware": "Radeon RX 6600",
   "id": "lc-disc-10879-c14111848-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06170",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 622.72,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6600; llama 7B Q4_0 | pp512 | 622.72 \u00b1 0.20; llama 7B Q4_0 | tg128 | 48.31 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14111848",
   "tg_tokens": 128,
   "tps": 48.31,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-15",
   "flags": [],
   "hardware": "Radeon RX 9070",
   "id": "lc-disc-10879-c14111848-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100; config=b06170",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3164.1,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070; llama 7B Q4_0 | pp512 | 3164.10 \u00b1 66.84; llama 7B Q4_0 | tg128 | 119.71 \u00b1 3.40",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14111848",
   "tg_tokens": 128,
   "tps": 119.71,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-15",
   "flags": [],
   "hardware": "Radeon RX 9070",
   "id": "lc-disc-10879-c14111848-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06170",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2859.98,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070; llama 7B Q4_0 | pp512 | 2859.98 \u00b1 31.53; llama 7B Q4_0 | tg128 | 119.51 \u00b1 0.13",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14111848",
   "tg_tokens": 128,
   "tps": 119.51,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-17",
   "flags": [],
   "hardware": "M1",
   "id": "lc-disc-10879-c14130306-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; 20.35 W measured total system power (hwmon), tg test; config=b06182",
   "params": null,
   "power_w": 20.35,
   "pp_tokens": 512,
   "pp_tps": 31.31,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M1; llama 7B Q4_0 | pp512 | 31.31 \u00b1 0.01; llama 7B Q4_0 | tg128 | 12.41 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14130306",
   "tg_tokens": 128,
   "tps": 12.41,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-17",
   "flags": [],
   "hardware": "M1",
   "id": "lc-disc-10879-c14130306-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06182",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 28.65,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M1; llama 7B Q4_0 | pp512 | 28.65 \u00b1 0.00; llama 7B Q4_0 | tg128 | 12.38 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14130306",
   "tg_tokens": 128,
   "tps": 12.38,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-17",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "lc-disc-10879-c14130339-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; 36.5 W measured total system power (hwmon), tg test; config=b06182",
   "params": null,
   "power_w": 36.5,
   "pp_tokens": 512,
   "pp_tps": 62.7,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M2 Pro; llama 7B Q4_0 | pp512 | 62.70 \u00b1 0.03; llama 7B Q4_0 | tg128 | 20.95 \u00b1 0.11",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14130339",
   "tg_tokens": 128,
   "tps": 20.95,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-17",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "lc-disc-10879-c14130339-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06182",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 58.86,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M2 Pro; llama 7B Q4_0 | pp512 | 58.86 \u00b1 0.02; llama 7B Q4_0 | tg128 | 20.97 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14130339",
   "tg_tokens": 128,
   "tps": 20.97,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-10879-c14130537-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b06183",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1121.72,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; llama 7B Q4_0 | pp512 | 1121.72 \u00b1 4.04; llama 7B Q4_0 | tg128 | 46.69 \u00b1 0.13",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14130537",
   "tg_tokens": 128,
   "tps": 46.69,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-17",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-10879-c14130537-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b06183",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1347.94,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; llama 7B Q4_0 | pp512 | 1347.94 \u00b1 6.65; llama 7B Q4_0 | tg128 | 46.09 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14130537",
   "tg_tokens": 128,
   "tps": 46.09,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-17",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-10879-c14130537-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b06183",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 803.41,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; llama 7B Q4_0 | pp512 | 803.41 \u00b1 3.48; llama 7B Q4_0 | tg128 | 51.89 \u00b1 0.12",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14130537",
   "tg_tokens": 128,
   "tps": 51.89,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-17",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-10879-c14130537-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b06183",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 927.96,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; llama 7B Q4_0 | pp512 | 927.96 \u00b1 1.79; llama 7B Q4_0 | tg128 | 51.14 \u00b1 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14130537",
   "tg_tokens": 128,
   "tps": 51.14,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-17",
   "flags": [],
   "hardware": "Arc B570",
   "id": "lc-disc-10879-c14133224-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100; config=b06188",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 469.46,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc B570; llama 7B Q4_0 | pp512 | 469.46 \u00b1 1.06; llama 7B Q4_0 | tg128 | 42.70 \u00b1 0.46",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14133224",
   "tg_tokens": 128,
   "tps": 42.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-17",
   "flags": [],
   "hardware": "Arc B570",
   "id": "lc-disc-10879-c14133224-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b06188",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 285.36,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc B570; llama 7B Q4_0 | pp512 | 285.36 \u00b1 0.13; llama 7B Q4_0 | tg128 | 43.38 \u00b1 0.13",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14133224",
   "tg_tokens": 128,
   "tps": 43.38,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "Arc B570",
   "id": "lc-disc-10879-c14133224-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100; config=b06188",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 468.82,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc B570; llama 7B Q4_0 | pp512 | 468.82 \u00b1 0.50; llama 7B Q4_0 | tg128 | 31.31 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14133224",
   "tg_tokens": 128,
   "tps": 31.31,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "Arc B570",
   "id": "lc-disc-10879-c14133224-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b06188",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 285.34,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc B570; llama 7B Q4_0 | pp512 | 285.34 \u00b1 0.73; llama 7B Q4_0 | tg128 | 31.74 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14133224",
   "tg_tokens": 128,
   "tps": 31.74,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-18",
   "flags": [],
   "hardware": "Arc A750",
   "id": "lc-disc-10879-c14135790-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b06188",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 665.38,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A750; llama 7B Q4_0 | pp512 | 665.38 \u00b1 5.49; llama 7B Q4_0 | tg128 | 41.43 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14135790",
   "tg_tokens": 128,
   "tps": 41.43,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-18",
   "flags": [],
   "hardware": "Arc A750",
   "id": "lc-disc-10879-c14135790-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06188",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 312.51,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A750; llama 7B Q4_0 | pp512 | 312.51 \u00b1 1.74; llama 7B Q4_0 | tg128 | 42.15 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14135790",
   "tg_tokens": 128,
   "tps": 42.15,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-23",
   "flags": [],
   "hardware": "Radeon RX 7900 XT",
   "id": "lc-disc-10879-c14199059-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b06257",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1619.49,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XT; llama 7B Q4_0 | pp512 | 1619.49 \u00b1 8.06; llama 7B Q4_0 | tg128 | 135.65 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14199059",
   "tg_tokens": 128,
   "tps": 135.65,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-23",
   "flags": [],
   "hardware": "Radeon RX 7900 XT",
   "id": "lc-disc-10879-c14199059-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06257",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1619.74,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XT; llama 7B Q4_0 | pp512 | 1619.74 \u00b1 6.71; llama 7B Q4_0 | tg128 | 137.78 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14199059",
   "tg_tokens": 128,
   "tps": 137.78,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-23",
   "flags": [],
   "hardware": "Radeon RX 7900 XT",
   "id": "lc-disc-10879-c14199059-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100; config=b06257",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2076.96,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XT; llama 7B Q4_0 | pp512 | 2076.96 \u00b1 3.82; llama 7B Q4_0 | tg128 | 117.57 \u00b1 0.13",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14199059",
   "tg_tokens": 128,
   "tps": 117.57,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-23",
   "flags": [],
   "hardware": "Radeon RX 7900 XT",
   "id": "lc-disc-10879-c14199059-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b06257",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1662.15,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XT; llama 7B Q4_0 | pp512 | 1662.15 \u00b1 3.48; llama 7B Q4_0 | tg128 | 116.94 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14199059",
   "tg_tokens": 128,
   "tps": 116.94,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-11",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-10879-c14652697-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4046.36,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 4046.36 \u00b1 26.87; llama 7B Q4_0 | tg128 | 128.53 \u00b1 0.42",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14652697",
   "tg_tokens": 128,
   "tps": 128.53,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-11",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-10879-c14652697-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=0 ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4124.63,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 4124.63 \u00b1 13.33; llama 7B Q4_0 | tg128 | 127.38 \u00b1 1.33",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14652697",
   "tg_tokens": 128,
   "tps": 127.38,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-11",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-disc-10879-c14654488-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b06692",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 10381.64,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 5090; llama 7B Q4_0 | pp512 | 10381.64 \u00b1 508.84; llama 7B Q4_0 | tg128 | 263.63 \u00b1 0.91",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14654488",
   "tg_tokens": 128,
   "tps": 263.63,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-11",
   "flags": [],
   "hardware": "RTX 5090",
   "id": "lc-disc-10879-c14654488-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06692",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 11796.38,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 5090; llama 7B Q4_0 | pp512 | 11796.38 \u00b1 601.36; llama 7B Q4_0 | tg128 | 273.68 \u00b1 0.52",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14654488",
   "tg_tokens": 128,
   "tps": 273.68,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-16",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c14697101-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100; config=b06400",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1997.19,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 1997.19 \u00b1 7.76; llama 7B Q4_0 | tg128 | 163.00 \u00b1 0.76",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14697101",
   "tg_tokens": 128,
   "tps": 163.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-16",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c14697101-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b06400",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2044.99,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2044.99 \u00b1 14.06; llama 7B Q4_0 | tg128 | 166.25 \u00b1 0.19",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14697101",
   "tg_tokens": 128,
   "tps": 166.25,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-16",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c14697101-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100; config=b06400",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2254.0,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2254.00 \u00b1 9.48; llama 7B Q4_0 | tg128 | 162.16 \u00b1 0.11",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14697101",
   "tg_tokens": 128,
   "tps": 162.16,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-16",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-10879-c14697101-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b06400",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2281.62,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2281.62 \u00b1 25.37; llama 7B Q4_0 | tg128 | 165.08 \u00b1 0.27",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14697101",
   "tg_tokens": 128,
   "tps": 165.08,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-18",
   "flags": [],
   "hardware": "M2",
   "id": "lc-disc-10879-c14716853-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b06794",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 796.29,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M2; llama 7B Q4_0 | pp512 | 796.29 _ 5.84; llama 7B Q4_0 | tg128 | 105.06 \u00b1 0.27",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14716853",
   "tg_tokens": 128,
   "tps": 105.06,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-18",
   "flags": [],
   "hardware": "M2",
   "id": "lc-disc-10879-c14716853-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b06794",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 792.74,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "M2; llama 7B Q4_0 | pp512 | 792.74 \u00b1 4.30; llama 7B Q4_0 | tg128 | 109.21 \u00b1 0.72",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14716853",
   "tg_tokens": 128,
   "tps": 109.21,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-18",
   "flags": [],
   "hardware": "RTX 4080 SUPER",
   "id": "lc-disc-10879-c14717139-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b06729",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 7101.18,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 4080 Super; llama 7B Q4_0 | pp512 | 7101.18 \u00b1 269.79; llama 7B Q4_0 | tg128 | 147.13 \u00b1 5.64",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14717139",
   "tg_tokens": 128,
   "tps": 147.13,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-18",
   "flags": [],
   "hardware": "RTX 4080 SUPER",
   "id": "lc-disc-10879-c14717139-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06729",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 8007.37,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 4080 Super; llama 7B Q4_0 | pp512 | 8007.37 \u00b1 46.03; llama 7B Q4_0 | tg128 | 150.20 \u00b1 0.26",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14717139",
   "tg_tokens": 128,
   "tps": 150.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-19",
   "flags": [],
   "hardware": "Radeon 660M",
   "id": "lc-disc-10879-c14720349-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b06795",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 138.61,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 660M; llama 7B Q4_0 | pp512 | 138.61 \u00b1 0.15; llama 7B Q4_0 | tg128 | 16.64 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14720349",
   "tg_tokens": 128,
   "tps": 16.64,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-19",
   "flags": [],
   "hardware": "Radeon 660M",
   "id": "lc-disc-10879-c14720349-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06795",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 148.05,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 660M; llama 7B Q4_0 | pp512 | 148.05 \u00b1 0.05; llama 7B Q4_0 | tg128 | 16.56 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14720349",
   "tg_tokens": 128,
   "tps": 16.56,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-19",
   "flags": [],
   "hardware": "Radeon 680M",
   "id": "lc-disc-10879-c14720658-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b06795",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 240.89,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 680M; llama 7B Q4_0 | pp512 | 240.89 \u00b1 0.52; llama 7B Q4_0 | tg128 | 21.26 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14720658",
   "tg_tokens": 128,
   "tps": 21.26,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-19",
   "flags": [],
   "hardware": "Radeon 680M",
   "id": "lc-disc-10879-c14720658-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06795",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 277.91,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 680M; llama 7B Q4_0 | pp512 | 277.91 \u00b1 0.37; llama 7B Q4_0 | tg128 | 21.15 \u00b1 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14720658",
   "tg_tokens": 128,
   "tps": 21.15,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-20",
   "flags": [],
   "hardware": "RTX 3070",
   "id": "lc-disc-10879-c14731155-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b06783",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1689.4,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3070; llama 7B Q4_0 | pp512 | 1689.40 \u00b1 19.57; llama 7B Q4_0 | tg128 | 63.64 \u00b1 0.39",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14731155",
   "tg_tokens": 128,
   "tps": 63.64,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-20",
   "flags": [],
   "hardware": "RTX 3070",
   "id": "lc-disc-10879-c14731155-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06783",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1832.07,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3070; llama 7B Q4_0 | pp512 | 1832.07 \u00b1 57.14; llama 7B Q4_0 | tg128 | 62.92 \u00b1 0.37",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14731155",
   "tg_tokens": 128,
   "tps": 62.92,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-23",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-10879-c14767089-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b06829",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 972.06,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; llama 7B Q4_0 | pp512 | 972.06 \u00b1 3.92; llama 7B Q4_0 | tg128 | 52.28 \u00b1 0.14",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14767089",
   "tg_tokens": 128,
   "tps": 52.28,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-23",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-10879-c14767089-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06829",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1054.53,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; llama 7B Q4_0 | pp512 | 1054.53 \u00b1 8.67; llama 7B Q4_0 | tg128 | 51.62 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14767089",
   "tg_tokens": 128,
   "tps": 51.62,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-24",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-10879-c14768745-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=16; config=b06830",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1120.83,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; llama 7B Q4_0 | pp512 | 1120.83 \u00b1 2.57; llama 7B Q4_0 | tg128 | 53.21 \u00b1 0.38",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14768745",
   "tg_tokens": 128,
   "tps": 53.21,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-24",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-10879-c14768745-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=16; config=b06830",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1243.31,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; llama 7B Q4_0 | pp512 | 1243.31 \u00b1 2.12; llama 7B Q4_0 | tg128 | 52.69 \u00b1 0.10",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14768745",
   "tg_tokens": 128,
   "tps": 52.69,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-24",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-10879-c14768745-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=16; config=b06830",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1264.6,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; llama 7B Q4_0 | pp512 | 1264.60 \u00b1 2.36; llama 7B Q4_0 | tg128 | 52.50 \u00b1 0.13",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14768745",
   "tg_tokens": 128,
   "tps": 52.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-24",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-10879-c14768745-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=16; config=b06830",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1305.74,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; llama 7B Q4_0 | pp512 | 1305.74 \u00b1 2.58; llama 7B Q4_0 | tg128 | 52.03 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14768745",
   "tg_tokens": 128,
   "tps": 52.03,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-24",
   "flags": [],
   "hardware": "Radeon RX 480 G",
   "id": "lc-disc-10879-c14772967-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100; config=b06833",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 201.82,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 480 G; dev=VULKAN2; llama 7B Q4_0 | pp512 | 201.82 \u00b1 0.73; llama 7B Q4_0 | tg128 | 36.49 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14772967",
   "tg_tokens": 128,
   "tps": 36.49,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-24",
   "flags": [],
   "hardware": "Radeon RX 480 G",
   "id": "lc-disc-10879-c14772967-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b06833",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 194.52,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 480 G; dev=VULKAN2; llama 7B Q4_0 | pp512 | 194.52 \u00b1 0.61; llama 7B Q4_0 | tg128 | 37.23 \u00b1 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14772967",
   "tg_tokens": 128,
   "tps": 37.23,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-23",
   "flags": [],
   "hardware": "A2000",
   "id": "lc-disc-10879-c14199068-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b06257",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1259.66,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A2000; llama 7B Q4_0 | pp512 | 1259.66 \u00b1 4.67; llama 7B Q4_0 | tg128 | 51.21 \u00b1 0.30",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14199068",
   "tg_tokens": 128,
   "tps": 51.21,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-23",
   "flags": [],
   "hardware": "A2000",
   "id": "lc-disc-10879-c14199068-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06257",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1411.42,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A2000; llama 7B Q4_0 | pp512 | 1411.42 \u00b1 8.17; llama 7B Q4_0 | tg128 | 51.96 \u00b1 0.29",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14199068",
   "tg_tokens": 128,
   "tps": 51.96,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-23",
   "flags": [],
   "hardware": "A2000",
   "id": "lc-disc-10879-c14199068-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100; config=b06257",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1245.19,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A2000; llama 7B Q4_0 | pp512 | 1245.19 \u00b1 8.76; llama 7B Q4_0 | tg128 | 45.52 \u00b1 0.54",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14199068",
   "tg_tokens": 128,
   "tps": 45.52,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-23",
   "flags": [],
   "hardware": "A2000",
   "id": "lc-disc-10879-c14199068-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b06257",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1361.85,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A2000; llama 7B Q4_0 | pp512 | 1361.85 \u00b1 3.26; llama 7B Q4_0 | tg128 | 45.69 \u00b1 0.20",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14199068",
   "tg_tokens": 128,
   "tps": 45.69,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-27",
   "flags": [],
   "hardware": "Radeon RX 580",
   "id": "lc-disc-10879-c14229112-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 100.35,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 580; llama 7B Q4_0 | pp512 | 100.35 \u00b1 0.29; llama 7B Q4_0 | tg128 | 16.08 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14229112",
   "tg_tokens": 128,
   "tps": 16.08,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-27",
   "flags": [],
   "hardware": "Radeon RX 580",
   "id": "lc-disc-10879-c14229112-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 85.37,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 580; llama 7B Q4_0 | pp512 | 85.37 \u00b1 0.57; llama 7B Q4_0 | tg128 | 13.99 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14229112",
   "tg_tokens": 128,
   "tps": 13.99,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-26",
   "flags": [],
   "hardware": "Radeon RX 580",
   "id": "lc-disc-10879-c15089176-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b06763",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 180.49,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 580; llama 7B Q4_0 | pp512 | 180.49 \u00b1 2.95; llama 7B Q4_0 | tg128 | 35.79 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15089176",
   "tg_tokens": 128,
   "tps": 35.79,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-26",
   "flags": [],
   "hardware": "Radeon RX 580",
   "id": "lc-disc-10879-c15089176-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06763",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 157.61,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 580; llama 7B Q4_0 | pp512 | 157.61 \u00b1 0.72; llama 7B Q4_0 | tg128 | 35.59 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-15089176",
   "tg_tokens": 128,
   "tps": 35.59,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-27",
   "flags": [],
   "hardware": "RTX 2080 Ti",
   "id": "lc-disc-10879-c14234638-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1888.24,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 2080 Ti; llama 7B Q4_0 | pp512 | 1888.24 \u00b1 9.20; llama 7B Q4_0 | tg128 | 97.58 \u00b1 6.60",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14234638",
   "tg_tokens": 128,
   "tps": 97.58,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-27",
   "flags": [],
   "hardware": "RTX 2080 Ti",
   "id": "lc-disc-10879-c14234638-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1936.25,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 2080 Ti; llama 7B Q4_0 | pp512 | 1936.25 \u00b1 32.08; llama 7B Q4_0 | tg128 | 100.99 \u00b1 0.24",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14234638",
   "tg_tokens": 128,
   "tps": 100.99,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-01",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc-disc-10879-c14278864-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 912.47,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3090; llama 7B Q4_0 | pp512 | 912.47 \u00b1 1.06; llama 7B Q4_0 | tg128 | 106.03 \u00b1 0.89",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14278864",
   "tg_tokens": 128,
   "tps": 106.03,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-01",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc-disc-10879-c14278864-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 783.94,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3090; llama 7B Q4_0 | pp512 | 783.94 \u00b1 0.77; llama 7B Q4_0 | tg128 | 108.45 \u00b1 0.48",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14278864",
   "tg_tokens": 128,
   "tps": 108.45,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-01",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc-disc-10879-c14278864-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 997.21,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3090; llama 7B Q4_0 | pp512 | 997.21 \u00b1 0.39; llama 7B Q4_0 | tg128 | 83.55 \u00b1 0.12",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14278864",
   "tg_tokens": 128,
   "tps": 83.55,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-01",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc-disc-10879-c14278864-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 426.56,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3090; llama 7B Q4_0 | pp512 | 426.56 \u00b1 0.41; llama 7B Q4_0 | tg128 | 73.58 \u00b1 0.26",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14278864",
   "tg_tokens": 128,
   "tps": 73.58,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-01",
   "flags": [],
   "hardware": "Radeon RX 6800 XT",
   "id": "lc-disc-10879-c14278864-llama7b-Q4_0-5",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1752.92,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6800 XT; llama 7B Q4_0 | pp512 | 1752.92 \u00b1 1.71; llama 7B Q4_0 | tg128 | 100.32 \u00b1 0.97",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14278864",
   "tg_tokens": 128,
   "tps": 100.32,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-01",
   "flags": [],
   "hardware": "Radeon RX 6800 XT",
   "id": "lc-disc-10879-c14278864-llama7b-Q4_0-6",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1704.79,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6800 XT; llama 7B Q4_0 | pp512 | 1704.79 \u00b1 0.71; llama 7B Q4_0 | tg128 | 100.50 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14278864",
   "tg_tokens": 128,
   "tps": 100.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-01",
   "flags": [],
   "hardware": "Radeon RX 6800 XT",
   "id": "lc-disc-10879-c14278864-llama7b-Q4_0-7",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1706.8,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6800 XT; llama 7B Q4_0 | pp512 | 1706.80 \u00b1 3.55; llama 7B Q4_0 | tg128 | 88.31 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14278864",
   "tg_tokens": 128,
   "tps": 88.31,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-01",
   "flags": [],
   "hardware": "Radeon RX 6800 XT",
   "id": "lc-disc-10879-c14278864-llama7b-Q4_0-8",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1199.27,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6800 XT; llama 7B Q4_0 | pp512 | 1199.27 \u00b1 0.90; llama 7B Q4_0 | tg128 | 84.88 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14278864",
   "tg_tokens": 128,
   "tps": 84.88,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-01",
   "flags": [],
   "hardware": "Radeon RX 6800 XT",
   "id": "lc-disc-10879-c14278864-llama7b-Q4_0-9",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 488.06,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6800 XT; llama 7B Q4_0 | pp512 | 488.06 \u00b1 0.27; llama 7B Q4_0 | tg128 | 59.36 \u00b1 0.16",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14278864",
   "tg_tokens": 128,
   "tps": 59.36,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-01",
   "flags": [],
   "hardware": "Radeon RX 6800 XT",
   "id": "lc-disc-10879-c14278864-llama7b-Q4_0-10",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 484.37,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6800 XT; llama 7B Q4_0 | pp512 | 484.37 \u00b1 0.27; llama 7B Q4_0 | tg128 | 59.22 \u00b1 0.15",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14278864",
   "tg_tokens": 128,
   "tps": 59.22,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-01",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc-disc-10879-c14278864-llama7b-Q4_0-11",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4409.02,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3090; llama 7B Q4_0 | pp512 | 4409.02 \u00b1 80.38; llama 7B Q4_0 | tg128 | 157.82 \u00b1 0.27",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14278864",
   "tg_tokens": 128,
   "tps": 157.82,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-01",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc-disc-10879-c14278864-llama7b-Q4_0-12",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4892.33,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3090; llama 7B Q4_0 | pp512 | 4892.33 \u00b1 39.53; llama 7B Q4_0 | tg128 | 156.57 \u00b1 0.46",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14278864",
   "tg_tokens": 128,
   "tps": 156.57,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-01",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc-disc-10879-c14278864-llama7b-Q4_0-13",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 5100.68,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3090; llama 7B Q4_0 | pp512 | 5100.68 \u00b1 10.65; llama 7B Q4_0 | tg128 | 159.36 \u00b1 0.35",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14278864",
   "tg_tokens": 128,
   "tps": 159.36,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-01",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc-disc-10879-c14278864-llama7b-Q4_0-14",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 5781.6,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 3090; llama 7B Q4_0 | pp512 | 5781.60 \u00b1 5.86; llama 7B Q4_0 | tg128 | 163.69 \u00b1 0.28",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14278864",
   "tg_tokens": 128,
   "tps": 163.69,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-01",
   "flags": [],
   "hardware": "Arc A770",
   "id": "lc-disc-10879-c14278864-llama7b-Q4_0-15",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 762.02,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A770; llama 7B Q4_0 | pp512 | 762.02 \u00b1 0.72; llama 7B Q4_0 | tg128 | 49.55 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14278864",
   "tg_tokens": 128,
   "tps": 49.55,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-01",
   "flags": [],
   "hardware": "Arc A770",
   "id": "lc-disc-10879-c14278864-llama7b-Q4_0-16",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 244.24,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A770; llama 7B Q4_0 | pp512 | 244.24 \u00b1 0.22; llama 7B Q4_0 | tg128 | 50.39 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14278864",
   "tg_tokens": 128,
   "tps": 50.39,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-07",
   "flags": [],
   "hardware": "Radeon RX 580 S",
   "id": "lc-disc-10879-c14333193-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100; config=b06405",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 131.42,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 580 S; llama 7B Q4_0 | pp512 | 131.42 \u00b1 4.36; llama 7B Q4_0 | tg128 | 12.72 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14333193",
   "tg_tokens": 128,
   "tps": 12.72,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-07",
   "flags": [],
   "hardware": "Radeon RX 580 S",
   "id": "lc-disc-10879-c14333193-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b06405",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 177.33,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 580 S; llama 7B Q4_0 | pp512 | 177.33 \u00b1 0.53; llama 7B Q4_0 | tg128 | 11.62 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14333193",
   "tg_tokens": 128,
   "tps": 11.62,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-08",
   "flags": [],
   "hardware": "Radeon RX 470 G",
   "id": "lc-disc-10879-c14335461-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b06293",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 192.0,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 470 G; llama 7B Q4_0 | pp512 | 192.00 \u00b1 0.88; llama 7B Q4_0 | tg128 | 38.21 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14335461",
   "tg_tokens": 128,
   "tps": 38.21,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-08",
   "flags": [],
   "hardware": "Radeon RX 470 G",
   "id": "lc-disc-10879-c14335461-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06293",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 182.13,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 470 G; llama 7B Q4_0 | pp512 | 182.13 \u00b1 0.35; llama 7B Q4_0 | tg128 | 37.15 \u00b1 0.20",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14335461",
   "tg_tokens": 128,
   "tps": 37.15,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-09",
   "flags": [],
   "hardware": "Radeon 660M",
   "id": "lc-disc-10879-c14347628-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b06425",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 136.56,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 660M; llama 7B Q4_0 | pp512 | 136.56 \u00b1 0.23; llama 7B Q4_0 | tg128 | 16.20 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14347628",
   "tg_tokens": 128,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-09",
   "flags": [],
   "hardware": "Radeon 660M",
   "id": "lc-disc-10879-c14347628-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06425",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 147.9,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 660M; llama 7B Q4_0 | pp512 | 147.90 \u00b1 0.07; llama 7B Q4_0 | tg128 | 16.11 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14347628",
   "tg_tokens": 128,
   "tps": 16.11,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-09",
   "flags": [],
   "hardware": "Radeon RX 6900 XT",
   "id": "lc-disc-10879-c14351000-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b06428",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1901.2,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6900 XT; llama 7B Q4_0 | pp512 | 1901.20 \u00b1 36.70; llama 7B Q4_0 | tg128 | 108.00 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14351000",
   "tg_tokens": 128,
   "tps": 108.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-09",
   "flags": [],
   "hardware": "Radeon RX 6900 XT",
   "id": "lc-disc-10879-c14351000-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06428",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1761.93,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6900 XT; llama 7B Q4_0 | pp512 | 1761.93 \u00b1 4.75; llama 7B Q4_0 | tg128 | 106.15 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14351000",
   "tg_tokens": 128,
   "tps": 106.15,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-10",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-10879-c14366456-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b06127",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 869.4,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; llama 7B Q4_0 | pp512 | 869.40 \u00b1 4.06; llama 7B Q4_0 | tg128 | 52.01 \u00b1 0.32",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14366456",
   "tg_tokens": 128,
   "tps": 52.01,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-10",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-10879-c14366456-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06127",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 951.25,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; llama 7B Q4_0 | pp512 | 951.25 \u00b1 3.96; llama 7B Q4_0 | tg128 | 51.71 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14366456",
   "tg_tokens": 128,
   "tps": 51.71,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-11",
   "flags": [],
   "hardware": "Radeon RX 7900 GRE",
   "id": "lc-disc-10879-c14377891-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b06447",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1420.2,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 GRE; llama 7B Q4_0 | pp512 | 1420.20 \u00b1 8.85; llama 7B Q4_0 | tg128 | 111.32 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14377891",
   "tg_tokens": 128,
   "tps": 111.32,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-11",
   "flags": [],
   "hardware": "Radeon RX 7900 GRE",
   "id": "lc-disc-10879-c14377891-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06447",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1465.45,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 GRE; llama 7B Q4_0 | pp512 | 1465.45 \u00b1 13.53; llama 7B Q4_0 | tg128 | 101.73 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14377891",
   "tg_tokens": 128,
   "tps": 101.73,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-21",
   "flags": [],
   "hardware": "Arc B570",
   "id": "lc-disc-10879-c14466678-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b06527",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 913.95,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc B570; llama 7B Q4_0 | pp512 | 913.95 \u00b1 0.90; llama 7B Q4_0 | tg128 | 49.64 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14466678",
   "tg_tokens": 128,
   "tps": 49.64,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-21",
   "flags": [],
   "hardware": "Arc B570",
   "id": "lc-disc-10879-c14466678-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06527",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 288.51,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc B570; llama 7B Q4_0 | pp512 | 288.51 \u00b1 0.09; llama 7B Q4_0 | tg128 | 50.49 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14466678",
   "tg_tokens": 128,
   "tps": 50.49,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-21",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-10879-c14467315-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1288.96,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; llama 7B Q4_0 | pp512 | 1288.96 \u00b1 6.49; llama 7B Q4_0 | tg128 | 53.59 \u00b1 0.38",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14467315",
   "tg_tokens": 128,
   "tps": 53.59,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-21",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-10879-c14467315-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1357.07,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; llama 7B Q4_0 | pp512 | 1357.07 \u00b1 10.94; llama 7B Q4_0 | tg128 | 53.00 \u00b1 0.13",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14467315",
   "tg_tokens": 128,
   "tps": 53.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-21",
   "flags": [],
   "hardware": "A770",
   "id": "lc-disc-10879-c14467566-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b06528",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 903.49,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A770; llama 7B Q4_0 | pp512 | 903.49 + 3.16; llama 7B Q4_0 | tg128 | 50.90 + 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14467566",
   "tg_tokens": 128,
   "tps": 50.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-21",
   "flags": [],
   "hardware": "A770",
   "id": "lc-disc-10879-c14467566-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06528",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 363.82,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A770; llama 7B Q4_0 | pp512 | 363.82 + 0.95; llama 7B Q4_0 | tg128 | 48.30 + 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14467566",
   "tg_tokens": 128,
   "tps": 48.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-21",
   "flags": [],
   "hardware": "A770",
   "id": "lc-disc-10879-c14467566-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100; config=b06528",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 986.03,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A770; llama 7B Q4_0 | pp512 | 986.03 + 6.03; llama 7B Q4_0 | tg128 | 52.04 + 0.20",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14467566",
   "tg_tokens": 128,
   "tps": 52.04,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-21",
   "flags": [],
   "hardware": "A770",
   "id": "lc-disc-10879-c14467566-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06528",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 372.22,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A770; llama 7B Q4_0 | pp512 | 372.22 + 0.79; llama 7B Q4_0 | tg128 | 49.66 + 0.17",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14467566",
   "tg_tokens": 128,
   "tps": 49.66,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-21",
   "flags": [
    "contradiction"
   ],
   "hardware": "A770",
   "id": "lc-disc-10879-c14467566-llama7b-Q4_0-5",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100; config=b06528",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1866.14,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A770; llama 7B Q4_0 | pp512 | 1866.14 + 2.85; llama 7B Q4_0 | tg128 | 52.62 + 0.30",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14467566",
   "tg_tokens": 128,
   "tps": 52.62,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-21",
   "flags": [
    "contradiction"
   ],
   "hardware": "A770",
   "id": "lc-disc-10879-c14467566-llama7b-Q4_0-6",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b06528",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 480.14,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A770; llama 7B Q4_0 | pp512 | 480.14 + 4.71; llama 7B Q4_0 | tg128 | 13.23 + 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14467566",
   "tg_tokens": 128,
   "tps": 13.23,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-21",
   "flags": [],
   "hardware": "A770",
   "id": "lc-disc-10879-c14467566-llama7b-Q4_0-7",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100; config=b06528",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2564.25,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A770; llama 7B Q4_0 | pp512 | 2564.25 + 3.71; llama 7B Q4_0 | tg128 | 72.54 + 0.12",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14467566",
   "tg_tokens": 128,
   "tps": 72.54,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-21",
   "flags": [],
   "hardware": "A770",
   "id": "lc-disc-10879-c14467566-llama7b-Q4_0-8",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b06528",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2545.16,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A770; llama 7B Q4_0 | pp512 | 2545.16 + 40.50; llama 7B Q4_0 | tg128 | 71.32 + 0.16",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14467566",
   "tg_tokens": 128,
   "tps": 71.32,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-22",
   "flags": [],
   "hardware": "A770",
   "id": "lc-disc-10879-c14475092-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b06528",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2416.16,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A770; llama 7B Q4_0 | pp512 | 2416.16 \u00b1 3.84; llama 7B Q4_0 | tg128 | 82.92 \u00b1 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14475092",
   "tg_tokens": 128,
   "tps": 82.92,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-22",
   "flags": [],
   "hardware": "A770",
   "id": "lc-disc-10879-c14475092-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06528",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2417.17,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A770; llama 7B Q4_0 | pp512 | 2417.17 \u00b1 2.11; llama 7B Q4_0 | tg128 | 81.60 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14475092",
   "tg_tokens": 128,
   "tps": 81.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-21",
   "flags": [],
   "hardware": "Arc A770",
   "id": "lc-disc-10879-c14467742-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b06529",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 904.11,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A770; llama 7B Q4_0 | pp512 | 904.11 \u00b1 5.74; llama 7B Q4_0 | tg128 | 48.46 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14467742",
   "tg_tokens": 128,
   "tps": 48.46,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-21",
   "flags": [],
   "hardware": "Arc A770",
   "id": "lc-disc-10879-c14467742-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06529",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 295.72,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A770; llama 7B Q4_0 | pp512 | 295.72 \u00b1 0.46; llama 7B Q4_0 | tg128 | 49.34 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14467742",
   "tg_tokens": 128,
   "tps": 49.34,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-21",
   "flags": [],
   "hardware": "Radeon RX 480 G",
   "id": "lc-disc-10879-c14467951-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100; config=b06403",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 248.66,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 480 G; llama 7B Q4_0 | pp512 | 248.66 \u00b1 0.28; llama 7B Q4_0 | tg128 | 34.71 \u00b1 0.14",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14467951",
   "tg_tokens": 128,
   "tps": 34.71,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-21",
   "flags": [],
   "hardware": "Radeon RX 480 G",
   "id": "lc-disc-10879-c14467951-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b06403",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 237.27,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 480 G; llama 7B Q4_0 | pp512 | 237.27 \u00b1 0.45; llama 7B Q4_0 | tg128 | 27.42 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14467951",
   "tg_tokens": 128,
   "tps": 27.42,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-21",
   "flags": [],
   "hardware": "Radeon 890M",
   "id": "lc-disc-10879-c14469067-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100; config=b06527",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 354.86,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 890M; llama 7B Q4_0 | pp512 | 354.86 \u00b1 0.57; llama 7B Q4_0 | tg128 | 19.39 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14469067",
   "tg_tokens": 128,
   "tps": 19.39,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-21",
   "flags": [],
   "hardware": "Radeon 890M",
   "id": "lc-disc-10879-c14469067-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b06527",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 380.03,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 890M; llama 7B Q4_0 | pp512 | 380.03 \u00b1 6.45; llama 7B Q4_0 | tg128 | 19.29 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14469067",
   "tg_tokens": 128,
   "tps": 19.29,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-05",
   "flags": [],
   "hardware": "Radeon 890M",
   "id": "lc-disc-10879-c16010708-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100; config=b08202",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 287.18,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 890M; llama 7B Q4_0 | pp512 | 287.18 \u00b1 1.04; llama 7B Q4_0 | tg128 | 18.76 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-16010708",
   "tg_tokens": 128,
   "tps": 18.76,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-05",
   "flags": [],
   "hardware": "Radeon 890M",
   "id": "lc-disc-10879-c16010708-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b08202",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 314.97,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 890M; llama 7B Q4_0 | pp512 | 314.97 \u00b1 0.70; llama 7B Q4_0 | tg128 | 19.78 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-16010708",
   "tg_tokens": 128,
   "tps": 19.78,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-23",
   "flags": [],
   "hardware": "Radeon RX 5700 XT",
   "id": "lc-disc-10879-c14487258-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b06559",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 492.13,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 5700 XT; llama 7B Q4_0 | pp512 | 492.13 \u00b1 0.35; llama 7B Q4_0 | tg128 | 71.58 \u00b1 0.24",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14487258",
   "tg_tokens": 128,
   "tps": 71.58,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-23",
   "flags": [],
   "hardware": "Radeon RX 5700 XT",
   "id": "lc-disc-10879-c14487258-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06559",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 474.89,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 5700 XT; llama 7B Q4_0 | pp512 | 474.89 \u00b1 0.23; llama 7B Q4_0 | tg128 | 71.66 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14487258",
   "tg_tokens": 128,
   "tps": 71.66,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-27",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-disc-10879-c14527618-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b06559",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1573.21,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7800 XT; llama 7B Q4_0 | pp512 | 1573.21 \u00b1 8.65; llama 7B Q4_0 | tg128 | 120.25 \u00b1 0.16",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14527618",
   "tg_tokens": 128,
   "tps": 120.25,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-27",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-disc-10879-c14527618-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06559",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1638.62,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7800 XT; llama 7B Q4_0 | pp512 | 1638.62 \u00b1 4.86; llama 7B Q4_0 | tg128 | 121.58 \u00b1 0.16",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14527618",
   "tg_tokens": 128,
   "tps": 121.58,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-27",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-disc-10879-c14527618-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100; config=b06559",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2053.11,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7800 XT; llama 7B Q4_0 | pp512 | 2053.11 \u00b1 6.42; llama 7B Q4_0 | tg128 | 116.97 \u00b1 0.34",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14527618",
   "tg_tokens": 128,
   "tps": 116.97,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-27",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-disc-10879-c14527618-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06559",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1967.78,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7800 XT; llama 7B Q4_0 | pp512 | 1967.78 \u00b1 8.20; llama 7B Q4_0 | tg128 | 115.71 \u00b1 0.25",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14527618",
   "tg_tokens": 128,
   "tps": 115.71,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-27",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-disc-10879-c14527618-llama7b-Q4_0-5",
   "model": "llama 7B",
   "notes": "threads=100; config=b06559",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1659.94,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7800 XT; llama 7B Q4_0 | pp512 | 1659.94 \u00b1 13.21; llama 7B Q4_0 | tg128 | 120.12 \u00b1 0.41",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14527618",
   "tg_tokens": 128,
   "tps": 120.12,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-27",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-disc-10879-c14527618-llama7b-Q4_0-6",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06559",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1595.43,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7800 XT; llama 7B Q4_0 | pp512 | 1595.43 \u00b1 14.64; llama 7B Q4_0 | tg128 | 121.40 \u00b1 0.34",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14527618",
   "tg_tokens": 128,
   "tps": 121.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-27",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-disc-10879-c14527618-llama7b-Q4_0-7",
   "model": "llama 7B",
   "notes": "threads=100; config=b06559",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1501.02,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7800 XT; llama 7B Q4_0 | pp512 | 1501.02 \u00b1 20.32; llama 7B Q4_0 | tg128 | 116.61 \u00b1 0.34",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14527618",
   "tg_tokens": 128,
   "tps": 116.61,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-27",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-disc-10879-c14527618-llama7b-Q4_0-8",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06559",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 881.05,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7800 XT; llama 7B Q4_0 | pp512 | 881.05 \u00b1 15.74; llama 7B Q4_0 | tg128 | 115.55 \u00b1 0.32",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14527618",
   "tg_tokens": 128,
   "tps": 115.55,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-27",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-disc-10879-c14527618-llama7b-Q4_0-9",
   "model": "llama 7B",
   "notes": "threads=100; config=b06559",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1002.13,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7800 XT; llama 7B Q4_0 | pp512 | 1002.13 \u00b1 2.27; llama 7B Q4_0 | tg128 | 122.52 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14527618",
   "tg_tokens": 128,
   "tps": 122.52,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-27",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-disc-10879-c14527618-llama7b-Q4_0-10",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06559",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 994.51,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7800 XT; llama 7B Q4_0 | pp512 | 994.51 \u00b1 1.11; llama 7B Q4_0 | tg128 | 123.89 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14527618",
   "tg_tokens": 128,
   "tps": 123.89,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-27",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-disc-10879-c14527618-llama7b-Q4_0-11",
   "model": "llama 7B",
   "notes": "threads=100; config=b06559",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1003.78,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7800 XT; llama 7B Q4_0 | pp512 | 1003.78 \u00b1 1.21; llama 7B Q4_0 | tg128 | 113.29 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14527618",
   "tg_tokens": 128,
   "tps": 113.29,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-27",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-disc-10879-c14527618-llama7b-Q4_0-12",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06559",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 728.81,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7800 XT; llama 7B Q4_0 | pp512 | 728.81 \u00b1 0.35; llama 7B Q4_0 | tg128 | 111.97 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14527618",
   "tg_tokens": 128,
   "tps": 111.97,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-08",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-disc-10879-c14625791-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b06713",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1919.21,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7800 XT; llama 7B Q4_0 | pp512 | 1919.21 \u00b1 12.01; llama 7B Q4_0 | tg128 | 119.89 \u00b1 0.23",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14625791",
   "tg_tokens": 128,
   "tps": 119.89,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-08",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-disc-10879-c14625791-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06713",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1998.36,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7800 XT; llama 7B Q4_0 | pp512 | 1998.36 \u00b1 2.28; llama 7B Q4_0 | tg128 | 121.38 \u00b1 0.26",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14625791",
   "tg_tokens": 128,
   "tps": 121.38,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-23",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-10879-c14491167-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3586.22,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 3586.22 \u00b1 11.07; llama 7B Q4_0 | tg128 | 128.49 \u00b1 0.28",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14491167",
   "tg_tokens": 128,
   "tps": 128.49,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-23",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-10879-c14491167-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=0 ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3633.48,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 3633.48 \u00b1 5.84; llama 7B Q4_0 | tg128 | 127.69 \u00b1 0.18",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14491167",
   "tg_tokens": 128,
   "tps": 127.69,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-23",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-10879-c14491167-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3525.21,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 3525.21 \u00b1 29.21; llama 7B Q4_0 | tg128 | 138.45 \u00b1 0.27",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14491167",
   "tg_tokens": 128,
   "tps": 138.45,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-23",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-10879-c14491167-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3895.62,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 3895.62 \u00b1 3.17; llama 7B Q4_0 | tg128 | 138.84 \u00b1 0.18",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14491167",
   "tg_tokens": 128,
   "tps": 138.84,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-24",
   "flags": [],
   "hardware": "Radeon RX 7900 XT",
   "id": "lc-disc-10879-c14500502-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b06478",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 92.18,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XT (RADV NAVI31) (radv) | uma: 0 | fp16: 1 | bf16: 0 | warp size:; llama 7B Q4_0 | pp512 | 92.18 \u00b1 1.50; llama 7B Q4_0 | tg128 | 71.52 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14500502",
   "tg_tokens": 128,
   "tps": 71.52,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-24",
   "flags": [],
   "hardware": "Radeon RX 7900 XT",
   "id": "lc-disc-10879-c14500502-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06478",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 172.66,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XT (RADV NAVI31) (radv) | uma: 0 | fp16: 1 | bf16: 0 | warp size:; llama 7B Q4_0 | pp512 | 172.66 \u00b1 0.13; llama 7B Q4_0 | tg128 | 77.00 \u00b1 0.17",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14500502",
   "tg_tokens": 128,
   "tps": 77.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-24",
   "flags": [],
   "hardware": "Radeon RX 7900 XT",
   "id": "lc-disc-10879-c14500502-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 main_gpu=2 ngl=100 sm=none; config=b06478",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 722.38,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XT; llama 7B Q4_0 | pp512 | 722.38 \u00b1 2.32; llama 7B Q4_0 | tg128 | 55.41 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14500502",
   "tg_tokens": 128,
   "tps": 55.41,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-24",
   "flags": [],
   "hardware": "Radeon RX 7900 XT",
   "id": "lc-disc-10879-c14500502-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 main_gpu=2 ngl=100 sm=none; config=b06478",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 721.44,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XT; llama 7B Q4_0 | pp512 | 721.44 \u00b1 1.71; llama 7B Q4_0 | tg128 | 60.18 \u00b1 0.10",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14500502",
   "tg_tokens": 128,
   "tps": 60.18,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-24",
   "flags": [],
   "hardware": "RTX 4070 Ti SUPER",
   "id": "lc-disc-10879-c14502081-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b06570",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 6099.18,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "NVIDIA GeForce RTX 4070 Ti SUPER (NVIDIA) | uma: 0 | fp16: 1 | bf16: 1 | warp si; llama 7B Q4_0 | pp512 | 6099.18 \u00b1 154.30; llama 7B Q4_0 | tg128 | 129.45 \u00b1 0.18",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14502081",
   "tg_tokens": 128,
   "tps": 129.45,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-24",
   "flags": [],
   "hardware": "RTX 4070 Ti SUPER",
   "id": "lc-disc-10879-c14502081-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06570",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 6801.18,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "NVIDIA GeForce RTX 4070 Ti SUPER (NVIDIA) | uma: 0 | fp16: 1 | bf16: 1 | warp si; llama 7B Q4_0 | pp512 | 6801.18 \u00b1 40.12; llama 7B Q4_0 | tg128 | 135.81 \u00b1 4.29",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14502081",
   "tg_tokens": 128,
   "tps": 135.81,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-24",
   "flags": [],
   "hardware": "A5000",
   "id": "lc-disc-10879-c14502081-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100; config=b06570",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3641.55,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "NVIDIA RTX A5000 (NVIDIA) | uma: 0 | fp16: 1 | bf16: 1 | warp size: 32 | shared; llama 7B Q4_0 | pp512 | 3641.55 \u00b1 9.05; llama 7B Q4_0 | tg128 | 139.89 \u00b1 0.69",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14502081",
   "tg_tokens": 128,
   "tps": 139.89,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-24",
   "flags": [],
   "hardware": "A5000",
   "id": "lc-disc-10879-c14502081-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06570",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4071.22,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "NVIDIA RTX A5000 (NVIDIA) | uma: 0 | fp16: 1 | bf16: 1 | warp size: 32 | shared; llama 7B Q4_0 | pp512 | 4071.22 \u00b1 13.13; llama 7B Q4_0 | tg128 | 140.43 \u00b1 0.22",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14502081",
   "tg_tokens": 128,
   "tps": 140.43,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-24",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc-disc-10879-c14502143-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b06570",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 9452.03,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "NVIDIA GeForce RTX 4090 (NVIDIA) | uma: 0 | fp16: 1 | bf16: 1 | warp size: 32 |; llama 7B Q4_0 | pp512 | 9452.03 \u00b1 187.70; llama 7B Q4_0 | tg128 | 187.97 \u00b1 0.21",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14502143",
   "tg_tokens": 128,
   "tps": 187.97,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-24",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "lc-disc-10879-c14502143-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06570",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 10830.41,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "NVIDIA GeForce RTX 4090 (NVIDIA) | uma: 0 | fp16: 1 | bf16: 1 | warp size: 32 |; llama 7B Q4_0 | pp512 | 10830.41 \u00b1 36.25; llama 7B Q4_0 | tg128 | 190.10 \u00b1 0.31",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14502143",
   "tg_tokens": 128,
   "tps": 190.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-24",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc-disc-10879-c14502143-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100; config=b06570",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4298.97,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "NVIDIA GeForce RTX 3090 (NVIDIA) | uma: 0 | fp16: 1 | bf16: 1 | warp size: 32 |; llama 7B Q4_0 | pp512 | 4298.97 \u00b1 10.59; llama 7B Q4_0 | tg128 | 160.13 \u00b1 0.25",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14502143",
   "tg_tokens": 128,
   "tps": 160.13,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-24",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "lc-disc-10879-c14502143-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06570",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4732.33,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "NVIDIA GeForce RTX 3090 (NVIDIA) | uma: 0 | fp16: 1 | bf16: 1 | warp size: 32 |; llama 7B Q4_0 | pp512 | 4732.33 \u00b1 4.80; llama 7B Q4_0 | tg128 | 162.28 \u00b1 0.21",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14502143",
   "tg_tokens": 128,
   "tps": 162.28,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-16",
   "flags": [],
   "hardware": "Radeon RX 6700 XT",
   "id": "lc-disc-10879-c14415689-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1051.2,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6700 XT; llama 7B Q4_0 | pp512 | 1051.20 \u00b1 0.98; llama 7B Q4_0 | tg128 | 83.88 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14415689",
   "tg_tokens": 128,
   "tps": 83.88,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-16",
   "flags": [],
   "hardware": "Radeon RX 6700 XT",
   "id": "lc-disc-10879-c14415689-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1010.9,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6700 XT; llama 7B Q4_0 | pp512 | 1010.90 \u00b1 12.89; llama 7B Q4_0 | tg128 | 81.86 \u00b1 0.19",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14415689",
   "tg_tokens": 128,
   "tps": 81.86,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-20",
   "flags": [],
   "hardware": "Arc B580",
   "id": "lc-disc-10879-c14464433-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b06527",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 620.94,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc B580; llama 7B Q4_0 | pp512 | 620.94 \u00b1 15.33; llama 7B Q4_0 | tg128 | 70.14 \u00b1 0.28",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14464433",
   "tg_tokens": 128,
   "tps": 70.14,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-20",
   "flags": [],
   "hardware": "Arc B580",
   "id": "lc-disc-10879-c14464433-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06527",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 419.49,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc B580; llama 7B Q4_0 | pp512 | 419.49 \u00b1 3.37; llama 7B Q4_0 | tg128 | 72.00 \u00b1 0.24",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-14464433",
   "tg_tokens": 128,
   "tps": 72.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-01",
   "flags": [],
   "hardware": "GeForce RTX 3090 Ti",
   "id": "lc-disc-15013-c13963584-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b06060",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 6567.49,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 3090 Ti; llama 7B Q4_0 | pp512 | 6567.49 \u00b1 20.30; llama 7B Q4_0 | tg128 | 171.19 \u00b1 3.98",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-13963584",
   "tg_tokens": 128,
   "tps": 171.19,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-01",
   "flags": [],
   "hardware": "GeForce RTX 3090 Ti",
   "id": "lc-disc-15013-c13963584-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06060",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 6924.01,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 3090 Ti; llama 7B Q4_0 | pp512 | 6924.01 \u00b1 10.76; llama 7B Q4_0 | tg128 | 172.26 \u00b1 1.31",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-13963584",
   "tg_tokens": 128,
   "tps": 172.26,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-01",
   "flags": [],
   "hardware": "GeForce RTX 3080",
   "id": "lc-disc-15013-c13963584-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; config=b06060",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 5013.86,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 3080; llama 7B Q4_0 | pp512 | 5013.86 \u00b1 24.80; llama 7B Q4_0 | tg128 | 139.65 \u00b1 0.99",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-13963584",
   "tg_tokens": 128,
   "tps": 139.65,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-01",
   "flags": [],
   "hardware": "GeForce RTX 3080",
   "id": "lc-disc-15013-c13963584-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06060",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 5569.56,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 3080; llama 7B Q4_0 | pp512 | 5569.56 \u00b1 14.04; llama 7B Q4_0 | tg128 | 139.95 \u00b1 0.95",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-13963584",
   "tg_tokens": 128,
   "tps": 139.95,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-01",
   "flags": [],
   "hardware": "GeForce RTX 4070 Ti Super",
   "id": "lc-disc-15013-c13964341-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b00647",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 6924.53,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 4070 Ti Super; llama 7B Q4_0 | pp512 | 6924.53 \u00b1 13.87; llama 7B Q4_0 | tg128 | 132.26 \u00b1 0.16",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-13964341",
   "tg_tokens": 128,
   "tps": 132.26,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-01",
   "flags": [],
   "hardware": "GeForce RTX 4070 Ti Super",
   "id": "lc-disc-15013-c13964341-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b00647",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 7612.32,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 4070 Ti Super; llama 7B Q4_0 | pp512 | 7612.32 \u00b1 37.35; llama 7B Q4_0 | tg128 | 132.85 \u00b1 0.31",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-13964341",
   "tg_tokens": 128,
   "tps": 132.85,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-07",
   "flags": [],
   "hardware": "GeForce RTX 4080",
   "id": "lc-disc-15013-c14032320-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b00647",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 8031.64,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 4080; llama 7B Q4_0 | pp512 | 8031.64 \u00b1 26.49; llama 7B Q4_0 | tg128 | 142.49 \u00b1 0.16",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14032320",
   "tg_tokens": 128,
   "tps": 142.49,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-07",
   "flags": [],
   "hardware": "GeForce RTX 4080",
   "id": "lc-disc-15013-c14032320-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b00647",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 9205.93,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 4080; llama 7B Q4_0 | pp512 | 9205.93 \u00b1 22.31; llama 7B Q4_0 | tg128 | 143.47 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14032320",
   "tg_tokens": 128,
   "tps": 143.47,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-02",
   "flags": [],
   "hardware": "GeForce RTX 3090",
   "id": "lc-disc-15013-c13968120-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b06060",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 5405.83,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 3090; llama 7B Q4_0 | pp512 | 5405.83 \u00b1 5.80; llama 7B Q4_0 | tg128 | 151.04 \u00b1 0.24",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-13968120",
   "tg_tokens": 128,
   "tps": 151.04,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-02",
   "flags": [],
   "hardware": "GeForce RTX 3090",
   "id": "lc-disc-15013-c13968120-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b06060",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 5932.44,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 3090; llama 7B Q4_0 | pp512 | 5932.44 \u00b1 10.87; llama 7B Q4_0 | tg128 | 155.36 \u00b1 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-13968120",
   "tg_tokens": 128,
   "tps": 155.36,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-02",
   "flags": [],
   "hardware": "GeForce RTX 5090",
   "id": "lc-disc-15013-c13968120-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b06060",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 14751.98,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5090; llama 7B Q4_0 | pp512 | 14751.98 \u00b1 136.24; llama 7B Q4_0 | tg128 | 239.62 \u00b1 0.37",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-13968120",
   "tg_tokens": 128,
   "tps": 239.62,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-02",
   "flags": [],
   "hardware": "GeForce RTX 5090",
   "id": "lc-disc-15013-c13968120-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b06060",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 16041.54,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5090; llama 7B Q4_0 | pp512 | 16041.54 \u00b1 85.27; llama 7B Q4_0 | tg128 | 248.57 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-13968120",
   "tg_tokens": 128,
   "tps": 248.57,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-02",
   "flags": [],
   "hardware": "GeForce RTX 5060 Ti",
   "id": "lc-disc-15013-c13972383-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b06002",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3737.25,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5060 Ti; llama 7B Q4_0 | pp512 | 3737.25 \u00b1 6.79; llama 7B Q4_0 | tg128 | 90.94 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-13972383",
   "tg_tokens": 128,
   "tps": 90.94,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-02",
   "flags": [],
   "hardware": "GeForce RTX 5060 Ti",
   "id": "lc-disc-15013-c13972383-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06002",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4195.53,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5060 Ti; llama 7B Q4_0 | pp512 | 4195.53 \u00b1 1.98; llama 7B Q4_0 | tg128 | 93.46 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-13972383",
   "tg_tokens": 128,
   "tps": 93.46,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-02",
   "flags": [],
   "hardware": "GeForce RTX 4060 Ti",
   "id": "lc-disc-15013-c13972383-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; config=b06002",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3394.63,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 4060 Ti; llama 7B Q4_0 | pp512 | 3394.63 \u00b1 7.44; llama 7B Q4_0 | tg128 | 63.86 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-13972383",
   "tg_tokens": 128,
   "tps": 63.86,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-02",
   "flags": [],
   "hardware": "GeForce RTX 4060 Ti",
   "id": "lc-disc-15013-c13972383-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06002",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3803.45,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 4060 Ti; llama 7B Q4_0 | pp512 | 3803.45 \u00b1 70.80; llama 7B Q4_0 | tg128 | 64.03 \u00b1 0.53",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-13972383",
   "tg_tokens": 128,
   "tps": 64.03,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-06",
   "flags": [],
   "hardware": "GeForce RTX 4090",
   "id": "lc-disc-15013-c14016393-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b06098",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 14770.63,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 4090; llama 7B Q4_0 | pp512 | 14770.63 \u00b1 102.93; llama 7B Q4_0 | tg128 | 188.96 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14016393",
   "tg_tokens": 128,
   "tps": 188.96,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-06",
   "flags": [],
   "hardware": "GeForce RTX 4090",
   "id": "lc-disc-15013-c14016393-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; config=b06098",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 11992.7,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 4090; llama 7B Q4_0 | pp512 | 11992.70 \u00b1 107.99; llama 7B Q4_0 | tg128 | 186.21 \u00b1 0.13",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14016393",
   "tg_tokens": 128,
   "tps": 186.21,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-11",
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "lc-disc-15013-c17259436-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 12742.48,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX PRO 6000 Blackwell; llama 7B Q4_0 | pp512 | 12742.48 \u00b1 285.22; llama 7B Q4_0 | tg128 | 260.42 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-17259436",
   "tg_tokens": 128,
   "tps": 260.42,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-11",
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "lc-disc-15013-c17259436-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 14463.16,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX PRO 6000 Blackwell; llama 7B Q4_0 | pp512 | 14463.16 \u00b1 488.65; llama 7B Q4_0 | tg128 | 289.30 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-17259436",
   "tg_tokens": 128,
   "tps": 289.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-19",
   "flags": [],
   "hardware": "GeForce RTX 5060",
   "id": "lc-disc-15013-c17364014-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b09715",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3269.08,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5060; llama 7B Q4_0 | pp512 | 3269.08 \u00b1 45.26; llama 7B Q4_0 | tg128 | 96.70 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-17364014",
   "tg_tokens": 128,
   "tps": 96.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-19",
   "flags": [],
   "hardware": "GeForce RTX 5060",
   "id": "lc-disc-15013-c17364014-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b09715",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3799.47,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5060; llama 7B Q4_0 | pp512 | 3799.47 \u00b1 30.82; llama 7B Q4_0 | tg128 | 100.61 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-17364014",
   "tg_tokens": 128,
   "tps": 100.61,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-07-15",
   "flags": [],
   "hardware": "GeForce RTX 3050",
   "id": "lc-disc-15013-c17647599-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b10015",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1147.01,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 3050; llama 7B Q4_0 | pp512 | 1147.01 \u00b1 4.37; llama 7B Q4_0 | pp1024 | 1053.73 \u00b1 0.96; llama 7B Q4_0 | pp2048 | 931.68 \u00b1 1.77; llama 7B Q4_0 | tg128 | 37.39 \u00b1 0.00; llama 7B Q4_0 | tg256 | 37.33 \u00b1 0.02; llama 7B Q4_0 | tg512 | 35.96 \u00b1 0.07; llama 7B Q4_0 | tg1024 | 35.33 \u00b1 0.03; llama 7B Q4_0 | tg2048 | 33.52 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-17647599",
   "tg_tokens": 128,
   "tps": 37.39,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-07-15",
   "flags": [],
   "hardware": "GeForce RTX 3050",
   "id": "lc-disc-15013-c17647599-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10015",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1258.25,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 3050; llama 7B Q4_0 | pp512 | 1258.25 \u00b1 5.38; llama 7B Q4_0 | pp1024 | 1231.52 \u00b1 1.33; llama 7B Q4_0 | pp2048 | 1179.21 \u00b1 2.37; llama 7B Q4_0 | tg128 | 38.51 \u00b1 0.03; llama 7B Q4_0 | tg256 | 38.51 \u00b1 0.02; llama 7B Q4_0 | tg512 | 37.76 \u00b1 0.04; llama 7B Q4_0 | tg1024 | 36.57 \u00b1 0.02; llama 7B Q4_0 | tg2048 | 34.44 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-17647599",
   "tg_tokens": 128,
   "tps": 38.51,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-07-30",
   "flags": [],
   "hardware": "A100",
   "id": "lc-disc-15013-c17838548-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b10149",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2698.95,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A100; llama 7B Q4_0 | pp512 | 2698.95 \u00b1 33.63; llama 7B Q4_0 | pp1024 | 2614.09 \u00b1 0.51; llama 7B Q4_0 | pp2048 | 2501.37 \u00b1 0.22; llama 7B Q4_0 | pp4096 | 2312.87 \u00b1 0.16; llama 7B Q4_0 | pp8192 | 1997.14 \u00b1 0.17; llama 7B Q4_0 | pp16384 | 1568.06 \u00b1 0.61; llama 7B Q4_0 | pp32768 | 1014.71 \u00b1 18.82; llama 7B Q4_0 | tg128 | 141.32 \u00b1 1.87; llama 7B Q4_0 | tg256 | 140.42 \u00b1 0.62; llama 7B Q4_0 | tg512 | 135.18 \u00b1 0.29; llama 7B Q4_0 | tg1024 | 132.47 \u00b1 0.58; llama 7B Q4_0 | tg2048 | 129.09 \u00b1 0.12",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-17838548",
   "tg_tokens": 128,
   "tps": 141.32,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-07-30",
   "flags": [],
   "hardware": "A100",
   "id": "lc-disc-15013-c17838548-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b10149",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3016.29,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A100; llama 7B Q4_0 | pp512 | 3016.29 \u00b1 30.51; llama 7B Q4_0 | pp1024 | 2966.77 \u00b1 7.45; llama 7B Q4_0 | pp2048 | 2901.18 \u00b1 2.75; llama 7B Q4_0 | pp4096 | 2738.75 \u00b1 4.77; llama 7B Q4_0 | pp8192 | 2457.51 \u00b1 7.97; llama 7B Q4_0 | pp16384 | 2021.53 \u00b1 4.60; llama 7B Q4_0 | pp32768 | 1482.42 \u00b1 0.88; llama 7B Q4_0 | tg128 | 155.72 \u00b1 0.96; llama 7B Q4_0 | tg256 | 152.81 \u00b1 1.51; llama 7B Q4_0 | tg512 | 148.71 \u00b1 1.20; llama 7B Q4_0 | tg1024 | 143.72 \u00b1 0.41; llama 7B Q4_0 | tg2048 | 137.10 \u00b1 1.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-17838548",
   "tg_tokens": 128,
   "tps": 155.72,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-08",
   "flags": [],
   "hardware": "GeForce RTX 3080",
   "id": "lc-disc-15013-c17940386-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b10326",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2875.02,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 3080; llama 7B Q4_0 | pp512 | 2875.02 \u00b1 85.93; llama 7B Q4_0 | tg128 | 90.60 \u00b1 0.20",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-17940386",
   "tg_tokens": 128,
   "tps": 90.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-08",
   "flags": [],
   "hardware": "GeForce RTX 3080",
   "id": "lc-disc-15013-c17940386-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b10326",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3275.48,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 3080; llama 7B Q4_0 | pp512 | 3275.48 \u00b1 85.45; llama 7B Q4_0 | tg128 | 93.42 \u00b1 0.34",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-17940386",
   "tg_tokens": 128,
   "tps": 93.42,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-21",
   "flags": [],
   "hardware": "GeForce RTX 3060",
   "id": "lc-disc-15013-c18108605-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2111.39,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 3060; llama 7B Q4_0 | pp512 | 2111.39 \u00b1 44.37; llama 7B Q4_0 | tg128 | 72.09 \u00b1 0.11",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-18108605",
   "tg_tokens": 128,
   "tps": 72.09,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-21",
   "flags": [],
   "hardware": "GeForce RTX 3060",
   "id": "lc-disc-15013-c18108605-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2407.93,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 3060; llama 7B Q4_0 | pp512 | 2407.93 \u00b1 10.64; llama 7B Q4_0 | tg128 | 75.58 \u00b1 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-18108605",
   "tg_tokens": 128,
   "tps": 75.58,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-07",
   "flags": [],
   "hardware": "GeForce RTX 3090",
   "id": "lc-disc-15013-c18332268-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b10809",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 5153.91,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 3090; dev=CUDA0; llama 7B Q4_0 | pp512 | 5153.91 \u00b1 429.71; llama 7B Q4_0 | tg128 | 168.95 \u00b1 0.66",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-18332268",
   "tg_tokens": 128,
   "tps": 168.95,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-07",
   "flags": [],
   "hardware": "GeForce RTX 3090",
   "id": "lc-disc-15013-c18332268-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10809",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 5973.72,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 3090; dev=CUDA0; llama 7B Q4_0 | pp512 | 5973.72 \u00b1 397.62; llama 7B Q4_0 | tg128 | 177.75 \u00b1 0.51",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-18332268",
   "tg_tokens": 128,
   "tps": 177.75,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-07",
   "flags": [],
   "hardware": "GeForce RTX 3090",
   "id": "lc-disc-15013-c18332268-qwen35moe35ba3b-Q5_K - Medium",
   "model": "qwen35moe 35B.A3B",
   "notes": "threads=99; config=b10809",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4058.5,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "GeForce RTX 3090; qwen35moe 35B.A3B Q5_K - Medium | pp512 | 4058.50 \u00b1 918.92; qwen35moe 35B.A3B Q5_K - Medium | pp1024 | 3150.20 \u00b1 541.54; qwen35moe 35B.A3B Q5_K - Medium | pp2048 | 3564.54 \u00b1 20.37; qwen35moe 35B.A3B Q5_K - Medium | tg512 | 156.16 \u00b1 2.01; qwen35moe 35B.A3B Q5_K - Medium | tg1024 | 157.41 \u00b1 0.90; qwen35moe 35B.A3B Q5_K - Medium | tg2048 | 158.99 \u00b1 0.85",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-18332268",
   "tg_tokens": 512,
   "tps": 156.16,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-07",
   "flags": [],
   "hardware": "GeForce RTX 3090",
   "id": "lc-disc-15013-c18332268-qwen35moe35ba3b-Q5_K - Medium-2",
   "model": "qwen35moe 35B.A3B",
   "notes": "threads=99; run 2 of 6 in this comment (separate llama-bench run; build/flags differ); config=b10809",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4305.41,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "GeForce RTX 3090; qwen35moe 35B.A3B Q5_K - Medium | pp512 | 4305.41 \u00b1 198.46; qwen35moe 35B.A3B Q5_K - Medium | pp1024 | 5180.27 \u00b1 31.87; qwen35moe 35B.A3B Q5_K - Medium | pp2048 | 5066.67 \u00b1 49.86; qwen35moe 35B.A3B Q5_K - Medium | tg512 | 157.32 \u00b1 2.26; qwen35moe 35B.A3B Q5_K - Medium | tg1024 | 161.14 \u00b1 1.01; qwen35moe 35B.A3B Q5_K - Medium | tg2048 | 162.56 \u00b1 0.61",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-18332268",
   "tg_tokens": 512,
   "tps": 157.32,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-07",
   "flags": [],
   "hardware": "GeForce RTX 3090",
   "id": "lc-disc-15013-c18332268-qwen35moe35ba3b-Q5_K - Medium-3",
   "model": "qwen35moe 35B.A3B",
   "notes": "threads=99; run 3 of 6 in this comment (separate llama-bench run; build/flags differ); config=b10809",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4424.02,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "GeForce RTX 3090; qwen35moe 35B.A3B Q5_K - Medium | pp512 | 4424.02 \u00b1 232.64; qwen35moe 35B.A3B Q5_K - Medium | pp1024 | 5192.66 \u00b1 27.38; qwen35moe 35B.A3B Q5_K - Medium | pp2048 | 5269.75 \u00b1 18.77; qwen35moe 35B.A3B Q5_K - Medium | tg512 | 157.14 \u00b1 2.00; qwen35moe 35B.A3B Q5_K - Medium | tg1024 | 160.49 \u00b1 1.17; qwen35moe 35B.A3B Q5_K - Medium | tg2048 | 160.93 \u00b1 0.53",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-18332268",
   "tg_tokens": 512,
   "tps": 157.14,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-07",
   "flags": [],
   "hardware": "GeForce RTX 3090",
   "id": "lc-disc-15013-c18332268-qwen35moe35ba3b-Q5_K - Medium-4",
   "model": "qwen35moe 35B.A3B",
   "notes": "threads=99; run 4 of 6 in this comment (separate llama-bench run; build/flags differ); config=b10809",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4384.76,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "GeForce RTX 3090; qwen35moe 35B.A3B Q5_K - Medium | pp512 | 4384.76 \u00b1 228.01; qwen35moe 35B.A3B Q5_K - Medium | pp1024 | 3256.04 \u00b1 374.82; qwen35moe 35B.A3B Q5_K - Medium | pp2048 | 3610.90 \u00b1 13.66; qwen35moe 35B.A3B Q5_K - Medium | tg512 | 158.25 \u00b1 2.09; qwen35moe 35B.A3B Q5_K - Medium | tg1024 | 157.59 \u00b1 0.92; qwen35moe 35B.A3B Q5_K - Medium | tg2048 | 158.77 \u00b1 0.67",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-18332268",
   "tg_tokens": 512,
   "tps": 158.25,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-07",
   "flags": [],
   "hardware": "GeForce RTX 3090",
   "id": "lc-disc-15013-c18332268-qwen35moe35ba3b-Q5_K - Medium-5",
   "model": "qwen35moe 35B.A3B",
   "notes": "threads=99; run 5 of 6 in this comment (separate llama-bench run; build/flags differ); config=b10809",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4364.06,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "GeForce RTX 3090; qwen35moe 35B.A3B Q5_K - Medium | pp512 | 4364.06 \u00b1 295.32; qwen35moe 35B.A3B Q5_K - Medium | pp1024 | 5212.24 \u00b1 32.57; qwen35moe 35B.A3B Q5_K - Medium | pp2048 | 5452.14 \u00b1 20.84; qwen35moe 35B.A3B Q5_K - Medium | tg512 | 147.76 \u00b1 0.51; qwen35moe 35B.A3B Q5_K - Medium | tg1024 | 158.01 \u00b1 0.91; qwen35moe 35B.A3B Q5_K - Medium | tg2048 | 160.34 \u00b1 0.48",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-18332268",
   "tg_tokens": 512,
   "tps": 147.76,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-07",
   "flags": [],
   "hardware": "GeForce RTX 3090",
   "id": "lc-disc-15013-c18332268-qwen35moe35ba3b-Q5_K - Medium-6",
   "model": "qwen35moe 35B.A3B",
   "notes": "threads=99; run 6 of 6 in this comment (separate llama-bench run; build/flags differ); config=b10809",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4358.85,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "GeForce RTX 3090; qwen35moe 35B.A3B Q5_K - Medium | pp512 | 4358.85 \u00b1 264.67; qwen35moe 35B.A3B Q5_K - Medium | pp1024 | 5190.63 \u00b1 36.27; qwen35moe 35B.A3B Q5_K - Medium | pp2048 | 5483.24 \u00b1 26.08; qwen35moe 35B.A3B Q5_K - Medium | tg512 | 156.64 \u00b1 2.14; qwen35moe 35B.A3B Q5_K - Medium | tg1024 | 158.12 \u00b1 1.28; qwen35moe 35B.A3B Q5_K - Medium | tg2048 | 160.72 \u00b1 0.59",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-18332268",
   "tg_tokens": 512,
   "tps": 156.64,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-07",
   "flags": [],
   "hardware": "GeForce RTX 3090",
   "id": "lc-disc-15013-c18332268-qwen3527b-Q5_K - Medium",
   "model": "qwen35 27B",
   "notes": "threads=99; config=b10809",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1715.55,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "GeForce RTX 3090; qwen35 27B Q5_K - Medium | pp512 | 1715.55 \u00b1 154.81; qwen35 27B Q5_K - Medium | pp1024 | 1753.64 \u00b1 68.05; qwen35 27B Q5_K - Medium | pp2048 | 1778.94 \u00b1 4.82; qwen35 27B Q5_K - Medium | tg512 | 77.24 \u00b1 0.88; qwen35 27B Q5_K - Medium | tg1024 | 76.89 \u00b1 0.40; qwen35 27B Q5_K - Medium | tg2048 | 76.99 \u00b1 0.15",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-18332268",
   "tg_tokens": 512,
   "tps": 77.24,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-07",
   "flags": [],
   "hardware": "GeForce RTX 3090",
   "id": "lc-disc-15013-c18332268-qwen3527b-Q5_K - Medium-2",
   "model": "qwen35 27B",
   "notes": "threads=99; run 2 of 6 in this comment (separate llama-bench run; build/flags differ); config=b10809",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2045.61,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "GeForce RTX 3090; qwen35 27B Q5_K - Medium | pp512 | 2045.61 \u00b1 16.03; qwen35 27B Q5_K - Medium | pp1024 | 2042.48 \u00b1 6.17; qwen35 27B Q5_K - Medium | pp2048 | 1997.27 \u00b1 2.07; qwen35 27B Q5_K - Medium | tg512 | 75.34 \u00b1 0.73; qwen35 27B Q5_K - Medium | tg1024 | 74.88 \u00b1 0.39; qwen35 27B Q5_K - Medium | tg2048 | 77.01 \u00b1 0.20",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-18332268",
   "tg_tokens": 512,
   "tps": 75.34,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-07",
   "flags": [],
   "hardware": "GeForce RTX 3090",
   "id": "lc-disc-15013-c18332268-qwen3527b-Q5_K - Medium-3",
   "model": "qwen35 27B",
   "notes": "threads=99; run 3 of 6 in this comment (separate llama-bench run; build/flags differ); config=b10809",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2047.57,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "GeForce RTX 3090; qwen35 27B Q5_K - Medium | pp512 | 2047.57 \u00b1 14.78; qwen35 27B Q5_K - Medium | pp1024 | 2001.32 \u00b1 5.14; qwen35 27B Q5_K - Medium | pp2048 | 1988.01 \u00b1 2.44; qwen35 27B Q5_K - Medium | tg512 | 74.75 \u00b1 0.64; qwen35 27B Q5_K - Medium | tg1024 | 75.82 \u00b1 0.31; qwen35 27B Q5_K - Medium | tg2048 | 75.76 \u00b1 0.14",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-18332268",
   "tg_tokens": 512,
   "tps": 74.75,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-07",
   "flags": [],
   "hardware": "GeForce RTX 3090",
   "id": "lc-disc-15013-c18332268-qwen3527b-Q5_K - Medium-4",
   "model": "qwen35 27B",
   "notes": "threads=99; run 4 of 6 in this comment (separate llama-bench run; build/flags differ); config=b10809",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2045.01,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "GeForce RTX 3090; qwen35 27B Q5_K - Medium | pp512 | 2045.01 \u00b1 14.64; qwen35 27B Q5_K - Medium | pp1024 | 1800.12 \u00b1 8.62; qwen35 27B Q5_K - Medium | pp2048 | 1854.42 \u00b1 3.20; qwen35 27B Q5_K - Medium | tg512 | 75.18 \u00b1 0.66; qwen35 27B Q5_K - Medium | tg1024 | 75.20 \u00b1 0.29; qwen35 27B Q5_K - Medium | tg2048 | 75.02 \u00b1 0.14",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-18332268",
   "tg_tokens": 512,
   "tps": 75.18,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-07",
   "flags": [],
   "hardware": "GeForce RTX 3090",
   "id": "lc-disc-15013-c18332268-qwen3527b-Q5_K - Medium-5",
   "model": "qwen35 27B",
   "notes": "threads=99; run 5 of 6 in this comment (separate llama-bench run; build/flags differ); config=b10809",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2051.13,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "GeForce RTX 3090; qwen35 27B Q5_K - Medium | pp512 | 2051.13 \u00b1 14.58; qwen35 27B Q5_K - Medium | pp1024 | 1989.40 \u00b1 7.18; qwen35 27B Q5_K - Medium | pp2048 | 1998.75 \u00b1 4.43; qwen35 27B Q5_K - Medium | tg512 | 74.48 \u00b1 0.80; qwen35 27B Q5_K - Medium | tg1024 | 74.76 \u00b1 0.26; qwen35 27B Q5_K - Medium | tg2048 | 75.98 \u00b1 0.17",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-18332268",
   "tg_tokens": 512,
   "tps": 74.48,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-07",
   "flags": [],
   "hardware": "GeForce RTX 3090",
   "id": "lc-disc-15013-c18332268-qwen3527b-Q5_K - Medium-6",
   "model": "qwen35 27B",
   "notes": "threads=99; run 6 of 6 in this comment (separate llama-bench run; build/flags differ); config=b10809",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2050.92,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "GeForce RTX 3090; qwen35 27B Q5_K - Medium | pp512 | 2050.92 \u00b1 14.39; qwen35 27B Q5_K - Medium | pp1024 | 2028.85 \u00b1 10.91; qwen35 27B Q5_K - Medium | pp2048 | 1924.18 \u00b1 2.81; qwen35 27B Q5_K - Medium | tg512 | 74.61 \u00b1 0.67; qwen35 27B Q5_K - Medium | tg1024 | 74.40 \u00b1 0.40; qwen35 27B Q5_K - Medium | tg2048 | 75.12 \u00b1 0.14",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-18332268",
   "tg_tokens": 512,
   "tps": 74.61,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "GeForce RTX 3090",
   "id": "lc-disc-15013-c18332268-qwen3527b-Q5_K - Medium-7",
   "model": "qwen35 27B",
   "notes": "threads=99; config=b10809",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1379.75,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "GeForce RTX 3090; qwen35 27B Q5_K - Medium | pp512 | 1379.75 \u00b1 20.00; qwen35 27B Q5_K - Medium | pp1024 | 2111.77 \u00b1 3.10; qwen35 27B Q5_K - Medium | tg512 | 38.57 \u00b1 0.02; qwen35 27B Q5_K - Medium | tg1024 | 38.53 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-18332268",
   "tg_tokens": 512,
   "tps": 38.57,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-07",
   "flags": [],
   "hardware": "GeForce RTX 3090",
   "id": "lc-disc-15013-c18332268-qwen3527b-Q5_K - Medium-8",
   "model": "qwen35 27B",
   "notes": "threads=99; run 2 of 12 in this comment (separate llama-bench run; build/flags differ); config=b10809",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1401.5,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "GeForce RTX 3090; qwen35 27B Q5_K - Medium | pp512 | 1401.50 \u00b1 14.02; qwen35 27B Q5_K - Medium | pp1024 | 2161.21 \u00b1 4.30; qwen35 27B Q5_K - Medium | tg512 | 38.99 \u00b1 0.02; qwen35 27B Q5_K - Medium | tg1024 | 38.95 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-18332268",
   "tg_tokens": 512,
   "tps": 38.99,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-07",
   "flags": [],
   "hardware": "GeForce RTX 3090",
   "id": "lc-disc-15013-c18332268-qwen3527b-Q5_K - Medium-9",
   "model": "qwen35 27B",
   "notes": "threads=99; run 3 of 12 in this comment (separate llama-bench run; build/flags differ); config=b10809",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1380.63,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "GeForce RTX 3090; qwen35 27B Q5_K - Medium | pp512 | 1380.63 \u00b1 15.36; qwen35 27B Q5_K - Medium | pp1024 | 1367.03 \u00b1 6.14; qwen35 27B Q5_K - Medium | tg512 | 38.54 \u00b1 0.03; qwen35 27B Q5_K - Medium | tg1024 | 38.53 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-18332268",
   "tg_tokens": 512,
   "tps": 38.54,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-07",
   "flags": [],
   "hardware": "GeForce RTX 3090",
   "id": "lc-disc-15013-c18332268-qwen3527b-Q5_K - Medium-10",
   "model": "qwen35 27B",
   "notes": "threads=99; run 4 of 12 in this comment (separate llama-bench run; build/flags differ); config=b10809",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1401.56,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "GeForce RTX 3090; qwen35 27B Q5_K - Medium | pp512 | 1401.56 \u00b1 12.60; qwen35 27B Q5_K - Medium | pp1024 | 1410.39 \u00b1 6.90; qwen35 27B Q5_K - Medium | tg512 | 38.99 \u00b1 0.02; qwen35 27B Q5_K - Medium | tg1024 | 38.95 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-18332268",
   "tg_tokens": 512,
   "tps": 38.99,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-07",
   "flags": [],
   "hardware": "GeForce RTX 3090",
   "id": "lc-disc-15013-c18332268-qwen3527b-Q5_K - Medium-11",
   "model": "qwen35 27B",
   "notes": "threads=99; run 5 of 12 in this comment (separate llama-bench run; build/flags differ); config=b10809",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1377.26,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "GeForce RTX 3090; qwen35 27B Q5_K - Medium | pp512 | 1377.26 \u00b1 15.14; qwen35 27B Q5_K - Medium | pp1024 | 1365.77 \u00b1 6.45; qwen35 27B Q5_K - Medium | tg512 | 38.57 \u00b1 0.02; qwen35 27B Q5_K - Medium | tg1024 | 38.53 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-18332268",
   "tg_tokens": 512,
   "tps": 38.57,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-07",
   "flags": [],
   "hardware": "GeForce RTX 3090",
   "id": "lc-disc-15013-c18332268-qwen3527b-Q5_K - Medium-12",
   "model": "qwen35 27B",
   "notes": "threads=99; run 6 of 12 in this comment (separate llama-bench run; build/flags differ); config=b10809",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1397.03,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "GeForce RTX 3090; qwen35 27B Q5_K - Medium | pp512 | 1397.03 \u00b1 15.84; qwen35 27B Q5_K - Medium | pp1024 | 1409.52 \u00b1 7.15; qwen35 27B Q5_K - Medium | tg512 | 38.98 \u00b1 0.02; qwen35 27B Q5_K - Medium | tg1024 | 38.95 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-18332268",
   "tg_tokens": 512,
   "tps": 38.98,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-07",
   "flags": [],
   "hardware": "GeForce RTX 3090",
   "id": "lc-disc-15013-c18332268-qwen3527b-Q5_K - Medium-13",
   "model": "qwen35 27B",
   "notes": "threads=99; run 7 of 12 in this comment (separate llama-bench run; build/flags differ); config=b10809",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1379.84,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "GeForce RTX 3090; qwen35 27B Q5_K - Medium | pp512 | 1379.84 \u00b1 15.01; qwen35 27B Q5_K - Medium | pp1024 | 2113.72 \u00b1 3.67; qwen35 27B Q5_K - Medium | tg512 | 38.54 \u00b1 0.03; qwen35 27B Q5_K - Medium | tg1024 | 38.51 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-18332268",
   "tg_tokens": 512,
   "tps": 38.54,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-07",
   "flags": [],
   "hardware": "GeForce RTX 3090",
   "id": "lc-disc-15013-c18332268-qwen3527b-Q5_K - Medium-14",
   "model": "qwen35 27B",
   "notes": "threads=99; run 8 of 12 in this comment (separate llama-bench run; build/flags differ); config=b10809",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1404.01,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "GeForce RTX 3090; qwen35 27B Q5_K - Medium | pp512 | 1404.01 \u00b1 11.14; qwen35 27B Q5_K - Medium | pp1024 | 2161.43 \u00b1 7.76; qwen35 27B Q5_K - Medium | tg512 | 38.98 \u00b1 0.02; qwen35 27B Q5_K - Medium | tg1024 | 38.95 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-18332268",
   "tg_tokens": 512,
   "tps": 38.98,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-07",
   "flags": [],
   "hardware": "GeForce RTX 3090",
   "id": "lc-disc-15013-c18332268-qwen3527b-Q5_K - Medium-15",
   "model": "qwen35 27B",
   "notes": "threads=99; run 9 of 12 in this comment (separate llama-bench run; build/flags differ); config=b10809",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1382.0,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "GeForce RTX 3090; qwen35 27B Q5_K - Medium | pp512 | 1382.00 \u00b1 14.09; qwen35 27B Q5_K - Medium | pp1024 | 1370.45 \u00b1 6.92; qwen35 27B Q5_K - Medium | tg512 | 38.56 \u00b1 0.03; qwen35 27B Q5_K - Medium | tg1024 | 38.53 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-18332268",
   "tg_tokens": 512,
   "tps": 38.56,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-07",
   "flags": [],
   "hardware": "GeForce RTX 3090",
   "id": "lc-disc-15013-c18332268-qwen3527b-Q5_K - Medium-16",
   "model": "qwen35 27B",
   "notes": "threads=99; run 10 of 12 in this comment (separate llama-bench run; build/flags differ); config=b10809",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1403.46,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "GeForce RTX 3090; qwen35 27B Q5_K - Medium | pp512 | 1403.46 \u00b1 18.44; qwen35 27B Q5_K - Medium | pp1024 | 1415.92 \u00b1 7.07; qwen35 27B Q5_K - Medium | tg512 | 39.00 \u00b1 0.02; qwen35 27B Q5_K - Medium | tg1024 | 38.94 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-18332268",
   "tg_tokens": 512,
   "tps": 39.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-07",
   "flags": [],
   "hardware": "GeForce RTX 3090",
   "id": "lc-disc-15013-c18332268-qwen3527b-Q5_K - Medium-17",
   "model": "qwen35 27B",
   "notes": "threads=99; run 11 of 12 in this comment (separate llama-bench run; build/flags differ); config=b10809",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1382.78,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "GeForce RTX 3090; qwen35 27B Q5_K - Medium | pp512 | 1382.78 \u00b1 15.48; qwen35 27B Q5_K - Medium | pp1024 | 1370.52 \u00b1 7.41; qwen35 27B Q5_K - Medium | tg512 | 38.54 \u00b1 0.02; qwen35 27B Q5_K - Medium | tg1024 | 38.53 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-18332268",
   "tg_tokens": 512,
   "tps": 38.54,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-07",
   "flags": [],
   "hardware": "GeForce RTX 3090",
   "id": "lc-disc-15013-c18332268-qwen3527b-Q5_K - Medium-18",
   "model": "qwen35 27B",
   "notes": "threads=99; run 12 of 12 in this comment (separate llama-bench run; build/flags differ); config=b10809",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1403.68,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "GeForce RTX 3090; qwen35 27B Q5_K - Medium | pp512 | 1403.68 \u00b1 16.50; qwen35 27B Q5_K - Medium | pp1024 | 1415.26 \u00b1 8.18; qwen35 27B Q5_K - Medium | tg512 | 38.98 \u00b1 0.02; qwen35 27B Q5_K - Medium | tg1024 | 38.95 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-18332268",
   "tg_tokens": 512,
   "tps": 38.98,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-07",
   "flags": [],
   "hardware": "GeForce RTX 3090",
   "id": "lc-disc-15013-c18332268-qwen35moe35ba3b-Q5_K - Medium-7",
   "model": "qwen35moe 35B.A3B",
   "notes": "threads=99; config=b10809",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3586.48,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "GeForce RTX 3090; qwen35moe 35B.A3B Q5_K - Medium | pp512 | 3586.48 \u00b1 135.65; qwen35moe 35B.A3B Q5_K - Medium | pp1024 | 5282.06 \u00b1 25.54; qwen35moe 35B.A3B Q5_K - Medium | pp2048 | 6333.17 \u00b1 18.76; qwen35moe 35B.A3B Q5_K - Medium | tg512 | 148.41 \u00b1 0.70; qwen35moe 35B.A3B Q5_K - Medium | tg1024 | 148.13 \u00b1 0.15; qwen35moe 35B.A3B Q5_K - Medium | tg2048 | 147.42 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-18332268",
   "tg_tokens": 512,
   "tps": 148.41,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-07",
   "flags": [],
   "hardware": "GeForce RTX 3090",
   "id": "lc-disc-15013-c18332268-qwen35moe35ba3b-Q5_K - Medium-8",
   "model": "qwen35moe 35B.A3B",
   "notes": "threads=99; run 2 of 12 in this comment (separate llama-bench run; build/flags differ); config=b10809",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3583.97,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "GeForce RTX 3090; qwen35moe 35B.A3B Q5_K - Medium | pp512 | 3583.97 \u00b1 80.24; qwen35moe 35B.A3B Q5_K - Medium | pp1024 | 5417.14 \u00b1 28.36; qwen35moe 35B.A3B Q5_K - Medium | pp2048 | 6618.72 \u00b1 33.39; qwen35moe 35B.A3B Q5_K - Medium | tg512 | 150.93 \u00b1 0.72; qwen35moe 35B.A3B Q5_K - Medium | tg1024 | 150.89 \u00b1 0.10; qwen35moe 35B.A3B Q5_K - Medium | tg2048 | 150.62 \u00b1 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-18332268",
   "tg_tokens": 512,
   "tps": 150.93,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-07",
   "flags": [],
   "hardware": "GeForce RTX 3090",
   "id": "lc-disc-15013-c18332268-qwen35moe35ba3b-Q5_K - Medium-9",
   "model": "qwen35moe 35B.A3B",
   "notes": "threads=99; run 3 of 12 in this comment (separate llama-bench run; build/flags differ); config=b10809",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3549.89,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "GeForce RTX 3090; qwen35moe 35B.A3B Q5_K - Medium | pp512 | 3549.89 \u00b1 94.04; qwen35moe 35B.A3B Q5_K - Medium | pp1024 | 4272.02 \u00b1 54.79; qwen35moe 35B.A3B Q5_K - Medium | pp2048 | 4707.78 \u00b1 40.58; qwen35moe 35B.A3B Q5_K - Medium | tg512 | 148.69 \u00b1 0.88; qwen35moe 35B.A3B Q5_K - Medium | tg1024 | 147.93 \u00b1 0.09; qwen35moe 35B.A3B Q5_K - Medium | tg2048 | 146.86 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-18332268",
   "tg_tokens": 512,
   "tps": 148.69,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-07",
   "flags": [],
   "hardware": "GeForce RTX 3090",
   "id": "lc-disc-15013-c18332268-qwen35moe35ba3b-Q5_K - Medium-10",
   "model": "qwen35moe 35B.A3B",
   "notes": "threads=99; run 4 of 12 in this comment (separate llama-bench run; build/flags differ); config=b10809",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3580.28,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "GeForce RTX 3090; qwen35moe 35B.A3B Q5_K - Medium | pp512 | 3580.28 \u00b1 64.98; qwen35moe 35B.A3B Q5_K - Medium | pp1024 | 4472.54 \u00b1 54.13; qwen35moe 35B.A3B Q5_K - Medium | pp2048 | 5069.91 \u00b1 39.70; qwen35moe 35B.A3B Q5_K - Medium | tg512 | 150.48 \u00b1 0.75; qwen35moe 35B.A3B Q5_K - Medium | tg1024 | 150.16 \u00b1 0.06; qwen35moe 35B.A3B Q5_K - Medium | tg2048 | 149.88 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-18332268",
   "tg_tokens": 512,
   "tps": 150.48,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-07",
   "flags": [],
   "hardware": "GeForce RTX 3090",
   "id": "lc-disc-15013-c18332268-qwen35moe35ba3b-Q5_K - Medium-11",
   "model": "qwen35moe 35B.A3B",
   "notes": "threads=99; run 5 of 12 in this comment (separate llama-bench run; build/flags differ); config=b10809",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3587.97,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "GeForce RTX 3090; qwen35moe 35B.A3B Q5_K - Medium | pp512 | 3587.97 \u00b1 72.78; qwen35moe 35B.A3B Q5_K - Medium | pp1024 | 4314.58 \u00b1 51.78; qwen35moe 35B.A3B Q5_K - Medium | pp2048 | 4720.32 \u00b1 33.94; qwen35moe 35B.A3B Q5_K - Medium | tg512 | 148.36 \u00b1 0.74; qwen35moe 35B.A3B Q5_K - Medium | tg1024 | 147.62 \u00b1 0.07; qwen35moe 35B.A3B Q5_K - Medium | tg2048 | 146.76 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-18332268",
   "tg_tokens": 512,
   "tps": 148.36,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-07",
   "flags": [],
   "hardware": "GeForce RTX 3090",
   "id": "lc-disc-15013-c18332268-qwen35moe35ba3b-Q5_K - Medium-12",
   "model": "qwen35moe 35B.A3B",
   "notes": "threads=99; run 6 of 12 in this comment (separate llama-bench run; build/flags differ); config=b10809",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3619.07,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "GeForce RTX 3090; qwen35moe 35B.A3B Q5_K - Medium | pp512 | 3619.07 \u00b1 94.51; qwen35moe 35B.A3B Q5_K - Medium | pp1024 | 4464.09 \u00b1 57.35; qwen35moe 35B.A3B Q5_K - Medium | pp2048 | 5072.60 \u00b1 60.84; qwen35moe 35B.A3B Q5_K - Medium | tg512 | 150.31 \u00b1 0.60; qwen35moe 35B.A3B Q5_K - Medium | tg1024 | 150.16 \u00b1 0.08; qwen35moe 35B.A3B Q5_K - Medium | tg2048 | 149.85 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-18332268",
   "tg_tokens": 512,
   "tps": 150.31,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-07",
   "flags": [],
   "hardware": "GeForce RTX 3090",
   "id": "lc-disc-15013-c18332268-qwen35moe35ba3b-Q5_K - Medium-13",
   "model": "qwen35moe 35B.A3B",
   "notes": "threads=99; run 7 of 12 in this comment (separate llama-bench run; build/flags differ); config=b10809",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3531.31,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "GeForce RTX 3090; qwen35moe 35B.A3B Q5_K - Medium | pp512 | 3531.31 \u00b1 85.42; qwen35moe 35B.A3B Q5_K - Medium | pp1024 | 5281.08 \u00b1 57.10; qwen35moe 35B.A3B Q5_K - Medium | pp2048 | 6323.76 \u00b1 16.85; qwen35moe 35B.A3B Q5_K - Medium | tg512 | 148.13 \u00b1 0.87; qwen35moe 35B.A3B Q5_K - Medium | tg1024 | 147.65 \u00b1 0.26; qwen35moe 35B.A3B Q5_K - Medium | tg2048 | 147.20 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-18332268",
   "tg_tokens": 512,
   "tps": 148.13,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-07",
   "flags": [],
   "hardware": "GeForce RTX 3090",
   "id": "lc-disc-15013-c18332268-qwen35moe35ba3b-Q5_K - Medium-14",
   "model": "qwen35moe 35B.A3B",
   "notes": "threads=99; run 8 of 12 in this comment (separate llama-bench run; build/flags differ); config=b10809",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3600.54,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "GeForce RTX 3090; qwen35moe 35B.A3B Q5_K - Medium | pp512 | 3600.54 \u00b1 88.32; qwen35moe 35B.A3B Q5_K - Medium | pp1024 | 5409.63 \u00b1 56.75; qwen35moe 35B.A3B Q5_K - Medium | pp2048 | 6627.92 \u00b1 22.80; qwen35moe 35B.A3B Q5_K - Medium | tg512 | 151.13 \u00b1 0.83; qwen35moe 35B.A3B Q5_K - Medium | tg1024 | 150.77 \u00b1 0.12; qwen35moe 35B.A3B Q5_K - Medium | tg2048 | 150.69 \u00b1 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-18332268",
   "tg_tokens": 512,
   "tps": 151.13,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-07",
   "flags": [],
   "hardware": "GeForce RTX 3090",
   "id": "lc-disc-15013-c18332268-qwen35moe35ba3b-Q5_K - Medium-15",
   "model": "qwen35moe 35B.A3B",
   "notes": "threads=99; run 9 of 12 in this comment (separate llama-bench run; build/flags differ); config=b10809",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3530.83,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "GeForce RTX 3090; qwen35moe 35B.A3B Q5_K - Medium | pp512 | 3530.83 \u00b1 87.59; qwen35moe 35B.A3B Q5_K - Medium | pp1024 | 4294.76 \u00b1 73.61; qwen35moe 35B.A3B Q5_K - Medium | pp2048 | 4727.64 \u00b1 34.42; qwen35moe 35B.A3B Q5_K - Medium | tg512 | 148.79 \u00b1 0.85; qwen35moe 35B.A3B Q5_K - Medium | tg1024 | 147.84 \u00b1 0.12; qwen35moe 35B.A3B Q5_K - Medium | tg2048 | 147.05 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-18332268",
   "tg_tokens": 512,
   "tps": 148.79,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-07",
   "flags": [],
   "hardware": "GeForce RTX 3090",
   "id": "lc-disc-15013-c18332268-qwen35moe35ba3b-Q5_K - Medium-16",
   "model": "qwen35moe 35B.A3B",
   "notes": "threads=99; run 10 of 12 in this comment (separate llama-bench run; build/flags differ); config=b10809",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3601.63,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "GeForce RTX 3090; qwen35moe 35B.A3B Q5_K - Medium | pp512 | 3601.63 \u00b1 86.00; qwen35moe 35B.A3B Q5_K - Medium | pp1024 | 4458.63 \u00b1 77.60; qwen35moe 35B.A3B Q5_K - Medium | pp2048 | 5079.90 \u00b1 43.59; qwen35moe 35B.A3B Q5_K - Medium | tg512 | 151.01 \u00b1 0.81; qwen35moe 35B.A3B Q5_K - Medium | tg1024 | 150.56 \u00b1 0.07; qwen35moe 35B.A3B Q5_K - Medium | tg2048 | 150.28 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-18332268",
   "tg_tokens": 512,
   "tps": 151.01,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-07",
   "flags": [],
   "hardware": "GeForce RTX 3090",
   "id": "lc-disc-15013-c18332268-qwen35moe35ba3b-Q5_K - Medium-17",
   "model": "qwen35moe 35B.A3B",
   "notes": "threads=99; run 11 of 12 in this comment (separate llama-bench run; build/flags differ); config=b10809",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3560.95,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "GeForce RTX 3090; qwen35moe 35B.A3B Q5_K - Medium | pp512 | 3560.95 \u00b1 101.47; qwen35moe 35B.A3B Q5_K - Medium | pp1024 | 4278.19 \u00b1 61.81; qwen35moe 35B.A3B Q5_K - Medium | pp2048 | 4716.25 \u00b1 41.81; qwen35moe 35B.A3B Q5_K - Medium | tg512 | 148.50 \u00b1 0.82; qwen35moe 35B.A3B Q5_K - Medium | tg1024 | 147.88 \u00b1 0.11; qwen35moe 35B.A3B Q5_K - Medium | tg2048 | 146.73 \u00b1 0.10",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-18332268",
   "tg_tokens": 512,
   "tps": 148.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-07",
   "flags": [],
   "hardware": "GeForce RTX 3090",
   "id": "lc-disc-15013-c18332268-qwen35moe35ba3b-Q5_K - Medium-18",
   "model": "qwen35moe 35B.A3B",
   "notes": "threads=99; run 12 of 12 in this comment (separate llama-bench run; build/flags differ); config=b10809",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3579.26,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "GeForce RTX 3090; qwen35moe 35B.A3B Q5_K - Medium | pp512 | 3579.26 \u00b1 89.88; qwen35moe 35B.A3B Q5_K - Medium | pp1024 | 4495.46 \u00b1 77.53; qwen35moe 35B.A3B Q5_K - Medium | pp2048 | 5054.23 \u00b1 33.29; qwen35moe 35B.A3B Q5_K - Medium | tg512 | 150.40 \u00b1 0.58; qwen35moe 35B.A3B Q5_K - Medium | tg1024 | 150.07 \u00b1 0.04; qwen35moe 35B.A3B Q5_K - Medium | tg2048 | 149.76 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-18332268",
   "tg_tokens": 512,
   "tps": 150.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-16",
   "flags": [],
   "hardware": "GeForce RTX 5090",
   "id": "lc-disc-15013-c18458510-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; 150.0 W measured GPU power 140-160 W during the run (midpoint 150.0 W, approximate); fa=0 ngl=99",
   "params": null,
   "power_w": 150.0,
   "pp_tokens": 512,
   "pp_tps": 6663.79,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5090; llama 7B Q4_0 | pp512 | 6663.79 \u00b1 211.62; llama 7B Q4_0 | tg128 | 157.84 \u00b1 1.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-18458510",
   "tg_tokens": 128,
   "tps": 157.84,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-16",
   "flags": [],
   "hardware": "GeForce RTX 5090",
   "id": "lc-disc-15013-c18458510-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 7603.2,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5090; llama 7B Q4_0 | pp512 | 7603.20 \u00b1 230.26; llama 7B Q4_0 | tg128 | 164.31 \u00b1 0.71",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-18458510",
   "tg_tokens": 128,
   "tps": 164.31,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-19",
   "flags": [],
   "hardware": "GeForce RTX 3060 Ti",
   "id": "lc-disc-15013-c18520153-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2675.22,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 3060 Ti; llama 7B Q4_0 | pp512 | 2675.22 \u00b1 45.49; llama 7B Q4_0 | pp1024 | 2463.94 \u00b1 34.48; llama 7B Q4_0 | pp2048 | 2231.53 \u00b1 5.82; llama 7B Q4_0 | pp4096 | 1896.88 \u00b1 1.14; llama 7B Q4_0 | tg128 | 92.23 \u00b1 0.08; llama 7B Q4_0 | tg256 | 92.26 \u00b1 0.02; llama 7B Q4_0 | tg512 | 88.73 \u00b1 0.10",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-18520153",
   "tg_tokens": 128,
   "tps": 92.23,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-19",
   "flags": [],
   "hardware": "GeForce RTX 3060 Ti",
   "id": "lc-disc-15013-c18520153-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2990.99,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 3060 Ti; llama 7B Q4_0 | pp512 | 2990.99 \u00b1 49.56; llama 7B Q4_0 | pp1024 | 2931.36 \u00b1 36.17; llama 7B Q4_0 | pp2048 | 2814.70 \u00b1 12.93; llama 7B Q4_0 | pp4096 | 2598.43 \u00b1 1.66; llama 7B Q4_0 | tg128 | 96.50 \u00b1 0.09; llama 7B Q4_0 | tg256 | 96.48 \u00b1 0.01; llama 7B Q4_0 | tg512 | 94.76 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-18520153",
   "tg_tokens": 128,
   "tps": 96.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-11",
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "lc-disc-15013-c14075193-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 14401.14,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX PRO 6000 Blackwell; llama 7B Q4_0 | pp512 | 14401.14 \u00b1 17.25; llama 7B Q4_0 | tg128 | 268.52 \u00b1 0.53",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14075193",
   "tg_tokens": 128,
   "tps": 268.52,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-11",
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "lc-disc-15013-c14075193-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 16652.18,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX PRO 6000 Blackwell; llama 7B Q4_0 | pp512 | 16652.18 \u00b1 13.96; llama 7B Q4_0 | tg128 | 275.76 \u00b1 0.21",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14075193",
   "tg_tokens": 128,
   "tps": 275.76,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-14",
   "flags": [],
   "hardware": "A4500",
   "id": "lc-disc-15013-c14103173-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2827.2,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A4500; llama 7B Q4_0 | pp512 | 2827.20 \u00b1 66.43; llama 7B Q4_0 | tg128 | 97.32 \u00b1 2.80",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14103173",
   "tg_tokens": 128,
   "tps": 97.32,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-14",
   "flags": [],
   "hardware": "A4500",
   "id": "lc-disc-15013-c14103173-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3453.1,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A4500; llama 7B Q4_0 | pp512 | 3453.10 \u00b1 49.19; llama 7B Q4_0 | tg128 | 103.00 \u00b1 0.25",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14103173",
   "tg_tokens": 128,
   "tps": 103.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-14",
   "flags": [],
   "hardware": "A4500",
   "id": "lc-disc-15013-c14103173-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2960.88,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A4500; llama 7B Q4_0 | pp512 | 2960.88 \u00b1 47.70; llama 7B Q4_0 | pp1024 | 2796.79 \u00b1 10.68; llama 7B Q4_0 | pp2048 | 2615.76 \u00b1 11.62; llama 7B Q4_0 | pp4096 | 2288.56 \u00b1 45.39; llama 7B Q4_0 | pp8192 | 1855.83 \u00b1 8.06; llama 7B Q4_0 | pp16384 | 1304.95 \u00b1 6.67; llama 7B Q4_0 | tg128 | 98.45 \u00b1 0.85; llama 7B Q4_0 | tg256 | 96.63 \u00b1 2.34; llama 7B Q4_0 | tg512 | 92.77 \u00b1 0.75; llama 7B Q4_0 | tg1024 | 91.48 \u00b1 1.19; llama 7B Q4_0 | tg2048 | 87.42 \u00b1 0.67",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14103173",
   "tg_tokens": 128,
   "tps": 98.45,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-14",
   "flags": [],
   "hardware": "A4500",
   "id": "lc-disc-15013-c14103173-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3382.11,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A4500; llama 7B Q4_0 | pp512 | 3382.11 \u00b1 35.06; llama 7B Q4_0 | pp1024 | 3414.61 \u00b1 36.23; llama 7B Q4_0 | pp2048 | 3334.50 \u00b1 10.57; llama 7B Q4_0 | pp4096 | 3170.51 \u00b1 15.71; llama 7B Q4_0 | pp8192 | 2898.26 \u00b1 9.63; llama 7B Q4_0 | pp16384 | 2440.37 \u00b1 5.62; llama 7B Q4_0 | tg128 | 95.87 \u00b1 3.39; llama 7B Q4_0 | tg256 | 96.22 \u00b1 1.43; llama 7B Q4_0 | tg512 | 96.10 \u00b1 1.12; llama 7B Q4_0 | tg1024 | 93.69 \u00b1 1.60; llama 7B Q4_0 | tg2048 | 90.40 \u00b1 0.58",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14103173",
   "tg_tokens": 128,
   "tps": 95.87,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-14",
   "flags": [],
   "hardware": "T4",
   "id": "lc-disc-15013-c14104555-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b06160",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1219.06,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "T4; llama 7B Q4_0 | pp512 | 1219.06 \u00b1 4.18; llama 7B Q4_0 | tg128 | 46.38 \u00b1 0.73",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14104555",
   "tg_tokens": 128,
   "tps": 46.38,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-14",
   "flags": [],
   "hardware": "T4",
   "id": "lc-disc-15013-c14104555-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06160",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1309.73,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "T4; llama 7B Q4_0 | pp512 | 1309.73 \u00b1 1.02; llama 7B Q4_0 | tg128 | 44.03 \u00b1 0.57",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14104555",
   "tg_tokens": 128,
   "tps": 44.03,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-17",
   "flags": [],
   "hardware": "A4000",
   "id": "lc-disc-15013-c14130794-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b06183",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2684.06,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A4000; llama 7B Q4_0 | pp512 | 2684.06 \u00b1 15.28; llama 7B Q4_0 | pp1024 | 2511.22 \u00b1 7.71; llama 7B Q4_0 | pp2048 | 2318.32 \u00b1 6.04; llama 7B Q4_0 | pp4096 | 1997.56 \u00b1 5.20; llama 7B Q4_0 | pp8192 | 1528.87 \u00b1 8.78; llama 7B Q4_0 | pp16384 | 1038.57 \u00b1 3.93; llama 7B Q4_0 | tg128 | 83.77 \u00b1 0.37; llama 7B Q4_0 | tg256 | 82.51 \u00b1 0.06; llama 7B Q4_0 | tg512 | 79.50 \u00b1 0.03; llama 7B Q4_0 | tg1024 | 78.09 \u00b1 0.04; llama 7B Q4_0 | tg2048 | 74.20 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14130794",
   "tg_tokens": 128,
   "tps": 83.77,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-17",
   "flags": [],
   "hardware": "A4000",
   "id": "lc-disc-15013-c14130794-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06183",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2807.83,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A4000; llama 7B Q4_0 | pp512 | 2807.83 \u00b1 52.44; llama 7B Q4_0 | pp1024 | 2780.21 \u00b1 25.00; llama 7B Q4_0 | pp2048 | 2729.44 \u00b1 9.02; llama 7B Q4_0 | pp4096 | 2612.68 \u00b1 7.70; llama 7B Q4_0 | pp8192 | 2394.07 \u00b1 1.86; llama 7B Q4_0 | pp16384 | 2024.94 \u00b1 1.74; llama 7B Q4_0 | tg128 | 85.17 \u00b1 0.66; llama 7B Q4_0 | tg256 | 85.52 \u00b1 0.01; llama 7B Q4_0 | tg512 | 83.62 \u00b1 0.08; llama 7B Q4_0 | tg1024 | 81.28 \u00b1 0.04; llama 7B Q4_0 | tg2048 | 76.88 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14130794",
   "tg_tokens": 128,
   "tps": 85.17,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-18",
   "flags": [],
   "hardware": "GeForce RTX 3060",
   "id": "lc-disc-15013-c14145116-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1921.04,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 3060; llama 7B Q4_0 | pp512 | 1921.04 \u00b1 13.93; llama 7B Q4_0 | tg128 | 71.66 \u00b1 0.65",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14145116",
   "tg_tokens": 128,
   "tps": 71.66,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-18",
   "flags": [],
   "hardware": "GeForce RTX 3060",
   "id": "lc-disc-15013-c14145116-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2343.25,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 3060; llama 7B Q4_0 | pp512 | 2343.25 \u00b1 27.63; llama 7B Q4_0 | tg128 | 73.75 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14145116",
   "tg_tokens": 128,
   "tps": 73.75,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-18",
   "flags": [],
   "hardware": "GeForce RTX 3060",
   "id": "lc-disc-15013-c14145116-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2028.81,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 3060; llama 7B Q4_0 | pp512 | 2028.81 \u00b1 6.86; llama 7B Q4_0 | tg128 | 69.01 \u00b1 0.32",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14145116",
   "tg_tokens": 128,
   "tps": 69.01,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-18",
   "flags": [],
   "hardware": "GeForce RTX 3060",
   "id": "lc-disc-15013-c14145116-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2239.81,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 3060; llama 7B Q4_0 | pp512 | 2239.81 \u00b1 4.18; llama 7B Q4_0 | tg128 | 71.17 \u00b1 0.57",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14145116",
   "tg_tokens": 128,
   "tps": 71.17,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-23",
   "flags": [],
   "hardware": "GeForce RTX 3060",
   "id": "lc-disc-15013-c14195721-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b06251",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2020.41,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 3060; llama 7B Q4_0 | pp512 | 2020.41 \u00b1 66.40; llama 7B Q4_0 | tg128 | 71.14 \u00b1 1.99",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14195721",
   "tg_tokens": 128,
   "tps": 71.14,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-23",
   "flags": [],
   "hardware": "GeForce RTX 3060",
   "id": "lc-disc-15013-c14195721-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06251",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2401.11,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 3060; llama 7B Q4_0 | pp512 | 2401.11 \u00b1 20.15; llama 7B Q4_0 | tg128 | 74.98 \u00b1 0.12",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14195721",
   "tg_tokens": 128,
   "tps": 74.98,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-19",
   "flags": [],
   "hardware": "A4000",
   "id": "lc-disc-15013-c14149404-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b06195",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 309.3,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A4000; llama 7B Q4_0 | pp512 | 309.30 \u00b1 0.05; llama 7B Q4_0 | pp1024 | 294.86 \u00b1 0.42; llama 7B Q4_0 | pp2048 | 274.91 \u00b1 0.26; llama 7B Q4_0 | tg128 | 23.63 \u00b1 0.00; llama 7B Q4_0 | tg256 | 22.90 \u00b1 0.02; llama 7B Q4_0 | tg512 | 21.61 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14149404",
   "tg_tokens": 128,
   "tps": 23.63,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-19",
   "flags": [],
   "hardware": "A4000",
   "id": "lc-disc-15013-c14149404-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b06195",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 311.55,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A4000; llama 7B Q4_0 | pp512 | 311.55 \u00b1 0.19; llama 7B Q4_0 | pp1024 | 300.80 \u00b1 0.20; llama 7B Q4_0 | pp2048 | 280.92 \u00b1 0.15; llama 7B Q4_0 | tg128 | 23.76 \u00b1 0.01; llama 7B Q4_0 | tg256 | 23.73 \u00b1 0.01; llama 7B Q4_0 | tg512 | 23.44 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14149404",
   "tg_tokens": 128,
   "tps": 23.76,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-19",
   "flags": [],
   "hardware": "A5000",
   "id": "lc-disc-15013-c14150776-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b06189",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4028.16,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A5000; llama 7B Q4_0 | pp512 | 4028.16 \u00b1 19.14; llama 7B Q4_0 | pp1024 | 3776.79 \u00b1 10.56; llama 7B Q4_0 | pp2048 | 3485.38 \u00b1 4.50; llama 7B Q4_0 | pp4096 | 3059.41 \u00b1 8.20; llama 7B Q4_0 | pp8192 | 2443.12 \u00b1 4.43; llama 7B Q4_0 | pp16384 | 1720.63 \u00b1 1.01; llama 7B Q4_0 | pp32768 | 986.24 \u00b1 0.24; llama 7B Q4_0 | tg128 | 130.07 \u00b1 2.74; llama 7B Q4_0 | tg256 | 129.03 \u00b1 0.16; llama 7B Q4_0 | tg512 | 124.03 \u00b1 0.04; llama 7B Q4_0 | tg1024 | 122.15 \u00b1 0.08; llama 7B Q4_0 | tg2048 | 116.65 \u00b1 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14150776",
   "tg_tokens": 128,
   "tps": 130.07,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-19",
   "flags": [],
   "hardware": "A5000",
   "id": "lc-disc-15013-c14150776-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06189",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4552.15,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A5000; llama 7B Q4_0 | pp512 | 4552.15 \u00b1 9.68; llama 7B Q4_0 | pp1024 | 4500.20 \u00b1 15.28; llama 7B Q4_0 | pp2048 | 4407.07 \u00b1 2.32; llama 7B Q4_0 | pp4096 | 4217.01 \u00b1 2.15; llama 7B Q4_0 | pp8192 | 3860.93 \u00b1 0.92; llama 7B Q4_0 | pp16384 | 3286.47 \u00b1 0.96; llama 7B Q4_0 | pp32768 | 2521.55 \u00b1 0.25; llama 7B Q4_0 | tg128 | 135.83 \u00b1 0.11; llama 7B Q4_0 | tg256 | 135.35 \u00b1 0.06; llama 7B Q4_0 | tg512 | 132.90 \u00b1 0.21; llama 7B Q4_0 | tg1024 | 129.18 \u00b1 0.08; llama 7B Q4_0 | tg2048 | 122.99 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14150776",
   "tg_tokens": 128,
   "tps": 135.83,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-20",
   "flags": [],
   "hardware": "GeForce RTX 5070",
   "id": "lc-disc-15013-c14161275-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 5184.75,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5070; llama 7B Q4_0 | pp512 | 5184.75 \u00b1 18.70; llama 7B Q4_0 | tg128 | 127.54 \u00b1 0.46",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14161275",
   "tg_tokens": 128,
   "tps": 127.54,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-20",
   "flags": [],
   "hardware": "GeForce RTX 5070",
   "id": "lc-disc-15013-c14161275-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 5783.44,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5070; llama 7B Q4_0 | pp512 | 5783.44 \u00b1 36.95; llama 7B Q4_0 | tg128 | 128.21 \u00b1 2.52",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14161275",
   "tg_tokens": 128,
   "tps": 128.21,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-21",
   "flags": [],
   "hardware": "GeForce RTX 3060",
   "id": "lc-disc-15013-c14182757-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b06195",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2137.5,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 3060; llama 7B Q4_0 | pp512 | 2137.50 \u00b1 10.12; llama 7B Q4_0 | tg128 | 75.57 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14182757",
   "tg_tokens": 128,
   "tps": 75.57,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-21",
   "flags": [],
   "hardware": "GeForce RTX 3060",
   "id": "lc-disc-15013-c14182757-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06195",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2407.67,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 3060; llama 7B Q4_0 | pp512 | 2407.67 \u00b1 3.73; llama 7B Q4_0 | tg128 | 76.92 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14182757",
   "tg_tokens": 128,
   "tps": 76.92,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-06",
   "flags": [],
   "hardware": "GeForce RTX 5060 Ti",
   "id": "lc-disc-15013-c16020010-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b08214",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3912.3,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5060 Ti; llama 7B Q4_0 | pp512 | 3912.30 \u00b1 28.13; llama 7B Q4_0 | tg128 | 93.90 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-16020010",
   "tg_tokens": 128,
   "tps": 93.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-06",
   "flags": [],
   "hardware": "GeForce RTX 5060 Ti",
   "id": "lc-disc-15013-c16020010-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b08214",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4613.86,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5060 Ti; llama 7B Q4_0 | pp512 | 4613.86 \u00b1 19.07; llama 7B Q4_0 | tg128 | 97.32 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-16020010",
   "tg_tokens": 128,
   "tps": 97.32,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-30",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-disc-15013-c16376369-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b08580",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3293.78,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GB10; llama 7B Q4_0 | pp512 | 3293.78 \u00b1 32.18; llama 7B Q4_0 | tg128 | 54.39 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-16376369",
   "tg_tokens": 128,
   "tps": 54.39,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-30",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-disc-15013-c16376369-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b08580",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3920.63,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GB10; llama 7B Q4_0 | pp512 | 3920.63 \u00b1 60.14; llama 7B Q4_0 | tg128 | 54.91 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-16376369",
   "tg_tokens": 128,
   "tps": 54.91,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-30",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-disc-15013-c16376603-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b08580",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3306.61,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GB10; dev=CUDA0; llama 7B Q4_0 | pp512 | 3306.61 \u00b1 59.73; llama 7B Q4_0 | tg128 | 59.06 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-16376603",
   "tg_tokens": 128,
   "tps": 59.06,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-30",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-disc-15013-c16376603-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b08580",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4016.46,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GB10; dev=CUDA0; llama 7B Q4_0 | pp512 | 4016.46 \u00b1 43.89; llama 7B Q4_0 | tg128 | 59.62 \u00b1 0.41",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-16376603",
   "tg_tokens": 128,
   "tps": 59.62,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-30",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-disc-15013-c16378182-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b08582",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3261.99,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GB10; llama 7B Q4_0 | pp512 | 3261.99 \u00b1 71.57; llama 7B Q4_0 | tg128 | 58.38 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-16378182",
   "tg_tokens": 128,
   "tps": 58.38,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-30",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-disc-15013-c16378182-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b08582",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3954.97,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GB10; llama 7B Q4_0 | pp512 | 3954.97 \u00b1 53.51; llama 7B Q4_0 | tg128 | 59.05 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-16378182",
   "tg_tokens": 128,
   "tps": 59.05,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-04-02",
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "lc-disc-15013-c16420189-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b08631",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 15928.07,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX PRO 6000 Blackwell; llama 7B Q4_0 | pp512 | 15928.07 \u00b1 800.27; llama 7B Q4_0 | pp1024 | 21464.62 \u00b1 21.04; llama 7B Q4_0 | pp2048 | 26467.24 \u00b1 10.96; llama 7B Q4_0 | pp4096 | 26840.74 \u00b1 26.38; llama 7B Q4_0 | pp8192 | 23064.67 \u00b1 4.40; llama 7B Q4_0 | pp16384 | 16623.47 \u00b1 4.44; llama 7B Q4_0 | pp32768 | 9726.40 \u00b1 2.77; llama 7B Q4_0 | tg128 | 264.66 \u00b1 0.20; llama 7B Q4_0 | tg256 | 264.72 \u00b1 0.05; llama 7B Q4_0 | tg512 | 255.24 \u00b1 0.67; llama 7B Q4_0 | tg1024 | 252.82 \u00b1 0.14; llama 7B Q4_0 | tg2048 | 243.36 \u00b1 0.16",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-16420189",
   "tg_tokens": 128,
   "tps": 264.66,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-04-02",
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "lc-disc-15013-c16420189-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b08631",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 18012.64,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX PRO 6000 Blackwell; llama 7B Q4_0 | pp512 | 18012.64 \u00b1 298.36; llama 7B Q4_0 | pp1024 | 19095.11 \u00b1 7.86; llama 7B Q4_0 | pp2048 | 19860.15 \u00b1 16.29; llama 7B Q4_0 | pp4096 | 19489.79 \u00b1 8.16; llama 7B Q4_0 | pp8192 | 17749.22 \u00b1 3.45; llama 7B Q4_0 | pp16384 | 14526.55 \u00b1 6.11; llama 7B Q4_0 | pp32768 | 9486.20 \u00b1 1.76; llama 7B Q4_0 | tg128 | 275.79 \u00b1 0.19; llama 7B Q4_0 | tg256 | 276.16 \u00b1 0.08; llama 7B Q4_0 | tg512 | 271.15 \u00b1 0.37; llama 7B Q4_0 | tg1024 | 265.54 \u00b1 0.20; llama 7B Q4_0 | tg2048 | 253.86 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-16420189",
   "tg_tokens": 128,
   "tps": 275.79,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-04-03",
   "flags": [],
   "hardware": "GeForce RTX 3080",
   "id": "lc-disc-15013-c16432553-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b08642",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 5438.27,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 3080; dev=CUDA0; llama 7B Q4_0 | pp512 | 5438.27 \u00b1 154.07; llama 7B Q4_0 | tg128 | 143.88 \u00b1 1.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-16432553",
   "tg_tokens": 128,
   "tps": 143.88,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-03",
   "flags": [],
   "hardware": "GeForce RTX 3090",
   "id": "lc-disc-15013-c16797292-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b09011",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 5349.43,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 3090; llama 7B Q4_0 | pp512 | 5349.43 \u00b1 78.79; llama 7B Q4_0 | tg128 | 168.83 \u00b1 0.20",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-16797292",
   "tg_tokens": 128,
   "tps": 168.83,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-03",
   "flags": [],
   "hardware": "GeForce RTX 3090",
   "id": "lc-disc-15013-c16797292-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b09011",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 5992.1,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 3090; llama 7B Q4_0 | pp512 | 5992.10 \u00b1 102.40; llama 7B Q4_0 | tg128 | 177.79 \u00b1 0.13",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-16797292",
   "tg_tokens": 128,
   "tps": 177.79,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-08",
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "lc-disc-15013-c16854411-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 12242.46,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX PRO 6000 Blackwell; llama 7B Q4_0 | pp512 | 12242.46 \u00b1 390.40; llama 7B Q4_0 | tg128 | 271.26 \u00b1 0.43",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-16854411",
   "tg_tokens": 128,
   "tps": 271.26,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-08",
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "lc-disc-15013-c16854411-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 13403.7,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX PRO 6000 Blackwell; llama 7B Q4_0 | pp512 | 13403.70 \u00b1 209.97; llama 7B Q4_0 | tg128 | 287.29 \u00b1 0.21",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-16854411",
   "tg_tokens": 128,
   "tps": 287.29,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-08",
   "flags": [],
   "hardware": "GeForce RTX 5060 Ti",
   "id": "lc-disc-15013-c16854411-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3733.17,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5060 Ti; llama 7B Q4_0 | pp512 | 3733.17 + 78.77; llama 7B Q4_0 | tg128 | 92.52 + 0.18",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-16854411",
   "tg_tokens": 128,
   "tps": 92.52,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-08",
   "flags": [],
   "hardware": "GeForce RTX 5060 Ti",
   "id": "lc-disc-15013-c16854411-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4427.42,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5060 Ti; llama 7B Q4_0 | pp512 | 4427.42 + 58.40; llama 7B Q4_0 | tg128 | 95.69 + 0.27",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-16854411",
   "tg_tokens": 128,
   "tps": 95.69,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-08",
   "flags": [],
   "hardware": "GeForce RTX 5060 Ti",
   "id": "lc-disc-15013-c16854411-llama7b-Q4_0-5",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3541.12,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5060 Ti; llama 7B Q4_0 | pp512 | 3541.12 + 73.58; llama 7B Q4_0 | tg128 | 92.75 + 0.12",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-16854411",
   "tg_tokens": 128,
   "tps": 92.75,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-08",
   "flags": [],
   "hardware": "GeForce RTX 5060 Ti",
   "id": "lc-disc-15013-c16854411-llama7b-Q4_0-6",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4430.01,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5060 Ti; llama 7B Q4_0 | pp512 | 4430.01 + 77.59; llama 7B Q4_0 | tg128 | 95.10 + 1.32",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-16854411",
   "tg_tokens": 128,
   "tps": 95.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "GeForce RTX 5060 Ti",
   "id": "lc-disc-15013-c16854411-llama7b-Q4_0-7",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 416.06,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5060 Ti; llama 7B Q4_0 | pp512 | 416.06 + 0.26; llama 7B Q4_0 | tg128 | 52.85 + 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-16854411",
   "tg_tokens": 128,
   "tps": 52.85,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "GeForce RTX 5060 Ti",
   "id": "lc-disc-15013-c16854411-llama7b-Q4_0-8",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 428.35,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5060 Ti; llama 7B Q4_0 | pp512 | 428.35 + 0.07; llama 7B Q4_0 | tg128 | 53.65 + 0.10",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-16854411",
   "tg_tokens": 128,
   "tps": 53.65,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "GeForce RTX 5060 Ti",
   "id": "lc-disc-15013-c16854411-llama7b-Q4_0-9",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 913.16,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5060 Ti; llama 7B Q4_0 | pp512 | 913.16 + 2.56; llama 7B Q4_0 | tg128 | 51.30 + 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-16854411",
   "tg_tokens": 128,
   "tps": 51.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "GeForce RTX 5060 Ti",
   "id": "lc-disc-15013-c16854411-llama7b-Q4_0-10",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1037.02,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5060 Ti; llama 7B Q4_0 | pp512 | 1037.02 + 1.01; llama 7B Q4_0 | tg128 | 53.96 + 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-16854411",
   "tg_tokens": 128,
   "tps": 53.96,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-08",
   "flags": [],
   "hardware": "GeForce RTX 5060 Ti",
   "id": "lc-disc-15013-c16854411-llama7b-Q4_0-11",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3359.51,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5060 Ti; llama 7B Q4_0 | pp512 | 3359.51 + 21.54; llama 7B Q4_0 | tg128 | 90.61 + 1.23",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-16854411",
   "tg_tokens": 128,
   "tps": 90.61,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-08",
   "flags": [],
   "hardware": "GeForce RTX 5060 Ti",
   "id": "lc-disc-15013-c16854411-llama7b-Q4_0-12",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3731.36,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5060 Ti; llama 7B Q4_0 | pp512 | 3731.36 + 102.20; llama 7B Q4_0 | tg128 | 95.84 + 0.36",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-16854411",
   "tg_tokens": 128,
   "tps": 95.84,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "GeForce RTX 5060 Ti",
   "id": "lc-disc-15013-c16854411-llama7b-Q4_0-13",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 188.88,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5060 Ti; llama 7B Q4_0 | pp512 | 188.88 + 0.01; llama 7B Q4_0 | tg128 | 34.19 + 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-16854411",
   "tg_tokens": 128,
   "tps": 34.19,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "GeForce RTX 5060 Ti",
   "id": "lc-disc-15013-c16854411-llama7b-Q4_0-14",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 189.5,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5060 Ti; llama 7B Q4_0 | pp512 | 189.50 + 0.01; llama 7B Q4_0 | tg128 | 34.71 + 0.10",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-16854411",
   "tg_tokens": 128,
   "tps": 34.71,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "GeForce RTX 5060 Ti",
   "id": "lc-disc-15013-c16854411-llama7b-Q4_0-15",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 518.53,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5060 Ti; llama 7B Q4_0 | pp512 | 518.53 + 0.22; llama 7B Q4_0 | tg128 | 63.79 + 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-16854411",
   "tg_tokens": 128,
   "tps": 63.79,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "GeForce RTX 5060 Ti",
   "id": "lc-disc-15013-c16854411-llama7b-Q4_0-16",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 572.26,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5060 Ti; llama 7B Q4_0 | pp512 | 572.26 + 0.19; llama 7B Q4_0 | tg128 | 66.59 + 0.32",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-16854411",
   "tg_tokens": 128,
   "tps": 66.59,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-13",
   "flags": [],
   "hardware": "RTX PRO 4000 Blackwell",
   "id": "lc-disc-15013-c16909094-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3628.68,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX PRO 4000 Blackwell; llama 7B Q4_0 | pp512 | 3628.68 \u00b1 42.39; llama 7B Q4_0 | tg128 | 89.73 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-16909094",
   "tg_tokens": 128,
   "tps": 89.73,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-13",
   "flags": [],
   "hardware": "RTX PRO 4000 Blackwell",
   "id": "lc-disc-15013-c16909094-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4078.52,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX PRO 4000 Blackwell; llama 7B Q4_0 | pp512 | 4078.52 \u00b1 15.41; llama 7B Q4_0 | tg128 | 92.54 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-16909094",
   "tg_tokens": 128,
   "tps": 92.54,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "A100",
   "id": "lc-disc-15013-c16909094-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 626.32,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A100; llama 7B Q4_0 | pp512 | 626.32 \u00b1 3.29; llama 7B Q4_0 | tg128 | 65.48 \u00b1 0.10",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-16909094",
   "tg_tokens": 128,
   "tps": 65.48,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "A100",
   "id": "lc-disc-15013-c16909094-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 653.01,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A100; llama 7B Q4_0 | pp512 | 653.01 \u00b1 1.92; llama 7B Q4_0 | tg128 | 65.91 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-16909094",
   "tg_tokens": 128,
   "tps": 65.91,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-26",
   "flags": [],
   "hardware": "A100",
   "id": "lc-disc-15013-c17057819-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; ngl=99; config=b08863",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1175.43,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A100; llama 7B Q4_0 | pp512 | 1175.43 \u00b1 13.69; llama 7B Q4_0 | tg128 | 128.41 \u00b1 0.10",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-17057819",
   "tg_tokens": 128,
   "tps": 128.41,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-25",
   "flags": [],
   "hardware": "A5500",
   "id": "lc-disc-15013-c17054190-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2180.23,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A5500; llama 7B Q4_0 | pp512 | 2180.23 \u00b1 55.75; llama 7B Q4_0 | tg128 | 74.68 \u00b1 0.84",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-17054190",
   "tg_tokens": 128,
   "tps": 74.68,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-25",
   "flags": [],
   "hardware": "A5500",
   "id": "lc-disc-15013-c17054190-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2338.14,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A5500; llama 7B Q4_0 | pp512 | 2338.14 \u00b1 53.16; llama 7B Q4_0 | tg128 | 74.66 \u00b1 1.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-17054190",
   "tg_tokens": 128,
   "tps": 74.66,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-25",
   "flags": [],
   "hardware": "Quadro RTX 5000",
   "id": "lc-disc-15013-c17054190-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1840.52,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Quadro RTX 5000; llama 7B Q4_0 | pp512 | 1840.52 \u00b1 22.87; llama 7B Q4_0 | tg128 | 74.76 \u00b1 0.34",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-17054190",
   "tg_tokens": 128,
   "tps": 74.76,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-25",
   "flags": [],
   "hardware": "Quadro RTX 5000",
   "id": "lc-disc-15013-c17054190-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1970.1,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Quadro RTX 5000; llama 7B Q4_0 | pp512 | 1970.10 \u00b1 12.48; llama 7B Q4_0 | tg128 | 77.74 \u00b1 0.32",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-17054190",
   "tg_tokens": 128,
   "tps": 77.74,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-25",
   "flags": [
    "contradiction"
   ],
   "hardware": "Quadro RTX 5000",
   "id": "lc-disc-15013-c17054190-llama7b-Q4_0-5",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 216.34,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Quadro RTX 5000; llama 7B Q4_0 | pp512 | 216.34 \u00b1 0.58; llama 7B Q4_0 | tg128 | 16.74 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-17054190",
   "tg_tokens": 128,
   "tps": 16.74,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-25",
   "flags": [
    "contradiction"
   ],
   "hardware": "Quadro RTX 5000",
   "id": "lc-disc-15013-c17054190-llama7b-Q4_0-6",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 218.32,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Quadro RTX 5000; llama 7B Q4_0 | pp512 | 218.32 \u00b1 0.06; llama 7B Q4_0 | tg128 | 17.84 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-17054190",
   "tg_tokens": 128,
   "tps": 17.84,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-25",
   "flags": [],
   "hardware": "Quadro RTX 5000",
   "id": "lc-disc-15013-c17054190-llama7b-Q4_0-7",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 545.12,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Quadro RTX 5000; llama 7B Q4_0 | pp512 | 545.12 \u00b1 1.57; llama 7B Q4_0 | tg128 | 34.71 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-17054190",
   "tg_tokens": 128,
   "tps": 34.71,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-25",
   "flags": [],
   "hardware": "Quadro RTX 5000",
   "id": "lc-disc-15013-c17054190-llama7b-Q4_0-8",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 549.43,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Quadro RTX 5000; llama 7B Q4_0 | pp512 | 549.43 \u00b1 0.57; llama 7B Q4_0 | tg128 | 36.11 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-17054190",
   "tg_tokens": 128,
   "tps": 36.11,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-26",
   "flags": [],
   "hardware": "V100",
   "id": "lc-disc-15013-c17059084-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99 sm=none; config=b09281",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 354.07,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "V100; llama 7B Q4_0 | pp512 | 354.07 \u00b1 0.03; llama 7B Q4_0 | tg128 | 99.84 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-17059084",
   "tg_tokens": 128,
   "tps": 99.84,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-26",
   "flags": [],
   "hardware": "V100",
   "id": "lc-disc-15013-c17059084-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99 sm=none; config=b09281",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 380.4,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "V100; llama 7B Q4_0 | pp512 | 380.40 \u00b1 0.03; llama 7B Q4_0 | tg128 | 109.49 \u00b1 0.12",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-17059084",
   "tg_tokens": 128,
   "tps": 109.49,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-26",
   "flags": [],
   "hardware": "V100",
   "id": "lc-disc-15013-c17059084-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; config=b09281",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 977.33,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "V100; llama 7B Q4_0 | pp512 | 977.33 \u00b1 0.83; llama 7B Q4_0 | tg128 | 99.95 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-17059084",
   "tg_tokens": 128,
   "tps": 99.95,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-26",
   "flags": [],
   "hardware": "V100",
   "id": "lc-disc-15013-c17059084-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b09281",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1156.76,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "V100; llama 7B Q4_0 | pp512 | 1156.76 \u00b1 6.05; llama 7B Q4_0 | tg128 | 109.46 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-17059084",
   "tg_tokens": 128,
   "tps": 109.46,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-26",
   "flags": [
    "contradiction"
   ],
   "hardware": "GeForce RTX 5090",
   "id": "lc-disc-15013-c17067742-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b09253",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 5056.95,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5090; llama 7B Q4_0 | pp512 | 5056.95 \u00b1 58.31; llama 7B Q4_0 | tg128 | 112.36 \u00b1 8.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-17067742",
   "tg_tokens": 128,
   "tps": 112.36,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-26",
   "flags": [
    "contradiction"
   ],
   "hardware": "GeForce RTX 5090",
   "id": "lc-disc-15013-c17067742-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b09253",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 5007.9,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5090; llama 7B Q4_0 | pp512 | 5007.90 \u00b1 72.08; llama 7B Q4_0 | tg128 | 117.16 \u00b1 0.57",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-17067742",
   "tg_tokens": 128,
   "tps": 117.16,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-26",
   "flags": [],
   "hardware": "GeForce RTX 5090",
   "id": "lc-disc-15013-c17067742-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b09253",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 6667.06,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5090; llama 7B Q4_0 | pp512 | 6667.06 \u00b1 123.40; llama 7B Q4_0 | tg128 | 156.49 \u00b1 0.98",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-17067742",
   "tg_tokens": 128,
   "tps": 156.49,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-26",
   "flags": [],
   "hardware": "GeForce RTX 5090",
   "id": "lc-disc-15013-c17067742-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b09253",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 7641.89,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5090; llama 7B Q4_0 | pp512 | 7641.89 \u00b1 275.60; llama 7B Q4_0 | tg128 | 158.14 \u00b1 4.70",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-17067742",
   "tg_tokens": 128,
   "tps": 158.14,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-23",
   "flags": [],
   "hardware": "GeForce RTX 3080",
   "id": "lc-disc-15013-c14196402-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=8; config=b06256",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4293.66,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 3080; llama 7B Q4_0 | pp512 | 4293.66 \u00b1 24.28; llama 7B Q4_0 | tg128 | 132.18 \u00b1 0.18",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14196402",
   "tg_tokens": 128,
   "tps": 132.18,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-23",
   "flags": [],
   "hardware": "GeForce RTX 3080",
   "id": "lc-disc-15013-c14196402-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=8; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06256",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4901.77,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 3080; llama 7B Q4_0 | pp512 | 4901.77 \u00b1 11.63; llama 7B Q4_0 | tg128 | 135.07 \u00b1 0.20",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14196402",
   "tg_tokens": 128,
   "tps": 135.07,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-23",
   "flags": [],
   "hardware": "P100",
   "id": "lc-disc-15013-c14196835-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99 sm=none; config=b06256",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 703.27,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "P100; llama 7B Q4_0 | pp512 | 703.27 \u00b1 3.21; llama 7B Q4_0 | tg128 | 50.20 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14196835",
   "tg_tokens": 128,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-23",
   "flags": [],
   "hardware": "P100",
   "id": "lc-disc-15013-c14196835-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99 sm=none; config=b06256",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 735.19,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "P100; llama 7B Q4_0 | pp512 | 735.19 \u00b1 3.72; llama 7B Q4_0 | tg128 | 51.08 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14196835",
   "tg_tokens": 128,
   "tps": 51.08,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-24",
   "flags": [],
   "hardware": "GeForce RTX 4090",
   "id": "lc-disc-15013-c14200008-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b06123",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 10293.86,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 4090; llama 7B Q4_0 | pp512 | 10293.86 \u00b1 134.72; llama 7B Q4_0 | tg128 | 189.33 \u00b1 0.19",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14200008",
   "tg_tokens": 128,
   "tps": 189.33,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-24",
   "flags": [],
   "hardware": "GeForce RTX 4090",
   "id": "lc-disc-15013-c14200008-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06123",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 12506.97,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 4090; llama 7B Q4_0 | pp512 | 12506.97 \u00b1 11.51; llama 7B Q4_0 | tg128 | 191.57 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14200008",
   "tg_tokens": 128,
   "tps": 191.57,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-24",
   "flags": [],
   "hardware": "GeForce RTX 4090",
   "id": "lc-disc-15013-c14200008-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; config=b06123",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 10259.68,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 4090; llama 7B Q4_0 | pp512 | 10259.68 \u00b1 153.00; llama 7B Q4_0 | pp1024 | 12555.82 \u00b1 28.04; llama 7B Q4_0 | pp2048 | 13183.82 \u00b1 20.22; llama 7B Q4_0 | pp4096 | 11987.23 \u00b1 6.71; llama 7B Q4_0 | pp8192 | 9442.55 \u00b1 4.48; llama 7B Q4_0 | pp16384 | 6051.97 \u00b1 1.31; llama 7B Q4_0 | pp32768 | 3492.30 \u00b1 0.90; llama 7B Q4_0 | tg128 | 187.83 \u00b1 2.06; llama 7B Q4_0 | tg256 | 186.96 \u00b1 0.14; llama 7B Q4_0 | tg512 | 183.08 \u00b1 0.05; llama 7B Q4_0 | tg1024 | 179.21 \u00b1 0.22; llama 7B Q4_0 | tg2048 | 170.17 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14200008",
   "tg_tokens": 128,
   "tps": 187.83,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-24",
   "flags": [],
   "hardware": "GeForce RTX 4090",
   "id": "lc-disc-15013-c14200008-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06123",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 12415.31,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 4090; llama 7B Q4_0 | pp512 | 12415.31 \u00b1 20.22; llama 7B Q4_0 | pp1024 | 16077.52 \u00b1 34.16; llama 7B Q4_0 | pp2048 | 18410.38 \u00b1 45.13; llama 7B Q4_0 | pp4096 | 19138.08 \u00b1 19.12; llama 7B Q4_0 | pp8192 | 16430.60 \u00b1 2.81; llama 7B Q4_0 | pp16384 | 12031.88 \u00b1 9.03; llama 7B Q4_0 | pp32768 | 8000.82 \u00b1 2.65; llama 7B Q4_0 | tg128 | 189.34 \u00b1 4.04; llama 7B Q4_0 | tg256 | 191.15 \u00b1 0.00; llama 7B Q4_0 | tg512 | 188.01 \u00b1 0.20; llama 7B Q4_0 | tg1024 | 182.59 \u00b1 0.07; llama 7B Q4_0 | tg2048 | 173.16 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14200008",
   "tg_tokens": 128,
   "tps": 189.34,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-27",
   "flags": [],
   "hardware": "GeForce RTX 4090",
   "id": "lc-disc-15013-c14234178-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 10712.64,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 4090; llama 7B Q4_0 | pp512 | 10712.64 \u00b1 135.22; llama 7B Q4_0 | pp1024 | 9831.51 \u00b1 120.33; llama 7B Q4_0 | pp2048 | 8534.83 \u00b1 94.59; llama 7B Q4_0 | pp4096 | 6939.80 \u00b1 24.46; llama 7B Q4_0 | pp8192 | 5036.04 \u00b1 6.41; llama 7B Q4_0 | pp16384 | 3012.11 \u00b1 1.49; llama 7B Q4_0 | tg128 | 167.53 \u00b1 1.89; llama 7B Q4_0 | tg256 | 166.85 \u00b1 0.19; llama 7B Q4_0 | tg512 | 163.54 \u00b1 0.09; llama 7B Q4_0 | tg1024 | 159.41 \u00b1 0.17; llama 7B Q4_0 | tg2048 | 150.76 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14234178",
   "tg_tokens": 128,
   "tps": 167.53,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-27",
   "flags": [],
   "hardware": "GeForce RTX 4090",
   "id": "lc-disc-15013-c14234178-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 12416.64,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 4090; llama 7B Q4_0 | pp512 | 12416.64 \u00b1 153.67; llama 7B Q4_0 | pp1024 | 12362.39 \u00b1 96.52; llama 7B Q4_0 | pp2048 | 11764.82 \u00b1 198.87; llama 7B Q4_0 | pp4096 | 11016.56 \u00b1 151.08; llama 7B Q4_0 | pp8192 | 9918.19 \u00b1 27.37; llama 7B Q4_0 | pp16384 | 8449.90 \u00b1 5.87; llama 7B Q4_0 | tg128 | 169.74 \u00b1 2.16; llama 7B Q4_0 | tg256 | 170.65 \u00b1 0.06; llama 7B Q4_0 | tg512 | 167.57 \u00b1 0.18; llama 7B Q4_0 | tg1024 | 162.95 \u00b1 0.10; llama 7B Q4_0 | tg2048 | 154.41 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14234178",
   "tg_tokens": 128,
   "tps": 169.74,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-27",
   "flags": [],
   "hardware": "GeForce RTX 4090",
   "id": "lc-disc-15013-c14234178-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 11773.24,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 4090; llama 7B Q4_0 | pp512 | 11773.24 \u00b1 91.37; llama 7B Q4_0 | pp1024 | 10885.59 \u00b1 89.51; llama 7B Q4_0 | pp2048 | 9388.77 \u00b1 8.90; llama 7B Q4_0 | pp4096 | 7400.17 \u00b1 3.67; llama 7B Q4_0 | pp8192 | 5216.83 \u00b1 0.57; llama 7B Q4_0 | pp16384 | 3050.56 \u00b1 0.83; llama 7B Q4_0 | tg128 | 167.52 \u00b1 1.75; llama 7B Q4_0 | tg256 | 166.90 \u00b1 0.22; llama 7B Q4_0 | tg512 | 163.71 \u00b1 0.09; llama 7B Q4_0 | tg1024 | 159.68 \u00b1 0.17; llama 7B Q4_0 | tg2048 | 151.05 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14234178",
   "tg_tokens": 128,
   "tps": 167.52,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-27",
   "flags": [],
   "hardware": "GeForce RTX 4090",
   "id": "lc-disc-15013-c14234178-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 14017.58,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 4090; llama 7B Q4_0 | pp512 | 14017.58 \u00b1 118.81; llama 7B Q4_0 | pp1024 | 14209.33 \u00b1 21.17; llama 7B Q4_0 | pp2048 | 13843.49 \u00b1 134.12; llama 7B Q4_0 | pp4096 | 13251.65 \u00b1 133.79; llama 7B Q4_0 | pp8192 | 12011.00 \u00b1 38.03; llama 7B Q4_0 | pp16384 | 10170.22 \u00b1 6.53; llama 7B Q4_0 | tg128 | 170.12 \u00b1 2.22; llama 7B Q4_0 | tg256 | 171.05 \u00b1 0.04; llama 7B Q4_0 | tg512 | 167.87 \u00b1 0.21; llama 7B Q4_0 | tg1024 | 163.23 \u00b1 0.12; llama 7B Q4_0 | tg2048 | 154.68 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14234178",
   "tg_tokens": 128,
   "tps": 170.12,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-27",
   "flags": [],
   "hardware": "GeForce RTX 4090",
   "id": "lc-disc-15013-c14234178-llama7b-Q4_0-5",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 10712.64,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 4090; llama 7B Q4_0 | pp512 | 10712.64 \u00b1 135.22; llama 7B Q4_0 | pp1024 | 9831.51 \u00b1 120.33; llama 7B Q4_0 | pp2048 | 8534.83 \u00b1 94.59; llama 7B Q4_0 | pp4096 | 6939.80 \u00b1 24.46; llama 7B Q4_0 | pp8192 | 5036.04 \u00b1 6.41; llama 7B Q4_0 | pp16384 | 3012.11 \u00b1 1.49; llama 7B Q4_0 | tg128 | 167.53 \u00b1 1.89; llama 7B Q4_0 | tg256 | 166.85 \u00b1 0.19; llama 7B Q4_0 | tg512 | 163.54 \u00b1 0.09; llama 7B Q4_0 | tg1024 | 159.41 \u00b1 0.17; llama 7B Q4_0 | tg2048 | 150.76 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14234178",
   "tg_tokens": 128,
   "tps": 167.53,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-27",
   "flags": [],
   "hardware": "GeForce RTX 4090",
   "id": "lc-disc-15013-c14234178-llama7b-Q4_0-6",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 12416.64,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 4090; llama 7B Q4_0 | pp512 | 12416.64 \u00b1 153.67; llama 7B Q4_0 | pp1024 | 12362.39 \u00b1 96.52; llama 7B Q4_0 | pp2048 | 11764.82 \u00b1 198.87; llama 7B Q4_0 | pp4096 | 11016.56 \u00b1 151.08; llama 7B Q4_0 | pp8192 | 9918.19 \u00b1 27.37; llama 7B Q4_0 | pp16384 | 8449.90 \u00b1 5.87; llama 7B Q4_0 | tg128 | 169.74 \u00b1 2.16; llama 7B Q4_0 | tg256 | 170.65 \u00b1 0.06; llama 7B Q4_0 | tg512 | 167.57 \u00b1 0.18; llama 7B Q4_0 | tg1024 | 162.95 \u00b1 0.10; llama 7B Q4_0 | tg2048 | 154.41 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14234178",
   "tg_tokens": 128,
   "tps": 169.74,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-28",
   "flags": [],
   "hardware": "GeForce RTX 5080",
   "id": "lc-disc-15013-c14244491-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b06307",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 8297.36,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5080; llama 7B Q4_0 | pp512 | 8297.36 \u00b1 9.50; llama 7B Q4_0 | pp1024 | 7526.74 \u00b1 5.23; llama 7B Q4_0 | pp2048 | 6764.02 \u00b1 3.40; llama 7B Q4_0 | pp4096 | 5672.04 \u00b1 2.65; llama 7B Q4_0 | pp8192 | 4302.98 \u00b1 1.93; llama 7B Q4_0 | pp16384 | 2764.80 \u00b1 1.63; llama 7B Q4_0 | tg128 | 181.99 \u00b1 0.42; llama 7B Q4_0 | tg256 | 179.77 \u00b1 0.11; llama 7B Q4_0 | tg512 | 174.80 \u00b1 0.24; llama 7B Q4_0 | tg1024 | 171.27 \u00b1 0.04; llama 7B Q4_0 | tg2048 | 163.08 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14244491",
   "tg_tokens": 128,
   "tps": 181.99,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-28",
   "flags": [],
   "hardware": "GeForce RTX 5080",
   "id": "lc-disc-15013-c14244491-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06307",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 9487.7,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5080; llama 7B Q4_0 | pp512 | 9487.70 \u00b1 21.89; llama 7B Q4_0 | pp1024 | 9359.28 \u00b1 12.61; llama 7B Q4_0 | pp2048 | 8744.84 \u00b1 54.69; llama 7B Q4_0 | pp4096 | 8287.33 \u00b1 65.44; llama 7B Q4_0 | pp8192 | 7596.55 \u00b1 34.52; llama 7B Q4_0 | pp16384 | 6392.95 \u00b1 26.51; llama 7B Q4_0 | tg128 | 184.68 \u00b1 0.05; llama 7B Q4_0 | tg256 | 183.91 \u00b1 0.01; llama 7B Q4_0 | tg512 | 180.83 \u00b1 0.14; llama 7B Q4_0 | tg1024 | 175.42 \u00b1 0.08; llama 7B Q4_0 | tg2048 | 166.30 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14244491",
   "tg_tokens": 128,
   "tps": 184.68,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-01",
   "flags": [],
   "hardware": "A6000",
   "id": "lc-disc-15013-c14274344-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b06342",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4913.93,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A6000; llama 7B Q4_0 | pp512 | 4913.93 \u00b1 6.79; llama 7B Q4_0 | pp1024 | 4619.66 \u00b1 5.08; llama 7B Q4_0 | pp2048 | 4313.47 \u00b1 8.60; llama 7B Q4_0 | pp4096 | 3740.25 \u00b1 4.44; llama 7B Q4_0 | pp8192 | 2899.87 \u00b1 3.97; llama 7B Q4_0 | pp16384 | 1981.68 \u00b1 3.73; llama 7B Q4_0 | pp32768 | 1091.41 \u00b1 0.73; llama 7B Q4_0 | tg128 | 138.73 \u00b1 2.75; llama 7B Q4_0 | tg256 | 137.87 \u00b1 0.15; llama 7B Q4_0 | tg512 | 133.61 \u00b1 0.08; llama 7B Q4_0 | tg1024 | 131.28 \u00b1 0.01; llama 7B Q4_0 | tg2048 | 125.16 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14274344",
   "tg_tokens": 128,
   "tps": 138.73,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-01",
   "flags": [],
   "hardware": "A6000",
   "id": "lc-disc-15013-c14274344-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06342",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 5662.39,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A6000; llama 7B Q4_0 | pp512 | 5662.39 \u00b1 13.87; llama 7B Q4_0 | pp1024 | 5600.69 \u00b1 18.81; llama 7B Q4_0 | pp2048 | 5497.16 \u00b1 15.85; llama 7B Q4_0 | pp4096 | 5035.14 \u00b1 4.70; llama 7B Q4_0 | pp8192 | 4599.80 \u00b1 2.69; llama 7B Q4_0 | pp16384 | 3887.28 \u00b1 1.35; llama 7B Q4_0 | pp32768 | 2954.14 \u00b1 0.52; llama 7B Q4_0 | tg128 | 144.87 \u00b1 0.18; llama 7B Q4_0 | tg256 | 143.27 \u00b1 0.05; llama 7B Q4_0 | tg512 | 140.72 \u00b1 0.20; llama 7B Q4_0 | tg1024 | 136.76 \u00b1 0.12; llama 7B Q4_0 | tg2048 | 130.11 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14274344",
   "tg_tokens": 128,
   "tps": 144.87,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-05",
   "flags": [],
   "hardware": "A100",
   "id": "lc-disc-15013-c14322852-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b06392",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4883.69,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A100; llama 7B Q4_0 | pp512 | 4883.69 \u00b1 5.30; llama 7B Q4_0 | pp1024 | 4760.67 \u00b1 1.58; llama 7B Q4_0 | pp2048 | 4572.24 \u00b1 2.23; llama 7B Q4_0 | pp4096 | 4230.73 \u00b1 0.58; llama 7B Q4_0 | pp8192 | 3664.65 \u00b1 0.33; llama 7B Q4_0 | pp16384 | 2851.50 \u00b1 4.65; llama 7B Q4_0 | pp32768 | 1867.40 \u00b1 1.92; llama 7B Q4_0 | tg128 | 185.91 \u00b1 7.07; llama 7B Q4_0 | tg256 | 186.56 \u00b1 0.17; llama 7B Q4_0 | tg512 | 180.46 \u00b1 0.09; llama 7B Q4_0 | tg1024 | 178.87 \u00b1 0.06; llama 7B Q4_0 | tg2048 | 173.15 \u00b1 0.08; llama 7B Q4_0 | tg4096 | 163.02 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14322852",
   "tg_tokens": 128,
   "tps": 185.91,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-05",
   "flags": [],
   "hardware": "A100",
   "id": "lc-disc-15013-c14322852-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06392",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 5295.94,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A100; llama 7B Q4_0 | pp512 | 5295.94 \u00b1 5.49; llama 7B Q4_0 | pp1024 | 5315.97 \u00b1 0.22; llama 7B Q4_0 | pp2048 | 5265.85 \u00b1 0.65; llama 7B Q4_0 | pp4096 | 5118.09 \u00b1 3.22; llama 7B Q4_0 | pp8192 | 4822.77 \u00b1 1.59; llama 7B Q4_0 | pp16384 | 4306.81 \u00b1 0.93; llama 7B Q4_0 | pp32768 | 3544.53 \u00b1 4.04; llama 7B Q4_0 | tg128 | 196.26 \u00b1 2.41; llama 7B Q4_0 | tg256 | 197.44 \u00b1 0.06; llama 7B Q4_0 | tg512 | 193.92 \u00b1 0.40; llama 7B Q4_0 | tg1024 | 190.27 \u00b1 0.15; llama 7B Q4_0 | tg2048 | 182.79 \u00b1 0.20; llama 7B Q4_0 | tg4096 | 170.53 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14322852",
   "tg_tokens": 128,
   "tps": 196.26,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-05",
   "flags": [],
   "hardware": "A100",
   "id": "lc-disc-15013-c14322920-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b06392",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4849.53,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A100; llama 7B Q4_0 | pp512 | 4849.53 \u00b1 8.94; llama 7B Q4_0 | pp1024 | 4755.22 \u00b1 8.14; llama 7B Q4_0 | pp2048 | 4569.57 \u00b1 6.91; llama 7B Q4_0 | pp4096 | 4239.42 \u00b1 5.94; llama 7B Q4_0 | pp8192 | 3663.99 \u00b1 1.79; llama 7B Q4_0 | pp16384 | 2894.84 \u00b1 0.85; llama 7B Q4_0 | pp32784 | 1918.93 \u00b1 0.37; llama 7B Q4_0 | tg128 | 190.88 \u00b1 0.33; llama 7B Q4_0 | tg256 | 188.06 \u00b1 0.29; llama 7B Q4_0 | tg512 | 182.03 \u00b1 0.10; llama 7B Q4_0 | tg1024 | 180.78 \u00b1 0.71; llama 7B Q4_0 | tg2048 | 175.78 \u00b1 0.16; llama 7B Q4_0 | tg4096 | 165.85 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14322920",
   "tg_tokens": 128,
   "tps": 190.88,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-05",
   "flags": [],
   "hardware": "A100",
   "id": "lc-disc-15013-c14322920-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06392",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 5285.96,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A100; llama 7B Q4_0 | pp512 | 5285.96 \u00b1 6.58; llama 7B Q4_0 | pp1024 | 5308.42 \u00b1 8.91; llama 7B Q4_0 | pp2048 | 5282.45 \u00b1 5.67; llama 7B Q4_0 | pp4096 | 5150.01 \u00b1 5.38; llama 7B Q4_0 | pp8192 | 4883.68 \u00b1 3.31; llama 7B Q4_0 | pp16384 | 4426.85 \u00b1 0.97; llama 7B Q4_0 | pp32784 | 3708.18 \u00b1 1.05; llama 7B Q4_0 | tg128 | 200.90 \u00b1 0.12; llama 7B Q4_0 | tg256 | 200.80 \u00b1 0.11; llama 7B Q4_0 | tg512 | 197.55 \u00b1 0.33; llama 7B Q4_0 | tg1024 | 193.66 \u00b1 0.29; llama 7B Q4_0 | tg2048 | 186.59 \u00b1 0.08; llama 7B Q4_0 | tg4096 | 174.47 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14322920",
   "tg_tokens": 128,
   "tps": 200.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-05",
   "flags": [
    "contradiction"
   ],
   "hardware": "H100",
   "id": "lc-disc-15013-c14322934-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b06392",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 7350.98,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "H100; llama 7B Q4_0 | pp512 | 7350.98 \u00b1 63.64; llama 7B Q4_0 | pp1024 | 7277.24 \u00b1 3.96; llama 7B Q4_0 | pp2048 | 6976.45 \u00b1 8.06; llama 7B Q4_0 | pp4096 | 6370.87 \u00b1 11.90; llama 7B Q4_0 | pp8192 | 5298.29 \u00b1 4.18; llama 7B Q4_0 | pp16384 | 3836.34 \u00b1 6.00; llama 7B Q4_0 | pp32784 | 2435.55 \u00b1 3.19; llama 7B Q4_0 | tg128 | 219.23 \u00b1 0.55; llama 7B Q4_0 | tg256 | 215.16 \u00b1 0.89; llama 7B Q4_0 | tg512 | 208.08 \u00b1 0.40; llama 7B Q4_0 | tg1024 | 206.23 \u00b1 0.86; llama 7B Q4_0 | tg2048 | 200.75 \u00b1 0.09; llama 7B Q4_0 | tg4096 | 189.94 \u00b1 0.45",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14322934",
   "tg_tokens": 128,
   "tps": 219.23,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-05",
   "flags": [
    "contradiction"
   ],
   "hardware": "H100",
   "id": "lc-disc-15013-c14322934-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b06392",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 8457.52,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "H100; llama 7B Q4_0 | pp512 | 8457.52 \u00b1 18.97; llama 7B Q4_0 | pp1024 | 8513.08 \u00b1 13.55; llama 7B Q4_0 | pp2048 | 8442.74 \u00b1 29.20; llama 7B Q4_0 | pp4096 | 8110.29 \u00b1 7.67; llama 7B Q4_0 | pp8192 | 7335.59 \u00b1 8.79; llama 7B Q4_0 | pp16384 | 6214.56 \u00b1 4.93; llama 7B Q4_0 | pp32784 | 4837.63 \u00b1 2.62; llama 7B Q4_0 | tg128 | 229.85 \u00b1 0.07; llama 7B Q4_0 | tg256 | 230.12 \u00b1 0.09; llama 7B Q4_0 | tg512 | 227.18 \u00b1 0.37; llama 7B Q4_0 | tg1024 | 207.57 \u00b1 0.66; llama 7B Q4_0 | tg2048 | 213.82 \u00b1 0.35; llama 7B Q4_0 | tg4096 | 200.16 \u00b1 0.21",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14322934",
   "tg_tokens": 128,
   "tps": 229.85,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-05",
   "flags": [],
   "hardware": "H100",
   "id": "lc-disc-15013-c14322969-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b06392",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 9918.34,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "H100; llama 7B Q4_0 | pp512 | 9918.34 \u00b1 176.97; llama 7B Q4_0 | pp1024 | 9794.27 \u00b1 76.14; llama 7B Q4_0 | pp2048 | 9366.63 \u00b1 50.02; llama 7B Q4_0 | pp4096 | 8568.65 \u00b1 3.37; llama 7B Q4_0 | pp8192 | 7152.63 \u00b1 2.61; llama 7B Q4_0 | pp16384 | 5331.01 \u00b1 1.40; llama 7B Q4_0 | pp32784 | 3456.25 \u00b1 1.25; llama 7B Q4_0 | tg128 | 267.81 \u00b1 1.54; llama 7B Q4_0 | tg256 | 263.37 \u00b1 0.26; llama 7B Q4_0 | tg512 | 255.22 \u00b1 0.18; llama 7B Q4_0 | tg1024 | 253.44 \u00b1 0.25; llama 7B Q4_0 | tg2048 | 247.29 \u00b1 0.14; llama 7B Q4_0 | tg4096 | 236.07 \u00b1 0.10",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14322969",
   "tg_tokens": 128,
   "tps": 267.81,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-05",
   "flags": [],
   "hardware": "H100",
   "id": "lc-disc-15013-c14322969-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b06392",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 11263.29,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "H100; llama 7B Q4_0 | pp512 | 11263.29 \u00b1 98.34; llama 7B Q4_0 | pp1024 | 11454.74 \u00b1 8.49; llama 7B Q4_0 | pp2048 | 11331.01 \u00b1 10.68; llama 7B Q4_0 | pp4096 | 10935.28 \u00b1 34.86; llama 7B Q4_0 | pp8192 | 10340.75 \u00b1 4.30; llama 7B Q4_0 | pp16384 | 9211.07 \u00b1 2.62; llama 7B Q4_0 | pp32784 | 7407.12 \u00b1 6.49; llama 7B Q4_0 | tg128 | 280.74 \u00b1 1.17; llama 7B Q4_0 | tg256 | 280.73 \u00b1 0.03; llama 7B Q4_0 | tg512 | 277.60 \u00b1 0.51; llama 7B Q4_0 | tg1024 | 274.97 \u00b1 0.41; llama 7B Q4_0 | tg2048 | 264.73 \u00b1 0.31; llama 7B Q4_0 | tg4096 | 248.77 \u00b1 0.19",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14322969",
   "tg_tokens": 128,
   "tps": 280.74,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-05",
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "lc-disc-15013-c14323205-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b06392",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 16618.98,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX PRO 6000 Blackwell; llama 7B Q4_0 | pp512 | 16618.98 \u00b1 20.66; llama 7B Q4_0 | pp1024 | 16590.31 \u00b1 28.12; llama 7B Q4_0 | pp2048 | 16255.31 \u00b1 42.25; llama 7B Q4_0 | pp4096 | 15571.01 \u00b1 44.15; llama 7B Q4_0 | pp8192 | 14231.99 \u00b1 25.94; llama 7B Q4_0 | pp16384 | 11797.81 \u00b1 27.10; llama 7B Q4_0 | pp32768 | 7462.29 \u00b1 27.29; llama 7B Q4_0 | tg128 | 281.11 \u00b1 0.41; llama 7B Q4_0 | tg256 | 280.98 \u00b1 0.11; llama 7B Q4_0 | tg512 | 277.08 \u00b1 0.48; llama 7B Q4_0 | tg1024 | 270.57 \u00b1 0.17; llama 7B Q4_0 | tg2048 | 256.81 \u00b1 0.21",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14323205",
   "tg_tokens": 128,
   "tps": 281.11,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-05",
   "flags": [],
   "hardware": "RTX Pro 6000 Blackwell",
   "id": "lc-disc-15013-c14323205-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; config=b06392",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 14854.63,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX PRO 6000 Blackwell; llama 7B Q4_0 | pp512 | 14854.63 \u00b1 22.73; llama 7B Q4_0 | pp1024 | 14046.00 \u00b1 42.45; llama 7B Q4_0 | pp2048 | 12772.94 \u00b1 18.17; llama 7B Q4_0 | pp4096 | 10429.76 \u00b1 11.19; llama 7B Q4_0 | pp8192 | 7786.16 \u00b1 3.03; llama 7B Q4_0 | pp16384 | 5128.03 \u00b1 3.09; llama 7B Q4_0 | pp32768 | 2845.66 \u00b1 2.95; llama 7B Q4_0 | tg128 | 274.20 \u00b1 0.14; llama 7B Q4_0 | tg256 | 271.15 \u00b1 0.27; llama 7B Q4_0 | tg512 | 264.13 \u00b1 0.14; llama 7B Q4_0 | tg1024 | 260.42 \u00b1 0.12; llama 7B Q4_0 | tg2048 | 249.18 \u00b1 0.11",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14323205",
   "tg_tokens": 128,
   "tps": 274.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-16",
   "flags": [],
   "hardware": "GeForce RTX 5080",
   "id": "lc-disc-15013-c15269201-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b07342",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 9644.43,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5080; llama 7B Q4_0 | pp512 | 9644.43 \u00b1 93.39; llama 7B Q4_0 | pp1024 | 9613.14 \u00b1 47.35; llama 7B Q4_0 | pp2048 | 9353.44 \u00b1 35.83; llama 7B Q4_0 | pp4096 | 8731.62 \u00b1 26.45; llama 7B Q4_0 | pp16384 | 6034.36 \u00b1 4.19; llama 7B Q4_0 | tg128 | 183.04 \u00b1 0.50",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15269201",
   "tg_tokens": 128,
   "tps": 183.04,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-16",
   "flags": [],
   "hardware": "GeForce RTX 5080",
   "id": "lc-disc-15013-c15269201-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07342",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 7405.11,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5080; llama 7B Q4_0 | pp512 | 7405.11 \u00b1 125.96; llama 7B Q4_0 | pp1024 | 6748.15 \u00b1 71.84; llama 7B Q4_0 | pp2048 | 5839.27 \u00b1 17.09; llama 7B Q4_0 | pp4096 | 4680.56 \u00b1 9.23; llama 7B Q4_0 | pp16384 | 2086.91 \u00b1 1.08; llama 7B Q4_0 | tg128 | 177.99 \u00b1 0.44",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15269201",
   "tg_tokens": 128,
   "tps": 177.99,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-17",
   "flags": [],
   "hardware": "V100",
   "id": "lc-disc-15013-c15284016-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b07455",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3261.46,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "V100; llama 7B Q4_0 | pp512 | 3261.46 \u00b1 0.77; llama 7B Q4_0 | pp1024 | 3171.60 \u00b1 5.28; llama 7B Q4_0 | pp2048 | 2990.51 \u00b1 0.87; llama 7B Q4_0 | pp4096 | 2677.52 \u00b1 1.44; llama 7B Q4_0 | pp8192 | 2193.69 \u00b1 0.60; llama 7B Q4_0 | pp16384 | 1554.80 \u00b1 0.22; llama 7B Q4_0 | tg128 | 140.23 \u00b1 0.02; llama 7B Q4_0 | tg256 | 139.46 \u00b1 0.02; llama 7B Q4_0 | tg512 | 134.37 \u00b1 0.02; llama 7B Q4_0 | tg1024 | 134.53 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15284016",
   "tg_tokens": 128,
   "tps": 140.23,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-17",
   "flags": [],
   "hardware": "V100",
   "id": "lc-disc-15013-c15284016-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07455",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3537.71,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "V100; llama 7B Q4_0 | pp512 | 3537.71 \u00b1 2.89; llama 7B Q4_0 | pp1024 | 3483.48 \u00b1 4.54; llama 7B Q4_0 | pp2048 | 3403.40 \u00b1 4.02; llama 7B Q4_0 | pp4096 | 3255.39 \u00b1 2.25; llama 7B Q4_0 | pp8192 | 2972.18 \u00b1 1.25; llama 7B Q4_0 | pp16384 | 2510.32 \u00b1 0.64; llama 7B Q4_0 | tg128 | 152.33 \u00b1 0.08; llama 7B Q4_0 | tg256 | 151.61 \u00b1 0.04; llama 7B Q4_0 | tg512 | 149.94 \u00b1 0.03; llama 7B Q4_0 | tg1024 | 145.94 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15284016",
   "tg_tokens": 128,
   "tps": 152.33,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-17",
   "flags": [],
   "hardware": "V100",
   "id": "lc-disc-15013-c15284016-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; config=b07455",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3160.43,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "V100; llama 7B Q4_0 | pp512 | 3160.43 \u00b1 2.59; llama 7B Q4_0 | pp1024 | 3051.85 \u00b1 1.46; llama 7B Q4_0 | pp2048 | 2885.46 \u00b1 0.32; llama 7B Q4_0 | pp4096 | 2567.17 \u00b1 0.27; llama 7B Q4_0 | pp8192 | 2088.20 \u00b1 0.55; llama 7B Q4_0 | pp16384 | 1435.41 \u00b1 0.59; llama 7B Q4_0 | pp32768 | 787.45 \u00b1 0.36; llama 7B Q4_0 | tg128 | 133.06 \u00b1 0.03; llama 7B Q4_0 | tg256 | 133.34 \u00b1 0.00; llama 7B Q4_0 | tg512 | 128.80 \u00b1 0.01; llama 7B Q4_0 | tg1024 | 128.46 \u00b1 0.06; llama 7B Q4_0 | tg2048 | 124.14 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15284016",
   "tg_tokens": 128,
   "tps": 133.06,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-17",
   "flags": [],
   "hardware": "V100",
   "id": "lc-disc-15013-c15284016-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07455",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3434.83,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "V100; llama 7B Q4_0 | pp512 | 3434.83 \u00b1 0.75; llama 7B Q4_0 | pp1024 | 3376.63 \u00b1 6.88; llama 7B Q4_0 | pp2048 | 3286.03 \u00b1 5.56; llama 7B Q4_0 | pp4096 | 3158.13 \u00b1 3.73; llama 7B Q4_0 | pp8192 | 2892.14 \u00b1 2.53; llama 7B Q4_0 | pp16384 | 2467.85 \u00b1 0.59; llama 7B Q4_0 | pp32768 | 1899.52 \u00b1 1.26; llama 7B Q4_0 | tg128 | 145.45 \u00b1 0.01; llama 7B Q4_0 | tg256 | 144.88 \u00b1 0.01; llama 7B Q4_0 | tg512 | 143.24 \u00b1 0.05; llama 7B Q4_0 | tg1024 | 139.56 \u00b1 0.00; llama 7B Q4_0 | tg2048 | 133.02 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15284016",
   "tg_tokens": 128,
   "tps": 145.45,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-22",
   "flags": [],
   "hardware": "GeForce RTX 3080",
   "id": "lc-disc-15013-c15321427-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; 149.0 W measured nvidia-smi power.draw, short runs (approximate); fa=0 ngl=99; config=b07509",
   "params": null,
   "power_w": 149.0,
   "pp_tokens": 512,
   "pp_tps": 2970.97,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 3080; llama 7B Q4_0 | pp512 | 2970.97 \u00b1 14.10; llama 7B Q4_0 | pp1024 | 2776.18 \u00b1 10.49; llama 7B Q4_0 | pp2048 | 2553.65 \u00b1 5.01; llama 7B Q4_0 | pp4096 | 2213.43 \u00b1 4.22; llama 7B Q4_0 | tg128 | 90.78 \u00b1 0.08; llama 7B Q4_0 | tg256 | 90.62 \u00b1 0.15; llama 7B Q4_0 | tg512 | 87.45 \u00b1 0.16; llama 7B Q4_0 | tg1024 | 85.32 \u00b1 0.20; llama 7B Q4_0 | tg2048 | 81.04 \u00b1 0.16",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15321427",
   "tg_tokens": 128,
   "tps": 90.78,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-22",
   "flags": [],
   "hardware": "GeForce RTX 3080",
   "id": "lc-disc-15013-c15321427-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b07509",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3211.59,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 3080; llama 7B Q4_0 | pp512 | 3211.59 \u00b1 49.51; llama 7B Q4_0 | pp1024 | 3219.32 \u00b1 45.43; llama 7B Q4_0 | pp2048 | 3076.62 \u00b1 10.88; llama 7B Q4_0 | pp4096 | 2905.61 \u00b1 3.25; llama 7B Q4_0 | tg128 | 92.45 \u00b1 0.32; llama 7B Q4_0 | tg256 | 92.72 \u00b1 0.22; llama 7B Q4_0 | tg512 | 90.96 \u00b1 0.11; llama 7B Q4_0 | tg1024 | 88.24 \u00b1 0.06; llama 7B Q4_0 | tg2048 | 83.41 \u00b1 0.28",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15321427",
   "tg_tokens": 128,
   "tps": 92.45,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-01-22",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "lc-disc-15013-c15568954-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b07789",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 11443.35,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX PRO 6000; llama 7B Q4_0 | pp512 | 11443.35 \u00b1 163.44; llama 7B Q4_0 | tg128 | 270.00 \u00b1 0.85",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15568954",
   "tg_tokens": 128,
   "tps": 270.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-01-22",
   "flags": [],
   "hardware": "RTX PRO 6000",
   "id": "lc-disc-15013-c15568954-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07789",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 12231.9,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX PRO 6000; llama 7B Q4_0 | pp512 | 12231.90 \u00b1 260.54; llama 7B Q4_0 | tg128 | 269.96 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15568954",
   "tg_tokens": 128,
   "tps": 269.96,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-01-22",
   "flags": [],
   "hardware": "NVIDIA RTX 4500",
   "id": "lc-disc-15013-c15571400-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 5512.71,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "NVIDIA RTX 4500; llama 7B Q4_0 | pp512 | 5512.71 \u00b1 37.58; llama 7B Q4_0 | tg128 | 92.94 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15571400",
   "tg_tokens": 128,
   "tps": 92.94,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-01-22",
   "flags": [],
   "hardware": "NVIDIA RTX 4500",
   "id": "lc-disc-15013-c15571400-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 6490.28,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "NVIDIA RTX 4500; llama 7B Q4_0 | pp512 | 6490.28 \u00b1 33.89; llama 7B Q4_0 | tg128 | 93.88 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15571400",
   "tg_tokens": 128,
   "tps": 93.88,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-02-11",
   "flags": [],
   "hardware": "GeForce RTX 4060",
   "id": "lc-disc-15013-c15767585-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b07993",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2466.31,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 4060; llama 7B Q4_0 | pp512 | 2466.31 + 4.85; llama 7B Q4_0 | tg128 | 58.22 + 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15767585",
   "tg_tokens": 128,
   "tps": 58.22,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-02-11",
   "flags": [],
   "hardware": "GeForce RTX 4060",
   "id": "lc-disc-15013-c15767585-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07993",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2928.74,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 4060; llama 7B Q4_0 | pp512 | 2928.74 + 4.50; llama 7B Q4_0 | tg128 | 59.15 + 0.43",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15767585",
   "tg_tokens": 128,
   "tps": 59.15,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-02-11",
   "flags": [],
   "hardware": "GeForce RTX 4060",
   "id": "lc-disc-15013-c15767585-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; config=b07993",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2464.58,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 4060; llama 7B Q4_0 | pp512 | 2464.58 + 3.29; llama 7B Q4_0 | tg128 | 57.87 + 0.46; llama 7B Q4_0 | tg256 | 57.16 + 0.10; llama 7B Q4_0 | tg512 | 56.47 + 0.45; llama 7B Q4_0 | tg1024 | 54.72 + 0.61; llama 7B Q4_0 | tg2048 | 52.28 + 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15767585",
   "tg_tokens": 128,
   "tps": 57.87,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-02-11",
   "flags": [],
   "hardware": "GeForce RTX 4060",
   "id": "lc-disc-15013-c15767585-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07993",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2676.6,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 4060; llama 7B Q4_0 | pp512 | 2676.60 + 112.08; llama 7B Q4_0 | tg128 | 59.55 + 0.02; llama 7B Q4_0 | tg256 | 59.55 + 0.01; llama 7B Q4_0 | tg512 | 58.47 + 0.03; llama 7B Q4_0 | tg1024 | 55.96 + 0.16; llama 7B Q4_0 | tg2048 | 52.79 + 0.10",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15767585",
   "tg_tokens": 128,
   "tps": 59.55,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-02-16",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-disc-15013-c15827423-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b08076",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3241.51,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GB10; dev=CUDA0; llama 7B Q4_0 | pp512 | 3241.51 \u00b1 28.66; llama 7B Q4_0 | tg128 | 58.13 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15827423",
   "tg_tokens": 128,
   "tps": 58.13,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-02-16",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-disc-15013-c15827423-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b08076",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3861.19,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GB10; dev=CUDA0; llama 7B Q4_0 | pp512 | 3861.19 \u00b1 36.01; llama 7B Q4_0 | tg128 | 58.14 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15827423",
   "tg_tokens": 128,
   "tps": 58.14,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-02-19",
   "flags": [],
   "hardware": "GeForce RTX 3060",
   "id": "lc-disc-15013-c15860728-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b08099",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2070.83,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 3060; llama 7B Q4_0 | pp512 | 2070.83 \u00b1 10.45; llama 7B Q4_0 | tg128 | 70.52 \u00b1 0.19",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15860728",
   "tg_tokens": 128,
   "tps": 70.52,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-02-19",
   "flags": [],
   "hardware": "GeForce RTX 3060",
   "id": "lc-disc-15013-c15860728-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b08099",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2346.54,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 3060; llama 7B Q4_0 | pp512 | 2346.54 \u00b1 17.52; llama 7B Q4_0 | tg128 | 73.74 \u00b1 0.19",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15860728",
   "tg_tokens": 128,
   "tps": 73.74,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-02-28",
   "flags": [],
   "hardware": "RTX Pro 5000 Blackwell",
   "id": "lc-disc-15013-c15954419-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b08180",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 9952.91,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX PRO 5000 Blackwell; llama 7B Q4_0 | pp512 | 9952.91 \u00b1 548.93; llama 7B Q4_0 | pp1024 | 9182.67 \u00b1 80.42; llama 7B Q4_0 | pp2048 | 8199.69 \u00b1 52.66; llama 7B Q4_0 | pp4096 | 6855.22 \u00b1 12.20; llama 7B Q4_0 | pp8192 | 5151.20 \u00b1 10.88; llama 7B Q4_0 | pp16384 | 3450.67 \u00b1 9.95; llama 7B Q4_0 | pp32768 | 1978.52 \u00b1 7.40; llama 7B Q4_0 | tg128 | 213.54 \u00b1 0.26; llama 7B Q4_0 | tg256 | 213.07 \u00b1 0.13; llama 7B Q4_0 | tg512 | 206.08 \u00b1 0.27; llama 7B Q4_0 | tg1024 | 204.25 \u00b1 0.09; llama 7B Q4_0 | tg2048 | 196.40 \u00b1 0.10",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15954419",
   "tg_tokens": 128,
   "tps": 213.54,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-02-28",
   "flags": [],
   "hardware": "RTX Pro 5000 Blackwell",
   "id": "lc-disc-15013-c15954419-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b08180",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 11046.55,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX PRO 5000 Blackwell; llama 7B Q4_0 | pp512 | 11046.55 \u00b1 298.03; llama 7B Q4_0 | pp1024 | 10059.62 \u00b1 26.27; llama 7B Q4_0 | pp2048 | 9754.84 \u00b1 17.54; llama 7B Q4_0 | pp4096 | 9179.72 \u00b1 5.18; llama 7B Q4_0 | pp8192 | 7939.43 \u00b1 3.62; llama 7B Q4_0 | pp16384 | 6242.84 \u00b1 1.12; llama 7B Q4_0 | pp32768 | 4405.04 \u00b1 1.25; llama 7B Q4_0 | tg128 | 226.31 \u00b1 0.28; llama 7B Q4_0 | tg256 | 225.05 \u00b1 0.09; llama 7B Q4_0 | tg512 | 221.08 \u00b1 0.24; llama 7B Q4_0 | tg1024 | 215.66 \u00b1 0.13; llama 7B Q4_0 | tg2048 | 205.76 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15954419",
   "tg_tokens": 128,
   "tps": 226.31,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-15",
   "flags": [],
   "hardware": "Quadro RTX 6000",
   "id": "lc-disc-15013-c14398960-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b06475",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2751.18,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Quadro RTX 6000; llama 7B Q4_0 | pp512 | 2751.18 \u00b1 19.43; llama 7B Q4_0 | tg128 | 102.77 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14398960",
   "tg_tokens": 128,
   "tps": 102.77,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-15",
   "flags": [],
   "hardware": "Quadro RTX 6000",
   "id": "lc-disc-15013-c14398960-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06475",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3053.96,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Quadro RTX 6000; llama 7B Q4_0 | pp512 | 3053.96 \u00b1 1.37; llama 7B Q4_0 | tg128 | 104.38 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14398960",
   "tg_tokens": 128,
   "tps": 104.38,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-20",
   "flags": [],
   "hardware": "NVIDIA RTX 3500",
   "id": "lc-disc-15013-c14463481-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b06477",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1406.43,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "NVIDIA RTX 3500; llama 7B Q4_0 | pp512 | 1406.43 \u00b1 52.64; llama 7B Q4_0 | tg128 | 30.23 \u00b1 0.23",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14463481",
   "tg_tokens": 128,
   "tps": 30.23,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-20",
   "flags": [],
   "hardware": "NVIDIA RTX 3500",
   "id": "lc-disc-15013-c14463481-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b06477",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1610.14,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "NVIDIA RTX 3500; llama 7B Q4_0 | pp512 | 1610.14 \u00b1 32.13; llama 7B Q4_0 | tg128 | 28.75 \u00b1 0.21",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14463481",
   "tg_tokens": 128,
   "tps": 28.75,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-20",
   "flags": [],
   "hardware": "NVIDIA RTX 3500",
   "id": "lc-disc-15013-c14463481-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; config=b06477",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3353.26,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "NVIDIA RTX 3500; llama 7B Q4_0 | pp512 | 3353.26 \u00b1 20.38; llama 7B Q4_0 | tg128 | 81.17 \u00b1 3.62",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14463481",
   "tg_tokens": 128,
   "tps": 81.17,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-20",
   "flags": [],
   "hardware": "NVIDIA RTX 3500",
   "id": "lc-disc-15013-c14463481-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06477",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3757.32,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "NVIDIA RTX 3500; llama 7B Q4_0 | pp512 | 3757.32 \u00b1 37.79; llama 7B Q4_0 | tg128 | 80.18 \u00b1 1.34",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14463481",
   "tg_tokens": 128,
   "tps": 80.18,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-22",
   "flags": [],
   "hardware": "V100",
   "id": "lc-disc-15013-c14472498-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b06533",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3042.64,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "V100; llama 7B Q4_0 | pp512 | 3042.64 \u00b1 40.71; llama 7B Q4_0 | tg128 | 129.08 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14472498",
   "tg_tokens": 128,
   "tps": 129.08,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-22",
   "flags": [],
   "hardware": "V100",
   "id": "lc-disc-15013-c14472498-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b06533",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2973.78,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "V100; llama 7B Q4_0 | pp512 | 2973.78 \u00b1 3.62; llama 7B Q4_0 | tg128 | 134.76 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14472498",
   "tg_tokens": 128,
   "tps": 134.76,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-22",
   "flags": [],
   "hardware": "V100",
   "id": "lc-disc-15013-c14472498-gemma312b-Q4_K - Medium",
   "model": "gemma3 12B",
   "notes": "threads=99; config=b06533",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1857.73,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "V100; gemma3 12B Q4_K - Medium | pp512 | 1857.73 \u00b1 4.03; gemma3 12B Q4_K - Medium | tg128 | 67.34 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14472498",
   "tg_tokens": 128,
   "tps": 67.34,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-22",
   "flags": [],
   "hardware": "V100",
   "id": "lc-disc-15013-c14472498-gemma312b-Q4_K - Medium-2",
   "model": "gemma3 12B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06533",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1771.04,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "V100; gemma3 12B Q4_K - Medium | pp512 | 1771.04 \u00b1 1.70; gemma3 12B Q4_K - Medium | tg128 | 68.94 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14472498",
   "tg_tokens": 128,
   "tps": 68.94,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-22",
   "flags": [],
   "hardware": "V100",
   "id": "lc-disc-15013-c14472498-qwen314b-Q4_K - Medium",
   "model": "Qwen3 14B",
   "notes": "threads=99; config=b06533",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1536.27,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "V100; qwen3 14B Q4_K - Medium | pp512 | 1536.27 \u00b1 2.28; qwen3 14B Q4_K - Medium | tg128 | 64.87 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14472498",
   "tg_tokens": 128,
   "tps": 64.87,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-22",
   "flags": [],
   "hardware": "V100",
   "id": "lc-disc-15013-c14472498-qwen314b-Q4_K - Medium-2",
   "model": "Qwen3 14B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06533",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1474.29,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "V100; qwen3 14B Q4_K - Medium | pp512 | 1474.29 \u00b1 2.45; qwen3 14B Q4_K - Medium | tg128 | 67.03 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14472498",
   "tg_tokens": 128,
   "tps": 67.03,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-22",
   "flags": [],
   "hardware": "V100",
   "id": "lc-disc-15013-c14472498-qwen332b-Q4_K - Medium",
   "model": "Qwen3 32B",
   "notes": "threads=99; config=b06533",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 642.41,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "V100; qwen3 32B Q4_K - Medium | pp512 | 642.41 \u00b1 0.47; qwen3 32B Q4_K - Medium | tg128 | 30.74 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14472498",
   "tg_tokens": 128,
   "tps": 30.74,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-22",
   "flags": [],
   "hardware": "V100",
   "id": "lc-disc-15013-c14472498-qwen332b-Q4_K - Medium-2",
   "model": "Qwen3 32B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06533",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 613.13,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "V100; qwen3 32B Q4_K - Medium | pp512 | 613.13 \u00b1 1.09; qwen3 32B Q4_K - Medium | tg128 | 31.80 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14472498",
   "tg_tokens": 128,
   "tps": 31.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-22",
   "flags": [],
   "hardware": "V100",
   "id": "lc-disc-15013-c14472498-qwen270b-Q4_0",
   "model": "qwen2 70B",
   "notes": "threads=99; config=b06533",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 335.74,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "V100; qwen2 70B Q4_0 | pp512 | 335.74 \u00b1 0.45; qwen2 70B Q4_0 | tg128 | 17.22 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14472498",
   "tg_tokens": 128,
   "tps": 17.22,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-22",
   "flags": [],
   "hardware": "V100",
   "id": "lc-disc-15013-c14472498-qwen270b-Q4_0-2",
   "model": "qwen2 70B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06533",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 324.02,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "V100; qwen2 70B Q4_0 | pp512 | 324.02 \u00b1 0.38; qwen2 70B Q4_0 | tg128 | 17.72 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14472498",
   "tg_tokens": 128,
   "tps": 17.72,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-22",
   "flags": [],
   "hardware": "NVIDIA RTX 6000",
   "id": "lc-disc-15013-c14472609-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b06475",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 9229.23,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "NVIDIA RTX 6000; llama 7B Q4_0 | pp512 | 9229.23 \u00b1 101.78; llama 7B Q4_0 | tg128 | 176.07 \u00b1 0.26",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14472609",
   "tg_tokens": 128,
   "tps": 176.07,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-22",
   "flags": [],
   "hardware": "NVIDIA RTX 6000",
   "id": "lc-disc-15013-c14472609-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06475",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 10576.85,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "NVIDIA RTX 6000; llama 7B Q4_0 | pp512 | 10576.85 \u00b1 530.21; llama 7B Q4_0 | tg128 | 179.47 \u00b1 0.32",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14472609",
   "tg_tokens": 128,
   "tps": 179.47,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-26",
   "flags": [],
   "hardware": "GeForce RTX 5090",
   "id": "lc-disc-15013-c14525481-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b06594",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 14142.08,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5090; llama 7B Q4_0 | pp512 | 14142.08 \u00b1 52.87; llama 7B Q4_0 | tg128 | 277.21 \u00b1 0.96",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14525481",
   "tg_tokens": 128,
   "tps": 277.21,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-26",
   "flags": [],
   "hardware": "GeForce RTX 5090",
   "id": "lc-disc-15013-c14525481-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06594",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 16195.14,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5090; llama 7B Q4_0 | pp512 | 16195.14 \u00b1 141.11; llama 7B Q4_0 | tg128 | 286.51 \u00b1 0.13",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14525481",
   "tg_tokens": 128,
   "tps": 286.51,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-30",
   "flags": [],
   "hardware": "NVIDIA RTX 4000",
   "id": "lc-disc-15013-c14553456-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b06638",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2779.77,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "NVIDIA RTX 4000; llama 7B Q4_0 | pp512 | 2779.77 \u00b1 9.91; llama 7B Q4_0 | tg128 | 61.83 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14553456",
   "tg_tokens": 128,
   "tps": 61.83,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-30",
   "flags": [],
   "hardware": "NVIDIA RTX 4000",
   "id": "lc-disc-15013-c14553456-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06638",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3171.86,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "NVIDIA RTX 4000; llama 7B Q4_0 | pp512 | 3171.86 \u00b1 4.34; llama 7B Q4_0 | tg128 | 61.37 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14553456",
   "tg_tokens": 128,
   "tps": 61.37,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-15",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-disc-15013-c14687095-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b06767",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3062.31,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GB10; llama 7B Q4_0 | pp512 | 3062.31 \u00b1 11.02; llama 7B Q4_0 | tg128 | 57.21 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14687095",
   "tg_tokens": 128,
   "tps": 57.21,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-15",
   "flags": [],
   "hardware": "GB10",
   "id": "lc-disc-15013-c14687095-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06767",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3661.37,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GB10; llama 7B Q4_0 | pp512 | 3661.37 \u00b1 38.66; llama 7B Q4_0 | tg128 | 56.74 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14687095",
   "tg_tokens": 128,
   "tps": 56.74,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-17",
   "flags": [],
   "hardware": "GeForce RTX 3070",
   "id": "lc-disc-15013-c14705424-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b06783",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1729.14,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 3070; llama 7B Q4_0 | pp512 | 1729.14 \u00b1 38.71; llama 7B Q4_0 | tg128 | 59.72 \u00b1 0.78",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14705424",
   "tg_tokens": 128,
   "tps": 59.72,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-17",
   "flags": [],
   "hardware": "GeForce RTX 3070",
   "id": "lc-disc-15013-c14705424-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06783",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1904.62,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 3070; llama 7B Q4_0 | pp512 | 1904.62 \u00b1 16.19; llama 7B Q4_0 | tg128 | 57.36 \u00b1 0.98",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14705424",
   "tg_tokens": 128,
   "tps": 57.36,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-17",
   "flags": [],
   "hardware": "GeForce RTX 3070",
   "id": "lc-disc-15013-c14705424-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; config=b06783",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1735.28,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 3070; llama 7B Q4_0 | pp512 | 1735.28 \u00b1 36.61; llama 7B Q4_0 | tg128 | 62.17 \u00b1 0.98",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14705424",
   "tg_tokens": 128,
   "tps": 62.17,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-17",
   "flags": [],
   "hardware": "GeForce RTX 3070",
   "id": "lc-disc-15013-c14705424-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06783",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1918.79,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 3070; llama 7B Q4_0 | pp512 | 1918.79 \u00b1 24.64; llama 7B Q4_0 | tg128 | 59.52 \u00b1 0.34",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14705424",
   "tg_tokens": 128,
   "tps": 59.52,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-18",
   "flags": [],
   "hardware": "GeForce RTX 4080 Super",
   "id": "lc-disc-15013-c14717261-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b06729",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 8125.15,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 4080 Super; dev=CUDA0; llama 7B Q4_0 | pp512 | 8125.15 \u00b1 41.05; llama 7B Q4_0 | tg128 | 148.33 \u00b1 0.20",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14717261",
   "tg_tokens": 128,
   "tps": 148.33,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-18",
   "flags": [],
   "hardware": "GeForce RTX 4080 Super",
   "id": "lc-disc-15013-c14717261-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06729",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 9439.01,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 4080 Super; dev=CUDA0; llama 7B Q4_0 | pp512 | 9439.01 \u00b1 56.75; llama 7B Q4_0 | tg128 | 147.48 \u00b1 1.41",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14717261",
   "tg_tokens": 128,
   "tps": 147.48,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-23",
   "flags": [],
   "hardware": "GeForce RTX 5090",
   "id": "lc-disc-15013-c14761847-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b06824",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 14073.41,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5090; llama 7B Q4_0 | pp512 | 14073.41 \u00b1 115.16; llama 7B Q4_0 | tg128 | 290.02 \u00b1 1.10",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14761847",
   "tg_tokens": 128,
   "tps": 290.02,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-23",
   "flags": [],
   "hardware": "GeForce RTX 5090",
   "id": "lc-disc-15013-c14761847-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06824",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 14970.15,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5090; llama 7B Q4_0 | pp512 | 14970.15 \u00b1 381.06; llama 7B Q4_0 | tg128 | 300.40 \u00b1 0.28",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14761847",
   "tg_tokens": 128,
   "tps": 300.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-07",
   "flags": [],
   "hardware": "GeForce RTX 5090",
   "id": "lc-disc-15013-c14901568-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b06972",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 13951.14,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5090; llama 7B Q4_0 | pp512 | 13951.14 \u00b1 187.85; llama 7B Q4_0 | tg128 | 298.18 \u00b1 0.34",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14901568",
   "tg_tokens": 128,
   "tps": 298.18,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-07",
   "flags": [],
   "hardware": "GeForce RTX 5090",
   "id": "lc-disc-15013-c14901568-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b06972",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 15915.61,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5090; llama 7B Q4_0 | pp512 | 15915.61 \u00b1 274.93; llama 7B Q4_0 | tg128 | 283.69 \u00b1 0.42",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14901568",
   "tg_tokens": 128,
   "tps": 283.69,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-07",
   "flags": [],
   "hardware": "GeForce RTX 5090",
   "id": "lc-disc-15013-c14905268-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=1 ngl=99; config=b06972",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 15094.35,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5090; llama 7B Q4_0 | pp512 | 15094.35 \u00b1 131.34; llama 7B Q4_0 | tg128 | 266.70 \u00b1 1.24",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14905268",
   "tg_tokens": 128,
   "tps": 266.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-07",
   "flags": [],
   "hardware": "GeForce RTX 5090",
   "id": "lc-disc-15013-c14905268-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=0 ngl=99; config=b06972",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 13143.45,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5090; llama 7B Q4_0 | pp512 | 13143.45 \u00b1 88.72; llama 7B Q4_0 | tg128 | 280.28 \u00b1 1.13",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14905268",
   "tg_tokens": 128,
   "tps": 280.28,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-07",
   "flags": [],
   "hardware": "GeForce RTX 5090",
   "id": "lc-disc-15013-c14905345-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b06972",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 16053.65,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5090; llama 7B Q4_0 | pp512 | 16053.65 \u00b1 139.58; llama 7B Q4_0 | tg128 | 284.03 \u00b1 3.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14905345",
   "tg_tokens": 128,
   "tps": 284.03,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-07",
   "flags": [],
   "hardware": "GeForce RTX 5090",
   "id": "lc-disc-15013-c14905345-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06972",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 14013.74,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5090; llama 7B Q4_0 | pp512 | 14013.74 \u00b1 21.71; llama 7B Q4_0 | tg128 | 299.85 \u00b1 0.40",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14905345",
   "tg_tokens": 128,
   "tps": 299.85,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-07",
   "flags": [],
   "hardware": "GeForce RTX 5090",
   "id": "lc-disc-15013-c14905345-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; config=b06972",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 14036.03,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5090; llama 7B Q4_0 | pp512 | 14036.03 \u00b1 65.48; llama 7B Q4_0 | tg128 | 299.92 \u00b1 0.26",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14905345",
   "tg_tokens": 128,
   "tps": 299.92,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-07",
   "flags": [],
   "hardware": "GeForce RTX 5090",
   "id": "lc-disc-15013-c14905345-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06972",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 15602.75,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5090; llama 7B Q4_0 | pp512 | 15602.75 \u00b1 616.25; llama 7B Q4_0 | tg128 | 285.51 \u00b1 0.20",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14905345",
   "tg_tokens": 128,
   "tps": 285.51,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-08",
   "flags": [],
   "hardware": "GeForce RTX 5090",
   "id": "lc-disc-15013-c14911245-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b06989",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 16070.88,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5090; llama 7B Q4_0 | pp512 | 16070.88 \u00b1 187.65; llama 7B Q4_0 | tg128 | 312.33 \u00b1 0.46",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14911245",
   "tg_tokens": 128,
   "tps": 312.33,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-08",
   "flags": [],
   "hardware": "GeForce RTX 5090",
   "id": "lc-disc-15013-c14911245-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06989",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 13964.52,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5090; llama 7B Q4_0 | pp512 | 13964.52 \u00b1 185.20; llama 7B Q4_0 | tg128 | 294.95 \u00b1 0.25",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14911245",
   "tg_tokens": 128,
   "tps": 294.95,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-19",
   "flags": [],
   "hardware": "V100",
   "id": "lc-disc-15013-c15016423-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3235.72,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "V100; llama 7B Q4_0 | pp512 | 3235.72 \u00b1 1.52; llama 7B Q4_0 | pp1024 | 3132.70 \u00b1 6.63; llama 7B Q4_0 | pp2048 | 2962.06 \u00b1 4.09; llama 7B Q4_0 | pp4096 | 2651.09 \u00b1 0.88; llama 7B Q4_0 | pp8192 | 2168.78 \u00b1 0.38; llama 7B Q4_0 | pp16384 | 1540.14 \u00b1 0.65; llama 7B Q4_0 | tg128 | 140.00 \u00b1 0.02; llama 7B Q4_0 | tg256 | 139.84 \u00b1 0.02; llama 7B Q4_0 | tg512 | 134.57 \u00b1 0.01; llama 7B Q4_0 | tg1024 | 134.65 \u00b1 0.08; llama 7B Q4_0 | tg2048 | 130.17 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15016423",
   "tg_tokens": 128,
   "tps": 140.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-19",
   "flags": [],
   "hardware": "V100",
   "id": "lc-disc-15013-c15016423-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3133.71,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "V100; llama 7B Q4_0 | pp512 | 3133.71 \u00b1 1.13; llama 7B Q4_0 | pp1024 | 2948.59 \u00b1 0.38; llama 7B Q4_0 | pp2048 | 2645.42 \u00b1 0.59; llama 7B Q4_0 | pp4096 | 2200.23 \u00b1 0.53; llama 7B Q4_0 | pp8192 | 1645.94 \u00b1 0.31; llama 7B Q4_0 | pp16384 | 1094.26 \u00b1 0.06; llama 7B Q4_0 | tg128 | 151.88 \u00b1 0.02; llama 7B Q4_0 | tg256 | 151.64 \u00b1 0.01; llama 7B Q4_0 | tg512 | 148.94 \u00b1 0.02; llama 7B Q4_0 | tg1024 | 143.44 \u00b1 0.01; llama 7B Q4_0 | tg2048 | 135.61 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15016423",
   "tg_tokens": 128,
   "tps": 151.88,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-19",
   "flags": [],
   "hardware": "V100",
   "id": "lc-disc-15013-c15016423-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3235.01,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Tesla V100-SXM2-16GB, compute capability 7.0, VMM: yes\n  Device; llama 7B Q4_0 | pp512 | 3235.01 \u00b1 0.88; llama 7B Q4_0 | tg128 | 139.89 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15016423",
   "tg_tokens": 128,
   "tps": 139.89,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-19",
   "flags": [],
   "hardware": "V100",
   "id": "lc-disc-15013-c15016423-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3134.81,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Tesla V100-SXM2-16GB, compute capability 7.0, VMM: yes\n  Device; llama 7B Q4_0 | pp512 | 3134.81 \u00b1 0.91; llama 7B Q4_0 | tg128 | 152.36 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15016423",
   "tg_tokens": 128,
   "tps": 152.36,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-19",
   "flags": [],
   "hardware": "V100",
   "id": "lc-disc-15013-c15016423-llama7b-Q4_0-5",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3149.71,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Tesla V100-SXM2-16GB, compute capability 7.0, VMM: yes\n   \n  \n \n \n model \n size; llama 7B Q4_0 | pp512 | 3149.71 \u00b1 12.20; llama 7B Q4_0 | tg128 | 139.11 \u00b1 0.10",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15016423",
   "tg_tokens": 128,
   "tps": 139.11,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-19",
   "flags": [],
   "hardware": "V100",
   "id": "lc-disc-15013-c15016423-llama7b-Q4_0-6",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3075.88,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Tesla V100-SXM2-16GB, compute capability 7.0, VMM: yes\n   \n  \n \n \n model \n size; llama 7B Q4_0 | pp512 | 3075.88 \u00b1 6.83; llama 7B Q4_0 | tg128 | 150.68 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15016423",
   "tg_tokens": 128,
   "tps": 150.68,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-21",
   "flags": [],
   "hardware": "Quadro RTX 4000",
   "id": "lc-disc-15013-c15039665-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b07108",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1536.89,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Quadro RTX 4000; llama 7B Q4_0 | pp512 | 1536.89 \u00b1 0.90; llama 7B Q4_0 | tg128 | 65.62 \u00b1 0.62",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15039665",
   "tg_tokens": 128,
   "tps": 65.62,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-21",
   "flags": [],
   "hardware": "Quadro RTX 4000",
   "id": "lc-disc-15013-c15039665-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07108",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1662.8,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Quadro RTX 4000; llama 7B Q4_0 | pp512 | 1662.80 \u00b1 2.04; llama 7B Q4_0 | tg128 | 67.62 \u00b1 0.67",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15039665",
   "tg_tokens": 128,
   "tps": 67.62,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-21",
   "flags": [],
   "hardware": "RTX PRO 4000 Blackwell",
   "id": "lc-disc-15013-c15039841-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b07108",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4992.83,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX PRO 4000 Blackwell; llama 7B Q4_0 | pp512 | 4992.83 \u00b1 113.52; llama 7B Q4_0 | pp1024 | 4456.87 \u00b1 14.01; llama 7B Q4_0 | pp2048 | 4035.09 \u00b1 7.14; llama 7B Q4_0 | pp4096 | 3390.09 \u00b1 10.19; llama 7B Q4_0 | pp8192 | 2672.00 \u00b1 12.17; llama 7B Q4_0 | pp16384 | 1712.53 \u00b1 8.62; llama 7B Q4_0 | pp32768 | 964.31 \u00b1 3.19; llama 7B Q4_0 | tg128 | 131.66 \u00b1 0.20; llama 7B Q4_0 | tg256 | 130.65 \u00b1 0.36; llama 7B Q4_0 | tg512 | 126.07 \u00b1 0.66; llama 7B Q4_0 | tg1024 | 122.89 \u00b1 0.37; llama 7B Q4_0 | tg2048 | 116.63 \u00b1 0.17",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15039841",
   "tg_tokens": 128,
   "tps": 131.66,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-21",
   "flags": [],
   "hardware": "RTX PRO 4000 Blackwell",
   "id": "lc-disc-15013-c15039841-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07108",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 5674.44,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX PRO 4000 Blackwell; llama 7B Q4_0 | pp512 | 5674.44 \u00b1 139.53; llama 7B Q4_0 | pp1024 | 5387.95 \u00b1 122.72; llama 7B Q4_0 | pp2048 | 5443.67 \u00b1 9.65; llama 7B Q4_0 | pp4096 | 5088.33 \u00b1 10.32; llama 7B Q4_0 | pp8192 | 4316.83 \u00b1 11.68; llama 7B Q4_0 | pp16384 | 3314.15 \u00b1 13.59; llama 7B Q4_0 | pp32768 | 2287.67 \u00b1 6.39; llama 7B Q4_0 | tg128 | 136.38 \u00b1 0.13; llama 7B Q4_0 | tg256 | 135.83 \u00b1 0.22; llama 7B Q4_0 | tg512 | 133.39 \u00b1 0.27; llama 7B Q4_0 | tg1024 | 126.10 \u00b1 0.42; llama 7B Q4_0 | tg2048 | 117.90 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15039841",
   "tg_tokens": 128,
   "tps": 136.38,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-23",
   "flags": [],
   "hardware": "GeForce RTX 3060",
   "id": "lc-disc-15013-c15053921-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b07130",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2118.17,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 3060; llama 7B Q4_0 | pp512 | 2118.17 \u00b1 19.94; llama 7B Q4_0 | tg128 | 75.17 \u00b1 0.15",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15053921",
   "tg_tokens": 128,
   "tps": 75.17,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-23",
   "flags": [],
   "hardware": "GeForce RTX 3060",
   "id": "lc-disc-15013-c15053921-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07130",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2452.67,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 3060; llama 7B Q4_0 | pp512 | 2452.67 \u00b1 22.63; llama 7B Q4_0 | tg128 | 78.19 \u00b1 0.40",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15053921",
   "tg_tokens": 128,
   "tps": 78.19,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-26",
   "flags": [],
   "hardware": "P100",
   "id": "lc-disc-15013-c15088555-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b07146",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 760.8,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "P100; llama 7B Q4_0 | pp512 | 760.80 \u00b1 2.92; llama 7B Q4_0 | pp1024 | 734.28 \u00b1 0.96; llama 7B Q4_0 | pp2048 | 710.52 \u00b1 3.99; llama 7B Q4_0 | pp4096 | 654.64 \u00b1 0.90; llama 7B Q4_0 | pp8192 | 568.03 \u00b1 0.49; llama 7B Q4_0 | pp16384 | 417.02 \u00b1 0.21; llama 7B Q4_0 | tg128 | 58.35 \u00b1 0.00; llama 7B Q4_0 | tg256 | 58.35 \u00b1 0.00; llama 7B Q4_0 | tg512 | 56.34 \u00b1 0.01; llama 7B Q4_0 | tg1024 | 54.06 \u00b1 0.00; llama 7B Q4_0 | tg2048 | 50.35 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15088555",
   "tg_tokens": 128,
   "tps": 58.35,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-26",
   "flags": [],
   "hardware": "P100",
   "id": "lc-disc-15013-c15088555-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07146",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 787.36,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "P100; llama 7B Q4_0 | pp512 | 787.36 \u00b1 3.27; llama 7B Q4_0 | pp1024 | 765.73 \u00b1 2.99; llama 7B Q4_0 | pp2048 | 760.03 \u00b1 2.44; llama 7B Q4_0 | pp4096 | 690.63 \u00b1 0.64; llama 7B Q4_0 | pp8192 | 620.79 \u00b1 0.38; llama 7B Q4_0 | pp16384 | 502.56 \u00b1 0.06; llama 7B Q4_0 | tg128 | 61.99 \u00b1 0.00; llama 7B Q4_0 | tg256 | 62.01 \u00b1 0.00; llama 7B Q4_0 | tg512 | 61.55 \u00b1 0.00; llama 7B Q4_0 | tg1024 | 60.59 \u00b1 0.00; llama 7B Q4_0 | tg2048 | 58.64 \u00b1 0.26",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15088555",
   "tg_tokens": 128,
   "tps": 61.99,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-26",
   "flags": [],
   "hardware": "V100",
   "id": "lc-disc-15013-c15088555-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b07146",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 767.12,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "V100; llama 7B Q4_0 | pp512 | 767.12 \u00b1 1.59; llama 7B Q4_0 | tg128 | 58.33 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15088555",
   "tg_tokens": 128,
   "tps": 58.33,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-26",
   "flags": [],
   "hardware": "V100",
   "id": "lc-disc-15013-c15088555-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b07146",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 788.45,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "V100; llama 7B Q4_0 | pp512 | 788.45 \u00b1 4.06; llama 7B Q4_0 | tg128 | 62.06 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15088555",
   "tg_tokens": 128,
   "tps": 62.06,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-27",
   "flags": [],
   "hardware": "P100",
   "id": "lc-disc-15013-c15095402-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b07169",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 476.77,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "P100; llama 7B Q4_0 | pp512 | 476.77 \u00b1 1.09; llama 7B Q4_0 | tg128 | 50.17 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15095402",
   "tg_tokens": 128,
   "tps": 50.17,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-27",
   "flags": [],
   "hardware": "P100",
   "id": "lc-disc-15013-c15095402-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b07169",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 492.2,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "P100; llama 7B Q4_0 | pp512 | 492.20 \u00b1 0.70; llama 7B Q4_0 | tg128 | 50.66 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15095402",
   "tg_tokens": 128,
   "tps": 50.66,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-27",
   "flags": [],
   "hardware": "P100",
   "id": "lc-disc-15013-c15095402-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b07169",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1006.05,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "P100; llama 7B Q4_0 | pp512 | 1006.05 \u00b1 0.57; llama 7B Q4_0 | tg128 | 51.54 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15095402",
   "tg_tokens": 128,
   "tps": 51.54,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-27",
   "flags": [],
   "hardware": "P100",
   "id": "lc-disc-15013-c15095402-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b07169",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1076.52,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "P100; llama 7B Q4_0 | pp512 | 1076.52 \u00b1 1.13; llama 7B Q4_0 | tg128 | 52.24 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15095402",
   "tg_tokens": 128,
   "tps": 52.24,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-04",
   "flags": [],
   "hardware": "GeForce RTX 5090",
   "id": "lc-disc-15013-c15160999-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b07268",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 14131.49,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5090; llama 7B Q4_0 | pp512 | 14131.49 \u00b1 88.72; llama 7B Q4_0 | tg128 | 297.98 \u00b1 0.21",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15160999",
   "tg_tokens": 128,
   "tps": 297.98,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-04",
   "flags": [],
   "hardware": "GeForce RTX 5090",
   "id": "lc-disc-15013-c15160999-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07268",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 14805.32,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5090; llama 7B Q4_0 | pp512 | 14805.32 \u00b1 468.93; llama 7B Q4_0 | tg128 | 314.00 \u00b1 0.54",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15160999",
   "tg_tokens": 128,
   "tps": 314.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-04",
   "flags": [],
   "hardware": "GeForce RTX 5090",
   "id": "lc-disc-15013-c15160999-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; config=b07268",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 14988.21,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5090; llama 7B Q4_0 | pp512 | 14988.21 \u00b1 144.13; llama 7B Q4_0 | tg128 | 297.87 \u00b1 0.88",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15160999",
   "tg_tokens": 128,
   "tps": 297.87,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-04",
   "flags": [],
   "hardware": "GeForce RTX 5090",
   "id": "lc-disc-15013-c15160999-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07268",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 17107.92,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5090; llama 7B Q4_0 | pp512 | 17107.92 \u00b1 226.89; llama 7B Q4_0 | tg128 | 315.30 \u00b1 0.34",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15160999",
   "tg_tokens": 128,
   "tps": 315.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-07",
   "flags": [],
   "hardware": "GeForce RTX 5070 Ti",
   "id": "lc-disc-15013-c15183261-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b07293",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 6952.38,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5070 Ti; llama 7B Q4_0 | pp512 | 6952.38 \u00b1 13.73; llama 7B Q4_0 | pp1024 | 6180.29 \u00b1 15.42; llama 7B Q4_0 | pp2048 | 5360.57 \u00b1 7.33; llama 7B Q4_0 | pp4096 | 4283.14 \u00b1 2.50; llama 7B Q4_0 | pp8192 | 3058.94 \u00b1 5.50; llama 7B Q4_0 | pp16384 | 1902.90 \u00b1 1.10; llama 7B Q4_0 | tg128 | 176.85 \u00b1 0.07; llama 7B Q4_0 | tg256 | 176.63 \u00b1 0.09; llama 7B Q4_0 | tg512 | 171.28 \u00b1 0.23; llama 7B Q4_0 | tg1024 | 162.42 \u00b1 11.07; llama 7B Q4_0 | tg2048 | 159.21 \u00b1 0.14",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15183261",
   "tg_tokens": 128,
   "tps": 176.85,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-07",
   "flags": [],
   "hardware": "GeForce RTX 5070 Ti",
   "id": "lc-disc-15013-c15183261-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07293",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 8419.56,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5070 Ti; llama 7B Q4_0 | pp512 | 8419.56 \u00b1 35.50; llama 7B Q4_0 | pp1024 | 8329.72 \u00b1 31.84; llama 7B Q4_0 | pp2048 | 8062.84 \u00b1 10.80; llama 7B Q4_0 | pp4096 | 7448.91 \u00b1 3.95; llama 7B Q4_0 | pp8192 | 6257.10 \u00b1 1.91; llama 7B Q4_0 | pp16384 | 4768.77 \u00b1 0.92; llama 7B Q4_0 | tg128 | 182.43 \u00b1 0.09; llama 7B Q4_0 | tg256 | 182.39 \u00b1 0.04; llama 7B Q4_0 | tg512 | 179.20 \u00b1 0.18; llama 7B Q4_0 | tg1024 | 173.58 \u00b1 0.76; llama 7B Q4_0 | tg2048 | 163.71 \u00b1 0.67",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15183261",
   "tg_tokens": 128,
   "tps": 182.43,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-09",
   "flags": [],
   "hardware": "GeForce RTX 5090",
   "id": "lc-disc-15013-c15209171-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b07330",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 13451.73,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5090; llama 7B Q4_0 | pp512 | 13451.73 + 366.50; llama 7B Q4_0 | pp1024 | 13860.22 + 123.86; llama 7B Q4_0 | pp2048 | 12532.71 + 50.11; llama 7B Q4_0 | pp4096 | 10923.40 + 9.74; llama 7B Q4_0 | pp8192 | 8379.23 + 83.53; llama 7B Q4_0 | pp16384 | 5752.52 + 33.85; llama 7B Q4_0 | pp32768 | 3254.13 + 12.67; llama 7B Q4_0 | tg128 | 287.55 + 1.13; llama 7B Q4_0 | tg256 | 287.93 + 0.17; llama 7B Q4_0 | tg512 | 276.98 + 0.25; llama 7B Q4_0 | tg1024 | 264.65 + 23.08; llama 7B Q4_0 | tg2048 | 265.40 + 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15209171",
   "tg_tokens": 128,
   "tps": 287.55,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-09",
   "flags": [],
   "hardware": "GeForce RTX 5090",
   "id": "lc-disc-15013-c15209171-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07330",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 15505.79,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 5090; llama 7B Q4_0 | pp512 | 15505.79 + 3664.49; llama 7B Q4_0 | pp1024 | 16908.90 + 226.38; llama 7B Q4_0 | pp2048 | 16316.27 + 44.05; llama 7B Q4_0 | pp4096 | 15335.63 + 46.98; llama 7B Q4_0 | pp8192 | 13534.84 + 26.00; llama 7B Q4_0 | pp16384 | 10939.17 + 11.40; llama 7B Q4_0 | pp32768 | 7901.57 + 7.32; llama 7B Q4_0 | tg128 | 301.92 + 1.14; llama 7B Q4_0 | tg256 | 302.10 + 0.29; llama 7B Q4_0 | tg512 | 297.87 + 0.10; llama 7B Q4_0 | tg1024 | 291.19 + 0.15; llama 7B Q4_0 | tg2048 | 278.36 + 0.14",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15209171",
   "tg_tokens": 128,
   "tps": 301.92,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "GeForce RTX 3090",
   "id": "lc-disc-15013-c15235019-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b07358",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4425.82,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 3090; llama 7B Q4_0 | pp512 | 4425.82 \u00b1 90.96; llama 7B Q4_0 | tg128 | 153.01 \u00b1 0.48",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15235019",
   "tg_tokens": 128,
   "tps": 153.01,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "GeForce RTX 3090",
   "id": "lc-disc-15013-c15235019-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07358",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 5051.48,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 3090; llama 7B Q4_0 | pp512 | 5051.48 \u00b1 39.51; llama 7B Q4_0 | tg128 | 160.67 \u00b1 0.21",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15235019",
   "tg_tokens": 128,
   "tps": 160.67,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "GeForce RTX 3090",
   "id": "lc-disc-15013-c15235019-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; config=b07358",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 5134.61,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 3090; llama 7B Q4_0 | pp512 | 5134.61 \u00b1 17.22; llama 7B Q4_0 | tg128 | 167.39 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15235019",
   "tg_tokens": 128,
   "tps": 167.39,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "GeForce RTX 3090",
   "id": "lc-disc-15013-c15235019-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07358",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 5874.09,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "GeForce RTX 3090; llama 7B Q4_0 | pp512 | 5874.09 \u00b1 27.18; llama 7B Q4_0 | tg128 | 175.48 \u00b1 0.11",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-15235019",
   "tg_tokens": 128,
   "tps": 175.48,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-25",
   "flags": [],
   "hardware": "B200",
   "id": "lc-disc-15013-c14781873-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b06839",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 10531.52,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "B200; llama 7B Q4_0 | pp512 | 10531.52 \u00b1 11.74; llama 7B Q4_0 | tg128 | 297.68 \u00b1 0.36",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14781873",
   "tg_tokens": 128,
   "tps": 297.68,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-25",
   "flags": [],
   "hardware": "B200",
   "id": "lc-disc-15013-c14781873-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06839",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 11510.56,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "B200; llama 7B Q4_0 | pp512 | 11510.56 \u00b1 28.49; llama 7B Q4_0 | tg128 | 315.85 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14781873",
   "tg_tokens": 128,
   "tps": 315.85,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-14",
   "flags": [],
   "hardware": "RTX PRO 4500 Blackwell",
   "id": "lc-disc-15013-c14964129-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b07054",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 7251.66,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX PRO 4500 Blackwell; llama 7B Q4_0 | pp512 | 7251.66 \u00b1 92.40; llama 7B Q4_0 | pp1024 | 6613.14 \u00b1 16.85; llama 7B Q4_0 | pp2048 | 5948.02 \u00b1 1.74; llama 7B Q4_0 | pp4096 | 4999.28 \u00b1 2.08; llama 7B Q4_0 | pp8192 | 3795.29 \u00b1 3.71; llama 7B Q4_0 | pp16384 | 2499.65 \u00b1 3.33; llama 7B Q4_0 | pp32768 | 1409.56 \u00b1 1.03; llama 7B Q4_0 | tg128 | 168.90 \u00b1 0.20; llama 7B Q4_0 | tg256 | 168.89 \u00b1 0.12; llama 7B Q4_0 | tg512 | 162.83 \u00b1 0.50; llama 7B Q4_0 | tg1024 | 154.06 \u00b1 0.15; llama 7B Q4_0 | tg2048 | 146.59 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14964129",
   "tg_tokens": 128,
   "tps": 168.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-14",
   "flags": [],
   "hardware": "RTX PRO 4500 Blackwell",
   "id": "lc-disc-15013-c14964129-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07054",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 8149.86,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX PRO 4500 Blackwell; llama 7B Q4_0 | pp512 | 8149.86 \u00b1 70.22; llama 7B Q4_0 | pp1024 | 8070.27 \u00b1 7.09; llama 7B Q4_0 | pp2048 | 7883.82 \u00b1 3.22; llama 7B Q4_0 | pp4096 | 7506.30 \u00b1 4.09; llama 7B Q4_0 | pp8192 | 6464.12 \u00b1 1.35; llama 7B Q4_0 | pp16384 | 4882.67 \u00b1 0.22; llama 7B Q4_0 | pp32768 | 3348.21 \u00b1 0.35; llama 7B Q4_0 | tg128 | 175.57 \u00b1 0.04; llama 7B Q4_0 | tg256 | 175.49 \u00b1 0.06; llama 7B Q4_0 | tg512 | 171.91 \u00b1 0.22; llama 7B Q4_0 | tg1024 | 166.47 \u00b1 0.16; llama 7B Q4_0 | tg2048 | 157.23 \u00b1 0.14",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14964129",
   "tg_tokens": 128,
   "tps": 175.57,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-14",
   "flags": [],
   "hardware": "RTX PRO 4500 Blackwell",
   "id": "lc-disc-15013-c14964129-gemma312b-Q4_K - Medium",
   "model": "gemma3 12B",
   "notes": "threads=99; config=b07054",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4075.46,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "RTX PRO 4500 Blackwell; gemma3 12B Q4_K - Medium | pp512 | 4075.46 \u00b1 50.50; gemma3 12B Q4_K - Medium | tg128 | 86.00 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14964129",
   "tg_tokens": 128,
   "tps": 86.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-14",
   "flags": [],
   "hardware": "RTX PRO 4500 Blackwell",
   "id": "lc-disc-15013-c14964129-gemma312b-Q4_K - Medium-2",
   "model": "gemma3 12B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07054",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4364.73,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "RTX PRO 4500 Blackwell; gemma3 12B Q4_K - Medium | pp512 | 4364.73 \u00b1 56.68; gemma3 12B Q4_K - Medium | tg128 | 88.38 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14964129",
   "tg_tokens": 128,
   "tps": 88.38,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-14",
   "flags": [],
   "hardware": "RTX PRO 4500 Blackwell",
   "id": "lc-disc-15013-c14964129-gemma327b-Q4_K - Medium",
   "model": "gemma3 27B",
   "notes": "threads=99; config=b07054",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1874.25,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "RTX PRO 4500 Blackwell; gemma3 27B Q4_K - Medium | pp512 | 1874.25 \u00b1 11.13; gemma3 27B Q4_K - Medium | tg128 | 41.80 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14964129",
   "tg_tokens": 128,
   "tps": 41.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-14",
   "flags": [],
   "hardware": "RTX PRO 4500 Blackwell",
   "id": "lc-disc-15013-c14964129-gemma327b-Q4_K - Medium-2",
   "model": "gemma3 27B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07054",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1999.51,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "RTX PRO 4500 Blackwell; gemma3 27B Q4_K - Medium | pp512 | 1999.51 \u00b1 12.91; gemma3 27B Q4_K - Medium | tg128 | 42.77 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14964129",
   "tg_tokens": 128,
   "tps": 42.77,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-14",
   "flags": [],
   "hardware": "RTX PRO 4500 Blackwell",
   "id": "lc-disc-15013-c14964129-qwen314b-Q4_K - Medium",
   "model": "Qwen3 14B",
   "notes": "threads=99; config=b07054",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3583.82,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "RTX PRO 4500 Blackwell; qwen3 14B Q4_K - Medium | pp512 | 3583.82 \u00b1 32.70; qwen3 14B Q4_K - Medium | tg128 | 79.37 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14964129",
   "tg_tokens": 128,
   "tps": 79.37,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-14",
   "flags": [],
   "hardware": "RTX PRO 4500 Blackwell",
   "id": "lc-disc-15013-c14964129-qwen314b-Q4_K - Medium-2",
   "model": "Qwen3 14B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07054",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3861.85,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "RTX PRO 4500 Blackwell; qwen3 14B Q4_K - Medium | pp512 | 3861.85 \u00b1 39.88; qwen3 14B Q4_K - Medium | tg128 | 82.32 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14964129",
   "tg_tokens": 128,
   "tps": 82.32,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-14",
   "flags": [],
   "hardware": "RTX PRO 4500 Blackwell",
   "id": "lc-disc-15013-c14964129-qwen332b-Q4_K - Medium",
   "model": "Qwen3 32B",
   "notes": "threads=99; config=b07054",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1514.35,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "RTX PRO 4500 Blackwell; qwen3 32B Q4_K - Medium | pp512 | 1514.35 \u00b1 5.90; qwen3 32B Q4_K - Medium | tg128 | 36.38 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14964129",
   "tg_tokens": 128,
   "tps": 36.38,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-14",
   "flags": [],
   "hardware": "RTX PRO 4500 Blackwell",
   "id": "lc-disc-15013-c14964129-qwen332b-Q4_K - Medium-2",
   "model": "Qwen3 32B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07054",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1675.78,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "RTX PRO 4500 Blackwell; qwen3 32B Q4_K - Medium | pp512 | 1675.78 \u00b1 7.49; qwen3 32B Q4_K - Medium | tg128 | 37.93 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14964129",
   "tg_tokens": 128,
   "tps": 37.93,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-14",
   "flags": [],
   "hardware": "RTX PRO 4500 Blackwell",
   "id": "lc-disc-15013-c14964129-gptoss20b-Q4_K - Medium",
   "model": "gpt-oss 20B",
   "notes": "threads=99; config=b07054",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 5812.88,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "RTX PRO 4500 Blackwell; gpt-oss 20B Q4_K - Medium | pp512 | 5812.88 \u00b1 50.48; gpt-oss 20B Q4_K - Medium | tg128 | 233.97 \u00b1 0.18",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14964129",
   "tg_tokens": 128,
   "tps": 233.97,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-14",
   "flags": [],
   "hardware": "RTX PRO 4500 Blackwell",
   "id": "lc-disc-15013-c14964129-gptoss20b-Q4_K - Medium-2",
   "model": "gpt-oss 20B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07054",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 6760.29,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "RTX PRO 4500 Blackwell; gpt-oss 20B Q4_K - Medium | pp512 | 6760.29 \u00b1 81.52; gpt-oss 20B Q4_K - Medium | tg128 | 240.75 \u00b1 0.27",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14964129",
   "tg_tokens": 128,
   "tps": 240.75,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-15",
   "flags": [],
   "hardware": "A100",
   "id": "lc-disc-15013-c14978771-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4967.14,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A100; llama 7B Q4_0 | pp512 | 4967.14 \u00b1 6.60; llama 7B Q4_0 | pp1024 | 4825.37 \u00b1 3.67; llama 7B Q4_0 | pp2048 | 4614.06 \u00b1 2.22; llama 7B Q4_0 | pp4096 | 4267.27 \u00b1 3.34; llama 7B Q4_0 | pp8192 | 3698.20 \u00b1 1.43; llama 7B Q4_0 | pp16384 | 2898.56 \u00b1 0.39; llama 7B Q4_0 | pp32768 | 1919.90 \u00b1 0.15; llama 7B Q4_0 | tg128 | 200.05 \u00b1 0.07; llama 7B Q4_0 | tg256 | 199.80 \u00b1 0.04; llama 7B Q4_0 | tg512 | 191.28 \u00b1 0.53; llama 7B Q4_0 | tg1024 | 191.02 \u00b1 0.21; llama 7B Q4_0 | tg2048 | 186.58 \u00b1 0.13",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14978771",
   "tg_tokens": 128,
   "tps": 200.05,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-15",
   "flags": [],
   "hardware": "A100",
   "id": "lc-disc-15013-c14978771-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 5422.2,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "A100; llama 7B Q4_0 | pp512 | 5422.20 \u00b1 13.26; llama 7B Q4_0 | pp1024 | 5420.45 \u00b1 2.57; llama 7B Q4_0 | pp2048 | 5363.99 \u00b1 3.64; llama 7B Q4_0 | pp4096 | 5231.95 \u00b1 1.57; llama 7B Q4_0 | pp8192 | 4953.44 \u00b1 0.58; llama 7B Q4_0 | pp16384 | 4480.15 \u00b1 1.61; llama 7B Q4_0 | pp32768 | 3749.27 \u00b1 0.51; llama 7B Q4_0 | tg128 | 215.11 \u00b1 0.05; llama 7B Q4_0 | tg256 | 215.23 \u00b1 0.02; llama 7B Q4_0 | tg512 | 210.92 \u00b1 0.51; llama 7B Q4_0 | tg1024 | 206.79 \u00b1 0.24; llama 7B Q4_0 | tg2048 | 199.11 \u00b1 0.12",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14978771",
   "tg_tokens": 128,
   "tps": 215.11,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-15",
   "flags": [],
   "hardware": "H100",
   "id": "lc-disc-15013-c14978771-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 10162.9,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "H100; llama 7B Q4_0 | pp512 | 10162.90 \u00b1 104.35; llama 7B Q4_0 | pp1024 | 10045.98 \u00b1 13.86; llama 7B Q4_0 | pp2048 | 9547.35 \u00b1 5.20; llama 7B Q4_0 | pp4096 | 8691.14 \u00b1 6.41; llama 7B Q4_0 | pp8192 | 7285.19 \u00b1 2.95; llama 7B Q4_0 | pp16384 | 5414.49 \u00b1 1.57; llama 7B Q4_0 | pp32768 | 3505.00 \u00b1 0.48; llama 7B Q4_0 | tg128 | 303.08 \u00b1 0.80; llama 7B Q4_0 | tg256 | 302.57 \u00b1 0.08; llama 7B Q4_0 | tg512 | 288.82 \u00b1 0.62; llama 7B Q4_0 | tg1024 | 289.61 \u00b1 0.37; llama 7B Q4_0 | tg2048 | 284.65 \u00b1 0.18",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14978771",
   "tg_tokens": 128,
   "tps": 303.08,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-15",
   "flags": [],
   "hardware": "H100",
   "id": "lc-disc-15013-c14978771-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 11564.03,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "H100; llama 7B Q4_0 | pp512 | 11564.03 \u00b1 130.42; llama 7B Q4_0 | pp1024 | 11631.87 \u00b1 6.59; llama 7B Q4_0 | pp2048 | 11496.66 \u00b1 90.78; llama 7B Q4_0 | pp4096 | 11111.17 \u00b1 30.72; llama 7B Q4_0 | pp8192 | 10499.58 \u00b1 14.11; llama 7B Q4_0 | pp16384 | 9367.99 \u00b1 6.00; llama 7B Q4_0 | pp32768 | 7499.86 \u00b1 1.27; llama 7B Q4_0 | tg128 | 328.70 \u00b1 1.27; llama 7B Q4_0 | tg256 | 327.80 \u00b1 0.09; llama 7B Q4_0 | tg512 | 321.77 \u00b1 0.78; llama 7B Q4_0 | tg1024 | 315.81 \u00b1 0.45; llama 7B Q4_0 | tg2048 | 303.94 \u00b1 0.25",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14978771",
   "tg_tokens": 128,
   "tps": 328.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-15",
   "flags": [],
   "hardware": "H200",
   "id": "lc-disc-15013-c14978771-llama7b-Q4_0-5",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 10428.65,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "H200; llama 7B Q4_0 | pp512 | 10428.65 \u00b1 12.38; llama 7B Q4_0 | pp1024 | 10152.13 \u00b1 8.13; llama 7B Q4_0 | pp2048 | 9704.39 \u00b1 4.68; llama 7B Q4_0 | pp4096 | 8897.22 \u00b1 8.35; llama 7B Q4_0 | pp8192 | 7629.36 \u00b1 0.54; llama 7B Q4_0 | pp16384 | 5826.12 \u00b1 0.77; llama 7B Q4_0 | pp32768 | 3862.13 \u00b1 0.53; llama 7B Q4_0 | tg128 | 307.79 \u00b1 0.26; llama 7B Q4_0 | tg256 | 308.14 \u00b1 0.03; llama 7B Q4_0 | tg512 | 293.97 \u00b1 0.94; llama 7B Q4_0 | tg1024 | 297.78 \u00b1 0.55; llama 7B Q4_0 | tg2048 | 294.64 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14978771",
   "tg_tokens": 128,
   "tps": 307.79,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-15",
   "flags": [],
   "hardware": "H200",
   "id": "lc-disc-15013-c14978771-llama7b-Q4_0-6",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 11633.96,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "H200; llama 7B Q4_0 | pp512 | 11633.96 \u00b1 14.96; llama 7B Q4_0 | pp1024 | 11687.43 \u00b1 6.29; llama 7B Q4_0 | pp2048 | 11598.09 \u00b1 7.56; llama 7B Q4_0 | pp4096 | 11325.42 \u00b1 7.31; llama 7B Q4_0 | pp8192 | 10735.38 \u00b1 2.95; llama 7B Q4_0 | pp16384 | 9665.42 \u00b1 1.50; llama 7B Q4_0 | pp32768 | 7976.59 \u00b1 2.99; llama 7B Q4_0 | tg128 | 337.60 \u00b1 1.40; llama 7B Q4_0 | tg256 | 334.97 \u00b1 0.22; llama 7B Q4_0 | tg512 | 331.70 \u00b1 1.11; llama 7B Q4_0 | tg1024 | 326.04 \u00b1 0.41; llama 7B Q4_0 | tg2048 | 317.30 \u00b1 0.20",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15013 (performance thread, NVIDIA CUDA)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-14978771",
   "tg_tokens": 128,
   "tps": 337.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-01",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-disc-15021-c13963577-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2151.81,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7800 XT; llama 7B Q4_0 | pp512 | 2151.81 + 17.94; llama 7B Q4_0 | tg128 | 100.94 + 0.10",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-13963577",
   "tg_tokens": 128,
   "tps": 100.94,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-01",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-disc-15021-c13963577-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2304.63,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7800 XT; llama 7B Q4_0 | pp512 | 2304.63 + 2.85; llama 7B Q4_0 | tg128 | 95.99 + 0.21",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-13963577",
   "tg_tokens": 128,
   "tps": 95.99,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-01",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-disc-15021-c13963577-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2145.6,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7800 XT; llama 7B Q4_0 | pp512 | 2145.60 + 23.14; llama 7B Q4_0 | tg128 | 96.89 + 0.22",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-13963577",
   "tg_tokens": 128,
   "tps": 96.89,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-01",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-disc-15021-c13963577-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2063.66,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7800 XT; llama 7B Q4_0 | pp512 | 2063.66 + 2.92; llama 7B Q4_0 | tg128 | 96.03 + 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-13963577",
   "tg_tokens": 128,
   "tps": 96.03,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-08",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-disc-15021-c14048789-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b06104",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2109.38,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7800 XT; llama 7B Q4_0 | pp512 | 2109.38 + 15.79; llama 7B Q4_0 | pp1024 | 1749.56 + 12.69; llama 7B Q4_0 | pp2048 | 1165.15 + 1.02; llama 7B Q4_0 | pp4096 | 997.83 + 0.53; llama 7B Q4_0 | pp8192 | 789.89 + 0.46; llama 7B Q4_0 | pp16384 | 196.02 + 0.96; llama 7B Q4_0 | tg128 | 99.55 + 0.09; llama 7B Q4_0 | tg256 | 98.16 + 0.15; llama 7B Q4_0 | tg512 | 90.29 + 0.07; llama 7B Q4_0 | tg1024 | 80.56 + 0.11; llama 7B Q4_0 | tg2048 | 62.58 + 0.18",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14048789",
   "tg_tokens": 128,
   "tps": 99.55,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-08",
   "flags": [],
   "hardware": "Radeon RX 7800 XT",
   "id": "lc-disc-15021-c14048789-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06104",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2296.72,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7800 XT; llama 7B Q4_0 | pp512 | 2296.72 + 3.40; llama 7B Q4_0 | pp1024 | 2225.68 + 2.84; llama 7B Q4_0 | pp2048 | 2069.86 + 2.06; llama 7B Q4_0 | pp4096 | 1814.41 + 2.23; llama 7B Q4_0 | pp8192 | 1423.62 + 0.94; llama 7B Q4_0 | pp16384 | 992.13 + 0.81; llama 7B Q4_0 | tg128 | 96.80 + 1.30; llama 7B Q4_0 | tg256 | 95.98 + 0.60; llama 7B Q4_0 | tg512 | 95.92 + 0.27; llama 7B Q4_0 | tg1024 | 91.30 + 0.74; llama 7B Q4_0 | tg2048 | 85.44 + 0.29",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14048789",
   "tg_tokens": 128,
   "tps": 96.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-01",
   "flags": [],
   "hardware": "Radeon RX 7900 XT",
   "id": "lc-disc-15021-c13963609-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b06060",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2967.12,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XT; llama 7B Q4_0 | pp512 | 2967.12 \u00b1 31.25; llama 7B Q4_0 | tg128 | 116.00 \u00b1 0.10",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-13963609",
   "tg_tokens": 128,
   "tps": 116.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-01",
   "flags": [],
   "hardware": "Radeon RX 7900 XT",
   "id": "lc-disc-15021-c13963609-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06060",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3163.24,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XT; llama 7B Q4_0 | pp512 | 3163.24 \u00b1 4.06; llama 7B Q4_0 | tg128 | 112.75 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-13963609",
   "tg_tokens": 128,
   "tps": 112.75,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-01",
   "flags": [],
   "hardware": "Radeon RX 7600 XT",
   "id": "lc-disc-15021-c13964669-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b06060",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1099.64,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7600 XT; llama 7B Q4_0 | pp512 | 1099.64 \u00b1 2.05; llama 7B Q4_0 | tg128 | 48.58 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-13964669",
   "tg_tokens": 128,
   "tps": 48.58,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-01",
   "flags": [],
   "hardware": "Radeon RX 7600 XT",
   "id": "lc-disc-15021-c13964669-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06060",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1199.16,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7600 XT; llama 7B Q4_0 | pp512 | 1199.16 \u00b1 1.07; llama 7B Q4_0 | tg128 | 47.65 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-13964669",
   "tg_tokens": 128,
   "tps": 47.65,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-01",
   "flags": [],
   "hardware": "Radeon RX 7600 XT",
   "id": "lc-disc-15021-c13964669-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b06060",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 606.24,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7600 XT; llama 7B Q4_0 | pp512 | 606.24 \u00b1 0.31; llama 7B Q4_0 | tg128 | 52.84 \u00b1 0.44",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-13964669",
   "tg_tokens": 128,
   "tps": 52.84,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-01",
   "flags": [],
   "hardware": "Radeon RX 7600 XT",
   "id": "lc-disc-15021-c13964669-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b06060",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 612.33,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7600 XT; llama 7B Q4_0 | pp512 | 612.33 \u00b1 0.53; llama 7B Q4_0 | tg128 | 53.70 \u00b1 0.33",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-13964669",
   "tg_tokens": 128,
   "tps": 53.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-02",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c13972703-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99 sm=none; config=b06075",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1803.65,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 1803.65 \u00b1 2.54; llama 7B Q4_0 | tg128 | 74.66 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-13972703",
   "tg_tokens": 128,
   "tps": 74.66,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-02",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c13972703-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99 sm=none; config=b06075",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1256.86,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 1256.86 \u00b1 0.55; llama 7B Q4_0 | tg128 | 70.83 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-13972703",
   "tg_tokens": 128,
   "tps": 70.83,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-02",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c13972703-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; config=b06075",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3243.15,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3243.15 \u00b1 10.32; llama 7B Q4_0 | tg128 | 125.84 \u00b1 0.11",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-13972703",
   "tg_tokens": 128,
   "tps": 125.84,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-02",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c13972703-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06075",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3557.68,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3557.68 \u00b1 13.45; llama 7B Q4_0 | tg128 | 122.71 \u00b1 0.11",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-13972703",
   "tg_tokens": 128,
   "tps": 122.71,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-02",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c13972703-llama7b-Q4_0-5",
   "model": "llama 7B",
   "notes": "threads=99; config=b06075",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3369.65,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3369.65 \u00b1 10.61; llama 7B Q4_0 | tg128 | 122.06 \u00b1 0.14",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-13972703",
   "tg_tokens": 128,
   "tps": 122.06,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-02",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c13972703-llama7b-Q4_0-6",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06075",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3573.3,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3573.30 \u00b1 14.31; llama 7B Q4_0 | tg128 | 118.71 \u00b1 0.14",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-13972703",
   "tg_tokens": 128,
   "tps": 118.71,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-17",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c14432902-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b01005",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2817.81,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 2817.81 \u00b1 26.69; llama 7B Q4_0 | tg128 | 112.18 \u00b1 0.45",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14432902",
   "tg_tokens": 128,
   "tps": 112.18,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-17",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c14432902-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b01005",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3053.77,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3053.77 \u00b1 17.38; llama 7B Q4_0 | tg128 | 110.70 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14432902",
   "tg_tokens": 128,
   "tps": 110.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-06",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c14019297-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b06099",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3434.01,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3434.01 \u00b1 38.33; llama 7B Q4_0 | tg128 | 153.91 \u00b1 0.18",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14019297",
   "tg_tokens": 128,
   "tps": 153.91,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-06",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c14019297-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b06099",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3633.86,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3633.86 \u00b1 10.29; llama 7B Q4_0 | tg128 | 145.23 \u00b1 0.10",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14019297",
   "tg_tokens": 128,
   "tps": 145.23,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-08",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c14045285-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b06118",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3386.75,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3386.75 \u00b1 5.33; llama 7B Q4_0 | tg128 | 128.25 \u00b1 0.24",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14045285",
   "tg_tokens": 128,
   "tps": 128.25,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-08",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c14045285-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b06118",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3674.25,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3674.25 \u00b1 11.35; llama 7B Q4_0 | tg128 | 124.61 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14045285",
   "tg_tokens": 128,
   "tps": 124.61,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-11",
   "flags": [],
   "hardware": "Radeon RX 6800",
   "id": "lc-disc-15021-c14073230-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b06123",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1447.07,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6800; llama 7B Q4_0 | pp512 | 1447.07 \u00b1 1.36; llama 7B Q4_0 | tg128 | 83.92 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14073230",
   "tg_tokens": 128,
   "tps": 83.92,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-04",
   "flags": [],
   "hardware": "Radeon RX 6800",
   "id": "lc-disc-15021-c17180754-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b06123",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1382.74,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6800; llama 7B Q4_0 | pp512 | 1382.74 \u00b1 21.98; llama 7B Q4_0 | tg128 | 79.78 \u00b1 0.12",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-17180754",
   "tg_tokens": 128,
   "tps": 79.78,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-04",
   "flags": [],
   "hardware": "Radeon RX 6800",
   "id": "lc-disc-15021-c17180754-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06123",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1675.31,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6800; llama 7B Q4_0 | pp512 | 1675.31 \u00b1 13.21; llama 7B Q4_0 | tg128 | 86.91 \u00b1 0.25",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-17180754",
   "tg_tokens": 128,
   "tps": 86.91,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-13",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c14091623-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b06146",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3473.24,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3473.24 \u00b1 12.30; llama 7B Q4_0 | tg128 | 132.17 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14091623",
   "tg_tokens": 128,
   "tps": 132.17,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-13",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c14091623-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b06146",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3698.73,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3698.73 \u00b1 17.60; llama 7B Q4_0 | tg128 | 127.43 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14091623",
   "tg_tokens": 128,
   "tps": 127.43,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-15",
   "flags": [],
   "hardware": "Radeon RX 6900 XT",
   "id": "lc-disc-15021-c14113551-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99 sm=none",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1824.47,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6900 XT; llama 7B Q4_0 | pp512 | 1824.47 \u00b1 1.02; llama 7B Q4_0 | tg128 | 83.02 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14113551",
   "tg_tokens": 128,
   "tps": 83.02,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-15",
   "flags": [],
   "hardware": "Radeon RX 6900 XT",
   "id": "lc-disc-15021-c14113551-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1250.68,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6900 XT; llama 7B Q4_0 | pp512 | 1250.68 \u00b1 0.73; llama 7B Q4_0 | tg128 | 80.45 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14113551",
   "tg_tokens": 128,
   "tps": 80.45,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-17",
   "flags": [],
   "hardware": "Radeon RX 9070",
   "id": "lc-disc-15021-c14130030-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b06183",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2361.1,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070; llama 7B Q4_0 | pp512 | 2361.10 \u00b1 0.88; llama 7B Q4_0 | tg128 | 99.39 \u00b1 0.57",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14130030",
   "tg_tokens": 128,
   "tps": 99.39,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-17",
   "flags": [],
   "hardware": "Radeon RX 9070",
   "id": "lc-disc-15021-c14130030-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06183",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1147.66,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070; llama 7B Q4_0 | pp512 | 1147.66 \u00b1 1.06; llama 7B Q4_0 | tg128 | 97.10 \u00b1 0.33",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14130030",
   "tg_tokens": 128,
   "tps": 97.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-05",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-15021-c17192251-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4733.95,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 4733.95 \u00b1 129.90; llama 7B Q4_0 | tg128 | 116.95 \u00b1 0.85",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-17192251",
   "tg_tokens": 128,
   "tps": 116.95,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-05",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-15021-c17192251-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 5541.12,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 5541.12 \u00b1 60.59; llama 7B Q4_0 | tg128 | 128.38 \u00b1 0.28",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-17192251",
   "tg_tokens": 128,
   "tps": 128.38,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-05",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-15021-c17192251-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 5673.65,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 5673.65 \u00b1 64.11; llama 7B Q4_0 | tg128 | 136.64 \u00b1 0.61",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-17192251",
   "tg_tokens": 128,
   "tps": 136.64,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-05",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-15021-c17192251-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 6004.06,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 6004.06 \u00b1 24.44; llama 7B Q4_0 | tg128 | 142.88 \u00b1 0.26",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-17192251",
   "tg_tokens": 128,
   "tps": 142.88,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-07-14",
   "flags": [],
   "hardware": "Radeon RX 7600",
   "id": "lc-disc-15021-c17632509-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1006.77,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7600; llama 7B Q4_0 | pp512 | 1006.77 \u00b1 38.87; llama 7B Q4_0 | tg128 | 53.27 \u00b1 0.75",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-17632509",
   "tg_tokens": 128,
   "tps": 53.27,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-07-14",
   "flags": [],
   "hardware": "Radeon RX 7600",
   "id": "lc-disc-15021-c17632509-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1308.31,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7600; llama 7B Q4_0 | pp512 | 1308.31 \u00b1 16.20; llama 7B Q4_0 | tg128 | 58.73 \u00b1 0.15",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-17632509",
   "tg_tokens": 128,
   "tps": 58.73,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": null,
   "ctx": null,
   "date": "2026-07-14",
   "flags": [],
   "hardware": "Radeon RX 7600",
   "id": "lc-disc-15021-c17632509-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=6",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 44.12,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7600; llama 7B Q4_0 | pp512 | 44.12 \u00b1 3.71; llama 7B Q4_0 | tg128 | 9.72 \u00b1 0.13",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-17632509",
   "tg_tokens": 128,
   "tps": 9.72,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CPU)",
   "batch": null,
   "ctx": null,
   "date": "2026-07-14",
   "flags": [],
   "hardware": "Radeon RX 7600",
   "id": "lc-disc-15021-c17632509-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=6; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 42.94,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7600; llama 7B Q4_0 | pp512 | 42.94 \u00b1 1.90; llama 7B Q4_0 | tg128 | 9.72 \u00b1 0.14",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-17632509",
   "tg_tokens": 128,
   "tps": 9.72,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-11",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-15021-c17979194-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b10376",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1063.6,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; llama 7B Q4_0 | pp512 | 1063.60 \u00b1 32.40; llama 7B Q4_0 | tg128 | 43.79 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-17979194",
   "tg_tokens": 128,
   "tps": 43.79,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-11",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-15021-c17979194-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b10376",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1143.51,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; llama 7B Q4_0 | pp512 | 1143.51 \u00b1 30.27; llama 7B Q4_0 | tg128 | 49.10 \u00b1 0.23",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-17979194",
   "tg_tokens": 128,
   "tps": 49.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-23",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c18122692-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b10590",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3875.83,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; dev=ROCM1; llama 7B Q4_0 | pp512 | 3875.83 \u00b1 206.27; llama 7B Q4_0 | tg128 | 147.06 \u00b1 0.37",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-18122692",
   "tg_tokens": 128,
   "tps": 147.06,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-23",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c18122692-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10590",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4190.07,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; dev=ROCM1; llama 7B Q4_0 | pp512 | 4190.07 \u00b1 147.38; llama 7B Q4_0 | tg128 | 153.88 \u00b1 0.29",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-18122692",
   "tg_tokens": 128,
   "tps": 153.88,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-23",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c18122692-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; config=b10590",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3537.94,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3537.94 \u00b1 40.60; llama 7B Q4_0 | tg128 | 136.05 \u00b1 0.21",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-18122692",
   "tg_tokens": 128,
   "tps": 136.05,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-23",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c18122692-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10590",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3764.9,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3764.90 \u00b1 32.17; llama 7B Q4_0 | tg128 | 142.17 \u00b1 0.16",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-18122692",
   "tg_tokens": 128,
   "tps": 142.17,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-23",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c18122692-llama7b-Q4_0-5",
   "model": "llama 7B",
   "notes": "threads=99; config=b10590",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3316.94,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3316.94 \u00b1 5.90; llama 7B Q4_0 | tg128 | 132.75 \u00b1 1.11",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-18122692",
   "tg_tokens": 128,
   "tps": 132.75,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-23",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c18122692-gemma431b-Q8_0",
   "model": "Gemma4 31B",
   "notes": "threads=99; fa=layer ngl=99; config=b10590",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 920.55,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Radeon RX 7900 XTX; gemma4 31B Q8_0 | pp512 | 920.55 \u00b1 2.70; gemma4 31B Q8_0 | tg128 | 22.97 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-18122692",
   "tg_tokens": 128,
   "tps": 22.97,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-23",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c18122692-gemma431b-Q8_0-2",
   "model": "Gemma4 31B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10590",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 960.35,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Radeon RX 7900 XTX; gemma4 31B Q8_0 | pp512 | 960.35 \u00b1 5.27; gemma4 31B Q8_0 | tg128 | 23.43 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-18122692",
   "tg_tokens": 128,
   "tps": 23.43,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-23",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c18122692-gemma431b-Q8_0-3",
   "model": "Gemma4 31B",
   "notes": "threads=99; config=b10590",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1114.17,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Radeon RX 7900 XTX; gemma4 31B Q8_0 | pp512 | 1114.17 \u00b1 0.82; gemma4 31B Q8_0 | tg128 | 32.35 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-18122692",
   "tg_tokens": 128,
   "tps": 32.35,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-10",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c18388777-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b10868",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4145.55,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 4145.55 \u00b1 90.53; llama 7B Q4_0 | tg128 | 155.69 \u00b1 1.68",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-18388777",
   "tg_tokens": 128,
   "tps": 155.69,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-10",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c18388777-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10868",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4403.4,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 4403.40 \u00b1 27.96; llama 7B Q4_0 | tg128 | 163.77 \u00b1 0.56",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-18388777",
   "tg_tokens": 128,
   "tps": 163.77,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-24",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-15021-c14203563-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2908.66,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 2908.66 \u00b1 8.00; llama 7B Q4_0 | tg128 | 83.63 \u00b1 0.36",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14203563",
   "tg_tokens": 128,
   "tps": 83.63,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-24",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-15021-c14203563-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1341.03,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 1341.03 \u00b1 0.28; llama 7B Q4_0 | tg128 | 85.43 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14203563",
   "tg_tokens": 128,
   "tps": 85.43,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-24",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-15021-c14203563-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2055.15,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 2055.15 \u00b1 11.24; llama 7B Q4_0 | tg128 | 128.80 \u00b1 0.22",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14203563",
   "tg_tokens": 128,
   "tps": 128.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-08-24",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-15021-c14203563-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1964.92,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 1964.92 \u00b1 2.13; llama 7B Q4_0 | tg128 | 130.12 \u00b1 0.24",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14203563",
   "tg_tokens": 128,
   "tps": 130.12,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-08",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-15021-c14908936-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3750.46,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 3750.46 \u00b1 544.36; llama 7B Q4_0 | tg128 | 86.16 \u00b1 0.12",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14908936",
   "tg_tokens": 128,
   "tps": 86.16,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-08",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-15021-c14908936-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4543.73,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 4543.73 \u00b1 47.67; llama 7B Q4_0 | tg128 | 89.51 \u00b1 0.20",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14908936",
   "tg_tokens": 128,
   "tps": 89.51,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-03",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-15021-c14300982-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3538.39,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 3538.39 \u00b1 513.67; llama 7B Q4_0 | tg128 | 87.43 \u00b1 0.30",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14300982",
   "tg_tokens": 128,
   "tps": 87.43,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-03",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-15021-c14300982-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4131.22,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 4131.22 \u00b1 6.50; llama 7B Q4_0 | tg128 | 87.57 \u00b1 0.99",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14300982",
   "tg_tokens": 128,
   "tps": 87.57,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-05",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-15021-c16016014-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b08211",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3845.97,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 3845.97 \u00b1 870.69; llama 7B Q4_0 | tg128 | 94.32 \u00b1 0.25",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-16016014",
   "tg_tokens": 128,
   "tps": 94.32,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-05",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-15021-c16016014-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b08211",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 5025.74,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 5025.74 \u00b1 12.64; llama 7B Q4_0 | tg128 | 97.27 \u00b1 0.69",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-16016014",
   "tg_tokens": 128,
   "tps": 97.27,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-06",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-15021-c16016175-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b08211",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3840.62,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 3840.62 \u00b1 864.42; llama 7B Q4_0 | tg128 | 98.11 \u00b1 0.38",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-16016175",
   "tg_tokens": 128,
   "tps": 98.11,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-06",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-15021-c16016175-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b08211",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 5009.54,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 5009.54 \u00b1 25.99; llama 7B Q4_0 | tg128 | 102.44 \u00b1 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-16016175",
   "tg_tokens": 128,
   "tps": 102.44,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-04-09",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-15021-c16498281-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b08720",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4306.74,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 4306.74 \u00b1 507.47; llama 7B Q4_0 | tg128 | 121.04 \u00b1 0.81",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-16498281",
   "tg_tokens": 128,
   "tps": 121.04,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-04-09",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-15021-c16498281-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b08720",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 5380.18,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 5380.18 \u00b1 29.55; llama 7B Q4_0 | tg128 | 127.26 \u00b1 0.34",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-16498281",
   "tg_tokens": 128,
   "tps": 127.26,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-09",
   "flags": [],
   "hardware": "Radeon RX 6900 XT",
   "id": "lc-disc-15021-c14348419-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b06428",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1889.84,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6900 XT; llama 7B Q4_0 | pp512 | 1889.84 \u00b1 31.21; llama 7B Q4_0 | tg128 | 88.49 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14348419",
   "tg_tokens": 128,
   "tps": 88.49,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-09",
   "flags": [],
   "hardware": "Radeon RX 6900 XT",
   "id": "lc-disc-15021-c14348419-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06428",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1948.31,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6900 XT; llama 7B Q4_0 | pp512 | 1948.31 \u00b1 13.51; llama 7B Q4_0 | tg128 | 85.04 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14348419",
   "tg_tokens": 128,
   "tps": 85.04,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-09",
   "flags": [],
   "hardware": "Radeon RX 6900 XT",
   "id": "lc-disc-15021-c14348419-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; config=b06428",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1901.2,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6900 XT; llama 7B Q4_0 | pp512 | 1901.20 \u00b1 36.70; llama 7B Q4_0 | tg128 | 108.00 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14348419",
   "tg_tokens": 128,
   "tps": 108.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-09",
   "flags": [],
   "hardware": "Radeon RX 6900 XT",
   "id": "lc-disc-15021-c14348419-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06428",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1761.93,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 6900 XT; llama 7B Q4_0 | pp512 | 1761.93 \u00b1 4.75; llama 7B Q4_0 | tg128 | 106.15 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14348419",
   "tg_tokens": 128,
   "tps": 106.15,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-14",
   "flags": [],
   "hardware": "Radeon RX 9060 XT",
   "id": "lc-disc-15021-c14397076-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b06470",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1419.67,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9060 XT; llama 7B Q4_0 | pp512 | 1419.67 \u00b1 3.64; llama 7B Q4_0 | tg128 | 67.58 \u00b1 0.24",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14397076",
   "tg_tokens": 128,
   "tps": 67.58,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-14",
   "flags": [],
   "hardware": "Radeon RX 9060 XT",
   "id": "lc-disc-15021-c14397076-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06470",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1479.27,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9060 XT; llama 7B Q4_0 | pp512 | 1479.27 \u00b1 0.71; llama 7B Q4_0 | tg128 | 65.42 \u00b1 0.19",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14397076",
   "tg_tokens": 128,
   "tps": 65.42,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-14",
   "flags": [],
   "hardware": "Radeon RX 9060 XT",
   "id": "lc-disc-15021-c14397076-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b06470",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2056.43,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9060 XT; llama 7B Q4_0 | pp512 | 2056.43 \u00b1 2.23; llama 7B Q4_0 | tg128 | 71.95 \u00b1 1.32",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14397076",
   "tg_tokens": 128,
   "tps": 71.95,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-14",
   "flags": [],
   "hardware": "Radeon RX 9060 XT",
   "id": "lc-disc-15021-c14397076-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b06470",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1840.64,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9060 XT; llama 7B Q4_0 | pp512 | 1840.64 \u00b1 3.76; llama 7B Q4_0 | tg128 | 72.24 \u00b1 0.13",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14397076",
   "tg_tokens": 128,
   "tps": 72.24,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-23",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-15021-c14482499-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b06552",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 351.36,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; llama 7B Q4_0 | pp512 | 351.36 \u00b1 0.67; llama 7B Q4_0 | tg128 | 47.97 \u00b1 0.33",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14482499",
   "tg_tokens": 128,
   "tps": 47.97,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-23",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-15021-c14482499-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b06552",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 366.08,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; llama 7B Q4_0 | pp512 | 366.08 \u00b1 1.44; llama 7B Q4_0 | tg128 | 48.97 \u00b1 0.15",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14482499",
   "tg_tokens": 128,
   "tps": 48.97,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-23",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-15021-c14490924-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=1 ngl=99; config=b06560",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4065.43,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 4065.43 \u00b1 5.08; llama 7B Q4_0 | tg128 | 89.90 \u00b1 0.22",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14490924",
   "tg_tokens": 128,
   "tps": 89.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-09-23",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-15021-c14490924-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=0 ngl=99; config=b06560",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4027.14,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 4027.14 \u00b1 4.65; llama 7B Q4_0 | tg128 | 89.17 \u00b1 0.17",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14490924",
   "tg_tokens": 128,
   "tps": 89.17,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-05",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-15021-c14599306-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=1 ngl=99; config=b06692",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4055.36,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 4055.36 \u00b1 11.83; llama 7B Q4_0 | tg128 | 93.26 \u00b1 0.22",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14599306",
   "tg_tokens": 128,
   "tps": 93.26,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-05",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-15021-c14599306-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=0 ngl=99; config=b06692",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4029.45,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 4029.45 \u00b1 10.03; llama 7B Q4_0 | tg128 | 88.96 \u00b1 0.10",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14599306",
   "tg_tokens": 128,
   "tps": 88.96,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-15",
   "flags": [],
   "hardware": "Radeon RX 7900 GRE",
   "id": "lc-disc-15021-c14690824-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1456.98,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 GRE; llama 7B Q4_0 | pp512 | 1456.98 \u00b1 12.39; llama 7B Q4_0 | tg128 | 96.07 \u00b1 0.10",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14690824",
   "tg_tokens": 128,
   "tps": 96.07,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-15",
   "flags": [],
   "hardware": "Radeon RX 7900 GRE",
   "id": "lc-disc-15021-c14690824-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1598.79,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 GRE; llama 7B Q4_0 | pp512 | 1598.79 \u00b1 11.48; llama 7B Q4_0 | tg128 | 97.53 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14690824",
   "tg_tokens": 128,
   "tps": 97.53,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-15",
   "flags": [],
   "hardware": "Radeon RX 7900 GRE",
   "id": "lc-disc-15021-c14690824-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1604.81,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 GRE; llama 7B Q4_0 | pp512 | 1604.81 \u00b1 7.74; llama 7B Q4_0 | tg128 | 111.31 \u00b1 1.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14690824",
   "tg_tokens": 128,
   "tps": 111.31,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-15",
   "flags": [],
   "hardware": "Radeon RX 7900 GRE",
   "id": "lc-disc-15021-c14690824-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1632.95,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 GRE; llama 7B Q4_0 | pp512 | 1632.95 \u00b1 6.15; llama 7B Q4_0 | tg128 | 112.59 \u00b1 0.45",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14690824",
   "tg_tokens": 128,
   "tps": 112.59,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-18",
   "flags": [],
   "hardware": "Radeon RX 7900 GRE",
   "id": "lc-disc-15021-c16191225-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b08400",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2708.46,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 GRE; llama 7B Q4_0 | pp512 | 2708.46 \u00b1 27.72; llama 7B Q4_0 | tg128 | 96.53 \u00b1 0.43",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-16191225",
   "tg_tokens": 128,
   "tps": 96.53,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-18",
   "flags": [],
   "hardware": "Radeon RX 7900 GRE",
   "id": "lc-disc-15021-c16191225-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b08400",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2982.56,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 GRE; llama 7B Q4_0 | pp512 | 2982.56 \u00b1 16.99; llama 7B Q4_0 | tg128 | 102.38 \u00b1 0.33",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-16191225",
   "tg_tokens": 128,
   "tps": 102.38,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-18",
   "flags": [],
   "hardware": "Radeon RX 7900 GRE",
   "id": "lc-disc-15021-c16191225-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; config=b08400",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2504.25,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 GRE; llama 7B Q4_0 | pp512 | 2504.25 \u00b1 15.36; llama 7B Q4_0 | tg128 | 110.91 \u00b1 0.14",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-16191225",
   "tg_tokens": 128,
   "tps": 110.91,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-18",
   "flags": [],
   "hardware": "Radeon RX 7900 GRE",
   "id": "lc-disc-15021-c16191225-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b08400",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2711.43,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 GRE; llama 7B Q4_0 | pp512 | 2711.43 \u00b1 4.44; llama 7B Q4_0 | tg128 | 116.70 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-16191225",
   "tg_tokens": 128,
   "tps": 116.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-23",
   "flags": [],
   "hardware": "Radeon RX 9070",
   "id": "lc-disc-15021-c14765174-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b06827",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3901.21,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070; llama 7B Q4_0 | pp512 | 3901.21 \u00b1 44.65; llama 7B Q4_0 | tg128 | 115.11 \u00b1 0.51",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14765174",
   "tg_tokens": 128,
   "tps": 115.11,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-23",
   "flags": [],
   "hardware": "Radeon RX 9070",
   "id": "lc-disc-15021-c14765174-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06827",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4074.98,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070; llama 7B Q4_0 | pp512 | 4074.98 \u00b1 3.27; llama 7B Q4_0 | tg128 | 115.84 \u00b1 0.71",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14765174",
   "tg_tokens": 128,
   "tps": 115.84,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-23",
   "flags": [],
   "hardware": "Radeon RX 9070",
   "id": "lc-disc-15021-c14765174-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; config=b06827",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2381.77,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070; llama 7B Q4_0 | pp512 | 2381.77 \u00b1 3.68; llama 7B Q4_0 | tg128 | 114.48 \u00b1 0.60",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14765174",
   "tg_tokens": 128,
   "tps": 114.48,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-23",
   "flags": [],
   "hardware": "Radeon RX 9070",
   "id": "lc-disc-15021-c14765174-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06827",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2452.68,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070; llama 7B Q4_0 | pp512 | 2452.68 \u00b1 1.33; llama 7B Q4_0 | tg128 | 115.32 \u00b1 0.52",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14765174",
   "tg_tokens": 128,
   "tps": 115.32,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-28",
   "flags": [],
   "hardware": "Radeon 890M",
   "id": "lc-disc-15021-c14807647-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b06854",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 163.25,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 890M; llama 7B Q4_0 | pp512 | 163.25 \u00b1 13.86; llama 7B Q4_0 | tg128 | 12.97 \u00b1 1.63",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14807647",
   "tg_tokens": 128,
   "tps": 12.97,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-28",
   "flags": [],
   "hardware": "Radeon 890M",
   "id": "lc-disc-15021-c14807647-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06854",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 213.31,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 890M; llama 7B Q4_0 | pp512 | 213.31 \u00b1 14.05; llama 7B Q4_0 | tg128 | 16.16 \u00b1 1.41",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14807647",
   "tg_tokens": 128,
   "tps": 16.16,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-28",
   "flags": [],
   "hardware": "Radeon 880M",
   "id": "lc-disc-15021-c14807647-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; config=b06854",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2632.74,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 880M; llama 7B Q4_0 | pp512 | 2632.74 \u00b1 17.12; llama 7B Q4_0 | tg128 | 132.12 \u00b1 0.23",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14807647",
   "tg_tokens": 128,
   "tps": 132.12,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-28",
   "flags": [],
   "hardware": "Radeon 880M",
   "id": "lc-disc-15021-c14807647-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06854",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2727.83,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 880M; llama 7B Q4_0 | pp512 | 2727.83 \u00b1 1.38; llama 7B Q4_0 | tg128 | 132.63 \u00b1 0.24",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14807647",
   "tg_tokens": 128,
   "tps": 132.63,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-28",
   "flags": [],
   "hardware": "Radeon 880M",
   "id": "lc-disc-15021-c14807647-gptoss120bmxfp4moe-unknown",
   "model": "gpt-oss 120B MXFP4 MoE",
   "notes": "threads=99; config=b06854",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1180.28,
   "provenance": "community",
   "quant": "as-published",
   "quote": "Radeon 880M; gpt-oss 120B MXFP4 MoE | pp512 | 1180.28 \u00b1 16.68; gpt-oss 120B MXFP4 MoE | tg128 | 97.09 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14807647",
   "tg_tokens": 128,
   "tps": 97.09,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-10-28",
   "flags": [],
   "hardware": "Radeon 880M",
   "id": "lc-disc-15021-c14807647-gptoss120bmxfp4moe-unknown-2",
   "model": "gpt-oss 120B MXFP4 MoE",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06854",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1223.08,
   "provenance": "community",
   "quant": "as-published",
   "quote": "Radeon 880M; gpt-oss 120B MXFP4 MoE | pp512 | 1223.08 \u00b1 16.13; gpt-oss 120B MXFP4 MoE | tg128 | 110.00 \u00b1 0.14",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14807647",
   "tg_tokens": 128,
   "tps": 110.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-22",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-15021-c15320008-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b07501",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4944.22,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 4944.22 \u00b1 40.34; llama 7B Q4_0 | tg128 | 97.21 \u00b1 0.25",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-15320008",
   "tg_tokens": 128,
   "tps": 97.21,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-22",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-15021-c15320008-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07501",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 5077.32,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 5077.32 \u00b1 14.66; llama 7B Q4_0 | tg128 | 101.61 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-15320008",
   "tg_tokens": 128,
   "tps": 101.61,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-02-16",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c15825763-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1253.3,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; dev=ROCM0; llama 7B Q4_0 | pp512 | 1253.30 \u00b1 76.39; llama 7B Q4_0 | tg128 | 129.33 \u00b1 0.10",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-15825763",
   "tg_tokens": 128,
   "tps": 129.33,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-02-16",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c15825763-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1155.75,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; dev=ROCM0; llama 7B Q4_0 | pp512 | 1155.75 \u00b1 1.66; llama 7B Q4_0 | tg128 | 134.46 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-15825763",
   "tg_tokens": 128,
   "tps": 134.46,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-02-16",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c15825763-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3292.37,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; dev=VULKAN0; llama 7B Q4_0 | pp512 | 3292.37 \u00b1 27.90; llama 7B Q4_0 | tg128 | 154.32 \u00b1 0.24",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-15825763",
   "tg_tokens": 128,
   "tps": 154.32,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-02-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c15825763-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3453.79,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; dev=VULKAN0; llama 7B Q4_0 | pp512 | 3453.79 \u00b1 47.32; llama 7B Q4_0 | tg128 | 161.46 \u00b1 0.27",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-15825763",
   "tg_tokens": 128,
   "tps": 161.46,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-01",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c15958764-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100; config=b08182",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1313.1,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 1313.10 \u00b1 3.53; llama 7B Q4_0 | tg128 | 129.69 \u00b1 0.13",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-15958764",
   "tg_tokens": 128,
   "tps": 129.69,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-01",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c15958764-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b08182",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1151.72,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 1151.72 \u00b1 1.08; llama 7B Q4_0 | tg128 | 134.87 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-15958764",
   "tg_tokens": 128,
   "tps": 134.87,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-01",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c15958764-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; config=b08182",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1520.24,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 1520.24 \u00b1 31.48; llama 7B Q4_0 | tg128 | 92.28 \u00b1 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-15958764",
   "tg_tokens": 128,
   "tps": 92.28,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-01",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c15958764-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b08182",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1601.25,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 1601.25 \u00b1 5.70; llama 7B Q4_0 | tg128 | 99.96 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-15958764",
   "tg_tokens": 128,
   "tps": 99.96,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-01",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c15958764-llama7b-Q4_0-5",
   "model": "llama 7B",
   "notes": "threads=99; config=b08182",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4932.17,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 4932.17 \u00b1 67.62; llama 7B Q4_0 | tg128 | 95.04 \u00b1 0.33",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-15958764",
   "tg_tokens": 128,
   "tps": 95.04,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-01",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c15958764-llama7b-Q4_0-6",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b08182",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 5025.17,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 5025.17 \u00b1 19.49; llama 7B Q4_0 | tg128 | 98.71 \u00b1 0.10",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-15958764",
   "tg_tokens": 128,
   "tps": 98.71,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-05",
   "flags": [],
   "hardware": "Radeon 890M",
   "id": "lc-disc-15021-c16010668-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100; config=b08202",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 121.84,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 890M; llama 7B Q4_0 | pp512 | 121.84 \u00b1 0.05; llama 7B Q4_0 | tg128 | 18.24 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-16010668",
   "tg_tokens": 128,
   "tps": 18.24,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-05",
   "flags": [],
   "hardware": "Radeon 890M",
   "id": "lc-disc-15021-c16010668-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100; config=b08202",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 110.39,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 890M; llama 7B Q4_0 | pp512 | 110.39 \u00b1 0.08; llama 7B Q4_0 | tg128 | 19.08 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-16010668",
   "tg_tokens": 128,
   "tps": 19.08,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-18",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c16191303-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b08400",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3623.53,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3623.53 \u00b1 22.80; llama 7B Q4_0 | tg128 | 139.89 \u00b1 0.10",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-16191303",
   "tg_tokens": 128,
   "tps": 139.89,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-18",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c16191303-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b08400",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3966.95,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3966.95 \u00b1 19.81; llama 7B Q4_0 | tg128 | 145.32 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-16191303",
   "tg_tokens": 128,
   "tps": 145.32,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-18",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c16191303-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b08400",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3164.42,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3164.42 \u00b1 29.55; llama 7B Q4_0 | tg128 | 161.53 \u00b1 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-16191303",
   "tg_tokens": 128,
   "tps": 161.53,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-18",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c16191303-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b08400",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3374.48,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3374.48 \u00b1 12.33; llama 7B Q4_0 | tg128 | 170.44 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-16191303",
   "tg_tokens": 128,
   "tps": 170.44,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-18",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c16191303-llama7b-Q4_0-5",
   "model": "llama 7B",
   "notes": "threads=99; config=b08400",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3552.47,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3552.47 \u00b1 39.58; llama 7B Q4_0 | tg128 | 144.84 \u00b1 0.36",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-16191303",
   "tg_tokens": 128,
   "tps": 144.84,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-18",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c16191303-llama7b-Q4_0-6",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b08400",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3458.31,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3458.31 \u00b1 33.75; llama 7B Q4_0 | tg128 | 150.40 \u00b1 0.87",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-16191303",
   "tg_tokens": 128,
   "tps": 150.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-18",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c16191303-llama7b-Q4_0-7",
   "model": "llama 7B",
   "notes": "threads=99; config=b08400",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 6695.94,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 6695.94 \u00b1 211.86; llama 7B Q4_0 | tg128 | 308.88 \u00b1 0.73",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-16191303",
   "tg_tokens": 128,
   "tps": 308.88,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-18",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c16191303-llama7b-Q4_0-8",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b08400",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 6508.13,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 6508.13 \u00b1 118.87; llama 7B Q4_0 | tg128 | 310.91 \u00b1 1.26",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-16191303",
   "tg_tokens": 128,
   "tps": 310.91,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-20",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c16218055-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3245.11,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3245.11 \u00b1 107.60; llama 7B Q4_0 | tg128 | 134.46 \u00b1 0.27",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-16218055",
   "tg_tokens": 128,
   "tps": 134.46,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-20",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c16218055-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3321.11,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3321.11 \u00b1 85.61; llama 7B Q4_0 | tg128 | 139.81 \u00b1 0.40",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-16218055",
   "tg_tokens": 128,
   "tps": 139.81,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-20",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c16218055-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 6695.4,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 6695.40 \u00b1 228.49; llama 7B Q4_0 | tg128 | 298.50 \u00b1 6.80",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-16218055",
   "tg_tokens": 128,
   "tps": 298.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-20",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c16218055-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 6287.12,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 6287.12 \u00b1 324.69; llama 7B Q4_0 | tg128 | 278.22 \u00b1 13.25",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-16218055",
   "tg_tokens": 128,
   "tps": 278.22,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-25",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-15021-c16312072-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b08514",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 5167.28,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 5167.28 \u00b1 106.65; llama 7B Q4_0 | tg128 | 116.50 \u00b1 0.32",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-16312072",
   "tg_tokens": 128,
   "tps": 116.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-25",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-15021-c16312072-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=0 ngl=99; config=b08514",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4753.07,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 4753.07 \u00b1 57.42; llama 7B Q4_0 | tg128 | 106.70 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-16312072",
   "tg_tokens": 128,
   "tps": 106.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-25",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-15021-c16312072-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; config=b08514",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4824.33,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 4824.33 \u00b1 27.00; llama 7B Q4_0 | tg128 | 119.97 \u00b1 0.50",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-16312072",
   "tg_tokens": 128,
   "tps": 119.97,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-03-25",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-15021-c16312072-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=0 ngl=99; config=b08514",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4661.14,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 4661.14 \u00b1 53.96; llama 7B Q4_0 | tg128 | 109.81 \u00b1 0.30",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-16312072",
   "tg_tokens": 128,
   "tps": 109.81,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-04-01",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-15021-c16413942-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b08611",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4766.48,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 4766.48 \u00b1 51.84; llama 7B Q4_0 | tg128 | 108.45 \u00b1 0.49",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-16413942",
   "tg_tokens": 128,
   "tps": 108.45,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-04-01",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-15021-c16413942-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b08611",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1728.29,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 1728.29 \u00b1 26.50; llama 7B Q4_0 | tg128 | 32.78 \u00b1 1.77",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-16413942",
   "tg_tokens": 128,
   "tps": 32.78,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-04-03",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-15021-c16436823-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1106.09,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; llama 7B Q4_0 | pp512 | 1106.09 \u00b1 17.99; llama 7B Q4_0 | pp1024 | 943.35 \u00b1 31.11; llama 7B Q4_0 | pp2048 | 825.25 \u00b1 0.63; llama 7B Q4_0 | pp4096 | 678.23 \u00b1 1.59; llama 7B Q4_0 | pp8192 | 457.73 \u00b1 0.57; llama 7B Q4_0 | pp16384 | 285.17 \u00b1 0.16; llama 7B Q4_0 | pp32768 | 149.55 \u00b1 0.68; llama 7B Q4_0 | tg128 | 44.15 \u00b1 0.03; llama 7B Q4_0 | tg256 | 44.06 \u00b1 0.08; llama 7B Q4_0 | tg512 | 40.62 \u00b1 0.01; llama 7B Q4_0 | tg1024 | 34.24 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-16436823",
   "tg_tokens": 128,
   "tps": 44.15,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-04-03",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-15021-c16436823-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1223.09,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; llama 7B Q4_0 | pp512 | 1223.09 \u00b1 32.28; llama 7B Q4_0 | pp1024 | 1222.64 \u00b1 1.41; llama 7B Q4_0 | pp2048 | 1155.04 \u00b1 0.86; llama 7B Q4_0 | pp4096 | 1027.97 \u00b1 1.46; llama 7B Q4_0 | tg128 | 49.50 \u00b1 0.06; llama 7B Q4_0 | tg256 | 49.49 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-16436823",
   "tg_tokens": 128,
   "tps": 49.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-04-03",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-15021-c16436823-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 8192,
   "pp_tps": 843.04,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; llama 7B Q4_0 | pp8192 | 843.04 \u00b1 2.52; llama 7B Q4_0 | tg1024 | 46.15 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-16436823",
   "tg_tokens": 1024,
   "tps": 46.15,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-04-20",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-15021-c16629634-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b08851",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 5166.48,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 5166.48 \u00b1 100.00; llama 7B Q4_0 | tg128 | 116.46 \u00b1 0.50",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-16629634",
   "tg_tokens": 128,
   "tps": 116.46,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-04-20",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-15021-c16629634-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=0 ngl=99; config=b08851",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4732.96,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 4732.96 \u00b1 59.88; llama 7B Q4_0 | tg128 | 106.63 \u00b1 0.30",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-16629634",
   "tg_tokens": 128,
   "tps": 106.63,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-09",
   "flags": [],
   "hardware": "Radeon RX 7600S",
   "id": "lc-disc-15021-c16865527-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 783.23,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7600S; llama 7B Q4_0 | pp512 | 783.23 \u00b1 5.71; llama 7B Q4_0 | tg128 | 45.02 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-16865527",
   "tg_tokens": 128,
   "tps": 45.02,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-09",
   "flags": [],
   "hardware": "Radeon RX 7600S",
   "id": "lc-disc-15021-c16865527-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 915.55,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7600S; llama 7B Q4_0 | pp512 | 915.55 \u00b1 4.70; llama 7B Q4_0 | tg128 | 49.62 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-16865527",
   "tg_tokens": 128,
   "tps": 49.62,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-13",
   "flags": [],
   "hardware": "Radeon 890M",
   "id": "lc-disc-15021-c16902593-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b09127",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 417.37,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 890M; llama 7B Q4_0 | pp512 | 417.37 \u00b1 2.25; llama 7B Q4_0 | pp1024 | 387.80 \u00b1 3.20; llama 7B Q4_0 | pp2048 | 322.97 \u00b1 4.57; llama 7B Q4_0 | pp4096 | 267.51 \u00b1 5.69; llama 7B Q4_0 | pp8192 | 201.26 \u00b1 0.19; llama 7B Q4_0 | pp16384 | 131.99 \u00b1 0.13; llama 7B Q4_0 | pp32768 | 65.26 \u00b1 0.07; llama 7B Q4_0 | tg128 | 16.12 \u00b1 0.10; llama 7B Q4_0 | tg256 | 15.88 \u00b1 0.05; llama 7B Q4_0 | tg512 | 15.14 \u00b1 0.02; llama 7B Q4_0 | tg1024 | 14.15 \u00b1 0.01; llama 7B Q4_0 | tg2048 | 12.64 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-16902593",
   "tg_tokens": 128,
   "tps": 16.12,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-13",
   "flags": [],
   "hardware": "Radeon 890M",
   "id": "lc-disc-15021-c16902593-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b09127",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 421.3,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 890M; llama 7B Q4_0 | pp512 | 421.30 \u00b1 6.79; llama 7B Q4_0 | pp1024 | 406.40 \u00b1 0.99; llama 7B Q4_0 | pp2048 | 386.99 \u00b1 0.62; llama 7B Q4_0 | pp4096 | 342.35 \u00b1 0.22; llama 7B Q4_0 | pp8192 | 265.00 \u00b1 0.24; llama 7B Q4_0 | pp16384 | 170.29 \u00b1 0.09; llama 7B Q4_0 | pp32768 | 96.47 \u00b1 0.08; llama 7B Q4_0 | tg128 | 16.73 \u00b1 0.11; llama 7B Q4_0 | tg256 | 17.05 \u00b1 0.06; llama 7B Q4_0 | tg512 | 16.50 \u00b1 0.03; llama 7B Q4_0 | tg1024 | 15.92 \u00b1 0.01; llama 7B Q4_0 | tg2048 | 14.97 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-16902593",
   "tg_tokens": 128,
   "tps": 16.73,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-05",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c14881921-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; config=b06957",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3552.27,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3552.27 \u00b1 101.96; llama 7B Q4_0 | tg128 | 167.11 \u00b1 0.50",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14881921",
   "tg_tokens": 128,
   "tps": 167.11,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-05",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c14881921-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b06957",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3874.25,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3874.25 \u00b1 11.92; llama 7B Q4_0 | tg128 | 170.12 \u00b1 0.56",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14881921",
   "tg_tokens": 128,
   "tps": 170.12,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-15",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-15021-c14975200-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=1 ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4336.99,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 4336.99 \u00b1 9.52; llama 7B Q4_0 | tg128 | 102.55 \u00b1 0.25",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14975200",
   "tg_tokens": 128,
   "tps": 102.55,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-15",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-15021-c14975200-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=0 ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4244.92,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 4244.92 \u00b1 8.97; llama 7B Q4_0 | tg128 | 98.13 \u00b1 0.17",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-14975200",
   "tg_tokens": 128,
   "tps": 98.13,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-26",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-15021-c15084265-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=1 ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 5055.19,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 5055.19 \u00b1 109.58; llama 7B Q4_0 | tg128 | 101.27 \u00b1 0.27",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-15084265",
   "tg_tokens": 128,
   "tps": 101.27,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-26",
   "flags": [],
   "hardware": "Radeon RX 9070 XT",
   "id": "lc-disc-15021-c15084265-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=0 ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4903.51,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9070 XT; llama 7B Q4_0 | pp512 | 4903.51 \u00b1 96.36; llama 7B Q4_0 | tg128 | 97.28 \u00b1 0.13",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-15084265",
   "tg_tokens": 128,
   "tps": 97.28,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-26",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-15021-c15089622-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b07166",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 908.35,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; llama 7B Q4_0 | pp512 | 908.35 \u00b1 2.08; llama 7B Q4_0 | tg128 | 46.51 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-15089622",
   "tg_tokens": 128,
   "tps": 46.51,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-26",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-15021-c15089622-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b07166",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1012.79,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; llama 7B Q4_0 | pp512 | 1012.79 \u00b1 2.51; llama 7B Q4_0 | tg128 | 50.59 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-15089622",
   "tg_tokens": 128,
   "tps": 50.59,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-26",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-15021-c15089622-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; config=b07166",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4443.54,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; dev=ROCM0; llama 7B Q4_0 | pp512 | 4443.54 \u00b1 339.25; llama 7B Q4_0 | tg128 | 93.84 \u00b1 0.26",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-15089622",
   "tg_tokens": 128,
   "tps": 93.84,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-26",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-15021-c15089622-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07166",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4773.07,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; dev=ROCM0; llama 7B Q4_0 | pp512 | 4773.07 \u00b1 49.30; llama 7B Q4_0 | tg128 | 97.98 \u00b1 0.13",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-15089622",
   "tg_tokens": 128,
   "tps": 97.98,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-26",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-15021-c15089622-llama7b-Q4_0-5",
   "model": "llama 7B",
   "notes": "threads=99; config=b07166",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3585.2,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; dev=ROCM1; llama 7B Q4_0 | pp512 | 3585.20 \u00b1 646.73; llama 7B Q4_0 | tg128 | 66.27 \u00b1 2.99",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-15089622",
   "tg_tokens": 128,
   "tps": 66.27,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-11-26",
   "flags": [
    "contradiction"
   ],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-15021-c15089622-llama7b-Q4_0-6",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07166",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4727.48,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; dev=ROCM1; llama 7B Q4_0 | pp512 | 4727.48 \u00b1 60.77; llama 7B Q4_0 | tg128 | 68.67 \u00b1 1.56",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-15089622",
   "tg_tokens": 128,
   "tps": 68.67,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-10",
   "flags": [],
   "hardware": "Radeon 660M",
   "id": "lc-disc-15021-c15223284-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b07349",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 120.52,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 660M; llama 7B Q4_0 | pp512 | 120.52 \u00b1 0.18; llama 7B Q4_0 | pp1024 | 112.86 \u00b1 0.24; llama 7B Q4_0 | pp2048 | 101.98 \u00b1 0.07; llama 7B Q4_0 | pp4096 | 86.82 \u00b1 0.24; llama 7B Q4_0 | tg128 | 12.72 \u00b1 0.13; llama 7B Q4_0 | tg256 | 12.46 \u00b1 0.21; llama 7B Q4_0 | tg512 | 11.89 \u00b1 0.19; llama 7B Q4_0 | tg1024 | 10.84 \u00b1 0.19; llama 7B Q4_0 | tg2048 | 9.38 \u00b1 0.16",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-15223284",
   "tg_tokens": 128,
   "tps": 12.72,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-10",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c15223284-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; config=b07349",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3429.68,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3429.68 \u00b1 63.23; llama 7B Q4_0 | tg128 | 119.00 \u00b1 0.19",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-15223284",
   "tg_tokens": 128,
   "tps": 119.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-10",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c15223284-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07349",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3730.42,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3730.42 \u00b1 42.87; llama 7B Q4_0 | tg128 | 126.28 \u00b1 0.15",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-15223284",
   "tg_tokens": 128,
   "tps": 126.28,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-10",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c15223284-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; config=b07349",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3541.67,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3541.67 \u00b1 39.92; llama 7B Q4_0 | tg128 | 131.70 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-15223284",
   "tg_tokens": 128,
   "tps": 131.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-10",
   "flags": [],
   "hardware": "Radeon RX 7900 XTX",
   "id": "lc-disc-15021-c15223284-llama7b-Q4_0-5",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07349",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3782.88,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 7900 XTX; llama 7B Q4_0 | pp512 | 3782.88 \u00b1 56.81; llama 7B Q4_0 | tg128 | 138.91 \u00b1 0.20",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-15223284",
   "tg_tokens": 128,
   "tps": 138.91,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-15021-c15234500-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99; config=b07358",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1352.52,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; llama 7B Q4_0 | pp512 | 1352.52 \u00b1 2.06; llama 7B Q4_0 | tg128 | 46.46 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-15234500",
   "tg_tokens": 128,
   "tps": 46.46,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-11",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-15021-c15234500-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99; config=b07358",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1576.59,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; llama 7B Q4_0 | pp512 | 1576.59 \u00b1 3.20; llama 7B Q4_0 | tg128 | 50.57 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-15234500",
   "tg_tokens": 128,
   "tps": 50.57,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-15",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-15021-c15257105-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=999; config=b07370",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1464.06,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; llama 7B Q4_0 | pp512 | 1464.06 \u00b1 13.51; llama 7B Q4_0 | tg128 | 50.29 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-15257105",
   "tg_tokens": 128,
   "tps": 50.29,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-15",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-15021-c15257105-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=999; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b07370",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1168.63,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; llama 7B Q4_0 | pp512 | 1168.63 \u00b1 2.49; llama 7B Q4_0 | tg128 | 45.53 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-15257105",
   "tg_tokens": 128,
   "tps": 45.53,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-20",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-15021-c15304050-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=1 ngl=99; config=b06867",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 998.73,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; llama 7B Q4_0 | pp512 | 998.73 \u00b1 0.60; llama 7B Q4_0 | tg128 | 50.30 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-15304050",
   "tg_tokens": 128,
   "tps": 50.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-02-25",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-15021-c15926380-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=999; config=b08155",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1488.56,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; llama 7B Q4_0 | pp512 | 1488.56 \u00b1 27.84; llama 7B Q4_0 | tg128 | 50.40 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-15926380",
   "tg_tokens": 128,
   "tps": 50.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-02-25",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-15021-c15926380-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=999; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b08155",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1201.51,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; llama 7B Q4_0 | pp512 | 1201.51 \u00b1 1.58; llama 7B Q4_0 | tg128 | 46.01 \u00b1 0.10",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-15926380",
   "tg_tokens": 128,
   "tps": 46.01,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-15",
   "flags": [],
   "hardware": "Radeon RX 9060 XT",
   "id": "lc-disc-15021-c15263439-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2427.16,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9060 XT; llama 7B Q4_0 | pp512 | 2427.16 \u00b1 21.65; llama 7B Q4_0 | tg128 | 63.98 \u00b1 0.17",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-15263439",
   "tg_tokens": 128,
   "tps": 63.98,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2025-12-15",
   "flags": [],
   "hardware": "Radeon RX 9060 XT",
   "id": "lc-disc-15021-c15263439-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2534.4,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9060 XT; llama 7B Q4_0 | pp512 | 2534.40 \u00b1 24.79; llama 7B Q4_0 | tg128 | 66.48 \u00b1 0.54",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-15263439",
   "tg_tokens": 128,
   "tps": 66.48,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-22",
   "flags": [],
   "hardware": "Arc A770",
   "id": "lc-disc-23313-c17015806-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=0 ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 912.36,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A770; llama 7B Q4_0 | pp512 | 912.36 \u00b1 3.15; llama 7B Q4_0 | tg128 | 34.47 \u00b1 0.15",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17015806",
   "tg_tokens": 128,
   "tps": 34.47,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-22",
   "flags": [],
   "hardware": "Arc A770",
   "id": "lc-disc-23313-c17015806-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 692.82,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A770; llama 7B Q4_0 | pp512 | 692.82 \u00b1 2.47; llama 7B Q4_0 | tg128 | 39.09 \u00b1 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17015806",
   "tg_tokens": 128,
   "tps": 39.09,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-27",
   "flags": [
    "contradiction"
   ],
   "hardware": "Arc B580",
   "id": "lc-disc-23313-c17077818-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; ngl=99; config=b09368",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1721.91,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc B580; llama 7B Q4_0 | pp512 | 1721.91 \u00b1 35.90; llama 7B Q4_0 | tg128 | 65.67 \u00b1 0.46",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17077818",
   "tg_tokens": 128,
   "tps": 65.67,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-27",
   "flags": [],
   "hardware": "Arc B580",
   "id": "lc-disc-23313-c17077818-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; ngl=99; config=b09368",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2063.52,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc B580; llama 7B Q4_0 | pp512 | 2063.52 \u00b1 3.52; llama 7B Q4_0 | tg128 | 73.76 \u00b1 0.13",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17077818",
   "tg_tokens": 128,
   "tps": 73.76,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-28",
   "flags": [],
   "hardware": "Arc Pro B50",
   "id": "lc-disc-23313-c17083412-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b09368",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 424.5,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B50; llama 7B Q4_0 | pp512 | 424.50 \u00b1 1.34; llama 7B Q4_0 | tg128 | 45.92 \u00b1 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17083412",
   "tg_tokens": 128,
   "tps": 45.92,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-28",
   "flags": [],
   "hardware": "Arc Pro B50",
   "id": "lc-disc-23313-c17083412-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; config=b09368",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 397.09,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B50; llama 7B Q4_0 | pp512 | 397.09 \u00b1 1.02; llama 7B Q4_0 | tg128 | 45.74 \u00b1 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17083412",
   "tg_tokens": 128,
   "tps": 45.74,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-28",
   "flags": [],
   "hardware": "Arc Pro B50",
   "id": "lc-disc-23313-c17083412-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b09368",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 397.73,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B50; llama 7B Q4_0 | pp512 | 397.73 \u00b1 0.23; llama 7B Q4_0 | tg128 | 47.80 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17083412",
   "tg_tokens": 128,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-28",
   "flags": [],
   "hardware": "Arc Pro B50",
   "id": "lc-disc-23313-c17083412-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; config=b09368",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 590.01,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B50; llama 7B Q4_0 | pp512 | 590.01 \u00b1 0.93; llama 7B Q4_0 | tg128 | 40.13 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17083412",
   "tg_tokens": 128,
   "tps": 40.13,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-05-28",
   "flags": [],
   "hardware": "Arc Pro B50",
   "id": "lc-disc-23313-c17083412-llama7b-Q4_0-5",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b09368",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 581.78,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B50; llama 7B Q4_0 | pp512 | 581.78 \u00b1 1.61; llama 7B Q4_0 | tg128 | 41.82 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17083412",
   "tg_tokens": 128,
   "tps": 41.82,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-04",
   "flags": [],
   "hardware": "Arc B580",
   "id": "lc-disc-23313-c17176270-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=-1; config=b09501",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1737.0,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc B580; llama 7B Q4_0 | pp512 | 1737.00 \u00b1 16.94; llama 7B Q4_0 | tg128 | 78.15 \u00b1 0.22",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17176270",
   "tg_tokens": 128,
   "tps": 78.15,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-04",
   "flags": [],
   "hardware": "Arc B580",
   "id": "lc-disc-23313-c17176270-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=-1; config=b09501",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1542.23,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc B580; llama 7B Q4_0 | pp512 | 1542.23 \u00b1 24.72; llama 7B Q4_0 | tg128 | 83.39 \u00b1 0.23",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17176270",
   "tg_tokens": 128,
   "tps": 83.39,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-04",
   "flags": [],
   "hardware": "Arc B580",
   "id": "lc-disc-23313-c17176270-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=-1; config=b09501",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1851.35,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc B580; dev=SYCL0; llama 7B Q4_0 | pp512 | 1851.35 \u00b1 1.16; llama 7B Q4_0 | tg128 | 77.92 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17176270",
   "tg_tokens": 128,
   "tps": 77.92,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-04",
   "flags": [],
   "hardware": "Arc B580",
   "id": "lc-disc-23313-c17176270-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=-1; config=b09501",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1631.3,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc B580; dev=SYCL0; llama 7B Q4_0 | pp512 | 1631.30 \u00b1 1.21; llama 7B Q4_0 | tg128 | 82.85 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17176270",
   "tg_tokens": 128,
   "tps": 82.85,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-04",
   "flags": [],
   "hardware": "Arc B580",
   "id": "lc-disc-23313-c17176270-llama7b-Q4_0-5",
   "model": "llama 7B",
   "notes": "threads=-1; config=b09501",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2015.51,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc B580; dev=SYCL1; llama 7B Q4_0 | pp512 | 2015.51 \u00b1 15.87; llama 7B Q4_0 | tg128 | 82.49 \u00b1 0.38",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17176270",
   "tg_tokens": 128,
   "tps": 82.49,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-04",
   "flags": [],
   "hardware": "Arc B580",
   "id": "lc-disc-23313-c17176270-llama7b-Q4_0-6",
   "model": "llama 7B",
   "notes": "threads=-1; config=b09501",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1806.18,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc B580; dev=SYCL1; llama 7B Q4_0 | pp512 | 1806.18 \u00b1 8.68; llama 7B Q4_0 | tg128 | 86.54 \u00b1 0.33",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17176270",
   "tg_tokens": 128,
   "tps": 86.54,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-04",
   "flags": [
    "contradiction"
   ],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c17176270-llama7b-Q4_0-7",
   "model": "llama 7B",
   "notes": "threads=-1; fa=0 ngl=-1; config=b09501",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1581.29,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B60; llama 7B Q4_0 | pp512 | 1581.29 \u00b1 43.72; llama 7B Q4_0 | tg128 | 53.56 \u00b1 14.29",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17176270",
   "tg_tokens": 128,
   "tps": 53.56,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-04",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c17176270-llama7b-Q4_0-8",
   "model": "llama 7B",
   "notes": "threads=-1; fa=0 ngl=-1; config=b09501",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1667.43,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B60; llama 7B Q4_0 | pp512 | 1667.43 \u00b1 11.14; llama 7B Q4_0 | tg128 | 65.94 \u00b1 0.47",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17176270",
   "tg_tokens": 128,
   "tps": 65.94,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-04",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c17176270-llama7b-Q4_0-9",
   "model": "llama 7B",
   "notes": "threads=-1; config=b09501",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1626.71,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B60; llama 7B Q4_0 | pp512 | 1626.71 \u00b1 2.51; llama 7B Q4_0 | tg128 | 69.46 \u00b1 0.41",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17176270",
   "tg_tokens": 128,
   "tps": 69.46,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-04",
   "flags": [
    "contradiction"
   ],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c17176270-llama7b-Q4_0-10",
   "model": "llama 7B",
   "notes": "threads=-1; config=b09501",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1621.0,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B60; dev=VULKAN1; llama 7B Q4_0 | pp512 | 1621.00 \u00b1 2.51; llama 7B Q4_0 | tg128 | 68.07 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17176270",
   "tg_tokens": 128,
   "tps": 68.07,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-04",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c17176270-llama7b-Q4_0-11",
   "model": "llama 7B",
   "notes": "threads=-1; config=b09501",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1572.29,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B60; dev=VULKAN1; llama 7B Q4_0 | pp512 | 1572.29 \u00b1 2.51; llama 7B Q4_0 | tg128 | 71.42 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17176270",
   "tg_tokens": 128,
   "tps": 71.42,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-04",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c17176270-llama7b-Q4_0-12",
   "model": "llama 7B",
   "notes": "threads=-1; config=b09501",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1929.93,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B60; dev=VULKAN0; llama 7B Q4_0 | pp512 | 1929.93 \u00b1 17.92; llama 7B Q4_0 | tg128 | 73.01 \u00b1 0.44",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17176270",
   "tg_tokens": 128,
   "tps": 73.01,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-04",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c17176270-llama7b-Q4_0-13",
   "model": "llama 7B",
   "notes": "threads=-1; config=b09501",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1841.31,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B60; dev=VULKAN0; llama 7B Q4_0 | pp512 | 1841.31 \u00b1 22.24; llama 7B Q4_0 | tg128 | 76.83 \u00b1 0.18",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17176270",
   "tg_tokens": 128,
   "tps": 76.83,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-05",
   "flags": [],
   "hardware": "Arc A380",
   "id": "lc-disc-23313-c17194767-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=-1; config=b09501",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 495.16,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A380; llama 7B Q4_0 | pp512 | 495.16 \u00b1 1.43; llama 7B Q4_0 | tg128 | 23.84 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17194767",
   "tg_tokens": 128,
   "tps": 23.84,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-05",
   "flags": [],
   "hardware": "Arc A380",
   "id": "lc-disc-23313-c17194767-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b09501",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 362.69,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A380; llama 7B Q4_0 | pp512 | 362.69 \u00b1 0.27; llama 7B Q4_0 | tg128 | 25.57 \u00b1 0.20",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17194767",
   "tg_tokens": 128,
   "tps": 25.57,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-05",
   "flags": [],
   "hardware": "Arc A380",
   "id": "lc-disc-23313-c17194767-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=-1; config=b09501",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 226.11,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A380; llama 7B Q4_0 | pp512 | 226.11 \u00b1 2.39; llama 7B Q4_0 | tg128 | 23.81 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17194767",
   "tg_tokens": 128,
   "tps": 23.81,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-05",
   "flags": [],
   "hardware": "Arc A380",
   "id": "lc-disc-23313-c17194767-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b09501",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 213.39,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A380; llama 7B Q4_0 | pp512 | 213.39 \u00b1 8.38; llama 7B Q4_0 | tg128 | 25.81 \u00b1 0.41",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17194767",
   "tg_tokens": 128,
   "tps": 25.81,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-05",
   "flags": [],
   "hardware": "Arc A380",
   "id": "lc-disc-23313-c17197062-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=-1; fa=0 ngl=-1; config=b09534",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 478.53,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A380; llama 7B Q4_0 | pp512 | 478.53 \u00b1 0.65; llama 7B Q4_0 | tg128 | 18.19 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17197062",
   "tg_tokens": 128,
   "tps": 18.19,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-05",
   "flags": [],
   "hardware": "Arc A380",
   "id": "lc-disc-23313-c17197062-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=-1; config=b09534",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 351.23,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A380; llama 7B Q4_0 | pp512 | 351.23 \u00b1 0.62; llama 7B Q4_0 | tg128 | 19.12 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17197062",
   "tg_tokens": 128,
   "tps": 19.12,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-05",
   "flags": [],
   "hardware": "Arc A380",
   "id": "lc-disc-23313-c17197062-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=-1; fa=0 ngl=-1; config=b09534",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 220.98,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A380; llama 7B Q4_0 | pp512 | 220.98 \u00b1 0.30; llama 7B Q4_0 | tg128 | 18.16 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17197062",
   "tg_tokens": 128,
   "tps": 18.16,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-05",
   "flags": [],
   "hardware": "Arc A380",
   "id": "lc-disc-23313-c17197062-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=-1; config=b09534",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 216.06,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A380; llama 7B Q4_0 | pp512 | 216.06 \u00b1 0.11; llama 7B Q4_0 | tg128 | 19.70 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17197062",
   "tg_tokens": 128,
   "tps": 19.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-05",
   "flags": [],
   "hardware": "Arc A380",
   "id": "lc-disc-23313-c17197062-llama7b-Q4_0-5",
   "model": "llama 7B",
   "notes": "threads=-1; config=b09534",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 491.34,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A380; llama 7B Q4_0 | pp512 | 491.34 \u00b1 1.26; llama 7B Q4_0 | tg128 | 23.97 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17197062",
   "tg_tokens": 128,
   "tps": 23.97,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-05",
   "flags": [],
   "hardware": "Arc A380",
   "id": "lc-disc-23313-c17197062-llama7b-Q4_0-6",
   "model": "llama 7B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b09534",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 358.39,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A380; llama 7B Q4_0 | pp512 | 358.39 \u00b1 0.28; llama 7B Q4_0 | tg128 | 25.63 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17197062",
   "tg_tokens": 128,
   "tps": 25.63,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-05",
   "flags": [],
   "hardware": "Arc A380",
   "id": "lc-disc-23313-c17197062-llama7b-Q4_0-7",
   "model": "llama 7B",
   "notes": "threads=-1; config=b09534",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 224.16,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A380; llama 7B Q4_0 | pp512 | 224.16 \u00b1 0.64; llama 7B Q4_0 | tg128 | 24.04 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17197062",
   "tg_tokens": 128,
   "tps": 24.04,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-05",
   "flags": [],
   "hardware": "Arc A380",
   "id": "lc-disc-23313-c17197062-llama7b-Q4_0-8",
   "model": "llama 7B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b09534",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 220.1,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A380; llama 7B Q4_0 | pp512 | 220.10 \u00b1 0.21; llama 7B Q4_0 | tg128 | 26.57 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17197062",
   "tg_tokens": 128,
   "tps": 26.57,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-12",
   "flags": [],
   "hardware": "Arc A380",
   "id": "lc-disc-23313-c17285273-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; fa=1 ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 276.43,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A380; llama 7B Q4_0 | pp512 | 276.43 \u00b1 0.23; llama 7B Q4_0 | tg128 | 16.17 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17285273",
   "tg_tokens": 128,
   "tps": 16.17,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-12",
   "flags": [],
   "hardware": "Arc A380",
   "id": "lc-disc-23313-c17285273-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=0 ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 347.28,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A380; llama 7B Q4_0 | pp512 | 347.28 \u00b1 0.06; llama 7B Q4_0 | tg128 | 14.95 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17285273",
   "tg_tokens": 128,
   "tps": 14.95,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-12",
   "flags": [],
   "hardware": "Arc A380",
   "id": "lc-disc-23313-c17285273-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99; fa=1 ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 170.14,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A380; llama 7B Q4_0 | pp512 | 170.14 \u00b1 0.26; llama 7B Q4_0 | tg128 | 15.97 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17285273",
   "tg_tokens": 128,
   "tps": 15.97,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-12",
   "flags": [],
   "hardware": "Arc A380",
   "id": "lc-disc-23313-c17285273-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=0 ngl=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 201.62,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A380; llama 7B Q4_0 | pp512 | 201.62 \u00b1 0.06; llama 7B Q4_0 | tg128 | 14.96 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17285273",
   "tg_tokens": 128,
   "tps": 14.96,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-12",
   "flags": [],
   "hardware": "Arc A380",
   "id": "lc-disc-23313-c17285273-llama7b-Q4_0-5",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 259.51,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A380; llama 7B Q4_0 | pp512 | 259.51 \u00b1 0.42; llama 7B Q4_0 | tg128 | 17.57 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17285273",
   "tg_tokens": 128,
   "tps": 17.57,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-12",
   "flags": [],
   "hardware": "Arc A380",
   "id": "lc-disc-23313-c17285273-llama7b-Q4_0-6",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 247.03,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A380; llama 7B Q4_0 | pp512 | 247.03 \u00b1 0.24; llama 7B Q4_0 | tg128 | 17.25 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17285273",
   "tg_tokens": 128,
   "tps": 17.25,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (OpenVINO)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-12",
   "flags": [],
   "hardware": "Arc A380",
   "id": "lc-disc-23313-c17285273-llama7b-Q4_0-7",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1497.21,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A380; llama 7B Q4_0 | pp512 | 1497.21 \u00b1 3.36; llama 7B Q4_0 | tg128 | 18.18 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17285273",
   "tg_tokens": 128,
   "tps": 18.18,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-13",
   "flags": [],
   "hardware": "Arc A380",
   "id": "lc-disc-23313-c17288632-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 441.14,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A380; llama 7B Q4_0 | pp512 | 441.14 \u00b1 0.30; llama 7B Q4_0 | tg128 | 19.76 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17288632",
   "tg_tokens": 128,
   "tps": 19.76,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-13",
   "flags": [],
   "hardware": "Arc A380",
   "id": "lc-disc-23313-c17288632-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 351.23,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A380; llama 7B Q4_0 | pp512 | 351.23 \u00b1 0.09; llama 7B Q4_0 | tg128 | 18.62 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17288632",
   "tg_tokens": 128,
   "tps": 18.62,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-13",
   "flags": [],
   "hardware": "Arc A380",
   "id": "lc-disc-23313-c17288632-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 227.06,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A380; llama 7B Q4_0 | pp512 | 227.06 \u00b1 0.46; llama 7B Q4_0 | tg128 | 20.26 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17288632",
   "tg_tokens": 128,
   "tps": 20.26,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-13",
   "flags": [],
   "hardware": "Arc A380",
   "id": "lc-disc-23313-c17288632-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 203.17,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A380; llama 7B Q4_0 | pp512 | 203.17 \u00b1 0.08; llama 7B Q4_0 | tg128 | 18.54 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17288632",
   "tg_tokens": 128,
   "tps": 18.54,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (OpenVINO)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-13",
   "flags": [],
   "hardware": "Arc A380",
   "id": "lc-disc-23313-c17288632-llama7b-Q4_0-5",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1513.67,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A380; llama 7B Q4_0 | pp512 | 1513.67 \u00b1 2.71; llama 7B Q4_0 | tg128 | 19.21 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17288632",
   "tg_tokens": 128,
   "tps": 19.21,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-13",
   "flags": [],
   "hardware": "Arc A380",
   "id": "lc-disc-23313-c17288632-llama7b-Q4_0-6",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 263.7,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A380; llama 7B Q4_0 | pp512 | 263.70 \u00b1 0.17; llama 7B Q4_0 | tg128 | 17.87 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17288632",
   "tg_tokens": 128,
   "tps": 17.87,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-13",
   "flags": [],
   "hardware": "Arc A380",
   "id": "lc-disc-23313-c17288632-llama7b-Q4_0-7",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 249.81,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A380; llama 7B Q4_0 | pp512 | 249.81 \u00b1 0.12; llama 7B Q4_0 | tg128 | 17.55 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17288632",
   "tg_tokens": 128,
   "tps": 17.55,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-12",
   "flags": [],
   "hardware": "Arc B580",
   "id": "lc-disc-23313-c17983413-gemma4b-Q4_K - Medium",
   "model": "gemma4 ?B",
   "notes": "threads=99; config=b02026",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1133.28,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "Arc B580; gemma4 ?B Q4_K - Medium | pp512 | 1133.28 \u00b1 3.24; gemma4 ?B Q4_K - Medium | tg128 | 37.00 \u00b1 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17983413",
   "tg_tokens": 128,
   "tps": 37.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-12",
   "flags": [],
   "hardware": "Arc B580",
   "id": "lc-disc-23313-c17983413-gemma4b-Q4_K - Medium-2",
   "model": "gemma4 ?B",
   "notes": "threads=99; config=b02026",
   "params": null,
   "power_w": null,
   "pp_tokens": 4096,
   "pp_tps": 1303.98,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "Arc B580; gemma4 ?B Q4_K - Medium | pp4096 | 1303.98 \u00b1 0.60; gemma4 ?B Q4_K - Medium | tg512 | 35.59 \u00b1 0.11",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17983413",
   "tg_tokens": 512,
   "tps": 35.59,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-12",
   "flags": [],
   "hardware": "Arc B580",
   "id": "lc-disc-23313-c17983413-granitemoe3b-Q8_0",
   "model": "granitemoe 3B",
   "notes": "threads=99; config=b02026",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3420.6,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc B580; granitemoe 3B Q8_0 | pp512 | 3420.60 \u00b1 34.31; granitemoe 3B Q8_0 | tg128 | 73.04 \u00b1 0.14",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17983413",
   "tg_tokens": 128,
   "tps": 73.04,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-12",
   "flags": [],
   "hardware": "Arc B580",
   "id": "lc-disc-23313-c17983413-granitemoe3b-Q8_0-2",
   "model": "granitemoe 3B",
   "notes": "threads=99; config=b02026",
   "params": null,
   "power_w": null,
   "pp_tokens": 4096,
   "pp_tps": 3611.53,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc B580; granitemoe 3B Q8_0 | pp4096 | 3611.53 \u00b1 24.25; granitemoe 3B Q8_0 | tg512 | 71.02 \u00b1 0.47",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17983413",
   "tg_tokens": 512,
   "tps": 71.02,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-17",
   "flags": [],
   "hardware": "Arc A770",
   "id": "lc-disc-23313-c18057036-museglimmer30b-Q4_K - Small",
   "model": "muse-glimmer 30B",
   "notes": "threads=99; fa=1 ngl=99 sm=layer",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 528.15,
   "provenance": "community",
   "quant": "Q4_K - Small",
   "quote": "Arc A770; dev=SYCL0/SYCL1; muse-glimmer 30B Q4_K - Small | pp512 | 528.15 \u00b1 0.76; muse-glimmer 30B Q4_K - Small | tg128 | 13.66 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18057036",
   "tg_tokens": 128,
   "tps": 13.66,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-17",
   "flags": [],
   "hardware": "Arc A770",
   "id": "lc-disc-23313-c18057036-museglimmer30b-Q4_K - Small-2",
   "model": "muse-glimmer 30B",
   "notes": "threads=99; run 2 of 3 in this comment (separate llama-bench run; build/flags differ); fa=0 ngl=99 sm=layer",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 598.6,
   "provenance": "community",
   "quant": "Q4_K - Small",
   "quote": "Arc A770; dev=SYCL0/SYCL1; muse-glimmer 30B Q4_K - Small | pp512 | 598.60 \u00b1 0.46; muse-glimmer 30B Q4_K - Small | tg128 | 12.57 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18057036",
   "tg_tokens": 128,
   "tps": 12.57,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-17",
   "flags": [],
   "hardware": "Arc A770",
   "id": "lc-disc-23313-c18057036-museglimmer30b-Q4_K - Small-3",
   "model": "muse-glimmer 30B",
   "notes": "threads=99; run 3 of 3 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=99 sm=tensor",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 445.85,
   "provenance": "community",
   "quant": "Q4_K - Small",
   "quote": "Arc A770; dev=SYCL0/SYCL1; muse-glimmer 30B Q4_K - Small | pp512 | 445.85 \u00b1 0.69; muse-glimmer 30B Q4_K - Small | tg128 | 8.69 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18057036",
   "tg_tokens": 128,
   "tps": 8.69,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-17",
   "flags": [],
   "hardware": "Arc A770",
   "id": "lc-disc-23313-c18057036-museglimmer30b-Q4_K - Small-4",
   "model": "muse-glimmer 30B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 544.62,
   "provenance": "community",
   "quant": "Q4_K - Small",
   "quote": "Arc A770; dev=SYCL0/SYCL1; muse-glimmer 30B Q4_K - Small | pp512 | 544.62 \u00b1 0.29; muse-glimmer 30B Q4_K - Small | tg128 | 18.68 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18057036",
   "tg_tokens": 128,
   "tps": 18.68,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-17",
   "flags": [],
   "hardware": "Arc A770",
   "id": "lc-disc-23313-c18057036-museglimmer30b-Q4_K - Small-5",
   "model": "muse-glimmer 30B",
   "notes": "threads=99; run 2 of 3 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 623.17,
   "provenance": "community",
   "quant": "Q4_K - Small",
   "quote": "Arc A770; dev=SYCL0/SYCL1; muse-glimmer 30B Q4_K - Small | pp512 | 623.17 \u00b1 0.25; muse-glimmer 30B Q4_K - Small | tg128 | 17.88 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18057036",
   "tg_tokens": 128,
   "tps": 17.88,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-17",
   "flags": [],
   "hardware": "Arc A770",
   "id": "lc-disc-23313-c18057036-museglimmer30b-Q4_K - Small-6",
   "model": "muse-glimmer 30B",
   "notes": "threads=99; run 3 of 3 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 467.16,
   "provenance": "community",
   "quant": "Q4_K - Small",
   "quote": "Arc A770; dev=SYCL0/SYCL1; muse-glimmer 30B Q4_K - Small | pp512 | 467.16 \u00b1 0.61; muse-glimmer 30B Q4_K - Small | tg128 | 9.98 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18057036",
   "tg_tokens": 128,
   "tps": 9.98,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-17",
   "flags": [],
   "hardware": "Arc A770",
   "id": "lc-disc-23313-c18057036-museglimmer30b-Q4_K - Small-7",
   "model": "muse-glimmer 30B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 546.94,
   "provenance": "community",
   "quant": "Q4_K - Small",
   "quote": "Arc A770; dev=SYCL0/SYCL1; muse-glimmer 30B Q4_K - Small | pp512 | 546.94 \u00b1 0.47; muse-glimmer 30B Q4_K - Small | tg128 | 18.68 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18057036",
   "tg_tokens": 128,
   "tps": 18.68,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-17",
   "flags": [],
   "hardware": "Arc A770",
   "id": "lc-disc-23313-c18057036-museglimmer30b-Q4_K - Small-8",
   "model": "muse-glimmer 30B",
   "notes": "threads=99; run 2 of 3 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 625.05,
   "provenance": "community",
   "quant": "Q4_K - Small",
   "quote": "Arc A770; dev=SYCL0/SYCL1; muse-glimmer 30B Q4_K - Small | pp512 | 625.05 \u00b1 1.07; muse-glimmer 30B Q4_K - Small | tg128 | 17.90 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18057036",
   "tg_tokens": 128,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-17",
   "flags": [],
   "hardware": "Arc A770",
   "id": "lc-disc-23313-c18057036-museglimmer30b-Q4_K - Small-9",
   "model": "muse-glimmer 30B",
   "notes": "threads=99; run 3 of 3 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 580.37,
   "provenance": "community",
   "quant": "Q4_K - Small",
   "quote": "Arc A770; dev=SYCL0/SYCL1; muse-glimmer 30B Q4_K - Small | pp512 | 580.37 \u00b1 0.55; muse-glimmer 30B Q4_K - Small | tg128 | 10.12 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18057036",
   "tg_tokens": 128,
   "tps": 10.12,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "Arc A770",
   "id": "lc-disc-23313-c18057036-museglimmer30b-Q4_K - Small-10",
   "model": "muse-glimmer 30B",
   "notes": "threads=99; fa=1 ngl=99 sm=tensor",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 547.77,
   "provenance": "community",
   "quant": "Q4_K - Small",
   "quote": "Arc A770; dev=SYCL0/SYCL1; muse-glimmer 30B Q4_K - Small | pp512 | 547.77 \u00b1 0.75; muse-glimmer 30B Q4_K - Small | tg128 | 9.82 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18057036",
   "tg_tokens": 128,
   "tps": 9.82,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-17",
   "flags": [],
   "hardware": "Arc A770",
   "id": "lc-disc-23313-c18057036-qwen3527b-Q4_K - Medium",
   "model": "qwen35 27B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 452.7,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "Arc A770; dev=SYCL0/SYCL1; qwen35 27B Q4_K - Medium | pp512 | 452.70 \u00b1 0.65; qwen35 27B Q4_K - Medium | tg128 | 17.40 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18057036",
   "tg_tokens": 128,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-17",
   "flags": [],
   "hardware": "Arc A770",
   "id": "lc-disc-23313-c18057036-qwen3527b-Q4_K - Medium-2",
   "model": "qwen35 27B",
   "notes": "threads=99; run 2 of 3 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 524.5,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "Arc A770; dev=SYCL0/SYCL1; qwen35 27B Q4_K - Medium | pp512 | 524.50 \u00b1 1.87; qwen35 27B Q4_K - Medium | tg128 | 17.63 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18057036",
   "tg_tokens": 128,
   "tps": 17.63,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-17",
   "flags": [],
   "hardware": "Arc A770",
   "id": "lc-disc-23313-c18057036-qwen3527b-Q4_K - Medium-3",
   "model": "qwen35 27B",
   "notes": "threads=99; run 3 of 3 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 516.6,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "Arc A770; dev=SYCL0/SYCL1; qwen35 27B Q4_K - Medium | pp512 | 516.60 \u00b1 1.24; qwen35 27B Q4_K - Medium | tg128 | 8.61 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18057036",
   "tg_tokens": 128,
   "tps": 8.61,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-21",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18103363-qwen3527b-Q4_K - Medium",
   "model": "qwen35 27B",
   "notes": "threads=99; config=b10553",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 788.86,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "Arc Pro B70; qwen35 27B Q4_K - Medium | pp512 | 788.86 \u00b1 2.85; qwen35 27B Q4_K - Medium | pp1024 | 805.80 \u00b1 1.28; qwen35 27B Q4_K - Medium | tg512 | 20.53 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18103363",
   "tg_tokens": 512,
   "tps": 20.53,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-21",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18103363-qwen3527b-Q5_K - Medium",
   "model": "qwen35 27B",
   "notes": "threads=99; config=b10553",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 949.64,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "Arc Pro B70; qwen35 27B Q5_K - Medium | pp512 | 949.64 \u00b1 4.58; qwen35 27B Q5_K - Medium | pp1024 | 970.65 \u00b1 3.16; qwen35 27B Q5_K - Medium | tg512 | 21.41 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18103363",
   "tg_tokens": 512,
   "tps": 21.41,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-21",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18103363-qwen3527b-Q6_K",
   "model": "qwen35 27B",
   "notes": "threads=99; config=b10553",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 950.58,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "Arc Pro B70; qwen35 27B Q6_K | pp512 | 950.58 \u00b1 3.76; qwen35 27B Q6_K | pp1024 | 965.97 \u00b1 0.65; qwen35 27B Q6_K | tg512 | 19.75 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18103363",
   "tg_tokens": 512,
   "tps": 19.75,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-26",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18155234-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=-1",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3009.58,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B70; llama 7B Q4_0 | pp512 | 3009.58 \u00b1 6.21; llama 7B Q4_0 | tg128 | 106.99 \u00b1 0.25",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18155234",
   "tg_tokens": 128,
   "tps": 106.99,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-26",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18155234-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1434.13,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B70; llama 7B Q4_0 | pp512 | 1434.13 \u00b1 0.58; llama 7B Q4_0 | tg128 | 112.08 \u00b1 0.17",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18155234",
   "tg_tokens": 128,
   "tps": 112.08,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-26",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18155234-qwen3527b-Q8_0",
   "model": "qwen35 27B",
   "notes": "threads=-1",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1046.98,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B70; qwen35 27B Q8_0 | pp512 | 1046.98 \u00b1 3.37; qwen35 27B Q8_0 | tg128 | 16.66 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18155234",
   "tg_tokens": 128,
   "tps": 16.66,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-26",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18155234-qwen3527b-Q8_0-2",
   "model": "qwen35 27B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 572.1,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B70; qwen35 27B Q8_0 | pp512 | 572.10 \u00b1 0.28; qwen35 27B Q8_0 | tg128 | 16.63 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18155234",
   "tg_tokens": 128,
   "tps": 16.63,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-26",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18155234-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=-1",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 971.17,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B70; llama 7B Q4_0 | pp512 | 971.17 \u00b1 1.37; llama 7B Q4_0 | tg128 | 106.69 \u00b1 0.11",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18155234",
   "tg_tokens": 128,
   "tps": 106.69,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-26",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18155234-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1133.3,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B70; llama 7B Q4_0 | pp512 | 1133.30 \u00b1 0.83; llama 7B Q4_0 | tg128 | 111.30 \u00b1 0.22",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18155234",
   "tg_tokens": 128,
   "tps": 111.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-26",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18155234-qwen3527b-Q8_0-3",
   "model": "qwen35 27B",
   "notes": "threads=-1",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 335.62,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B70; qwen35 27B Q8_0 | pp512 | 335.62 \u00b1 0.42; qwen35 27B Q8_0 | tg128 | 16.64 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18155234",
   "tg_tokens": 128,
   "tps": 16.64,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-26",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18155234-qwen3527b-Q8_0-4",
   "model": "qwen35 27B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 300.36,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B70; qwen35 27B Q8_0 | pp512 | 300.36 \u00b1 0.13; qwen35 27B Q8_0 | tg128 | 16.60 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18155234",
   "tg_tokens": 128,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-26",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18155234-llama7b-Q4_0-5",
   "model": "llama 7B",
   "notes": "threads=-1",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4271.32,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B70; dev=VULKAN1; llama 7B Q4_0 | pp512 | 4271.32 \u00b1 6.30; llama 7B Q4_0 | tg128 | 112.43 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18155234",
   "tg_tokens": 128,
   "tps": 112.43,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-26",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18155234-llama7b-Q4_0-6",
   "model": "llama 7B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 4651.26,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B70; dev=VULKAN1; llama 7B Q4_0 | pp512 | 4651.26 \u00b1 15.15; llama 7B Q4_0 | tg128 | 116.67 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18155234",
   "tg_tokens": 128,
   "tps": 116.67,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-26",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18155234-qwen3527b-Q8_0-5",
   "model": "qwen35 27B",
   "notes": "threads=-1",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1140.11,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B70; dev=VULKAN1; qwen35 27B Q8_0 | pp512 | 1140.11 \u00b1 3.26; qwen35 27B Q8_0 | tg128 | 18.56 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18155234",
   "tg_tokens": 128,
   "tps": 18.56,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-26",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18155234-qwen3527b-Q8_0-6",
   "model": "qwen35 27B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1138.09,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B70; dev=VULKAN1; qwen35 27B Q8_0 | pp512 | 1138.09 \u00b1 1.59; qwen35 27B Q8_0 | tg128 | 18.60 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18155234",
   "tg_tokens": 128,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-26",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18164236-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=-1; config=b10636",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 970.74,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B70; llama 7B Q4_0 | pp512 | 970.74  \u00b1 0.70; llama 7B Q4_0 | tg128 | 105.45  \u00b1 0.17",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18164236",
   "tg_tokens": 128,
   "tps": 105.45,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-26",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18164236-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10636",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1135.88,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B70; llama 7B Q4_0 | pp512 | 1135.88  \u00b1 1.97; llama 7B Q4_0 | tg128 | 109.99  \u00b1 0.11",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18164236",
   "tg_tokens": 128,
   "tps": 109.99,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-26",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18164236-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=-1; config=b10636",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2830.14,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B70; llama 7B Q4_0 | pp512 | 2830.14 \u00b1 46.23; llama 7B Q4_0 | tg128 | 105.64 \u00b1 0.22",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18164236",
   "tg_tokens": 128,
   "tps": 105.64,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-26",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18164236-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10636",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1382.27,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B70; llama 7B Q4_0 | pp512 | 1382.27 \u00b1 1.99; llama 7B Q4_0 | tg128 | 109.72 \u00b1 0.14",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18164236",
   "tg_tokens": 128,
   "tps": 109.72,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-26",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18164236-mistral314b-Q4_0",
   "model": "mistral3 14B",
   "notes": "threads=99; config=b10636",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 972.13,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B70; mistral3 14B Q4_0 | pp512 | 972.13 \u00b1 3.02; mistral3 14B Q4_0 | tg128 | 34.24 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18164236",
   "tg_tokens": 128,
   "tps": 34.24,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-26",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18164236-mistral314b-Q5_K - Medium",
   "model": "mistral3 14B",
   "notes": "threads=99; config=b10636",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1073.66,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "Arc Pro B70; mistral3 14B Q5_K - Medium | pp512 | 1073.66 \u00b1 4.43; mistral3 14B Q5_K - Medium | tg128 | 30.04 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18164236",
   "tg_tokens": 128,
   "tps": 30.04,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-26",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18164236-mistral314b-Q6_K",
   "model": "mistral3 14B",
   "notes": "threads=99; config=b10636",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1199.46,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "Arc Pro B70; mistral3 14B Q6_K | pp512 | 1199.46 \u00b1 12.20; mistral3 14B Q6_K | tg128 | 26.85 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18164236",
   "tg_tokens": 128,
   "tps": 26.85,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-26",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18164236-mistral314b-Q8_0",
   "model": "mistral3 14B",
   "notes": "threads=99; config=b10636",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1189.47,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B70; mistral3 14B Q8_0 | pp512 | 1189.47 \u00b1 2.36; mistral3 14B Q8_0 | tg128 | 18.87 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18164236",
   "tg_tokens": 128,
   "tps": 18.87,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-26",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18164236-mistral314b-Q4_K - Medium",
   "model": "mistral3 14B",
   "notes": "threads=99; config=b10636",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1043.96,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "Arc Pro B70; mistral3 14B Q4_K - Medium | pp512 | 1043.96 \u00b1 2.12; mistral3 14B Q4_K - Medium | tg128 | 32.33 \u00b1 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18164236",
   "tg_tokens": 128,
   "tps": 32.33,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-26",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18164236-mistral314b-Q5_K - Medium-2",
   "model": "mistral3 14B",
   "notes": "threads=99; config=b10636",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1243.84,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "Arc Pro B70; mistral3 14B Q5_K - Medium | pp512 | 1243.84 \u00b1 11.72; mistral3 14B Q5_K - Medium | tg128 | 30.08 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18164236",
   "tg_tokens": 128,
   "tps": 30.08,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-26",
   "flags": [
    "contradiction"
   ],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18164236-mistral314b-Q6_K-2",
   "model": "mistral3 14B",
   "notes": "threads=99; config=b10636",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1194.97,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "Arc Pro B70; mistral3 14B Q6_K | pp512 | 1194.97 \u00b1 10.06; mistral3 14B Q6_K | tg128 | 24.07 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18164236",
   "tg_tokens": 128,
   "tps": 24.07,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-26",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18164236-mistral314b-Q4_0-2",
   "model": "mistral3 14B",
   "notes": "threads=99; config=b10636",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1034.09,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B70; mistral3 14B Q4_0 | pp512 | 1034.09 \u00b1 0.69; mistral3 14B Q4_0 | tg128 | 34.72 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18164236",
   "tg_tokens": 128,
   "tps": 34.72,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-26",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18164236-mistral314b-Q4_0-3",
   "model": "mistral3 14B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10636",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 423.46,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B70; mistral3 14B Q4_0 | pp512 | 423.46 \u00b1 0.82; mistral3 14B Q4_0 | tg128 | 35.58 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18164236",
   "tg_tokens": 128,
   "tps": 35.58,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-27",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18174698-qwen3527b-IQ3_XXS - 3.0625 bpw",
   "model": "qwen35 27B",
   "notes": "threads=6; config=b10644",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 445.32,
   "provenance": "community",
   "quant": "IQ3_XXS - 3.0625 bpw",
   "quote": "Arc Pro B70; qwen35 27B IQ3_XXS - 3.0625 bpw | pp512 | 445.32 \u00b1 0.53; qwen35 27B IQ3_XXS - 3.0625 bpw | tg128 | 10.28 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18174698",
   "tg_tokens": 128,
   "tps": 10.28,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-27",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18174698-qwen3527b-IQ3_XXS - 3.0625 bpw-2",
   "model": "qwen35 27B",
   "notes": "threads=6; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10644",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 436.31,
   "provenance": "community",
   "quant": "IQ3_XXS - 3.0625 bpw",
   "quote": "Arc Pro B70; qwen35 27B IQ3_XXS - 3.0625 bpw | pp512 | 436.31 \u00b1 0.21; qwen35 27B IQ3_XXS - 3.0625 bpw | tg128 | 10.23 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18174698",
   "tg_tokens": 128,
   "tps": 10.23,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-27",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18174698-qwen3527b-Q4_0",
   "model": "qwen35 27B",
   "notes": "threads=6; config=b10644",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 812.57,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B70; qwen35 27B Q4_0 | pp512 | 812.57 \u00b1 2.08; qwen35 27B Q4_0 | tg128 | 26.89 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18174698",
   "tg_tokens": 128,
   "tps": 26.89,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-27",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18174698-qwen3527b-Q4_0-2",
   "model": "qwen35 27B",
   "notes": "threads=6; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10644",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 354.67,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B70; qwen35 27B Q4_0 | pp512 | 354.67 \u00b1 1.02; qwen35 27B Q4_0 | tg128 | 26.77 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18174698",
   "tg_tokens": 128,
   "tps": 26.77,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-27",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18174698-qwen3527b-Q4_K - Medium",
   "model": "qwen35 27B",
   "notes": "threads=6; config=b10644",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 867.21,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "Arc Pro B70; qwen35 27B Q4_K - Medium | pp512 | 867.21 \u00b1 1.71; qwen35 27B Q4_K - Medium | tg128 | 24.97 \u00b1 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18174698",
   "tg_tokens": 128,
   "tps": 24.97,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-27",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18174698-qwen3527b-Q4_K - Medium-2",
   "model": "qwen35 27B",
   "notes": "threads=6; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10644",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 756.78,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "Arc Pro B70; qwen35 27B Q4_K - Medium | pp512 | 756.78 \u00b1 38.06; qwen35 27B Q4_K - Medium | tg128 | 24.86 \u00b1 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18174698",
   "tg_tokens": 128,
   "tps": 24.86,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-27",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18174698-qwen3527b-Q5_K - Medium",
   "model": "qwen35 27B",
   "notes": "threads=6; config=b10644",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1031.14,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "Arc Pro B70; qwen35 27B Q5_K - Medium | pp512 | 1031.14 \u00b1 4.23; qwen35 27B Q5_K - Medium | tg128 | 22.24 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18174698",
   "tg_tokens": 128,
   "tps": 22.24,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-27",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18174698-qwen3527b-Q5_K - Medium-2",
   "model": "qwen35 27B",
   "notes": "threads=6; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10644",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 833.6,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "Arc Pro B70; qwen35 27B Q5_K - Medium | pp512 | 833.60 \u00b1 0.72; qwen35 27B Q5_K - Medium | tg128 | 22.11 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18174698",
   "tg_tokens": 128,
   "tps": 22.11,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-27",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18174698-qwen3527b-Q6_K",
   "model": "qwen35 27B",
   "notes": "threads=6; config=b10644",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 984.56,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "Arc Pro B70; qwen35 27B Q6_K | pp512 | 984.56 \u00b1 5.09; qwen35 27B Q6_K | tg128 | 20.08 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18174698",
   "tg_tokens": 128,
   "tps": 20.08,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-27",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18174698-qwen3527b-Q6_K-2",
   "model": "qwen35 27B",
   "notes": "threads=6; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10644",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 715.19,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "Arc Pro B70; qwen35 27B Q6_K | pp512 | 715.19 \u00b1 1.40; qwen35 27B Q6_K | tg128 | 19.93 \u00b1 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18174698",
   "tg_tokens": 128,
   "tps": 19.93,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-27",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18174698-qwen3527b-Q4_0-3",
   "model": "qwen35 27B",
   "notes": "threads=6; config=b10644",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 815.63,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B70; qwen35 27B Q4_0 | pp512 | 815.63 \u00b1 1.66; qwen35 27B Q4_0 | tg128 | 27.59 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18174698",
   "tg_tokens": 128,
   "tps": 27.59,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-27",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18174698-qwen3527b-Q4_0-4",
   "model": "qwen35 27B",
   "notes": "threads=6; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10644",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 358.96,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B70; qwen35 27B Q4_0 | pp512 | 358.96 \u00b1 1.03; qwen35 27B Q4_0 | tg128 | 27.40 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18174698",
   "tg_tokens": 128,
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-27",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18174698-qwen3527b-Q5_K - Medium-3",
   "model": "qwen35 27B",
   "notes": "threads=6; config=b10644",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 978.1,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "Arc Pro B70; qwen35 27B Q5_K - Medium | pp512 | 978.10 \u00b1 6.44; qwen35 27B Q5_K - Medium | tg128 | 22.31 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18174698",
   "tg_tokens": 128,
   "tps": 22.31,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-27",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18174698-qwen3527b-Q5_K - Medium-4",
   "model": "qwen35 27B",
   "notes": "threads=6; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10644",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 843.67,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "Arc Pro B70; qwen35 27B Q5_K - Medium | pp512 | 843.67 \u00b1 2.55; qwen35 27B Q5_K - Medium | tg128 | 22.24 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18174698",
   "tg_tokens": 128,
   "tps": 22.24,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-27",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18174698-qwen3527b-Q6_K-3",
   "model": "qwen35 27B",
   "notes": "threads=6; config=b10644",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 973.08,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "Arc Pro B70; qwen35 27B Q6_K | pp512 | 973.08 \u00b1 5.22; qwen35 27B Q6_K | tg128 | 21.21 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18174698",
   "tg_tokens": 128,
   "tps": 21.21,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-27",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18174698-qwen3527b-Q6_K-4",
   "model": "qwen35 27B",
   "notes": "threads=6; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10644",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 789.89,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "Arc Pro B70; qwen35 27B Q6_K | pp512 | 789.89 \u00b1 2.14; qwen35 27B Q6_K | tg128 | 21.15 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18174698",
   "tg_tokens": 128,
   "tps": 21.15,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18195166-qwen3527b-Q4_K - Medium",
   "model": "qwen35 27B",
   "notes": "threads=-1; config=b10644",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 875.38,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "Arc Pro B70; qwen35 27B Q4_K - Medium | pp512 | 875.38 \u00b1 1.83; qwen35 27B Q4_K - Medium | tg128 | 22.92 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195166",
   "tg_tokens": 128,
   "tps": 22.92,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18195166-qwen3527b-Q4_K - Medium-2",
   "model": "qwen35 27B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10644",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 853.39,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "Arc Pro B70; qwen35 27B Q4_K - Medium | pp512 | 853.39 \u00b1 0.43; qwen35 27B Q4_K - Medium | tg128 | 22.86 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195166",
   "tg_tokens": 128,
   "tps": 22.86,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18195166-qwen3527b-Q5_K - Medium",
   "model": "qwen35 27B",
   "notes": "threads=-1; config=b10644",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1024.19,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "Arc Pro B70; qwen35 27B Q5_K - Medium | pp512 | 1024.19 \u00b1 2.41; qwen35 27B Q5_K - Medium | tg128 | 23.22 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195166",
   "tg_tokens": 128,
   "tps": 23.22,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18195166-qwen3527b-Q5_K - Medium-2",
   "model": "qwen35 27B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10644",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 914.54,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "Arc Pro B70; qwen35 27B Q5_K - Medium | pp512 | 914.54 \u00b1 0.78; qwen35 27B Q5_K - Medium | tg128 | 23.17 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195166",
   "tg_tokens": 128,
   "tps": 23.17,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18195166-qwen3527b-Q6_K",
   "model": "qwen35 27B",
   "notes": "threads=-1; fa=0 ngl=-1; config=b10644",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1035.09,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "Arc Pro B70; qwen35 27B Q6_K | pp512 | 1035.09 \u00b1 3.09; qwen35 27B Q6_K | tg128 | 18.91 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195166",
   "tg_tokens": 128,
   "tps": 18.91,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18195166-qwen3527b-Q6_K-2",
   "model": "qwen35 27B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=-1; config=b10644",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 688.23,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "Arc Pro B70; qwen35 27B Q6_K | pp512 | 688.23 \u00b1 0.50; qwen35 27B Q6_K | tg128 | 18.86 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195166",
   "tg_tokens": 128,
   "tps": 18.86,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-27",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18176229-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=-1; config=b10650",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1869.16,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B60; llama 7B Q4_0 | pp512 | 1869.16 \u00b1 9.67; llama 7B Q4_0 | tg128 | 80.03 \u00b1 0.12",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18176229",
   "tg_tokens": 128,
   "tps": 80.03,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-27",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18176229-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10650",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 928.91,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B60; llama 7B Q4_0 | pp512 | 928.91 \u00b1 0.39; llama 7B Q4_0 | tg128 | 84.44 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18176229",
   "tg_tokens": 128,
   "tps": 84.44,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-27",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18176229-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=-1; config=b10650",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 570.18,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B60; llama 7B Q4_0 | pp512 | 570.18 \u00b1 0.60; llama 7B Q4_0 | tg128 | 80.03 \u00b1 0.12",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18176229",
   "tg_tokens": 128,
   "tps": 80.03,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-27",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18176229-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10650",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 664.82,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B60; llama 7B Q4_0 | pp512 | 664.82 \u00b1 0.13; llama 7B Q4_0 | tg128 | 84.12 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18176229",
   "tg_tokens": 128,
   "tps": 84.12,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-qwen35moe35ba3b-Q4_K - Medium",
   "model": "qwen35moe 35B.A3B",
   "notes": "threads=-1; config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 954.41,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "Arc Pro B60; qwen35moe 35B.A3B Q4_K - Medium | pp512 | 954.41 \u00b1 17.24; qwen35moe 35B.A3B Q4_K - Medium | tg128 | 81.67 \u00b1 0.10",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 81.67,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-qwen35moe35ba3b-Q4_K - Medium-2",
   "model": "qwen35moe 35B.A3B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 965.33,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "Arc Pro B60; qwen35moe 35B.A3B Q4_K - Medium | pp512 | 965.33 \u00b1 7.23; qwen35moe 35B.A3B Q4_K - Medium | tg128 | 81.84 \u00b1 0.15",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 81.84,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-qwen35moe35ba3b-Q4_K - Medium-3",
   "model": "qwen35moe 35B.A3B",
   "notes": "threads=-1; config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 776.14,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "Arc Pro B60; qwen35moe 35B.A3B Q4_K - Medium | pp512 | 776.14 \u00b1 6.89; qwen35moe 35B.A3B Q4_K - Medium | tg128 | 81.68 \u00b1 0.11",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 81.68,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-qwen35moe35ba3b-Q4_K - Medium-4",
   "model": "qwen35moe 35B.A3B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 774.92,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "Arc Pro B60; qwen35moe 35B.A3B Q4_K - Medium | pp512 | 774.92 \u00b1 4.85; qwen35moe 35B.A3B Q4_K - Medium | tg128 | 81.63 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 81.63,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-qwen359b-Q6_K",
   "model": "qwen35 9B",
   "notes": "threads=-1; config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1958.82,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "Arc Pro B60; qwen35 9B Q6_K | pp512 | 1958.82 \u00b1 1.32; qwen35 9B Q6_K | tg128 | 51.10 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 51.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-qwen359b-Q6_K-2",
   "model": "qwen35 9B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1630.46,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "Arc Pro B60; qwen35 9B Q6_K | pp512 | 1630.46 \u00b1 1.61; qwen35 9B Q6_K | tg128 | 51.21 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 51.21,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-qwen359b-Q6_K-3",
   "model": "qwen35 9B",
   "notes": "threads=-1; config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 677.24,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "Arc Pro B60; qwen35 9B Q6_K | pp512 | 677.24 \u00b1 0.26; qwen35 9B Q6_K | tg128 | 51.14 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 51.14,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-qwen359b-Q6_K-4",
   "model": "qwen35 9B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 647.03,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "Arc Pro B60; qwen35 9B Q6_K | pp512 | 647.03 \u00b1 0.15; qwen35 9B Q6_K | tg128 | 51.31 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 51.31,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-exaone432b-IQ4_XS - 4.25 bpw",
   "model": "exaone4 32B",
   "notes": "threads=-1; config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 362.96,
   "provenance": "community",
   "quant": "IQ4_XS - 4.25 bpw",
   "quote": "Arc Pro B60; exaone4 32B IQ4_XS - 4.25 bpw | pp512 | 362.96 \u00b1 1.10; exaone4 32B IQ4_XS - 4.25 bpw | tg128 | 11.10 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 11.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-exaone432b-IQ4_XS - 4.25 bpw-2",
   "model": "exaone4 32B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 382.48,
   "provenance": "community",
   "quant": "IQ4_XS - 4.25 bpw",
   "quote": "Arc Pro B60; exaone4 32B IQ4_XS - 4.25 bpw | pp512 | 382.48 \u00b1 0.10; exaone4 32B IQ4_XS - 4.25 bpw | tg128 | 11.26 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 11.26,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-exaone432b-IQ4_XS - 4.25 bpw-3",
   "model": "exaone4 32B",
   "notes": "threads=-1; config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 140.3,
   "provenance": "community",
   "quant": "IQ4_XS - 4.25 bpw",
   "quote": "Arc Pro B60; exaone4 32B IQ4_XS - 4.25 bpw | pp512 | 140.30 \u00b1 0.05; exaone4 32B IQ4_XS - 4.25 bpw | tg128 | 11.14 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 11.14,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-exaone432b-IQ4_XS - 4.25 bpw-4",
   "model": "exaone4 32B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 143.11,
   "provenance": "community",
   "quant": "IQ4_XS - 4.25 bpw",
   "quote": "Arc Pro B60; exaone4 32B IQ4_XS - 4.25 bpw | pp512 | 143.11 \u00b1 0.02; exaone4 32B IQ4_XS - 4.25 bpw | tg128 | 11.30 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 11.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-exaone432b-Q4_K - Medium",
   "model": "exaone4 32B",
   "notes": "threads=-1; config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 546.71,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "Arc Pro B60; exaone4 32B Q4_K - Medium | pp512 | 546.71 \u00b1 1.63; exaone4 32B Q4_K - Medium | tg128 | 19.03 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 19.03,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-exaone432b-Q4_K - Medium-2",
   "model": "exaone4 32B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 562.66,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "Arc Pro B60; exaone4 32B Q4_K - Medium | pp512 | 562.66 \u00b1 0.12; exaone4 32B Q4_K - Medium | tg128 | 19.40 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 19.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-exaone432b-Q4_K - Medium-3",
   "model": "exaone4 32B",
   "notes": "threads=-1; config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 163.21,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "Arc Pro B60; exaone4 32B Q4_K - Medium | pp512 | 163.21 \u00b1 0.10; exaone4 32B Q4_K - Medium | tg128 | 19.02 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 19.02,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-exaone432b-Q4_K - Medium-4",
   "model": "exaone4 32B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 165.62,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "Arc Pro B60; exaone4 32B Q4_K - Medium | pp512 | 165.62 \u00b1 0.02; exaone4 32B Q4_K - Medium | tg128 | 19.37 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 19.37,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-gemma4b-Q4_0",
   "model": "gemma4 ?B",
   "notes": "threads=-1; config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1166.0,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B60; gemma4 ?B Q4_0 | pp512 | 1166.00 \u00b1 0.19; gemma4 ?B Q4_0 | tg128 | 43.51 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 43.51,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-gemma4b-Q4_0-2",
   "model": "gemma4 ?B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 543.74,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B60; gemma4 ?B Q4_0 | pp512 | 543.74 \u00b1 0.08; gemma4 ?B Q4_0 | tg128 | 42.55 \u00b1 0.10",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 42.55,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-gemma4b-Q4_0-3",
   "model": "gemma4 ?B",
   "notes": "threads=-1; config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 335.89,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B60; gemma4 ?B Q4_0 | pp512 | 335.89 \u00b1 0.03; gemma4 ?B Q4_0 | tg128 | 43.60 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 43.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-gemma4b-Q4_0-4",
   "model": "gemma4 ?B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 384.54,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B60; gemma4 ?B Q4_0 | pp512 | 384.54 \u00b1 0.03; gemma4 ?B Q4_0 | tg128 | 42.39 \u00b1 0.10",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 42.39,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-gemma426ba4b-Q4_0",
   "model": "gemma4 26B.A4B",
   "notes": "threads=-1; config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1070.72,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B60; gemma4 26B.A4B Q4_0 | pp512 | 1070.72 \u00b1 11.16; gemma4 26B.A4B Q4_0 | tg128 | 45.20 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 45.2,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-gemma426ba4b-Q4_0-2",
   "model": "gemma4 26B.A4B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 960.33,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B60; gemma4 26B.A4B Q4_0 | pp512 | 960.33 \u00b1 5.55; gemma4 26B.A4B Q4_0 | tg128 | 45.19 \u00b1 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 45.19,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-gemma426ba4b-Q4_0-3",
   "model": "gemma4 26B.A4B",
   "notes": "threads=-1; config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 599.37,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B60; gemma4 26B.A4B Q4_0 | pp512 | 599.37 \u00b1 2.94; gemma4 26B.A4B Q4_0 | tg128 | 45.23 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 45.23,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-gemma426ba4b-Q4_0-4",
   "model": "gemma4 26B.A4B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 612.37,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B60; gemma4 26B.A4B Q4_0 | pp512 | 612.37 \u00b1 4.02; gemma4 26B.A4B Q4_0 | tg128 | 45.21 \u00b1 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 45.21,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-gemma431b-Q4_0",
   "model": "Gemma4 31B",
   "notes": "threads=-1; config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 472.95,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B60; gemma4 31B Q4_0 | pp512 | 472.95 \u00b1 4.60; gemma4 31B Q4_0 | tg128 | 18.94 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 18.94,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-gemma431b-Q4_0-2",
   "model": "Gemma4 31B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 210.16,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B60; gemma4 31B Q4_0 | pp512 | 210.16 \u00b1 0.04; gemma4 31B Q4_0 | tg128 | 19.63 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 19.63,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-gemma431b-Q4_0-3",
   "model": "Gemma4 31B",
   "notes": "threads=-1; config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 127.18,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B60; gemma4 31B Q4_0 | pp512 | 127.18 \u00b1 0.27; gemma4 31B Q4_0 | tg128 | 18.92 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 18.92,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-gemma431b-Q4_0-4",
   "model": "Gemma4 31B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 146.16,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B60; gemma4 31B Q4_0 | pp512 | 146.16 \u00b1 0.06; gemma4 31B Q4_0 | tg128 | 19.64 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 19.64,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-graniteb-Q4_K - Medium",
   "model": "granite ?B",
   "notes": "threads=-1; config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 580.79,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "Arc Pro B60; granite ?B Q4_K - Medium | pp512 | 580.79 \u00b1 36.96; granite ?B Q4_K - Medium | tg128 | 20.21 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 20.21,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-graniteb-Q4_K - Medium-2",
   "model": "granite ?B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 600.37,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "Arc Pro B60; granite ?B Q4_K - Medium | pp512 | 600.37 \u00b1 0.16; granite ?B Q4_K - Medium | tg128 | 21.01 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 21.01,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-graniteb-Q4_K - Medium-3",
   "model": "granite ?B",
   "notes": "threads=-1; config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 178.83,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "Arc Pro B60; granite ?B Q4_K - Medium | pp512 | 178.83 \u00b1 0.50; granite ?B Q4_K - Medium | tg128 | 20.21 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 20.21,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-graniteb-Q4_K - Medium-4",
   "model": "granite ?B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 180.7,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "Arc Pro B60; granite ?B Q4_K - Medium | pp512 | 180.70 \u00b1 0.05; granite ?B Q4_K - Medium | tg128 | 20.98 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 20.98,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-granite3b-Q6_K",
   "model": "granite 3B",
   "notes": "threads=-1; config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1907.81,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "Arc Pro B60; granite 3B Q6_K | pp512 | 1907.81 \u00b1 1.35; granite 3B Q6_K | tg128 | 49.60 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-granite3b-Q6_K-2",
   "model": "granite 3B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1717.91,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "Arc Pro B60; granite 3B Q6_K | pp512 | 1717.91 \u00b1 0.62; granite 3B Q6_K | tg128 | 52.46 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 52.46,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-granite3b-Q6_K-3",
   "model": "granite 3B",
   "notes": "threads=-1; config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 604.43,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "Arc Pro B60; granite 3B Q6_K | pp512 | 604.43 \u00b1 0.20; granite 3B Q6_K | tg128 | 49.60 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-granite3b-Q6_K-4",
   "model": "granite 3B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 597.19,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "Arc Pro B60; granite 3B Q6_K | pp512 | 597.19 \u00b1 0.09; granite 3B Q6_K | tg128 | 52.36 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 52.36,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-laguna30ba3b-Q4_K - Medium",
   "model": "laguna 30B.A3B",
   "notes": "threads=-1; config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 811.97,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "Arc Pro B60; laguna 30B.A3B Q4_K - Medium | pp512 | 811.97 \u00b1 25.74; laguna 30B.A3B Q4_K - Medium | tg128 | 60.12 \u00b1 1.44",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 60.12,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-laguna30ba3b-Q4_K - Medium-2",
   "model": "laguna 30B.A3B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 888.17,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "Arc Pro B60; laguna 30B.A3B Q4_K - Medium | pp512 | 888.17 \u00b1 30.62; laguna 30B.A3B Q4_K - Medium | tg128 | 91.50 \u00b1 0.16",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 91.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-laguna30ba3b-Q4_K - Medium-3",
   "model": "laguna 30B.A3B",
   "notes": "threads=-1; config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 690.0,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "Arc Pro B60; laguna 30B.A3B Q4_K - Medium | pp512 | 690.00 \u00b1 4.41; laguna 30B.A3B Q4_K - Medium | tg128 | 61.06 \u00b1 1.36",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 61.06,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-laguna30ba3b-Q4_K - Medium-4",
   "model": "laguna 30B.A3B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 745.03,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "Arc Pro B60; laguna 30B.A3B Q4_K - Medium | pp512 | 745.03 \u00b1 3.70; laguna 30B.A3B Q4_K - Medium | tg128 | 91.28 \u00b1 0.14",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 91.28,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-museglimmer30b-Q4_K - Medium",
   "model": "muse-glimmer 30B",
   "notes": "threads=-1; config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 649.88,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "Arc Pro B60; muse-glimmer 30B Q4_K - Medium | pp512 | 649.88 \u00b1 4.80; muse-glimmer 30B Q4_K - Medium | tg128 | 23.94 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 23.94,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-museglimmer30b-Q4_K - Medium-2",
   "model": "muse-glimmer 30B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 686.91,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "Arc Pro B60; muse-glimmer 30B Q4_K - Medium | pp512 | 686.91 \u00b1 0.35; muse-glimmer 30B Q4_K - Medium | tg128 | 24.34 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 24.34,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-museglimmer30b-Q4_K - Medium-3",
   "model": "muse-glimmer 30B",
   "notes": "threads=-1; config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 202.45,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "Arc Pro B60; muse-glimmer 30B Q4_K - Medium | pp512 | 202.45 \u00b1 0.50; muse-glimmer 30B Q4_K - Medium | tg128 | 23.92 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 23.92,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-museglimmer30b-Q4_K - Medium-4",
   "model": "muse-glimmer 30B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 206.45,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "Arc Pro B60; muse-glimmer 30B Q4_K - Medium | pp512 | 206.45 \u00b1 0.03; muse-glimmer 30B Q4_K - Medium | tg128 | 24.33 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 24.33,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-cohere2moe30ba3b-IQ4_NL - 4.5 bpw",
   "model": "cohere2moe 30B.A3B",
   "notes": "threads=-1; config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 605.7,
   "provenance": "community",
   "quant": "IQ4_NL - 4.5 bpw",
   "quote": "Arc Pro B60; cohere2moe 30B.A3B IQ4_NL - 4.5 bpw | pp512 | 605.70 \u00b1 6.50; cohere2moe 30B.A3B IQ4_NL - 4.5 bpw | tg128 | 30.51 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 30.51,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-cohere2moe30ba3b-IQ4_NL - 4.5 bpw-2",
   "model": "cohere2moe 30B.A3B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 616.59,
   "provenance": "community",
   "quant": "IQ4_NL - 4.5 bpw",
   "quote": "Arc Pro B60; cohere2moe 30B.A3B IQ4_NL - 4.5 bpw | pp512 | 616.59 \u00b1 3.82; cohere2moe 30B.A3B IQ4_NL - 4.5 bpw | tg128 | 31.31 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 31.31,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-cohere2moe30ba3b-IQ4_NL - 4.5 bpw-3",
   "model": "cohere2moe 30B.A3B",
   "notes": "threads=-1; config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 520.1,
   "provenance": "community",
   "quant": "IQ4_NL - 4.5 bpw",
   "quote": "Arc Pro B60; cohere2moe 30B.A3B IQ4_NL - 4.5 bpw | pp512 | 520.10 \u00b1 6.93; cohere2moe 30B.A3B IQ4_NL - 4.5 bpw | tg128 | 30.54 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 30.54,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-cohere2moe30ba3b-IQ4_NL - 4.5 bpw-4",
   "model": "cohere2moe 30B.A3B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 531.97,
   "provenance": "community",
   "quant": "IQ4_NL - 4.5 bpw",
   "quote": "Arc Pro B60; cohere2moe 30B.A3B IQ4_NL - 4.5 bpw | pp512 | 531.97 \u00b1 1.50; cohere2moe 30B.A3B IQ4_NL - 4.5 bpw | tg128 | 31.28 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 31.28,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-cohere2moe30ba3b-Q4_K - Medium",
   "model": "cohere2moe 30B.A3B",
   "notes": "threads=-1; config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1005.12,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "Arc Pro B60; cohere2moe 30B.A3B Q4_K - Medium | pp512 | 1005.12 \u00b1 21.49; cohere2moe 30B.A3B Q4_K - Medium | tg128 | 66.58 \u00b1 1.53",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 66.58,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-cohere2moe30ba3b-Q4_K - Medium-2",
   "model": "cohere2moe 30B.A3B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1048.0,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "Arc Pro B60; cohere2moe 30B.A3B Q4_K - Medium | pp512 | 1048.00 \u00b1 6.18; cohere2moe 30B.A3B Q4_K - Medium | tg128 | 70.86 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 70.86,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-cohere2moe30ba3b-Q4_K - Medium-3",
   "model": "cohere2moe 30B.A3B",
   "notes": "threads=-1; config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 779.19,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "Arc Pro B60; cohere2moe 30B.A3B Q4_K - Medium | pp512 | 779.19 \u00b1 6.76; cohere2moe 30B.A3B Q4_K - Medium | tg128 | 67.51 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 67.51,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-cohere2moe30ba3b-Q4_K - Medium-4",
   "model": "cohere2moe 30B.A3B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 795.24,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "Arc Pro B60; cohere2moe 30B.A3B Q4_K - Medium | pp512 | 795.24 \u00b1 2.43; cohere2moe 30B.A3B Q4_K - Medium | tg128 | 70.59 \u00b1 0.11",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 70.59,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-qwen3527b-Q4_K - Medium",
   "model": "qwen35 27B",
   "notes": "threads=-1; config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 508.73,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "Arc Pro B60; qwen35 27B Q4_K - Medium | pp512 | 508.73 \u00b1 3.47; qwen35 27B Q4_K - Medium | tg128 | 14.92 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 14.92,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-qwen3527b-Q4_K - Medium-2",
   "model": "qwen35 27B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 507.51,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "Arc Pro B60; qwen35 27B Q4_K - Medium | pp512 | 507.51 \u00b1 0.26; qwen35 27B Q4_K - Medium | tg128 | 14.92 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 14.92,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-qwen3527b-Q4_K - Medium-3",
   "model": "qwen35 27B",
   "notes": "threads=-1; config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 185.63,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "Arc Pro B60; qwen35 27B Q4_K - Medium | pp512 | 185.63 \u00b1 0.43; qwen35 27B Q4_K - Medium | tg128 | 14.92 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 14.92,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-qwen3527b-Q4_K - Medium-4",
   "model": "qwen35 27B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 186.28,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "Arc Pro B60; qwen35 27B Q4_K - Medium | pp512 | 186.28 \u00b1 0.03; qwen35 27B Q4_K - Medium | tg128 | 14.92 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 14.92,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-qwen3527b-Q4_K - Small",
   "model": "qwen35 27B",
   "notes": "threads=-1; config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 456.99,
   "provenance": "community",
   "quant": "Q4_K - Small",
   "quote": "Arc Pro B60; qwen35 27B Q4_K - Small | pp512 | 456.99 \u00b1 3.26; qwen35 27B Q4_K - Small | tg128 | 13.25 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 13.25,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-qwen3527b-Q4_K - Small-2",
   "model": "qwen35 27B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 459.73,
   "provenance": "community",
   "quant": "Q4_K - Small",
   "quote": "Arc Pro B60; qwen35 27B Q4_K - Small | pp512 | 459.73 \u00b1 2.47; qwen35 27B Q4_K - Small | tg128 | 13.26 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 13.26,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-qwen3527b-Q4_K - Small-3",
   "model": "qwen35 27B",
   "notes": "threads=-1; config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 179.83,
   "provenance": "community",
   "quant": "Q4_K - Small",
   "quote": "Arc Pro B60; qwen35 27B Q4_K - Small | pp512 | 179.83 \u00b1 0.13; qwen35 27B Q4_K - Small | tg128 | 13.25 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 13.25,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-08-29",
   "flags": [],
   "hardware": "Arc Pro B60",
   "id": "lc-disc-23313-c18195362-qwen3527b-Q4_K - Small-4",
   "model": "qwen35 27B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10665",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 180.7,
   "provenance": "community",
   "quant": "Q4_K - Small",
   "quote": "Arc Pro B60; qwen35 27B Q4_K - Small | pp512 | 180.70 \u00b1 0.06; qwen35 27B Q4_K - Small | tg128 | 13.25 \u00b1 0.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18195362",
   "tg_tokens": 128,
   "tps": 13.25,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-31",
   "flags": [],
   "hardware": "Arc A770",
   "id": "lc-disc-23313-c18216776-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99; config=b10714",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 822.07,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A770; llama 7B Q4_0 | pp512 | 822.07 + 3.10; llama 7B Q4_0 | tg128 | 43.18 + 0.07",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18216776",
   "tg_tokens": 128,
   "tps": 43.18,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-31",
   "flags": [],
   "hardware": "Arc A770",
   "id": "lc-disc-23313-c18216776-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b10714",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 697.13,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc A770; llama 7B Q4_0 | pp512 | 697.13 + 1.56; llama 7B Q4_0 | tg128 | 41.48 + 0.41",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18216776",
   "tg_tokens": 128,
   "tps": 41.48,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-01",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18229981-qwen3527b-Q6_K",
   "model": "qwen35 27B",
   "notes": "threads=-1",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 979.33,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "Arc Pro B70; qwen35 27B Q6_K | pp512 | 979.33 \u00b1 4.64; qwen35 27B Q6_K | tg128 | 21.25 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18229981",
   "tg_tokens": 128,
   "tps": 21.25,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-01",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18229981-qwen3527b-Q6_K-2",
   "model": "qwen35 27B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 793.28,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "Arc Pro B70; qwen35 27B Q6_K | pp512 | 793.28 \u00b1 1.67; qwen35 27B Q6_K | tg128 | 21.15 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18229981",
   "tg_tokens": 128,
   "tps": 21.15,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-01",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18229981-qwen3527b-Q5_K - Medium",
   "model": "qwen35 27B",
   "notes": "threads=-1",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 970.49,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "Arc Pro B70; qwen35 27B Q5_K - Medium | pp512 | 970.49 \u00b1 3.90; qwen35 27B Q5_K - Medium | tg128 | 22.30 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18229981",
   "tg_tokens": 128,
   "tps": 22.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-01",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18229981-qwen3527b-Q5_K - Medium-2",
   "model": "qwen35 27B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 840.86,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "Arc Pro B70; qwen35 27B Q5_K - Medium | pp512 | 840.86 \u00b1 4.11; qwen35 27B Q5_K - Medium | tg128 | 22.22 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18229981",
   "tg_tokens": 128,
   "tps": 22.22,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-01",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18229981-qwen3527b-Q5_K - Medium-3",
   "model": "qwen35 27B",
   "notes": "threads=-1",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 858.88,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "Arc Pro B70; qwen35 27B Q5_K - Medium | pp512 | 858.88 \u00b1 22.25; qwen35 27B Q5_K - Medium | tg128 | 22.22 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18229981",
   "tg_tokens": 128,
   "tps": 22.22,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-01",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18229981-qwen3527b-Q5_K - Medium-4",
   "model": "qwen35 27B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 815.38,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "Arc Pro B70; qwen35 27B Q5_K - Medium | pp512 | 815.38 \u00b1 6.21; qwen35 27B Q5_K - Medium | tg128 | 22.25 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18229981",
   "tg_tokens": 128,
   "tps": 22.25,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-01",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18229981-qwen3527b-Q5_K - Small",
   "model": "qwen35 27B",
   "notes": "threads=-1",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 914.77,
   "provenance": "community",
   "quant": "Q5_K - Small",
   "quote": "Arc Pro B70; qwen35 27B Q5_K - Small | pp512 | 914.77 \u00b1 2.72; qwen35 27B Q5_K - Small | tg128 | 22.40 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18229981",
   "tg_tokens": 128,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-01",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18229981-qwen3527b-Q5_K - Small-2",
   "model": "qwen35 27B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 845.78,
   "provenance": "community",
   "quant": "Q5_K - Small",
   "quote": "Arc Pro B70; qwen35 27B Q5_K - Small | pp512 | 845.78 \u00b1 1.73; qwen35 27B Q5_K - Small | tg128 | 22.33 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18229981",
   "tg_tokens": 128,
   "tps": 22.33,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-01",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18229981-qwen3527b-Q4_K - Medium",
   "model": "qwen35 27B",
   "notes": "threads=-1",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 811.78,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "Arc Pro B70; qwen35 27B Q4_K - Medium | pp512 | 811.78 \u00b1 1.40; qwen35 27B Q4_K - Medium | tg128 | 21.39 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18229981",
   "tg_tokens": 128,
   "tps": 21.39,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-01",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18229981-qwen3527b-Q4_K - Medium-2",
   "model": "qwen35 27B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 784.04,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "Arc Pro B70; qwen35 27B Q4_K - Medium | pp512 | 784.04 \u00b1 1.72; qwen35 27B Q4_K - Medium | tg128 | 21.33 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18229981",
   "tg_tokens": 128,
   "tps": 21.33,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-01",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18229981-qwen3527b-Q4_0",
   "model": "qwen35 27B",
   "notes": "threads=-1",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 813.56,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B70; qwen35 27B Q4_0 | pp512 | 813.56 \u00b1 1.92; qwen35 27B Q4_0 | tg128 | 27.48 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18229981",
   "tg_tokens": 128,
   "tps": 27.48,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-01",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18229981-qwen3527b-Q4_0-2",
   "model": "qwen35 27B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 355.72,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B70; qwen35 27B Q4_0 | pp512 | 355.72 \u00b1 0.43; qwen35 27B Q4_0 | tg128 | 27.37 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18229981",
   "tg_tokens": 128,
   "tps": 27.37,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-01",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18229981-qwen3527b-IQ4_XS - 4.25 bpw",
   "model": "qwen35 27B",
   "notes": "threads=-1",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 583.02,
   "provenance": "community",
   "quant": "IQ4_XS - 4.25 bpw",
   "quote": "Arc Pro B70; qwen35 27B IQ4_XS - 4.25 bpw | pp512 | 583.02 \u00b1 0.61; qwen35 27B IQ4_XS - 4.25 bpw | tg128 | 16.40 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18229981",
   "tg_tokens": 128,
   "tps": 16.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-01",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18229981-qwen3527b-IQ4_XS - 4.25 bpw-2",
   "model": "qwen35 27B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 586.82,
   "provenance": "community",
   "quant": "IQ4_XS - 4.25 bpw",
   "quote": "Arc Pro B70; qwen35 27B IQ4_XS - 4.25 bpw | pp512 | 586.82 \u00b1 0.83; qwen35 27B IQ4_XS - 4.25 bpw | tg128 | 16.34 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18229981",
   "tg_tokens": 128,
   "tps": 16.34,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-01",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18229981-qwen3527b-Q5_K - Medium-5",
   "model": "qwen35 27B",
   "notes": "threads=-1",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 948.86,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "Arc Pro B70; qwen35 27B Q5_K - Medium | pp512 | 948.86 \u00b1 4.32; qwen35 27B Q5_K - Medium | tg128 | 22.31 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18229981",
   "tg_tokens": 128,
   "tps": 22.31,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-01",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18229981-qwen3527b-Q5_K - Medium-6",
   "model": "qwen35 27B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 791.62,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "Arc Pro B70; qwen35 27B Q5_K - Medium | pp512 | 791.62 \u00b1 1.74; qwen35 27B Q5_K - Medium | tg128 | 22.17 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18229981",
   "tg_tokens": 128,
   "tps": 22.17,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-01",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18229981-qwen3527b-Q4_0-3",
   "model": "qwen35 27B",
   "notes": "threads=-1",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 792.89,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B70; qwen35 27B Q4_0 | pp512 | 792.89 \u00b1 2.42; qwen35 27B Q4_0 | tg128 | 27.53 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18229981",
   "tg_tokens": 128,
   "tps": 27.53,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-01",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18229981-qwen3527b-Q4_0-4",
   "model": "qwen35 27B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 353.01,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B70; qwen35 27B Q4_0 | pp512 | 353.01 \u00b1 1.91; qwen35 27B Q4_0 | tg128 | 27.36 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18229981",
   "tg_tokens": 128,
   "tps": 27.36,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-01",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18229981-qwen3527b-IQ4_XS - 4.25 bpw-3",
   "model": "qwen35 27B",
   "notes": "threads=-1",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 571.95,
   "provenance": "community",
   "quant": "IQ4_XS - 4.25 bpw",
   "quote": "Arc Pro B70; qwen35 27B IQ4_XS - 4.25 bpw | pp512 | 571.95 \u00b1 0.63; qwen35 27B IQ4_XS - 4.25 bpw | tg128 | 16.43 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18229981",
   "tg_tokens": 128,
   "tps": 16.43,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-01",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18229981-qwen3527b-IQ4_XS - 4.25 bpw-4",
   "model": "qwen35 27B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 560.78,
   "provenance": "community",
   "quant": "IQ4_XS - 4.25 bpw",
   "quote": "Arc Pro B70; qwen35 27B IQ4_XS - 4.25 bpw | pp512 | 560.78 \u00b1 0.38; qwen35 27B IQ4_XS - 4.25 bpw | tg128 | 16.40 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18229981",
   "tg_tokens": 128,
   "tps": 16.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-01",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18229981-qwen3527b-Q5_K - Medium-7",
   "model": "qwen35 27B",
   "notes": "threads=-1",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 311.0,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "Arc Pro B70; qwen35 27B Q5_K - Medium | pp512 | 311.00 \u00b1 0.43; qwen35 27B Q5_K - Medium | tg128 | 22.26 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18229981",
   "tg_tokens": 128,
   "tps": 22.26,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-01",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18229981-qwen3527b-Q5_K - Medium-8",
   "model": "qwen35 27B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 304.89,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "Arc Pro B70; qwen35 27B Q5_K - Medium | pp512 | 304.89 \u00b1 0.13; qwen35 27B Q5_K - Medium | tg128 | 22.17 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18229981",
   "tg_tokens": 128,
   "tps": 22.17,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-01",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18229981-qwen3527b-Q4_0-5",
   "model": "qwen35 27B",
   "notes": "threads=-1",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 239.5,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B70; qwen35 27B Q4_0 | pp512 | 239.50 \u00b1 2.02; qwen35 27B Q4_0 | tg128 | 27.07 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18229981",
   "tg_tokens": 128,
   "tps": 27.07,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-01",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18229981-qwen3527b-Q4_0-6",
   "model": "qwen35 27B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 270.79,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B70; qwen35 27B Q4_0 | pp512 | 270.79 \u00b1 0.47; qwen35 27B Q4_0 | tg128 | 27.05 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18229981",
   "tg_tokens": 128,
   "tps": 27.05,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-01",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18229981-qwen3527b-IQ4_XS - 4.25 bpw-5",
   "model": "qwen35 27B",
   "notes": "threads=-1",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 274.41,
   "provenance": "community",
   "quant": "IQ4_XS - 4.25 bpw",
   "quote": "Arc Pro B70; qwen35 27B IQ4_XS - 4.25 bpw | pp512 | 274.41 \u00b1 0.40; qwen35 27B IQ4_XS - 4.25 bpw | tg128 | 16.51 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18229981",
   "tg_tokens": 128,
   "tps": 16.51,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-01",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18229981-qwen3527b-IQ4_XS - 4.25 bpw-6",
   "model": "qwen35 27B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 273.2,
   "provenance": "community",
   "quant": "IQ4_XS - 4.25 bpw",
   "quote": "Arc Pro B70; qwen35 27B IQ4_XS - 4.25 bpw | pp512 | 273.20 \u00b1 0.13; qwen35 27B IQ4_XS - 4.25 bpw | tg128 | 16.42 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18229981",
   "tg_tokens": 128,
   "tps": 16.42,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-16",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18467211-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=-1",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3271.16,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B70; llama 7B Q4_0 | pp512 | 3271.16 \u00b1 11.59; llama 7B Q4_0 | tg128 | 111.13 \u00b1 0.11",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18467211",
   "tg_tokens": 128,
   "tps": 111.13,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-16",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18467211-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1324.76,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc Pro B70; llama 7B Q4_0 | pp512 | 1324.76 \u00b1 2.34; llama 7B Q4_0 | tg128 | 107.11 \u00b1 0.23",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18467211",
   "tg_tokens": 128,
   "tps": 107.11,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-13",
   "flags": [],
   "hardware": "Arc B580",
   "id": "lc-disc-23313-c17291611-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1954.15,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc B580; llama 7B Q4_0 | pp512 | 1954.15 \u00b1 5.44; llama 7B Q4_0 | tg128 | 83.67 \u00b1 0.14",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17291611",
   "tg_tokens": 128,
   "tps": 83.67,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-13",
   "flags": [],
   "hardware": "Arc B580",
   "id": "lc-disc-23313-c17291611-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 893.17,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc B580; llama 7B Q4_0 | pp512 | 893.17 \u00b1 0.77; llama 7B Q4_0 | tg128 | 88.13 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17291611",
   "tg_tokens": 128,
   "tps": 88.13,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-13",
   "flags": [],
   "hardware": "Arc B580",
   "id": "lc-disc-23313-c17291611-gemma412bq4km-unknown",
   "model": "gemma4 12B Q4_KM",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1320.02,
   "provenance": "community",
   "quant": "as-published",
   "quote": "Arc B580; gemma4 12B Q4_KM | pp512 | 1320.02 \u00b1 0.78; gemma4 12B Q4_KM | tg128 | 41.34 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17291611",
   "tg_tokens": 128,
   "tps": 41.34,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-13",
   "flags": [],
   "hardware": "Arc B580",
   "id": "lc-disc-23313-c17291611-gemma412bq4km-unknown-2",
   "model": "gemma4 12B Q4_KM",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1062.58,
   "provenance": "community",
   "quant": "as-published",
   "quote": "Arc B580; gemma4 12B Q4_KM | pp512 | 1062.58 \u00b1 3.35; gemma4 12B Q4_KM | tg128 | 39.96 \u00b1 0.09",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17291611",
   "tg_tokens": 128,
   "tps": 39.96,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-13",
   "flags": [],
   "hardware": "Arc B580",
   "id": "lc-disc-23313-c17291611-gemma4e4b-Q4_0",
   "model": "gemma4 E4B",
   "notes": "threads=99",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2763.42,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc B580; gemma4 E4B Q4_0 | pp512 | 2763.42 \u00b1 2.33; gemma4 E4B Q4_0 | tg128 | 82.30 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17291611",
   "tg_tokens": 128,
   "tps": 82.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-06-13",
   "flags": [],
   "hardware": "Arc B580",
   "id": "lc-disc-23313-c17291611-gemma4e4b-Q4_0-2",
   "model": "gemma4 E4B",
   "notes": "threads=99; run 2 of 2 in this comment (separate llama-bench run; build/flags differ)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1349.62,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Arc B580; gemma4 E4B Q4_0 | pp512 | 1349.62 \u00b1 4.12; gemma4 E4B Q4_0 | tg128 | 81.08 \u00b1 0.24",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17291611",
   "tg_tokens": 128,
   "tps": 81.08,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-07-10",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c17596982-qwen3527b-Q5_K - Medium",
   "model": "qwen35 27B",
   "notes": "threads=-1; config=b09959",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 935.68,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "Arc Pro B70; qwen35 27B Q5_K - Medium | pp512 | 935.68 \u00b1 1.32; qwen35 27B Q5_K - Medium | tg128 | 22.88 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17596982",
   "tg_tokens": 128,
   "tps": 22.88,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-07-10",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c17596982-qwen3527b-Q5_K - Medium-2",
   "model": "qwen35 27B",
   "notes": "threads=-1; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); config=b09959",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 837.0,
   "provenance": "community",
   "quant": "Q5_K - Medium",
   "quote": "Arc Pro B70; qwen35 27B Q5_K - Medium | pp512 | 837.00 \u00b1 1.14; qwen35 27B Q5_K - Medium | tg128 | 22.81 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-17596982",
   "tg_tokens": 128,
   "tps": 22.81,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08-18",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-10879-c18068837-q36apex-rocm-bare",
   "model": "Qwen3.6-35B-A3B-APEX-MTP-I-Balanced",
   "notes": "Strix Halo iGPU, llama.cpp master 2026-08, chat completions decode, single request; ROCm (HIP) build, no speculative decoding; config=rocm-bare",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q5_K_M",
   "quote": "Hardware: Ryzen AI MAX+ 395 / Radeon 8060S (gfx1151) / 128 GB LPDDR5X unified memory, no dedicated VRAM.; Method: llama.cpp master (2026-08), separate Vulkan (RADV) + ROCm (HIP) builds.; llama-bench -p512 -n128 -b512 -t16, plus llama-server /v1/chat/completions max_tokens=200. Single request, serial, \u00b15%/cell. Decode throughput (tok/s):; Qwen3.6-35B-A3B-APEX-MTP-I-Balanced MoE Q5_K_M 56.41 66.86 69.47 80.89",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18068837",
   "tg_tokens": null,
   "tps": "56.41",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08-18",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-10879-c18068837-q36apex-rocm-mtp",
   "model": "Qwen3.6-35B-A3B-APEX-MTP-I-Balanced",
   "notes": "Strix Halo iGPU, llama.cpp master 2026-08, chat completions decode, single request; ROCm (HIP) build + MTP speculative decoding; config=rocm-mtp",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q5_K_M",
   "quote": "Hardware: Ryzen AI MAX+ 395 / Radeon 8060S (gfx1151) / 128 GB LPDDR5X unified memory, no dedicated VRAM.; Method: llama.cpp master (2026-08), separate Vulkan (RADV) + ROCm (HIP) builds.; llama-bench -p512 -n128 -b512 -t16, plus llama-server /v1/chat/completions max_tokens=200. Single request, serial, \u00b15%/cell. Decode throughput (tok/s):; Qwen3.6-35B-A3B-APEX-MTP-I-Balanced MoE Q5_K_M 56.41 66.86 69.47 80.89",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18068837",
   "tg_tokens": null,
   "tps": "69.47",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08-18",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-10879-c18068837-q36apex-vulkan-bare",
   "model": "Qwen3.6-35B-A3B-APEX-MTP-I-Balanced",
   "notes": "Strix Halo iGPU, llama.cpp master 2026-08, chat completions decode, single request; Vulkan (RADV) build, no speculative decoding; config=vulkan-bare",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q5_K_M",
   "quote": "Hardware: Ryzen AI MAX+ 395 / Radeon 8060S (gfx1151) / 128 GB LPDDR5X unified memory, no dedicated VRAM.; Method: llama.cpp master (2026-08), separate Vulkan (RADV) + ROCm (HIP) builds.; llama-bench -p512 -n128 -b512 -t16, plus llama-server /v1/chat/completions max_tokens=200. Single request, serial, \u00b15%/cell. Decode throughput (tok/s):; Qwen3.6-35B-A3B-APEX-MTP-I-Balanced MoE Q5_K_M 56.41 66.86 69.47 80.89",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18068837",
   "tg_tokens": null,
   "tps": "66.86",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08-18",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-10879-c18068837-q36apex-vulkan-mtp",
   "model": "Qwen3.6-35B-A3B-APEX-MTP-I-Balanced",
   "notes": "Strix Halo iGPU, llama.cpp master 2026-08, chat completions decode, single request; Vulkan (RADV) build + MTP speculative decoding; config=vulkan-mtp",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q5_K_M",
   "quote": "Hardware: Ryzen AI MAX+ 395 / Radeon 8060S (gfx1151) / 128 GB LPDDR5X unified memory, no dedicated VRAM.; Method: llama.cpp master (2026-08), separate Vulkan (RADV) + ROCm (HIP) builds.; llama-bench -p512 -n128 -b512 -t16, plus llama-server /v1/chat/completions max_tokens=200. Single request, serial, \u00b15%/cell. Decode throughput (tok/s):; Qwen3.6-35B-A3B-APEX-MTP-I-Balanced MoE Q5_K_M 56.41 66.86 69.47 80.89",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18068837",
   "tg_tokens": null,
   "tps": "80.89",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08-18",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-10879-c18068837-hymt2-rocm-bare",
   "model": "Hy-MT2-30B-A3B-APEX-Imatrix-I-Balanced",
   "notes": "Strix Halo iGPU, llama.cpp master 2026-08, chat completions decode, single request; ROCm (HIP) build, no speculative decoding; config=rocm-bare",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "Hardware: Ryzen AI MAX+ 395 / Radeon 8060S (gfx1151) / 128 GB LPDDR5X unified memory, no dedicated VRAM.; Method: llama.cpp master (2026-08), separate Vulkan (RADV) + ROCm (HIP) builds.; llama-bench -p512 -n128 -b512 -t16, plus llama-server /v1/chat/completions max_tokens=200. Single request, serial, \u00b15%/cell. Decode throughput (tok/s):; Hy-MT2-30B-A3B-APEX-Imatrix-I-Balanced MoE Q6_K 59.19 74.28 \u2014 \u2014",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18068837",
   "tg_tokens": null,
   "tps": "59.19",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08-18",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-10879-c18068837-hymt2-vulkan-bare",
   "model": "Hy-MT2-30B-A3B-APEX-Imatrix-I-Balanced",
   "notes": "Strix Halo iGPU, llama.cpp master 2026-08, chat completions decode, single request; Vulkan (RADV) build, no speculative decoding; config=vulkan-bare",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "Hardware: Ryzen AI MAX+ 395 / Radeon 8060S (gfx1151) / 128 GB LPDDR5X unified memory, no dedicated VRAM.; Method: llama.cpp master (2026-08), separate Vulkan (RADV) + ROCm (HIP) builds.; llama-bench -p512 -n128 -b512 -t16, plus llama-server /v1/chat/completions max_tokens=200. Single request, serial, \u00b15%/cell. Decode throughput (tok/s):; Hy-MT2-30B-A3B-APEX-Imatrix-I-Balanced MoE Q6_K 59.19 74.28 \u2014 \u2014",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18068837",
   "tg_tokens": null,
   "tps": "74.28",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08-18",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-10879-c18068837-qwopus-rocm-bare",
   "model": "Qwopus3.6-27B-Coder-Compat-MTP",
   "notes": "Strix Halo iGPU, llama.cpp master 2026-08, chat completions decode, single request; ROCm (HIP) build, no speculative decoding; config=rocm-bare",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q5_K_M",
   "quote": "Hardware: Ryzen AI MAX+ 395 / Radeon 8060S (gfx1151) / 128 GB LPDDR5X unified memory, no dedicated VRAM.; Method: llama.cpp master (2026-08), separate Vulkan (RADV) + ROCm (HIP) builds.; llama-bench -p512 -n128 -b512 -t16, plus llama-server /v1/chat/completions max_tokens=200. Single request, serial, \u00b15%/cell. Decode throughput (tok/s):; Qwopus3.6-27B-Coder-Compat-MTP-Q5_K_M Dense Q5_K_M 10.94 11.23 19.38 22.38",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18068837",
   "tg_tokens": null,
   "tps": "10.94",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08-18",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-10879-c18068837-qwopus-rocm-mtp",
   "model": "Qwopus3.6-27B-Coder-Compat-MTP",
   "notes": "Strix Halo iGPU, llama.cpp master 2026-08, chat completions decode, single request; ROCm (HIP) build + MTP speculative decoding; config=rocm-mtp",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q5_K_M",
   "quote": "Hardware: Ryzen AI MAX+ 395 / Radeon 8060S (gfx1151) / 128 GB LPDDR5X unified memory, no dedicated VRAM.; Method: llama.cpp master (2026-08), separate Vulkan (RADV) + ROCm (HIP) builds.; llama-bench -p512 -n128 -b512 -t16, plus llama-server /v1/chat/completions max_tokens=200. Single request, serial, \u00b15%/cell. Decode throughput (tok/s):; Qwopus3.6-27B-Coder-Compat-MTP-Q5_K_M Dense Q5_K_M 10.94 11.23 19.38 22.38",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18068837",
   "tg_tokens": null,
   "tps": "19.38",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08-18",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-10879-c18068837-qwopus-vulkan-bare",
   "model": "Qwopus3.6-27B-Coder-Compat-MTP",
   "notes": "Strix Halo iGPU, llama.cpp master 2026-08, chat completions decode, single request; Vulkan (RADV) build, no speculative decoding; config=vulkan-bare",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q5_K_M",
   "quote": "Hardware: Ryzen AI MAX+ 395 / Radeon 8060S (gfx1151) / 128 GB LPDDR5X unified memory, no dedicated VRAM.; Method: llama.cpp master (2026-08), separate Vulkan (RADV) + ROCm (HIP) builds.; llama-bench -p512 -n128 -b512 -t16, plus llama-server /v1/chat/completions max_tokens=200. Single request, serial, \u00b15%/cell. Decode throughput (tok/s):; Qwopus3.6-27B-Coder-Compat-MTP-Q5_K_M Dense Q5_K_M 10.94 11.23 19.38 22.38",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18068837",
   "tg_tokens": null,
   "tps": "11.23",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08-18",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-10879-c18068837-qwopus-vulkan-mtp",
   "model": "Qwopus3.6-27B-Coder-Compat-MTP",
   "notes": "Strix Halo iGPU, llama.cpp master 2026-08, chat completions decode, single request; Vulkan (RADV) build + MTP speculative decoding; config=vulkan-mtp",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q5_K_M",
   "quote": "Hardware: Ryzen AI MAX+ 395 / Radeon 8060S (gfx1151) / 128 GB LPDDR5X unified memory, no dedicated VRAM.; Method: llama.cpp master (2026-08), separate Vulkan (RADV) + ROCm (HIP) builds.; llama-bench -p512 -n128 -b512 -t16, plus llama-server /v1/chat/completions max_tokens=200. Single request, serial, \u00b15%/cell. Decode throughput (tok/s):; Qwopus3.6-27B-Coder-Compat-MTP-Q5_K_M Dense Q5_K_M 10.94 11.23 19.38 22.38",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18068837",
   "tg_tokens": null,
   "tps": "22.38",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08-18",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-10879-c18068837-q38ad-rocm-bare",
   "model": "Qwen3.8-27B-AD",
   "notes": "Strix Halo iGPU, llama.cpp master 2026-08, chat completions decode, single request; ROCm (HIP) build, no speculative decoding; config=rocm-bare",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q5_K_M",
   "quote": "Hardware: Ryzen AI MAX+ 395 / Radeon 8060S (gfx1151) / 128 GB LPDDR5X unified memory, no dedicated VRAM.; Method: llama.cpp master (2026-08), separate Vulkan (RADV) + ROCm (HIP) builds.; llama-bench -p512 -n128 -b512 -t16, plus llama-server /v1/chat/completions max_tokens=200. Single request, serial, \u00b15%/cell. Decode throughput (tok/s):; Qwen3.8-27B-AD-Q5_K_M Dense Q5_K_M 10.52 10.76 14.09 16.98",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18068837",
   "tg_tokens": null,
   "tps": "10.52",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08-18",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-10879-c18068837-q38ad-rocm-mtp",
   "model": "Qwen3.8-27B-AD",
   "notes": "Strix Halo iGPU, llama.cpp master 2026-08, chat completions decode, single request; ROCm (HIP) build + MTP speculative decoding; config=rocm-mtp",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q5_K_M",
   "quote": "Hardware: Ryzen AI MAX+ 395 / Radeon 8060S (gfx1151) / 128 GB LPDDR5X unified memory, no dedicated VRAM.; Method: llama.cpp master (2026-08), separate Vulkan (RADV) + ROCm (HIP) builds.; llama-bench -p512 -n128 -b512 -t16, plus llama-server /v1/chat/completions max_tokens=200. Single request, serial, \u00b15%/cell. Decode throughput (tok/s):; Qwen3.8-27B-AD-Q5_K_M Dense Q5_K_M 10.52 10.76 14.09 16.98",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18068837",
   "tg_tokens": null,
   "tps": "14.09",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08-18",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-10879-c18068837-q38ad-vulkan-bare",
   "model": "Qwen3.8-27B-AD",
   "notes": "Strix Halo iGPU, llama.cpp master 2026-08, chat completions decode, single request; Vulkan (RADV) build, no speculative decoding; config=vulkan-bare",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q5_K_M",
   "quote": "Hardware: Ryzen AI MAX+ 395 / Radeon 8060S (gfx1151) / 128 GB LPDDR5X unified memory, no dedicated VRAM.; Method: llama.cpp master (2026-08), separate Vulkan (RADV) + ROCm (HIP) builds.; llama-bench -p512 -n128 -b512 -t16, plus llama-server /v1/chat/completions max_tokens=200. Single request, serial, \u00b15%/cell. Decode throughput (tok/s):; Qwen3.8-27B-AD-Q5_K_M Dense Q5_K_M 10.52 10.76 14.09 16.98",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18068837",
   "tg_tokens": null,
   "tps": "10.76",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": 1,
   "ctx": null,
   "date": "2026-08-18",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-10879-c18068837-q38ad-vulkan-mtp",
   "model": "Qwen3.8-27B-AD",
   "notes": "Strix Halo iGPU, llama.cpp master 2026-08, chat completions decode, single request; Vulkan (RADV) build + MTP speculative decoding; config=vulkan-mtp",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q5_K_M",
   "quote": "Hardware: Ryzen AI MAX+ 395 / Radeon 8060S (gfx1151) / 128 GB LPDDR5X unified memory, no dedicated VRAM.; Method: llama.cpp master (2026-08), separate Vulkan (RADV) + ROCm (HIP) builds.; llama-bench -p512 -n128 -b512 -t16, plus llama-server /v1/chat/completions max_tokens=200. Single request, serial, \u00b15%/cell. Decode throughput (tok/s):; Qwen3.8-27B-AD-Q5_K_M Dense Q5_K_M 10.52 10.76 14.09 16.98",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18068837",
   "tg_tokens": null,
   "tps": "16.98",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-20",
   "flags": [],
   "hardware": "AMD Custom GPU 0932",
   "id": "lc-disc-10879-c18099068-llama7b-fa0",
   "model": "llama 7B",
   "notes": "Steam Deck (AMD Custom GPU 0932, VANGOGH iGPU), llama-b10360, -ngl 100, flash attention off; config=fa0",
   "params": "6.74B",
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "144.31",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Custom GPU 0932 (RADV VANGOGH); build: 48d22e2 (10360; deck@steamdeck llama-b10360; llama-2-7b.Q4_0.gguf llama 7B Q4_0 3.56 GiB 6.74 B Vulkan 100 0 pp512 144.31 \u00b1 2.74; llama-2-7b.Q4_0.gguf llama 7B Q4_0 3.56 GiB 6.74 B Vulkan 100 0 tg128 17.52 \u00b1 0.15",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18099068",
   "tg_tokens": 128,
   "tps": "17.52",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-20",
   "flags": [],
   "hardware": "AMD Custom GPU 0932",
   "id": "lc-disc-10879-c18099068-llama7b-fa1",
   "model": "llama 7B",
   "notes": "Steam Deck (AMD Custom GPU 0932, VANGOGH iGPU), llama-b10360, -ngl 100, flash attention on; config=fa1",
   "params": "6.74B",
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "164.34",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Custom GPU 0932 (RADV VANGOGH); build: 48d22e2 (10360; deck@steamdeck llama-b10360; llama-2-7b.Q4_0.gguf llama 7B Q4_0 3.56 GiB 6.74 B Vulkan 100 1 pp512 164.34 \u00b1 1.65; llama-2-7b.Q4_0.gguf llama 7B Q4_0 3.56 GiB 6.74 B Vulkan 100 1 tg128 18.40 \u00b1 0.14",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18099068",
   "tg_tokens": 128,
   "tps": "18.40",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-20",
   "flags": [],
   "hardware": "AMD Custom GPU 0932",
   "id": "lc-disc-10879-c18099068-q38-27b-fa0",
   "model": "Qwen3.8-27B",
   "notes": "Steam Deck (AMD Custom GPU 0932, VANGOGH iGPU), llama-b10360, -ngl 100, flash attention off; config=fa0",
   "params": "27.32B",
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "19.32",
   "provenance": "community",
   "quant": "Q4_K - Small",
   "quote": "AMD Custom GPU 0932 (RADV VANGOGH); build: 48d22e2 (10360; deck@steamdeck llama-b10360; Qwen3.8-27B-UD-IQ2_XXS.gguf qwen35 27B Q4_K - Small 8.38 GiB 27.32 B Vulkan 100 0 pp512 19.32 \u00b1 0.68; Qwen3.8-27B-UD-IQ2_XXS.gguf qwen35 27B Q4_K - Small 8.38 GiB 27.32 B Vulkan 100 0 tg128 3.81 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18099068",
   "tg_tokens": 128,
   "tps": "3.81",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-20",
   "flags": [],
   "hardware": "AMD Custom GPU 0932",
   "id": "lc-disc-10879-c18099068-q38-27b-fa1",
   "model": "Qwen3.8-27B",
   "notes": "Steam Deck (AMD Custom GPU 0932, VANGOGH iGPU), llama-b10360, -ngl 100, flash attention on; config=fa1",
   "params": "27.32B",
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "19.83",
   "provenance": "community",
   "quant": "Q4_K - Small",
   "quote": "AMD Custom GPU 0932 (RADV VANGOGH); build: 48d22e2 (10360; deck@steamdeck llama-b10360; Qwen3.8-27B-UD-IQ2_XXS.gguf qwen35 27B Q4_K - Small 8.38 GiB 27.32 B Vulkan 100 1 pp512 19.83 \u00b1 0.16; Qwen3.8-27B-UD-IQ2_XXS.gguf qwen35 27B Q4_K - Small 8.38 GiB 27.32 B Vulkan 100 1 tg128 3.81 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18099068",
   "tg_tokens": 128,
   "tps": "3.81",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-20",
   "flags": [],
   "hardware": "AMD Custom GPU 0932",
   "id": "lc-disc-10879-c18099068-q38-9b-fa0",
   "model": "Qwen3.8-9B",
   "notes": "Steam Deck (AMD Custom GPU 0932, VANGOGH iGPU), llama-b10360, -ngl 100, flash attention off; config=fa0",
   "params": "9.20B",
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "101.16",
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "AMD Custom GPU 0932 (RADV VANGOGH); build: 48d22e2 (10360; deck@steamdeck llama-b10360; Qwen3.8-9B-Q4_K_M.gguf qwen35 9B Q4_K - Medium 5.37 GiB 9.20 B Vulkan 100 0 pp512 101.16 \u00b1 0.56; Qwen3.8-9B-Q4_K_M.gguf qwen35 9B Q4_K - Medium 5.37 GiB 9.20 B Vulkan 100 0 tg128 12.22 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18099068",
   "tg_tokens": 128,
   "tps": "12.22",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-20",
   "flags": [],
   "hardware": "AMD Custom GPU 0932",
   "id": "lc-disc-10879-c18099068-q38-9b-fa1",
   "model": "Qwen3.8-9B",
   "notes": "Steam Deck (AMD Custom GPU 0932, VANGOGH iGPU), llama-b10360, -ngl 100, flash attention on; config=fa1",
   "params": "9.20B",
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "100.16",
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "AMD Custom GPU 0932 (RADV VANGOGH); build: 48d22e2 (10360; deck@steamdeck llama-b10360; Qwen3.8-9B-Q4_K_M.gguf qwen35 9B Q4_K - Medium 5.37 GiB 9.20 B Vulkan 100 1 pp512 100.16 \u00b1 0.59; Qwen3.8-9B-Q4_K_M.gguf qwen35 9B Q4_K - Medium 5.37 GiB 9.20 B Vulkan 100 1 tg128 12.08 \u00b1 0.31",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18099068",
   "tg_tokens": 128,
   "tps": "12.08",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-20",
   "flags": [],
   "hardware": "AMD Custom GPU 0932",
   "id": "lc-disc-10879-c18099068-q35-4b-fa0",
   "model": "Qwen3.5-4B",
   "notes": "Steam Deck (AMD Custom GPU 0932, VANGOGH iGPU), llama-b10360, -ngl 100, flash attention off; config=fa0",
   "params": "4.21B",
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "189.22",
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "AMD Custom GPU 0932 (RADV VANGOGH); build: 48d22e2 (10360; deck@steamdeck llama-b10360; Qwen3.5-4B-Q4_K_M.gguf qwen35 4B Q4_K - Medium 2.54 GiB 4.21 B Vulkan 100 0 pp512 189.22 \u00b1 1.31; Qwen3.5-4B-Q4_K_M.gguf qwen35 4B Q4_K - Medium 2.54 GiB 4.21 B Vulkan 100 0 tg128 22.55 \u00b1 0.13",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18099068",
   "tg_tokens": 128,
   "tps": "22.55",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-20",
   "flags": [],
   "hardware": "AMD Custom GPU 0932",
   "id": "lc-disc-10879-c18099068-q35-4b-fa1",
   "model": "Qwen3.5-4B",
   "notes": "Steam Deck (AMD Custom GPU 0932, VANGOGH iGPU), llama-b10360, -ngl 100, flash attention on; config=fa1",
   "params": "4.21B",
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "190.20",
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "AMD Custom GPU 0932 (RADV VANGOGH); build: 48d22e2 (10360; deck@steamdeck llama-b10360; Qwen3.5-4B-Q4_K_M.gguf qwen35 4B Q4_K - Medium 2.54 GiB 4.21 B Vulkan 100 1 pp512 190.20 \u00b1 1.28; Qwen3.5-4B-Q4_K_M.gguf qwen35 4B Q4_K - Medium 2.54 GiB 4.21 B Vulkan 100 1 tg128 22.67 \u00b1 0.11",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18099068",
   "tg_tokens": 128,
   "tps": "22.67",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-20",
   "flags": [],
   "hardware": "AMD Custom GPU 0932",
   "id": "lc-disc-10879-c18099068-ornith-fa0",
   "model": "Ornith-1.5-9B",
   "notes": "Steam Deck (AMD Custom GPU 0932, VANGOGH iGPU), llama-b10360, -ngl 100, flash attention off; config=fa0",
   "params": "8.95B",
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "93.88",
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "AMD Custom GPU 0932 (RADV VANGOGH); build: 48d22e2 (10360; deck@steamdeck llama-b10360; Ornith-1.5-9B-Q4_K_M.gguf qwen35 9B Q4_K - Medium 5.23 GiB 8.95 B Vulkan 100 0 pp512 93.88 \u00b1 2.10; Ornith-1.5-9B-Q4_K_M.gguf qwen35 9B Q4_K - Medium 5.23 GiB 8.95 B Vulkan 100 0 tg128 11.22 \u00b1 0.21",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18099068",
   "tg_tokens": 128,
   "tps": "11.22",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-20",
   "flags": [],
   "hardware": "AMD Custom GPU 0932",
   "id": "lc-disc-10879-c18099068-ornith-fa1",
   "model": "Ornith-1.5-9B",
   "notes": "Steam Deck (AMD Custom GPU 0932, VANGOGH iGPU), llama-b10360, -ngl 100, flash attention on; config=fa1",
   "params": "8.95B",
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "98.66",
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "AMD Custom GPU 0932 (RADV VANGOGH); build: 48d22e2 (10360; deck@steamdeck llama-b10360; Ornith-1.5-9B-Q4_K_M.gguf qwen35 9B Q4_K - Medium 5.23 GiB 8.95 B Vulkan 100 1 pp512 98.66 \u00b1 0.68; Ornith-1.5-9B-Q4_K_M.gguf qwen35 9B Q4_K - Medium 5.23 GiB 8.95 B Vulkan 100 1 tg128 11.48 \u00b1 0.13",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18099068",
   "tg_tokens": 128,
   "tps": "11.48",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-20",
   "flags": [],
   "hardware": "AMD Custom GPU 0932",
   "id": "lc-disc-10879-c18099068-agentworld-fa0",
   "model": "Qwen-AgentWorld-35B-A3B",
   "notes": "Steam Deck (AMD Custom GPU 0932, VANGOGH iGPU), llama-b10360, -ngl 100, flash attention off; config=fa0",
   "params": "34.66B",
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "86.75",
   "provenance": "community",
   "quant": "IQ2_XXS - 2.0625 bpw",
   "quote": "AMD Custom GPU 0932 (RADV VANGOGH); build: 48d22e2 (10360; deck@steamdeck llama-b10360; Qwen-AgentWorld-35B-A3B-UD-IQ2_XXS.gguf qwen35moe 35B.A3B IQ2_XXS - 2.0625 bpw 10.70 GiB 34.66 B Vulkan 100 0 pp512 86.75 \u00b1 2.00; Qwen-AgentWorld-35B-A3B-UD-IQ2_XXS.gguf qwen35moe 35B.A3B IQ2_XXS - 2.0625 bpw 10.70 GiB 34.66 B Vulkan 100 0 tg128 23.01 \u00b1 0.49",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18099068",
   "tg_tokens": 128,
   "tps": "23.01",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-20",
   "flags": [],
   "hardware": "AMD Custom GPU 0932",
   "id": "lc-disc-10879-c18099068-agentworld-fa1",
   "model": "Qwen-AgentWorld-35B-A3B",
   "notes": "Steam Deck (AMD Custom GPU 0932, VANGOGH iGPU), llama-b10360, -ngl 100, flash attention on; config=fa1",
   "params": "34.66B",
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "88.23",
   "provenance": "community",
   "quant": "IQ2_XXS - 2.0625 bpw",
   "quote": "AMD Custom GPU 0932 (RADV VANGOGH); build: 48d22e2 (10360; deck@steamdeck llama-b10360; Qwen-AgentWorld-35B-A3B-UD-IQ2_XXS.gguf qwen35moe 35B.A3B IQ2_XXS - 2.0625 bpw 10.70 GiB 34.66 B Vulkan 100 1 pp512 88.23 \u00b1 2.20; Qwen-AgentWorld-35B-A3B-UD-IQ2_XXS.gguf qwen35moe 35B.A3B IQ2_XXS - 2.0625 bpw 10.70 GiB 34.66 B Vulkan 100 1 tg128 23.09 \u00b1 0.37",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18099068",
   "tg_tokens": 128,
   "tps": "23.09",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-20",
   "flags": [],
   "hardware": "AMD Custom GPU 0932",
   "id": "lc-disc-10879-c18099068-muse-fa0",
   "model": "muse-glimmer 30B",
   "notes": "Steam Deck (AMD Custom GPU 0932, VANGOGH iGPU), llama-b10360, -ngl 100, flash attention off; config=fa0",
   "params": "27.85B",
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "19.63",
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "AMD Custom GPU 0932 (RADV VANGOGH); build: 48d22e2 (10360; deck@steamdeck llama-b10360; Muse-Glimmer-30B-UD-IQ2_XXS.gguf muse-glimmer 30B Q4_K - Medium 10.00 GiB 27.85 B Vulkan 100 0 pp512 19.63 \u00b1 0.22; Muse-Glimmer-30B-UD-IQ2_XXS.gguf muse-glimmer 30B Q4_K - Medium 10.00 GiB 27.85 B Vulkan 100 0 tg128 3.77 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18099068",
   "tg_tokens": 128,
   "tps": "3.77",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-20",
   "flags": [],
   "hardware": "AMD Custom GPU 0932",
   "id": "lc-disc-10879-c18099068-muse-fa1",
   "model": "muse-glimmer 30B",
   "notes": "Steam Deck (AMD Custom GPU 0932, VANGOGH iGPU), llama-b10360, -ngl 100, flash attention on; config=fa1",
   "params": "27.85B",
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "19.77",
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "AMD Custom GPU 0932 (RADV VANGOGH); build: 48d22e2 (10360; deck@steamdeck llama-b10360; Muse-Glimmer-30B-UD-IQ2_XXS.gguf muse-glimmer 30B Q4_K - Medium 10.00 GiB 27.85 B Vulkan 100 1 pp512 19.77 \u00b1 0.19; Muse-Glimmer-30B-UD-IQ2_XXS.gguf muse-glimmer 30B Q4_K - Medium 10.00 GiB 27.85 B Vulkan 100 1 tg128 3.83 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18099068",
   "tg_tokens": 128,
   "tps": "3.83",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-20",
   "flags": [],
   "hardware": "AMD Custom GPU 0932",
   "id": "lc-disc-10879-c18099068-gemma31-fa0",
   "model": "gemma-4-31B-it",
   "notes": "Steam Deck (AMD Custom GPU 0932, VANGOGH iGPU), llama-b10360, -ngl 100, flash attention off; config=fa0",
   "params": "30.70B",
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "17.15",
   "provenance": "community",
   "quant": "IQ2_XXS - 2.0625 bpw",
   "quote": "AMD Custom GPU 0932 (RADV VANGOGH); build: 48d22e2 (10360; deck@steamdeck llama-b10360; gemma-4-31B-it-UD-IQ2_XXS.gguf gemma4 31B IQ2_XXS - 2.0625 bpw 7.93 GiB 30.70 B Vulkan 100 0 pp512 17.15 \u00b1 0.08; gemma-4-31B-it-UD-IQ2_XXS.gguf gemma4 31B IQ2_XXS - 2.0625 bpw 7.93 GiB 30.70 B Vulkan 100 0 tg128 3.50 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18099068",
   "tg_tokens": 128,
   "tps": "3.50",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-20",
   "flags": [],
   "hardware": "AMD Custom GPU 0932",
   "id": "lc-disc-10879-c18099068-gemma31-fa1",
   "model": "gemma-4-31B-it",
   "notes": "Steam Deck (AMD Custom GPU 0932, VANGOGH iGPU), llama-b10360, -ngl 100, flash attention on; config=fa1",
   "params": "30.70B",
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "17.22",
   "provenance": "community",
   "quant": "IQ2_XXS - 2.0625 bpw",
   "quote": "AMD Custom GPU 0932 (RADV VANGOGH); build: 48d22e2 (10360; deck@steamdeck llama-b10360; gemma-4-31B-it-UD-IQ2_XXS.gguf gemma4 31B IQ2_XXS - 2.0625 bpw 7.93 GiB 30.70 B Vulkan 100 1 pp512 17.22 \u00b1 0.06; gemma-4-31B-it-UD-IQ2_XXS.gguf gemma4 31B IQ2_XXS - 2.0625 bpw 7.93 GiB 30.70 B Vulkan 100 1 tg128 3.56 \u00b1 0.01",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18099068",
   "tg_tokens": 128,
   "tps": "3.56",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-20",
   "flags": [],
   "hardware": "AMD Custom GPU 0932",
   "id": "lc-disc-10879-c18099068-gemma12qat-fa0",
   "model": "gemma-4-12B-it-qat",
   "notes": "Steam Deck (AMD Custom GPU 0932, VANGOGH iGPU), llama-b10360, -ngl 100, flash attention off; config=fa0",
   "params": "11.91B",
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "84.66",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Custom GPU 0932 (RADV VANGOGH); build: 48d22e2 (10360; deck@steamdeck llama-b10360; gemma-4-12B-it-qat-UD-Q4_K_XL.gguf gemma4 ?B Q4_0 6.24 GiB 11.91 B Vulkan 100 0 pp512 84.66 \u00b1 0.61; gemma-4-12B-it-qat-UD-Q4_K_XL.gguf gemma4 ?B Q4_0 6.24 GiB 11.91 B Vulkan 100 0 tg128 9.06 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18099068",
   "tg_tokens": 128,
   "tps": "9.06",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-20",
   "flags": [],
   "hardware": "AMD Custom GPU 0932",
   "id": "lc-disc-10879-c18099068-gemma12qat-fa1",
   "model": "gemma-4-12B-it-qat",
   "notes": "Steam Deck (AMD Custom GPU 0932, VANGOGH iGPU), llama-b10360, -ngl 100, flash attention on; config=fa1",
   "params": "11.91B",
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "90.83",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Custom GPU 0932 (RADV VANGOGH); build: 48d22e2 (10360; deck@steamdeck llama-b10360; gemma-4-12B-it-qat-UD-Q4_K_XL.gguf gemma4 ?B Q4_0 6.24 GiB 11.91 B Vulkan 100 1 pp512 90.83 \u00b1 0.38; gemma-4-12B-it-qat-UD-Q4_K_XL.gguf gemma4 ?B Q4_0 6.24 GiB 11.91 B Vulkan 100 1 tg128 9.23 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18099068",
   "tg_tokens": 128,
   "tps": "9.23",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-20",
   "flags": [],
   "hardware": "AMD Custom GPU 0932",
   "id": "lc-disc-10879-c18099068-gemmae4b-fa0",
   "model": "Gemma 4 E4B",
   "notes": "Steam Deck (AMD Custom GPU 0932, VANGOGH iGPU), llama-b10360, -ngl 100, flash attention off; config=fa0",
   "params": "7.52B",
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "163.93",
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "AMD Custom GPU 0932 (RADV VANGOGH); build: 48d22e2 (10360; deck@steamdeck llama-b10360; gemma-4-E4B-it-UD-Q4_K_XL.gguf gemma4 E4B Q4_K - Medium 4.74 GiB 7.52 B Vulkan 100 0 pp512 163.93 \u00b1 5.22; gemma-4-E4B-it-UD-Q4_K_XL.gguf gemma4 E4B Q4_K - Medium 4.74 GiB 7.52 B Vulkan 100 0 tg128 17.80 \u00b1 0.13",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18099068",
   "tg_tokens": 128,
   "tps": "17.80",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-20",
   "flags": [],
   "hardware": "AMD Custom GPU 0932",
   "id": "lc-disc-10879-c18099068-gemmae4b-fa1",
   "model": "Gemma 4 E4B",
   "notes": "Steam Deck (AMD Custom GPU 0932, VANGOGH iGPU), llama-b10360, -ngl 100, flash attention on; config=fa1",
   "params": "7.52B",
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "161.37",
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "AMD Custom GPU 0932 (RADV VANGOGH); build: 48d22e2 (10360; deck@steamdeck llama-b10360; gemma-4-E4B-it-UD-Q4_K_XL.gguf gemma4 E4B Q4_K - Medium 4.74 GiB 7.52 B Vulkan 100 1 pp512 161.37 \u00b1 1.09; gemma-4-E4B-it-UD-Q4_K_XL.gguf gemma4 E4B Q4_K - Medium 4.74 GiB 7.52 B Vulkan 100 1 tg128 17.78 \u00b1 0.67",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18099068",
   "tg_tokens": 128,
   "tps": "17.78",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-20",
   "flags": [],
   "hardware": "AMD Custom GPU 0932",
   "id": "lc-disc-10879-c18099068-nemotron-fa0",
   "model": "NVIDIA-Nemotron-3-Nano-4B",
   "notes": "Steam Deck (AMD Custom GPU 0932, VANGOGH iGPU), llama-b10360, -ngl 100, flash attention off; config=fa0",
   "params": "3.97B",
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "195.39",
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "AMD Custom GPU 0932 (RADV VANGOGH); build: 48d22e2 (10360; deck@steamdeck llama-b10360; NVIDIA-Nemotron-3-Nano-4B-UD-Q4_K_XL.gguf nemotron_h ?B Q4_K - Medium 2.91 GiB 3.97 B Vulkan 100 0 pp512 195.39 \u00b1 1.30; NVIDIA-Nemotron-3-Nano-4B-UD-Q4_K_XL.gguf nemotron_h ?B Q4_K - Medium 2.91 GiB 3.97 B Vulkan 100 0 tg128 18.30 \u00b1 0.09",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18099068",
   "tg_tokens": 128,
   "tps": "18.30",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-20",
   "flags": [],
   "hardware": "AMD Custom GPU 0932",
   "id": "lc-disc-10879-c18099068-nemotron-fa1",
   "model": "NVIDIA-Nemotron-3-Nano-4B",
   "notes": "Steam Deck (AMD Custom GPU 0932, VANGOGH iGPU), llama-b10360, -ngl 100, flash attention on; config=fa1",
   "params": "3.97B",
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "192.69",
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "AMD Custom GPU 0932 (RADV VANGOGH); build: 48d22e2 (10360; deck@steamdeck llama-b10360; NVIDIA-Nemotron-3-Nano-4B-UD-Q4_K_XL.gguf nemotron_h ?B Q4_K - Medium 2.91 GiB 3.97 B Vulkan 100 1 pp512 192.69 \u00b1 1.81; NVIDIA-Nemotron-3-Nano-4B-UD-Q4_K_XL.gguf nemotron_h ?B Q4_K - Medium 2.91 GiB 3.97 B Vulkan 100 1 tg128 18.46 \u00b1 0.22",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18099068",
   "tg_tokens": 128,
   "tps": "18.46",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-20",
   "flags": [],
   "hardware": "AMD Custom GPU 0932",
   "id": "lc-disc-10879-c18099068-nanbeige-fa0",
   "model": "Nanbeige4.2-3B",
   "notes": "Steam Deck (AMD Custom GPU 0932, VANGOGH iGPU), llama-b10360, -ngl 100, flash attention off; config=fa0",
   "params": "4.17B",
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "123.81",
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "AMD Custom GPU 0932 (RADV VANGOGH); build: 48d22e2 (10360; deck@steamdeck llama-b10360; Nanbeige4.2-3B-Q4_K_M.gguf nanbeige ?B Q4_K - Medium 2.39 GiB 4.17 B Vulkan 100 0 pp512 123.81 \u00b1 0.68; Nanbeige4.2-3B-Q4_K_M.gguf nanbeige ?B Q4_K - Medium 2.39 GiB 4.17 B Vulkan 100 0 tg128 15.27 \u00b1 0.44",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18099068",
   "tg_tokens": 128,
   "tps": "15.27",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-08-20",
   "flags": [],
   "hardware": "AMD Custom GPU 0932",
   "id": "lc-disc-10879-c18099068-nanbeige-fa1",
   "model": "Nanbeige4.2-3B",
   "notes": "Steam Deck (AMD Custom GPU 0932, VANGOGH iGPU), llama-b10360, -ngl 100, flash attention on; config=fa1",
   "params": "4.17B",
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "126.28",
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "AMD Custom GPU 0932 (RADV VANGOGH); build: 48d22e2 (10360; deck@steamdeck llama-b10360; Nanbeige4.2-3B-Q4_K_M.gguf nanbeige ?B Q4_K - Medium 2.39 GiB 4.17 B Vulkan 100 1 pp512 126.28 \u00b1 2.13; Nanbeige4.2-3B-Q4_K_M.gguf nanbeige ?B Q4_K - Medium 2.39 GiB 4.17 B Vulkan 100 1 tg128 16.36 \u00b1 0.20",
   "ram_gb": null,
   "retrieved": "2026-10-01",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18099068",
   "tg_tokens": 128,
   "tps": "16.36",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "RX 6800",
   "id": "lc-disc-10879-c18706574-fa0",
   "model": "llama 7B",
   "notes": "config=fa0",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "1593.97",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Radeon RX 6800 (16 GB GDDR6, 256-bit), Arch Linux (kernel 7.2.5), Mesa RADV 26.2.2, Ryzen 7 5700X, Resizable BAR on; build: f1cee99; llama 7B Q4_0 3.56 GiB 6.74 B Vulkan 100 0 pp512 1593.97 \u00b1 1.38; llama 7B Q4_0 3.56 GiB 6.74 B Vulkan 100 0 tg128 101.47 \u00b1 0.35",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #10879 (performance thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18706574",
   "tg_tokens": 128,
   "tps": "101.47",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "RX 6800",
   "id": "lc-disc-10879-c18706574-fa1",
   "model": "llama 7B",
   "notes": "config=fa1",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "1597.95",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Radeon RX 6800 (16 GB GDDR6, 256-bit), Arch Linux (kernel 7.2.5), Mesa RADV 26.2.2, Ryzen 7 5700X, Resizable BAR on; build: f1cee99; llama 7B Q4_0 3.56 GiB 6.74 B Vulkan 100 1 pp512 1597.95 \u00b1 1.63; llama 7B Q4_0 3.56 GiB 6.74 B Vulkan 100 1 tg128 106.49 \u00b1 0.23",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #10879 (performance thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18706574",
   "tg_tokens": 128,
   "tps": "106.49",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "RX 6800",
   "id": "lc-disc-15021-c18706576-fa0",
   "model": "llama 7B",
   "notes": "config=fa0",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "1510.13",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RX 6800 (16 GB GDDR6, 256-bit, gfx1030), Arch Linux (kernel 7.2.5), ROCm 7.2.4 (Arch packages), Ryzen 7 5700X, Resizable BAR on; build: f1cee99; llama 7B Q4_0 3.56 GiB 6.74 B ROCm 99 0 pp512 1510.13 \u00b1 8.51; llama 7B Q4_0 3.56 GiB 6.74 B ROCm 99 0 tg128 86.03 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #15021 (performance thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-18706576",
   "tg_tokens": 128,
   "tps": "86.03",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "RX 6800",
   "id": "lc-disc-15021-c18706576-fa1",
   "model": "llama 7B",
   "notes": "config=fa1",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "1738.62",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RX 6800 (16 GB GDDR6, 256-bit, gfx1030), Arch Linux (kernel 7.2.5), ROCm 7.2.4 (Arch packages), Ryzen 7 5700X, Resizable BAR on; build: f1cee99; llama 7B Q4_0 3.56 GiB 6.74 B ROCm 99 1 pp512 1738.62 \u00b1 9.53; llama 7B Q4_0 3.56 GiB 6.74 B ROCm 99 1 tg128 93.6 \u00b1 0.08",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #15021 (performance thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-18706576",
   "tg_tokens": 128,
   "tps": "93.6",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-12",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "lc-disc-15021-c18407971-rocm-fa0",
   "model": "llama 7B",
   "notes": "config=rocm-fa0",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "4665.08",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Device: AMD Radeon AI PRO R9700, gfx1201 (RDNA4), 32 GB; ROCm: TheRock 10.1.0a20260828 (HIP 7.16.26343); Vulkan: Mesa RADV 26.1.8; llama.cpp: 8ea2902; llama 7B Q4_0 3.56 GiB 6.74 B ROCm 99 0 pp512 4665.08 \u00b1 528.61; llama 7B Q4_0 3.56 GiB 6.74 B ROCm 99 0 tg128 112.72 \u00b1 0.52",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #15021 (performance thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-18407971",
   "tg_tokens": 128,
   "tps": "112.72",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-12",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "lc-disc-15021-c18407971-rocm-fa1",
   "model": "llama 7B",
   "notes": "config=rocm-fa1",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "5039.78",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Device: AMD Radeon AI PRO R9700, gfx1201 (RDNA4), 32 GB; ROCm: TheRock 10.1.0a20260828 (HIP 7.16.26343); Vulkan: Mesa RADV 26.1.8; llama.cpp: 8ea2902; llama 7B Q4_0 3.56 GiB 6.74 B ROCm 99 1 pp512 5039.78 \u00b1 151.04; llama 7B Q4_0 3.56 GiB 6.74 B ROCm 99 1 tg128 121.62 \u00b1 0.38",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #15021 (performance thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-18407971",
   "tg_tokens": 128,
   "tps": "121.62",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-12",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "lc-disc-15021-c18407971-rocm-pp8192",
   "model": "llama 7B",
   "notes": "config=rocm-pp8192",
   "params": null,
   "power_w": null,
   "pp_tokens": 8192,
   "pp_tps": "2755.01",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Device: AMD Radeon AI PRO R9700, gfx1201 (RDNA4), 32 GB; ROCm: TheRock 10.1.0a20260828 (HIP 7.16.26343); Vulkan: Mesa RADV 26.1.8; llama.cpp: 8ea2902; llama 7B Q4_0 3.56 GiB 6.74 B ROCm 99 1 pp8192 2755.01 \u00b1 7.01; llama 7B Q4_0 3.56 GiB 6.74 B ROCm 99 1 tg128 121.29 \u00b1 0.35",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #15021 (performance thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-18407971",
   "tg_tokens": 128,
   "tps": "121.29",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-12",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "lc-disc-15021-c18407971-vk-fa0",
   "model": "llama 7B",
   "notes": "config=vk-fa0",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "5033.0",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Device: AMD Radeon AI PRO R9700, gfx1201 (RDNA4), 32 GB; ROCm: TheRock 10.1.0a20260828 (HIP 7.16.26343); Vulkan: Mesa RADV 26.1.8; llama.cpp: 8ea2902; llama 7B Q4_0 3.56 GiB 6.74 B Vulkan 99 0 pp512 5033.0 \u00b1 298.57; llama 7B Q4_0 3.56 GiB 6.74 B Vulkan 99 0 tg128 125.96 \u00b1 0.31",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #15021 (performance thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-18407971",
   "tg_tokens": 128,
   "tps": "125.96",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-12",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "lc-disc-15021-c18407971-vk-fa1",
   "model": "llama 7B",
   "notes": "config=vk-fa1",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "5411.28",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Device: AMD Radeon AI PRO R9700, gfx1201 (RDNA4), 32 GB; ROCm: TheRock 10.1.0a20260828 (HIP 7.16.26343); Vulkan: Mesa RADV 26.1.8; llama.cpp: 8ea2902; llama 7B Q4_0 3.56 GiB 6.74 B Vulkan 99 1 pp512 5411.28 \u00b1 12.85; llama 7B Q4_0 3.56 GiB 6.74 B Vulkan 99 1 tg128 131.19 \u00b1 0.11",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #15021 (performance thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-18407971",
   "tg_tokens": 128,
   "tps": "131.19",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-12",
   "flags": [],
   "hardware": "Radeon AI PRO R9700",
   "id": "lc-disc-15021-c18407971-vk-pp8192",
   "model": "llama 7B",
   "notes": "config=vk-pp8192",
   "params": null,
   "power_w": null,
   "pp_tokens": 8192,
   "pp_tps": "4235.87",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Device: AMD Radeon AI PRO R9700, gfx1201 (RDNA4), 32 GB; ROCm: TheRock 10.1.0a20260828 (HIP 7.16.26343); Vulkan: Mesa RADV 26.1.8; llama.cpp: 8ea2902; llama 7B Q4_0 3.56 GiB 6.74 B Vulkan 99 1 pp8192 4235.87 \u00b1 5.47; llama 7B Q4_0 3.56 GiB 6.74 B Vulkan 99 1 tg128 130.97 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #15021 (performance thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-18407971",
   "tg_tokens": 128,
   "tps": "130.97",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-12",
   "flags": [],
   "hardware": "Arc(TM) B390",
   "id": "lc-disc-10879-c18413704-fa0-short",
   "model": "llama 8B",
   "notes": "config=fa0-short",
   "params": "8.03B",
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "1380.07",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "ggml_vulkan: 0 = Intel(R) Arc(TM) B390 GPU (Intel Corporation) | uma: 1 | fp16: 1 | matrix cores: KHR_coopmat; build: 481c65f09 (10903); llama 8B Q4_0 4.35 GiB 8.03 B Vulkan 99 0 pp512 1380.07 \u00b1 10.98; llama 8B Q4_0 4.35 GiB 8.03 B Vulkan 99 0 tg128 24.45 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #10879 (performance thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18413704",
   "tg_tokens": 128,
   "tps": "24.45",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-12",
   "flags": [],
   "hardware": "Arc(TM) B390",
   "id": "lc-disc-10879-c18413704-fa1-short",
   "model": "llama 8B",
   "notes": "config=fa1-short",
   "params": "8.03B",
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "1332.41",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "ggml_vulkan: 0 = Intel(R) Arc(TM) B390 GPU (Intel Corporation) | uma: 1 | fp16: 1 | matrix cores: KHR_coopmat; build: 481c65f09 (10903); llama 8B Q4_0 4.35 GiB 8.03 B Vulkan 99 1 pp512 1332.41 \u00b1 5.02; llama 8B Q4_0 4.35 GiB 8.03 B Vulkan 99 1 tg128 23.93 \u00b1 0.16",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #10879 (performance thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18413704",
   "tg_tokens": 128,
   "tps": "23.93",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-12",
   "flags": [],
   "hardware": "Arc(TM) B390",
   "id": "lc-disc-10879-c18413704-fa0-long",
   "model": "llama 8B",
   "notes": "config=fa0-long",
   "params": "8.03B",
   "power_w": null,
   "pp_tokens": 8192,
   "pp_tps": "733.92",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "ggml_vulkan: 0 = Intel(R) Arc(TM) B390 GPU (Intel Corporation) | uma: 1 | fp16: 1 | matrix cores: KHR_coopmat; build: 481c65f09 (10903); llama 8B Q4_0 4.35 GiB 8.03 B Vulkan 99 0 pp8192 733.92 \u00b1 1.71; llama 8B Q4_0 4.35 GiB 8.03 B Vulkan 99 0 tg2048 23.47 \u00b1 0.22",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #10879 (performance thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18413704",
   "tg_tokens": 2048,
   "tps": "23.47",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-12",
   "flags": [],
   "hardware": "Arc(TM) B390",
   "id": "lc-disc-10879-c18413704-fa1-long",
   "model": "llama 8B",
   "notes": "config=fa1-long",
   "params": "8.03B",
   "power_w": null,
   "pp_tokens": 8192,
   "pp_tps": "314.3",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "ggml_vulkan: 0 = Intel(R) Arc(TM) B390 GPU (Intel Corporation) | uma: 1 | fp16: 1 | matrix cores: KHR_coopmat; build: 481c65f09 (10903); llama 8B Q4_0 4.35 GiB 8.03 B Vulkan 99 1 pp8192 314.3 \u00b1 0.44; llama 8B Q4_0 4.35 GiB 8.03 B Vulkan 99 1 tg2048 22.04 \u00b1 0.12",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #10879 (performance thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18413704",
   "tg_tokens": 2048,
   "tps": "22.04",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-13",
   "flags": [],
   "hardware": "Arc(TM) 140V",
   "id": "lc-disc-10879-c18422050-fa0",
   "model": "llama 7B",
   "notes": "config=fa0",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "864.99",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "ggml_vulkan: 0 = Intel(R) Arc(TM) 140V GPU (16GB) (Intel Corporation) | uma: 1 | fp16: 1; build: 4a8993735 (10941); llama 7B Q4_0 3.56 GiB 6.74 B Vulkan -1 0 pp512 864.99 \u00b1 7.49; llama 7B Q4_0 3.56 GiB 6.74 B Vulkan -1 0 tg128 24.37 \u00b1 0.17",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #10879 (performance thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18422050",
   "tg_tokens": 128,
   "tps": "24.37",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-13",
   "flags": [],
   "hardware": "Arc(TM) 140V",
   "id": "lc-disc-10879-c18422050-fa1",
   "model": "llama 7B",
   "notes": "config=fa1",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "867.27",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "ggml_vulkan: 0 = Intel(R) Arc(TM) 140V GPU (16GB) (Intel Corporation) | uma: 1 | fp16: 1; build: 4a8993735 (10941); llama 7B Q4_0 3.56 GiB 6.74 B Vulkan -1 1 pp512 867.27 \u00b1 2.35; llama 7B Q4_0 3.56 GiB 6.74 B Vulkan -1 1 tg128 25.12 \u00b1 0.10",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #10879 (performance thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18422050",
   "tg_tokens": 128,
   "tps": "25.12",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "P102-100",
   "id": "lc-disc-15013-c18693775-fa0",
   "model": "llama 7B",
   "notes": "config=fa0-5gb",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "927.54",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Device 0: NVIDIA P102-100, compute capability 6.1, VMM: yes, VRAM: 5119 MiB; build: 0c1e57098 (11312); llama 7B Q4_0 3.56 GiB 6.74 B CUDA 99 0 pp512 927.54 \u00b1 6.49; llama 7B Q4_0 3.56 GiB 6.74 B CUDA 99 0 tg128 46.45 \u00b1 0.13",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #15013 (performance thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-18693775",
   "tg_tokens": 128,
   "tps": "46.45",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "P102-100",
   "id": "lc-disc-15013-c18693775-fa1",
   "model": "llama 7B",
   "notes": "config=fa1-5gb",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "1029.43",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Device 0: NVIDIA P102-100, compute capability 6.1, VMM: yes, VRAM: 5119 MiB; build: 0c1e57098 (11312); llama 7B Q4_0 3.56 GiB 6.74 B CUDA 99 1 pp512 1029.43 \u00b1 7.16; llama 7B Q4_0 3.56 GiB 6.74 B CUDA 99 1 tg128 49.08 \u00b1 0.12",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #15013 (performance thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-18693775",
   "tg_tokens": 128,
   "tps": "49.08",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "P102-100",
   "id": "lc-disc-15013-c18695871-fa0",
   "model": "llama 7B",
   "notes": "config=fa0-10gb",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "935.02",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Device 0: NVIDIA P102-100, compute capability 6.1, VMM: yes, VRAM: 10239 MiB; build: 0c1e57098 (11312); llama 7B Q4_0 3.56 GiB 6.74 B CUDA 99 0 pp512 935.02 \u00b1 3.74; llama 7B Q4_0 3.56 GiB 6.74 B CUDA 99 0 tg128 46.77 \u00b1 0.16",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #15013 (performance thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-18695871",
   "tg_tokens": 128,
   "tps": "46.77",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "P102-100",
   "id": "lc-disc-15013-c18695871-fa1",
   "model": "llama 7B",
   "notes": "config=fa1-10gb",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "1032.13",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Device 0: NVIDIA P102-100, compute capability 6.1, VMM: yes, VRAM: 10239 MiB; build: 0c1e57098 (11312); llama 7B Q4_0 3.56 GiB 6.74 B CUDA 99 1 pp512 1032.13 \u00b1 1.14; llama 7B Q4_0 3.56 GiB 6.74 B CUDA 99 1 tg128 49.59 \u00b1 0.12",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #15013 (performance thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15013#discussioncomment-18695871",
   "tg_tokens": 128,
   "tps": "49.59",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": 12,
   "ctx": null,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "lc-disc-4167-c18683605-quant-f16",
   "model": "llama 7B",
   "notes": "config=quant-f16",
   "params": "6.74B",
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "5207.14",
   "provenance": "community",
   "quant": "F16",
   "quote": "Mac Studio M5 Ultra 36/80 (256GB); Chip: Apple M5 Ultra; Total Number of Cores: 36 (12 Super and 24 Performance); Memory: 256 GB; c1d0e7a00 llama.cpp : bump version to 0.3.0; llama 7B F16 12.55 GiB 6.74 B MTL,BLAS 12 pp512 5207.14 \u00b1 11.61; llama 7B F16 12.55 GiB 6.74 B MTL,BLAS 12 tg128 73.6 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #4167 (performance thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-18683605",
   "tg_tokens": 128,
   "tps": "73.6",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": 12,
   "ctx": null,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "lc-disc-4167-c18683605-quant-q8_0",
   "model": "llama 7B",
   "notes": "config=quant-q8_0",
   "params": "6.74B",
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "4910.55",
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Mac Studio M5 Ultra 36/80 (256GB); Chip: Apple M5 Ultra; Total Number of Cores: 36 (12 Super and 24 Performance); Memory: 256 GB; c1d0e7a00 llama.cpp : bump version to 0.3.0; llama 7B Q8_0 6.67 GiB 6.74 B MTL,BLAS 12 pp512 4910.55 \u00b1 14.90; llama 7B Q8_0 6.67 GiB 6.74 B MTL,BLAS 12 tg128 119.01 \u00b1 0.14",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #4167 (performance thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-18683605",
   "tg_tokens": 128,
   "tps": "119.01",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": 12,
   "ctx": null,
   "date": "2026-09-30",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "lc-disc-4167-c18683605-quant-q4_0",
   "model": "llama 7B",
   "notes": "config=quant-q4_0",
   "params": "6.74B",
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "4944.66",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Mac Studio M5 Ultra 36/80 (256GB); Chip: Apple M5 Ultra; Total Number of Cores: 36 (12 Super and 24 Performance); Memory: 256 GB; c1d0e7a00 llama.cpp : bump version to 0.3.0; llama 7B Q4_0 3.53 GiB 6.74 B MTL,BLAS 12 pp512 4944.66 \u00b1 17.61; llama 7B Q4_0 3.53 GiB 6.74 B MTL,BLAS 12 tg128 179.1 \u00b1 2.45",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #4167 (performance thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-18683605",
   "tg_tokens": 128,
   "tps": "179.1",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": 2,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "M6",
   "id": "lc-disc-4167-c18694932-quant-f16",
   "model": "llama 7B",
   "notes": "config=quant-f16",
   "params": "6.74B",
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "952.53",
   "provenance": "community",
   "quant": "F16",
   "quote": "Mac mini M6, 12c GPU, 32 GB RAM (BW 170GB/s), macOS 27.0.1; build: 0c1e57098; llama 7B F16 12.55 GiB 6.74 B BLAS,MTL 2 pp512 952.53 \u00b1 1.53; llama 7B F16 12.55 GiB 6.74 B BLAS,MTL 2 tg128 11.59 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #4167 (performance thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-18694932",
   "tg_tokens": 128,
   "tps": "11.59",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": 2,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "M6",
   "id": "lc-disc-4167-c18694932-quant-q8_0",
   "model": "llama 7B",
   "notes": "config=quant-q8_0",
   "params": "6.74B",
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "970.46",
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Mac mini M6, 12c GPU, 32 GB RAM (BW 170GB/s), macOS 27.0.1; build: 0c1e57098; llama 7B Q8_0 12.55 GiB 6.74 B BLAS,MTL 2 pp512 970.46 \u00b1 2.67; llama 7B Q8_0 12.55 GiB 6.74 B BLAS,MTL 2 tg128 21.01 \u00b1 0.05",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #4167 (performance thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-18694932",
   "tg_tokens": 128,
   "tps": "21.01",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": 2,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "M6",
   "id": "lc-disc-4167-c18694932-quant-q4_0",
   "model": "llama 7B",
   "notes": "config=quant-q4_0",
   "params": "6.74B",
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "988.93",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Mac mini M6, 12c GPU, 32 GB RAM (BW 170GB/s), macOS 27.0.1; build: 0c1e57098; llama 7B Q4_0 12.55 GiB 6.74 B BLAS,MTL 2 pp512 988.93 \u00b1 5.22; llama 7B Q4_0 12.55 GiB 6.74 B BLAS,MTL 2 tg128 35.74 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #4167 (performance thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-18694932",
   "tg_tokens": 128,
   "tps": "35.74",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "M2 Max (30 GPU)",
   "id": "lc-disc-4167-c18631101-code-off",
   "model": "Qwen3.8-27B",
   "notes": "config=code-off",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ3_S",
   "quote": "I tested DFlash2 on the same M2 Max (30 GPU, 32GB); Qwen3.8-27B dense, UD-IQ3_S; Built from xsn/dflash2 (4e2f54f, Metal); --spec-draft-n-max 7; code 10.53 10.18 0.97x 76%",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #4167 (performance thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-18631101",
   "tg_tokens": 128,
   "tps": "10.53",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "M2 Max (30 GPU)",
   "id": "lc-disc-4167-c18631101-code-dflash",
   "model": "Qwen3.8-27B",
   "notes": "config=code-dflash",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ3_S",
   "quote": "I tested DFlash2 on the same M2 Max (30 GPU, 32GB); Qwen3.8-27B dense, UD-IQ3_S; Built from xsn/dflash2 (4e2f54f, Metal); --spec-draft-n-max 7; code 10.53 10.18 0.97x 76%",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #4167 (performance thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-18631101",
   "tg_tokens": 128,
   "tps": "10.18",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "M2 Max (30 GPU)",
   "id": "lc-disc-4167-c18631101-math-off",
   "model": "Qwen3.8-27B",
   "notes": "config=math-off",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ3_S",
   "quote": "I tested DFlash2 on the same M2 Max (30 GPU, 32GB); Qwen3.8-27B dense, UD-IQ3_S; Built from xsn/dflash2 (4e2f54f, Metal); --spec-draft-n-max 7; math 10.3 9.81 0.95x 72%",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #4167 (performance thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-18631101",
   "tg_tokens": 128,
   "tps": "10.3",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "M2 Max (30 GPU)",
   "id": "lc-disc-4167-c18631101-math-dflash",
   "model": "Qwen3.8-27B",
   "notes": "config=math-dflash",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ3_S",
   "quote": "I tested DFlash2 on the same M2 Max (30 GPU, 32GB); Qwen3.8-27B dense, UD-IQ3_S; Built from xsn/dflash2 (4e2f54f, Metal); --spec-draft-n-max 7; math 10.3 9.81 0.95x 72%",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #4167 (performance thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-18631101",
   "tg_tokens": 128,
   "tps": "9.81",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "M2 Max (30 GPU)",
   "id": "lc-disc-4167-c18631101-prose-off",
   "model": "Qwen3.8-27B",
   "notes": "config=prose-off",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ3_S",
   "quote": "I tested DFlash2 on the same M2 Max (30 GPU, 32GB); Qwen3.8-27B dense, UD-IQ3_S; Built from xsn/dflash2 (4e2f54f, Metal); --spec-draft-n-max 7; prose 10.29 6.13 0.60x 35%",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #4167 (performance thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-18631101",
   "tg_tokens": 128,
   "tps": "10.29",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-28",
   "flags": [],
   "hardware": "M2 Max (30 GPU)",
   "id": "lc-disc-4167-c18631101-prose-dflash",
   "model": "Qwen3.8-27B",
   "notes": "config=prose-dflash",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ3_S",
   "quote": "I tested DFlash2 on the same M2 Max (30 GPU, 32GB); Qwen3.8-27B dense, UD-IQ3_S; Built from xsn/dflash2 (4e2f54f, Metal); --spec-draft-n-max 7; prose 10.29 6.13 0.60x 35%",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #4167 (performance thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-18631101",
   "tg_tokens": 128,
   "tps": "6.13",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "RX 6800",
   "id": "lc-disc-10879-c18706574-vulkan-gptoss",
   "model": "gpt-oss 20B",
   "notes": "config=vulkan-gptoss",
   "params": "23.6B",
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "1883",
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "AMD Radeon RX 6800 (16 GB GDDR6, 256-bit), Arch Linux (kernel 7.2.5), Mesa RADV 26.2.2, Ryzen 7 5700X, Resizable BAR on; Larger models on the same card ( -ngl 99 -fa 1 , pp512 / tg128); gpt-oss-20b MXFP4 1883 / 145.9 2022 / 121.2",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18706574",
   "tg_tokens": 128,
   "tps": "145.9",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "RX 6800",
   "id": "lc-disc-10879-c18706574-rocm-gptoss",
   "model": "gpt-oss 20B",
   "notes": "config=rocm-gptoss",
   "params": "23.6B",
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "2022",
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "AMD Radeon RX 6800 (16 GB GDDR6, 256-bit), Arch Linux (kernel 7.2.5), Mesa RADV 26.2.2, Ryzen 7 5700X, Resizable BAR on; Larger models on the same card ( -ngl 99 -fa 1 , pp512 / tg128); gpt-oss-20b MXFP4 1883 / 145.9 2022 / 121.2",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18706574",
   "tg_tokens": 128,
   "tps": "121.2",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "RX 6800",
   "id": "lc-disc-10879-c18706574-vulkan-qwen38",
   "model": "Qwen3.8-27B",
   "notes": "config=vulkan-qwen38",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "356",
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "AMD Radeon RX 6800 (16 GB GDDR6, 256-bit), Arch Linux (kernel 7.2.5), Mesa RADV 26.2.2, Ryzen 7 5700X, Resizable BAR on; Larger models on the same card ( -ngl 99 -fa 1 , pp512 / tg128); Qwen3.8-27B UD-IQ4_XS 356 / 28.4 395 / 26.1",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18706574",
   "tg_tokens": 128,
   "tps": "28.4",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "RX 6800",
   "id": "lc-disc-10879-c18706574-rocm-qwen38",
   "model": "Qwen3.8-27B",
   "notes": "config=rocm-qwen38",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "395",
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "AMD Radeon RX 6800 (16 GB GDDR6, 256-bit), Arch Linux (kernel 7.2.5), Mesa RADV 26.2.2, Ryzen 7 5700X, Resizable BAR on; Larger models on the same card ( -ngl 99 -fa 1 , pp512 / tg128); Qwen3.8-27B UD-IQ4_XS 356 / 28.4 395 / 26.1",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18706574",
   "tg_tokens": 128,
   "tps": "26.1",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "RX 6800",
   "id": "lc-disc-10879-c18706574-vulkan-gemma4",
   "model": "Gemma 4 12B",
   "notes": "config=vulkan-gemma4",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "799",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "AMD Radeon RX 6800 (16 GB GDDR6, 256-bit), Arch Linux (kernel 7.2.5), Mesa RADV 26.2.2, Ryzen 7 5700X, Resizable BAR on; Larger models on the same card ( -ngl 99 -fa 1 , pp512 / tg128); Gemma 4 12B Q4_K_M 799 / 52.1 736 / 48.9",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18706574",
   "tg_tokens": 128,
   "tps": "52.1",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "RX 6800",
   "id": "lc-disc-10879-c18706574-rocm-gemma4",
   "model": "Gemma 4 12B",
   "notes": "config=rocm-gemma4",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "736",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "AMD Radeon RX 6800 (16 GB GDDR6, 256-bit), Arch Linux (kernel 7.2.5), Mesa RADV 26.2.2, Ryzen 7 5700X, Resizable BAR on; Larger models on the same card ( -ngl 99 -fa 1 , pp512 / tg128); Gemma 4 12B Q4_K_M 799 / 52.1 736 / 48.9",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18706574",
   "tg_tokens": 128,
   "tps": "48.9",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-27593-f16-off",
   "model": "Qwen3.8-27B",
   "notes": "config=f16-off | llama.cpp discussion #27593, Arc Pro B70 (BMG G31, 32 GB), Qwen3.8-27B Q8_0, master 63b64a50a, SYCL (oneAPI 2026.1).",
   "params": null,
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": 389.2,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B70 Qwen3.8-27B Q8_0: -DGGML_SYCL_F16=OFF (CMake default) pp2048 389.2 pp8192 385.3 tg128 15.80",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #27593 (Arc Pro B70 tuning thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/27593",
   "tg_tokens": 128,
   "tps": 15.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-27593-f16-on",
   "model": "Qwen3.8-27B",
   "notes": "config=f16-on | llama.cpp discussion #27593, Arc Pro B70 (BMG G31, 32 GB), Qwen3.8-27B Q8_0, master 63b64a50a, SYCL (oneAPI 2026.1).",
   "params": null,
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": 1446.8,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B70 Qwen3.8-27B Q8_0: -DGGML_SYCL_F16=ON pp2048 1446.8 pp8192 1388.2 tg128 15.79",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #27593 (Arc Pro B70 tuning thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/27593",
   "tg_tokens": 128,
   "tps": 15.79,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-27593-mtp-short-nmax-off",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-short-nmax-off | llama.cpp discussion #27593, Arc Pro B70 (BMG G31, 32 GB), Qwen3.8-27B Q8_0, master 63b64a50a, SYCL (oneAPI 2026.1).",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B70 Qwen3.8-27B Q8_0: --spec-draft-n-max off short answer (22 tokens) 15.73 t/s, acceptance None",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #27593 (Arc Pro B70 tuning thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/27593",
   "tg_tokens": 22,
   "tps": 15.73,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-27593-mtp-short-nmax-1",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-short-nmax-1 | llama.cpp discussion #27593, Arc Pro B70 (BMG G31, 32 GB), Qwen3.8-27B Q8_0, master 63b64a50a, SYCL (oneAPI 2026.1).",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B70 Qwen3.8-27B Q8_0: --spec-draft-n-max 1 short answer (22 tokens) 26.69 t/s, acceptance 0.909",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #27593 (Arc Pro B70 tuning thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/27593",
   "tg_tokens": 22,
   "tps": 26.69,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-27593-mtp-short-nmax-2",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-short-nmax-2 | llama.cpp discussion #27593, Arc Pro B70 (BMG G31, 32 GB), Qwen3.8-27B Q8_0, master 63b64a50a, SYCL (oneAPI 2026.1).",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B70 Qwen3.8-27B Q8_0: --spec-draft-n-max 2 short answer (22 tokens) 33.52 t/s, acceptance 0.812",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #27593 (Arc Pro B70 tuning thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/27593",
   "tg_tokens": 22,
   "tps": 33.52,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-27593-mtp-short-nmax-3",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-short-nmax-3 | llama.cpp discussion #27593, Arc Pro B70 (BMG G31, 32 GB), Qwen3.8-27B Q8_0, master 63b64a50a, SYCL (oneAPI 2026.1).",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B70 Qwen3.8-27B Q8_0: --spec-draft-n-max 3 short answer (22 tokens) 41.35 t/s, acceptance 0.833",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #27593 (Arc Pro B70 tuning thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/27593",
   "tg_tokens": 22,
   "tps": 41.35,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-27593-mtp-short-nmax-4",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-short-nmax-4 | llama.cpp discussion #27593, Arc Pro B70 (BMG G31, 32 GB), Qwen3.8-27B Q8_0, master 63b64a50a, SYCL (oneAPI 2026.1).",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B70 Qwen3.8-27B Q8_0: --spec-draft-n-max 4 short answer (22 tokens) 46.11 t/s, acceptance 0.8",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #27593 (Arc Pro B70 tuning thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/27593",
   "tg_tokens": 22,
   "tps": 46.11,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-27593-mtp-short-nmax-6",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-short-nmax-6 | llama.cpp discussion #27593, Arc Pro B70 (BMG G31, 32 GB), Qwen3.8-27B Q8_0, master 63b64a50a, SYCL (oneAPI 2026.1).",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B70 Qwen3.8-27B Q8_0: --spec-draft-n-max 6 short answer (22 tokens) 50.92 t/s, acceptance 0.708",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #27593 (Arc Pro B70 tuning thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/27593",
   "tg_tokens": 22,
   "tps": 50.92,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-27593-mtp-long-nmax-off",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-long-nmax-off | llama.cpp discussion #27593, Arc Pro B70 (BMG G31, 32 GB), Qwen3.8-27B Q8_0, master 63b64a50a, SYCL (oneAPI 2026.1).",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B70 Qwen3.8-27B Q8_0: --spec-draft-n-max off long answer (512 tokens) 15.67 t/s, acceptance None",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #27593 (Arc Pro B70 tuning thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/27593",
   "tg_tokens": 512,
   "tps": 15.67,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-27593-mtp-long-nmax-1",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-long-nmax-1 | llama.cpp discussion #27593, Arc Pro B70 (BMG G31, 32 GB), Qwen3.8-27B Q8_0, master 63b64a50a, SYCL (oneAPI 2026.1).",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B70 Qwen3.8-27B Q8_0: --spec-draft-n-max 1 long answer (512 tokens) 25.99 t/s, acceptance 0.868",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #27593 (Arc Pro B70 tuning thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/27593",
   "tg_tokens": 512,
   "tps": 25.99,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-27593-mtp-long-nmax-2",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-long-nmax-2 | llama.cpp discussion #27593, Arc Pro B70 (BMG G31, 32 GB), Qwen3.8-27B Q8_0, master 63b64a50a, SYCL (oneAPI 2026.1).",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B70 Qwen3.8-27B Q8_0: --spec-draft-n-max 2 long answer (512 tokens) 30.62 t/s, acceptance 0.701",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #27593 (Arc Pro B70 tuning thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/27593",
   "tg_tokens": 512,
   "tps": 30.62,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-27593-mtp-long-nmax-3",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-long-nmax-3 | llama.cpp discussion #27593, Arc Pro B70 (BMG G31, 32 GB), Qwen3.8-27B Q8_0, master 63b64a50a, SYCL (oneAPI 2026.1).",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B70 Qwen3.8-27B Q8_0: --spec-draft-n-max 3 long answer (512 tokens) 33.53 t/s, acceptance 0.615",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #27593 (Arc Pro B70 tuning thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/27593",
   "tg_tokens": 512,
   "tps": 33.53,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-27593-mtp-long-nmax-4",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-long-nmax-4 | llama.cpp discussion #27593, Arc Pro B70 (BMG G31, 32 GB), Qwen3.8-27B Q8_0, master 63b64a50a, SYCL (oneAPI 2026.1).",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B70 Qwen3.8-27B Q8_0: --spec-draft-n-max 4 long answer (512 tokens) 33.23 t/s, acceptance 0.507",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #27593 (Arc Pro B70 tuning thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/27593",
   "tg_tokens": 512,
   "tps": 33.23,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-27593-mtp-long-nmax-6",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-long-nmax-6 | llama.cpp discussion #27593, Arc Pro B70 (BMG G31, 32 GB), Qwen3.8-27B Q8_0, master 63b64a50a, SYCL (oneAPI 2026.1).",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B70 Qwen3.8-27B Q8_0: --spec-draft-n-max 6 long answer (512 tokens) 32.5 t/s, acceptance 0.395",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #27593 (Arc Pro B70 tuning thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/27593",
   "tg_tokens": 512,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-27593-b2048-ub512",
   "model": "Qwen3.8-27B",
   "notes": "config=b2048-ub512 | llama.cpp discussion #27593, Arc Pro B70 (BMG G31, 32 GB), Qwen3.8-27B Q8_0, master 63b64a50a, SYCL (oneAPI 2026.1).",
   "params": null,
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": 1074.8,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B70 Qwen3.8-27B Q8_0: -b 2048 -ub 512 pp2048 1074.8",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #27593 (Arc Pro B70 tuning thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/27593",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-27593-b2048-ub1024",
   "model": "Qwen3.8-27B",
   "notes": "config=b2048-ub1024 | llama.cpp discussion #27593, Arc Pro B70 (BMG G31, 32 GB), Qwen3.8-27B Q8_0, master 63b64a50a, SYCL (oneAPI 2026.1).",
   "params": null,
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": 1304.7,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B70 Qwen3.8-27B Q8_0: -b 2048 -ub 1024 pp2048 1304.7",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #27593 (Arc Pro B70 tuning thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/27593",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-27593-b2048-ub2048",
   "model": "Qwen3.8-27B",
   "notes": "config=b2048-ub2048 | llama.cpp discussion #27593, Arc Pro B70 (BMG G31, 32 GB), Qwen3.8-27B Q8_0, master 63b64a50a, SYCL (oneAPI 2026.1).",
   "params": null,
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": 1448.7,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B70 Qwen3.8-27B Q8_0: -b 2048 -ub 2048 pp2048 1448.7",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #27593 (Arc Pro B70 tuning thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/27593",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-27593-b4096-ub512",
   "model": "Qwen3.8-27B",
   "notes": "config=b4096-ub512 | llama.cpp discussion #27593, Arc Pro B70 (BMG G31, 32 GB), Qwen3.8-27B Q8_0, master 63b64a50a, SYCL (oneAPI 2026.1).",
   "params": null,
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": 1066.1,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B70 Qwen3.8-27B Q8_0: -b 4096 -ub 512 pp2048 1066.1",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #27593 (Arc Pro B70 tuning thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/27593",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-27593-b4096-ub1024",
   "model": "Qwen3.8-27B",
   "notes": "config=b4096-ub1024 | llama.cpp discussion #27593, Arc Pro B70 (BMG G31, 32 GB), Qwen3.8-27B Q8_0, master 63b64a50a, SYCL (oneAPI 2026.1).",
   "params": null,
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": 1302.4,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B70 Qwen3.8-27B Q8_0: -b 4096 -ub 1024 pp2048 1302.4",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #27593 (Arc Pro B70 tuning thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/27593",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-27593-b4096-ub2048",
   "model": "Qwen3.8-27B",
   "notes": "config=b4096-ub2048 | llama.cpp discussion #27593, Arc Pro B70 (BMG G31, 32 GB), Qwen3.8-27B Q8_0, master 63b64a50a, SYCL (oneAPI 2026.1).",
   "params": null,
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": 1444.8,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B70 Qwen3.8-27B Q8_0: -b 4096 -ub 2048 pp2048 1444.8",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #27593 (Arc Pro B70 tuning thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/27593",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-27593-b4096-ub4096",
   "model": "Qwen3.8-27B",
   "notes": "config=b4096-ub4096 | llama.cpp discussion #27593, Arc Pro B70 (BMG G31, 32 GB), Qwen3.8-27B Q8_0, master 63b64a50a, SYCL (oneAPI 2026.1).",
   "params": null,
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": 1445.1,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B70 Qwen3.8-27B Q8_0: -b 4096 -ub 4096 pp2048 1445.1",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #27593 (Arc Pro B70 tuning thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/27593",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-27593-fa0",
   "model": "Qwen3.8-27B",
   "notes": "config=fa0 | llama.cpp discussion #27593, Arc Pro B70 (BMG G31, 32 GB), Qwen3.8-27B Q8_0, master 63b64a50a, SYCL (oneAPI 2026.1).",
   "params": null,
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": 1361.6,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B70 Qwen3.8-27B Q8_0: -fa 0 pp2048 1361.6 pp8192 1270.3 tg128 15.81",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #27593 (Arc Pro B70 tuning thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/27593",
   "tg_tokens": 128,
   "tps": 15.81,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-27593-fa1",
   "model": "Qwen3.8-27B",
   "notes": "config=fa1 | llama.cpp discussion #27593, Arc Pro B70 (BMG G31, 32 GB), Qwen3.8-27B Q8_0, master 63b64a50a, SYCL (oneAPI 2026.1).",
   "params": null,
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": 1446.8,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B70 Qwen3.8-27B Q8_0: -fa 1 pp2048 1446.8 pp8192 1388.2 tg128 15.79",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #27593 (Arc Pro B70 tuning thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/27593",
   "tg_tokens": 128,
   "tps": 15.79,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 49152,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-27593-kv-f16",
   "model": "Qwen3.8-27B",
   "notes": "config=kv-f16 | llama.cpp discussion #27593, Arc Pro B70 (BMG G31, 32 GB), Qwen3.8-27B Q8_0, master 63b64a50a, SYCL (oneAPI 2026.1).",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1115.7,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B70 Qwen3.8-27B Q8_0: KV f16, 49152 context, prefill 1115.7, tg128 15.79",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #27593 (Arc Pro B70 tuning thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/27593",
   "tg_tokens": 128,
   "tps": 15.79,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 65536,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-27593-kv-q8_0",
   "model": "Qwen3.8-27B",
   "notes": "config=kv-q8_0 | llama.cpp discussion #27593, Arc Pro B70 (BMG G31, 32 GB), Qwen3.8-27B Q8_0, master 63b64a50a, SYCL (oneAPI 2026.1).",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1032.6,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B70 Qwen3.8-27B Q8_0: KV q8_0, 65536 context, prefill 1032.6, tg128 15.72",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #27593 (Arc Pro B70 tuning thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/27593",
   "tg_tokens": 128,
   "tps": 15.72,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 131072,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-27593-kv-q4_0",
   "model": "Qwen3.8-27B",
   "notes": "config=kv-q4_0 | llama.cpp discussion #27593, Arc Pro B70 (BMG G31, 32 GB), Qwen3.8-27B Q8_0, master 63b64a50a, SYCL (oneAPI 2026.1).",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 795.3,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B70 Qwen3.8-27B Q8_0: KV q4_0, 131072 context, prefill 795.3, tg128 15.69",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #27593 (Arc Pro B70 tuning thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/27593",
   "tg_tokens": 128,
   "tps": 15.69,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-27593-env-baseline",
   "model": "Qwen3.8-27B",
   "notes": "config=env-baseline | llama.cpp discussion #27593, Arc Pro B70 (BMG G31, 32 GB), Qwen3.8-27B Q8_0, master 63b64a50a, SYCL (oneAPI 2026.1).",
   "params": null,
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": 1438.3,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B70 Qwen3.8-27B Q8_0: baseline -ngl 99 -b 4096 -ub 2048 -fa 1 pp2048 1438.3 tg128 15.79",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #27593 (Arc Pro B70 tuning thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/27593",
   "tg_tokens": 128,
   "tps": 15.79,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-27593-env-enable-opt-0",
   "model": "Qwen3.8-27B",
   "notes": "config=env-enable-opt-0 | llama.cpp discussion #27593, Arc Pro B70 (BMG G31, 32 GB), Qwen3.8-27B Q8_0, master 63b64a50a, SYCL (oneAPI 2026.1).",
   "params": null,
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": 1442.4,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B70 Qwen3.8-27B Q8_0: GGML_SYCL_ENABLE_OPT=0 pp2048 1442.4 tg128 5.07",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #27593 (Arc Pro B70 tuning thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/27593",
   "tg_tokens": 128,
   "tps": 5.07,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-27593-env-dmmv-1",
   "model": "Qwen3.8-27B",
   "notes": "config=env-dmmv-1 | llama.cpp discussion #27593, Arc Pro B70 (BMG G31, 32 GB), Qwen3.8-27B Q8_0, master 63b64a50a, SYCL (oneAPI 2026.1).",
   "params": null,
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": 1448.7,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B70 Qwen3.8-27B Q8_0: GGML_SYCL_PRIORITIZE_DMMV=1 pp2048 1448.7 tg128 13.86",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #27593 (Arc Pro B70 tuning thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/27593",
   "tg_tokens": 128,
   "tps": 13.86,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-27593-env-onednn-0",
   "model": "Qwen3.8-27B",
   "notes": "config=env-onednn-0 | llama.cpp discussion #27593, Arc Pro B70 (BMG G31, 32 GB), Qwen3.8-27B Q8_0, master 63b64a50a, SYCL (oneAPI 2026.1).",
   "params": null,
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": 1372.1,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B70 Qwen3.8-27B Q8_0: GGML_SYCL_FA_ONEDNN=0 pp2048 1372.1 tg128 16.34",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #27593 (Arc Pro B70 tuning thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/27593",
   "tg_tokens": 128,
   "tps": 16.34,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-27593-env-l0v2-0",
   "model": "Qwen3.8-27B",
   "notes": "config=env-l0v2-0 | llama.cpp discussion #27593, Arc Pro B70 (BMG G31, 32 GB), Qwen3.8-27B Q8_0, master 63b64a50a, SYCL (oneAPI 2026.1).",
   "params": null,
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": 1450.2,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B70 Qwen3.8-27B Q8_0: SYCL_UR_USE_LEVEL_ZERO_V2=0 pp2048 1450.2 tg128 16.01",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #27593 (Arc Pro B70 tuning thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/27593",
   "tg_tokens": 128,
   "tps": 16.01,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-27593-env-onednn-0+l0v2-0",
   "model": "Qwen3.8-27B",
   "notes": "config=env-onednn-0+l0v2-0 | llama.cpp discussion #27593, Arc Pro B70 (BMG G31, 32 GB), Qwen3.8-27B Q8_0, master 63b64a50a, SYCL (oneAPI 2026.1).",
   "params": null,
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": 1372.2,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B70 Qwen3.8-27B Q8_0: GGML_SYCL_FA_ONEDNN=0 + SYCL_UR_USE_LEVEL_ZERO_V2=0 pp2048 1372.2 tg128 16.59",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #27593 (Arc Pro B70 tuning thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/27593",
   "tg_tokens": 128,
   "tps": 16.59,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-27593-env-fusion-0",
   "model": "Qwen3.8-27B",
   "notes": "config=env-fusion-0 | llama.cpp discussion #27593, Arc Pro B70 (BMG G31, 32 GB), Qwen3.8-27B Q8_0, master 63b64a50a, SYCL (oneAPI 2026.1).",
   "params": null,
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": 1401.0,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B70 Qwen3.8-27B Q8_0: GGML_SYCL_ENABLE_FUSION=0 pp2048 1401.0 tg128 15.52",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #27593 (Arc Pro B70 tuning thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/27593",
   "tg_tokens": 128,
   "tps": 15.52,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-27593-env-usm-system-1",
   "model": "Qwen3.8-27B",
   "notes": "config=env-usm-system-1 | llama.cpp discussion #27593, Arc Pro B70 (BMG G31, 32 GB), Qwen3.8-27B Q8_0, master 63b64a50a, SYCL (oneAPI 2026.1).",
   "params": null,
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": 1434.1,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B70 Qwen3.8-27B Q8_0: GGML_SYCL_USM_SYSTEM=1 pp2048 1434.1 tg128 15.82",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #27593 (Arc Pro B70 tuning thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/27593",
   "tg_tokens": 128,
   "tps": 15.82,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-27593-env-dnn-0",
   "model": "Qwen3.8-27B",
   "notes": "config=env-dnn-0 | llama.cpp discussion #27593, Arc Pro B70 (BMG G31, 32 GB), Qwen3.8-27B Q8_0, master 63b64a50a, SYCL (oneAPI 2026.1).",
   "params": null,
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": 1437.9,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B70 Qwen3.8-27B Q8_0: GGML_SYCL_ENABLE_DNN=0 pp2048 1437.9 tg128 15.79",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #27593 (Arc Pro B70 tuning thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/27593",
   "tg_tokens": 128,
   "tps": 15.79,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-27593-env-vmm-0",
   "model": "Qwen3.8-27B",
   "notes": "config=env-vmm-0 | llama.cpp discussion #27593, Arc Pro B70 (BMG G31, 32 GB), Qwen3.8-27B Q8_0, master 63b64a50a, SYCL (oneAPI 2026.1).",
   "params": null,
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": 1443.9,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B70 Qwen3.8-27B Q8_0: GGML_SYCL_ENABLE_VMM=0 pp2048 1443.9 tg128 15.8",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #27593 (Arc Pro B70 tuning thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/27593",
   "tg_tokens": 128,
   "tps": 15.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-27593-env-esimd-0",
   "model": "Qwen3.8-27B",
   "notes": "config=env-esimd-0 | llama.cpp discussion #27593, Arc Pro B70 (BMG G31, 32 GB), Qwen3.8-27B Q8_0, master 63b64a50a, SYCL (oneAPI 2026.1).",
   "params": null,
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": 1447.7,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B70 Qwen3.8-27B Q8_0: GGML_SYCL_ENABLE_ESIMD=0 pp2048 1447.7 tg128 15.79",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #27593 (Arc Pro B70 tuning thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/27593",
   "tg_tokens": 128,
   "tps": 15.79,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-27593-env-mkl-fa-0",
   "model": "Qwen3.8-27B",
   "notes": "config=env-mkl-fa-0 | llama.cpp discussion #27593, Arc Pro B70 (BMG G31, 32 GB), Qwen3.8-27B Q8_0, master 63b64a50a, SYCL (oneAPI 2026.1).",
   "params": null,
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": 1449.3,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B70 Qwen3.8-27B Q8_0: GGML_SYCL_ENABLE_MKL_FA=0 pp2048 1449.3 tg128 15.79",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #27593 (Arc Pro B70 tuning thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/27593",
   "tg_tokens": 128,
   "tps": 15.79,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-27593-env-graph-1",
   "model": "Qwen3.8-27B",
   "notes": "config=env-graph-1 | llama.cpp discussion #27593, Arc Pro B70 (BMG G31, 32 GB), Qwen3.8-27B Q8_0, master 63b64a50a, SYCL (oneAPI 2026.1).",
   "params": null,
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": 1450.0,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B70 Qwen3.8-27B Q8_0: GGML_SYCL_ENABLE_GRAPH=1 pp2048 1450.0 tg128 15.79",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #27593 (Arc Pro B70 tuning thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/27593",
   "tg_tokens": 128,
   "tps": 15.79,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-27593-env-async-mem-0",
   "model": "Qwen3.8-27B",
   "notes": "config=env-async-mem-0 | llama.cpp discussion #27593, Arc Pro B70 (BMG G31, 32 GB), Qwen3.8-27B Q8_0, master 63b64a50a, SYCL (oneAPI 2026.1).",
   "params": null,
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": 1448.1,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B70 Qwen3.8-27B Q8_0: GGML_SYCL_USE_ASYNC_MEM_OP=0 pp2048 1448.1 tg128 15.79",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #27593 (Arc Pro B70 tuning thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/27593",
   "tg_tokens": 128,
   "tps": 15.79,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-27593-env-no-pinned-1",
   "model": "Qwen3.8-27B",
   "notes": "config=env-no-pinned-1 | llama.cpp discussion #27593, Arc Pro B70 (BMG G31, 32 GB), Qwen3.8-27B Q8_0, master 63b64a50a, SYCL (oneAPI 2026.1).",
   "params": null,
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": 1449.1,
   "provenance": "community",
   "quant": "Q8_0",
   "quote": "Arc Pro B70 Qwen3.8-27B Q8_0: GGML_SYCL_NO_PINNED=1 pp2048 1449.1 tg128 15.79",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #27593 (Arc Pro B70 tuning thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/27593",
   "tg_tokens": 128,
   "tps": 15.79,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 262144,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-27593-c18692471-single-b70",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp discussion #27593 comment 18692471 (AdamN1): Arc Pro B70 llama-server, -b 2048 -fa 1, --spec-type draft-mtp -n-max 3, kv_unified, Qwen3.8-27B-UD-Q4_K_XL; progress-log PP rate at 4096 tokens.",
   "params": null,
   "power_w": null,
   "pp_tokens": 4096,
   "pp_tps": 1149.66,
   "provenance": "community",
   "quant": "Q4_K_XL (UD)",
   "quote": "Arc Pro B70 (single card, n_slots 4, ctx 262144, KV q8_0, MTP k=3) Qwen3.8-27B-UD-Q4_K_XL: prompt processing n_tokens 4096, 1149.66 tokens per second; n_tokens 64121, 869.93 tokens per second",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #27593 (Arc Pro B70 tuning thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/27593#discussioncomment-18692471",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": 262144,
   "date": "2026-10-01",
   "flags": [],
   "hardware": "2x Arc Pro B70",
   "id": "lc-disc-27593-c18692471-2x-b70",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp discussion #27593 comment 18692471 (AdamN1): 2x Arc Pro B70 llama-server, -b 2048 -fa 1, --spec-type draft-mtp -n-max 3, kv_unified, Qwen3.8-27B-UD-Q4_K_XL; progress-log PP rate at 4096 tokens.",
   "params": null,
   "power_w": null,
   "pp_tokens": 4096,
   "pp_tps": 855.54,
   "provenance": "community",
   "quant": "Q4_K_XL (UD)",
   "quote": "2x Arc Pro B70 (tensor 50,50 over PCIe gen4 x16) Qwen3.8-27B-UD-Q4_K_XL: prompt processing n_tokens 4096, 855.54 tokens per second; n_tokens 64121, 667.61 tokens per second",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #27593 (Arc Pro B70 tuning thread)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/27593#discussioncomment-18692471",
   "tg_tokens": null,
   "tps": null,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18467211-llama7b-fa1",
   "model": "llama 7B",
   "notes": "config=llama7b-fa1 | llama.cpp discussion #23313 comment 18467211 (cbelcher): TheBloke llama-2-7b Q4_0, Arc Pro B70 passed through to a Proxmox VM (Xeon E5-2699 v4, 64 GB DDR4), oneAPI 2026.1.1, llama.cpp 0.4.1-dev built with -DGGML_SYCL_F16=ON.",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3271.16,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "llama 7B Q4_0 on Arc Pro B70 (VFIO passthrough, Proxmox VM), SYCL F16 build, llama-bench: fa 1 pp512 3271.16 tg128 111.13",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel Arc)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18467211",
   "tg_tokens": 128,
   "tps": 111.13,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18467211-llama7b-fa0",
   "model": "llama 7B",
   "notes": "config=llama7b-fa0 | llama.cpp discussion #23313 comment 18467211 (cbelcher): same VM setup, fa=0 arm.",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1324.76,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "llama 7B Q4_0 on Arc Pro B70 (VFIO passthrough, Proxmox VM), SYCL F16 build, llama-bench: fa 0 pp512 1324.76 tg128 107.11",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel Arc)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18467211",
   "tg_tokens": 128,
   "tps": 107.11,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18629941-d0",
   "model": "Qwen3.8-27B",
   "notes": "config=d0 | llama.cpp discussion #23313 comment 18629941 (NickM-27): qwen35 27B Q6_K, SYCL, prompt-depth decay at fixed pp2048/tg256.",
   "params": null,
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": 1212.47,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "qwen35 27B Q6_K (Qwen3.8-27B) on Arc Pro B70, latest SYCL, n_ubatch 1024: pp2048 1212.47 tg256 21.32",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel Arc)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18629941",
   "tg_tokens": 256,
   "tps": 21.32,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18629941-d8192",
   "model": "Qwen3.8-27B",
   "notes": "config=d8192 | llama.cpp discussion #23313 comment 18629941 (NickM-27): qwen35 27B Q6_K, SYCL, prompt-depth decay at fixed pp2048/tg256.",
   "params": null,
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": 1029.24,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "qwen35 27B Q6_K (Qwen3.8-27B) on Arc Pro B70, latest SYCL, n_ubatch 1024: pp2048 @ d8192 1029.24 tg256 @ d8192 19.6",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel Arc)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18629941",
   "tg_tokens": 256,
   "tps": 19.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18629941-d16384",
   "model": "Qwen3.8-27B",
   "notes": "config=d16384 | llama.cpp discussion #23313 comment 18629941 (NickM-27): qwen35 27B Q6_K, SYCL, prompt-depth decay at fixed pp2048/tg256.",
   "params": null,
   "power_w": null,
   "pp_tokens": 2048,
   "pp_tps": 968.59,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "qwen35 27B Q6_K (Qwen3.8-27B) on Arc Pro B70, latest SYCL, n_ubatch 1024: pp2048 @ d16384 968.59 tg256 @ d16384 18.15",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel Arc)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18629941",
   "tg_tokens": 256,
   "tps": 18.15,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18629941-mtp-code_python",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-code_python | llama.cpp discussion #23313 comment 18629941 (NickM-27): mtp-bench (multi-token-prediction spec decode), code_python prompt type, draft acceptance rate 0.936.",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "qwen35 27B Q6_K (Qwen3.8-27B) on Arc Pro B70, latest SYCL: mtp-bench code_python rate=0.936 tok/s=40.9",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel Arc)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18629941",
   "tg_tokens": null,
   "tps": 40.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18629941-mtp-code_cpp",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-code_cpp | llama.cpp discussion #23313 comment 18629941 (NickM-27): mtp-bench (multi-token-prediction spec decode), code_cpp prompt type, draft acceptance rate 0.875.",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "qwen35 27B Q6_K (Qwen3.8-27B) on Arc Pro B70, latest SYCL: mtp-bench code_cpp rate=0.875 tok/s=38.5",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel Arc)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18629941",
   "tg_tokens": null,
   "tps": 38.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18629941-mtp-explain_concept",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-explain_concept | llama.cpp discussion #23313 comment 18629941 (NickM-27): mtp-bench (multi-token-prediction spec decode), explain_concept prompt type, draft acceptance rate 0.687.",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "qwen35 27B Q6_K (Qwen3.8-27B) on Arc Pro B70, latest SYCL: mtp-bench explain_concept rate=0.687 tok/s=29.0",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel Arc)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18629941",
   "tg_tokens": null,
   "tps": 29.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18629941-mtp-summarize",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-summarize | llama.cpp discussion #23313 comment 18629941 (NickM-27): mtp-bench (multi-token-prediction spec decode), summarize prompt type, draft acceptance rate 0.708.",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "qwen35 27B Q6_K (Qwen3.8-27B) on Arc Pro B70, latest SYCL: mtp-bench summarize rate=0.708 tok/s=32.1",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel Arc)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18629941",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18629941-mtp-qa_factual",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-qa_factual | llama.cpp discussion #23313 comment 18629941 (NickM-27): mtp-bench (multi-token-prediction spec decode), qa_factual prompt type, draft acceptance rate 0.767.",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "qwen35 27B Q6_K (Qwen3.8-27B) on Arc Pro B70, latest SYCL: mtp-bench qa_factual rate=0.767 tok/s=32.9",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel Arc)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18629941",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18629941-mtp-translation",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-translation | llama.cpp discussion #23313 comment 18629941 (NickM-27): mtp-bench (multi-token-prediction spec decode), translation prompt type, draft acceptance rate 0.765.",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "qwen35 27B Q6_K (Qwen3.8-27B) on Arc Pro B70, latest SYCL: mtp-bench translation rate=0.765 tok/s=29.5",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel Arc)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18629941",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18629941-mtp-creative_short",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-creative_short | llama.cpp discussion #23313 comment 18629941 (NickM-27): mtp-bench (multi-token-prediction spec decode), creative_short prompt type, draft acceptance rate 0.9.",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "qwen35 27B Q6_K (Qwen3.8-27B) on Arc Pro B70, latest SYCL: mtp-bench creative_short rate=0.9 tok/s=28.1",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel Arc)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18629941",
   "tg_tokens": null,
   "tps": 28.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18629941-mtp-stepwise_math",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-stepwise_math | llama.cpp discussion #23313 comment 18629941 (NickM-27): mtp-bench (multi-token-prediction spec decode), stepwise_math prompt type, draft acceptance rate 0.823.",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "qwen35 27B Q6_K (Qwen3.8-27B) on Arc Pro B70, latest SYCL: mtp-bench stepwise_math rate=0.823 tok/s=35.5",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel Arc)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18629941",
   "tg_tokens": null,
   "tps": 35.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "Arc Pro B70",
   "id": "lc-disc-23313-c18629941-mtp-long_code_review",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-long_code_review | llama.cpp discussion #23313 comment 18629941 (NickM-27): mtp-bench (multi-token-prediction spec decode), long_code_review prompt type, draft acceptance rate 0.653.",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q6_K",
   "quote": "qwen35 27B Q6_K (Qwen3.8-27B) on Arc Pro B70, latest SYCL: mtp-bench long_code_review rate=0.653 tok/s=28.0",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel Arc)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18629941",
   "tg_tokens": null,
   "tps": 28.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-22",
   "flags": [],
   "hardware": "M2 Max (30 GPU)",
   "id": "lc-disc-4167-c18555918-q36a3b",
   "model": "Qwen3.6-35B-A3B",
   "notes": "config=q36a3b; M2 Max 32GB; pp4096/tg128; MoE 3B active, larger file than the dense 27B but ~6x faster prefill",
   "params": null,
   "power_w": null,
   "pp_tokens": 4096,
   "pp_tps": "671.40",
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "M2 Max (30 GPU, 32GB), GPU otherwise idle; Qwen3.6-35B-A3B (MoE, 3B active) Q4_K_M 20.22 GiB pp4096 671.40 \u00b1 11.20, tg128 58.56 \u00b1 0.82",
   "ram_gb": 32,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-18555918",
   "tg_tokens": 128,
   "tps": "58.56",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-22",
   "flags": [],
   "hardware": "M2 Max (30 GPU)",
   "id": "lc-disc-4167-c18555918-q38-27b",
   "model": "Qwen3.8-27B",
   "notes": "config=iq3s-pp4096; M2 Max 32GB; pp4096/tg128; dense 27B, architecture swamps quantization (5.4x slower generation than the MoE 35B-A3B)",
   "params": null,
   "power_w": null,
   "pp_tokens": 4096,
   "pp_tps": "106.78",
   "provenance": "community",
   "quant": "IQ3_S",
   "quote": "M2 Max (30 GPU, 32GB); Qwen3.8-27B (dense) IQ3_S 11.20 GiB pp4096 106.78 \u00b1 0.34, tg128 10.82 \u00b1 0.06",
   "ram_gb": 32,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-18555918",
   "tg_tokens": 128,
   "tps": "10.82",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-22",
   "flags": [],
   "hardware": "M2 Max (30 GPU)",
   "id": "lc-disc-4167-c18555918-mtp-nodraft",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-nodraft; MTP off baseline; 32k window q8_0 KV",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "M2 Max (30 GPU, 32GB); Qwen3.8-27B, --spec-type draft-mtp, 32k window, q8_0 KV cache, ollama; no draft head tg128 46.2",
   "ram_gb": 32,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-18555918",
   "tg_tokens": 128,
   "tps": "46.2",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-22",
   "flags": [],
   "hardware": "M2 Max (30 GPU)",
   "id": "lc-disc-4167-c18555918-mtp-n2",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-n2; MTP depth 2; already below the no-draft baseline",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "M2 Max (30 GPU, 32GB); Qwen3.8-27B, --spec-type draft-mtp, 32k window, q8_0 KV cache, ollama; draft-mtp (n=2) tg128 34.6",
   "ram_gb": 32,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-18555918",
   "tg_tokens": 128,
   "tps": "34.6",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-22",
   "flags": [],
   "hardware": "M2 Max (30 GPU)",
   "id": "lc-disc-4167-c18555918-mtp-n4",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-n4; MTP depth 4; monotonically worse with depth",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "M2 Max (30 GPU, 32GB); Qwen3.8-27B, --spec-type draft-mtp, 32k window, q8_0 KV cache, ollama; deeper drafting n=4: 29.4",
   "ram_gb": 32,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-18555918",
   "tg_tokens": 128,
   "tps": "29.4",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-22",
   "flags": [],
   "hardware": "M2 Max (30 GPU)",
   "id": "lc-disc-4167-c18555918-mtp-n6",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-n6; MTP depth 6; monotonically worse with depth",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "M2 Max (30 GPU, 32GB); Qwen3.8-27B, --spec-type draft-mtp, 32k window, q8_0 KV cache, ollama; deeper drafting n=6: 20.1",
   "ram_gb": 32,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-18555918",
   "tg_tokens": 128,
   "tps": "20.1",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Metal)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-22",
   "flags": [],
   "hardware": "M2 Max (30 GPU)",
   "id": "lc-disc-4167-c18555918-mtp-n8",
   "model": "Qwen3.8-27B",
   "notes": "config=mtp-n8; MTP depth 8; monotonically worse with depth",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unspecified",
   "quote": "M2 Max (30 GPU, 32GB); Qwen3.8-27B, --spec-type draft-mtp, 32k window, q8_0 KV cache, ollama; deeper drafting n=8: 17.1",
   "ram_gb": 32,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #4167 (performance thread, Apple Silicon)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-18555918",
   "tg_tokens": 128,
   "tps": "17.1",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-10",
   "flags": [],
   "hardware": "V620",
   "id": "lc-disc-15021-c18378967-v620-fa0",
   "model": "llama 7B",
   "notes": "config=fa0; single V620 (32GB gfx1030), ROCm v10, llama.cpp 434ddbbc0 (10884); GPU1 replicate: pp512 1819.51 tg128 91.48",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "1846.20",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Radeon Pro V620, gfx1030, ROCm v10, single GPU (HIP_VISIBLE_DEVICES=0); build 434ddbbc0 (10884); llama 7B Q4_0 3.56 GiB 6.74 B ROCm fa0 pp512 1846.20 \u00b1 93.09 tg128 91.70 \u00b1 0.59",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-18378967",
   "tg_tokens": 128,
   "tps": "91.70",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-10",
   "flags": [],
   "hardware": "V620",
   "id": "lc-disc-15021-c18378967-v620-fa1",
   "model": "llama 7B",
   "notes": "config=fa1; single V620; GPU1 replicate: pp512 2141.85 tg128 98.44",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "2173.13",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "AMD Radeon Pro V620, gfx1030, ROCm v10, single GPU; build 434ddbbc0 (10884); llama 7B Q4_0 ROCm fa1 pp512 2173.13 \u00b1 103.19 tg128 98.78 \u00b1 0.53",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-18378967",
   "tg_tokens": 128,
   "tps": "98.78",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-10",
   "flags": [],
   "hardware": "2x V620",
   "id": "lc-disc-15021-c18378967-2xv620-fa0",
   "model": "llama 7B",
   "notes": "config=fa0; 2x V620 dual-card; prefill holds but decode drops well below single card",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "1801.75",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "2x V620 both cards at the same time (found 2 ROCm devices, gfx1030), ROCm v10; build 434ddbbc0 (10884); llama 7B Q4_0 ROCm fa0 pp512 1801.75 \u00b1 30.74 tg128 64.85 \u00b1 0.11",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-18378967",
   "tg_tokens": 128,
   "tps": "64.85",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-10",
   "flags": [],
   "hardware": "2x V620",
   "id": "lc-disc-15021-c18378967-2xv620-fa1",
   "model": "llama 7B",
   "notes": "config=fa1; 2x V620 dual-card; fa on recovers some prefill but decode stays ~29% below single card",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "2146.36",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "2x V620 both cards, ROCm v10; build 434ddbbc0 (10884); llama 7B Q4_0 ROCm fa1 pp512 2146.36 \u00b1 29.46 tg128 71.38 \u00b1 0.11",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-18378967",
   "tg_tokens": 128,
   "tps": "71.38",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-17",
   "flags": [],
   "hardware": "Arc A770",
   "id": "lc-disc-23313-c18474926-fa0",
   "model": "llama 7B",
   "notes": "config=fa0; Arc A770 SYCL, -ctk f16 -ctv f16, build 2cdae802e (10714)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "826.15",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Intel Arc A770, i7-13700K, Ubuntu 24.04, 64GB DDR5; build 2cdae802e (10714); llama 7B Q4_0 3.56 GiB 6.74 B SYCL fa0 pp512 826.15 \u00b1 1.07 tg128 47.81 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18474926",
   "tg_tokens": 128,
   "tps": "47.81",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (SYCL)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-17",
   "flags": [],
   "hardware": "Arc A770",
   "id": "lc-disc-23313-c18474926-fa1",
   "model": "llama 7B",
   "notes": "config=fa1; Arc A770 SYCL; fa on: prefill +18.6%, decode +13.7%",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": "979.76",
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Intel Arc A770, i7-13700K, Ubuntu 24.04; build 2cdae802e (10714); llama 7B Q4_0 SYCL fa1 pp512 979.76 \u00b1 2.92 tg128 54.38 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #23313 (performance thread, Intel SYCL)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/23313#discussioncomment-18474926",
   "tg_tokens": 128,
   "tps": "54.38",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "Radeon RX 9060 XT",
   "id": "lc-disc-10879-c18709094-rx9060xt-llama7b-fa0",
   "model": "llama 7B",
   "notes": "config=fa0",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2844.8,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9060 XT; llama 7B Q4_0 fa 0 | pp512 | 2844.80 \u00b1 132.10; llama 7B Q4_0 fa 0 | tg128 | 71.07 \u00b1 0.10",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18709094",
   "tg_tokens": 128,
   "tps": 71.07,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "Radeon RX 9060 XT",
   "id": "lc-disc-10879-c18709094-rx9060xt-llama7b-fa1",
   "model": "llama 7B",
   "notes": "config=fa1",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3228.15,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9060 XT; llama 7B Q4_0 fa 1 | pp512 | 3228.15 \u00b1 16.15; llama 7B Q4_0 fa 1 | tg128 | 74.01 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18709094",
   "tg_tokens": 128,
   "tps": 74.01,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "Radeon RX 9060 XT",
   "id": "lc-disc-10879-c18709094-rx9060xt-gptoss20b",
   "model": "gpt-oss 20B",
   "notes": "larger-model table, -ngl 99 -fa 1, pp512/tg128 (Vulkan column; HIP in same table)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3464,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Radeon RX 9060 XT; gpt-oss-20b MXFP4 fa 1 | pp512 3464; tg128 105.8",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18709094",
   "tg_tokens": 128,
   "tps": 105.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "Radeon RX 9060 XT",
   "id": "lc-disc-10879-c18709094-rx9060xt-qwen38-27b",
   "model": "Qwen3.8-27B",
   "notes": "larger-model table, -ngl 99 -fa 1, pp512/tg128 (Vulkan column; HIP in same table)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 709,
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "Radeon RX 9060 XT; Qwen3.8-27B UD-IQ4_XS fa 1 | pp512 709; tg128 20.1",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18709094",
   "tg_tokens": 128,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "Radeon RX 9060 XT",
   "id": "lc-disc-10879-c18709094-rx9060xt-gemma4-12b",
   "model": "Gemma 4 12B",
   "notes": "larger-model table, -ngl 99 -fa 1, pp512/tg128 (Vulkan column; HIP in same table)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1509,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Radeon RX 9060 XT; Gemma 4 12B Q4_K_M fa 1 | pp512 1509; tg128 39.0",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18709094",
   "tg_tokens": 128,
   "tps": 39.0,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "Radeon RX 9060 XT",
   "id": "lc-disc-15021-c18709095-rx9060xt-llama7b-fa0",
   "model": "llama 7B",
   "notes": "config=fa0",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 2641.07,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9060 XT; llama 7B Q4_0 fa 0 | pp512 | 2641.07 \u00b1 48.67; llama 7B Q4_0 fa 0 | tg128 | 67.27 \u00b1 0.15",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-18709095",
   "tg_tokens": 128,
   "tps": 67.27,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "Radeon RX 9060 XT",
   "id": "lc-disc-15021-c18709095-rx9060xt-llama7b-fa1",
   "model": "llama 7B",
   "notes": "config=fa1",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3014.51,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon RX 9060 XT; llama 7B Q4_0 fa 1 | pp512 | 3014.51 \u00b1 23.33; llama 7B Q4_0 fa 1 | tg128 | 70.18 \u00b1 0.06",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-18709095",
   "tg_tokens": 128,
   "tps": 70.18,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "Radeon RX 9060 XT",
   "id": "lc-disc-15021-c18709095-rx9060xt-gptoss20b",
   "model": "gpt-oss 20B",
   "notes": "larger-model table, -ngl 99 -fa 1, pp512/tg128 (HIP column; Vulkan in same table)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 3372,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Radeon RX 9060 XT; gpt-oss-20b MXFP4 fa 1 | pp512 3372; tg128 92.8",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-18709095",
   "tg_tokens": 128,
   "tps": 92.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "Radeon RX 9060 XT",
   "id": "lc-disc-15021-c18709095-rx9060xt-qwen38-27b",
   "model": "Qwen3.8-27B",
   "notes": "larger-model table, -ngl 99 -fa 1, pp512/tg128 (HIP column; Vulkan in same table)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 738,
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "Radeon RX 9060 XT; Qwen3.8-27B UD-IQ4_XS fa 1 | pp512 738; tg128 19.5",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-18709095",
   "tg_tokens": 128,
   "tps": 19.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "Radeon RX 9060 XT",
   "id": "lc-disc-15021-c18709095-rx9060xt-gemma4-12b",
   "model": "Gemma 4 12B",
   "notes": "larger-model table, -ngl 99 -fa 1, pp512/tg128 (HIP column; Vulkan in same table)",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1467,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Radeon RX 9060 XT; Gemma 4 12B Q4_K_M fa 1 | pp512 1467; tg128 37.6",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #15021 (performance thread, AMD ROCm)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/15021#discussioncomment-18709095",
   "tg_tokens": 128,
   "tps": 37.6,
   "ttft_s": null
  },
  {
   "backend": "CPU",
   "batch": null,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "Ryzen AI Max+ 395",
   "id": "lc-disc-10879-c18711544-llama7b-cpu-fa0",
   "model": "llama 7B",
   "notes": "config=fa0; Strix Halo (Ryzen AI Max+ 395) CPU, 16 threads, fa off; llama 7B Q4_0, llama-bench",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 300.27,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Ryzen AI Max+ 395; llama 7B Q4_0 | pp512 | 300.27 \u00b1 1.06; llama 7B Q4_0 | tg128 | 27.68 \u00b1 0.03",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18711544",
   "tg_tokens": 128,
   "tps": 27.68,
   "ttft_s": null
  },
  {
   "backend": "CPU",
   "batch": null,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "Ryzen AI Max+ 395",
   "id": "lc-disc-10879-c18711544-llama7b-cpu-fa1",
   "model": "llama 7B",
   "notes": "config=fa1; Strix Halo (Ryzen AI Max+ 395) CPU, 16 threads, fa on; llama 7B Q4_0, llama-bench",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 340.11,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Ryzen AI Max+ 395; llama 7B Q4_0 | pp512 | 340.11 \u00b1 3.08; llama 7B Q4_0 | tg128 | 29.52 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18711544",
   "tg_tokens": 128,
   "tps": 29.52,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-10879-c18711544-llama7b-rocm-fa0",
   "model": "llama 7B",
   "notes": "config=fa0; Strix Halo Radeon 8060S iGPU, ROCm, ngl 100, fa off; llama 7B Q4_0, llama-bench",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1313.76,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; llama 7B Q4_0 | pp512 | 1313.76 \u00b1 7.76; llama 7B Q4_0 | tg128 | 47.36 \u00b1 0.07",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18711544",
   "tg_tokens": 128,
   "tps": 47.36,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (ROCm)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-10879-c18711544-llama7b-rocm-fa1",
   "model": "llama 7B",
   "notes": "config=fa1; Strix Halo Radeon 8060S iGPU, ROCm, ngl 100, fa on; llama 7B Q4_0, llama-bench",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1456.15,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; llama 7B Q4_0 | pp512 | 1456.15 \u00b1 7.22; llama 7B Q4_0 | tg128 | 51.33 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18711544",
   "tg_tokens": 128,
   "tps": 51.33,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-10879-c18711544-llama7b-vk-fa0",
   "model": "llama 7B",
   "notes": "config=fa0; Strix Halo Radeon 8060S iGPU, Vulkan (RADV strix_halo), ngl 100, fa off; llama 7B Q4_0, llama-bench",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1487.08,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; llama 7B Q4_0 | pp512 | 1487.08 \u00b1 16.51; llama 7B Q4_0 | tg128 | 51.81 \u00b1 0.13",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18711544",
   "tg_tokens": 128,
   "tps": 51.81,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-02",
   "flags": [],
   "hardware": "Radeon 8060S",
   "id": "lc-disc-10879-c18711544-llama7b-vk-fa1",
   "model": "llama 7B",
   "notes": "config=fa1; Strix Halo Radeon 8060S iGPU, Vulkan (RADV strix_halo), ngl 100, fa on; llama 7B Q4_0, llama-bench",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1704.3,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 8060S; llama 7B Q4_0 | pp512 | 1704.3 \u00b1 16.9; llama 7B Q4_0 | tg128 | 54.38 \u00b1 0.3",
   "ram_gb": null,
   "retrieved": "2026-10-02",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18711544",
   "tg_tokens": 128,
   "tps": 54.38,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-04",
   "flags": [],
   "hardware": "RTX 5070",
   "id": "lc-disc-29930-chttps://github.com/ggml-org/llama.cpp/discussions/29930-qwen38flashnext-UD-IQ3_XXS",
   "model": "Qwen3.8-Flash-Next",
   "notes": "up from ~7 tok/s on inherited setup; 14-15 tok/s in conversation, 10.15 tok/s on a 4892-token coding prompt, 10.41 tok/s over 10k tokens (16 min); methodology and benchmark scripts in the hashyy fork repo",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ3_XXS",
   "quote": "RTX 5070 12GB + 32GB DDR4-2400, Ryzen 5 5600GT, Windows; Qwen3.8-Flash-Next 177B (UD-IQ3_XXS), expert-streaming fork (hashyy); benchmark about 11.5 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-04",
   "source_name": "llama.cpp discussions #29930 (show-and-tell, RTX 5070 expert streaming)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/29930",
   "tg_tokens": null,
   "tps": 11.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-25",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc-29429-2x3090-glm53-3.0bit-stock",
   "model": "GLM-5.3-Flash",
   "notes": "config=stock; neurall/llama.cpp fork README table (discussion #29429); short chat, single stream, temp 0, model in RAM; paired fork row 22.4 tok/s (1.9x)",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "3.0-bit",
   "quote": "Machine A (2x RTX 3090, Ryzen 7 3700X, 125 GB DDR4-3200), GLM-5.3-Flash 3.0-bit (106 GB), short chat decode, upstream stock llama.cpp 11.9 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-04",
   "source_name": "neurall/llama.cpp fork README (MoE expert cache benchmark, discussion #29429)",
   "source_url": "https://raw.githubusercontent.com/neurall/llama.cpp/142568af7f/README.md",
   "tg_tokens": null,
   "tps": 11.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-25",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc-29429-2x3090-glm53-3.0bit-fork",
   "model": "GLM-5.3-Flash",
   "notes": "config=fork; neurall/llama.cpp MoE expert cache (builds on csantiago78 PR #27861); short chat, single stream, temp 0, model in RAM; stock baseline 11.9 tok/s",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "3.0-bit",
   "quote": "Machine A (2x RTX 3090, Ryzen 7 3700X, 125 GB DDR4-3200), GLM-5.3-Flash 3.0-bit (106 GB), short chat decode, this fork 22.4 tok/s (1.9x over stock)",
   "ram_gb": null,
   "retrieved": "2026-10-04",
   "source_name": "neurall/llama.cpp fork README (MoE expert cache benchmark, discussion #29429)",
   "source_url": "https://raw.githubusercontent.com/neurall/llama.cpp/142568af7f/README.md",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-25",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc-29429-2x3090-mimo26-iq3xxs-stock",
   "model": "MiMo-V2.6-Flash",
   "notes": "config=stock; neurall/llama.cpp fork README table (discussion #29429); short chat, single stream, temp 0; paired fork row 10.9 tok/s (2.4x)",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ3_XXS",
   "quote": "Machine A (2x RTX 3090, Ryzen 7 3700X, 125 GB DDR4-3200), MiMo-V2.6-Flash IQ3_XXS (132 GB, bigger than RAM), short chat decode, upstream stock 4.6 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-04",
   "source_name": "neurall/llama.cpp fork README (MoE expert cache benchmark, discussion #29429)",
   "source_url": "https://raw.githubusercontent.com/neurall/llama.cpp/142568af7f/README.md",
   "tg_tokens": null,
   "tps": 4.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-25",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc-29429-2x3090-mimo26-iq3xxs-fork",
   "model": "MiMo-V2.6-Flash",
   "notes": "config=fork; neurall/llama.cpp MoE expert cache (builds on csantiago78 PR #27861); short chat, single stream, temp 0; stock baseline 4.6 tok/s",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ3_XXS",
   "quote": "Machine A (2x RTX 3090, Ryzen 7 3700X, 125 GB DDR4-3200), MiMo-V2.6-Flash IQ3_XXS (132 GB, bigger than RAM), short chat decode, this fork 10.9 tok/s (2.4x over stock)",
   "ram_gb": null,
   "retrieved": "2026-10-04",
   "source_name": "neurall/llama.cpp fork README (MoE expert cache benchmark, discussion #29429)",
   "source_url": "https://raw.githubusercontent.com/neurall/llama.cpp/142568af7f/README.md",
   "tg_tokens": null,
   "tps": 10.9,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-25",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc-29429-2x3090-qwen38next-udiq4xs-stock",
   "model": "Qwen3.8-Flash-Next",
   "notes": "config=stock; neurall/llama.cpp fork README table (discussion #29429); short chat, single stream, temp 0; paired fork row 46.5 tok/s (1.7x)",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "Machine A (2x RTX 3090, Ryzen 7 3700X, 125 GB DDR4-3200), Qwen3.8-Flash-Next UD-IQ4_XS (88 GB), short chat decode, upstream stock 27.7 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-04",
   "source_name": "neurall/llama.cpp fork README (MoE expert cache benchmark, discussion #29429)",
   "source_url": "https://raw.githubusercontent.com/neurall/llama.cpp/142568af7f/README.md",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-25",
   "flags": [],
   "hardware": "2x RTX 3090",
   "id": "lc-29429-2x3090-qwen38next-udiq4xs-fork",
   "model": "Qwen3.8-Flash-Next",
   "notes": "config=fork; neurall/llama.cpp MoE expert cache (builds on csantiago78 PR #27861); short chat, single stream, temp 0; stock baseline 27.7 tok/s",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "UD-IQ4_XS",
   "quote": "Machine A (2x RTX 3090, Ryzen 7 3700X, 125 GB DDR4-3200), Qwen3.8-Flash-Next UD-IQ4_XS (88 GB), short chat decode, this fork 46.5 tok/s (1.7x over stock)",
   "ram_gb": null,
   "retrieved": "2026-10-04",
   "source_name": "neurall/llama.cpp fork README (MoE expert cache benchmark, discussion #29429)",
   "source_url": "https://raw.githubusercontent.com/neurall/llama.cpp/142568af7f/README.md",
   "tg_tokens": null,
   "tps": 46.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-25",
   "flags": [],
   "hardware": "4x RTX 3090",
   "id": "lc-29429-4x3090-glm53-4gpu-stock",
   "model": "GLM-5.3-Flash",
   "notes": "config=4gpu-stock; neurall/llama.cpp fork README table (discussion #29429); rented 4x 3090 box, 85% in VRAM on 4 GPUs, first run only; short chat, single stream, temp 0",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "3.0-bit",
   "quote": "Machine D (4x RTX 3090, EPYC 7B12 64 cores, 256 GB DDR4, rented), GLM-5.3-Flash 3.0-bit (117.5 GB), short chat decode, upstream stock 24.7 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-04",
   "source_name": "neurall/llama.cpp fork README (MoE expert cache benchmark, discussion #29429)",
   "source_url": "https://raw.githubusercontent.com/neurall/llama.cpp/142568af7f/README.md",
   "tg_tokens": null,
   "tps": 24.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-25",
   "flags": [],
   "hardware": "4x RTX 3090",
   "id": "lc-29429-4x3090-glm53-4gpu-fork",
   "model": "GLM-5.3-Flash",
   "notes": "config=4gpu-fork; neurall/llama.cpp MoE expert cache; warm (second) run with saved state; stock baseline 24.7 tok/s; -t 16 variant reaches 33.8 tok/s (1.4x)",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "3.0-bit",
   "quote": "Machine D (4x RTX 3090, EPYC 7B12 64 cores, 256 GB DDR4, rented), GLM-5.3-Flash 3.0-bit (117.5 GB), short chat decode warm run, this fork 31.5 tok/s (1.3x over stock)",
   "ram_gb": null,
   "retrieved": "2026-10-04",
   "source_name": "neurall/llama.cpp fork README (MoE expert cache benchmark, discussion #29429)",
   "source_url": "https://raw.githubusercontent.com/neurall/llama.cpp/142568af7f/README.md",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-25",
   "flags": [],
   "hardware": "4x RTX 3090",
   "id": "lc-29429-4x3090-glm53-3of4-stock",
   "model": "GLM-5.3-Flash",
   "notes": "config=3of4-stock; neurall/llama.cpp fork README table (discussion #29429); 3 of the 4 GPUs (66% in VRAM); short chat, single stream, temp 0; paired fork row 32.6 tok/s (1.6x)",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "3.0-bit",
   "quote": "Machine D (4x RTX 3090, EPYC 7B12 64 cores, 256 GB DDR4, rented), GLM-5.3-Flash 3.0-bit, 3 of the 4 GPUs, short chat decode, upstream stock 20.1 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-04",
   "source_name": "neurall/llama.cpp fork README (MoE expert cache benchmark, discussion #29429)",
   "source_url": "https://raw.githubusercontent.com/neurall/llama.cpp/142568af7f/README.md",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (CUDA)",
   "batch": null,
   "ctx": null,
   "date": "2026-09-25",
   "flags": [],
   "hardware": "4x RTX 3090",
   "id": "lc-29429-4x3090-glm53-3of4-fork",
   "model": "GLM-5.3-Flash",
   "notes": "config=3of4-fork; neurall/llama.cpp MoE expert cache; 3 of the 4 GPUs (best ratio per the README); stock baseline 20.1 tok/s",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "3.0-bit",
   "quote": "Machine D (4x RTX 3090, EPYC 7B12 64 cores, 256 GB DDR4, rented), GLM-5.3-Flash 3.0-bit, 3 of the 4 GPUs, short chat decode, this fork 32.6 tok/s (1.6x over stock)",
   "ram_gb": null,
   "retrieved": "2026-10-04",
   "source_name": "neurall/llama.cpp fork README (MoE expert cache benchmark, discussion #29429)",
   "source_url": "https://raw.githubusercontent.com/neurall/llama.cpp/142568af7f/README.md",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "RTX 4060",
   "id": "lc-disc-10879-c18847326-llama7b-Q4_0",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100 sm=none; config=b11540",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1793.16,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 4060; llama 7B Q4_0 | pp512 | 1793.16 \u00b1 29.09; llama 7B Q4_0 | tg128 | 52.58 \u00b1 0.18",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18847326",
   "tg_tokens": 128,
   "tps": 52.58,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "RTX 4060",
   "id": "lc-disc-10879-c18847326-llama7b-Q4_0-2",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100 sm=none; config=b11540",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 1908.24,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "RTX 4060; llama 7B Q4_0 | pp512 | 1908.24 \u00b1 27.95; llama 7B Q4_0 | tg128 | 53.80 \u00b1 0.31",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18847326",
   "tg_tokens": 128,
   "tps": 53.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "Radeon 680M",
   "id": "lc-disc-10879-c18847326-llama7b-Q4_0-3",
   "model": "llama 7B",
   "notes": "threads=100; fa=0 ngl=100 sm=none; radv_perftest=nogttspill; config=b11540",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 260.18,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 680M; llama 7B Q4_0 | pp512 | 260.18 \u00b1 0.60; llama 7B Q4_0 | tg128 | 14.58 \u00b1 0.04",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18847326",
   "tg_tokens": 128,
   "tps": 14.58,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp (Vulkan)",
   "batch": null,
   "ctx": null,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "Radeon 680M",
   "id": "lc-disc-10879-c18847326-llama7b-Q4_0-4",
   "model": "llama 7B",
   "notes": "threads=100; run 2 of 2 in this comment (separate llama-bench run; build/flags differ); fa=1 ngl=100 sm=none; radv_perftest=nogttspill; config=b11540",
   "params": null,
   "power_w": null,
   "pp_tokens": 512,
   "pp_tps": 325.54,
   "provenance": "community",
   "quant": "Q4_0",
   "quote": "Radeon 680M; llama 7B Q4_0 | pp512 | 325.54 \u00b1 73.80; llama 7B Q4_0 | tg128 | 15.42 \u00b1 0.02",
   "ram_gb": null,
   "retrieved": "2026-10-14",
   "source_name": "llama.cpp discussions #10879 (performance thread, Vulkan)",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/10879#discussioncomment-18847326",
   "tg_tokens": 128,
   "tps": 15.42,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "llmcheck-deepseek-v4-flash-m5-max-q2-k-mlx",
   "model": "DeepSeek-V4-Flash",
   "notes": null,
   "params": "284B-A13B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "DeepSeek V4 Flash,284B-A13B,Q2_K,M5 Max,128,MLX,39,0.4,2026-08",
   "ram_gb": 128,
   "retrieved": "2026-10-03",
   "source_name": "LLMCheck Apple Silicon LLM Benchmark Database (CC BY 4.0)",
   "source_url": "https://modelfit.io/blog/deepseek-v4-flash-0731-128gb-macbook/",
   "tps": 39,
   "ttft_s": 0.4
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "M3 Max",
   "id": "llmcheck-deepseek-v4-flash-m3-max-q2-k-mlx",
   "model": "DeepSeek-V4-Flash",
   "notes": null,
   "params": "284B-A13B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q2_K",
   "quote": "DeepSeek V4 Flash,284B-A13B,Q2_K,M3 Max,128,MLX,27,0.6,2026-08",
   "ram_gb": 128,
   "retrieved": "2026-10-03",
   "source_name": "LLMCheck Apple Silicon LLM Benchmark Database (CC BY 4.0)",
   "source_url": "https://modelfit.io/blog/deepseek-v4-flash-0731-128gb-macbook/",
   "tps": 27,
   "ttft_s": 0.6
  },
  {
   "backend": "LM Studio",
   "batch": null,
   "ctx": null,
   "date": "2026-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "llmcheck-glm-5-2-m3-ultra-iq1-s-lm-studio",
   "model": "GLM 5.2",
   "notes": null,
   "params": "753B-A40B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "IQ1_S",
   "quote": "GLM 5.2,753B-A40B,IQ1_S,M3 Ultra,256,LM Studio,22,1.5,2026-07",
   "ram_gb": 256,
   "retrieved": "2026-10-03",
   "source_name": "LLMCheck Apple Silicon LLM Benchmark Database (CC BY 4.0)",
   "source_url": "https://codersera.com/blog/how-to-run-glm-5-2-locally-2026/",
   "tps": 22,
   "ttft_s": 1.5
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": "2026-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "llmcheck-glm-5-2-m3-ultra-q4-k-m-mlx",
   "model": "GLM 5.2",
   "notes": null,
   "params": "753B-A40B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "GLM 5.2,753B-A40B,Q4_K_M,M3 Ultra,512,MLX,15,2.0,2026-07",
   "ram_gb": 512,
   "retrieved": "2026-10-03",
   "source_name": "LLMCheck Apple Silicon LLM Benchmark Database (CC BY 4.0)",
   "source_url": "https://kingy.ai/blog/glm-5-2-256gb-mac-studio-local/",
   "tps": 15,
   "ttft_s": 2.0
  },
  {
   "backend": "LM Studio",
   "batch": null,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "llmcheck-muse-glimmer-30b-m4-pro-q4-k-m-lm-studio",
   "model": "muse-glimmer 30B",
   "notes": null,
   "params": "30B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Muse Glimmer 30B,30B,Q4_K_M,M4 Pro,24,LM Studio,10,0.8,2026-08",
   "ram_gb": 24,
   "retrieved": "2026-10-03",
   "source_name": "LLMCheck Apple Silicon LLM Benchmark Database (CC BY 4.0)",
   "source_url": "https://kingy.ai/blog/muse-glimmer-private-offline-ai-agent/",
   "tps": 10,
   "ttft_s": 0.8
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "llmcheck-muse-glimmer-30b-m5-max-q4-k-m-mlx",
   "model": "muse-glimmer 30B",
   "notes": null,
   "params": "30B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "Muse Glimmer 30B,30B,Q4_K_M,M5 Max,64,MLX,27,0.4,2026-08",
   "ram_gb": 64,
   "retrieved": "2026-10-03",
   "source_name": "LLMCheck Apple Silicon LLM Benchmark Database (CC BY 4.0)",
   "source_url": "https://huggingface.co/blog/muse-glimmer",
   "tps": 27,
   "ttft_s": 0.4
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "llmcheck-muse-glimmer-30b-m4-max-q4-k-m-mlx",
   "model": "muse-glimmer 30B",
   "notes": null,
   "params": "30B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "Muse Glimmer 30B,30B,Q4_K_M,M4 Max,48,MLX,24,0.5,2026-08",
   "ram_gb": 48,
   "retrieved": "2026-10-03",
   "source_name": "LLMCheck Apple Silicon LLM Benchmark Database (CC BY 4.0)",
   "source_url": "https://huggingface.co/blog/muse-glimmer",
   "tps": 24,
   "ttft_s": 0.5
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "llmcheck-lfm2-5-2-6b-m5-max-q4-k-m-mlx",
   "model": "LFM2.5-2.6B",
   "notes": null,
   "params": "2.6B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "LFM2.5-2.6B,2.6B,Q4_K_M,M5 Max,64,MLX,220,0.1,2026-08",
   "ram_gb": 64,
   "retrieved": "2026-10-03",
   "source_name": "LLMCheck Apple Silicon LLM Benchmark Database (CC BY 4.0)",
   "source_url": "https://www.liquid.ai/blog/lfm2-5-2-6b",
   "tps": 220,
   "ttft_s": 0.1
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": "2026-09-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "llmcheck-maple-preview-20b-a1b-m5-pro-ternary-mlx",
   "model": "Maple Preview 20B-A1B",
   "notes": null,
   "params": "20B-A1B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Ternary",
   "quote": "Maple Preview 20B-A1B,20B-A1B,Ternary,M5 Pro,48,MLX,387,None,2026-09-17",
   "ram_gb": 48,
   "retrieved": "2026-10-03",
   "source_name": "LLMCheck Apple Silicon LLM Benchmark Database (CC BY 4.0)",
   "source_url": "https://llmcheck.net/data/benchmarks.json",
   "tps": 387,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "M4",
   "id": "llmcheck-maple-preview-20b-a1b-m4-ternary-mlx",
   "model": "Maple Preview 20B-A1B",
   "notes": null,
   "params": "20B-A1B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Ternary",
   "quote": "Maple Preview 20B-A1B,20B-A1B,Ternary,M4,16,MLX,200,0.1,2026-08",
   "ram_gb": 16,
   "retrieved": "2026-10-03",
   "source_name": "LLMCheck Apple Silicon LLM Benchmark Database (CC BY 4.0)",
   "source_url": "https://huggingface.co/deepgrove/maple-preview",
   "tps": 200,
   "ttft_s": 0.1
  },
  {
   "backend": "LM Studio",
   "batch": null,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "llmcheck-qwen3-8-27b-m5-ultra-q4-k-m-lm-studio",
   "model": "Qwen3.8-27B",
   "notes": null,
   "params": "27.8B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "Qwen3.8-27B,27.8B,Q4_K_M,M5 Ultra,256,LM Studio,55,0.5,2026-09",
   "ram_gb": 256,
   "retrieved": "2026-10-03",
   "source_name": "LLMCheck Apple Silicon LLM Benchmark Database (CC BY 4.0)",
   "source_url": "https://www.bgr.com/2264114/mac-studio-m5-ultra-review/",
   "tps": 55,
   "ttft_s": 0.5
  },
  {
   "backend": "LM Studio",
   "batch": null,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "llmcheck-qwen3-8-27b-m3-ultra-q4-k-m-lm-studio",
   "model": "Qwen3.8-27B",
   "notes": null,
   "params": "27.8B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "Qwen3.8-27B,27.8B,Q4_K_M,M3 Ultra,96,LM Studio,40,None,2026-09",
   "ram_gb": 96,
   "retrieved": "2026-10-03",
   "source_name": "LLMCheck Apple Silicon LLM Benchmark Database (CC BY 4.0)",
   "source_url": "https://www.bgr.com/2264114/mac-studio-m5-ultra-review/",
   "tps": 40,
   "ttft_s": null
  },
  {
   "backend": "LM Studio",
   "batch": null,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "llmcheck-qwen-3-5-122b-a10b-m5-ultra-unstated-lm-studio",
   "model": "Qwen 3.5 122B-A10B",
   "notes": null,
   "params": "122B-A10B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "Qwen 3.5 122B-A10B,122B-A10B,unstated,M5 Ultra,256,LM Studio,80,None,2026-09",
   "ram_gb": 256,
   "retrieved": "2026-10-03",
   "source_name": "LLMCheck Apple Silicon LLM Benchmark Database (CC BY 4.0)",
   "source_url": "https://www.bgr.com/2264114/mac-studio-m5-ultra-review/",
   "tps": 80,
   "ttft_s": null
  },
  {
   "backend": "LM Studio",
   "batch": null,
   "ctx": null,
   "date": "2026-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "llmcheck-qwen-3-5-122b-a10b-m3-ultra-unstated-lm-studio",
   "model": "Qwen 3.5 122B-A10B",
   "notes": null,
   "params": "122B-A10B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "unstated",
   "quote": "Qwen 3.5 122B-A10B,122B-A10B,unstated,M3 Ultra,96,LM Studio,60,None,2026-09",
   "ram_gb": 96,
   "retrieved": "2026-10-03",
   "source_name": "LLMCheck Apple Silicon LLM Benchmark Database (CC BY 4.0)",
   "source_url": "https://www.bgr.com/2264114/mac-studio-m5-ultra-review/",
   "tps": 60,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": "2026-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "llmcheck-qwen-3-6-35b-a3b-m5-max-4-bit-mlx",
   "model": "Qwen 3.6-35B-A3B",
   "notes": "reclassified from estimate to measured by llmcheck (2026-08)",
   "params": "35B-A3B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "4-bit",
   "quote": "Qwen 3.6-35B-A3B,35B-A3B,4-bit,M5 Max,128,MLX,126,None,2026-08",
   "ram_gb": 128,
   "retrieved": "2026-10-05",
   "source_name": "LLMCheck Apple Silicon LLM Benchmark Database (CC BY 4.0)",
   "source_url": "https://omlx.ai/benchmarks/performance/nx08c47t",
   "tg_tokens": null,
   "tps": 126,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "M6",
   "id": "llmcheck-qwen-3-6-35b-a3b-m6-4-bit-mlx",
   "model": "Qwen 3.6-35B-A3B",
   "notes": "reclassified from estimate to measured by llmcheck (2026-10)",
   "params": "35B-A3B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "4-bit",
   "quote": "Qwen 3.6-35B-A3B,35B-A3B,4-bit,M6,32,MLX,64,None,2026-10",
   "ram_gb": 32,
   "retrieved": "2026-10-05",
   "source_name": "LLMCheck Apple Silicon LLM Benchmark Database (CC BY 4.0)",
   "source_url": "https://github.com/HDZucht/Mac-Mini-M6-LLM-benchmarks",
   "tg_tokens": null,
   "tps": 64,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "M6",
   "id": "llmcheck-nemotron-3-5-lightning-m6-q4-k-m-llama-cpp",
   "model": "Nemotron-3.5-Lightning",
   "notes": "reclassified from estimate to measured by llmcheck (2026-10)",
   "params": "30B-A3B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Nemotron 3.5 Lightning,30B-A3B,Q4_K_M,M6,32,llama.cpp,45,None,2026-10",
   "ram_gb": 32,
   "retrieved": "2026-10-05",
   "source_name": "LLMCheck Apple Silicon LLM Benchmark Database (CC BY 4.0)",
   "source_url": "https://github.com/HDZucht/Mac-Mini-M6-LLM-benchmarks",
   "tg_tokens": null,
   "tps": 45,
   "ttft_s": null
  },
  {
   "backend": "LM Studio",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "M6",
   "id": "llmcheck-gemma-4-26b-a4b-m6-qat-lm-studio",
   "model": "Gemma 4 26B-A4B",
   "notes": "reclassified from estimate to measured by llmcheck (2026-10)",
   "params": "26B-A4B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "QAT",
   "quote": "Gemma 4 26B-A4B,26B-A4B,QAT,M6,32,LM Studio,48,None,2026-10",
   "ram_gb": 32,
   "retrieved": "2026-10-05",
   "source_name": "LLMCheck Apple Silicon LLM Benchmark Database (CC BY 4.0)",
   "source_url": "https://github.com/HDZucht/Mac-Mini-M6-LLM-benchmarks",
   "tg_tokens": null,
   "tps": 48,
   "ttft_s": null
  },
  {
   "backend": "LM Studio",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "M6",
   "id": "llmcheck-gemma-4-31b-m6-qat-lm-studio",
   "model": "Gemma-4-31B",
   "notes": "reclassified from estimate to measured by llmcheck (2026-10)",
   "params": "31B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "QAT",
   "quote": "Gemma 4 31B,31B,QAT,M6,32,LM Studio,8,None,2026-10",
   "ram_gb": 32,
   "retrieved": "2026-10-05",
   "source_name": "LLMCheck Apple Silicon LLM Benchmark Database (CC BY 4.0)",
   "source_url": "https://github.com/HDZucht/Mac-Mini-M6-LLM-benchmarks",
   "tg_tokens": null,
   "tps": 8,
   "ttft_s": null
  },
  {
   "backend": "LM Studio",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "M6",
   "id": "llmcheck-muse-glimmer-30b-m6-unstated-lm-studio",
   "model": "muse-glimmer 30B",
   "notes": "reclassified from estimate to measured by llmcheck (2026-10)",
   "params": "30B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unstated",
   "quote": "Muse Glimmer 30B,30B,unstated,M6,32,LM Studio,9,None,2026-10",
   "ram_gb": 32,
   "retrieved": "2026-10-05",
   "source_name": "LLMCheck Apple Silicon LLM Benchmark Database (CC BY 4.0)",
   "source_url": "https://github.com/HDZucht/Mac-Mini-M6-LLM-benchmarks",
   "tg_tokens": null,
   "tps": 9,
   "ttft_s": null
  },
  {
   "backend": "LM Studio",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "M6",
   "id": "llmcheck-gpt-oss-20b-m6-unstated-lm-studio",
   "model": "gpt-oss 20B",
   "notes": "reclassified from estimate to community by llmcheck (HDZucht M6 card)",
   "params": "21B-A3.6B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "unstated",
   "quote": "GPT-oss 20B,21B-A3.6B,unstated,M6,32,LM Studio,45,None,2026-10",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "LLMCheck Apple Silicon LLM Benchmark Database (CC BY 4.0)",
   "source_url": "https://github.com/HDZucht/Mac-Mini-M6-LLM-benchmarks",
   "tg_tokens": null,
   "tps": 45,
   "ttft_s": null
  },
  {
   "backend": "LM Studio",
   "batch": null,
   "ctx": null,
   "date": "2026-10",
   "flags": [],
   "hardware": "M6",
   "id": "llmcheck-gemma-4-12b-m6-qat-lm-studio",
   "model": "Gemma 4 12B",
   "notes": "reclassified from estimate to measured by llmcheck (2026-10)",
   "params": "12B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "QAT",
   "quote": "Gemma 4 12B,12B,QAT,M6,32,LM Studio,19,None,2026-10",
   "ram_gb": 32,
   "retrieved": "2026-10-05",
   "source_name": "LLMCheck Apple Silicon LLM Benchmark Database (CC BY 4.0)",
   "source_url": "https://github.com/HDZucht/Mac-Mini-M6-LLM-benchmarks",
   "tg_tokens": null,
   "tps": 19,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": "2026-01",
   "flags": [],
   "hardware": "M4 Max",
   "id": "llmcheck-qwen-3-30b-a3b-m4-max-4-bit-mlx",
   "model": "Qwen 3 30B-A3B",
   "notes": "reclassified from estimate to measured by llmcheck (2026-01)",
   "params": "30B-A3B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "4-bit",
   "quote": "Qwen 3 30B-A3B,30B-A3B,4-bit,M4 Max,128,MLX,110,None,2026-01",
   "ram_gb": 128,
   "retrieved": "2026-10-05",
   "source_name": "LLMCheck Apple Silicon LLM Benchmark Database (CC BY 4.0)",
   "source_url": "https://arxiv.org/abs/2601.19139",
   "tg_tokens": null,
   "tps": 110,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": "2026-01",
   "flags": [],
   "hardware": "M4 Max",
   "id": "llmcheck-qwen-3-30b-a3b-m4-max-4-bit-llama-cpp",
   "model": "Qwen 3 30B-A3B",
   "notes": "reclassified from estimate to measured by llmcheck (2026-01)",
   "params": "30B-A3B",
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "4-bit",
   "quote": "Qwen 3 30B-A3B,30B-A3B,4-bit,M4 Max,128,llama.cpp,90,None,2026-01",
   "ram_gb": 128,
   "retrieved": "2026-10-05",
   "source_name": "LLMCheck Apple Silicon LLM Benchmark Database (CC BY 4.0)",
   "source_url": "https://llmcheck.net/data/benchmarks.json",
   "tg_tokens": null,
   "tps": 90,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp llama-bench",
   "batch": null,
   "ctx": 4096,
   "date": "2026-08-17",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "llmconfigurator-rtx-4090-qwen3-30b-a3b-q4-k-xl-ctx4096",
   "model": "Qwen3 30B A3B",
   "notes": "publisher: Hardware Corner",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K_XL",
   "quote": "rtx-4090,qwen3-30b-a3b,Q4_K_XL,ctx 4096,195.78 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "LLM Configurator measured benchmarks (CC BY 4.0)",
   "source_url": "https://www.hardware-corner.net/rtx-4090-llm-benchmarks/",
   "tps": 195.78,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp llama-bench",
   "batch": null,
   "ctx": 57344,
   "date": "2026-08-17",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "llmconfigurator-rtx-4090-qwen3-30b-a3b-q4-k-xl-ctx57344",
   "model": "Qwen3 30B A3B",
   "notes": "publisher: Hardware Corner",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K_XL",
   "quote": "rtx-4090,qwen3-30b-a3b,Q4_K_XL,ctx 57344,74.64 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "LLM Configurator measured benchmarks (CC BY 4.0)",
   "source_url": "https://www.hardware-corner.net/rtx-4090-llm-benchmarks/",
   "tps": 74.64,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp llama-bench",
   "batch": null,
   "ctx": 4096,
   "date": "2026-08-17",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "llmconfigurator-rtx-4090-qwen3-8b-q4-k-m-ctx4096",
   "model": "Qwen3 8B",
   "notes": "publisher: Hardware Corner",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "rtx-4090,qwen3-8b,Q4_K_M,ctx 4096,141.3 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "LLM Configurator measured benchmarks (CC BY 4.0)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4090/",
   "tps": 141.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp llama-bench",
   "batch": null,
   "ctx": 16384,
   "date": "2026-08-17",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "llmconfigurator-rtx-4090-qwen3-8b-q4-k-m-ctx16384",
   "model": "Qwen3 8B",
   "notes": "publisher: Hardware Corner",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "rtx-4090,qwen3-8b,Q4_K_M,ctx 16384,108 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "LLM Configurator measured benchmarks (CC BY 4.0)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4090/",
   "tps": 108,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp llama-bench",
   "batch": null,
   "ctx": 32768,
   "date": "2026-08-17",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "llmconfigurator-rtx-4090-qwen3-8b-q4-k-m-ctx32768",
   "model": "Qwen3 8B",
   "notes": "publisher: Hardware Corner",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "rtx-4090,qwen3-8b,Q4_K_M,ctx 32768,82.3 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "LLM Configurator measured benchmarks (CC BY 4.0)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4090/",
   "tps": 82.3,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp llama-bench",
   "batch": null,
   "ctx": 65536,
   "date": "2026-08-17",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "llmconfigurator-rtx-4090-qwen3-8b-q4-k-m-ctx65536",
   "model": "Qwen3 8B",
   "notes": "publisher: Hardware Corner",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "rtx-4090,qwen3-8b,Q4_K_M,ctx 65536,56.1 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "LLM Configurator measured benchmarks (CC BY 4.0)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4090/",
   "tps": 56.1,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp llama-bench",
   "batch": null,
   "ctx": 131072,
   "date": "2026-08-17",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "llmconfigurator-rtx-4090-qwen3-8b-q4-k-m-ctx131072",
   "model": "Qwen3 8B",
   "notes": "publisher: Hardware Corner",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "rtx-4090,qwen3-8b,Q4_K_M,ctx 131072,33.8 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "LLM Configurator measured benchmarks (CC BY 4.0)",
   "source_url": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4090/",
   "tps": 33.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp llama-bench",
   "batch": null,
   "ctx": 4096,
   "date": "2026-08-17",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "llmconfigurator-rtx-4090-llama-3-1-8b-q4-k-xl-ctx4096",
   "model": "Llama 3.1 8B",
   "notes": "publisher: Hardware Corner",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K_XL",
   "quote": "rtx-4090,llama-3.1-8b,Q4_K_XL,ctx 4096,131 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "LLM Configurator measured benchmarks (CC BY 4.0)",
   "source_url": "https://www.hardware-corner.net/rtx-4090-llm-benchmarks/",
   "tps": 131,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp llama-bench",
   "batch": null,
   "ctx": 65536,
   "date": "2026-08-17",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "llmconfigurator-rtx-4090-llama-3-1-8b-q4-k-xl-ctx65536",
   "model": "Llama 3.1 8B",
   "notes": "publisher: Hardware Corner",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K_XL",
   "quote": "rtx-4090,llama-3.1-8b,Q4_K_XL,ctx 65536,53.07 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "LLM Configurator measured benchmarks (CC BY 4.0)",
   "source_url": "https://www.hardware-corner.net/rtx-4090-llm-benchmarks/",
   "tps": 53.07,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp llama-bench",
   "batch": null,
   "ctx": 4096,
   "date": "2026-08-17",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "llmconfigurator-rtx-4090-qwen3-14b-q4-k-xl-ctx4096",
   "model": "Qwen3 14B",
   "notes": "publisher: Hardware Corner",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K_XL",
   "quote": "rtx-4090,qwen3-14b,Q4_K_XL,ctx 4096,82.82 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "LLM Configurator measured benchmarks (CC BY 4.0)",
   "source_url": "https://www.hardware-corner.net/rtx-4090-llm-benchmarks/",
   "tps": 82.82,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp llama-bench",
   "batch": null,
   "ctx": 65536,
   "date": "2026-08-17",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "llmconfigurator-rtx-4090-qwen3-14b-q4-k-xl-ctx65536",
   "model": "Qwen3 14B",
   "notes": "publisher: Hardware Corner",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K_XL",
   "quote": "rtx-4090,qwen3-14b,Q4_K_XL,ctx 65536,38.73 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "LLM Configurator measured benchmarks (CC BY 4.0)",
   "source_url": "https://www.hardware-corner.net/rtx-4090-llm-benchmarks/",
   "tps": 38.73,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp llama-bench",
   "batch": null,
   "ctx": 4096,
   "date": "2026-08-17",
   "flags": [],
   "hardware": "RTX 4090",
   "id": "llmconfigurator-rtx-4090-qwen-2-5-7b-q4-k-m-ctx4096",
   "model": "qwen-2.5-7b",
   "notes": "publisher: Mustafa.net",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "rtx-4090,qwen-2.5-7b,Q4_K_M,ctx 4096,135 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "LLM Configurator measured benchmarks (CC BY 4.0)",
   "source_url": "https://mustafa.net/llm-tokens-per-second-benchmarks/",
   "tps": 135,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp llama-bench",
   "batch": null,
   "ctx": 4096,
   "date": "2026-08-17",
   "flags": [],
   "hardware": "RTX 3090",
   "id": "llmconfigurator-rtx-3090-qwen-2-5-7b-q4-k-m-ctx4096",
   "model": "qwen-2.5-7b",
   "notes": "publisher: Mustafa.net",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "rtx-3090,qwen-2.5-7b,Q4_K_M,ctx 4096,95 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "LLM Configurator measured benchmarks (CC BY 4.0)",
   "source_url": "https://mustafa.net/llm-tokens-per-second-benchmarks/",
   "tps": 95,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp llama-bench",
   "batch": null,
   "ctx": 4096,
   "date": "2026-08-17",
   "flags": [],
   "hardware": "M3 Max",
   "id": "llmconfigurator-apple-m3-max-qwen-2-5-7b-q4-k-m-ctx4096",
   "model": "qwen-2.5-7b",
   "notes": "publisher: Mustafa.net",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K_M",
   "quote": "apple-m3-max,qwen-2.5-7b,Q4_K_M,ctx 4096,40 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "LLM Configurator measured benchmarks (CC BY 4.0)",
   "source_url": "https://mustafa.net/llm-tokens-per-second-benchmarks/",
   "tps": 40,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gmlxlf27-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark on M3 Ultra (80c) 512 GB; peak memory 66.3 GB at this context; identical duplicate traces of the same session: vlb9nyb0, vx4ljpey, cavk7p1r, lhzzunu0, 43mgnpxt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 897.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 897.8 52.3 66.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gmlxlf27",
   "source_url": "https://omlx.ai/benchmarks/performance/gmlxlf27",
   "tg_tokens": null,
   "tps": 52.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gmlxlf27-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark on M3 Ultra (80c) 512 GB; peak memory 66.7 GB at this context; identical duplicate traces of the same session: vlb9nyb0, vx4ljpey, cavk7p1r, lhzzunu0, 43mgnpxt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 935.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 935.9 47.9 66.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gmlxlf27",
   "source_url": "https://omlx.ai/benchmarks/performance/gmlxlf27",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gmlxlf27-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark on M3 Ultra (80c) 512 GB; peak memory 67.0 GB at this context; identical duplicate traces of the same session: vlb9nyb0, vx4ljpey, cavk7p1r, lhzzunu0, 43mgnpxt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 920.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 920.4 47.4 67.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gmlxlf27",
   "source_url": "https://omlx.ai/benchmarks/performance/gmlxlf27",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gmlxlf27-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark on M3 Ultra (80c) 512 GB; peak memory 67.7 GB at this context; identical duplicate traces of the same session: vlb9nyb0, vx4ljpey, cavk7p1r, lhzzunu0, 43mgnpxt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 862.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 862.3 43.1 67.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gmlxlf27",
   "source_url": "https://omlx.ai/benchmarks/performance/gmlxlf27",
   "tg_tokens": null,
   "tps": 43.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gmlxlf27-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark on M3 Ultra (80c) 512 GB; peak memory 69.1 GB at this context; identical duplicate traces of the same session: vlb9nyb0, vx4ljpey, cavk7p1r, lhzzunu0, 43mgnpxt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 746.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 746.1 36.4 69.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gmlxlf27",
   "source_url": "https://omlx.ai/benchmarks/performance/gmlxlf27",
   "tg_tokens": null,
   "tps": 36.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gmlxlf27-ctx65536",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark on M3 Ultra (80c) 512 GB; peak memory 71.9 GB at this context; identical duplicate traces of the same session: vlb9nyb0, vx4ljpey, cavk7p1r, lhzzunu0, 43mgnpxt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 570.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 570.2 27.7 71.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gmlxlf27",
   "source_url": "https://omlx.ai/benchmarks/performance/gmlxlf27",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-glkostsz-ctx1024",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark on M3 Ultra (80c) 512 GB; peak memory 302.7 GB at this context; identical duplicate traces of the same session: 0ogsu5db, au03cy2k, q66jpzr6, rb7hx514; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 454.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 1k 454.0 28.5 302.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/glkostsz",
   "source_url": "https://omlx.ai/benchmarks/performance/glkostsz",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-glkostsz-ctx4096",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark on M3 Ultra (80c) 512 GB; peak memory 302.9 GB at this context; identical duplicate traces of the same session: 0ogsu5db, au03cy2k, q66jpzr6, rb7hx514; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 487.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 4k 487.1 27.2 302.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/glkostsz",
   "source_url": "https://omlx.ai/benchmarks/performance/glkostsz",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-glkostsz-ctx8192",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark on M3 Ultra (80c) 512 GB; peak memory 303.3 GB at this context; identical duplicate traces of the same session: 0ogsu5db, au03cy2k, q66jpzr6, rb7hx514; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 484.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 8k 484.9 26.0 303.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/glkostsz",
   "source_url": "https://omlx.ai/benchmarks/performance/glkostsz",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-glkostsz-ctx16384",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark on M3 Ultra (80c) 512 GB; peak memory 304.0 GB at this context; identical duplicate traces of the same session: 0ogsu5db, au03cy2k, q66jpzr6, rb7hx514; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 462.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 16k 462.0 23.6 304.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/glkostsz",
   "source_url": "https://omlx.ai/benchmarks/performance/glkostsz",
   "tg_tokens": null,
   "tps": 23.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-glkostsz-ctx32768",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark on M3 Ultra (80c) 512 GB; peak memory 305.5 GB at this context; identical duplicate traces of the same session: 0ogsu5db, au03cy2k, q66jpzr6, rb7hx514; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 422.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 32k 422.5 20.2 305.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/glkostsz",
   "source_url": "https://omlx.ai/benchmarks/performance/glkostsz",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-pmdh45qm-ctx1024",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark on M3 Ultra (80c) 512 GB; peak memory 227.2 GB at this context; identical duplicate traces of the same session: 6f5a4eko, rf82ex1p, jye2fo4m, s15twedr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 600.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 600.4 34.2 227.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pmdh45qm",
   "source_url": "https://omlx.ai/benchmarks/performance/pmdh45qm",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-pmdh45qm-ctx4096",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark on M3 Ultra (80c) 512 GB; peak memory 228.2 GB at this context; identical duplicate traces of the same session: 6f5a4eko, rf82ex1p, jye2fo4m, s15twedr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 707.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 707.9 30.1 228.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pmdh45qm",
   "source_url": "https://omlx.ai/benchmarks/performance/pmdh45qm",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-pmdh45qm-ctx8192",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark on M3 Ultra (80c) 512 GB; peak memory 229.1 GB at this context; identical duplicate traces of the same session: 6f5a4eko, rf82ex1p, jye2fo4m, s15twedr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 663.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 663.9 25.9 229.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pmdh45qm",
   "source_url": "https://omlx.ai/benchmarks/performance/pmdh45qm",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-pmdh45qm-ctx16384",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark on M3 Ultra (80c) 512 GB; peak memory 231.1 GB at this context; identical duplicate traces of the same session: 6f5a4eko, rf82ex1p, jye2fo4m, s15twedr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 564.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k 564.2 21.1 231.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pmdh45qm",
   "source_url": "https://omlx.ai/benchmarks/performance/pmdh45qm",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-pmdh45qm-ctx32768",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark on M3 Ultra (80c) 512 GB; peak memory 235.0 GB at this context; identical duplicate traces of the same session: 6f5a4eko, rf82ex1p, jye2fo4m, s15twedr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 427.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 32k 427.8 14.6 235.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pmdh45qm",
   "source_url": "https://omlx.ai/benchmarks/performance/pmdh45qm",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ped6lbh2-ctx1024",
   "model": "GLM-5",
   "notes": "oMLX community benchmark on M3 Ultra (80c) 512 GB; peak memory 391.8 GB at this context; identical duplicate traces of the same session: w2ncir8g, dig8n87z, 314vh14k; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 186.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 186.0 16.6 391.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ped6lbh2",
   "source_url": "https://omlx.ai/benchmarks/performance/ped6lbh2",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ped6lbh2-ctx4096",
   "model": "GLM-5",
   "notes": "oMLX community benchmark on M3 Ultra (80c) 512 GB; peak memory 394.1 GB at this context; identical duplicate traces of the same session: w2ncir8g, dig8n87z, 314vh14k; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 180.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 180.1 13.8 394.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ped6lbh2",
   "source_url": "https://omlx.ai/benchmarks/performance/ped6lbh2",
   "tg_tokens": null,
   "tps": 13.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ped6lbh2-ctx8192",
   "model": "GLM-5",
   "notes": "oMLX community benchmark on M3 Ultra (80c) 512 GB; peak memory 396.7 GB at this context; identical duplicate traces of the same session: w2ncir8g, dig8n87z, 314vh14k; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 154.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 154.2 13.1 396.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ped6lbh2",
   "source_url": "https://omlx.ai/benchmarks/performance/ped6lbh2",
   "tg_tokens": null,
   "tps": 13.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ped6lbh2-ctx16384",
   "model": "GLM-5",
   "notes": "oMLX community benchmark on M3 Ultra (80c) 512 GB; peak memory 402.7 GB at this context; identical duplicate traces of the same session: w2ncir8g, dig8n87z, 314vh14k; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 117.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 117.3 12.0 402.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ped6lbh2",
   "source_url": "https://omlx.ai/benchmarks/performance/ped6lbh2",
   "tg_tokens": null,
   "tps": 12.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ped6lbh2-ctx32768",
   "model": "GLM-5",
   "notes": "oMLX community benchmark on M3 Ultra (80c) 512 GB; peak memory 415.4 GB at this context; identical duplicate traces of the same session: w2ncir8g, dig8n87z, 314vh14k; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 77.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 77.3 10.4 415.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ped6lbh2",
   "source_url": "https://omlx.ai/benchmarks/performance/ped6lbh2",
   "tg_tokens": null,
   "tps": 10.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hizr47ls-ctx1024",
   "model": "Huihui-Qwen3.8-27B-abliterated",
   "notes": "MTPLX build with MTP heads enabled; oMLX community benchmark trace, single run at 1k context; peak memory 33.8 GB; TTFT 1274 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 804.0,
   "provenance": "community",
   "quant": "Q8",
   "quote": "Huihui-Qwen3.8-27B-abliterated-MTPLX-Q8 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k tokens 804.0 PP tok/s 49.8 TG tok/s 1274 TTFT (ms) 33.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hizr47ls",
   "source_url": "https://omlx.ai/benchmarks/performance/hizr47ls",
   "tg_tokens": null,
   "tps": 49.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-y73j2owy-ctx131072",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 65.5 GB; TTFT 82528 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1588.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 128k tokens 1,588 PP tok/s 51.3 TG tok/s 82528 TTFT (ms) 65.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y73j2owy",
   "source_url": "https://omlx.ai/benchmarks/performance/y73j2owy",
   "tg_tokens": null,
   "tps": 51.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-gkk8vpx0-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: rxbp96yc; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 225.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 225.0 19.0 5.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gkk8vpx0",
   "source_url": "https://omlx.ai/benchmarks/performance/gkk8vpx0",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-gkk8vpx0-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: rxbp96yc; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 239.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 239.4 19.4 6.2 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gkk8vpx0",
   "source_url": "https://omlx.ai/benchmarks/performance/gkk8vpx0",
   "tg_tokens": null,
   "tps": 19.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-5aiww3ow-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "community MTPLX build with MTP heads; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: h0jneyhw; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 277.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed-oMLX on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 1k 277.9 46.8 22.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5aiww3ow",
   "source_url": "https://omlx.ai/benchmarks/performance/5aiww3ow",
   "tg_tokens": null,
   "tps": 46.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-5aiww3ow-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "community MTPLX build with MTP heads; oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: h0jneyhw; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 286.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed-oMLX on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 4k 286.4 41.7 26.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5aiww3ow",
   "source_url": "https://omlx.ai/benchmarks/performance/5aiww3ow",
   "tg_tokens": null,
   "tps": 41.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-81yp3qxz-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "community MTPLX build with MTP heads; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: vxygr040; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 285.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed-oMLX on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 4k 285.3 42.0 26.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/81yp3qxz",
   "source_url": "https://omlx.ai/benchmarks/performance/81yp3qxz",
   "tg_tokens": null,
   "tps": 42.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-81yp3qxz-ctx16384",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "community MTPLX build with MTP heads; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: vxygr040; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 268.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed-oMLX on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 16k 268.6 44.2 28.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/81yp3qxz",
   "source_url": "https://omlx.ai/benchmarks/performance/81yp3qxz",
   "tg_tokens": null,
   "tps": 44.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-yrplfaht-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "community MTPLX build with MTP heads; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: fgxc9kks; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 286.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed-oMLX on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 4k 286.9 39.9 26.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yrplfaht",
   "source_url": "https://omlx.ai/benchmarks/performance/yrplfaht",
   "tg_tokens": null,
   "tps": 39.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-yrplfaht-ctx16384",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "community MTPLX build with MTP heads; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: fgxc9kks; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 273.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed-oMLX on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 16k 273.6 43.0 28.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yrplfaht",
   "source_url": "https://omlx.ai/benchmarks/performance/yrplfaht",
   "tg_tokens": null,
   "tps": 43.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-r9z3owfy-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: hyn7bzvn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 507.3,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 507.3 14.2 21.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r9z3owfy",
   "source_url": "https://omlx.ai/benchmarks/performance/r9z3owfy",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-r9z3owfy-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: hyn7bzvn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 453.9,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 453.9 13.7 21.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r9z3owfy",
   "source_url": "https://omlx.ai/benchmarks/performance/r9z3owfy",
   "tg_tokens": null,
   "tps": 13.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-93i8c9sx-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: c31tq5pg, ev03nt9i; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 228.6,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 1k 228.6 29.0 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/93i8c9sx",
   "source_url": "https://omlx.ai/benchmarks/performance/93i8c9sx",
   "tg_tokens": null,
   "tps": 29.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-93i8c9sx-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: c31tq5pg, ev03nt9i; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 245.4,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 4k 245.4 28.3 20.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/93i8c9sx",
   "source_url": "https://omlx.ai/benchmarks/performance/93i8c9sx",
   "tg_tokens": null,
   "tps": 28.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-93i8c9sx-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: c31tq5pg, ev03nt9i; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 239.8,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 8k 239.8 27.7 20.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/93i8c9sx",
   "source_url": "https://omlx.ai/benchmarks/performance/93i8c9sx",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-93i8c9sx-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: c31tq5pg, ev03nt9i; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 229.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 16k 229.0 26.0 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/93i8c9sx",
   "source_url": "https://omlx.ai/benchmarks/performance/93i8c9sx",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-izpi8f74-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: ewfym865; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 273.8,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 1k 273.8 42.2 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/izpi8f74",
   "source_url": "https://omlx.ai/benchmarks/performance/izpi8f74",
   "tg_tokens": null,
   "tps": 42.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-izpi8f74-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: ewfym865; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 291.4,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 4k 291.4 43.0 23.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/izpi8f74",
   "source_url": "https://omlx.ai/benchmarks/performance/izpi8f74",
   "tg_tokens": null,
   "tps": 43.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-b9kw2q58-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.7 GB at this context; duplicate traces of the same session: d8hclmdc, xc8piny5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 210.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 210.2 27.9 33.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b9kw2q58",
   "source_url": "https://omlx.ai/benchmarks/performance/b9kw2q58",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-b9kw2q58-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: d8hclmdc, xc8piny5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 214.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 214.3 25.8 37.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b9kw2q58",
   "source_url": "https://omlx.ai/benchmarks/performance/b9kw2q58",
   "tg_tokens": null,
   "tps": 25.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-b9kw2q58-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.1 GB at this context; duplicate traces of the same session: d8hclmdc, xc8piny5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 149.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 149.6 15.4 43.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b9kw2q58",
   "source_url": "https://omlx.ai/benchmarks/performance/b9kw2q58",
   "tg_tokens": null,
   "tps": 15.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ta01qzx6-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 195k context; peak memory 96.5 GB; TTFT 91925 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2176.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 195k tokens 2,176 PP tok/s 50.0 TG tok/s 91925 TTFT (ms) 96.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ta01qzx6",
   "source_url": "https://omlx.ai/benchmarks/performance/ta01qzx6",
   "tg_tokens": null,
   "tps": 50.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-lorkjlol-ctx1024",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 12.1 GB at this context; duplicate traces of the same session: 7vi3j31s; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1877.0,
   "provenance": "community",
   "quant": "UD-MLX-3bit",
   "quote": "gemma-4-26b-a4b-it-UD-MLX-3bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 3bit; 1k 1,877 67.9 12.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lorkjlol",
   "source_url": "https://omlx.ai/benchmarks/performance/lorkjlol",
   "tg_tokens": null,
   "tps": 67.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-lorkjlol-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 12.3 GB at this context; duplicate traces of the same session: 7vi3j31s; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1958.0,
   "provenance": "community",
   "quant": "UD-MLX-3bit",
   "quote": "gemma-4-26b-a4b-it-UD-MLX-3bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 3bit; 4k 1,958 61.9 12.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lorkjlol",
   "source_url": "https://omlx.ai/benchmarks/performance/lorkjlol",
   "tg_tokens": null,
   "tps": 61.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ghd4mymd-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: p9ujrg1n; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 471.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mlx-community--Qwen3.8-27B-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 471.0 9.7 29.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ghd4mymd",
   "source_url": "https://omlx.ai/benchmarks/performance/ghd4mymd",
   "tg_tokens": null,
   "tps": 9.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ghd4mymd-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: p9ujrg1n; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 435.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mlx-community--Qwen3.8-27B-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 16k 435.6 9.3 30.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ghd4mymd",
   "source_url": "https://omlx.ai/benchmarks/performance/ghd4mymd",
   "tg_tokens": null,
   "tps": 9.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-3oc0icev-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 174.5 GB; TTFT 16806 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1950.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k tokens 1,950 PP tok/s 85.6 TG tok/s 16806 TTFT (ms) 174.5 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3oc0icev",
   "source_url": "https://omlx.ai/benchmarks/performance/3oc0icev",
   "tg_tokens": null,
   "tps": 85.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ibm0l1ax-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 23.1 GB; TTFT 58544 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1119.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 64k tokens 1,119 PP tok/s 89.3 TG tok/s 58544 TTFT (ms) 23.1 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ibm0l1ax",
   "source_url": "https://omlx.ai/benchmarks/performance/ibm0l1ax",
   "tg_tokens": null,
   "tps": 89.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-9ey95s4r-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 20.8 GB; TTFT 25219 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1299.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 32k tokens 1,299 PP tok/s 96.9 TG tok/s 25219 TTFT (ms) 20.8 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ey95s4r",
   "source_url": "https://omlx.ai/benchmarks/performance/9ey95s4r",
   "tg_tokens": null,
   "tps": 96.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-o0qtu8rz-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: ijsokdxs, lr6txaa0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 217.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 217.3 33.4 33.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o0qtu8rz",
   "source_url": "https://omlx.ai/benchmarks/performance/o0qtu8rz",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-o0qtu8rz-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: ijsokdxs, lr6txaa0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 220.0 25.7 37.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o0qtu8rz",
   "source_url": "https://omlx.ai/benchmarks/performance/o0qtu8rz",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-o0qtu8rz-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.0 GB at this context; duplicate traces of the same session: ijsokdxs, lr6txaa0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 142.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 142.9 18.2 43.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o0qtu8rz",
   "source_url": "https://omlx.ai/benchmarks/performance/o0qtu8rz",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-mxe30jah-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.1 GB at this context; duplicate traces of the same session: ws73upph; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 650.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 650.7 87.7 99.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mxe30jah",
   "source_url": "https://omlx.ai/benchmarks/performance/mxe30jah",
   "tg_tokens": null,
   "tps": 87.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-mxe30jah-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.9 GB at this context; duplicate traces of the same session: ws73upph; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 730.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 730.0 71.3 100.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mxe30jah",
   "source_url": "https://omlx.ai/benchmarks/performance/mxe30jah",
   "tg_tokens": null,
   "tps": 71.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1pcoe9m2-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 195k context; peak memory 96.7 GB; TTFT 101514 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1970.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 195k tokens 1,970 PP tok/s 45.8 TG tok/s 101514 TTFT (ms) 96.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1pcoe9m2",
   "source_url": "https://omlx.ai/benchmarks/performance/1pcoe9m2",
   "tg_tokens": null,
   "tps": 45.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0pf7d2jx-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 195k context; peak memory 96.6 GB; TTFT 101217 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1976.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 195k tokens 1,976 PP tok/s 45.2 TG tok/s 101217 TTFT (ms) 96.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0pf7d2jx",
   "source_url": "https://omlx.ai/benchmarks/performance/0pf7d2jx",
   "tg_tokens": null,
   "tps": 45.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mu464l99-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 195k context; peak memory 96.1 GB; TTFT 104095 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1921.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 195k tokens 1,921 PP tok/s 45.7 TG tok/s 104095 TTFT (ms) 96.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mu464l99",
   "source_url": "https://omlx.ai/benchmarks/performance/mu464l99",
   "tg_tokens": null,
   "tps": 45.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-w3b5y9x6-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 195k context; peak memory 96.5 GB; TTFT 97049 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2061.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 195k tokens 2,061 PP tok/s 49.4 TG tok/s 97049 TTFT (ms) 96.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w3b5y9x6",
   "source_url": "https://omlx.ai/benchmarks/performance/w3b5y9x6",
   "tg_tokens": null,
   "tps": 49.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-w1xonjvm-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace, single run at 1k context; peak memory 16.0 GB; TTFT 8905 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 115.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 1k tokens 115.0 PP tok/s 20.0 TG tok/s 8905 TTFT (ms) 16.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w1xonjvm",
   "source_url": "https://omlx.ai/benchmarks/performance/w1xonjvm",
   "tg_tokens": null,
   "tps": 20.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-b1gpzfcg-ctx131072",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 103.3 GB; TTFT 181539 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 722.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 128k tokens 722.0 PP tok/s 32.4 TG tok/s 181539 TTFT (ms) 103.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b1gpzfcg",
   "source_url": "https://omlx.ai/benchmarks/performance/b1gpzfcg",
   "tg_tokens": null,
   "tps": 32.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wwbkf1s0-ctx131072",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 103.2 GB; TTFT 181564 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 721.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 128k tokens 721.9 PP tok/s 33.8 TG tok/s 181564 TTFT (ms) 103.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wwbkf1s0",
   "source_url": "https://omlx.ai/benchmarks/performance/wwbkf1s0",
   "tg_tokens": null,
   "tps": 33.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-ial7kabv-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: 5fvml94a; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 193.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 16 GB \u00b7 4bit; 1k 193.0 33.5 5.9 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ial7kabv",
   "source_url": "https://omlx.ai/benchmarks/performance/ial7kabv",
   "tg_tokens": null,
   "tps": 33.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-ial7kabv-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: 5fvml94a; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 188.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 16 GB \u00b7 4bit; 4k 188.7 32.5 6.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ial7kabv",
   "source_url": "https://omlx.ai/benchmarks/performance/ial7kabv",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-rw1hoeuz-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.2 GB at this context; duplicate traces of the same session: 7de9inuf; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 731.7,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 4k 731.7 54.5 32.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rw1hoeuz",
   "source_url": "https://omlx.ai/benchmarks/performance/rw1hoeuz",
   "tg_tokens": null,
   "tps": 54.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-rw1hoeuz-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.4 GB at this context; duplicate traces of the same session: 7de9inuf; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 438.7,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 32k 438.7 47.3 34.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rw1hoeuz",
   "source_url": "https://omlx.ai/benchmarks/performance/rw1hoeuz",
   "tg_tokens": null,
   "tps": 47.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-d75rkjff-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.2 GB at this context; duplicate traces of the same session: uqxp3mcg; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 734.2,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 4k 734.2 33.1 32.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d75rkjff",
   "source_url": "https://omlx.ai/benchmarks/performance/d75rkjff",
   "tg_tokens": null,
   "tps": 33.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-d75rkjff-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.4 GB at this context; duplicate traces of the same session: uqxp3mcg; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 433.3,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 32k 433.3 46.8 34.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d75rkjff",
   "source_url": "https://omlx.ai/benchmarks/performance/d75rkjff",
   "tg_tokens": null,
   "tps": 46.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ddx21ucq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: a7c36ecm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 127.4 26.4 18.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ddx21ucq",
   "source_url": "https://omlx.ai/benchmarks/performance/ddx21ucq",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ddx21ucq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: a7c36ecm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 130.3 28.4 19.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ddx21ucq",
   "source_url": "https://omlx.ai/benchmarks/performance/ddx21ucq",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rbgzwyf7-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: chonyuwu, ol8sfqde, rjvlrqxw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 327.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 327.3 23.1 28.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rbgzwyf7",
   "source_url": "https://omlx.ai/benchmarks/performance/rbgzwyf7",
   "tg_tokens": null,
   "tps": 23.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rbgzwyf7-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.2 GB at this context; duplicate traces of the same session: chonyuwu, ol8sfqde, rjvlrqxw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 337.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 337.7 22.8 32.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rbgzwyf7",
   "source_url": "https://omlx.ai/benchmarks/performance/rbgzwyf7",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rbgzwyf7-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.2 GB at this context; duplicate traces of the same session: chonyuwu, ol8sfqde, rjvlrqxw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 332.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 332.6 22.2 33.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rbgzwyf7",
   "source_url": "https://omlx.ai/benchmarks/performance/rbgzwyf7",
   "tg_tokens": null,
   "tps": 22.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rbgzwyf7-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.4 GB at this context; duplicate traces of the same session: chonyuwu, ol8sfqde, rjvlrqxw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 321.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 16k 321.9 21.7 32.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rbgzwyf7",
   "source_url": "https://omlx.ai/benchmarks/performance/rbgzwyf7",
   "tg_tokens": null,
   "tps": 21.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-u6hm82p0-ctx16384",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 32.9 GB; TTFT 24311 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 673.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ6e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 16k tokens 673.9 PP tok/s 59.6 TG tok/s 24311 TTFT (ms) 32.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u6hm82p0",
   "source_url": "https://omlx.ai/benchmarks/performance/u6hm82p0",
   "tg_tokens": null,
   "tps": 59.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-xqltjtpo-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 3hkr2x8v, pqi2kxlz, ye45yn0j; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 105.9 15.6 18.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xqltjtpo",
   "source_url": "https://omlx.ai/benchmarks/performance/xqltjtpo",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-xqltjtpo-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 3hkr2x8v, pqi2kxlz, ye45yn0j; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 151.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 151.3 15.0 19.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xqltjtpo",
   "source_url": "https://omlx.ai/benchmarks/performance/xqltjtpo",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-xqltjtpo-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 3hkr2x8v, pqi2kxlz, ye45yn0j; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 115.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 115.0 14.5 19.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xqltjtpo",
   "source_url": "https://omlx.ai/benchmarks/performance/xqltjtpo",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-xqltjtpo-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 3hkr2x8v, pqi2kxlz, ye45yn0j; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 96.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 96.1 12.8 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xqltjtpo",
   "source_url": "https://omlx.ai/benchmarks/performance/xqltjtpo",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-83bgxyo1-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: ilknboft, jg59xbr1, r4rpzc2d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2066.0,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Ornith-1.5-35B-A3B-mlx-nvfp4 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 2,066 115.1 22.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/83bgxyo1",
   "source_url": "https://omlx.ai/benchmarks/performance/83bgxyo1",
   "tg_tokens": null,
   "tps": 115.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-83bgxyo1-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: ilknboft, jg59xbr1, r4rpzc2d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1698.0,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Ornith-1.5-35B-A3B-mlx-nvfp4 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,698 105.1 23.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/83bgxyo1",
   "source_url": "https://omlx.ai/benchmarks/performance/83bgxyo1",
   "tg_tokens": null,
   "tps": 105.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-83bgxyo1-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: ilknboft, jg59xbr1, r4rpzc2d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 934.2,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Ornith-1.5-35B-A3B-mlx-nvfp4 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 934.2 75.9 24.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/83bgxyo1",
   "source_url": "https://omlx.ai/benchmarks/performance/83bgxyo1",
   "tg_tokens": null,
   "tps": 75.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-83bgxyo1-ctx199680",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: ilknboft, jg59xbr1, r4rpzc2d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 418.0,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Ornith-1.5-35B-A3B-mlx-nvfp4 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 195k 418.0 34.0 28.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/83bgxyo1",
   "source_url": "https://omlx.ai/benchmarks/performance/83bgxyo1",
   "tg_tokens": null,
   "tps": 34.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-skf9vdcv-ctx131072",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 103.2 GB; TTFT 182014 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 720.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 128k tokens 720.1 PP tok/s 34.0 TG tok/s 182014 TTFT (ms) 103.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/skf9vdcv",
   "source_url": "https://omlx.ai/benchmarks/performance/skf9vdcv",
   "tg_tokens": null,
   "tps": 34.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-fhiqx9px-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: cr9gkj7l; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 719.6,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 4k 719.6 55.4 29.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fhiqx9px",
   "source_url": "https://omlx.ai/benchmarks/performance/fhiqx9px",
   "tg_tokens": null,
   "tps": 55.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-fhiqx9px-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: cr9gkj7l; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 448.7,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 32k 448.7 46.1 32.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fhiqx9px",
   "source_url": "https://omlx.ai/benchmarks/performance/fhiqx9px",
   "tg_tokens": null,
   "tps": 46.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-w4hxkkuf-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: gf00iz2i; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 72.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 72.9 14.4 18.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w4hxkkuf",
   "source_url": "https://omlx.ai/benchmarks/performance/w4hxkkuf",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-w4hxkkuf-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: gf00iz2i; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 77.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 77.2 15.5 22.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w4hxkkuf",
   "source_url": "https://omlx.ai/benchmarks/performance/w4hxkkuf",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-k13o5edp-ctx1024",
   "model": "GLM-5",
   "notes": "oMLX community benchmark trace; peak memory 391.8 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 186.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 186.0 16.6 391.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k13o5edp",
   "source_url": "https://omlx.ai/benchmarks/performance/k13o5edp",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-k13o5edp-ctx4096",
   "model": "GLM-5",
   "notes": "oMLX community benchmark trace; peak memory 394.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 180.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 180.1 13.8 394.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k13o5edp",
   "source_url": "https://omlx.ai/benchmarks/performance/k13o5edp",
   "tg_tokens": null,
   "tps": 13.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-k13o5edp-ctx8192",
   "model": "GLM-5",
   "notes": "oMLX community benchmark trace; peak memory 396.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 154.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 154.2 13.1 396.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k13o5edp",
   "source_url": "https://omlx.ai/benchmarks/performance/k13o5edp",
   "tg_tokens": null,
   "tps": 13.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-k13o5edp-ctx16384",
   "model": "GLM-5",
   "notes": "oMLX community benchmark trace; peak memory 402.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 117.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 117.3 12.0 402.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k13o5edp",
   "source_url": "https://omlx.ai/benchmarks/performance/k13o5edp",
   "tg_tokens": null,
   "tps": 12.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-k13o5edp-ctx32768",
   "model": "GLM-5",
   "notes": "oMLX community benchmark trace; peak memory 415.4 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 77.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 77.3 10.4 415.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k13o5edp",
   "source_url": "https://omlx.ai/benchmarks/performance/k13o5edp",
   "tg_tokens": null,
   "tps": 10.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-scvy0oc3-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 61.6 GB at this context; duplicate traces of the same session: bg1g56a9, ejsmx46o, ipzlht96, u1zg06ml; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 464.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 1k 464.3 53.4 61.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/scvy0oc3",
   "source_url": "https://omlx.ai/benchmarks/performance/scvy0oc3",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-scvy0oc3-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 61.7 GB at this context; duplicate traces of the same session: bg1g56a9, ejsmx46o, ipzlht96, u1zg06ml; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 443.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 4k 443.4 50.2 61.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/scvy0oc3",
   "source_url": "https://omlx.ai/benchmarks/performance/scvy0oc3",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-scvy0oc3-ctx8192",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 62.0 GB at this context; duplicate traces of the same session: bg1g56a9, ejsmx46o, ipzlht96, u1zg06ml; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 427.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 8k 427.1 46.1 62.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/scvy0oc3",
   "source_url": "https://omlx.ai/benchmarks/performance/scvy0oc3",
   "tg_tokens": null,
   "tps": 46.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-scvy0oc3-ctx16384",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 62.4 GB at this context; duplicate traces of the same session: bg1g56a9, ejsmx46o, ipzlht96, u1zg06ml; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 412.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 16k 412.9 38.1 62.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/scvy0oc3",
   "source_url": "https://omlx.ai/benchmarks/performance/scvy0oc3",
   "tg_tokens": null,
   "tps": 38.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-scvy0oc3-ctx32768",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 63.3 GB at this context; duplicate traces of the same session: bg1g56a9, ejsmx46o, ipzlht96, u1zg06ml; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 379.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 32k 379.4 25.5 63.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/scvy0oc3",
   "source_url": "https://omlx.ai/benchmarks/performance/scvy0oc3",
   "tg_tokens": null,
   "tps": 25.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-6b8emmf6-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 61.6 GB at this context; duplicate traces of the same session: bgwyirz3, l7j9n9kx, n6v3m0w2, o2wbgz1d, p9nbpjyd; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 751.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 1k 751.3 57.9 61.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6b8emmf6",
   "source_url": "https://omlx.ai/benchmarks/performance/6b8emmf6",
   "tg_tokens": null,
   "tps": 57.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-6b8emmf6-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 61.7 GB at this context; duplicate traces of the same session: bgwyirz3, l7j9n9kx, n6v3m0w2, o2wbgz1d, p9nbpjyd; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 966.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 4k 966.4 56.2 61.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6b8emmf6",
   "source_url": "https://omlx.ai/benchmarks/performance/6b8emmf6",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-6b8emmf6-ctx8192",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 62.0 GB at this context; duplicate traces of the same session: bgwyirz3, l7j9n9kx, n6v3m0w2, o2wbgz1d, p9nbpjyd; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 986.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 8k 986.0 53.4 62.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6b8emmf6",
   "source_url": "https://omlx.ai/benchmarks/performance/6b8emmf6",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-6b8emmf6-ctx16384",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 62.4 GB at this context; duplicate traces of the same session: bgwyirz3, l7j9n9kx, n6v3m0w2, o2wbgz1d, p9nbpjyd; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 956.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 16k 956.5 47.0 62.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6b8emmf6",
   "source_url": "https://omlx.ai/benchmarks/performance/6b8emmf6",
   "tg_tokens": null,
   "tps": 47.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-6b8emmf6-ctx32768",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 63.3 GB at this context; duplicate traces of the same session: bgwyirz3, l7j9n9kx, n6v3m0w2, o2wbgz1d, p9nbpjyd; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 881.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 32k 881.3 39.1 63.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6b8emmf6",
   "source_url": "https://omlx.ai/benchmarks/performance/6b8emmf6",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-6b8emmf6-ctx65536",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 65.1 GB at this context; duplicate traces of the same session: bgwyirz3, l7j9n9kx, n6v3m0w2, o2wbgz1d, p9nbpjyd; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 731.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 64k 731.4 28.9 65.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6b8emmf6",
   "source_url": "https://omlx.ai/benchmarks/performance/6b8emmf6",
   "tg_tokens": null,
   "tps": 28.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jm19565y-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.0 GB at this context; duplicate traces of the same session: 40u5dytk, dklk8tmi; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1298.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 1,298 73.9 43.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jm19565y",
   "source_url": "https://omlx.ai/benchmarks/performance/jm19565y",
   "tg_tokens": null,
   "tps": 73.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jm19565y-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.2 GB at this context; duplicate traces of the same session: 40u5dytk, dklk8tmi; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1558.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 1,558 71.0 43.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jm19565y",
   "source_url": "https://omlx.ai/benchmarks/performance/jm19565y",
   "tg_tokens": null,
   "tps": 71.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jm19565y-ctx8192",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.4 GB at this context; duplicate traces of the same session: 40u5dytk, dklk8tmi; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1552.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,552 66.4 43.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jm19565y",
   "source_url": "https://omlx.ai/benchmarks/performance/jm19565y",
   "tg_tokens": null,
   "tps": 66.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jm19565y-ctx16384",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.9 GB at this context; duplicate traces of the same session: 40u5dytk, dklk8tmi; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1461.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 1,461 58.9 43.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jm19565y",
   "source_url": "https://omlx.ai/benchmarks/performance/jm19565y",
   "tg_tokens": null,
   "tps": 58.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jm19565y-ctx32768",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 44.7 GB at this context; duplicate traces of the same session: 40u5dytk, dklk8tmi; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1281.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 1,281 49.4 44.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jm19565y",
   "source_url": "https://omlx.ai/benchmarks/performance/jm19565y",
   "tg_tokens": null,
   "tps": 49.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-jm19565y-ctx65536",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 46.5 GB at this context; duplicate traces of the same session: 40u5dytk, dklk8tmi; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1007.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 1,007 36.9 46.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jm19565y",
   "source_url": "https://omlx.ai/benchmarks/performance/jm19565y",
   "tg_tokens": null,
   "tps": 36.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-yx3zefz3-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 80.1 GB at this context; duplicate traces of the same session: 9kt8xfk2, f94m87x4, u2k59i3j, y1fs7wsv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1442.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 1,442 58.8 80.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yx3zefz3",
   "source_url": "https://omlx.ai/benchmarks/performance/yx3zefz3",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-yx3zefz3-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 80.3 GB at this context; duplicate traces of the same session: 9kt8xfk2, f94m87x4, u2k59i3j, y1fs7wsv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1872.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 1,872 58.5 80.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yx3zefz3",
   "source_url": "https://omlx.ai/benchmarks/performance/yx3zefz3",
   "tg_tokens": null,
   "tps": 58.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-yx3zefz3-ctx8192",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 80.5 GB at this context; duplicate traces of the same session: 9kt8xfk2, f94m87x4, u2k59i3j, y1fs7wsv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1888.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 1,888 55.2 80.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yx3zefz3",
   "source_url": "https://omlx.ai/benchmarks/performance/yx3zefz3",
   "tg_tokens": null,
   "tps": 55.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-yx3zefz3-ctx16384",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 80.9 GB at this context; duplicate traces of the same session: 9kt8xfk2, f94m87x4, u2k59i3j, y1fs7wsv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1783.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k 1,783 49.9 80.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yx3zefz3",
   "source_url": "https://omlx.ai/benchmarks/performance/yx3zefz3",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-yx3zefz3-ctx32768",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 81.8 GB at this context; duplicate traces of the same session: 9kt8xfk2, f94m87x4, u2k59i3j, y1fs7wsv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1549.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 32k 1,549 42.5 81.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yx3zefz3",
   "source_url": "https://omlx.ai/benchmarks/performance/yx3zefz3",
   "tg_tokens": null,
   "tps": 42.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-kjqbv5gt-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.3 GB at this context; duplicate traces of the same session: 0e4ftp2f, 2xv9i0r9, 47gtj6nx, 5fzqvnit; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 229.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 229.0 34.8 66.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kjqbv5gt",
   "source_url": "https://omlx.ai/benchmarks/performance/kjqbv5gt",
   "tg_tokens": null,
   "tps": 34.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-kjqbv5gt-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.7 GB at this context; duplicate traces of the same session: 0e4ftp2f, 2xv9i0r9, 47gtj6nx, 5fzqvnit; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 196.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 196.2 31.0 66.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kjqbv5gt",
   "source_url": "https://omlx.ai/benchmarks/performance/kjqbv5gt",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-kjqbv5gt-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.0 GB at this context; duplicate traces of the same session: 0e4ftp2f, 2xv9i0r9, 47gtj6nx, 5fzqvnit; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 192.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 192.2 27.8 67.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kjqbv5gt",
   "source_url": "https://omlx.ai/benchmarks/performance/kjqbv5gt",
   "tg_tokens": null,
   "tps": 27.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-kjqbv5gt-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.7 GB at this context; duplicate traces of the same session: 0e4ftp2f, 2xv9i0r9, 47gtj6nx, 5fzqvnit; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 183.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 183.0 22.5 67.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kjqbv5gt",
   "source_url": "https://omlx.ai/benchmarks/performance/kjqbv5gt",
   "tg_tokens": null,
   "tps": 22.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-kjqbv5gt-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 69.1 GB at this context; duplicate traces of the same session: 0e4ftp2f, 2xv9i0r9, 47gtj6nx, 5fzqvnit; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 170.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 32k 170.3 16.8 69.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kjqbv5gt",
   "source_url": "https://omlx.ai/benchmarks/performance/kjqbv5gt",
   "tg_tokens": null,
   "tps": 16.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-0mvxl5m1-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.3 GB at this context; duplicate traces of the same session: 2k6ofxl2, hfzzor6h, kv66u88z, u28zirxu, vrw9n3xu; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 439.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 1k 439.6 47.8 66.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0mvxl5m1",
   "source_url": "https://omlx.ai/benchmarks/performance/0mvxl5m1",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-0mvxl5m1-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.7 GB at this context; duplicate traces of the same session: 2k6ofxl2, hfzzor6h, kv66u88z, u28zirxu, vrw9n3xu; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 463.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 4k 463.3 45.6 66.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0mvxl5m1",
   "source_url": "https://omlx.ai/benchmarks/performance/0mvxl5m1",
   "tg_tokens": null,
   "tps": 45.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-0mvxl5m1-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.0 GB at this context; duplicate traces of the same session: 2k6ofxl2, hfzzor6h, kv66u88z, u28zirxu, vrw9n3xu; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 459.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 8k 459.9 41.3 67.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0mvxl5m1",
   "source_url": "https://omlx.ai/benchmarks/performance/0mvxl5m1",
   "tg_tokens": null,
   "tps": 41.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-0mvxl5m1-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.7 GB at this context; duplicate traces of the same session: 2k6ofxl2, hfzzor6h, kv66u88z, u28zirxu, vrw9n3xu; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 444.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 16k 444.0 36.3 67.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0mvxl5m1",
   "source_url": "https://omlx.ai/benchmarks/performance/0mvxl5m1",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-0mvxl5m1-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 69.1 GB at this context; duplicate traces of the same session: 2k6ofxl2, hfzzor6h, kv66u88z, u28zirxu, vrw9n3xu; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 409.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 32k 409.4 28.3 69.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0mvxl5m1",
   "source_url": "https://omlx.ai/benchmarks/performance/0mvxl5m1",
   "tg_tokens": null,
   "tps": 28.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-0mvxl5m1-ctx65536",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 71.9 GB at this context; duplicate traces of the same session: 2k6ofxl2, hfzzor6h, kv66u88z, u28zirxu, vrw9n3xu; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 331.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 64k 331.7 19.0 71.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0mvxl5m1",
   "source_url": "https://omlx.ai/benchmarks/performance/0mvxl5m1",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vayuu764-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.3 GB at this context; duplicate traces of the same session: lm7yt1ag; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 904.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 904.5 52.5 66.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vayuu764",
   "source_url": "https://omlx.ai/benchmarks/performance/vayuu764",
   "tg_tokens": null,
   "tps": 52.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vayuu764-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.7 GB at this context; duplicate traces of the same session: lm7yt1ag; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 951.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 951.6 52.0 66.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vayuu764",
   "source_url": "https://omlx.ai/benchmarks/performance/vayuu764",
   "tg_tokens": null,
   "tps": 52.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5u8amia4-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 123.2 GB at this context; duplicate traces of the same session: 77dim65a, ho7w5nty, kbpdn60e, z1bl3y33; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 837.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 837.2 39.1 123.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5u8amia4",
   "source_url": "https://omlx.ai/benchmarks/performance/5u8amia4",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5u8amia4-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 123.5 GB at this context; duplicate traces of the same session: 77dim65a, ho7w5nty, kbpdn60e, z1bl3y33; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 895.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 895.1 37.8 123.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5u8amia4",
   "source_url": "https://omlx.ai/benchmarks/performance/5u8amia4",
   "tg_tokens": null,
   "tps": 37.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5u8amia4-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 123.9 GB at this context; duplicate traces of the same session: 77dim65a, ho7w5nty, kbpdn60e, z1bl3y33; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 876.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 876.6 35.7 123.9 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5u8amia4",
   "source_url": "https://omlx.ai/benchmarks/performance/5u8amia4",
   "tg_tokens": null,
   "tps": 35.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5u8amia4-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 124.5 GB at this context; duplicate traces of the same session: 77dim65a, ho7w5nty, kbpdn60e, z1bl3y33; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 826.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k 826.4 32.5 124.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5u8amia4",
   "source_url": "https://omlx.ai/benchmarks/performance/5u8amia4",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5u8amia4-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 125.9 GB at this context; duplicate traces of the same session: 77dim65a, ho7w5nty, kbpdn60e, z1bl3y33; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 717.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 32k 717.8 27.9 125.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5u8amia4",
   "source_url": "https://omlx.ai/benchmarks/performance/5u8amia4",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-h08jztak-ctx1024",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 754ncwmf, 76f7vm8m, iugrfis1, yiokj60f; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 375.2,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 1k 375.2 26.7 22.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h08jztak",
   "source_url": "https://omlx.ai/benchmarks/performance/h08jztak",
   "tg_tokens": null,
   "tps": 26.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-h08jztak-ctx4096",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: 754ncwmf, 76f7vm8m, iugrfis1, yiokj60f; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 401.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 4k 401.1 25.4 22.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h08jztak",
   "source_url": "https://omlx.ai/benchmarks/performance/h08jztak",
   "tg_tokens": null,
   "tps": 25.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-h08jztak-ctx8192",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: 754ncwmf, 76f7vm8m, iugrfis1, yiokj60f; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 400.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 8k 400.8 24.4 22.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h08jztak",
   "source_url": "https://omlx.ai/benchmarks/performance/h08jztak",
   "tg_tokens": null,
   "tps": 24.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-h08jztak-ctx16384",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: 754ncwmf, 76f7vm8m, iugrfis1, yiokj60f; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 392.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 16k 392.6 22.4 23.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h08jztak",
   "source_url": "https://omlx.ai/benchmarks/performance/h08jztak",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-h08jztak-ctx32768",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: 754ncwmf, 76f7vm8m, iugrfis1, yiokj60f; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 370.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 32k 370.0 19.5 25.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h08jztak",
   "source_url": "https://omlx.ai/benchmarks/performance/h08jztak",
   "tg_tokens": null,
   "tps": 19.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-itu5wwop-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.6 GB at this context; duplicate traces of the same session: 7188vfes; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 218.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-9B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 6bit; 1k 218.0 15.7 8.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/itu5wwop",
   "source_url": "https://omlx.ai/benchmarks/performance/itu5wwop",
   "tg_tokens": null,
   "tps": 15.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-itu5wwop-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.8 GB at this context; duplicate traces of the same session: 7188vfes; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 218.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-9B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 6bit; 4k 218.9 15.3 8.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/itu5wwop",
   "source_url": "https://omlx.ai/benchmarks/performance/itu5wwop",
   "tg_tokens": null,
   "tps": 15.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-a4z2odp6-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 8.6 GB; TTFT 4721 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 216.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-9B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 6bit; 1k tokens 216.9 PP tok/s 15.7 TG tok/s 4721 TTFT (ms) 8.6 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a4z2odp6",
   "source_url": "https://omlx.ai/benchmarks/performance/a4z2odp6",
   "tg_tokens": null,
   "tps": 15.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-7a3wu6hq-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.2 GB at this context; duplicate traces of the same session: ak90o81b; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 215.4,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Qwen3.5-9B-mxfp4 on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 215.4 26.1 7.2 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7a3wu6hq",
   "source_url": "https://omlx.ai/benchmarks/performance/7a3wu6hq",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-7a3wu6hq-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.7 GB at this context; duplicate traces of the same session: ak90o81b; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 214.5,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Qwen3.5-9B-mxfp4 on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 214.5 25.3 7.7 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7a3wu6hq",
   "source_url": "https://omlx.ai/benchmarks/performance/7a3wu6hq",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-w2lll5k0-ctx1024",
   "model": "gpt-oss 120B",
   "notes": "oMLX community benchmark trace; peak memory 58.0 GB at this context; duplicate traces of the same session: 3gd322b3, 5oxq1gs1, nqo8pz28, vfwc00k8; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 187.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gpt-oss-120b on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 187.7 54.9 58.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w2lll5k0",
   "source_url": "https://omlx.ai/benchmarks/performance/w2lll5k0",
   "tg_tokens": null,
   "tps": 54.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-w2lll5k0-ctx4096",
   "model": "gpt-oss 120B",
   "notes": "oMLX community benchmark trace; peak memory 58.1 GB at this context; duplicate traces of the same session: 3gd322b3, 5oxq1gs1, nqo8pz28, vfwc00k8; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 182.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gpt-oss-120b on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 182.0 44.7 58.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w2lll5k0",
   "source_url": "https://omlx.ai/benchmarks/performance/w2lll5k0",
   "tg_tokens": null,
   "tps": 44.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-w2lll5k0-ctx8192",
   "model": "gpt-oss 120B",
   "notes": "oMLX community benchmark trace; peak memory 58.3 GB at this context; duplicate traces of the same session: 3gd322b3, 5oxq1gs1, nqo8pz28, vfwc00k8; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 183.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gpt-oss-120b on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 183.9 37.1 58.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w2lll5k0",
   "source_url": "https://omlx.ai/benchmarks/performance/w2lll5k0",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-w2lll5k0-ctx16384",
   "model": "gpt-oss 120B",
   "notes": "oMLX community benchmark trace; peak memory 58.6 GB at this context; duplicate traces of the same session: 3gd322b3, 5oxq1gs1, nqo8pz28, vfwc00k8; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 177.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gpt-oss-120b on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 177.1 22.7 58.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w2lll5k0",
   "source_url": "https://omlx.ai/benchmarks/performance/w2lll5k0",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-w2lll5k0-ctx32768",
   "model": "gpt-oss 120B",
   "notes": "oMLX community benchmark trace; peak memory 59.2 GB at this context; duplicate traces of the same session: 3gd322b3, 5oxq1gs1, nqo8pz28, vfwc00k8; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 165.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gpt-oss-120b on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 32k 165.3 14.6 59.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w2lll5k0",
   "source_url": "https://omlx.ai/benchmarks/performance/w2lll5k0",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-e4gy4svv-ctx1024",
   "model": "gpt-oss 120B",
   "notes": "oMLX community benchmark trace; peak memory 115.8 GB at this context; duplicate traces of the same session: ivuw3lz5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 569.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gpt-oss-120b on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 569.3 60.7 115.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e4gy4svv",
   "source_url": "https://omlx.ai/benchmarks/performance/e4gy4svv",
   "tg_tokens": null,
   "tps": 60.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-e4gy4svv-ctx4096",
   "model": "gpt-oss 120B",
   "notes": "oMLX community benchmark trace; peak memory 115.9 GB at this context; duplicate traces of the same session: ivuw3lz5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 631.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gpt-oss-120b on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 631.9 54.2 115.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e4gy4svv",
   "source_url": "https://omlx.ai/benchmarks/performance/e4gy4svv",
   "tg_tokens": null,
   "tps": 54.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-osqlhjud-ctx1024",
   "model": "gpt-oss 120B",
   "notes": "MXFP4 experts with Q8 attention; oMLX community benchmark trace; peak memory 59.2 GB at this context; duplicate traces of the same session: 06kdatf0, i54z07a7, vtfd1nia, wjn65xp9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 613.2,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-120b-MXFP4-Q8 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 613.2 68.4 59.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/osqlhjud",
   "source_url": "https://omlx.ai/benchmarks/performance/osqlhjud",
   "tg_tokens": null,
   "tps": 68.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-osqlhjud-ctx4096",
   "model": "gpt-oss 120B",
   "notes": "MXFP4 experts with Q8 attention; oMLX community benchmark trace; peak memory 59.3 GB at this context; duplicate traces of the same session: 06kdatf0, i54z07a7, vtfd1nia, wjn65xp9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 685.5,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-120b-MXFP4-Q8 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 685.5 64.1 59.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/osqlhjud",
   "source_url": "https://omlx.ai/benchmarks/performance/osqlhjud",
   "tg_tokens": null,
   "tps": 64.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-osqlhjud-ctx8192",
   "model": "gpt-oss 120B",
   "notes": "MXFP4 experts with Q8 attention; oMLX community benchmark trace; peak memory 59.5 GB at this context; duplicate traces of the same session: 06kdatf0, i54z07a7, vtfd1nia, wjn65xp9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 687.8,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-120b-MXFP4-Q8 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 687.8 57.8 59.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/osqlhjud",
   "source_url": "https://omlx.ai/benchmarks/performance/osqlhjud",
   "tg_tokens": null,
   "tps": 57.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-osqlhjud-ctx16384",
   "model": "gpt-oss 120B",
   "notes": "MXFP4 experts with Q8 attention; oMLX community benchmark trace; peak memory 59.8 GB at this context; duplicate traces of the same session: 06kdatf0, i54z07a7, vtfd1nia, wjn65xp9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 663.6,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-120b-MXFP4-Q8 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 663.6 47.5 59.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/osqlhjud",
   "source_url": "https://omlx.ai/benchmarks/performance/osqlhjud",
   "tg_tokens": null,
   "tps": 47.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-osqlhjud-ctx32768",
   "model": "gpt-oss 120B",
   "notes": "MXFP4 experts with Q8 attention; oMLX community benchmark trace; peak memory 60.4 GB at this context; duplicate traces of the same session: 06kdatf0, i54z07a7, vtfd1nia, wjn65xp9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 600.7,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-120b-MXFP4-Q8 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 600.7 35.2 60.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/osqlhjud",
   "source_url": "https://omlx.ai/benchmarks/performance/osqlhjud",
   "tg_tokens": null,
   "tps": 35.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-w2ncir8g-ctx1024",
   "model": "GLM-5",
   "notes": "oMLX community benchmark trace; peak memory 391.8 GB at this context; duplicate traces of the same session: 314vh14k, dig8n87z; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 186.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 186.0 16.6 391.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w2ncir8g",
   "source_url": "https://omlx.ai/benchmarks/performance/w2ncir8g",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-w2ncir8g-ctx4096",
   "model": "GLM-5",
   "notes": "oMLX community benchmark trace; peak memory 394.1 GB at this context; duplicate traces of the same session: 314vh14k, dig8n87z; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 180.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 180.1 13.8 394.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w2ncir8g",
   "source_url": "https://omlx.ai/benchmarks/performance/w2ncir8g",
   "tg_tokens": null,
   "tps": 13.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-w2ncir8g-ctx8192",
   "model": "GLM-5",
   "notes": "oMLX community benchmark trace; peak memory 396.7 GB at this context; duplicate traces of the same session: 314vh14k, dig8n87z; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 154.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 154.2 13.1 396.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w2ncir8g",
   "source_url": "https://omlx.ai/benchmarks/performance/w2ncir8g",
   "tg_tokens": null,
   "tps": 13.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-w2ncir8g-ctx16384",
   "model": "GLM-5",
   "notes": "oMLX community benchmark trace; peak memory 402.7 GB at this context; duplicate traces of the same session: 314vh14k, dig8n87z; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 117.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 117.3 12.0 402.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w2ncir8g",
   "source_url": "https://omlx.ai/benchmarks/performance/w2ncir8g",
   "tg_tokens": null,
   "tps": 12.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-w2ncir8g-ctx32768",
   "model": "GLM-5",
   "notes": "oMLX community benchmark trace; peak memory 415.4 GB at this context; duplicate traces of the same session: 314vh14k, dig8n87z; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 77.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 77.3 10.4 415.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w2ncir8g",
   "source_url": "https://omlx.ai/benchmarks/performance/w2ncir8g",
   "tg_tokens": null,
   "tps": 10.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-i9zl3yj0-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.1 GB at this context; duplicate traces of the same session: fqqdyr66, t9shfw5c; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2283.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,283 127.2 177.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i9zl3yj0",
   "source_url": "https://omlx.ai/benchmarks/performance/i9zl3yj0",
   "tg_tokens": null,
   "tps": 127.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-i9zl3yj0-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.6 GB at this context; duplicate traces of the same session: fqqdyr66, t9shfw5c; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2252.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,252 88.0 177.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i9zl3yj0",
   "source_url": "https://omlx.ai/benchmarks/performance/i9zl3yj0",
   "tg_tokens": null,
   "tps": 88.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-i9zl3yj0-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.8 GB at this context; duplicate traces of the same session: fqqdyr66, t9shfw5c; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2202.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,202 140.2 178.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i9zl3yj0",
   "source_url": "https://omlx.ai/benchmarks/performance/i9zl3yj0",
   "tg_tokens": null,
   "tps": 140.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-c2vsozrs-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.1 GB at this context; duplicate traces of the same session: 9sc95md1, goj6q9vt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2221.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,221 100.0 178.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c2vsozrs",
   "source_url": "https://omlx.ai/benchmarks/performance/c2vsozrs",
   "tg_tokens": null,
   "tps": 100.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-c2vsozrs-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.6 GB at this context; duplicate traces of the same session: 9sc95md1, goj6q9vt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2198.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,198 74.4 178.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c2vsozrs",
   "source_url": "https://omlx.ai/benchmarks/performance/c2vsozrs",
   "tg_tokens": null,
   "tps": 74.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-c2vsozrs-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 179.9 GB at this context; duplicate traces of the same session: 9sc95md1, goj6q9vt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2155.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,155 126.4 179.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c2vsozrs",
   "source_url": "https://omlx.ai/benchmarks/performance/c2vsozrs",
   "tg_tokens": null,
   "tps": 126.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-kp0yq5zn-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.1 GB at this context; duplicate traces of the same session: rpi89413, v2rhfwv0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2225.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,225 108.7 178.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kp0yq5zn",
   "source_url": "https://omlx.ai/benchmarks/performance/kp0yq5zn",
   "tg_tokens": null,
   "tps": 108.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-kp0yq5zn-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.6 GB at this context; duplicate traces of the same session: rpi89413, v2rhfwv0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2199.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,199 80.0 178.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kp0yq5zn",
   "source_url": "https://omlx.ai/benchmarks/performance/kp0yq5zn",
   "tg_tokens": null,
   "tps": 80.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-kp0yq5zn-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 179.9 GB at this context; duplicate traces of the same session: rpi89413, v2rhfwv0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2156.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,156 113.8 179.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kp0yq5zn",
   "source_url": "https://omlx.ai/benchmarks/performance/kp0yq5zn",
   "tg_tokens": null,
   "tps": 113.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dqv8j3yb-ctx65536",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 4.5 GB at this context; duplicate traces of the same session: wwnhcy1a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2901.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM5-2B-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,901 87.5 4.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dqv8j3yb",
   "source_url": "https://omlx.ai/benchmarks/performance/dqv8j3yb",
   "tg_tokens": null,
   "tps": 87.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dqv8j3yb-ctx199680",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 9.7 GB at this context; duplicate traces of the same session: wwnhcy1a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 991.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM5-2B-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 991.2 32.5 9.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dqv8j3yb",
   "source_url": "https://omlx.ai/benchmarks/performance/dqv8j3yb",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-2314huk0-ctx1024",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 94.1 GB at this context; duplicate traces of the same session: 2kpxm1c1, jnc9tgs9, s6lyqz2e; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 299.7,
   "provenance": "community",
   "quant": "3bit",
   "quote": "MiniMax-M2.5 on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 3bit; 1k 299.7 42.1 94.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2314huk0",
   "source_url": "https://omlx.ai/benchmarks/performance/2314huk0",
   "tg_tokens": null,
   "tps": 42.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-2314huk0-ctx4096",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 95.1 GB at this context; duplicate traces of the same session: 2kpxm1c1, jnc9tgs9, s6lyqz2e; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 337.6,
   "provenance": "community",
   "quant": "3bit",
   "quote": "MiniMax-M2.5 on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 3bit; 4k 337.6 35.7 95.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2314huk0",
   "source_url": "https://omlx.ai/benchmarks/performance/2314huk0",
   "tg_tokens": null,
   "tps": 35.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-2314huk0-ctx8192",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 96.0 GB at this context; duplicate traces of the same session: 2kpxm1c1, jnc9tgs9, s6lyqz2e; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 320.5,
   "provenance": "community",
   "quant": "3bit",
   "quote": "MiniMax-M2.5 on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 3bit; 8k 320.5 29.5 96.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2314huk0",
   "source_url": "https://omlx.ai/benchmarks/performance/2314huk0",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-2314huk0-ctx16384",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 98.0 GB at this context; duplicate traces of the same session: 2kpxm1c1, jnc9tgs9, s6lyqz2e; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 282.6,
   "provenance": "community",
   "quant": "3bit",
   "quote": "MiniMax-M2.5 on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 3bit; 16k 282.6 22.4 98.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2314huk0",
   "source_url": "https://omlx.ai/benchmarks/performance/2314huk0",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-8v14iy81-ctx1024",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 174.0 GB at this context; duplicate traces of the same session: 2we8lcjq, bzeg02vo, e48q54i5, wjvi0q0a, xbfycdsw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 496.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 1k 496.7 40.0 174.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8v14iy81",
   "source_url": "https://omlx.ai/benchmarks/performance/8v14iy81",
   "tg_tokens": null,
   "tps": 40.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-8v14iy81-ctx4096",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 175.0 GB at this context; duplicate traces of the same session: 2we8lcjq, bzeg02vo, e48q54i5, wjvi0q0a, xbfycdsw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 554.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 4k 554.8 33.9 175.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8v14iy81",
   "source_url": "https://omlx.ai/benchmarks/performance/8v14iy81",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-8v14iy81-ctx8192",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 175.9 GB at this context; duplicate traces of the same session: 2we8lcjq, bzeg02vo, e48q54i5, wjvi0q0a, xbfycdsw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 514.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 8k 514.0 28.9 175.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8v14iy81",
   "source_url": "https://omlx.ai/benchmarks/performance/8v14iy81",
   "tg_tokens": null,
   "tps": 28.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-8v14iy81-ctx16384",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 177.8 GB at this context; duplicate traces of the same session: 2we8lcjq, bzeg02vo, e48q54i5, wjvi0q0a, xbfycdsw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 436.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 16k 436.3 23.5 177.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8v14iy81",
   "source_url": "https://omlx.ai/benchmarks/performance/8v14iy81",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-8v14iy81-ctx32768",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 181.7 GB at this context; duplicate traces of the same session: 2we8lcjq, bzeg02vo, e48q54i5, wjvi0q0a, xbfycdsw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 330.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 32k 330.8 16.1 181.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8v14iy81",
   "source_url": "https://omlx.ai/benchmarks/performance/8v14iy81",
   "tg_tokens": null,
   "tps": 16.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-8v14iy81-ctx65536",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 189.6 GB at this context; duplicate traces of the same session: 2we8lcjq, bzeg02vo, e48q54i5, wjvi0q0a, xbfycdsw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 221.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 64k 221.1 9.6 189.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8v14iy81",
   "source_url": "https://omlx.ai/benchmarks/performance/8v14iy81",
   "tg_tokens": null,
   "tps": 9.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-6f5a4eko-ctx1024",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 227.2 GB at this context; duplicate traces of the same session: jye2fo4m, rf82ex1p, s15twedr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 600.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 600.4 34.2 227.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6f5a4eko",
   "source_url": "https://omlx.ai/benchmarks/performance/6f5a4eko",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-6f5a4eko-ctx4096",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 228.2 GB at this context; duplicate traces of the same session: jye2fo4m, rf82ex1p, s15twedr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 707.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 707.9 30.1 228.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6f5a4eko",
   "source_url": "https://omlx.ai/benchmarks/performance/6f5a4eko",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-6f5a4eko-ctx8192",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 229.1 GB at this context; duplicate traces of the same session: jye2fo4m, rf82ex1p, s15twedr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 663.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 663.9 25.9 229.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6f5a4eko",
   "source_url": "https://omlx.ai/benchmarks/performance/6f5a4eko",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-6f5a4eko-ctx16384",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 231.1 GB at this context; duplicate traces of the same session: jye2fo4m, rf82ex1p, s15twedr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 564.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k 564.2 21.1 231.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6f5a4eko",
   "source_url": "https://omlx.ai/benchmarks/performance/6f5a4eko",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-6f5a4eko-ctx32768",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 235.0 GB at this context; duplicate traces of the same session: jye2fo4m, rf82ex1p, s15twedr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 427.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 32k 427.8 14.6 235.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6f5a4eko",
   "source_url": "https://omlx.ai/benchmarks/performance/6f5a4eko",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-r4rpzc2d-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: ilknboft, jg59xbr1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2066.0,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Ornith-1.5-35B-A3B-mlx-nvfp4 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 2,066 115.1 22.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r4rpzc2d",
   "source_url": "https://omlx.ai/benchmarks/performance/r4rpzc2d",
   "tg_tokens": null,
   "tps": 115.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-r4rpzc2d-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: ilknboft, jg59xbr1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1698.0,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Ornith-1.5-35B-A3B-mlx-nvfp4 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,698 105.1 23.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r4rpzc2d",
   "source_url": "https://omlx.ai/benchmarks/performance/r4rpzc2d",
   "tg_tokens": null,
   "tps": 105.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-r4rpzc2d-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: ilknboft, jg59xbr1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 934.2,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Ornith-1.5-35B-A3B-mlx-nvfp4 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 934.2 75.9 24.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r4rpzc2d",
   "source_url": "https://omlx.ai/benchmarks/performance/r4rpzc2d",
   "tg_tokens": null,
   "tps": 75.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-r4rpzc2d-ctx199680",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: ilknboft, jg59xbr1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 418.0,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Ornith-1.5-35B-A3B-mlx-nvfp4 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 195k 418.0 34.0 28.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r4rpzc2d",
   "source_url": "https://omlx.ai/benchmarks/performance/r4rpzc2d",
   "tg_tokens": null,
   "tps": 34.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-u1zg06ml-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 61.6 GB at this context; duplicate traces of the same session: bg1g56a9, ejsmx46o, ipzlht96; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 464.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 1k 464.3 53.4 61.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u1zg06ml",
   "source_url": "https://omlx.ai/benchmarks/performance/u1zg06ml",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-u1zg06ml-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 61.7 GB at this context; duplicate traces of the same session: bg1g56a9, ejsmx46o, ipzlht96; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 443.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 4k 443.4 50.2 61.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u1zg06ml",
   "source_url": "https://omlx.ai/benchmarks/performance/u1zg06ml",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-u1zg06ml-ctx8192",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 62.0 GB at this context; duplicate traces of the same session: bg1g56a9, ejsmx46o, ipzlht96; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 427.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 8k 427.1 46.1 62.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u1zg06ml",
   "source_url": "https://omlx.ai/benchmarks/performance/u1zg06ml",
   "tg_tokens": null,
   "tps": 46.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-u1zg06ml-ctx16384",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 62.4 GB at this context; duplicate traces of the same session: bg1g56a9, ejsmx46o, ipzlht96; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 412.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 16k 412.9 38.1 62.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u1zg06ml",
   "source_url": "https://omlx.ai/benchmarks/performance/u1zg06ml",
   "tg_tokens": null,
   "tps": 38.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-u1zg06ml-ctx32768",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 63.3 GB at this context; duplicate traces of the same session: bg1g56a9, ejsmx46o, ipzlht96; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 379.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 32k 379.4 25.5 63.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u1zg06ml",
   "source_url": "https://omlx.ai/benchmarks/performance/u1zg06ml",
   "tg_tokens": null,
   "tps": 25.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-o2wbgz1d-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 61.6 GB at this context; duplicate traces of the same session: bgwyirz3, l7j9n9kx, n6v3m0w2, p9nbpjyd; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 751.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 1k 751.3 57.9 61.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o2wbgz1d",
   "source_url": "https://omlx.ai/benchmarks/performance/o2wbgz1d",
   "tg_tokens": null,
   "tps": 57.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-o2wbgz1d-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 61.7 GB at this context; duplicate traces of the same session: bgwyirz3, l7j9n9kx, n6v3m0w2, p9nbpjyd; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 966.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 4k 966.4 56.2 61.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o2wbgz1d",
   "source_url": "https://omlx.ai/benchmarks/performance/o2wbgz1d",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-o2wbgz1d-ctx8192",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 62.0 GB at this context; duplicate traces of the same session: bgwyirz3, l7j9n9kx, n6v3m0w2, p9nbpjyd; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 986.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 8k 986.0 53.4 62.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o2wbgz1d",
   "source_url": "https://omlx.ai/benchmarks/performance/o2wbgz1d",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-o2wbgz1d-ctx16384",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 62.4 GB at this context; duplicate traces of the same session: bgwyirz3, l7j9n9kx, n6v3m0w2, p9nbpjyd; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 956.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 16k 956.5 47.0 62.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o2wbgz1d",
   "source_url": "https://omlx.ai/benchmarks/performance/o2wbgz1d",
   "tg_tokens": null,
   "tps": 47.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-o2wbgz1d-ctx32768",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 63.3 GB at this context; duplicate traces of the same session: bgwyirz3, l7j9n9kx, n6v3m0w2, p9nbpjyd; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 881.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 32k 881.3 39.1 63.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o2wbgz1d",
   "source_url": "https://omlx.ai/benchmarks/performance/o2wbgz1d",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-o2wbgz1d-ctx65536",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 65.1 GB at this context; duplicate traces of the same session: bgwyirz3, l7j9n9kx, n6v3m0w2, p9nbpjyd; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 731.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 64k 731.4 28.9 65.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o2wbgz1d",
   "source_url": "https://omlx.ai/benchmarks/performance/o2wbgz1d",
   "tg_tokens": null,
   "tps": 28.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-40u5dytk-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.0 GB at this context; duplicate traces of the same session: dklk8tmi, fdttset4, rihr0r0i, scn0knor; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1298.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 1,298 73.9 43.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/40u5dytk",
   "source_url": "https://omlx.ai/benchmarks/performance/40u5dytk",
   "tg_tokens": null,
   "tps": 73.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-40u5dytk-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.2 GB at this context; duplicate traces of the same session: dklk8tmi, fdttset4, rihr0r0i, scn0knor; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1558.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 1,558 71.0 43.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/40u5dytk",
   "source_url": "https://omlx.ai/benchmarks/performance/40u5dytk",
   "tg_tokens": null,
   "tps": 71.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-40u5dytk-ctx8192",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.4 GB at this context; duplicate traces of the same session: dklk8tmi, fdttset4, rihr0r0i, scn0knor; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1552.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,552 66.4 43.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/40u5dytk",
   "source_url": "https://omlx.ai/benchmarks/performance/40u5dytk",
   "tg_tokens": null,
   "tps": 66.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-40u5dytk-ctx16384",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.9 GB at this context; duplicate traces of the same session: dklk8tmi, fdttset4, rihr0r0i, scn0knor; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1461.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 1,461 58.9 43.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/40u5dytk",
   "source_url": "https://omlx.ai/benchmarks/performance/40u5dytk",
   "tg_tokens": null,
   "tps": 58.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-40u5dytk-ctx32768",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 44.7 GB at this context; duplicate traces of the same session: dklk8tmi, fdttset4, rihr0r0i, scn0knor; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1281.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 1,281 49.4 44.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/40u5dytk",
   "source_url": "https://omlx.ai/benchmarks/performance/40u5dytk",
   "tg_tokens": null,
   "tps": 49.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-40u5dytk-ctx65536",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 46.5 GB at this context; duplicate traces of the same session: dklk8tmi, fdttset4, rihr0r0i, scn0knor; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1007.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 1,007 36.9 46.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/40u5dytk",
   "source_url": "https://omlx.ai/benchmarks/performance/40u5dytk",
   "tg_tokens": null,
   "tps": 36.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rj4m0w2o-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 80.1 GB at this context; duplicate traces of the same session: 75x2gbeu, jrf4vgpr, oylwhkj1, rhiba0g8, xill9rj9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1236.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 1,236 63.9 80.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rj4m0w2o",
   "source_url": "https://omlx.ai/benchmarks/performance/rj4m0w2o",
   "tg_tokens": null,
   "tps": 63.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rj4m0w2o-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 80.3 GB at this context; duplicate traces of the same session: 75x2gbeu, jrf4vgpr, oylwhkj1, rhiba0g8, xill9rj9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1509.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 1,509 61.9 80.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rj4m0w2o",
   "source_url": "https://omlx.ai/benchmarks/performance/rj4m0w2o",
   "tg_tokens": null,
   "tps": 61.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rj4m0w2o-ctx8192",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 80.5 GB at this context; duplicate traces of the same session: 75x2gbeu, jrf4vgpr, oylwhkj1, rhiba0g8, xill9rj9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1521.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 1,521 56.5 80.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rj4m0w2o",
   "source_url": "https://omlx.ai/benchmarks/performance/rj4m0w2o",
   "tg_tokens": null,
   "tps": 56.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rj4m0w2o-ctx16384",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 80.9 GB at this context; duplicate traces of the same session: 75x2gbeu, jrf4vgpr, oylwhkj1, rhiba0g8, xill9rj9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1434.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 16k 1,434 52.0 80.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rj4m0w2o",
   "source_url": "https://omlx.ai/benchmarks/performance/rj4m0w2o",
   "tg_tokens": null,
   "tps": 52.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rj4m0w2o-ctx32768",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 81.8 GB at this context; duplicate traces of the same session: 75x2gbeu, jrf4vgpr, oylwhkj1, rhiba0g8, xill9rj9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1262.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 32k 1,262 45.8 81.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rj4m0w2o",
   "source_url": "https://omlx.ai/benchmarks/performance/rj4m0w2o",
   "tg_tokens": null,
   "tps": 45.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rj4m0w2o-ctx65536",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 83.6 GB at this context; duplicate traces of the same session: 75x2gbeu, jrf4vgpr, oylwhkj1, rhiba0g8, xill9rj9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 994.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 64k 994.8 34.7 83.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rj4m0w2o",
   "source_url": "https://omlx.ai/benchmarks/performance/rj4m0w2o",
   "tg_tokens": null,
   "tps": 34.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-f94m87x4-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 80.1 GB at this context; duplicate traces of the same session: 9kt8xfk2, u2k59i3j, y1fs7wsv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1442.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 1,442 58.8 80.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f94m87x4",
   "source_url": "https://omlx.ai/benchmarks/performance/f94m87x4",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-f94m87x4-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 80.3 GB at this context; duplicate traces of the same session: 9kt8xfk2, u2k59i3j, y1fs7wsv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1872.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 1,872 58.5 80.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f94m87x4",
   "source_url": "https://omlx.ai/benchmarks/performance/f94m87x4",
   "tg_tokens": null,
   "tps": 58.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-f94m87x4-ctx8192",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 80.5 GB at this context; duplicate traces of the same session: 9kt8xfk2, u2k59i3j, y1fs7wsv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1888.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 1,888 55.2 80.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f94m87x4",
   "source_url": "https://omlx.ai/benchmarks/performance/f94m87x4",
   "tg_tokens": null,
   "tps": 55.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-f94m87x4-ctx16384",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 80.9 GB at this context; duplicate traces of the same session: 9kt8xfk2, u2k59i3j, y1fs7wsv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1783.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k 1,783 49.9 80.9 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f94m87x4",
   "source_url": "https://omlx.ai/benchmarks/performance/f94m87x4",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-f94m87x4-ctx32768",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 81.8 GB at this context; duplicate traces of the same session: 9kt8xfk2, u2k59i3j, y1fs7wsv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1549.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 32k 1,549 42.5 81.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f94m87x4",
   "source_url": "https://omlx.ai/benchmarks/performance/f94m87x4",
   "tg_tokens": null,
   "tps": 42.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-0e4ftp2f-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.3 GB at this context; duplicate traces of the same session: 2xv9i0r9, 47gtj6nx, 5fzqvnit; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 229.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 229.0 34.8 66.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0e4ftp2f",
   "source_url": "https://omlx.ai/benchmarks/performance/0e4ftp2f",
   "tg_tokens": null,
   "tps": 34.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-0e4ftp2f-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.7 GB at this context; duplicate traces of the same session: 2xv9i0r9, 47gtj6nx, 5fzqvnit; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 196.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 196.2 31.0 66.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0e4ftp2f",
   "source_url": "https://omlx.ai/benchmarks/performance/0e4ftp2f",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-0e4ftp2f-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.0 GB at this context; duplicate traces of the same session: 2xv9i0r9, 47gtj6nx, 5fzqvnit; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 192.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 192.2 27.8 67.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0e4ftp2f",
   "source_url": "https://omlx.ai/benchmarks/performance/0e4ftp2f",
   "tg_tokens": null,
   "tps": 27.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-0e4ftp2f-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.7 GB at this context; duplicate traces of the same session: 2xv9i0r9, 47gtj6nx, 5fzqvnit; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 183.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 183.0 22.5 67.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0e4ftp2f",
   "source_url": "https://omlx.ai/benchmarks/performance/0e4ftp2f",
   "tg_tokens": null,
   "tps": 22.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-0e4ftp2f-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 69.1 GB at this context; duplicate traces of the same session: 2xv9i0r9, 47gtj6nx, 5fzqvnit; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 170.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 32k 170.3 16.8 69.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0e4ftp2f",
   "source_url": "https://omlx.ai/benchmarks/performance/0e4ftp2f",
   "tg_tokens": null,
   "tps": 16.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-kv66u88z-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.3 GB at this context; duplicate traces of the same session: 2k6ofxl2, hfzzor6h, u28zirxu, vrw9n3xu; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 439.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 1k 439.6 47.8 66.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kv66u88z",
   "source_url": "https://omlx.ai/benchmarks/performance/kv66u88z",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-kv66u88z-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.7 GB at this context; duplicate traces of the same session: 2k6ofxl2, hfzzor6h, u28zirxu, vrw9n3xu; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 463.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 4k 463.3 45.6 66.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kv66u88z",
   "source_url": "https://omlx.ai/benchmarks/performance/kv66u88z",
   "tg_tokens": null,
   "tps": 45.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-kv66u88z-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.0 GB at this context; duplicate traces of the same session: 2k6ofxl2, hfzzor6h, u28zirxu, vrw9n3xu; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 459.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 8k 459.9 41.3 67.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kv66u88z",
   "source_url": "https://omlx.ai/benchmarks/performance/kv66u88z",
   "tg_tokens": null,
   "tps": 41.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-kv66u88z-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.7 GB at this context; duplicate traces of the same session: 2k6ofxl2, hfzzor6h, u28zirxu, vrw9n3xu; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 444.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 16k 444.0 36.3 67.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kv66u88z",
   "source_url": "https://omlx.ai/benchmarks/performance/kv66u88z",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-kv66u88z-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 69.1 GB at this context; duplicate traces of the same session: 2k6ofxl2, hfzzor6h, u28zirxu, vrw9n3xu; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 409.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 32k 409.4 28.3 69.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kv66u88z",
   "source_url": "https://omlx.ai/benchmarks/performance/kv66u88z",
   "tg_tokens": null,
   "tps": 28.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-kv66u88z-ctx65536",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 71.9 GB at this context; duplicate traces of the same session: 2k6ofxl2, hfzzor6h, u28zirxu, vrw9n3xu; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 331.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 64k 331.7 19.0 71.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kv66u88z",
   "source_url": "https://omlx.ai/benchmarks/performance/kv66u88z",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-op4450vs-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.3 GB at this context; duplicate traces of the same session: bt6bvgoc, ejitl07v, jgotjhpb, n1ur8xge, nwfc7wtx; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 704.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 704.1 53.8 66.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/op4450vs",
   "source_url": "https://omlx.ai/benchmarks/performance/op4450vs",
   "tg_tokens": null,
   "tps": 53.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-op4450vs-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.7 GB at this context; duplicate traces of the same session: bt6bvgoc, ejitl07v, jgotjhpb, n1ur8xge, nwfc7wtx; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 729.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 729.1 51.2 66.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/op4450vs",
   "source_url": "https://omlx.ai/benchmarks/performance/op4450vs",
   "tg_tokens": null,
   "tps": 51.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-op4450vs-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.0 GB at this context; duplicate traces of the same session: bt6bvgoc, ejitl07v, jgotjhpb, n1ur8xge, nwfc7wtx; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 712.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 712.9 48.2 67.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/op4450vs",
   "source_url": "https://omlx.ai/benchmarks/performance/op4450vs",
   "tg_tokens": null,
   "tps": 48.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-op4450vs-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.7 GB at this context; duplicate traces of the same session: bt6bvgoc, ejitl07v, jgotjhpb, n1ur8xge, nwfc7wtx; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 670.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 670.3 41.2 67.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/op4450vs",
   "source_url": "https://omlx.ai/benchmarks/performance/op4450vs",
   "tg_tokens": null,
   "tps": 41.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-op4450vs-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 69.1 GB at this context; duplicate traces of the same session: bt6bvgoc, ejitl07v, jgotjhpb, n1ur8xge, nwfc7wtx; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 590.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 590.2 35.0 69.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/op4450vs",
   "source_url": "https://omlx.ai/benchmarks/performance/op4450vs",
   "tg_tokens": null,
   "tps": 35.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-op4450vs-ctx65536",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 71.9 GB at this context; duplicate traces of the same session: bt6bvgoc, ejitl07v, jgotjhpb, n1ur8xge, nwfc7wtx; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 466.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 466.9 26.0 71.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/op4450vs",
   "source_url": "https://omlx.ai/benchmarks/performance/op4450vs",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vlb9nyb0-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.3 GB at this context; duplicate traces of the same session: 43mgnpxt, cavk7p1r, lhzzunu0, vx4ljpey; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 897.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 897.8 52.3 66.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vlb9nyb0",
   "source_url": "https://omlx.ai/benchmarks/performance/vlb9nyb0",
   "tg_tokens": null,
   "tps": 52.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vlb9nyb0-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.7 GB at this context; duplicate traces of the same session: 43mgnpxt, cavk7p1r, lhzzunu0, vx4ljpey; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 935.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 935.9 47.9 66.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vlb9nyb0",
   "source_url": "https://omlx.ai/benchmarks/performance/vlb9nyb0",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vlb9nyb0-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.0 GB at this context; duplicate traces of the same session: 43mgnpxt, cavk7p1r, lhzzunu0, vx4ljpey; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 920.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 920.4 47.4 67.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vlb9nyb0",
   "source_url": "https://omlx.ai/benchmarks/performance/vlb9nyb0",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vlb9nyb0-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.7 GB at this context; duplicate traces of the same session: 43mgnpxt, cavk7p1r, lhzzunu0, vx4ljpey; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 862.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 862.3 43.1 67.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vlb9nyb0",
   "source_url": "https://omlx.ai/benchmarks/performance/vlb9nyb0",
   "tg_tokens": null,
   "tps": 43.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vlb9nyb0-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 69.1 GB at this context; duplicate traces of the same session: 43mgnpxt, cavk7p1r, lhzzunu0, vx4ljpey; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 746.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 746.1 36.4 69.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vlb9nyb0",
   "source_url": "https://omlx.ai/benchmarks/performance/vlb9nyb0",
   "tg_tokens": null,
   "tps": 36.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vlb9nyb0-ctx65536",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 71.9 GB at this context; duplicate traces of the same session: 43mgnpxt, cavk7p1r, lhzzunu0, vx4ljpey; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 570.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 570.2 27.7 71.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vlb9nyb0",
   "source_url": "https://omlx.ai/benchmarks/performance/vlb9nyb0",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-lm7yt1ag-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.3 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 904.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 904.5 52.5 66.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lm7yt1ag",
   "source_url": "https://omlx.ai/benchmarks/performance/lm7yt1ag",
   "tg_tokens": null,
   "tps": 52.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-lm7yt1ag-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 951.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 951.6 52.0 66.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lm7yt1ag",
   "source_url": "https://omlx.ai/benchmarks/performance/lm7yt1ag",
   "tg_tokens": null,
   "tps": 52.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-kbpdn60e-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 123.2 GB at this context; duplicate traces of the same session: 77dim65a, ho7w5nty, z1bl3y33; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 837.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 837.2 39.1 123.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kbpdn60e",
   "source_url": "https://omlx.ai/benchmarks/performance/kbpdn60e",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-kbpdn60e-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 123.5 GB at this context; duplicate traces of the same session: 77dim65a, ho7w5nty, z1bl3y33; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 895.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 895.1 37.8 123.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kbpdn60e",
   "source_url": "https://omlx.ai/benchmarks/performance/kbpdn60e",
   "tg_tokens": null,
   "tps": 37.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-kbpdn60e-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 123.9 GB at this context; duplicate traces of the same session: 77dim65a, ho7w5nty, z1bl3y33; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 876.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 876.6 35.7 123.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kbpdn60e",
   "source_url": "https://omlx.ai/benchmarks/performance/kbpdn60e",
   "tg_tokens": null,
   "tps": 35.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-kbpdn60e-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 124.5 GB at this context; duplicate traces of the same session: 77dim65a, ho7w5nty, z1bl3y33; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 826.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k 826.4 32.5 124.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kbpdn60e",
   "source_url": "https://omlx.ai/benchmarks/performance/kbpdn60e",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-kbpdn60e-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 125.9 GB at this context; duplicate traces of the same session: 77dim65a, ho7w5nty, z1bl3y33; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 717.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 32k 717.8 27.9 125.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kbpdn60e",
   "source_url": "https://omlx.ai/benchmarks/performance/kbpdn60e",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-ak0mp87u-ctx1024",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 5raey9l1, 5y2fk7fi, 8rwegtmb, i0ji024h, xz6ox1sf; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 226.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 1k 226.5 33.4 16.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ak0mp87u",
   "source_url": "https://omlx.ai/benchmarks/performance/ak0mp87u",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-ak0mp87u-ctx4096",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 5raey9l1, 5y2fk7fi, 8rwegtmb, i0ji024h, xz6ox1sf; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 228.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 4k 228.9 31.3 17.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ak0mp87u",
   "source_url": "https://omlx.ai/benchmarks/performance/ak0mp87u",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-ak0mp87u-ctx8192",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: 5raey9l1, 5y2fk7fi, 8rwegtmb, i0ji024h, xz6ox1sf; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 227.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 8k 227.0 29.3 17.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ak0mp87u",
   "source_url": "https://omlx.ai/benchmarks/performance/ak0mp87u",
   "tg_tokens": null,
   "tps": 29.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-ak0mp87u-ctx16384",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: 5raey9l1, 5y2fk7fi, 8rwegtmb, i0ji024h, xz6ox1sf; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 222.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 16k 222.3 26.6 18.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ak0mp87u",
   "source_url": "https://omlx.ai/benchmarks/performance/ak0mp87u",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-ak0mp87u-ctx32768",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 5raey9l1, 5y2fk7fi, 8rwegtmb, i0ji024h, xz6ox1sf; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 212.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 32k 212.9 21.9 20.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ak0mp87u",
   "source_url": "https://omlx.ai/benchmarks/performance/ak0mp87u",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-ak0mp87u-ctx65536",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 5raey9l1, 5y2fk7fi, 8rwegtmb, i0ji024h, xz6ox1sf; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 192.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 64k 192.2 16.8 23.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ak0mp87u",
   "source_url": "https://omlx.ai/benchmarks/performance/ak0mp87u",
   "tg_tokens": null,
   "tps": 16.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rf0le76b-ctx1024",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 8tvgq62t, ah2c9b47, iff2zzcf, jkbcbaj9, w3gm3t09; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 322.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 322.7 38.0 16.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rf0le76b",
   "source_url": "https://omlx.ai/benchmarks/performance/rf0le76b",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rf0le76b-ctx4096",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 8tvgq62t, ah2c9b47, iff2zzcf, jkbcbaj9, w3gm3t09; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 324.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 324.2 35.3 17.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rf0le76b",
   "source_url": "https://omlx.ai/benchmarks/performance/rf0le76b",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rf0le76b-ctx8192",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: 8tvgq62t, ah2c9b47, iff2zzcf, jkbcbaj9, w3gm3t09; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 319.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 319.7 33.6 17.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rf0le76b",
   "source_url": "https://omlx.ai/benchmarks/performance/rf0le76b",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rf0le76b-ctx16384",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: 8tvgq62t, ah2c9b47, iff2zzcf, jkbcbaj9, w3gm3t09; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 310.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 310.0 30.0 18.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rf0le76b",
   "source_url": "https://omlx.ai/benchmarks/performance/rf0le76b",
   "tg_tokens": null,
   "tps": 30.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rf0le76b-ctx32768",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 8tvgq62t, ah2c9b47, iff2zzcf, jkbcbaj9, w3gm3t09; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 291.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 291.3 24.9 20.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rf0le76b",
   "source_url": "https://omlx.ai/benchmarks/performance/rf0le76b",
   "tg_tokens": null,
   "tps": 24.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rf0le76b-ctx65536",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 8tvgq62t, ah2c9b47, iff2zzcf, jkbcbaj9, w3gm3t09; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 258.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 258.7 18.1 23.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rf0le76b",
   "source_url": "https://omlx.ai/benchmarks/performance/rf0le76b",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-754ncwmf-ctx1024",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 76f7vm8m, iugrfis1, yiokj60f; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 375.2,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 1k 375.2 26.7 22.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/754ncwmf",
   "source_url": "https://omlx.ai/benchmarks/performance/754ncwmf",
   "tg_tokens": null,
   "tps": 26.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-754ncwmf-ctx4096",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: 76f7vm8m, iugrfis1, yiokj60f; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 401.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 4k 401.1 25.4 22.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/754ncwmf",
   "source_url": "https://omlx.ai/benchmarks/performance/754ncwmf",
   "tg_tokens": null,
   "tps": 25.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-754ncwmf-ctx8192",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: 76f7vm8m, iugrfis1, yiokj60f; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 400.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 8k 400.8 24.4 22.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/754ncwmf",
   "source_url": "https://omlx.ai/benchmarks/performance/754ncwmf",
   "tg_tokens": null,
   "tps": 24.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-754ncwmf-ctx16384",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: 76f7vm8m, iugrfis1, yiokj60f; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 392.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 16k 392.6 22.4 23.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/754ncwmf",
   "source_url": "https://omlx.ai/benchmarks/performance/754ncwmf",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-754ncwmf-ctx32768",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: 76f7vm8m, iugrfis1, yiokj60f; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 370.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 32k 370.0 19.5 25.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/754ncwmf",
   "source_url": "https://omlx.ai/benchmarks/performance/754ncwmf",
   "tg_tokens": null,
   "tps": 19.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rwbkz9nr-ctx1024",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 210.4 GB at this context; duplicate traces of the same session: 4o3fa12n, 6xmkfdgf, k0jw9l2d, rdlvb7oi, vive2iyg; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 381.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 381.0 37.0 210.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rwbkz9nr",
   "source_url": "https://omlx.ai/benchmarks/performance/rwbkz9nr",
   "tg_tokens": null,
   "tps": 37.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rwbkz9nr-ctx4096",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 210.7 GB at this context; duplicate traces of the same session: 4o3fa12n, 6xmkfdgf, k0jw9l2d, rdlvb7oi, vive2iyg; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 404.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 404.5 35.3 210.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rwbkz9nr",
   "source_url": "https://omlx.ai/benchmarks/performance/rwbkz9nr",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rwbkz9nr-ctx8192",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 211.0 GB at this context; duplicate traces of the same session: 4o3fa12n, 6xmkfdgf, k0jw9l2d, rdlvb7oi, vive2iyg; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 400.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 400.7 31.3 211.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rwbkz9nr",
   "source_url": "https://omlx.ai/benchmarks/performance/rwbkz9nr",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rwbkz9nr-ctx16384",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 211.8 GB at this context; duplicate traces of the same session: 4o3fa12n, 6xmkfdgf, k0jw9l2d, rdlvb7oi, vive2iyg; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 384.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 384.0 24.8 211.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rwbkz9nr",
   "source_url": "https://omlx.ai/benchmarks/performance/rwbkz9nr",
   "tg_tokens": null,
   "tps": 24.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rwbkz9nr-ctx32768",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 213.2 GB at this context; duplicate traces of the same session: 4o3fa12n, 6xmkfdgf, k0jw9l2d, rdlvb7oi, vive2iyg; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 350.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 350.4 17.8 213.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rwbkz9nr",
   "source_url": "https://omlx.ai/benchmarks/performance/rwbkz9nr",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rwbkz9nr-ctx65536",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 216.2 GB at this context; duplicate traces of the same session: 4o3fa12n, 6xmkfdgf, k0jw9l2d, rdlvb7oi, vive2iyg; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 294.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 294.3 11.7 216.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rwbkz9nr",
   "source_url": "https://omlx.ai/benchmarks/performance/rwbkz9nr",
   "tg_tokens": null,
   "tps": 11.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0ogsu5db-ctx1024",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 302.7 GB at this context; duplicate traces of the same session: au03cy2k, q66jpzr6, rb7hx514; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 454.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 1k 454.0 28.5 302.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ogsu5db",
   "source_url": "https://omlx.ai/benchmarks/performance/0ogsu5db",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0ogsu5db-ctx4096",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 302.9 GB at this context; duplicate traces of the same session: au03cy2k, q66jpzr6, rb7hx514; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 487.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 4k 487.1 27.2 302.9 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ogsu5db",
   "source_url": "https://omlx.ai/benchmarks/performance/0ogsu5db",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0ogsu5db-ctx8192",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 303.3 GB at this context; duplicate traces of the same session: au03cy2k, q66jpzr6, rb7hx514; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 484.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 8k 484.9 26.0 303.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ogsu5db",
   "source_url": "https://omlx.ai/benchmarks/performance/0ogsu5db",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0ogsu5db-ctx16384",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 304.0 GB at this context; duplicate traces of the same session: au03cy2k, q66jpzr6, rb7hx514; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 462.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 16k 462.0 23.6 304.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ogsu5db",
   "source_url": "https://omlx.ai/benchmarks/performance/0ogsu5db",
   "tg_tokens": null,
   "tps": 23.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0ogsu5db-ctx32768",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 305.5 GB at this context; duplicate traces of the same session: au03cy2k, q66jpzr6, rb7hx514; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 422.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 32k 422.5 20.2 305.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ogsu5db",
   "source_url": "https://omlx.ai/benchmarks/performance/0ogsu5db",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-fsmx6den-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.7 GB at this context; duplicate traces of the same session: a6fmmdir, gnjkbfpd, gp94nvfp, t8kof1zr, xpnt88fb; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 775.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 1k 775.8 74.3 8.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fsmx6den",
   "source_url": "https://omlx.ai/benchmarks/performance/fsmx6den",
   "tg_tokens": null,
   "tps": 74.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-fsmx6den-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.9 GB at this context; duplicate traces of the same session: a6fmmdir, gnjkbfpd, gp94nvfp, t8kof1zr, xpnt88fb; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 783.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 4k 783.9 72.1 8.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fsmx6den",
   "source_url": "https://omlx.ai/benchmarks/performance/fsmx6den",
   "tg_tokens": null,
   "tps": 72.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-fsmx6den-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 9.2 GB at this context; duplicate traces of the same session: a6fmmdir, gnjkbfpd, gp94nvfp, t8kof1zr, xpnt88fb; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 776.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 8k 776.7 68.7 9.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fsmx6den",
   "source_url": "https://omlx.ai/benchmarks/performance/fsmx6den",
   "tg_tokens": null,
   "tps": 68.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-fsmx6den-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 9.6 GB at this context; duplicate traces of the same session: a6fmmdir, gnjkbfpd, gp94nvfp, t8kof1zr, xpnt88fb; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 758.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 16k 758.8 62.8 9.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fsmx6den",
   "source_url": "https://omlx.ai/benchmarks/performance/fsmx6den",
   "tg_tokens": null,
   "tps": 62.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-fsmx6den-ctx32768",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.6 GB at this context; duplicate traces of the same session: a6fmmdir, gnjkbfpd, gp94nvfp, t8kof1zr, xpnt88fb; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 724.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 32k 724.1 52.6 10.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fsmx6den",
   "source_url": "https://omlx.ai/benchmarks/performance/fsmx6den",
   "tg_tokens": null,
   "tps": 52.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-fsmx6den-ctx65536",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: a6fmmdir, gnjkbfpd, gp94nvfp, t8kof1zr, xpnt88fb; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 653.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 64k 653.8 39.8 12.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fsmx6den",
   "source_url": "https://omlx.ai/benchmarks/performance/fsmx6den",
   "tg_tokens": null,
   "tps": 39.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-7188vfes-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.6 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 218.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-9B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 6bit; 1k 218.0 15.7 8.6 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7188vfes",
   "source_url": "https://omlx.ai/benchmarks/performance/7188vfes",
   "tg_tokens": null,
   "tps": 15.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-7188vfes-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.8 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 218.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-9B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 6bit; 4k 218.9 15.3 8.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7188vfes",
   "source_url": "https://omlx.ai/benchmarks/performance/7188vfes",
   "tg_tokens": null,
   "tps": 15.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-sio1c77w-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 8.6 GB; TTFT 4715 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 217.2,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-9B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 6bit; 1k tokens 217.2 PP tok/s 15.8 TG tok/s 4715 TTFT (ms) 8.6 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sio1c77w",
   "source_url": "https://omlx.ai/benchmarks/performance/sio1c77w",
   "tg_tokens": null,
   "tps": 15.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-57bqk5df-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 8.6 GB; TTFT 4714 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 217.2,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-9B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 6bit; 1k tokens 217.2 PP tok/s 15.7 TG tok/s 4714 TTFT (ms) 8.6 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/57bqk5df",
   "source_url": "https://omlx.ai/benchmarks/performance/57bqk5df",
   "tg_tokens": null,
   "tps": 15.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-b9rhjswd-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 8.6 GB; TTFT 4706 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 217.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-9B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 6bit; 1k tokens 217.6 PP tok/s 15.7 TG tok/s 4706 TTFT (ms) 8.6 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b9rhjswd",
   "source_url": "https://omlx.ai/benchmarks/performance/b9rhjswd",
   "tg_tokens": null,
   "tps": 15.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-5fvml94a-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 193.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 16 GB \u00b7 4bit; 1k 193.0 33.5 5.9 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5fvml94a",
   "source_url": "https://omlx.ai/benchmarks/performance/5fvml94a",
   "tg_tokens": null,
   "tps": 33.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-5fvml94a-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 188.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 16 GB \u00b7 4bit; 4k 188.7 32.5 6.4 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5fvml94a",
   "source_url": "https://omlx.ai/benchmarks/performance/5fvml94a",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-rxbp96yc-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 225.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 225.0 19.0 5.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rxbp96yc",
   "source_url": "https://omlx.ai/benchmarks/performance/rxbp96yc",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-rxbp96yc-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 239.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 239.4 19.4 6.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rxbp96yc",
   "source_url": "https://omlx.ai/benchmarks/performance/rxbp96yc",
   "tg_tokens": null,
   "tps": 19.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-axjkl93n-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.2 GB at this context; duplicate traces of the same session: 3lfp52kg; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 169.7,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Qwen3.5-9B-mxfp4 on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 169.7 21.7 7.2 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/axjkl93n",
   "source_url": "https://omlx.ai/benchmarks/performance/axjkl93n",
   "tg_tokens": null,
   "tps": 21.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-axjkl93n-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.7 GB at this context; duplicate traces of the same session: 3lfp52kg; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 169.0,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Qwen3.5-9B-mxfp4 on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 169.0 20.7 7.7 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/axjkl93n",
   "source_url": "https://omlx.ai/benchmarks/performance/axjkl93n",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-ak90o81b-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.2 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 215.4,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Qwen3.5-9B-mxfp4 on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 215.4 26.1 7.2 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ak90o81b",
   "source_url": "https://omlx.ai/benchmarks/performance/ak90o81b",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-ak90o81b-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.7 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 214.5,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Qwen3.5-9B-mxfp4 on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 214.5 25.3 7.7 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ak90o81b",
   "source_url": "https://omlx.ai/benchmarks/performance/ak90o81b",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-3j3xuiru-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 7bx99kw6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 744.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 744.3 108.4 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3j3xuiru",
   "source_url": "https://omlx.ai/benchmarks/performance/3j3xuiru",
   "tg_tokens": null,
   "tps": 108.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-3j3xuiru-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 7bx99kw6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2064.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,064 105.4 20.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3j3xuiru",
   "source_url": "https://omlx.ai/benchmarks/performance/3j3xuiru",
   "tg_tokens": null,
   "tps": 105.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-uqxp3mcg-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.2 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 734.2,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 4k 734.2 33.1 32.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uqxp3mcg",
   "source_url": "https://omlx.ai/benchmarks/performance/uqxp3mcg",
   "tg_tokens": null,
   "tps": 33.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-uqxp3mcg-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.4 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 433.3,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 32k 433.3 46.8 34.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uqxp3mcg",
   "source_url": "https://omlx.ai/benchmarks/performance/uqxp3mcg",
   "tg_tokens": null,
   "tps": 46.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-7de9inuf-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.2 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 731.7,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 4k 731.7 54.5 32.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7de9inuf",
   "source_url": "https://omlx.ai/benchmarks/performance/7de9inuf",
   "tg_tokens": null,
   "tps": 54.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-7de9inuf-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.4 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 438.7,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 32k 438.7 47.3 34.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7de9inuf",
   "source_url": "https://omlx.ai/benchmarks/performance/7de9inuf",
   "tg_tokens": null,
   "tps": 47.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-cr9gkj7l-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 719.6,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 4k 719.6 55.4 29.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cr9gkj7l",
   "source_url": "https://omlx.ai/benchmarks/performance/cr9gkj7l",
   "tg_tokens": null,
   "tps": 55.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-cr9gkj7l-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 448.7,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 32k 448.7 46.1 32.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cr9gkj7l",
   "source_url": "https://omlx.ai/benchmarks/performance/cr9gkj7l",
   "tg_tokens": null,
   "tps": 46.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-w4fd4fuk-ctx1024",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 65k3qlf9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 237.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Abliterated-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 237.6 15.7 28.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w4fd4fuk",
   "source_url": "https://omlx.ai/benchmarks/performance/w4fd4fuk",
   "tg_tokens": null,
   "tps": 15.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-w4fd4fuk-ctx4096",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 32.2 GB at this context; duplicate traces of the same session: 65k3qlf9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 182.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Abliterated-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 182.1 13.9 32.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w4fd4fuk",
   "source_url": "https://omlx.ai/benchmarks/performance/w4fd4fuk",
   "tg_tokens": null,
   "tps": 13.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lan28zdm-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: p46idvkv, wufyjzmw, y79kne58; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 928.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 928.8 33.2 16.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lan28zdm",
   "source_url": "https://omlx.ai/benchmarks/performance/lan28zdm",
   "tg_tokens": null,
   "tps": 33.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lan28zdm-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: p46idvkv, wufyjzmw, y79kne58; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1008.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,008 32.8 17.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lan28zdm",
   "source_url": "https://omlx.ai/benchmarks/performance/lan28zdm",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lan28zdm-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: p46idvkv, wufyjzmw, y79kne58; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 878.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 878.4 32.0 17.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lan28zdm",
   "source_url": "https://omlx.ai/benchmarks/performance/lan28zdm",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lan28zdm-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: p46idvkv, wufyjzmw, y79kne58; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 797.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 797.7 30.8 17.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lan28zdm",
   "source_url": "https://omlx.ai/benchmarks/performance/lan28zdm",
   "tg_tokens": null,
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-h0jneyhw-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "community MTPLX build with MTP heads; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 277.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed-oMLX on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 1k 277.9 46.8 22.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h0jneyhw",
   "source_url": "https://omlx.ai/benchmarks/performance/h0jneyhw",
   "tg_tokens": null,
   "tps": 46.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-h0jneyhw-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "community MTPLX build with MTP heads; oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 286.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed-oMLX on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 4k 286.4 41.7 26.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h0jneyhw",
   "source_url": "https://omlx.ai/benchmarks/performance/h0jneyhw",
   "tg_tokens": null,
   "tps": 41.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-vxygr040-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "community MTPLX build with MTP heads; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 285.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed-oMLX on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 4k 285.3 42.0 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vxygr040",
   "source_url": "https://omlx.ai/benchmarks/performance/vxygr040",
   "tg_tokens": null,
   "tps": 42.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-vxygr040-ctx16384",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "community MTPLX build with MTP heads; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 268.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed-oMLX on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 16k 268.6 44.2 28.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vxygr040",
   "source_url": "https://omlx.ai/benchmarks/performance/vxygr040",
   "tg_tokens": null,
   "tps": 44.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-fgxc9kks-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "community MTPLX build with MTP heads; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 286.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed-oMLX on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 4k 286.9 39.9 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fgxc9kks",
   "source_url": "https://omlx.ai/benchmarks/performance/fgxc9kks",
   "tg_tokens": null,
   "tps": 39.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-fgxc9kks-ctx16384",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "community MTPLX build with MTP heads; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 273.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed-oMLX on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 16k 273.6 43.0 28.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fgxc9kks",
   "source_url": "https://omlx.ai/benchmarks/performance/fgxc9kks",
   "tg_tokens": null,
   "tps": 43.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8kar5htx-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: 3t9jllvp, 5h3m9gbn, bqy4j59f, hhxc5sho, j0dficgz, nme46ce0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 893.6,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 1k 893.6 38.2 13.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8kar5htx",
   "source_url": "https://omlx.ai/benchmarks/performance/8kar5htx",
   "tg_tokens": null,
   "tps": 38.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8kar5htx-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: 3t9jllvp, 5h3m9gbn, bqy4j59f, hhxc5sho, j0dficgz, nme46ce0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 935.7,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 4k 935.7 38.0 14.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8kar5htx",
   "source_url": "https://omlx.ai/benchmarks/performance/8kar5htx",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8kar5htx-ctx8192",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: 3t9jllvp, 5h3m9gbn, bqy4j59f, hhxc5sho, j0dficgz, nme46ce0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 852.6,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 8k 852.6 35.3 14.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8kar5htx",
   "source_url": "https://omlx.ai/benchmarks/performance/8kar5htx",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8kar5htx-ctx16384",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 15.3 GB at this context; duplicate traces of the same session: 3t9jllvp, 5h3m9gbn, bqy4j59f, hhxc5sho, j0dficgz, nme46ce0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 735.4,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 16k 735.4 32.5 15.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8kar5htx",
   "source_url": "https://omlx.ai/benchmarks/performance/8kar5htx",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8kar5htx-ctx32768",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: 3t9jllvp, 5h3m9gbn, bqy4j59f, hhxc5sho, j0dficgz, nme46ce0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 699.2,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 32k 699.2 32.1 16.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8kar5htx",
   "source_url": "https://omlx.ai/benchmarks/performance/8kar5htx",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8kar5htx-ctx65536",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: 3t9jllvp, 5h3m9gbn, bqy4j59f, hhxc5sho, j0dficgz, nme46ce0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 680.6,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 64k 680.6 27.9 18.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8kar5htx",
   "source_url": "https://omlx.ai/benchmarks/performance/8kar5htx",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8kar5htx-ctx131072",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 3t9jllvp, 5h3m9gbn, bqy4j59f, hhxc5sho, j0dficgz, nme46ce0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 549.4,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 128k 549.4 21.5 22.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8kar5htx",
   "source_url": "https://omlx.ai/benchmarks/performance/8kar5htx",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-5xf6vclp-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 5ejt4gud; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 454.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 454.2 18.2 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5xf6vclp",
   "source_url": "https://omlx.ai/benchmarks/performance/5xf6vclp",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-5xf6vclp-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 5ejt4gud; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 500.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 500.8 17.8 17.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5xf6vclp",
   "source_url": "https://omlx.ai/benchmarks/performance/5xf6vclp",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-hyn7bzvn-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 507.3,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 507.3 14.2 21.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hyn7bzvn",
   "source_url": "https://omlx.ai/benchmarks/performance/hyn7bzvn",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-hyn7bzvn-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 453.9,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 453.9 13.7 21.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hyn7bzvn",
   "source_url": "https://omlx.ai/benchmarks/performance/hyn7bzvn",
   "tg_tokens": null,
   "tps": 13.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-c31tq5pg-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: ev03nt9i; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 228.6,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 1k 228.6 29.0 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c31tq5pg",
   "source_url": "https://omlx.ai/benchmarks/performance/c31tq5pg",
   "tg_tokens": null,
   "tps": 29.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-c31tq5pg-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: ev03nt9i; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 245.4,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 4k 245.4 28.3 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c31tq5pg",
   "source_url": "https://omlx.ai/benchmarks/performance/c31tq5pg",
   "tg_tokens": null,
   "tps": 28.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-c31tq5pg-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: ev03nt9i; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 239.8,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 8k 239.8 27.7 20.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c31tq5pg",
   "source_url": "https://omlx.ai/benchmarks/performance/c31tq5pg",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-c31tq5pg-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: ev03nt9i; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 229.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 16k 229.0 26.0 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c31tq5pg",
   "source_url": "https://omlx.ai/benchmarks/performance/c31tq5pg",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-ewfym865-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 273.8,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 1k 273.8 42.2 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ewfym865",
   "source_url": "https://omlx.ai/benchmarks/performance/ewfym865",
   "tg_tokens": null,
   "tps": 42.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-ewfym865-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 291.4,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 4k 291.4 43.0 23.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ewfym865",
   "source_url": "https://omlx.ai/benchmarks/performance/ewfym865",
   "tg_tokens": null,
   "tps": 43.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-gf00iz2i-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 72.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 72.9 14.4 18.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gf00iz2i",
   "source_url": "https://omlx.ai/benchmarks/performance/gf00iz2i",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-gf00iz2i-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 77.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 77.2 15.5 22.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gf00iz2i",
   "source_url": "https://omlx.ai/benchmarks/performance/gf00iz2i",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-a7c36ecm-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 127.4 26.4 18.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a7c36ecm",
   "source_url": "https://omlx.ai/benchmarks/performance/a7c36ecm",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-a7c36ecm-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 130.3 28.4 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a7c36ecm",
   "source_url": "https://omlx.ai/benchmarks/performance/a7c36ecm",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-98oeeyrn-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: t10uju16; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1192.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 64k 1,192 86.5 23.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/98oeeyrn",
   "source_url": "https://omlx.ai/benchmarks/performance/98oeeyrn",
   "tg_tokens": null,
   "tps": 86.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-98oeeyrn-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.6 GB at this context; duplicate traces of the same session: t10uju16; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 936.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 128k 936.6 74.6 27.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/98oeeyrn",
   "source_url": "https://omlx.ai/benchmarks/performance/98oeeyrn",
   "tg_tokens": null,
   "tps": 74.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-xc8piny5-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.7 GB at this context; duplicate traces of the same session: d8hclmdc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 210.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 210.2 27.9 33.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xc8piny5",
   "source_url": "https://omlx.ai/benchmarks/performance/xc8piny5",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-xc8piny5-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: d8hclmdc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 214.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 214.3 25.8 37.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xc8piny5",
   "source_url": "https://omlx.ai/benchmarks/performance/xc8piny5",
   "tg_tokens": null,
   "tps": 25.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-xc8piny5-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.1 GB at this context; duplicate traces of the same session: d8hclmdc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 149.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 149.6 15.4 43.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xc8piny5",
   "source_url": "https://omlx.ai/benchmarks/performance/xc8piny5",
   "tg_tokens": null,
   "tps": 15.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ijsokdxs-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: lr6txaa0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 217.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 217.3 33.4 33.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ijsokdxs",
   "source_url": "https://omlx.ai/benchmarks/performance/ijsokdxs",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ijsokdxs-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: lr6txaa0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 220.0 25.7 37.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ijsokdxs",
   "source_url": "https://omlx.ai/benchmarks/performance/ijsokdxs",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ijsokdxs-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.0 GB at this context; duplicate traces of the same session: lr6txaa0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 142.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 142.9 18.2 43.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ijsokdxs",
   "source_url": "https://omlx.ai/benchmarks/performance/ijsokdxs",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rjvlrqxw-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: chonyuwu, ol8sfqde; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 327.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 327.3 23.1 28.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rjvlrqxw",
   "source_url": "https://omlx.ai/benchmarks/performance/rjvlrqxw",
   "tg_tokens": null,
   "tps": 23.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rjvlrqxw-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.2 GB at this context; duplicate traces of the same session: chonyuwu, ol8sfqde; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 337.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 337.7 22.8 32.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rjvlrqxw",
   "source_url": "https://omlx.ai/benchmarks/performance/rjvlrqxw",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rjvlrqxw-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.2 GB at this context; duplicate traces of the same session: chonyuwu, ol8sfqde; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 332.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 332.6 22.2 33.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rjvlrqxw",
   "source_url": "https://omlx.ai/benchmarks/performance/rjvlrqxw",
   "tg_tokens": null,
   "tps": 22.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rjvlrqxw-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.4 GB at this context; duplicate traces of the same session: chonyuwu, ol8sfqde; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 321.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 16k 321.9 21.7 32.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rjvlrqxw",
   "source_url": "https://omlx.ai/benchmarks/performance/rjvlrqxw",
   "tg_tokens": null,
   "tps": 21.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9ewa5glr-ctx65536",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.2 GB at this context; duplicate traces of the same session: bun5fio0, g8yms4n0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 854.9,
   "provenance": "community",
   "quant": "mxfp4+mxfp8",
   "quote": "Qwen3.8-Flash-Next-Uncensored-MLX-mixed-mxfp4_mxfp8_mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 854.9 32.2 103.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ewa5glr",
   "source_url": "https://omlx.ai/benchmarks/performance/9ewa5glr",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9ewa5glr-ctx131072",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.0 GB at this context; duplicate traces of the same session: bun5fio0, g8yms4n0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 818.2,
   "provenance": "community",
   "quant": "mxfp4+mxfp8",
   "quote": "Qwen3.8-Flash-Next-Uncensored-MLX-mixed-mxfp4_mxfp8_mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 818.2 37.9 110.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ewa5glr",
   "source_url": "https://omlx.ai/benchmarks/performance/9ewa5glr",
   "tg_tokens": null,
   "tps": 37.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9ewa5glr-ctx199680",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 120.0 GB at this context; duplicate traces of the same session: bun5fio0, g8yms4n0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 780.0,
   "provenance": "community",
   "quant": "mxfp4+mxfp8",
   "quote": "Qwen3.8-Flash-Next-Uncensored-MLX-mixed-mxfp4_mxfp8_mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 195k 780.0 37.4 120.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ewa5glr",
   "source_url": "https://omlx.ai/benchmarks/performance/9ewa5glr",
   "tg_tokens": null,
   "tps": 37.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ws73upph-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.1 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 650.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 650.7 87.7 99.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ws73upph",
   "source_url": "https://omlx.ai/benchmarks/performance/ws73upph",
   "tg_tokens": null,
   "tps": 87.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ws73upph-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.9 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 730.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 730.0 71.3 100.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ws73upph",
   "source_url": "https://omlx.ai/benchmarks/performance/ws73upph",
   "tg_tokens": null,
   "tps": 71.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-k5r5mli1-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: 8ei1yfsr, re3tiqd6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 525.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 525.3 56.3 71.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k5r5mli1",
   "source_url": "https://omlx.ai/benchmarks/performance/k5r5mli1",
   "tg_tokens": null,
   "tps": 56.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-k5r5mli1-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: 8ei1yfsr, re3tiqd6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 592.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 592.8 53.4 71.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k5r5mli1",
   "source_url": "https://omlx.ai/benchmarks/performance/k5r5mli1",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-k5r5mli1-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: 8ei1yfsr, re3tiqd6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 465.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 465.7 35.0 71.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k5r5mli1",
   "source_url": "https://omlx.ai/benchmarks/performance/k5r5mli1",
   "tg_tokens": null,
   "tps": 35.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-k3lg0q8m-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: c6fnc70f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 505.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 505.9 59.8 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k3lg0q8m",
   "source_url": "https://omlx.ai/benchmarks/performance/k3lg0q8m",
   "tg_tokens": null,
   "tps": 59.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-k3lg0q8m-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: c6fnc70f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 538.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 538.4 54.9 71.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k3lg0q8m",
   "source_url": "https://omlx.ai/benchmarks/performance/k3lg0q8m",
   "tg_tokens": null,
   "tps": 54.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-fumvnysw-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: ztljos0k; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1576.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,576 54.4 16.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fumvnysw",
   "source_url": "https://omlx.ai/benchmarks/performance/fumvnysw",
   "tg_tokens": null,
   "tps": 54.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-fumvnysw-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: ztljos0k; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1768.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 1,768 53.5 17.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fumvnysw",
   "source_url": "https://omlx.ai/benchmarks/performance/fumvnysw",
   "tg_tokens": null,
   "tps": 53.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-j49kw3yo-ctx1024",
   "model": "clef",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: z5kfpqtv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 428.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "clef-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 428.1 37.6 16.0 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j49kw3yo",
   "source_url": "https://omlx.ai/benchmarks/performance/j49kw3yo",
   "tg_tokens": null,
   "tps": 37.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-j49kw3yo-ctx4096",
   "model": "clef",
   "notes": "oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: z5kfpqtv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 472.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "clef-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 472.5 36.8 19.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j49kw3yo",
   "source_url": "https://omlx.ai/benchmarks/performance/j49kw3yo",
   "tg_tokens": null,
   "tps": 36.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-7vi3j31s-ctx1024",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 12.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1877.0,
   "provenance": "community",
   "quant": "UD-MLX-3bit",
   "quote": "gemma-4-26b-a4b-it-UD-MLX-3bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 3bit; 1k 1,877 67.9 12.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7vi3j31s",
   "source_url": "https://omlx.ai/benchmarks/performance/7vi3j31s",
   "tg_tokens": null,
   "tps": 67.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-7vi3j31s-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 12.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1958.0,
   "provenance": "community",
   "quant": "UD-MLX-3bit",
   "quote": "gemma-4-26b-a4b-it-UD-MLX-3bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 3bit; 4k 1,958 61.9 12.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7vi3j31s",
   "source_url": "https://omlx.ai/benchmarks/performance/7vi3j31s",
   "tg_tokens": null,
   "tps": 61.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-vfwc00k8-ctx1024",
   "model": "gpt-oss 120B",
   "notes": "oMLX community benchmark trace; peak memory 58.0 GB at this context; duplicate traces of the same session: 3gd322b3, 5oxq1gs1, nqo8pz28; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 187.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gpt-oss-120b on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 187.7 54.9 58.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vfwc00k8",
   "source_url": "https://omlx.ai/benchmarks/performance/vfwc00k8",
   "tg_tokens": null,
   "tps": 54.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-vfwc00k8-ctx4096",
   "model": "gpt-oss 120B",
   "notes": "oMLX community benchmark trace; peak memory 58.1 GB at this context; duplicate traces of the same session: 3gd322b3, 5oxq1gs1, nqo8pz28; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 182.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gpt-oss-120b on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 182.0 44.7 58.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vfwc00k8",
   "source_url": "https://omlx.ai/benchmarks/performance/vfwc00k8",
   "tg_tokens": null,
   "tps": 44.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-vfwc00k8-ctx8192",
   "model": "gpt-oss 120B",
   "notes": "oMLX community benchmark trace; peak memory 58.3 GB at this context; duplicate traces of the same session: 3gd322b3, 5oxq1gs1, nqo8pz28; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 183.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gpt-oss-120b on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 183.9 37.1 58.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vfwc00k8",
   "source_url": "https://omlx.ai/benchmarks/performance/vfwc00k8",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-vfwc00k8-ctx16384",
   "model": "gpt-oss 120B",
   "notes": "oMLX community benchmark trace; peak memory 58.6 GB at this context; duplicate traces of the same session: 3gd322b3, 5oxq1gs1, nqo8pz28; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 177.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gpt-oss-120b on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 177.1 22.7 58.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vfwc00k8",
   "source_url": "https://omlx.ai/benchmarks/performance/vfwc00k8",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-vfwc00k8-ctx32768",
   "model": "gpt-oss 120B",
   "notes": "oMLX community benchmark trace; peak memory 59.2 GB at this context; duplicate traces of the same session: 3gd322b3, 5oxq1gs1, nqo8pz28; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 165.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gpt-oss-120b on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 32k 165.3 14.6 59.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vfwc00k8",
   "source_url": "https://omlx.ai/benchmarks/performance/vfwc00k8",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ivuw3lz5-ctx1024",
   "model": "gpt-oss 120B",
   "notes": "oMLX community benchmark trace; peak memory 115.8 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 569.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gpt-oss-120b on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 569.3 60.7 115.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ivuw3lz5",
   "source_url": "https://omlx.ai/benchmarks/performance/ivuw3lz5",
   "tg_tokens": null,
   "tps": 60.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ivuw3lz5-ctx4096",
   "model": "gpt-oss 120B",
   "notes": "oMLX community benchmark trace; peak memory 115.9 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 631.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gpt-oss-120b on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 631.9 54.2 115.9 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ivuw3lz5",
   "source_url": "https://omlx.ai/benchmarks/performance/ivuw3lz5",
   "tg_tokens": null,
   "tps": 54.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-wz0dtthr-ctx8192",
   "model": "gpt-oss 120B",
   "notes": "MXFP4 experts with Q8 attention; oMLX community benchmark trace; peak memory 59.5 GB at this context; duplicate traces of the same session: sl2clrx6; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 365.9,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-120b-MXFP4-Q8 on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 8k 365.9 43.7 59.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wz0dtthr",
   "source_url": "https://omlx.ai/benchmarks/performance/wz0dtthr",
   "tg_tokens": null,
   "tps": 43.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-wz0dtthr-ctx16384",
   "model": "gpt-oss 120B",
   "notes": "MXFP4 experts with Q8 attention; oMLX community benchmark trace; peak memory 59.8 GB at this context; duplicate traces of the same session: sl2clrx6; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 361.6,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-120b-MXFP4-Q8 on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 16k 361.6 38.4 59.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wz0dtthr",
   "source_url": "https://omlx.ai/benchmarks/performance/wz0dtthr",
   "tg_tokens": null,
   "tps": 38.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-a3v6r61h-ctx1024",
   "model": "gpt-oss 120B",
   "notes": "MXFP4 experts with Q8 attention; oMLX community benchmark trace; peak memory 59.2 GB at this context; duplicate traces of the same session: oawwywpl; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 314.1,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-120b-MXFP4-Q8 on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 1k 314.1 67.7 59.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a3v6r61h",
   "source_url": "https://omlx.ai/benchmarks/performance/a3v6r61h",
   "tg_tokens": null,
   "tps": 67.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-a3v6r61h-ctx4096",
   "model": "gpt-oss 120B",
   "notes": "MXFP4 experts with Q8 attention; oMLX community benchmark trace; peak memory 59.3 GB at this context; duplicate traces of the same session: oawwywpl; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 363.7,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-120b-MXFP4-Q8 on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 4k 363.7 60.1 59.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a3v6r61h",
   "source_url": "https://omlx.ai/benchmarks/performance/a3v6r61h",
   "tg_tokens": null,
   "tps": 60.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-qxaj08j3-ctx32768",
   "model": "gpt-oss 120B",
   "notes": "MXFP4 experts with Q8 attention; oMLX community benchmark trace; peak memory 60.4 GB at this context; duplicate traces of the same session: rezzxtyy; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 337.1,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-120b-MXFP4-Q8 on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 32k 337.1 25.7 60.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qxaj08j3",
   "source_url": "https://omlx.ai/benchmarks/performance/qxaj08j3",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-qxaj08j3-ctx65536",
   "model": "gpt-oss 120B",
   "notes": "MXFP4 experts with Q8 attention; oMLX community benchmark trace; peak memory 61.6 GB at this context; duplicate traces of the same session: rezzxtyy; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 294.5,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-120b-MXFP4-Q8 on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 64k 294.5 15.4 61.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qxaj08j3",
   "source_url": "https://omlx.ai/benchmarks/performance/qxaj08j3",
   "tg_tokens": null,
   "tps": 15.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vtfd1nia-ctx1024",
   "model": "gpt-oss 120B",
   "notes": "MXFP4 experts with Q8 attention; oMLX community benchmark trace; peak memory 59.2 GB at this context; duplicate traces of the same session: 06kdatf0, i54z07a7, wjn65xp9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 613.2,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-120b-MXFP4-Q8 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 613.2 68.4 59.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vtfd1nia",
   "source_url": "https://omlx.ai/benchmarks/performance/vtfd1nia",
   "tg_tokens": null,
   "tps": 68.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vtfd1nia-ctx4096",
   "model": "gpt-oss 120B",
   "notes": "MXFP4 experts with Q8 attention; oMLX community benchmark trace; peak memory 59.3 GB at this context; duplicate traces of the same session: 06kdatf0, i54z07a7, wjn65xp9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 685.5,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-120b-MXFP4-Q8 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 685.5 64.1 59.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vtfd1nia",
   "source_url": "https://omlx.ai/benchmarks/performance/vtfd1nia",
   "tg_tokens": null,
   "tps": 64.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vtfd1nia-ctx8192",
   "model": "gpt-oss 120B",
   "notes": "MXFP4 experts with Q8 attention; oMLX community benchmark trace; peak memory 59.5 GB at this context; duplicate traces of the same session: 06kdatf0, i54z07a7, wjn65xp9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 687.8,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-120b-MXFP4-Q8 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 687.8 57.8 59.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vtfd1nia",
   "source_url": "https://omlx.ai/benchmarks/performance/vtfd1nia",
   "tg_tokens": null,
   "tps": 57.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vtfd1nia-ctx16384",
   "model": "gpt-oss 120B",
   "notes": "MXFP4 experts with Q8 attention; oMLX community benchmark trace; peak memory 59.8 GB at this context; duplicate traces of the same session: 06kdatf0, i54z07a7, wjn65xp9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 663.6,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-120b-MXFP4-Q8 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 663.6 47.5 59.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vtfd1nia",
   "source_url": "https://omlx.ai/benchmarks/performance/vtfd1nia",
   "tg_tokens": null,
   "tps": 47.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vtfd1nia-ctx32768",
   "model": "gpt-oss 120B",
   "notes": "MXFP4 experts with Q8 attention; oMLX community benchmark trace; peak memory 60.4 GB at this context; duplicate traces of the same session: 06kdatf0, i54z07a7, wjn65xp9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 600.7,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-120b-MXFP4-Q8 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 600.7 35.2 60.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vtfd1nia",
   "source_url": "https://omlx.ai/benchmarks/performance/vtfd1nia",
   "tg_tokens": null,
   "tps": 35.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-106yi11o-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: t9rdwkjo; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 562.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gpt-oss-20b-MLX-8bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 1k 562.4 51.9 21.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/106yi11o",
   "source_url": "https://omlx.ai/benchmarks/performance/106yi11o",
   "tg_tokens": null,
   "tps": 51.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-106yi11o-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: t9rdwkjo; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 626.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gpt-oss-20b-MLX-8bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 4k 626.9 49.7 21.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/106yi11o",
   "source_url": "https://omlx.ai/benchmarks/performance/106yi11o",
   "tg_tokens": null,
   "tps": 49.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-p9ujrg1n-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 471.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mlx-community--Qwen3.8-27B-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 471.0 9.7 29.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p9ujrg1n",
   "source_url": "https://omlx.ai/benchmarks/performance/p9ujrg1n",
   "tg_tokens": null,
   "tps": 9.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-p9ujrg1n-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 435.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mlx-community--Qwen3.8-27B-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 16k 435.6 9.3 30.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p9ujrg1n",
   "source_url": "https://omlx.ai/benchmarks/performance/p9ujrg1n",
   "tg_tokens": null,
   "tps": 9.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-pqi2kxlz-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 3hkr2x8v, ye45yn0j; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 105.9 15.6 18.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pqi2kxlz",
   "source_url": "https://omlx.ai/benchmarks/performance/pqi2kxlz",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-pqi2kxlz-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 3hkr2x8v, ye45yn0j; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 151.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 151.3 15.0 19.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pqi2kxlz",
   "source_url": "https://omlx.ai/benchmarks/performance/pqi2kxlz",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-pqi2kxlz-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 3hkr2x8v, ye45yn0j; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 115.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 115.0 14.5 19.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pqi2kxlz",
   "source_url": "https://omlx.ai/benchmarks/performance/pqi2kxlz",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-pqi2kxlz-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 3hkr2x8v, ye45yn0j; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 96.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 96.1 12.8 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pqi2kxlz",
   "source_url": "https://omlx.ai/benchmarks/performance/pqi2kxlz",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-w50542pa-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: c510mcpw, thcn3cyp, u28zoca4; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 106.0 15.6 18.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w50542pa",
   "source_url": "https://omlx.ai/benchmarks/performance/w50542pa",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-w50542pa-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: c510mcpw, thcn3cyp, u28zoca4; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 151.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 151.0 15.0 19.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w50542pa",
   "source_url": "https://omlx.ai/benchmarks/performance/w50542pa",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-w50542pa-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: c510mcpw, thcn3cyp, u28zoca4; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 151.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 151.6 14.5 19.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w50542pa",
   "source_url": "https://omlx.ai/benchmarks/performance/w50542pa",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-w50542pa-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: c510mcpw, thcn3cyp, u28zoca4; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 131.6 12.8 21.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w50542pa",
   "source_url": "https://omlx.ai/benchmarks/performance/w50542pa",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8s5lyopb-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "pyros-vault MTP build; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: 6yjomi11, agmyufp1, b1sluf0m, hj41ahwa, r2t9yutl, ziaa5bwx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 773.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "pyros-vault:Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 773.5 34.2 33.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8s5lyopb",
   "source_url": "https://omlx.ai/benchmarks/performance/8s5lyopb",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8s5lyopb-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "pyros-vault MTP build; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: 6yjomi11, agmyufp1, b1sluf0m, hj41ahwa, r2t9yutl, ziaa5bwx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 850.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "pyros-vault:Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 850.0 39.2 34.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8s5lyopb",
   "source_url": "https://omlx.ai/benchmarks/performance/8s5lyopb",
   "tg_tokens": null,
   "tps": 39.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8s5lyopb-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "pyros-vault MTP build; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: 6yjomi11, agmyufp1, b1sluf0m, hj41ahwa, r2t9yutl, ziaa5bwx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 823.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "pyros-vault:Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 823.6 50.2 35.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8s5lyopb",
   "source_url": "https://omlx.ai/benchmarks/performance/8s5lyopb",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8s5lyopb-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "pyros-vault MTP build; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: 6yjomi11, agmyufp1, b1sluf0m, hj41ahwa, r2t9yutl, ziaa5bwx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 752.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "pyros-vault:Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 752.8 53.2 35.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8s5lyopb",
   "source_url": "https://omlx.ai/benchmarks/performance/8s5lyopb",
   "tg_tokens": null,
   "tps": 53.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8s5lyopb-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "pyros-vault MTP build; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: 6yjomi11, agmyufp1, b1sluf0m, hj41ahwa, r2t9yutl, ziaa5bwx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 709.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "pyros-vault:Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 709.2 55.2 36.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8s5lyopb",
   "source_url": "https://omlx.ai/benchmarks/performance/8s5lyopb",
   "tg_tokens": null,
   "tps": 55.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8s5lyopb-ctx65536",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "pyros-vault MTP build; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: 6yjomi11, agmyufp1, b1sluf0m, hj41ahwa, r2t9yutl, ziaa5bwx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 617.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "pyros-vault:Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 617.3 41.5 39.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8s5lyopb",
   "source_url": "https://omlx.ai/benchmarks/performance/8s5lyopb",
   "tg_tokens": null,
   "tps": 41.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8s5lyopb-ctx131072",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "pyros-vault MTP build; oMLX community benchmark trace; peak memory 83.2 GB at this context; duplicate traces of the same session: 6yjomi11, agmyufp1, b1sluf0m, hj41ahwa, r2t9yutl, ziaa5bwx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 329.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "pyros-vault:Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 329.1 22.7 83.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8s5lyopb",
   "source_url": "https://omlx.ai/benchmarks/performance/8s5lyopb",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-dig8n87z-ctx1024",
   "model": "GLM-5",
   "notes": "oMLX community benchmark trace; peak memory 391.8 GB at this context; duplicate traces of the same session: 314vh14k; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 186.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 186.0 16.6 391.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dig8n87z",
   "source_url": "https://omlx.ai/benchmarks/performance/dig8n87z",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-dig8n87z-ctx4096",
   "model": "GLM-5",
   "notes": "oMLX community benchmark trace; peak memory 394.1 GB at this context; duplicate traces of the same session: 314vh14k; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 180.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 180.1 13.8 394.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dig8n87z",
   "source_url": "https://omlx.ai/benchmarks/performance/dig8n87z",
   "tg_tokens": null,
   "tps": 13.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-dig8n87z-ctx8192",
   "model": "GLM-5",
   "notes": "oMLX community benchmark trace; peak memory 396.7 GB at this context; duplicate traces of the same session: 314vh14k; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 154.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 154.2 13.1 396.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dig8n87z",
   "source_url": "https://omlx.ai/benchmarks/performance/dig8n87z",
   "tg_tokens": null,
   "tps": 13.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-dig8n87z-ctx16384",
   "model": "GLM-5",
   "notes": "oMLX community benchmark trace; peak memory 402.7 GB at this context; duplicate traces of the same session: 314vh14k; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 117.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 117.3 12.0 402.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dig8n87z",
   "source_url": "https://omlx.ai/benchmarks/performance/dig8n87z",
   "tg_tokens": null,
   "tps": 12.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-dig8n87z-ctx32768",
   "model": "GLM-5",
   "notes": "oMLX community benchmark trace; peak memory 415.4 GB at this context; duplicate traces of the same session: 314vh14k; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 77.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 77.3 10.4 415.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dig8n87z",
   "source_url": "https://omlx.ai/benchmarks/performance/dig8n87z",
   "tg_tokens": null,
   "tps": 10.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9xf0mntr-ctx1024",
   "model": "GLM-5",
   "notes": "oMLX community benchmark trace; peak memory 420.0 GB at this context; duplicate traces of the same session: 3ctrecp9, teqbm0rb, ugvdbq3h, xj3z6y6p; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 188.7,
   "provenance": "community",
   "quant": "4.8bit",
   "quote": "GLM-5-4.8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 188.7 16.6 420.0 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9xf0mntr",
   "source_url": "https://omlx.ai/benchmarks/performance/9xf0mntr",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9xf0mntr-ctx4096",
   "model": "GLM-5",
   "notes": "oMLX community benchmark trace; peak memory 422.3 GB at this context; duplicate traces of the same session: 3ctrecp9, teqbm0rb, ugvdbq3h, xj3z6y6p; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 182.6,
   "provenance": "community",
   "quant": "4.8bit",
   "quote": "GLM-5-4.8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 182.6 13.7 422.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9xf0mntr",
   "source_url": "https://omlx.ai/benchmarks/performance/9xf0mntr",
   "tg_tokens": null,
   "tps": 13.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9xf0mntr-ctx8192",
   "model": "GLM-5",
   "notes": "oMLX community benchmark trace; peak memory 424.9 GB at this context; duplicate traces of the same session: 3ctrecp9, teqbm0rb, ugvdbq3h, xj3z6y6p; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 156.3,
   "provenance": "community",
   "quant": "4.8bit",
   "quote": "GLM-5-4.8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 156.3 13.5 424.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9xf0mntr",
   "source_url": "https://omlx.ai/benchmarks/performance/9xf0mntr",
   "tg_tokens": null,
   "tps": 13.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9xf0mntr-ctx16384",
   "model": "GLM-5",
   "notes": "oMLX community benchmark trace; peak memory 430.9 GB at this context; duplicate traces of the same session: 3ctrecp9, teqbm0rb, ugvdbq3h, xj3z6y6p; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 119.8,
   "provenance": "community",
   "quant": "4.8bit",
   "quote": "GLM-5-4.8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 119.8 13.0 430.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9xf0mntr",
   "source_url": "https://omlx.ai/benchmarks/performance/9xf0mntr",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9xf0mntr-ctx32768",
   "model": "GLM-5",
   "notes": "oMLX community benchmark trace; peak memory 443.6 GB at this context; duplicate traces of the same session: 3ctrecp9, teqbm0rb, ugvdbq3h, xj3z6y6p; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 73.2,
   "provenance": "community",
   "quant": "4.8bit",
   "quote": "GLM-5-4.8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 73.2 12.3 443.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9xf0mntr",
   "source_url": "https://omlx.ai/benchmarks/performance/9xf0mntr",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-fqqdyr66-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.1 GB at this context; duplicate traces of the same session: t9shfw5c; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2283.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,283 127.2 177.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fqqdyr66",
   "source_url": "https://omlx.ai/benchmarks/performance/fqqdyr66",
   "tg_tokens": null,
   "tps": 127.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-fqqdyr66-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.6 GB at this context; duplicate traces of the same session: t9shfw5c; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2252.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,252 88.0 177.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fqqdyr66",
   "source_url": "https://omlx.ai/benchmarks/performance/fqqdyr66",
   "tg_tokens": null,
   "tps": 88.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-fqqdyr66-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.8 GB at this context; duplicate traces of the same session: t9shfw5c; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2202.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,202 140.2 178.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fqqdyr66",
   "source_url": "https://omlx.ai/benchmarks/performance/fqqdyr66",
   "tg_tokens": null,
   "tps": 140.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-9sc95md1-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.1 GB at this context; duplicate traces of the same session: goj6q9vt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2221.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,221 100.0 178.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9sc95md1",
   "source_url": "https://omlx.ai/benchmarks/performance/9sc95md1",
   "tg_tokens": null,
   "tps": 100.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-9sc95md1-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.6 GB at this context; duplicate traces of the same session: goj6q9vt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2198.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,198 74.4 178.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9sc95md1",
   "source_url": "https://omlx.ai/benchmarks/performance/9sc95md1",
   "tg_tokens": null,
   "tps": 74.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-9sc95md1-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 179.9 GB at this context; duplicate traces of the same session: goj6q9vt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2155.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,155 126.4 179.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9sc95md1",
   "source_url": "https://omlx.ai/benchmarks/performance/9sc95md1",
   "tg_tokens": null,
   "tps": 126.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-v2rhfwv0-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.1 GB at this context; duplicate traces of the same session: rpi89413; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2225.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,225 108.7 178.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v2rhfwv0",
   "source_url": "https://omlx.ai/benchmarks/performance/v2rhfwv0",
   "tg_tokens": null,
   "tps": 108.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-v2rhfwv0-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.6 GB at this context; duplicate traces of the same session: rpi89413; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2199.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,199 80.0 178.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v2rhfwv0",
   "source_url": "https://omlx.ai/benchmarks/performance/v2rhfwv0",
   "tg_tokens": null,
   "tps": 80.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-v2rhfwv0-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 179.9 GB at this context; duplicate traces of the same session: rpi89413; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2156.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,156 113.8 179.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v2rhfwv0",
   "source_url": "https://omlx.ai/benchmarks/performance/v2rhfwv0",
   "tg_tokens": null,
   "tps": 113.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wwnhcy1a-ctx65536",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 4.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2901.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM5-2B-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,901 87.5 4.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wwnhcy1a",
   "source_url": "https://omlx.ai/benchmarks/performance/wwnhcy1a",
   "tg_tokens": null,
   "tps": 87.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wwnhcy1a-ctx199680",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 9.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 991.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM5-2B-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 991.2 32.5 9.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wwnhcy1a",
   "source_url": "https://omlx.ai/benchmarks/performance/wwnhcy1a",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-jnc9tgs9-ctx1024",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 94.1 GB at this context; duplicate traces of the same session: 2kpxm1c1, s6lyqz2e; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 299.7,
   "provenance": "community",
   "quant": "3bit",
   "quote": "MiniMax-M2.5 on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 3bit; 1k 299.7 42.1 94.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jnc9tgs9",
   "source_url": "https://omlx.ai/benchmarks/performance/jnc9tgs9",
   "tg_tokens": null,
   "tps": 42.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-jnc9tgs9-ctx4096",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 95.1 GB at this context; duplicate traces of the same session: 2kpxm1c1, s6lyqz2e; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 337.6,
   "provenance": "community",
   "quant": "3bit",
   "quote": "MiniMax-M2.5 on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 3bit; 4k 337.6 35.7 95.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jnc9tgs9",
   "source_url": "https://omlx.ai/benchmarks/performance/jnc9tgs9",
   "tg_tokens": null,
   "tps": 35.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-jnc9tgs9-ctx8192",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 96.0 GB at this context; duplicate traces of the same session: 2kpxm1c1, s6lyqz2e; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 320.5,
   "provenance": "community",
   "quant": "3bit",
   "quote": "MiniMax-M2.5 on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 3bit; 8k 320.5 29.5 96.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jnc9tgs9",
   "source_url": "https://omlx.ai/benchmarks/performance/jnc9tgs9",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-jnc9tgs9-ctx16384",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 98.0 GB at this context; duplicate traces of the same session: 2kpxm1c1, s6lyqz2e; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 282.6,
   "provenance": "community",
   "quant": "3bit",
   "quote": "MiniMax-M2.5 on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 3bit; 16k 282.6 22.4 98.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jnc9tgs9",
   "source_url": "https://omlx.ai/benchmarks/performance/jnc9tgs9",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-wjvi0q0a-ctx1024",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 174.0 GB at this context; duplicate traces of the same session: 2we8lcjq, bzeg02vo, e48q54i5, xbfycdsw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 496.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 1k 496.7 40.0 174.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wjvi0q0a",
   "source_url": "https://omlx.ai/benchmarks/performance/wjvi0q0a",
   "tg_tokens": null,
   "tps": 40.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-wjvi0q0a-ctx4096",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 175.0 GB at this context; duplicate traces of the same session: 2we8lcjq, bzeg02vo, e48q54i5, xbfycdsw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 554.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 4k 554.8 33.9 175.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wjvi0q0a",
   "source_url": "https://omlx.ai/benchmarks/performance/wjvi0q0a",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-wjvi0q0a-ctx8192",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 175.9 GB at this context; duplicate traces of the same session: 2we8lcjq, bzeg02vo, e48q54i5, xbfycdsw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 514.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 8k 514.0 28.9 175.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wjvi0q0a",
   "source_url": "https://omlx.ai/benchmarks/performance/wjvi0q0a",
   "tg_tokens": null,
   "tps": 28.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-wjvi0q0a-ctx16384",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 177.8 GB at this context; duplicate traces of the same session: 2we8lcjq, bzeg02vo, e48q54i5, xbfycdsw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 436.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 16k 436.3 23.5 177.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wjvi0q0a",
   "source_url": "https://omlx.ai/benchmarks/performance/wjvi0q0a",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-wjvi0q0a-ctx32768",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 181.7 GB at this context; duplicate traces of the same session: 2we8lcjq, bzeg02vo, e48q54i5, xbfycdsw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 330.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 32k 330.8 16.1 181.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wjvi0q0a",
   "source_url": "https://omlx.ai/benchmarks/performance/wjvi0q0a",
   "tg_tokens": null,
   "tps": 16.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-wjvi0q0a-ctx65536",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 189.6 GB at this context; duplicate traces of the same session: 2we8lcjq, bzeg02vo, e48q54i5, xbfycdsw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 221.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 64k 221.1 9.6 189.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wjvi0q0a",
   "source_url": "https://omlx.ai/benchmarks/performance/wjvi0q0a",
   "tg_tokens": null,
   "tps": 9.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rf82ex1p-ctx1024",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 227.2 GB at this context; duplicate traces of the same session: jye2fo4m, s15twedr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 600.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 600.4 34.2 227.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rf82ex1p",
   "source_url": "https://omlx.ai/benchmarks/performance/rf82ex1p",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rf82ex1p-ctx4096",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 228.2 GB at this context; duplicate traces of the same session: jye2fo4m, s15twedr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 707.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 707.9 30.1 228.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rf82ex1p",
   "source_url": "https://omlx.ai/benchmarks/performance/rf82ex1p",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rf82ex1p-ctx8192",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 229.1 GB at this context; duplicate traces of the same session: jye2fo4m, s15twedr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 663.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 663.9 25.9 229.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rf82ex1p",
   "source_url": "https://omlx.ai/benchmarks/performance/rf82ex1p",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rf82ex1p-ctx16384",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 231.1 GB at this context; duplicate traces of the same session: jye2fo4m, s15twedr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 564.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k 564.2 21.1 231.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rf82ex1p",
   "source_url": "https://omlx.ai/benchmarks/performance/rf82ex1p",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rf82ex1p-ctx32768",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 235.0 GB at this context; duplicate traces of the same session: jye2fo4m, s15twedr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 427.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 32k 427.8 14.6 235.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rf82ex1p",
   "source_url": "https://omlx.ai/benchmarks/performance/rf82ex1p",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ilknboft-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: jg59xbr1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2066.0,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Ornith-1.5-35B-A3B-mlx-nvfp4 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 2,066 115.1 22.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ilknboft",
   "source_url": "https://omlx.ai/benchmarks/performance/ilknboft",
   "tg_tokens": null,
   "tps": 115.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ilknboft-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: jg59xbr1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1698.0,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Ornith-1.5-35B-A3B-mlx-nvfp4 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,698 105.1 23.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ilknboft",
   "source_url": "https://omlx.ai/benchmarks/performance/ilknboft",
   "tg_tokens": null,
   "tps": 105.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ilknboft-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: jg59xbr1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 934.2,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Ornith-1.5-35B-A3B-mlx-nvfp4 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 934.2 75.9 24.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ilknboft",
   "source_url": "https://omlx.ai/benchmarks/performance/ilknboft",
   "tg_tokens": null,
   "tps": 75.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ilknboft-ctx199680",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: jg59xbr1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 418.0,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Ornith-1.5-35B-A3B-mlx-nvfp4 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 195k 418.0 34.0 28.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ilknboft",
   "source_url": "https://omlx.ai/benchmarks/performance/ilknboft",
   "tg_tokens": null,
   "tps": 34.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-z62bghg3-ctx1024",
   "model": "Qwen2.5-Coder-14B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 8.5 GB at this context; duplicate traces of the same session: 9d8qdwu5; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 271.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-14B-Instruct on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 271.2 34.4 8.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z62bghg3",
   "source_url": "https://omlx.ai/benchmarks/performance/z62bghg3",
   "tg_tokens": null,
   "tps": 34.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-z62bghg3-ctx4096",
   "model": "Qwen2.5-Coder-14B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 9.2 GB at this context; duplicate traces of the same session: 9d8qdwu5; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 262.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-14B-Instruct on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 262.9 30.1 9.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z62bghg3",
   "source_url": "https://omlx.ai/benchmarks/performance/z62bghg3",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-8wbn5uhu-ctx1024",
   "model": "Qwen2.5-Coder-14B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 15.3 GB at this context; duplicate traces of the same session: 4prten73; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 270.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen2.5-Coder-14B-Instruct on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 1k 270.8 20.4 15.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8wbn5uhu",
   "source_url": "https://omlx.ai/benchmarks/performance/8wbn5uhu",
   "tg_tokens": null,
   "tps": 20.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-8wbn5uhu-ctx4096",
   "model": "Qwen2.5-Coder-14B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 4prten73; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 265.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen2.5-Coder-14B-Instruct on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k 265.9 18.8 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8wbn5uhu",
   "source_url": "https://omlx.ai/benchmarks/performance/8wbn5uhu",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ck4tk51q-ctx1024",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: zmqaz63e; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 462.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 462.0 58.1 16.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ck4tk51q",
   "source_url": "https://omlx.ai/benchmarks/performance/ck4tk51q",
   "tg_tokens": null,
   "tps": 58.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ck4tk51q-ctx4096",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: zmqaz63e; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 460.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 460.3 46.5 17.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ck4tk51q",
   "source_url": "https://omlx.ai/benchmarks/performance/ck4tk51q",
   "tg_tokens": null,
   "tps": 46.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-h2w5s6qk-ctx1024",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: 6v4q8j4y; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 433.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 1k 433.5 48.4 23.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h2w5s6qk",
   "source_url": "https://omlx.ai/benchmarks/performance/h2w5s6qk",
   "tg_tokens": null,
   "tps": 48.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-h2w5s6qk-ctx4096",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: 6v4q8j4y; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 443.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 4k 443.5 39.9 24.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h2w5s6qk",
   "source_url": "https://omlx.ai/benchmarks/performance/h2w5s6qk",
   "tg_tokens": null,
   "tps": 39.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-x7m6v6ru-ctx1024",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 30.8 GB at this context; duplicate traces of the same session: ltejhkkd; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 444.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 1k 444.5 44.7 30.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x7m6v6ru",
   "source_url": "https://omlx.ai/benchmarks/performance/x7m6v6ru",
   "tg_tokens": null,
   "tps": 44.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-x7m6v6ru-ctx4096",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 31.3 GB at this context; duplicate traces of the same session: ltejhkkd; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 462.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k 462.4 37.1 31.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x7m6v6ru",
   "source_url": "https://omlx.ai/benchmarks/performance/x7m6v6ru",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-4x2s7cj1-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.0 GB at this context; duplicate traces of the same session: 12jw4znp; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 400.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 1k 400.0 60.0 43.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4x2s7cj1",
   "source_url": "https://omlx.ai/benchmarks/performance/4x2s7cj1",
   "tg_tokens": null,
   "tps": 60.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-4x2s7cj1-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.2 GB at this context; duplicate traces of the same session: 12jw4znp; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 442.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 442.0 56.9 43.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4x2s7cj1",
   "source_url": "https://omlx.ai/benchmarks/performance/4x2s7cj1",
   "tg_tokens": null,
   "tps": 56.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-p4xs5kej-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.0 GB at this context; duplicate traces of the same session: ee6x45d0; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 402.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 1k 402.4 59.7 43.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p4xs5kej",
   "source_url": "https://omlx.ai/benchmarks/performance/p4xs5kej",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-p4xs5kej-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.2 GB at this context; duplicate traces of the same session: ee6x45d0; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 441.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 441.6 56.1 43.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p4xs5kej",
   "source_url": "https://omlx.ai/benchmarks/performance/p4xs5kej",
   "tg_tokens": null,
   "tps": 56.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ejsmx46o-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 61.6 GB at this context; duplicate traces of the same session: bg1g56a9, ipzlht96; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 464.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 1k 464.3 53.4 61.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ejsmx46o",
   "source_url": "https://omlx.ai/benchmarks/performance/ejsmx46o",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ejsmx46o-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 61.7 GB at this context; duplicate traces of the same session: bg1g56a9, ipzlht96; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 443.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 4k 443.4 50.2 61.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ejsmx46o",
   "source_url": "https://omlx.ai/benchmarks/performance/ejsmx46o",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ejsmx46o-ctx8192",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 62.0 GB at this context; duplicate traces of the same session: bg1g56a9, ipzlht96; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 427.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 8k 427.1 46.1 62.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ejsmx46o",
   "source_url": "https://omlx.ai/benchmarks/performance/ejsmx46o",
   "tg_tokens": null,
   "tps": 46.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ejsmx46o-ctx16384",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 62.4 GB at this context; duplicate traces of the same session: bg1g56a9, ipzlht96; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 412.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 16k 412.9 38.1 62.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ejsmx46o",
   "source_url": "https://omlx.ai/benchmarks/performance/ejsmx46o",
   "tg_tokens": null,
   "tps": 38.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ejsmx46o-ctx32768",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 63.3 GB at this context; duplicate traces of the same session: bg1g56a9, ipzlht96; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 379.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 32k 379.4 25.5 63.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ejsmx46o",
   "source_url": "https://omlx.ai/benchmarks/performance/ejsmx46o",
   "tg_tokens": null,
   "tps": 25.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-p9nbpjyd-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 61.6 GB at this context; duplicate traces of the same session: bgwyirz3, l7j9n9kx, n6v3m0w2; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 751.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 1k 751.3 57.9 61.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p9nbpjyd",
   "source_url": "https://omlx.ai/benchmarks/performance/p9nbpjyd",
   "tg_tokens": null,
   "tps": 57.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-p9nbpjyd-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 61.7 GB at this context; duplicate traces of the same session: bgwyirz3, l7j9n9kx, n6v3m0w2; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 966.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 4k 966.4 56.2 61.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p9nbpjyd",
   "source_url": "https://omlx.ai/benchmarks/performance/p9nbpjyd",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-p9nbpjyd-ctx8192",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 62.0 GB at this context; duplicate traces of the same session: bgwyirz3, l7j9n9kx, n6v3m0w2; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 986.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 8k 986.0 53.4 62.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p9nbpjyd",
   "source_url": "https://omlx.ai/benchmarks/performance/p9nbpjyd",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-p9nbpjyd-ctx16384",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 62.4 GB at this context; duplicate traces of the same session: bgwyirz3, l7j9n9kx, n6v3m0w2; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 956.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 16k 956.5 47.0 62.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p9nbpjyd",
   "source_url": "https://omlx.ai/benchmarks/performance/p9nbpjyd",
   "tg_tokens": null,
   "tps": 47.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-p9nbpjyd-ctx32768",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 63.3 GB at this context; duplicate traces of the same session: bgwyirz3, l7j9n9kx, n6v3m0w2; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 881.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 32k 881.3 39.1 63.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p9nbpjyd",
   "source_url": "https://omlx.ai/benchmarks/performance/p9nbpjyd",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-p9nbpjyd-ctx65536",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 65.1 GB at this context; duplicate traces of the same session: bgwyirz3, l7j9n9kx, n6v3m0w2; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 731.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 64k 731.4 28.9 65.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p9nbpjyd",
   "source_url": "https://omlx.ai/benchmarks/performance/p9nbpjyd",
   "tg_tokens": null,
   "tps": 28.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-dklk8tmi-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.0 GB at this context; duplicate traces of the same session: fdttset4, rihr0r0i, scn0knor; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1298.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 1,298 73.9 43.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dklk8tmi",
   "source_url": "https://omlx.ai/benchmarks/performance/dklk8tmi",
   "tg_tokens": null,
   "tps": 73.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-dklk8tmi-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.2 GB at this context; duplicate traces of the same session: fdttset4, rihr0r0i, scn0knor; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1558.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 1,558 71.0 43.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dklk8tmi",
   "source_url": "https://omlx.ai/benchmarks/performance/dklk8tmi",
   "tg_tokens": null,
   "tps": 71.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-dklk8tmi-ctx8192",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.4 GB at this context; duplicate traces of the same session: fdttset4, rihr0r0i, scn0knor; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1552.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,552 66.4 43.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dklk8tmi",
   "source_url": "https://omlx.ai/benchmarks/performance/dklk8tmi",
   "tg_tokens": null,
   "tps": 66.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-dklk8tmi-ctx16384",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.9 GB at this context; duplicate traces of the same session: fdttset4, rihr0r0i, scn0knor; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1461.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 1,461 58.9 43.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dklk8tmi",
   "source_url": "https://omlx.ai/benchmarks/performance/dklk8tmi",
   "tg_tokens": null,
   "tps": 58.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-dklk8tmi-ctx32768",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 44.7 GB at this context; duplicate traces of the same session: fdttset4, rihr0r0i, scn0knor; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1281.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 1,281 49.4 44.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dklk8tmi",
   "source_url": "https://omlx.ai/benchmarks/performance/dklk8tmi",
   "tg_tokens": null,
   "tps": 49.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-dklk8tmi-ctx65536",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 46.5 GB at this context; duplicate traces of the same session: fdttset4, rihr0r0i, scn0knor; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1007.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 1,007 36.9 46.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dklk8tmi",
   "source_url": "https://omlx.ai/benchmarks/performance/dklk8tmi",
   "tg_tokens": null,
   "tps": 36.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-oylwhkj1-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 80.1 GB at this context; duplicate traces of the same session: 75x2gbeu, jrf4vgpr, rhiba0g8, xill9rj9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1236.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 1,236 63.9 80.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oylwhkj1",
   "source_url": "https://omlx.ai/benchmarks/performance/oylwhkj1",
   "tg_tokens": null,
   "tps": 63.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-oylwhkj1-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 80.3 GB at this context; duplicate traces of the same session: 75x2gbeu, jrf4vgpr, rhiba0g8, xill9rj9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1509.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 1,509 61.9 80.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oylwhkj1",
   "source_url": "https://omlx.ai/benchmarks/performance/oylwhkj1",
   "tg_tokens": null,
   "tps": 61.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-oylwhkj1-ctx8192",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 80.5 GB at this context; duplicate traces of the same session: 75x2gbeu, jrf4vgpr, rhiba0g8, xill9rj9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1521.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 1,521 56.5 80.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oylwhkj1",
   "source_url": "https://omlx.ai/benchmarks/performance/oylwhkj1",
   "tg_tokens": null,
   "tps": 56.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-oylwhkj1-ctx16384",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 80.9 GB at this context; duplicate traces of the same session: 75x2gbeu, jrf4vgpr, rhiba0g8, xill9rj9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1434.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 16k 1,434 52.0 80.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oylwhkj1",
   "source_url": "https://omlx.ai/benchmarks/performance/oylwhkj1",
   "tg_tokens": null,
   "tps": 52.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-oylwhkj1-ctx32768",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 81.8 GB at this context; duplicate traces of the same session: 75x2gbeu, jrf4vgpr, rhiba0g8, xill9rj9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1262.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 32k 1,262 45.8 81.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oylwhkj1",
   "source_url": "https://omlx.ai/benchmarks/performance/oylwhkj1",
   "tg_tokens": null,
   "tps": 45.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-oylwhkj1-ctx65536",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 83.6 GB at this context; duplicate traces of the same session: 75x2gbeu, jrf4vgpr, rhiba0g8, xill9rj9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 994.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 64k 994.8 34.7 83.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oylwhkj1",
   "source_url": "https://omlx.ai/benchmarks/performance/oylwhkj1",
   "tg_tokens": null,
   "tps": 34.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-u2k59i3j-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 80.1 GB at this context; duplicate traces of the same session: 9kt8xfk2, y1fs7wsv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1442.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 1,442 58.8 80.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u2k59i3j",
   "source_url": "https://omlx.ai/benchmarks/performance/u2k59i3j",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-u2k59i3j-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 80.3 GB at this context; duplicate traces of the same session: 9kt8xfk2, y1fs7wsv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1872.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 1,872 58.5 80.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u2k59i3j",
   "source_url": "https://omlx.ai/benchmarks/performance/u2k59i3j",
   "tg_tokens": null,
   "tps": 58.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-u2k59i3j-ctx8192",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 80.5 GB at this context; duplicate traces of the same session: 9kt8xfk2, y1fs7wsv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1888.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 1,888 55.2 80.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u2k59i3j",
   "source_url": "https://omlx.ai/benchmarks/performance/u2k59i3j",
   "tg_tokens": null,
   "tps": 55.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-u2k59i3j-ctx16384",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 80.9 GB at this context; duplicate traces of the same session: 9kt8xfk2, y1fs7wsv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1783.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k 1,783 49.9 80.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u2k59i3j",
   "source_url": "https://omlx.ai/benchmarks/performance/u2k59i3j",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-u2k59i3j-ctx32768",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 81.8 GB at this context; duplicate traces of the same session: 9kt8xfk2, y1fs7wsv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1549.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 32k 1,549 42.5 81.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u2k59i3j",
   "source_url": "https://omlx.ai/benchmarks/performance/u2k59i3j",
   "tg_tokens": null,
   "tps": 42.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-7qprrfgz-ctx1024",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 1.5 GB at this context; duplicate traces of the same session: 3l7graoy, doabcv9b, od368nvv, vozr2mu4, y89wklsn; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 756.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 4bit; 1k 756.4 100.8 1.5 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7qprrfgz",
   "source_url": "https://omlx.ai/benchmarks/performance/7qprrfgz",
   "tg_tokens": null,
   "tps": 100.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-7qprrfgz-ctx4096",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 1.6 GB at this context; duplicate traces of the same session: 3l7graoy, doabcv9b, od368nvv, vozr2mu4, y89wklsn; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 734.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 4bit; 4k 734.1 91.3 1.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7qprrfgz",
   "source_url": "https://omlx.ai/benchmarks/performance/7qprrfgz",
   "tg_tokens": null,
   "tps": 91.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-7qprrfgz-ctx8192",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 1.6 GB at this context; duplicate traces of the same session: 3l7graoy, doabcv9b, od368nvv, vozr2mu4, y89wklsn; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 710.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 4bit; 8k 710.7 82.2 1.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7qprrfgz",
   "source_url": "https://omlx.ai/benchmarks/performance/7qprrfgz",
   "tg_tokens": null,
   "tps": 82.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-7qprrfgz-ctx16384",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 2.0 GB at this context; duplicate traces of the same session: 3l7graoy, doabcv9b, od368nvv, vozr2mu4, y89wklsn; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 647.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 4bit; 16k 647.6 68.2 2.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7qprrfgz",
   "source_url": "https://omlx.ai/benchmarks/performance/7qprrfgz",
   "tg_tokens": null,
   "tps": 68.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-7qprrfgz-ctx32768",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 2.6 GB at this context; duplicate traces of the same session: 3l7graoy, doabcv9b, od368nvv, vozr2mu4, y89wklsn; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 552.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 4bit; 32k 552.2 47.8 2.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7qprrfgz",
   "source_url": "https://omlx.ai/benchmarks/performance/7qprrfgz",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-7qprrfgz-ctx65536",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 4.1 GB at this context; duplicate traces of the same session: 3l7graoy, doabcv9b, od368nvv, vozr2mu4, y89wklsn; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 408.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 4bit; 64k 408.6 17.3 4.1 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7qprrfgz",
   "source_url": "https://omlx.ai/benchmarks/performance/7qprrfgz",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-piyx2eno-ctx1024",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 1.8 GB at this context; duplicate traces of the same session: 059q72vw, es1hp4a4, f00eert8, mxnj5jsm, wpqf311n; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 755.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 8bit; 1k 755.9 62.4 1.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/piyx2eno",
   "source_url": "https://omlx.ai/benchmarks/performance/piyx2eno",
   "tg_tokens": null,
   "tps": 62.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-piyx2eno-ctx4096",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 1.9 GB at this context; duplicate traces of the same session: 059q72vw, es1hp4a4, f00eert8, mxnj5jsm, wpqf311n; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 733.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 8bit; 4k 733.9 58.9 1.9 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/piyx2eno",
   "source_url": "https://omlx.ai/benchmarks/performance/piyx2eno",
   "tg_tokens": null,
   "tps": 58.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-piyx2eno-ctx8192",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 2.0 GB at this context; duplicate traces of the same session: 059q72vw, es1hp4a4, f00eert8, mxnj5jsm, wpqf311n; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 714.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 8bit; 8k 714.0 55.0 2.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/piyx2eno",
   "source_url": "https://omlx.ai/benchmarks/performance/piyx2eno",
   "tg_tokens": null,
   "tps": 55.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-piyx2eno-ctx16384",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 2.4 GB at this context; duplicate traces of the same session: 059q72vw, es1hp4a4, f00eert8, mxnj5jsm, wpqf311n; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 654.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 8bit; 16k 654.9 47.9 2.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/piyx2eno",
   "source_url": "https://omlx.ai/benchmarks/performance/piyx2eno",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-piyx2eno-ctx32768",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 3.0 GB at this context; duplicate traces of the same session: 059q72vw, es1hp4a4, f00eert8, mxnj5jsm, wpqf311n; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 554.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 8bit; 32k 554.7 37.3 3.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/piyx2eno",
   "source_url": "https://omlx.ai/benchmarks/performance/piyx2eno",
   "tg_tokens": null,
   "tps": 37.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-piyx2eno-ctx65536",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 4.4 GB at this context; duplicate traces of the same session: 059q72vw, es1hp4a4, f00eert8, mxnj5jsm, wpqf311n; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 414.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 8bit; 64k 414.9 24.8 4.4 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/piyx2eno",
   "source_url": "https://omlx.ai/benchmarks/performance/piyx2eno",
   "tg_tokens": null,
   "tps": 24.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-2xv9i0r9-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.3 GB at this context; duplicate traces of the same session: 47gtj6nx, 5fzqvnit; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 229.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 229.0 34.8 66.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2xv9i0r9",
   "source_url": "https://omlx.ai/benchmarks/performance/2xv9i0r9",
   "tg_tokens": null,
   "tps": 34.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-2xv9i0r9-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.7 GB at this context; duplicate traces of the same session: 47gtj6nx, 5fzqvnit; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 196.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 196.2 31.0 66.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2xv9i0r9",
   "source_url": "https://omlx.ai/benchmarks/performance/2xv9i0r9",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-2xv9i0r9-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.0 GB at this context; duplicate traces of the same session: 47gtj6nx, 5fzqvnit; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 192.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 192.2 27.8 67.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2xv9i0r9",
   "source_url": "https://omlx.ai/benchmarks/performance/2xv9i0r9",
   "tg_tokens": null,
   "tps": 27.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-2xv9i0r9-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.7 GB at this context; duplicate traces of the same session: 47gtj6nx, 5fzqvnit; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 183.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 183.0 22.5 67.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2xv9i0r9",
   "source_url": "https://omlx.ai/benchmarks/performance/2xv9i0r9",
   "tg_tokens": null,
   "tps": 22.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-2xv9i0r9-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 69.1 GB at this context; duplicate traces of the same session: 47gtj6nx, 5fzqvnit; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 170.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 32k 170.3 16.8 69.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2xv9i0r9",
   "source_url": "https://omlx.ai/benchmarks/performance/2xv9i0r9",
   "tg_tokens": null,
   "tps": 16.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-vrw9n3xu-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.3 GB at this context; duplicate traces of the same session: 2k6ofxl2, hfzzor6h, u28zirxu; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 439.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 1k 439.6 47.8 66.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vrw9n3xu",
   "source_url": "https://omlx.ai/benchmarks/performance/vrw9n3xu",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-vrw9n3xu-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.7 GB at this context; duplicate traces of the same session: 2k6ofxl2, hfzzor6h, u28zirxu; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 463.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 4k 463.3 45.6 66.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vrw9n3xu",
   "source_url": "https://omlx.ai/benchmarks/performance/vrw9n3xu",
   "tg_tokens": null,
   "tps": 45.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-vrw9n3xu-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.0 GB at this context; duplicate traces of the same session: 2k6ofxl2, hfzzor6h, u28zirxu; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 459.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 8k 459.9 41.3 67.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vrw9n3xu",
   "source_url": "https://omlx.ai/benchmarks/performance/vrw9n3xu",
   "tg_tokens": null,
   "tps": 41.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-vrw9n3xu-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.7 GB at this context; duplicate traces of the same session: 2k6ofxl2, hfzzor6h, u28zirxu; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 444.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 16k 444.0 36.3 67.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vrw9n3xu",
   "source_url": "https://omlx.ai/benchmarks/performance/vrw9n3xu",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-vrw9n3xu-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 69.1 GB at this context; duplicate traces of the same session: 2k6ofxl2, hfzzor6h, u28zirxu; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 409.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 32k 409.4 28.3 69.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vrw9n3xu",
   "source_url": "https://omlx.ai/benchmarks/performance/vrw9n3xu",
   "tg_tokens": null,
   "tps": 28.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-vrw9n3xu-ctx65536",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 71.9 GB at this context; duplicate traces of the same session: 2k6ofxl2, hfzzor6h, u28zirxu; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 331.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 64k 331.7 19.0 71.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vrw9n3xu",
   "source_url": "https://omlx.ai/benchmarks/performance/vrw9n3xu",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-nwfc7wtx-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.3 GB at this context; duplicate traces of the same session: bt6bvgoc, ejitl07v, jgotjhpb, n1ur8xge; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 704.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 704.1 53.8 66.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nwfc7wtx",
   "source_url": "https://omlx.ai/benchmarks/performance/nwfc7wtx",
   "tg_tokens": null,
   "tps": 53.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-nwfc7wtx-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.7 GB at this context; duplicate traces of the same session: bt6bvgoc, ejitl07v, jgotjhpb, n1ur8xge; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 729.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 729.1 51.2 66.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nwfc7wtx",
   "source_url": "https://omlx.ai/benchmarks/performance/nwfc7wtx",
   "tg_tokens": null,
   "tps": 51.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-nwfc7wtx-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.0 GB at this context; duplicate traces of the same session: bt6bvgoc, ejitl07v, jgotjhpb, n1ur8xge; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 712.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 712.9 48.2 67.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nwfc7wtx",
   "source_url": "https://omlx.ai/benchmarks/performance/nwfc7wtx",
   "tg_tokens": null,
   "tps": 48.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-nwfc7wtx-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.7 GB at this context; duplicate traces of the same session: bt6bvgoc, ejitl07v, jgotjhpb, n1ur8xge; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 670.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 670.3 41.2 67.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nwfc7wtx",
   "source_url": "https://omlx.ai/benchmarks/performance/nwfc7wtx",
   "tg_tokens": null,
   "tps": 41.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-nwfc7wtx-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 69.1 GB at this context; duplicate traces of the same session: bt6bvgoc, ejitl07v, jgotjhpb, n1ur8xge; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 590.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 590.2 35.0 69.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nwfc7wtx",
   "source_url": "https://omlx.ai/benchmarks/performance/nwfc7wtx",
   "tg_tokens": null,
   "tps": 35.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-nwfc7wtx-ctx65536",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 71.9 GB at this context; duplicate traces of the same session: bt6bvgoc, ejitl07v, jgotjhpb, n1ur8xge; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 466.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 466.9 26.0 71.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nwfc7wtx",
   "source_url": "https://omlx.ai/benchmarks/performance/nwfc7wtx",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-fxz58k75-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.3 GB at this context; duplicate traces of the same session: 919dis83, bzhe0vh3, d0ib6lbe, kvxrqtnz, tiqiwz3j; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 638.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 638.7 54.4 66.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fxz58k75",
   "source_url": "https://omlx.ai/benchmarks/performance/fxz58k75",
   "tg_tokens": null,
   "tps": 54.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-fxz58k75-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.7 GB at this context; duplicate traces of the same session: 919dis83, bzhe0vh3, d0ib6lbe, kvxrqtnz, tiqiwz3j; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 714.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 714.1 50.9 66.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fxz58k75",
   "source_url": "https://omlx.ai/benchmarks/performance/fxz58k75",
   "tg_tokens": null,
   "tps": 50.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-fxz58k75-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.0 GB at this context; duplicate traces of the same session: 919dis83, bzhe0vh3, d0ib6lbe, kvxrqtnz, tiqiwz3j; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 712.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 712.7 47.3 67.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fxz58k75",
   "source_url": "https://omlx.ai/benchmarks/performance/fxz58k75",
   "tg_tokens": null,
   "tps": 47.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-fxz58k75-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.7 GB at this context; duplicate traces of the same session: 919dis83, bzhe0vh3, d0ib6lbe, kvxrqtnz, tiqiwz3j; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 670.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 16k 670.7 35.9 67.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fxz58k75",
   "source_url": "https://omlx.ai/benchmarks/performance/fxz58k75",
   "tg_tokens": null,
   "tps": 35.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-fxz58k75-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 69.1 GB at this context; duplicate traces of the same session: 919dis83, bzhe0vh3, d0ib6lbe, kvxrqtnz, tiqiwz3j; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 587.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 587.5 24.4 69.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fxz58k75",
   "source_url": "https://omlx.ai/benchmarks/performance/fxz58k75",
   "tg_tokens": null,
   "tps": 24.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-fxz58k75-ctx65536",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 71.9 GB at this context; duplicate traces of the same session: 919dis83, bzhe0vh3, d0ib6lbe, kvxrqtnz, tiqiwz3j; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 458.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k 458.6 15.5 71.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fxz58k75",
   "source_url": "https://omlx.ai/benchmarks/performance/fxz58k75",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-uzfzs2vq-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 123.2 GB at this context; duplicate traces of the same session: i32xq8me; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 863.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 1k 863.0 42.9 123.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uzfzs2vq",
   "source_url": "https://omlx.ai/benchmarks/performance/uzfzs2vq",
   "tg_tokens": null,
   "tps": 42.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-uzfzs2vq-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 123.5 GB at this context; duplicate traces of the same session: i32xq8me; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 911.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 4k 911.3 41.4 123.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uzfzs2vq",
   "source_url": "https://omlx.ai/benchmarks/performance/uzfzs2vq",
   "tg_tokens": null,
   "tps": 41.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vx4ljpey-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.3 GB at this context; duplicate traces of the same session: 43mgnpxt, cavk7p1r, lhzzunu0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 897.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 897.8 52.3 66.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vx4ljpey",
   "source_url": "https://omlx.ai/benchmarks/performance/vx4ljpey",
   "tg_tokens": null,
   "tps": 52.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vx4ljpey-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.7 GB at this context; duplicate traces of the same session: 43mgnpxt, cavk7p1r, lhzzunu0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 935.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 935.9 47.9 66.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vx4ljpey",
   "source_url": "https://omlx.ai/benchmarks/performance/vx4ljpey",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vx4ljpey-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.0 GB at this context; duplicate traces of the same session: 43mgnpxt, cavk7p1r, lhzzunu0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 920.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 920.4 47.4 67.0 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vx4ljpey",
   "source_url": "https://omlx.ai/benchmarks/performance/vx4ljpey",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vx4ljpey-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.7 GB at this context; duplicate traces of the same session: 43mgnpxt, cavk7p1r, lhzzunu0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 862.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 862.3 43.1 67.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vx4ljpey",
   "source_url": "https://omlx.ai/benchmarks/performance/vx4ljpey",
   "tg_tokens": null,
   "tps": 43.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vx4ljpey-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 69.1 GB at this context; duplicate traces of the same session: 43mgnpxt, cavk7p1r, lhzzunu0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 746.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 746.1 36.4 69.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vx4ljpey",
   "source_url": "https://omlx.ai/benchmarks/performance/vx4ljpey",
   "tg_tokens": null,
   "tps": 36.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vx4ljpey-ctx65536",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 71.9 GB at this context; duplicate traces of the same session: 43mgnpxt, cavk7p1r, lhzzunu0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 570.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 570.2 27.7 71.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vx4ljpey",
   "source_url": "https://omlx.ai/benchmarks/performance/vx4ljpey",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-z1bl3y33-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 123.2 GB at this context; duplicate traces of the same session: 77dim65a, ho7w5nty; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 837.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 837.2 39.1 123.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z1bl3y33",
   "source_url": "https://omlx.ai/benchmarks/performance/z1bl3y33",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-z1bl3y33-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 123.5 GB at this context; duplicate traces of the same session: 77dim65a, ho7w5nty; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 895.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 895.1 37.8 123.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z1bl3y33",
   "source_url": "https://omlx.ai/benchmarks/performance/z1bl3y33",
   "tg_tokens": null,
   "tps": 37.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-z1bl3y33-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 123.9 GB at this context; duplicate traces of the same session: 77dim65a, ho7w5nty; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 876.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 876.6 35.7 123.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z1bl3y33",
   "source_url": "https://omlx.ai/benchmarks/performance/z1bl3y33",
   "tg_tokens": null,
   "tps": 35.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-z1bl3y33-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 124.5 GB at this context; duplicate traces of the same session: 77dim65a, ho7w5nty; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 826.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k 826.4 32.5 124.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z1bl3y33",
   "source_url": "https://omlx.ai/benchmarks/performance/z1bl3y33",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-z1bl3y33-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 125.9 GB at this context; duplicate traces of the same session: 77dim65a, ho7w5nty; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 717.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 32k 717.8 27.9 125.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z1bl3y33",
   "source_url": "https://omlx.ai/benchmarks/performance/z1bl3y33",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-xz6ox1sf-ctx1024",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 5raey9l1, 5y2fk7fi, 8rwegtmb, i0ji024h; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 226.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 1k 226.5 33.4 16.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xz6ox1sf",
   "source_url": "https://omlx.ai/benchmarks/performance/xz6ox1sf",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-xz6ox1sf-ctx4096",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 5raey9l1, 5y2fk7fi, 8rwegtmb, i0ji024h; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 228.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 4k 228.9 31.3 17.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xz6ox1sf",
   "source_url": "https://omlx.ai/benchmarks/performance/xz6ox1sf",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-xz6ox1sf-ctx8192",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: 5raey9l1, 5y2fk7fi, 8rwegtmb, i0ji024h; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 227.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 8k 227.0 29.3 17.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xz6ox1sf",
   "source_url": "https://omlx.ai/benchmarks/performance/xz6ox1sf",
   "tg_tokens": null,
   "tps": 29.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-xz6ox1sf-ctx16384",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: 5raey9l1, 5y2fk7fi, 8rwegtmb, i0ji024h; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 222.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 16k 222.3 26.6 18.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xz6ox1sf",
   "source_url": "https://omlx.ai/benchmarks/performance/xz6ox1sf",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-xz6ox1sf-ctx32768",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 5raey9l1, 5y2fk7fi, 8rwegtmb, i0ji024h; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 212.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 32k 212.9 21.9 20.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xz6ox1sf",
   "source_url": "https://omlx.ai/benchmarks/performance/xz6ox1sf",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-xz6ox1sf-ctx65536",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 5raey9l1, 5y2fk7fi, 8rwegtmb, i0ji024h; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 192.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 64k 192.2 16.8 23.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xz6ox1sf",
   "source_url": "https://omlx.ai/benchmarks/performance/xz6ox1sf",
   "tg_tokens": null,
   "tps": 16.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-8tvgq62t-ctx1024",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: ah2c9b47, iff2zzcf, jkbcbaj9, w3gm3t09; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 322.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 322.7 38.0 16.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8tvgq62t",
   "source_url": "https://omlx.ai/benchmarks/performance/8tvgq62t",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-8tvgq62t-ctx4096",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: ah2c9b47, iff2zzcf, jkbcbaj9, w3gm3t09; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 324.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 324.2 35.3 17.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8tvgq62t",
   "source_url": "https://omlx.ai/benchmarks/performance/8tvgq62t",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-8tvgq62t-ctx8192",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: ah2c9b47, iff2zzcf, jkbcbaj9, w3gm3t09; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 319.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 319.7 33.6 17.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8tvgq62t",
   "source_url": "https://omlx.ai/benchmarks/performance/8tvgq62t",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-8tvgq62t-ctx16384",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: ah2c9b47, iff2zzcf, jkbcbaj9, w3gm3t09; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 310.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 310.0 30.0 18.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8tvgq62t",
   "source_url": "https://omlx.ai/benchmarks/performance/8tvgq62t",
   "tg_tokens": null,
   "tps": 30.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-8tvgq62t-ctx32768",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: ah2c9b47, iff2zzcf, jkbcbaj9, w3gm3t09; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 291.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 291.3 24.9 20.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8tvgq62t",
   "source_url": "https://omlx.ai/benchmarks/performance/8tvgq62t",
   "tg_tokens": null,
   "tps": 24.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-8tvgq62t-ctx65536",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: ah2c9b47, iff2zzcf, jkbcbaj9, w3gm3t09; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 258.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 258.7 18.1 23.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8tvgq62t",
   "source_url": "https://omlx.ai/benchmarks/performance/8tvgq62t",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-yiokj60f-ctx1024",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 76f7vm8m, iugrfis1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 375.2,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 1k 375.2 26.7 22.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yiokj60f",
   "source_url": "https://omlx.ai/benchmarks/performance/yiokj60f",
   "tg_tokens": null,
   "tps": 26.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-yiokj60f-ctx4096",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: 76f7vm8m, iugrfis1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 401.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 4k 401.1 25.4 22.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yiokj60f",
   "source_url": "https://omlx.ai/benchmarks/performance/yiokj60f",
   "tg_tokens": null,
   "tps": 25.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-yiokj60f-ctx8192",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: 76f7vm8m, iugrfis1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 400.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 8k 400.8 24.4 22.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yiokj60f",
   "source_url": "https://omlx.ai/benchmarks/performance/yiokj60f",
   "tg_tokens": null,
   "tps": 24.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-yiokj60f-ctx16384",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: 76f7vm8m, iugrfis1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 392.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 16k 392.6 22.4 23.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yiokj60f",
   "source_url": "https://omlx.ai/benchmarks/performance/yiokj60f",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-yiokj60f-ctx32768",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: 76f7vm8m, iugrfis1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 370.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 32k 370.0 19.5 25.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yiokj60f",
   "source_url": "https://omlx.ai/benchmarks/performance/yiokj60f",
   "tg_tokens": null,
   "tps": 19.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-n58913j5-ctx1024",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: of4pm30v; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 238.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 238.7 28.8 15.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n58913j5",
   "source_url": "https://omlx.ai/benchmarks/performance/n58913j5",
   "tg_tokens": null,
   "tps": 28.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-n58913j5-ctx4096",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: of4pm30v; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 243.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 243.8 27.6 16.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n58913j5",
   "source_url": "https://omlx.ai/benchmarks/performance/n58913j5",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-zgli43rv-ctx65536",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 23.3 GB; TTFT 159348 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 411.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 64k tokens 411.3 PP tok/s 18.9 TG tok/s 159348 TTFT (ms) 23.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zgli43rv",
   "source_url": "https://omlx.ai/benchmarks/performance/zgli43rv",
   "tg_tokens": null,
   "tps": 18.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-mxylkvp8-ctx1024",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: byclik12; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 540.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 540.2 63.7 20.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mxylkvp8",
   "source_url": "https://omlx.ai/benchmarks/performance/mxylkvp8",
   "tg_tokens": null,
   "tps": 63.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-mxylkvp8-ctx4096",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: byclik12; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 585.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 585.0 60.1 20.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mxylkvp8",
   "source_url": "https://omlx.ai/benchmarks/performance/mxylkvp8",
   "tg_tokens": null,
   "tps": 60.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-p1rehpl3-ctx1024",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: 25m3innk; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 524.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-35B-A3B on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 1k 524.7 50.4 28.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p1rehpl3",
   "source_url": "https://omlx.ai/benchmarks/performance/p1rehpl3",
   "tg_tokens": null,
   "tps": 50.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-p1rehpl3-ctx4096",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.3 GB at this context; duplicate traces of the same session: 25m3innk; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 558.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-35B-A3B on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 4k 558.7 49.0 28.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p1rehpl3",
   "source_url": "https://omlx.ai/benchmarks/performance/p1rehpl3",
   "tg_tokens": null,
   "tps": 49.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rdlvb7oi-ctx1024",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 210.4 GB at this context; duplicate traces of the same session: 4o3fa12n, 6xmkfdgf, k0jw9l2d, vive2iyg; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 381.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 381.0 37.0 210.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rdlvb7oi",
   "source_url": "https://omlx.ai/benchmarks/performance/rdlvb7oi",
   "tg_tokens": null,
   "tps": 37.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rdlvb7oi-ctx4096",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 210.7 GB at this context; duplicate traces of the same session: 4o3fa12n, 6xmkfdgf, k0jw9l2d, vive2iyg; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 404.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 404.5 35.3 210.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rdlvb7oi",
   "source_url": "https://omlx.ai/benchmarks/performance/rdlvb7oi",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rdlvb7oi-ctx8192",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 211.0 GB at this context; duplicate traces of the same session: 4o3fa12n, 6xmkfdgf, k0jw9l2d, vive2iyg; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 400.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 400.7 31.3 211.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rdlvb7oi",
   "source_url": "https://omlx.ai/benchmarks/performance/rdlvb7oi",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rdlvb7oi-ctx16384",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 211.8 GB at this context; duplicate traces of the same session: 4o3fa12n, 6xmkfdgf, k0jw9l2d, vive2iyg; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 384.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 384.0 24.8 211.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rdlvb7oi",
   "source_url": "https://omlx.ai/benchmarks/performance/rdlvb7oi",
   "tg_tokens": null,
   "tps": 24.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rdlvb7oi-ctx32768",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 213.2 GB at this context; duplicate traces of the same session: 4o3fa12n, 6xmkfdgf, k0jw9l2d, vive2iyg; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 350.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 350.4 17.8 213.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rdlvb7oi",
   "source_url": "https://omlx.ai/benchmarks/performance/rdlvb7oi",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rdlvb7oi-ctx65536",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 216.2 GB at this context; duplicate traces of the same session: 4o3fa12n, 6xmkfdgf, k0jw9l2d, vive2iyg; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 294.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 294.3 11.7 216.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rdlvb7oi",
   "source_url": "https://omlx.ai/benchmarks/performance/rdlvb7oi",
   "tg_tokens": null,
   "tps": 11.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-au03cy2k-ctx1024",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 302.7 GB at this context; duplicate traces of the same session: q66jpzr6, rb7hx514; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 454.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 1k 454.0 28.5 302.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/au03cy2k",
   "source_url": "https://omlx.ai/benchmarks/performance/au03cy2k",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-au03cy2k-ctx4096",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 302.9 GB at this context; duplicate traces of the same session: q66jpzr6, rb7hx514; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 487.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 4k 487.1 27.2 302.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/au03cy2k",
   "source_url": "https://omlx.ai/benchmarks/performance/au03cy2k",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-au03cy2k-ctx8192",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 303.3 GB at this context; duplicate traces of the same session: q66jpzr6, rb7hx514; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 484.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 8k 484.9 26.0 303.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/au03cy2k",
   "source_url": "https://omlx.ai/benchmarks/performance/au03cy2k",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-au03cy2k-ctx16384",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 304.0 GB at this context; duplicate traces of the same session: q66jpzr6, rb7hx514; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 462.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 16k 462.0 23.6 304.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/au03cy2k",
   "source_url": "https://omlx.ai/benchmarks/performance/au03cy2k",
   "tg_tokens": null,
   "tps": 23.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-au03cy2k-ctx32768",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 305.5 GB at this context; duplicate traces of the same session: q66jpzr6, rb7hx514; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 422.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 32k 422.5 20.2 305.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/au03cy2k",
   "source_url": "https://omlx.ai/benchmarks/performance/au03cy2k",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-mq5sq5gu-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: yyrehuqw; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 338.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 338.1 31.9 3.9 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mq5sq5gu",
   "source_url": "https://omlx.ai/benchmarks/performance/mq5sq5gu",
   "tg_tokens": null,
   "tps": 31.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-mq5sq5gu-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: yyrehuqw; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 327.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 327.4 29.8 4.2 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mq5sq5gu",
   "source_url": "https://omlx.ai/benchmarks/performance/mq5sq5gu",
   "tg_tokens": null,
   "tps": 29.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-w2lccyfs-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: 6b5xaxkg; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 312.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 1k 312.5 56.6 6.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w2lccyfs",
   "source_url": "https://omlx.ai/benchmarks/performance/w2lccyfs",
   "tg_tokens": null,
   "tps": 56.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-w2lccyfs-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: 6b5xaxkg; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 317.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 4k 317.4 53.8 7.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w2lccyfs",
   "source_url": "https://omlx.ai/benchmarks/performance/w2lccyfs",
   "tg_tokens": null,
   "tps": 53.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-7hbe4e98-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.7 GB at this context; duplicate traces of the same session: 4tl4ubsq; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 304.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 1k 304.3 35.6 10.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7hbe4e98",
   "source_url": "https://omlx.ai/benchmarks/performance/7hbe4e98",
   "tg_tokens": null,
   "tps": 35.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-7hbe4e98-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.0 GB at this context; duplicate traces of the same session: 4tl4ubsq; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 319.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k 319.1 34.7 11.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7hbe4e98",
   "source_url": "https://omlx.ai/benchmarks/performance/7hbe4e98",
   "tg_tokens": null,
   "tps": 34.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-o5q2zwvz-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.7 GB at this context; duplicate traces of the same session: alvdhbe3; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 316.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 1k 316.3 35.9 10.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o5q2zwvz",
   "source_url": "https://omlx.ai/benchmarks/performance/o5q2zwvz",
   "tg_tokens": null,
   "tps": 35.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-o5q2zwvz-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.0 GB at this context; duplicate traces of the same session: alvdhbe3; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 321.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k 321.1 34.7 11.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o5q2zwvz",
   "source_url": "https://omlx.ai/benchmarks/performance/o5q2zwvz",
   "tg_tokens": null,
   "tps": 34.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-gp94nvfp-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.7 GB at this context; duplicate traces of the same session: a6fmmdir, gnjkbfpd, t8kof1zr, xpnt88fb; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 775.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 1k 775.8 74.3 8.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gp94nvfp",
   "source_url": "https://omlx.ai/benchmarks/performance/gp94nvfp",
   "tg_tokens": null,
   "tps": 74.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-gp94nvfp-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.9 GB at this context; duplicate traces of the same session: a6fmmdir, gnjkbfpd, t8kof1zr, xpnt88fb; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 783.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 4k 783.9 72.1 8.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gp94nvfp",
   "source_url": "https://omlx.ai/benchmarks/performance/gp94nvfp",
   "tg_tokens": null,
   "tps": 72.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-gp94nvfp-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 9.2 GB at this context; duplicate traces of the same session: a6fmmdir, gnjkbfpd, t8kof1zr, xpnt88fb; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 776.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 8k 776.7 68.7 9.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gp94nvfp",
   "source_url": "https://omlx.ai/benchmarks/performance/gp94nvfp",
   "tg_tokens": null,
   "tps": 68.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-gp94nvfp-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 9.6 GB at this context; duplicate traces of the same session: a6fmmdir, gnjkbfpd, t8kof1zr, xpnt88fb; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 758.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 16k 758.8 62.8 9.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gp94nvfp",
   "source_url": "https://omlx.ai/benchmarks/performance/gp94nvfp",
   "tg_tokens": null,
   "tps": 62.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-gp94nvfp-ctx32768",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.6 GB at this context; duplicate traces of the same session: a6fmmdir, gnjkbfpd, t8kof1zr, xpnt88fb; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 724.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 32k 724.1 52.6 10.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gp94nvfp",
   "source_url": "https://omlx.ai/benchmarks/performance/gp94nvfp",
   "tg_tokens": null,
   "tps": 52.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-gp94nvfp-ctx65536",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: a6fmmdir, gnjkbfpd, t8kof1zr, xpnt88fb; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 653.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 64k 653.8 39.8 12.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gp94nvfp",
   "source_url": "https://omlx.ai/benchmarks/performance/gp94nvfp",
   "tg_tokens": null,
   "tps": 39.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-uoy2dgmw-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: lagqqc8e; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 223.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 223.7 19.6 6.5 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uoy2dgmw",
   "source_url": "https://omlx.ai/benchmarks/performance/uoy2dgmw",
   "tg_tokens": null,
   "tps": 19.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-uoy2dgmw-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: lagqqc8e; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 224.8 20.7 6.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uoy2dgmw",
   "source_url": "https://omlx.ai/benchmarks/performance/uoy2dgmw",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-rlsu2zn7-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: uafhftej; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 225.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 225.0 21.7 6.5 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rlsu2zn7",
   "source_url": "https://omlx.ai/benchmarks/performance/rlsu2zn7",
   "tg_tokens": null,
   "tps": 21.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-rlsu2zn7-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: uafhftej; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 225.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 225.4 21.0 6.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rlsu2zn7",
   "source_url": "https://omlx.ai/benchmarks/performance/rlsu2zn7",
   "tg_tokens": null,
   "tps": 21.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-jk21vz3t-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: swdphr5s; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 218.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 218.1 19.8 6.5 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jk21vz3t",
   "source_url": "https://omlx.ai/benchmarks/performance/jk21vz3t",
   "tg_tokens": null,
   "tps": 19.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-jk21vz3t-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: swdphr5s; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 219.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 219.1 19.9 6.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jk21vz3t",
   "source_url": "https://omlx.ai/benchmarks/performance/jk21vz3t",
   "tg_tokens": null,
   "tps": 19.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-3lfp52kg-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.2 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 169.7,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Qwen3.5-9B-mxfp4 on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 169.7 21.7 7.2 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3lfp52kg",
   "source_url": "https://omlx.ai/benchmarks/performance/3lfp52kg",
   "tg_tokens": null,
   "tps": 21.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-3lfp52kg-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.7 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 169.0,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Qwen3.5-9B-mxfp4 on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 169.0 20.7 7.7 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3lfp52kg",
   "source_url": "https://omlx.ai/benchmarks/performance/3lfp52kg",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-7bx99kw6-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 744.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 744.3 108.4 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7bx99kw6",
   "source_url": "https://omlx.ai/benchmarks/performance/7bx99kw6",
   "tg_tokens": null,
   "tps": 108.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-7bx99kw6-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2064.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,064 105.4 20.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7bx99kw6",
   "source_url": "https://omlx.ai/benchmarks/performance/7bx99kw6",
   "tg_tokens": null,
   "tps": 105.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-65k3qlf9-ctx1024",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 237.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Abliterated-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 237.6 15.7 28.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/65k3qlf9",
   "source_url": "https://omlx.ai/benchmarks/performance/65k3qlf9",
   "tg_tokens": null,
   "tps": 15.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-65k3qlf9-ctx4096",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 32.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 182.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Abliterated-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 182.1 13.9 32.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/65k3qlf9",
   "source_url": "https://omlx.ai/benchmarks/performance/65k3qlf9",
   "tg_tokens": null,
   "tps": 13.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-p46idvkv-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: wufyjzmw, y79kne58; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 928.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 928.8 33.2 16.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p46idvkv",
   "source_url": "https://omlx.ai/benchmarks/performance/p46idvkv",
   "tg_tokens": null,
   "tps": 33.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-p46idvkv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: wufyjzmw, y79kne58; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1008.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,008 32.8 17.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p46idvkv",
   "source_url": "https://omlx.ai/benchmarks/performance/p46idvkv",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-p46idvkv-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: wufyjzmw, y79kne58; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 878.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 878.4 32.0 17.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p46idvkv",
   "source_url": "https://omlx.ai/benchmarks/performance/p46idvkv",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-p46idvkv-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: wufyjzmw, y79kne58; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 797.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 797.7 30.8 17.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p46idvkv",
   "source_url": "https://omlx.ai/benchmarks/performance/p46idvkv",
   "tg_tokens": null,
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3t9jllvp-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: 5h3m9gbn, bqy4j59f, hhxc5sho, j0dficgz, nme46ce0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 893.6,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 1k 893.6 38.2 13.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3t9jllvp",
   "source_url": "https://omlx.ai/benchmarks/performance/3t9jllvp",
   "tg_tokens": null,
   "tps": 38.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3t9jllvp-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: 5h3m9gbn, bqy4j59f, hhxc5sho, j0dficgz, nme46ce0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 935.7,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 4k 935.7 38.0 14.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3t9jllvp",
   "source_url": "https://omlx.ai/benchmarks/performance/3t9jllvp",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3t9jllvp-ctx8192",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: 5h3m9gbn, bqy4j59f, hhxc5sho, j0dficgz, nme46ce0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 852.6,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 8k 852.6 35.3 14.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3t9jllvp",
   "source_url": "https://omlx.ai/benchmarks/performance/3t9jllvp",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3t9jllvp-ctx16384",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 15.3 GB at this context; duplicate traces of the same session: 5h3m9gbn, bqy4j59f, hhxc5sho, j0dficgz, nme46ce0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 735.4,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 16k 735.4 32.5 15.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3t9jllvp",
   "source_url": "https://omlx.ai/benchmarks/performance/3t9jllvp",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3t9jllvp-ctx32768",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: 5h3m9gbn, bqy4j59f, hhxc5sho, j0dficgz, nme46ce0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 699.2,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 32k 699.2 32.1 16.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3t9jllvp",
   "source_url": "https://omlx.ai/benchmarks/performance/3t9jllvp",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3t9jllvp-ctx65536",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: 5h3m9gbn, bqy4j59f, hhxc5sho, j0dficgz, nme46ce0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 680.6,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 64k 680.6 27.9 18.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3t9jllvp",
   "source_url": "https://omlx.ai/benchmarks/performance/3t9jllvp",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3t9jllvp-ctx131072",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 5h3m9gbn, bqy4j59f, hhxc5sho, j0dficgz, nme46ce0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 549.4,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 128k 549.4 21.5 22.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3t9jllvp",
   "source_url": "https://omlx.ai/benchmarks/performance/3t9jllvp",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-5ejt4gud-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 454.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 454.2 18.2 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ejt4gud",
   "source_url": "https://omlx.ai/benchmarks/performance/5ejt4gud",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-5ejt4gud-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 500.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 500.8 17.8 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ejt4gud",
   "source_url": "https://omlx.ai/benchmarks/performance/5ejt4gud",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-ev03nt9i-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 228.6,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 1k 228.6 29.0 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ev03nt9i",
   "source_url": "https://omlx.ai/benchmarks/performance/ev03nt9i",
   "tg_tokens": null,
   "tps": 29.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-ev03nt9i-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 245.4,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 4k 245.4 28.3 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ev03nt9i",
   "source_url": "https://omlx.ai/benchmarks/performance/ev03nt9i",
   "tg_tokens": null,
   "tps": 28.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-ev03nt9i-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 239.8,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 8k 239.8 27.7 20.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ev03nt9i",
   "source_url": "https://omlx.ai/benchmarks/performance/ev03nt9i",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-ev03nt9i-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 229.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 16k 229.0 26.0 22.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ev03nt9i",
   "source_url": "https://omlx.ai/benchmarks/performance/ev03nt9i",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-t10uju16-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1192.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 64k 1,192 86.5 23.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t10uju16",
   "source_url": "https://omlx.ai/benchmarks/performance/t10uju16",
   "tg_tokens": null,
   "tps": 86.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-t10uju16-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.6 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 936.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 128k 936.6 74.6 27.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t10uju16",
   "source_url": "https://omlx.ai/benchmarks/performance/t10uju16",
   "tg_tokens": null,
   "tps": 74.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-d8hclmdc-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 210.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 210.2 27.9 33.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d8hclmdc",
   "source_url": "https://omlx.ai/benchmarks/performance/d8hclmdc",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-d8hclmdc-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 214.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 214.3 25.8 37.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d8hclmdc",
   "source_url": "https://omlx.ai/benchmarks/performance/d8hclmdc",
   "tg_tokens": null,
   "tps": 25.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-d8hclmdc-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 149.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 149.6 15.4 43.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d8hclmdc",
   "source_url": "https://omlx.ai/benchmarks/performance/d8hclmdc",
   "tg_tokens": null,
   "tps": 15.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-lr6txaa0-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 217.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 217.3 33.4 33.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lr6txaa0",
   "source_url": "https://omlx.ai/benchmarks/performance/lr6txaa0",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-lr6txaa0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 220.0 25.7 37.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lr6txaa0",
   "source_url": "https://omlx.ai/benchmarks/performance/lr6txaa0",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-lr6txaa0-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 142.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 142.9 18.2 43.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lr6txaa0",
   "source_url": "https://omlx.ai/benchmarks/performance/lr6txaa0",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ol8sfqde-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: chonyuwu; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 327.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 327.3 23.1 28.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ol8sfqde",
   "source_url": "https://omlx.ai/benchmarks/performance/ol8sfqde",
   "tg_tokens": null,
   "tps": 23.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ol8sfqde-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.2 GB at this context; duplicate traces of the same session: chonyuwu; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 337.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 337.7 22.8 32.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ol8sfqde",
   "source_url": "https://omlx.ai/benchmarks/performance/ol8sfqde",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ol8sfqde-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.2 GB at this context; duplicate traces of the same session: chonyuwu; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 332.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 332.6 22.2 33.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ol8sfqde",
   "source_url": "https://omlx.ai/benchmarks/performance/ol8sfqde",
   "tg_tokens": null,
   "tps": 22.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ol8sfqde-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.4 GB at this context; duplicate traces of the same session: chonyuwu; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 321.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 16k 321.9 21.7 32.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ol8sfqde",
   "source_url": "https://omlx.ai/benchmarks/performance/ol8sfqde",
   "tg_tokens": null,
   "tps": 21.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bun5fio0-ctx65536",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.2 GB at this context; duplicate traces of the same session: g8yms4n0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 854.9,
   "provenance": "community",
   "quant": "mxfp4+mxfp8",
   "quote": "Qwen3.8-Flash-Next-Uncensored-MLX-mixed-mxfp4_mxfp8_mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 854.9 32.2 103.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bun5fio0",
   "source_url": "https://omlx.ai/benchmarks/performance/bun5fio0",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bun5fio0-ctx131072",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.0 GB at this context; duplicate traces of the same session: g8yms4n0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 818.2,
   "provenance": "community",
   "quant": "mxfp4+mxfp8",
   "quote": "Qwen3.8-Flash-Next-Uncensored-MLX-mixed-mxfp4_mxfp8_mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 818.2 37.9 110.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bun5fio0",
   "source_url": "https://omlx.ai/benchmarks/performance/bun5fio0",
   "tg_tokens": null,
   "tps": 37.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bun5fio0-ctx199680",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 120.0 GB at this context; duplicate traces of the same session: g8yms4n0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 780.0,
   "provenance": "community",
   "quant": "mxfp4+mxfp8",
   "quote": "Qwen3.8-Flash-Next-Uncensored-MLX-mixed-mxfp4_mxfp8_mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 195k 780.0 37.4 120.0 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bun5fio0",
   "source_url": "https://omlx.ai/benchmarks/performance/bun5fio0",
   "tg_tokens": null,
   "tps": 37.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-8ei1yfsr-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: re3tiqd6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 525.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 525.3 56.3 71.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8ei1yfsr",
   "source_url": "https://omlx.ai/benchmarks/performance/8ei1yfsr",
   "tg_tokens": null,
   "tps": 56.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-8ei1yfsr-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: re3tiqd6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 592.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 592.8 53.4 71.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8ei1yfsr",
   "source_url": "https://omlx.ai/benchmarks/performance/8ei1yfsr",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-8ei1yfsr-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: re3tiqd6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 465.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 465.7 35.0 71.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8ei1yfsr",
   "source_url": "https://omlx.ai/benchmarks/performance/8ei1yfsr",
   "tg_tokens": null,
   "tps": 35.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-c6fnc70f-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 505.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 505.9 59.8 70.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c6fnc70f",
   "source_url": "https://omlx.ai/benchmarks/performance/c6fnc70f",
   "tg_tokens": null,
   "tps": 59.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-c6fnc70f-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 538.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 538.4 54.9 71.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c6fnc70f",
   "source_url": "https://omlx.ai/benchmarks/performance/c6fnc70f",
   "tg_tokens": null,
   "tps": 54.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ztljos0k-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1576.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,576 54.4 16.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ztljos0k",
   "source_url": "https://omlx.ai/benchmarks/performance/ztljos0k",
   "tg_tokens": null,
   "tps": 54.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ztljos0k-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1768.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 1,768 53.5 17.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ztljos0k",
   "source_url": "https://omlx.ai/benchmarks/performance/ztljos0k",
   "tg_tokens": null,
   "tps": 53.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-z5kfpqtv-ctx1024",
   "model": "clef",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 428.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "clef-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 428.1 37.6 16.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z5kfpqtv",
   "source_url": "https://omlx.ai/benchmarks/performance/z5kfpqtv",
   "tg_tokens": null,
   "tps": 37.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-z5kfpqtv-ctx4096",
   "model": "clef",
   "notes": "oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 472.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "clef-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 472.5 36.8 19.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z5kfpqtv",
   "source_url": "https://omlx.ai/benchmarks/performance/z5kfpqtv",
   "tg_tokens": null,
   "tps": 36.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-5oxq1gs1-ctx1024",
   "model": "gpt-oss 120B",
   "notes": "oMLX community benchmark trace; peak memory 58.0 GB at this context; duplicate traces of the same session: 3gd322b3, nqo8pz28; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 187.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gpt-oss-120b on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 187.7 54.9 58.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5oxq1gs1",
   "source_url": "https://omlx.ai/benchmarks/performance/5oxq1gs1",
   "tg_tokens": null,
   "tps": 54.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-5oxq1gs1-ctx4096",
   "model": "gpt-oss 120B",
   "notes": "oMLX community benchmark trace; peak memory 58.1 GB at this context; duplicate traces of the same session: 3gd322b3, nqo8pz28; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 182.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gpt-oss-120b on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 182.0 44.7 58.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5oxq1gs1",
   "source_url": "https://omlx.ai/benchmarks/performance/5oxq1gs1",
   "tg_tokens": null,
   "tps": 44.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-5oxq1gs1-ctx8192",
   "model": "gpt-oss 120B",
   "notes": "oMLX community benchmark trace; peak memory 58.3 GB at this context; duplicate traces of the same session: 3gd322b3, nqo8pz28; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 183.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gpt-oss-120b on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 183.9 37.1 58.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5oxq1gs1",
   "source_url": "https://omlx.ai/benchmarks/performance/5oxq1gs1",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-5oxq1gs1-ctx16384",
   "model": "gpt-oss 120B",
   "notes": "oMLX community benchmark trace; peak memory 58.6 GB at this context; duplicate traces of the same session: 3gd322b3, nqo8pz28; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 177.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gpt-oss-120b on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 177.1 22.7 58.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5oxq1gs1",
   "source_url": "https://omlx.ai/benchmarks/performance/5oxq1gs1",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-5oxq1gs1-ctx32768",
   "model": "gpt-oss 120B",
   "notes": "oMLX community benchmark trace; peak memory 59.2 GB at this context; duplicate traces of the same session: 3gd322b3, nqo8pz28; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 165.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gpt-oss-120b on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 32k 165.3 14.6 59.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5oxq1gs1",
   "source_url": "https://omlx.ai/benchmarks/performance/5oxq1gs1",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-sl2clrx6-ctx8192",
   "model": "gpt-oss 120B",
   "notes": "MXFP4 experts with Q8 attention; oMLX community benchmark trace; peak memory 59.5 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 365.9,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-120b-MXFP4-Q8 on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 8k 365.9 43.7 59.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sl2clrx6",
   "source_url": "https://omlx.ai/benchmarks/performance/sl2clrx6",
   "tg_tokens": null,
   "tps": 43.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-sl2clrx6-ctx16384",
   "model": "gpt-oss 120B",
   "notes": "MXFP4 experts with Q8 attention; oMLX community benchmark trace; peak memory 59.8 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 361.6,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-120b-MXFP4-Q8 on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 16k 361.6 38.4 59.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sl2clrx6",
   "source_url": "https://omlx.ai/benchmarks/performance/sl2clrx6",
   "tg_tokens": null,
   "tps": 38.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-oawwywpl-ctx1024",
   "model": "gpt-oss 120B",
   "notes": "MXFP4 experts with Q8 attention; oMLX community benchmark trace; peak memory 59.2 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 314.1,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-120b-MXFP4-Q8 on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 1k 314.1 67.7 59.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oawwywpl",
   "source_url": "https://omlx.ai/benchmarks/performance/oawwywpl",
   "tg_tokens": null,
   "tps": 67.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-oawwywpl-ctx4096",
   "model": "gpt-oss 120B",
   "notes": "MXFP4 experts with Q8 attention; oMLX community benchmark trace; peak memory 59.3 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 363.7,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-120b-MXFP4-Q8 on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 4k 363.7 60.1 59.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oawwywpl",
   "source_url": "https://omlx.ai/benchmarks/performance/oawwywpl",
   "tg_tokens": null,
   "tps": 60.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-rezzxtyy-ctx32768",
   "model": "gpt-oss 120B",
   "notes": "MXFP4 experts with Q8 attention; oMLX community benchmark trace; peak memory 60.4 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 337.1,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-120b-MXFP4-Q8 on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 32k 337.1 25.7 60.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rezzxtyy",
   "source_url": "https://omlx.ai/benchmarks/performance/rezzxtyy",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-rezzxtyy-ctx65536",
   "model": "gpt-oss 120B",
   "notes": "MXFP4 experts with Q8 attention; oMLX community benchmark trace; peak memory 61.6 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 294.5,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-120b-MXFP4-Q8 on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 64k 294.5 15.4 61.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rezzxtyy",
   "source_url": "https://omlx.ai/benchmarks/performance/rezzxtyy",
   "tg_tokens": null,
   "tps": 15.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-06kdatf0-ctx1024",
   "model": "gpt-oss 120B",
   "notes": "MXFP4 experts with Q8 attention; oMLX community benchmark trace; peak memory 59.2 GB at this context; duplicate traces of the same session: i54z07a7, wjn65xp9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 613.2,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-120b-MXFP4-Q8 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 613.2 68.4 59.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/06kdatf0",
   "source_url": "https://omlx.ai/benchmarks/performance/06kdatf0",
   "tg_tokens": null,
   "tps": 68.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-06kdatf0-ctx4096",
   "model": "gpt-oss 120B",
   "notes": "MXFP4 experts with Q8 attention; oMLX community benchmark trace; peak memory 59.3 GB at this context; duplicate traces of the same session: i54z07a7, wjn65xp9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 685.5,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-120b-MXFP4-Q8 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 685.5 64.1 59.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/06kdatf0",
   "source_url": "https://omlx.ai/benchmarks/performance/06kdatf0",
   "tg_tokens": null,
   "tps": 64.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-06kdatf0-ctx8192",
   "model": "gpt-oss 120B",
   "notes": "MXFP4 experts with Q8 attention; oMLX community benchmark trace; peak memory 59.5 GB at this context; duplicate traces of the same session: i54z07a7, wjn65xp9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 687.8,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-120b-MXFP4-Q8 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 687.8 57.8 59.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/06kdatf0",
   "source_url": "https://omlx.ai/benchmarks/performance/06kdatf0",
   "tg_tokens": null,
   "tps": 57.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-06kdatf0-ctx16384",
   "model": "gpt-oss 120B",
   "notes": "MXFP4 experts with Q8 attention; oMLX community benchmark trace; peak memory 59.8 GB at this context; duplicate traces of the same session: i54z07a7, wjn65xp9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 663.6,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-120b-MXFP4-Q8 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 663.6 47.5 59.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/06kdatf0",
   "source_url": "https://omlx.ai/benchmarks/performance/06kdatf0",
   "tg_tokens": null,
   "tps": 47.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-06kdatf0-ctx32768",
   "model": "gpt-oss 120B",
   "notes": "MXFP4 experts with Q8 attention; oMLX community benchmark trace; peak memory 60.4 GB at this context; duplicate traces of the same session: i54z07a7, wjn65xp9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 600.7,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-120b-MXFP4-Q8 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 600.7 35.2 60.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/06kdatf0",
   "source_url": "https://omlx.ai/benchmarks/performance/06kdatf0",
   "tg_tokens": null,
   "tps": 35.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-t9rdwkjo-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 562.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gpt-oss-20b-MLX-8bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 1k 562.4 51.9 21.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t9rdwkjo",
   "source_url": "https://omlx.ai/benchmarks/performance/t9rdwkjo",
   "tg_tokens": null,
   "tps": 51.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-t9rdwkjo-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 626.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gpt-oss-20b-MLX-8bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 4k 626.9 49.7 21.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t9rdwkjo",
   "source_url": "https://omlx.ai/benchmarks/performance/t9rdwkjo",
   "tg_tokens": null,
   "tps": 49.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-3hkr2x8v-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: ye45yn0j; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 105.9 15.6 18.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3hkr2x8v",
   "source_url": "https://omlx.ai/benchmarks/performance/3hkr2x8v",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-3hkr2x8v-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: ye45yn0j; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 151.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 151.3 15.0 19.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3hkr2x8v",
   "source_url": "https://omlx.ai/benchmarks/performance/3hkr2x8v",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-3hkr2x8v-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: ye45yn0j; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 115.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 115.0 14.5 19.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3hkr2x8v",
   "source_url": "https://omlx.ai/benchmarks/performance/3hkr2x8v",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-3hkr2x8v-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: ye45yn0j; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 96.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 96.1 12.8 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3hkr2x8v",
   "source_url": "https://omlx.ai/benchmarks/performance/3hkr2x8v",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-thcn3cyp-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: c510mcpw, u28zoca4; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 106.0 15.6 18.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/thcn3cyp",
   "source_url": "https://omlx.ai/benchmarks/performance/thcn3cyp",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-thcn3cyp-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: c510mcpw, u28zoca4; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 151.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 151.0 15.0 19.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/thcn3cyp",
   "source_url": "https://omlx.ai/benchmarks/performance/thcn3cyp",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-thcn3cyp-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: c510mcpw, u28zoca4; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 151.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 151.6 14.5 19.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/thcn3cyp",
   "source_url": "https://omlx.ai/benchmarks/performance/thcn3cyp",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-thcn3cyp-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: c510mcpw, u28zoca4; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 131.6 12.8 21.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/thcn3cyp",
   "source_url": "https://omlx.ai/benchmarks/performance/thcn3cyp",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6yjomi11-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "pyros-vault MTP build; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: agmyufp1, b1sluf0m, hj41ahwa, r2t9yutl, ziaa5bwx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 773.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "pyros-vault:Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 773.5 34.2 33.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6yjomi11",
   "source_url": "https://omlx.ai/benchmarks/performance/6yjomi11",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6yjomi11-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "pyros-vault MTP build; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: agmyufp1, b1sluf0m, hj41ahwa, r2t9yutl, ziaa5bwx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 850.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "pyros-vault:Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 850.0 39.2 34.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6yjomi11",
   "source_url": "https://omlx.ai/benchmarks/performance/6yjomi11",
   "tg_tokens": null,
   "tps": 39.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-6yjomi11-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "pyros-vault MTP build; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: agmyufp1, b1sluf0m, hj41ahwa, r2t9yutl, ziaa5bwx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 823.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "pyros-vault:Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 823.6 50.2 35.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6yjomi11",
   "source_url": "https://omlx.ai/benchmarks/performance/6yjomi11",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-6yjomi11-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "pyros-vault MTP build; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: agmyufp1, b1sluf0m, hj41ahwa, r2t9yutl, ziaa5bwx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 752.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "pyros-vault:Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 752.8 53.2 35.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6yjomi11",
   "source_url": "https://omlx.ai/benchmarks/performance/6yjomi11",
   "tg_tokens": null,
   "tps": 53.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-6yjomi11-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "pyros-vault MTP build; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: agmyufp1, b1sluf0m, hj41ahwa, r2t9yutl, ziaa5bwx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 709.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "pyros-vault:Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 709.2 55.2 36.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6yjomi11",
   "source_url": "https://omlx.ai/benchmarks/performance/6yjomi11",
   "tg_tokens": null,
   "tps": 55.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-6yjomi11-ctx65536",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "pyros-vault MTP build; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: agmyufp1, b1sluf0m, hj41ahwa, r2t9yutl, ziaa5bwx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 617.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "pyros-vault:Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 617.3 41.5 39.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6yjomi11",
   "source_url": "https://omlx.ai/benchmarks/performance/6yjomi11",
   "tg_tokens": null,
   "tps": 41.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-6yjomi11-ctx131072",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "pyros-vault MTP build; oMLX community benchmark trace; peak memory 83.2 GB at this context; duplicate traces of the same session: agmyufp1, b1sluf0m, hj41ahwa, r2t9yutl, ziaa5bwx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 329.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "pyros-vault:Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 329.1 22.7 83.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6yjomi11",
   "source_url": "https://omlx.ai/benchmarks/performance/6yjomi11",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-aaacem48-ctx1024",
   "model": "GLM-4.7-Flash",
   "notes": "oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: 3c0s4t7h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1160.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "GLM-4.7-Flash on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 1,160 60.2 30.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aaacem48",
   "source_url": "https://omlx.ai/benchmarks/performance/aaacem48",
   "tg_tokens": null,
   "tps": 60.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-aaacem48-ctx4096",
   "model": "GLM-4.7-Flash",
   "notes": "oMLX community benchmark trace; peak memory 31.3 GB at this context; duplicate traces of the same session: 3c0s4t7h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1091.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "GLM-4.7-Flash on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 1,091 55.3 31.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aaacem48",
   "source_url": "https://omlx.ai/benchmarks/performance/aaacem48",
   "tg_tokens": null,
   "tps": 55.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-314vh14k-ctx1024",
   "model": "GLM-5",
   "notes": "oMLX community benchmark trace; peak memory 391.8 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 186.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 186.0 16.6 391.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/314vh14k",
   "source_url": "https://omlx.ai/benchmarks/performance/314vh14k",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-314vh14k-ctx4096",
   "model": "GLM-5",
   "notes": "oMLX community benchmark trace; peak memory 394.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 180.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 180.1 13.8 394.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/314vh14k",
   "source_url": "https://omlx.ai/benchmarks/performance/314vh14k",
   "tg_tokens": null,
   "tps": 13.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-314vh14k-ctx8192",
   "model": "GLM-5",
   "notes": "oMLX community benchmark trace; peak memory 396.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 154.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 154.2 13.1 396.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/314vh14k",
   "source_url": "https://omlx.ai/benchmarks/performance/314vh14k",
   "tg_tokens": null,
   "tps": 13.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-314vh14k-ctx16384",
   "model": "GLM-5",
   "notes": "oMLX community benchmark trace; peak memory 402.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 117.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 117.3 12.0 402.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/314vh14k",
   "source_url": "https://omlx.ai/benchmarks/performance/314vh14k",
   "tg_tokens": null,
   "tps": 12.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-314vh14k-ctx32768",
   "model": "GLM-5",
   "notes": "oMLX community benchmark trace; peak memory 415.4 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 77.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 77.3 10.4 415.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/314vh14k",
   "source_url": "https://omlx.ai/benchmarks/performance/314vh14k",
   "tg_tokens": null,
   "tps": 10.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ugvdbq3h-ctx1024",
   "model": "GLM-5",
   "notes": "oMLX community benchmark trace; peak memory 420.0 GB at this context; duplicate traces of the same session: 3ctrecp9, teqbm0rb, xj3z6y6p; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 188.7,
   "provenance": "community",
   "quant": "4.8bit",
   "quote": "GLM-5-4.8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 188.7 16.6 420.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ugvdbq3h",
   "source_url": "https://omlx.ai/benchmarks/performance/ugvdbq3h",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ugvdbq3h-ctx4096",
   "model": "GLM-5",
   "notes": "oMLX community benchmark trace; peak memory 422.3 GB at this context; duplicate traces of the same session: 3ctrecp9, teqbm0rb, xj3z6y6p; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 182.6,
   "provenance": "community",
   "quant": "4.8bit",
   "quote": "GLM-5-4.8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 182.6 13.7 422.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ugvdbq3h",
   "source_url": "https://omlx.ai/benchmarks/performance/ugvdbq3h",
   "tg_tokens": null,
   "tps": 13.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ugvdbq3h-ctx8192",
   "model": "GLM-5",
   "notes": "oMLX community benchmark trace; peak memory 424.9 GB at this context; duplicate traces of the same session: 3ctrecp9, teqbm0rb, xj3z6y6p; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 156.3,
   "provenance": "community",
   "quant": "4.8bit",
   "quote": "GLM-5-4.8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 156.3 13.5 424.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ugvdbq3h",
   "source_url": "https://omlx.ai/benchmarks/performance/ugvdbq3h",
   "tg_tokens": null,
   "tps": 13.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ugvdbq3h-ctx16384",
   "model": "GLM-5",
   "notes": "oMLX community benchmark trace; peak memory 430.9 GB at this context; duplicate traces of the same session: 3ctrecp9, teqbm0rb, xj3z6y6p; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 119.8,
   "provenance": "community",
   "quant": "4.8bit",
   "quote": "GLM-5-4.8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 119.8 13.0 430.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ugvdbq3h",
   "source_url": "https://omlx.ai/benchmarks/performance/ugvdbq3h",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ugvdbq3h-ctx32768",
   "model": "GLM-5",
   "notes": "oMLX community benchmark trace; peak memory 443.6 GB at this context; duplicate traces of the same session: 3ctrecp9, teqbm0rb, xj3z6y6p; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 73.2,
   "provenance": "community",
   "quant": "4.8bit",
   "quote": "GLM-5-4.8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 73.2 12.3 443.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ugvdbq3h",
   "source_url": "https://omlx.ai/benchmarks/performance/ugvdbq3h",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-t9shfw5c-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2283.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,283 127.2 177.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t9shfw5c",
   "source_url": "https://omlx.ai/benchmarks/performance/t9shfw5c",
   "tg_tokens": null,
   "tps": 127.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-t9shfw5c-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.6 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2252.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,252 88.0 177.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t9shfw5c",
   "source_url": "https://omlx.ai/benchmarks/performance/t9shfw5c",
   "tg_tokens": null,
   "tps": 88.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-t9shfw5c-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.8 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2202.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,202 140.2 178.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t9shfw5c",
   "source_url": "https://omlx.ai/benchmarks/performance/t9shfw5c",
   "tg_tokens": null,
   "tps": 140.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-goj6q9vt-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2221.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,221 100.0 178.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/goj6q9vt",
   "source_url": "https://omlx.ai/benchmarks/performance/goj6q9vt",
   "tg_tokens": null,
   "tps": 100.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-goj6q9vt-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.6 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2198.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,198 74.4 178.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/goj6q9vt",
   "source_url": "https://omlx.ai/benchmarks/performance/goj6q9vt",
   "tg_tokens": null,
   "tps": 74.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-goj6q9vt-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 179.9 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2155.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,155 126.4 179.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/goj6q9vt",
   "source_url": "https://omlx.ai/benchmarks/performance/goj6q9vt",
   "tg_tokens": null,
   "tps": 126.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-rpi89413-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2225.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,225 108.7 178.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rpi89413",
   "source_url": "https://omlx.ai/benchmarks/performance/rpi89413",
   "tg_tokens": null,
   "tps": 108.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-rpi89413-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.6 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2199.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,199 80.0 178.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rpi89413",
   "source_url": "https://omlx.ai/benchmarks/performance/rpi89413",
   "tg_tokens": null,
   "tps": 80.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-rpi89413-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 179.9 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2156.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,156 113.8 179.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rpi89413",
   "source_url": "https://omlx.ai/benchmarks/performance/rpi89413",
   "tg_tokens": null,
   "tps": 113.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-40r63ctv-ctx1024",
   "model": "Llama-3.2-3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 2.3 GB at this context; duplicate traces of the same session: 4xl0yxqf; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 520.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Llama-3.2-3B-Instruct on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 520.4 45.2 2.3 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/40r63ctv",
   "source_url": "https://omlx.ai/benchmarks/performance/40r63ctv",
   "tg_tokens": null,
   "tps": 45.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-40r63ctv-ctx4096",
   "model": "Llama-3.2-3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 2.6 GB at this context; duplicate traces of the same session: 4xl0yxqf; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 478.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Llama-3.2-3B-Instruct on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 478.6 37.0 2.6 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/40r63ctv",
   "source_url": "https://omlx.ai/benchmarks/performance/40r63ctv",
   "tg_tokens": null,
   "tps": 37.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-s6lyqz2e-ctx1024",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 94.1 GB at this context; duplicate traces of the same session: 2kpxm1c1; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 299.7,
   "provenance": "community",
   "quant": "3bit",
   "quote": "MiniMax-M2.5 on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 3bit; 1k 299.7 42.1 94.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s6lyqz2e",
   "source_url": "https://omlx.ai/benchmarks/performance/s6lyqz2e",
   "tg_tokens": null,
   "tps": 42.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-s6lyqz2e-ctx4096",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 95.1 GB at this context; duplicate traces of the same session: 2kpxm1c1; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 337.6,
   "provenance": "community",
   "quant": "3bit",
   "quote": "MiniMax-M2.5 on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 3bit; 4k 337.6 35.7 95.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s6lyqz2e",
   "source_url": "https://omlx.ai/benchmarks/performance/s6lyqz2e",
   "tg_tokens": null,
   "tps": 35.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-s6lyqz2e-ctx8192",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 96.0 GB at this context; duplicate traces of the same session: 2kpxm1c1; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 320.5,
   "provenance": "community",
   "quant": "3bit",
   "quote": "MiniMax-M2.5 on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 3bit; 8k 320.5 29.5 96.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s6lyqz2e",
   "source_url": "https://omlx.ai/benchmarks/performance/s6lyqz2e",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-s6lyqz2e-ctx16384",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 98.0 GB at this context; duplicate traces of the same session: 2kpxm1c1; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 282.6,
   "provenance": "community",
   "quant": "3bit",
   "quote": "MiniMax-M2.5 on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 3bit; 16k 282.6 22.4 98.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s6lyqz2e",
   "source_url": "https://omlx.ai/benchmarks/performance/s6lyqz2e",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-e48q54i5-ctx1024",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 174.0 GB at this context; duplicate traces of the same session: 2we8lcjq, bzeg02vo, xbfycdsw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 496.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 1k 496.7 40.0 174.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e48q54i5",
   "source_url": "https://omlx.ai/benchmarks/performance/e48q54i5",
   "tg_tokens": null,
   "tps": 40.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-e48q54i5-ctx4096",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 175.0 GB at this context; duplicate traces of the same session: 2we8lcjq, bzeg02vo, xbfycdsw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 554.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 4k 554.8 33.9 175.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e48q54i5",
   "source_url": "https://omlx.ai/benchmarks/performance/e48q54i5",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-e48q54i5-ctx8192",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 175.9 GB at this context; duplicate traces of the same session: 2we8lcjq, bzeg02vo, xbfycdsw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 514.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 8k 514.0 28.9 175.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e48q54i5",
   "source_url": "https://omlx.ai/benchmarks/performance/e48q54i5",
   "tg_tokens": null,
   "tps": 28.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-e48q54i5-ctx16384",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 177.8 GB at this context; duplicate traces of the same session: 2we8lcjq, bzeg02vo, xbfycdsw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 436.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 16k 436.3 23.5 177.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e48q54i5",
   "source_url": "https://omlx.ai/benchmarks/performance/e48q54i5",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-e48q54i5-ctx32768",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 181.7 GB at this context; duplicate traces of the same session: 2we8lcjq, bzeg02vo, xbfycdsw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 330.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 32k 330.8 16.1 181.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e48q54i5",
   "source_url": "https://omlx.ai/benchmarks/performance/e48q54i5",
   "tg_tokens": null,
   "tps": 16.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-e48q54i5-ctx65536",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 189.6 GB at this context; duplicate traces of the same session: 2we8lcjq, bzeg02vo, xbfycdsw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 221.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 64k 221.1 9.6 189.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e48q54i5",
   "source_url": "https://omlx.ai/benchmarks/performance/e48q54i5",
   "tg_tokens": null,
   "tps": 9.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jye2fo4m-ctx1024",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 227.2 GB at this context; duplicate traces of the same session: s15twedr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 600.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 600.4 34.2 227.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jye2fo4m",
   "source_url": "https://omlx.ai/benchmarks/performance/jye2fo4m",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jye2fo4m-ctx4096",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 228.2 GB at this context; duplicate traces of the same session: s15twedr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 707.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 707.9 30.1 228.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jye2fo4m",
   "source_url": "https://omlx.ai/benchmarks/performance/jye2fo4m",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jye2fo4m-ctx8192",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 229.1 GB at this context; duplicate traces of the same session: s15twedr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 663.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 663.9 25.9 229.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jye2fo4m",
   "source_url": "https://omlx.ai/benchmarks/performance/jye2fo4m",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jye2fo4m-ctx16384",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 231.1 GB at this context; duplicate traces of the same session: s15twedr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 564.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k 564.2 21.1 231.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jye2fo4m",
   "source_url": "https://omlx.ai/benchmarks/performance/jye2fo4m",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jye2fo4m-ctx32768",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 235.0 GB at this context; duplicate traces of the same session: s15twedr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 427.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 32k 427.8 14.6 235.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jye2fo4m",
   "source_url": "https://omlx.ai/benchmarks/performance/jye2fo4m",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-jg59xbr1-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2066.0,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Ornith-1.5-35B-A3B-mlx-nvfp4 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 2,066 115.1 22.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jg59xbr1",
   "source_url": "https://omlx.ai/benchmarks/performance/jg59xbr1",
   "tg_tokens": null,
   "tps": 115.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-jg59xbr1-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1698.0,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Ornith-1.5-35B-A3B-mlx-nvfp4 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,698 105.1 23.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jg59xbr1",
   "source_url": "https://omlx.ai/benchmarks/performance/jg59xbr1",
   "tg_tokens": null,
   "tps": 105.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-jg59xbr1-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 934.2,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Ornith-1.5-35B-A3B-mlx-nvfp4 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 934.2 75.9 24.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jg59xbr1",
   "source_url": "https://omlx.ai/benchmarks/performance/jg59xbr1",
   "tg_tokens": null,
   "tps": 75.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-jg59xbr1-ctx199680",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 418.0,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Ornith-1.5-35B-A3B-mlx-nvfp4 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 195k 418.0 34.0 28.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jg59xbr1",
   "source_url": "https://omlx.ai/benchmarks/performance/jg59xbr1",
   "tg_tokens": null,
   "tps": 34.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-9d8qdwu5-ctx1024",
   "model": "Qwen2.5-Coder-14B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 8.5 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 271.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-14B-Instruct on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 271.2 34.4 8.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9d8qdwu5",
   "source_url": "https://omlx.ai/benchmarks/performance/9d8qdwu5",
   "tg_tokens": null,
   "tps": 34.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-9d8qdwu5-ctx4096",
   "model": "Qwen2.5-Coder-14B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 9.2 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 262.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-14B-Instruct on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 262.9 30.1 9.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9d8qdwu5",
   "source_url": "https://omlx.ai/benchmarks/performance/9d8qdwu5",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-4prten73-ctx1024",
   "model": "Qwen2.5-Coder-14B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 15.3 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 270.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen2.5-Coder-14B-Instruct on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 1k 270.8 20.4 15.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4prten73",
   "source_url": "https://omlx.ai/benchmarks/performance/4prten73",
   "tg_tokens": null,
   "tps": 20.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-4prten73-ctx4096",
   "model": "Qwen2.5-Coder-14B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 265.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen2.5-Coder-14B-Instruct on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k 265.9 18.8 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4prten73",
   "source_url": "https://omlx.ai/benchmarks/performance/4prten73",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-zmqaz63e-ctx1024",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 462.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 462.0 58.1 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zmqaz63e",
   "source_url": "https://omlx.ai/benchmarks/performance/zmqaz63e",
   "tg_tokens": null,
   "tps": 58.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-zmqaz63e-ctx4096",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 460.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 460.3 46.5 17.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zmqaz63e",
   "source_url": "https://omlx.ai/benchmarks/performance/zmqaz63e",
   "tg_tokens": null,
   "tps": 46.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-6v4q8j4y-ctx1024",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 433.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 1k 433.5 48.4 23.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6v4q8j4y",
   "source_url": "https://omlx.ai/benchmarks/performance/6v4q8j4y",
   "tg_tokens": null,
   "tps": 48.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-6v4q8j4y-ctx4096",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 443.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 4k 443.5 39.9 24.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6v4q8j4y",
   "source_url": "https://omlx.ai/benchmarks/performance/6v4q8j4y",
   "tg_tokens": null,
   "tps": 39.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ltejhkkd-ctx1024",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 30.8 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 444.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 1k 444.5 44.7 30.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ltejhkkd",
   "source_url": "https://omlx.ai/benchmarks/performance/ltejhkkd",
   "tg_tokens": null,
   "tps": 44.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-ltejhkkd-ctx4096",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 31.3 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 462.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k 462.4 37.1 31.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ltejhkkd",
   "source_url": "https://omlx.ai/benchmarks/performance/ltejhkkd",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-12jw4znp-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.0 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 400.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 1k 400.0 60.0 43.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/12jw4znp",
   "source_url": "https://omlx.ai/benchmarks/performance/12jw4znp",
   "tg_tokens": null,
   "tps": 60.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-12jw4znp-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.2 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 442.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 442.0 56.9 43.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/12jw4znp",
   "source_url": "https://omlx.ai/benchmarks/performance/12jw4znp",
   "tg_tokens": null,
   "tps": 56.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ee6x45d0-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.0 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 402.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 1k 402.4 59.7 43.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ee6x45d0",
   "source_url": "https://omlx.ai/benchmarks/performance/ee6x45d0",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ee6x45d0-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.2 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 441.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 441.6 56.1 43.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ee6x45d0",
   "source_url": "https://omlx.ai/benchmarks/performance/ee6x45d0",
   "tg_tokens": null,
   "tps": 56.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ipzlht96-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 61.6 GB at this context; duplicate traces of the same session: bg1g56a9; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 464.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 1k 464.3 53.4 61.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ipzlht96",
   "source_url": "https://omlx.ai/benchmarks/performance/ipzlht96",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ipzlht96-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 61.7 GB at this context; duplicate traces of the same session: bg1g56a9; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 443.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 4k 443.4 50.2 61.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ipzlht96",
   "source_url": "https://omlx.ai/benchmarks/performance/ipzlht96",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ipzlht96-ctx8192",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 62.0 GB at this context; duplicate traces of the same session: bg1g56a9; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 427.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 8k 427.1 46.1 62.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ipzlht96",
   "source_url": "https://omlx.ai/benchmarks/performance/ipzlht96",
   "tg_tokens": null,
   "tps": 46.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ipzlht96-ctx16384",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 62.4 GB at this context; duplicate traces of the same session: bg1g56a9; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 412.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 16k 412.9 38.1 62.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ipzlht96",
   "source_url": "https://omlx.ai/benchmarks/performance/ipzlht96",
   "tg_tokens": null,
   "tps": 38.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ipzlht96-ctx32768",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 63.3 GB at this context; duplicate traces of the same session: bg1g56a9; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 379.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 32k 379.4 25.5 63.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ipzlht96",
   "source_url": "https://omlx.ai/benchmarks/performance/ipzlht96",
   "tg_tokens": null,
   "tps": 25.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-l7j9n9kx-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 61.6 GB at this context; duplicate traces of the same session: bgwyirz3, n6v3m0w2; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 751.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 1k 751.3 57.9 61.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l7j9n9kx",
   "source_url": "https://omlx.ai/benchmarks/performance/l7j9n9kx",
   "tg_tokens": null,
   "tps": 57.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-l7j9n9kx-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 61.7 GB at this context; duplicate traces of the same session: bgwyirz3, n6v3m0w2; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 966.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 4k 966.4 56.2 61.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l7j9n9kx",
   "source_url": "https://omlx.ai/benchmarks/performance/l7j9n9kx",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-l7j9n9kx-ctx8192",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 62.0 GB at this context; duplicate traces of the same session: bgwyirz3, n6v3m0w2; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 986.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 8k 986.0 53.4 62.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l7j9n9kx",
   "source_url": "https://omlx.ai/benchmarks/performance/l7j9n9kx",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-l7j9n9kx-ctx16384",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 62.4 GB at this context; duplicate traces of the same session: bgwyirz3, n6v3m0w2; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 956.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 16k 956.5 47.0 62.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l7j9n9kx",
   "source_url": "https://omlx.ai/benchmarks/performance/l7j9n9kx",
   "tg_tokens": null,
   "tps": 47.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-l7j9n9kx-ctx32768",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 63.3 GB at this context; duplicate traces of the same session: bgwyirz3, n6v3m0w2; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 881.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 32k 881.3 39.1 63.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l7j9n9kx",
   "source_url": "https://omlx.ai/benchmarks/performance/l7j9n9kx",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-l7j9n9kx-ctx65536",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 65.1 GB at this context; duplicate traces of the same session: bgwyirz3, n6v3m0w2; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 731.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 64k 731.4 28.9 65.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l7j9n9kx",
   "source_url": "https://omlx.ai/benchmarks/performance/l7j9n9kx",
   "tg_tokens": null,
   "tps": 28.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-fdttset4-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.0 GB at this context; duplicate traces of the same session: rihr0r0i, scn0knor; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1298.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 1,298 73.9 43.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fdttset4",
   "source_url": "https://omlx.ai/benchmarks/performance/fdttset4",
   "tg_tokens": null,
   "tps": 73.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-fdttset4-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.2 GB at this context; duplicate traces of the same session: rihr0r0i, scn0knor; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1558.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 1,558 71.0 43.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fdttset4",
   "source_url": "https://omlx.ai/benchmarks/performance/fdttset4",
   "tg_tokens": null,
   "tps": 71.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-fdttset4-ctx8192",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.4 GB at this context; duplicate traces of the same session: rihr0r0i, scn0knor; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1552.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,552 66.4 43.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fdttset4",
   "source_url": "https://omlx.ai/benchmarks/performance/fdttset4",
   "tg_tokens": null,
   "tps": 66.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-fdttset4-ctx16384",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.9 GB at this context; duplicate traces of the same session: rihr0r0i, scn0knor; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1461.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 1,461 58.9 43.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fdttset4",
   "source_url": "https://omlx.ai/benchmarks/performance/fdttset4",
   "tg_tokens": null,
   "tps": 58.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-fdttset4-ctx32768",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 44.7 GB at this context; duplicate traces of the same session: rihr0r0i, scn0knor; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1281.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 1,281 49.4 44.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fdttset4",
   "source_url": "https://omlx.ai/benchmarks/performance/fdttset4",
   "tg_tokens": null,
   "tps": 49.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-fdttset4-ctx65536",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 46.5 GB at this context; duplicate traces of the same session: rihr0r0i, scn0knor; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1007.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 1,007 36.9 46.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fdttset4",
   "source_url": "https://omlx.ai/benchmarks/performance/fdttset4",
   "tg_tokens": null,
   "tps": 36.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-75x2gbeu-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 80.1 GB at this context; duplicate traces of the same session: jrf4vgpr, rhiba0g8, xill9rj9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1236.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 1,236 63.9 80.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/75x2gbeu",
   "source_url": "https://omlx.ai/benchmarks/performance/75x2gbeu",
   "tg_tokens": null,
   "tps": 63.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-75x2gbeu-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 80.3 GB at this context; duplicate traces of the same session: jrf4vgpr, rhiba0g8, xill9rj9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1509.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 1,509 61.9 80.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/75x2gbeu",
   "source_url": "https://omlx.ai/benchmarks/performance/75x2gbeu",
   "tg_tokens": null,
   "tps": 61.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-75x2gbeu-ctx8192",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 80.5 GB at this context; duplicate traces of the same session: jrf4vgpr, rhiba0g8, xill9rj9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1521.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 1,521 56.5 80.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/75x2gbeu",
   "source_url": "https://omlx.ai/benchmarks/performance/75x2gbeu",
   "tg_tokens": null,
   "tps": 56.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-75x2gbeu-ctx16384",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 80.9 GB at this context; duplicate traces of the same session: jrf4vgpr, rhiba0g8, xill9rj9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1434.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 16k 1,434 52.0 80.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/75x2gbeu",
   "source_url": "https://omlx.ai/benchmarks/performance/75x2gbeu",
   "tg_tokens": null,
   "tps": 52.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-75x2gbeu-ctx32768",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 81.8 GB at this context; duplicate traces of the same session: jrf4vgpr, rhiba0g8, xill9rj9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1262.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 32k 1,262 45.8 81.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/75x2gbeu",
   "source_url": "https://omlx.ai/benchmarks/performance/75x2gbeu",
   "tg_tokens": null,
   "tps": 45.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-75x2gbeu-ctx65536",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 83.6 GB at this context; duplicate traces of the same session: jrf4vgpr, rhiba0g8, xill9rj9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 994.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 64k 994.8 34.7 83.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/75x2gbeu",
   "source_url": "https://omlx.ai/benchmarks/performance/75x2gbeu",
   "tg_tokens": null,
   "tps": 34.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-y1fs7wsv-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 80.1 GB at this context; duplicate traces of the same session: 9kt8xfk2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1442.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 1,442 58.8 80.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y1fs7wsv",
   "source_url": "https://omlx.ai/benchmarks/performance/y1fs7wsv",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-y1fs7wsv-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 80.3 GB at this context; duplicate traces of the same session: 9kt8xfk2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1872.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 1,872 58.5 80.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y1fs7wsv",
   "source_url": "https://omlx.ai/benchmarks/performance/y1fs7wsv",
   "tg_tokens": null,
   "tps": 58.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-y1fs7wsv-ctx8192",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 80.5 GB at this context; duplicate traces of the same session: 9kt8xfk2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1888.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 1,888 55.2 80.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y1fs7wsv",
   "source_url": "https://omlx.ai/benchmarks/performance/y1fs7wsv",
   "tg_tokens": null,
   "tps": 55.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-y1fs7wsv-ctx16384",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 80.9 GB at this context; duplicate traces of the same session: 9kt8xfk2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1783.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k 1,783 49.9 80.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y1fs7wsv",
   "source_url": "https://omlx.ai/benchmarks/performance/y1fs7wsv",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-y1fs7wsv-ctx32768",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 81.8 GB at this context; duplicate traces of the same session: 9kt8xfk2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1549.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 32k 1,549 42.5 81.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y1fs7wsv",
   "source_url": "https://omlx.ai/benchmarks/performance/y1fs7wsv",
   "tg_tokens": null,
   "tps": 42.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-vozr2mu4-ctx1024",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 1.5 GB at this context; duplicate traces of the same session: 3l7graoy, doabcv9b, od368nvv, y89wklsn; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 756.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 4bit; 1k 756.4 100.8 1.5 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vozr2mu4",
   "source_url": "https://omlx.ai/benchmarks/performance/vozr2mu4",
   "tg_tokens": null,
   "tps": 100.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-vozr2mu4-ctx4096",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 1.6 GB at this context; duplicate traces of the same session: 3l7graoy, doabcv9b, od368nvv, y89wklsn; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 734.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 4bit; 4k 734.1 91.3 1.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vozr2mu4",
   "source_url": "https://omlx.ai/benchmarks/performance/vozr2mu4",
   "tg_tokens": null,
   "tps": 91.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-vozr2mu4-ctx8192",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 1.6 GB at this context; duplicate traces of the same session: 3l7graoy, doabcv9b, od368nvv, y89wklsn; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 710.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 4bit; 8k 710.7 82.2 1.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vozr2mu4",
   "source_url": "https://omlx.ai/benchmarks/performance/vozr2mu4",
   "tg_tokens": null,
   "tps": 82.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-vozr2mu4-ctx16384",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 2.0 GB at this context; duplicate traces of the same session: 3l7graoy, doabcv9b, od368nvv, y89wklsn; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 647.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 4bit; 16k 647.6 68.2 2.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vozr2mu4",
   "source_url": "https://omlx.ai/benchmarks/performance/vozr2mu4",
   "tg_tokens": null,
   "tps": 68.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-vozr2mu4-ctx32768",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 2.6 GB at this context; duplicate traces of the same session: 3l7graoy, doabcv9b, od368nvv, y89wklsn; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 552.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 4bit; 32k 552.2 47.8 2.6 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vozr2mu4",
   "source_url": "https://omlx.ai/benchmarks/performance/vozr2mu4",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-vozr2mu4-ctx65536",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 4.1 GB at this context; duplicate traces of the same session: 3l7graoy, doabcv9b, od368nvv, y89wklsn; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 408.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 4bit; 64k 408.6 17.3 4.1 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vozr2mu4",
   "source_url": "https://omlx.ai/benchmarks/performance/vozr2mu4",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-059q72vw-ctx1024",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 1.8 GB at this context; duplicate traces of the same session: es1hp4a4, f00eert8, mxnj5jsm, wpqf311n; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 755.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 8bit; 1k 755.9 62.4 1.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/059q72vw",
   "source_url": "https://omlx.ai/benchmarks/performance/059q72vw",
   "tg_tokens": null,
   "tps": 62.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-059q72vw-ctx4096",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 1.9 GB at this context; duplicate traces of the same session: es1hp4a4, f00eert8, mxnj5jsm, wpqf311n; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 733.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 8bit; 4k 733.9 58.9 1.9 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/059q72vw",
   "source_url": "https://omlx.ai/benchmarks/performance/059q72vw",
   "tg_tokens": null,
   "tps": 58.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-059q72vw-ctx8192",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 2.0 GB at this context; duplicate traces of the same session: es1hp4a4, f00eert8, mxnj5jsm, wpqf311n; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 714.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 8bit; 8k 714.0 55.0 2.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/059q72vw",
   "source_url": "https://omlx.ai/benchmarks/performance/059q72vw",
   "tg_tokens": null,
   "tps": 55.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-059q72vw-ctx16384",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 2.4 GB at this context; duplicate traces of the same session: es1hp4a4, f00eert8, mxnj5jsm, wpqf311n; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 654.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 8bit; 16k 654.9 47.9 2.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/059q72vw",
   "source_url": "https://omlx.ai/benchmarks/performance/059q72vw",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-059q72vw-ctx32768",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 3.0 GB at this context; duplicate traces of the same session: es1hp4a4, f00eert8, mxnj5jsm, wpqf311n; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 554.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 8bit; 32k 554.7 37.3 3.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/059q72vw",
   "source_url": "https://omlx.ai/benchmarks/performance/059q72vw",
   "tg_tokens": null,
   "tps": 37.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-059q72vw-ctx65536",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 4.4 GB at this context; duplicate traces of the same session: es1hp4a4, f00eert8, mxnj5jsm, wpqf311n; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 414.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 8bit; 64k 414.9 24.8 4.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/059q72vw",
   "source_url": "https://omlx.ai/benchmarks/performance/059q72vw",
   "tg_tokens": null,
   "tps": 24.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-wn6qfcj7-ctx1024",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 1.8 GB at this context; duplicate traces of the same session: 7bv8ke3w; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 713.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 8bit; 1k 713.6 46.2 1.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wn6qfcj7",
   "source_url": "https://omlx.ai/benchmarks/performance/wn6qfcj7",
   "tg_tokens": null,
   "tps": 46.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-wn6qfcj7-ctx4096",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 1.9 GB at this context; duplicate traces of the same session: 7bv8ke3w; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 707.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 8bit; 4k 707.8 44.7 1.9 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wn6qfcj7",
   "source_url": "https://omlx.ai/benchmarks/performance/wn6qfcj7",
   "tg_tokens": null,
   "tps": 44.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-on0kwjmb-ctx1024",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 1.8 GB at this context; duplicate traces of the same session: hj1h2047, udv2mwwl; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 726.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 8bit; 1k 726.8 48.2 1.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/on0kwjmb",
   "source_url": "https://omlx.ai/benchmarks/performance/on0kwjmb",
   "tg_tokens": null,
   "tps": 48.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-on0kwjmb-ctx4096",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 1.9 GB at this context; duplicate traces of the same session: hj1h2047, udv2mwwl; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 653.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 8bit; 4k 653.5 41.2 1.9 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/on0kwjmb",
   "source_url": "https://omlx.ai/benchmarks/performance/on0kwjmb",
   "tg_tokens": null,
   "tps": 41.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-on0kwjmb-ctx8192",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 2.0 GB at this context; duplicate traces of the same session: hj1h2047, udv2mwwl; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 646.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 8bit; 8k 646.1 45.1 2.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/on0kwjmb",
   "source_url": "https://omlx.ai/benchmarks/performance/on0kwjmb",
   "tg_tokens": null,
   "tps": 45.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-k4yuhp35-ctx65536",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 5.0 GB; TTFT 14065 ms; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4660.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.5-0.8B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 bf16; 64k tokens 4,660 PP tok/s 83.2 TG tok/s 14065 TTFT (ms) 5.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k4yuhp35",
   "source_url": "https://omlx.ai/benchmarks/performance/k4yuhp35",
   "tg_tokens": null,
   "tps": 83.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-0dakblwg-ctx1024",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 1.8 GB; TTFT 619 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1656.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 8bit; 1k tokens 1,656 PP tok/s 90.4 TG tok/s 619 TTFT (ms) 1.8 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0dakblwg",
   "source_url": "https://omlx.ai/benchmarks/performance/0dakblwg",
   "tg_tokens": null,
   "tps": 90.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-4vyestqf-ctx1024",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 1.8 GB; TTFT 499 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2051.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 8bit; 1k tokens 2,051 PP tok/s 99.8 TG tok/s 499 TTFT (ms) 1.8 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4vyestqf",
   "source_url": "https://omlx.ai/benchmarks/performance/4vyestqf",
   "tg_tokens": null,
   "tps": 99.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-47gtj6nx-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.3 GB at this context; duplicate traces of the same session: 5fzqvnit; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 229.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 229.0 34.8 66.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/47gtj6nx",
   "source_url": "https://omlx.ai/benchmarks/performance/47gtj6nx",
   "tg_tokens": null,
   "tps": 34.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-47gtj6nx-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.7 GB at this context; duplicate traces of the same session: 5fzqvnit; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 196.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 196.2 31.0 66.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/47gtj6nx",
   "source_url": "https://omlx.ai/benchmarks/performance/47gtj6nx",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-47gtj6nx-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.0 GB at this context; duplicate traces of the same session: 5fzqvnit; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 192.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 192.2 27.8 67.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/47gtj6nx",
   "source_url": "https://omlx.ai/benchmarks/performance/47gtj6nx",
   "tg_tokens": null,
   "tps": 27.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-47gtj6nx-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.7 GB at this context; duplicate traces of the same session: 5fzqvnit; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 183.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 183.0 22.5 67.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/47gtj6nx",
   "source_url": "https://omlx.ai/benchmarks/performance/47gtj6nx",
   "tg_tokens": null,
   "tps": 22.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-47gtj6nx-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 69.1 GB at this context; duplicate traces of the same session: 5fzqvnit; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 170.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 32k 170.3 16.8 69.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/47gtj6nx",
   "source_url": "https://omlx.ai/benchmarks/performance/47gtj6nx",
   "tg_tokens": null,
   "tps": 16.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-u28zirxu-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.3 GB at this context; duplicate traces of the same session: 2k6ofxl2, hfzzor6h; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 439.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 1k 439.6 47.8 66.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u28zirxu",
   "source_url": "https://omlx.ai/benchmarks/performance/u28zirxu",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-u28zirxu-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.7 GB at this context; duplicate traces of the same session: 2k6ofxl2, hfzzor6h; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 463.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 4k 463.3 45.6 66.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u28zirxu",
   "source_url": "https://omlx.ai/benchmarks/performance/u28zirxu",
   "tg_tokens": null,
   "tps": 45.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-u28zirxu-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.0 GB at this context; duplicate traces of the same session: 2k6ofxl2, hfzzor6h; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 459.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 8k 459.9 41.3 67.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u28zirxu",
   "source_url": "https://omlx.ai/benchmarks/performance/u28zirxu",
   "tg_tokens": null,
   "tps": 41.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-u28zirxu-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.7 GB at this context; duplicate traces of the same session: 2k6ofxl2, hfzzor6h; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 444.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 16k 444.0 36.3 67.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u28zirxu",
   "source_url": "https://omlx.ai/benchmarks/performance/u28zirxu",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-u28zirxu-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 69.1 GB at this context; duplicate traces of the same session: 2k6ofxl2, hfzzor6h; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 409.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 32k 409.4 28.3 69.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u28zirxu",
   "source_url": "https://omlx.ai/benchmarks/performance/u28zirxu",
   "tg_tokens": null,
   "tps": 28.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-u28zirxu-ctx65536",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 71.9 GB at this context; duplicate traces of the same session: 2k6ofxl2, hfzzor6h; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 331.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 64k 331.7 19.0 71.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u28zirxu",
   "source_url": "https://omlx.ai/benchmarks/performance/u28zirxu",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-n1ur8xge-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.3 GB at this context; duplicate traces of the same session: bt6bvgoc, ejitl07v, jgotjhpb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 704.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 704.1 53.8 66.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n1ur8xge",
   "source_url": "https://omlx.ai/benchmarks/performance/n1ur8xge",
   "tg_tokens": null,
   "tps": 53.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-n1ur8xge-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.7 GB at this context; duplicate traces of the same session: bt6bvgoc, ejitl07v, jgotjhpb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 729.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 729.1 51.2 66.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n1ur8xge",
   "source_url": "https://omlx.ai/benchmarks/performance/n1ur8xge",
   "tg_tokens": null,
   "tps": 51.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-n1ur8xge-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.0 GB at this context; duplicate traces of the same session: bt6bvgoc, ejitl07v, jgotjhpb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 712.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 712.9 48.2 67.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n1ur8xge",
   "source_url": "https://omlx.ai/benchmarks/performance/n1ur8xge",
   "tg_tokens": null,
   "tps": 48.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-n1ur8xge-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.7 GB at this context; duplicate traces of the same session: bt6bvgoc, ejitl07v, jgotjhpb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 670.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 670.3 41.2 67.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n1ur8xge",
   "source_url": "https://omlx.ai/benchmarks/performance/n1ur8xge",
   "tg_tokens": null,
   "tps": 41.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-n1ur8xge-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 69.1 GB at this context; duplicate traces of the same session: bt6bvgoc, ejitl07v, jgotjhpb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 590.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 590.2 35.0 69.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n1ur8xge",
   "source_url": "https://omlx.ai/benchmarks/performance/n1ur8xge",
   "tg_tokens": null,
   "tps": 35.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-n1ur8xge-ctx65536",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 71.9 GB at this context; duplicate traces of the same session: bt6bvgoc, ejitl07v, jgotjhpb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 466.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 466.9 26.0 71.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n1ur8xge",
   "source_url": "https://omlx.ai/benchmarks/performance/n1ur8xge",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-d0ib6lbe-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.3 GB at this context; duplicate traces of the same session: 919dis83, bzhe0vh3, kvxrqtnz, tiqiwz3j; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 638.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 638.7 54.4 66.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d0ib6lbe",
   "source_url": "https://omlx.ai/benchmarks/performance/d0ib6lbe",
   "tg_tokens": null,
   "tps": 54.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-d0ib6lbe-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.7 GB at this context; duplicate traces of the same session: 919dis83, bzhe0vh3, kvxrqtnz, tiqiwz3j; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 714.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 714.1 50.9 66.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d0ib6lbe",
   "source_url": "https://omlx.ai/benchmarks/performance/d0ib6lbe",
   "tg_tokens": null,
   "tps": 50.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-d0ib6lbe-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.0 GB at this context; duplicate traces of the same session: 919dis83, bzhe0vh3, kvxrqtnz, tiqiwz3j; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 712.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 712.7 47.3 67.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d0ib6lbe",
   "source_url": "https://omlx.ai/benchmarks/performance/d0ib6lbe",
   "tg_tokens": null,
   "tps": 47.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-d0ib6lbe-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.7 GB at this context; duplicate traces of the same session: 919dis83, bzhe0vh3, kvxrqtnz, tiqiwz3j; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 670.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 16k 670.7 35.9 67.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d0ib6lbe",
   "source_url": "https://omlx.ai/benchmarks/performance/d0ib6lbe",
   "tg_tokens": null,
   "tps": 35.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-d0ib6lbe-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 69.1 GB at this context; duplicate traces of the same session: 919dis83, bzhe0vh3, kvxrqtnz, tiqiwz3j; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 587.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 587.5 24.4 69.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d0ib6lbe",
   "source_url": "https://omlx.ai/benchmarks/performance/d0ib6lbe",
   "tg_tokens": null,
   "tps": 24.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-d0ib6lbe-ctx65536",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 71.9 GB at this context; duplicate traces of the same session: 919dis83, bzhe0vh3, kvxrqtnz, tiqiwz3j; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 458.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k 458.6 15.5 71.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d0ib6lbe",
   "source_url": "https://omlx.ai/benchmarks/performance/d0ib6lbe",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-i32xq8me-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 123.2 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 863.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 1k 863.0 42.9 123.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i32xq8me",
   "source_url": "https://omlx.ai/benchmarks/performance/i32xq8me",
   "tg_tokens": null,
   "tps": 42.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-i32xq8me-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 123.5 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 911.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 4k 911.3 41.4 123.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i32xq8me",
   "source_url": "https://omlx.ai/benchmarks/performance/i32xq8me",
   "tg_tokens": null,
   "tps": 41.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-cavk7p1r-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.3 GB at this context; duplicate traces of the same session: 43mgnpxt, lhzzunu0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 897.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 897.8 52.3 66.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cavk7p1r",
   "source_url": "https://omlx.ai/benchmarks/performance/cavk7p1r",
   "tg_tokens": null,
   "tps": 52.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-cavk7p1r-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.7 GB at this context; duplicate traces of the same session: 43mgnpxt, lhzzunu0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 935.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 935.9 47.9 66.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cavk7p1r",
   "source_url": "https://omlx.ai/benchmarks/performance/cavk7p1r",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-cavk7p1r-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.0 GB at this context; duplicate traces of the same session: 43mgnpxt, lhzzunu0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 920.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 920.4 47.4 67.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cavk7p1r",
   "source_url": "https://omlx.ai/benchmarks/performance/cavk7p1r",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-cavk7p1r-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.7 GB at this context; duplicate traces of the same session: 43mgnpxt, lhzzunu0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 862.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 862.3 43.1 67.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cavk7p1r",
   "source_url": "https://omlx.ai/benchmarks/performance/cavk7p1r",
   "tg_tokens": null,
   "tps": 43.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-cavk7p1r-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 69.1 GB at this context; duplicate traces of the same session: 43mgnpxt, lhzzunu0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 746.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 746.1 36.4 69.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cavk7p1r",
   "source_url": "https://omlx.ai/benchmarks/performance/cavk7p1r",
   "tg_tokens": null,
   "tps": 36.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-cavk7p1r-ctx65536",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 71.9 GB at this context; duplicate traces of the same session: 43mgnpxt, lhzzunu0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 570.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 570.2 27.7 71.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cavk7p1r",
   "source_url": "https://omlx.ai/benchmarks/performance/cavk7p1r",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-77dim65a-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 123.2 GB at this context; duplicate traces of the same session: ho7w5nty; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 837.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 837.2 39.1 123.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/77dim65a",
   "source_url": "https://omlx.ai/benchmarks/performance/77dim65a",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-77dim65a-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 123.5 GB at this context; duplicate traces of the same session: ho7w5nty; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 895.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 895.1 37.8 123.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/77dim65a",
   "source_url": "https://omlx.ai/benchmarks/performance/77dim65a",
   "tg_tokens": null,
   "tps": 37.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-77dim65a-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 123.9 GB at this context; duplicate traces of the same session: ho7w5nty; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 876.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 876.6 35.7 123.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/77dim65a",
   "source_url": "https://omlx.ai/benchmarks/performance/77dim65a",
   "tg_tokens": null,
   "tps": 35.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-77dim65a-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 124.5 GB at this context; duplicate traces of the same session: ho7w5nty; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 826.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k 826.4 32.5 124.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/77dim65a",
   "source_url": "https://omlx.ai/benchmarks/performance/77dim65a",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-77dim65a-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 125.9 GB at this context; duplicate traces of the same session: ho7w5nty; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 717.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 32k 717.8 27.9 125.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/77dim65a",
   "source_url": "https://omlx.ai/benchmarks/performance/77dim65a",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ko5y3ob0-ctx1024",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: lrmb0sb2; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 68.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 1k 68.0 15.4 16.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ko5y3ob0",
   "source_url": "https://omlx.ai/benchmarks/performance/ko5y3ob0",
   "tg_tokens": null,
   "tps": 15.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ko5y3ob0-ctx4096",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: lrmb0sb2; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 68.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k 68.0 14.4 17.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ko5y3ob0",
   "source_url": "https://omlx.ai/benchmarks/performance/ko5y3ob0",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-8rwegtmb-ctx1024",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 5raey9l1, 5y2fk7fi, i0ji024h; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 226.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 1k 226.5 33.4 16.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8rwegtmb",
   "source_url": "https://omlx.ai/benchmarks/performance/8rwegtmb",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-8rwegtmb-ctx4096",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 5raey9l1, 5y2fk7fi, i0ji024h; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 228.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 4k 228.9 31.3 17.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8rwegtmb",
   "source_url": "https://omlx.ai/benchmarks/performance/8rwegtmb",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-8rwegtmb-ctx8192",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: 5raey9l1, 5y2fk7fi, i0ji024h; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 227.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 8k 227.0 29.3 17.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8rwegtmb",
   "source_url": "https://omlx.ai/benchmarks/performance/8rwegtmb",
   "tg_tokens": null,
   "tps": 29.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-8rwegtmb-ctx16384",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: 5raey9l1, 5y2fk7fi, i0ji024h; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 222.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 16k 222.3 26.6 18.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8rwegtmb",
   "source_url": "https://omlx.ai/benchmarks/performance/8rwegtmb",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-8rwegtmb-ctx32768",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 5raey9l1, 5y2fk7fi, i0ji024h; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 212.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 32k 212.9 21.9 20.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8rwegtmb",
   "source_url": "https://omlx.ai/benchmarks/performance/8rwegtmb",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-8rwegtmb-ctx65536",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 5raey9l1, 5y2fk7fi, i0ji024h; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 192.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 64k 192.2 16.8 23.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8rwegtmb",
   "source_url": "https://omlx.ai/benchmarks/performance/8rwegtmb",
   "tg_tokens": null,
   "tps": 16.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ah2c9b47-ctx1024",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: iff2zzcf, jkbcbaj9, w3gm3t09; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 322.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 322.7 38.0 16.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ah2c9b47",
   "source_url": "https://omlx.ai/benchmarks/performance/ah2c9b47",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ah2c9b47-ctx4096",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: iff2zzcf, jkbcbaj9, w3gm3t09; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 324.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 324.2 35.3 17.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ah2c9b47",
   "source_url": "https://omlx.ai/benchmarks/performance/ah2c9b47",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ah2c9b47-ctx8192",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: iff2zzcf, jkbcbaj9, w3gm3t09; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 319.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 319.7 33.6 17.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ah2c9b47",
   "source_url": "https://omlx.ai/benchmarks/performance/ah2c9b47",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ah2c9b47-ctx16384",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: iff2zzcf, jkbcbaj9, w3gm3t09; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 310.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 310.0 30.0 18.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ah2c9b47",
   "source_url": "https://omlx.ai/benchmarks/performance/ah2c9b47",
   "tg_tokens": null,
   "tps": 30.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ah2c9b47-ctx32768",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: iff2zzcf, jkbcbaj9, w3gm3t09; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 291.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 291.3 24.9 20.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ah2c9b47",
   "source_url": "https://omlx.ai/benchmarks/performance/ah2c9b47",
   "tg_tokens": null,
   "tps": 24.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ah2c9b47-ctx65536",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: iff2zzcf, jkbcbaj9, w3gm3t09; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 258.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 258.7 18.1 23.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ah2c9b47",
   "source_url": "https://omlx.ai/benchmarks/performance/ah2c9b47",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-h0s1wf79-ctx1024",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: j3mx6mkn, nktcknsh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 239.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 239.4 27.6 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h0s1wf79",
   "source_url": "https://omlx.ai/benchmarks/performance/h0s1wf79",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-h0s1wf79-ctx4096",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: j3mx6mkn, nktcknsh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 199.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 199.2 23.6 17.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h0s1wf79",
   "source_url": "https://omlx.ai/benchmarks/performance/h0s1wf79",
   "tg_tokens": null,
   "tps": 23.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-h0s1wf79-ctx8192",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: j3mx6mkn, nktcknsh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 186.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 186.4 20.3 17.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h0s1wf79",
   "source_url": "https://omlx.ai/benchmarks/performance/h0s1wf79",
   "tg_tokens": null,
   "tps": 20.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-76f7vm8m-ctx1024",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: iugrfis1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 375.2,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 1k 375.2 26.7 22.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/76f7vm8m",
   "source_url": "https://omlx.ai/benchmarks/performance/76f7vm8m",
   "tg_tokens": null,
   "tps": 26.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-76f7vm8m-ctx4096",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: iugrfis1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 401.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 4k 401.1 25.4 22.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/76f7vm8m",
   "source_url": "https://omlx.ai/benchmarks/performance/76f7vm8m",
   "tg_tokens": null,
   "tps": 25.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-76f7vm8m-ctx8192",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: iugrfis1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 400.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 8k 400.8 24.4 22.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/76f7vm8m",
   "source_url": "https://omlx.ai/benchmarks/performance/76f7vm8m",
   "tg_tokens": null,
   "tps": 24.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-76f7vm8m-ctx16384",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: iugrfis1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 392.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 16k 392.6 22.4 23.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/76f7vm8m",
   "source_url": "https://omlx.ai/benchmarks/performance/76f7vm8m",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-76f7vm8m-ctx32768",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: iugrfis1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 370.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 32k 370.0 19.5 25.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/76f7vm8m",
   "source_url": "https://omlx.ai/benchmarks/performance/76f7vm8m",
   "tg_tokens": null,
   "tps": 19.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-of4pm30v-ctx1024",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 238.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 238.7 28.8 15.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/of4pm30v",
   "source_url": "https://omlx.ai/benchmarks/performance/of4pm30v",
   "tg_tokens": null,
   "tps": 28.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-of4pm30v-ctx4096",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 243.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 243.8 27.6 16.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/of4pm30v",
   "source_url": "https://omlx.ai/benchmarks/performance/of4pm30v",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-byclik12-ctx1024",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 540.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 540.2 63.7 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/byclik12",
   "source_url": "https://omlx.ai/benchmarks/performance/byclik12",
   "tg_tokens": null,
   "tps": 63.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-byclik12-ctx4096",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 585.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 585.0 60.1 20.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/byclik12",
   "source_url": "https://omlx.ai/benchmarks/performance/byclik12",
   "tg_tokens": null,
   "tps": 60.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-25m3innk-ctx1024",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 524.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-35B-A3B on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 1k 524.7 50.4 28.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/25m3innk",
   "source_url": "https://omlx.ai/benchmarks/performance/25m3innk",
   "tg_tokens": null,
   "tps": 50.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-25m3innk-ctx4096",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.3 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 558.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-35B-A3B on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 4k 558.7 49.0 28.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/25m3innk",
   "source_url": "https://omlx.ai/benchmarks/performance/25m3innk",
   "tg_tokens": null,
   "tps": 49.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-x8rtp3tm-ctx1024",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: rpqawcit; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1361.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,361 95.7 20.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x8rtp3tm",
   "source_url": "https://omlx.ai/benchmarks/performance/x8rtp3tm",
   "tg_tokens": null,
   "tps": 95.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-x8rtp3tm-ctx4096",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: rpqawcit; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1067.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,067 95.8 20.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x8rtp3tm",
   "source_url": "https://omlx.ai/benchmarks/performance/x8rtp3tm",
   "tg_tokens": null,
   "tps": 95.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-x8rtp3tm-ctx8192",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: rpqawcit; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1091.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,091 90.5 20.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x8rtp3tm",
   "source_url": "https://omlx.ai/benchmarks/performance/x8rtp3tm",
   "tg_tokens": null,
   "tps": 90.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-jsai2ted-ctx1024",
   "model": "Qwen3.5-35B-A3B-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 5idwaj24; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 418.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B-Heretic on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 1k 418.3 55.5 20.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jsai2ted",
   "source_url": "https://omlx.ai/benchmarks/performance/jsai2ted",
   "tg_tokens": null,
   "tps": 55.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-jsai2ted-ctx4096",
   "model": "Qwen3.5-35B-A3B-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 5idwaj24; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 439.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B-Heretic on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k 439.0 49.0 20.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jsai2ted",
   "source_url": "https://omlx.ai/benchmarks/performance/jsai2ted",
   "tg_tokens": null,
   "tps": 49.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vive2iyg-ctx1024",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 210.4 GB at this context; duplicate traces of the same session: 4o3fa12n, 6xmkfdgf, k0jw9l2d; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 381.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 381.0 37.0 210.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vive2iyg",
   "source_url": "https://omlx.ai/benchmarks/performance/vive2iyg",
   "tg_tokens": null,
   "tps": 37.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vive2iyg-ctx4096",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 210.7 GB at this context; duplicate traces of the same session: 4o3fa12n, 6xmkfdgf, k0jw9l2d; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 404.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 404.5 35.3 210.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vive2iyg",
   "source_url": "https://omlx.ai/benchmarks/performance/vive2iyg",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vive2iyg-ctx8192",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 211.0 GB at this context; duplicate traces of the same session: 4o3fa12n, 6xmkfdgf, k0jw9l2d; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 400.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 400.7 31.3 211.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vive2iyg",
   "source_url": "https://omlx.ai/benchmarks/performance/vive2iyg",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vive2iyg-ctx16384",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 211.8 GB at this context; duplicate traces of the same session: 4o3fa12n, 6xmkfdgf, k0jw9l2d; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 384.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 384.0 24.8 211.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vive2iyg",
   "source_url": "https://omlx.ai/benchmarks/performance/vive2iyg",
   "tg_tokens": null,
   "tps": 24.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vive2iyg-ctx32768",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 213.2 GB at this context; duplicate traces of the same session: 4o3fa12n, 6xmkfdgf, k0jw9l2d; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 350.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 350.4 17.8 213.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vive2iyg",
   "source_url": "https://omlx.ai/benchmarks/performance/vive2iyg",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vive2iyg-ctx65536",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 216.2 GB at this context; duplicate traces of the same session: 4o3fa12n, 6xmkfdgf, k0jw9l2d; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 294.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 294.3 11.7 216.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vive2iyg",
   "source_url": "https://omlx.ai/benchmarks/performance/vive2iyg",
   "tg_tokens": null,
   "tps": 11.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-q66jpzr6-ctx1024",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 302.7 GB at this context; duplicate traces of the same session: rb7hx514; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 454.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 1k 454.0 28.5 302.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q66jpzr6",
   "source_url": "https://omlx.ai/benchmarks/performance/q66jpzr6",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-q66jpzr6-ctx4096",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 302.9 GB at this context; duplicate traces of the same session: rb7hx514; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 487.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 4k 487.1 27.2 302.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q66jpzr6",
   "source_url": "https://omlx.ai/benchmarks/performance/q66jpzr6",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-q66jpzr6-ctx8192",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 303.3 GB at this context; duplicate traces of the same session: rb7hx514; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 484.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 8k 484.9 26.0 303.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q66jpzr6",
   "source_url": "https://omlx.ai/benchmarks/performance/q66jpzr6",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-q66jpzr6-ctx16384",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 304.0 GB at this context; duplicate traces of the same session: rb7hx514; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 462.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 16k 462.0 23.6 304.0 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q66jpzr6",
   "source_url": "https://omlx.ai/benchmarks/performance/q66jpzr6",
   "tg_tokens": null,
   "tps": 23.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-q66jpzr6-ctx32768",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 305.5 GB at this context; duplicate traces of the same session: rb7hx514; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 422.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 32k 422.5 20.2 305.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q66jpzr6",
   "source_url": "https://omlx.ai/benchmarks/performance/q66jpzr6",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-z60c2xaw-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.0 GB at this context; duplicate traces of the same session: 8owje5g3; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 292.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 292.6 53.8 4.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z60c2xaw",
   "source_url": "https://omlx.ai/benchmarks/performance/z60c2xaw",
   "tg_tokens": null,
   "tps": 53.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-z60c2xaw-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: 8owje5g3; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 292.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 292.5 49.3 4.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z60c2xaw",
   "source_url": "https://omlx.ai/benchmarks/performance/z60c2xaw",
   "tg_tokens": null,
   "tps": 49.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-3s2owzjd-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.0 GB at this context; duplicate traces of the same session: 0oga1cf0, j290pxfi, qni6te2s, xckibqvi, zwuvz3cp; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1428.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 1k 1,428 137.5 4.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3s2owzjd",
   "source_url": "https://omlx.ai/benchmarks/performance/3s2owzjd",
   "tg_tokens": null,
   "tps": 137.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-3s2owzjd-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: 0oga1cf0, j290pxfi, qni6te2s, xckibqvi, zwuvz3cp; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1440.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 4k 1,440 128.2 4.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3s2owzjd",
   "source_url": "https://omlx.ai/benchmarks/performance/3s2owzjd",
   "tg_tokens": null,
   "tps": 128.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-3s2owzjd-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.6 GB at this context; duplicate traces of the same session: 0oga1cf0, j290pxfi, qni6te2s, xckibqvi, zwuvz3cp; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1414.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 8k 1,414 117.4 4.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3s2owzjd",
   "source_url": "https://omlx.ai/benchmarks/performance/3s2owzjd",
   "tg_tokens": null,
   "tps": 117.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-3s2owzjd-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 5.3 GB at this context; duplicate traces of the same session: 0oga1cf0, j290pxfi, qni6te2s, xckibqvi, zwuvz3cp; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1355.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 16k 1,355 98.7 5.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3s2owzjd",
   "source_url": "https://omlx.ai/benchmarks/performance/3s2owzjd",
   "tg_tokens": null,
   "tps": 98.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-3s2owzjd-ctx32768",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: 0oga1cf0, j290pxfi, qni6te2s, xckibqvi, zwuvz3cp; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1247.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 32k 1,247 74.5 6.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3s2owzjd",
   "source_url": "https://omlx.ai/benchmarks/performance/3s2owzjd",
   "tg_tokens": null,
   "tps": 74.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-3s2owzjd-ctx65536",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 8.6 GB at this context; duplicate traces of the same session: 0oga1cf0, j290pxfi, qni6te2s, xckibqvi, zwuvz3cp; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1051.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 64k 1,051 50.4 8.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3s2owzjd",
   "source_url": "https://omlx.ai/benchmarks/performance/3s2owzjd",
   "tg_tokens": null,
   "tps": 50.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-ufzud5nf-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: 4vf3fhkn, 5t6c2eu1, cgvsvi3i, p2mm6l9j, u49qccjp; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1423.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 1k 1,423 99.9 5.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ufzud5nf",
   "source_url": "https://omlx.ai/benchmarks/performance/ufzud5nf",
   "tg_tokens": null,
   "tps": 99.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-ufzud5nf-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 6.1 GB at this context; duplicate traces of the same session: 4vf3fhkn, 5t6c2eu1, cgvsvi3i, p2mm6l9j, u49qccjp; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1448.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 4k 1,448 96.2 6.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ufzud5nf",
   "source_url": "https://omlx.ai/benchmarks/performance/ufzud5nf",
   "tg_tokens": null,
   "tps": 96.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-ufzud5nf-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: 4vf3fhkn, 5t6c2eu1, cgvsvi3i, p2mm6l9j, u49qccjp; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1421.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 8k 1,421 90.1 6.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ufzud5nf",
   "source_url": "https://omlx.ai/benchmarks/performance/ufzud5nf",
   "tg_tokens": null,
   "tps": 90.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-ufzud5nf-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 7.3 GB at this context; duplicate traces of the same session: 4vf3fhkn, 5t6c2eu1, cgvsvi3i, p2mm6l9j, u49qccjp; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1362.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 16k 1,362 80.1 7.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ufzud5nf",
   "source_url": "https://omlx.ai/benchmarks/performance/ufzud5nf",
   "tg_tokens": null,
   "tps": 80.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-ufzud5nf-ctx32768",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 7.9 GB at this context; duplicate traces of the same session: 4vf3fhkn, 5t6c2eu1, cgvsvi3i, p2mm6l9j, u49qccjp; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1255.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 32k 1,255 63.0 7.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ufzud5nf",
   "source_url": "https://omlx.ai/benchmarks/performance/ufzud5nf",
   "tg_tokens": null,
   "tps": 63.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-ufzud5nf-ctx65536",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 9.9 GB at this context; duplicate traces of the same session: 4vf3fhkn, 5t6c2eu1, cgvsvi3i, p2mm6l9j, u49qccjp; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1056.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 64k 1,056 45.5 9.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ufzud5nf",
   "source_url": "https://omlx.ai/benchmarks/performance/ufzud5nf",
   "tg_tokens": null,
   "tps": 45.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-yyrehuqw-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 338.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 338.1 31.9 3.9 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yyrehuqw",
   "source_url": "https://omlx.ai/benchmarks/performance/yyrehuqw",
   "tg_tokens": null,
   "tps": 31.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-yyrehuqw-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 327.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 327.4 29.8 4.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yyrehuqw",
   "source_url": "https://omlx.ai/benchmarks/performance/yyrehuqw",
   "tg_tokens": null,
   "tps": 29.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-rt8jg3e5-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 3.9 GB; TTFT 2558 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 400.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k tokens 400.3 PP tok/s 38.8 TG tok/s 2558 TTFT (ms) 3.9 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rt8jg3e5",
   "source_url": "https://omlx.ai/benchmarks/performance/rt8jg3e5",
   "tg_tokens": null,
   "tps": 38.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-ip9k1jp7-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: a6wnmd9u; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 378.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 378.3 35.1 3.9 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ip9k1jp7",
   "source_url": "https://omlx.ai/benchmarks/performance/ip9k1jp7",
   "tg_tokens": null,
   "tps": 35.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-ip9k1jp7-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: a6wnmd9u; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 380.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 380.6 34.1 4.2 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ip9k1jp7",
   "source_url": "https://omlx.ai/benchmarks/performance/ip9k1jp7",
   "tg_tokens": null,
   "tps": 34.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-zx6o7kdc-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.7 GB at this context; duplicate traces of the same session: 4umrbkhz; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 239.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 8bit; 1k 239.0 35.3 10.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zx6o7kdc",
   "source_url": "https://omlx.ai/benchmarks/performance/zx6o7kdc",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-zx6o7kdc-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.0 GB at this context; duplicate traces of the same session: 4umrbkhz; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 241.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 8bit; 4k 241.2 33.4 11.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zx6o7kdc",
   "source_url": "https://omlx.ai/benchmarks/performance/zx6o7kdc",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-cwtnpgxy-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.7 GB at this context; duplicate traces of the same session: 1n8qgmq9, 5fjbahy5, 6rngmu2d, hyclrwcc, ob8t8tj4; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 239.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 8bit; 1k 239.3 35.1 10.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cwtnpgxy",
   "source_url": "https://omlx.ai/benchmarks/performance/cwtnpgxy",
   "tg_tokens": null,
   "tps": 35.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-cwtnpgxy-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.0 GB at this context; duplicate traces of the same session: 1n8qgmq9, 5fjbahy5, 6rngmu2d, hyclrwcc, ob8t8tj4; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 242.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 8bit; 4k 242.1 33.3 11.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cwtnpgxy",
   "source_url": "https://omlx.ai/benchmarks/performance/cwtnpgxy",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-cwtnpgxy-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.2 GB at this context; duplicate traces of the same session: 1n8qgmq9, 5fjbahy5, 6rngmu2d, hyclrwcc, ob8t8tj4; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 236.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 8bit; 8k 236.2 30.3 11.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cwtnpgxy",
   "source_url": "https://omlx.ai/benchmarks/performance/cwtnpgxy",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-cwtnpgxy-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.7 GB at this context; duplicate traces of the same session: 1n8qgmq9, 5fjbahy5, 6rngmu2d, hyclrwcc, ob8t8tj4; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 234.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 8bit; 16k 234.4 26.2 11.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cwtnpgxy",
   "source_url": "https://omlx.ai/benchmarks/performance/cwtnpgxy",
   "tg_tokens": null,
   "tps": 26.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-cwtnpgxy-ctx32768",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: 1n8qgmq9, 5fjbahy5, 6rngmu2d, hyclrwcc, ob8t8tj4; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 8bit; 32k 224.8 20.3 12.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cwtnpgxy",
   "source_url": "https://omlx.ai/benchmarks/performance/cwtnpgxy",
   "tg_tokens": null,
   "tps": 20.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-cwtnpgxy-ctx65536",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: 1n8qgmq9, 5fjbahy5, 6rngmu2d, hyclrwcc, ob8t8tj4; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 8bit; 64k 205.0 13.9 14.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cwtnpgxy",
   "source_url": "https://omlx.ai/benchmarks/performance/cwtnpgxy",
   "tg_tokens": null,
   "tps": 13.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-6b5xaxkg-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 312.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 1k 312.5 56.6 6.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6b5xaxkg",
   "source_url": "https://omlx.ai/benchmarks/performance/6b5xaxkg",
   "tg_tokens": null,
   "tps": 56.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-6b5xaxkg-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 317.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 4k 317.4 53.8 7.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6b5xaxkg",
   "source_url": "https://omlx.ai/benchmarks/performance/6b5xaxkg",
   "tg_tokens": null,
   "tps": 53.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-4tl4ubsq-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.7 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 304.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 1k 304.3 35.6 10.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4tl4ubsq",
   "source_url": "https://omlx.ai/benchmarks/performance/4tl4ubsq",
   "tg_tokens": null,
   "tps": 35.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-4tl4ubsq-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.0 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 319.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k 319.1 34.7 11.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4tl4ubsq",
   "source_url": "https://omlx.ai/benchmarks/performance/4tl4ubsq",
   "tg_tokens": null,
   "tps": 34.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-alvdhbe3-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.7 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 316.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 1k 316.3 35.9 10.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/alvdhbe3",
   "source_url": "https://omlx.ai/benchmarks/performance/alvdhbe3",
   "tg_tokens": null,
   "tps": 35.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-alvdhbe3-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.0 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 321.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k 321.1 34.7 11.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/alvdhbe3",
   "source_url": "https://omlx.ai/benchmarks/performance/alvdhbe3",
   "tg_tokens": null,
   "tps": 34.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-ajlpo1sp-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: iha3k32k; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 158.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 158.6 32.3 6.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ajlpo1sp",
   "source_url": "https://omlx.ai/benchmarks/performance/ajlpo1sp",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-ajlpo1sp-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: iha3k32k; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 159.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 159.3 31.5 7.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ajlpo1sp",
   "source_url": "https://omlx.ai/benchmarks/performance/ajlpo1sp",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-gnjkbfpd-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.7 GB at this context; duplicate traces of the same session: a6fmmdir, t8kof1zr, xpnt88fb; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 775.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 1k 775.8 74.3 8.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gnjkbfpd",
   "source_url": "https://omlx.ai/benchmarks/performance/gnjkbfpd",
   "tg_tokens": null,
   "tps": 74.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-gnjkbfpd-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.9 GB at this context; duplicate traces of the same session: a6fmmdir, t8kof1zr, xpnt88fb; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 783.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 4k 783.9 72.1 8.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gnjkbfpd",
   "source_url": "https://omlx.ai/benchmarks/performance/gnjkbfpd",
   "tg_tokens": null,
   "tps": 72.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-gnjkbfpd-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 9.2 GB at this context; duplicate traces of the same session: a6fmmdir, t8kof1zr, xpnt88fb; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 776.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 8k 776.7 68.7 9.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gnjkbfpd",
   "source_url": "https://omlx.ai/benchmarks/performance/gnjkbfpd",
   "tg_tokens": null,
   "tps": 68.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-gnjkbfpd-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 9.6 GB at this context; duplicate traces of the same session: a6fmmdir, t8kof1zr, xpnt88fb; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 758.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 16k 758.8 62.8 9.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gnjkbfpd",
   "source_url": "https://omlx.ai/benchmarks/performance/gnjkbfpd",
   "tg_tokens": null,
   "tps": 62.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-gnjkbfpd-ctx32768",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.6 GB at this context; duplicate traces of the same session: a6fmmdir, t8kof1zr, xpnt88fb; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 724.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 32k 724.1 52.6 10.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gnjkbfpd",
   "source_url": "https://omlx.ai/benchmarks/performance/gnjkbfpd",
   "tg_tokens": null,
   "tps": 52.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-gnjkbfpd-ctx65536",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: a6fmmdir, t8kof1zr, xpnt88fb; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 653.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 64k 653.8 39.8 12.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gnjkbfpd",
   "source_url": "https://omlx.ai/benchmarks/performance/gnjkbfpd",
   "tg_tokens": null,
   "tps": 39.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-ksocjl8g-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.7 GB at this context; duplicate traces of the same session: 6o6zm8tl, 8g093y7c, mm823kic, voj8az1w, wxd1vexu; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 794.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 1k 794.0 64.3 10.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ksocjl8g",
   "source_url": "https://omlx.ai/benchmarks/performance/ksocjl8g",
   "tg_tokens": null,
   "tps": 64.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-ksocjl8g-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.0 GB at this context; duplicate traces of the same session: 6o6zm8tl, 8g093y7c, mm823kic, voj8az1w, wxd1vexu; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 804.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 4k 804.0 62.6 11.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ksocjl8g",
   "source_url": "https://omlx.ai/benchmarks/performance/ksocjl8g",
   "tg_tokens": null,
   "tps": 62.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-ksocjl8g-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.2 GB at this context; duplicate traces of the same session: 6o6zm8tl, 8g093y7c, mm823kic, voj8az1w, wxd1vexu; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 797.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 8k 797.4 59.7 11.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ksocjl8g",
   "source_url": "https://omlx.ai/benchmarks/performance/ksocjl8g",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-ksocjl8g-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.7 GB at this context; duplicate traces of the same session: 6o6zm8tl, 8g093y7c, mm823kic, voj8az1w, wxd1vexu; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 778.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 16k 778.7 54.0 11.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ksocjl8g",
   "source_url": "https://omlx.ai/benchmarks/performance/ksocjl8g",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-ksocjl8g-ctx32768",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: 6o6zm8tl, 8g093y7c, mm823kic, voj8az1w, wxd1vexu; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 742.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 32k 742.0 46.4 12.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ksocjl8g",
   "source_url": "https://omlx.ai/benchmarks/performance/ksocjl8g",
   "tg_tokens": null,
   "tps": 46.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-ksocjl8g-ctx65536",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: 6o6zm8tl, 8g093y7c, mm823kic, voj8az1w, wxd1vexu; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 668.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 64k 668.6 33.9 14.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ksocjl8g",
   "source_url": "https://omlx.ai/benchmarks/performance/ksocjl8g",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-lagqqc8e-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 223.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 223.7 19.6 6.5 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lagqqc8e",
   "source_url": "https://omlx.ai/benchmarks/performance/lagqqc8e",
   "tg_tokens": null,
   "tps": 19.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-lagqqc8e-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 224.8 20.7 6.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lagqqc8e",
   "source_url": "https://omlx.ai/benchmarks/performance/lagqqc8e",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-uafhftej-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 225.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 225.0 21.7 6.5 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uafhftej",
   "source_url": "https://omlx.ai/benchmarks/performance/uafhftej",
   "tg_tokens": null,
   "tps": 21.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-uafhftej-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 225.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 225.4 21.0 6.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uafhftej",
   "source_url": "https://omlx.ai/benchmarks/performance/uafhftej",
   "tg_tokens": null,
   "tps": 21.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-swdphr5s-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 218.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 218.1 19.8 6.5 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/swdphr5s",
   "source_url": "https://omlx.ai/benchmarks/performance/swdphr5s",
   "tg_tokens": null,
   "tps": 19.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-swdphr5s-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 219.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 219.1 19.9 6.8 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/swdphr5s",
   "source_url": "https://omlx.ai/benchmarks/performance/swdphr5s",
   "tg_tokens": null,
   "tps": 19.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-nhr893lp-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.7 GB at this context; duplicate traces of the same session: hqgt9fac; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 878.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 878.9 52.1 10.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nhr893lp",
   "source_url": "https://omlx.ai/benchmarks/performance/nhr893lp",
   "tg_tokens": null,
   "tps": 52.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-nhr893lp-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.0 GB at this context; duplicate traces of the same session: hqgt9fac; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 890.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 890.4 50.9 11.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nhr893lp",
   "source_url": "https://omlx.ai/benchmarks/performance/nhr893lp",
   "tg_tokens": null,
   "tps": 50.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-vwdm65mw-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: qe1o4j3u; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 445.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 445.0 48.9 6.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vwdm65mw",
   "source_url": "https://omlx.ai/benchmarks/performance/vwdm65mw",
   "tg_tokens": null,
   "tps": 48.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-vwdm65mw-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: qe1o4j3u; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 447.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 447.3 47.6 7.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vwdm65mw",
   "source_url": "https://omlx.ai/benchmarks/performance/vwdm65mw",
   "tg_tokens": null,
   "tps": 47.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-y79kne58-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: wufyjzmw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 928.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 928.8 33.2 16.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y79kne58",
   "source_url": "https://omlx.ai/benchmarks/performance/y79kne58",
   "tg_tokens": null,
   "tps": 33.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-y79kne58-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: wufyjzmw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1008.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,008 32.8 17.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y79kne58",
   "source_url": "https://omlx.ai/benchmarks/performance/y79kne58",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-y79kne58-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: wufyjzmw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 878.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 878.4 32.0 17.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y79kne58",
   "source_url": "https://omlx.ai/benchmarks/performance/y79kne58",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-y79kne58-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: wufyjzmw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 797.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 797.7 30.8 17.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y79kne58",
   "source_url": "https://omlx.ai/benchmarks/performance/y79kne58",
   "tg_tokens": null,
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5h3m9gbn-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: bqy4j59f, hhxc5sho, j0dficgz, nme46ce0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 893.6,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 1k 893.6 38.2 13.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5h3m9gbn",
   "source_url": "https://omlx.ai/benchmarks/performance/5h3m9gbn",
   "tg_tokens": null,
   "tps": 38.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5h3m9gbn-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: bqy4j59f, hhxc5sho, j0dficgz, nme46ce0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 935.7,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 4k 935.7 38.0 14.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5h3m9gbn",
   "source_url": "https://omlx.ai/benchmarks/performance/5h3m9gbn",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5h3m9gbn-ctx8192",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: bqy4j59f, hhxc5sho, j0dficgz, nme46ce0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 852.6,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 8k 852.6 35.3 14.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5h3m9gbn",
   "source_url": "https://omlx.ai/benchmarks/performance/5h3m9gbn",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5h3m9gbn-ctx16384",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 15.3 GB at this context; duplicate traces of the same session: bqy4j59f, hhxc5sho, j0dficgz, nme46ce0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 735.4,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 16k 735.4 32.5 15.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5h3m9gbn",
   "source_url": "https://omlx.ai/benchmarks/performance/5h3m9gbn",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5h3m9gbn-ctx32768",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: bqy4j59f, hhxc5sho, j0dficgz, nme46ce0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 699.2,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 32k 699.2 32.1 16.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5h3m9gbn",
   "source_url": "https://omlx.ai/benchmarks/performance/5h3m9gbn",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5h3m9gbn-ctx65536",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: bqy4j59f, hhxc5sho, j0dficgz, nme46ce0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 680.6,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 64k 680.6 27.9 18.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5h3m9gbn",
   "source_url": "https://omlx.ai/benchmarks/performance/5h3m9gbn",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5h3m9gbn-ctx131072",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: bqy4j59f, hhxc5sho, j0dficgz, nme46ce0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 549.4,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 128k 549.4 21.5 22.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5h3m9gbn",
   "source_url": "https://omlx.ai/benchmarks/performance/5h3m9gbn",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-chonyuwu-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 327.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 327.3 23.1 28.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/chonyuwu",
   "source_url": "https://omlx.ai/benchmarks/performance/chonyuwu",
   "tg_tokens": null,
   "tps": 23.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-chonyuwu-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.2 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 337.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 337.7 22.8 32.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/chonyuwu",
   "source_url": "https://omlx.ai/benchmarks/performance/chonyuwu",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-chonyuwu-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.2 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 332.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 332.6 22.2 33.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/chonyuwu",
   "source_url": "https://omlx.ai/benchmarks/performance/chonyuwu",
   "tg_tokens": null,
   "tps": 22.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-chonyuwu-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.4 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 321.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 16k 321.9 21.7 32.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/chonyuwu",
   "source_url": "https://omlx.ai/benchmarks/performance/chonyuwu",
   "tg_tokens": null,
   "tps": 21.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-g8yms4n0-ctx65536",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.2 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 854.9,
   "provenance": "community",
   "quant": "mxfp4+mxfp8",
   "quote": "Qwen3.8-Flash-Next-Uncensored-MLX-mixed-mxfp4_mxfp8_mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 854.9 32.2 103.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g8yms4n0",
   "source_url": "https://omlx.ai/benchmarks/performance/g8yms4n0",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-g8yms4n0-ctx131072",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.0 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 818.2,
   "provenance": "community",
   "quant": "mxfp4+mxfp8",
   "quote": "Qwen3.8-Flash-Next-Uncensored-MLX-mixed-mxfp4_mxfp8_mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 818.2 37.9 110.0 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g8yms4n0",
   "source_url": "https://omlx.ai/benchmarks/performance/g8yms4n0",
   "tg_tokens": null,
   "tps": 37.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-g8yms4n0-ctx199680",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 120.0 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 780.0,
   "provenance": "community",
   "quant": "mxfp4+mxfp8",
   "quote": "Qwen3.8-Flash-Next-Uncensored-MLX-mixed-mxfp4_mxfp8_mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 195k 780.0 37.4 120.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g8yms4n0",
   "source_url": "https://omlx.ai/benchmarks/performance/g8yms4n0",
   "tg_tokens": null,
   "tps": 37.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-re3tiqd6-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 525.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 525.3 56.3 71.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/re3tiqd6",
   "source_url": "https://omlx.ai/benchmarks/performance/re3tiqd6",
   "tg_tokens": null,
   "tps": 56.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-re3tiqd6-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 592.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 592.8 53.4 71.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/re3tiqd6",
   "source_url": "https://omlx.ai/benchmarks/performance/re3tiqd6",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-re3tiqd6-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 465.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 465.7 35.0 71.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/re3tiqd6",
   "source_url": "https://omlx.ai/benchmarks/performance/re3tiqd6",
   "tg_tokens": null,
   "tps": 35.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-i9i67eqs-ctx1024",
   "model": "Gemma 3 4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 3.6 GB at this context; duplicate traces of the same session: fra5t4az; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 901.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-3-4b-it-qat on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 901.8 34.9 3.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i9i67eqs",
   "source_url": "https://omlx.ai/benchmarks/performance/i9i67eqs",
   "tg_tokens": null,
   "tps": 34.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-i9i67eqs-ctx4096",
   "model": "Gemma 3 4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: fra5t4az; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 889.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-3-4b-it-qat on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 889.9 80.2 3.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i9i67eqs",
   "source_url": "https://omlx.ai/benchmarks/performance/i9i67eqs",
   "tg_tokens": null,
   "tps": 80.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-3gd322b3-ctx1024",
   "model": "gpt-oss 120B",
   "notes": "oMLX community benchmark trace; peak memory 58.0 GB at this context; duplicate traces of the same session: nqo8pz28; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 187.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gpt-oss-120b on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 187.7 54.9 58.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3gd322b3",
   "source_url": "https://omlx.ai/benchmarks/performance/3gd322b3",
   "tg_tokens": null,
   "tps": 54.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-3gd322b3-ctx4096",
   "model": "gpt-oss 120B",
   "notes": "oMLX community benchmark trace; peak memory 58.1 GB at this context; duplicate traces of the same session: nqo8pz28; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 182.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gpt-oss-120b on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 182.0 44.7 58.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3gd322b3",
   "source_url": "https://omlx.ai/benchmarks/performance/3gd322b3",
   "tg_tokens": null,
   "tps": 44.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-3gd322b3-ctx8192",
   "model": "gpt-oss 120B",
   "notes": "oMLX community benchmark trace; peak memory 58.3 GB at this context; duplicate traces of the same session: nqo8pz28; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 183.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gpt-oss-120b on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 183.9 37.1 58.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3gd322b3",
   "source_url": "https://omlx.ai/benchmarks/performance/3gd322b3",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-3gd322b3-ctx16384",
   "model": "gpt-oss 120B",
   "notes": "oMLX community benchmark trace; peak memory 58.6 GB at this context; duplicate traces of the same session: nqo8pz28; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 177.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gpt-oss-120b on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 177.1 22.7 58.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3gd322b3",
   "source_url": "https://omlx.ai/benchmarks/performance/3gd322b3",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-3gd322b3-ctx32768",
   "model": "gpt-oss 120B",
   "notes": "oMLX community benchmark trace; peak memory 59.2 GB at this context; duplicate traces of the same session: nqo8pz28; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 165.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gpt-oss-120b on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 32k 165.3 14.6 59.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3gd322b3",
   "source_url": "https://omlx.ai/benchmarks/performance/3gd322b3",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-i54z07a7-ctx1024",
   "model": "gpt-oss 120B",
   "notes": "MXFP4 experts with Q8 attention; oMLX community benchmark trace; peak memory 59.2 GB at this context; duplicate traces of the same session: wjn65xp9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 613.2,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-120b-MXFP4-Q8 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 613.2 68.4 59.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i54z07a7",
   "source_url": "https://omlx.ai/benchmarks/performance/i54z07a7",
   "tg_tokens": null,
   "tps": 68.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-i54z07a7-ctx4096",
   "model": "gpt-oss 120B",
   "notes": "MXFP4 experts with Q8 attention; oMLX community benchmark trace; peak memory 59.3 GB at this context; duplicate traces of the same session: wjn65xp9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 685.5,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-120b-MXFP4-Q8 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 685.5 64.1 59.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i54z07a7",
   "source_url": "https://omlx.ai/benchmarks/performance/i54z07a7",
   "tg_tokens": null,
   "tps": 64.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-i54z07a7-ctx8192",
   "model": "gpt-oss 120B",
   "notes": "MXFP4 experts with Q8 attention; oMLX community benchmark trace; peak memory 59.5 GB at this context; duplicate traces of the same session: wjn65xp9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 687.8,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-120b-MXFP4-Q8 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 687.8 57.8 59.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i54z07a7",
   "source_url": "https://omlx.ai/benchmarks/performance/i54z07a7",
   "tg_tokens": null,
   "tps": 57.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-i54z07a7-ctx16384",
   "model": "gpt-oss 120B",
   "notes": "MXFP4 experts with Q8 attention; oMLX community benchmark trace; peak memory 59.8 GB at this context; duplicate traces of the same session: wjn65xp9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 663.6,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-120b-MXFP4-Q8 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 663.6 47.5 59.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i54z07a7",
   "source_url": "https://omlx.ai/benchmarks/performance/i54z07a7",
   "tg_tokens": null,
   "tps": 47.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-i54z07a7-ctx32768",
   "model": "gpt-oss 120B",
   "notes": "MXFP4 experts with Q8 attention; oMLX community benchmark trace; peak memory 60.4 GB at this context; duplicate traces of the same session: wjn65xp9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 600.7,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-120b-MXFP4-Q8 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 600.7 35.2 60.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i54z07a7",
   "source_url": "https://omlx.ai/benchmarks/performance/i54z07a7",
   "tg_tokens": null,
   "tps": 35.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ye45yn0j-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 105.9 15.6 18.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ye45yn0j",
   "source_url": "https://omlx.ai/benchmarks/performance/ye45yn0j",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ye45yn0j-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 151.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 151.3 15.0 19.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ye45yn0j",
   "source_url": "https://omlx.ai/benchmarks/performance/ye45yn0j",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ye45yn0j-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 115.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 115.0 14.5 19.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ye45yn0j",
   "source_url": "https://omlx.ai/benchmarks/performance/ye45yn0j",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ye45yn0j-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 96.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 96.1 12.8 19.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ye45yn0j",
   "source_url": "https://omlx.ai/benchmarks/performance/ye45yn0j",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-u28zoca4-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: c510mcpw; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 106.0 15.6 18.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u28zoca4",
   "source_url": "https://omlx.ai/benchmarks/performance/u28zoca4",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-u28zoca4-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: c510mcpw; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 151.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 151.0 15.0 19.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u28zoca4",
   "source_url": "https://omlx.ai/benchmarks/performance/u28zoca4",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-u28zoca4-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: c510mcpw; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 151.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 151.6 14.5 19.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u28zoca4",
   "source_url": "https://omlx.ai/benchmarks/performance/u28zoca4",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-u28zoca4-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: c510mcpw; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 131.6 12.8 21.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u28zoca4",
   "source_url": "https://omlx.ai/benchmarks/performance/u28zoca4",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-agmyufp1-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "pyros-vault MTP build; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: b1sluf0m, hj41ahwa, r2t9yutl, ziaa5bwx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 773.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "pyros-vault:Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 773.5 34.2 33.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/agmyufp1",
   "source_url": "https://omlx.ai/benchmarks/performance/agmyufp1",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-agmyufp1-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "pyros-vault MTP build; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: b1sluf0m, hj41ahwa, r2t9yutl, ziaa5bwx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 850.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "pyros-vault:Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 850.0 39.2 34.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/agmyufp1",
   "source_url": "https://omlx.ai/benchmarks/performance/agmyufp1",
   "tg_tokens": null,
   "tps": 39.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-agmyufp1-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "pyros-vault MTP build; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: b1sluf0m, hj41ahwa, r2t9yutl, ziaa5bwx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 823.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "pyros-vault:Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 823.6 50.2 35.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/agmyufp1",
   "source_url": "https://omlx.ai/benchmarks/performance/agmyufp1",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-agmyufp1-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "pyros-vault MTP build; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: b1sluf0m, hj41ahwa, r2t9yutl, ziaa5bwx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 752.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "pyros-vault:Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 752.8 53.2 35.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/agmyufp1",
   "source_url": "https://omlx.ai/benchmarks/performance/agmyufp1",
   "tg_tokens": null,
   "tps": 53.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-agmyufp1-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "pyros-vault MTP build; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: b1sluf0m, hj41ahwa, r2t9yutl, ziaa5bwx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 709.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "pyros-vault:Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 709.2 55.2 36.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/agmyufp1",
   "source_url": "https://omlx.ai/benchmarks/performance/agmyufp1",
   "tg_tokens": null,
   "tps": 55.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-agmyufp1-ctx65536",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "pyros-vault MTP build; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: b1sluf0m, hj41ahwa, r2t9yutl, ziaa5bwx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 617.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "pyros-vault:Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 617.3 41.5 39.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/agmyufp1",
   "source_url": "https://omlx.ai/benchmarks/performance/agmyufp1",
   "tg_tokens": null,
   "tps": 41.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-agmyufp1-ctx131072",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "pyros-vault MTP build; oMLX community benchmark trace; peak memory 83.2 GB at this context; duplicate traces of the same session: b1sluf0m, hj41ahwa, r2t9yutl, ziaa5bwx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 329.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "pyros-vault:Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 329.1 22.7 83.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/agmyufp1",
   "source_url": "https://omlx.ai/benchmarks/performance/agmyufp1",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-3c0s4t7h-ctx1024",
   "model": "GLM-4.7-Flash",
   "notes": "oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1160.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "GLM-4.7-Flash on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 1,160 60.2 30.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3c0s4t7h",
   "source_url": "https://omlx.ai/benchmarks/performance/3c0s4t7h",
   "tg_tokens": null,
   "tps": 60.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-3c0s4t7h-ctx4096",
   "model": "GLM-4.7-Flash",
   "notes": "oMLX community benchmark trace; peak memory 31.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1091.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "GLM-4.7-Flash on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 1,091 55.3 31.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3c0s4t7h",
   "source_url": "https://omlx.ai/benchmarks/performance/3c0s4t7h",
   "tg_tokens": null,
   "tps": 55.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-82xgho4i-ctx1024",
   "model": "GLM-4.7-Flash",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 0q5i6fy9; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 608.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "GLM-4.7-Flash on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 608.4 49.8 23.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/82xgho4i",
   "source_url": "https://omlx.ai/benchmarks/performance/82xgho4i",
   "tg_tokens": null,
   "tps": 49.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-82xgho4i-ctx4096",
   "model": "GLM-4.7-Flash",
   "notes": "oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: 0q5i6fy9; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 546.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "GLM-4.7-Flash on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 546.4 42.5 24.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/82xgho4i",
   "source_url": "https://omlx.ai/benchmarks/performance/82xgho4i",
   "tg_tokens": null,
   "tps": 42.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3ctrecp9-ctx1024",
   "model": "GLM-5",
   "notes": "oMLX community benchmark trace; peak memory 420.0 GB at this context; duplicate traces of the same session: teqbm0rb, xj3z6y6p; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 188.7,
   "provenance": "community",
   "quant": "4.8bit",
   "quote": "GLM-5-4.8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 188.7 16.6 420.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ctrecp9",
   "source_url": "https://omlx.ai/benchmarks/performance/3ctrecp9",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3ctrecp9-ctx4096",
   "model": "GLM-5",
   "notes": "oMLX community benchmark trace; peak memory 422.3 GB at this context; duplicate traces of the same session: teqbm0rb, xj3z6y6p; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 182.6,
   "provenance": "community",
   "quant": "4.8bit",
   "quote": "GLM-5-4.8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 182.6 13.7 422.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ctrecp9",
   "source_url": "https://omlx.ai/benchmarks/performance/3ctrecp9",
   "tg_tokens": null,
   "tps": 13.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3ctrecp9-ctx8192",
   "model": "GLM-5",
   "notes": "oMLX community benchmark trace; peak memory 424.9 GB at this context; duplicate traces of the same session: teqbm0rb, xj3z6y6p; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 156.3,
   "provenance": "community",
   "quant": "4.8bit",
   "quote": "GLM-5-4.8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 156.3 13.5 424.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ctrecp9",
   "source_url": "https://omlx.ai/benchmarks/performance/3ctrecp9",
   "tg_tokens": null,
   "tps": 13.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3ctrecp9-ctx16384",
   "model": "GLM-5",
   "notes": "oMLX community benchmark trace; peak memory 430.9 GB at this context; duplicate traces of the same session: teqbm0rb, xj3z6y6p; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 119.8,
   "provenance": "community",
   "quant": "4.8bit",
   "quote": "GLM-5-4.8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 119.8 13.0 430.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ctrecp9",
   "source_url": "https://omlx.ai/benchmarks/performance/3ctrecp9",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3ctrecp9-ctx32768",
   "model": "GLM-5",
   "notes": "oMLX community benchmark trace; peak memory 443.6 GB at this context; duplicate traces of the same session: teqbm0rb, xj3z6y6p; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 73.2,
   "provenance": "community",
   "quant": "4.8bit",
   "quote": "GLM-5-4.8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 73.2 12.3 443.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ctrecp9",
   "source_url": "https://omlx.ai/benchmarks/performance/3ctrecp9",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-4xl0yxqf-ctx1024",
   "model": "Llama-3.2-3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 2.3 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 520.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Llama-3.2-3B-Instruct on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 520.4 45.2 2.3 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4xl0yxqf",
   "source_url": "https://omlx.ai/benchmarks/performance/4xl0yxqf",
   "tg_tokens": null,
   "tps": 45.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-4xl0yxqf-ctx4096",
   "model": "Llama-3.2-3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 2.6 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 478.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Llama-3.2-3B-Instruct on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 478.6 37.0 2.6 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4xl0yxqf",
   "source_url": "https://omlx.ai/benchmarks/performance/4xl0yxqf",
   "tg_tokens": null,
   "tps": 37.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-2kpxm1c1-ctx1024",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 94.1 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 299.7,
   "provenance": "community",
   "quant": "3bit",
   "quote": "MiniMax-M2.5 on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 3bit; 1k 299.7 42.1 94.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2kpxm1c1",
   "source_url": "https://omlx.ai/benchmarks/performance/2kpxm1c1",
   "tg_tokens": null,
   "tps": 42.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-2kpxm1c1-ctx4096",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 95.1 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 337.6,
   "provenance": "community",
   "quant": "3bit",
   "quote": "MiniMax-M2.5 on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 3bit; 4k 337.6 35.7 95.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2kpxm1c1",
   "source_url": "https://omlx.ai/benchmarks/performance/2kpxm1c1",
   "tg_tokens": null,
   "tps": 35.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-2kpxm1c1-ctx8192",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 96.0 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 320.5,
   "provenance": "community",
   "quant": "3bit",
   "quote": "MiniMax-M2.5 on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 3bit; 8k 320.5 29.5 96.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2kpxm1c1",
   "source_url": "https://omlx.ai/benchmarks/performance/2kpxm1c1",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-2kpxm1c1-ctx16384",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 98.0 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 282.6,
   "provenance": "community",
   "quant": "3bit",
   "quote": "MiniMax-M2.5 on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 3bit; 16k 282.6 22.4 98.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2kpxm1c1",
   "source_url": "https://omlx.ai/benchmarks/performance/2kpxm1c1",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bzeg02vo-ctx1024",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 174.0 GB at this context; duplicate traces of the same session: 2we8lcjq, xbfycdsw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 496.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 1k 496.7 40.0 174.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bzeg02vo",
   "source_url": "https://omlx.ai/benchmarks/performance/bzeg02vo",
   "tg_tokens": null,
   "tps": 40.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bzeg02vo-ctx4096",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 175.0 GB at this context; duplicate traces of the same session: 2we8lcjq, xbfycdsw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 554.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 4k 554.8 33.9 175.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bzeg02vo",
   "source_url": "https://omlx.ai/benchmarks/performance/bzeg02vo",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bzeg02vo-ctx8192",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 175.9 GB at this context; duplicate traces of the same session: 2we8lcjq, xbfycdsw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 514.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 8k 514.0 28.9 175.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bzeg02vo",
   "source_url": "https://omlx.ai/benchmarks/performance/bzeg02vo",
   "tg_tokens": null,
   "tps": 28.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bzeg02vo-ctx16384",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 177.8 GB at this context; duplicate traces of the same session: 2we8lcjq, xbfycdsw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 436.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 16k 436.3 23.5 177.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bzeg02vo",
   "source_url": "https://omlx.ai/benchmarks/performance/bzeg02vo",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bzeg02vo-ctx32768",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 181.7 GB at this context; duplicate traces of the same session: 2we8lcjq, xbfycdsw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 330.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 32k 330.8 16.1 181.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bzeg02vo",
   "source_url": "https://omlx.ai/benchmarks/performance/bzeg02vo",
   "tg_tokens": null,
   "tps": 16.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bzeg02vo-ctx65536",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 189.6 GB at this context; duplicate traces of the same session: 2we8lcjq, xbfycdsw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 221.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 64k 221.1 9.6 189.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bzeg02vo",
   "source_url": "https://omlx.ai/benchmarks/performance/bzeg02vo",
   "tg_tokens": null,
   "tps": 9.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-s15twedr-ctx1024",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 227.2 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 600.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 600.4 34.2 227.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s15twedr",
   "source_url": "https://omlx.ai/benchmarks/performance/s15twedr",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-s15twedr-ctx4096",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 228.2 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 707.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 707.9 30.1 228.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s15twedr",
   "source_url": "https://omlx.ai/benchmarks/performance/s15twedr",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-s15twedr-ctx8192",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 229.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 663.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 663.9 25.9 229.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s15twedr",
   "source_url": "https://omlx.ai/benchmarks/performance/s15twedr",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-s15twedr-ctx16384",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 231.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 564.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k 564.2 21.1 231.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s15twedr",
   "source_url": "https://omlx.ai/benchmarks/performance/s15twedr",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-s15twedr-ctx32768",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 235.0 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 427.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 32k 427.8 14.6 235.0 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s15twedr",
   "source_url": "https://omlx.ai/benchmarks/performance/s15twedr",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-bg1g56a9-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 61.6 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 464.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 1k 464.3 53.4 61.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bg1g56a9",
   "source_url": "https://omlx.ai/benchmarks/performance/bg1g56a9",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-bg1g56a9-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 61.7 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 443.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 4k 443.4 50.2 61.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bg1g56a9",
   "source_url": "https://omlx.ai/benchmarks/performance/bg1g56a9",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-bg1g56a9-ctx8192",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 62.0 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 427.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 8k 427.1 46.1 62.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bg1g56a9",
   "source_url": "https://omlx.ai/benchmarks/performance/bg1g56a9",
   "tg_tokens": null,
   "tps": 46.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-bg1g56a9-ctx16384",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 62.4 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 412.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 16k 412.9 38.1 62.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bg1g56a9",
   "source_url": "https://omlx.ai/benchmarks/performance/bg1g56a9",
   "tg_tokens": null,
   "tps": 38.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-bg1g56a9-ctx32768",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 63.3 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 379.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 32k 379.4 25.5 63.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bg1g56a9",
   "source_url": "https://omlx.ai/benchmarks/performance/bg1g56a9",
   "tg_tokens": null,
   "tps": 25.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-n6v3m0w2-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 61.6 GB at this context; duplicate traces of the same session: bgwyirz3; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 751.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 1k 751.3 57.9 61.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n6v3m0w2",
   "source_url": "https://omlx.ai/benchmarks/performance/n6v3m0w2",
   "tg_tokens": null,
   "tps": 57.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-n6v3m0w2-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 61.7 GB at this context; duplicate traces of the same session: bgwyirz3; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 966.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 4k 966.4 56.2 61.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n6v3m0w2",
   "source_url": "https://omlx.ai/benchmarks/performance/n6v3m0w2",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-n6v3m0w2-ctx8192",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 62.0 GB at this context; duplicate traces of the same session: bgwyirz3; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 986.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 8k 986.0 53.4 62.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n6v3m0w2",
   "source_url": "https://omlx.ai/benchmarks/performance/n6v3m0w2",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-n6v3m0w2-ctx16384",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 62.4 GB at this context; duplicate traces of the same session: bgwyirz3; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 956.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 16k 956.5 47.0 62.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n6v3m0w2",
   "source_url": "https://omlx.ai/benchmarks/performance/n6v3m0w2",
   "tg_tokens": null,
   "tps": 47.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-n6v3m0w2-ctx32768",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 63.3 GB at this context; duplicate traces of the same session: bgwyirz3; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 881.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 32k 881.3 39.1 63.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n6v3m0w2",
   "source_url": "https://omlx.ai/benchmarks/performance/n6v3m0w2",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-n6v3m0w2-ctx65536",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 65.1 GB at this context; duplicate traces of the same session: bgwyirz3; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 731.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 64k 731.4 28.9 65.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n6v3m0w2",
   "source_url": "https://omlx.ai/benchmarks/performance/n6v3m0w2",
   "tg_tokens": null,
   "tps": 28.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-scn0knor-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.0 GB at this context; duplicate traces of the same session: rihr0r0i; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1298.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 1,298 73.9 43.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/scn0knor",
   "source_url": "https://omlx.ai/benchmarks/performance/scn0knor",
   "tg_tokens": null,
   "tps": 73.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-scn0knor-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.2 GB at this context; duplicate traces of the same session: rihr0r0i; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1558.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 1,558 71.0 43.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/scn0knor",
   "source_url": "https://omlx.ai/benchmarks/performance/scn0knor",
   "tg_tokens": null,
   "tps": 71.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-scn0knor-ctx8192",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.4 GB at this context; duplicate traces of the same session: rihr0r0i; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1552.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,552 66.4 43.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/scn0knor",
   "source_url": "https://omlx.ai/benchmarks/performance/scn0knor",
   "tg_tokens": null,
   "tps": 66.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-scn0knor-ctx16384",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.9 GB at this context; duplicate traces of the same session: rihr0r0i; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1461.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 1,461 58.9 43.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/scn0knor",
   "source_url": "https://omlx.ai/benchmarks/performance/scn0knor",
   "tg_tokens": null,
   "tps": 58.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-scn0knor-ctx32768",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 44.7 GB at this context; duplicate traces of the same session: rihr0r0i; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1281.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 1,281 49.4 44.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/scn0knor",
   "source_url": "https://omlx.ai/benchmarks/performance/scn0knor",
   "tg_tokens": null,
   "tps": 49.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-scn0knor-ctx65536",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 46.5 GB at this context; duplicate traces of the same session: rihr0r0i; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1007.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 1,007 36.9 46.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/scn0knor",
   "source_url": "https://omlx.ai/benchmarks/performance/scn0knor",
   "tg_tokens": null,
   "tps": 36.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jrf4vgpr-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 80.1 GB at this context; duplicate traces of the same session: rhiba0g8, xill9rj9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1236.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 1,236 63.9 80.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jrf4vgpr",
   "source_url": "https://omlx.ai/benchmarks/performance/jrf4vgpr",
   "tg_tokens": null,
   "tps": 63.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-jrf4vgpr-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 80.3 GB at this context; duplicate traces of the same session: rhiba0g8, xill9rj9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1509.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 1,509 61.9 80.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jrf4vgpr",
   "source_url": "https://omlx.ai/benchmarks/performance/jrf4vgpr",
   "tg_tokens": null,
   "tps": 61.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jrf4vgpr-ctx8192",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 80.5 GB at this context; duplicate traces of the same session: rhiba0g8, xill9rj9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1521.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 1,521 56.5 80.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jrf4vgpr",
   "source_url": "https://omlx.ai/benchmarks/performance/jrf4vgpr",
   "tg_tokens": null,
   "tps": 56.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-jrf4vgpr-ctx16384",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 80.9 GB at this context; duplicate traces of the same session: rhiba0g8, xill9rj9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1434.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 16k 1,434 52.0 80.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jrf4vgpr",
   "source_url": "https://omlx.ai/benchmarks/performance/jrf4vgpr",
   "tg_tokens": null,
   "tps": 52.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jrf4vgpr-ctx32768",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 81.8 GB at this context; duplicate traces of the same session: rhiba0g8, xill9rj9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1262.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 32k 1,262 45.8 81.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jrf4vgpr",
   "source_url": "https://omlx.ai/benchmarks/performance/jrf4vgpr",
   "tg_tokens": null,
   "tps": 45.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-jrf4vgpr-ctx65536",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 83.6 GB at this context; duplicate traces of the same session: rhiba0g8, xill9rj9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 994.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 64k 994.8 34.7 83.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jrf4vgpr",
   "source_url": "https://omlx.ai/benchmarks/performance/jrf4vgpr",
   "tg_tokens": null,
   "tps": 34.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9kt8xfk2-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 80.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1442.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 1,442 58.8 80.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9kt8xfk2",
   "source_url": "https://omlx.ai/benchmarks/performance/9kt8xfk2",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9kt8xfk2-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 80.3 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1872.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 1,872 58.5 80.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9kt8xfk2",
   "source_url": "https://omlx.ai/benchmarks/performance/9kt8xfk2",
   "tg_tokens": null,
   "tps": 58.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9kt8xfk2-ctx8192",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 80.5 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1888.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 1,888 55.2 80.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9kt8xfk2",
   "source_url": "https://omlx.ai/benchmarks/performance/9kt8xfk2",
   "tg_tokens": null,
   "tps": 55.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9kt8xfk2-ctx16384",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 80.9 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1783.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k 1,783 49.9 80.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9kt8xfk2",
   "source_url": "https://omlx.ai/benchmarks/performance/9kt8xfk2",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9kt8xfk2-ctx32768",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 81.8 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1549.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 32k 1,549 42.5 81.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9kt8xfk2",
   "source_url": "https://omlx.ai/benchmarks/performance/9kt8xfk2",
   "tg_tokens": null,
   "tps": 42.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-qqnsjbs4-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.0 GB at this context; duplicate traces of the same session: 36gako4n; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 574.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 574.9 64.4 43.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qqnsjbs4",
   "source_url": "https://omlx.ai/benchmarks/performance/qqnsjbs4",
   "tg_tokens": null,
   "tps": 64.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-qqnsjbs4-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.2 GB at this context; duplicate traces of the same session: 36gako4n; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 660.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 660.7 60.6 43.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qqnsjbs4",
   "source_url": "https://omlx.ai/benchmarks/performance/qqnsjbs4",
   "tg_tokens": null,
   "tps": 60.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-y89wklsn-ctx1024",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 1.5 GB at this context; duplicate traces of the same session: 3l7graoy, doabcv9b, od368nvv; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 756.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 4bit; 1k 756.4 100.8 1.5 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y89wklsn",
   "source_url": "https://omlx.ai/benchmarks/performance/y89wklsn",
   "tg_tokens": null,
   "tps": 100.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-y89wklsn-ctx4096",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 1.6 GB at this context; duplicate traces of the same session: 3l7graoy, doabcv9b, od368nvv; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 734.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 4bit; 4k 734.1 91.3 1.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y89wklsn",
   "source_url": "https://omlx.ai/benchmarks/performance/y89wklsn",
   "tg_tokens": null,
   "tps": 91.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-y89wklsn-ctx8192",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 1.6 GB at this context; duplicate traces of the same session: 3l7graoy, doabcv9b, od368nvv; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 710.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 4bit; 8k 710.7 82.2 1.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y89wklsn",
   "source_url": "https://omlx.ai/benchmarks/performance/y89wklsn",
   "tg_tokens": null,
   "tps": 82.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-y89wklsn-ctx16384",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 2.0 GB at this context; duplicate traces of the same session: 3l7graoy, doabcv9b, od368nvv; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 647.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 4bit; 16k 647.6 68.2 2.0 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y89wklsn",
   "source_url": "https://omlx.ai/benchmarks/performance/y89wklsn",
   "tg_tokens": null,
   "tps": 68.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-y89wklsn-ctx32768",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 2.6 GB at this context; duplicate traces of the same session: 3l7graoy, doabcv9b, od368nvv; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 552.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 4bit; 32k 552.2 47.8 2.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y89wklsn",
   "source_url": "https://omlx.ai/benchmarks/performance/y89wklsn",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-y89wklsn-ctx65536",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 4.1 GB at this context; duplicate traces of the same session: 3l7graoy, doabcv9b, od368nvv; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 408.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 4bit; 64k 408.6 17.3 4.1 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y89wklsn",
   "source_url": "https://omlx.ai/benchmarks/performance/y89wklsn",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-es1hp4a4-ctx1024",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 1.8 GB at this context; duplicate traces of the same session: f00eert8, mxnj5jsm, wpqf311n; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 755.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 8bit; 1k 755.9 62.4 1.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/es1hp4a4",
   "source_url": "https://omlx.ai/benchmarks/performance/es1hp4a4",
   "tg_tokens": null,
   "tps": 62.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-es1hp4a4-ctx4096",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 1.9 GB at this context; duplicate traces of the same session: f00eert8, mxnj5jsm, wpqf311n; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 733.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 8bit; 4k 733.9 58.9 1.9 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/es1hp4a4",
   "source_url": "https://omlx.ai/benchmarks/performance/es1hp4a4",
   "tg_tokens": null,
   "tps": 58.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-es1hp4a4-ctx8192",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 2.0 GB at this context; duplicate traces of the same session: f00eert8, mxnj5jsm, wpqf311n; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 714.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 8bit; 8k 714.0 55.0 2.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/es1hp4a4",
   "source_url": "https://omlx.ai/benchmarks/performance/es1hp4a4",
   "tg_tokens": null,
   "tps": 55.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-es1hp4a4-ctx16384",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 2.4 GB at this context; duplicate traces of the same session: f00eert8, mxnj5jsm, wpqf311n; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 654.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 8bit; 16k 654.9 47.9 2.4 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/es1hp4a4",
   "source_url": "https://omlx.ai/benchmarks/performance/es1hp4a4",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-es1hp4a4-ctx32768",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 3.0 GB at this context; duplicate traces of the same session: f00eert8, mxnj5jsm, wpqf311n; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 554.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 8bit; 32k 554.7 37.3 3.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/es1hp4a4",
   "source_url": "https://omlx.ai/benchmarks/performance/es1hp4a4",
   "tg_tokens": null,
   "tps": 37.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-es1hp4a4-ctx65536",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 4.4 GB at this context; duplicate traces of the same session: f00eert8, mxnj5jsm, wpqf311n; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 414.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 8bit; 64k 414.9 24.8 4.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/es1hp4a4",
   "source_url": "https://omlx.ai/benchmarks/performance/es1hp4a4",
   "tg_tokens": null,
   "tps": 24.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-7bv8ke3w-ctx1024",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 1.8 GB at this context; duplicate traces of the same session: none; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 713.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 8bit; 1k 713.6 46.2 1.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7bv8ke3w",
   "source_url": "https://omlx.ai/benchmarks/performance/7bv8ke3w",
   "tg_tokens": null,
   "tps": 46.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-7bv8ke3w-ctx4096",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 1.9 GB at this context; duplicate traces of the same session: none; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 707.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 8bit; 4k 707.8 44.7 1.9 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7bv8ke3w",
   "source_url": "https://omlx.ai/benchmarks/performance/7bv8ke3w",
   "tg_tokens": null,
   "tps": 44.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-hj1h2047-ctx1024",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 1.8 GB at this context; duplicate traces of the same session: udv2mwwl; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 726.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 8bit; 1k 726.8 48.2 1.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hj1h2047",
   "source_url": "https://omlx.ai/benchmarks/performance/hj1h2047",
   "tg_tokens": null,
   "tps": 48.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-hj1h2047-ctx4096",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 1.9 GB at this context; duplicate traces of the same session: udv2mwwl; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 653.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 8bit; 4k 653.5 41.2 1.9 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hj1h2047",
   "source_url": "https://omlx.ai/benchmarks/performance/hj1h2047",
   "tg_tokens": null,
   "tps": 41.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-hj1h2047-ctx8192",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 2.0 GB at this context; duplicate traces of the same session: udv2mwwl; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 646.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 8bit; 8k 646.1 45.1 2.0 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hj1h2047",
   "source_url": "https://omlx.ai/benchmarks/performance/hj1h2047",
   "tg_tokens": null,
   "tps": 45.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-5fzqvnit-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.3 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 229.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 229.0 34.8 66.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5fzqvnit",
   "source_url": "https://omlx.ai/benchmarks/performance/5fzqvnit",
   "tg_tokens": null,
   "tps": 34.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-5fzqvnit-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.7 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 196.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 196.2 31.0 66.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5fzqvnit",
   "source_url": "https://omlx.ai/benchmarks/performance/5fzqvnit",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-5fzqvnit-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.0 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 192.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 192.2 27.8 67.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5fzqvnit",
   "source_url": "https://omlx.ai/benchmarks/performance/5fzqvnit",
   "tg_tokens": null,
   "tps": 27.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-5fzqvnit-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.7 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 183.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 183.0 22.5 67.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5fzqvnit",
   "source_url": "https://omlx.ai/benchmarks/performance/5fzqvnit",
   "tg_tokens": null,
   "tps": 22.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-5fzqvnit-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 69.1 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 170.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 32k 170.3 16.8 69.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5fzqvnit",
   "source_url": "https://omlx.ai/benchmarks/performance/5fzqvnit",
   "tg_tokens": null,
   "tps": 16.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-2k6ofxl2-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.3 GB at this context; duplicate traces of the same session: hfzzor6h; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 439.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 1k 439.6 47.8 66.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2k6ofxl2",
   "source_url": "https://omlx.ai/benchmarks/performance/2k6ofxl2",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-2k6ofxl2-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.7 GB at this context; duplicate traces of the same session: hfzzor6h; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 463.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 4k 463.3 45.6 66.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2k6ofxl2",
   "source_url": "https://omlx.ai/benchmarks/performance/2k6ofxl2",
   "tg_tokens": null,
   "tps": 45.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-2k6ofxl2-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.0 GB at this context; duplicate traces of the same session: hfzzor6h; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 459.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 8k 459.9 41.3 67.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2k6ofxl2",
   "source_url": "https://omlx.ai/benchmarks/performance/2k6ofxl2",
   "tg_tokens": null,
   "tps": 41.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-2k6ofxl2-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.7 GB at this context; duplicate traces of the same session: hfzzor6h; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 444.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 16k 444.0 36.3 67.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2k6ofxl2",
   "source_url": "https://omlx.ai/benchmarks/performance/2k6ofxl2",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-2k6ofxl2-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 69.1 GB at this context; duplicate traces of the same session: hfzzor6h; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 409.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 32k 409.4 28.3 69.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2k6ofxl2",
   "source_url": "https://omlx.ai/benchmarks/performance/2k6ofxl2",
   "tg_tokens": null,
   "tps": 28.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-2k6ofxl2-ctx65536",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 71.9 GB at this context; duplicate traces of the same session: hfzzor6h; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 331.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 64k 331.7 19.0 71.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2k6ofxl2",
   "source_url": "https://omlx.ai/benchmarks/performance/2k6ofxl2",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bt6bvgoc-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.3 GB at this context; duplicate traces of the same session: ejitl07v, jgotjhpb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 704.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 704.1 53.8 66.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bt6bvgoc",
   "source_url": "https://omlx.ai/benchmarks/performance/bt6bvgoc",
   "tg_tokens": null,
   "tps": 53.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bt6bvgoc-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.7 GB at this context; duplicate traces of the same session: ejitl07v, jgotjhpb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 729.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 729.1 51.2 66.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bt6bvgoc",
   "source_url": "https://omlx.ai/benchmarks/performance/bt6bvgoc",
   "tg_tokens": null,
   "tps": 51.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bt6bvgoc-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.0 GB at this context; duplicate traces of the same session: ejitl07v, jgotjhpb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 712.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 712.9 48.2 67.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bt6bvgoc",
   "source_url": "https://omlx.ai/benchmarks/performance/bt6bvgoc",
   "tg_tokens": null,
   "tps": 48.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bt6bvgoc-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.7 GB at this context; duplicate traces of the same session: ejitl07v, jgotjhpb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 670.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 670.3 41.2 67.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bt6bvgoc",
   "source_url": "https://omlx.ai/benchmarks/performance/bt6bvgoc",
   "tg_tokens": null,
   "tps": 41.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bt6bvgoc-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 69.1 GB at this context; duplicate traces of the same session: ejitl07v, jgotjhpb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 590.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 590.2 35.0 69.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bt6bvgoc",
   "source_url": "https://omlx.ai/benchmarks/performance/bt6bvgoc",
   "tg_tokens": null,
   "tps": 35.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bt6bvgoc-ctx65536",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 71.9 GB at this context; duplicate traces of the same session: ejitl07v, jgotjhpb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 466.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 466.9 26.0 71.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bt6bvgoc",
   "source_url": "https://omlx.ai/benchmarks/performance/bt6bvgoc",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-tiqiwz3j-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.3 GB at this context; duplicate traces of the same session: 919dis83, bzhe0vh3, kvxrqtnz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 638.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 638.7 54.4 66.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tiqiwz3j",
   "source_url": "https://omlx.ai/benchmarks/performance/tiqiwz3j",
   "tg_tokens": null,
   "tps": 54.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-tiqiwz3j-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.7 GB at this context; duplicate traces of the same session: 919dis83, bzhe0vh3, kvxrqtnz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 714.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 714.1 50.9 66.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tiqiwz3j",
   "source_url": "https://omlx.ai/benchmarks/performance/tiqiwz3j",
   "tg_tokens": null,
   "tps": 50.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-tiqiwz3j-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.0 GB at this context; duplicate traces of the same session: 919dis83, bzhe0vh3, kvxrqtnz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 712.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 712.7 47.3 67.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tiqiwz3j",
   "source_url": "https://omlx.ai/benchmarks/performance/tiqiwz3j",
   "tg_tokens": null,
   "tps": 47.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-tiqiwz3j-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.7 GB at this context; duplicate traces of the same session: 919dis83, bzhe0vh3, kvxrqtnz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 670.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 16k 670.7 35.9 67.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tiqiwz3j",
   "source_url": "https://omlx.ai/benchmarks/performance/tiqiwz3j",
   "tg_tokens": null,
   "tps": 35.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-tiqiwz3j-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 69.1 GB at this context; duplicate traces of the same session: 919dis83, bzhe0vh3, kvxrqtnz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 587.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 587.5 24.4 69.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tiqiwz3j",
   "source_url": "https://omlx.ai/benchmarks/performance/tiqiwz3j",
   "tg_tokens": null,
   "tps": 24.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-tiqiwz3j-ctx65536",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 71.9 GB at this context; duplicate traces of the same session: 919dis83, bzhe0vh3, kvxrqtnz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 458.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k 458.6 15.5 71.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tiqiwz3j",
   "source_url": "https://omlx.ai/benchmarks/performance/tiqiwz3j",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-lhzzunu0-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.3 GB at this context; duplicate traces of the same session: 43mgnpxt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 897.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 897.8 52.3 66.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lhzzunu0",
   "source_url": "https://omlx.ai/benchmarks/performance/lhzzunu0",
   "tg_tokens": null,
   "tps": 52.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-lhzzunu0-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.7 GB at this context; duplicate traces of the same session: 43mgnpxt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 935.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 935.9 47.9 66.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lhzzunu0",
   "source_url": "https://omlx.ai/benchmarks/performance/lhzzunu0",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-lhzzunu0-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.0 GB at this context; duplicate traces of the same session: 43mgnpxt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 920.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 920.4 47.4 67.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lhzzunu0",
   "source_url": "https://omlx.ai/benchmarks/performance/lhzzunu0",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-lhzzunu0-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.7 GB at this context; duplicate traces of the same session: 43mgnpxt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 862.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 862.3 43.1 67.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lhzzunu0",
   "source_url": "https://omlx.ai/benchmarks/performance/lhzzunu0",
   "tg_tokens": null,
   "tps": 43.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-lhzzunu0-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 69.1 GB at this context; duplicate traces of the same session: 43mgnpxt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 746.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 746.1 36.4 69.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lhzzunu0",
   "source_url": "https://omlx.ai/benchmarks/performance/lhzzunu0",
   "tg_tokens": null,
   "tps": 36.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-lhzzunu0-ctx65536",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 71.9 GB at this context; duplicate traces of the same session: 43mgnpxt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 570.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 570.2 27.7 71.9 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lhzzunu0",
   "source_url": "https://omlx.ai/benchmarks/performance/lhzzunu0",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ho7w5nty-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 123.2 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 837.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 837.2 39.1 123.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ho7w5nty",
   "source_url": "https://omlx.ai/benchmarks/performance/ho7w5nty",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ho7w5nty-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 123.5 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 895.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 895.1 37.8 123.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ho7w5nty",
   "source_url": "https://omlx.ai/benchmarks/performance/ho7w5nty",
   "tg_tokens": null,
   "tps": 37.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ho7w5nty-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 123.9 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 876.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 876.6 35.7 123.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ho7w5nty",
   "source_url": "https://omlx.ai/benchmarks/performance/ho7w5nty",
   "tg_tokens": null,
   "tps": 35.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ho7w5nty-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 124.5 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 826.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k 826.4 32.5 124.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ho7w5nty",
   "source_url": "https://omlx.ai/benchmarks/performance/ho7w5nty",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ho7w5nty-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 125.9 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 717.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 32k 717.8 27.9 125.9 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ho7w5nty",
   "source_url": "https://omlx.ai/benchmarks/performance/ho7w5nty",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-lrmb0sb2-ctx1024",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 68.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 1k 68.0 15.4 16.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lrmb0sb2",
   "source_url": "https://omlx.ai/benchmarks/performance/lrmb0sb2",
   "tg_tokens": null,
   "tps": 15.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-lrmb0sb2-ctx4096",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 68.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k 68.0 14.4 17.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lrmb0sb2",
   "source_url": "https://omlx.ai/benchmarks/performance/lrmb0sb2",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ne7w6qe7-ctx1024",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.1 GB at this context; duplicate traces of the same session: ht88yq6s; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 85.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-27B on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 1k 85.8 11.1 29.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ne7w6qe7",
   "source_url": "https://omlx.ai/benchmarks/performance/ne7w6qe7",
   "tg_tokens": null,
   "tps": 11.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ne7w6qe7-ctx4096",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: ht88yq6s; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 89.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-27B on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k 89.2 10.6 29.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ne7w6qe7",
   "source_url": "https://omlx.ai/benchmarks/performance/ne7w6qe7",
   "tg_tokens": null,
   "tps": 10.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-8v24oyi6-ctx1024",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: fwtc8lbd, wtdigojg, yo2wg2fa; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 180.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 180.4 32.2 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8v24oyi6",
   "source_url": "https://omlx.ai/benchmarks/performance/8v24oyi6",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-8v24oyi6-ctx4096",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: fwtc8lbd, wtdigojg, yo2wg2fa; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 182.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 182.4 30.4 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8v24oyi6",
   "source_url": "https://omlx.ai/benchmarks/performance/8v24oyi6",
   "tg_tokens": null,
   "tps": 30.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-8v24oyi6-ctx8192",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: fwtc8lbd, wtdigojg, yo2wg2fa; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 180.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 8k 180.9 28.4 17.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8v24oyi6",
   "source_url": "https://omlx.ai/benchmarks/performance/8v24oyi6",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-8v24oyi6-ctx16384",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: fwtc8lbd, wtdigojg, yo2wg2fa; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 176.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 16k 176.2 25.3 18.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8v24oyi6",
   "source_url": "https://omlx.ai/benchmarks/performance/8v24oyi6",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-27rg9ogk-ctx32768",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 20.1 GB; TTFT 193741 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 169.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 32k tokens 169.1 PP tok/s 20.5 TG tok/s 193741 TTFT (ms) 20.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/27rg9ogk",
   "source_url": "https://omlx.ai/benchmarks/performance/27rg9ogk",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-5y2fk7fi-ctx1024",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 5raey9l1, i0ji024h; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 226.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 1k 226.5 33.4 16.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5y2fk7fi",
   "source_url": "https://omlx.ai/benchmarks/performance/5y2fk7fi",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-5y2fk7fi-ctx4096",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 5raey9l1, i0ji024h; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 228.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 4k 228.9 31.3 17.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5y2fk7fi",
   "source_url": "https://omlx.ai/benchmarks/performance/5y2fk7fi",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-5y2fk7fi-ctx8192",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: 5raey9l1, i0ji024h; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 227.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 8k 227.0 29.3 17.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5y2fk7fi",
   "source_url": "https://omlx.ai/benchmarks/performance/5y2fk7fi",
   "tg_tokens": null,
   "tps": 29.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-5y2fk7fi-ctx16384",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: 5raey9l1, i0ji024h; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 222.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 16k 222.3 26.6 18.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5y2fk7fi",
   "source_url": "https://omlx.ai/benchmarks/performance/5y2fk7fi",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-5y2fk7fi-ctx32768",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 5raey9l1, i0ji024h; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 212.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 32k 212.9 21.9 20.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5y2fk7fi",
   "source_url": "https://omlx.ai/benchmarks/performance/5y2fk7fi",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-5y2fk7fi-ctx65536",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 5raey9l1, i0ji024h; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 192.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 64k 192.2 16.8 23.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5y2fk7fi",
   "source_url": "https://omlx.ai/benchmarks/performance/5y2fk7fi",
   "tg_tokens": null,
   "tps": 16.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jkbcbaj9-ctx1024",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: iff2zzcf, w3gm3t09; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 322.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 322.7 38.0 16.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jkbcbaj9",
   "source_url": "https://omlx.ai/benchmarks/performance/jkbcbaj9",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jkbcbaj9-ctx4096",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: iff2zzcf, w3gm3t09; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 324.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 324.2 35.3 17.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jkbcbaj9",
   "source_url": "https://omlx.ai/benchmarks/performance/jkbcbaj9",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jkbcbaj9-ctx8192",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: iff2zzcf, w3gm3t09; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 319.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 319.7 33.6 17.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jkbcbaj9",
   "source_url": "https://omlx.ai/benchmarks/performance/jkbcbaj9",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jkbcbaj9-ctx16384",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: iff2zzcf, w3gm3t09; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 310.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 310.0 30.0 18.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jkbcbaj9",
   "source_url": "https://omlx.ai/benchmarks/performance/jkbcbaj9",
   "tg_tokens": null,
   "tps": 30.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jkbcbaj9-ctx32768",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: iff2zzcf, w3gm3t09; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 291.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 291.3 24.9 20.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jkbcbaj9",
   "source_url": "https://omlx.ai/benchmarks/performance/jkbcbaj9",
   "tg_tokens": null,
   "tps": 24.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jkbcbaj9-ctx65536",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: iff2zzcf, w3gm3t09; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 258.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 258.7 18.1 23.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jkbcbaj9",
   "source_url": "https://omlx.ai/benchmarks/performance/jkbcbaj9",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-nktcknsh-ctx1024",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: j3mx6mkn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 239.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 239.4 27.6 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nktcknsh",
   "source_url": "https://omlx.ai/benchmarks/performance/nktcknsh",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-nktcknsh-ctx4096",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: j3mx6mkn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 199.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 199.2 23.6 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nktcknsh",
   "source_url": "https://omlx.ai/benchmarks/performance/nktcknsh",
   "tg_tokens": null,
   "tps": 23.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-nktcknsh-ctx8192",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: j3mx6mkn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 186.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 186.4 20.3 17.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nktcknsh",
   "source_url": "https://omlx.ai/benchmarks/performance/nktcknsh",
   "tg_tokens": null,
   "tps": 20.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-5zv9koqw-ctx1024",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: 4laq01bf; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 111.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 1k 111.1 21.1 15.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5zv9koqw",
   "source_url": "https://omlx.ai/benchmarks/performance/5zv9koqw",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-5zv9koqw-ctx4096",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: 4laq01bf; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 113.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k 113.8 20.1 16.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5zv9koqw",
   "source_url": "https://omlx.ai/benchmarks/performance/5zv9koqw",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-5zv9koqw-ctx8192",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 4laq01bf; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 113.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 8k 113.1 19.0 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5zv9koqw",
   "source_url": "https://omlx.ai/benchmarks/performance/5zv9koqw",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-5zv9koqw-ctx16384",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 17.5 GB at this context; duplicate traces of the same session: 4laq01bf; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 111.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 16k 111.0 17.3 17.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5zv9koqw",
   "source_url": "https://omlx.ai/benchmarks/performance/5zv9koqw",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-5zv9koqw-ctx32768",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 4laq01bf; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 32k 106.7 14.9 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5zv9koqw",
   "source_url": "https://omlx.ai/benchmarks/performance/5zv9koqw",
   "tg_tokens": null,
   "tps": 14.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-iugrfis1-ctx1024",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 375.2,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 1k 375.2 26.7 22.0 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iugrfis1",
   "source_url": "https://omlx.ai/benchmarks/performance/iugrfis1",
   "tg_tokens": null,
   "tps": 26.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-iugrfis1-ctx4096",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 401.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 4k 401.1 25.4 22.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iugrfis1",
   "source_url": "https://omlx.ai/benchmarks/performance/iugrfis1",
   "tg_tokens": null,
   "tps": 25.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-iugrfis1-ctx8192",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 400.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 8k 400.8 24.4 22.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iugrfis1",
   "source_url": "https://omlx.ai/benchmarks/performance/iugrfis1",
   "tg_tokens": null,
   "tps": 24.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-iugrfis1-ctx16384",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 392.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 16k 392.6 22.4 23.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iugrfis1",
   "source_url": "https://omlx.ai/benchmarks/performance/iugrfis1",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-iugrfis1-ctx32768",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 370.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 32k 370.0 19.5 25.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iugrfis1",
   "source_url": "https://omlx.ai/benchmarks/performance/iugrfis1",
   "tg_tokens": null,
   "tps": 19.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-wk1vunpj-ctx1024",
   "model": "Qwen3.5-2B",
   "notes": "oMLX community benchmark trace; peak memory 3.1 GB at this context; duplicate traces of the same session: p09e4cme; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1768.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-2B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 6bit; 1k 1,768 113.3 3.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wk1vunpj",
   "source_url": "https://omlx.ai/benchmarks/performance/wk1vunpj",
   "tg_tokens": null,
   "tps": 113.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-wk1vunpj-ctx4096",
   "model": "Qwen3.5-2B",
   "notes": "oMLX community benchmark trace; peak memory 3.2 GB at this context; duplicate traces of the same session: p09e4cme; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1797.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-2B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 6bit; 4k 1,797 109.1 3.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wk1vunpj",
   "source_url": "https://omlx.ai/benchmarks/performance/wk1vunpj",
   "tg_tokens": null,
   "tps": 109.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-lhp1vgf1-ctx1024",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 36.2 GB at this context; duplicate traces of the same session: zfc09bsq; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 519.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-35B-A3B on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 1k 519.6 48.0 36.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lhp1vgf1",
   "source_url": "https://omlx.ai/benchmarks/performance/lhp1vgf1",
   "tg_tokens": null,
   "tps": 48.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-lhp1vgf1-ctx4096",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 36.3 GB at this context; duplicate traces of the same session: zfc09bsq; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 566.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-35B-A3B on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k 566.5 43.5 36.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lhp1vgf1",
   "source_url": "https://omlx.ai/benchmarks/performance/lhp1vgf1",
   "tg_tokens": null,
   "tps": 43.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-rpqawcit-ctx1024",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: glgcvy99; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1361.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,361 95.7 20.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rpqawcit",
   "source_url": "https://omlx.ai/benchmarks/performance/rpqawcit",
   "tg_tokens": null,
   "tps": 95.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-rpqawcit-ctx4096",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: glgcvy99; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1067.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,067 95.8 20.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rpqawcit",
   "source_url": "https://omlx.ai/benchmarks/performance/rpqawcit",
   "tg_tokens": null,
   "tps": 95.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-rpqawcit-ctx8192",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: glgcvy99; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1091.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,091 90.5 20.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rpqawcit",
   "source_url": "https://omlx.ai/benchmarks/performance/rpqawcit",
   "tg_tokens": null,
   "tps": 90.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-6ehiuze9-ctx1024",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 0emkmm0e, 3xet8brb, femieeij, ptzhx9ln; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1652.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,652 93.4 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ehiuze9",
   "source_url": "https://omlx.ai/benchmarks/performance/6ehiuze9",
   "tg_tokens": null,
   "tps": 93.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-6ehiuze9-ctx4096",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 0emkmm0e, 3xet8brb, femieeij, ptzhx9ln; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1706.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,706 88.6 20.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ehiuze9",
   "source_url": "https://omlx.ai/benchmarks/performance/6ehiuze9",
   "tg_tokens": null,
   "tps": 88.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-6ehiuze9-ctx8192",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: 0emkmm0e, 3xet8brb, femieeij, ptzhx9ln; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1487.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,487 90.0 20.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ehiuze9",
   "source_url": "https://omlx.ai/benchmarks/performance/6ehiuze9",
   "tg_tokens": null,
   "tps": 90.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-6ehiuze9-ctx16384",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 0emkmm0e, 3xet8brb, femieeij, ptzhx9ln; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1227.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,227 63.7 20.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ehiuze9",
   "source_url": "https://omlx.ai/benchmarks/performance/6ehiuze9",
   "tg_tokens": null,
   "tps": 63.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-6ehiuze9-ctx32768",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: 0emkmm0e, 3xet8brb, femieeij, ptzhx9ln; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1026.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 1,026 42.5 21.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ehiuze9",
   "source_url": "https://omlx.ai/benchmarks/performance/6ehiuze9",
   "tg_tokens": null,
   "tps": 42.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-1txw1wzl-ctx1024",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: vpoz1qjs, xmv2d997; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 821.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 821.5 75.6 20.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1txw1wzl",
   "source_url": "https://omlx.ai/benchmarks/performance/1txw1wzl",
   "tg_tokens": null,
   "tps": 75.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-1txw1wzl-ctx4096",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: vpoz1qjs, xmv2d997; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 879.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 879.6 79.1 20.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1txw1wzl",
   "source_url": "https://omlx.ai/benchmarks/performance/1txw1wzl",
   "tg_tokens": null,
   "tps": 79.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-1txw1wzl-ctx8192",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: vpoz1qjs, xmv2d997; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 873.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 873.4 71.9 20.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1txw1wzl",
   "source_url": "https://omlx.ai/benchmarks/performance/1txw1wzl",
   "tg_tokens": null,
   "tps": 71.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-0tkxpczj-ctx1024",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 9bgkvghg, cq1j63z1, fxnkfdkj, tzmnrcg3, vy20rvxa; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 867.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 867.2 82.8 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0tkxpczj",
   "source_url": "https://omlx.ai/benchmarks/performance/0tkxpczj",
   "tg_tokens": null,
   "tps": 82.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-0tkxpczj-ctx4096",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 9bgkvghg, cq1j63z1, fxnkfdkj, tzmnrcg3, vy20rvxa; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 899.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 899.9 79.2 20.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0tkxpczj",
   "source_url": "https://omlx.ai/benchmarks/performance/0tkxpczj",
   "tg_tokens": null,
   "tps": 79.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-0tkxpczj-ctx8192",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: 9bgkvghg, cq1j63z1, fxnkfdkj, tzmnrcg3, vy20rvxa; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 876.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 876.2 74.9 20.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0tkxpczj",
   "source_url": "https://omlx.ai/benchmarks/performance/0tkxpczj",
   "tg_tokens": null,
   "tps": 74.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-0tkxpczj-ctx16384",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 9bgkvghg, cq1j63z1, fxnkfdkj, tzmnrcg3, vy20rvxa; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 805.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 805.6 57.7 20.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0tkxpczj",
   "source_url": "https://omlx.ai/benchmarks/performance/0tkxpczj",
   "tg_tokens": null,
   "tps": 57.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-0tkxpczj-ctx32768",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: 9bgkvghg, cq1j63z1, fxnkfdkj, tzmnrcg3, vy20rvxa; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 693.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 693.2 40.6 21.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0tkxpczj",
   "source_url": "https://omlx.ai/benchmarks/performance/0tkxpczj",
   "tg_tokens": null,
   "tps": 40.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-0tkxpczj-ctx65536",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 9bgkvghg, cq1j63z1, fxnkfdkj, tzmnrcg3, vy20rvxa; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 537.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 537.6 25.9 23.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0tkxpczj",
   "source_url": "https://omlx.ai/benchmarks/performance/0tkxpczj",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-5idwaj24-ctx1024",
   "model": "Qwen3.5-35B-A3B-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 418.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B-Heretic on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 1k 418.3 55.5 20.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5idwaj24",
   "source_url": "https://omlx.ai/benchmarks/performance/5idwaj24",
   "tg_tokens": null,
   "tps": 55.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-5idwaj24-ctx4096",
   "model": "Qwen3.5-35B-A3B-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 439.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B-Heretic on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k 439.0 49.0 20.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5idwaj24",
   "source_url": "https://omlx.ai/benchmarks/performance/5idwaj24",
   "tg_tokens": null,
   "tps": 49.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-6xmkfdgf-ctx1024",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 210.4 GB at this context; duplicate traces of the same session: 4o3fa12n, k0jw9l2d; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 381.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 381.0 37.0 210.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6xmkfdgf",
   "source_url": "https://omlx.ai/benchmarks/performance/6xmkfdgf",
   "tg_tokens": null,
   "tps": 37.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-6xmkfdgf-ctx4096",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 210.7 GB at this context; duplicate traces of the same session: 4o3fa12n, k0jw9l2d; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 404.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 404.5 35.3 210.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6xmkfdgf",
   "source_url": "https://omlx.ai/benchmarks/performance/6xmkfdgf",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-6xmkfdgf-ctx8192",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 211.0 GB at this context; duplicate traces of the same session: 4o3fa12n, k0jw9l2d; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 400.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 400.7 31.3 211.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6xmkfdgf",
   "source_url": "https://omlx.ai/benchmarks/performance/6xmkfdgf",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-6xmkfdgf-ctx16384",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 211.8 GB at this context; duplicate traces of the same session: 4o3fa12n, k0jw9l2d; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 384.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 384.0 24.8 211.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6xmkfdgf",
   "source_url": "https://omlx.ai/benchmarks/performance/6xmkfdgf",
   "tg_tokens": null,
   "tps": 24.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-6xmkfdgf-ctx32768",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 213.2 GB at this context; duplicate traces of the same session: 4o3fa12n, k0jw9l2d; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 350.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 350.4 17.8 213.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6xmkfdgf",
   "source_url": "https://omlx.ai/benchmarks/performance/6xmkfdgf",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-6xmkfdgf-ctx65536",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 216.2 GB at this context; duplicate traces of the same session: 4o3fa12n, k0jw9l2d; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 294.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 294.3 11.7 216.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6xmkfdgf",
   "source_url": "https://omlx.ai/benchmarks/performance/6xmkfdgf",
   "tg_tokens": null,
   "tps": 11.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rb7hx514-ctx1024",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 302.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 454.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 1k 454.0 28.5 302.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rb7hx514",
   "source_url": "https://omlx.ai/benchmarks/performance/rb7hx514",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rb7hx514-ctx4096",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 302.9 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 487.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 4k 487.1 27.2 302.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rb7hx514",
   "source_url": "https://omlx.ai/benchmarks/performance/rb7hx514",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rb7hx514-ctx8192",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 303.3 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 484.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 8k 484.9 26.0 303.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rb7hx514",
   "source_url": "https://omlx.ai/benchmarks/performance/rb7hx514",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rb7hx514-ctx16384",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 304.0 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 462.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 16k 462.0 23.6 304.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rb7hx514",
   "source_url": "https://omlx.ai/benchmarks/performance/rb7hx514",
   "tg_tokens": null,
   "tps": 23.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rb7hx514-ctx32768",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 305.5 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 422.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 32k 422.5 20.2 305.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rb7hx514",
   "source_url": "https://omlx.ai/benchmarks/performance/rb7hx514",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-8owje5g3-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 292.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 292.6 53.8 4.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8owje5g3",
   "source_url": "https://omlx.ai/benchmarks/performance/8owje5g3",
   "tg_tokens": null,
   "tps": 53.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-8owje5g3-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 292.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 292.5 49.3 4.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8owje5g3",
   "source_url": "https://omlx.ai/benchmarks/performance/8owje5g3",
   "tg_tokens": null,
   "tps": 49.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-n7waxvmp-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.0 GB at this context; duplicate traces of the same session: 0ysklzxp, aeq42ap1, lsp40v2a, o7u1nsfl, yswmz05e; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 736.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 1k 736.4 102.2 4.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n7waxvmp",
   "source_url": "https://omlx.ai/benchmarks/performance/n7waxvmp",
   "tg_tokens": null,
   "tps": 102.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-n7waxvmp-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: 0ysklzxp, aeq42ap1, lsp40v2a, o7u1nsfl, yswmz05e; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 734.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k 734.5 95.4 4.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n7waxvmp",
   "source_url": "https://omlx.ai/benchmarks/performance/n7waxvmp",
   "tg_tokens": null,
   "tps": 95.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-n7waxvmp-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.6 GB at this context; duplicate traces of the same session: 0ysklzxp, aeq42ap1, lsp40v2a, o7u1nsfl, yswmz05e; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 720.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 8k 720.3 87.0 4.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n7waxvmp",
   "source_url": "https://omlx.ai/benchmarks/performance/n7waxvmp",
   "tg_tokens": null,
   "tps": 87.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-n7waxvmp-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 5.3 GB at this context; duplicate traces of the same session: 0ysklzxp, aeq42ap1, lsp40v2a, o7u1nsfl, yswmz05e; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 691.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 16k 691.3 73.4 5.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n7waxvmp",
   "source_url": "https://omlx.ai/benchmarks/performance/n7waxvmp",
   "tg_tokens": null,
   "tps": 73.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-n7waxvmp-ctx32768",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: 0ysklzxp, aeq42ap1, lsp40v2a, o7u1nsfl, yswmz05e; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 576.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 32k 576.0 28.6 6.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n7waxvmp",
   "source_url": "https://omlx.ai/benchmarks/performance/n7waxvmp",
   "tg_tokens": null,
   "tps": 28.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-n7waxvmp-ctx65536",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 8.6 GB at this context; duplicate traces of the same session: 0ysklzxp, aeq42ap1, lsp40v2a, o7u1nsfl, yswmz05e; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 468.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 64k 468.7 18.2 8.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n7waxvmp",
   "source_url": "https://omlx.ai/benchmarks/performance/n7waxvmp",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-qni6te2s-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.0 GB at this context; duplicate traces of the same session: 0oga1cf0, j290pxfi, xckibqvi, zwuvz3cp; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1428.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 1k 1,428 137.5 4.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qni6te2s",
   "source_url": "https://omlx.ai/benchmarks/performance/qni6te2s",
   "tg_tokens": null,
   "tps": 137.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-qni6te2s-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: 0oga1cf0, j290pxfi, xckibqvi, zwuvz3cp; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1440.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 4k 1,440 128.2 4.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qni6te2s",
   "source_url": "https://omlx.ai/benchmarks/performance/qni6te2s",
   "tg_tokens": null,
   "tps": 128.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-qni6te2s-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.6 GB at this context; duplicate traces of the same session: 0oga1cf0, j290pxfi, xckibqvi, zwuvz3cp; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1414.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 8k 1,414 117.4 4.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qni6te2s",
   "source_url": "https://omlx.ai/benchmarks/performance/qni6te2s",
   "tg_tokens": null,
   "tps": 117.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-qni6te2s-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 5.3 GB at this context; duplicate traces of the same session: 0oga1cf0, j290pxfi, xckibqvi, zwuvz3cp; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1355.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 16k 1,355 98.7 5.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qni6te2s",
   "source_url": "https://omlx.ai/benchmarks/performance/qni6te2s",
   "tg_tokens": null,
   "tps": 98.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-qni6te2s-ctx32768",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: 0oga1cf0, j290pxfi, xckibqvi, zwuvz3cp; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1247.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 32k 1,247 74.5 6.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qni6te2s",
   "source_url": "https://omlx.ai/benchmarks/performance/qni6te2s",
   "tg_tokens": null,
   "tps": 74.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-qni6te2s-ctx65536",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 8.6 GB at this context; duplicate traces of the same session: 0oga1cf0, j290pxfi, xckibqvi, zwuvz3cp; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1051.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 64k 1,051 50.4 8.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qni6te2s",
   "source_url": "https://omlx.ai/benchmarks/performance/qni6te2s",
   "tg_tokens": null,
   "tps": 50.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-p2mm6l9j-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: 4vf3fhkn, 5t6c2eu1, cgvsvi3i, u49qccjp; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1423.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 1k 1,423 99.9 5.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p2mm6l9j",
   "source_url": "https://omlx.ai/benchmarks/performance/p2mm6l9j",
   "tg_tokens": null,
   "tps": 99.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-p2mm6l9j-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 6.1 GB at this context; duplicate traces of the same session: 4vf3fhkn, 5t6c2eu1, cgvsvi3i, u49qccjp; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1448.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 4k 1,448 96.2 6.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p2mm6l9j",
   "source_url": "https://omlx.ai/benchmarks/performance/p2mm6l9j",
   "tg_tokens": null,
   "tps": 96.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-p2mm6l9j-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: 4vf3fhkn, 5t6c2eu1, cgvsvi3i, u49qccjp; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1421.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 8k 1,421 90.1 6.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p2mm6l9j",
   "source_url": "https://omlx.ai/benchmarks/performance/p2mm6l9j",
   "tg_tokens": null,
   "tps": 90.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-p2mm6l9j-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 7.3 GB at this context; duplicate traces of the same session: 4vf3fhkn, 5t6c2eu1, cgvsvi3i, u49qccjp; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1362.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 16k 1,362 80.1 7.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p2mm6l9j",
   "source_url": "https://omlx.ai/benchmarks/performance/p2mm6l9j",
   "tg_tokens": null,
   "tps": 80.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-p2mm6l9j-ctx32768",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 7.9 GB at this context; duplicate traces of the same session: 4vf3fhkn, 5t6c2eu1, cgvsvi3i, u49qccjp; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1255.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 32k 1,255 63.0 7.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p2mm6l9j",
   "source_url": "https://omlx.ai/benchmarks/performance/p2mm6l9j",
   "tg_tokens": null,
   "tps": 63.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-p2mm6l9j-ctx65536",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 9.9 GB at this context; duplicate traces of the same session: 4vf3fhkn, 5t6c2eu1, cgvsvi3i, u49qccjp; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1056.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 64k 1,056 45.5 9.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p2mm6l9j",
   "source_url": "https://omlx.ai/benchmarks/performance/p2mm6l9j",
   "tg_tokens": null,
   "tps": 45.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-9ylnou44-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: b5hno3da; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 396.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 396.6 36.4 3.9 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ylnou44",
   "source_url": "https://omlx.ai/benchmarks/performance/9ylnou44",
   "tg_tokens": null,
   "tps": 36.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-9ylnou44-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: b5hno3da; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 393.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 393.2 34.2 4.2 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ylnou44",
   "source_url": "https://omlx.ai/benchmarks/performance/9ylnou44",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-4wxdnprn-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: dnohwwb2; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 407.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 407.7 38.6 3.9 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4wxdnprn",
   "source_url": "https://omlx.ai/benchmarks/performance/4wxdnprn",
   "tg_tokens": null,
   "tps": 38.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-4wxdnprn-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: dnohwwb2; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 405.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 405.2 36.6 4.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4wxdnprn",
   "source_url": "https://omlx.ai/benchmarks/performance/4wxdnprn",
   "tg_tokens": null,
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-a6wnmd9u-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 378.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 378.3 35.1 3.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a6wnmd9u",
   "source_url": "https://omlx.ai/benchmarks/performance/a6wnmd9u",
   "tg_tokens": null,
   "tps": 35.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-a6wnmd9u-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 380.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 380.6 34.1 4.2 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a6wnmd9u",
   "source_url": "https://omlx.ai/benchmarks/performance/a6wnmd9u",
   "tg_tokens": null,
   "tps": 34.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-89qqx7p1-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.0 GB at this context; duplicate traces of the same session: c1jz32zu, h7hh9kvw, j3udbnrh, nvxbhp1x, qsyh8wbm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 811.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 811.1 82.1 4.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/89qqx7p1",
   "source_url": "https://omlx.ai/benchmarks/performance/89qqx7p1",
   "tg_tokens": null,
   "tps": 82.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-89qqx7p1-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: c1jz32zu, h7hh9kvw, j3udbnrh, nvxbhp1x, qsyh8wbm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 798.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 798.0 76.8 4.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/89qqx7p1",
   "source_url": "https://omlx.ai/benchmarks/performance/89qqx7p1",
   "tg_tokens": null,
   "tps": 76.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-89qqx7p1-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.6 GB at this context; duplicate traces of the same session: c1jz32zu, h7hh9kvw, j3udbnrh, nvxbhp1x, qsyh8wbm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 781.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 781.9 72.4 4.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/89qqx7p1",
   "source_url": "https://omlx.ai/benchmarks/performance/89qqx7p1",
   "tg_tokens": null,
   "tps": 72.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-89qqx7p1-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 5.3 GB at this context; duplicate traces of the same session: c1jz32zu, h7hh9kvw, j3udbnrh, nvxbhp1x, qsyh8wbm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 737.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 737.3 62.8 5.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/89qqx7p1",
   "source_url": "https://omlx.ai/benchmarks/performance/89qqx7p1",
   "tg_tokens": null,
   "tps": 62.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-89qqx7p1-ctx32768",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: c1jz32zu, h7hh9kvw, j3udbnrh, nvxbhp1x, qsyh8wbm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 655.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 655.0 49.9 6.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/89qqx7p1",
   "source_url": "https://omlx.ai/benchmarks/performance/89qqx7p1",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-89qqx7p1-ctx65536",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 8.6 GB at this context; duplicate traces of the same session: c1jz32zu, h7hh9kvw, j3udbnrh, nvxbhp1x, qsyh8wbm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 534.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 534.5 36.3 8.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/89qqx7p1",
   "source_url": "https://omlx.ai/benchmarks/performance/89qqx7p1",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-g6l1twvq-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "OptiQ build; oMLX community benchmark trace; peak memory 4.8 GB at this context; duplicate traces of the same session: o4cjy9j3; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 254.9,
   "provenance": "community",
   "quant": "OptiQ",
   "quote": "Qwen3.5-4B-OptiQ on M4 (8c); M4 (8c) \u00b7 16 GB \u00b7 4bit; 1k 254.9 23.3 4.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g6l1twvq",
   "source_url": "https://omlx.ai/benchmarks/performance/g6l1twvq",
   "tg_tokens": null,
   "tps": 23.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-g6l1twvq-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "OptiQ build; oMLX community benchmark trace; peak memory 5.4 GB at this context; duplicate traces of the same session: o4cjy9j3; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 263.4,
   "provenance": "community",
   "quant": "OptiQ",
   "quote": "Qwen3.5-4B-OptiQ on M4 (8c); M4 (8c) \u00b7 16 GB \u00b7 4bit; 4k 263.4 22.1 5.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g6l1twvq",
   "source_url": "https://omlx.ai/benchmarks/performance/g6l1twvq",
   "tg_tokens": null,
   "tps": 22.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ob8t8tj4-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.7 GB at this context; duplicate traces of the same session: 1n8qgmq9, 5fjbahy5, 6rngmu2d, hyclrwcc; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 239.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 8bit; 1k 239.3 35.1 10.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ob8t8tj4",
   "source_url": "https://omlx.ai/benchmarks/performance/ob8t8tj4",
   "tg_tokens": null,
   "tps": 35.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ob8t8tj4-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.0 GB at this context; duplicate traces of the same session: 1n8qgmq9, 5fjbahy5, 6rngmu2d, hyclrwcc; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 242.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 8bit; 4k 242.1 33.3 11.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ob8t8tj4",
   "source_url": "https://omlx.ai/benchmarks/performance/ob8t8tj4",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ob8t8tj4-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.2 GB at this context; duplicate traces of the same session: 1n8qgmq9, 5fjbahy5, 6rngmu2d, hyclrwcc; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 236.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 8bit; 8k 236.2 30.3 11.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ob8t8tj4",
   "source_url": "https://omlx.ai/benchmarks/performance/ob8t8tj4",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ob8t8tj4-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.7 GB at this context; duplicate traces of the same session: 1n8qgmq9, 5fjbahy5, 6rngmu2d, hyclrwcc; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 234.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 8bit; 16k 234.4 26.2 11.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ob8t8tj4",
   "source_url": "https://omlx.ai/benchmarks/performance/ob8t8tj4",
   "tg_tokens": null,
   "tps": 26.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ob8t8tj4-ctx32768",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: 1n8qgmq9, 5fjbahy5, 6rngmu2d, hyclrwcc; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 8bit; 32k 224.8 20.3 12.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ob8t8tj4",
   "source_url": "https://omlx.ai/benchmarks/performance/ob8t8tj4",
   "tg_tokens": null,
   "tps": 20.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ob8t8tj4-ctx65536",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: 1n8qgmq9, 5fjbahy5, 6rngmu2d, hyclrwcc; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 8bit; 64k 205.0 13.9 14.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ob8t8tj4",
   "source_url": "https://omlx.ai/benchmarks/performance/ob8t8tj4",
   "tg_tokens": null,
   "tps": 13.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-4umrbkhz-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.7 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 239.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 8bit; 1k 239.0 35.3 10.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4umrbkhz",
   "source_url": "https://omlx.ai/benchmarks/performance/4umrbkhz",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-4umrbkhz-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.0 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 241.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 8bit; 4k 241.2 33.4 11.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4umrbkhz",
   "source_url": "https://omlx.ai/benchmarks/performance/4umrbkhz",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ssh6dgpy-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: 6kpz3oma; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 308.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 308.2 55.1 6.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ssh6dgpy",
   "source_url": "https://omlx.ai/benchmarks/performance/ssh6dgpy",
   "tg_tokens": null,
   "tps": 55.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ssh6dgpy-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: 6kpz3oma; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 311.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 311.9 53.2 7.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ssh6dgpy",
   "source_url": "https://omlx.ai/benchmarks/performance/ssh6dgpy",
   "tg_tokens": null,
   "tps": 53.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-iha3k32k-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 158.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 158.6 32.3 6.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iha3k32k",
   "source_url": "https://omlx.ai/benchmarks/performance/iha3k32k",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-iha3k32k-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 159.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 159.3 31.5 7.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iha3k32k",
   "source_url": "https://omlx.ai/benchmarks/performance/iha3k32k",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-fykkivoz-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: vfdq12ip; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 401.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 1k 401.8 58.2 6.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fykkivoz",
   "source_url": "https://omlx.ai/benchmarks/performance/fykkivoz",
   "tg_tokens": null,
   "tps": 58.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-fykkivoz-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.3 GB at this context; duplicate traces of the same session: vfdq12ip; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 408.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k 408.3 53.6 7.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fykkivoz",
   "source_url": "https://omlx.ai/benchmarks/performance/fykkivoz",
   "tg_tokens": null,
   "tps": 53.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-t8kof1zr-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.7 GB at this context; duplicate traces of the same session: a6fmmdir, xpnt88fb; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 775.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 1k 775.8 74.3 8.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t8kof1zr",
   "source_url": "https://omlx.ai/benchmarks/performance/t8kof1zr",
   "tg_tokens": null,
   "tps": 74.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-t8kof1zr-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.9 GB at this context; duplicate traces of the same session: a6fmmdir, xpnt88fb; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 783.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 4k 783.9 72.1 8.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t8kof1zr",
   "source_url": "https://omlx.ai/benchmarks/performance/t8kof1zr",
   "tg_tokens": null,
   "tps": 72.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-t8kof1zr-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 9.2 GB at this context; duplicate traces of the same session: a6fmmdir, xpnt88fb; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 776.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 8k 776.7 68.7 9.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t8kof1zr",
   "source_url": "https://omlx.ai/benchmarks/performance/t8kof1zr",
   "tg_tokens": null,
   "tps": 68.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-t8kof1zr-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 9.6 GB at this context; duplicate traces of the same session: a6fmmdir, xpnt88fb; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 758.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 16k 758.8 62.8 9.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t8kof1zr",
   "source_url": "https://omlx.ai/benchmarks/performance/t8kof1zr",
   "tg_tokens": null,
   "tps": 62.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-t8kof1zr-ctx32768",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.6 GB at this context; duplicate traces of the same session: a6fmmdir, xpnt88fb; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 724.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 32k 724.1 52.6 10.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t8kof1zr",
   "source_url": "https://omlx.ai/benchmarks/performance/t8kof1zr",
   "tg_tokens": null,
   "tps": 52.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-t8kof1zr-ctx65536",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: a6fmmdir, xpnt88fb; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 653.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 64k 653.8 39.8 12.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t8kof1zr",
   "source_url": "https://omlx.ai/benchmarks/performance/t8kof1zr",
   "tg_tokens": null,
   "tps": 39.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-mm823kic-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.7 GB at this context; duplicate traces of the same session: 6o6zm8tl, 8g093y7c, voj8az1w, wxd1vexu; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 794.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 1k 794.0 64.3 10.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mm823kic",
   "source_url": "https://omlx.ai/benchmarks/performance/mm823kic",
   "tg_tokens": null,
   "tps": 64.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-mm823kic-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.0 GB at this context; duplicate traces of the same session: 6o6zm8tl, 8g093y7c, voj8az1w, wxd1vexu; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 804.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 4k 804.0 62.6 11.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mm823kic",
   "source_url": "https://omlx.ai/benchmarks/performance/mm823kic",
   "tg_tokens": null,
   "tps": 62.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-mm823kic-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.2 GB at this context; duplicate traces of the same session: 6o6zm8tl, 8g093y7c, voj8az1w, wxd1vexu; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 797.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 8k 797.4 59.7 11.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mm823kic",
   "source_url": "https://omlx.ai/benchmarks/performance/mm823kic",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-mm823kic-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.7 GB at this context; duplicate traces of the same session: 6o6zm8tl, 8g093y7c, voj8az1w, wxd1vexu; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 778.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 16k 778.7 54.0 11.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mm823kic",
   "source_url": "https://omlx.ai/benchmarks/performance/mm823kic",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-mm823kic-ctx32768",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: 6o6zm8tl, 8g093y7c, voj8az1w, wxd1vexu; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 742.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 32k 742.0 46.4 12.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mm823kic",
   "source_url": "https://omlx.ai/benchmarks/performance/mm823kic",
   "tg_tokens": null,
   "tps": 46.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-mm823kic-ctx65536",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: 6o6zm8tl, 8g093y7c, voj8az1w, wxd1vexu; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 668.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 64k 668.6 33.9 14.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mm823kic",
   "source_url": "https://omlx.ai/benchmarks/performance/mm823kic",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-7ei9xvbp-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: jvcibpti; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 222.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 222.1 20.7 6.5 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ei9xvbp",
   "source_url": "https://omlx.ai/benchmarks/performance/7ei9xvbp",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-7ei9xvbp-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: jvcibpti; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 223.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 223.2 20.1 6.8 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ei9xvbp",
   "source_url": "https://omlx.ai/benchmarks/performance/7ei9xvbp",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-tr0839a7-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 6.5 GB; TTFT 4614 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 222.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k tokens 222.0 PP tok/s 20.6 TG tok/s 4614 TTFT (ms) 6.5 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tr0839a7",
   "source_url": "https://omlx.ai/benchmarks/performance/tr0839a7",
   "tg_tokens": null,
   "tps": 20.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-43qf0gin-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 6.5 GB; TTFT 4609 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 222.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k tokens 222.2 PP tok/s 20.7 TG tok/s 4609 TTFT (ms) 6.5 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/43qf0gin",
   "source_url": "https://omlx.ai/benchmarks/performance/43qf0gin",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-hqgt9fac-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 878.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 878.9 52.1 10.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hqgt9fac",
   "source_url": "https://omlx.ai/benchmarks/performance/hqgt9fac",
   "tg_tokens": null,
   "tps": 52.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-hqgt9fac-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 890.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 890.4 50.9 11.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hqgt9fac",
   "source_url": "https://omlx.ai/benchmarks/performance/hqgt9fac",
   "tg_tokens": null,
   "tps": 50.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-qe1o4j3u-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 445.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 445.0 48.9 6.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qe1o4j3u",
   "source_url": "https://omlx.ai/benchmarks/performance/qe1o4j3u",
   "tg_tokens": null,
   "tps": 48.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-qe1o4j3u-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 447.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 447.3 47.6 7.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qe1o4j3u",
   "source_url": "https://omlx.ai/benchmarks/performance/qe1o4j3u",
   "tg_tokens": null,
   "tps": 47.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-e2bpzn48-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: 381hens8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 414.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 414.7 44.5 6.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e2bpzn48",
   "source_url": "https://omlx.ai/benchmarks/performance/e2bpzn48",
   "tg_tokens": null,
   "tps": 44.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-e2bpzn48-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: 381hens8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 411.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 411.9 43.2 7.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e2bpzn48",
   "source_url": "https://omlx.ai/benchmarks/performance/e2bpzn48",
   "tg_tokens": null,
   "tps": 43.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wufyjzmw-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 928.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 928.8 33.2 16.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wufyjzmw",
   "source_url": "https://omlx.ai/benchmarks/performance/wufyjzmw",
   "tg_tokens": null,
   "tps": 33.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wufyjzmw-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1008.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,008 32.8 17.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wufyjzmw",
   "source_url": "https://omlx.ai/benchmarks/performance/wufyjzmw",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wufyjzmw-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 878.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 878.4 32.0 17.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wufyjzmw",
   "source_url": "https://omlx.ai/benchmarks/performance/wufyjzmw",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wufyjzmw-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 797.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 797.7 30.8 17.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wufyjzmw",
   "source_url": "https://omlx.ai/benchmarks/performance/wufyjzmw",
   "tg_tokens": null,
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-nme46ce0-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: bqy4j59f, hhxc5sho, j0dficgz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 893.6,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 1k 893.6 38.2 13.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nme46ce0",
   "source_url": "https://omlx.ai/benchmarks/performance/nme46ce0",
   "tg_tokens": null,
   "tps": 38.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-nme46ce0-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: bqy4j59f, hhxc5sho, j0dficgz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 935.7,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 4k 935.7 38.0 14.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nme46ce0",
   "source_url": "https://omlx.ai/benchmarks/performance/nme46ce0",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-nme46ce0-ctx8192",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: bqy4j59f, hhxc5sho, j0dficgz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 852.6,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 8k 852.6 35.3 14.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nme46ce0",
   "source_url": "https://omlx.ai/benchmarks/performance/nme46ce0",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-nme46ce0-ctx16384",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 15.3 GB at this context; duplicate traces of the same session: bqy4j59f, hhxc5sho, j0dficgz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 735.4,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 16k 735.4 32.5 15.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nme46ce0",
   "source_url": "https://omlx.ai/benchmarks/performance/nme46ce0",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-nme46ce0-ctx32768",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: bqy4j59f, hhxc5sho, j0dficgz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 699.2,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 32k 699.2 32.1 16.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nme46ce0",
   "source_url": "https://omlx.ai/benchmarks/performance/nme46ce0",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-nme46ce0-ctx65536",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: bqy4j59f, hhxc5sho, j0dficgz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 680.6,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 64k 680.6 27.9 18.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nme46ce0",
   "source_url": "https://omlx.ai/benchmarks/performance/nme46ce0",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-nme46ce0-ctx131072",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: bqy4j59f, hhxc5sho, j0dficgz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 549.4,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 128k 549.4 21.5 22.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nme46ce0",
   "source_url": "https://omlx.ai/benchmarks/performance/nme46ce0",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-fra5t4az-ctx1024",
   "model": "Gemma 3 4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 3.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 901.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-3-4b-it-qat on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 901.8 34.9 3.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fra5t4az",
   "source_url": "https://omlx.ai/benchmarks/performance/fra5t4az",
   "tg_tokens": null,
   "tps": 34.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-fra5t4az-ctx4096",
   "model": "Gemma 3 4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 889.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-3-4b-it-qat on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 889.9 80.2 3.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fra5t4az",
   "source_url": "https://omlx.ai/benchmarks/performance/fra5t4az",
   "tg_tokens": null,
   "tps": 80.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-nqo8pz28-ctx1024",
   "model": "gpt-oss 120B",
   "notes": "oMLX community benchmark trace; peak memory 58.0 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 187.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gpt-oss-120b on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 187.7 54.9 58.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nqo8pz28",
   "source_url": "https://omlx.ai/benchmarks/performance/nqo8pz28",
   "tg_tokens": null,
   "tps": 54.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-nqo8pz28-ctx4096",
   "model": "gpt-oss 120B",
   "notes": "oMLX community benchmark trace; peak memory 58.1 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 182.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gpt-oss-120b on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 182.0 44.7 58.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nqo8pz28",
   "source_url": "https://omlx.ai/benchmarks/performance/nqo8pz28",
   "tg_tokens": null,
   "tps": 44.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-nqo8pz28-ctx8192",
   "model": "gpt-oss 120B",
   "notes": "oMLX community benchmark trace; peak memory 58.3 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 183.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gpt-oss-120b on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 183.9 37.1 58.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nqo8pz28",
   "source_url": "https://omlx.ai/benchmarks/performance/nqo8pz28",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-nqo8pz28-ctx16384",
   "model": "gpt-oss 120B",
   "notes": "oMLX community benchmark trace; peak memory 58.6 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 177.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gpt-oss-120b on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 177.1 22.7 58.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nqo8pz28",
   "source_url": "https://omlx.ai/benchmarks/performance/nqo8pz28",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-nqo8pz28-ctx32768",
   "model": "gpt-oss 120B",
   "notes": "oMLX community benchmark trace; peak memory 59.2 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 165.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gpt-oss-120b on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 32k 165.3 14.6 59.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nqo8pz28",
   "source_url": "https://omlx.ai/benchmarks/performance/nqo8pz28",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-wjn65xp9-ctx1024",
   "model": "gpt-oss 120B",
   "notes": "MXFP4 experts with Q8 attention; oMLX community benchmark trace; peak memory 59.2 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 613.2,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-120b-MXFP4-Q8 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 613.2 68.4 59.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wjn65xp9",
   "source_url": "https://omlx.ai/benchmarks/performance/wjn65xp9",
   "tg_tokens": null,
   "tps": 68.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-wjn65xp9-ctx4096",
   "model": "gpt-oss 120B",
   "notes": "MXFP4 experts with Q8 attention; oMLX community benchmark trace; peak memory 59.3 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 685.5,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-120b-MXFP4-Q8 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 685.5 64.1 59.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wjn65xp9",
   "source_url": "https://omlx.ai/benchmarks/performance/wjn65xp9",
   "tg_tokens": null,
   "tps": 64.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-wjn65xp9-ctx8192",
   "model": "gpt-oss 120B",
   "notes": "MXFP4 experts with Q8 attention; oMLX community benchmark trace; peak memory 59.5 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 687.8,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-120b-MXFP4-Q8 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 687.8 57.8 59.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wjn65xp9",
   "source_url": "https://omlx.ai/benchmarks/performance/wjn65xp9",
   "tg_tokens": null,
   "tps": 57.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-wjn65xp9-ctx16384",
   "model": "gpt-oss 120B",
   "notes": "MXFP4 experts with Q8 attention; oMLX community benchmark trace; peak memory 59.8 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 663.6,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-120b-MXFP4-Q8 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 663.6 47.5 59.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wjn65xp9",
   "source_url": "https://omlx.ai/benchmarks/performance/wjn65xp9",
   "tg_tokens": null,
   "tps": 47.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-wjn65xp9-ctx32768",
   "model": "gpt-oss 120B",
   "notes": "MXFP4 experts with Q8 attention; oMLX community benchmark trace; peak memory 60.4 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 600.7,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-120b-MXFP4-Q8 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 600.7 35.2 60.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wjn65xp9",
   "source_url": "https://omlx.ai/benchmarks/performance/wjn65xp9",
   "tg_tokens": null,
   "tps": 35.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-c510mcpw-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 106.0 15.6 18.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c510mcpw",
   "source_url": "https://omlx.ai/benchmarks/performance/c510mcpw",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-c510mcpw-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 151.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 151.0 15.0 19.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c510mcpw",
   "source_url": "https://omlx.ai/benchmarks/performance/c510mcpw",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-c510mcpw-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 151.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 151.6 14.5 19.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c510mcpw",
   "source_url": "https://omlx.ai/benchmarks/performance/c510mcpw",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-c510mcpw-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 131.6 12.8 21.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c510mcpw",
   "source_url": "https://omlx.ai/benchmarks/performance/c510mcpw",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-b1sluf0m-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "pyros-vault MTP build; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: hj41ahwa, r2t9yutl, ziaa5bwx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 773.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "pyros-vault:Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 773.5 34.2 33.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b1sluf0m",
   "source_url": "https://omlx.ai/benchmarks/performance/b1sluf0m",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-b1sluf0m-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "pyros-vault MTP build; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: hj41ahwa, r2t9yutl, ziaa5bwx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 850.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "pyros-vault:Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 850.0 39.2 34.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b1sluf0m",
   "source_url": "https://omlx.ai/benchmarks/performance/b1sluf0m",
   "tg_tokens": null,
   "tps": 39.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-b1sluf0m-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "pyros-vault MTP build; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: hj41ahwa, r2t9yutl, ziaa5bwx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 823.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "pyros-vault:Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 823.6 50.2 35.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b1sluf0m",
   "source_url": "https://omlx.ai/benchmarks/performance/b1sluf0m",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-b1sluf0m-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "pyros-vault MTP build; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: hj41ahwa, r2t9yutl, ziaa5bwx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 752.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "pyros-vault:Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 752.8 53.2 35.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b1sluf0m",
   "source_url": "https://omlx.ai/benchmarks/performance/b1sluf0m",
   "tg_tokens": null,
   "tps": 53.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-b1sluf0m-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "pyros-vault MTP build; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: hj41ahwa, r2t9yutl, ziaa5bwx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 709.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "pyros-vault:Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 709.2 55.2 36.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b1sluf0m",
   "source_url": "https://omlx.ai/benchmarks/performance/b1sluf0m",
   "tg_tokens": null,
   "tps": 55.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-b1sluf0m-ctx65536",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "pyros-vault MTP build; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: hj41ahwa, r2t9yutl, ziaa5bwx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 617.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "pyros-vault:Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 617.3 41.5 39.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b1sluf0m",
   "source_url": "https://omlx.ai/benchmarks/performance/b1sluf0m",
   "tg_tokens": null,
   "tps": 41.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-b1sluf0m-ctx131072",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "pyros-vault MTP build; oMLX community benchmark trace; peak memory 83.2 GB at this context; duplicate traces of the same session: hj41ahwa, r2t9yutl, ziaa5bwx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 329.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "pyros-vault:Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 329.1 22.7 83.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b1sluf0m",
   "source_url": "https://omlx.ai/benchmarks/performance/b1sluf0m",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-0q5i6fy9-ctx1024",
   "model": "GLM-4.7-Flash",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 608.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "GLM-4.7-Flash on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 608.4 49.8 23.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0q5i6fy9",
   "source_url": "https://omlx.ai/benchmarks/performance/0q5i6fy9",
   "tg_tokens": null,
   "tps": 49.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-0q5i6fy9-ctx4096",
   "model": "GLM-4.7-Flash",
   "notes": "oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 546.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "GLM-4.7-Flash on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 546.4 42.5 24.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0q5i6fy9",
   "source_url": "https://omlx.ai/benchmarks/performance/0q5i6fy9",
   "tg_tokens": null,
   "tps": 42.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xj3z6y6p-ctx1024",
   "model": "GLM-5",
   "notes": "oMLX community benchmark trace; peak memory 420.0 GB at this context; duplicate traces of the same session: teqbm0rb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 188.7,
   "provenance": "community",
   "quant": "4.8bit",
   "quote": "GLM-5-4.8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 188.7 16.6 420.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xj3z6y6p",
   "source_url": "https://omlx.ai/benchmarks/performance/xj3z6y6p",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xj3z6y6p-ctx4096",
   "model": "GLM-5",
   "notes": "oMLX community benchmark trace; peak memory 422.3 GB at this context; duplicate traces of the same session: teqbm0rb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 182.6,
   "provenance": "community",
   "quant": "4.8bit",
   "quote": "GLM-5-4.8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 182.6 13.7 422.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xj3z6y6p",
   "source_url": "https://omlx.ai/benchmarks/performance/xj3z6y6p",
   "tg_tokens": null,
   "tps": 13.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xj3z6y6p-ctx8192",
   "model": "GLM-5",
   "notes": "oMLX community benchmark trace; peak memory 424.9 GB at this context; duplicate traces of the same session: teqbm0rb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 156.3,
   "provenance": "community",
   "quant": "4.8bit",
   "quote": "GLM-5-4.8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 156.3 13.5 424.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xj3z6y6p",
   "source_url": "https://omlx.ai/benchmarks/performance/xj3z6y6p",
   "tg_tokens": null,
   "tps": 13.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xj3z6y6p-ctx16384",
   "model": "GLM-5",
   "notes": "oMLX community benchmark trace; peak memory 430.9 GB at this context; duplicate traces of the same session: teqbm0rb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 119.8,
   "provenance": "community",
   "quant": "4.8bit",
   "quote": "GLM-5-4.8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 119.8 13.0 430.9 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xj3z6y6p",
   "source_url": "https://omlx.ai/benchmarks/performance/xj3z6y6p",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xj3z6y6p-ctx32768",
   "model": "GLM-5",
   "notes": "oMLX community benchmark trace; peak memory 443.6 GB at this context; duplicate traces of the same session: teqbm0rb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 73.2,
   "provenance": "community",
   "quant": "4.8bit",
   "quote": "GLM-5-4.8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 73.2 12.3 443.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xj3z6y6p",
   "source_url": "https://omlx.ai/benchmarks/performance/xj3z6y6p",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-qt1g5q7f-ctx1024",
   "model": "HY-MT1.5-1.8B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 3.7 GB; TTFT 1122 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 912.9,
   "provenance": "community",
   "quant": "unknown",
   "quote": "HY-MT1.5-1.8B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 unknown; 1k tokens 912.9 PP tok/s 26.5 TG tok/s 1122 TTFT (ms) 3.7 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qt1g5q7f",
   "source_url": "https://omlx.ai/benchmarks/performance/qt1g5q7f",
   "tg_tokens": null,
   "tps": 26.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xbfycdsw-ctx1024",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 174.0 GB at this context; duplicate traces of the same session: 2we8lcjq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 496.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 1k 496.7 40.0 174.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xbfycdsw",
   "source_url": "https://omlx.ai/benchmarks/performance/xbfycdsw",
   "tg_tokens": null,
   "tps": 40.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xbfycdsw-ctx4096",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 175.0 GB at this context; duplicate traces of the same session: 2we8lcjq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 554.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 4k 554.8 33.9 175.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xbfycdsw",
   "source_url": "https://omlx.ai/benchmarks/performance/xbfycdsw",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xbfycdsw-ctx8192",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 175.9 GB at this context; duplicate traces of the same session: 2we8lcjq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 514.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 8k 514.0 28.9 175.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xbfycdsw",
   "source_url": "https://omlx.ai/benchmarks/performance/xbfycdsw",
   "tg_tokens": null,
   "tps": 28.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xbfycdsw-ctx16384",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 177.8 GB at this context; duplicate traces of the same session: 2we8lcjq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 436.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 16k 436.3 23.5 177.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xbfycdsw",
   "source_url": "https://omlx.ai/benchmarks/performance/xbfycdsw",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xbfycdsw-ctx32768",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 181.7 GB at this context; duplicate traces of the same session: 2we8lcjq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 330.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 32k 330.8 16.1 181.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xbfycdsw",
   "source_url": "https://omlx.ai/benchmarks/performance/xbfycdsw",
   "tg_tokens": null,
   "tps": 16.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xbfycdsw-ctx65536",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 189.6 GB at this context; duplicate traces of the same session: 2we8lcjq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 221.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 64k 221.1 9.6 189.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xbfycdsw",
   "source_url": "https://omlx.ai/benchmarks/performance/xbfycdsw",
   "tg_tokens": null,
   "tps": 9.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-bgwyirz3-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 61.6 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 751.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 1k 751.3 57.9 61.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bgwyirz3",
   "source_url": "https://omlx.ai/benchmarks/performance/bgwyirz3",
   "tg_tokens": null,
   "tps": 57.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-bgwyirz3-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 61.7 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 966.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 4k 966.4 56.2 61.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bgwyirz3",
   "source_url": "https://omlx.ai/benchmarks/performance/bgwyirz3",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-bgwyirz3-ctx8192",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 62.0 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 986.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 8k 986.0 53.4 62.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bgwyirz3",
   "source_url": "https://omlx.ai/benchmarks/performance/bgwyirz3",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-bgwyirz3-ctx16384",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 62.4 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 956.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 16k 956.5 47.0 62.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bgwyirz3",
   "source_url": "https://omlx.ai/benchmarks/performance/bgwyirz3",
   "tg_tokens": null,
   "tps": 47.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-bgwyirz3-ctx32768",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 63.3 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 881.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 32k 881.3 39.1 63.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bgwyirz3",
   "source_url": "https://omlx.ai/benchmarks/performance/bgwyirz3",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-bgwyirz3-ctx65536",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 65.1 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 731.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 64k 731.4 28.9 65.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bgwyirz3",
   "source_url": "https://omlx.ai/benchmarks/performance/bgwyirz3",
   "tg_tokens": null,
   "tps": 28.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rihr0r0i-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.0 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1298.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 1,298 73.9 43.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rihr0r0i",
   "source_url": "https://omlx.ai/benchmarks/performance/rihr0r0i",
   "tg_tokens": null,
   "tps": 73.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rihr0r0i-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.2 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1558.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 1,558 71.0 43.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rihr0r0i",
   "source_url": "https://omlx.ai/benchmarks/performance/rihr0r0i",
   "tg_tokens": null,
   "tps": 71.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rihr0r0i-ctx8192",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.4 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1552.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,552 66.4 43.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rihr0r0i",
   "source_url": "https://omlx.ai/benchmarks/performance/rihr0r0i",
   "tg_tokens": null,
   "tps": 66.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rihr0r0i-ctx16384",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.9 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1461.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 1,461 58.9 43.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rihr0r0i",
   "source_url": "https://omlx.ai/benchmarks/performance/rihr0r0i",
   "tg_tokens": null,
   "tps": 58.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rihr0r0i-ctx32768",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 44.7 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1281.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 1,281 49.4 44.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rihr0r0i",
   "source_url": "https://omlx.ai/benchmarks/performance/rihr0r0i",
   "tg_tokens": null,
   "tps": 49.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rihr0r0i-ctx65536",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 46.5 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1007.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 1,007 36.9 46.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rihr0r0i",
   "source_url": "https://omlx.ai/benchmarks/performance/rihr0r0i",
   "tg_tokens": null,
   "tps": 36.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xill9rj9-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 80.1 GB at this context; duplicate traces of the same session: rhiba0g8; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1236.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 1,236 63.9 80.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xill9rj9",
   "source_url": "https://omlx.ai/benchmarks/performance/xill9rj9",
   "tg_tokens": null,
   "tps": 63.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xill9rj9-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 80.3 GB at this context; duplicate traces of the same session: rhiba0g8; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1509.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 1,509 61.9 80.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xill9rj9",
   "source_url": "https://omlx.ai/benchmarks/performance/xill9rj9",
   "tg_tokens": null,
   "tps": 61.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xill9rj9-ctx8192",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 80.5 GB at this context; duplicate traces of the same session: rhiba0g8; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1521.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 1,521 56.5 80.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xill9rj9",
   "source_url": "https://omlx.ai/benchmarks/performance/xill9rj9",
   "tg_tokens": null,
   "tps": 56.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xill9rj9-ctx16384",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 80.9 GB at this context; duplicate traces of the same session: rhiba0g8; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1434.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 16k 1,434 52.0 80.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xill9rj9",
   "source_url": "https://omlx.ai/benchmarks/performance/xill9rj9",
   "tg_tokens": null,
   "tps": 52.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xill9rj9-ctx32768",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 81.8 GB at this context; duplicate traces of the same session: rhiba0g8; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1262.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 32k 1,262 45.8 81.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xill9rj9",
   "source_url": "https://omlx.ai/benchmarks/performance/xill9rj9",
   "tg_tokens": null,
   "tps": 45.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xill9rj9-ctx65536",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 83.6 GB at this context; duplicate traces of the same session: rhiba0g8; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 994.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 64k 994.8 34.7 83.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xill9rj9",
   "source_url": "https://omlx.ai/benchmarks/performance/xill9rj9",
   "tg_tokens": null,
   "tps": 34.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-36gako4n-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 574.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 574.9 64.4 43.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/36gako4n",
   "source_url": "https://omlx.ai/benchmarks/performance/36gako4n",
   "tg_tokens": null,
   "tps": 64.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-36gako4n-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.2 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 660.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 660.7 60.6 43.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/36gako4n",
   "source_url": "https://omlx.ai/benchmarks/performance/36gako4n",
   "tg_tokens": null,
   "tps": 60.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-m7pdu55p-ctx1024",
   "model": "Qwen3-VL-8B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 6.0 GB at this context; duplicate traces of the same session: jca6774u; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-VL-8B-Instruct on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 224.6 21.3 6.0 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m7pdu55p",
   "source_url": "https://omlx.ai/benchmarks/performance/m7pdu55p",
   "tg_tokens": null,
   "tps": 21.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-m7pdu55p-ctx4096",
   "model": "Qwen3-VL-8B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: jca6774u; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 217.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-VL-8B-Instruct on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 217.1 18.9 6.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m7pdu55p",
   "source_url": "https://omlx.ai/benchmarks/performance/m7pdu55p",
   "tg_tokens": null,
   "tps": 18.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-od368nvv-ctx1024",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 1.5 GB at this context; duplicate traces of the same session: 3l7graoy, doabcv9b; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 756.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 4bit; 1k 756.4 100.8 1.5 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/od368nvv",
   "source_url": "https://omlx.ai/benchmarks/performance/od368nvv",
   "tg_tokens": null,
   "tps": 100.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-od368nvv-ctx4096",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 1.6 GB at this context; duplicate traces of the same session: 3l7graoy, doabcv9b; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 734.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 4bit; 4k 734.1 91.3 1.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/od368nvv",
   "source_url": "https://omlx.ai/benchmarks/performance/od368nvv",
   "tg_tokens": null,
   "tps": 91.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-od368nvv-ctx8192",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 1.6 GB at this context; duplicate traces of the same session: 3l7graoy, doabcv9b; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 710.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 4bit; 8k 710.7 82.2 1.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/od368nvv",
   "source_url": "https://omlx.ai/benchmarks/performance/od368nvv",
   "tg_tokens": null,
   "tps": 82.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-od368nvv-ctx16384",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 2.0 GB at this context; duplicate traces of the same session: 3l7graoy, doabcv9b; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 647.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 4bit; 16k 647.6 68.2 2.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/od368nvv",
   "source_url": "https://omlx.ai/benchmarks/performance/od368nvv",
   "tg_tokens": null,
   "tps": 68.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-od368nvv-ctx32768",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 2.6 GB at this context; duplicate traces of the same session: 3l7graoy, doabcv9b; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 552.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 4bit; 32k 552.2 47.8 2.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/od368nvv",
   "source_url": "https://omlx.ai/benchmarks/performance/od368nvv",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-od368nvv-ctx65536",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 4.1 GB at this context; duplicate traces of the same session: 3l7graoy, doabcv9b; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 408.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 4bit; 64k 408.6 17.3 4.1 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/od368nvv",
   "source_url": "https://omlx.ai/benchmarks/performance/od368nvv",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-mxnj5jsm-ctx1024",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 1.8 GB at this context; duplicate traces of the same session: f00eert8, wpqf311n; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 755.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 8bit; 1k 755.9 62.4 1.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mxnj5jsm",
   "source_url": "https://omlx.ai/benchmarks/performance/mxnj5jsm",
   "tg_tokens": null,
   "tps": 62.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-mxnj5jsm-ctx4096",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 1.9 GB at this context; duplicate traces of the same session: f00eert8, wpqf311n; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 733.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 8bit; 4k 733.9 58.9 1.9 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mxnj5jsm",
   "source_url": "https://omlx.ai/benchmarks/performance/mxnj5jsm",
   "tg_tokens": null,
   "tps": 58.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-mxnj5jsm-ctx8192",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 2.0 GB at this context; duplicate traces of the same session: f00eert8, wpqf311n; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 714.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 8bit; 8k 714.0 55.0 2.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mxnj5jsm",
   "source_url": "https://omlx.ai/benchmarks/performance/mxnj5jsm",
   "tg_tokens": null,
   "tps": 55.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-mxnj5jsm-ctx16384",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 2.4 GB at this context; duplicate traces of the same session: f00eert8, wpqf311n; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 654.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 8bit; 16k 654.9 47.9 2.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mxnj5jsm",
   "source_url": "https://omlx.ai/benchmarks/performance/mxnj5jsm",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-mxnj5jsm-ctx32768",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 3.0 GB at this context; duplicate traces of the same session: f00eert8, wpqf311n; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 554.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 8bit; 32k 554.7 37.3 3.0 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mxnj5jsm",
   "source_url": "https://omlx.ai/benchmarks/performance/mxnj5jsm",
   "tg_tokens": null,
   "tps": 37.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-mxnj5jsm-ctx65536",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 4.4 GB at this context; duplicate traces of the same session: f00eert8, wpqf311n; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 414.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 8bit; 64k 414.9 24.8 4.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mxnj5jsm",
   "source_url": "https://omlx.ai/benchmarks/performance/mxnj5jsm",
   "tg_tokens": null,
   "tps": 24.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-udv2mwwl-ctx1024",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 1.8 GB at this context; duplicate traces of the same session: none; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 726.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 8bit; 1k 726.8 48.2 1.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/udv2mwwl",
   "source_url": "https://omlx.ai/benchmarks/performance/udv2mwwl",
   "tg_tokens": null,
   "tps": 48.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-udv2mwwl-ctx4096",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 1.9 GB at this context; duplicate traces of the same session: none; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 653.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 8bit; 4k 653.5 41.2 1.9 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/udv2mwwl",
   "source_url": "https://omlx.ai/benchmarks/performance/udv2mwwl",
   "tg_tokens": null,
   "tps": 41.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-udv2mwwl-ctx8192",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 2.0 GB at this context; duplicate traces of the same session: none; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 646.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 8bit; 8k 646.1 45.1 2.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/udv2mwwl",
   "source_url": "https://omlx.ai/benchmarks/performance/udv2mwwl",
   "tg_tokens": null,
   "tps": 45.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-lqloqw71-ctx1024",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 3.3 GB; TTFT 783 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1308.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.5-0.8B on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 unknown; 1k tokens 1,308 PP tok/s 11.7 TG tok/s 783 TTFT (ms) 3.3 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lqloqw71",
   "source_url": "https://omlx.ai/benchmarks/performance/lqloqw71",
   "tg_tokens": null,
   "tps": 11.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-hfzzor6h-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.3 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 439.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 1k 439.6 47.8 66.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hfzzor6h",
   "source_url": "https://omlx.ai/benchmarks/performance/hfzzor6h",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-hfzzor6h-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.7 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 463.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 4k 463.3 45.6 66.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hfzzor6h",
   "source_url": "https://omlx.ai/benchmarks/performance/hfzzor6h",
   "tg_tokens": null,
   "tps": 45.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-hfzzor6h-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.0 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 459.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 8k 459.9 41.3 67.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hfzzor6h",
   "source_url": "https://omlx.ai/benchmarks/performance/hfzzor6h",
   "tg_tokens": null,
   "tps": 41.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-hfzzor6h-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.7 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 444.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 16k 444.0 36.3 67.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hfzzor6h",
   "source_url": "https://omlx.ai/benchmarks/performance/hfzzor6h",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-hfzzor6h-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 69.1 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 409.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 32k 409.4 28.3 69.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hfzzor6h",
   "source_url": "https://omlx.ai/benchmarks/performance/hfzzor6h",
   "tg_tokens": null,
   "tps": 28.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-hfzzor6h-ctx65536",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 71.9 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 331.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 64k 331.7 19.0 71.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hfzzor6h",
   "source_url": "https://omlx.ai/benchmarks/performance/hfzzor6h",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jgotjhpb-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.3 GB at this context; duplicate traces of the same session: ejitl07v; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 704.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 704.1 53.8 66.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jgotjhpb",
   "source_url": "https://omlx.ai/benchmarks/performance/jgotjhpb",
   "tg_tokens": null,
   "tps": 53.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jgotjhpb-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.7 GB at this context; duplicate traces of the same session: ejitl07v; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 729.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 729.1 51.2 66.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jgotjhpb",
   "source_url": "https://omlx.ai/benchmarks/performance/jgotjhpb",
   "tg_tokens": null,
   "tps": 51.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jgotjhpb-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.0 GB at this context; duplicate traces of the same session: ejitl07v; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 712.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 712.9 48.2 67.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jgotjhpb",
   "source_url": "https://omlx.ai/benchmarks/performance/jgotjhpb",
   "tg_tokens": null,
   "tps": 48.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jgotjhpb-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.7 GB at this context; duplicate traces of the same session: ejitl07v; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 670.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 670.3 41.2 67.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jgotjhpb",
   "source_url": "https://omlx.ai/benchmarks/performance/jgotjhpb",
   "tg_tokens": null,
   "tps": 41.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jgotjhpb-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 69.1 GB at this context; duplicate traces of the same session: ejitl07v; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 590.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 590.2 35.0 69.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jgotjhpb",
   "source_url": "https://omlx.ai/benchmarks/performance/jgotjhpb",
   "tg_tokens": null,
   "tps": 35.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jgotjhpb-ctx65536",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 71.9 GB at this context; duplicate traces of the same session: ejitl07v; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 466.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 466.9 26.0 71.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jgotjhpb",
   "source_url": "https://omlx.ai/benchmarks/performance/jgotjhpb",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-kvxrqtnz-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.3 GB at this context; duplicate traces of the same session: 919dis83, bzhe0vh3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 638.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 638.7 54.4 66.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kvxrqtnz",
   "source_url": "https://omlx.ai/benchmarks/performance/kvxrqtnz",
   "tg_tokens": null,
   "tps": 54.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-kvxrqtnz-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.7 GB at this context; duplicate traces of the same session: 919dis83, bzhe0vh3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 714.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 714.1 50.9 66.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kvxrqtnz",
   "source_url": "https://omlx.ai/benchmarks/performance/kvxrqtnz",
   "tg_tokens": null,
   "tps": 50.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-kvxrqtnz-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.0 GB at this context; duplicate traces of the same session: 919dis83, bzhe0vh3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 712.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 712.7 47.3 67.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kvxrqtnz",
   "source_url": "https://omlx.ai/benchmarks/performance/kvxrqtnz",
   "tg_tokens": null,
   "tps": 47.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-kvxrqtnz-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.7 GB at this context; duplicate traces of the same session: 919dis83, bzhe0vh3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 670.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 16k 670.7 35.9 67.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kvxrqtnz",
   "source_url": "https://omlx.ai/benchmarks/performance/kvxrqtnz",
   "tg_tokens": null,
   "tps": 35.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-kvxrqtnz-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 69.1 GB at this context; duplicate traces of the same session: 919dis83, bzhe0vh3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 587.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 587.5 24.4 69.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kvxrqtnz",
   "source_url": "https://omlx.ai/benchmarks/performance/kvxrqtnz",
   "tg_tokens": null,
   "tps": 24.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-kvxrqtnz-ctx65536",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 71.9 GB at this context; duplicate traces of the same session: 919dis83, bzhe0vh3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 458.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k 458.6 15.5 71.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kvxrqtnz",
   "source_url": "https://omlx.ai/benchmarks/performance/kvxrqtnz",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-43mgnpxt-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.3 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 897.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 897.8 52.3 66.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/43mgnpxt",
   "source_url": "https://omlx.ai/benchmarks/performance/43mgnpxt",
   "tg_tokens": null,
   "tps": 52.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-43mgnpxt-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 935.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 935.9 47.9 66.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/43mgnpxt",
   "source_url": "https://omlx.ai/benchmarks/performance/43mgnpxt",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-43mgnpxt-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.0 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 920.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 920.4 47.4 67.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/43mgnpxt",
   "source_url": "https://omlx.ai/benchmarks/performance/43mgnpxt",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-43mgnpxt-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 862.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 862.3 43.1 67.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/43mgnpxt",
   "source_url": "https://omlx.ai/benchmarks/performance/43mgnpxt",
   "tg_tokens": null,
   "tps": 43.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-43mgnpxt-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 69.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 746.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 746.1 36.4 69.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/43mgnpxt",
   "source_url": "https://omlx.ai/benchmarks/performance/43mgnpxt",
   "tg_tokens": null,
   "tps": 36.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-43mgnpxt-ctx65536",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 71.9 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 570.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 570.2 27.7 71.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/43mgnpxt",
   "source_url": "https://omlx.ai/benchmarks/performance/43mgnpxt",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ht88yq6s-ctx1024",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.1 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 85.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-27B on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 1k 85.8 11.1 29.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ht88yq6s",
   "source_url": "https://omlx.ai/benchmarks/performance/ht88yq6s",
   "tg_tokens": null,
   "tps": 11.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ht88yq6s-ctx4096",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 89.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-27B on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k 89.2 10.6 29.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ht88yq6s",
   "source_url": "https://omlx.ai/benchmarks/performance/ht88yq6s",
   "tg_tokens": null,
   "tps": 10.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-yo2wg2fa-ctx1024",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: fwtc8lbd, wtdigojg; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 180.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 180.4 32.2 16.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yo2wg2fa",
   "source_url": "https://omlx.ai/benchmarks/performance/yo2wg2fa",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-yo2wg2fa-ctx4096",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: fwtc8lbd, wtdigojg; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 182.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 182.4 30.4 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yo2wg2fa",
   "source_url": "https://omlx.ai/benchmarks/performance/yo2wg2fa",
   "tg_tokens": null,
   "tps": 30.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-yo2wg2fa-ctx8192",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: fwtc8lbd, wtdigojg; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 180.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 8k 180.9 28.4 17.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yo2wg2fa",
   "source_url": "https://omlx.ai/benchmarks/performance/yo2wg2fa",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-yo2wg2fa-ctx16384",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: fwtc8lbd, wtdigojg; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 176.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 16k 176.2 25.3 18.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yo2wg2fa",
   "source_url": "https://omlx.ai/benchmarks/performance/yo2wg2fa",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-5raey9l1-ctx1024",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: i0ji024h; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 226.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 1k 226.5 33.4 16.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5raey9l1",
   "source_url": "https://omlx.ai/benchmarks/performance/5raey9l1",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-5raey9l1-ctx4096",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: i0ji024h; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 228.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 4k 228.9 31.3 17.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5raey9l1",
   "source_url": "https://omlx.ai/benchmarks/performance/5raey9l1",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-5raey9l1-ctx8192",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: i0ji024h; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 227.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 8k 227.0 29.3 17.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5raey9l1",
   "source_url": "https://omlx.ai/benchmarks/performance/5raey9l1",
   "tg_tokens": null,
   "tps": 29.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-5raey9l1-ctx16384",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: i0ji024h; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 222.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 16k 222.3 26.6 18.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5raey9l1",
   "source_url": "https://omlx.ai/benchmarks/performance/5raey9l1",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-5raey9l1-ctx32768",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: i0ji024h; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 212.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 32k 212.9 21.9 20.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5raey9l1",
   "source_url": "https://omlx.ai/benchmarks/performance/5raey9l1",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-5raey9l1-ctx65536",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: i0ji024h; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 192.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 64k 192.2 16.8 23.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5raey9l1",
   "source_url": "https://omlx.ai/benchmarks/performance/5raey9l1",
   "tg_tokens": null,
   "tps": 16.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-us685980-ctx1024",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 2fbfl41g; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 226.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 1k 226.5 33.5 16.6 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/us685980",
   "source_url": "https://omlx.ai/benchmarks/performance/us685980",
   "tg_tokens": null,
   "tps": 33.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-us685980-ctx4096",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 2fbfl41g; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 228.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 4k 228.4 31.5 17.0 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/us685980",
   "source_url": "https://omlx.ai/benchmarks/performance/us685980",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-w3gm3t09-ctx1024",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: iff2zzcf; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 322.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 322.7 38.0 16.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w3gm3t09",
   "source_url": "https://omlx.ai/benchmarks/performance/w3gm3t09",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-w3gm3t09-ctx4096",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: iff2zzcf; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 324.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 324.2 35.3 17.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w3gm3t09",
   "source_url": "https://omlx.ai/benchmarks/performance/w3gm3t09",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-w3gm3t09-ctx8192",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: iff2zzcf; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 319.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 319.7 33.6 17.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w3gm3t09",
   "source_url": "https://omlx.ai/benchmarks/performance/w3gm3t09",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-w3gm3t09-ctx16384",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: iff2zzcf; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 310.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 310.0 30.0 18.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w3gm3t09",
   "source_url": "https://omlx.ai/benchmarks/performance/w3gm3t09",
   "tg_tokens": null,
   "tps": 30.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-w3gm3t09-ctx32768",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: iff2zzcf; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 291.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 291.3 24.9 20.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w3gm3t09",
   "source_url": "https://omlx.ai/benchmarks/performance/w3gm3t09",
   "tg_tokens": null,
   "tps": 24.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-w3gm3t09-ctx65536",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: iff2zzcf; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 258.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 258.7 18.1 23.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w3gm3t09",
   "source_url": "https://omlx.ai/benchmarks/performance/w3gm3t09",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-j3mx6mkn-ctx1024",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 239.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 239.4 27.6 16.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j3mx6mkn",
   "source_url": "https://omlx.ai/benchmarks/performance/j3mx6mkn",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-j3mx6mkn-ctx4096",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 199.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 199.2 23.6 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j3mx6mkn",
   "source_url": "https://omlx.ai/benchmarks/performance/j3mx6mkn",
   "tg_tokens": null,
   "tps": 23.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-j3mx6mkn-ctx8192",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 186.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 186.4 20.3 17.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j3mx6mkn",
   "source_url": "https://omlx.ai/benchmarks/performance/j3mx6mkn",
   "tg_tokens": null,
   "tps": 20.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-8c5n17xl-ctx1024",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: lsrf1njm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 240.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 240.5 20.5 22.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8c5n17xl",
   "source_url": "https://omlx.ai/benchmarks/performance/8c5n17xl",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-8c5n17xl-ctx4096",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: lsrf1njm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 235.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 235.0 19.5 23.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8c5n17xl",
   "source_url": "https://omlx.ai/benchmarks/performance/8c5n17xl",
   "tg_tokens": null,
   "tps": 19.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-l23r0h09-ctx1024",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 3grd36vj, 3rjvar3g, mhf4fjks, pfl2vdmm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 125.6 15.2 16.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l23r0h09",
   "source_url": "https://omlx.ai/benchmarks/performance/l23r0h09",
   "tg_tokens": null,
   "tps": 15.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-l23r0h09-ctx4096",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 3grd36vj, 3rjvar3g, mhf4fjks, pfl2vdmm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 125.9 14.8 17.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l23r0h09",
   "source_url": "https://omlx.ai/benchmarks/performance/l23r0h09",
   "tg_tokens": null,
   "tps": 14.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-l23r0h09-ctx8192",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: 3grd36vj, 3rjvar3g, mhf4fjks, pfl2vdmm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 124.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 124.3 14.2 17.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l23r0h09",
   "source_url": "https://omlx.ai/benchmarks/performance/l23r0h09",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-l23r0h09-ctx16384",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: 3grd36vj, 3rjvar3g, mhf4fjks, pfl2vdmm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 121.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 121.0 12.3 18.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l23r0h09",
   "source_url": "https://omlx.ai/benchmarks/performance/l23r0h09",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-l23r0h09-ctx32768",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 3grd36vj, 3rjvar3g, mhf4fjks, pfl2vdmm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 114.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 114.0 9.7 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l23r0h09",
   "source_url": "https://omlx.ai/benchmarks/performance/l23r0h09",
   "tg_tokens": null,
   "tps": 9.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-23f9ser1-ctx1024",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 16.6 GB; TTFT 8105 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 126.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k tokens 126.3 PP tok/s 15.4 TG tok/s 8105 TTFT (ms) 16.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/23f9ser1",
   "source_url": "https://omlx.ai/benchmarks/performance/23f9ser1",
   "tg_tokens": null,
   "tps": 15.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-duk8v6f6-ctx1024",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: vmky9zzs; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 128.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 1k 128.6 24.7 15.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/duk8v6f6",
   "source_url": "https://omlx.ai/benchmarks/performance/duk8v6f6",
   "tg_tokens": null,
   "tps": 24.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-duk8v6f6-ctx4096",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: vmky9zzs; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 133.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 4k 133.5 23.0 16.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/duk8v6f6",
   "source_url": "https://omlx.ai/benchmarks/performance/duk8v6f6",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-4laq01bf-ctx1024",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: ipsj46i8, mfwckz8w, qyady0u2; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 111.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 1k 111.1 21.1 15.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4laq01bf",
   "source_url": "https://omlx.ai/benchmarks/performance/4laq01bf",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-4laq01bf-ctx4096",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: ipsj46i8, mfwckz8w, qyady0u2; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 113.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k 113.8 20.1 16.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4laq01bf",
   "source_url": "https://omlx.ai/benchmarks/performance/4laq01bf",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-4laq01bf-ctx8192",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: ipsj46i8, mfwckz8w, qyady0u2; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 113.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 8k 113.1 19.0 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4laq01bf",
   "source_url": "https://omlx.ai/benchmarks/performance/4laq01bf",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-4laq01bf-ctx16384",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 17.5 GB at this context; duplicate traces of the same session: ipsj46i8, mfwckz8w, qyady0u2; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 111.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 16k 111.0 17.3 17.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4laq01bf",
   "source_url": "https://omlx.ai/benchmarks/performance/4laq01bf",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-4laq01bf-ctx32768",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: ipsj46i8, mfwckz8w, qyady0u2; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 32k 106.7 14.9 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4laq01bf",
   "source_url": "https://omlx.ai/benchmarks/performance/4laq01bf",
   "tg_tokens": null,
   "tps": 14.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-p09e4cme-ctx1024",
   "model": "Qwen3.5-2B",
   "notes": "oMLX community benchmark trace; peak memory 3.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1768.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-2B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 6bit; 1k 1,768 113.3 3.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p09e4cme",
   "source_url": "https://omlx.ai/benchmarks/performance/p09e4cme",
   "tg_tokens": null,
   "tps": 113.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-p09e4cme-ctx4096",
   "model": "Qwen3.5-2B",
   "notes": "oMLX community benchmark trace; peak memory 3.2 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1797.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-2B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 6bit; 4k 1,797 109.1 3.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p09e4cme",
   "source_url": "https://omlx.ai/benchmarks/performance/p09e4cme",
   "tg_tokens": null,
   "tps": 109.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-kny2vtr3-ctx1024",
   "model": "Qwen3.5-2B-Claude-4.6-Opus-Reasoning-Distilled",
   "notes": "oMLX community benchmark trace; peak memory 4.1 GB at this context; duplicate traces of the same session: q76ia64b; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 860.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.5-2B-Claude-4.6-Opus-Reasoning-Distilled on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 bf16; 1k 860.0 25.1 4.1 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kny2vtr3",
   "source_url": "https://omlx.ai/benchmarks/performance/kny2vtr3",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-kny2vtr3-ctx4096",
   "model": "Qwen3.5-2B-Claude-4.6-Opus-Reasoning-Distilled",
   "notes": "oMLX community benchmark trace; peak memory 4.3 GB at this context; duplicate traces of the same session: q76ia64b; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1059.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.5-2B-Claude-4.6-Opus-Reasoning-Distilled on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 bf16; 4k 1,059 24.6 4.3 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kny2vtr3",
   "source_url": "https://omlx.ai/benchmarks/performance/kny2vtr3",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-zfc09bsq-ctx1024",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 36.2 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 519.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-35B-A3B on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 1k 519.6 48.0 36.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zfc09bsq",
   "source_url": "https://omlx.ai/benchmarks/performance/zfc09bsq",
   "tg_tokens": null,
   "tps": 48.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-zfc09bsq-ctx4096",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 36.3 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 566.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-35B-A3B on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k 566.5 43.5 36.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zfc09bsq",
   "source_url": "https://omlx.ai/benchmarks/performance/zfc09bsq",
   "tg_tokens": null,
   "tps": 43.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-0emkmm0e-ctx1024",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 3xet8brb, femieeij, ptzhx9ln; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1652.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,652 93.4 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0emkmm0e",
   "source_url": "https://omlx.ai/benchmarks/performance/0emkmm0e",
   "tg_tokens": null,
   "tps": 93.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-0emkmm0e-ctx4096",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 3xet8brb, femieeij, ptzhx9ln; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1706.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,706 88.6 20.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0emkmm0e",
   "source_url": "https://omlx.ai/benchmarks/performance/0emkmm0e",
   "tg_tokens": null,
   "tps": 88.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-0emkmm0e-ctx8192",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: 3xet8brb, femieeij, ptzhx9ln; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1487.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,487 90.0 20.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0emkmm0e",
   "source_url": "https://omlx.ai/benchmarks/performance/0emkmm0e",
   "tg_tokens": null,
   "tps": 90.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-0emkmm0e-ctx16384",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 3xet8brb, femieeij, ptzhx9ln; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1227.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,227 63.7 20.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0emkmm0e",
   "source_url": "https://omlx.ai/benchmarks/performance/0emkmm0e",
   "tg_tokens": null,
   "tps": 63.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-0emkmm0e-ctx32768",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: 3xet8brb, femieeij, ptzhx9ln; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1026.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 1,026 42.5 21.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0emkmm0e",
   "source_url": "https://omlx.ai/benchmarks/performance/0emkmm0e",
   "tg_tokens": null,
   "tps": 42.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-glgcvy99-ctx1024",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1361.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,361 95.7 20.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/glgcvy99",
   "source_url": "https://omlx.ai/benchmarks/performance/glgcvy99",
   "tg_tokens": null,
   "tps": 95.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-glgcvy99-ctx4096",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1067.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,067 95.8 20.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/glgcvy99",
   "source_url": "https://omlx.ai/benchmarks/performance/glgcvy99",
   "tg_tokens": null,
   "tps": 95.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-glgcvy99-ctx8192",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1091.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,091 90.5 20.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/glgcvy99",
   "source_url": "https://omlx.ai/benchmarks/performance/glgcvy99",
   "tg_tokens": null,
   "tps": 90.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-xmv2d997-ctx1024",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: vpoz1qjs; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 821.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 821.5 75.6 20.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xmv2d997",
   "source_url": "https://omlx.ai/benchmarks/performance/xmv2d997",
   "tg_tokens": null,
   "tps": 75.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-xmv2d997-ctx4096",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: vpoz1qjs; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 879.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 879.6 79.1 20.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xmv2d997",
   "source_url": "https://omlx.ai/benchmarks/performance/xmv2d997",
   "tg_tokens": null,
   "tps": 79.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-xmv2d997-ctx8192",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: vpoz1qjs; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 873.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 873.4 71.9 20.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xmv2d997",
   "source_url": "https://omlx.ai/benchmarks/performance/xmv2d997",
   "tg_tokens": null,
   "tps": 71.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-vy20rvxa-ctx1024",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 9bgkvghg, cq1j63z1, fxnkfdkj, tzmnrcg3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 867.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 867.2 82.8 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vy20rvxa",
   "source_url": "https://omlx.ai/benchmarks/performance/vy20rvxa",
   "tg_tokens": null,
   "tps": 82.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-vy20rvxa-ctx4096",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 9bgkvghg, cq1j63z1, fxnkfdkj, tzmnrcg3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 899.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 899.9 79.2 20.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vy20rvxa",
   "source_url": "https://omlx.ai/benchmarks/performance/vy20rvxa",
   "tg_tokens": null,
   "tps": 79.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-vy20rvxa-ctx8192",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: 9bgkvghg, cq1j63z1, fxnkfdkj, tzmnrcg3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 876.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 876.2 74.9 20.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vy20rvxa",
   "source_url": "https://omlx.ai/benchmarks/performance/vy20rvxa",
   "tg_tokens": null,
   "tps": 74.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-vy20rvxa-ctx16384",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 9bgkvghg, cq1j63z1, fxnkfdkj, tzmnrcg3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 805.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 805.6 57.7 20.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vy20rvxa",
   "source_url": "https://omlx.ai/benchmarks/performance/vy20rvxa",
   "tg_tokens": null,
   "tps": 57.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-vy20rvxa-ctx32768",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: 9bgkvghg, cq1j63z1, fxnkfdkj, tzmnrcg3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 693.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 693.2 40.6 21.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vy20rvxa",
   "source_url": "https://omlx.ai/benchmarks/performance/vy20rvxa",
   "tg_tokens": null,
   "tps": 40.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-vy20rvxa-ctx65536",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 9bgkvghg, cq1j63z1, fxnkfdkj, tzmnrcg3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 537.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 537.6 25.9 23.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vy20rvxa",
   "source_url": "https://omlx.ai/benchmarks/performance/vy20rvxa",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-k0jw9l2d-ctx1024",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 210.4 GB at this context; duplicate traces of the same session: 4o3fa12n; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 381.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 381.0 37.0 210.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k0jw9l2d",
   "source_url": "https://omlx.ai/benchmarks/performance/k0jw9l2d",
   "tg_tokens": null,
   "tps": 37.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-k0jw9l2d-ctx4096",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 210.7 GB at this context; duplicate traces of the same session: 4o3fa12n; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 404.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 404.5 35.3 210.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k0jw9l2d",
   "source_url": "https://omlx.ai/benchmarks/performance/k0jw9l2d",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-k0jw9l2d-ctx8192",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 211.0 GB at this context; duplicate traces of the same session: 4o3fa12n; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 400.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 400.7 31.3 211.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k0jw9l2d",
   "source_url": "https://omlx.ai/benchmarks/performance/k0jw9l2d",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-k0jw9l2d-ctx16384",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 211.8 GB at this context; duplicate traces of the same session: 4o3fa12n; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 384.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 384.0 24.8 211.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k0jw9l2d",
   "source_url": "https://omlx.ai/benchmarks/performance/k0jw9l2d",
   "tg_tokens": null,
   "tps": 24.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-k0jw9l2d-ctx32768",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 213.2 GB at this context; duplicate traces of the same session: 4o3fa12n; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 350.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 350.4 17.8 213.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k0jw9l2d",
   "source_url": "https://omlx.ai/benchmarks/performance/k0jw9l2d",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-k0jw9l2d-ctx65536",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 216.2 GB at this context; duplicate traces of the same session: 4o3fa12n; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 294.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 294.3 11.7 216.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k0jw9l2d",
   "source_url": "https://omlx.ai/benchmarks/performance/k0jw9l2d",
   "tg_tokens": null,
   "tps": 11.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-yswmz05e-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.0 GB at this context; duplicate traces of the same session: 0ysklzxp, aeq42ap1, lsp40v2a, o7u1nsfl; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 736.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 1k 736.4 102.2 4.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yswmz05e",
   "source_url": "https://omlx.ai/benchmarks/performance/yswmz05e",
   "tg_tokens": null,
   "tps": 102.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-yswmz05e-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: 0ysklzxp, aeq42ap1, lsp40v2a, o7u1nsfl; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 734.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k 734.5 95.4 4.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yswmz05e",
   "source_url": "https://omlx.ai/benchmarks/performance/yswmz05e",
   "tg_tokens": null,
   "tps": 95.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-yswmz05e-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.6 GB at this context; duplicate traces of the same session: 0ysklzxp, aeq42ap1, lsp40v2a, o7u1nsfl; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 720.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 8k 720.3 87.0 4.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yswmz05e",
   "source_url": "https://omlx.ai/benchmarks/performance/yswmz05e",
   "tg_tokens": null,
   "tps": 87.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-yswmz05e-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 5.3 GB at this context; duplicate traces of the same session: 0ysklzxp, aeq42ap1, lsp40v2a, o7u1nsfl; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 691.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 16k 691.3 73.4 5.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yswmz05e",
   "source_url": "https://omlx.ai/benchmarks/performance/yswmz05e",
   "tg_tokens": null,
   "tps": 73.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-yswmz05e-ctx32768",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: 0ysklzxp, aeq42ap1, lsp40v2a, o7u1nsfl; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 576.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 32k 576.0 28.6 6.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yswmz05e",
   "source_url": "https://omlx.ai/benchmarks/performance/yswmz05e",
   "tg_tokens": null,
   "tps": 28.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-yswmz05e-ctx65536",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 8.6 GB at this context; duplicate traces of the same session: 0ysklzxp, aeq42ap1, lsp40v2a, o7u1nsfl; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 468.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 64k 468.7 18.2 8.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yswmz05e",
   "source_url": "https://omlx.ai/benchmarks/performance/yswmz05e",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-xckibqvi-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.0 GB at this context; duplicate traces of the same session: 0oga1cf0, j290pxfi, zwuvz3cp; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1428.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 1k 1,428 137.5 4.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xckibqvi",
   "source_url": "https://omlx.ai/benchmarks/performance/xckibqvi",
   "tg_tokens": null,
   "tps": 137.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-xckibqvi-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: 0oga1cf0, j290pxfi, zwuvz3cp; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1440.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 4k 1,440 128.2 4.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xckibqvi",
   "source_url": "https://omlx.ai/benchmarks/performance/xckibqvi",
   "tg_tokens": null,
   "tps": 128.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-xckibqvi-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.6 GB at this context; duplicate traces of the same session: 0oga1cf0, j290pxfi, zwuvz3cp; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1414.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 8k 1,414 117.4 4.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xckibqvi",
   "source_url": "https://omlx.ai/benchmarks/performance/xckibqvi",
   "tg_tokens": null,
   "tps": 117.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-xckibqvi-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 5.3 GB at this context; duplicate traces of the same session: 0oga1cf0, j290pxfi, zwuvz3cp; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1355.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 16k 1,355 98.7 5.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xckibqvi",
   "source_url": "https://omlx.ai/benchmarks/performance/xckibqvi",
   "tg_tokens": null,
   "tps": 98.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-xckibqvi-ctx32768",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: 0oga1cf0, j290pxfi, zwuvz3cp; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1247.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 32k 1,247 74.5 6.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xckibqvi",
   "source_url": "https://omlx.ai/benchmarks/performance/xckibqvi",
   "tg_tokens": null,
   "tps": 74.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-xckibqvi-ctx65536",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 8.6 GB at this context; duplicate traces of the same session: 0oga1cf0, j290pxfi, zwuvz3cp; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1051.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 64k 1,051 50.4 8.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xckibqvi",
   "source_url": "https://omlx.ai/benchmarks/performance/xckibqvi",
   "tg_tokens": null,
   "tps": 50.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-u49qccjp-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: 4vf3fhkn, 5t6c2eu1, cgvsvi3i; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1423.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 1k 1,423 99.9 5.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u49qccjp",
   "source_url": "https://omlx.ai/benchmarks/performance/u49qccjp",
   "tg_tokens": null,
   "tps": 99.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-u49qccjp-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 6.1 GB at this context; duplicate traces of the same session: 4vf3fhkn, 5t6c2eu1, cgvsvi3i; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1448.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 4k 1,448 96.2 6.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u49qccjp",
   "source_url": "https://omlx.ai/benchmarks/performance/u49qccjp",
   "tg_tokens": null,
   "tps": 96.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-u49qccjp-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: 4vf3fhkn, 5t6c2eu1, cgvsvi3i; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1421.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 8k 1,421 90.1 6.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u49qccjp",
   "source_url": "https://omlx.ai/benchmarks/performance/u49qccjp",
   "tg_tokens": null,
   "tps": 90.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-u49qccjp-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 7.3 GB at this context; duplicate traces of the same session: 4vf3fhkn, 5t6c2eu1, cgvsvi3i; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1362.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 16k 1,362 80.1 7.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u49qccjp",
   "source_url": "https://omlx.ai/benchmarks/performance/u49qccjp",
   "tg_tokens": null,
   "tps": 80.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-u49qccjp-ctx32768",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 7.9 GB at this context; duplicate traces of the same session: 4vf3fhkn, 5t6c2eu1, cgvsvi3i; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1255.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 32k 1,255 63.0 7.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u49qccjp",
   "source_url": "https://omlx.ai/benchmarks/performance/u49qccjp",
   "tg_tokens": null,
   "tps": 63.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-u49qccjp-ctx65536",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 9.9 GB at this context; duplicate traces of the same session: 4vf3fhkn, 5t6c2eu1, cgvsvi3i; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1056.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 64k 1,056 45.5 9.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u49qccjp",
   "source_url": "https://omlx.ai/benchmarks/performance/u49qccjp",
   "tg_tokens": null,
   "tps": 45.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-dnohwwb2-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 407.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 407.7 38.6 3.9 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dnohwwb2",
   "source_url": "https://omlx.ai/benchmarks/performance/dnohwwb2",
   "tg_tokens": null,
   "tps": 38.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-dnohwwb2-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 405.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 405.2 36.6 4.2 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dnohwwb2",
   "source_url": "https://omlx.ai/benchmarks/performance/dnohwwb2",
   "tg_tokens": null,
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-b5hno3da-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 396.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 396.6 36.4 3.9 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b5hno3da",
   "source_url": "https://omlx.ai/benchmarks/performance/b5hno3da",
   "tg_tokens": null,
   "tps": 36.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-b5hno3da-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 393.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 393.2 34.2 4.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b5hno3da",
   "source_url": "https://omlx.ai/benchmarks/performance/b5hno3da",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-bkoocnbv-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: o03mk72e; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 406.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 406.4 37.9 3.9 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bkoocnbv",
   "source_url": "https://omlx.ai/benchmarks/performance/bkoocnbv",
   "tg_tokens": null,
   "tps": 37.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-bkoocnbv-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: o03mk72e; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 403.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 403.4 35.5 4.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bkoocnbv",
   "source_url": "https://omlx.ai/benchmarks/performance/bkoocnbv",
   "tg_tokens": null,
   "tps": 35.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-0z5umpj8-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: jtr6rvgo; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 397.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 397.8 37.1 3.9 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0z5umpj8",
   "source_url": "https://omlx.ai/benchmarks/performance/0z5umpj8",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-0z5umpj8-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: jtr6rvgo; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 396.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 396.2 34.2 4.2 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0z5umpj8",
   "source_url": "https://omlx.ai/benchmarks/performance/0z5umpj8",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-zgqpr95d-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 3.9 GB; TTFT 2556 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 400.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k tokens 400.6 PP tok/s 36.9 TG tok/s 2556 TTFT (ms) 3.9 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zgqpr95d",
   "source_url": "https://omlx.ai/benchmarks/performance/zgqpr95d",
   "tg_tokens": null,
   "tps": 36.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-1gnof7xh-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: 91h2g4t3, exfl6ecd, s11jretg, tir4w3vr, tirqyj3o; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 374.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 374.9 33.7 3.9 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1gnof7xh",
   "source_url": "https://omlx.ai/benchmarks/performance/1gnof7xh",
   "tg_tokens": null,
   "tps": 33.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-1gnof7xh-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: 91h2g4t3, exfl6ecd, s11jretg, tir4w3vr, tirqyj3o; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 373.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 373.3 32.0 4.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1gnof7xh",
   "source_url": "https://omlx.ai/benchmarks/performance/1gnof7xh",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-1gnof7xh-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.6 GB at this context; duplicate traces of the same session: 91h2g4t3, exfl6ecd, s11jretg, tir4w3vr, tirqyj3o; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 364.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 8k 364.6 29.5 4.6 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1gnof7xh",
   "source_url": "https://omlx.ai/benchmarks/performance/1gnof7xh",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-1gnof7xh-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.8 GB at this context; duplicate traces of the same session: 91h2g4t3, exfl6ecd, s11jretg, tir4w3vr, tirqyj3o; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 340.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 16k 340.6 25.2 4.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1gnof7xh",
   "source_url": "https://omlx.ai/benchmarks/performance/1gnof7xh",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-1gnof7xh-ctx32768",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: 91h2g4t3, exfl6ecd, s11jretg, tir4w3vr, tirqyj3o; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 299.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 32k 299.7 20.3 5.9 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1gnof7xh",
   "source_url": "https://omlx.ai/benchmarks/performance/1gnof7xh",
   "tg_tokens": null,
   "tps": 20.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-1gnof7xh-ctx65536",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 8.0 GB at this context; duplicate traces of the same session: 91h2g4t3, exfl6ecd, s11jretg, tir4w3vr, tirqyj3o; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 231.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 64k 231.3 13.6 8.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1gnof7xh",
   "source_url": "https://omlx.ai/benchmarks/performance/1gnof7xh",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-925jucnl-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: mhaypknn, rkslpdmh; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 406.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 406.1 37.4 3.9 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/925jucnl",
   "source_url": "https://omlx.ai/benchmarks/performance/925jucnl",
   "tg_tokens": null,
   "tps": 37.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-925jucnl-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: mhaypknn, rkslpdmh; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 403.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 403.8 36.0 4.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/925jucnl",
   "source_url": "https://omlx.ai/benchmarks/performance/925jucnl",
   "tg_tokens": null,
   "tps": 36.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-925jucnl-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.6 GB at this context; duplicate traces of the same session: mhaypknn, rkslpdmh; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 385.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 8k 385.2 31.8 4.6 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/925jucnl",
   "source_url": "https://omlx.ai/benchmarks/performance/925jucnl",
   "tg_tokens": null,
   "tps": 31.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-crppjgnt-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 3.9 GB; TTFT 2557 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 400.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k tokens 400.4 PP tok/s 37.4 TG tok/s 2557 TTFT (ms) 3.9 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/crppjgnt",
   "source_url": "https://omlx.ai/benchmarks/performance/crppjgnt",
   "tg_tokens": null,
   "tps": 37.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-h7hh9kvw-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.0 GB at this context; duplicate traces of the same session: c1jz32zu, j3udbnrh, nvxbhp1x, qsyh8wbm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 811.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 811.1 82.1 4.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h7hh9kvw",
   "source_url": "https://omlx.ai/benchmarks/performance/h7hh9kvw",
   "tg_tokens": null,
   "tps": 82.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-h7hh9kvw-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: c1jz32zu, j3udbnrh, nvxbhp1x, qsyh8wbm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 798.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 798.0 76.8 4.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h7hh9kvw",
   "source_url": "https://omlx.ai/benchmarks/performance/h7hh9kvw",
   "tg_tokens": null,
   "tps": 76.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-h7hh9kvw-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.6 GB at this context; duplicate traces of the same session: c1jz32zu, j3udbnrh, nvxbhp1x, qsyh8wbm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 781.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 781.9 72.4 4.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h7hh9kvw",
   "source_url": "https://omlx.ai/benchmarks/performance/h7hh9kvw",
   "tg_tokens": null,
   "tps": 72.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-h7hh9kvw-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 5.3 GB at this context; duplicate traces of the same session: c1jz32zu, j3udbnrh, nvxbhp1x, qsyh8wbm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 737.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 737.3 62.8 5.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h7hh9kvw",
   "source_url": "https://omlx.ai/benchmarks/performance/h7hh9kvw",
   "tg_tokens": null,
   "tps": 62.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-h7hh9kvw-ctx32768",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: c1jz32zu, j3udbnrh, nvxbhp1x, qsyh8wbm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 655.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 655.0 49.9 6.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h7hh9kvw",
   "source_url": "https://omlx.ai/benchmarks/performance/h7hh9kvw",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-h7hh9kvw-ctx65536",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 8.6 GB at this context; duplicate traces of the same session: c1jz32zu, j3udbnrh, nvxbhp1x, qsyh8wbm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 534.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 534.5 36.3 8.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h7hh9kvw",
   "source_url": "https://omlx.ai/benchmarks/performance/h7hh9kvw",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-o4cjy9j3-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "OptiQ build; oMLX community benchmark trace; peak memory 4.8 GB at this context; duplicate traces of the same session: none; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 254.9,
   "provenance": "community",
   "quant": "OptiQ",
   "quote": "Qwen3.5-4B-OptiQ on M4 (8c); M4 (8c) \u00b7 16 GB \u00b7 4bit; 1k 254.9 23.3 4.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o4cjy9j3",
   "source_url": "https://omlx.ai/benchmarks/performance/o4cjy9j3",
   "tg_tokens": null,
   "tps": 23.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-o4cjy9j3-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "OptiQ build; oMLX community benchmark trace; peak memory 5.4 GB at this context; duplicate traces of the same session: none; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 263.4,
   "provenance": "community",
   "quant": "OptiQ",
   "quote": "Qwen3.5-4B-OptiQ on M4 (8c); M4 (8c) \u00b7 16 GB \u00b7 4bit; 4k 263.4 22.1 5.4 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o4cjy9j3",
   "source_url": "https://omlx.ai/benchmarks/performance/o4cjy9j3",
   "tg_tokens": null,
   "tps": 22.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-5fjbahy5-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.7 GB at this context; duplicate traces of the same session: 1n8qgmq9, 6rngmu2d, hyclrwcc; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 239.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 8bit; 1k 239.3 35.1 10.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5fjbahy5",
   "source_url": "https://omlx.ai/benchmarks/performance/5fjbahy5",
   "tg_tokens": null,
   "tps": 35.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-5fjbahy5-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.0 GB at this context; duplicate traces of the same session: 1n8qgmq9, 6rngmu2d, hyclrwcc; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 242.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 8bit; 4k 242.1 33.3 11.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5fjbahy5",
   "source_url": "https://omlx.ai/benchmarks/performance/5fjbahy5",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-5fjbahy5-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.2 GB at this context; duplicate traces of the same session: 1n8qgmq9, 6rngmu2d, hyclrwcc; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 236.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 8bit; 8k 236.2 30.3 11.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5fjbahy5",
   "source_url": "https://omlx.ai/benchmarks/performance/5fjbahy5",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-5fjbahy5-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.7 GB at this context; duplicate traces of the same session: 1n8qgmq9, 6rngmu2d, hyclrwcc; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 234.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 8bit; 16k 234.4 26.2 11.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5fjbahy5",
   "source_url": "https://omlx.ai/benchmarks/performance/5fjbahy5",
   "tg_tokens": null,
   "tps": 26.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-5fjbahy5-ctx32768",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: 1n8qgmq9, 6rngmu2d, hyclrwcc; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 8bit; 32k 224.8 20.3 12.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5fjbahy5",
   "source_url": "https://omlx.ai/benchmarks/performance/5fjbahy5",
   "tg_tokens": null,
   "tps": 20.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-5fjbahy5-ctx65536",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: 1n8qgmq9, 6rngmu2d, hyclrwcc; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 8bit; 64k 205.0 13.9 14.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5fjbahy5",
   "source_url": "https://omlx.ai/benchmarks/performance/5fjbahy5",
   "tg_tokens": null,
   "tps": 13.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-6kpz3oma-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 308.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 308.2 55.1 6.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6kpz3oma",
   "source_url": "https://omlx.ai/benchmarks/performance/6kpz3oma",
   "tg_tokens": null,
   "tps": 55.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-6kpz3oma-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 311.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 311.9 53.2 7.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6kpz3oma",
   "source_url": "https://omlx.ai/benchmarks/performance/6kpz3oma",
   "tg_tokens": null,
   "tps": 53.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-mlfronh1-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: opjss9el; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 138.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 1k 138.0 30.0 6.7 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mlfronh1",
   "source_url": "https://omlx.ai/benchmarks/performance/mlfronh1",
   "tg_tokens": null,
   "tps": 30.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-mlfronh1-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: opjss9el; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 138.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 4k 138.6 28.7 7.0 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mlfronh1",
   "source_url": "https://omlx.ai/benchmarks/performance/mlfronh1",
   "tg_tokens": null,
   "tps": 28.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-vfdq12ip-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 401.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 1k 401.8 58.2 6.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vfdq12ip",
   "source_url": "https://omlx.ai/benchmarks/performance/vfdq12ip",
   "tg_tokens": null,
   "tps": 58.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-vfdq12ip-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.3 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 408.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k 408.3 53.6 7.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vfdq12ip",
   "source_url": "https://omlx.ai/benchmarks/performance/vfdq12ip",
   "tg_tokens": null,
   "tps": 53.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-1a589yg5-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: kxuicg18; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 131.1 17.3 6.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1a589yg5",
   "source_url": "https://omlx.ai/benchmarks/performance/1a589yg5",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-1a589yg5-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: kxuicg18; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 130.7 16.9 7.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1a589yg5",
   "source_url": "https://omlx.ai/benchmarks/performance/1a589yg5",
   "tg_tokens": null,
   "tps": 16.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-xpnt88fb-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.7 GB at this context; duplicate traces of the same session: a6fmmdir; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 775.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 1k 775.8 74.3 8.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xpnt88fb",
   "source_url": "https://omlx.ai/benchmarks/performance/xpnt88fb",
   "tg_tokens": null,
   "tps": 74.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-xpnt88fb-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.9 GB at this context; duplicate traces of the same session: a6fmmdir; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 783.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 4k 783.9 72.1 8.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xpnt88fb",
   "source_url": "https://omlx.ai/benchmarks/performance/xpnt88fb",
   "tg_tokens": null,
   "tps": 72.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-xpnt88fb-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 9.2 GB at this context; duplicate traces of the same session: a6fmmdir; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 776.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 8k 776.7 68.7 9.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xpnt88fb",
   "source_url": "https://omlx.ai/benchmarks/performance/xpnt88fb",
   "tg_tokens": null,
   "tps": 68.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-xpnt88fb-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 9.6 GB at this context; duplicate traces of the same session: a6fmmdir; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 758.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 16k 758.8 62.8 9.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xpnt88fb",
   "source_url": "https://omlx.ai/benchmarks/performance/xpnt88fb",
   "tg_tokens": null,
   "tps": 62.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-xpnt88fb-ctx32768",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.6 GB at this context; duplicate traces of the same session: a6fmmdir; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 724.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 32k 724.1 52.6 10.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xpnt88fb",
   "source_url": "https://omlx.ai/benchmarks/performance/xpnt88fb",
   "tg_tokens": null,
   "tps": 52.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-xpnt88fb-ctx65536",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: a6fmmdir; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 653.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 64k 653.8 39.8 12.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xpnt88fb",
   "source_url": "https://omlx.ai/benchmarks/performance/xpnt88fb",
   "tg_tokens": null,
   "tps": 39.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-8g093y7c-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.7 GB at this context; duplicate traces of the same session: 6o6zm8tl, voj8az1w, wxd1vexu; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 794.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 1k 794.0 64.3 10.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8g093y7c",
   "source_url": "https://omlx.ai/benchmarks/performance/8g093y7c",
   "tg_tokens": null,
   "tps": 64.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-8g093y7c-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.0 GB at this context; duplicate traces of the same session: 6o6zm8tl, voj8az1w, wxd1vexu; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 804.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 4k 804.0 62.6 11.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8g093y7c",
   "source_url": "https://omlx.ai/benchmarks/performance/8g093y7c",
   "tg_tokens": null,
   "tps": 62.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-8g093y7c-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.2 GB at this context; duplicate traces of the same session: 6o6zm8tl, voj8az1w, wxd1vexu; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 797.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 8k 797.4 59.7 11.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8g093y7c",
   "source_url": "https://omlx.ai/benchmarks/performance/8g093y7c",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-8g093y7c-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.7 GB at this context; duplicate traces of the same session: 6o6zm8tl, voj8az1w, wxd1vexu; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 778.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 16k 778.7 54.0 11.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8g093y7c",
   "source_url": "https://omlx.ai/benchmarks/performance/8g093y7c",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-8g093y7c-ctx32768",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: 6o6zm8tl, voj8az1w, wxd1vexu; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 742.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 32k 742.0 46.4 12.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8g093y7c",
   "source_url": "https://omlx.ai/benchmarks/performance/8g093y7c",
   "tg_tokens": null,
   "tps": 46.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-8g093y7c-ctx65536",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: 6o6zm8tl, voj8az1w, wxd1vexu; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 668.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 64k 668.6 33.9 14.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8g093y7c",
   "source_url": "https://omlx.ai/benchmarks/performance/8g093y7c",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-vbwrj8jg-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: f4y5m3yt; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 722.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 1k 722.9 97.0 6.7 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vbwrj8jg",
   "source_url": "https://omlx.ai/benchmarks/performance/vbwrj8jg",
   "tg_tokens": null,
   "tps": 97.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-vbwrj8jg-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: f4y5m3yt; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 786.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 4k 786.3 93.7 7.0 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vbwrj8jg",
   "source_url": "https://omlx.ai/benchmarks/performance/vbwrj8jg",
   "tg_tokens": null,
   "tps": 93.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-lis7wrjs-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: mk5718cc, r7jnwmp2, rr8vgmtb, v9s757f5, zd9pv0ve; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 727.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 1k 727.6 96.7 6.7 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lis7wrjs",
   "source_url": "https://omlx.ai/benchmarks/performance/lis7wrjs",
   "tg_tokens": null,
   "tps": 96.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-lis7wrjs-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: mk5718cc, r7jnwmp2, rr8vgmtb, v9s757f5, zd9pv0ve; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 789.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 4k 789.6 93.2 7.0 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lis7wrjs",
   "source_url": "https://omlx.ai/benchmarks/performance/lis7wrjs",
   "tg_tokens": null,
   "tps": 93.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-lis7wrjs-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.1 GB at this context; duplicate traces of the same session: mk5718cc, r7jnwmp2, rr8vgmtb, v9s757f5, zd9pv0ve; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 792.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 8k 792.6 87.6 7.1 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lis7wrjs",
   "source_url": "https://omlx.ai/benchmarks/performance/lis7wrjs",
   "tg_tokens": null,
   "tps": 87.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-lis7wrjs-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.6 GB at this context; duplicate traces of the same session: mk5718cc, r7jnwmp2, rr8vgmtb, v9s757f5, zd9pv0ve; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 774.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 16k 774.4 76.6 7.6 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lis7wrjs",
   "source_url": "https://omlx.ai/benchmarks/performance/lis7wrjs",
   "tg_tokens": null,
   "tps": 76.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-lis7wrjs-ctx32768",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.7 GB at this context; duplicate traces of the same session: mk5718cc, r7jnwmp2, rr8vgmtb, v9s757f5, zd9pv0ve; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 738.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 32k 738.1 61.9 8.7 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lis7wrjs",
   "source_url": "https://omlx.ai/benchmarks/performance/lis7wrjs",
   "tg_tokens": null,
   "tps": 61.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-lis7wrjs-ctx65536",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.8 GB at this context; duplicate traces of the same session: mk5718cc, r7jnwmp2, rr8vgmtb, v9s757f5, zd9pv0ve; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 665.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 64k 665.7 44.3 10.8 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lis7wrjs",
   "source_url": "https://omlx.ai/benchmarks/performance/lis7wrjs",
   "tg_tokens": null,
   "tps": 44.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-h80g2jms-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: p13i2hbr; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 174.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M3 (10c); M3 (10c) \u00b7 16 GB \u00b7 4bit; 1k 174.7 17.7 6.5 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h80g2jms",
   "source_url": "https://omlx.ai/benchmarks/performance/h80g2jms",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-h80g2jms-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: p13i2hbr; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 157.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M3 (10c); M3 (10c) \u00b7 16 GB \u00b7 4bit; 4k 157.4 17.6 6.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h80g2jms",
   "source_url": "https://omlx.ai/benchmarks/performance/h80g2jms",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-u3d2a38i-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: g209ho7m; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 222.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 222.5 21.0 6.5 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u3d2a38i",
   "source_url": "https://omlx.ai/benchmarks/performance/u3d2a38i",
   "tg_tokens": null,
   "tps": 21.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-u3d2a38i-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: g209ho7m; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 222.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 222.7 20.1 6.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u3d2a38i",
   "source_url": "https://omlx.ai/benchmarks/performance/u3d2a38i",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-covniwh2-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: fbepb9zp; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 224.1 20.2 6.5 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/covniwh2",
   "source_url": "https://omlx.ai/benchmarks/performance/covniwh2",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-covniwh2-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: fbepb9zp; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 224.2 19.8 6.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/covniwh2",
   "source_url": "https://omlx.ai/benchmarks/performance/covniwh2",
   "tg_tokens": null,
   "tps": 19.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-jvcibpti-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 222.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 222.1 20.7 6.5 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jvcibpti",
   "source_url": "https://omlx.ai/benchmarks/performance/jvcibpti",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-jvcibpti-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 223.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 223.2 20.1 6.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jvcibpti",
   "source_url": "https://omlx.ai/benchmarks/performance/jvcibpti",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-ajvahpc3-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: ocokv3ff; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 196.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 196.5 20.9 6.5 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ajvahpc3",
   "source_url": "https://omlx.ai/benchmarks/performance/ajvahpc3",
   "tg_tokens": null,
   "tps": 20.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-ajvahpc3-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: ocokv3ff; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 196.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 196.4 20.3 6.8 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ajvahpc3",
   "source_url": "https://omlx.ai/benchmarks/performance/ajvahpc3",
   "tg_tokens": null,
   "tps": 20.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-381hens8-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 414.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 414.7 44.5 6.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/381hens8",
   "source_url": "https://omlx.ai/benchmarks/performance/381hens8",
   "tg_tokens": null,
   "tps": 44.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-381hens8-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 411.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 411.9 43.2 7.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/381hens8",
   "source_url": "https://omlx.ai/benchmarks/performance/381hens8",
   "tg_tokens": null,
   "tps": 43.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-hwy2e99j-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: i6exorv3; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 231.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 231.9 27.2 6.5 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hwy2e99j",
   "source_url": "https://omlx.ai/benchmarks/performance/hwy2e99j",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-hwy2e99j-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: i6exorv3; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 235.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 235.2 25.9 6.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hwy2e99j",
   "source_url": "https://omlx.ai/benchmarks/performance/hwy2e99j",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-p9tfsk2o-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: 41hzqa36; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 231.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 231.9 25.3 6.5 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p9tfsk2o",
   "source_url": "https://omlx.ai/benchmarks/performance/p9tfsk2o",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-p9tfsk2o-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: 41hzqa36; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 232.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 232.5 25.0 6.8 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p9tfsk2o",
   "source_url": "https://omlx.ai/benchmarks/performance/p9tfsk2o",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-j0dficgz-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: bqy4j59f, hhxc5sho; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 893.6,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 1k 893.6 38.2 13.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j0dficgz",
   "source_url": "https://omlx.ai/benchmarks/performance/j0dficgz",
   "tg_tokens": null,
   "tps": 38.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-j0dficgz-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: bqy4j59f, hhxc5sho; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 935.7,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 4k 935.7 38.0 14.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j0dficgz",
   "source_url": "https://omlx.ai/benchmarks/performance/j0dficgz",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-j0dficgz-ctx8192",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: bqy4j59f, hhxc5sho; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 852.6,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 8k 852.6 35.3 14.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j0dficgz",
   "source_url": "https://omlx.ai/benchmarks/performance/j0dficgz",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-j0dficgz-ctx16384",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 15.3 GB at this context; duplicate traces of the same session: bqy4j59f, hhxc5sho; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 735.4,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 16k 735.4 32.5 15.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j0dficgz",
   "source_url": "https://omlx.ai/benchmarks/performance/j0dficgz",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-j0dficgz-ctx32768",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: bqy4j59f, hhxc5sho; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 699.2,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 32k 699.2 32.1 16.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j0dficgz",
   "source_url": "https://omlx.ai/benchmarks/performance/j0dficgz",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-j0dficgz-ctx65536",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: bqy4j59f, hhxc5sho; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 680.6,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 64k 680.6 27.9 18.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j0dficgz",
   "source_url": "https://omlx.ai/benchmarks/performance/j0dficgz",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-j0dficgz-ctx131072",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: bqy4j59f, hhxc5sho; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 549.4,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 128k 549.4 21.5 22.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j0dficgz",
   "source_url": "https://omlx.ai/benchmarks/performance/j0dficgz",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-r2t9yutl-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "pyros-vault MTP build; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: hj41ahwa, ziaa5bwx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 773.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "pyros-vault:Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 773.5 34.2 33.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r2t9yutl",
   "source_url": "https://omlx.ai/benchmarks/performance/r2t9yutl",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-r2t9yutl-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "pyros-vault MTP build; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: hj41ahwa, ziaa5bwx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 850.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "pyros-vault:Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 850.0 39.2 34.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r2t9yutl",
   "source_url": "https://omlx.ai/benchmarks/performance/r2t9yutl",
   "tg_tokens": null,
   "tps": 39.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-r2t9yutl-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "pyros-vault MTP build; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: hj41ahwa, ziaa5bwx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 823.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "pyros-vault:Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 823.6 50.2 35.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r2t9yutl",
   "source_url": "https://omlx.ai/benchmarks/performance/r2t9yutl",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-r2t9yutl-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "pyros-vault MTP build; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: hj41ahwa, ziaa5bwx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 752.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "pyros-vault:Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 752.8 53.2 35.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r2t9yutl",
   "source_url": "https://omlx.ai/benchmarks/performance/r2t9yutl",
   "tg_tokens": null,
   "tps": 53.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-r2t9yutl-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "pyros-vault MTP build; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: hj41ahwa, ziaa5bwx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 709.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "pyros-vault:Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 709.2 55.2 36.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r2t9yutl",
   "source_url": "https://omlx.ai/benchmarks/performance/r2t9yutl",
   "tg_tokens": null,
   "tps": 55.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-r2t9yutl-ctx65536",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "pyros-vault MTP build; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: hj41ahwa, ziaa5bwx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 617.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "pyros-vault:Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 617.3 41.5 39.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r2t9yutl",
   "source_url": "https://omlx.ai/benchmarks/performance/r2t9yutl",
   "tg_tokens": null,
   "tps": 41.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-r2t9yutl-ctx131072",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "pyros-vault MTP build; oMLX community benchmark trace; peak memory 83.2 GB at this context; duplicate traces of the same session: hj41ahwa, ziaa5bwx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 329.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "pyros-vault:Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 329.1 22.7 83.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r2t9yutl",
   "source_url": "https://omlx.ai/benchmarks/performance/r2t9yutl",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-teqbm0rb-ctx1024",
   "model": "GLM-5",
   "notes": "oMLX community benchmark trace; peak memory 420.0 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 188.7,
   "provenance": "community",
   "quant": "4.8bit",
   "quote": "GLM-5-4.8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 188.7 16.6 420.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/teqbm0rb",
   "source_url": "https://omlx.ai/benchmarks/performance/teqbm0rb",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-teqbm0rb-ctx4096",
   "model": "GLM-5",
   "notes": "oMLX community benchmark trace; peak memory 422.3 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 182.6,
   "provenance": "community",
   "quant": "4.8bit",
   "quote": "GLM-5-4.8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 182.6 13.7 422.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/teqbm0rb",
   "source_url": "https://omlx.ai/benchmarks/performance/teqbm0rb",
   "tg_tokens": null,
   "tps": 13.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-teqbm0rb-ctx8192",
   "model": "GLM-5",
   "notes": "oMLX community benchmark trace; peak memory 424.9 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 156.3,
   "provenance": "community",
   "quant": "4.8bit",
   "quote": "GLM-5-4.8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 156.3 13.5 424.9 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/teqbm0rb",
   "source_url": "https://omlx.ai/benchmarks/performance/teqbm0rb",
   "tg_tokens": null,
   "tps": 13.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-teqbm0rb-ctx16384",
   "model": "GLM-5",
   "notes": "oMLX community benchmark trace; peak memory 430.9 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 119.8,
   "provenance": "community",
   "quant": "4.8bit",
   "quote": "GLM-5-4.8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 119.8 13.0 430.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/teqbm0rb",
   "source_url": "https://omlx.ai/benchmarks/performance/teqbm0rb",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-teqbm0rb-ctx32768",
   "model": "GLM-5",
   "notes": "oMLX community benchmark trace; peak memory 443.6 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 73.2,
   "provenance": "community",
   "quant": "4.8bit",
   "quote": "GLM-5-4.8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 73.2 12.3 443.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/teqbm0rb",
   "source_url": "https://omlx.ai/benchmarks/performance/teqbm0rb",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2we8lcjq-ctx1024",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 174.0 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 496.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 1k 496.7 40.0 174.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2we8lcjq",
   "source_url": "https://omlx.ai/benchmarks/performance/2we8lcjq",
   "tg_tokens": null,
   "tps": 40.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2we8lcjq-ctx4096",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 175.0 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 554.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 4k 554.8 33.9 175.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2we8lcjq",
   "source_url": "https://omlx.ai/benchmarks/performance/2we8lcjq",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2we8lcjq-ctx8192",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 175.9 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 514.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 8k 514.0 28.9 175.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2we8lcjq",
   "source_url": "https://omlx.ai/benchmarks/performance/2we8lcjq",
   "tg_tokens": null,
   "tps": 28.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2we8lcjq-ctx16384",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 177.8 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 436.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 16k 436.3 23.5 177.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2we8lcjq",
   "source_url": "https://omlx.ai/benchmarks/performance/2we8lcjq",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2we8lcjq-ctx32768",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 181.7 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 330.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 32k 330.8 16.1 181.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2we8lcjq",
   "source_url": "https://omlx.ai/benchmarks/performance/2we8lcjq",
   "tg_tokens": null,
   "tps": 16.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2we8lcjq-ctx65536",
   "model": "MiniMax M2.5",
   "notes": "oMLX community benchmark trace; peak memory 189.6 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 221.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "MiniMax-M2.5 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 64k 221.1 9.6 189.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2we8lcjq",
   "source_url": "https://omlx.ai/benchmarks/performance/2we8lcjq",
   "tg_tokens": null,
   "tps": 9.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rhiba0g8-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 80.1 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1236.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 1,236 63.9 80.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rhiba0g8",
   "source_url": "https://omlx.ai/benchmarks/performance/rhiba0g8",
   "tg_tokens": null,
   "tps": 63.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rhiba0g8-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 80.3 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1509.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 1,509 61.9 80.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rhiba0g8",
   "source_url": "https://omlx.ai/benchmarks/performance/rhiba0g8",
   "tg_tokens": null,
   "tps": 61.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rhiba0g8-ctx8192",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 80.5 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1521.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 1,521 56.5 80.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rhiba0g8",
   "source_url": "https://omlx.ai/benchmarks/performance/rhiba0g8",
   "tg_tokens": null,
   "tps": 56.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rhiba0g8-ctx16384",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 80.9 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1434.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 16k 1,434 52.0 80.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rhiba0g8",
   "source_url": "https://omlx.ai/benchmarks/performance/rhiba0g8",
   "tg_tokens": null,
   "tps": 52.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rhiba0g8-ctx32768",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 81.8 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1262.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 32k 1,262 45.8 81.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rhiba0g8",
   "source_url": "https://omlx.ai/benchmarks/performance/rhiba0g8",
   "tg_tokens": null,
   "tps": 45.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rhiba0g8-ctx65536",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 83.6 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 994.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 64k 994.8 34.7 83.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rhiba0g8",
   "source_url": "https://omlx.ai/benchmarks/performance/rhiba0g8",
   "tg_tokens": null,
   "tps": 34.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-jca6774u-ctx1024",
   "model": "Qwen3-VL-8B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 6.0 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-VL-8B-Instruct on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 224.6 21.3 6.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jca6774u",
   "source_url": "https://omlx.ai/benchmarks/performance/jca6774u",
   "tg_tokens": null,
   "tps": 21.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-jca6774u-ctx4096",
   "model": "Qwen3-VL-8B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 217.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-VL-8B-Instruct on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 217.1 18.9 6.4 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jca6774u",
   "source_url": "https://omlx.ai/benchmarks/performance/jca6774u",
   "tg_tokens": null,
   "tps": 18.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-doabcv9b-ctx1024",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 1.5 GB at this context; duplicate traces of the same session: 3l7graoy; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 756.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 4bit; 1k 756.4 100.8 1.5 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/doabcv9b",
   "source_url": "https://omlx.ai/benchmarks/performance/doabcv9b",
   "tg_tokens": null,
   "tps": 100.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-doabcv9b-ctx4096",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 1.6 GB at this context; duplicate traces of the same session: 3l7graoy; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 734.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 4bit; 4k 734.1 91.3 1.6 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/doabcv9b",
   "source_url": "https://omlx.ai/benchmarks/performance/doabcv9b",
   "tg_tokens": null,
   "tps": 91.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-doabcv9b-ctx8192",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 1.6 GB at this context; duplicate traces of the same session: 3l7graoy; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 710.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 4bit; 8k 710.7 82.2 1.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/doabcv9b",
   "source_url": "https://omlx.ai/benchmarks/performance/doabcv9b",
   "tg_tokens": null,
   "tps": 82.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-doabcv9b-ctx16384",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 2.0 GB at this context; duplicate traces of the same session: 3l7graoy; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 647.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 4bit; 16k 647.6 68.2 2.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/doabcv9b",
   "source_url": "https://omlx.ai/benchmarks/performance/doabcv9b",
   "tg_tokens": null,
   "tps": 68.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-doabcv9b-ctx32768",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 2.6 GB at this context; duplicate traces of the same session: 3l7graoy; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 552.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 4bit; 32k 552.2 47.8 2.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/doabcv9b",
   "source_url": "https://omlx.ai/benchmarks/performance/doabcv9b",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-doabcv9b-ctx65536",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 4.1 GB at this context; duplicate traces of the same session: 3l7graoy; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 408.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 4bit; 64k 408.6 17.3 4.1 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/doabcv9b",
   "source_url": "https://omlx.ai/benchmarks/performance/doabcv9b",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-wpqf311n-ctx1024",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 1.8 GB at this context; duplicate traces of the same session: f00eert8; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 755.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 8bit; 1k 755.9 62.4 1.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wpqf311n",
   "source_url": "https://omlx.ai/benchmarks/performance/wpqf311n",
   "tg_tokens": null,
   "tps": 62.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-wpqf311n-ctx4096",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 1.9 GB at this context; duplicate traces of the same session: f00eert8; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 733.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 8bit; 4k 733.9 58.9 1.9 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wpqf311n",
   "source_url": "https://omlx.ai/benchmarks/performance/wpqf311n",
   "tg_tokens": null,
   "tps": 58.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-wpqf311n-ctx8192",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 2.0 GB at this context; duplicate traces of the same session: f00eert8; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 714.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 8bit; 8k 714.0 55.0 2.0 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wpqf311n",
   "source_url": "https://omlx.ai/benchmarks/performance/wpqf311n",
   "tg_tokens": null,
   "tps": 55.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-wpqf311n-ctx16384",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 2.4 GB at this context; duplicate traces of the same session: f00eert8; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 654.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 8bit; 16k 654.9 47.9 2.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wpqf311n",
   "source_url": "https://omlx.ai/benchmarks/performance/wpqf311n",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-wpqf311n-ctx32768",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 3.0 GB at this context; duplicate traces of the same session: f00eert8; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 554.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 8bit; 32k 554.7 37.3 3.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wpqf311n",
   "source_url": "https://omlx.ai/benchmarks/performance/wpqf311n",
   "tg_tokens": null,
   "tps": 37.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-wpqf311n-ctx65536",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 4.4 GB at this context; duplicate traces of the same session: f00eert8; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 414.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 8bit; 64k 414.9 24.8 4.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wpqf311n",
   "source_url": "https://omlx.ai/benchmarks/performance/wpqf311n",
   "tg_tokens": null,
   "tps": 24.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ejitl07v-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.3 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 704.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 704.1 53.8 66.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ejitl07v",
   "source_url": "https://omlx.ai/benchmarks/performance/ejitl07v",
   "tg_tokens": null,
   "tps": 53.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ejitl07v-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.7 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 729.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 729.1 51.2 66.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ejitl07v",
   "source_url": "https://omlx.ai/benchmarks/performance/ejitl07v",
   "tg_tokens": null,
   "tps": 51.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ejitl07v-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.0 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 712.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 712.9 48.2 67.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ejitl07v",
   "source_url": "https://omlx.ai/benchmarks/performance/ejitl07v",
   "tg_tokens": null,
   "tps": 48.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ejitl07v-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.7 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 670.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 670.3 41.2 67.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ejitl07v",
   "source_url": "https://omlx.ai/benchmarks/performance/ejitl07v",
   "tg_tokens": null,
   "tps": 41.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ejitl07v-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 69.1 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 590.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 590.2 35.0 69.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ejitl07v",
   "source_url": "https://omlx.ai/benchmarks/performance/ejitl07v",
   "tg_tokens": null,
   "tps": 35.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ejitl07v-ctx65536",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 71.9 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 466.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 466.9 26.0 71.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ejitl07v",
   "source_url": "https://omlx.ai/benchmarks/performance/ejitl07v",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bzhe0vh3-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.3 GB at this context; duplicate traces of the same session: 919dis83; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 638.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 638.7 54.4 66.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bzhe0vh3",
   "source_url": "https://omlx.ai/benchmarks/performance/bzhe0vh3",
   "tg_tokens": null,
   "tps": 54.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bzhe0vh3-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.7 GB at this context; duplicate traces of the same session: 919dis83; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 714.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 714.1 50.9 66.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bzhe0vh3",
   "source_url": "https://omlx.ai/benchmarks/performance/bzhe0vh3",
   "tg_tokens": null,
   "tps": 50.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bzhe0vh3-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.0 GB at this context; duplicate traces of the same session: 919dis83; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 712.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 712.7 47.3 67.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bzhe0vh3",
   "source_url": "https://omlx.ai/benchmarks/performance/bzhe0vh3",
   "tg_tokens": null,
   "tps": 47.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-bzhe0vh3-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.7 GB at this context; duplicate traces of the same session: 919dis83; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 670.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 16k 670.7 35.9 67.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bzhe0vh3",
   "source_url": "https://omlx.ai/benchmarks/performance/bzhe0vh3",
   "tg_tokens": null,
   "tps": 35.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-bzhe0vh3-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 69.1 GB at this context; duplicate traces of the same session: 919dis83; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 587.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 587.5 24.4 69.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bzhe0vh3",
   "source_url": "https://omlx.ai/benchmarks/performance/bzhe0vh3",
   "tg_tokens": null,
   "tps": 24.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-bzhe0vh3-ctx65536",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 71.9 GB at this context; duplicate traces of the same session: 919dis83; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 458.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k 458.6 15.5 71.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bzhe0vh3",
   "source_url": "https://omlx.ai/benchmarks/performance/bzhe0vh3",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-wtdigojg-ctx1024",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: fwtc8lbd; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 180.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 180.4 32.2 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wtdigojg",
   "source_url": "https://omlx.ai/benchmarks/performance/wtdigojg",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-wtdigojg-ctx4096",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: fwtc8lbd; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 182.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 182.4 30.4 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wtdigojg",
   "source_url": "https://omlx.ai/benchmarks/performance/wtdigojg",
   "tg_tokens": null,
   "tps": 30.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-wtdigojg-ctx8192",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: fwtc8lbd; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 180.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 8k 180.9 28.4 17.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wtdigojg",
   "source_url": "https://omlx.ai/benchmarks/performance/wtdigojg",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-wtdigojg-ctx16384",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: fwtc8lbd; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 176.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 16k 176.2 25.3 18.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wtdigojg",
   "source_url": "https://omlx.ai/benchmarks/performance/wtdigojg",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-i0ji024h-ctx1024",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 226.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 1k 226.5 33.4 16.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i0ji024h",
   "source_url": "https://omlx.ai/benchmarks/performance/i0ji024h",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-i0ji024h-ctx4096",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 228.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 4k 228.9 31.3 17.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i0ji024h",
   "source_url": "https://omlx.ai/benchmarks/performance/i0ji024h",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-i0ji024h-ctx8192",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 227.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 8k 227.0 29.3 17.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i0ji024h",
   "source_url": "https://omlx.ai/benchmarks/performance/i0ji024h",
   "tg_tokens": null,
   "tps": 29.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-i0ji024h-ctx16384",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 222.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 16k 222.3 26.6 18.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i0ji024h",
   "source_url": "https://omlx.ai/benchmarks/performance/i0ji024h",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-i0ji024h-ctx32768",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 212.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 32k 212.9 21.9 20.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i0ji024h",
   "source_url": "https://omlx.ai/benchmarks/performance/i0ji024h",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-i0ji024h-ctx65536",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 192.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 64k 192.2 16.8 23.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i0ji024h",
   "source_url": "https://omlx.ai/benchmarks/performance/i0ji024h",
   "tg_tokens": null,
   "tps": 16.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-2fbfl41g-ctx1024",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 226.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 1k 226.5 33.5 16.6 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2fbfl41g",
   "source_url": "https://omlx.ai/benchmarks/performance/2fbfl41g",
   "tg_tokens": null,
   "tps": 33.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-2fbfl41g-ctx4096",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 228.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 4k 228.4 31.5 17.0 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2fbfl41g",
   "source_url": "https://omlx.ai/benchmarks/performance/2fbfl41g",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-iff2zzcf-ctx1024",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 322.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 322.7 38.0 16.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iff2zzcf",
   "source_url": "https://omlx.ai/benchmarks/performance/iff2zzcf",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-iff2zzcf-ctx4096",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 324.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 324.2 35.3 17.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iff2zzcf",
   "source_url": "https://omlx.ai/benchmarks/performance/iff2zzcf",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-iff2zzcf-ctx8192",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 319.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 319.7 33.6 17.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iff2zzcf",
   "source_url": "https://omlx.ai/benchmarks/performance/iff2zzcf",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-iff2zzcf-ctx16384",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 310.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 310.0 30.0 18.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iff2zzcf",
   "source_url": "https://omlx.ai/benchmarks/performance/iff2zzcf",
   "tg_tokens": null,
   "tps": 30.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-iff2zzcf-ctx32768",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 291.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 291.3 24.9 20.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iff2zzcf",
   "source_url": "https://omlx.ai/benchmarks/performance/iff2zzcf",
   "tg_tokens": null,
   "tps": 24.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-iff2zzcf-ctx65536",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 258.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 258.7 18.1 23.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iff2zzcf",
   "source_url": "https://omlx.ai/benchmarks/performance/iff2zzcf",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-lsrf1njm-ctx1024",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 240.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 240.5 20.5 22.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lsrf1njm",
   "source_url": "https://omlx.ai/benchmarks/performance/lsrf1njm",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-lsrf1njm-ctx4096",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 235.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 235.0 19.5 23.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lsrf1njm",
   "source_url": "https://omlx.ai/benchmarks/performance/lsrf1njm",
   "tg_tokens": null,
   "tps": 19.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-3rjvar3g-ctx1024",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 3grd36vj, mhf4fjks, pfl2vdmm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 125.6 15.2 16.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3rjvar3g",
   "source_url": "https://omlx.ai/benchmarks/performance/3rjvar3g",
   "tg_tokens": null,
   "tps": 15.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-3rjvar3g-ctx4096",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 3grd36vj, mhf4fjks, pfl2vdmm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 125.9 14.8 17.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3rjvar3g",
   "source_url": "https://omlx.ai/benchmarks/performance/3rjvar3g",
   "tg_tokens": null,
   "tps": 14.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-3rjvar3g-ctx8192",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: 3grd36vj, mhf4fjks, pfl2vdmm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 124.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 124.3 14.2 17.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3rjvar3g",
   "source_url": "https://omlx.ai/benchmarks/performance/3rjvar3g",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-3rjvar3g-ctx16384",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: 3grd36vj, mhf4fjks, pfl2vdmm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 121.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 121.0 12.3 18.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3rjvar3g",
   "source_url": "https://omlx.ai/benchmarks/performance/3rjvar3g",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-3rjvar3g-ctx32768",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 3grd36vj, mhf4fjks, pfl2vdmm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 114.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 114.0 9.7 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3rjvar3g",
   "source_url": "https://omlx.ai/benchmarks/performance/3rjvar3g",
   "tg_tokens": null,
   "tps": 9.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-vmky9zzs-ctx1024",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 128.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 1k 128.6 24.7 15.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vmky9zzs",
   "source_url": "https://omlx.ai/benchmarks/performance/vmky9zzs",
   "tg_tokens": null,
   "tps": 24.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-vmky9zzs-ctx4096",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 133.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 4k 133.5 23.0 16.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vmky9zzs",
   "source_url": "https://omlx.ai/benchmarks/performance/vmky9zzs",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ipsj46i8-ctx1024",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: mfwckz8w, qyady0u2; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 111.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 1k 111.1 21.1 15.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ipsj46i8",
   "source_url": "https://omlx.ai/benchmarks/performance/ipsj46i8",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ipsj46i8-ctx4096",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: mfwckz8w, qyady0u2; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 113.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k 113.8 20.1 16.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ipsj46i8",
   "source_url": "https://omlx.ai/benchmarks/performance/ipsj46i8",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ipsj46i8-ctx8192",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: mfwckz8w, qyady0u2; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 113.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 8k 113.1 19.0 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ipsj46i8",
   "source_url": "https://omlx.ai/benchmarks/performance/ipsj46i8",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ipsj46i8-ctx16384",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 17.5 GB at this context; duplicate traces of the same session: mfwckz8w, qyady0u2; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 111.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 16k 111.0 17.3 17.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ipsj46i8",
   "source_url": "https://omlx.ai/benchmarks/performance/ipsj46i8",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ipsj46i8-ctx32768",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: mfwckz8w, qyady0u2; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 32k 106.7 14.9 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ipsj46i8",
   "source_url": "https://omlx.ai/benchmarks/performance/ipsj46i8",
   "tg_tokens": null,
   "tps": 14.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-q76ia64b-ctx1024",
   "model": "Qwen3.5-2B-Claude-4.6-Opus-Reasoning-Distilled",
   "notes": "oMLX community benchmark trace; peak memory 4.1 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 860.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.5-2B-Claude-4.6-Opus-Reasoning-Distilled on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 bf16; 1k 860.0 25.1 4.1 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q76ia64b",
   "source_url": "https://omlx.ai/benchmarks/performance/q76ia64b",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-q76ia64b-ctx4096",
   "model": "Qwen3.5-2B-Claude-4.6-Opus-Reasoning-Distilled",
   "notes": "oMLX community benchmark trace; peak memory 4.3 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1059.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.5-2B-Claude-4.6-Opus-Reasoning-Distilled on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 bf16; 4k 1,059 24.6 4.3 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q76ia64b",
   "source_url": "https://omlx.ai/benchmarks/performance/q76ia64b",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-3xet8brb-ctx1024",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: femieeij, ptzhx9ln; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1652.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,652 93.4 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3xet8brb",
   "source_url": "https://omlx.ai/benchmarks/performance/3xet8brb",
   "tg_tokens": null,
   "tps": 93.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-3xet8brb-ctx4096",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: femieeij, ptzhx9ln; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1706.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,706 88.6 20.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3xet8brb",
   "source_url": "https://omlx.ai/benchmarks/performance/3xet8brb",
   "tg_tokens": null,
   "tps": 88.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-3xet8brb-ctx8192",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: femieeij, ptzhx9ln; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1487.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,487 90.0 20.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3xet8brb",
   "source_url": "https://omlx.ai/benchmarks/performance/3xet8brb",
   "tg_tokens": null,
   "tps": 90.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-3xet8brb-ctx16384",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: femieeij, ptzhx9ln; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1227.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,227 63.7 20.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3xet8brb",
   "source_url": "https://omlx.ai/benchmarks/performance/3xet8brb",
   "tg_tokens": null,
   "tps": 63.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-3xet8brb-ctx32768",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: femieeij, ptzhx9ln; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1026.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 1,026 42.5 21.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3xet8brb",
   "source_url": "https://omlx.ai/benchmarks/performance/3xet8brb",
   "tg_tokens": null,
   "tps": 42.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-vpoz1qjs-ctx1024",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 821.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 821.5 75.6 20.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vpoz1qjs",
   "source_url": "https://omlx.ai/benchmarks/performance/vpoz1qjs",
   "tg_tokens": null,
   "tps": 75.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-vpoz1qjs-ctx4096",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 879.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 879.6 79.1 20.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vpoz1qjs",
   "source_url": "https://omlx.ai/benchmarks/performance/vpoz1qjs",
   "tg_tokens": null,
   "tps": 79.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-vpoz1qjs-ctx8192",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 873.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 873.4 71.9 20.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vpoz1qjs",
   "source_url": "https://omlx.ai/benchmarks/performance/vpoz1qjs",
   "tg_tokens": null,
   "tps": 71.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-tzmnrcg3-ctx1024",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 9bgkvghg, cq1j63z1, fxnkfdkj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 867.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 867.2 82.8 20.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tzmnrcg3",
   "source_url": "https://omlx.ai/benchmarks/performance/tzmnrcg3",
   "tg_tokens": null,
   "tps": 82.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-tzmnrcg3-ctx4096",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 9bgkvghg, cq1j63z1, fxnkfdkj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 899.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 899.9 79.2 20.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tzmnrcg3",
   "source_url": "https://omlx.ai/benchmarks/performance/tzmnrcg3",
   "tg_tokens": null,
   "tps": 79.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-tzmnrcg3-ctx8192",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: 9bgkvghg, cq1j63z1, fxnkfdkj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 876.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 876.2 74.9 20.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tzmnrcg3",
   "source_url": "https://omlx.ai/benchmarks/performance/tzmnrcg3",
   "tg_tokens": null,
   "tps": 74.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-tzmnrcg3-ctx16384",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 9bgkvghg, cq1j63z1, fxnkfdkj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 805.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 805.6 57.7 20.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tzmnrcg3",
   "source_url": "https://omlx.ai/benchmarks/performance/tzmnrcg3",
   "tg_tokens": null,
   "tps": 57.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-tzmnrcg3-ctx32768",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: 9bgkvghg, cq1j63z1, fxnkfdkj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 693.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 693.2 40.6 21.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tzmnrcg3",
   "source_url": "https://omlx.ai/benchmarks/performance/tzmnrcg3",
   "tg_tokens": null,
   "tps": 40.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-tzmnrcg3-ctx65536",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 9bgkvghg, cq1j63z1, fxnkfdkj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 537.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 537.6 25.9 23.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tzmnrcg3",
   "source_url": "https://omlx.ai/benchmarks/performance/tzmnrcg3",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4o3fa12n-ctx1024",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 210.4 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 381.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 381.0 37.0 210.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4o3fa12n",
   "source_url": "https://omlx.ai/benchmarks/performance/4o3fa12n",
   "tg_tokens": null,
   "tps": 37.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4o3fa12n-ctx4096",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 210.7 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 404.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 404.5 35.3 210.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4o3fa12n",
   "source_url": "https://omlx.ai/benchmarks/performance/4o3fa12n",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4o3fa12n-ctx8192",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 211.0 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 400.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 400.7 31.3 211.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4o3fa12n",
   "source_url": "https://omlx.ai/benchmarks/performance/4o3fa12n",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4o3fa12n-ctx16384",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 211.8 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 384.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 384.0 24.8 211.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4o3fa12n",
   "source_url": "https://omlx.ai/benchmarks/performance/4o3fa12n",
   "tg_tokens": null,
   "tps": 24.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4o3fa12n-ctx32768",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 213.2 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 350.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 350.4 17.8 213.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4o3fa12n",
   "source_url": "https://omlx.ai/benchmarks/performance/4o3fa12n",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4o3fa12n-ctx65536",
   "model": "Qwen3.5-397B-A17B",
   "notes": "oMLX community benchmark trace; peak memory 216.2 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 294.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-397B-A17B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 294.3 11.7 216.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4o3fa12n",
   "source_url": "https://omlx.ai/benchmarks/performance/4o3fa12n",
   "tg_tokens": null,
   "tps": 11.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-o7u1nsfl-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.0 GB at this context; duplicate traces of the same session: 0ysklzxp, aeq42ap1, lsp40v2a; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 736.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 1k 736.4 102.2 4.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o7u1nsfl",
   "source_url": "https://omlx.ai/benchmarks/performance/o7u1nsfl",
   "tg_tokens": null,
   "tps": 102.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-o7u1nsfl-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: 0ysklzxp, aeq42ap1, lsp40v2a; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 734.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k 734.5 95.4 4.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o7u1nsfl",
   "source_url": "https://omlx.ai/benchmarks/performance/o7u1nsfl",
   "tg_tokens": null,
   "tps": 95.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-o7u1nsfl-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.6 GB at this context; duplicate traces of the same session: 0ysklzxp, aeq42ap1, lsp40v2a; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 720.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 8k 720.3 87.0 4.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o7u1nsfl",
   "source_url": "https://omlx.ai/benchmarks/performance/o7u1nsfl",
   "tg_tokens": null,
   "tps": 87.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-o7u1nsfl-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 5.3 GB at this context; duplicate traces of the same session: 0ysklzxp, aeq42ap1, lsp40v2a; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 691.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 16k 691.3 73.4 5.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o7u1nsfl",
   "source_url": "https://omlx.ai/benchmarks/performance/o7u1nsfl",
   "tg_tokens": null,
   "tps": 73.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-o7u1nsfl-ctx32768",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: 0ysklzxp, aeq42ap1, lsp40v2a; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 576.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 32k 576.0 28.6 6.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o7u1nsfl",
   "source_url": "https://omlx.ai/benchmarks/performance/o7u1nsfl",
   "tg_tokens": null,
   "tps": 28.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-o7u1nsfl-ctx65536",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 8.6 GB at this context; duplicate traces of the same session: 0ysklzxp, aeq42ap1, lsp40v2a; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 468.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 64k 468.7 18.2 8.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o7u1nsfl",
   "source_url": "https://omlx.ai/benchmarks/performance/o7u1nsfl",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-zwuvz3cp-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.0 GB at this context; duplicate traces of the same session: 0oga1cf0, j290pxfi; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1428.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 1k 1,428 137.5 4.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zwuvz3cp",
   "source_url": "https://omlx.ai/benchmarks/performance/zwuvz3cp",
   "tg_tokens": null,
   "tps": 137.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-zwuvz3cp-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: 0oga1cf0, j290pxfi; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1440.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 4k 1,440 128.2 4.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zwuvz3cp",
   "source_url": "https://omlx.ai/benchmarks/performance/zwuvz3cp",
   "tg_tokens": null,
   "tps": 128.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-zwuvz3cp-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.6 GB at this context; duplicate traces of the same session: 0oga1cf0, j290pxfi; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1414.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 8k 1,414 117.4 4.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zwuvz3cp",
   "source_url": "https://omlx.ai/benchmarks/performance/zwuvz3cp",
   "tg_tokens": null,
   "tps": 117.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-zwuvz3cp-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 5.3 GB at this context; duplicate traces of the same session: 0oga1cf0, j290pxfi; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1355.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 16k 1,355 98.7 5.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zwuvz3cp",
   "source_url": "https://omlx.ai/benchmarks/performance/zwuvz3cp",
   "tg_tokens": null,
   "tps": 98.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-zwuvz3cp-ctx32768",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: 0oga1cf0, j290pxfi; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1247.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 32k 1,247 74.5 6.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zwuvz3cp",
   "source_url": "https://omlx.ai/benchmarks/performance/zwuvz3cp",
   "tg_tokens": null,
   "tps": 74.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-zwuvz3cp-ctx65536",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 8.6 GB at this context; duplicate traces of the same session: 0oga1cf0, j290pxfi; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1051.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 64k 1,051 50.4 8.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zwuvz3cp",
   "source_url": "https://omlx.ai/benchmarks/performance/zwuvz3cp",
   "tg_tokens": null,
   "tps": 50.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-4vf3fhkn-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: 5t6c2eu1, cgvsvi3i; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1423.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 1k 1,423 99.9 5.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4vf3fhkn",
   "source_url": "https://omlx.ai/benchmarks/performance/4vf3fhkn",
   "tg_tokens": null,
   "tps": 99.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-4vf3fhkn-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 6.1 GB at this context; duplicate traces of the same session: 5t6c2eu1, cgvsvi3i; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1448.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 4k 1,448 96.2 6.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4vf3fhkn",
   "source_url": "https://omlx.ai/benchmarks/performance/4vf3fhkn",
   "tg_tokens": null,
   "tps": 96.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-4vf3fhkn-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: 5t6c2eu1, cgvsvi3i; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1421.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 8k 1,421 90.1 6.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4vf3fhkn",
   "source_url": "https://omlx.ai/benchmarks/performance/4vf3fhkn",
   "tg_tokens": null,
   "tps": 90.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-4vf3fhkn-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 7.3 GB at this context; duplicate traces of the same session: 5t6c2eu1, cgvsvi3i; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1362.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 16k 1,362 80.1 7.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4vf3fhkn",
   "source_url": "https://omlx.ai/benchmarks/performance/4vf3fhkn",
   "tg_tokens": null,
   "tps": 80.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-4vf3fhkn-ctx32768",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 7.9 GB at this context; duplicate traces of the same session: 5t6c2eu1, cgvsvi3i; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1255.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 32k 1,255 63.0 7.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4vf3fhkn",
   "source_url": "https://omlx.ai/benchmarks/performance/4vf3fhkn",
   "tg_tokens": null,
   "tps": 63.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-4vf3fhkn-ctx65536",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 9.9 GB at this context; duplicate traces of the same session: 5t6c2eu1, cgvsvi3i; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1056.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 64k 1,056 45.5 9.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4vf3fhkn",
   "source_url": "https://omlx.ai/benchmarks/performance/4vf3fhkn",
   "tg_tokens": null,
   "tps": 45.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-o03mk72e-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 406.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 406.4 37.9 3.9 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o03mk72e",
   "source_url": "https://omlx.ai/benchmarks/performance/o03mk72e",
   "tg_tokens": null,
   "tps": 37.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-o03mk72e-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 403.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 403.4 35.5 4.2 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o03mk72e",
   "source_url": "https://omlx.ai/benchmarks/performance/o03mk72e",
   "tg_tokens": null,
   "tps": 35.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-jtr6rvgo-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 397.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 397.8 37.1 3.9 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jtr6rvgo",
   "source_url": "https://omlx.ai/benchmarks/performance/jtr6rvgo",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-jtr6rvgo-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 396.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 396.2 34.2 4.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jtr6rvgo",
   "source_url": "https://omlx.ai/benchmarks/performance/jtr6rvgo",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-exfl6ecd-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: 91h2g4t3, s11jretg, tir4w3vr, tirqyj3o; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 374.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 374.9 33.7 3.9 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/exfl6ecd",
   "source_url": "https://omlx.ai/benchmarks/performance/exfl6ecd",
   "tg_tokens": null,
   "tps": 33.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-exfl6ecd-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: 91h2g4t3, s11jretg, tir4w3vr, tirqyj3o; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 373.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 373.3 32.0 4.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/exfl6ecd",
   "source_url": "https://omlx.ai/benchmarks/performance/exfl6ecd",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-exfl6ecd-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.6 GB at this context; duplicate traces of the same session: 91h2g4t3, s11jretg, tir4w3vr, tirqyj3o; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 364.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 8k 364.6 29.5 4.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/exfl6ecd",
   "source_url": "https://omlx.ai/benchmarks/performance/exfl6ecd",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-exfl6ecd-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.8 GB at this context; duplicate traces of the same session: 91h2g4t3, s11jretg, tir4w3vr, tirqyj3o; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 340.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 16k 340.6 25.2 4.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/exfl6ecd",
   "source_url": "https://omlx.ai/benchmarks/performance/exfl6ecd",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-exfl6ecd-ctx32768",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: 91h2g4t3, s11jretg, tir4w3vr, tirqyj3o; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 299.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 32k 299.7 20.3 5.9 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/exfl6ecd",
   "source_url": "https://omlx.ai/benchmarks/performance/exfl6ecd",
   "tg_tokens": null,
   "tps": 20.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-exfl6ecd-ctx65536",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 8.0 GB at this context; duplicate traces of the same session: 91h2g4t3, s11jretg, tir4w3vr, tirqyj3o; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 231.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 64k 231.3 13.6 8.0 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/exfl6ecd",
   "source_url": "https://omlx.ai/benchmarks/performance/exfl6ecd",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-mhaypknn-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: rkslpdmh; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 406.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 406.1 37.4 3.9 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mhaypknn",
   "source_url": "https://omlx.ai/benchmarks/performance/mhaypknn",
   "tg_tokens": null,
   "tps": 37.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-mhaypknn-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: rkslpdmh; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 403.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 403.8 36.0 4.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mhaypknn",
   "source_url": "https://omlx.ai/benchmarks/performance/mhaypknn",
   "tg_tokens": null,
   "tps": 36.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-mhaypknn-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.6 GB at this context; duplicate traces of the same session: rkslpdmh; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 385.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 8k 385.2 31.8 4.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mhaypknn",
   "source_url": "https://omlx.ai/benchmarks/performance/mhaypknn",
   "tg_tokens": null,
   "tps": 31.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-c1jz32zu-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.0 GB at this context; duplicate traces of the same session: j3udbnrh, nvxbhp1x, qsyh8wbm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 811.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 811.1 82.1 4.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c1jz32zu",
   "source_url": "https://omlx.ai/benchmarks/performance/c1jz32zu",
   "tg_tokens": null,
   "tps": 82.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-c1jz32zu-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: j3udbnrh, nvxbhp1x, qsyh8wbm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 798.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 798.0 76.8 4.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c1jz32zu",
   "source_url": "https://omlx.ai/benchmarks/performance/c1jz32zu",
   "tg_tokens": null,
   "tps": 76.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-c1jz32zu-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.6 GB at this context; duplicate traces of the same session: j3udbnrh, nvxbhp1x, qsyh8wbm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 781.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 781.9 72.4 4.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c1jz32zu",
   "source_url": "https://omlx.ai/benchmarks/performance/c1jz32zu",
   "tg_tokens": null,
   "tps": 72.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-c1jz32zu-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 5.3 GB at this context; duplicate traces of the same session: j3udbnrh, nvxbhp1x, qsyh8wbm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 737.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 737.3 62.8 5.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c1jz32zu",
   "source_url": "https://omlx.ai/benchmarks/performance/c1jz32zu",
   "tg_tokens": null,
   "tps": 62.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-c1jz32zu-ctx32768",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: j3udbnrh, nvxbhp1x, qsyh8wbm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 655.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 655.0 49.9 6.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c1jz32zu",
   "source_url": "https://omlx.ai/benchmarks/performance/c1jz32zu",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-c1jz32zu-ctx65536",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 8.6 GB at this context; duplicate traces of the same session: j3udbnrh, nvxbhp1x, qsyh8wbm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 534.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 534.5 36.3 8.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c1jz32zu",
   "source_url": "https://omlx.ai/benchmarks/performance/c1jz32zu",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-hyclrwcc-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.7 GB at this context; duplicate traces of the same session: 1n8qgmq9, 6rngmu2d; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 239.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 8bit; 1k 239.3 35.1 10.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hyclrwcc",
   "source_url": "https://omlx.ai/benchmarks/performance/hyclrwcc",
   "tg_tokens": null,
   "tps": 35.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-hyclrwcc-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.0 GB at this context; duplicate traces of the same session: 1n8qgmq9, 6rngmu2d; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 242.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 8bit; 4k 242.1 33.3 11.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hyclrwcc",
   "source_url": "https://omlx.ai/benchmarks/performance/hyclrwcc",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-hyclrwcc-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.2 GB at this context; duplicate traces of the same session: 1n8qgmq9, 6rngmu2d; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 236.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 8bit; 8k 236.2 30.3 11.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hyclrwcc",
   "source_url": "https://omlx.ai/benchmarks/performance/hyclrwcc",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-hyclrwcc-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.7 GB at this context; duplicate traces of the same session: 1n8qgmq9, 6rngmu2d; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 234.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 8bit; 16k 234.4 26.2 11.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hyclrwcc",
   "source_url": "https://omlx.ai/benchmarks/performance/hyclrwcc",
   "tg_tokens": null,
   "tps": 26.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-hyclrwcc-ctx32768",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: 1n8qgmq9, 6rngmu2d; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 8bit; 32k 224.8 20.3 12.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hyclrwcc",
   "source_url": "https://omlx.ai/benchmarks/performance/hyclrwcc",
   "tg_tokens": null,
   "tps": 20.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-hyclrwcc-ctx65536",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: 1n8qgmq9, 6rngmu2d; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 8bit; 64k 205.0 13.9 14.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hyclrwcc",
   "source_url": "https://omlx.ai/benchmarks/performance/hyclrwcc",
   "tg_tokens": null,
   "tps": 13.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-opjss9el-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: none; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 138.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 1k 138.0 30.0 6.7 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/opjss9el",
   "source_url": "https://omlx.ai/benchmarks/performance/opjss9el",
   "tg_tokens": null,
   "tps": 30.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-opjss9el-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: none; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 138.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 4k 138.6 28.7 7.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/opjss9el",
   "source_url": "https://omlx.ai/benchmarks/performance/opjss9el",
   "tg_tokens": null,
   "tps": 28.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-kxuicg18-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 131.1 17.3 6.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kxuicg18",
   "source_url": "https://omlx.ai/benchmarks/performance/kxuicg18",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-kxuicg18-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 130.7 16.9 7.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kxuicg18",
   "source_url": "https://omlx.ai/benchmarks/performance/kxuicg18",
   "tg_tokens": null,
   "tps": 16.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-a6fmmdir-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.7 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 775.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 1k 775.8 74.3 8.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a6fmmdir",
   "source_url": "https://omlx.ai/benchmarks/performance/a6fmmdir",
   "tg_tokens": null,
   "tps": 74.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-a6fmmdir-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.9 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 783.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 4k 783.9 72.1 8.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a6fmmdir",
   "source_url": "https://omlx.ai/benchmarks/performance/a6fmmdir",
   "tg_tokens": null,
   "tps": 72.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-a6fmmdir-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 9.2 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 776.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 8k 776.7 68.7 9.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a6fmmdir",
   "source_url": "https://omlx.ai/benchmarks/performance/a6fmmdir",
   "tg_tokens": null,
   "tps": 68.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-a6fmmdir-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 9.6 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 758.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 16k 758.8 62.8 9.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a6fmmdir",
   "source_url": "https://omlx.ai/benchmarks/performance/a6fmmdir",
   "tg_tokens": null,
   "tps": 62.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-a6fmmdir-ctx32768",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.6 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 724.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 32k 724.1 52.6 10.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a6fmmdir",
   "source_url": "https://omlx.ai/benchmarks/performance/a6fmmdir",
   "tg_tokens": null,
   "tps": 52.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-a6fmmdir-ctx65536",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 653.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 64k 653.8 39.8 12.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a6fmmdir",
   "source_url": "https://omlx.ai/benchmarks/performance/a6fmmdir",
   "tg_tokens": null,
   "tps": 39.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-voj8az1w-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.7 GB at this context; duplicate traces of the same session: 6o6zm8tl, wxd1vexu; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 794.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 1k 794.0 64.3 10.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/voj8az1w",
   "source_url": "https://omlx.ai/benchmarks/performance/voj8az1w",
   "tg_tokens": null,
   "tps": 64.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-voj8az1w-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.0 GB at this context; duplicate traces of the same session: 6o6zm8tl, wxd1vexu; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 804.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 4k 804.0 62.6 11.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/voj8az1w",
   "source_url": "https://omlx.ai/benchmarks/performance/voj8az1w",
   "tg_tokens": null,
   "tps": 62.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-voj8az1w-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.2 GB at this context; duplicate traces of the same session: 6o6zm8tl, wxd1vexu; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 797.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 8k 797.4 59.7 11.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/voj8az1w",
   "source_url": "https://omlx.ai/benchmarks/performance/voj8az1w",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-voj8az1w-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.7 GB at this context; duplicate traces of the same session: 6o6zm8tl, wxd1vexu; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 778.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 16k 778.7 54.0 11.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/voj8az1w",
   "source_url": "https://omlx.ai/benchmarks/performance/voj8az1w",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-voj8az1w-ctx32768",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: 6o6zm8tl, wxd1vexu; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 742.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 32k 742.0 46.4 12.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/voj8az1w",
   "source_url": "https://omlx.ai/benchmarks/performance/voj8az1w",
   "tg_tokens": null,
   "tps": 46.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-voj8az1w-ctx65536",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: 6o6zm8tl, wxd1vexu; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 668.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 64k 668.6 33.9 14.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/voj8az1w",
   "source_url": "https://omlx.ai/benchmarks/performance/voj8az1w",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-f4y5m3yt-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 722.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 1k 722.9 97.0 6.7 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f4y5m3yt",
   "source_url": "https://omlx.ai/benchmarks/performance/f4y5m3yt",
   "tg_tokens": null,
   "tps": 97.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-f4y5m3yt-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 786.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 4k 786.3 93.7 7.0 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f4y5m3yt",
   "source_url": "https://omlx.ai/benchmarks/performance/f4y5m3yt",
   "tg_tokens": null,
   "tps": 93.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-v9s757f5-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: mk5718cc, r7jnwmp2, rr8vgmtb, zd9pv0ve; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 727.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 1k 727.6 96.7 6.7 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v9s757f5",
   "source_url": "https://omlx.ai/benchmarks/performance/v9s757f5",
   "tg_tokens": null,
   "tps": 96.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-v9s757f5-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: mk5718cc, r7jnwmp2, rr8vgmtb, zd9pv0ve; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 789.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 4k 789.6 93.2 7.0 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v9s757f5",
   "source_url": "https://omlx.ai/benchmarks/performance/v9s757f5",
   "tg_tokens": null,
   "tps": 93.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-v9s757f5-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.1 GB at this context; duplicate traces of the same session: mk5718cc, r7jnwmp2, rr8vgmtb, zd9pv0ve; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 792.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 8k 792.6 87.6 7.1 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v9s757f5",
   "source_url": "https://omlx.ai/benchmarks/performance/v9s757f5",
   "tg_tokens": null,
   "tps": 87.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-v9s757f5-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.6 GB at this context; duplicate traces of the same session: mk5718cc, r7jnwmp2, rr8vgmtb, zd9pv0ve; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 774.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 16k 774.4 76.6 7.6 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v9s757f5",
   "source_url": "https://omlx.ai/benchmarks/performance/v9s757f5",
   "tg_tokens": null,
   "tps": 76.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-v9s757f5-ctx32768",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.7 GB at this context; duplicate traces of the same session: mk5718cc, r7jnwmp2, rr8vgmtb, zd9pv0ve; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 738.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 32k 738.1 61.9 8.7 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v9s757f5",
   "source_url": "https://omlx.ai/benchmarks/performance/v9s757f5",
   "tg_tokens": null,
   "tps": 61.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-v9s757f5-ctx65536",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.8 GB at this context; duplicate traces of the same session: mk5718cc, r7jnwmp2, rr8vgmtb, zd9pv0ve; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 665.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 64k 665.7 44.3 10.8 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v9s757f5",
   "source_url": "https://omlx.ai/benchmarks/performance/v9s757f5",
   "tg_tokens": null,
   "tps": 44.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-p13i2hbr-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 174.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M3 (10c); M3 (10c) \u00b7 16 GB \u00b7 4bit; 1k 174.7 17.7 6.5 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p13i2hbr",
   "source_url": "https://omlx.ai/benchmarks/performance/p13i2hbr",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-p13i2hbr-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 157.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M3 (10c); M3 (10c) \u00b7 16 GB \u00b7 4bit; 4k 157.4 17.6 6.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p13i2hbr",
   "source_url": "https://omlx.ai/benchmarks/performance/p13i2hbr",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-g209ho7m-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 222.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 222.5 21.0 6.5 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g209ho7m",
   "source_url": "https://omlx.ai/benchmarks/performance/g209ho7m",
   "tg_tokens": null,
   "tps": 21.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-g209ho7m-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 222.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 222.7 20.1 6.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g209ho7m",
   "source_url": "https://omlx.ai/benchmarks/performance/g209ho7m",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-fbepb9zp-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 224.1 20.2 6.5 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fbepb9zp",
   "source_url": "https://omlx.ai/benchmarks/performance/fbepb9zp",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-fbepb9zp-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 224.2 19.8 6.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fbepb9zp",
   "source_url": "https://omlx.ai/benchmarks/performance/fbepb9zp",
   "tg_tokens": null,
   "tps": 19.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-ocokv3ff-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 196.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 196.5 20.9 6.5 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ocokv3ff",
   "source_url": "https://omlx.ai/benchmarks/performance/ocokv3ff",
   "tg_tokens": null,
   "tps": 20.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-ocokv3ff-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 196.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 196.4 20.3 6.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ocokv3ff",
   "source_url": "https://omlx.ai/benchmarks/performance/ocokv3ff",
   "tg_tokens": null,
   "tps": 20.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-i6exorv3-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 231.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 231.9 27.2 6.5 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i6exorv3",
   "source_url": "https://omlx.ai/benchmarks/performance/i6exorv3",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-i6exorv3-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 235.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 235.2 25.9 6.8 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i6exorv3",
   "source_url": "https://omlx.ai/benchmarks/performance/i6exorv3",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-41hzqa36-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 231.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 231.9 25.3 6.5 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/41hzqa36",
   "source_url": "https://omlx.ai/benchmarks/performance/41hzqa36",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-41hzqa36-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 232.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 232.5 25.0 6.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/41hzqa36",
   "source_url": "https://omlx.ai/benchmarks/performance/41hzqa36",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hhxc5sho-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: bqy4j59f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 893.6,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 1k 893.6 38.2 13.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hhxc5sho",
   "source_url": "https://omlx.ai/benchmarks/performance/hhxc5sho",
   "tg_tokens": null,
   "tps": 38.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hhxc5sho-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: bqy4j59f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 935.7,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 4k 935.7 38.0 14.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hhxc5sho",
   "source_url": "https://omlx.ai/benchmarks/performance/hhxc5sho",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hhxc5sho-ctx8192",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: bqy4j59f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 852.6,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 8k 852.6 35.3 14.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hhxc5sho",
   "source_url": "https://omlx.ai/benchmarks/performance/hhxc5sho",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hhxc5sho-ctx16384",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 15.3 GB at this context; duplicate traces of the same session: bqy4j59f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 735.4,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 16k 735.4 32.5 15.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hhxc5sho",
   "source_url": "https://omlx.ai/benchmarks/performance/hhxc5sho",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hhxc5sho-ctx32768",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: bqy4j59f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 699.2,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 32k 699.2 32.1 16.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hhxc5sho",
   "source_url": "https://omlx.ai/benchmarks/performance/hhxc5sho",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hhxc5sho-ctx65536",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: bqy4j59f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 680.6,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 64k 680.6 27.9 18.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hhxc5sho",
   "source_url": "https://omlx.ai/benchmarks/performance/hhxc5sho",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hhxc5sho-ctx131072",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: bqy4j59f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 549.4,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 128k 549.4 21.5 22.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hhxc5sho",
   "source_url": "https://omlx.ai/benchmarks/performance/hhxc5sho",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ziaa5bwx-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "pyros-vault MTP build; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: hj41ahwa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 773.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "pyros-vault:Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 773.5 34.2 33.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ziaa5bwx",
   "source_url": "https://omlx.ai/benchmarks/performance/ziaa5bwx",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ziaa5bwx-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "pyros-vault MTP build; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: hj41ahwa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 850.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "pyros-vault:Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 850.0 39.2 34.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ziaa5bwx",
   "source_url": "https://omlx.ai/benchmarks/performance/ziaa5bwx",
   "tg_tokens": null,
   "tps": 39.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ziaa5bwx-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "pyros-vault MTP build; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: hj41ahwa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 823.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "pyros-vault:Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 823.6 50.2 35.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ziaa5bwx",
   "source_url": "https://omlx.ai/benchmarks/performance/ziaa5bwx",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ziaa5bwx-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "pyros-vault MTP build; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: hj41ahwa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 752.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "pyros-vault:Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 752.8 53.2 35.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ziaa5bwx",
   "source_url": "https://omlx.ai/benchmarks/performance/ziaa5bwx",
   "tg_tokens": null,
   "tps": 53.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ziaa5bwx-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "pyros-vault MTP build; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: hj41ahwa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 709.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "pyros-vault:Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 709.2 55.2 36.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ziaa5bwx",
   "source_url": "https://omlx.ai/benchmarks/performance/ziaa5bwx",
   "tg_tokens": null,
   "tps": 55.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ziaa5bwx-ctx65536",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "pyros-vault MTP build; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: hj41ahwa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 617.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "pyros-vault:Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 617.3 41.5 39.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ziaa5bwx",
   "source_url": "https://omlx.ai/benchmarks/performance/ziaa5bwx",
   "tg_tokens": null,
   "tps": 41.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ziaa5bwx-ctx131072",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "pyros-vault MTP build; oMLX community benchmark trace; peak memory 83.2 GB at this context; duplicate traces of the same session: hj41ahwa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 329.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "pyros-vault:Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 329.1 22.7 83.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ziaa5bwx",
   "source_url": "https://omlx.ai/benchmarks/performance/ziaa5bwx",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-z7g3r9xn-ctx1024",
   "model": "Ornith-1.5-35B-A3B-BigBang",
   "notes": "KaedeTai MTP build; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: z1nee43w; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1158.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "KaedeTai--Ornith-1.5-35B-A3B-BigBang-MTP-mlx-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 1k 1,158 98.5 21.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z7g3r9xn",
   "source_url": "https://omlx.ai/benchmarks/performance/z7g3r9xn",
   "tg_tokens": null,
   "tps": 98.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-z7g3r9xn-ctx4096",
   "model": "Ornith-1.5-35B-A3B-BigBang",
   "notes": "KaedeTai MTP build; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: z1nee43w; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1109.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "KaedeTai--Ornith-1.5-35B-A3B-BigBang-MTP-mlx-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 4k 1,109 106.1 21.9 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z7g3r9xn",
   "source_url": "https://omlx.ai/benchmarks/performance/z7g3r9xn",
   "tg_tokens": null,
   "tps": 106.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-nch4qhlv-ctx4096",
   "model": "LFM2.5-8B-A1B",
   "notes": "oMLX community benchmark trace; peak memory 5.3 GB at this context; duplicate traces of the same session: palfu0y6; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3428.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-8B-A1B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 3,428 104.5 5.3 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nch4qhlv",
   "source_url": "https://omlx.ai/benchmarks/performance/nch4qhlv",
   "tg_tokens": null,
   "tps": 104.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-nch4qhlv-ctx16384",
   "model": "LFM2.5-8B-A1B",
   "notes": "oMLX community benchmark trace; peak memory 5.5 GB at this context; duplicate traces of the same session: palfu0y6; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3179.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-8B-A1B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 16k 3,179 91.9 5.5 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nch4qhlv",
   "source_url": "https://omlx.ai/benchmarks/performance/nch4qhlv",
   "tg_tokens": null,
   "tps": 91.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-xewwn0s7-ctx32768",
   "model": "LFM2.5-8B-A1B",
   "notes": "oMLX community benchmark trace; peak memory 5.6 GB at this context; duplicate traces of the same session: wgwpwgon; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2898.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-8B-A1B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 32k 2,898 80.0 5.6 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xewwn0s7",
   "source_url": "https://omlx.ai/benchmarks/performance/xewwn0s7",
   "tg_tokens": null,
   "tps": 80.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-xewwn0s7-ctx131072",
   "model": "LFM2.5-8B-A1B",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: wgwpwgon; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1175.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-8B-A1B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 128k 1,175 19.5 6.7 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xewwn0s7",
   "source_url": "https://omlx.ai/benchmarks/performance/xewwn0s7",
   "tg_tokens": null,
   "tps": 19.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-3l7graoy-ctx1024",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 1.5 GB at this context; duplicate traces of the same session: none; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 756.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 4bit; 1k 756.4 100.8 1.5 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3l7graoy",
   "source_url": "https://omlx.ai/benchmarks/performance/3l7graoy",
   "tg_tokens": null,
   "tps": 100.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-3l7graoy-ctx4096",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 1.6 GB at this context; duplicate traces of the same session: none; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 734.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 4bit; 4k 734.1 91.3 1.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3l7graoy",
   "source_url": "https://omlx.ai/benchmarks/performance/3l7graoy",
   "tg_tokens": null,
   "tps": 91.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-3l7graoy-ctx8192",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 1.6 GB at this context; duplicate traces of the same session: none; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 710.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 4bit; 8k 710.7 82.2 1.6 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3l7graoy",
   "source_url": "https://omlx.ai/benchmarks/performance/3l7graoy",
   "tg_tokens": null,
   "tps": 82.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-3l7graoy-ctx16384",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 2.0 GB at this context; duplicate traces of the same session: none; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 647.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 4bit; 16k 647.6 68.2 2.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3l7graoy",
   "source_url": "https://omlx.ai/benchmarks/performance/3l7graoy",
   "tg_tokens": null,
   "tps": 68.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-3l7graoy-ctx32768",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 2.6 GB at this context; duplicate traces of the same session: none; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 552.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 4bit; 32k 552.2 47.8 2.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3l7graoy",
   "source_url": "https://omlx.ai/benchmarks/performance/3l7graoy",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-3l7graoy-ctx65536",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 4.1 GB at this context; duplicate traces of the same session: none; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 408.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 4bit; 64k 408.6 17.3 4.1 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3l7graoy",
   "source_url": "https://omlx.ai/benchmarks/performance/3l7graoy",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-f00eert8-ctx1024",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 1.8 GB at this context; duplicate traces of the same session: none; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 755.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 8bit; 1k 755.9 62.4 1.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f00eert8",
   "source_url": "https://omlx.ai/benchmarks/performance/f00eert8",
   "tg_tokens": null,
   "tps": 62.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-f00eert8-ctx4096",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 1.9 GB at this context; duplicate traces of the same session: none; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 733.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 8bit; 4k 733.9 58.9 1.9 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f00eert8",
   "source_url": "https://omlx.ai/benchmarks/performance/f00eert8",
   "tg_tokens": null,
   "tps": 58.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-f00eert8-ctx8192",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 2.0 GB at this context; duplicate traces of the same session: none; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 714.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 8bit; 8k 714.0 55.0 2.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f00eert8",
   "source_url": "https://omlx.ai/benchmarks/performance/f00eert8",
   "tg_tokens": null,
   "tps": 55.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-f00eert8-ctx16384",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 2.4 GB at this context; duplicate traces of the same session: none; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 654.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 8bit; 16k 654.9 47.9 2.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f00eert8",
   "source_url": "https://omlx.ai/benchmarks/performance/f00eert8",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-f00eert8-ctx32768",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 3.0 GB at this context; duplicate traces of the same session: none; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 554.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 8bit; 32k 554.7 37.3 3.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f00eert8",
   "source_url": "https://omlx.ai/benchmarks/performance/f00eert8",
   "tg_tokens": null,
   "tps": 37.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-f00eert8-ctx65536",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 4.4 GB at this context; duplicate traces of the same session: none; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 414.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 8bit; 64k 414.9 24.8 4.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f00eert8",
   "source_url": "https://omlx.ai/benchmarks/performance/f00eert8",
   "tg_tokens": null,
   "tps": 24.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-gtt02lr7-ctx1024",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 1.8 GB; TTFT 465 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2200.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 8bit; 1k tokens 2,200 PP tok/s 109.2 TG tok/s 465 TTFT (ms) 1.8 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gtt02lr7",
   "source_url": "https://omlx.ai/benchmarks/performance/gtt02lr7",
   "tg_tokens": null,
   "tps": 109.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-919dis83-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.3 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 638.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 638.7 54.4 66.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/919dis83",
   "source_url": "https://omlx.ai/benchmarks/performance/919dis83",
   "tg_tokens": null,
   "tps": 54.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-919dis83-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.7 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 714.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 714.1 50.9 66.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/919dis83",
   "source_url": "https://omlx.ai/benchmarks/performance/919dis83",
   "tg_tokens": null,
   "tps": 50.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-919dis83-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.0 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 712.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 712.7 47.3 67.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/919dis83",
   "source_url": "https://omlx.ai/benchmarks/performance/919dis83",
   "tg_tokens": null,
   "tps": 47.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-919dis83-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.7 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 670.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 16k 670.7 35.9 67.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/919dis83",
   "source_url": "https://omlx.ai/benchmarks/performance/919dis83",
   "tg_tokens": null,
   "tps": 35.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-919dis83-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 69.1 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 587.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 587.5 24.4 69.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/919dis83",
   "source_url": "https://omlx.ai/benchmarks/performance/919dis83",
   "tg_tokens": null,
   "tps": 24.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-919dis83-ctx65536",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 71.9 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 458.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k 458.6 15.5 71.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/919dis83",
   "source_url": "https://omlx.ai/benchmarks/performance/919dis83",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5aeflnjy-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1717.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 1,717 53.1 95.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5aeflnjy",
   "source_url": "https://omlx.ai/benchmarks/performance/5aeflnjy",
   "tg_tokens": null,
   "tps": 53.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5aeflnjy-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1725.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 1,725 51.2 95.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5aeflnjy",
   "source_url": "https://omlx.ai/benchmarks/performance/5aeflnjy",
   "tg_tokens": null,
   "tps": 51.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5aeflnjy-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1646.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 1,646 46.9 95.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5aeflnjy",
   "source_url": "https://omlx.ai/benchmarks/performance/5aeflnjy",
   "tg_tokens": null,
   "tps": 46.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5aeflnjy-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1476.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 1,476 45.6 95.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5aeflnjy",
   "source_url": "https://omlx.ai/benchmarks/performance/5aeflnjy",
   "tg_tokens": null,
   "tps": 45.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-x3s2lhvm-ctx65536",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 102.4 GB; TTFT 58153 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1127.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k tokens 1,127 PP tok/s 45.9 TG tok/s 58153 TTFT (ms) 102.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x3s2lhvm",
   "source_url": "https://omlx.ai/benchmarks/performance/x3s2lhvm",
   "tg_tokens": null,
   "tps": 45.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-fwtc8lbd-ctx1024",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 180.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 180.4 32.2 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fwtc8lbd",
   "source_url": "https://omlx.ai/benchmarks/performance/fwtc8lbd",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-fwtc8lbd-ctx4096",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 182.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 182.4 30.4 17.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fwtc8lbd",
   "source_url": "https://omlx.ai/benchmarks/performance/fwtc8lbd",
   "tg_tokens": null,
   "tps": 30.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-fwtc8lbd-ctx8192",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 180.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 8k 180.9 28.4 17.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fwtc8lbd",
   "source_url": "https://omlx.ai/benchmarks/performance/fwtc8lbd",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-fwtc8lbd-ctx16384",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 176.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 16k 176.2 25.3 18.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fwtc8lbd",
   "source_url": "https://omlx.ai/benchmarks/performance/fwtc8lbd",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-pfl2vdmm-ctx1024",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 3grd36vj, mhf4fjks; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 125.6 15.2 16.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pfl2vdmm",
   "source_url": "https://omlx.ai/benchmarks/performance/pfl2vdmm",
   "tg_tokens": null,
   "tps": 15.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-pfl2vdmm-ctx4096",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 3grd36vj, mhf4fjks; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 125.9 14.8 17.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pfl2vdmm",
   "source_url": "https://omlx.ai/benchmarks/performance/pfl2vdmm",
   "tg_tokens": null,
   "tps": 14.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-pfl2vdmm-ctx8192",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: 3grd36vj, mhf4fjks; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 124.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 124.3 14.2 17.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pfl2vdmm",
   "source_url": "https://omlx.ai/benchmarks/performance/pfl2vdmm",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-pfl2vdmm-ctx16384",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: 3grd36vj, mhf4fjks; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 121.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 121.0 12.3 18.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pfl2vdmm",
   "source_url": "https://omlx.ai/benchmarks/performance/pfl2vdmm",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-pfl2vdmm-ctx32768",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 3grd36vj, mhf4fjks; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 114.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 114.0 9.7 20.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pfl2vdmm",
   "source_url": "https://omlx.ai/benchmarks/performance/pfl2vdmm",
   "tg_tokens": null,
   "tps": 9.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-mfwckz8w-ctx1024",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: qyady0u2; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 111.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 1k 111.1 21.1 15.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mfwckz8w",
   "source_url": "https://omlx.ai/benchmarks/performance/mfwckz8w",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-mfwckz8w-ctx4096",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: qyady0u2; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 113.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k 113.8 20.1 16.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mfwckz8w",
   "source_url": "https://omlx.ai/benchmarks/performance/mfwckz8w",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-mfwckz8w-ctx8192",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: qyady0u2; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 113.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 8k 113.1 19.0 16.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mfwckz8w",
   "source_url": "https://omlx.ai/benchmarks/performance/mfwckz8w",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-mfwckz8w-ctx16384",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 17.5 GB at this context; duplicate traces of the same session: qyady0u2; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 111.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 16k 111.0 17.3 17.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mfwckz8w",
   "source_url": "https://omlx.ai/benchmarks/performance/mfwckz8w",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-mfwckz8w-ctx32768",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: qyady0u2; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 32k 106.7 14.9 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mfwckz8w",
   "source_url": "https://omlx.ai/benchmarks/performance/mfwckz8w",
   "tg_tokens": null,
   "tps": 14.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-1y6f1t7a-ctx1024",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 36.2 GB at this context; duplicate traces of the same session: x9i9s0nc; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 519.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-35B-A3B on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 1k 519.9 48.3 36.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1y6f1t7a",
   "source_url": "https://omlx.ai/benchmarks/performance/1y6f1t7a",
   "tg_tokens": null,
   "tps": 48.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-1y6f1t7a-ctx4096",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 36.3 GB at this context; duplicate traces of the same session: x9i9s0nc; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 574.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-35B-A3B on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k 574.8 46.9 36.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1y6f1t7a",
   "source_url": "https://omlx.ai/benchmarks/performance/1y6f1t7a",
   "tg_tokens": null,
   "tps": 46.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ptzhx9ln-ctx1024",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: femieeij; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1652.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,652 93.4 20.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ptzhx9ln",
   "source_url": "https://omlx.ai/benchmarks/performance/ptzhx9ln",
   "tg_tokens": null,
   "tps": 93.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ptzhx9ln-ctx4096",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: femieeij; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1706.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,706 88.6 20.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ptzhx9ln",
   "source_url": "https://omlx.ai/benchmarks/performance/ptzhx9ln",
   "tg_tokens": null,
   "tps": 88.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ptzhx9ln-ctx8192",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: femieeij; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1487.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,487 90.0 20.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ptzhx9ln",
   "source_url": "https://omlx.ai/benchmarks/performance/ptzhx9ln",
   "tg_tokens": null,
   "tps": 90.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ptzhx9ln-ctx16384",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: femieeij; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1227.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,227 63.7 20.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ptzhx9ln",
   "source_url": "https://omlx.ai/benchmarks/performance/ptzhx9ln",
   "tg_tokens": null,
   "tps": 63.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ptzhx9ln-ctx32768",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: femieeij; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1026.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 1,026 42.5 21.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ptzhx9ln",
   "source_url": "https://omlx.ai/benchmarks/performance/ptzhx9ln",
   "tg_tokens": null,
   "tps": 42.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-9bgkvghg-ctx1024",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: cq1j63z1, fxnkfdkj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 867.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 867.2 82.8 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9bgkvghg",
   "source_url": "https://omlx.ai/benchmarks/performance/9bgkvghg",
   "tg_tokens": null,
   "tps": 82.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-9bgkvghg-ctx4096",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: cq1j63z1, fxnkfdkj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 899.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 899.9 79.2 20.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9bgkvghg",
   "source_url": "https://omlx.ai/benchmarks/performance/9bgkvghg",
   "tg_tokens": null,
   "tps": 79.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-9bgkvghg-ctx8192",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: cq1j63z1, fxnkfdkj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 876.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 876.2 74.9 20.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9bgkvghg",
   "source_url": "https://omlx.ai/benchmarks/performance/9bgkvghg",
   "tg_tokens": null,
   "tps": 74.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-9bgkvghg-ctx16384",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: cq1j63z1, fxnkfdkj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 805.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 805.6 57.7 20.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9bgkvghg",
   "source_url": "https://omlx.ai/benchmarks/performance/9bgkvghg",
   "tg_tokens": null,
   "tps": 57.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-9bgkvghg-ctx32768",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: cq1j63z1, fxnkfdkj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 693.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 693.2 40.6 21.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9bgkvghg",
   "source_url": "https://omlx.ai/benchmarks/performance/9bgkvghg",
   "tg_tokens": null,
   "tps": 40.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-9bgkvghg-ctx65536",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: cq1j63z1, fxnkfdkj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 537.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 537.6 25.9 23.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9bgkvghg",
   "source_url": "https://omlx.ai/benchmarks/performance/9bgkvghg",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-7ybrofc4-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: oia8i4m4; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 145.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 4bit; 1k 145.3 22.2 3.9 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ybrofc4",
   "source_url": "https://omlx.ai/benchmarks/performance/7ybrofc4",
   "tg_tokens": null,
   "tps": 22.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-7ybrofc4-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: oia8i4m4; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 144.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 4bit; 4k 144.5 21.0 4.2 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ybrofc4",
   "source_url": "https://omlx.ai/benchmarks/performance/7ybrofc4",
   "tg_tokens": null,
   "tps": 21.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-9x15iu0r-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.0 GB at this context; duplicate traces of the same session: v6gs4f1n; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 253.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 1k 253.0 47.5 4.0 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9x15iu0r",
   "source_url": "https://omlx.ai/benchmarks/performance/9x15iu0r",
   "tg_tokens": null,
   "tps": 47.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-9x15iu0r-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: v6gs4f1n; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 255.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 4k 255.0 46.5 4.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9x15iu0r",
   "source_url": "https://omlx.ai/benchmarks/performance/9x15iu0r",
   "tg_tokens": null,
   "tps": 46.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-lsp40v2a-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.0 GB at this context; duplicate traces of the same session: 0ysklzxp, aeq42ap1; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 736.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 1k 736.4 102.2 4.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lsp40v2a",
   "source_url": "https://omlx.ai/benchmarks/performance/lsp40v2a",
   "tg_tokens": null,
   "tps": 102.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-lsp40v2a-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: 0ysklzxp, aeq42ap1; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 734.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k 734.5 95.4 4.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lsp40v2a",
   "source_url": "https://omlx.ai/benchmarks/performance/lsp40v2a",
   "tg_tokens": null,
   "tps": 95.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-lsp40v2a-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.6 GB at this context; duplicate traces of the same session: 0ysklzxp, aeq42ap1; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 720.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 8k 720.3 87.0 4.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lsp40v2a",
   "source_url": "https://omlx.ai/benchmarks/performance/lsp40v2a",
   "tg_tokens": null,
   "tps": 87.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-lsp40v2a-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 5.3 GB at this context; duplicate traces of the same session: 0ysklzxp, aeq42ap1; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 691.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 16k 691.3 73.4 5.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lsp40v2a",
   "source_url": "https://omlx.ai/benchmarks/performance/lsp40v2a",
   "tg_tokens": null,
   "tps": 73.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-lsp40v2a-ctx32768",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: 0ysklzxp, aeq42ap1; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 576.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 32k 576.0 28.6 6.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lsp40v2a",
   "source_url": "https://omlx.ai/benchmarks/performance/lsp40v2a",
   "tg_tokens": null,
   "tps": 28.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-lsp40v2a-ctx65536",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 8.6 GB at this context; duplicate traces of the same session: 0ysklzxp, aeq42ap1; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 468.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 64k 468.7 18.2 8.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lsp40v2a",
   "source_url": "https://omlx.ai/benchmarks/performance/lsp40v2a",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-0oga1cf0-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.0 GB at this context; duplicate traces of the same session: j290pxfi; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1428.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 1k 1,428 137.5 4.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0oga1cf0",
   "source_url": "https://omlx.ai/benchmarks/performance/0oga1cf0",
   "tg_tokens": null,
   "tps": 137.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-0oga1cf0-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: j290pxfi; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1440.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 4k 1,440 128.2 4.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0oga1cf0",
   "source_url": "https://omlx.ai/benchmarks/performance/0oga1cf0",
   "tg_tokens": null,
   "tps": 128.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-0oga1cf0-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.6 GB at this context; duplicate traces of the same session: j290pxfi; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1414.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 8k 1,414 117.4 4.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0oga1cf0",
   "source_url": "https://omlx.ai/benchmarks/performance/0oga1cf0",
   "tg_tokens": null,
   "tps": 117.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-0oga1cf0-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 5.3 GB at this context; duplicate traces of the same session: j290pxfi; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1355.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 16k 1,355 98.7 5.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0oga1cf0",
   "source_url": "https://omlx.ai/benchmarks/performance/0oga1cf0",
   "tg_tokens": null,
   "tps": 98.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-0oga1cf0-ctx32768",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: j290pxfi; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1247.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 32k 1,247 74.5 6.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0oga1cf0",
   "source_url": "https://omlx.ai/benchmarks/performance/0oga1cf0",
   "tg_tokens": null,
   "tps": 74.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-0oga1cf0-ctx65536",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 8.6 GB at this context; duplicate traces of the same session: j290pxfi; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1051.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 64k 1,051 50.4 8.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0oga1cf0",
   "source_url": "https://omlx.ai/benchmarks/performance/0oga1cf0",
   "tg_tokens": null,
   "tps": 50.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-5t6c2eu1-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: cgvsvi3i; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1423.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 1k 1,423 99.9 5.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5t6c2eu1",
   "source_url": "https://omlx.ai/benchmarks/performance/5t6c2eu1",
   "tg_tokens": null,
   "tps": 99.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-5t6c2eu1-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 6.1 GB at this context; duplicate traces of the same session: cgvsvi3i; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1448.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 4k 1,448 96.2 6.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5t6c2eu1",
   "source_url": "https://omlx.ai/benchmarks/performance/5t6c2eu1",
   "tg_tokens": null,
   "tps": 96.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-5t6c2eu1-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: cgvsvi3i; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1421.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 8k 1,421 90.1 6.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5t6c2eu1",
   "source_url": "https://omlx.ai/benchmarks/performance/5t6c2eu1",
   "tg_tokens": null,
   "tps": 90.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-5t6c2eu1-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 7.3 GB at this context; duplicate traces of the same session: cgvsvi3i; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1362.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 16k 1,362 80.1 7.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5t6c2eu1",
   "source_url": "https://omlx.ai/benchmarks/performance/5t6c2eu1",
   "tg_tokens": null,
   "tps": 80.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-5t6c2eu1-ctx32768",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 7.9 GB at this context; duplicate traces of the same session: cgvsvi3i; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1255.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 32k 1,255 63.0 7.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5t6c2eu1",
   "source_url": "https://omlx.ai/benchmarks/performance/5t6c2eu1",
   "tg_tokens": null,
   "tps": 63.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-5t6c2eu1-ctx65536",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 9.9 GB at this context; duplicate traces of the same session: cgvsvi3i; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1056.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 64k 1,056 45.5 9.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5t6c2eu1",
   "source_url": "https://omlx.ai/benchmarks/performance/5t6c2eu1",
   "tg_tokens": null,
   "tps": 45.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-id732h0q-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 334.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M3 (10c); M3 (10c) \u00b7 16 GB \u00b7 4bit; 1k 334.5 30.8 3.9 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/id732h0q",
   "source_url": "https://omlx.ai/benchmarks/performance/id732h0q",
   "tg_tokens": null,
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-id732h0q-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 296.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M3 (10c); M3 (10c) \u00b7 16 GB \u00b7 4bit; 4k 296.7 29.5 4.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/id732h0q",
   "source_url": "https://omlx.ai/benchmarks/performance/id732h0q",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-tirqyj3o-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: 91h2g4t3, s11jretg, tir4w3vr; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 374.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 374.9 33.7 3.9 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tirqyj3o",
   "source_url": "https://omlx.ai/benchmarks/performance/tirqyj3o",
   "tg_tokens": null,
   "tps": 33.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-tirqyj3o-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: 91h2g4t3, s11jretg, tir4w3vr; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 373.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 373.3 32.0 4.2 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tirqyj3o",
   "source_url": "https://omlx.ai/benchmarks/performance/tirqyj3o",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-tirqyj3o-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.6 GB at this context; duplicate traces of the same session: 91h2g4t3, s11jretg, tir4w3vr; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 364.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 8k 364.6 29.5 4.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tirqyj3o",
   "source_url": "https://omlx.ai/benchmarks/performance/tirqyj3o",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-tirqyj3o-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.8 GB at this context; duplicate traces of the same session: 91h2g4t3, s11jretg, tir4w3vr; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 340.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 16k 340.6 25.2 4.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tirqyj3o",
   "source_url": "https://omlx.ai/benchmarks/performance/tirqyj3o",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-tirqyj3o-ctx32768",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: 91h2g4t3, s11jretg, tir4w3vr; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 299.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 32k 299.7 20.3 5.9 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tirqyj3o",
   "source_url": "https://omlx.ai/benchmarks/performance/tirqyj3o",
   "tg_tokens": null,
   "tps": 20.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-tirqyj3o-ctx65536",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 8.0 GB at this context; duplicate traces of the same session: 91h2g4t3, s11jretg, tir4w3vr; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 231.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 64k 231.3 13.6 8.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tirqyj3o",
   "source_url": "https://omlx.ai/benchmarks/performance/tirqyj3o",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-rkslpdmh-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 406.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 406.1 37.4 3.9 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rkslpdmh",
   "source_url": "https://omlx.ai/benchmarks/performance/rkslpdmh",
   "tg_tokens": null,
   "tps": 37.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-rkslpdmh-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 403.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 403.8 36.0 4.2 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rkslpdmh",
   "source_url": "https://omlx.ai/benchmarks/performance/rkslpdmh",
   "tg_tokens": null,
   "tps": 36.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-rkslpdmh-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.6 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 385.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 8k 385.2 31.8 4.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rkslpdmh",
   "source_url": "https://omlx.ai/benchmarks/performance/rkslpdmh",
   "tg_tokens": null,
   "tps": 31.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-v05wvgce-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 3.9 GB; TTFT 6457 ms; duplicate traces of the same session: none; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 158.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (8c); M4 (8c) \u00b7 16 GB \u00b7 4bit; 1k tokens 158.6 PP tok/s 19.9 TG tok/s 6457 TTFT (ms) 3.9 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v05wvgce",
   "source_url": "https://omlx.ai/benchmarks/performance/v05wvgce",
   "tg_tokens": null,
   "tps": 19.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-j3udbnrh-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.0 GB at this context; duplicate traces of the same session: nvxbhp1x, qsyh8wbm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 811.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 811.1 82.1 4.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j3udbnrh",
   "source_url": "https://omlx.ai/benchmarks/performance/j3udbnrh",
   "tg_tokens": null,
   "tps": 82.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-j3udbnrh-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: nvxbhp1x, qsyh8wbm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 798.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 798.0 76.8 4.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j3udbnrh",
   "source_url": "https://omlx.ai/benchmarks/performance/j3udbnrh",
   "tg_tokens": null,
   "tps": 76.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-j3udbnrh-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.6 GB at this context; duplicate traces of the same session: nvxbhp1x, qsyh8wbm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 781.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 781.9 72.4 4.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j3udbnrh",
   "source_url": "https://omlx.ai/benchmarks/performance/j3udbnrh",
   "tg_tokens": null,
   "tps": 72.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-j3udbnrh-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 5.3 GB at this context; duplicate traces of the same session: nvxbhp1x, qsyh8wbm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 737.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 737.3 62.8 5.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j3udbnrh",
   "source_url": "https://omlx.ai/benchmarks/performance/j3udbnrh",
   "tg_tokens": null,
   "tps": 62.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-j3udbnrh-ctx32768",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: nvxbhp1x, qsyh8wbm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 655.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 655.0 49.9 6.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j3udbnrh",
   "source_url": "https://omlx.ai/benchmarks/performance/j3udbnrh",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-j3udbnrh-ctx65536",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 8.6 GB at this context; duplicate traces of the same session: nvxbhp1x, qsyh8wbm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 534.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 534.5 36.3 8.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j3udbnrh",
   "source_url": "https://omlx.ai/benchmarks/performance/j3udbnrh",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-7vz8r9rx-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: znelvuh1; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 77.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 4bit; 1k 77.8 12.6 6.5 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7vz8r9rx",
   "source_url": "https://omlx.ai/benchmarks/performance/7vz8r9rx",
   "tg_tokens": null,
   "tps": 12.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-7vz8r9rx-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: znelvuh1; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 77.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 4bit; 4k 77.5 12.1 6.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7vz8r9rx",
   "source_url": "https://omlx.ai/benchmarks/performance/7vz8r9rx",
   "tg_tokens": null,
   "tps": 12.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-1n8qgmq9-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.7 GB at this context; duplicate traces of the same session: 6rngmu2d; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 239.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 8bit; 1k 239.3 35.1 10.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1n8qgmq9",
   "source_url": "https://omlx.ai/benchmarks/performance/1n8qgmq9",
   "tg_tokens": null,
   "tps": 35.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-1n8qgmq9-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.0 GB at this context; duplicate traces of the same session: 6rngmu2d; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 242.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 8bit; 4k 242.1 33.3 11.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1n8qgmq9",
   "source_url": "https://omlx.ai/benchmarks/performance/1n8qgmq9",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-1n8qgmq9-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.2 GB at this context; duplicate traces of the same session: 6rngmu2d; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 236.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 8bit; 8k 236.2 30.3 11.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1n8qgmq9",
   "source_url": "https://omlx.ai/benchmarks/performance/1n8qgmq9",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-1n8qgmq9-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.7 GB at this context; duplicate traces of the same session: 6rngmu2d; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 234.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 8bit; 16k 234.4 26.2 11.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1n8qgmq9",
   "source_url": "https://omlx.ai/benchmarks/performance/1n8qgmq9",
   "tg_tokens": null,
   "tps": 26.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-1n8qgmq9-ctx32768",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: 6rngmu2d; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 8bit; 32k 224.8 20.3 12.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1n8qgmq9",
   "source_url": "https://omlx.ai/benchmarks/performance/1n8qgmq9",
   "tg_tokens": null,
   "tps": 20.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-1n8qgmq9-ctx65536",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: 6rngmu2d; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 8bit; 64k 205.0 13.9 14.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1n8qgmq9",
   "source_url": "https://omlx.ai/benchmarks/performance/1n8qgmq9",
   "tg_tokens": null,
   "tps": 13.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-s5zje522-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: k9e5j7ms; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 138.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 1k 138.1 30.0 6.7 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s5zje522",
   "source_url": "https://omlx.ai/benchmarks/performance/s5zje522",
   "tg_tokens": null,
   "tps": 30.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-s5zje522-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: k9e5j7ms; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 138.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 4k 138.6 28.6 7.0 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s5zje522",
   "source_url": "https://omlx.ai/benchmarks/performance/s5zje522",
   "tg_tokens": null,
   "tps": 28.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-xogtwhgu-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.7 GB at this context; duplicate traces of the same session: 1h65skb8, 5qvzekfa, 6m1wd63n, 8w117y2g; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 8bit; 1k 127.6 12.3 10.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xogtwhgu",
   "source_url": "https://omlx.ai/benchmarks/performance/xogtwhgu",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-xogtwhgu-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.0 GB at this context; duplicate traces of the same session: 1h65skb8, 5qvzekfa, 6m1wd63n, 8w117y2g; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 132.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 8bit; 4k 132.0 12.0 11.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xogtwhgu",
   "source_url": "https://omlx.ai/benchmarks/performance/xogtwhgu",
   "tg_tokens": null,
   "tps": 12.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-xogtwhgu-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.2 GB at this context; duplicate traces of the same session: 1h65skb8, 5qvzekfa, 6m1wd63n, 8w117y2g; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 8bit; 8k 130.7 11.5 11.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xogtwhgu",
   "source_url": "https://omlx.ai/benchmarks/performance/xogtwhgu",
   "tg_tokens": null,
   "tps": 11.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-xogtwhgu-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.7 GB at this context; duplicate traces of the same session: 1h65skb8, 5qvzekfa, 6m1wd63n, 8w117y2g; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 8bit; 16k 127.5 10.6 11.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xogtwhgu",
   "source_url": "https://omlx.ai/benchmarks/performance/xogtwhgu",
   "tg_tokens": null,
   "tps": 10.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-xogtwhgu-ctx32768",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: 1h65skb8, 5qvzekfa, 6m1wd63n, 8w117y2g; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 90.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 8bit; 32k 90.4 8.6 12.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xogtwhgu",
   "source_url": "https://omlx.ai/benchmarks/performance/xogtwhgu",
   "tg_tokens": null,
   "tps": 8.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-wxd1vexu-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.7 GB at this context; duplicate traces of the same session: 6o6zm8tl; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 794.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 1k 794.0 64.3 10.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wxd1vexu",
   "source_url": "https://omlx.ai/benchmarks/performance/wxd1vexu",
   "tg_tokens": null,
   "tps": 64.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-wxd1vexu-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.0 GB at this context; duplicate traces of the same session: 6o6zm8tl; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 804.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 4k 804.0 62.6 11.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wxd1vexu",
   "source_url": "https://omlx.ai/benchmarks/performance/wxd1vexu",
   "tg_tokens": null,
   "tps": 62.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-wxd1vexu-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.2 GB at this context; duplicate traces of the same session: 6o6zm8tl; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 797.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 8k 797.4 59.7 11.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wxd1vexu",
   "source_url": "https://omlx.ai/benchmarks/performance/wxd1vexu",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-wxd1vexu-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.7 GB at this context; duplicate traces of the same session: 6o6zm8tl; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 778.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 16k 778.7 54.0 11.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wxd1vexu",
   "source_url": "https://omlx.ai/benchmarks/performance/wxd1vexu",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-wxd1vexu-ctx32768",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: 6o6zm8tl; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 742.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 32k 742.0 46.4 12.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wxd1vexu",
   "source_url": "https://omlx.ai/benchmarks/performance/wxd1vexu",
   "tg_tokens": null,
   "tps": 46.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-wxd1vexu-ctx65536",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: 6o6zm8tl; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 668.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 64k 668.6 33.9 14.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wxd1vexu",
   "source_url": "https://omlx.ai/benchmarks/performance/wxd1vexu",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-zd9pv0ve-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: mk5718cc, r7jnwmp2, rr8vgmtb; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 727.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 1k 727.6 96.7 6.7 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zd9pv0ve",
   "source_url": "https://omlx.ai/benchmarks/performance/zd9pv0ve",
   "tg_tokens": null,
   "tps": 96.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-zd9pv0ve-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: mk5718cc, r7jnwmp2, rr8vgmtb; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 789.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 4k 789.6 93.2 7.0 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zd9pv0ve",
   "source_url": "https://omlx.ai/benchmarks/performance/zd9pv0ve",
   "tg_tokens": null,
   "tps": 93.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-zd9pv0ve-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.1 GB at this context; duplicate traces of the same session: mk5718cc, r7jnwmp2, rr8vgmtb; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 792.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 8k 792.6 87.6 7.1 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zd9pv0ve",
   "source_url": "https://omlx.ai/benchmarks/performance/zd9pv0ve",
   "tg_tokens": null,
   "tps": 87.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-zd9pv0ve-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.6 GB at this context; duplicate traces of the same session: mk5718cc, r7jnwmp2, rr8vgmtb; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 774.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 16k 774.4 76.6 7.6 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zd9pv0ve",
   "source_url": "https://omlx.ai/benchmarks/performance/zd9pv0ve",
   "tg_tokens": null,
   "tps": 76.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-zd9pv0ve-ctx32768",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.7 GB at this context; duplicate traces of the same session: mk5718cc, r7jnwmp2, rr8vgmtb; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 738.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 32k 738.1 61.9 8.7 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zd9pv0ve",
   "source_url": "https://omlx.ai/benchmarks/performance/zd9pv0ve",
   "tg_tokens": null,
   "tps": 61.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-zd9pv0ve-ctx65536",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.8 GB at this context; duplicate traces of the same session: mk5718cc, r7jnwmp2, rr8vgmtb; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 665.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 64k 665.7 44.3 10.8 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zd9pv0ve",
   "source_url": "https://omlx.ai/benchmarks/performance/zd9pv0ve",
   "tg_tokens": null,
   "tps": 44.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-ub1ea8hz-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: crssj6pi; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 155.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 155.3 19.0 6.5 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ub1ea8hz",
   "source_url": "https://omlx.ai/benchmarks/performance/ub1ea8hz",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-ub1ea8hz-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: crssj6pi; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 155.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 155.6 17.8 6.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ub1ea8hz",
   "source_url": "https://omlx.ai/benchmarks/performance/ub1ea8hz",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-x32s7apl-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: rwmboxzn; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 726.4,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 4k 726.4 53.7 29.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x32s7apl",
   "source_url": "https://omlx.ai/benchmarks/performance/x32s7apl",
   "tg_tokens": null,
   "tps": 53.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-x32s7apl-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: rwmboxzn; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 455.5,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 32k 455.5 44.9 32.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x32s7apl",
   "source_url": "https://omlx.ai/benchmarks/performance/x32s7apl",
   "tg_tokens": null,
   "tps": 44.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-h8ei32x4-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 453.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 453.3 33.8 19.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h8ei32x4",
   "source_url": "https://omlx.ai/benchmarks/performance/h8ei32x4",
   "tg_tokens": null,
   "tps": 33.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-h8ei32x4-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 430.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 430.4 31.8 21.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h8ei32x4",
   "source_url": "https://omlx.ai/benchmarks/performance/h8ei32x4",
   "tg_tokens": null,
   "tps": 31.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-3gemr7oz-ctx1024",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "MPTLS optimized speed build; oMLX community benchmark trace, single run at 1k context; peak memory 20.1 GB; TTFT 2221 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 461.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Abliterated-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k tokens 461.1 PP tok/s 14.7 TG tok/s 2221 TTFT (ms) 20.1 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3gemr7oz",
   "source_url": "https://omlx.ai/benchmarks/performance/3gemr7oz",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bqy4j59f-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 893.6,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 1k 893.6 38.2 13.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bqy4j59f",
   "source_url": "https://omlx.ai/benchmarks/performance/bqy4j59f",
   "tg_tokens": null,
   "tps": 38.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bqy4j59f-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 935.7,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 4k 935.7 38.0 14.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bqy4j59f",
   "source_url": "https://omlx.ai/benchmarks/performance/bqy4j59f",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bqy4j59f-ctx8192",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 852.6,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 8k 852.6 35.3 14.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bqy4j59f",
   "source_url": "https://omlx.ai/benchmarks/performance/bqy4j59f",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bqy4j59f-ctx16384",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 15.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 735.4,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 16k 735.4 32.5 15.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bqy4j59f",
   "source_url": "https://omlx.ai/benchmarks/performance/bqy4j59f",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bqy4j59f-ctx32768",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 699.2,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 32k 699.2 32.1 16.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bqy4j59f",
   "source_url": "https://omlx.ai/benchmarks/performance/bqy4j59f",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bqy4j59f-ctx65536",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 680.6,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 64k 680.6 27.9 18.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bqy4j59f",
   "source_url": "https://omlx.ai/benchmarks/performance/bqy4j59f",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bqy4j59f-ctx131072",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 549.4,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 128k 549.4 21.5 22.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bqy4j59f",
   "source_url": "https://omlx.ai/benchmarks/performance/bqy4j59f",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-k5vs731p-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 228.6,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 1k 228.6 29.0 16.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k5vs731p",
   "source_url": "https://omlx.ai/benchmarks/performance/k5vs731p",
   "tg_tokens": null,
   "tps": 29.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-k5vs731p-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 245.4,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 4k 245.4 28.3 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k5vs731p",
   "source_url": "https://omlx.ai/benchmarks/performance/k5vs731p",
   "tg_tokens": null,
   "tps": 28.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-k5vs731p-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 239.8,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 8k 239.8 27.7 20.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k5vs731p",
   "source_url": "https://omlx.ai/benchmarks/performance/k5vs731p",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-k5vs731p-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 229.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 16k 229.0 26.0 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k5vs731p",
   "source_url": "https://omlx.ai/benchmarks/performance/k5vs731p",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-d7wb68vd-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 412.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 412.3 61.1 24.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d7wb68vd",
   "source_url": "https://omlx.ai/benchmarks/performance/d7wb68vd",
   "tg_tokens": null,
   "tps": 61.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-d7wb68vd-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 396.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 396.4 63.3 27.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d7wb68vd",
   "source_url": "https://omlx.ai/benchmarks/performance/d7wb68vd",
   "tg_tokens": null,
   "tps": 63.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-t3bj3g02-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: n64tq8ac; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 910.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 910.6 32.0 17.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t3bj3g02",
   "source_url": "https://omlx.ai/benchmarks/performance/t3bj3g02",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-t3bj3g02-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: n64tq8ac; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 931.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 931.4 35.8 18.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t3bj3g02",
   "source_url": "https://omlx.ai/benchmarks/performance/t3bj3g02",
   "tg_tokens": null,
   "tps": 35.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-9gjj73p1-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: kn01zcvj, z1tgsltk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 216.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 216.8 32.8 33.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9gjj73p1",
   "source_url": "https://omlx.ai/benchmarks/performance/9gjj73p1",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-9gjj73p1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: kn01zcvj, z1tgsltk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 218.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 218.1 27.9 37.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9gjj73p1",
   "source_url": "https://omlx.ai/benchmarks/performance/9gjj73p1",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-9gjj73p1-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 40.6 GB at this context; duplicate traces of the same session: kn01zcvj, z1tgsltk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 186.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 186.2 32.0 40.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9gjj73p1",
   "source_url": "https://omlx.ai/benchmarks/performance/9gjj73p1",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-h48v8nf7-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 135.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 135.0 23.7 33.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h48v8nf7",
   "source_url": "https://omlx.ai/benchmarks/performance/h48v8nf7",
   "tg_tokens": null,
   "tps": 23.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-h48v8nf7-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 221.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 221.6 26.6 37.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h48v8nf7",
   "source_url": "https://omlx.ai/benchmarks/performance/h48v8nf7",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-w9neqy0o-ctx1024",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: 5iyg0ao2, 60q4e4fn, tggxugfz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 842.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ6e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 842.1 88.0 29.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w9neqy0o",
   "source_url": "https://omlx.ai/benchmarks/performance/w9neqy0o",
   "tg_tokens": null,
   "tps": 88.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-w9neqy0o-ctx4096",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.9 GB at this context; duplicate traces of the same session: 5iyg0ao2, 60q4e4fn, tggxugfz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 989.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ6e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 989.3 70.8 31.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w9neqy0o",
   "source_url": "https://omlx.ai/benchmarks/performance/w9neqy0o",
   "tg_tokens": null,
   "tps": 70.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-w9neqy0o-ctx8192",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.2 GB at this context; duplicate traces of the same session: 5iyg0ao2, 60q4e4fn, tggxugfz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 866.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ6e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 8k 866.7 94.2 32.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w9neqy0o",
   "source_url": "https://omlx.ai/benchmarks/performance/w9neqy0o",
   "tg_tokens": null,
   "tps": 94.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-w9neqy0o-ctx16384",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.7 GB at this context; duplicate traces of the same session: 5iyg0ao2, 60q4e4fn, tggxugfz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 711.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ6e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 16k 711.5 81.4 32.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w9neqy0o",
   "source_url": "https://omlx.ai/benchmarks/performance/w9neqy0o",
   "tg_tokens": null,
   "tps": 81.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-lq322bkn-ctx16384",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "MTP + PLE; oMLX community benchmark trace, single run at 16k context; peak memory 43.1 GB; TTFT 62066 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 264.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-REAP-288-oQ4e-BF16-MTP-PLE on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k tokens 264.0 PP tok/s 24.4 TG tok/s 62066 TTFT (ms) 43.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lq322bkn",
   "source_url": "https://omlx.ai/benchmarks/performance/lq322bkn",
   "tg_tokens": null,
   "tps": 24.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-gsyyvmr2-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "MTP + PLE; oMLX community benchmark trace, single run at 4k context; peak memory 43.4 GB; TTFT 17005 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 240.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-REAP-288-oQ4e-BF16-MTP-PLE on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k tokens 240.9 PP tok/s 27.9 TG tok/s 17005 TTFT (ms) 43.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gsyyvmr2",
   "source_url": "https://omlx.ai/benchmarks/performance/gsyyvmr2",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3elp66wz-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 195k context; peak memory 96.7 GB; TTFT 98664 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2027.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 195k tokens 2,027 PP tok/s 45.9 TG tok/s 98664 TTFT (ms) 96.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3elp66wz",
   "source_url": "https://omlx.ai/benchmarks/performance/3elp66wz",
   "tg_tokens": null,
   "tps": 45.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-r74fhc0f-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: bt839ka2; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1042.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 1,042 50.4 14.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r74fhc0f",
   "source_url": "https://omlx.ai/benchmarks/performance/r74fhc0f",
   "tg_tokens": null,
   "tps": 50.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-r74fhc0f-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.5 GB at this context; duplicate traces of the same session: bt839ka2; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1318.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,318 49.0 14.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r74fhc0f",
   "source_url": "https://omlx.ai/benchmarks/performance/r74fhc0f",
   "tg_tokens": null,
   "tps": 49.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hv9wsqs7-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: b2s9t2j3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3631.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 3,631 112.5 14.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hv9wsqs7",
   "source_url": "https://omlx.ai/benchmarks/performance/hv9wsqs7",
   "tg_tokens": null,
   "tps": 112.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hv9wsqs7-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: b2s9t2j3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4033.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 4,033 113.6 14.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hv9wsqs7",
   "source_url": "https://omlx.ai/benchmarks/performance/hv9wsqs7",
   "tg_tokens": null,
   "tps": 113.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-8tftpk5s-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 5dh8u8jj, o2207fu5; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 561.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 1k 561.8 65.9 19.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8tftpk5s",
   "source_url": "https://omlx.ai/benchmarks/performance/8tftpk5s",
   "tg_tokens": null,
   "tps": 65.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-8tftpk5s-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 5dh8u8jj, o2207fu5; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 565.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k 565.2 66.2 20.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8tftpk5s",
   "source_url": "https://omlx.ai/benchmarks/performance/8tftpk5s",
   "tg_tokens": null,
   "tps": 66.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-8tftpk5s-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: 5dh8u8jj, o2207fu5; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 442.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 16k 442.6 54.5 20.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8tftpk5s",
   "source_url": "https://omlx.ai/benchmarks/performance/8tftpk5s",
   "tg_tokens": null,
   "tps": 54.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-6q5mg9s0-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "scottlowry MTP build; oMLX community benchmark trace; peak memory 30.0 GB at this context; duplicate traces of the same session: wysrquee, zq5bhoyg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 702.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "models--scottlowry--Ornith-1.5-35B-A3B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 702.9 74.2 30.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6q5mg9s0",
   "source_url": "https://omlx.ai/benchmarks/performance/6q5mg9s0",
   "tg_tokens": null,
   "tps": 74.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-6q5mg9s0-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "scottlowry MTP build; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: wysrquee, zq5bhoyg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 788.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "models--scottlowry--Ornith-1.5-35B-A3B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 788.3 69.2 32.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6q5mg9s0",
   "source_url": "https://omlx.ai/benchmarks/performance/6q5mg9s0",
   "tg_tokens": null,
   "tps": 69.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-6q5mg9s0-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "scottlowry MTP build; oMLX community benchmark trace; peak memory 32.3 GB at this context; duplicate traces of the same session: wysrquee, zq5bhoyg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 753.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "models--scottlowry--Ornith-1.5-35B-A3B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 6bit; 8k 753.7 56.7 32.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6q5mg9s0",
   "source_url": "https://omlx.ai/benchmarks/performance/6q5mg9s0",
   "tg_tokens": null,
   "tps": 56.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-a4xo9hjs-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "scottlowry MTP build; oMLX community benchmark trace, single run at 1k context; peak memory 30.0 GB; TTFT 1444 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 709.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "models--scottlowry--Ornith-1.5-35B-A3B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k tokens 709.4 PP tok/s 64.1 TG tok/s 1444 TTFT (ms) 30.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a4xo9hjs",
   "source_url": "https://omlx.ai/benchmarks/performance/a4xo9hjs",
   "tg_tokens": null,
   "tps": 64.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-eg4cmpft-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "scottlowry MTP build; oMLX community benchmark trace, single run at 1k context; peak memory 30.3 GB; TTFT 1702 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 601.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "models--scottlowry--Ornith-1.5-35B-A3B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k tokens 601.6 PP tok/s 44.8 TG tok/s 1702 TTFT (ms) 30.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eg4cmpft",
   "source_url": "https://omlx.ai/benchmarks/performance/eg4cmpft",
   "tg_tokens": null,
   "tps": 44.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-a4hq2cod-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "scottlowry MTP build; oMLX community benchmark trace, single run at 1k context; peak memory 30.3 GB; TTFT 1470 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 696.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "models--scottlowry--Ornith-1.5-35B-A3B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k tokens 696.8 PP tok/s 60.8 TG tok/s 1470 TTFT (ms) 30.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a4hq2cod",
   "source_url": "https://omlx.ai/benchmarks/performance/a4hq2cod",
   "tg_tokens": null,
   "tps": 60.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hj41ahwa-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "pyros-vault MTP build; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 773.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "pyros-vault:Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 773.5 34.2 33.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hj41ahwa",
   "source_url": "https://omlx.ai/benchmarks/performance/hj41ahwa",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hj41ahwa-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "pyros-vault MTP build; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 850.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "pyros-vault:Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 850.0 39.2 34.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hj41ahwa",
   "source_url": "https://omlx.ai/benchmarks/performance/hj41ahwa",
   "tg_tokens": null,
   "tps": 39.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hj41ahwa-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "pyros-vault MTP build; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 823.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "pyros-vault:Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 823.6 50.2 35.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hj41ahwa",
   "source_url": "https://omlx.ai/benchmarks/performance/hj41ahwa",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hj41ahwa-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "pyros-vault MTP build; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 752.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "pyros-vault:Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 752.8 53.2 35.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hj41ahwa",
   "source_url": "https://omlx.ai/benchmarks/performance/hj41ahwa",
   "tg_tokens": null,
   "tps": 53.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hj41ahwa-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "pyros-vault MTP build; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 709.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "pyros-vault:Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 709.2 55.2 36.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hj41ahwa",
   "source_url": "https://omlx.ai/benchmarks/performance/hj41ahwa",
   "tg_tokens": null,
   "tps": 55.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hj41ahwa-ctx65536",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "pyros-vault MTP build; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 617.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "pyros-vault:Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 617.3 41.5 39.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hj41ahwa",
   "source_url": "https://omlx.ai/benchmarks/performance/hj41ahwa",
   "tg_tokens": null,
   "tps": 41.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hj41ahwa-ctx131072",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "pyros-vault MTP build; oMLX community benchmark trace; peak memory 83.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 329.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "pyros-vault:Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 329.1 22.7 83.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hj41ahwa",
   "source_url": "https://omlx.ai/benchmarks/performance/hj41ahwa",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rzplusoi-ctx4096",
   "model": "Bonsai-27B",
   "notes": "oMLX community benchmark trace; peak memory 10.3 GB at this context; duplicate traces of the same session: b7xpzzy7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 377.5,
   "provenance": "community",
   "quant": "1bit",
   "quote": "Bonsai-27B-mlx-1bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 377.5 56.2 10.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rzplusoi",
   "source_url": "https://omlx.ai/benchmarks/performance/rzplusoi",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rzplusoi-ctx16384",
   "model": "Bonsai-27B",
   "notes": "oMLX community benchmark trace; peak memory 11.9 GB at this context; duplicate traces of the same session: b7xpzzy7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 363.8,
   "provenance": "community",
   "quant": "1bit",
   "quote": "Bonsai-27B-mlx-1bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 16k 363.8 48.9 11.9 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rzplusoi",
   "source_url": "https://omlx.ai/benchmarks/performance/rzplusoi",
   "tg_tokens": null,
   "tps": 48.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-z1nee43w-ctx1024",
   "model": "Ornith-1.5-35B-A3B-BigBang",
   "notes": "KaedeTai MTP build; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1158.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "KaedeTai--Ornith-1.5-35B-A3B-BigBang-MTP-mlx-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 1k 1,158 98.5 21.1 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z1nee43w",
   "source_url": "https://omlx.ai/benchmarks/performance/z1nee43w",
   "tg_tokens": null,
   "tps": 98.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-z1nee43w-ctx4096",
   "model": "Ornith-1.5-35B-A3B-BigBang",
   "notes": "KaedeTai MTP build; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1109.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "KaedeTai--Ornith-1.5-35B-A3B-BigBang-MTP-mlx-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 4k 1,109 106.1 21.9 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z1nee43w",
   "source_url": "https://omlx.ai/benchmarks/performance/z1nee43w",
   "tg_tokens": null,
   "tps": 106.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-palfu0y6-ctx4096",
   "model": "LFM2.5-8B-A1B",
   "notes": "oMLX community benchmark trace; peak memory 5.3 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3428.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-8B-A1B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 3,428 104.5 5.3 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/palfu0y6",
   "source_url": "https://omlx.ai/benchmarks/performance/palfu0y6",
   "tg_tokens": null,
   "tps": 104.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-palfu0y6-ctx16384",
   "model": "LFM2.5-8B-A1B",
   "notes": "oMLX community benchmark trace; peak memory 5.5 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3179.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-8B-A1B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 16k 3,179 91.9 5.5 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/palfu0y6",
   "source_url": "https://omlx.ai/benchmarks/performance/palfu0y6",
   "tg_tokens": null,
   "tps": 91.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-wgwpwgon-ctx32768",
   "model": "LFM2.5-8B-A1B",
   "notes": "oMLX community benchmark trace; peak memory 5.6 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2898.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-8B-A1B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 32k 2,898 80.0 5.6 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wgwpwgon",
   "source_url": "https://omlx.ai/benchmarks/performance/wgwpwgon",
   "tg_tokens": null,
   "tps": 80.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-wgwpwgon-ctx131072",
   "model": "LFM2.5-8B-A1B",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1175.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-8B-A1B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 128k 1,175 19.5 6.7 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wgwpwgon",
   "source_url": "https://omlx.ai/benchmarks/performance/wgwpwgon",
   "tg_tokens": null,
   "tps": 19.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-g3cav3pn-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 215 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4767.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 4,767 PP tok/s 150.4 TG tok/s 215 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g3cav3pn",
   "source_url": "https://omlx.ai/benchmarks/performance/g3cav3pn",
   "tg_tokens": null,
   "tps": 150.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-oxq6bssg-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 220 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4652.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 4,652 PP tok/s 142.9 TG tok/s 220 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oxq6bssg",
   "source_url": "https://omlx.ai/benchmarks/performance/oxq6bssg",
   "tg_tokens": null,
   "tps": 142.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-e200w4ty-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5154.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 5,154 107.4 19.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e200w4ty",
   "source_url": "https://omlx.ai/benchmarks/performance/e200w4ty",
   "tg_tokens": null,
   "tps": 107.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-e200w4ty-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6354.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,354 123.7 20.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e200w4ty",
   "source_url": "https://omlx.ai/benchmarks/performance/e200w4ty",
   "tg_tokens": null,
   "tps": 123.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-89d7uxnb-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5173.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 5,173 99.7 20.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/89d7uxnb",
   "source_url": "https://omlx.ai/benchmarks/performance/89d7uxnb",
   "tg_tokens": null,
   "tps": 99.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-89d7uxnb-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6353.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,353 132.0 21.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/89d7uxnb",
   "source_url": "https://omlx.ai/benchmarks/performance/89d7uxnb",
   "tg_tokens": null,
   "tps": 132.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-h7n9y7k7-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 220 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4653.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 4,653 PP tok/s 143.9 TG tok/s 220 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h7n9y7k7",
   "source_url": "https://omlx.ai/benchmarks/performance/h7n9y7k7",
   "tg_tokens": null,
   "tps": 143.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-djd0re78-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 215 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4760.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 4,760 PP tok/s 143.1 TG tok/s 215 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/djd0re78",
   "source_url": "https://omlx.ai/benchmarks/performance/djd0re78",
   "tg_tokens": null,
   "tps": 143.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-d2yv0p5s-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.4 GB at this context; duplicate traces of the same session: 802d4tjq, e1jngcsh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1717.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 1,717 53.1 95.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d2yv0p5s",
   "source_url": "https://omlx.ai/benchmarks/performance/d2yv0p5s",
   "tg_tokens": null,
   "tps": 53.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-d2yv0p5s-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.3 GB at this context; duplicate traces of the same session: 802d4tjq, e1jngcsh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1725.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 1,725 51.2 95.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d2yv0p5s",
   "source_url": "https://omlx.ai/benchmarks/performance/d2yv0p5s",
   "tg_tokens": null,
   "tps": 51.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-d2yv0p5s-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.5 GB at this context; duplicate traces of the same session: 802d4tjq, e1jngcsh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1646.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 1,646 46.9 95.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d2yv0p5s",
   "source_url": "https://omlx.ai/benchmarks/performance/d2yv0p5s",
   "tg_tokens": null,
   "tps": 46.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-d2yv0p5s-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.9 GB at this context; duplicate traces of the same session: 802d4tjq, e1jngcsh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1476.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 1,476 45.6 95.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d2yv0p5s",
   "source_url": "https://omlx.ai/benchmarks/performance/d2yv0p5s",
   "tg_tokens": null,
   "tps": 45.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ibyrresn-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.4 GB at this context; duplicate traces of the same session: 19zntg5k, gynfr246, xf1l2fhf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1712.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 1,712 53.3 95.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ibyrresn",
   "source_url": "https://omlx.ai/benchmarks/performance/ibyrresn",
   "tg_tokens": null,
   "tps": 53.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ibyrresn-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.3 GB at this context; duplicate traces of the same session: 19zntg5k, gynfr246, xf1l2fhf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1731.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 1,731 50.1 95.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ibyrresn",
   "source_url": "https://omlx.ai/benchmarks/performance/ibyrresn",
   "tg_tokens": null,
   "tps": 50.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ibyrresn-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.5 GB at this context; duplicate traces of the same session: 19zntg5k, gynfr246, xf1l2fhf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1651.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 1,651 48.8 95.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ibyrresn",
   "source_url": "https://omlx.ai/benchmarks/performance/ibyrresn",
   "tg_tokens": null,
   "tps": 48.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ibyrresn-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.9 GB at this context; duplicate traces of the same session: 19zntg5k, gynfr246, xf1l2fhf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1484.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 1,484 46.1 95.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ibyrresn",
   "source_url": "https://omlx.ai/benchmarks/performance/ibyrresn",
   "tg_tokens": null,
   "tps": 46.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bmv2nz5c-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.4 GB at this context; duplicate traces of the same session: bplfbxvz, u0bwi2wc, wwp9blxe; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1709.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 1,709 53.0 95.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bmv2nz5c",
   "source_url": "https://omlx.ai/benchmarks/performance/bmv2nz5c",
   "tg_tokens": null,
   "tps": 53.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bmv2nz5c-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.3 GB at this context; duplicate traces of the same session: bplfbxvz, u0bwi2wc, wwp9blxe; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1727.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 1,727 49.7 95.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bmv2nz5c",
   "source_url": "https://omlx.ai/benchmarks/performance/bmv2nz5c",
   "tg_tokens": null,
   "tps": 49.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bmv2nz5c-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.5 GB at this context; duplicate traces of the same session: bplfbxvz, u0bwi2wc, wwp9blxe; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1648.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 1,648 48.3 95.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bmv2nz5c",
   "source_url": "https://omlx.ai/benchmarks/performance/bmv2nz5c",
   "tg_tokens": null,
   "tps": 48.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bmv2nz5c-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.9 GB at this context; duplicate traces of the same session: bplfbxvz, u0bwi2wc, wwp9blxe; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1481.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 1,481 44.9 95.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bmv2nz5c",
   "source_url": "https://omlx.ai/benchmarks/performance/bmv2nz5c",
   "tg_tokens": null,
   "tps": 44.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xei64v36-ctx65536",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 102.4 GB; TTFT 58195 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1126.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k tokens 1,126 PP tok/s 39.7 TG tok/s 58195 TTFT (ms) 102.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xei64v36",
   "source_url": "https://omlx.ai/benchmarks/performance/xei64v36",
   "tg_tokens": null,
   "tps": 39.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-3grd36vj-ctx1024",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: mhf4fjks; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 125.6 15.2 16.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3grd36vj",
   "source_url": "https://omlx.ai/benchmarks/performance/3grd36vj",
   "tg_tokens": null,
   "tps": 15.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-3grd36vj-ctx4096",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: mhf4fjks; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 125.9 14.8 17.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3grd36vj",
   "source_url": "https://omlx.ai/benchmarks/performance/3grd36vj",
   "tg_tokens": null,
   "tps": 14.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-3grd36vj-ctx8192",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: mhf4fjks; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 124.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 124.3 14.2 17.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3grd36vj",
   "source_url": "https://omlx.ai/benchmarks/performance/3grd36vj",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-3grd36vj-ctx16384",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: mhf4fjks; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 121.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 121.0 12.3 18.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3grd36vj",
   "source_url": "https://omlx.ai/benchmarks/performance/3grd36vj",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-3grd36vj-ctx32768",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: mhf4fjks; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 114.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 114.0 9.7 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3grd36vj",
   "source_url": "https://omlx.ai/benchmarks/performance/3grd36vj",
   "tg_tokens": null,
   "tps": 9.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-fkqsk61y-ctx1024",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: l84ddqfd, wp91qzvf; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 82.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 1k 82.0 15.2 16.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fkqsk61y",
   "source_url": "https://omlx.ai/benchmarks/performance/fkqsk61y",
   "tg_tokens": null,
   "tps": 15.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-fkqsk61y-ctx4096",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: l84ddqfd, wp91qzvf; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 77.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k 77.3 14.8 17.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fkqsk61y",
   "source_url": "https://omlx.ai/benchmarks/performance/fkqsk61y",
   "tg_tokens": null,
   "tps": 14.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-fkqsk61y-ctx16384",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: l84ddqfd, wp91qzvf; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 69.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 16k 69.0 13.4 18.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fkqsk61y",
   "source_url": "https://omlx.ai/benchmarks/performance/fkqsk61y",
   "tg_tokens": null,
   "tps": 13.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-qyady0u2-ctx1024",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 111.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 1k 111.1 21.1 15.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qyady0u2",
   "source_url": "https://omlx.ai/benchmarks/performance/qyady0u2",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-qyady0u2-ctx4096",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 113.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k 113.8 20.1 16.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qyady0u2",
   "source_url": "https://omlx.ai/benchmarks/performance/qyady0u2",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-qyady0u2-ctx8192",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 113.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 8k 113.1 19.0 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qyady0u2",
   "source_url": "https://omlx.ai/benchmarks/performance/qyady0u2",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-qyady0u2-ctx16384",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 17.5 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 111.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 16k 111.0 17.3 17.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qyady0u2",
   "source_url": "https://omlx.ai/benchmarks/performance/qyady0u2",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-qyady0u2-ctx32768",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "Claude-4.6-Opus distill; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 32k 106.7 14.9 19.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qyady0u2",
   "source_url": "https://omlx.ai/benchmarks/performance/qyady0u2",
   "tg_tokens": null,
   "tps": 14.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-x9i9s0nc-ctx1024",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 36.2 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 519.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-35B-A3B on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 1k 519.9 48.3 36.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x9i9s0nc",
   "source_url": "https://omlx.ai/benchmarks/performance/x9i9s0nc",
   "tg_tokens": null,
   "tps": 48.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-x9i9s0nc-ctx4096",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 36.3 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 574.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-35B-A3B on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k 574.8 46.9 36.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x9i9s0nc",
   "source_url": "https://omlx.ai/benchmarks/performance/x9i9s0nc",
   "tg_tokens": null,
   "tps": 46.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-femieeij-ctx1024",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1652.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,652 93.4 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/femieeij",
   "source_url": "https://omlx.ai/benchmarks/performance/femieeij",
   "tg_tokens": null,
   "tps": 93.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-femieeij-ctx4096",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1706.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,706 88.6 20.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/femieeij",
   "source_url": "https://omlx.ai/benchmarks/performance/femieeij",
   "tg_tokens": null,
   "tps": 88.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-femieeij-ctx8192",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1487.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,487 90.0 20.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/femieeij",
   "source_url": "https://omlx.ai/benchmarks/performance/femieeij",
   "tg_tokens": null,
   "tps": 90.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-femieeij-ctx16384",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1227.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,227 63.7 20.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/femieeij",
   "source_url": "https://omlx.ai/benchmarks/performance/femieeij",
   "tg_tokens": null,
   "tps": 63.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-femieeij-ctx32768",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1026.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 1,026 42.5 21.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/femieeij",
   "source_url": "https://omlx.ai/benchmarks/performance/femieeij",
   "tg_tokens": null,
   "tps": 42.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-cq1j63z1-ctx1024",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: fxnkfdkj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 867.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 867.2 82.8 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cq1j63z1",
   "source_url": "https://omlx.ai/benchmarks/performance/cq1j63z1",
   "tg_tokens": null,
   "tps": 82.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-cq1j63z1-ctx4096",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: fxnkfdkj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 899.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 899.9 79.2 20.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cq1j63z1",
   "source_url": "https://omlx.ai/benchmarks/performance/cq1j63z1",
   "tg_tokens": null,
   "tps": 79.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-cq1j63z1-ctx8192",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: fxnkfdkj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 876.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 876.2 74.9 20.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cq1j63z1",
   "source_url": "https://omlx.ai/benchmarks/performance/cq1j63z1",
   "tg_tokens": null,
   "tps": 74.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-cq1j63z1-ctx16384",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: fxnkfdkj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 805.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 805.6 57.7 20.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cq1j63z1",
   "source_url": "https://omlx.ai/benchmarks/performance/cq1j63z1",
   "tg_tokens": null,
   "tps": 57.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-cq1j63z1-ctx32768",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: fxnkfdkj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 693.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 693.2 40.6 21.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cq1j63z1",
   "source_url": "https://omlx.ai/benchmarks/performance/cq1j63z1",
   "tg_tokens": null,
   "tps": 40.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-cq1j63z1-ctx65536",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: fxnkfdkj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 537.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 537.6 25.9 23.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cq1j63z1",
   "source_url": "https://omlx.ai/benchmarks/performance/cq1j63z1",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-oia8i4m4-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: none; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 145.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 4bit; 1k 145.3 22.2 3.9 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oia8i4m4",
   "source_url": "https://omlx.ai/benchmarks/performance/oia8i4m4",
   "tg_tokens": null,
   "tps": 22.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-oia8i4m4-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: none; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 144.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 4bit; 4k 144.5 21.0 4.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oia8i4m4",
   "source_url": "https://omlx.ai/benchmarks/performance/oia8i4m4",
   "tg_tokens": null,
   "tps": 21.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-v6gs4f1n-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.0 GB at this context; duplicate traces of the same session: none; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 253.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 1k 253.0 47.5 4.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v6gs4f1n",
   "source_url": "https://omlx.ai/benchmarks/performance/v6gs4f1n",
   "tg_tokens": null,
   "tps": 47.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-v6gs4f1n-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: none; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 255.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 4k 255.0 46.5 4.2 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v6gs4f1n",
   "source_url": "https://omlx.ai/benchmarks/performance/v6gs4f1n",
   "tg_tokens": null,
   "tps": 46.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-aeq42ap1-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.0 GB at this context; duplicate traces of the same session: 0ysklzxp; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 736.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 1k 736.4 102.2 4.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aeq42ap1",
   "source_url": "https://omlx.ai/benchmarks/performance/aeq42ap1",
   "tg_tokens": null,
   "tps": 102.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-aeq42ap1-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: 0ysklzxp; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 734.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k 734.5 95.4 4.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aeq42ap1",
   "source_url": "https://omlx.ai/benchmarks/performance/aeq42ap1",
   "tg_tokens": null,
   "tps": 95.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-aeq42ap1-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.6 GB at this context; duplicate traces of the same session: 0ysklzxp; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 720.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 8k 720.3 87.0 4.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aeq42ap1",
   "source_url": "https://omlx.ai/benchmarks/performance/aeq42ap1",
   "tg_tokens": null,
   "tps": 87.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-aeq42ap1-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 5.3 GB at this context; duplicate traces of the same session: 0ysklzxp; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 691.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 16k 691.3 73.4 5.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aeq42ap1",
   "source_url": "https://omlx.ai/benchmarks/performance/aeq42ap1",
   "tg_tokens": null,
   "tps": 73.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-aeq42ap1-ctx32768",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: 0ysklzxp; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 576.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 32k 576.0 28.6 6.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aeq42ap1",
   "source_url": "https://omlx.ai/benchmarks/performance/aeq42ap1",
   "tg_tokens": null,
   "tps": 28.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-aeq42ap1-ctx65536",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 8.6 GB at this context; duplicate traces of the same session: 0ysklzxp; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 468.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 64k 468.7 18.2 8.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aeq42ap1",
   "source_url": "https://omlx.ai/benchmarks/performance/aeq42ap1",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-j290pxfi-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.0 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1428.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 1k 1,428 137.5 4.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j290pxfi",
   "source_url": "https://omlx.ai/benchmarks/performance/j290pxfi",
   "tg_tokens": null,
   "tps": 137.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-j290pxfi-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1440.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 4k 1,440 128.2 4.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j290pxfi",
   "source_url": "https://omlx.ai/benchmarks/performance/j290pxfi",
   "tg_tokens": null,
   "tps": 128.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-j290pxfi-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.6 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1414.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 8k 1,414 117.4 4.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j290pxfi",
   "source_url": "https://omlx.ai/benchmarks/performance/j290pxfi",
   "tg_tokens": null,
   "tps": 117.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-j290pxfi-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 5.3 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1355.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 16k 1,355 98.7 5.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j290pxfi",
   "source_url": "https://omlx.ai/benchmarks/performance/j290pxfi",
   "tg_tokens": null,
   "tps": 98.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-j290pxfi-ctx32768",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1247.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 32k 1,247 74.5 6.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j290pxfi",
   "source_url": "https://omlx.ai/benchmarks/performance/j290pxfi",
   "tg_tokens": null,
   "tps": 74.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-j290pxfi-ctx65536",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 8.6 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1051.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 64k 1,051 50.4 8.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j290pxfi",
   "source_url": "https://omlx.ai/benchmarks/performance/j290pxfi",
   "tg_tokens": null,
   "tps": 50.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-cgvsvi3i-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1423.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 1k 1,423 99.9 5.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cgvsvi3i",
   "source_url": "https://omlx.ai/benchmarks/performance/cgvsvi3i",
   "tg_tokens": null,
   "tps": 99.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-cgvsvi3i-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 6.1 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1448.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 4k 1,448 96.2 6.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cgvsvi3i",
   "source_url": "https://omlx.ai/benchmarks/performance/cgvsvi3i",
   "tg_tokens": null,
   "tps": 96.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-cgvsvi3i-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1421.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 8k 1,421 90.1 6.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cgvsvi3i",
   "source_url": "https://omlx.ai/benchmarks/performance/cgvsvi3i",
   "tg_tokens": null,
   "tps": 90.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-cgvsvi3i-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 7.3 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1362.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 16k 1,362 80.1 7.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cgvsvi3i",
   "source_url": "https://omlx.ai/benchmarks/performance/cgvsvi3i",
   "tg_tokens": null,
   "tps": 80.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-cgvsvi3i-ctx32768",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 7.9 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1255.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 32k 1,255 63.0 7.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cgvsvi3i",
   "source_url": "https://omlx.ai/benchmarks/performance/cgvsvi3i",
   "tg_tokens": null,
   "tps": 63.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-cgvsvi3i-ctx65536",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 9.9 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1056.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-4B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 64k 1,056 45.5 9.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cgvsvi3i",
   "source_url": "https://omlx.ai/benchmarks/performance/cgvsvi3i",
   "tg_tokens": null,
   "tps": 45.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-s11jretg-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: 91h2g4t3, tir4w3vr; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 374.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 374.9 33.7 3.9 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s11jretg",
   "source_url": "https://omlx.ai/benchmarks/performance/s11jretg",
   "tg_tokens": null,
   "tps": 33.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-s11jretg-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: 91h2g4t3, tir4w3vr; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 373.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 373.3 32.0 4.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s11jretg",
   "source_url": "https://omlx.ai/benchmarks/performance/s11jretg",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-s11jretg-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.6 GB at this context; duplicate traces of the same session: 91h2g4t3, tir4w3vr; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 364.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 8k 364.6 29.5 4.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s11jretg",
   "source_url": "https://omlx.ai/benchmarks/performance/s11jretg",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-s11jretg-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.8 GB at this context; duplicate traces of the same session: 91h2g4t3, tir4w3vr; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 340.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 16k 340.6 25.2 4.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s11jretg",
   "source_url": "https://omlx.ai/benchmarks/performance/s11jretg",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-s11jretg-ctx32768",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: 91h2g4t3, tir4w3vr; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 299.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 32k 299.7 20.3 5.9 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s11jretg",
   "source_url": "https://omlx.ai/benchmarks/performance/s11jretg",
   "tg_tokens": null,
   "tps": 20.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-s11jretg-ctx65536",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 8.0 GB at this context; duplicate traces of the same session: 91h2g4t3, tir4w3vr; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 231.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 64k 231.3 13.6 8.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s11jretg",
   "source_url": "https://omlx.ai/benchmarks/performance/s11jretg",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-nvxbhp1x-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.0 GB at this context; duplicate traces of the same session: qsyh8wbm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 811.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 811.1 82.1 4.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nvxbhp1x",
   "source_url": "https://omlx.ai/benchmarks/performance/nvxbhp1x",
   "tg_tokens": null,
   "tps": 82.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-nvxbhp1x-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: qsyh8wbm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 798.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 798.0 76.8 4.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nvxbhp1x",
   "source_url": "https://omlx.ai/benchmarks/performance/nvxbhp1x",
   "tg_tokens": null,
   "tps": 76.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-nvxbhp1x-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.6 GB at this context; duplicate traces of the same session: qsyh8wbm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 781.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 781.9 72.4 4.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nvxbhp1x",
   "source_url": "https://omlx.ai/benchmarks/performance/nvxbhp1x",
   "tg_tokens": null,
   "tps": 72.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-nvxbhp1x-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 5.3 GB at this context; duplicate traces of the same session: qsyh8wbm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 737.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 737.3 62.8 5.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nvxbhp1x",
   "source_url": "https://omlx.ai/benchmarks/performance/nvxbhp1x",
   "tg_tokens": null,
   "tps": 62.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-nvxbhp1x-ctx32768",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: qsyh8wbm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 655.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 655.0 49.9 6.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nvxbhp1x",
   "source_url": "https://omlx.ai/benchmarks/performance/nvxbhp1x",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-nvxbhp1x-ctx65536",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 8.6 GB at this context; duplicate traces of the same session: qsyh8wbm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 534.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 534.5 36.3 8.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nvxbhp1x",
   "source_url": "https://omlx.ai/benchmarks/performance/nvxbhp1x",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-znelvuh1-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: none; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 77.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 4bit; 1k 77.8 12.6 6.5 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/znelvuh1",
   "source_url": "https://omlx.ai/benchmarks/performance/znelvuh1",
   "tg_tokens": null,
   "tps": 12.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-znelvuh1-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: none; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 77.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 4bit; 4k 77.5 12.1 6.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/znelvuh1",
   "source_url": "https://omlx.ai/benchmarks/performance/znelvuh1",
   "tg_tokens": null,
   "tps": 12.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-6rngmu2d-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.7 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 239.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 8bit; 1k 239.3 35.1 10.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6rngmu2d",
   "source_url": "https://omlx.ai/benchmarks/performance/6rngmu2d",
   "tg_tokens": null,
   "tps": 35.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-6rngmu2d-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.0 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 242.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 8bit; 4k 242.1 33.3 11.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6rngmu2d",
   "source_url": "https://omlx.ai/benchmarks/performance/6rngmu2d",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-6rngmu2d-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.2 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 236.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 8bit; 8k 236.2 30.3 11.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6rngmu2d",
   "source_url": "https://omlx.ai/benchmarks/performance/6rngmu2d",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-6rngmu2d-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.7 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 234.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 8bit; 16k 234.4 26.2 11.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6rngmu2d",
   "source_url": "https://omlx.ai/benchmarks/performance/6rngmu2d",
   "tg_tokens": null,
   "tps": 26.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-6rngmu2d-ctx32768",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 8bit; 32k 224.8 20.3 12.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6rngmu2d",
   "source_url": "https://omlx.ai/benchmarks/performance/6rngmu2d",
   "tg_tokens": null,
   "tps": 20.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-6rngmu2d-ctx65536",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 8bit; 64k 205.0 13.9 14.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6rngmu2d",
   "source_url": "https://omlx.ai/benchmarks/performance/6rngmu2d",
   "tg_tokens": null,
   "tps": 13.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-k9e5j7ms-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: none; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 138.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 1k 138.1 30.0 6.7 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k9e5j7ms",
   "source_url": "https://omlx.ai/benchmarks/performance/k9e5j7ms",
   "tg_tokens": null,
   "tps": 30.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-k9e5j7ms-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: none; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 138.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 4k 138.6 28.6 7.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k9e5j7ms",
   "source_url": "https://omlx.ai/benchmarks/performance/k9e5j7ms",
   "tg_tokens": null,
   "tps": 28.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-6m1wd63n-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.7 GB at this context; duplicate traces of the same session: 1h65skb8, 5qvzekfa, 8w117y2g; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 8bit; 1k 127.6 12.3 10.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6m1wd63n",
   "source_url": "https://omlx.ai/benchmarks/performance/6m1wd63n",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-6m1wd63n-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.0 GB at this context; duplicate traces of the same session: 1h65skb8, 5qvzekfa, 8w117y2g; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 132.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 8bit; 4k 132.0 12.0 11.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6m1wd63n",
   "source_url": "https://omlx.ai/benchmarks/performance/6m1wd63n",
   "tg_tokens": null,
   "tps": 12.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-6m1wd63n-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.2 GB at this context; duplicate traces of the same session: 1h65skb8, 5qvzekfa, 8w117y2g; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 8bit; 8k 130.7 11.5 11.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6m1wd63n",
   "source_url": "https://omlx.ai/benchmarks/performance/6m1wd63n",
   "tg_tokens": null,
   "tps": 11.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-6m1wd63n-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.7 GB at this context; duplicate traces of the same session: 1h65skb8, 5qvzekfa, 8w117y2g; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 8bit; 16k 127.5 10.6 11.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6m1wd63n",
   "source_url": "https://omlx.ai/benchmarks/performance/6m1wd63n",
   "tg_tokens": null,
   "tps": 10.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-6m1wd63n-ctx32768",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: 1h65skb8, 5qvzekfa, 8w117y2g; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 90.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 8bit; 32k 90.4 8.6 12.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6m1wd63n",
   "source_url": "https://omlx.ai/benchmarks/performance/6m1wd63n",
   "tg_tokens": null,
   "tps": 8.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-6o6zm8tl-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.7 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 794.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 1k 794.0 64.3 10.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6o6zm8tl",
   "source_url": "https://omlx.ai/benchmarks/performance/6o6zm8tl",
   "tg_tokens": null,
   "tps": 64.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-6o6zm8tl-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.0 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 804.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 4k 804.0 62.6 11.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6o6zm8tl",
   "source_url": "https://omlx.ai/benchmarks/performance/6o6zm8tl",
   "tg_tokens": null,
   "tps": 62.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-6o6zm8tl-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.2 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 797.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 8k 797.4 59.7 11.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6o6zm8tl",
   "source_url": "https://omlx.ai/benchmarks/performance/6o6zm8tl",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-6o6zm8tl-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.7 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 778.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 16k 778.7 54.0 11.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6o6zm8tl",
   "source_url": "https://omlx.ai/benchmarks/performance/6o6zm8tl",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-6o6zm8tl-ctx32768",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 742.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 32k 742.0 46.4 12.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6o6zm8tl",
   "source_url": "https://omlx.ai/benchmarks/performance/6o6zm8tl",
   "tg_tokens": null,
   "tps": 46.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-6o6zm8tl-ctx65536",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 668.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 8bit; 64k 668.6 33.9 14.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6o6zm8tl",
   "source_url": "https://omlx.ai/benchmarks/performance/6o6zm8tl",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-rr8vgmtb-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: mk5718cc, r7jnwmp2; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 727.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 1k 727.6 96.7 6.7 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rr8vgmtb",
   "source_url": "https://omlx.ai/benchmarks/performance/rr8vgmtb",
   "tg_tokens": null,
   "tps": 96.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-rr8vgmtb-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: mk5718cc, r7jnwmp2; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 789.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 4k 789.6 93.2 7.0 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rr8vgmtb",
   "source_url": "https://omlx.ai/benchmarks/performance/rr8vgmtb",
   "tg_tokens": null,
   "tps": 93.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-rr8vgmtb-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.1 GB at this context; duplicate traces of the same session: mk5718cc, r7jnwmp2; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 792.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 8k 792.6 87.6 7.1 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rr8vgmtb",
   "source_url": "https://omlx.ai/benchmarks/performance/rr8vgmtb",
   "tg_tokens": null,
   "tps": 87.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-rr8vgmtb-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.6 GB at this context; duplicate traces of the same session: mk5718cc, r7jnwmp2; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 774.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 16k 774.4 76.6 7.6 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rr8vgmtb",
   "source_url": "https://omlx.ai/benchmarks/performance/rr8vgmtb",
   "tg_tokens": null,
   "tps": 76.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-rr8vgmtb-ctx32768",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.7 GB at this context; duplicate traces of the same session: mk5718cc, r7jnwmp2; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 738.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 32k 738.1 61.9 8.7 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rr8vgmtb",
   "source_url": "https://omlx.ai/benchmarks/performance/rr8vgmtb",
   "tg_tokens": null,
   "tps": 61.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-rr8vgmtb-ctx65536",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.8 GB at this context; duplicate traces of the same session: mk5718cc, r7jnwmp2; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 665.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 64k 665.7 44.3 10.8 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rr8vgmtb",
   "source_url": "https://omlx.ai/benchmarks/performance/rr8vgmtb",
   "tg_tokens": null,
   "tps": 44.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-crssj6pi-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 155.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 155.3 19.0 6.5 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/crssj6pi",
   "source_url": "https://omlx.ai/benchmarks/performance/crssj6pi",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-crssj6pi-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 155.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 155.6 17.8 6.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/crssj6pi",
   "source_url": "https://omlx.ai/benchmarks/performance/crssj6pi",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-rwmboxzn-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 726.4,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 4k 726.4 53.7 29.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rwmboxzn",
   "source_url": "https://omlx.ai/benchmarks/performance/rwmboxzn",
   "tg_tokens": null,
   "tps": 53.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-rwmboxzn-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 455.5,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 32k 455.5 44.9 32.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rwmboxzn",
   "source_url": "https://omlx.ai/benchmarks/performance/rwmboxzn",
   "tg_tokens": null,
   "tps": 44.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-mfqt1inn-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 453.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 453.3 33.8 19.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mfqt1inn",
   "source_url": "https://omlx.ai/benchmarks/performance/mfqt1inn",
   "tg_tokens": null,
   "tps": 33.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-mfqt1inn-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 430.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 430.4 31.8 21.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mfqt1inn",
   "source_url": "https://omlx.ai/benchmarks/performance/mfqt1inn",
   "tg_tokens": null,
   "tps": 31.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-u31tsm68-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: lj35qhvt; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 79.5,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 79.5 13.0 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u31tsm68",
   "source_url": "https://omlx.ai/benchmarks/performance/u31tsm68",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-u31tsm68-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: lj35qhvt; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 80.4,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 80.4 11.0 25.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u31tsm68",
   "source_url": "https://omlx.ai/benchmarks/performance/u31tsm68",
   "tg_tokens": null,
   "tps": 11.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-8d4iut1w-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 412.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 412.3 61.1 24.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8d4iut1w",
   "source_url": "https://omlx.ai/benchmarks/performance/8d4iut1w",
   "tg_tokens": null,
   "tps": 61.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-8d4iut1w-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 396.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 396.4 63.3 27.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8d4iut1w",
   "source_url": "https://omlx.ai/benchmarks/performance/8d4iut1w",
   "tg_tokens": null,
   "tps": 63.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n64tq8ac-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 910.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 910.6 32.0 17.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n64tq8ac",
   "source_url": "https://omlx.ai/benchmarks/performance/n64tq8ac",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n64tq8ac-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 931.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 931.4 35.8 18.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n64tq8ac",
   "source_url": "https://omlx.ai/benchmarks/performance/n64tq8ac",
   "tg_tokens": null,
   "tps": 35.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bnntiaco-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 40o2d3uj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1125.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,125 58.0 17.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bnntiaco",
   "source_url": "https://omlx.ai/benchmarks/performance/bnntiaco",
   "tg_tokens": null,
   "tps": 58.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bnntiaco-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 40o2d3uj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1336.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,336 93.0 19.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bnntiaco",
   "source_url": "https://omlx.ai/benchmarks/performance/bnntiaco",
   "tg_tokens": null,
   "tps": 93.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-z1tgsltk-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: kn01zcvj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 216.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 216.8 32.8 33.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z1tgsltk",
   "source_url": "https://omlx.ai/benchmarks/performance/z1tgsltk",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-z1tgsltk-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: kn01zcvj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 218.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 218.1 27.9 37.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z1tgsltk",
   "source_url": "https://omlx.ai/benchmarks/performance/z1tgsltk",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-z1tgsltk-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 40.6 GB at this context; duplicate traces of the same session: kn01zcvj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 186.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 186.2 32.0 40.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z1tgsltk",
   "source_url": "https://omlx.ai/benchmarks/performance/z1tgsltk",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-ya9hlawm-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 135.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 135.0 23.7 33.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ya9hlawm",
   "source_url": "https://omlx.ai/benchmarks/performance/ya9hlawm",
   "tg_tokens": null,
   "tps": 23.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ya9hlawm-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 221.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 221.6 26.6 37.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ya9hlawm",
   "source_url": "https://omlx.ai/benchmarks/performance/ya9hlawm",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9cttvdxm-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.1 GB at this context; duplicate traces of the same session: 3dtdicn0, jkowgmac, x342yu55; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 914.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 914.1 29.3 35.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9cttvdxm",
   "source_url": "https://omlx.ai/benchmarks/performance/9cttvdxm",
   "tg_tokens": null,
   "tps": 29.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9cttvdxm-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: 3dtdicn0, jkowgmac, x342yu55; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 922.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 922.8 25.9 35.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9cttvdxm",
   "source_url": "https://omlx.ai/benchmarks/performance/9cttvdxm",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9cttvdxm-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: 3dtdicn0, jkowgmac, x342yu55; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 901.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 901.1 36.3 35.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9cttvdxm",
   "source_url": "https://omlx.ai/benchmarks/performance/9cttvdxm",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9cttvdxm-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: 3dtdicn0, jkowgmac, x342yu55; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 845.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 845.1 28.4 36.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9cttvdxm",
   "source_url": "https://omlx.ai/benchmarks/performance/9cttvdxm",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gs1tj5df-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.1 GB at this context; duplicate traces of the same session: chol024p, p6c126fe, purne41h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 914.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 914.8 39.5 35.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gs1tj5df",
   "source_url": "https://omlx.ai/benchmarks/performance/gs1tj5df",
   "tg_tokens": null,
   "tps": 39.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gs1tj5df-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: chol024p, p6c126fe, purne41h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 923.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 923.9 26.6 35.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gs1tj5df",
   "source_url": "https://omlx.ai/benchmarks/performance/gs1tj5df",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gs1tj5df-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: chol024p, p6c126fe, purne41h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 901.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 901.2 28.2 35.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gs1tj5df",
   "source_url": "https://omlx.ai/benchmarks/performance/gs1tj5df",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gs1tj5df-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: chol024p, p6c126fe, purne41h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 844.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 844.4 25.9 36.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gs1tj5df",
   "source_url": "https://omlx.ai/benchmarks/performance/gs1tj5df",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zu6exgpi-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.1 GB at this context; duplicate traces of the same session: alt2dmzj, gy0lzrga, u0x3kuj4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 916.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 916.4 38.1 35.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zu6exgpi",
   "source_url": "https://omlx.ai/benchmarks/performance/zu6exgpi",
   "tg_tokens": null,
   "tps": 38.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zu6exgpi-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: alt2dmzj, gy0lzrga, u0x3kuj4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 923.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 923.8 27.2 35.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zu6exgpi",
   "source_url": "https://omlx.ai/benchmarks/performance/zu6exgpi",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zu6exgpi-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: alt2dmzj, gy0lzrga, u0x3kuj4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 901.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 901.5 27.4 35.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zu6exgpi",
   "source_url": "https://omlx.ai/benchmarks/performance/zu6exgpi",
   "tg_tokens": null,
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zu6exgpi-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: alt2dmzj, gy0lzrga, u0x3kuj4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 844.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 844.0 28.5 36.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zu6exgpi",
   "source_url": "https://omlx.ai/benchmarks/performance/zu6exgpi",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-60q4e4fn-ctx1024",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: 5iyg0ao2, tggxugfz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 842.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ6e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 842.1 88.0 29.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/60q4e4fn",
   "source_url": "https://omlx.ai/benchmarks/performance/60q4e4fn",
   "tg_tokens": null,
   "tps": 88.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-60q4e4fn-ctx4096",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.9 GB at this context; duplicate traces of the same session: 5iyg0ao2, tggxugfz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 989.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ6e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 989.3 70.8 31.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/60q4e4fn",
   "source_url": "https://omlx.ai/benchmarks/performance/60q4e4fn",
   "tg_tokens": null,
   "tps": 70.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-60q4e4fn-ctx8192",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.2 GB at this context; duplicate traces of the same session: 5iyg0ao2, tggxugfz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 866.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ6e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 8k 866.7 94.2 32.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/60q4e4fn",
   "source_url": "https://omlx.ai/benchmarks/performance/60q4e4fn",
   "tg_tokens": null,
   "tps": 94.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-60q4e4fn-ctx16384",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.7 GB at this context; duplicate traces of the same session: 5iyg0ao2, tggxugfz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 711.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ6e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 16k 711.5 81.4 32.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/60q4e4fn",
   "source_url": "https://omlx.ai/benchmarks/performance/60q4e4fn",
   "tg_tokens": null,
   "tps": 81.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5xdnwg0o-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: s79lihtn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 931.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 931.0 73.4 70.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5xdnwg0o",
   "source_url": "https://omlx.ai/benchmarks/performance/5xdnwg0o",
   "tg_tokens": null,
   "tps": 73.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5xdnwg0o-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: s79lihtn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1701.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,701 66.0 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5xdnwg0o",
   "source_url": "https://omlx.ai/benchmarks/performance/5xdnwg0o",
   "tg_tokens": null,
   "tps": 66.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n35k0hk0-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.9 GB; TTFT 966 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1060.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,060 PP tok/s 69.2 TG tok/s 966 TTFT (ms) 70.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n35k0hk0",
   "source_url": "https://omlx.ai/benchmarks/performance/n35k0hk0",
   "tg_tokens": null,
   "tps": 69.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hi095tsp-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 87.2 GB at this context; duplicate traces of the same session: cnkqwckr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1397.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 1,397 61.8 87.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hi095tsp",
   "source_url": "https://omlx.ai/benchmarks/performance/hi095tsp",
   "tg_tokens": null,
   "tps": 61.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hi095tsp-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 87.8 GB at this context; duplicate traces of the same session: cnkqwckr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2059.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 2,059 56.5 87.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hi095tsp",
   "source_url": "https://omlx.ai/benchmarks/performance/hi095tsp",
   "tg_tokens": null,
   "tps": 56.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-66e06q4y-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 87.1 GB at this context; duplicate traces of the same session: iiqlm9iz, sn7zinna, vmupds27; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1439.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 1,439 63.2 87.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/66e06q4y",
   "source_url": "https://omlx.ai/benchmarks/performance/66e06q4y",
   "tg_tokens": null,
   "tps": 63.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-66e06q4y-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 87.8 GB at this context; duplicate traces of the same session: iiqlm9iz, sn7zinna, vmupds27; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2078.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 2,078 66.1 87.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/66e06q4y",
   "source_url": "https://omlx.ai/benchmarks/performance/66e06q4y",
   "tg_tokens": null,
   "tps": 66.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-66e06q4y-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 88.9 GB at this context; duplicate traces of the same session: iiqlm9iz, sn7zinna, vmupds27; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2444.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 2,444 60.4 88.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/66e06q4y",
   "source_url": "https://omlx.ai/benchmarks/performance/66e06q4y",
   "tg_tokens": null,
   "tps": 60.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-66e06q4y-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 89.5 GB at this context; duplicate traces of the same session: iiqlm9iz, sn7zinna, vmupds27; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2456.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 2,456 60.3 89.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/66e06q4y",
   "source_url": "https://omlx.ai/benchmarks/performance/66e06q4y",
   "tg_tokens": null,
   "tps": 60.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mepfwgr4-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 87.3 GB at this context; duplicate traces of the same session: 5n3l3ofa, i4898hgl, j8hi3657; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1036.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 1,036 57.2 87.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mepfwgr4",
   "source_url": "https://omlx.ai/benchmarks/performance/mepfwgr4",
   "tg_tokens": null,
   "tps": 57.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mepfwgr4-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 87.8 GB at this context; duplicate traces of the same session: 5n3l3ofa, i4898hgl, j8hi3657; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1918.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 1,918 56.6 87.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mepfwgr4",
   "source_url": "https://omlx.ai/benchmarks/performance/mepfwgr4",
   "tg_tokens": null,
   "tps": 56.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mepfwgr4-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 88.9 GB at this context; duplicate traces of the same session: 5n3l3ofa, i4898hgl, j8hi3657; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2339.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 2,339 56.1 88.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mepfwgr4",
   "source_url": "https://omlx.ai/benchmarks/performance/mepfwgr4",
   "tg_tokens": null,
   "tps": 56.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mepfwgr4-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 89.5 GB at this context; duplicate traces of the same session: 5n3l3ofa, i4898hgl, j8hi3657; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2480.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 2,480 56.4 89.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mepfwgr4",
   "source_url": "https://omlx.ai/benchmarks/performance/mepfwgr4",
   "tg_tokens": null,
   "tps": 56.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-bt839ka2-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1042.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 1,042 50.4 14.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bt839ka2",
   "source_url": "https://omlx.ai/benchmarks/performance/bt839ka2",
   "tg_tokens": null,
   "tps": 50.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-bt839ka2-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.5 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1318.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,318 49.0 14.5 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bt839ka2",
   "source_url": "https://omlx.ai/benchmarks/performance/bt839ka2",
   "tg_tokens": null,
   "tps": 49.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-b2s9t2j3-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3631.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 3,631 112.5 14.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b2s9t2j3",
   "source_url": "https://omlx.ai/benchmarks/performance/b2s9t2j3",
   "tg_tokens": null,
   "tps": 112.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-b2s9t2j3-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4033.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 4,033 113.6 14.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b2s9t2j3",
   "source_url": "https://omlx.ai/benchmarks/performance/b2s9t2j3",
   "tg_tokens": null,
   "tps": 113.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-dgbnn70f-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.1 GB at this context; duplicate traces of the same session: m4pizi09; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 495.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 1k 495.1 52.2 14.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dgbnn70f",
   "source_url": "https://omlx.ai/benchmarks/performance/dgbnn70f",
   "tg_tokens": null,
   "tps": 52.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-dgbnn70f-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: m4pizi09; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 490.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k 490.7 50.2 14.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dgbnn70f",
   "source_url": "https://omlx.ai/benchmarks/performance/dgbnn70f",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-dgbnn70f-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.9 GB at this context; duplicate traces of the same session: m4pizi09; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 456.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 16k 456.8 42.1 14.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dgbnn70f",
   "source_url": "https://omlx.ai/benchmarks/performance/dgbnn70f",
   "tg_tokens": null,
   "tps": 42.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-lb0vetyp-ctx1024",
   "model": "Gemma-4-31B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 4kjjp3pk, jjvej6hr; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 66.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31b-it-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 1k 66.6 11.4 17.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lb0vetyp",
   "source_url": "https://omlx.ai/benchmarks/performance/lb0vetyp",
   "tg_tokens": null,
   "tps": 11.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-lb0vetyp-ctx4096",
   "model": "Gemma-4-31B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 4kjjp3pk, jjvej6hr; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 66.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31b-it-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k 66.6 11.1 18.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lb0vetyp",
   "source_url": "https://omlx.ai/benchmarks/performance/lb0vetyp",
   "tg_tokens": null,
   "tps": 11.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-lb0vetyp-ctx16384",
   "model": "Gemma-4-31B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: 4kjjp3pk, jjvej6hr; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 63.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31b-it-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 16k 63.1 9.6 20.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lb0vetyp",
   "source_url": "https://omlx.ai/benchmarks/performance/lb0vetyp",
   "tg_tokens": null,
   "tps": 9.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-o2207fu5-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 5dh8u8jj; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 561.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 1k 561.8 65.9 19.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o2207fu5",
   "source_url": "https://omlx.ai/benchmarks/performance/o2207fu5",
   "tg_tokens": null,
   "tps": 65.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-o2207fu5-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 5dh8u8jj; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 565.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k 565.2 66.2 20.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o2207fu5",
   "source_url": "https://omlx.ai/benchmarks/performance/o2207fu5",
   "tg_tokens": null,
   "tps": 66.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-o2207fu5-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: 5dh8u8jj; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 442.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 16k 442.6 54.5 20.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o2207fu5",
   "source_url": "https://omlx.ai/benchmarks/performance/o2207fu5",
   "tg_tokens": null,
   "tps": 54.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-wysrquee-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "scottlowry MTP build; oMLX community benchmark trace; peak memory 30.0 GB at this context; duplicate traces of the same session: zq5bhoyg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 702.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "models--scottlowry--Ornith-1.5-35B-A3B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 702.9 74.2 30.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wysrquee",
   "source_url": "https://omlx.ai/benchmarks/performance/wysrquee",
   "tg_tokens": null,
   "tps": 74.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-wysrquee-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "scottlowry MTP build; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: zq5bhoyg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 788.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "models--scottlowry--Ornith-1.5-35B-A3B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 788.3 69.2 32.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wysrquee",
   "source_url": "https://omlx.ai/benchmarks/performance/wysrquee",
   "tg_tokens": null,
   "tps": 69.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-wysrquee-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "scottlowry MTP build; oMLX community benchmark trace; peak memory 32.3 GB at this context; duplicate traces of the same session: zq5bhoyg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 753.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "models--scottlowry--Ornith-1.5-35B-A3B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 6bit; 8k 753.7 56.7 32.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wysrquee",
   "source_url": "https://omlx.ai/benchmarks/performance/wysrquee",
   "tg_tokens": null,
   "tps": 56.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-b7xpzzy7-ctx4096",
   "model": "Bonsai-27B",
   "notes": "oMLX community benchmark trace; peak memory 10.3 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 377.5,
   "provenance": "community",
   "quant": "1bit",
   "quote": "Bonsai-27B-mlx-1bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 377.5 56.2 10.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b7xpzzy7",
   "source_url": "https://omlx.ai/benchmarks/performance/b7xpzzy7",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-b7xpzzy7-ctx16384",
   "model": "Bonsai-27B",
   "notes": "oMLX community benchmark trace; peak memory 11.9 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 363.8,
   "provenance": "community",
   "quant": "1bit",
   "quote": "Bonsai-27B-mlx-1bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 16k 363.8 48.9 11.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b7xpzzy7",
   "source_url": "https://omlx.ai/benchmarks/performance/b7xpzzy7",
   "tg_tokens": null,
   "tps": 48.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0rxzkj0t-ctx16384",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.4 GB at this context; duplicate traces of the same session: 0jmj5tkk, qfmh3bdj, ra3j8y09; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 449.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "DeepSeek-V4-Flash-0731-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 16k 449.8 37.1 156.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0rxzkj0t",
   "source_url": "https://omlx.ai/benchmarks/performance/0rxzkj0t",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0rxzkj0t-ctx32768",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.5 GB at this context; duplicate traces of the same session: 0jmj5tkk, qfmh3bdj, ra3j8y09; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 450.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "DeepSeek-V4-Flash-0731-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 32k 450.3 45.1 156.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0rxzkj0t",
   "source_url": "https://omlx.ai/benchmarks/performance/0rxzkj0t",
   "tg_tokens": null,
   "tps": 45.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0rxzkj0t-ctx65536",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.8 GB at this context; duplicate traces of the same session: 0jmj5tkk, qfmh3bdj, ra3j8y09; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 426.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "DeepSeek-V4-Flash-0731-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 64k 426.2 36.5 156.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0rxzkj0t",
   "source_url": "https://omlx.ai/benchmarks/performance/0rxzkj0t",
   "tg_tokens": null,
   "tps": 36.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0rxzkj0t-ctx131072",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 157.1 GB at this context; duplicate traces of the same session: 0jmj5tkk, qfmh3bdj, ra3j8y09; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 385.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "DeepSeek-V4-Flash-0731-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 128k 385.6 35.1 157.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0rxzkj0t",
   "source_url": "https://omlx.ai/benchmarks/performance/0rxzkj0t",
   "tg_tokens": null,
   "tps": 35.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-a1xjithc-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5173.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 5,173 99.7 20.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a1xjithc",
   "source_url": "https://omlx.ai/benchmarks/performance/a1xjithc",
   "tg_tokens": null,
   "tps": 99.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-a1xjithc-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6353.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,353 132.0 21.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a1xjithc",
   "source_url": "https://omlx.ai/benchmarks/performance/a1xjithc",
   "tg_tokens": null,
   "tps": 132.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zb2v8ns6-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5154.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 5,154 107.4 19.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zb2v8ns6",
   "source_url": "https://omlx.ai/benchmarks/performance/zb2v8ns6",
   "tg_tokens": null,
   "tps": 107.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zb2v8ns6-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6354.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,354 123.7 20.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zb2v8ns6",
   "source_url": "https://omlx.ai/benchmarks/performance/zb2v8ns6",
   "tg_tokens": null,
   "tps": 123.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-19zntg5k-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.4 GB at this context; duplicate traces of the same session: gynfr246, xf1l2fhf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1712.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 1,712 53.3 95.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/19zntg5k",
   "source_url": "https://omlx.ai/benchmarks/performance/19zntg5k",
   "tg_tokens": null,
   "tps": 53.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-19zntg5k-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.3 GB at this context; duplicate traces of the same session: gynfr246, xf1l2fhf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1731.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 1,731 50.1 95.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/19zntg5k",
   "source_url": "https://omlx.ai/benchmarks/performance/19zntg5k",
   "tg_tokens": null,
   "tps": 50.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-19zntg5k-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.5 GB at this context; duplicate traces of the same session: gynfr246, xf1l2fhf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1651.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 1,651 48.8 95.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/19zntg5k",
   "source_url": "https://omlx.ai/benchmarks/performance/19zntg5k",
   "tg_tokens": null,
   "tps": 48.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-19zntg5k-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.9 GB at this context; duplicate traces of the same session: gynfr246, xf1l2fhf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1484.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 1,484 46.1 95.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/19zntg5k",
   "source_url": "https://omlx.ai/benchmarks/performance/19zntg5k",
   "tg_tokens": null,
   "tps": 46.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-802d4tjq-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.4 GB at this context; duplicate traces of the same session: e1jngcsh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1717.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 1,717 53.1 95.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/802d4tjq",
   "source_url": "https://omlx.ai/benchmarks/performance/802d4tjq",
   "tg_tokens": null,
   "tps": 53.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-802d4tjq-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.3 GB at this context; duplicate traces of the same session: e1jngcsh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1725.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 1,725 51.2 95.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/802d4tjq",
   "source_url": "https://omlx.ai/benchmarks/performance/802d4tjq",
   "tg_tokens": null,
   "tps": 51.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-802d4tjq-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.5 GB at this context; duplicate traces of the same session: e1jngcsh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1646.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 1,646 46.9 95.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/802d4tjq",
   "source_url": "https://omlx.ai/benchmarks/performance/802d4tjq",
   "tg_tokens": null,
   "tps": 46.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-802d4tjq-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.9 GB at this context; duplicate traces of the same session: e1jngcsh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1476.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 1,476 45.6 95.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/802d4tjq",
   "source_url": "https://omlx.ai/benchmarks/performance/802d4tjq",
   "tg_tokens": null,
   "tps": 45.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wwp9blxe-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.4 GB at this context; duplicate traces of the same session: bplfbxvz, u0bwi2wc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1709.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 1,709 53.0 95.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wwp9blxe",
   "source_url": "https://omlx.ai/benchmarks/performance/wwp9blxe",
   "tg_tokens": null,
   "tps": 53.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wwp9blxe-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.3 GB at this context; duplicate traces of the same session: bplfbxvz, u0bwi2wc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1727.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 1,727 49.7 95.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wwp9blxe",
   "source_url": "https://omlx.ai/benchmarks/performance/wwp9blxe",
   "tg_tokens": null,
   "tps": 49.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wwp9blxe-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.5 GB at this context; duplicate traces of the same session: bplfbxvz, u0bwi2wc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1648.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 1,648 48.3 95.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wwp9blxe",
   "source_url": "https://omlx.ai/benchmarks/performance/wwp9blxe",
   "tg_tokens": null,
   "tps": 48.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wwp9blxe-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.9 GB at this context; duplicate traces of the same session: bplfbxvz, u0bwi2wc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1481.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 1,481 44.9 95.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wwp9blxe",
   "source_url": "https://omlx.ai/benchmarks/performance/wwp9blxe",
   "tg_tokens": null,
   "tps": 44.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-036tptd6-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.9 GB at this context; duplicate traces of the same session: kibimees, ommqe00g, qbz7e0al; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1666.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 1,666 58.6 100.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/036tptd6",
   "source_url": "https://omlx.ai/benchmarks/performance/036tptd6",
   "tg_tokens": null,
   "tps": 58.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-036tptd6-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.8 GB at this context; duplicate traces of the same session: kibimees, ommqe00g, qbz7e0al; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1657.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 1,657 64.9 100.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/036tptd6",
   "source_url": "https://omlx.ai/benchmarks/performance/036tptd6",
   "tg_tokens": null,
   "tps": 64.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-036tptd6-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.0 GB at this context; duplicate traces of the same session: kibimees, ommqe00g, qbz7e0al; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1574.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 1,574 60.5 101.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/036tptd6",
   "source_url": "https://omlx.ai/benchmarks/performance/036tptd6",
   "tg_tokens": null,
   "tps": 60.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-036tptd6-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: kibimees, ommqe00g, qbz7e0al; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1407.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 1,407 55.9 101.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/036tptd6",
   "source_url": "https://omlx.ai/benchmarks/performance/036tptd6",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-27vu02k4-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.9 GB at this context; duplicate traces of the same session: e4zqm2j0, p6oyzjro, wk6mv5n4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1672.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 1,672 63.6 100.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/27vu02k4",
   "source_url": "https://omlx.ai/benchmarks/performance/27vu02k4",
   "tg_tokens": null,
   "tps": 63.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-27vu02k4-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.8 GB at this context; duplicate traces of the same session: e4zqm2j0, p6oyzjro, wk6mv5n4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1654.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 1,654 61.0 100.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/27vu02k4",
   "source_url": "https://omlx.ai/benchmarks/performance/27vu02k4",
   "tg_tokens": null,
   "tps": 61.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-27vu02k4-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.0 GB at this context; duplicate traces of the same session: e4zqm2j0, p6oyzjro, wk6mv5n4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1573.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 1,573 60.5 101.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/27vu02k4",
   "source_url": "https://omlx.ai/benchmarks/performance/27vu02k4",
   "tg_tokens": null,
   "tps": 60.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-27vu02k4-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: e4zqm2j0, p6oyzjro, wk6mv5n4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1411.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 1,411 56.7 101.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/27vu02k4",
   "source_url": "https://omlx.ai/benchmarks/performance/27vu02k4",
   "tg_tokens": null,
   "tps": 56.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jm9bmp3w-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.9 GB at this context; duplicate traces of the same session: 185q07hf, a3ok0fr2, hl6sujug; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1665.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 1,665 60.4 100.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jm9bmp3w",
   "source_url": "https://omlx.ai/benchmarks/performance/jm9bmp3w",
   "tg_tokens": null,
   "tps": 60.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jm9bmp3w-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.8 GB at this context; duplicate traces of the same session: 185q07hf, a3ok0fr2, hl6sujug; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1647.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 1,647 64.0 100.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jm9bmp3w",
   "source_url": "https://omlx.ai/benchmarks/performance/jm9bmp3w",
   "tg_tokens": null,
   "tps": 64.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jm9bmp3w-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.0 GB at this context; duplicate traces of the same session: 185q07hf, a3ok0fr2, hl6sujug; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1578.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 1,578 61.8 101.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jm9bmp3w",
   "source_url": "https://omlx.ai/benchmarks/performance/jm9bmp3w",
   "tg_tokens": null,
   "tps": 61.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jm9bmp3w-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: 185q07hf, a3ok0fr2, hl6sujug; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1411.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 1,411 52.2 101.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jm9bmp3w",
   "source_url": "https://omlx.ai/benchmarks/performance/jm9bmp3w",
   "tg_tokens": null,
   "tps": 52.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7niqowrk-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.2 GB at this context; duplicate traces of the same session: 2njxgwlk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1699.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 1,699 51.2 96.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7niqowrk",
   "source_url": "https://omlx.ai/benchmarks/performance/7niqowrk",
   "tg_tokens": null,
   "tps": 51.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7niqowrk-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.1 GB at this context; duplicate traces of the same session: 2njxgwlk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1701.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 1,701 48.9 96.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7niqowrk",
   "source_url": "https://omlx.ai/benchmarks/performance/7niqowrk",
   "tg_tokens": null,
   "tps": 48.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7niqowrk-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.3 GB at this context; duplicate traces of the same session: 2njxgwlk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1624.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 1,624 47.0 96.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7niqowrk",
   "source_url": "https://omlx.ai/benchmarks/performance/7niqowrk",
   "tg_tokens": null,
   "tps": 47.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7niqowrk-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.7 GB at this context; duplicate traces of the same session: 2njxgwlk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1463.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 1,463 44.5 96.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7niqowrk",
   "source_url": "https://omlx.ai/benchmarks/performance/7niqowrk",
   "tg_tokens": null,
   "tps": 44.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jk92esg9-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1700.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 1,700 51.1 96.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jk92esg9",
   "source_url": "https://omlx.ai/benchmarks/performance/jk92esg9",
   "tg_tokens": null,
   "tps": 51.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jk92esg9-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1694.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 1,694 48.3 96.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jk92esg9",
   "source_url": "https://omlx.ai/benchmarks/performance/jk92esg9",
   "tg_tokens": null,
   "tps": 48.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jk92esg9-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1627.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 1,627 46.0 96.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jk92esg9",
   "source_url": "https://omlx.ai/benchmarks/performance/jk92esg9",
   "tg_tokens": null,
   "tps": 46.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jk92esg9-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1457.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 1,457 44.3 96.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jk92esg9",
   "source_url": "https://omlx.ai/benchmarks/performance/jk92esg9",
   "tg_tokens": null,
   "tps": 44.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-mhf4fjks-ctx1024",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 125.6 15.2 16.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mhf4fjks",
   "source_url": "https://omlx.ai/benchmarks/performance/mhf4fjks",
   "tg_tokens": null,
   "tps": 15.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-mhf4fjks-ctx4096",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 125.9 14.8 17.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mhf4fjks",
   "source_url": "https://omlx.ai/benchmarks/performance/mhf4fjks",
   "tg_tokens": null,
   "tps": 14.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-mhf4fjks-ctx8192",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 124.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 124.3 14.2 17.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mhf4fjks",
   "source_url": "https://omlx.ai/benchmarks/performance/mhf4fjks",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-mhf4fjks-ctx16384",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 121.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 121.0 12.3 18.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mhf4fjks",
   "source_url": "https://omlx.ai/benchmarks/performance/mhf4fjks",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-mhf4fjks-ctx32768",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 114.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 114.0 9.7 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mhf4fjks",
   "source_url": "https://omlx.ai/benchmarks/performance/mhf4fjks",
   "tg_tokens": null,
   "tps": 9.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-l84ddqfd-ctx1024",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: wp91qzvf; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 82.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 1k 82.0 15.2 16.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l84ddqfd",
   "source_url": "https://omlx.ai/benchmarks/performance/l84ddqfd",
   "tg_tokens": null,
   "tps": 15.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-l84ddqfd-ctx4096",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: wp91qzvf; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 77.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k 77.3 14.8 17.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l84ddqfd",
   "source_url": "https://omlx.ai/benchmarks/performance/l84ddqfd",
   "tg_tokens": null,
   "tps": 14.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-l84ddqfd-ctx16384",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: wp91qzvf; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 69.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 16k 69.0 13.4 18.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l84ddqfd",
   "source_url": "https://omlx.ai/benchmarks/performance/l84ddqfd",
   "tg_tokens": null,
   "tps": 13.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-fxnkfdkj-ctx1024",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 867.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 867.2 82.8 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fxnkfdkj",
   "source_url": "https://omlx.ai/benchmarks/performance/fxnkfdkj",
   "tg_tokens": null,
   "tps": 82.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-fxnkfdkj-ctx4096",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 899.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 899.9 79.2 20.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fxnkfdkj",
   "source_url": "https://omlx.ai/benchmarks/performance/fxnkfdkj",
   "tg_tokens": null,
   "tps": 79.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-fxnkfdkj-ctx8192",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 876.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 876.2 74.9 20.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fxnkfdkj",
   "source_url": "https://omlx.ai/benchmarks/performance/fxnkfdkj",
   "tg_tokens": null,
   "tps": 74.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-fxnkfdkj-ctx16384",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 805.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 805.6 57.7 20.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fxnkfdkj",
   "source_url": "https://omlx.ai/benchmarks/performance/fxnkfdkj",
   "tg_tokens": null,
   "tps": 57.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-fxnkfdkj-ctx32768",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 693.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 693.2 40.6 21.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fxnkfdkj",
   "source_url": "https://omlx.ai/benchmarks/performance/fxnkfdkj",
   "tg_tokens": null,
   "tps": 40.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-fxnkfdkj-ctx65536",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 537.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 537.6 25.9 23.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fxnkfdkj",
   "source_url": "https://omlx.ai/benchmarks/performance/fxnkfdkj",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-0ysklzxp-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.0 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 736.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 1k 736.4 102.2 4.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ysklzxp",
   "source_url": "https://omlx.ai/benchmarks/performance/0ysklzxp",
   "tg_tokens": null,
   "tps": 102.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-0ysklzxp-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 734.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k 734.5 95.4 4.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ysklzxp",
   "source_url": "https://omlx.ai/benchmarks/performance/0ysklzxp",
   "tg_tokens": null,
   "tps": 95.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-0ysklzxp-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.6 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 720.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 8k 720.3 87.0 4.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ysklzxp",
   "source_url": "https://omlx.ai/benchmarks/performance/0ysklzxp",
   "tg_tokens": null,
   "tps": 87.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-0ysklzxp-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 5.3 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 691.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 16k 691.3 73.4 5.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ysklzxp",
   "source_url": "https://omlx.ai/benchmarks/performance/0ysklzxp",
   "tg_tokens": null,
   "tps": 73.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-0ysklzxp-ctx32768",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 576.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 32k 576.0 28.6 6.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ysklzxp",
   "source_url": "https://omlx.ai/benchmarks/performance/0ysklzxp",
   "tg_tokens": null,
   "tps": 28.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-0ysklzxp-ctx65536",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 8.6 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 468.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 64k 468.7 18.2 8.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ysklzxp",
   "source_url": "https://omlx.ai/benchmarks/performance/0ysklzxp",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-91h2g4t3-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: tir4w3vr; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 374.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 374.9 33.7 3.9 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/91h2g4t3",
   "source_url": "https://omlx.ai/benchmarks/performance/91h2g4t3",
   "tg_tokens": null,
   "tps": 33.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-91h2g4t3-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: tir4w3vr; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 373.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 373.3 32.0 4.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/91h2g4t3",
   "source_url": "https://omlx.ai/benchmarks/performance/91h2g4t3",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-91h2g4t3-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.6 GB at this context; duplicate traces of the same session: tir4w3vr; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 364.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 8k 364.6 29.5 4.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/91h2g4t3",
   "source_url": "https://omlx.ai/benchmarks/performance/91h2g4t3",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-91h2g4t3-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.8 GB at this context; duplicate traces of the same session: tir4w3vr; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 340.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 16k 340.6 25.2 4.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/91h2g4t3",
   "source_url": "https://omlx.ai/benchmarks/performance/91h2g4t3",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-91h2g4t3-ctx32768",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: tir4w3vr; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 299.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 32k 299.7 20.3 5.9 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/91h2g4t3",
   "source_url": "https://omlx.ai/benchmarks/performance/91h2g4t3",
   "tg_tokens": null,
   "tps": 20.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-91h2g4t3-ctx65536",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 8.0 GB at this context; duplicate traces of the same session: tir4w3vr; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 231.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 64k 231.3 13.6 8.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/91h2g4t3",
   "source_url": "https://omlx.ai/benchmarks/performance/91h2g4t3",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-qsyh8wbm-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 811.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 811.1 82.1 4.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qsyh8wbm",
   "source_url": "https://omlx.ai/benchmarks/performance/qsyh8wbm",
   "tg_tokens": null,
   "tps": 82.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-qsyh8wbm-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 798.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 798.0 76.8 4.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qsyh8wbm",
   "source_url": "https://omlx.ai/benchmarks/performance/qsyh8wbm",
   "tg_tokens": null,
   "tps": 76.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-qsyh8wbm-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 781.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 781.9 72.4 4.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qsyh8wbm",
   "source_url": "https://omlx.ai/benchmarks/performance/qsyh8wbm",
   "tg_tokens": null,
   "tps": 72.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-qsyh8wbm-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 5.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 737.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 737.3 62.8 5.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qsyh8wbm",
   "source_url": "https://omlx.ai/benchmarks/performance/qsyh8wbm",
   "tg_tokens": null,
   "tps": 62.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-qsyh8wbm-ctx32768",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 655.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 655.0 49.9 6.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qsyh8wbm",
   "source_url": "https://omlx.ai/benchmarks/performance/qsyh8wbm",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-qsyh8wbm-ctx65536",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 8.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 534.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 534.5 36.3 8.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qsyh8wbm",
   "source_url": "https://omlx.ai/benchmarks/performance/qsyh8wbm",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-8w117y2g-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.7 GB at this context; duplicate traces of the same session: 1h65skb8, 5qvzekfa; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 8bit; 1k 127.6 12.3 10.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8w117y2g",
   "source_url": "https://omlx.ai/benchmarks/performance/8w117y2g",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-8w117y2g-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.0 GB at this context; duplicate traces of the same session: 1h65skb8, 5qvzekfa; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 132.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 8bit; 4k 132.0 12.0 11.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8w117y2g",
   "source_url": "https://omlx.ai/benchmarks/performance/8w117y2g",
   "tg_tokens": null,
   "tps": 12.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-8w117y2g-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.2 GB at this context; duplicate traces of the same session: 1h65skb8, 5qvzekfa; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 8bit; 8k 130.7 11.5 11.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8w117y2g",
   "source_url": "https://omlx.ai/benchmarks/performance/8w117y2g",
   "tg_tokens": null,
   "tps": 11.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-8w117y2g-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.7 GB at this context; duplicate traces of the same session: 1h65skb8, 5qvzekfa; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 8bit; 16k 127.5 10.6 11.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8w117y2g",
   "source_url": "https://omlx.ai/benchmarks/performance/8w117y2g",
   "tg_tokens": null,
   "tps": 10.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-8w117y2g-ctx32768",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: 1h65skb8, 5qvzekfa; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 90.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 8bit; 32k 90.4 8.6 12.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8w117y2g",
   "source_url": "https://omlx.ai/benchmarks/performance/8w117y2g",
   "tg_tokens": null,
   "tps": 8.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-mk5718cc-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: r7jnwmp2; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 727.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 1k 727.6 96.7 6.7 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mk5718cc",
   "source_url": "https://omlx.ai/benchmarks/performance/mk5718cc",
   "tg_tokens": null,
   "tps": 96.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-mk5718cc-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: r7jnwmp2; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 789.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 4k 789.6 93.2 7.0 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mk5718cc",
   "source_url": "https://omlx.ai/benchmarks/performance/mk5718cc",
   "tg_tokens": null,
   "tps": 93.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-mk5718cc-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.1 GB at this context; duplicate traces of the same session: r7jnwmp2; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 792.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 8k 792.6 87.6 7.1 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mk5718cc",
   "source_url": "https://omlx.ai/benchmarks/performance/mk5718cc",
   "tg_tokens": null,
   "tps": 87.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-mk5718cc-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.6 GB at this context; duplicate traces of the same session: r7jnwmp2; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 774.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 16k 774.4 76.6 7.6 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mk5718cc",
   "source_url": "https://omlx.ai/benchmarks/performance/mk5718cc",
   "tg_tokens": null,
   "tps": 76.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-mk5718cc-ctx32768",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.7 GB at this context; duplicate traces of the same session: r7jnwmp2; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 738.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 32k 738.1 61.9 8.7 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mk5718cc",
   "source_url": "https://omlx.ai/benchmarks/performance/mk5718cc",
   "tg_tokens": null,
   "tps": 61.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-mk5718cc-ctx65536",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.8 GB at this context; duplicate traces of the same session: r7jnwmp2; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 665.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 64k 665.7 44.3 10.8 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mk5718cc",
   "source_url": "https://omlx.ai/benchmarks/performance/mk5718cc",
   "tg_tokens": null,
   "tps": 44.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-jzkl9svu-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: 12gj2lle, 3jtaloiq; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 284.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 1k 284.7 47.3 5.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jzkl9svu",
   "source_url": "https://omlx.ai/benchmarks/performance/jzkl9svu",
   "tg_tokens": null,
   "tps": 47.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-jzkl9svu-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: 12gj2lle, 3jtaloiq; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 276.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k 276.2 45.9 6.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jzkl9svu",
   "source_url": "https://omlx.ai/benchmarks/performance/jzkl9svu",
   "tg_tokens": null,
   "tps": 45.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-jzkl9svu-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 12gj2lle, 3jtaloiq; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 241.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 16k 241.1 8.8 19.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jzkl9svu",
   "source_url": "https://omlx.ai/benchmarks/performance/jzkl9svu",
   "tg_tokens": null,
   "tps": 8.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-vfdt8szz-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: jnb7sytn, vrkofb71; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3006.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,006 105.5 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vfdt8szz",
   "source_url": "https://omlx.ai/benchmarks/performance/vfdt8szz",
   "tg_tokens": null,
   "tps": 105.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-vfdt8szz-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: jnb7sytn, vrkofb71; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3041.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 3,041 93.7 20.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vfdt8szz",
   "source_url": "https://omlx.ai/benchmarks/performance/vfdt8szz",
   "tg_tokens": null,
   "tps": 93.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-vfdt8szz-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: jnb7sytn, vrkofb71; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1350.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 128k 1,350 46.8 22.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vfdt8szz",
   "source_url": "https://omlx.ai/benchmarks/performance/vfdt8szz",
   "tg_tokens": null,
   "tps": 46.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-sekv52u6-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 18npewwm; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1121.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.6-35B-A3B-mxfp4 on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 1,121 64.5 19.6 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sekv52u6",
   "source_url": "https://omlx.ai/benchmarks/performance/sekv52u6",
   "tg_tokens": null,
   "tps": 64.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-sekv52u6-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: 18npewwm; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1589.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.6-35B-A3B-mxfp4 on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 1,589 87.7 20.3 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sekv52u6",
   "source_url": "https://omlx.ai/benchmarks/performance/sekv52u6",
   "tg_tokens": null,
   "tps": 87.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-htncesyo-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 1k context; peak memory 21.5 GB; TTFT 230 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4453.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 4,453 PP tok/s 238.4 TG tok/s 230 TTFT (ms) 21.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/htncesyo",
   "source_url": "https://omlx.ai/benchmarks/performance/htncesyo",
   "tg_tokens": null,
   "tps": 238.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-mdwzy570-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 32k context; peak memory 41.9 GB; TTFT 73322 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 446.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 32k tokens 446.9 PP tok/s 46.0 TG tok/s 73322 TTFT (ms) 41.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mdwzy570",
   "source_url": "https://omlx.ai/benchmarks/performance/mdwzy570",
   "tg_tokens": null,
   "tps": 46.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-lj35qhvt-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 79.5,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 79.5 13.0 23.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lj35qhvt",
   "source_url": "https://omlx.ai/benchmarks/performance/lj35qhvt",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-lj35qhvt-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 80.4,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 80.4 11.0 25.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lj35qhvt",
   "source_url": "https://omlx.ai/benchmarks/performance/lj35qhvt",
   "tg_tokens": null,
   "tps": 11.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-znf0r58s-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.2 GB at this context; duplicate traces of the same session: unhuhbqn; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.8,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 204.8 16.9 16.2 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/znf0r58s",
   "source_url": "https://omlx.ai/benchmarks/performance/znf0r58s",
   "tg_tokens": null,
   "tps": 16.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-znf0r58s-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: unhuhbqn; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 212.5,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 212.5 17.5 17.3 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/znf0r58s",
   "source_url": "https://omlx.ai/benchmarks/performance/znf0r58s",
   "tg_tokens": null,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-40o2d3uj-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1125.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,125 58.0 17.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/40o2d3uj",
   "source_url": "https://omlx.ai/benchmarks/performance/40o2d3uj",
   "tg_tokens": null,
   "tps": 58.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-40o2d3uj-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1336.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,336 93.0 19.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/40o2d3uj",
   "source_url": "https://omlx.ai/benchmarks/performance/40o2d3uj",
   "tg_tokens": null,
   "tps": 93.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-kn01zcvj-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 216.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 216.8 32.8 33.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kn01zcvj",
   "source_url": "https://omlx.ai/benchmarks/performance/kn01zcvj",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-kn01zcvj-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 218.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 218.1 27.9 37.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kn01zcvj",
   "source_url": "https://omlx.ai/benchmarks/performance/kn01zcvj",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-kn01zcvj-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 40.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 186.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 186.2 32.0 40.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kn01zcvj",
   "source_url": "https://omlx.ai/benchmarks/performance/kn01zcvj",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-chol024p-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.1 GB at this context; duplicate traces of the same session: p6c126fe, purne41h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 914.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 914.8 39.5 35.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/chol024p",
   "source_url": "https://omlx.ai/benchmarks/performance/chol024p",
   "tg_tokens": null,
   "tps": 39.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-chol024p-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: p6c126fe, purne41h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 923.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 923.9 26.6 35.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/chol024p",
   "source_url": "https://omlx.ai/benchmarks/performance/chol024p",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-chol024p-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: p6c126fe, purne41h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 901.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 901.2 28.2 35.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/chol024p",
   "source_url": "https://omlx.ai/benchmarks/performance/chol024p",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-chol024p-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: p6c126fe, purne41h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 844.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 844.4 25.9 36.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/chol024p",
   "source_url": "https://omlx.ai/benchmarks/performance/chol024p",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-alt2dmzj-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.1 GB at this context; duplicate traces of the same session: gy0lzrga, u0x3kuj4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 916.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 916.4 38.1 35.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/alt2dmzj",
   "source_url": "https://omlx.ai/benchmarks/performance/alt2dmzj",
   "tg_tokens": null,
   "tps": 38.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-alt2dmzj-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: gy0lzrga, u0x3kuj4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 923.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 923.8 27.2 35.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/alt2dmzj",
   "source_url": "https://omlx.ai/benchmarks/performance/alt2dmzj",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-alt2dmzj-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: gy0lzrga, u0x3kuj4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 901.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 901.5 27.4 35.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/alt2dmzj",
   "source_url": "https://omlx.ai/benchmarks/performance/alt2dmzj",
   "tg_tokens": null,
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-alt2dmzj-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: gy0lzrga, u0x3kuj4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 844.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 844.0 28.5 36.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/alt2dmzj",
   "source_url": "https://omlx.ai/benchmarks/performance/alt2dmzj",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3dtdicn0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.1 GB at this context; duplicate traces of the same session: jkowgmac, x342yu55; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 914.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 914.1 29.3 35.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3dtdicn0",
   "source_url": "https://omlx.ai/benchmarks/performance/3dtdicn0",
   "tg_tokens": null,
   "tps": 29.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3dtdicn0-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: jkowgmac, x342yu55; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 922.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 922.8 25.9 35.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3dtdicn0",
   "source_url": "https://omlx.ai/benchmarks/performance/3dtdicn0",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3dtdicn0-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: jkowgmac, x342yu55; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 901.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 901.1 36.3 35.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3dtdicn0",
   "source_url": "https://omlx.ai/benchmarks/performance/3dtdicn0",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3dtdicn0-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: jkowgmac, x342yu55; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 845.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 845.1 28.4 36.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3dtdicn0",
   "source_url": "https://omlx.ai/benchmarks/performance/3dtdicn0",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-tggxugfz-ctx1024",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: 5iyg0ao2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 842.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ6e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 842.1 88.0 29.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tggxugfz",
   "source_url": "https://omlx.ai/benchmarks/performance/tggxugfz",
   "tg_tokens": null,
   "tps": 88.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-tggxugfz-ctx4096",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.9 GB at this context; duplicate traces of the same session: 5iyg0ao2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 989.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ6e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 989.3 70.8 31.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tggxugfz",
   "source_url": "https://omlx.ai/benchmarks/performance/tggxugfz",
   "tg_tokens": null,
   "tps": 70.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-tggxugfz-ctx8192",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.2 GB at this context; duplicate traces of the same session: 5iyg0ao2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 866.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ6e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 8k 866.7 94.2 32.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tggxugfz",
   "source_url": "https://omlx.ai/benchmarks/performance/tggxugfz",
   "tg_tokens": null,
   "tps": 94.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-tggxugfz-ctx16384",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.7 GB at this context; duplicate traces of the same session: 5iyg0ao2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 711.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ6e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 16k 711.5 81.4 32.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tggxugfz",
   "source_url": "https://omlx.ai/benchmarks/performance/tggxugfz",
   "tg_tokens": null,
   "tps": 81.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-jelruthx-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.7 GB; TTFT 2422 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 422.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 422.9 PP tok/s 39.7 TG tok/s 2422 TTFT (ms) 70.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jelruthx",
   "source_url": "https://omlx.ai/benchmarks/performance/jelruthx",
   "tg_tokens": null,
   "tps": 39.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-s79lihtn-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 931.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 931.0 73.4 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s79lihtn",
   "source_url": "https://omlx.ai/benchmarks/performance/s79lihtn",
   "tg_tokens": null,
   "tps": 73.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-s79lihtn-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1701.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,701 66.0 70.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s79lihtn",
   "source_url": "https://omlx.ai/benchmarks/performance/s79lihtn",
   "tg_tokens": null,
   "tps": 66.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vmupds27-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 87.1 GB at this context; duplicate traces of the same session: iiqlm9iz, sn7zinna; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1439.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 1,439 63.2 87.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vmupds27",
   "source_url": "https://omlx.ai/benchmarks/performance/vmupds27",
   "tg_tokens": null,
   "tps": 63.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vmupds27-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 87.8 GB at this context; duplicate traces of the same session: iiqlm9iz, sn7zinna; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2078.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 2,078 66.1 87.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vmupds27",
   "source_url": "https://omlx.ai/benchmarks/performance/vmupds27",
   "tg_tokens": null,
   "tps": 66.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vmupds27-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 88.9 GB at this context; duplicate traces of the same session: iiqlm9iz, sn7zinna; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2444.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 2,444 60.4 88.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vmupds27",
   "source_url": "https://omlx.ai/benchmarks/performance/vmupds27",
   "tg_tokens": null,
   "tps": 60.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vmupds27-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 89.5 GB at this context; duplicate traces of the same session: iiqlm9iz, sn7zinna; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2456.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 2,456 60.3 89.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vmupds27",
   "source_url": "https://omlx.ai/benchmarks/performance/vmupds27",
   "tg_tokens": null,
   "tps": 60.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cnkqwckr-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 87.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1397.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 1,397 61.8 87.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cnkqwckr",
   "source_url": "https://omlx.ai/benchmarks/performance/cnkqwckr",
   "tg_tokens": null,
   "tps": 61.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cnkqwckr-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 87.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2059.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 2,059 56.5 87.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cnkqwckr",
   "source_url": "https://omlx.ai/benchmarks/performance/cnkqwckr",
   "tg_tokens": null,
   "tps": 56.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-j8hi3657-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 87.3 GB at this context; duplicate traces of the same session: 5n3l3ofa, i4898hgl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1036.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 1,036 57.2 87.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j8hi3657",
   "source_url": "https://omlx.ai/benchmarks/performance/j8hi3657",
   "tg_tokens": null,
   "tps": 57.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-j8hi3657-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 87.8 GB at this context; duplicate traces of the same session: 5n3l3ofa, i4898hgl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1918.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 1,918 56.6 87.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j8hi3657",
   "source_url": "https://omlx.ai/benchmarks/performance/j8hi3657",
   "tg_tokens": null,
   "tps": 56.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-j8hi3657-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 88.9 GB at this context; duplicate traces of the same session: 5n3l3ofa, i4898hgl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2339.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 2,339 56.1 88.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j8hi3657",
   "source_url": "https://omlx.ai/benchmarks/performance/j8hi3657",
   "tg_tokens": null,
   "tps": 56.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-j8hi3657-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 89.5 GB at this context; duplicate traces of the same session: 5n3l3ofa, i4898hgl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2480.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 2,480 56.4 89.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j8hi3657",
   "source_url": "https://omlx.ai/benchmarks/performance/j8hi3657",
   "tg_tokens": null,
   "tps": 56.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-m4pizi09-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.1 GB at this context; duplicate traces of the same session: vx0t12lo; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 495.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 1k 495.1 52.2 14.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m4pizi09",
   "source_url": "https://omlx.ai/benchmarks/performance/m4pizi09",
   "tg_tokens": null,
   "tps": 52.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-m4pizi09-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: vx0t12lo; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 490.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k 490.7 50.2 14.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m4pizi09",
   "source_url": "https://omlx.ai/benchmarks/performance/m4pizi09",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-m4pizi09-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.9 GB at this context; duplicate traces of the same session: vx0t12lo; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 456.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 16k 456.8 42.1 14.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m4pizi09",
   "source_url": "https://omlx.ai/benchmarks/performance/m4pizi09",
   "tg_tokens": null,
   "tps": 42.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-4kjjp3pk-ctx1024",
   "model": "Gemma-4-31B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: jjvej6hr; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 66.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31b-it-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 1k 66.6 11.4 17.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4kjjp3pk",
   "source_url": "https://omlx.ai/benchmarks/performance/4kjjp3pk",
   "tg_tokens": null,
   "tps": 11.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-4kjjp3pk-ctx4096",
   "model": "Gemma-4-31B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: jjvej6hr; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 66.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31b-it-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k 66.6 11.1 18.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4kjjp3pk",
   "source_url": "https://omlx.ai/benchmarks/performance/4kjjp3pk",
   "tg_tokens": null,
   "tps": 11.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-4kjjp3pk-ctx16384",
   "model": "Gemma-4-31B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: jjvej6hr; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 63.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31b-it-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 16k 63.1 9.6 20.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4kjjp3pk",
   "source_url": "https://omlx.ai/benchmarks/performance/4kjjp3pk",
   "tg_tokens": null,
   "tps": 9.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-xkqka989-ctx1024",
   "model": "gemma4 E4B",
   "notes": "OptiQ build; oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: 0tho7xm7, q8nyy2rv; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 806.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-OptiQ-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 1k 806.1 52.2 6.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xkqka989",
   "source_url": "https://omlx.ai/benchmarks/performance/xkqka989",
   "tg_tokens": null,
   "tps": 52.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-xkqka989-ctx4096",
   "model": "gemma4 E4B",
   "notes": "OptiQ build; oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: 0tho7xm7, q8nyy2rv; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 801.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-OptiQ-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k 801.6 49.5 6.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xkqka989",
   "source_url": "https://omlx.ai/benchmarks/performance/xkqka989",
   "tg_tokens": null,
   "tps": 49.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-xkqka989-ctx16384",
   "model": "gemma4 E4B",
   "notes": "OptiQ build; oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: 0tho7xm7, q8nyy2rv; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 794.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-OptiQ-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 16k 794.8 42.1 7.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xkqka989",
   "source_url": "https://omlx.ai/benchmarks/performance/xkqka989",
   "tg_tokens": null,
   "tps": 42.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-xt9oj34l-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: 2j2o12e4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1039.0,
   "provenance": "community",
   "quant": "MXFP4-Q4LH",
   "quote": "gpt-oss-20b-MXFP4-Q4LH on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 1,039 85.8 11.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xt9oj34l",
   "source_url": "https://omlx.ai/benchmarks/performance/xt9oj34l",
   "tg_tokens": null,
   "tps": 85.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-xt9oj34l-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: 2j2o12e4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1072.0,
   "provenance": "community",
   "quant": "MXFP4-Q4LH",
   "quote": "gpt-oss-20b-MXFP4-Q4LH on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,072 88.7 11.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xt9oj34l",
   "source_url": "https://omlx.ai/benchmarks/performance/xt9oj34l",
   "tg_tokens": null,
   "tps": 88.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-iefz4g9i-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: y61nygak; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1017.0,
   "provenance": "community",
   "quant": "MXFP4-Q4LH",
   "quote": "gpt-oss-20b-MXFP4-Q4LH on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 1,017 105.0 11.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iefz4g9i",
   "source_url": "https://omlx.ai/benchmarks/performance/iefz4g9i",
   "tg_tokens": null,
   "tps": 105.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-iefz4g9i-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: y61nygak; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1155.0,
   "provenance": "community",
   "quant": "MXFP4-Q4LH",
   "quote": "gpt-oss-20b-MXFP4-Q4LH on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,155 94.7 11.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iefz4g9i",
   "source_url": "https://omlx.ai/benchmarks/performance/iefz4g9i",
   "tg_tokens": null,
   "tps": 94.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-52jkvrz0-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: qa1avrux; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 963.4,
   "provenance": "community",
   "quant": "MXFP4-Q4LH",
   "quote": "gpt-oss-20b-MXFP4-Q4LH on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 963.4 89.6 11.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/52jkvrz0",
   "source_url": "https://omlx.ai/benchmarks/performance/52jkvrz0",
   "tg_tokens": null,
   "tps": 89.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-52jkvrz0-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: qa1avrux; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1027.0,
   "provenance": "community",
   "quant": "MXFP4-Q4LH",
   "quote": "gpt-oss-20b-MXFP4-Q4LH on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,027 31.5 22.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/52jkvrz0",
   "source_url": "https://omlx.ai/benchmarks/performance/52jkvrz0",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-kpki5wim-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: qrevsbxl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 880.1,
   "provenance": "community",
   "quant": "MXFP4-Q4LH",
   "quote": "gpt-oss-20b-MXFP4-Q4LH on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 880.1 41.4 11.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kpki5wim",
   "source_url": "https://omlx.ai/benchmarks/performance/kpki5wim",
   "tg_tokens": null,
   "tps": 41.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-kpki5wim-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: qrevsbxl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 944.9,
   "provenance": "community",
   "quant": "MXFP4-Q4LH",
   "quote": "gpt-oss-20b-MXFP4-Q4LH on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 944.9 36.6 11.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kpki5wim",
   "source_url": "https://omlx.ai/benchmarks/performance/kpki5wim",
   "tg_tokens": null,
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-bmkl119v-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: tvxzkq44; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1013.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 1,013 81.3 11.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bmkl119v",
   "source_url": "https://omlx.ai/benchmarks/performance/bmkl119v",
   "tg_tokens": null,
   "tps": 81.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-bmkl119v-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: tvxzkq44; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1020.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,020 80.5 11.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bmkl119v",
   "source_url": "https://omlx.ai/benchmarks/performance/bmkl119v",
   "tg_tokens": null,
   "tps": 80.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-yvd7esg2-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: epqwsa2y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 897.2,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 897.2 84.0 11.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yvd7esg2",
   "source_url": "https://omlx.ai/benchmarks/performance/yvd7esg2",
   "tg_tokens": null,
   "tps": 84.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-yvd7esg2-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: epqwsa2y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1086.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,086 82.4 11.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yvd7esg2",
   "source_url": "https://omlx.ai/benchmarks/performance/yvd7esg2",
   "tg_tokens": null,
   "tps": 82.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-06qa1524-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: dvtsdiwg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 905.6,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 905.6 83.8 11.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/06qa1524",
   "source_url": "https://omlx.ai/benchmarks/performance/06qa1524",
   "tg_tokens": null,
   "tps": 83.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-06qa1524-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: dvtsdiwg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 989.2,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 989.2 84.9 11.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/06qa1524",
   "source_url": "https://omlx.ai/benchmarks/performance/06qa1524",
   "tg_tokens": null,
   "tps": 84.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-vkhby126-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: ckxz0tkj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 878.5,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 878.5 77.1 11.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vkhby126",
   "source_url": "https://omlx.ai/benchmarks/performance/vkhby126",
   "tg_tokens": null,
   "tps": 77.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-vkhby126-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: ckxz0tkj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 918.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 918.0 83.9 11.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vkhby126",
   "source_url": "https://omlx.ai/benchmarks/performance/vkhby126",
   "tg_tokens": null,
   "tps": 83.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-dz5od9jn-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: seqypyh0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 862.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 862.0 36.3 11.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dz5od9jn",
   "source_url": "https://omlx.ai/benchmarks/performance/dz5od9jn",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-dz5od9jn-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: seqypyh0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 878.3,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 878.3 38.4 11.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dz5od9jn",
   "source_url": "https://omlx.ai/benchmarks/performance/dz5od9jn",
   "tg_tokens": null,
   "tps": 38.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-5dh8u8jj-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 561.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 1k 561.8 65.9 19.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5dh8u8jj",
   "source_url": "https://omlx.ai/benchmarks/performance/5dh8u8jj",
   "tg_tokens": null,
   "tps": 65.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-5dh8u8jj-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 565.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k 565.2 66.2 20.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5dh8u8jj",
   "source_url": "https://omlx.ai/benchmarks/performance/5dh8u8jj",
   "tg_tokens": null,
   "tps": 66.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-5dh8u8jj-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 442.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 16k 442.6 54.5 20.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5dh8u8jj",
   "source_url": "https://omlx.ai/benchmarks/performance/5dh8u8jj",
   "tg_tokens": null,
   "tps": 54.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-zq5bhoyg-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "scottlowry MTP build; oMLX community benchmark trace; peak memory 30.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 702.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "models--scottlowry--Ornith-1.5-35B-A3B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 702.9 74.2 30.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zq5bhoyg",
   "source_url": "https://omlx.ai/benchmarks/performance/zq5bhoyg",
   "tg_tokens": null,
   "tps": 74.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-zq5bhoyg-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "scottlowry MTP build; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 788.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "models--scottlowry--Ornith-1.5-35B-A3B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 788.3 69.2 32.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zq5bhoyg",
   "source_url": "https://omlx.ai/benchmarks/performance/zq5bhoyg",
   "tg_tokens": null,
   "tps": 69.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-zq5bhoyg-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "scottlowry MTP build; oMLX community benchmark trace; peak memory 32.3 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 753.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "models--scottlowry--Ornith-1.5-35B-A3B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 6bit; 8k 753.7 56.7 32.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zq5bhoyg",
   "source_url": "https://omlx.ai/benchmarks/performance/zq5bhoyg",
   "tg_tokens": null,
   "tps": 56.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-jfwewk6h-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 1pe7rxgz, 8og5un2c, emwv4wld; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 105.9 15.5 18.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jfwewk6h",
   "source_url": "https://omlx.ai/benchmarks/performance/jfwewk6h",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-jfwewk6h-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 1pe7rxgz, 8og5un2c, emwv4wld; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 150.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 150.8 14.7 19.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jfwewk6h",
   "source_url": "https://omlx.ai/benchmarks/performance/jfwewk6h",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-jfwewk6h-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 1pe7rxgz, 8og5un2c, emwv4wld; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 115.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 115.0 14.5 19.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jfwewk6h",
   "source_url": "https://omlx.ai/benchmarks/performance/jfwewk6h",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-jfwewk6h-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 1pe7rxgz, 8og5un2c, emwv4wld; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 96.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 96.0 12.8 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jfwewk6h",
   "source_url": "https://omlx.ai/benchmarks/performance/jfwewk6h",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-qfmh3bdj-ctx16384",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.4 GB at this context; duplicate traces of the same session: 0jmj5tkk, ra3j8y09; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 449.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "DeepSeek-V4-Flash-0731-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 16k 449.8 37.1 156.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qfmh3bdj",
   "source_url": "https://omlx.ai/benchmarks/performance/qfmh3bdj",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-qfmh3bdj-ctx32768",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.5 GB at this context; duplicate traces of the same session: 0jmj5tkk, ra3j8y09; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 450.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "DeepSeek-V4-Flash-0731-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 32k 450.3 45.1 156.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qfmh3bdj",
   "source_url": "https://omlx.ai/benchmarks/performance/qfmh3bdj",
   "tg_tokens": null,
   "tps": 45.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-qfmh3bdj-ctx65536",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.8 GB at this context; duplicate traces of the same session: 0jmj5tkk, ra3j8y09; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 426.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "DeepSeek-V4-Flash-0731-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 64k 426.2 36.5 156.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qfmh3bdj",
   "source_url": "https://omlx.ai/benchmarks/performance/qfmh3bdj",
   "tg_tokens": null,
   "tps": 36.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-qfmh3bdj-ctx131072",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 157.1 GB at this context; duplicate traces of the same session: 0jmj5tkk, ra3j8y09; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 385.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "DeepSeek-V4-Flash-0731-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 128k 385.6 35.1 157.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qfmh3bdj",
   "source_url": "https://omlx.ai/benchmarks/performance/qfmh3bdj",
   "tg_tokens": null,
   "tps": 35.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-q8xco2dk-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.5 GB at this context; duplicate traces of the same session: e5o5pios, ewuuqdr0, tshe7lwb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 394.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 394.3 27.6 175.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q8xco2dk",
   "source_url": "https://omlx.ai/benchmarks/performance/q8xco2dk",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-q8xco2dk-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.6 GB at this context; duplicate traces of the same session: e5o5pios, ewuuqdr0, tshe7lwb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 402.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 402.8 26.5 175.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q8xco2dk",
   "source_url": "https://omlx.ai/benchmarks/performance/q8xco2dk",
   "tg_tokens": null,
   "tps": 26.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-q8xco2dk-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 176.0 GB at this context; duplicate traces of the same session: e5o5pios, ewuuqdr0, tshe7lwb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 397.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 397.0 25.0 176.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q8xco2dk",
   "source_url": "https://omlx.ai/benchmarks/performance/q8xco2dk",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-q8xco2dk-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 176.7 GB at this context; duplicate traces of the same session: e5o5pios, ewuuqdr0, tshe7lwb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 388.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 128k 388.8 21.5 176.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q8xco2dk",
   "source_url": "https://omlx.ai/benchmarks/performance/q8xco2dk",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ta8mbot1-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 221 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4634.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 4,634 PP tok/s 144.0 TG tok/s 221 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ta8mbot1",
   "source_url": "https://omlx.ai/benchmarks/performance/ta8mbot1",
   "tg_tokens": null,
   "tps": 144.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-xf1l2fhf-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.4 GB at this context; duplicate traces of the same session: gynfr246; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1712.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 1,712 53.3 95.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xf1l2fhf",
   "source_url": "https://omlx.ai/benchmarks/performance/xf1l2fhf",
   "tg_tokens": null,
   "tps": 53.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-xf1l2fhf-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.3 GB at this context; duplicate traces of the same session: gynfr246; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1731.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 1,731 50.1 95.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xf1l2fhf",
   "source_url": "https://omlx.ai/benchmarks/performance/xf1l2fhf",
   "tg_tokens": null,
   "tps": 50.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-xf1l2fhf-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.5 GB at this context; duplicate traces of the same session: gynfr246; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1651.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 1,651 48.8 95.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xf1l2fhf",
   "source_url": "https://omlx.ai/benchmarks/performance/xf1l2fhf",
   "tg_tokens": null,
   "tps": 48.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-xf1l2fhf-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.9 GB at this context; duplicate traces of the same session: gynfr246; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1484.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 1,484 46.1 95.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xf1l2fhf",
   "source_url": "https://omlx.ai/benchmarks/performance/xf1l2fhf",
   "tg_tokens": null,
   "tps": 46.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-e1jngcsh-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1717.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 1,717 53.1 95.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e1jngcsh",
   "source_url": "https://omlx.ai/benchmarks/performance/e1jngcsh",
   "tg_tokens": null,
   "tps": 53.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-e1jngcsh-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1725.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 1,725 51.2 95.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e1jngcsh",
   "source_url": "https://omlx.ai/benchmarks/performance/e1jngcsh",
   "tg_tokens": null,
   "tps": 51.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-e1jngcsh-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1646.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 1,646 46.9 95.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e1jngcsh",
   "source_url": "https://omlx.ai/benchmarks/performance/e1jngcsh",
   "tg_tokens": null,
   "tps": 46.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-e1jngcsh-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1476.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 1,476 45.6 95.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e1jngcsh",
   "source_url": "https://omlx.ai/benchmarks/performance/e1jngcsh",
   "tg_tokens": null,
   "tps": 45.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-u0bwi2wc-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.4 GB at this context; duplicate traces of the same session: bplfbxvz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1709.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 1,709 53.0 95.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u0bwi2wc",
   "source_url": "https://omlx.ai/benchmarks/performance/u0bwi2wc",
   "tg_tokens": null,
   "tps": 53.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-u0bwi2wc-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.3 GB at this context; duplicate traces of the same session: bplfbxvz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1727.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 1,727 49.7 95.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u0bwi2wc",
   "source_url": "https://omlx.ai/benchmarks/performance/u0bwi2wc",
   "tg_tokens": null,
   "tps": 49.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-u0bwi2wc-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.5 GB at this context; duplicate traces of the same session: bplfbxvz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1648.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 1,648 48.3 95.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u0bwi2wc",
   "source_url": "https://omlx.ai/benchmarks/performance/u0bwi2wc",
   "tg_tokens": null,
   "tps": 48.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-u0bwi2wc-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.9 GB at this context; duplicate traces of the same session: bplfbxvz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1481.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 1,481 44.9 95.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u0bwi2wc",
   "source_url": "https://omlx.ai/benchmarks/performance/u0bwi2wc",
   "tg_tokens": null,
   "tps": 44.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ommqe00g-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.9 GB at this context; duplicate traces of the same session: kibimees, qbz7e0al; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1666.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 1,666 58.6 100.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ommqe00g",
   "source_url": "https://omlx.ai/benchmarks/performance/ommqe00g",
   "tg_tokens": null,
   "tps": 58.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ommqe00g-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.8 GB at this context; duplicate traces of the same session: kibimees, qbz7e0al; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1657.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 1,657 64.9 100.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ommqe00g",
   "source_url": "https://omlx.ai/benchmarks/performance/ommqe00g",
   "tg_tokens": null,
   "tps": 64.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ommqe00g-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.0 GB at this context; duplicate traces of the same session: kibimees, qbz7e0al; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1574.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 1,574 60.5 101.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ommqe00g",
   "source_url": "https://omlx.ai/benchmarks/performance/ommqe00g",
   "tg_tokens": null,
   "tps": 60.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ommqe00g-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: kibimees, qbz7e0al; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1407.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 1,407 55.9 101.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ommqe00g",
   "source_url": "https://omlx.ai/benchmarks/performance/ommqe00g",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-p6oyzjro-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.9 GB at this context; duplicate traces of the same session: e4zqm2j0, wk6mv5n4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1672.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 1,672 63.6 100.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p6oyzjro",
   "source_url": "https://omlx.ai/benchmarks/performance/p6oyzjro",
   "tg_tokens": null,
   "tps": 63.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-p6oyzjro-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.8 GB at this context; duplicate traces of the same session: e4zqm2j0, wk6mv5n4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1654.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 1,654 61.0 100.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p6oyzjro",
   "source_url": "https://omlx.ai/benchmarks/performance/p6oyzjro",
   "tg_tokens": null,
   "tps": 61.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-p6oyzjro-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.0 GB at this context; duplicate traces of the same session: e4zqm2j0, wk6mv5n4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1573.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 1,573 60.5 101.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p6oyzjro",
   "source_url": "https://omlx.ai/benchmarks/performance/p6oyzjro",
   "tg_tokens": null,
   "tps": 60.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-p6oyzjro-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: e4zqm2j0, wk6mv5n4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1411.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 1,411 56.7 101.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p6oyzjro",
   "source_url": "https://omlx.ai/benchmarks/performance/p6oyzjro",
   "tg_tokens": null,
   "tps": 56.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hl6sujug-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.9 GB at this context; duplicate traces of the same session: 185q07hf, a3ok0fr2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1665.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 1,665 60.4 100.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hl6sujug",
   "source_url": "https://omlx.ai/benchmarks/performance/hl6sujug",
   "tg_tokens": null,
   "tps": 60.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hl6sujug-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.8 GB at this context; duplicate traces of the same session: 185q07hf, a3ok0fr2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1647.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 1,647 64.0 100.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hl6sujug",
   "source_url": "https://omlx.ai/benchmarks/performance/hl6sujug",
   "tg_tokens": null,
   "tps": 64.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hl6sujug-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.0 GB at this context; duplicate traces of the same session: 185q07hf, a3ok0fr2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1578.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 1,578 61.8 101.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hl6sujug",
   "source_url": "https://omlx.ai/benchmarks/performance/hl6sujug",
   "tg_tokens": null,
   "tps": 61.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hl6sujug-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: 185q07hf, a3ok0fr2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1411.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 1,411 52.2 101.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hl6sujug",
   "source_url": "https://omlx.ai/benchmarks/performance/hl6sujug",
   "tg_tokens": null,
   "tps": 52.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2njxgwlk-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.2 GB at this context; duplicate traces of the same session: no8zb9r7, nvisbf52; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1699.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 1,699 51.2 96.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2njxgwlk",
   "source_url": "https://omlx.ai/benchmarks/performance/2njxgwlk",
   "tg_tokens": null,
   "tps": 51.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2njxgwlk-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.1 GB at this context; duplicate traces of the same session: no8zb9r7, nvisbf52; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1701.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 1,701 48.9 96.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2njxgwlk",
   "source_url": "https://omlx.ai/benchmarks/performance/2njxgwlk",
   "tg_tokens": null,
   "tps": 48.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2njxgwlk-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.3 GB at this context; duplicate traces of the same session: no8zb9r7, nvisbf52; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1624.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 1,624 47.0 96.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2njxgwlk",
   "source_url": "https://omlx.ai/benchmarks/performance/2njxgwlk",
   "tg_tokens": null,
   "tps": 47.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2njxgwlk-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.7 GB at this context; duplicate traces of the same session: no8zb9r7, nvisbf52; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1463.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 1,463 44.5 96.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2njxgwlk",
   "source_url": "https://omlx.ai/benchmarks/performance/2njxgwlk",
   "tg_tokens": null,
   "tps": 44.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hl50em4x-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.2 GB at this context; duplicate traces of the same session: sr3hr72s, xbrgpmbj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1700.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 1,700 51.1 96.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hl50em4x",
   "source_url": "https://omlx.ai/benchmarks/performance/hl50em4x",
   "tg_tokens": null,
   "tps": 51.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hl50em4x-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.1 GB at this context; duplicate traces of the same session: sr3hr72s, xbrgpmbj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1694.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 1,694 48.3 96.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hl50em4x",
   "source_url": "https://omlx.ai/benchmarks/performance/hl50em4x",
   "tg_tokens": null,
   "tps": 48.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hl50em4x-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.3 GB at this context; duplicate traces of the same session: sr3hr72s, xbrgpmbj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1627.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 1,627 46.0 96.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hl50em4x",
   "source_url": "https://omlx.ai/benchmarks/performance/hl50em4x",
   "tg_tokens": null,
   "tps": 46.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hl50em4x-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.7 GB at this context; duplicate traces of the same session: sr3hr72s, xbrgpmbj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1457.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 1,457 44.3 96.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hl50em4x",
   "source_url": "https://omlx.ai/benchmarks/performance/hl50em4x",
   "tg_tokens": null,
   "tps": 44.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n92zvcay-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.2 GB at this context; duplicate traces of the same session: 3yn3nds4, 4w4icx5o, a4w8ebx1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1703.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 1,703 51.1 96.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n92zvcay",
   "source_url": "https://omlx.ai/benchmarks/performance/n92zvcay",
   "tg_tokens": null,
   "tps": 51.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n92zvcay-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.1 GB at this context; duplicate traces of the same session: 3yn3nds4, 4w4icx5o, a4w8ebx1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1703.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 1,703 48.6 96.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n92zvcay",
   "source_url": "https://omlx.ai/benchmarks/performance/n92zvcay",
   "tg_tokens": null,
   "tps": 48.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n92zvcay-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.3 GB at this context; duplicate traces of the same session: 3yn3nds4, 4w4icx5o, a4w8ebx1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1628.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 1,628 45.6 96.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n92zvcay",
   "source_url": "https://omlx.ai/benchmarks/performance/n92zvcay",
   "tg_tokens": null,
   "tps": 45.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n92zvcay-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.7 GB at this context; duplicate traces of the same session: 3yn3nds4, 4w4icx5o, a4w8ebx1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1470.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 1,470 44.6 96.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n92zvcay",
   "source_url": "https://omlx.ai/benchmarks/performance/n92zvcay",
   "tg_tokens": null,
   "tps": 44.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-wp91qzvf-ctx1024",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 82.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 1k 82.0 15.2 16.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wp91qzvf",
   "source_url": "https://omlx.ai/benchmarks/performance/wp91qzvf",
   "tg_tokens": null,
   "tps": 15.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-wp91qzvf-ctx4096",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 77.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k 77.3 14.8 17.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wp91qzvf",
   "source_url": "https://omlx.ai/benchmarks/performance/wp91qzvf",
   "tg_tokens": null,
   "tps": 14.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-wp91qzvf-ctx16384",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 69.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 16k 69.0 13.4 18.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wp91qzvf",
   "source_url": "https://omlx.ai/benchmarks/performance/wp91qzvf",
   "tg_tokens": null,
   "tps": 13.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-tir4w3vr-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 374.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 374.9 33.7 3.9 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tir4w3vr",
   "source_url": "https://omlx.ai/benchmarks/performance/tir4w3vr",
   "tg_tokens": null,
   "tps": 33.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-tir4w3vr-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 373.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 373.3 32.0 4.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tir4w3vr",
   "source_url": "https://omlx.ai/benchmarks/performance/tir4w3vr",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-tir4w3vr-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.6 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 364.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 8k 364.6 29.5 4.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tir4w3vr",
   "source_url": "https://omlx.ai/benchmarks/performance/tir4w3vr",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-tir4w3vr-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.8 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 340.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 16k 340.6 25.2 4.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tir4w3vr",
   "source_url": "https://omlx.ai/benchmarks/performance/tir4w3vr",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-tir4w3vr-ctx32768",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 299.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 32k 299.7 20.3 5.9 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tir4w3vr",
   "source_url": "https://omlx.ai/benchmarks/performance/tir4w3vr",
   "tg_tokens": null,
   "tps": 20.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-tir4w3vr-ctx65536",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 8.0 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 231.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 64k 231.3 13.6 8.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tir4w3vr",
   "source_url": "https://omlx.ai/benchmarks/performance/tir4w3vr",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-1h65skb8-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.7 GB at this context; duplicate traces of the same session: 5qvzekfa; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 8bit; 1k 127.6 12.3 10.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1h65skb8",
   "source_url": "https://omlx.ai/benchmarks/performance/1h65skb8",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-1h65skb8-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.0 GB at this context; duplicate traces of the same session: 5qvzekfa; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 132.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 8bit; 4k 132.0 12.0 11.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1h65skb8",
   "source_url": "https://omlx.ai/benchmarks/performance/1h65skb8",
   "tg_tokens": null,
   "tps": 12.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-1h65skb8-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.2 GB at this context; duplicate traces of the same session: 5qvzekfa; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 8bit; 8k 130.7 11.5 11.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1h65skb8",
   "source_url": "https://omlx.ai/benchmarks/performance/1h65skb8",
   "tg_tokens": null,
   "tps": 11.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-1h65skb8-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.7 GB at this context; duplicate traces of the same session: 5qvzekfa; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 8bit; 16k 127.5 10.6 11.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1h65skb8",
   "source_url": "https://omlx.ai/benchmarks/performance/1h65skb8",
   "tg_tokens": null,
   "tps": 10.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-1h65skb8-ctx32768",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: 5qvzekfa; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 90.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 8bit; 32k 90.4 8.6 12.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1h65skb8",
   "source_url": "https://omlx.ai/benchmarks/performance/1h65skb8",
   "tg_tokens": null,
   "tps": 8.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-r7jnwmp2-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 727.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 1k 727.6 96.7 6.7 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r7jnwmp2",
   "source_url": "https://omlx.ai/benchmarks/performance/r7jnwmp2",
   "tg_tokens": null,
   "tps": 96.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-r7jnwmp2-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 789.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 4k 789.6 93.2 7.0 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r7jnwmp2",
   "source_url": "https://omlx.ai/benchmarks/performance/r7jnwmp2",
   "tg_tokens": null,
   "tps": 93.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-r7jnwmp2-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.1 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 792.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 8k 792.6 87.6 7.1 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r7jnwmp2",
   "source_url": "https://omlx.ai/benchmarks/performance/r7jnwmp2",
   "tg_tokens": null,
   "tps": 87.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-r7jnwmp2-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.6 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 774.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 16k 774.4 76.6 7.6 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r7jnwmp2",
   "source_url": "https://omlx.ai/benchmarks/performance/r7jnwmp2",
   "tg_tokens": null,
   "tps": 76.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-r7jnwmp2-ctx32768",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.7 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 738.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 32k 738.1 61.9 8.7 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r7jnwmp2",
   "source_url": "https://omlx.ai/benchmarks/performance/r7jnwmp2",
   "tg_tokens": null,
   "tps": 61.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-r7jnwmp2-ctx65536",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.8 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 665.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 64k 665.7 44.3 10.8 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r7jnwmp2",
   "source_url": "https://omlx.ai/benchmarks/performance/r7jnwmp2",
   "tg_tokens": null,
   "tps": 44.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-12gj2lle-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: 3jtaloiq; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 284.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 1k 284.7 47.3 5.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/12gj2lle",
   "source_url": "https://omlx.ai/benchmarks/performance/12gj2lle",
   "tg_tokens": null,
   "tps": 47.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-12gj2lle-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: 3jtaloiq; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 276.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k 276.2 45.9 6.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/12gj2lle",
   "source_url": "https://omlx.ai/benchmarks/performance/12gj2lle",
   "tg_tokens": null,
   "tps": 45.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-12gj2lle-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 3jtaloiq; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 241.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 16k 241.1 8.8 19.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/12gj2lle",
   "source_url": "https://omlx.ai/benchmarks/performance/12gj2lle",
   "tg_tokens": null,
   "tps": 8.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-e9o2szi3-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: hkahu15r; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 837.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 16 GB \u00b7 4bit; 1k 837.7 26.4 5.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e9o2szi3",
   "source_url": "https://omlx.ai/benchmarks/performance/e9o2szi3",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-e9o2szi3-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: hkahu15r; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 885.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 16 GB \u00b7 4bit; 4k 885.3 25.5 6.2 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e9o2szi3",
   "source_url": "https://omlx.ai/benchmarks/performance/e9o2szi3",
   "tg_tokens": null,
   "tps": 25.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-vrkofb71-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: jnb7sytn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3006.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,006 105.5 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vrkofb71",
   "source_url": "https://omlx.ai/benchmarks/performance/vrkofb71",
   "tg_tokens": null,
   "tps": 105.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-vrkofb71-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: jnb7sytn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3041.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 3,041 93.7 20.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vrkofb71",
   "source_url": "https://omlx.ai/benchmarks/performance/vrkofb71",
   "tg_tokens": null,
   "tps": 93.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-vrkofb71-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: jnb7sytn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1350.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 128k 1,350 46.8 22.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vrkofb71",
   "source_url": "https://omlx.ai/benchmarks/performance/vrkofb71",
   "tg_tokens": null,
   "tps": 46.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-pvmolawo-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V13-dequantized",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.5 GB at this context; duplicate traces of the same session: l8wf6fi5; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1945.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V13-dequantized-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 1,945 149.6 37.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pvmolawo",
   "source_url": "https://omlx.ai/benchmarks/performance/pvmolawo",
   "tg_tokens": null,
   "tps": 149.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-pvmolawo-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V13-dequantized",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.1 GB at this context; duplicate traces of the same session: l8wf6fi5; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2304.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V13-dequantized-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 2,304 127.5 39.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pvmolawo",
   "source_url": "https://omlx.ai/benchmarks/performance/pvmolawo",
   "tg_tokens": null,
   "tps": 127.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-18npewwm-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1121.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.6-35B-A3B-mxfp4 on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 1,121 64.5 19.6 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/18npewwm",
   "source_url": "https://omlx.ai/benchmarks/performance/18npewwm",
   "tg_tokens": null,
   "tps": 64.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-18npewwm-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1589.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.6-35B-A3B-mxfp4 on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 1,589 87.7 20.3 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/18npewwm",
   "source_url": "https://omlx.ai/benchmarks/performance/18npewwm",
   "tg_tokens": null,
   "tps": 87.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mwh65nzm-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 4k context; peak memory 22.1 GB; TTFT 650 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6304.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 6,304 PP tok/s 187.4 TG tok/s 650 TTFT (ms) 22.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mwh65nzm",
   "source_url": "https://omlx.ai/benchmarks/performance/mwh65nzm",
   "tg_tokens": null,
   "tps": 187.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3hag6n8q-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 4k context; peak memory 22.1 GB; TTFT 655 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6250.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 6,250 PP tok/s 177.1 TG tok/s 655 TTFT (ms) 22.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3hag6n8q",
   "source_url": "https://omlx.ai/benchmarks/performance/3hag6n8q",
   "tg_tokens": null,
   "tps": 177.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-dc8ouhns-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: d9zfmiis; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 105.4 18.0 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dc8ouhns",
   "source_url": "https://omlx.ai/benchmarks/performance/dc8ouhns",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-dc8ouhns-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: d9zfmiis; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 111.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 111.3 19.2 22.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dc8ouhns",
   "source_url": "https://omlx.ai/benchmarks/performance/dc8ouhns",
   "tg_tokens": null,
   "tps": 19.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-idqxok6j-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: ilfux3gp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 465.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 465.8 18.2 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/idqxok6j",
   "source_url": "https://omlx.ai/benchmarks/performance/idqxok6j",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-idqxok6j-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: ilfux3gp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 459.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 459.5 17.8 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/idqxok6j",
   "source_url": "https://omlx.ai/benchmarks/performance/idqxok6j",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-unhuhbqn-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.2 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.8,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 204.8 16.9 16.2 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/unhuhbqn",
   "source_url": "https://omlx.ai/benchmarks/performance/unhuhbqn",
   "tg_tokens": null,
   "tps": 16.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-unhuhbqn-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 212.5,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 212.5 17.5 17.3 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/unhuhbqn",
   "source_url": "https://omlx.ai/benchmarks/performance/unhuhbqn",
   "tg_tokens": null,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-7vi9ot1b-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: j6pmudnh; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 154.4,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 154.4 19.6 16.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7vi9ot1b",
   "source_url": "https://omlx.ai/benchmarks/performance/7vi9ot1b",
   "tg_tokens": null,
   "tps": 19.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-7vi9ot1b-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: j6pmudnh; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 163.4,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 163.4 18.8 20.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7vi9ot1b",
   "source_url": "https://omlx.ai/benchmarks/performance/7vi9ot1b",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-ynhx5vgx-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 65.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k 65.3 6.5 16.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ynhx5vgx",
   "source_url": "https://omlx.ai/benchmarks/performance/ynhx5vgx",
   "tg_tokens": null,
   "tps": 6.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-ynhx5vgx-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 66.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 66.7 6.3 17.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ynhx5vgx",
   "source_url": "https://omlx.ai/benchmarks/performance/ynhx5vgx",
   "tg_tokens": null,
   "tps": 6.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gy0lzrga-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.1 GB at this context; duplicate traces of the same session: u0x3kuj4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 916.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 916.4 38.1 35.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gy0lzrga",
   "source_url": "https://omlx.ai/benchmarks/performance/gy0lzrga",
   "tg_tokens": null,
   "tps": 38.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gy0lzrga-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: u0x3kuj4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 923.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 923.8 27.2 35.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gy0lzrga",
   "source_url": "https://omlx.ai/benchmarks/performance/gy0lzrga",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gy0lzrga-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: u0x3kuj4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 901.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 901.5 27.4 35.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gy0lzrga",
   "source_url": "https://omlx.ai/benchmarks/performance/gy0lzrga",
   "tg_tokens": null,
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gy0lzrga-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: u0x3kuj4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 844.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 844.0 28.5 36.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gy0lzrga",
   "source_url": "https://omlx.ai/benchmarks/performance/gy0lzrga",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jkowgmac-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.1 GB at this context; duplicate traces of the same session: x342yu55; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 914.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 914.1 29.3 35.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jkowgmac",
   "source_url": "https://omlx.ai/benchmarks/performance/jkowgmac",
   "tg_tokens": null,
   "tps": 29.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jkowgmac-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: x342yu55; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 922.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 922.8 25.9 35.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jkowgmac",
   "source_url": "https://omlx.ai/benchmarks/performance/jkowgmac",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jkowgmac-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: x342yu55; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 901.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 901.1 36.3 35.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jkowgmac",
   "source_url": "https://omlx.ai/benchmarks/performance/jkowgmac",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jkowgmac-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: x342yu55; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 845.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 845.1 28.4 36.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jkowgmac",
   "source_url": "https://omlx.ai/benchmarks/performance/jkowgmac",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-p6c126fe-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.1 GB at this context; duplicate traces of the same session: purne41h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 914.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 914.8 39.5 35.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p6c126fe",
   "source_url": "https://omlx.ai/benchmarks/performance/p6c126fe",
   "tg_tokens": null,
   "tps": 39.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-p6c126fe-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: purne41h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 923.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 923.9 26.6 35.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p6c126fe",
   "source_url": "https://omlx.ai/benchmarks/performance/p6c126fe",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-p6c126fe-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: purne41h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 901.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 901.2 28.2 35.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p6c126fe",
   "source_url": "https://omlx.ai/benchmarks/performance/p6c126fe",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-p6c126fe-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: purne41h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 844.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 844.4 25.9 36.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p6c126fe",
   "source_url": "https://omlx.ai/benchmarks/performance/p6c126fe",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-5iyg0ao2-ctx1024",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 842.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ6e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 842.1 88.0 29.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5iyg0ao2",
   "source_url": "https://omlx.ai/benchmarks/performance/5iyg0ao2",
   "tg_tokens": null,
   "tps": 88.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-5iyg0ao2-ctx4096",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.9 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 989.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ6e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 989.3 70.8 31.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5iyg0ao2",
   "source_url": "https://omlx.ai/benchmarks/performance/5iyg0ao2",
   "tg_tokens": null,
   "tps": 70.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-5iyg0ao2-ctx8192",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.2 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 866.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ6e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 8k 866.7 94.2 32.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5iyg0ao2",
   "source_url": "https://omlx.ai/benchmarks/performance/5iyg0ao2",
   "tg_tokens": null,
   "tps": 94.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-5iyg0ao2-ctx16384",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.7 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 711.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ6e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 16k 711.5 81.4 32.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5iyg0ao2",
   "source_url": "https://omlx.ai/benchmarks/performance/5iyg0ao2",
   "tg_tokens": null,
   "tps": 81.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3isid3jp-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 108.5 GB at this context; duplicate traces of the same session: 53o4qubr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 865.9,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ4-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 865.9 47.9 108.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3isid3jp",
   "source_url": "https://omlx.ai/benchmarks/performance/3isid3jp",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3isid3jp-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 109.3 GB at this context; duplicate traces of the same session: 53o4qubr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 969.4,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ4-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 969.4 45.3 109.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3isid3jp",
   "source_url": "https://omlx.ai/benchmarks/performance/3isid3jp",
   "tg_tokens": null,
   "tps": 45.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-29z0ph1u-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: ko47l5tz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 335.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 335.2 78.4 70.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/29z0ph1u",
   "source_url": "https://omlx.ai/benchmarks/performance/29z0ph1u",
   "tg_tokens": null,
   "tps": 78.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-29z0ph1u-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: ko47l5tz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 932.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 932.3 51.2 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/29z0ph1u",
   "source_url": "https://omlx.ai/benchmarks/performance/29z0ph1u",
   "tg_tokens": null,
   "tps": 51.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-sn7zinna-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 87.1 GB at this context; duplicate traces of the same session: iiqlm9iz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1439.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 1,439 63.2 87.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sn7zinna",
   "source_url": "https://omlx.ai/benchmarks/performance/sn7zinna",
   "tg_tokens": null,
   "tps": 63.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-sn7zinna-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 87.8 GB at this context; duplicate traces of the same session: iiqlm9iz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2078.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 2,078 66.1 87.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sn7zinna",
   "source_url": "https://omlx.ai/benchmarks/performance/sn7zinna",
   "tg_tokens": null,
   "tps": 66.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-sn7zinna-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 88.9 GB at this context; duplicate traces of the same session: iiqlm9iz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2444.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 2,444 60.4 88.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sn7zinna",
   "source_url": "https://omlx.ai/benchmarks/performance/sn7zinna",
   "tg_tokens": null,
   "tps": 60.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-sn7zinna-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 89.5 GB at this context; duplicate traces of the same session: iiqlm9iz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2456.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 2,456 60.3 89.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sn7zinna",
   "source_url": "https://omlx.ai/benchmarks/performance/sn7zinna",
   "tg_tokens": null,
   "tps": 60.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i4898hgl-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 87.3 GB at this context; duplicate traces of the same session: 5n3l3ofa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1036.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 1,036 57.2 87.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i4898hgl",
   "source_url": "https://omlx.ai/benchmarks/performance/i4898hgl",
   "tg_tokens": null,
   "tps": 57.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i4898hgl-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 87.8 GB at this context; duplicate traces of the same session: 5n3l3ofa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1918.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 1,918 56.6 87.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i4898hgl",
   "source_url": "https://omlx.ai/benchmarks/performance/i4898hgl",
   "tg_tokens": null,
   "tps": 56.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i4898hgl-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 88.9 GB at this context; duplicate traces of the same session: 5n3l3ofa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2339.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 2,339 56.1 88.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i4898hgl",
   "source_url": "https://omlx.ai/benchmarks/performance/i4898hgl",
   "tg_tokens": null,
   "tps": 56.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i4898hgl-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 89.5 GB at this context; duplicate traces of the same session: 5n3l3ofa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2480.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 2,480 56.4 89.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i4898hgl",
   "source_url": "https://omlx.ai/benchmarks/performance/i4898hgl",
   "tg_tokens": null,
   "tps": 56.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-h5nkf3st-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace, single run at 8k context; peak memory 143.4 GB; TTFT 8019 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1022.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 8k tokens 1,022 PP tok/s 75.2 TG tok/s 8019 TTFT (ms) 143.4 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h5nkf3st",
   "source_url": "https://omlx.ai/benchmarks/performance/h5nkf3st",
   "tg_tokens": null,
   "tps": 75.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ldh79fgi-ctx8192",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "drowzeys-keys-Mac abliterated build; oMLX community benchmark trace; peak memory 157.5 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 466.1,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "drowzeys-keys-Mac-DeepSeek-V4-Flash-0731-MXFP4-MLX-Abliterated on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 466.1 44.5 157.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ldh79fgi",
   "source_url": "https://omlx.ai/benchmarks/performance/ldh79fgi",
   "tg_tokens": null,
   "tps": 44.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ldh79fgi-ctx32768",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "drowzeys-keys-Mac abliterated build; oMLX community benchmark trace; peak memory 161.9 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 445.1,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "drowzeys-keys-Mac-DeepSeek-V4-Flash-0731-MXFP4-MLX-Abliterated on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 445.1 45.2 161.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ldh79fgi",
   "source_url": "https://omlx.ai/benchmarks/performance/ldh79fgi",
   "tg_tokens": null,
   "tps": 45.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-vx0t12lo-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.1 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 495.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 1k 495.1 52.2 14.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vx0t12lo",
   "source_url": "https://omlx.ai/benchmarks/performance/vx0t12lo",
   "tg_tokens": null,
   "tps": 52.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-vx0t12lo-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 490.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k 490.7 50.2 14.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vx0t12lo",
   "source_url": "https://omlx.ai/benchmarks/performance/vx0t12lo",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-vx0t12lo-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.9 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 456.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 16k 456.8 42.1 14.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vx0t12lo",
   "source_url": "https://omlx.ai/benchmarks/performance/vx0t12lo",
   "tg_tokens": null,
   "tps": 42.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-jjvej6hr-ctx1024",
   "model": "Gemma-4-31B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 66.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31b-it-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 1k 66.6 11.4 17.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jjvej6hr",
   "source_url": "https://omlx.ai/benchmarks/performance/jjvej6hr",
   "tg_tokens": null,
   "tps": 11.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-jjvej6hr-ctx4096",
   "model": "Gemma-4-31B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 66.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31b-it-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k 66.6 11.1 18.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jjvej6hr",
   "source_url": "https://omlx.ai/benchmarks/performance/jjvej6hr",
   "tg_tokens": null,
   "tps": 11.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-jjvej6hr-ctx16384",
   "model": "Gemma-4-31B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 63.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31b-it-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 16k 63.1 9.6 20.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jjvej6hr",
   "source_url": "https://omlx.ai/benchmarks/performance/jjvej6hr",
   "tg_tokens": null,
   "tps": 9.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-q8nyy2rv-ctx1024",
   "model": "gemma4 E4B",
   "notes": "OptiQ build; oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: 0tho7xm7; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 806.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-OptiQ-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 1k 806.1 52.2 6.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q8nyy2rv",
   "source_url": "https://omlx.ai/benchmarks/performance/q8nyy2rv",
   "tg_tokens": null,
   "tps": 52.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-q8nyy2rv-ctx4096",
   "model": "gemma4 E4B",
   "notes": "OptiQ build; oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: 0tho7xm7; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 801.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-OptiQ-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k 801.6 49.5 6.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q8nyy2rv",
   "source_url": "https://omlx.ai/benchmarks/performance/q8nyy2rv",
   "tg_tokens": null,
   "tps": 49.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-q8nyy2rv-ctx16384",
   "model": "gemma4 E4B",
   "notes": "OptiQ build; oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: 0tho7xm7; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 794.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-OptiQ-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 16k 794.8 42.1 7.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q8nyy2rv",
   "source_url": "https://omlx.ai/benchmarks/performance/q8nyy2rv",
   "tg_tokens": null,
   "tps": 42.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-2j2o12e4-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1039.0,
   "provenance": "community",
   "quant": "MXFP4-Q4LH",
   "quote": "gpt-oss-20b-MXFP4-Q4LH on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 1,039 85.8 11.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2j2o12e4",
   "source_url": "https://omlx.ai/benchmarks/performance/2j2o12e4",
   "tg_tokens": null,
   "tps": 85.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-2j2o12e4-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1072.0,
   "provenance": "community",
   "quant": "MXFP4-Q4LH",
   "quote": "gpt-oss-20b-MXFP4-Q4LH on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,072 88.7 11.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2j2o12e4",
   "source_url": "https://omlx.ai/benchmarks/performance/2j2o12e4",
   "tg_tokens": null,
   "tps": 88.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-qa1avrux-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 963.4,
   "provenance": "community",
   "quant": "MXFP4-Q4LH",
   "quote": "gpt-oss-20b-MXFP4-Q4LH on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 963.4 89.6 11.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qa1avrux",
   "source_url": "https://omlx.ai/benchmarks/performance/qa1avrux",
   "tg_tokens": null,
   "tps": 89.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-qa1avrux-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1027.0,
   "provenance": "community",
   "quant": "MXFP4-Q4LH",
   "quote": "gpt-oss-20b-MXFP4-Q4LH on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,027 31.5 22.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qa1avrux",
   "source_url": "https://omlx.ai/benchmarks/performance/qa1avrux",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-y61nygak-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1017.0,
   "provenance": "community",
   "quant": "MXFP4-Q4LH",
   "quote": "gpt-oss-20b-MXFP4-Q4LH on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 1,017 105.0 11.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y61nygak",
   "source_url": "https://omlx.ai/benchmarks/performance/y61nygak",
   "tg_tokens": null,
   "tps": 105.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-y61nygak-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1155.0,
   "provenance": "community",
   "quant": "MXFP4-Q4LH",
   "quote": "gpt-oss-20b-MXFP4-Q4LH on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,155 94.7 11.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y61nygak",
   "source_url": "https://omlx.ai/benchmarks/performance/y61nygak",
   "tg_tokens": null,
   "tps": 94.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-qrevsbxl-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 880.1,
   "provenance": "community",
   "quant": "MXFP4-Q4LH",
   "quote": "gpt-oss-20b-MXFP4-Q4LH on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 880.1 41.4 11.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qrevsbxl",
   "source_url": "https://omlx.ai/benchmarks/performance/qrevsbxl",
   "tg_tokens": null,
   "tps": 41.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-qrevsbxl-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 944.9,
   "provenance": "community",
   "quant": "MXFP4-Q4LH",
   "quote": "gpt-oss-20b-MXFP4-Q4LH on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 944.9 36.6 11.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qrevsbxl",
   "source_url": "https://omlx.ai/benchmarks/performance/qrevsbxl",
   "tg_tokens": null,
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-xed9xp30-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: hmdfh1ev; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1058.0,
   "provenance": "community",
   "quant": "MXFP4-Q4LH",
   "quote": "gpt-oss-20b-MXFP4-Q4LH on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 1,058 89.6 11.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xed9xp30",
   "source_url": "https://omlx.ai/benchmarks/performance/xed9xp30",
   "tg_tokens": null,
   "tps": 89.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-xed9xp30-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: hmdfh1ev; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1072.0,
   "provenance": "community",
   "quant": "MXFP4-Q4LH",
   "quote": "gpt-oss-20b-MXFP4-Q4LH on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,072 85.9 11.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xed9xp30",
   "source_url": "https://omlx.ai/benchmarks/performance/xed9xp30",
   "tg_tokens": null,
   "tps": 85.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-aas75jsz-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: k40y4pv6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 975.0,
   "provenance": "community",
   "quant": "MXFP4-Q4LH",
   "quote": "gpt-oss-20b-MXFP4-Q4LH on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 975.0 89.7 11.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aas75jsz",
   "source_url": "https://omlx.ai/benchmarks/performance/aas75jsz",
   "tg_tokens": null,
   "tps": 89.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-aas75jsz-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: k40y4pv6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 965.9,
   "provenance": "community",
   "quant": "MXFP4-Q4LH",
   "quote": "gpt-oss-20b-MXFP4-Q4LH on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 965.9 84.7 11.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aas75jsz",
   "source_url": "https://omlx.ai/benchmarks/performance/aas75jsz",
   "tg_tokens": null,
   "tps": 84.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-p120ji35-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: k1wdrw76; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 839.9,
   "provenance": "community",
   "quant": "MXFP4-Q4LH",
   "quote": "gpt-oss-20b-MXFP4-Q4LH on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 839.9 94.1 11.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p120ji35",
   "source_url": "https://omlx.ai/benchmarks/performance/p120ji35",
   "tg_tokens": null,
   "tps": 94.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-p120ji35-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: k1wdrw76; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1025.0,
   "provenance": "community",
   "quant": "MXFP4-Q4LH",
   "quote": "gpt-oss-20b-MXFP4-Q4LH on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,025 83.1 11.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p120ji35",
   "source_url": "https://omlx.ai/benchmarks/performance/p120ji35",
   "tg_tokens": null,
   "tps": 83.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-5mwrlhs0-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: jq17kvsd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1368.0,
   "provenance": "community",
   "quant": "MXFP4-Q4LH",
   "quote": "gpt-oss-20b-MXFP4-Q4LH on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 1,368 108.1 11.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5mwrlhs0",
   "source_url": "https://omlx.ai/benchmarks/performance/5mwrlhs0",
   "tg_tokens": null,
   "tps": 108.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-5mwrlhs0-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: jq17kvsd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1351.0,
   "provenance": "community",
   "quant": "MXFP4-Q4LH",
   "quote": "gpt-oss-20b-MXFP4-Q4LH on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,351 103.3 11.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5mwrlhs0",
   "source_url": "https://omlx.ai/benchmarks/performance/5mwrlhs0",
   "tg_tokens": null,
   "tps": 103.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-lylprg58-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: y0o6ij5s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1361.0,
   "provenance": "community",
   "quant": "MXFP4-Q4LH",
   "quote": "gpt-oss-20b-MXFP4-Q4LH on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 1,361 107.3 11.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lylprg58",
   "source_url": "https://omlx.ai/benchmarks/performance/lylprg58",
   "tg_tokens": null,
   "tps": 107.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-lylprg58-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: y0o6ij5s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1348.0,
   "provenance": "community",
   "quant": "MXFP4-Q4LH",
   "quote": "gpt-oss-20b-MXFP4-Q4LH on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,348 103.5 11.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lylprg58",
   "source_url": "https://omlx.ai/benchmarks/performance/lylprg58",
   "tg_tokens": null,
   "tps": 103.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ustyy3ni-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: 8xidrlhm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1368.0,
   "provenance": "community",
   "quant": "MXFP4-Q4LH",
   "quote": "gpt-oss-20b-MXFP4-Q4LH on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 1,368 107.0 11.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ustyy3ni",
   "source_url": "https://omlx.ai/benchmarks/performance/ustyy3ni",
   "tg_tokens": null,
   "tps": 107.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ustyy3ni-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: 8xidrlhm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1374.0,
   "provenance": "community",
   "quant": "MXFP4-Q4LH",
   "quote": "gpt-oss-20b-MXFP4-Q4LH on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,374 102.7 11.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ustyy3ni",
   "source_url": "https://omlx.ai/benchmarks/performance/ustyy3ni",
   "tg_tokens": null,
   "tps": 102.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-tvxzkq44-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1013.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 1,013 81.3 11.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tvxzkq44",
   "source_url": "https://omlx.ai/benchmarks/performance/tvxzkq44",
   "tg_tokens": null,
   "tps": 81.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-tvxzkq44-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1020.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,020 80.5 11.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tvxzkq44",
   "source_url": "https://omlx.ai/benchmarks/performance/tvxzkq44",
   "tg_tokens": null,
   "tps": 80.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-epqwsa2y-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 897.2,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 897.2 84.0 11.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/epqwsa2y",
   "source_url": "https://omlx.ai/benchmarks/performance/epqwsa2y",
   "tg_tokens": null,
   "tps": 84.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-epqwsa2y-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1086.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,086 82.4 11.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/epqwsa2y",
   "source_url": "https://omlx.ai/benchmarks/performance/epqwsa2y",
   "tg_tokens": null,
   "tps": 82.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-dvtsdiwg-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 905.6,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 905.6 83.8 11.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dvtsdiwg",
   "source_url": "https://omlx.ai/benchmarks/performance/dvtsdiwg",
   "tg_tokens": null,
   "tps": 83.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-dvtsdiwg-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 989.2,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 989.2 84.9 11.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dvtsdiwg",
   "source_url": "https://omlx.ai/benchmarks/performance/dvtsdiwg",
   "tg_tokens": null,
   "tps": 84.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-ckxz0tkj-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 878.5,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 878.5 77.1 11.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ckxz0tkj",
   "source_url": "https://omlx.ai/benchmarks/performance/ckxz0tkj",
   "tg_tokens": null,
   "tps": 77.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-ckxz0tkj-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 918.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 918.0 83.9 11.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ckxz0tkj",
   "source_url": "https://omlx.ai/benchmarks/performance/ckxz0tkj",
   "tg_tokens": null,
   "tps": 83.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-seqypyh0-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 862.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 862.0 36.3 11.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/seqypyh0",
   "source_url": "https://omlx.ai/benchmarks/performance/seqypyh0",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-seqypyh0-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 878.3,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 878.3 38.4 11.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/seqypyh0",
   "source_url": "https://omlx.ai/benchmarks/performance/seqypyh0",
   "tg_tokens": null,
   "tps": 38.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-oh5bqj5o-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: t0454j17; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 996.6,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 996.6 86.4 11.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oh5bqj5o",
   "source_url": "https://omlx.ai/benchmarks/performance/oh5bqj5o",
   "tg_tokens": null,
   "tps": 86.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-oh5bqj5o-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: t0454j17; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1047.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,047 78.3 11.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oh5bqj5o",
   "source_url": "https://omlx.ai/benchmarks/performance/oh5bqj5o",
   "tg_tokens": null,
   "tps": 78.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-p7l7dwb5-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: pmo45t7e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 906.8,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 906.8 87.0 11.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p7l7dwb5",
   "source_url": "https://omlx.ai/benchmarks/performance/p7l7dwb5",
   "tg_tokens": null,
   "tps": 87.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-p7l7dwb5-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: pmo45t7e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 974.5,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 974.5 74.4 11.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p7l7dwb5",
   "source_url": "https://omlx.ai/benchmarks/performance/p7l7dwb5",
   "tg_tokens": null,
   "tps": 74.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-w1fj3mb2-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.7 GB at this context; duplicate traces of the same session: 58ug8aa2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 454.5,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 454.5 48.0 11.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w1fj3mb2",
   "source_url": "https://omlx.ai/benchmarks/performance/w1fj3mb2",
   "tg_tokens": null,
   "tps": 48.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-w1fj3mb2-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: 58ug8aa2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 354.6,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 354.6 32.0 11.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w1fj3mb2",
   "source_url": "https://omlx.ai/benchmarks/performance/w1fj3mb2",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-cix1kf9o-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: j3mlbacb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1373.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 1,373 69.8 11.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cix1kf9o",
   "source_url": "https://omlx.ai/benchmarks/performance/cix1kf9o",
   "tg_tokens": null,
   "tps": 69.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-cix1kf9o-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.9 GB at this context; duplicate traces of the same session: j3mlbacb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 946.8,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 946.8 67.5 11.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cix1kf9o",
   "source_url": "https://omlx.ai/benchmarks/performance/cix1kf9o",
   "tg_tokens": null,
   "tps": 67.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-p4whm16j-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: 711v1as1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1369.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 1,369 99.3 11.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p4whm16j",
   "source_url": "https://omlx.ai/benchmarks/performance/p4whm16j",
   "tg_tokens": null,
   "tps": 99.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-p4whm16j-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: 711v1as1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1376.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,376 96.8 11.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p4whm16j",
   "source_url": "https://omlx.ai/benchmarks/performance/p4whm16j",
   "tg_tokens": null,
   "tps": 96.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-s95aggyj-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: 1n29iz2p; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1372.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 1,372 98.7 11.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s95aggyj",
   "source_url": "https://omlx.ai/benchmarks/performance/s95aggyj",
   "tg_tokens": null,
   "tps": 98.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-s95aggyj-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: 1n29iz2p; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1378.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,378 96.1 11.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s95aggyj",
   "source_url": "https://omlx.ai/benchmarks/performance/s95aggyj",
   "tg_tokens": null,
   "tps": 96.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-bi98x9e8-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 11.8 GB; TTFT 748 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1369.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k tokens 1,369 PP tok/s 100.3 TG tok/s 748 TTFT (ms) 11.8 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bi98x9e8",
   "source_url": "https://omlx.ai/benchmarks/performance/bi98x9e8",
   "tg_tokens": null,
   "tps": 100.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-emwv4wld-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 1pe7rxgz, 8og5un2c; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 105.9 15.5 18.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/emwv4wld",
   "source_url": "https://omlx.ai/benchmarks/performance/emwv4wld",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-emwv4wld-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 1pe7rxgz, 8og5un2c; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 150.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 150.8 14.7 19.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/emwv4wld",
   "source_url": "https://omlx.ai/benchmarks/performance/emwv4wld",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-emwv4wld-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 1pe7rxgz, 8og5un2c; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 115.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 115.0 14.5 19.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/emwv4wld",
   "source_url": "https://omlx.ai/benchmarks/performance/emwv4wld",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-emwv4wld-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 1pe7rxgz, 8og5un2c; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 96.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 96.0 12.8 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/emwv4wld",
   "source_url": "https://omlx.ai/benchmarks/performance/emwv4wld",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-3zf1n8i7-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 5y2n1rjs; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 105.4 15.1 18.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3zf1n8i7",
   "source_url": "https://omlx.ai/benchmarks/performance/3zf1n8i7",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-3zf1n8i7-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 5y2n1rjs; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 134.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 134.7 15.0 19.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3zf1n8i7",
   "source_url": "https://omlx.ai/benchmarks/performance/3zf1n8i7",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-3zf1n8i7-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: 5y2n1rjs; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 151.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 151.5 14.2 19.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3zf1n8i7",
   "source_url": "https://omlx.ai/benchmarks/performance/3zf1n8i7",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-3zf1n8i7-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 5y2n1rjs; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 97.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 97.6 12.8 19.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3zf1n8i7",
   "source_url": "https://omlx.ai/benchmarks/performance/3zf1n8i7",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3d1kmxzy-ctx8192",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.3 GB at this context; duplicate traces of the same session: ldw0yn8a; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 470.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 470.2 45.0 156.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3d1kmxzy",
   "source_url": "https://omlx.ai/benchmarks/performance/3d1kmxzy",
   "tg_tokens": null,
   "tps": 45.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3d1kmxzy-ctx32768",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.5 GB at this context; duplicate traces of the same session: ldw0yn8a; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 448.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 448.5 44.9 156.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3d1kmxzy",
   "source_url": "https://omlx.ai/benchmarks/performance/3d1kmxzy",
   "tg_tokens": null,
   "tps": 44.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0jmj5tkk-ctx16384",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.4 GB at this context; duplicate traces of the same session: ra3j8y09; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 449.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "DeepSeek-V4-Flash-0731-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 16k 449.8 37.1 156.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0jmj5tkk",
   "source_url": "https://omlx.ai/benchmarks/performance/0jmj5tkk",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0jmj5tkk-ctx32768",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.5 GB at this context; duplicate traces of the same session: ra3j8y09; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 450.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "DeepSeek-V4-Flash-0731-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 32k 450.3 45.1 156.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0jmj5tkk",
   "source_url": "https://omlx.ai/benchmarks/performance/0jmj5tkk",
   "tg_tokens": null,
   "tps": 45.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0jmj5tkk-ctx65536",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.8 GB at this context; duplicate traces of the same session: ra3j8y09; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 426.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "DeepSeek-V4-Flash-0731-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 64k 426.2 36.5 156.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0jmj5tkk",
   "source_url": "https://omlx.ai/benchmarks/performance/0jmj5tkk",
   "tg_tokens": null,
   "tps": 36.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0jmj5tkk-ctx131072",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 157.1 GB at this context; duplicate traces of the same session: ra3j8y09; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 385.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "DeepSeek-V4-Flash-0731-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 128k 385.6 35.1 157.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0jmj5tkk",
   "source_url": "https://omlx.ai/benchmarks/performance/0jmj5tkk",
   "tg_tokens": null,
   "tps": 35.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ewuuqdr0-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.5 GB at this context; duplicate traces of the same session: e5o5pios, tshe7lwb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 394.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 394.3 27.6 175.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ewuuqdr0",
   "source_url": "https://omlx.ai/benchmarks/performance/ewuuqdr0",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ewuuqdr0-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.6 GB at this context; duplicate traces of the same session: e5o5pios, tshe7lwb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 402.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 402.8 26.5 175.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ewuuqdr0",
   "source_url": "https://omlx.ai/benchmarks/performance/ewuuqdr0",
   "tg_tokens": null,
   "tps": 26.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ewuuqdr0-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 176.0 GB at this context; duplicate traces of the same session: e5o5pios, tshe7lwb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 397.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 397.0 25.0 176.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ewuuqdr0",
   "source_url": "https://omlx.ai/benchmarks/performance/ewuuqdr0",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ewuuqdr0-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 176.7 GB at this context; duplicate traces of the same session: e5o5pios, tshe7lwb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 388.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 128k 388.8 21.5 176.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ewuuqdr0",
   "source_url": "https://omlx.ai/benchmarks/performance/ewuuqdr0",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-az5fi4yu-ctx1024",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "fp16 attention; oMLX community benchmark trace; peak memory 8.0 GB at this context; duplicate traces of the same session: 4wesagw6; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 366.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-MLX-oQ6e-fp16 on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 6bit; 1k 366.2 34.2 8.0 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/az5fi4yu",
   "source_url": "https://omlx.ai/benchmarks/performance/az5fi4yu",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-az5fi4yu-ctx8192",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "fp16 attention; oMLX community benchmark trace; peak memory 8.7 GB at this context; duplicate traces of the same session: 4wesagw6; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 382.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-MLX-oQ6e-fp16 on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 6bit; 8k 382.7 32.9 8.7 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/az5fi4yu",
   "source_url": "https://omlx.ai/benchmarks/performance/az5fi4yu",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-1ziszmta-ctx1024",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 8.4 GB at this context; duplicate traces of the same session: 1b6lb5kt; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 358.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 358.1 45.0 8.4 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ziszmta",
   "source_url": "https://omlx.ai/benchmarks/performance/1ziszmta",
   "tg_tokens": null,
   "tps": 45.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-1ziszmta-ctx8192",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 8.7 GB at this context; duplicate traces of the same session: 1b6lb5kt; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 479.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 8k 479.9 40.1 8.7 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ziszmta",
   "source_url": "https://omlx.ai/benchmarks/performance/1ziszmta",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-j5tyxrtl-ctx1024",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 2.2 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1565.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiniCPM5-2B-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 1,565 132.4 2.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j5tyxrtl",
   "source_url": "https://omlx.ai/benchmarks/performance/j5tyxrtl",
   "tg_tokens": null,
   "tps": 132.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-j5tyxrtl-ctx4096",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 2.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1423.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiniCPM5-2B-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,423 116.2 2.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j5tyxrtl",
   "source_url": "https://omlx.ai/benchmarks/performance/j5tyxrtl",
   "tg_tokens": null,
   "tps": 116.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-j5tyxrtl-ctx8192",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 2.4 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1181.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiniCPM5-2B-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 1,181 105.2 2.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j5tyxrtl",
   "source_url": "https://omlx.ai/benchmarks/performance/j5tyxrtl",
   "tg_tokens": null,
   "tps": 105.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-j5tyxrtl-ctx16384",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 2.7 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 886.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiniCPM5-2B-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 886.3 89.3 2.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j5tyxrtl",
   "source_url": "https://omlx.ai/benchmarks/performance/j5tyxrtl",
   "tg_tokens": null,
   "tps": 89.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-j5tyxrtl-ctx32768",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 3.5 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 621.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiniCPM5-2B-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 621.7 65.8 3.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j5tyxrtl",
   "source_url": "https://omlx.ai/benchmarks/performance/j5tyxrtl",
   "tg_tokens": null,
   "tps": 65.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-90g7a435-ctx8192",
   "model": "Ornith-1.0-9B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: oxev93j3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1553.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.0-9B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 1,553 68.2 12.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/90g7a435",
   "source_url": "https://omlx.ai/benchmarks/performance/90g7a435",
   "tg_tokens": null,
   "tps": 68.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-90g7a435-ctx32768",
   "model": "Ornith-1.0-9B",
   "notes": "oMLX community benchmark trace; peak memory 13.8 GB at this context; duplicate traces of the same session: oxev93j3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1342.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.0-9B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 32k 1,342 61.5 13.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/90g7a435",
   "source_url": "https://omlx.ai/benchmarks/performance/90g7a435",
   "tg_tokens": null,
   "tps": 61.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-8fx9ya46-ctx8192",
   "model": "Ornith-1.0-9B",
   "notes": "oMLX community benchmark trace; peak memory 13.0 GB at this context; duplicate traces of the same session: dhe5hhgr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1437.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.0-9B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 1,437 62.4 13.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8fx9ya46",
   "source_url": "https://omlx.ai/benchmarks/performance/8fx9ya46",
   "tg_tokens": null,
   "tps": 62.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-8fx9ya46-ctx32768",
   "model": "Ornith-1.0-9B",
   "notes": "oMLX community benchmark trace; peak memory 14.2 GB at this context; duplicate traces of the same session: dhe5hhgr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1301.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.0-9B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 32k 1,301 52.5 14.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8fx9ya46",
   "source_url": "https://omlx.ai/benchmarks/performance/8fx9ya46",
   "tg_tokens": null,
   "tps": 52.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mmgyfazl-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 220 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4664.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 4,664 PP tok/s 124.6 TG tok/s 220 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mmgyfazl",
   "source_url": "https://omlx.ai/benchmarks/performance/mmgyfazl",
   "tg_tokens": null,
   "tps": 124.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qh3dlenf-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 227 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4518.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 4,518 PP tok/s 129.3 TG tok/s 227 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qh3dlenf",
   "source_url": "https://omlx.ai/benchmarks/performance/qh3dlenf",
   "tg_tokens": null,
   "tps": 129.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-b9x0n799-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 219 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4683.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 4,683 PP tok/s 125.3 TG tok/s 219 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b9x0n799",
   "source_url": "https://omlx.ai/benchmarks/performance/b9x0n799",
   "tg_tokens": null,
   "tps": 125.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-h1gsy7pw-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 228 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4492.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 4,492 PP tok/s 137.3 TG tok/s 228 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h1gsy7pw",
   "source_url": "https://omlx.ai/benchmarks/performance/h1gsy7pw",
   "tg_tokens": null,
   "tps": 137.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-355mrf08-ctx1024",
   "model": "Qwen3-Next-80B-A3B-Thinking-Uncensored",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 43.1 GB; TTFT 852 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1202.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Next-80B-A3B-Thinking-Uncensored-mlx-4Bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,202 PP tok/s 55.1 TG tok/s 852 TTFT (ms) 43.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/355mrf08",
   "source_url": "https://omlx.ai/benchmarks/performance/355mrf08",
   "tg_tokens": null,
   "tps": 55.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-1woqbp47-ctx32768",
   "model": "Qwen3-Next-80B-A3B-Thinking-Uncensored",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 44.7 GB; TTFT 37085 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 883.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Next-80B-A3B-Thinking-Uncensored-mlx-4Bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k tokens 883.6 PP tok/s 50.7 TG tok/s 37085 TTFT (ms) 44.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1woqbp47",
   "source_url": "https://omlx.ai/benchmarks/performance/1woqbp47",
   "tg_tokens": null,
   "tps": 50.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-gynfr246-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1712.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 1,712 53.3 95.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gynfr246",
   "source_url": "https://omlx.ai/benchmarks/performance/gynfr246",
   "tg_tokens": null,
   "tps": 53.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-gynfr246-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1731.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 1,731 50.1 95.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gynfr246",
   "source_url": "https://omlx.ai/benchmarks/performance/gynfr246",
   "tg_tokens": null,
   "tps": 50.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-gynfr246-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1651.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 1,651 48.8 95.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gynfr246",
   "source_url": "https://omlx.ai/benchmarks/performance/gynfr246",
   "tg_tokens": null,
   "tps": 48.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-gynfr246-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1484.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 1,484 46.1 95.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gynfr246",
   "source_url": "https://omlx.ai/benchmarks/performance/gynfr246",
   "tg_tokens": null,
   "tps": 46.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bplfbxvz-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1709.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 1,709 53.0 95.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bplfbxvz",
   "source_url": "https://omlx.ai/benchmarks/performance/bplfbxvz",
   "tg_tokens": null,
   "tps": 53.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bplfbxvz-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1727.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 1,727 49.7 95.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bplfbxvz",
   "source_url": "https://omlx.ai/benchmarks/performance/bplfbxvz",
   "tg_tokens": null,
   "tps": 49.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bplfbxvz-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1648.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 1,648 48.3 95.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bplfbxvz",
   "source_url": "https://omlx.ai/benchmarks/performance/bplfbxvz",
   "tg_tokens": null,
   "tps": 48.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bplfbxvz-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1481.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 1,481 44.9 95.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bplfbxvz",
   "source_url": "https://omlx.ai/benchmarks/performance/bplfbxvz",
   "tg_tokens": null,
   "tps": 44.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-04bbw9w8-ctx65536",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 96.8 GB at this context; duplicate traces of the same session: bkugem8q, y9u05uw2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1209.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k 1,209 39.7 96.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/04bbw9w8",
   "source_url": "https://omlx.ai/benchmarks/performance/04bbw9w8",
   "tg_tokens": null,
   "tps": 39.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-04bbw9w8-ctx131072",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 98.1 GB at this context; duplicate traces of the same session: bkugem8q, y9u05uw2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 881.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 128k 881.3 31.3 98.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/04bbw9w8",
   "source_url": "https://omlx.ai/benchmarks/performance/04bbw9w8",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-04bbw9w8-ctx199680",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 99.7 GB at this context; duplicate traces of the same session: bkugem8q, y9u05uw2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 688.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 195k 688.3 26.1 99.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/04bbw9w8",
   "source_url": "https://omlx.ai/benchmarks/performance/04bbw9w8",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-r9x72u7r-ctx65536",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 96.8 GB at this context; duplicate traces of the same session: n7qkkjl7, vinifbzt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1201.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k 1,201 39.3 96.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r9x72u7r",
   "source_url": "https://omlx.ai/benchmarks/performance/r9x72u7r",
   "tg_tokens": null,
   "tps": 39.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-r9x72u7r-ctx131072",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 98.1 GB at this context; duplicate traces of the same session: n7qkkjl7, vinifbzt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 885.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 128k 885.5 32.1 98.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r9x72u7r",
   "source_url": "https://omlx.ai/benchmarks/performance/r9x72u7r",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-r9x72u7r-ctx199680",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 99.7 GB at this context; duplicate traces of the same session: n7qkkjl7, vinifbzt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 690.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 195k 690.3 26.3 99.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r9x72u7r",
   "source_url": "https://omlx.ai/benchmarks/performance/r9x72u7r",
   "tg_tokens": null,
   "tps": 26.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-kibimees-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.9 GB at this context; duplicate traces of the same session: qbz7e0al; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1666.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 1,666 58.6 100.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kibimees",
   "source_url": "https://omlx.ai/benchmarks/performance/kibimees",
   "tg_tokens": null,
   "tps": 58.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-kibimees-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.8 GB at this context; duplicate traces of the same session: qbz7e0al; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1657.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 1,657 64.9 100.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kibimees",
   "source_url": "https://omlx.ai/benchmarks/performance/kibimees",
   "tg_tokens": null,
   "tps": 64.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-kibimees-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.0 GB at this context; duplicate traces of the same session: qbz7e0al; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1574.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 1,574 60.5 101.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kibimees",
   "source_url": "https://omlx.ai/benchmarks/performance/kibimees",
   "tg_tokens": null,
   "tps": 60.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-kibimees-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: qbz7e0al; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1407.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 1,407 55.9 101.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kibimees",
   "source_url": "https://omlx.ai/benchmarks/performance/kibimees",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-e4zqm2j0-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.9 GB at this context; duplicate traces of the same session: wk6mv5n4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1672.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 1,672 63.6 100.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e4zqm2j0",
   "source_url": "https://omlx.ai/benchmarks/performance/e4zqm2j0",
   "tg_tokens": null,
   "tps": 63.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-e4zqm2j0-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.8 GB at this context; duplicate traces of the same session: wk6mv5n4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1654.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 1,654 61.0 100.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e4zqm2j0",
   "source_url": "https://omlx.ai/benchmarks/performance/e4zqm2j0",
   "tg_tokens": null,
   "tps": 61.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-e4zqm2j0-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.0 GB at this context; duplicate traces of the same session: wk6mv5n4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1573.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 1,573 60.5 101.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e4zqm2j0",
   "source_url": "https://omlx.ai/benchmarks/performance/e4zqm2j0",
   "tg_tokens": null,
   "tps": 60.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-e4zqm2j0-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: wk6mv5n4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1411.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 1,411 56.7 101.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e4zqm2j0",
   "source_url": "https://omlx.ai/benchmarks/performance/e4zqm2j0",
   "tg_tokens": null,
   "tps": 56.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-a3ok0fr2-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.9 GB at this context; duplicate traces of the same session: 185q07hf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1665.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 1,665 60.4 100.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a3ok0fr2",
   "source_url": "https://omlx.ai/benchmarks/performance/a3ok0fr2",
   "tg_tokens": null,
   "tps": 60.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-a3ok0fr2-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.8 GB at this context; duplicate traces of the same session: 185q07hf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1647.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 1,647 64.0 100.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a3ok0fr2",
   "source_url": "https://omlx.ai/benchmarks/performance/a3ok0fr2",
   "tg_tokens": null,
   "tps": 64.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-a3ok0fr2-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.0 GB at this context; duplicate traces of the same session: 185q07hf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1578.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 1,578 61.8 101.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a3ok0fr2",
   "source_url": "https://omlx.ai/benchmarks/performance/a3ok0fr2",
   "tg_tokens": null,
   "tps": 61.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-a3ok0fr2-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: 185q07hf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1411.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 1,411 52.2 101.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a3ok0fr2",
   "source_url": "https://omlx.ai/benchmarks/performance/a3ok0fr2",
   "tg_tokens": null,
   "tps": 52.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nvisbf52-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.2 GB at this context; duplicate traces of the same session: no8zb9r7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1699.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 1,699 51.2 96.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nvisbf52",
   "source_url": "https://omlx.ai/benchmarks/performance/nvisbf52",
   "tg_tokens": null,
   "tps": 51.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nvisbf52-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.1 GB at this context; duplicate traces of the same session: no8zb9r7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1701.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 1,701 48.9 96.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nvisbf52",
   "source_url": "https://omlx.ai/benchmarks/performance/nvisbf52",
   "tg_tokens": null,
   "tps": 48.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nvisbf52-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.3 GB at this context; duplicate traces of the same session: no8zb9r7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1624.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 1,624 47.0 96.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nvisbf52",
   "source_url": "https://omlx.ai/benchmarks/performance/nvisbf52",
   "tg_tokens": null,
   "tps": 47.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nvisbf52-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.7 GB at this context; duplicate traces of the same session: no8zb9r7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1463.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 1,463 44.5 96.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nvisbf52",
   "source_url": "https://omlx.ai/benchmarks/performance/nvisbf52",
   "tg_tokens": null,
   "tps": 44.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xbrgpmbj-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.2 GB at this context; duplicate traces of the same session: sr3hr72s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1700.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 1,700 51.1 96.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xbrgpmbj",
   "source_url": "https://omlx.ai/benchmarks/performance/xbrgpmbj",
   "tg_tokens": null,
   "tps": 51.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xbrgpmbj-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.1 GB at this context; duplicate traces of the same session: sr3hr72s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1694.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 1,694 48.3 96.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xbrgpmbj",
   "source_url": "https://omlx.ai/benchmarks/performance/xbrgpmbj",
   "tg_tokens": null,
   "tps": 48.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xbrgpmbj-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.3 GB at this context; duplicate traces of the same session: sr3hr72s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1627.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 1,627 46.0 96.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xbrgpmbj",
   "source_url": "https://omlx.ai/benchmarks/performance/xbrgpmbj",
   "tg_tokens": null,
   "tps": 46.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xbrgpmbj-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.7 GB at this context; duplicate traces of the same session: sr3hr72s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1457.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 1,457 44.3 96.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xbrgpmbj",
   "source_url": "https://omlx.ai/benchmarks/performance/xbrgpmbj",
   "tg_tokens": null,
   "tps": 44.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4w4icx5o-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.2 GB at this context; duplicate traces of the same session: 3yn3nds4, a4w8ebx1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1703.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 1,703 51.1 96.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4w4icx5o",
   "source_url": "https://omlx.ai/benchmarks/performance/4w4icx5o",
   "tg_tokens": null,
   "tps": 51.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4w4icx5o-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.1 GB at this context; duplicate traces of the same session: 3yn3nds4, a4w8ebx1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1703.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 1,703 48.6 96.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4w4icx5o",
   "source_url": "https://omlx.ai/benchmarks/performance/4w4icx5o",
   "tg_tokens": null,
   "tps": 48.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4w4icx5o-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.3 GB at this context; duplicate traces of the same session: 3yn3nds4, a4w8ebx1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1628.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 1,628 45.6 96.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4w4icx5o",
   "source_url": "https://omlx.ai/benchmarks/performance/4w4icx5o",
   "tg_tokens": null,
   "tps": 45.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4w4icx5o-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.7 GB at this context; duplicate traces of the same session: 3yn3nds4, a4w8ebx1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1470.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 1,470 44.6 96.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4w4icx5o",
   "source_url": "https://omlx.ai/benchmarks/performance/4w4icx5o",
   "tg_tokens": null,
   "tps": 44.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-5qvzekfa-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.7 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 8bit; 1k 127.6 12.3 10.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5qvzekfa",
   "source_url": "https://omlx.ai/benchmarks/performance/5qvzekfa",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-5qvzekfa-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 132.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 8bit; 4k 132.0 12.0 11.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5qvzekfa",
   "source_url": "https://omlx.ai/benchmarks/performance/5qvzekfa",
   "tg_tokens": null,
   "tps": 12.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-5qvzekfa-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.2 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 8bit; 8k 130.7 11.5 11.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5qvzekfa",
   "source_url": "https://omlx.ai/benchmarks/performance/5qvzekfa",
   "tg_tokens": null,
   "tps": 11.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-5qvzekfa-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.7 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 8bit; 16k 127.5 10.6 11.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5qvzekfa",
   "source_url": "https://omlx.ai/benchmarks/performance/5qvzekfa",
   "tg_tokens": null,
   "tps": 10.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-03-09",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-5qvzekfa-ctx32768",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 90.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 8bit; 32k 90.4 8.6 12.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5qvzekfa",
   "source_url": "https://omlx.ai/benchmarks/performance/5qvzekfa",
   "tg_tokens": null,
   "tps": 8.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-3jtaloiq-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 284.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 1k 284.7 47.3 5.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3jtaloiq",
   "source_url": "https://omlx.ai/benchmarks/performance/3jtaloiq",
   "tg_tokens": null,
   "tps": 47.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-3jtaloiq-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 276.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k 276.2 45.9 6.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3jtaloiq",
   "source_url": "https://omlx.ai/benchmarks/performance/3jtaloiq",
   "tg_tokens": null,
   "tps": 45.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-3jtaloiq-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 241.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 16k 241.1 8.8 19.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3jtaloiq",
   "source_url": "https://omlx.ai/benchmarks/performance/3jtaloiq",
   "tg_tokens": null,
   "tps": 8.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-hkahu15r-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 837.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 16 GB \u00b7 4bit; 1k 837.7 26.4 5.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hkahu15r",
   "source_url": "https://omlx.ai/benchmarks/performance/hkahu15r",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-hkahu15r-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 885.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 16 GB \u00b7 4bit; 4k 885.3 25.5 6.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hkahu15r",
   "source_url": "https://omlx.ai/benchmarks/performance/hkahu15r",
   "tg_tokens": null,
   "tps": 25.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-w9sp0t7k-ctx4096",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 09ntp2z5; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 97.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-27B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 97.7 18.9 19.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w9sp0t7k",
   "source_url": "https://omlx.ai/benchmarks/performance/w9sp0t7k",
   "tg_tokens": null,
   "tps": 18.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-w9sp0t7k-ctx16384",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 09ntp2z5; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 93.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-27B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 16k 93.8 17.6 21.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w9sp0t7k",
   "source_url": "https://omlx.ai/benchmarks/performance/w9sp0t7k",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-kl51vdlf-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: edb17jze; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 694.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 694.3 84.8 21.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kl51vdlf",
   "source_url": "https://omlx.ai/benchmarks/performance/kl51vdlf",
   "tg_tokens": null,
   "tps": 84.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-kl51vdlf-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: edb17jze; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 606.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 16k 606.5 74.1 22.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kl51vdlf",
   "source_url": "https://omlx.ai/benchmarks/performance/kl51vdlf",
   "tg_tokens": null,
   "tps": 74.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-jnb7sytn-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3006.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,006 105.5 20.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jnb7sytn",
   "source_url": "https://omlx.ai/benchmarks/performance/jnb7sytn",
   "tg_tokens": null,
   "tps": 105.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-jnb7sytn-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3041.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 3,041 93.7 20.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jnb7sytn",
   "source_url": "https://omlx.ai/benchmarks/performance/jnb7sytn",
   "tg_tokens": null,
   "tps": 93.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-jnb7sytn-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1350.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 128k 1,350 46.8 22.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jnb7sytn",
   "source_url": "https://omlx.ai/benchmarks/performance/jnb7sytn",
   "tg_tokens": null,
   "tps": 46.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-l8wf6fi5-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V13-dequantized",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.5 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1945.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V13-dequantized-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 1,945 149.6 37.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l8wf6fi5",
   "source_url": "https://omlx.ai/benchmarks/performance/l8wf6fi5",
   "tg_tokens": null,
   "tps": 149.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-l8wf6fi5-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V13-dequantized",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.1 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2304.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V13-dequantized-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 2,304 127.5 39.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l8wf6fi5",
   "source_url": "https://omlx.ai/benchmarks/performance/l8wf6fi5",
   "tg_tokens": null,
   "tps": 127.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-d9zfmiis-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 105.4 18.0 19.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d9zfmiis",
   "source_url": "https://omlx.ai/benchmarks/performance/d9zfmiis",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-d9zfmiis-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 111.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 111.3 19.2 22.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d9zfmiis",
   "source_url": "https://omlx.ai/benchmarks/performance/d9zfmiis",
   "tg_tokens": null,
   "tps": 19.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ilfux3gp-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 465.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 465.8 18.2 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ilfux3gp",
   "source_url": "https://omlx.ai/benchmarks/performance/ilfux3gp",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ilfux3gp-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 459.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 459.5 17.8 17.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ilfux3gp",
   "source_url": "https://omlx.ai/benchmarks/performance/ilfux3gp",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-b0x76m13-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 42.9 GB at this context; duplicate traces of the same session: 3iqllwal; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 440.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 440.6 37.9 42.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b0x76m13",
   "source_url": "https://omlx.ai/benchmarks/performance/b0x76m13",
   "tg_tokens": null,
   "tps": 37.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-b0x76m13-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 45.4 GB at this context; duplicate traces of the same session: 3iqllwal; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 380.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 32k 380.1 62.4 45.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b0x76m13",
   "source_url": "https://omlx.ai/benchmarks/performance/b0x76m13",
   "tg_tokens": null,
   "tps": 62.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-n76fs97w-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: d07wfsxo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 217.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-fp16-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 217.6 47.5 22.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n76fs97w",
   "source_url": "https://omlx.ai/benchmarks/performance/n76fs97w",
   "tg_tokens": null,
   "tps": 47.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-n76fs97w-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: d07wfsxo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 85.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-fp16-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 85.8 39.9 25.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n76fs97w",
   "source_url": "https://omlx.ai/benchmarks/performance/n76fs97w",
   "tg_tokens": null,
   "tps": 39.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-j6pmudnh-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 154.4,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 154.4 19.6 16.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j6pmudnh",
   "source_url": "https://omlx.ai/benchmarks/performance/j6pmudnh",
   "tg_tokens": null,
   "tps": 19.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-j6pmudnh-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 163.4,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 163.4 18.8 20.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j6pmudnh",
   "source_url": "https://omlx.ai/benchmarks/performance/j6pmudnh",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-3bgmdgnq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 65.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k 65.3 6.5 16.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3bgmdgnq",
   "source_url": "https://omlx.ai/benchmarks/performance/3bgmdgnq",
   "tg_tokens": null,
   "tps": 6.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-3bgmdgnq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 66.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 66.7 6.3 17.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3bgmdgnq",
   "source_url": "https://omlx.ai/benchmarks/performance/3bgmdgnq",
   "tg_tokens": null,
   "tps": 6.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-csk6ivwz-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: 3vq1q0gy, m1u78jiq, rmm80x3i, s9p8o4ub, vt0ioyhc; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1146.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 1k 1,146 80.1 34.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/csk6ivwz",
   "source_url": "https://omlx.ai/benchmarks/performance/csk6ivwz",
   "tg_tokens": null,
   "tps": 80.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-csk6ivwz-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: 3vq1q0gy, m1u78jiq, rmm80x3i, s9p8o4ub, vt0ioyhc; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1401.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 4k 1,401 69.7 35.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/csk6ivwz",
   "source_url": "https://omlx.ai/benchmarks/performance/csk6ivwz",
   "tg_tokens": null,
   "tps": 69.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-csk6ivwz-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: 3vq1q0gy, m1u78jiq, rmm80x3i, s9p8o4ub, vt0ioyhc; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1404.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 8k 1,404 80.4 36.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/csk6ivwz",
   "source_url": "https://omlx.ai/benchmarks/performance/csk6ivwz",
   "tg_tokens": null,
   "tps": 80.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-csk6ivwz-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 36.6 GB at this context; duplicate traces of the same session: 3vq1q0gy, m1u78jiq, rmm80x3i, s9p8o4ub, vt0ioyhc; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1365.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 16k 1,365 91.7 36.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/csk6ivwz",
   "source_url": "https://omlx.ai/benchmarks/performance/csk6ivwz",
   "tg_tokens": null,
   "tps": 91.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-csk6ivwz-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 37.8 GB at this context; duplicate traces of the same session: 3vq1q0gy, m1u78jiq, rmm80x3i, s9p8o4ub, vt0ioyhc; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1275.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 32k 1,275 93.1 37.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/csk6ivwz",
   "source_url": "https://omlx.ai/benchmarks/performance/csk6ivwz",
   "tg_tokens": null,
   "tps": 93.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-csk6ivwz-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 40.1 GB at this context; duplicate traces of the same session: 3vq1q0gy, m1u78jiq, rmm80x3i, s9p8o4ub, vt0ioyhc; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1119.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 64k 1,119 72.5 40.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/csk6ivwz",
   "source_url": "https://omlx.ai/benchmarks/performance/csk6ivwz",
   "tg_tokens": null,
   "tps": 72.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lzdlvr16-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.0 GB at this context; duplicate traces of the same session: fh670599; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 440.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 440.1 55.1 43.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lzdlvr16",
   "source_url": "https://omlx.ai/benchmarks/performance/lzdlvr16",
   "tg_tokens": null,
   "tps": 55.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lzdlvr16-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 45.4 GB at this context; duplicate traces of the same session: fh670599; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 380.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 32k 380.8 62.6 45.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lzdlvr16",
   "source_url": "https://omlx.ai/benchmarks/performance/lzdlvr16",
   "tg_tokens": null,
   "tps": 62.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-u0x3kuj4-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 916.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 916.4 38.1 35.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u0x3kuj4",
   "source_url": "https://omlx.ai/benchmarks/performance/u0x3kuj4",
   "tg_tokens": null,
   "tps": 38.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-u0x3kuj4-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 923.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 923.8 27.2 35.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u0x3kuj4",
   "source_url": "https://omlx.ai/benchmarks/performance/u0x3kuj4",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-u0x3kuj4-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 901.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 901.5 27.4 35.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u0x3kuj4",
   "source_url": "https://omlx.ai/benchmarks/performance/u0x3kuj4",
   "tg_tokens": null,
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-u0x3kuj4-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 844.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 844.0 28.5 36.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u0x3kuj4",
   "source_url": "https://omlx.ai/benchmarks/performance/u0x3kuj4",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-x342yu55-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 914.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 914.1 29.3 35.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x342yu55",
   "source_url": "https://omlx.ai/benchmarks/performance/x342yu55",
   "tg_tokens": null,
   "tps": 29.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-x342yu55-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 922.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 922.8 25.9 35.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x342yu55",
   "source_url": "https://omlx.ai/benchmarks/performance/x342yu55",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-x342yu55-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 901.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 901.1 36.3 35.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x342yu55",
   "source_url": "https://omlx.ai/benchmarks/performance/x342yu55",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-x342yu55-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 845.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 845.1 28.4 36.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x342yu55",
   "source_url": "https://omlx.ai/benchmarks/performance/x342yu55",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-purne41h-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 914.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 914.8 39.5 35.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/purne41h",
   "source_url": "https://omlx.ai/benchmarks/performance/purne41h",
   "tg_tokens": null,
   "tps": 39.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-purne41h-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 923.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 923.9 26.6 35.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/purne41h",
   "source_url": "https://omlx.ai/benchmarks/performance/purne41h",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-purne41h-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 901.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 901.2 28.2 35.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/purne41h",
   "source_url": "https://omlx.ai/benchmarks/performance/purne41h",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-purne41h-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 844.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 844.4 25.9 36.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/purne41h",
   "source_url": "https://omlx.ai/benchmarks/performance/purne41h",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ltnr6ygk-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 749.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 749.1 42.3 39.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ltnr6ygk",
   "source_url": "https://omlx.ai/benchmarks/performance/ltnr6ygk",
   "tg_tokens": null,
   "tps": 42.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ltnr6ygk-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 597.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 597.5 39.1 43.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ltnr6ygk",
   "source_url": "https://omlx.ai/benchmarks/performance/ltnr6ygk",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ltnr6ygk-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 49.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 491.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 195k 491.9 26.9 49.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ltnr6ygk",
   "source_url": "https://omlx.ai/benchmarks/performance/ltnr6ygk",
   "tg_tokens": null,
   "tps": 26.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-53o4qubr-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 108.5 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 865.9,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ4-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 865.9 47.9 108.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/53o4qubr",
   "source_url": "https://omlx.ai/benchmarks/performance/53o4qubr",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-53o4qubr-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 109.3 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 969.4,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ4-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 969.4 45.3 109.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/53o4qubr",
   "source_url": "https://omlx.ai/benchmarks/performance/53o4qubr",
   "tg_tokens": null,
   "tps": 45.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-4ks7k0oj-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.3 GB at this context; duplicate traces of the same session: d0j5uvk2, qr20h47d, x9pxv48h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 431.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 431.8 52.9 69.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ks7k0oj",
   "source_url": "https://omlx.ai/benchmarks/performance/4ks7k0oj",
   "tg_tokens": null,
   "tps": 52.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-4ks7k0oj-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.9 GB at this context; duplicate traces of the same session: d0j5uvk2, qr20h47d, x9pxv48h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 484.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 484.6 48.7 69.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ks7k0oj",
   "source_url": "https://omlx.ai/benchmarks/performance/4ks7k0oj",
   "tg_tokens": null,
   "tps": 48.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-4ks7k0oj-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.9 GB at this context; duplicate traces of the same session: d0j5uvk2, qr20h47d, x9pxv48h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 633.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 633.1 53.5 69.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ks7k0oj",
   "source_url": "https://omlx.ai/benchmarks/performance/4ks7k0oj",
   "tg_tokens": null,
   "tps": 53.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-4ks7k0oj-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.1 GB at this context; duplicate traces of the same session: d0j5uvk2, qr20h47d, x9pxv48h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 580.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 580.7 54.5 70.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ks7k0oj",
   "source_url": "https://omlx.ai/benchmarks/performance/4ks7k0oj",
   "tg_tokens": null,
   "tps": 54.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ko47l5tz-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 335.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 335.2 78.4 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ko47l5tz",
   "source_url": "https://omlx.ai/benchmarks/performance/ko47l5tz",
   "tg_tokens": null,
   "tps": 78.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ko47l5tz-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 932.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 932.3 51.2 70.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ko47l5tz",
   "source_url": "https://omlx.ai/benchmarks/performance/ko47l5tz",
   "tg_tokens": null,
   "tps": 51.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5n3l3ofa-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 87.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1036.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 1,036 57.2 87.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5n3l3ofa",
   "source_url": "https://omlx.ai/benchmarks/performance/5n3l3ofa",
   "tg_tokens": null,
   "tps": 57.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5n3l3ofa-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 87.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1918.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 1,918 56.6 87.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5n3l3ofa",
   "source_url": "https://omlx.ai/benchmarks/performance/5n3l3ofa",
   "tg_tokens": null,
   "tps": 56.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5n3l3ofa-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 88.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2339.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 2,339 56.1 88.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5n3l3ofa",
   "source_url": "https://omlx.ai/benchmarks/performance/5n3l3ofa",
   "tg_tokens": null,
   "tps": 56.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5n3l3ofa-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 89.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2480.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 2,480 56.4 89.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5n3l3ofa",
   "source_url": "https://omlx.ai/benchmarks/performance/5n3l3ofa",
   "tg_tokens": null,
   "tps": 56.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-iiqlm9iz-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 87.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1439.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 1,439 63.2 87.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iiqlm9iz",
   "source_url": "https://omlx.ai/benchmarks/performance/iiqlm9iz",
   "tg_tokens": null,
   "tps": 63.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-iiqlm9iz-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 87.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2078.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 2,078 66.1 87.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iiqlm9iz",
   "source_url": "https://omlx.ai/benchmarks/performance/iiqlm9iz",
   "tg_tokens": null,
   "tps": 66.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-iiqlm9iz-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 88.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2444.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 2,444 60.4 88.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iiqlm9iz",
   "source_url": "https://omlx.ai/benchmarks/performance/iiqlm9iz",
   "tg_tokens": null,
   "tps": 60.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-iiqlm9iz-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 89.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2456.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 2,456 60.3 89.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iiqlm9iz",
   "source_url": "https://omlx.ai/benchmarks/performance/iiqlm9iz",
   "tg_tokens": null,
   "tps": 60.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-oy6f2wyy-ctx4096",
   "model": "clef-flash",
   "notes": "oMLX community benchmark trace; peak memory 10.5 GB at this context; duplicate traces of the same session: 0rpwu95t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3148.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "clef-flash-8bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 8bit; 4k 3,148 57.2 10.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oy6f2wyy",
   "source_url": "https://omlx.ai/benchmarks/performance/oy6f2wyy",
   "tg_tokens": null,
   "tps": 57.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-oy6f2wyy-ctx16384",
   "model": "clef-flash",
   "notes": "oMLX community benchmark trace; peak memory 10.9 GB at this context; duplicate traces of the same session: 0rpwu95t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3140.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "clef-flash-8bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 8bit; 16k 3,140 55.1 10.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oy6f2wyy",
   "source_url": "https://omlx.ai/benchmarks/performance/oy6f2wyy",
   "tg_tokens": null,
   "tps": 55.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-90mnwilp-ctx8192",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "drowzeys-keys-Mac abliterated build; oMLX community benchmark trace; peak memory 157.5 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 466.1,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "drowzeys-keys-Mac-DeepSeek-V4-Flash-0731-MXFP4-MLX-Abliterated on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 466.1 44.5 157.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/90mnwilp",
   "source_url": "https://omlx.ai/benchmarks/performance/90mnwilp",
   "tg_tokens": null,
   "tps": 44.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-90mnwilp-ctx32768",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "drowzeys-keys-Mac abliterated build; oMLX community benchmark trace; peak memory 161.9 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 445.1,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "drowzeys-keys-Mac-DeepSeek-V4-Flash-0731-MXFP4-MLX-Abliterated on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 445.1 45.2 161.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/90mnwilp",
   "source_url": "https://omlx.ai/benchmarks/performance/90mnwilp",
   "tg_tokens": null,
   "tps": 45.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-0tho7xm7-ctx1024",
   "model": "gemma4 E4B",
   "notes": "OptiQ build; oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 806.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-OptiQ-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 1k 806.1 52.2 6.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0tho7xm7",
   "source_url": "https://omlx.ai/benchmarks/performance/0tho7xm7",
   "tg_tokens": null,
   "tps": 52.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-0tho7xm7-ctx4096",
   "model": "gemma4 E4B",
   "notes": "OptiQ build; oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 801.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-OptiQ-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k 801.6 49.5 6.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0tho7xm7",
   "source_url": "https://omlx.ai/benchmarks/performance/0tho7xm7",
   "tg_tokens": null,
   "tps": 49.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-0tho7xm7-ctx16384",
   "model": "gemma4 E4B",
   "notes": "OptiQ build; oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 794.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-OptiQ-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 16k 794.8 42.1 7.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0tho7xm7",
   "source_url": "https://omlx.ai/benchmarks/performance/0tho7xm7",
   "tg_tokens": null,
   "tps": 42.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-hmdfh1ev-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1058.0,
   "provenance": "community",
   "quant": "MXFP4-Q4LH",
   "quote": "gpt-oss-20b-MXFP4-Q4LH on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 1,058 89.6 11.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hmdfh1ev",
   "source_url": "https://omlx.ai/benchmarks/performance/hmdfh1ev",
   "tg_tokens": null,
   "tps": 89.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-hmdfh1ev-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1072.0,
   "provenance": "community",
   "quant": "MXFP4-Q4LH",
   "quote": "gpt-oss-20b-MXFP4-Q4LH on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,072 85.9 11.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hmdfh1ev",
   "source_url": "https://omlx.ai/benchmarks/performance/hmdfh1ev",
   "tg_tokens": null,
   "tps": 85.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-k1wdrw76-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 839.9,
   "provenance": "community",
   "quant": "MXFP4-Q4LH",
   "quote": "gpt-oss-20b-MXFP4-Q4LH on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 839.9 94.1 11.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k1wdrw76",
   "source_url": "https://omlx.ai/benchmarks/performance/k1wdrw76",
   "tg_tokens": null,
   "tps": 94.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-k1wdrw76-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1025.0,
   "provenance": "community",
   "quant": "MXFP4-Q4LH",
   "quote": "gpt-oss-20b-MXFP4-Q4LH on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,025 83.1 11.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k1wdrw76",
   "source_url": "https://omlx.ai/benchmarks/performance/k1wdrw76",
   "tg_tokens": null,
   "tps": 83.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-k40y4pv6-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 975.0,
   "provenance": "community",
   "quant": "MXFP4-Q4LH",
   "quote": "gpt-oss-20b-MXFP4-Q4LH on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 975.0 89.7 11.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k40y4pv6",
   "source_url": "https://omlx.ai/benchmarks/performance/k40y4pv6",
   "tg_tokens": null,
   "tps": 89.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-k40y4pv6-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 965.9,
   "provenance": "community",
   "quant": "MXFP4-Q4LH",
   "quote": "gpt-oss-20b-MXFP4-Q4LH on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 965.9 84.7 11.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k40y4pv6",
   "source_url": "https://omlx.ai/benchmarks/performance/k40y4pv6",
   "tg_tokens": null,
   "tps": 84.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-jq17kvsd-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1368.0,
   "provenance": "community",
   "quant": "MXFP4-Q4LH",
   "quote": "gpt-oss-20b-MXFP4-Q4LH on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 1,368 108.1 11.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jq17kvsd",
   "source_url": "https://omlx.ai/benchmarks/performance/jq17kvsd",
   "tg_tokens": null,
   "tps": 108.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-jq17kvsd-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1351.0,
   "provenance": "community",
   "quant": "MXFP4-Q4LH",
   "quote": "gpt-oss-20b-MXFP4-Q4LH on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,351 103.3 11.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jq17kvsd",
   "source_url": "https://omlx.ai/benchmarks/performance/jq17kvsd",
   "tg_tokens": null,
   "tps": 103.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-y0o6ij5s-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1361.0,
   "provenance": "community",
   "quant": "MXFP4-Q4LH",
   "quote": "gpt-oss-20b-MXFP4-Q4LH on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 1,361 107.3 11.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y0o6ij5s",
   "source_url": "https://omlx.ai/benchmarks/performance/y0o6ij5s",
   "tg_tokens": null,
   "tps": 107.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-y0o6ij5s-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1348.0,
   "provenance": "community",
   "quant": "MXFP4-Q4LH",
   "quote": "gpt-oss-20b-MXFP4-Q4LH on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,348 103.5 11.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y0o6ij5s",
   "source_url": "https://omlx.ai/benchmarks/performance/y0o6ij5s",
   "tg_tokens": null,
   "tps": 103.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-8xidrlhm-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1368.0,
   "provenance": "community",
   "quant": "MXFP4-Q4LH",
   "quote": "gpt-oss-20b-MXFP4-Q4LH on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 1,368 107.0 11.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8xidrlhm",
   "source_url": "https://omlx.ai/benchmarks/performance/8xidrlhm",
   "tg_tokens": null,
   "tps": 107.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-8xidrlhm-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1374.0,
   "provenance": "community",
   "quant": "MXFP4-Q4LH",
   "quote": "gpt-oss-20b-MXFP4-Q4LH on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,374 102.7 11.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8xidrlhm",
   "source_url": "https://omlx.ai/benchmarks/performance/8xidrlhm",
   "tg_tokens": null,
   "tps": 102.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-pmo45t7e-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 906.8,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 906.8 87.0 11.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pmo45t7e",
   "source_url": "https://omlx.ai/benchmarks/performance/pmo45t7e",
   "tg_tokens": null,
   "tps": 87.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-pmo45t7e-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 974.5,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 974.5 74.4 11.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pmo45t7e",
   "source_url": "https://omlx.ai/benchmarks/performance/pmo45t7e",
   "tg_tokens": null,
   "tps": 74.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-t0454j17-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 996.6,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 996.6 86.4 11.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t0454j17",
   "source_url": "https://omlx.ai/benchmarks/performance/t0454j17",
   "tg_tokens": null,
   "tps": 86.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-t0454j17-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1047.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,047 78.3 11.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t0454j17",
   "source_url": "https://omlx.ai/benchmarks/performance/t0454j17",
   "tg_tokens": null,
   "tps": 78.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-58ug8aa2-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 454.5,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 454.5 48.0 11.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/58ug8aa2",
   "source_url": "https://omlx.ai/benchmarks/performance/58ug8aa2",
   "tg_tokens": null,
   "tps": 48.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-58ug8aa2-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 354.6,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 354.6 32.0 11.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/58ug8aa2",
   "source_url": "https://omlx.ai/benchmarks/performance/58ug8aa2",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-j3mlbacb-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1373.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 1,373 69.8 11.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j3mlbacb",
   "source_url": "https://omlx.ai/benchmarks/performance/j3mlbacb",
   "tg_tokens": null,
   "tps": 69.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-j3mlbacb-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 946.8,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 946.8 67.5 11.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j3mlbacb",
   "source_url": "https://omlx.ai/benchmarks/performance/j3mlbacb",
   "tg_tokens": null,
   "tps": 67.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-1n29iz2p-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1372.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 1,372 98.7 11.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1n29iz2p",
   "source_url": "https://omlx.ai/benchmarks/performance/1n29iz2p",
   "tg_tokens": null,
   "tps": 98.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-1n29iz2p-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1378.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,378 96.1 11.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1n29iz2p",
   "source_url": "https://omlx.ai/benchmarks/performance/1n29iz2p",
   "tg_tokens": null,
   "tps": 96.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-711v1as1-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1369.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 1,369 99.3 11.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/711v1as1",
   "source_url": "https://omlx.ai/benchmarks/performance/711v1as1",
   "tg_tokens": null,
   "tps": 99.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-711v1as1-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1376.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,376 96.8 11.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/711v1as1",
   "source_url": "https://omlx.ai/benchmarks/performance/711v1as1",
   "tg_tokens": null,
   "tps": 96.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vb7iqz2d-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: 4i1gfl9r; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2063.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.5-4B-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 2,063 170.1 3.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vb7iqz2d",
   "source_url": "https://omlx.ai/benchmarks/performance/vb7iqz2d",
   "tg_tokens": null,
   "tps": 170.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vb7iqz2d-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.7 GB at this context; duplicate traces of the same session: 4i1gfl9r; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2172.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.5-4B-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 2,172 163.6 4.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vb7iqz2d",
   "source_url": "https://omlx.ai/benchmarks/performance/vb7iqz2d",
   "tg_tokens": null,
   "tps": 163.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gqr2k4hj-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.8 GB at this context; duplicate traces of the same session: a090ohsb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1914.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.5-4B-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 1,914 163.6 3.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gqr2k4hj",
   "source_url": "https://omlx.ai/benchmarks/performance/gqr2k4hj",
   "tg_tokens": null,
   "tps": 163.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gqr2k4hj-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.6 GB at this context; duplicate traces of the same session: a090ohsb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2256.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.5-4B-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 2,256 159.3 4.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gqr2k4hj",
   "source_url": "https://omlx.ai/benchmarks/performance/gqr2k4hj",
   "tg_tokens": null,
   "tps": 159.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-8og5un2c-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 1pe7rxgz; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 105.9 15.5 18.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8og5un2c",
   "source_url": "https://omlx.ai/benchmarks/performance/8og5un2c",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-8og5un2c-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 1pe7rxgz; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 150.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 150.8 14.7 19.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8og5un2c",
   "source_url": "https://omlx.ai/benchmarks/performance/8og5un2c",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-8og5un2c-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 1pe7rxgz; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 115.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 115.0 14.5 19.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8og5un2c",
   "source_url": "https://omlx.ai/benchmarks/performance/8og5un2c",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-8og5un2c-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 1pe7rxgz; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 96.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 96.0 12.8 19.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8og5un2c",
   "source_url": "https://omlx.ai/benchmarks/performance/8og5un2c",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-5y2n1rjs-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 0w05qp6k, d2d84wd2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 105.4 15.1 18.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5y2n1rjs",
   "source_url": "https://omlx.ai/benchmarks/performance/5y2n1rjs",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-5y2n1rjs-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 0w05qp6k, d2d84wd2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 134.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 134.7 15.0 19.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5y2n1rjs",
   "source_url": "https://omlx.ai/benchmarks/performance/5y2n1rjs",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-5y2n1rjs-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: 0w05qp6k, d2d84wd2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 151.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 151.5 14.2 19.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5y2n1rjs",
   "source_url": "https://omlx.ai/benchmarks/performance/5y2n1rjs",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-5y2n1rjs-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 0w05qp6k, d2d84wd2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 97.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 97.6 12.8 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5y2n1rjs",
   "source_url": "https://omlx.ai/benchmarks/performance/5y2n1rjs",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-gdn5oo3n-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: gzyg4eih, p9txj8t4, zszu42ci; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 105.2 14.6 18.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gdn5oo3n",
   "source_url": "https://omlx.ai/benchmarks/performance/gdn5oo3n",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-gdn5oo3n-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: gzyg4eih, p9txj8t4, zszu42ci; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 113.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 113.1 14.4 19.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gdn5oo3n",
   "source_url": "https://omlx.ai/benchmarks/performance/gdn5oo3n",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-gdn5oo3n-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: gzyg4eih, p9txj8t4, zszu42ci; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 106.5 14.3 17.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gdn5oo3n",
   "source_url": "https://omlx.ai/benchmarks/performance/gdn5oo3n",
   "tg_tokens": null,
   "tps": 14.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-gdn5oo3n-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: gzyg4eih, p9txj8t4, zszu42ci; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 96.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 96.0 12.7 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gdn5oo3n",
   "source_url": "https://omlx.ai/benchmarks/performance/gdn5oo3n",
   "tg_tokens": null,
   "tps": 12.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-jcs2if36-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: q2pv67xn; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 105.4 15.0 18.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jcs2if36",
   "source_url": "https://omlx.ai/benchmarks/performance/jcs2if36",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-jcs2if36-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: q2pv67xn; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 125.0 14.5 19.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jcs2if36",
   "source_url": "https://omlx.ai/benchmarks/performance/jcs2if36",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-jcs2if36-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: q2pv67xn; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 106.2 13.8 17.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jcs2if36",
   "source_url": "https://omlx.ai/benchmarks/performance/jcs2if36",
   "tg_tokens": null,
   "tps": 13.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-jcs2if36-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: q2pv67xn; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 95.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 95.9 12.3 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jcs2if36",
   "source_url": "https://omlx.ai/benchmarks/performance/jcs2if36",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ldw0yn8a-ctx8192",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.3 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 470.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 470.2 45.0 156.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ldw0yn8a",
   "source_url": "https://omlx.ai/benchmarks/performance/ldw0yn8a",
   "tg_tokens": null,
   "tps": 45.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ldw0yn8a-ctx32768",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.5 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 448.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 448.5 44.9 156.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ldw0yn8a",
   "source_url": "https://omlx.ai/benchmarks/performance/ldw0yn8a",
   "tg_tokens": null,
   "tps": 44.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ra3j8y09-ctx16384",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.4 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 449.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "DeepSeek-V4-Flash-0731-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 16k 449.8 37.1 156.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ra3j8y09",
   "source_url": "https://omlx.ai/benchmarks/performance/ra3j8y09",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ra3j8y09-ctx32768",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.5 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 450.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "DeepSeek-V4-Flash-0731-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 32k 450.3 45.1 156.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ra3j8y09",
   "source_url": "https://omlx.ai/benchmarks/performance/ra3j8y09",
   "tg_tokens": null,
   "tps": 45.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ra3j8y09-ctx65536",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.8 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 426.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "DeepSeek-V4-Flash-0731-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 64k 426.2 36.5 156.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ra3j8y09",
   "source_url": "https://omlx.ai/benchmarks/performance/ra3j8y09",
   "tg_tokens": null,
   "tps": 36.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ra3j8y09-ctx131072",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 157.1 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 385.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "DeepSeek-V4-Flash-0731-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 128k 385.6 35.1 157.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ra3j8y09",
   "source_url": "https://omlx.ai/benchmarks/performance/ra3j8y09",
   "tg_tokens": null,
   "tps": 35.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-75ug0jmf-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace, single run at 1k context; peak memory 21.8 GB; TTFT 5109 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 200.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ4e on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k tokens 200.4 PP tok/s 15.4 TG tok/s 5109 TTFT (ms) 21.8 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/75ug0jmf",
   "source_url": "https://omlx.ai/benchmarks/performance/75ug0jmf",
   "tg_tokens": null,
   "tps": 15.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-06mhnvzq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace, single run at 1k context; peak memory 17.8 GB; TTFT 4225 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 242.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ4e on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k tokens 242.4 PP tok/s 14.2 TG tok/s 4225 TTFT (ms) 17.8 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/06mhnvzq",
   "source_url": "https://omlx.ai/benchmarks/performance/06mhnvzq",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-fdp91wrb-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "Alis build; oMLX community benchmark trace; peak memory 246.8 GB at this context; duplicate traces of the same session: lg48st6x, mr3r4332, xmbu727z; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 478.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "GLM-5.3-Flash-Alis-MLX-6bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 1k 478.1 25.1 246.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fdp91wrb",
   "source_url": "https://omlx.ai/benchmarks/performance/fdp91wrb",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-fdp91wrb-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "Alis build; oMLX community benchmark trace; peak memory 248.4 GB at this context; duplicate traces of the same session: lg48st6x, mr3r4332, xmbu727z; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 543.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "GLM-5.3-Flash-Alis-MLX-6bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 4k 543.9 27.4 248.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fdp91wrb",
   "source_url": "https://omlx.ai/benchmarks/performance/fdp91wrb",
   "tg_tokens": null,
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-fdp91wrb-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "Alis build; oMLX community benchmark trace; peak memory 249.1 GB at this context; duplicate traces of the same session: lg48st6x, mr3r4332, xmbu727z; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 528.2,
   "provenance": "community",
   "quant": "6bit",
   "quote": "GLM-5.3-Flash-Alis-MLX-6bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 8k 528.2 35.1 249.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fdp91wrb",
   "source_url": "https://omlx.ai/benchmarks/performance/fdp91wrb",
   "tg_tokens": null,
   "tps": 35.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-fdp91wrb-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "Alis build; oMLX community benchmark trace; peak memory 249.8 GB at this context; duplicate traces of the same session: lg48st6x, mr3r4332, xmbu727z; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 519.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "GLM-5.3-Flash-Alis-MLX-6bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 16k 519.5 40.6 249.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fdp91wrb",
   "source_url": "https://omlx.ai/benchmarks/performance/fdp91wrb",
   "tg_tokens": null,
   "tps": 40.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-e5o5pios-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.5 GB at this context; duplicate traces of the same session: tshe7lwb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 394.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 394.3 27.6 175.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e5o5pios",
   "source_url": "https://omlx.ai/benchmarks/performance/e5o5pios",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-e5o5pios-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.6 GB at this context; duplicate traces of the same session: tshe7lwb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 402.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 402.8 26.5 175.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e5o5pios",
   "source_url": "https://omlx.ai/benchmarks/performance/e5o5pios",
   "tg_tokens": null,
   "tps": 26.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-e5o5pios-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 176.0 GB at this context; duplicate traces of the same session: tshe7lwb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 397.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 397.0 25.0 176.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e5o5pios",
   "source_url": "https://omlx.ai/benchmarks/performance/e5o5pios",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-e5o5pios-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 176.7 GB at this context; duplicate traces of the same session: tshe7lwb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 388.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 128k 388.8 21.5 176.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e5o5pios",
   "source_url": "https://omlx.ai/benchmarks/performance/e5o5pios",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-s0s1iqgm-ctx1024",
   "model": "KAT-Coder-V2.5-Dev",
   "notes": "MTP build; oMLX community benchmark trace, single run at 1k context; peak memory 21.8 GB; TTFT 480 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2133.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "KAT-Coder-V2.5-Dev-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k tokens 2,133 PP tok/s 48.1 TG tok/s 480 TTFT (ms) 21.8 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s0s1iqgm",
   "source_url": "https://omlx.ai/benchmarks/performance/s0s1iqgm",
   "tg_tokens": null,
   "tps": 48.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-4wesagw6-ctx1024",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "fp16 attention; oMLX community benchmark trace; peak memory 8.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 366.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-MLX-oQ6e-fp16 on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 6bit; 1k 366.2 34.2 8.0 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4wesagw6",
   "source_url": "https://omlx.ai/benchmarks/performance/4wesagw6",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-4wesagw6-ctx8192",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "fp16 attention; oMLX community benchmark trace; peak memory 8.7 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 382.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-MLX-oQ6e-fp16 on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 6bit; 8k 382.7 32.9 8.7 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4wesagw6",
   "source_url": "https://omlx.ai/benchmarks/performance/4wesagw6",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-1b6lb5kt-ctx1024",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 8.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 358.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 358.1 45.0 8.4 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1b6lb5kt",
   "source_url": "https://omlx.ai/benchmarks/performance/1b6lb5kt",
   "tg_tokens": null,
   "tps": 45.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-1b6lb5kt-ctx8192",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 8.7 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 479.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 8k 479.9 40.1 8.7 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1b6lb5kt",
   "source_url": "https://omlx.ai/benchmarks/performance/1b6lb5kt",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ledc7gs6-ctx1024",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 5.2 GB at this context; duplicate traces of the same session: iw3lxzjb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1619.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "MiniCPM5-2B on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 unknown; 1k 1,619 50.2 5.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ledc7gs6",
   "source_url": "https://omlx.ai/benchmarks/performance/ledc7gs6",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ledc7gs6-ctx4096",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 5.5 GB at this context; duplicate traces of the same session: iw3lxzjb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1526.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "MiniCPM5-2B on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 unknown; 4k 1,526 47.8 5.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ledc7gs6",
   "source_url": "https://omlx.ai/benchmarks/performance/ledc7gs6",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-7qv02hq8-ctx1024",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 2.2 GB at this context; duplicate traces of the same session: ct33eimh, hn1mo7qb, su7rwtmi; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1565.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiniCPM5-2B-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 1,565 132.4 2.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7qv02hq8",
   "source_url": "https://omlx.ai/benchmarks/performance/7qv02hq8",
   "tg_tokens": null,
   "tps": 132.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-7qv02hq8-ctx4096",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 2.3 GB at this context; duplicate traces of the same session: ct33eimh, hn1mo7qb, su7rwtmi; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1423.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiniCPM5-2B-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,423 116.2 2.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7qv02hq8",
   "source_url": "https://omlx.ai/benchmarks/performance/7qv02hq8",
   "tg_tokens": null,
   "tps": 116.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-7qv02hq8-ctx8192",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 2.4 GB at this context; duplicate traces of the same session: ct33eimh, hn1mo7qb, su7rwtmi; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1181.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiniCPM5-2B-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 1,181 105.2 2.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7qv02hq8",
   "source_url": "https://omlx.ai/benchmarks/performance/7qv02hq8",
   "tg_tokens": null,
   "tps": 105.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-7qv02hq8-ctx16384",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 2.7 GB at this context; duplicate traces of the same session: ct33eimh, hn1mo7qb, su7rwtmi; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 886.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiniCPM5-2B-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 886.3 89.3 2.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7qv02hq8",
   "source_url": "https://omlx.ai/benchmarks/performance/7qv02hq8",
   "tg_tokens": null,
   "tps": 89.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-7qv02hq8-ctx32768",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 3.5 GB at this context; duplicate traces of the same session: ct33eimh, hn1mo7qb, su7rwtmi; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 621.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiniCPM5-2B-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 621.7 65.8 3.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7qv02hq8",
   "source_url": "https://omlx.ai/benchmarks/performance/7qv02hq8",
   "tg_tokens": null,
   "tps": 65.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ebe51d40-ctx4096",
   "model": "Nex-N2.5-mini-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 1cj2cp5a; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 858.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Nex-N2.5-mini-Uncensored-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 858.4 56.6 22.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ebe51d40",
   "source_url": "https://omlx.ai/benchmarks/performance/ebe51d40",
   "tg_tokens": null,
   "tps": 56.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ebe51d40-ctx16384",
   "model": "Nex-N2.5-mini-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: 1cj2cp5a; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 723.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Nex-N2.5-mini-Uncensored-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 723.7 51.1 22.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ebe51d40",
   "source_url": "https://omlx.ai/benchmarks/performance/ebe51d40",
   "tg_tokens": null,
   "tps": 51.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-oxev93j3-ctx8192",
   "model": "Ornith-1.0-9B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1553.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.0-9B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 1,553 68.2 12.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oxev93j3",
   "source_url": "https://omlx.ai/benchmarks/performance/oxev93j3",
   "tg_tokens": null,
   "tps": 68.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-oxev93j3-ctx32768",
   "model": "Ornith-1.0-9B",
   "notes": "oMLX community benchmark trace; peak memory 13.8 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1342.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.0-9B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 32k 1,342 61.5 13.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oxev93j3",
   "source_url": "https://omlx.ai/benchmarks/performance/oxev93j3",
   "tg_tokens": null,
   "tps": 61.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-dhe5hhgr-ctx8192",
   "model": "Ornith-1.0-9B",
   "notes": "oMLX community benchmark trace; peak memory 13.0 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1437.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.0-9B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 1,437 62.4 13.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dhe5hhgr",
   "source_url": "https://omlx.ai/benchmarks/performance/dhe5hhgr",
   "tg_tokens": null,
   "tps": 62.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-dhe5hhgr-ctx32768",
   "model": "Ornith-1.0-9B",
   "notes": "oMLX community benchmark trace; peak memory 14.2 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1301.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.0-9B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 32k 1,301 52.5 14.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dhe5hhgr",
   "source_url": "https://omlx.ai/benchmarks/performance/dhe5hhgr",
   "tg_tokens": null,
   "tps": 52.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-rzklr86h-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2044.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 2,044 113.0 22.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rzklr86h",
   "source_url": "https://omlx.ai/benchmarks/performance/rzklr86h",
   "tg_tokens": null,
   "tps": 113.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-rzklr86h-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1703.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,703 106.4 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rzklr86h",
   "source_url": "https://omlx.ai/benchmarks/performance/rzklr86h",
   "tg_tokens": null,
   "tps": 106.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-rzklr86h-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 940.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 940.7 75.5 24.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rzklr86h",
   "source_url": "https://omlx.ai/benchmarks/performance/rzklr86h",
   "tg_tokens": null,
   "tps": 75.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-rzklr86h-ctx199680",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 365.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 195k 365.4 26.6 27.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rzklr86h",
   "source_url": "https://omlx.ai/benchmarks/performance/rzklr86h",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bkugem8q-ctx65536",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 96.8 GB at this context; duplicate traces of the same session: y9u05uw2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1209.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k 1,209 39.7 96.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bkugem8q",
   "source_url": "https://omlx.ai/benchmarks/performance/bkugem8q",
   "tg_tokens": null,
   "tps": 39.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bkugem8q-ctx131072",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 98.1 GB at this context; duplicate traces of the same session: y9u05uw2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 881.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 128k 881.3 31.3 98.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bkugem8q",
   "source_url": "https://omlx.ai/benchmarks/performance/bkugem8q",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bkugem8q-ctx199680",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 99.7 GB at this context; duplicate traces of the same session: y9u05uw2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 688.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 195k 688.3 26.1 99.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bkugem8q",
   "source_url": "https://omlx.ai/benchmarks/performance/bkugem8q",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-vinifbzt-ctx65536",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 96.8 GB at this context; duplicate traces of the same session: n7qkkjl7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1201.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k 1,201 39.3 96.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vinifbzt",
   "source_url": "https://omlx.ai/benchmarks/performance/vinifbzt",
   "tg_tokens": null,
   "tps": 39.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-vinifbzt-ctx131072",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 98.1 GB at this context; duplicate traces of the same session: n7qkkjl7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 885.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 128k 885.5 32.1 98.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vinifbzt",
   "source_url": "https://omlx.ai/benchmarks/performance/vinifbzt",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-vinifbzt-ctx199680",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 99.7 GB at this context; duplicate traces of the same session: n7qkkjl7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 690.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 195k 690.3 26.3 99.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vinifbzt",
   "source_url": "https://omlx.ai/benchmarks/performance/vinifbzt",
   "tg_tokens": null,
   "tps": 26.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-edrct5yz-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "48GB RAM variant; oMLX community benchmark trace, single run at 1k context; peak memory 44.9 GB; TTFT 1933 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 529.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B-RAM-48GB-MLX on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 529.8 PP tok/s 44.5 TG tok/s 1933 TTFT (ms) 44.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/edrct5yz",
   "source_url": "https://omlx.ai/benchmarks/performance/edrct5yz",
   "tg_tokens": null,
   "tps": 44.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qbz7e0al-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1666.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 1,666 58.6 100.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qbz7e0al",
   "source_url": "https://omlx.ai/benchmarks/performance/qbz7e0al",
   "tg_tokens": null,
   "tps": 58.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qbz7e0al-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1657.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 1,657 64.9 100.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qbz7e0al",
   "source_url": "https://omlx.ai/benchmarks/performance/qbz7e0al",
   "tg_tokens": null,
   "tps": 64.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qbz7e0al-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1574.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 1,574 60.5 101.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qbz7e0al",
   "source_url": "https://omlx.ai/benchmarks/performance/qbz7e0al",
   "tg_tokens": null,
   "tps": 60.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qbz7e0al-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1407.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 1,407 55.9 101.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qbz7e0al",
   "source_url": "https://omlx.ai/benchmarks/performance/qbz7e0al",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wk6mv5n4-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1672.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 1,672 63.6 100.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wk6mv5n4",
   "source_url": "https://omlx.ai/benchmarks/performance/wk6mv5n4",
   "tg_tokens": null,
   "tps": 63.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wk6mv5n4-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1654.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 1,654 61.0 100.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wk6mv5n4",
   "source_url": "https://omlx.ai/benchmarks/performance/wk6mv5n4",
   "tg_tokens": null,
   "tps": 61.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wk6mv5n4-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1573.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 1,573 60.5 101.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wk6mv5n4",
   "source_url": "https://omlx.ai/benchmarks/performance/wk6mv5n4",
   "tg_tokens": null,
   "tps": 60.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wk6mv5n4-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1411.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 1,411 56.7 101.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wk6mv5n4",
   "source_url": "https://omlx.ai/benchmarks/performance/wk6mv5n4",
   "tg_tokens": null,
   "tps": 56.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-185q07hf-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1665.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 1,665 60.4 100.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/185q07hf",
   "source_url": "https://omlx.ai/benchmarks/performance/185q07hf",
   "tg_tokens": null,
   "tps": 60.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-185q07hf-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1647.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 1,647 64.0 100.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/185q07hf",
   "source_url": "https://omlx.ai/benchmarks/performance/185q07hf",
   "tg_tokens": null,
   "tps": 64.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-185q07hf-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1578.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 1,578 61.8 101.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/185q07hf",
   "source_url": "https://omlx.ai/benchmarks/performance/185q07hf",
   "tg_tokens": null,
   "tps": 61.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-185q07hf-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1411.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 1,411 52.2 101.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/185q07hf",
   "source_url": "https://omlx.ai/benchmarks/performance/185q07hf",
   "tg_tokens": null,
   "tps": 52.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-no8zb9r7-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1699.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 1,699 51.2 96.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/no8zb9r7",
   "source_url": "https://omlx.ai/benchmarks/performance/no8zb9r7",
   "tg_tokens": null,
   "tps": 51.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-no8zb9r7-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1701.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 1,701 48.9 96.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/no8zb9r7",
   "source_url": "https://omlx.ai/benchmarks/performance/no8zb9r7",
   "tg_tokens": null,
   "tps": 48.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-no8zb9r7-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1624.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 1,624 47.0 96.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/no8zb9r7",
   "source_url": "https://omlx.ai/benchmarks/performance/no8zb9r7",
   "tg_tokens": null,
   "tps": 47.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-no8zb9r7-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1463.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 1,463 44.5 96.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/no8zb9r7",
   "source_url": "https://omlx.ai/benchmarks/performance/no8zb9r7",
   "tg_tokens": null,
   "tps": 44.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-sr3hr72s-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1700.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 1,700 51.1 96.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sr3hr72s",
   "source_url": "https://omlx.ai/benchmarks/performance/sr3hr72s",
   "tg_tokens": null,
   "tps": 51.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-sr3hr72s-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1694.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 1,694 48.3 96.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sr3hr72s",
   "source_url": "https://omlx.ai/benchmarks/performance/sr3hr72s",
   "tg_tokens": null,
   "tps": 48.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-sr3hr72s-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1627.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 1,627 46.0 96.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sr3hr72s",
   "source_url": "https://omlx.ai/benchmarks/performance/sr3hr72s",
   "tg_tokens": null,
   "tps": 46.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-sr3hr72s-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1457.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 1,457 44.3 96.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sr3hr72s",
   "source_url": "https://omlx.ai/benchmarks/performance/sr3hr72s",
   "tg_tokens": null,
   "tps": 44.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3yn3nds4-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.2 GB at this context; duplicate traces of the same session: a4w8ebx1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1703.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 1,703 51.1 96.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3yn3nds4",
   "source_url": "https://omlx.ai/benchmarks/performance/3yn3nds4",
   "tg_tokens": null,
   "tps": 51.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3yn3nds4-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.1 GB at this context; duplicate traces of the same session: a4w8ebx1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1703.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 1,703 48.6 96.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3yn3nds4",
   "source_url": "https://omlx.ai/benchmarks/performance/3yn3nds4",
   "tg_tokens": null,
   "tps": 48.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3yn3nds4-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.3 GB at this context; duplicate traces of the same session: a4w8ebx1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1628.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 1,628 45.6 96.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3yn3nds4",
   "source_url": "https://omlx.ai/benchmarks/performance/3yn3nds4",
   "tg_tokens": null,
   "tps": 45.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3yn3nds4-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.7 GB at this context; duplicate traces of the same session: a4w8ebx1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1470.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 1,470 44.6 96.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3yn3nds4",
   "source_url": "https://omlx.ai/benchmarks/performance/3yn3nds4",
   "tg_tokens": null,
   "tps": 44.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-omsk8hmv-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.6 GB at this context; duplicate traces of the same session: t6qyc68l; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 391.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 391.0 34.5 3.6 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/omsk8hmv",
   "source_url": "https://omlx.ai/benchmarks/performance/omsk8hmv",
   "tg_tokens": null,
   "tps": 34.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-omsk8hmv-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.8 GB at this context; duplicate traces of the same session: t6qyc68l; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 527.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 527.9 32.9 3.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/omsk8hmv",
   "source_url": "https://omlx.ai/benchmarks/performance/omsk8hmv",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-peav03wz-ctx32768",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 5.2 GB; TTFT 82847 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 395.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 32k tokens 395.5 PP tok/s 22.1 TG tok/s 82847 TTFT (ms) 5.2 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/peav03wz",
   "source_url": "https://omlx.ai/benchmarks/performance/peav03wz",
   "tg_tokens": null,
   "tps": 22.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-shs6urx9-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 4.4 GB; TTFT 35325 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 463.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 16k tokens 463.8 PP tok/s 27.1 TG tok/s 35325 TTFT (ms) 4.4 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/shs6urx9",
   "source_url": "https://omlx.ai/benchmarks/performance/shs6urx9",
   "tg_tokens": null,
   "tps": 27.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-wbglqtak-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 3.8 GB; TTFT 10150 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 403.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k tokens 403.6 PP tok/s 32.8 TG tok/s 10150 TTFT (ms) 3.8 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wbglqtak",
   "source_url": "https://omlx.ai/benchmarks/performance/wbglqtak",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-skwz7s1n-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 3.8 GB; TTFT 7843 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 522.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k tokens 522.3 PP tok/s 32.6 TG tok/s 7843 TTFT (ms) 3.8 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/skwz7s1n",
   "source_url": "https://omlx.ai/benchmarks/performance/skwz7s1n",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-09ntp2z5-ctx4096",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 97.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-27B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 97.7 18.9 19.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/09ntp2z5",
   "source_url": "https://omlx.ai/benchmarks/performance/09ntp2z5",
   "tg_tokens": null,
   "tps": 18.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-09ntp2z5-ctx16384",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 93.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-27B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 16k 93.8 17.6 21.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/09ntp2z5",
   "source_url": "https://omlx.ai/benchmarks/performance/09ntp2z5",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-uwd6wa0h-ctx1024",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: hzv2a7ii; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 261.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 261.3 29.2 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uwd6wa0h",
   "source_url": "https://omlx.ai/benchmarks/performance/uwd6wa0h",
   "tg_tokens": null,
   "tps": 29.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-uwd6wa0h-ctx4096",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: hzv2a7ii; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 232.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 232.9 26.8 19.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uwd6wa0h",
   "source_url": "https://omlx.ai/benchmarks/performance/uwd6wa0h",
   "tg_tokens": null,
   "tps": 26.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-q7mh0rff-ctx1024",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 5xs8v2lb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 255.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-27B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 255.8 16.2 28.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q7mh0rff",
   "source_url": "https://omlx.ai/benchmarks/performance/q7mh0rff",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-q7mh0rff-ctx4096",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.7 GB at this context; duplicate traces of the same session: 5xs8v2lb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 213.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-27B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 213.1 15.5 31.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q7mh0rff",
   "source_url": "https://omlx.ai/benchmarks/performance/q7mh0rff",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-edb17jze-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 694.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 694.3 84.8 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/edb17jze",
   "source_url": "https://omlx.ai/benchmarks/performance/edb17jze",
   "tg_tokens": null,
   "tps": 84.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-edb17jze-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 606.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 16k 606.5 74.1 22.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/edb17jze",
   "source_url": "https://omlx.ai/benchmarks/performance/edb17jze",
   "tg_tokens": null,
   "tps": 74.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-hjxzxky0-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.1 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 658.5,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 4k 658.5 39.6 29.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hjxzxky0",
   "source_url": "https://omlx.ai/benchmarks/performance/hjxzxky0",
   "tg_tokens": null,
   "tps": 39.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M1 Max",
   "id": "omlx-hjxzxky0-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.5 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 412.4,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 32k 412.4 8.5 30.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hjxzxky0",
   "source_url": "https://omlx.ai/benchmarks/performance/hjxzxky0",
   "tg_tokens": null,
   "tps": 8.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-047aw1n7-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 64k context; peak memory 40.4 GB; TTFT 200073 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 327.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 64k tokens 327.6 PP tok/s 37.1 TG tok/s 200073 TTFT (ms) 40.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/047aw1n7",
   "source_url": "https://omlx.ai/benchmarks/performance/047aw1n7",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tzfu54i7-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: s22czx4n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1018.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,018 33.2 17.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tzfu54i7",
   "source_url": "https://omlx.ai/benchmarks/performance/tzfu54i7",
   "tg_tokens": null,
   "tps": 33.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tzfu54i7-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: s22czx4n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 992.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 992.4 31.2 18.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tzfu54i7",
   "source_url": "https://omlx.ai/benchmarks/performance/tzfu54i7",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-b1zlb2ay-ctx8192",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "MPTLS optimized speed build; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: z11pu0jw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 338.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Abliterated-MTPLX-Optimized-Speed on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 338.0 53.8 28.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b1zlb2ay",
   "source_url": "https://omlx.ai/benchmarks/performance/b1zlb2ay",
   "tg_tokens": null,
   "tps": 53.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-b1zlb2ay-ctx32768",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "MPTLS optimized speed build; oMLX community benchmark trace; peak memory 31.1 GB at this context; duplicate traces of the same session: z11pu0jw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 299.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Abliterated-MTPLX-Optimized-Speed on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 299.9 69.6 31.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b1zlb2ay",
   "source_url": "https://omlx.ai/benchmarks/performance/b1zlb2ay",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ehefo2ee-ctx8192",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "MPTLS optimized speed build; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: 7t2rx9uu; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 343.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Abliterated-MTPLX-Optimized-Speed on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 343.2 31.1 24.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ehefo2ee",
   "source_url": "https://omlx.ai/benchmarks/performance/ehefo2ee",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ehefo2ee-ctx32768",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "MPTLS optimized speed build; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: 7t2rx9uu; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 309.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Abliterated-MTPLX-Optimized-Speed on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 309.2 27.7 26.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ehefo2ee",
   "source_url": "https://omlx.ai/benchmarks/performance/ehefo2ee",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-3iqllwal-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 42.9 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 440.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 440.6 37.9 42.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3iqllwal",
   "source_url": "https://omlx.ai/benchmarks/performance/3iqllwal",
   "tg_tokens": null,
   "tps": 37.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3iqllwal-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 45.4 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 380.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 32k 380.1 62.4 45.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3iqllwal",
   "source_url": "https://omlx.ai/benchmarks/performance/3iqllwal",
   "tg_tokens": null,
   "tps": 62.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-d07wfsxo-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 217.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-fp16-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 217.6 47.5 22.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d07wfsxo",
   "source_url": "https://omlx.ai/benchmarks/performance/d07wfsxo",
   "tg_tokens": null,
   "tps": 47.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-d07wfsxo-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 85.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-fp16-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 85.8 39.9 25.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d07wfsxo",
   "source_url": "https://omlx.ai/benchmarks/performance/d07wfsxo",
   "tg_tokens": null,
   "tps": 39.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-m1u78jiq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: 3vq1q0gy, rmm80x3i, s9p8o4ub, vt0ioyhc; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1146.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 1k 1,146 80.1 34.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m1u78jiq",
   "source_url": "https://omlx.ai/benchmarks/performance/m1u78jiq",
   "tg_tokens": null,
   "tps": 80.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-m1u78jiq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: 3vq1q0gy, rmm80x3i, s9p8o4ub, vt0ioyhc; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1401.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 4k 1,401 69.7 35.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m1u78jiq",
   "source_url": "https://omlx.ai/benchmarks/performance/m1u78jiq",
   "tg_tokens": null,
   "tps": 69.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-m1u78jiq-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: 3vq1q0gy, rmm80x3i, s9p8o4ub, vt0ioyhc; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1404.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 8k 1,404 80.4 36.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m1u78jiq",
   "source_url": "https://omlx.ai/benchmarks/performance/m1u78jiq",
   "tg_tokens": null,
   "tps": 80.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-m1u78jiq-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 36.6 GB at this context; duplicate traces of the same session: 3vq1q0gy, rmm80x3i, s9p8o4ub, vt0ioyhc; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1365.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 16k 1,365 91.7 36.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m1u78jiq",
   "source_url": "https://omlx.ai/benchmarks/performance/m1u78jiq",
   "tg_tokens": null,
   "tps": 91.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-m1u78jiq-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 37.8 GB at this context; duplicate traces of the same session: 3vq1q0gy, rmm80x3i, s9p8o4ub, vt0ioyhc; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1275.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 32k 1,275 93.1 37.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m1u78jiq",
   "source_url": "https://omlx.ai/benchmarks/performance/m1u78jiq",
   "tg_tokens": null,
   "tps": 93.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-m1u78jiq-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 40.1 GB at this context; duplicate traces of the same session: 3vq1q0gy, rmm80x3i, s9p8o4ub, vt0ioyhc; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1119.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 64k 1,119 72.5 40.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m1u78jiq",
   "source_url": "https://omlx.ai/benchmarks/performance/m1u78jiq",
   "tg_tokens": null,
   "tps": 72.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-fh670599-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.0 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 440.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 440.1 55.1 43.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fh670599",
   "source_url": "https://omlx.ai/benchmarks/performance/fh670599",
   "tg_tokens": null,
   "tps": 55.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-fh670599-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 45.4 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 380.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 32k 380.8 62.6 45.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fh670599",
   "source_url": "https://omlx.ai/benchmarks/performance/fh670599",
   "tg_tokens": null,
   "tps": 62.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-z2o06ccn-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.0 GB at this context; duplicate traces of the same session: j3u9q2t7; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 432.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 432.8 33.4 43.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z2o06ccn",
   "source_url": "https://omlx.ai/benchmarks/performance/z2o06ccn",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-z2o06ccn-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 45.4 GB at this context; duplicate traces of the same session: j3u9q2t7; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 378.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 32k 378.6 34.1 45.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z2o06ccn",
   "source_url": "https://omlx.ai/benchmarks/performance/z2o06ccn",
   "tg_tokens": null,
   "tps": 34.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2x6qg0dh-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.2 GB at this context; duplicate traces of the same session: 1wdzlgec; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 749.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 749.1 42.3 39.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2x6qg0dh",
   "source_url": "https://omlx.ai/benchmarks/performance/2x6qg0dh",
   "tg_tokens": null,
   "tps": 42.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2x6qg0dh-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.8 GB at this context; duplicate traces of the same session: 1wdzlgec; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 597.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 597.5 39.1 43.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2x6qg0dh",
   "source_url": "https://omlx.ai/benchmarks/performance/2x6qg0dh",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2x6qg0dh-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 49.0 GB at this context; duplicate traces of the same session: 1wdzlgec; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 491.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 195k 491.9 26.9 49.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2x6qg0dh",
   "source_url": "https://omlx.ai/benchmarks/performance/2x6qg0dh",
   "tg_tokens": null,
   "tps": 26.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-nwcq2zm4-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.2 GB at this context; duplicate traces of the same session: 5rtd0h9q, h4tt3b62; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 738.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 738.6 42.3 39.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nwcq2zm4",
   "source_url": "https://omlx.ai/benchmarks/performance/nwcq2zm4",
   "tg_tokens": null,
   "tps": 42.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-nwcq2zm4-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.8 GB at this context; duplicate traces of the same session: 5rtd0h9q, h4tt3b62; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 599.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 599.3 41.6 43.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nwcq2zm4",
   "source_url": "https://omlx.ai/benchmarks/performance/nwcq2zm4",
   "tg_tokens": null,
   "tps": 41.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nwcq2zm4-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 49.0 GB at this context; duplicate traces of the same session: 5rtd0h9q, h4tt3b62; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 493.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 195k 493.0 27.0 49.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nwcq2zm4",
   "source_url": "https://omlx.ai/benchmarks/performance/nwcq2zm4",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-x9pxv48h-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.3 GB at this context; duplicate traces of the same session: d0j5uvk2, qr20h47d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 431.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 431.8 52.9 69.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x9pxv48h",
   "source_url": "https://omlx.ai/benchmarks/performance/x9pxv48h",
   "tg_tokens": null,
   "tps": 52.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-x9pxv48h-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.9 GB at this context; duplicate traces of the same session: d0j5uvk2, qr20h47d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 484.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 484.6 48.7 69.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x9pxv48h",
   "source_url": "https://omlx.ai/benchmarks/performance/x9pxv48h",
   "tg_tokens": null,
   "tps": 48.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-x9pxv48h-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.9 GB at this context; duplicate traces of the same session: d0j5uvk2, qr20h47d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 633.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 633.1 53.5 69.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x9pxv48h",
   "source_url": "https://omlx.ai/benchmarks/performance/x9pxv48h",
   "tg_tokens": null,
   "tps": 53.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-x9pxv48h-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.1 GB at this context; duplicate traces of the same session: d0j5uvk2, qr20h47d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 580.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 580.7 54.5 70.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x9pxv48h",
   "source_url": "https://omlx.ai/benchmarks/performance/x9pxv48h",
   "tg_tokens": null,
   "tps": 54.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-xiexi38t-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 1k context; peak memory 22.5 GB; TTFT 1741 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 588.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k tokens 588.1 PP tok/s 34.8 TG tok/s 1741 TTFT (ms) 22.5 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xiexi38t",
   "source_url": "https://omlx.ai/benchmarks/performance/xiexi38t",
   "tg_tokens": null,
   "tps": 34.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-expdcumt-ctx1024",
   "model": "clef",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 392fmqgn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 248.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "clef-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 248.6 16.1 28.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/expdcumt",
   "source_url": "https://omlx.ai/benchmarks/performance/expdcumt",
   "tg_tokens": null,
   "tps": 16.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-expdcumt-ctx4096",
   "model": "clef",
   "notes": "oMLX community benchmark trace; peak memory 31.7 GB at this context; duplicate traces of the same session: 392fmqgn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 199.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "clef-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 199.9 14.6 31.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/expdcumt",
   "source_url": "https://omlx.ai/benchmarks/performance/expdcumt",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0rpwu95t-ctx4096",
   "model": "clef-flash",
   "notes": "oMLX community benchmark trace; peak memory 10.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3148.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "clef-flash-8bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 8bit; 4k 3,148 57.2 10.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0rpwu95t",
   "source_url": "https://omlx.ai/benchmarks/performance/0rpwu95t",
   "tg_tokens": null,
   "tps": 57.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0rpwu95t-ctx16384",
   "model": "clef-flash",
   "notes": "oMLX community benchmark trace; peak memory 10.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3140.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "clef-flash-8bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 8bit; 16k 3,140 55.1 10.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0rpwu95t",
   "source_url": "https://omlx.ai/benchmarks/performance/0rpwu95t",
   "tg_tokens": null,
   "tps": 55.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-g9xkyssw-ctx1024",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 7.1 GB at this context; duplicate traces of the same session: 0c4ytr08; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 503.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-MLX-4bit on M5 (10c); M5 (10c) \u00b7 16 GB \u00b7 4bit; 1k 503.5 16.5 7.1 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g9xkyssw",
   "source_url": "https://omlx.ai/benchmarks/performance/g9xkyssw",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-g9xkyssw-ctx4096",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 7.5 GB at this context; duplicate traces of the same session: 0c4ytr08; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 506.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-MLX-4bit on M5 (10c); M5 (10c) \u00b7 16 GB \u00b7 4bit; 4k 506.8 16.5 7.5 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g9xkyssw",
   "source_url": "https://omlx.ai/benchmarks/performance/g9xkyssw",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-43v1b0sr-ctx131072",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace, single run at 128k context; peak memory 22.4 GB; TTFT 216682 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 604.9,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "gemma-4-26B-A4B-it-qat-nvfp4 on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 128k tokens 604.9 PP tok/s 19.2 TG tok/s 216682 TTFT (ms) 22.4 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/43v1b0sr",
   "source_url": "https://omlx.ai/benchmarks/performance/43v1b0sr",
   "tg_tokens": null,
   "tps": 19.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4i1gfl9r-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2063.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.5-4B-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 2,063 170.1 3.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4i1gfl9r",
   "source_url": "https://omlx.ai/benchmarks/performance/4i1gfl9r",
   "tg_tokens": null,
   "tps": 170.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4i1gfl9r-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.7 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2172.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.5-4B-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 2,172 163.6 4.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4i1gfl9r",
   "source_url": "https://omlx.ai/benchmarks/performance/4i1gfl9r",
   "tg_tokens": null,
   "tps": 163.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-a090ohsb-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.8 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1914.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.5-4B-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 1,914 163.6 3.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a090ohsb",
   "source_url": "https://omlx.ai/benchmarks/performance/a090ohsb",
   "tg_tokens": null,
   "tps": 163.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-a090ohsb-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.6 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2256.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.5-4B-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 2,256 159.3 4.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a090ohsb",
   "source_url": "https://omlx.ai/benchmarks/performance/a090ohsb",
   "tg_tokens": null,
   "tps": 159.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-dp47b749-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: wlab5sqe; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 217.4,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "mlx-community--Qwen3.8-27B-nvfp4 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 217.4 27.2 19.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dp47b749",
   "source_url": "https://omlx.ai/benchmarks/performance/dp47b749",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-dp47b749-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: wlab5sqe; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 202.7,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "mlx-community--Qwen3.8-27B-nvfp4 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 202.7 24.9 20.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dp47b749",
   "source_url": "https://omlx.ai/benchmarks/performance/dp47b749",
   "tg_tokens": null,
   "tps": 24.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-1pe7rxgz-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 105.9 15.5 18.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1pe7rxgz",
   "source_url": "https://omlx.ai/benchmarks/performance/1pe7rxgz",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-1pe7rxgz-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 150.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 150.8 14.7 19.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1pe7rxgz",
   "source_url": "https://omlx.ai/benchmarks/performance/1pe7rxgz",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-1pe7rxgz-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 115.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 115.0 14.5 19.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1pe7rxgz",
   "source_url": "https://omlx.ai/benchmarks/performance/1pe7rxgz",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-1pe7rxgz-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 96.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 96.0 12.8 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1pe7rxgz",
   "source_url": "https://omlx.ai/benchmarks/performance/1pe7rxgz",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-d2d84wd2-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 0w05qp6k; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 105.4 15.1 18.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d2d84wd2",
   "source_url": "https://omlx.ai/benchmarks/performance/d2d84wd2",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-d2d84wd2-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 0w05qp6k; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 134.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 134.7 15.0 19.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d2d84wd2",
   "source_url": "https://omlx.ai/benchmarks/performance/d2d84wd2",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-d2d84wd2-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: 0w05qp6k; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 151.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 151.5 14.2 19.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d2d84wd2",
   "source_url": "https://omlx.ai/benchmarks/performance/d2d84wd2",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-d2d84wd2-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 0w05qp6k; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 97.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 97.6 12.8 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d2d84wd2",
   "source_url": "https://omlx.ai/benchmarks/performance/d2d84wd2",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-gzyg4eih-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: p9txj8t4, zszu42ci; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 105.2 14.6 18.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gzyg4eih",
   "source_url": "https://omlx.ai/benchmarks/performance/gzyg4eih",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-gzyg4eih-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: p9txj8t4, zszu42ci; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 113.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 113.1 14.4 19.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gzyg4eih",
   "source_url": "https://omlx.ai/benchmarks/performance/gzyg4eih",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-gzyg4eih-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: p9txj8t4, zszu42ci; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 106.5 14.3 17.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gzyg4eih",
   "source_url": "https://omlx.ai/benchmarks/performance/gzyg4eih",
   "tg_tokens": null,
   "tps": 14.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-gzyg4eih-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: p9txj8t4, zszu42ci; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 96.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 96.0 12.7 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gzyg4eih",
   "source_url": "https://omlx.ai/benchmarks/performance/gzyg4eih",
   "tg_tokens": null,
   "tps": 12.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-q2pv67xn-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 6x0z3p0t, y9niffuu; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 105.4 15.0 18.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q2pv67xn",
   "source_url": "https://omlx.ai/benchmarks/performance/q2pv67xn",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-q2pv67xn-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 6x0z3p0t, y9niffuu; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 125.0 14.5 19.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q2pv67xn",
   "source_url": "https://omlx.ai/benchmarks/performance/q2pv67xn",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-q2pv67xn-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: 6x0z3p0t, y9niffuu; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 106.2 13.8 17.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q2pv67xn",
   "source_url": "https://omlx.ai/benchmarks/performance/q2pv67xn",
   "tg_tokens": null,
   "tps": 13.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-q2pv67xn-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 6x0z3p0t, y9niffuu; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 95.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 95.9 12.3 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q2pv67xn",
   "source_url": "https://omlx.ai/benchmarks/performance/q2pv67xn",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-04r2rykf-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: gvgg80gc, j15hd5nf, l6ailkmx; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 105.3 15.0 18.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/04r2rykf",
   "source_url": "https://omlx.ai/benchmarks/performance/04r2rykf",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-04r2rykf-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: gvgg80gc, j15hd5nf, l6ailkmx; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 150.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 150.8 14.5 19.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/04r2rykf",
   "source_url": "https://omlx.ai/benchmarks/performance/04r2rykf",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-04r2rykf-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: gvgg80gc, j15hd5nf, l6ailkmx; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 114.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 114.4 13.7 19.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/04r2rykf",
   "source_url": "https://omlx.ai/benchmarks/performance/04r2rykf",
   "tg_tokens": null,
   "tps": 13.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-04r2rykf-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: gvgg80gc, j15hd5nf, l6ailkmx; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 96.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 96.0 12.4 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/04r2rykf",
   "source_url": "https://omlx.ai/benchmarks/performance/04r2rykf",
   "tg_tokens": null,
   "tps": 12.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-8yipord4-ctx1024",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: fn9hwir2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 320.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 320.6 71.8 18.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8yipord4",
   "source_url": "https://omlx.ai/benchmarks/performance/8yipord4",
   "tg_tokens": null,
   "tps": 71.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-8yipord4-ctx4096",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: fn9hwir2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 345.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 345.2 65.3 22.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8yipord4",
   "source_url": "https://omlx.ai/benchmarks/performance/8yipord4",
   "tg_tokens": null,
   "tps": 65.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-8yipord4-ctx8192",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: fn9hwir2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 339.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 339.9 71.0 22.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8yipord4",
   "source_url": "https://omlx.ai/benchmarks/performance/8yipord4",
   "tg_tokens": null,
   "tps": 71.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-p7fs4kiu-ctx32768",
   "model": "dsv4-flash-vision-exp",
   "notes": "oMLX community benchmark trace; peak memory 150.1 GB at this context; duplicate traces of the same session: u5db2s02; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1142.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "dsv4-flash-vision-exp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 1,142 34.3 150.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p7fs4kiu",
   "source_url": "https://omlx.ai/benchmarks/performance/p7fs4kiu",
   "tg_tokens": null,
   "tps": 34.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-p7fs4kiu-ctx65536",
   "model": "dsv4-flash-vision-exp",
   "notes": "oMLX community benchmark trace; peak memory 150.3 GB at this context; duplicate traces of the same session: u5db2s02; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1126.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "dsv4-flash-vision-exp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 1,126 33.9 150.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p7fs4kiu",
   "source_url": "https://omlx.ai/benchmarks/performance/p7fs4kiu",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-p7fs4kiu-ctx131072",
   "model": "dsv4-flash-vision-exp",
   "notes": "oMLX community benchmark trace; peak memory 150.7 GB at this context; duplicate traces of the same session: u5db2s02; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1096.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "dsv4-flash-vision-exp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 1,096 33.3 150.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p7fs4kiu",
   "source_url": "https://omlx.ai/benchmarks/performance/p7fs4kiu",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-lg48st6x-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "Alis build; oMLX community benchmark trace; peak memory 246.8 GB at this context; duplicate traces of the same session: mr3r4332, xmbu727z; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 478.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "GLM-5.3-Flash-Alis-MLX-6bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 1k 478.1 25.1 246.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lg48st6x",
   "source_url": "https://omlx.ai/benchmarks/performance/lg48st6x",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lg48st6x-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "Alis build; oMLX community benchmark trace; peak memory 248.4 GB at this context; duplicate traces of the same session: mr3r4332, xmbu727z; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 543.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "GLM-5.3-Flash-Alis-MLX-6bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 4k 543.9 27.4 248.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lg48st6x",
   "source_url": "https://omlx.ai/benchmarks/performance/lg48st6x",
   "tg_tokens": null,
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-lg48st6x-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "Alis build; oMLX community benchmark trace; peak memory 249.1 GB at this context; duplicate traces of the same session: mr3r4332, xmbu727z; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 528.2,
   "provenance": "community",
   "quant": "6bit",
   "quote": "GLM-5.3-Flash-Alis-MLX-6bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 8k 528.2 35.1 249.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lg48st6x",
   "source_url": "https://omlx.ai/benchmarks/performance/lg48st6x",
   "tg_tokens": null,
   "tps": 35.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-lg48st6x-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "Alis build; oMLX community benchmark trace; peak memory 249.8 GB at this context; duplicate traces of the same session: mr3r4332, xmbu727z; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 519.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "GLM-5.3-Flash-Alis-MLX-6bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 16k 519.5 40.6 249.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lg48st6x",
   "source_url": "https://omlx.ai/benchmarks/performance/lg48st6x",
   "tg_tokens": null,
   "tps": 40.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-tshe7lwb-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.5 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 394.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 394.3 27.6 175.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tshe7lwb",
   "source_url": "https://omlx.ai/benchmarks/performance/tshe7lwb",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-tshe7lwb-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.6 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 402.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 402.8 26.5 175.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tshe7lwb",
   "source_url": "https://omlx.ai/benchmarks/performance/tshe7lwb",
   "tg_tokens": null,
   "tps": 26.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-tshe7lwb-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 176.0 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 397.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 397.0 25.0 176.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tshe7lwb",
   "source_url": "https://omlx.ai/benchmarks/performance/tshe7lwb",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-tshe7lwb-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 176.7 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 388.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 128k 388.8 21.5 176.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tshe7lwb",
   "source_url": "https://omlx.ai/benchmarks/performance/tshe7lwb",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-iw3lxzjb-ctx1024",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 5.2 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1619.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "MiniCPM5-2B on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 unknown; 1k 1,619 50.2 5.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iw3lxzjb",
   "source_url": "https://omlx.ai/benchmarks/performance/iw3lxzjb",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-iw3lxzjb-ctx4096",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 5.5 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1526.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "MiniCPM5-2B on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 unknown; 4k 1,526 47.8 5.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iw3lxzjb",
   "source_url": "https://omlx.ai/benchmarks/performance/iw3lxzjb",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ct33eimh-ctx1024",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 2.2 GB at this context; duplicate traces of the same session: hn1mo7qb, su7rwtmi; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1565.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiniCPM5-2B-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 1,565 132.4 2.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ct33eimh",
   "source_url": "https://omlx.ai/benchmarks/performance/ct33eimh",
   "tg_tokens": null,
   "tps": 132.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ct33eimh-ctx4096",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 2.3 GB at this context; duplicate traces of the same session: hn1mo7qb, su7rwtmi; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1423.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiniCPM5-2B-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,423 116.2 2.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ct33eimh",
   "source_url": "https://omlx.ai/benchmarks/performance/ct33eimh",
   "tg_tokens": null,
   "tps": 116.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ct33eimh-ctx8192",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 2.4 GB at this context; duplicate traces of the same session: hn1mo7qb, su7rwtmi; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1181.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiniCPM5-2B-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 1,181 105.2 2.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ct33eimh",
   "source_url": "https://omlx.ai/benchmarks/performance/ct33eimh",
   "tg_tokens": null,
   "tps": 105.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ct33eimh-ctx16384",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 2.7 GB at this context; duplicate traces of the same session: hn1mo7qb, su7rwtmi; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 886.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiniCPM5-2B-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 886.3 89.3 2.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ct33eimh",
   "source_url": "https://omlx.ai/benchmarks/performance/ct33eimh",
   "tg_tokens": null,
   "tps": 89.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ct33eimh-ctx32768",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 3.5 GB at this context; duplicate traces of the same session: hn1mo7qb, su7rwtmi; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 621.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiniCPM5-2B-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 621.7 65.8 3.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ct33eimh",
   "source_url": "https://omlx.ai/benchmarks/performance/ct33eimh",
   "tg_tokens": null,
   "tps": 65.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-1cj2cp5a-ctx4096",
   "model": "Nex-N2.5-mini-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 858.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Nex-N2.5-mini-Uncensored-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 858.4 56.6 22.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1cj2cp5a",
   "source_url": "https://omlx.ai/benchmarks/performance/1cj2cp5a",
   "tg_tokens": null,
   "tps": 56.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-1cj2cp5a-ctx16384",
   "model": "Nex-N2.5-mini-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 723.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Nex-N2.5-mini-Uncensored-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 723.7 51.1 22.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1cj2cp5a",
   "source_url": "https://omlx.ai/benchmarks/performance/1cj2cp5a",
   "tg_tokens": null,
   "tps": 51.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-e57twrl1-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.2 GB; TTFT 957 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1070.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-4bit-MLX on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 1,070 PP tok/s 54.3 TG tok/s 957 TTFT (ms) 19.2 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e57twrl1",
   "source_url": "https://omlx.ai/benchmarks/performance/e57twrl1",
   "tg_tokens": null,
   "tps": 54.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-ny3gd1qa-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 2r4vj7rw; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1095.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-4bit-MLX on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k 1,095 54.1 19.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ny3gd1qa",
   "source_url": "https://omlx.ai/benchmarks/performance/ny3gd1qa",
   "tg_tokens": null,
   "tps": 54.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-ny3gd1qa-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 2r4vj7rw; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1473.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-4bit-MLX on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 1,473 51.9 19.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ny3gd1qa",
   "source_url": "https://omlx.ai/benchmarks/performance/ny3gd1qa",
   "tg_tokens": null,
   "tps": 51.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-u0zup6y0-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: og74z87l, rtkk9h7t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2044.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 2,044 113.0 22.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u0zup6y0",
   "source_url": "https://omlx.ai/benchmarks/performance/u0zup6y0",
   "tg_tokens": null,
   "tps": 113.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-u0zup6y0-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: og74z87l, rtkk9h7t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1703.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,703 106.4 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u0zup6y0",
   "source_url": "https://omlx.ai/benchmarks/performance/u0zup6y0",
   "tg_tokens": null,
   "tps": 106.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-u0zup6y0-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: og74z87l, rtkk9h7t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 940.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 940.7 75.5 24.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u0zup6y0",
   "source_url": "https://omlx.ai/benchmarks/performance/u0zup6y0",
   "tg_tokens": null,
   "tps": 75.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-u0zup6y0-ctx199680",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: og74z87l, rtkk9h7t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 365.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 195k 365.4 26.6 27.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u0zup6y0",
   "source_url": "https://omlx.ai/benchmarks/performance/u0zup6y0",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-kee4wlz7-ctx1024",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 1.4 GB; TTFT 237 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4321.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-0.8B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 4,321 PP tok/s 213.5 TG tok/s 237 TTFT (ms) 1.4 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kee4wlz7",
   "source_url": "https://omlx.ai/benchmarks/performance/kee4wlz7",
   "tg_tokens": null,
   "tps": 213.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-y9u05uw2-ctx65536",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 96.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1209.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k 1,209 39.7 96.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y9u05uw2",
   "source_url": "https://omlx.ai/benchmarks/performance/y9u05uw2",
   "tg_tokens": null,
   "tps": 39.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-y9u05uw2-ctx131072",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 98.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 881.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 128k 881.3 31.3 98.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y9u05uw2",
   "source_url": "https://omlx.ai/benchmarks/performance/y9u05uw2",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-y9u05uw2-ctx199680",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 99.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 688.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 195k 688.3 26.1 99.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y9u05uw2",
   "source_url": "https://omlx.ai/benchmarks/performance/y9u05uw2",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-n7qkkjl7-ctx65536",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 96.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1201.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k 1,201 39.3 96.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n7qkkjl7",
   "source_url": "https://omlx.ai/benchmarks/performance/n7qkkjl7",
   "tg_tokens": null,
   "tps": 39.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-n7qkkjl7-ctx131072",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 98.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 885.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 128k 885.5 32.1 98.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n7qkkjl7",
   "source_url": "https://omlx.ai/benchmarks/performance/n7qkkjl7",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-n7qkkjl7-ctx199680",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 99.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 690.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 195k 690.3 26.3 99.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n7qkkjl7",
   "source_url": "https://omlx.ai/benchmarks/performance/n7qkkjl7",
   "tg_tokens": null,
   "tps": 26.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-a4w8ebx1-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1703.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 1,703 51.1 96.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a4w8ebx1",
   "source_url": "https://omlx.ai/benchmarks/performance/a4w8ebx1",
   "tg_tokens": null,
   "tps": 51.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-a4w8ebx1-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1703.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 1,703 48.6 96.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a4w8ebx1",
   "source_url": "https://omlx.ai/benchmarks/performance/a4w8ebx1",
   "tg_tokens": null,
   "tps": 48.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-a4w8ebx1-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1628.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 1,628 45.6 96.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a4w8ebx1",
   "source_url": "https://omlx.ai/benchmarks/performance/a4w8ebx1",
   "tg_tokens": null,
   "tps": 45.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-a4w8ebx1-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1470.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-122B-A10B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 1,470 44.6 96.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a4w8ebx1",
   "source_url": "https://omlx.ai/benchmarks/performance/a4w8ebx1",
   "tg_tokens": null,
   "tps": 44.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-t6qyc68l-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.6 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 391.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 391.0 34.5 3.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t6qyc68l",
   "source_url": "https://omlx.ai/benchmarks/performance/t6qyc68l",
   "tg_tokens": null,
   "tps": 34.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-t6qyc68l-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.8 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 527.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 527.9 32.9 3.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t6qyc68l",
   "source_url": "https://omlx.ai/benchmarks/performance/t6qyc68l",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-zwavcrc6-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: d50rwh2e; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 208.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 208.3 16.2 6.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zwavcrc6",
   "source_url": "https://omlx.ai/benchmarks/performance/zwavcrc6",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-zwavcrc6-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: d50rwh2e; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 201.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 16k 201.7 14.3 6.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zwavcrc6",
   "source_url": "https://omlx.ai/benchmarks/performance/zwavcrc6",
   "tg_tokens": null,
   "tps": 14.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-hzv2a7ii-ctx1024",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 261.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 261.3 29.2 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hzv2a7ii",
   "source_url": "https://omlx.ai/benchmarks/performance/hzv2a7ii",
   "tg_tokens": null,
   "tps": 29.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-hzv2a7ii-ctx4096",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 232.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 232.9 26.8 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hzv2a7ii",
   "source_url": "https://omlx.ai/benchmarks/performance/hzv2a7ii",
   "tg_tokens": null,
   "tps": 26.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-5xs8v2lb-ctx1024",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 255.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-27B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 255.8 16.2 28.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5xs8v2lb",
   "source_url": "https://omlx.ai/benchmarks/performance/5xs8v2lb",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-5xs8v2lb-ctx4096",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 213.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-27B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 213.1 15.5 31.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5xs8v2lb",
   "source_url": "https://omlx.ai/benchmarks/performance/5xs8v2lb",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-7ry3f72o-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: i91lha7a; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 877.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 877.4 60.5 35.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ry3f72o",
   "source_url": "https://omlx.ai/benchmarks/performance/7ry3f72o",
   "tg_tokens": null,
   "tps": 60.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-7ry3f72o-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: i91lha7a; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1030.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 1,030 59.0 37.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ry3f72o",
   "source_url": "https://omlx.ai/benchmarks/performance/7ry3f72o",
   "tg_tokens": null,
   "tps": 59.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-2wxvtklt-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.1 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 658.5,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 4k 658.5 39.6 29.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2wxvtklt",
   "source_url": "https://omlx.ai/benchmarks/performance/2wxvtklt",
   "tg_tokens": null,
   "tps": 39.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M1 Max",
   "id": "omlx-2wxvtklt-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.5 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 412.4,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 32k 412.4 8.5 30.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2wxvtklt",
   "source_url": "https://omlx.ai/benchmarks/performance/2wxvtklt",
   "tg_tokens": null,
   "tps": 8.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-y7qdw2oy-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.1 GB at this context; duplicate traces of the same session: k15rmmf8; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 660.6,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 4k 660.6 36.6 29.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y7qdw2oy",
   "source_url": "https://omlx.ai/benchmarks/performance/y7qdw2oy",
   "tg_tokens": null,
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M1 Max",
   "id": "omlx-y7qdw2oy-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.5 GB at this context; duplicate traces of the same session: k15rmmf8; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 404.9,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 32k 404.9 8.4 30.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y7qdw2oy",
   "source_url": "https://omlx.ai/benchmarks/performance/y7qdw2oy",
   "tg_tokens": null,
   "tps": 8.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-s22czx4n-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1018.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,018 33.2 17.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s22czx4n",
   "source_url": "https://omlx.ai/benchmarks/performance/s22czx4n",
   "tg_tokens": null,
   "tps": 33.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-s22czx4n-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 992.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 992.4 31.2 18.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s22czx4n",
   "source_url": "https://omlx.ai/benchmarks/performance/s22czx4n",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1377dgxh-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 4i7e2znm; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 392.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 392.3 18.2 16.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1377dgxh",
   "source_url": "https://omlx.ai/benchmarks/performance/1377dgxh",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1377dgxh-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 4i7e2znm; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 417.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 417.6 17.8 17.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1377dgxh",
   "source_url": "https://omlx.ai/benchmarks/performance/1377dgxh",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-z11pu0jw-ctx8192",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "MPTLS optimized speed build; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 338.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Abliterated-MTPLX-Optimized-Speed on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 338.0 53.8 28.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z11pu0jw",
   "source_url": "https://omlx.ai/benchmarks/performance/z11pu0jw",
   "tg_tokens": null,
   "tps": 53.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-z11pu0jw-ctx32768",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "MPTLS optimized speed build; oMLX community benchmark trace; peak memory 31.1 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 299.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Abliterated-MTPLX-Optimized-Speed on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 299.9 69.6 31.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z11pu0jw",
   "source_url": "https://omlx.ai/benchmarks/performance/z11pu0jw",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-7t2rx9uu-ctx8192",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "MPTLS optimized speed build; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 343.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Abliterated-MTPLX-Optimized-Speed on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 343.2 31.1 24.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7t2rx9uu",
   "source_url": "https://omlx.ai/benchmarks/performance/7t2rx9uu",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-7t2rx9uu-ctx32768",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "MPTLS optimized speed build; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 309.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Abliterated-MTPLX-Optimized-Speed on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 309.2 27.7 26.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7t2rx9uu",
   "source_url": "https://omlx.ai/benchmarks/performance/7t2rx9uu",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-fn9hwir2-ctx1024",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: emh3sv5c; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 320.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 320.6 71.8 18.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fn9hwir2",
   "source_url": "https://omlx.ai/benchmarks/performance/fn9hwir2",
   "tg_tokens": null,
   "tps": 71.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-fn9hwir2-ctx4096",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: emh3sv5c; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 345.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 345.2 65.3 22.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fn9hwir2",
   "source_url": "https://omlx.ai/benchmarks/performance/fn9hwir2",
   "tg_tokens": null,
   "tps": 65.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-fn9hwir2-ctx8192",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: emh3sv5c; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 339.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 339.9 71.0 22.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fn9hwir2",
   "source_url": "https://omlx.ai/benchmarks/performance/fn9hwir2",
   "tg_tokens": null,
   "tps": 71.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3gwatkfd-ctx1024",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: 9p4bjotg, asmcv08y, ytz2rl7v; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 313.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ6e-MTP-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 1k 313.4 55.7 26.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3gwatkfd",
   "source_url": "https://omlx.ai/benchmarks/performance/3gwatkfd",
   "tg_tokens": null,
   "tps": 55.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3gwatkfd-ctx4096",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 30.0 GB at this context; duplicate traces of the same session: 9p4bjotg, asmcv08y, ytz2rl7v; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 340.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ6e-MTP-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 4k 340.3 48.4 30.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3gwatkfd",
   "source_url": "https://omlx.ai/benchmarks/performance/3gwatkfd",
   "tg_tokens": null,
   "tps": 48.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3gwatkfd-ctx8192",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 30.5 GB at this context; duplicate traces of the same session: 9p4bjotg, asmcv08y, ytz2rl7v; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 330.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ6e-MTP-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 8k 330.6 60.1 30.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3gwatkfd",
   "source_url": "https://omlx.ai/benchmarks/performance/3gwatkfd",
   "tg_tokens": null,
   "tps": 60.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3gwatkfd-ctx16384",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 31.4 GB at this context; duplicate traces of the same session: 9p4bjotg, asmcv08y, ytz2rl7v; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 319.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ6e-MTP-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 16k 319.8 63.5 31.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3gwatkfd",
   "source_url": "https://omlx.ai/benchmarks/performance/3gwatkfd",
   "tg_tokens": null,
   "tps": 63.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-ceaiwisp-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS bare speed FP16 build; oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: dfjmy5l6, lsykoi7r, n94hmver, qu3u4w3m; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 231.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Bare-Speed-FP16 on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 1k 231.3 30.7 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ceaiwisp",
   "source_url": "https://omlx.ai/benchmarks/performance/ceaiwisp",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-ceaiwisp-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS bare speed FP16 build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: dfjmy5l6, lsykoi7r, n94hmver, qu3u4w3m; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 246.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Bare-Speed-FP16 on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 4k 246.4 29.9 19.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ceaiwisp",
   "source_url": "https://omlx.ai/benchmarks/performance/ceaiwisp",
   "tg_tokens": null,
   "tps": 29.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-ceaiwisp-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS bare speed FP16 build; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: dfjmy5l6, lsykoi7r, n94hmver, qu3u4w3m; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 240.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Bare-Speed-FP16 on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 8k 240.9 29.1 20.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ceaiwisp",
   "source_url": "https://omlx.ai/benchmarks/performance/ceaiwisp",
   "tg_tokens": null,
   "tps": 29.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-ceaiwisp-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS bare speed FP16 build; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: dfjmy5l6, lsykoi7r, n94hmver, qu3u4w3m; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 229.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Bare-Speed-FP16 on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 16k 229.8 27.3 21.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ceaiwisp",
   "source_url": "https://omlx.ai/benchmarks/performance/ceaiwisp",
   "tg_tokens": null,
   "tps": 27.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-ceaiwisp-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS bare speed FP16 build; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: dfjmy5l6, lsykoi7r, n94hmver, qu3u4w3m; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 209.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Bare-Speed-FP16 on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 32k 209.9 24.5 22.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ceaiwisp",
   "source_url": "https://omlx.ai/benchmarks/performance/ceaiwisp",
   "tg_tokens": null,
   "tps": 24.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-iv4h6wjp-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized speed FP16 build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 2ydj1h0q, bzphuyno, qp2f2en9, xpfma15c; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 230.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed-FP16 on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 1k 230.3 25.8 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iv4h6wjp",
   "source_url": "https://omlx.ai/benchmarks/performance/iv4h6wjp",
   "tg_tokens": null,
   "tps": 25.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-iv4h6wjp-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized speed FP16 build; oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: 2ydj1h0q, bzphuyno, qp2f2en9, xpfma15c; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 232.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed-FP16 on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 4k 232.9 25.4 23.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iv4h6wjp",
   "source_url": "https://omlx.ai/benchmarks/performance/iv4h6wjp",
   "tg_tokens": null,
   "tps": 25.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-iv4h6wjp-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized speed FP16 build; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: 2ydj1h0q, bzphuyno, qp2f2en9, xpfma15c; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 228.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed-FP16 on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 8k 228.0 24.9 24.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iv4h6wjp",
   "source_url": "https://omlx.ai/benchmarks/performance/iv4h6wjp",
   "tg_tokens": null,
   "tps": 24.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-iv4h6wjp-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized speed FP16 build; oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: 2ydj1h0q, bzphuyno, qp2f2en9, xpfma15c; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 218.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed-FP16 on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 16k 218.6 23.5 25.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iv4h6wjp",
   "source_url": "https://omlx.ai/benchmarks/performance/iv4h6wjp",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-iv4h6wjp-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized speed FP16 build; oMLX community benchmark trace; peak memory 26.9 GB at this context; duplicate traces of the same session: 2ydj1h0q, bzphuyno, qp2f2en9, xpfma15c; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 200.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed-FP16 on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 32k 200.8 21.9 26.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iv4h6wjp",
   "source_url": "https://omlx.ai/benchmarks/performance/iv4h6wjp",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dq9o3gpx-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 14.7 GB at this context; duplicate traces of the same session: 3x7l0mgl, ajcsh2vn, f2itb8w9, f2lblzjy, ybylsd7j, ysedkpbk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 799.5,
   "provenance": "community",
   "quant": "oQ3.5",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3.5 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 1k 799.5 32.5 14.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dq9o3gpx",
   "source_url": "https://omlx.ai/benchmarks/performance/dq9o3gpx",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dq9o3gpx-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 15.5 GB at this context; duplicate traces of the same session: 3x7l0mgl, ajcsh2vn, f2itb8w9, f2lblzjy, ybylsd7j, ysedkpbk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 770.2,
   "provenance": "community",
   "quant": "oQ3.5",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3.5 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 4k 770.2 32.5 15.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dq9o3gpx",
   "source_url": "https://omlx.ai/benchmarks/performance/dq9o3gpx",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dq9o3gpx-ctx8192",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: 3x7l0mgl, ajcsh2vn, f2itb8w9, f2lblzjy, ybylsd7j, ysedkpbk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 691.7,
   "provenance": "community",
   "quant": "oQ3.5",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3.5 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 8k 691.7 30.6 15.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dq9o3gpx",
   "source_url": "https://omlx.ai/benchmarks/performance/dq9o3gpx",
   "tg_tokens": null,
   "tps": 30.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dq9o3gpx-ctx16384",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: 3x7l0mgl, ajcsh2vn, f2itb8w9, f2lblzjy, ybylsd7j, ysedkpbk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 650.9,
   "provenance": "community",
   "quant": "oQ3.5",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3.5 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 16k 650.9 28.3 16.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dq9o3gpx",
   "source_url": "https://omlx.ai/benchmarks/performance/dq9o3gpx",
   "tg_tokens": null,
   "tps": 28.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dq9o3gpx-ctx32768",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 3x7l0mgl, ajcsh2vn, f2itb8w9, f2lblzjy, ybylsd7j, ysedkpbk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 604.0,
   "provenance": "community",
   "quant": "oQ3.5",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3.5 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 32k 604.0 25.7 17.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dq9o3gpx",
   "source_url": "https://omlx.ai/benchmarks/performance/dq9o3gpx",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dq9o3gpx-ctx65536",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 3x7l0mgl, ajcsh2vn, f2itb8w9, f2lblzjy, ybylsd7j, ysedkpbk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 540.1,
   "provenance": "community",
   "quant": "oQ3.5",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3.5 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 64k 540.1 22.8 19.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dq9o3gpx",
   "source_url": "https://omlx.ai/benchmarks/performance/dq9o3gpx",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dq9o3gpx-ctx131072",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 3x7l0mgl, ajcsh2vn, f2itb8w9, f2lblzjy, ybylsd7j, ysedkpbk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 448.0,
   "provenance": "community",
   "quant": "oQ3.5",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3.5 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 128k 448.0 18.3 23.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dq9o3gpx",
   "source_url": "https://omlx.ai/benchmarks/performance/dq9o3gpx",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6m3px64c-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: tyfrh0na; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 738.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 738.9 17.5 28.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6m3px64c",
   "source_url": "https://omlx.ai/benchmarks/performance/6m3px64c",
   "tg_tokens": null,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6m3px64c-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: tyfrh0na; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 851.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 851.1 17.3 29.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6m3px64c",
   "source_url": "https://omlx.ai/benchmarks/performance/6m3px64c",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dgaoslt1-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 30vegd14; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 798.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 798.7 17.4 28.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dgaoslt1",
   "source_url": "https://omlx.ai/benchmarks/performance/dgaoslt1",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dgaoslt1-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: 30vegd14; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 844.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 844.2 17.9 29.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dgaoslt1",
   "source_url": "https://omlx.ai/benchmarks/performance/dgaoslt1",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fu3381rs-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 77i11bgd, cvmk6872; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 862.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 862.8 47.5 19.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fu3381rs",
   "source_url": "https://omlx.ai/benchmarks/performance/fu3381rs",
   "tg_tokens": null,
   "tps": 47.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fu3381rs-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 77i11bgd, cvmk6872; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 770.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 770.6 73.1 20.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fu3381rs",
   "source_url": "https://omlx.ai/benchmarks/performance/fu3381rs",
   "tg_tokens": null,
   "tps": 73.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fu3381rs-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.3 GB at this context; duplicate traces of the same session: 77i11bgd, cvmk6872; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 571.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 571.1 44.3 27.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fu3381rs",
   "source_url": "https://omlx.ai/benchmarks/performance/fu3381rs",
   "tg_tokens": null,
   "tps": 44.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-l9v31akg-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 195k context; peak memory 35.6 GB; TTFT 684648 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 292.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 195k tokens 292.1 PP tok/s 20.2 TG tok/s 684648 TTFT (ms) 35.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l9v31akg",
   "source_url": "https://omlx.ai/benchmarks/performance/l9v31akg",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-vt0ioyhc-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: 3vq1q0gy, rmm80x3i, s9p8o4ub; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1146.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 1k 1,146 80.1 34.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vt0ioyhc",
   "source_url": "https://omlx.ai/benchmarks/performance/vt0ioyhc",
   "tg_tokens": null,
   "tps": 80.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-vt0ioyhc-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: 3vq1q0gy, rmm80x3i, s9p8o4ub; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1401.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 4k 1,401 69.7 35.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vt0ioyhc",
   "source_url": "https://omlx.ai/benchmarks/performance/vt0ioyhc",
   "tg_tokens": null,
   "tps": 69.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-vt0ioyhc-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: 3vq1q0gy, rmm80x3i, s9p8o4ub; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1404.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 8k 1,404 80.4 36.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vt0ioyhc",
   "source_url": "https://omlx.ai/benchmarks/performance/vt0ioyhc",
   "tg_tokens": null,
   "tps": 80.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-vt0ioyhc-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 36.6 GB at this context; duplicate traces of the same session: 3vq1q0gy, rmm80x3i, s9p8o4ub; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1365.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 16k 1,365 91.7 36.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vt0ioyhc",
   "source_url": "https://omlx.ai/benchmarks/performance/vt0ioyhc",
   "tg_tokens": null,
   "tps": 91.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-vt0ioyhc-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 37.8 GB at this context; duplicate traces of the same session: 3vq1q0gy, rmm80x3i, s9p8o4ub; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1275.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 32k 1,275 93.1 37.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vt0ioyhc",
   "source_url": "https://omlx.ai/benchmarks/performance/vt0ioyhc",
   "tg_tokens": null,
   "tps": 93.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-vt0ioyhc-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 40.1 GB at this context; duplicate traces of the same session: 3vq1q0gy, rmm80x3i, s9p8o4ub; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1119.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 64k 1,119 72.5 40.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vt0ioyhc",
   "source_url": "https://omlx.ai/benchmarks/performance/vt0ioyhc",
   "tg_tokens": null,
   "tps": 72.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-j3u9q2t7-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.0 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 432.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 432.8 33.4 43.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j3u9q2t7",
   "source_url": "https://omlx.ai/benchmarks/performance/j3u9q2t7",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-j3u9q2t7-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 45.4 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 378.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 32k 378.6 34.1 45.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j3u9q2t7",
   "source_url": "https://omlx.ai/benchmarks/performance/j3u9q2t7",
   "tg_tokens": null,
   "tps": 34.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1wdzlgec-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 749.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 749.1 42.3 39.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1wdzlgec",
   "source_url": "https://omlx.ai/benchmarks/performance/1wdzlgec",
   "tg_tokens": null,
   "tps": 42.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1wdzlgec-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 597.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 597.5 39.1 43.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1wdzlgec",
   "source_url": "https://omlx.ai/benchmarks/performance/1wdzlgec",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1wdzlgec-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 49.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 491.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 195k 491.9 26.9 49.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1wdzlgec",
   "source_url": "https://omlx.ai/benchmarks/performance/1wdzlgec",
   "tg_tokens": null,
   "tps": 26.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-h4tt3b62-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.2 GB at this context; duplicate traces of the same session: 5rtd0h9q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 738.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 738.6 42.3 39.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h4tt3b62",
   "source_url": "https://omlx.ai/benchmarks/performance/h4tt3b62",
   "tg_tokens": null,
   "tps": 42.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-h4tt3b62-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.8 GB at this context; duplicate traces of the same session: 5rtd0h9q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 599.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 599.3 41.6 43.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h4tt3b62",
   "source_url": "https://omlx.ai/benchmarks/performance/h4tt3b62",
   "tg_tokens": null,
   "tps": 41.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-h4tt3b62-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 49.0 GB at this context; duplicate traces of the same session: 5rtd0h9q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 493.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 195k 493.0 27.0 49.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h4tt3b62",
   "source_url": "https://omlx.ai/benchmarks/performance/h4tt3b62",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-31uuen02-ctx1024",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 29.9 GB; TTFT 1495 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 684.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ6e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k tokens 684.9 PP tok/s 65.1 TG tok/s 1495 TTFT (ms) 29.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/31uuen02",
   "source_url": "https://omlx.ai/benchmarks/performance/31uuen02",
   "tg_tokens": null,
   "tps": 65.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-vsm0qlv0-ctx1024",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 29.9 GB; TTFT 1444 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 709.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ6e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k tokens 709.3 PP tok/s 69.8 TG tok/s 1444 TTFT (ms) 29.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vsm0qlv0",
   "source_url": "https://omlx.ai/benchmarks/performance/vsm0qlv0",
   "tg_tokens": null,
   "tps": 69.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-aw9xon68-ctx1024",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 29.9 GB; TTFT 1461 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 701.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ6e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k tokens 701.1 PP tok/s 74.8 TG tok/s 1461 TTFT (ms) 29.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aw9xon68",
   "source_url": "https://omlx.ai/benchmarks/performance/aw9xon68",
   "tg_tokens": null,
   "tps": 74.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-qr20h47d-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.3 GB at this context; duplicate traces of the same session: d0j5uvk2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 431.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 431.8 52.9 69.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qr20h47d",
   "source_url": "https://omlx.ai/benchmarks/performance/qr20h47d",
   "tg_tokens": null,
   "tps": 52.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-qr20h47d-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.9 GB at this context; duplicate traces of the same session: d0j5uvk2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 484.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 484.6 48.7 69.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qr20h47d",
   "source_url": "https://omlx.ai/benchmarks/performance/qr20h47d",
   "tg_tokens": null,
   "tps": 48.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-qr20h47d-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.9 GB at this context; duplicate traces of the same session: d0j5uvk2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 633.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 633.1 53.5 69.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qr20h47d",
   "source_url": "https://omlx.ai/benchmarks/performance/qr20h47d",
   "tg_tokens": null,
   "tps": 53.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-qr20h47d-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.1 GB at this context; duplicate traces of the same session: d0j5uvk2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 580.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 580.7 54.5 70.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qr20h47d",
   "source_url": "https://omlx.ai/benchmarks/performance/qr20h47d",
   "tg_tokens": null,
   "tps": 54.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-opu7aya2-ctx131072",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 128k context; peak memory 36.7 GB; TTFT 937829 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 139.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ6e-MTP on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 6bit; 128k tokens 139.8 PP tok/s 20.7 TG tok/s 937829 TTFT (ms) 36.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/opu7aya2",
   "source_url": "https://omlx.ai/benchmarks/performance/opu7aya2",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-392fmqgn-ctx1024",
   "model": "clef",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 248.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "clef-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 248.6 16.1 28.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/392fmqgn",
   "source_url": "https://omlx.ai/benchmarks/performance/392fmqgn",
   "tg_tokens": null,
   "tps": 16.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-392fmqgn-ctx4096",
   "model": "clef",
   "notes": "oMLX community benchmark trace; peak memory 31.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 199.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "clef-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 199.9 14.6 31.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/392fmqgn",
   "source_url": "https://omlx.ai/benchmarks/performance/392fmqgn",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-u5db2s02-ctx32768",
   "model": "dsv4-flash-vision-exp",
   "notes": "oMLX community benchmark trace; peak memory 150.1 GB at this context; duplicate traces of the same session: 2ab6iwfw; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1142.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "dsv4-flash-vision-exp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 1,142 34.3 150.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u5db2s02",
   "source_url": "https://omlx.ai/benchmarks/performance/u5db2s02",
   "tg_tokens": null,
   "tps": 34.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-u5db2s02-ctx65536",
   "model": "dsv4-flash-vision-exp",
   "notes": "oMLX community benchmark trace; peak memory 150.3 GB at this context; duplicate traces of the same session: 2ab6iwfw; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1126.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "dsv4-flash-vision-exp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 1,126 33.9 150.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u5db2s02",
   "source_url": "https://omlx.ai/benchmarks/performance/u5db2s02",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-u5db2s02-ctx131072",
   "model": "dsv4-flash-vision-exp",
   "notes": "oMLX community benchmark trace; peak memory 150.7 GB at this context; duplicate traces of the same session: 2ab6iwfw; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1096.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "dsv4-flash-vision-exp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 1,096 33.3 150.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u5db2s02",
   "source_url": "https://omlx.ai/benchmarks/performance/u5db2s02",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-0c4ytr08-ctx1024",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 7.1 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 503.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-MLX-4bit on M5 (10c); M5 (10c) \u00b7 16 GB \u00b7 4bit; 1k 503.5 16.5 7.1 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0c4ytr08",
   "source_url": "https://omlx.ai/benchmarks/performance/0c4ytr08",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-0c4ytr08-ctx4096",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 7.5 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 506.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-MLX-4bit on M5 (10c); M5 (10c) \u00b7 16 GB \u00b7 4bit; 4k 506.8 16.5 7.5 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0c4ytr08",
   "source_url": "https://omlx.ai/benchmarks/performance/0c4ytr08",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-3mpi7vg1-ctx1024",
   "model": "gemma-4-26B-A4B-it-uncensored-heretic-ara",
   "notes": "oMLX community benchmark trace; peak memory 15.1 GB at this context; duplicate traces of the same session: hx2153np, x4lhh5io; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1542.0,
   "provenance": "community",
   "quant": "mixed 4+6bit",
   "quote": "gemma-4-26B-A4B-it-uncensored-heretic-ara-MLX-4bit-mixed_4_6 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,542 110.2 15.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3mpi7vg1",
   "source_url": "https://omlx.ai/benchmarks/performance/3mpi7vg1",
   "tg_tokens": null,
   "tps": 110.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-3mpi7vg1-ctx4096",
   "model": "gemma-4-26B-A4B-it-uncensored-heretic-ara",
   "notes": "oMLX community benchmark trace; peak memory 15.4 GB at this context; duplicate traces of the same session: hx2153np, x4lhh5io; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1497.0,
   "provenance": "community",
   "quant": "mixed 4+6bit",
   "quote": "gemma-4-26B-A4B-it-uncensored-heretic-ara-MLX-4bit-mixed_4_6 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,497 106.7 15.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3mpi7vg1",
   "source_url": "https://omlx.ai/benchmarks/performance/3mpi7vg1",
   "tg_tokens": null,
   "tps": 106.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-3mpi7vg1-ctx8192",
   "model": "gemma-4-26B-A4B-it-uncensored-heretic-ara",
   "notes": "oMLX community benchmark trace; peak memory 15.6 GB at this context; duplicate traces of the same session: hx2153np, x4lhh5io; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1299.0,
   "provenance": "community",
   "quant": "mixed 4+6bit",
   "quote": "gemma-4-26B-A4B-it-uncensored-heretic-ara-MLX-4bit-mixed_4_6 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,299 101.7 15.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3mpi7vg1",
   "source_url": "https://omlx.ai/benchmarks/performance/3mpi7vg1",
   "tg_tokens": null,
   "tps": 101.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-432b2jqo-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1318.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26b-a4b-it-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 1,318 78.3 26.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/432b2jqo",
   "source_url": "https://omlx.ai/benchmarks/performance/432b2jqo",
   "tg_tokens": null,
   "tps": 78.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-432b2jqo-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1320.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26b-a4b-it-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 1,320 70.8 26.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/432b2jqo",
   "source_url": "https://omlx.ai/benchmarks/performance/432b2jqo",
   "tg_tokens": null,
   "tps": 70.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-wlab5sqe-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 217.4,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "mlx-community--Qwen3.8-27B-nvfp4 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 217.4 27.2 19.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wlab5sqe",
   "source_url": "https://omlx.ai/benchmarks/performance/wlab5sqe",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-wlab5sqe-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 202.7,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "mlx-community--Qwen3.8-27B-nvfp4 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 202.7 24.9 20.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wlab5sqe",
   "source_url": "https://omlx.ai/benchmarks/performance/wlab5sqe",
   "tg_tokens": null,
   "tps": 24.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-cmykdrmz-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "scottlowry MTP build; oMLX community benchmark trace, single run at 4k context; peak memory 32.1 GB; TTFT 4974 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 823.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "models--scottlowry--Ornith-1.5-35B-A3B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k tokens 823.4 PP tok/s 67.2 TG tok/s 4974 TTFT (ms) 32.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cmykdrmz",
   "source_url": "https://omlx.ai/benchmarks/performance/cmykdrmz",
   "tg_tokens": null,
   "tps": 67.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-3391cncm-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "scottlowry MTP build; oMLX community benchmark trace, single run at 4k context; peak memory 32.1 GB; TTFT 4970 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 824.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "models--scottlowry--Ornith-1.5-35B-A3B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k tokens 824.1 PP tok/s 58.3 TG tok/s 4970 TTFT (ms) 32.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3391cncm",
   "source_url": "https://omlx.ai/benchmarks/performance/3391cncm",
   "tg_tokens": null,
   "tps": 58.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ngkbyv8g-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "scottlowry MTP build; oMLX community benchmark trace, single run at 4k context; peak memory 32.1 GB; TTFT 4966 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 824.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "models--scottlowry--Ornith-1.5-35B-A3B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k tokens 824.9 PP tok/s 72.6 TG tok/s 4966 TTFT (ms) 32.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ngkbyv8g",
   "source_url": "https://omlx.ai/benchmarks/performance/ngkbyv8g",
   "tg_tokens": null,
   "tps": 72.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-0w05qp6k-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 105.4 15.1 18.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0w05qp6k",
   "source_url": "https://omlx.ai/benchmarks/performance/0w05qp6k",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-0w05qp6k-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 134.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 134.7 15.0 19.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0w05qp6k",
   "source_url": "https://omlx.ai/benchmarks/performance/0w05qp6k",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-0w05qp6k-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 151.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 151.5 14.2 19.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0w05qp6k",
   "source_url": "https://omlx.ai/benchmarks/performance/0w05qp6k",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-0w05qp6k-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 97.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 97.6 12.8 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0w05qp6k",
   "source_url": "https://omlx.ai/benchmarks/performance/0w05qp6k",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-zszu42ci-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: p9txj8t4; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 105.2 14.6 18.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zszu42ci",
   "source_url": "https://omlx.ai/benchmarks/performance/zszu42ci",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-zszu42ci-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: p9txj8t4; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 113.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 113.1 14.4 19.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zszu42ci",
   "source_url": "https://omlx.ai/benchmarks/performance/zszu42ci",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-zszu42ci-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: p9txj8t4; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 106.5 14.3 17.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zszu42ci",
   "source_url": "https://omlx.ai/benchmarks/performance/zszu42ci",
   "tg_tokens": null,
   "tps": 14.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-zszu42ci-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: p9txj8t4; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 96.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 96.0 12.7 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zszu42ci",
   "source_url": "https://omlx.ai/benchmarks/performance/zszu42ci",
   "tg_tokens": null,
   "tps": 12.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-y9niffuu-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 6x0z3p0t; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 105.4 15.0 18.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y9niffuu",
   "source_url": "https://omlx.ai/benchmarks/performance/y9niffuu",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-y9niffuu-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 6x0z3p0t; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 125.0 14.5 19.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y9niffuu",
   "source_url": "https://omlx.ai/benchmarks/performance/y9niffuu",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-y9niffuu-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: 6x0z3p0t; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 106.2 13.8 17.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y9niffuu",
   "source_url": "https://omlx.ai/benchmarks/performance/y9niffuu",
   "tg_tokens": null,
   "tps": 13.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-y9niffuu-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 6x0z3p0t; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 95.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 95.9 12.3 19.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y9niffuu",
   "source_url": "https://omlx.ai/benchmarks/performance/y9niffuu",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-l6ailkmx-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: gvgg80gc, j15hd5nf; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 105.3 15.0 18.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l6ailkmx",
   "source_url": "https://omlx.ai/benchmarks/performance/l6ailkmx",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-l6ailkmx-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: gvgg80gc, j15hd5nf; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 150.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 150.8 14.5 19.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l6ailkmx",
   "source_url": "https://omlx.ai/benchmarks/performance/l6ailkmx",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-l6ailkmx-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: gvgg80gc, j15hd5nf; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 114.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 114.4 13.7 19.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l6ailkmx",
   "source_url": "https://omlx.ai/benchmarks/performance/l6ailkmx",
   "tg_tokens": null,
   "tps": 13.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-l6ailkmx-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: gvgg80gc, j15hd5nf; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 96.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 96.0 12.4 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l6ailkmx",
   "source_url": "https://omlx.ai/benchmarks/performance/l6ailkmx",
   "tg_tokens": null,
   "tps": 12.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-h7osce9e-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: orcmu142; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 105.9 15.4 18.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h7osce9e",
   "source_url": "https://omlx.ai/benchmarks/performance/h7osce9e",
   "tg_tokens": null,
   "tps": 15.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-h7osce9e-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: orcmu142; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 151.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 151.0 15.0 19.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h7osce9e",
   "source_url": "https://omlx.ai/benchmarks/performance/h7osce9e",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-h7osce9e-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: orcmu142; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 151.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 151.6 14.5 19.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h7osce9e",
   "source_url": "https://omlx.ai/benchmarks/performance/h7osce9e",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-h7osce9e-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: orcmu142; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 131.1 12.8 21.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-06",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h7osce9e",
   "source_url": "https://omlx.ai/benchmarks/performance/h7osce9e",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5yk14uew-ctx1024",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 155.8 GB at this context; duplicate traces of the same session: qi1ft6q9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 530.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 530.5 42.6 155.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5yk14uew",
   "source_url": "https://omlx.ai/benchmarks/performance/5yk14uew",
   "tg_tokens": null,
   "tps": 42.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5yk14uew-ctx4096",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.3 GB at this context; duplicate traces of the same session: qi1ft6q9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 612.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 612.2 37.6 156.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5yk14uew",
   "source_url": "https://omlx.ai/benchmarks/performance/5yk14uew",
   "tg_tokens": null,
   "tps": 37.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-2buamrga-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace, single run at 16k context; peak memory 194.2 GB; TTFT 132750 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 123.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M1 (80c); M1 (80c) \u00b7 256 GB \u00b7 unknown; 16k tokens 123.4 PP tok/s 9.9 TG tok/s 132750 TTFT (ms) 194.2 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2buamrga",
   "source_url": "https://omlx.ai/benchmarks/performance/2buamrga",
   "tg_tokens": null,
   "tps": 9.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-mr3r4332-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "Alis build; oMLX community benchmark trace; peak memory 246.8 GB at this context; duplicate traces of the same session: xmbu727z; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 478.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "GLM-5.3-Flash-Alis-MLX-6bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 1k 478.1 25.1 246.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mr3r4332",
   "source_url": "https://omlx.ai/benchmarks/performance/mr3r4332",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-mr3r4332-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "Alis build; oMLX community benchmark trace; peak memory 248.4 GB at this context; duplicate traces of the same session: xmbu727z; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 543.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "GLM-5.3-Flash-Alis-MLX-6bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 4k 543.9 27.4 248.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mr3r4332",
   "source_url": "https://omlx.ai/benchmarks/performance/mr3r4332",
   "tg_tokens": null,
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-mr3r4332-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "Alis build; oMLX community benchmark trace; peak memory 249.1 GB at this context; duplicate traces of the same session: xmbu727z; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 528.2,
   "provenance": "community",
   "quant": "6bit",
   "quote": "GLM-5.3-Flash-Alis-MLX-6bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 8k 528.2 35.1 249.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mr3r4332",
   "source_url": "https://omlx.ai/benchmarks/performance/mr3r4332",
   "tg_tokens": null,
   "tps": 35.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-mr3r4332-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "Alis build; oMLX community benchmark trace; peak memory 249.8 GB at this context; duplicate traces of the same session: xmbu727z; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 519.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "GLM-5.3-Flash-Alis-MLX-6bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 16k 519.5 40.6 249.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mr3r4332",
   "source_url": "https://omlx.ai/benchmarks/performance/mr3r4332",
   "tg_tokens": null,
   "tps": 40.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-khfwtyjz-ctx4096",
   "model": "Laguna-S-2.1",
   "notes": "fast build; oMLX community benchmark trace; peak memory 40.0 GB at this context; duplicate traces of the same session: r9ckmjqw; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 201.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Laguna-S-2.1-oQ4e-fast on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 201.8 14.8 40.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/khfwtyjz",
   "source_url": "https://omlx.ai/benchmarks/performance/khfwtyjz",
   "tg_tokens": null,
   "tps": 14.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-khfwtyjz-ctx16384",
   "model": "Laguna-S-2.1",
   "notes": "fast build; oMLX community benchmark trace; peak memory 40.6 GB at this context; duplicate traces of the same session: r9ckmjqw; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 216.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Laguna-S-2.1-oQ4e-fast on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 216.7 15.5 40.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/khfwtyjz",
   "source_url": "https://omlx.ai/benchmarks/performance/khfwtyjz",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-su7rwtmi-ctx1024",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 2.2 GB at this context; duplicate traces of the same session: hn1mo7qb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1565.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiniCPM5-2B-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 1,565 132.4 2.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/su7rwtmi",
   "source_url": "https://omlx.ai/benchmarks/performance/su7rwtmi",
   "tg_tokens": null,
   "tps": 132.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-su7rwtmi-ctx4096",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 2.3 GB at this context; duplicate traces of the same session: hn1mo7qb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1423.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiniCPM5-2B-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,423 116.2 2.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/su7rwtmi",
   "source_url": "https://omlx.ai/benchmarks/performance/su7rwtmi",
   "tg_tokens": null,
   "tps": 116.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-su7rwtmi-ctx8192",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 2.4 GB at this context; duplicate traces of the same session: hn1mo7qb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1181.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiniCPM5-2B-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 1,181 105.2 2.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/su7rwtmi",
   "source_url": "https://omlx.ai/benchmarks/performance/su7rwtmi",
   "tg_tokens": null,
   "tps": 105.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-su7rwtmi-ctx16384",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 2.7 GB at this context; duplicate traces of the same session: hn1mo7qb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 886.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiniCPM5-2B-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 886.3 89.3 2.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/su7rwtmi",
   "source_url": "https://omlx.ai/benchmarks/performance/su7rwtmi",
   "tg_tokens": null,
   "tps": 89.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-su7rwtmi-ctx32768",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 3.5 GB at this context; duplicate traces of the same session: hn1mo7qb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 621.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiniCPM5-2B-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 621.7 65.8 3.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/su7rwtmi",
   "source_url": "https://omlx.ai/benchmarks/performance/su7rwtmi",
   "tg_tokens": null,
   "tps": 65.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-2r4vj7rw-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1095.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-4bit-MLX on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k 1,095 54.1 19.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2r4vj7rw",
   "source_url": "https://omlx.ai/benchmarks/performance/2r4vj7rw",
   "tg_tokens": null,
   "tps": 54.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-2r4vj7rw-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1473.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-4bit-MLX on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 1,473 51.9 19.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2r4vj7rw",
   "source_url": "https://omlx.ai/benchmarks/performance/2r4vj7rw",
   "tg_tokens": null,
   "tps": 51.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-rtkk9h7t-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: og74z87l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2044.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 2,044 113.0 22.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rtkk9h7t",
   "source_url": "https://omlx.ai/benchmarks/performance/rtkk9h7t",
   "tg_tokens": null,
   "tps": 113.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-rtkk9h7t-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: og74z87l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1703.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,703 106.4 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rtkk9h7t",
   "source_url": "https://omlx.ai/benchmarks/performance/rtkk9h7t",
   "tg_tokens": null,
   "tps": 106.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-rtkk9h7t-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: og74z87l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 940.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 940.7 75.5 24.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rtkk9h7t",
   "source_url": "https://omlx.ai/benchmarks/performance/rtkk9h7t",
   "tg_tokens": null,
   "tps": 75.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-rtkk9h7t-ctx199680",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: og74z87l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 365.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 195k 365.4 26.6 27.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rtkk9h7t",
   "source_url": "https://omlx.ai/benchmarks/performance/rtkk9h7t",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-0q8i5unf-ctx1024",
   "model": "Ornith-1.5-35B-A3B-uncensored",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 35.5 GB; TTFT 1764 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 580.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-uncensored-MLX-8bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 1k tokens 580.6 PP tok/s 58.6 TG tok/s 1764 TTFT (ms) 35.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0q8i5unf",
   "source_url": "https://omlx.ai/benchmarks/performance/0q8i5unf",
   "tg_tokens": null,
   "tps": 58.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-qeec2a8o-ctx1024",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 15.5 GB; TTFT 4477 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 228.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled-MLX-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 228.7 PP tok/s 8.4 TG tok/s 4477 TTFT (ms) 15.5 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qeec2a8o",
   "source_url": "https://omlx.ai/benchmarks/performance/qeec2a8o",
   "tg_tokens": null,
   "tps": 8.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-d50rwh2e-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 208.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 208.3 16.2 6.2 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d50rwh2e",
   "source_url": "https://omlx.ai/benchmarks/performance/d50rwh2e",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-d50rwh2e-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 201.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 16k 201.7 14.3 6.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d50rwh2e",
   "source_url": "https://omlx.ai/benchmarks/performance/d50rwh2e",
   "tg_tokens": null,
   "tps": 14.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-i91lha7a-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 877.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 877.4 60.5 35.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i91lha7a",
   "source_url": "https://omlx.ai/benchmarks/performance/i91lha7a",
   "tg_tokens": null,
   "tps": 60.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-i91lha7a-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1030.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 1,030 59.0 37.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i91lha7a",
   "source_url": "https://omlx.ai/benchmarks/performance/i91lha7a",
   "tg_tokens": null,
   "tps": 59.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ou6p4rn7-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: hhri84yk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 963.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 963.2 66.3 21.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ou6p4rn7",
   "source_url": "https://omlx.ai/benchmarks/performance/ou6p4rn7",
   "tg_tokens": null,
   "tps": 66.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ou6p4rn7-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: hhri84yk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 868.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 868.9 59.8 21.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ou6p4rn7",
   "source_url": "https://omlx.ai/benchmarks/performance/ou6p4rn7",
   "tg_tokens": null,
   "tps": 59.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-k15rmmf8-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.1 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 660.6,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 4k 660.6 36.6 29.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k15rmmf8",
   "source_url": "https://omlx.ai/benchmarks/performance/k15rmmf8",
   "tg_tokens": null,
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M1 Max",
   "id": "omlx-k15rmmf8-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.5 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 404.9,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 32k 404.9 8.4 30.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k15rmmf8",
   "source_url": "https://omlx.ai/benchmarks/performance/k15rmmf8",
   "tg_tokens": null,
   "tps": 8.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-4i7e2znm-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 392.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 392.3 18.2 16.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4i7e2znm",
   "source_url": "https://omlx.ai/benchmarks/performance/4i7e2znm",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-4i7e2znm-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 417.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 417.6 17.8 17.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4i7e2znm",
   "source_url": "https://omlx.ai/benchmarks/performance/4i7e2znm",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-wv6dzvbs-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.6 GB at this context; duplicate traces of the same session: u2g0ptq5; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 44.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 1k 44.3 7.0 33.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wv6dzvbs",
   "source_url": "https://omlx.ai/benchmarks/performance/wv6dzvbs",
   "tg_tokens": null,
   "tps": 7.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-wv6dzvbs-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 37.6 GB at this context; duplicate traces of the same session: u2g0ptq5; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 42.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 4k 42.7 6.1 37.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wv6dzvbs",
   "source_url": "https://omlx.ai/benchmarks/performance/wv6dzvbs",
   "tg_tokens": null,
   "tps": 6.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-emh3sv5c-ctx1024",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 320.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 320.6 71.8 18.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/emh3sv5c",
   "source_url": "https://omlx.ai/benchmarks/performance/emh3sv5c",
   "tg_tokens": null,
   "tps": 71.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-emh3sv5c-ctx4096",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 345.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 345.2 65.3 22.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/emh3sv5c",
   "source_url": "https://omlx.ai/benchmarks/performance/emh3sv5c",
   "tg_tokens": null,
   "tps": 65.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-emh3sv5c-ctx8192",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 339.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 339.9 71.0 22.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/emh3sv5c",
   "source_url": "https://omlx.ai/benchmarks/performance/emh3sv5c",
   "tg_tokens": null,
   "tps": 71.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-asmcv08y-ctx1024",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: 9p4bjotg, ytz2rl7v; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 313.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ6e-MTP-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 1k 313.4 55.7 26.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/asmcv08y",
   "source_url": "https://omlx.ai/benchmarks/performance/asmcv08y",
   "tg_tokens": null,
   "tps": 55.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-asmcv08y-ctx4096",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 30.0 GB at this context; duplicate traces of the same session: 9p4bjotg, ytz2rl7v; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 340.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ6e-MTP-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 4k 340.3 48.4 30.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/asmcv08y",
   "source_url": "https://omlx.ai/benchmarks/performance/asmcv08y",
   "tg_tokens": null,
   "tps": 48.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-asmcv08y-ctx8192",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 30.5 GB at this context; duplicate traces of the same session: 9p4bjotg, ytz2rl7v; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 330.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ6e-MTP-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 8k 330.6 60.1 30.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/asmcv08y",
   "source_url": "https://omlx.ai/benchmarks/performance/asmcv08y",
   "tg_tokens": null,
   "tps": 60.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-asmcv08y-ctx16384",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 31.4 GB at this context; duplicate traces of the same session: 9p4bjotg, ytz2rl7v; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 319.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ6e-MTP-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 16k 319.8 63.5 31.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/asmcv08y",
   "source_url": "https://omlx.ai/benchmarks/performance/asmcv08y",
   "tg_tokens": null,
   "tps": 63.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-roc0ytfs-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: vfam05gd; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 228.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 228.4 7.8 17.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/roc0ytfs",
   "source_url": "https://omlx.ai/benchmarks/performance/roc0ytfs",
   "tg_tokens": null,
   "tps": 7.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-roc0ytfs-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: vfam05gd; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 199.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 16k 199.5 7.7 19.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/roc0ytfs",
   "source_url": "https://omlx.ai/benchmarks/performance/roc0ytfs",
   "tg_tokens": null,
   "tps": 7.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9404jf5e-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 4eipt3s8, clm20hpc, e3nr3bo5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 925.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 925.7 33.4 16.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9404jf5e",
   "source_url": "https://omlx.ai/benchmarks/performance/9404jf5e",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9404jf5e-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 4eipt3s8, clm20hpc, e3nr3bo5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1011.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,011 31.1 17.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9404jf5e",
   "source_url": "https://omlx.ai/benchmarks/performance/9404jf5e",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9404jf5e-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 4eipt3s8, clm20hpc, e3nr3bo5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 929.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 929.9 31.6 17.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9404jf5e",
   "source_url": "https://omlx.ai/benchmarks/performance/9404jf5e",
   "tg_tokens": null,
   "tps": 31.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9404jf5e-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 4eipt3s8, clm20hpc, e3nr3bo5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 814.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 814.7 30.4 17.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9404jf5e",
   "source_url": "https://omlx.ai/benchmarks/performance/9404jf5e",
   "tg_tokens": null,
   "tps": 30.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-dfjmy5l6-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS bare speed FP16 build; oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: lsykoi7r, n94hmver, qu3u4w3m; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 231.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Bare-Speed-FP16 on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 1k 231.3 30.7 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dfjmy5l6",
   "source_url": "https://omlx.ai/benchmarks/performance/dfjmy5l6",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-dfjmy5l6-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS bare speed FP16 build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: lsykoi7r, n94hmver, qu3u4w3m; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 246.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Bare-Speed-FP16 on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 4k 246.4 29.9 19.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dfjmy5l6",
   "source_url": "https://omlx.ai/benchmarks/performance/dfjmy5l6",
   "tg_tokens": null,
   "tps": 29.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-dfjmy5l6-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS bare speed FP16 build; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: lsykoi7r, n94hmver, qu3u4w3m; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 240.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Bare-Speed-FP16 on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 8k 240.9 29.1 20.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dfjmy5l6",
   "source_url": "https://omlx.ai/benchmarks/performance/dfjmy5l6",
   "tg_tokens": null,
   "tps": 29.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-dfjmy5l6-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS bare speed FP16 build; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: lsykoi7r, n94hmver, qu3u4w3m; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 229.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Bare-Speed-FP16 on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 16k 229.8 27.3 21.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dfjmy5l6",
   "source_url": "https://omlx.ai/benchmarks/performance/dfjmy5l6",
   "tg_tokens": null,
   "tps": 27.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-dfjmy5l6-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS bare speed FP16 build; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: lsykoi7r, n94hmver, qu3u4w3m; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 209.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Bare-Speed-FP16 on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 32k 209.9 24.5 22.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dfjmy5l6",
   "source_url": "https://omlx.ai/benchmarks/performance/dfjmy5l6",
   "tg_tokens": null,
   "tps": 24.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-2ydj1h0q-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized speed FP16 build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: bzphuyno, qp2f2en9, xpfma15c; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 230.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed-FP16 on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 1k 230.3 25.8 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ydj1h0q",
   "source_url": "https://omlx.ai/benchmarks/performance/2ydj1h0q",
   "tg_tokens": null,
   "tps": 25.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-2ydj1h0q-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized speed FP16 build; oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: bzphuyno, qp2f2en9, xpfma15c; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 232.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed-FP16 on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 4k 232.9 25.4 23.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ydj1h0q",
   "source_url": "https://omlx.ai/benchmarks/performance/2ydj1h0q",
   "tg_tokens": null,
   "tps": 25.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-2ydj1h0q-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized speed FP16 build; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: bzphuyno, qp2f2en9, xpfma15c; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 228.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed-FP16 on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 8k 228.0 24.9 24.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ydj1h0q",
   "source_url": "https://omlx.ai/benchmarks/performance/2ydj1h0q",
   "tg_tokens": null,
   "tps": 24.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-2ydj1h0q-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized speed FP16 build; oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: bzphuyno, qp2f2en9, xpfma15c; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 218.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed-FP16 on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 16k 218.6 23.5 25.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ydj1h0q",
   "source_url": "https://omlx.ai/benchmarks/performance/2ydj1h0q",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-2ydj1h0q-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized speed FP16 build; oMLX community benchmark trace; peak memory 26.9 GB at this context; duplicate traces of the same session: bzphuyno, qp2f2en9, xpfma15c; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 200.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed-FP16 on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 32k 200.8 21.9 26.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ydj1h0q",
   "source_url": "https://omlx.ai/benchmarks/performance/2ydj1h0q",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-i50hake3-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 12.8 GB at this context; duplicate traces of the same session: 2c6rh7mc, 8jsbe21z, ct3802rn, ed6nqr13, l7n5mzmy, xu24dxik; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 896.5,
   "provenance": "community",
   "quant": "oQ2.7",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ2.7 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 1k 896.5 44.3 12.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i50hake3",
   "source_url": "https://omlx.ai/benchmarks/performance/i50hake3",
   "tg_tokens": null,
   "tps": 44.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-i50hake3-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: 2c6rh7mc, 8jsbe21z, ct3802rn, ed6nqr13, l7n5mzmy, xu24dxik; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 452.6,
   "provenance": "community",
   "quant": "oQ2.7",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ2.7 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 4k 452.6 43.3 13.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i50hake3",
   "source_url": "https://omlx.ai/benchmarks/performance/i50hake3",
   "tg_tokens": null,
   "tps": 43.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-i50hake3-ctx8192",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 13.9 GB at this context; duplicate traces of the same session: 2c6rh7mc, 8jsbe21z, ct3802rn, ed6nqr13, l7n5mzmy, xu24dxik; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 350.0,
   "provenance": "community",
   "quant": "oQ2.7",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ2.7 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 8k 350.0 42.0 13.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i50hake3",
   "source_url": "https://omlx.ai/benchmarks/performance/i50hake3",
   "tg_tokens": null,
   "tps": 42.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-i50hake3-ctx16384",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: 2c6rh7mc, 8jsbe21z, ct3802rn, ed6nqr13, l7n5mzmy, xu24dxik; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 314.5,
   "provenance": "community",
   "quant": "oQ2.7",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ2.7 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 16k 314.5 40.1 14.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i50hake3",
   "source_url": "https://omlx.ai/benchmarks/performance/i50hake3",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-i50hake3-ctx32768",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 15.4 GB at this context; duplicate traces of the same session: 2c6rh7mc, 8jsbe21z, ct3802rn, ed6nqr13, l7n5mzmy, xu24dxik; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 296.8,
   "provenance": "community",
   "quant": "oQ2.7",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ2.7 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 32k 296.8 37.0 15.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i50hake3",
   "source_url": "https://omlx.ai/benchmarks/performance/i50hake3",
   "tg_tokens": null,
   "tps": 37.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-i50hake3-ctx65536",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: 2c6rh7mc, 8jsbe21z, ct3802rn, ed6nqr13, l7n5mzmy, xu24dxik; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 269.8,
   "provenance": "community",
   "quant": "oQ2.7",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ2.7 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 64k 269.8 28.2 17.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i50hake3",
   "source_url": "https://omlx.ai/benchmarks/performance/i50hake3",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-i50hake3-ctx131072",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 2c6rh7mc, 8jsbe21z, ct3802rn, ed6nqr13, l7n5mzmy, xu24dxik; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 244.1,
   "provenance": "community",
   "quant": "oQ2.7",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ2.7 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 128k 244.1 24.7 21.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i50hake3",
   "source_url": "https://omlx.ai/benchmarks/performance/i50hake3",
   "tg_tokens": null,
   "tps": 24.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ajcsh2vn-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 14.7 GB at this context; duplicate traces of the same session: 3x7l0mgl, f2itb8w9, f2lblzjy, ybylsd7j, ysedkpbk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 799.5,
   "provenance": "community",
   "quant": "oQ3.5",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3.5 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 1k 799.5 32.5 14.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ajcsh2vn",
   "source_url": "https://omlx.ai/benchmarks/performance/ajcsh2vn",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ajcsh2vn-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 15.5 GB at this context; duplicate traces of the same session: 3x7l0mgl, f2itb8w9, f2lblzjy, ybylsd7j, ysedkpbk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 770.2,
   "provenance": "community",
   "quant": "oQ3.5",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3.5 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 4k 770.2 32.5 15.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ajcsh2vn",
   "source_url": "https://omlx.ai/benchmarks/performance/ajcsh2vn",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ajcsh2vn-ctx8192",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: 3x7l0mgl, f2itb8w9, f2lblzjy, ybylsd7j, ysedkpbk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 691.7,
   "provenance": "community",
   "quant": "oQ3.5",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3.5 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 8k 691.7 30.6 15.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ajcsh2vn",
   "source_url": "https://omlx.ai/benchmarks/performance/ajcsh2vn",
   "tg_tokens": null,
   "tps": 30.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ajcsh2vn-ctx16384",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: 3x7l0mgl, f2itb8w9, f2lblzjy, ybylsd7j, ysedkpbk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 650.9,
   "provenance": "community",
   "quant": "oQ3.5",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3.5 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 16k 650.9 28.3 16.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ajcsh2vn",
   "source_url": "https://omlx.ai/benchmarks/performance/ajcsh2vn",
   "tg_tokens": null,
   "tps": 28.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ajcsh2vn-ctx32768",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 3x7l0mgl, f2itb8w9, f2lblzjy, ybylsd7j, ysedkpbk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 604.0,
   "provenance": "community",
   "quant": "oQ3.5",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3.5 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 32k 604.0 25.7 17.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ajcsh2vn",
   "source_url": "https://omlx.ai/benchmarks/performance/ajcsh2vn",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ajcsh2vn-ctx65536",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 3x7l0mgl, f2itb8w9, f2lblzjy, ybylsd7j, ysedkpbk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 540.1,
   "provenance": "community",
   "quant": "oQ3.5",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3.5 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 64k 540.1 22.8 19.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ajcsh2vn",
   "source_url": "https://omlx.ai/benchmarks/performance/ajcsh2vn",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ajcsh2vn-ctx131072",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 3x7l0mgl, f2itb8w9, f2lblzjy, ybylsd7j, ysedkpbk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 448.0,
   "provenance": "community",
   "quant": "oQ3.5",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ3.5 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 128k 448.0 18.3 23.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ajcsh2vn",
   "source_url": "https://omlx.ai/benchmarks/performance/ajcsh2vn",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-u2xuta6t-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: 15rwzkhd; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 133.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 133.5 14.8 17.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u2xuta6t",
   "source_url": "https://omlx.ai/benchmarks/performance/u2xuta6t",
   "tg_tokens": null,
   "tps": 14.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-u2xuta6t-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: 15rwzkhd; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 128.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 128.1 13.5 18.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u2xuta6t",
   "source_url": "https://omlx.ai/benchmarks/performance/u2xuta6t",
   "tg_tokens": null,
   "tps": 13.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tyfrh0na-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 738.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 738.9 17.5 28.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tyfrh0na",
   "source_url": "https://omlx.ai/benchmarks/performance/tyfrh0na",
   "tg_tokens": null,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tyfrh0na-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 851.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 851.1 17.3 29.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tyfrh0na",
   "source_url": "https://omlx.ai/benchmarks/performance/tyfrh0na",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-30vegd14-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 798.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 798.7 17.4 28.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/30vegd14",
   "source_url": "https://omlx.ai/benchmarks/performance/30vegd14",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-30vegd14-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 844.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 844.2 17.9 29.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/30vegd14",
   "source_url": "https://omlx.ai/benchmarks/performance/30vegd14",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-xpohsw6v-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: qey1mvne; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 247.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 247.4 51.2 21.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xpohsw6v",
   "source_url": "https://omlx.ai/benchmarks/performance/xpohsw6v",
   "tg_tokens": null,
   "tps": 51.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-xpohsw6v-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.0 GB at this context; duplicate traces of the same session: qey1mvne; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 184.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 184.9 11.5 39.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xpohsw6v",
   "source_url": "https://omlx.ai/benchmarks/performance/xpohsw6v",
   "tg_tokens": null,
   "tps": 11.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cvmk6872-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 77i11bgd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 862.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 862.8 47.5 19.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cvmk6872",
   "source_url": "https://omlx.ai/benchmarks/performance/cvmk6872",
   "tg_tokens": null,
   "tps": 47.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cvmk6872-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 77i11bgd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 770.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 770.6 73.1 20.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cvmk6872",
   "source_url": "https://omlx.ai/benchmarks/performance/cvmk6872",
   "tg_tokens": null,
   "tps": 73.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cvmk6872-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.3 GB at this context; duplicate traces of the same session: 77i11bgd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 571.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 571.1 44.3 27.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cvmk6872",
   "source_url": "https://omlx.ai/benchmarks/performance/cvmk6872",
   "tg_tokens": null,
   "tps": 44.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-3vq1q0gy-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: rmm80x3i, s9p8o4ub; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1146.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 1k 1,146 80.1 34.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3vq1q0gy",
   "source_url": "https://omlx.ai/benchmarks/performance/3vq1q0gy",
   "tg_tokens": null,
   "tps": 80.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-3vq1q0gy-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: rmm80x3i, s9p8o4ub; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1401.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 4k 1,401 69.7 35.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3vq1q0gy",
   "source_url": "https://omlx.ai/benchmarks/performance/3vq1q0gy",
   "tg_tokens": null,
   "tps": 69.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-3vq1q0gy-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: rmm80x3i, s9p8o4ub; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1404.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 8k 1,404 80.4 36.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3vq1q0gy",
   "source_url": "https://omlx.ai/benchmarks/performance/3vq1q0gy",
   "tg_tokens": null,
   "tps": 80.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-3vq1q0gy-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 36.6 GB at this context; duplicate traces of the same session: rmm80x3i, s9p8o4ub; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1365.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 16k 1,365 91.7 36.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3vq1q0gy",
   "source_url": "https://omlx.ai/benchmarks/performance/3vq1q0gy",
   "tg_tokens": null,
   "tps": 91.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-3vq1q0gy-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 37.8 GB at this context; duplicate traces of the same session: rmm80x3i, s9p8o4ub; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1275.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 32k 1,275 93.1 37.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3vq1q0gy",
   "source_url": "https://omlx.ai/benchmarks/performance/3vq1q0gy",
   "tg_tokens": null,
   "tps": 93.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-3vq1q0gy-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 40.1 GB at this context; duplicate traces of the same session: rmm80x3i, s9p8o4ub; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1119.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 64k 1,119 72.5 40.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3vq1q0gy",
   "source_url": "https://omlx.ai/benchmarks/performance/3vq1q0gy",
   "tg_tokens": null,
   "tps": 72.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5rtd0h9q-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 738.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 738.6 42.3 39.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5rtd0h9q",
   "source_url": "https://omlx.ai/benchmarks/performance/5rtd0h9q",
   "tg_tokens": null,
   "tps": 42.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5rtd0h9q-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 599.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 599.3 41.6 43.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5rtd0h9q",
   "source_url": "https://omlx.ai/benchmarks/performance/5rtd0h9q",
   "tg_tokens": null,
   "tps": 41.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5rtd0h9q-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 49.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 493.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 195k 493.0 27.0 49.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5rtd0h9q",
   "source_url": "https://omlx.ai/benchmarks/performance/5rtd0h9q",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-riongc0b-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 73.5 GB at this context; duplicate traces of the same session: 38c5pwrc, dzqjaajh, jldn9ml2, rnxro9f5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 870.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 870.8 69.5 73.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/riongc0b",
   "source_url": "https://omlx.ai/benchmarks/performance/riongc0b",
   "tg_tokens": null,
   "tps": 69.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-riongc0b-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 74.0 GB at this context; duplicate traces of the same session: 38c5pwrc, dzqjaajh, jldn9ml2, rnxro9f5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2203.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,203 70.8 74.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/riongc0b",
   "source_url": "https://omlx.ai/benchmarks/performance/riongc0b",
   "tg_tokens": null,
   "tps": 70.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-riongc0b-ctx65536",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 78.0 GB at this context; duplicate traces of the same session: 38c5pwrc, dzqjaajh, jldn9ml2, rnxro9f5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2959.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,959 62.7 78.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/riongc0b",
   "source_url": "https://omlx.ai/benchmarks/performance/riongc0b",
   "tg_tokens": null,
   "tps": 62.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-riongc0b-ctx131072",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 81.2 GB at this context; duplicate traces of the same session: 38c5pwrc, dzqjaajh, jldn9ml2, rnxro9f5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2847.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,847 69.5 81.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/riongc0b",
   "source_url": "https://omlx.ai/benchmarks/performance/riongc0b",
   "tg_tokens": null,
   "tps": 69.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-riongc0b-ctx199680",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 86.2 GB at this context; duplicate traces of the same session: 38c5pwrc, dzqjaajh, jldn9ml2, rnxro9f5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2642.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,642 59.4 86.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/riongc0b",
   "source_url": "https://omlx.ai/benchmarks/performance/riongc0b",
   "tg_tokens": null,
   "tps": 59.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-e4a5o58q-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 73.5 GB at this context; duplicate traces of the same session: 1cxod9v4, 1x0ffrej, co6z7e0m, lbk03hb4, ungjll42, zxg1wqhv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 986.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 986.0 54.1 73.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e4a5o58q",
   "source_url": "https://omlx.ai/benchmarks/performance/e4a5o58q",
   "tg_tokens": null,
   "tps": 54.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-e4a5o58q-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 73.8 GB at this context; duplicate traces of the same session: 1cxod9v4, 1x0ffrej, co6z7e0m, lbk03hb4, ungjll42, zxg1wqhv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1940.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,940 63.0 73.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e4a5o58q",
   "source_url": "https://omlx.ai/benchmarks/performance/e4a5o58q",
   "tg_tokens": null,
   "tps": 63.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-e4a5o58q-ctx8192",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 74.6 GB at this context; duplicate traces of the same session: 1cxod9v4, 1x0ffrej, co6z7e0m, lbk03hb4, ungjll42, zxg1wqhv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2265.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,265 49.6 74.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e4a5o58q",
   "source_url": "https://omlx.ai/benchmarks/performance/e4a5o58q",
   "tg_tokens": null,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-e4a5o58q-ctx32768",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 76.4 GB at this context; duplicate traces of the same session: 1cxod9v4, 1x0ffrej, co6z7e0m, lbk03hb4, ungjll42, zxg1wqhv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2553.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,553 57.0 76.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e4a5o58q",
   "source_url": "https://omlx.ai/benchmarks/performance/e4a5o58q",
   "tg_tokens": null,
   "tps": 57.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-e4a5o58q-ctx65536",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 78.1 GB at this context; duplicate traces of the same session: 1cxod9v4, 1x0ffrej, co6z7e0m, lbk03hb4, ungjll42, zxg1wqhv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2534.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,534 28.2 78.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e4a5o58q",
   "source_url": "https://omlx.ai/benchmarks/performance/e4a5o58q",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-e4a5o58q-ctx131072",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 81.2 GB at this context; duplicate traces of the same session: 1cxod9v4, 1x0ffrej, co6z7e0m, lbk03hb4, ungjll42, zxg1wqhv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2463.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,463 58.2 81.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e4a5o58q",
   "source_url": "https://omlx.ai/benchmarks/performance/e4a5o58q",
   "tg_tokens": null,
   "tps": 58.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-e4a5o58q-ctx199680",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 86.1 GB at this context; duplicate traces of the same session: 1cxod9v4, 1x0ffrej, co6z7e0m, lbk03hb4, ungjll42, zxg1wqhv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2337.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,337 53.5 86.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e4a5o58q",
   "source_url": "https://omlx.ai/benchmarks/performance/e4a5o58q",
   "tg_tokens": null,
   "tps": 53.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2rp68cbq-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 69.1 GB at this context; duplicate traces of the same session: a5e6y7y1, f440d8h9, if33kp3n, o62q4w3p, pfufmx2f, sr7voyfu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1056.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,056 60.2 69.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2rp68cbq",
   "source_url": "https://omlx.ai/benchmarks/performance/2rp68cbq",
   "tg_tokens": null,
   "tps": 60.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2rp68cbq-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 70.1 GB at this context; duplicate traces of the same session: a5e6y7y1, f440d8h9, if33kp3n, o62q4w3p, pfufmx2f, sr7voyfu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2426.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,426 65.1 70.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2rp68cbq",
   "source_url": "https://omlx.ai/benchmarks/performance/2rp68cbq",
   "tg_tokens": null,
   "tps": 65.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2rp68cbq-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: a5e6y7y1, f440d8h9, if33kp3n, o62q4w3p, pfufmx2f, sr7voyfu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2752.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,752 65.3 70.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2rp68cbq",
   "source_url": "https://omlx.ai/benchmarks/performance/2rp68cbq",
   "tg_tokens": null,
   "tps": 65.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2rp68cbq-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: a5e6y7y1, f440d8h9, if33kp3n, o62q4w3p, pfufmx2f, sr7voyfu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2708.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,708 68.0 71.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2rp68cbq",
   "source_url": "https://omlx.ai/benchmarks/performance/2rp68cbq",
   "tg_tokens": null,
   "tps": 68.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2rp68cbq-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 73.2 GB at this context; duplicate traces of the same session: a5e6y7y1, f440d8h9, if33kp3n, o62q4w3p, pfufmx2f, sr7voyfu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2596.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,596 65.7 73.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2rp68cbq",
   "source_url": "https://omlx.ai/benchmarks/performance/2rp68cbq",
   "tg_tokens": null,
   "tps": 65.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2rp68cbq-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 76.6 GB at this context; duplicate traces of the same session: a5e6y7y1, f440d8h9, if33kp3n, o62q4w3p, pfufmx2f, sr7voyfu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2528.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,528 67.1 76.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2rp68cbq",
   "source_url": "https://omlx.ai/benchmarks/performance/2rp68cbq",
   "tg_tokens": null,
   "tps": 67.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2rp68cbq-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 80.8 GB at this context; duplicate traces of the same session: a5e6y7y1, f440d8h9, if33kp3n, o62q4w3p, pfufmx2f, sr7voyfu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2429.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,429 63.6 80.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2rp68cbq",
   "source_url": "https://omlx.ai/benchmarks/performance/2rp68cbq",
   "tg_tokens": null,
   "tps": 63.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-d0j5uvk2-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 431.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 431.8 52.9 69.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d0j5uvk2",
   "source_url": "https://omlx.ai/benchmarks/performance/d0j5uvk2",
   "tg_tokens": null,
   "tps": 52.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-d0j5uvk2-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 484.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 484.6 48.7 69.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d0j5uvk2",
   "source_url": "https://omlx.ai/benchmarks/performance/d0j5uvk2",
   "tg_tokens": null,
   "tps": 48.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-d0j5uvk2-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 633.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 633.1 53.5 69.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d0j5uvk2",
   "source_url": "https://omlx.ai/benchmarks/performance/d0j5uvk2",
   "tg_tokens": null,
   "tps": 53.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-d0j5uvk2-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 580.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 580.7 54.5 70.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d0j5uvk2",
   "source_url": "https://omlx.ai/benchmarks/performance/d0j5uvk2",
   "tg_tokens": null,
   "tps": 54.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-300v2ogr-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.0 GB at this context; duplicate traces of the same session: 5xrp0ndq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 195.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 195.1 10.4 43.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/300v2ogr",
   "source_url": "https://omlx.ai/benchmarks/performance/300v2ogr",
   "tg_tokens": null,
   "tps": 10.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-300v2ogr-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.1 GB at this context; duplicate traces of the same session: 5xrp0ndq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 153.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 153.0 10.8 43.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/300v2ogr",
   "source_url": "https://omlx.ai/benchmarks/performance/300v2ogr",
   "tg_tokens": null,
   "tps": 10.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-p47x67f5-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 42.4 GB at this context; duplicate traces of the same session: ci1nmk6s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 77.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 77.6 9.0 42.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p47x67f5",
   "source_url": "https://omlx.ai/benchmarks/performance/p47x67f5",
   "tg_tokens": null,
   "tps": 9.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-p47x67f5-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.2 GB at this context; duplicate traces of the same session: ci1nmk6s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 117.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 117.1 10.8 43.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p47x67f5",
   "source_url": "https://omlx.ai/benchmarks/performance/p47x67f5",
   "tg_tokens": null,
   "tps": 10.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-tvuj7xvf-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 40.8 GB at this context; duplicate traces of the same session: 0w6bir0y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 78.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 78.0 10.2 40.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tvuj7xvf",
   "source_url": "https://omlx.ai/benchmarks/performance/tvuj7xvf",
   "tg_tokens": null,
   "tps": 10.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-tvuj7xvf-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 41.5 GB at this context; duplicate traces of the same session: 0w6bir0y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 149.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 149.7 11.3 41.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tvuj7xvf",
   "source_url": "https://omlx.ai/benchmarks/performance/tvuj7xvf",
   "tg_tokens": null,
   "tps": 11.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-2ab6iwfw-ctx32768",
   "model": "dsv4-flash-vision-exp",
   "notes": "oMLX community benchmark trace; peak memory 150.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1142.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "dsv4-flash-vision-exp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 1,142 34.3 150.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ab6iwfw",
   "source_url": "https://omlx.ai/benchmarks/performance/2ab6iwfw",
   "tg_tokens": null,
   "tps": 34.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-2ab6iwfw-ctx65536",
   "model": "dsv4-flash-vision-exp",
   "notes": "oMLX community benchmark trace; peak memory 150.3 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1126.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "dsv4-flash-vision-exp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 1,126 33.9 150.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ab6iwfw",
   "source_url": "https://omlx.ai/benchmarks/performance/2ab6iwfw",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-2ab6iwfw-ctx131072",
   "model": "dsv4-flash-vision-exp",
   "notes": "oMLX community benchmark trace; peak memory 150.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1096.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "dsv4-flash-vision-exp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 1,096 33.3 150.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ab6iwfw",
   "source_url": "https://omlx.ai/benchmarks/performance/2ab6iwfw",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-x4lhh5io-ctx1024",
   "model": "gemma-4-26B-A4B-it-uncensored-heretic-ara",
   "notes": "oMLX community benchmark trace; peak memory 15.1 GB at this context; duplicate traces of the same session: hx2153np; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1542.0,
   "provenance": "community",
   "quant": "mixed 4+6bit",
   "quote": "gemma-4-26B-A4B-it-uncensored-heretic-ara-MLX-4bit-mixed_4_6 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,542 110.2 15.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x4lhh5io",
   "source_url": "https://omlx.ai/benchmarks/performance/x4lhh5io",
   "tg_tokens": null,
   "tps": 110.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-x4lhh5io-ctx4096",
   "model": "gemma-4-26B-A4B-it-uncensored-heretic-ara",
   "notes": "oMLX community benchmark trace; peak memory 15.4 GB at this context; duplicate traces of the same session: hx2153np; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1497.0,
   "provenance": "community",
   "quant": "mixed 4+6bit",
   "quote": "gemma-4-26B-A4B-it-uncensored-heretic-ara-MLX-4bit-mixed_4_6 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,497 106.7 15.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x4lhh5io",
   "source_url": "https://omlx.ai/benchmarks/performance/x4lhh5io",
   "tg_tokens": null,
   "tps": 106.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-x4lhh5io-ctx8192",
   "model": "gemma-4-26B-A4B-it-uncensored-heretic-ara",
   "notes": "oMLX community benchmark trace; peak memory 15.6 GB at this context; duplicate traces of the same session: hx2153np; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1299.0,
   "provenance": "community",
   "quant": "mixed 4+6bit",
   "quote": "gemma-4-26B-A4B-it-uncensored-heretic-ara-MLX-4bit-mixed_4_6 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,299 101.7 15.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x4lhh5io",
   "source_url": "https://omlx.ai/benchmarks/performance/x4lhh5io",
   "tg_tokens": null,
   "tps": 101.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-xc4atsxm-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1318.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26b-a4b-it-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 1,318 78.3 26.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xc4atsxm",
   "source_url": "https://omlx.ai/benchmarks/performance/xc4atsxm",
   "tg_tokens": null,
   "tps": 78.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-xc4atsxm-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1320.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26b-a4b-it-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 1,320 70.8 26.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xc4atsxm",
   "source_url": "https://omlx.ai/benchmarks/performance/xc4atsxm",
   "tg_tokens": null,
   "tps": 70.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-p9txj8t4-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 105.2 14.6 18.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p9txj8t4",
   "source_url": "https://omlx.ai/benchmarks/performance/p9txj8t4",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-p9txj8t4-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 113.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 113.1 14.4 19.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p9txj8t4",
   "source_url": "https://omlx.ai/benchmarks/performance/p9txj8t4",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-p9txj8t4-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 106.5 14.3 17.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p9txj8t4",
   "source_url": "https://omlx.ai/benchmarks/performance/p9txj8t4",
   "tg_tokens": null,
   "tps": 14.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-p9txj8t4-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 96.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 96.0 12.7 19.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p9txj8t4",
   "source_url": "https://omlx.ai/benchmarks/performance/p9txj8t4",
   "tg_tokens": null,
   "tps": 12.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-6x0z3p0t-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 105.4 15.0 18.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6x0z3p0t",
   "source_url": "https://omlx.ai/benchmarks/performance/6x0z3p0t",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-6x0z3p0t-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 125.0 14.5 19.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6x0z3p0t",
   "source_url": "https://omlx.ai/benchmarks/performance/6x0z3p0t",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-6x0z3p0t-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 106.2 13.8 17.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6x0z3p0t",
   "source_url": "https://omlx.ai/benchmarks/performance/6x0z3p0t",
   "tg_tokens": null,
   "tps": 13.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-6x0z3p0t-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 95.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 95.9 12.3 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6x0z3p0t",
   "source_url": "https://omlx.ai/benchmarks/performance/6x0z3p0t",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-gvgg80gc-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: j15hd5nf; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 105.3 15.0 18.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gvgg80gc",
   "source_url": "https://omlx.ai/benchmarks/performance/gvgg80gc",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-gvgg80gc-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: j15hd5nf; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 150.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 150.8 14.5 19.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gvgg80gc",
   "source_url": "https://omlx.ai/benchmarks/performance/gvgg80gc",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-gvgg80gc-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: j15hd5nf; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 114.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 114.4 13.7 19.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gvgg80gc",
   "source_url": "https://omlx.ai/benchmarks/performance/gvgg80gc",
   "tg_tokens": null,
   "tps": 13.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-gvgg80gc-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: j15hd5nf; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 96.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 96.0 12.4 19.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gvgg80gc",
   "source_url": "https://omlx.ai/benchmarks/performance/gvgg80gc",
   "tg_tokens": null,
   "tps": 12.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-orcmu142-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 2gj5qsom, azc6k454; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 105.9 15.4 18.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/orcmu142",
   "source_url": "https://omlx.ai/benchmarks/performance/orcmu142",
   "tg_tokens": null,
   "tps": 15.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-orcmu142-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 2gj5qsom, azc6k454; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 151.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 151.0 15.0 19.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/orcmu142",
   "source_url": "https://omlx.ai/benchmarks/performance/orcmu142",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-orcmu142-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: 2gj5qsom, azc6k454; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 151.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 151.6 14.5 19.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/orcmu142",
   "source_url": "https://omlx.ai/benchmarks/performance/orcmu142",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-orcmu142-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 2gj5qsom, azc6k454; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 131.1 12.8 21.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/orcmu142",
   "source_url": "https://omlx.ai/benchmarks/performance/orcmu142",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-1j0j4zw7-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: 04mlzsmw, blx0hhfb, kd1h2rex; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 105.9 15.5 18.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1j0j4zw7",
   "source_url": "https://omlx.ai/benchmarks/performance/1j0j4zw7",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-1j0j4zw7-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 04mlzsmw, blx0hhfb, kd1h2rex; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 121.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 121.2 15.0 19.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1j0j4zw7",
   "source_url": "https://omlx.ai/benchmarks/performance/1j0j4zw7",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-1j0j4zw7-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: 04mlzsmw, blx0hhfb, kd1h2rex; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 106.5 14.3 17.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1j0j4zw7",
   "source_url": "https://omlx.ai/benchmarks/performance/1j0j4zw7",
   "tg_tokens": null,
   "tps": 14.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-1j0j4zw7-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 04mlzsmw, blx0hhfb, kd1h2rex; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 96.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 96.0 12.8 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1j0j4zw7",
   "source_url": "https://omlx.ai/benchmarks/performance/1j0j4zw7",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ydea020u-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 172.8 GB at this context; duplicate traces of the same session: fi9e2yv9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 446.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 446.7 46.5 172.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ydea020u",
   "source_url": "https://omlx.ai/benchmarks/performance/ydea020u",
   "tg_tokens": null,
   "tps": 46.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ydea020u-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 173.7 GB at this context; duplicate traces of the same session: fi9e2yv9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 575.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 575.9 39.9 173.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ydea020u",
   "source_url": "https://omlx.ai/benchmarks/performance/ydea020u",
   "tg_tokens": null,
   "tps": 39.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-h4tjgr0u-ctx1024",
   "model": "Muse Glimmer",
   "notes": "oMLX community benchmark trace; peak memory 28.2 GB at this context; duplicate traces of the same session: 4g02exm6, i4qqptnv, jfwqjf5g, pcb3m6u4, td13xte0, v1b3y42g; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 93.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Muse-Glimmer-30B-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 1k 93.6 10.4 28.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h4tjgr0u",
   "source_url": "https://omlx.ai/benchmarks/performance/h4tjgr0u",
   "tg_tokens": null,
   "tps": 10.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-h4tjgr0u-ctx4096",
   "model": "Muse Glimmer",
   "notes": "oMLX community benchmark trace; peak memory 28.4 GB at this context; duplicate traces of the same session: 4g02exm6, i4qqptnv, jfwqjf5g, pcb3m6u4, td13xte0, v1b3y42g; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 80.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Muse-Glimmer-30B-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k 80.2 10.0 28.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h4tjgr0u",
   "source_url": "https://omlx.ai/benchmarks/performance/h4tjgr0u",
   "tg_tokens": null,
   "tps": 10.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-h4tjgr0u-ctx8192",
   "model": "Muse Glimmer",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 4g02exm6, i4qqptnv, jfwqjf5g, pcb3m6u4, td13xte0, v1b3y42g; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 78.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Muse-Glimmer-30B-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 8k 78.3 9.8 28.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h4tjgr0u",
   "source_url": "https://omlx.ai/benchmarks/performance/h4tjgr0u",
   "tg_tokens": null,
   "tps": 9.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-h4tjgr0u-ctx16384",
   "model": "Muse Glimmer",
   "notes": "oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: 4g02exm6, i4qqptnv, jfwqjf5g, pcb3m6u4, td13xte0, v1b3y42g; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 76.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Muse-Glimmer-30B-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 16k 76.3 9.6 28.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h4tjgr0u",
   "source_url": "https://omlx.ai/benchmarks/performance/h4tjgr0u",
   "tg_tokens": null,
   "tps": 9.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-h4tjgr0u-ctx32768",
   "model": "Muse Glimmer",
   "notes": "oMLX community benchmark trace; peak memory 28.8 GB at this context; duplicate traces of the same session: 4g02exm6, i4qqptnv, jfwqjf5g, pcb3m6u4, td13xte0, v1b3y42g; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 74.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Muse-Glimmer-30B-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 32k 74.0 9.2 28.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h4tjgr0u",
   "source_url": "https://omlx.ai/benchmarks/performance/h4tjgr0u",
   "tg_tokens": null,
   "tps": 9.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-h4tjgr0u-ctx65536",
   "model": "Muse Glimmer",
   "notes": "oMLX community benchmark trace; peak memory 29.2 GB at this context; duplicate traces of the same session: 4g02exm6, i4qqptnv, jfwqjf5g, pcb3m6u4, td13xte0, v1b3y42g; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 70.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Muse-Glimmer-30B-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 64k 70.7 8.5 29.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h4tjgr0u",
   "source_url": "https://omlx.ai/benchmarks/performance/h4tjgr0u",
   "tg_tokens": null,
   "tps": 8.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-h4tjgr0u-ctx131072",
   "model": "Muse Glimmer",
   "notes": "oMLX community benchmark trace; peak memory 30.0 GB at this context; duplicate traces of the same session: 4g02exm6, i4qqptnv, jfwqjf5g, pcb3m6u4, td13xte0, v1b3y42g; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 65.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Muse-Glimmer-30B-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 128k 65.4 7.7 30.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h4tjgr0u",
   "source_url": "https://omlx.ai/benchmarks/performance/h4tjgr0u",
   "tg_tokens": null,
   "tps": 7.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-js8ia3fv-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: bul9cayk, fbgeckjy; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1142.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-4bit-MLX on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k 1,142 56.2 19.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/js8ia3fv",
   "source_url": "https://omlx.ai/benchmarks/performance/js8ia3fv",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-js8ia3fv-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: bul9cayk, fbgeckjy; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1479.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-4bit-MLX on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 1,479 55.0 19.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/js8ia3fv",
   "source_url": "https://omlx.ai/benchmarks/performance/js8ia3fv",
   "tg_tokens": null,
   "tps": 55.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-js8ia3fv-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: bul9cayk, fbgeckjy; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1433.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-4bit-MLX on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 8k 1,433 52.6 19.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/js8ia3fv",
   "source_url": "https://omlx.ai/benchmarks/performance/js8ia3fv",
   "tg_tokens": null,
   "tps": 52.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-js8ia3fv-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: bul9cayk, fbgeckjy; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1269.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-4bit-MLX on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 16k 1,269 48.1 20.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/js8ia3fv",
   "source_url": "https://omlx.ai/benchmarks/performance/js8ia3fv",
   "tg_tokens": null,
   "tps": 48.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-qnjz8n3v-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: p1i0riae; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 766.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 1k 766.0 70.4 35.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qnjz8n3v",
   "source_url": "https://omlx.ai/benchmarks/performance/qnjz8n3v",
   "tg_tokens": null,
   "tps": 70.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-qnjz8n3v-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 38.1 GB at this context; duplicate traces of the same session: p1i0riae; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 741.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 8k 741.3 61.8 38.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qnjz8n3v",
   "source_url": "https://omlx.ai/benchmarks/performance/qnjz8n3v",
   "tg_tokens": null,
   "tps": 61.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-zs48svdn-ctx1024",
   "model": "Q_pi",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: g2gfnbud; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 255.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Q_pi on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 255.2 16.2 28.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zs48svdn",
   "source_url": "https://omlx.ai/benchmarks/performance/zs48svdn",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-zs48svdn-ctx4096",
   "model": "Q_pi",
   "notes": "oMLX community benchmark trace; peak memory 31.7 GB at this context; duplicate traces of the same session: g2gfnbud; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 218.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Q_pi on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 218.2 15.3 31.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zs48svdn",
   "source_url": "https://omlx.ai/benchmarks/performance/zs48svdn",
   "tg_tokens": null,
   "tps": 15.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-oqn7eq7v-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: ao1h1zde; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1095.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 16 GB \u00b7 4bit; 4k 1,095 29.1 6.3 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oqn7eq7v",
   "source_url": "https://omlx.ai/benchmarks/performance/oqn7eq7v",
   "tg_tokens": null,
   "tps": 29.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-oqn7eq7v-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.6 GB at this context; duplicate traces of the same session: ao1h1zde; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1087.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 16 GB \u00b7 4bit; 16k 1,087 26.6 6.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oqn7eq7v",
   "source_url": "https://omlx.ai/benchmarks/performance/oqn7eq7v",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xrw8vs7x-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3709.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 3,709 130.6 19.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xrw8vs7x",
   "source_url": "https://omlx.ai/benchmarks/performance/xrw8vs7x",
   "tg_tokens": null,
   "tps": 130.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xrw8vs7x-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4576.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 4,576 117.9 20.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xrw8vs7x",
   "source_url": "https://omlx.ai/benchmarks/performance/xrw8vs7x",
   "tg_tokens": null,
   "tps": 117.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-78qfyntr-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 258.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 258.1 29.4 16.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/78qfyntr",
   "source_url": "https://omlx.ai/benchmarks/performance/78qfyntr",
   "tg_tokens": null,
   "tps": 29.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-78qfyntr-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 80.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 189.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 189.7 10.6 80.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/78qfyntr",
   "source_url": "https://omlx.ai/benchmarks/performance/78qfyntr",
   "tg_tokens": null,
   "tps": 10.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4hhvk3hk-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 834.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 834.2 40.1 19.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4hhvk3hk",
   "source_url": "https://omlx.ai/benchmarks/performance/4hhvk3hk",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4hhvk3hk-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 874.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 874.3 45.0 21.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4hhvk3hk",
   "source_url": "https://omlx.ai/benchmarks/performance/4hhvk3hk",
   "tg_tokens": null,
   "tps": 45.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-xz49ixf1-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: r7a6cr0h; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 112.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 1k 112.1 14.7 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xz49ixf1",
   "source_url": "https://omlx.ai/benchmarks/performance/xz49ixf1",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-xz49ixf1-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: r7a6cr0h; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 105.4 13.7 19.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xz49ixf1",
   "source_url": "https://omlx.ai/benchmarks/performance/xz49ixf1",
   "tg_tokens": null,
   "tps": 13.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-wz934g5c-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: a568njap, iadnkdlq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 257.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 257.3 31.0 19.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wz934g5c",
   "source_url": "https://omlx.ai/benchmarks/performance/wz934g5c",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-wz934g5c-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: a568njap, iadnkdlq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 250.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 250.7 32.0 19.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wz934g5c",
   "source_url": "https://omlx.ai/benchmarks/performance/wz934g5c",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-wz934g5c-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: a568njap, iadnkdlq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 226.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 226.0 26.0 22.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wz934g5c",
   "source_url": "https://omlx.ai/benchmarks/performance/wz934g5c",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-4i7wy00o-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: uufeymew, zr0toah6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 341.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 341.0 34.4 24.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4i7wy00o",
   "source_url": "https://omlx.ai/benchmarks/performance/4i7wy00o",
   "tg_tokens": null,
   "tps": 34.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-4i7wy00o-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: uufeymew, zr0toah6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 330.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 330.7 44.9 25.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4i7wy00o",
   "source_url": "https://omlx.ai/benchmarks/performance/4i7wy00o",
   "tg_tokens": null,
   "tps": 44.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-4i7wy00o-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.9 GB at this context; duplicate traces of the same session: uufeymew, zr0toah6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 285.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 285.4 34.5 28.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4i7wy00o",
   "source_url": "https://omlx.ai/benchmarks/performance/4i7wy00o",
   "tg_tokens": null,
   "tps": 34.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-w6snp6qs-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 29.9 GB; TTFT 1620834 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 80.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 128k tokens 80.9 PP tok/s 19.1 TG tok/s 1620834 TTFT (ms) 29.9 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w6snp6qs",
   "source_url": "https://omlx.ai/benchmarks/performance/w6snp6qs",
   "tg_tokens": null,
   "tps": 19.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-4tbdwznv-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 29.9 GB; TTFT 1626764 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 80.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 128k tokens 80.6 PP tok/s 7.7 TG tok/s 1626764 TTFT (ms) 29.9 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4tbdwznv",
   "source_url": "https://omlx.ai/benchmarks/performance/4tbdwznv",
   "tg_tokens": null,
   "tps": 7.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-vj1zy202-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 41.2 GB at this context; duplicate traces of the same session: 0v70qptd, a3zztsw5, babklw5s, jhsbatyb; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 357.6,
   "provenance": "community",
   "quant": "oQ6e-fp16",
   "quote": "Qwen3.8-27B-oQ6e-fp16-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 6bit; 1k 357.6 22.4 41.2 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vj1zy202",
   "source_url": "https://omlx.ai/benchmarks/performance/vj1zy202",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-vj1zy202-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 44.4 GB at this context; duplicate traces of the same session: 0v70qptd, a3zztsw5, babklw5s, jhsbatyb; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 478.8,
   "provenance": "community",
   "quant": "oQ6e-fp16",
   "quote": "Qwen3.8-27B-oQ6e-fp16-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 6bit; 4k 478.8 29.6 44.4 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vj1zy202",
   "source_url": "https://omlx.ai/benchmarks/performance/vj1zy202",
   "tg_tokens": null,
   "tps": 29.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-vj1zy202-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 45.2 GB at this context; duplicate traces of the same session: 0v70qptd, a3zztsw5, babklw5s, jhsbatyb; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 523.6,
   "provenance": "community",
   "quant": "oQ6e-fp16",
   "quote": "Qwen3.8-27B-oQ6e-fp16-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 6bit; 8k 523.6 35.9 45.2 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vj1zy202",
   "source_url": "https://omlx.ai/benchmarks/performance/vj1zy202",
   "tg_tokens": null,
   "tps": 35.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-vj1zy202-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 46.3 GB at this context; duplicate traces of the same session: 0v70qptd, a3zztsw5, babklw5s, jhsbatyb; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 493.4,
   "provenance": "community",
   "quant": "oQ6e-fp16",
   "quote": "Qwen3.8-27B-oQ6e-fp16-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 6bit; 16k 493.4 32.9 46.3 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vj1zy202",
   "source_url": "https://omlx.ai/benchmarks/performance/vj1zy202",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-vj1zy202-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 48.2 GB at this context; duplicate traces of the same session: 0v70qptd, a3zztsw5, babklw5s, jhsbatyb; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 437.7,
   "provenance": "community",
   "quant": "oQ6e-fp16",
   "quote": "Qwen3.8-27B-oQ6e-fp16-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 6bit; 32k 437.7 46.2 48.2 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vj1zy202",
   "source_url": "https://omlx.ai/benchmarks/performance/vj1zy202",
   "tg_tokens": null,
   "tps": 46.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-zpjeyspt-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace, single run at 4k context; peak memory 121.1 GB; TTFT 2206 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1857.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 5bit; 4k tokens 1,857 PP tok/s 101.6 TG tok/s 2206 TTFT (ms) 121.1 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zpjeyspt",
   "source_url": "https://omlx.ai/benchmarks/performance/zpjeyspt",
   "tg_tokens": null,
   "tps": 101.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ygfydz81-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 29.1 GB at this context; duplicate traces of the same session: 71o95iky, aggu68sh, gfy9i6oe, ke27wd6i, lxedjpxo, z3iysrly, zu46tb2r; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 418.1,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 418.1 26.8 29.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ygfydz81",
   "source_url": "https://omlx.ai/benchmarks/performance/ygfydz81",
   "tg_tokens": null,
   "tps": 26.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ygfydz81-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 30.0 GB at this context; duplicate traces of the same session: 71o95iky, aggu68sh, gfy9i6oe, ke27wd6i, lxedjpxo, z3iysrly, zu46tb2r; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 456.1,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 456.1 27.5 30.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ygfydz81",
   "source_url": "https://omlx.ai/benchmarks/performance/ygfydz81",
   "tg_tokens": null,
   "tps": 27.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ygfydz81-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: 71o95iky, aggu68sh, gfy9i6oe, ke27wd6i, lxedjpxo, z3iysrly, zu46tb2r; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 444.4,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 444.4 26.4 30.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ygfydz81",
   "source_url": "https://omlx.ai/benchmarks/performance/ygfydz81",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ygfydz81-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 30.9 GB at this context; duplicate traces of the same session: 71o95iky, aggu68sh, gfy9i6oe, ke27wd6i, lxedjpxo, z3iysrly, zu46tb2r; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 443.4,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 16k 443.4 27.5 30.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ygfydz81",
   "source_url": "https://omlx.ai/benchmarks/performance/ygfydz81",
   "tg_tokens": null,
   "tps": 27.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ygfydz81-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: 71o95iky, aggu68sh, gfy9i6oe, ke27wd6i, lxedjpxo, z3iysrly, zu46tb2r; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 408.9,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 32k 408.9 23.9 32.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ygfydz81",
   "source_url": "https://omlx.ai/benchmarks/performance/ygfydz81",
   "tg_tokens": null,
   "tps": 23.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ygfydz81-ctx65536",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 35.0 GB at this context; duplicate traces of the same session: 71o95iky, aggu68sh, gfy9i6oe, ke27wd6i, lxedjpxo, z3iysrly, zu46tb2r; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 356.2,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 64k 356.2 19.1 35.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ygfydz81",
   "source_url": "https://omlx.ai/benchmarks/performance/ygfydz81",
   "tg_tokens": null,
   "tps": 19.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ygfydz81-ctx131072",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 40.5 GB at this context; duplicate traces of the same session: 71o95iky, aggu68sh, gfy9i6oe, ke27wd6i, lxedjpxo, z3iysrly, zu46tb2r; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 306.3,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 128k 306.3 15.7 40.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ygfydz81",
   "source_url": "https://omlx.ai/benchmarks/performance/ygfydz81",
   "tg_tokens": null,
   "tps": 15.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ygfydz81-ctx199680",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 45.6 GB at this context; duplicate traces of the same session: 71o95iky, aggu68sh, gfy9i6oe, ke27wd6i, lxedjpxo, z3iysrly, zu46tb2r; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 239.7,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 195k 239.7 11.3 45.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ygfydz81",
   "source_url": "https://omlx.ai/benchmarks/performance/ygfydz81",
   "tg_tokens": null,
   "tps": 11.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-hb85d0mb-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: 88rc93dy; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 679.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26b-a4b-it-8bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 1k 679.2 56.9 25.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hb85d0mb",
   "source_url": "https://omlx.ai/benchmarks/performance/hb85d0mb",
   "tg_tokens": null,
   "tps": 56.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-hb85d0mb-ctx8192",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: 88rc93dy; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 599.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26b-a4b-it-8bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 8k 599.7 53.0 26.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hb85d0mb",
   "source_url": "https://omlx.ai/benchmarks/performance/hb85d0mb",
   "tg_tokens": null,
   "tps": 53.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-eok1dfra-ctx65536",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 9.9 GB; TTFT 28347 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2312.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-E4B-it-MLX-8bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 8bit; 64k tokens 2,312 PP tok/s 30.6 TG tok/s 28347 TTFT (ms) 9.9 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eok1dfra",
   "source_url": "https://omlx.ai/benchmarks/performance/eok1dfra",
   "tg_tokens": null,
   "tps": 30.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-yg0g294q-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: cq57q791, kj30o5i6, kkj4j8n2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 105.9 15.5 18.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yg0g294q",
   "source_url": "https://omlx.ai/benchmarks/performance/yg0g294q",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-yg0g294q-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: cq57q791, kj30o5i6, kkj4j8n2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 119.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 119.4 15.0 19.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yg0g294q",
   "source_url": "https://omlx.ai/benchmarks/performance/yg0g294q",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-yg0g294q-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: cq57q791, kj30o5i6, kkj4j8n2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 106.6 14.4 17.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yg0g294q",
   "source_url": "https://omlx.ai/benchmarks/performance/yg0g294q",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-yg0g294q-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: cq57q791, kj30o5i6, kkj4j8n2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 96.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 96.1 12.8 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yg0g294q",
   "source_url": "https://omlx.ai/benchmarks/performance/yg0g294q",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-u6pol3q5-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 18.0 GB at this context; duplicate traces of the same session: 3uj6jy0e, ouupmdl1, uibv8z91; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 105.7 15.4 18.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u6pol3q5",
   "source_url": "https://omlx.ai/benchmarks/performance/u6pol3q5",
   "tg_tokens": null,
   "tps": 15.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-u6pol3q5-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: 3uj6jy0e, ouupmdl1, uibv8z91; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 122.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 122.0 14.7 18.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u6pol3q5",
   "source_url": "https://omlx.ai/benchmarks/performance/u6pol3q5",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-u6pol3q5-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: 3uj6jy0e, ouupmdl1, uibv8z91; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 106.6 14.3 17.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u6pol3q5",
   "source_url": "https://omlx.ai/benchmarks/performance/u6pol3q5",
   "tg_tokens": null,
   "tps": 14.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-u6pol3q5-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 3uj6jy0e, ouupmdl1, uibv8z91; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 96.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 96.1 12.8 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u6pol3q5",
   "source_url": "https://omlx.ai/benchmarks/performance/u6pol3q5",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-56tw1mqx-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: l6i96e42, ufu1bmkt, ycfzwqed; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 105.7 15.4 18.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/56tw1mqx",
   "source_url": "https://omlx.ai/benchmarks/performance/56tw1mqx",
   "tg_tokens": null,
   "tps": 15.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-56tw1mqx-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: l6i96e42, ufu1bmkt, ycfzwqed; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 152.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 152.3 15.0 19.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/56tw1mqx",
   "source_url": "https://omlx.ai/benchmarks/performance/56tw1mqx",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-56tw1mqx-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: l6i96e42, ufu1bmkt, ycfzwqed; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 155.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 155.5 14.5 19.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/56tw1mqx",
   "source_url": "https://omlx.ai/benchmarks/performance/56tw1mqx",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-56tw1mqx-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: l6i96e42, ufu1bmkt, ycfzwqed; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 99.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 99.4 12.8 20.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/56tw1mqx",
   "source_url": "https://omlx.ai/benchmarks/performance/56tw1mqx",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-316or6y7-ctx16384",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 147.4 GB at this context; duplicate traces of the same session: 8dfd1hya; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 457.0,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "DeepSeek-V4-Flash-0731-MXFP4-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 512 GB \u00b7 4bit; 16k 457.0 24.0 147.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/316or6y7",
   "source_url": "https://omlx.ai/benchmarks/performance/316or6y7",
   "tg_tokens": null,
   "tps": 24.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-316or6y7-ctx32768",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 147.5 GB at this context; duplicate traces of the same session: 8dfd1hya; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 462.2,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "DeepSeek-V4-Flash-0731-MXFP4-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 512 GB \u00b7 4bit; 32k 462.2 23.4 147.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/316or6y7",
   "source_url": "https://omlx.ai/benchmarks/performance/316or6y7",
   "tg_tokens": null,
   "tps": 23.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-49xgxdoh-ctx16384",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 157.5 GB at this context; duplicate traces of the same session: 6a92bj19; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 463.2,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "DeepSeek-V4-Flash-0731-MXFP4-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 512 GB \u00b7 4bit; 16k 463.2 32.9 157.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/49xgxdoh",
   "source_url": "https://omlx.ai/benchmarks/performance/49xgxdoh",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-49xgxdoh-ctx32768",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 157.7 GB at this context; duplicate traces of the same session: 6a92bj19; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 449.2,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "DeepSeek-V4-Flash-0731-MXFP4-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 512 GB \u00b7 4bit; 32k 449.2 46.4 157.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/49xgxdoh",
   "source_url": "https://omlx.ai/benchmarks/performance/49xgxdoh",
   "tg_tokens": null,
   "tps": 46.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-tvopjvpt-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1142.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-4bit-MLX on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k 1,142 56.2 19.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tvopjvpt",
   "source_url": "https://omlx.ai/benchmarks/performance/tvopjvpt",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-tvopjvpt-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1479.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-4bit-MLX on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 1,479 55.0 19.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tvopjvpt",
   "source_url": "https://omlx.ai/benchmarks/performance/tvopjvpt",
   "tg_tokens": null,
   "tps": 55.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-tvopjvpt-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1433.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-4bit-MLX on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 8k 1,433 52.6 19.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tvopjvpt",
   "source_url": "https://omlx.ai/benchmarks/performance/tvopjvpt",
   "tg_tokens": null,
   "tps": 52.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-tvopjvpt-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1269.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-4bit-MLX on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 16k 1,269 48.1 20.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tvopjvpt",
   "source_url": "https://omlx.ai/benchmarks/performance/tvopjvpt",
   "tg_tokens": null,
   "tps": 48.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-ibrqlc17-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: jtoz1wan; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1111.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-4bit-MLX on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k 1,111 55.3 19.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ibrqlc17",
   "source_url": "https://omlx.ai/benchmarks/performance/ibrqlc17",
   "tg_tokens": null,
   "tps": 55.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-ibrqlc17-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: jtoz1wan; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1469.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-4bit-MLX on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 1,469 52.1 19.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ibrqlc17",
   "source_url": "https://omlx.ai/benchmarks/performance/ibrqlc17",
   "tg_tokens": null,
   "tps": 52.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-eey7uk15-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 22.7 GB; TTFT 44434 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 737.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k tokens 737.5 PP tok/s 66.7 TG tok/s 44434 TTFT (ms) 22.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eey7uk15",
   "source_url": "https://omlx.ai/benchmarks/performance/eey7uk15",
   "tg_tokens": null,
   "tps": 66.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qr94ya0a-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: i8tl6q1e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3604.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 3,604 128.9 19.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qr94ya0a",
   "source_url": "https://omlx.ai/benchmarks/performance/qr94ya0a",
   "tg_tokens": null,
   "tps": 128.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qr94ya0a-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: i8tl6q1e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4468.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 4,468 117.7 20.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qr94ya0a",
   "source_url": "https://omlx.ai/benchmarks/performance/qr94ya0a",
   "tg_tokens": null,
   "tps": 117.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-dsylteyd-ctx131072",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 128k context; peak memory 28.0 GB; TTFT 196452 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 667.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k tokens 667.2 PP tok/s 59.4 TG tok/s 196452 TTFT (ms) 28.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dsylteyd",
   "source_url": "https://omlx.ai/benchmarks/performance/dsylteyd",
   "tg_tokens": null,
   "tps": 59.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-w6g8m7cm-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 33cnrvs4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1744.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,744 132.9 21.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w6g8m7cm",
   "source_url": "https://omlx.ai/benchmarks/performance/w6g8m7cm",
   "tg_tokens": null,
   "tps": 132.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-w6g8m7cm-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 33cnrvs4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2064.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,064 137.4 23.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w6g8m7cm",
   "source_url": "https://omlx.ai/benchmarks/performance/w6g8m7cm",
   "tg_tokens": null,
   "tps": 137.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-w6g8m7cm-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 33cnrvs4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1946.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,946 137.4 23.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w6g8m7cm",
   "source_url": "https://omlx.ai/benchmarks/performance/w6g8m7cm",
   "tg_tokens": null,
   "tps": 137.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-w6g8m7cm-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 33cnrvs4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1731.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,731 138.6 23.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w6g8m7cm",
   "source_url": "https://omlx.ai/benchmarks/performance/w6g8m7cm",
   "tg_tokens": null,
   "tps": 138.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-wqs93npr-ctx1024",
   "model": "Qwen3.5-2B",
   "notes": "oMLX community benchmark trace; peak memory 2.5 GB at this context; duplicate traces of the same session: xq9kvjun; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 989.9,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.5-2B-OptiQ-4bit on M2 Pro (16c); M2 Pro (16c) \u00b7 16 GB \u00b7 4bit; 1k 989.9 98.8 2.5 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wqs93npr",
   "source_url": "https://omlx.ai/benchmarks/performance/wqs93npr",
   "tg_tokens": null,
   "tps": 98.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-wqs93npr-ctx4096",
   "model": "Qwen3.5-2B",
   "notes": "oMLX community benchmark trace; peak memory 2.8 GB at this context; duplicate traces of the same session: xq9kvjun; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 976.9,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.5-2B-OptiQ-4bit on M2 Pro (16c); M2 Pro (16c) \u00b7 16 GB \u00b7 4bit; 4k 976.9 95.0 2.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wqs93npr",
   "source_url": "https://omlx.ai/benchmarks/performance/wqs93npr",
   "tg_tokens": null,
   "tps": 95.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-xc84eioz-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 3.9 GB; TTFT 9373 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 437.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.5-4B-oQ4e-fp16-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k tokens 437.0 PP tok/s 34.8 TG tok/s 9373 TTFT (ms) 3.9 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xc84eioz",
   "source_url": "https://omlx.ai/benchmarks/performance/xc84eioz",
   "tg_tokens": null,
   "tps": 34.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-5ln3uidw-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: ndzw0evc; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 216.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M2 Pro (16c); M2 Pro (16c) \u00b7 16 GB \u00b7 4bit; 1k 216.2 36.6 5.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ln3uidw",
   "source_url": "https://omlx.ai/benchmarks/performance/5ln3uidw",
   "tg_tokens": null,
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-5ln3uidw-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: ndzw0evc; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 208.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M2 Pro (16c); M2 Pro (16c) \u00b7 16 GB \u00b7 4bit; 4k 208.4 35.3 6.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ln3uidw",
   "source_url": "https://omlx.ai/benchmarks/performance/5ln3uidw",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zarn78dy-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3709.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 3,709 130.6 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zarn78dy",
   "source_url": "https://omlx.ai/benchmarks/performance/zarn78dy",
   "tg_tokens": null,
   "tps": 130.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zarn78dy-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4576.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 4,576 117.9 20.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zarn78dy",
   "source_url": "https://omlx.ai/benchmarks/performance/zarn78dy",
   "tg_tokens": null,
   "tps": 117.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-s2nukc8t-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: 6tiq91p5; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 267.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 267.7 39.8 20.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s2nukc8t",
   "source_url": "https://omlx.ai/benchmarks/performance/s2nukc8t",
   "tg_tokens": null,
   "tps": 39.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-s2nukc8t-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 6tiq91p5; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 284.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 284.6 36.1 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s2nukc8t",
   "source_url": "https://omlx.ai/benchmarks/performance/s2nukc8t",
   "tg_tokens": null,
   "tps": 36.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-s2nukc8t-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 6tiq91p5; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 264.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 264.3 32.5 22.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s2nukc8t",
   "source_url": "https://omlx.ai/benchmarks/performance/s2nukc8t",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-5gw3ct6v-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: xti0vbd0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 224.3 19.3 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5gw3ct6v",
   "source_url": "https://omlx.ai/benchmarks/performance/5gw3ct6v",
   "tg_tokens": null,
   "tps": 19.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-5gw3ct6v-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: xti0vbd0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 178.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 178.9 15.7 20.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5gw3ct6v",
   "source_url": "https://omlx.ai/benchmarks/performance/5gw3ct6v",
   "tg_tokens": null,
   "tps": 15.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-tvh25pmr-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 258.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 258.1 29.4 16.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tvh25pmr",
   "source_url": "https://omlx.ai/benchmarks/performance/tvh25pmr",
   "tg_tokens": null,
   "tps": 29.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-tvh25pmr-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 80.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 189.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 189.7 10.6 80.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tvh25pmr",
   "source_url": "https://omlx.ai/benchmarks/performance/tvh25pmr",
   "tg_tokens": null,
   "tps": 10.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-b7w0skmn-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 834.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 834.2 40.1 19.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b7w0skmn",
   "source_url": "https://omlx.ai/benchmarks/performance/b7w0skmn",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-b7w0skmn-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 874.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 874.3 45.0 21.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b7w0skmn",
   "source_url": "https://omlx.ai/benchmarks/performance/b7w0skmn",
   "tg_tokens": null,
   "tps": 45.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-zc5j0q06-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: a2dqv8gi; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 206.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 206.2 22.4 16.0 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zc5j0q06",
   "source_url": "https://omlx.ai/benchmarks/performance/zc5j0q06",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-zc5j0q06-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: a2dqv8gi; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 209.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 209.8 21.8 17.3 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zc5j0q06",
   "source_url": "https://omlx.ai/benchmarks/performance/zc5j0q06",
   "tg_tokens": null,
   "tps": 21.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ll2jbuw3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: covp390k, dnedwstn; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 35.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 35.7 32.3 25.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ll2jbuw3",
   "source_url": "https://omlx.ai/benchmarks/performance/ll2jbuw3",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ll2jbuw3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: covp390k, dnedwstn; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 439.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 439.1 91.6 24.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ll2jbuw3",
   "source_url": "https://omlx.ai/benchmarks/performance/ll2jbuw3",
   "tg_tokens": null,
   "tps": 91.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ll2jbuw3-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: covp390k, dnedwstn; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 451.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 451.2 49.6 25.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ll2jbuw3",
   "source_url": "https://omlx.ai/benchmarks/performance/ll2jbuw3",
   "tg_tokens": null,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-un7rxjzv-ctx1024",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 29.9 GB; TTFT 1418 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 722.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ6e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k tokens 722.3 PP tok/s 80.1 TG tok/s 1418 TTFT (ms) 29.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/un7rxjzv",
   "source_url": "https://omlx.ai/benchmarks/performance/un7rxjzv",
   "tg_tokens": null,
   "tps": 80.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-suvov1m5-ctx1024",
   "model": "Qwen3.8-9B-Distill",
   "notes": "oMLX community benchmark trace; peak memory 9.9 GB at this context; duplicate traces of the same session: oka17spl; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 370.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-9B-Distill-MLX-8bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 8bit; 1k 370.2 28.4 9.9 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/suvov1m5",
   "source_url": "https://omlx.ai/benchmarks/performance/suvov1m5",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-suvov1m5-ctx4096",
   "model": "Qwen3.8-9B-Distill",
   "notes": "oMLX community benchmark trace; peak memory 10.5 GB at this context; duplicate traces of the same session: oka17spl; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 380.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-9B-Distill-MLX-8bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 8bit; 4k 380.6 28.0 10.5 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/suvov1m5",
   "source_url": "https://omlx.ai/benchmarks/performance/suvov1m5",
   "tg_tokens": null,
   "tps": 28.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wqqx4ai4-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 74.7 GB at this context; duplicate traces of the same session: 1esk99x6, 8bkije1n, hnkv4u3d, tnelqgf3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1143.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-Uncensored on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,143 60.1 74.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wqqx4ai4",
   "source_url": "https://omlx.ai/benchmarks/performance/wqqx4ai4",
   "tg_tokens": null,
   "tps": 60.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wqqx4ai4-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 74.9 GB at this context; duplicate traces of the same session: 1esk99x6, 8bkije1n, hnkv4u3d, tnelqgf3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1704.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-Uncensored on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,704 69.5 74.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wqqx4ai4",
   "source_url": "https://omlx.ai/benchmarks/performance/wqqx4ai4",
   "tg_tokens": null,
   "tps": 69.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wqqx4ai4-ctx65536",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 79.2 GB at this context; duplicate traces of the same session: 1esk99x6, 8bkije1n, hnkv4u3d, tnelqgf3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2099.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-Uncensored on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,099 74.3 79.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wqqx4ai4",
   "source_url": "https://omlx.ai/benchmarks/performance/wqqx4ai4",
   "tg_tokens": null,
   "tps": 74.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wqqx4ai4-ctx131072",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 84.3 GB at this context; duplicate traces of the same session: 1esk99x6, 8bkije1n, hnkv4u3d, tnelqgf3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2031.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-Uncensored on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,031 63.4 84.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wqqx4ai4",
   "source_url": "https://omlx.ai/benchmarks/performance/wqqx4ai4",
   "tg_tokens": null,
   "tps": 63.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wqqx4ai4-ctx199680",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 87.4 GB at this context; duplicate traces of the same session: 1esk99x6, 8bkije1n, hnkv4u3d, tnelqgf3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1945.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-Uncensored on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 1,945 56.9 87.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wqqx4ai4",
   "source_url": "https://omlx.ai/benchmarks/performance/wqqx4ai4",
   "tg_tokens": null,
   "tps": 56.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ahpyuj0z-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 72.0 GB at this context; duplicate traces of the same session: 2oc8q3t0, g5szi2x7, lor20srr, rod9gg08, wm1kox2m, z1adpvuv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 658.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-Uncensored on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 658.9 38.4 72.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ahpyuj0z",
   "source_url": "https://omlx.ai/benchmarks/performance/ahpyuj0z",
   "tg_tokens": null,
   "tps": 38.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ahpyuj0z-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 72.3 GB at this context; duplicate traces of the same session: 2oc8q3t0, g5szi2x7, lor20srr, rod9gg08, wm1kox2m, z1adpvuv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1414.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-Uncensored on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,414 37.6 72.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ahpyuj0z",
   "source_url": "https://omlx.ai/benchmarks/performance/ahpyuj0z",
   "tg_tokens": null,
   "tps": 37.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ahpyuj0z-ctx8192",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: 2oc8q3t0, g5szi2x7, lor20srr, rod9gg08, wm1kox2m, z1adpvuv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1926.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-Uncensored on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,926 41.3 72.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ahpyuj0z",
   "source_url": "https://omlx.ai/benchmarks/performance/ahpyuj0z",
   "tg_tokens": null,
   "tps": 41.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ahpyuj0z-ctx16384",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 73.6 GB at this context; duplicate traces of the same session: 2oc8q3t0, g5szi2x7, lor20srr, rod9gg08, wm1kox2m, z1adpvuv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2088.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-Uncensored on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,088 48.0 73.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ahpyuj0z",
   "source_url": "https://omlx.ai/benchmarks/performance/ahpyuj0z",
   "tg_tokens": null,
   "tps": 48.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ahpyuj0z-ctx65536",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 75.7 GB at this context; duplicate traces of the same session: 2oc8q3t0, g5szi2x7, lor20srr, rod9gg08, wm1kox2m, z1adpvuv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2087.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-Uncensored on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,087 45.8 75.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ahpyuj0z",
   "source_url": "https://omlx.ai/benchmarks/performance/ahpyuj0z",
   "tg_tokens": null,
   "tps": 45.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ahpyuj0z-ctx131072",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 79.5 GB at this context; duplicate traces of the same session: 2oc8q3t0, g5szi2x7, lor20srr, rod9gg08, wm1kox2m, z1adpvuv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2033.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-Uncensored on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,033 48.2 79.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ahpyuj0z",
   "source_url": "https://omlx.ai/benchmarks/performance/ahpyuj0z",
   "tg_tokens": null,
   "tps": 48.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ahpyuj0z-ctx199680",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 83.8 GB at this context; duplicate traces of the same session: 2oc8q3t0, g5szi2x7, lor20srr, rod9gg08, wm1kox2m, z1adpvuv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1949.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-Uncensored on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 1,949 45.0 83.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ahpyuj0z",
   "source_url": "https://omlx.ai/benchmarks/performance/ahpyuj0z",
   "tg_tokens": null,
   "tps": 45.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rjljbjnf-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.3 GB at this context; duplicate traces of the same session: 0dm3rhhr, 56m0z6x4, 5xztqxet, l0vpzmf4, tqzk2hfx; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 596.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 596.8 64.5 69.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rjljbjnf",
   "source_url": "https://omlx.ai/benchmarks/performance/rjljbjnf",
   "tg_tokens": null,
   "tps": 64.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rjljbjnf-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.9 GB at this context; duplicate traces of the same session: 0dm3rhhr, 56m0z6x4, 5xztqxet, l0vpzmf4, tqzk2hfx; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 840.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 840.8 67.6 69.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rjljbjnf",
   "source_url": "https://omlx.ai/benchmarks/performance/rjljbjnf",
   "tg_tokens": null,
   "tps": 67.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rjljbjnf-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.9 GB at this context; duplicate traces of the same session: 0dm3rhhr, 56m0z6x4, 5xztqxet, l0vpzmf4, tqzk2hfx; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1087.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 1,087 67.8 69.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rjljbjnf",
   "source_url": "https://omlx.ai/benchmarks/performance/rjljbjnf",
   "tg_tokens": null,
   "tps": 67.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rjljbjnf-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: 0dm3rhhr, 56m0z6x4, 5xztqxet, l0vpzmf4, tqzk2hfx; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1077.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 1,077 69.3 70.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rjljbjnf",
   "source_url": "https://omlx.ai/benchmarks/performance/rjljbjnf",
   "tg_tokens": null,
   "tps": 69.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rjljbjnf-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: 0dm3rhhr, 56m0z6x4, 5xztqxet, l0vpzmf4, tqzk2hfx; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1061.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k 1,061 62.4 71.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rjljbjnf",
   "source_url": "https://omlx.ai/benchmarks/performance/rjljbjnf",
   "tg_tokens": null,
   "tps": 62.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rjljbjnf-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.9 GB at this context; duplicate traces of the same session: 0dm3rhhr, 56m0z6x4, 5xztqxet, l0vpzmf4, tqzk2hfx; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1040.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 128k 1,040 62.1 74.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rjljbjnf",
   "source_url": "https://omlx.ai/benchmarks/performance/rjljbjnf",
   "tg_tokens": null,
   "tps": 62.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-6hhr6ob1-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 41.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 195.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 195.9 10.6 41.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6hhr6ob1",
   "source_url": "https://omlx.ai/benchmarks/performance/6hhr6ob1",
   "tg_tokens": null,
   "tps": 10.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-6hhr6ob1-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 41.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 126.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 126.7 10.7 41.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6hhr6ob1",
   "source_url": "https://omlx.ai/benchmarks/performance/6hhr6ob1",
   "tg_tokens": null,
   "tps": 10.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ab0axuoa-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 12.3 GB at this context; duplicate traces of the same session: 2jhxh6hc; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 388.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 8bit; 4k 388.0 28.6 12.3 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ab0axuoa",
   "source_url": "https://omlx.ai/benchmarks/performance/ab0axuoa",
   "tg_tokens": null,
   "tps": 28.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ab0axuoa-ctx16384",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 12.9 GB at this context; duplicate traces of the same session: 2jhxh6hc; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 399.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 8bit; 16k 399.5 23.3 12.9 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ab0axuoa",
   "source_url": "https://omlx.ai/benchmarks/performance/ab0axuoa",
   "tg_tokens": null,
   "tps": 23.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-m9zbs0dz-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: hbmqkp4q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3796.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 3,796 107.8 14.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m9zbs0dz",
   "source_url": "https://omlx.ai/benchmarks/performance/m9zbs0dz",
   "tg_tokens": null,
   "tps": 107.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-m9zbs0dz-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: hbmqkp4q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3052.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 3,052 104.2 14.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m9zbs0dz",
   "source_url": "https://omlx.ai/benchmarks/performance/m9zbs0dz",
   "tg_tokens": null,
   "tps": 104.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-mal93t2m-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: bnrsb9dl, fspv0evq, wh6io4cb; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 106.2 15.6 18.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mal93t2m",
   "source_url": "https://omlx.ai/benchmarks/performance/mal93t2m",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-mal93t2m-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: bnrsb9dl, fspv0evq, wh6io4cb; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 121.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 121.8 15.0 18.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mal93t2m",
   "source_url": "https://omlx.ai/benchmarks/performance/mal93t2m",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-mal93t2m-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: bnrsb9dl, fspv0evq, wh6io4cb; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 106.7 14.5 17.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mal93t2m",
   "source_url": "https://omlx.ai/benchmarks/performance/mal93t2m",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-mal93t2m-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: bnrsb9dl, fspv0evq, wh6io4cb; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 96.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 96.1 12.8 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mal93t2m",
   "source_url": "https://omlx.ai/benchmarks/performance/mal93t2m",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-b3nb4sik-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 172.8 GB at this context; duplicate traces of the same session: 4xp04u8q; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 362.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 362.9 37.9 172.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b3nb4sik",
   "source_url": "https://omlx.ai/benchmarks/performance/b3nb4sik",
   "tg_tokens": null,
   "tps": 37.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-b3nb4sik-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 174.2 GB at this context; duplicate traces of the same session: 4xp04u8q; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 389.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 389.1 35.9 174.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b3nb4sik",
   "source_url": "https://omlx.ai/benchmarks/performance/b3nb4sik",
   "tg_tokens": null,
   "tps": 35.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-hz6vgrz3-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 1f2r2mgb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1744.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,744 132.9 21.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hz6vgrz3",
   "source_url": "https://omlx.ai/benchmarks/performance/hz6vgrz3",
   "tg_tokens": null,
   "tps": 132.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-hz6vgrz3-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 1f2r2mgb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2064.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,064 137.4 23.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hz6vgrz3",
   "source_url": "https://omlx.ai/benchmarks/performance/hz6vgrz3",
   "tg_tokens": null,
   "tps": 137.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-hz6vgrz3-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 1f2r2mgb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1946.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,946 137.4 23.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hz6vgrz3",
   "source_url": "https://omlx.ai/benchmarks/performance/hz6vgrz3",
   "tg_tokens": null,
   "tps": 137.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-hz6vgrz3-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 1f2r2mgb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1731.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,731 138.6 23.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hz6vgrz3",
   "source_url": "https://omlx.ai/benchmarks/performance/hz6vgrz3",
   "tg_tokens": null,
   "tps": 138.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-dzw59r2a-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: h8xd4oui, numwen31, zv87mnb8; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 427.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 427.9 37.9 3.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dzw59r2a",
   "source_url": "https://omlx.ai/benchmarks/performance/dzw59r2a",
   "tg_tokens": null,
   "tps": 37.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-dzw59r2a-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.7 GB at this context; duplicate traces of the same session: h8xd4oui, numwen31, zv87mnb8; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 435.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 435.8 36.5 3.7 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dzw59r2a",
   "source_url": "https://omlx.ai/benchmarks/performance/dzw59r2a",
   "tg_tokens": null,
   "tps": 36.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-dzw59r2a-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.0 GB at this context; duplicate traces of the same session: h8xd4oui, numwen31, zv87mnb8; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 424.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 8k 424.2 34.5 4.0 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dzw59r2a",
   "source_url": "https://omlx.ai/benchmarks/performance/dzw59r2a",
   "tg_tokens": null,
   "tps": 34.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-dzw59r2a-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.4 GB at this context; duplicate traces of the same session: h8xd4oui, numwen31, zv87mnb8; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 395.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 16k 395.6 30.9 4.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dzw59r2a",
   "source_url": "https://omlx.ai/benchmarks/performance/dzw59r2a",
   "tg_tokens": null,
   "tps": 30.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-2u13l9gq-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: yrt40ff7; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 859.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 859.4 22.9 6.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2u13l9gq",
   "source_url": "https://omlx.ai/benchmarks/performance/2u13l9gq",
   "tg_tokens": null,
   "tps": 22.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-2u13l9gq-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.6 GB at this context; duplicate traces of the same session: yrt40ff7; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 838.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 16k 838.6 22.8 6.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2u13l9gq",
   "source_url": "https://omlx.ai/benchmarks/performance/2u13l9gq",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8rwxycre-ctx65536",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 8.3 GB; TTFT 23350 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2807.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k tokens 2,807 PP tok/s 74.5 TG tok/s 23350 TTFT (ms) 8.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8rwxycre",
   "source_url": "https://omlx.ai/benchmarks/performance/8rwxycre",
   "tg_tokens": null,
   "tps": 74.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-08eagg25-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 9qsqyjhx; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 565.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 565.5 68.6 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/08eagg25",
   "source_url": "https://omlx.ai/benchmarks/performance/08eagg25",
   "tg_tokens": null,
   "tps": 68.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-08eagg25-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 9qsqyjhx; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 604.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 604.3 66.3 21.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/08eagg25",
   "source_url": "https://omlx.ai/benchmarks/performance/08eagg25",
   "tg_tokens": null,
   "tps": 66.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-90eixai2-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 267.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 267.7 39.8 20.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/90eixai2",
   "source_url": "https://omlx.ai/benchmarks/performance/90eixai2",
   "tg_tokens": null,
   "tps": 39.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-90eixai2-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 284.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 284.6 36.1 21.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/90eixai2",
   "source_url": "https://omlx.ai/benchmarks/performance/90eixai2",
   "tg_tokens": null,
   "tps": 36.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-90eixai2-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 264.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 264.3 32.5 22.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/90eixai2",
   "source_url": "https://omlx.ai/benchmarks/performance/90eixai2",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-e6ooekye-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 20.5 GB; TTFT 3873 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 264.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k tokens 264.4 PP tok/s 40.1 TG tok/s 3873 TTFT (ms) 20.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e6ooekye",
   "source_url": "https://omlx.ai/benchmarks/performance/e6ooekye",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-syx0c4gc-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 8k context; peak memory 21.3 GB; TTFT 2374 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3451.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k tokens 3,451 PP tok/s 114.1 TG tok/s 2374 TTFT (ms) 21.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/syx0c4gc",
   "source_url": "https://omlx.ai/benchmarks/performance/syx0c4gc",
   "tg_tokens": null,
   "tps": 114.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rbkzrtth-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 20.6 GB; TTFT 1205 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 850.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 850.0 PP tok/s 121.8 TG tok/s 1205 TTFT (ms) 20.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rbkzrtth",
   "source_url": "https://omlx.ai/benchmarks/performance/rbkzrtth",
   "tg_tokens": null,
   "tps": 121.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qed57all-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 20.6 GB; TTFT 210 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4883.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 4,883 PP tok/s 121.4 TG tok/s 210 TTFT (ms) 20.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qed57all",
   "source_url": "https://omlx.ai/benchmarks/performance/qed57all",
   "tg_tokens": null,
   "tps": 121.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-pqrmyuty-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 0s8hpcqv, 24dq03ca; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 91.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 91.3 10.2 19.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pqrmyuty",
   "source_url": "https://omlx.ai/benchmarks/performance/pqrmyuty",
   "tg_tokens": null,
   "tps": 10.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-pqrmyuty-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 0s8hpcqv, 24dq03ca; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 79.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 79.5 8.9 21.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pqrmyuty",
   "source_url": "https://omlx.ai/benchmarks/performance/pqrmyuty",
   "tg_tokens": null,
   "tps": 8.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-pqrmyuty-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.7 GB at this context; duplicate traces of the same session: 0s8hpcqv, 24dq03ca; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 51.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 128k 51.7 5.7 28.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pqrmyuty",
   "source_url": "https://omlx.ai/benchmarks/performance/pqrmyuty",
   "tg_tokens": null,
   "tps": 5.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1cn0glg2-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: r0uglvkg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 852.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 852.1 33.5 16.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1cn0glg2",
   "source_url": "https://omlx.ai/benchmarks/performance/1cn0glg2",
   "tg_tokens": null,
   "tps": 33.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1cn0glg2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: r0uglvkg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 966.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 966.1 32.9 17.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1cn0glg2",
   "source_url": "https://omlx.ai/benchmarks/performance/1cn0glg2",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2axdky8s-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: ras8jl4h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 941.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 941.7 33.9 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2axdky8s",
   "source_url": "https://omlx.ai/benchmarks/performance/2axdky8s",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2axdky8s-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: ras8jl4h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1019.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,019 33.3 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2axdky8s",
   "source_url": "https://omlx.ai/benchmarks/performance/2axdky8s",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5r8v5en3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: cotnvv4j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 940.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 940.4 30.3 34.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5r8v5en3",
   "source_url": "https://omlx.ai/benchmarks/performance/5r8v5en3",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5r8v5en3-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 35.4 GB at this context; duplicate traces of the same session: cotnvv4j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 827.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 827.6 30.2 35.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5r8v5en3",
   "source_url": "https://omlx.ai/benchmarks/performance/5r8v5en3",
   "tg_tokens": null,
   "tps": 30.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-tgl8sos5-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 23.1 GB; TTFT 82171 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 797.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k tokens 797.6 PP tok/s 57.1 TG tok/s 82171 TTFT (ms) 23.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tgl8sos5",
   "source_url": "https://omlx.ai/benchmarks/performance/tgl8sos5",
   "tg_tokens": null,
   "tps": 57.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qn4j6rc7-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: b1c5xp9y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 913.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 913.6 67.5 17.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qn4j6rc7",
   "source_url": "https://omlx.ai/benchmarks/performance/qn4j6rc7",
   "tg_tokens": null,
   "tps": 67.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qn4j6rc7-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: b1c5xp9y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 990.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 990.2 73.7 18.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qn4j6rc7",
   "source_url": "https://omlx.ai/benchmarks/performance/qn4j6rc7",
   "tg_tokens": null,
   "tps": 73.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-dttnniln-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 29.9 GB; TTFT 1617270 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 81.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 128k tokens 81.0 PP tok/s 20.8 TG tok/s 1617270 TTFT (ms) 29.9 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dttnniln",
   "source_url": "https://omlx.ai/benchmarks/performance/dttnniln",
   "tg_tokens": null,
   "tps": 20.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2blgo9x1-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 105.2 GB at this context; duplicate traces of the same session: xhfv90zd; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 854.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 854.3 37.6 105.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2blgo9x1",
   "source_url": "https://omlx.ai/benchmarks/performance/2blgo9x1",
   "tg_tokens": null,
   "tps": 37.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2blgo9x1-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 107.0 GB at this context; duplicate traces of the same session: xhfv90zd; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 803.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 803.3 36.0 107.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2blgo9x1",
   "source_url": "https://omlx.ai/benchmarks/performance/2blgo9x1",
   "tg_tokens": null,
   "tps": 36.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-c5u3g7hk-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 69.1 GB at this context; duplicate traces of the same session: 3icwurn3, hg3irvb4, rbyjc0jq, rstre9t2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1300.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,300 76.0 69.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c5u3g7hk",
   "source_url": "https://omlx.ai/benchmarks/performance/c5u3g7hk",
   "tg_tokens": null,
   "tps": 76.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-c5u3g7hk-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 3icwurn3, hg3irvb4, rbyjc0jq, rstre9t2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2924.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,924 70.6 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c5u3g7hk",
   "source_url": "https://omlx.ai/benchmarks/performance/c5u3g7hk",
   "tg_tokens": null,
   "tps": 70.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-c5u3g7hk-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 73.3 GB at this context; duplicate traces of the same session: 3icwurn3, hg3irvb4, rbyjc0jq, rstre9t2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2868.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,868 68.6 73.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c5u3g7hk",
   "source_url": "https://omlx.ai/benchmarks/performance/c5u3g7hk",
   "tg_tokens": null,
   "tps": 68.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-c5u3g7hk-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 76.7 GB at this context; duplicate traces of the same session: 3icwurn3, hg3irvb4, rbyjc0jq, rstre9t2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2779.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,779 68.2 76.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c5u3g7hk",
   "source_url": "https://omlx.ai/benchmarks/performance/c5u3g7hk",
   "tg_tokens": null,
   "tps": 68.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-c5u3g7hk-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 80.8 GB at this context; duplicate traces of the same session: 3icwurn3, hg3irvb4, rbyjc0jq, rstre9t2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2650.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,650 66.3 80.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c5u3g7hk",
   "source_url": "https://omlx.ai/benchmarks/performance/c5u3g7hk",
   "tg_tokens": null,
   "tps": 66.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ez41qy2z-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 41.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 195.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 195.9 10.6 41.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ez41qy2z",
   "source_url": "https://omlx.ai/benchmarks/performance/ez41qy2z",
   "tg_tokens": null,
   "tps": 10.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ez41qy2z-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 41.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 126.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 126.7 10.7 41.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ez41qy2z",
   "source_url": "https://omlx.ai/benchmarks/performance/ez41qy2z",
   "tg_tokens": null,
   "tps": 10.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-q9ommemt-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 41.6 GB at this context; duplicate traces of the same session: vqucz1n6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 198.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 198.3 8.9 41.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q9ommemt",
   "source_url": "https://omlx.ai/benchmarks/performance/q9ommemt",
   "tg_tokens": null,
   "tps": 8.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-q9ommemt-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 41.6 GB at this context; duplicate traces of the same session: vqucz1n6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 134.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 134.4 11.1 41.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q9ommemt",
   "source_url": "https://omlx.ai/benchmarks/performance/q9ommemt",
   "tg_tokens": null,
   "tps": 11.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-v7xwlqby-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 41.6 GB; TTFT 68807 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 119.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k tokens 119.1 PP tok/s 11.6 TG tok/s 68807 TTFT (ms) 41.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v7xwlqby",
   "source_url": "https://omlx.ai/benchmarks/performance/v7xwlqby",
   "tg_tokens": null,
   "tps": 11.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-jeno3ykm-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.3 GB at this context; duplicate traces of the same session: mj4vg4gc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 187.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 187.4 9.7 43.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jeno3ykm",
   "source_url": "https://omlx.ai/benchmarks/performance/jeno3ykm",
   "tg_tokens": null,
   "tps": 9.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-jeno3ykm-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.1 GB at this context; duplicate traces of the same session: mj4vg4gc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 126.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 126.0 12.8 43.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jeno3ykm",
   "source_url": "https://omlx.ai/benchmarks/performance/jeno3ykm",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ovtzddmm-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.0 GB at this context; duplicate traces of the same session: 207ibtai, 6bnfegxu, aemyezws, mh798jnm, r4ppxhmx, wd5h5ma1; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 525.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 525.1 123.6 99.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ovtzddmm",
   "source_url": "https://omlx.ai/benchmarks/performance/ovtzddmm",
   "tg_tokens": null,
   "tps": 123.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ovtzddmm-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.5 GB at this context; duplicate traces of the same session: 207ibtai, 6bnfegxu, aemyezws, mh798jnm, r4ppxhmx, wd5h5ma1; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1973.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,973 112.9 99.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ovtzddmm",
   "source_url": "https://omlx.ai/benchmarks/performance/ovtzddmm",
   "tg_tokens": null,
   "tps": 112.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ovtzddmm-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.4 GB at this context; duplicate traces of the same session: 207ibtai, 6bnfegxu, aemyezws, mh798jnm, r4ppxhmx, wd5h5ma1; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2406.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 2,406 113.5 100.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ovtzddmm",
   "source_url": "https://omlx.ai/benchmarks/performance/ovtzddmm",
   "tg_tokens": null,
   "tps": 113.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ovtzddmm-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.2 GB at this context; duplicate traces of the same session: 207ibtai, 6bnfegxu, aemyezws, mh798jnm, r4ppxhmx, wd5h5ma1; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4834.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 4,834 103.8 101.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ovtzddmm",
   "source_url": "https://omlx.ai/benchmarks/performance/ovtzddmm",
   "tg_tokens": null,
   "tps": 103.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ovtzddmm-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.3 GB at this context; duplicate traces of the same session: 207ibtai, 6bnfegxu, aemyezws, mh798jnm, r4ppxhmx, wd5h5ma1; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4749.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 4,749 79.4 102.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ovtzddmm",
   "source_url": "https://omlx.ai/benchmarks/performance/ovtzddmm",
   "tg_tokens": null,
   "tps": 79.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ovtzddmm-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.7 GB at this context; duplicate traces of the same session: 207ibtai, 6bnfegxu, aemyezws, mh798jnm, r4ppxhmx, wd5h5ma1; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4624.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 4,624 89.2 104.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ovtzddmm",
   "source_url": "https://omlx.ai/benchmarks/performance/ovtzddmm",
   "tg_tokens": null,
   "tps": 89.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ovtzddmm-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 108.2 GB at this context; duplicate traces of the same session: 207ibtai, 6bnfegxu, aemyezws, mh798jnm, r4ppxhmx, wd5h5ma1; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4477.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 4,477 94.5 108.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ovtzddmm",
   "source_url": "https://omlx.ai/benchmarks/performance/ovtzddmm",
   "tg_tokens": null,
   "tps": 94.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ynykawq4-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: 75doqdl4, hm8xu1vz, qty1lycj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 106.0 15.5 18.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ynykawq4",
   "source_url": "https://omlx.ai/benchmarks/performance/ynykawq4",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ynykawq4-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: 75doqdl4, hm8xu1vz, qty1lycj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 121.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 121.2 15.0 18.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ynykawq4",
   "source_url": "https://omlx.ai/benchmarks/performance/ynykawq4",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ynykawq4-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: 75doqdl4, hm8xu1vz, qty1lycj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 106.6 14.3 17.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ynykawq4",
   "source_url": "https://omlx.ai/benchmarks/performance/ynykawq4",
   "tg_tokens": null,
   "tps": 14.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ynykawq4-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 75doqdl4, hm8xu1vz, qty1lycj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 96.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 96.0 12.8 19.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ynykawq4",
   "source_url": "https://omlx.ai/benchmarks/performance/ynykawq4",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-9unne5xz-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: 3d4ctfw6, trpt1v06; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 106.0 15.5 18.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9unne5xz",
   "source_url": "https://omlx.ai/benchmarks/performance/9unne5xz",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-9unne5xz-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: 3d4ctfw6, trpt1v06; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 121.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 121.3 15.0 18.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9unne5xz",
   "source_url": "https://omlx.ai/benchmarks/performance/9unne5xz",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-9unne5xz-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: 3d4ctfw6, trpt1v06; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 106.7 14.5 17.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9unne5xz",
   "source_url": "https://omlx.ai/benchmarks/performance/9unne5xz",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-9unne5xz-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 3d4ctfw6, trpt1v06; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 96.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 96.1 12.8 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9unne5xz",
   "source_url": "https://omlx.ai/benchmarks/performance/9unne5xz",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-y7ydt4g2-ctx4096",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "prism-ml MLX build; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: dtw52of8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 126.4,
   "provenance": "community",
   "quant": "2bit",
   "quote": "prism-ml--Ternary-Bonsai-2-27B-mlx-2bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 unknown; 4k 126.4 18.1 19.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y7ydt4g2",
   "source_url": "https://omlx.ai/benchmarks/performance/y7ydt4g2",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-y7ydt4g2-ctx16384",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "prism-ml MLX build; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: dtw52of8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 117.5,
   "provenance": "community",
   "quant": "2bit",
   "quote": "prism-ml--Ternary-Bonsai-2-27B-mlx-2bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 117.5 16.9 19.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y7ydt4g2",
   "source_url": "https://omlx.ai/benchmarks/performance/y7ydt4g2",
   "tg_tokens": null,
   "tps": 16.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-lmzaf115-ctx4096",
   "model": "Laguna-S-2.1",
   "notes": "oMLX community benchmark trace; peak memory 45.9 GB at this context; duplicate traces of the same session: flsrl0yl; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 113.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Laguna-S-2.1-oQ6e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 113.0 8.5 45.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lmzaf115",
   "source_url": "https://omlx.ai/benchmarks/performance/lmzaf115",
   "tg_tokens": null,
   "tps": 8.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-lmzaf115-ctx16384",
   "model": "Laguna-S-2.1",
   "notes": "oMLX community benchmark trace; peak memory 46.5 GB at this context; duplicate traces of the same session: flsrl0yl; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 115.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Laguna-S-2.1-oQ6e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 16k 115.2 9.0 46.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lmzaf115",
   "source_url": "https://omlx.ai/benchmarks/performance/lmzaf115",
   "tg_tokens": null,
   "tps": 9.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fskjq5lo-ctx4096",
   "model": "Laguna-S-2.1",
   "notes": "oMLX community benchmark trace; peak memory 50.0 GB at this context; duplicate traces of the same session: 6fzmtqo2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 115.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Laguna-S-2.1-oQ6e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 115.6 9.3 50.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fskjq5lo",
   "source_url": "https://omlx.ai/benchmarks/performance/fskjq5lo",
   "tg_tokens": null,
   "tps": 9.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fskjq5lo-ctx16384",
   "model": "Laguna-S-2.1",
   "notes": "oMLX community benchmark trace; peak memory 50.7 GB at this context; duplicate traces of the same session: 6fzmtqo2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 122.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Laguna-S-2.1-oQ6e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 16k 122.4 9.3 50.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fskjq5lo",
   "source_url": "https://omlx.ai/benchmarks/performance/fskjq5lo",
   "tg_tokens": null,
   "tps": 9.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-awj3o1mn-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: kmj5qqyo, lmimjmnl; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1118.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 1k 1,118 82.7 20.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/awj3o1mn",
   "source_url": "https://omlx.ai/benchmarks/performance/awj3o1mn",
   "tg_tokens": null,
   "tps": 82.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-awj3o1mn-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: kmj5qqyo, lmimjmnl; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1746.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 4k 1,746 77.7 21.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/awj3o1mn",
   "source_url": "https://omlx.ai/benchmarks/performance/awj3o1mn",
   "tg_tokens": null,
   "tps": 77.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-awj3o1mn-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: kmj5qqyo, lmimjmnl; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1638.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 8k 1,638 74.6 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/awj3o1mn",
   "source_url": "https://omlx.ai/benchmarks/performance/awj3o1mn",
   "tg_tokens": null,
   "tps": 74.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-gkqb1y8q-ctx32768",
   "model": "Qwen2.5-Coder-7B-Instruct",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 6.3 GB; TTFT 24048 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1363.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-7B-Instruct-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k tokens 1,363 PP tok/s 42.0 TG tok/s 24048 TTFT (ms) 6.3 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gkqb1y8q",
   "source_url": "https://omlx.ai/benchmarks/performance/gkqb1y8q",
   "tg_tokens": null,
   "tps": 42.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-c3a2k8xa-ctx32768",
   "model": "Qwen2.5-Coder-7B-Instruct",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 6.3 GB; TTFT 24730 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1325.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-7B-Instruct-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k tokens 1,325 PP tok/s 42.1 TG tok/s 24730 TTFT (ms) 6.3 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c3a2k8xa",
   "source_url": "https://omlx.ai/benchmarks/performance/c3a2k8xa",
   "tg_tokens": null,
   "tps": 42.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-ymqinqte-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: v0qnu9cy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 221.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 221.2 22.1 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ymqinqte",
   "source_url": "https://omlx.ai/benchmarks/performance/ymqinqte",
   "tg_tokens": null,
   "tps": 22.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-ymqinqte-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: v0qnu9cy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 229.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 229.4 22.2 19.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ymqinqte",
   "source_url": "https://omlx.ai/benchmarks/performance/ymqinqte",
   "tg_tokens": null,
   "tps": 22.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-wxl5oqjw-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: hy0em3vl; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 335.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 335.1 38.9 16.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wxl5oqjw",
   "source_url": "https://omlx.ai/benchmarks/performance/wxl5oqjw",
   "tg_tokens": null,
   "tps": 38.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-wxl5oqjw-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: hy0em3vl; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 340.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 340.1 37.6 19.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wxl5oqjw",
   "source_url": "https://omlx.ai/benchmarks/performance/wxl5oqjw",
   "tg_tokens": null,
   "tps": 37.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-aq09m8wr-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: w8216aa2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 506.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 506.9 17.7 17.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aq09m8wr",
   "source_url": "https://omlx.ai/benchmarks/performance/aq09m8wr",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-aq09m8wr-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: w8216aa2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 493.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 493.3 16.6 17.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aq09m8wr",
   "source_url": "https://omlx.ai/benchmarks/performance/aq09m8wr",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-jp3k1htt-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 8agpagay, kwy5zxjd, nvfzuj45, rov8lysb, rxyk35vb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1580.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 4bit; 1k 1,580 55.1 16.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jp3k1htt",
   "source_url": "https://omlx.ai/benchmarks/performance/jp3k1htt",
   "tg_tokens": null,
   "tps": 55.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-jp3k1htt-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 8agpagay, kwy5zxjd, nvfzuj45, rov8lysb, rxyk35vb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1858.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 4bit; 4k 1,858 53.7 17.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jp3k1htt",
   "source_url": "https://omlx.ai/benchmarks/performance/jp3k1htt",
   "tg_tokens": null,
   "tps": 53.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-jp3k1htt-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 8agpagay, kwy5zxjd, nvfzuj45, rov8lysb, rxyk35vb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1867.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 4bit; 8k 1,867 53.2 17.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jp3k1htt",
   "source_url": "https://omlx.ai/benchmarks/performance/jp3k1htt",
   "tg_tokens": null,
   "tps": 53.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-jp3k1htt-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 8agpagay, kwy5zxjd, nvfzuj45, rov8lysb, rxyk35vb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1828.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 4bit; 16k 1,828 50.6 17.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jp3k1htt",
   "source_url": "https://omlx.ai/benchmarks/performance/jp3k1htt",
   "tg_tokens": null,
   "tps": 50.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-jp3k1htt-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 8agpagay, kwy5zxjd, nvfzuj45, rov8lysb, rxyk35vb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1718.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 4bit; 32k 1,718 47.6 18.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jp3k1htt",
   "source_url": "https://omlx.ai/benchmarks/performance/jp3k1htt",
   "tg_tokens": null,
   "tps": 47.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-jp3k1htt-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 8agpagay, kwy5zxjd, nvfzuj45, rov8lysb, rxyk35vb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1524.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 4bit; 64k 1,524 42.2 20.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jp3k1htt",
   "source_url": "https://omlx.ai/benchmarks/performance/jp3k1htt",
   "tg_tokens": null,
   "tps": 42.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-pb96ktmj-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: q22z5964; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1580.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 4bit; 1k 1,580 55.1 16.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pb96ktmj",
   "source_url": "https://omlx.ai/benchmarks/performance/pb96ktmj",
   "tg_tokens": null,
   "tps": 55.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-pb96ktmj-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: q22z5964; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1800.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 4bit; 4k 1,800 53.7 17.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pb96ktmj",
   "source_url": "https://omlx.ai/benchmarks/performance/pb96ktmj",
   "tg_tokens": null,
   "tps": 53.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-7pmek6id-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 1tyrpf14; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 111.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 1k 111.4 11.3 28.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7pmek6id",
   "source_url": "https://omlx.ai/benchmarks/performance/7pmek6id",
   "tg_tokens": null,
   "tps": 11.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-7pmek6id-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.2 GB at this context; duplicate traces of the same session: 1tyrpf14; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 110.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k 110.0 11.0 32.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7pmek6id",
   "source_url": "https://omlx.ai/benchmarks/performance/7pmek6id",
   "tg_tokens": null,
   "tps": 11.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-16ry0kfc-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 22.9 GB; TTFT 53094 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 77.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k tokens 77.1 PP tok/s 16.9 TG tok/s 53094 TTFT (ms) 22.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/16ry0kfc",
   "source_url": "https://omlx.ai/benchmarks/performance/16ry0kfc",
   "tg_tokens": null,
   "tps": 16.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-zxekyw0k-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 22.8 GB; TTFT 53122 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 77.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k tokens 77.1 PP tok/s 14.0 TG tok/s 53122 TTFT (ms) 22.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zxekyw0k",
   "source_url": "https://omlx.ai/benchmarks/performance/zxekyw0k",
   "tg_tokens": null,
   "tps": 14.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ch9xmuef-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 23.1 GB; TTFT 82118 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 798.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k tokens 798.1 PP tok/s 58.5 TG tok/s 82118 TTFT (ms) 23.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ch9xmuef",
   "source_url": "https://omlx.ai/benchmarks/performance/ch9xmuef",
   "tg_tokens": null,
   "tps": 58.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-rz27g4iz-ctx1024",
   "model": "Qwen3.8-27B-OBLITERATED",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 328m3o7k; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 124.1,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-OBLITERATED-OptiQ-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 124.1 12.0 19.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rz27g4iz",
   "source_url": "https://omlx.ai/benchmarks/performance/rz27g4iz",
   "tg_tokens": null,
   "tps": 12.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-rz27g4iz-ctx4096",
   "model": "Qwen3.8-27B-OBLITERATED",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 328m3o7k; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 119.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-OBLITERATED-OptiQ-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 119.0 11.6 23.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rz27g4iz",
   "source_url": "https://omlx.ai/benchmarks/performance/rz27g4iz",
   "tg_tokens": null,
   "tps": 11.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-g1n9zpzn-ctx1024",
   "model": "Qwen3.8-27B-OBLITERATED-V3-Sharp",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-OBLITERATED-V3-Sharp-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 125.2 26.2 18.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g1n9zpzn",
   "source_url": "https://omlx.ai/benchmarks/performance/g1n9zpzn",
   "tg_tokens": null,
   "tps": 26.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-g1n9zpzn-ctx4096",
   "model": "Qwen3.8-27B-OBLITERATED-V3-Sharp",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 128.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-OBLITERATED-V3-Sharp-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 128.7 21.7 22.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g1n9zpzn",
   "source_url": "https://omlx.ai/benchmarks/performance/g1n9zpzn",
   "tg_tokens": null,
   "tps": 21.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fkzulwc8-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 86.3 GB at this context; duplicate traces of the same session: 04pdv4j9, 60dnrahb, mdh03hc2, p2j03u09, q2m37x0q, r3em3a0p, tvfgr7d4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 865.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 865.4 37.6 86.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fkzulwc8",
   "source_url": "https://omlx.ai/benchmarks/performance/fkzulwc8",
   "tg_tokens": null,
   "tps": 37.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fkzulwc8-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 86.7 GB at this context; duplicate traces of the same session: 04pdv4j9, 60dnrahb, mdh03hc2, p2j03u09, q2m37x0q, r3em3a0p, tvfgr7d4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1687.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 1,687 50.6 86.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fkzulwc8",
   "source_url": "https://omlx.ai/benchmarks/performance/fkzulwc8",
   "tg_tokens": null,
   "tps": 50.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fkzulwc8-ctx8192",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 87.8 GB at this context; duplicate traces of the same session: 04pdv4j9, 60dnrahb, mdh03hc2, p2j03u09, q2m37x0q, r3em3a0p, tvfgr7d4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2084.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 2,084 46.3 87.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fkzulwc8",
   "source_url": "https://omlx.ai/benchmarks/performance/fkzulwc8",
   "tg_tokens": null,
   "tps": 46.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fkzulwc8-ctx16384",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 89.0 GB at this context; duplicate traces of the same session: 04pdv4j9, 60dnrahb, mdh03hc2, p2j03u09, q2m37x0q, r3em3a0p, tvfgr7d4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2400.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 2,400 44.0 89.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fkzulwc8",
   "source_url": "https://omlx.ai/benchmarks/performance/fkzulwc8",
   "tg_tokens": null,
   "tps": 44.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fkzulwc8-ctx32768",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 89.9 GB at this context; duplicate traces of the same session: 04pdv4j9, 60dnrahb, mdh03hc2, p2j03u09, q2m37x0q, r3em3a0p, tvfgr7d4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2319.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 2,319 68.6 89.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fkzulwc8",
   "source_url": "https://omlx.ai/benchmarks/performance/fkzulwc8",
   "tg_tokens": null,
   "tps": 68.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fkzulwc8-ctx65536",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 91.2 GB at this context; duplicate traces of the same session: 04pdv4j9, 60dnrahb, mdh03hc2, p2j03u09, q2m37x0q, r3em3a0p, tvfgr7d4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2236.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 2,236 68.6 91.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fkzulwc8",
   "source_url": "https://omlx.ai/benchmarks/performance/fkzulwc8",
   "tg_tokens": null,
   "tps": 68.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fkzulwc8-ctx131072",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 94.3 GB at this context; duplicate traces of the same session: 04pdv4j9, 60dnrahb, mdh03hc2, p2j03u09, q2m37x0q, r3em3a0p, tvfgr7d4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2155.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 2,155 64.3 94.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fkzulwc8",
   "source_url": "https://omlx.ai/benchmarks/performance/fkzulwc8",
   "tg_tokens": null,
   "tps": 64.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fkzulwc8-ctx199680",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 99.6 GB at this context; duplicate traces of the same session: 04pdv4j9, 60dnrahb, mdh03hc2, p2j03u09, q2m37x0q, r3em3a0p, tvfgr7d4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2043.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 195k 2,043 51.5 99.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fkzulwc8",
   "source_url": "https://omlx.ai/benchmarks/performance/fkzulwc8",
   "tg_tokens": null,
   "tps": 51.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-8x4qreb7-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 73.2 GB; TTFT 10611 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3088.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 32k tokens 3,088 PP tok/s 79.7 TG tok/s 10611 TTFT (ms) 73.2 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8x4qreb7",
   "source_url": "https://omlx.ai/benchmarks/performance/8x4qreb7",
   "tg_tokens": null,
   "tps": 79.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-vou758am-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: gfvrx4xs, tm30vm2b, vm0vel5b; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1129.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k 1,129 80.9 70.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vou758am",
   "source_url": "https://omlx.ai/benchmarks/performance/vou758am",
   "tg_tokens": null,
   "tps": 80.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-vou758am-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: gfvrx4xs, tm30vm2b, vm0vel5b; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2206.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 4k 2,206 75.2 70.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vou758am",
   "source_url": "https://omlx.ai/benchmarks/performance/vou758am",
   "tg_tokens": null,
   "tps": 75.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-vou758am-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: gfvrx4xs, tm30vm2b, vm0vel5b; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2369.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 8k 2,369 80.1 71.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vou758am",
   "source_url": "https://omlx.ai/benchmarks/performance/vou758am",
   "tg_tokens": null,
   "tps": 80.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-vou758am-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.2 GB at this context; duplicate traces of the same session: gfvrx4xs, tm30vm2b, vm0vel5b; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4102.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 32k 4,102 42.1 73.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vou758am",
   "source_url": "https://omlx.ai/benchmarks/performance/vou758am",
   "tg_tokens": null,
   "tps": 42.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-6wxmmrrp-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.2 GB at this context; duplicate traces of the same session: sbqobzg6, u37729ga; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3169.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 32k 3,169 96.3 73.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6wxmmrrp",
   "source_url": "https://omlx.ai/benchmarks/performance/6wxmmrrp",
   "tg_tokens": null,
   "tps": 96.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-6wxmmrrp-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.0 GB at this context; duplicate traces of the same session: sbqobzg6, u37729ga; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3705.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 64k 3,705 65.1 74.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6wxmmrrp",
   "source_url": "https://omlx.ai/benchmarks/performance/6wxmmrrp",
   "tg_tokens": null,
   "tps": 65.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-6wxmmrrp-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.7 GB at this context; duplicate traces of the same session: sbqobzg6, u37729ga; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2538.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 128k 2,538 34.2 78.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6wxmmrrp",
   "source_url": "https://omlx.ai/benchmarks/performance/6wxmmrrp",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-h0rykpyr-ctx4096",
   "model": "Spark-X2.5-4B-Abliterated-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 3.0 GB at this context; duplicate traces of the same session: l8ih4bcy; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 152.1,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Spark-X2.5-4B-Abliterated-Uncensored-MLX-oQ4e-fp16 on M3 (10c); M3 (10c) \u00b7 16 GB \u00b7 4bit; 4k 152.1 11.4 3.0 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h0rykpyr",
   "source_url": "https://omlx.ai/benchmarks/performance/h0rykpyr",
   "tg_tokens": null,
   "tps": 11.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-h0rykpyr-ctx16384",
   "model": "Spark-X2.5-4B-Abliterated-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: l8ih4bcy; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.3,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Spark-X2.5-4B-Abliterated-Uncensored-MLX-oQ4e-fp16 on M3 (10c); M3 (10c) \u00b7 16 GB \u00b7 4bit; 16k 105.3 4.7 3.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h0rykpyr",
   "source_url": "https://omlx.ai/benchmarks/performance/h0rykpyr",
   "tg_tokens": null,
   "tps": 4.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-bhuekvvf-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 197.5 GB at this context; duplicate traces of the same session: jdfmyll1; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 530.8,
   "provenance": "community",
   "quant": "oQ3e-2b27-CED",
   "quote": "keys-Mac-oMLX-0.7.0.dev2-DeepSeek-V4.1-Flash-oQ3e-2b27-CED-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 unknown; 1k 530.8 24.6 197.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhuekvvf",
   "source_url": "https://omlx.ai/benchmarks/performance/bhuekvvf",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-bhuekvvf-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 198.0 GB at this context; duplicate traces of the same session: jdfmyll1; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 865.5,
   "provenance": "community",
   "quant": "oQ3e-2b27-CED",
   "quote": "keys-Mac-oMLX-0.7.0.dev2-DeepSeek-V4.1-Flash-oQ3e-2b27-CED-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 unknown; 4k 865.5 32.3 198.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhuekvvf",
   "source_url": "https://omlx.ai/benchmarks/performance/bhuekvvf",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-jd9e61as-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 106.0 15.5 18.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jd9e61as",
   "source_url": "https://omlx.ai/benchmarks/performance/jd9e61as",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-jd9e61as-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 121.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 121.3 15.0 18.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jd9e61as",
   "source_url": "https://omlx.ai/benchmarks/performance/jd9e61as",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-jd9e61as-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 106.7 14.5 17.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jd9e61as",
   "source_url": "https://omlx.ai/benchmarks/performance/jd9e61as",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-jd9e61as-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 96.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 96.1 12.8 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jd9e61as",
   "source_url": "https://omlx.ai/benchmarks/performance/jd9e61as",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-636lgmrx-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: krhimo2b, pearghsk, xjoky125; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 105.2 15.4 18.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/636lgmrx",
   "source_url": "https://omlx.ai/benchmarks/performance/636lgmrx",
   "tg_tokens": null,
   "tps": 15.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-636lgmrx-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: krhimo2b, pearghsk, xjoky125; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 150.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 150.9 15.0 19.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/636lgmrx",
   "source_url": "https://omlx.ai/benchmarks/performance/636lgmrx",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-636lgmrx-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: krhimo2b, pearghsk, xjoky125; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 151.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 151.5 14.3 19.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/636lgmrx",
   "source_url": "https://omlx.ai/benchmarks/performance/636lgmrx",
   "tg_tokens": null,
   "tps": 14.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-636lgmrx-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: krhimo2b, pearghsk, xjoky125; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 131.6 12.8 21.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/636lgmrx",
   "source_url": "https://omlx.ai/benchmarks/performance/636lgmrx",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-12c0kksk-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 2u1zmetb, 4kkx6150, vy5tru1m; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 105.4 15.4 18.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/12c0kksk",
   "source_url": "https://omlx.ai/benchmarks/performance/12c0kksk",
   "tg_tokens": null,
   "tps": 15.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-12c0kksk-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 2u1zmetb, 4kkx6150, vy5tru1m; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 150.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 150.7 15.0 19.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/12c0kksk",
   "source_url": "https://omlx.ai/benchmarks/performance/12c0kksk",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-12c0kksk-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: 2u1zmetb, 4kkx6150, vy5tru1m; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 151.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 151.5 14.5 19.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/12c0kksk",
   "source_url": "https://omlx.ai/benchmarks/performance/12c0kksk",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-12c0kksk-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 2u1zmetb, 4kkx6150, vy5tru1m; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 131.2 12.8 21.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/12c0kksk",
   "source_url": "https://omlx.ai/benchmarks/performance/12c0kksk",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-wb0ivwhg-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 174.4 GB; TTFT 8261 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1983.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k tokens 1,983 PP tok/s 70.5 TG tok/s 8261 TTFT (ms) 174.4 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wb0ivwhg",
   "source_url": "https://omlx.ai/benchmarks/performance/wb0ivwhg",
   "tg_tokens": null,
   "tps": 70.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-goix9ghl-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 174.3 GB; TTFT 4160 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1969.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k tokens 1,969 PP tok/s 70.7 TG tok/s 4160 TTFT (ms) 174.3 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/goix9ghl",
   "source_url": "https://omlx.ai/benchmarks/performance/goix9ghl",
   "tg_tokens": null,
   "tps": 70.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-apa7thpy-ctx1024",
   "model": "K2-Horizon-3.7B",
   "notes": "oMLX community benchmark trace; peak memory 3.5 GB at this context; duplicate traces of the same session: 1ko97zne, jsacxc6w, vmm3coj1; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 996.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "K2-Horizon-3.7B-MLX-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 1k 996.3 105.3 3.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/apa7thpy",
   "source_url": "https://omlx.ai/benchmarks/performance/apa7thpy",
   "tg_tokens": null,
   "tps": 105.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-apa7thpy-ctx4096",
   "model": "K2-Horizon-3.7B",
   "notes": "oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: 1ko97zne, jsacxc6w, vmm3coj1; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 937.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "K2-Horizon-3.7B-MLX-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 4k 937.0 88.3 3.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/apa7thpy",
   "source_url": "https://omlx.ai/benchmarks/performance/apa7thpy",
   "tg_tokens": null,
   "tps": 88.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-apa7thpy-ctx8192",
   "model": "K2-Horizon-3.7B",
   "notes": "oMLX community benchmark trace; peak memory 4.5 GB at this context; duplicate traces of the same session: 1ko97zne, jsacxc6w, vmm3coj1; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 827.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "K2-Horizon-3.7B-MLX-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 8k 827.7 72.7 4.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/apa7thpy",
   "source_url": "https://omlx.ai/benchmarks/performance/apa7thpy",
   "tg_tokens": null,
   "tps": 72.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-apa7thpy-ctx16384",
   "model": "K2-Horizon-3.7B",
   "notes": "oMLX community benchmark trace; peak memory 5.4 GB at this context; duplicate traces of the same session: 1ko97zne, jsacxc6w, vmm3coj1; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 675.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "K2-Horizon-3.7B-MLX-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 16k 675.5 36.8 5.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/apa7thpy",
   "source_url": "https://omlx.ai/benchmarks/performance/apa7thpy",
   "tg_tokens": null,
   "tps": 36.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-zw6r6yd2-ctx1024",
   "model": "K2-Horizon-7B",
   "notes": "oMLX community benchmark trace; peak memory 5.5 GB at this context; duplicate traces of the same session: 1eadl98v; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 588.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "K2-Horizon-7B-MLX-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 1k 588.1 74.9 5.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zw6r6yd2",
   "source_url": "https://omlx.ai/benchmarks/performance/zw6r6yd2",
   "tg_tokens": null,
   "tps": 74.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-zw6r6yd2-ctx4096",
   "model": "K2-Horizon-7B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: 1eadl98v; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 564.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "K2-Horizon-7B-MLX-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 4k 564.0 66.9 5.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zw6r6yd2",
   "source_url": "https://omlx.ai/benchmarks/performance/zw6r6yd2",
   "tg_tokens": null,
   "tps": 66.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-zw6r6yd2-ctx8192",
   "model": "K2-Horizon-7B",
   "notes": "oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: 1eadl98v; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 524.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "K2-Horizon-7B-MLX-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 8k 524.8 57.0 6.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zw6r6yd2",
   "source_url": "https://omlx.ai/benchmarks/performance/zw6r6yd2",
   "tg_tokens": null,
   "tps": 57.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-zw6r6yd2-ctx16384",
   "model": "K2-Horizon-7B",
   "notes": "oMLX community benchmark trace; peak memory 7.5 GB at this context; duplicate traces of the same session: 1eadl98v; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 457.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "K2-Horizon-7B-MLX-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 16k 457.2 44.3 7.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zw6r6yd2",
   "source_url": "https://omlx.ai/benchmarks/performance/zw6r6yd2",
   "tg_tokens": null,
   "tps": 44.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-z3z25nvo-ctx8192",
   "model": "llama-3-1-8b-instruct",
   "notes": "oMLX community benchmark trace, single run at 8k context; peak memory 5.7 GB; TTFT 7835 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1046.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Llama-3.1-8B-Instruct-4bit on M6 (12c); M6 (12c) \u00b7 16 GB \u00b7 4bit; 8k tokens 1,046 PP tok/s 25.4 TG tok/s 7835 TTFT (ms) 5.7 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z3z25nvo",
   "source_url": "https://omlx.ai/benchmarks/performance/z3z25nvo",
   "tg_tokens": null,
   "tps": 25.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mubf0rfb-ctx4096",
   "model": "Muse Glimmer",
   "notes": "oMLX community benchmark trace; peak memory 15.5 GB at this context; duplicate traces of the same session: 7pch51de, 92dmddme, nrtknrev, x087p4c6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 955.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-30B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 955.0 32.4 15.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mubf0rfb",
   "source_url": "https://omlx.ai/benchmarks/performance/mubf0rfb",
   "tg_tokens": null,
   "tps": 32.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mubf0rfb-ctx16384",
   "model": "Muse Glimmer",
   "notes": "oMLX community benchmark trace; peak memory 15.6 GB at this context; duplicate traces of the same session: 7pch51de, 92dmddme, nrtknrev, x087p4c6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 775.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-30B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 775.4 31.4 15.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mubf0rfb",
   "source_url": "https://omlx.ai/benchmarks/performance/mubf0rfb",
   "tg_tokens": null,
   "tps": 31.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mubf0rfb-ctx32768",
   "model": "Muse Glimmer",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: 7pch51de, 92dmddme, nrtknrev, x087p4c6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 749.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-30B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 749.7 29.5 15.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mubf0rfb",
   "source_url": "https://omlx.ai/benchmarks/performance/mubf0rfb",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mubf0rfb-ctx65536",
   "model": "Muse Glimmer",
   "notes": "oMLX community benchmark trace; peak memory 16.2 GB at this context; duplicate traces of the same session: 7pch51de, 92dmddme, nrtknrev, x087p4c6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 724.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-30B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 724.5 27.8 16.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mubf0rfb",
   "source_url": "https://omlx.ai/benchmarks/performance/mubf0rfb",
   "tg_tokens": null,
   "tps": 27.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mubf0rfb-ctx131072",
   "model": "Muse Glimmer",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 7pch51de, 92dmddme, nrtknrev, x087p4c6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 649.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-30B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 649.8 25.1 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mubf0rfb",
   "source_url": "https://omlx.ai/benchmarks/performance/mubf0rfb",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M2",
   "id": "omlx-bw2c6phr-ctx1024",
   "model": "Qwen3-0.6B",
   "notes": "oMLX community benchmark trace; peak memory 1.3 GB at this context; duplicate traces of the same session: 3ew89dag; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1749.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-0.6B-8bit on M2 (10c); M2 (10c) \u00b7 16 GB \u00b7 8bit; 1k 1,749 86.8 1.3 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bw2c6phr",
   "source_url": "https://omlx.ai/benchmarks/performance/bw2c6phr",
   "tg_tokens": null,
   "tps": 86.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M2",
   "id": "omlx-bw2c6phr-ctx4096",
   "model": "Qwen3-0.6B",
   "notes": "oMLX community benchmark trace; peak memory 1.6 GB at this context; duplicate traces of the same session: 3ew89dag; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1317.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-0.6B-8bit on M2 (10c); M2 (10c) \u00b7 16 GB \u00b7 8bit; 4k 1,317 60.7 1.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bw2c6phr",
   "source_url": "https://omlx.ai/benchmarks/performance/bw2c6phr",
   "tg_tokens": null,
   "tps": 60.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-zdqebfmg-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: fjacelgk; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 276.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 1k 276.1 42.7 3.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zdqebfmg",
   "source_url": "https://omlx.ai/benchmarks/performance/zdqebfmg",
   "tg_tokens": null,
   "tps": 42.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-zdqebfmg-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.0 GB at this context; duplicate traces of the same session: fjacelgk; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 295.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 4k 295.1 39.9 4.0 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zdqebfmg",
   "source_url": "https://omlx.ai/benchmarks/performance/zdqebfmg",
   "tg_tokens": null,
   "tps": 39.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ruebie8h-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 4eqpqd3y; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 887.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 887.9 81.4 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ruebie8h",
   "source_url": "https://omlx.ai/benchmarks/performance/ruebie8h",
   "tg_tokens": null,
   "tps": 81.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ruebie8h-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 4eqpqd3y; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1020.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,020 79.4 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ruebie8h",
   "source_url": "https://omlx.ai/benchmarks/performance/ruebie8h",
   "tg_tokens": null,
   "tps": 79.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-twmvnfht-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic-MTPLX",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: lld4q3l7; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 370.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MTPLX-4bit-FP16 on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 370.5 54.1 19.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/twmvnfht",
   "source_url": "https://omlx.ai/benchmarks/performance/twmvnfht",
   "tg_tokens": null,
   "tps": 54.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-twmvnfht-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic-MTPLX",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: lld4q3l7; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 472.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MTPLX-4bit-FP16 on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 472.5 52.3 20.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/twmvnfht",
   "source_url": "https://omlx.ai/benchmarks/performance/twmvnfht",
   "tg_tokens": null,
   "tps": 52.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-h21ev63y-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: 78vbu71e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1876.0,
   "provenance": "community",
   "quant": "VQ-5.4bpw",
   "quote": "Qwen3.6-35B-A3B-VQ-5.4bpw on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 2bit; 1k 1,876 82.9 23.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h21ev63y",
   "source_url": "https://omlx.ai/benchmarks/performance/h21ev63y",
   "tg_tokens": null,
   "tps": 82.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-h21ev63y-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 78vbu71e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2656.0,
   "provenance": "community",
   "quant": "VQ-5.4bpw",
   "quote": "Qwen3.6-35B-A3B-VQ-5.4bpw on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 2bit; 16k 2,656 78.6 23.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h21ev63y",
   "source_url": "https://omlx.ai/benchmarks/performance/h21ev63y",
   "tg_tokens": null,
   "tps": 78.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-6ttvd2k3-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: cmil9gbo, llbg0ptg, qbjpeaju; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1320.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 1,320 91.0 21.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ttvd2k3",
   "source_url": "https://omlx.ai/benchmarks/performance/6ttvd2k3",
   "tg_tokens": null,
   "tps": 91.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-6ttvd2k3-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: cmil9gbo, llbg0ptg, qbjpeaju; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2440.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 2,440 148.9 23.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ttvd2k3",
   "source_url": "https://omlx.ai/benchmarks/performance/6ttvd2k3",
   "tg_tokens": null,
   "tps": 148.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-6ttvd2k3-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: cmil9gbo, llbg0ptg, qbjpeaju; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1267.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k 1,267 134.3 25.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ttvd2k3",
   "source_url": "https://omlx.ai/benchmarks/performance/6ttvd2k3",
   "tg_tokens": null,
   "tps": 134.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-6ttvd2k3-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: cmil9gbo, llbg0ptg, qbjpeaju; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 828.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 128k 828.0 109.3 27.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ttvd2k3",
   "source_url": "https://omlx.ai/benchmarks/performance/6ttvd2k3",
   "tg_tokens": null,
   "tps": 109.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-62lp3dk7-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: ihm4hgqc, mpa0uslg, sov527kx; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2117.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 2,117 184.5 21.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/62lp3dk7",
   "source_url": "https://omlx.ai/benchmarks/performance/62lp3dk7",
   "tg_tokens": null,
   "tps": 184.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-62lp3dk7-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: ihm4hgqc, mpa0uslg, sov527kx; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2520.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 2,520 151.1 23.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/62lp3dk7",
   "source_url": "https://omlx.ai/benchmarks/performance/62lp3dk7",
   "tg_tokens": null,
   "tps": 151.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-62lp3dk7-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: ihm4hgqc, mpa0uslg, sov527kx; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1268.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k 1,268 133.8 25.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/62lp3dk7",
   "source_url": "https://omlx.ai/benchmarks/performance/62lp3dk7",
   "tg_tokens": null,
   "tps": 133.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-62lp3dk7-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: ihm4hgqc, mpa0uslg, sov527kx; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 828.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 128k 828.1 112.7 27.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/62lp3dk7",
   "source_url": "https://omlx.ai/benchmarks/performance/62lp3dk7",
   "tg_tokens": null,
   "tps": 112.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-v2ncr9st-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 3ipz57tz, 9iq67ci0, d36u1wqc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2092.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 2,092 176.6 21.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v2ncr9st",
   "source_url": "https://omlx.ai/benchmarks/performance/v2ncr9st",
   "tg_tokens": null,
   "tps": 176.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-v2ncr9st-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: 3ipz57tz, 9iq67ci0, d36u1wqc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2352.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 2,352 159.0 23.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v2ncr9st",
   "source_url": "https://omlx.ai/benchmarks/performance/v2ncr9st",
   "tg_tokens": null,
   "tps": 159.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-v2ncr9st-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: 3ipz57tz, 9iq67ci0, d36u1wqc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1269.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k 1,269 140.7 25.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v2ncr9st",
   "source_url": "https://omlx.ai/benchmarks/performance/v2ncr9st",
   "tg_tokens": null,
   "tps": 140.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-v2ncr9st-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: 3ipz57tz, 9iq67ci0, d36u1wqc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 828.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 128k 828.4 109.3 27.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v2ncr9st",
   "source_url": "https://omlx.ai/benchmarks/performance/v2ncr9st",
   "tg_tokens": null,
   "tps": 109.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5u26uysh-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: dbgci9vf, lznbamhu, wizj11hz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2108.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 2,108 188.7 21.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5u26uysh",
   "source_url": "https://omlx.ai/benchmarks/performance/5u26uysh",
   "tg_tokens": null,
   "tps": 188.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5u26uysh-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: dbgci9vf, lznbamhu, wizj11hz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2557.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 2,557 145.7 23.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5u26uysh",
   "source_url": "https://omlx.ai/benchmarks/performance/5u26uysh",
   "tg_tokens": null,
   "tps": 145.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5u26uysh-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: dbgci9vf, lznbamhu, wizj11hz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1270.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k 1,270 133.5 25.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5u26uysh",
   "source_url": "https://omlx.ai/benchmarks/performance/5u26uysh",
   "tg_tokens": null,
   "tps": 133.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5u26uysh-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: dbgci9vf, lznbamhu, wizj11hz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 828.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 128k 828.2 112.4 27.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5u26uysh",
   "source_url": "https://omlx.ai/benchmarks/performance/5u26uysh",
   "tg_tokens": null,
   "tps": 112.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-d4xs2o7n-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: akxw3e6j; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1031.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 6bit; 1k 1,031 65.3 29.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d4xs2o7n",
   "source_url": "https://omlx.ai/benchmarks/performance/d4xs2o7n",
   "tg_tokens": null,
   "tps": 65.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-d4xs2o7n-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: akxw3e6j; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1260.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 6bit; 4k 1,260 78.5 32.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d4xs2o7n",
   "source_url": "https://omlx.ai/benchmarks/performance/d4xs2o7n",
   "tg_tokens": null,
   "tps": 78.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7897qz5z-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: jrck5lqd, xnih3zpd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 987.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 987.4 33.1 17.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7897qz5z",
   "source_url": "https://omlx.ai/benchmarks/performance/7897qz5z",
   "tg_tokens": null,
   "tps": 33.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7897qz5z-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: jrck5lqd, xnih3zpd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 983.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 983.3 31.2 17.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7897qz5z",
   "source_url": "https://omlx.ai/benchmarks/performance/7897qz5z",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7897qz5z-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: jrck5lqd, xnih3zpd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 643.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 643.5 21.6 24.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7897qz5z",
   "source_url": "https://omlx.ai/benchmarks/performance/7897qz5z",
   "tg_tokens": null,
   "tps": 21.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-cagyr83d-ctx1024",
   "model": "Qwen3.8-27B-Atlassian-v2-goose",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: fpnfl7bo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1151.0,
   "provenance": "community",
   "quant": "Q4",
   "quote": "Qwen3.8-27B-Atlassian-v2-goose-Q4-mlx on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,151 33.4 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cagyr83d",
   "source_url": "https://omlx.ai/benchmarks/performance/cagyr83d",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-cagyr83d-ctx16384",
   "model": "Qwen3.8-27B-Atlassian-v2-goose",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: fpnfl7bo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1398.0,
   "provenance": "community",
   "quant": "Q4",
   "quote": "Qwen3.8-27B-Atlassian-v2-goose-Q4-mlx on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,398 31.2 19.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cagyr83d",
   "source_url": "https://omlx.ai/benchmarks/performance/cagyr83d",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-su0bj1h9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 04x3wr0x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 252.2,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 252.2 20.8 22.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/su0bj1h9",
   "source_url": "https://omlx.ai/benchmarks/performance/su0bj1h9",
   "tg_tokens": null,
   "tps": 20.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-su0bj1h9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: 04x3wr0x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 271.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 271.1 20.6 25.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/su0bj1h9",
   "source_url": "https://omlx.ai/benchmarks/performance/su0bj1h9",
   "tg_tokens": null,
   "tps": 20.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3xsyyw3s-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: 1yp8p8bt, hova41qu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 951.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 951.0 18.8 29.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3xsyyw3s",
   "source_url": "https://omlx.ai/benchmarks/performance/3xsyyw3s",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3xsyyw3s-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: 1yp8p8bt, hova41qu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 931.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 931.6 17.7 30.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3xsyyw3s",
   "source_url": "https://omlx.ai/benchmarks/performance/3xsyyw3s",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3xsyyw3s-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: 1yp8p8bt, hova41qu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 620.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 620.9 14.2 37.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3xsyyw3s",
   "source_url": "https://omlx.ai/benchmarks/performance/3xsyyw3s",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-vmry2t9s-ctx1024",
   "model": "Qwen3.8-27B-OBLITERATED-V3-Sharp",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-OBLITERATED-V3-Sharp-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 125.2 26.2 18.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vmry2t9s",
   "source_url": "https://omlx.ai/benchmarks/performance/vmry2t9s",
   "tg_tokens": null,
   "tps": 26.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-vmry2t9s-ctx4096",
   "model": "Qwen3.8-27B-OBLITERATED-V3-Sharp",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 128.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-OBLITERATED-V3-Sharp-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 128.7 21.7 22.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vmry2t9s",
   "source_url": "https://omlx.ai/benchmarks/performance/vmry2t9s",
   "tg_tokens": null,
   "tps": 21.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-c6y7yk3w-ctx1024",
   "model": "Qwen3.8-27B-OBLITERATED-V3-Sharp",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 18.6 GB; TTFT 8180 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-OBLITERATED-V3-Sharp-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k tokens 125.2 PP tok/s 24.8 TG tok/s 8180 TTFT (ms) 18.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c6y7yk3w",
   "source_url": "https://omlx.ai/benchmarks/performance/c6y7yk3w",
   "tg_tokens": null,
   "tps": 24.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ms733779-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 3ynhm3y1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 328.6,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 328.6 37.2 16.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ms733779",
   "source_url": "https://omlx.ai/benchmarks/performance/ms733779",
   "tg_tokens": null,
   "tps": 37.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ms733779-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 3ynhm3y1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 339.5,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 339.5 35.7 20.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ms733779",
   "source_url": "https://omlx.ai/benchmarks/performance/ms733779",
   "tg_tokens": null,
   "tps": 35.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-uz9j3ye8-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 131.7 11.8 21.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uz9j3ye8",
   "source_url": "https://omlx.ai/benchmarks/performance/uz9j3ye8",
   "tg_tokens": null,
   "tps": 11.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-uz9j3ye8-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 185.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 185.0 8.4 22.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uz9j3ye8",
   "source_url": "https://omlx.ai/benchmarks/performance/uz9j3ye8",
   "tg_tokens": null,
   "tps": 8.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-04nenqxv-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 103.8 GB; TTFT 10974 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 746.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k tokens 746.5 PP tok/s 74.9 TG tok/s 10974 TTFT (ms) 103.8 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/04nenqxv",
   "source_url": "https://omlx.ai/benchmarks/performance/04nenqxv",
   "tg_tokens": null,
   "tps": 74.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-2rz3j419-ctx1024",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 3.8 GB at this context; duplicate traces of the same session: 0fs5t0kd; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1380.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E2B-it-MLX-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 1k 1,380 54.0 3.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2rz3j419",
   "source_url": "https://omlx.ai/benchmarks/performance/2rz3j419",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-2rz3j419-ctx4096",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 3.8 GB at this context; duplicate traces of the same session: 0fs5t0kd; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1605.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E2B-it-MLX-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 4k 1,605 53.8 3.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2rz3j419",
   "source_url": "https://omlx.ai/benchmarks/performance/2rz3j419",
   "tg_tokens": null,
   "tps": 53.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-u84kbb8g-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 172.6 GB at this context; duplicate traces of the same session: mhn99wge; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 372.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 372.8 38.0 172.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u84kbb8g",
   "source_url": "https://omlx.ai/benchmarks/performance/u84kbb8g",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-u84kbb8g-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 173.7 GB at this context; duplicate traces of the same session: mhn99wge; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 450.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 450.5 32.2 173.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u84kbb8g",
   "source_url": "https://omlx.ai/benchmarks/performance/u84kbb8g",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-sq3owyyo-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 172.3 GB at this context; duplicate traces of the same session: ih38qvir; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 199.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 199.0 27.1 172.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sq3owyyo",
   "source_url": "https://omlx.ai/benchmarks/performance/sq3owyyo",
   "tg_tokens": null,
   "tps": 27.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-sq3owyyo-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 173.1 GB at this context; duplicate traces of the same session: ih38qvir; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 290.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 290.2 22.4 173.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sq3owyyo",
   "source_url": "https://omlx.ai/benchmarks/performance/sq3owyyo",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ludf6pm1-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 172.5 GB at this context; duplicate traces of the same session: 8qynhqny; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 371.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 371.8 38.0 172.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ludf6pm1",
   "source_url": "https://omlx.ai/benchmarks/performance/ludf6pm1",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ludf6pm1-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 173.7 GB at this context; duplicate traces of the same session: 8qynhqny; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 450.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 450.3 10.2 173.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ludf6pm1",
   "source_url": "https://omlx.ai/benchmarks/performance/ludf6pm1",
   "tg_tokens": null,
   "tps": 10.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-l2i97xz3-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 173.2 GB at this context; duplicate traces of the same session: s1qyx4qz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 170.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 170.9 20.4 173.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l2i97xz3",
   "source_url": "https://omlx.ai/benchmarks/performance/l2i97xz3",
   "tg_tokens": null,
   "tps": 20.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-l2i97xz3-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 173.4 GB at this context; duplicate traces of the same session: s1qyx4qz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 231.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 231.0 32.3 173.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l2i97xz3",
   "source_url": "https://omlx.ai/benchmarks/performance/l2i97xz3",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-0gp0ayge-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 173.0 GB; TTFT 15924 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 257.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k tokens 257.2 PP tok/s 32.2 TG tok/s 15924 TTFT (ms) 173.0 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0gp0ayge",
   "source_url": "https://omlx.ai/benchmarks/performance/0gp0ayge",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-oggc8567-ctx1024",
   "model": "K2-Horizon-7B",
   "notes": "oMLX community benchmark trace; peak memory 5.5 GB at this context; duplicate traces of the same session: iwlevb1x; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 588.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "K2-Horizon-7B-MLX-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 1k 588.1 74.9 5.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oggc8567",
   "source_url": "https://omlx.ai/benchmarks/performance/oggc8567",
   "tg_tokens": null,
   "tps": 74.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-oggc8567-ctx4096",
   "model": "K2-Horizon-7B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: iwlevb1x; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 564.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "K2-Horizon-7B-MLX-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 4k 564.0 66.9 5.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oggc8567",
   "source_url": "https://omlx.ai/benchmarks/performance/oggc8567",
   "tg_tokens": null,
   "tps": 66.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-oggc8567-ctx8192",
   "model": "K2-Horizon-7B",
   "notes": "oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: iwlevb1x; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 524.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "K2-Horizon-7B-MLX-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 8k 524.8 57.0 6.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oggc8567",
   "source_url": "https://omlx.ai/benchmarks/performance/oggc8567",
   "tg_tokens": null,
   "tps": 57.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-oggc8567-ctx16384",
   "model": "K2-Horizon-7B",
   "notes": "oMLX community benchmark trace; peak memory 7.5 GB at this context; duplicate traces of the same session: iwlevb1x; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 457.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "K2-Horizon-7B-MLX-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 16k 457.2 44.3 7.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oggc8567",
   "source_url": "https://omlx.ai/benchmarks/performance/oggc8567",
   "tg_tokens": null,
   "tps": 44.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-ddmimdbn-ctx1024",
   "model": "llama-3-1-8b-instruct",
   "notes": "oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: 8bvyv16y; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 998.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Llama-3.1-8B-Instruct-4bit on M6 (12c); M6 (12c) \u00b7 16 GB \u00b7 4bit; 1k 998.1 30.9 4.9 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ddmimdbn",
   "source_url": "https://omlx.ai/benchmarks/performance/ddmimdbn",
   "tg_tokens": null,
   "tps": 30.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-ddmimdbn-ctx4096",
   "model": "llama-3-1-8b-instruct",
   "notes": "oMLX community benchmark trace; peak memory 5.2 GB at this context; duplicate traces of the same session: 8bvyv16y; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1077.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Llama-3.1-8B-Instruct-4bit on M6 (12c); M6 (12c) \u00b7 16 GB \u00b7 4bit; 4k 1,077 28.5 5.2 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ddmimdbn",
   "source_url": "https://omlx.ai/benchmarks/performance/ddmimdbn",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-d5d6709m-ctx1024",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: 9xtdlg0t, rr472klo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 897.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 897.6 87.9 6.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d5d6709m",
   "source_url": "https://omlx.ai/benchmarks/performance/d5d6709m",
   "tg_tokens": null,
   "tps": 87.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-d5d6709m-ctx8192",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.0 GB at this context; duplicate traces of the same session: 9xtdlg0t, rr472klo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1166.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,166 83.4 8.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d5d6709m",
   "source_url": "https://omlx.ai/benchmarks/performance/d5d6709m",
   "tg_tokens": null,
   "tps": 83.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-d5d6709m-ctx32768",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "oMLX community benchmark trace; peak memory 9.2 GB at this context; duplicate traces of the same session: 9xtdlg0t, rr472klo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1030.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,030 71.5 9.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d5d6709m",
   "source_url": "https://omlx.ai/benchmarks/performance/d5d6709m",
   "tg_tokens": null,
   "tps": 71.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-g387y1ft-ctx1024",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: 2fc2mfkt, karkmeyt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 914.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 914.4 87.6 6.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g387y1ft",
   "source_url": "https://omlx.ai/benchmarks/performance/g387y1ft",
   "tg_tokens": null,
   "tps": 87.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-g387y1ft-ctx8192",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.0 GB at this context; duplicate traces of the same session: 2fc2mfkt, karkmeyt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1203.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,203 83.7 8.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g387y1ft",
   "source_url": "https://omlx.ai/benchmarks/performance/g387y1ft",
   "tg_tokens": null,
   "tps": 83.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-g387y1ft-ctx32768",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "oMLX community benchmark trace; peak memory 9.3 GB at this context; duplicate traces of the same session: 2fc2mfkt, karkmeyt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1031.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,031 71.6 9.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g387y1ft",
   "source_url": "https://omlx.ai/benchmarks/performance/g387y1ft",
   "tg_tokens": null,
   "tps": 71.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-zzswj1i1-ctx1024",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: 5zmkqask, n5ox1ute; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 903.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-oQ8e on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 903.0 52.9 11.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zzswj1i1",
   "source_url": "https://omlx.ai/benchmarks/performance/zzswj1i1",
   "tg_tokens": null,
   "tps": 52.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-zzswj1i1-ctx8192",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: 5zmkqask, n5ox1ute; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1159.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-oQ8e on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 1,159 51.0 12.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zzswj1i1",
   "source_url": "https://omlx.ai/benchmarks/performance/zzswj1i1",
   "tg_tokens": null,
   "tps": 51.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-zzswj1i1-ctx32768",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "oMLX community benchmark trace; peak memory 14.0 GB at this context; duplicate traces of the same session: 5zmkqask, n5ox1ute; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 978.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-oQ8e on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 978.1 46.2 14.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zzswj1i1",
   "source_url": "https://omlx.ai/benchmarks/performance/zzswj1i1",
   "tg_tokens": null,
   "tps": 46.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-qry2o9yd-ctx4096",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 24.4 GB; TTFT 3496 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1172.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-OptiQ-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k tokens 1,172 PP tok/s 91.5 TG tok/s 3496 TTFT (ms) 24.4 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qry2o9yd",
   "source_url": "https://omlx.ai/benchmarks/performance/qry2o9yd",
   "tg_tokens": null,
   "tps": 91.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n17dot8y-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 215 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4758.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 4,758 PP tok/s 144.3 TG tok/s 215 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n17dot8y",
   "source_url": "https://omlx.ai/benchmarks/performance/n17dot8y",
   "tg_tokens": null,
   "tps": 144.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jx961yx4-ctx4096",
   "model": "Qwen3-VL-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: covmaf9s, e879yf0z, h4inz9ch, i1h0iv9s, xfgvevjx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4873.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-VL-30B-A3B-Instruct-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 4,873 114.9 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jx961yx4",
   "source_url": "https://omlx.ai/benchmarks/performance/jx961yx4",
   "tg_tokens": null,
   "tps": 114.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jx961yx4-ctx16384",
   "model": "Qwen3-VL-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: covmaf9s, e879yf0z, h4inz9ch, i1h0iv9s, xfgvevjx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4023.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-VL-30B-A3B-Instruct-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 4,023 91.6 18.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jx961yx4",
   "source_url": "https://omlx.ai/benchmarks/performance/jx961yx4",
   "tg_tokens": null,
   "tps": 91.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jx961yx4-ctx32768",
   "model": "Qwen3-VL-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: covmaf9s, e879yf0z, h4inz9ch, i1h0iv9s, xfgvevjx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2422.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-VL-30B-A3B-Instruct-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 2,422 68.1 19.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jx961yx4",
   "source_url": "https://omlx.ai/benchmarks/performance/jx961yx4",
   "tg_tokens": null,
   "tps": 68.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jx961yx4-ctx65536",
   "model": "Qwen3-VL-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: covmaf9s, e879yf0z, h4inz9ch, i1h0iv9s, xfgvevjx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1369.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-VL-30B-A3B-Instruct-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 1,369 43.8 22.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jx961yx4",
   "source_url": "https://omlx.ai/benchmarks/performance/jx961yx4",
   "tg_tokens": null,
   "tps": 43.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jx961yx4-ctx131072",
   "model": "Qwen3-VL-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: covmaf9s, e879yf0z, h4inz9ch, i1h0iv9s, xfgvevjx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 776.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-VL-30B-A3B-Instruct-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 776.3 27.4 28.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jx961yx4",
   "source_url": "https://omlx.ai/benchmarks/performance/jx961yx4",
   "tg_tokens": null,
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jx961yx4-ctx199680",
   "model": "Qwen3-VL-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: covmaf9s, e879yf0z, h4inz9ch, i1h0iv9s, xfgvevjx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 522.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-VL-30B-A3B-Instruct-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 195k 522.7 19.4 34.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jx961yx4",
   "source_url": "https://omlx.ai/benchmarks/performance/jx961yx4",
   "tg_tokens": null,
   "tps": 19.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-yuuhyv20-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: d40je8lk; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1459.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 1,459 109.2 19.4 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yuuhyv20",
   "source_url": "https://omlx.ai/benchmarks/performance/yuuhyv20",
   "tg_tokens": null,
   "tps": 109.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-yuuhyv20-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: d40je8lk; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1602.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 1,602 105.0 20.1 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yuuhyv20",
   "source_url": "https://omlx.ai/benchmarks/performance/yuuhyv20",
   "tg_tokens": null,
   "tps": 105.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-m7xj11b8-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 64k context; peak memory 42.6 GB; TTFT 212050 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 309.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 64k tokens 309.1 PP tok/s 37.5 TG tok/s 212050 TTFT (ms) 42.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m7xj11b8",
   "source_url": "https://omlx.ai/benchmarks/performance/m7xj11b8",
   "tg_tokens": null,
   "tps": 37.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-m4cpoyhz-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: dc5h0rzl, rwnmtyqr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1318.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,318 33.0 18.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m4cpoyhz",
   "source_url": "https://omlx.ai/benchmarks/performance/m4cpoyhz",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-m4cpoyhz-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: dc5h0rzl, rwnmtyqr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1454.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,454 26.1 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m4cpoyhz",
   "source_url": "https://omlx.ai/benchmarks/performance/m4cpoyhz",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-m4cpoyhz-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 26.2 GB at this context; duplicate traces of the same session: dc5h0rzl, rwnmtyqr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 786.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 786.3 20.9 26.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m4cpoyhz",
   "source_url": "https://omlx.ai/benchmarks/performance/m4cpoyhz",
   "tg_tokens": null,
   "tps": 20.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-9zx4gfwg-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 8k context; peak memory 17.4 GB; TTFT 5395 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1518.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k tokens 1,518 PP tok/s 52.7 TG tok/s 5395 TTFT (ms) 17.4 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9zx4gfwg",
   "source_url": "https://omlx.ai/benchmarks/performance/9zx4gfwg",
   "tg_tokens": null,
   "tps": 52.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xczslnu2-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 325.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 325.6 23.1 28.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xczslnu2",
   "source_url": "https://omlx.ai/benchmarks/performance/xczslnu2",
   "tg_tokens": null,
   "tps": 23.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xczslnu2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.7 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 341.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 341.6 22.8 31.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xczslnu2",
   "source_url": "https://omlx.ai/benchmarks/performance/xczslnu2",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ks8ewob8-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: i5ho2y94, l82af15o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 919.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 919.2 18.3 29.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ks8ewob8",
   "source_url": "https://omlx.ai/benchmarks/performance/ks8ewob8",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ks8ewob8-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: i5ho2y94, l82af15o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 916.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 16k 916.3 17.6 30.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ks8ewob8",
   "source_url": "https://omlx.ai/benchmarks/performance/ks8ewob8",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ks8ewob8-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: i5ho2y94, l82af15o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 587.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 128k 587.8 14.1 37.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ks8ewob8",
   "source_url": "https://omlx.ai/benchmarks/performance/ks8ewob8",
   "tg_tokens": null,
   "tps": 14.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-s1287xa6-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 477.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 477.0 14.5 20.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s1287xa6",
   "source_url": "https://omlx.ai/benchmarks/performance/s1287xa6",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-s1287xa6-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 502.9,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 502.9 14.3 21.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s1287xa6",
   "source_url": "https://omlx.ai/benchmarks/performance/s1287xa6",
   "tg_tokens": null,
   "tps": 14.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hcoadks4-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 7oph21ba; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 781.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 781.0 31.1 16.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hcoadks4",
   "source_url": "https://omlx.ai/benchmarks/performance/hcoadks4",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hcoadks4-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 7oph21ba; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 913.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 913.9 30.5 17.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hcoadks4",
   "source_url": "https://omlx.ai/benchmarks/performance/hcoadks4",
   "tg_tokens": null,
   "tps": 30.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-9ldu6yg8-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: iwtvn1v4; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 159.7,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 1k 159.7 20.3 16.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ldu6yg8",
   "source_url": "https://omlx.ai/benchmarks/performance/9ldu6yg8",
   "tg_tokens": null,
   "tps": 20.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-9ldu6yg8-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: iwtvn1v4; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 165.4,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 4k 165.4 19.8 17.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ldu6yg8",
   "source_url": "https://omlx.ai/benchmarks/performance/9ldu6yg8",
   "tg_tokens": null,
   "tps": 19.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-uk5k6rtm-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: wwj9m27c; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 433.4,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 433.4 17.3 20.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uk5k6rtm",
   "source_url": "https://omlx.ai/benchmarks/performance/uk5k6rtm",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-uk5k6rtm-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: wwj9m27c; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 491.2,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 491.2 16.1 18.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uk5k6rtm",
   "source_url": "https://omlx.ai/benchmarks/performance/uk5k6rtm",
   "tg_tokens": null,
   "tps": 16.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-8cs8c0sy-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 9wrvktqd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 220.4 45.2 18.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8cs8c0sy",
   "source_url": "https://omlx.ai/benchmarks/performance/8cs8c0sy",
   "tg_tokens": null,
   "tps": 45.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-8cs8c0sy-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 9wrvktqd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 220.9 40.7 22.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8cs8c0sy",
   "source_url": "https://omlx.ai/benchmarks/performance/8cs8c0sy",
   "tg_tokens": null,
   "tps": 40.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-sr17cyq9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.9 GB at this context; duplicate traces of the same session: gd7cor7j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 887.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 887.4 58.0 17.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sr17cyq9",
   "source_url": "https://omlx.ai/benchmarks/performance/sr17cyq9",
   "tg_tokens": null,
   "tps": 58.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-sr17cyq9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: gd7cor7j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 984.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 984.6 60.5 18.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sr17cyq9",
   "source_url": "https://omlx.ai/benchmarks/performance/sr17cyq9",
   "tg_tokens": null,
   "tps": 60.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0qmtia4l-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 131.7 11.8 21.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0qmtia4l",
   "source_url": "https://omlx.ai/benchmarks/performance/0qmtia4l",
   "tg_tokens": null,
   "tps": 11.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0qmtia4l-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 185.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 185.0 8.4 22.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0qmtia4l",
   "source_url": "https://omlx.ai/benchmarks/performance/0qmtia4l",
   "tg_tokens": null,
   "tps": 8.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3l2gyjxd-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 44.9 GB at this context; duplicate traces of the same session: 8obvgn7h; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 821.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 821.5 34.8 44.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3l2gyjxd",
   "source_url": "https://omlx.ai/benchmarks/performance/3l2gyjxd",
   "tg_tokens": null,
   "tps": 34.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "outlier"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3l2gyjxd-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 44.9 GB at this context; duplicate traces of the same session: 8obvgn7h; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1270.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,270 37.6 44.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3l2gyjxd",
   "source_url": "https://omlx.ai/benchmarks/performance/3l2gyjxd",
   "tg_tokens": null,
   "tps": 37.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fznuc3xc-ctx8192",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 8k context; peak memory 45.3 GB; TTFT 6827 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1200.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k tokens 1,200 PP tok/s 33.0 TG tok/s 6827 TTFT (ms) 45.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fznuc3xc",
   "source_url": "https://omlx.ai/benchmarks/performance/fznuc3xc",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-r3k508hv-ctx4096",
   "model": "Ternary-Bonsai-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 10.4 GB at this context; duplicate traces of the same session: 8c18s0x7; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 80.1,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-27B-mlx-2bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 2bit; 4k 80.1 13.4 10.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r3k508hv",
   "source_url": "https://omlx.ai/benchmarks/performance/r3k508hv",
   "tg_tokens": null,
   "tps": 13.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-r3k508hv-ctx16384",
   "model": "Ternary-Bonsai-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 11.7 GB at this context; duplicate traces of the same session: 8c18s0x7; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 77.6,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-27B-mlx-2bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 2bit; 16k 77.6 12.1 11.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r3k508hv",
   "source_url": "https://omlx.ai/benchmarks/performance/r3k508hv",
   "tg_tokens": null,
   "tps": 12.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-drsfuywj-ctx1024",
   "model": "FrogNano-4B-2609",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: z6er5qaq; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 415.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "FrogNano-4B-2609-oQ4e-mtp-MLX on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k 415.3 36.2 3.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/drsfuywj",
   "source_url": "https://omlx.ai/benchmarks/performance/drsfuywj",
   "tg_tokens": null,
   "tps": 36.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-drsfuywj-ctx4096",
   "model": "FrogNano-4B-2609",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: z6er5qaq; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 433.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "FrogNano-4B-2609-oQ4e-mtp-MLX on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 433.1 34.6 3.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/drsfuywj",
   "source_url": "https://omlx.ai/benchmarks/performance/drsfuywj",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-kkxzs3o0-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.0 GB at this context; duplicate traces of the same session: 54gfvila, adifmbqi; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2389.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,389 120.9 175.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kkxzs3o0",
   "source_url": "https://omlx.ai/benchmarks/performance/kkxzs3o0",
   "tg_tokens": null,
   "tps": 120.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-kkxzs3o0-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.8 GB at this context; duplicate traces of the same session: 54gfvila, adifmbqi; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2368.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,368 81.9 177.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kkxzs3o0",
   "source_url": "https://omlx.ai/benchmarks/performance/kkxzs3o0",
   "tg_tokens": null,
   "tps": 81.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-kkxzs3o0-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 183.6 GB at this context; duplicate traces of the same session: 54gfvila, adifmbqi; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2322.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,322 134.4 183.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kkxzs3o0",
   "source_url": "https://omlx.ai/benchmarks/performance/kkxzs3o0",
   "tg_tokens": null,
   "tps": 134.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-3tge0dkt-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.0 GB at this context; duplicate traces of the same session: 0269i14f, 4rolxno4, 7cu2t4p9, 9v9mymze, ghpok24d; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1511.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,511 45.4 174.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3tge0dkt",
   "source_url": "https://omlx.ai/benchmarks/performance/3tge0dkt",
   "tg_tokens": null,
   "tps": 45.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-3tge0dkt-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.4 GB at this context; duplicate traces of the same session: 0269i14f, 4rolxno4, 7cu2t4p9, 9v9mymze, ghpok24d; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1974.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,974 45.6 175.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3tge0dkt",
   "source_url": "https://omlx.ai/benchmarks/performance/3tge0dkt",
   "tg_tokens": null,
   "tps": 45.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-3tge0dkt-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 176.3 GB at this context; duplicate traces of the same session: 0269i14f, 4rolxno4, 7cu2t4p9, 9v9mymze, ghpok24d; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1981.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 1,981 59.1 176.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3tge0dkt",
   "source_url": "https://omlx.ai/benchmarks/performance/3tge0dkt",
   "tg_tokens": null,
   "tps": 59.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-3tge0dkt-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.0 GB at this context; duplicate traces of the same session: 0269i14f, 4rolxno4, 7cu2t4p9, 9v9mymze, ghpok24d; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1964.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,964 82.3 177.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3tge0dkt",
   "source_url": "https://omlx.ai/benchmarks/performance/3tge0dkt",
   "tg_tokens": null,
   "tps": 82.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-3tge0dkt-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.4 GB at this context; duplicate traces of the same session: 0269i14f, 4rolxno4, 7cu2t4p9, 9v9mymze, ghpok24d; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1951.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,951 93.5 178.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3tge0dkt",
   "source_url": "https://omlx.ai/benchmarks/performance/3tge0dkt",
   "tg_tokens": null,
   "tps": 93.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-3tge0dkt-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 181.3 GB at this context; duplicate traces of the same session: 0269i14f, 4rolxno4, 7cu2t4p9, 9v9mymze, ghpok24d; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1916.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,916 58.9 181.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3tge0dkt",
   "source_url": "https://omlx.ai/benchmarks/performance/3tge0dkt",
   "tg_tokens": null,
   "tps": 58.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-3tge0dkt-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 187.0 GB at this context; duplicate traces of the same session: 0269i14f, 4rolxno4, 7cu2t4p9, 9v9mymze, ghpok24d; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1881.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 1,881 81.1 187.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3tge0dkt",
   "source_url": "https://omlx.ai/benchmarks/performance/3tge0dkt",
   "tg_tokens": null,
   "tps": 81.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-e5vgxi41-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 179.9 GB at this context; duplicate traces of the same session: 4i7fbu7c, zpluvv4h; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2345.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,345 103.7 179.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e5vgxi41",
   "source_url": "https://omlx.ai/benchmarks/performance/e5vgxi41",
   "tg_tokens": null,
   "tps": 103.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-e5vgxi41-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 182.8 GB at this context; duplicate traces of the same session: 4i7fbu7c, zpluvv4h; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2316.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,316 85.7 182.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e5vgxi41",
   "source_url": "https://omlx.ai/benchmarks/performance/e5vgxi41",
   "tg_tokens": null,
   "tps": 85.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-e5vgxi41-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 188.6 GB at this context; duplicate traces of the same session: 4i7fbu7c, zpluvv4h; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2271.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,271 119.8 188.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e5vgxi41",
   "source_url": "https://omlx.ai/benchmarks/performance/e5vgxi41",
   "tg_tokens": null,
   "tps": 119.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-td5i5otv-ctx1024",
   "model": "llama-3-1-8b-instruct",
   "notes": "oMLX community benchmark trace; peak memory 5.0 GB at this context; duplicate traces of the same session: qopa208d; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 462.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Llama-3.1-8B-Instruct-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 24 GB \u00b7 4bit; 1k 462.4 48.9 5.0 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/td5i5otv",
   "source_url": "https://omlx.ai/benchmarks/performance/td5i5otv",
   "tg_tokens": null,
   "tps": 48.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-td5i5otv-ctx4096",
   "model": "llama-3-1-8b-instruct",
   "notes": "oMLX community benchmark trace; peak memory 5.4 GB at this context; duplicate traces of the same session: qopa208d; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 442.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Llama-3.1-8B-Instruct-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 24 GB \u00b7 4bit; 4k 442.3 47.6 5.4 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/td5i5otv",
   "source_url": "https://omlx.ai/benchmarks/performance/td5i5otv",
   "tg_tokens": null,
   "tps": 47.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-17haciah-ctx16384",
   "model": "MiMo-V2.6-Qwen-9B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: d0mygahw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1389.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "MiMo-V2.6-Qwen-9B-MTPLX-Optimized-Speed on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 16k 1,389 78.1 13.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/17haciah",
   "source_url": "https://omlx.ai/benchmarks/performance/17haciah",
   "tg_tokens": null,
   "tps": 78.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-17haciah-ctx131072",
   "model": "MiMo-V2.6-Qwen-9B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: d0mygahw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 772.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "MiMo-V2.6-Qwen-9B-MTPLX-Optimized-Speed on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 128k 772.3 41.2 19.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/17haciah",
   "source_url": "https://omlx.ai/benchmarks/performance/17haciah",
   "tg_tokens": null,
   "tps": 41.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-np290cz3-ctx1024",
   "model": "Qwen2.5-Coder-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 4.8 GB at this context; duplicate traces of the same session: e78t31nu; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 247.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-7B-Instruct-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 16 GB \u00b7 4bit; 1k 247.4 34.5 4.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/np290cz3",
   "source_url": "https://omlx.ai/benchmarks/performance/np290cz3",
   "tg_tokens": null,
   "tps": 34.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-np290cz3-ctx4096",
   "model": "Qwen2.5-Coder-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: e78t31nu; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 287.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-7B-Instruct-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 16 GB \u00b7 4bit; 4k 287.4 29.9 4.9 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/np290cz3",
   "source_url": "https://omlx.ai/benchmarks/performance/np290cz3",
   "tg_tokens": null,
   "tps": 29.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-qifq3vmb-ctx1024",
   "model": "Qwen3 8B",
   "notes": "oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: glci00bb; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 755.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-8B-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 755.6 28.0 4.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qifq3vmb",
   "source_url": "https://omlx.ai/benchmarks/performance/qifq3vmb",
   "tg_tokens": null,
   "tps": 28.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-qifq3vmb-ctx4096",
   "model": "Qwen3 8B",
   "notes": "oMLX community benchmark trace; peak memory 5.4 GB at this context; duplicate traces of the same session: glci00bb; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 747.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-8B-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 747.7 25.7 5.4 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qifq3vmb",
   "source_url": "https://omlx.ai/benchmarks/performance/qifq3vmb",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-6h2l6gpi-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 22.1 GB; TTFT 46195 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 709.3,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.6-35B-A3B-nvfp4 on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k tokens 709.3 PP tok/s 52.3 TG tok/s 46195 TTFT (ms) 22.1 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6h2l6gpi",
   "source_url": "https://omlx.ai/benchmarks/performance/6h2l6gpi",
   "tg_tokens": null,
   "tps": 52.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ig4vhpr0-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 21.6 GB; TTFT 49374 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 663.7,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.6-35B-A3B-nvfp4 on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k tokens 663.7 PP tok/s 26.6 TG tok/s 49374 TTFT (ms) 21.6 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ig4vhpr0",
   "source_url": "https://omlx.ai/benchmarks/performance/ig4vhpr0",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-z7blri0o-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 21.6 GB; TTFT 49241 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 665.5,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.6-35B-A3B-nvfp4 on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k tokens 665.5 PP tok/s 23.6 TG tok/s 49241 TTFT (ms) 21.6 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z7blri0o",
   "source_url": "https://omlx.ai/benchmarks/performance/z7blri0o",
   "tg_tokens": null,
   "tps": 23.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-xm9iqrnl-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 8dvtgb7h; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 106.7 17.1 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xm9iqrnl",
   "source_url": "https://omlx.ai/benchmarks/performance/xm9iqrnl",
   "tg_tokens": null,
   "tps": 17.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-xm9iqrnl-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 8dvtgb7h; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 93.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 93.4 15.4 19.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xm9iqrnl",
   "source_url": "https://omlx.ai/benchmarks/performance/xm9iqrnl",
   "tg_tokens": null,
   "tps": 15.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9uzc824h-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: aul1ysci; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 861.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 861.1 33.7 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9uzc824h",
   "source_url": "https://omlx.ai/benchmarks/performance/9uzc824h",
   "tg_tokens": null,
   "tps": 33.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9uzc824h-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: aul1ysci; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 982.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 982.0 33.0 17.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9uzc824h",
   "source_url": "https://omlx.ai/benchmarks/performance/9uzc824h",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-it5ig1f5-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 325.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 325.6 23.1 28.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/it5ig1f5",
   "source_url": "https://omlx.ai/benchmarks/performance/it5ig1f5",
   "tg_tokens": null,
   "tps": 23.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-it5ig1f5-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.7 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 341.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 341.6 22.8 31.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/it5ig1f5",
   "source_url": "https://omlx.ai/benchmarks/performance/it5ig1f5",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-rki05br9-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 477.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 477.0 14.5 20.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rki05br9",
   "source_url": "https://omlx.ai/benchmarks/performance/rki05br9",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-rki05br9-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 502.9,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 502.9 14.3 21.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rki05br9",
   "source_url": "https://omlx.ai/benchmarks/performance/rki05br9",
   "tg_tokens": null,
   "tps": 14.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-kna35dnx-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: r28q2xp7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 215.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 215.1 19.7 16.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kna35dnx",
   "source_url": "https://omlx.ai/benchmarks/performance/kna35dnx",
   "tg_tokens": null,
   "tps": 19.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-kna35dnx-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: r28q2xp7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 219.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 219.5 18.9 19.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kna35dnx",
   "source_url": "https://omlx.ai/benchmarks/performance/kna35dnx",
   "tg_tokens": null,
   "tps": 18.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-zlzaqav2-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: eqo7srfp, mkpaf2se; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 209.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 209.4 43.2 18.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zlzaqav2",
   "source_url": "https://omlx.ai/benchmarks/performance/zlzaqav2",
   "tg_tokens": null,
   "tps": 43.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-zlzaqav2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: eqo7srfp, mkpaf2se; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 225.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 225.7 42.1 22.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zlzaqav2",
   "source_url": "https://omlx.ai/benchmarks/performance/zlzaqav2",
   "tg_tokens": null,
   "tps": 42.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-zlzaqav2-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.6 GB at this context; duplicate traces of the same session: eqo7srfp, mkpaf2se; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 153.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 153.2 19.8 27.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zlzaqav2",
   "source_url": "https://omlx.ai/benchmarks/performance/zlzaqav2",
   "tg_tokens": null,
   "tps": 19.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-h2gjiev7-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 24.6 GB; TTFT 408490 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 160.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 64k tokens 160.4 PP tok/s 14.4 TG tok/s 408490 TTFT (ms) 24.6 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h2gjiev7",
   "source_url": "https://omlx.ai/benchmarks/performance/h2gjiev7",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-j675y361-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: 1kykphj0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 468.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 468.9 23.6 25.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j675y361",
   "source_url": "https://omlx.ai/benchmarks/performance/j675y361",
   "tg_tokens": null,
   "tps": 23.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-j675y361-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: 1kykphj0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 437.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 16k 437.6 10.0 26.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j675y361",
   "source_url": "https://omlx.ai/benchmarks/performance/j675y361",
   "tg_tokens": null,
   "tps": 10.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-8brybv27-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: 2tftcvn5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 453.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 453.7 24.7 27.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8brybv27",
   "source_url": "https://omlx.ai/benchmarks/performance/8brybv27",
   "tg_tokens": null,
   "tps": 24.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-8brybv27-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: 2tftcvn5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 438.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 16k 438.5 34.7 28.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8brybv27",
   "source_url": "https://omlx.ai/benchmarks/performance/8brybv27",
   "tg_tokens": null,
   "tps": 34.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-squ85chb-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace, single run at 195k context; peak memory 151.5 GB; TTFT 204795 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 976.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 195k tokens 976.6 PP tok/s 50.6 TG tok/s 204795 TTFT (ms) 151.5 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/squ85chb",
   "source_url": "https://omlx.ai/benchmarks/performance/squ85chb",
   "tg_tokens": null,
   "tps": 50.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rmqiixyc-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 144.3 GB at this context; duplicate traces of the same session: bybh6790; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1017.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 32k 1,017 69.3 144.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rmqiixyc",
   "source_url": "https://omlx.ai/benchmarks/performance/rmqiixyc",
   "tg_tokens": null,
   "tps": 69.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rmqiixyc-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 146.7 GB at this context; duplicate traces of the same session: bybh6790; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1003.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 64k 1,003 58.3 146.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rmqiixyc",
   "source_url": "https://omlx.ai/benchmarks/performance/rmqiixyc",
   "tg_tokens": null,
   "tps": 58.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-5wro9bxz-ctx1024",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 16.6 GB; TTFT 7215 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 141.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k tokens 141.9 PP tok/s 19.8 TG tok/s 7215 TTFT (ms) 16.6 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5wro9bxz",
   "source_url": "https://omlx.ai/benchmarks/performance/5wro9bxz",
   "tg_tokens": null,
   "tps": 19.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-z21la38g-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 31.8 GB at this context; duplicate traces of the same session: 1h4bs8xm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 119.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mlx-community--Qwen3.8-27B-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 119.5 7.8 31.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z21la38g",
   "source_url": "https://omlx.ai/benchmarks/performance/z21la38g",
   "tg_tokens": null,
   "tps": 7.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-z21la38g-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 33.6 GB at this context; duplicate traces of the same session: 1h4bs8xm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mlx-community--Qwen3.8-27B-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 16k 105.8 7.6 33.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z21la38g",
   "source_url": "https://omlx.ai/benchmarks/performance/z21la38g",
   "tg_tokens": null,
   "tps": 7.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-if9wdbw4-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; rapid-mlx build; oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: sece75c4; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 141.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "rapid-mlx--Qwen3.8-27B-4bit-MTP-MLX on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 141.5 20.6 16.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/if9wdbw4",
   "source_url": "https://omlx.ai/benchmarks/performance/if9wdbw4",
   "tg_tokens": null,
   "tps": 20.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-if9wdbw4-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; rapid-mlx build; oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: sece75c4; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "rapid-mlx--Qwen3.8-27B-4bit-MTP-MLX on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 131.2 18.7 19.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/if9wdbw4",
   "source_url": "https://omlx.ai/benchmarks/performance/if9wdbw4",
   "tg_tokens": null,
   "tps": 18.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-2w8pujei-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: lmwmiwg4; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 91.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "scottlowry--Swift-1.5-Qwen3.8-27b-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 1k 91.6 16.1 26.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2w8pujei",
   "source_url": "https://omlx.ai/benchmarks/performance/2w8pujei",
   "tg_tokens": null,
   "tps": 16.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-2w8pujei-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: lmwmiwg4; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 90.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "scottlowry--Swift-1.5-Qwen3.8-27b-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 4k 90.8 16.8 27.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2w8pujei",
   "source_url": "https://omlx.ai/benchmarks/performance/2w8pujei",
   "tg_tokens": null,
   "tps": 16.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-re2pcg9t-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.0 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1511.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,511 45.4 174.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/re2pcg9t",
   "source_url": "https://omlx.ai/benchmarks/performance/re2pcg9t",
   "tg_tokens": null,
   "tps": 45.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-re2pcg9t-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.4 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1974.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,974 45.6 175.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/re2pcg9t",
   "source_url": "https://omlx.ai/benchmarks/performance/re2pcg9t",
   "tg_tokens": null,
   "tps": 45.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-re2pcg9t-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 176.3 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1981.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 1,981 59.1 176.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/re2pcg9t",
   "source_url": "https://omlx.ai/benchmarks/performance/re2pcg9t",
   "tg_tokens": null,
   "tps": 59.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-re2pcg9t-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.0 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1964.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,964 82.3 177.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/re2pcg9t",
   "source_url": "https://omlx.ai/benchmarks/performance/re2pcg9t",
   "tg_tokens": null,
   "tps": 82.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-re2pcg9t-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.4 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1951.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,951 93.5 178.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/re2pcg9t",
   "source_url": "https://omlx.ai/benchmarks/performance/re2pcg9t",
   "tg_tokens": null,
   "tps": 93.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-re2pcg9t-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 181.3 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1916.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,916 58.9 181.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/re2pcg9t",
   "source_url": "https://omlx.ai/benchmarks/performance/re2pcg9t",
   "tg_tokens": null,
   "tps": 58.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-re2pcg9t-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 187.0 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1881.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 1,881 81.1 187.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/re2pcg9t",
   "source_url": "https://omlx.ai/benchmarks/performance/re2pcg9t",
   "tg_tokens": null,
   "tps": 81.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-w9rrrgab-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.4 GB at this context; duplicate traces of the same session: 0ga6aqdt, avh80cw6, jlb22sku, o01j09ja, tkp5doda, wfw7jo1p; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1301.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,301 62.3 173.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w9rrrgab",
   "source_url": "https://omlx.ai/benchmarks/performance/w9rrrgab",
   "tg_tokens": null,
   "tps": 62.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-w9rrrgab-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.3 GB at this context; duplicate traces of the same session: 0ga6aqdt, avh80cw6, jlb22sku, o01j09ja, tkp5doda, wfw7jo1p; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1959.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,959 56.9 174.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w9rrrgab",
   "source_url": "https://omlx.ai/benchmarks/performance/w9rrrgab",
   "tg_tokens": null,
   "tps": 56.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-w9rrrgab-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.8 GB at this context; duplicate traces of the same session: 0ga6aqdt, avh80cw6, jlb22sku, o01j09ja, tkp5doda, wfw7jo1p; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1962.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 1,962 57.3 174.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w9rrrgab",
   "source_url": "https://omlx.ai/benchmarks/performance/w9rrrgab",
   "tg_tokens": null,
   "tps": 57.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-w9rrrgab-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.9 GB at this context; duplicate traces of the same session: 0ga6aqdt, avh80cw6, jlb22sku, o01j09ja, tkp5doda, wfw7jo1p; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1955.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,955 56.9 174.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w9rrrgab",
   "source_url": "https://omlx.ai/benchmarks/performance/w9rrrgab",
   "tg_tokens": null,
   "tps": 56.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-w9rrrgab-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.2 GB at this context; duplicate traces of the same session: 0ga6aqdt, avh80cw6, jlb22sku, o01j09ja, tkp5doda, wfw7jo1p; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1951.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,951 56.8 175.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w9rrrgab",
   "source_url": "https://omlx.ai/benchmarks/performance/w9rrrgab",
   "tg_tokens": null,
   "tps": 56.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-w9rrrgab-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.5 GB at this context; duplicate traces of the same session: 0ga6aqdt, avh80cw6, jlb22sku, o01j09ja, tkp5doda, wfw7jo1p; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1919.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,919 55.9 175.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w9rrrgab",
   "source_url": "https://omlx.ai/benchmarks/performance/w9rrrgab",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-w9rrrgab-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 176.2 GB at this context; duplicate traces of the same session: 0ga6aqdt, avh80cw6, jlb22sku, o01j09ja, tkp5doda, wfw7jo1p; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1884.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 1,884 53.6 176.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w9rrrgab",
   "source_url": "https://omlx.ai/benchmarks/performance/w9rrrgab",
   "tg_tokens": null,
   "tps": 53.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-phvokigm-ctx1024",
   "model": "Qwen3.5-35B-A3B-Claude-4.6-Opus-abliterated",
   "notes": "Huihui build; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: tdl6nyec; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 468.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Huihui-Qwen3.5-35B-A3B-Claude-4.6-Opus-abliterated-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k 468.8 48.3 19.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/phvokigm",
   "source_url": "https://omlx.ai/benchmarks/performance/phvokigm",
   "tg_tokens": null,
   "tps": 48.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-phvokigm-ctx4096",
   "model": "Qwen3.5-35B-A3B-Claude-4.6-Opus-abliterated",
   "notes": "Huihui build; oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: tdl6nyec; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 517.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Huihui-Qwen3.5-35B-A3B-Claude-4.6-Opus-abliterated-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 517.2 46.1 20.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/phvokigm",
   "source_url": "https://omlx.ai/benchmarks/performance/phvokigm",
   "tg_tokens": null,
   "tps": 46.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-7axq4iin-ctx1024",
   "model": "Llama-3-Groq-8B-Tool-Use",
   "notes": "oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: qowkq3l1; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 236.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Llama-3-Groq-8B-Tool-Use-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k 236.4 22.8 4.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7axq4iin",
   "source_url": "https://omlx.ai/benchmarks/performance/7axq4iin",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-7axq4iin-ctx4096",
   "model": "Llama-3-Groq-8B-Tool-Use",
   "notes": "oMLX community benchmark trace; peak memory 5.2 GB at this context; duplicate traces of the same session: qowkq3l1; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 226.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Llama-3-Groq-8B-Tool-Use-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 226.8 20.8 5.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7axq4iin",
   "source_url": "https://omlx.ai/benchmarks/performance/7axq4iin",
   "tg_tokens": null,
   "tps": 20.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-n8br9jr6-ctx1024",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.6 GB at this context; duplicate traces of the same session: eigh1kkd; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 229.5,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-OptiQ-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k 229.5 16.3 7.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n8br9jr6",
   "source_url": "https://omlx.ai/benchmarks/performance/n8br9jr6",
   "tg_tokens": null,
   "tps": 16.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-n8br9jr6-ctx4096",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.1 GB at this context; duplicate traces of the same session: eigh1kkd; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 236.8,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-OptiQ-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 236.8 16.0 8.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n8br9jr6",
   "source_url": "https://omlx.ai/benchmarks/performance/n8br9jr6",
   "tg_tokens": null,
   "tps": 16.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-40nw0we3-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "48GB RAM variant; oMLX community benchmark trace; peak memory 44.9 GB at this context; duplicate traces of the same session: 5yv8fx2x, ex4jjcpu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 229.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B-RAM-48GB-MLX on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 229.1 29.6 44.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/40nw0we3",
   "source_url": "https://omlx.ai/benchmarks/performance/40nw0we3",
   "tg_tokens": null,
   "tps": 29.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-40nw0we3-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "48GB RAM variant; oMLX community benchmark trace; peak memory 49.1 GB at this context; duplicate traces of the same session: 5yv8fx2x, ex4jjcpu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 244.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B-RAM-48GB-MLX on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 244.0 28.9 49.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/40nw0we3",
   "source_url": "https://omlx.ai/benchmarks/performance/40nw0we3",
   "tg_tokens": null,
   "tps": 28.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-40nw0we3-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "48GB RAM variant; oMLX community benchmark trace; peak memory 49.8 GB at this context; duplicate traces of the same session: 5yv8fx2x, ex4jjcpu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 231.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B-RAM-48GB-MLX on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 231.2 28.0 49.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/40nw0we3",
   "source_url": "https://omlx.ai/benchmarks/performance/40nw0we3",
   "tg_tokens": null,
   "tps": 28.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-nmovl4ni-ctx1024",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: xgwjqqfy; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 65.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled-MLX-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k 65.3 6.7 15.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nmovl4ni",
   "source_url": "https://omlx.ai/benchmarks/performance/nmovl4ni",
   "tg_tokens": null,
   "tps": 6.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-nmovl4ni-ctx4096",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: xgwjqqfy; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 66.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled-MLX-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 66.1 6.6 17.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nmovl4ni",
   "source_url": "https://omlx.ai/benchmarks/performance/nmovl4ni",
   "tg_tokens": null,
   "tps": 6.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3b784pbq-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 65.8 GB at this context; duplicate traces of the same session: 4myh21mn, fhh33ohd, msuswddl, o4vnr12w, upqg4z6x, yi5qc2tp; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1574.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.6-35B-A3B-bf16 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 bf16; 1k 1,574 73.9 65.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3b784pbq",
   "source_url": "https://omlx.ai/benchmarks/performance/3b784pbq",
   "tg_tokens": null,
   "tps": 73.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3b784pbq-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 67.5 GB at this context; duplicate traces of the same session: 4myh21mn, fhh33ohd, msuswddl, o4vnr12w, upqg4z6x, yi5qc2tp; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2648.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.6-35B-A3B-bf16 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 bf16; 4k 2,648 72.8 67.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3b784pbq",
   "source_url": "https://omlx.ai/benchmarks/performance/3b784pbq",
   "tg_tokens": null,
   "tps": 72.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3b784pbq-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 67.6 GB at this context; duplicate traces of the same session: 4myh21mn, fhh33ohd, msuswddl, o4vnr12w, upqg4z6x, yi5qc2tp; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2438.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.6-35B-A3B-bf16 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 bf16; 8k 2,438 72.9 67.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3b784pbq",
   "source_url": "https://omlx.ai/benchmarks/performance/3b784pbq",
   "tg_tokens": null,
   "tps": 72.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3b784pbq-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 67.9 GB at this context; duplicate traces of the same session: 4myh21mn, fhh33ohd, msuswddl, o4vnr12w, upqg4z6x, yi5qc2tp; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2259.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.6-35B-A3B-bf16 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 bf16; 16k 2,259 68.3 67.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3b784pbq",
   "source_url": "https://omlx.ai/benchmarks/performance/3b784pbq",
   "tg_tokens": null,
   "tps": 68.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3b784pbq-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 68.5 GB at this context; duplicate traces of the same session: 4myh21mn, fhh33ohd, msuswddl, o4vnr12w, upqg4z6x, yi5qc2tp; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1873.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.6-35B-A3B-bf16 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 bf16; 32k 1,873 63.4 68.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3b784pbq",
   "source_url": "https://omlx.ai/benchmarks/performance/3b784pbq",
   "tg_tokens": null,
   "tps": 63.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3b784pbq-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 69.7 GB at this context; duplicate traces of the same session: 4myh21mn, fhh33ohd, msuswddl, o4vnr12w, upqg4z6x, yi5qc2tp; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1387.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.6-35B-A3B-bf16 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 bf16; 64k 1,387 56.9 69.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3b784pbq",
   "source_url": "https://omlx.ai/benchmarks/performance/3b784pbq",
   "tg_tokens": null,
   "tps": 56.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3b784pbq-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 4myh21mn, fhh33ohd, msuswddl, o4vnr12w, upqg4z6x, yi5qc2tp; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 910.1,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.6-35B-A3B-bf16 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 bf16; 128k 910.1 46.8 71.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3b784pbq",
   "source_url": "https://omlx.ai/benchmarks/performance/3b784pbq",
   "tg_tokens": null,
   "tps": 46.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ffwz4dij-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: gc3m6pco; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1189.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 1k 1,189 32.9 28.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ffwz4dij",
   "source_url": "https://omlx.ai/benchmarks/performance/ffwz4dij",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ffwz4dij-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: gc3m6pco; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1384.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 4k 1,384 32.7 29.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ffwz4dij",
   "source_url": "https://omlx.ai/benchmarks/performance/ffwz4dij",
   "tg_tokens": null,
   "tps": 32.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-nsjpapbx-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: d0zj7c1i; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 404.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 404.5 18.1 16.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nsjpapbx",
   "source_url": "https://omlx.ai/benchmarks/performance/nsjpapbx",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-nsjpapbx-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: d0zj7c1i; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 422.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 422.5 17.8 17.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nsjpapbx",
   "source_url": "https://omlx.ai/benchmarks/performance/nsjpapbx",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3px42fic-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 852.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 852.8 18.4 29.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3px42fic",
   "source_url": "https://omlx.ai/benchmarks/performance/3px42fic",
   "tg_tokens": null,
   "tps": 18.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3px42fic-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 841.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 841.3 17.6 30.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3px42fic",
   "source_url": "https://omlx.ai/benchmarks/performance/3px42fic",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-smmv2zt7-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; bf16; oMLX community benchmark trace, single run at 1k context; peak memory 33.9 GB; TTFT 2597 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 394.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Uncensored-oQ8e-bf16-mtp-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 1k tokens 394.3 PP tok/s 51.4 TG tok/s 2597 TTFT (ms) 33.9 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/smmv2zt7",
   "source_url": "https://omlx.ai/benchmarks/performance/smmv2zt7",
   "tg_tokens": null,
   "tps": 51.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-pwu7feys-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 102.0 GB at this context; duplicate traces of the same session: 79es67ek, 7znt4feb, asb0ybzk, bkvcdz13, uzmrxlr1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 690.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-fp16 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 1k 690.9 50.5 102.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pwu7feys",
   "source_url": "https://omlx.ai/benchmarks/performance/pwu7feys",
   "tg_tokens": null,
   "tps": 50.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-pwu7feys-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 104.4 GB at this context; duplicate traces of the same session: 79es67ek, 7znt4feb, asb0ybzk, bkvcdz13, uzmrxlr1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 861.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-fp16 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 4k 861.8 49.3 104.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pwu7feys",
   "source_url": "https://omlx.ai/benchmarks/performance/pwu7feys",
   "tg_tokens": null,
   "tps": 49.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-pwu7feys-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 104.5 GB at this context; duplicate traces of the same session: 79es67ek, 7znt4feb, asb0ybzk, bkvcdz13, uzmrxlr1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 859.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-fp16 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 8k 859.5 56.4 104.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pwu7feys",
   "source_url": "https://omlx.ai/benchmarks/performance/pwu7feys",
   "tg_tokens": null,
   "tps": 56.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-pwu7feys-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 105.3 GB at this context; duplicate traces of the same session: 79es67ek, 7znt4feb, asb0ybzk, bkvcdz13, uzmrxlr1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 844.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-fp16 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 32k 844.1 39.8 105.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pwu7feys",
   "source_url": "https://omlx.ai/benchmarks/performance/pwu7feys",
   "tg_tokens": null,
   "tps": 39.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-pwu7feys-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 106.5 GB at this context; duplicate traces of the same session: 79es67ek, 7znt4feb, asb0ybzk, bkvcdz13, uzmrxlr1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 829.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-fp16 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 64k 829.6 44.8 106.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pwu7feys",
   "source_url": "https://omlx.ai/benchmarks/performance/pwu7feys",
   "tg_tokens": null,
   "tps": 44.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-pwu7feys-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 109.3 GB at this context; duplicate traces of the same session: 79es67ek, 7znt4feb, asb0ybzk, bkvcdz13, uzmrxlr1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 752.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-fp16 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 128k 752.7 38.2 109.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pwu7feys",
   "source_url": "https://omlx.ai/benchmarks/performance/pwu7feys",
   "tg_tokens": null,
   "tps": 38.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-frqf6077-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 73.9 GB at this context; duplicate traces of the same session: fk5ggoc2, jfumy2jx, t1ppxls5, v97uaz02; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 399.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-fp16 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 1k 399.1 22.2 73.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/frqf6077",
   "source_url": "https://omlx.ai/benchmarks/performance/frqf6077",
   "tg_tokens": null,
   "tps": 22.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-frqf6077-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 81.1 GB at this context; duplicate traces of the same session: fk5ggoc2, jfumy2jx, t1ppxls5, v97uaz02; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 498.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-fp16 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 4k 498.5 21.9 81.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/frqf6077",
   "source_url": "https://omlx.ai/benchmarks/performance/frqf6077",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-frqf6077-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 84.1 GB at this context; duplicate traces of the same session: fk5ggoc2, jfumy2jx, t1ppxls5, v97uaz02; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 550.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-fp16 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 8k 550.4 21.4 84.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/frqf6077",
   "source_url": "https://omlx.ai/benchmarks/performance/frqf6077",
   "tg_tokens": null,
   "tps": 21.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-frqf6077-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 84.8 GB at this context; duplicate traces of the same session: fk5ggoc2, jfumy2jx, t1ppxls5, v97uaz02; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 536.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-fp16 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 16k 536.3 19.3 84.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/frqf6077",
   "source_url": "https://omlx.ai/benchmarks/performance/frqf6077",
   "tg_tokens": null,
   "tps": 19.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-frqf6077-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 90.3 GB at this context; duplicate traces of the same session: fk5ggoc2, jfumy2jx, t1ppxls5, v97uaz02; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 517.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-fp16 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 32k 517.6 18.6 90.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/frqf6077",
   "source_url": "https://omlx.ai/benchmarks/performance/frqf6077",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-frqf6077-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 92.2 GB at this context; duplicate traces of the same session: fk5ggoc2, jfumy2jx, t1ppxls5, v97uaz02; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 417.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-fp16 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 64k 417.5 25.2 92.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/frqf6077",
   "source_url": "https://omlx.ai/benchmarks/performance/frqf6077",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-frqf6077-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 97.0 GB at this context; duplicate traces of the same session: fk5ggoc2, jfumy2jx, t1ppxls5, v97uaz02; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 466.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-fp16 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 128k 466.7 22.4 97.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/frqf6077",
   "source_url": "https://omlx.ai/benchmarks/performance/frqf6077",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-0eyx83ep-ctx1024",
   "model": "Qwopus3.8-27B-MTPLX",
   "notes": "Quality FP16 build; oMLX community benchmark trace; peak memory 34.2 GB at this context; duplicate traces of the same session: tsxvs5ll; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 215.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwopus3.8-27B-MTPLX-8bit-Quality-FP16 on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 8bit; 1k 215.3 20.0 34.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0eyx83ep",
   "source_url": "https://omlx.ai/benchmarks/performance/0eyx83ep",
   "tg_tokens": null,
   "tps": 20.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-0eyx83ep-ctx4096",
   "model": "Qwopus3.8-27B-MTPLX",
   "notes": "Quality FP16 build; oMLX community benchmark trace; peak memory 37.9 GB at this context; duplicate traces of the same session: tsxvs5ll; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 228.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwopus3.8-27B-MTPLX-8bit-Quality-FP16 on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 8bit; 4k 228.8 32.5 37.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0eyx83ep",
   "source_url": "https://omlx.ai/benchmarks/performance/0eyx83ep",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-4khief5l-ctx1024",
   "model": "clef",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: 5iwxgbiv; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 65.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "clef-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k 65.6 6.8 15.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4khief5l",
   "source_url": "https://omlx.ai/benchmarks/performance/4khief5l",
   "tg_tokens": null,
   "tps": 6.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-4khief5l-ctx4096",
   "model": "clef",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: 5iwxgbiv; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 67.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "clef-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 67.6 6.7 17.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4khief5l",
   "source_url": "https://omlx.ai/benchmarks/performance/4khief5l",
   "tg_tokens": null,
   "tps": 6.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-lvhdln3d-ctx1024",
   "model": "clef-flash",
   "notes": "generic trace name; oMLX community benchmark trace; peak memory 5.7 GB at this context; duplicate traces of the same session: 3slgstwr; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 228.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "clef-flash-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k 228.0 20.3 5.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lvhdln3d",
   "source_url": "https://omlx.ai/benchmarks/performance/lvhdln3d",
   "tg_tokens": null,
   "tps": 20.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-lvhdln3d-ctx4096",
   "model": "clef-flash",
   "notes": "generic trace name; oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: 3slgstwr; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 240.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "clef-flash-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 240.4 20.9 6.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lvhdln3d",
   "source_url": "https://omlx.ai/benchmarks/performance/lvhdln3d",
   "tg_tokens": null,
   "tps": 20.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-kdp3a8zj-ctx1024",
   "model": "mlx-4bit",
   "notes": "generic trace name; model identity from the session title; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 35ji16r4; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 685.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 685.6 107.7 19.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kdp3a8zj",
   "source_url": "https://omlx.ai/benchmarks/performance/kdp3a8zj",
   "tg_tokens": null,
   "tps": 107.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-06",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-kdp3a8zj-ctx4096",
   "model": "mlx-4bit",
   "notes": "generic trace name; model identity from the session title; oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 35ji16r4; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1622.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 1,622 103.9 20.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kdp3a8zj",
   "source_url": "https://omlx.ai/benchmarks/performance/kdp3a8zj",
   "tg_tokens": null,
   "tps": 103.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-vmayg127-ctx1024",
   "model": "DeepSeek-V4-Flash",
   "notes": "oMLX community benchmark trace; peak memory 142.5 GB at this context; duplicate traces of the same session: a44f3svn; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1204.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-V4-Flash-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,204 46.1 142.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vmayg127",
   "source_url": "https://omlx.ai/benchmarks/performance/vmayg127",
   "tg_tokens": null,
   "tps": 46.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-vmayg127-ctx4096",
   "model": "DeepSeek-V4-Flash",
   "notes": "oMLX community benchmark trace; peak memory 142.9 GB at this context; duplicate traces of the same session: a44f3svn; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1474.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-V4-Flash-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,474 41.1 142.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vmayg127",
   "source_url": "https://omlx.ai/benchmarks/performance/vmayg127",
   "tg_tokens": null,
   "tps": 41.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-odrydk5q-ctx1024",
   "model": "DeepSeek-V4-Flash",
   "notes": "oMLX community benchmark trace; peak memory 142.5 GB at this context; duplicate traces of the same session: 62q22zzx; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 437.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-V4-Flash-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 437.2 45.2 142.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/odrydk5q",
   "source_url": "https://omlx.ai/benchmarks/performance/odrydk5q",
   "tg_tokens": null,
   "tps": 45.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-odrydk5q-ctx4096",
   "model": "DeepSeek-V4-Flash",
   "notes": "oMLX community benchmark trace; peak memory 142.9 GB at this context; duplicate traces of the same session: 62q22zzx; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 924.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-V4-Flash-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 924.4 39.0 142.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/odrydk5q",
   "source_url": "https://omlx.ai/benchmarks/performance/odrydk5q",
   "tg_tokens": null,
   "tps": 39.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-lfe86sp0-ctx4096",
   "model": "Devstral-Small-2-24B-Instruct-2512",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 20.1 GB; TTFT 16905 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 242.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Devstral-Small-2-24B-Instruct-2512-6bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k tokens 242.3 PP tok/s 19.9 TG tok/s 16905 TTFT (ms) 20.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lfe86sp0",
   "source_url": "https://omlx.ai/benchmarks/performance/lfe86sp0",
   "tg_tokens": null,
   "tps": 19.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-e9nl7e7q-ctx1024",
   "model": "FrogNano-4B-2609",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 5.2 GB at this context; duplicate traces of the same session: 80n8239l; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 412.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "FrogNano-4B-2609-oQ8e-mtp-MLX on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 8bit; 1k 412.4 21.4 5.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e9nl7e7q",
   "source_url": "https://omlx.ai/benchmarks/performance/e9nl7e7q",
   "tg_tokens": null,
   "tps": 21.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-e9nl7e7q-ctx4096",
   "model": "FrogNano-4B-2609",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 5.7 GB at this context; duplicate traces of the same session: 80n8239l; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 423.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "FrogNano-4B-2609-oQ8e-mtp-MLX on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 8bit; 4k 423.8 21.3 5.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e9nl7e7q",
   "source_url": "https://omlx.ai/benchmarks/performance/e9nl7e7q",
   "tg_tokens": null,
   "tps": 21.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-u00175ie-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 173.9 GB; TTFT 9272 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 441.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k tokens 441.8 PP tok/s 40.4 TG tok/s 9272 TTFT (ms) 173.9 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u00175ie",
   "source_url": "https://omlx.ai/benchmarks/performance/u00175ie",
   "tg_tokens": null,
   "tps": 40.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-g4gbtfbd-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 173.9 GB; TTFT 9105 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 449.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k tokens 449.9 PP tok/s 32.4 TG tok/s 9105 TTFT (ms) 173.9 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g4gbtfbd",
   "source_url": "https://omlx.ai/benchmarks/performance/g4gbtfbd",
   "tg_tokens": null,
   "tps": 32.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-g5jprlmm-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 173.9 GB; TTFT 9275 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 441.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k tokens 441.6 PP tok/s 40.2 TG tok/s 9275 TTFT (ms) 173.9 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g5jprlmm",
   "source_url": "https://omlx.ai/benchmarks/performance/g5jprlmm",
   "tg_tokens": null,
   "tps": 40.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-cqznxq3c-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 173.9 GB; TTFT 9287 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 441.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k tokens 441.1 PP tok/s 38.7 TG tok/s 9287 TTFT (ms) 173.9 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cqznxq3c",
   "source_url": "https://omlx.ai/benchmarks/performance/cqznxq3c",
   "tg_tokens": null,
   "tps": 38.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-yc8o6w11-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 173.9 GB; TTFT 9282 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 441.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k tokens 441.3 PP tok/s 34.0 TG tok/s 9282 TTFT (ms) 173.9 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yc8o6w11",
   "source_url": "https://omlx.ai/benchmarks/performance/yc8o6w11",
   "tg_tokens": null,
   "tps": 34.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-7qh400mt-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 173.9 GB; TTFT 9276 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 441.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k tokens 441.6 PP tok/s 41.6 TG tok/s 9276 TTFT (ms) 173.9 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7qh400mt",
   "source_url": "https://omlx.ai/benchmarks/performance/7qh400mt",
   "tg_tokens": null,
   "tps": 41.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ha3k4qdk-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 173.9 GB; TTFT 9209 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 444.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k tokens 444.8 PP tok/s 32.3 TG tok/s 9209 TTFT (ms) 173.9 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ha3k4qdk",
   "source_url": "https://omlx.ai/benchmarks/performance/ha3k4qdk",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-v30l8x1f-ctx1024",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.8 GB at this context; duplicate traces of the same session: bzyhh5rk; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 172.1,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-OptiQ-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 1k 172.1 25.3 7.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v30l8x1f",
   "source_url": "https://omlx.ai/benchmarks/performance/v30l8x1f",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-v30l8x1f-ctx4096",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.3 GB at this context; duplicate traces of the same session: bzyhh5rk; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 170.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-OptiQ-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 4k 170.0 24.6 8.3 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v30l8x1f",
   "source_url": "https://omlx.ai/benchmarks/performance/v30l8x1f",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mpine6dp-ctx1024",
   "model": "Mistral-Small-4-119B-2603",
   "notes": "oMLX community benchmark trace; peak memory 63.0 GB at this context; duplicate traces of the same session: 15vbdu5s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1507.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Mistral-Small-4-119B-2603-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,507 86.8 63.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mpine6dp",
   "source_url": "https://omlx.ai/benchmarks/performance/mpine6dp",
   "tg_tokens": null,
   "tps": 86.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mpine6dp-ctx4096",
   "model": "Mistral-Small-4-119B-2603",
   "notes": "oMLX community benchmark trace; peak memory 63.3 GB at this context; duplicate traces of the same session: 15vbdu5s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2556.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Mistral-Small-4-119B-2603-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,556 78.7 63.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mpine6dp",
   "source_url": "https://omlx.ai/benchmarks/performance/mpine6dp",
   "tg_tokens": null,
   "tps": 78.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-b8stczyc-ctx4096",
   "model": "Qwen3.6-14B-A3B-FableVibes",
   "notes": "oMLX community benchmark trace; peak memory 9.4 GB at this context; duplicate traces of the same session: 0054wcgo, uoxhtzvz; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 429.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-14B-A3B-FableVibes-oQ4e on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 429.4 34.7 9.4 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b8stczyc",
   "source_url": "https://omlx.ai/benchmarks/performance/b8stczyc",
   "tg_tokens": null,
   "tps": 34.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-b8stczyc-ctx8192",
   "model": "Qwen3.6-14B-A3B-FableVibes",
   "notes": "oMLX community benchmark trace; peak memory 9.5 GB at this context; duplicate traces of the same session: 0054wcgo, uoxhtzvz; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 425.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-14B-A3B-FableVibes-oQ4e on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 8k 425.3 33.6 9.5 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b8stczyc",
   "source_url": "https://omlx.ai/benchmarks/performance/b8stczyc",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-b8stczyc-ctx16384",
   "model": "Qwen3.6-14B-A3B-FableVibes",
   "notes": "oMLX community benchmark trace; peak memory 9.8 GB at this context; duplicate traces of the same session: 0054wcgo, uoxhtzvz; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 355.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-14B-A3B-FableVibes-oQ4e on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 16k 355.5 26.7 9.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b8stczyc",
   "source_url": "https://omlx.ai/benchmarks/performance/b8stczyc",
   "tg_tokens": null,
   "tps": 26.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-c80yyzok-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "DWQ build; oMLX community benchmark trace, single run at 1k context; peak memory 20.3 GB; TTFT 624 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1642.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit-DWQ on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 8bit; 1k tokens 1,642 PP tok/s 49.6 TG tok/s 624 TTFT (ms) 20.3 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c80yyzok",
   "source_url": "https://omlx.ai/benchmarks/performance/c80yyzok",
   "tg_tokens": null,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-lb69s1wv-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "DWQ build; oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: rqbkjuls; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1592.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit-DWQ on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 8bit; 1k 1,592 50.8 20.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lb69s1wv",
   "source_url": "https://omlx.ai/benchmarks/performance/lb69s1wv",
   "tg_tokens": null,
   "tps": 50.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-lb69s1wv-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "DWQ build; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: rqbkjuls; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1992.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit-DWQ on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 8bit; 4k 1,992 49.6 20.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lb69s1wv",
   "source_url": "https://omlx.ai/benchmarks/performance/lb69s1wv",
   "tg_tokens": null,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-m660odul-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: aoluvdap, t5rg8h3g, v924mric; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 773.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 773.7 19.1 28.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m660odul",
   "source_url": "https://omlx.ai/benchmarks/performance/m660odul",
   "tg_tokens": null,
   "tps": 19.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-m660odul-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.8 GB at this context; duplicate traces of the same session: aoluvdap, t5rg8h3g, v924mric; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 935.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 935.8 18.2 29.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m660odul",
   "source_url": "https://omlx.ai/benchmarks/performance/m660odul",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-m660odul-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.3 GB at this context; duplicate traces of the same session: aoluvdap, t5rg8h3g, v924mric; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 880.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 880.9 17.6 31.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m660odul",
   "source_url": "https://omlx.ai/benchmarks/performance/m660odul",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-m660odul-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: aoluvdap, t5rg8h3g, v924mric; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 632.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 632.3 14.5 37.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m660odul",
   "source_url": "https://omlx.ai/benchmarks/performance/m660odul",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-18q4b5wx-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 852.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 852.8 18.4 29.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/18q4b5wx",
   "source_url": "https://omlx.ai/benchmarks/performance/18q4b5wx",
   "tg_tokens": null,
   "tps": 18.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-18q4b5wx-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 841.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 841.3 17.6 30.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/18q4b5wx",
   "source_url": "https://omlx.ai/benchmarks/performance/18q4b5wx",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-fqtj006x-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized speed FP16 build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: cz4gzozw; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 109.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed-FP16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 109.9 12.5 23.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fqtj006x",
   "source_url": "https://omlx.ai/benchmarks/performance/fqtj006x",
   "tg_tokens": null,
   "tps": 12.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-fqtj006x-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized speed FP16 build; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: cz4gzozw; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 99.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed-FP16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 99.3 13.4 27.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fqtj006x",
   "source_url": "https://omlx.ai/benchmarks/performance/fqtj006x",
   "tg_tokens": null,
   "tps": 13.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-r24x5sfc-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Speed build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: t6qh5r4l; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 65.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Speed on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 65.7 12.4 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r24x5sfc",
   "source_url": "https://omlx.ai/benchmarks/performance/r24x5sfc",
   "tg_tokens": null,
   "tps": 12.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-r24x5sfc-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Speed build; oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: t6qh5r4l; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 63.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Speed on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 63.4 9.8 27.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r24x5sfc",
   "source_url": "https://omlx.ai/benchmarks/performance/r24x5sfc",
   "tg_tokens": null,
   "tps": 9.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-pxt2l5a9-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: nk3feznb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 474.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 474.9 17.6 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pxt2l5a9",
   "source_url": "https://omlx.ai/benchmarks/performance/pxt2l5a9",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-pxt2l5a9-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: nk3feznb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 459.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 459.3 17.3 17.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pxt2l5a9",
   "source_url": "https://omlx.ai/benchmarks/performance/pxt2l5a9",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-q28o024r-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: l9ujch9f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 266.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 266.1 43.8 20.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q28o024r",
   "source_url": "https://omlx.ai/benchmarks/performance/q28o024r",
   "tg_tokens": null,
   "tps": 43.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-q28o024r-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: l9ujch9f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 228.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 228.8 30.4 21.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q28o024r",
   "source_url": "https://omlx.ai/benchmarks/performance/q28o024r",
   "tg_tokens": null,
   "tps": 30.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-9dzybrzf-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 120.5 GB at this context; duplicate traces of the same session: l1p7azlb; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 635.5,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 5bit; 1k 635.5 113.1 120.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9dzybrzf",
   "source_url": "https://omlx.ai/benchmarks/performance/9dzybrzf",
   "tg_tokens": null,
   "tps": 113.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-9dzybrzf-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 121.0 GB at this context; duplicate traces of the same session: l1p7azlb; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1629.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 5bit; 4k 1,629 56.7 121.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9dzybrzf",
   "source_url": "https://omlx.ai/benchmarks/performance/9dzybrzf",
   "tg_tokens": null,
   "tps": 56.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-c9wywg3d-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 73.9 GB at this context; duplicate traces of the same session: fqap5zh7; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 399.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-fp16 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 1k 399.1 22.2 73.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c9wywg3d",
   "source_url": "https://omlx.ai/benchmarks/performance/c9wywg3d",
   "tg_tokens": null,
   "tps": 22.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-c9wywg3d-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 81.1 GB at this context; duplicate traces of the same session: fqap5zh7; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 498.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-fp16 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 4k 498.5 21.9 81.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c9wywg3d",
   "source_url": "https://omlx.ai/benchmarks/performance/c9wywg3d",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-c9wywg3d-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 84.1 GB at this context; duplicate traces of the same session: fqap5zh7; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 550.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-fp16 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 8k 550.4 21.4 84.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c9wywg3d",
   "source_url": "https://omlx.ai/benchmarks/performance/c9wywg3d",
   "tg_tokens": null,
   "tps": 21.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-c9wywg3d-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 84.8 GB at this context; duplicate traces of the same session: fqap5zh7; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 536.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-fp16 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 16k 536.3 19.3 84.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c9wywg3d",
   "source_url": "https://omlx.ai/benchmarks/performance/c9wywg3d",
   "tg_tokens": null,
   "tps": 19.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-c9wywg3d-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 90.3 GB at this context; duplicate traces of the same session: fqap5zh7; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 517.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-fp16 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 32k 517.6 18.6 90.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c9wywg3d",
   "source_url": "https://omlx.ai/benchmarks/performance/c9wywg3d",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-c9wywg3d-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 92.2 GB at this context; duplicate traces of the same session: fqap5zh7; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 417.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-fp16 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 64k 417.5 25.2 92.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c9wywg3d",
   "source_url": "https://omlx.ai/benchmarks/performance/c9wywg3d",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-c9wywg3d-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 97.0 GB at this context; duplicate traces of the same session: fqap5zh7; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 466.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-fp16 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 128k 466.7 22.4 97.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c9wywg3d",
   "source_url": "https://omlx.ai/benchmarks/performance/c9wywg3d",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-d5q2dsp0-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 84.7 GB; TTFT 1081 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 947.6,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k tokens 947.6 PP tok/s 68.3 TG tok/s 1081 TTFT (ms) 84.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d5q2dsp0",
   "source_url": "https://omlx.ai/benchmarks/performance/d5q2dsp0",
   "tg_tokens": null,
   "tps": 68.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-cb041lgd-ctx4096",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: ic3vgf1z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 913.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 913.8 66.6 20.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cb041lgd",
   "source_url": "https://omlx.ai/benchmarks/performance/cb041lgd",
   "tg_tokens": null,
   "tps": 66.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-cb041lgd-ctx16384",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: ic3vgf1z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 949.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 949.9 77.9 20.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cb041lgd",
   "source_url": "https://omlx.ai/benchmarks/performance/cb041lgd",
   "tg_tokens": null,
   "tps": 77.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tmmimi5f-ctx4096",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: 9cvahv5q, ky5d60px, tcywc70y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1001.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,001 58.8 18.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tmmimi5f",
   "source_url": "https://omlx.ai/benchmarks/performance/tmmimi5f",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-tmmimi5f-ctx16384",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 9cvahv5q, ky5d60px, tcywc70y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 969.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 969.6 78.7 19.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tmmimi5f",
   "source_url": "https://omlx.ai/benchmarks/performance/tmmimi5f",
   "tg_tokens": null,
   "tps": 78.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-tmmimi5f-ctx65536",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: 9cvahv5q, ky5d60px, tcywc70y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 790.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 790.4 55.8 23.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tmmimi5f",
   "source_url": "https://omlx.ai/benchmarks/performance/tmmimi5f",
   "tg_tokens": null,
   "tps": 55.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-tmmimi5f-ctx131072",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: 9cvahv5q, ky5d60px, tcywc70y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 627.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 627.0 50.5 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tmmimi5f",
   "source_url": "https://omlx.ai/benchmarks/performance/tmmimi5f",
   "tg_tokens": null,
   "tps": 50.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jw7gp28j-ctx4096",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: wet2qlzp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 998.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 998.7 68.0 18.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jw7gp28j",
   "source_url": "https://omlx.ai/benchmarks/performance/jw7gp28j",
   "tg_tokens": null,
   "tps": 68.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jw7gp28j-ctx16384",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: wet2qlzp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 969.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 969.7 81.2 19.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jw7gp28j",
   "source_url": "https://omlx.ai/benchmarks/performance/jw7gp28j",
   "tg_tokens": null,
   "tps": 81.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-1cj99mn6-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 20.3 GB; TTFT 2852 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1436.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "gemma-4-26b-a4b-it-6bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k tokens 1,436 PP tok/s 90.5 TG tok/s 2852 TTFT (ms) 20.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1cj99mn6",
   "source_url": "https://omlx.ai/benchmarks/performance/1cj99mn6",
   "tg_tokens": null,
   "tps": 90.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-7hu1upm3-ctx1024",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: 3c5s6vdb, 7jbuvy8d, d6qjk8ok, z6bvjrqt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1522.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,522 60.6 177.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7hu1upm3",
   "source_url": "https://omlx.ai/benchmarks/performance/7hu1upm3",
   "tg_tokens": null,
   "tps": 60.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-7hu1upm3-ctx4096",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: 3c5s6vdb, 7jbuvy8d, d6qjk8ok, z6bvjrqt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2248.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 2,248 60.1 178.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7hu1upm3",
   "source_url": "https://omlx.ai/benchmarks/performance/7hu1upm3",
   "tg_tokens": null,
   "tps": 60.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-7hu1upm3-ctx8192",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 178.9 GB at this context; duplicate traces of the same session: 3c5s6vdb, 7jbuvy8d, d6qjk8ok, z6bvjrqt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2308.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 2,308 79.3 178.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7hu1upm3",
   "source_url": "https://omlx.ai/benchmarks/performance/7hu1upm3",
   "tg_tokens": null,
   "tps": 79.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-7hu1upm3-ctx16384",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 179.1 GB at this context; duplicate traces of the same session: 3c5s6vdb, 7jbuvy8d, d6qjk8ok, z6bvjrqt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2311.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 2,311 83.5 179.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7hu1upm3",
   "source_url": "https://omlx.ai/benchmarks/performance/7hu1upm3",
   "tg_tokens": null,
   "tps": 83.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-7hu1upm3-ctx32768",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 179.4 GB at this context; duplicate traces of the same session: 3c5s6vdb, 7jbuvy8d, d6qjk8ok, z6bvjrqt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2300.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,300 83.4 179.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7hu1upm3",
   "source_url": "https://omlx.ai/benchmarks/performance/7hu1upm3",
   "tg_tokens": null,
   "tps": 83.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-69l3za1n-ctx1024",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 178.0 GB at this context; duplicate traces of the same session: 00g4sj3g, 783pkjpd, d7t0gbue, g7j9pjx2, rehmols9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1817.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,817 52.4 178.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/69l3za1n",
   "source_url": "https://omlx.ai/benchmarks/performance/69l3za1n",
   "tg_tokens": null,
   "tps": 52.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-69l3za1n-ctx4096",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 179.3 GB at this context; duplicate traces of the same session: 00g4sj3g, 783pkjpd, d7t0gbue, g7j9pjx2, rehmols9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2366.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 2,366 96.7 179.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/69l3za1n",
   "source_url": "https://omlx.ai/benchmarks/performance/69l3za1n",
   "tg_tokens": null,
   "tps": 96.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-69l3za1n-ctx8192",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 180.3 GB at this context; duplicate traces of the same session: 00g4sj3g, 783pkjpd, d7t0gbue, g7j9pjx2, rehmols9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2384.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 2,384 54.9 180.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/69l3za1n",
   "source_url": "https://omlx.ai/benchmarks/performance/69l3za1n",
   "tg_tokens": null,
   "tps": 54.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-69l3za1n-ctx16384",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 181.0 GB at this context; duplicate traces of the same session: 00g4sj3g, 783pkjpd, d7t0gbue, g7j9pjx2, rehmols9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2367.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 2,367 48.9 181.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/69l3za1n",
   "source_url": "https://omlx.ai/benchmarks/performance/69l3za1n",
   "tg_tokens": null,
   "tps": 48.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-69l3za1n-ctx32768",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 182.4 GB at this context; duplicate traces of the same session: 00g4sj3g, 783pkjpd, d7t0gbue, g7j9pjx2, rehmols9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2354.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,354 33.6 182.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/69l3za1n",
   "source_url": "https://omlx.ai/benchmarks/performance/69l3za1n",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-69l3za1n-ctx65536",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 185.3 GB at this context; duplicate traces of the same session: 00g4sj3g, 783pkjpd, d7t0gbue, g7j9pjx2, rehmols9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2317.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,317 81.6 185.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/69l3za1n",
   "source_url": "https://omlx.ai/benchmarks/performance/69l3za1n",
   "tg_tokens": null,
   "tps": 81.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-0eur30ns-ctx1024",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 178.0 GB at this context; duplicate traces of the same session: 2kjf06p1, 7kupong4, t1k85jqp, xwcoxa88, yi3tavaa; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 751.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 751.1 38.7 178.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0eur30ns",
   "source_url": "https://omlx.ai/benchmarks/performance/0eur30ns",
   "tg_tokens": null,
   "tps": 38.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-0eur30ns-ctx4096",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 179.3 GB at this context; duplicate traces of the same session: 2kjf06p1, 7kupong4, t1k85jqp, xwcoxa88, yi3tavaa; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1406.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 1,406 43.1 179.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0eur30ns",
   "source_url": "https://omlx.ai/benchmarks/performance/0eur30ns",
   "tg_tokens": null,
   "tps": 43.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-0eur30ns-ctx8192",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 180.3 GB at this context; duplicate traces of the same session: 2kjf06p1, 7kupong4, t1k85jqp, xwcoxa88, yi3tavaa; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2382.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 2,382 89.4 180.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0eur30ns",
   "source_url": "https://omlx.ai/benchmarks/performance/0eur30ns",
   "tg_tokens": null,
   "tps": 89.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-0eur30ns-ctx16384",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 181.0 GB at this context; duplicate traces of the same session: 2kjf06p1, 7kupong4, t1k85jqp, xwcoxa88, yi3tavaa; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2372.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 2,372 98.3 181.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0eur30ns",
   "source_url": "https://omlx.ai/benchmarks/performance/0eur30ns",
   "tg_tokens": null,
   "tps": 98.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-0eur30ns-ctx32768",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 182.4 GB at this context; duplicate traces of the same session: 2kjf06p1, 7kupong4, t1k85jqp, xwcoxa88, yi3tavaa; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2353.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,353 96.8 182.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0eur30ns",
   "source_url": "https://omlx.ai/benchmarks/performance/0eur30ns",
   "tg_tokens": null,
   "tps": 96.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-0eur30ns-ctx65536",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 185.3 GB at this context; duplicate traces of the same session: 2kjf06p1, 7kupong4, t1k85jqp, xwcoxa88, yi3tavaa; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2319.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,319 55.7 185.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0eur30ns",
   "source_url": "https://omlx.ai/benchmarks/performance/0eur30ns",
   "tg_tokens": null,
   "tps": 55.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-9iibldmb-ctx1024",
   "model": "Qwen3-4B-Instruct-2507",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 2.9 GB; TTFT 1858 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 551.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-4B-Instruct-2507-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k tokens 551.2 PP tok/s 76.5 TG tok/s 1858 TTFT (ms) 2.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9iibldmb",
   "source_url": "https://omlx.ai/benchmarks/performance/9iibldmb",
   "tg_tokens": null,
   "tps": 76.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-wz6kz7ee-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1366.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 1,366 44.8 4.2 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wz6kz7ee",
   "source_url": "https://omlx.ai/benchmarks/performance/wz6kz7ee",
   "tg_tokens": null,
   "tps": 44.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-wz6kz7ee-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.5 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1383.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 16k 1,383 37.3 4.5 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wz6kz7ee",
   "source_url": "https://omlx.ai/benchmarks/performance/wz6kz7ee",
   "tg_tokens": null,
   "tps": 37.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-91d0ingc-ctx1024",
   "model": "Qwen3.5-VL-9B",
   "notes": "CRACK build; oMLX community benchmark trace; peak memory 5.7 GB at this context; duplicate traces of the same session: x7kxpeg2; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 725.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-VL-9B-4bitCRACK on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k 725.8 26.5 5.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/91d0ingc",
   "source_url": "https://omlx.ai/benchmarks/performance/91d0ingc",
   "tg_tokens": null,
   "tps": 26.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-91d0ingc-ctx4096",
   "model": "Qwen3.5-VL-9B",
   "notes": "CRACK build; oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: x7kxpeg2; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 728.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-VL-9B-4bitCRACK on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 728.2 25.7 6.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/91d0ingc",
   "source_url": "https://omlx.ai/benchmarks/performance/91d0ingc",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-qtdnn054-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: lhb8nh34; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 968.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k 968.8 57.0 19.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qtdnn054",
   "source_url": "https://omlx.ai/benchmarks/performance/qtdnn054",
   "tg_tokens": null,
   "tps": 57.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-qtdnn054-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: lhb8nh34; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 754.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 754.4 54.6 20.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qtdnn054",
   "source_url": "https://omlx.ai/benchmarks/performance/qtdnn054",
   "tg_tokens": null,
   "tps": 54.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-z4premv0-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 27.9 GB; TTFT 1497 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 684.2,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-6bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 6bit; 1k tokens 684.2 PP tok/s 73.9 TG tok/s 1497 TTFT (ms) 27.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z4premv0",
   "source_url": "https://omlx.ai/benchmarks/performance/z4premv0",
   "tg_tokens": null,
   "tps": 73.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-z6rt48ni-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 21.9 GB; TTFT 666 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1538.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 1,538 PP tok/s 90.8 TG tok/s 666 TTFT (ms) 21.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z6rt48ni",
   "source_url": "https://omlx.ai/benchmarks/performance/z6rt48ni",
   "tg_tokens": null,
   "tps": 90.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-kybmgby8-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4419.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,419 222.3 21.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kybmgby8",
   "source_url": "https://omlx.ai/benchmarks/performance/kybmgby8",
   "tg_tokens": null,
   "tps": 222.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kybmgby8-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4954.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 4,954 164.1 22.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kybmgby8",
   "source_url": "https://omlx.ai/benchmarks/performance/kybmgby8",
   "tg_tokens": null,
   "tps": 164.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-h3os8ygs-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 20.1 GB; TTFT 663 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1544.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 1,544 PP tok/s 102.0 TG tok/s 663 TTFT (ms) 20.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h3os8ygs",
   "source_url": "https://omlx.ai/benchmarks/performance/h3os8ygs",
   "tg_tokens": null,
   "tps": 102.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-t2hmjc9y-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 21.4 GB; TTFT 694 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1475.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 1,475 PP tok/s 90.0 TG tok/s 694 TTFT (ms) 21.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t2hmjc9y",
   "source_url": "https://omlx.ai/benchmarks/performance/t2hmjc9y",
   "tg_tokens": null,
   "tps": 90.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-qmn1gr23-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 21.4 GB; TTFT 692 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1479.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 1,479 PP tok/s 145.6 TG tok/s 692 TTFT (ms) 21.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qmn1gr23",
   "source_url": "https://omlx.ai/benchmarks/performance/qmn1gr23",
   "tg_tokens": null,
   "tps": 145.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-z2nz9x6m-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: p4rhqey9; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 463.6,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 463.6 14.3 20.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z2nz9x6m",
   "source_url": "https://omlx.ai/benchmarks/performance/z2nz9x6m",
   "tg_tokens": null,
   "tps": 14.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-z2nz9x6m-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: p4rhqey9; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 492.6,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 492.6 14.2 21.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z2nz9x6m",
   "source_url": "https://omlx.ai/benchmarks/performance/z2nz9x6m",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8jnpg0hh-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 996.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 996.6 66.9 17.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8jnpg0hh",
   "source_url": "https://omlx.ai/benchmarks/performance/8jnpg0hh",
   "tg_tokens": null,
   "tps": 66.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8jnpg0hh-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1076.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,076 53.0 19.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8jnpg0hh",
   "source_url": "https://omlx.ai/benchmarks/performance/8jnpg0hh",
   "tg_tokens": null,
   "tps": 53.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-cdhd0mpu-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 19.6 GB; TTFT 37253 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 439.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k tokens 439.8 PP tok/s 28.8 TG tok/s 37253 TTFT (ms) 19.6 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cdhd0mpu",
   "source_url": "https://omlx.ai/benchmarks/performance/cdhd0mpu",
   "tg_tokens": null,
   "tps": 28.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-r3j29jsw-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 19.9 GB; TTFT 28390 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 577.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k tokens 577.1 PP tok/s 24.7 TG tok/s 28390 TTFT (ms) 19.9 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r3j29jsw",
   "source_url": "https://omlx.ai/benchmarks/performance/r3j29jsw",
   "tg_tokens": null,
   "tps": 24.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-xx2j3ade-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 49.8 GB; TTFT 52754 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 77.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k tokens 77.6 PP tok/s 4.8 TG tok/s 52754 TTFT (ms) 49.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xx2j3ade",
   "source_url": "https://omlx.ai/benchmarks/performance/xx2j3ade",
   "tg_tokens": null,
   "tps": 4.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pfo3f7to-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: 2xcf5l4v, a9axmzo2, cecfgyaj, hq9p9znv, umsc49qm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1251.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,251 77.4 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pfo3f7to",
   "source_url": "https://omlx.ai/benchmarks/performance/pfo3f7to",
   "tg_tokens": null,
   "tps": 77.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pfo3f7to-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: 2xcf5l4v, a9axmzo2, cecfgyaj, hq9p9znv, umsc49qm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2220.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,220 67.5 71.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pfo3f7to",
   "source_url": "https://omlx.ai/benchmarks/performance/pfo3f7to",
   "tg_tokens": null,
   "tps": 67.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-pfo3f7to-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 2xcf5l4v, a9axmzo2, cecfgyaj, hq9p9znv, umsc49qm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2619.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,619 89.1 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pfo3f7to",
   "source_url": "https://omlx.ai/benchmarks/performance/pfo3f7to",
   "tg_tokens": null,
   "tps": 89.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pfo3f7to-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.9 GB at this context; duplicate traces of the same session: 2xcf5l4v, a9axmzo2, cecfgyaj, hq9p9znv, umsc49qm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2809.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,809 61.9 72.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pfo3f7to",
   "source_url": "https://omlx.ai/benchmarks/performance/pfo3f7to",
   "tg_tokens": null,
   "tps": 61.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pfo3f7to-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.7 GB at this context; duplicate traces of the same session: 2xcf5l4v, a9axmzo2, cecfgyaj, hq9p9znv, umsc49qm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2780.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,780 53.6 73.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pfo3f7to",
   "source_url": "https://omlx.ai/benchmarks/performance/pfo3f7to",
   "tg_tokens": null,
   "tps": 53.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pfo3f7to-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 76.0 GB at this context; duplicate traces of the same session: 2xcf5l4v, a9axmzo2, cecfgyaj, hq9p9znv, umsc49qm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2462.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,462 67.6 76.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pfo3f7to",
   "source_url": "https://omlx.ai/benchmarks/performance/pfo3f7to",
   "tg_tokens": null,
   "tps": 67.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dwtsksut-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: 7jdpovgh, mw05dmlo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 990.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 990.1 64.9 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dwtsksut",
   "source_url": "https://omlx.ai/benchmarks/performance/dwtsksut",
   "tg_tokens": null,
   "tps": 64.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dwtsksut-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 7jdpovgh, mw05dmlo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2321.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,321 81.7 71.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dwtsksut",
   "source_url": "https://omlx.ai/benchmarks/performance/dwtsksut",
   "tg_tokens": null,
   "tps": 81.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dwtsksut-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.7 GB at this context; duplicate traces of the same session: 7jdpovgh, mw05dmlo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2558.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,558 52.2 73.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dwtsksut",
   "source_url": "https://omlx.ai/benchmarks/performance/dwtsksut",
   "tg_tokens": null,
   "tps": 52.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-xjowgsub-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 2z2x2gad, 3lvhu50d, 3n9b7ueb, 778k3rh7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 101.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 101.1 29.9 18.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xjowgsub",
   "source_url": "https://omlx.ai/benchmarks/performance/xjowgsub",
   "tg_tokens": null,
   "tps": 29.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-xjowgsub-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 2z2x2gad, 3lvhu50d, 3n9b7ueb, 778k3rh7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 109.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 109.3 25.2 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xjowgsub",
   "source_url": "https://omlx.ai/benchmarks/performance/xjowgsub",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-xjowgsub-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 2z2x2gad, 3lvhu50d, 3n9b7ueb, 778k3rh7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 119.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 119.2 28.7 20.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xjowgsub",
   "source_url": "https://omlx.ai/benchmarks/performance/xjowgsub",
   "tg_tokens": null,
   "tps": 28.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-xjowgsub-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 2z2x2gad, 3lvhu50d, 3n9b7ueb, 778k3rh7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 113.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 113.5 31.9 20.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xjowgsub",
   "source_url": "https://omlx.ai/benchmarks/performance/xjowgsub",
   "tg_tokens": null,
   "tps": 31.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-xjowgsub-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: 2z2x2gad, 3lvhu50d, 3n9b7ueb, 778k3rh7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 103.2 27.3 22.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xjowgsub",
   "source_url": "https://omlx.ai/benchmarks/performance/xjowgsub",
   "tg_tokens": null,
   "tps": 27.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-xjowgsub-ctx65536",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.2 GB at this context; duplicate traces of the same session: 2z2x2gad, 3lvhu50d, 3n9b7ueb, 778k3rh7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 91.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 91.6 18.2 25.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xjowgsub",
   "source_url": "https://omlx.ai/benchmarks/performance/xjowgsub",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-8e0ra0hg-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: a0ppdnvp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1427.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "gemma-4-26b-a4b-it-6bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 1,427 91.0 20.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8e0ra0hg",
   "source_url": "https://omlx.ai/benchmarks/performance/8e0ra0hg",
   "tg_tokens": null,
   "tps": 91.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-8e0ra0hg-ctx16384",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: a0ppdnvp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1151.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "gemma-4-26b-a4b-it-6bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 1,151 80.9 20.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8e0ra0hg",
   "source_url": "https://omlx.ai/benchmarks/performance/8e0ra0hg",
   "tg_tokens": null,
   "tps": 80.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-kzzc06yd-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 177.4 GB; TTFT 16662 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1967.0,
   "provenance": "community",
   "quant": "mixed 4/8bit",
   "quote": "GLM-5.3-Flash-MLX-mixed-4_8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k tokens 1,967 PP tok/s 55.9 TG tok/s 16662 TTFT (ms) 177.4 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kzzc06yd",
   "source_url": "https://omlx.ai/benchmarks/performance/kzzc06yd",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-zgwhou78-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1366.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 1,366 44.8 4.2 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zgwhou78",
   "source_url": "https://omlx.ai/benchmarks/performance/zgwhou78",
   "tg_tokens": null,
   "tps": 44.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-zgwhou78-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.5 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1383.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 16k 1,383 37.3 4.5 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zgwhou78",
   "source_url": "https://omlx.ai/benchmarks/performance/zgwhou78",
   "tg_tokens": null,
   "tps": 37.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-40amoi90-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: 5uf41mwv, 9j37rfdj, ffdwtdeu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1758.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 1,758 57.7 5.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/40amoi90",
   "source_url": "https://omlx.ai/benchmarks/performance/40amoi90",
   "tg_tokens": null,
   "tps": 57.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-40amoi90-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: 5uf41mwv, 9j37rfdj, ffdwtdeu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1788.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,788 56.5 6.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/40amoi90",
   "source_url": "https://omlx.ai/benchmarks/performance/40amoi90",
   "tg_tokens": null,
   "tps": 56.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-40amoi90-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: 5uf41mwv, 9j37rfdj, ffdwtdeu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1822.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 1,822 55.0 6.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/40amoi90",
   "source_url": "https://omlx.ai/benchmarks/performance/40amoi90",
   "tg_tokens": null,
   "tps": 55.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-40amoi90-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: 5uf41mwv, 9j37rfdj, ffdwtdeu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1759.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 1,759 51.6 6.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/40amoi90",
   "source_url": "https://omlx.ai/benchmarks/performance/40amoi90",
   "tg_tokens": null,
   "tps": 51.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-202lztk2-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 7pzuxtzh, odfsefyg, tt174tk4; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 862.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 862.6 86.8 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/202lztk2",
   "source_url": "https://omlx.ai/benchmarks/performance/202lztk2",
   "tg_tokens": null,
   "tps": 86.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-202lztk2-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: 7pzuxtzh, odfsefyg, tt174tk4; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 747.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 747.2 74.9 20.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/202lztk2",
   "source_url": "https://omlx.ai/benchmarks/performance/202lztk2",
   "tg_tokens": null,
   "tps": 74.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-202lztk2-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: 7pzuxtzh, odfsefyg, tt174tk4; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 615.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 615.1 63.9 21.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/202lztk2",
   "source_url": "https://omlx.ai/benchmarks/performance/202lztk2",
   "tg_tokens": null,
   "tps": 63.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-202lztk2-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 7pzuxtzh, odfsefyg, tt174tk4; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 451.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 64k 451.9 48.2 22.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/202lztk2",
   "source_url": "https://omlx.ai/benchmarks/performance/202lztk2",
   "tg_tokens": null,
   "tps": 48.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-nt1ezu0i-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: zi98q8ej; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 964.9,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.6-35B-A3B-5bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 5bit; 4k 964.9 74.4 24.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nt1ezu0i",
   "source_url": "https://omlx.ai/benchmarks/performance/nt1ezu0i",
   "tg_tokens": null,
   "tps": 74.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-nt1ezu0i-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: zi98q8ej; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 928.3,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.6-35B-A3B-5bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 5bit; 8k 928.3 70.3 24.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nt1ezu0i",
   "source_url": "https://omlx.ai/benchmarks/performance/nt1ezu0i",
   "tg_tokens": null,
   "tps": 70.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Max",
   "id": "omlx-lgbvo54x-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 25.7 GB; TTFT 4298 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 238.3,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 238.3 PP tok/s 27.0 TG tok/s 4298 TTFT (ms) 25.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lgbvo54x",
   "source_url": "https://omlx.ai/benchmarks/performance/lgbvo54x",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ytosvl7v-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 21.9 GB; TTFT 664 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1542.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 1,542 PP tok/s 90.1 TG tok/s 664 TTFT (ms) 21.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ytosvl7v",
   "source_url": "https://omlx.ai/benchmarks/performance/ytosvl7v",
   "tg_tokens": null,
   "tps": 90.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-09cdix1b-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 6jlkt49f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4418.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,418 202.8 21.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/09cdix1b",
   "source_url": "https://omlx.ai/benchmarks/performance/09cdix1b",
   "tg_tokens": null,
   "tps": 202.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-09cdix1b-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 6jlkt49f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4476.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 4,476 164.3 22.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/09cdix1b",
   "source_url": "https://omlx.ai/benchmarks/performance/09cdix1b",
   "tg_tokens": null,
   "tps": 164.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-32j8ux20-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4419.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,419 222.3 21.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/32j8ux20",
   "source_url": "https://omlx.ai/benchmarks/performance/32j8ux20",
   "tg_tokens": null,
   "tps": 222.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-32j8ux20-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4954.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 4,954 164.1 22.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/32j8ux20",
   "source_url": "https://omlx.ai/benchmarks/performance/32j8ux20",
   "tg_tokens": null,
   "tps": 164.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hqgvfnq6-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 77y1hqh1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4434.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,434 224.0 21.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hqgvfnq6",
   "source_url": "https://omlx.ai/benchmarks/performance/hqgvfnq6",
   "tg_tokens": null,
   "tps": 224.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hqgvfnq6-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 77y1hqh1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6279.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,279 183.4 22.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hqgvfnq6",
   "source_url": "https://omlx.ai/benchmarks/performance/hqgvfnq6",
   "tg_tokens": null,
   "tps": 183.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-x4jgmois-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: ntvqx7dh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 468.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 468.0 18.2 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x4jgmois",
   "source_url": "https://omlx.ai/benchmarks/performance/x4jgmois",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-x4jgmois-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: ntvqx7dh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 505.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 505.3 17.8 17.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x4jgmois",
   "source_url": "https://omlx.ai/benchmarks/performance/x4jgmois",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-bh6bktpy-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: 8fjhmxs0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 6bit; 4k 125.3 24.1 24.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bh6bktpy",
   "source_url": "https://omlx.ai/benchmarks/performance/bh6bktpy",
   "tg_tokens": null,
   "tps": 24.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-bh6bktpy-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: 8fjhmxs0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 124.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 6bit; 8k 124.7 23.5 25.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bh6bktpy",
   "source_url": "https://omlx.ai/benchmarks/performance/bh6bktpy",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-x47es9f4-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 996.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 996.6 66.9 17.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x47es9f4",
   "source_url": "https://omlx.ai/benchmarks/performance/x47es9f4",
   "tg_tokens": null,
   "tps": 66.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-x47es9f4-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1076.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,076 53.0 19.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x47es9f4",
   "source_url": "https://omlx.ai/benchmarks/performance/x47es9f4",
   "tg_tokens": null,
   "tps": 53.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1zfuqqiu-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: lj34wxng; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 908.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 908.0 76.7 17.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1zfuqqiu",
   "source_url": "https://omlx.ai/benchmarks/performance/1zfuqqiu",
   "tg_tokens": null,
   "tps": 76.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1zfuqqiu-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: lj34wxng; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 964.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 964.6 63.7 18.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1zfuqqiu",
   "source_url": "https://omlx.ai/benchmarks/performance/1zfuqqiu",
   "tg_tokens": null,
   "tps": 63.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mnd3k9p2-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: g1hsdeww; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 723.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 723.2 70.1 17.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mnd3k9p2",
   "source_url": "https://omlx.ai/benchmarks/performance/mnd3k9p2",
   "tg_tokens": null,
   "tps": 70.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mnd3k9p2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: g1hsdeww; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 730.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 730.9 54.0 18.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mnd3k9p2",
   "source_url": "https://omlx.ai/benchmarks/performance/mnd3k9p2",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7fkwd3hd-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 19.2 GB; TTFT 3062 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1338.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 1,338 PP tok/s 60.3 TG tok/s 3062 TTFT (ms) 19.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7fkwd3hd",
   "source_url": "https://omlx.ai/benchmarks/performance/7fkwd3hd",
   "tg_tokens": null,
   "tps": 60.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4thgd1wy-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: ue3xcxsa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1128.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,128 59.9 17.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4thgd1wy",
   "source_url": "https://omlx.ai/benchmarks/performance/4thgd1wy",
   "tg_tokens": null,
   "tps": 59.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4thgd1wy-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: ue3xcxsa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1367.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,367 68.7 19.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4thgd1wy",
   "source_url": "https://omlx.ai/benchmarks/performance/4thgd1wy",
   "tg_tokens": null,
   "tps": 68.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qtcuy2n0-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 19.6 GB; TTFT 39891 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 410.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k tokens 410.7 PP tok/s 39.9 TG tok/s 39891 TTFT (ms) 19.6 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qtcuy2n0",
   "source_url": "https://omlx.ai/benchmarks/performance/qtcuy2n0",
   "tg_tokens": null,
   "tps": 39.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-o7zntjcz-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 19.0 GB; TTFT 20294 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 403.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k tokens 403.7 PP tok/s 24.0 TG tok/s 20294 TTFT (ms) 19.0 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o7zntjcz",
   "source_url": "https://omlx.ai/benchmarks/performance/o7zntjcz",
   "tg_tokens": null,
   "tps": 24.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-76guqois-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 19.0 GB; TTFT 20591 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 397.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k tokens 397.9 PP tok/s 30.8 TG tok/s 20591 TTFT (ms) 19.0 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/76guqois",
   "source_url": "https://omlx.ai/benchmarks/performance/76guqois",
   "tg_tokens": null,
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7obvyigp-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: rg7qjncj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 903.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 903.1 51.6 27.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7obvyigp",
   "source_url": "https://omlx.ai/benchmarks/performance/7obvyigp",
   "tg_tokens": null,
   "tps": 51.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7obvyigp-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: rg7qjncj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 766.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 766.2 60.0 28.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7obvyigp",
   "source_url": "https://omlx.ai/benchmarks/performance/7obvyigp",
   "tg_tokens": null,
   "tps": 60.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-e4h5t3hi-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: pepdnzdy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 846.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 846.9 39.6 27.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e4h5t3hi",
   "source_url": "https://omlx.ai/benchmarks/performance/e4h5t3hi",
   "tg_tokens": null,
   "tps": 39.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-e4h5t3hi-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: pepdnzdy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 711.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 711.4 50.8 28.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e4h5t3hi",
   "source_url": "https://omlx.ai/benchmarks/performance/e4h5t3hi",
   "tg_tokens": null,
   "tps": 50.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-tqkkj8e9-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: 226c7rcj, 52txngqu, aozx3y8s, wkq2wl4p; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 627.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 8k 627.2 45.8 71.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tqkkj8e9",
   "source_url": "https://omlx.ai/benchmarks/performance/tqkkj8e9",
   "tg_tokens": null,
   "tps": 45.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-tqkkj8e9-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 226c7rcj, 52txngqu, aozx3y8s, wkq2wl4p; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 626.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 16k 626.9 57.3 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tqkkj8e9",
   "source_url": "https://omlx.ai/benchmarks/performance/tqkkj8e9",
   "tg_tokens": null,
   "tps": 57.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-tqkkj8e9-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.4 GB at this context; duplicate traces of the same session: 226c7rcj, 52txngqu, aozx3y8s, wkq2wl4p; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 624.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 32k 624.0 43.1 72.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tqkkj8e9",
   "source_url": "https://omlx.ai/benchmarks/performance/tqkkj8e9",
   "tg_tokens": null,
   "tps": 43.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-tqkkj8e9-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.5 GB at this context; duplicate traces of the same session: 226c7rcj, 52txngqu, aozx3y8s, wkq2wl4p; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 616.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 64k 616.6 47.3 74.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tqkkj8e9",
   "source_url": "https://omlx.ai/benchmarks/performance/tqkkj8e9",
   "tg_tokens": null,
   "tps": 47.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-tqkkj8e9-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.6 GB at this context; duplicate traces of the same session: 226c7rcj, 52txngqu, aozx3y8s, wkq2wl4p; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 515.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 128k 515.8 44.0 78.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tqkkj8e9",
   "source_url": "https://omlx.ai/benchmarks/performance/tqkkj8e9",
   "tg_tokens": null,
   "tps": 44.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-udyfzreo-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.8 GB at this context; duplicate traces of the same session: dwfizm13, v3unjx80, w32o4h7k, zqhlq28c; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 540.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 1k 540.5 51.7 100.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/udyfzreo",
   "source_url": "https://omlx.ai/benchmarks/performance/udyfzreo",
   "tg_tokens": null,
   "tps": 51.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-udyfzreo-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: dwfizm13, v3unjx80, w32o4h7k, zqhlq28c; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 633.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 4k 633.2 56.4 101.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/udyfzreo",
   "source_url": "https://omlx.ai/benchmarks/performance/udyfzreo",
   "tg_tokens": null,
   "tps": 56.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-udyfzreo-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: dwfizm13, v3unjx80, w32o4h7k, zqhlq28c; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 628.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 8k 628.8 55.4 101.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/udyfzreo",
   "source_url": "https://omlx.ai/benchmarks/performance/udyfzreo",
   "tg_tokens": null,
   "tps": 55.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-udyfzreo-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.6 GB at this context; duplicate traces of the same session: dwfizm13, v3unjx80, w32o4h7k, zqhlq28c; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 627.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 16k 627.3 59.1 101.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/udyfzreo",
   "source_url": "https://omlx.ai/benchmarks/performance/udyfzreo",
   "tg_tokens": null,
   "tps": 59.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-udyfzreo-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.2 GB at this context; duplicate traces of the same session: dwfizm13, v3unjx80, w32o4h7k, zqhlq28c; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 623.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 32k 623.6 40.4 102.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/udyfzreo",
   "source_url": "https://omlx.ai/benchmarks/performance/udyfzreo",
   "tg_tokens": null,
   "tps": 40.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-udyfzreo-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.3 GB at this context; duplicate traces of the same session: dwfizm13, v3unjx80, w32o4h7k, zqhlq28c; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 616.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 64k 616.4 49.2 104.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/udyfzreo",
   "source_url": "https://omlx.ai/benchmarks/performance/udyfzreo",
   "tg_tokens": null,
   "tps": 49.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-udyfzreo-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 108.4 GB at this context; duplicate traces of the same session: dwfizm13, v3unjx80, w32o4h7k, zqhlq28c; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 603.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 128k 603.3 44.3 108.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/udyfzreo",
   "source_url": "https://omlx.ai/benchmarks/performance/udyfzreo",
   "tg_tokens": null,
   "tps": 44.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-xfwt6qoz-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace, single run at 128k context; peak memory 152.5 GB; TTFT 34288 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3823.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 128k tokens 3,823 PP tok/s 96.6 TG tok/s 34288 TTFT (ms) 152.5 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xfwt6qoz",
   "source_url": "https://omlx.ai/benchmarks/performance/xfwt6qoz",
   "tg_tokens": null,
   "tps": 96.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-1rdlqqtl-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 101.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 101.1 29.9 18.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1rdlqqtl",
   "source_url": "https://omlx.ai/benchmarks/performance/1rdlqqtl",
   "tg_tokens": null,
   "tps": 29.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-1rdlqqtl-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 109.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 109.3 25.2 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1rdlqqtl",
   "source_url": "https://omlx.ai/benchmarks/performance/1rdlqqtl",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-1rdlqqtl-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 119.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 119.2 28.7 20.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1rdlqqtl",
   "source_url": "https://omlx.ai/benchmarks/performance/1rdlqqtl",
   "tg_tokens": null,
   "tps": 28.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-1rdlqqtl-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 113.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 113.5 31.9 20.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1rdlqqtl",
   "source_url": "https://omlx.ai/benchmarks/performance/1rdlqqtl",
   "tg_tokens": null,
   "tps": 31.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-1rdlqqtl-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 103.2 27.3 22.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1rdlqqtl",
   "source_url": "https://omlx.ai/benchmarks/performance/1rdlqqtl",
   "tg_tokens": null,
   "tps": 27.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-1rdlqqtl-ctx65536",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.2 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 91.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 91.6 18.2 25.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1rdlqqtl",
   "source_url": "https://omlx.ai/benchmarks/performance/1rdlqqtl",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-u0p1526o-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 21.4 GB; TTFT 94903 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 345.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 32k tokens 345.3 PP tok/s 49.0 TG tok/s 94903 TTFT (ms) 21.4 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u0p1526o",
   "source_url": "https://omlx.ai/benchmarks/performance/u0p1526o",
   "tg_tokens": null,
   "tps": 49.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-8pjt1r7x-ctx1024",
   "model": "Hy-MT2-1.8B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 1.8 GB at this context; duplicate traces of the same session: 9279ml6s, udvp0wyo, zmr25fwg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 805.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Hy-MT2-1.8B-mlx-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 805.3 104.5 1.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8pjt1r7x",
   "source_url": "https://omlx.ai/benchmarks/performance/8pjt1r7x",
   "tg_tokens": null,
   "tps": 104.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-8pjt1r7x-ctx4096",
   "model": "Hy-MT2-1.8B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 1.9 GB at this context; duplicate traces of the same session: 9279ml6s, udvp0wyo, zmr25fwg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 749.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Hy-MT2-1.8B-mlx-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 749.4 90.0 1.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8pjt1r7x",
   "source_url": "https://omlx.ai/benchmarks/performance/8pjt1r7x",
   "tg_tokens": null,
   "tps": 90.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-8pjt1r7x-ctx8192",
   "model": "Hy-MT2-1.8B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 2.1 GB at this context; duplicate traces of the same session: 9279ml6s, udvp0wyo, zmr25fwg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 676.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Hy-MT2-1.8B-mlx-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 8k 676.2 76.5 2.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8pjt1r7x",
   "source_url": "https://omlx.ai/benchmarks/performance/8pjt1r7x",
   "tg_tokens": null,
   "tps": 76.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-8pjt1r7x-ctx16384",
   "model": "Hy-MT2-1.8B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 2.6 GB at this context; duplicate traces of the same session: 9279ml6s, udvp0wyo, zmr25fwg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 538.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Hy-MT2-1.8B-mlx-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k 538.8 57.1 2.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8pjt1r7x",
   "source_url": "https://omlx.ai/benchmarks/performance/8pjt1r7x",
   "tg_tokens": null,
   "tps": 57.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-6idkf8z4-ctx1024",
   "model": "Hy-MT2-1.8B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 1.8 GB at this context; duplicate traces of the same session: qbpx2a3a, wnx477s3; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 815.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Hy-MT2-1.8B-mlx-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 815.8 103.7 1.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6idkf8z4",
   "source_url": "https://omlx.ai/benchmarks/performance/6idkf8z4",
   "tg_tokens": null,
   "tps": 103.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-6idkf8z4-ctx4096",
   "model": "Hy-MT2-1.8B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 1.9 GB at this context; duplicate traces of the same session: qbpx2a3a, wnx477s3; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 776.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Hy-MT2-1.8B-mlx-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 776.9 89.5 1.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6idkf8z4",
   "source_url": "https://omlx.ai/benchmarks/performance/6idkf8z4",
   "tg_tokens": null,
   "tps": 89.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-6idkf8z4-ctx8192",
   "model": "Hy-MT2-1.8B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 2.1 GB at this context; duplicate traces of the same session: qbpx2a3a, wnx477s3; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 681.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Hy-MT2-1.8B-mlx-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 8k 681.7 74.2 2.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6idkf8z4",
   "source_url": "https://omlx.ai/benchmarks/performance/6idkf8z4",
   "tg_tokens": null,
   "tps": 74.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-axr6vpv4-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 36.2 GB; TTFT 2392 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1713.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k tokens 1,713 PP tok/s 101.4 TG tok/s 2392 TTFT (ms) 36.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/axr6vpv4",
   "source_url": "https://omlx.ai/benchmarks/performance/axr6vpv4",
   "tg_tokens": null,
   "tps": 101.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-c8zu053v-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 6xxkzsqt, ia3d7nc1, m8tsr0f2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 347.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 347.9 52.3 19.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c8zu053v",
   "source_url": "https://omlx.ai/benchmarks/performance/c8zu053v",
   "tg_tokens": null,
   "tps": 52.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-c8zu053v-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 6xxkzsqt, ia3d7nc1, m8tsr0f2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 394.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 394.6 51.0 20.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c8zu053v",
   "source_url": "https://omlx.ai/benchmarks/performance/c8zu053v",
   "tg_tokens": null,
   "tps": 51.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-c8zu053v-ctx8192",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: 6xxkzsqt, ia3d7nc1, m8tsr0f2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 383.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 8k 383.1 47.4 20.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c8zu053v",
   "source_url": "https://omlx.ai/benchmarks/performance/c8zu053v",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-c8zu053v-ctx16384",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: 6xxkzsqt, ia3d7nc1, m8tsr0f2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k 323.1 43.2 20.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c8zu053v",
   "source_url": "https://omlx.ai/benchmarks/performance/c8zu053v",
   "tg_tokens": null,
   "tps": 43.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-krrhdu9b-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 1f8ylcuw; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 904.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 904.9 102.9 21.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/krrhdu9b",
   "source_url": "https://omlx.ai/benchmarks/performance/krrhdu9b",
   "tg_tokens": null,
   "tps": 102.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-krrhdu9b-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: 1f8ylcuw; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1046.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,046 91.3 23.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/krrhdu9b",
   "source_url": "https://omlx.ai/benchmarks/performance/krrhdu9b",
   "tg_tokens": null,
   "tps": 91.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q5yyhgfz-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 30.9 GB; TTFT 3702 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4425.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k tokens 4,425 PP tok/s 113.6 TG tok/s 3702 TTFT (ms) 30.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q5yyhgfz",
   "source_url": "https://omlx.ai/benchmarks/performance/q5yyhgfz",
   "tg_tokens": null,
   "tps": 113.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0owh98rk-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 53.7 GB at this context; duplicate traces of the same session: gki3e5jm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 861.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.8-27B on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 861.0 28.0 53.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0owh98rk",
   "source_url": "https://omlx.ai/benchmarks/performance/0owh98rk",
   "tg_tokens": null,
   "tps": 28.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0owh98rk-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 55.0 GB at this context; duplicate traces of the same session: gki3e5jm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1038.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.8-27B on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 4k 1,038 23.0 55.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0owh98rk",
   "source_url": "https://omlx.ai/benchmarks/performance/0owh98rk",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-21j9xzfm-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: bsw4zw7s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 834.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 834.5 19.1 28.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/21j9xzfm",
   "source_url": "https://omlx.ai/benchmarks/performance/21j9xzfm",
   "tg_tokens": null,
   "tps": 19.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-21j9xzfm-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: bsw4zw7s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 961.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 961.1 19.0 29.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/21j9xzfm",
   "source_url": "https://omlx.ai/benchmarks/performance/21j9xzfm",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-60t5a86b-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace, single run at 4k context; peak memory 27.0 GB; TTFT 50160 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 81.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k tokens 81.7 PP tok/s 12.9 TG tok/s 50160 TTFT (ms) 27.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/60t5a86b",
   "source_url": "https://omlx.ai/benchmarks/performance/60t5a86b",
   "tg_tokens": null,
   "tps": 12.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-8xea6rc5-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace, single run at 4k context; peak memory 27.0 GB; TTFT 50161 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 81.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k tokens 81.7 PP tok/s 12.6 TG tok/s 50161 TTFT (ms) 27.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8xea6rc5",
   "source_url": "https://omlx.ai/benchmarks/performance/8xea6rc5",
   "tg_tokens": null,
   "tps": 12.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-5deu6ice-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace, single run at 4k context; peak memory 27.0 GB; TTFT 50272 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 81.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k tokens 81.5 PP tok/s 10.2 TG tok/s 50272 TTFT (ms) 27.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5deu6ice",
   "source_url": "https://omlx.ai/benchmarks/performance/5deu6ice",
   "tg_tokens": null,
   "tps": 10.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-or072v3v-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized speed FP16 build; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: r249mzbj; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 110.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed-FP16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 110.3 14.1 24.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/or072v3v",
   "source_url": "https://omlx.ai/benchmarks/performance/or072v3v",
   "tg_tokens": null,
   "tps": 14.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-or072v3v-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized speed FP16 build; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: r249mzbj; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 99.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed-FP16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 99.1 12.4 28.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/or072v3v",
   "source_url": "https://omlx.ai/benchmarks/performance/or072v3v",
   "tg_tokens": null,
   "tps": 12.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qu81oe6g-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 52.0 GB at this context; duplicate traces of the same session: eha48pte; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 891.8,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.8-27B-bf16 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 bf16; 1k 891.8 10.4 52.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qu81oe6g",
   "source_url": "https://omlx.ai/benchmarks/performance/qu81oe6g",
   "tg_tokens": null,
   "tps": 10.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qu81oe6g-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 53.3 GB at this context; duplicate traces of the same session: eha48pte; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1069.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.8-27B-bf16 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 bf16; 4k 1,069 10.4 53.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qu81oe6g",
   "source_url": "https://omlx.ai/benchmarks/performance/qu81oe6g",
   "tg_tokens": null,
   "tps": 10.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ytggevyl-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: 3qehqmkv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 836.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 836.9 41.1 26.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ytggevyl",
   "source_url": "https://omlx.ai/benchmarks/performance/ytggevyl",
   "tg_tokens": null,
   "tps": 41.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ytggevyl-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 3qehqmkv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 898.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 898.9 63.8 28.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ytggevyl",
   "source_url": "https://omlx.ai/benchmarks/performance/ytggevyl",
   "tg_tokens": null,
   "tps": 63.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-voup7o9i-ctx1024",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: ciatm207; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1670.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 1,670 57.1 5.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/voup7o9i",
   "source_url": "https://omlx.ai/benchmarks/performance/voup7o9i",
   "tg_tokens": null,
   "tps": 57.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-voup7o9i-ctx4096",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: ciatm207; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1813.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,813 56.0 6.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/voup7o9i",
   "source_url": "https://omlx.ai/benchmarks/performance/voup7o9i",
   "tg_tokens": null,
   "tps": 56.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-usg5sc6y-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 48.4 GB; TTFT 45626 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 89.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k tokens 89.8 PP tok/s 9.4 TG tok/s 45626 TTFT (ms) 48.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/usg5sc6y",
   "source_url": "https://omlx.ai/benchmarks/performance/usg5sc6y",
   "tg_tokens": null,
   "tps": 9.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-drbqbc1d-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 49.8 GB; TTFT 52890 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 77.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k tokens 77.4 PP tok/s 8.9 TG tok/s 52890 TTFT (ms) 49.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/drbqbc1d",
   "source_url": "https://omlx.ai/benchmarks/performance/drbqbc1d",
   "tg_tokens": null,
   "tps": 8.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-4ldacihh-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.8 GB at this context; duplicate traces of the same session: 7z04djx5; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 540.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 1k 540.5 51.7 100.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ldacihh",
   "source_url": "https://omlx.ai/benchmarks/performance/4ldacihh",
   "tg_tokens": null,
   "tps": 51.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-4ldacihh-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: 7z04djx5; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 633.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 4k 633.2 56.4 101.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ldacihh",
   "source_url": "https://omlx.ai/benchmarks/performance/4ldacihh",
   "tg_tokens": null,
   "tps": 56.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-4ldacihh-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: 7z04djx5; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 628.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 8k 628.8 55.4 101.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ldacihh",
   "source_url": "https://omlx.ai/benchmarks/performance/4ldacihh",
   "tg_tokens": null,
   "tps": 55.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-4ldacihh-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.6 GB at this context; duplicate traces of the same session: 7z04djx5; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 627.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 16k 627.3 59.1 101.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ldacihh",
   "source_url": "https://omlx.ai/benchmarks/performance/4ldacihh",
   "tg_tokens": null,
   "tps": 59.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-4ldacihh-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.2 GB at this context; duplicate traces of the same session: 7z04djx5; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 623.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 32k 623.6 40.4 102.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ldacihh",
   "source_url": "https://omlx.ai/benchmarks/performance/4ldacihh",
   "tg_tokens": null,
   "tps": 40.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-4ldacihh-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.3 GB at this context; duplicate traces of the same session: 7z04djx5; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 616.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 64k 616.4 49.2 104.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ldacihh",
   "source_url": "https://omlx.ai/benchmarks/performance/4ldacihh",
   "tg_tokens": null,
   "tps": 49.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-4ldacihh-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 108.4 GB at this context; duplicate traces of the same session: 7z04djx5; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 603.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 128k 603.3 44.3 108.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ldacihh",
   "source_url": "https://omlx.ai/benchmarks/performance/4ldacihh",
   "tg_tokens": null,
   "tps": 44.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-tdinjkkl-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: 44k00ohh, okt47e09; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 575.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 575.6 67.5 70.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tdinjkkl",
   "source_url": "https://omlx.ai/benchmarks/performance/tdinjkkl",
   "tg_tokens": null,
   "tps": 67.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-tdinjkkl-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: 44k00ohh, okt47e09; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 701.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 701.2 59.9 71.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tdinjkkl",
   "source_url": "https://omlx.ai/benchmarks/performance/tdinjkkl",
   "tg_tokens": null,
   "tps": 59.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-tdinjkkl-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: 44k00ohh, okt47e09; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1046.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 1,046 63.4 71.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tdinjkkl",
   "source_url": "https://omlx.ai/benchmarks/performance/tdinjkkl",
   "tg_tokens": null,
   "tps": 63.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xhusd3rs-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 102.8 GB; TTFT 5652 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2899.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k tokens 2,899 PP tok/s 48.4 TG tok/s 5652 TTFT (ms) 102.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xhusd3rs",
   "source_url": "https://omlx.ai/benchmarks/performance/xhusd3rs",
   "tg_tokens": null,
   "tps": 48.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ase533f1-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 195k context; peak memory 109.7 GB; TTFT 79229 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2524.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k tokens 2,524 PP tok/s 62.6 TG tok/s 79229 TTFT (ms) 109.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ase533f1",
   "source_url": "https://omlx.ai/benchmarks/performance/ase533f1",
   "tg_tokens": null,
   "tps": 62.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3dtq0qhr-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 100.5 GB; TTFT 2823 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 362.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 362.7 PP tok/s 49.3 TG tok/s 2823 TTFT (ms) 100.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3dtq0qhr",
   "source_url": "https://omlx.ai/benchmarks/performance/3dtq0qhr",
   "tg_tokens": null,
   "tps": 49.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-2n9kt8e0-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 74.2 GB at this context; duplicate traces of the same session: 271ilq5m, 64wh4572, jb0ad7o3, l1wx5fma, xxqn0i6n; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 512.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-fp16 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 1k 512.7 20.3 74.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2n9kt8e0",
   "source_url": "https://omlx.ai/benchmarks/performance/2n9kt8e0",
   "tg_tokens": null,
   "tps": 20.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-2n9kt8e0-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 81.1 GB at this context; duplicate traces of the same session: 271ilq5m, 64wh4572, jb0ad7o3, l1wx5fma, xxqn0i6n; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 423.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-fp16 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 4k 423.8 21.0 81.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2n9kt8e0",
   "source_url": "https://omlx.ai/benchmarks/performance/2n9kt8e0",
   "tg_tokens": null,
   "tps": 21.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-2n9kt8e0-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 84.1 GB at this context; duplicate traces of the same session: 271ilq5m, 64wh4572, jb0ad7o3, l1wx5fma, xxqn0i6n; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 552.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-fp16 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 8k 552.3 27.1 84.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2n9kt8e0",
   "source_url": "https://omlx.ai/benchmarks/performance/2n9kt8e0",
   "tg_tokens": null,
   "tps": 27.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-2n9kt8e0-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 84.8 GB at this context; duplicate traces of the same session: 271ilq5m, 64wh4572, jb0ad7o3, l1wx5fma, xxqn0i6n; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 538.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-fp16 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 16k 538.2 26.9 84.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2n9kt8e0",
   "source_url": "https://omlx.ai/benchmarks/performance/2n9kt8e0",
   "tg_tokens": null,
   "tps": 26.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-2n9kt8e0-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 92.2 GB at this context; duplicate traces of the same session: 271ilq5m, 64wh4572, jb0ad7o3, l1wx5fma, xxqn0i6n; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 516.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-fp16 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 64k 516.9 22.3 92.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2n9kt8e0",
   "source_url": "https://omlx.ai/benchmarks/performance/2n9kt8e0",
   "tg_tokens": null,
   "tps": 22.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-2n9kt8e0-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 97.0 GB at this context; duplicate traces of the same session: 271ilq5m, 64wh4572, jb0ad7o3, l1wx5fma, xxqn0i6n; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 421.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-fp16 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 128k 421.4 21.9 97.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2n9kt8e0",
   "source_url": "https://omlx.ai/benchmarks/performance/2n9kt8e0",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-gb2usmqj-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 1k context; peak memory 22.5 GB; TTFT 866 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1182.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k tokens 1,182 PP tok/s 57.6 TG tok/s 866 TTFT (ms) 22.5 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gb2usmqj",
   "source_url": "https://omlx.ai/benchmarks/performance/gb2usmqj",
   "tg_tokens": null,
   "tps": 57.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-uppv4dca-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 1k context; peak memory 20.0 GB; TTFT 461 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2222.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k tokens 2,222 PP tok/s 48.9 TG tok/s 461 TTFT (ms) 20.0 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uppv4dca",
   "source_url": "https://omlx.ai/benchmarks/performance/uppv4dca",
   "tg_tokens": null,
   "tps": 48.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-yxcr22ik-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 1k context; peak memory 11.3 GB; TTFT 4077 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 251.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k tokens 251.2 PP tok/s 25.0 TG tok/s 4077 TTFT (ms) 11.3 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yxcr22ik",
   "source_url": "https://omlx.ai/benchmarks/performance/yxcr22ik",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-5y4k1mgy-ctx32768",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: r2unfzie; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 394.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 32k 394.8 32.4 15.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5y4k1mgy",
   "source_url": "https://omlx.ai/benchmarks/performance/5y4k1mgy",
   "tg_tokens": null,
   "tps": 32.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-5y4k1mgy-ctx65536",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: r2unfzie; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 353.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 64k 353.2 22.4 17.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5y4k1mgy",
   "source_url": "https://omlx.ai/benchmarks/performance/5y4k1mgy",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-55y89x7d-ctx65536",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 22.8 GB; TTFT 71093 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 921.8,
   "provenance": "community",
   "quant": "unknown",
   "quote": "MiMo-V2.6-Distill-Qwen-9B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 unknown; 64k tokens 921.8 PP tok/s 34.8 TG tok/s 71093 TTFT (ms) 22.8 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/55y89x7d",
   "source_url": "https://omlx.ai/benchmarks/performance/55y89x7d",
   "tg_tokens": null,
   "tps": 34.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-t9wap4xq-ctx65536",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 12.8 GB; TTFT 70753 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 926.3,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-OptiQ-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k tokens 926.3 PP tok/s 65.7 TG tok/s 70753 TTFT (ms) 12.8 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t9wap4xq",
   "source_url": "https://omlx.ai/benchmarks/performance/t9wap4xq",
   "tg_tokens": null,
   "tps": 65.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-092zn0oq-ctx1024",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 156.7 GB at this context; duplicate traces of the same session: 16cv6ene, 7pekbaxe, ot4x41zj, pckke4yk, pm6ev9dr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 409.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-V2.6-Flash-RL-MLX-4bit-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 409.1 51.7 156.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/092zn0oq",
   "source_url": "https://omlx.ai/benchmarks/performance/092zn0oq",
   "tg_tokens": null,
   "tps": 51.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-092zn0oq-ctx4096",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 157.8 GB at this context; duplicate traces of the same session: 16cv6ene, 7pekbaxe, ot4x41zj, pckke4yk, pm6ev9dr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 480.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-V2.6-Flash-RL-MLX-4bit-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 480.7 51.5 157.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/092zn0oq",
   "source_url": "https://omlx.ai/benchmarks/performance/092zn0oq",
   "tg_tokens": null,
   "tps": 51.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-092zn0oq-ctx8192",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 158.9 GB at this context; duplicate traces of the same session: 16cv6ene, 7pekbaxe, ot4x41zj, pckke4yk, pm6ev9dr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 471.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-V2.6-Flash-RL-MLX-4bit-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 471.5 47.7 158.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/092zn0oq",
   "source_url": "https://omlx.ai/benchmarks/performance/092zn0oq",
   "tg_tokens": null,
   "tps": 47.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-092zn0oq-ctx16384",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 161.1 GB at this context; duplicate traces of the same session: 16cv6ene, 7pekbaxe, ot4x41zj, pckke4yk, pm6ev9dr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 447.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-V2.6-Flash-RL-MLX-4bit-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 447.7 41.3 161.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/092zn0oq",
   "source_url": "https://omlx.ai/benchmarks/performance/092zn0oq",
   "tg_tokens": null,
   "tps": 41.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-092zn0oq-ctx32768",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 165.4 GB at this context; duplicate traces of the same session: 16cv6ene, 7pekbaxe, ot4x41zj, pckke4yk, pm6ev9dr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 389.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-V2.6-Flash-RL-MLX-4bit-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 389.3 40.9 165.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/092zn0oq",
   "source_url": "https://omlx.ai/benchmarks/performance/092zn0oq",
   "tg_tokens": null,
   "tps": 40.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-092zn0oq-ctx65536",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 174.1 GB at this context; duplicate traces of the same session: 16cv6ene, 7pekbaxe, ot4x41zj, pckke4yk, pm6ev9dr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 304.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-V2.6-Flash-RL-MLX-4bit-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 304.0 33.0 174.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/092zn0oq",
   "source_url": "https://omlx.ai/benchmarks/performance/092zn0oq",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ii5afuji-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Nail build; oMLX community benchmark trace; peak memory 30.9 GB at this context; duplicate traces of the same session: 1bf1kh0t; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 554.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX-6bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 1k 554.0 52.4 30.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ii5afuji",
   "source_url": "https://omlx.ai/benchmarks/performance/ii5afuji",
   "tg_tokens": null,
   "tps": 52.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-ii5afuji-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Nail build; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: 1bf1kh0t; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 558.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX-6bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 4k 558.9 34.7 32.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ii5afuji",
   "source_url": "https://omlx.ai/benchmarks/performance/ii5afuji",
   "tg_tokens": null,
   "tps": 34.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-r0kkhwjt-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Nail build; oMLX community benchmark trace; peak memory 33.2 GB at this context; duplicate traces of the same session: bz48q9h2; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 547.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX-6bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 1k 547.5 45.5 33.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r0kkhwjt",
   "source_url": "https://omlx.ai/benchmarks/performance/r0kkhwjt",
   "tg_tokens": null,
   "tps": 45.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-r0kkhwjt-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Nail build; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: bz48q9h2; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 582.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX-6bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 4k 582.3 31.3 34.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r0kkhwjt",
   "source_url": "https://omlx.ai/benchmarks/performance/r0kkhwjt",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-xoi5zbi2-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Nail build; oMLX community benchmark trace; peak memory 30.7 GB at this context; duplicate traces of the same session: bbspgq67; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 514.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX-6bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 1k 514.9 51.2 30.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xoi5zbi2",
   "source_url": "https://omlx.ai/benchmarks/performance/xoi5zbi2",
   "tg_tokens": null,
   "tps": 51.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-xoi5zbi2-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Nail build; oMLX community benchmark trace; peak memory 32.5 GB at this context; duplicate traces of the same session: bbspgq67; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 566.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX-6bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 4k 566.9 49.7 32.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xoi5zbi2",
   "source_url": "https://omlx.ai/benchmarks/performance/xoi5zbi2",
   "tg_tokens": null,
   "tps": 49.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-rvpjpcsh-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: izszql0q, llgn7cyq, s1pawpa6, uxjc8j11; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 274.0,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Ornith-1.5-35B-A3B-oQ3e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 3bit; 1k 274.0 17.2 5.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rvpjpcsh",
   "source_url": "https://omlx.ai/benchmarks/performance/rvpjpcsh",
   "tg_tokens": null,
   "tps": 17.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-rvpjpcsh-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: izszql0q, llgn7cyq, s1pawpa6, uxjc8j11; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 503.8,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Ornith-1.5-35B-A3B-oQ3e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 3bit; 4k 503.8 17.9 6.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rvpjpcsh",
   "source_url": "https://omlx.ai/benchmarks/performance/rvpjpcsh",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-rvpjpcsh-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: izszql0q, llgn7cyq, s1pawpa6, uxjc8j11; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 500.5,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Ornith-1.5-35B-A3B-oQ3e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 3bit; 16k 500.5 15.0 7.0 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rvpjpcsh",
   "source_url": "https://omlx.ai/benchmarks/performance/rvpjpcsh",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-rvpjpcsh-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.3 GB at this context; duplicate traces of the same session: izszql0q, llgn7cyq, s1pawpa6, uxjc8j11; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 484.1,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Ornith-1.5-35B-A3B-oQ3e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 3bit; 32k 484.1 15.3 7.3 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rvpjpcsh",
   "source_url": "https://omlx.ai/benchmarks/performance/rvpjpcsh",
   "tg_tokens": null,
   "tps": 15.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-rvpjpcsh-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.9 GB at this context; duplicate traces of the same session: izszql0q, llgn7cyq, s1pawpa6, uxjc8j11; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 282.9,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Ornith-1.5-35B-A3B-oQ3e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 3bit; 64k 282.9 14.8 7.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rvpjpcsh",
   "source_url": "https://omlx.ai/benchmarks/performance/rvpjpcsh",
   "tg_tokens": null,
   "tps": 14.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-rvpjpcsh-ctx131072",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 9.4 GB at this context; duplicate traces of the same session: izszql0q, llgn7cyq, s1pawpa6, uxjc8j11; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 349.1,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Ornith-1.5-35B-A3B-oQ3e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 3bit; 128k 349.1 12.3 9.4 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rvpjpcsh",
   "source_url": "https://omlx.ai/benchmarks/performance/rvpjpcsh",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ksrkmdhj-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: 6x7hen7g, j3ifr5mi, pwe4h0we, sqvc9ibj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4203.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 4,203 101.4 35.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ksrkmdhj",
   "source_url": "https://omlx.ai/benchmarks/performance/ksrkmdhj",
   "tg_tokens": null,
   "tps": 101.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ksrkmdhj-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 36.7 GB at this context; duplicate traces of the same session: 6x7hen7g, j3ifr5mi, pwe4h0we, sqvc9ibj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4688.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 32k 4,688 88.0 36.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ksrkmdhj",
   "source_url": "https://omlx.ai/benchmarks/performance/ksrkmdhj",
   "tg_tokens": null,
   "tps": 88.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ksrkmdhj-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: 6x7hen7g, j3ifr5mi, pwe4h0we, sqvc9ibj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3668.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 64k 3,668 76.8 37.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ksrkmdhj",
   "source_url": "https://omlx.ai/benchmarks/performance/ksrkmdhj",
   "tg_tokens": null,
   "tps": 76.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ksrkmdhj-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 38.6 GB at this context; duplicate traces of the same session: 6x7hen7g, j3ifr5mi, pwe4h0we, sqvc9ibj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2519.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 128k 2,519 62.6 38.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ksrkmdhj",
   "source_url": "https://omlx.ai/benchmarks/performance/ksrkmdhj",
   "tg_tokens": null,
   "tps": 62.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ksrkmdhj-ctx199680",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 39.9 GB at this context; duplicate traces of the same session: 6x7hen7g, j3ifr5mi, pwe4h0we, sqvc9ibj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1886.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 195k 1,886 52.6 39.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ksrkmdhj",
   "source_url": "https://omlx.ai/benchmarks/performance/ksrkmdhj",
   "tg_tokens": null,
   "tps": 52.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-67y3sbel-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 8hl1yl80; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 834.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 834.5 19.1 28.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/67y3sbel",
   "source_url": "https://omlx.ai/benchmarks/performance/67y3sbel",
   "tg_tokens": null,
   "tps": 19.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-67y3sbel-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: 8hl1yl80; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 948.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 948.4 18.9 29.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/67y3sbel",
   "source_url": "https://omlx.ai/benchmarks/performance/67y3sbel",
   "tg_tokens": null,
   "tps": 18.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-arfekec8-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: u9rjbdmj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 932.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 932.5 34.0 16.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/arfekec8",
   "source_url": "https://omlx.ai/benchmarks/performance/arfekec8",
   "tg_tokens": null,
   "tps": 34.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-arfekec8-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: u9rjbdmj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1016.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,016 33.4 17.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/arfekec8",
   "source_url": "https://omlx.ai/benchmarks/performance/arfekec8",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-16u6xp1b-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.0 GB at this context; duplicate traces of the same session: 2wutvhdd, c57kdp9h; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 320.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 320.3 57.6 39.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/16u6xp1b",
   "source_url": "https://omlx.ai/benchmarks/performance/16u6xp1b",
   "tg_tokens": null,
   "tps": 57.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-16u6xp1b-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 42.3 GB at this context; duplicate traces of the same session: 2wutvhdd, c57kdp9h; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 411.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 411.5 50.6 42.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/16u6xp1b",
   "source_url": "https://omlx.ai/benchmarks/performance/16u6xp1b",
   "tg_tokens": null,
   "tps": 50.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-16u6xp1b-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 42.8 GB at this context; duplicate traces of the same session: 2wutvhdd, c57kdp9h; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 413.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 413.6 55.3 42.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/16u6xp1b",
   "source_url": "https://omlx.ai/benchmarks/performance/16u6xp1b",
   "tg_tokens": null,
   "tps": 55.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-6nbrzvra-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: s8e77fln, sut0nme0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 318.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 318.5 56.6 33.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6nbrzvra",
   "source_url": "https://omlx.ai/benchmarks/performance/6nbrzvra",
   "tg_tokens": null,
   "tps": 56.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-6nbrzvra-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: s8e77fln, sut0nme0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 331.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 331.6 53.1 37.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6nbrzvra",
   "source_url": "https://omlx.ai/benchmarks/performance/6nbrzvra",
   "tg_tokens": null,
   "tps": 53.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-6nbrzvra-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.8 GB at this context; duplicate traces of the same session: s8e77fln, sut0nme0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 326.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 326.8 52.6 37.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6nbrzvra",
   "source_url": "https://omlx.ai/benchmarks/performance/6nbrzvra",
   "tg_tokens": null,
   "tps": 52.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5cevmpng-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.2 GB at this context; duplicate traces of the same session: 640omggm, r4eco78d; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 337.2,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 337.2 40.4 15.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5cevmpng",
   "source_url": "https://omlx.ai/benchmarks/performance/5cevmpng",
   "tg_tokens": null,
   "tps": 40.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5cevmpng-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 640omggm, r4eco78d; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 345.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 345.0 39.0 18.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5cevmpng",
   "source_url": "https://omlx.ai/benchmarks/performance/5cevmpng",
   "tg_tokens": null,
   "tps": 39.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5cevmpng-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: 640omggm, r4eco78d; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 339.1,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 339.1 38.1 18.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5cevmpng",
   "source_url": "https://omlx.ai/benchmarks/performance/5cevmpng",
   "tg_tokens": null,
   "tps": 38.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-s0o0q8zx-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP320",
   "notes": "MTP Vision build; oMLX community benchmark trace, single run at 1k context; peak memory 30.3 GB; TTFT 9340 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 109.6,
   "provenance": "community",
   "quant": "oQ3e-DWQ",
   "quote": "Qwen3.8-Flash-Next-REAP320-oQ3e-DWQ-MTP-Vision-MLX on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 3bit; 1k tokens 109.6 PP tok/s 13.5 TG tok/s 9340 TTFT (ms) 30.3 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s0o0q8zx",
   "source_url": "https://omlx.ai/benchmarks/performance/s0o0q8zx",
   "tg_tokens": null,
   "tps": 13.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-gvi3jwp8-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP320",
   "notes": "MTP Vision build; oMLX community benchmark trace, single run at 1k context; peak memory 7.2 GB; TTFT 9630 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.3,
   "provenance": "community",
   "quant": "oQ3e-DWQ",
   "quote": "Qwen3.8-Flash-Next-REAP320-oQ3e-DWQ-MTP-Vision-MLX on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 3bit; 1k tokens 106.3 PP tok/s 5.3 TG tok/s 9630 TTFT (ms) 7.2 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gvi3jwp8",
   "source_url": "https://omlx.ai/benchmarks/performance/gvi3jwp8",
   "tg_tokens": null,
   "tps": 5.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-q5wz9tez-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 160.7 GB at this context; duplicate traces of the same session: jh1qvnn0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 370.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 4k 370.4 18.9 160.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q5wz9tez",
   "source_url": "https://omlx.ai/benchmarks/performance/q5wz9tez",
   "tg_tokens": null,
   "tps": 18.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-q5wz9tez-ctx16384",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 152.8 GB at this context; duplicate traces of the same session: jh1qvnn0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 723.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 16k 723.9 20.1 152.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q5wz9tez",
   "source_url": "https://omlx.ai/benchmarks/performance/q5wz9tez",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-l59gy01p-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 160.4 GB; TTFT 3026 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 338.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 1k tokens 338.4 PP tok/s 15.3 TG tok/s 3026 TTFT (ms) 160.4 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l59gy01p",
   "source_url": "https://omlx.ai/benchmarks/performance/l59gy01p",
   "tg_tokens": null,
   "tps": 15.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-mzg8bq5t-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 104.3 GB at this context; duplicate traces of the same session: 4bac8hi7; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 868.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 4bit; 4k 868.5 54.6 104.3 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mzg8bq5t",
   "source_url": "https://omlx.ai/benchmarks/performance/mzg8bq5t",
   "tg_tokens": null,
   "tps": 54.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-mzg8bq5t-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 104.5 GB at this context; duplicate traces of the same session: 4bac8hi7; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 860.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 4bit; 8k 860.2 55.2 104.5 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mzg8bq5t",
   "source_url": "https://omlx.ai/benchmarks/performance/mzg8bq5t",
   "tg_tokens": null,
   "tps": 55.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nwdboz22-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 55.9 GB; TTFT 24230 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 676.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k tokens 676.2 PP tok/s 13.9 TG tok/s 24230 TTFT (ms) 55.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nwdboz22",
   "source_url": "https://omlx.ai/benchmarks/performance/nwdboz22",
   "tg_tokens": null,
   "tps": 13.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i5toarlo-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 53.0 GB at this context; duplicate traces of the same session: 9gm00one; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 135.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 135.1 14.6 53.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i5toarlo",
   "source_url": "https://omlx.ai/benchmarks/performance/i5toarlo",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i5toarlo-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 55.8 GB at this context; duplicate traces of the same session: 9gm00one; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 698.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 698.1 11.6 55.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i5toarlo",
   "source_url": "https://omlx.ai/benchmarks/performance/i5toarlo",
   "tg_tokens": null,
   "tps": 11.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-695e6r7q-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 1k context; peak memory 22.5 GB; TTFT 846 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1211.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k tokens 1,211 PP tok/s 49.0 TG tok/s 846 TTFT (ms) 22.5 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/695e6r7q",
   "source_url": "https://omlx.ai/benchmarks/performance/695e6r7q",
   "tg_tokens": null,
   "tps": 49.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-l7h9jb7m-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 1k context; peak memory 20.0 GB; TTFT 498 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2058.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k tokens 2,058 PP tok/s 56.4 TG tok/s 498 TTFT (ms) 20.0 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l7h9jb7m",
   "source_url": "https://omlx.ai/benchmarks/performance/l7h9jb7m",
   "tg_tokens": null,
   "tps": 56.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-18oojlme-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 1k context; peak memory 11.3 GB; TTFT 4595 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 222.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k tokens 222.9 PP tok/s 21.9 TG tok/s 4595 TTFT (ms) 11.3 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/18oojlme",
   "source_url": "https://omlx.ai/benchmarks/performance/18oojlme",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-fp75gvq7-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 1k context; peak memory 11.3 GB; TTFT 4665 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 219.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k tokens 219.5 PP tok/s 22.8 TG tok/s 4665 TTFT (ms) 11.3 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fp75gvq7",
   "source_url": "https://omlx.ai/benchmarks/performance/fp75gvq7",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-q601a6uu-ctx1024",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 154.3 GB at this context; duplicate traces of the same session: au8jcpoe, bb73wg3i, hn9co94y, l9isbbeg, uwa9nnyb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 427.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-V2.6-Flash-RL-MLX-4bit-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 427.9 55.5 154.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q601a6uu",
   "source_url": "https://omlx.ai/benchmarks/performance/q601a6uu",
   "tg_tokens": null,
   "tps": 55.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-q601a6uu-ctx4096",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 155.5 GB at this context; duplicate traces of the same session: au8jcpoe, bb73wg3i, hn9co94y, l9isbbeg, uwa9nnyb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 484.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-V2.6-Flash-RL-MLX-4bit-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 484.4 54.5 155.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q601a6uu",
   "source_url": "https://omlx.ai/benchmarks/performance/q601a6uu",
   "tg_tokens": null,
   "tps": 54.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-q601a6uu-ctx8192",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 156.6 GB at this context; duplicate traces of the same session: au8jcpoe, bb73wg3i, hn9co94y, l9isbbeg, uwa9nnyb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 473.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-V2.6-Flash-RL-MLX-4bit-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 473.1 52.8 156.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q601a6uu",
   "source_url": "https://omlx.ai/benchmarks/performance/q601a6uu",
   "tg_tokens": null,
   "tps": 52.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-q601a6uu-ctx16384",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 158.7 GB at this context; duplicate traces of the same session: au8jcpoe, bb73wg3i, hn9co94y, l9isbbeg, uwa9nnyb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 450.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-V2.6-Flash-RL-MLX-4bit-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 450.1 50.1 158.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q601a6uu",
   "source_url": "https://omlx.ai/benchmarks/performance/q601a6uu",
   "tg_tokens": null,
   "tps": 50.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-q601a6uu-ctx32768",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 163.1 GB at this context; duplicate traces of the same session: au8jcpoe, bb73wg3i, hn9co94y, l9isbbeg, uwa9nnyb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 402.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-V2.6-Flash-RL-MLX-4bit-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 402.8 44.2 163.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q601a6uu",
   "source_url": "https://omlx.ai/benchmarks/performance/q601a6uu",
   "tg_tokens": null,
   "tps": 44.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-q601a6uu-ctx65536",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 171.8 GB at this context; duplicate traces of the same session: au8jcpoe, bb73wg3i, hn9co94y, l9isbbeg, uwa9nnyb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 330.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-V2.6-Flash-RL-MLX-4bit-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 330.3 37.1 171.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q601a6uu",
   "source_url": "https://omlx.ai/benchmarks/performance/q601a6uu",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-m8bihd04-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 274.0,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Ornith-1.5-35B-A3B-oQ3e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 3bit; 1k 274.0 17.2 5.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m8bihd04",
   "source_url": "https://omlx.ai/benchmarks/performance/m8bihd04",
   "tg_tokens": null,
   "tps": 17.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-m8bihd04-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 503.8,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Ornith-1.5-35B-A3B-oQ3e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 3bit; 4k 503.8 17.9 6.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m8bihd04",
   "source_url": "https://omlx.ai/benchmarks/performance/m8bihd04",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-m8bihd04-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 500.5,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Ornith-1.5-35B-A3B-oQ3e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 3bit; 16k 500.5 15.0 7.0 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m8bihd04",
   "source_url": "https://omlx.ai/benchmarks/performance/m8bihd04",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-m8bihd04-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.3 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 484.1,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Ornith-1.5-35B-A3B-oQ3e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 3bit; 32k 484.1 15.3 7.3 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m8bihd04",
   "source_url": "https://omlx.ai/benchmarks/performance/m8bihd04",
   "tg_tokens": null,
   "tps": 15.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-m8bihd04-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.9 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 282.9,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Ornith-1.5-35B-A3B-oQ3e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 3bit; 64k 282.9 14.8 7.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m8bihd04",
   "source_url": "https://omlx.ai/benchmarks/performance/m8bihd04",
   "tg_tokens": null,
   "tps": 14.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-m8bihd04-ctx131072",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 9.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 349.1,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Ornith-1.5-35B-A3B-oQ3e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 3bit; 128k 349.1 12.3 9.4 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m8bihd04",
   "source_url": "https://omlx.ai/benchmarks/performance/m8bihd04",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-yuv7wnru-ctx1024",
   "model": "Qwen3-Coder-Next-REAP-48B-A3B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 27.0 GB at this context; duplicate traces of the same session: 0tbf3hly, 3ljbfpq1, 6aotiugi, 7v9j9pe8, c2ta95k2, fz5g0sw0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 696.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next-REAP-48B-A3B-mlx-4Bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 696.9 56.8 27.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yuv7wnru",
   "source_url": "https://omlx.ai/benchmarks/performance/yuv7wnru",
   "tg_tokens": null,
   "tps": 56.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-yuv7wnru-ctx4096",
   "model": "Qwen3-Coder-Next-REAP-48B-A3B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: 0tbf3hly, 3ljbfpq1, 6aotiugi, 7v9j9pe8, c2ta95k2, fz5g0sw0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 764.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next-REAP-48B-A3B-mlx-4Bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 764.5 55.0 27.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yuv7wnru",
   "source_url": "https://omlx.ai/benchmarks/performance/yuv7wnru",
   "tg_tokens": null,
   "tps": 55.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-yuv7wnru-ctx8192",
   "model": "Qwen3-Coder-Next-REAP-48B-A3B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: 0tbf3hly, 3ljbfpq1, 6aotiugi, 7v9j9pe8, c2ta95k2, fz5g0sw0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 722.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next-REAP-48B-A3B-mlx-4Bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 722.0 51.9 27.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yuv7wnru",
   "source_url": "https://omlx.ai/benchmarks/performance/yuv7wnru",
   "tg_tokens": null,
   "tps": 51.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-yuv7wnru-ctx16384",
   "model": "Qwen3-Coder-Next-REAP-48B-A3B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: 0tbf3hly, 3ljbfpq1, 6aotiugi, 7v9j9pe8, c2ta95k2, fz5g0sw0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 655.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next-REAP-48B-A3B-mlx-4Bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 655.9 50.5 28.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yuv7wnru",
   "source_url": "https://omlx.ai/benchmarks/performance/yuv7wnru",
   "tg_tokens": null,
   "tps": 50.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-yuv7wnru-ctx32768",
   "model": "Qwen3-Coder-Next-REAP-48B-A3B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: 0tbf3hly, 3ljbfpq1, 6aotiugi, 7v9j9pe8, c2ta95k2, fz5g0sw0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 538.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next-REAP-48B-A3B-mlx-4Bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 538.6 43.2 28.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yuv7wnru",
   "source_url": "https://omlx.ai/benchmarks/performance/yuv7wnru",
   "tg_tokens": null,
   "tps": 43.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-yuv7wnru-ctx65536",
   "model": "Qwen3-Coder-Next-REAP-48B-A3B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: 0tbf3hly, 3ljbfpq1, 6aotiugi, 7v9j9pe8, c2ta95k2, fz5g0sw0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 414.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next-REAP-48B-A3B-mlx-4Bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 414.3 32.1 29.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yuv7wnru",
   "source_url": "https://omlx.ai/benchmarks/performance/yuv7wnru",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-yuv7wnru-ctx131072",
   "model": "Qwen3-Coder-Next-REAP-48B-A3B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 31.7 GB at this context; duplicate traces of the same session: 0tbf3hly, 3ljbfpq1, 6aotiugi, 7v9j9pe8, c2ta95k2, fz5g0sw0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 267.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next-REAP-48B-A3B-mlx-4Bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 128k 267.1 21.9 31.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yuv7wnru",
   "source_url": "https://omlx.ai/benchmarks/performance/yuv7wnru",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-min16qd2-ctx1024",
   "model": "Qwen3-Coder-Next-REAP-48B-A3B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 27.0 GB at this context; duplicate traces of the same session: cw0qz3rs; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 662.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next-REAP-48B-A3B-mlx-4Bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 662.1 58.3 27.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/min16qd2",
   "source_url": "https://omlx.ai/benchmarks/performance/min16qd2",
   "tg_tokens": null,
   "tps": 58.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-min16qd2-ctx4096",
   "model": "Qwen3-Coder-Next-REAP-48B-A3B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: cw0qz3rs; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 767.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next-REAP-48B-A3B-mlx-4Bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 767.5 56.5 27.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/min16qd2",
   "source_url": "https://omlx.ai/benchmarks/performance/min16qd2",
   "tg_tokens": null,
   "tps": 56.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-likr5pbd-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.0 GB at this context; duplicate traces of the same session: 7hrqcwqh; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 340.3,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Qwen3.5-9B-MLX-MXFP4 on M5 (10c); M5 (10c) \u00b7 16 GB \u00b7 4bit; 4k 340.3 14.8 6.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/likr5pbd",
   "source_url": "https://omlx.ai/benchmarks/performance/likr5pbd",
   "tg_tokens": null,
   "tps": 14.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-likr5pbd-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: 7hrqcwqh; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 327.8,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Qwen3.5-9B-MLX-MXFP4 on M5 (10c); M5 (10c) \u00b7 16 GB \u00b7 4bit; 16k 327.8 13.7 6.3 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/likr5pbd",
   "source_url": "https://omlx.ai/benchmarks/performance/likr5pbd",
   "tg_tokens": null,
   "tps": 13.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-sxygzckr-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: 8v1cjpni, 93djy0ei, eoibftvo, sjq964kv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4128.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 4,128 102.6 35.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sxygzckr",
   "source_url": "https://omlx.ai/benchmarks/performance/sxygzckr",
   "tg_tokens": null,
   "tps": 102.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-sxygzckr-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 36.7 GB at this context; duplicate traces of the same session: 8v1cjpni, 93djy0ei, eoibftvo, sjq964kv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4660.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 32k 4,660 89.1 36.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sxygzckr",
   "source_url": "https://omlx.ai/benchmarks/performance/sxygzckr",
   "tg_tokens": null,
   "tps": 89.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-sxygzckr-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: 8v1cjpni, 93djy0ei, eoibftvo, sjq964kv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3660.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 64k 3,660 77.1 37.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sxygzckr",
   "source_url": "https://omlx.ai/benchmarks/performance/sxygzckr",
   "tg_tokens": null,
   "tps": 77.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-sxygzckr-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 38.6 GB at this context; duplicate traces of the same session: 8v1cjpni, 93djy0ei, eoibftvo, sjq964kv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2522.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 128k 2,522 62.9 38.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sxygzckr",
   "source_url": "https://omlx.ai/benchmarks/performance/sxygzckr",
   "tg_tokens": null,
   "tps": 62.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-sxygzckr-ctx199680",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 39.9 GB at this context; duplicate traces of the same session: 8v1cjpni, 93djy0ei, eoibftvo, sjq964kv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1874.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 195k 1,874 52.9 39.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sxygzckr",
   "source_url": "https://omlx.ai/benchmarks/performance/sxygzckr",
   "tg_tokens": null,
   "tps": 52.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fjv62ow7-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: r8x33n2s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 877.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 877.7 35.3 32.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fjv62ow7",
   "source_url": "https://omlx.ai/benchmarks/performance/fjv62ow7",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fjv62ow7-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.7 GB at this context; duplicate traces of the same session: r8x33n2s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 951.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 951.0 35.9 33.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fjv62ow7",
   "source_url": "https://omlx.ai/benchmarks/performance/fjv62ow7",
   "tg_tokens": null,
   "tps": 35.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0k93sp3p-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: kc7oj15j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 895.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 895.6 18.8 29.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0k93sp3p",
   "source_url": "https://omlx.ai/benchmarks/performance/0k93sp3p",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0k93sp3p-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: kc7oj15j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 933.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 16k 933.0 18.1 30.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0k93sp3p",
   "source_url": "https://omlx.ai/benchmarks/performance/0k93sp3p",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dhenqwgy-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: 0cb7ow72; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 970.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 970.0 51.6 19.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dhenqwgy",
   "source_url": "https://omlx.ai/benchmarks/performance/dhenqwgy",
   "tg_tokens": null,
   "tps": 51.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dhenqwgy-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 0cb7ow72; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1008.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,008 55.4 21.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dhenqwgy",
   "source_url": "https://omlx.ai/benchmarks/performance/dhenqwgy",
   "tg_tokens": null,
   "tps": 55.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-08lgdmf9-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.4 GB at this context; duplicate traces of the same session: 0v67es6g, 1d08a1z5, 3qt12e68, awwa5n18, mkv7prvl, uhgah2us; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2623.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 2,623 155.9 100.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/08lgdmf9",
   "source_url": "https://omlx.ai/benchmarks/performance/08lgdmf9",
   "tg_tokens": null,
   "tps": 155.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-08lgdmf9-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.0 GB at this context; duplicate traces of the same session: 0v67es6g, 1d08a1z5, 3qt12e68, awwa5n18, mkv7prvl, uhgah2us; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4290.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 4,290 132.9 101.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/08lgdmf9",
   "source_url": "https://omlx.ai/benchmarks/performance/08lgdmf9",
   "tg_tokens": null,
   "tps": 132.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-08lgdmf9-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.2 GB at this context; duplicate traces of the same session: 0v67es6g, 1d08a1z5, 3qt12e68, awwa5n18, mkv7prvl, uhgah2us; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5870.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 5,870 149.2 102.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/08lgdmf9",
   "source_url": "https://omlx.ai/benchmarks/performance/08lgdmf9",
   "tg_tokens": null,
   "tps": 149.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-08lgdmf9-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.2 GB at this context; duplicate traces of the same session: 0v67es6g, 1d08a1z5, 3qt12e68, awwa5n18, mkv7prvl, uhgah2us; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5880.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 5,880 125.5 103.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/08lgdmf9",
   "source_url": "https://omlx.ai/benchmarks/performance/08lgdmf9",
   "tg_tokens": null,
   "tps": 125.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-08lgdmf9-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.3 GB at this context; duplicate traces of the same session: 0v67es6g, 1d08a1z5, 3qt12e68, awwa5n18, mkv7prvl, uhgah2us; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5790.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 5,790 127.5 104.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/08lgdmf9",
   "source_url": "https://omlx.ai/benchmarks/performance/08lgdmf9",
   "tg_tokens": null,
   "tps": 127.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-08lgdmf9-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 119.1 GB at this context; duplicate traces of the same session: 0v67es6g, 1d08a1z5, 3qt12e68, awwa5n18, mkv7prvl, uhgah2us; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5595.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,595 30.7 119.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/08lgdmf9",
   "source_url": "https://omlx.ai/benchmarks/performance/08lgdmf9",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-08lgdmf9-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 130.6 GB at this context; duplicate traces of the same session: 0v67es6g, 1d08a1z5, 3qt12e68, awwa5n18, mkv7prvl, uhgah2us; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2893.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 2,893 24.6 130.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/08lgdmf9",
   "source_url": "https://omlx.ai/benchmarks/performance/08lgdmf9",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-bz056fx9-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 100.4 GB; TTFT 373 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2743.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k tokens 2,743 PP tok/s 122.1 TG tok/s 373 TTFT (ms) 100.4 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bz056fx9",
   "source_url": "https://omlx.ai/benchmarks/performance/bz056fx9",
   "tg_tokens": null,
   "tps": 122.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ne02fx87-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: 975ru7k1, pem2nui6, xdq1ji6c, yml9lp0d; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3076.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 3,076 127.1 100.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ne02fx87",
   "source_url": "https://omlx.ai/benchmarks/performance/ne02fx87",
   "tg_tokens": null,
   "tps": 127.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ne02fx87-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.0 GB at this context; duplicate traces of the same session: 975ru7k1, pem2nui6, xdq1ji6c, yml9lp0d; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5123.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 5,123 104.2 101.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ne02fx87",
   "source_url": "https://omlx.ai/benchmarks/performance/ne02fx87",
   "tg_tokens": null,
   "tps": 104.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ne02fx87-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.3 GB at this context; duplicate traces of the same session: 975ru7k1, pem2nui6, xdq1ji6c, yml9lp0d; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5547.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 5,547 126.4 104.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ne02fx87",
   "source_url": "https://omlx.ai/benchmarks/performance/ne02fx87",
   "tg_tokens": null,
   "tps": 126.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ne02fx87-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 108.1 GB at this context; duplicate traces of the same session: 975ru7k1, pem2nui6, xdq1ji6c, yml9lp0d; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5301.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,301 109.5 108.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ne02fx87",
   "source_url": "https://omlx.ai/benchmarks/performance/ne02fx87",
   "tg_tokens": null,
   "tps": 109.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ne02fx87-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.7 GB at this context; duplicate traces of the same session: 975ru7k1, pem2nui6, xdq1ji6c, yml9lp0d; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5139.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 5,139 89.4 110.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ne02fx87",
   "source_url": "https://omlx.ai/benchmarks/performance/ne02fx87",
   "tg_tokens": null,
   "tps": 89.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-nqyen2wa-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: q90x1awk, sd6w6rwh, xfoi65ry; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3005.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 3,005 153.1 100.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nqyen2wa",
   "source_url": "https://omlx.ai/benchmarks/performance/nqyen2wa",
   "tg_tokens": null,
   "tps": 153.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-nqyen2wa-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.1 GB at this context; duplicate traces of the same session: q90x1awk, sd6w6rwh, xfoi65ry; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5658.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 5,658 157.0 102.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nqyen2wa",
   "source_url": "https://omlx.ai/benchmarks/performance/nqyen2wa",
   "tg_tokens": null,
   "tps": 157.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-nqyen2wa-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.3 GB at this context; duplicate traces of the same session: q90x1awk, sd6w6rwh, xfoi65ry; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4846.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 4,846 125.1 104.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nqyen2wa",
   "source_url": "https://omlx.ai/benchmarks/performance/nqyen2wa",
   "tg_tokens": null,
   "tps": 125.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-nqyen2wa-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.3 GB at this context; duplicate traces of the same session: q90x1awk, sd6w6rwh, xfoi65ry; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4760.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 4,760 88.2 110.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nqyen2wa",
   "source_url": "https://omlx.ai/benchmarks/performance/nqyen2wa",
   "tg_tokens": null,
   "tps": 88.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-7o9iblup-ctx1024",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: 3xmpc4qf, 9avz7vpj, krm4c5c7, mpsg3dfq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1336.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 1,336 75.9 25.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7o9iblup",
   "source_url": "https://omlx.ai/benchmarks/performance/7o9iblup",
   "tg_tokens": null,
   "tps": 75.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-7o9iblup-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: 3xmpc4qf, 9avz7vpj, krm4c5c7, mpsg3dfq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1338.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 1,338 74.7 26.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7o9iblup",
   "source_url": "https://omlx.ai/benchmarks/performance/7o9iblup",
   "tg_tokens": null,
   "tps": 74.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-7o9iblup-ctx8192",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: 3xmpc4qf, 9avz7vpj, krm4c5c7, mpsg3dfq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1342.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 1,342 72.0 26.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7o9iblup",
   "source_url": "https://omlx.ai/benchmarks/performance/7o9iblup",
   "tg_tokens": null,
   "tps": 72.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-7o9iblup-ctx16384",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: 3xmpc4qf, 9avz7vpj, krm4c5c7, mpsg3dfq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1210.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 1,210 67.0 26.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7o9iblup",
   "source_url": "https://omlx.ai/benchmarks/performance/7o9iblup",
   "tg_tokens": null,
   "tps": 67.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-7o9iblup-ctx32768",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: 3xmpc4qf, 9avz7vpj, krm4c5c7, mpsg3dfq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1018.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 1,018 60.0 27.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7o9iblup",
   "source_url": "https://omlx.ai/benchmarks/performance/7o9iblup",
   "tg_tokens": null,
   "tps": 60.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-7o9iblup-ctx65536",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 29.1 GB at this context; duplicate traces of the same session: 3xmpc4qf, 9avz7vpj, krm4c5c7, mpsg3dfq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 804.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 804.2 45.7 29.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7o9iblup",
   "source_url": "https://omlx.ai/benchmarks/performance/7o9iblup",
   "tg_tokens": null,
   "tps": 45.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-7o9iblup-ctx131072",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 32.4 GB at this context; duplicate traces of the same session: 3xmpc4qf, 9avz7vpj, krm4c5c7, mpsg3dfq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 477.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 477.1 13.9 32.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7o9iblup",
   "source_url": "https://omlx.ai/benchmarks/performance/7o9iblup",
   "tg_tokens": null,
   "tps": 13.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-7o9iblup-ctx199680",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 35.9 GB at this context; duplicate traces of the same session: 3xmpc4qf, 9avz7vpj, krm4c5c7, mpsg3dfq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 288.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 195k 288.3 23.5 35.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7o9iblup",
   "source_url": "https://omlx.ai/benchmarks/performance/7o9iblup",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-e66joshc-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Nail build; oMLX community benchmark trace; peak memory 30.7 GB at this context; duplicate traces of the same session: ikcre03r; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 567.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX-6bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 1k 567.3 51.7 30.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e66joshc",
   "source_url": "https://omlx.ai/benchmarks/performance/e66joshc",
   "tg_tokens": null,
   "tps": 51.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-e66joshc-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Nail build; oMLX community benchmark trace; peak memory 32.5 GB at this context; duplicate traces of the same session: ikcre03r; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 583.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX-6bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 4k 583.3 50.0 32.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e66joshc",
   "source_url": "https://omlx.ai/benchmarks/performance/e66joshc",
   "tg_tokens": null,
   "tps": 50.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-hfmvfjiq-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 7.1 GB at this context; duplicate traces of the same session: 6ejsguba, 6rqn6t0s, 9wefr8lb, o9hafmwa, u5e76ftj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 259.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 259.4 16.4 7.1 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hfmvfjiq",
   "source_url": "https://omlx.ai/benchmarks/performance/hfmvfjiq",
   "tg_tokens": null,
   "tps": 16.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-hfmvfjiq-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 7.9 GB at this context; duplicate traces of the same session: 6ejsguba, 6rqn6t0s, 9wefr8lb, o9hafmwa, u5e76ftj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 486.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 486.3 16.5 7.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hfmvfjiq",
   "source_url": "https://omlx.ai/benchmarks/performance/hfmvfjiq",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-hfmvfjiq-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 8.2 GB at this context; duplicate traces of the same session: 6ejsguba, 6rqn6t0s, 9wefr8lb, o9hafmwa, u5e76ftj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 464.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 16k 464.2 14.5 8.2 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hfmvfjiq",
   "source_url": "https://omlx.ai/benchmarks/performance/hfmvfjiq",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-hfmvfjiq-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 8.5 GB at this context; duplicate traces of the same session: 6ejsguba, 6rqn6t0s, 9wefr8lb, o9hafmwa, u5e76ftj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 454.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 32k 454.3 14.4 8.5 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hfmvfjiq",
   "source_url": "https://omlx.ai/benchmarks/performance/hfmvfjiq",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-hfmvfjiq-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 9.1 GB at this context; duplicate traces of the same session: 6ejsguba, 6rqn6t0s, 9wefr8lb, o9hafmwa, u5e76ftj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 410.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 64k 410.6 13.7 9.1 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hfmvfjiq",
   "source_url": "https://omlx.ai/benchmarks/performance/hfmvfjiq",
   "tg_tokens": null,
   "tps": 13.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-hfmvfjiq-ctx131072",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 10.6 GB at this context; duplicate traces of the same session: 6ejsguba, 6rqn6t0s, 9wefr8lb, o9hafmwa, u5e76ftj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 336.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 128k 336.4 10.8 10.6 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hfmvfjiq",
   "source_url": "https://omlx.ai/benchmarks/performance/hfmvfjiq",
   "tg_tokens": null,
   "tps": 10.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-v6izgop4-ctx1024",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: ssrtmajh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 219.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-27B-MLX-8bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 219.9 12.9 28.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v6izgop4",
   "source_url": "https://omlx.ai/benchmarks/performance/v6izgop4",
   "tg_tokens": null,
   "tps": 12.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-v6izgop4-ctx4096",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.7 GB at this context; duplicate traces of the same session: ssrtmajh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 225.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-27B-MLX-8bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 225.3 12.8 31.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v6izgop4",
   "source_url": "https://omlx.ai/benchmarks/performance/v6izgop4",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-t861ir8u-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 69.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 757.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 757.0 79.6 69.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t861ir8u",
   "source_url": "https://omlx.ai/benchmarks/performance/t861ir8u",
   "tg_tokens": null,
   "tps": 79.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-t861ir8u-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 70.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1364.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,364 69.6 70.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t861ir8u",
   "source_url": "https://omlx.ai/benchmarks/performance/t861ir8u",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-zwv8k7wi-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: 28ejwkir, a3tzjs1s, g0g4p3fg, gbji5ihz, liqptd3h, vnh8kh8f; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2820.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 2,820 151.6 100.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zwv8k7wi",
   "source_url": "https://omlx.ai/benchmarks/performance/zwv8k7wi",
   "tg_tokens": null,
   "tps": 151.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-zwv8k7wi-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.0 GB at this context; duplicate traces of the same session: 28ejwkir, a3tzjs1s, g0g4p3fg, gbji5ihz, liqptd3h, vnh8kh8f; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5081.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 5,081 157.7 101.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zwv8k7wi",
   "source_url": "https://omlx.ai/benchmarks/performance/zwv8k7wi",
   "tg_tokens": null,
   "tps": 157.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-zwv8k7wi-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.1 GB at this context; duplicate traces of the same session: 28ejwkir, a3tzjs1s, g0g4p3fg, gbji5ihz, liqptd3h, vnh8kh8f; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5663.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 5,663 149.1 102.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zwv8k7wi",
   "source_url": "https://omlx.ai/benchmarks/performance/zwv8k7wi",
   "tg_tokens": null,
   "tps": 149.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-zwv8k7wi-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.1 GB at this context; duplicate traces of the same session: 28ejwkir, a3tzjs1s, g0g4p3fg, gbji5ihz, liqptd3h, vnh8kh8f; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5647.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 5,647 158.9 103.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zwv8k7wi",
   "source_url": "https://omlx.ai/benchmarks/performance/zwv8k7wi",
   "tg_tokens": null,
   "tps": 158.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-zwv8k7wi-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.4 GB at this context; duplicate traces of the same session: 28ejwkir, a3tzjs1s, g0g4p3fg, gbji5ihz, liqptd3h, vnh8kh8f; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5515.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 5,515 142.7 104.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zwv8k7wi",
   "source_url": "https://omlx.ai/benchmarks/performance/zwv8k7wi",
   "tg_tokens": null,
   "tps": 142.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-zwv8k7wi-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 108.8 GB at this context; duplicate traces of the same session: 28ejwkir, a3tzjs1s, g0g4p3fg, gbji5ihz, liqptd3h, vnh8kh8f; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5319.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,319 109.1 108.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zwv8k7wi",
   "source_url": "https://omlx.ai/benchmarks/performance/zwv8k7wi",
   "tg_tokens": null,
   "tps": 109.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-zwv8k7wi-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.7 GB at this context; duplicate traces of the same session: 28ejwkir, a3tzjs1s, g0g4p3fg, gbji5ihz, liqptd3h, vnh8kh8f; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5144.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 5,144 96.6 110.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zwv8k7wi",
   "source_url": "https://omlx.ai/benchmarks/performance/zwv8k7wi",
   "tg_tokens": null,
   "tps": 96.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-xf1326m9-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: huxz7vmt, ox7thwsp, rrd0j5v8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2850.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 2,850 134.3 100.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xf1326m9",
   "source_url": "https://omlx.ai/benchmarks/performance/xf1326m9",
   "tg_tokens": null,
   "tps": 134.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-xf1326m9-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.0 GB at this context; duplicate traces of the same session: huxz7vmt, ox7thwsp, rrd0j5v8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4892.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 4,892 137.1 101.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xf1326m9",
   "source_url": "https://omlx.ai/benchmarks/performance/xf1326m9",
   "tg_tokens": null,
   "tps": 137.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-xf1326m9-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.1 GB at this context; duplicate traces of the same session: huxz7vmt, ox7thwsp, rrd0j5v8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5644.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 5,644 156.6 102.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xf1326m9",
   "source_url": "https://omlx.ai/benchmarks/performance/xf1326m9",
   "tg_tokens": null,
   "tps": 156.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-xf1326m9-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.4 GB at this context; duplicate traces of the same session: huxz7vmt, ox7thwsp, rrd0j5v8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5529.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 5,529 135.1 104.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xf1326m9",
   "source_url": "https://omlx.ai/benchmarks/performance/xf1326m9",
   "tg_tokens": null,
   "tps": 135.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-thvplmg2-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 100.7 GB; TTFT 379 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2700.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k tokens 2,700 PP tok/s 119.3 TG tok/s 379 TTFT (ms) 100.7 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/thvplmg2",
   "source_url": "https://omlx.ai/benchmarks/performance/thvplmg2",
   "tg_tokens": null,
   "tps": 119.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-x20k2n2x-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 100.8 GB; TTFT 369 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2775.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k tokens 2,775 PP tok/s 63.5 TG tok/s 369 TTFT (ms) 100.8 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x20k2n2x",
   "source_url": "https://omlx.ai/benchmarks/performance/x20k2n2x",
   "tg_tokens": null,
   "tps": 63.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-spr1clds-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 100.7 GB; TTFT 392 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2613.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k tokens 2,613 PP tok/s 113.5 TG tok/s 392 TTFT (ms) 100.7 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/spr1clds",
   "source_url": "https://omlx.ai/benchmarks/performance/spr1clds",
   "tg_tokens": null,
   "tps": 113.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ie0461z1-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.4 GB at this context; duplicate traces of the same session: 8mxjff5s, buw6i88y, c00k9jse, i2m3zo93, k1ole3sw, l3spsny1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2873.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 2,873 131.9 100.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ie0461z1",
   "source_url": "https://omlx.ai/benchmarks/performance/ie0461z1",
   "tg_tokens": null,
   "tps": 131.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ie0461z1-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.0 GB at this context; duplicate traces of the same session: 8mxjff5s, buw6i88y, c00k9jse, i2m3zo93, k1ole3sw, l3spsny1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4881.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 4,881 150.1 101.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ie0461z1",
   "source_url": "https://omlx.ai/benchmarks/performance/ie0461z1",
   "tg_tokens": null,
   "tps": 150.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ie0461z1-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.1 GB at this context; duplicate traces of the same session: 8mxjff5s, buw6i88y, c00k9jse, i2m3zo93, k1ole3sw, l3spsny1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5636.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 5,636 136.3 102.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ie0461z1",
   "source_url": "https://omlx.ai/benchmarks/performance/ie0461z1",
   "tg_tokens": null,
   "tps": 136.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ie0461z1-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.1 GB at this context; duplicate traces of the same session: 8mxjff5s, buw6i88y, c00k9jse, i2m3zo93, k1ole3sw, l3spsny1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5644.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 5,644 123.1 103.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ie0461z1",
   "source_url": "https://omlx.ai/benchmarks/performance/ie0461z1",
   "tg_tokens": null,
   "tps": 123.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ie0461z1-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.3 GB at this context; duplicate traces of the same session: 8mxjff5s, buw6i88y, c00k9jse, i2m3zo93, k1ole3sw, l3spsny1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5510.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 5,510 129.6 104.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ie0461z1",
   "source_url": "https://omlx.ai/benchmarks/performance/ie0461z1",
   "tg_tokens": null,
   "tps": 129.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ie0461z1-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 108.1 GB at this context; duplicate traces of the same session: 8mxjff5s, buw6i88y, c00k9jse, i2m3zo93, k1ole3sw, l3spsny1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5344.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,344 105.6 108.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ie0461z1",
   "source_url": "https://omlx.ai/benchmarks/performance/ie0461z1",
   "tg_tokens": null,
   "tps": 105.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ie0461z1-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.7 GB at this context; duplicate traces of the same session: 8mxjff5s, buw6i88y, c00k9jse, i2m3zo93, k1ole3sw, l3spsny1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5218.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 5,218 92.0 110.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ie0461z1",
   "source_url": "https://omlx.ai/benchmarks/performance/ie0461z1",
   "tg_tokens": null,
   "tps": 92.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-dli7y8p6-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: cr14s5uw, gu2j49ui, k7s7qk4u; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2933.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 2,933 144.6 100.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dli7y8p6",
   "source_url": "https://omlx.ai/benchmarks/performance/dli7y8p6",
   "tg_tokens": null,
   "tps": 144.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-dli7y8p6-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.0 GB at this context; duplicate traces of the same session: cr14s5uw, gu2j49ui, k7s7qk4u; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4958.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 4,958 140.8 101.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dli7y8p6",
   "source_url": "https://omlx.ai/benchmarks/performance/dli7y8p6",
   "tg_tokens": null,
   "tps": 140.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-dli7y8p6-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.1 GB at this context; duplicate traces of the same session: cr14s5uw, gu2j49ui, k7s7qk4u; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5658.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 5,658 159.1 102.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dli7y8p6",
   "source_url": "https://omlx.ai/benchmarks/performance/dli7y8p6",
   "tg_tokens": null,
   "tps": 159.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-dli7y8p6-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.5 GB at this context; duplicate traces of the same session: cr14s5uw, gu2j49ui, k7s7qk4u; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5654.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 5,654 177.2 102.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dli7y8p6",
   "source_url": "https://omlx.ai/benchmarks/performance/dli7y8p6",
   "tg_tokens": null,
   "tps": 177.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-p9gcn2aa-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: 0d8mgoyj, isschd1d, ky9owg8k, syz3l4g3, w03vwc7d; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2890.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 2,890 144.1 100.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p9gcn2aa",
   "source_url": "https://omlx.ai/benchmarks/performance/p9gcn2aa",
   "tg_tokens": null,
   "tps": 144.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-p9gcn2aa-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.0 GB at this context; duplicate traces of the same session: 0d8mgoyj, isschd1d, ky9owg8k, syz3l4g3, w03vwc7d; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4905.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 4,905 132.7 101.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p9gcn2aa",
   "source_url": "https://omlx.ai/benchmarks/performance/p9gcn2aa",
   "tg_tokens": null,
   "tps": 132.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-p9gcn2aa-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.1 GB at this context; duplicate traces of the same session: 0d8mgoyj, isschd1d, ky9owg8k, syz3l4g3, w03vwc7d; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5627.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 5,627 147.7 102.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p9gcn2aa",
   "source_url": "https://omlx.ai/benchmarks/performance/p9gcn2aa",
   "tg_tokens": null,
   "tps": 147.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-p9gcn2aa-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.3 GB at this context; duplicate traces of the same session: 0d8mgoyj, isschd1d, ky9owg8k, syz3l4g3, w03vwc7d; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5553.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 5,553 127.5 104.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p9gcn2aa",
   "source_url": "https://omlx.ai/benchmarks/performance/p9gcn2aa",
   "tg_tokens": null,
   "tps": 127.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-p9gcn2aa-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 108.1 GB at this context; duplicate traces of the same session: 0d8mgoyj, isschd1d, ky9owg8k, syz3l4g3, w03vwc7d; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5379.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,379 106.2 108.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p9gcn2aa",
   "source_url": "https://omlx.ai/benchmarks/performance/p9gcn2aa",
   "tg_tokens": null,
   "tps": 106.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-p9gcn2aa-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.7 GB at this context; duplicate traces of the same session: 0d8mgoyj, isschd1d, ky9owg8k, syz3l4g3, w03vwc7d; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5226.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 5,226 90.3 110.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p9gcn2aa",
   "source_url": "https://omlx.ai/benchmarks/performance/p9gcn2aa",
   "tg_tokens": null,
   "tps": 90.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-azb221f4-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 100.6 GB; TTFT 342 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2997.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k tokens 2,997 PP tok/s 145.5 TG tok/s 342 TTFT (ms) 100.6 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/azb221f4",
   "source_url": "https://omlx.ai/benchmarks/performance/azb221f4",
   "tg_tokens": null,
   "tps": 145.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-u1ik20tn-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 137.6 GB at this context; duplicate traces of the same session: tn4wg69p; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 972.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 128k 972.7 56.9 137.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u1ik20tn",
   "source_url": "https://omlx.ai/benchmarks/performance/u1ik20tn",
   "tg_tokens": null,
   "tps": 56.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-u1ik20tn-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 138.6 GB at this context; duplicate traces of the same session: tn4wg69p; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 984.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 195k 984.7 52.6 138.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u1ik20tn",
   "source_url": "https://omlx.ai/benchmarks/performance/u1ik20tn",
   "tg_tokens": null,
   "tps": 52.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-pklsr7op-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 195k context; peak memory 138.5 GB; TTFT 206018 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 970.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 195k tokens 970.8 PP tok/s 46.2 TG tok/s 206018 TTFT (ms) 138.5 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pklsr7op",
   "source_url": "https://omlx.ai/benchmarks/performance/pklsr7op",
   "tg_tokens": null,
   "tps": 46.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-i06jcfln-ctx1024",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: d3kfammq, sa9tch22; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1336.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 1,336 75.9 25.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i06jcfln",
   "source_url": "https://omlx.ai/benchmarks/performance/i06jcfln",
   "tg_tokens": null,
   "tps": 75.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-i06jcfln-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: d3kfammq, sa9tch22; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1338.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 1,338 74.7 26.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i06jcfln",
   "source_url": "https://omlx.ai/benchmarks/performance/i06jcfln",
   "tg_tokens": null,
   "tps": 74.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-i06jcfln-ctx8192",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: d3kfammq, sa9tch22; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1342.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 1,342 72.0 26.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i06jcfln",
   "source_url": "https://omlx.ai/benchmarks/performance/i06jcfln",
   "tg_tokens": null,
   "tps": 72.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-i06jcfln-ctx16384",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: d3kfammq, sa9tch22; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1210.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 1,210 67.0 26.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i06jcfln",
   "source_url": "https://omlx.ai/benchmarks/performance/i06jcfln",
   "tg_tokens": null,
   "tps": 67.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-i06jcfln-ctx32768",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: d3kfammq, sa9tch22; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1018.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 1,018 60.0 27.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i06jcfln",
   "source_url": "https://omlx.ai/benchmarks/performance/i06jcfln",
   "tg_tokens": null,
   "tps": 60.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-i06jcfln-ctx65536",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 29.1 GB at this context; duplicate traces of the same session: d3kfammq, sa9tch22; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 804.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 804.2 45.7 29.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i06jcfln",
   "source_url": "https://omlx.ai/benchmarks/performance/i06jcfln",
   "tg_tokens": null,
   "tps": 45.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-i06jcfln-ctx131072",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 32.4 GB at this context; duplicate traces of the same session: d3kfammq, sa9tch22; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 477.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 477.1 13.9 32.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i06jcfln",
   "source_url": "https://omlx.ai/benchmarks/performance/i06jcfln",
   "tg_tokens": null,
   "tps": 13.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-i06jcfln-ctx199680",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 35.9 GB at this context; duplicate traces of the same session: d3kfammq, sa9tch22; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 288.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 195k 288.3 23.5 35.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i06jcfln",
   "source_url": "https://omlx.ai/benchmarks/performance/i06jcfln",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-1fu7vr1q-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.7 GB at this context; duplicate traces of the same session: 1mdbpwex; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1613.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 16 GB \u00b7 4bit; 1k 1,613 44.2 4.7 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1fu7vr1q",
   "source_url": "https://omlx.ai/benchmarks/performance/1fu7vr1q",
   "tg_tokens": null,
   "tps": 44.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-1fu7vr1q-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.7 GB at this context; duplicate traces of the same session: 1mdbpwex; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2317.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 16 GB \u00b7 4bit; 4k 2,317 42.7 4.7 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1fu7vr1q",
   "source_url": "https://omlx.ai/benchmarks/performance/1fu7vr1q",
   "tg_tokens": null,
   "tps": 42.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-bsw6ei2u-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: w8nn5vvq; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 81.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ6e on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 6bit; 1k 81.1 9.2 28.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bsw6ei2u",
   "source_url": "https://omlx.ai/benchmarks/performance/bsw6ei2u",
   "tg_tokens": null,
   "tps": 9.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-bsw6ei2u-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace; peak memory 32.5 GB at this context; duplicate traces of the same session: w8nn5vvq; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 76.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ6e on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 6bit; 4k 76.0 8.9 32.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bsw6ei2u",
   "source_url": "https://omlx.ai/benchmarks/performance/bsw6ei2u",
   "tg_tokens": null,
   "tps": 8.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-rg9y8vcc-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 29.4 GB at this context; duplicate traces of the same session: qcxs0fpq; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 868.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-6bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 6bit; 4k 868.5 72.1 29.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rg9y8vcc",
   "source_url": "https://omlx.ai/benchmarks/performance/rg9y8vcc",
   "tg_tokens": null,
   "tps": 72.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-rg9y8vcc-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 30.4 GB at this context; duplicate traces of the same session: qcxs0fpq; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 721.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-6bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 6bit; 16k 721.6 59.8 30.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rg9y8vcc",
   "source_url": "https://omlx.ai/benchmarks/performance/rg9y8vcc",
   "tg_tokens": null,
   "tps": 59.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-z53bqmc6-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MLX Serve build; oMLX community benchmark trace, single run at 64k context; peak memory 21.2 GB; TTFT 18021 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3637.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-Serve-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k tokens 3,637 PP tok/s 100.4 TG tok/s 18021 TTFT (ms) 21.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z53bqmc6",
   "source_url": "https://omlx.ai/benchmarks/performance/z53bqmc6",
   "tg_tokens": null,
   "tps": 100.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-2kafibc0-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: i6qoeqfs; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 832.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 832.4 109.3 19.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2kafibc0",
   "source_url": "https://omlx.ai/benchmarks/performance/2kafibc0",
   "tg_tokens": null,
   "tps": 109.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-2kafibc0-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: i6qoeqfs; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1951.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 1,951 104.6 20.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2kafibc0",
   "source_url": "https://omlx.ai/benchmarks/performance/2kafibc0",
   "tg_tokens": null,
   "tps": 104.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-gj0pil0y-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 16k context; peak memory 22.3 GB; TTFT 9255 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1770.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k tokens 1,770 PP tok/s 102.8 TG tok/s 9255 TTFT (ms) 22.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gj0pil0y",
   "source_url": "https://omlx.ai/benchmarks/performance/gj0pil0y",
   "tg_tokens": null,
   "tps": 102.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-mgebv8nc-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: znizw2k5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 907.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 907.1 117.3 21.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mgebv8nc",
   "source_url": "https://omlx.ai/benchmarks/performance/mgebv8nc",
   "tg_tokens": null,
   "tps": 117.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-mgebv8nc-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: znizw2k5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1039.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,039 94.2 23.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mgebv8nc",
   "source_url": "https://omlx.ai/benchmarks/performance/mgebv8nc",
   "tg_tokens": null,
   "tps": 94.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-59n0fgiy-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 23.4 GB; TTFT 71199 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 230.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k tokens 230.1 PP tok/s 57.2 TG tok/s 71199 TTFT (ms) 23.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/59n0fgiy",
   "source_url": "https://omlx.ai/benchmarks/performance/59n0fgiy",
   "tg_tokens": null,
   "tps": 57.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ahfltjug-ctx131072",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 75.4 GB; TTFT 54588 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2401.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k tokens 2,401 PP tok/s 58.9 TG tok/s 54588 TTFT (ms) 75.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ahfltjug",
   "source_url": "https://omlx.ai/benchmarks/performance/ahfltjug",
   "tg_tokens": null,
   "tps": 58.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-l3oplkb0-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 69.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 757.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 757.0 79.6 69.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l3oplkb0",
   "source_url": "https://omlx.ai/benchmarks/performance/l3oplkb0",
   "tg_tokens": null,
   "tps": 79.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-l3oplkb0-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 70.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1364.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,364 69.6 70.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l3oplkb0",
   "source_url": "https://omlx.ai/benchmarks/performance/l3oplkb0",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wpqqfqad-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 69.5 GB; TTFT 3511 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1167.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 1,167 PP tok/s 62.3 TG tok/s 3511 TTFT (ms) 69.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wpqqfqad",
   "source_url": "https://omlx.ai/benchmarks/performance/wpqqfqad",
   "tg_tokens": null,
   "tps": 62.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-k041i0pw-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: h4jadvvo, rm8u5zz2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1579.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,579 60.5 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k041i0pw",
   "source_url": "https://omlx.ai/benchmarks/performance/k041i0pw",
   "tg_tokens": null,
   "tps": 60.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-k041i0pw-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.7 GB at this context; duplicate traces of the same session: h4jadvvo, rm8u5zz2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2371.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,371 82.3 72.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k041i0pw",
   "source_url": "https://omlx.ai/benchmarks/performance/k041i0pw",
   "tg_tokens": null,
   "tps": 82.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-k041i0pw-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.5 GB at this context; duplicate traces of the same session: h4jadvvo, rm8u5zz2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2407.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,407 79.5 74.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k041i0pw",
   "source_url": "https://omlx.ai/benchmarks/performance/k041i0pw",
   "tg_tokens": null,
   "tps": 79.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-scziz6nv-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.2 GB at this context; duplicate traces of the same session: t1zhotxw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1193.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,193 69.6 69.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/scziz6nv",
   "source_url": "https://omlx.ai/benchmarks/performance/scziz6nv",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-scziz6nv-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.3 GB at this context; duplicate traces of the same session: t1zhotxw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2193.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,193 65.9 69.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/scziz6nv",
   "source_url": "https://omlx.ai/benchmarks/performance/scziz6nv",
   "tg_tokens": null,
   "tps": 65.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-scziz6nv-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.1 GB at this context; duplicate traces of the same session: t1zhotxw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2031.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,031 64.2 70.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/scziz6nv",
   "source_url": "https://omlx.ai/benchmarks/performance/scziz6nv",
   "tg_tokens": null,
   "tps": 64.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-scziz6nv-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: t1zhotxw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2429.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,429 70.2 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/scziz6nv",
   "source_url": "https://omlx.ai/benchmarks/performance/scziz6nv",
   "tg_tokens": null,
   "tps": 70.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-scziz6nv-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.4 GB at this context; duplicate traces of the same session: t1zhotxw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2310.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,310 26.1 71.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/scziz6nv",
   "source_url": "https://omlx.ai/benchmarks/performance/scziz6nv",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-scziz6nv-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.4 GB at this context; duplicate traces of the same session: t1zhotxw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2293.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,293 63.7 72.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/scziz6nv",
   "source_url": "https://omlx.ai/benchmarks/performance/scziz6nv",
   "tg_tokens": null,
   "tps": 63.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-scziz6nv-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.9 GB at this context; duplicate traces of the same session: t1zhotxw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2218.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,218 63.9 74.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/scziz6nv",
   "source_url": "https://omlx.ai/benchmarks/performance/scziz6nv",
   "tg_tokens": null,
   "tps": 63.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-9q4x220x-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: 3nq0xvgm, 7ub3wxl1, fh5tff8o, w6et842m; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2768.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 2,768 131.4 100.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9q4x220x",
   "source_url": "https://omlx.ai/benchmarks/performance/9q4x220x",
   "tg_tokens": null,
   "tps": 131.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-9q4x220x-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.2 GB at this context; duplicate traces of the same session: 3nq0xvgm, 7ub3wxl1, fh5tff8o, w6et842m; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5636.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 5,636 155.3 102.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9q4x220x",
   "source_url": "https://omlx.ai/benchmarks/performance/9q4x220x",
   "tg_tokens": null,
   "tps": 155.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-9q4x220x-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.4 GB at this context; duplicate traces of the same session: 3nq0xvgm, 7ub3wxl1, fh5tff8o, w6et842m; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5800.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 5,800 139.0 104.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9q4x220x",
   "source_url": "https://omlx.ai/benchmarks/performance/9q4x220x",
   "tg_tokens": null,
   "tps": 139.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-9q4x220x-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 108.9 GB at this context; duplicate traces of the same session: 3nq0xvgm, 7ub3wxl1, fh5tff8o, w6et842m; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5596.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,596 112.8 108.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9q4x220x",
   "source_url": "https://omlx.ai/benchmarks/performance/9q4x220x",
   "tg_tokens": null,
   "tps": 112.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-9q4x220x-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.7 GB at this context; duplicate traces of the same session: 3nq0xvgm, 7ub3wxl1, fh5tff8o, w6et842m; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5407.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 5,407 89.5 110.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9q4x220x",
   "source_url": "https://omlx.ai/benchmarks/performance/9q4x220x",
   "tg_tokens": null,
   "tps": 89.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-90k9004a-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.0 GB at this context; duplicate traces of the same session: 3kgszx9w, cpmx28sk, scfh5k3v, v5gmstut; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2951.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 2,951 124.6 99.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/90k9004a",
   "source_url": "https://omlx.ai/benchmarks/performance/90k9004a",
   "tg_tokens": null,
   "tps": 124.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-90k9004a-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.4 GB at this context; duplicate traces of the same session: 3kgszx9w, cpmx28sk, scfh5k3v, v5gmstut; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5695.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 5,695 117.5 100.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/90k9004a",
   "source_url": "https://omlx.ai/benchmarks/performance/90k9004a",
   "tg_tokens": null,
   "tps": 117.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-90k9004a-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.3 GB at this context; duplicate traces of the same session: 3kgszx9w, cpmx28sk, scfh5k3v, v5gmstut; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5738.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 5,738 100.2 102.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/90k9004a",
   "source_url": "https://omlx.ai/benchmarks/performance/90k9004a",
   "tg_tokens": null,
   "tps": 100.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-90k9004a-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.7 GB at this context; duplicate traces of the same session: 3kgszx9w, cpmx28sk, scfh5k3v, v5gmstut; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5562.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,562 98.2 104.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/90k9004a",
   "source_url": "https://omlx.ai/benchmarks/performance/90k9004a",
   "tg_tokens": null,
   "tps": 98.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-90k9004a-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 108.2 GB at this context; duplicate traces of the same session: 3kgszx9w, cpmx28sk, scfh5k3v, v5gmstut; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5382.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 5,382 96.6 108.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/90k9004a",
   "source_url": "https://omlx.ai/benchmarks/performance/90k9004a",
   "tg_tokens": null,
   "tps": 96.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-86yrszvm-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: 4ojk1w93, 8kb7shz7, 9yplml9x, jqb6yjto, yverzv9f; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2818.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 2,818 158.2 100.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/86yrszvm",
   "source_url": "https://omlx.ai/benchmarks/performance/86yrszvm",
   "tg_tokens": null,
   "tps": 158.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-86yrszvm-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.1 GB at this context; duplicate traces of the same session: 4ojk1w93, 8kb7shz7, 9yplml9x, jqb6yjto, yverzv9f; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5628.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 5,628 161.3 102.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/86yrszvm",
   "source_url": "https://omlx.ai/benchmarks/performance/86yrszvm",
   "tg_tokens": null,
   "tps": 161.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-86yrszvm-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.6 GB at this context; duplicate traces of the same session: 4ojk1w93, 8kb7shz7, 9yplml9x, jqb6yjto, yverzv9f; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5645.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 5,645 113.7 102.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/86yrszvm",
   "source_url": "https://omlx.ai/benchmarks/performance/86yrszvm",
   "tg_tokens": null,
   "tps": 113.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-86yrszvm-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.4 GB at this context; duplicate traces of the same session: 4ojk1w93, 8kb7shz7, 9yplml9x, jqb6yjto, yverzv9f; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5509.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 5,509 145.3 104.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/86yrszvm",
   "source_url": "https://omlx.ai/benchmarks/performance/86yrszvm",
   "tg_tokens": null,
   "tps": 145.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-86yrszvm-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 108.9 GB at this context; duplicate traces of the same session: 4ojk1w93, 8kb7shz7, 9yplml9x, jqb6yjto, yverzv9f; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5322.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,322 111.9 108.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/86yrszvm",
   "source_url": "https://omlx.ai/benchmarks/performance/86yrszvm",
   "tg_tokens": null,
   "tps": 111.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-86yrszvm-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.7 GB at this context; duplicate traces of the same session: 4ojk1w93, 8kb7shz7, 9yplml9x, jqb6yjto, yverzv9f; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5148.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 5,148 89.2 110.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/86yrszvm",
   "source_url": "https://omlx.ai/benchmarks/performance/86yrszvm",
   "tg_tokens": null,
   "tps": 89.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-l29hys7d-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: 5fa3keky, fixtr9ed; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3010.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 3,010 159.8 100.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l29hys7d",
   "source_url": "https://omlx.ai/benchmarks/performance/l29hys7d",
   "tg_tokens": null,
   "tps": 159.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-l29hys7d-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.1 GB at this context; duplicate traces of the same session: 5fa3keky, fixtr9ed; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5646.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 5,646 127.0 102.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l29hys7d",
   "source_url": "https://omlx.ai/benchmarks/performance/l29hys7d",
   "tg_tokens": null,
   "tps": 127.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-l29hys7d-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.4 GB at this context; duplicate traces of the same session: 5fa3keky, fixtr9ed; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5541.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 5,541 138.8 104.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l29hys7d",
   "source_url": "https://omlx.ai/benchmarks/performance/l29hys7d",
   "tg_tokens": null,
   "tps": 138.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-7pqdtp27-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 100.6 GB; TTFT 372 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2754.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k tokens 2,754 PP tok/s 127.0 TG tok/s 372 TTFT (ms) 100.6 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7pqdtp27",
   "source_url": "https://omlx.ai/benchmarks/performance/7pqdtp27",
   "tg_tokens": null,
   "tps": 127.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-bv1v45uo-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.7 GB at this context; duplicate traces of the same session: 6wd8s1k8, kczyti2p, serqax3x, xtpt1mgb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2694.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 2,694 115.7 100.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bv1v45uo",
   "source_url": "https://omlx.ai/benchmarks/performance/bv1v45uo",
   "tg_tokens": null,
   "tps": 115.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-bv1v45uo-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.1 GB at this context; duplicate traces of the same session: 6wd8s1k8, kczyti2p, serqax3x, xtpt1mgb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5466.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 5,466 132.4 102.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bv1v45uo",
   "source_url": "https://omlx.ai/benchmarks/performance/bv1v45uo",
   "tg_tokens": null,
   "tps": 132.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-bv1v45uo-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.3 GB at this context; duplicate traces of the same session: 6wd8s1k8, kczyti2p, serqax3x, xtpt1mgb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5524.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 5,524 131.1 104.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bv1v45uo",
   "source_url": "https://omlx.ai/benchmarks/performance/bv1v45uo",
   "tg_tokens": null,
   "tps": 131.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-bv1v45uo-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.1 GB at this context; duplicate traces of the same session: 6wd8s1k8, kczyti2p, serqax3x, xtpt1mgb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5355.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,355 113.5 109.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bv1v45uo",
   "source_url": "https://omlx.ai/benchmarks/performance/bv1v45uo",
   "tg_tokens": null,
   "tps": 113.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-bv1v45uo-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.7 GB at this context; duplicate traces of the same session: 6wd8s1k8, kczyti2p, serqax3x, xtpt1mgb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5195.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 5,195 84.9 110.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bv1v45uo",
   "source_url": "https://omlx.ai/benchmarks/performance/bv1v45uo",
   "tg_tokens": null,
   "tps": 84.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-00cusxro-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.4 GB at this context; duplicate traces of the same session: g64t6bt7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3073.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 3,073 132.7 100.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/00cusxro",
   "source_url": "https://omlx.ai/benchmarks/performance/00cusxro",
   "tg_tokens": null,
   "tps": 132.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-00cusxro-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.1 GB at this context; duplicate traces of the same session: g64t6bt7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5664.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 5,664 158.8 102.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/00cusxro",
   "source_url": "https://omlx.ai/benchmarks/performance/00cusxro",
   "tg_tokens": null,
   "tps": 158.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ebuzypqc-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.4 GB at this context; duplicate traces of the same session: xv5wvnfb, zaeuc2p7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2742.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 2,742 122.9 100.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ebuzypqc",
   "source_url": "https://omlx.ai/benchmarks/performance/ebuzypqc",
   "tg_tokens": null,
   "tps": 122.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ebuzypqc-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.1 GB at this context; duplicate traces of the same session: xv5wvnfb, zaeuc2p7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5475.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 5,475 123.7 102.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ebuzypqc",
   "source_url": "https://omlx.ai/benchmarks/performance/ebuzypqc",
   "tg_tokens": null,
   "tps": 123.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ebuzypqc-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.2 GB at this context; duplicate traces of the same session: xv5wvnfb, zaeuc2p7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5544.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 5,544 128.3 104.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ebuzypqc",
   "source_url": "https://omlx.ai/benchmarks/performance/ebuzypqc",
   "tg_tokens": null,
   "tps": 128.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ebuzypqc-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 108.8 GB at this context; duplicate traces of the same session: xv5wvnfb, zaeuc2p7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5358.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,358 109.0 108.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ebuzypqc",
   "source_url": "https://omlx.ai/benchmarks/performance/ebuzypqc",
   "tg_tokens": null,
   "tps": 109.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ebuzypqc-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.6 GB at this context; duplicate traces of the same session: xv5wvnfb, zaeuc2p7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5193.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 5,193 90.2 110.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ebuzypqc",
   "source_url": "https://omlx.ai/benchmarks/performance/ebuzypqc",
   "tg_tokens": null,
   "tps": 90.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gv2vn9ui-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 150.3 GB at this context; duplicate traces of the same session: h5atmauq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 999.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 128k 999.3 53.7 150.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gv2vn9ui",
   "source_url": "https://omlx.ai/benchmarks/performance/gv2vn9ui",
   "tg_tokens": null,
   "tps": 53.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gv2vn9ui-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 151.7 GB at this context; duplicate traces of the same session: h5atmauq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 986.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 195k 986.3 48.4 151.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gv2vn9ui",
   "source_url": "https://omlx.ai/benchmarks/performance/gv2vn9ui",
   "tg_tokens": null,
   "tps": 48.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-07lkfs6p-ctx1024",
   "model": "GLM-5.3-Flash-Abliterated",
   "notes": "MLX; oMLX community benchmark trace, single run at 1k context; peak memory 177.6 GB; TTFT 2198 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 466.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-Abliterated-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k tokens 466.0 PP tok/s 42.8 TG tok/s 2198 TTFT (ms) 177.6 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/07lkfs6p",
   "source_url": "https://omlx.ai/benchmarks/performance/07lkfs6p",
   "tg_tokens": null,
   "tps": 42.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-wrf7e6g5-ctx1024",
   "model": "GLM-5.3",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 399.5 GB; TTFT 5184 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 197.5,
   "provenance": "community",
   "quant": "mixed 4/8bit",
   "quote": "GLM-5.3-mixed-4-8 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k tokens 197.5 PP tok/s 14.7 TG tok/s 5184 TTFT (ms) 399.5 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wrf7e6g5",
   "source_url": "https://omlx.ai/benchmarks/performance/wrf7e6g5",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-rj1s2wzv-ctx4096",
   "model": "LFM2.5-8B-A1B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 9.3 GB at this context; duplicate traces of the same session: m44qy0vy, ynuukiky; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 12888.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-8B-A1B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 12,888 174.9 9.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rj1s2wzv",
   "source_url": "https://omlx.ai/benchmarks/performance/rj1s2wzv",
   "tg_tokens": null,
   "tps": 174.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-rj1s2wzv-ctx16384",
   "model": "LFM2.5-8B-A1B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 9.5 GB at this context; duplicate traces of the same session: m44qy0vy, ynuukiky; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 12166.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-8B-A1B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 12,166 152.6 9.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rj1s2wzv",
   "source_url": "https://omlx.ai/benchmarks/performance/rj1s2wzv",
   "tg_tokens": null,
   "tps": 152.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-rj1s2wzv-ctx65536",
   "model": "LFM2.5-8B-A1B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 10.1 GB at this context; duplicate traces of the same session: m44qy0vy, ynuukiky; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 9844.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-8B-A1B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 9,844 108.5 10.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rj1s2wzv",
   "source_url": "https://omlx.ai/benchmarks/performance/rj1s2wzv",
   "tg_tokens": null,
   "tps": 108.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fo0v7tqd-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 35.5 GB; TTFT 239 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4294.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k tokens 4,294 PP tok/s 106.8 TG tok/s 239 TTFT (ms) 35.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fo0v7tqd",
   "source_url": "https://omlx.ai/benchmarks/performance/fo0v7tqd",
   "tg_tokens": null,
   "tps": 106.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3xmc50vb-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 65.8 GB at this context; duplicate traces of the same session: ijujp231; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2341.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Ornith-1.5-35B-A3B-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 2,341 72.3 65.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3xmc50vb",
   "source_url": "https://omlx.ai/benchmarks/performance/3xmc50vb",
   "tg_tokens": null,
   "tps": 72.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3xmc50vb-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 66.5 GB at this context; duplicate traces of the same session: ijujp231; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4179.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Ornith-1.5-35B-A3B-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 4k 4,179 70.7 66.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3xmc50vb",
   "source_url": "https://omlx.ai/benchmarks/performance/3xmc50vb",
   "tg_tokens": null,
   "tps": 70.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-vj4q65jp-ctx4096",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: ffg5vugq; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1736.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,736 60.5 16.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vj4q65jp",
   "source_url": "https://omlx.ai/benchmarks/performance/vj4q65jp",
   "tg_tokens": null,
   "tps": 60.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-vj4q65jp-ctx16384",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 18.0 GB at this context; duplicate traces of the same session: ffg5vugq; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1194.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 1,194 40.4 18.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vj4q65jp",
   "source_url": "https://omlx.ai/benchmarks/performance/vj4q65jp",
   "tg_tokens": null,
   "tps": 40.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nuziaxxm-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MLX Serve build; oMLX community benchmark trace, single run at 64k context; peak memory 21.2 GB; TTFT 17710 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3701.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-Serve-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k tokens 3,701 PP tok/s 100.9 TG tok/s 17710 TTFT (ms) 21.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nuziaxxm",
   "source_url": "https://omlx.ai/benchmarks/performance/nuziaxxm",
   "tg_tokens": null,
   "tps": 100.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-j7f3xm3n-ctx16384",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final",
   "notes": "dequantized fp16 MTP build; oMLX community benchmark trace, single run at 16k context; peak memory 40.3 GB; TTFT 9875 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1659.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized-oQ8e-fp16-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k tokens 1,659 PP tok/s 137.9 TG tok/s 9875 TTFT (ms) 40.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j7f3xm3n",
   "source_url": "https://omlx.ai/benchmarks/performance/j7f3xm3n",
   "tg_tokens": null,
   "tps": 137.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-zh9zvjuz-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: g78fz2qq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2316.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 2,316 190.3 23.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zh9zvjuz",
   "source_url": "https://omlx.ai/benchmarks/performance/zh9zvjuz",
   "tg_tokens": null,
   "tps": 190.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-zh9zvjuz-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: g78fz2qq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2073.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 2,073 174.5 23.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zh9zvjuz",
   "source_url": "https://omlx.ai/benchmarks/performance/zh9zvjuz",
   "tg_tokens": null,
   "tps": 174.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-yq9dpjx8-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: g22t0s7o; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 389.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 389.4 18.1 16.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yq9dpjx8",
   "source_url": "https://omlx.ai/benchmarks/performance/yq9dpjx8",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-yq9dpjx8-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: g22t0s7o; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 414.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 414.4 17.9 17.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yq9dpjx8",
   "source_url": "https://omlx.ai/benchmarks/performance/yq9dpjx8",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-t5djlkvq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 263.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 263.7 9.8 15.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t5djlkvq",
   "source_url": "https://omlx.ai/benchmarks/performance/t5djlkvq",
   "tg_tokens": null,
   "tps": 9.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-t5djlkvq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 314.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 314.5 9.7 17.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t5djlkvq",
   "source_url": "https://omlx.ai/benchmarks/performance/t5djlkvq",
   "tg_tokens": null,
   "tps": 9.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-t5djlkvq-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 311.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 8k 311.4 9.5 17.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t5djlkvq",
   "source_url": "https://omlx.ai/benchmarks/performance/t5djlkvq",
   "tg_tokens": null,
   "tps": 9.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-ersdf09r-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: awkxi1uy, u1yziugy; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 198.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 198.2 15.3 19.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ersdf09r",
   "source_url": "https://omlx.ai/benchmarks/performance/ersdf09r",
   "tg_tokens": null,
   "tps": 15.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-ersdf09r-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: awkxi1uy, u1yziugy; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 308.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 308.3 19.2 21.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ersdf09r",
   "source_url": "https://omlx.ai/benchmarks/performance/ersdf09r",
   "tg_tokens": null,
   "tps": 19.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-ersdf09r-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: awkxi1uy, u1yziugy; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 297.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 8k 297.9 18.5 22.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ersdf09r",
   "source_url": "https://omlx.ai/benchmarks/performance/ersdf09r",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-reqtpbbv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MLX Serve build; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 8otb4mrs, vs5fs0ms; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1013.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-Serve-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,013 33.2 18.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/reqtpbbv",
   "source_url": "https://omlx.ai/benchmarks/performance/reqtpbbv",
   "tg_tokens": null,
   "tps": 33.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-reqtpbbv-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MLX Serve build; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: 8otb4mrs, vs5fs0ms; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 884.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-Serve-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 884.6 31.4 19.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/reqtpbbv",
   "source_url": "https://omlx.ai/benchmarks/performance/reqtpbbv",
   "tg_tokens": null,
   "tps": 31.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-reqtpbbv-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MLX Serve build; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: 8otb4mrs, vs5fs0ms; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 690.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-Serve-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 690.4 26.1 22.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/reqtpbbv",
   "source_url": "https://omlx.ai/benchmarks/performance/reqtpbbv",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-eub8dqlz-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MLX Serve build; oMLX community benchmark trace; peak memory 30.7 GB at this context; duplicate traces of the same session: 3mzhh36l, axmvkh3a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 947.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-Serve-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 947.7 18.9 30.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eub8dqlz",
   "source_url": "https://omlx.ai/benchmarks/performance/eub8dqlz",
   "tg_tokens": null,
   "tps": 18.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-eub8dqlz-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MLX Serve build; oMLX community benchmark trace; peak memory 31.4 GB at this context; duplicate traces of the same session: 3mzhh36l, axmvkh3a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 805.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-Serve-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 805.0 17.8 31.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eub8dqlz",
   "source_url": "https://omlx.ai/benchmarks/performance/eub8dqlz",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-eub8dqlz-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MLX Serve build; oMLX community benchmark trace; peak memory 34.4 GB at this context; duplicate traces of the same session: 3mzhh36l, axmvkh3a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 661.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-Serve-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 661.4 15.6 34.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eub8dqlz",
   "source_url": "https://omlx.ai/benchmarks/performance/eub8dqlz",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1117yd9c-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: g9tvv30z; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 444.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 444.1 73.3 25.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1117yd9c",
   "source_url": "https://omlx.ai/benchmarks/performance/1117yd9c",
   "tg_tokens": null,
   "tps": 73.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1117yd9c-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: g9tvv30z; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 423.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 423.5 83.0 26.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1117yd9c",
   "source_url": "https://omlx.ai/benchmarks/performance/1117yd9c",
   "tg_tokens": null,
   "tps": 83.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-wxgl7s95-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 2hhl2l3t, ph0810on; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 408.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 408.3 53.6 21.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wxgl7s95",
   "source_url": "https://omlx.ai/benchmarks/performance/wxgl7s95",
   "tg_tokens": null,
   "tps": 53.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-wxgl7s95-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: 2hhl2l3t, ph0810on; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 442.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 442.6 72.0 24.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wxgl7s95",
   "source_url": "https://omlx.ai/benchmarks/performance/wxgl7s95",
   "tg_tokens": null,
   "tps": 72.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-wxgl7s95-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: 2hhl2l3t, ph0810on; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 517.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 517.0 100.0 25.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wxgl7s95",
   "source_url": "https://omlx.ai/benchmarks/performance/wxgl7s95",
   "tg_tokens": null,
   "tps": 100.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-wxgl7s95-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: 2hhl2l3t, ph0810on; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 511.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 511.9 88.9 25.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wxgl7s95",
   "source_url": "https://omlx.ai/benchmarks/performance/wxgl7s95",
   "tg_tokens": null,
   "tps": 88.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-hdu10taj-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: s4x92614; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1289.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k 1,289 87.2 17.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hdu10taj",
   "source_url": "https://omlx.ai/benchmarks/performance/hdu10taj",
   "tg_tokens": null,
   "tps": 87.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-hdu10taj-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: s4x92614; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1306.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 4k 1,306 93.1 18.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hdu10taj",
   "source_url": "https://omlx.ai/benchmarks/performance/hdu10taj",
   "tg_tokens": null,
   "tps": 93.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ta0jjt0p-ctx16384",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 143.6 GB at this context; duplicate traces of the same session: kh43hpkj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1339.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 16k 1,339 85.7 143.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ta0jjt0p",
   "source_url": "https://omlx.ai/benchmarks/performance/ta0jjt0p",
   "tg_tokens": null,
   "tps": 85.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ta0jjt0p-ctx131072",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 151.3 GB at this context; duplicate traces of the same session: kh43hpkj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1280.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 128k 1,280 64.6 151.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ta0jjt0p",
   "source_url": "https://omlx.ai/benchmarks/performance/ta0jjt0p",
   "tg_tokens": null,
   "tps": 64.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-054vk6wz-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.2 GB at this context; duplicate traces of the same session: 1j1t2dn7, 2dg4jcro, 4t0r465f, f5e7tsl8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1193.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,193 69.6 69.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/054vk6wz",
   "source_url": "https://omlx.ai/benchmarks/performance/054vk6wz",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-054vk6wz-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.3 GB at this context; duplicate traces of the same session: 1j1t2dn7, 2dg4jcro, 4t0r465f, f5e7tsl8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2193.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,193 65.9 69.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/054vk6wz",
   "source_url": "https://omlx.ai/benchmarks/performance/054vk6wz",
   "tg_tokens": null,
   "tps": 65.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-054vk6wz-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.1 GB at this context; duplicate traces of the same session: 1j1t2dn7, 2dg4jcro, 4t0r465f, f5e7tsl8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2031.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,031 64.2 70.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/054vk6wz",
   "source_url": "https://omlx.ai/benchmarks/performance/054vk6wz",
   "tg_tokens": null,
   "tps": 64.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-054vk6wz-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 1j1t2dn7, 2dg4jcro, 4t0r465f, f5e7tsl8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2429.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,429 70.2 70.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/054vk6wz",
   "source_url": "https://omlx.ai/benchmarks/performance/054vk6wz",
   "tg_tokens": null,
   "tps": 70.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-054vk6wz-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.4 GB at this context; duplicate traces of the same session: 1j1t2dn7, 2dg4jcro, 4t0r465f, f5e7tsl8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2310.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,310 26.1 71.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/054vk6wz",
   "source_url": "https://omlx.ai/benchmarks/performance/054vk6wz",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-054vk6wz-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.4 GB at this context; duplicate traces of the same session: 1j1t2dn7, 2dg4jcro, 4t0r465f, f5e7tsl8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2293.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,293 63.7 72.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/054vk6wz",
   "source_url": "https://omlx.ai/benchmarks/performance/054vk6wz",
   "tg_tokens": null,
   "tps": 63.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-054vk6wz-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.9 GB at this context; duplicate traces of the same session: 1j1t2dn7, 2dg4jcro, 4t0r465f, f5e7tsl8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2218.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,218 63.9 74.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/054vk6wz",
   "source_url": "https://omlx.ai/benchmarks/performance/054vk6wz",
   "tg_tokens": null,
   "tps": 63.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-veqvzv3h-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.4 GB at this context; duplicate traces of the same session: skihjadj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2742.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 2,742 122.9 100.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/veqvzv3h",
   "source_url": "https://omlx.ai/benchmarks/performance/veqvzv3h",
   "tg_tokens": null,
   "tps": 122.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-veqvzv3h-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.1 GB at this context; duplicate traces of the same session: skihjadj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5475.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 5,475 123.7 102.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/veqvzv3h",
   "source_url": "https://omlx.ai/benchmarks/performance/veqvzv3h",
   "tg_tokens": null,
   "tps": 123.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-veqvzv3h-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.2 GB at this context; duplicate traces of the same session: skihjadj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5544.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 5,544 128.3 104.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/veqvzv3h",
   "source_url": "https://omlx.ai/benchmarks/performance/veqvzv3h",
   "tg_tokens": null,
   "tps": 128.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-veqvzv3h-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 108.8 GB at this context; duplicate traces of the same session: skihjadj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5358.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,358 109.0 108.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/veqvzv3h",
   "source_url": "https://omlx.ai/benchmarks/performance/veqvzv3h",
   "tg_tokens": null,
   "tps": 109.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-veqvzv3h-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.6 GB at this context; duplicate traces of the same session: skihjadj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5193.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 5,193 90.2 110.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/veqvzv3h",
   "source_url": "https://omlx.ai/benchmarks/performance/veqvzv3h",
   "tg_tokens": null,
   "tps": 90.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-gj8fo0ts-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 109.0 GB; TTFT 1381 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 741.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k tokens 741.5 PP tok/s 38.9 TG tok/s 1381 TTFT (ms) 109.0 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gj8fo0ts",
   "source_url": "https://omlx.ai/benchmarks/performance/gj8fo0ts",
   "tg_tokens": null,
   "tps": 38.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-z03qh61h-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 195k context; peak memory 110.5 GB; TTFT 38826 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5151.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k tokens 5,151 PP tok/s 103.9 TG tok/s 38826 TTFT (ms) 110.5 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z03qh61h",
   "source_url": "https://omlx.ai/benchmarks/performance/z03qh61h",
   "tg_tokens": null,
   "tps": 103.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-5jkow92t-ctx1024",
   "model": "bitnet-b1.58-2B",
   "notes": "oMLX community benchmark trace; peak memory 1.9 GB at this context; duplicate traces of the same session: jr3ij8ev, ry586nzi; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 814.2,
   "provenance": "community",
   "quant": "4T",
   "quote": "bitnet-b1.58-2B-4T on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 unknown; 1k 814.2 170.7 1.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5jkow92t",
   "source_url": "https://omlx.ai/benchmarks/performance/5jkow92t",
   "tg_tokens": null,
   "tps": 170.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-5jkow92t-ctx4096",
   "model": "bitnet-b1.58-2B",
   "notes": "oMLX community benchmark trace; peak memory 2.1 GB at this context; duplicate traces of the same session: jr3ij8ev, ry586nzi; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 802.1,
   "provenance": "community",
   "quant": "4T",
   "quote": "bitnet-b1.58-2B-4T on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 unknown; 4k 802.1 148.4 2.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5jkow92t",
   "source_url": "https://omlx.ai/benchmarks/performance/5jkow92t",
   "tg_tokens": null,
   "tps": 148.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-5jkow92t-ctx8192",
   "model": "bitnet-b1.58-2B",
   "notes": "oMLX community benchmark trace; peak memory 2.4 GB at this context; duplicate traces of the same session: jr3ij8ev, ry586nzi; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 790.1,
   "provenance": "community",
   "quant": "4T",
   "quote": "bitnet-b1.58-2B-4T on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 unknown; 8k 790.1 130.5 2.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5jkow92t",
   "source_url": "https://omlx.ai/benchmarks/performance/5jkow92t",
   "tg_tokens": null,
   "tps": 130.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-zs37dzso-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: dmg5brkc, f4536193; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3873.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 3,873 112.8 14.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zs37dzso",
   "source_url": "https://omlx.ai/benchmarks/performance/zs37dzso",
   "tg_tokens": null,
   "tps": 112.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-zs37dzso-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.9 GB at this context; duplicate traces of the same session: dmg5brkc, f4536193; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3497.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 3,497 94.2 14.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zs37dzso",
   "source_url": "https://omlx.ai/benchmarks/performance/zs37dzso",
   "tg_tokens": null,
   "tps": 94.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zs37dzso-ctx65536",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: dmg5brkc, f4536193; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2526.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,526 63.0 17.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zs37dzso",
   "source_url": "https://omlx.ai/benchmarks/performance/zs37dzso",
   "tg_tokens": null,
   "tps": 63.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-9h2jdi9i-ctx1024",
   "model": "translategemma-4b-it",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 2.9 GB; TTFT 678 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1511.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "translategemma-4b-it-4bit on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 1,511 PP tok/s 112.6 TG tok/s 678 TTFT (ms) 2.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9h2jdi9i",
   "source_url": "https://omlx.ai/benchmarks/performance/9h2jdi9i",
   "tg_tokens": null,
   "tps": 112.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-m9cijvtc-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: racanhjz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 430.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ6e on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 1k 430.2 25.7 26.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m9cijvtc",
   "source_url": "https://omlx.ai/benchmarks/performance/m9cijvtc",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-m9cijvtc-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: racanhjz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 466.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ6e on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 4k 466.3 26.5 27.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m9cijvtc",
   "source_url": "https://omlx.ai/benchmarks/performance/m9cijvtc",
   "tg_tokens": null,
   "tps": 26.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-x8dlf0ka-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Jundot build; oMLX community benchmark trace; peak memory 100.0 GB at this context; duplicate traces of the same session: amdys2ag; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 993.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 993.5 67.2 100.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x8dlf0ka",
   "source_url": "https://omlx.ai/benchmarks/performance/x8dlf0ka",
   "tg_tokens": null,
   "tps": 67.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-x8dlf0ka-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Jundot build; oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: amdys2ag; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1052.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 1,052 49.9 101.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x8dlf0ka",
   "source_url": "https://omlx.ai/benchmarks/performance/x8dlf0ka",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9f8o17x6-ctx16384",
   "model": "Mistral-Small-3.2-24B-Instruct-2506",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 15.8 GB; TTFT 19905 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 823.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Mistral-Small-3.2-24B-Instruct-2506-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k tokens 823.1 PP tok/s 31.5 TG tok/s 19905 TTFT (ms) 15.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9f8o17x6",
   "source_url": "https://omlx.ai/benchmarks/performance/9f8o17x6",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hhz8i59g-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: qh0xmsee; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4203.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 4,203 98.2 35.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hhz8i59g",
   "source_url": "https://omlx.ai/benchmarks/performance/hhz8i59g",
   "tg_tokens": null,
   "tps": 98.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hhz8i59g-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 36.2 GB at this context; duplicate traces of the same session: qh0xmsee; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6046.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 6,046 95.0 36.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hhz8i59g",
   "source_url": "https://omlx.ai/benchmarks/performance/hhz8i59g",
   "tg_tokens": null,
   "tps": 95.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-6cbyjsll-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 21.9 GB; TTFT 1384 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 740.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 740.0 PP tok/s 113.3 TG tok/s 1384 TTFT (ms) 21.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6cbyjsll",
   "source_url": "https://omlx.ai/benchmarks/performance/6cbyjsll",
   "tg_tokens": null,
   "tps": 113.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-m50l0tgb-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 0wrcck1i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4544.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,544 108.6 21.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m50l0tgb",
   "source_url": "https://omlx.ai/benchmarks/performance/m50l0tgb",
   "tg_tokens": null,
   "tps": 108.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-m50l0tgb-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: 0wrcck1i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6197.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,197 102.0 22.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m50l0tgb",
   "source_url": "https://omlx.ai/benchmarks/performance/m50l0tgb",
   "tg_tokens": null,
   "tps": 102.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-hv81q4q7-ctx1024",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "MLX; oMLX community benchmark trace; peak memory 30.8 GB at this context; duplicate traces of the same session: l60tqd1g; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 599.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-MLX-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 1k 599.8 47.7 30.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hv81q4q7",
   "source_url": "https://omlx.ai/benchmarks/performance/hv81q4q7",
   "tg_tokens": null,
   "tps": 47.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-hv81q4q7-ctx4096",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "MLX; oMLX community benchmark trace; peak memory 31.3 GB at this context; duplicate traces of the same session: l60tqd1g; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 661.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-MLX-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k 661.9 41.8 31.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hv81q4q7",
   "source_url": "https://omlx.ai/benchmarks/performance/hv81q4q7",
   "tg_tokens": null,
   "tps": 41.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-p0zs5doc-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: 467hqwoe; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1022.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 16 GB \u00b7 4bit; 1k 1,022 29.7 5.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p0zs5doc",
   "source_url": "https://omlx.ai/benchmarks/performance/p0zs5doc",
   "tg_tokens": null,
   "tps": 29.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-p0zs5doc-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: 467hqwoe; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1121.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 16 GB \u00b7 4bit; 4k 1,121 29.1 6.2 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p0zs5doc",
   "source_url": "https://omlx.ai/benchmarks/performance/p0zs5doc",
   "tg_tokens": null,
   "tps": 29.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3lejxqxc-ctx1024",
   "model": "Qwen3.6-27B-Coder-uncensored",
   "notes": "oMLX community benchmark trace; peak memory 15.2 GB at this context; duplicate traces of the same session: 4u0r11j2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 271.4,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Qwen3.6-27B-Coder-uncensored-MXFP4 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 271.4 34.9 15.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3lejxqxc",
   "source_url": "https://omlx.ai/benchmarks/performance/3lejxqxc",
   "tg_tokens": null,
   "tps": 34.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3lejxqxc-ctx4096",
   "model": "Qwen3.6-27B-Coder-uncensored",
   "notes": "oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 4u0r11j2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 255.6,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Qwen3.6-27B-Coder-uncensored-MXFP4 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 255.6 34.8 18.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3lejxqxc",
   "source_url": "https://omlx.ai/benchmarks/performance/3lejxqxc",
   "tg_tokens": null,
   "tps": 34.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-zlyjvcka-ctx4096",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1625.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-text-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,625 48.8 20.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zlyjvcka",
   "source_url": "https://omlx.ai/benchmarks/performance/zlyjvcka",
   "tg_tokens": null,
   "tps": 48.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-zlyjvcka-ctx8192",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1598.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-text-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 1,598 47.4 20.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zlyjvcka",
   "source_url": "https://omlx.ai/benchmarks/performance/zlyjvcka",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-zlyjvcka-ctx16384",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1422.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-text-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 1,422 43.2 21.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zlyjvcka",
   "source_url": "https://omlx.ai/benchmarks/performance/zlyjvcka",
   "tg_tokens": null,
   "tps": 43.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M6",
   "id": "omlx-apjdgcsx-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: r20j3ceu; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 128.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 16 GB \u00b7 4bit; 1k 128.3 11.4 6.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/apjdgcsx",
   "source_url": "https://omlx.ai/benchmarks/performance/apjdgcsx",
   "tg_tokens": null,
   "tps": 11.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M6",
   "id": "omlx-apjdgcsx-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: r20j3ceu; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 229.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 16 GB \u00b7 4bit; 4k 229.6 10.2 6.9 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/apjdgcsx",
   "source_url": "https://omlx.ai/benchmarks/performance/apjdgcsx",
   "tg_tokens": null,
   "tps": 10.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-lcbkwg6j-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: ul91mu9b; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 563.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 563.8 16.3 18.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lcbkwg6j",
   "source_url": "https://omlx.ai/benchmarks/performance/lcbkwg6j",
   "tg_tokens": null,
   "tps": 16.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-lcbkwg6j-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: ul91mu9b; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 628.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 628.1 15.7 19.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lcbkwg6j",
   "source_url": "https://omlx.ai/benchmarks/performance/lcbkwg6j",
   "tg_tokens": null,
   "tps": 15.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-i56ltsu3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: vupbl73n; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 263.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 263.7 9.8 15.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i56ltsu3",
   "source_url": "https://omlx.ai/benchmarks/performance/i56ltsu3",
   "tg_tokens": null,
   "tps": 9.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-i56ltsu3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: vupbl73n; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 314.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 314.5 9.7 17.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i56ltsu3",
   "source_url": "https://omlx.ai/benchmarks/performance/i56ltsu3",
   "tg_tokens": null,
   "tps": 9.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-i56ltsu3-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: vupbl73n; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 311.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 8k 311.4 9.5 17.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i56ltsu3",
   "source_url": "https://omlx.ai/benchmarks/performance/i56ltsu3",
   "tg_tokens": null,
   "tps": 9.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-dzjql46t-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: obhl8lm1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 256.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 8bit; 1k 256.3 16.7 28.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dzjql46t",
   "source_url": "https://omlx.ai/benchmarks/performance/dzjql46t",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-dzjql46t-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: obhl8lm1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 254.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 8bit; 4k 254.7 16.4 29.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dzjql46t",
   "source_url": "https://omlx.ai/benchmarks/performance/dzjql46t",
   "tg_tokens": null,
   "tps": 16.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4sd0ce5c-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MLX Serve build; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 66w9z1j8, 8mrvg0xa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 986.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-Serve-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 986.8 29.4 18.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4sd0ce5c",
   "source_url": "https://omlx.ai/benchmarks/performance/4sd0ce5c",
   "tg_tokens": null,
   "tps": 29.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4sd0ce5c-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MLX Serve build; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: 66w9z1j8, 8mrvg0xa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 877.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-Serve-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 877.6 28.0 19.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4sd0ce5c",
   "source_url": "https://omlx.ai/benchmarks/performance/4sd0ce5c",
   "tg_tokens": null,
   "tps": 28.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4sd0ce5c-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MLX Serve build; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: 66w9z1j8, 8mrvg0xa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 673.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-Serve-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 673.8 26.0 22.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4sd0ce5c",
   "source_url": "https://omlx.ai/benchmarks/performance/4sd0ce5c",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-r15f0n84-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 408.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 408.3 53.6 21.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r15f0n84",
   "source_url": "https://omlx.ai/benchmarks/performance/r15f0n84",
   "tg_tokens": null,
   "tps": 53.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-r15f0n84-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 442.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 442.6 72.0 24.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r15f0n84",
   "source_url": "https://omlx.ai/benchmarks/performance/r15f0n84",
   "tg_tokens": null,
   "tps": 72.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-r15f0n84-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 517.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 517.0 100.0 25.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r15f0n84",
   "source_url": "https://omlx.ai/benchmarks/performance/r15f0n84",
   "tg_tokens": null,
   "tps": 100.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-r15f0n84-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 511.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 511.9 88.9 25.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r15f0n84",
   "source_url": "https://omlx.ai/benchmarks/performance/r15f0n84",
   "tg_tokens": null,
   "tps": 88.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-zwmtp8xg-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: c6o9i575; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 126.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 126.9 21.5 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zwmtp8xg",
   "source_url": "https://omlx.ai/benchmarks/performance/zwmtp8xg",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-zwmtp8xg-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: c6o9i575; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 151.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 151.0 27.4 24.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zwmtp8xg",
   "source_url": "https://omlx.ai/benchmarks/performance/zwmtp8xg",
   "tg_tokens": null,
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-jzjhnbo1-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 4b3kgtsy; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 128.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 128.6 29.8 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jzjhnbo1",
   "source_url": "https://omlx.ai/benchmarks/performance/jzjhnbo1",
   "tg_tokens": null,
   "tps": 29.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-jzjhnbo1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: 4b3kgtsy; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 147.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 147.5 31.7 24.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jzjhnbo1",
   "source_url": "https://omlx.ai/benchmarks/performance/jzjhnbo1",
   "tg_tokens": null,
   "tps": 31.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lbt61ael-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 70.9 GB; TTFT 2223 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1843.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 1,843 PP tok/s 66.1 TG tok/s 2223 TTFT (ms) 70.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lbt61ael",
   "source_url": "https://omlx.ai/benchmarks/performance/lbt61ael",
   "tg_tokens": null,
   "tps": 66.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-xevosmna-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: bxyuohpl, g9hc74yn, vstdwdd4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3028.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 3,028 134.2 100.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xevosmna",
   "source_url": "https://omlx.ai/benchmarks/performance/xevosmna",
   "tg_tokens": null,
   "tps": 134.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-xevosmna-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.5 GB at this context; duplicate traces of the same session: bxyuohpl, g9hc74yn, vstdwdd4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5610.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 5,610 169.5 102.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xevosmna",
   "source_url": "https://omlx.ai/benchmarks/performance/xevosmna",
   "tg_tokens": null,
   "tps": 169.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-xevosmna-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.3 GB at this context; duplicate traces of the same session: bxyuohpl, g9hc74yn, vstdwdd4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5491.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 5,491 141.3 104.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xevosmna",
   "source_url": "https://omlx.ai/benchmarks/performance/xevosmna",
   "tg_tokens": null,
   "tps": 141.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-xevosmna-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 108.8 GB at this context; duplicate traces of the same session: bxyuohpl, g9hc74yn, vstdwdd4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5298.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,298 114.1 108.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xevosmna",
   "source_url": "https://omlx.ai/benchmarks/performance/xevosmna",
   "tg_tokens": null,
   "tps": 114.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-eq9qxhg0-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace, single run at 16k context; peak memory 24.9 GB; TTFT 20618 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 794.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k tokens 794.6 PP tok/s 24.8 TG tok/s 20618 TTFT (ms) 24.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eq9qxhg0",
   "source_url": "https://omlx.ai/benchmarks/performance/eq9qxhg0",
   "tg_tokens": null,
   "tps": 24.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-thcrj191-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace, single run at 16k context; peak memory 24.9 GB; TTFT 19911 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 822.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k tokens 822.9 PP tok/s 37.2 TG tok/s 19911 TTFT (ms) 24.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/thcrj191",
   "source_url": "https://omlx.ai/benchmarks/performance/thcrj191",
   "tg_tokens": null,
   "tps": 37.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ddtcnqs0-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace, single run at 16k context; peak memory 24.8 GB; TTFT 25382 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 645.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k tokens 645.5 PP tok/s 19.0 TG tok/s 25382 TTFT (ms) 24.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ddtcnqs0",
   "source_url": "https://omlx.ai/benchmarks/performance/ddtcnqs0",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4tq2ou47-ctx16384",
   "model": "clef",
   "notes": "oMLX community benchmark trace; peak memory 38.2 GB at this context; duplicate traces of the same session: 47pvpseq, vqlimanr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 430.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "clef-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 16k 430.9 21.3 38.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4tq2ou47",
   "source_url": "https://omlx.ai/benchmarks/performance/4tq2ou47",
   "tg_tokens": null,
   "tps": 21.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4tq2ou47-ctx65536",
   "model": "clef",
   "notes": "oMLX community benchmark trace; peak memory 42.3 GB at this context; duplicate traces of the same session: 47pvpseq, vqlimanr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 333.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "clef-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 64k 333.7 18.6 42.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4tq2ou47",
   "source_url": "https://omlx.ai/benchmarks/performance/4tq2ou47",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4tq2ou47-ctx131072",
   "model": "clef",
   "notes": "oMLX community benchmark trace; peak memory 46.3 GB at this context; duplicate traces of the same session: 47pvpseq, vqlimanr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 257.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "clef-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 128k 257.3 15.9 46.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4tq2ou47",
   "source_url": "https://omlx.ai/benchmarks/performance/4tq2ou47",
   "tg_tokens": null,
   "tps": 15.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xapm0e25-ctx1024",
   "model": "clef",
   "notes": "generic trace name; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: owb3unmi; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 334.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "clef-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 334.5 37.5 22.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xapm0e25",
   "source_url": "https://omlx.ai/benchmarks/performance/xapm0e25",
   "tg_tokens": null,
   "tps": 37.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xapm0e25-ctx4096",
   "model": "clef",
   "notes": "generic trace name; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: owb3unmi; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 358.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "clef-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 358.1 36.2 27.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xapm0e25",
   "source_url": "https://omlx.ai/benchmarks/performance/xapm0e25",
   "tg_tokens": null,
   "tps": 36.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-yobpfrqv-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 15.2 GB; TTFT 4335 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3780.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k tokens 3,780 PP tok/s 100.5 TG tok/s 4335 TTFT (ms) 15.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yobpfrqv",
   "source_url": "https://omlx.ai/benchmarks/performance/yobpfrqv",
   "tg_tokens": null,
   "tps": 100.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qptvc2f6-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 15.2 GB; TTFT 4400 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3723.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k tokens 3,723 PP tok/s 95.1 TG tok/s 4400 TTFT (ms) 15.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qptvc2f6",
   "source_url": "https://omlx.ai/benchmarks/performance/qptvc2f6",
   "tg_tokens": null,
   "tps": 95.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zsca3u39-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 17.1 GB; TTFT 4915 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3334.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k tokens 3,334 PP tok/s 89.0 TG tok/s 4915 TTFT (ms) 17.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zsca3u39",
   "source_url": "https://omlx.ai/benchmarks/performance/zsca3u39",
   "tg_tokens": null,
   "tps": 89.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-220xhlvm-ctx1024",
   "model": "Hemmingway-1",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: iotkn1w1; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 82.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Hemmingway-1-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 1k 82.3 16.5 17.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/220xhlvm",
   "source_url": "https://omlx.ai/benchmarks/performance/220xhlvm",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-220xhlvm-ctx4096",
   "model": "Hemmingway-1",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: iotkn1w1; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 79.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Hemmingway-1-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k 79.8 14.6 18.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/220xhlvm",
   "source_url": "https://omlx.ai/benchmarks/performance/220xhlvm",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-vwx4dq38-ctx4096",
   "model": "Muse Glimmer",
   "notes": "Abliterated MM build; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: mpyk2m5z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 137.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Muse-Glimmer-30B-Abliterated-MM-oQ5e on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 5bit; 4k 137.4 11.3 19.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vwx4dq38",
   "source_url": "https://omlx.ai/benchmarks/performance/vwx4dq38",
   "tg_tokens": null,
   "tps": 11.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-vwx4dq38-ctx16384",
   "model": "Muse Glimmer",
   "notes": "Abliterated MM build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: mpyk2m5z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 146.3,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Muse-Glimmer-30B-Abliterated-MM-oQ5e on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 5bit; 16k 146.3 7.7 19.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vwx4dq38",
   "source_url": "https://omlx.ai/benchmarks/performance/vwx4dq38",
   "tg_tokens": null,
   "tps": 7.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rm0y1zq4-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: v22mx3d3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4593.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,593 100.8 21.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rm0y1zq4",
   "source_url": "https://omlx.ai/benchmarks/performance/rm0y1zq4",
   "tg_tokens": null,
   "tps": 100.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rm0y1zq4-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: v22mx3d3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5947.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 5,947 96.1 22.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rm0y1zq4",
   "source_url": "https://omlx.ai/benchmarks/performance/rm0y1zq4",
   "tg_tokens": null,
   "tps": 96.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-r0a7tg12-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 64k context; peak memory 21.9 GB; TTFT 18364 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3569.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k tokens 3,569 PP tok/s 94.0 TG tok/s 18364 TTFT (ms) 21.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r0a7tg12",
   "source_url": "https://omlx.ai/benchmarks/performance/r0a7tg12",
   "tg_tokens": null,
   "tps": 94.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dnlfe65t-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 64k context; peak memory 21.9 GB; TTFT 17815 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3679.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k tokens 3,679 PP tok/s 96.1 TG tok/s 17815 TTFT (ms) 21.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dnlfe65t",
   "source_url": "https://omlx.ai/benchmarks/performance/dnlfe65t",
   "tg_tokens": null,
   "tps": 96.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-39x3qeoo-ctx4096",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 46doc7ga; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4597.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 4,597 113.0 20.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/39x3qeoo",
   "source_url": "https://omlx.ai/benchmarks/performance/39x3qeoo",
   "tg_tokens": null,
   "tps": 113.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-39x3qeoo-ctx16384",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: 46doc7ga; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4123.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 4,123 97.5 20.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/39x3qeoo",
   "source_url": "https://omlx.ai/benchmarks/performance/39x3qeoo",
   "tg_tokens": null,
   "tps": 97.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-a16k5jkm-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 9.9 GB at this context; duplicate traces of the same session: j8gdg641; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 194.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B-MLX-8bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 8bit; 1k 194.9 19.5 9.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a16k5jkm",
   "source_url": "https://omlx.ai/benchmarks/performance/a16k5jkm",
   "tg_tokens": null,
   "tps": 19.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-a16k5jkm-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 10.6 GB at this context; duplicate traces of the same session: j8gdg641; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 199.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B-MLX-8bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 8bit; 4k 199.9 18.7 10.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a16k5jkm",
   "source_url": "https://omlx.ai/benchmarks/performance/a16k5jkm",
   "tg_tokens": null,
   "tps": 18.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-zt50xlsm-ctx4096",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 6cu005zv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1625.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-text-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,625 48.8 20.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zt50xlsm",
   "source_url": "https://omlx.ai/benchmarks/performance/zt50xlsm",
   "tg_tokens": null,
   "tps": 48.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-zt50xlsm-ctx8192",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 6cu005zv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1598.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-text-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 1,598 47.4 20.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zt50xlsm",
   "source_url": "https://omlx.ai/benchmarks/performance/zt50xlsm",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-zt50xlsm-ctx16384",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 6cu005zv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1422.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-text-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 1,422 43.2 21.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zt50xlsm",
   "source_url": "https://omlx.ai/benchmarks/performance/zt50xlsm",
   "tg_tokens": null,
   "tps": 43.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-rdp7nk7a-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace, single run at 16k context; peak memory 30.8 GB; TTFT 25046 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 654.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 5bit; 16k tokens 654.2 PP tok/s 69.7 TG tok/s 25046 TTFT (ms) 30.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rdp7nk7a",
   "source_url": "https://omlx.ai/benchmarks/performance/rdp7nk7a",
   "tg_tokens": null,
   "tps": 69.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-yjvsfy6t-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace, single run at 16k context; peak memory 30.8 GB; TTFT 25209 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 649.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 5bit; 16k tokens 649.9 PP tok/s 64.4 TG tok/s 25209 TTFT (ms) 30.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yjvsfy6t",
   "source_url": "https://omlx.ai/benchmarks/performance/yjvsfy6t",
   "tg_tokens": null,
   "tps": 64.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-3z2egm6t-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 3o9dyvg5; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 207.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 207.7 22.3 16.0 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3z2egm6t",
   "source_url": "https://omlx.ai/benchmarks/performance/3z2egm6t",
   "tg_tokens": null,
   "tps": 22.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-3z2egm6t-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 3o9dyvg5; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 215.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 215.3 21.5 17.0 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3z2egm6t",
   "source_url": "https://omlx.ai/benchmarks/performance/3z2egm6t",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ryq38ptz-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: d266i9hi; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 709.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 709.0 17.5 18.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ryq38ptz",
   "source_url": "https://omlx.ai/benchmarks/performance/ryq38ptz",
   "tg_tokens": null,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ryq38ptz-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: d266i9hi; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 727.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 727.2 16.5 19.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ryq38ptz",
   "source_url": "https://omlx.ai/benchmarks/performance/ryq38ptz",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-g4txsaqq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: gaub39gr; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 616.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 616.4 18.3 17.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g4txsaqq",
   "source_url": "https://omlx.ai/benchmarks/performance/g4txsaqq",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-g4txsaqq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: gaub39gr; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 731.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 731.5 18.0 18.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g4txsaqq",
   "source_url": "https://omlx.ai/benchmarks/performance/g4txsaqq",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-xj0xliqm-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: vtxqfyz9; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 616.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 616.6 18.3 17.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xj0xliqm",
   "source_url": "https://omlx.ai/benchmarks/performance/xj0xliqm",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-xj0xliqm-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: vtxqfyz9; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 734.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 734.1 17.7 18.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xj0xliqm",
   "source_url": "https://omlx.ai/benchmarks/performance/xj0xliqm",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-v63k8t0b-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: rmeujws1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 618.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 618.0 18.2 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v63k8t0b",
   "source_url": "https://omlx.ai/benchmarks/performance/v63k8t0b",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-v63k8t0b-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: rmeujws1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 733.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 733.8 18.0 18.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v63k8t0b",
   "source_url": "https://omlx.ai/benchmarks/performance/v63k8t0b",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-jljupl4r-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 1673zeo6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 617.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 617.8 18.3 17.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jljupl4r",
   "source_url": "https://omlx.ai/benchmarks/performance/jljupl4r",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-jljupl4r-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 1673zeo6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 725.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 725.8 17.9 18.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jljupl4r",
   "source_url": "https://omlx.ai/benchmarks/performance/jljupl4r",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-0zdminak-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: s2sh0yhu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 217.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 217.2 21.8 15.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0zdminak",
   "source_url": "https://omlx.ai/benchmarks/performance/0zdminak",
   "tg_tokens": null,
   "tps": 21.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-0zdminak-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: s2sh0yhu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 156.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 156.2 21.1 17.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0zdminak",
   "source_url": "https://omlx.ai/benchmarks/performance/0zdminak",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-yvleninv-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: usrxh3g2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 131.4 15.1 16.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yvleninv",
   "source_url": "https://omlx.ai/benchmarks/performance/yvleninv",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-yvleninv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: usrxh3g2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 137.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 137.9 14.8 19.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yvleninv",
   "source_url": "https://omlx.ai/benchmarks/performance/yvleninv",
   "tg_tokens": null,
   "tps": 14.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-2v1zqbnv-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: oqimwboq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 128.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 128.6 27.1 18.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2v1zqbnv",
   "source_url": "https://omlx.ai/benchmarks/performance/2v1zqbnv",
   "tg_tokens": null,
   "tps": 27.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-2v1zqbnv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: oqimwboq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 135.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 135.2 21.4 22.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2v1zqbnv",
   "source_url": "https://omlx.ai/benchmarks/performance/2v1zqbnv",
   "tg_tokens": null,
   "tps": 21.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-w6kk2ayx-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: fsu36yho; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 128.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 128.2 26.0 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w6kk2ayx",
   "source_url": "https://omlx.ai/benchmarks/performance/w6kk2ayx",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-w6kk2ayx-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: fsu36yho; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 149.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 149.5 17.3 24.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w6kk2ayx",
   "source_url": "https://omlx.ai/benchmarks/performance/w6kk2ayx",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-kp7qb7t6-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: zg8j3ylg, zkk5zs9c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 717.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 717.0 55.1 71.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kp7qb7t6",
   "source_url": "https://omlx.ai/benchmarks/performance/kp7qb7t6",
   "tg_tokens": null,
   "tps": 55.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-kp7qb7t6-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.2 GB at this context; duplicate traces of the same session: zg8j3ylg, zkk5zs9c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 787.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 787.9 54.6 78.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kp7qb7t6",
   "source_url": "https://omlx.ai/benchmarks/performance/kp7qb7t6",
   "tg_tokens": null,
   "tps": 54.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-kp7qb7t6-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.9 GB at this context; duplicate traces of the same session: zg8j3ylg, zkk5zs9c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 777.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 777.2 51.9 79.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kp7qb7t6",
   "source_url": "https://omlx.ai/benchmarks/performance/kp7qb7t6",
   "tg_tokens": null,
   "tps": 51.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-j5jhyzvz-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.4 GB at this context; duplicate traces of the same session: hs1off48; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2490.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 2,490 129.4 100.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j5jhyzvz",
   "source_url": "https://omlx.ai/benchmarks/performance/j5jhyzvz",
   "tg_tokens": null,
   "tps": 129.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-j5jhyzvz-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.3 GB at this context; duplicate traces of the same session: hs1off48; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4366.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 4,366 93.9 110.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j5jhyzvz",
   "source_url": "https://omlx.ai/benchmarks/performance/j5jhyzvz",
   "tg_tokens": null,
   "tps": 93.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-95f8wekg-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.5 GB at this context; duplicate traces of the same session: u6re15b1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5755.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 5,755 168.9 102.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/95f8wekg",
   "source_url": "https://omlx.ai/benchmarks/performance/95f8wekg",
   "tg_tokens": null,
   "tps": 168.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-95f8wekg-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.3 GB at this context; duplicate traces of the same session: u6re15b1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5782.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 5,782 124.6 104.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/95f8wekg",
   "source_url": "https://omlx.ai/benchmarks/performance/95f8wekg",
   "tg_tokens": null,
   "tps": 124.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-95f8wekg-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 108.9 GB at this context; duplicate traces of the same session: u6re15b1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5561.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,561 113.1 108.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/95f8wekg",
   "source_url": "https://omlx.ai/benchmarks/performance/95f8wekg",
   "tg_tokens": null,
   "tps": 113.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-95f8wekg-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.7 GB at this context; duplicate traces of the same session: u6re15b1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5389.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 5,389 101.3 110.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/95f8wekg",
   "source_url": "https://omlx.ai/benchmarks/performance/95f8wekg",
   "tg_tokens": null,
   "tps": 101.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-gvk2ubva-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.5 GB at this context; duplicate traces of the same session: ftxdut6p; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4976.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 4,976 142.4 102.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gvk2ubva",
   "source_url": "https://omlx.ai/benchmarks/performance/gvk2ubva",
   "tg_tokens": null,
   "tps": 142.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-gvk2ubva-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.3 GB at this context; duplicate traces of the same session: ftxdut6p; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5712.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 5,712 131.6 104.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gvk2ubva",
   "source_url": "https://omlx.ai/benchmarks/performance/gvk2ubva",
   "tg_tokens": null,
   "tps": 131.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-gvk2ubva-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.0 GB at this context; duplicate traces of the same session: ftxdut6p; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5522.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,522 110.8 109.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gvk2ubva",
   "source_url": "https://omlx.ai/benchmarks/performance/gvk2ubva",
   "tg_tokens": null,
   "tps": 110.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-gvk2ubva-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.7 GB at this context; duplicate traces of the same session: ftxdut6p; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5376.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 5,376 89.3 110.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gvk2ubva",
   "source_url": "https://omlx.ai/benchmarks/performance/gvk2ubva",
   "tg_tokens": null,
   "tps": 89.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-l6phnrvd-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: tm3vzsig; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 761.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 761.8 58.4 14.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l6phnrvd",
   "source_url": "https://omlx.ai/benchmarks/performance/l6phnrvd",
   "tg_tokens": null,
   "tps": 58.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-l6phnrvd-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.7 GB at this context; duplicate traces of the same session: tm3vzsig; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 715.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 715.6 57.0 14.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l6phnrvd",
   "source_url": "https://omlx.ai/benchmarks/performance/l6phnrvd",
   "tg_tokens": null,
   "tps": 57.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ojtv6z3u-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 17.1 GB; TTFT 6487 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2526.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k tokens 2,526 PP tok/s 89.2 TG tok/s 6487 TTFT (ms) 17.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ojtv6z3u",
   "source_url": "https://omlx.ai/benchmarks/performance/ojtv6z3u",
   "tg_tokens": null,
   "tps": 89.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9bc3e8lp-ctx32768",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 16.0 GB; TTFT 11163 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2935.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k tokens 2,935 PP tok/s 79.2 TG tok/s 11163 TTFT (ms) 16.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9bc3e8lp",
   "source_url": "https://omlx.ai/benchmarks/performance/9bc3e8lp",
   "tg_tokens": null,
   "tps": 79.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-aov0rzra-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: gnny969u; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 929.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 929.0 92.8 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aov0rzra",
   "source_url": "https://omlx.ai/benchmarks/performance/aov0rzra",
   "tg_tokens": null,
   "tps": 92.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-aov0rzra-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: gnny969u; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1040.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,040 90.2 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aov0rzra",
   "source_url": "https://omlx.ai/benchmarks/performance/aov0rzra",
   "tg_tokens": null,
   "tps": 90.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-aov0rzra-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: gnny969u; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 927.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 927.1 83.5 21.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aov0rzra",
   "source_url": "https://omlx.ai/benchmarks/performance/aov0rzra",
   "tg_tokens": null,
   "tps": 83.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-rbcjea7h-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 910.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 910.4 92.9 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rbcjea7h",
   "source_url": "https://omlx.ai/benchmarks/performance/rbcjea7h",
   "tg_tokens": null,
   "tps": 92.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-rbcjea7h-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1082.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,082 88.5 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rbcjea7h",
   "source_url": "https://omlx.ai/benchmarks/performance/rbcjea7h",
   "tg_tokens": null,
   "tps": 88.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-rbcjea7h-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1007.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 1,007 84.0 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rbcjea7h",
   "source_url": "https://omlx.ai/benchmarks/performance/rbcjea7h",
   "tg_tokens": null,
   "tps": 84.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-zpzshrqz-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: 769herbw, cogcf1i4, hya6ekq7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 855.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 855.3 59.6 35.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zpzshrqz",
   "source_url": "https://omlx.ai/benchmarks/performance/zpzshrqz",
   "tg_tokens": null,
   "tps": 59.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-zpzshrqz-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.6 GB at this context; duplicate traces of the same session: 769herbw, cogcf1i4, hya6ekq7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 979.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 979.9 52.8 37.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zpzshrqz",
   "source_url": "https://omlx.ai/benchmarks/performance/zpzshrqz",
   "tg_tokens": null,
   "tps": 52.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-zpzshrqz-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 38.1 GB at this context; duplicate traces of the same session: 769herbw, cogcf1i4, hya6ekq7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 839.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 16k 839.4 47.3 38.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zpzshrqz",
   "source_url": "https://omlx.ai/benchmarks/performance/zpzshrqz",
   "tg_tokens": null,
   "tps": 47.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-zpzshrqz-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 39.0 GB at this context; duplicate traces of the same session: 769herbw, cogcf1i4, hya6ekq7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 653.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 32k 653.5 39.4 39.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zpzshrqz",
   "source_url": "https://omlx.ai/benchmarks/performance/zpzshrqz",
   "tg_tokens": null,
   "tps": 39.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-dxmskmar-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 776.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 776.2 68.5 20.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dxmskmar",
   "source_url": "https://omlx.ai/benchmarks/performance/dxmskmar",
   "tg_tokens": null,
   "tps": 68.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-dxmskmar-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 837.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 837.3 66.1 21.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dxmskmar",
   "source_url": "https://omlx.ai/benchmarks/performance/dxmskmar",
   "tg_tokens": null,
   "tps": 66.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ni3sigtg-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: k3mcgygn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 111.8,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 111.8 18.6 21.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ni3sigtg",
   "source_url": "https://omlx.ai/benchmarks/performance/ni3sigtg",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ni3sigtg-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: k3mcgygn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 124.2,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 124.2 19.9 22.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ni3sigtg",
   "source_url": "https://omlx.ai/benchmarks/performance/ni3sigtg",
   "tg_tokens": null,
   "tps": 19.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qj58ss53-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: 3r817r16, 4qtls08e, 4xous7i8, 533cnz6d, 5a6xfy25, qsupq1k6, v1n0g50r; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 494.2,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 494.2 14.1 16.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qj58ss53",
   "source_url": "https://omlx.ai/benchmarks/performance/qj58ss53",
   "tg_tokens": null,
   "tps": 14.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qj58ss53-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 3r817r16, 4qtls08e, 4xous7i8, 533cnz6d, 5a6xfy25, qsupq1k6, v1n0g50r; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 515.4,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 515.4 14.1 17.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qj58ss53",
   "source_url": "https://omlx.ai/benchmarks/performance/qj58ss53",
   "tg_tokens": null,
   "tps": 14.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qj58ss53-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 3r817r16, 4qtls08e, 4xous7i8, 533cnz6d, 5a6xfy25, qsupq1k6, v1n0g50r; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 517.1,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 517.1 11.8 17.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qj58ss53",
   "source_url": "https://omlx.ai/benchmarks/performance/qj58ss53",
   "tg_tokens": null,
   "tps": 11.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qj58ss53-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: 3r817r16, 4qtls08e, 4xous7i8, 533cnz6d, 5a6xfy25, qsupq1k6, v1n0g50r; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 486.5,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 486.5 13.4 18.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qj58ss53",
   "source_url": "https://omlx.ai/benchmarks/performance/qj58ss53",
   "tg_tokens": null,
   "tps": 13.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qj58ss53-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 3r817r16, 4qtls08e, 4xous7i8, 533cnz6d, 5a6xfy25, qsupq1k6, v1n0g50r; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 464.3,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 464.3 14.5 19.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qj58ss53",
   "source_url": "https://omlx.ai/benchmarks/performance/qj58ss53",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qj58ss53-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 3r817r16, 4qtls08e, 4xous7i8, 533cnz6d, 5a6xfy25, qsupq1k6, v1n0g50r; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 404.7,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 404.7 13.6 21.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qj58ss53",
   "source_url": "https://omlx.ai/benchmarks/performance/qj58ss53",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qj58ss53-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: 3r817r16, 4qtls08e, 4xous7i8, 533cnz6d, 5a6xfy25, qsupq1k6, v1n0g50r; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.5,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 128k 323.5 12.7 25.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qj58ss53",
   "source_url": "https://omlx.ai/benchmarks/performance/qj58ss53",
   "tg_tokens": null,
   "tps": 12.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qj58ss53-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.3 GB at this context; duplicate traces of the same session: 3r817r16, 4qtls08e, 4xous7i8, 533cnz6d, 5a6xfy25, qsupq1k6, v1n0g50r; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 252.2,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 195k 252.2 11.3 29.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qj58ss53",
   "source_url": "https://omlx.ai/benchmarks/performance/qj58ss53",
   "tg_tokens": null,
   "tps": 11.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-hrj0whrq-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.6 GB at this context; duplicate traces of the same session: fujez7ht; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 342.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 342.3 78.9 30.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hrj0whrq",
   "source_url": "https://omlx.ai/benchmarks/performance/hrj0whrq",
   "tg_tokens": null,
   "tps": 78.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hrj0whrq-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.2 GB at this context; duplicate traces of the same session: fujez7ht; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 330.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 330.1 83.0 32.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hrj0whrq",
   "source_url": "https://omlx.ai/benchmarks/performance/hrj0whrq",
   "tg_tokens": null,
   "tps": 83.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rnp8a43y-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: hou7glh8; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 443.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 443.4 74.7 25.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rnp8a43y",
   "source_url": "https://omlx.ai/benchmarks/performance/rnp8a43y",
   "tg_tokens": null,
   "tps": 74.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rnp8a43y-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: hou7glh8; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 423.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 423.5 78.9 26.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rnp8a43y",
   "source_url": "https://omlx.ai/benchmarks/performance/rnp8a43y",
   "tg_tokens": null,
   "tps": 78.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fwrvssp7-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 1f0y1i7c, ahc9y2rs, g8pg6pld; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 933.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 933.3 31.2 17.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fwrvssp7",
   "source_url": "https://omlx.ai/benchmarks/performance/fwrvssp7",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fwrvssp7-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: 1f0y1i7c, ahc9y2rs, g8pg6pld; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 823.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 823.3 29.4 19.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fwrvssp7",
   "source_url": "https://omlx.ai/benchmarks/performance/fwrvssp7",
   "tg_tokens": null,
   "tps": 29.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fwrvssp7-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.2 GB at this context; duplicate traces of the same session: 1f0y1i7c, ahc9y2rs, g8pg6pld; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 580.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 580.9 25.0 27.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fwrvssp7",
   "source_url": "https://omlx.ai/benchmarks/performance/fwrvssp7",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fwrvssp7-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 48.3 GB at this context; duplicate traces of the same session: 1f0y1i7c, ahc9y2rs, g8pg6pld; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 348.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 348.3 17.1 48.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fwrvssp7",
   "source_url": "https://omlx.ai/benchmarks/performance/fwrvssp7",
   "tg_tokens": null,
   "tps": 17.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-u1pl3xsx-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: thwfcyal; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 344.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 344.5 25.5 17.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u1pl3xsx",
   "source_url": "https://omlx.ai/benchmarks/performance/u1pl3xsx",
   "tg_tokens": null,
   "tps": 25.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-u1pl3xsx-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.6 GB at this context; duplicate traces of the same session: thwfcyal; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 296.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 296.4 17.2 27.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u1pl3xsx",
   "source_url": "https://omlx.ai/benchmarks/performance/u1pl3xsx",
   "tg_tokens": null,
   "tps": 17.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-9xbihb8b-ctx1024",
   "model": "Qwen3.8-9B-Distill",
   "notes": "oMLX community benchmark trace; peak memory 11.3 GB at this context; duplicate traces of the same session: d4ei7jx3; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 372.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-9B-Distill-MLX-8bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 8bit; 1k 372.7 27.1 11.3 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9xbihb8b",
   "source_url": "https://omlx.ai/benchmarks/performance/9xbihb8b",
   "tg_tokens": null,
   "tps": 27.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-9xbihb8b-ctx8192",
   "model": "Qwen3.8-9B-Distill",
   "notes": "oMLX community benchmark trace; peak memory 11.6 GB at this context; duplicate traces of the same session: d4ei7jx3; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 463.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-9B-Distill-MLX-8bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 8bit; 8k 463.8 26.0 11.6 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9xbihb8b",
   "source_url": "https://omlx.ai/benchmarks/performance/9xbihb8b",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-d7t7zezk-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 86.5 GB at this context; duplicate traces of the same session: utvkwud8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 855.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 855.0 63.5 86.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d7t7zezk",
   "source_url": "https://omlx.ai/benchmarks/performance/d7t7zezk",
   "tg_tokens": null,
   "tps": 63.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-d7t7zezk-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 86.7 GB at this context; duplicate traces of the same session: utvkwud8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1707.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 1,707 62.2 86.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d7t7zezk",
   "source_url": "https://omlx.ai/benchmarks/performance/d7t7zezk",
   "tg_tokens": null,
   "tps": 62.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-uvasfbd3-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace, single run at 4k context; peak memory 86.7 GB; TTFT 2471 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1658.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k tokens 1,658 PP tok/s 94.1 TG tok/s 2471 TTFT (ms) 86.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uvasfbd3",
   "source_url": "https://omlx.ai/benchmarks/performance/uvasfbd3",
   "tg_tokens": null,
   "tps": 94.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-s9ukhsfp-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: qh7py9z4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1162.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,162 70.3 70.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s9ukhsfp",
   "source_url": "https://omlx.ai/benchmarks/performance/s9ukhsfp",
   "tg_tokens": null,
   "tps": 70.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-s9ukhsfp-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: qh7py9z4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2136.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,136 72.9 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s9ukhsfp",
   "source_url": "https://omlx.ai/benchmarks/performance/s9ukhsfp",
   "tg_tokens": null,
   "tps": 72.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-sz2sj3a5-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: izojzybf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 163.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 163.7 14.9 22.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sz2sj3a5",
   "source_url": "https://omlx.ai/benchmarks/performance/sz2sj3a5",
   "tg_tokens": null,
   "tps": 14.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-sz2sj3a5-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: izojzybf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 275.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 275.9 12.6 23.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sz2sj3a5",
   "source_url": "https://omlx.ai/benchmarks/performance/sz2sj3a5",
   "tg_tokens": null,
   "tps": 12.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-atie4pgh-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.5 GB at this context; duplicate traces of the same session: kw164m40; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4976.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 4,976 142.4 102.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/atie4pgh",
   "source_url": "https://omlx.ai/benchmarks/performance/atie4pgh",
   "tg_tokens": null,
   "tps": 142.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-atie4pgh-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.3 GB at this context; duplicate traces of the same session: kw164m40; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5712.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 5,712 131.6 104.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/atie4pgh",
   "source_url": "https://omlx.ai/benchmarks/performance/atie4pgh",
   "tg_tokens": null,
   "tps": 131.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-atie4pgh-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.0 GB at this context; duplicate traces of the same session: kw164m40; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5522.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,522 110.8 109.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/atie4pgh",
   "source_url": "https://omlx.ai/benchmarks/performance/atie4pgh",
   "tg_tokens": null,
   "tps": 110.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-atie4pgh-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.7 GB at this context; duplicate traces of the same session: kw164m40; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5376.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 5,376 89.3 110.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/atie4pgh",
   "source_url": "https://omlx.ai/benchmarks/performance/atie4pgh",
   "tg_tokens": null,
   "tps": 89.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-6ed8tc2c-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.5 GB at this context; duplicate traces of the same session: gq8xeuzs; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5755.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 5,755 168.9 102.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ed8tc2c",
   "source_url": "https://omlx.ai/benchmarks/performance/6ed8tc2c",
   "tg_tokens": null,
   "tps": 168.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-6ed8tc2c-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.3 GB at this context; duplicate traces of the same session: gq8xeuzs; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5782.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 5,782 124.6 104.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ed8tc2c",
   "source_url": "https://omlx.ai/benchmarks/performance/6ed8tc2c",
   "tg_tokens": null,
   "tps": 124.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-6ed8tc2c-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 108.9 GB at this context; duplicate traces of the same session: gq8xeuzs; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5561.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,561 113.1 108.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ed8tc2c",
   "source_url": "https://omlx.ai/benchmarks/performance/6ed8tc2c",
   "tg_tokens": null,
   "tps": 113.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-6ed8tc2c-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.7 GB at this context; duplicate traces of the same session: gq8xeuzs; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5389.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 5,389 101.3 110.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ed8tc2c",
   "source_url": "https://omlx.ai/benchmarks/performance/6ed8tc2c",
   "tg_tokens": null,
   "tps": 101.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-25z0mili-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.2 GB at this context; duplicate traces of the same session: gcdcqctn, ov7p9vun, rifnjxv4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4818.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 4,818 135.8 102.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/25z0mili",
   "source_url": "https://omlx.ai/benchmarks/performance/25z0mili",
   "tg_tokens": null,
   "tps": 135.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-25z0mili-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.3 GB at this context; duplicate traces of the same session: gcdcqctn, ov7p9vun, rifnjxv4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5471.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 5,471 141.9 104.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/25z0mili",
   "source_url": "https://omlx.ai/benchmarks/performance/25z0mili",
   "tg_tokens": null,
   "tps": 141.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-25z0mili-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 108.8 GB at this context; duplicate traces of the same session: gcdcqctn, ov7p9vun, rifnjxv4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5298.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,298 111.5 108.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/25z0mili",
   "source_url": "https://omlx.ai/benchmarks/performance/25z0mili",
   "tg_tokens": null,
   "tps": 111.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-25z0mili-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.6 GB at this context; duplicate traces of the same session: gcdcqctn, ov7p9vun, rifnjxv4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5175.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 5,175 96.8 110.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/25z0mili",
   "source_url": "https://omlx.ai/benchmarks/performance/25z0mili",
   "tg_tokens": null,
   "tps": 96.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-oqdvrwu8-ctx4096",
   "model": "Qwythos-9B-v2",
   "notes": "MLX; oMLX community benchmark trace, single run at 4k context; peak memory 10.5 GB; TTFT 8058 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 508.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwythos-9B-v2-MLX-8bit on M2 Max (38c); M2 Max (38c) \u00b7 32 GB \u00b7 8bit; 4k tokens 508.3 PP tok/s 38.0 TG tok/s 8058 TTFT (ms) 10.5 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oqdvrwu8",
   "source_url": "https://omlx.ai/benchmarks/performance/oqdvrwu8",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ppj1n610-ctx4096",
   "model": "Qwythos-9B-v2",
   "notes": "MLX; oMLX community benchmark trace, single run at 4k context; peak memory 10.5 GB; TTFT 8065 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 507.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwythos-9B-v2-MLX-8bit on M2 Max (38c); M2 Max (38c) \u00b7 32 GB \u00b7 8bit; 4k tokens 507.9 PP tok/s 38.0 TG tok/s 8065 TTFT (ms) 10.5 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ppj1n610",
   "source_url": "https://omlx.ai/benchmarks/performance/ppj1n610",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-8phfggcu-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: 5bv9skma; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1991.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E4B-it-MLX-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 1,991 64.7 6.2 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8phfggcu",
   "source_url": "https://omlx.ai/benchmarks/performance/8phfggcu",
   "tg_tokens": null,
   "tps": 64.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-8phfggcu-ctx16384",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: 5bv9skma; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1932.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E4B-it-MLX-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 1,932 58.0 6.4 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8phfggcu",
   "source_url": "https://omlx.ai/benchmarks/performance/8phfggcu",
   "tg_tokens": null,
   "tps": 58.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-vq1sm80j-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace; peak memory 32.5 GB at this context; duplicate traces of the same session: 2gltif13; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 80.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ6e on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 6bit; 1k 80.2 9.1 32.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vq1sm80j",
   "source_url": "https://omlx.ai/benchmarks/performance/vq1sm80j",
   "tg_tokens": null,
   "tps": 9.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-vq1sm80j-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace; peak memory 36.6 GB at this context; duplicate traces of the same session: 2gltif13; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 75.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ6e on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 6bit; 4k 75.0 8.8 36.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vq1sm80j",
   "source_url": "https://omlx.ai/benchmarks/performance/vq1sm80j",
   "tg_tokens": null,
   "tps": 8.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-7ytipdgw-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: gdlt5h8i; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 910.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 910.4 92.9 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ytipdgw",
   "source_url": "https://omlx.ai/benchmarks/performance/7ytipdgw",
   "tg_tokens": null,
   "tps": 92.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-7ytipdgw-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: gdlt5h8i; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1082.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,082 88.5 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ytipdgw",
   "source_url": "https://omlx.ai/benchmarks/performance/7ytipdgw",
   "tg_tokens": null,
   "tps": 88.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-7ytipdgw-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: gdlt5h8i; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1007.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 1,007 84.0 21.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ytipdgw",
   "source_url": "https://omlx.ai/benchmarks/performance/7ytipdgw",
   "tg_tokens": null,
   "tps": 84.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-3q5oxmt0-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 929.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 929.0 92.8 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3q5oxmt0",
   "source_url": "https://omlx.ai/benchmarks/performance/3q5oxmt0",
   "tg_tokens": null,
   "tps": 92.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-3q5oxmt0-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1040.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,040 90.2 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3q5oxmt0",
   "source_url": "https://omlx.ai/benchmarks/performance/3q5oxmt0",
   "tg_tokens": null,
   "tps": 90.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-3q5oxmt0-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 927.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 927.1 83.5 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3q5oxmt0",
   "source_url": "https://omlx.ai/benchmarks/performance/3q5oxmt0",
   "tg_tokens": null,
   "tps": 83.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-7ogmj06y-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 3cvt7pac, ifvls0bk, muzbmzq1, y6hv5i3m; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 941.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 941.8 93.2 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ogmj06y",
   "source_url": "https://omlx.ai/benchmarks/performance/7ogmj06y",
   "tg_tokens": null,
   "tps": 93.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-7ogmj06y-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 3cvt7pac, ifvls0bk, muzbmzq1, y6hv5i3m; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1083.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,083 90.8 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ogmj06y",
   "source_url": "https://omlx.ai/benchmarks/performance/7ogmj06y",
   "tg_tokens": null,
   "tps": 90.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-7ogmj06y-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 3cvt7pac, ifvls0bk, muzbmzq1, y6hv5i3m; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1023.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 1,023 25.2 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ogmj06y",
   "source_url": "https://omlx.ai/benchmarks/performance/7ogmj06y",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-7ogmj06y-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: 3cvt7pac, ifvls0bk, muzbmzq1, y6hv5i3m; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 412.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 412.2 23.1 22.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ogmj06y",
   "source_url": "https://omlx.ai/benchmarks/performance/7ogmj06y",
   "tg_tokens": null,
   "tps": 23.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-7ogmj06y-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 3cvt7pac, ifvls0bk, muzbmzq1, y6hv5i3m; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 334.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 334.7 37.5 23.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ogmj06y",
   "source_url": "https://omlx.ai/benchmarks/performance/7ogmj06y",
   "tg_tokens": null,
   "tps": 37.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-5bzgkiy6-ctx4096",
   "model": "Qwen2.5-Coder-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: 2v8zkj9s; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1193.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-7B-Instruct-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,193 33.9 4.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5bzgkiy6",
   "source_url": "https://omlx.ai/benchmarks/performance/5bzgkiy6",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-5bzgkiy6-ctx16384",
   "model": "Qwen2.5-Coder-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 5.5 GB at this context; duplicate traces of the same session: 2v8zkj9s; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1038.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-7B-Instruct-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 1,038 28.5 5.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5bzgkiy6",
   "source_url": "https://omlx.ai/benchmarks/performance/5bzgkiy6",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-bnl7oiwd-ctx4096",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "MLX; oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: mcxbpprq; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1716.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,716 60.1 17.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bnl7oiwd",
   "source_url": "https://omlx.ai/benchmarks/performance/bnl7oiwd",
   "tg_tokens": null,
   "tps": 60.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-bnl7oiwd-ctx16384",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "MLX; oMLX community benchmark trace; peak memory 18.0 GB at this context; duplicate traces of the same session: mcxbpprq; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1165.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 1,165 40.2 18.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bnl7oiwd",
   "source_url": "https://omlx.ai/benchmarks/performance/bnl7oiwd",
   "tg_tokens": null,
   "tps": 40.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-91ivcfyh-ctx1024",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: fcdsaald, rhruql0i, yt4nrx4l; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 483.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 483.3 17.3 15.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/91ivcfyh",
   "source_url": "https://omlx.ai/benchmarks/performance/91ivcfyh",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-91ivcfyh-ctx4096",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: fcdsaald, rhruql0i, yt4nrx4l; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 514.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 514.0 17.4 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/91ivcfyh",
   "source_url": "https://omlx.ai/benchmarks/performance/91ivcfyh",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-91ivcfyh-ctx8192",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: fcdsaald, rhruql0i, yt4nrx4l; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 492.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 492.4 16.5 17.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/91ivcfyh",
   "source_url": "https://omlx.ai/benchmarks/performance/91ivcfyh",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-91ivcfyh-ctx16384",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: fcdsaald, rhruql0i, yt4nrx4l; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 449.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 449.8 16.7 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/91ivcfyh",
   "source_url": "https://omlx.ai/benchmarks/performance/91ivcfyh",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-2wk332cl-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 776.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 776.2 68.5 20.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2wk332cl",
   "source_url": "https://omlx.ai/benchmarks/performance/2wk332cl",
   "tg_tokens": null,
   "tps": 68.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-2wk332cl-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 837.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 837.3 66.1 21.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2wk332cl",
   "source_url": "https://omlx.ai/benchmarks/performance/2wk332cl",
   "tg_tokens": null,
   "tps": 66.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-qj1x0ttv-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: xhpivk2d; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 749.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 749.5 105.4 21.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qj1x0ttv",
   "source_url": "https://omlx.ai/benchmarks/performance/qj1x0ttv",
   "tg_tokens": null,
   "tps": 105.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-qj1x0ttv-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: xhpivk2d; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 815.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 815.9 83.3 22.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qj1x0ttv",
   "source_url": "https://omlx.ai/benchmarks/performance/qj1x0ttv",
   "tg_tokens": null,
   "tps": 83.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-1and7b1z-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: r1q2mmzv; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 755.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 755.2 79.1 20.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1and7b1z",
   "source_url": "https://omlx.ai/benchmarks/performance/1and7b1z",
   "tg_tokens": null,
   "tps": 79.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-1and7b1z-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: r1q2mmzv; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 842.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 842.5 75.8 20.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1and7b1z",
   "source_url": "https://omlx.ai/benchmarks/performance/1and7b1z",
   "tg_tokens": null,
   "tps": 75.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2dj7j7c0-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: e7xlkrkd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 923.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 923.9 57.8 23.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2dj7j7c0",
   "source_url": "https://omlx.ai/benchmarks/performance/2dj7j7c0",
   "tg_tokens": null,
   "tps": 57.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2dj7j7c0-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.5 GB at this context; duplicate traces of the same session: e7xlkrkd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 570.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 570.4 32.2 32.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2dj7j7c0",
   "source_url": "https://omlx.ai/benchmarks/performance/2dj7j7c0",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-h6pbdw8y-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: j7k87lhn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1161.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,161 45.3 19.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h6pbdw8y",
   "source_url": "https://omlx.ai/benchmarks/performance/h6pbdw8y",
   "tg_tokens": null,
   "tps": 45.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-h6pbdw8y-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: j7k87lhn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1325.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,325 73.8 19.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h6pbdw8y",
   "source_url": "https://omlx.ai/benchmarks/performance/h6pbdw8y",
   "tg_tokens": null,
   "tps": 73.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-stehrry2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: tb91jwix; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 646.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 646.7 23.3 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/stehrry2",
   "source_url": "https://omlx.ai/benchmarks/performance/stehrry2",
   "tg_tokens": null,
   "tps": 23.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-stehrry2-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: tb91jwix; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 675.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 675.0 38.8 19.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/stehrry2",
   "source_url": "https://omlx.ai/benchmarks/performance/stehrry2",
   "tg_tokens": null,
   "tps": 38.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-kewk10am-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace, single run at 64k context; peak memory 115.6 GB; TTFT 68287 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 959.7,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ4-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k tokens 959.7 PP tok/s 63.8 TG tok/s 68287 TTFT (ms) 115.6 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kewk10am",
   "source_url": "https://omlx.ai/benchmarks/performance/kewk10am",
   "tg_tokens": null,
   "tps": 63.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-jm8410lm-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.9 GB at this context; duplicate traces of the same session: 1m2nbxco, 67zzz9n5, 8nt4rpyj, su6ygsn2, vanf8t9b; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2585.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 2,585 102.6 100.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jm8410lm",
   "source_url": "https://omlx.ai/benchmarks/performance/jm8410lm",
   "tg_tokens": null,
   "tps": 102.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-jm8410lm-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.0 GB at this context; duplicate traces of the same session: 1m2nbxco, 67zzz9n5, 8nt4rpyj, su6ygsn2, vanf8t9b; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4342.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 4,342 101.6 101.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jm8410lm",
   "source_url": "https://omlx.ai/benchmarks/performance/jm8410lm",
   "tg_tokens": null,
   "tps": 101.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-jm8410lm-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.1 GB at this context; duplicate traces of the same session: 1m2nbxco, 67zzz9n5, 8nt4rpyj, su6ygsn2, vanf8t9b; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4711.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 4,711 109.6 102.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jm8410lm",
   "source_url": "https://omlx.ai/benchmarks/performance/jm8410lm",
   "tg_tokens": null,
   "tps": 109.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-jm8410lm-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 105.3 GB at this context; duplicate traces of the same session: 1m2nbxco, 67zzz9n5, 8nt4rpyj, su6ygsn2, vanf8t9b; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4629.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 4,629 87.4 105.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jm8410lm",
   "source_url": "https://omlx.ai/benchmarks/performance/jm8410lm",
   "tg_tokens": null,
   "tps": 87.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-jm8410lm-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.5 GB at this context; duplicate traces of the same session: 1m2nbxco, 67zzz9n5, 8nt4rpyj, su6ygsn2, vanf8t9b; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4492.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 4,492 102.0 109.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jm8410lm",
   "source_url": "https://omlx.ai/benchmarks/performance/jm8410lm",
   "tg_tokens": null,
   "tps": 102.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-jm8410lm-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.7 GB at this context; duplicate traces of the same session: 1m2nbxco, 67zzz9n5, 8nt4rpyj, su6ygsn2, vanf8t9b; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4364.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 4,364 82.1 110.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jm8410lm",
   "source_url": "https://omlx.ai/benchmarks/performance/jm8410lm",
   "tg_tokens": null,
   "tps": 82.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-y0dufcm8-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.4 GB at this context; duplicate traces of the same session: 4d275ys7, gdujq99s, gs339ogs, xa1vn11l, y29iueol, yv1tvx4d; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2339.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 2,339 111.2 100.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y0dufcm8",
   "source_url": "https://omlx.ai/benchmarks/performance/y0dufcm8",
   "tg_tokens": null,
   "tps": 111.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-y0dufcm8-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.1 GB at this context; duplicate traces of the same session: 4d275ys7, gdujq99s, gs339ogs, xa1vn11l, y29iueol, yv1tvx4d; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4701.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 4,701 106.2 102.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y0dufcm8",
   "source_url": "https://omlx.ai/benchmarks/performance/y0dufcm8",
   "tg_tokens": null,
   "tps": 106.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-y0dufcm8-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.6 GB at this context; duplicate traces of the same session: 4d275ys7, gdujq99s, gs339ogs, xa1vn11l, y29iueol, yv1tvx4d; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4730.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 4,730 106.3 102.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y0dufcm8",
   "source_url": "https://omlx.ai/benchmarks/performance/y0dufcm8",
   "tg_tokens": null,
   "tps": 106.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-y0dufcm8-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.1 GB at this context; duplicate traces of the same session: 4d275ys7, gdujq99s, gs339ogs, xa1vn11l, y29iueol, yv1tvx4d; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4696.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 4,696 105.5 103.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y0dufcm8",
   "source_url": "https://omlx.ai/benchmarks/performance/y0dufcm8",
   "tg_tokens": null,
   "tps": 105.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-y0dufcm8-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.3 GB at this context; duplicate traces of the same session: 4d275ys7, gdujq99s, gs339ogs, xa1vn11l, y29iueol, yv1tvx4d; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4613.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 4,613 114.3 104.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y0dufcm8",
   "source_url": "https://omlx.ai/benchmarks/performance/y0dufcm8",
   "tg_tokens": null,
   "tps": 114.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-y0dufcm8-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 108.5 GB at this context; duplicate traces of the same session: 4d275ys7, gdujq99s, gs339ogs, xa1vn11l, y29iueol, yv1tvx4d; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4477.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 4,477 97.3 108.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y0dufcm8",
   "source_url": "https://omlx.ai/benchmarks/performance/y0dufcm8",
   "tg_tokens": null,
   "tps": 97.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-y0dufcm8-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.7 GB at this context; duplicate traces of the same session: 4d275ys7, gdujq99s, gs339ogs, xa1vn11l, y29iueol, yv1tvx4d; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4347.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 4,347 86.6 110.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y0dufcm8",
   "source_url": "https://omlx.ai/benchmarks/performance/y0dufcm8",
   "tg_tokens": null,
   "tps": 86.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-bhi3oq8e-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: 0xhyjict, 3vftdx0s, f9ql7m3g, oekn3lvu; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3536.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 4k 3,536 101.7 143.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhi3oq8e",
   "source_url": "https://omlx.ai/benchmarks/performance/bhi3oq8e",
   "tg_tokens": null,
   "tps": 101.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-bhi3oq8e-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 144.5 GB at this context; duplicate traces of the same session: 0xhyjict, 3vftdx0s, f9ql7m3g, oekn3lvu; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3971.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 8k 3,971 115.2 144.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhi3oq8e",
   "source_url": "https://omlx.ai/benchmarks/performance/bhi3oq8e",
   "tg_tokens": null,
   "tps": 115.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-bhi3oq8e-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 146.8 GB at this context; duplicate traces of the same session: 0xhyjict, 3vftdx0s, f9ql7m3g, oekn3lvu; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3908.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 64k 3,908 107.1 146.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhi3oq8e",
   "source_url": "https://omlx.ai/benchmarks/performance/bhi3oq8e",
   "tg_tokens": null,
   "tps": 107.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-bhi3oq8e-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 150.8 GB at this context; duplicate traces of the same session: 0xhyjict, 3vftdx0s, f9ql7m3g, oekn3lvu; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3811.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 128k 3,811 95.0 150.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhi3oq8e",
   "source_url": "https://omlx.ai/benchmarks/performance/bhi3oq8e",
   "tg_tokens": null,
   "tps": 95.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-bhi3oq8e-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 153.3 GB at this context; duplicate traces of the same session: 0xhyjict, 3vftdx0s, f9ql7m3g, oekn3lvu; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3712.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 195k 3,712 82.0 153.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhi3oq8e",
   "source_url": "https://omlx.ai/benchmarks/performance/bhi3oq8e",
   "tg_tokens": null,
   "tps": 82.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wpr2c8b8-ctx199680",
   "model": "gpt-oss 20B",
   "notes": "mlx-community; oMLX community benchmark trace, single run at 195k context; peak memory 17.4 GB; TTFT 126263 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1584.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "mlx-community--gpt-oss-20b-MXFP4-Q8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k tokens 1,584 PP tok/s 36.3 TG tok/s 126263 TTFT (ms) 17.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wpr2c8b8",
   "source_url": "https://omlx.ai/benchmarks/performance/wpr2c8b8",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-be7jqdno-ctx4096",
   "model": "DeepHermes-3-Llama-3-8B-Preview",
   "notes": "oMLX community benchmark trace; peak memory 5.4 GB at this context; duplicate traces of the same session: 2afea5dx; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1374.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepHermes-3-Llama-3-8B-Preview-4Bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 1,374 53.4 5.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/be7jqdno",
   "source_url": "https://omlx.ai/benchmarks/performance/be7jqdno",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-be7jqdno-ctx16384",
   "model": "DeepHermes-3-Llama-3-8B-Preview",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: 2afea5dx; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1205.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepHermes-3-Llama-3-8B-Preview-4Bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 1,205 41.2 6.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/be7jqdno",
   "source_url": "https://omlx.ai/benchmarks/performance/be7jqdno",
   "tg_tokens": null,
   "tps": 41.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-83ccfqj1-ctx4096",
   "model": "DeepHermes-3-Llama-3-8B-Preview",
   "notes": "oMLX community benchmark trace; peak memory 5.4 GB at this context; duplicate traces of the same session: lxhmtgd8; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1393.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepHermes-3-Llama-3-8B-Preview-4Bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 1,393 55.0 5.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/83ccfqj1",
   "source_url": "https://omlx.ai/benchmarks/performance/83ccfqj1",
   "tg_tokens": null,
   "tps": 55.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-83ccfqj1-ctx16384",
   "model": "DeepHermes-3-Llama-3-8B-Preview",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: lxhmtgd8; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1127.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepHermes-3-Llama-3-8B-Preview-4Bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 1,127 37.2 6.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/83ccfqj1",
   "source_url": "https://omlx.ai/benchmarks/performance/83ccfqj1",
   "tg_tokens": null,
   "tps": 37.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-zr5pq3d8-ctx16384",
   "model": "DeepSeek-Coder-V2-Lite-Instruct",
   "notes": "MLX; oMLX community benchmark trace, single run at 16k context; peak memory 14.9 GB; TTFT 38879 ms; duplicate traces of the same session: none; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 421.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-Coder-V2-Lite-Instruct-4bit-mlx on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 4bit; 16k tokens 421.4 PP tok/s 22.4 TG tok/s 38879 TTFT (ms) 14.9 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zr5pq3d8",
   "source_url": "https://omlx.ai/benchmarks/performance/zr5pq3d8",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-r1210zzd-ctx1024",
   "model": "DeepSeek-Coder-V2-Lite-Instruct",
   "notes": "MLX; oMLX community benchmark trace; peak memory 9.1 GB at this context; duplicate traces of the same session: zcw7i764; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 576.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-Coder-V2-Lite-Instruct-4bit-mlx on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 4bit; 1k 576.0 63.0 9.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r1210zzd",
   "source_url": "https://omlx.ai/benchmarks/performance/r1210zzd",
   "tg_tokens": null,
   "tps": 63.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-r1210zzd-ctx4096",
   "model": "DeepSeek-Coder-V2-Lite-Instruct",
   "notes": "MLX; oMLX community benchmark trace; peak memory 10.2 GB at this context; duplicate traces of the same session: zcw7i764; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 546.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-Coder-V2-Lite-Instruct-4bit-mlx on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 4bit; 4k 546.9 46.7 10.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r1210zzd",
   "source_url": "https://omlx.ai/benchmarks/performance/r1210zzd",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-0ugi2tp9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace; peak memory 32.7 GB at this context; duplicate traces of the same session: ok0xbc9h; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 81.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ6e on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 6bit; 1k 81.3 7.3 32.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ugi2tp9",
   "source_url": "https://omlx.ai/benchmarks/performance/0ugi2tp9",
   "tg_tokens": null,
   "tps": 7.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-0ugi2tp9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace; peak memory 36.7 GB at this context; duplicate traces of the same session: ok0xbc9h; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 76.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ6e on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 6bit; 4k 76.0 6.1 36.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ugi2tp9",
   "source_url": "https://omlx.ai/benchmarks/performance/0ugi2tp9",
   "tg_tokens": null,
   "tps": 6.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-q7f5zp9b-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.2 GB at this context; duplicate traces of the same session: 17gja1pf, nrbftzmq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2277.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,277 127.5 173.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q7f5zp9b",
   "source_url": "https://omlx.ai/benchmarks/performance/q7f5zp9b",
   "tg_tokens": null,
   "tps": 127.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-q7f5zp9b-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.7 GB at this context; duplicate traces of the same session: 17gja1pf, nrbftzmq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2251.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,251 92.6 173.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q7f5zp9b",
   "source_url": "https://omlx.ai/benchmarks/performance/q7f5zp9b",
   "tg_tokens": null,
   "tps": 92.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-q7f5zp9b-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.9 GB at this context; duplicate traces of the same session: 17gja1pf, nrbftzmq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2200.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,200 126.5 174.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q7f5zp9b",
   "source_url": "https://omlx.ai/benchmarks/performance/q7f5zp9b",
   "tg_tokens": null,
   "tps": 126.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-zkc3n4uo-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.1 GB at this context; duplicate traces of the same session: 0f7z50w2, yxaouzbc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2232.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,232 119.1 178.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zkc3n4uo",
   "source_url": "https://omlx.ai/benchmarks/performance/zkc3n4uo",
   "tg_tokens": null,
   "tps": 119.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-zkc3n4uo-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.6 GB at this context; duplicate traces of the same session: 0f7z50w2, yxaouzbc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2203.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,203 91.7 178.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zkc3n4uo",
   "source_url": "https://omlx.ai/benchmarks/performance/zkc3n4uo",
   "tg_tokens": null,
   "tps": 91.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-zkc3n4uo-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 179.9 GB at this context; duplicate traces of the same session: 0f7z50w2, yxaouzbc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2158.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,158 132.1 179.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zkc3n4uo",
   "source_url": "https://omlx.ai/benchmarks/performance/zkc3n4uo",
   "tg_tokens": null,
   "tps": 132.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mh9d7434-ctx4096",
   "model": "Hermes-4-70B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 71.9 GB at this context; duplicate traces of the same session: oyjxypqf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 299.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Hermes-4-70B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 299.8 7.0 71.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mh9d7434",
   "source_url": "https://omlx.ai/benchmarks/performance/mh9d7434",
   "tg_tokens": null,
   "tps": 7.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mh9d7434-ctx65536",
   "model": "Hermes-4-70B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 90.5 GB at this context; duplicate traces of the same session: oyjxypqf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 165.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Hermes-4-70B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 165.6 4.4 90.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mh9d7434",
   "source_url": "https://omlx.ai/benchmarks/performance/mh9d7434",
   "tg_tokens": null,
   "tps": 4.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-tyr2i19s-ctx8192",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace, single run at 8k context; peak memory 17.4 GB; TTFT 32100 ms; duplicate traces of the same session: none; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 255.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 4bit; 8k tokens 255.2 PP tok/s 26.7 TG tok/s 32100 TTFT (ms) 17.4 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tyr2i19s",
   "source_url": "https://omlx.ai/benchmarks/performance/tyr2i19s",
   "tg_tokens": null,
   "tps": 26.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-gbbq3n1n-ctx1024",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 16.6 GB; TTFT 3451 ms; duplicate traces of the same session: none; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 296.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 4bit; 1k tokens 296.7 PP tok/s 39.8 TG tok/s 3451 TTFT (ms) 16.6 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gbbq3n1n",
   "source_url": "https://omlx.ai/benchmarks/performance/gbbq3n1n",
   "tg_tokens": null,
   "tps": 39.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-cslkm7bl-ctx1024",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 16.6 GB; TTFT 3225 ms; duplicate traces of the same session: none; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 317.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 4bit; 1k tokens 317.5 PP tok/s 39.5 TG tok/s 3225 TTFT (ms) 16.6 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cslkm7bl",
   "source_url": "https://omlx.ai/benchmarks/performance/cslkm7bl",
   "tg_tokens": null,
   "tps": 39.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-1ccbfhr8-ctx4096",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: mkjxjle4; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 81.4,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-27B-OptiQ-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 81.4 12.1 20.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ccbfhr8",
   "source_url": "https://omlx.ai/benchmarks/performance/1ccbfhr8",
   "tg_tokens": null,
   "tps": 12.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-1ccbfhr8-ctx16384",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: mkjxjle4; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 77.4,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-27B-OptiQ-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 77.4 11.2 22.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ccbfhr8",
   "source_url": "https://omlx.ai/benchmarks/performance/1ccbfhr8",
   "tg_tokens": null,
   "tps": 11.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-yuep47f6-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: 7n19ciq0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1703.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 1,703 167.2 24.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yuep47f6",
   "source_url": "https://omlx.ai/benchmarks/performance/yuep47f6",
   "tg_tokens": null,
   "tps": 167.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-yuep47f6-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.6 GB at this context; duplicate traces of the same session: 7n19ciq0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1256.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 1,256 136.9 25.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yuep47f6",
   "source_url": "https://omlx.ai/benchmarks/performance/yuep47f6",
   "tg_tokens": null,
   "tps": 136.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rwgrohlq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: dvf206w1, gbbs0hge, pt5r1as3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 973.9,
   "provenance": "community",
   "quant": "AWQ-5.0bpw",
   "quote": "Qwen3.8-27B-AWQ-5.0bpw on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 4k 973.9 58.2 21.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rwgrohlq",
   "source_url": "https://omlx.ai/benchmarks/performance/rwgrohlq",
   "tg_tokens": null,
   "tps": 58.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-rwgrohlq-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: dvf206w1, gbbs0hge, pt5r1as3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 901.7,
   "provenance": "community",
   "quant": "AWQ-5.0bpw",
   "quote": "Qwen3.8-27B-AWQ-5.0bpw on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 901.7 75.9 22.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rwgrohlq",
   "source_url": "https://omlx.ai/benchmarks/performance/rwgrohlq",
   "tg_tokens": null,
   "tps": 75.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-rwgrohlq-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: dvf206w1, gbbs0hge, pt5r1as3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 694.0,
   "provenance": "community",
   "quant": "AWQ-5.0bpw",
   "quote": "Qwen3.8-27B-AWQ-5.0bpw on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 64k 694.0 55.0 25.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rwgrohlq",
   "source_url": "https://omlx.ai/benchmarks/performance/rwgrohlq",
   "tg_tokens": null,
   "tps": 55.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-rwgrohlq-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 35.4 GB at this context; duplicate traces of the same session: dvf206w1, gbbs0hge, pt5r1as3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 474.8,
   "provenance": "community",
   "quant": "AWQ-5.0bpw",
   "quote": "Qwen3.8-27B-AWQ-5.0bpw on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 195k 474.8 32.8 35.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rwgrohlq",
   "source_url": "https://omlx.ai/benchmarks/performance/rwgrohlq",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-98ztc97t-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 41673o4a, 4ko51tvl, p6zxretf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 996.6,
   "provenance": "community",
   "quant": "AWQ-5.0bpw",
   "quote": "Qwen3.8-27B-AWQ-5.0bpw on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 4k 996.6 66.8 21.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/98ztc97t",
   "source_url": "https://omlx.ai/benchmarks/performance/98ztc97t",
   "tg_tokens": null,
   "tps": 66.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-98ztc97t-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: 41673o4a, 4ko51tvl, p6zxretf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 903.0,
   "provenance": "community",
   "quant": "AWQ-5.0bpw",
   "quote": "Qwen3.8-27B-AWQ-5.0bpw on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 903.0 73.6 22.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/98ztc97t",
   "source_url": "https://omlx.ai/benchmarks/performance/98ztc97t",
   "tg_tokens": null,
   "tps": 73.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-98ztc97t-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: 41673o4a, 4ko51tvl, p6zxretf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 688.8,
   "provenance": "community",
   "quant": "AWQ-5.0bpw",
   "quote": "Qwen3.8-27B-AWQ-5.0bpw on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 64k 688.8 54.2 25.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/98ztc97t",
   "source_url": "https://omlx.ai/benchmarks/performance/98ztc97t",
   "tg_tokens": null,
   "tps": 54.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-98ztc97t-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 35.4 GB at this context; duplicate traces of the same session: 41673o4a, 4ko51tvl, p6zxretf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 463.3,
   "provenance": "community",
   "quant": "AWQ-5.0bpw",
   "quote": "Qwen3.8-27B-AWQ-5.0bpw on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 195k 463.3 35.5 35.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/98ztc97t",
   "source_url": "https://omlx.ai/benchmarks/performance/98ztc97t",
   "tg_tokens": null,
   "tps": 35.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-y9mlrgnv-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.6 GB at this context; duplicate traces of the same session: zx7tcsnr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 210.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 210.1 23.8 32.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y9mlrgnv",
   "source_url": "https://omlx.ai/benchmarks/performance/y9mlrgnv",
   "tg_tokens": null,
   "tps": 23.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-y9mlrgnv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.0 GB at this context; duplicate traces of the same session: zx7tcsnr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 276.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 276.3 26.4 36.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y9mlrgnv",
   "source_url": "https://omlx.ai/benchmarks/performance/y9mlrgnv",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-81adk6w4-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: 1atk03cy, 5rl33wbz, s1g8m2da; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 925.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 925.2 50.0 27.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/81adk6w4",
   "source_url": "https://omlx.ai/benchmarks/performance/81adk6w4",
   "tg_tokens": null,
   "tps": 50.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-81adk6w4-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: 1atk03cy, 5rl33wbz, s1g8m2da; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 801.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 801.2 64.0 28.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/81adk6w4",
   "source_url": "https://omlx.ai/benchmarks/performance/81adk6w4",
   "tg_tokens": null,
   "tps": 64.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-81adk6w4-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: 1atk03cy, 5rl33wbz, s1g8m2da; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 629.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k 629.3 45.1 32.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/81adk6w4",
   "source_url": "https://omlx.ai/benchmarks/performance/81adk6w4",
   "tg_tokens": null,
   "tps": 45.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-81adk6w4-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 41.5 GB at this context; duplicate traces of the same session: 1atk03cy, 5rl33wbz, s1g8m2da; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 449.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 195k 449.7 32.7 41.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/81adk6w4",
   "source_url": "https://omlx.ai/benchmarks/performance/81adk6w4",
   "tg_tokens": null,
   "tps": 32.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4bdx55e1-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 80.0 GB; TTFT 63962 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1025.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k tokens 1,025 PP tok/s 70.8 TG tok/s 63962 TTFT (ms) 80.0 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4bdx55e1",
   "source_url": "https://omlx.ai/benchmarks/performance/4bdx55e1",
   "tg_tokens": null,
   "tps": 70.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-yd5scotq-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: 5dh8n2gb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 537.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 537.1 48.8 70.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yd5scotq",
   "source_url": "https://omlx.ai/benchmarks/performance/yd5scotq",
   "tg_tokens": null,
   "tps": 48.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-yd5scotq-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: 5dh8n2gb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 653.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 653.9 53.0 71.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yd5scotq",
   "source_url": "https://omlx.ai/benchmarks/performance/yd5scotq",
   "tg_tokens": null,
   "tps": 53.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qv7pp4ih-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.6 GB; TTFT 1073 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 954.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 954.2 PP tok/s 77.8 TG tok/s 1073 TTFT (ms) 70.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qv7pp4ih",
   "source_url": "https://omlx.ai/benchmarks/performance/qv7pp4ih",
   "tg_tokens": null,
   "tps": 77.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-f5akp4p9-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 69.2 GB; TTFT 2676 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 382.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 382.7 PP tok/s 61.6 TG tok/s 2676 TTFT (ms) 69.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f5akp4p9",
   "source_url": "https://omlx.ai/benchmarks/performance/f5akp4p9",
   "tg_tokens": null,
   "tps": 61.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jahb01k1-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.0 GB at this context; duplicate traces of the same session: cenzbaog; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1929.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,929 81.7 99.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jahb01k1",
   "source_url": "https://omlx.ai/benchmarks/performance/jahb01k1",
   "tg_tokens": null,
   "tps": 81.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jahb01k1-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.5 GB at this context; duplicate traces of the same session: cenzbaog; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2849.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,849 77.2 99.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jahb01k1",
   "source_url": "https://omlx.ai/benchmarks/performance/jahb01k1",
   "tg_tokens": null,
   "tps": 77.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-nj9lzlac-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 100.8 GB at this context; duplicate traces of the same session: n1k0nrn3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 325.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 325.9 46.8 100.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nj9lzlac",
   "source_url": "https://omlx.ai/benchmarks/performance/nj9lzlac",
   "tg_tokens": null,
   "tps": 46.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-nj9lzlac-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: n1k0nrn3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 383.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 383.1 44.4 101.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nj9lzlac",
   "source_url": "https://omlx.ai/benchmarks/performance/nj9lzlac",
   "tg_tokens": null,
   "tps": 44.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-n40dxshh-ctx1024",
   "model": "qwen3.5-4b-agentic-coder-v4",
   "notes": "MLX; oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: w9lzjplo; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 349.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "qwen3.5-4b-agentic-coder-v4-4bit-mlx on M1 Pro (16c); M1 Pro (16c) \u00b7 16 GB \u00b7 4bit; 1k 349.0 54.3 3.4 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n40dxshh",
   "source_url": "https://omlx.ai/benchmarks/performance/n40dxshh",
   "tg_tokens": null,
   "tps": 54.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-n40dxshh-ctx4096",
   "model": "qwen3.5-4b-agentic-coder-v4",
   "notes": "MLX; oMLX community benchmark trace; peak memory 4.0 GB at this context; duplicate traces of the same session: w9lzjplo; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 341.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "qwen3.5-4b-agentic-coder-v4-4bit-mlx on M1 Pro (16c); M1 Pro (16c) \u00b7 16 GB \u00b7 4bit; 4k 341.9 52.5 4.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n40dxshh",
   "source_url": "https://omlx.ai/benchmarks/performance/n40dxshh",
   "tg_tokens": null,
   "tps": 52.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-fl0ue2mh-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.7 GB at this context; duplicate traces of the same session: qa47m1e5, z0dpck8j; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 696.6,
   "provenance": "community",
   "quant": "mixed 4/8bit",
   "quote": "GLM-5.3-Flash-MLX-mixed-4_8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 696.6 41.4 175.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fl0ue2mh",
   "source_url": "https://omlx.ai/benchmarks/performance/fl0ue2mh",
   "tg_tokens": null,
   "tps": 41.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-fl0ue2mh-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.0 GB at this context; duplicate traces of the same session: qa47m1e5, z0dpck8j; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 614.1,
   "provenance": "community",
   "quant": "mixed 4/8bit",
   "quote": "GLM-5.3-Flash-MLX-mixed-4_8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 614.1 35.1 178.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fl0ue2mh",
   "source_url": "https://omlx.ai/benchmarks/performance/fl0ue2mh",
   "tg_tokens": null,
   "tps": 35.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-fl0ue2mh-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.1 GB at this context; duplicate traces of the same session: qa47m1e5, z0dpck8j; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 583.6,
   "provenance": "community",
   "quant": "mixed 4/8bit",
   "quote": "GLM-5.3-Flash-MLX-mixed-4_8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 583.6 35.4 178.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fl0ue2mh",
   "source_url": "https://omlx.ai/benchmarks/performance/fl0ue2mh",
   "tg_tokens": null,
   "tps": 35.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-exrvtdi6-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.8 GB at this context; duplicate traces of the same session: x1x8j5dq; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 687.0,
   "provenance": "community",
   "quant": "mixed 4/8bit",
   "quote": "GLM-5.3-Flash-MLX-mixed-4_8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 687.0 41.5 175.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/exrvtdi6",
   "source_url": "https://omlx.ai/benchmarks/performance/exrvtdi6",
   "tg_tokens": null,
   "tps": 41.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-exrvtdi6-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.0 GB at this context; duplicate traces of the same session: x1x8j5dq; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 596.1,
   "provenance": "community",
   "quant": "mixed 4/8bit",
   "quote": "GLM-5.3-Flash-MLX-mixed-4_8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 596.1 35.2 178.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/exrvtdi6",
   "source_url": "https://omlx.ai/benchmarks/performance/exrvtdi6",
   "tg_tokens": null,
   "tps": 35.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-x5ya2mtc-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Huihui abliterated build; oMLX community benchmark trace, single run at 8k context; peak memory 70.9 GB; TTFT 3502 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2339.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.8-Flash-Next-abliterated-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k tokens 2,339 PP tok/s 73.5 TG tok/s 3502 TTFT (ms) 70.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x5ya2mtc",
   "source_url": "https://omlx.ai/benchmarks/performance/x5ya2mtc",
   "tg_tokens": null,
   "tps": 73.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2ntdezlj-ctx1024",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: 04rgwmep, 08ozy2o3, 19xormjc, 9pln8hxq, cywmj7yn, rtcyemu2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3560.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 3,560 163.8 18.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ntdezlj",
   "source_url": "https://omlx.ai/benchmarks/performance/2ntdezlj",
   "tg_tokens": null,
   "tps": 163.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2ntdezlj-ctx4096",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 04rgwmep, 08ozy2o3, 19xormjc, 9pln8hxq, cywmj7yn, rtcyemu2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4070.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 4,070 157.9 19.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ntdezlj",
   "source_url": "https://omlx.ai/benchmarks/performance/2ntdezlj",
   "tg_tokens": null,
   "tps": 157.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2ntdezlj-ctx8192",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 04rgwmep, 08ozy2o3, 19xormjc, 9pln8hxq, cywmj7yn, rtcyemu2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4002.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 4,002 152.6 19.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ntdezlj",
   "source_url": "https://omlx.ai/benchmarks/performance/2ntdezlj",
   "tg_tokens": null,
   "tps": 152.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2ntdezlj-ctx16384",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 04rgwmep, 08ozy2o3, 19xormjc, 9pln8hxq, cywmj7yn, rtcyemu2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3380.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 3,380 139.1 19.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ntdezlj",
   "source_url": "https://omlx.ai/benchmarks/performance/2ntdezlj",
   "tg_tokens": null,
   "tps": 139.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2ntdezlj-ctx32768",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 04rgwmep, 08ozy2o3, 19xormjc, 9pln8hxq, cywmj7yn, rtcyemu2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3159.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 3,159 122.2 20.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ntdezlj",
   "source_url": "https://omlx.ai/benchmarks/performance/2ntdezlj",
   "tg_tokens": null,
   "tps": 122.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2ntdezlj-ctx65536",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 04rgwmep, 08ozy2o3, 19xormjc, 9pln8hxq, cywmj7yn, rtcyemu2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2796.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,796 100.2 20.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ntdezlj",
   "source_url": "https://omlx.ai/benchmarks/performance/2ntdezlj",
   "tg_tokens": null,
   "tps": 100.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2ntdezlj-ctx131072",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 04rgwmep, 08ozy2o3, 19xormjc, 9pln8hxq, cywmj7yn, rtcyemu2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2281.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,281 73.7 20.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ntdezlj",
   "source_url": "https://omlx.ai/benchmarks/performance/2ntdezlj",
   "tg_tokens": null,
   "tps": 73.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2ntdezlj-ctx199680",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: 04rgwmep, 08ozy2o3, 19xormjc, 9pln8hxq, cywmj7yn, rtcyemu2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1886.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 1,886 57.0 21.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ntdezlj",
   "source_url": "https://omlx.ai/benchmarks/performance/2ntdezlj",
   "tg_tokens": null,
   "tps": 57.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-324542jl-ctx1024",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: 02p6zxyo, 0ta645xu, 1a0iv18m, 27uzyfx2, jue2uef2, v4t67n0g, wxa7nd77; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1563.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,563 130.3 22.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/324542jl",
   "source_url": "https://omlx.ai/benchmarks/performance/324542jl",
   "tg_tokens": null,
   "tps": 130.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-324542jl-ctx4096",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: 02p6zxyo, 0ta645xu, 1a0iv18m, 27uzyfx2, jue2uef2, v4t67n0g, wxa7nd77; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2219.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,219 126.7 23.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/324542jl",
   "source_url": "https://omlx.ai/benchmarks/performance/324542jl",
   "tg_tokens": null,
   "tps": 126.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-324542jl-ctx8192",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 02p6zxyo, 0ta645xu, 1a0iv18m, 27uzyfx2, jue2uef2, v4t67n0g, wxa7nd77; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3352.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 3,352 124.4 23.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/324542jl",
   "source_url": "https://omlx.ai/benchmarks/performance/324542jl",
   "tg_tokens": null,
   "tps": 124.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-324542jl-ctx16384",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 02p6zxyo, 0ta645xu, 1a0iv18m, 27uzyfx2, jue2uef2, v4t67n0g, wxa7nd77; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3351.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 3,351 120.8 23.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/324542jl",
   "source_url": "https://omlx.ai/benchmarks/performance/324542jl",
   "tg_tokens": null,
   "tps": 120.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-324542jl-ctx32768",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: 02p6zxyo, 0ta645xu, 1a0iv18m, 27uzyfx2, jue2uef2, v4t67n0g, wxa7nd77; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3176.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 3,176 114.0 24.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/324542jl",
   "source_url": "https://omlx.ai/benchmarks/performance/324542jl",
   "tg_tokens": null,
   "tps": 114.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-324542jl-ctx65536",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: 02p6zxyo, 0ta645xu, 1a0iv18m, 27uzyfx2, jue2uef2, v4t67n0g, wxa7nd77; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2901.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,901 102.3 24.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/324542jl",
   "source_url": "https://omlx.ai/benchmarks/performance/324542jl",
   "tg_tokens": null,
   "tps": 102.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-324542jl-ctx131072",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: 02p6zxyo, 0ta645xu, 1a0iv18m, 27uzyfx2, jue2uef2, v4t67n0g, wxa7nd77; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2370.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,370 85.6 24.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/324542jl",
   "source_url": "https://omlx.ai/benchmarks/performance/324542jl",
   "tg_tokens": null,
   "tps": 85.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-324542jl-ctx199680",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: 02p6zxyo, 0ta645xu, 1a0iv18m, 27uzyfx2, jue2uef2, v4t67n0g, wxa7nd77; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1927.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 1,927 71.6 25.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/324542jl",
   "source_url": "https://omlx.ai/benchmarks/performance/324542jl",
   "tg_tokens": null,
   "tps": 71.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ous51vfq-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 61.7 GB at this context; duplicate traces of the same session: 6bvppawi, pqp47lwh, z3k0qqud; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2541.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next-6bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 6bit; 1k 2,541 91.0 61.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ous51vfq",
   "source_url": "https://omlx.ai/benchmarks/performance/ous51vfq",
   "tg_tokens": null,
   "tps": 91.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ous51vfq-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 62.5 GB at this context; duplicate traces of the same session: 6bvppawi, pqp47lwh, z3k0qqud; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4338.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next-6bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 6bit; 4k 4,338 94.4 62.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ous51vfq",
   "source_url": "https://omlx.ai/benchmarks/performance/ous51vfq",
   "tg_tokens": null,
   "tps": 94.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ous51vfq-ctx8192",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 62.6 GB at this context; duplicate traces of the same session: 6bvppawi, pqp47lwh, z3k0qqud; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4935.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next-6bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 6bit; 8k 4,935 92.7 62.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ous51vfq",
   "source_url": "https://omlx.ai/benchmarks/performance/ous51vfq",
   "tg_tokens": null,
   "tps": 92.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ous51vfq-ctx32768",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 63.1 GB at this context; duplicate traces of the same session: 6bvppawi, pqp47lwh, z3k0qqud; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4284.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next-6bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 6bit; 32k 4,284 85.4 63.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ous51vfq",
   "source_url": "https://omlx.ai/benchmarks/performance/ous51vfq",
   "tg_tokens": null,
   "tps": 85.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ous51vfq-ctx65536",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 63.9 GB at this context; duplicate traces of the same session: 6bvppawi, pqp47lwh, z3k0qqud; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3569.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next-6bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 6bit; 64k 3,569 75.7 63.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ous51vfq",
   "source_url": "https://omlx.ai/benchmarks/performance/ous51vfq",
   "tg_tokens": null,
   "tps": 75.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hhhqgdi3-ctx65536",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V13",
   "notes": "dequantized fp16 MTP build; oMLX community benchmark trace, single run at 64k context; peak memory 31.6 GB; TTFT 20181 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3247.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V13-dequantized-oQ6e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k tokens 3,247 PP tok/s 83.0 TG tok/s 20181 TTFT (ms) 31.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hhhqgdi3",
   "source_url": "https://omlx.ai/benchmarks/performance/hhhqgdi3",
   "tg_tokens": null,
   "tps": 83.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-q8ye0q8g-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: t3c5xcx8; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 625.4,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-text-oQ4 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 625.4 53.3 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q8ye0q8g",
   "source_url": "https://omlx.ai/benchmarks/performance/q8ye0q8g",
   "tg_tokens": null,
   "tps": 53.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-q8ye0q8g-ctx16384",
   "model": "Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive",
   "notes": "oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: t3c5xcx8; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 458.9,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-text-oQ4 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 458.9 46.3 23.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q8ye0q8g",
   "source_url": "https://omlx.ai/benchmarks/performance/q8ye0q8g",
   "tg_tokens": null,
   "tps": 46.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gmjydle1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: 2y1uy3bx, 566yh9ux, i81zg04c, k3344ebm, u8w882h2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 982.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 982.3 33.0 17.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gmjydle1",
   "source_url": "https://omlx.ai/benchmarks/performance/gmjydle1",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gmjydle1-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 2y1uy3bx, 566yh9ux, i81zg04c, k3344ebm, u8w882h2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1013.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,013 32.5 17.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gmjydle1",
   "source_url": "https://omlx.ai/benchmarks/performance/gmjydle1",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gmjydle1-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 2y1uy3bx, 566yh9ux, i81zg04c, k3344ebm, u8w882h2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 983.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 983.3 31.3 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gmjydle1",
   "source_url": "https://omlx.ai/benchmarks/performance/gmjydle1",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gmjydle1-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 2y1uy3bx, 566yh9ux, i81zg04c, k3344ebm, u8w882h2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 927.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 927.4 29.7 18.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gmjydle1",
   "source_url": "https://omlx.ai/benchmarks/performance/gmjydle1",
   "tg_tokens": null,
   "tps": 29.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gmjydle1-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 2y1uy3bx, 566yh9ux, i81zg04c, k3344ebm, u8w882h2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 813.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 813.5 26.3 20.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gmjydle1",
   "source_url": "https://omlx.ai/benchmarks/performance/gmjydle1",
   "tg_tokens": null,
   "tps": 26.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gmjydle1-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: 2y1uy3bx, 566yh9ux, i81zg04c, k3344ebm, u8w882h2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 650.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 650.7 21.7 24.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gmjydle1",
   "source_url": "https://omlx.ai/benchmarks/performance/gmjydle1",
   "tg_tokens": null,
   "tps": 21.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-4pyhepeq-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: nlnste3q, qej1r2ex, yzcvoae0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 550.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 550.4 30.1 17.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4pyhepeq",
   "source_url": "https://omlx.ai/benchmarks/performance/4pyhepeq",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-4pyhepeq-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: nlnste3q, qej1r2ex, yzcvoae0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 697.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 697.8 45.9 19.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4pyhepeq",
   "source_url": "https://omlx.ai/benchmarks/performance/4pyhepeq",
   "tg_tokens": null,
   "tps": 45.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-4pyhepeq-ctx8192",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: nlnste3q, qej1r2ex, yzcvoae0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 717.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 717.8 22.4 19.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4pyhepeq",
   "source_url": "https://omlx.ai/benchmarks/performance/4pyhepeq",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-4pyhepeq-ctx16384",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: nlnste3q, qej1r2ex, yzcvoae0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 698.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 698.8 29.7 19.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4pyhepeq",
   "source_url": "https://omlx.ai/benchmarks/performance/4pyhepeq",
   "tg_tokens": null,
   "tps": 29.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-1rf72qud-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 6210tu9u; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 102.9 22.7 17.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1rf72qud",
   "source_url": "https://omlx.ai/benchmarks/performance/1rf72qud",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-1rf72qud-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: 6210tu9u; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 102.8 22.2 19.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1rf72qud",
   "source_url": "https://omlx.ai/benchmarks/performance/1rf72qud",
   "tg_tokens": null,
   "tps": 22.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-a1ibjtzc-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: asqee5oj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1176.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,176 80.9 70.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a1ibjtzc",
   "source_url": "https://omlx.ai/benchmarks/performance/a1ibjtzc",
   "tg_tokens": null,
   "tps": 80.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-a1ibjtzc-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: asqee5oj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2137.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,137 89.4 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a1ibjtzc",
   "source_url": "https://omlx.ai/benchmarks/performance/a1ibjtzc",
   "tg_tokens": null,
   "tps": 89.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n2zuonu9-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: rk9bg69f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1744.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,744 79.4 100.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n2zuonu9",
   "source_url": "https://omlx.ai/benchmarks/performance/n2zuonu9",
   "tg_tokens": null,
   "tps": 79.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n2zuonu9-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.0 GB at this context; duplicate traces of the same session: rk9bg69f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2124.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,124 91.4 101.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n2zuonu9",
   "source_url": "https://omlx.ai/benchmarks/performance/n2zuonu9",
   "tg_tokens": null,
   "tps": 91.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-rxjiwfq3-ctx1024",
   "model": "Gemma 4 12B",
   "notes": "oMLX community benchmark trace; peak memory 12.8 GB at this context; duplicate traces of the same session: exngyuj3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 274.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-12B-it-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 1k 274.5 17.3 12.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rxjiwfq3",
   "source_url": "https://omlx.ai/benchmarks/performance/rxjiwfq3",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-rxjiwfq3-ctx4096",
   "model": "Gemma 4 12B",
   "notes": "oMLX community benchmark trace; peak memory 13.2 GB at this context; duplicate traces of the same session: exngyuj3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 271.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-12B-it-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 4k 271.6 17.2 13.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rxjiwfq3",
   "source_url": "https://omlx.ai/benchmarks/performance/rxjiwfq3",
   "tg_tokens": null,
   "tps": 17.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8dmnzb5u-ctx65536",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 13.5 GB; TTFT 18985 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3452.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k tokens 3,452 PP tok/s 75.3 TG tok/s 18985 TTFT (ms) 13.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8dmnzb5u",
   "source_url": "https://omlx.ai/benchmarks/performance/8dmnzb5u",
   "tg_tokens": null,
   "tps": 75.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-yls3vniu-ctx4096",
   "model": "Qwen3.8-27B-aeon-ultimate-uncensored",
   "notes": "vision MTP build; oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: 8n9nkin4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 842.5,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "qwen3.8-27b-aeon-ultimate-uncensored-mxfp4-vision-mtplx on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 842.5 31.7 16.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yls3vniu",
   "source_url": "https://omlx.ai/benchmarks/performance/yls3vniu",
   "tg_tokens": null,
   "tps": 31.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-yls3vniu-ctx65536",
   "model": "Qwen3.8-27B-aeon-ultimate-uncensored",
   "notes": "vision MTP build; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: 8n9nkin4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 574.8,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "qwen3.8-27b-aeon-ultimate-uncensored-mxfp4-vision-mtplx on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 574.8 23.2 24.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yls3vniu",
   "source_url": "https://omlx.ai/benchmarks/performance/yls3vniu",
   "tg_tokens": null,
   "tps": 23.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-vo2o73pp-ctx1024",
   "model": "Qwen3.5-9B-GLM5.1-Distill-v1",
   "notes": "MLX; oMLX community benchmark trace; peak memory 5.7 GB at this context; duplicate traces of the same session: 30wjqi6h; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 711.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MLX-Qwen3.5-9B-GLM5.1-Distill-v1-4bit on M5 (10c); M5 (10c) \u00b7 16 GB \u00b7 4bit; 1k 711.7 26.2 5.7 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vo2o73pp",
   "source_url": "https://omlx.ai/benchmarks/performance/vo2o73pp",
   "tg_tokens": null,
   "tps": 26.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-vo2o73pp-ctx4096",
   "model": "Qwen3.5-9B-GLM5.1-Distill-v1",
   "notes": "MLX; oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: 30wjqi6h; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 712.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MLX-Qwen3.5-9B-GLM5.1-Distill-v1-4bit on M5 (10c); M5 (10c) \u00b7 16 GB \u00b7 4bit; 4k 712.3 25.6 6.3 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vo2o73pp",
   "source_url": "https://omlx.ai/benchmarks/performance/vo2o73pp",
   "tg_tokens": null,
   "tps": 25.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-1e31ymcw-ctx1024",
   "model": "Qwen3.5-9B-GLM5.1-Distill-v1",
   "notes": "MLX; oMLX community benchmark trace; peak memory 5.7 GB at this context; duplicate traces of the same session: 0vanl2dh; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 701.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MLX-Qwen3.5-9B-GLM5.1-Distill-v1-4bit on M5 (10c); M5 (10c) \u00b7 16 GB \u00b7 4bit; 1k 701.7 24.8 5.7 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1e31ymcw",
   "source_url": "https://omlx.ai/benchmarks/performance/1e31ymcw",
   "tg_tokens": null,
   "tps": 24.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-1e31ymcw-ctx4096",
   "model": "Qwen3.5-9B-GLM5.1-Distill-v1",
   "notes": "MLX; oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: 0vanl2dh; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 473.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MLX-Qwen3.5-9B-GLM5.1-Distill-v1-4bit on M5 (10c); M5 (10c) \u00b7 16 GB \u00b7 4bit; 4k 473.2 24.2 6.3 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1e31ymcw",
   "source_url": "https://omlx.ai/benchmarks/performance/1e31ymcw",
   "tg_tokens": null,
   "tps": 24.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ml51d61m-ctx1024",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3560.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 3,560 163.8 18.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ml51d61m",
   "source_url": "https://omlx.ai/benchmarks/performance/ml51d61m",
   "tg_tokens": null,
   "tps": 163.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ml51d61m-ctx4096",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4070.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 4,070 157.9 19.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ml51d61m",
   "source_url": "https://omlx.ai/benchmarks/performance/ml51d61m",
   "tg_tokens": null,
   "tps": 157.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ml51d61m-ctx8192",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4002.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 4,002 152.6 19.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ml51d61m",
   "source_url": "https://omlx.ai/benchmarks/performance/ml51d61m",
   "tg_tokens": null,
   "tps": 152.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ml51d61m-ctx16384",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3380.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 3,380 139.1 19.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ml51d61m",
   "source_url": "https://omlx.ai/benchmarks/performance/ml51d61m",
   "tg_tokens": null,
   "tps": 139.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ml51d61m-ctx32768",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3159.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 3,159 122.2 20.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ml51d61m",
   "source_url": "https://omlx.ai/benchmarks/performance/ml51d61m",
   "tg_tokens": null,
   "tps": 122.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ml51d61m-ctx65536",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2796.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,796 100.2 20.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ml51d61m",
   "source_url": "https://omlx.ai/benchmarks/performance/ml51d61m",
   "tg_tokens": null,
   "tps": 100.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ml51d61m-ctx131072",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2281.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,281 73.7 20.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ml51d61m",
   "source_url": "https://omlx.ai/benchmarks/performance/ml51d61m",
   "tg_tokens": null,
   "tps": 73.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ml51d61m-ctx199680",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1886.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 1,886 57.0 21.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ml51d61m",
   "source_url": "https://omlx.ai/benchmarks/performance/ml51d61m",
   "tg_tokens": null,
   "tps": 57.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-hqj52s7w-ctx1024",
   "model": "Qwen2.5-Coder-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 4.8 GB at this context; duplicate traces of the same session: 4xaapvcz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 492.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-7B-Instruct-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 492.0 55.8 4.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hqj52s7w",
   "source_url": "https://omlx.ai/benchmarks/performance/hqj52s7w",
   "tg_tokens": null,
   "tps": 55.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-hqj52s7w-ctx4096",
   "model": "Qwen2.5-Coder-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: 4xaapvcz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 483.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-7B-Instruct-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 483.9 52.7 4.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hqj52s7w",
   "source_url": "https://omlx.ai/benchmarks/performance/hqj52s7w",
   "tg_tokens": null,
   "tps": 52.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-brycodap-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 61.7 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2541.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next-6bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 6bit; 1k 2,541 91.0 61.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/brycodap",
   "source_url": "https://omlx.ai/benchmarks/performance/brycodap",
   "tg_tokens": null,
   "tps": 91.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-brycodap-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 62.5 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4338.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next-6bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 6bit; 4k 4,338 94.4 62.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/brycodap",
   "source_url": "https://omlx.ai/benchmarks/performance/brycodap",
   "tg_tokens": null,
   "tps": 94.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-brycodap-ctx8192",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 62.6 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4935.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next-6bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 6bit; 8k 4,935 92.7 62.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/brycodap",
   "source_url": "https://omlx.ai/benchmarks/performance/brycodap",
   "tg_tokens": null,
   "tps": 92.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-brycodap-ctx32768",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 63.1 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4284.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next-6bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 6bit; 32k 4,284 85.4 63.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/brycodap",
   "source_url": "https://omlx.ai/benchmarks/performance/brycodap",
   "tg_tokens": null,
   "tps": 85.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-brycodap-ctx65536",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 63.9 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3569.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next-6bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 6bit; 64k 3,569 75.7 63.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/brycodap",
   "source_url": "https://omlx.ai/benchmarks/performance/brycodap",
   "tg_tokens": null,
   "tps": 75.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-j8fugmtn-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 7gul1xtl, o0ggpjq3, u8wtvrnq, y6b1jyfo; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 851.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 851.4 91.0 19.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j8fugmtn",
   "source_url": "https://omlx.ai/benchmarks/performance/j8fugmtn",
   "tg_tokens": null,
   "tps": 91.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-j8fugmtn-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 7gul1xtl, o0ggpjq3, u8wtvrnq, y6b1jyfo; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 978.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 978.4 84.4 20.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j8fugmtn",
   "source_url": "https://omlx.ai/benchmarks/performance/j8fugmtn",
   "tg_tokens": null,
   "tps": 84.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-j8fugmtn-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: 7gul1xtl, o0ggpjq3, u8wtvrnq, y6b1jyfo; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 843.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 843.5 72.5 20.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j8fugmtn",
   "source_url": "https://omlx.ai/benchmarks/performance/j8fugmtn",
   "tg_tokens": null,
   "tps": 72.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-j8fugmtn-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 7gul1xtl, o0ggpjq3, u8wtvrnq, y6b1jyfo; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 715.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 715.4 66.7 20.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j8fugmtn",
   "source_url": "https://omlx.ai/benchmarks/performance/j8fugmtn",
   "tg_tokens": null,
   "tps": 66.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-j8fugmtn-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 7gul1xtl, o0ggpjq3, u8wtvrnq, y6b1jyfo; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 541.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 541.3 51.7 21.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j8fugmtn",
   "source_url": "https://omlx.ai/benchmarks/performance/j8fugmtn",
   "tg_tokens": null,
   "tps": 51.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-lnuqbuc9-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 9v1xexq1, we57ihys; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 797.2,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 797.2 88.6 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lnuqbuc9",
   "source_url": "https://omlx.ai/benchmarks/performance/lnuqbuc9",
   "tg_tokens": null,
   "tps": 88.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-lnuqbuc9-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 9v1xexq1, we57ihys; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 944.2,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 944.2 74.4 23.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lnuqbuc9",
   "source_url": "https://omlx.ai/benchmarks/performance/lnuqbuc9",
   "tg_tokens": null,
   "tps": 74.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-lnuqbuc9-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: 9v1xexq1, we57ihys; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 792.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 792.0 99.9 24.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lnuqbuc9",
   "source_url": "https://omlx.ai/benchmarks/performance/lnuqbuc9",
   "tg_tokens": null,
   "tps": 99.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-lnuqbuc9-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: 9v1xexq1, we57ihys; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 736.6,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 736.6 82.2 23.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lnuqbuc9",
   "source_url": "https://omlx.ai/benchmarks/performance/lnuqbuc9",
   "tg_tokens": null,
   "tps": 82.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-lnuqbuc9-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: 9v1xexq1, we57ihys; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 616.1,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 32k 616.1 105.2 24.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lnuqbuc9",
   "source_url": "https://omlx.ai/benchmarks/performance/lnuqbuc9",
   "tg_tokens": null,
   "tps": 105.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-q08rgfiw-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: whamikrs; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1402.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-oQ8-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 8bit; 1k 1,402 61.2 35.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q08rgfiw",
   "source_url": "https://omlx.ai/benchmarks/performance/q08rgfiw",
   "tg_tokens": null,
   "tps": 61.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-q08rgfiw-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: whamikrs; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1634.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-oQ8-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 8bit; 4k 1,634 59.7 37.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q08rgfiw",
   "source_url": "https://omlx.ai/benchmarks/performance/q08rgfiw",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Pro",
   "id": "omlx-3zzvezs9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: kmyn2b2i; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 1k 102.3 8.7 16.0 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3zzvezs9",
   "source_url": "https://omlx.ai/benchmarks/performance/3zzvezs9",
   "tg_tokens": null,
   "tps": 8.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Pro",
   "id": "omlx-3zzvezs9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: kmyn2b2i; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 4k 105.8 8.7 17.0 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3zzvezs9",
   "source_url": "https://omlx.ai/benchmarks/performance/3zzvezs9",
   "tg_tokens": null,
   "tps": 8.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qhwtg1i9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: xxlzwcf5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 471.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 471.6 37.4 17.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qhwtg1i9",
   "source_url": "https://omlx.ai/benchmarks/performance/qhwtg1i9",
   "tg_tokens": null,
   "tps": 37.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qhwtg1i9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: xxlzwcf5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 463.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 463.7 51.5 18.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qhwtg1i9",
   "source_url": "https://omlx.ai/benchmarks/performance/qhwtg1i9",
   "tg_tokens": null,
   "tps": 51.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-4o8v2fa3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: f3j0mbr0, t9tjnxif; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 252.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 252.9 20.7 22.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4o8v2fa3",
   "source_url": "https://omlx.ai/benchmarks/performance/4o8v2fa3",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-4o8v2fa3-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: f3j0mbr0, t9tjnxif; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 267.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 8k 267.3 20.1 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4o8v2fa3",
   "source_url": "https://omlx.ai/benchmarks/performance/4o8v2fa3",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-4o8v2fa3-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: f3j0mbr0, t9tjnxif; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 255.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k 255.6 19.4 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4o8v2fa3",
   "source_url": "https://omlx.ai/benchmarks/performance/4o8v2fa3",
   "tg_tokens": null,
   "tps": 19.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cl89yluy-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 687.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 687.4 39.6 26.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cl89yluy",
   "source_url": "https://omlx.ai/benchmarks/performance/cl89yluy",
   "tg_tokens": null,
   "tps": 39.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cl89yluy-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 870.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 870.0 50.1 27.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cl89yluy",
   "source_url": "https://omlx.ai/benchmarks/performance/cl89yluy",
   "tg_tokens": null,
   "tps": 50.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yry6wo5t-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MLX; oMLX community benchmark trace; peak memory 75.3 GB at this context; duplicate traces of the same session: 58yvc2sk, 90d3xz3a, fpo7nx6d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 592.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 592.5 38.0 75.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yry6wo5t",
   "source_url": "https://omlx.ai/benchmarks/performance/yry6wo5t",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yry6wo5t-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MLX; oMLX community benchmark trace; peak memory 75.5 GB at this context; duplicate traces of the same session: 58yvc2sk, 90d3xz3a, fpo7nx6d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1232.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,232 37.0 75.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yry6wo5t",
   "source_url": "https://omlx.ai/benchmarks/performance/yry6wo5t",
   "tg_tokens": null,
   "tps": 37.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yry6wo5t-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MLX; oMLX community benchmark trace; peak memory 76.0 GB at this context; duplicate traces of the same session: 58yvc2sk, 90d3xz3a, fpo7nx6d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1697.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,697 39.3 76.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yry6wo5t",
   "source_url": "https://omlx.ai/benchmarks/performance/yry6wo5t",
   "tg_tokens": null,
   "tps": 39.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-yry6wo5t-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MLX; oMLX community benchmark trace; peak memory 76.7 GB at this context; duplicate traces of the same session: 58yvc2sk, 90d3xz3a, fpo7nx6d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2083.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,083 40.0 76.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yry6wo5t",
   "source_url": "https://omlx.ai/benchmarks/performance/yry6wo5t",
   "tg_tokens": null,
   "tps": 40.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-yaata3bu-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: eqecc5ng, vhdchzhu, yijspm34; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 665.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 665.3 40.9 71.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yaata3bu",
   "source_url": "https://omlx.ai/benchmarks/performance/yaata3bu",
   "tg_tokens": null,
   "tps": 40.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-yaata3bu-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: eqecc5ng, vhdchzhu, yijspm34; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 570.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 570.7 63.8 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yaata3bu",
   "source_url": "https://omlx.ai/benchmarks/performance/yaata3bu",
   "tg_tokens": null,
   "tps": 63.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-yaata3bu-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: eqecc5ng, vhdchzhu, yijspm34; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 543.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 543.3 58.9 72.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yaata3bu",
   "source_url": "https://omlx.ai/benchmarks/performance/yaata3bu",
   "tg_tokens": null,
   "tps": 58.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-yaata3bu-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 75.0 GB at this context; duplicate traces of the same session: eqecc5ng, vhdchzhu, yijspm34; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 425.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 425.9 48.0 75.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yaata3bu",
   "source_url": "https://omlx.ai/benchmarks/performance/yaata3bu",
   "tg_tokens": null,
   "tps": 48.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-vl90pri9-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: 5nmevipn, f1i1v6ua; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 572.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 572.7 31.8 71.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vl90pri9",
   "source_url": "https://omlx.ai/benchmarks/performance/vl90pri9",
   "tg_tokens": null,
   "tps": 31.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-vl90pri9-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.4 GB at this context; duplicate traces of the same session: 5nmevipn, f1i1v6ua; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 494.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 494.6 43.4 72.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vl90pri9",
   "source_url": "https://omlx.ai/benchmarks/performance/vl90pri9",
   "tg_tokens": null,
   "tps": 43.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-vl90pri9-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.0 GB at this context; duplicate traces of the same session: 5nmevipn, f1i1v6ua; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 314.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 314.1 48.0 74.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vl90pri9",
   "source_url": "https://omlx.ai/benchmarks/performance/vl90pri9",
   "tg_tokens": null,
   "tps": 48.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-fr4gyj2z-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: q3ng06zh; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 682.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 1k 682.4 59.9 14.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fr4gyj2z",
   "source_url": "https://omlx.ai/benchmarks/performance/fr4gyj2z",
   "tg_tokens": null,
   "tps": 59.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-fr4gyj2z-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: q3ng06zh; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 635.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 4k 635.5 58.3 14.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fr4gyj2z",
   "source_url": "https://omlx.ai/benchmarks/performance/fr4gyj2z",
   "tg_tokens": null,
   "tps": 58.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-iko808vo-ctx1024",
   "model": "Gemma 4 26B-A4B-it-uncensored",
   "notes": "fp16; oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: 01w6ija8; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1160.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "gemma-4-26B-A4B-it-uncensored-oQ5-fp16 on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 5bit; 1k 1,160 49.6 17.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iko808vo",
   "source_url": "https://omlx.ai/benchmarks/performance/iko808vo",
   "tg_tokens": null,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-iko808vo-ctx4096",
   "model": "Gemma 4 26B-A4B-it-uncensored",
   "notes": "fp16; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: 01w6ija8; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1148.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "gemma-4-26B-A4B-it-uncensored-oQ5-fp16 on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 5bit; 4k 1,148 22.4 17.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iko808vo",
   "source_url": "https://omlx.ai/benchmarks/performance/iko808vo",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-xbd6fac6-ctx1024",
   "model": "Gemma 4 Ortenzya-The-Creative-Wordsmith-31B-it-uncensored-heretic",
   "notes": "MLX; oMLX community benchmark trace; peak memory 31.8 GB at this context; duplicate traces of the same session: r59x20t4; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 113.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-Ortenzya-The-Creative-Wordsmith-31B-it-uncensored-heretic-mlx-8Bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 8bit; 1k 113.3 14.9 31.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xbd6fac6",
   "source_url": "https://omlx.ai/benchmarks/performance/xbd6fac6",
   "tg_tokens": null,
   "tps": 14.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-xbd6fac6-ctx4096",
   "model": "Gemma 4 Ortenzya-The-Creative-Wordsmith-31B-it-uncensored-heretic",
   "notes": "MLX; oMLX community benchmark trace; peak memory 33.0 GB at this context; duplicate traces of the same session: r59x20t4; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 110.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-Ortenzya-The-Creative-Wordsmith-31B-it-uncensored-heretic-mlx-8Bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 8bit; 4k 110.0 14.6 33.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xbd6fac6",
   "source_url": "https://omlx.ai/benchmarks/performance/xbd6fac6",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-cb1omtij-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: jeytbp1w, pq5w7ek2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 245.9,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "qwen3.8-27b-oq6-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 245.9 46.0 26.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cb1omtij",
   "source_url": "https://omlx.ai/benchmarks/performance/cb1omtij",
   "tg_tokens": null,
   "tps": 46.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-cb1omtij-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.1 GB at this context; duplicate traces of the same session: jeytbp1w, pq5w7ek2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 261.3,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "qwen3.8-27b-oq6-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 261.3 48.0 30.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cb1omtij",
   "source_url": "https://omlx.ai/benchmarks/performance/cb1omtij",
   "tg_tokens": null,
   "tps": 48.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-cb1omtij-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.2 GB at this context; duplicate traces of the same session: jeytbp1w, pq5w7ek2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 230.2,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "qwen3.8-27b-oq6-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 230.2 51.5 33.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cb1omtij",
   "source_url": "https://omlx.ai/benchmarks/performance/cb1omtij",
   "tg_tokens": null,
   "tps": 51.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-c51u25p4-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 197.2,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "qwen3.8-27b-oq6-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 64k 197.2 37.5 35.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c51u25p4",
   "source_url": "https://omlx.ai/benchmarks/performance/c51u25p4",
   "tg_tokens": null,
   "tps": 37.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-c51u25p4-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 40.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 160.8,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "qwen3.8-27b-oq6-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 128k 160.8 32.5 40.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c51u25p4",
   "source_url": "https://omlx.ai/benchmarks/performance/c51u25p4",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-bricbuvm-ctx1024",
   "model": "supergemma4-26b-abliterated-multimodal",
   "notes": "MLX; oMLX community benchmark trace; peak memory 5.1 GB at this context; duplicate traces of the same session: hxzrty4s; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 167.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "supergemma4-26b-abliterated-multimodal-mlx-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k 167.7 17.1 5.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bricbuvm",
   "source_url": "https://omlx.ai/benchmarks/performance/bricbuvm",
   "tg_tokens": null,
   "tps": 17.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-bricbuvm-ctx4096",
   "model": "supergemma4-26b-abliterated-multimodal",
   "notes": "MLX; oMLX community benchmark trace; peak memory 5.5 GB at this context; duplicate traces of the same session: hxzrty4s; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "supergemma4-26b-abliterated-multimodal-mlx-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 224.4 19.6 5.5 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bricbuvm",
   "source_url": "https://omlx.ai/benchmarks/performance/bricbuvm",
   "tg_tokens": null,
   "tps": 19.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-5j8a4tt1-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: 7oudglz8; config=19c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 240.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M2 Pro (19c); M2 Pro (19c) \u00b7 32 GB \u00b7 4bit; 4k 240.5 34.3 6.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5j8a4tt1",
   "source_url": "https://omlx.ai/benchmarks/performance/5j8a4tt1",
   "tg_tokens": null,
   "tps": 34.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-5j8a4tt1-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.1 GB at this context; duplicate traces of the same session: 7oudglz8; config=19c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 217.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M2 Pro (19c); M2 Pro (19c) \u00b7 32 GB \u00b7 4bit; 16k 217.6 30.1 7.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5j8a4tt1",
   "source_url": "https://omlx.ai/benchmarks/performance/5j8a4tt1",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-jxh08znr-ctx1024",
   "model": "Qwen3.6-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.8 GB at this context; duplicate traces of the same session: hvq8jp6b; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 44.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-27B-oQ4e-mtp on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 44.6 9.9 16.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jxh08znr",
   "source_url": "https://omlx.ai/benchmarks/performance/jxh08znr",
   "tg_tokens": null,
   "tps": 9.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-jxh08znr-ctx4096",
   "model": "Qwen3.6-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.2 GB at this context; duplicate traces of the same session: hvq8jp6b; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 47.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-27B-oQ4e-mtp on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 47.3 9.7 18.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jxh08znr",
   "source_url": "https://omlx.ai/benchmarks/performance/jxh08znr",
   "tg_tokens": null,
   "tps": 9.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-q9qo7f5b-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 9ks714j1; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 797.2,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 797.2 88.6 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q9qo7f5b",
   "source_url": "https://omlx.ai/benchmarks/performance/q9qo7f5b",
   "tg_tokens": null,
   "tps": 88.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-q9qo7f5b-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 9ks714j1; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 944.2,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 944.2 74.4 23.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q9qo7f5b",
   "source_url": "https://omlx.ai/benchmarks/performance/q9qo7f5b",
   "tg_tokens": null,
   "tps": 74.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-q9qo7f5b-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: 9ks714j1; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 792.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 792.0 99.9 24.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q9qo7f5b",
   "source_url": "https://omlx.ai/benchmarks/performance/q9qo7f5b",
   "tg_tokens": null,
   "tps": 99.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-q9qo7f5b-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: 9ks714j1; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 736.6,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 736.6 82.2 23.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q9qo7f5b",
   "source_url": "https://omlx.ai/benchmarks/performance/q9qo7f5b",
   "tg_tokens": null,
   "tps": 82.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-q9qo7f5b-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: 9ks714j1; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 616.1,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 32k 616.1 105.2 24.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q9qo7f5b",
   "source_url": "https://omlx.ai/benchmarks/performance/q9qo7f5b",
   "tg_tokens": null,
   "tps": 105.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-cmhdznr1-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; MLX; oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: frl0kymf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 812.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit-MTP-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 812.4 51.6 33.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cmhdznr1",
   "source_url": "https://omlx.ai/benchmarks/performance/cmhdznr1",
   "tg_tokens": null,
   "tps": 51.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-cmhdznr1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; MLX; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: frl0kymf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 936.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit-MTP-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 936.8 47.1 34.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cmhdznr1",
   "source_url": "https://omlx.ai/benchmarks/performance/cmhdznr1",
   "tg_tokens": null,
   "tps": 47.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ldw9mcas-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; MLX; oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: 69mg5wuz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 807.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit-MTP-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 807.0 46.6 33.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ldw9mcas",
   "source_url": "https://omlx.ai/benchmarks/performance/ldw9mcas",
   "tg_tokens": null,
   "tps": 46.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ldw9mcas-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; MLX; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: 69mg5wuz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 927.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit-MTP-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 927.7 39.6 34.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ldw9mcas",
   "source_url": "https://omlx.ai/benchmarks/performance/ldw9mcas",
   "tg_tokens": null,
   "tps": 39.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-xm9l9v6o-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace, single run at 128k context; peak memory 43.7 GB; TTFT 247733 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 529.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k tokens 529.1 PP tok/s 41.2 TG tok/s 247733 TTFT (ms) 43.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xm9l9v6o",
   "source_url": "https://omlx.ai/benchmarks/performance/xm9l9v6o",
   "tg_tokens": null,
   "tps": 41.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-uzfxsoh2-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: wvnnde0f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 810.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 810.0 56.0 33.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uzfxsoh2",
   "source_url": "https://omlx.ai/benchmarks/performance/uzfxsoh2",
   "tg_tokens": null,
   "tps": 56.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-uzfxsoh2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: wvnnde0f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 912.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 912.9 47.0 34.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uzfxsoh2",
   "source_url": "https://omlx.ai/benchmarks/performance/uzfxsoh2",
   "tg_tokens": null,
   "tps": 47.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-vlytozpx-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 16yqsq75, g4a1l7qx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 260.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 260.6 28.2 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vlytozpx",
   "source_url": "https://omlx.ai/benchmarks/performance/vlytozpx",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-vlytozpx-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 16yqsq75, g4a1l7qx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 276.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 276.1 27.4 19.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vlytozpx",
   "source_url": "https://omlx.ai/benchmarks/performance/vlytozpx",
   "tg_tokens": null,
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-vlytozpx-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 16yqsq75, g4a1l7qx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 261.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 261.0 25.4 21.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vlytozpx",
   "source_url": "https://omlx.ai/benchmarks/performance/vlytozpx",
   "tg_tokens": null,
   "tps": 25.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2t2uqw9l-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: clump3u4, mign6vpd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 255.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 255.0 37.9 18.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2t2uqw9l",
   "source_url": "https://omlx.ai/benchmarks/performance/2t2uqw9l",
   "tg_tokens": null,
   "tps": 37.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2t2uqw9l-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: clump3u4, mign6vpd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 270.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 270.0 37.8 22.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2t2uqw9l",
   "source_url": "https://omlx.ai/benchmarks/performance/2t2uqw9l",
   "tg_tokens": null,
   "tps": 37.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2t2uqw9l-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: clump3u4, mign6vpd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 254.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 254.9 37.1 23.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2t2uqw9l",
   "source_url": "https://omlx.ai/benchmarks/performance/2t2uqw9l",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-0wr1lzm7-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: eu67zi5m, jl4wnnan; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 260.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 260.9 28.2 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0wr1lzm7",
   "source_url": "https://omlx.ai/benchmarks/performance/0wr1lzm7",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-0wr1lzm7-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: eu67zi5m, jl4wnnan; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 276.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 276.1 27.3 19.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0wr1lzm7",
   "source_url": "https://omlx.ai/benchmarks/performance/0wr1lzm7",
   "tg_tokens": null,
   "tps": 27.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-0wr1lzm7-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: eu67zi5m, jl4wnnan; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 261.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 261.1 25.4 21.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0wr1lzm7",
   "source_url": "https://omlx.ai/benchmarks/performance/0wr1lzm7",
   "tg_tokens": null,
   "tps": 25.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-qtx9caa9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 021pb51t, onml0kpw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 253.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 253.9 59.3 18.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qtx9caa9",
   "source_url": "https://omlx.ai/benchmarks/performance/qtx9caa9",
   "tg_tokens": null,
   "tps": 59.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-qtx9caa9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 021pb51t, onml0kpw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 269.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 269.9 57.4 22.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qtx9caa9",
   "source_url": "https://omlx.ai/benchmarks/performance/qtx9caa9",
   "tg_tokens": null,
   "tps": 57.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-qtx9caa9-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 021pb51t, onml0kpw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 254.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 254.9 68.6 23.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qtx9caa9",
   "source_url": "https://omlx.ai/benchmarks/performance/qtx9caa9",
   "tg_tokens": null,
   "tps": 68.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ry3q0zcc-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 687.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 687.4 39.6 26.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ry3q0zcc",
   "source_url": "https://omlx.ai/benchmarks/performance/ry3q0zcc",
   "tg_tokens": null,
   "tps": 39.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ry3q0zcc-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 870.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 870.0 50.1 27.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ry3q0zcc",
   "source_url": "https://omlx.ai/benchmarks/performance/ry3q0zcc",
   "tg_tokens": null,
   "tps": 50.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mep0blk1-ctx1024",
   "model": "Gemma-4-31B",
   "notes": "oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: yvqyfn3k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 643.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31b-it-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 643.6 15.5 32.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mep0blk1",
   "source_url": "https://omlx.ai/benchmarks/performance/mep0blk1",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mep0blk1-ctx4096",
   "model": "Gemma-4-31B",
   "notes": "oMLX community benchmark trace; peak memory 33.1 GB at this context; duplicate traces of the same session: yvqyfn3k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 691.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31b-it-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 691.6 15.1 33.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mep0blk1",
   "source_url": "https://omlx.ai/benchmarks/performance/mep0blk1",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-m5rymx9p-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 197.2,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "qwen3.8-27b-oq6-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 64k 197.2 37.5 35.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m5rymx9p",
   "source_url": "https://omlx.ai/benchmarks/performance/m5rymx9p",
   "tg_tokens": null,
   "tps": 37.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-m5rymx9p-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 40.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 160.8,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "qwen3.8-27b-oq6-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 128k 160.8 32.5 40.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m5rymx9p",
   "source_url": "https://omlx.ai/benchmarks/performance/m5rymx9p",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-d5gzzxzz-ctx1024",
   "model": "Bonsai-2-27B-CRACK-Ternary",
   "notes": "CRACK build; oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: hq6dw6hp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 90.8,
   "provenance": "community",
   "quant": "ternary oMLX",
   "quote": "Bonsai-2-27B-CRACK-Ternary-oMLX on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 2bit; 1k 90.8 18.0 11.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d5gzzxzz",
   "source_url": "https://omlx.ai/benchmarks/performance/d5gzzxzz",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-d5gzzxzz-ctx4096",
   "model": "Bonsai-2-27B-CRACK-Ternary",
   "notes": "CRACK build; oMLX community benchmark trace; peak memory 14.5 GB at this context; duplicate traces of the same session: hq6dw6hp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 91.5,
   "provenance": "community",
   "quant": "ternary oMLX",
   "quote": "Bonsai-2-27B-CRACK-Ternary-oMLX on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 2bit; 4k 91.5 17.4 14.5 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d5gzzxzz",
   "source_url": "https://omlx.ai/benchmarks/performance/d5gzzxzz",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-vwezqr1a-ctx1024",
   "model": "DeepSeek-R1-Distill-Llama-70B",
   "notes": "oMLX community benchmark trace; peak memory 37.8 GB at this context; duplicate traces of the same session: l6uo3d1t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 349.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-R1-Distill-Llama-70B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 349.8 13.5 37.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vwezqr1a",
   "source_url": "https://omlx.ai/benchmarks/performance/vwezqr1a",
   "tg_tokens": null,
   "tps": 13.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-vwezqr1a-ctx4096",
   "model": "DeepSeek-R1-Distill-Llama-70B",
   "notes": "oMLX community benchmark trace; peak memory 39.0 GB at this context; duplicate traces of the same session: l6uo3d1t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 368.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-R1-Distill-Llama-70B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 368.0 13.0 39.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vwezqr1a",
   "source_url": "https://omlx.ai/benchmarks/performance/vwezqr1a",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-so64aiod-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: 7u78pz2o; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 764.6,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "Ornith-1.5-35B-A3B-oQ2e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 2bit; 1k 764.6 77.3 12.7 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/so64aiod",
   "source_url": "https://omlx.ai/benchmarks/performance/so64aiod",
   "tg_tokens": null,
   "tps": 77.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-so64aiod-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 13.3 GB at this context; duplicate traces of the same session: 7u78pz2o; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 840.6,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "Ornith-1.5-35B-A3B-oQ2e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 2bit; 4k 840.6 72.8 13.3 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/so64aiod",
   "source_url": "https://omlx.ai/benchmarks/performance/so64aiod",
   "tg_tokens": null,
   "tps": 72.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-t3g0h1uv-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: 0vn7ut93, 5o3v93kg, b0lcw3p1, dgmoogi1, je4juxyo, tmjauklz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 795.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 795.7 51.8 37.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t3g0h1uv",
   "source_url": "https://omlx.ai/benchmarks/performance/t3g0h1uv",
   "tg_tokens": null,
   "tps": 51.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-t3g0h1uv-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.0 GB at this context; duplicate traces of the same session: 0vn7ut93, 5o3v93kg, b0lcw3p1, dgmoogi1, je4juxyo, tmjauklz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2979.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 2,979 62.0 38.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t3g0h1uv",
   "source_url": "https://omlx.ai/benchmarks/performance/t3g0h1uv",
   "tg_tokens": null,
   "tps": 62.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-t3g0h1uv-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.1 GB at this context; duplicate traces of the same session: 0vn7ut93, 5o3v93kg, b0lcw3p1, dgmoogi1, je4juxyo, tmjauklz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2906.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 2,906 55.5 38.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t3g0h1uv",
   "source_url": "https://omlx.ai/benchmarks/performance/t3g0h1uv",
   "tg_tokens": null,
   "tps": 55.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-t3g0h1uv-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.3 GB at this context; duplicate traces of the same session: 0vn7ut93, 5o3v93kg, b0lcw3p1, dgmoogi1, je4juxyo, tmjauklz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2690.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 16k 2,690 76.9 38.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t3g0h1uv",
   "source_url": "https://omlx.ai/benchmarks/performance/t3g0h1uv",
   "tg_tokens": null,
   "tps": 76.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-t3g0h1uv-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.7 GB at this context; duplicate traces of the same session: 0vn7ut93, 5o3v93kg, b0lcw3p1, dgmoogi1, je4juxyo, tmjauklz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2272.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 32k 2,272 40.7 38.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t3g0h1uv",
   "source_url": "https://omlx.ai/benchmarks/performance/t3g0h1uv",
   "tg_tokens": null,
   "tps": 40.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-t3g0h1uv-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.4 GB at this context; duplicate traces of the same session: 0vn7ut93, 5o3v93kg, b0lcw3p1, dgmoogi1, je4juxyo, tmjauklz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1704.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 64k 1,704 33.3 39.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t3g0h1uv",
   "source_url": "https://omlx.ai/benchmarks/performance/t3g0h1uv",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-t3g0h1uv-ctx131072",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 40.9 GB at this context; duplicate traces of the same session: 0vn7ut93, 5o3v93kg, b0lcw3p1, dgmoogi1, je4juxyo, tmjauklz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1132.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 128k 1,132 29.0 40.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t3g0h1uv",
   "source_url": "https://omlx.ai/benchmarks/performance/t3g0h1uv",
   "tg_tokens": null,
   "tps": 29.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-2gafeap4-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; FP16; MLX; oMLX community benchmark trace, single run at 64k context; peak memory 26.7 GB; TTFT 124713 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 525.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-FP16-MTP-MLX on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 64k tokens 525.5 PP tok/s 42.6 TG tok/s 124713 TTFT (ms) 26.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2gafeap4",
   "source_url": "https://omlx.ai/benchmarks/performance/2gafeap4",
   "tg_tokens": null,
   "tps": 42.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-inpv5ly3-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 64k context; peak memory 42.6 GB; TTFT 219692 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 298.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 64k tokens 298.3 PP tok/s 36.7 TG tok/s 219692 TTFT (ms) 42.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/inpv5ly3",
   "source_url": "https://omlx.ai/benchmarks/performance/inpv5ly3",
   "tg_tokens": null,
   "tps": 36.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-tx4ya3g7-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 16.1 GB; TTFT 15508 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 66.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 66.0 PP tok/s 3.8 TG tok/s 15508 TTFT (ms) 16.1 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tx4ya3g7",
   "source_url": "https://omlx.ai/benchmarks/performance/tx4ya3g7",
   "tg_tokens": null,
   "tps": 3.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-rlse5s4o-ctx4096",
   "model": "Laguna-S-2.1",
   "notes": "fast build; oMLX community benchmark trace; peak memory 40.0 GB at this context; duplicate traces of the same session: qu189w62; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 201.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Laguna-S-2.1-oQ4e-fast on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 201.0 14.7 40.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rlse5s4o",
   "source_url": "https://omlx.ai/benchmarks/performance/rlse5s4o",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-rlse5s4o-ctx16384",
   "model": "Laguna-S-2.1",
   "notes": "fast build; oMLX community benchmark trace; peak memory 40.6 GB at this context; duplicate traces of the same session: qu189w62; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 209.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Laguna-S-2.1-oQ4e-fast on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 209.3 14.8 40.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rlse5s4o",
   "source_url": "https://omlx.ai/benchmarks/performance/rlse5s4o",
   "tg_tokens": null,
   "tps": 14.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-rrwypibn-ctx4096",
   "model": "Laguna-S-2.1",
   "notes": "fast build; oMLX community benchmark trace; peak memory 40.0 GB at this context; duplicate traces of the same session: jtahrxhz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 203.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Laguna-S-2.1-oQ4e-fast on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 203.1 15.2 40.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rrwypibn",
   "source_url": "https://omlx.ai/benchmarks/performance/rrwypibn",
   "tg_tokens": null,
   "tps": 15.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-rrwypibn-ctx16384",
   "model": "Laguna-S-2.1",
   "notes": "fast build; oMLX community benchmark trace; peak memory 40.7 GB at this context; duplicate traces of the same session: jtahrxhz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 211.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Laguna-S-2.1-oQ4e-fast on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 211.6 14.7 40.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rrwypibn",
   "source_url": "https://omlx.ai/benchmarks/performance/rrwypibn",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-uzcjnc5q-ctx1024",
   "model": "NeoHorse-1-4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 3.8 GB at this context; duplicate traces of the same session: vvx9pzpw; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 409.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NeoHorse-1-4B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 409.0 52.6 3.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uzcjnc5q",
   "source_url": "https://omlx.ai/benchmarks/performance/uzcjnc5q",
   "tg_tokens": null,
   "tps": 52.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-uzcjnc5q-ctx4096",
   "model": "NeoHorse-1-4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.0 GB at this context; duplicate traces of the same session: vvx9pzpw; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 541.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NeoHorse-1-4B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 541.1 48.3 4.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uzcjnc5q",
   "source_url": "https://omlx.ai/benchmarks/performance/uzcjnc5q",
   "tg_tokens": null,
   "tps": 48.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-18zmbr1o-ctx1024",
   "model": "NeoHorse-1-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: 31kxi64s; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 223.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NeoHorse-1-9B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 223.7 33.4 6.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/18zmbr1o",
   "source_url": "https://omlx.ai/benchmarks/performance/18zmbr1o",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-18zmbr1o-ctx4096",
   "model": "NeoHorse-1-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: 31kxi64s; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 231.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NeoHorse-1-9B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 231.3 28.3 6.7 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/18zmbr1o",
   "source_url": "https://omlx.ai/benchmarks/performance/18zmbr1o",
   "tg_tokens": null,
   "tps": 28.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-r0hxpn0r-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 35.5 GB; TTFT 247 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4139.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k tokens 4,139 PP tok/s 99.7 TG tok/s 247 TTFT (ms) 35.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r0hxpn0r",
   "source_url": "https://omlx.ai/benchmarks/performance/r0hxpn0r",
   "tg_tokens": null,
   "tps": 99.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ge9l61je-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "refusal-ablated MTP build; MLX; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: tnl2qohl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2008.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-refusal-ablated-MLX-oQ8-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 2,008 93.6 37.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ge9l61je",
   "source_url": "https://omlx.ai/benchmarks/performance/ge9l61je",
   "tg_tokens": null,
   "tps": 93.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ge9l61je-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "refusal-ablated MTP build; MLX; oMLX community benchmark trace; peak memory 38.0 GB at this context; duplicate traces of the same session: tnl2qohl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4873.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-refusal-ablated-MLX-oQ8-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 4,873 124.2 38.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ge9l61je",
   "source_url": "https://omlx.ai/benchmarks/performance/ge9l61je",
   "tg_tokens": null,
   "tps": 124.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-y2qewli6-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 3i7m8x5r, 644o6p1b, ex3qimke, g0rhf4lu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 888.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 888.9 88.8 21.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y2qewli6",
   "source_url": "https://omlx.ai/benchmarks/performance/y2qewli6",
   "tg_tokens": null,
   "tps": 88.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-y2qewli6-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 3i7m8x5r, 644o6p1b, ex3qimke, g0rhf4lu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 940.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 940.9 69.0 22.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y2qewli6",
   "source_url": "https://omlx.ai/benchmarks/performance/y2qewli6",
   "tg_tokens": null,
   "tps": 69.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-y2qewli6-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: 3i7m8x5r, 644o6p1b, ex3qimke, g0rhf4lu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 839.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 839.9 81.0 22.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y2qewli6",
   "source_url": "https://omlx.ai/benchmarks/performance/y2qewli6",
   "tg_tokens": null,
   "tps": 81.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-y2qewli6-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: 3i7m8x5r, 644o6p1b, ex3qimke, g0rhf4lu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 691.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 691.0 74.6 23.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y2qewli6",
   "source_url": "https://omlx.ai/benchmarks/performance/y2qewli6",
   "tg_tokens": null,
   "tps": 74.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-y2qewli6-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: 3i7m8x5r, 644o6p1b, ex3qimke, g0rhf4lu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 506.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 506.1 65.3 24.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y2qewli6",
   "source_url": "https://omlx.ai/benchmarks/performance/y2qewli6",
   "tg_tokens": null,
   "tps": 65.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xn0nomqz-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: 5ki9eu5g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 862.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 862.4 34.5 33.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xn0nomqz",
   "source_url": "https://omlx.ai/benchmarks/performance/xn0nomqz",
   "tg_tokens": null,
   "tps": 34.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xn0nomqz-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: 5ki9eu5g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 944.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 944.6 31.7 34.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xn0nomqz",
   "source_url": "https://omlx.ai/benchmarks/performance/xn0nomqz",
   "tg_tokens": null,
   "tps": 31.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-4z689png-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 16.1 GB; TTFT 15415 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 66.4,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 66.4 PP tok/s 4.4 TG tok/s 15415 TTFT (ms) 16.1 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4z689png",
   "source_url": "https://omlx.ai/benchmarks/performance/4z689png",
   "tg_tokens": null,
   "tps": 4.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-hcrtjeh5-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 16.1 GB; TTFT 15411 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 66.4,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 66.4 PP tok/s 4.1 TG tok/s 15411 TTFT (ms) 16.1 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hcrtjeh5",
   "source_url": "https://omlx.ai/benchmarks/performance/hcrtjeh5",
   "tg_tokens": null,
   "tps": 4.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-af3a57cn-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 16.1 GB; TTFT 15410 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 66.4,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 66.4 PP tok/s 4.5 TG tok/s 15410 TTFT (ms) 16.1 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/af3a57cn",
   "source_url": "https://omlx.ai/benchmarks/performance/af3a57cn",
   "tg_tokens": null,
   "tps": 4.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-r7e2y6di-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 464bc2t6, cijjll71, kg4gcjni, rt5psltn, z5rphhkp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 127.4 29.1 21.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r7e2y6di",
   "source_url": "https://omlx.ai/benchmarks/performance/r7e2y6di",
   "tg_tokens": null,
   "tps": 29.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-r7e2y6di-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: 464bc2t6, cijjll71, kg4gcjni, rt5psltn, z5rphhkp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 144.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 144.8 44.2 21.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r7e2y6di",
   "source_url": "https://omlx.ai/benchmarks/performance/r7e2y6di",
   "tg_tokens": null,
   "tps": 44.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-r7e2y6di-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 464bc2t6, cijjll71, kg4gcjni, rt5psltn, z5rphhkp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 131.0 19.8 20.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r7e2y6di",
   "source_url": "https://omlx.ai/benchmarks/performance/r7e2y6di",
   "tg_tokens": null,
   "tps": 19.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-r7e2y6di-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 464bc2t6, cijjll71, kg4gcjni, rt5psltn, z5rphhkp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 126.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 126.8 27.4 20.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r7e2y6di",
   "source_url": "https://omlx.ai/benchmarks/performance/r7e2y6di",
   "tg_tokens": null,
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-r7e2y6di-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: 464bc2t6, cijjll71, kg4gcjni, rt5psltn, z5rphhkp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 117.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 117.7 25.3 22.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r7e2y6di",
   "source_url": "https://omlx.ai/benchmarks/performance/r7e2y6di",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-r7e2y6di-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.2 GB at this context; duplicate traces of the same session: 464bc2t6, cijjll71, kg4gcjni, rt5psltn, z5rphhkp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 102.3 19.8 25.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r7e2y6di",
   "source_url": "https://omlx.ai/benchmarks/performance/r7e2y6di",
   "tg_tokens": null,
   "tps": 19.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-e0fofiyw-ctx1024",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "MLX; oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: f7yeqd3r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 879.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-35B-A3B-Distill-mlx-8Bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 879.6 99.0 35.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e0fofiyw",
   "source_url": "https://omlx.ai/benchmarks/performance/e0fofiyw",
   "tg_tokens": null,
   "tps": 99.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-e0fofiyw-ctx4096",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "MLX; oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: f7yeqd3r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2778.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-35B-A3B-Distill-mlx-8Bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 2,778 97.9 36.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e0fofiyw",
   "source_url": "https://omlx.ai/benchmarks/performance/e0fofiyw",
   "tg_tokens": null,
   "tps": 97.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2y56eps3-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 42.3 GB; TTFT 4921 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 832.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 832.4 PP tok/s 36.2 TG tok/s 4921 TTFT (ms) 42.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2y56eps3",
   "source_url": "https://omlx.ai/benchmarks/performance/2y56eps3",
   "tg_tokens": null,
   "tps": 36.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-a59gmki1-ctx32768",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 44.8 GB; TTFT 23072 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1420.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k tokens 1,420 PP tok/s 38.8 TG tok/s 23072 TTFT (ms) 44.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a59gmki1",
   "source_url": "https://omlx.ai/benchmarks/performance/a59gmki1",
   "tg_tokens": null,
   "tps": 38.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cc65rzni-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: qg2ednx4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1202.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,202 76.2 70.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cc65rzni",
   "source_url": "https://omlx.ai/benchmarks/performance/cc65rzni",
   "tg_tokens": null,
   "tps": 76.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cc65rzni-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: qg2ednx4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2104.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,104 69.8 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cc65rzni",
   "source_url": "https://omlx.ai/benchmarks/performance/cc65rzni",
   "tg_tokens": null,
   "tps": 69.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-l3v9xrun-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: htu0d793; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1128.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,128 84.0 70.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l3v9xrun",
   "source_url": "https://omlx.ai/benchmarks/performance/l3v9xrun",
   "tg_tokens": null,
   "tps": 84.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-l3v9xrun-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: htu0d793; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2051.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,051 69.6 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l3v9xrun",
   "source_url": "https://omlx.ai/benchmarks/performance/l3v9xrun",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-elx6jwir-ctx4096",
   "model": "Laguna-S-2.1",
   "notes": "oMLX community benchmark trace; peak memory 45.9 GB at this context; duplicate traces of the same session: 6uoqe9j9; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 114.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Laguna-S-2.1-oQ6e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 114.6 8.9 45.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/elx6jwir",
   "source_url": "https://omlx.ai/benchmarks/performance/elx6jwir",
   "tg_tokens": null,
   "tps": 8.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-elx6jwir-ctx16384",
   "model": "Laguna-S-2.1",
   "notes": "oMLX community benchmark trace; peak memory 46.6 GB at this context; duplicate traces of the same session: 6uoqe9j9; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 116.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Laguna-S-2.1-oQ6e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 16k 116.6 8.9 46.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/elx6jwir",
   "source_url": "https://omlx.ai/benchmarks/performance/elx6jwir",
   "tg_tokens": null,
   "tps": 8.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-62rbrmh1-ctx1024",
   "model": "NeoHorse-1-4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 3.8 GB at this context; duplicate traces of the same session: 9hx8b4o6; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 409.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NeoHorse-1-4B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 409.0 54.5 3.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/62rbrmh1",
   "source_url": "https://omlx.ai/benchmarks/performance/62rbrmh1",
   "tg_tokens": null,
   "tps": 54.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-62rbrmh1-ctx4096",
   "model": "NeoHorse-1-4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.0 GB at this context; duplicate traces of the same session: 9hx8b4o6; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 539.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NeoHorse-1-4B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 539.6 47.9 4.0 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/62rbrmh1",
   "source_url": "https://omlx.ai/benchmarks/performance/62rbrmh1",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-6qqdrq21-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: bab5104k, fw72wdeu, x2nqgflm, yscl37e5, z2sjizhy, ze3tw4ts; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1976.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 1,976 83.1 37.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6qqdrq21",
   "source_url": "https://omlx.ai/benchmarks/performance/6qqdrq21",
   "tg_tokens": null,
   "tps": 83.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-6qqdrq21-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.0 GB at this context; duplicate traces of the same session: bab5104k, fw72wdeu, x2nqgflm, yscl37e5, z2sjizhy, ze3tw4ts; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2994.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 2,994 88.7 38.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6qqdrq21",
   "source_url": "https://omlx.ai/benchmarks/performance/6qqdrq21",
   "tg_tokens": null,
   "tps": 88.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-6qqdrq21-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.1 GB at this context; duplicate traces of the same session: bab5104k, fw72wdeu, x2nqgflm, yscl37e5, z2sjizhy, ze3tw4ts; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2913.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 2,913 80.9 38.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6qqdrq21",
   "source_url": "https://omlx.ai/benchmarks/performance/6qqdrq21",
   "tg_tokens": null,
   "tps": 80.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-6qqdrq21-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.3 GB at this context; duplicate traces of the same session: bab5104k, fw72wdeu, x2nqgflm, yscl37e5, z2sjizhy, ze3tw4ts; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2698.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 16k 2,698 72.3 38.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6qqdrq21",
   "source_url": "https://omlx.ai/benchmarks/performance/6qqdrq21",
   "tg_tokens": null,
   "tps": 72.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-6qqdrq21-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.7 GB at this context; duplicate traces of the same session: bab5104k, fw72wdeu, x2nqgflm, yscl37e5, z2sjizhy, ze3tw4ts; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2294.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 32k 2,294 63.9 38.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6qqdrq21",
   "source_url": "https://omlx.ai/benchmarks/performance/6qqdrq21",
   "tg_tokens": null,
   "tps": 63.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-6qqdrq21-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.4 GB at this context; duplicate traces of the same session: bab5104k, fw72wdeu, x2nqgflm, yscl37e5, z2sjizhy, ze3tw4ts; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1733.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 64k 1,733 41.3 39.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6qqdrq21",
   "source_url": "https://omlx.ai/benchmarks/performance/6qqdrq21",
   "tg_tokens": null,
   "tps": 41.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-6qqdrq21-ctx131072",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 40.9 GB at this context; duplicate traces of the same session: bab5104k, fw72wdeu, x2nqgflm, yscl37e5, z2sjizhy, ze3tw4ts; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1138.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 128k 1,138 27.8 40.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6qqdrq21",
   "source_url": "https://omlx.ai/benchmarks/performance/6qqdrq21",
   "tg_tokens": null,
   "tps": 27.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-5f4vvb3c-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.6 GB at this context; duplicate traces of the same session: 18p580zu; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 378.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 378.7 48.8 6.6 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5f4vvb3c",
   "source_url": "https://omlx.ai/benchmarks/performance/5f4vvb3c",
   "tg_tokens": null,
   "tps": 48.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-5f4vvb3c-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: 18p580zu; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 532.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 532.1 47.7 6.7 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5f4vvb3c",
   "source_url": "https://omlx.ai/benchmarks/performance/5f4vvb3c",
   "tg_tokens": null,
   "tps": 47.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fo52qdxb-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 0e2tds33, 3y1nd5ub, lk1u577g; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1657.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.6-35B-A3B-mxfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 1,657 95.9 18.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fo52qdxb",
   "source_url": "https://omlx.ai/benchmarks/performance/fo52qdxb",
   "tg_tokens": null,
   "tps": 95.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fo52qdxb-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 0e2tds33, 3y1nd5ub, lk1u577g; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2358.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.6-35B-A3B-mxfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 2,358 82.9 19.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fo52qdxb",
   "source_url": "https://omlx.ai/benchmarks/performance/fo52qdxb",
   "tg_tokens": null,
   "tps": 82.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fo52qdxb-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: 0e2tds33, 3y1nd5ub, lk1u577g; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1311.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.6-35B-A3B-mxfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 1,311 49.9 24.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fo52qdxb",
   "source_url": "https://omlx.ai/benchmarks/performance/fo52qdxb",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fo52qdxb-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: 0e2tds33, 3y1nd5ub, lk1u577g; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 835.7,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.6-35B-A3B-mxfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 128k 835.7 33.0 29.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fo52qdxb",
   "source_url": "https://omlx.ai/benchmarks/performance/fo52qdxb",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-zuwte087-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 16.1 GB; TTFT 15412 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 66.4,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 66.4 PP tok/s 4.5 TG tok/s 15412 TTFT (ms) 16.1 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zuwte087",
   "source_url": "https://omlx.ai/benchmarks/performance/zuwte087",
   "tg_tokens": null,
   "tps": 4.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-x4t2e6dg-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 15.0 GB; TTFT 15530 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 65.9,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 65.9 PP tok/s 7.2 TG tok/s 15530 TTFT (ms) 15.0 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x4t2e6dg",
   "source_url": "https://omlx.ai/benchmarks/performance/x4t2e6dg",
   "tg_tokens": null,
   "tps": 7.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-0zs9388g-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 29.1 GB at this context; duplicate traces of the same session: 0uqj6ceu, 3md1gp0j, 6c064q9u, am32l5gd, apihyvz2, dzvm5in3, qde2pfew; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 418.3,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 418.3 27.9 29.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0zs9388g",
   "source_url": "https://omlx.ai/benchmarks/performance/0zs9388g",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-0zs9388g-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 30.0 GB at this context; duplicate traces of the same session: 0uqj6ceu, 3md1gp0j, 6c064q9u, am32l5gd, apihyvz2, dzvm5in3, qde2pfew; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 460.3,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 460.3 25.3 30.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0zs9388g",
   "source_url": "https://omlx.ai/benchmarks/performance/0zs9388g",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-0zs9388g-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: 0uqj6ceu, 3md1gp0j, 6c064q9u, am32l5gd, apihyvz2, dzvm5in3, qde2pfew; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 469.3,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 469.3 30.6 30.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0zs9388g",
   "source_url": "https://omlx.ai/benchmarks/performance/0zs9388g",
   "tg_tokens": null,
   "tps": 30.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-0zs9388g-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 30.9 GB at this context; duplicate traces of the same session: 0uqj6ceu, 3md1gp0j, 6c064q9u, am32l5gd, apihyvz2, dzvm5in3, qde2pfew; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 456.7,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 16k 456.7 29.4 30.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0zs9388g",
   "source_url": "https://omlx.ai/benchmarks/performance/0zs9388g",
   "tg_tokens": null,
   "tps": 29.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-0zs9388g-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: 0uqj6ceu, 3md1gp0j, 6c064q9u, am32l5gd, apihyvz2, dzvm5in3, qde2pfew; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 428.1,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 32k 428.1 25.2 32.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0zs9388g",
   "source_url": "https://omlx.ai/benchmarks/performance/0zs9388g",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-0zs9388g-ctx65536",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 35.0 GB at this context; duplicate traces of the same session: 0uqj6ceu, 3md1gp0j, 6c064q9u, am32l5gd, apihyvz2, dzvm5in3, qde2pfew; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 359.5,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 64k 359.5 20.7 35.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0zs9388g",
   "source_url": "https://omlx.ai/benchmarks/performance/0zs9388g",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-0zs9388g-ctx131072",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 40.5 GB at this context; duplicate traces of the same session: 0uqj6ceu, 3md1gp0j, 6c064q9u, am32l5gd, apihyvz2, dzvm5in3, qde2pfew; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 287.6,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 128k 287.6 15.3 40.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0zs9388g",
   "source_url": "https://omlx.ai/benchmarks/performance/0zs9388g",
   "tg_tokens": null,
   "tps": 15.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-0zs9388g-ctx199680",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 45.6 GB at this context; duplicate traces of the same session: 0uqj6ceu, 3md1gp0j, 6c064q9u, am32l5gd, apihyvz2, dzvm5in3, qde2pfew; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 237.6,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 195k 237.6 11.0 45.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0zs9388g",
   "source_url": "https://omlx.ai/benchmarks/performance/0zs9388g",
   "tg_tokens": null,
   "tps": 11.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-poypj5gu-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: i3iqp24y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1210.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,210 90.6 70.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/poypj5gu",
   "source_url": "https://omlx.ai/benchmarks/performance/poypj5gu",
   "tg_tokens": null,
   "tps": 90.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-poypj5gu-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: i3iqp24y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2112.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,112 86.9 71.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/poypj5gu",
   "source_url": "https://omlx.ai/benchmarks/performance/poypj5gu",
   "tg_tokens": null,
   "tps": 86.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ll0nr7wj-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: 8qmaw8fd, dtqm0jhr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1203.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,203 88.7 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ll0nr7wj",
   "source_url": "https://omlx.ai/benchmarks/performance/ll0nr7wj",
   "tg_tokens": null,
   "tps": 88.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ll0nr7wj-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 8qmaw8fd, dtqm0jhr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2100.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,100 85.3 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ll0nr7wj",
   "source_url": "https://omlx.ai/benchmarks/performance/ll0nr7wj",
   "tg_tokens": null,
   "tps": 85.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ll0nr7wj-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 73.3 GB at this context; duplicate traces of the same session: 8qmaw8fd, dtqm0jhr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2654.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,654 77.5 73.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ll0nr7wj",
   "source_url": "https://omlx.ai/benchmarks/performance/ll0nr7wj",
   "tg_tokens": null,
   "tps": 77.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-azqtq8j8-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: 3pnrmdel, v8u106wh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1165.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,165 89.5 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/azqtq8j8",
   "source_url": "https://omlx.ai/benchmarks/performance/azqtq8j8",
   "tg_tokens": null,
   "tps": 89.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-azqtq8j8-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 3pnrmdel, v8u106wh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2083.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,083 75.2 70.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/azqtq8j8",
   "source_url": "https://omlx.ai/benchmarks/performance/azqtq8j8",
   "tg_tokens": null,
   "tps": 75.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-azqtq8j8-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: 3pnrmdel, v8u106wh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2657.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,657 81.0 73.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/azqtq8j8",
   "source_url": "https://omlx.ai/benchmarks/performance/azqtq8j8",
   "tg_tokens": null,
   "tps": 81.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-qstnq82e-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 4.7 GB at this context; duplicate traces of the same session: jul2j1rh; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2196.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 2,196 39.6 4.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qstnq82e",
   "source_url": "https://omlx.ai/benchmarks/performance/qstnq82e",
   "tg_tokens": null,
   "tps": 39.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-qstnq82e-ctx16384",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 5.4 GB at this context; duplicate traces of the same session: jul2j1rh; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2045.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 16k 2,045 33.9 5.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qstnq82e",
   "source_url": "https://omlx.ai/benchmarks/performance/qstnq82e",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-qz3s7go0-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "mlx-community; oMLX community benchmark trace; peak memory 4.7 GB at this context; duplicate traces of the same session: iqn105mp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1053.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--gemma-4-e4b-it-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 1,053 69.3 4.7 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qz3s7go0",
   "source_url": "https://omlx.ai/benchmarks/performance/qz3s7go0",
   "tg_tokens": null,
   "tps": 69.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-qz3s7go0-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "mlx-community; oMLX community benchmark trace; peak memory 4.7 GB at this context; duplicate traces of the same session: iqn105mp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1063.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--gemma-4-e4b-it-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 1,063 67.4 4.7 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qz3s7go0",
   "source_url": "https://omlx.ai/benchmarks/performance/qz3s7go0",
   "tg_tokens": null,
   "tps": 67.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-fmp29odl-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.2 GB at this context; duplicate traces of the same session: hf6q0ezb, jw2rh0b3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2276.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,276 125.0 173.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fmp29odl",
   "source_url": "https://omlx.ai/benchmarks/performance/fmp29odl",
   "tg_tokens": null,
   "tps": 125.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-fmp29odl-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.7 GB at this context; duplicate traces of the same session: hf6q0ezb, jw2rh0b3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2254.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,254 82.9 173.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fmp29odl",
   "source_url": "https://omlx.ai/benchmarks/performance/fmp29odl",
   "tg_tokens": null,
   "tps": 82.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-fmp29odl-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.9 GB at this context; duplicate traces of the same session: hf6q0ezb, jw2rh0b3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2198.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,198 126.8 174.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fmp29odl",
   "source_url": "https://omlx.ai/benchmarks/performance/fmp29odl",
   "tg_tokens": null,
   "tps": 126.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-h3d0mc23-ctx4096",
   "model": "Laguna-S-2.1",
   "notes": "oMLX community benchmark trace; peak memory 45.9 GB at this context; duplicate traces of the same session: htnjum3k; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 108.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Laguna-S-2.1-oQ6e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 108.4 8.6 45.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h3d0mc23",
   "source_url": "https://omlx.ai/benchmarks/performance/h3d0mc23",
   "tg_tokens": null,
   "tps": 8.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-h3d0mc23-ctx16384",
   "model": "Laguna-S-2.1",
   "notes": "oMLX community benchmark trace; peak memory 46.6 GB at this context; duplicate traces of the same session: htnjum3k; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 115.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Laguna-S-2.1-oQ6e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 16k 115.8 9.2 46.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h3d0mc23",
   "source_url": "https://omlx.ai/benchmarks/performance/h3d0mc23",
   "tg_tokens": null,
   "tps": 9.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-ox8fun61-ctx1024",
   "model": "Nanbeige4.2-3B",
   "notes": "oMLX community benchmark trace; peak memory 3.0 GB at this context; duplicate traces of the same session: b5ou36vd; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 245.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Nanbeige4.2-3B-oQ4e on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 245.4 21.0 3.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ox8fun61",
   "source_url": "https://omlx.ai/benchmarks/performance/ox8fun61",
   "tg_tokens": null,
   "tps": 21.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-ox8fun61-ctx4096",
   "model": "Nanbeige4.2-3B",
   "notes": "oMLX community benchmark trace; peak memory 3.5 GB at this context; duplicate traces of the same session: b5ou36vd; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Nanbeige4.2-3B-oQ4e on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 220.9 17.1 3.5 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ox8fun61",
   "source_url": "https://omlx.ai/benchmarks/performance/ox8fun61",
   "tg_tokens": null,
   "tps": 17.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ltta3hej-ctx1024",
   "model": "Occamy-1.0",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: 3bh2egv6, na7vgntb, ob3is23n, qxhac4y1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4117.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Occamy-1.0-oQ8e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 4,117 94.3 35.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ltta3hej",
   "source_url": "https://omlx.ai/benchmarks/performance/ltta3hej",
   "tg_tokens": null,
   "tps": 94.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ltta3hej-ctx32768",
   "model": "Occamy-1.0",
   "notes": "oMLX community benchmark trace; peak memory 36.7 GB at this context; duplicate traces of the same session: 3bh2egv6, na7vgntb, ob3is23n, qxhac4y1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4410.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Occamy-1.0-oQ8e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 4,410 83.5 36.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ltta3hej",
   "source_url": "https://omlx.ai/benchmarks/performance/ltta3hej",
   "tg_tokens": null,
   "tps": 83.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ltta3hej-ctx65536",
   "model": "Occamy-1.0",
   "notes": "oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: 3bh2egv6, na7vgntb, ob3is23n, qxhac4y1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3207.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Occamy-1.0-oQ8e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 3,207 71.2 37.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ltta3hej",
   "source_url": "https://omlx.ai/benchmarks/performance/ltta3hej",
   "tg_tokens": null,
   "tps": 71.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ltta3hej-ctx131072",
   "model": "Occamy-1.0",
   "notes": "oMLX community benchmark trace; peak memory 38.5 GB at this context; duplicate traces of the same session: 3bh2egv6, na7vgntb, ob3is23n, qxhac4y1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2173.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Occamy-1.0-oQ8e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 2,173 58.2 38.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ltta3hej",
   "source_url": "https://omlx.ai/benchmarks/performance/ltta3hej",
   "tg_tokens": null,
   "tps": 58.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ltta3hej-ctx199680",
   "model": "Occamy-1.0",
   "notes": "oMLX community benchmark trace; peak memory 39.9 GB at this context; duplicate traces of the same session: 3bh2egv6, na7vgntb, ob3is23n, qxhac4y1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1593.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Occamy-1.0-oQ8e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 195k 1,593 49.5 39.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ltta3hej",
   "source_url": "https://omlx.ai/benchmarks/performance/ltta3hej",
   "tg_tokens": null,
   "tps": 49.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-oiaa73qc-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: 0hqi3fo0, v0wfirx3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 446.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 4k 446.6 23.5 32.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oiaa73qc",
   "source_url": "https://omlx.ai/benchmarks/performance/oiaa73qc",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-oiaa73qc-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.7 GB at this context; duplicate traces of the same session: 0hqi3fo0, v0wfirx3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 424.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 16k 424.1 22.4 32.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oiaa73qc",
   "source_url": "https://omlx.ai/benchmarks/performance/oiaa73qc",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-oiaa73qc-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.2 GB at this context; duplicate traces of the same session: 0hqi3fo0, v0wfirx3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 400.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 32k 400.9 21.6 34.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oiaa73qc",
   "source_url": "https://omlx.ai/benchmarks/performance/oiaa73qc",
   "tg_tokens": null,
   "tps": 21.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ked7cmsy-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 28.3 GB at this context; duplicate traces of the same session: q34zkv80; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 115.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ8e-1M on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 1k 115.3 12.1 28.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ked7cmsy",
   "source_url": "https://omlx.ai/benchmarks/performance/ked7cmsy",
   "tg_tokens": null,
   "tps": 12.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ked7cmsy-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: q34zkv80; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 114.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ8e-1M on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 4k 114.8 11.9 32.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ked7cmsy",
   "source_url": "https://omlx.ai/benchmarks/performance/ked7cmsy",
   "tg_tokens": null,
   "tps": 11.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-luels52c-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: f790lhd0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1019.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,019 62.3 101.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/luels52c",
   "source_url": "https://omlx.ai/benchmarks/performance/luels52c",
   "tg_tokens": null,
   "tps": 62.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-luels52c-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.7 GB at this context; duplicate traces of the same session: f790lhd0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 954.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 954.0 60.6 103.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/luels52c",
   "source_url": "https://omlx.ai/benchmarks/performance/luels52c",
   "tg_tokens": null,
   "tps": 60.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ktv09rhp-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "mlx-community; oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: rgta6r55; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1049.0,
   "provenance": "community",
   "quant": "qat-OptiQ-4bit",
   "quote": "mlx-community--gemma-4-e4b-it-qat-OptiQ-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 1,049 55.9 6.8 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ktv09rhp",
   "source_url": "https://omlx.ai/benchmarks/performance/ktv09rhp",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ktv09rhp-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "mlx-community; oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: rgta6r55; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1060.0,
   "provenance": "community",
   "quant": "qat-OptiQ-4bit",
   "quote": "mlx-community--gemma-4-e4b-it-qat-OptiQ-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 1,060 54.1 6.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ktv09rhp",
   "source_url": "https://omlx.ai/benchmarks/performance/ktv09rhp",
   "tg_tokens": null,
   "tps": 54.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-85g3483e-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: 17mp04m8, 7p827z2a, cix08mpc, dcexsf74, ebuqdm3v, uf0vobd8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1898.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 1,898 77.0 37.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/85g3483e",
   "source_url": "https://omlx.ai/benchmarks/performance/85g3483e",
   "tg_tokens": null,
   "tps": 77.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-85g3483e-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.0 GB at this context; duplicate traces of the same session: 17mp04m8, 7p827z2a, cix08mpc, dcexsf74, ebuqdm3v, uf0vobd8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2908.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 2,908 80.1 38.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/85g3483e",
   "source_url": "https://omlx.ai/benchmarks/performance/85g3483e",
   "tg_tokens": null,
   "tps": 80.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-85g3483e-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.1 GB at this context; duplicate traces of the same session: 17mp04m8, 7p827z2a, cix08mpc, dcexsf74, ebuqdm3v, uf0vobd8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2936.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 2,936 84.0 38.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/85g3483e",
   "source_url": "https://omlx.ai/benchmarks/performance/85g3483e",
   "tg_tokens": null,
   "tps": 84.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-85g3483e-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.3 GB at this context; duplicate traces of the same session: 17mp04m8, 7p827z2a, cix08mpc, dcexsf74, ebuqdm3v, uf0vobd8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2719.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 16k 2,719 81.2 38.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/85g3483e",
   "source_url": "https://omlx.ai/benchmarks/performance/85g3483e",
   "tg_tokens": null,
   "tps": 81.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-85g3483e-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.7 GB at this context; duplicate traces of the same session: 17mp04m8, 7p827z2a, cix08mpc, dcexsf74, ebuqdm3v, uf0vobd8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2296.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 32k 2,296 62.5 38.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/85g3483e",
   "source_url": "https://omlx.ai/benchmarks/performance/85g3483e",
   "tg_tokens": null,
   "tps": 62.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-85g3483e-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.4 GB at this context; duplicate traces of the same session: 17mp04m8, 7p827z2a, cix08mpc, dcexsf74, ebuqdm3v, uf0vobd8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1695.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 64k 1,695 39.8 39.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/85g3483e",
   "source_url": "https://omlx.ai/benchmarks/performance/85g3483e",
   "tg_tokens": null,
   "tps": 39.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-85g3483e-ctx131072",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 40.9 GB at this context; duplicate traces of the same session: 17mp04m8, 7p827z2a, cix08mpc, dcexsf74, ebuqdm3v, uf0vobd8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1128.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 128k 1,128 26.7 40.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/85g3483e",
   "source_url": "https://omlx.ai/benchmarks/performance/85g3483e",
   "tg_tokens": null,
   "tps": 26.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-d4hnhsbf-ctx1024",
   "model": "Qwen2.5-1.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 1.5 GB at this context; duplicate traces of the same session: 163njlhe; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 522.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-1.5B-Instruct-4bit on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 4bit; 1k 522.1 40.5 1.5 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d4hnhsbf",
   "source_url": "https://omlx.ai/benchmarks/performance/d4hnhsbf",
   "tg_tokens": null,
   "tps": 40.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-d4hnhsbf-ctx4096",
   "model": "Qwen2.5-1.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 1.6 GB at this context; duplicate traces of the same session: 163njlhe; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 492.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-1.5B-Instruct-4bit on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 4bit; 4k 492.3 36.0 1.6 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d4hnhsbf",
   "source_url": "https://omlx.ai/benchmarks/performance/d4hnhsbf",
   "tg_tokens": null,
   "tps": 36.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-92o11kdj-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; fp16; oMLX 0.7.0; oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: 06j4w64d, 6rkp15tv, dcokj110, pei08uqk; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 156.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Uncensored-oQ8e-fp16-mtp-omlx-0.7.0 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 156.5 16.0 34.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/92o11kdj",
   "source_url": "https://omlx.ai/benchmarks/performance/92o11kdj",
   "tg_tokens": null,
   "tps": 16.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-92o11kdj-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; fp16; oMLX 0.7.0; oMLX community benchmark trace; peak memory 38.5 GB at this context; duplicate traces of the same session: 06j4w64d, 6rkp15tv, dcokj110, pei08uqk; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 139.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Uncensored-oQ8e-fp16-mtp-omlx-0.7.0 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 139.4 14.7 38.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/92o11kdj",
   "source_url": "https://omlx.ai/benchmarks/performance/92o11kdj",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-92o11kdj-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; fp16; oMLX 0.7.0; oMLX community benchmark trace; peak memory 39.5 GB at this context; duplicate traces of the same session: 06j4w64d, 6rkp15tv, dcokj110, pei08uqk; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 137.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Uncensored-oQ8e-fp16-mtp-omlx-0.7.0 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 8k 137.1 14.9 39.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/92o11kdj",
   "source_url": "https://omlx.ai/benchmarks/performance/92o11kdj",
   "tg_tokens": null,
   "tps": 14.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-92o11kdj-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; fp16; oMLX 0.7.0; oMLX community benchmark trace; peak memory 41.2 GB at this context; duplicate traces of the same session: 06j4w64d, 6rkp15tv, dcokj110, pei08uqk; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Uncensored-oQ8e-fp16-mtp-omlx-0.7.0 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 16k 130.2 17.6 41.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/92o11kdj",
   "source_url": "https://omlx.ai/benchmarks/performance/92o11kdj",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-92o11kdj-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; fp16; oMLX 0.7.0; oMLX community benchmark trace; peak memory 42.8 GB at this context; duplicate traces of the same session: 06j4w64d, 6rkp15tv, dcokj110, pei08uqk; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 120.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Uncensored-oQ8e-fp16-mtp-omlx-0.7.0 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 32k 120.6 16.1 42.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/92o11kdj",
   "source_url": "https://omlx.ai/benchmarks/performance/92o11kdj",
   "tg_tokens": null,
   "tps": 16.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kqcy9y83-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.6 GB; TTFT 989 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1035.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,035 PP tok/s 67.8 TG tok/s 989 TTFT (ms) 70.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kqcy9y83",
   "source_url": "https://omlx.ai/benchmarks/performance/kqcy9y83",
   "tg_tokens": null,
   "tps": 67.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gsk1f7fp-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.2 GB at this context; duplicate traces of the same session: g27g9wzm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 984.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 984.1 67.4 69.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gsk1f7fp",
   "source_url": "https://omlx.ai/benchmarks/performance/gsk1f7fp",
   "tg_tokens": null,
   "tps": 67.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gsk1f7fp-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.4 GB at this context; duplicate traces of the same session: g27g9wzm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1777.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,777 66.3 69.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gsk1f7fp",
   "source_url": "https://omlx.ai/benchmarks/performance/gsk1f7fp",
   "tg_tokens": null,
   "tps": 66.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-vs0ssx0m-ctx1024",
   "model": "Gemma-4-31B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: ejw6uasp; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 159.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "gemma-4-31B-it-MLX-5bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 5bit; 1k 159.0 14.8 21.4 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vs0ssx0m",
   "source_url": "https://omlx.ai/benchmarks/performance/vs0ssx0m",
   "tg_tokens": null,
   "tps": 14.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-vs0ssx0m-ctx4096",
   "model": "Gemma-4-31B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: ejw6uasp; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 147.3,
   "provenance": "community",
   "quant": "5bit",
   "quote": "gemma-4-31B-it-MLX-5bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 5bit; 4k 147.3 14.3 22.3 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vs0ssx0m",
   "source_url": "https://omlx.ai/benchmarks/performance/vs0ssx0m",
   "tg_tokens": null,
   "tps": 14.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-2svjsnfs-ctx4096",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 3.7 GB; TTFT 1029 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3980.0,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "gemma-4-e2b-nvfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k tokens 3,980 PP tok/s 94.4 TG tok/s 1029 TTFT (ms) 3.7 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2svjsnfs",
   "source_url": "https://omlx.ai/benchmarks/performance/2svjsnfs",
   "tg_tokens": null,
   "tps": 94.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ek1240b9-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: m9n31nub; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2507.0,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "gemma-4-e4b-it-nvfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 2,507 53.8 6.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ek1240b9",
   "source_url": "https://omlx.ai/benchmarks/performance/ek1240b9",
   "tg_tokens": null,
   "tps": 53.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ek1240b9-ctx16384",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: m9n31nub; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3491.0,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "gemma-4-e4b-it-nvfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 3,491 48.5 6.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ek1240b9",
   "source_url": "https://omlx.ai/benchmarks/performance/ek1240b9",
   "tg_tokens": null,
   "tps": 48.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-aufslik3-ctx4096",
   "model": "Qwen2.5-Coder-1.5B-Instruct",
   "notes": "mlx-community; oMLX community benchmark trace, single run at 4k context; peak memory 1.6 GB; TTFT 834 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4912.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen2.5-Coder-1.5B-Instruct-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 unknown; 4k tokens 4,912 PP tok/s 261.3 TG tok/s 834 TTFT (ms) 1.6 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aufslik3",
   "source_url": "https://omlx.ai/benchmarks/performance/aufslik3",
   "tg_tokens": null,
   "tps": 261.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-mwzxnm1a-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "MLX; mlx-community; oMLX community benchmark trace, single run at 4k context; peak memory 7.2 GB; TTFT 3336 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1228.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.5-9B-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k tokens 1,228 PP tok/s 111.3 TG tok/s 3336 TTFT (ms) 7.2 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mwzxnm1a",
   "source_url": "https://omlx.ai/benchmarks/performance/mwzxnm1a",
   "tg_tokens": null,
   "tps": 111.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-qyjyj07u-ctx4096",
   "model": "Qwen3.6-27B",
   "notes": "ThinkingCap build; mlx-community; oMLX community benchmark trace, single run at 4k context; peak memory 22.5 GB; TTFT 11845 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 345.8,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "mlx-community--ThinkingCap-Qwen3.6-27B-OptiQ-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k tokens 345.8 PP tok/s 32.8 TG tok/s 11845 TTFT (ms) 22.5 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qyjyj07u",
   "source_url": "https://omlx.ai/benchmarks/performance/qyjyj07u",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-vympsvy8-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 183.6 GB; TTFT 34161 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1919.0,
   "provenance": "community",
   "quant": "mixed 4/8bit",
   "quote": "GLM-5.3-Flash-MLX-mixed-4_8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k tokens 1,919 PP tok/s 62.4 TG tok/s 34161 TTFT (ms) 183.6 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vympsvy8",
   "source_url": "https://omlx.ai/benchmarks/performance/vympsvy8",
   "tg_tokens": null,
   "tps": 62.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-l10wvgao-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 183.6 GB; TTFT 34167 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1918.0,
   "provenance": "community",
   "quant": "mixed 4/8bit",
   "quote": "GLM-5.3-Flash-MLX-mixed-4_8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k tokens 1,918 PP tok/s 63.2 TG tok/s 34167 TTFT (ms) 183.6 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l10wvgao",
   "source_url": "https://omlx.ai/benchmarks/performance/l10wvgao",
   "tg_tokens": null,
   "tps": 63.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-a54on2vv-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 183.6 GB; TTFT 34177 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1918.0,
   "provenance": "community",
   "quant": "mixed 4/8bit",
   "quote": "GLM-5.3-Flash-MLX-mixed-4_8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k tokens 1,918 PP tok/s 49.9 TG tok/s 34177 TTFT (ms) 183.6 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a54on2vv",
   "source_url": "https://omlx.ai/benchmarks/performance/a54on2vv",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-3rv5zooh-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 189.3 GB; TTFT 69651 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1882.0,
   "provenance": "community",
   "quant": "mixed 4/8bit",
   "quote": "GLM-5.3-Flash-MLX-mixed-4_8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k tokens 1,882 PP tok/s 91.5 TG tok/s 69651 TTFT (ms) 189.3 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3rv5zooh",
   "source_url": "https://omlx.ai/benchmarks/performance/3rv5zooh",
   "tg_tokens": null,
   "tps": 91.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-01fjfeho-ctx4096",
   "model": "Laguna-S-2.1",
   "notes": "fast build; oMLX community benchmark trace; peak memory 40.0 GB at this context; duplicate traces of the same session: v5s94p16; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 194.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Laguna-S-2.1-oQ4e-fast on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 194.3 14.3 40.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/01fjfeho",
   "source_url": "https://omlx.ai/benchmarks/performance/01fjfeho",
   "tg_tokens": null,
   "tps": 14.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-01fjfeho-ctx16384",
   "model": "Laguna-S-2.1",
   "notes": "fast build; oMLX community benchmark trace; peak memory 40.6 GB at this context; duplicate traces of the same session: v5s94p16; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 208.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Laguna-S-2.1-oQ4e-fast on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 208.6 14.7 40.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/01fjfeho",
   "source_url": "https://omlx.ai/benchmarks/performance/01fjfeho",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-eyftb3du-ctx4096",
   "model": "Laguna-S-2.1",
   "notes": "fast build; oMLX community benchmark trace; peak memory 40.0 GB at this context; duplicate traces of the same session: sfa7h062; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 199.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Laguna-S-2.1-oQ4e-fast on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 199.2 14.7 40.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eyftb3du",
   "source_url": "https://omlx.ai/benchmarks/performance/eyftb3du",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-eyftb3du-ctx16384",
   "model": "Laguna-S-2.1",
   "notes": "fast build; oMLX community benchmark trace; peak memory 40.6 GB at this context; duplicate traces of the same session: sfa7h062; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Laguna-S-2.1-oQ4e-fast on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 205.5 14.8 40.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eyftb3du",
   "source_url": "https://omlx.ai/benchmarks/performance/eyftb3du",
   "tg_tokens": null,
   "tps": 14.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-joazf6wb-ctx1024",
   "model": "Qwen2.5-VL-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 4.5 GB at this context; duplicate traces of the same session: 0de4b8ee; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 145.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-VL-7B-Instruct-4bit on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 4bit; 1k 145.0 13.7 4.5 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/joazf6wb",
   "source_url": "https://omlx.ai/benchmarks/performance/joazf6wb",
   "tg_tokens": null,
   "tps": 13.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-joazf6wb-ctx4096",
   "model": "Qwen2.5-VL-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: 0de4b8ee; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 140.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-VL-7B-Instruct-4bit on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 4bit; 4k 140.6 13.0 4.9 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/joazf6wb",
   "source_url": "https://omlx.ai/benchmarks/performance/joazf6wb",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-hzssxo4q-ctx1024",
   "model": "Qwen2.5-VL-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 4.8 GB at this context; duplicate traces of the same session: 4sceap93; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 449.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-VL-7B-Instruct-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 449.4 65.7 4.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hzssxo4q",
   "source_url": "https://omlx.ai/benchmarks/performance/hzssxo4q",
   "tg_tokens": null,
   "tps": 65.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-hzssxo4q-ctx4096",
   "model": "Qwen2.5-VL-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: 4sceap93; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 436.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-VL-7B-Instruct-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 436.3 59.5 4.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hzssxo4q",
   "source_url": "https://omlx.ai/benchmarks/performance/hzssxo4q",
   "tg_tokens": null,
   "tps": 59.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-gowzl4hh-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 0hwkwgwk, 5mc8vsar, 9kko94z7, szz23gz9, tb6i1naj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 857.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 857.8 88.1 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gowzl4hh",
   "source_url": "https://omlx.ai/benchmarks/performance/gowzl4hh",
   "tg_tokens": null,
   "tps": 88.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-gowzl4hh-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 0hwkwgwk, 5mc8vsar, 9kko94z7, szz23gz9, tb6i1naj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 831.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 831.3 82.2 20.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gowzl4hh",
   "source_url": "https://omlx.ai/benchmarks/performance/gowzl4hh",
   "tg_tokens": null,
   "tps": 82.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-gowzl4hh-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: 0hwkwgwk, 5mc8vsar, 9kko94z7, szz23gz9, tb6i1naj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 746.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 746.0 75.4 20.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gowzl4hh",
   "source_url": "https://omlx.ai/benchmarks/performance/gowzl4hh",
   "tg_tokens": null,
   "tps": 75.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-gowzl4hh-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: 0hwkwgwk, 5mc8vsar, 9kko94z7, szz23gz9, tb6i1naj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 613.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 613.1 63.1 21.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gowzl4hh",
   "source_url": "https://omlx.ai/benchmarks/performance/gowzl4hh",
   "tg_tokens": null,
   "tps": 63.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-gowzl4hh-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 0hwkwgwk, 5mc8vsar, 9kko94z7, szz23gz9, tb6i1naj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 448.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 64k 448.6 47.2 22.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gowzl4hh",
   "source_url": "https://omlx.ai/benchmarks/performance/gowzl4hh",
   "tg_tokens": null,
   "tps": 47.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-gowzl4hh-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: 0hwkwgwk, 5mc8vsar, 9kko94z7, szz23gz9, tb6i1naj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 280.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 128k 280.6 32.5 24.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gowzl4hh",
   "source_url": "https://omlx.ai/benchmarks/performance/gowzl4hh",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-nzg8mg6i-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "ms build; oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 6vnm6zpu; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 84.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-ms on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 84.2 15.5 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nzg8mg6i",
   "source_url": "https://omlx.ai/benchmarks/performance/nzg8mg6i",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-nzg8mg6i-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "ms build; oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 6vnm6zpu; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 78.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-ms on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 78.0 15.1 20.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nzg8mg6i",
   "source_url": "https://omlx.ai/benchmarks/performance/nzg8mg6i",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-do5kqbzu-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: rds4sgko; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 450.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 450.3 30.2 25.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/do5kqbzu",
   "source_url": "https://omlx.ai/benchmarks/performance/do5kqbzu",
   "tg_tokens": null,
   "tps": 30.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-do5kqbzu-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: rds4sgko; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 459.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 459.6 23.0 27.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/do5kqbzu",
   "source_url": "https://omlx.ai/benchmarks/performance/do5kqbzu",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-iq1o8pjk-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: 768ok4b2, ak8o25lv, hshie487, oxe6du3i, p00qqrcl, rk4i5q5o; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1566.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 4bit; 1k 1,566 111.7 17.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iq1o8pjk",
   "source_url": "https://omlx.ai/benchmarks/performance/iq1o8pjk",
   "tg_tokens": null,
   "tps": 111.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-iq1o8pjk-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 768ok4b2, ak8o25lv, hshie487, oxe6du3i, p00qqrcl, rk4i5q5o; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1899.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 4bit; 4k 1,899 100.8 18.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iq1o8pjk",
   "source_url": "https://omlx.ai/benchmarks/performance/iq1o8pjk",
   "tg_tokens": null,
   "tps": 100.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-iq1o8pjk-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 768ok4b2, ak8o25lv, hshie487, oxe6du3i, p00qqrcl, rk4i5q5o; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1908.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 4bit; 8k 1,908 125.5 19.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iq1o8pjk",
   "source_url": "https://omlx.ai/benchmarks/performance/iq1o8pjk",
   "tg_tokens": null,
   "tps": 125.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-iq1o8pjk-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 768ok4b2, ak8o25lv, hshie487, oxe6du3i, p00qqrcl, rk4i5q5o; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1833.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 4bit; 16k 1,833 111.7 19.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iq1o8pjk",
   "source_url": "https://omlx.ai/benchmarks/performance/iq1o8pjk",
   "tg_tokens": null,
   "tps": 111.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-iq1o8pjk-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 768ok4b2, ak8o25lv, hshie487, oxe6du3i, p00qqrcl, rk4i5q5o; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1729.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 4bit; 32k 1,729 130.9 20.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iq1o8pjk",
   "source_url": "https://omlx.ai/benchmarks/performance/iq1o8pjk",
   "tg_tokens": null,
   "tps": 130.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-iq1o8pjk-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: 768ok4b2, ak8o25lv, hshie487, oxe6du3i, p00qqrcl, rk4i5q5o; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1494.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 4bit; 64k 1,494 99.9 23.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iq1o8pjk",
   "source_url": "https://omlx.ai/benchmarks/performance/iq1o8pjk",
   "tg_tokens": null,
   "tps": 99.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-iq1o8pjk-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: 768ok4b2, ak8o25lv, hshie487, oxe6du3i, p00qqrcl, rk4i5q5o; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1214.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 4bit; 128k 1,214 92.6 27.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iq1o8pjk",
   "source_url": "https://omlx.ai/benchmarks/performance/iq1o8pjk",
   "tg_tokens": null,
   "tps": 92.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-7ww6ap7s-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: jdwj50fw; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1529.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 4bit; 4k 1,529 54.1 18.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ww6ap7s",
   "source_url": "https://omlx.ai/benchmarks/performance/7ww6ap7s",
   "tg_tokens": null,
   "tps": 54.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-7ww6ap7s-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: jdwj50fw; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1841.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 4bit; 16k 1,841 116.0 19.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ww6ap7s",
   "source_url": "https://omlx.ai/benchmarks/performance/7ww6ap7s",
   "tg_tokens": null,
   "tps": 116.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-epvfmbka-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.2 GB at this context; duplicate traces of the same session: ewi301qp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1315.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,315 82.2 69.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/epvfmbka",
   "source_url": "https://omlx.ai/benchmarks/performance/epvfmbka",
   "tg_tokens": null,
   "tps": 82.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-epvfmbka-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.4 GB at this context; duplicate traces of the same session: ewi301qp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2055.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,055 73.9 69.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/epvfmbka",
   "source_url": "https://omlx.ai/benchmarks/performance/epvfmbka",
   "tg_tokens": null,
   "tps": 73.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-c44q0t9u-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 69.5 GB; TTFT 2363 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1734.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 1,734 PP tok/s 57.5 TG tok/s 2363 TTFT (ms) 69.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c44q0t9u",
   "source_url": "https://omlx.ai/benchmarks/performance/c44q0t9u",
   "tg_tokens": null,
   "tps": 57.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xiw3rqig-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 195k context; peak memory 82.0 GB; TTFT 115572 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1731.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k tokens 1,731 PP tok/s 57.7 TG tok/s 115572 TTFT (ms) 82.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xiw3rqig",
   "source_url": "https://omlx.ai/benchmarks/performance/xiw3rqig",
   "tg_tokens": null,
   "tps": 57.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-aao2vqbd-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 69.2 GB; TTFT 873 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1173.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,173 PP tok/s 76.6 TG tok/s 873 TTFT (ms) 69.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aao2vqbd",
   "source_url": "https://omlx.ai/benchmarks/performance/aao2vqbd",
   "tg_tokens": null,
   "tps": 76.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fqjg1fmq-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: tpj3pzav; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1276.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,276 83.2 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fqjg1fmq",
   "source_url": "https://omlx.ai/benchmarks/performance/fqjg1fmq",
   "tg_tokens": null,
   "tps": 83.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fqjg1fmq-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: tpj3pzav; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1986.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,986 83.8 70.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fqjg1fmq",
   "source_url": "https://omlx.ai/benchmarks/performance/fqjg1fmq",
   "tg_tokens": null,
   "tps": 83.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0t18yauy-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 69.2 GB; TTFT 1898 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 539.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 539.6 PP tok/s 61.2 TG tok/s 1898 TTFT (ms) 69.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0t18yauy",
   "source_url": "https://omlx.ai/benchmarks/performance/0t18yauy",
   "tg_tokens": null,
   "tps": 61.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-mbyps5hv-ctx1024",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 80.0 GB at this context; duplicate traces of the same session: 0xopeej6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 290.0,
   "provenance": "community",
   "quant": "2.4bit mixed",
   "quote": "DeepSeek-V4-Flash-0731-2.4bit-mixed on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 2bit; 1k 290.0 34.6 80.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mbyps5hv",
   "source_url": "https://omlx.ai/benchmarks/performance/mbyps5hv",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-mbyps5hv-ctx4096",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 82.1 GB at this context; duplicate traces of the same session: 0xopeej6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 267.0,
   "provenance": "community",
   "quant": "2.4bit mixed",
   "quote": "DeepSeek-V4-Flash-0731-2.4bit-mixed on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 2bit; 4k 267.0 31.2 82.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mbyps5hv",
   "source_url": "https://omlx.ai/benchmarks/performance/mbyps5hv",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-s6cc1kam-ctx1024",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 80.0 GB at this context; duplicate traces of the same session: qu81bebz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 261.6,
   "provenance": "community",
   "quant": "2.4bit mixed",
   "quote": "DeepSeek-V4-Flash-0731-2.4bit-mixed on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 2bit; 1k 261.6 34.8 80.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s6cc1kam",
   "source_url": "https://omlx.ai/benchmarks/performance/s6cc1kam",
   "tg_tokens": null,
   "tps": 34.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-s6cc1kam-ctx4096",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 82.1 GB at this context; duplicate traces of the same session: qu81bebz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 261.7,
   "provenance": "community",
   "quant": "2.4bit mixed",
   "quote": "DeepSeek-V4-Flash-0731-2.4bit-mixed on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 2bit; 4k 261.7 31.9 82.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s6cc1kam",
   "source_url": "https://omlx.ai/benchmarks/performance/s6cc1kam",
   "tg_tokens": null,
   "tps": 31.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-5j7jpv4h-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.0 GB at this context; duplicate traces of the same session: h4vmnqwf, vjuev91t; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2387.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,387 125.0 175.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5j7jpv4h",
   "source_url": "https://omlx.ai/benchmarks/performance/5j7jpv4h",
   "tg_tokens": null,
   "tps": 125.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-5j7jpv4h-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.8 GB at this context; duplicate traces of the same session: h4vmnqwf, vjuev91t; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2365.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,365 84.3 177.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5j7jpv4h",
   "source_url": "https://omlx.ai/benchmarks/performance/5j7jpv4h",
   "tg_tokens": null,
   "tps": 84.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-5j7jpv4h-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 183.6 GB at this context; duplicate traces of the same session: h4vmnqwf, vjuev91t; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2317.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,317 121.4 183.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5j7jpv4h",
   "source_url": "https://omlx.ai/benchmarks/performance/5j7jpv4h",
   "tg_tokens": null,
   "tps": 121.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2ytfewbo-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.5 GB at this context; duplicate traces of the same session: 6qrqkxpv, fy5p3qrb, letqe1ag, s89jx4qq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 398.0,
   "provenance": "community",
   "quant": "mixed 4/8bit",
   "quote": "GLM-5.3-Flash-MLX-mixed-4_8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 398.0 32.3 174.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ytfewbo",
   "source_url": "https://omlx.ai/benchmarks/performance/2ytfewbo",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2ytfewbo-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.6 GB at this context; duplicate traces of the same session: 6qrqkxpv, fy5p3qrb, letqe1ag, s89jx4qq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 480.5,
   "provenance": "community",
   "quant": "mixed 4/8bit",
   "quote": "GLM-5.3-Flash-MLX-mixed-4_8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 480.5 30.9 175.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ytfewbo",
   "source_url": "https://omlx.ai/benchmarks/performance/2ytfewbo",
   "tg_tokens": null,
   "tps": 30.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2ytfewbo-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 176.8 GB at this context; duplicate traces of the same session: 6qrqkxpv, fy5p3qrb, letqe1ag, s89jx4qq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 469.2,
   "provenance": "community",
   "quant": "mixed 4/8bit",
   "quote": "GLM-5.3-Flash-MLX-mixed-4_8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 469.2 30.4 176.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ytfewbo",
   "source_url": "https://omlx.ai/benchmarks/performance/2ytfewbo",
   "tg_tokens": null,
   "tps": 30.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2ytfewbo-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 176.3 GB at this context; duplicate traces of the same session: 6qrqkxpv, fy5p3qrb, letqe1ag, s89jx4qq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 374.1,
   "provenance": "community",
   "quant": "mixed 4/8bit",
   "quote": "GLM-5.3-Flash-MLX-mixed-4_8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 374.1 30.5 176.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ytfewbo",
   "source_url": "https://omlx.ai/benchmarks/performance/2ytfewbo",
   "tg_tokens": null,
   "tps": 30.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2ytfewbo-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 176.6 GB at this context; duplicate traces of the same session: 6qrqkxpv, fy5p3qrb, letqe1ag, s89jx4qq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 284.9,
   "provenance": "community",
   "quant": "mixed 4/8bit",
   "quote": "GLM-5.3-Flash-MLX-mixed-4_8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 284.9 30.3 176.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ytfewbo",
   "source_url": "https://omlx.ai/benchmarks/performance/2ytfewbo",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-s1zxo4x1-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 179.9 GB at this context; duplicate traces of the same session: 1dxy3o3h, s391nio8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2342.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,342 88.8 179.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s1zxo4x1",
   "source_url": "https://omlx.ai/benchmarks/performance/s1zxo4x1",
   "tg_tokens": null,
   "tps": 88.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-s1zxo4x1-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 182.8 GB at this context; duplicate traces of the same session: 1dxy3o3h, s391nio8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2303.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,303 73.8 182.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s1zxo4x1",
   "source_url": "https://omlx.ai/benchmarks/performance/s1zxo4x1",
   "tg_tokens": null,
   "tps": 73.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-s1zxo4x1-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 188.6 GB at this context; duplicate traces of the same session: 1dxy3o3h, s391nio8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2263.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,263 122.6 188.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s1zxo4x1",
   "source_url": "https://omlx.ai/benchmarks/performance/s1zxo4x1",
   "tg_tokens": null,
   "tps": 122.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-qyoisjmt-ctx1024",
   "model": "Ling-3.0-tiny",
   "notes": "fp16; oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: ewpc3do0; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 611.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ling-3.0-tiny-oQ6e-fp16 on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 6bit; 1k 611.7 41.0 6.9 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qyoisjmt",
   "source_url": "https://omlx.ai/benchmarks/performance/qyoisjmt",
   "tg_tokens": null,
   "tps": 41.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-qyoisjmt-ctx8192",
   "model": "Ling-3.0-tiny",
   "notes": "fp16; oMLX community benchmark trace; peak memory 7.6 GB at this context; duplicate traces of the same session: ewpc3do0; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 644.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ling-3.0-tiny-oQ6e-fp16 on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 6bit; 8k 644.6 36.7 7.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qyoisjmt",
   "source_url": "https://omlx.ai/benchmarks/performance/qyoisjmt",
   "tg_tokens": null,
   "tps": 36.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-sdfv5wh0-ctx1024",
   "model": "MiniCPM5-2B",
   "notes": "fp16; oMLX community benchmark trace; peak memory 2.3 GB at this context; duplicate traces of the same session: pvngynrw; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 404.6,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "MiniCPM5-2B-oQ5-fp16 on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 5bit; 1k 404.6 26.6 2.3 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sdfv5wh0",
   "source_url": "https://omlx.ai/benchmarks/performance/sdfv5wh0",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-sdfv5wh0-ctx8192",
   "model": "MiniCPM5-2B",
   "notes": "fp16; oMLX community benchmark trace; peak memory 2.6 GB at this context; duplicate traces of the same session: pvngynrw; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 333.1,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "MiniCPM5-2B-oQ5-fp16 on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 5bit; 8k 333.1 16.5 2.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sdfv5wh0",
   "source_url": "https://omlx.ai/benchmarks/performance/sdfv5wh0",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-enbfy2v1-ctx1024",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 8dpckvz1, 9xy1g0ic, v2xfjsxz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 440.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-27B-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 440.8 13.0 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/enbfy2v1",
   "source_url": "https://omlx.ai/benchmarks/performance/enbfy2v1",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-enbfy2v1-ctx4096",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 8dpckvz1, 9xy1g0ic, v2xfjsxz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 484.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-27B-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 484.3 12.7 23.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/enbfy2v1",
   "source_url": "https://omlx.ai/benchmarks/performance/enbfy2v1",
   "tg_tokens": null,
   "tps": 12.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-enbfy2v1-ctx32768",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: 8dpckvz1, 9xy1g0ic, v2xfjsxz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 430.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-27B-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 32k 430.1 10.9 25.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/enbfy2v1",
   "source_url": "https://omlx.ai/benchmarks/performance/enbfy2v1",
   "tg_tokens": null,
   "tps": 10.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-enbfy2v1-ctx65536",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace; peak memory 27.0 GB at this context; duplicate traces of the same session: 8dpckvz1, 9xy1g0ic, v2xfjsxz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 372.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-27B-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 64k 372.0 9.6 27.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/enbfy2v1",
   "source_url": "https://omlx.ai/benchmarks/performance/enbfy2v1",
   "tg_tokens": null,
   "tps": 9.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-9wy903ot-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: dmle9irm; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 598.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 598.5 99.2 19.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9wy903ot",
   "source_url": "https://omlx.ai/benchmarks/performance/9wy903ot",
   "tg_tokens": null,
   "tps": 99.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-9wy903ot-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: dmle9irm; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1624.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 1,624 97.1 20.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9wy903ot",
   "source_url": "https://omlx.ai/benchmarks/performance/9wy903ot",
   "tg_tokens": null,
   "tps": 97.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yxzq9mfe-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: lzkio40r, o53zq1ge, t91eadzu, yvwy4gwr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4855.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,855 123.4 20.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yxzq9mfe",
   "source_url": "https://omlx.ai/benchmarks/performance/yxzq9mfe",
   "tg_tokens": null,
   "tps": 123.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yxzq9mfe-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: lzkio40r, o53zq1ge, t91eadzu, yvwy4gwr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4635.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 4,635 104.6 21.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yxzq9mfe",
   "source_url": "https://omlx.ai/benchmarks/performance/yxzq9mfe",
   "tg_tokens": null,
   "tps": 104.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yxzq9mfe-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: lzkio40r, o53zq1ge, t91eadzu, yvwy4gwr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3538.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 3,538 88.6 22.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yxzq9mfe",
   "source_url": "https://omlx.ai/benchmarks/performance/yxzq9mfe",
   "tg_tokens": null,
   "tps": 88.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yxzq9mfe-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: lzkio40r, o53zq1ge, t91eadzu, yvwy4gwr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2431.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,431 69.2 23.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yxzq9mfe",
   "source_url": "https://omlx.ai/benchmarks/performance/yxzq9mfe",
   "tg_tokens": null,
   "tps": 69.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yxzq9mfe-ctx199680",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: lzkio40r, o53zq1ge, t91eadzu, yvwy4gwr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1834.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 1,834 56.5 24.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yxzq9mfe",
   "source_url": "https://omlx.ai/benchmarks/performance/yxzq9mfe",
   "tg_tokens": null,
   "tps": 56.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-0m6i5dwc-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: xn81fwcn; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 114.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 114.8 16.7 25.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0m6i5dwc",
   "source_url": "https://omlx.ai/benchmarks/performance/0m6i5dwc",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-0m6i5dwc-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.2 GB at this context; duplicate traces of the same session: xn81fwcn; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 112.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 16k 112.7 14.9 27.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0m6i5dwc",
   "source_url": "https://omlx.ai/benchmarks/performance/0m6i5dwc",
   "tg_tokens": null,
   "tps": 14.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-erdr1jko-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MLX; oMLX community benchmark trace; peak memory 31.2 GB at this context; duplicate traces of the same session: ijtc84ph; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 95.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 95.1 9.3 31.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/erdr1jko",
   "source_url": "https://omlx.ai/benchmarks/performance/erdr1jko",
   "tg_tokens": null,
   "tps": 9.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-erdr1jko-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MLX; oMLX community benchmark trace; peak memory 31.6 GB at this context; duplicate traces of the same session: ijtc84ph; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 92.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 92.7 6.8 31.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/erdr1jko",
   "source_url": "https://omlx.ai/benchmarks/performance/erdr1jko",
   "tg_tokens": null,
   "tps": 6.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-xqe4in56-ctx1024",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: y940yobs; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 240.1,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 2bit; 1k 240.1 22.3 11.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xqe4in56",
   "source_url": "https://omlx.ai/benchmarks/performance/xqe4in56",
   "tg_tokens": null,
   "tps": 22.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-xqe4in56-ctx4096",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 14.2 GB at this context; duplicate traces of the same session: y940yobs; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 278.9,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 2bit; 4k 278.9 20.1 14.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xqe4in56",
   "source_url": "https://omlx.ai/benchmarks/performance/xqe4in56",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-1mdgigc4-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 2iue0own, 817gvi5j, bx817ssx; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 960.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 960.3 63.2 20.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1mdgigc4",
   "source_url": "https://omlx.ai/benchmarks/performance/1mdgigc4",
   "tg_tokens": null,
   "tps": 63.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M2 Max",
   "id": "omlx-1mdgigc4-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 2iue0own, 817gvi5j, bx817ssx; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 963.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 963.4 16.2 21.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1mdgigc4",
   "source_url": "https://omlx.ai/benchmarks/performance/1mdgigc4",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M2 Max",
   "id": "omlx-1mdgigc4-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 2iue0own, 817gvi5j, bx817ssx; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 809.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 809.1 16.0 22.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1mdgigc4",
   "source_url": "https://omlx.ai/benchmarks/performance/1mdgigc4",
   "tg_tokens": null,
   "tps": 16.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M2 Max",
   "id": "omlx-1mdgigc4-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: 2iue0own, 817gvi5j, bx817ssx; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 635.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 32k 635.5 13.5 22.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1mdgigc4",
   "source_url": "https://omlx.ai/benchmarks/performance/1mdgigc4",
   "tg_tokens": null,
   "tps": 13.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-f0i9a8y5-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: dwoc5jr0, hd99ohb1, tk58753i; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 952.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 952.6 52.4 20.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f0i9a8y5",
   "source_url": "https://omlx.ai/benchmarks/performance/f0i9a8y5",
   "tg_tokens": null,
   "tps": 52.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M2 Max",
   "id": "omlx-f0i9a8y5-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: dwoc5jr0, hd99ohb1, tk58753i; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 964.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 964.2 19.3 21.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f0i9a8y5",
   "source_url": "https://omlx.ai/benchmarks/performance/f0i9a8y5",
   "tg_tokens": null,
   "tps": 19.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M2 Max",
   "id": "omlx-f0i9a8y5-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: dwoc5jr0, hd99ohb1, tk58753i; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 809.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 809.5 12.8 22.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f0i9a8y5",
   "source_url": "https://omlx.ai/benchmarks/performance/f0i9a8y5",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M2 Max",
   "id": "omlx-f0i9a8y5-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: dwoc5jr0, hd99ohb1, tk58753i; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 636.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 32k 636.1 13.4 22.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f0i9a8y5",
   "source_url": "https://omlx.ai/benchmarks/performance/f0i9a8y5",
   "tg_tokens": null,
   "tps": 13.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-n7976sv8-ctx131072",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 83.0 GB; TTFT 674324 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 194.4,
   "provenance": "community",
   "quant": "2.4bit mixed",
   "quote": "DeepSeek-V4-Flash-0731-2.4bit-mixed on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 2bit; 128k tokens 194.4 PP tok/s 28.4 TG tok/s 674324 TTFT (ms) 83.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n7976sv8",
   "source_url": "https://omlx.ai/benchmarks/performance/n7976sv8",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-djzz59n1-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: z0w2qxhn; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2280.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 2,280 106.5 19.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/djzz59n1",
   "source_url": "https://omlx.ai/benchmarks/performance/djzz59n1",
   "tg_tokens": null,
   "tps": 106.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-djzz59n1-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: z0w2qxhn; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2882.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,882 103.0 20.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/djzz59n1",
   "source_url": "https://omlx.ai/benchmarks/performance/djzz59n1",
   "tg_tokens": null,
   "tps": 103.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-tcwdvgtd-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: hvlwt8fl; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 798.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 798.6 26.5 5.8 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tcwdvgtd",
   "source_url": "https://omlx.ai/benchmarks/performance/tcwdvgtd",
   "tg_tokens": null,
   "tps": 26.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-tcwdvgtd-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: hvlwt8fl; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 872.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 872.4 24.6 6.2 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tcwdvgtd",
   "source_url": "https://omlx.ai/benchmarks/performance/tcwdvgtd",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-tgohnkpi-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: htlibpd9; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 82.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 1k 82.3 15.6 16.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tgohnkpi",
   "source_url": "https://omlx.ai/benchmarks/performance/tgohnkpi",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-tgohnkpi-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: htlibpd9; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 80.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k 80.2 15.0 17.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tgohnkpi",
   "source_url": "https://omlx.ai/benchmarks/performance/tgohnkpi",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-wqsuhcuc-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: sjybwdjy; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1216.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,216 55.6 16.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wqsuhcuc",
   "source_url": "https://omlx.ai/benchmarks/performance/wqsuhcuc",
   "tg_tokens": null,
   "tps": 55.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-wqsuhcuc-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: sjybwdjy; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1439.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,439 54.0 17.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wqsuhcuc",
   "source_url": "https://omlx.ai/benchmarks/performance/wqsuhcuc",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-3qt8zdcc-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 15.0 GB; TTFT 15541 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 65.9,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 65.9 PP tok/s 7.2 TG tok/s 15541 TTFT (ms) 15.0 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3qt8zdcc",
   "source_url": "https://omlx.ai/benchmarks/performance/3qt8zdcc",
   "tg_tokens": null,
   "tps": 7.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-l37iwkho-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.6 GB at this context; duplicate traces of the same session: xdbulti9; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 214.9,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 214.9 25.3 34.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l37iwkho",
   "source_url": "https://omlx.ai/benchmarks/performance/l37iwkho",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-l37iwkho-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.5 GB at this context; duplicate traces of the same session: xdbulti9; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 152.4,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 16k 152.4 33.0 34.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l37iwkho",
   "source_url": "https://omlx.ai/benchmarks/performance/l37iwkho",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2gtrqth2-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 75fgor5z, 9xfysv51, b1y6pglj, mpq92kt1, poo5p924; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 324.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 324.9 21.3 21.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2gtrqth2",
   "source_url": "https://omlx.ai/benchmarks/performance/2gtrqth2",
   "tg_tokens": null,
   "tps": 21.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2gtrqth2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: 75fgor5z, 9xfysv51, b1y6pglj, mpq92kt1, poo5p924; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 439.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 439.5 55.1 24.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2gtrqth2",
   "source_url": "https://omlx.ai/benchmarks/performance/2gtrqth2",
   "tg_tokens": null,
   "tps": 55.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2gtrqth2-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: 75fgor5z, 9xfysv51, b1y6pglj, mpq92kt1, poo5p924; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 487.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 487.3 72.0 25.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2gtrqth2",
   "source_url": "https://omlx.ai/benchmarks/performance/2gtrqth2",
   "tg_tokens": null,
   "tps": 72.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2gtrqth2-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: 75fgor5z, 9xfysv51, b1y6pglj, mpq92kt1, poo5p924; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 483.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 483.2 73.7 25.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2gtrqth2",
   "source_url": "https://omlx.ai/benchmarks/performance/2gtrqth2",
   "tg_tokens": null,
   "tps": 73.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2gtrqth2-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.3 GB at this context; duplicate traces of the same session: 75fgor5z, 9xfysv51, b1y6pglj, mpq92kt1, poo5p924; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 449.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 449.8 53.8 27.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2gtrqth2",
   "source_url": "https://omlx.ai/benchmarks/performance/2gtrqth2",
   "tg_tokens": null,
   "tps": 53.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2gtrqth2-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.6 GB at this context; duplicate traces of the same session: 75fgor5z, 9xfysv51, b1y6pglj, mpq92kt1, poo5p924; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 371.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 371.1 57.9 30.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2gtrqth2",
   "source_url": "https://omlx.ai/benchmarks/performance/2gtrqth2",
   "tg_tokens": null,
   "tps": 57.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-nvazg73a-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: kme0pyzj, s55fxjpt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 658.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 658.1 50.3 71.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nvazg73a",
   "source_url": "https://omlx.ai/benchmarks/performance/nvazg73a",
   "tg_tokens": null,
   "tps": 50.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-nvazg73a-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.5 GB at this context; duplicate traces of the same session: kme0pyzj, s55fxjpt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 636.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 636.2 60.1 74.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nvazg73a",
   "source_url": "https://omlx.ai/benchmarks/performance/nvazg73a",
   "tg_tokens": null,
   "tps": 60.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-nvazg73a-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.3 GB at this context; duplicate traces of the same session: kme0pyzj, s55fxjpt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 565.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 565.0 51.2 78.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nvazg73a",
   "source_url": "https://omlx.ai/benchmarks/performance/nvazg73a",
   "tg_tokens": null,
   "tps": 51.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-zafvs5mv-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 7i2a7d53, cx7323rh, u3r31izo; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 957.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 957.3 100.3 19.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zafvs5mv",
   "source_url": "https://omlx.ai/benchmarks/performance/zafvs5mv",
   "tg_tokens": null,
   "tps": 100.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-zafvs5mv-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 7i2a7d53, cx7323rh, u3r31izo; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1024.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 1,024 89.3 22.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zafvs5mv",
   "source_url": "https://omlx.ai/benchmarks/performance/zafvs5mv",
   "tg_tokens": null,
   "tps": 89.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-zafvs5mv-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 7i2a7d53, cx7323rh, u3r31izo; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 871.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 871.9 82.2 21.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zafvs5mv",
   "source_url": "https://omlx.ai/benchmarks/performance/zafvs5mv",
   "tg_tokens": null,
   "tps": 82.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-zafvs5mv-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: 7i2a7d53, cx7323rh, u3r31izo; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 694.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 32k 694.8 72.7 21.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zafvs5mv",
   "source_url": "https://omlx.ai/benchmarks/performance/zafvs5mv",
   "tg_tokens": null,
   "tps": 72.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-rl2ie6zi-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 6jq4r0ad, 923g1ica, xlstows7; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 956.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 956.3 99.5 19.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rl2ie6zi",
   "source_url": "https://omlx.ai/benchmarks/performance/rl2ie6zi",
   "tg_tokens": null,
   "tps": 99.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-rl2ie6zi-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 6jq4r0ad, 923g1ica, xlstows7; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1024.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 1,024 89.6 22.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rl2ie6zi",
   "source_url": "https://omlx.ai/benchmarks/performance/rl2ie6zi",
   "tg_tokens": null,
   "tps": 89.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-rl2ie6zi-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 6jq4r0ad, 923g1ica, xlstows7; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 872.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 872.8 81.2 21.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rl2ie6zi",
   "source_url": "https://omlx.ai/benchmarks/performance/rl2ie6zi",
   "tg_tokens": null,
   "tps": 81.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-rl2ie6zi-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: 6jq4r0ad, 923g1ica, xlstows7; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 695.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 32k 695.2 72.3 21.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rl2ie6zi",
   "source_url": "https://omlx.ai/benchmarks/performance/rl2ie6zi",
   "tg_tokens": null,
   "tps": 72.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-op8prffy-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 3.8 GB; TTFT 7296 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 561.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k tokens 561.4 PP tok/s 36.4 TG tok/s 7296 TTFT (ms) 3.8 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/op8prffy",
   "source_url": "https://omlx.ai/benchmarks/performance/op8prffy",
   "tg_tokens": null,
   "tps": 36.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-czwqyxhg-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 1zd5acof; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 209.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 209.1 7.6 17.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/czwqyxhg",
   "source_url": "https://omlx.ai/benchmarks/performance/czwqyxhg",
   "tg_tokens": null,
   "tps": 7.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-czwqyxhg-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 1zd5acof; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 170.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 16k 170.1 6.7 17.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/czwqyxhg",
   "source_url": "https://omlx.ai/benchmarks/performance/czwqyxhg",
   "tg_tokens": null,
   "tps": 6.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-m3ubuhbh-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: l4sp0z3h; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1066.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 1k 1,066 33.2 28.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m3ubuhbh",
   "source_url": "https://omlx.ai/benchmarks/performance/m3ubuhbh",
   "tg_tokens": null,
   "tps": 33.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-m3ubuhbh-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: l4sp0z3h; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1340.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 4k 1,340 32.9 29.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m3ubuhbh",
   "source_url": "https://omlx.ai/benchmarks/performance/m3ubuhbh",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-umv8yhc3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 38.9 GB at this context; duplicate traces of the same session: tysz7e1z; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 153.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 4k 153.8 20.5 38.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/umv8yhc3",
   "source_url": "https://omlx.ai/benchmarks/performance/umv8yhc3",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-umv8yhc3-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 40.7 GB at this context; duplicate traces of the same session: tysz7e1z; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 143.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 16k 143.6 17.9 40.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/umv8yhc3",
   "source_url": "https://omlx.ai/benchmarks/performance/umv8yhc3",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-qa62htrr-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: j0ij8jfg, v04vbqym, zitufbld; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 656.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 656.0 71.4 71.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qa62htrr",
   "source_url": "https://omlx.ai/benchmarks/performance/qa62htrr",
   "tg_tokens": null,
   "tps": 71.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-qa62htrr-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 72.3 GB at this context; duplicate traces of the same session: j0ij8jfg, v04vbqym, zitufbld; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 645.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 645.1 64.0 72.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qa62htrr",
   "source_url": "https://omlx.ai/benchmarks/performance/qa62htrr",
   "tg_tokens": null,
   "tps": 64.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-qa62htrr-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 74.5 GB at this context; duplicate traces of the same session: j0ij8jfg, v04vbqym, zitufbld; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 610.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 610.6 53.8 74.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qa62htrr",
   "source_url": "https://omlx.ai/benchmarks/performance/qa62htrr",
   "tg_tokens": null,
   "tps": 53.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-qa62htrr-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 78.6 GB at this context; duplicate traces of the same session: j0ij8jfg, v04vbqym, zitufbld; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 535.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 535.0 48.8 78.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qa62htrr",
   "source_url": "https://omlx.ai/benchmarks/performance/qa62htrr",
   "tg_tokens": null,
   "tps": 48.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-lktojoqv-ctx8192",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.9 GB at this context; duplicate traces of the same session: p1nyil1j; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 475.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 8k 475.1 46.9 14.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lktojoqv",
   "source_url": "https://omlx.ai/benchmarks/performance/lktojoqv",
   "tg_tokens": null,
   "tps": 46.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-lktojoqv-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 15.2 GB at this context; duplicate traces of the same session: p1nyil1j; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 467.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 16k 467.0 42.4 15.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lktojoqv",
   "source_url": "https://omlx.ai/benchmarks/performance/lktojoqv",
   "tg_tokens": null,
   "tps": 42.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2",
   "id": "omlx-vmtschc6-ctx4096",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 3.0 GB at this context; duplicate traces of the same session: kn7btslm; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1260.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e2b-it-4bit on M2 (10c); M2 (10c) \u00b7 8 GB \u00b7 4bit; 4k 1,260 50.9 3.0 GB view",
   "ram_gb": 8,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vmtschc6",
   "source_url": "https://omlx.ai/benchmarks/performance/vmtschc6",
   "tg_tokens": null,
   "tps": 50.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2",
   "id": "omlx-vmtschc6-ctx16384",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 3.1 GB at this context; duplicate traces of the same session: kn7btslm; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1295.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e2b-it-4bit on M2 (10c); M2 (10c) \u00b7 8 GB \u00b7 4bit; 16k 1,295 43.6 3.1 GB current",
   "ram_gb": 8,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vmtschc6",
   "source_url": "https://omlx.ai/benchmarks/performance/vmtschc6",
   "tg_tokens": null,
   "tps": 43.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-9gk9frpb-ctx1024",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.3 GB at this context; duplicate traces of the same session: eld857iv, erjdwupp, j38y1sz5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-27B-8bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 8bit; 1k 104.5 8.7 28.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9gk9frpb",
   "source_url": "https://omlx.ai/benchmarks/performance/9gk9frpb",
   "tg_tokens": null,
   "tps": 8.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-9gk9frpb-ctx4096",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: eld857iv, erjdwupp, j38y1sz5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-27B-8bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 8bit; 4k 106.0 8.6 29.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9gk9frpb",
   "source_url": "https://omlx.ai/benchmarks/performance/9gk9frpb",
   "tg_tokens": null,
   "tps": 8.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-9gk9frpb-ctx16384",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace; peak memory 30.7 GB at this context; duplicate traces of the same session: eld857iv, erjdwupp, j38y1sz5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 96.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-27B-8bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 8bit; 16k 96.5 8.3 30.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9gk9frpb",
   "source_url": "https://omlx.ai/benchmarks/performance/9gk9frpb",
   "tg_tokens": null,
   "tps": 8.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-9gk9frpb-ctx32768",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: eld857iv, erjdwupp, j38y1sz5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 93.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-27B-8bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 8bit; 32k 93.6 7.8 32.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9gk9frpb",
   "source_url": "https://omlx.ai/benchmarks/performance/9gk9frpb",
   "tg_tokens": null,
   "tps": 7.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-o5wboito-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: lql4g3fj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 737.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 737.0 108.3 21.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o5wboito",
   "source_url": "https://omlx.ai/benchmarks/performance/o5wboito",
   "tg_tokens": null,
   "tps": 108.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-o5wboito-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: lql4g3fj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 805.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 805.1 87.1 22.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o5wboito",
   "source_url": "https://omlx.ai/benchmarks/performance/o5wboito",
   "tg_tokens": null,
   "tps": 87.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-fsnai9u9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.6 GB; TTFT 10076 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 101.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k tokens 101.6 PP tok/s 16.3 TG tok/s 10076 TTFT (ms) 19.6 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fsnai9u9",
   "source_url": "https://omlx.ai/benchmarks/performance/fsnai9u9",
   "tg_tokens": null,
   "tps": 16.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-5sddcwb8-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: iiem1ryu; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 382.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 382.1 16.6 17.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5sddcwb8",
   "source_url": "https://omlx.ai/benchmarks/performance/5sddcwb8",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-5sddcwb8-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: iiem1ryu; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 373.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 373.7 15.7 17.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5sddcwb8",
   "source_url": "https://omlx.ai/benchmarks/performance/5sddcwb8",
   "tg_tokens": null,
   "tps": 15.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-9ilvvnm5-ctx32768",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace, single run at 32k context; peak memory 16.0 GB; TTFT 77438 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 423.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 32k tokens 423.2 PP tok/s 34.6 TG tok/s 77438 TTFT (ms) 16.0 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ilvvnm5",
   "source_url": "https://omlx.ai/benchmarks/performance/9ilvvnm5",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-trjavvjo-ctx32768",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 16.2 GB; TTFT 76708 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 427.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 32k tokens 427.2 PP tok/s 32.7 TG tok/s 76708 TTFT (ms) 16.2 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/trjavvjo",
   "source_url": "https://omlx.ai/benchmarks/performance/trjavvjo",
   "tg_tokens": null,
   "tps": 32.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-4043sjxp-ctx8192",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 15.0 GB at this context; duplicate traces of the same session: zipwbasd; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 479.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 8k 479.9 43.0 15.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4043sjxp",
   "source_url": "https://omlx.ai/benchmarks/performance/4043sjxp",
   "tg_tokens": null,
   "tps": 43.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-4043sjxp-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 15.4 GB at this context; duplicate traces of the same session: zipwbasd; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 463.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 16k 463.9 39.5 15.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4043sjxp",
   "source_url": "https://omlx.ai/benchmarks/performance/4043sjxp",
   "tg_tokens": null,
   "tps": 39.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-4xt0oyfg-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 926t3nve, pku5crbg, rfvgie0g; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 954.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 954.5 100.1 19.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4xt0oyfg",
   "source_url": "https://omlx.ai/benchmarks/performance/4xt0oyfg",
   "tg_tokens": null,
   "tps": 100.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-4xt0oyfg-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 926t3nve, pku5crbg, rfvgie0g; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 949.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 949.6 87.8 21.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4xt0oyfg",
   "source_url": "https://omlx.ai/benchmarks/performance/4xt0oyfg",
   "tg_tokens": null,
   "tps": 87.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-4xt0oyfg-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 926t3nve, pku5crbg, rfvgie0g; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 824.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 824.7 81.3 22.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4xt0oyfg",
   "source_url": "https://omlx.ai/benchmarks/performance/4xt0oyfg",
   "tg_tokens": null,
   "tps": 81.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-4xt0oyfg-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 926t3nve, pku5crbg, rfvgie0g; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 639.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 32k 639.1 71.7 23.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4xt0oyfg",
   "source_url": "https://omlx.ai/benchmarks/performance/4xt0oyfg",
   "tg_tokens": null,
   "tps": 71.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-7r8cid9s-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 12.3 GB at this context; duplicate traces of the same session: juzcekjl; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 482.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 8k 482.4 21.0 12.3 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7r8cid9s",
   "source_url": "https://omlx.ai/benchmarks/performance/7r8cid9s",
   "tg_tokens": null,
   "tps": 21.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-7r8cid9s-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: juzcekjl; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 454.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 32k 454.2 20.4 12.7 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7r8cid9s",
   "source_url": "https://omlx.ai/benchmarks/performance/7r8cid9s",
   "tg_tokens": null,
   "tps": 20.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-s5j7it3p-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MLX; oMLX community benchmark trace, single run at 1k context; peak memory 75.4 GB; TTFT 1574 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 650.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 650.4 PP tok/s 41.4 TG tok/s 1574 TTFT (ms) 75.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s5j7it3p",
   "source_url": "https://omlx.ai/benchmarks/performance/s5j7it3p",
   "tg_tokens": null,
   "tps": 41.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-utu6mckb-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 4wdabpps, 513kkkog, m5ziqmad; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 957.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 957.8 99.7 19.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/utu6mckb",
   "source_url": "https://omlx.ai/benchmarks/performance/utu6mckb",
   "tg_tokens": null,
   "tps": 99.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-utu6mckb-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: 4wdabpps, 513kkkog, m5ziqmad; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 941.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 941.8 89.3 21.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/utu6mckb",
   "source_url": "https://omlx.ai/benchmarks/performance/utu6mckb",
   "tg_tokens": null,
   "tps": 89.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-utu6mckb-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: 4wdabpps, 513kkkog, m5ziqmad; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 810.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 810.4 82.1 22.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/utu6mckb",
   "source_url": "https://omlx.ai/benchmarks/performance/utu6mckb",
   "tg_tokens": null,
   "tps": 82.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-utu6mckb-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 4wdabpps, 513kkkog, m5ziqmad; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 630.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 32k 630.6 72.5 23.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/utu6mckb",
   "source_url": "https://omlx.ai/benchmarks/performance/utu6mckb",
   "tg_tokens": null,
   "tps": 72.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ou42mqig-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; Jundot build; oMLX community benchmark trace, single run at 128k context; peak memory 53.3 GB; TTFT 415272 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 315.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Jundot--Qwen3.8-27B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 128k tokens 315.6 PP tok/s 17.6 TG tok/s 415272 TTFT (ms) 53.3 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ou42mqig",
   "source_url": "https://omlx.ai/benchmarks/performance/ou42mqig",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-tw6h3lb8-ctx4096",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: y5ngtudo; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1876.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 1,876 145.2 19.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tw6h3lb8",
   "source_url": "https://omlx.ai/benchmarks/performance/tw6h3lb8",
   "tg_tokens": null,
   "tps": 145.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-tw6h3lb8-ctx16384",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: y5ngtudo; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1953.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 1,953 135.1 19.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tw6h3lb8",
   "source_url": "https://omlx.ai/benchmarks/performance/tw6h3lb8",
   "tg_tokens": null,
   "tps": 135.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-s51mrhne-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "Alis build; oMLX community benchmark trace, single run at 1k context; peak memory 173.4 GB; TTFT 2208 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 463.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-Alis-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k tokens 463.8 PP tok/s 30.9 TG tok/s 2208 TTFT (ms) 173.4 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s51mrhne",
   "source_url": "https://omlx.ai/benchmarks/performance/s51mrhne",
   "tg_tokens": null,
   "tps": 30.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-l5n4h0pa-ctx1024",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 2.5 GB at this context; duplicate traces of the same session: agwz8r69; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1848.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM5-2B-OptiQ-4bit on M5 (8c); M5 (8c) \u00b7 16 GB \u00b7 4bit; 1k 1,848 70.3 2.5 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l5n4h0pa",
   "source_url": "https://omlx.ai/benchmarks/performance/l5n4h0pa",
   "tg_tokens": null,
   "tps": 70.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-l5n4h0pa-ctx4096",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 2.6 GB at this context; duplicate traces of the same session: agwz8r69; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1854.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM5-2B-OptiQ-4bit on M5 (8c); M5 (8c) \u00b7 16 GB \u00b7 4bit; 4k 1,854 63.6 2.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l5n4h0pa",
   "source_url": "https://omlx.ai/benchmarks/performance/l5n4h0pa",
   "tg_tokens": null,
   "tps": 63.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-08ojni69-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: ek5k9k83; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 580.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 580.6 63.5 20.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/08ojni69",
   "source_url": "https://omlx.ai/benchmarks/performance/08ojni69",
   "tg_tokens": null,
   "tps": 63.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-08ojni69-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: ek5k9k83; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 593.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 593.1 60.2 22.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/08ojni69",
   "source_url": "https://omlx.ai/benchmarks/performance/08ojni69",
   "tg_tokens": null,
   "tps": 60.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-rggxhk1d-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 10.2 GB at this context; duplicate traces of the same session: k7ekw7n8, u2mt9h1f; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 496.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 8k 496.6 17.4 10.2 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rggxhk1d",
   "source_url": "https://omlx.ai/benchmarks/performance/rggxhk1d",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-rggxhk1d-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 10.6 GB at this context; duplicate traces of the same session: k7ekw7n8, u2mt9h1f; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 454.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 32k 454.2 16.2 10.6 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rggxhk1d",
   "source_url": "https://omlx.ai/benchmarks/performance/rggxhk1d",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-rggxhk1d-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 11.2 GB at this context; duplicate traces of the same session: k7ekw7n8, u2mt9h1f; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 414.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 64k 414.0 15.2 11.2 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rggxhk1d",
   "source_url": "https://omlx.ai/benchmarks/performance/rggxhk1d",
   "tg_tokens": null,
   "tps": 15.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mq3tk2ki-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 16.9 GB; TTFT 1061 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 964.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 964.8 PP tok/s 47.9 TG tok/s 1061 TTFT (ms) 16.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mq3tk2ki",
   "source_url": "https://omlx.ai/benchmarks/performance/mq3tk2ki",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8ac9gql2-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: 3bxsszvx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 864.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 864.0 24.0 29.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8ac9gql2",
   "source_url": "https://omlx.ai/benchmarks/performance/8ac9gql2",
   "tg_tokens": null,
   "tps": 24.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8ac9gql2-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 31.0 GB at this context; duplicate traces of the same session: 3bxsszvx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 789.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 789.0 23.6 31.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8ac9gql2",
   "source_url": "https://omlx.ai/benchmarks/performance/8ac9gql2",
   "tg_tokens": null,
   "tps": 23.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ehxlk9sa-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: yttmbxpb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 867.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 867.8 33.7 29.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ehxlk9sa",
   "source_url": "https://omlx.ai/benchmarks/performance/ehxlk9sa",
   "tg_tokens": null,
   "tps": 33.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ehxlk9sa-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 31.1 GB at this context; duplicate traces of the same session: yttmbxpb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 853.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 853.2 26.3 31.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ehxlk9sa",
   "source_url": "https://omlx.ai/benchmarks/performance/ehxlk9sa",
   "tg_tokens": null,
   "tps": 26.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-b6tyrzra-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: u2r8pfez; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 870.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 870.2 74.1 17.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b6tyrzra",
   "source_url": "https://omlx.ai/benchmarks/performance/b6tyrzra",
   "tg_tokens": null,
   "tps": 74.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-b6tyrzra-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: u2r8pfez; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 847.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 847.5 63.5 18.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b6tyrzra",
   "source_url": "https://omlx.ai/benchmarks/performance/b6tyrzra",
   "tg_tokens": null,
   "tps": 63.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-vlz5u846-ctx1024",
   "model": "Qwen3.8-4B-Distill",
   "notes": "oMLX community benchmark trace; peak memory 3.5 GB at this context; duplicate traces of the same session: s3004vl5; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1040.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-4B-Distill-MLX-4bit on M5 (8c); M5 (8c) \u00b7 16 GB \u00b7 4bit; 1k 1,040 46.9 3.5 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vlz5u846",
   "source_url": "https://omlx.ai/benchmarks/performance/vlz5u846",
   "tg_tokens": null,
   "tps": 46.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-vlz5u846-ctx4096",
   "model": "Qwen3.8-4B-Distill",
   "notes": "oMLX community benchmark trace; peak memory 3.8 GB at this context; duplicate traces of the same session: s3004vl5; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1011.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-4B-Distill-MLX-4bit on M5 (8c); M5 (8c) \u00b7 16 GB \u00b7 4bit; 4k 1,011 44.8 3.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vlz5u846",
   "source_url": "https://omlx.ai/benchmarks/performance/vlz5u846",
   "tg_tokens": null,
   "tps": 44.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-41poigvk-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP320",
   "notes": "MTP Vision build; oMLX community benchmark trace; peak memory 66.5 GB at this context; duplicate traces of the same session: cy2kn6oj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 503.2,
   "provenance": "community",
   "quant": "oQ3e-DWQ",
   "quote": "Qwen3.8-Flash-Next-REAP320-oQ3e-DWQ-MTP-Vision-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 1k 503.2 75.6 66.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/41poigvk",
   "source_url": "https://omlx.ai/benchmarks/performance/41poigvk",
   "tg_tokens": null,
   "tps": 75.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-41poigvk-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP320",
   "notes": "MTP Vision build; oMLX community benchmark trace; peak memory 67.1 GB at this context; duplicate traces of the same session: cy2kn6oj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1162.0,
   "provenance": "community",
   "quant": "oQ3e-DWQ",
   "quote": "Qwen3.8-Flash-Next-REAP320-oQ3e-DWQ-MTP-Vision-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 4k 1,162 55.3 67.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/41poigvk",
   "source_url": "https://omlx.ai/benchmarks/performance/41poigvk",
   "tg_tokens": null,
   "tps": 55.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-e956m9pb-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build (TF); oMLX community benchmark trace, single run at 1k context; peak memory 75.4 GB; TTFT 1107 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 924.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-Uncensored-TF-4bit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 924.9 PP tok/s 40.2 TG tok/s 1107 TTFT (ms) 75.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e956m9pb",
   "source_url": "https://omlx.ai/benchmarks/performance/e956m9pb",
   "tg_tokens": null,
   "tps": 40.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-7sjxzi12-ctx1024",
   "model": "clef-flash",
   "notes": "generic trace name; oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: 7kgxyo2a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2471.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "clef-flash-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,471 100.6 5.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7sjxzi12",
   "source_url": "https://omlx.ai/benchmarks/performance/7sjxzi12",
   "tg_tokens": null,
   "tps": 100.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-7sjxzi12-ctx4096",
   "model": "clef-flash",
   "notes": "generic trace name; oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: 7kgxyo2a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2816.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "clef-flash-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,816 99.4 6.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7sjxzi12",
   "source_url": "https://omlx.ai/benchmarks/performance/7sjxzi12",
   "tg_tokens": null,
   "tps": 99.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-q4totk5c-ctx4096",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: nllpi8vn, vyoeuav2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2124.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 2,124 135.6 21.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q4totk5c",
   "source_url": "https://omlx.ai/benchmarks/performance/q4totk5c",
   "tg_tokens": null,
   "tps": 135.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-q4totk5c-ctx16384",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: nllpi8vn, vyoeuav2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4065.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 4,065 127.1 21.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q4totk5c",
   "source_url": "https://omlx.ai/benchmarks/performance/q4totk5c",
   "tg_tokens": null,
   "tps": 127.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-q4totk5c-ctx131072",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: nllpi8vn, vyoeuav2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2332.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 2,332 84.6 21.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q4totk5c",
   "source_url": "https://omlx.ai/benchmarks/performance/q4totk5c",
   "tg_tokens": null,
   "tps": 84.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-t09gwryr-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: xyf5tw42; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1117.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k 1,117 48.7 19.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t09gwryr",
   "source_url": "https://omlx.ai/benchmarks/performance/t09gwryr",
   "tg_tokens": null,
   "tps": 48.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-t09gwryr-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: xyf5tw42; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1721.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 1,721 47.3 20.5 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t09gwryr",
   "source_url": "https://omlx.ai/benchmarks/performance/t09gwryr",
   "tg_tokens": null,
   "tps": 47.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-ojadpj83-ctx1024",
   "model": "Qwen3.5-35B-A3B-Uncensored",
   "notes": "FernflowerAI build; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: q666778r; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1365.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B-Uncensored-FernflowerAI-MLX-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k 1,365 59.3 19.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ojadpj83",
   "source_url": "https://omlx.ai/benchmarks/performance/ojadpj83",
   "tg_tokens": null,
   "tps": 59.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-ojadpj83-ctx4096",
   "model": "Qwen3.5-35B-A3B-Uncensored",
   "notes": "FernflowerAI build; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: q666778r; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1689.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B-Uncensored-FernflowerAI-MLX-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 1,689 58.1 19.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ojadpj83",
   "source_url": "https://omlx.ai/benchmarks/performance/ojadpj83",
   "tg_tokens": null,
   "tps": 58.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-21skc6ml-ctx4096",
   "model": "Qwen3.5-35B-A3B-Uncensored",
   "notes": "FernflowerAI build; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: n05fmpez; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1719.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B-Uncensored-FernflowerAI-MLX-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 1,719 57.9 19.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/21skc6ml",
   "source_url": "https://omlx.ai/benchmarks/performance/21skc6ml",
   "tg_tokens": null,
   "tps": 57.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-21skc6ml-ctx16384",
   "model": "Qwen3.5-35B-A3B-Uncensored",
   "notes": "FernflowerAI build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: n05fmpez; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1554.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B-Uncensored-FernflowerAI-MLX-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 16k 1,554 50.2 20.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/21skc6ml",
   "source_url": "https://omlx.ai/benchmarks/performance/21skc6ml",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-oo5p089v-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: h4757055; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1322.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k 1,322 40.6 20.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oo5p089v",
   "source_url": "https://omlx.ai/benchmarks/performance/oo5p089v",
   "tg_tokens": null,
   "tps": 40.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-oo5p089v-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: h4757055; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1619.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 1,619 40.1 21.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oo5p089v",
   "source_url": "https://omlx.ai/benchmarks/performance/oo5p089v",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-y3wbxb04-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: d8sq19q8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 723.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 723.2 75.7 20.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y3wbxb04",
   "source_url": "https://omlx.ai/benchmarks/performance/y3wbxb04",
   "tg_tokens": null,
   "tps": 75.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-y3wbxb04-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: d8sq19q8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1877.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,877 73.4 21.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y3wbxb04",
   "source_url": "https://omlx.ai/benchmarks/performance/y3wbxb04",
   "tg_tokens": null,
   "tps": 73.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-a5xgiqqn-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: u248f5ur; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5788.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 5,788 215.4 22.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a5xgiqqn",
   "source_url": "https://omlx.ai/benchmarks/performance/a5xgiqqn",
   "tg_tokens": null,
   "tps": 215.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-a5xgiqqn-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: u248f5ur; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5060.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 5,060 173.1 22.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a5xgiqqn",
   "source_url": "https://omlx.ai/benchmarks/performance/a5xgiqqn",
   "tg_tokens": null,
   "tps": 173.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-v6umnsbd-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: jbem3u47, rjzy4ozm; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 95.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 95.7 18.6 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v6umnsbd",
   "source_url": "https://omlx.ai/benchmarks/performance/v6umnsbd",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-v6umnsbd-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: jbem3u47, rjzy4ozm; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 89.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 89.9 17.3 21.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v6umnsbd",
   "source_url": "https://omlx.ai/benchmarks/performance/v6umnsbd",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-v6umnsbd-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.7 GB at this context; duplicate traces of the same session: jbem3u47, rjzy4ozm; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 53.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 128k 53.6 10.3 28.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v6umnsbd",
   "source_url": "https://omlx.ai/benchmarks/performance/v6umnsbd",
   "tg_tokens": null,
   "tps": 10.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ls2qn8jw-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: ul6z60v8, xaoigq37; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1420.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 4k 1,420 32.9 29.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ls2qn8jw",
   "source_url": "https://omlx.ai/benchmarks/performance/ls2qn8jw",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ls2qn8jw-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.3 GB at this context; duplicate traces of the same session: ul6z60v8, xaoigq37; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1310.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 32k 1,310 30.0 31.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ls2qn8jw",
   "source_url": "https://omlx.ai/benchmarks/performance/ls2qn8jw",
   "tg_tokens": null,
   "tps": 30.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ls2qn8jw-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.3 GB at this context; duplicate traces of the same session: ul6z60v8, xaoigq37; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1152.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 64k 1,152 27.8 33.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ls2qn8jw",
   "source_url": "https://omlx.ai/benchmarks/performance/ls2qn8jw",
   "tg_tokens": null,
   "tps": 27.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-iix2z792-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 28.3 GB at this context; duplicate traces of the same session: otffabik; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 413.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 413.3 9.7 28.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iix2z792",
   "source_url": "https://omlx.ai/benchmarks/performance/iix2z792",
   "tg_tokens": null,
   "tps": 9.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-iix2z792-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: otffabik; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 470.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 470.8 9.7 29.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iix2z792",
   "source_url": "https://omlx.ai/benchmarks/performance/iix2z792",
   "tg_tokens": null,
   "tps": 9.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qmg5emp9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: l2wm2ys7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 986.9,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 986.9 60.5 18.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qmg5emp9",
   "source_url": "https://omlx.ai/benchmarks/performance/qmg5emp9",
   "tg_tokens": null,
   "tps": 60.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qmg5emp9-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: l2wm2ys7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 886.6,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 886.6 55.8 19.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qmg5emp9",
   "source_url": "https://omlx.ai/benchmarks/performance/qmg5emp9",
   "tg_tokens": null,
   "tps": 55.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qmg5emp9-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: l2wm2ys7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 817.8,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 817.8 41.1 21.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qmg5emp9",
   "source_url": "https://omlx.ai/benchmarks/performance/qmg5emp9",
   "tg_tokens": null,
   "tps": 41.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0bdc21it-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 0f1h0g7m, 69nx2tdt, cv2d16vr, g0yydxi1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1014.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,014 68.3 17.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0bdc21it",
   "source_url": "https://omlx.ai/benchmarks/performance/0bdc21it",
   "tg_tokens": null,
   "tps": 68.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0bdc21it-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 0f1h0g7m, 69nx2tdt, cv2d16vr, g0yydxi1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1361.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,361 63.3 19.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0bdc21it",
   "source_url": "https://omlx.ai/benchmarks/performance/0bdc21it",
   "tg_tokens": null,
   "tps": 63.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0bdc21it-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 0f1h0g7m, 69nx2tdt, cv2d16vr, g0yydxi1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1338.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,338 66.3 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0bdc21it",
   "source_url": "https://omlx.ai/benchmarks/performance/0bdc21it",
   "tg_tokens": null,
   "tps": 66.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0bdc21it-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 0f1h0g7m, 69nx2tdt, cv2d16vr, g0yydxi1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1259.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,259 74.6 19.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0bdc21it",
   "source_url": "https://omlx.ai/benchmarks/performance/0bdc21it",
   "tg_tokens": null,
   "tps": 74.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0bdc21it-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 0f1h0g7m, 69nx2tdt, cv2d16vr, g0yydxi1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1080.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,080 73.2 21.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0bdc21it",
   "source_url": "https://omlx.ai/benchmarks/performance/0bdc21it",
   "tg_tokens": null,
   "tps": 73.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-sb5m37x6-ctx32768",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 73.5 GB; TTFT 19879 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1648.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k tokens 1,648 PP tok/s 78.7 TG tok/s 19879 TTFT (ms) 73.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sb5m37x6",
   "source_url": "https://omlx.ai/benchmarks/performance/sb5m37x6",
   "tg_tokens": null,
   "tps": 78.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-6sgfx588-ctx16384",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 73.0 GB; TTFT 10006 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1637.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k tokens 1,637 PP tok/s 84.7 TG tok/s 10006 TTFT (ms) 73.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6sgfx588",
   "source_url": "https://omlx.ai/benchmarks/performance/6sgfx588",
   "tg_tokens": null,
   "tps": 84.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-h9zdft4a-ctx16384",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: 9a2zwwfm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1695.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,695 76.0 72.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h9zdft4a",
   "source_url": "https://omlx.ai/benchmarks/performance/h9zdft4a",
   "tg_tokens": null,
   "tps": 76.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-h9zdft4a-ctx32768",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: 9a2zwwfm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1607.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,607 86.1 73.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h9zdft4a",
   "source_url": "https://omlx.ai/benchmarks/performance/h9zdft4a",
   "tg_tokens": null,
   "tps": 86.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-vnz27eud-ctx16384",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 73.2 GB; TTFT 10067 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1628.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k tokens 1,628 PP tok/s 87.8 TG tok/s 10067 TTFT (ms) 73.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vnz27eud",
   "source_url": "https://omlx.ai/benchmarks/performance/vnz27eud",
   "tg_tokens": null,
   "tps": 87.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-slqco3nm-ctx16384",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 73.2 GB; TTFT 9994 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1639.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k tokens 1,639 PP tok/s 76.1 TG tok/s 9994 TTFT (ms) 73.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/slqco3nm",
   "source_url": "https://omlx.ai/benchmarks/performance/slqco3nm",
   "tg_tokens": null,
   "tps": 76.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-rql69pqi-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: 138asxxv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 670.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 670.3 51.6 71.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rql69pqi",
   "source_url": "https://omlx.ai/benchmarks/performance/rql69pqi",
   "tg_tokens": null,
   "tps": 51.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-rql69pqi-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.0 GB at this context; duplicate traces of the same session: 138asxxv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 758.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 758.7 66.4 72.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rql69pqi",
   "source_url": "https://omlx.ai/benchmarks/performance/rql69pqi",
   "tg_tokens": null,
   "tps": 66.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-v02jcaoa-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: yr004zpl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 496.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 496.4 45.8 70.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v02jcaoa",
   "source_url": "https://omlx.ai/benchmarks/performance/v02jcaoa",
   "tg_tokens": null,
   "tps": 45.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-v02jcaoa-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: yr004zpl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 588.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 588.8 71.6 71.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v02jcaoa",
   "source_url": "https://omlx.ai/benchmarks/performance/v02jcaoa",
   "tg_tokens": null,
   "tps": 71.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-muw7gc0r-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: 69wkpatj, ul0dssxr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 377.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 377.4 46.3 71.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/muw7gc0r",
   "source_url": "https://omlx.ai/benchmarks/performance/muw7gc0r",
   "tg_tokens": null,
   "tps": 46.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-muw7gc0r-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 72.0 GB at this context; duplicate traces of the same session: 69wkpatj, ul0dssxr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 442.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 442.7 51.4 72.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/muw7gc0r",
   "source_url": "https://omlx.ai/benchmarks/performance/muw7gc0r",
   "tg_tokens": null,
   "tps": 51.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-muw7gc0r-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 73.1 GB at this context; duplicate traces of the same session: 69wkpatj, ul0dssxr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 492.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 492.0 47.7 73.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/muw7gc0r",
   "source_url": "https://omlx.ai/benchmarks/performance/muw7gc0r",
   "tg_tokens": null,
   "tps": 47.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-8jo2z1p5-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: sm32dvvj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 489.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 489.4 49.6 70.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8jo2z1p5",
   "source_url": "https://omlx.ai/benchmarks/performance/8jo2z1p5",
   "tg_tokens": null,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-8jo2z1p5-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: sm32dvvj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 599.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 599.3 69.2 71.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8jo2z1p5",
   "source_url": "https://omlx.ai/benchmarks/performance/8jo2z1p5",
   "tg_tokens": null,
   "tps": 69.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-pf3fzi6a-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build (mlx); oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 9fj0oy4b; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1148.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ornith-1.5-35b-a3b-mtp-mlx-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 1k 1,148 115.8 21.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pf3fzi6a",
   "source_url": "https://omlx.ai/benchmarks/performance/pf3fzi6a",
   "tg_tokens": null,
   "tps": 115.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-pf3fzi6a-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build (mlx); oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 9fj0oy4b; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1259.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ornith-1.5-35b-a3b-mtp-mlx-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 4k 1,259 114.5 21.8 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pf3fzi6a",
   "source_url": "https://omlx.ai/benchmarks/performance/pf3fzi6a",
   "tg_tokens": null,
   "tps": 114.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-wu3psapl-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; Jundot build; oMLX community benchmark trace, single run at 128k context; peak memory 56.5 GB; TTFT 399075 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 328.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Jundot--Qwen3.8-27B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 128k tokens 328.4 PP tok/s 14.6 TG tok/s 399075 TTFT (ms) 56.5 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wu3psapl",
   "source_url": "https://omlx.ai/benchmarks/performance/wu3psapl",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-mfg06zo5-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; Jundot build; oMLX community benchmark trace, single run at 128k context; peak memory 43.8 GB; TTFT 485556 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 269.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Jundot--Qwen3.8-27B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 128k tokens 269.9 PP tok/s 2.5 TG tok/s 485556 TTFT (ms) 43.8 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mfg06zo5",
   "source_url": "https://omlx.ai/benchmarks/performance/mfg06zo5",
   "tg_tokens": null,
   "tps": 2.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-rznqi0mj-ctx1024",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: oy4lezar; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3556.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 3,556 164.4 18.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rznqi0mj",
   "source_url": "https://omlx.ai/benchmarks/performance/rznqi0mj",
   "tg_tokens": null,
   "tps": 164.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-rznqi0mj-ctx4096",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: oy4lezar; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4070.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 4,070 158.5 19.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rznqi0mj",
   "source_url": "https://omlx.ai/benchmarks/performance/rznqi0mj",
   "tg_tokens": null,
   "tps": 158.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-67nvda9y-ctx1024",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: 8vjaj4yt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3427.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 3,427 131.3 22.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/67nvda9y",
   "source_url": "https://omlx.ai/benchmarks/performance/67nvda9y",
   "tg_tokens": null,
   "tps": 131.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-67nvda9y-ctx4096",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: 8vjaj4yt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4015.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 4,015 129.4 23.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/67nvda9y",
   "source_url": "https://omlx.ai/benchmarks/performance/67nvda9y",
   "tg_tokens": null,
   "tps": 129.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-uohe7m0x-ctx1024",
   "model": "Nemotron-Cascade-2-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: 5r5llf83, atiq38ia, go9qcvhl, mfl4gqfi, tspl9jui; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 479.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Nemotron-Cascade-2-30B-A3B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 479.7 66.6 18.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uohe7m0x",
   "source_url": "https://omlx.ai/benchmarks/performance/uohe7m0x",
   "tg_tokens": null,
   "tps": 66.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-uohe7m0x-ctx4096",
   "model": "Nemotron-Cascade-2-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 5r5llf83, atiq38ia, go9qcvhl, mfl4gqfi, tspl9jui; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 542.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Nemotron-Cascade-2-30B-A3B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 542.5 65.0 19.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uohe7m0x",
   "source_url": "https://omlx.ai/benchmarks/performance/uohe7m0x",
   "tg_tokens": null,
   "tps": 65.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-uohe7m0x-ctx8192",
   "model": "Nemotron-Cascade-2-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 5r5llf83, atiq38ia, go9qcvhl, mfl4gqfi, tspl9jui; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 533.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Nemotron-Cascade-2-30B-A3B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 8k 533.5 62.5 19.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uohe7m0x",
   "source_url": "https://omlx.ai/benchmarks/performance/uohe7m0x",
   "tg_tokens": null,
   "tps": 62.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-uohe7m0x-ctx16384",
   "model": "Nemotron-Cascade-2-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 5r5llf83, atiq38ia, go9qcvhl, mfl4gqfi, tspl9jui; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 506.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Nemotron-Cascade-2-30B-A3B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k 506.8 59.1 19.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uohe7m0x",
   "source_url": "https://omlx.ai/benchmarks/performance/uohe7m0x",
   "tg_tokens": null,
   "tps": 59.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-uohe7m0x-ctx32768",
   "model": "Nemotron-Cascade-2-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 5r5llf83, atiq38ia, go9qcvhl, mfl4gqfi, tspl9jui; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 461.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Nemotron-Cascade-2-30B-A3B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 32k 461.7 51.9 20.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uohe7m0x",
   "source_url": "https://omlx.ai/benchmarks/performance/uohe7m0x",
   "tg_tokens": null,
   "tps": 51.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-uohe7m0x-ctx65536",
   "model": "Nemotron-Cascade-2-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 5r5llf83, atiq38ia, go9qcvhl, mfl4gqfi, tspl9jui; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 392.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Nemotron-Cascade-2-30B-A3B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 64k 392.0 39.9 20.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uohe7m0x",
   "source_url": "https://omlx.ai/benchmarks/performance/uohe7m0x",
   "tg_tokens": null,
   "tps": 39.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wwq9teiq-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 30.9 GB at this context; duplicate traces of the same session: mmv1bmxu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 923.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 923.9 71.8 30.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wwq9teiq",
   "source_url": "https://omlx.ai/benchmarks/performance/wwq9teiq",
   "tg_tokens": null,
   "tps": 71.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wwq9teiq-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 32.3 GB at this context; duplicate traces of the same session: mmv1bmxu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 746.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 746.1 16.5 32.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wwq9teiq",
   "source_url": "https://omlx.ai/benchmarks/performance/wwq9teiq",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-j72p0hzv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: c6dxl0gv, yc0yln1c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 986.3,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 986.3 61.4 18.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j72p0hzv",
   "source_url": "https://omlx.ai/benchmarks/performance/j72p0hzv",
   "tg_tokens": null,
   "tps": 61.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-j72p0hzv-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: c6dxl0gv, yc0yln1c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 876.5,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 876.5 67.9 19.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j72p0hzv",
   "source_url": "https://omlx.ai/benchmarks/performance/j72p0hzv",
   "tg_tokens": null,
   "tps": 67.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-j72p0hzv-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: c6dxl0gv, yc0yln1c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 791.3,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 791.3 12.6 21.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j72p0hzv",
   "source_url": "https://omlx.ai/benchmarks/performance/j72p0hzv",
   "tg_tokens": null,
   "tps": 12.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dy6y78sr-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 986.9,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 986.9 60.5 18.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dy6y78sr",
   "source_url": "https://omlx.ai/benchmarks/performance/dy6y78sr",
   "tg_tokens": null,
   "tps": 60.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dy6y78sr-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 886.6,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 886.6 55.8 19.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dy6y78sr",
   "source_url": "https://omlx.ai/benchmarks/performance/dy6y78sr",
   "tg_tokens": null,
   "tps": 55.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dy6y78sr-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 817.8,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 817.8 41.1 21.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dy6y78sr",
   "source_url": "https://omlx.ai/benchmarks/performance/dy6y78sr",
   "tg_tokens": null,
   "tps": 41.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hk40dtz5-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 0r9ao36r, 7w62y3hp, hpz9ygp8, xmi83nb6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 955.6,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 955.6 49.7 17.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hk40dtz5",
   "source_url": "https://omlx.ai/benchmarks/performance/hk40dtz5",
   "tg_tokens": null,
   "tps": 49.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hk40dtz5-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 0r9ao36r, 7w62y3hp, hpz9ygp8, xmi83nb6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 929.3,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 929.3 55.9 19.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hk40dtz5",
   "source_url": "https://omlx.ai/benchmarks/performance/hk40dtz5",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hk40dtz5-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 0r9ao36r, 7w62y3hp, hpz9ygp8, xmi83nb6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 881.9,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 881.9 59.8 20.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hk40dtz5",
   "source_url": "https://omlx.ai/benchmarks/performance/hk40dtz5",
   "tg_tokens": null,
   "tps": 59.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hk40dtz5-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 0r9ao36r, 7w62y3hp, hpz9ygp8, xmi83nb6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 783.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 783.0 43.6 21.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hk40dtz5",
   "source_url": "https://omlx.ai/benchmarks/performance/hk40dtz5",
   "tg_tokens": null,
   "tps": 43.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hk40dtz5-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 0r9ao36r, 7w62y3hp, hpz9ygp8, xmi83nb6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 708.2,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 708.2 28.0 19.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hk40dtz5",
   "source_url": "https://omlx.ai/benchmarks/performance/hk40dtz5",
   "tg_tokens": null,
   "tps": 28.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-98wmv06x-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 9kd1dbxy, hs112q6u, ov5s7kd4, pv94fkhw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 950.3,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 950.3 52.7 17.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/98wmv06x",
   "source_url": "https://omlx.ai/benchmarks/performance/98wmv06x",
   "tg_tokens": null,
   "tps": 52.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-98wmv06x-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 9kd1dbxy, hs112q6u, ov5s7kd4, pv94fkhw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 961.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 961.0 59.7 19.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/98wmv06x",
   "source_url": "https://omlx.ai/benchmarks/performance/98wmv06x",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-98wmv06x-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 9kd1dbxy, hs112q6u, ov5s7kd4, pv94fkhw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 788.3,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 788.3 31.4 17.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/98wmv06x",
   "source_url": "https://omlx.ai/benchmarks/performance/98wmv06x",
   "tg_tokens": null,
   "tps": 31.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-98wmv06x-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: 9kd1dbxy, hs112q6u, ov5s7kd4, pv94fkhw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 816.5,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 816.5 29.6 18.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/98wmv06x",
   "source_url": "https://omlx.ai/benchmarks/performance/98wmv06x",
   "tg_tokens": null,
   "tps": 29.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-98wmv06x-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 9kd1dbxy, hs112q6u, ov5s7kd4, pv94fkhw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 739.9,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 739.9 28.4 19.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/98wmv06x",
   "source_url": "https://omlx.ai/benchmarks/performance/98wmv06x",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mdokvhsi-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 5xc5jifa, 8zh0qh4h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 929.5,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 929.5 47.1 20.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mdokvhsi",
   "source_url": "https://omlx.ai/benchmarks/performance/mdokvhsi",
   "tg_tokens": null,
   "tps": 47.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mdokvhsi-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: 5xc5jifa, 8zh0qh4h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 958.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 958.0 49.1 21.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mdokvhsi",
   "source_url": "https://omlx.ai/benchmarks/performance/mdokvhsi",
   "tg_tokens": null,
   "tps": 49.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mdokvhsi-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: 5xc5jifa, 8zh0qh4h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 855.3,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 855.3 53.5 22.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mdokvhsi",
   "source_url": "https://omlx.ai/benchmarks/performance/mdokvhsi",
   "tg_tokens": null,
   "tps": 53.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mdokvhsi-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: 5xc5jifa, 8zh0qh4h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 748.8,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 748.8 41.4 24.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mdokvhsi",
   "source_url": "https://omlx.ai/benchmarks/performance/mdokvhsi",
   "tg_tokens": null,
   "tps": 41.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mdokvhsi-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: 5xc5jifa, 8zh0qh4h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 707.2,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 707.2 8.5 26.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mdokvhsi",
   "source_url": "https://omlx.ai/benchmarks/performance/mdokvhsi",
   "tg_tokens": null,
   "tps": 8.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mdokvhsi-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.3 GB at this context; duplicate traces of the same session: 5xc5jifa, 8zh0qh4h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 573.4,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 573.4 4.1 31.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mdokvhsi",
   "source_url": "https://omlx.ai/benchmarks/performance/mdokvhsi",
   "tg_tokens": null,
   "tps": 4.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mdokvhsi-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 39.6 GB at this context; duplicate traces of the same session: 5xc5jifa, 8zh0qh4h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 455.7,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 455.7 3.3 39.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mdokvhsi",
   "source_url": "https://omlx.ai/benchmarks/performance/mdokvhsi",
   "tg_tokens": null,
   "tps": 3.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-rurrswfw-ctx1024",
   "model": "Qwen3.8-35B-A3B-Distill-Heretic",
   "notes": "MTP speculative decoding enabled; fp16 build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 9vtrznya, aad5cifw, f3blm8xe, gvw954r7, u0602144, zmezm0v4; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 726.3,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-35B-A3B-Distill-Heretic-oQ4-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 726.3 73.1 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rurrswfw",
   "source_url": "https://omlx.ai/benchmarks/performance/rurrswfw",
   "tg_tokens": null,
   "tps": 73.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-rurrswfw-ctx4096",
   "model": "Qwen3.8-35B-A3B-Distill-Heretic",
   "notes": "MTP speculative decoding enabled; fp16 build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 9vtrznya, aad5cifw, f3blm8xe, gvw954r7, u0602144, zmezm0v4; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 835.8,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-35B-A3B-Distill-Heretic-oQ4-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 835.8 69.2 22.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rurrswfw",
   "source_url": "https://omlx.ai/benchmarks/performance/rurrswfw",
   "tg_tokens": null,
   "tps": 69.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-rurrswfw-ctx8192",
   "model": "Qwen3.8-35B-A3B-Distill-Heretic",
   "notes": "MTP speculative decoding enabled; fp16 build; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: 9vtrznya, aad5cifw, f3blm8xe, gvw954r7, u0602144, zmezm0v4; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 840.6,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-35B-A3B-Distill-Heretic-oQ4-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 8k 840.6 65.9 22.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rurrswfw",
   "source_url": "https://omlx.ai/benchmarks/performance/rurrswfw",
   "tg_tokens": null,
   "tps": 65.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-rurrswfw-ctx16384",
   "model": "Qwen3.8-35B-A3B-Distill-Heretic",
   "notes": "MTP speculative decoding enabled; fp16 build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 9vtrznya, aad5cifw, f3blm8xe, gvw954r7, u0602144, zmezm0v4; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 741.1,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-35B-A3B-Distill-Heretic-oQ4-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k 741.1 61.1 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rurrswfw",
   "source_url": "https://omlx.ai/benchmarks/performance/rurrswfw",
   "tg_tokens": null,
   "tps": 61.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-rurrswfw-ctx32768",
   "model": "Qwen3.8-35B-A3B-Distill-Heretic",
   "notes": "MTP speculative decoding enabled; fp16 build; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: 9vtrznya, aad5cifw, f3blm8xe, gvw954r7, u0602144, zmezm0v4; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 596.3,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-35B-A3B-Distill-Heretic-oQ4-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 32k 596.3 54.8 24.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rurrswfw",
   "source_url": "https://omlx.ai/benchmarks/performance/rurrswfw",
   "tg_tokens": null,
   "tps": 54.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-rurrswfw-ctx65536",
   "model": "Qwen3.8-35B-A3B-Distill-Heretic",
   "notes": "MTP speculative decoding enabled; fp16 build; oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: 9vtrznya, aad5cifw, f3blm8xe, gvw954r7, u0602144, zmezm0v4; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 432.8,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-35B-A3B-Distill-Heretic-oQ4-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 64k 432.8 43.9 24.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rurrswfw",
   "source_url": "https://omlx.ai/benchmarks/performance/rurrswfw",
   "tg_tokens": null,
   "tps": 43.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-rurrswfw-ctx131072",
   "model": "Qwen3.8-35B-A3B-Distill-Heretic",
   "notes": "MTP speculative decoding enabled; fp16 build; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: 9vtrznya, aad5cifw, f3blm8xe, gvw954r7, u0602144, zmezm0v4; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 280.8,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-35B-A3B-Distill-Heretic-oQ4-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 128k 280.8 33.3 26.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rurrswfw",
   "source_url": "https://omlx.ai/benchmarks/performance/rurrswfw",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-46avxx9x-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: ndxoo01q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 681.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 681.7 73.6 100.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/46avxx9x",
   "source_url": "https://omlx.ai/benchmarks/performance/46avxx9x",
   "tg_tokens": null,
   "tps": 73.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-46avxx9x-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: ndxoo01q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 818.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 818.0 65.7 101.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/46avxx9x",
   "source_url": "https://omlx.ai/benchmarks/performance/46avxx9x",
   "tg_tokens": null,
   "tps": 65.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-kjhyccgp-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.8 GB at this context; duplicate traces of the same session: 8bazs3ye; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 686.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 686.1 72.0 100.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kjhyccgp",
   "source_url": "https://omlx.ai/benchmarks/performance/kjhyccgp",
   "tg_tokens": null,
   "tps": 72.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-kjhyccgp-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: 8bazs3ye; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 818.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 818.4 32.9 101.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kjhyccgp",
   "source_url": "https://omlx.ai/benchmarks/performance/kjhyccgp",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-oud6g5zm-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 100.5 GB; TTFT 1495 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 684.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 684.9 PP tok/s 53.1 TG tok/s 1495 TTFT (ms) 100.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oud6g5zm",
   "source_url": "https://omlx.ai/benchmarks/performance/oud6g5zm",
   "tg_tokens": null,
   "tps": 53.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-y4spj8c3-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.9 GB at this context; duplicate traces of the same session: z8l5yg6i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 513.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 513.6 64.0 100.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y4spj8c3",
   "source_url": "https://omlx.ai/benchmarks/performance/y4spj8c3",
   "tg_tokens": null,
   "tps": 64.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-y4spj8c3-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: z8l5yg6i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 816.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 816.5 22.1 101.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y4spj8c3",
   "source_url": "https://omlx.ai/benchmarks/performance/y4spj8c3",
   "tg_tokens": null,
   "tps": 22.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ea4nlvi0-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: a5qqg16s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 669.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 669.2 60.2 100.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ea4nlvi0",
   "source_url": "https://omlx.ai/benchmarks/performance/ea4nlvi0",
   "tg_tokens": null,
   "tps": 60.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ea4nlvi0-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: a5qqg16s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 814.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 814.6 67.8 101.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ea4nlvi0",
   "source_url": "https://omlx.ai/benchmarks/performance/ea4nlvi0",
   "tg_tokens": null,
   "tps": 67.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-catc0d40-ctx1024",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 155.8 GB at this context; duplicate traces of the same session: 57o1trhf, 7ofwr642, av5mtyhz, d0k0j1mu, zj0fquy9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 449.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 449.1 29.1 155.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/catc0d40",
   "source_url": "https://omlx.ai/benchmarks/performance/catc0d40",
   "tg_tokens": null,
   "tps": 29.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-catc0d40-ctx4096",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.3 GB at this context; duplicate traces of the same session: 57o1trhf, 7ofwr642, av5mtyhz, d0k0j1mu, zj0fquy9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 484.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 484.2 32.1 156.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/catc0d40",
   "source_url": "https://omlx.ai/benchmarks/performance/catc0d40",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-catc0d40-ctx8192",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 164.5 GB at this context; duplicate traces of the same session: 57o1trhf, 7ofwr642, av5mtyhz, d0k0j1mu, zj0fquy9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 474.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 474.9 4.2 164.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/catc0d40",
   "source_url": "https://omlx.ai/benchmarks/performance/catc0d40",
   "tg_tokens": null,
   "tps": 4.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-catc0d40-ctx16384",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 164.6 GB at this context; duplicate traces of the same session: 57o1trhf, 7ofwr642, av5mtyhz, d0k0j1mu, zj0fquy9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 465.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 465.7 2.1 164.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/catc0d40",
   "source_url": "https://omlx.ai/benchmarks/performance/catc0d40",
   "tg_tokens": null,
   "tps": 2.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-catc0d40-ctx32768",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.5 GB at this context; duplicate traces of the same session: 57o1trhf, 7ofwr642, av5mtyhz, d0k0j1mu, zj0fquy9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 452.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 452.1 26.9 156.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/catc0d40",
   "source_url": "https://omlx.ai/benchmarks/performance/catc0d40",
   "tg_tokens": null,
   "tps": 26.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-catc0d40-ctx65536",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 165.7 GB at this context; duplicate traces of the same session: 57o1trhf, 7ofwr642, av5mtyhz, d0k0j1mu, zj0fquy9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 428.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 428.0 0.5 165.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/catc0d40",
   "source_url": "https://omlx.ai/benchmarks/performance/catc0d40",
   "tg_tokens": null,
   "tps": 0.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-iuqrvx76-ctx1024",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; Jundot build; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 2mqj6cqe, 2rgy5vhx, kktsdbkf, xmthmwg6, xp97xdfi, yi1jcsqt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 703.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot:gemma-4-31B-it-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 703.9 48.3 21.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iuqrvx76",
   "source_url": "https://omlx.ai/benchmarks/performance/iuqrvx76",
   "tg_tokens": null,
   "tps": 48.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-iuqrvx76-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; Jundot build; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 2mqj6cqe, 2rgy5vhx, kktsdbkf, xmthmwg6, xp97xdfi, yi1jcsqt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 727.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot:gemma-4-31B-it-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 727.8 45.2 21.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iuqrvx76",
   "source_url": "https://omlx.ai/benchmarks/performance/iuqrvx76",
   "tg_tokens": null,
   "tps": 45.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-iuqrvx76-ctx8192",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; Jundot build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 2mqj6cqe, 2rgy5vhx, kktsdbkf, xmthmwg6, xp97xdfi, yi1jcsqt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 665.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot:gemma-4-31B-it-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 665.4 38.3 22.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iuqrvx76",
   "source_url": "https://omlx.ai/benchmarks/performance/iuqrvx76",
   "tg_tokens": null,
   "tps": 38.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-iuqrvx76-ctx16384",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; Jundot build; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: 2mqj6cqe, 2rgy5vhx, kktsdbkf, xmthmwg6, xp97xdfi, yi1jcsqt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 601.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot:gemma-4-31B-it-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 601.3 45.8 23.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iuqrvx76",
   "source_url": "https://omlx.ai/benchmarks/performance/iuqrvx76",
   "tg_tokens": null,
   "tps": 45.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-iuqrvx76-ctx32768",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; Jundot build; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: 2mqj6cqe, 2rgy5vhx, kktsdbkf, xmthmwg6, xp97xdfi, yi1jcsqt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 564.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot:gemma-4-31B-it-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 564.9 27.6 25.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iuqrvx76",
   "source_url": "https://omlx.ai/benchmarks/performance/iuqrvx76",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-iuqrvx76-ctx65536",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; Jundot build; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: 2mqj6cqe, 2rgy5vhx, kktsdbkf, xmthmwg6, xp97xdfi, yi1jcsqt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 468.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot:gemma-4-31B-it-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 468.4 20.7 29.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iuqrvx76",
   "source_url": "https://omlx.ai/benchmarks/performance/iuqrvx76",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-iuqrvx76-ctx131072",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; Jundot build; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: 2mqj6cqe, 2rgy5vhx, kktsdbkf, xmthmwg6, xp97xdfi, yi1jcsqt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 327.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot:gemma-4-31B-it-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 327.5 15.0 39.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iuqrvx76",
   "source_url": "https://omlx.ai/benchmarks/performance/iuqrvx76",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-2go2wjjw-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 6v1jitf0; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 557.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 557.7 62.1 20.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2go2wjjw",
   "source_url": "https://omlx.ai/benchmarks/performance/2go2wjjw",
   "tg_tokens": null,
   "tps": 62.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-2go2wjjw-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 6v1jitf0; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 588.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 588.4 59.6 22.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2go2wjjw",
   "source_url": "https://omlx.ai/benchmarks/performance/2go2wjjw",
   "tg_tokens": null,
   "tps": 59.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-yndn4a62-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 585.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 585.0 84.6 20.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yndn4a62",
   "source_url": "https://omlx.ai/benchmarks/performance/yndn4a62",
   "tg_tokens": null,
   "tps": 84.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-yndn4a62-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1672.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 1,672 82.3 20.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yndn4a62",
   "source_url": "https://omlx.ai/benchmarks/performance/yndn4a62",
   "tg_tokens": null,
   "tps": 82.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-2sa63ofj-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.5 GB at this context; duplicate traces of the same session: 7tqa0nqg; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 363.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 6bit; 4k 363.6 42.4 8.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2sa63ofj",
   "source_url": "https://omlx.ai/benchmarks/performance/2sa63ofj",
   "tg_tokens": null,
   "tps": 42.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-2sa63ofj-ctx16384",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 9.1 GB at this context; duplicate traces of the same session: 7tqa0nqg; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 329.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 6bit; 16k 329.9 41.1 9.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2sa63ofj",
   "source_url": "https://omlx.ai/benchmarks/performance/2sa63ofj",
   "tg_tokens": null,
   "tps": 41.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-kf8vtbm6-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 3.4 GB; TTFT 11148 ms; duplicate traces of the same session: none; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 91.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M1 (7c); M1 (7c) \u00b7 8 GB \u00b7 4bit; 1k tokens 91.9 PP tok/s 13.0 TG tok/s 11148 TTFT (ms) 3.4 Peak mem",
   "ram_gb": 8,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kf8vtbm6",
   "source_url": "https://omlx.ai/benchmarks/performance/kf8vtbm6",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-rpuy3034-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 3.4 GB; TTFT 10916 ms; duplicate traces of the same session: none; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 93.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M1 (7c); M1 (7c) \u00b7 8 GB \u00b7 4bit; 1k tokens 93.8 PP tok/s 13.2 TG tok/s 10916 TTFT (ms) 3.4 Peak mem",
   "ram_gb": 8,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rpuy3034",
   "source_url": "https://omlx.ai/benchmarks/performance/rpuy3034",
   "tg_tokens": null,
   "tps": 13.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-c0lh8u9c-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 2th3jicp, 3tfddckn, aow31qwu, b5rzjakv, jgdgpq7g, ktpc21ch, m9wfqibn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 934.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 934.9 33.5 18.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c0lh8u9c",
   "source_url": "https://omlx.ai/benchmarks/performance/c0lh8u9c",
   "tg_tokens": null,
   "tps": 33.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-c0lh8u9c-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 2th3jicp, 3tfddckn, aow31qwu, b5rzjakv, jgdgpq7g, ktpc21ch, m9wfqibn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 686.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 686.3 33.1 19.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c0lh8u9c",
   "source_url": "https://omlx.ai/benchmarks/performance/c0lh8u9c",
   "tg_tokens": null,
   "tps": 33.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-c0lh8u9c-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 2th3jicp, 3tfddckn, aow31qwu, b5rzjakv, jgdgpq7g, ktpc21ch, m9wfqibn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 624.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 624.3 32.0 20.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c0lh8u9c",
   "source_url": "https://omlx.ai/benchmarks/performance/c0lh8u9c",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-c0lh8u9c-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 2th3jicp, 3tfddckn, aow31qwu, b5rzjakv, jgdgpq7g, ktpc21ch, m9wfqibn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 603.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 603.2 30.0 20.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c0lh8u9c",
   "source_url": "https://omlx.ai/benchmarks/performance/c0lh8u9c",
   "tg_tokens": null,
   "tps": 30.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-c0lh8u9c-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: 2th3jicp, 3tfddckn, aow31qwu, b5rzjakv, jgdgpq7g, ktpc21ch, m9wfqibn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 571.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 571.4 27.4 21.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c0lh8u9c",
   "source_url": "https://omlx.ai/benchmarks/performance/c0lh8u9c",
   "tg_tokens": null,
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-c0lh8u9c-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 2th3jicp, 3tfddckn, aow31qwu, b5rzjakv, jgdgpq7g, ktpc21ch, m9wfqibn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 517.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 517.9 23.9 23.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c0lh8u9c",
   "source_url": "https://omlx.ai/benchmarks/performance/c0lh8u9c",
   "tg_tokens": null,
   "tps": 23.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-c0lh8u9c-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 27.6 GB at this context; duplicate traces of the same session: 2th3jicp, 3tfddckn, aow31qwu, b5rzjakv, jgdgpq7g, ktpc21ch, m9wfqibn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 431.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 431.3 18.4 27.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c0lh8u9c",
   "source_url": "https://omlx.ai/benchmarks/performance/c0lh8u9c",
   "tg_tokens": null,
   "tps": 18.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-c0lh8u9c-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.6 GB at this context; duplicate traces of the same session: 2th3jicp, 3tfddckn, aow31qwu, b5rzjakv, jgdgpq7g, ktpc21ch, m9wfqibn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 366.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 366.3 15.5 31.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c0lh8u9c",
   "source_url": "https://omlx.ai/benchmarks/performance/c0lh8u9c",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-5wxzs1ev-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 2pg65ahp; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 135.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 1k 135.6 14.2 22.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5wxzs1ev",
   "source_url": "https://omlx.ai/benchmarks/performance/5wxzs1ev",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-5wxzs1ev-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: 2pg65ahp; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 124.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 4k 124.5 13.7 25.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5wxzs1ev",
   "source_url": "https://omlx.ai/benchmarks/performance/5wxzs1ev",
   "tg_tokens": null,
   "tps": 13.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hdu31k47-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: z0duw977; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 858.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 858.0 62.7 29.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hdu31k47",
   "source_url": "https://omlx.ai/benchmarks/performance/hdu31k47",
   "tg_tokens": null,
   "tps": 62.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hdu31k47-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 31.1 GB at this context; duplicate traces of the same session: z0duw977; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 863.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 863.6 32.6 31.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hdu31k47",
   "source_url": "https://omlx.ai/benchmarks/performance/hdu31k47",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-p9triwb7-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: socpsa6e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 876.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 876.2 45.0 29.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p9triwb7",
   "source_url": "https://omlx.ai/benchmarks/performance/p9triwb7",
   "tg_tokens": null,
   "tps": 45.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-p9triwb7-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 31.1 GB at this context; duplicate traces of the same session: socpsa6e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 840.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 840.0 39.3 31.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p9triwb7",
   "source_url": "https://omlx.ai/benchmarks/performance/p9triwb7",
   "tg_tokens": null,
   "tps": 39.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9rk5xct4-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 828.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 828.9 46.2 29.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9rk5xct4",
   "source_url": "https://omlx.ai/benchmarks/performance/9rk5xct4",
   "tg_tokens": null,
   "tps": 46.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9rk5xct4-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 31.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 884.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 884.5 34.0 31.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9rk5xct4",
   "source_url": "https://omlx.ai/benchmarks/performance/9rk5xct4",
   "tg_tokens": null,
   "tps": 34.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1svlcptz-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: hv8kb86e, qugsadwo, quui8apj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 929.5,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 929.5 47.1 20.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1svlcptz",
   "source_url": "https://omlx.ai/benchmarks/performance/1svlcptz",
   "tg_tokens": null,
   "tps": 47.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1svlcptz-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: hv8kb86e, qugsadwo, quui8apj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 958.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 958.0 49.1 21.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1svlcptz",
   "source_url": "https://omlx.ai/benchmarks/performance/1svlcptz",
   "tg_tokens": null,
   "tps": 49.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1svlcptz-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: hv8kb86e, qugsadwo, quui8apj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 855.3,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 855.3 53.5 22.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1svlcptz",
   "source_url": "https://omlx.ai/benchmarks/performance/1svlcptz",
   "tg_tokens": null,
   "tps": 53.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1svlcptz-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: hv8kb86e, qugsadwo, quui8apj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 748.8,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 748.8 41.4 24.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1svlcptz",
   "source_url": "https://omlx.ai/benchmarks/performance/1svlcptz",
   "tg_tokens": null,
   "tps": 41.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1svlcptz-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: hv8kb86e, qugsadwo, quui8apj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 707.2,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 707.2 8.5 26.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1svlcptz",
   "source_url": "https://omlx.ai/benchmarks/performance/1svlcptz",
   "tg_tokens": null,
   "tps": 8.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1svlcptz-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.3 GB at this context; duplicate traces of the same session: hv8kb86e, qugsadwo, quui8apj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 573.4,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 573.4 4.1 31.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1svlcptz",
   "source_url": "https://omlx.ai/benchmarks/performance/1svlcptz",
   "tg_tokens": null,
   "tps": 4.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1svlcptz-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 39.6 GB at this context; duplicate traces of the same session: hv8kb86e, qugsadwo, quui8apj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 455.7,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 455.7 3.3 39.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1svlcptz",
   "source_url": "https://omlx.ai/benchmarks/performance/1svlcptz",
   "tg_tokens": null,
   "tps": 3.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-4wo40ocb-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 41.7 GB; TTFT 15031 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 272.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k tokens 272.5 PP tok/s 33.0 TG tok/s 15031 TTFT (ms) 41.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4wo40ocb",
   "source_url": "https://omlx.ai/benchmarks/performance/4wo40ocb",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-9h69rw49-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 99.1 GB; TTFT 1573 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 651.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 651.1 PP tok/s 53.9 TG tok/s 1573 TTFT (ms) 99.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9h69rw49",
   "source_url": "https://omlx.ai/benchmarks/performance/9h69rw49",
   "tg_tokens": null,
   "tps": 53.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lm1s6s7c-ctx1024",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 7.4 GB at this context; duplicate traces of the same session: 13xfj45o; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 919.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-MLX-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 919.6 48.0 7.4 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lm1s6s7c",
   "source_url": "https://omlx.ai/benchmarks/performance/lm1s6s7c",
   "tg_tokens": null,
   "tps": 48.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lm1s6s7c-ctx4096",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 7.7 GB at this context; duplicate traces of the same session: 13xfj45o; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1253.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-MLX-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 1,253 46.5 7.7 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lm1s6s7c",
   "source_url": "https://omlx.ai/benchmarks/performance/lm1s6s7c",
   "tg_tokens": null,
   "tps": 46.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-1lxndf7s-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 9numkt1w, gp2hh2g3, t0hoztln; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 953.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 953.8 100.5 19.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1lxndf7s",
   "source_url": "https://omlx.ai/benchmarks/performance/1lxndf7s",
   "tg_tokens": null,
   "tps": 100.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-1lxndf7s-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 9numkt1w, gp2hh2g3, t0hoztln; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1024.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 1,024 90.1 22.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1lxndf7s",
   "source_url": "https://omlx.ai/benchmarks/performance/1lxndf7s",
   "tg_tokens": null,
   "tps": 90.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-1lxndf7s-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 9numkt1w, gp2hh2g3, t0hoztln; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 872.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 872.8 82.2 21.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1lxndf7s",
   "source_url": "https://omlx.ai/benchmarks/performance/1lxndf7s",
   "tg_tokens": null,
   "tps": 82.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-1lxndf7s-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: 9numkt1w, gp2hh2g3, t0hoztln; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 694.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 32k 694.9 73.0 21.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1lxndf7s",
   "source_url": "https://omlx.ai/benchmarks/performance/1lxndf7s",
   "tg_tokens": null,
   "tps": 73.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-6p0nsdhr-ctx1024",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 163.3 GB at this context; duplicate traces of the same session: 2pi69q07, 5hg509vw, 9ed4xnbn, bpq6a083, naor4n44; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 356.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 356.1 15.7 163.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6p0nsdhr",
   "source_url": "https://omlx.ai/benchmarks/performance/6p0nsdhr",
   "tg_tokens": null,
   "tps": 15.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-6p0nsdhr-ctx4096",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.3 GB at this context; duplicate traces of the same session: 2pi69q07, 5hg509vw, 9ed4xnbn, bpq6a083, naor4n44; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 446.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 446.5 29.7 156.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6p0nsdhr",
   "source_url": "https://omlx.ai/benchmarks/performance/6p0nsdhr",
   "tg_tokens": null,
   "tps": 29.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-6p0nsdhr-ctx8192",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.3 GB at this context; duplicate traces of the same session: 2pi69q07, 5hg509vw, 9ed4xnbn, bpq6a083, naor4n44; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 474.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 474.8 27.7 156.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6p0nsdhr",
   "source_url": "https://omlx.ai/benchmarks/performance/6p0nsdhr",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-6p0nsdhr-ctx16384",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 164.6 GB at this context; duplicate traces of the same session: 2pi69q07, 5hg509vw, 9ed4xnbn, bpq6a083, naor4n44; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 466.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 466.3 2.1 164.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6p0nsdhr",
   "source_url": "https://omlx.ai/benchmarks/performance/6p0nsdhr",
   "tg_tokens": null,
   "tps": 2.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-6p0nsdhr-ctx32768",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.5 GB at this context; duplicate traces of the same session: 2pi69q07, 5hg509vw, 9ed4xnbn, bpq6a083, naor4n44; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 452.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 452.0 26.1 156.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6p0nsdhr",
   "source_url": "https://omlx.ai/benchmarks/performance/6p0nsdhr",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-6p0nsdhr-ctx65536",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 165.7 GB at this context; duplicate traces of the same session: 2pi69q07, 5hg509vw, 9ed4xnbn, bpq6a083, naor4n44; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 427.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 427.8 0.5 165.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6p0nsdhr",
   "source_url": "https://omlx.ai/benchmarks/performance/6p0nsdhr",
   "tg_tokens": null,
   "tps": 0.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-7w9amqoi-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: t8a8j3vd; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 79.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ6e on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 6bit; 1k 79.2 8.9 26.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7w9amqoi",
   "source_url": "https://omlx.ai/benchmarks/performance/7w9amqoi",
   "tg_tokens": null,
   "tps": 8.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-7w9amqoi-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace; peak memory 30.9 GB at this context; duplicate traces of the same session: t8a8j3vd; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 74.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ6e on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 6bit; 4k 74.7 8.3 30.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7w9amqoi",
   "source_url": "https://omlx.ai/benchmarks/performance/7w9amqoi",
   "tg_tokens": null,
   "tps": 8.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-omqdm0q9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: q3f0ip5l; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 79.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ6e on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 6bit; 1k 79.9 9.0 26.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/omqdm0q9",
   "source_url": "https://omlx.ai/benchmarks/performance/omqdm0q9",
   "tg_tokens": null,
   "tps": 9.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-omqdm0q9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace; peak memory 30.9 GB at this context; duplicate traces of the same session: q3f0ip5l; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 74.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ6e on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 6bit; 4k 74.5 8.5 30.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/omqdm0q9",
   "source_url": "https://omlx.ai/benchmarks/performance/omqdm0q9",
   "tg_tokens": null,
   "tps": 8.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-vih45dbj-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace, single run at 195k context; peak memory 177.7 GB; TTFT 93400 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2141.0,
   "provenance": "community",
   "quant": "mixed 4/8bit",
   "quote": "GLM-5.3-Flash-MLX-mixed-4_8bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k tokens 2,141 PP tok/s 50.6 TG tok/s 93400 TTFT (ms) 177.7 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vih45dbj",
   "source_url": "https://omlx.ai/benchmarks/performance/vih45dbj",
   "tg_tokens": null,
   "tps": 50.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-yhz7me44-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 174.2 GB; TTFT 2095 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 488.7,
   "provenance": "community",
   "quant": "mixed 4/8bit",
   "quote": "GLM-5.3-Flash-MLX-mixed-4_8bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k tokens 488.7 PP tok/s 61.2 TG tok/s 2095 TTFT (ms) 174.2 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yhz7me44",
   "source_url": "https://omlx.ai/benchmarks/performance/yhz7me44",
   "tg_tokens": null,
   "tps": 61.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-47uwwdal-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 585.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 585.0 84.6 20.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/47uwwdal",
   "source_url": "https://omlx.ai/benchmarks/performance/47uwwdal",
   "tg_tokens": null,
   "tps": 84.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-47uwwdal-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1672.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 1,672 82.3 20.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/47uwwdal",
   "source_url": "https://omlx.ai/benchmarks/performance/47uwwdal",
   "tg_tokens": null,
   "tps": 82.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-03u71rlj-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: 2oknfl1n; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 84.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 84.9 15.0 16.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/03u71rlj",
   "source_url": "https://omlx.ai/benchmarks/performance/03u71rlj",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-03u71rlj-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 2oknfl1n; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 78.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 78.9 14.5 20.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/03u71rlj",
   "source_url": "https://omlx.ai/benchmarks/performance/03u71rlj",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-l05sc5jm-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 828.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 828.9 46.2 29.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l05sc5jm",
   "source_url": "https://omlx.ai/benchmarks/performance/l05sc5jm",
   "tg_tokens": null,
   "tps": 46.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-l05sc5jm-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 31.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 884.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 884.5 34.0 31.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l05sc5jm",
   "source_url": "https://omlx.ai/benchmarks/performance/l05sc5jm",
   "tg_tokens": null,
   "tps": 34.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-equ2sv1y-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: lk2dde7e; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 125.0 31.6 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/equ2sv1y",
   "source_url": "https://omlx.ai/benchmarks/performance/equ2sv1y",
   "tg_tokens": null,
   "tps": 31.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-equ2sv1y-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: lk2dde7e; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 150.7,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 150.7 39.5 23.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/equ2sv1y",
   "source_url": "https://omlx.ai/benchmarks/performance/equ2sv1y",
   "tg_tokens": null,
   "tps": 39.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-0mb9jvag-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 150.6,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 150.6 20.2 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0mb9jvag",
   "source_url": "https://omlx.ai/benchmarks/performance/0mb9jvag",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-0mb9jvag-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 150.5,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 150.5 16.8 23.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0mb9jvag",
   "source_url": "https://omlx.ai/benchmarks/performance/0mb9jvag",
   "tg_tokens": null,
   "tps": 16.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-8nmmmkpx-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 153.1,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 153.1 18.8 20.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8nmmmkpx",
   "source_url": "https://omlx.ai/benchmarks/performance/8nmmmkpx",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-8nmmmkpx-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 142.7,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 142.7 18.3 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8nmmmkpx",
   "source_url": "https://omlx.ai/benchmarks/performance/8nmmmkpx",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-3plcl8s0-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: grde4zuu; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 73.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 73.3 13.1 18.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3plcl8s0",
   "source_url": "https://omlx.ai/benchmarks/performance/3plcl8s0",
   "tg_tokens": null,
   "tps": 13.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-3plcl8s0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: grde4zuu; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 63.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 63.9 13.6 22.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3plcl8s0",
   "source_url": "https://omlx.ai/benchmarks/performance/3plcl8s0",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-x7t8v5k2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace, single run at 4k context; peak memory 37.9 GB; TTFT 27833 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 147.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp-fp16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k tokens 147.2 PP tok/s 15.1 TG tok/s 27833 TTFT (ms) 37.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x7t8v5k2",
   "source_url": "https://omlx.ai/benchmarks/performance/x7t8v5k2",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-yrd3bnz6-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: 4r8e9qk2, 698m8i7b; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 142.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp-fp16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 1k 142.3 15.6 33.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yrd3bnz6",
   "source_url": "https://omlx.ai/benchmarks/performance/yrd3bnz6",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-yrd3bnz6-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 37.7 GB at this context; duplicate traces of the same session: 4r8e9qk2, 698m8i7b; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 147.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp-fp16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k 147.0 15.3 37.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yrd3bnz6",
   "source_url": "https://omlx.ai/benchmarks/performance/yrd3bnz6",
   "tg_tokens": null,
   "tps": 15.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-yrd3bnz6-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 38.8 GB at this context; duplicate traces of the same session: 4r8e9qk2, 698m8i7b; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 142.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp-fp16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 8k 142.0 18.4 38.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yrd3bnz6",
   "source_url": "https://omlx.ai/benchmarks/performance/yrd3bnz6",
   "tg_tokens": null,
   "tps": 18.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-9ftkscgp-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace, single run at 4k context; peak memory 37.9 GB; TTFT 29041 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 141.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp-fp16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k tokens 141.0 PP tok/s 16.0 TG tok/s 29041 TTFT (ms) 37.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ftkscgp",
   "source_url": "https://omlx.ai/benchmarks/performance/9ftkscgp",
   "tg_tokens": null,
   "tps": 16.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-89ymts27-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.0 GB at this context; duplicate traces of the same session: njog4zrk, p1ks6d89; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 419.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 419.4 81.8 99.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/89ymts27",
   "source_url": "https://omlx.ai/benchmarks/performance/89ymts27",
   "tg_tokens": null,
   "tps": 81.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-89ymts27-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.5 GB at this context; duplicate traces of the same session: njog4zrk, p1ks6d89; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1323.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,323 72.6 99.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/89ymts27",
   "source_url": "https://omlx.ai/benchmarks/performance/89ymts27",
   "tg_tokens": null,
   "tps": 72.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-89ymts27-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.4 GB at this context; duplicate traces of the same session: njog4zrk, p1ks6d89; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1978.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,978 74.9 100.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/89ymts27",
   "source_url": "https://omlx.ai/benchmarks/performance/89ymts27",
   "tg_tokens": null,
   "tps": 74.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-pexciugh-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.7 GB at this context; duplicate traces of the same session: 43hyy2bd, lwy4fp3t, wn4i0ssb, yto1if4o; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2908.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 2,908 99.9 100.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pexciugh",
   "source_url": "https://omlx.ai/benchmarks/performance/pexciugh",
   "tg_tokens": null,
   "tps": 99.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-pexciugh-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.5 GB at this context; duplicate traces of the same session: 43hyy2bd, lwy4fp3t, wn4i0ssb, yto1if4o; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5644.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 5,644 101.5 102.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pexciugh",
   "source_url": "https://omlx.ai/benchmarks/performance/pexciugh",
   "tg_tokens": null,
   "tps": 101.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-pexciugh-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.1 GB at this context; duplicate traces of the same session: 43hyy2bd, lwy4fp3t, wn4i0ssb, yto1if4o; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5624.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 5,624 99.9 103.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pexciugh",
   "source_url": "https://omlx.ai/benchmarks/performance/pexciugh",
   "tg_tokens": null,
   "tps": 99.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-pexciugh-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.4 GB at this context; duplicate traces of the same session: 43hyy2bd, lwy4fp3t, wn4i0ssb, yto1if4o; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5497.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 5,497 80.5 104.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pexciugh",
   "source_url": "https://omlx.ai/benchmarks/performance/pexciugh",
   "tg_tokens": null,
   "tps": 80.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-pexciugh-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 108.8 GB at this context; duplicate traces of the same session: 43hyy2bd, lwy4fp3t, wn4i0ssb, yto1if4o; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5331.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,331 62.6 108.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pexciugh",
   "source_url": "https://omlx.ai/benchmarks/performance/pexciugh",
   "tg_tokens": null,
   "tps": 62.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-rtcr84zf-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.7 GB at this context; duplicate traces of the same session: 02w6zmlr, hbmntrjp, vi3g3lt2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2848.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 2,848 156.9 100.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rtcr84zf",
   "source_url": "https://omlx.ai/benchmarks/performance/rtcr84zf",
   "tg_tokens": null,
   "tps": 156.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-rtcr84zf-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.1 GB at this context; duplicate traces of the same session: 02w6zmlr, hbmntrjp, vi3g3lt2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5631.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 5,631 127.6 103.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rtcr84zf",
   "source_url": "https://omlx.ai/benchmarks/performance/rtcr84zf",
   "tg_tokens": null,
   "tps": 127.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-rtcr84zf-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.4 GB at this context; duplicate traces of the same session: 02w6zmlr, hbmntrjp, vi3g3lt2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5510.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 5,510 134.8 104.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rtcr84zf",
   "source_url": "https://omlx.ai/benchmarks/performance/rtcr84zf",
   "tg_tokens": null,
   "tps": 134.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-rtcr84zf-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 108.6 GB at this context; duplicate traces of the same session: 02w6zmlr, hbmntrjp, vi3g3lt2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5338.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,338 117.3 108.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rtcr84zf",
   "source_url": "https://omlx.ai/benchmarks/performance/rtcr84zf",
   "tg_tokens": null,
   "tps": 117.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ysvsqgvq-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.2 GB at this context; duplicate traces of the same session: m2wi8j6h; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5477.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 5,477 155.0 102.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ysvsqgvq",
   "source_url": "https://omlx.ai/benchmarks/performance/ysvsqgvq",
   "tg_tokens": null,
   "tps": 155.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ysvsqgvq-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.1 GB at this context; duplicate traces of the same session: m2wi8j6h; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5603.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 5,603 105.3 103.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ysvsqgvq",
   "source_url": "https://omlx.ai/benchmarks/performance/ysvsqgvq",
   "tg_tokens": null,
   "tps": 105.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-cgcqqiv6-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.4 GB at this context; duplicate traces of the same session: vqhicp60; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5514.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 5,514 118.4 102.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cgcqqiv6",
   "source_url": "https://omlx.ai/benchmarks/performance/cgcqqiv6",
   "tg_tokens": null,
   "tps": 118.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-cgcqqiv6-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.1 GB at this context; duplicate traces of the same session: vqhicp60; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5619.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 5,619 105.4 103.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cgcqqiv6",
   "source_url": "https://omlx.ai/benchmarks/performance/cgcqqiv6",
   "tg_tokens": null,
   "tps": 105.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-4vzkigf7-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: gchk1mmw, tgg2eukv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 269.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 269.0 85.5 22.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4vzkigf7",
   "source_url": "https://omlx.ai/benchmarks/performance/4vzkigf7",
   "tg_tokens": null,
   "tps": 85.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-4vzkigf7-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: gchk1mmw, tgg2eukv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 254.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 254.7 58.3 23.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4vzkigf7",
   "source_url": "https://omlx.ai/benchmarks/performance/4vzkigf7",
   "tg_tokens": null,
   "tps": 58.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-4vzkigf7-ctx65536",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: gchk1mmw, tgg2eukv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 205.6 36.7 27.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4vzkigf7",
   "source_url": "https://omlx.ai/benchmarks/performance/4vzkigf7",
   "tg_tokens": null,
   "tps": 36.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-60lt3qro-ctx32768",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace, single run at 32k context; peak memory 16.0 GB; TTFT 51382 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 637.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 32k tokens 637.7 PP tok/s 20.3 TG tok/s 51382 TTFT (ms) 16.0 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/60lt3qro",
   "source_url": "https://omlx.ai/benchmarks/performance/60lt3qro",
   "tg_tokens": null,
   "tps": 20.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-p20k9icp-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: y02ho85y; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 81.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ6e on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 6bit; 1k 81.2 6.9 27.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p20k9icp",
   "source_url": "https://omlx.ai/benchmarks/performance/p20k9icp",
   "tg_tokens": null,
   "tps": 6.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-p20k9icp-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace; peak memory 31.6 GB at this context; duplicate traces of the same session: y02ho85y; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 76.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ6e on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 6bit; 4k 76.2 6.3 31.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p20k9icp",
   "source_url": "https://omlx.ai/benchmarks/performance/p20k9icp",
   "tg_tokens": null,
   "tps": 6.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-xullipgc-ctx16384",
   "model": "LFM2.5-2.6B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 4.9 GB; TTFT 17688 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 926.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-2.6B-MLX-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 16k tokens 926.3 PP tok/s 82.5 TG tok/s 17688 TTFT (ms) 4.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xullipgc",
   "source_url": "https://omlx.ai/benchmarks/performance/xullipgc",
   "tg_tokens": null,
   "tps": 82.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ey7ge67w-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: 8vdxo5sl, hw705cim, lq9xmgaw, xkzlhn8a; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 392.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 392.3 58.6 5.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ey7ge67w",
   "source_url": "https://omlx.ai/benchmarks/performance/ey7ge67w",
   "tg_tokens": null,
   "tps": 58.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ey7ge67w-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.4 GB at this context; duplicate traces of the same session: 8vdxo5sl, hw705cim, lq9xmgaw, xkzlhn8a; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 365.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 365.6 51.1 7.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ey7ge67w",
   "source_url": "https://omlx.ai/benchmarks/performance/ey7ge67w",
   "tg_tokens": null,
   "tps": 51.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ey7ge67w-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.9 GB at this context; duplicate traces of the same session: 8vdxo5sl, hw705cim, lq9xmgaw, xkzlhn8a; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 353.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 353.1 48.0 7.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ey7ge67w",
   "source_url": "https://omlx.ai/benchmarks/performance/ey7ge67w",
   "tg_tokens": null,
   "tps": 48.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-ey7ge67w-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.9 GB at this context; duplicate traces of the same session: 8vdxo5sl, hw705cim, lq9xmgaw, xkzlhn8a; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 331.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 331.7 45.3 8.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ey7ge67w",
   "source_url": "https://omlx.ai/benchmarks/performance/ey7ge67w",
   "tg_tokens": null,
   "tps": 45.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ey7ge67w-ctx32768",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.3 GB at this context; duplicate traces of the same session: 8vdxo5sl, hw705cim, lq9xmgaw, xkzlhn8a; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 295.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 32k 295.0 42.1 10.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ey7ge67w",
   "source_url": "https://omlx.ai/benchmarks/performance/ey7ge67w",
   "tg_tokens": null,
   "tps": 42.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-5wbv33xt-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: ugot63ji; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 462.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 462.2 50.2 5.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5wbv33xt",
   "source_url": "https://omlx.ai/benchmarks/performance/5wbv33xt",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-5wbv33xt-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: ugot63ji; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 478.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 478.3 49.5 6.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5wbv33xt",
   "source_url": "https://omlx.ai/benchmarks/performance/5wbv33xt",
   "tg_tokens": null,
   "tps": 49.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ow3042bl-ctx1024",
   "model": "Qwen3.6-27B",
   "notes": "UD (Unsloth Dynamic) build; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 9fa8iqp6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 375.3,
   "provenance": "community",
   "quant": "Q4_K_XL",
   "quote": "Qwen3.6-27B-UD-Q4_K_XL-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 375.3 12.7 21.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ow3042bl",
   "source_url": "https://omlx.ai/benchmarks/performance/ow3042bl",
   "tg_tokens": null,
   "tps": 12.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ow3042bl-ctx4096",
   "model": "Qwen3.6-27B",
   "notes": "UD (Unsloth Dynamic) build; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: 9fa8iqp6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 469.8,
   "provenance": "community",
   "quant": "Q4_K_XL",
   "quote": "Qwen3.6-27B-UD-Q4_K_XL-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 469.8 12.7 22.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ow3042bl",
   "source_url": "https://omlx.ai/benchmarks/performance/ow3042bl",
   "tg_tokens": null,
   "tps": 12.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-705c0p2p-ctx1024",
   "model": "Qwen3.6-35B-A3B-Fable-Holo3.1-Qwopus-KAT-Coder-C-qx86-hi",
   "notes": "community merge; oMLX community benchmark trace; peak memory 37.1 GB at this context; duplicate traces of the same session: ki0hl2z8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2030.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-Fable-Holo3.1-Qwopus-KAT-Coder-C-qx86-hi-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 2,030 67.3 37.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/705c0p2p",
   "source_url": "https://omlx.ai/benchmarks/performance/705c0p2p",
   "tg_tokens": null,
   "tps": 67.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-705c0p2p-ctx4096",
   "model": "Qwen3.6-35B-A3B-Fable-Holo3.1-Qwopus-KAT-Coder-C-qx86-hi",
   "notes": "community merge; oMLX community benchmark trace; peak memory 37.9 GB at this context; duplicate traces of the same session: ki0hl2z8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2952.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-Fable-Holo3.1-Qwopus-KAT-Coder-C-qx86-hi-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 2,952 66.1 37.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/705c0p2p",
   "source_url": "https://omlx.ai/benchmarks/performance/705c0p2p",
   "tg_tokens": null,
   "tps": 66.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-humm509u-ctx1024",
   "model": "Qwen3.6-35B-A3B-Fable-Holo3.1-Qwopus-KAT-Coder-C-qx86-hi",
   "notes": "community merge; oMLX community benchmark trace; peak memory 37.1 GB at this context; duplicate traces of the same session: 7q7pbmo0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 678.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-Fable-Holo3.1-Qwopus-KAT-Coder-C-qx86-hi-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 678.5 70.0 37.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/humm509u",
   "source_url": "https://omlx.ai/benchmarks/performance/humm509u",
   "tg_tokens": null,
   "tps": 70.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-humm509u-ctx4096",
   "model": "Qwen3.6-35B-A3B-Fable-Holo3.1-Qwopus-KAT-Coder-C-qx86-hi",
   "notes": "community merge; oMLX community benchmark trace; peak memory 37.9 GB at this context; duplicate traces of the same session: 7q7pbmo0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1937.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-Fable-Holo3.1-Qwopus-KAT-Coder-C-qx86-hi-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 1,937 65.1 37.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/humm509u",
   "source_url": "https://omlx.ai/benchmarks/performance/humm509u",
   "tg_tokens": null,
   "tps": 65.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-5keork3u-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "VL build, FP16 attention; oMLX community benchmark trace; peak memory 26.9 GB at this context; duplicate traces of the same session: 5xcavqs2; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1242.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "Qwen3.6-35B-A3B-MLX-VL-oQ5-FP16 on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 5bit; 16k 1,242 65.5 26.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5keork3u",
   "source_url": "https://omlx.ai/benchmarks/performance/5keork3u",
   "tg_tokens": null,
   "tps": 65.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-5keork3u-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "VL build, FP16 attention; oMLX community benchmark trace; peak memory 30.0 GB at this context; duplicate traces of the same session: 5xcavqs2; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 471.1,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "Qwen3.6-35B-A3B-MLX-VL-oQ5-FP16 on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 5bit; 128k 471.1 40.4 30.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5keork3u",
   "source_url": "https://omlx.ai/benchmarks/performance/5keork3u",
   "tg_tokens": null,
   "tps": 40.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-g5nguv0s-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: ypuopg5d; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 416.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 416.5 10.1 28.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g5nguv0s",
   "source_url": "https://omlx.ai/benchmarks/performance/g5nguv0s",
   "tg_tokens": null,
   "tps": 10.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-g5nguv0s-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: ypuopg5d; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 474.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 474.4 10.0 29.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g5nguv0s",
   "source_url": "https://omlx.ai/benchmarks/performance/g5nguv0s",
   "tg_tokens": null,
   "tps": 10.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-c9yaxw3d-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: 6hd2so61; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 891.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 891.0 32.4 17.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c9yaxw3d",
   "source_url": "https://omlx.ai/benchmarks/performance/c9yaxw3d",
   "tg_tokens": null,
   "tps": 32.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-c9yaxw3d-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 6hd2so61; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 980.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 980.0 31.0 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c9yaxw3d",
   "source_url": "https://omlx.ai/benchmarks/performance/c9yaxw3d",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-heiq6poo-ctx131072",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 24.9 GB; TTFT 208973 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 627.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k tokens 627.2 PP tok/s 21.2 TG tok/s 208973 TTFT (ms) 24.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/heiq6poo",
   "source_url": "https://omlx.ai/benchmarks/performance/heiq6poo",
   "tg_tokens": null,
   "tps": 21.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-9d26et8l-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 150.6,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 150.6 20.2 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9d26et8l",
   "source_url": "https://omlx.ai/benchmarks/performance/9d26et8l",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-9d26et8l-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 150.5,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 150.5 16.8 23.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9d26et8l",
   "source_url": "https://omlx.ai/benchmarks/performance/9d26et8l",
   "tg_tokens": null,
   "tps": 16.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-e8ebugux-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: zmf9b44e; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 125.0 28.5 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e8ebugux",
   "source_url": "https://omlx.ai/benchmarks/performance/e8ebugux",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-e8ebugux-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: zmf9b44e; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 150.9,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 150.9 25.5 23.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e8ebugux",
   "source_url": "https://omlx.ai/benchmarks/performance/e8ebugux",
   "tg_tokens": null,
   "tps": 25.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-6mg4q7i1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 153.1,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 153.1 18.8 20.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6mg4q7i1",
   "source_url": "https://omlx.ai/benchmarks/performance/6mg4q7i1",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-6mg4q7i1-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 142.7,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 142.7 18.3 21.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6mg4q7i1",
   "source_url": "https://omlx.ai/benchmarks/performance/6mg4q7i1",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-380or2oj-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "SAGE-PATHS MTP build; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: 2nuc4778; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 341.3,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-SAGE-PATHS-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 4k 341.3 71.5 25.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/380or2oj",
   "source_url": "https://omlx.ai/benchmarks/performance/380or2oj",
   "tg_tokens": null,
   "tps": 71.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-380or2oj-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "SAGE-PATHS MTP build; oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: 2nuc4778; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 336.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-SAGE-PATHS-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 8k 336.7 41.5 25.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/380or2oj",
   "source_url": "https://omlx.ai/benchmarks/performance/380or2oj",
   "tg_tokens": null,
   "tps": 41.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-380or2oj-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "SAGE-PATHS MTP build; oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: 2nuc4778; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 324.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-SAGE-PATHS-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 16k 324.0 49.1 26.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/380or2oj",
   "source_url": "https://omlx.ai/benchmarks/performance/380or2oj",
   "tg_tokens": null,
   "tps": 49.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-380or2oj-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "SAGE-PATHS MTP build; oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: 2nuc4778; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 301.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-SAGE-PATHS-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 32k 301.0 56.7 28.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/380or2oj",
   "source_url": "https://omlx.ai/benchmarks/performance/380or2oj",
   "tg_tokens": null,
   "tps": 56.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-d1rjdgrw-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace, single run at 4k context; peak memory 37.9 GB; TTFT 28262 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 144.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp-fp16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k tokens 144.9 PP tok/s 18.2 TG tok/s 28262 TTFT (ms) 37.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d1rjdgrw",
   "source_url": "https://omlx.ai/benchmarks/performance/d1rjdgrw",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-v1ngd5dg-ctx1024",
   "model": "Qwen3.8-27B-pi",
   "notes": "oMLX community benchmark trace; peak memory 53.3 GB at this context; duplicate traces of the same session: 621lg7ar; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 330.3,
   "provenance": "community",
   "quant": "FP8",
   "quote": "Qwen3.8-27B-pi-FP8 on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 unknown; 1k 330.3 11.8 53.3 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v1ngd5dg",
   "source_url": "https://omlx.ai/benchmarks/performance/v1ngd5dg",
   "tg_tokens": null,
   "tps": 11.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-v1ngd5dg-ctx4096",
   "model": "Qwen3.8-27B-pi",
   "notes": "oMLX community benchmark trace; peak memory 56.9 GB at this context; duplicate traces of the same session: 621lg7ar; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 372.3,
   "provenance": "community",
   "quant": "FP8",
   "quote": "Qwen3.8-27B-pi-FP8 on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 unknown; 4k 372.3 11.4 56.9 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v1ngd5dg",
   "source_url": "https://omlx.ai/benchmarks/performance/v1ngd5dg",
   "tg_tokens": null,
   "tps": 11.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wfy8ap0s-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: h9fz5yvo, spxn59tn, tnp8n2ls; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1235.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,235 86.6 71.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wfy8ap0s",
   "source_url": "https://omlx.ai/benchmarks/performance/wfy8ap0s",
   "tg_tokens": null,
   "tps": 86.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wfy8ap0s-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: h9fz5yvo, spxn59tn, tnp8n2ls; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2087.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,087 71.6 71.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wfy8ap0s",
   "source_url": "https://omlx.ai/benchmarks/performance/wfy8ap0s",
   "tg_tokens": null,
   "tps": 71.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wfy8ap0s-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.0 GB at this context; duplicate traces of the same session: h9fz5yvo, spxn59tn, tnp8n2ls; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2086.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,086 69.7 73.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wfy8ap0s",
   "source_url": "https://omlx.ai/benchmarks/performance/wfy8ap0s",
   "tg_tokens": null,
   "tps": 69.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wfy8ap0s-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.0 GB at this context; duplicate traces of the same session: h9fz5yvo, spxn59tn, tnp8n2ls; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1688.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 1,688 65.0 79.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wfy8ap0s",
   "source_url": "https://omlx.ai/benchmarks/performance/wfy8ap0s",
   "tg_tokens": null,
   "tps": 65.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fieyk2ax-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: 8e10wpkh, ckkzkxxl, qzdybk0o, y22ep11b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 930.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 930.5 67.6 71.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fieyk2ax",
   "source_url": "https://omlx.ai/benchmarks/performance/fieyk2ax",
   "tg_tokens": null,
   "tps": 67.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fieyk2ax-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: 8e10wpkh, ckkzkxxl, qzdybk0o, y22ep11b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1981.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,981 57.8 71.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fieyk2ax",
   "source_url": "https://omlx.ai/benchmarks/performance/fieyk2ax",
   "tg_tokens": null,
   "tps": 57.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fieyk2ax-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.0 GB at this context; duplicate traces of the same session: 8e10wpkh, ckkzkxxl, qzdybk0o, y22ep11b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2064.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,064 61.9 73.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fieyk2ax",
   "source_url": "https://omlx.ai/benchmarks/performance/fieyk2ax",
   "tg_tokens": null,
   "tps": 61.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fieyk2ax-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.3 GB at this context; duplicate traces of the same session: 8e10wpkh, ckkzkxxl, qzdybk0o, y22ep11b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1683.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 1,683 50.7 79.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fieyk2ax",
   "source_url": "https://omlx.ai/benchmarks/performance/fieyk2ax",
   "tg_tokens": null,
   "tps": 50.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fieyk2ax-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.9 GB at this context; duplicate traces of the same session: 8e10wpkh, ckkzkxxl, qzdybk0o, y22ep11b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1359.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 1,359 47.4 80.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fieyk2ax",
   "source_url": "https://omlx.ai/benchmarks/performance/fieyk2ax",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-igyk87vh-ctx1024",
   "model": "Ternary-Bonsai-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 8.9 GB at this context; duplicate traces of the same session: g3ii0jyg; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 475.7,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-27B-mlx-2bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 2bit; 1k 475.7 32.4 8.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/igyk87vh",
   "source_url": "https://omlx.ai/benchmarks/performance/igyk87vh",
   "tg_tokens": null,
   "tps": 32.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-igyk87vh-ctx4096",
   "model": "Ternary-Bonsai-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 10.2 GB at this context; duplicate traces of the same session: g3ii0jyg; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 506.0,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-27B-mlx-2bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 2bit; 4k 506.0 31.6 10.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/igyk87vh",
   "source_url": "https://omlx.ai/benchmarks/performance/igyk87vh",
   "tg_tokens": null,
   "tps": 31.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-w0dbmvbx-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "QAT OptiQ build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: lhqr2bju; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3822.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 3,822 96.3 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w0dbmvbx",
   "source_url": "https://omlx.ai/benchmarks/performance/w0dbmvbx",
   "tg_tokens": null,
   "tps": 96.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-w0dbmvbx-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "QAT OptiQ build; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: lhqr2bju; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4028.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 4,028 94.4 20.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w0dbmvbx",
   "source_url": "https://omlx.ai/benchmarks/performance/w0dbmvbx",
   "tg_tokens": null,
   "tps": 94.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9k5argnz-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "QAT OptiQ build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: bayhzdk1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3638.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 3,638 95.7 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9k5argnz",
   "source_url": "https://omlx.ai/benchmarks/performance/9k5argnz",
   "tg_tokens": null,
   "tps": 95.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9k5argnz-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "QAT OptiQ build; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: bayhzdk1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3999.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 3,999 94.6 20.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9k5argnz",
   "source_url": "https://omlx.ai/benchmarks/performance/9k5argnz",
   "tg_tokens": null,
   "tps": 94.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1mphzyih-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; Jundot build; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: f0xx6dly; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2984.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,984 74.3 21.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1mphzyih",
   "source_url": "https://omlx.ai/benchmarks/performance/1mphzyih",
   "tg_tokens": null,
   "tps": 74.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1mphzyih-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; Jundot build; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: f0xx6dly; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2205.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 2,205 63.1 21.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1mphzyih",
   "source_url": "https://omlx.ai/benchmarks/performance/1mphzyih",
   "tg_tokens": null,
   "tps": 63.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-gv3guu5n-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; Jundot build; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: x8tb3isy; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2971.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,971 75.8 21.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gv3guu5n",
   "source_url": "https://omlx.ai/benchmarks/performance/gv3guu5n",
   "tg_tokens": null,
   "tps": 75.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-gv3guu5n-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; Jundot build; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: x8tb3isy; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2212.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 2,212 68.7 21.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gv3guu5n",
   "source_url": "https://omlx.ai/benchmarks/performance/gv3guu5n",
   "tg_tokens": null,
   "tps": 68.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-bq9yj6vv-ctx1024",
   "model": "MiMo-V2.6-Flash-MOPD",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 156.7 GB; TTFT 545 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1878.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiMo-V2.6-Flash-MOPD-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k tokens 1,878 PP tok/s 78.7 TG tok/s 545 TTFT (ms) 156.7 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bq9yj6vv",
   "source_url": "https://omlx.ai/benchmarks/performance/bq9yj6vv",
   "tg_tokens": null,
   "tps": 78.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-2uvmi29q-ctx1024",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 156.7 GB; TTFT 1855 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 552.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "MiMo-V2.6-Flash-RL-mxfp4-q8 on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k tokens 552.0 PP tok/s 71.4 TG tok/s 1855 TTFT (ms) 156.7 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2uvmi29q",
   "source_url": "https://omlx.ai/benchmarks/performance/2uvmi29q",
   "tg_tokens": null,
   "tps": 71.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-b7dqu060-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 0upjhkkd, 81fv1s9p, spud2cau, x6ap9bvf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1210.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,210 92.1 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b7dqu060",
   "source_url": "https://omlx.ai/benchmarks/performance/b7dqu060",
   "tg_tokens": null,
   "tps": 92.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-b7dqu060-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 0upjhkkd, 81fv1s9p, spud2cau, x6ap9bvf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1697.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,697 92.6 21.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b7dqu060",
   "source_url": "https://omlx.ai/benchmarks/performance/b7dqu060",
   "tg_tokens": null,
   "tps": 92.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-b7dqu060-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 0upjhkkd, 81fv1s9p, spud2cau, x6ap9bvf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1619.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,619 82.8 22.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b7dqu060",
   "source_url": "https://omlx.ai/benchmarks/performance/b7dqu060",
   "tg_tokens": null,
   "tps": 82.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-b7dqu060-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 0upjhkkd, 81fv1s9p, spud2cau, x6ap9bvf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1371.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,371 72.8 21.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b7dqu060",
   "source_url": "https://omlx.ai/benchmarks/performance/b7dqu060",
   "tg_tokens": null,
   "tps": 72.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-b7dqu060-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: 0upjhkkd, 81fv1s9p, spud2cau, x6ap9bvf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 860.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 860.3 47.9 21.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b7dqu060",
   "source_url": "https://omlx.ai/benchmarks/performance/b7dqu060",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-axdju853-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 2dn619z9, dkcb17mo, iwyjq73v, qbpr5124; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 693.1,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 693.1 63.4 21.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/axdju853",
   "source_url": "https://omlx.ai/benchmarks/performance/axdju853",
   "tg_tokens": null,
   "tps": 63.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-axdju853-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 2dn619z9, dkcb17mo, iwyjq73v, qbpr5124; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 783.1,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 783.1 59.5 23.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/axdju853",
   "source_url": "https://omlx.ai/benchmarks/performance/axdju853",
   "tg_tokens": null,
   "tps": 59.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-axdju853-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: 2dn619z9, dkcb17mo, iwyjq73v, qbpr5124; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 710.3,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 710.3 57.6 24.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/axdju853",
   "source_url": "https://omlx.ai/benchmarks/performance/axdju853",
   "tg_tokens": null,
   "tps": 57.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-axdju853-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: 2dn619z9, dkcb17mo, iwyjq73v, qbpr5124; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 604.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 604.0 54.0 25.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/axdju853",
   "source_url": "https://omlx.ai/benchmarks/performance/axdju853",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-axdju853-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: 2dn619z9, dkcb17mo, iwyjq73v, qbpr5124; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 468.2,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 32k 468.2 49.6 26.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/axdju853",
   "source_url": "https://omlx.ai/benchmarks/performance/axdju853",
   "tg_tokens": null,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5swumzq8-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 38.0 GB at this context; duplicate traces of the same session: 24d1q83d, 2m2v2n3n, 6qas9lig, izog7n0i, jhqy8vf7, jl7xmn3t, pgpbz2eo; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2150.0,
   "provenance": "community",
   "quant": "mixed 9bit",
   "quote": "Qwen3.6-35B-A3B-MLX-mixed-9bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 2,150 93.7 38.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5swumzq8",
   "source_url": "https://omlx.ai/benchmarks/performance/5swumzq8",
   "tg_tokens": null,
   "tps": 93.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5swumzq8-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 40.0 GB at this context; duplicate traces of the same session: 24d1q83d, 2m2v2n3n, 6qas9lig, izog7n0i, jhqy8vf7, jl7xmn3t, pgpbz2eo; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2530.0,
   "provenance": "community",
   "quant": "mixed 9bit",
   "quote": "Qwen3.6-35B-A3B-MLX-mixed-9bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 2,530 92.1 40.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5swumzq8",
   "source_url": "https://omlx.ai/benchmarks/performance/5swumzq8",
   "tg_tokens": null,
   "tps": 92.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5swumzq8-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 40.6 GB at this context; duplicate traces of the same session: 24d1q83d, 2m2v2n3n, 6qas9lig, izog7n0i, jhqy8vf7, jl7xmn3t, pgpbz2eo; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2427.0,
   "provenance": "community",
   "quant": "mixed 9bit",
   "quote": "Qwen3.6-35B-A3B-MLX-mixed-9bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 8k 2,427 91.1 40.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5swumzq8",
   "source_url": "https://omlx.ai/benchmarks/performance/5swumzq8",
   "tg_tokens": null,
   "tps": 91.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5swumzq8-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 40.0 GB at this context; duplicate traces of the same session: 24d1q83d, 2m2v2n3n, 6qas9lig, izog7n0i, jhqy8vf7, jl7xmn3t, pgpbz2eo; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2190.0,
   "provenance": "community",
   "quant": "mixed 9bit",
   "quote": "Qwen3.6-35B-A3B-MLX-mixed-9bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 16k 2,190 84.3 40.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5swumzq8",
   "source_url": "https://omlx.ai/benchmarks/performance/5swumzq8",
   "tg_tokens": null,
   "tps": 84.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5swumzq8-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 40.3 GB at this context; duplicate traces of the same session: 24d1q83d, 2m2v2n3n, 6qas9lig, izog7n0i, jhqy8vf7, jl7xmn3t, pgpbz2eo; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1804.0,
   "provenance": "community",
   "quant": "mixed 9bit",
   "quote": "Qwen3.6-35B-A3B-MLX-mixed-9bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 32k 1,804 78.1 40.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5swumzq8",
   "source_url": "https://omlx.ai/benchmarks/performance/5swumzq8",
   "tg_tokens": null,
   "tps": 78.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5swumzq8-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 40.9 GB at this context; duplicate traces of the same session: 24d1q83d, 2m2v2n3n, 6qas9lig, izog7n0i, jhqy8vf7, jl7xmn3t, pgpbz2eo; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1325.0,
   "provenance": "community",
   "quant": "mixed 9bit",
   "quote": "Qwen3.6-35B-A3B-MLX-mixed-9bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 64k 1,325 67.3 40.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5swumzq8",
   "source_url": "https://omlx.ai/benchmarks/performance/5swumzq8",
   "tg_tokens": null,
   "tps": 67.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5swumzq8-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 42.1 GB at this context; duplicate traces of the same session: 24d1q83d, 2m2v2n3n, 6qas9lig, izog7n0i, jhqy8vf7, jl7xmn3t, pgpbz2eo; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 862.0,
   "provenance": "community",
   "quant": "mixed 9bit",
   "quote": "Qwen3.6-35B-A3B-MLX-mixed-9bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 128k 862.0 55.1 42.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5swumzq8",
   "source_url": "https://omlx.ai/benchmarks/performance/5swumzq8",
   "tg_tokens": null,
   "tps": 55.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5swumzq8-ctx199680",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 43.5 GB at this context; duplicate traces of the same session: 24d1q83d, 2m2v2n3n, 6qas9lig, izog7n0i, jhqy8vf7, jl7xmn3t, pgpbz2eo; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 629.6,
   "provenance": "community",
   "quant": "mixed 9bit",
   "quote": "Qwen3.6-35B-A3B-MLX-mixed-9bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 195k 629.6 46.3 43.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5swumzq8",
   "source_url": "https://omlx.ai/benchmarks/performance/5swumzq8",
   "tg_tokens": null,
   "tps": 46.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-6ixcl7lb-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: hgoog5tn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1758.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 1,758 64.4 21.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ixcl7lb",
   "source_url": "https://omlx.ai/benchmarks/performance/6ixcl7lb",
   "tg_tokens": null,
   "tps": 64.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-6ixcl7lb-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: hgoog5tn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2569.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 2,569 52.3 22.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ixcl7lb",
   "source_url": "https://omlx.ai/benchmarks/performance/6ixcl7lb",
   "tg_tokens": null,
   "tps": 52.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-itwqswg8-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: ey8g3nhf; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1161.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "Qwen3.6-35B-A3B-oQ5-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 5bit; 16k 1,161 66.0 27.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/itwqswg8",
   "source_url": "https://omlx.ai/benchmarks/performance/itwqswg8",
   "tg_tokens": null,
   "tps": 66.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-itwqswg8-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 31.1 GB at this context; duplicate traces of the same session: ey8g3nhf; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 427.7,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "Qwen3.6-35B-A3B-oQ5-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 5bit; 128k 427.7 42.7 31.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/itwqswg8",
   "source_url": "https://omlx.ai/benchmarks/performance/itwqswg8",
   "tg_tokens": null,
   "tps": 42.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-mqx4tpq2-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 38.2 GB at this context; duplicate traces of the same session: 31g75od8, fz63qiwa; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1188.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 1,188 106.4 38.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mqx4tpq2",
   "source_url": "https://omlx.ai/benchmarks/performance/mqx4tpq2",
   "tg_tokens": null,
   "tps": 106.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-mqx4tpq2-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 40.3 GB at this context; duplicate traces of the same session: 31g75od8, fz63qiwa; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2421.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 2,421 135.5 40.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mqx4tpq2",
   "source_url": "https://omlx.ai/benchmarks/performance/mqx4tpq2",
   "tg_tokens": null,
   "tps": 135.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-mqx4tpq2-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 41.0 GB at this context; duplicate traces of the same session: 31g75od8, fz63qiwa; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2278.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 8k 2,278 140.4 41.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mqx4tpq2",
   "source_url": "https://omlx.ai/benchmarks/performance/mqx4tpq2",
   "tg_tokens": null,
   "tps": 140.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-mqx4tpq2-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 40.3 GB at this context; duplicate traces of the same session: 31g75od8, fz63qiwa; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2039.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 16k 2,039 167.6 40.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mqx4tpq2",
   "source_url": "https://omlx.ai/benchmarks/performance/mqx4tpq2",
   "tg_tokens": null,
   "tps": 167.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-mqx4tpq2-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 40.7 GB at this context; duplicate traces of the same session: 31g75od8, fz63qiwa; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1704.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 32k 1,704 145.1 40.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mqx4tpq2",
   "source_url": "https://omlx.ai/benchmarks/performance/mqx4tpq2",
   "tg_tokens": null,
   "tps": 145.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-mqx4tpq2-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 41.4 GB at this context; duplicate traces of the same session: 31g75od8, fz63qiwa; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1236.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 64k 1,236 122.2 41.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mqx4tpq2",
   "source_url": "https://omlx.ai/benchmarks/performance/mqx4tpq2",
   "tg_tokens": null,
   "tps": 122.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-mqx4tpq2-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 43.0 GB at this context; duplicate traces of the same session: 31g75od8, fz63qiwa; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 793.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 128k 793.1 104.8 43.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mqx4tpq2",
   "source_url": "https://omlx.ai/benchmarks/performance/mqx4tpq2",
   "tg_tokens": null,
   "tps": 104.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-mqx4tpq2-ctx199680",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 44.8 GB at this context; duplicate traces of the same session: 31g75od8, fz63qiwa; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 575.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 195k 575.6 66.6 44.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mqx4tpq2",
   "source_url": "https://omlx.ai/benchmarks/performance/mqx4tpq2",
   "tg_tokens": null,
   "tps": 66.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-r0fow63i-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "SAGE-PATHS MTP build; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: jiqnwtp2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 341.3,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-SAGE-PATHS-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 4k 341.3 71.5 25.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r0fow63i",
   "source_url": "https://omlx.ai/benchmarks/performance/r0fow63i",
   "tg_tokens": null,
   "tps": 71.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-r0fow63i-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "SAGE-PATHS MTP build; oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: jiqnwtp2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 336.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-SAGE-PATHS-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 8k 336.7 41.5 25.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r0fow63i",
   "source_url": "https://omlx.ai/benchmarks/performance/r0fow63i",
   "tg_tokens": null,
   "tps": 41.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-r0fow63i-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "SAGE-PATHS MTP build; oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: jiqnwtp2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 324.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-SAGE-PATHS-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 16k 324.0 49.1 26.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r0fow63i",
   "source_url": "https://omlx.ai/benchmarks/performance/r0fow63i",
   "tg_tokens": null,
   "tps": 49.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-r0fow63i-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "SAGE-PATHS MTP build; oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: jiqnwtp2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 301.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-SAGE-PATHS-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 32k 301.0 56.7 28.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r0fow63i",
   "source_url": "https://omlx.ai/benchmarks/performance/r0fow63i",
   "tg_tokens": null,
   "tps": 56.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jcad1kzq-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 43.5 GB; TTFT 280059 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 468.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k tokens 468.0 PP tok/s 37.1 TG tok/s 280059 TTFT (ms) 43.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jcad1kzq",
   "source_url": "https://omlx.ai/benchmarks/performance/jcad1kzq",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-mqnzn9p7-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: qvp0nzjr, v0awgvhd; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 877.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 16k 877.2 61.9 71.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mqnzn9p7",
   "source_url": "https://omlx.ai/benchmarks/performance/mqnzn9p7",
   "tg_tokens": null,
   "tps": 61.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-mqnzn9p7-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.4 GB at this context; duplicate traces of the same session: qvp0nzjr, v0awgvhd; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 996.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 128k 996.9 56.7 78.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mqnzn9p7",
   "source_url": "https://omlx.ai/benchmarks/performance/mqnzn9p7",
   "tg_tokens": null,
   "tps": 56.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-mqnzn9p7-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.8 GB at this context; duplicate traces of the same session: qvp0nzjr, v0awgvhd; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 953.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 195k 953.3 48.6 79.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mqnzn9p7",
   "source_url": "https://omlx.ai/benchmarks/performance/mqnzn9p7",
   "tg_tokens": null,
   "tps": 48.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6lsfar9p-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: dqx2ekr3, dt0dk1am; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1266.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,266 76.3 70.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6lsfar9p",
   "source_url": "https://omlx.ai/benchmarks/performance/6lsfar9p",
   "tg_tokens": null,
   "tps": 76.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6lsfar9p-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: dqx2ekr3, dt0dk1am; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2092.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,092 70.5 71.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6lsfar9p",
   "source_url": "https://omlx.ai/benchmarks/performance/6lsfar9p",
   "tg_tokens": null,
   "tps": 70.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6lsfar9p-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.9 GB at this context; duplicate traces of the same session: dqx2ekr3, dt0dk1am; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1534.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 1,534 60.9 78.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6lsfar9p",
   "source_url": "https://omlx.ai/benchmarks/performance/6lsfar9p",
   "tg_tokens": null,
   "tps": 60.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ea5catdq-ctx1024",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 2x187098, nwjd17vy, rbbyk8hy, uq19x1jn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 899.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 899.5 31.9 16.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ea5catdq",
   "source_url": "https://omlx.ai/benchmarks/performance/ea5catdq",
   "tg_tokens": null,
   "tps": 31.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ea5catdq-ctx4096",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 2x187098, nwjd17vy, rbbyk8hy, uq19x1jn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 930.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 930.9 31.5 17.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ea5catdq",
   "source_url": "https://omlx.ai/benchmarks/performance/ea5catdq",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ea5catdq-ctx8192",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 2x187098, nwjd17vy, rbbyk8hy, uq19x1jn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 928.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 928.5 30.8 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ea5catdq",
   "source_url": "https://omlx.ai/benchmarks/performance/ea5catdq",
   "tg_tokens": null,
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ea5catdq-ctx16384",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: 2x187098, nwjd17vy, rbbyk8hy, uq19x1jn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 829.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 829.4 29.7 18.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ea5catdq",
   "source_url": "https://omlx.ai/benchmarks/performance/ea5catdq",
   "tg_tokens": null,
   "tps": 29.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ea5catdq-ctx32768",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 2x187098, nwjd17vy, rbbyk8hy, uq19x1jn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 771.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 771.7 28.4 19.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ea5catdq",
   "source_url": "https://omlx.ai/benchmarks/performance/ea5catdq",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-nguowvil-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: th2ld4u3; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 655.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 655.3 68.2 14.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nguowvil",
   "source_url": "https://omlx.ai/benchmarks/performance/nguowvil",
   "tg_tokens": null,
   "tps": 68.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-nguowvil-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: th2ld4u3; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 629.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 629.5 65.6 14.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nguowvil",
   "source_url": "https://omlx.ai/benchmarks/performance/nguowvil",
   "tg_tokens": null,
   "tps": 65.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-xauwi7z3-ctx1024",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: jsraj3lo, thaq8x0g, uce6jm7t, uvlm5snc; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 96.6,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 5bit; 1k 96.6 11.3 19.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xauwi7z3",
   "source_url": "https://omlx.ai/benchmarks/performance/xauwi7z3",
   "tg_tokens": null,
   "tps": 11.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-xauwi7z3-ctx4096",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: jsraj3lo, thaq8x0g, uce6jm7t, uvlm5snc; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 98.1,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 5bit; 4k 98.1 10.5 22.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xauwi7z3",
   "source_url": "https://omlx.ai/benchmarks/performance/xauwi7z3",
   "tg_tokens": null,
   "tps": 10.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-xauwi7z3-ctx8192",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: jsraj3lo, thaq8x0g, uce6jm7t, uvlm5snc; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 94.5,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 5bit; 8k 94.5 10.3 24.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xauwi7z3",
   "source_url": "https://omlx.ai/benchmarks/performance/xauwi7z3",
   "tg_tokens": null,
   "tps": 10.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-xauwi7z3-ctx16384",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: jsraj3lo, thaq8x0g, uce6jm7t, uvlm5snc; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 91.5,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 5bit; 16k 91.5 10.1 25.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xauwi7z3",
   "source_url": "https://omlx.ai/benchmarks/performance/xauwi7z3",
   "tg_tokens": null,
   "tps": 10.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-xauwi7z3-ctx32768",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 26.9 GB at this context; duplicate traces of the same session: jsraj3lo, thaq8x0g, uce6jm7t, uvlm5snc; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 83.5,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 5bit; 32k 83.5 9.4 26.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xauwi7z3",
   "source_url": "https://omlx.ai/benchmarks/performance/xauwi7z3",
   "tg_tokens": null,
   "tps": 9.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-7ia3rr6s-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.4 GB at this context; duplicate traces of the same session: 90ciq61d, pma3568d; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1619.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-MLX-oQ4-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,619 83.0 178.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ia3rr6s",
   "source_url": "https://omlx.ai/benchmarks/performance/7ia3rr6s",
   "tg_tokens": null,
   "tps": 83.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-7ia3rr6s-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.9 GB at this context; duplicate traces of the same session: 90ciq61d, pma3568d; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1622.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-MLX-oQ4-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,622 73.2 178.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ia3rr6s",
   "source_url": "https://omlx.ai/benchmarks/performance/7ia3rr6s",
   "tg_tokens": null,
   "tps": 73.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-7ia3rr6s-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.5 GB at this context; duplicate traces of the same session: 90ciq61d, pma3568d; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1594.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-MLX-oQ4-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,594 65.9 179.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ia3rr6s",
   "source_url": "https://omlx.ai/benchmarks/performance/7ia3rr6s",
   "tg_tokens": null,
   "tps": 65.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-zvgghd7k-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.6 GB at this context; duplicate traces of the same session: 5cuk82lr, 7200bxl3, atdphn9a, mi34778s, mqh5nf5s, srhrgd4u, vrmk7ka1; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 459.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 459.2 37.3 175.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zvgghd7k",
   "source_url": "https://omlx.ai/benchmarks/performance/zvgghd7k",
   "tg_tokens": null,
   "tps": 37.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-zvgghd7k-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.0 GB at this context; duplicate traces of the same session: 5cuk82lr, 7200bxl3, atdphn9a, mi34778s, mqh5nf5s, srhrgd4u, vrmk7ka1; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1221.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,221 41.6 177.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zvgghd7k",
   "source_url": "https://omlx.ai/benchmarks/performance/zvgghd7k",
   "tg_tokens": null,
   "tps": 41.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-zvgghd7k-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.9 GB at this context; duplicate traces of the same session: 5cuk82lr, 7200bxl3, atdphn9a, mi34778s, mqh5nf5s, srhrgd4u, vrmk7ka1; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1960.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 1,960 72.4 177.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zvgghd7k",
   "source_url": "https://omlx.ai/benchmarks/performance/zvgghd7k",
   "tg_tokens": null,
   "tps": 72.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-zvgghd7k-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.6 GB at this context; duplicate traces of the same session: 5cuk82lr, 7200bxl3, atdphn9a, mi34778s, mqh5nf5s, srhrgd4u, vrmk7ka1; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1971.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,971 99.1 178.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zvgghd7k",
   "source_url": "https://omlx.ai/benchmarks/performance/zvgghd7k",
   "tg_tokens": null,
   "tps": 99.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-zvgghd7k-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 179.9 GB at this context; duplicate traces of the same session: 5cuk82lr, 7200bxl3, atdphn9a, mi34778s, mqh5nf5s, srhrgd4u, vrmk7ka1; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1953.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,953 87.4 179.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zvgghd7k",
   "source_url": "https://omlx.ai/benchmarks/performance/zvgghd7k",
   "tg_tokens": null,
   "tps": 87.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-zvgghd7k-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 182.8 GB at this context; duplicate traces of the same session: 5cuk82lr, 7200bxl3, atdphn9a, mi34778s, mqh5nf5s, srhrgd4u, vrmk7ka1; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1925.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,925 60.5 182.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zvgghd7k",
   "source_url": "https://omlx.ai/benchmarks/performance/zvgghd7k",
   "tg_tokens": null,
   "tps": 60.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-zvgghd7k-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 188.6 GB at this context; duplicate traces of the same session: 5cuk82lr, 7200bxl3, atdphn9a, mi34778s, mqh5nf5s, srhrgd4u, vrmk7ka1; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1887.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 1,887 96.1 188.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zvgghd7k",
   "source_url": "https://omlx.ai/benchmarks/performance/zvgghd7k",
   "tg_tokens": null,
   "tps": 96.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-zvgghd7k-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 194.4 GB at this context; duplicate traces of the same session: 5cuk82lr, 7200bxl3, atdphn9a, mi34778s, mqh5nf5s, srhrgd4u, vrmk7ka1; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1835.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 1,835 42.6 194.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zvgghd7k",
   "source_url": "https://omlx.ai/benchmarks/performance/zvgghd7k",
   "tg_tokens": null,
   "tps": 42.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-1pqwilrr-ctx1024",
   "model": "Gemma-4-Dark-Thoughts-V2-31B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: aomydsfo, hu001q1z; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 360.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Gemma-4-Dark-Thoughts-V2-31B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 360.5 13.3 18.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1pqwilrr",
   "source_url": "https://omlx.ai/benchmarks/performance/1pqwilrr",
   "tg_tokens": null,
   "tps": 13.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-1pqwilrr-ctx8192",
   "model": "Gemma-4-Dark-Thoughts-V2-31B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: aomydsfo, hu001q1z; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 326.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Gemma-4-Dark-Thoughts-V2-31B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 326.2 12.4 19.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1pqwilrr",
   "source_url": "https://omlx.ai/benchmarks/performance/1pqwilrr",
   "tg_tokens": null,
   "tps": 12.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-1pqwilrr-ctx32768",
   "model": "Gemma-4-Dark-Thoughts-V2-31B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: aomydsfo, hu001q1z; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 289.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Gemma-4-Dark-Thoughts-V2-31B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 289.5 9.7 23.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1pqwilrr",
   "source_url": "https://omlx.ai/benchmarks/performance/1pqwilrr",
   "tg_tokens": null,
   "tps": 9.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-nxov87mp-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 3qchaybt, 5dzva0b3, bf3zlxa1, tyfv3dx0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1567.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,567 105.4 20.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nxov87mp",
   "source_url": "https://omlx.ai/benchmarks/performance/nxov87mp",
   "tg_tokens": null,
   "tps": 105.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-nxov87mp-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 3qchaybt, 5dzva0b3, bf3zlxa1, tyfv3dx0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1613.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,613 96.8 21.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nxov87mp",
   "source_url": "https://omlx.ai/benchmarks/performance/nxov87mp",
   "tg_tokens": null,
   "tps": 96.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-nxov87mp-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 3qchaybt, 5dzva0b3, bf3zlxa1, tyfv3dx0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1531.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,531 34.0 21.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nxov87mp",
   "source_url": "https://omlx.ai/benchmarks/performance/nxov87mp",
   "tg_tokens": null,
   "tps": 34.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Max",
   "id": "omlx-nxov87mp-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 3qchaybt, 5dzva0b3, bf3zlxa1, tyfv3dx0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1236.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,236 20.3 22.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nxov87mp",
   "source_url": "https://omlx.ai/benchmarks/performance/nxov87mp",
   "tg_tokens": null,
   "tps": 20.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Max",
   "id": "omlx-nxov87mp-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 3qchaybt, 5dzva0b3, bf3zlxa1, tyfv3dx0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 923.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 923.7 10.9 22.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nxov87mp",
   "source_url": "https://omlx.ai/benchmarks/performance/nxov87mp",
   "tg_tokens": null,
   "tps": 10.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-8t1kyovq-ctx1024",
   "model": "Qwen-AgentWorld-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: vekcwtod; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2271.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen-AgentWorld-35B-A3B-oQ4 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,271 80.0 20.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8t1kyovq",
   "source_url": "https://omlx.ai/benchmarks/performance/8t1kyovq",
   "tg_tokens": null,
   "tps": 80.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-8t1kyovq-ctx4096",
   "model": "Qwen-AgentWorld-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: vekcwtod; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3165.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen-AgentWorld-35B-A3B-oQ4 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,165 74.6 20.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8t1kyovq",
   "source_url": "https://omlx.ai/benchmarks/performance/8t1kyovq",
   "tg_tokens": null,
   "tps": 74.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-5gatp7xj-ctx1024",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 6fy1lkic; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2566.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,566 98.6 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5gatp7xj",
   "source_url": "https://omlx.ai/benchmarks/performance/5gatp7xj",
   "tg_tokens": null,
   "tps": 98.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-5gatp7xj-ctx4096",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 6fy1lkic; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3276.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,276 96.6 20.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5gatp7xj",
   "source_url": "https://omlx.ai/benchmarks/performance/5gatp7xj",
   "tg_tokens": null,
   "tps": 96.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-tyxl5aga-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 38.2 GB at this context; duplicate traces of the same session: 3cff09f8, 6f6bmtua, d0ouuf9h, nij7v773; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1188.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 1,188 106.4 38.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tyxl5aga",
   "source_url": "https://omlx.ai/benchmarks/performance/tyxl5aga",
   "tg_tokens": null,
   "tps": 106.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-tyxl5aga-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 40.3 GB at this context; duplicate traces of the same session: 3cff09f8, 6f6bmtua, d0ouuf9h, nij7v773; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2421.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 2,421 135.5 40.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tyxl5aga",
   "source_url": "https://omlx.ai/benchmarks/performance/tyxl5aga",
   "tg_tokens": null,
   "tps": 135.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-tyxl5aga-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 41.0 GB at this context; duplicate traces of the same session: 3cff09f8, 6f6bmtua, d0ouuf9h, nij7v773; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2278.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 8k 2,278 140.4 41.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tyxl5aga",
   "source_url": "https://omlx.ai/benchmarks/performance/tyxl5aga",
   "tg_tokens": null,
   "tps": 140.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-tyxl5aga-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 40.3 GB at this context; duplicate traces of the same session: 3cff09f8, 6f6bmtua, d0ouuf9h, nij7v773; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2039.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 16k 2,039 167.6 40.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tyxl5aga",
   "source_url": "https://omlx.ai/benchmarks/performance/tyxl5aga",
   "tg_tokens": null,
   "tps": 167.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-tyxl5aga-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 40.7 GB at this context; duplicate traces of the same session: 3cff09f8, 6f6bmtua, d0ouuf9h, nij7v773; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1704.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 32k 1,704 145.1 40.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tyxl5aga",
   "source_url": "https://omlx.ai/benchmarks/performance/tyxl5aga",
   "tg_tokens": null,
   "tps": 145.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-tyxl5aga-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 41.4 GB at this context; duplicate traces of the same session: 3cff09f8, 6f6bmtua, d0ouuf9h, nij7v773; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1236.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 64k 1,236 122.2 41.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tyxl5aga",
   "source_url": "https://omlx.ai/benchmarks/performance/tyxl5aga",
   "tg_tokens": null,
   "tps": 122.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-tyxl5aga-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 43.0 GB at this context; duplicate traces of the same session: 3cff09f8, 6f6bmtua, d0ouuf9h, nij7v773; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 793.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 128k 793.1 104.8 43.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tyxl5aga",
   "source_url": "https://omlx.ai/benchmarks/performance/tyxl5aga",
   "tg_tokens": null,
   "tps": 104.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-tyxl5aga-ctx199680",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 44.8 GB at this context; duplicate traces of the same session: 3cff09f8, 6f6bmtua, d0ouuf9h, nij7v773; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 575.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 195k 575.6 66.6 44.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tyxl5aga",
   "source_url": "https://omlx.ai/benchmarks/performance/tyxl5aga",
   "tg_tokens": null,
   "tps": 66.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-7ocpmoss-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 3912nula; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 101.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 1k 101.3 10.7 28.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ocpmoss",
   "source_url": "https://omlx.ai/benchmarks/performance/7ocpmoss",
   "tg_tokens": null,
   "tps": 10.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-7ocpmoss-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.2 GB at this context; duplicate traces of the same session: 3912nula; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 89.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 4k 89.0 9.9 32.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ocpmoss",
   "source_url": "https://omlx.ai/benchmarks/performance/7ocpmoss",
   "tg_tokens": null,
   "tps": 9.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ztodjjfw-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: nsqwawx7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 465.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 465.8 17.5 15.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ztodjjfw",
   "source_url": "https://omlx.ai/benchmarks/performance/ztodjjfw",
   "tg_tokens": null,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ztodjjfw-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: nsqwawx7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 473.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 473.3 17.2 17.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ztodjjfw",
   "source_url": "https://omlx.ai/benchmarks/performance/ztodjjfw",
   "tg_tokens": null,
   "tps": 17.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "outlier"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-9hfllj3j-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: e19xcjmp, ev40pe0m, ou7ierrm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 436.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 436.6 42.2 29.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9hfllj3j",
   "source_url": "https://omlx.ai/benchmarks/performance/9hfllj3j",
   "tg_tokens": null,
   "tps": 42.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-9hfllj3j-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 31.4 GB at this context; duplicate traces of the same session: e19xcjmp, ev40pe0m, ou7ierrm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 437.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 437.3 27.6 31.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9hfllj3j",
   "source_url": "https://omlx.ai/benchmarks/performance/9hfllj3j",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-9hfllj3j-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 32.6 GB at this context; duplicate traces of the same session: e19xcjmp, ev40pe0m, ou7ierrm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 470.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 470.3 26.8 32.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9hfllj3j",
   "source_url": "https://omlx.ai/benchmarks/performance/9hfllj3j",
   "tg_tokens": null,
   "tps": 26.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "outlier"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-9hfllj3j-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 34.3 GB at this context; duplicate traces of the same session: e19xcjmp, ev40pe0m, ou7ierrm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 451.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 16k 451.5 33.9 34.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9hfllj3j",
   "source_url": "https://omlx.ai/benchmarks/performance/9hfllj3j",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-rkwwyg99-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: qrvfhzss; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 439.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 439.5 35.2 29.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rkwwyg99",
   "source_url": "https://omlx.ai/benchmarks/performance/rkwwyg99",
   "tg_tokens": null,
   "tps": 35.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-rkwwyg99-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 31.4 GB at this context; duplicate traces of the same session: qrvfhzss; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 466.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 466.9 27.0 31.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rkwwyg99",
   "source_url": "https://omlx.ai/benchmarks/performance/rkwwyg99",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nktyb24e-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.9 GB at this context; duplicate traces of the same session: kfty29er; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1033.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 128k 1,033 57.3 74.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nktyb24e",
   "source_url": "https://omlx.ai/benchmarks/performance/nktyb24e",
   "tg_tokens": null,
   "tps": 57.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nktyb24e-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.9 GB at this context; duplicate traces of the same session: kfty29er; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 994.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 195k 994.7 49.3 78.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nktyb24e",
   "source_url": "https://omlx.ai/benchmarks/performance/nktyb24e",
   "tg_tokens": null,
   "tps": 49.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5l9k3jct-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.5 GB at this context; duplicate traces of the same session: zzspbf5a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1107.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,107 75.2 70.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5l9k3jct",
   "source_url": "https://omlx.ai/benchmarks/performance/5l9k3jct",
   "tg_tokens": null,
   "tps": 75.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5l9k3jct-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: zzspbf5a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1998.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,998 73.1 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5l9k3jct",
   "source_url": "https://omlx.ai/benchmarks/performance/5l9k3jct",
   "tg_tokens": null,
   "tps": 73.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-677smb2t-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 172.5 GB; TTFT 2269 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 451.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k tokens 451.3 PP tok/s 39.8 TG tok/s 2269 TTFT (ms) 172.5 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/677smb2t",
   "source_url": "https://omlx.ai/benchmarks/performance/677smb2t",
   "tg_tokens": null,
   "tps": 39.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ckuw0w9o-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.5 GB at this context; duplicate traces of the same session: t3j8bs52; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 566.6,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-MLX-oQ4-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 566.6 51.2 177.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ckuw0w9o",
   "source_url": "https://omlx.ai/benchmarks/performance/ckuw0w9o",
   "tg_tokens": null,
   "tps": 51.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ckuw0w9o-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: t3j8bs52; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1088.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-MLX-oQ4-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,088 58.1 178.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ckuw0w9o",
   "source_url": "https://omlx.ai/benchmarks/performance/ckuw0w9o",
   "tg_tokens": null,
   "tps": 58.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-37tr5q5h-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 195k context; peak memory 177.0 GB; TTFT 91071 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2196.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k tokens 2,196 PP tok/s 51.8 TG tok/s 91071 TTFT (ms) 177.0 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/37tr5q5h",
   "source_url": "https://omlx.ai/benchmarks/performance/37tr5q5h",
   "tg_tokens": null,
   "tps": 51.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-g0hgavc9-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; Jundot build; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: mhp7stnl; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3278.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,278 83.2 20.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g0hgavc9",
   "source_url": "https://omlx.ai/benchmarks/performance/g0hgavc9",
   "tg_tokens": null,
   "tps": 83.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-g0hgavc9-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; Jundot build; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: mhp7stnl; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2556.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 2,556 68.8 21.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g0hgavc9",
   "source_url": "https://omlx.ai/benchmarks/performance/g0hgavc9",
   "tg_tokens": null,
   "tps": 68.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-r3t7dg5p-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: 2r06q2td, 92wt7c98, be5730jp, gbl4z0om; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1529.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,529 69.2 19.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r3t7dg5p",
   "source_url": "https://omlx.ai/benchmarks/performance/r3t7dg5p",
   "tg_tokens": null,
   "tps": 69.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-r3t7dg5p-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: 2r06q2td, 92wt7c98, be5730jp, gbl4z0om; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1609.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,609 103.9 20.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r3t7dg5p",
   "source_url": "https://omlx.ai/benchmarks/performance/r3t7dg5p",
   "tg_tokens": null,
   "tps": 103.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Max",
   "id": "omlx-r3t7dg5p-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 2r06q2td, 92wt7c98, be5730jp, gbl4z0om; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1501.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,501 19.7 21.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r3t7dg5p",
   "source_url": "https://omlx.ai/benchmarks/performance/r3t7dg5p",
   "tg_tokens": null,
   "tps": 19.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Max",
   "id": "omlx-r3t7dg5p-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 2r06q2td, 92wt7c98, be5730jp, gbl4z0om; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1159.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,159 14.2 21.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r3t7dg5p",
   "source_url": "https://omlx.ai/benchmarks/performance/r3t7dg5p",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Max",
   "id": "omlx-r3t7dg5p-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: 2r06q2td, 92wt7c98, be5730jp, gbl4z0om; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 867.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 867.2 12.5 22.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r3t7dg5p",
   "source_url": "https://omlx.ai/benchmarks/performance/r3t7dg5p",
   "tg_tokens": null,
   "tps": 12.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-tm42fcf5-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "REAP-19B build; oMLX community benchmark trace, single run at 64k context; peak memory 15.7 GB; TTFT 189892 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 345.1,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit-REAP-19B on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 64k tokens 345.1 PP tok/s 46.5 TG tok/s 189892 TTFT (ms) 15.7 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tm42fcf5",
   "source_url": "https://omlx.ai/benchmarks/performance/tm42fcf5",
   "tg_tokens": null,
   "tps": 46.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-cc2fz0fi-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: dw8vccdl; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 466.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 466.5 50.6 5.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cc2fz0fi",
   "source_url": "https://omlx.ai/benchmarks/performance/cc2fz0fi",
   "tg_tokens": null,
   "tps": 50.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-cc2fz0fi-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: dw8vccdl; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 481.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 481.5 49.0 6.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cc2fz0fi",
   "source_url": "https://omlx.ai/benchmarks/performance/cc2fz0fi",
   "tg_tokens": null,
   "tps": 49.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-meoof2h9-ctx1024",
   "model": "Qwen3.6-27B",
   "notes": "UD (Unsloth Dynamic) build; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: ny6c1m8z; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 467.2,
   "provenance": "community",
   "quant": "Q4_K_XL",
   "quote": "Qwen3.6-27B-UD-Q4_K_XL-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 467.2 13.4 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/meoof2h9",
   "source_url": "https://omlx.ai/benchmarks/performance/meoof2h9",
   "tg_tokens": null,
   "tps": 13.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-meoof2h9-ctx4096",
   "model": "Qwen3.6-27B",
   "notes": "UD (Unsloth Dynamic) build; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: ny6c1m8z; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 505.6,
   "provenance": "community",
   "quant": "Q4_K_XL",
   "quote": "Qwen3.6-27B-UD-Q4_K_XL-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 505.6 12.7 22.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/meoof2h9",
   "source_url": "https://omlx.ai/benchmarks/performance/meoof2h9",
   "tg_tokens": null,
   "tps": 12.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0631hwcn-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 218 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4694.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 4,694 PP tok/s 142.9 TG tok/s 218 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0631hwcn",
   "source_url": "https://omlx.ai/benchmarks/performance/0631hwcn",
   "tg_tokens": null,
   "tps": 142.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-r6ms7xhy-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 225 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4546.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 4,546 PP tok/s 138.3 TG tok/s 225 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r6ms7xhy",
   "source_url": "https://omlx.ai/benchmarks/performance/r6ms7xhy",
   "tg_tokens": null,
   "tps": 138.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-mfu5mt77-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 22.9 GB; TTFT 403165 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 81.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 32k tokens 81.3 PP tok/s 10.5 TG tok/s 403165 TTFT (ms) 22.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mfu5mt77",
   "source_url": "https://omlx.ai/benchmarks/performance/mfu5mt77",
   "tg_tokens": null,
   "tps": 10.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-4dgkazes-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 16.0 GB; TTFT 11160 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 91.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k tokens 91.8 PP tok/s 12.6 TG tok/s 11160 TTFT (ms) 16.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4dgkazes",
   "source_url": "https://omlx.ai/benchmarks/performance/4dgkazes",
   "tg_tokens": null,
   "tps": 12.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-860d2jpz-ctx1024",
   "model": "Qwen3.8-27B-JANG_4D-CRACK",
   "notes": "community fine-tune; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 8e7ocmc6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 579.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-JANG_4D-CRACK on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 579.5 25.2 19.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/860d2jpz",
   "source_url": "https://omlx.ai/benchmarks/performance/860d2jpz",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-860d2jpz-ctx4096",
   "model": "Qwen3.8-27B-JANG_4D-CRACK",
   "notes": "community fine-tune; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 8e7ocmc6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 627.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-JANG_4D-CRACK on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 627.4 23.3 22.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/860d2jpz",
   "source_url": "https://omlx.ai/benchmarks/performance/860d2jpz",
   "tg_tokens": null,
   "tps": 23.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-muf1j23e-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: 3uzys69f; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 107.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 1k 107.2 14.8 15.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/muf1j23e",
   "source_url": "https://omlx.ai/benchmarks/performance/muf1j23e",
   "tg_tokens": null,
   "tps": 14.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-muf1j23e-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: 3uzys69f; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 105.4 13.3 19.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/muf1j23e",
   "source_url": "https://omlx.ai/benchmarks/performance/muf1j23e",
   "tg_tokens": null,
   "tps": 13.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-2l710ctc-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: c0ccdg7y; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 200.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 200.5 32.0 19.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2l710ctc",
   "source_url": "https://omlx.ai/benchmarks/performance/2l710ctc",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-2l710ctc-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: c0ccdg7y; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 187.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 16k 187.0 29.6 21.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2l710ctc",
   "source_url": "https://omlx.ai/benchmarks/performance/2l710ctc",
   "tg_tokens": null,
   "tps": 29.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "outlier"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ymefzami-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: b7n4raxe; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 429.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 429.0 45.6 29.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ymefzami",
   "source_url": "https://omlx.ai/benchmarks/performance/ymefzami",
   "tg_tokens": null,
   "tps": 45.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ymefzami-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 31.4 GB at this context; duplicate traces of the same session: b7n4raxe; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 453.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 453.7 26.8 31.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ymefzami",
   "source_url": "https://omlx.ai/benchmarks/performance/ymefzami",
   "tg_tokens": null,
   "tps": 26.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-9scoescv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 30.9 GB at this context; duplicate traces of the same session: 4pjh0s6q; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 469.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 469.7 20.7 30.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9scoescv",
   "source_url": "https://omlx.ai/benchmarks/performance/9scoescv",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-9scoescv-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 32.3 GB at this context; duplicate traces of the same session: 4pjh0s6q; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 444.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 16k 444.9 28.8 32.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9scoescv",
   "source_url": "https://omlx.ai/benchmarks/performance/9scoescv",
   "tg_tokens": null,
   "tps": 28.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-v9m63nln-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Heretic",
   "notes": "fp16 MTP build; oMLX community benchmark trace, single run at 4k context; peak memory 38.8 GB; TTFT 27849 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 147.1,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-27B-TURBO-Fable-Heretic-oQ8-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k tokens 147.1 PP tok/s 23.1 TG tok/s 27849 TTFT (ms) 38.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v9m63nln",
   "source_url": "https://omlx.ai/benchmarks/performance/v9m63nln",
   "tg_tokens": null,
   "tps": 23.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-451prfrs-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: pyw2gi4p, zshlcm3e; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 124.7,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 124.7 28.3 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/451prfrs",
   "source_url": "https://omlx.ai/benchmarks/performance/451prfrs",
   "tg_tokens": null,
   "tps": 28.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-451prfrs-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: pyw2gi4p, zshlcm3e; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 150.5,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 150.5 38.0 23.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/451prfrs",
   "source_url": "https://omlx.ai/benchmarks/performance/451prfrs",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-451prfrs-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: pyw2gi4p, zshlcm3e; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 140.1,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 140.1 20.9 24.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/451prfrs",
   "source_url": "https://omlx.ai/benchmarks/performance/451prfrs",
   "tg_tokens": null,
   "tps": 20.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-yiofj6pa-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace, single run at 4k context; peak memory 37.9 GB; TTFT 28312 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 144.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp-fp16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k tokens 144.7 PP tok/s 22.8 TG tok/s 28312 TTFT (ms) 37.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yiofj6pa",
   "source_url": "https://omlx.ai/benchmarks/performance/yiofj6pa",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-2n87j4xo-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace, single run at 8k context; peak memory 39.1 GB; TTFT 59277 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 138.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp-fp16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 8k tokens 138.2 PP tok/s 16.2 TG tok/s 59277 TTFT (ms) 39.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2n87j4xo",
   "source_url": "https://omlx.ai/benchmarks/performance/2n87j4xo",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-v61vfcqj-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 52.0 GB at this context; duplicate traces of the same session: q182wxng, qk18nx82; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 799.8,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 2bit; 16k 799.8 46.6 52.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v61vfcqj",
   "source_url": "https://omlx.ai/benchmarks/performance/v61vfcqj",
   "tg_tokens": null,
   "tps": 46.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-v61vfcqj-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 52.6 GB at this context; duplicate traces of the same session: q182wxng, qk18nx82; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1024.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 2bit; 32k 1,024 43.0 52.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v61vfcqj",
   "source_url": "https://omlx.ai/benchmarks/performance/v61vfcqj",
   "tg_tokens": null,
   "tps": 43.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-v61vfcqj-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 54.6 GB at this context; duplicate traces of the same session: q182wxng, qk18nx82; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1011.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 2bit; 64k 1,011 46.7 54.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v61vfcqj",
   "source_url": "https://omlx.ai/benchmarks/performance/v61vfcqj",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-v61vfcqj-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 59.3 GB at this context; duplicate traces of the same session: q182wxng, qk18nx82; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 982.5,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 2bit; 128k 982.5 52.4 59.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v61vfcqj",
   "source_url": "https://omlx.ai/benchmarks/performance/v61vfcqj",
   "tg_tokens": null,
   "tps": 52.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mug6c5w6-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.5 GB at this context; duplicate traces of the same session: b72k0oo6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1022.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,022 81.7 70.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mug6c5w6",
   "source_url": "https://omlx.ai/benchmarks/performance/mug6c5w6",
   "tg_tokens": null,
   "tps": 81.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mug6c5w6-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: b72k0oo6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1977.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,977 80.5 70.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mug6c5w6",
   "source_url": "https://omlx.ai/benchmarks/performance/mug6c5w6",
   "tg_tokens": null,
   "tps": 80.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-87lb0kai-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.5 GB at this context; duplicate traces of the same session: 1z5kqwss; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1109.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,109 85.2 70.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/87lb0kai",
   "source_url": "https://omlx.ai/benchmarks/performance/87lb0kai",
   "tg_tokens": null,
   "tps": 85.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-87lb0kai-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 1z5kqwss; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1996.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,996 82.8 70.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/87lb0kai",
   "source_url": "https://omlx.ai/benchmarks/performance/87lb0kai",
   "tg_tokens": null,
   "tps": 82.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-x00ynand-ctx4096",
   "model": "gemma-4-12b-coder-fable5-composer2.5",
   "notes": "community coder fine-tune; oMLX community benchmark trace, single run at 4k context; peak memory 7.6 GB; TTFT 29847 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 137.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12b-coder-fable5-composer2.5-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k tokens 137.2 PP tok/s 13.6 TG tok/s 29847 TTFT (ms) 7.6 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x00ynand",
   "source_url": "https://omlx.ai/benchmarks/performance/x00ynand",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-j7a785iz-ctx8192",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: bi8vehzw; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 369.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 369.9 12.9 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j7a785iz",
   "source_url": "https://omlx.ai/benchmarks/performance/j7a785iz",
   "tg_tokens": null,
   "tps": 12.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-j7a785iz-ctx32768",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: bi8vehzw; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 289.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 289.4 9.7 23.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j7a785iz",
   "source_url": "https://omlx.ai/benchmarks/performance/j7a785iz",
   "tg_tokens": null,
   "tps": 9.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-mwbrraay-ctx4096",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 3.0 GB; TTFT 1745 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2347.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e2b-it-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k tokens 2,347 PP tok/s 65.3 TG tok/s 1745 TTFT (ms) 3.0 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mwbrraay",
   "source_url": "https://omlx.ai/benchmarks/performance/mwbrraay",
   "tg_tokens": null,
   "tps": 65.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ymofkx31-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 195k context; peak memory 194.4 GB; TTFT 91225 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2192.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k tokens 2,192 PP tok/s 36.5 TG tok/s 91225 TTFT (ms) 194.4 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ymofkx31",
   "source_url": "https://omlx.ai/benchmarks/performance/ymofkx31",
   "tg_tokens": null,
   "tps": 36.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-8r7tmoqv-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: 8gkedjug, tmb1ew6o; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 688.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 688.9 52.0 19.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8r7tmoqv",
   "source_url": "https://omlx.ai/benchmarks/performance/8r7tmoqv",
   "tg_tokens": null,
   "tps": 52.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-8r7tmoqv-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: 8gkedjug, tmb1ew6o; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 717.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 717.1 44.5 20.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8r7tmoqv",
   "source_url": "https://omlx.ai/benchmarks/performance/8r7tmoqv",
   "tg_tokens": null,
   "tps": 44.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-8r7tmoqv-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: 8gkedjug, tmb1ew6o; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 431.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 32k 431.0 12.3 22.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8r7tmoqv",
   "source_url": "https://omlx.ai/benchmarks/performance/8r7tmoqv",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-cfxsmc45-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: jndx0rza; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 765.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 1k 765.6 93.6 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cfxsmc45",
   "source_url": "https://omlx.ai/benchmarks/performance/cfxsmc45",
   "tg_tokens": null,
   "tps": 93.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-cfxsmc45-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: jndx0rza; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 781.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 781.3 90.9 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cfxsmc45",
   "source_url": "https://omlx.ai/benchmarks/performance/cfxsmc45",
   "tg_tokens": null,
   "tps": 90.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-4e18uloj-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: j9m1856p; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 725.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 1k 725.8 95.8 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4e18uloj",
   "source_url": "https://omlx.ai/benchmarks/performance/4e18uloj",
   "tg_tokens": null,
   "tps": 95.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-4e18uloj-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: j9m1856p; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 779.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 779.9 90.8 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4e18uloj",
   "source_url": "https://omlx.ai/benchmarks/performance/4e18uloj",
   "tg_tokens": null,
   "tps": 90.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-le4eat7z-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 9vpwfuiu; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2127.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 2,127 74.8 21.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/le4eat7z",
   "source_url": "https://omlx.ai/benchmarks/performance/le4eat7z",
   "tg_tokens": null,
   "tps": 74.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-le4eat7z-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: 9vpwfuiu; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2648.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 2,648 71.6 22.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/le4eat7z",
   "source_url": "https://omlx.ai/benchmarks/performance/le4eat7z",
   "tg_tokens": null,
   "tps": 71.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-nsqvx17s-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 21.9 GB; TTFT 474 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2162.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k tokens 2,162 PP tok/s 76.0 TG tok/s 474 TTFT (ms) 21.9 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nsqvx17s",
   "source_url": "https://omlx.ai/benchmarks/performance/nsqvx17s",
   "tg_tokens": null,
   "tps": 76.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-09ryl86a-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 21.8 GB; TTFT 2309 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 443.5,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k tokens 443.5 PP tok/s 73.4 TG tok/s 2309 TTFT (ms) 21.8 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/09ryl86a",
   "source_url": "https://omlx.ai/benchmarks/performance/09ryl86a",
   "tg_tokens": null,
   "tps": 73.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-8izcmbk4-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: fmmrduao; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1163.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "Qwen3.6-35B-A3B-oQ5-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 5bit; 16k 1,163 63.9 27.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8izcmbk4",
   "source_url": "https://omlx.ai/benchmarks/performance/8izcmbk4",
   "tg_tokens": null,
   "tps": 63.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-8izcmbk4-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 31.1 GB at this context; duplicate traces of the same session: fmmrduao; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 433.6,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "Qwen3.6-35B-A3B-oQ5-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 5bit; 128k 433.6 45.0 31.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8izcmbk4",
   "source_url": "https://omlx.ai/benchmarks/performance/8izcmbk4",
   "tg_tokens": null,
   "tps": 45.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-wudfs3p0-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 28.9 GB; TTFT 2458777 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 53.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 128k tokens 53.3 PP tok/s 7.3 TG tok/s 2458777 TTFT (ms) 28.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wudfs3p0",
   "source_url": "https://omlx.ai/benchmarks/performance/wudfs3p0",
   "tg_tokens": null,
   "tps": 7.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-hfkw1gwc-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: uo49gsqq; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 111.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 111.7 19.2 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hfkw1gwc",
   "source_url": "https://omlx.ai/benchmarks/performance/hfkw1gwc",
   "tg_tokens": null,
   "tps": 19.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-hfkw1gwc-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: uo49gsqq; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 104.1 18.6 19.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hfkw1gwc",
   "source_url": "https://omlx.ai/benchmarks/performance/hfkw1gwc",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-9ihlvto9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 2i0zpw1r; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1198.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8Bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 1k 1,198 33.2 28.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ihlvto9",
   "source_url": "https://omlx.ai/benchmarks/performance/9ihlvto9",
   "tg_tokens": null,
   "tps": 33.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-9ihlvto9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: 2i0zpw1r; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1461.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8Bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 4k 1,461 33.0 29.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ihlvto9",
   "source_url": "https://omlx.ai/benchmarks/performance/9ihlvto9",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-d6jtrw34-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: f4fm6v7i; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1203.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8Bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 1k 1,203 57.4 33.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d6jtrw34",
   "source_url": "https://omlx.ai/benchmarks/performance/d6jtrw34",
   "tg_tokens": null,
   "tps": 57.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-d6jtrw34-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: f4fm6v7i; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1443.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8Bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 4k 1,443 54.2 34.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d6jtrw34",
   "source_url": "https://omlx.ai/benchmarks/performance/d6jtrw34",
   "tg_tokens": null,
   "tps": 54.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-z6z5jfwm-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: rtgh0f29; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 107.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 1k 107.9 13.0 18.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z6z5jfwm",
   "source_url": "https://omlx.ai/benchmarks/performance/z6z5jfwm",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-z6z5jfwm-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: rtgh0f29; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 100.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 100.6 11.6 22.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z6z5jfwm",
   "source_url": "https://omlx.ai/benchmarks/performance/z6z5jfwm",
   "tg_tokens": null,
   "tps": 11.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-xyhoglu3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: pjat2dqi; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 1k 105.2 19.9 18.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xyhoglu3",
   "source_url": "https://omlx.ai/benchmarks/performance/xyhoglu3",
   "tg_tokens": null,
   "tps": 19.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-xyhoglu3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: pjat2dqi; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 101.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 101.6 16.7 22.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xyhoglu3",
   "source_url": "https://omlx.ai/benchmarks/performance/xyhoglu3",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-rfszhqm7-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 1k 106.1 17.0 18.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rfszhqm7",
   "source_url": "https://omlx.ai/benchmarks/performance/rfszhqm7",
   "tg_tokens": null,
   "tps": 17.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-rfszhqm7-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 100.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 100.2 17.1 22.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rfszhqm7",
   "source_url": "https://omlx.ai/benchmarks/performance/rfszhqm7",
   "tg_tokens": null,
   "tps": 17.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-mv9ph3hm-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 186fabfi, 8j02oc4k, o4bins6c, vccj8dqz; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 101.2,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 101.2 12.8 19.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mv9ph3hm",
   "source_url": "https://omlx.ai/benchmarks/performance/mv9ph3hm",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-mv9ph3hm-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: 186fabfi, 8j02oc4k, o4bins6c, vccj8dqz; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 96.1,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 96.1 12.4 23.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mv9ph3hm",
   "source_url": "https://omlx.ai/benchmarks/performance/mv9ph3hm",
   "tg_tokens": null,
   "tps": 12.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-mv9ph3hm-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: 186fabfi, 8j02oc4k, o4bins6c, vccj8dqz; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 93.9,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 93.9 12.1 24.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mv9ph3hm",
   "source_url": "https://omlx.ai/benchmarks/performance/mv9ph3hm",
   "tg_tokens": null,
   "tps": 12.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-mv9ph3hm-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: 186fabfi, 8j02oc4k, o4bins6c, vccj8dqz; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 90.3,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 90.3 11.6 26.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mv9ph3hm",
   "source_url": "https://omlx.ai/benchmarks/performance/mv9ph3hm",
   "tg_tokens": null,
   "tps": 11.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-mv9ph3hm-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: 186fabfi, 8j02oc4k, o4bins6c, vccj8dqz; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 83.7,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 32k 83.7 10.7 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mv9ph3hm",
   "source_url": "https://omlx.ai/benchmarks/performance/mv9ph3hm",
   "tg_tokens": null,
   "tps": 10.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-w1ai0swv-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: smzhgktp; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 455.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 455.4 109.3 30.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w1ai0swv",
   "source_url": "https://omlx.ai/benchmarks/performance/w1ai0swv",
   "tg_tokens": null,
   "tps": 109.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-w1ai0swv-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.4 GB at this context; duplicate traces of the same session: smzhgktp; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 407.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 407.2 91.3 32.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w1ai0swv",
   "source_url": "https://omlx.ai/benchmarks/performance/w1ai0swv",
   "tg_tokens": null,
   "tps": 91.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-u1b0wa4d-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: qvs3c03n; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 673.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 673.7 31.0 19.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u1b0wa4d",
   "source_url": "https://omlx.ai/benchmarks/performance/u1b0wa4d",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-u1b0wa4d-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: qvs3c03n; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 681.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 681.0 47.0 19.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u1b0wa4d",
   "source_url": "https://omlx.ai/benchmarks/performance/u1b0wa4d",
   "tg_tokens": null,
   "tps": 47.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-sewkfgjr-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "SAGE-PATHS MTP build; oMLX community benchmark trace; peak memory 30.6 GB at this context; duplicate traces of the same session: glv2irj2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 263.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-SAGE-PATHS-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 64k 263.0 34.0 30.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sewkfgjr",
   "source_url": "https://omlx.ai/benchmarks/performance/sewkfgjr",
   "tg_tokens": null,
   "tps": 34.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-sewkfgjr-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "SAGE-PATHS MTP build; oMLX community benchmark trace; peak memory 35.1 GB at this context; duplicate traces of the same session: glv2irj2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 210.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-SAGE-PATHS-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 128k 210.2 32.2 35.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sewkfgjr",
   "source_url": "https://omlx.ai/benchmarks/performance/sewkfgjr",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-stc7spyy-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 52.0 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 799.8,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 2bit; 16k 799.8 46.6 52.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/stc7spyy",
   "source_url": "https://omlx.ai/benchmarks/performance/stc7spyy",
   "tg_tokens": null,
   "tps": 46.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-stc7spyy-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 52.6 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1024.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 2bit; 32k 1,024 43.0 52.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/stc7spyy",
   "source_url": "https://omlx.ai/benchmarks/performance/stc7spyy",
   "tg_tokens": null,
   "tps": 43.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-stc7spyy-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 54.6 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1011.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 2bit; 64k 1,011 46.7 54.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/stc7spyy",
   "source_url": "https://omlx.ai/benchmarks/performance/stc7spyy",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-stc7spyy-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 59.3 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 982.5,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 2bit; 128k 982.5 52.4 59.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/stc7spyy",
   "source_url": "https://omlx.ai/benchmarks/performance/stc7spyy",
   "tg_tokens": null,
   "tps": 52.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-s8wwnrhg-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 74.9 GB; TTFT 56759 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2309.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k tokens 2,309 PP tok/s 61.1 TG tok/s 56759 TTFT (ms) 74.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s8wwnrhg",
   "source_url": "https://omlx.ai/benchmarks/performance/s8wwnrhg",
   "tg_tokens": null,
   "tps": 61.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pkriy70b-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 195k context; peak memory 78.4 GB; TTFT 93726 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2134.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k tokens 2,134 PP tok/s 60.9 TG tok/s 93726 TTFT (ms) 78.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pkriy70b",
   "source_url": "https://omlx.ai/benchmarks/performance/pkriy70b",
   "tg_tokens": null,
   "tps": 60.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-lewjve4o-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: 672nk30i, gtnc87qu; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4140.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 4,140 99.9 101.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lewjve4o",
   "source_url": "https://omlx.ai/benchmarks/performance/lewjve4o",
   "tg_tokens": null,
   "tps": 99.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-lewjve4o-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.1 GB at this context; duplicate traces of the same session: 672nk30i, gtnc87qu; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4682.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 4,682 106.7 102.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lewjve4o",
   "source_url": "https://omlx.ai/benchmarks/performance/lewjve4o",
   "tg_tokens": null,
   "tps": 106.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-lewjve4o-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.6 GB at this context; duplicate traces of the same session: 672nk30i, gtnc87qu; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4720.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 4,720 123.7 102.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lewjve4o",
   "source_url": "https://omlx.ai/benchmarks/performance/lewjve4o",
   "tg_tokens": null,
   "tps": 123.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-vfhuxeqp-ctx1024",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "MTP speculative decoding enabled; Shiftedx build; oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: ff1aigpp; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1274.0,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Shiftedx--Qwen3.8-27B-Abliterated-MLX-MXFP4-MTP on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,274 87.7 16.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vfhuxeqp",
   "source_url": "https://omlx.ai/benchmarks/performance/vfhuxeqp",
   "tg_tokens": null,
   "tps": 87.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-vfhuxeqp-ctx4096",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "MTP speculative decoding enabled; Shiftedx build; oMLX community benchmark trace; peak memory 17.9 GB at this context; duplicate traces of the same session: ff1aigpp; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1792.0,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Shiftedx--Qwen3.8-27B-Abliterated-MLX-MXFP4-MTP on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 1,792 82.0 17.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vfhuxeqp",
   "source_url": "https://omlx.ai/benchmarks/performance/vfhuxeqp",
   "tg_tokens": null,
   "tps": 82.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-hiumxjhx-ctx1024",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: v54a039q; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 657.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 1k 657.4 58.1 25.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hiumxjhx",
   "source_url": "https://omlx.ai/benchmarks/performance/hiumxjhx",
   "tg_tokens": null,
   "tps": 58.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-hiumxjhx-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: v54a039q; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 613.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 4k 613.5 50.0 26.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hiumxjhx",
   "source_url": "https://omlx.ai/benchmarks/performance/hiumxjhx",
   "tg_tokens": null,
   "tps": 50.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ahswjkmp-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "AXQ build, MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.0 GB at this context; duplicate traces of the same session: 38jae1ra; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2106.0,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "AX-Tiel-Coder-35B-A3B-MLX-AXQ-MXFP8-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 2,106 103.0 35.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ahswjkmp",
   "source_url": "https://omlx.ai/benchmarks/performance/ahswjkmp",
   "tg_tokens": null,
   "tps": 103.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ahswjkmp-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "AXQ build, MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.7 GB at this context; duplicate traces of the same session: 38jae1ra; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4478.0,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "AX-Tiel-Coder-35B-A3B-MLX-AXQ-MXFP8-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 4,478 101.4 35.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ahswjkmp",
   "source_url": "https://omlx.ai/benchmarks/performance/ahswjkmp",
   "tg_tokens": null,
   "tps": 101.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-208vhlv0-ctx1024",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: 7q3cz7s5, bzhx9f6r, mwvzrfbq, p1kk3i37, pd2t03zo, v4asp3zk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1857.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Mellum2-12B-A2.5B-Instruct-mxfp4 on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,857 135.8 6.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/208vhlv0",
   "source_url": "https://omlx.ai/benchmarks/performance/208vhlv0",
   "tg_tokens": null,
   "tps": 135.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-208vhlv0-ctx4096",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: 7q3cz7s5, bzhx9f6r, mwvzrfbq, p1kk3i37, pd2t03zo, v4asp3zk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1737.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Mellum2-12B-A2.5B-Instruct-mxfp4 on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,737 134.4 6.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/208vhlv0",
   "source_url": "https://omlx.ai/benchmarks/performance/208vhlv0",
   "tg_tokens": null,
   "tps": 134.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-208vhlv0-ctx8192",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: 7q3cz7s5, bzhx9f6r, mwvzrfbq, p1kk3i37, pd2t03zo, v4asp3zk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1870.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Mellum2-12B-A2.5B-Instruct-mxfp4 on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,870 124.5 6.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/208vhlv0",
   "source_url": "https://omlx.ai/benchmarks/performance/208vhlv0",
   "tg_tokens": null,
   "tps": 124.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-208vhlv0-ctx16384",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: 7q3cz7s5, bzhx9f6r, mwvzrfbq, p1kk3i37, pd2t03zo, v4asp3zk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1800.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Mellum2-12B-A2.5B-Instruct-mxfp4 on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,800 106.7 6.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/208vhlv0",
   "source_url": "https://omlx.ai/benchmarks/performance/208vhlv0",
   "tg_tokens": null,
   "tps": 106.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-208vhlv0-ctx32768",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 7.2 GB at this context; duplicate traces of the same session: 7q3cz7s5, bzhx9f6r, mwvzrfbq, p1kk3i37, pd2t03zo, v4asp3zk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1289.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Mellum2-12B-A2.5B-Instruct-mxfp4 on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,289 84.0 7.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/208vhlv0",
   "source_url": "https://omlx.ai/benchmarks/performance/208vhlv0",
   "tg_tokens": null,
   "tps": 84.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-208vhlv0-ctx65536",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 7.8 GB at this context; duplicate traces of the same session: 7q3cz7s5, bzhx9f6r, mwvzrfbq, p1kk3i37, pd2t03zo, v4asp3zk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 990.6,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Mellum2-12B-A2.5B-Instruct-mxfp4 on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 990.6 50.6 7.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/208vhlv0",
   "source_url": "https://omlx.ai/benchmarks/performance/208vhlv0",
   "tg_tokens": null,
   "tps": 50.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-208vhlv0-ctx131072",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 9.5 GB at this context; duplicate traces of the same session: 7q3cz7s5, bzhx9f6r, mwvzrfbq, p1kk3i37, pd2t03zo, v4asp3zk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 804.8,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Mellum2-12B-A2.5B-Instruct-mxfp4 on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 804.8 34.4 9.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/208vhlv0",
   "source_url": "https://omlx.ai/benchmarks/performance/208vhlv0",
   "tg_tokens": null,
   "tps": 34.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-n36x82bn-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: 2gwyswm7, 8d1mbx9i; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 663.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 663.6 58.6 19.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n36x82bn",
   "source_url": "https://omlx.ai/benchmarks/performance/n36x82bn",
   "tg_tokens": null,
   "tps": 58.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-n36x82bn-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: 2gwyswm7, 8d1mbx9i; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 719.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 719.3 50.8 20.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n36x82bn",
   "source_url": "https://omlx.ai/benchmarks/performance/n36x82bn",
   "tg_tokens": null,
   "tps": 50.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-n36x82bn-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 2gwyswm7, 8d1mbx9i; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 560.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 560.5 14.8 21.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n36x82bn",
   "source_url": "https://omlx.ai/benchmarks/performance/n36x82bn",
   "tg_tokens": null,
   "tps": 14.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-hkh0qch9-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 36.3 GB at this context; duplicate traces of the same session: c4ua4gt9, rbkh77rk; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 666.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 1k 666.6 41.5 36.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hkh0qch9",
   "source_url": "https://omlx.ai/benchmarks/performance/hkh0qch9",
   "tg_tokens": null,
   "tps": 41.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-hkh0qch9-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: c4ua4gt9, rbkh77rk; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 732.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k 732.6 47.1 37.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hkh0qch9",
   "source_url": "https://omlx.ai/benchmarks/performance/hkh0qch9",
   "tg_tokens": null,
   "tps": 47.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-hkh0qch9-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 38.0 GB at this context; duplicate traces of the same session: c4ua4gt9, rbkh77rk; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 571.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 16k 571.9 12.4 38.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hkh0qch9",
   "source_url": "https://omlx.ai/benchmarks/performance/hkh0qch9",
   "tg_tokens": null,
   "tps": 12.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-upmvcmt4-ctx16384",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "MLX; oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: 8tjhfwte; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3823.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 3,823 86.1 18.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/upmvcmt4",
   "source_url": "https://omlx.ai/benchmarks/performance/upmvcmt4",
   "tg_tokens": null,
   "tps": 86.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-upmvcmt4-ctx65536",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "MLX; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: 8tjhfwte; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1242.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 64k 1,242 43.9 22.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/upmvcmt4",
   "source_url": "https://omlx.ai/benchmarks/performance/upmvcmt4",
   "tg_tokens": null,
   "tps": 43.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-uzjx0611-ctx4096",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: zn9d1007; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 202.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-27B-MLX-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 202.4 32.1 19.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uzjx0611",
   "source_url": "https://omlx.ai/benchmarks/performance/uzjx0611",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-uzjx0611-ctx16384",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: zn9d1007; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 187.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-27B-MLX-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 16k 187.7 29.6 21.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uzjx0611",
   "source_url": "https://omlx.ai/benchmarks/performance/uzjx0611",
   "tg_tokens": null,
   "tps": 29.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-5dj91x21-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: b54vr2nn; config=19c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 465.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M2 Pro (19c); M2 Pro (19c) \u00b7 32 GB \u00b7 4bit; 1k 465.1 65.0 19.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5dj91x21",
   "source_url": "https://omlx.ai/benchmarks/performance/5dj91x21",
   "tg_tokens": null,
   "tps": 65.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-5dj91x21-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: b54vr2nn; config=19c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 483.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M2 Pro (19c); M2 Pro (19c) \u00b7 32 GB \u00b7 4bit; 4k 483.1 47.7 20.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5dj91x21",
   "source_url": "https://omlx.ai/benchmarks/performance/5dj91x21",
   "tg_tokens": null,
   "tps": 47.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-5i2fywzg-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 22.5 GB; TTFT 38865 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1686.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k tokens 1,686 PP tok/s 51.4 TG tok/s 38865 TTFT (ms) 22.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5i2fywzg",
   "source_url": "https://omlx.ai/benchmarks/performance/5i2fywzg",
   "tg_tokens": null,
   "tps": 51.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-frn517pd-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 21.8 GB; TTFT 15626 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2097.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k tokens 2,097 PP tok/s 61.0 TG tok/s 15626 TTFT (ms) 21.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/frn517pd",
   "source_url": "https://omlx.ai/benchmarks/performance/frn517pd",
   "tg_tokens": null,
   "tps": 61.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-c1x3niv4-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 1k 106.1 17.0 18.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c1x3niv4",
   "source_url": "https://omlx.ai/benchmarks/performance/c1x3niv4",
   "tg_tokens": null,
   "tps": 17.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-c1x3niv4-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 100.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 100.2 17.1 22.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c1x3niv4",
   "source_url": "https://omlx.ai/benchmarks/performance/c1x3niv4",
   "tg_tokens": null,
   "tps": 17.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-0thr0nri-ctx1024",
   "model": "Qwen3.8-3.6-27B-blend",
   "notes": "blend of Qwen3.8 and Qwen3.6 27B; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: e8q6miz1, hn6h4r3w, hy3x96j1, udkdodh7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 215.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-3.6-27B-blend-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 215.4 28.8 19.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0thr0nri",
   "source_url": "https://omlx.ai/benchmarks/performance/0thr0nri",
   "tg_tokens": null,
   "tps": 28.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-0thr0nri-ctx4096",
   "model": "Qwen3.8-3.6-27B-blend",
   "notes": "blend of Qwen3.8 and Qwen3.6 27B; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: e8q6miz1, hn6h4r3w, hy3x96j1, udkdodh7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 186.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-3.6-27B-blend-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 186.3 26.5 22.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0thr0nri",
   "source_url": "https://omlx.ai/benchmarks/performance/0thr0nri",
   "tg_tokens": null,
   "tps": 26.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-0thr0nri-ctx8192",
   "model": "Qwen3.8-3.6-27B-blend",
   "notes": "blend of Qwen3.8 and Qwen3.6 27B; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: e8q6miz1, hn6h4r3w, hy3x96j1, udkdodh7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 170.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-3.6-27B-blend-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 170.2 19.3 23.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0thr0nri",
   "source_url": "https://omlx.ai/benchmarks/performance/0thr0nri",
   "tg_tokens": null,
   "tps": 19.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-0thr0nri-ctx16384",
   "model": "Qwen3.8-3.6-27B-blend",
   "notes": "blend of Qwen3.8 and Qwen3.6 27B; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: e8q6miz1, hn6h4r3w, hy3x96j1, udkdodh7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 174.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-3.6-27B-blend-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 174.3 20.8 23.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0thr0nri",
   "source_url": "https://omlx.ai/benchmarks/performance/0thr0nri",
   "tg_tokens": null,
   "tps": 20.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-0thr0nri-ctx32768",
   "model": "Qwen3.8-3.6-27B-blend",
   "notes": "blend of Qwen3.8 and Qwen3.6 27B; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: e8q6miz1, hn6h4r3w, hy3x96j1, udkdodh7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 163.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-3.6-27B-blend-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 163.6 20.7 24.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0thr0nri",
   "source_url": "https://omlx.ai/benchmarks/performance/0thr0nri",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-8jqnvbhp-ctx1024",
   "model": "Qwen3.8-3.6-27B-blend",
   "notes": "blend of Qwen3.8 and Qwen3.6 27B; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: duk0zky8, p691ty3k, u2yiy6v4, uc3hay7n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 223.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-3.6-27B-blend-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 223.4 34.6 19.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8jqnvbhp",
   "source_url": "https://omlx.ai/benchmarks/performance/8jqnvbhp",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-8jqnvbhp-ctx4096",
   "model": "Qwen3.8-3.6-27B-blend",
   "notes": "blend of Qwen3.8 and Qwen3.6 27B; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: duk0zky8, p691ty3k, u2yiy6v4, uc3hay7n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 186.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-3.6-27B-blend-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 186.6 25.4 22.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8jqnvbhp",
   "source_url": "https://omlx.ai/benchmarks/performance/8jqnvbhp",
   "tg_tokens": null,
   "tps": 25.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-8jqnvbhp-ctx8192",
   "model": "Qwen3.8-3.6-27B-blend",
   "notes": "blend of Qwen3.8 and Qwen3.6 27B; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: duk0zky8, p691ty3k, u2yiy6v4, uc3hay7n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 175.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-3.6-27B-blend-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 175.7 21.9 23.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8jqnvbhp",
   "source_url": "https://omlx.ai/benchmarks/performance/8jqnvbhp",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-8jqnvbhp-ctx16384",
   "model": "Qwen3.8-3.6-27B-blend",
   "notes": "blend of Qwen3.8 and Qwen3.6 27B; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: duk0zky8, p691ty3k, u2yiy6v4, uc3hay7n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 197.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-3.6-27B-blend-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 197.2 26.2 23.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8jqnvbhp",
   "source_url": "https://omlx.ai/benchmarks/performance/8jqnvbhp",
   "tg_tokens": null,
   "tps": 26.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-8jqnvbhp-ctx32768",
   "model": "Qwen3.8-3.6-27B-blend",
   "notes": "blend of Qwen3.8 and Qwen3.6 27B; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: duk0zky8, p691ty3k, u2yiy6v4, uc3hay7n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 171.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-3.6-27B-blend-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 171.6 22.2 24.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8jqnvbhp",
   "source_url": "https://omlx.ai/benchmarks/performance/8jqnvbhp",
   "tg_tokens": null,
   "tps": 22.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-7vhnq8yv-ctx199680",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 195k context; peak memory 121.3 GB; TTFT 260096 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 768.9,
   "provenance": "community",
   "quant": "mxfp4+mxfp8",
   "quote": "Qwen3.8-Flash-Next-Uncensored-MLX-mixed-mxfp4_mxfp8_mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 195k tokens 768.9 PP tok/s 35.4 TG tok/s 260096 TTFT (ms) 121.3 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7vhnq8yv",
   "source_url": "https://omlx.ai/benchmarks/performance/7vhnq8yv",
   "tg_tokens": null,
   "tps": 35.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-9fdjbg93-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.2 GB at this context; duplicate traces of the same session: dmxaobg8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 555.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 555.1 56.1 71.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9fdjbg93",
   "source_url": "https://omlx.ai/benchmarks/performance/9fdjbg93",
   "tg_tokens": null,
   "tps": 56.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-9fdjbg93-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: dmxaobg8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 461.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 461.5 55.0 72.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9fdjbg93",
   "source_url": "https://omlx.ai/benchmarks/performance/9fdjbg93",
   "tg_tokens": null,
   "tps": 55.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4ulwokm6-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 195k context; peak memory 80.5 GB; TTFT 89813 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2227.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k tokens 2,227 PP tok/s 61.3 TG tok/s 89813 TTFT (ms) 80.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ulwokm6",
   "source_url": "https://omlx.ai/benchmarks/performance/4ulwokm6",
   "tg_tokens": null,
   "tps": 61.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-rbfppjwc-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 195k context; peak memory 78.2 GB; TTFT 94347 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2120.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 195k tokens 2,120 PP tok/s 69.9 TG tok/s 94347 TTFT (ms) 78.2 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rbfppjwc",
   "source_url": "https://omlx.ai/benchmarks/performance/rbfppjwc",
   "tg_tokens": null,
   "tps": 69.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-i50vinks-ctx1024",
   "model": "UI-Venus-2-9B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 5.9 GB; TTFT 4784 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 214.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "UI-Venus-2-9B-mlx-4bit on M2 Pro (16c); M2 Pro (16c) \u00b7 16 GB \u00b7 4bit; 1k tokens 214.0 PP tok/s 33.5 TG tok/s 4784 TTFT (ms) 5.9 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i50vinks",
   "source_url": "https://omlx.ai/benchmarks/performance/i50vinks",
   "tg_tokens": null,
   "tps": 33.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8g128bnk-ctx1024",
   "model": "Gemma 4 12B",
   "notes": "oMLX community benchmark trace; peak memory 12.8 GB at this context; duplicate traces of the same session: ssoqrndr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 943.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-12B-it-8bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 8bit; 1k 943.1 38.6 12.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8g128bnk",
   "source_url": "https://omlx.ai/benchmarks/performance/8g128bnk",
   "tg_tokens": null,
   "tps": 38.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8g128bnk-ctx4096",
   "model": "Gemma 4 12B",
   "notes": "oMLX community benchmark trace; peak memory 13.2 GB at this context; duplicate traces of the same session: ssoqrndr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1433.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-12B-it-8bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 8bit; 4k 1,433 38.0 13.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8g128bnk",
   "source_url": "https://omlx.ai/benchmarks/performance/8g128bnk",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-2bo2yspj-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: 74kb0jqv, jpu0u7oy, qw3zq2s2, xlru529e; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 831.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 831.5 74.7 14.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2bo2yspj",
   "source_url": "https://omlx.ai/benchmarks/performance/2bo2yspj",
   "tg_tokens": null,
   "tps": 74.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-2bo2yspj-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: 74kb0jqv, jpu0u7oy, qw3zq2s2, xlru529e; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 811.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 811.6 71.4 14.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2bo2yspj",
   "source_url": "https://omlx.ai/benchmarks/performance/2bo2yspj",
   "tg_tokens": null,
   "tps": 71.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-2bo2yspj-ctx8192",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: 74kb0jqv, jpu0u7oy, qw3zq2s2, xlru529e; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 815.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 815.9 66.9 14.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2bo2yspj",
   "source_url": "https://omlx.ai/benchmarks/performance/2bo2yspj",
   "tg_tokens": null,
   "tps": 66.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-2bo2yspj-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 15.2 GB at this context; duplicate traces of the same session: 74kb0jqv, jpu0u7oy, qw3zq2s2, xlru529e; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 778.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 778.3 60.5 15.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2bo2yspj",
   "source_url": "https://omlx.ai/benchmarks/performance/2bo2yspj",
   "tg_tokens": null,
   "tps": 60.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-2bo2yspj-ctx32768",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 74kb0jqv, jpu0u7oy, qw3zq2s2, xlru529e; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 713.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 32k 713.6 50.4 16.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2bo2yspj",
   "source_url": "https://omlx.ai/benchmarks/performance/2bo2yspj",
   "tg_tokens": null,
   "tps": 50.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ffrd7sxx-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 8ueqd82y, y3im37w3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1716.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 1,716 137.8 19.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ffrd7sxx",
   "source_url": "https://omlx.ai/benchmarks/performance/ffrd7sxx",
   "tg_tokens": null,
   "tps": 137.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ffrd7sxx-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 8ueqd82y, y3im37w3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2019.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 2,019 135.0 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ffrd7sxx",
   "source_url": "https://omlx.ai/benchmarks/performance/ffrd7sxx",
   "tg_tokens": null,
   "tps": 135.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ffrd7sxx-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: 8ueqd82y, y3im37w3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1424.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 1,424 102.5 20.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ffrd7sxx",
   "source_url": "https://omlx.ai/benchmarks/performance/ffrd7sxx",
   "tg_tokens": null,
   "tps": 102.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-yijojv0w-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 5tk629v8, oqt7m7t8, vkqbvt3e; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2125.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 2,125 72.6 21.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yijojv0w",
   "source_url": "https://omlx.ai/benchmarks/performance/yijojv0w",
   "tg_tokens": null,
   "tps": 72.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-yijojv0w-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: 5tk629v8, oqt7m7t8, vkqbvt3e; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2426.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 2,426 62.3 22.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yijojv0w",
   "source_url": "https://omlx.ai/benchmarks/performance/yijojv0w",
   "tg_tokens": null,
   "tps": 62.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-yijojv0w-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 5tk629v8, oqt7m7t8, vkqbvt3e; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2241.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 2,241 65.1 22.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yijojv0w",
   "source_url": "https://omlx.ai/benchmarks/performance/yijojv0w",
   "tg_tokens": null,
   "tps": 65.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-yijojv0w-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: 5tk629v8, oqt7m7t8, vkqbvt3e; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1927.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 1,927 57.2 23.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yijojv0w",
   "source_url": "https://omlx.ai/benchmarks/performance/yijojv0w",
   "tg_tokens": null,
   "tps": 57.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-4z3azm3i-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 2o91reye; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2099.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 2,099 72.6 21.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4z3azm3i",
   "source_url": "https://omlx.ai/benchmarks/performance/4z3azm3i",
   "tg_tokens": null,
   "tps": 72.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-4z3azm3i-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: 2o91reye; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2586.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 2,586 69.6 22.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4z3azm3i",
   "source_url": "https://omlx.ai/benchmarks/performance/4z3azm3i",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-4z3azm3i-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 2o91reye; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2332.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 2,332 64.8 22.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4z3azm3i",
   "source_url": "https://omlx.ai/benchmarks/performance/4z3azm3i",
   "tg_tokens": null,
   "tps": 64.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-4z3azm3i-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: 2o91reye; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1903.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 1,903 58.1 23.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4z3azm3i",
   "source_url": "https://omlx.ai/benchmarks/performance/4z3azm3i",
   "tg_tokens": null,
   "tps": 58.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-m7msypvn-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 6hpmwha0, 7djangvc, ahcraijq; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2089.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 2,089 112.1 21.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m7msypvn",
   "source_url": "https://omlx.ai/benchmarks/performance/m7msypvn",
   "tg_tokens": null,
   "tps": 112.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-m7msypvn-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 6hpmwha0, 7djangvc, ahcraijq; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2536.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 2,536 85.0 22.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m7msypvn",
   "source_url": "https://omlx.ai/benchmarks/performance/m7msypvn",
   "tg_tokens": null,
   "tps": 85.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-m7msypvn-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 6hpmwha0, 7djangvc, ahcraijq; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2291.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 2,291 86.0 22.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m7msypvn",
   "source_url": "https://omlx.ai/benchmarks/performance/m7msypvn",
   "tg_tokens": null,
   "tps": 86.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-m7msypvn-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 6hpmwha0, 7djangvc, ahcraijq; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1866.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 1,866 87.3 22.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m7msypvn",
   "source_url": "https://omlx.ai/benchmarks/performance/m7msypvn",
   "tg_tokens": null,
   "tps": 87.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-m0iyn3cj-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: bnep20l4, mdysi9c3, zprqau3i, zsr496mx; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2089.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 2,089 123.0 21.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m0iyn3cj",
   "source_url": "https://omlx.ai/benchmarks/performance/m0iyn3cj",
   "tg_tokens": null,
   "tps": 123.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-m0iyn3cj-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: bnep20l4, mdysi9c3, zprqau3i, zsr496mx; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2569.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,569 117.6 22.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m0iyn3cj",
   "source_url": "https://omlx.ai/benchmarks/performance/m0iyn3cj",
   "tg_tokens": null,
   "tps": 117.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-m0iyn3cj-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: bnep20l4, mdysi9c3, zprqau3i, zsr496mx; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2544.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 2,544 84.9 22.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m0iyn3cj",
   "source_url": "https://omlx.ai/benchmarks/performance/m0iyn3cj",
   "tg_tokens": null,
   "tps": 84.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-m0iyn3cj-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: bnep20l4, mdysi9c3, zprqau3i, zsr496mx; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2293.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 2,293 67.0 22.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m0iyn3cj",
   "source_url": "https://omlx.ai/benchmarks/performance/m0iyn3cj",
   "tg_tokens": null,
   "tps": 67.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-m0iyn3cj-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: bnep20l4, mdysi9c3, zprqau3i, zsr496mx; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1865.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 1,865 84.8 22.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m0iyn3cj",
   "source_url": "https://omlx.ai/benchmarks/performance/m0iyn3cj",
   "tg_tokens": null,
   "tps": 84.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-g1bvxjpn-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: a5uuvddc; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2107.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 2,107 118.2 21.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g1bvxjpn",
   "source_url": "https://omlx.ai/benchmarks/performance/g1bvxjpn",
   "tg_tokens": null,
   "tps": 118.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-g1bvxjpn-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: a5uuvddc; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2548.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 2,548 80.9 22.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g1bvxjpn",
   "source_url": "https://omlx.ai/benchmarks/performance/g1bvxjpn",
   "tg_tokens": null,
   "tps": 80.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-g1bvxjpn-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: a5uuvddc; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2284.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 2,284 87.1 22.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g1bvxjpn",
   "source_url": "https://omlx.ai/benchmarks/performance/g1bvxjpn",
   "tg_tokens": null,
   "tps": 87.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-g1bvxjpn-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: a5uuvddc; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1829.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 1,829 79.9 22.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g1bvxjpn",
   "source_url": "https://omlx.ai/benchmarks/performance/g1bvxjpn",
   "tg_tokens": null,
   "tps": 79.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-dd9e3v2i-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.0 GB at this context; duplicate traces of the same session: b89k7mzm, nkfpwlqp, u7iblqy0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1641.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 6bit; 1k 1,641 111.1 30.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dd9e3v2i",
   "source_url": "https://omlx.ai/benchmarks/performance/dd9e3v2i",
   "tg_tokens": null,
   "tps": 111.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-dd9e3v2i-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.7 GB at this context; duplicate traces of the same session: b89k7mzm, nkfpwlqp, u7iblqy0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2069.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 6bit; 8k 2,069 71.9 30.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dd9e3v2i",
   "source_url": "https://omlx.ai/benchmarks/performance/dd9e3v2i",
   "tg_tokens": null,
   "tps": 71.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-dd9e3v2i-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.9 GB at this context; duplicate traces of the same session: b89k7mzm, nkfpwlqp, u7iblqy0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1910.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 6bit; 16k 1,910 76.9 30.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dd9e3v2i",
   "source_url": "https://omlx.ai/benchmarks/performance/dd9e3v2i",
   "tg_tokens": null,
   "tps": 76.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-dd9e3v2i-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.2 GB at this context; duplicate traces of the same session: b89k7mzm, nkfpwlqp, u7iblqy0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1597.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 6bit; 32k 1,597 74.7 31.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dd9e3v2i",
   "source_url": "https://omlx.ai/benchmarks/performance/dd9e3v2i",
   "tg_tokens": null,
   "tps": 74.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0vvlxlhu-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 805.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 805.4 38.7 33.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0vvlxlhu",
   "source_url": "https://omlx.ai/benchmarks/performance/0vvlxlhu",
   "tg_tokens": null,
   "tps": 38.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0vvlxlhu-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 919.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 919.4 41.4 34.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0vvlxlhu",
   "source_url": "https://omlx.ai/benchmarks/performance/0vvlxlhu",
   "tg_tokens": null,
   "tps": 41.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-rba7sxjt-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: 5d0f1v4e, af0xg3f2, atguxyif; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 293.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 1k 293.1 22.4 72.8 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rba7sxjt",
   "source_url": "https://omlx.ai/benchmarks/performance/rba7sxjt",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-rba7sxjt-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.2 GB at this context; duplicate traces of the same session: 5d0f1v4e, af0xg3f2, atguxyif; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 457.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 4k 457.8 22.4 73.2 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rba7sxjt",
   "source_url": "https://omlx.ai/benchmarks/performance/rba7sxjt",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-rba7sxjt-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.0 GB at this context; duplicate traces of the same session: 5d0f1v4e, af0xg3f2, atguxyif; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 585.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 8k 585.9 7.1 73.0 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rba7sxjt",
   "source_url": "https://omlx.ai/benchmarks/performance/rba7sxjt",
   "tg_tokens": null,
   "tps": 7.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-rba7sxjt-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.5 GB at this context; duplicate traces of the same session: 5d0f1v4e, af0xg3f2, atguxyif; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 668.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 16k 668.9 25.6 73.5 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rba7sxjt",
   "source_url": "https://omlx.ai/benchmarks/performance/rba7sxjt",
   "tg_tokens": null,
   "tps": 25.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-a5mmzuj7-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 101.2 GB; TTFT 5656 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 724.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 724.2 PP tok/s 58.8 TG tok/s 5656 TTFT (ms) 101.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a5mmzuj7",
   "source_url": "https://omlx.ai/benchmarks/performance/a5mmzuj7",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zohp2hh0-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.2 GB at this context; duplicate traces of the same session: e3o04zqk, ic2k8zvo, lnpjsrw4, lt5nqqae, pz52rmrp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 913.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 913.3 62.4 69.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zohp2hh0",
   "source_url": "https://omlx.ai/benchmarks/performance/zohp2hh0",
   "tg_tokens": null,
   "tps": 62.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zohp2hh0-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.3 GB at this context; duplicate traces of the same session: e3o04zqk, ic2k8zvo, lnpjsrw4, lt5nqqae, pz52rmrp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1663.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,663 39.5 69.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zohp2hh0",
   "source_url": "https://omlx.ai/benchmarks/performance/zohp2hh0",
   "tg_tokens": null,
   "tps": 39.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zohp2hh0-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.1 GB at this context; duplicate traces of the same session: e3o04zqk, ic2k8zvo, lnpjsrw4, lt5nqqae, pz52rmrp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1236.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,236 63.0 70.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zohp2hh0",
   "source_url": "https://omlx.ai/benchmarks/performance/zohp2hh0",
   "tg_tokens": null,
   "tps": 63.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zohp2hh0-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.4 GB at this context; duplicate traces of the same session: e3o04zqk, ic2k8zvo, lnpjsrw4, lt5nqqae, pz52rmrp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1622.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,622 20.0 71.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zohp2hh0",
   "source_url": "https://omlx.ai/benchmarks/performance/zohp2hh0",
   "tg_tokens": null,
   "tps": 20.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zohp2hh0-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.4 GB at this context; duplicate traces of the same session: e3o04zqk, ic2k8zvo, lnpjsrw4, lt5nqqae, pz52rmrp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1416.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 1,416 44.0 72.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zohp2hh0",
   "source_url": "https://omlx.ai/benchmarks/performance/zohp2hh0",
   "tg_tokens": null,
   "tps": 44.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zohp2hh0-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.9 GB at this context; duplicate traces of the same session: e3o04zqk, ic2k8zvo, lnpjsrw4, lt5nqqae, pz52rmrp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1053.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 1,053 43.1 74.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zohp2hh0",
   "source_url": "https://omlx.ai/benchmarks/performance/zohp2hh0",
   "tg_tokens": null,
   "tps": 43.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-2bw45zre-ctx4096",
   "model": "Ternary-Bonsai-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 12.6 GB at this context; duplicate traces of the same session: nw3gve82; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 321.2,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-27B-mlx-2bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 2bit; 4k 321.2 42.3 12.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2bw45zre",
   "source_url": "https://omlx.ai/benchmarks/performance/2bw45zre",
   "tg_tokens": null,
   "tps": 42.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-2bw45zre-ctx16384",
   "model": "Ternary-Bonsai-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: nw3gve82; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 301.5,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-27B-mlx-2bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 2bit; 16k 301.5 38.2 14.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2bw45zre",
   "source_url": "https://omlx.ai/benchmarks/performance/2bw45zre",
   "tg_tokens": null,
   "tps": 38.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-hkpa20ll-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: wibwoaom; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1064.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 1,064 78.9 14.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hkpa20ll",
   "source_url": "https://omlx.ai/benchmarks/performance/hkpa20ll",
   "tg_tokens": null,
   "tps": 78.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-hkpa20ll-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 15.2 GB at this context; duplicate traces of the same session: wibwoaom; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1141.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 16k 1,141 66.1 15.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hkpa20ll",
   "source_url": "https://omlx.ai/benchmarks/performance/hkpa20ll",
   "tg_tokens": null,
   "tps": 66.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-w1mby4qp-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 6.1 GB at this context; duplicate traces of the same session: 7rpxbn3v; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 990.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E4B-it-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 990.8 48.5 6.1 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w1mby4qp",
   "source_url": "https://omlx.ai/benchmarks/performance/w1mby4qp",
   "tg_tokens": null,
   "tps": 48.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-w1mby4qp-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: 7rpxbn3v; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1044.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E4B-it-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 1,044 47.3 6.2 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w1mby4qp",
   "source_url": "https://omlx.ai/benchmarks/performance/w1mby4qp",
   "tg_tokens": null,
   "tps": 47.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-i3o7cbct-ctx1024",
   "model": "Nemotron-Cascade-2-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 32.7 GB at this context; duplicate traces of the same session: 3ezr26pc, 7f8kqduy, d3ype31x, hw31kv8s, kvvxtwa6; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 743.2,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Nemotron-Cascade-2-30B-A3B-mxfp8 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 743.2 70.3 32.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i3o7cbct",
   "source_url": "https://omlx.ai/benchmarks/performance/i3o7cbct",
   "tg_tokens": null,
   "tps": 70.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-i3o7cbct-ctx4096",
   "model": "Nemotron-Cascade-2-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 33.6 GB at this context; duplicate traces of the same session: 3ezr26pc, 7f8kqduy, d3ype31x, hw31kv8s, kvvxtwa6; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 655.2,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Nemotron-Cascade-2-30B-A3B-mxfp8 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 655.2 62.8 33.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i3o7cbct",
   "source_url": "https://omlx.ai/benchmarks/performance/i3o7cbct",
   "tg_tokens": null,
   "tps": 62.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-i3o7cbct-ctx8192",
   "model": "Nemotron-Cascade-2-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 33.6 GB at this context; duplicate traces of the same session: 3ezr26pc, 7f8kqduy, d3ype31x, hw31kv8s, kvvxtwa6; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 632.8,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Nemotron-Cascade-2-30B-A3B-mxfp8 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 8k 632.8 60.7 33.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i3o7cbct",
   "source_url": "https://omlx.ai/benchmarks/performance/i3o7cbct",
   "tg_tokens": null,
   "tps": 60.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-i3o7cbct-ctx16384",
   "model": "Nemotron-Cascade-2-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 33.7 GB at this context; duplicate traces of the same session: 3ezr26pc, 7f8kqduy, d3ype31x, hw31kv8s, kvvxtwa6; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 646.9,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Nemotron-Cascade-2-30B-A3B-mxfp8 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 16k 646.9 50.4 33.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i3o7cbct",
   "source_url": "https://omlx.ai/benchmarks/performance/i3o7cbct",
   "tg_tokens": null,
   "tps": 50.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-i3o7cbct-ctx32768",
   "model": "Nemotron-Cascade-2-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: 3ezr26pc, 7f8kqduy, d3ype31x, hw31kv8s, kvvxtwa6; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 593.4,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Nemotron-Cascade-2-30B-A3B-mxfp8 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 32k 593.4 54.4 33.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i3o7cbct",
   "source_url": "https://omlx.ai/benchmarks/performance/i3o7cbct",
   "tg_tokens": null,
   "tps": 54.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-i3o7cbct-ctx65536",
   "model": "Nemotron-Cascade-2-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 34.0 GB at this context; duplicate traces of the same session: 3ezr26pc, 7f8kqduy, d3ype31x, hw31kv8s, kvvxtwa6; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 530.3,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Nemotron-Cascade-2-30B-A3B-mxfp8 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 64k 530.3 47.0 34.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i3o7cbct",
   "source_url": "https://omlx.ai/benchmarks/performance/i3o7cbct",
   "tg_tokens": null,
   "tps": 47.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-svwci1xp-ctx4096",
   "model": "Qwen3 8B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 5.4 GB; TTFT 18578 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-8B-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k tokens 220.5 PP tok/s 18.8 TG tok/s 18578 TTFT (ms) 5.4 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/svwci1xp",
   "source_url": "https://omlx.ai/benchmarks/performance/svwci1xp",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-niugqa4o-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 44.0 GB at this context; duplicate traces of the same session: ltn2uads; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1042.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next-MLX-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 1,042 60.5 44.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/niugqa4o",
   "source_url": "https://omlx.ai/benchmarks/performance/niugqa4o",
   "tg_tokens": null,
   "tps": 60.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-niugqa4o-ctx16384",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 44.2 GB at this context; duplicate traces of the same session: ltn2uads; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 879.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next-MLX-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 16k 879.3 57.1 44.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/niugqa4o",
   "source_url": "https://omlx.ai/benchmarks/performance/niugqa4o",
   "tg_tokens": null,
   "tps": 57.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fji65joq-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 0w8ewnxm; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2099.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 2,099 72.6 21.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fji65joq",
   "source_url": "https://omlx.ai/benchmarks/performance/fji65joq",
   "tg_tokens": null,
   "tps": 72.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fji65joq-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: 0w8ewnxm; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2586.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 2,586 69.6 22.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fji65joq",
   "source_url": "https://omlx.ai/benchmarks/performance/fji65joq",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fji65joq-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 0w8ewnxm; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2332.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 2,332 64.8 22.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fji65joq",
   "source_url": "https://omlx.ai/benchmarks/performance/fji65joq",
   "tg_tokens": null,
   "tps": 64.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fji65joq-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: 0w8ewnxm; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1903.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 1,903 58.1 23.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fji65joq",
   "source_url": "https://omlx.ai/benchmarks/performance/fji65joq",
   "tg_tokens": null,
   "tps": 58.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-z7ep5jtv-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 5qu2sg98; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2107.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 2,107 118.2 21.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z7ep5jtv",
   "source_url": "https://omlx.ai/benchmarks/performance/z7ep5jtv",
   "tg_tokens": null,
   "tps": 118.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-z7ep5jtv-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 5qu2sg98; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2548.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 2,548 80.9 22.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z7ep5jtv",
   "source_url": "https://omlx.ai/benchmarks/performance/z7ep5jtv",
   "tg_tokens": null,
   "tps": 80.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-z7ep5jtv-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 5qu2sg98; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2284.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 2,284 87.1 22.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z7ep5jtv",
   "source_url": "https://omlx.ai/benchmarks/performance/z7ep5jtv",
   "tg_tokens": null,
   "tps": 87.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-z7ep5jtv-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 5qu2sg98; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1829.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 1,829 79.9 22.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z7ep5jtv",
   "source_url": "https://omlx.ai/benchmarks/performance/z7ep5jtv",
   "tg_tokens": null,
   "tps": 79.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-g0v85j84-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 3vk2cgyo, 8j69p8iz, yjfwtn77; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1130.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 1,130 58.4 21.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g0v85j84",
   "source_url": "https://omlx.ai/benchmarks/performance/g0v85j84",
   "tg_tokens": null,
   "tps": 58.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-g0v85j84-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 3vk2cgyo, 8j69p8iz, yjfwtn77; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2500.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 2,500 50.6 22.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g0v85j84",
   "source_url": "https://omlx.ai/benchmarks/performance/g0v85j84",
   "tg_tokens": null,
   "tps": 50.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-g0v85j84-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 3vk2cgyo, 8j69p8iz, yjfwtn77; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2267.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 2,267 79.2 22.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g0v85j84",
   "source_url": "https://omlx.ai/benchmarks/performance/g0v85j84",
   "tg_tokens": null,
   "tps": 79.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-g0v85j84-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 3vk2cgyo, 8j69p8iz, yjfwtn77; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1844.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 1,844 82.7 22.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g0v85j84",
   "source_url": "https://omlx.ai/benchmarks/performance/g0v85j84",
   "tg_tokens": null,
   "tps": 82.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-y4lcskay-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: 16g7xv3t, vc50ghls; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1440.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 6bit; 1k 1,440 110.2 29.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y4lcskay",
   "source_url": "https://omlx.ai/benchmarks/performance/y4lcskay",
   "tg_tokens": null,
   "tps": 110.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-y4lcskay-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.7 GB at this context; duplicate traces of the same session: 16g7xv3t, vc50ghls; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2080.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 6bit; 8k 2,080 70.4 30.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y4lcskay",
   "source_url": "https://omlx.ai/benchmarks/performance/y4lcskay",
   "tg_tokens": null,
   "tps": 70.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-y4lcskay-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.9 GB at this context; duplicate traces of the same session: 16g7xv3t, vc50ghls; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1887.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 6bit; 16k 1,887 78.5 30.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y4lcskay",
   "source_url": "https://omlx.ai/benchmarks/performance/y4lcskay",
   "tg_tokens": null,
   "tps": 78.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-16fjtx6e-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 805.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 805.4 38.7 33.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/16fjtx6e",
   "source_url": "https://omlx.ai/benchmarks/performance/16fjtx6e",
   "tg_tokens": null,
   "tps": 38.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-16fjtx6e-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 919.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 919.4 41.4 34.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/16fjtx6e",
   "source_url": "https://omlx.ai/benchmarks/performance/16fjtx6e",
   "tg_tokens": null,
   "tps": 41.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7412wjgz-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: m3ekafmy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 794.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 794.8 41.8 33.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7412wjgz",
   "source_url": "https://omlx.ai/benchmarks/performance/7412wjgz",
   "tg_tokens": null,
   "tps": 41.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7412wjgz-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: m3ekafmy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 909.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 909.2 34.5 34.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7412wjgz",
   "source_url": "https://omlx.ai/benchmarks/performance/7412wjgz",
   "tg_tokens": null,
   "tps": 34.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-sod3rt3g-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 101.5 GB; TTFT 83418 ms; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 785.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 64k tokens 785.6 PP tok/s 55.7 TG tok/s 83418 TTFT (ms) 101.5 Peak mem",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sod3rt3g",
   "source_url": "https://omlx.ai/benchmarks/performance/sod3rt3g",
   "tg_tokens": null,
   "tps": 55.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-oxokoq12-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: r6jnog4n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 790.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 790.6 69.4 101.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oxokoq12",
   "source_url": "https://omlx.ai/benchmarks/performance/oxokoq12",
   "tg_tokens": null,
   "tps": 69.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-oxokoq12-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: r6jnog4n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 645.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 645.6 53.8 101.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oxokoq12",
   "source_url": "https://omlx.ai/benchmarks/performance/oxokoq12",
   "tg_tokens": null,
   "tps": 53.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2iolu6mv-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 73.4 GB; TTFT 14124 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2320.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k tokens 2,320 PP tok/s 16.1 TG tok/s 14124 TTFT (ms) 73.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2iolu6mv",
   "source_url": "https://omlx.ai/benchmarks/performance/2iolu6mv",
   "tg_tokens": null,
   "tps": 16.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ee2ohwk0-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.7 GB; TTFT 1158 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 884.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 884.4 PP tok/s 37.1 TG tok/s 1158 TTFT (ms) 70.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ee2ohwk0",
   "source_url": "https://omlx.ai/benchmarks/performance/ee2ohwk0",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-jotb48a3-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; Swift 1.5 build; oMLX community benchmark trace, single run at 64k context; peak memory 106.4 GB; TTFT 101333 ms; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 646.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 64k tokens 646.7 PP tok/s 49.8 TG tok/s 101333 TTFT (ms) 106.4 Peak mem",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jotb48a3",
   "source_url": "https://omlx.ai/benchmarks/performance/jotb48a3",
   "tg_tokens": null,
   "tps": 49.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-yxdicgaj-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 3.8 GB; TTFT 9586 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 427.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.5-4B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k tokens 427.3 PP tok/s 34.7 TG tok/s 9586 TTFT (ms) 3.8 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yxdicgaj",
   "source_url": "https://omlx.ai/benchmarks/performance/yxdicgaj",
   "tg_tokens": null,
   "tps": 34.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-y4hxh321-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 179.4 GB; TTFT 14536 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2254.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-MLX-oQ4-MTP on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k tokens 2,254 PP tok/s 86.1 TG tok/s 14536 TTFT (ms) 179.4 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y4hxh321",
   "source_url": "https://omlx.ai/benchmarks/performance/y4hxh321",
   "tg_tokens": null,
   "tps": 86.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ro07kota-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 178.2 GB; TTFT 1979 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2069.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-MLX-oQ4-MTP on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k tokens 2,069 PP tok/s 56.3 TG tok/s 1979 TTFT (ms) 178.2 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ro07kota",
   "source_url": "https://omlx.ai/benchmarks/performance/ro07kota",
   "tg_tokens": null,
   "tps": 56.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-3aytttr2-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 177.3 GB; TTFT 799 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1282.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-MLX-oQ4-MTP on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k tokens 1,282 PP tok/s 64.0 TG tok/s 799 TTFT (ms) 177.3 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3aytttr2",
   "source_url": "https://omlx.ai/benchmarks/performance/3aytttr2",
   "tg_tokens": null,
   "tps": 64.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-picc3cvr-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 17.9 GB; TTFT 127988 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 128.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 16k tokens 128.0 PP tok/s 16.5 TG tok/s 127988 TTFT (ms) 17.9 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/picc3cvr",
   "source_url": "https://omlx.ai/benchmarks/performance/picc3cvr",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-tvvooh8j-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: du75a84b; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 400.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 400.0 17.2 16.0 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tvvooh8j",
   "source_url": "https://omlx.ai/benchmarks/performance/tvvooh8j",
   "tg_tokens": null,
   "tps": 17.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-tvvooh8j-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: du75a84b; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 416.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 416.4 17.0 17.0 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tvvooh8j",
   "source_url": "https://omlx.ai/benchmarks/performance/tvvooh8j",
   "tg_tokens": null,
   "tps": 17.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0yogn1rb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: ugre4c98; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 368.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 368.9 17.1 16.0 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0yogn1rb",
   "source_url": "https://omlx.ai/benchmarks/performance/0yogn1rb",
   "tg_tokens": null,
   "tps": 17.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0yogn1rb-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: ugre4c98; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 396.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 396.3 16.7 17.0 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0yogn1rb",
   "source_url": "https://omlx.ai/benchmarks/performance/0yogn1rb",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-2e5nb4gd-ctx1024",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: lcoipeyy; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1557.0,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Qwen3.8-27B-Abliterated-MLX-MXFP4-MTP on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,557 80.1 16.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2e5nb4gd",
   "source_url": "https://omlx.ai/benchmarks/performance/2e5nb4gd",
   "tg_tokens": null,
   "tps": 80.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-2e5nb4gd-ctx4096",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.9 GB at this context; duplicate traces of the same session: lcoipeyy; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1844.0,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Qwen3.8-27B-Abliterated-MLX-MXFP4-MTP on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 1,844 84.4 17.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2e5nb4gd",
   "source_url": "https://omlx.ai/benchmarks/performance/2e5nb4gd",
   "tg_tokens": null,
   "tps": 84.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-uqzen2fe-ctx1024",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: yaojb0e5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1492.0,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Qwen3.8-27B-Abliterated-MLX-MXFP4-MTP on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,492 46.1 16.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uqzen2fe",
   "source_url": "https://omlx.ai/benchmarks/performance/uqzen2fe",
   "tg_tokens": null,
   "tps": 46.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-uqzen2fe-ctx4096",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.9 GB at this context; duplicate traces of the same session: yaojb0e5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1792.0,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Qwen3.8-27B-Abliterated-MLX-MXFP4-MTP on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 1,792 49.7 17.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uqzen2fe",
   "source_url": "https://omlx.ai/benchmarks/performance/uqzen2fe",
   "tg_tokens": null,
   "tps": 49.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wfk4v4bs-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: uohar72q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1009.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,009 27.2 21.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wfk4v4bs",
   "source_url": "https://omlx.ai/benchmarks/performance/wfk4v4bs",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wfk4v4bs-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: uohar72q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 980.5,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 980.5 26.0 21.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wfk4v4bs",
   "source_url": "https://omlx.ai/benchmarks/performance/wfk4v4bs",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-1o471nxc-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 105.1 GB at this context; duplicate traces of the same session: jew1dbxd; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2514.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 2,514 48.7 105.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1o471nxc",
   "source_url": "https://omlx.ai/benchmarks/performance/1o471nxc",
   "tg_tokens": null,
   "tps": 48.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-1o471nxc-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 105.5 GB at this context; duplicate traces of the same session: jew1dbxd; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3805.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 3,805 47.0 105.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1o471nxc",
   "source_url": "https://omlx.ai/benchmarks/performance/1o471nxc",
   "tg_tokens": null,
   "tps": 47.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-myrdi7rj-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 5fb6s304; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1178.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,178 69.3 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/myrdi7rj",
   "source_url": "https://omlx.ai/benchmarks/performance/myrdi7rj",
   "tg_tokens": null,
   "tps": 69.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-myrdi7rj-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 5fb6s304; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2396.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,396 68.0 71.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/myrdi7rj",
   "source_url": "https://omlx.ai/benchmarks/performance/myrdi7rj",
   "tg_tokens": null,
   "tps": 68.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-q1xsq2s4-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.4 GB at this context; duplicate traces of the same session: 7iqf7n3w; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2280.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-MLX-oQ4-MTP on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,280 79.3 179.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q1xsq2s4",
   "source_url": "https://omlx.ai/benchmarks/performance/q1xsq2s4",
   "tg_tokens": null,
   "tps": 79.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-q1xsq2s4-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.3 GB at this context; duplicate traces of the same session: 7iqf7n3w; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2199.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-MLX-oQ4-MTP on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,199 83.0 181.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q1xsq2s4",
   "source_url": "https://omlx.ai/benchmarks/performance/q1xsq2s4",
   "tg_tokens": null,
   "tps": 83.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-nlj2yqka-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.4 GB at this context; duplicate traces of the same session: dnri7ytp; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2277.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-MLX-oQ4-MTP on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,277 61.7 179.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nlj2yqka",
   "source_url": "https://omlx.ai/benchmarks/performance/nlj2yqka",
   "tg_tokens": null,
   "tps": 61.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-nlj2yqka-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.3 GB at this context; duplicate traces of the same session: dnri7ytp; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2210.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-MLX-oQ4-MTP on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,210 70.3 181.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nlj2yqka",
   "source_url": "https://omlx.ai/benchmarks/performance/nlj2yqka",
   "tg_tokens": null,
   "tps": 70.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-yd63h13n-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 195k context; peak memory 182.6 GB; TTFT 92350 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2166.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-MLX-oQ4-MTP on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k tokens 2,166 PP tok/s 75.8 TG tok/s 92350 TTFT (ms) 182.6 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yd63h13n",
   "source_url": "https://omlx.ai/benchmarks/performance/yd63h13n",
   "tg_tokens": null,
   "tps": 75.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ymujx71m-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.3 GB at this context; duplicate traces of the same session: e8fk68c1, pqn0xl94; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1474.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-MLX-oQ4-MTP on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,474 59.7 177.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ymujx71m",
   "source_url": "https://omlx.ai/benchmarks/performance/ymujx71m",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ymujx71m-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: e8fk68c1, pqn0xl94; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2285.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-MLX-oQ4-MTP on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 2,285 60.1 178.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ymujx71m",
   "source_url": "https://omlx.ai/benchmarks/performance/ymujx71m",
   "tg_tokens": null,
   "tps": 60.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ymujx71m-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.9 GB at this context; duplicate traces of the same session: e8fk68c1, pqn0xl94; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2291.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-MLX-oQ4-MTP on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 2,291 69.4 178.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ymujx71m",
   "source_url": "https://omlx.ai/benchmarks/performance/ymujx71m",
   "tg_tokens": null,
   "tps": 69.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-3gi16lqk-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.1 GB at this context; duplicate traces of the same session: nj0inr3s, y5cns3x7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2246.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-MLX-oQ4-MTP on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,246 75.0 180.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3gi16lqk",
   "source_url": "https://omlx.ai/benchmarks/performance/3gi16lqk",
   "tg_tokens": null,
   "tps": 75.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-3gi16lqk-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.3 GB at this context; duplicate traces of the same session: nj0inr3s, y5cns3x7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2213.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-MLX-oQ4-MTP on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,213 75.9 181.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3gi16lqk",
   "source_url": "https://omlx.ai/benchmarks/performance/3gi16lqk",
   "tg_tokens": null,
   "tps": 75.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-3gi16lqk-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 182.6 GB at this context; duplicate traces of the same session: nj0inr3s, y5cns3x7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2150.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-MLX-oQ4-MTP on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 2,150 57.8 182.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3gi16lqk",
   "source_url": "https://omlx.ai/benchmarks/performance/3gi16lqk",
   "tg_tokens": null,
   "tps": 57.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-x2504sd7-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build, MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 087e6oev, 1ofsjnsz, b6d1qz85, bw0ii514, kfjso5yr, w9q85epe; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 742.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fixed-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k 742.5 42.7 21.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x2504sd7",
   "source_url": "https://omlx.ai/benchmarks/performance/x2504sd7",
   "tg_tokens": null,
   "tps": 42.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-07",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-x2504sd7-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build, MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 087e6oev, 1ofsjnsz, b6d1qz85, bw0ii514, kfjso5yr, w9q85epe; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 906.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fixed-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 906.9 41.0 21.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x2504sd7",
   "source_url": "https://omlx.ai/benchmarks/performance/x2504sd7",
   "tg_tokens": null,
   "tps": 41.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-x2504sd7-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build, MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 087e6oev, 1ofsjnsz, b6d1qz85, bw0ii514, kfjso5yr, w9q85epe; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 906.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fixed-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 8k 906.6 43.3 21.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x2504sd7",
   "source_url": "https://omlx.ai/benchmarks/performance/x2504sd7",
   "tg_tokens": null,
   "tps": 43.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-x2504sd7-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build, MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 087e6oev, 1ofsjnsz, b6d1qz85, bw0ii514, kfjso5yr, w9q85epe; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 845.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fixed-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 16k 845.7 37.8 22.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x2504sd7",
   "source_url": "https://omlx.ai/benchmarks/performance/x2504sd7",
   "tg_tokens": null,
   "tps": 37.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-x2504sd7-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build, MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: 087e6oev, 1ofsjnsz, b6d1qz85, bw0ii514, kfjso5yr, w9q85epe; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 754.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fixed-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 32k 754.5 34.4 22.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x2504sd7",
   "source_url": "https://omlx.ai/benchmarks/performance/x2504sd7",
   "tg_tokens": null,
   "tps": 34.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-x2504sd7-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build, MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 087e6oev, 1ofsjnsz, b6d1qz85, bw0ii514, kfjso5yr, w9q85epe; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 593.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fixed-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 64k 593.8 27.4 23.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x2504sd7",
   "source_url": "https://omlx.ai/benchmarks/performance/x2504sd7",
   "tg_tokens": null,
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-07",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-x2504sd7-ctx131072",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build, MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: 087e6oev, 1ofsjnsz, b6d1qz85, bw0ii514, kfjso5yr, w9q85epe; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 394.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fixed-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 128k 394.9 18.5 24.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x2504sd7",
   "source_url": "https://omlx.ai/benchmarks/performance/x2504sd7",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-nwuhb6k3-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 18.9 GB; TTFT 293712 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 111.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 32k tokens 111.6 PP tok/s 15.3 TG tok/s 293712 TTFT (ms) 18.9 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nwuhb6k3",
   "source_url": "https://omlx.ai/benchmarks/performance/nwuhb6k3",
   "tg_tokens": null,
   "tps": 15.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-wk2yff0i-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: 2xcot9s9, zaln3vx9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1829.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 4bit; 4k 1,829 52.7 17.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wk2yff0i",
   "source_url": "https://omlx.ai/benchmarks/performance/wk2yff0i",
   "tg_tokens": null,
   "tps": 52.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-wk2yff0i-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 2xcot9s9, zaln3vx9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1805.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 4bit; 16k 1,805 49.9 17.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wk2yff0i",
   "source_url": "https://omlx.ai/benchmarks/performance/wk2yff0i",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-wk2yff0i-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: 2xcot9s9, zaln3vx9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1184.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 4bit; 128k 1,184 34.7 24.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wk2yff0i",
   "source_url": "https://omlx.ai/benchmarks/performance/wk2yff0i",
   "tg_tokens": null,
   "tps": 34.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-5l83i92l-ctx1024",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: mxpsojay; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1550.0,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Qwen3.8-27B-Abliterated-MLX-MXFP4-MTP on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,550 94.5 16.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5l83i92l",
   "source_url": "https://omlx.ai/benchmarks/performance/5l83i92l",
   "tg_tokens": null,
   "tps": 94.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-5l83i92l-ctx4096",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.9 GB at this context; duplicate traces of the same session: mxpsojay; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1833.0,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Qwen3.8-27B-Abliterated-MLX-MXFP4-MTP on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 1,833 87.8 17.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5l83i92l",
   "source_url": "https://omlx.ai/benchmarks/performance/5l83i92l",
   "tg_tokens": null,
   "tps": 87.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-7z5hd3it-ctx1024",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: 4d2aqh17; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1558.0,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Qwen3.8-27B-Abliterated-MLX-MXFP4-MTP on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,558 80.4 16.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7z5hd3it",
   "source_url": "https://omlx.ai/benchmarks/performance/7z5hd3it",
   "tg_tokens": null,
   "tps": 80.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-7z5hd3it-ctx4096",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.9 GB at this context; duplicate traces of the same session: 4d2aqh17; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1862.0,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Qwen3.8-27B-Abliterated-MLX-MXFP4-MTP on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 1,862 88.3 17.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7z5hd3it",
   "source_url": "https://omlx.ai/benchmarks/performance/7z5hd3it",
   "tg_tokens": null,
   "tps": 88.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-t8382y3i-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: q55fo3nl; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 428.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 428.9 10.0 28.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t8382y3i",
   "source_url": "https://omlx.ai/benchmarks/performance/t8382y3i",
   "tg_tokens": null,
   "tps": 10.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-t8382y3i-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: q55fo3nl; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 421.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 421.7 9.8 29.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t8382y3i",
   "source_url": "https://omlx.ai/benchmarks/performance/t8382y3i",
   "tg_tokens": null,
   "tps": 9.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-jgyhj42p-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: qx86ft3j; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 457.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 457.2 14.3 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jgyhj42p",
   "source_url": "https://omlx.ai/benchmarks/performance/jgyhj42p",
   "tg_tokens": null,
   "tps": 14.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-jgyhj42p-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: qx86ft3j; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 493.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 493.9 14.2 21.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jgyhj42p",
   "source_url": "https://omlx.ai/benchmarks/performance/jgyhj42p",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-5ofnkgue-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: ue5jalek; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 516.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 516.2 13.6 71.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ofnkgue",
   "source_url": "https://omlx.ai/benchmarks/performance/5ofnkgue",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-5ofnkgue-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 75.2 GB at this context; duplicate traces of the same session: ue5jalek; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 742.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 742.3 50.2 75.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ofnkgue",
   "source_url": "https://omlx.ai/benchmarks/performance/5ofnkgue",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-bxbm686b-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.2 GB at this context; duplicate traces of the same session: 3d352mot; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 749.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.5-9B-OptiQ-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 749.0 18.7 8.2 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bxbm686b",
   "source_url": "https://omlx.ai/benchmarks/performance/bxbm686b",
   "tg_tokens": null,
   "tps": 18.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-bxbm686b-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.5 GB at this context; duplicate traces of the same session: 3d352mot; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 734.2,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.5-9B-OptiQ-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 16k 734.2 17.5 8.5 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bxbm686b",
   "source_url": "https://omlx.ai/benchmarks/performance/bxbm686b",
   "tg_tokens": null,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q1fyqrsc-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: 02hu80oj, e57604qf, jonj0w7c, mc0wjpxa, p3jms5lp, smbsmq0k, sogj22nu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 666.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 666.2 33.2 16.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q1fyqrsc",
   "source_url": "https://omlx.ai/benchmarks/performance/q1fyqrsc",
   "tg_tokens": null,
   "tps": 33.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q1fyqrsc-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 02hu80oj, e57604qf, jonj0w7c, mc0wjpxa, p3jms5lp, smbsmq0k, sogj22nu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 983.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 983.6 32.8 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q1fyqrsc",
   "source_url": "https://omlx.ai/benchmarks/performance/q1fyqrsc",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q1fyqrsc-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 02hu80oj, e57604qf, jonj0w7c, mc0wjpxa, p3jms5lp, smbsmq0k, sogj22nu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 996.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 996.4 32.1 17.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q1fyqrsc",
   "source_url": "https://omlx.ai/benchmarks/performance/q1fyqrsc",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q1fyqrsc-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 02hu80oj, e57604qf, jonj0w7c, mc0wjpxa, p3jms5lp, smbsmq0k, sogj22nu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 972.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 972.8 30.8 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q1fyqrsc",
   "source_url": "https://omlx.ai/benchmarks/performance/q1fyqrsc",
   "tg_tokens": null,
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q1fyqrsc-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 02hu80oj, e57604qf, jonj0w7c, mc0wjpxa, p3jms5lp, smbsmq0k, sogj22nu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 904.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 904.5 29.2 18.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q1fyqrsc",
   "source_url": "https://omlx.ai/benchmarks/performance/q1fyqrsc",
   "tg_tokens": null,
   "tps": 29.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q1fyqrsc-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 02hu80oj, e57604qf, jonj0w7c, mc0wjpxa, p3jms5lp, smbsmq0k, sogj22nu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 791.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 791.4 25.8 20.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q1fyqrsc",
   "source_url": "https://omlx.ai/benchmarks/performance/q1fyqrsc",
   "tg_tokens": null,
   "tps": 25.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q1fyqrsc-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: 02hu80oj, e57604qf, jonj0w7c, mc0wjpxa, p3jms5lp, smbsmq0k, sogj22nu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 631.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 631.0 21.2 24.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q1fyqrsc",
   "source_url": "https://omlx.ai/benchmarks/performance/q1fyqrsc",
   "tg_tokens": null,
   "tps": 21.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-q1fyqrsc-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.0 GB at this context; duplicate traces of the same session: 02hu80oj, e57604qf, jonj0w7c, mc0wjpxa, p3jms5lp, smbsmq0k, sogj22nu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 521.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 195k 521.0 18.0 29.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q1fyqrsc",
   "source_url": "https://omlx.ai/benchmarks/performance/q1fyqrsc",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-am4miqiq-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 68.4 GB; TTFT 33310 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 245.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k tokens 245.9 PP tok/s 41.6 TG tok/s 33310 TTFT (ms) 68.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/am4miqiq",
   "source_url": "https://omlx.ai/benchmarks/performance/am4miqiq",
   "tg_tokens": null,
   "tps": 41.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-6k5d1jm3-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 59.2 GB; TTFT 657565 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 99.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k tokens 99.7 PP tok/s 16.9 TG tok/s 657565 TTFT (ms) 59.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6k5d1jm3",
   "source_url": "https://omlx.ai/benchmarks/performance/6k5d1jm3",
   "tg_tokens": null,
   "tps": 16.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-hferqjyy-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 146.7 GB at this context; duplicate traces of the same session: 617e4ngt, xvdlyblf; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4446.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 64k 4,446 105.8 146.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hferqjyy",
   "source_url": "https://omlx.ai/benchmarks/performance/hferqjyy",
   "tg_tokens": null,
   "tps": 105.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-hferqjyy-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 151.5 GB at this context; duplicate traces of the same session: 617e4ngt, xvdlyblf; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4324.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 128k 4,324 88.5 151.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hferqjyy",
   "source_url": "https://omlx.ai/benchmarks/performance/hferqjyy",
   "tg_tokens": null,
   "tps": 88.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-hferqjyy-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 153.2 GB at this context; duplicate traces of the same session: 617e4ngt, xvdlyblf; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4150.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 195k 4,150 91.4 153.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hferqjyy",
   "source_url": "https://omlx.ai/benchmarks/performance/hferqjyy",
   "tg_tokens": null,
   "tps": 91.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-9lht4xya-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 146.7 GB at this context; duplicate traces of the same session: 5p2nf1xb, 7ym5pang; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4455.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 64k 4,455 118.8 146.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9lht4xya",
   "source_url": "https://omlx.ai/benchmarks/performance/9lht4xya",
   "tg_tokens": null,
   "tps": 118.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-9lht4xya-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 151.5 GB at this context; duplicate traces of the same session: 5p2nf1xb, 7ym5pang; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4343.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 128k 4,343 96.5 151.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9lht4xya",
   "source_url": "https://omlx.ai/benchmarks/performance/9lht4xya",
   "tg_tokens": null,
   "tps": 96.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-9lht4xya-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 153.2 GB at this context; duplicate traces of the same session: 5p2nf1xb, 7ym5pang; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4221.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 195k 4,221 79.7 153.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9lht4xya",
   "source_url": "https://omlx.ai/benchmarks/performance/9lht4xya",
   "tg_tokens": null,
   "tps": 79.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-s7rhn2pa-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 144.5 GB at this context; duplicate traces of the same session: g57gorsx, x5oy3e3d; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4376.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 8k 4,376 97.5 144.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s7rhn2pa",
   "source_url": "https://omlx.ai/benchmarks/performance/s7rhn2pa",
   "tg_tokens": null,
   "tps": 97.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-s7rhn2pa-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.1 GB at this context; duplicate traces of the same session: g57gorsx, x5oy3e3d; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4546.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 16k 4,546 122.1 145.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s7rhn2pa",
   "source_url": "https://omlx.ai/benchmarks/performance/s7rhn2pa",
   "tg_tokens": null,
   "tps": 122.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-s7rhn2pa-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.7 GB at this context; duplicate traces of the same session: g57gorsx, x5oy3e3d; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4538.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 32k 4,538 113.8 145.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s7rhn2pa",
   "source_url": "https://omlx.ai/benchmarks/performance/s7rhn2pa",
   "tg_tokens": null,
   "tps": 113.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-9w7ncfx4-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace, single run at 8k context; peak memory 144.5 GB; TTFT 2031 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4035.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 8k tokens 4,035 PP tok/s 144.6 TG tok/s 2031 TTFT (ms) 144.5 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9w7ncfx4",
   "source_url": "https://omlx.ai/benchmarks/performance/9w7ncfx4",
   "tg_tokens": null,
   "tps": 144.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-hc0aj1dh-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.5 GB at this context; duplicate traces of the same session: 0jcte21w, 49xxfgjr, ivk7x3lh, mwf3pdca; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2058.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 2,058 107.1 142.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hc0aj1dh",
   "source_url": "https://omlx.ai/benchmarks/performance/hc0aj1dh",
   "tg_tokens": null,
   "tps": 107.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-hc0aj1dh-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: 0jcte21w, 49xxfgjr, ivk7x3lh, mwf3pdca; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3901.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 3,901 128.8 143.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hc0aj1dh",
   "source_url": "https://omlx.ai/benchmarks/performance/hc0aj1dh",
   "tg_tokens": null,
   "tps": 128.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-hc0aj1dh-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 146.8 GB at this context; duplicate traces of the same session: 0jcte21w, 49xxfgjr, ivk7x3lh, mwf3pdca; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4480.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 64k 4,480 116.8 146.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hc0aj1dh",
   "source_url": "https://omlx.ai/benchmarks/performance/hc0aj1dh",
   "tg_tokens": null,
   "tps": 116.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-hc0aj1dh-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 150.9 GB at this context; duplicate traces of the same session: 0jcte21w, 49xxfgjr, ivk7x3lh, mwf3pdca; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4358.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 128k 4,358 105.6 150.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hc0aj1dh",
   "source_url": "https://omlx.ai/benchmarks/performance/hc0aj1dh",
   "tg_tokens": null,
   "tps": 105.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-hc0aj1dh-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 153.3 GB at this context; duplicate traces of the same session: 0jcte21w, 49xxfgjr, ivk7x3lh, mwf3pdca; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4232.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 195k 4,232 87.4 153.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hc0aj1dh",
   "source_url": "https://omlx.ai/benchmarks/performance/hc0aj1dh",
   "tg_tokens": null,
   "tps": 87.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-2ffnbbis-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; mlx-community build; oMLX community benchmark trace; peak memory 67.5 GB at this context; duplicate traces of the same session: kj7r7fwg, ooxn31a5; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 303.3,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "mlx-community--Qwen3.5-122B-A10B-oQ4-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 303.3 36.2 67.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ffnbbis",
   "source_url": "https://omlx.ai/benchmarks/performance/2ffnbbis",
   "tg_tokens": null,
   "tps": 36.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-2ffnbbis-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; mlx-community build; oMLX community benchmark trace; peak memory 71.2 GB at this context; duplicate traces of the same session: kj7r7fwg, ooxn31a5; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 335.3,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "mlx-community--Qwen3.5-122B-A10B-oQ4-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 335.3 35.0 71.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ffnbbis",
   "source_url": "https://omlx.ai/benchmarks/performance/2ffnbbis",
   "tg_tokens": null,
   "tps": 35.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-2ffnbbis-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; mlx-community build; oMLX community benchmark trace; peak memory 72.3 GB at this context; duplicate traces of the same session: kj7r7fwg, ooxn31a5; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 328.5,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "mlx-community--Qwen3.5-122B-A10B-oQ4-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 328.5 35.1 72.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ffnbbis",
   "source_url": "https://omlx.ai/benchmarks/performance/2ffnbbis",
   "tg_tokens": null,
   "tps": 35.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-1175qcb1-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; mlx-community build; oMLX community benchmark trace; peak memory 68.9 GB at this context; duplicate traces of the same session: 1kxw0h9c, 8ra4of7t, d2dvxhfm, hc120m8t; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 292.6,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "mlx-community--Qwen3.5-122B-A10B-oQ4-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 292.6 39.8 68.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1175qcb1",
   "source_url": "https://omlx.ai/benchmarks/performance/1175qcb1",
   "tg_tokens": null,
   "tps": 39.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-1175qcb1-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; mlx-community build; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: 1kxw0h9c, 8ra4of7t, d2dvxhfm, hc120m8t; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 327.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "mlx-community--Qwen3.5-122B-A10B-oQ4-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 327.0 36.4 72.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1175qcb1",
   "source_url": "https://omlx.ai/benchmarks/performance/1175qcb1",
   "tg_tokens": null,
   "tps": 36.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-1175qcb1-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; mlx-community build; oMLX community benchmark trace; peak memory 74.0 GB at this context; duplicate traces of the same session: 1kxw0h9c, 8ra4of7t, d2dvxhfm, hc120m8t; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 319.8,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "mlx-community--Qwen3.5-122B-A10B-oQ4-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 319.8 32.7 74.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1175qcb1",
   "source_url": "https://omlx.ai/benchmarks/performance/1175qcb1",
   "tg_tokens": null,
   "tps": 32.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-1175qcb1-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; mlx-community build; oMLX community benchmark trace; peak memory 72.9 GB at this context; duplicate traces of the same session: 1kxw0h9c, 8ra4of7t, d2dvxhfm, hc120m8t; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 280.8,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "mlx-community--Qwen3.5-122B-A10B-oQ4-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 280.8 29.7 72.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1175qcb1",
   "source_url": "https://omlx.ai/benchmarks/performance/1175qcb1",
   "tg_tokens": null,
   "tps": 29.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-1175qcb1-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; mlx-community build; oMLX community benchmark trace; peak memory 72.9 GB at this context; duplicate traces of the same session: 1kxw0h9c, 8ra4of7t, d2dvxhfm, hc120m8t; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 231.2,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "mlx-community--Qwen3.5-122B-A10B-oQ4-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 32k 231.2 28.2 72.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1175qcb1",
   "source_url": "https://omlx.ai/benchmarks/performance/1175qcb1",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ziai8icn-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: kt0grbsq; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1009.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 1k 1,009 118.1 35.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ziai8icn",
   "source_url": "https://omlx.ai/benchmarks/performance/ziai8icn",
   "tg_tokens": null,
   "tps": 118.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ziai8icn-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: kt0grbsq; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3486.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 4k 3,486 122.7 36.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ziai8icn",
   "source_url": "https://omlx.ai/benchmarks/performance/ziai8icn",
   "tg_tokens": null,
   "tps": 122.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-39kxfthp-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 53.7 GB; TTFT 17463 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1876.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 32k tokens 1,876 PP tok/s 76.3 TG tok/s 17463 TTFT (ms) 53.7 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/39kxfthp",
   "source_url": "https://omlx.ai/benchmarks/performance/39kxfthp",
   "tg_tokens": null,
   "tps": 76.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-23s5xvxb-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 38.3 GB; TTFT 17228 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1902.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k tokens 1,902 PP tok/s 89.1 TG tok/s 17228 TTFT (ms) 38.3 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/23s5xvxb",
   "source_url": "https://omlx.ai/benchmarks/performance/23s5xvxb",
   "tg_tokens": null,
   "tps": 89.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lzep5oeu-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 37.4 GB; TTFT 7103 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2307.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 16k tokens 2,307 PP tok/s 97.6 TG tok/s 7103 TTFT (ms) 37.4 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lzep5oeu",
   "source_url": "https://omlx.ai/benchmarks/performance/lzep5oeu",
   "tg_tokens": null,
   "tps": 97.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-l2009ajw-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: 5wps44pd; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1300.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 4k 1,300 33.0 29.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l2009ajw",
   "source_url": "https://omlx.ai/benchmarks/performance/l2009ajw",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-l2009ajw-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: 5wps44pd; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1391.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 16k 1,391 31.5 30.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l2009ajw",
   "source_url": "https://omlx.ai/benchmarks/performance/l2009ajw",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-m13lp2xx-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: 49rdzxns, kzs8s7gw, v3cvk167, xgip131e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 768.7,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-27B-oQ6-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 768.7 53.1 26.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m13lp2xx",
   "source_url": "https://omlx.ai/benchmarks/performance/m13lp2xx",
   "tg_tokens": null,
   "tps": 53.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-m13lp2xx-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: 49rdzxns, kzs8s7gw, v3cvk167, xgip131e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 850.4,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-27B-oQ6-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 850.4 59.0 27.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m13lp2xx",
   "source_url": "https://omlx.ai/benchmarks/performance/m13lp2xx",
   "tg_tokens": null,
   "tps": 59.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-m13lp2xx-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: 49rdzxns, kzs8s7gw, v3cvk167, xgip131e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 850.4,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-27B-oQ6-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 850.4 53.6 28.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m13lp2xx",
   "source_url": "https://omlx.ai/benchmarks/performance/m13lp2xx",
   "tg_tokens": null,
   "tps": 53.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-m13lp2xx-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.9 GB at this context; duplicate traces of the same session: 49rdzxns, kzs8s7gw, v3cvk167, xgip131e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 609.4,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-27B-oQ6-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k 609.4 43.1 31.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m13lp2xx",
   "source_url": "https://omlx.ai/benchmarks/performance/m13lp2xx",
   "tg_tokens": null,
   "tps": 43.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-m13lp2xx-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.6 GB at this context; duplicate traces of the same session: 49rdzxns, kzs8s7gw, v3cvk167, xgip131e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 501.8,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-27B-oQ6-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 128k 501.8 39.7 36.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m13lp2xx",
   "source_url": "https://omlx.ai/benchmarks/performance/m13lp2xx",
   "tg_tokens": null,
   "tps": 39.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-qbad5llq-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.4 GB at this context; duplicate traces of the same session: hznrqc9j; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2408.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 2,408 53.1 100.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qbad5llq",
   "source_url": "https://omlx.ai/benchmarks/performance/qbad5llq",
   "tg_tokens": null,
   "tps": 53.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-qbad5llq-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.0 GB at this context; duplicate traces of the same session: hznrqc9j; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2991.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 2,991 44.1 101.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qbad5llq",
   "source_url": "https://omlx.ai/benchmarks/performance/qbad5llq",
   "tg_tokens": null,
   "tps": 44.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-8vn2z5k4-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.0 GB at this context; duplicate traces of the same session: rhbcavfj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 324.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 324.3 73.1 99.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8vn2z5k4",
   "source_url": "https://omlx.ai/benchmarks/performance/8vn2z5k4",
   "tg_tokens": null,
   "tps": 73.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-8vn2z5k4-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.5 GB at this context; duplicate traces of the same session: rhbcavfj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1445.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 1,445 55.2 99.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8vn2z5k4",
   "source_url": "https://omlx.ai/benchmarks/performance/8vn2z5k4",
   "tg_tokens": null,
   "tps": 55.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-3tq8mqmq-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.9 GB at this context; duplicate traces of the same session: 9gregdu3, bpiptd1d, kmwd9d3u, opjpxshk, peyind4n, ws3hq90e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 247.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 247.1 47.9 39.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3tq8mqmq",
   "source_url": "https://omlx.ai/benchmarks/performance/3tq8mqmq",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-3tq8mqmq-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.4 GB at this context; duplicate traces of the same session: 9gregdu3, bpiptd1d, kmwd9d3u, opjpxshk, peyind4n, ws3hq90e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 330.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 330.0 38.4 43.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3tq8mqmq",
   "source_url": "https://omlx.ai/benchmarks/performance/3tq8mqmq",
   "tg_tokens": null,
   "tps": 38.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-3tq8mqmq-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.8 GB at this context; duplicate traces of the same session: 9gregdu3, bpiptd1d, kmwd9d3u, opjpxshk, peyind4n, ws3hq90e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 325.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 325.8 26.6 43.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3tq8mqmq",
   "source_url": "https://omlx.ai/benchmarks/performance/3tq8mqmq",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-3tq8mqmq-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 44.9 GB at this context; duplicate traces of the same session: 9gregdu3, bpiptd1d, kmwd9d3u, opjpxshk, peyind4n, ws3hq90e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 310.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 310.0 37.6 44.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3tq8mqmq",
   "source_url": "https://omlx.ai/benchmarks/performance/3tq8mqmq",
   "tg_tokens": null,
   "tps": 37.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-3tq8mqmq-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 46.7 GB at this context; duplicate traces of the same session: 9gregdu3, bpiptd1d, kmwd9d3u, opjpxshk, peyind4n, ws3hq90e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 281.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 281.9 34.1 46.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3tq8mqmq",
   "source_url": "https://omlx.ai/benchmarks/performance/3tq8mqmq",
   "tg_tokens": null,
   "tps": 34.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-3tq8mqmq-ctx65536",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 49.3 GB at this context; duplicate traces of the same session: 9gregdu3, bpiptd1d, kmwd9d3u, opjpxshk, peyind4n, ws3hq90e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 238.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 238.8 31.6 49.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3tq8mqmq",
   "source_url": "https://omlx.ai/benchmarks/performance/3tq8mqmq",
   "tg_tokens": null,
   "tps": 31.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-3tq8mqmq-ctx131072",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 53.8 GB at this context; duplicate traces of the same session: 9gregdu3, bpiptd1d, kmwd9d3u, opjpxshk, peyind4n, ws3hq90e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 184.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 184.4 31.6 53.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3tq8mqmq",
   "source_url": "https://omlx.ai/benchmarks/performance/3tq8mqmq",
   "tg_tokens": null,
   "tps": 31.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-1rp60op4-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 7.9 GB at this context; duplicate traces of the same session: 13qqte98; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1908.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-E4B-it-MLX-8bit on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 8bit; 1k 1,908 28.9 7.9 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1rp60op4",
   "source_url": "https://omlx.ai/benchmarks/performance/1rp60op4",
   "tg_tokens": null,
   "tps": 28.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-1rp60op4-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 7.9 GB at this context; duplicate traces of the same session: 13qqte98; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2527.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-E4B-it-MLX-8bit on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 8bit; 4k 2,527 28.2 7.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1rp60op4",
   "source_url": "https://omlx.ai/benchmarks/performance/1rp60op4",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-rw1ovbat-ctx199680",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace, single run at 195k context; peak memory 8.3 GB; TTFT 96003 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2083.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "gemma-4-e2b-it-5bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 5bit; 195k tokens 2,083 PP tok/s 27.3 TG tok/s 96003 TTFT (ms) 8.3 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rw1ovbat",
   "source_url": "https://omlx.ai/benchmarks/performance/rw1ovbat",
   "tg_tokens": null,
   "tps": 27.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-9eugr4t0-ctx65536",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 4.9 GB; TTFT 14609 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4486.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "gemma-4-e2b-it-5bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 5bit; 64k tokens 4,486 PP tok/s 47.9 TG tok/s 14609 TTFT (ms) 4.9 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9eugr4t0",
   "source_url": "https://omlx.ai/benchmarks/performance/9eugr4t0",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-3lod2sne-ctx32768",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 4.1 GB; TTFT 6604 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4962.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "gemma-4-e2b-it-5bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 5bit; 32k tokens 4,962 PP tok/s 59.6 TG tok/s 6604 TTFT (ms) 4.1 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3lod2sne",
   "source_url": "https://omlx.ai/benchmarks/performance/3lod2sne",
   "tg_tokens": null,
   "tps": 59.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-yj2gyr5v-ctx131072",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 6.5 GB; TTFT 45914 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2855.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "gemma-4-e2b-it-5bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 5bit; 128k tokens 2,855 PP tok/s 34.1 TG tok/s 45914 TTFT (ms) 6.5 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yj2gyr5v",
   "source_url": "https://omlx.ai/benchmarks/performance/yj2gyr5v",
   "tg_tokens": null,
   "tps": 34.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-i67mwmd2-ctx1024",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: vcma497a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 367.2,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Muse-Glimmer-30B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 367.2 37.1 26.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i67mwmd2",
   "source_url": "https://omlx.ai/benchmarks/performance/i67mwmd2",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-i67mwmd2-ctx4096",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 27.3 GB at this context; duplicate traces of the same session: vcma497a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 663.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Muse-Glimmer-30B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 663.8 40.2 27.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i67mwmd2",
   "source_url": "https://omlx.ai/benchmarks/performance/i67mwmd2",
   "tg_tokens": null,
   "tps": 40.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zg7kqwqp-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 224 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4574.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 4,574 PP tok/s 150.8 TG tok/s 224 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zg7kqwqp",
   "source_url": "https://omlx.ai/benchmarks/performance/zg7kqwqp",
   "tg_tokens": null,
   "tps": 150.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8ochq8nd-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: i9pe7007; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4601.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,601 183.3 21.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8ochq8nd",
   "source_url": "https://omlx.ai/benchmarks/performance/8ochq8nd",
   "tg_tokens": null,
   "tps": 183.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8ochq8nd-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: i9pe7007; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6378.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,378 199.1 22.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8ochq8nd",
   "source_url": "https://omlx.ai/benchmarks/performance/8ochq8nd",
   "tg_tokens": null,
   "tps": 199.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-76kt8k4d-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 22.3 GB; TTFT 12585 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 81.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-6Bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 1k tokens 81.4 PP tok/s 3.7 TG tok/s 12585 TTFT (ms) 22.3 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/76kt8k4d",
   "source_url": "https://omlx.ai/benchmarks/performance/76kt8k4d",
   "tg_tokens": null,
   "tps": 3.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-r9dsvzhi-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: us1b7h29; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1135.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,135 72.0 17.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r9dsvzhi",
   "source_url": "https://omlx.ai/benchmarks/performance/r9dsvzhi",
   "tg_tokens": null,
   "tps": 72.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-r9dsvzhi-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: us1b7h29; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1381.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,381 69.6 19.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r9dsvzhi",
   "source_url": "https://omlx.ai/benchmarks/performance/r9dsvzhi",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-dycpsjvq-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: dmvb8quy; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2906.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 2,906 100.4 100.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dycpsjvq",
   "source_url": "https://omlx.ai/benchmarks/performance/dycpsjvq",
   "tg_tokens": null,
   "tps": 100.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-dycpsjvq-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.0 GB at this context; duplicate traces of the same session: dmvb8quy; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5011.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 5,011 104.9 101.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dycpsjvq",
   "source_url": "https://omlx.ai/benchmarks/performance/dycpsjvq",
   "tg_tokens": null,
   "tps": 104.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-7j0ie35t-ctx32768",
   "model": "supergemma4-26b-uncensored",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 15.8 GB; TTFT 71788 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 456.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "supergemma4-26b-uncensored-mlx-4bit-v2 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 32k tokens 456.5 PP tok/s 35.3 TG tok/s 71788 TTFT (ms) 15.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7j0ie35t",
   "source_url": "https://omlx.ai/benchmarks/performance/7j0ie35t",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-amzmvi8t-ctx1024",
   "model": "DeepSeek-V4-Flash",
   "notes": "oMLX community benchmark trace; peak memory 142.5 GB at this context; duplicate traces of the same session: 3sc75ldl, 7r81ct6o, ebmisc4o, illu4qwu, mvsmdjzz, n523g172, pmt6867t; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 590.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-V4-Flash-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 590.5 44.8 142.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/amzmvi8t",
   "source_url": "https://omlx.ai/benchmarks/performance/amzmvi8t",
   "tg_tokens": null,
   "tps": 44.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-amzmvi8t-ctx4096",
   "model": "DeepSeek-V4-Flash",
   "notes": "oMLX community benchmark trace; peak memory 142.9 GB at this context; duplicate traces of the same session: 3sc75ldl, 7r81ct6o, ebmisc4o, illu4qwu, mvsmdjzz, n523g172, pmt6867t; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1258.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-V4-Flash-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,258 40.8 142.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/amzmvi8t",
   "source_url": "https://omlx.ai/benchmarks/performance/amzmvi8t",
   "tg_tokens": null,
   "tps": 40.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-amzmvi8t-ctx8192",
   "model": "DeepSeek-V4-Flash",
   "notes": "oMLX community benchmark trace; peak memory 142.9 GB at this context; duplicate traces of the same session: 3sc75ldl, 7r81ct6o, ebmisc4o, illu4qwu, mvsmdjzz, n523g172, pmt6867t; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1416.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-V4-Flash-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 1,416 39.5 142.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/amzmvi8t",
   "source_url": "https://omlx.ai/benchmarks/performance/amzmvi8t",
   "tg_tokens": null,
   "tps": 39.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-amzmvi8t-ctx16384",
   "model": "DeepSeek-V4-Flash",
   "notes": "oMLX community benchmark trace; peak memory 143.0 GB at this context; duplicate traces of the same session: 3sc75ldl, 7r81ct6o, ebmisc4o, illu4qwu, mvsmdjzz, n523g172, pmt6867t; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1362.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-V4-Flash-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,362 39.3 143.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/amzmvi8t",
   "source_url": "https://omlx.ai/benchmarks/performance/amzmvi8t",
   "tg_tokens": null,
   "tps": 39.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-amzmvi8t-ctx32768",
   "model": "DeepSeek-V4-Flash",
   "notes": "oMLX community benchmark trace; peak memory 143.1 GB at this context; duplicate traces of the same session: 3sc75ldl, 7r81ct6o, ebmisc4o, illu4qwu, mvsmdjzz, n523g172, pmt6867t; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1278.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-V4-Flash-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,278 37.9 143.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/amzmvi8t",
   "source_url": "https://omlx.ai/benchmarks/performance/amzmvi8t",
   "tg_tokens": null,
   "tps": 37.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-amzmvi8t-ctx65536",
   "model": "DeepSeek-V4-Flash",
   "notes": "oMLX community benchmark trace; peak memory 143.4 GB at this context; duplicate traces of the same session: 3sc75ldl, 7r81ct6o, ebmisc4o, illu4qwu, mvsmdjzz, n523g172, pmt6867t; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1154.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-V4-Flash-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,154 36.4 143.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/amzmvi8t",
   "source_url": "https://omlx.ai/benchmarks/performance/amzmvi8t",
   "tg_tokens": null,
   "tps": 36.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-amzmvi8t-ctx131072",
   "model": "DeepSeek-V4-Flash",
   "notes": "oMLX community benchmark trace; peak memory 143.8 GB at this context; duplicate traces of the same session: 3sc75ldl, 7r81ct6o, ebmisc4o, illu4qwu, mvsmdjzz, n523g172, pmt6867t; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 950.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-V4-Flash-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 950.8 34.0 143.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/amzmvi8t",
   "source_url": "https://omlx.ai/benchmarks/performance/amzmvi8t",
   "tg_tokens": null,
   "tps": 34.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-amzmvi8t-ctx199680",
   "model": "DeepSeek-V4-Flash",
   "notes": "oMLX community benchmark trace; peak memory 144.0 GB at this context; duplicate traces of the same session: 3sc75ldl, 7r81ct6o, ebmisc4o, illu4qwu, mvsmdjzz, n523g172, pmt6867t; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 808.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-V4-Flash-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 808.9 33.7 144.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/amzmvi8t",
   "source_url": "https://omlx.ai/benchmarks/performance/amzmvi8t",
   "tg_tokens": null,
   "tps": 33.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-wggol4kl-ctx1024",
   "model": "LFM2-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: 2rpxecrg, 4553rqbk, 4oly2wxq, yswk15ax; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 924.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2-24B-A2B-MLX-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 1k 924.8 67.6 24.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wggol4kl",
   "source_url": "https://omlx.ai/benchmarks/performance/wggol4kl",
   "tg_tokens": null,
   "tps": 67.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-wggol4kl-ctx4096",
   "model": "LFM2-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: 2rpxecrg, 4553rqbk, 4oly2wxq, yswk15ax; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1044.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2-24B-A2B-MLX-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k 1,044 65.8 24.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wggol4kl",
   "source_url": "https://omlx.ai/benchmarks/performance/wggol4kl",
   "tg_tokens": null,
   "tps": 65.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-wggol4kl-ctx8192",
   "model": "LFM2-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: 2rpxecrg, 4553rqbk, 4oly2wxq, yswk15ax; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 815.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2-24B-A2B-MLX-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 8k 815.4 63.0 25.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wggol4kl",
   "source_url": "https://omlx.ai/benchmarks/performance/wggol4kl",
   "tg_tokens": null,
   "tps": 63.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-wggol4kl-ctx16384",
   "model": "LFM2-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: 2rpxecrg, 4553rqbk, 4oly2wxq, yswk15ax; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 764.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2-24B-A2B-MLX-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 16k 764.3 59.0 26.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wggol4kl",
   "source_url": "https://omlx.ai/benchmarks/performance/wggol4kl",
   "tg_tokens": null,
   "tps": 59.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-wggol4kl-ctx32768",
   "model": "LFM2-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: 2rpxecrg, 4553rqbk, 4oly2wxq, yswk15ax; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 684.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2-24B-A2B-MLX-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 32k 684.0 53.4 28.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wggol4kl",
   "source_url": "https://omlx.ai/benchmarks/performance/wggol4kl",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-63sj0yo9-ctx1024",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 156.3 GB at this context; duplicate traces of the same session: sz7gho11, t7m1ovqm; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1844.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-V2.6-Flash-RL-MLX-4bit-MTP on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,844 84.9 156.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/63sj0yo9",
   "source_url": "https://omlx.ai/benchmarks/performance/63sj0yo9",
   "tg_tokens": null,
   "tps": 84.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-63sj0yo9-ctx4096",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 157.6 GB at this context; duplicate traces of the same session: sz7gho11, t7m1ovqm; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3207.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-V2.6-Flash-RL-MLX-4bit-MTP on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 3,207 104.8 157.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/63sj0yo9",
   "source_url": "https://omlx.ai/benchmarks/performance/63sj0yo9",
   "tg_tokens": null,
   "tps": 104.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-63sj0yo9-ctx8192",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 158.9 GB at this context; duplicate traces of the same session: sz7gho11, t7m1ovqm; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3381.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-V2.6-Flash-RL-MLX-4bit-MTP on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 3,381 92.9 158.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/63sj0yo9",
   "source_url": "https://omlx.ai/benchmarks/performance/63sj0yo9",
   "tg_tokens": null,
   "tps": 92.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-d5cyrurj-ctx32768",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 160.1 GB; TTFT 14071 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2329.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-V2.6-Flash-RL-MLX-4bit-MTP on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k tokens 2,329 PP tok/s 68.6 TG tok/s 14071 TTFT (ms) 160.1 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d5cyrurj",
   "source_url": "https://omlx.ai/benchmarks/performance/d5cyrurj",
   "tg_tokens": null,
   "tps": 68.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-efhr1p75-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 61.7 GB at this context; duplicate traces of the same session: pcduoqjj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1980.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 1,980 76.8 61.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/efhr1p75",
   "source_url": "https://omlx.ai/benchmarks/performance/efhr1p75",
   "tg_tokens": null,
   "tps": 76.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-efhr1p75-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 62.5 GB at this context; duplicate traces of the same session: pcduoqjj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3090.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 3,090 75.4 62.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/efhr1p75",
   "source_url": "https://omlx.ai/benchmarks/performance/efhr1p75",
   "tg_tokens": null,
   "tps": 75.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-rgyehuta-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: 19o17o43; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 955.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MTP-mlx-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 955.0 104.5 35.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rgyehuta",
   "source_url": "https://omlx.ai/benchmarks/performance/rgyehuta",
   "tg_tokens": null,
   "tps": 104.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-rgyehuta-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: 19o17o43; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2910.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MTP-mlx-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 2,910 103.5 36.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rgyehuta",
   "source_url": "https://omlx.ai/benchmarks/performance/rgyehuta",
   "tg_tokens": null,
   "tps": 103.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-u5gdu1i3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: 7s8s40wd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 829.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 829.3 41.3 29.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u5gdu1i3",
   "source_url": "https://omlx.ai/benchmarks/performance/u5gdu1i3",
   "tg_tokens": null,
   "tps": 41.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-u5gdu1i3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.1 GB at this context; duplicate traces of the same session: 7s8s40wd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 869.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 869.0 35.4 31.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u5gdu1i3",
   "source_url": "https://omlx.ai/benchmarks/performance/u5gdu1i3",
   "tg_tokens": null,
   "tps": 35.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-h7877x3q-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: k8h4szd7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 887.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 887.8 58.3 23.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h7877x3q",
   "source_url": "https://omlx.ai/benchmarks/performance/h7877x3q",
   "tg_tokens": null,
   "tps": 58.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-h7877x3q-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: k8h4szd7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 909.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 909.0 51.3 24.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h7877x3q",
   "source_url": "https://omlx.ai/benchmarks/performance/h7877x3q",
   "tg_tokens": null,
   "tps": 51.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ysxvnd7p-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: tlirgnad; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 102.9 22.0 25.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ysxvnd7p",
   "source_url": "https://omlx.ai/benchmarks/performance/ysxvnd7p",
   "tg_tokens": null,
   "tps": 22.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ysxvnd7p-ctx16384",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 26.9 GB at this context; duplicate traces of the same session: tlirgnad; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 98.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 98.3 23.5 26.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ysxvnd7p",
   "source_url": "https://omlx.ai/benchmarks/performance/ysxvnd7p",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qv1almxv-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "oMLX community benchmark trace; peak memory 52.9 GB at this context; duplicate traces of the same session: mb537hgi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 594.1,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.8-27B-MTPLX-bf16 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 bf16; 1k 594.1 21.6 52.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qv1almxv",
   "source_url": "https://omlx.ai/benchmarks/performance/qv1almxv",
   "tg_tokens": null,
   "tps": 21.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qv1almxv-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "oMLX community benchmark trace; peak memory 54.2 GB at this context; duplicate traces of the same session: mb537hgi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 949.2,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.8-27B-MTPLX-bf16 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 bf16; 4k 949.2 20.3 54.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qv1almxv",
   "source_url": "https://omlx.ai/benchmarks/performance/qv1almxv",
   "tg_tokens": null,
   "tps": 20.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wwu3m7k3-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 69.7 GB; TTFT 1866 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 548.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 548.8 PP tok/s 79.1 TG tok/s 1866 TTFT (ms) 69.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wwu3m7k3",
   "source_url": "https://omlx.ai/benchmarks/performance/wwu3m7k3",
   "tg_tokens": null,
   "tps": 79.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ggjb8vzf-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 69.7 GB at this context; duplicate traces of the same session: jw4tkez8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2067.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,067 78.4 69.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ggjb8vzf",
   "source_url": "https://omlx.ai/benchmarks/performance/ggjb8vzf",
   "tg_tokens": null,
   "tps": 78.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ggjb8vzf-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 70.2 GB at this context; duplicate traces of the same session: jw4tkez8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2982.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,982 48.7 70.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ggjb8vzf",
   "source_url": "https://omlx.ai/benchmarks/performance/ggjb8vzf",
   "tg_tokens": null,
   "tps": 48.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-5y09sqpk-ctx1024",
   "model": "DeepSeek-R1-0528-Qwen3-8B",
   "notes": "oMLX community benchmark trace; peak memory 9.2 GB at this context; duplicate traces of the same session: cwiqgz36, fqk99cxg, inhm2sq1, oejyzw0u, se0d0ge4, tvntwlwy, zwnlm2nf; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2213.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "DeepSeek-R1-0528-Qwen3-8B-MLX-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 1k 2,213 101.6 9.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5y09sqpk",
   "source_url": "https://omlx.ai/benchmarks/performance/5y09sqpk",
   "tg_tokens": null,
   "tps": 101.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-5y09sqpk-ctx4096",
   "model": "DeepSeek-R1-0528-Qwen3-8B",
   "notes": "oMLX community benchmark trace; peak memory 9.3 GB at this context; duplicate traces of the same session: cwiqgz36, fqk99cxg, inhm2sq1, oejyzw0u, se0d0ge4, tvntwlwy, zwnlm2nf; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4783.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "DeepSeek-R1-0528-Qwen3-8B-MLX-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 4k 4,783 98.7 9.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5y09sqpk",
   "source_url": "https://omlx.ai/benchmarks/performance/5y09sqpk",
   "tg_tokens": null,
   "tps": 98.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-5y09sqpk-ctx8192",
   "model": "DeepSeek-R1-0528-Qwen3-8B",
   "notes": "oMLX community benchmark trace; peak memory 9.8 GB at this context; duplicate traces of the same session: cwiqgz36, fqk99cxg, inhm2sq1, oejyzw0u, se0d0ge4, tvntwlwy, zwnlm2nf; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4535.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "DeepSeek-R1-0528-Qwen3-8B-MLX-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 8k 4,535 93.2 9.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5y09sqpk",
   "source_url": "https://omlx.ai/benchmarks/performance/5y09sqpk",
   "tg_tokens": null,
   "tps": 93.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-5y09sqpk-ctx16384",
   "model": "DeepSeek-R1-0528-Qwen3-8B",
   "notes": "oMLX community benchmark trace; peak memory 10.9 GB at this context; duplicate traces of the same session: cwiqgz36, fqk99cxg, inhm2sq1, oejyzw0u, se0d0ge4, tvntwlwy, zwnlm2nf; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4070.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "DeepSeek-R1-0528-Qwen3-8B-MLX-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 16k 4,070 83.0 10.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5y09sqpk",
   "source_url": "https://omlx.ai/benchmarks/performance/5y09sqpk",
   "tg_tokens": null,
   "tps": 83.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-5y09sqpk-ctx32768",
   "model": "DeepSeek-R1-0528-Qwen3-8B",
   "notes": "oMLX community benchmark trace; peak memory 13.2 GB at this context; duplicate traces of the same session: cwiqgz36, fqk99cxg, inhm2sq1, oejyzw0u, se0d0ge4, tvntwlwy, zwnlm2nf; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3257.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "DeepSeek-R1-0528-Qwen3-8B-MLX-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 32k 3,257 69.3 13.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5y09sqpk",
   "source_url": "https://omlx.ai/benchmarks/performance/5y09sqpk",
   "tg_tokens": null,
   "tps": 69.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-5y09sqpk-ctx65536",
   "model": "DeepSeek-R1-0528-Qwen3-8B",
   "notes": "oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: cwiqgz36, fqk99cxg, inhm2sq1, oejyzw0u, se0d0ge4, tvntwlwy, zwnlm2nf; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2353.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "DeepSeek-R1-0528-Qwen3-8B-MLX-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 64k 2,353 53.1 17.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5y09sqpk",
   "source_url": "https://omlx.ai/benchmarks/performance/5y09sqpk",
   "tg_tokens": null,
   "tps": 53.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-5y09sqpk-ctx131072",
   "model": "DeepSeek-R1-0528-Qwen3-8B",
   "notes": "oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: cwiqgz36, fqk99cxg, inhm2sq1, oejyzw0u, se0d0ge4, tvntwlwy, zwnlm2nf; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1480.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "DeepSeek-R1-0528-Qwen3-8B-MLX-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 128k 1,480 35.9 26.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5y09sqpk",
   "source_url": "https://omlx.ai/benchmarks/performance/5y09sqpk",
   "tg_tokens": null,
   "tps": 35.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-5y09sqpk-ctx199680",
   "model": "DeepSeek-R1-0528-Qwen3-8B",
   "notes": "oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: cwiqgz36, fqk99cxg, inhm2sq1, oejyzw0u, se0d0ge4, tvntwlwy, zwnlm2nf; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1052.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "DeepSeek-R1-0528-Qwen3-8B-MLX-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 195k 1,052 27.1 36.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5y09sqpk",
   "source_url": "https://omlx.ai/benchmarks/performance/5y09sqpk",
   "tg_tokens": null,
   "tps": 27.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-763r6k18-ctx1024",
   "model": "Huihui-Qwen3.8-27B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: efua5ngc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 806.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Huihui-Qwen3.8-27B-abliterated-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 806.5 41.1 33.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/763r6k18",
   "source_url": "https://omlx.ai/benchmarks/performance/763r6k18",
   "tg_tokens": null,
   "tps": 41.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-763r6k18-ctx4096",
   "model": "Huihui-Qwen3.8-27B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: efua5ngc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 915.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Huihui-Qwen3.8-27B-abliterated-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 915.2 42.8 34.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/763r6k18",
   "source_url": "https://omlx.ai/benchmarks/performance/763r6k18",
   "tg_tokens": null,
   "tps": 42.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-6f8qk80u-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: 0jn8zihu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3911.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 3,911 126.2 37.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6f8qk80u",
   "source_url": "https://omlx.ai/benchmarks/performance/6f8qk80u",
   "tg_tokens": null,
   "tps": 126.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6f8qk80u-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.0 GB at this context; duplicate traces of the same session: 0jn8zihu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5758.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 5,758 114.6 38.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6f8qk80u",
   "source_url": "https://omlx.ai/benchmarks/performance/6f8qk80u",
   "tg_tokens": null,
   "tps": 114.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9n1oiz9r-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: hibsjram; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2356.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 2,356 66.3 37.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9n1oiz9r",
   "source_url": "https://omlx.ai/benchmarks/performance/9n1oiz9r",
   "tg_tokens": null,
   "tps": 66.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9n1oiz9r-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.0 GB at this context; duplicate traces of the same session: hibsjram; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5719.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 5,719 131.2 38.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9n1oiz9r",
   "source_url": "https://omlx.ai/benchmarks/performance/9n1oiz9r",
   "tg_tokens": null,
   "tps": 131.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-yo8cde6n-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 1f9axmr3, 5yw13jhq; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 568.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 1k 568.2 68.9 19.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yo8cde6n",
   "source_url": "https://omlx.ai/benchmarks/performance/yo8cde6n",
   "tg_tokens": null,
   "tps": 68.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-yo8cde6n-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 1f9axmr3, 5yw13jhq; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 574.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k 574.2 66.8 20.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yo8cde6n",
   "source_url": "https://omlx.ai/benchmarks/performance/yo8cde6n",
   "tg_tokens": null,
   "tps": 66.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-yo8cde6n-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: 1f9axmr3, 5yw13jhq; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 524.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 8k 524.8 61.1 20.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yo8cde6n",
   "source_url": "https://omlx.ai/benchmarks/performance/yo8cde6n",
   "tg_tokens": null,
   "tps": 61.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-v30ha6at-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 53.7 GB at this context; duplicate traces of the same session: 0g2y1dzw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 858.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.8-27B on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 858.0 28.5 53.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v30ha6at",
   "source_url": "https://omlx.ai/benchmarks/performance/v30ha6at",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-v30ha6at-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 55.0 GB at this context; duplicate traces of the same session: 0g2y1dzw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1040.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.8-27B on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 4k 1,040 23.5 55.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v30ha6at",
   "source_url": "https://omlx.ai/benchmarks/performance/v30ha6at",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-umb1hgue-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: zl2sy735; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 956.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 956.7 32.8 17.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/umb1hgue",
   "source_url": "https://omlx.ai/benchmarks/performance/umb1hgue",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-umb1hgue-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: zl2sy735; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 902.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 902.7 30.8 17.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/umb1hgue",
   "source_url": "https://omlx.ai/benchmarks/performance/umb1hgue",
   "tg_tokens": null,
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-wjtnfnwx-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 18.5 GB; TTFT 4693 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 218.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 218.2 PP tok/s 39.6 TG tok/s 4693 TTFT (ms) 18.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wjtnfnwx",
   "source_url": "https://omlx.ai/benchmarks/performance/wjtnfnwx",
   "tg_tokens": null,
   "tps": 39.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-973ic7gc-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 1j10g1c1; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1108.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 6bit; 1k 1,108 36.6 22.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/973ic7gc",
   "source_url": "https://omlx.ai/benchmarks/performance/973ic7gc",
   "tg_tokens": null,
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-973ic7gc-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: 1j10g1c1; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1378.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 6bit; 8k 1,378 35.8 24.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/973ic7gc",
   "source_url": "https://omlx.ai/benchmarks/performance/973ic7gc",
   "tg_tokens": null,
   "tps": 35.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-8zptusyq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: nznkc955; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1139.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 6bit; 1k 1,139 38.9 26.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8zptusyq",
   "source_url": "https://omlx.ai/benchmarks/performance/8zptusyq",
   "tg_tokens": null,
   "tps": 38.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-8zptusyq-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: nznkc955; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1337.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 6bit; 8k 1,337 80.9 28.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8zptusyq",
   "source_url": "https://omlx.ai/benchmarks/performance/8zptusyq",
   "tg_tokens": null,
   "tps": 80.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-278ku5lu-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.2 GB at this context; duplicate traces of the same session: 0a62lwjg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 803.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 803.6 51.0 39.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/278ku5lu",
   "source_url": "https://omlx.ai/benchmarks/performance/278ku5lu",
   "tg_tokens": null,
   "tps": 51.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-278ku5lu-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 40.2 GB at this context; duplicate traces of the same session: 0a62lwjg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 730.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 730.3 43.4 40.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/278ku5lu",
   "source_url": "https://omlx.ai/benchmarks/performance/278ku5lu",
   "tg_tokens": null,
   "tps": 43.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0ozme02u-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 70.3 GB; TTFT 703 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1458.0,
   "provenance": "community",
   "quant": "GSQ-RCO-Q2",
   "quote": "Qwen3.8-Flash-Next-GSQ-RCO-Q2-mlx on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,458 PP tok/s 71.3 TG tok/s 703 TTFT (ms) 70.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ozme02u",
   "source_url": "https://omlx.ai/benchmarks/performance/0ozme02u",
   "tg_tokens": null,
   "tps": 71.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-aq30ozj1-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 74.8 GB at this context; duplicate traces of the same session: fmqkjlqg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 606.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 606.5 53.4 74.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aq30ozj1",
   "source_url": "https://omlx.ai/benchmarks/performance/aq30ozj1",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-aq30ozj1-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 78.8 GB at this context; duplicate traces of the same session: fmqkjlqg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 537.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 537.1 52.4 78.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aq30ozj1",
   "source_url": "https://omlx.ai/benchmarks/performance/aq30ozj1",
   "tg_tokens": null,
   "tps": 52.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-gcye8d5f-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 75.0 GB at this context; duplicate traces of the same session: 0qz33djc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 635.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 635.1 48.1 75.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gcye8d5f",
   "source_url": "https://omlx.ai/benchmarks/performance/gcye8d5f",
   "tg_tokens": null,
   "tps": 48.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-gcye8d5f-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 78.8 GB at this context; duplicate traces of the same session: 0qz33djc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 556.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 556.5 46.2 78.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gcye8d5f",
   "source_url": "https://omlx.ai/benchmarks/performance/gcye8d5f",
   "tg_tokens": null,
   "tps": 46.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-qdxzm536-ctx1024",
   "model": "gpt-oss 120B",
   "notes": "MXFP4 experts with Q8 attention; oMLX community benchmark trace; peak memory 59.5 GB at this context; duplicate traces of the same session: en1xtexc, xiuzsnqb; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 858.1,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-120b-MXFP4-Q8 on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k 858.1 97.2 59.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qdxzm536",
   "source_url": "https://omlx.ai/benchmarks/performance/qdxzm536",
   "tg_tokens": null,
   "tps": 97.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-qdxzm536-ctx8192",
   "model": "gpt-oss 120B",
   "notes": "MXFP4 experts with Q8 attention; oMLX community benchmark trace; peak memory 59.7 GB at this context; duplicate traces of the same session: en1xtexc, xiuzsnqb; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3148.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-120b-MXFP4-Q8 on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 8k 3,148 90.2 59.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qdxzm536",
   "source_url": "https://omlx.ai/benchmarks/performance/qdxzm536",
   "tg_tokens": null,
   "tps": 90.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-qdxzm536-ctx32768",
   "model": "gpt-oss 120B",
   "notes": "MXFP4 experts with Q8 attention; oMLX community benchmark trace; peak memory 60.5 GB at this context; duplicate traces of the same session: en1xtexc, xiuzsnqb; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2941.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-120b-MXFP4-Q8 on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 32k 2,941 75.4 60.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qdxzm536",
   "source_url": "https://omlx.ai/benchmarks/performance/qdxzm536",
   "tg_tokens": null,
   "tps": 75.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-0jnagrn6-ctx8192",
   "model": "FrogNano-4B-2609",
   "notes": "oMLX community benchmark trace, single run at 8k context; peak memory 12.1 GB; TTFT 13771 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 594.9,
   "provenance": "community",
   "quant": "4-bit",
   "quote": "FrogNano-4B-2609 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 unknown; 8k tokens 594.9 PP tok/s 50.5 TG tok/s 13771 TTFT (ms) 12.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0jnagrn6",
   "source_url": "https://omlx.ai/benchmarks/performance/0jnagrn6",
   "tg_tokens": null,
   "tps": 50.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-87mgzxh3-ctx32768",
   "model": "GLM-4.7-Flash",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 41.5 GB; TTFT 62989 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 520.2,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-4.7-Flash-oQ4 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k tokens 520.2 PP tok/s 37.5 TG tok/s 62989 TTFT (ms) 41.5 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/87mgzxh3",
   "source_url": "https://omlx.ai/benchmarks/performance/87mgzxh3",
   "tg_tokens": null,
   "tps": 37.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-k4t40qjn-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: 0jtdyg0k; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 196.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M3 (10c); M3 (10c) \u00b7 16 GB \u00b7 4bit; 1k 196.0 19.0 5.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k4t40qjn",
   "source_url": "https://omlx.ai/benchmarks/performance/k4t40qjn",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-k4t40qjn-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: 0jtdyg0k; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 195.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M3 (10c); M3 (10c) \u00b7 16 GB \u00b7 4bit; 4k 195.7 17.0 6.3 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k4t40qjn",
   "source_url": "https://omlx.ai/benchmarks/performance/k4t40qjn",
   "tg_tokens": null,
   "tps": 17.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-3jadgtgd-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 34.5 GB at this context; duplicate traces of the same session: 6fa141fw, f86e6qvd, re1bakhi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1631.0,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Qwen3.6-35B-A3B-mxfp8 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 1,631 96.4 34.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3jadgtgd",
   "source_url": "https://omlx.ai/benchmarks/performance/3jadgtgd",
   "tg_tokens": null,
   "tps": 96.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-3jadgtgd-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 36.4 GB at this context; duplicate traces of the same session: 6fa141fw, f86e6qvd, re1bakhi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2042.0,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Qwen3.6-35B-A3B-mxfp8 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 2,042 94.1 36.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3jadgtgd",
   "source_url": "https://omlx.ai/benchmarks/performance/3jadgtgd",
   "tg_tokens": null,
   "tps": 94.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-3jadgtgd-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 36.5 GB at this context; duplicate traces of the same session: 6fa141fw, f86e6qvd, re1bakhi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1947.0,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Qwen3.6-35B-A3B-mxfp8 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 8k 1,947 89.8 36.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3jadgtgd",
   "source_url": "https://omlx.ai/benchmarks/performance/3jadgtgd",
   "tg_tokens": null,
   "tps": 89.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-3jadgtgd-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: 6fa141fw, f86e6qvd, re1bakhi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1717.0,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Qwen3.6-35B-A3B-mxfp8 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 16k 1,717 85.4 36.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3jadgtgd",
   "source_url": "https://omlx.ai/benchmarks/performance/3jadgtgd",
   "tg_tokens": null,
   "tps": 85.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-uakpbeow-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 21.4 GB; TTFT 695 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1473.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 1,473 PP tok/s 148.2 TG tok/s 695 TTFT (ms) 21.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uakpbeow",
   "source_url": "https://omlx.ai/benchmarks/performance/uakpbeow",
   "tg_tokens": null,
   "tps": 148.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-o2w8b3cr-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 20.1 GB; TTFT 665 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1540.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 1,540 PP tok/s 100.5 TG tok/s 665 TTFT (ms) 20.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o2w8b3cr",
   "source_url": "https://omlx.ai/benchmarks/performance/o2w8b3cr",
   "tg_tokens": null,
   "tps": 100.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-e9n1o51f-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: 4hkkse8e, s7h0gkke; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 901.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 901.1 60.4 17.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e9n1o51f",
   "source_url": "https://omlx.ai/benchmarks/performance/e9n1o51f",
   "tg_tokens": null,
   "tps": 60.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-e9n1o51f-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 4hkkse8e, s7h0gkke; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 984.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 984.6 74.6 18.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e9n1o51f",
   "source_url": "https://omlx.ai/benchmarks/performance/e9n1o51f",
   "tg_tokens": null,
   "tps": 74.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-e9n1o51f-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 4hkkse8e, s7h0gkke; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 902.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 902.3 74.7 19.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e9n1o51f",
   "source_url": "https://omlx.ai/benchmarks/performance/e9n1o51f",
   "tg_tokens": null,
   "tps": 74.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lt79wdh9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 33gqshcp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1133.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,133 59.5 17.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lt79wdh9",
   "source_url": "https://omlx.ai/benchmarks/performance/lt79wdh9",
   "tg_tokens": null,
   "tps": 59.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lt79wdh9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 33gqshcp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1376.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,376 61.1 19.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lt79wdh9",
   "source_url": "https://omlx.ai/benchmarks/performance/lt79wdh9",
   "tg_tokens": null,
   "tps": 61.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-i6zpfhim-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 195k context; peak memory 71.7 GB; TTFT 69237 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2889.0,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ3-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 3bit; 195k tokens 2,889 PP tok/s 57.3 TG tok/s 69237 TTFT (ms) 71.7 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i6zpfhim",
   "source_url": "https://omlx.ai/benchmarks/performance/i6zpfhim",
   "tg_tokens": null,
   "tps": 57.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-u020y2m2-ctx4096",
   "model": "Qwen3.8-27B-Uncensored-HauhauCS-Aggressive",
   "notes": "itrejomx build; oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: 99w1f3j6; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 409.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "itrejomx--Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTPLX-6bit on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 4k 409.8 26.7 25.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u020y2m2",
   "source_url": "https://omlx.ai/benchmarks/performance/u020y2m2",
   "tg_tokens": null,
   "tps": 26.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-u020y2m2-ctx16384",
   "model": "Qwen3.8-27B-Uncensored-HauhauCS-Aggressive",
   "notes": "itrejomx build; oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: 99w1f3j6; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 385.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "itrejomx--Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTPLX-6bit on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 16k 385.7 25.3 27.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u020y2m2",
   "source_url": "https://omlx.ai/benchmarks/performance/u020y2m2",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-o0idksk6-ctx4096",
   "model": "Qwen3.8-27B-Uncensored-HauhauCS-Aggressive",
   "notes": "itrejomx build; oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: 41i6lidz; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 408.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "itrejomx--Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTPLX-6bit on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 4k 408.5 26.5 25.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o0idksk6",
   "source_url": "https://omlx.ai/benchmarks/performance/o0idksk6",
   "tg_tokens": null,
   "tps": 26.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-o0idksk6-ctx16384",
   "model": "Qwen3.8-27B-Uncensored-HauhauCS-Aggressive",
   "notes": "itrejomx build; oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: 41i6lidz; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 384.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "itrejomx--Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTPLX-6bit on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 16k 384.9 25.2 27.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o0idksk6",
   "source_url": "https://omlx.ai/benchmarks/performance/o0idksk6",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-fgp62jkq-ctx4096",
   "model": "Qwen3.8-27B-Uncensored-HauhauCS-Aggressive",
   "notes": "itrejomx build; oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: jvhkig90; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 393.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "itrejomx--Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTPLX-6bit on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 4k 393.9 26.5 25.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fgp62jkq",
   "source_url": "https://omlx.ai/benchmarks/performance/fgp62jkq",
   "tg_tokens": null,
   "tps": 26.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-fgp62jkq-ctx16384",
   "model": "Qwen3.8-27B-Uncensored-HauhauCS-Aggressive",
   "notes": "itrejomx build; oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: jvhkig90; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 377.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "itrejomx--Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTPLX-6bit on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 6bit; 16k 377.3 25.3 27.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fgp62jkq",
   "source_url": "https://omlx.ai/benchmarks/performance/fgp62jkq",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-2xwqwpds-ctx1024",
   "model": "DeepSeek-R1-0528-Qwen3-8B",
   "notes": "oMLX community benchmark trace; peak memory 8.9 GB at this context; duplicate traces of the same session: mg8rmzi3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1542.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "DeepSeek-R1-0528-Qwen3-8B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 8bit; 1k 1,542 30.9 8.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2xwqwpds",
   "source_url": "https://omlx.ai/benchmarks/performance/2xwqwpds",
   "tg_tokens": null,
   "tps": 30.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-2xwqwpds-ctx4096",
   "model": "DeepSeek-R1-0528-Qwen3-8B",
   "notes": "oMLX community benchmark trace; peak memory 9.3 GB at this context; duplicate traces of the same session: mg8rmzi3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1314.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "DeepSeek-R1-0528-Qwen3-8B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 8bit; 4k 1,314 27.0 9.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2xwqwpds",
   "source_url": "https://omlx.ai/benchmarks/performance/2xwqwpds",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-99s60m5a-ctx32768",
   "model": "Ling-3.0-tiny",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 33.4 GB; TTFT 8092 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4049.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ling-3.0-tiny on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 unknown; 32k tokens 4,049 PP tok/s 102.9 TG tok/s 8092 TTFT (ms) 33.4 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/99s60m5a",
   "source_url": "https://omlx.ai/benchmarks/performance/99s60m5a",
   "tg_tokens": null,
   "tps": 102.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-etzvvebp-ctx32768",
   "model": "Ling-3.0-tiny",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 33.4 GB; TTFT 8123 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4034.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ling-3.0-tiny on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 unknown; 32k tokens 4,034 PP tok/s 101.8 TG tok/s 8123 TTFT (ms) 33.4 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/etzvvebp",
   "source_url": "https://omlx.ai/benchmarks/performance/etzvvebp",
   "tg_tokens": null,
   "tps": 101.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-fxpqgr4c-ctx32768",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 35.8 GB; TTFT 30483 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1075.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "MiMo-V2.6-Distill-Qwen-9B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 unknown; 32k tokens 1,075 PP tok/s 37.7 TG tok/s 30483 TTFT (ms) 35.8 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fxpqgr4c",
   "source_url": "https://omlx.ai/benchmarks/performance/fxpqgr4c",
   "tg_tokens": null,
   "tps": 37.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jnemfkzv-ctx32768",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 25.8 GB; TTFT 30342 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1080.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-OptiQ-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k tokens 1,080 PP tok/s 77.6 TG tok/s 30342 TTFT (ms) 25.8 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jnemfkzv",
   "source_url": "https://omlx.ai/benchmarks/performance/jnemfkzv",
   "tg_tokens": null,
   "tps": 77.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-woc80ojw-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 16k context; peak memory 37.6 GB; TTFT 13774 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1190.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 8bit; 16k tokens 1,190 PP tok/s 33.4 TG tok/s 13774 TTFT (ms) 37.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/woc80ojw",
   "source_url": "https://omlx.ai/benchmarks/performance/woc80ojw",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-f64hh7vu-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: d3848w3n; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 305.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 305.9 36.9 33.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f64hh7vu",
   "source_url": "https://omlx.ai/benchmarks/performance/f64hh7vu",
   "tg_tokens": null,
   "tps": 36.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-f64hh7vu-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: d3848w3n; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 334.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 334.3 47.7 37.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f64hh7vu",
   "source_url": "https://omlx.ai/benchmarks/performance/f64hh7vu",
   "tg_tokens": null,
   "tps": 47.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-qakki22e-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 57.3 GB; TTFT 5965 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 171.7,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ3e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 3bit; 1k tokens 171.7 PP tok/s 8.2 TG tok/s 5965 TTFT (ms) 57.3 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qakki22e",
   "source_url": "https://omlx.ai/benchmarks/performance/qakki22e",
   "tg_tokens": null,
   "tps": 8.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-09shw58c-ctx1024",
   "model": "FrogNano-4B-2609",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.5 GB at this context; duplicate traces of the same session: hap6w5lp; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 725.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "FrogNano-4B-2609-oQ8e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 1k 725.2 57.8 7.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/09shw58c",
   "source_url": "https://omlx.ai/benchmarks/performance/09shw58c",
   "tg_tokens": null,
   "tps": 57.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-09shw58c-ctx4096",
   "model": "FrogNano-4B-2609",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 9.1 GB at this context; duplicate traces of the same session: hap6w5lp; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 736.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "FrogNano-4B-2609-oQ8e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 4k 736.1 64.3 9.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/09shw58c",
   "source_url": "https://omlx.ai/benchmarks/performance/09shw58c",
   "tg_tokens": null,
   "tps": 64.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-x3jo1bb1-ctx131072",
   "model": "FrogNano-4B-2609",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 17.0 GB; TTFT 479645 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 273.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "FrogNano-4B-2609-oQ8e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 128k tokens 273.3 PP tok/s 31.9 TG tok/s 479645 TTFT (ms) 17.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x3jo1bb1",
   "source_url": "https://omlx.ai/benchmarks/performance/x3jo1bb1",
   "tg_tokens": null,
   "tps": 31.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-006pf7l4-ctx32768",
   "model": "MiniCPM5-1B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 18.2 GB; TTFT 9196 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3563.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM5-1B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 unknown; 32k tokens 3,563 PP tok/s 164.8 TG tok/s 9196 TTFT (ms) 18.2 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/006pf7l4",
   "source_url": "https://omlx.ai/benchmarks/performance/006pf7l4",
   "tg_tokens": null,
   "tps": 164.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6h0kwbr5-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.8 GB at this context; duplicate traces of the same session: 9fryuzvn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 862.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 862.7 33.7 28.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6h0kwbr5",
   "source_url": "https://omlx.ai/benchmarks/performance/6h0kwbr5",
   "tg_tokens": null,
   "tps": 33.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6h0kwbr5-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.8 GB at this context; duplicate traces of the same session: 9fryuzvn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 995.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 995.4 33.1 29.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6h0kwbr5",
   "source_url": "https://omlx.ai/benchmarks/performance/6h0kwbr5",
   "tg_tokens": null,
   "tps": 33.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-kv0cyoqv-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 6aw9q77w; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 223.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 1k 223.3 34.6 19.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kv0cyoqv",
   "source_url": "https://omlx.ai/benchmarks/performance/kv0cyoqv",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-kv0cyoqv-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: 6aw9q77w; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 240.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 4k 240.1 33.3 23.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kv0cyoqv",
   "source_url": "https://omlx.ai/benchmarks/performance/kv0cyoqv",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-5ya9b5cd-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: kwvzeqq2; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 217.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 1k 217.8 29.0 19.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ya9b5cd",
   "source_url": "https://omlx.ai/benchmarks/performance/5ya9b5cd",
   "tg_tokens": null,
   "tps": 29.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-5ya9b5cd-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: kwvzeqq2; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 238.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 4k 238.8 25.9 23.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ya9b5cd",
   "source_url": "https://omlx.ai/benchmarks/performance/5ya9b5cd",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-j8cj3it3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 1enjwl8u, bv5ok7j3, ul5uoka1; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 158.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 1k 158.0 17.7 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j8cj3it3",
   "source_url": "https://omlx.ai/benchmarks/performance/j8cj3it3",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-j8cj3it3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: 1enjwl8u, bv5ok7j3, ul5uoka1; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 163.3,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k 163.3 18.3 23.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j8cj3it3",
   "source_url": "https://omlx.ai/benchmarks/performance/j8cj3it3",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-j8cj3it3-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: 1enjwl8u, bv5ok7j3, ul5uoka1; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 156.8,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 8k 156.8 23.2 24.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j8cj3it3",
   "source_url": "https://omlx.ai/benchmarks/performance/j8cj3it3",
   "tg_tokens": null,
   "tps": 23.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-j8cj3it3-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: 1enjwl8u, bv5ok7j3, ul5uoka1; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 151.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 16k 151.0 28.1 25.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j8cj3it3",
   "source_url": "https://omlx.ai/benchmarks/performance/j8cj3it3",
   "tg_tokens": null,
   "tps": 28.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pl1x8t76-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 0njljqbn, 1c7kihc2, dj0tub2p, ks8j4a93, oiggi4cp, zak1rda4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 892.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 892.4 49.2 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pl1x8t76",
   "source_url": "https://omlx.ai/benchmarks/performance/pl1x8t76",
   "tg_tokens": null,
   "tps": 49.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pl1x8t76-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 0njljqbn, 1c7kihc2, dj0tub2p, ks8j4a93, oiggi4cp, zak1rda4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1375.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,375 59.1 19.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pl1x8t76",
   "source_url": "https://omlx.ai/benchmarks/performance/pl1x8t76",
   "tg_tokens": null,
   "tps": 59.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pl1x8t76-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 0njljqbn, 1c7kihc2, dj0tub2p, ks8j4a93, oiggi4cp, zak1rda4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1413.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,413 67.6 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pl1x8t76",
   "source_url": "https://omlx.ai/benchmarks/performance/pl1x8t76",
   "tg_tokens": null,
   "tps": 67.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-pl1x8t76-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 0njljqbn, 1c7kihc2, dj0tub2p, ks8j4a93, oiggi4cp, zak1rda4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1383.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,383 85.3 19.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pl1x8t76",
   "source_url": "https://omlx.ai/benchmarks/performance/pl1x8t76",
   "tg_tokens": null,
   "tps": 85.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pl1x8t76-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 0njljqbn, 1c7kihc2, dj0tub2p, ks8j4a93, oiggi4cp, zak1rda4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1263.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 1,263 78.9 21.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pl1x8t76",
   "source_url": "https://omlx.ai/benchmarks/performance/pl1x8t76",
   "tg_tokens": null,
   "tps": 78.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-pl1x8t76-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 0njljqbn, 1c7kihc2, dj0tub2p, ks8j4a93, oiggi4cp, zak1rda4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1056.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 1,056 58.9 23.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pl1x8t76",
   "source_url": "https://omlx.ai/benchmarks/performance/pl1x8t76",
   "tg_tokens": null,
   "tps": 58.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-pl1x8t76-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: 0njljqbn, 1c7kihc2, dj0tub2p, ks8j4a93, oiggi4cp, zak1rda4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 782.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 782.0 51.7 28.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pl1x8t76",
   "source_url": "https://omlx.ai/benchmarks/performance/pl1x8t76",
   "tg_tokens": null,
   "tps": 51.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-e9vr4f9l-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.4 GB at this context; duplicate traces of the same session: 7in01epq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 318.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 318.0 58.1 38.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e9vr4f9l",
   "source_url": "https://omlx.ai/benchmarks/performance/e9vr4f9l",
   "tg_tokens": null,
   "tps": 58.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-e9vr4f9l-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 41.7 GB at this context; duplicate traces of the same session: 7in01epq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 393.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 393.7 47.6 41.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e9vr4f9l",
   "source_url": "https://omlx.ai/benchmarks/performance/e9vr4f9l",
   "tg_tokens": null,
   "tps": 47.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-vc71up3m-ctx1024",
   "model": "Qwen3.8-35B-A3B-Distill-Heretic",
   "notes": "MTP speculative decoding enabled; fp16 build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: ji4cgf9h; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 740.2,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-35B-A3B-Distill-Heretic-oQ4-fp16-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 740.2 71.5 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vc71up3m",
   "source_url": "https://omlx.ai/benchmarks/performance/vc71up3m",
   "tg_tokens": null,
   "tps": 71.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-vc71up3m-ctx4096",
   "model": "Qwen3.8-35B-A3B-Distill-Heretic",
   "notes": "MTP speculative decoding enabled; fp16 build; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: ji4cgf9h; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 942.3,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-35B-A3B-Distill-Heretic-oQ4-fp16-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 942.3 69.3 21.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vc71up3m",
   "source_url": "https://omlx.ai/benchmarks/performance/vc71up3m",
   "tg_tokens": null,
   "tps": 69.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mashpk83-ctx16384",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 71.4 GB at this context; duplicate traces of the same session: 0gocq6gf, 9ldszff3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1317.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,317 55.2 71.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mashpk83",
   "source_url": "https://omlx.ai/benchmarks/performance/mashpk83",
   "tg_tokens": null,
   "tps": 55.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mashpk83-ctx32768",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 0gocq6gf, 9ldszff3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1357.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,357 43.1 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mashpk83",
   "source_url": "https://omlx.ai/benchmarks/performance/mashpk83",
   "tg_tokens": null,
   "tps": 43.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mashpk83-ctx65536",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 72.9 GB at this context; duplicate traces of the same session: 0gocq6gf, 9ldszff3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1343.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 1,343 45.4 72.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mashpk83",
   "source_url": "https://omlx.ai/benchmarks/performance/mashpk83",
   "tg_tokens": null,
   "tps": 45.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fbhrt5hf-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: 8b2q7g52, ktoscn7f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1028.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,028 55.9 70.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fbhrt5hf",
   "source_url": "https://omlx.ai/benchmarks/performance/fbhrt5hf",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fbhrt5hf-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.4 GB at this context; duplicate traces of the same session: 8b2q7g52, ktoscn7f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1402.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,402 56.1 71.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fbhrt5hf",
   "source_url": "https://omlx.ai/benchmarks/performance/fbhrt5hf",
   "tg_tokens": null,
   "tps": 56.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fbhrt5hf-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.4 GB at this context; duplicate traces of the same session: 8b2q7g52, ktoscn7f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1413.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 1,413 53.5 72.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fbhrt5hf",
   "source_url": "https://omlx.ai/benchmarks/performance/fbhrt5hf",
   "tg_tokens": null,
   "tps": 53.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-d03h17ud-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: gaoug17k, h5hxs95t, ifs1mocm, kmv7detx, q9etv219, w0geyaws; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1382.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,382 78.3 16.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d03h17ud",
   "source_url": "https://omlx.ai/benchmarks/performance/d03h17ud",
   "tg_tokens": null,
   "tps": 78.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-d03h17ud-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: gaoug17k, h5hxs95t, ifs1mocm, kmv7detx, q9etv219, w0geyaws; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4248.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 4,248 102.8 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d03h17ud",
   "source_url": "https://omlx.ai/benchmarks/performance/d03h17ud",
   "tg_tokens": null,
   "tps": 102.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-d03h17ud-ctx8192",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: gaoug17k, h5hxs95t, ifs1mocm, kmv7detx, q9etv219, w0geyaws; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4155.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 4,155 117.1 16.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d03h17ud",
   "source_url": "https://omlx.ai/benchmarks/performance/d03h17ud",
   "tg_tokens": null,
   "tps": 117.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-d03h17ud-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: gaoug17k, h5hxs95t, ifs1mocm, kmv7detx, q9etv219, w0geyaws; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3949.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 3,949 113.9 17.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d03h17ud",
   "source_url": "https://omlx.ai/benchmarks/performance/d03h17ud",
   "tg_tokens": null,
   "tps": 113.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-d03h17ud-ctx32768",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: gaoug17k, h5hxs95t, ifs1mocm, kmv7detx, q9etv219, w0geyaws; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3496.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 3,496 93.3 18.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d03h17ud",
   "source_url": "https://omlx.ai/benchmarks/performance/d03h17ud",
   "tg_tokens": null,
   "tps": 93.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-d03h17ud-ctx65536",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: gaoug17k, h5hxs95t, ifs1mocm, kmv7detx, q9etv219, w0geyaws; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2770.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 2,770 76.6 19.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d03h17ud",
   "source_url": "https://omlx.ai/benchmarks/performance/d03h17ud",
   "tg_tokens": null,
   "tps": 76.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-d03h17ud-ctx131072",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: gaoug17k, h5hxs95t, ifs1mocm, kmv7detx, q9etv219, w0geyaws; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1888.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 1,888 51.4 23.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d03h17ud",
   "source_url": "https://omlx.ai/benchmarks/performance/d03h17ud",
   "tg_tokens": null,
   "tps": 51.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-8m5axbpe-ctx1024",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 2.0 GB at this context; duplicate traces of the same session: 2oc6pri2; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 326.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiniCPM5-2B-oQ4e on M1 (8c); M1 (8c) \u00b7 8 GB \u00b7 4bit; 1k 326.5 37.3 2.0 GB view",
   "ram_gb": 8,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8m5axbpe",
   "source_url": "https://omlx.ai/benchmarks/performance/8m5axbpe",
   "tg_tokens": null,
   "tps": 37.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-8m5axbpe-ctx4096",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 2.1 GB at this context; duplicate traces of the same session: 2oc6pri2; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 290.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiniCPM5-2B-oQ4e on M1 (8c); M1 (8c) \u00b7 8 GB \u00b7 4bit; 4k 290.6 32.1 2.1 GB current",
   "ram_gb": 8,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8m5axbpe",
   "source_url": "https://omlx.ai/benchmarks/performance/8m5axbpe",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-57hbk7ff-ctx1024",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: ic9nqgjg; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1475.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 1,475 67.3 19.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/57hbk7ff",
   "source_url": "https://omlx.ai/benchmarks/performance/57hbk7ff",
   "tg_tokens": null,
   "tps": 67.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-57hbk7ff-ctx4096",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: ic9nqgjg; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1608.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 1,608 67.2 20.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/57hbk7ff",
   "source_url": "https://omlx.ai/benchmarks/performance/57hbk7ff",
   "tg_tokens": null,
   "tps": 67.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-jnpuus7e-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: ehqc85q7, pjpynyfa; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 487.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 487.6 26.2 17.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jnpuus7e",
   "source_url": "https://omlx.ai/benchmarks/performance/jnpuus7e",
   "tg_tokens": null,
   "tps": 26.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-jnpuus7e-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: ehqc85q7, pjpynyfa; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 591.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 591.9 26.8 19.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jnpuus7e",
   "source_url": "https://omlx.ai/benchmarks/performance/jnpuus7e",
   "tg_tokens": null,
   "tps": 26.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-jnpuus7e-ctx8192",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: ehqc85q7, pjpynyfa; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 604.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 604.7 27.9 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jnpuus7e",
   "source_url": "https://omlx.ai/benchmarks/performance/jnpuus7e",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xmsa9e0b-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: 6wzhutt0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 837.5,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 837.5 50.9 26.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xmsa9e0b",
   "source_url": "https://omlx.ai/benchmarks/performance/xmsa9e0b",
   "tg_tokens": null,
   "tps": 50.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xmsa9e0b-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: 6wzhutt0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 731.6,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 731.6 53.4 27.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xmsa9e0b",
   "source_url": "https://omlx.ai/benchmarks/performance/xmsa9e0b",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-efmxlw2h-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 38.9 GB at this context; duplicate traces of the same session: 1ip51x2u; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 159.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 4k 159.9 19.5 38.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/efmxlw2h",
   "source_url": "https://omlx.ai/benchmarks/performance/efmxlw2h",
   "tg_tokens": null,
   "tps": 19.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-efmxlw2h-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 40.7 GB at this context; duplicate traces of the same session: 1ip51x2u; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 149.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 16k 149.3 25.5 40.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/efmxlw2h",
   "source_url": "https://omlx.ai/benchmarks/performance/efmxlw2h",
   "tg_tokens": null,
   "tps": 25.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1hetg1fu-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: gdo23o6y; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 318.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 318.4 54.5 33.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1hetg1fu",
   "source_url": "https://omlx.ai/benchmarks/performance/1hetg1fu",
   "tg_tokens": null,
   "tps": 54.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1hetg1fu-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: gdo23o6y; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 334.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 334.3 49.7 37.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1hetg1fu",
   "source_url": "https://omlx.ai/benchmarks/performance/1hetg1fu",
   "tg_tokens": null,
   "tps": 49.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lnr6fu0k-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: 0ybs8ngz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 316.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 316.8 40.6 29.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lnr6fu0k",
   "source_url": "https://omlx.ai/benchmarks/performance/lnr6fu0k",
   "tg_tokens": null,
   "tps": 40.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lnr6fu0k-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.1 GB at this context; duplicate traces of the same session: 0ybs8ngz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 331.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 331.9 39.3 31.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lnr6fu0k",
   "source_url": "https://omlx.ai/benchmarks/performance/lnr6fu0k",
   "tg_tokens": null,
   "tps": 39.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-fpmpdj9g-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.4 GB at this context; duplicate traces of the same session: 28w3alcs; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 318.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 318.2 50.4 38.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fpmpdj9g",
   "source_url": "https://omlx.ai/benchmarks/performance/fpmpdj9g",
   "tg_tokens": null,
   "tps": 50.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-fpmpdj9g-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 41.7 GB at this context; duplicate traces of the same session: 28w3alcs; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 410.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 410.8 51.9 41.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fpmpdj9g",
   "source_url": "https://omlx.ai/benchmarks/performance/fpmpdj9g",
   "tg_tokens": null,
   "tps": 51.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2sd31dmj-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: amcf90f1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 315.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 315.6 34.8 29.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2sd31dmj",
   "source_url": "https://omlx.ai/benchmarks/performance/2sd31dmj",
   "tg_tokens": null,
   "tps": 34.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2sd31dmj-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.1 GB at this context; duplicate traces of the same session: amcf90f1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 332.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 332.2 37.8 31.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2sd31dmj",
   "source_url": "https://omlx.ai/benchmarks/performance/2sd31dmj",
   "tg_tokens": null,
   "tps": 37.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ztoaijbi-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: 7inubk47; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1809.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,809 68.9 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ztoaijbi",
   "source_url": "https://omlx.ai/benchmarks/performance/ztoaijbi",
   "tg_tokens": null,
   "tps": 68.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ztoaijbi-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.7 GB at this context; duplicate traces of the same session: 7inubk47; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2527.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,527 69.7 72.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ztoaijbi",
   "source_url": "https://omlx.ai/benchmarks/performance/ztoaijbi",
   "tg_tokens": null,
   "tps": 69.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mr8amf2z-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 70.9 GB; TTFT 2454 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1670.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 1,670 PP tok/s 64.3 TG tok/s 2454 TTFT (ms) 70.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mr8amf2z",
   "source_url": "https://omlx.ai/benchmarks/performance/mr8amf2z",
   "tg_tokens": null,
   "tps": 64.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jsfrs9qk-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.5 GB at this context; duplicate traces of the same session: jbpfdklz, qb4gsw26; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1109.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,109 81.0 70.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jsfrs9qk",
   "source_url": "https://omlx.ai/benchmarks/performance/jsfrs9qk",
   "tg_tokens": null,
   "tps": 81.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jsfrs9qk-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: jbpfdklz, qb4gsw26; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2057.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,057 73.0 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jsfrs9qk",
   "source_url": "https://omlx.ai/benchmarks/performance/jsfrs9qk",
   "tg_tokens": null,
   "tps": 73.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jsfrs9qk-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.7 GB at this context; duplicate traces of the same session: jbpfdklz, qb4gsw26; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1934.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 1,934 58.8 78.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jsfrs9qk",
   "source_url": "https://omlx.ai/benchmarks/performance/jsfrs9qk",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kbp1lmis-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.5 GB; TTFT 1165 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 879.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 879.0 PP tok/s 55.1 TG tok/s 1165 TTFT (ms) 70.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kbp1lmis",
   "source_url": "https://omlx.ai/benchmarks/performance/kbp1lmis",
   "tg_tokens": null,
   "tps": 55.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-3oscqc6i-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: hd0ivzeh; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 223.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-9B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 223.4 36.4 7.0 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3oscqc6i",
   "source_url": "https://omlx.ai/benchmarks/performance/3oscqc6i",
   "tg_tokens": null,
   "tps": 36.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-3oscqc6i-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.5 GB at this context; duplicate traces of the same session: hd0ivzeh; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 230.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-9B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 230.6 34.7 7.5 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3oscqc6i",
   "source_url": "https://omlx.ai/benchmarks/performance/3oscqc6i",
   "tg_tokens": null,
   "tps": 34.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-cfjkikcr-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.0 GB at this context; duplicate traces of the same session: t83bv0gv; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 844.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 844.9 124.3 99.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cfjkikcr",
   "source_url": "https://omlx.ai/benchmarks/performance/cfjkikcr",
   "tg_tokens": null,
   "tps": 124.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-cfjkikcr-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.5 GB at this context; duplicate traces of the same session: t83bv0gv; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1473.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,473 114.3 99.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cfjkikcr",
   "source_url": "https://omlx.ai/benchmarks/performance/cfjkikcr",
   "tg_tokens": null,
   "tps": 114.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-7909oa7v-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: 1ahxs9dm, elvupj2o, ky3231c9, qhuv5ihn; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1436.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 1,436 50.7 14.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7909oa7v",
   "source_url": "https://omlx.ai/benchmarks/performance/7909oa7v",
   "tg_tokens": null,
   "tps": 50.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-7909oa7v-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 14.5 GB at this context; duplicate traces of the same session: 1ahxs9dm, elvupj2o, ky3231c9, qhuv5ihn; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1442.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,442 47.7 14.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7909oa7v",
   "source_url": "https://omlx.ai/benchmarks/performance/7909oa7v",
   "tg_tokens": null,
   "tps": 47.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-7909oa7v-ctx8192",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 14.7 GB at this context; duplicate traces of the same session: 1ahxs9dm, elvupj2o, ky3231c9, qhuv5ihn; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1390.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 8k 1,390 46.1 14.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7909oa7v",
   "source_url": "https://omlx.ai/benchmarks/performance/7909oa7v",
   "tg_tokens": null,
   "tps": 46.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-7909oa7v-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 15.1 GB at this context; duplicate traces of the same session: 1ahxs9dm, elvupj2o, ky3231c9, qhuv5ihn; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1286.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 1,286 40.3 15.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7909oa7v",
   "source_url": "https://omlx.ai/benchmarks/performance/7909oa7v",
   "tg_tokens": null,
   "tps": 40.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-7909oa7v-ctx32768",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: 1ahxs9dm, elvupj2o, ky3231c9, qhuv5ihn; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1092.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 32k 1,092 32.9 15.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7909oa7v",
   "source_url": "https://omlx.ai/benchmarks/performance/7909oa7v",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-e5msym4x-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.8 GB at this context; duplicate traces of the same session: 4w0746se; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1057.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 1,057 39.2 3.8 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e5msym4x",
   "source_url": "https://omlx.ai/benchmarks/performance/e5msym4x",
   "tg_tokens": null,
   "tps": 39.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-e5msym4x-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.3 GB at this context; duplicate traces of the same session: 4w0746se; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1309.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 1,309 41.8 4.3 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e5msym4x",
   "source_url": "https://omlx.ai/benchmarks/performance/e5msym4x",
   "tg_tokens": null,
   "tps": 41.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6f1yqkyk-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: 1ur67nle, 43zpobeb, 6tccdidp, imfm2slr, shr8ktll, w8fh86ma, wcq243rp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3034.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 3,034 105.0 5.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6f1yqkyk",
   "source_url": "https://omlx.ai/benchmarks/performance/6f1yqkyk",
   "tg_tokens": null,
   "tps": 105.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-6f1yqkyk-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: 1ur67nle, 43zpobeb, 6tccdidp, imfm2slr, shr8ktll, w8fh86ma, wcq243rp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3567.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 3,567 103.2 6.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6f1yqkyk",
   "source_url": "https://omlx.ai/benchmarks/performance/6f1yqkyk",
   "tg_tokens": null,
   "tps": 103.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-6f1yqkyk-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: 1ur67nle, 43zpobeb, 6tccdidp, imfm2slr, shr8ktll, w8fh86ma, wcq243rp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3553.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 3,553 100.9 6.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6f1yqkyk",
   "source_url": "https://omlx.ai/benchmarks/performance/6f1yqkyk",
   "tg_tokens": null,
   "tps": 100.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-6f1yqkyk-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: 1ur67nle, 43zpobeb, 6tccdidp, imfm2slr, shr8ktll, w8fh86ma, wcq243rp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3433.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 3,433 95.7 6.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6f1yqkyk",
   "source_url": "https://omlx.ai/benchmarks/performance/6f1yqkyk",
   "tg_tokens": null,
   "tps": 95.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-6f1yqkyk-ctx32768",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.2 GB at this context; duplicate traces of the same session: 1ur67nle, 43zpobeb, 6tccdidp, imfm2slr, shr8ktll, w8fh86ma, wcq243rp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3189.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 3,189 88.0 7.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6f1yqkyk",
   "source_url": "https://omlx.ai/benchmarks/performance/6f1yqkyk",
   "tg_tokens": null,
   "tps": 88.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-6f1yqkyk-ctx65536",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.2 GB at this context; duplicate traces of the same session: 1ur67nle, 43zpobeb, 6tccdidp, imfm2slr, shr8ktll, w8fh86ma, wcq243rp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2778.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 2,778 74.5 8.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6f1yqkyk",
   "source_url": "https://omlx.ai/benchmarks/performance/6f1yqkyk",
   "tg_tokens": null,
   "tps": 74.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-6f1yqkyk-ctx131072",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.2 GB at this context; duplicate traces of the same session: 1ur67nle, 43zpobeb, 6tccdidp, imfm2slr, shr8ktll, w8fh86ma, wcq243rp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2186.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 2,186 57.0 10.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6f1yqkyk",
   "source_url": "https://omlx.ai/benchmarks/performance/6f1yqkyk",
   "tg_tokens": null,
   "tps": 57.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-6f1yqkyk-ctx199680",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 12.3 GB at this context; duplicate traces of the same session: 1ur67nle, 43zpobeb, 6tccdidp, imfm2slr, shr8ktll, w8fh86ma, wcq243rp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1776.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 195k 1,776 45.8 12.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6f1yqkyk",
   "source_url": "https://omlx.ai/benchmarks/performance/6f1yqkyk",
   "tg_tokens": null,
   "tps": 45.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-x0bm4vox-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 10w1ap0s, 59ixd216, 8t96j84t, dpxasall, ey4x0lso; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 922.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 922.7 153.4 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x0bm4vox",
   "source_url": "https://omlx.ai/benchmarks/performance/x0bm4vox",
   "tg_tokens": null,
   "tps": 153.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-x0bm4vox-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 10w1ap0s, 59ixd216, 8t96j84t, dpxasall, ey4x0lso; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3062.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 3,062 149.8 20.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x0bm4vox",
   "source_url": "https://omlx.ai/benchmarks/performance/x0bm4vox",
   "tg_tokens": null,
   "tps": 149.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-x0bm4vox-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 10w1ap0s, 59ixd216, 8t96j84t, dpxasall, ey4x0lso; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4987.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 4,987 123.7 20.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x0bm4vox",
   "source_url": "https://omlx.ai/benchmarks/performance/x0bm4vox",
   "tg_tokens": null,
   "tps": 123.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-x0bm4vox-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 10w1ap0s, 59ixd216, 8t96j84t, dpxasall, ey4x0lso; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3835.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 3,835 102.8 21.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x0bm4vox",
   "source_url": "https://omlx.ai/benchmarks/performance/x0bm4vox",
   "tg_tokens": null,
   "tps": 102.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-x0bm4vox-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: 10w1ap0s, 59ixd216, 8t96j84t, dpxasall, ey4x0lso; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2600.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 2,600 78.8 22.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x0bm4vox",
   "source_url": "https://omlx.ai/benchmarks/performance/x0bm4vox",
   "tg_tokens": null,
   "tps": 78.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-x0bm4vox-ctx199680",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: 10w1ap0s, 59ixd216, 8t96j84t, dpxasall, ey4x0lso; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1914.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 195k 1,914 62.4 23.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x0bm4vox",
   "source_url": "https://omlx.ai/benchmarks/performance/x0bm4vox",
   "tg_tokens": null,
   "tps": 62.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-7es9hfog-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: agtn58y6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 254.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 254.0 28.9 15.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7es9hfog",
   "source_url": "https://omlx.ai/benchmarks/performance/7es9hfog",
   "tg_tokens": null,
   "tps": 28.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-7es9hfog-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: agtn58y6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 272.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 272.6 27.9 19.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7es9hfog",
   "source_url": "https://omlx.ai/benchmarks/performance/7es9hfog",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-0yw9v2gu-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.4 GB at this context; duplicate traces of the same session: 2qnln154; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 248.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 248.9 16.4 28.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0yw9v2gu",
   "source_url": "https://omlx.ai/benchmarks/performance/0yw9v2gu",
   "tg_tokens": null,
   "tps": 16.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-0yw9v2gu-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.7 GB at this context; duplicate traces of the same session: 2qnln154; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 252.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 252.0 16.3 31.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0yw9v2gu",
   "source_url": "https://omlx.ai/benchmarks/performance/0yw9v2gu",
   "tg_tokens": null,
   "tps": 16.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-mikhfjz9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 38.9 GB at this context; duplicate traces of the same session: 7z8iyflz; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 159.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 4k 159.9 20.3 38.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mikhfjz9",
   "source_url": "https://omlx.ai/benchmarks/performance/mikhfjz9",
   "tg_tokens": null,
   "tps": 20.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-mikhfjz9-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 40.7 GB at this context; duplicate traces of the same session: 7z8iyflz; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 149.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 16k 149.2 25.3 40.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mikhfjz9",
   "source_url": "https://omlx.ai/benchmarks/performance/mikhfjz9",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-v5z2baic-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 38.9 GB at this context; duplicate traces of the same session: 59bwewcn; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 159.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 4k 159.9 21.6 38.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v5z2baic",
   "source_url": "https://omlx.ai/benchmarks/performance/v5z2baic",
   "tg_tokens": null,
   "tps": 21.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-v5z2baic-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 40.7 GB at this context; duplicate traces of the same session: 59bwewcn; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 149.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 16k 149.3 25.4 40.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v5z2baic",
   "source_url": "https://omlx.ai/benchmarks/performance/v5z2baic",
   "tg_tokens": null,
   "tps": 25.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-9fe5nkd0-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.0 GB at this context; duplicate traces of the same session: f5v6df6d, h5pqvjak, puqqu2w7, qtmdiupd, w6ngyb5e; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2763.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 2,763 125.1 99.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9fe5nkd0",
   "source_url": "https://omlx.ai/benchmarks/performance/9fe5nkd0",
   "tg_tokens": null,
   "tps": 125.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-9fe5nkd0-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.5 GB at this context; duplicate traces of the same session: f5v6df6d, h5pqvjak, puqqu2w7, qtmdiupd, w6ngyb5e; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4453.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 4,453 114.1 99.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9fe5nkd0",
   "source_url": "https://omlx.ai/benchmarks/performance/9fe5nkd0",
   "tg_tokens": null,
   "tps": 114.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-9fe5nkd0-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.4 GB at this context; duplicate traces of the same session: f5v6df6d, h5pqvjak, puqqu2w7, qtmdiupd, w6ngyb5e; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4891.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 4,891 111.9 100.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9fe5nkd0",
   "source_url": "https://omlx.ai/benchmarks/performance/9fe5nkd0",
   "tg_tokens": null,
   "tps": 111.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-9fe5nkd0-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.3 GB at this context; duplicate traces of the same session: f5v6df6d, h5pqvjak, puqqu2w7, qtmdiupd, w6ngyb5e; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4736.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 4,736 80.3 102.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9fe5nkd0",
   "source_url": "https://omlx.ai/benchmarks/performance/9fe5nkd0",
   "tg_tokens": null,
   "tps": 80.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-9fe5nkd0-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.7 GB at this context; duplicate traces of the same session: f5v6df6d, h5pqvjak, puqqu2w7, qtmdiupd, w6ngyb5e; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4653.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 4,653 95.4 104.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9fe5nkd0",
   "source_url": "https://omlx.ai/benchmarks/performance/9fe5nkd0",
   "tg_tokens": null,
   "tps": 95.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-9fe5nkd0-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 108.2 GB at this context; duplicate traces of the same session: f5v6df6d, h5pqvjak, puqqu2w7, qtmdiupd, w6ngyb5e; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4519.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 4,519 95.5 108.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9fe5nkd0",
   "source_url": "https://omlx.ai/benchmarks/performance/9fe5nkd0",
   "tg_tokens": null,
   "tps": 95.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-8e1hje7t-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.2 GB at this context; duplicate traces of the same session: lwry1wmq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 730.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 730.2 60.7 180.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8e1hje7t",
   "source_url": "https://omlx.ai/benchmarks/performance/8e1hje7t",
   "tg_tokens": null,
   "tps": 60.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-8e1hje7t-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.8 GB at this context; duplicate traces of the same session: lwry1wmq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1099.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 1,099 51.8 180.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8e1hje7t",
   "source_url": "https://omlx.ai/benchmarks/performance/8e1hje7t",
   "tg_tokens": null,
   "tps": 51.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-1wao6x99-ctx1024",
   "model": "DeepSeek-R1-0528-Qwen3-8B",
   "notes": "oMLX community benchmark trace; peak memory 9.2 GB at this context; duplicate traces of the same session: 7vpv6dos, c93n65ox, cx2tnvze, nb3jvroq, q4rth7jn, x2jvbefg, xe4re8vy; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3021.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "DeepSeek-R1-0528-Qwen3-8B-MLX-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 1k 3,021 103.5 9.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1wao6x99",
   "source_url": "https://omlx.ai/benchmarks/performance/1wao6x99",
   "tg_tokens": null,
   "tps": 103.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-1wao6x99-ctx4096",
   "model": "DeepSeek-R1-0528-Qwen3-8B",
   "notes": "oMLX community benchmark trace; peak memory 9.3 GB at this context; duplicate traces of the same session: 7vpv6dos, c93n65ox, cx2tnvze, nb3jvroq, q4rth7jn, x2jvbefg, xe4re8vy; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4876.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "DeepSeek-R1-0528-Qwen3-8B-MLX-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 4k 4,876 100.3 9.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1wao6x99",
   "source_url": "https://omlx.ai/benchmarks/performance/1wao6x99",
   "tg_tokens": null,
   "tps": 100.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-1wao6x99-ctx8192",
   "model": "DeepSeek-R1-0528-Qwen3-8B",
   "notes": "oMLX community benchmark trace; peak memory 9.8 GB at this context; duplicate traces of the same session: 7vpv6dos, c93n65ox, cx2tnvze, nb3jvroq, q4rth7jn, x2jvbefg, xe4re8vy; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4621.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "DeepSeek-R1-0528-Qwen3-8B-MLX-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 8k 4,621 94.2 9.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1wao6x99",
   "source_url": "https://omlx.ai/benchmarks/performance/1wao6x99",
   "tg_tokens": null,
   "tps": 94.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-1wao6x99-ctx16384",
   "model": "DeepSeek-R1-0528-Qwen3-8B",
   "notes": "oMLX community benchmark trace; peak memory 10.9 GB at this context; duplicate traces of the same session: 7vpv6dos, c93n65ox, cx2tnvze, nb3jvroq, q4rth7jn, x2jvbefg, xe4re8vy; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4139.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "DeepSeek-R1-0528-Qwen3-8B-MLX-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 16k 4,139 84.3 10.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1wao6x99",
   "source_url": "https://omlx.ai/benchmarks/performance/1wao6x99",
   "tg_tokens": null,
   "tps": 84.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-1wao6x99-ctx32768",
   "model": "DeepSeek-R1-0528-Qwen3-8B",
   "notes": "oMLX community benchmark trace; peak memory 13.2 GB at this context; duplicate traces of the same session: 7vpv6dos, c93n65ox, cx2tnvze, nb3jvroq, q4rth7jn, x2jvbefg, xe4re8vy; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3294.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "DeepSeek-R1-0528-Qwen3-8B-MLX-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 32k 3,294 70.2 13.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1wao6x99",
   "source_url": "https://omlx.ai/benchmarks/performance/1wao6x99",
   "tg_tokens": null,
   "tps": 70.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-1wao6x99-ctx65536",
   "model": "DeepSeek-R1-0528-Qwen3-8B",
   "notes": "oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 7vpv6dos, c93n65ox, cx2tnvze, nb3jvroq, q4rth7jn, x2jvbefg, xe4re8vy; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2363.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "DeepSeek-R1-0528-Qwen3-8B-MLX-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 64k 2,363 53.1 17.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1wao6x99",
   "source_url": "https://omlx.ai/benchmarks/performance/1wao6x99",
   "tg_tokens": null,
   "tps": 53.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-1wao6x99-ctx131072",
   "model": "DeepSeek-R1-0528-Qwen3-8B",
   "notes": "oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: 7vpv6dos, c93n65ox, cx2tnvze, nb3jvroq, q4rth7jn, x2jvbefg, xe4re8vy; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1476.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "DeepSeek-R1-0528-Qwen3-8B-MLX-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 128k 1,476 35.8 26.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1wao6x99",
   "source_url": "https://omlx.ai/benchmarks/performance/1wao6x99",
   "tg_tokens": null,
   "tps": 35.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-1wao6x99-ctx199680",
   "model": "DeepSeek-R1-0528-Qwen3-8B",
   "notes": "oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: 7vpv6dos, c93n65ox, cx2tnvze, nb3jvroq, q4rth7jn, x2jvbefg, xe4re8vy; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1056.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "DeepSeek-R1-0528-Qwen3-8B-MLX-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 195k 1,056 26.9 36.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1wao6x99",
   "source_url": "https://omlx.ai/benchmarks/performance/1wao6x99",
   "tg_tokens": null,
   "tps": 26.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-9bafp646-ctx1024",
   "model": "Qwen3 14B",
   "notes": "oMLX community benchmark trace; peak memory 8.5 GB at this context; duplicate traces of the same session: ilaa8lww, q4d2mnhx; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 248.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-14B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 248.0 28.5 8.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9bafp646",
   "source_url": "https://omlx.ai/benchmarks/performance/9bafp646",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-9bafp646-ctx4096",
   "model": "Qwen3 14B",
   "notes": "oMLX community benchmark trace; peak memory 9.0 GB at this context; duplicate traces of the same session: ilaa8lww, q4d2mnhx; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 240.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-14B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 240.6 26.6 9.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9bafp646",
   "source_url": "https://omlx.ai/benchmarks/performance/9bafp646",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-9bafp646-ctx8192",
   "model": "Qwen3 14B",
   "notes": "oMLX community benchmark trace; peak memory 9.6 GB at this context; duplicate traces of the same session: ilaa8lww, q4d2mnhx; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 228.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-14B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 228.0 24.3 9.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9bafp646",
   "source_url": "https://omlx.ai/benchmarks/performance/9bafp646",
   "tg_tokens": null,
   "tps": 24.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-m88v3uot-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: hd0ixhad; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 223.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-9B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 223.6 33.1 7.0 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m88v3uot",
   "source_url": "https://omlx.ai/benchmarks/performance/m88v3uot",
   "tg_tokens": null,
   "tps": 33.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-m88v3uot-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.5 GB at this context; duplicate traces of the same session: hd0ixhad; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 230.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-9B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 230.8 33.4 7.5 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m88v3uot",
   "source_url": "https://omlx.ai/benchmarks/performance/m88v3uot",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-i5vuqudd-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: jo41w0lr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3405.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 3,405 93.1 24.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i5vuqudd",
   "source_url": "https://omlx.ai/benchmarks/performance/i5vuqudd",
   "tg_tokens": null,
   "tps": 93.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-i5vuqudd-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: jo41w0lr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5147.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 5,147 91.2 24.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i5vuqudd",
   "source_url": "https://omlx.ai/benchmarks/performance/i5vuqudd",
   "tg_tokens": null,
   "tps": 91.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-afpb810i-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: 1w4hxf8e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5848.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 5,848 95.8 24.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/afpb810i",
   "source_url": "https://omlx.ai/benchmarks/performance/afpb810i",
   "tg_tokens": null,
   "tps": 95.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-afpb810i-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: 1w4hxf8e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5014.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 5,014 94.1 24.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/afpb810i",
   "source_url": "https://omlx.ai/benchmarks/performance/afpb810i",
   "tg_tokens": null,
   "tps": 94.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8xo3z19c-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: my1wj1z7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6010.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 6,010 180.2 22.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8xo3z19c",
   "source_url": "https://omlx.ai/benchmarks/performance/8xo3z19c",
   "tg_tokens": null,
   "tps": 180.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8xo3z19c-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: my1wj1z7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5235.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 5,235 157.2 22.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8xo3z19c",
   "source_url": "https://omlx.ai/benchmarks/performance/8xo3z19c",
   "tg_tokens": null,
   "tps": 157.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-db8rrchu-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 22.1 GB; TTFT 2337 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3505.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k tokens 3,505 PP tok/s 62.0 TG tok/s 2337 TTFT (ms) 22.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/db8rrchu",
   "source_url": "https://omlx.ai/benchmarks/performance/db8rrchu",
   "tg_tokens": null,
   "tps": 62.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-sfbs3ee0-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: h84pphw9; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 113.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 1k 113.5 19.8 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sfbs3ee0",
   "source_url": "https://omlx.ai/benchmarks/performance/sfbs3ee0",
   "tg_tokens": null,
   "tps": 19.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-sfbs3ee0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: h84pphw9; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 103.9 19.0 19.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sfbs3ee0",
   "source_url": "https://omlx.ai/benchmarks/performance/sfbs3ee0",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-69fh4ott-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 20.4 GB; TTFT 272523 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 60.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 16k tokens 60.1 PP tok/s 5.0 TG tok/s 272523 TTFT (ms) 20.4 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/69fh4ott",
   "source_url": "https://omlx.ai/benchmarks/performance/69fh4ott",
   "tg_tokens": null,
   "tps": 5.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-b3ey630k-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.0 GB at this context; duplicate traces of the same session: f3jcr53i, ktcozotq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 256.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 64k 256.6 47.2 43.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b3ey630k",
   "source_url": "https://omlx.ai/benchmarks/performance/b3ey630k",
   "tg_tokens": null,
   "tps": 47.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-b3ey630k-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 47.5 GB at this context; duplicate traces of the same session: f3jcr53i, ktcozotq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 206.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 128k 206.9 46.7 47.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b3ey630k",
   "source_url": "https://omlx.ai/benchmarks/performance/b3ey630k",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-b3ey630k-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 52.5 GB at this context; duplicate traces of the same session: f3jcr53i, ktcozotq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 171.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 195k 171.3 30.5 52.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b3ey630k",
   "source_url": "https://omlx.ai/benchmarks/performance/b3ey630k",
   "tg_tokens": null,
   "tps": 30.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-0uxspf88-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 105.1 GB at this context; duplicate traces of the same session: wvs0xdmt; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2491.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 2,491 48.3 105.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0uxspf88",
   "source_url": "https://omlx.ai/benchmarks/performance/0uxspf88",
   "tg_tokens": null,
   "tps": 48.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-0uxspf88-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 105.5 GB at this context; duplicate traces of the same session: wvs0xdmt; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2111.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 2,111 49.2 105.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0uxspf88",
   "source_url": "https://omlx.ai/benchmarks/performance/0uxspf88",
   "tg_tokens": null,
   "tps": 49.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-el0cp2v4-ctx65536",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: li9kg58t; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 841.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 64k 841.7 23.9 17.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/el0cp2v4",
   "source_url": "https://omlx.ai/benchmarks/performance/el0cp2v4",
   "tg_tokens": null,
   "tps": 23.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-el0cp2v4-ctx131072",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: li9kg58t; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 568.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 128k 568.7 14.0 20.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/el0cp2v4",
   "source_url": "https://omlx.ai/benchmarks/performance/el0cp2v4",
   "tg_tokens": null,
   "tps": 14.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-kpaz1reu-ctx65536",
   "model": "GLM-4.7-Flash",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 51.7 GB; TTFT 226900 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 288.8,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-4.7-Flash-oQ4 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k tokens 288.8 PP tok/s 21.3 TG tok/s 226900 TTFT (ms) 51.7 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kpaz1reu",
   "source_url": "https://omlx.ai/benchmarks/performance/kpaz1reu",
   "tg_tokens": null,
   "tps": 21.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-bm5laore-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.7 GB at this context; duplicate traces of the same session: 68iwmprd, pz5boniw; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 569.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 569.0 27.1 174.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bm5laore",
   "source_url": "https://omlx.ai/benchmarks/performance/bm5laore",
   "tg_tokens": null,
   "tps": 27.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-bm5laore-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.5 GB at this context; duplicate traces of the same session: 68iwmprd, pz5boniw; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 536.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 536.7 26.9 175.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bm5laore",
   "source_url": "https://omlx.ai/benchmarks/performance/bm5laore",
   "tg_tokens": null,
   "tps": 26.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-bm5laore-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 191.6 GB at this context; duplicate traces of the same session: 68iwmprd, pz5boniw; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 448.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 448.0 19.8 191.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bm5laore",
   "source_url": "https://omlx.ai/benchmarks/performance/bm5laore",
   "tg_tokens": null,
   "tps": 19.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ua57gr3n-ctx65536",
   "model": "Ling-3.0-tiny",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 36.2 GB; TTFT 21916 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2990.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ling-3.0-tiny on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 unknown; 64k tokens 2,990 PP tok/s 89.9 TG tok/s 21916 TTFT (ms) 36.2 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ua57gr3n",
   "source_url": "https://omlx.ai/benchmarks/performance/ua57gr3n",
   "tg_tokens": null,
   "tps": 89.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-izj740ri-ctx65536",
   "model": "Ling-3.0-tiny",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 36.2 GB; TTFT 21805 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3006.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ling-3.0-tiny on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 unknown; 64k tokens 3,006 PP tok/s 89.7 TG tok/s 21805 TTFT (ms) 36.2 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/izj740ri",
   "source_url": "https://omlx.ai/benchmarks/performance/izj740ri",
   "tg_tokens": null,
   "tps": 89.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-knlzkk1s-ctx65536",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 37.8 GB; TTFT 70591 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 928.4,
   "provenance": "community",
   "quant": "unknown",
   "quote": "MiMo-V2.6-Distill-Qwen-9B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 unknown; 64k tokens 928.4 PP tok/s 35.0 TG tok/s 70591 TTFT (ms) 37.8 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/knlzkk1s",
   "source_url": "https://omlx.ai/benchmarks/performance/knlzkk1s",
   "tg_tokens": null,
   "tps": 35.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-9pg9ebpw-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 21.1 GB; TTFT 37143 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 441.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 16k tokens 441.1 PP tok/s 28.4 TG tok/s 37143 TTFT (ms) 21.1 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9pg9ebpw",
   "source_url": "https://omlx.ai/benchmarks/performance/9pg9ebpw",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-a9kcs4f2-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 22.5 GB; TTFT 38728 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 423.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 16k tokens 423.1 PP tok/s 38.3 TG tok/s 38728 TTFT (ms) 22.5 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a9kcs4f2",
   "source_url": "https://omlx.ai/benchmarks/performance/a9kcs4f2",
   "tg_tokens": null,
   "tps": 38.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-424fd7fi-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 22.5 GB; TTFT 38780 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 422.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 16k tokens 422.5 PP tok/s 39.7 TG tok/s 38780 TTFT (ms) 22.5 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/424fd7fi",
   "source_url": "https://omlx.ai/benchmarks/performance/424fd7fi",
   "tg_tokens": null,
   "tps": 39.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-dyyw0yln-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 22.5 GB; TTFT 39165 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 418.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 16k tokens 418.3 PP tok/s 35.6 TG tok/s 39165 TTFT (ms) 22.5 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dyyw0yln",
   "source_url": "https://omlx.ai/benchmarks/performance/dyyw0yln",
   "tg_tokens": null,
   "tps": 35.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-yw5wmay0-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 0b9o1758, 2p2ho7j4, 784asgvg, 7khnh2hi, abcjr0eg, wwqrcng8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 911.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 911.4 106.9 21.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yw5wmay0",
   "source_url": "https://omlx.ai/benchmarks/performance/yw5wmay0",
   "tg_tokens": null,
   "tps": 106.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-yw5wmay0-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 0b9o1758, 2p2ho7j4, 784asgvg, 7khnh2hi, abcjr0eg, wwqrcng8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 945.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 945.3 70.8 22.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yw5wmay0",
   "source_url": "https://omlx.ai/benchmarks/performance/yw5wmay0",
   "tg_tokens": null,
   "tps": 70.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-yw5wmay0-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: 0b9o1758, 2p2ho7j4, 784asgvg, 7khnh2hi, abcjr0eg, wwqrcng8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 843.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 843.6 92.3 22.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yw5wmay0",
   "source_url": "https://omlx.ai/benchmarks/performance/yw5wmay0",
   "tg_tokens": null,
   "tps": 92.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-yw5wmay0-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: 0b9o1758, 2p2ho7j4, 784asgvg, 7khnh2hi, abcjr0eg, wwqrcng8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 689.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 689.1 81.9 23.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yw5wmay0",
   "source_url": "https://omlx.ai/benchmarks/performance/yw5wmay0",
   "tg_tokens": null,
   "tps": 81.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-yw5wmay0-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: 0b9o1758, 2p2ho7j4, 784asgvg, 7khnh2hi, abcjr0eg, wwqrcng8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 501.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 501.0 66.7 24.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yw5wmay0",
   "source_url": "https://omlx.ai/benchmarks/performance/yw5wmay0",
   "tg_tokens": null,
   "tps": 66.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-yw5wmay0-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.2 GB at this context; duplicate traces of the same session: 0b9o1758, 2p2ho7j4, 784asgvg, 7khnh2hi, abcjr0eg, wwqrcng8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 328.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 128k 328.2 47.3 26.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yw5wmay0",
   "source_url": "https://omlx.ai/benchmarks/performance/yw5wmay0",
   "tg_tokens": null,
   "tps": 47.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-yw5wmay0-ctx199680",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.7 GB at this context; duplicate traces of the same session: 0b9o1758, 2p2ho7j4, 784asgvg, 7khnh2hi, abcjr0eg, wwqrcng8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 240.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 195k 240.3 28.3 28.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yw5wmay0",
   "source_url": "https://omlx.ai/benchmarks/performance/yw5wmay0",
   "tg_tokens": null,
   "tps": 28.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-r6jvxu9m-ctx16384",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace, single run at 16k context; peak memory 21.3 GB; TTFT 268978 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 60.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 16k tokens 60.9 PP tok/s 10.8 TG tok/s 268978 TTFT (ms) 21.3 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r6jvxu9m",
   "source_url": "https://omlx.ai/benchmarks/performance/r6jvxu9m",
   "tg_tokens": null,
   "tps": 10.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-07js11w5-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 17.1 GB; TTFT 46698 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 350.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 16k tokens 350.8 PP tok/s 22.4 TG tok/s 46698 TTFT (ms) 17.1 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/07js11w5",
   "source_url": "https://omlx.ai/benchmarks/performance/07js11w5",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-r27xizj2-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 15.3 GB; TTFT 46817 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 350.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 16k tokens 350.0 PP tok/s 23.9 TG tok/s 46817 TTFT (ms) 15.3 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r27xizj2",
   "source_url": "https://omlx.ai/benchmarks/performance/r27xizj2",
   "tg_tokens": null,
   "tps": 23.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-mjpj0doa-ctx65536",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 27.8 GB; TTFT 70293 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 932.3,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-OptiQ-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k tokens 932.3 PP tok/s 66.7 TG tok/s 70293 TTFT (ms) 27.8 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mjpj0doa",
   "source_url": "https://omlx.ai/benchmarks/performance/mjpj0doa",
   "tg_tokens": null,
   "tps": 66.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-dekl42ie-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 6.9 GB; TTFT 71644 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 228.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-9B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 16k tokens 228.7 PP tok/s 17.0 TG tok/s 71644 TTFT (ms) 6.9 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dekl42ie",
   "source_url": "https://omlx.ai/benchmarks/performance/dekl42ie",
   "tg_tokens": null,
   "tps": 17.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-sgncqvy1-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 8.2 GB; TTFT 76899 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 213.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-9B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 16k tokens 213.1 PP tok/s 27.2 TG tok/s 76899 TTFT (ms) 8.2 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sgncqvy1",
   "source_url": "https://omlx.ai/benchmarks/performance/sgncqvy1",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-2eb602bb-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 08jczncp, iurbk3bm; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 402.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 402.4 17.2 17.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2eb602bb",
   "source_url": "https://omlx.ai/benchmarks/performance/2eb602bb",
   "tg_tokens": null,
   "tps": 17.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-2eb602bb-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 08jczncp, iurbk3bm; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 353.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 353.1 16.0 17.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2eb602bb",
   "source_url": "https://omlx.ai/benchmarks/performance/2eb602bb",
   "tg_tokens": null,
   "tps": 16.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-2eb602bb-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 08jczncp, iurbk3bm; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 281.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 64k 281.0 13.0 20.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2eb602bb",
   "source_url": "https://omlx.ai/benchmarks/performance/2eb602bb",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-pcboptuv-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: zod816o6; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 548.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 548.7 26.8 17.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pcboptuv",
   "source_url": "https://omlx.ai/benchmarks/performance/pcboptuv",
   "tg_tokens": null,
   "tps": 26.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-pcboptuv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: zod816o6; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 610.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 610.6 25.8 18.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pcboptuv",
   "source_url": "https://omlx.ai/benchmarks/performance/pcboptuv",
   "tg_tokens": null,
   "tps": 25.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-gm7n3dfb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: e5y2jcyu; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 392.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 392.0 20.9 17.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gm7n3dfb",
   "source_url": "https://omlx.ai/benchmarks/performance/gm7n3dfb",
   "tg_tokens": null,
   "tps": 20.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-gm7n3dfb-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: e5y2jcyu; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 411.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 411.4 30.4 18.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gm7n3dfb",
   "source_url": "https://omlx.ai/benchmarks/performance/gm7n3dfb",
   "tg_tokens": null,
   "tps": 30.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-y2x97plg-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: mrjy88at; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 549.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 549.9 26.8 17.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y2x97plg",
   "source_url": "https://omlx.ai/benchmarks/performance/y2x97plg",
   "tg_tokens": null,
   "tps": 26.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-y2x97plg-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: mrjy88at; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 615.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 615.3 29.1 18.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y2x97plg",
   "source_url": "https://omlx.ai/benchmarks/performance/y2x97plg",
   "tg_tokens": null,
   "tps": 29.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-80309gt0-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: j1385hzw; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 405.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 405.9 27.9 16.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/80309gt0",
   "source_url": "https://omlx.ai/benchmarks/performance/80309gt0",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-80309gt0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: j1385hzw; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 392.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 392.7 26.1 18.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/80309gt0",
   "source_url": "https://omlx.ai/benchmarks/performance/80309gt0",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-n90q71ea-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 39.9 GB; TTFT 160051 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 16k tokens 102.4 PP tok/s 18.5 TG tok/s 160051 TTFT (ms) 39.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n90q71ea",
   "source_url": "https://omlx.ai/benchmarks/performance/n90q71ea",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-6pzpr2ks-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 5234t7sp, 9n3vcv5l; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1618.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 1,618 94.2 19.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6pzpr2ks",
   "source_url": "https://omlx.ai/benchmarks/performance/6pzpr2ks",
   "tg_tokens": null,
   "tps": 94.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-6pzpr2ks-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 5234t7sp, 9n3vcv5l; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1965.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 1,965 95.7 20.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6pzpr2ks",
   "source_url": "https://omlx.ai/benchmarks/performance/6pzpr2ks",
   "tg_tokens": null,
   "tps": 95.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-6pzpr2ks-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 5234t7sp, 9n3vcv5l; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1542.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 1,542 82.8 22.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6pzpr2ks",
   "source_url": "https://omlx.ai/benchmarks/performance/6pzpr2ks",
   "tg_tokens": null,
   "tps": 82.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-7gejdprf-ctx1024",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: fk51mym6, hxm8pu2i; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 397.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Alis-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 397.3 47.6 19.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7gejdprf",
   "source_url": "https://omlx.ai/benchmarks/performance/7gejdprf",
   "tg_tokens": null,
   "tps": 47.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-7gejdprf-ctx4096",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: fk51mym6, hxm8pu2i; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 426.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Alis-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 426.1 49.8 23.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7gejdprf",
   "source_url": "https://omlx.ai/benchmarks/performance/7gejdprf",
   "tg_tokens": null,
   "tps": 49.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-7gejdprf-ctx32768",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: fk51mym6, hxm8pu2i; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 351.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Alis-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 351.6 58.7 26.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7gejdprf",
   "source_url": "https://omlx.ai/benchmarks/performance/7gejdprf",
   "tg_tokens": null,
   "tps": 58.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ip5ipp5x-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 24.6 GB; TTFT 172386 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 95.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 16k tokens 95.0 PP tok/s 20.6 TG tok/s 172386 TTFT (ms) 24.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ip5ipp5x",
   "source_url": "https://omlx.ai/benchmarks/performance/ip5ipp5x",
   "tg_tokens": null,
   "tps": 20.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-08",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-m30e0kiq-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 195k context; peak memory 21.2 GB; TTFT 288037 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 694.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 195k tokens 694.4 PP tok/s 9.2 TG tok/s 288037 TTFT (ms) 21.2 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m30e0kiq",
   "source_url": "https://omlx.ai/benchmarks/performance/m30e0kiq",
   "tg_tokens": null,
   "tps": 9.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-7w808wkd-ctx1024",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 77hlwkmm, mbtk9sjc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 372.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Alis-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 372.8 50.8 19.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7w808wkd",
   "source_url": "https://omlx.ai/benchmarks/performance/7w808wkd",
   "tg_tokens": null,
   "tps": 50.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-7w808wkd-ctx4096",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: 77hlwkmm, mbtk9sjc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 414.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Alis-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 414.8 44.9 23.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7w808wkd",
   "source_url": "https://omlx.ai/benchmarks/performance/7w808wkd",
   "tg_tokens": null,
   "tps": 44.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-7w808wkd-ctx32768",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: 77hlwkmm, mbtk9sjc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 356.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Alis-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 356.5 59.5 26.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7w808wkd",
   "source_url": "https://omlx.ai/benchmarks/performance/7w808wkd",
   "tg_tokens": null,
   "tps": 59.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-q0mpulsd-ctx1024",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: jatszo12, nplerc79; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 391.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Alis-MLX-6bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 391.3 43.8 27.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q0mpulsd",
   "source_url": "https://omlx.ai/benchmarks/performance/q0mpulsd",
   "tg_tokens": null,
   "tps": 43.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-q0mpulsd-ctx4096",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 31.0 GB at this context; duplicate traces of the same session: jatszo12, nplerc79; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 425.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Alis-MLX-6bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 425.6 36.8 31.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q0mpulsd",
   "source_url": "https://omlx.ai/benchmarks/performance/q0mpulsd",
   "tg_tokens": null,
   "tps": 36.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-q0mpulsd-ctx32768",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 34.1 GB at this context; duplicate traces of the same session: jatszo12, nplerc79; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 363.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Alis-MLX-6bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 363.1 47.0 34.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q0mpulsd",
   "source_url": "https://omlx.ai/benchmarks/performance/q0mpulsd",
   "tg_tokens": null,
   "tps": 47.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5eqo1zwr-ctx1024",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: 6oqab9y2, 8sdzi7to; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 377.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Alis-MLX-6bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 377.5 44.3 27.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5eqo1zwr",
   "source_url": "https://omlx.ai/benchmarks/performance/5eqo1zwr",
   "tg_tokens": null,
   "tps": 44.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5eqo1zwr-ctx4096",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 31.0 GB at this context; duplicate traces of the same session: 6oqab9y2, 8sdzi7to; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 408.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Alis-MLX-6bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 408.3 36.7 31.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5eqo1zwr",
   "source_url": "https://omlx.ai/benchmarks/performance/5eqo1zwr",
   "tg_tokens": null,
   "tps": 36.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5eqo1zwr-ctx32768",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 34.1 GB at this context; duplicate traces of the same session: 6oqab9y2, 8sdzi7to; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 353.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Alis-MLX-6bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 353.3 47.0 34.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5eqo1zwr",
   "source_url": "https://omlx.ai/benchmarks/performance/5eqo1zwr",
   "tg_tokens": null,
   "tps": 47.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ghkd0djj-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: a39ipuko; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 865.8,
   "provenance": "community",
   "quant": "Q4",
   "quote": "Qwen3.8-27B-Q4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 865.8 33.9 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ghkd0djj",
   "source_url": "https://omlx.ai/benchmarks/performance/ghkd0djj",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ghkd0djj-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: a39ipuko; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 975.8,
   "provenance": "community",
   "quant": "Q4",
   "quote": "Qwen3.8-27B-Q4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 975.8 33.0 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ghkd0djj",
   "source_url": "https://omlx.ai/benchmarks/performance/ghkd0djj",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0cyleajc-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 5y3g8xad, bf4p3fkf, qiy9f9af, qkopipy1, roqiq8p1, uv2fyttf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 551.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 551.7 43.2 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0cyleajc",
   "source_url": "https://omlx.ai/benchmarks/performance/0cyleajc",
   "tg_tokens": null,
   "tps": 43.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0cyleajc-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 5y3g8xad, bf4p3fkf, qiy9f9af, qkopipy1, roqiq8p1, uv2fyttf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 663.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 663.6 32.8 19.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0cyleajc",
   "source_url": "https://omlx.ai/benchmarks/performance/0cyleajc",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0cyleajc-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 5y3g8xad, bf4p3fkf, qiy9f9af, qkopipy1, roqiq8p1, uv2fyttf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 689.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 689.6 14.7 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0cyleajc",
   "source_url": "https://omlx.ai/benchmarks/performance/0cyleajc",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0cyleajc-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 5y3g8xad, bf4p3fkf, qiy9f9af, qkopipy1, roqiq8p1, uv2fyttf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 642.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 642.0 23.4 19.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0cyleajc",
   "source_url": "https://omlx.ai/benchmarks/performance/0cyleajc",
   "tg_tokens": null,
   "tps": 23.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0cyleajc-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 5y3g8xad, bf4p3fkf, qiy9f9af, qkopipy1, roqiq8p1, uv2fyttf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 598.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 598.6 17.5 21.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0cyleajc",
   "source_url": "https://omlx.ai/benchmarks/performance/0cyleajc",
   "tg_tokens": null,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0cyleajc-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 5y3g8xad, bf4p3fkf, qiy9f9af, qkopipy1, roqiq8p1, uv2fyttf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 496.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 496.0 12.1 23.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0cyleajc",
   "source_url": "https://omlx.ai/benchmarks/performance/0cyleajc",
   "tg_tokens": null,
   "tps": 12.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0cyleajc-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 5y3g8xad, bf4p3fkf, qiy9f9af, qkopipy1, roqiq8p1, uv2fyttf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 390.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 128k 390.9 9.7 28.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0cyleajc",
   "source_url": "https://omlx.ai/benchmarks/performance/0cyleajc",
   "tg_tokens": null,
   "tps": 9.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-c7ijna1a-ctx32768",
   "model": "Qwen3.8-35B-A3B-Distill-Heretic",
   "notes": "MTP speculative decoding enabled; fp16 build; oMLX community benchmark trace; peak memory 26.2 GB at this context; duplicate traces of the same session: yvtrvsmp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 633.5,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-35B-A3B-Distill-Heretic-oQ4-fp16-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 633.5 71.1 26.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c7ijna1a",
   "source_url": "https://omlx.ai/benchmarks/performance/c7ijna1a",
   "tg_tokens": null,
   "tps": 71.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-c7ijna1a-ctx65536",
   "model": "Qwen3.8-35B-A3B-Distill-Heretic",
   "notes": "MTP speculative decoding enabled; fp16 build; oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: yvtrvsmp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 445.7,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-35B-A3B-Distill-Heretic-oQ4-fp16-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 445.7 61.5 27.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c7ijna1a",
   "source_url": "https://omlx.ai/benchmarks/performance/c7ijna1a",
   "tg_tokens": null,
   "tps": 61.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-29ooofbo-ctx16384",
   "model": "Qwen3.8-35B-A3B-Distill-Heretic",
   "notes": "MTP speculative decoding enabled; fp16 build; oMLX community benchmark trace, single run at 16k context; peak memory 25.1 GB; TTFT 19535 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 838.7,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-35B-A3B-Distill-Heretic-oQ4-fp16-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k tokens 838.7 PP tok/s 60.7 TG tok/s 19535 TTFT (ms) 25.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/29ooofbo",
   "source_url": "https://omlx.ai/benchmarks/performance/29ooofbo",
   "tg_tokens": null,
   "tps": 60.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-97otwmgo-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: 6e43hgef, lpg0zg75, uqa02aa5, z150s874; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1175.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,175 80.8 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/97otwmgo",
   "source_url": "https://omlx.ai/benchmarks/performance/97otwmgo",
   "tg_tokens": null,
   "tps": 80.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-97otwmgo-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 6e43hgef, lpg0zg75, uqa02aa5, z150s874; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2477.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,477 86.2 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/97otwmgo",
   "source_url": "https://omlx.ai/benchmarks/performance/97otwmgo",
   "tg_tokens": null,
   "tps": 86.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-97otwmgo-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: 6e43hgef, lpg0zg75, uqa02aa5, z150s874; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2577.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,577 88.6 73.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/97otwmgo",
   "source_url": "https://omlx.ai/benchmarks/performance/97otwmgo",
   "tg_tokens": null,
   "tps": 88.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-97otwmgo-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.6 GB at this context; duplicate traces of the same session: 6e43hgef, lpg0zg75, uqa02aa5, z150s874; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2342.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,342 90.2 74.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/97otwmgo",
   "source_url": "https://omlx.ai/benchmarks/performance/97otwmgo",
   "tg_tokens": null,
   "tps": 90.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-97otwmgo-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.0 GB at this context; duplicate traces of the same session: 6e43hgef, lpg0zg75, uqa02aa5, z150s874; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2210.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,210 65.5 79.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/97otwmgo",
   "source_url": "https://omlx.ai/benchmarks/performance/97otwmgo",
   "tg_tokens": null,
   "tps": 65.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-luxqc53r-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: em5n0e26, kh2z2bil, nq58l37c, uhaidvrw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1188.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,188 88.8 70.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/luxqc53r",
   "source_url": "https://omlx.ai/benchmarks/performance/luxqc53r",
   "tg_tokens": null,
   "tps": 88.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-luxqc53r-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: em5n0e26, kh2z2bil, nq58l37c, uhaidvrw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2116.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,116 83.0 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/luxqc53r",
   "source_url": "https://omlx.ai/benchmarks/performance/luxqc53r",
   "tg_tokens": null,
   "tps": 83.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-luxqc53r-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: em5n0e26, kh2z2bil, nq58l37c, uhaidvrw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2652.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,652 84.5 72.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/luxqc53r",
   "source_url": "https://omlx.ai/benchmarks/performance/luxqc53r",
   "tg_tokens": null,
   "tps": 84.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-luxqc53r-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.6 GB at this context; duplicate traces of the same session: em5n0e26, kh2z2bil, nq58l37c, uhaidvrw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2385.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,385 82.6 74.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/luxqc53r",
   "source_url": "https://omlx.ai/benchmarks/performance/luxqc53r",
   "tg_tokens": null,
   "tps": 82.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-luxqc53r-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.9 GB at this context; duplicate traces of the same session: em5n0e26, kh2z2bil, nq58l37c, uhaidvrw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2229.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,229 71.2 78.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/luxqc53r",
   "source_url": "https://omlx.ai/benchmarks/performance/luxqc53r",
   "tg_tokens": null,
   "tps": 71.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kmxdhls5-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: htw3xhrp, sfx5r2im, wydah23g, yua4gypy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1237.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,237 73.8 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kmxdhls5",
   "source_url": "https://omlx.ai/benchmarks/performance/kmxdhls5",
   "tg_tokens": null,
   "tps": 73.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kmxdhls5-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: htw3xhrp, sfx5r2im, wydah23g, yua4gypy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2118.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,118 72.5 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kmxdhls5",
   "source_url": "https://omlx.ai/benchmarks/performance/kmxdhls5",
   "tg_tokens": null,
   "tps": 72.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kmxdhls5-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: htw3xhrp, sfx5r2im, wydah23g, yua4gypy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2321.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,321 61.7 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kmxdhls5",
   "source_url": "https://omlx.ai/benchmarks/performance/kmxdhls5",
   "tg_tokens": null,
   "tps": 61.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kmxdhls5-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: htw3xhrp, sfx5r2im, wydah23g, yua4gypy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2592.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,592 78.7 72.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kmxdhls5",
   "source_url": "https://omlx.ai/benchmarks/performance/kmxdhls5",
   "tg_tokens": null,
   "tps": 78.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kmxdhls5-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: htw3xhrp, sfx5r2im, wydah23g, yua4gypy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2348.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,348 56.4 73.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kmxdhls5",
   "source_url": "https://omlx.ai/benchmarks/performance/kmxdhls5",
   "tg_tokens": null,
   "tps": 56.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cs05rfop-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: y0p7t0vz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1198.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,198 82.8 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cs05rfop",
   "source_url": "https://omlx.ai/benchmarks/performance/cs05rfop",
   "tg_tokens": null,
   "tps": 82.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cs05rfop-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: y0p7t0vz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2065.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,065 61.5 70.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cs05rfop",
   "source_url": "https://omlx.ai/benchmarks/performance/cs05rfop",
   "tg_tokens": null,
   "tps": 61.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-08",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-qlqav02u-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 195k context; peak memory 77.0 GB; TTFT 167983 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1191.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 195k tokens 1,191 PP tok/s 24.6 TG tok/s 167983 TTFT (ms) 77.0 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qlqav02u",
   "source_url": "https://omlx.ai/benchmarks/performance/qlqav02u",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-08",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-efgbep0h-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 195k context; peak memory 65.3 GB; TTFT 562151 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 355.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 195k tokens 355.8 PP tok/s 28.9 TG tok/s 562151 TTFT (ms) 65.3 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/efgbep0h",
   "source_url": "https://omlx.ai/benchmarks/performance/efgbep0h",
   "tg_tokens": null,
   "tps": 28.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-25lc7c09-ctx1024",
   "model": "AliceAI-T5-35B-A0.6B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: ta5yeu6z; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3892.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "AliceAI-T5-35B-A0.6B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 3,892 151.2 18.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/25lc7c09",
   "source_url": "https://omlx.ai/benchmarks/performance/25lc7c09",
   "tg_tokens": null,
   "tps": 151.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-25lc7c09-ctx4096",
   "model": "AliceAI-T5-35B-A0.6B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: ta5yeu6z; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3705.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "AliceAI-T5-35B-A0.6B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 3,705 126.0 19.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/25lc7c09",
   "source_url": "https://omlx.ai/benchmarks/performance/25lc7c09",
   "tg_tokens": null,
   "tps": 126.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-l74szbj9-ctx1024",
   "model": "gemma-4-26B-A4B-it-heretic-ara",
   "notes": "oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: xvmon0zs; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 998.2,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Gemma-4-26B-A4B-it-Heretic-ARA-6bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 998.2 40.3 23.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l74szbj9",
   "source_url": "https://omlx.ai/benchmarks/performance/l74szbj9",
   "tg_tokens": null,
   "tps": 40.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-l74szbj9-ctx4096",
   "model": "gemma-4-26B-A4B-it-heretic-ara",
   "notes": "oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: xvmon0zs; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1237.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Gemma-4-26B-A4B-it-Heretic-ARA-6bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 1,237 33.1 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l74szbj9",
   "source_url": "https://omlx.ai/benchmarks/performance/l74szbj9",
   "tg_tokens": null,
   "tps": 33.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-dlwuj31s-ctx1024",
   "model": "Huihui-Qwen3.8-27B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: lad6rk91; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.8-27B-abliterated-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 125.1 13.7 16.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dlwuj31s",
   "source_url": "https://omlx.ai/benchmarks/performance/dlwuj31s",
   "tg_tokens": null,
   "tps": 13.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-dlwuj31s-ctx4096",
   "model": "Huihui-Qwen3.8-27B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: lad6rk91; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 121.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.8-27B-abliterated-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 121.8 13.2 17.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dlwuj31s",
   "source_url": "https://omlx.ai/benchmarks/performance/dlwuj31s",
   "tg_tokens": null,
   "tps": 13.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-qpe6gth3-ctx1024",
   "model": "Ling-3.0-tiny",
   "notes": "fp16; oMLX community benchmark trace; peak memory 5.2 GB at this context; duplicate traces of the same session: chia4req; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1723.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-tiny-oQ4e-fp16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 1,723 108.4 5.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qpe6gth3",
   "source_url": "https://omlx.ai/benchmarks/performance/qpe6gth3",
   "tg_tokens": null,
   "tps": 108.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-qpe6gth3-ctx4096",
   "model": "Ling-3.0-tiny",
   "notes": "fp16; oMLX community benchmark trace; peak memory 5.7 GB at this context; duplicate traces of the same session: chia4req; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2035.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-tiny-oQ4e-fp16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 2,035 106.4 5.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qpe6gth3",
   "source_url": "https://omlx.ai/benchmarks/performance/qpe6gth3",
   "tg_tokens": null,
   "tps": 106.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-pw5j0987-ctx1024",
   "model": "Qwable-v2",
   "notes": "FP16; DWQ; MTP; Vision; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 0x3ueb5p; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 748.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwable-v2-oQ4e-FP16-DWQ-MTP-Vision-MLX on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 748.4 74.2 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pw5j0987",
   "source_url": "https://omlx.ai/benchmarks/performance/pw5j0987",
   "tg_tokens": null,
   "tps": 74.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-pw5j0987-ctx4096",
   "model": "Qwable-v2",
   "notes": "FP16; DWQ; MTP; Vision; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 0x3ueb5p; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 894.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwable-v2-oQ4e-FP16-DWQ-MTP-Vision-MLX on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 894.6 69.8 22.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pw5j0987",
   "source_url": "https://omlx.ai/benchmarks/performance/pw5j0987",
   "tg_tokens": null,
   "tps": 69.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-4aucwclo-ctx1024",
   "model": "Qwen3.8-4B-Distill",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: szjzyedd; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 750.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-4B-Distill-MLX-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 750.4 89.4 4.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4aucwclo",
   "source_url": "https://omlx.ai/benchmarks/performance/4aucwclo",
   "tg_tokens": null,
   "tps": 89.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-4aucwclo-ctx4096",
   "model": "Qwen3.8-4B-Distill",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 6.0 GB at this context; duplicate traces of the same session: szjzyedd; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1013.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-4B-Distill-MLX-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 1,013 83.1 6.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4aucwclo",
   "source_url": "https://omlx.ai/benchmarks/performance/4aucwclo",
   "tg_tokens": null,
   "tps": 83.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-upolt5g2-ctx1024",
   "model": "Qwen3.8-9B-Distill",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 6.1 GB at this context; duplicate traces of the same session: d1vtscdz; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 413.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-9B-Distill-MLX-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 413.0 56.5 6.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/upolt5g2",
   "source_url": "https://omlx.ai/benchmarks/performance/upolt5g2",
   "tg_tokens": null,
   "tps": 56.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-upolt5g2-ctx4096",
   "model": "Qwen3.8-9B-Distill",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.7 GB at this context; duplicate traces of the same session: d1vtscdz; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 441.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-9B-Distill-MLX-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 441.8 46.6 7.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/upolt5g2",
   "source_url": "https://omlx.ai/benchmarks/performance/upolt5g2",
   "tg_tokens": null,
   "tps": 46.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-zrtekk5n-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 195k context; peak memory 75.3 GB; TTFT 149634 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1337.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 195k tokens 1,337 PP tok/s 32.7 TG tok/s 149634 TTFT (ms) 75.3 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zrtekk5n",
   "source_url": "https://omlx.ai/benchmarks/performance/zrtekk5n",
   "tg_tokens": null,
   "tps": 32.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-p5brjw46-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "scottlowry MTP build; oMLX community benchmark trace; peak memory 30.0 GB at this context; duplicate traces of the same session: 0p9279rp, abpz9el9, gnyulisj, k769vcob, snimc55g; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 713.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "models--scottlowry--Ornith-1.5-35B-A3B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 713.4 61.4 30.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p5brjw46",
   "source_url": "https://omlx.ai/benchmarks/performance/p5brjw46",
   "tg_tokens": null,
   "tps": 61.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-p5brjw46-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "scottlowry MTP build; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: 0p9279rp, abpz9el9, gnyulisj, k769vcob, snimc55g; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 812.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "models--scottlowry--Ornith-1.5-35B-A3B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 812.1 63.3 32.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p5brjw46",
   "source_url": "https://omlx.ai/benchmarks/performance/p5brjw46",
   "tg_tokens": null,
   "tps": 63.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-p5brjw46-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "scottlowry MTP build; oMLX community benchmark trace; peak memory 32.3 GB at this context; duplicate traces of the same session: 0p9279rp, abpz9el9, gnyulisj, k769vcob, snimc55g; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 772.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "models--scottlowry--Ornith-1.5-35B-A3B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 6bit; 8k 772.0 57.5 32.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p5brjw46",
   "source_url": "https://omlx.ai/benchmarks/performance/p5brjw46",
   "tg_tokens": null,
   "tps": 57.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-p5brjw46-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "scottlowry MTP build; oMLX community benchmark trace; peak memory 32.9 GB at this context; duplicate traces of the same session: 0p9279rp, abpz9el9, gnyulisj, k769vcob, snimc55g; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 690.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "models--scottlowry--Ornith-1.5-35B-A3B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 6bit; 16k 690.1 53.1 32.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p5brjw46",
   "source_url": "https://omlx.ai/benchmarks/performance/p5brjw46",
   "tg_tokens": null,
   "tps": 53.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-p5brjw46-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "scottlowry MTP build; oMLX community benchmark trace; peak memory 34.1 GB at this context; duplicate traces of the same session: 0p9279rp, abpz9el9, gnyulisj, k769vcob, snimc55g; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 563.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "models--scottlowry--Ornith-1.5-35B-A3B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 6bit; 32k 563.4 47.8 34.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p5brjw46",
   "source_url": "https://omlx.ai/benchmarks/performance/p5brjw46",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-p5brjw46-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "scottlowry MTP build; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: 0p9279rp, abpz9el9, gnyulisj, k769vcob, snimc55g; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 410.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "models--scottlowry--Ornith-1.5-35B-A3B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 6bit; 64k 410.3 40.2 35.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p5brjw46",
   "source_url": "https://omlx.ai/benchmarks/performance/p5brjw46",
   "tg_tokens": null,
   "tps": 40.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-z176di2q-ctx1024",
   "model": "DeepSeek-R1-0528-Qwen3-8B",
   "notes": "oMLX community benchmark trace; peak memory 8.9 GB at this context; duplicate traces of the same session: fguzmnxe; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 907.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "DeepSeek-R1-0528-Qwen3-8B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 907.1 54.8 8.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z176di2q",
   "source_url": "https://omlx.ai/benchmarks/performance/z176di2q",
   "tg_tokens": null,
   "tps": 54.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-z176di2q-ctx4096",
   "model": "DeepSeek-R1-0528-Qwen3-8B",
   "notes": "oMLX community benchmark trace; peak memory 9.3 GB at this context; duplicate traces of the same session: fguzmnxe; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 814.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "DeepSeek-R1-0528-Qwen3-8B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 814.9 52.0 9.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z176di2q",
   "source_url": "https://omlx.ai/benchmarks/performance/z176di2q",
   "tg_tokens": null,
   "tps": 52.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-if6x2wg6-ctx65536",
   "model": "MiniCPM5-1B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 18.9 GB; TTFT 31033 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2112.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM5-1B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 unknown; 64k tokens 2,112 PP tok/s 122.9 TG tok/s 31033 TTFT (ms) 18.9 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/if6x2wg6",
   "source_url": "https://omlx.ai/benchmarks/performance/if6x2wg6",
   "tg_tokens": null,
   "tps": 122.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1px2t0y5-ctx65536",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 22.7 GB; TTFT 60883 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1076.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "MiniCPM5-2B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 unknown; 64k tokens 1,076 PP tok/s 63.0 TG tok/s 60883 TTFT (ms) 22.7 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1px2t0y5",
   "source_url": "https://omlx.ai/benchmarks/performance/1px2t0y5",
   "tg_tokens": null,
   "tps": 63.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-hm79wv01-ctx65536",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 22.7 GB; TTFT 60895 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1076.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "MiniCPM5-2B-MLX-bf16 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 bf16; 64k tokens 1,076 PP tok/s 62.5 TG tok/s 60895 TTFT (ms) 22.7 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hm79wv01",
   "source_url": "https://omlx.ai/benchmarks/performance/hm79wv01",
   "tg_tokens": null,
   "tps": 62.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-uog5w9kb-ctx131072",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: k168yx0t; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1125.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 128k 1,125 52.6 24.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uog5w9kb",
   "source_url": "https://omlx.ai/benchmarks/performance/uog5w9kb",
   "tg_tokens": null,
   "tps": 52.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-uog5w9kb-ctx199680",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: k168yx0t; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 853.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 195k 853.0 43.3 26.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uog5w9kb",
   "source_url": "https://omlx.ai/benchmarks/performance/uog5w9kb",
   "tg_tokens": null,
   "tps": 43.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-wnji3sk4-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 2bdrgx14; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 224.6 22.5 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wnji3sk4",
   "source_url": "https://omlx.ai/benchmarks/performance/wnji3sk4",
   "tg_tokens": null,
   "tps": 22.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-wnji3sk4-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 2bdrgx14; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 237.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 237.9 21.9 19.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wnji3sk4",
   "source_url": "https://omlx.ai/benchmarks/performance/wnji3sk4",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4kyvss7z-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: ikzq3f2u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 811.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 811.0 24.8 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4kyvss7z",
   "source_url": "https://omlx.ai/benchmarks/performance/4kyvss7z",
   "tg_tokens": null,
   "tps": 24.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4kyvss7z-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: ikzq3f2u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 937.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 937.4 24.5 23.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4kyvss7z",
   "source_url": "https://omlx.ai/benchmarks/performance/4kyvss7z",
   "tg_tokens": null,
   "tps": 24.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rn17zn5d-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: 1yv81yh6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 912.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 912.4 72.9 17.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rn17zn5d",
   "source_url": "https://omlx.ai/benchmarks/performance/rn17zn5d",
   "tg_tokens": null,
   "tps": 72.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rn17zn5d-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 1yv81yh6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1004.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,004 59.7 18.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rn17zn5d",
   "source_url": "https://omlx.ai/benchmarks/performance/rn17zn5d",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fz9wcnn4-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: 402flupb, 58tq4chr, c72ek1bj, rarj2a8j, sgxgrelj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 947.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 947.3 58.7 71.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fz9wcnn4",
   "source_url": "https://omlx.ai/benchmarks/performance/fz9wcnn4",
   "tg_tokens": null,
   "tps": 58.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fz9wcnn4-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: 402flupb, 58tq4chr, c72ek1bj, rarj2a8j, sgxgrelj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2242.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,242 63.9 71.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fz9wcnn4",
   "source_url": "https://omlx.ai/benchmarks/performance/fz9wcnn4",
   "tg_tokens": null,
   "tps": 63.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fz9wcnn4-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: 402flupb, 58tq4chr, c72ek1bj, rarj2a8j, sgxgrelj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2604.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,604 77.2 72.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fz9wcnn4",
   "source_url": "https://omlx.ai/benchmarks/performance/fz9wcnn4",
   "tg_tokens": null,
   "tps": 77.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fz9wcnn4-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: 402flupb, 58tq4chr, c72ek1bj, rarj2a8j, sgxgrelj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2515.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,515 88.2 73.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fz9wcnn4",
   "source_url": "https://omlx.ai/benchmarks/performance/fz9wcnn4",
   "tg_tokens": null,
   "tps": 88.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fz9wcnn4-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 75.7 GB at this context; duplicate traces of the same session: 402flupb, 58tq4chr, c72ek1bj, rarj2a8j, sgxgrelj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2369.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,369 49.7 75.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fz9wcnn4",
   "source_url": "https://omlx.ai/benchmarks/performance/fz9wcnn4",
   "tg_tokens": null,
   "tps": 49.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fz9wcnn4-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 80.0 GB at this context; duplicate traces of the same session: 402flupb, 58tq4chr, c72ek1bj, rarj2a8j, sgxgrelj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2313.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,313 46.7 80.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fz9wcnn4",
   "source_url": "https://omlx.ai/benchmarks/performance/fz9wcnn4",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9npndmmb-ctx65536",
   "model": "Muse Glimmer",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 33.5 GB; TTFT 221504 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 295.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-30B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k tokens 295.9 PP tok/s 33.0 TG tok/s 221504 TTFT (ms) 33.5 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9npndmmb",
   "source_url": "https://omlx.ai/benchmarks/performance/9npndmmb",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-k5n96asy-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 1ennv939, g7lcjdyw, jyjabmmq, xzf1ww2j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4713.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,713 147.6 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k5n96asy",
   "source_url": "https://omlx.ai/benchmarks/performance/k5n96asy",
   "tg_tokens": null,
   "tps": 147.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-k5n96asy-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 1ennv939, g7lcjdyw, jyjabmmq, xzf1ww2j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4331.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 4,331 118.2 20.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k5n96asy",
   "source_url": "https://omlx.ai/benchmarks/performance/k5n96asy",
   "tg_tokens": null,
   "tps": 118.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-k5n96asy-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 1ennv939, g7lcjdyw, jyjabmmq, xzf1ww2j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3230.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 3,230 99.2 21.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k5n96asy",
   "source_url": "https://omlx.ai/benchmarks/performance/k5n96asy",
   "tg_tokens": null,
   "tps": 99.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-k5n96asy-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: 1ennv939, g7lcjdyw, jyjabmmq, xzf1ww2j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2231.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,231 74.8 22.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k5n96asy",
   "source_url": "https://omlx.ai/benchmarks/performance/k5n96asy",
   "tg_tokens": null,
   "tps": 74.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-k5n96asy-ctx199680",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: 1ennv939, g7lcjdyw, jyjabmmq, xzf1ww2j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1683.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 1,683 57.7 23.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k5n96asy",
   "source_url": "https://omlx.ai/benchmarks/performance/k5n96asy",
   "tg_tokens": null,
   "tps": 57.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5gvsprok-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 226 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4528.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 4,528 PP tok/s 146.8 TG tok/s 226 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5gvsprok",
   "source_url": "https://omlx.ai/benchmarks/performance/5gvsprok",
   "tg_tokens": null,
   "tps": 146.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-g2g98irq-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 73lb2whd, k7c6uc8a, lqnna415, pcs3tqkm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4564.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,564 146.5 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g2g98irq",
   "source_url": "https://omlx.ai/benchmarks/performance/g2g98irq",
   "tg_tokens": null,
   "tps": 146.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-g2g98irq-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 73lb2whd, k7c6uc8a, lqnna415, pcs3tqkm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4493.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 4,493 118.5 20.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g2g98irq",
   "source_url": "https://omlx.ai/benchmarks/performance/g2g98irq",
   "tg_tokens": null,
   "tps": 118.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-g2g98irq-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 73lb2whd, k7c6uc8a, lqnna415, pcs3tqkm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3311.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 3,311 99.0 21.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g2g98irq",
   "source_url": "https://omlx.ai/benchmarks/performance/g2g98irq",
   "tg_tokens": null,
   "tps": 99.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-g2g98irq-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: 73lb2whd, k7c6uc8a, lqnna415, pcs3tqkm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2268.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,268 75.6 22.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g2g98irq",
   "source_url": "https://omlx.ai/benchmarks/performance/g2g98irq",
   "tg_tokens": null,
   "tps": 75.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-g2g98irq-ctx199680",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: 73lb2whd, k7c6uc8a, lqnna415, pcs3tqkm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1694.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 1,694 58.9 23.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g2g98irq",
   "source_url": "https://omlx.ai/benchmarks/performance/g2g98irq",
   "tg_tokens": null,
   "tps": 58.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-bbz2z9cp-ctx1024",
   "model": "Qwen3.6-35B-A3B-Heretic-Splash",
   "notes": "oMLX community benchmark trace; peak memory not reported at this context; duplicate traces of the same session: p5fcbi5z, wcazyzxe; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1387.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.6-35B-A3B-Heretic-Splash on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 unknown; 1k 1,387 173.2 \u2014 view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bbz2z9cp",
   "source_url": "https://omlx.ai/benchmarks/performance/bbz2z9cp",
   "tg_tokens": null,
   "tps": 173.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-bbz2z9cp-ctx4096",
   "model": "Qwen3.6-35B-A3B-Heretic-Splash",
   "notes": "oMLX community benchmark trace; peak memory not reported at this context; duplicate traces of the same session: p5fcbi5z, wcazyzxe; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1560.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.6-35B-A3B-Heretic-Splash on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 unknown; 4k 1,560 137.3 \u2014 view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bbz2z9cp",
   "source_url": "https://omlx.ai/benchmarks/performance/bbz2z9cp",
   "tg_tokens": null,
   "tps": 137.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-bbz2z9cp-ctx8192",
   "model": "Qwen3.6-35B-A3B-Heretic-Splash",
   "notes": "oMLX community benchmark trace; peak memory not reported at this context; duplicate traces of the same session: p5fcbi5z, wcazyzxe; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 793.8,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.6-35B-A3B-Heretic-Splash on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 unknown; 8k 793.8 27.4 \u2014 current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bbz2z9cp",
   "source_url": "https://omlx.ai/benchmarks/performance/bbz2z9cp",
   "tg_tokens": null,
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-vfofx8if-ctx32768",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: tc12cxuu, yxd0ak1i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1248.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,248 73.0 24.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vfofx8if",
   "source_url": "https://omlx.ai/benchmarks/performance/vfofx8if",
   "tg_tokens": null,
   "tps": 73.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-vfofx8if-ctx65536",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.7 GB at this context; duplicate traces of the same session: tc12cxuu, yxd0ak1i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 842.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 842.4 104.6 25.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vfofx8if",
   "source_url": "https://omlx.ai/benchmarks/performance/vfofx8if",
   "tg_tokens": null,
   "tps": 104.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-vfofx8if-ctx131072",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: tc12cxuu, yxd0ak1i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 511.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 511.0 74.6 27.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vfofx8if",
   "source_url": "https://omlx.ai/benchmarks/performance/vfofx8if",
   "tg_tokens": null,
   "tps": 74.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-wxerkjcn-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 537.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 537.2 59.9 22.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wxerkjcn",
   "source_url": "https://omlx.ai/benchmarks/performance/wxerkjcn",
   "tg_tokens": null,
   "tps": 59.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-wxerkjcn-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 458.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k 458.0 52.2 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wxerkjcn",
   "source_url": "https://omlx.ai/benchmarks/performance/wxerkjcn",
   "tg_tokens": null,
   "tps": 52.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-wxerkjcn-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 243.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 64k 243.2 37.1 24.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wxerkjcn",
   "source_url": "https://omlx.ai/benchmarks/performance/wxerkjcn",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jeibbbrg-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 883.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 883.3 59.9 24.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jeibbbrg",
   "source_url": "https://omlx.ai/benchmarks/performance/jeibbbrg",
   "tg_tokens": null,
   "tps": 59.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jeibbbrg-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 25.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 882.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 882.9 48.9 25.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jeibbbrg",
   "source_url": "https://omlx.ai/benchmarks/performance/jeibbbrg",
   "tg_tokens": null,
   "tps": 48.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pqpif0zi-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 28.4 GB; TTFT 6078 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 673.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 673.9 PP tok/s 25.5 TG tok/s 6078 TTFT (ms) 28.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pqpif0zi",
   "source_url": "https://omlx.ai/benchmarks/performance/pqpif0zi",
   "tg_tokens": null,
   "tps": 25.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4xs5eybl-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: 151ajwna, 6acgtc8c, 89btdj46; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 914.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 914.2 67.4 17.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4xs5eybl",
   "source_url": "https://omlx.ai/benchmarks/performance/4xs5eybl",
   "tg_tokens": null,
   "tps": 67.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4xs5eybl-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 151ajwna, 6acgtc8c, 89btdj46; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1007.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,007 48.5 18.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4xs5eybl",
   "source_url": "https://omlx.ai/benchmarks/performance/4xs5eybl",
   "tg_tokens": null,
   "tps": 48.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4xs5eybl-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 151ajwna, 6acgtc8c, 89btdj46; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1004.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,004 51.6 19.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4xs5eybl",
   "source_url": "https://omlx.ai/benchmarks/performance/4xs5eybl",
   "tg_tokens": null,
   "tps": 51.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4xs5eybl-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 151ajwna, 6acgtc8c, 89btdj46; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 974.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 974.4 51.3 19.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4xs5eybl",
   "source_url": "https://omlx.ai/benchmarks/performance/4xs5eybl",
   "tg_tokens": null,
   "tps": 51.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-n7m61umr-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 1r0xadou, 5ntcqz79, i0w3ks9x; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 401.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 401.6 31.2 18.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n7m61umr",
   "source_url": "https://omlx.ai/benchmarks/performance/n7m61umr",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-n7m61umr-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 1r0xadou, 5ntcqz79, i0w3ks9x; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 399.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 399.0 22.6 19.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n7m61umr",
   "source_url": "https://omlx.ai/benchmarks/performance/n7m61umr",
   "tg_tokens": null,
   "tps": 22.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-n7m61umr-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 1r0xadou, 5ntcqz79, i0w3ks9x; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 382.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 382.3 34.7 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n7m61umr",
   "source_url": "https://omlx.ai/benchmarks/performance/n7m61umr",
   "tg_tokens": null,
   "tps": 34.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-n7m61umr-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 1r0xadou, 5ntcqz79, i0w3ks9x; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 356.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 356.0 30.7 20.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n7m61umr",
   "source_url": "https://omlx.ai/benchmarks/performance/n7m61umr",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kjfbskrv-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: 22s6n1i8, 2936t3e3, 2sbo695d, 6npwewu7, jy696s0b, wxpvn7wx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1003.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,003 78.9 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kjfbskrv",
   "source_url": "https://omlx.ai/benchmarks/performance/kjfbskrv",
   "tg_tokens": null,
   "tps": 78.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-kjfbskrv-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: 22s6n1i8, 2936t3e3, 2sbo695d, 6npwewu7, jy696s0b, wxpvn7wx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1865.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,865 101.7 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kjfbskrv",
   "source_url": "https://omlx.ai/benchmarks/performance/kjfbskrv",
   "tg_tokens": null,
   "tps": 101.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kjfbskrv-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: 22s6n1i8, 2936t3e3, 2sbo695d, 6npwewu7, jy696s0b, wxpvn7wx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2582.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,582 77.2 72.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kjfbskrv",
   "source_url": "https://omlx.ai/benchmarks/performance/kjfbskrv",
   "tg_tokens": null,
   "tps": 77.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kjfbskrv-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: 22s6n1i8, 2936t3e3, 2sbo695d, 6npwewu7, jy696s0b, wxpvn7wx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2525.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,525 78.2 73.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kjfbskrv",
   "source_url": "https://omlx.ai/benchmarks/performance/kjfbskrv",
   "tg_tokens": null,
   "tps": 78.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kjfbskrv-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.6 GB at this context; duplicate traces of the same session: 22s6n1i8, 2936t3e3, 2sbo695d, 6npwewu7, jy696s0b, wxpvn7wx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2360.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,360 64.1 74.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kjfbskrv",
   "source_url": "https://omlx.ai/benchmarks/performance/kjfbskrv",
   "tg_tokens": null,
   "tps": 64.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kjfbskrv-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.7 GB at this context; duplicate traces of the same session: 22s6n1i8, 2936t3e3, 2sbo695d, 6npwewu7, jy696s0b, wxpvn7wx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2277.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,277 53.4 78.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kjfbskrv",
   "source_url": "https://omlx.ai/benchmarks/performance/kjfbskrv",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-kjfbskrv-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.9 GB at this context; duplicate traces of the same session: 22s6n1i8, 2936t3e3, 2sbo695d, 6npwewu7, jy696s0b, wxpvn7wx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2277.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,277 66.7 80.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kjfbskrv",
   "source_url": "https://omlx.ai/benchmarks/performance/kjfbskrv",
   "tg_tokens": null,
   "tps": 66.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ruvumw4b-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.7 GB; TTFT 1049 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 976.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 976.1 PP tok/s 63.2 TG tok/s 1049 TTFT (ms) 70.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ruvumw4b",
   "source_url": "https://omlx.ai/benchmarks/performance/ruvumw4b",
   "tg_tokens": null,
   "tps": 63.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8r5svxlv-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 195k context; peak memory 80.6 GB; TTFT 77182 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2591.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k tokens 2,591 PP tok/s 59.4 TG tok/s 77182 TTFT (ms) 80.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8r5svxlv",
   "source_url": "https://omlx.ai/benchmarks/performance/8r5svxlv",
   "tg_tokens": null,
   "tps": 59.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-e4rky355-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 78.5 GB; TTFT 58972 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2223.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 128k tokens 2,223 PP tok/s 54.0 TG tok/s 58972 TTFT (ms) 78.5 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e4rky355",
   "source_url": "https://omlx.ai/benchmarks/performance/e4rky355",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-4of4l631-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 74.9 GB; TTFT 52771 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2484.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 128k tokens 2,484 PP tok/s 58.8 TG tok/s 52771 TTFT (ms) 74.9 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4of4l631",
   "source_url": "https://omlx.ai/benchmarks/performance/4of4l631",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-bp12smc6-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 78.1 GB; TTFT 62201 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2107.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 128k tokens 2,107 PP tok/s 55.9 TG tok/s 62201 TTFT (ms) 78.1 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bp12smc6",
   "source_url": "https://omlx.ai/benchmarks/performance/bp12smc6",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gi1f3p2l-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 184.8 GB at this context; duplicate traces of the same session: 52lc5u8h; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1226.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 32k 1,226 26.4 184.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gi1f3p2l",
   "source_url": "https://omlx.ai/benchmarks/performance/gi1f3p2l",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gi1f3p2l-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 186.4 GB at this context; duplicate traces of the same session: 52lc5u8h; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1252.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 64k 1,252 52.7 186.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gi1f3p2l",
   "source_url": "https://omlx.ai/benchmarks/performance/gi1f3p2l",
   "tg_tokens": null,
   "tps": 52.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7k7wg8f4-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: h5nnfxi8, wplipa8r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5734.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 5,734 143.7 22.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7k7wg8f4",
   "source_url": "https://omlx.ai/benchmarks/performance/7k7wg8f4",
   "tg_tokens": null,
   "tps": 143.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-7k7wg8f4-ctx16384",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: h5nnfxi8, wplipa8r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5562.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 5,562 198.1 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7k7wg8f4",
   "source_url": "https://omlx.ai/benchmarks/performance/7k7wg8f4",
   "tg_tokens": null,
   "tps": 198.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-7k7wg8f4-ctx65536",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: h5nnfxi8, wplipa8r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3601.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 3,601 124.3 23.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7k7wg8f4",
   "source_url": "https://omlx.ai/benchmarks/performance/7k7wg8f4",
   "tg_tokens": null,
   "tps": 124.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7k7wg8f4-ctx131072",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: h5nnfxi8, wplipa8r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2403.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 2,403 106.0 24.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7k7wg8f4",
   "source_url": "https://omlx.ai/benchmarks/performance/7k7wg8f4",
   "tg_tokens": null,
   "tps": 106.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-93ejxh6c-ctx1024",
   "model": "Ling-3.0-tiny-heretic",
   "notes": "oMLX community benchmark trace; peak memory 5.0 GB at this context; duplicate traces of the same session: 1yvd2398; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1032.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ling-3.0-tiny-heretic-4bit-lmh4 on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 1,032 102.5 5.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/93ejxh6c",
   "source_url": "https://omlx.ai/benchmarks/performance/93ejxh6c",
   "tg_tokens": null,
   "tps": 102.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-93ejxh6c-ctx4096",
   "model": "Ling-3.0-tiny-heretic",
   "notes": "oMLX community benchmark trace; peak memory 5.5 GB at this context; duplicate traces of the same session: 1yvd2398; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1077.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ling-3.0-tiny-heretic-4bit-lmh4 on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 1,077 96.2 5.5 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/93ejxh6c",
   "source_url": "https://omlx.ai/benchmarks/performance/93ejxh6c",
   "tg_tokens": null,
   "tps": 96.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xxebuld7-ctx1024",
   "model": "Nex-N2.5-mini-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: euycnzq2, s2ns3pft; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1963.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Nex-N2.5-mini-Uncensored-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 1,963 108.6 19.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xxebuld7",
   "source_url": "https://omlx.ai/benchmarks/performance/xxebuld7",
   "tg_tokens": null,
   "tps": 108.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xxebuld7-ctx4096",
   "model": "Nex-N2.5-mini-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: euycnzq2, s2ns3pft; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2708.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Nex-N2.5-mini-Uncensored-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 2,708 107.7 21.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xxebuld7",
   "source_url": "https://omlx.ai/benchmarks/performance/xxebuld7",
   "tg_tokens": null,
   "tps": 107.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xxebuld7-ctx8192",
   "model": "Nex-N2.5-mini-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: euycnzq2, s2ns3pft; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2554.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Nex-N2.5-mini-Uncensored-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 2,554 100.3 21.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xxebuld7",
   "source_url": "https://omlx.ai/benchmarks/performance/xxebuld7",
   "tg_tokens": null,
   "tps": 100.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xxebuld7-ctx16384",
   "model": "Nex-N2.5-mini-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: euycnzq2, s2ns3pft; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2300.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Nex-N2.5-mini-Uncensored-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 16k 2,300 94.6 21.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xxebuld7",
   "source_url": "https://omlx.ai/benchmarks/performance/xxebuld7",
   "tg_tokens": null,
   "tps": 94.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xxebuld7-ctx32768",
   "model": "Nex-N2.5-mini-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: euycnzq2, s2ns3pft; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1895.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Nex-N2.5-mini-Uncensored-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 1,895 86.9 22.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xxebuld7",
   "source_url": "https://omlx.ai/benchmarks/performance/xxebuld7",
   "tg_tokens": null,
   "tps": 86.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xxebuld7-ctx65536",
   "model": "Nex-N2.5-mini-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: euycnzq2, s2ns3pft; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1400.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Nex-N2.5-mini-Uncensored-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k 1,400 74.5 23.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xxebuld7",
   "source_url": "https://omlx.ai/benchmarks/performance/xxebuld7",
   "tg_tokens": null,
   "tps": 74.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fv9qrg3g-ctx1024",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 4tlocimk, lmo2peeo, olmuk47g, pharoa0a; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 759.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 759.3 103.2 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fv9qrg3g",
   "source_url": "https://omlx.ai/benchmarks/performance/fv9qrg3g",
   "tg_tokens": null,
   "tps": 103.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fv9qrg3g-ctx4096",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 4tlocimk, lmo2peeo, olmuk47g, pharoa0a; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2003.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,003 100.9 20.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fv9qrg3g",
   "source_url": "https://omlx.ai/benchmarks/performance/fv9qrg3g",
   "tg_tokens": null,
   "tps": 100.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fv9qrg3g-ctx8192",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: 4tlocimk, lmo2peeo, olmuk47g, pharoa0a; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3330.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 3,330 100.8 20.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fv9qrg3g",
   "source_url": "https://omlx.ai/benchmarks/performance/fv9qrg3g",
   "tg_tokens": null,
   "tps": 100.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fv9qrg3g-ctx16384",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: 4tlocimk, lmo2peeo, olmuk47g, pharoa0a; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3022.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 3,022 87.7 20.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fv9qrg3g",
   "source_url": "https://omlx.ai/benchmarks/performance/fv9qrg3g",
   "tg_tokens": null,
   "tps": 87.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fv9qrg3g-ctx32768",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 4tlocimk, lmo2peeo, olmuk47g, pharoa0a; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2584.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 2,584 81.0 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fv9qrg3g",
   "source_url": "https://omlx.ai/benchmarks/performance/fv9qrg3g",
   "tg_tokens": null,
   "tps": 81.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-in01f665-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: myn6dse6; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 537.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 537.2 59.9 22.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/in01f665",
   "source_url": "https://omlx.ai/benchmarks/performance/in01f665",
   "tg_tokens": null,
   "tps": 59.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-in01f665-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: myn6dse6; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 458.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k 458.0 52.2 23.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/in01f665",
   "source_url": "https://omlx.ai/benchmarks/performance/in01f665",
   "tg_tokens": null,
   "tps": 52.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-in01f665-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: myn6dse6; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 243.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 64k 243.2 37.1 24.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/in01f665",
   "source_url": "https://omlx.ai/benchmarks/performance/in01f665",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-rn7xel3b-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 883.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 883.3 59.9 24.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rn7xel3b",
   "source_url": "https://omlx.ai/benchmarks/performance/rn7xel3b",
   "tg_tokens": null,
   "tps": 59.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-rn7xel3b-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 25.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 882.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 882.9 48.9 25.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rn7xel3b",
   "source_url": "https://omlx.ai/benchmarks/performance/rn7xel3b",
   "tg_tokens": null,
   "tps": 48.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bgdek66q-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: hwtkgb9m, yhz5zr2x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 826.2,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Uncensored-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 826.2 23.9 22.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bgdek66q",
   "source_url": "https://omlx.ai/benchmarks/performance/bgdek66q",
   "tg_tokens": null,
   "tps": 23.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bgdek66q-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: hwtkgb9m, yhz5zr2x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 908.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Uncensored-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 908.0 24.4 23.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bgdek66q",
   "source_url": "https://omlx.ai/benchmarks/performance/bgdek66q",
   "tg_tokens": null,
   "tps": 24.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bgdek66q-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: hwtkgb9m, yhz5zr2x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 929.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Uncensored-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k 929.3 22.5 24.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bgdek66q",
   "source_url": "https://omlx.ai/benchmarks/performance/bgdek66q",
   "tg_tokens": null,
   "tps": 22.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-hd9b8787-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 128.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 128.6 15.0 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hd9b8787",
   "source_url": "https://omlx.ai/benchmarks/performance/hd9b8787",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-hd9b8787-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 134.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 134.4 14.2 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hd9b8787",
   "source_url": "https://omlx.ai/benchmarks/performance/hd9b8787",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6mdjrs58-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 16hmirqd, 8ws7ltix, appb5ykm, jt53z8eh, rj9o0rh2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1121.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,121 62.2 17.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6mdjrs58",
   "source_url": "https://omlx.ai/benchmarks/performance/6mdjrs58",
   "tg_tokens": null,
   "tps": 62.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6mdjrs58-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 16hmirqd, 8ws7ltix, appb5ykm, jt53z8eh, rj9o0rh2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1271.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,271 51.7 19.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6mdjrs58",
   "source_url": "https://omlx.ai/benchmarks/performance/6mdjrs58",
   "tg_tokens": null,
   "tps": 51.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6mdjrs58-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 16hmirqd, 8ws7ltix, appb5ykm, jt53z8eh, rj9o0rh2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1194.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,194 56.7 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6mdjrs58",
   "source_url": "https://omlx.ai/benchmarks/performance/6mdjrs58",
   "tg_tokens": null,
   "tps": 56.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6mdjrs58-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 16hmirqd, 8ws7ltix, appb5ykm, jt53z8eh, rj9o0rh2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1074.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,074 77.7 19.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6mdjrs58",
   "source_url": "https://omlx.ai/benchmarks/performance/6mdjrs58",
   "tg_tokens": null,
   "tps": 77.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6mdjrs58-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 16hmirqd, 8ws7ltix, appb5ykm, jt53z8eh, rj9o0rh2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 924.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 924.4 71.5 21.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6mdjrs58",
   "source_url": "https://omlx.ai/benchmarks/performance/6mdjrs58",
   "tg_tokens": null,
   "tps": 71.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6mdjrs58-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 16hmirqd, 8ws7ltix, appb5ykm, jt53z8eh, rj9o0rh2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 726.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 726.5 38.9 23.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6mdjrs58",
   "source_url": "https://omlx.ai/benchmarks/performance/6mdjrs58",
   "tg_tokens": null,
   "tps": 38.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-x59u8o0k-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 0m9lyemt, ps31o7su; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 589.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 589.7 48.1 17.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x59u8o0k",
   "source_url": "https://omlx.ai/benchmarks/performance/x59u8o0k",
   "tg_tokens": null,
   "tps": 48.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-x59u8o0k-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 0m9lyemt, ps31o7su; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 792.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 792.7 41.6 19.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x59u8o0k",
   "source_url": "https://omlx.ai/benchmarks/performance/x59u8o0k",
   "tg_tokens": null,
   "tps": 41.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-x59u8o0k-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 0m9lyemt, ps31o7su; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 835.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 835.1 47.9 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x59u8o0k",
   "source_url": "https://omlx.ai/benchmarks/performance/x59u8o0k",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fihpc42j-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 24.6 GB; TTFT 21439 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 191.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 191.1 PP tok/s 9.0 TG tok/s 21439 TTFT (ms) 24.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fihpc42j",
   "source_url": "https://omlx.ai/benchmarks/performance/fihpc42j",
   "tg_tokens": null,
   "tps": 9.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bke8gjnw-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 25.2 GB; TTFT 17756 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 230.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 230.7 PP tok/s 10.1 TG tok/s 17756 TTFT (ms) 25.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bke8gjnw",
   "source_url": "https://omlx.ai/benchmarks/performance/bke8gjnw",
   "tg_tokens": null,
   "tps": 10.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0fukhzdc-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 0c7e3w5h, ie5pa03l, vej1an8p; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 396.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 396.4 39.6 18.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0fukhzdc",
   "source_url": "https://omlx.ai/benchmarks/performance/0fukhzdc",
   "tg_tokens": null,
   "tps": 39.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0fukhzdc-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 0c7e3w5h, ie5pa03l, vej1an8p; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 396.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 396.9 17.5 19.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0fukhzdc",
   "source_url": "https://omlx.ai/benchmarks/performance/0fukhzdc",
   "tg_tokens": null,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0fukhzdc-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 0c7e3w5h, ie5pa03l, vej1an8p; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 378.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 378.0 24.5 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0fukhzdc",
   "source_url": "https://omlx.ai/benchmarks/performance/0fukhzdc",
   "tg_tokens": null,
   "tps": 24.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0fukhzdc-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 0c7e3w5h, ie5pa03l, vej1an8p; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 354.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 354.3 17.4 20.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0fukhzdc",
   "source_url": "https://omlx.ai/benchmarks/performance/0fukhzdc",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-5i1byeci-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 1vnwcix6, j6jq5m2b, jep3w426; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 402.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 402.2 50.7 18.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5i1byeci",
   "source_url": "https://omlx.ai/benchmarks/performance/5i1byeci",
   "tg_tokens": null,
   "tps": 50.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-5i1byeci-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 1vnwcix6, j6jq5m2b, jep3w426; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 399.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 399.2 25.0 19.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5i1byeci",
   "source_url": "https://omlx.ai/benchmarks/performance/5i1byeci",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-5i1byeci-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 1vnwcix6, j6jq5m2b, jep3w426; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 369.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 369.6 33.3 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5i1byeci",
   "source_url": "https://omlx.ai/benchmarks/performance/5i1byeci",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-5i1byeci-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 1vnwcix6, j6jq5m2b, jep3w426; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 355.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 355.5 39.7 20.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5i1byeci",
   "source_url": "https://omlx.ai/benchmarks/performance/5i1byeci",
   "tg_tokens": null,
   "tps": 39.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-rrehi6g4-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: dwk0c4mk, jljshmw2, nbqbsokk, xa6benn5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 481.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 481.0 35.8 20.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rrehi6g4",
   "source_url": "https://omlx.ai/benchmarks/performance/rrehi6g4",
   "tg_tokens": null,
   "tps": 35.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-rrehi6g4-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: dwk0c4mk, jljshmw2, nbqbsokk, xa6benn5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 499.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 499.1 25.7 21.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rrehi6g4",
   "source_url": "https://omlx.ai/benchmarks/performance/rrehi6g4",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-rrehi6g4-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: dwk0c4mk, jljshmw2, nbqbsokk, xa6benn5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 497.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 497.0 24.3 22.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rrehi6g4",
   "source_url": "https://omlx.ai/benchmarks/performance/rrehi6g4",
   "tg_tokens": null,
   "tps": 24.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-rrehi6g4-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: dwk0c4mk, jljshmw2, nbqbsokk, xa6benn5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 464.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 464.8 14.2 24.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rrehi6g4",
   "source_url": "https://omlx.ai/benchmarks/performance/rrehi6g4",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-rrehi6g4-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: dwk0c4mk, jljshmw2, nbqbsokk, xa6benn5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 434.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 434.3 8.5 26.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rrehi6g4",
   "source_url": "https://omlx.ai/benchmarks/performance/rrehi6g4",
   "tg_tokens": null,
   "tps": 8.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-bssm1a2d-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: uymhwhp6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 317.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 317.1 51.7 33.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bssm1a2d",
   "source_url": "https://omlx.ai/benchmarks/performance/bssm1a2d",
   "tg_tokens": null,
   "tps": 51.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bssm1a2d-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: uymhwhp6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 333.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 333.8 53.3 37.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bssm1a2d",
   "source_url": "https://omlx.ai/benchmarks/performance/bssm1a2d",
   "tg_tokens": null,
   "tps": 53.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-mynkjrwf-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.0 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 318.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 318.1 58.8 39.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mynkjrwf",
   "source_url": "https://omlx.ai/benchmarks/performance/mynkjrwf",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-mynkjrwf-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 42.3 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 406.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 406.2 48.9 42.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mynkjrwf",
   "source_url": "https://omlx.ai/benchmarks/performance/mynkjrwf",
   "tg_tokens": null,
   "tps": 48.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-g9sqlp6r-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 75.0 GB; TTFT 55579 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2358.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 128k tokens 2,358 PP tok/s 63.1 TG tok/s 55579 TTFT (ms) 75.0 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g9sqlp6r",
   "source_url": "https://omlx.ai/benchmarks/performance/g9sqlp6r",
   "tg_tokens": null,
   "tps": 63.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mmkx0udr-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5734.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 5,734 143.7 22.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mmkx0udr",
   "source_url": "https://omlx.ai/benchmarks/performance/mmkx0udr",
   "tg_tokens": null,
   "tps": 143.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mmkx0udr-ctx16384",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5562.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 5,562 198.1 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mmkx0udr",
   "source_url": "https://omlx.ai/benchmarks/performance/mmkx0udr",
   "tg_tokens": null,
   "tps": 198.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mmkx0udr-ctx65536",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3601.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 3,601 124.3 23.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mmkx0udr",
   "source_url": "https://omlx.ai/benchmarks/performance/mmkx0udr",
   "tg_tokens": null,
   "tps": 124.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mmkx0udr-ctx131072",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2403.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 2,403 106.0 24.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mmkx0udr",
   "source_url": "https://omlx.ai/benchmarks/performance/mmkx0udr",
   "tg_tokens": null,
   "tps": 106.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-k2ncez1x-ctx1024",
   "model": "K2-Horizon-MoVA-36B-A4B",
   "notes": "FP16; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 9d1de2sr; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 339.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "K2-Horizon-MoVA-36B-A4B-oQ4e-FP16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 339.1 36.6 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k2ncez1x",
   "source_url": "https://omlx.ai/benchmarks/performance/k2ncez1x",
   "tg_tokens": null,
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-k2ncez1x-ctx4096",
   "model": "K2-Horizon-MoVA-36B-A4B",
   "notes": "FP16; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 9d1de2sr; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 312.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "K2-Horizon-MoVA-36B-A4B-oQ4e-FP16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 312.6 32.9 22.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k2ncez1x",
   "source_url": "https://omlx.ai/benchmarks/performance/k2ncez1x",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-drb4letw-ctx1024",
   "model": "Nex-N2.5-mini-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: v72h5l1w, x06975tn; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1963.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Nex-N2.5-mini-Uncensored-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 1,963 108.6 19.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/drb4letw",
   "source_url": "https://omlx.ai/benchmarks/performance/drb4letw",
   "tg_tokens": null,
   "tps": 108.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-drb4letw-ctx4096",
   "model": "Nex-N2.5-mini-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: v72h5l1w, x06975tn; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2708.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Nex-N2.5-mini-Uncensored-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 2,708 107.7 21.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/drb4letw",
   "source_url": "https://omlx.ai/benchmarks/performance/drb4letw",
   "tg_tokens": null,
   "tps": 107.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-drb4letw-ctx8192",
   "model": "Nex-N2.5-mini-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: v72h5l1w, x06975tn; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2554.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Nex-N2.5-mini-Uncensored-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 2,554 100.3 21.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/drb4letw",
   "source_url": "https://omlx.ai/benchmarks/performance/drb4letw",
   "tg_tokens": null,
   "tps": 100.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-drb4letw-ctx16384",
   "model": "Nex-N2.5-mini-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: v72h5l1w, x06975tn; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2300.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Nex-N2.5-mini-Uncensored-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 16k 2,300 94.6 21.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/drb4letw",
   "source_url": "https://omlx.ai/benchmarks/performance/drb4letw",
   "tg_tokens": null,
   "tps": 94.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-drb4letw-ctx32768",
   "model": "Nex-N2.5-mini-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: v72h5l1w, x06975tn; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1895.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Nex-N2.5-mini-Uncensored-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 1,895 86.9 22.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/drb4letw",
   "source_url": "https://omlx.ai/benchmarks/performance/drb4letw",
   "tg_tokens": null,
   "tps": 86.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-drb4letw-ctx65536",
   "model": "Nex-N2.5-mini-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: v72h5l1w, x06975tn; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1400.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Nex-N2.5-mini-Uncensored-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k 1,400 74.5 23.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/drb4letw",
   "source_url": "https://omlx.ai/benchmarks/performance/drb4letw",
   "tg_tokens": null,
   "tps": 74.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ra0ghqay-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 39.3 GB; TTFT 46663 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1404.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k tokens 1,404 PP tok/s 80.2 TG tok/s 46663 TTFT (ms) 39.3 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ra0ghqay",
   "source_url": "https://omlx.ai/benchmarks/performance/ra0ghqay",
   "tg_tokens": null,
   "tps": 80.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9odbyyrk-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 55.4 GB; TTFT 47145 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1390.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 64k tokens 1,390 PP tok/s 67.9 TG tok/s 47145 TTFT (ms) 55.4 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9odbyyrk",
   "source_url": "https://omlx.ai/benchmarks/performance/9odbyyrk",
   "tg_tokens": null,
   "tps": 67.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-j1inamp5-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 55.4 GB; TTFT 47126 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1391.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 64k tokens 1,391 PP tok/s 67.5 TG tok/s 47126 TTFT (ms) 55.4 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j1inamp5",
   "source_url": "https://omlx.ai/benchmarks/performance/j1inamp5",
   "tg_tokens": null,
   "tps": 67.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-szd8mop7-ctx65536",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 37.8 GB; TTFT 70518 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 929.3,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Ornith-1.5-9B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 unknown; 64k tokens 929.3 PP tok/s 35.0 TG tok/s 70518 TTFT (ms) 37.8 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/szd8mop7",
   "source_url": "https://omlx.ai/benchmarks/performance/szd8mop7",
   "tg_tokens": null,
   "tps": 35.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-k5mpp0ef-ctx65536",
   "model": "Qwen-AgentWorld-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 40.1 GB; TTFT 46792 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1401.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen-AgentWorld-35B-A3B-MLX-oQ4 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k tokens 1,401 PP tok/s 74.7 TG tok/s 46792 TTFT (ms) 40.1 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k5mpp0ef",
   "source_url": "https://omlx.ai/benchmarks/performance/k5mpp0ef",
   "tg_tokens": null,
   "tps": 74.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-cc7ij3k8-ctx1024",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "FP16; oMLX community benchmark trace; peak memory 12.8 GB at this context; duplicate traces of the same session: 5b99cw1d; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 813.8,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B-FP16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 813.8 68.6 12.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cc7ij3k8",
   "source_url": "https://omlx.ai/benchmarks/performance/cc7ij3k8",
   "tg_tokens": null,
   "tps": 68.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-cc7ij3k8-ctx4096",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "FP16; oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: 5b99cw1d; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 907.8,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B-FP16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 907.8 67.6 14.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cc7ij3k8",
   "source_url": "https://omlx.ai/benchmarks/performance/cc7ij3k8",
   "tg_tokens": null,
   "tps": 67.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-r0jo4300-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: vyr5p7og, wgymx2r8; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 588.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 588.7 59.0 22.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r0jo4300",
   "source_url": "https://omlx.ai/benchmarks/performance/r0jo4300",
   "tg_tokens": null,
   "tps": 59.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-r0jo4300-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: vyr5p7og, wgymx2r8; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 152.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 128k 152.0 27.0 26.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r0jo4300",
   "source_url": "https://omlx.ai/benchmarks/performance/r0jo4300",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-r0jo4300-ctx199680",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: vyr5p7og, wgymx2r8; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 109.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 195k 109.4 20.7 27.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r0jo4300",
   "source_url": "https://omlx.ai/benchmarks/performance/r0jo4300",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-n51g8jt9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: 02icvfli; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 804.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 804.9 54.0 33.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n51g8jt9",
   "source_url": "https://omlx.ai/benchmarks/performance/n51g8jt9",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n51g8jt9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: 02icvfli; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 924.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 924.7 41.8 34.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n51g8jt9",
   "source_url": "https://omlx.ai/benchmarks/performance/n51g8jt9",
   "tg_tokens": null,
   "tps": 41.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-gjuapyns-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 128.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 128.6 15.0 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gjuapyns",
   "source_url": "https://omlx.ai/benchmarks/performance/gjuapyns",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-gjuapyns-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 134.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 134.4 14.2 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gjuapyns",
   "source_url": "https://omlx.ai/benchmarks/performance/gjuapyns",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ne2y77ke-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: qxty2nv9; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 122.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 122.9 24.0 20.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ne2y77ke",
   "source_url": "https://omlx.ai/benchmarks/performance/ne2y77ke",
   "tg_tokens": null,
   "tps": 24.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ne2y77ke-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: qxty2nv9; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 185.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 185.0 30.5 24.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ne2y77ke",
   "source_url": "https://omlx.ai/benchmarks/performance/ne2y77ke",
   "tg_tokens": null,
   "tps": 30.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tbcox7hf-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: g5ufsxr3, ggeapf4t, odjbzzpj, yno62si1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1130.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,130 72.9 17.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tbcox7hf",
   "source_url": "https://omlx.ai/benchmarks/performance/tbcox7hf",
   "tg_tokens": null,
   "tps": 72.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tbcox7hf-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: g5ufsxr3, ggeapf4t, odjbzzpj, yno62si1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1234.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,234 54.7 19.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tbcox7hf",
   "source_url": "https://omlx.ai/benchmarks/performance/tbcox7hf",
   "tg_tokens": null,
   "tps": 54.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tbcox7hf-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: g5ufsxr3, ggeapf4t, odjbzzpj, yno62si1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1087.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,087 62.0 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tbcox7hf",
   "source_url": "https://omlx.ai/benchmarks/performance/tbcox7hf",
   "tg_tokens": null,
   "tps": 62.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tbcox7hf-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: g5ufsxr3, ggeapf4t, odjbzzpj, yno62si1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1027.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,027 74.0 19.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tbcox7hf",
   "source_url": "https://omlx.ai/benchmarks/performance/tbcox7hf",
   "tg_tokens": null,
   "tps": 74.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tbcox7hf-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: g5ufsxr3, ggeapf4t, odjbzzpj, yno62si1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 901.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 901.1 67.5 21.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tbcox7hf",
   "source_url": "https://omlx.ai/benchmarks/performance/tbcox7hf",
   "tg_tokens": null,
   "tps": 67.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-6935dwua-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 9q59uh29, sw2xsmra; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 584.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 584.1 40.2 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6935dwua",
   "source_url": "https://omlx.ai/benchmarks/performance/6935dwua",
   "tg_tokens": null,
   "tps": 40.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-6935dwua-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 9q59uh29, sw2xsmra; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 683.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 683.4 34.4 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6935dwua",
   "source_url": "https://omlx.ai/benchmarks/performance/6935dwua",
   "tg_tokens": null,
   "tps": 34.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-6935dwua-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 9q59uh29, sw2xsmra; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 707.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 707.5 32.3 19.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6935dwua",
   "source_url": "https://omlx.ai/benchmarks/performance/6935dwua",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0ui5c52i-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: az9btmiu, pjqqgk52; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 582.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 582.4 37.9 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ui5c52i",
   "source_url": "https://omlx.ai/benchmarks/performance/0ui5c52i",
   "tg_tokens": null,
   "tps": 37.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-0ui5c52i-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: az9btmiu, pjqqgk52; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 683.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 683.4 61.4 19.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ui5c52i",
   "source_url": "https://omlx.ai/benchmarks/performance/0ui5c52i",
   "tg_tokens": null,
   "tps": 61.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0ui5c52i-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: az9btmiu, pjqqgk52; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 701.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 701.6 23.1 19.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ui5c52i",
   "source_url": "https://omlx.ai/benchmarks/performance/0ui5c52i",
   "tg_tokens": null,
   "tps": 23.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-gx8wb4k5-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 0npf5fxe, ig7lfmee; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 604.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 604.1 42.5 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gx8wb4k5",
   "source_url": "https://omlx.ai/benchmarks/performance/gx8wb4k5",
   "tg_tokens": null,
   "tps": 42.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-gx8wb4k5-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 0npf5fxe, ig7lfmee; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 701.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 701.6 41.9 19.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gx8wb4k5",
   "source_url": "https://omlx.ai/benchmarks/performance/gx8wb4k5",
   "tg_tokens": null,
   "tps": 41.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-gx8wb4k5-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 0npf5fxe, ig7lfmee; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 720.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 720.5 27.2 19.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gx8wb4k5",
   "source_url": "https://omlx.ai/benchmarks/performance/gx8wb4k5",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-lk4yi2c8-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: duj3pkxx, x3qs2ngi; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 601.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 601.1 39.4 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lk4yi2c8",
   "source_url": "https://omlx.ai/benchmarks/performance/lk4yi2c8",
   "tg_tokens": null,
   "tps": 39.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-lk4yi2c8-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: duj3pkxx, x3qs2ngi; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 675.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 675.1 46.0 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lk4yi2c8",
   "source_url": "https://omlx.ai/benchmarks/performance/lk4yi2c8",
   "tg_tokens": null,
   "tps": 46.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-lk4yi2c8-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: duj3pkxx, x3qs2ngi; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 701.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 701.3 25.7 19.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lk4yi2c8",
   "source_url": "https://omlx.ai/benchmarks/performance/lk4yi2c8",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-c4np18mx-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 8tmzw7vf, z19bw2qn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 598.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 598.2 37.1 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c4np18mx",
   "source_url": "https://omlx.ai/benchmarks/performance/c4np18mx",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-c4np18mx-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 8tmzw7vf, z19bw2qn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 669.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 669.3 31.0 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c4np18mx",
   "source_url": "https://omlx.ai/benchmarks/performance/c4np18mx",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-c4np18mx-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 8tmzw7vf, z19bw2qn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 693.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 693.4 35.6 19.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c4np18mx",
   "source_url": "https://omlx.ai/benchmarks/performance/c4np18mx",
   "tg_tokens": null,
   "tps": 35.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3m5m9488-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 597.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 597.8 40.3 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3m5m9488",
   "source_url": "https://omlx.ai/benchmarks/performance/3m5m9488",
   "tg_tokens": null,
   "tps": 40.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3m5m9488-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 692.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 692.6 34.0 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3m5m9488",
   "source_url": "https://omlx.ai/benchmarks/performance/3m5m9488",
   "tg_tokens": null,
   "tps": 34.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3m5m9488-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 710.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 710.4 37.7 19.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3m5m9488",
   "source_url": "https://omlx.ai/benchmarks/performance/3m5m9488",
   "tg_tokens": null,
   "tps": 37.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3m5m9488-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 696.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 696.5 41.9 20.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3m5m9488",
   "source_url": "https://omlx.ai/benchmarks/performance/3m5m9488",
   "tg_tokens": null,
   "tps": 41.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3m5m9488-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 630.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 630.2 41.5 23.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3m5m9488",
   "source_url": "https://omlx.ai/benchmarks/performance/3m5m9488",
   "tg_tokens": null,
   "tps": 41.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-pixaw279-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.0 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 318.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 318.1 58.8 39.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pixaw279",
   "source_url": "https://omlx.ai/benchmarks/performance/pixaw279",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-pixaw279-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 42.3 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 406.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 406.2 48.9 42.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pixaw279",
   "source_url": "https://omlx.ai/benchmarks/performance/pixaw279",
   "tg_tokens": null,
   "tps": 48.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-swr6xaxe-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.8 GB at this context; duplicate traces of the same session: wlj81so6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 89.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 89.2 10.7 16.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/swr6xaxe",
   "source_url": "https://omlx.ai/benchmarks/performance/swr6xaxe",
   "tg_tokens": null,
   "tps": 10.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-swr6xaxe-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: wlj81so6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 267.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 267.3 9.5 17.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/swr6xaxe",
   "source_url": "https://omlx.ai/benchmarks/performance/swr6xaxe",
   "tg_tokens": null,
   "tps": 9.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-cizgw27f-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.4 GB at this context; duplicate traces of the same session: zht92sw7; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1616.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 4k 1,616 79.9 69.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cizgw27f",
   "source_url": "https://omlx.ai/benchmarks/performance/cizgw27f",
   "tg_tokens": null,
   "tps": 79.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-cizgw27f-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.4 GB at this context; duplicate traces of the same session: zht92sw7; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2917.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 32k 2,917 77.7 71.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cizgw27f",
   "source_url": "https://omlx.ai/benchmarks/performance/cizgw27f",
   "tg_tokens": null,
   "tps": 77.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-cizgw27f-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.9 GB at this context; duplicate traces of the same session: zht92sw7; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2760.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 128k 2,760 67.9 74.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cizgw27f",
   "source_url": "https://omlx.ai/benchmarks/performance/cizgw27f",
   "tg_tokens": null,
   "tps": 67.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2okwnc5a-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 252.4 GB at this context; duplicate traces of the same session: 70qb8im3, bx3n4fc0, nef98fmr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 492.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-Q8mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 16k 492.7 61.4 252.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2okwnc5a",
   "source_url": "https://omlx.ai/benchmarks/performance/2okwnc5a",
   "tg_tokens": null,
   "tps": 61.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2okwnc5a-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 252.6 GB at this context; duplicate traces of the same session: 70qb8im3, bx3n4fc0, nef98fmr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 513.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-Q8mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 32k 513.4 46.1 252.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2okwnc5a",
   "source_url": "https://omlx.ai/benchmarks/performance/2okwnc5a",
   "tg_tokens": null,
   "tps": 46.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2okwnc5a-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 253.3 GB at this context; duplicate traces of the same session: 70qb8im3, bx3n4fc0, nef98fmr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 506.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-Q8mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 64k 506.2 37.7 253.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2okwnc5a",
   "source_url": "https://omlx.ai/benchmarks/performance/2okwnc5a",
   "tg_tokens": null,
   "tps": 37.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2okwnc5a-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 254.6 GB at this context; duplicate traces of the same session: 70qb8im3, bx3n4fc0, nef98fmr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 496.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-Q8mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 128k 496.1 63.2 254.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2okwnc5a",
   "source_url": "https://omlx.ai/benchmarks/performance/2okwnc5a",
   "tg_tokens": null,
   "tps": 63.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-j1n80unc-ctx65536",
   "model": "Nex-N2.5-mini",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 40.1 GB; TTFT 46742 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1402.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Nex-N2.5-mini-oQ4 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k tokens 1,402 PP tok/s 76.7 TG tok/s 46742 TTFT (ms) 40.1 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j1n80unc",
   "source_url": "https://omlx.ai/benchmarks/performance/j1n80unc",
   "tg_tokens": null,
   "tps": 76.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-4iuj27su-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: cg2aiims, l30qviev, mcdiqxz3; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1521.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 1,521 102.7 20.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4iuj27su",
   "source_url": "https://omlx.ai/benchmarks/performance/4iuj27su",
   "tg_tokens": null,
   "tps": 102.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-4iuj27su-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: cg2aiims, l30qviev, mcdiqxz3; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2804.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 2,804 97.3 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4iuj27su",
   "source_url": "https://omlx.ai/benchmarks/performance/4iuj27su",
   "tg_tokens": null,
   "tps": 97.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-4iuj27su-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: cg2aiims, l30qviev, mcdiqxz3; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2563.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 2,563 89.8 20.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4iuj27su",
   "source_url": "https://omlx.ai/benchmarks/performance/4iuj27su",
   "tg_tokens": null,
   "tps": 89.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-4iuj27su-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: cg2aiims, l30qviev, mcdiqxz3; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2160.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 2,160 77.2 20.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4iuj27su",
   "source_url": "https://omlx.ai/benchmarks/performance/4iuj27su",
   "tg_tokens": null,
   "tps": 77.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-94o1fko8-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: olpbyx00; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 241.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 241.2 20.9 6.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/94o1fko8",
   "source_url": "https://omlx.ai/benchmarks/performance/94o1fko8",
   "tg_tokens": null,
   "tps": 20.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-94o1fko8-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: olpbyx00; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 231.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 16k 231.2 18.5 6.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/94o1fko8",
   "source_url": "https://omlx.ai/benchmarks/performance/94o1fko8",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-oolra7s5-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 6z2tv0mx, y1ckqmre; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1667.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 1,667 100.8 19.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oolra7s5",
   "source_url": "https://omlx.ai/benchmarks/performance/oolra7s5",
   "tg_tokens": null,
   "tps": 100.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-oolra7s5-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 6z2tv0mx, y1ckqmre; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2026.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 2,026 95.7 20.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oolra7s5",
   "source_url": "https://omlx.ai/benchmarks/performance/oolra7s5",
   "tg_tokens": null,
   "tps": 95.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-oolra7s5-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 6z2tv0mx, y1ckqmre; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1516.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 1,516 80.0 22.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oolra7s5",
   "source_url": "https://omlx.ai/benchmarks/performance/oolra7s5",
   "tg_tokens": null,
   "tps": 80.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-wdaqgfx3-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 39.9 GB; TTFT 46681 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1404.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k tokens 1,404 PP tok/s 76.6 TG tok/s 46681 TTFT (ms) 39.9 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wdaqgfx3",
   "source_url": "https://omlx.ai/benchmarks/performance/wdaqgfx3",
   "tg_tokens": null,
   "tps": 76.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-746pg3cx-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 42.6 GB; TTFT 239365 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 273.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k tokens 273.8 PP tok/s 29.3 TG tok/s 239365 TTFT (ms) 42.6 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/746pg3cx",
   "source_url": "https://omlx.ai/benchmarks/performance/746pg3cx",
   "tg_tokens": null,
   "tps": 29.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jtmypbxq-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.5 GB at this context; duplicate traces of the same session: bv1tbe0i; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 491.5,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 491.5 76.2 34.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jtmypbxq",
   "source_url": "https://omlx.ai/benchmarks/performance/jtmypbxq",
   "tg_tokens": null,
   "tps": 76.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jtmypbxq-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.7 GB at this context; duplicate traces of the same session: bv1tbe0i; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 340.2,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k 340.2 60.1 39.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jtmypbxq",
   "source_url": "https://omlx.ai/benchmarks/performance/jtmypbxq",
   "tg_tokens": null,
   "tps": 60.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-c35z7b8x-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: 46ap7qso; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 346.5,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 346.5 80.4 23.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c35z7b8x",
   "source_url": "https://omlx.ai/benchmarks/performance/c35z7b8x",
   "tg_tokens": null,
   "tps": 80.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-c35z7b8x-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 46ap7qso; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 265.2,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k 265.2 60.4 28.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c35z7b8x",
   "source_url": "https://omlx.ai/benchmarks/performance/c35z7b8x",
   "tg_tokens": null,
   "tps": 60.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5pt0jrul-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: bx7utjyv; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 322.9,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 322.9 74.7 19.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5pt0jrul",
   "source_url": "https://omlx.ai/benchmarks/performance/5pt0jrul",
   "tg_tokens": null,
   "tps": 74.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5pt0jrul-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: bx7utjyv; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 356.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 356.0 68.1 22.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5pt0jrul",
   "source_url": "https://omlx.ai/benchmarks/performance/5pt0jrul",
   "tg_tokens": null,
   "tps": 68.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-wwhern2p-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.6 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 310.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 310.0 41.2 30.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wwhern2p",
   "source_url": "https://omlx.ai/benchmarks/performance/wwhern2p",
   "tg_tokens": null,
   "tps": 41.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-wwhern2p-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.6 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 490.4,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 490.4 65.2 33.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wwhern2p",
   "source_url": "https://omlx.ai/benchmarks/performance/wwhern2p",
   "tg_tokens": null,
   "tps": 65.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7er1o5fn-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: 0otjt7hp, xd3qtus8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1138.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,138 39.5 22.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7er1o5fn",
   "source_url": "https://omlx.ai/benchmarks/performance/7er1o5fn",
   "tg_tokens": null,
   "tps": 39.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7er1o5fn-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: 0otjt7hp, xd3qtus8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1149.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,149 50.9 23.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7er1o5fn",
   "source_url": "https://omlx.ai/benchmarks/performance/7er1o5fn",
   "tg_tokens": null,
   "tps": 50.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7er1o5fn-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: 0otjt7hp, xd3qtus8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1134.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,134 52.0 23.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7er1o5fn",
   "source_url": "https://omlx.ai/benchmarks/performance/7er1o5fn",
   "tg_tokens": null,
   "tps": 52.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-es4ynma9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 18.8 GB; TTFT 18479 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 221.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 221.7 PP tok/s 14.8 TG tok/s 18479 TTFT (ms) 18.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/es4ynma9",
   "source_url": "https://omlx.ai/benchmarks/performance/es4ynma9",
   "tg_tokens": null,
   "tps": 14.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-vx5ec4yy-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 5jo8qhmx, tds4zv4d, z8q4oacs; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 597.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 597.8 40.3 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vx5ec4yy",
   "source_url": "https://omlx.ai/benchmarks/performance/vx5ec4yy",
   "tg_tokens": null,
   "tps": 40.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-vx5ec4yy-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 5jo8qhmx, tds4zv4d, z8q4oacs; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 692.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 692.6 34.0 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vx5ec4yy",
   "source_url": "https://omlx.ai/benchmarks/performance/vx5ec4yy",
   "tg_tokens": null,
   "tps": 34.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-vx5ec4yy-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 5jo8qhmx, tds4zv4d, z8q4oacs; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 710.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 710.4 37.7 19.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vx5ec4yy",
   "source_url": "https://omlx.ai/benchmarks/performance/vx5ec4yy",
   "tg_tokens": null,
   "tps": 37.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-vx5ec4yy-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 5jo8qhmx, tds4zv4d, z8q4oacs; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 696.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 696.5 41.9 20.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vx5ec4yy",
   "source_url": "https://omlx.ai/benchmarks/performance/vx5ec4yy",
   "tg_tokens": null,
   "tps": 41.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-vx5ec4yy-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: 5jo8qhmx, tds4zv4d, z8q4oacs; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 630.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 630.2 41.5 23.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vx5ec4yy",
   "source_url": "https://omlx.ai/benchmarks/performance/vx5ec4yy",
   "tg_tokens": null,
   "tps": 41.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-lfi3gl95-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 2eqrrghn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 581.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 581.8 37.1 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lfi3gl95",
   "source_url": "https://omlx.ai/benchmarks/performance/lfi3gl95",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-lfi3gl95-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 2eqrrghn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 682.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 682.7 55.6 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lfi3gl95",
   "source_url": "https://omlx.ai/benchmarks/performance/lfi3gl95",
   "tg_tokens": null,
   "tps": 55.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-lfi3gl95-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 2eqrrghn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 708.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 708.5 40.9 19.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lfi3gl95",
   "source_url": "https://omlx.ai/benchmarks/performance/lfi3gl95",
   "tg_tokens": null,
   "tps": 40.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-opvj0axn-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 105.2 GB; TTFT 1283 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 798.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k tokens 798.1 PP tok/s 36.4 TG tok/s 1283 TTFT (ms) 105.2 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/opvj0axn",
   "source_url": "https://omlx.ai/benchmarks/performance/opvj0axn",
   "tg_tokens": null,
   "tps": 36.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7vdjh66o-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.0 GB at this context; duplicate traces of the same session: ds5mp3yf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 74.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 74.0 8.7 39.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7vdjh66o",
   "source_url": "https://omlx.ai/benchmarks/performance/7vdjh66o",
   "tg_tokens": null,
   "tps": 8.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7vdjh66o-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 41.7 GB at this context; duplicate traces of the same session: ds5mp3yf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 226.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 226.2 8.7 41.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7vdjh66o",
   "source_url": "https://omlx.ai/benchmarks/performance/7vdjh66o",
   "tg_tokens": null,
   "tps": 8.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-c0uxi6bn-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.4 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1616.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 4k 1,616 79.9 69.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c0uxi6bn",
   "source_url": "https://omlx.ai/benchmarks/performance/c0uxi6bn",
   "tg_tokens": null,
   "tps": 79.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-c0uxi6bn-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.4 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2917.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 32k 2,917 77.7 71.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c0uxi6bn",
   "source_url": "https://omlx.ai/benchmarks/performance/c0uxi6bn",
   "tg_tokens": null,
   "tps": 77.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-c0uxi6bn-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.9 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2760.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 128k 2,760 67.9 74.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c0uxi6bn",
   "source_url": "https://omlx.ai/benchmarks/performance/c0uxi6bn",
   "tg_tokens": null,
   "tps": 67.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-buslzy25-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 45.8 GB at this context; duplicate traces of the same session: jbisfojr, lev94rnu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 282.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-fp16-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 282.1 32.8 45.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/buslzy25",
   "source_url": "https://omlx.ai/benchmarks/performance/buslzy25",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-buslzy25-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 47.4 GB at this context; duplicate traces of the same session: jbisfojr, lev94rnu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 252.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-fp16-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 252.8 45.0 47.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/buslzy25",
   "source_url": "https://omlx.ai/benchmarks/performance/buslzy25",
   "tg_tokens": null,
   "tps": 45.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-buslzy25-ctx65536",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 50.1 GB at this context; duplicate traces of the same session: jbisfojr, lev94rnu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 213.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-fp16-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 213.6 27.5 50.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/buslzy25",
   "source_url": "https://omlx.ai/benchmarks/performance/buslzy25",
   "tg_tokens": null,
   "tps": 27.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pjh46z9j-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; Swift 1.5 build; oMLX community benchmark trace, single run at 16k context; peak memory 49.6 GB; TTFT 75068 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 218.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k tokens 218.3 PP tok/s 7.4 TG tok/s 75068 TTFT (ms) 49.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pjh46z9j",
   "source_url": "https://omlx.ai/benchmarks/performance/pjh46z9j",
   "tg_tokens": null,
   "tps": 7.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-p89ymuor-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 29.4 GB at this context; duplicate traces of the same session: y7jihb0w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 79.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 79.2 7.9 29.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p89ymuor",
   "source_url": "https://omlx.ai/benchmarks/performance/p89ymuor",
   "tg_tokens": null,
   "tps": 7.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-p89ymuor-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 32.2 GB at this context; duplicate traces of the same session: y7jihb0w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 286.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 286.8 8.7 32.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p89ymuor",
   "source_url": "https://omlx.ai/benchmarks/performance/p89ymuor",
   "tg_tokens": null,
   "tps": 8.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-sjls09hu-ctx16384",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: pze9p9w2, z5g6u8re; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 570.4,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 2bit; 16k 570.4 37.4 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sjls09hu",
   "source_url": "https://omlx.ai/benchmarks/performance/sjls09hu",
   "tg_tokens": null,
   "tps": 37.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-sjls09hu-ctx32768",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: pze9p9w2, z5g6u8re; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 489.1,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 2bit; 32k 489.1 31.5 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sjls09hu",
   "source_url": "https://omlx.ai/benchmarks/performance/sjls09hu",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-sjls09hu-ctx65536",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: pze9p9w2, z5g6u8re; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 363.5,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 2bit; 64k 363.5 24.1 21.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sjls09hu",
   "source_url": "https://omlx.ai/benchmarks/performance/sjls09hu",
   "tg_tokens": null,
   "tps": 24.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-qqg2mwow-ctx1024",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP; FP16; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 9hyxkiua; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 759.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP-FP16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 759.0 74.1 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qqg2mwow",
   "source_url": "https://omlx.ai/benchmarks/performance/qqg2mwow",
   "tg_tokens": null,
   "tps": 74.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-qqg2mwow-ctx4096",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP; FP16; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 9hyxkiua; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 769.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP-FP16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 769.9 69.7 22.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qqg2mwow",
   "source_url": "https://omlx.ai/benchmarks/performance/qqg2mwow",
   "tg_tokens": null,
   "tps": 69.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-53nu1fdg-ctx32768",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 240.7 GB; TTFT 56514 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 579.8,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-oQ3e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 32k tokens 579.8 PP tok/s 36.7 TG tok/s 56514 TTFT (ms) 240.7 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/53nu1fdg",
   "source_url": "https://omlx.ai/benchmarks/performance/53nu1fdg",
   "tg_tokens": null,
   "tps": 36.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-13kqa4yp-ctx1024",
   "model": "K2-Horizon-MoVA-36B-A4B",
   "notes": "FP16; oMLX community benchmark trace; peak memory 25.6 GB at this context; duplicate traces of the same session: eddm4ig4; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 424.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "K2-Horizon-MoVA-36B-A4B-4bit-FP16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 424.2 31.5 25.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/13kqa4yp",
   "source_url": "https://omlx.ai/benchmarks/performance/13kqa4yp",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-13kqa4yp-ctx4096",
   "model": "K2-Horizon-MoVA-36B-A4B",
   "notes": "FP16; oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: eddm4ig4; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 408.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "K2-Horizon-MoVA-36B-A4B-4bit-FP16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 408.1 28.6 26.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/13kqa4yp",
   "source_url": "https://omlx.ai/benchmarks/performance/13kqa4yp",
   "tg_tokens": null,
   "tps": 28.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-w01pgj7z-ctx1024",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: cd4gi2uw, u0ix98w5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2759.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ6 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 2,759 133.0 26.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w01pgj7z",
   "source_url": "https://omlx.ai/benchmarks/performance/w01pgj7z",
   "tg_tokens": null,
   "tps": 133.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-w01pgj7z-ctx8192",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: cd4gi2uw, u0ix98w5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3607.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ6 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 3,607 112.9 27.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w01pgj7z",
   "source_url": "https://omlx.ai/benchmarks/performance/w01pgj7z",
   "tg_tokens": null,
   "tps": 112.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-w01pgj7z-ctx199680",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: cd4gi2uw, u0ix98w5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2197.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ6 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 195k 2,197 72.3 28.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w01pgj7z",
   "source_url": "https://omlx.ai/benchmarks/performance/w01pgj7z",
   "tg_tokens": null,
   "tps": 72.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-wo75ljkr-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: kqj11738; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2150.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 2,150 91.5 35.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wo75ljkr",
   "source_url": "https://omlx.ai/benchmarks/performance/wo75ljkr",
   "tg_tokens": null,
   "tps": 91.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-wo75ljkr-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: kqj11738; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3369.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 3,369 89.7 37.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wo75ljkr",
   "source_url": "https://omlx.ai/benchmarks/performance/wo75ljkr",
   "tg_tokens": null,
   "tps": 89.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xk88y41r-ctx65536",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 62.7 GB; TTFT 65375 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1003.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k tokens 1,003 PP tok/s 55.9 TG tok/s 65375 TTFT (ms) 62.7 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xk88y41r",
   "source_url": "https://omlx.ai/benchmarks/performance/xk88y41r",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rkjek0pt-ctx65536",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 28.9 GB; TTFT 47553 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1378.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.5-4B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 unknown; 64k tokens 1,378 PP tok/s 54.3 TG tok/s 47553 TTFT (ms) 28.9 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rkjek0pt",
   "source_url": "https://omlx.ai/benchmarks/performance/rkjek0pt",
   "tg_tokens": null,
   "tps": 54.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-z0cctc91-ctx65536",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 37.8 GB; TTFT 70549 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 928.9,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.5-9B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 unknown; 64k tokens 928.9 PP tok/s 34.9 TG tok/s 70549 TTFT (ms) 37.8 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z0cctc91",
   "source_url": "https://omlx.ai/benchmarks/performance/z0cctc91",
   "tg_tokens": null,
   "tps": 34.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1e7mf6wv-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 39.3 GB; TTFT 46633 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1405.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k tokens 1,405 PP tok/s 79.4 TG tok/s 46633 TTFT (ms) 39.3 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1e7mf6wv",
   "source_url": "https://omlx.ai/benchmarks/performance/1e7mf6wv",
   "tg_tokens": null,
   "tps": 79.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-iz6u57v2-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 55.4 GB; TTFT 47112 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1391.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 64k tokens 1,391 PP tok/s 65.1 TG tok/s 47112 TTFT (ms) 55.4 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iz6u57v2",
   "source_url": "https://omlx.ai/benchmarks/performance/iz6u57v2",
   "tg_tokens": null,
   "tps": 65.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-8jb0uulg-ctx65536",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 39.3 GB; TTFT 46627 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1406.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k tokens 1,406 PP tok/s 79.9 TG tok/s 46627 TTFT (ms) 39.3 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8jb0uulg",
   "source_url": "https://omlx.ai/benchmarks/performance/8jb0uulg",
   "tg_tokens": null,
   "tps": 79.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-kh1aawll-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: ya2wp4bq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2663.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 2,663 91.9 35.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kh1aawll",
   "source_url": "https://omlx.ai/benchmarks/performance/kh1aawll",
   "tg_tokens": null,
   "tps": 91.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-kh1aawll-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: ya2wp4bq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3381.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 3,381 89.9 37.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kh1aawll",
   "source_url": "https://omlx.ai/benchmarks/performance/kh1aawll",
   "tg_tokens": null,
   "tps": 89.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-zejrsyfa-ctx1024",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "FP16; oMLX community benchmark trace; peak memory 12.8 GB at this context; duplicate traces of the same session: 28qo3xw5; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 792.2,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B-FP16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 792.2 61.5 12.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zejrsyfa",
   "source_url": "https://omlx.ai/benchmarks/performance/zejrsyfa",
   "tg_tokens": null,
   "tps": 61.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-zejrsyfa-ctx4096",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "FP16; oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: 28qo3xw5; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 663.1,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B-FP16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 663.1 64.6 14.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zejrsyfa",
   "source_url": "https://omlx.ai/benchmarks/performance/zejrsyfa",
   "tg_tokens": null,
   "tps": 64.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-aix61noo-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: r69m096x; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1368.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,368 55.5 16.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aix61noo",
   "source_url": "https://omlx.ai/benchmarks/performance/aix61noo",
   "tg_tokens": null,
   "tps": 55.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-aix61noo-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: r69m096x; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1559.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,559 54.0 17.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aix61noo",
   "source_url": "https://omlx.ai/benchmarks/performance/aix61noo",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-aix61noo-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: r69m096x; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1552.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 1,552 52.9 17.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aix61noo",
   "source_url": "https://omlx.ai/benchmarks/performance/aix61noo",
   "tg_tokens": null,
   "tps": 52.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-aix61noo-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: r69m096x; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1504.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,504 50.7 17.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aix61noo",
   "source_url": "https://omlx.ai/benchmarks/performance/aix61noo",
   "tg_tokens": null,
   "tps": 50.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-wdh8fmz1-ctx1024",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.8 GB; TTFT 2660 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 384.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Alis-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k tokens 384.9 PP tok/s 55.7 TG tok/s 2660 TTFT (ms) 19.8 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wdh8fmz1",
   "source_url": "https://omlx.ai/benchmarks/performance/wdh8fmz1",
   "tg_tokens": null,
   "tps": 55.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-lrpxqn7p-ctx1024",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 377.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Alis-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 377.0 53.2 19.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lrpxqn7p",
   "source_url": "https://omlx.ai/benchmarks/performance/lrpxqn7p",
   "tg_tokens": null,
   "tps": 53.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lrpxqn7p-ctx4096",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 405.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Alis-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 405.3 44.7 23.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lrpxqn7p",
   "source_url": "https://omlx.ai/benchmarks/performance/lrpxqn7p",
   "tg_tokens": null,
   "tps": 44.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-lrpxqn7p-ctx32768",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 344.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Alis-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 344.6 68.5 26.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lrpxqn7p",
   "source_url": "https://omlx.ai/benchmarks/performance/lrpxqn7p",
   "tg_tokens": null,
   "tps": 68.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-44a5ky9m-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: lyfdpo3a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 808.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 808.0 50.0 33.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/44a5ky9m",
   "source_url": "https://omlx.ai/benchmarks/performance/44a5ky9m",
   "tg_tokens": null,
   "tps": 50.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-44a5ky9m-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: lyfdpo3a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 895.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 895.1 44.9 34.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/44a5ky9m",
   "source_url": "https://omlx.ai/benchmarks/performance/44a5ky9m",
   "tg_tokens": null,
   "tps": 44.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wufbjtf0-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: 2vh2vtj4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 807.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 807.5 49.1 33.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wufbjtf0",
   "source_url": "https://omlx.ai/benchmarks/performance/wufbjtf0",
   "tg_tokens": null,
   "tps": 49.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wufbjtf0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: 2vh2vtj4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 923.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 923.3 41.0 34.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wufbjtf0",
   "source_url": "https://omlx.ai/benchmarks/performance/wufbjtf0",
   "tg_tokens": null,
   "tps": 41.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-vnp9gadu-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.6 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 310.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 310.0 41.2 30.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vnp9gadu",
   "source_url": "https://omlx.ai/benchmarks/performance/vnp9gadu",
   "tg_tokens": null,
   "tps": 41.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-vnp9gadu-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.6 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 490.4,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 490.4 65.2 33.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vnp9gadu",
   "source_url": "https://omlx.ai/benchmarks/performance/vnp9gadu",
   "tg_tokens": null,
   "tps": 65.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-b6qm5wl3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 2hbfynia; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 321.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 321.4 45.5 18.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b6qm5wl3",
   "source_url": "https://omlx.ai/benchmarks/performance/b6qm5wl3",
   "tg_tokens": null,
   "tps": 45.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-b6qm5wl3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 2hbfynia; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 345.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 345.6 61.7 22.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b6qm5wl3",
   "source_url": "https://omlx.ai/benchmarks/performance/b6qm5wl3",
   "tg_tokens": null,
   "tps": 61.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-418pxxt9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 68p5d997, jgrwv2ht, p7ln41gs; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 400.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 400.8 39.7 18.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/418pxxt9",
   "source_url": "https://omlx.ai/benchmarks/performance/418pxxt9",
   "tg_tokens": null,
   "tps": 39.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-418pxxt9-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 68p5d997, jgrwv2ht, p7ln41gs; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 399.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 399.6 25.0 19.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/418pxxt9",
   "source_url": "https://omlx.ai/benchmarks/performance/418pxxt9",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-418pxxt9-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 68p5d997, jgrwv2ht, p7ln41gs; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 384.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 384.4 33.5 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/418pxxt9",
   "source_url": "https://omlx.ai/benchmarks/performance/418pxxt9",
   "tg_tokens": null,
   "tps": 33.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-418pxxt9-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 68p5d997, jgrwv2ht, p7ln41gs; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 358.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 358.7 40.7 20.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/418pxxt9",
   "source_url": "https://omlx.ai/benchmarks/performance/418pxxt9",
   "tg_tokens": null,
   "tps": 40.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-12v9o6wj-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 581.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 581.8 37.1 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/12v9o6wj",
   "source_url": "https://omlx.ai/benchmarks/performance/12v9o6wj",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-12v9o6wj-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 682.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 682.7 55.6 19.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/12v9o6wj",
   "source_url": "https://omlx.ai/benchmarks/performance/12v9o6wj",
   "tg_tokens": null,
   "tps": 55.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-12v9o6wj-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 708.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 708.5 40.9 19.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/12v9o6wj",
   "source_url": "https://omlx.ai/benchmarks/performance/12v9o6wj",
   "tg_tokens": null,
   "tps": 40.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-umj5docu-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 287.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 287.7 23.5 18.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/umj5docu",
   "source_url": "https://omlx.ai/benchmarks/performance/umj5docu",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-umj5docu-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 300.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 300.8 20.8 19.5 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/umj5docu",
   "source_url": "https://omlx.ai/benchmarks/performance/umj5docu",
   "tg_tokens": null,
   "tps": 20.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-n0qsvr91-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: axp0qbr0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 317.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 317.9 42.0 33.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n0qsvr91",
   "source_url": "https://omlx.ai/benchmarks/performance/n0qsvr91",
   "tg_tokens": null,
   "tps": 42.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-n0qsvr91-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: axp0qbr0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 333.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 333.1 48.9 37.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n0qsvr91",
   "source_url": "https://omlx.ai/benchmarks/performance/n0qsvr91",
   "tg_tokens": null,
   "tps": 48.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wtymrw51-ctx1024",
   "model": "Qwen3.8-Flash-Next-Dynamic",
   "notes": "BF16; PLE; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 82.2 GB at this context; duplicate traces of the same session: pc4xot10, xddo664s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1306.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Dynamic-oQ5e-BF16-PLE-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,306 67.6 82.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wtymrw51",
   "source_url": "https://omlx.ai/benchmarks/performance/wtymrw51",
   "tg_tokens": null,
   "tps": 67.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wtymrw51-ctx8192",
   "model": "Qwen3.8-Flash-Next-Dynamic",
   "notes": "BF16; PLE; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 83.1 GB at this context; duplicate traces of the same session: pc4xot10, xddo664s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1967.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Dynamic-oQ5e-BF16-PLE-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,967 73.6 83.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wtymrw51",
   "source_url": "https://omlx.ai/benchmarks/performance/wtymrw51",
   "tg_tokens": null,
   "tps": 73.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wtymrw51-ctx199680",
   "model": "Qwen3.8-Flash-Next-Dynamic",
   "notes": "BF16; PLE; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 91.3 GB at this context; duplicate traces of the same session: pc4xot10, xddo664s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2053.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Dynamic-oQ5e-BF16-PLE-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,053 52.2 91.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wtymrw51",
   "source_url": "https://omlx.ai/benchmarks/performance/wtymrw51",
   "tg_tokens": null,
   "tps": 52.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-vr64pssq-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: l9hx3mgi; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 117.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 117.1 12.1 26.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vr64pssq",
   "source_url": "https://omlx.ai/benchmarks/performance/vr64pssq",
   "tg_tokens": null,
   "tps": 12.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-vr64pssq-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.2 GB at this context; duplicate traces of the same session: l9hx3mgi; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 286.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 286.0 12.5 27.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vr64pssq",
   "source_url": "https://omlx.ai/benchmarks/performance/vr64pssq",
   "tg_tokens": null,
   "tps": 12.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xxqrb3sj-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.4 GB at this context; duplicate traces of the same session: pmsqruun; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 70.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 70.4 8.8 29.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xxqrb3sj",
   "source_url": "https://omlx.ai/benchmarks/performance/xxqrb3sj",
   "tg_tokens": null,
   "tps": 8.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xxqrb3sj-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.2 GB at this context; duplicate traces of the same session: pmsqruun; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 256.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 256.9 7.1 32.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xxqrb3sj",
   "source_url": "https://omlx.ai/benchmarks/performance/xxqrb3sj",
   "tg_tokens": null,
   "tps": 7.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0v8v99cv-ctx1024",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 80.0 GB at this context; duplicate traces of the same session: znkgte2a; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 565.4,
   "provenance": "community",
   "quant": "2.4bit mixed",
   "quote": "DeepSeek-V4-Flash-0731-2.4bit-mixed on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 2bit; 1k 565.4 31.4 80.0 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0v8v99cv",
   "source_url": "https://omlx.ai/benchmarks/performance/0v8v99cv",
   "tg_tokens": null,
   "tps": 31.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0v8v99cv-ctx4096",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 80.8 GB at this context; duplicate traces of the same session: znkgte2a; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 615.7,
   "provenance": "community",
   "quant": "2.4bit mixed",
   "quote": "DeepSeek-V4-Flash-0731-2.4bit-mixed on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 2bit; 4k 615.7 28.6 80.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0v8v99cv",
   "source_url": "https://omlx.ai/benchmarks/performance/0v8v99cv",
   "tg_tokens": null,
   "tps": 28.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-nctg3f3f-ctx32768",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 240.7 GB; TTFT 57171 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 573.2,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-oQ3e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 32k tokens 573.2 PP tok/s 38.4 TG tok/s 57171 TTFT (ms) 240.7 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nctg3f3f",
   "source_url": "https://omlx.ai/benchmarks/performance/nctg3f3f",
   "tg_tokens": null,
   "tps": 38.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-evcsre5c-ctx16384",
   "model": "Ling-3.0-tiny",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 6.6 GB; TTFT 18754 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 873.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-tiny-oQ4e on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 4bit; 16k tokens 873.6 PP tok/s 63.9 TG tok/s 18754 TTFT (ms) 6.6 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/evcsre5c",
   "source_url": "https://omlx.ai/benchmarks/performance/evcsre5c",
   "tg_tokens": null,
   "tps": 63.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-exwibgsy-ctx1024",
   "model": "Nex-N2.5-mini-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: kaqurka2; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 558.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Nex-N2.5-mini-Uncensored-oQ8e on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 1k 558.1 58.9 35.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/exwibgsy",
   "source_url": "https://omlx.ai/benchmarks/performance/exwibgsy",
   "tg_tokens": null,
   "tps": 58.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-exwibgsy-ctx4096",
   "model": "Nex-N2.5-mini-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 37.6 GB at this context; duplicate traces of the same session: kaqurka2; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 616.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Nex-N2.5-mini-Uncensored-oQ8e on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 4k 616.5 57.0 37.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/exwibgsy",
   "source_url": "https://omlx.ai/benchmarks/performance/exwibgsy",
   "tg_tokens": null,
   "tps": 57.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-bey829p1-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: cg9x73rk; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 899.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 899.4 53.0 21.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bey829p1",
   "source_url": "https://omlx.ai/benchmarks/performance/bey829p1",
   "tg_tokens": null,
   "tps": 53.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-bey829p1-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: cg9x73rk; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1671.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,671 50.9 21.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bey829p1",
   "source_url": "https://omlx.ai/benchmarks/performance/bey829p1",
   "tg_tokens": null,
   "tps": 50.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ivr92ilx-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.1 GB; TTFT 597 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1715.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k tokens 1,715 PP tok/s 99.2 TG tok/s 597 TTFT (ms) 19.1 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ivr92ilx",
   "source_url": "https://omlx.ai/benchmarks/performance/ivr92ilx",
   "tg_tokens": null,
   "tps": 99.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-57nv20cb-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: f9f9g897, ve2ai5p4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1684.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 1,684 98.7 19.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/57nv20cb",
   "source_url": "https://omlx.ai/benchmarks/performance/57nv20cb",
   "tg_tokens": null,
   "tps": 98.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-57nv20cb-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: f9f9g897, ve2ai5p4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2010.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 2,010 93.4 20.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/57nv20cb",
   "source_url": "https://omlx.ai/benchmarks/performance/57nv20cb",
   "tg_tokens": null,
   "tps": 93.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-57nv20cb-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: f9f9g897, ve2ai5p4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1504.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 1,504 78.9 22.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/57nv20cb",
   "source_url": "https://omlx.ai/benchmarks/performance/57nv20cb",
   "tg_tokens": null,
   "tps": 78.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3pqccjl5-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 00lkrou9, q01cevij; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1666.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 1,666 98.2 19.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3pqccjl5",
   "source_url": "https://omlx.ai/benchmarks/performance/3pqccjl5",
   "tg_tokens": null,
   "tps": 98.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-3pqccjl5-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 00lkrou9, q01cevij; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2023.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 2,023 93.9 20.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3pqccjl5",
   "source_url": "https://omlx.ai/benchmarks/performance/3pqccjl5",
   "tg_tokens": null,
   "tps": 93.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-3pqccjl5-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 00lkrou9, q01cevij; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1510.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 1,510 79.3 22.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3pqccjl5",
   "source_url": "https://omlx.ai/benchmarks/performance/3pqccjl5",
   "tg_tokens": null,
   "tps": 79.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-brdyln4g-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 37.2 GB; TTFT 627 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1634.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k tokens 1,634 PP tok/s 99.0 TG tok/s 627 TTFT (ms) 37.2 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/brdyln4g",
   "source_url": "https://omlx.ai/benchmarks/performance/brdyln4g",
   "tg_tokens": null,
   "tps": 99.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-sb9563gp-ctx32768",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: jqgazn3m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1233.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,233 144.8 24.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sb9563gp",
   "source_url": "https://omlx.ai/benchmarks/performance/sb9563gp",
   "tg_tokens": null,
   "tps": 144.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-sb9563gp-ctx65536",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.7 GB at this context; duplicate traces of the same session: jqgazn3m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 823.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 823.2 103.9 25.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sb9563gp",
   "source_url": "https://omlx.ai/benchmarks/performance/sb9563gp",
   "tg_tokens": null,
   "tps": 103.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-sb9563gp-ctx131072",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: jqgazn3m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 453.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 453.6 63.8 27.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sb9563gp",
   "source_url": "https://omlx.ai/benchmarks/performance/sb9563gp",
   "tg_tokens": null,
   "tps": 63.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-6sjlsyzo-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 17.3 GB; TTFT 52050 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 78.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k tokens 78.7 PP tok/s 15.1 TG tok/s 52050 TTFT (ms) 17.3 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6sjlsyzo",
   "source_url": "https://omlx.ai/benchmarks/performance/6sjlsyzo",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-3fm27foq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: yte3v4ct; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1368.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,368 55.5 16.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3fm27foq",
   "source_url": "https://omlx.ai/benchmarks/performance/3fm27foq",
   "tg_tokens": null,
   "tps": 55.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-3fm27foq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: yte3v4ct; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1559.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,559 54.0 17.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3fm27foq",
   "source_url": "https://omlx.ai/benchmarks/performance/3fm27foq",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-3fm27foq-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: yte3v4ct; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1552.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 1,552 52.9 17.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3fm27foq",
   "source_url": "https://omlx.ai/benchmarks/performance/3fm27foq",
   "tg_tokens": null,
   "tps": 52.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-3fm27foq-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: yte3v4ct; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1504.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,504 50.7 17.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3fm27foq",
   "source_url": "https://omlx.ai/benchmarks/performance/3fm27foq",
   "tg_tokens": null,
   "tps": 50.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-sl04ujhq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: g5eo4ifo; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1356.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,356 55.5 16.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sl04ujhq",
   "source_url": "https://omlx.ai/benchmarks/performance/sl04ujhq",
   "tg_tokens": null,
   "tps": 55.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-sl04ujhq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: g5eo4ifo; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1560.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,560 54.0 17.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sl04ujhq",
   "source_url": "https://omlx.ai/benchmarks/performance/sl04ujhq",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-i3v4ap0j-ctx1024",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 6lxyo3yu; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 377.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Alis-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 377.0 53.2 19.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i3v4ap0j",
   "source_url": "https://omlx.ai/benchmarks/performance/i3v4ap0j",
   "tg_tokens": null,
   "tps": 53.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-i3v4ap0j-ctx4096",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 6lxyo3yu; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 405.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Alis-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 405.3 44.7 23.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i3v4ap0j",
   "source_url": "https://omlx.ai/benchmarks/performance/i3v4ap0j",
   "tg_tokens": null,
   "tps": 44.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-i3v4ap0j-ctx32768",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: 6lxyo3yu; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 344.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Alis-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 344.6 68.5 26.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i3v4ap0j",
   "source_url": "https://omlx.ai/benchmarks/performance/i3v4ap0j",
   "tg_tokens": null,
   "tps": 68.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-b3awk3r5-ctx1024",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: jywoqdvc, q21nxo7h; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 391.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Alis-MLX-6bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 391.8 41.2 27.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b3awk3r5",
   "source_url": "https://omlx.ai/benchmarks/performance/b3awk3r5",
   "tg_tokens": null,
   "tps": 41.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-b3awk3r5-ctx4096",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 31.0 GB at this context; duplicate traces of the same session: jywoqdvc, q21nxo7h; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 417.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Alis-MLX-6bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 417.7 36.3 31.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b3awk3r5",
   "source_url": "https://omlx.ai/benchmarks/performance/b3awk3r5",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-b3awk3r5-ctx32768",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 34.1 GB at this context; duplicate traces of the same session: jywoqdvc, q21nxo7h; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 354.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Alis-MLX-6bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 354.6 46.4 34.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b3awk3r5",
   "source_url": "https://omlx.ai/benchmarks/performance/b3awk3r5",
   "tg_tokens": null,
   "tps": 46.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-r2ybx7oo-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.8 GB at this context; duplicate traces of the same session: ofpduu8t; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 337.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 337.2 22.9 31.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r2ybx7oo",
   "source_url": "https://omlx.ai/benchmarks/performance/r2ybx7oo",
   "tg_tokens": null,
   "tps": 22.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-r2ybx7oo-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.9 GB at this context; duplicate traces of the same session: ofpduu8t; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 16k 323.6 21.8 32.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r2ybx7oo",
   "source_url": "https://omlx.ai/benchmarks/performance/r2ybx7oo",
   "tg_tokens": null,
   "tps": 21.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ogluk6fk-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 42.4 GB; TTFT 238771 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 274.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k tokens 274.5 PP tok/s 27.3 TG tok/s 238771 TTFT (ms) 42.4 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ogluk6fk",
   "source_url": "https://omlx.ai/benchmarks/performance/ogluk6fk",
   "tg_tokens": null,
   "tps": 27.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-1mkwl0m8-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 9zgostk5, qc38zzo5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 460.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 460.9 55.9 21.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1mkwl0m8",
   "source_url": "https://omlx.ai/benchmarks/performance/1mkwl0m8",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1mkwl0m8-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: 9zgostk5, qc38zzo5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 431.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 431.0 36.1 23.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1mkwl0m8",
   "source_url": "https://omlx.ai/benchmarks/performance/1mkwl0m8",
   "tg_tokens": null,
   "tps": 36.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1mkwl0m8-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 9zgostk5, qc38zzo5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 447.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 447.1 25.5 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1mkwl0m8",
   "source_url": "https://omlx.ai/benchmarks/performance/1mkwl0m8",
   "tg_tokens": null,
   "tps": 25.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-gb0xmbmg-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 1jxjqdjt, hrqn002w; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 553.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 553.8 44.5 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gb0xmbmg",
   "source_url": "https://omlx.ai/benchmarks/performance/gb0xmbmg",
   "tg_tokens": null,
   "tps": 44.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-gb0xmbmg-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 1jxjqdjt, hrqn002w; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 641.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 641.7 28.0 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gb0xmbmg",
   "source_url": "https://omlx.ai/benchmarks/performance/gb0xmbmg",
   "tg_tokens": null,
   "tps": 28.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-gb0xmbmg-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 1jxjqdjt, hrqn002w; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 676.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 676.3 21.7 19.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gb0xmbmg",
   "source_url": "https://omlx.ai/benchmarks/performance/gb0xmbmg",
   "tg_tokens": null,
   "tps": 21.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-mw61omxw-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 287.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 287.7 23.5 18.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mw61omxw",
   "source_url": "https://omlx.ai/benchmarks/performance/mw61omxw",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-mw61omxw-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 300.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 300.8 20.8 19.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mw61omxw",
   "source_url": "https://omlx.ai/benchmarks/performance/mw61omxw",
   "tg_tokens": null,
   "tps": 20.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-4m52s0k4-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: duau4r2f; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 286.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 286.3 17.8 18.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4m52s0k4",
   "source_url": "https://omlx.ai/benchmarks/performance/4m52s0k4",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-4m52s0k4-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: duau4r2f; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 297.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 297.7 18.9 19.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4m52s0k4",
   "source_url": "https://omlx.ai/benchmarks/performance/4m52s0k4",
   "tg_tokens": null,
   "tps": 18.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-req7couh-ctx65536",
   "model": "Qwopus3.6-27B-Coder",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 42.4 GB; TTFT 238831 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 274.4,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwopus3.6-27B-Coder-oQ4-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k tokens 274.4 PP tok/s 27.3 TG tok/s 238831 TTFT (ms) 42.4 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/req7couh",
   "source_url": "https://omlx.ai/benchmarks/performance/req7couh",
   "tg_tokens": null,
   "tps": 27.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-l47h65vp-ctx65536",
   "model": "gemma-4-26B-A4B-it-heretic",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 44.8 GB; TTFT 51870 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1264.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-heretic-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k tokens 1,264 PP tok/s 54.2 TG tok/s 51870 TTFT (ms) 44.8 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l47h65vp",
   "source_url": "https://omlx.ai/benchmarks/performance/l47h65vp",
   "tg_tokens": null,
   "tps": 54.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-d34fci52-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.2 GB at this context; duplicate traces of the same session: yk2vz1vb; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1269.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 1,269 40.5 16.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d34fci52",
   "source_url": "https://omlx.ai/benchmarks/performance/d34fci52",
   "tg_tokens": null,
   "tps": 40.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-d34fci52-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: yk2vz1vb; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1208.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,208 44.7 16.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d34fci52",
   "source_url": "https://omlx.ai/benchmarks/performance/d34fci52",
   "tg_tokens": null,
   "tps": 44.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4d1bkwjc-ctx1024",
   "model": "Qwen3-VL-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 6xaulw19; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2416.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3-VL-30B-A3B-Instruct-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 2,416 104.7 16.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4d1bkwjc",
   "source_url": "https://omlx.ai/benchmarks/performance/4d1bkwjc",
   "tg_tokens": null,
   "tps": 104.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4d1bkwjc-ctx4096",
   "model": "Qwen3-VL-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: 6xaulw19; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2597.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3-VL-30B-A3B-Instruct-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 2,597 97.5 17.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4d1bkwjc",
   "source_url": "https://omlx.ai/benchmarks/performance/4d1bkwjc",
   "tg_tokens": null,
   "tps": 97.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-sr0db9ut-ctx32768",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 333.1 GB; TTFT 89923 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 364.4,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-oQ3e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 32k tokens 364.4 PP tok/s 37.9 TG tok/s 89923 TTFT (ms) 333.1 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sr0db9ut",
   "source_url": "https://omlx.ai/benchmarks/performance/sr0db9ut",
   "tg_tokens": null,
   "tps": 37.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-u1o7za0v-ctx32768",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 240.7 GB; TTFT 56301 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 582.0,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-oQ3e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 32k tokens 582.0 PP tok/s 38.0 TG tok/s 56301 TTFT (ms) 240.7 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u1o7za0v",
   "source_url": "https://omlx.ai/benchmarks/performance/u1o7za0v",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-wvtesgs5-ctx32768",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace, single run at 32k context; peak memory 289.3 GB; TTFT 51692 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 633.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 32k tokens 633.9 PP tok/s 33.8 TG tok/s 51692 TTFT (ms) 289.3 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wvtesgs5",
   "source_url": "https://omlx.ai/benchmarks/performance/wvtesgs5",
   "tg_tokens": null,
   "tps": 33.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-u0uovqpz-ctx1024",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: 4fmrq4nb, 53x6288q, 6kxv0puc, 7eje0aaj, p27cnz9t, xojhgjy4; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3677.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Mellum2-12B-A2.5B-Instruct-mxfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 3,677 118.8 6.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u0uovqpz",
   "source_url": "https://omlx.ai/benchmarks/performance/u0uovqpz",
   "tg_tokens": null,
   "tps": 118.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-u0uovqpz-ctx4096",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: 4fmrq4nb, 53x6288q, 6kxv0puc, 7eje0aaj, p27cnz9t, xojhgjy4; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3888.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Mellum2-12B-A2.5B-Instruct-mxfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 3,888 119.0 6.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u0uovqpz",
   "source_url": "https://omlx.ai/benchmarks/performance/u0uovqpz",
   "tg_tokens": null,
   "tps": 119.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-u0uovqpz-ctx8192",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: 4fmrq4nb, 53x6288q, 6kxv0puc, 7eje0aaj, p27cnz9t, xojhgjy4; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3877.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Mellum2-12B-A2.5B-Instruct-mxfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 3,877 108.5 6.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u0uovqpz",
   "source_url": "https://omlx.ai/benchmarks/performance/u0uovqpz",
   "tg_tokens": null,
   "tps": 108.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-u0uovqpz-ctx16384",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: 4fmrq4nb, 53x6288q, 6kxv0puc, 7eje0aaj, p27cnz9t, xojhgjy4; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3767.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Mellum2-12B-A2.5B-Instruct-mxfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 3,767 94.0 6.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u0uovqpz",
   "source_url": "https://omlx.ai/benchmarks/performance/u0uovqpz",
   "tg_tokens": null,
   "tps": 94.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-u0uovqpz-ctx32768",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 7.2 GB at this context; duplicate traces of the same session: 4fmrq4nb, 53x6288q, 6kxv0puc, 7eje0aaj, p27cnz9t, xojhgjy4; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3330.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Mellum2-12B-A2.5B-Instruct-mxfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 3,330 72.6 7.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u0uovqpz",
   "source_url": "https://omlx.ai/benchmarks/performance/u0uovqpz",
   "tg_tokens": null,
   "tps": 72.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-u0uovqpz-ctx65536",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 7.8 GB at this context; duplicate traces of the same session: 4fmrq4nb, 53x6288q, 6kxv0puc, 7eje0aaj, p27cnz9t, xojhgjy4; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2646.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Mellum2-12B-A2.5B-Instruct-mxfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 2,646 51.8 7.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u0uovqpz",
   "source_url": "https://omlx.ai/benchmarks/performance/u0uovqpz",
   "tg_tokens": null,
   "tps": 51.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-u0uovqpz-ctx131072",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 9.5 GB at this context; duplicate traces of the same session: 4fmrq4nb, 53x6288q, 6kxv0puc, 7eje0aaj, p27cnz9t, xojhgjy4; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1922.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Mellum2-12B-A2.5B-Instruct-mxfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 128k 1,922 32.4 9.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u0uovqpz",
   "source_url": "https://omlx.ai/benchmarks/performance/u0uovqpz",
   "tg_tokens": null,
   "tps": 32.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-sj4xgmgd-ctx1024",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: 48z2cjox, 8vfxoy7n, btxw5zwq, sk4z2rlq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1073.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Mellum2-12B-A2.5B-Instruct-mxfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 1,073 128.8 6.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sj4xgmgd",
   "source_url": "https://omlx.ai/benchmarks/performance/sj4xgmgd",
   "tg_tokens": null,
   "tps": 128.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-sj4xgmgd-ctx4096",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: 48z2cjox, 8vfxoy7n, btxw5zwq, sk4z2rlq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2680.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Mellum2-12B-A2.5B-Instruct-mxfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 2,680 125.0 6.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sj4xgmgd",
   "source_url": "https://omlx.ai/benchmarks/performance/sj4xgmgd",
   "tg_tokens": null,
   "tps": 125.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-sj4xgmgd-ctx8192",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: 48z2cjox, 8vfxoy7n, btxw5zwq, sk4z2rlq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3980.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Mellum2-12B-A2.5B-Instruct-mxfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 3,980 125.2 6.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sj4xgmgd",
   "source_url": "https://omlx.ai/benchmarks/performance/sj4xgmgd",
   "tg_tokens": null,
   "tps": 125.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-sj4xgmgd-ctx16384",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: 48z2cjox, 8vfxoy7n, btxw5zwq, sk4z2rlq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3807.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Mellum2-12B-A2.5B-Instruct-mxfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 3,807 116.8 6.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sj4xgmgd",
   "source_url": "https://omlx.ai/benchmarks/performance/sj4xgmgd",
   "tg_tokens": null,
   "tps": 116.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-sj4xgmgd-ctx32768",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 7.2 GB at this context; duplicate traces of the same session: 48z2cjox, 8vfxoy7n, btxw5zwq, sk4z2rlq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3373.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Mellum2-12B-A2.5B-Instruct-mxfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 3,373 105.1 7.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sj4xgmgd",
   "source_url": "https://omlx.ai/benchmarks/performance/sj4xgmgd",
   "tg_tokens": null,
   "tps": 105.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-f5msu2b2-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build, MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: odya2g8r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1584.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fixed-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,584 98.7 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f5msu2b2",
   "source_url": "https://omlx.ai/benchmarks/performance/f5msu2b2",
   "tg_tokens": null,
   "tps": 98.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-f5msu2b2-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build, MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: odya2g8r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1956.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fixed-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,956 142.8 23.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f5msu2b2",
   "source_url": "https://omlx.ai/benchmarks/performance/f5msu2b2",
   "tg_tokens": null,
   "tps": 142.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-gw0zst00-ctx32768",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1233.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,233 144.8 24.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gw0zst00",
   "source_url": "https://omlx.ai/benchmarks/performance/gw0zst00",
   "tg_tokens": null,
   "tps": 144.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-gw0zst00-ctx65536",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 823.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 823.2 103.9 25.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gw0zst00",
   "source_url": "https://omlx.ai/benchmarks/performance/gw0zst00",
   "tg_tokens": null,
   "tps": 103.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-gw0zst00-ctx131072",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 453.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 453.6 63.8 27.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gw0zst00",
   "source_url": "https://omlx.ai/benchmarks/performance/gw0zst00",
   "tg_tokens": null,
   "tps": 63.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-778plsde-ctx1024",
   "model": "Qwen3.6-40B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 41.0 GB at this context; duplicate traces of the same session: dk6hxbzv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 79.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-40B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-Thinking-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 79.2 7.4 41.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/778plsde",
   "source_url": "https://omlx.ai/benchmarks/performance/778plsde",
   "tg_tokens": null,
   "tps": 7.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-778plsde-ctx4096",
   "model": "Qwen3.6-40B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 45.8 GB at this context; duplicate traces of the same session: dk6hxbzv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 84.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-40B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-Thinking-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 84.2 7.6 45.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/778plsde",
   "source_url": "https://omlx.ai/benchmarks/performance/778plsde",
   "tg_tokens": null,
   "tps": 7.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-n1ehmp22-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 07vdi654; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 322.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 322.6 45.5 19.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n1ehmp22",
   "source_url": "https://omlx.ai/benchmarks/performance/n1ehmp22",
   "tg_tokens": null,
   "tps": 45.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-n1ehmp22-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 07vdi654; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 337.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 337.7 45.7 21.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n1ehmp22",
   "source_url": "https://omlx.ai/benchmarks/performance/n1ehmp22",
   "tg_tokens": null,
   "tps": 45.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-zhpczjk7-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: xdx5mqih; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 334.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 334.7 39.4 16.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zhpczjk7",
   "source_url": "https://omlx.ai/benchmarks/performance/zhpczjk7",
   "tg_tokens": null,
   "tps": 39.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-zhpczjk7-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: xdx5mqih; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 352.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 352.9 38.1 19.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zhpczjk7",
   "source_url": "https://omlx.ai/benchmarks/performance/zhpczjk7",
   "tg_tokens": null,
   "tps": 38.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-ahsdey84-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 77.2,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 77.2 11.9 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ahsdey84",
   "source_url": "https://omlx.ai/benchmarks/performance/ahsdey84",
   "tg_tokens": null,
   "tps": 11.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-ahsdey84-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 85.1,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 85.1 14.8 23.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ahsdey84",
   "source_url": "https://omlx.ai/benchmarks/performance/ahsdey84",
   "tg_tokens": null,
   "tps": 14.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-na9oy9md-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: fx24j5xz; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 101.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 101.0 12.4 18.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/na9oy9md",
   "source_url": "https://omlx.ai/benchmarks/performance/na9oy9md",
   "tg_tokens": null,
   "tps": 12.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-na9oy9md-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: fx24j5xz; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 85.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 85.3 14.7 22.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/na9oy9md",
   "source_url": "https://omlx.ai/benchmarks/performance/na9oy9md",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-bs722r1g-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 3dchqd6g, lnr4ty9k; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 587.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 587.5 36.9 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bs722r1g",
   "source_url": "https://omlx.ai/benchmarks/performance/bs722r1g",
   "tg_tokens": null,
   "tps": 36.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-bs722r1g-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 3dchqd6g, lnr4ty9k; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 688.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 688.9 43.1 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bs722r1g",
   "source_url": "https://omlx.ai/benchmarks/performance/bs722r1g",
   "tg_tokens": null,
   "tps": 43.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-bs722r1g-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 3dchqd6g, lnr4ty9k; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 638.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 638.3 21.8 19.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bs722r1g",
   "source_url": "https://omlx.ai/benchmarks/performance/bs722r1g",
   "tg_tokens": null,
   "tps": 21.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-y4euw1m5-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: p08r008b, wxspp5sg; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 470.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 470.7 30.0 22.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y4euw1m5",
   "source_url": "https://omlx.ai/benchmarks/performance/y4euw1m5",
   "tg_tokens": null,
   "tps": 30.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-y4euw1m5-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: p08r008b, wxspp5sg; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 509.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 509.3 52.9 24.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y4euw1m5",
   "source_url": "https://omlx.ai/benchmarks/performance/y4euw1m5",
   "tg_tokens": null,
   "tps": 52.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-y4euw1m5-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: p08r008b, wxspp5sg; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 522.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 522.9 23.2 24.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y4euw1m5",
   "source_url": "https://omlx.ai/benchmarks/performance/y4euw1m5",
   "tg_tokens": null,
   "tps": 23.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-xe24ye2g-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: en01amdp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 246.6,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 5bit; 1k 246.6 43.0 23.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xe24ye2g",
   "source_url": "https://omlx.ai/benchmarks/performance/xe24ye2g",
   "tg_tokens": null,
   "tps": 43.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-xe24ye2g-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: en01amdp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 263.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 5bit; 4k 263.9 49.4 26.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xe24ye2g",
   "source_url": "https://omlx.ai/benchmarks/performance/xe24ye2g",
   "tg_tokens": null,
   "tps": 49.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-pcmblx5f-ctx1024",
   "model": "Qwen3.8-3.6-27B-blend",
   "notes": "blend of Qwen3.8 and Qwen3.6 27B; oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: 0e2un210, 4252xhwo, liw11mtt, uc32tvyt; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 466.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-3.6-27B-blend-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 466.6 27.1 16.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pcmblx5f",
   "source_url": "https://omlx.ai/benchmarks/performance/pcmblx5f",
   "tg_tokens": null,
   "tps": 27.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-pcmblx5f-ctx4096",
   "model": "Qwen3.8-3.6-27B-blend",
   "notes": "blend of Qwen3.8 and Qwen3.6 27B; oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: 0e2un210, 4252xhwo, liw11mtt, uc32tvyt; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 475.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-3.6-27B-blend-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 475.4 25.0 18.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pcmblx5f",
   "source_url": "https://omlx.ai/benchmarks/performance/pcmblx5f",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-pcmblx5f-ctx8192",
   "model": "Qwen3.8-3.6-27B-blend",
   "notes": "blend of Qwen3.8 and Qwen3.6 27B; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 0e2un210, 4252xhwo, liw11mtt, uc32tvyt; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 469.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-3.6-27B-blend-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 469.9 24.3 18.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pcmblx5f",
   "source_url": "https://omlx.ai/benchmarks/performance/pcmblx5f",
   "tg_tokens": null,
   "tps": 24.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-pcmblx5f-ctx16384",
   "model": "Qwen3.8-3.6-27B-blend",
   "notes": "blend of Qwen3.8 and Qwen3.6 27B; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: 0e2un210, 4252xhwo, liw11mtt, uc32tvyt; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 435.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-3.6-27B-blend-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 435.2 27.2 18.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pcmblx5f",
   "source_url": "https://omlx.ai/benchmarks/performance/pcmblx5f",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-pcmblx5f-ctx32768",
   "model": "Qwen3.8-3.6-27B-blend",
   "notes": "blend of Qwen3.8 and Qwen3.6 27B; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 0e2un210, 4252xhwo, liw11mtt, uc32tvyt; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 422.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-3.6-27B-blend-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 422.3 25.3 19.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pcmblx5f",
   "source_url": "https://omlx.ai/benchmarks/performance/pcmblx5f",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-qijimft9-ctx1024",
   "model": "Swift-Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: 3gtd79go, rm1rbl56; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 292.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Swift-Qwen3.8-27B on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 292.8 56.1 18.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qijimft9",
   "source_url": "https://omlx.ai/benchmarks/performance/qijimft9",
   "tg_tokens": null,
   "tps": 56.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-qijimft9-ctx8192",
   "model": "Swift-Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 3gtd79go, rm1rbl56; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 319.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Swift-Qwen3.8-27B on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 8k 319.1 61.3 23.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qijimft9",
   "source_url": "https://omlx.ai/benchmarks/performance/qijimft9",
   "tg_tokens": null,
   "tps": 61.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-qijimft9-ctx16384",
   "model": "Swift-Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: 3gtd79go, rm1rbl56; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 304.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Swift-Qwen3.8-27B on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 16k 304.5 59.0 24.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qijimft9",
   "source_url": "https://omlx.ai/benchmarks/performance/qijimft9",
   "tg_tokens": null,
   "tps": 59.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-v86evzy9-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3387.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 3,387 115.0 14.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v86evzy9",
   "source_url": "https://omlx.ai/benchmarks/performance/v86evzy9",
   "tg_tokens": null,
   "tps": 115.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-v86evzy9-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4221.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 4,221 118.4 14.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v86evzy9",
   "source_url": "https://omlx.ai/benchmarks/performance/v86evzy9",
   "tg_tokens": null,
   "tps": 118.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-cy7fl2ag-ctx1024",
   "model": "gemma-4-26B-A4B-it-heretic",
   "notes": "oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: 59lpzgwb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4060.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-heretic-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 4,060 119.7 14.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cy7fl2ag",
   "source_url": "https://omlx.ai/benchmarks/performance/cy7fl2ag",
   "tg_tokens": null,
   "tps": 119.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-cy7fl2ag-ctx4096",
   "model": "gemma-4-26B-A4B-it-heretic",
   "notes": "oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: 59lpzgwb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4508.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-heretic-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 4,508 114.0 14.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cy7fl2ag",
   "source_url": "https://omlx.ai/benchmarks/performance/cy7fl2ag",
   "tg_tokens": null,
   "tps": 114.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-a8752c3t-ctx1024",
   "model": "DeepSeek-R1-0528-Qwen3-8B",
   "notes": "oMLX community benchmark trace; peak memory 8.9 GB at this context; duplicate traces of the same session: lexc89px; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 918.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "DeepSeek-R1-0528-Qwen3-8B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 918.0 54.0 8.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a8752c3t",
   "source_url": "https://omlx.ai/benchmarks/performance/a8752c3t",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-a8752c3t-ctx4096",
   "model": "DeepSeek-R1-0528-Qwen3-8B",
   "notes": "oMLX community benchmark trace; peak memory 9.3 GB at this context; duplicate traces of the same session: lexc89px; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 820.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "DeepSeek-R1-0528-Qwen3-8B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 820.4 51.8 9.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a8752c3t",
   "source_url": "https://omlx.ai/benchmarks/performance/a8752c3t",
   "tg_tokens": null,
   "tps": 51.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-j5xyu3m6-ctx32768",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace, single run at 32k context; peak memory 289.3 GB; TTFT 53811 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 608.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 32k tokens 608.9 PP tok/s 34.3 TG tok/s 53811 TTFT (ms) 289.3 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j5xyu3m6",
   "source_url": "https://omlx.ai/benchmarks/performance/j5xyu3m6",
   "tg_tokens": null,
   "tps": 34.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ftbyr7u8-ctx32768",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace, single run at 32k context; peak memory 404.6 GB; TTFT 86326 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 379.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 32k tokens 379.6 PP tok/s 33.6 TG tok/s 86326 TTFT (ms) 404.6 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ftbyr7u8",
   "source_url": "https://omlx.ai/benchmarks/performance/ftbyr7u8",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-weeovtx3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 829.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ6e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 829.1 56.6 26.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/weeovtx3",
   "source_url": "https://omlx.ai/benchmarks/performance/weeovtx3",
   "tg_tokens": null,
   "tps": 56.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-weeovtx3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 918.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ6e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 918.3 47.5 27.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/weeovtx3",
   "source_url": "https://omlx.ai/benchmarks/performance/weeovtx3",
   "tg_tokens": null,
   "tps": 47.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-et8fq5pi-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 858.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ6e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 858.0 24.5 22.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/et8fq5pi",
   "source_url": "https://omlx.ai/benchmarks/performance/et8fq5pi",
   "tg_tokens": null,
   "tps": 24.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-et8fq5pi-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 924.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ6e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 924.5 24.2 23.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/et8fq5pi",
   "source_url": "https://omlx.ai/benchmarks/performance/et8fq5pi",
   "tg_tokens": null,
   "tps": 24.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-hdb7l6i5-ctx1024",
   "model": "Llama 3.1 8B",
   "notes": "oMLX community benchmark trace; peak memory 5.0 GB at this context; duplicate traces of the same session: pxc7e73v; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 563.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Meta-Llama-3.1-8B-Instruct-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 1k 563.0 70.3 5.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hdb7l6i5",
   "source_url": "https://omlx.ai/benchmarks/performance/hdb7l6i5",
   "tg_tokens": null,
   "tps": 70.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-hdb7l6i5-ctx4096",
   "model": "Llama 3.1 8B",
   "notes": "oMLX community benchmark trace; peak memory 5.4 GB at this context; duplicate traces of the same session: pxc7e73v; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 535.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Meta-Llama-3.1-8B-Instruct-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 4k 535.5 61.9 5.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hdb7l6i5",
   "source_url": "https://omlx.ai/benchmarks/performance/hdb7l6i5",
   "tg_tokens": null,
   "tps": 61.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-curkea1d-ctx1024",
   "model": "Muse Glimmer",
   "notes": "oMLX community benchmark trace; peak memory 15.2 GB at this context; duplicate traces of the same session: szlz46lz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 230.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-30B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 230.3 18.8 15.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/curkea1d",
   "source_url": "https://omlx.ai/benchmarks/performance/curkea1d",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-curkea1d-ctx4096",
   "model": "Muse Glimmer",
   "notes": "oMLX community benchmark trace; peak memory 15.5 GB at this context; duplicate traces of the same session: szlz46lz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 214.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-30B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 214.6 18.7 15.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/curkea1d",
   "source_url": "https://omlx.ai/benchmarks/performance/curkea1d",
   "tg_tokens": null,
   "tps": 18.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-9sbu061g-ctx1024",
   "model": "Muse Glimmer",
   "notes": "oMLX community benchmark trace; peak memory 15.2 GB at this context; duplicate traces of the same session: yezhbpsx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 254.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-30B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 254.9 28.3 15.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9sbu061g",
   "source_url": "https://omlx.ai/benchmarks/performance/9sbu061g",
   "tg_tokens": null,
   "tps": 28.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-9sbu061g-ctx4096",
   "model": "Muse Glimmer",
   "notes": "oMLX community benchmark trace; peak memory 15.5 GB at this context; duplicate traces of the same session: yezhbpsx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 252.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-30B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 252.5 27.7 15.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9sbu061g",
   "source_url": "https://omlx.ai/benchmarks/performance/9sbu061g",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-cqplcoju-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 5.7 GB at this context; duplicate traces of the same session: h7r52nh2, k1b1u5lh; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 421.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-4B-MLX-8bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 8bit; 4k 421.7 21.1 5.7 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cqplcoju",
   "source_url": "https://omlx.ai/benchmarks/performance/cqplcoju",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-cqplcoju-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: h7r52nh2, k1b1u5lh; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 382.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-4B-MLX-8bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 8bit; 16k 382.1 19.2 6.3 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cqplcoju",
   "source_url": "https://omlx.ai/benchmarks/performance/cqplcoju",
   "tg_tokens": null,
   "tps": 19.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-cqplcoju-ctx32768",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 7.1 GB at this context; duplicate traces of the same session: h7r52nh2, k1b1u5lh; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 330.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-4B-MLX-8bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 8bit; 32k 330.3 16.5 7.1 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cqplcoju",
   "source_url": "https://omlx.ai/benchmarks/performance/cqplcoju",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-v78qwdas-ctx1024",
   "model": "Qwen3.6-35B-A3B-MAYA",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: ofrx20ak; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 930.6,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.6-35B-A3B-MAYA-oQ3.5e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 3bit; 1k 930.6 54.6 19.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v78qwdas",
   "source_url": "https://omlx.ai/benchmarks/performance/v78qwdas",
   "tg_tokens": null,
   "tps": 54.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-v78qwdas-ctx4096",
   "model": "Qwen3.6-35B-A3B-MAYA",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: ofrx20ak; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1060.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.6-35B-A3B-MAYA-oQ3.5e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 3bit; 4k 1,060 72.7 20.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v78qwdas",
   "source_url": "https://omlx.ai/benchmarks/performance/v78qwdas",
   "tg_tokens": null,
   "tps": 72.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ok1vc242-ctx1024",
   "model": "Qwen3.6-40B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 41.0 GB at this context; duplicate traces of the same session: hv2sse10; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 156.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-40B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-Thinking-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 156.5 10.9 41.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ok1vc242",
   "source_url": "https://omlx.ai/benchmarks/performance/ok1vc242",
   "tg_tokens": null,
   "tps": 10.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ok1vc242-ctx4096",
   "model": "Qwen3.6-40B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 45.8 GB at this context; duplicate traces of the same session: hv2sse10; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-40B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-Thinking-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 125.4 8.7 45.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ok1vc242",
   "source_url": "https://omlx.ai/benchmarks/performance/ok1vc242",
   "tg_tokens": null,
   "tps": 8.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-wjymik7n-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 335.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 335.6 39.5 16.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wjymik7n",
   "source_url": "https://omlx.ai/benchmarks/performance/wjymik7n",
   "tg_tokens": null,
   "tps": 39.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-wjymik7n-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 358.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 358.0 38.2 19.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wjymik7n",
   "source_url": "https://omlx.ai/benchmarks/performance/wjymik7n",
   "tg_tokens": null,
   "tps": 38.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-nq0njwj4-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 17.1 GB; TTFT 36993 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 110.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k tokens 110.7 PP tok/s 12.4 TG tok/s 36993 TTFT (ms) 17.1 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nq0njwj4",
   "source_url": "https://omlx.ai/benchmarks/performance/nq0njwj4",
   "tg_tokens": null,
   "tps": 12.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5621z73k-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 33.6 GB; TTFT 84882 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 772.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 64k tokens 772.1 PP tok/s 16.1 TG tok/s 84882 TTFT (ms) 33.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5621z73k",
   "source_url": "https://omlx.ai/benchmarks/performance/5621z73k",
   "tg_tokens": null,
   "tps": 16.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vpuz6s76-ctx1024",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: 27yzs0td, g5xcsmoc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 388.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Alis-MLX-6bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 388.3 44.7 27.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vpuz6s76",
   "source_url": "https://omlx.ai/benchmarks/performance/vpuz6s76",
   "tg_tokens": null,
   "tps": 44.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vpuz6s76-ctx4096",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 31.0 GB at this context; duplicate traces of the same session: 27yzs0td, g5xcsmoc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 413.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Alis-MLX-6bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 413.8 40.2 31.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vpuz6s76",
   "source_url": "https://omlx.ai/benchmarks/performance/vpuz6s76",
   "tg_tokens": null,
   "tps": 40.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vpuz6s76-ctx32768",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 34.1 GB at this context; duplicate traces of the same session: 27yzs0td, g5xcsmoc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 354.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Alis-MLX-6bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 354.4 46.7 34.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vpuz6s76",
   "source_url": "https://omlx.ai/benchmarks/performance/vpuz6s76",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-wz433si3-ctx1024",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 27.5 GB; TTFT 2654 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 385.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Alis-MLX-6bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k tokens 385.8 PP tok/s 41.9 TG tok/s 2654 TTFT (ms) 27.5 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wz433si3",
   "source_url": "https://omlx.ai/benchmarks/performance/wz433si3",
   "tg_tokens": null,
   "tps": 41.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ro7zd5w5-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 873.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 873.5 33.5 16.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ro7zd5w5",
   "source_url": "https://omlx.ai/benchmarks/performance/ro7zd5w5",
   "tg_tokens": null,
   "tps": 33.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ro7zd5w5-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 981.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 981.0 32.9 17.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ro7zd5w5",
   "source_url": "https://omlx.ai/benchmarks/performance/ro7zd5w5",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-0et2neel-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 4iji2xip; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 327.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 327.0 37.1 16.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0et2neel",
   "source_url": "https://omlx.ai/benchmarks/performance/0et2neel",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-0et2neel-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 4iji2xip; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 361.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 361.9 36.2 19.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0et2neel",
   "source_url": "https://omlx.ai/benchmarks/performance/0et2neel",
   "tg_tokens": null,
   "tps": 36.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-933wkolr-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 77.2,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 77.2 11.9 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/933wkolr",
   "source_url": "https://omlx.ai/benchmarks/performance/933wkolr",
   "tg_tokens": null,
   "tps": 11.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-933wkolr-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 85.1,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 85.1 14.8 23.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/933wkolr",
   "source_url": "https://omlx.ai/benchmarks/performance/933wkolr",
   "tg_tokens": null,
   "tps": 14.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-052lxzvf-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 4djc5prv; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 569.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 569.6 34.9 19.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/052lxzvf",
   "source_url": "https://omlx.ai/benchmarks/performance/052lxzvf",
   "tg_tokens": null,
   "tps": 34.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-052lxzvf-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 4djc5prv; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 565.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 565.6 38.8 19.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/052lxzvf",
   "source_url": "https://omlx.ai/benchmarks/performance/052lxzvf",
   "tg_tokens": null,
   "tps": 38.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-appbggqd-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: i4kl2ho6; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 365.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 365.0 18.6 19.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/appbggqd",
   "source_url": "https://omlx.ai/benchmarks/performance/appbggqd",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-appbggqd-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: i4kl2ho6; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 415.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 415.7 18.8 20.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/appbggqd",
   "source_url": "https://omlx.ai/benchmarks/performance/appbggqd",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vjsmevj6-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 72.6 GB; TTFT 601 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1703.0,
   "provenance": "community",
   "quant": "GSQ-RCO-Q2",
   "quote": "Qwen3.8-Flash-Next-GSQ-RCO-Q2-mlx on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,703 PP tok/s 79.3 TG tok/s 601 TTFT (ms) 72.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vjsmevj6",
   "source_url": "https://omlx.ai/benchmarks/performance/vjsmevj6",
   "tg_tokens": null,
   "tps": 79.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-2e8e5n5g-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.5 GB at this context; duplicate traces of the same session: 3xezw7j9, wh6l9eg7; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1470.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 4k 1,470 86.8 69.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2e8e5n5g",
   "source_url": "https://omlx.ai/benchmarks/performance/2e8e5n5g",
   "tg_tokens": null,
   "tps": 86.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-2e8e5n5g-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.4 GB at this context; duplicate traces of the same session: 3xezw7j9, wh6l9eg7; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2355.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 32k 2,355 82.1 71.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2e8e5n5g",
   "source_url": "https://omlx.ai/benchmarks/performance/2e8e5n5g",
   "tg_tokens": null,
   "tps": 82.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-2e8e5n5g-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.8 GB at this context; duplicate traces of the same session: 3xezw7j9, wh6l9eg7; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2517.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 128k 2,517 74.7 74.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2e8e5n5g",
   "source_url": "https://omlx.ai/benchmarks/performance/2e8e5n5g",
   "tg_tokens": null,
   "tps": 74.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-ix6k2x9b-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: 03l531hd; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 5bit; 4k 323.2 51.3 27.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ix6k2x9b",
   "source_url": "https://omlx.ai/benchmarks/performance/ix6k2x9b",
   "tg_tokens": null,
   "tps": 51.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-ix6k2x9b-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: 03l531hd; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 316.6,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 5bit; 8k 316.6 60.9 28.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ix6k2x9b",
   "source_url": "https://omlx.ai/benchmarks/performance/ix6k2x9b",
   "tg_tokens": null,
   "tps": 60.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8kqpq77i-ctx16384",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: 39bbhds9, 6i65w8jr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3182.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 3,182 122.6 21.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8kqpq77i",
   "source_url": "https://omlx.ai/benchmarks/performance/8kqpq77i",
   "tg_tokens": null,
   "tps": 122.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8kqpq77i-ctx32768",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 39bbhds9, 6i65w8jr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4701.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 4,701 114.8 21.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8kqpq77i",
   "source_url": "https://omlx.ai/benchmarks/performance/8kqpq77i",
   "tg_tokens": null,
   "tps": 114.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8kqpq77i-ctx65536",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 39bbhds9, 6i65w8jr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3747.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 3,747 96.6 21.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8kqpq77i",
   "source_url": "https://omlx.ai/benchmarks/performance/8kqpq77i",
   "tg_tokens": null,
   "tps": 96.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-yfeush4r-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: wqnrqicd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1435.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,435 111.9 14.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yfeush4r",
   "source_url": "https://omlx.ai/benchmarks/performance/yfeush4r",
   "tg_tokens": null,
   "tps": 111.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-yfeush4r-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: wqnrqicd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1390.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,390 107.4 14.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yfeush4r",
   "source_url": "https://omlx.ai/benchmarks/performance/yfeush4r",
   "tg_tokens": null,
   "tps": 107.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-u6e9taj0-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3387.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 3,387 115.0 14.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u6e9taj0",
   "source_url": "https://omlx.ai/benchmarks/performance/u6e9taj0",
   "tg_tokens": null,
   "tps": 115.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-u6e9taj0-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4221.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 4,221 118.4 14.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u6e9taj0",
   "source_url": "https://omlx.ai/benchmarks/performance/u6e9taj0",
   "tg_tokens": null,
   "tps": 118.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-h7h9lnqk-ctx4096",
   "model": "DeepSeek-V4-Flash",
   "notes": "oMLX community benchmark trace; peak memory 142.3 GB at this context; duplicate traces of the same session: llsogv7r; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 525.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-V4-Flash on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 525.2 28.9 142.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h7h9lnqk",
   "source_url": "https://omlx.ai/benchmarks/performance/h7h9lnqk",
   "tg_tokens": null,
   "tps": 28.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-h7h9lnqk-ctx16384",
   "model": "DeepSeek-V4-Flash",
   "notes": "oMLX community benchmark trace; peak memory 142.7 GB at this context; duplicate traces of the same session: llsogv7r; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 502.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-V4-Flash on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 502.2 28.0 142.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h7h9lnqk",
   "source_url": "https://omlx.ai/benchmarks/performance/h7h9lnqk",
   "tg_tokens": null,
   "tps": 28.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-am7jnewu-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 829.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ6e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 829.1 56.6 26.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/am7jnewu",
   "source_url": "https://omlx.ai/benchmarks/performance/am7jnewu",
   "tg_tokens": null,
   "tps": 56.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-am7jnewu-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 918.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ6e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 918.3 47.5 27.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/am7jnewu",
   "source_url": "https://omlx.ai/benchmarks/performance/am7jnewu",
   "tg_tokens": null,
   "tps": 47.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4b5mjz32-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 858.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ6e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 858.0 24.5 22.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4b5mjz32",
   "source_url": "https://omlx.ai/benchmarks/performance/4b5mjz32",
   "tg_tokens": null,
   "tps": 24.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4b5mjz32-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 924.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ6e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 924.5 24.2 23.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4b5mjz32",
   "source_url": "https://omlx.ai/benchmarks/performance/4b5mjz32",
   "tg_tokens": null,
   "tps": 24.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-zmxn61rz-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: 94j1nyrd, aba2gd91; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 833.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ6e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 833.2 57.3 26.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zmxn61rz",
   "source_url": "https://omlx.ai/benchmarks/performance/zmxn61rz",
   "tg_tokens": null,
   "tps": 57.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zmxn61rz-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: 94j1nyrd, aba2gd91; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 910.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ6e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 910.9 45.7 27.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zmxn61rz",
   "source_url": "https://omlx.ai/benchmarks/performance/zmxn61rz",
   "tg_tokens": null,
   "tps": 45.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-zmxn61rz-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace; peak memory 31.9 GB at this context; duplicate traces of the same session: 94j1nyrd, aba2gd91; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 625.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ6e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k 625.5 45.1 31.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zmxn61rz",
   "source_url": "https://omlx.ai/benchmarks/performance/zmxn61rz",
   "tg_tokens": null,
   "tps": 45.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-blh1gbx9-ctx4096",
   "model": "Huihui-Ornith-1.5-35B-A3B-abliterated",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.9 GB at this context; duplicate traces of the same session: 25pbflsj, euti486g, gmc3ep10, oyqyuhuj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1015.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Huihui-Ornith-1.5-35B-A3B-abliterated-oQ5e-fp16-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 5bit; 4k 1,015 69.2 28.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/blh1gbx9",
   "source_url": "https://omlx.ai/benchmarks/performance/blh1gbx9",
   "tg_tokens": null,
   "tps": 69.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-blh1gbx9-ctx16384",
   "model": "Huihui-Ornith-1.5-35B-A3B-abliterated",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.8 GB at this context; duplicate traces of the same session: 25pbflsj, euti486g, gmc3ep10, oyqyuhuj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 845.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Huihui-Ornith-1.5-35B-A3B-abliterated-oQ5e-fp16-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 5bit; 16k 845.1 62.8 29.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/blh1gbx9",
   "source_url": "https://omlx.ai/benchmarks/performance/blh1gbx9",
   "tg_tokens": null,
   "tps": 62.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-blh1gbx9-ctx32768",
   "model": "Huihui-Ornith-1.5-35B-A3B-abliterated",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.1 GB at this context; duplicate traces of the same session: 25pbflsj, euti486g, gmc3ep10, oyqyuhuj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 639.3,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Huihui-Ornith-1.5-35B-A3B-abliterated-oQ5e-fp16-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 5bit; 32k 639.3 68.7 31.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/blh1gbx9",
   "source_url": "https://omlx.ai/benchmarks/performance/blh1gbx9",
   "tg_tokens": null,
   "tps": 68.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-blh1gbx9-ctx65536",
   "model": "Huihui-Ornith-1.5-35B-A3B-abliterated",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.0 GB at this context; duplicate traces of the same session: 25pbflsj, euti486g, gmc3ep10, oyqyuhuj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 469.6,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Huihui-Ornith-1.5-35B-A3B-abliterated-oQ5e-fp16-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 5bit; 64k 469.6 53.1 33.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/blh1gbx9",
   "source_url": "https://omlx.ai/benchmarks/performance/blh1gbx9",
   "tg_tokens": null,
   "tps": 53.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-blh1gbx9-ctx131072",
   "model": "Huihui-Ornith-1.5-35B-A3B-abliterated",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: 25pbflsj, euti486g, gmc3ep10, oyqyuhuj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 315.6,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Huihui-Ornith-1.5-35B-A3B-abliterated-oQ5e-fp16-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 5bit; 128k 315.6 36.7 35.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/blh1gbx9",
   "source_url": "https://omlx.ai/benchmarks/performance/blh1gbx9",
   "tg_tokens": null,
   "tps": 36.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-kv59kf39-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: lmaxo7b9, qunb7dzf; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 743.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 743.3 69.5 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kv59kf39",
   "source_url": "https://omlx.ai/benchmarks/performance/kv59kf39",
   "tg_tokens": null,
   "tps": 69.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-kv59kf39-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: lmaxo7b9, qunb7dzf; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 785.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 785.8 66.7 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kv59kf39",
   "source_url": "https://omlx.ai/benchmarks/performance/kv59kf39",
   "tg_tokens": null,
   "tps": 66.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-kv59kf39-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: lmaxo7b9, qunb7dzf; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 521.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 32k 521.3 48.9 22.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kv59kf39",
   "source_url": "https://omlx.ai/benchmarks/performance/kv59kf39",
   "tg_tokens": null,
   "tps": 48.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-hdzdvng3-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: o5phscxh; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 375.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 375.4 46.0 6.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hdzdvng3",
   "source_url": "https://omlx.ai/benchmarks/performance/hdzdvng3",
   "tg_tokens": null,
   "tps": 46.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-hdzdvng3-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.2 GB at this context; duplicate traces of the same session: o5phscxh; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 359.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 359.8 40.4 7.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hdzdvng3",
   "source_url": "https://omlx.ai/benchmarks/performance/hdzdvng3",
   "tg_tokens": null,
   "tps": 40.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-n2eya3g6-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 335.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 335.6 39.5 16.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n2eya3g6",
   "source_url": "https://omlx.ai/benchmarks/performance/n2eya3g6",
   "tg_tokens": null,
   "tps": 39.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-n2eya3g6-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 358.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 358.0 38.2 19.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n2eya3g6",
   "source_url": "https://omlx.ai/benchmarks/performance/n2eya3g6",
   "tg_tokens": null,
   "tps": 38.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hzsvr953-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: xmtlpq4z; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 335.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 335.5 39.5 16.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hzsvr953",
   "source_url": "https://omlx.ai/benchmarks/performance/hzsvr953",
   "tg_tokens": null,
   "tps": 39.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hzsvr953-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: xmtlpq4z; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 355.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 355.5 38.2 19.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hzsvr953",
   "source_url": "https://omlx.ai/benchmarks/performance/hzsvr953",
   "tg_tokens": null,
   "tps": 38.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wccxirz4-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: 8ce9np2t, pvumr59r, q7akk9jk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 859.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 859.0 44.8 17.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wccxirz4",
   "source_url": "https://omlx.ai/benchmarks/performance/wccxirz4",
   "tg_tokens": null,
   "tps": 44.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wccxirz4-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: 8ce9np2t, pvumr59r, q7akk9jk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 844.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 844.5 42.5 19.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wccxirz4",
   "source_url": "https://omlx.ai/benchmarks/performance/wccxirz4",
   "tg_tokens": null,
   "tps": 42.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wccxirz4-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 8ce9np2t, pvumr59r, q7akk9jk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 743.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 743.6 18.4 21.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wccxirz4",
   "source_url": "https://omlx.ai/benchmarks/performance/wccxirz4",
   "tg_tokens": null,
   "tps": 18.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wccxirz4-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: 8ce9np2t, pvumr59r, q7akk9jk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 439.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 439.5 3.8 29.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wccxirz4",
   "source_url": "https://omlx.ai/benchmarks/performance/wccxirz4",
   "tg_tokens": null,
   "tps": 3.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-065h3gur-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.3 GB at this context; duplicate traces of the same session: 2pxt4u8x; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 125.0 8.0 28.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/065h3gur",
   "source_url": "https://omlx.ai/benchmarks/performance/065h3gur",
   "tg_tokens": null,
   "tps": 8.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-065h3gur-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.8 GB at this context; duplicate traces of the same session: 2pxt4u8x; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 109.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 109.8 8.2 31.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/065h3gur",
   "source_url": "https://omlx.ai/benchmarks/performance/065h3gur",
   "tg_tokens": null,
   "tps": 8.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2b5yywx2-ctx1024",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: fc11k1cf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 946.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 946.5 31.7 16.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2b5yywx2",
   "source_url": "https://omlx.ai/benchmarks/performance/2b5yywx2",
   "tg_tokens": null,
   "tps": 31.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2b5yywx2-ctx4096",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: fc11k1cf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 984.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 984.4 31.6 17.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2b5yywx2",
   "source_url": "https://omlx.ai/benchmarks/performance/2b5yywx2",
   "tg_tokens": null,
   "tps": 31.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n74y40aw-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 873.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 873.5 33.5 16.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n74y40aw",
   "source_url": "https://omlx.ai/benchmarks/performance/n74y40aw",
   "tg_tokens": null,
   "tps": 33.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n74y40aw-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 981.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 981.0 32.9 17.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n74y40aw",
   "source_url": "https://omlx.ai/benchmarks/performance/n74y40aw",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kc4cbzhf-ctx32768",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: bh6nnmbl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4771.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 4,771 114.9 21.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kc4cbzhf",
   "source_url": "https://omlx.ai/benchmarks/performance/kc4cbzhf",
   "tg_tokens": null,
   "tps": 114.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kc4cbzhf-ctx65536",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: bh6nnmbl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3795.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 3,795 96.3 21.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kc4cbzhf",
   "source_url": "https://omlx.ai/benchmarks/performance/kc4cbzhf",
   "tg_tokens": null,
   "tps": 96.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-z9yd2foi-ctx65536",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: pia78a04; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3462.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 3,462 133.5 23.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z9yd2foi",
   "source_url": "https://omlx.ai/benchmarks/performance/z9yd2foi",
   "tg_tokens": null,
   "tps": 133.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-z9yd2foi-ctx131072",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: pia78a04; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2259.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 2,259 104.8 24.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z9yd2foi",
   "source_url": "https://omlx.ai/benchmarks/performance/z9yd2foi",
   "tg_tokens": null,
   "tps": 104.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4x8l12cj-ctx32768",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 2n71t68j, f1fgitfk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4423.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 4,423 154.8 22.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4x8l12cj",
   "source_url": "https://omlx.ai/benchmarks/performance/4x8l12cj",
   "tg_tokens": null,
   "tps": 154.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4x8l12cj-ctx65536",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 2n71t68j, f1fgitfk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3404.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 3,404 124.7 23.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4x8l12cj",
   "source_url": "https://omlx.ai/benchmarks/performance/4x8l12cj",
   "tg_tokens": null,
   "tps": 124.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4x8l12cj-ctx131072",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: 2n71t68j, f1fgitfk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2245.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 2,245 107.4 24.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4x8l12cj",
   "source_url": "https://omlx.ai/benchmarks/performance/4x8l12cj",
   "tg_tokens": null,
   "tps": 107.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cwphw8hu-ctx32768",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 1evvae5v; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4548.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 4,548 159.9 22.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cwphw8hu",
   "source_url": "https://omlx.ai/benchmarks/performance/cwphw8hu",
   "tg_tokens": null,
   "tps": 159.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-cwphw8hu-ctx65536",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 1evvae5v; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3534.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 3,534 130.1 23.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cwphw8hu",
   "source_url": "https://omlx.ai/benchmarks/performance/cwphw8hu",
   "tg_tokens": null,
   "tps": 130.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-swhcr302-ctx32768",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: lu6cbjjl, qdwqo0z7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2200.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 2,200 80.4 23.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/swhcr302",
   "source_url": "https://omlx.ai/benchmarks/performance/swhcr302",
   "tg_tokens": null,
   "tps": 80.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-swhcr302-ctx65536",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: lu6cbjjl, qdwqo0z7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1557.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 1,557 64.1 24.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/swhcr302",
   "source_url": "https://omlx.ai/benchmarks/performance/swhcr302",
   "tg_tokens": null,
   "tps": 64.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-swhcr302-ctx131072",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: lu6cbjjl, qdwqo0z7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 936.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 128k 936.2 50.9 26.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/swhcr302",
   "source_url": "https://omlx.ai/benchmarks/performance/swhcr302",
   "tg_tokens": null,
   "tps": 50.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-fyld6w7w-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: 7801eibv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1430.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,430 111.8 14.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fyld6w7w",
   "source_url": "https://omlx.ai/benchmarks/performance/fyld6w7w",
   "tg_tokens": null,
   "tps": 111.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-fyld6w7w-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: 7801eibv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1457.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,457 107.4 14.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fyld6w7w",
   "source_url": "https://omlx.ai/benchmarks/performance/fyld6w7w",
   "tg_tokens": null,
   "tps": 107.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-5e9430d1-ctx1024",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.7 GB at this context; duplicate traces of the same session: j18544f2; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 318.4,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-OptiQ-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 318.4 45.5 7.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5e9430d1",
   "source_url": "https://omlx.ai/benchmarks/performance/5e9430d1",
   "tg_tokens": null,
   "tps": 45.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-5e9430d1-ctx4096",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.3 GB at this context; duplicate traces of the same session: j18544f2; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.5,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-OptiQ-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 323.5 43.9 8.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5e9430d1",
   "source_url": "https://omlx.ai/benchmarks/performance/5e9430d1",
   "tg_tokens": null,
   "tps": 43.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-e90vwh4s-ctx1024",
   "model": "Nemotron-Cascade-2-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 32.7 GB at this context; duplicate traces of the same session: 23fwtjht, 9fr7dug4, n3rfc8lw, u1p8czqf, zd02mlfq; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 786.7,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Nemotron-Cascade-2-30B-A3B-mxfp8 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 786.7 71.4 32.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e90vwh4s",
   "source_url": "https://omlx.ai/benchmarks/performance/e90vwh4s",
   "tg_tokens": null,
   "tps": 71.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-e90vwh4s-ctx4096",
   "model": "Nemotron-Cascade-2-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 33.6 GB at this context; duplicate traces of the same session: 23fwtjht, 9fr7dug4, n3rfc8lw, u1p8czqf, zd02mlfq; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 691.7,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Nemotron-Cascade-2-30B-A3B-mxfp8 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 691.7 64.1 33.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e90vwh4s",
   "source_url": "https://omlx.ai/benchmarks/performance/e90vwh4s",
   "tg_tokens": null,
   "tps": 64.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-e90vwh4s-ctx8192",
   "model": "Nemotron-Cascade-2-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 33.6 GB at this context; duplicate traces of the same session: 23fwtjht, 9fr7dug4, n3rfc8lw, u1p8czqf, zd02mlfq; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 657.2,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Nemotron-Cascade-2-30B-A3B-mxfp8 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 8k 657.2 57.2 33.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e90vwh4s",
   "source_url": "https://omlx.ai/benchmarks/performance/e90vwh4s",
   "tg_tokens": null,
   "tps": 57.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-e90vwh4s-ctx16384",
   "model": "Nemotron-Cascade-2-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 33.7 GB at this context; duplicate traces of the same session: 23fwtjht, 9fr7dug4, n3rfc8lw, u1p8czqf, zd02mlfq; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 629.1,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Nemotron-Cascade-2-30B-A3B-mxfp8 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 16k 629.1 59.1 33.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e90vwh4s",
   "source_url": "https://omlx.ai/benchmarks/performance/e90vwh4s",
   "tg_tokens": null,
   "tps": 59.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-e90vwh4s-ctx32768",
   "model": "Nemotron-Cascade-2-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: 23fwtjht, 9fr7dug4, n3rfc8lw, u1p8czqf, zd02mlfq; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 568.8,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Nemotron-Cascade-2-30B-A3B-mxfp8 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 32k 568.8 43.1 33.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e90vwh4s",
   "source_url": "https://omlx.ai/benchmarks/performance/e90vwh4s",
   "tg_tokens": null,
   "tps": 43.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-e90vwh4s-ctx65536",
   "model": "Nemotron-Cascade-2-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 34.0 GB at this context; duplicate traces of the same session: 23fwtjht, 9fr7dug4, n3rfc8lw, u1p8czqf, zd02mlfq; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 488.2,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Nemotron-Cascade-2-30B-A3B-mxfp8 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 64k 488.2 46.0 34.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e90vwh4s",
   "source_url": "https://omlx.ai/benchmarks/performance/e90vwh4s",
   "tg_tokens": null,
   "tps": 46.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-pq59fnw5-ctx4096",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.6 GB at this context; duplicate traces of the same session: 57a062py, eya9jvme, jo19ghxp, w5pz0urn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 987.1,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-oQ8-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 987.1 73.7 39.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pq59fnw5",
   "source_url": "https://omlx.ai/benchmarks/performance/pq59fnw5",
   "tg_tokens": null,
   "tps": 73.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-pq59fnw5-ctx16384",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 40.6 GB at this context; duplicate traces of the same session: 57a062py, eya9jvme, jo19ghxp, w5pz0urn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 799.4,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-oQ8-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 16k 799.4 53.0 40.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pq59fnw5",
   "source_url": "https://omlx.ai/benchmarks/performance/pq59fnw5",
   "tg_tokens": null,
   "tps": 53.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-pq59fnw5-ctx32768",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 42.0 GB at this context; duplicate traces of the same session: 57a062py, eya9jvme, jo19ghxp, w5pz0urn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 646.8,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-oQ8-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 32k 646.8 48.8 42.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pq59fnw5",
   "source_url": "https://omlx.ai/benchmarks/performance/pq59fnw5",
   "tg_tokens": null,
   "tps": 48.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-pq59fnw5-ctx65536",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 44.0 GB at this context; duplicate traces of the same session: 57a062py, eya9jvme, jo19ghxp, w5pz0urn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 460.5,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-oQ8-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 64k 460.5 25.7 44.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pq59fnw5",
   "source_url": "https://omlx.ai/benchmarks/performance/pq59fnw5",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-pq59fnw5-ctx131072",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 46.8 GB at this context; duplicate traces of the same session: 57a062py, eya9jvme, jo19ghxp, w5pz0urn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 295.5,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-oQ8-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 128k 295.5 23.6 46.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pq59fnw5",
   "source_url": "https://omlx.ai/benchmarks/performance/pq59fnw5",
   "tg_tokens": null,
   "tps": 23.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-7rwp7787-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: 850lt1wo, lwgdh2sg, rez1s2qw, wllrr4y5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 995.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 995.0 86.3 32.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7rwp7787",
   "source_url": "https://omlx.ai/benchmarks/performance/7rwp7787",
   "tg_tokens": null,
   "tps": 86.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-7rwp7787-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.9 GB at this context; duplicate traces of the same session: 850lt1wo, lwgdh2sg, rez1s2qw, wllrr4y5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 814.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 16k 814.9 64.8 32.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7rwp7787",
   "source_url": "https://omlx.ai/benchmarks/performance/7rwp7787",
   "tg_tokens": null,
   "tps": 64.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-7rwp7787-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.2 GB at this context; duplicate traces of the same session: 850lt1wo, lwgdh2sg, rez1s2qw, wllrr4y5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 654.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 32k 654.6 61.6 34.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7rwp7787",
   "source_url": "https://omlx.ai/benchmarks/performance/7rwp7787",
   "tg_tokens": null,
   "tps": 61.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-7rwp7787-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: 850lt1wo, lwgdh2sg, rez1s2qw, wllrr4y5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 497.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 64k 497.9 51.6 36.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7rwp7787",
   "source_url": "https://omlx.ai/benchmarks/performance/7rwp7787",
   "tg_tokens": null,
   "tps": 51.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-7rwp7787-ctx131072",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.6 GB at this context; duplicate traces of the same session: 850lt1wo, lwgdh2sg, rez1s2qw, wllrr4y5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 331.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 128k 331.7 37.8 38.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7rwp7787",
   "source_url": "https://omlx.ai/benchmarks/performance/7rwp7787",
   "tg_tokens": null,
   "tps": 37.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-sinjsoi9-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: w3bqaoq2; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 759.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 1k 759.9 56.9 35.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sinjsoi9",
   "source_url": "https://omlx.ai/benchmarks/performance/sinjsoi9",
   "tg_tokens": null,
   "tps": 56.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-sinjsoi9-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.5 GB at this context; duplicate traces of the same session: w3bqaoq2; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 555.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k 555.0 54.7 37.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sinjsoi9",
   "source_url": "https://omlx.ai/benchmarks/performance/sinjsoi9",
   "tg_tokens": null,
   "tps": 54.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-sinjsoi9-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.2 GB at this context; duplicate traces of the same session: w3bqaoq2; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 450.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 32k 450.4 45.6 38.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sinjsoi9",
   "source_url": "https://omlx.ai/benchmarks/performance/sinjsoi9",
   "tg_tokens": null,
   "tps": 45.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-550csd7j-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: l3ngwnqd, xqsaiixj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 740.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 740.9 66.5 20.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/550csd7j",
   "source_url": "https://omlx.ai/benchmarks/performance/550csd7j",
   "tg_tokens": null,
   "tps": 66.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-550csd7j-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: l3ngwnqd, xqsaiixj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 773.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 773.6 64.8 22.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/550csd7j",
   "source_url": "https://omlx.ai/benchmarks/performance/550csd7j",
   "tg_tokens": null,
   "tps": 64.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-550csd7j-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: l3ngwnqd, xqsaiixj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 740.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 8k 740.0 62.0 23.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/550csd7j",
   "source_url": "https://omlx.ai/benchmarks/performance/550csd7j",
   "tg_tokens": null,
   "tps": 62.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1ud7mhs9-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 33.4 GB; TTFT 18668 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1755.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 32k tokens 1,755 PP tok/s 139.9 TG tok/s 18668 TTFT (ms) 33.4 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ud7mhs9",
   "source_url": "https://omlx.ai/benchmarks/performance/1ud7mhs9",
   "tg_tokens": null,
   "tps": 139.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1y57h1hv-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 33.4 GB; TTFT 18656 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1756.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 32k tokens 1,756 PP tok/s 143.2 TG tok/s 18656 TTFT (ms) 33.4 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1y57h1hv",
   "source_url": "https://omlx.ai/benchmarks/performance/1y57h1hv",
   "tg_tokens": null,
   "tps": 143.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-aceur6pf-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 33.5 GB; TTFT 19225 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1705.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 32k tokens 1,705 PP tok/s 141.0 TG tok/s 19225 TTFT (ms) 33.5 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aceur6pf",
   "source_url": "https://omlx.ai/benchmarks/performance/aceur6pf",
   "tg_tokens": null,
   "tps": 141.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1623q51u-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 33.5 GB; TTFT 19605 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1671.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 32k tokens 1,671 PP tok/s 139.9 TG tok/s 19605 TTFT (ms) 33.5 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1623q51u",
   "source_url": "https://omlx.ai/benchmarks/performance/1623q51u",
   "tg_tokens": null,
   "tps": 139.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9zjhlcxi-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MLX Serve build; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: 04jbqibl, wedx87yf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 973.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-Serve-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 973.7 45.5 22.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9zjhlcxi",
   "source_url": "https://omlx.ai/benchmarks/performance/9zjhlcxi",
   "tg_tokens": null,
   "tps": 45.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9zjhlcxi-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MLX Serve build; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: 04jbqibl, wedx87yf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 855.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-Serve-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 855.4 18.3 24.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9zjhlcxi",
   "source_url": "https://omlx.ai/benchmarks/performance/9zjhlcxi",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9zjhlcxi-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MLX Serve build; oMLX community benchmark trace; peak memory 29.1 GB at this context; duplicate traces of the same session: 04jbqibl, wedx87yf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 653.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-Serve-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 653.8 3.2 29.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9zjhlcxi",
   "source_url": "https://omlx.ai/benchmarks/performance/9zjhlcxi",
   "tg_tokens": null,
   "tps": 3.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-clyow6l7-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MLX Serve build; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 344v4hoe, tue9wgr2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 981.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-Serve-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 981.6 32.5 18.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/clyow6l7",
   "source_url": "https://omlx.ai/benchmarks/performance/clyow6l7",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-clyow6l7-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MLX Serve build; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: 344v4hoe, tue9wgr2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 818.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-Serve-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 818.5 31.0 19.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/clyow6l7",
   "source_url": "https://omlx.ai/benchmarks/performance/clyow6l7",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-clyow6l7-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MLX Serve build; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: 344v4hoe, tue9wgr2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 626.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-Serve-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 626.6 25.0 22.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/clyow6l7",
   "source_url": "https://omlx.ai/benchmarks/performance/clyow6l7",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-v7w6f2jk-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: 9uw8zxs0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 321.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 321.8 40.5 16.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v7w6f2jk",
   "source_url": "https://omlx.ai/benchmarks/performance/v7w6f2jk",
   "tg_tokens": null,
   "tps": 40.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-v7w6f2jk-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: 9uw8zxs0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 352.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 352.3 44.9 20.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v7w6f2jk",
   "source_url": "https://omlx.ai/benchmarks/performance/v7w6f2jk",
   "tg_tokens": null,
   "tps": 44.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-h0y0ho3b-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: s56jann1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 330.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 330.3 36.6 16.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h0y0ho3b",
   "source_url": "https://omlx.ai/benchmarks/performance/h0y0ho3b",
   "tg_tokens": null,
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-h0y0ho3b-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: s56jann1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 361.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 361.9 35.7 19.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h0y0ho3b",
   "source_url": "https://omlx.ai/benchmarks/performance/h0y0ho3b",
   "tg_tokens": null,
   "tps": 35.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-enx7rdpn-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 329.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 329.2 37.1 16.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/enx7rdpn",
   "source_url": "https://omlx.ai/benchmarks/performance/enx7rdpn",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-enx7rdpn-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 360.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 360.7 36.3 19.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/enx7rdpn",
   "source_url": "https://omlx.ai/benchmarks/performance/enx7rdpn",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7pt9du7d-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.9 GB at this context; duplicate traces of the same session: jw2vb5xc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 416.1,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Qwen3.8-27B-mxfp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 416.1 11.8 28.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7pt9du7d",
   "source_url": "https://omlx.ai/benchmarks/performance/7pt9du7d",
   "tg_tokens": null,
   "tps": 11.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7pt9du7d-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: jw2vb5xc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 500.5,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Qwen3.8-27B-mxfp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 500.5 13.0 29.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7pt9du7d",
   "source_url": "https://omlx.ai/benchmarks/performance/7pt9du7d",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-36rja9a4-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 27.9 GB; TTFT 329132 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 398.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 128k tokens 398.2 PP tok/s 27.0 TG tok/s 329132 TTFT (ms) 27.9 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/36rja9a4",
   "source_url": "https://omlx.ai/benchmarks/performance/36rja9a4",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-83cv3nm9-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 27.9 GB; TTFT 330598 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 396.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 128k tokens 396.5 PP tok/s 20.1 TG tok/s 330598 TTFT (ms) 27.9 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/83cv3nm9",
   "source_url": "https://omlx.ai/benchmarks/performance/83cv3nm9",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-gbaqmpp0-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 4.7 GB at this context; duplicate traces of the same session: 59uyknk6, jpoi3fc1, wtypvme5; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 726.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 1k 726.8 57.5 4.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gbaqmpp0",
   "source_url": "https://omlx.ai/benchmarks/performance/gbaqmpp0",
   "tg_tokens": null,
   "tps": 57.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-gbaqmpp0-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 4.7 GB at this context; duplicate traces of the same session: 59uyknk6, jpoi3fc1, wtypvme5; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 809.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k 809.7 54.2 4.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gbaqmpp0",
   "source_url": "https://omlx.ai/benchmarks/performance/gbaqmpp0",
   "tg_tokens": null,
   "tps": 54.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-gbaqmpp0-ctx8192",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 4.8 GB at this context; duplicate traces of the same session: 59uyknk6, jpoi3fc1, wtypvme5; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 833.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 8k 833.6 50.4 4.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gbaqmpp0",
   "source_url": "https://omlx.ai/benchmarks/performance/gbaqmpp0",
   "tg_tokens": null,
   "tps": 50.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-gbaqmpp0-ctx16384",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: 59uyknk6, jpoi3fc1, wtypvme5; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 809.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 16k 809.9 45.1 4.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gbaqmpp0",
   "source_url": "https://omlx.ai/benchmarks/performance/gbaqmpp0",
   "tg_tokens": null,
   "tps": 45.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ukgq3cw0-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 25.5 GB; TTFT 16461 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 995.3,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k tokens 995.3 PP tok/s 75.3 TG tok/s 16461 TTFT (ms) 25.5 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ukgq3cw0",
   "source_url": "https://omlx.ai/benchmarks/performance/ukgq3cw0",
   "tg_tokens": null,
   "tps": 75.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ph4f3g52-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 2ouvdnyc; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 874.3,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 874.3 49.7 22.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ph4f3g52",
   "source_url": "https://omlx.ai/benchmarks/performance/ph4f3g52",
   "tg_tokens": null,
   "tps": 49.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ph4f3g52-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: 2ouvdnyc; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1411.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 1,411 71.9 24.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ph4f3g52",
   "source_url": "https://omlx.ai/benchmarks/performance/ph4f3g52",
   "tg_tokens": null,
   "tps": 71.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-75mjyuh6-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 759.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 1k 759.9 56.9 35.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/75mjyuh6",
   "source_url": "https://omlx.ai/benchmarks/performance/75mjyuh6",
   "tg_tokens": null,
   "tps": 56.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-75mjyuh6-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.5 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 555.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k 555.0 54.7 37.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/75mjyuh6",
   "source_url": "https://omlx.ai/benchmarks/performance/75mjyuh6",
   "tg_tokens": null,
   "tps": 54.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-75mjyuh6-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.2 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 450.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 32k 450.4 45.6 38.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/75mjyuh6",
   "source_url": "https://omlx.ai/benchmarks/performance/75mjyuh6",
   "tg_tokens": null,
   "tps": 45.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-126kie6v-ctx1024",
   "model": "Qwen2.5-Coder-14B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 8.5 GB at this context; duplicate traces of the same session: xji2hdgn; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 129.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-14B-Instruct-MLX-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 4bit; 1k 129.1 19.5 8.5 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/126kie6v",
   "source_url": "https://omlx.ai/benchmarks/performance/126kie6v",
   "tg_tokens": null,
   "tps": 19.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-126kie6v-ctx4096",
   "model": "Qwen2.5-Coder-14B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 9.1 GB at this context; duplicate traces of the same session: xji2hdgn; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 123.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-14B-Instruct-MLX-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 4bit; 4k 123.4 17.6 9.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/126kie6v",
   "source_url": "https://omlx.ai/benchmarks/performance/126kie6v",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-bjnggm7r-ctx1024",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: kgm3ea8k; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 84.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 84.1 15.5 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bjnggm7r",
   "source_url": "https://omlx.ai/benchmarks/performance/bjnggm7r",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-bjnggm7r-ctx4096",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: kgm3ea8k; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 77.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 77.0 15.1 20.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bjnggm7r",
   "source_url": "https://omlx.ai/benchmarks/performance/bjnggm7r",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-3equi6j2-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: 86vllq1j, ni7a054f; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 715.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 715.0 69.7 20.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3equi6j2",
   "source_url": "https://omlx.ai/benchmarks/performance/3equi6j2",
   "tg_tokens": null,
   "tps": 69.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-3equi6j2-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: 86vllq1j, ni7a054f; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 827.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 827.3 67.5 22.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3equi6j2",
   "source_url": "https://omlx.ai/benchmarks/performance/3equi6j2",
   "tg_tokens": null,
   "tps": 67.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-3equi6j2-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: 86vllq1j, ni7a054f; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 788.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 8k 788.6 65.0 23.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3equi6j2",
   "source_url": "https://omlx.ai/benchmarks/performance/3equi6j2",
   "tg_tokens": null,
   "tps": 65.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-ajwr2apo-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 0b66m8dz; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 83.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 83.9 15.8 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ajwr2apo",
   "source_url": "https://omlx.ai/benchmarks/performance/ajwr2apo",
   "tg_tokens": null,
   "tps": 15.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ajwr2apo-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 0b66m8dz; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 79.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 79.4 15.4 20.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ajwr2apo",
   "source_url": "https://omlx.ai/benchmarks/performance/ajwr2apo",
   "tg_tokens": null,
   "tps": 15.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xrh3whas-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: jxmje30t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 665.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 665.0 36.6 29.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xrh3whas",
   "source_url": "https://omlx.ai/benchmarks/performance/xrh3whas",
   "tg_tokens": null,
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xrh3whas-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.4 GB at this context; duplicate traces of the same session: jxmje30t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 570.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 570.8 37.3 31.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xrh3whas",
   "source_url": "https://omlx.ai/benchmarks/performance/xrh3whas",
   "tg_tokens": null,
   "tps": 37.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ilrgg1gi-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MLX Serve build; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: fcifch6j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 681.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-Serve-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 681.8 32.6 21.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ilrgg1gi",
   "source_url": "https://omlx.ai/benchmarks/performance/ilrgg1gi",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ilrgg1gi-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MLX Serve build; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: fcifch6j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 595.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-Serve-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 595.4 29.4 23.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ilrgg1gi",
   "source_url": "https://omlx.ai/benchmarks/performance/ilrgg1gi",
   "tg_tokens": null,
   "tps": 29.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-s100cdk5-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 329.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 329.2 37.1 16.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s100cdk5",
   "source_url": "https://omlx.ai/benchmarks/performance/s100cdk5",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-s100cdk5-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 360.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 360.7 36.3 19.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s100cdk5",
   "source_url": "https://omlx.ai/benchmarks/performance/s100cdk5",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-5v983qwe-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 27.8 GB; TTFT 88384 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 185.4,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k tokens 185.4 PP tok/s 22.7 TG tok/s 88384 TTFT (ms) 27.8 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5v983qwe",
   "source_url": "https://omlx.ai/benchmarks/performance/5v983qwe",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ponynl7o-ctx1024",
   "model": "Qwythos-9B-v3-Claude-Mythos-5-1M",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 12.2 GB at this context; duplicate traces of the same session: aqkattr4; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 414.8,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwythos-9B-v3-Claude-Mythos-5-1M-MLX-oQ8-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 1k 414.8 46.0 12.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ponynl7o",
   "source_url": "https://omlx.ai/benchmarks/performance/ponynl7o",
   "tg_tokens": null,
   "tps": 46.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ponynl7o-ctx4096",
   "model": "Qwythos-9B-v3-Claude-Mythos-5-1M",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 12.9 GB at this context; duplicate traces of the same session: aqkattr4; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 449.5,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwythos-9B-v3-Claude-Mythos-5-1M-MLX-oQ8-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 4k 449.5 46.6 12.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ponynl7o",
   "source_url": "https://omlx.ai/benchmarks/performance/ponynl7o",
   "tg_tokens": null,
   "tps": 46.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-o4mct281-ctx4096",
   "model": "Occamy-1.0-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 22o3zkb3, c4s7gx8c, ckqfs6x3, h5sj0sbs; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1026.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "occamy-1.0-abliterated-mtp-mlx-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,026 96.4 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o4mct281",
   "source_url": "https://omlx.ai/benchmarks/performance/o4mct281",
   "tg_tokens": null,
   "tps": 96.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-o4mct281-ctx16384",
   "model": "Occamy-1.0-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: 22o3zkb3, c4s7gx8c, ckqfs6x3, h5sj0sbs; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 861.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "occamy-1.0-abliterated-mtp-mlx-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 861.9 81.2 24.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o4mct281",
   "source_url": "https://omlx.ai/benchmarks/performance/o4mct281",
   "tg_tokens": null,
   "tps": 81.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-o4mct281-ctx32768",
   "model": "Occamy-1.0-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: 22o3zkb3, c4s7gx8c, ckqfs6x3, h5sj0sbs; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 701.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "occamy-1.0-abliterated-mtp-mlx-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 701.6 67.5 25.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o4mct281",
   "source_url": "https://omlx.ai/benchmarks/performance/o4mct281",
   "tg_tokens": null,
   "tps": 67.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-o4mct281-ctx65536",
   "model": "Occamy-1.0-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.6 GB at this context; duplicate traces of the same session: 22o3zkb3, c4s7gx8c, ckqfs6x3, h5sj0sbs; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 511.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "occamy-1.0-abliterated-mtp-mlx-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 511.0 41.9 27.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o4mct281",
   "source_url": "https://omlx.ai/benchmarks/performance/o4mct281",
   "tg_tokens": null,
   "tps": 41.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-o4mct281-ctx131072",
   "model": "Occamy-1.0-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.2 GB at this context; duplicate traces of the same session: 22o3zkb3, c4s7gx8c, ckqfs6x3, h5sj0sbs; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 330.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "occamy-1.0-abliterated-mtp-mlx-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 128k 330.0 28.8 30.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o4mct281",
   "source_url": "https://omlx.ai/benchmarks/performance/o4mct281",
   "tg_tokens": null,
   "tps": 28.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-3j12i4ra-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: hfe2uaf6; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 80.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ6e on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 6bit; 1k 80.8 9.1 28.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3j12i4ra",
   "source_url": "https://omlx.ai/benchmarks/performance/3j12i4ra",
   "tg_tokens": null,
   "tps": 9.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-3j12i4ra-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace; peak memory 32.3 GB at this context; duplicate traces of the same session: hfe2uaf6; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 89.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ6e on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 6bit; 4k 89.6 8.7 32.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3j12i4ra",
   "source_url": "https://omlx.ai/benchmarks/performance/3j12i4ra",
   "tg_tokens": null,
   "tps": 8.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-gndffpy4-ctx1024",
   "model": "FrogNano-4B-2609",
   "notes": "jang build; oMLX community benchmark trace; peak memory 4.3 GB at this context; duplicate traces of the same session: y2zotui3; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 393.0,
   "provenance": "community",
   "quant": "gptq",
   "quote": "FrogNano-4B-2609-gptq-mlx-jang on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 393.0 26.2 4.3 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gndffpy4",
   "source_url": "https://omlx.ai/benchmarks/performance/gndffpy4",
   "tg_tokens": null,
   "tps": 26.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-gndffpy4-ctx4096",
   "model": "FrogNano-4B-2609",
   "notes": "jang build; oMLX community benchmark trace; peak memory 4.4 GB at this context; duplicate traces of the same session: y2zotui3; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 526.5,
   "provenance": "community",
   "quant": "gptq",
   "quote": "FrogNano-4B-2609-gptq-mlx-jang on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 526.5 27.0 4.4 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gndffpy4",
   "source_url": "https://omlx.ai/benchmarks/performance/gndffpy4",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-ezy81mrn-ctx1024",
   "model": "FrogNano-4B-2609",
   "notes": "jang build; oMLX community benchmark trace; peak memory 4.0 GB at this context; duplicate traces of the same session: g2s0m39o; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 409.4,
   "provenance": "community",
   "quant": "gptq",
   "quote": "FrogNano-4B-2609-gptq-mlx-jang on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 409.4 28.5 4.0 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ezy81mrn",
   "source_url": "https://omlx.ai/benchmarks/performance/ezy81mrn",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-ezy81mrn-ctx4096",
   "model": "FrogNano-4B-2609",
   "notes": "jang build; oMLX community benchmark trace; peak memory 4.4 GB at this context; duplicate traces of the same session: g2s0m39o; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 427.3,
   "provenance": "community",
   "quant": "gptq",
   "quote": "FrogNano-4B-2609-gptq-mlx-jang on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 427.3 27.4 4.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ezy81mrn",
   "source_url": "https://omlx.ai/benchmarks/performance/ezy81mrn",
   "tg_tokens": null,
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-uyzxxuay-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 24.7 GB; TTFT 485436 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 270.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 128k tokens 270.0 PP tok/s 29.8 TG tok/s 485436 TTFT (ms) 24.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uyzxxuay",
   "source_url": "https://omlx.ai/benchmarks/performance/uyzxxuay",
   "tg_tokens": null,
   "tps": 29.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-w7wrmlv9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: mhiyq248; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 127.8 15.3 15.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w7wrmlv9",
   "source_url": "https://omlx.ai/benchmarks/performance/w7wrmlv9",
   "tg_tokens": null,
   "tps": 15.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-w7wrmlv9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: mhiyq248; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 136.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 136.6 15.0 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w7wrmlv9",
   "source_url": "https://omlx.ai/benchmarks/performance/w7wrmlv9",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-jgw46wnk-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 3qzqap80, gpb6c7nq, ihrm7wxo, n3sd8ijq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 592.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 592.7 33.0 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jgw46wnk",
   "source_url": "https://omlx.ai/benchmarks/performance/jgw46wnk",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-jgw46wnk-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 3qzqap80, gpb6c7nq, ihrm7wxo, n3sd8ijq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 694.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 694.3 36.5 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jgw46wnk",
   "source_url": "https://omlx.ai/benchmarks/performance/jgw46wnk",
   "tg_tokens": null,
   "tps": 36.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-jgw46wnk-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 3qzqap80, gpb6c7nq, ihrm7wxo, n3sd8ijq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 716.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 716.6 45.3 19.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jgw46wnk",
   "source_url": "https://omlx.ai/benchmarks/performance/jgw46wnk",
   "tg_tokens": null,
   "tps": 45.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-jgw46wnk-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 3qzqap80, gpb6c7nq, ihrm7wxo, n3sd8ijq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 693.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 693.9 44.2 20.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jgw46wnk",
   "source_url": "https://omlx.ai/benchmarks/performance/jgw46wnk",
   "tg_tokens": null,
   "tps": 44.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-jgw46wnk-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: 3qzqap80, gpb6c7nq, ihrm7wxo, n3sd8ijq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 635.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 635.2 42.3 23.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jgw46wnk",
   "source_url": "https://omlx.ai/benchmarks/performance/jgw46wnk",
   "tg_tokens": null,
   "tps": 42.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-idfj8gyc-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 60lpdbb4, ecu995pg, nrxs5yyn, xwrfkr9a; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 591.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 591.6 42.2 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/idfj8gyc",
   "source_url": "https://omlx.ai/benchmarks/performance/idfj8gyc",
   "tg_tokens": null,
   "tps": 42.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-idfj8gyc-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 60lpdbb4, ecu995pg, nrxs5yyn, xwrfkr9a; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 694.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 694.7 49.9 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/idfj8gyc",
   "source_url": "https://omlx.ai/benchmarks/performance/idfj8gyc",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-idfj8gyc-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 60lpdbb4, ecu995pg, nrxs5yyn, xwrfkr9a; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 710.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 710.7 26.3 19.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/idfj8gyc",
   "source_url": "https://omlx.ai/benchmarks/performance/idfj8gyc",
   "tg_tokens": null,
   "tps": 26.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-idfj8gyc-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 60lpdbb4, ecu995pg, nrxs5yyn, xwrfkr9a; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 677.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 677.4 33.8 20.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/idfj8gyc",
   "source_url": "https://omlx.ai/benchmarks/performance/idfj8gyc",
   "tg_tokens": null,
   "tps": 33.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-idfj8gyc-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: 60lpdbb4, ecu995pg, nrxs5yyn, xwrfkr9a; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 633.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 633.7 25.9 23.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/idfj8gyc",
   "source_url": "https://omlx.ai/benchmarks/performance/idfj8gyc",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-44zio6kh-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 8srgxsrz, nzzbvvem; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 588.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 588.0 33.0 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/44zio6kh",
   "source_url": "https://omlx.ai/benchmarks/performance/44zio6kh",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-44zio6kh-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 8srgxsrz, nzzbvvem; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 694.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 694.5 57.0 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/44zio6kh",
   "source_url": "https://omlx.ai/benchmarks/performance/44zio6kh",
   "tg_tokens": null,
   "tps": 57.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-44zio6kh-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 8srgxsrz, nzzbvvem; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 709.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 709.9 22.7 19.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/44zio6kh",
   "source_url": "https://omlx.ai/benchmarks/performance/44zio6kh",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-6o3975ou-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: m3djev7d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1254.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,254 90.6 14.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6o3975ou",
   "source_url": "https://omlx.ai/benchmarks/performance/6o3975ou",
   "tg_tokens": null,
   "tps": 90.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-6o3975ou-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: m3djev7d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1268.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,268 87.7 14.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6o3975ou",
   "source_url": "https://omlx.ai/benchmarks/performance/6o3975ou",
   "tg_tokens": null,
   "tps": 87.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-91ogy3py-ctx4096",
   "model": "Ornith-1.5-35B-A3B-V2-MTPLX",
   "notes": "oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: 3to9xvk2, c2a15r0e, myun7mkj, yqxi116h; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1027.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ornith-ai-Ornith-1.5-35B-A3B-V2-MTPLX on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,027 79.7 24.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/91ogy3py",
   "source_url": "https://omlx.ai/benchmarks/performance/91ogy3py",
   "tg_tokens": null,
   "tps": 79.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-91ogy3py-ctx16384",
   "model": "Ornith-1.5-35B-A3B-V2-MTPLX",
   "notes": "oMLX community benchmark trace; peak memory 25.2 GB at this context; duplicate traces of the same session: 3to9xvk2, c2a15r0e, myun7mkj, yqxi116h; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 861.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ornith-ai-Ornith-1.5-35B-A3B-V2-MTPLX on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 861.9 67.4 25.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/91ogy3py",
   "source_url": "https://omlx.ai/benchmarks/performance/91ogy3py",
   "tg_tokens": null,
   "tps": 67.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-91ogy3py-ctx32768",
   "model": "Ornith-1.5-35B-A3B-V2-MTPLX",
   "notes": "oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: 3to9xvk2, c2a15r0e, myun7mkj, yqxi116h; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 701.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ornith-ai-Ornith-1.5-35B-A3B-V2-MTPLX on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 701.3 60.9 26.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/91ogy3py",
   "source_url": "https://omlx.ai/benchmarks/performance/91ogy3py",
   "tg_tokens": null,
   "tps": 60.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-91ogy3py-ctx65536",
   "model": "Ornith-1.5-35B-A3B-V2-MTPLX",
   "notes": "oMLX community benchmark trace; peak memory 28.4 GB at this context; duplicate traces of the same session: 3to9xvk2, c2a15r0e, myun7mkj, yqxi116h; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 511.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ornith-ai-Ornith-1.5-35B-A3B-V2-MTPLX on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 511.4 44.1 28.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/91ogy3py",
   "source_url": "https://omlx.ai/benchmarks/performance/91ogy3py",
   "tg_tokens": null,
   "tps": 44.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-91ogy3py-ctx131072",
   "model": "Ornith-1.5-35B-A3B-V2-MTPLX",
   "notes": "oMLX community benchmark trace; peak memory 30.9 GB at this context; duplicate traces of the same session: 3to9xvk2, c2a15r0e, myun7mkj, yqxi116h; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 334.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ornith-ai-Ornith-1.5-35B-A3B-V2-MTPLX on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 128k 334.5 26.4 30.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/91ogy3py",
   "source_url": "https://omlx.ai/benchmarks/performance/91ogy3py",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-121iyvtl-ctx1024",
   "model": "FrogNano-4B-2609",
   "notes": "jang build; oMLX community benchmark trace; peak memory 4.3 GB at this context; duplicate traces of the same session: itsu3u15; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 394.8,
   "provenance": "community",
   "quant": "gptq",
   "quote": "FrogNano-4B-2609-gptq-mlx-jang on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 394.8 27.1 4.3 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/121iyvtl",
   "source_url": "https://omlx.ai/benchmarks/performance/121iyvtl",
   "tg_tokens": null,
   "tps": 27.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-121iyvtl-ctx4096",
   "model": "FrogNano-4B-2609",
   "notes": "jang build; oMLX community benchmark trace; peak memory 4.4 GB at this context; duplicate traces of the same session: itsu3u15; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 519.0,
   "provenance": "community",
   "quant": "gptq",
   "quote": "FrogNano-4B-2609-gptq-mlx-jang on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 519.0 26.4 4.4 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/121iyvtl",
   "source_url": "https://omlx.ai/benchmarks/performance/121iyvtl",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ze53e58n-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.2 GB at this context; duplicate traces of the same session: d86mgd6n, n90nisno; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2342.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,342 126.1 173.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ze53e58n",
   "source_url": "https://omlx.ai/benchmarks/performance/ze53e58n",
   "tg_tokens": null,
   "tps": 126.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ze53e58n-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.7 GB at this context; duplicate traces of the same session: d86mgd6n, n90nisno; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2319.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,319 89.1 173.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ze53e58n",
   "source_url": "https://omlx.ai/benchmarks/performance/ze53e58n",
   "tg_tokens": null,
   "tps": 89.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ze53e58n-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.9 GB at this context; duplicate traces of the same session: d86mgd6n, n90nisno; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2262.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,262 141.0 174.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ze53e58n",
   "source_url": "https://omlx.ai/benchmarks/performance/ze53e58n",
   "tg_tokens": null,
   "tps": 141.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-xxt0kd0v-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: 9ss7jjfv, d69klxxa, tlk787dl, zfsz8kn7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 935.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 935.7 73.3 32.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xxt0kd0v",
   "source_url": "https://omlx.ai/benchmarks/performance/xxt0kd0v",
   "tg_tokens": null,
   "tps": 73.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-xxt0kd0v-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.9 GB at this context; duplicate traces of the same session: 9ss7jjfv, d69klxxa, tlk787dl, zfsz8kn7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 798.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 16k 798.0 61.6 32.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xxt0kd0v",
   "source_url": "https://omlx.ai/benchmarks/performance/xxt0kd0v",
   "tg_tokens": null,
   "tps": 61.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-xxt0kd0v-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.2 GB at this context; duplicate traces of the same session: 9ss7jjfv, d69klxxa, tlk787dl, zfsz8kn7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 658.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 32k 658.7 52.0 34.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xxt0kd0v",
   "source_url": "https://omlx.ai/benchmarks/performance/xxt0kd0v",
   "tg_tokens": null,
   "tps": 52.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-xxt0kd0v-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: 9ss7jjfv, d69klxxa, tlk787dl, zfsz8kn7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 498.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 64k 498.5 38.8 36.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xxt0kd0v",
   "source_url": "https://omlx.ai/benchmarks/performance/xxt0kd0v",
   "tg_tokens": null,
   "tps": 38.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-xxt0kd0v-ctx131072",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.6 GB at this context; duplicate traces of the same session: 9ss7jjfv, d69klxxa, tlk787dl, zfsz8kn7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 331.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 128k 331.3 25.7 38.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xxt0kd0v",
   "source_url": "https://omlx.ai/benchmarks/performance/xxt0kd0v",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-2wbx259r-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: mhdp0qr1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 822.2,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 6bit; 1k 822.2 57.0 29.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2wbx259r",
   "source_url": "https://omlx.ai/benchmarks/performance/2wbx259r",
   "tg_tokens": null,
   "tps": 57.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-2wbx259r-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: mhdp0qr1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1372.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 6bit; 4k 1,372 78.9 32.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2wbx259r",
   "source_url": "https://omlx.ai/benchmarks/performance/2wbx259r",
   "tg_tokens": null,
   "tps": 78.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-pyal3up3-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 33.1 GB; TTFT 24253 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1351.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k tokens 1,351 PP tok/s 130.9 TG tok/s 24253 TTFT (ms) 33.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pyal3up3",
   "source_url": "https://omlx.ai/benchmarks/performance/pyal3up3",
   "tg_tokens": null,
   "tps": 130.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-yq84l6yi-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 33.1 GB; TTFT 24782 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1322.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k tokens 1,322 PP tok/s 92.9 TG tok/s 24782 TTFT (ms) 33.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yq84l6yi",
   "source_url": "https://omlx.ai/benchmarks/performance/yq84l6yi",
   "tg_tokens": null,
   "tps": 92.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bj6hu54y-ctx4096",
   "model": "Qwen3.8-27B-Coder390-LynnStyle-Mixed",
   "notes": "A4toA8 build; oMLX community benchmark trace, single run at 4k context; peak memory 20.6 GB; TTFT 4226 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 969.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Coder390-LynnStyle-Mixed-MLX-A4toA8 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 969.3 PP tok/s 26.7 TG tok/s 4226 TTFT (ms) 20.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bj6hu54y",
   "source_url": "https://omlx.ai/benchmarks/performance/bj6hu54y",
   "tg_tokens": null,
   "tps": 26.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mn7ml4k4-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 26.6 GB; TTFT 10656 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 384.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 384.4 PP tok/s 23.6 TG tok/s 10656 TTFT (ms) 26.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mn7ml4k4",
   "source_url": "https://omlx.ai/benchmarks/performance/mn7ml4k4",
   "tg_tokens": null,
   "tps": 23.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0iirgtsc-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: fjqcgghw; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 577.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 577.7 31.1 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0iirgtsc",
   "source_url": "https://omlx.ai/benchmarks/performance/0iirgtsc",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0iirgtsc-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: fjqcgghw; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 670.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 670.3 34.6 19.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0iirgtsc",
   "source_url": "https://omlx.ai/benchmarks/performance/0iirgtsc",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3i1u1ldz-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 6z5bpfyg, pp3cy7gu, vtw9pdrl, zyswkg7z; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 579.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 579.5 38.4 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3i1u1ldz",
   "source_url": "https://omlx.ai/benchmarks/performance/3i1u1ldz",
   "tg_tokens": null,
   "tps": 38.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3i1u1ldz-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 6z5bpfyg, pp3cy7gu, vtw9pdrl, zyswkg7z; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 666.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 666.8 25.2 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3i1u1ldz",
   "source_url": "https://omlx.ai/benchmarks/performance/3i1u1ldz",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3i1u1ldz-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 6z5bpfyg, pp3cy7gu, vtw9pdrl, zyswkg7z; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 690.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 690.1 31.9 19.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3i1u1ldz",
   "source_url": "https://omlx.ai/benchmarks/performance/3i1u1ldz",
   "tg_tokens": null,
   "tps": 31.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3i1u1ldz-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 6z5bpfyg, pp3cy7gu, vtw9pdrl, zyswkg7z; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 685.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 685.4 41.3 20.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3i1u1ldz",
   "source_url": "https://omlx.ai/benchmarks/performance/3i1u1ldz",
   "tg_tokens": null,
   "tps": 41.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3i1u1ldz-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: 6z5bpfyg, pp3cy7gu, vtw9pdrl, zyswkg7z; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 567.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 567.2 21.4 23.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3i1u1ldz",
   "source_url": "https://omlx.ai/benchmarks/performance/3i1u1ldz",
   "tg_tokens": null,
   "tps": 21.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-is0ovk4b-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 2ifrof9v, 4m4oyw2a, 8che0yg3, m02f4bkj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 583.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 583.1 32.1 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/is0ovk4b",
   "source_url": "https://omlx.ai/benchmarks/performance/is0ovk4b",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-is0ovk4b-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 2ifrof9v, 4m4oyw2a, 8che0yg3, m02f4bkj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 692.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 692.0 33.9 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/is0ovk4b",
   "source_url": "https://omlx.ai/benchmarks/performance/is0ovk4b",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-is0ovk4b-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 2ifrof9v, 4m4oyw2a, 8che0yg3, m02f4bkj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 707.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 707.9 31.7 19.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/is0ovk4b",
   "source_url": "https://omlx.ai/benchmarks/performance/is0ovk4b",
   "tg_tokens": null,
   "tps": 31.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-is0ovk4b-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 2ifrof9v, 4m4oyw2a, 8che0yg3, m02f4bkj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 680.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 680.6 46.1 20.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/is0ovk4b",
   "source_url": "https://omlx.ai/benchmarks/performance/is0ovk4b",
   "tg_tokens": null,
   "tps": 46.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-is0ovk4b-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: 2ifrof9v, 4m4oyw2a, 8che0yg3, m02f4bkj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 633.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 633.3 45.9 23.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/is0ovk4b",
   "source_url": "https://omlx.ai/benchmarks/performance/is0ovk4b",
   "tg_tokens": null,
   "tps": 45.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-loem59tm-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "Jundot build; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 25.3 GB; TTFT 143185 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 228.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp-Jundot on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k tokens 228.9 PP tok/s 17.5 TG tok/s 143185 TTFT (ms) 25.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/loem59tm",
   "source_url": "https://omlx.ai/benchmarks/performance/loem59tm",
   "tg_tokens": null,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-exe6khgq-ctx1024",
   "model": "Qwen3.8-Flash-Next-RCO256",
   "notes": "oMLX community benchmark trace; peak memory 30.2 GB at this context; duplicate traces of the same session: 765i386o, z36f3r9d; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 684.7,
   "provenance": "community",
   "quant": "oQ3n2",
   "quote": "Qwen3.8-Flash-Next-RCO256-oQ3n2-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 3bit; 1k 684.7 43.0 30.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/exe6khgq",
   "source_url": "https://omlx.ai/benchmarks/performance/exe6khgq",
   "tg_tokens": null,
   "tps": 43.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-exe6khgq-ctx4096",
   "model": "Qwen3.8-Flash-Next-RCO256",
   "notes": "oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: 765i386o, z36f3r9d; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 921.2,
   "provenance": "community",
   "quant": "oQ3n2",
   "quote": "Qwen3.8-Flash-Next-RCO256-oQ3n2-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 3bit; 4k 921.2 36.8 30.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/exe6khgq",
   "source_url": "https://omlx.ai/benchmarks/performance/exe6khgq",
   "tg_tokens": null,
   "tps": 36.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-exe6khgq-ctx65536",
   "model": "Qwen3.8-Flash-Next-RCO256",
   "notes": "oMLX community benchmark trace; peak memory 32.4 GB at this context; duplicate traces of the same session: 765i386o, z36f3r9d; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 953.8,
   "provenance": "community",
   "quant": "oQ3n2",
   "quote": "Qwen3.8-Flash-Next-RCO256-oQ3n2-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 3bit; 64k 953.8 35.7 32.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/exe6khgq",
   "source_url": "https://omlx.ai/benchmarks/performance/exe6khgq",
   "tg_tokens": null,
   "tps": 35.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-u79uoufy-ctx1024",
   "model": "Qwen3.8-Flash-Next-RCO256",
   "notes": "oMLX community benchmark trace; peak memory 30.2 GB at this context; duplicate traces of the same session: oe5mi61p, zxz8h28h; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 607.7,
   "provenance": "community",
   "quant": "oQ3n2",
   "quote": "Qwen3.8-Flash-Next-RCO256-oQ3n2-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 3bit; 1k 607.7 38.7 30.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u79uoufy",
   "source_url": "https://omlx.ai/benchmarks/performance/u79uoufy",
   "tg_tokens": null,
   "tps": 38.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-u79uoufy-ctx4096",
   "model": "Qwen3.8-Flash-Next-RCO256",
   "notes": "oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: oe5mi61p, zxz8h28h; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 956.0,
   "provenance": "community",
   "quant": "oQ3n2",
   "quote": "Qwen3.8-Flash-Next-RCO256-oQ3n2-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 3bit; 4k 956.0 38.6 30.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u79uoufy",
   "source_url": "https://omlx.ai/benchmarks/performance/u79uoufy",
   "tg_tokens": null,
   "tps": 38.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-u79uoufy-ctx65536",
   "model": "Qwen3.8-Flash-Next-RCO256",
   "notes": "oMLX community benchmark trace; peak memory 32.4 GB at this context; duplicate traces of the same session: oe5mi61p, zxz8h28h; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1027.0,
   "provenance": "community",
   "quant": "oQ3n2",
   "quote": "Qwen3.8-Flash-Next-RCO256-oQ3n2-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 3bit; 64k 1,027 36.8 32.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u79uoufy",
   "source_url": "https://omlx.ai/benchmarks/performance/u79uoufy",
   "tg_tokens": null,
   "tps": 36.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-oo7lcj4g-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 30.4 GB at this context; duplicate traces of the same session: q7526nx6; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 733.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-6bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 733.0 53.4 30.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oo7lcj4g",
   "source_url": "https://omlx.ai/benchmarks/performance/oo7lcj4g",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-oo7lcj4g-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 32.6 GB at this context; duplicate traces of the same session: q7526nx6; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 819.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-6bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 6bit; 8k 819.9 50.3 32.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oo7lcj4g",
   "source_url": "https://omlx.ai/benchmarks/performance/oo7lcj4g",
   "tg_tokens": null,
   "tps": 50.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-3ics0f9e-ctx1024",
   "model": "K2-Horizon-3.7B",
   "notes": "oMLX community benchmark trace; peak memory 5.7 GB at this context; duplicate traces of the same session: 0o8vjqen; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 785.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "K2-Horizon-3.7B-MLX-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 1k 785.0 40.9 5.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ics0f9e",
   "source_url": "https://omlx.ai/benchmarks/performance/3ics0f9e",
   "tg_tokens": null,
   "tps": 40.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-3ics0f9e-ctx8192",
   "model": "K2-Horizon-3.7B",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: 0o8vjqen; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 647.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "K2-Horizon-3.7B-MLX-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 8k 647.8 27.8 6.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ics0f9e",
   "source_url": "https://omlx.ai/benchmarks/performance/3ics0f9e",
   "tg_tokens": null,
   "tps": 27.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-5da1lgu9-ctx131072",
   "model": "Q_pi_S",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 40.2 GB; TTFT 1059038 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 123.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Q_pi_S on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 128k tokens 123.8 PP tok/s 25.6 TG tok/s 1059038 TTFT (ms) 40.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5da1lgu9",
   "source_url": "https://omlx.ai/benchmarks/performance/5da1lgu9",
   "tg_tokens": null,
   "tps": 25.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-0zhbefvl-ctx131072",
   "model": "Q_pi_S",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 40.2 GB; TTFT 1064856 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 123.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Q_pi_S on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 128k tokens 123.1 PP tok/s 20.2 TG tok/s 1064856 TTFT (ms) 40.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0zhbefvl",
   "source_url": "https://omlx.ai/benchmarks/performance/0zhbefvl",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-f41w7mm3-ctx4096",
   "model": "Qwen3 8B",
   "notes": "oMLX community benchmark trace; peak memory 5.4 GB at this context; duplicate traces of the same session: pcu8ytdc; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 225.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-8B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 225.1 20.3 5.4 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f41w7mm3",
   "source_url": "https://omlx.ai/benchmarks/performance/f41w7mm3",
   "tg_tokens": null,
   "tps": 20.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-f41w7mm3-ctx16384",
   "model": "Qwen3 8B",
   "notes": "oMLX community benchmark trace; peak memory 7.1 GB at this context; duplicate traces of the same session: pcu8ytdc; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 174.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-8B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 16k 174.0 13.3 7.1 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f41w7mm3",
   "source_url": "https://omlx.ai/benchmarks/performance/f41w7mm3",
   "tg_tokens": null,
   "tps": 13.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-gkv876pw-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: qds4n98a; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 241.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 241.2 21.3 6.3 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gkv876pw",
   "source_url": "https://omlx.ai/benchmarks/performance/gkv876pw",
   "tg_tokens": null,
   "tps": 21.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-gkv876pw-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: qds4n98a; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 228.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 16k 228.2 18.6 6.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gkv876pw",
   "source_url": "https://omlx.ai/benchmarks/performance/gkv876pw",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zy1l5mqw-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 6.9 GB; TTFT 931 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4400.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 4,400 PP tok/s 66.7 TG tok/s 931 TTFT (ms) 6.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zy1l5mqw",
   "source_url": "https://omlx.ai/benchmarks/performance/zy1l5mqw",
   "tg_tokens": null,
   "tps": 66.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-5iz1l0b4-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: kvypw4bt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1603.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.6-35B-A3B-5bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 5bit; 1k 1,603 94.9 23.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5iz1l0b4",
   "source_url": "https://omlx.ai/benchmarks/performance/5iz1l0b4",
   "tg_tokens": null,
   "tps": 94.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-5iz1l0b4-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: kvypw4bt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2038.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.6-35B-A3B-5bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 5bit; 4k 2,038 96.0 24.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5iz1l0b4",
   "source_url": "https://omlx.ai/benchmarks/performance/5iz1l0b4",
   "tg_tokens": null,
   "tps": 96.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-05wrf0bg-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 33.4 GB; TTFT 18624 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1760.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 32k tokens 1,760 PP tok/s 142.4 TG tok/s 18624 TTFT (ms) 33.4 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/05wrf0bg",
   "source_url": "https://omlx.ai/benchmarks/performance/05wrf0bg",
   "tg_tokens": null,
   "tps": 142.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wnn0uicj-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 33.5 GB; TTFT 84997 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 771.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 64k tokens 771.0 PP tok/s 16.0 TG tok/s 84997 TTFT (ms) 33.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wnn0uicj",
   "source_url": "https://omlx.ai/benchmarks/performance/wnn0uicj",
   "tg_tokens": null,
   "tps": 16.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-sb5dp2wg-ctx4096",
   "model": "Qwen3.8-27B-Coder390-LynnStyle-Mixed",
   "notes": "A4toA8 build; MTPFix; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 705m8172, h2ez4ley; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 936.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Coder390-LynnStyle-Mixed-MLX-A4toA8-MTPFix on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 936.2 52.0 20.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sb5dp2wg",
   "source_url": "https://omlx.ai/benchmarks/performance/sb5dp2wg",
   "tg_tokens": null,
   "tps": 52.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-sb5dp2wg-ctx16384",
   "model": "Qwen3.8-27B-Coder390-LynnStyle-Mixed",
   "notes": "A4toA8 build; MTPFix; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 705m8172, h2ez4ley; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 894.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Coder390-LynnStyle-Mixed-MLX-A4toA8-MTPFix on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 894.1 71.1 22.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sb5dp2wg",
   "source_url": "https://omlx.ai/benchmarks/performance/sb5dp2wg",
   "tg_tokens": null,
   "tps": 71.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-sb5dp2wg-ctx131072",
   "model": "Qwen3.8-27B-Coder390-LynnStyle-Mixed",
   "notes": "A4toA8 build; MTPFix; oMLX community benchmark trace; peak memory 37.9 GB at this context; duplicate traces of the same session: 705m8172, h2ez4ley; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 554.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Coder390-LynnStyle-Mixed-MLX-A4toA8-MTPFix on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 554.3 46.0 37.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sb5dp2wg",
   "source_url": "https://omlx.ai/benchmarks/performance/sb5dp2wg",
   "tg_tokens": null,
   "tps": 46.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-7zrwvnsr-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 16.6 GB; TTFT 8473 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 120.8,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k tokens 120.8 PP tok/s 14.5 TG tok/s 8473 TTFT (ms) 16.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7zrwvnsr",
   "source_url": "https://omlx.ai/benchmarks/performance/7zrwvnsr",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-7qhzx87p-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 21.5 GB; TTFT 63374 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 129.3,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k tokens 129.3 PP tok/s 13.6 TG tok/s 63374 TTFT (ms) 21.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7qhzx87p",
   "source_url": "https://omlx.ai/benchmarks/performance/7qhzx87p",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-ihwhvn0y-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 8636 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 118.6,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k tokens 118.6 PP tok/s 15.4 TG tok/s 8636 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ihwhvn0y",
   "source_url": "https://omlx.ai/benchmarks/performance/ihwhvn0y",
   "tg_tokens": null,
   "tps": 15.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-4busvqc1-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 16.6 GB; TTFT 8386 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 122.1,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k tokens 122.1 PP tok/s 14.2 TG tok/s 8386 TTFT (ms) 16.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4busvqc1",
   "source_url": "https://omlx.ai/benchmarks/performance/4busvqc1",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-bwyitof1-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 8182 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.1,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k tokens 125.1 PP tok/s 16.9 TG tok/s 8182 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bwyitof1",
   "source_url": "https://omlx.ai/benchmarks/performance/bwyitof1",
   "tg_tokens": null,
   "tps": 16.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-46gw399m-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 24.6 GB; TTFT 64249 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.5,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k tokens 127.5 PP tok/s 18.5 TG tok/s 64249 TTFT (ms) 24.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/46gw399m",
   "source_url": "https://omlx.ai/benchmarks/performance/46gw399m",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rgloecx5-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 19.1 GB; TTFT 3072 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1334.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 1,334 PP tok/s 58.9 TG tok/s 3072 TTFT (ms) 19.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rgloecx5",
   "source_url": "https://omlx.ai/benchmarks/performance/rgloecx5",
   "tg_tokens": null,
   "tps": 58.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ppu6jmtm-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 19.1 GB; TTFT 3170 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1292.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 1,292 PP tok/s 58.0 TG tok/s 3170 TTFT (ms) 19.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ppu6jmtm",
   "source_url": "https://omlx.ai/benchmarks/performance/ppu6jmtm",
   "tg_tokens": null,
   "tps": 58.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gc5xee78-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 18.8 GB; TTFT 4569 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 896.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 896.6 PP tok/s 63.5 TG tok/s 4569 TTFT (ms) 18.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gc5xee78",
   "source_url": "https://omlx.ai/benchmarks/performance/gc5xee78",
   "tg_tokens": null,
   "tps": 63.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-b2o22b4e-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 42.6 GB; TTFT 54534 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 150.2,
   "provenance": "community",
   "quant": "oQ2",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ2-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 2bit; 8k tokens 150.2 PP tok/s 15.7 TG tok/s 54534 TTFT (ms) 42.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b2o22b4e",
   "source_url": "https://omlx.ai/benchmarks/performance/b2o22b4e",
   "tg_tokens": null,
   "tps": 15.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3gduhaxg-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 49.7 GB; TTFT 178481 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 183.6,
   "provenance": "community",
   "quant": "oQ2",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ2-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 2bit; 32k tokens 183.6 PP tok/s 12.4 TG tok/s 178481 TTFT (ms) 49.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3gduhaxg",
   "source_url": "https://omlx.ai/benchmarks/performance/3gduhaxg",
   "tg_tokens": null,
   "tps": 12.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-44a9ry0v-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 1k context; peak memory 20.1 GB; TTFT 1131 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 905.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k tokens 905.0 PP tok/s 79.4 TG tok/s 1131 TTFT (ms) 20.1 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/44a9ry0v",
   "source_url": "https://omlx.ai/benchmarks/performance/44a9ry0v",
   "tg_tokens": null,
   "tps": 79.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-lhixgp4t-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 36.4 GB; TTFT 444784 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 294.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-6bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 6bit; 128k tokens 294.7 PP tok/s 25.3 TG tok/s 444784 TTFT (ms) 36.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lhixgp4t",
   "source_url": "https://omlx.ai/benchmarks/performance/lhixgp4t",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-bvbriiwx-ctx131072",
   "model": "Q_pi_S",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 40.2 GB; TTFT 961576 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 136.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Q_pi_S on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 128k tokens 136.3 PP tok/s 25.7 TG tok/s 961576 TTFT (ms) 40.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bvbriiwx",
   "source_url": "https://omlx.ai/benchmarks/performance/bvbriiwx",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-uiciv7py-ctx1024",
   "model": "Qwen3 8B",
   "notes": "oMLX community benchmark trace; peak memory 5.0 GB at this context; duplicate traces of the same session: pzfzjmn4; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 394.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-8B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 1k 394.8 60.5 5.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uiciv7py",
   "source_url": "https://omlx.ai/benchmarks/performance/uiciv7py",
   "tg_tokens": null,
   "tps": 60.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-uiciv7py-ctx4096",
   "model": "Qwen3 8B",
   "notes": "oMLX community benchmark trace; peak memory 5.5 GB at this context; duplicate traces of the same session: pzfzjmn4; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 376.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-8B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 4k 376.3 54.5 5.5 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uiciv7py",
   "source_url": "https://omlx.ai/benchmarks/performance/uiciv7py",
   "tg_tokens": null,
   "tps": 54.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-99u1penf-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: ronifvgw; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 106.4 15.7 15.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/99u1penf",
   "source_url": "https://omlx.ai/benchmarks/performance/99u1penf",
   "tg_tokens": null,
   "tps": 15.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-99u1penf-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: ronifvgw; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 103.9 15.2 19.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/99u1penf",
   "source_url": "https://omlx.ai/benchmarks/performance/99u1penf",
   "tg_tokens": null,
   "tps": 15.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ptidubsg-ctx1024",
   "model": "Qwen3.8-27B-Coder390-LynnStyle-Mixed",
   "notes": "A4toA8 build; MTPFix; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: wgterbv2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 883.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Coder390-LynnStyle-Mixed-MLX-A4toA8-MTPFix on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 883.7 39.1 20.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ptidubsg",
   "source_url": "https://omlx.ai/benchmarks/performance/ptidubsg",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ptidubsg-ctx199680",
   "model": "Qwen3.8-27B-Coder390-LynnStyle-Mixed",
   "notes": "A4toA8 build; MTPFix; oMLX community benchmark trace; peak memory 36.4 GB at this context; duplicate traces of the same session: wgterbv2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 445.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Coder390-LynnStyle-Mixed-MLX-A4toA8-MTPFix on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 195k 445.6 14.0 36.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ptidubsg",
   "source_url": "https://omlx.ai/benchmarks/performance/ptidubsg",
   "tg_tokens": null,
   "tps": 14.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-rmn1rs2o-ctx1024",
   "model": "Qwen3.8-27B-Coder390-LynnStyle-Mixed",
   "notes": "A4toA8 build; MTPFix; oMLX community benchmark trace, single run at 1k context; peak memory 20.1 GB; TTFT 1197 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 855.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Coder390-LynnStyle-Mixed-MLX-A4toA8-MTPFix on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 855.4 PP tok/s 54.7 TG tok/s 1197 TTFT (ms) 20.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rmn1rs2o",
   "source_url": "https://omlx.ai/benchmarks/performance/rmn1rs2o",
   "tg_tokens": null,
   "tps": 54.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-vvnjpfu0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: c3uhiixn; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 67.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 67.7 6.6 17.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vvnjpfu0",
   "source_url": "https://omlx.ai/benchmarks/performance/vvnjpfu0",
   "tg_tokens": null,
   "tps": 6.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-vvnjpfu0-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: c3uhiixn; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 62.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 16k 62.9 6.1 18.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vvnjpfu0",
   "source_url": "https://omlx.ai/benchmarks/performance/vvnjpfu0",
   "tg_tokens": null,
   "tps": 6.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-2xnjbk81-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: 7ep5tbdc; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 67.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 67.3 6.6 17.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2xnjbk81",
   "source_url": "https://omlx.ai/benchmarks/performance/2xnjbk81",
   "tg_tokens": null,
   "tps": 6.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-2xnjbk81-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 7ep5tbdc; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 63.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 16k 63.3 6.1 18.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2xnjbk81",
   "source_url": "https://omlx.ai/benchmarks/performance/2xnjbk81",
   "tg_tokens": null,
   "tps": 6.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-445gb4av-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 9001 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 113.8,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k tokens 113.8 PP tok/s 18.6 TG tok/s 9001 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/445gb4av",
   "source_url": "https://omlx.ai/benchmarks/performance/445gb4av",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-rsqsgmeh-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 8557 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 119.7,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k tokens 119.7 PP tok/s 19.8 TG tok/s 8557 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rsqsgmeh",
   "source_url": "https://omlx.ai/benchmarks/performance/rsqsgmeh",
   "tg_tokens": null,
   "tps": 19.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-3yxza48w-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 655vw01n, i2dofblr; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 102.6 23.7 18.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3yxza48w",
   "source_url": "https://omlx.ai/benchmarks/performance/3yxza48w",
   "tg_tokens": null,
   "tps": 23.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-3yxza48w-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: 655vw01n, i2dofblr; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 103.0 25.4 22.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3yxza48w",
   "source_url": "https://omlx.ai/benchmarks/performance/3yxza48w",
   "tg_tokens": null,
   "tps": 25.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-3yxza48w-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 655vw01n, i2dofblr; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 101.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 8k 101.5 27.5 23.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3yxza48w",
   "source_url": "https://omlx.ai/benchmarks/performance/3yxza48w",
   "tg_tokens": null,
   "tps": 27.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-btg504dh-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: p38anvbt, sh31z0be; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 525.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 525.5 43.2 24.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/btg504dh",
   "source_url": "https://omlx.ai/benchmarks/performance/btg504dh",
   "tg_tokens": null,
   "tps": 43.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-btg504dh-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: p38anvbt, sh31z0be; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 450.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 450.2 40.6 24.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/btg504dh",
   "source_url": "https://omlx.ai/benchmarks/performance/btg504dh",
   "tg_tokens": null,
   "tps": 40.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-btg504dh-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: p38anvbt, sh31z0be; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 344.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 344.9 28.6 28.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/btg504dh",
   "source_url": "https://omlx.ai/benchmarks/performance/btg504dh",
   "tg_tokens": null,
   "tps": 28.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-19z2nvk1-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 28.0 GB; TTFT 168205 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 389.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k tokens 389.6 PP tok/s 32.3 TG tok/s 168205 TTFT (ms) 28.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/19z2nvk1",
   "source_url": "https://omlx.ai/benchmarks/performance/19z2nvk1",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-pjtga02s-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: gi8u9di6, q9nyrhr0, xevba35b; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 460.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 460.6 30.9 17.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pjtga02s",
   "source_url": "https://omlx.ai/benchmarks/performance/pjtga02s",
   "tg_tokens": null,
   "tps": 30.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-pjtga02s-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: gi8u9di6, q9nyrhr0, xevba35b; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 465.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 465.7 25.8 18.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pjtga02s",
   "source_url": "https://omlx.ai/benchmarks/performance/pjtga02s",
   "tg_tokens": null,
   "tps": 25.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-pjtga02s-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: gi8u9di6, q9nyrhr0, xevba35b; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 365.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 365.1 15.7 22.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pjtga02s",
   "source_url": "https://omlx.ai/benchmarks/performance/pjtga02s",
   "tg_tokens": null,
   "tps": 15.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-pjtga02s-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: gi8u9di6, q9nyrhr0, xevba35b; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 270.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 128k 270.2 10.7 27.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pjtga02s",
   "source_url": "https://omlx.ai/benchmarks/performance/pjtga02s",
   "tg_tokens": null,
   "tps": 10.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-kkg75eiq-ctx1024",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "vision; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: 2rskrdt9; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 841.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ5e-vision on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 5bit; 1k 841.0 67.5 24.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kkg75eiq",
   "source_url": "https://omlx.ai/benchmarks/performance/kkg75eiq",
   "tg_tokens": null,
   "tps": 67.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-kkg75eiq-ctx32768",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "vision; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: 2rskrdt9; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 659.8,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ5e-vision on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 5bit; 32k 659.8 51.2 25.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kkg75eiq",
   "source_url": "https://omlx.ai/benchmarks/performance/kkg75eiq",
   "tg_tokens": null,
   "tps": 51.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-rh0nv41a-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 45.4 GB; TTFT 159209 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.8,
   "provenance": "community",
   "quant": "oQ2",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ2-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 2bit; 32k tokens 205.8 PP tok/s 15.9 TG tok/s 159209 TTFT (ms) 45.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rh0nv41a",
   "source_url": "https://omlx.ai/benchmarks/performance/rh0nv41a",
   "tg_tokens": null,
   "tps": 15.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-84yj3uav-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 45.4 GB; TTFT 158086 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 207.3,
   "provenance": "community",
   "quant": "oQ2",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ2-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 2bit; 32k tokens 207.3 PP tok/s 17.8 TG tok/s 158086 TTFT (ms) 45.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/84yj3uav",
   "source_url": "https://omlx.ai/benchmarks/performance/84yj3uav",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-dwq9zmiw-ctx16384",
   "model": "Occamy-1.0-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: 08zhdh61, nfd3kacn, x2cxevt5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 865.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "occamy-1.0-abliterated-mtp-mlx-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 865.1 91.4 24.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dwq9zmiw",
   "source_url": "https://omlx.ai/benchmarks/performance/dwq9zmiw",
   "tg_tokens": null,
   "tps": 91.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-dwq9zmiw-ctx32768",
   "model": "Occamy-1.0-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: 08zhdh61, nfd3kacn, x2cxevt5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 703.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "occamy-1.0-abliterated-mtp-mlx-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 703.4 82.4 25.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dwq9zmiw",
   "source_url": "https://omlx.ai/benchmarks/performance/dwq9zmiw",
   "tg_tokens": null,
   "tps": 82.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-dwq9zmiw-ctx65536",
   "model": "Occamy-1.0-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.6 GB at this context; duplicate traces of the same session: 08zhdh61, nfd3kacn, x2cxevt5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 512.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "occamy-1.0-abliterated-mtp-mlx-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 512.3 62.7 27.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dwq9zmiw",
   "source_url": "https://omlx.ai/benchmarks/performance/dwq9zmiw",
   "tg_tokens": null,
   "tps": 62.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-dwq9zmiw-ctx131072",
   "model": "Occamy-1.0-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.2 GB at this context; duplicate traces of the same session: 08zhdh61, nfd3kacn, x2cxevt5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 334.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "occamy-1.0-abliterated-mtp-mlx-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 128k 334.7 39.5 30.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dwq9zmiw",
   "source_url": "https://omlx.ai/benchmarks/performance/dwq9zmiw",
   "tg_tokens": null,
   "tps": 39.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-4gife81o-ctx4096",
   "model": "Ornith-1.5-35B-A3B-V2-MTPLX",
   "notes": "oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: jb2uxfd1, mv2ztmam, xn1radra, xv9gltiq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1034.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ornith-ai-Ornith-1.5-35B-A3B-V2-MTPLX on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,034 111.7 24.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4gife81o",
   "source_url": "https://omlx.ai/benchmarks/performance/4gife81o",
   "tg_tokens": null,
   "tps": 111.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-4gife81o-ctx16384",
   "model": "Ornith-1.5-35B-A3B-V2-MTPLX",
   "notes": "oMLX community benchmark trace; peak memory 25.2 GB at this context; duplicate traces of the same session: jb2uxfd1, mv2ztmam, xn1radra, xv9gltiq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 865.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ornith-ai-Ornith-1.5-35B-A3B-V2-MTPLX on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 865.0 71.5 25.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4gife81o",
   "source_url": "https://omlx.ai/benchmarks/performance/4gife81o",
   "tg_tokens": null,
   "tps": 71.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-4gife81o-ctx32768",
   "model": "Ornith-1.5-35B-A3B-V2-MTPLX",
   "notes": "oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: jb2uxfd1, mv2ztmam, xn1radra, xv9gltiq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 700.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ornith-ai-Ornith-1.5-35B-A3B-V2-MTPLX on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 700.8 78.6 26.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4gife81o",
   "source_url": "https://omlx.ai/benchmarks/performance/4gife81o",
   "tg_tokens": null,
   "tps": 78.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-4gife81o-ctx65536",
   "model": "Ornith-1.5-35B-A3B-V2-MTPLX",
   "notes": "oMLX community benchmark trace; peak memory 28.4 GB at this context; duplicate traces of the same session: jb2uxfd1, mv2ztmam, xn1radra, xv9gltiq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 512.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ornith-ai-Ornith-1.5-35B-A3B-V2-MTPLX on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 512.4 58.2 28.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4gife81o",
   "source_url": "https://omlx.ai/benchmarks/performance/4gife81o",
   "tg_tokens": null,
   "tps": 58.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-4gife81o-ctx131072",
   "model": "Ornith-1.5-35B-A3B-V2-MTPLX",
   "notes": "oMLX community benchmark trace; peak memory 30.9 GB at this context; duplicate traces of the same session: jb2uxfd1, mv2ztmam, xn1radra, xv9gltiq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 334.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ornith-ai-Ornith-1.5-35B-A3B-V2-MTPLX on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 128k 334.9 40.1 30.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4gife81o",
   "source_url": "https://omlx.ai/benchmarks/performance/4gife81o",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jt87svxp-ctx4096",
   "model": "Qwen3.6-35B",
   "notes": "oMLX community benchmark trace; peak memory 36.4 GB at this context; duplicate traces of the same session: c3achfr5, tw4v2x9c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3911.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "qwen36-35b on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 3,911 82.9 36.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jt87svxp",
   "source_url": "https://omlx.ai/benchmarks/performance/jt87svxp",
   "tg_tokens": null,
   "tps": 82.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jt87svxp-ctx16384",
   "model": "Qwen3.6-35B",
   "notes": "oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: c3achfr5, tw4v2x9c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3425.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "qwen36-35b on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 3,425 77.8 37.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jt87svxp",
   "source_url": "https://omlx.ai/benchmarks/performance/jt87svxp",
   "tg_tokens": null,
   "tps": 77.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jt87svxp-ctx131072",
   "model": "Qwen3.6-35B",
   "notes": "oMLX community benchmark trace; peak memory 46.6 GB at this context; duplicate traces of the same session: c3achfr5, tw4v2x9c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 710.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "qwen36-35b on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 710.9 33.9 46.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jt87svxp",
   "source_url": "https://omlx.ai/benchmarks/performance/jt87svxp",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-dva2k45n-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 6.0 GB at this context; duplicate traces of the same session: udu7e9y4; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 178.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-9B-oQ4e-mtp on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 178.6 31.7 6.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dva2k45n",
   "source_url": "https://omlx.ai/benchmarks/performance/dva2k45n",
   "tg_tokens": null,
   "tps": 31.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-dva2k45n-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: udu7e9y4; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 188.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-9B-oQ4e-mtp on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 188.6 31.5 6.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dva2k45n",
   "source_url": "https://omlx.ai/benchmarks/performance/dva2k45n",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-1v0v5ylk-ctx131072",
   "model": "Q_pi_S",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 40.2 GB; TTFT 954071 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 137.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Q_pi_S on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 128k tokens 137.4 PP tok/s 23.4 TG tok/s 954071 TTFT (ms) 40.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1v0v5ylk",
   "source_url": "https://omlx.ai/benchmarks/performance/1v0v5ylk",
   "tg_tokens": null,
   "tps": 23.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-jx32wqwx-ctx4096",
   "model": "Qwen3 8B",
   "notes": "oMLX community benchmark trace; peak memory 5.5 GB at this context; duplicate traces of the same session: q2ljn3o1; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 378.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-8B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 4k 378.6 54.0 5.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jx32wqwx",
   "source_url": "https://omlx.ai/benchmarks/performance/jx32wqwx",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-jx32wqwx-ctx65536",
   "model": "Qwen3 8B",
   "notes": "oMLX community benchmark trace; peak memory 13.8 GB at this context; duplicate traces of the same session: q2ljn3o1; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 171.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-8B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 64k 171.1 19.0 13.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jx32wqwx",
   "source_url": "https://omlx.ai/benchmarks/performance/jx32wqwx",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-hj5jjkle-ctx4096",
   "model": "Qwen3 8B",
   "notes": "oMLX community benchmark trace; peak memory 5.5 GB at this context; duplicate traces of the same session: u1zhv6r0; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 378.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-8B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 4k 378.3 54.4 5.5 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hj5jjkle",
   "source_url": "https://omlx.ai/benchmarks/performance/hj5jjkle",
   "tg_tokens": null,
   "tps": 54.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-hj5jjkle-ctx32768",
   "model": "Qwen3 8B",
   "notes": "oMLX community benchmark trace; peak memory 9.3 GB at this context; duplicate traces of the same session: u1zhv6r0; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 242.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-8B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 32k 242.2 28.2 9.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hj5jjkle",
   "source_url": "https://omlx.ai/benchmarks/performance/hj5jjkle",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-a60ahc3j-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: hel8xnqe; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 80.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k 80.4 14.7 17.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a60ahc3j",
   "source_url": "https://omlx.ai/benchmarks/performance/a60ahc3j",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-a60ahc3j-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: hel8xnqe; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 72.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 16k 72.1 13.3 18.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a60ahc3j",
   "source_url": "https://omlx.ai/benchmarks/performance/a60ahc3j",
   "tg_tokens": null,
   "tps": 13.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-s857yfqu-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 23.2 GB; TTFT 185322 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 353.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k tokens 353.6 PP tok/s 30.3 TG tok/s 185322 TTFT (ms) 23.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s857yfqu",
   "source_url": "https://omlx.ai/benchmarks/performance/s857yfqu",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ir79xmg6-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 45.4 GB; TTFT 165207 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 198.3,
   "provenance": "community",
   "quant": "oQ2",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ2-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 2bit; 32k tokens 198.3 PP tok/s 12.8 TG tok/s 165207 TTFT (ms) 45.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ir79xmg6",
   "source_url": "https://omlx.ai/benchmarks/performance/ir79xmg6",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-gzdorw9x-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.3 GB at this context; duplicate traces of the same session: 06k9gcxj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 470.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 470.1 55.7 69.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gzdorw9x",
   "source_url": "https://omlx.ai/benchmarks/performance/gzdorw9x",
   "tg_tokens": null,
   "tps": 55.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-gzdorw9x-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.9 GB at this context; duplicate traces of the same session: 06k9gcxj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 842.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 842.2 52.9 69.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gzdorw9x",
   "source_url": "https://omlx.ai/benchmarks/performance/gzdorw9x",
   "tg_tokens": null,
   "tps": 52.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-u7fba59b-ctx4096",
   "model": "Ternary-Bonsai-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 10.1 GB at this context; duplicate traces of the same session: x0jxpp8i; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 233.1,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-27B-mlx-2bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 2bit; 4k 233.1 17.0 10.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u7fba59b",
   "source_url": "https://omlx.ai/benchmarks/performance/u7fba59b",
   "tg_tokens": null,
   "tps": 17.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-u7fba59b-ctx16384",
   "model": "Ternary-Bonsai-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 10.7 GB at this context; duplicate traces of the same session: x0jxpp8i; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 212.9,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-27B-mlx-2bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 2bit; 16k 212.9 14.4 10.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u7fba59b",
   "source_url": "https://omlx.ai/benchmarks/performance/u7fba59b",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-qd7lh1d2-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "fp16; oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: vui92wpo; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 385.3,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "gemma-4-26B-A4B-it-oQ3-fp16 on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 3bit; 1k 385.3 36.5 11.5 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qd7lh1d2",
   "source_url": "https://omlx.ai/benchmarks/performance/qd7lh1d2",
   "tg_tokens": null,
   "tps": 36.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-qd7lh1d2-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "fp16; oMLX community benchmark trace; peak memory 11.7 GB at this context; duplicate traces of the same session: vui92wpo; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 369.4,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "gemma-4-26B-A4B-it-oQ3-fp16 on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 3bit; 4k 369.4 36.1 11.7 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qd7lh1d2",
   "source_url": "https://omlx.ai/benchmarks/performance/qd7lh1d2",
   "tg_tokens": null,
   "tps": 36.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-uhkappt1-ctx1024",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO",
   "notes": "MTP speculative decoding enabled; DFlash2; fp16; oMLX community benchmark trace; peak memory 34.6 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 189.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "nerkyor:Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 189.8 22.0 34.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uhkappt1",
   "source_url": "https://omlx.ai/benchmarks/performance/uhkappt1",
   "tg_tokens": null,
   "tps": 22.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-uhkappt1-ctx4096",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO",
   "notes": "MTP speculative decoding enabled; DFlash2; fp16; oMLX community benchmark trace; peak memory 38.5 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 196.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "nerkyor:Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 196.6 23.5 38.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uhkappt1",
   "source_url": "https://omlx.ai/benchmarks/performance/uhkappt1",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-0673ulbt-ctx65536",
   "model": "Ornith-1.5-35B-A3B-V2-MTPLX",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 28.0 GB; TTFT 126919 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 516.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ornith-ai-Ornith-1.5-35B-A3B-V2-MTPLX on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k tokens 516.4 PP tok/s 57.2 TG tok/s 126919 TTFT (ms) 28.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0673ulbt",
   "source_url": "https://omlx.ai/benchmarks/performance/0673ulbt",
   "tg_tokens": null,
   "tps": 57.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-0ar2v38d-ctx1024",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: jxxpx7su, l2quh723; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 588.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 588.6 92.4 20.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ar2v38d",
   "source_url": "https://omlx.ai/benchmarks/performance/0ar2v38d",
   "tg_tokens": null,
   "tps": 92.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-0ar2v38d-ctx4096",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: jxxpx7su, l2quh723; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1760.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 1,760 89.3 20.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ar2v38d",
   "source_url": "https://omlx.ai/benchmarks/performance/0ar2v38d",
   "tg_tokens": null,
   "tps": 89.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-0ar2v38d-ctx8192",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: jxxpx7su, l2quh723; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3309.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 3,309 86.5 20.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ar2v38d",
   "source_url": "https://omlx.ai/benchmarks/performance/0ar2v38d",
   "tg_tokens": null,
   "tps": 86.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-4edcn80h-ctx1024",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.7 GB at this context; duplicate traces of the same session: 8e8yjt9f; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 389.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 389.6 41.3 177.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4edcn80h",
   "source_url": "https://omlx.ai/benchmarks/performance/4edcn80h",
   "tg_tokens": null,
   "tps": 41.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-4edcn80h-ctx4096",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.8 GB at this context; duplicate traces of the same session: 8e8yjt9f; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 543.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 543.4 31.8 178.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4edcn80h",
   "source_url": "https://omlx.ai/benchmarks/performance/4edcn80h",
   "tg_tokens": null,
   "tps": 31.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-3wbvrwlk-ctx1024",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 7.7 GB; TTFT 1257 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 814.9,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-OptiQ-4bit on M5 (10c); M5 (10c) \u00b7 16 GB \u00b7 4bit; 1k tokens 814.9 PP tok/s 20.7 TG tok/s 1257 TTFT (ms) 7.7 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3wbvrwlk",
   "source_url": "https://omlx.ai/benchmarks/performance/3wbvrwlk",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-gvr38si7-ctx1024",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: 1drmfr5n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3753.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-oQ8-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 3,753 98.3 35.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gvr38si7",
   "source_url": "https://omlx.ai/benchmarks/performance/gvr38si7",
   "tg_tokens": null,
   "tps": 98.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-gvr38si7-ctx4096",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: 1drmfr5n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4726.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-oQ8-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 4,726 97.6 36.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gvr38si7",
   "source_url": "https://omlx.ai/benchmarks/performance/gvr38si7",
   "tg_tokens": null,
   "tps": 97.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-rrobiedz-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 23.6 GB; TTFT 91851 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 356.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 32k tokens 356.8 PP tok/s 50.6 TG tok/s 91851 TTFT (ms) 23.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rrobiedz",
   "source_url": "https://omlx.ai/benchmarks/performance/rrobiedz",
   "tg_tokens": null,
   "tps": 50.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-sqn80fci-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 21.6 GB; TTFT 37770 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 433.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k tokens 433.8 PP tok/s 29.7 TG tok/s 37770 TTFT (ms) 21.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sqn80fci",
   "source_url": "https://omlx.ai/benchmarks/performance/sqn80fci",
   "tg_tokens": null,
   "tps": 29.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-z8ubjhoi-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 23.0 GB; TTFT 35458 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 462.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k tokens 462.1 PP tok/s 57.2 TG tok/s 35458 TTFT (ms) 23.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z8ubjhoi",
   "source_url": "https://omlx.ai/benchmarks/performance/z8ubjhoi",
   "tg_tokens": null,
   "tps": 57.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-tbymk1oi-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 23.0 GB; TTFT 35425 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 462.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k tokens 462.5 PP tok/s 57.2 TG tok/s 35425 TTFT (ms) 23.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tbymk1oi",
   "source_url": "https://omlx.ai/benchmarks/performance/tbymk1oi",
   "tg_tokens": null,
   "tps": 57.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-y0jvhved-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 1795 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 570.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k tokens 570.4 PP tok/s 71.1 TG tok/s 1795 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y0jvhved",
   "source_url": "https://omlx.ai/benchmarks/performance/y0jvhved",
   "tg_tokens": null,
   "tps": 71.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5",
   "id": "omlx-42r9doa6-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 7.2 GB; TTFT 5553 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 184.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 (10c); M5 (10c) \u00b7 16 GB \u00b7 4bit; 1k tokens 184.4 PP tok/s 15.4 TG tok/s 5553 TTFT (ms) 7.2 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/42r9doa6",
   "source_url": "https://omlx.ai/benchmarks/performance/42r9doa6",
   "tg_tokens": null,
   "tps": 15.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-aazbcr61-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 30.9 GB; TTFT 3884 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4218.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k tokens 4,218 PP tok/s 174.8 TG tok/s 3884 TTFT (ms) 30.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aazbcr61",
   "source_url": "https://omlx.ai/benchmarks/performance/aazbcr61",
   "tg_tokens": null,
   "tps": 174.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-k714p3q4-ctx1024",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: i7hewbix; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1123.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,123 70.4 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k714p3q4",
   "source_url": "https://omlx.ai/benchmarks/performance/k714p3q4",
   "tg_tokens": null,
   "tps": 70.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-k714p3q4-ctx4096",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: i7hewbix; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1331.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,331 64.8 19.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k714p3q4",
   "source_url": "https://omlx.ai/benchmarks/performance/k714p3q4",
   "tg_tokens": null,
   "tps": 64.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-whgr0qc9-ctx1024",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: w4ov3v14; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 886.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 886.4 44.7 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/whgr0qc9",
   "source_url": "https://omlx.ai/benchmarks/performance/whgr0qc9",
   "tg_tokens": null,
   "tps": 44.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-whgr0qc9-ctx4096",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: w4ov3v14; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1312.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,312 61.9 19.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/whgr0qc9",
   "source_url": "https://omlx.ai/benchmarks/performance/whgr0qc9",
   "tg_tokens": null,
   "tps": 61.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-62ga69t6-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: 0oyodrf5; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 67.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 67.8 6.6 17.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/62ga69t6",
   "source_url": "https://omlx.ai/benchmarks/performance/62ga69t6",
   "tg_tokens": null,
   "tps": 6.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-62ga69t6-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 0oyodrf5; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 63.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 16k 63.6 6.1 18.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/62ga69t6",
   "source_url": "https://omlx.ai/benchmarks/performance/62ga69t6",
   "tg_tokens": null,
   "tps": 6.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-9u3yrvix-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 23.6 GB; TTFT 192779 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 340.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k tokens 340.0 PP tok/s 31.5 TG tok/s 192779 TTFT (ms) 23.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9u3yrvix",
   "source_url": "https://omlx.ai/benchmarks/performance/9u3yrvix",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-brae2gbt-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "pi build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: dkpf2nsk; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 218.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-pi-4bit-mtp-mlx on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 218.3 7.8 17.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/brae2gbt",
   "source_url": "https://omlx.ai/benchmarks/performance/brae2gbt",
   "tg_tokens": null,
   "tps": 7.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-brae2gbt-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "pi build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: dkpf2nsk; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 202.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-pi-4bit-mtp-mlx on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 16k 202.5 7.4 17.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/brae2gbt",
   "source_url": "https://omlx.ai/benchmarks/performance/brae2gbt",
   "tg_tokens": null,
   "tps": 7.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-zug6txzp-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.3 GB at this context; duplicate traces of the same session: 2yub04xb, 4dtvns2a, luzx0wh3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 600.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 600.1 70.1 69.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zug6txzp",
   "source_url": "https://omlx.ai/benchmarks/performance/zug6txzp",
   "tg_tokens": null,
   "tps": 70.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-zug6txzp-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.9 GB at this context; duplicate traces of the same session: 2yub04xb, 4dtvns2a, luzx0wh3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 739.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 739.8 61.5 69.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zug6txzp",
   "source_url": "https://omlx.ai/benchmarks/performance/zug6txzp",
   "tg_tokens": null,
   "tps": 61.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-zug6txzp-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.9 GB at this context; duplicate traces of the same session: 2yub04xb, 4dtvns2a, luzx0wh3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 840.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 840.6 64.0 69.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zug6txzp",
   "source_url": "https://omlx.ai/benchmarks/performance/zug6txzp",
   "tg_tokens": null,
   "tps": 64.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-zug6txzp-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: 2yub04xb, 4dtvns2a, luzx0wh3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 835.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 835.8 56.1 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zug6txzp",
   "source_url": "https://omlx.ai/benchmarks/performance/zug6txzp",
   "tg_tokens": null,
   "tps": 56.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gylym1xz-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 122.2 GB at this context; duplicate traces of the same session: dcpx5cb0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1088.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "mlx-community--Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 5bit; 1k 1,088 75.7 122.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gylym1xz",
   "source_url": "https://omlx.ai/benchmarks/performance/gylym1xz",
   "tg_tokens": null,
   "tps": 75.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gylym1xz-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 124.0 GB at this context; duplicate traces of the same session: dcpx5cb0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1004.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "mlx-community--Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 5bit; 4k 1,004 86.3 124.0 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gylym1xz",
   "source_url": "https://omlx.ai/benchmarks/performance/gylym1xz",
   "tg_tokens": null,
   "tps": 86.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-h74bj136-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 122.2 GB at this context; duplicate traces of the same session: wvr6g91p; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1085.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "mlx-community--Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 5bit; 1k 1,085 82.5 122.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h74bj136",
   "source_url": "https://omlx.ai/benchmarks/performance/h74bj136",
   "tg_tokens": null,
   "tps": 82.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-h74bj136-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 124.0 GB at this context; duplicate traces of the same session: wvr6g91p; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 997.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "mlx-community--Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 5bit; 4k 997.7 92.6 124.0 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h74bj136",
   "source_url": "https://omlx.ai/benchmarks/performance/h74bj136",
   "tg_tokens": null,
   "tps": 92.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-33slqquk-ctx1024",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO",
   "notes": "MTP speculative decoding enabled; DFlash2; fp16; oMLX community benchmark trace; peak memory 28.3 GB at this context; duplicate traces of the same session: 11wc68z9; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 196.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "nerkyor:Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 196.1 12.3 28.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/33slqquk",
   "source_url": "https://omlx.ai/benchmarks/performance/33slqquk",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-33slqquk-ctx4096",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO",
   "notes": "MTP speculative decoding enabled; DFlash2; fp16; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: 11wc68z9; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 201.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "nerkyor:Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 201.4 12.1 32.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/33slqquk",
   "source_url": "https://omlx.ai/benchmarks/performance/33slqquk",
   "tg_tokens": null,
   "tps": 12.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-e4fabqfs-ctx1024",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO",
   "notes": "MTP speculative decoding enabled; DFlash2; fp16; oMLX community benchmark trace; peak memory 34.6 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 189.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "nerkyor:Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 189.8 22.0 34.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e4fabqfs",
   "source_url": "https://omlx.ai/benchmarks/performance/e4fabqfs",
   "tg_tokens": null,
   "tps": 22.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-e4fabqfs-ctx4096",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO",
   "notes": "MTP speculative decoding enabled; DFlash2; fp16; oMLX community benchmark trace; peak memory 38.5 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 196.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "nerkyor:Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 196.6 23.5 38.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e4fabqfs",
   "source_url": "https://omlx.ai/benchmarks/performance/e4fabqfs",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-6ro31wja-ctx199680",
   "model": "GLM-5.3",
   "notes": "oMLX community benchmark trace, single run at 195k context; peak memory 177.5 GB; TTFT 109310 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1830.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "omlx-glm53-oq4e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k tokens 1,830 PP tok/s 50.4 TG tok/s 109310 TTFT (ms) 177.5 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ro31wja",
   "source_url": "https://omlx.ai/benchmarks/performance/6ro31wja",
   "tg_tokens": null,
   "tps": 50.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-sjqakai4-ctx1024",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: e4yqjzzg, h57vpbov; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 99.4,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 5bit; 1k 99.4 9.5 22.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sjqakai4",
   "source_url": "https://omlx.ai/benchmarks/performance/sjqakai4",
   "tg_tokens": null,
   "tps": 9.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-sjqakai4-ctx4096",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: e4yqjzzg, h57vpbov; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 101.1,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 5bit; 4k 101.1 8.6 24.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sjqakai4",
   "source_url": "https://omlx.ai/benchmarks/performance/sjqakai4",
   "tg_tokens": null,
   "tps": 8.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-sjqakai4-ctx8192",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: e4yqjzzg, h57vpbov; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 97.9,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 5bit; 8k 97.9 9.2 25.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sjqakai4",
   "source_url": "https://omlx.ai/benchmarks/performance/sjqakai4",
   "tg_tokens": null,
   "tps": 9.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9yg666qp-ctx1024",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 35.5 GB; TTFT 255 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4018.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-oQ8-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k tokens 4,018 PP tok/s 101.1 TG tok/s 255 TTFT (ms) 35.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9yg666qp",
   "source_url": "https://omlx.ai/benchmarks/performance/9yg666qp",
   "tg_tokens": null,
   "tps": 101.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-3k5dxhs4-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: 3x4fire3, 6cf03gdv, e4y01ytf, h9wmqqer, ufdyl654, vgk0kru0, y9siv1zl; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 552.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 1k 552.3 51.1 35.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3k5dxhs4",
   "source_url": "https://omlx.ai/benchmarks/performance/3k5dxhs4",
   "tg_tokens": null,
   "tps": 51.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-3k5dxhs4-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.6 GB at this context; duplicate traces of the same session: 3x4fire3, 6cf03gdv, e4y01ytf, h9wmqqer, ufdyl654, vgk0kru0, y9siv1zl; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 609.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 4k 609.4 48.5 37.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3k5dxhs4",
   "source_url": "https://omlx.ai/benchmarks/performance/3k5dxhs4",
   "tg_tokens": null,
   "tps": 48.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-3k5dxhs4-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 38.1 GB at this context; duplicate traces of the same session: 3x4fire3, 6cf03gdv, e4y01ytf, h9wmqqer, ufdyl654, vgk0kru0, y9siv1zl; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 556.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 8k 556.5 46.4 38.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3k5dxhs4",
   "source_url": "https://omlx.ai/benchmarks/performance/3k5dxhs4",
   "tg_tokens": null,
   "tps": 46.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-3k5dxhs4-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 39.1 GB at this context; duplicate traces of the same session: 3x4fire3, 6cf03gdv, e4y01ytf, h9wmqqer, ufdyl654, vgk0kru0, y9siv1zl; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 473.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 16k 473.6 42.4 39.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3k5dxhs4",
   "source_url": "https://omlx.ai/benchmarks/performance/3k5dxhs4",
   "tg_tokens": null,
   "tps": 42.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-3k5dxhs4-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 39.8 GB at this context; duplicate traces of the same session: 3x4fire3, 6cf03gdv, e4y01ytf, h9wmqqer, ufdyl654, vgk0kru0, y9siv1zl; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 363.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 32k 363.0 36.3 39.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3k5dxhs4",
   "source_url": "https://omlx.ai/benchmarks/performance/3k5dxhs4",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-3k5dxhs4-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 40.3 GB at this context; duplicate traces of the same session: 3x4fire3, 6cf03gdv, e4y01ytf, h9wmqqer, ufdyl654, vgk0kru0, y9siv1zl; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 248.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 64k 248.9 26.9 40.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3k5dxhs4",
   "source_url": "https://omlx.ai/benchmarks/performance/3k5dxhs4",
   "tg_tokens": null,
   "tps": 26.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-3k5dxhs4-ctx131072",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 41.6 GB at this context; duplicate traces of the same session: 3x4fire3, 6cf03gdv, e4y01ytf, h9wmqqer, ufdyl654, vgk0kru0, y9siv1zl; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 153.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 128k 153.3 16.9 41.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3k5dxhs4",
   "source_url": "https://omlx.ai/benchmarks/performance/3k5dxhs4",
   "tg_tokens": null,
   "tps": 16.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-3k5dxhs4-ctx199680",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 42.9 GB at this context; duplicate traces of the same session: 3x4fire3, 6cf03gdv, e4y01ytf, h9wmqqer, ufdyl654, vgk0kru0, y9siv1zl; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 111.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 195k 111.0 12.2 42.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3k5dxhs4",
   "source_url": "https://omlx.ai/benchmarks/performance/3k5dxhs4",
   "tg_tokens": null,
   "tps": 12.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-kci738u0-ctx4096",
   "model": "Ornith-1.5-35B-A3B-uncensored",
   "notes": "oMLX community benchmark trace; peak memory 37.5 GB at this context; duplicate traces of the same session: hzgnp4sw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1605.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-uncensored-MLX-8bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 1,605 76.6 37.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kci738u0",
   "source_url": "https://omlx.ai/benchmarks/performance/kci738u0",
   "tg_tokens": null,
   "tps": 76.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-kci738u0-ctx16384",
   "model": "Ornith-1.5-35B-A3B-uncensored",
   "notes": "oMLX community benchmark trace; peak memory 37.8 GB at this context; duplicate traces of the same session: hzgnp4sw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1525.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-uncensored-MLX-8bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 1,525 70.4 37.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kci738u0",
   "source_url": "https://omlx.ai/benchmarks/performance/kci738u0",
   "tg_tokens": null,
   "tps": 70.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-393zt872-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: u71s6ium; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.5-9B-oQ4-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 224.0 29.6 7.0 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/393zt872",
   "source_url": "https://omlx.ai/benchmarks/performance/393zt872",
   "tg_tokens": null,
   "tps": 29.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-393zt872-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.7 GB at this context; duplicate traces of the same session: u71s6ium; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 221.1,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.5-9B-oQ4-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 221.1 32.0 7.7 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/393zt872",
   "source_url": "https://omlx.ai/benchmarks/performance/393zt872",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-klo47sje-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: wea7oywk; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 234.2,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.5-9B-oQ4-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 234.2 20.8 5.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/klo47sje",
   "source_url": "https://omlx.ai/benchmarks/performance/klo47sje",
   "tg_tokens": null,
   "tps": 20.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-klo47sje-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: wea7oywk; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 234.3,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.5-9B-oQ4-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 234.3 20.2 6.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/klo47sje",
   "source_url": "https://omlx.ai/benchmarks/performance/klo47sje",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-g3jgqs3u-ctx65536",
   "model": "Qwen3.6-35B-A3B-Abliterated-Heretic",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 26.3 GB; TTFT 86539 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 757.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Abliterated-Heretic-MLX-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 4bit; 64k tokens 757.3 PP tok/s 58.8 TG tok/s 86539 TTFT (ms) 26.3 Peak mem",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g3jgqs3u",
   "source_url": "https://omlx.ai/benchmarks/performance/g3jgqs3u",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1cnou7r2-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 39.1 GB at this context; duplicate traces of the same session: ipm3q12c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 725.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 725.7 40.3 39.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1cnou7r2",
   "source_url": "https://omlx.ai/benchmarks/performance/1cnou7r2",
   "tg_tokens": null,
   "tps": 40.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1cnou7r2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 40.2 GB at this context; duplicate traces of the same session: ipm3q12c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 627.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 627.7 50.7 40.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1cnou7r2",
   "source_url": "https://omlx.ai/benchmarks/performance/1cnou7r2",
   "tg_tokens": null,
   "tps": 50.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-gxcpaq4t-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 22.2 GB; TTFT 8149 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.7,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k tokens 125.7 PP tok/s 18.2 TG tok/s 8149 TTFT (ms) 22.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gxcpaq4t",
   "source_url": "https://omlx.ai/benchmarks/performance/gxcpaq4t",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-h9hz1vps-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 25.9 GB; TTFT 108219 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 302.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k tokens 302.8 PP tok/s 98.4 TG tok/s 108219 TTFT (ms) 25.9 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h9hz1vps",
   "source_url": "https://omlx.ai/benchmarks/performance/h9hz1vps",
   "tg_tokens": null,
   "tps": 98.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-u87tvkzd-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 18.5 GB; TTFT 3149 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 325.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k tokens 325.2 PP tok/s 59.5 TG tok/s 3149 TTFT (ms) 18.5 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u87tvkzd",
   "source_url": "https://omlx.ai/benchmarks/performance/u87tvkzd",
   "tg_tokens": null,
   "tps": 59.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-a4xsl1hm-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 25.9 GB; TTFT 108212 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 302.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k tokens 302.8 PP tok/s 98.0 TG tok/s 108212 TTFT (ms) 25.9 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a4xsl1hm",
   "source_url": "https://omlx.ai/benchmarks/performance/a4xsl1hm",
   "tg_tokens": null,
   "tps": 98.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-71e54u3d-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 25.9 GB; TTFT 108217 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 302.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k tokens 302.8 PP tok/s 60.9 TG tok/s 108217 TTFT (ms) 25.9 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/71e54u3d",
   "source_url": "https://omlx.ai/benchmarks/performance/71e54u3d",
   "tg_tokens": null,
   "tps": 60.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5q2fwb9j-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 28.6 GB; TTFT 84720 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 386.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k tokens 386.8 PP tok/s 63.7 TG tok/s 84720 TTFT (ms) 28.6 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5q2fwb9j",
   "source_url": "https://omlx.ai/benchmarks/performance/5q2fwb9j",
   "tg_tokens": null,
   "tps": 63.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-xzog68na-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: 2z3wkztb; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 65.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k 65.5 6.6 16.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xzog68na",
   "source_url": "https://omlx.ai/benchmarks/performance/xzog68na",
   "tg_tokens": null,
   "tps": 6.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-xzog68na-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: 2z3wkztb; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 67.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 67.4 6.5 17.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xzog68na",
   "source_url": "https://omlx.ai/benchmarks/performance/xzog68na",
   "tg_tokens": null,
   "tps": 6.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-s9ujfcyz-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: guq9o3dq; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 212.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-fp16-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 212.3 7.7 17.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s9ujfcyz",
   "source_url": "https://omlx.ai/benchmarks/performance/s9ujfcyz",
   "tg_tokens": null,
   "tps": 7.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-s9ujfcyz-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: guq9o3dq; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 171.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-fp16-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 16k 171.0 7.2 18.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s9ujfcyz",
   "source_url": "https://omlx.ai/benchmarks/performance/s9ujfcyz",
   "tg_tokens": null,
   "tps": 7.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ztv6d38e-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: t2wxa9od; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 810.1,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 810.1 48.2 33.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ztv6d38e",
   "source_url": "https://omlx.ai/benchmarks/performance/ztv6d38e",
   "tg_tokens": null,
   "tps": 48.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ztv6d38e-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: t2wxa9od; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 925.9,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 925.9 43.0 34.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ztv6d38e",
   "source_url": "https://omlx.ai/benchmarks/performance/ztv6d38e",
   "tg_tokens": null,
   "tps": 43.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-hsq0phtd-ctx8192",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: pckhr9o8; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 512.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 8bit; 8k 512.2 30.1 17.1 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hsq0phtd",
   "source_url": "https://omlx.ai/benchmarks/performance/hsq0phtd",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-hsq0phtd-ctx16384",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 17.5 GB at this context; duplicate traces of the same session: pckhr9o8; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 562.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 8bit; 16k 562.6 32.4 17.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hsq0phtd",
   "source_url": "https://omlx.ai/benchmarks/performance/hsq0phtd",
   "tg_tokens": null,
   "tps": 32.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-jg2mxc3j-ctx8192",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 12.4 GB at this context; duplicate traces of the same session: ey1nxydz; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 381.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 8bit; 8k 381.4 19.6 12.4 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jg2mxc3j",
   "source_url": "https://omlx.ai/benchmarks/performance/jg2mxc3j",
   "tg_tokens": null,
   "tps": 19.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-jg2mxc3j-ctx16384",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 12.9 GB at this context; duplicate traces of the same session: ey1nxydz; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 405.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 8bit; 16k 405.3 22.2 12.9 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jg2mxc3j",
   "source_url": "https://omlx.ai/benchmarks/performance/jg2mxc3j",
   "tg_tokens": null,
   "tps": 22.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-vm9prx6n-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: kvo2rfd9; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 538.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 538.7 40.5 4.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vm9prx6n",
   "source_url": "https://omlx.ai/benchmarks/performance/vm9prx6n",
   "tg_tokens": null,
   "tps": 40.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-vm9prx6n-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: kvo2rfd9; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 516.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 516.7 36.5 4.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vm9prx6n",
   "source_url": "https://omlx.ai/benchmarks/performance/vm9prx6n",
   "tg_tokens": null,
   "tps": 36.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-d1x18ysz-ctx4096",
   "model": "Qwen3.6-27B",
   "notes": "ThinkingCap build; mlx-community; oMLX community benchmark trace, single run at 4k context; peak memory 22.5 GB; TTFT 11743 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 348.8,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "mlx-community--ThinkingCap-Qwen3.6-27B-OptiQ-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k tokens 348.8 PP tok/s 32.8 TG tok/s 11743 TTFT (ms) 22.5 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d1x18ysz",
   "source_url": "https://omlx.ai/benchmarks/performance/d1x18ysz",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-wl4jfddc-ctx1024",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: z4gsr2iu; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 99.3,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 5bit; 1k 99.3 8.6 22.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wl4jfddc",
   "source_url": "https://omlx.ai/benchmarks/performance/wl4jfddc",
   "tg_tokens": null,
   "tps": 8.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-wl4jfddc-ctx4096",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: z4gsr2iu; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.1,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 5bit; 4k 104.1 8.3 24.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wl4jfddc",
   "source_url": "https://omlx.ai/benchmarks/performance/wl4jfddc",
   "tg_tokens": null,
   "tps": 8.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-wl4jfddc-ctx8192",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: z4gsr2iu; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.5,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 5bit; 8k 102.5 9.8 25.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wl4jfddc",
   "source_url": "https://omlx.ai/benchmarks/performance/wl4jfddc",
   "tg_tokens": null,
   "tps": 9.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-6fv6iko9-ctx1024",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 80.0 GB at this context; duplicate traces of the same session: 2jyk141q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 290.4,
   "provenance": "community",
   "quant": "2.4bit mixed",
   "quote": "DeepSeek-V4-Flash-0731-2.4bit-mixed on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 2bit; 1k 290.4 34.0 80.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6fv6iko9",
   "source_url": "https://omlx.ai/benchmarks/performance/6fv6iko9",
   "tg_tokens": null,
   "tps": 34.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-6fv6iko9-ctx4096",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 82.1 GB at this context; duplicate traces of the same session: 2jyk141q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 268.1,
   "provenance": "community",
   "quant": "2.4bit mixed",
   "quote": "DeepSeek-V4-Flash-0731-2.4bit-mixed on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 2bit; 4k 268.1 30.8 82.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6fv6iko9",
   "source_url": "https://omlx.ai/benchmarks/performance/6fv6iko9",
   "tg_tokens": null,
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-gr8lxfsh-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.2 GB at this context; duplicate traces of the same session: 8kzs1r2v, hkn5mmfp; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2339.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,339 128.3 173.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gr8lxfsh",
   "source_url": "https://omlx.ai/benchmarks/performance/gr8lxfsh",
   "tg_tokens": null,
   "tps": 128.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-gr8lxfsh-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.7 GB at this context; duplicate traces of the same session: 8kzs1r2v, hkn5mmfp; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2315.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,315 86.9 173.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gr8lxfsh",
   "source_url": "https://omlx.ai/benchmarks/performance/gr8lxfsh",
   "tg_tokens": null,
   "tps": 86.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-gr8lxfsh-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.9 GB at this context; duplicate traces of the same session: 8kzs1r2v, hkn5mmfp; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2258.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,258 125.8 174.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gr8lxfsh",
   "source_url": "https://omlx.ai/benchmarks/performance/gr8lxfsh",
   "tg_tokens": null,
   "tps": 125.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-w40ewce3-ctx1024",
   "model": "Ling-3.0-tiny",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 5.2 GB; TTFT 563 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1819.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-tiny-oQ4e on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k tokens 1,819 PP tok/s 134.1 TG tok/s 563 TTFT (ms) 5.2 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w40ewce3",
   "source_url": "https://omlx.ai/benchmarks/performance/w40ewce3",
   "tg_tokens": null,
   "tps": 134.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-g9bzcscw-ctx1024",
   "model": "Mellum-4b-base",
   "notes": "oMLX community benchmark trace; peak memory 3.0 GB at this context; duplicate traces of the same session: py4u5pib; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2657.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Mellum-4b-base-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 2,657 89.1 3.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g9bzcscw",
   "source_url": "https://omlx.ai/benchmarks/performance/g9bzcscw",
   "tg_tokens": null,
   "tps": 89.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-g9bzcscw-ctx4096",
   "model": "Mellum-4b-base",
   "notes": "oMLX community benchmark trace; peak memory 4.1 GB at this context; duplicate traces of the same session: py4u5pib; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2501.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Mellum-4b-base-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,501 59.1 4.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g9bzcscw",
   "source_url": "https://omlx.ai/benchmarks/performance/g9bzcscw",
   "tg_tokens": null,
   "tps": 59.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-kjeiph4v-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: m8w8olch; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 794.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 794.8 72.1 5.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kjeiph4v",
   "source_url": "https://omlx.ai/benchmarks/performance/kjeiph4v",
   "tg_tokens": null,
   "tps": 72.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-kjeiph4v-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.2 GB at this context; duplicate traces of the same session: m8w8olch; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 849.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 849.5 70.0 7.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kjeiph4v",
   "source_url": "https://omlx.ai/benchmarks/performance/kjeiph4v",
   "tg_tokens": null,
   "tps": 70.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-029bd7at-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: teaaiys3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1558.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,558 106.5 19.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/029bd7at",
   "source_url": "https://omlx.ai/benchmarks/performance/029bd7at",
   "tg_tokens": null,
   "tps": 106.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-029bd7at-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: teaaiys3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1842.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,842 105.8 21.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/029bd7at",
   "source_url": "https://omlx.ai/benchmarks/performance/029bd7at",
   "tg_tokens": null,
   "tps": 105.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-imwq2usy-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: bd0p7szi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1562.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,562 107.4 19.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/imwq2usy",
   "source_url": "https://omlx.ai/benchmarks/performance/imwq2usy",
   "tg_tokens": null,
   "tps": 107.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-imwq2usy-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: bd0p7szi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1845.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,845 106.5 21.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/imwq2usy",
   "source_url": "https://omlx.ai/benchmarks/performance/imwq2usy",
   "tg_tokens": null,
   "tps": 106.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-nj18nhpc-ctx65536",
   "model": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed",
   "notes": "oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: 7jed70v5; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 761.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 4bit; 64k 761.4 66.5 22.4 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nj18nhpc",
   "source_url": "https://omlx.ai/benchmarks/performance/nj18nhpc",
   "tg_tokens": null,
   "tps": 66.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-nj18nhpc-ctx131072",
   "model": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed",
   "notes": "oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 7jed70v5; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 477.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 4bit; 128k 477.9 51.0 23.6 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nj18nhpc",
   "source_url": "https://omlx.ai/benchmarks/performance/nj18nhpc",
   "tg_tokens": null,
   "tps": 51.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-x8re9vqn-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 7kdjakpw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 771.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 771.2 19.0 28.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x8re9vqn",
   "source_url": "https://omlx.ai/benchmarks/performance/x8re9vqn",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-x8re9vqn-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: 7kdjakpw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 918.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 918.8 18.9 29.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x8re9vqn",
   "source_url": "https://omlx.ai/benchmarks/performance/x8re9vqn",
   "tg_tokens": null,
   "tps": 18.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-rfxvc7aq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 6obpq627; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 225.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 225.5 23.1 16.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rfxvc7aq",
   "source_url": "https://omlx.ai/benchmarks/performance/rfxvc7aq",
   "tg_tokens": null,
   "tps": 23.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-rfxvc7aq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 6obpq627; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 235.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 235.1 22.2 19.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rfxvc7aq",
   "source_url": "https://omlx.ai/benchmarks/performance/rfxvc7aq",
   "tg_tokens": null,
   "tps": 22.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-7u5j4m98-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: jbnt335z; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 110.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 110.8 19.5 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7u5j4m98",
   "source_url": "https://omlx.ai/benchmarks/performance/7u5j4m98",
   "tg_tokens": null,
   "tps": 19.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-7u5j4m98-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: jbnt335z; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 103.6 18.9 19.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7u5j4m98",
   "source_url": "https://omlx.ai/benchmarks/performance/7u5j4m98",
   "tg_tokens": null,
   "tps": 18.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-6xiwtghx-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: ocqsribu; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 325.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 325.2 71.1 21.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6xiwtghx",
   "source_url": "https://omlx.ai/benchmarks/performance/6xiwtghx",
   "tg_tokens": null,
   "tps": 71.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-6xiwtghx-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: ocqsribu; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 423.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 423.2 98.6 24.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6xiwtghx",
   "source_url": "https://omlx.ai/benchmarks/performance/6xiwtghx",
   "tg_tokens": null,
   "tps": 98.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-wxsztud1-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: jk65st8o; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 325.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 325.1 72.6 21.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wxsztud1",
   "source_url": "https://omlx.ai/benchmarks/performance/wxsztud1",
   "tg_tokens": null,
   "tps": 72.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-wxsztud1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: jk65st8o; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 423.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 423.6 91.5 24.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wxsztud1",
   "source_url": "https://omlx.ai/benchmarks/performance/wxsztud1",
   "tg_tokens": null,
   "tps": 91.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-mvn7kvxf-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 1d62ded6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 325.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 325.1 59.0 21.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mvn7kvxf",
   "source_url": "https://omlx.ai/benchmarks/performance/mvn7kvxf",
   "tg_tokens": null,
   "tps": 59.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-mvn7kvxf-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: 1d62ded6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 424.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 424.0 102.2 24.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mvn7kvxf",
   "source_url": "https://omlx.ai/benchmarks/performance/mvn7kvxf",
   "tg_tokens": null,
   "tps": 102.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-b7bra8i4-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: ob5vq8aq; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 64.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 64.7 14.0 19.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b7bra8i4",
   "source_url": "https://omlx.ai/benchmarks/performance/b7bra8i4",
   "tg_tokens": null,
   "tps": 14.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-b7bra8i4-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: ob5vq8aq; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 60.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 16k 60.5 17.0 21.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b7bra8i4",
   "source_url": "https://omlx.ai/benchmarks/performance/b7bra8i4",
   "tg_tokens": null,
   "tps": 17.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-of1687je-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 1iqe4cmk, blbr941f, d2u7kwcm, p26kkwd4; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 126.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 126.8 24.6 20.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/of1687je",
   "source_url": "https://omlx.ai/benchmarks/performance/of1687je",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-of1687je-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: 1iqe4cmk, blbr941f, d2u7kwcm, p26kkwd4; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 181.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 181.5 27.5 24.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/of1687je",
   "source_url": "https://omlx.ai/benchmarks/performance/of1687je",
   "tg_tokens": null,
   "tps": 27.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-of1687je-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: 1iqe4cmk, blbr941f, d2u7kwcm, p26kkwd4; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 184.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 184.8 33.2 25.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/of1687je",
   "source_url": "https://omlx.ai/benchmarks/performance/of1687je",
   "tg_tokens": null,
   "tps": 33.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-of1687je-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: 1iqe4cmk, blbr941f, d2u7kwcm, p26kkwd4; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 124.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 124.5 34.2 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/of1687je",
   "source_url": "https://omlx.ai/benchmarks/performance/of1687je",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-of1687je-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: 1iqe4cmk, blbr941f, d2u7kwcm, p26kkwd4; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 79.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 128k 79.5 21.6 32.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/of1687je",
   "source_url": "https://omlx.ai/benchmarks/performance/of1687je",
   "tg_tokens": null,
   "tps": 21.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-zmx50wzh-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: 1cutsqmq; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 255.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 255.8 19.1 17.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zmx50wzh",
   "source_url": "https://omlx.ai/benchmarks/performance/zmx50wzh",
   "tg_tokens": null,
   "tps": 19.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-zmx50wzh-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: 1cutsqmq; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 205.1 19.3 18.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zmx50wzh",
   "source_url": "https://omlx.ai/benchmarks/performance/zmx50wzh",
   "tg_tokens": null,
   "tps": 19.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-edcophne-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 75.3 GB at this context; duplicate traces of the same session: 42oyt0vg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 457.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 457.0 32.4 75.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/edcophne",
   "source_url": "https://omlx.ai/benchmarks/performance/edcophne",
   "tg_tokens": null,
   "tps": 32.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-edcophne-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 75.8 GB at this context; duplicate traces of the same session: 42oyt0vg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 472.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 472.5 31.7 75.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/edcophne",
   "source_url": "https://omlx.ai/benchmarks/performance/edcophne",
   "tg_tokens": null,
   "tps": 31.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-yq1tsrl8-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 32.8 GB; TTFT 16728 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 61.2,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Qwen3.8-Flash-Next-MLX-2bit-MTP on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 2bit; 1k tokens 61.2 PP tok/s 3.9 TG tok/s 16728 TTFT (ms) 32.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yq1tsrl8",
   "source_url": "https://omlx.ai/benchmarks/performance/yq1tsrl8",
   "tg_tokens": null,
   "tps": 3.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-xvjs3s4k-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 590.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 590.1 65.5 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xvjs3s4k",
   "source_url": "https://omlx.ai/benchmarks/performance/xvjs3s4k",
   "tg_tokens": null,
   "tps": 65.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-xvjs3s4k-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 698.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 698.6 52.3 71.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xvjs3s4k",
   "source_url": "https://omlx.ai/benchmarks/performance/xvjs3s4k",
   "tg_tokens": null,
   "tps": 52.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-lpl7ahcb-ctx1024",
   "model": "ThinkingCap-Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: r47lwjkx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 249.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "ThinkingCap-Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 249.5 37.7 33.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lpl7ahcb",
   "source_url": "https://omlx.ai/benchmarks/performance/lpl7ahcb",
   "tg_tokens": null,
   "tps": 37.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-lpl7ahcb-ctx4096",
   "model": "ThinkingCap-Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: r47lwjkx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 254.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "ThinkingCap-Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 254.9 31.9 37.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lpl7ahcb",
   "source_url": "https://omlx.ai/benchmarks/performance/lpl7ahcb",
   "tg_tokens": null,
   "tps": 31.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-wk4hxexp-ctx1024",
   "model": "gemma-4-26B-A4B-it-OptiQ-4bit-REAP-14B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 10.2 GB; TTFT 1511 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 677.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-OptiQ-4bit-REAP-14B on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k tokens 677.7 PP tok/s 53.2 TG tok/s 1511 TTFT (ms) 10.2 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wk4hxexp",
   "source_url": "https://omlx.ai/benchmarks/performance/wk4hxexp",
   "tg_tokens": null,
   "tps": 53.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-uhq929d9-ctx1024",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 99.3,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 5bit; 1k 99.3 8.6 22.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uhq929d9",
   "source_url": "https://omlx.ai/benchmarks/performance/uhq929d9",
   "tg_tokens": null,
   "tps": 8.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-uhq929d9-ctx4096",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.1,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 5bit; 4k 104.1 8.3 24.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uhq929d9",
   "source_url": "https://omlx.ai/benchmarks/performance/uhq929d9",
   "tg_tokens": null,
   "tps": 8.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-uhq929d9-ctx8192",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.5,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 5bit; 8k 102.5 9.8 25.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uhq929d9",
   "source_url": "https://omlx.ai/benchmarks/performance/uhq929d9",
   "tg_tokens": null,
   "tps": 9.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-3uyfel70-ctx4096",
   "model": "AREX-2",
   "notes": "fp16; oMLX community benchmark trace; peak memory 32.3 GB at this context; duplicate traces of the same session: l3rifz7f; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 162.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "AREX-2-oQ8e-fp16 on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 4k 162.8 12.0 32.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3uyfel70",
   "source_url": "https://omlx.ai/benchmarks/performance/3uyfel70",
   "tg_tokens": null,
   "tps": 12.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-3uyfel70-ctx16384",
   "model": "AREX-2",
   "notes": "fp16; oMLX community benchmark trace; peak memory 34.6 GB at this context; duplicate traces of the same session: l3rifz7f; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 152.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "AREX-2-oQ8e-fp16 on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 16k 152.3 11.5 34.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3uyfel70",
   "source_url": "https://omlx.ai/benchmarks/performance/3uyfel70",
   "tg_tokens": null,
   "tps": 11.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-wo61shxn-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.0 GB at this context; duplicate traces of the same session: o59dlv3j, x3ze0din; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1960.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,960 99.4 175.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wo61shxn",
   "source_url": "https://omlx.ai/benchmarks/performance/wo61shxn",
   "tg_tokens": null,
   "tps": 99.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-wo61shxn-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.8 GB at this context; duplicate traces of the same session: o59dlv3j, x3ze0din; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1958.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,958 67.4 177.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wo61shxn",
   "source_url": "https://omlx.ai/benchmarks/performance/wo61shxn",
   "tg_tokens": null,
   "tps": 67.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-wo61shxn-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 183.7 GB at this context; duplicate traces of the same session: o59dlv3j, x3ze0din; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1926.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 1,926 95.5 183.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wo61shxn",
   "source_url": "https://omlx.ai/benchmarks/performance/wo61shxn",
   "tg_tokens": null,
   "tps": 95.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-9w1euw3r-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 170.7 GB at this context; duplicate traces of the same session: bjjg4pyb, vi1uxcuh; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1563.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,563 53.2 170.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9w1euw3r",
   "source_url": "https://omlx.ai/benchmarks/performance/9w1euw3r",
   "tg_tokens": null,
   "tps": 53.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-9w1euw3r-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 172.0 GB at this context; duplicate traces of the same session: bjjg4pyb, vi1uxcuh; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2023.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 2,023 47.4 172.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9w1euw3r",
   "source_url": "https://omlx.ai/benchmarks/performance/9w1euw3r",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-9w1euw3r-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 183.7 GB at this context; duplicate traces of the same session: bjjg4pyb, vi1uxcuh; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1915.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 1,915 99.8 183.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9w1euw3r",
   "source_url": "https://omlx.ai/benchmarks/performance/9w1euw3r",
   "tg_tokens": null,
   "tps": 99.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-60ao5var-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 170.7 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1337.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,337 49.7 170.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/60ao5var",
   "source_url": "https://omlx.ai/benchmarks/performance/60ao5var",
   "tg_tokens": null,
   "tps": 49.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-60ao5var-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 172.0 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2039.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 2,039 47.3 172.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/60ao5var",
   "source_url": "https://omlx.ai/benchmarks/performance/60ao5var",
   "tg_tokens": null,
   "tps": 47.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-oyv5wb7u-ctx1024",
   "model": "muse-glimmer 30B",
   "notes": "OptiQ; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 2fizo0ee, 3pkjv7ew, j3q3tka8, kkcmd70p, minq5xlb, qksx3f9q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 254.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-30B-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 254.2 23.2 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oyv5wb7u",
   "source_url": "https://omlx.ai/benchmarks/performance/oyv5wb7u",
   "tg_tokens": null,
   "tps": 23.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-oyv5wb7u-ctx4096",
   "model": "muse-glimmer 30B",
   "notes": "OptiQ; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: 2fizo0ee, 3pkjv7ew, j3q3tka8, kkcmd70p, minq5xlb, qksx3f9q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 215.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-30B-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 215.8 21.6 19.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oyv5wb7u",
   "source_url": "https://omlx.ai/benchmarks/performance/oyv5wb7u",
   "tg_tokens": null,
   "tps": 21.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-oyv5wb7u-ctx8192",
   "model": "muse-glimmer 30B",
   "notes": "OptiQ; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 2fizo0ee, 3pkjv7ew, j3q3tka8, kkcmd70p, minq5xlb, qksx3f9q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 201.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-30B-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 201.3 22.2 19.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oyv5wb7u",
   "source_url": "https://omlx.ai/benchmarks/performance/oyv5wb7u",
   "tg_tokens": null,
   "tps": 22.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-oyv5wb7u-ctx16384",
   "model": "muse-glimmer 30B",
   "notes": "OptiQ; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 2fizo0ee, 3pkjv7ew, j3q3tka8, kkcmd70p, minq5xlb, qksx3f9q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 200.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-30B-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 200.8 21.2 19.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oyv5wb7u",
   "source_url": "https://omlx.ai/benchmarks/performance/oyv5wb7u",
   "tg_tokens": null,
   "tps": 21.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-oyv5wb7u-ctx32768",
   "model": "muse-glimmer 30B",
   "notes": "OptiQ; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 2fizo0ee, 3pkjv7ew, j3q3tka8, kkcmd70p, minq5xlb, qksx3f9q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 188.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-30B-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 188.5 19.1 19.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oyv5wb7u",
   "source_url": "https://omlx.ai/benchmarks/performance/oyv5wb7u",
   "tg_tokens": null,
   "tps": 19.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-oyv5wb7u-ctx65536",
   "model": "muse-glimmer 30B",
   "notes": "OptiQ; oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: 2fizo0ee, 3pkjv7ew, j3q3tka8, kkcmd70p, minq5xlb, qksx3f9q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 170.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-30B-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 170.3 17.2 20.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oyv5wb7u",
   "source_url": "https://omlx.ai/benchmarks/performance/oyv5wb7u",
   "tg_tokens": null,
   "tps": 17.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-oyv5wb7u-ctx131072",
   "model": "muse-glimmer 30B",
   "notes": "OptiQ; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 2fizo0ee, 3pkjv7ew, j3q3tka8, kkcmd70p, minq5xlb, qksx3f9q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 150.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-30B-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 150.5 15.2 21.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oyv5wb7u",
   "source_url": "https://omlx.ai/benchmarks/performance/oyv5wb7u",
   "tg_tokens": null,
   "tps": 15.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-svgf8mlm-ctx1024",
   "model": "NeoHorse-1-4B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 3.9 GB; TTFT 1487 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 688.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NeoHorse-1-4B-oQ4e on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k tokens 688.8 PP tok/s 75.9 TG tok/s 1487 TTFT (ms) 3.9 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/svgf8mlm",
   "source_url": "https://omlx.ai/benchmarks/performance/svgf8mlm",
   "tg_tokens": null,
   "tps": 75.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-vn0jrcln-ctx1024",
   "model": "Qwen2.5-Coder-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 4.8 GB at this context; duplicate traces of the same session: srnhvg17; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 297.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-7B-Instruct-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 297.5 38.9 4.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vn0jrcln",
   "source_url": "https://omlx.ai/benchmarks/performance/vn0jrcln",
   "tg_tokens": null,
   "tps": 38.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-vn0jrcln-ctx4096",
   "model": "Qwen2.5-Coder-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: srnhvg17; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 284.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-7B-Instruct-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 284.6 35.6 4.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vn0jrcln",
   "source_url": "https://omlx.ai/benchmarks/performance/vn0jrcln",
   "tg_tokens": null,
   "tps": 35.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-8z0j63k3-ctx1024",
   "model": "Qwen3.5-9B-Claude-4.6-HighIQ-INSTRUCT-HERETIC-UNCENSORED",
   "notes": "oMLX community benchmark trace; peak memory 9.7 GB at this context; duplicate traces of the same session: 9p1hxgwk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1508.0,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Qwen3.5-9B-Claude-4.6-HighIQ-INSTRUCT-HERETIC-UNCENSORED-MLX-mxfp8 on M5 Pro (20c); M5 Pro (20c) \u00b7 24 GB \u00b7 8bit; 1k 1,508 32.9 9.7 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8z0j63k3",
   "source_url": "https://omlx.ai/benchmarks/performance/8z0j63k3",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-8z0j63k3-ctx4096",
   "model": "Qwen3.5-9B-Claude-4.6-HighIQ-INSTRUCT-HERETIC-UNCENSORED",
   "notes": "oMLX community benchmark trace; peak memory 10.2 GB at this context; duplicate traces of the same session: 9p1hxgwk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1624.0,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Qwen3.5-9B-Claude-4.6-HighIQ-INSTRUCT-HERETIC-UNCENSORED-MLX-mxfp8 on M5 Pro (20c); M5 Pro (20c) \u00b7 24 GB \u00b7 8bit; 4k 1,624 32.3 10.2 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8z0j63k3",
   "source_url": "https://omlx.ai/benchmarks/performance/8z0j63k3",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-xuvg7q5c-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: 2ibk3m8f; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 382.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-9B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 1k 382.2 56.7 5.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xuvg7q5c",
   "source_url": "https://omlx.ai/benchmarks/performance/xuvg7q5c",
   "tg_tokens": null,
   "tps": 56.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-xuvg7q5c-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 6.6 GB at this context; duplicate traces of the same session: 2ibk3m8f; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 397.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-9B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 4k 397.3 55.2 6.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xuvg7q5c",
   "source_url": "https://omlx.ai/benchmarks/performance/xuvg7q5c",
   "tg_tokens": null,
   "tps": 55.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-gyjxf87v-ctx65536",
   "model": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed",
   "notes": "FP16; oMLX community benchmark trace, single run at 64k context; peak memory 22.3 GB; TTFT 54754 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1197.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed-FP16 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 4bit; 64k tokens 1,197 PP tok/s 71.6 TG tok/s 54754 TTFT (ms) 22.3 Peak mem",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gyjxf87v",
   "source_url": "https://omlx.ai/benchmarks/performance/gyjxf87v",
   "tg_tokens": null,
   "tps": 71.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-9lp921la-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 8468 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 120.9,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k tokens 120.9 PP tok/s 23.4 TG tok/s 8468 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9lp921la",
   "source_url": "https://omlx.ai/benchmarks/performance/9lp921la",
   "tg_tokens": null,
   "tps": 23.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-1ov1z251-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 8527 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 120.1,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k tokens 120.1 PP tok/s 17.8 TG tok/s 8527 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ov1z251",
   "source_url": "https://omlx.ai/benchmarks/performance/1ov1z251",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-cmkrlv5y-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 22.2 GB; TTFT 8484 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 120.7,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k tokens 120.7 PP tok/s 16.5 TG tok/s 8484 TTFT (ms) 22.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cmkrlv5y",
   "source_url": "https://omlx.ai/benchmarks/performance/cmkrlv5y",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-elo8s4j0-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 8391 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 122.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k tokens 122.0 PP tok/s 20.4 TG tok/s 8391 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/elo8s4j0",
   "source_url": "https://omlx.ai/benchmarks/performance/elo8s4j0",
   "tg_tokens": null,
   "tps": 20.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ri73jb2t-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: qsz0d4tq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 289.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 289.3 17.5 17.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ri73jb2t",
   "source_url": "https://omlx.ai/benchmarks/performance/ri73jb2t",
   "tg_tokens": null,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ri73jb2t-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: qsz0d4tq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 390.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 390.5 15.6 19.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ri73jb2t",
   "source_url": "https://omlx.ai/benchmarks/performance/ri73jb2t",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-z6q69ykl-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: tllauadp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 748.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 748.5 38.6 33.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z6q69ykl",
   "source_url": "https://omlx.ai/benchmarks/performance/z6q69ykl",
   "tg_tokens": null,
   "tps": 38.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-z6q69ykl-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: tllauadp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 925.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 925.1 44.4 34.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z6q69ykl",
   "source_url": "https://omlx.ai/benchmarks/performance/z6q69ykl",
   "tg_tokens": null,
   "tps": 44.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-86zx2zcs-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: fvgz14iw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 767.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 767.9 18.8 28.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/86zx2zcs",
   "source_url": "https://omlx.ai/benchmarks/performance/86zx2zcs",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-86zx2zcs-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: fvgz14iw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 898.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 898.9 18.6 29.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/86zx2zcs",
   "source_url": "https://omlx.ai/benchmarks/performance/86zx2zcs",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-b75j26xc-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 122.3 GB at this context; duplicate traces of the same session: kq5bdzvx; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 823.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 1k 823.1 64.9 122.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b75j26xc",
   "source_url": "https://omlx.ai/benchmarks/performance/b75j26xc",
   "tg_tokens": null,
   "tps": 64.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-b75j26xc-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 122.8 GB at this context; duplicate traces of the same session: kq5bdzvx; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1041.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 4k 1,041 80.9 122.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b75j26xc",
   "source_url": "https://omlx.ai/benchmarks/performance/b75j26xc",
   "tg_tokens": null,
   "tps": 80.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-t911z9gy-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: xktstih3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 533.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 533.4 44.2 70.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t911z9gy",
   "source_url": "https://omlx.ai/benchmarks/performance/t911z9gy",
   "tg_tokens": null,
   "tps": 44.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-t911z9gy-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: xktstih3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 646.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 646.5 45.0 71.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t911z9gy",
   "source_url": "https://omlx.ai/benchmarks/performance/t911z9gy",
   "tg_tokens": null,
   "tps": 45.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-7by0qadq-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 590.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 590.1 65.5 70.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7by0qadq",
   "source_url": "https://omlx.ai/benchmarks/performance/7by0qadq",
   "tg_tokens": null,
   "tps": 65.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-7by0qadq-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 698.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 698.6 52.3 71.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7by0qadq",
   "source_url": "https://omlx.ai/benchmarks/performance/7by0qadq",
   "tg_tokens": null,
   "tps": 52.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-fmdgfq5h-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 21.2 GB; TTFT 87544 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 187.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k tokens 187.2 PP tok/s 16.3 TG tok/s 87544 TTFT (ms) 21.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fmdgfq5h",
   "source_url": "https://omlx.ai/benchmarks/performance/fmdgfq5h",
   "tg_tokens": null,
   "tps": 16.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-c925d85e-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace, single run at 1k context; peak memory 70.8 GB; TTFT 1972 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 519.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 519.3 PP tok/s 52.9 TG tok/s 1972 TTFT (ms) 70.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c925d85e",
   "source_url": "https://omlx.ai/benchmarks/performance/c925d85e",
   "tg_tokens": null,
   "tps": 52.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-fkly82yd-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: u69rb3jw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 523.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 523.1 59.3 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fkly82yd",
   "source_url": "https://omlx.ai/benchmarks/performance/fkly82yd",
   "tg_tokens": null,
   "tps": 59.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-fkly82yd-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: u69rb3jw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 668.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 668.5 51.9 71.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fkly82yd",
   "source_url": "https://omlx.ai/benchmarks/performance/fkly82yd",
   "tg_tokens": null,
   "tps": 51.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-2dgvxyfk-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: iqc3eh0k; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 641.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 1k 641.8 53.9 14.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2dgvxyfk",
   "source_url": "https://omlx.ai/benchmarks/performance/2dgvxyfk",
   "tg_tokens": null,
   "tps": 53.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-2dgvxyfk-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: iqc3eh0k; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 656.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 4k 656.9 51.2 14.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2dgvxyfk",
   "source_url": "https://omlx.ai/benchmarks/performance/2dgvxyfk",
   "tg_tokens": null,
   "tps": 51.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-u75ivaa3-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 170.7 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1337.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,337 49.7 170.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u75ivaa3",
   "source_url": "https://omlx.ai/benchmarks/performance/u75ivaa3",
   "tg_tokens": null,
   "tps": 49.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-u75ivaa3-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 172.0 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2039.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 2,039 47.3 172.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u75ivaa3",
   "source_url": "https://omlx.ai/benchmarks/performance/u75ivaa3",
   "tg_tokens": null,
   "tps": 47.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-oyyo1wxm-ctx1024",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: r2v0fm4g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1253.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,253 106.0 22.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oyyo1wxm",
   "source_url": "https://omlx.ai/benchmarks/performance/oyyo1wxm",
   "tg_tokens": null,
   "tps": 106.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-oyyo1wxm-ctx4096",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: r2v0fm4g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1491.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,491 122.3 23.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oyyo1wxm",
   "source_url": "https://omlx.ai/benchmarks/performance/oyyo1wxm",
   "tg_tokens": null,
   "tps": 122.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-pgpckj30-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 8ssrb5wn, dik1whhm, l1ruw6ys, txyfuqfq, wpo5in2g, yrsx0j50, zmspy0br; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 528.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 528.8 61.1 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pgpckj30",
   "source_url": "https://omlx.ai/benchmarks/performance/pgpckj30",
   "tg_tokens": null,
   "tps": 61.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-pgpckj30-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 8ssrb5wn, dik1whhm, l1ruw6ys, txyfuqfq, wpo5in2g, yrsx0j50, zmspy0br; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 583.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 583.8 59.6 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pgpckj30",
   "source_url": "https://omlx.ai/benchmarks/performance/pgpckj30",
   "tg_tokens": null,
   "tps": 59.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-pgpckj30-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 8ssrb5wn, dik1whhm, l1ruw6ys, txyfuqfq, wpo5in2g, yrsx0j50, zmspy0br; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 539.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 8k 539.5 55.6 22.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pgpckj30",
   "source_url": "https://omlx.ai/benchmarks/performance/pgpckj30",
   "tg_tokens": null,
   "tps": 55.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-pgpckj30-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: 8ssrb5wn, dik1whhm, l1ruw6ys, txyfuqfq, wpo5in2g, yrsx0j50, zmspy0br; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 461.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k 461.9 49.6 23.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pgpckj30",
   "source_url": "https://omlx.ai/benchmarks/performance/pgpckj30",
   "tg_tokens": null,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-pgpckj30-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 8ssrb5wn, dik1whhm, l1ruw6ys, txyfuqfq, wpo5in2g, yrsx0j50, zmspy0br; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 356.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 32k 356.1 42.4 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pgpckj30",
   "source_url": "https://omlx.ai/benchmarks/performance/pgpckj30",
   "tg_tokens": null,
   "tps": 42.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-pgpckj30-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: 8ssrb5wn, dik1whhm, l1ruw6ys, txyfuqfq, wpo5in2g, yrsx0j50, zmspy0br; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 245.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 64k 245.4 31.2 24.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pgpckj30",
   "source_url": "https://omlx.ai/benchmarks/performance/pgpckj30",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-pgpckj30-ctx131072",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: 8ssrb5wn, dik1whhm, l1ruw6ys, txyfuqfq, wpo5in2g, yrsx0j50, zmspy0br; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 153.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 128k 153.2 19.5 25.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pgpckj30",
   "source_url": "https://omlx.ai/benchmarks/performance/pgpckj30",
   "tg_tokens": null,
   "tps": 19.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-pgpckj30-ctx199680",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: 8ssrb5wn, dik1whhm, l1ruw6ys, txyfuqfq, wpo5in2g, yrsx0j50, zmspy0br; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 109.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 195k 109.5 14.6 26.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pgpckj30",
   "source_url": "https://omlx.ai/benchmarks/performance/pgpckj30",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-lyb29m5y-ctx1024",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 4ky6qtry; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 880.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 880.9 92.7 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lyb29m5y",
   "source_url": "https://omlx.ai/benchmarks/performance/lyb29m5y",
   "tg_tokens": null,
   "tps": 92.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-lyb29m5y-ctx4096",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 4ky6qtry; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1059.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,059 87.7 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lyb29m5y",
   "source_url": "https://omlx.ai/benchmarks/performance/lyb29m5y",
   "tg_tokens": null,
   "tps": 87.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-6m8rgime-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: o5puzkz7; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 361.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 361.0 55.8 3.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6m8rgime",
   "source_url": "https://omlx.ai/benchmarks/performance/6m8rgime",
   "tg_tokens": null,
   "tps": 55.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-6m8rgime-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.0 GB at this context; duplicate traces of the same session: o5puzkz7; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 341.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 341.5 53.2 4.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6m8rgime",
   "source_url": "https://omlx.ai/benchmarks/performance/6m8rgime",
   "tg_tokens": null,
   "tps": 53.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-0lc6un8f-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace, single run at 8k context; peak memory 4.0 GB; TTFT 19246 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 425.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 8k tokens 425.6 PP tok/s 34.5 TG tok/s 19246 TTFT (ms) 4.0 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0lc6un8f",
   "source_url": "https://omlx.ai/benchmarks/performance/0lc6un8f",
   "tg_tokens": null,
   "tps": 34.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-sa9n12u5-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: 6kmekgpd, ntzovrtj; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 426.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 426.3 37.6 3.4 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sa9n12u5",
   "source_url": "https://omlx.ai/benchmarks/performance/sa9n12u5",
   "tg_tokens": null,
   "tps": 37.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-sa9n12u5-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.8 GB at this context; duplicate traces of the same session: 6kmekgpd, ntzovrtj; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 432.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 432.9 36.3 3.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sa9n12u5",
   "source_url": "https://omlx.ai/benchmarks/performance/sa9n12u5",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-sa9n12u5-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.0 GB at this context; duplicate traces of the same session: 6kmekgpd, ntzovrtj; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 422.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 8k 422.8 32.9 4.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sa9n12u5",
   "source_url": "https://omlx.ai/benchmarks/performance/sa9n12u5",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-a0dt7lxk-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: b1grvu6s, vj2z19ig, zklx1cuw; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 429.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 429.0 37.4 3.4 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a0dt7lxk",
   "source_url": "https://omlx.ai/benchmarks/performance/a0dt7lxk",
   "tg_tokens": null,
   "tps": 37.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-a0dt7lxk-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.8 GB at this context; duplicate traces of the same session: b1grvu6s, vj2z19ig, zklx1cuw; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 437.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 437.5 36.3 3.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a0dt7lxk",
   "source_url": "https://omlx.ai/benchmarks/performance/a0dt7lxk",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-a0dt7lxk-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.0 GB at this context; duplicate traces of the same session: b1grvu6s, vj2z19ig, zklx1cuw; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 420.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 8k 420.1 34.7 4.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a0dt7lxk",
   "source_url": "https://omlx.ai/benchmarks/performance/a0dt7lxk",
   "tg_tokens": null,
   "tps": 34.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-a0dt7lxk-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.4 GB at this context; duplicate traces of the same session: b1grvu6s, vj2z19ig, zklx1cuw; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 380.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 16k 380.9 30.7 4.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a0dt7lxk",
   "source_url": "https://omlx.ai/benchmarks/performance/a0dt7lxk",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-u6g7ivo5-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.9 GB at this context; duplicate traces of the same session: 17z8px1p, 3576eof4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4536.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k 4,536 131.9 30.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u6g7ivo5",
   "source_url": "https://omlx.ai/benchmarks/performance/u6g7ivo5",
   "tg_tokens": null,
   "tps": 131.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-u6g7ivo5-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: 17z8px1p, 3576eof4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3159.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 64k 3,159 127.7 32.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u6g7ivo5",
   "source_url": "https://omlx.ai/benchmarks/performance/u6g7ivo5",
   "tg_tokens": null,
   "tps": 127.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-u6g7ivo5-ctx199680",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.3 GB at this context; duplicate traces of the same session: 17z8px1p, 3576eof4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1690.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 195k 1,690 69.7 35.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u6g7ivo5",
   "source_url": "https://omlx.ai/benchmarks/performance/u6g7ivo5",
   "tg_tokens": null,
   "tps": 69.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-zmk5r13a-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 15.5 GB at this context; duplicate traces of the same session: 7v4t4j92; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 283.1,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.8-27B-3bit-mtp-mlx on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 3bit; 1k 283.1 20.8 15.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zmk5r13a",
   "source_url": "https://omlx.ai/benchmarks/performance/zmk5r13a",
   "tg_tokens": null,
   "tps": 20.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-zmk5r13a-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.7 GB at this context; duplicate traces of the same session: 7v4t4j92; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 297.6,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.8-27B-3bit-mtp-mlx on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 3bit; 4k 297.6 19.3 16.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zmk5r13a",
   "source_url": "https://omlx.ai/benchmarks/performance/zmk5r13a",
   "tg_tokens": null,
   "tps": 19.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-0x2l2k91-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 24.9 GB; TTFT 134391 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 975.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k tokens 975.3 PP tok/s 33.9 TG tok/s 134391 TTFT (ms) 24.9 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0x2l2k91",
   "source_url": "https://omlx.ai/benchmarks/performance/0x2l2k91",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-30x97j9p-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 20.9 GB; TTFT 53530 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1224.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k tokens 1,224 PP tok/s 41.3 TG tok/s 53530 TTFT (ms) 20.9 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/30x97j9p",
   "source_url": "https://omlx.ai/benchmarks/performance/30x97j9p",
   "tg_tokens": null,
   "tps": 41.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-i7rqn7k3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 76xn789w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 224.5 22.7 16.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i7rqn7k3",
   "source_url": "https://omlx.ai/benchmarks/performance/i7rqn7k3",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-i7rqn7k3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 76xn789w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 225.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 225.6 22.1 19.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i7rqn7k3",
   "source_url": "https://omlx.ai/benchmarks/performance/i7rqn7k3",
   "tg_tokens": null,
   "tps": 22.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ebwrph1h-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: y09hf6je; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 325.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 325.5 78.8 21.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ebwrph1h",
   "source_url": "https://omlx.ai/benchmarks/performance/ebwrph1h",
   "tg_tokens": null,
   "tps": 78.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ebwrph1h-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: y09hf6je; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 424.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 424.0 102.6 24.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ebwrph1h",
   "source_url": "https://omlx.ai/benchmarks/performance/ebwrph1h",
   "tg_tokens": null,
   "tps": 102.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-2l6xceuf-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: v0o75456; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 278.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 278.9 9.0 16.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2l6xceuf",
   "source_url": "https://omlx.ai/benchmarks/performance/2l6xceuf",
   "tg_tokens": null,
   "tps": 9.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-2l6xceuf-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: v0o75456; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 298.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 298.1 8.9 17.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2l6xceuf",
   "source_url": "https://omlx.ai/benchmarks/performance/2l6xceuf",
   "tg_tokens": null,
   "tps": 8.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-h7iim0mt-ctx199680",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace, single run at 195k context; peak memory 61.1 GB; TTFT 46788 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4275.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 195k tokens 4,275 PP tok/s 71.8 TG tok/s 46788 TTFT (ms) 61.1 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h7iim0mt",
   "source_url": "https://omlx.ai/benchmarks/performance/h7iim0mt",
   "tg_tokens": null,
   "tps": 71.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-le6vkjgw-ctx16384",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 25.1 GB; TTFT 38522 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 425.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k tokens 425.3 PP tok/s 54.1 TG tok/s 38522 TTFT (ms) 25.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/le6vkjgw",
   "source_url": "https://omlx.ai/benchmarks/performance/le6vkjgw",
   "tg_tokens": null,
   "tps": 54.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-18kwhs1w-ctx16384",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 25.1 GB; TTFT 38621 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 424.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k tokens 424.2 PP tok/s 48.3 TG tok/s 38621 TTFT (ms) 25.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/18kwhs1w",
   "source_url": "https://omlx.ai/benchmarks/performance/18kwhs1w",
   "tg_tokens": null,
   "tps": 48.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ovriwo31-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 4t8jea8z, suhugkit; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 774.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 774.1 91.1 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ovriwo31",
   "source_url": "https://omlx.ai/benchmarks/performance/ovriwo31",
   "tg_tokens": null,
   "tps": 91.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ovriwo31-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 4t8jea8z, suhugkit; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 886.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 886.1 87.1 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ovriwo31",
   "source_url": "https://omlx.ai/benchmarks/performance/ovriwo31",
   "tg_tokens": null,
   "tps": 87.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ovriwo31-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: 4t8jea8z, suhugkit; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 580.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 32k 580.1 62.8 23.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ovriwo31",
   "source_url": "https://omlx.ai/benchmarks/performance/ovriwo31",
   "tg_tokens": null,
   "tps": 62.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-eqp70z2x-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: kmg5aqnx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1490.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,490 107.4 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eqp70z2x",
   "source_url": "https://omlx.ai/benchmarks/performance/eqp70z2x",
   "tg_tokens": null,
   "tps": 107.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-eqp70z2x-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: kmg5aqnx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1845.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,845 102.9 21.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eqp70z2x",
   "source_url": "https://omlx.ai/benchmarks/performance/eqp70z2x",
   "tg_tokens": null,
   "tps": 102.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9d9r6rnw-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 8fhypiwq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 972.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 972.7 49.5 18.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9d9r6rnw",
   "source_url": "https://omlx.ai/benchmarks/performance/9d9r6rnw",
   "tg_tokens": null,
   "tps": 49.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9d9r6rnw-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 8fhypiwq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 950.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 950.5 37.0 19.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9d9r6rnw",
   "source_url": "https://omlx.ai/benchmarks/performance/9d9r6rnw",
   "tg_tokens": null,
   "tps": 37.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-h3orjrax-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: nlia9mud; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 132.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k 132.6 20.6 19.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h3orjrax",
   "source_url": "https://omlx.ai/benchmarks/performance/h3orjrax",
   "tg_tokens": null,
   "tps": 20.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-h3orjrax-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: nlia9mud; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 122.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 16k 122.1 18.3 21.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h3orjrax",
   "source_url": "https://omlx.ai/benchmarks/performance/h3orjrax",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-bsw763s1-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: komgrcuq; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 142.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 1k 142.7 20.7 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bsw763s1",
   "source_url": "https://omlx.ai/benchmarks/performance/bsw763s1",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-bsw763s1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: komgrcuq; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 132.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k 132.2 20.6 19.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bsw763s1",
   "source_url": "https://omlx.ai/benchmarks/performance/bsw763s1",
   "tg_tokens": null,
   "tps": 20.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-d57mj3ia-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: 1tl3odaf; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 290.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 290.1 9.5 15.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d57mj3ia",
   "source_url": "https://omlx.ai/benchmarks/performance/d57mj3ia",
   "tg_tokens": null,
   "tps": 9.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-d57mj3ia-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 1tl3odaf; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 309.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 309.6 9.4 17.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d57mj3ia",
   "source_url": "https://omlx.ai/benchmarks/performance/d57mj3ia",
   "tg_tokens": null,
   "tps": 9.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-qa58plyq-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: tbyb80dg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 92.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 92.2 17.1 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qa58plyq",
   "source_url": "https://omlx.ai/benchmarks/performance/qa58plyq",
   "tg_tokens": null,
   "tps": 17.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-qa58plyq-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: tbyb80dg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 103.0 17.6 21.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qa58plyq",
   "source_url": "https://omlx.ai/benchmarks/performance/qa58plyq",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-mr6v49o3-ctx1024",
   "model": "Qwen3.8-27B-heretic-ara",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 8oeip2f0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-27B-heretic-ara-oQ3.5e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 3bit; 1k 127.0 21.4 16.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mr6v49o3",
   "source_url": "https://omlx.ai/benchmarks/performance/mr6v49o3",
   "tg_tokens": null,
   "tps": 21.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-mr6v49o3-ctx8192",
   "model": "Qwen3.8-27B-heretic-ara",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.9 GB at this context; duplicate traces of the same session: 8oeip2f0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.1,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-27B-heretic-ara-oQ3.5e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 3bit; 8k 125.1 18.6 17.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mr6v49o3",
   "source_url": "https://omlx.ai/benchmarks/performance/mr6v49o3",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-020j5m7u-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: sekbyx3y; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 269.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 269.8 18.8 17.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/020j5m7u",
   "source_url": "https://omlx.ai/benchmarks/performance/020j5m7u",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-020j5m7u-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: sekbyx3y; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 220.8 19.5 18.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/020j5m7u",
   "source_url": "https://omlx.ai/benchmarks/performance/020j5m7u",
   "tg_tokens": null,
   "tps": 19.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-6ysr3rbb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: 5j78eezd, enppdnjl, lwhw5jmi, m6oet5x3, qwnvyjq2, rc759flg, ymbchyka; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 832.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 832.5 59.7 26.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ysr3rbb",
   "source_url": "https://omlx.ai/benchmarks/performance/6ysr3rbb",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-6ysr3rbb-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: 5j78eezd, enppdnjl, lwhw5jmi, m6oet5x3, qwnvyjq2, rc759flg, ymbchyka; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 925.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 925.7 77.2 27.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ysr3rbb",
   "source_url": "https://omlx.ai/benchmarks/performance/6ysr3rbb",
   "tg_tokens": null,
   "tps": 77.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6ysr3rbb-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: 5j78eezd, enppdnjl, lwhw5jmi, m6oet5x3, qwnvyjq2, rc759flg, ymbchyka; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 923.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 8k 923.6 39.5 28.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ysr3rbb",
   "source_url": "https://omlx.ai/benchmarks/performance/6ysr3rbb",
   "tg_tokens": null,
   "tps": 39.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6ysr3rbb-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: 5j78eezd, enppdnjl, lwhw5jmi, m6oet5x3, qwnvyjq2, rc759flg, ymbchyka; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 901.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k 901.9 56.4 28.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ysr3rbb",
   "source_url": "https://omlx.ai/benchmarks/performance/6ysr3rbb",
   "tg_tokens": null,
   "tps": 56.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6ysr3rbb-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: 5j78eezd, enppdnjl, lwhw5jmi, m6oet5x3, qwnvyjq2, rc759flg, ymbchyka; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 846.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 846.7 51.8 29.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ysr3rbb",
   "source_url": "https://omlx.ai/benchmarks/performance/6ysr3rbb",
   "tg_tokens": null,
   "tps": 51.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6ysr3rbb-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: 5j78eezd, enppdnjl, lwhw5jmi, m6oet5x3, qwnvyjq2, rc759flg, ymbchyka; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 744.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 64k 744.9 39.6 32.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ysr3rbb",
   "source_url": "https://omlx.ai/benchmarks/performance/6ysr3rbb",
   "tg_tokens": null,
   "tps": 39.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-6ysr3rbb-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: 5j78eezd, enppdnjl, lwhw5jmi, m6oet5x3, qwnvyjq2, rc759flg, ymbchyka; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 598.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 128k 598.8 43.2 36.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ysr3rbb",
   "source_url": "https://omlx.ai/benchmarks/performance/6ysr3rbb",
   "tg_tokens": null,
   "tps": 43.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-6ysr3rbb-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 42.1 GB at this context; duplicate traces of the same session: 5j78eezd, enppdnjl, lwhw5jmi, m6oet5x3, qwnvyjq2, rc759flg, ymbchyka; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 496.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 195k 496.2 38.3 42.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ysr3rbb",
   "source_url": "https://omlx.ai/benchmarks/performance/6ysr3rbb",
   "tg_tokens": null,
   "tps": 38.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-827hntl4-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "RCO; DWQ; oMLX community benchmark trace; peak memory 14.0 GB at this context; duplicate traces of the same session: sb4fb1cw; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 121.3,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Swift-1.5-Qwen3.8-27b-RCO-DWQ-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 3bit; 1k 121.3 21.8 14.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/827hntl4",
   "source_url": "https://omlx.ai/benchmarks/performance/827hntl4",
   "tg_tokens": null,
   "tps": 21.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-827hntl4-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "RCO; DWQ; oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: sb4fb1cw; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 122.2,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Swift-1.5-Qwen3.8-27b-RCO-DWQ-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 3bit; 8k 122.2 18.4 15.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/827hntl4",
   "source_url": "https://omlx.ai/benchmarks/performance/827hntl4",
   "tg_tokens": null,
   "tps": 18.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-bqlwjc6d-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "RCO; DWQ; oMLX community benchmark trace; peak memory 14.7 GB at this context; duplicate traces of the same session: 9fuew736; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 128.7,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Swift-1.5-Qwen3.8-27b-RCO-DWQ-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 3bit; 1k 128.7 17.4 14.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bqlwjc6d",
   "source_url": "https://omlx.ai/benchmarks/performance/bqlwjc6d",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-bqlwjc6d-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "RCO; DWQ; oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 9fuew736; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.3,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Swift-1.5-Qwen3.8-27b-RCO-DWQ-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 3bit; 8k 125.3 15.5 16.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bqlwjc6d",
   "source_url": "https://omlx.ai/benchmarks/performance/bqlwjc6d",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-5fbv44ny-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: xrb8rsph; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 126.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 126.0 29.3 18.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5fbv44ny",
   "source_url": "https://omlx.ai/benchmarks/performance/5fbv44ny",
   "tg_tokens": null,
   "tps": 29.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-5fbv44ny-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: xrb8rsph; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 127.8 16.3 20.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5fbv44ny",
   "source_url": "https://omlx.ai/benchmarks/performance/5fbv44ny",
   "tg_tokens": null,
   "tps": 16.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-9kuwqrka-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.7 GB at this context; duplicate traces of the same session: 29im37ec, 31wfdl6p; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1654.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 1,654 48.7 11.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9kuwqrka",
   "source_url": "https://omlx.ai/benchmarks/performance/9kuwqrka",
   "tg_tokens": null,
   "tps": 48.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-9kuwqrka-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.7 GB at this context; duplicate traces of the same session: 29im37ec, 31wfdl6p; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1879.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,879 47.5 11.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9kuwqrka",
   "source_url": "https://omlx.ai/benchmarks/performance/9kuwqrka",
   "tg_tokens": null,
   "tps": 47.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-9kuwqrka-ctx8192",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: 29im37ec, 31wfdl6p; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1848.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 8k 1,848 45.4 11.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9kuwqrka",
   "source_url": "https://omlx.ai/benchmarks/performance/9kuwqrka",
   "tg_tokens": null,
   "tps": 45.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-e01mwa0j-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.7 GB at this context; duplicate traces of the same session: omzevpjt; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1657.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 1,657 49.6 11.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e01mwa0j",
   "source_url": "https://omlx.ai/benchmarks/performance/e01mwa0j",
   "tg_tokens": null,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-e01mwa0j-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.7 GB at this context; duplicate traces of the same session: omzevpjt; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1887.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,887 47.4 11.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e01mwa0j",
   "source_url": "https://omlx.ai/benchmarks/performance/e01mwa0j",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-si6e1p33-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.7 GB at this context; duplicate traces of the same session: z942m0xe; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 681.9,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 681.9 47.3 11.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/si6e1p33",
   "source_url": "https://omlx.ai/benchmarks/performance/si6e1p33",
   "tg_tokens": null,
   "tps": 47.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-si6e1p33-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.7 GB at this context; duplicate traces of the same session: z942m0xe; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1769.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,769 45.1 11.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/si6e1p33",
   "source_url": "https://omlx.ai/benchmarks/performance/si6e1p33",
   "tg_tokens": null,
   "tps": 45.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-gdst3rlb-ctx1024",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "prism-ml MLX build; oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: yavr88jb; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 152.2,
   "provenance": "community",
   "quant": "2bit",
   "quote": "prism-ml--Ternary-Bonsai-2-27B-mlx-2bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 unknown; 1k 152.2 11.7 11.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gdst3rlb",
   "source_url": "https://omlx.ai/benchmarks/performance/gdst3rlb",
   "tg_tokens": null,
   "tps": 11.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-gdst3rlb-ctx4096",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "prism-ml MLX build; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: yavr88jb; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 157.0,
   "provenance": "community",
   "quant": "2bit",
   "quote": "prism-ml--Ternary-Bonsai-2-27B-mlx-2bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 unknown; 4k 157.0 11.6 13.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gdst3rlb",
   "source_url": "https://omlx.ai/benchmarks/performance/gdst3rlb",
   "tg_tokens": null,
   "tps": 11.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-05b39hw6-ctx4096",
   "model": "AREX-2",
   "notes": "fp16; oMLX community benchmark trace; peak memory 32.3 GB at this context; duplicate traces of the same session: tq05brqf; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 162.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "AREX-2-oQ8e-fp16 on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 4k 162.5 11.8 32.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/05b39hw6",
   "source_url": "https://omlx.ai/benchmarks/performance/05b39hw6",
   "tg_tokens": null,
   "tps": 11.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-05b39hw6-ctx16384",
   "model": "AREX-2",
   "notes": "fp16; oMLX community benchmark trace; peak memory 34.6 GB at this context; duplicate traces of the same session: tq05brqf; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 149.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "AREX-2-oQ8e-fp16 on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 16k 149.8 11.3 34.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/05b39hw6",
   "source_url": "https://omlx.ai/benchmarks/performance/05b39hw6",
   "tg_tokens": null,
   "tps": 11.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-s15rqlq8-ctx4096",
   "model": "AREX-2",
   "notes": "fp16; oMLX community benchmark trace; peak memory 32.3 GB at this context; duplicate traces of the same session: a4py8lv6; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 162.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "AREX-2-oQ8e-fp16 on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 4k 162.5 11.8 32.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s15rqlq8",
   "source_url": "https://omlx.ai/benchmarks/performance/s15rqlq8",
   "tg_tokens": null,
   "tps": 11.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-s15rqlq8-ctx16384",
   "model": "AREX-2",
   "notes": "fp16; oMLX community benchmark trace; peak memory 34.6 GB at this context; duplicate traces of the same session: a4py8lv6; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 152.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "AREX-2-oQ8e-fp16 on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 16k 152.0 11.3 34.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s15rqlq8",
   "source_url": "https://omlx.ai/benchmarks/performance/s15rqlq8",
   "tg_tokens": null,
   "tps": 11.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5wtibism-ctx1024",
   "model": "Huihui-Qwen3.6-35B-A3B-Claude-4.6-Opus-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: jody9lob; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3853.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Huihui-Qwen3.6-35B-A3B-Claude-4.6-Opus-abliterated-mlx-oQ8-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 3,853 96.3 35.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5wtibism",
   "source_url": "https://omlx.ai/benchmarks/performance/5wtibism",
   "tg_tokens": null,
   "tps": 96.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5wtibism-ctx4096",
   "model": "Huihui-Qwen3.6-35B-A3B-Claude-4.6-Opus-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: jody9lob; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2148.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Huihui-Qwen3.6-35B-A3B-Claude-4.6-Opus-abliterated-mlx-oQ8-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 2,148 94.2 36.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5wtibism",
   "source_url": "https://omlx.ai/benchmarks/performance/5wtibism",
   "tg_tokens": null,
   "tps": 94.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-i8itkyaq-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: v3n4vnu8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3064.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 3,064 146.1 20.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i8itkyaq",
   "source_url": "https://omlx.ai/benchmarks/performance/i8itkyaq",
   "tg_tokens": null,
   "tps": 146.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-i8itkyaq-ctx131072",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: v3n4vnu8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2252.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 2,252 78.0 22.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i8itkyaq",
   "source_url": "https://omlx.ai/benchmarks/performance/i8itkyaq",
   "tg_tokens": null,
   "tps": 78.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gxgx16i7-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 5.9 GB; TTFT 445 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2299.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 2,299 PP tok/s 43.4 TG tok/s 445 TTFT (ms) 5.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gxgx16i7",
   "source_url": "https://omlx.ai/benchmarks/performance/gxgx16i7",
   "tg_tokens": null,
   "tps": 43.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9ieiid3p-ctx131072",
   "model": "Qwen3.6-35B-A3B-Abliterated-Heretic",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 26.5 GB; TTFT 54331 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2413.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Abliterated-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k tokens 2,413 PP tok/s 72.6 TG tok/s 54331 TTFT (ms) 26.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ieiid3p",
   "source_url": "https://omlx.ai/benchmarks/performance/9ieiid3p",
   "tg_tokens": null,
   "tps": 72.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1xksk41u-ctx65536",
   "model": "Qwen3.6-35B-A3B-Abliterated-Heretic",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 25.2 GB; TTFT 18712 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3502.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Abliterated-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k tokens 3,502 PP tok/s 94.0 TG tok/s 18712 TTFT (ms) 25.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1xksk41u",
   "source_url": "https://omlx.ai/benchmarks/performance/1xksk41u",
   "tg_tokens": null,
   "tps": 94.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-boh0xaxz-ctx1024",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 12.6 GB at this context; duplicate traces of the same session: 45zm2lc0; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 414.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 414.5 49.1 12.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/boh0xaxz",
   "source_url": "https://omlx.ai/benchmarks/performance/boh0xaxz",
   "tg_tokens": null,
   "tps": 49.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-boh0xaxz-ctx4096",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.2 GB at this context; duplicate traces of the same session: 45zm2lc0; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1097.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,097 49.8 13.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/boh0xaxz",
   "source_url": "https://omlx.ai/benchmarks/performance/boh0xaxz",
   "tg_tokens": null,
   "tps": 49.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cyld7vo9-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: 47wi04zr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 785.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 785.6 3.9 22.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cyld7vo9",
   "source_url": "https://omlx.ai/benchmarks/performance/cyld7vo9",
   "tg_tokens": null,
   "tps": 3.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cyld7vo9-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.6 GB at this context; duplicate traces of the same session: 47wi04zr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 616.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 616.7 3.2 32.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cyld7vo9",
   "source_url": "https://omlx.ai/benchmarks/performance/cyld7vo9",
   "tg_tokens": null,
   "tps": 3.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-99w7k6ex-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: s8lo9zgu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 127.5 15.1 16.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/99w7k6ex",
   "source_url": "https://omlx.ai/benchmarks/performance/99w7k6ex",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-99w7k6ex-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: s8lo9zgu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 126.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 126.7 14.7 17.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/99w7k6ex",
   "source_url": "https://omlx.ai/benchmarks/performance/99w7k6ex",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1or3p45t-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 16.6 GB; TTFT 1109 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 923.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 923.4 PP tok/s 32.3 TG tok/s 1109 TTFT (ms) 16.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1or3p45t",
   "source_url": "https://omlx.ai/benchmarks/performance/1or3p45t",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-621hy4e7-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 2dj09d6n, odsq6rpg, zqbnl165; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 918.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 918.0 32.3 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/621hy4e7",
   "source_url": "https://omlx.ai/benchmarks/performance/621hy4e7",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-621hy4e7-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 2dj09d6n, odsq6rpg, zqbnl165; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 999.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 999.8 31.8 17.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/621hy4e7",
   "source_url": "https://omlx.ai/benchmarks/performance/621hy4e7",
   "tg_tokens": null,
   "tps": 31.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-621hy4e7-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 2dj09d6n, odsq6rpg, zqbnl165; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 808.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 808.4 25.7 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/621hy4e7",
   "source_url": "https://omlx.ai/benchmarks/performance/621hy4e7",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-621hy4e7-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: 2dj09d6n, odsq6rpg, zqbnl165; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 651.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 651.6 21.2 25.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/621hy4e7",
   "source_url": "https://omlx.ai/benchmarks/performance/621hy4e7",
   "tg_tokens": null,
   "tps": 21.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-x1w1nvif-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: ikxexx8d, xiw7pw66; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1267.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,267 48.8 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x1w1nvif",
   "source_url": "https://omlx.ai/benchmarks/performance/x1w1nvif",
   "tg_tokens": null,
   "tps": 48.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-x1w1nvif-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: ikxexx8d, xiw7pw66; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1165.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,165 74.4 19.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x1w1nvif",
   "source_url": "https://omlx.ai/benchmarks/performance/x1w1nvif",
   "tg_tokens": null,
   "tps": 74.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-x1w1nvif-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: ikxexx8d, xiw7pw66; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 556.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 556.9 30.4 27.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x1w1nvif",
   "source_url": "https://omlx.ai/benchmarks/performance/x1w1nvif",
   "tg_tokens": null,
   "tps": 30.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cfd2jz6r-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: i59l0l5q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1021.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,021 67.9 70.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cfd2jz6r",
   "source_url": "https://omlx.ai/benchmarks/performance/cfd2jz6r",
   "tg_tokens": null,
   "tps": 67.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cfd2jz6r-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: i59l0l5q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1880.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,880 93.9 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cfd2jz6r",
   "source_url": "https://omlx.ai/benchmarks/performance/cfd2jz6r",
   "tg_tokens": null,
   "tps": 93.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-bl8ts83v-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "RCO; DWQ; oMLX community benchmark trace; peak memory 13.0 GB at this context; duplicate traces of the same session: ua9j0772; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 132.0,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Swift-1.5-Qwen3.8-27b-RCO-DWQ-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 3bit; 1k 132.0 18.4 13.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bl8ts83v",
   "source_url": "https://omlx.ai/benchmarks/performance/bl8ts83v",
   "tg_tokens": null,
   "tps": 18.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-bl8ts83v-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "RCO; DWQ; oMLX community benchmark trace; peak memory 14.2 GB at this context; duplicate traces of the same session: ua9j0772; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 120.9,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Swift-1.5-Qwen3.8-27b-RCO-DWQ-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 3bit; 8k 120.9 17.0 14.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bl8ts83v",
   "source_url": "https://omlx.ai/benchmarks/performance/bl8ts83v",
   "tg_tokens": null,
   "tps": 17.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-dkppwc6j-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 24.2 GB; TTFT 115955 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 141.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k tokens 141.3 PP tok/s 20.0 TG tok/s 115955 TTFT (ms) 24.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dkppwc6j",
   "source_url": "https://omlx.ai/benchmarks/performance/dkppwc6j",
   "tg_tokens": null,
   "tps": 20.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fmk4hx5z-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 1bg90l1t, 5q0rr6rs; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1324.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,324 65.1 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fmk4hx5z",
   "source_url": "https://omlx.ai/benchmarks/performance/fmk4hx5z",
   "tg_tokens": null,
   "tps": 65.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fmk4hx5z-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 1bg90l1t, 5q0rr6rs; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1108.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,108 79.3 19.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fmk4hx5z",
   "source_url": "https://omlx.ai/benchmarks/performance/fmk4hx5z",
   "tg_tokens": null,
   "tps": 79.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fmk4hx5z-ctx131072",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: 1bg90l1t, 5q0rr6rs; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 523.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 523.3 31.9 27.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fmk4hx5z",
   "source_url": "https://omlx.ai/benchmarks/performance/fmk4hx5z",
   "tg_tokens": null,
   "tps": 31.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-8lv0s4ft-ctx1024",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 266lf56d, 57x8zc8c; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 104.9 14.2 21.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8lv0s4ft",
   "source_url": "https://omlx.ai/benchmarks/performance/8lv0s4ft",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-8lv0s4ft-ctx4096",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: 266lf56d, 57x8zc8c; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 155.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 155.8 14.0 24.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8lv0s4ft",
   "source_url": "https://omlx.ai/benchmarks/performance/8lv0s4ft",
   "tg_tokens": null,
   "tps": 14.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-8lv0s4ft-ctx32768",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: 266lf56d, 57x8zc8c; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 123.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 32k 123.6 13.4 27.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8lv0s4ft",
   "source_url": "https://omlx.ai/benchmarks/performance/8lv0s4ft",
   "tg_tokens": null,
   "tps": 13.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-s3ph7dfe-ctx1024",
   "model": "Qwythos-9B-v2",
   "notes": "OptiQ; oMLX community benchmark trace; peak memory 7.7 GB at this context; duplicate traces of the same session: 3799tyfu, cy9aiocf, fiog2imc, kuoc2tkf, nstvw0a0; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 341.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwythos-9B-v2-OptiQ-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 341.9 24.7 7.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s3ph7dfe",
   "source_url": "https://omlx.ai/benchmarks/performance/s3ph7dfe",
   "tg_tokens": null,
   "tps": 24.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-s3ph7dfe-ctx4096",
   "model": "Qwythos-9B-v2",
   "notes": "OptiQ; oMLX community benchmark trace; peak memory 9.2 GB at this context; duplicate traces of the same session: 3799tyfu, cy9aiocf, fiog2imc, kuoc2tkf, nstvw0a0; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 356.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwythos-9B-v2-OptiQ-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 356.7 24.4 9.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s3ph7dfe",
   "source_url": "https://omlx.ai/benchmarks/performance/s3ph7dfe",
   "tg_tokens": null,
   "tps": 24.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-s3ph7dfe-ctx8192",
   "model": "Qwythos-9B-v2",
   "notes": "OptiQ; oMLX community benchmark trace; peak memory 9.6 GB at this context; duplicate traces of the same session: 3799tyfu, cy9aiocf, fiog2imc, kuoc2tkf, nstvw0a0; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 342.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwythos-9B-v2-OptiQ-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 342.8 24.4 9.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s3ph7dfe",
   "source_url": "https://omlx.ai/benchmarks/performance/s3ph7dfe",
   "tg_tokens": null,
   "tps": 24.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-s3ph7dfe-ctx16384",
   "model": "Qwythos-9B-v2",
   "notes": "OptiQ; oMLX community benchmark trace; peak memory 10.3 GB at this context; duplicate traces of the same session: 3799tyfu, cy9aiocf, fiog2imc, kuoc2tkf, nstvw0a0; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 315.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwythos-9B-v2-OptiQ-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 315.9 22.4 10.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s3ph7dfe",
   "source_url": "https://omlx.ai/benchmarks/performance/s3ph7dfe",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-s3ph7dfe-ctx32768",
   "model": "Qwythos-9B-v2",
   "notes": "OptiQ; oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: 3799tyfu, cy9aiocf, fiog2imc, kuoc2tkf, nstvw0a0; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 301.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwythos-9B-v2-OptiQ-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 32k 301.4 18.7 11.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s3ph7dfe",
   "source_url": "https://omlx.ai/benchmarks/performance/s3ph7dfe",
   "tg_tokens": null,
   "tps": 18.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-s3ph7dfe-ctx65536",
   "model": "Qwythos-9B-v2",
   "notes": "OptiQ; oMLX community benchmark trace; peak memory 12.9 GB at this context; duplicate traces of the same session: 3799tyfu, cy9aiocf, fiog2imc, kuoc2tkf, nstvw0a0; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 253.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwythos-9B-v2-OptiQ-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 64k 253.8 16.3 12.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s3ph7dfe",
   "source_url": "https://omlx.ai/benchmarks/performance/s3ph7dfe",
   "tg_tokens": null,
   "tps": 16.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-pxn8dk7z-ctx1024",
   "model": "Qwythos-9B-v2",
   "notes": "OptiQ; oMLX community benchmark trace; peak memory 7.7 GB at this context; duplicate traces of the same session: 6umvbud1, nndkcdut, qroxtync; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 485.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwythos-9B-v2-OptiQ-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 485.1 37.8 7.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pxn8dk7z",
   "source_url": "https://omlx.ai/benchmarks/performance/pxn8dk7z",
   "tg_tokens": null,
   "tps": 37.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-pxn8dk7z-ctx4096",
   "model": "Qwythos-9B-v2",
   "notes": "OptiQ; oMLX community benchmark trace; peak memory 9.2 GB at this context; duplicate traces of the same session: 6umvbud1, nndkcdut, qroxtync; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 423.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwythos-9B-v2-OptiQ-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 423.8 40.9 9.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pxn8dk7z",
   "source_url": "https://omlx.ai/benchmarks/performance/pxn8dk7z",
   "tg_tokens": null,
   "tps": 40.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-pxn8dk7z-ctx8192",
   "model": "Qwythos-9B-v2",
   "notes": "OptiQ; oMLX community benchmark trace; peak memory 9.6 GB at this context; duplicate traces of the same session: 6umvbud1, nndkcdut, qroxtync; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 391.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwythos-9B-v2-OptiQ-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 391.3 38.7 9.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pxn8dk7z",
   "source_url": "https://omlx.ai/benchmarks/performance/pxn8dk7z",
   "tg_tokens": null,
   "tps": 38.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-pxn8dk7z-ctx16384",
   "model": "Qwythos-9B-v2",
   "notes": "OptiQ; oMLX community benchmark trace; peak memory 10.3 GB at this context; duplicate traces of the same session: 6umvbud1, nndkcdut, qroxtync; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 379.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwythos-9B-v2-OptiQ-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 379.0 36.3 10.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pxn8dk7z",
   "source_url": "https://omlx.ai/benchmarks/performance/pxn8dk7z",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-pxn8dk7z-ctx32768",
   "model": "Qwythos-9B-v2",
   "notes": "OptiQ; oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: 6umvbud1, nndkcdut, qroxtync; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 340.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwythos-9B-v2-OptiQ-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 32k 340.4 33.0 11.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pxn8dk7z",
   "source_url": "https://omlx.ai/benchmarks/performance/pxn8dk7z",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-pxn8dk7z-ctx65536",
   "model": "Qwythos-9B-v2",
   "notes": "OptiQ; oMLX community benchmark trace; peak memory 12.9 GB at this context; duplicate traces of the same session: 6umvbud1, nndkcdut, qroxtync; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 293.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwythos-9B-v2-OptiQ-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 64k 293.1 25.3 12.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pxn8dk7z",
   "source_url": "https://omlx.ai/benchmarks/performance/pxn8dk7z",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-573st713-ctx4096",
   "model": "GLM-4.7-Flash",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 24.3 GB; TTFT 3769 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1087.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "GLM-4.7-Flash-6bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k tokens 1,087 PP tok/s 64.1 TG tok/s 3769 TTFT (ms) 24.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/573st713",
   "source_url": "https://omlx.ai/benchmarks/performance/573st713",
   "tg_tokens": null,
   "tps": 64.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-p3oys412-ctx1024",
   "model": "Huihui-Qwen3.6-35B-A3B-Claude-4.6-Opus-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4232.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Huihui-Qwen3.6-35B-A3B-Claude-4.6-Opus-abliterated-mlx-oQ8-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 4,232 99.7 35.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p3oys412",
   "source_url": "https://omlx.ai/benchmarks/performance/p3oys412",
   "tg_tokens": null,
   "tps": 99.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-p3oys412-ctx4096",
   "model": "Huihui-Qwen3.6-35B-A3B-Claude-4.6-Opus-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5824.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Huihui-Qwen3.6-35B-A3B-Claude-4.6-Opus-abliterated-mlx-oQ8-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 5,824 95.1 36.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p3oys412",
   "source_url": "https://omlx.ai/benchmarks/performance/p3oys412",
   "tg_tokens": null,
   "tps": 95.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-p3oys412-ctx8192",
   "model": "Huihui-Qwen3.6-35B-A3B-Claude-4.6-Opus-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5699.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Huihui-Qwen3.6-35B-A3B-Claude-4.6-Opus-abliterated-mlx-oQ8-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 5,699 95.0 36.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p3oys412",
   "source_url": "https://omlx.ai/benchmarks/performance/p3oys412",
   "tg_tokens": null,
   "tps": 95.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-4radrrhy-ctx1024",
   "model": "PaddleOCR-VL-1.6",
   "notes": "oMLX community benchmark trace; peak memory 1.2 GB at this context; duplicate traces of the same session: hlr12uvf; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4633.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "PaddleOCR-VL-1.6-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 4,633 287.8 1.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4radrrhy",
   "source_url": "https://omlx.ai/benchmarks/performance/4radrrhy",
   "tg_tokens": null,
   "tps": 287.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-4radrrhy-ctx4096",
   "model": "PaddleOCR-VL-1.6",
   "notes": "oMLX community benchmark trace; peak memory 1.4 GB at this context; duplicate traces of the same session: hlr12uvf; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3385.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "PaddleOCR-VL-1.6-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 3,385 205.9 1.4 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4radrrhy",
   "source_url": "https://omlx.ai/benchmarks/performance/4radrrhy",
   "tg_tokens": null,
   "tps": 205.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yfx0ipzk-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 5.9 GB; TTFT 337 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3037.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 3,037 PP tok/s 100.8 TG tok/s 337 TTFT (ms) 5.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yfx0ipzk",
   "source_url": "https://omlx.ai/benchmarks/performance/yfx0ipzk",
   "tg_tokens": null,
   "tps": 100.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5",
   "id": "omlx-ne1dlhse-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 7.2 GB at this context; duplicate traces of the same session: yr09ncuq; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 177.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 (10c); M5 (10c) \u00b7 16 GB \u00b7 4bit; 1k 177.1 14.7 7.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ne1dlhse",
   "source_url": "https://omlx.ai/benchmarks/performance/ne1dlhse",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5",
   "id": "omlx-ne1dlhse-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 8.0 GB at this context; duplicate traces of the same session: yr09ncuq; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 236.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 (10c); M5 (10c) \u00b7 16 GB \u00b7 4bit; 4k 236.9 13.6 8.0 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ne1dlhse",
   "source_url": "https://omlx.ai/benchmarks/performance/ne1dlhse",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-3x6j0del-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 253.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 253.9 26.8 17.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3x6j0del",
   "source_url": "https://omlx.ai/benchmarks/performance/3x6j0del",
   "tg_tokens": null,
   "tps": 26.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-3x6j0del-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 244.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 244.7 25.2 17.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3x6j0del",
   "source_url": "https://omlx.ai/benchmarks/performance/3x6j0del",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-d2a6wwp0-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 16.0 GB; TTFT 819 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1250.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k tokens 1,250 PP tok/s 54.7 TG tok/s 819 TTFT (ms) 16.0 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d2a6wwp0",
   "source_url": "https://omlx.ai/benchmarks/performance/d2a6wwp0",
   "tg_tokens": null,
   "tps": 54.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-enr7resl-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 30dlfg0x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 894.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 894.9 40.1 20.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/enr7resl",
   "source_url": "https://omlx.ai/benchmarks/performance/enr7resl",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-enr7resl-ctx65536",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: 30dlfg0x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 775.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 775.6 42.5 22.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/enr7resl",
   "source_url": "https://omlx.ai/benchmarks/performance/enr7resl",
   "tg_tokens": null,
   "tps": 42.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-oczgdrv2-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 9b0oz2at; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 194.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 1k 194.2 35.2 18.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oczgdrv2",
   "source_url": "https://omlx.ai/benchmarks/performance/oczgdrv2",
   "tg_tokens": null,
   "tps": 35.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-oczgdrv2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 9b0oz2at; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 197.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 4k 197.4 25.5 22.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oczgdrv2",
   "source_url": "https://omlx.ai/benchmarks/performance/oczgdrv2",
   "tg_tokens": null,
   "tps": 25.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-isrojlr4-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: fjgtcn9h; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 123.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 123.8 28.0 18.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/isrojlr4",
   "source_url": "https://omlx.ai/benchmarks/performance/isrojlr4",
   "tg_tokens": null,
   "tps": 28.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-isrojlr4-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: fjgtcn9h; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 129.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 129.6 31.7 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/isrojlr4",
   "source_url": "https://omlx.ai/benchmarks/performance/isrojlr4",
   "tg_tokens": null,
   "tps": 31.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-69d5s4zw-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.5 GB at this context; duplicate traces of the same session: 3vunbx6o; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 230.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k 230.5 7.7 16.5 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/69d5s4zw",
   "source_url": "https://omlx.ai/benchmarks/performance/69d5s4zw",
   "tg_tokens": null,
   "tps": 7.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-69d5s4zw-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 3vunbx6o; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 244.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 244.1 7.7 17.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/69d5s4zw",
   "source_url": "https://omlx.ai/benchmarks/performance/69d5s4zw",
   "tg_tokens": null,
   "tps": 7.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rkei32x6-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 4zd9873h, a0sbjv1h, g634v14w, ludv3awm, queriw27; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1143.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,143 73.1 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rkei32x6",
   "source_url": "https://omlx.ai/benchmarks/performance/rkei32x6",
   "tg_tokens": null,
   "tps": 73.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rkei32x6-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 4zd9873h, a0sbjv1h, g634v14w, ludv3awm, queriw27; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1365.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,365 66.8 19.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rkei32x6",
   "source_url": "https://omlx.ai/benchmarks/performance/rkei32x6",
   "tg_tokens": null,
   "tps": 66.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rkei32x6-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 4zd9873h, a0sbjv1h, g634v14w, ludv3awm, queriw27; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1417.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,417 39.6 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rkei32x6",
   "source_url": "https://omlx.ai/benchmarks/performance/rkei32x6",
   "tg_tokens": null,
   "tps": 39.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rkei32x6-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 4zd9873h, a0sbjv1h, g634v14w, ludv3awm, queriw27; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1356.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,356 62.6 19.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rkei32x6",
   "source_url": "https://omlx.ai/benchmarks/performance/rkei32x6",
   "tg_tokens": null,
   "tps": 62.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rkei32x6-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 4zd9873h, a0sbjv1h, g634v14w, ludv3awm, queriw27; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1258.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 1,258 56.1 21.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rkei32x6",
   "source_url": "https://omlx.ai/benchmarks/performance/rkei32x6",
   "tg_tokens": null,
   "tps": 56.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rkei32x6-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 4zd9873h, a0sbjv1h, g634v14w, ludv3awm, queriw27; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1044.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 1,044 46.5 23.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rkei32x6",
   "source_url": "https://omlx.ai/benchmarks/performance/rkei32x6",
   "tg_tokens": null,
   "tps": 46.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-ghjpx4wb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: 13j72ay7; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 8bit; 1k 204.7 28.0 33.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ghjpx4wb",
   "source_url": "https://omlx.ai/benchmarks/performance/ghjpx4wb",
   "tg_tokens": null,
   "tps": 28.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-ghjpx4wb-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.5 GB at this context; duplicate traces of the same session: 13j72ay7; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 211.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 8bit; 4k 211.6 30.3 37.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ghjpx4wb",
   "source_url": "https://omlx.ai/benchmarks/performance/ghjpx4wb",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-rbh2f914-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 105.1 GB; TTFT 409 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2505.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k tokens 2,505 PP tok/s 48.8 TG tok/s 409 TTFT (ms) 105.1 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rbh2f914",
   "source_url": "https://omlx.ai/benchmarks/performance/rbh2f914",
   "tg_tokens": null,
   "tps": 48.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-w1f5l7j5-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 105.1 GB; TTFT 409 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2503.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k tokens 2,503 PP tok/s 48.9 TG tok/s 409 TTFT (ms) 105.1 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w1f5l7j5",
   "source_url": "https://omlx.ai/benchmarks/performance/w1f5l7j5",
   "tg_tokens": null,
   "tps": 48.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-f0ysl7um-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: htj07qy5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 248.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 248.6 60.1 18.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f0ysl7um",
   "source_url": "https://omlx.ai/benchmarks/performance/f0ysl7um",
   "tg_tokens": null,
   "tps": 60.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-f0ysl7um-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: htj07qy5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 224.3 50.9 22.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f0ysl7um",
   "source_url": "https://omlx.ai/benchmarks/performance/f0ysl7um",
   "tg_tokens": null,
   "tps": 50.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-epqhdwst-ctx1024",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: 6459e4s3; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 154.7,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 2bit; 1k 154.7 12.0 11.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/epqhdwst",
   "source_url": "https://omlx.ai/benchmarks/performance/epqhdwst",
   "tg_tokens": null,
   "tps": 12.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-epqhdwst-ctx4096",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: 6459e4s3; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 154.5,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 2bit; 4k 154.5 11.7 13.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/epqhdwst",
   "source_url": "https://omlx.ai/benchmarks/performance/epqhdwst",
   "tg_tokens": null,
   "tps": 11.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kfbflwf8-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: u1kw1ihh, xnwyjmfv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3645.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 3,645 125.5 20.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kfbflwf8",
   "source_url": "https://omlx.ai/benchmarks/performance/kfbflwf8",
   "tg_tokens": null,
   "tps": 125.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kfbflwf8-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: u1kw1ihh, xnwyjmfv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4569.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 4,569 119.3 20.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kfbflwf8",
   "source_url": "https://omlx.ai/benchmarks/performance/kfbflwf8",
   "tg_tokens": null,
   "tps": 119.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kfbflwf8-ctx8192",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: u1kw1ihh, xnwyjmfv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4368.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 4,368 115.9 21.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kfbflwf8",
   "source_url": "https://omlx.ai/benchmarks/performance/kfbflwf8",
   "tg_tokens": null,
   "tps": 115.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fdyos7c2-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 5spz8dto; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2002.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,002 124.3 20.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fdyos7c2",
   "source_url": "https://omlx.ai/benchmarks/performance/fdyos7c2",
   "tg_tokens": null,
   "tps": 124.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fdyos7c2-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 5spz8dto; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4215.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 4,215 118.8 20.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fdyos7c2",
   "source_url": "https://omlx.ai/benchmarks/performance/fdyos7c2",
   "tg_tokens": null,
   "tps": 118.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1twvdk13-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: a19bd93o, bln45wpn, ekmwytlz, i6ct4vm3, xyb0ys2n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3545.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 3,545 84.1 16.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1twvdk13",
   "source_url": "https://omlx.ai/benchmarks/performance/1twvdk13",
   "tg_tokens": null,
   "tps": 84.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1twvdk13-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: a19bd93o, bln45wpn, ekmwytlz, i6ct4vm3, xyb0ys2n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4023.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 4,023 81.3 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1twvdk13",
   "source_url": "https://omlx.ai/benchmarks/performance/1twvdk13",
   "tg_tokens": null,
   "tps": 81.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1twvdk13-ctx8192",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: a19bd93o, bln45wpn, ekmwytlz, i6ct4vm3, xyb0ys2n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3960.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 3,960 81.5 16.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1twvdk13",
   "source_url": "https://omlx.ai/benchmarks/performance/1twvdk13",
   "tg_tokens": null,
   "tps": 81.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1twvdk13-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: a19bd93o, bln45wpn, ekmwytlz, i6ct4vm3, xyb0ys2n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3866.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 3,866 89.1 17.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1twvdk13",
   "source_url": "https://omlx.ai/benchmarks/performance/1twvdk13",
   "tg_tokens": null,
   "tps": 89.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1twvdk13-ctx32768",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: a19bd93o, bln45wpn, ekmwytlz, i6ct4vm3, xyb0ys2n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3463.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 3,463 87.3 18.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1twvdk13",
   "source_url": "https://omlx.ai/benchmarks/performance/1twvdk13",
   "tg_tokens": null,
   "tps": 87.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1twvdk13-ctx65536",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: a19bd93o, bln45wpn, ekmwytlz, i6ct4vm3, xyb0ys2n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2765.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 2,765 67.7 19.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1twvdk13",
   "source_url": "https://omlx.ai/benchmarks/performance/1twvdk13",
   "tg_tokens": null,
   "tps": 67.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-e9s11gg1-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.0 GB at this context; duplicate traces of the same session: afo0hzk7; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1252.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k 1,252 45.4 14.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e9s11gg1",
   "source_url": "https://omlx.ai/benchmarks/performance/e9s11gg1",
   "tg_tokens": null,
   "tps": 45.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-e9s11gg1-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: afo0hzk7; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1151.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 1,151 43.5 14.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e9s11gg1",
   "source_url": "https://omlx.ai/benchmarks/performance/e9s11gg1",
   "tg_tokens": null,
   "tps": 43.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-gbdj1wcp-ctx1024",
   "model": "Qwythos-9B-v2",
   "notes": "OptiQ; oMLX community benchmark trace; peak memory 7.7 GB at this context; duplicate traces of the same session: y23dxvcp; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 485.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwythos-9B-v2-OptiQ-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 485.1 37.8 7.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gbdj1wcp",
   "source_url": "https://omlx.ai/benchmarks/performance/gbdj1wcp",
   "tg_tokens": null,
   "tps": 37.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-gbdj1wcp-ctx4096",
   "model": "Qwythos-9B-v2",
   "notes": "OptiQ; oMLX community benchmark trace; peak memory 9.2 GB at this context; duplicate traces of the same session: y23dxvcp; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 423.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwythos-9B-v2-OptiQ-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 423.8 40.9 9.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gbdj1wcp",
   "source_url": "https://omlx.ai/benchmarks/performance/gbdj1wcp",
   "tg_tokens": null,
   "tps": 40.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-gbdj1wcp-ctx8192",
   "model": "Qwythos-9B-v2",
   "notes": "OptiQ; oMLX community benchmark trace; peak memory 9.6 GB at this context; duplicate traces of the same session: y23dxvcp; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 391.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwythos-9B-v2-OptiQ-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 391.3 38.7 9.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gbdj1wcp",
   "source_url": "https://omlx.ai/benchmarks/performance/gbdj1wcp",
   "tg_tokens": null,
   "tps": 38.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-gbdj1wcp-ctx16384",
   "model": "Qwythos-9B-v2",
   "notes": "OptiQ; oMLX community benchmark trace; peak memory 10.3 GB at this context; duplicate traces of the same session: y23dxvcp; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 379.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwythos-9B-v2-OptiQ-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 379.0 36.3 10.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gbdj1wcp",
   "source_url": "https://omlx.ai/benchmarks/performance/gbdj1wcp",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-gbdj1wcp-ctx32768",
   "model": "Qwythos-9B-v2",
   "notes": "OptiQ; oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: y23dxvcp; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 340.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwythos-9B-v2-OptiQ-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 32k 340.4 33.0 11.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gbdj1wcp",
   "source_url": "https://omlx.ai/benchmarks/performance/gbdj1wcp",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-gbdj1wcp-ctx65536",
   "model": "Qwythos-9B-v2",
   "notes": "OptiQ; oMLX community benchmark trace; peak memory 12.9 GB at this context; duplicate traces of the same session: y23dxvcp; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 293.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwythos-9B-v2-OptiQ-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 64k 293.1 25.3 12.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gbdj1wcp",
   "source_url": "https://omlx.ai/benchmarks/performance/gbdj1wcp",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-bmz85ku2-ctx1024",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: obke0xrf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1768.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,768 120.0 20.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bmz85ku2",
   "source_url": "https://omlx.ai/benchmarks/performance/bmz85ku2",
   "tg_tokens": null,
   "tps": 120.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-bmz85ku2-ctx4096",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: obke0xrf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2119.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,119 112.2 21.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bmz85ku2",
   "source_url": "https://omlx.ai/benchmarks/performance/bmz85ku2",
   "tg_tokens": null,
   "tps": 112.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-o6ngbi7q-ctx1024",
   "model": "gemma-4-26B-A4B-it-heretic-ara",
   "notes": "oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: 31ygb1he; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1087.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Gemma-4-26B-A4B-it-Heretic-ARA-6bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 1,087 45.0 23.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o6ngbi7q",
   "source_url": "https://omlx.ai/benchmarks/performance/o6ngbi7q",
   "tg_tokens": null,
   "tps": 45.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-o6ngbi7q-ctx4096",
   "model": "gemma-4-26B-A4B-it-heretic-ara",
   "notes": "oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 31ygb1he; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1236.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Gemma-4-26B-A4B-it-Heretic-ARA-6bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 1,236 39.0 23.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o6ngbi7q",
   "source_url": "https://omlx.ai/benchmarks/performance/o6ngbi7q",
   "tg_tokens": null,
   "tps": 39.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-zandbs6r-ctx1024",
   "model": "Gemma-4-31B-heretic-ara",
   "notes": "Google Template build; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: syj8jy60; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 181.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Gemma-4-31B-Heretic-ARA-6bit-GoogleTemplate on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 181.7 14.2 26.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zandbs6r",
   "source_url": "https://omlx.ai/benchmarks/performance/zandbs6r",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-zandbs6r-ctx4096",
   "model": "Gemma-4-31B-heretic-ara",
   "notes": "Google Template build; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: syj8jy60; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 188.2,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Gemma-4-31B-Heretic-ARA-6bit-GoogleTemplate on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 188.2 14.6 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zandbs6r",
   "source_url": "https://omlx.ai/benchmarks/performance/zandbs6r",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-97s3gzu3-ctx1024",
   "model": "Huihui-Qwen3.6-35B-A3B-Claude-4.6-Opus-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: ym7q5vh1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4232.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Huihui-Qwen3.6-35B-A3B-Claude-4.6-Opus-abliterated-mlx-oQ8-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 4,232 99.7 35.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/97s3gzu3",
   "source_url": "https://omlx.ai/benchmarks/performance/97s3gzu3",
   "tg_tokens": null,
   "tps": 99.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-97s3gzu3-ctx4096",
   "model": "Huihui-Qwen3.6-35B-A3B-Claude-4.6-Opus-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: ym7q5vh1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5824.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Huihui-Qwen3.6-35B-A3B-Claude-4.6-Opus-abliterated-mlx-oQ8-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 5,824 95.1 36.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/97s3gzu3",
   "source_url": "https://omlx.ai/benchmarks/performance/97s3gzu3",
   "tg_tokens": null,
   "tps": 95.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-97s3gzu3-ctx8192",
   "model": "Huihui-Qwen3.6-35B-A3B-Claude-4.6-Opus-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.2 GB at this context; duplicate traces of the same session: ym7q5vh1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5699.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Huihui-Qwen3.6-35B-A3B-Claude-4.6-Opus-abliterated-mlx-oQ8-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 5,699 95.0 36.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/97s3gzu3",
   "source_url": "https://omlx.ai/benchmarks/performance/97s3gzu3",
   "tg_tokens": null,
   "tps": 95.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-nfo7xxbi-ctx1024",
   "model": "Huihui-Qwen3.6-35B-A3B-Claude-4.6-Opus-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 35.5 GB; TTFT 249 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4106.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Huihui-Qwen3.6-35B-A3B-Claude-4.6-Opus-abliterated-mlx-oQ8-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k tokens 4,106 PP tok/s 93.1 TG tok/s 249 TTFT (ms) 35.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nfo7xxbi",
   "source_url": "https://omlx.ai/benchmarks/performance/nfo7xxbi",
   "tg_tokens": null,
   "tps": 93.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2c0x6vmf-ctx1024",
   "model": "Huihui-Qwen3.6-35B-A3B-Claude-4.6-Opus-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 35.5 GB; TTFT 237 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4325.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Huihui-Qwen3.6-35B-A3B-Claude-4.6-Opus-abliterated-mlx-oQ8-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k tokens 4,325 PP tok/s 95.6 TG tok/s 237 TTFT (ms) 35.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2c0x6vmf",
   "source_url": "https://omlx.ai/benchmarks/performance/2c0x6vmf",
   "tg_tokens": null,
   "tps": 95.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-xqf7ynfd-ctx1024",
   "model": "LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated-v3",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 9.4 GB; TTFT 676 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1516.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated-v3-8bit-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 8bit; 1k tokens 1,516 PP tok/s 104.7 TG tok/s 676 TTFT (ms) 9.4 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xqf7ynfd",
   "source_url": "https://omlx.ai/benchmarks/performance/xqf7ynfd",
   "tg_tokens": null,
   "tps": 104.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-3xamsop5-ctx1024",
   "model": "LFM2.5-8B-A1B",
   "notes": "oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: tb2rwihj, z4930rf8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3549.0,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "LFM2.5-8B-A1B-MXFP4 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 3,549 272.8 4.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3xamsop5",
   "source_url": "https://omlx.ai/benchmarks/performance/3xamsop5",
   "tg_tokens": null,
   "tps": 272.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-3xamsop5-ctx4096",
   "model": "LFM2.5-8B-A1B",
   "notes": "oMLX community benchmark trace; peak memory 5.1 GB at this context; duplicate traces of the same session: tb2rwihj, z4930rf8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4071.0,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "LFM2.5-8B-A1B-MXFP4 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 4,071 264.3 5.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3xamsop5",
   "source_url": "https://omlx.ai/benchmarks/performance/3xamsop5",
   "tg_tokens": null,
   "tps": 264.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-3xamsop5-ctx8192",
   "model": "LFM2.5-8B-A1B",
   "notes": "oMLX community benchmark trace; peak memory 5.1 GB at this context; duplicate traces of the same session: tb2rwihj, z4930rf8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3968.0,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "LFM2.5-8B-A1B-MXFP4 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 3,968 256.2 5.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3xamsop5",
   "source_url": "https://omlx.ai/benchmarks/performance/3xamsop5",
   "tg_tokens": null,
   "tps": 256.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-23qumohy-ctx1024",
   "model": "LFM2.5-VL-3B",
   "notes": "oMLX community benchmark trace; peak memory 3.1 GB at this context; duplicate traces of the same session: jf4dp6p7, t4gtytlo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2560.0,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "LFM2.5-VL-3B-MXFP4 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 2,560 250.4 3.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/23qumohy",
   "source_url": "https://omlx.ai/benchmarks/performance/23qumohy",
   "tg_tokens": null,
   "tps": 250.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-23qumohy-ctx4096",
   "model": "LFM2.5-VL-3B",
   "notes": "oMLX community benchmark trace; peak memory 3.3 GB at this context; duplicate traces of the same session: jf4dp6p7, t4gtytlo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2661.0,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "LFM2.5-VL-3B-MXFP4 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 2,661 236.4 3.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/23qumohy",
   "source_url": "https://omlx.ai/benchmarks/performance/23qumohy",
   "tg_tokens": null,
   "tps": 236.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-23qumohy-ctx8192",
   "model": "LFM2.5-VL-3B",
   "notes": "oMLX community benchmark trace; peak memory 3.3 GB at this context; duplicate traces of the same session: jf4dp6p7, t4gtytlo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2599.0,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "LFM2.5-VL-3B-MXFP4 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 2,599 225.4 3.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/23qumohy",
   "source_url": "https://omlx.ai/benchmarks/performance/23qumohy",
   "tg_tokens": null,
   "tps": 225.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-4j4rj2l6-ctx131072",
   "model": "LFM2.5-VL-450M",
   "notes": "oMLX community benchmark trace; peak memory 2.4 GB at this context; duplicate traces of the same session: qk1k8iwa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4713.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-VL-450M-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 128k 4,713 174.2 2.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4j4rj2l6",
   "source_url": "https://omlx.ai/benchmarks/performance/4j4rj2l6",
   "tg_tokens": null,
   "tps": 174.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-4j4rj2l6-ctx199680",
   "model": "LFM2.5-VL-450M",
   "notes": "oMLX community benchmark trace; peak memory 3.2 GB at this context; duplicate traces of the same session: qk1k8iwa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3290.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-VL-450M-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 195k 3,290 122.2 3.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4j4rj2l6",
   "source_url": "https://omlx.ai/benchmarks/performance/4j4rj2l6",
   "tg_tokens": null,
   "tps": 122.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-te7tqdcn-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: us5eteqc; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1753.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 1,753 57.5 20.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/te7tqdcn",
   "source_url": "https://omlx.ai/benchmarks/performance/te7tqdcn",
   "tg_tokens": null,
   "tps": 57.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-te7tqdcn-ctx131072",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: us5eteqc; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 734.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 128k 734.6 26.1 22.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/te7tqdcn",
   "source_url": "https://omlx.ai/benchmarks/performance/te7tqdcn",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-emuvhewq-ctx1024",
   "model": "Qwen3 14B",
   "notes": "oMLX community benchmark trace; peak memory 15.3 GB at this context; duplicate traces of the same session: 0ouhw6l9; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 641.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-14B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 8bit; 1k 641.4 15.8 15.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/emuvhewq",
   "source_url": "https://omlx.ai/benchmarks/performance/emuvhewq",
   "tg_tokens": null,
   "tps": 15.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-emuvhewq-ctx4096",
   "model": "Qwen3 14B",
   "notes": "oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: 0ouhw6l9; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 691.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-14B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 8bit; 4k 691.2 14.6 15.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/emuvhewq",
   "source_url": "https://omlx.ai/benchmarks/performance/emuvhewq",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-4u7wn5al-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: 64gvcwrz; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 763.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 763.7 25.8 6.3 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4u7wn5al",
   "source_url": "https://omlx.ai/benchmarks/performance/4u7wn5al",
   "tg_tokens": null,
   "tps": 25.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-4u7wn5al-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.6 GB at this context; duplicate traces of the same session: 64gvcwrz; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 738.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 16k 738.7 24.0 6.6 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4u7wn5al",
   "source_url": "https://omlx.ai/benchmarks/performance/4u7wn5al",
   "tg_tokens": null,
   "tps": 24.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-w9s36r60-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 6lmu524l; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 641.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k 641.6 56.8 19.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w9s36r60",
   "source_url": "https://omlx.ai/benchmarks/performance/w9s36r60",
   "tg_tokens": null,
   "tps": 56.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-w9s36r60-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 6lmu524l; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1321.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 1,321 55.0 19.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w9s36r60",
   "source_url": "https://omlx.ai/benchmarks/performance/w9s36r60",
   "tg_tokens": null,
   "tps": 55.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-57lm71dj-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: hix7dq7d; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 576.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 576.1 71.3 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/57lm71dj",
   "source_url": "https://omlx.ai/benchmarks/performance/57lm71dj",
   "tg_tokens": null,
   "tps": 71.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-57lm71dj-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: hix7dq7d; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 592.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 592.2 68.9 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/57lm71dj",
   "source_url": "https://omlx.ai/benchmarks/performance/57lm71dj",
   "tg_tokens": null,
   "tps": 68.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-pfnip1pc-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 8ytrb1l8; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 547.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 547.3 71.3 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pfnip1pc",
   "source_url": "https://omlx.ai/benchmarks/performance/pfnip1pc",
   "tg_tokens": null,
   "tps": 71.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-pfnip1pc-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 8ytrb1l8; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 588.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 588.8 69.3 21.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pfnip1pc",
   "source_url": "https://omlx.ai/benchmarks/performance/pfnip1pc",
   "tg_tokens": null,
   "tps": 69.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-g7wfzdm4-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: sl9nqqu6; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 567.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 1k 567.7 58.5 35.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g7wfzdm4",
   "source_url": "https://omlx.ai/benchmarks/performance/g7wfzdm4",
   "tg_tokens": null,
   "tps": 58.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-g7wfzdm4-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.6 GB at this context; duplicate traces of the same session: sl9nqqu6; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 616.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 4k 616.9 56.7 37.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g7wfzdm4",
   "source_url": "https://omlx.ai/benchmarks/performance/g7wfzdm4",
   "tg_tokens": null,
   "tps": 56.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-v3s41nxu-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 8nedgv9s; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1850.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,850 70.0 20.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v3s41nxu",
   "source_url": "https://omlx.ai/benchmarks/performance/v3s41nxu",
   "tg_tokens": null,
   "tps": 70.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-v3s41nxu-ctx16384",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 8nedgv9s; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1683.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 1,683 61.6 20.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v3s41nxu",
   "source_url": "https://omlx.ai/benchmarks/performance/v3s41nxu",
   "tg_tokens": null,
   "tps": 61.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-f162q8os-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: jb0gostg; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1843.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,843 70.0 20.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f162q8os",
   "source_url": "https://omlx.ai/benchmarks/performance/f162q8os",
   "tg_tokens": null,
   "tps": 70.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-f162q8os-ctx16384",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: jb0gostg; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1683.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 1,683 61.6 20.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f162q8os",
   "source_url": "https://omlx.ai/benchmarks/performance/f162q8os",
   "tg_tokens": null,
   "tps": 61.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-lefumc6z-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 81pnp113; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1839.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,839 69.4 20.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lefumc6z",
   "source_url": "https://omlx.ai/benchmarks/performance/lefumc6z",
   "tg_tokens": null,
   "tps": 69.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-lefumc6z-ctx16384",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 81pnp113; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1681.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 1,681 61.6 20.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lefumc6z",
   "source_url": "https://omlx.ai/benchmarks/performance/lefumc6z",
   "tg_tokens": null,
   "tps": 61.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-0flyk6ry-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: h21t4rat; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1813.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,813 69.8 20.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0flyk6ry",
   "source_url": "https://omlx.ai/benchmarks/performance/0flyk6ry",
   "tg_tokens": null,
   "tps": 69.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-0flyk6ry-ctx16384",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: h21t4rat; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1677.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 1,677 61.3 20.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0flyk6ry",
   "source_url": "https://omlx.ai/benchmarks/performance/0flyk6ry",
   "tg_tokens": null,
   "tps": 61.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-gt6i8ny5-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: kt25b23b; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1843.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,843 69.5 20.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gt6i8ny5",
   "source_url": "https://omlx.ai/benchmarks/performance/gt6i8ny5",
   "tg_tokens": null,
   "tps": 69.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-gt6i8ny5-ctx16384",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: kt25b23b; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1681.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 1,681 61.7 20.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gt6i8ny5",
   "source_url": "https://omlx.ai/benchmarks/performance/gt6i8ny5",
   "tg_tokens": null,
   "tps": 61.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-4wdh2nlx-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 253.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 253.9 26.8 17.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4wdh2nlx",
   "source_url": "https://omlx.ai/benchmarks/performance/4wdh2nlx",
   "tg_tokens": null,
   "tps": 26.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-4wdh2nlx-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 244.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 244.7 25.2 17.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4wdh2nlx",
   "source_url": "https://omlx.ai/benchmarks/performance/4wdh2nlx",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-go9mfbz5-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 1kop1gbk; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 84.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 84.4 15.5 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/go9mfbz5",
   "source_url": "https://omlx.ai/benchmarks/performance/go9mfbz5",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-go9mfbz5-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 1kop1gbk; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 78.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 78.5 15.0 20.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/go9mfbz5",
   "source_url": "https://omlx.ai/benchmarks/performance/go9mfbz5",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-nctpt34u-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 13.3 GB; TTFT 7601 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 134.7,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-27B-oQ2.7e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 2bit; 1k tokens 134.7 PP tok/s 16.0 TG tok/s 7601 TTFT (ms) 13.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nctpt34u",
   "source_url": "https://omlx.ai/benchmarks/performance/nctpt34u",
   "tg_tokens": null,
   "tps": 16.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-kli63kih-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 13.3 GB; TTFT 7536 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 135.9,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-27B-oQ2.7e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 2bit; 1k tokens 135.9 PP tok/s 15.8 TG tok/s 7536 TTFT (ms) 13.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kli63kih",
   "source_url": "https://omlx.ai/benchmarks/performance/kli63kih",
   "tg_tokens": null,
   "tps": 15.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-olvvw46o-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 13.3 GB; TTFT 7412 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 138.2,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-27B-oQ2.7e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 2bit; 1k tokens 138.2 PP tok/s 16.0 TG tok/s 7412 TTFT (ms) 13.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/olvvw46o",
   "source_url": "https://omlx.ai/benchmarks/performance/olvvw46o",
   "tg_tokens": null,
   "tps": 16.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-lnnd70ad-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 13.3 GB; TTFT 7337 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 139.6,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-27B-oQ2.7e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 2bit; 1k tokens 139.6 PP tok/s 16.1 TG tok/s 7337 TTFT (ms) 13.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lnnd70ad",
   "source_url": "https://omlx.ai/benchmarks/performance/lnnd70ad",
   "tg_tokens": null,
   "tps": 16.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-08a92dvb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 16.9 GB; TTFT 9383 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 109.1,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k tokens 109.1 PP tok/s 18.5 TG tok/s 9383 TTFT (ms) 16.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/08a92dvb",
   "source_url": "https://omlx.ai/benchmarks/performance/08a92dvb",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-v2xapvb9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: cuhflaqk; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 226.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 226.3 17.3 19.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v2xapvb9",
   "source_url": "https://omlx.ai/benchmarks/performance/v2xapvb9",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-v2xapvb9-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: cuhflaqk; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 220.5 22.4 19.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v2xapvb9",
   "source_url": "https://omlx.ai/benchmarks/performance/v2xapvb9",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-pzw1o1yh-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: 7e8l6464; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 82.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 1k 82.9 14.6 33.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pzw1o1yh",
   "source_url": "https://omlx.ai/benchmarks/performance/pzw1o1yh",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-pzw1o1yh-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.0 GB at this context; duplicate traces of the same session: 7e8l6464; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 81.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 4k 81.6 14.8 38.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pzw1o1yh",
   "source_url": "https://omlx.ai/benchmarks/performance/pzw1o1yh",
   "tg_tokens": null,
   "tps": 14.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-yl5zv86f-ctx16384",
   "model": "Qwen3.8-Flash-Next-REAP320",
   "notes": "DWQ MTP Vision build; oMLX community benchmark trace, single run at 16k context; peak memory 51.5 GB; TTFT 120922 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 135.5,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-Flash-Next-REAP320-oQ3e-fp16-DWQ-MTP-Vision-MLX on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 3bit; 16k tokens 135.5 PP tok/s 8.5 TG tok/s 120922 TTFT (ms) 51.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yl5zv86f",
   "source_url": "https://omlx.ai/benchmarks/performance/yl5zv86f",
   "tg_tokens": null,
   "tps": 8.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ubn0drz5-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 73.8 GB at this context; duplicate traces of the same session: 5zzna3qo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 423.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 423.3 18.8 73.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ubn0drz5",
   "source_url": "https://omlx.ai/benchmarks/performance/ubn0drz5",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ubn0drz5-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 81.1 GB at this context; duplicate traces of the same session: 5zzna3qo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 472.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 472.5 20.3 81.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ubn0drz5",
   "source_url": "https://omlx.ai/benchmarks/performance/ubn0drz5",
   "tg_tokens": null,
   "tps": 20.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-4pjtd5rq-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: fhsqm5h7, p430uylf; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 659.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 1k 659.3 76.2 100.5 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4pjtd5rq",
   "source_url": "https://omlx.ai/benchmarks/performance/4pjtd5rq",
   "tg_tokens": null,
   "tps": 76.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-4pjtd5rq-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.4 GB at this context; duplicate traces of the same session: fhsqm5h7, p430uylf; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 665.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 4k 665.4 67.0 102.4 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4pjtd5rq",
   "source_url": "https://omlx.ai/benchmarks/performance/4pjtd5rq",
   "tg_tokens": null,
   "tps": 67.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-4pjtd5rq-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.9 GB at this context; duplicate traces of the same session: fhsqm5h7, p430uylf; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 626.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 8k 626.6 73.9 103.9 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4pjtd5rq",
   "source_url": "https://omlx.ai/benchmarks/performance/4pjtd5rq",
   "tg_tokens": null,
   "tps": 73.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-egaepo44-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: b0qedgnm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 253.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 253.6 26.5 16.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/egaepo44",
   "source_url": "https://omlx.ai/benchmarks/performance/egaepo44",
   "tg_tokens": null,
   "tps": 26.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-egaepo44-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: b0qedgnm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 211.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 211.6 22.0 19.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/egaepo44",
   "source_url": "https://omlx.ai/benchmarks/performance/egaepo44",
   "tg_tokens": null,
   "tps": 22.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-98kkogni-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: gc6b0l36; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 244.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 244.3 48.8 18.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/98kkogni",
   "source_url": "https://omlx.ai/benchmarks/performance/98kkogni",
   "tg_tokens": null,
   "tps": 48.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-98kkogni-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: gc6b0l36; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 222.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 222.6 47.1 22.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/98kkogni",
   "source_url": "https://omlx.ai/benchmarks/performance/98kkogni",
   "tg_tokens": null,
   "tps": 47.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-4tljl41h-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 7zcjsc17, hcfu42qh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 252.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 252.1 47.2 21.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4tljl41h",
   "source_url": "https://omlx.ai/benchmarks/performance/4tljl41h",
   "tg_tokens": null,
   "tps": 47.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-4tljl41h-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: 7zcjsc17, hcfu42qh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 282.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 282.1 39.9 25.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4tljl41h",
   "source_url": "https://omlx.ai/benchmarks/performance/4tljl41h",
   "tg_tokens": null,
   "tps": 39.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-4tljl41h-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: 7zcjsc17, hcfu42qh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 277.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 277.7 50.7 25.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4tljl41h",
   "source_url": "https://omlx.ai/benchmarks/performance/4tljl41h",
   "tg_tokens": null,
   "tps": 50.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-qs8lz79x-ctx1024",
   "model": "Swift-Qwen3.8-27b-oQ3e",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 16.2 GB at this context; duplicate traces of the same session: egq9j4rv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 121.5,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Swift-Qwen3.8-27b-oQ3e-fp16-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 3bit; 1k 121.5 19.0 16.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qs8lz79x",
   "source_url": "https://omlx.ai/benchmarks/performance/qs8lz79x",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-qs8lz79x-ctx8192",
   "model": "Swift-Qwen3.8-27b-oQ3e",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 17.5 GB at this context; duplicate traces of the same session: egq9j4rv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 120.2,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Swift-Qwen3.8-27b-oQ3e-fp16-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 3bit; 8k 120.2 22.3 17.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qs8lz79x",
   "source_url": "https://omlx.ai/benchmarks/performance/qs8lz79x",
   "tg_tokens": null,
   "tps": 22.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-mnn6ogvr-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; Swift 1.5 build; oMLX community benchmark trace; peak memory 101.8 GB at this context; duplicate traces of the same session: aegzfx3l, p595jsqd, w3nf2nrm; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 792.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 1k 792.2 55.0 101.8 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mnn6ogvr",
   "source_url": "https://omlx.ai/benchmarks/performance/mnn6ogvr",
   "tg_tokens": null,
   "tps": 55.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-mnn6ogvr-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; Swift 1.5 build; oMLX community benchmark trace; peak memory 104.4 GB at this context; duplicate traces of the same session: aegzfx3l, p595jsqd, w3nf2nrm; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 809.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 4k 809.4 62.9 104.4 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mnn6ogvr",
   "source_url": "https://omlx.ai/benchmarks/performance/mnn6ogvr",
   "tg_tokens": null,
   "tps": 62.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-mnn6ogvr-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; Swift 1.5 build; oMLX community benchmark trace; peak memory 105.4 GB at this context; duplicate traces of the same session: aegzfx3l, p595jsqd, w3nf2nrm; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 771.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 8k 771.5 46.0 105.4 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mnn6ogvr",
   "source_url": "https://omlx.ai/benchmarks/performance/mnn6ogvr",
   "tg_tokens": null,
   "tps": 46.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-mnn6ogvr-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; Swift 1.5 build; oMLX community benchmark trace; peak memory 105.8 GB at this context; duplicate traces of the same session: aegzfx3l, p595jsqd, w3nf2nrm; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 740.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 16k 740.8 56.6 105.8 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mnn6ogvr",
   "source_url": "https://omlx.ai/benchmarks/performance/mnn6ogvr",
   "tg_tokens": null,
   "tps": 56.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-9isuns18-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: 1zbf3epw, bvhooiut, vqkiorkt; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 653.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 1k 653.6 77.4 100.6 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9isuns18",
   "source_url": "https://omlx.ai/benchmarks/performance/9isuns18",
   "tg_tokens": null,
   "tps": 77.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-9isuns18-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 102.4 GB at this context; duplicate traces of the same session: 1zbf3epw, bvhooiut, vqkiorkt; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 659.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 4k 659.6 59.3 102.4 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9isuns18",
   "source_url": "https://omlx.ai/benchmarks/performance/9isuns18",
   "tg_tokens": null,
   "tps": 59.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-9isuns18-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 103.9 GB at this context; duplicate traces of the same session: 1zbf3epw, bvhooiut, vqkiorkt; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 619.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 8k 619.2 77.4 103.9 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9isuns18",
   "source_url": "https://omlx.ai/benchmarks/performance/9isuns18",
   "tg_tokens": null,
   "tps": 77.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-9isuns18-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 104.2 GB at this context; duplicate traces of the same session: 1zbf3epw, bvhooiut, vqkiorkt; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 602.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 16k 602.9 75.1 104.2 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9isuns18",
   "source_url": "https://omlx.ai/benchmarks/performance/9isuns18",
   "tg_tokens": null,
   "tps": 75.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-m6afw28u-ctx1024",
   "model": "clef-flash",
   "notes": "generic trace name; oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: s8pigt2r; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 201.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "clef-flash-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k 201.9 21.2 5.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m6afw28u",
   "source_url": "https://omlx.ai/benchmarks/performance/m6afw28u",
   "tg_tokens": null,
   "tps": 21.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-m6afw28u-ctx4096",
   "model": "clef-flash",
   "notes": "generic trace name; oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: s8pigt2r; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 207.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "clef-flash-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 207.3 20.5 6.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m6afw28u",
   "source_url": "https://omlx.ai/benchmarks/performance/m6afw28u",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-80onvkdj-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: 0quxg4bv; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1508.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 1,508 50.6 14.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/80onvkdj",
   "source_url": "https://omlx.ai/benchmarks/performance/80onvkdj",
   "tg_tokens": null,
   "tps": 50.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-80onvkdj-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 14.5 GB at this context; duplicate traces of the same session: 0quxg4bv; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1560.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,560 49.9 14.5 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/80onvkdj",
   "source_url": "https://omlx.ai/benchmarks/performance/80onvkdj",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-86j4plxh-ctx4096",
   "model": "Maple Preview",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 6.3 GB; TTFT 1098 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3731.0,
   "provenance": "community",
   "quant": "2bit",
   "quote": "maple-preview-2bit-mlx on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 2bit; 4k tokens 3,731 PP tok/s 328.8 TG tok/s 1098 TTFT (ms) 6.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/86j4plxh",
   "source_url": "https://omlx.ai/benchmarks/performance/86j4plxh",
   "tg_tokens": null,
   "tps": 328.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-0v2dujkg-ctx1024",
   "model": "Qwen3-VL-2B-Instruct",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 1.7 GB; TTFT 223 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4599.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3-VL-2B-Instruct-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k tokens 4,599 PP tok/s 189.1 TG tok/s 223 TTFT (ms) 1.7 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0v2dujkg",
   "source_url": "https://omlx.ai/benchmarks/performance/0v2dujkg",
   "tg_tokens": null,
   "tps": 189.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-87jn4xje-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.7 GB at this context; duplicate traces of the same session: 0wyg11gu, 74ukzw34, l7sd94vu; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 194.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ornith-ai--Ornith-1.5-9B-MLX-4bit on M3 (10c); M3 (10c) \u00b7 16 GB \u00b7 4bit; 1k 194.7 18.8 5.7 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/87jn4xje",
   "source_url": "https://omlx.ai/benchmarks/performance/87jn4xje",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-87jn4xje-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: 0wyg11gu, 74ukzw34, l7sd94vu; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 202.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ornith-ai--Ornith-1.5-9B-MLX-4bit on M3 (10c); M3 (10c) \u00b7 16 GB \u00b7 4bit; 4k 202.5 18.7 6.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/87jn4xje",
   "source_url": "https://omlx.ai/benchmarks/performance/87jn4xje",
   "tg_tokens": null,
   "tps": 18.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-87jn4xje-ctx8192",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: 0wyg11gu, 74ukzw34, l7sd94vu; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 195.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ornith-ai--Ornith-1.5-9B-MLX-4bit on M3 (10c); M3 (10c) \u00b7 16 GB \u00b7 4bit; 8k 195.0 18.1 6.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/87jn4xje",
   "source_url": "https://omlx.ai/benchmarks/performance/87jn4xje",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-87jn4xje-ctx16384",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: 0wyg11gu, 74ukzw34, l7sd94vu; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 177.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ornith-ai--Ornith-1.5-9B-MLX-4bit on M3 (10c); M3 (10c) \u00b7 16 GB \u00b7 4bit; 16k 177.6 15.4 6.9 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/87jn4xje",
   "source_url": "https://omlx.ai/benchmarks/performance/87jn4xje",
   "tg_tokens": null,
   "tps": 15.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-in6yb5gj-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.5 GB at this context; duplicate traces of the same session: syde1l9v; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 468.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "ornith-ai--Ornith-1.5-9B-MLX-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 24 GB \u00b7 8bit; 4k 468.9 27.9 10.5 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/in6yb5gj",
   "source_url": "https://omlx.ai/benchmarks/performance/in6yb5gj",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-in6yb5gj-ctx16384",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.0 GB at this context; duplicate traces of the same session: syde1l9v; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 441.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "ornith-ai--Ornith-1.5-9B-MLX-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 24 GB \u00b7 8bit; 16k 441.3 26.3 11.0 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/in6yb5gj",
   "source_url": "https://omlx.ai/benchmarks/performance/in6yb5gj",
   "tg_tokens": null,
   "tps": 26.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-qbrr4f77-ctx1024",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "prism-ml MLX build; oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: h8mnpezd; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 126.7,
   "provenance": "community",
   "quant": "2bit",
   "quote": "prism-ml--Ternary-Bonsai-2-27B-mlx-2bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 unknown; 1k 126.7 10.1 11.5 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qbrr4f77",
   "source_url": "https://omlx.ai/benchmarks/performance/qbrr4f77",
   "tg_tokens": null,
   "tps": 10.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-qbrr4f77-ctx4096",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "prism-ml MLX build; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: h8mnpezd; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 123.8,
   "provenance": "community",
   "quant": "2bit",
   "quote": "prism-ml--Ternary-Bonsai-2-27B-mlx-2bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 unknown; 4k 123.8 9.6 13.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qbrr4f77",
   "source_url": "https://omlx.ai/benchmarks/performance/qbrr4f77",
   "tg_tokens": null,
   "tps": 9.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ckfp1mvi-ctx1024",
   "model": "Nanbeige4.2-3B",
   "notes": "oMLX community benchmark trace; peak memory 8.3 GB at this context; duplicate traces of the same session: 8r1h350r, n0o6ozpn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1586.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Nanbeige4.2-3B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 unknown; 1k 1,586 25.2 8.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ckfp1mvi",
   "source_url": "https://omlx.ai/benchmarks/performance/ckfp1mvi",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ckfp1mvi-ctx4096",
   "model": "Nanbeige4.2-3B",
   "notes": "oMLX community benchmark trace; peak memory 9.0 GB at this context; duplicate traces of the same session: 8r1h350r, n0o6ozpn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1515.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Nanbeige4.2-3B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 unknown; 4k 1,515 25.7 9.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ckfp1mvi",
   "source_url": "https://omlx.ai/benchmarks/performance/ckfp1mvi",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ckfp1mvi-ctx32768",
   "model": "Nanbeige4.2-3B",
   "notes": "oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: 8r1h350r, n0o6ozpn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 752.1,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Nanbeige4.2-3B on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 unknown; 32k 752.1 19.7 13.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ckfp1mvi",
   "source_url": "https://omlx.ai/benchmarks/performance/ckfp1mvi",
   "tg_tokens": null,
   "tps": 19.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-socaytyx-ctx1024",
   "model": "Qwen2.5-Coder-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 4.5 GB at this context; duplicate traces of the same session: m7fcwzh6; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1132.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-7B-Instruct-4bit on M6 (12c); M6 (12c) \u00b7 16 GB \u00b7 4bit; 1k 1,132 32.0 4.5 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/socaytyx",
   "source_url": "https://omlx.ai/benchmarks/performance/socaytyx",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-socaytyx-ctx4096",
   "model": "Qwen2.5-Coder-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: m7fcwzh6; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1163.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-7B-Instruct-4bit on M6 (12c); M6 (12c) \u00b7 16 GB \u00b7 4bit; 4k 1,163 31.0 4.9 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/socaytyx",
   "source_url": "https://omlx.ai/benchmarks/performance/socaytyx",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2",
   "id": "omlx-p4f1d13a-ctx4096",
   "model": "Qwen3 14B",
   "notes": "oMLX community benchmark trace; peak memory 8.9 GB at this context; duplicate traces of the same session: jehzq4b4; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 70.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-14B-4bit on M2 (10c); M2 (10c) \u00b7 16 GB \u00b7 4bit; 4k 70.7 10.1 8.9 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p4f1d13a",
   "source_url": "https://omlx.ai/benchmarks/performance/p4f1d13a",
   "tg_tokens": null,
   "tps": 10.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2",
   "id": "omlx-p4f1d13a-ctx8192",
   "model": "Qwen3 14B",
   "notes": "oMLX community benchmark trace; peak memory 9.5 GB at this context; duplicate traces of the same session: jehzq4b4; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 67.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-14B-4bit on M2 (10c); M2 (10c) \u00b7 16 GB \u00b7 4bit; 8k 67.1 9.2 9.5 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p4f1d13a",
   "source_url": "https://omlx.ai/benchmarks/performance/p4f1d13a",
   "tg_tokens": null,
   "tps": 9.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-q6ya9xs1-ctx4096",
   "model": "Qwen3 4B",
   "notes": "oMLX community benchmark trace; peak memory 3.2 GB at this context; duplicate traces of the same session: ss8x9929; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 305.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-4B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 305.8 29.4 3.2 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q6ya9xs1",
   "source_url": "https://omlx.ai/benchmarks/performance/q6ya9xs1",
   "tg_tokens": null,
   "tps": 29.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-q6ya9xs1-ctx16384",
   "model": "Qwen3 4B",
   "notes": "oMLX community benchmark trace; peak memory 4.8 GB at this context; duplicate traces of the same session: ss8x9929; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 211.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-4B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 16k 211.8 16.2 4.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q6ya9xs1",
   "source_url": "https://omlx.ai/benchmarks/performance/q6ya9xs1",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-drhf6zax-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.7 GB at this context; duplicate traces of the same session: c13g7q8q; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1345.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 8k 1,345 42.0 4.7 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/drhf6zax",
   "source_url": "https://omlx.ai/benchmarks/performance/drhf6zax",
   "tg_tokens": null,
   "tps": 42.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-drhf6zax-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 5.0 GB at this context; duplicate traces of the same session: c13g7q8q; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1261.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 16k 1,261 48.8 5.0 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/drhf6zax",
   "source_url": "https://omlx.ai/benchmarks/performance/drhf6zax",
   "tg_tokens": null,
   "tps": 48.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-jiqgece3-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: tftl6ypm; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 862.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 862.3 25.7 6.3 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jiqgece3",
   "source_url": "https://omlx.ai/benchmarks/performance/jiqgece3",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-jiqgece3-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.6 GB at this context; duplicate traces of the same session: tftl6ypm; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 843.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 16k 843.1 23.4 6.6 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jiqgece3",
   "source_url": "https://omlx.ai/benchmarks/performance/jiqgece3",
   "tg_tokens": null,
   "tps": 23.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-xbs2r76k-ctx4096",
   "model": "Qwen3.5-9B-Uncensored-HauhauCS-Aggressive",
   "notes": "oMLX community benchmark trace; peak memory 6.0 GB at this context; duplicate traces of the same session: lzorlqak; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1159.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.5-9B-Uncensored-HauhauCS-Aggressive-MLX-mxfp4 on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,159 33.0 6.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xbs2r76k",
   "source_url": "https://omlx.ai/benchmarks/performance/xbs2r76k",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-xbs2r76k-ctx16384",
   "model": "Qwen3.5-9B-Uncensored-HauhauCS-Aggressive",
   "notes": "oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: lzorlqak; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1102.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.5-9B-Uncensored-HauhauCS-Aggressive-MLX-mxfp4 on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 1,102 30.2 6.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xbs2r76k",
   "source_url": "https://omlx.ai/benchmarks/performance/xbs2r76k",
   "tg_tokens": null,
   "tps": 30.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-mhlen505-ctx1024",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: 46nmabcw; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 475.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 475.1 18.0 15.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mhlen505",
   "source_url": "https://omlx.ai/benchmarks/performance/mhlen505",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-mhlen505-ctx4096",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 46nmabcw; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 487.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 487.0 17.7 17.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mhlen505",
   "source_url": "https://omlx.ai/benchmarks/performance/mhlen505",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-igsjkzzh-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1955.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 1,955 99.2 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/igsjkzzh",
   "source_url": "https://omlx.ai/benchmarks/performance/igsjkzzh",
   "tg_tokens": null,
   "tps": 99.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-igsjkzzh-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2417.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,417 97.0 20.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/igsjkzzh",
   "source_url": "https://omlx.ai/benchmarks/performance/igsjkzzh",
   "tg_tokens": null,
   "tps": 97.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-z75a6nr1-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: eiajp7a8; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 581.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 1k 581.6 57.5 35.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z75a6nr1",
   "source_url": "https://omlx.ai/benchmarks/performance/z75a6nr1",
   "tg_tokens": null,
   "tps": 57.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-z75a6nr1-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.6 GB at this context; duplicate traces of the same session: eiajp7a8; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 605.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 4k 605.4 55.8 37.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z75a6nr1",
   "source_url": "https://omlx.ai/benchmarks/performance/z75a6nr1",
   "tg_tokens": null,
   "tps": 55.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-o1bk5xyh-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 5wnr3u0t; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 194.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k 194.1 18.5 17.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o1bk5xyh",
   "source_url": "https://omlx.ai/benchmarks/performance/o1bk5xyh",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-o1bk5xyh-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: 5wnr3u0t; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 185.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 185.4 15.7 18.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o1bk5xyh",
   "source_url": "https://omlx.ai/benchmarks/performance/o1bk5xyh",
   "tg_tokens": null,
   "tps": 15.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-84f101bx-ctx1024",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 3c64858c, fxjk8y78, lm5gyjct, r4ngqpdh, x5pacpv3; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 862.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 862.0 102.1 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/84f101bx",
   "source_url": "https://omlx.ai/benchmarks/performance/84f101bx",
   "tg_tokens": null,
   "tps": 102.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-84f101bx-ctx4096",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: 3c64858c, fxjk8y78, lm5gyjct, r4ngqpdh, x5pacpv3; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1107.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 1,107 92.4 24.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/84f101bx",
   "source_url": "https://omlx.ai/benchmarks/performance/84f101bx",
   "tg_tokens": null,
   "tps": 92.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-84f101bx-ctx8192",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: 3c64858c, fxjk8y78, lm5gyjct, r4ngqpdh, x5pacpv3; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1033.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 1,033 102.7 24.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/84f101bx",
   "source_url": "https://omlx.ai/benchmarks/performance/84f101bx",
   "tg_tokens": null,
   "tps": 102.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-84f101bx-ctx16384",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: 3c64858c, fxjk8y78, lm5gyjct, r4ngqpdh, x5pacpv3; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 907.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 907.0 89.7 25.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/84f101bx",
   "source_url": "https://omlx.ai/benchmarks/performance/84f101bx",
   "tg_tokens": null,
   "tps": 89.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-84f101bx-ctx32768",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: 3c64858c, fxjk8y78, lm5gyjct, r4ngqpdh, x5pacpv3; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 728.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 32k 728.3 88.1 26.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/84f101bx",
   "source_url": "https://omlx.ai/benchmarks/performance/84f101bx",
   "tg_tokens": null,
   "tps": 88.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-84f101bx-ctx65536",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: 3c64858c, fxjk8y78, lm5gyjct, r4ngqpdh, x5pacpv3; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 527.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 64k 527.7 64.4 27.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/84f101bx",
   "source_url": "https://omlx.ai/benchmarks/performance/84f101bx",
   "tg_tokens": null,
   "tps": 64.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-e6q3wtxb-ctx1024",
   "model": "Qwen3.8-Flash-Next-Dynamic",
   "notes": "BF16; PLE; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 82.0 GB; TTFT 731 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1400.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Dynamic-oQ5e-BF16-PLE-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,400 PP tok/s 74.3 TG tok/s 731 TTFT (ms) 82.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e6q3wtxb",
   "source_url": "https://omlx.ai/benchmarks/performance/e6q3wtxb",
   "tg_tokens": null,
   "tps": 74.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yyp1v5xa-ctx8192",
   "model": "Qwen3.8-Flash-Next-Dynamic",
   "notes": "BF16; PLE; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 83.2 GB; TTFT 4104 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1996.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Dynamic-oQ5e-BF16-PLE-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k tokens 1,996 PP tok/s 57.5 TG tok/s 4104 TTFT (ms) 83.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yyp1v5xa",
   "source_url": "https://omlx.ai/benchmarks/performance/yyp1v5xa",
   "tg_tokens": null,
   "tps": 57.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3cy48miu-ctx1024",
   "model": "Qwen3.8-Flash-Next-Dynamic",
   "notes": "BF16; PLE; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 82.0 GB at this context; duplicate traces of the same session: 2504i09n, xcuj9mc9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1304.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Dynamic-oQ5e-BF16-PLE-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,304 51.4 82.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3cy48miu",
   "source_url": "https://omlx.ai/benchmarks/performance/3cy48miu",
   "tg_tokens": null,
   "tps": 51.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3cy48miu-ctx16384",
   "model": "Qwen3.8-Flash-Next-Dynamic",
   "notes": "BF16; PLE; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 84.1 GB at this context; duplicate traces of the same session: 2504i09n, xcuj9mc9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2168.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Dynamic-oQ5e-BF16-PLE-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,168 73.2 84.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3cy48miu",
   "source_url": "https://omlx.ai/benchmarks/performance/3cy48miu",
   "tg_tokens": null,
   "tps": 73.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3cy48miu-ctx199680",
   "model": "Qwen3.8-Flash-Next-Dynamic",
   "notes": "BF16; PLE; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 91.3 GB at this context; duplicate traces of the same session: 2504i09n, xcuj9mc9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2049.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Dynamic-oQ5e-BF16-PLE-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,049 47.8 91.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3cy48miu",
   "source_url": "https://omlx.ai/benchmarks/performance/3cy48miu",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-01gjkopx-ctx1024",
   "model": "Qwen3.8-Flash-Next-Dynamic",
   "notes": "BF16; PLE; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 81.9 GB; TTFT 747 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1370.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Dynamic-oQ5e-BF16-PLE-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,370 PP tok/s 60.8 TG tok/s 747 TTFT (ms) 81.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/01gjkopx",
   "source_url": "https://omlx.ai/benchmarks/performance/01gjkopx",
   "tg_tokens": null,
   "tps": 60.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-aa9slivq-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 77.7 GB; TTFT 2206 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 464.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-MLX-4bit-MTP on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 464.1 PP tok/s 44.5 TG tok/s 2206 TTFT (ms) 77.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aa9slivq",
   "source_url": "https://omlx.ai/benchmarks/performance/aa9slivq",
   "tg_tokens": null,
   "tps": 44.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ucemn2zh-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 74.6,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 1k 74.6 9.8 26.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ucemn2zh",
   "source_url": "https://omlx.ai/benchmarks/performance/ucemn2zh",
   "tg_tokens": null,
   "tps": 9.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ucemn2zh-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 27.3 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 189.5,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 4k 189.5 7.2 27.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ucemn2zh",
   "source_url": "https://omlx.ai/benchmarks/performance/ucemn2zh",
   "tg_tokens": null,
   "tps": 7.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-5pkj1oyn-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 73.8 GB at this context; duplicate traces of the same session: upaqhmi6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 327.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 327.7 17.2 73.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5pkj1oyn",
   "source_url": "https://omlx.ai/benchmarks/performance/5pkj1oyn",
   "tg_tokens": null,
   "tps": 17.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-5pkj1oyn-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 81.1 GB at this context; duplicate traces of the same session: upaqhmi6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 429.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 429.7 17.0 81.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5pkj1oyn",
   "source_url": "https://omlx.ai/benchmarks/performance/5pkj1oyn",
   "tg_tokens": null,
   "tps": 17.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-zmfio2ip-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.2 GB at this context; duplicate traces of the same session: 79hr0qky; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 395.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 395.5 48.9 69.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zmfio2ip",
   "source_url": "https://omlx.ai/benchmarks/performance/zmfio2ip",
   "tg_tokens": null,
   "tps": 48.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-zmfio2ip-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.9 GB at this context; duplicate traces of the same session: 79hr0qky; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 586.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 586.2 38.1 69.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zmfio2ip",
   "source_url": "https://omlx.ai/benchmarks/performance/zmfio2ip",
   "tg_tokens": null,
   "tps": 38.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-agovu0sv-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 99.0 GB; TTFT 1065 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 961.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k tokens 961.6 PP tok/s 124.6 TG tok/s 1065 TTFT (ms) 99.0 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/agovu0sv",
   "source_url": "https://omlx.ai/benchmarks/performance/agovu0sv",
   "tg_tokens": null,
   "tps": 124.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-y6fi232z-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: 0uj77xpc, 5z0wv4qu, 8ezj2yoh, mbz8nkgw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3894.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 3,894 120.3 16.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y6fi232z",
   "source_url": "https://omlx.ai/benchmarks/performance/y6fi232z",
   "tg_tokens": null,
   "tps": 120.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-y6fi232z-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 0uj77xpc, 5z0wv4qu, 8ezj2yoh, mbz8nkgw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4049.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 4,049 131.3 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y6fi232z",
   "source_url": "https://omlx.ai/benchmarks/performance/y6fi232z",
   "tg_tokens": null,
   "tps": 131.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-y6fi232z-ctx8192",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: 0uj77xpc, 5z0wv4qu, 8ezj2yoh, mbz8nkgw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3977.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 3,977 121.8 16.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y6fi232z",
   "source_url": "https://omlx.ai/benchmarks/performance/y6fi232z",
   "tg_tokens": null,
   "tps": 121.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-y6fi232z-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 0uj77xpc, 5z0wv4qu, 8ezj2yoh, mbz8nkgw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3753.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 3,753 110.0 17.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y6fi232z",
   "source_url": "https://omlx.ai/benchmarks/performance/y6fi232z",
   "tg_tokens": null,
   "tps": 110.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-y6fi232z-ctx32768",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: 0uj77xpc, 5z0wv4qu, 8ezj2yoh, mbz8nkgw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3405.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 3,405 100.2 18.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y6fi232z",
   "source_url": "https://omlx.ai/benchmarks/performance/y6fi232z",
   "tg_tokens": null,
   "tps": 100.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-q96ubftn-ctx1024",
   "model": "LFM2-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 13.3 GB at this context; duplicate traces of the same session: byneltm6, dn6kcplo, j3cjpvzc; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 913.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2-24B-A2B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 913.2 87.8 13.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q96ubftn",
   "source_url": "https://omlx.ai/benchmarks/performance/q96ubftn",
   "tg_tokens": null,
   "tps": 87.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-q96ubftn-ctx4096",
   "model": "LFM2-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: byneltm6, dn6kcplo, j3cjpvzc; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1020.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2-24B-A2B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 1,020 84.8 13.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q96ubftn",
   "source_url": "https://omlx.ai/benchmarks/performance/q96ubftn",
   "tg_tokens": null,
   "tps": 84.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-q96ubftn-ctx8192",
   "model": "LFM2-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: byneltm6, dn6kcplo, j3cjpvzc; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 853.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2-24B-A2B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 853.2 81.7 14.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q96ubftn",
   "source_url": "https://omlx.ai/benchmarks/performance/q96ubftn",
   "tg_tokens": null,
   "tps": 81.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-q96ubftn-ctx16384",
   "model": "LFM2-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 15.3 GB at this context; duplicate traces of the same session: byneltm6, dn6kcplo, j3cjpvzc; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 795.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2-24B-A2B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 795.7 75.3 15.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q96ubftn",
   "source_url": "https://omlx.ai/benchmarks/performance/q96ubftn",
   "tg_tokens": null,
   "tps": 75.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-k8w9w5h4-ctx1024",
   "model": "LFM2-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: 38vb6c17, 5qm8kfun, 6pljt0az; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 933.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2-24B-A2B-MLX-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 1k 933.2 68.0 24.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k8w9w5h4",
   "source_url": "https://omlx.ai/benchmarks/performance/k8w9w5h4",
   "tg_tokens": null,
   "tps": 68.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-k8w9w5h4-ctx4096",
   "model": "LFM2-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: 38vb6c17, 5qm8kfun, 6pljt0az; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1044.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2-24B-A2B-MLX-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k 1,044 66.4 24.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k8w9w5h4",
   "source_url": "https://omlx.ai/benchmarks/performance/k8w9w5h4",
   "tg_tokens": null,
   "tps": 66.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-k8w9w5h4-ctx8192",
   "model": "LFM2-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: 38vb6c17, 5qm8kfun, 6pljt0az; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 838.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2-24B-A2B-MLX-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 8k 838.2 63.2 25.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k8w9w5h4",
   "source_url": "https://omlx.ai/benchmarks/performance/k8w9w5h4",
   "tg_tokens": null,
   "tps": 63.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-k8w9w5h4-ctx16384",
   "model": "LFM2-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: 38vb6c17, 5qm8kfun, 6pljt0az; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 782.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2-24B-A2B-MLX-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 16k 782.7 59.1 26.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k8w9w5h4",
   "source_url": "https://omlx.ai/benchmarks/performance/k8w9w5h4",
   "tg_tokens": null,
   "tps": 59.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-46nv8x7n-ctx1024",
   "model": "Llama 3.3 70B",
   "notes": "oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: si4rcuwv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 163.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Llama-3.3-70B-Instruct-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 163.8 9.0 70.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/46nv8x7n",
   "source_url": "https://omlx.ai/benchmarks/performance/46nv8x7n",
   "tg_tokens": null,
   "tps": 9.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-46nv8x7n-ctx4096",
   "model": "Llama 3.3 70B",
   "notes": "oMLX community benchmark trace; peak memory 71.9 GB at this context; duplicate traces of the same session: si4rcuwv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 164.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Llama-3.3-70B-Instruct-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 164.5 8.7 71.9 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/46nv8x7n",
   "source_url": "https://omlx.ai/benchmarks/performance/46nv8x7n",
   "tg_tokens": null,
   "tps": 8.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-kfw27vyv-ctx32768",
   "model": "Qwen3-0.6B",
   "notes": "Base model; oMLX community benchmark trace, single run at 32k context; peak memory 5.0 GB; TTFT 7659 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4278.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3-0.6B-Base on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 unknown; 32k tokens 4,278 PP tok/s 77.9 TG tok/s 7659 TTFT (ms) 5.0 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kfw27vyv",
   "source_url": "https://omlx.ai/benchmarks/performance/kfw27vyv",
   "tg_tokens": null,
   "tps": 77.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-u90wfa49-ctx1024",
   "model": "Qwen3-4B-Instruct-2507",
   "notes": "oMLX community benchmark trace; peak memory 2.8 GB at this context; duplicate traces of the same session: kmddo9q7; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 420.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-4B-Instruct-2507-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 420.8 37.9 2.8 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u90wfa49",
   "source_url": "https://omlx.ai/benchmarks/performance/u90wfa49",
   "tg_tokens": null,
   "tps": 37.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-u90wfa49-ctx4096",
   "model": "Qwen3-4B-Instruct-2507",
   "notes": "oMLX community benchmark trace; peak memory 3.2 GB at this context; duplicate traces of the same session: kmddo9q7; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 381.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-4B-Instruct-2507-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 381.9 31.8 3.2 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u90wfa49",
   "source_url": "https://omlx.ai/benchmarks/performance/u90wfa49",
   "tg_tokens": null,
   "tps": 31.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-tbk9c13u-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1955.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 1,955 99.2 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tbk9c13u",
   "source_url": "https://omlx.ai/benchmarks/performance/tbk9c13u",
   "tg_tokens": null,
   "tps": 99.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-tbk9c13u-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2417.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,417 97.0 20.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tbk9c13u",
   "source_url": "https://omlx.ai/benchmarks/performance/tbk9c13u",
   "tg_tokens": null,
   "tps": 97.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qek5iyd4-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: k1u8fl2b; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2053.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 2,053 80.5 27.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qek5iyd4",
   "source_url": "https://omlx.ai/benchmarks/performance/qek5iyd4",
   "tg_tokens": null,
   "tps": 80.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qek5iyd4-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: k1u8fl2b; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2679.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 2,679 78.6 28.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qek5iyd4",
   "source_url": "https://omlx.ai/benchmarks/performance/qek5iyd4",
   "tg_tokens": null,
   "tps": 78.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qek5iyd4-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.7 GB at this context; duplicate traces of the same session: k1u8fl2b; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2108.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 32k 2,108 63.3 28.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qek5iyd4",
   "source_url": "https://omlx.ai/benchmarks/performance/qek5iyd4",
   "tg_tokens": null,
   "tps": 63.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-lohestth-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1489.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 1,489 82.8 27.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lohestth",
   "source_url": "https://omlx.ai/benchmarks/performance/lohestth",
   "tg_tokens": null,
   "tps": 82.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-lohestth-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2453.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 2,453 80.6 28.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lohestth",
   "source_url": "https://omlx.ai/benchmarks/performance/lohestth",
   "tg_tokens": null,
   "tps": 80.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-thl3xc18-ctx131072",
   "model": "Qwen3.6-35B-A3B-MLX-VL",
   "notes": "FP16 build; oMLX community benchmark trace, single run at 128k context; peak memory 24.2 GB; TTFT 168297 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 778.8,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-MLX-VL-oQ4-FP16 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 4bit; 128k tokens 778.8 PP tok/s 56.5 TG tok/s 168297 TTFT (ms) 24.2 Peak mem",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/thl3xc18",
   "source_url": "https://omlx.ai/benchmarks/performance/thl3xc18",
   "tg_tokens": null,
   "tps": 56.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-i6ge1eyf-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: gfzdk0ok; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1857.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,857 69.2 20.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i6ge1eyf",
   "source_url": "https://omlx.ai/benchmarks/performance/i6ge1eyf",
   "tg_tokens": null,
   "tps": 69.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-i6ge1eyf-ctx16384",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: gfzdk0ok; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1685.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 1,685 61.0 20.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i6ge1eyf",
   "source_url": "https://omlx.ai/benchmarks/performance/i6ge1eyf",
   "tg_tokens": null,
   "tps": 61.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-pptqnte3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 3bvg86kt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 258.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 258.4 29.5 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pptqnte3",
   "source_url": "https://omlx.ai/benchmarks/performance/pptqnte3",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-pptqnte3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 3bvg86kt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 273.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 273.7 28.5 19.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pptqnte3",
   "source_url": "https://omlx.ai/benchmarks/performance/pptqnte3",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-v1894rcd-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: oaznk5s1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 445.9,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 445.9 14.4 20.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v1894rcd",
   "source_url": "https://omlx.ai/benchmarks/performance/v1894rcd",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-v1894rcd-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: oaznk5s1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 501.5,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 501.5 14.4 21.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v1894rcd",
   "source_url": "https://omlx.ai/benchmarks/performance/v1894rcd",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-bwh8g997-ctx16384",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace, single run at 16k context; peak memory 6.6 GB; TTFT 15237 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1075.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k tokens 1,075 PP tok/s 25.9 TG tok/s 15237 TTFT (ms) 6.6 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bwh8g997",
   "source_url": "https://omlx.ai/benchmarks/performance/bwh8g997",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-oy1fxz9q-ctx16384",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace, single run at 16k context; peak memory 7.2 GB; TTFT 15887 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1031.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k tokens 1,031 PP tok/s 44.6 TG tok/s 15887 TTFT (ms) 7.2 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oy1fxz9q",
   "source_url": "https://omlx.ai/benchmarks/performance/oy1fxz9q",
   "tg_tokens": null,
   "tps": 44.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-e79hxi17-ctx16384",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace, single run at 16k context; peak memory 7.2 GB; TTFT 15926 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1029.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k tokens 1,029 PP tok/s 43.7 TG tok/s 15926 TTFT (ms) 7.2 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e79hxi17",
   "source_url": "https://omlx.ai/benchmarks/performance/e79hxi17",
   "tg_tokens": null,
   "tps": 43.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-rp0tbzpd-ctx16384",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace, single run at 16k context; peak memory 7.2 GB; TTFT 15975 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1026.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k tokens 1,026 PP tok/s 41.4 TG tok/s 15975 TTFT (ms) 7.2 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rp0tbzpd",
   "source_url": "https://omlx.ai/benchmarks/performance/rp0tbzpd",
   "tg_tokens": null,
   "tps": 41.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-7usskbp9-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 71.5 GB; TTFT 7565 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 541.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 541.5 PP tok/s 48.8 TG tok/s 7565 TTFT (ms) 71.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7usskbp9",
   "source_url": "https://omlx.ai/benchmarks/performance/7usskbp9",
   "tg_tokens": null,
   "tps": 48.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2wq4qj34-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 74.6,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 1k 74.6 9.8 26.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2wq4qj34",
   "source_url": "https://omlx.ai/benchmarks/performance/2wq4qj34",
   "tg_tokens": null,
   "tps": 9.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2wq4qj34-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 27.3 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 189.5,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 4k 189.5 7.2 27.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2wq4qj34",
   "source_url": "https://omlx.ai/benchmarks/performance/2wq4qj34",
   "tg_tokens": null,
   "tps": 7.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-e6ltxzbl-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 75.2 GB; TTFT 87590 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 748.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k tokens 748.2 PP tok/s 52.4 TG tok/s 87590 TTFT (ms) 75.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e6ltxzbl",
   "source_url": "https://omlx.ai/benchmarks/performance/e6ltxzbl",
   "tg_tokens": null,
   "tps": 52.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-4aafjvjt-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.3 GB at this context; duplicate traces of the same session: 5jwapiih, 9qwm6muk, epk9007y, grcf8p4n, iephfwqn, qhz0dqz9; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 115.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 1k 115.6 10.6 28.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4aafjvjt",
   "source_url": "https://omlx.ai/benchmarks/performance/4aafjvjt",
   "tg_tokens": null,
   "tps": 10.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-4aafjvjt-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.4 GB at this context; duplicate traces of the same session: 5jwapiih, 9qwm6muk, epk9007y, grcf8p4n, iephfwqn, qhz0dqz9; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 116.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 4k 116.4 10.3 32.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4aafjvjt",
   "source_url": "https://omlx.ai/benchmarks/performance/4aafjvjt",
   "tg_tokens": null,
   "tps": 10.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-4aafjvjt-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.6 GB at this context; duplicate traces of the same session: 5jwapiih, 9qwm6muk, epk9007y, grcf8p4n, iephfwqn, qhz0dqz9; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 113.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 8k 113.7 10.0 33.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4aafjvjt",
   "source_url": "https://omlx.ai/benchmarks/performance/4aafjvjt",
   "tg_tokens": null,
   "tps": 10.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-4aafjvjt-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.2 GB at this context; duplicate traces of the same session: 5jwapiih, 9qwm6muk, epk9007y, grcf8p4n, iephfwqn, qhz0dqz9; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 109.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 16k 109.8 9.5 34.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4aafjvjt",
   "source_url": "https://omlx.ai/benchmarks/performance/4aafjvjt",
   "tg_tokens": null,
   "tps": 9.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-4aafjvjt-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: 5jwapiih, 9qwm6muk, epk9007y, grcf8p4n, iephfwqn, qhz0dqz9; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 32k 102.7 9.0 35.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4aafjvjt",
   "source_url": "https://omlx.ai/benchmarks/performance/4aafjvjt",
   "tg_tokens": null,
   "tps": 9.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-4aafjvjt-ctx65536",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: 5jwapiih, 9qwm6muk, epk9007y, grcf8p4n, iephfwqn, qhz0dqz9; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 88.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 64k 88.3 6.9 37.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4aafjvjt",
   "source_url": "https://omlx.ai/benchmarks/performance/4aafjvjt",
   "tg_tokens": null,
   "tps": 6.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-4aafjvjt-ctx131072",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 41.2 GB at this context; duplicate traces of the same session: 5jwapiih, 9qwm6muk, epk9007y, grcf8p4n, iephfwqn, qhz0dqz9; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 68.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 128k 68.4 5.2 41.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4aafjvjt",
   "source_url": "https://omlx.ai/benchmarks/performance/4aafjvjt",
   "tg_tokens": null,
   "tps": 5.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-gb6e4qy1-ctx1024",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: kmqiqzj9; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 779.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 8bit; 1k 779.1 58.9 25.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gb6e4qy1",
   "source_url": "https://omlx.ai/benchmarks/performance/gb6e4qy1",
   "tg_tokens": null,
   "tps": 58.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-gb6e4qy1-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: kmqiqzj9; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 781.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 8bit; 4k 781.2 57.3 26.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gb6e4qy1",
   "source_url": "https://omlx.ai/benchmarks/performance/gb6e4qy1",
   "tg_tokens": null,
   "tps": 57.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-2ddn5em3-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: 51h3y7q5, 62wkh0sz, vf00blh9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1507.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "gemma-4-26B-A4B-it-oQ5 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 5bit; 1k 1,507 93.2 18.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ddn5em3",
   "source_url": "https://omlx.ai/benchmarks/performance/2ddn5em3",
   "tg_tokens": null,
   "tps": 93.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-2ddn5em3-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 51h3y7q5, 62wkh0sz, vf00blh9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1445.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "gemma-4-26B-A4B-it-oQ5 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 5bit; 4k 1,445 116.9 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ddn5em3",
   "source_url": "https://omlx.ai/benchmarks/performance/2ddn5em3",
   "tg_tokens": null,
   "tps": 116.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2ddn5em3-ctx8192",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 51h3y7q5, 62wkh0sz, vf00blh9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1379.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "gemma-4-26B-A4B-it-oQ5 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 5bit; 8k 1,379 76.5 20.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ddn5em3",
   "source_url": "https://omlx.ai/benchmarks/performance/2ddn5em3",
   "tg_tokens": null,
   "tps": 76.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2ddn5em3-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 51h3y7q5, 62wkh0sz, vf00blh9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1283.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "gemma-4-26B-A4B-it-oQ5 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 5bit; 16k 1,283 55.9 21.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ddn5em3",
   "source_url": "https://omlx.ai/benchmarks/performance/2ddn5em3",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-vl7cckh1-ctx1024",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: o00a5ysy; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 116.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-MLX-8bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 8bit; 1k 116.5 10.2 32.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vl7cckh1",
   "source_url": "https://omlx.ai/benchmarks/performance/vl7cckh1",
   "tg_tokens": null,
   "tps": 10.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-vl7cckh1-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 33.1 GB at this context; duplicate traces of the same session: o00a5ysy; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 111.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-MLX-8bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 8bit; 4k 111.9 10.1 33.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vl7cckh1",
   "source_url": "https://omlx.ai/benchmarks/performance/vl7cckh1",
   "tg_tokens": null,
   "tps": 10.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gyrupzsy-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.9 GB at this context; duplicate traces of the same session: 5nd63wc9, d3mxlo2j, j9gx19yr, sfl2l7u1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 576.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 576.1 34.2 178.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gyrupzsy",
   "source_url": "https://omlx.ai/benchmarks/performance/gyrupzsy",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gyrupzsy-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.2 GB at this context; duplicate traces of the same session: 5nd63wc9, d3mxlo2j, j9gx19yr, sfl2l7u1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 529.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 529.9 48.0 179.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gyrupzsy",
   "source_url": "https://omlx.ai/benchmarks/performance/gyrupzsy",
   "tg_tokens": null,
   "tps": 48.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gyrupzsy-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.1 GB at this context; duplicate traces of the same session: 5nd63wc9, d3mxlo2j, j9gx19yr, sfl2l7u1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 515.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 515.4 43.5 180.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gyrupzsy",
   "source_url": "https://omlx.ai/benchmarks/performance/gyrupzsy",
   "tg_tokens": null,
   "tps": 43.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gyrupzsy-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.4 GB at this context; duplicate traces of the same session: 5nd63wc9, d3mxlo2j, j9gx19yr, sfl2l7u1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 504.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 504.7 41.3 181.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gyrupzsy",
   "source_url": "https://omlx.ai/benchmarks/performance/gyrupzsy",
   "tg_tokens": null,
   "tps": 41.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gyrupzsy-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 182.8 GB at this context; duplicate traces of the same session: 5nd63wc9, d3mxlo2j, j9gx19yr, sfl2l7u1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 491.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 195k 491.9 29.4 182.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gyrupzsy",
   "source_url": "https://omlx.ai/benchmarks/performance/gyrupzsy",
   "tg_tokens": null,
   "tps": 29.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-bxo4pus3-ctx1024",
   "model": "Huihui-Qwen3.5-9B-abliterated",
   "notes": "oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: qwka5nun, uhhuv82k, zor0lt4b; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 223.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Huihui-Qwen3.5-9B-abliterated-mlx-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 223.6 25.1 6.3 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bxo4pus3",
   "source_url": "https://omlx.ai/benchmarks/performance/bxo4pus3",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-bxo4pus3-ctx4096",
   "model": "Huihui-Qwen3.5-9B-abliterated",
   "notes": "oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: qwka5nun, uhhuv82k, zor0lt4b; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 228.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Huihui-Qwen3.5-9B-abliterated-mlx-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 228.8 26.1 7.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bxo4pus3",
   "source_url": "https://omlx.ai/benchmarks/performance/bxo4pus3",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-bxo4pus3-ctx8192",
   "model": "Huihui-Qwen3.5-9B-abliterated",
   "notes": "oMLX community benchmark trace; peak memory 7.2 GB at this context; duplicate traces of the same session: qwka5nun, uhhuv82k, zor0lt4b; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Huihui-Qwen3.5-9B-abliterated-mlx-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 8k 220.9 24.6 7.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bxo4pus3",
   "source_url": "https://omlx.ai/benchmarks/performance/bxo4pus3",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-bxo4pus3-ctx16384",
   "model": "Huihui-Qwen3.5-9B-abliterated",
   "notes": "oMLX community benchmark trace; peak memory 7.8 GB at this context; duplicate traces of the same session: qwka5nun, uhhuv82k, zor0lt4b; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 203.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Huihui-Qwen3.5-9B-abliterated-mlx-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 16k 203.2 16.3 7.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bxo4pus3",
   "source_url": "https://omlx.ai/benchmarks/performance/bxo4pus3",
   "tg_tokens": null,
   "tps": 16.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xn6a8afh-ctx1024",
   "model": "Nanbeige4.2-3B",
   "notes": "oMLX community benchmark trace; peak memory 3.1 GB at this context; duplicate traces of the same session: c3aphnk4, rjs1xx1k; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1127.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Nanbeige4.2-3B-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,127 54.3 3.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xn6a8afh",
   "source_url": "https://omlx.ai/benchmarks/performance/xn6a8afh",
   "tg_tokens": null,
   "tps": 54.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xn6a8afh-ctx4096",
   "model": "Nanbeige4.2-3B",
   "notes": "oMLX community benchmark trace; peak memory 3.6 GB at this context; duplicate traces of the same session: c3aphnk4, rjs1xx1k; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1500.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Nanbeige4.2-3B-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 1,500 46.6 3.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xn6a8afh",
   "source_url": "https://omlx.ai/benchmarks/performance/xn6a8afh",
   "tg_tokens": null,
   "tps": 46.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xn6a8afh-ctx32768",
   "model": "Nanbeige4.2-3B",
   "notes": "oMLX community benchmark trace; peak memory 8.3 GB at this context; duplicate traces of the same session: c3aphnk4, rjs1xx1k; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 746.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Nanbeige4.2-3B-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 746.3 35.2 8.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xn6a8afh",
   "source_url": "https://omlx.ai/benchmarks/performance/xn6a8afh",
   "tg_tokens": null,
   "tps": 35.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-zqrp2o2z-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: boo28vhc, v96yh3rx; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 289.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 289.4 46.3 5.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zqrp2o2z",
   "source_url": "https://omlx.ai/benchmarks/performance/zqrp2o2z",
   "tg_tokens": null,
   "tps": 46.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-zqrp2o2z-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.5 GB at this context; duplicate traces of the same session: boo28vhc, v96yh3rx; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 269.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 269.9 44.7 7.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zqrp2o2z",
   "source_url": "https://omlx.ai/benchmarks/performance/zqrp2o2z",
   "tg_tokens": null,
   "tps": 44.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-zqrp2o2z-ctx8192",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.0 GB at this context; duplicate traces of the same session: boo28vhc, v96yh3rx; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 262.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 8k 262.1 42.4 8.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zqrp2o2z",
   "source_url": "https://omlx.ai/benchmarks/performance/zqrp2o2z",
   "tg_tokens": null,
   "tps": 42.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-zqrp2o2z-ctx16384",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 9.1 GB at this context; duplicate traces of the same session: boo28vhc, v96yh3rx; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 247.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k 247.3 39.6 9.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zqrp2o2z",
   "source_url": "https://omlx.ai/benchmarks/performance/zqrp2o2z",
   "tg_tokens": null,
   "tps": 39.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-zqrp2o2z-ctx32768",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 9.9 GB at this context; duplicate traces of the same session: boo28vhc, v96yh3rx; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 32k 220.3 34.3 9.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zqrp2o2z",
   "source_url": "https://omlx.ai/benchmarks/performance/zqrp2o2z",
   "tg_tokens": null,
   "tps": 34.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-zqrp2o2z-ctx65536",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.9 GB at this context; duplicate traces of the same session: boo28vhc, v96yh3rx; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 180.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 64k 180.5 26.8 10.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zqrp2o2z",
   "source_url": "https://omlx.ai/benchmarks/performance/zqrp2o2z",
   "tg_tokens": null,
   "tps": 26.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-zqrp2o2z-ctx131072",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 12.9 GB at this context; duplicate traces of the same session: boo28vhc, v96yh3rx; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 133.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 128k 133.7 18.8 12.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zqrp2o2z",
   "source_url": "https://omlx.ai/benchmarks/performance/zqrp2o2z",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-n61adkv7-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 10.0 GB at this context; duplicate traces of the same session: p32zcjj0; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2371.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B-MLX-8bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 8bit; 1k 2,371 47.8 10.0 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n61adkv7",
   "source_url": "https://omlx.ai/benchmarks/performance/n61adkv7",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-n61adkv7-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 10.5 GB at this context; duplicate traces of the same session: p32zcjj0; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2672.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B-MLX-8bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 8bit; 4k 2,672 47.0 10.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n61adkv7",
   "source_url": "https://omlx.ai/benchmarks/performance/n61adkv7",
   "tg_tokens": null,
   "tps": 47.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-74bl6oce-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2053.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 2,053 80.5 27.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/74bl6oce",
   "source_url": "https://omlx.ai/benchmarks/performance/74bl6oce",
   "tg_tokens": null,
   "tps": 80.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-74bl6oce-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2679.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 2,679 78.6 28.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/74bl6oce",
   "source_url": "https://omlx.ai/benchmarks/performance/74bl6oce",
   "tg_tokens": null,
   "tps": 78.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-74bl6oce-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.7 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2108.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 32k 2,108 63.3 28.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/74bl6oce",
   "source_url": "https://omlx.ai/benchmarks/performance/74bl6oce",
   "tg_tokens": null,
   "tps": 63.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-zn2ys8uo-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1489.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 1,489 82.8 27.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zn2ys8uo",
   "source_url": "https://omlx.ai/benchmarks/performance/zn2ys8uo",
   "tg_tokens": null,
   "tps": 82.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-zn2ys8uo-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2453.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 2,453 80.6 28.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zn2ys8uo",
   "source_url": "https://omlx.ai/benchmarks/performance/zn2ys8uo",
   "tg_tokens": null,
   "tps": 80.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-f4f9okty-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: bkb1tq86, ku53b9dr, o0kf0thn, val9knuy, z5zhuaq8; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 695.3,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 695.3 97.6 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f4f9okty",
   "source_url": "https://omlx.ai/benchmarks/performance/f4f9okty",
   "tg_tokens": null,
   "tps": 97.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-f4f9okty-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: bkb1tq86, ku53b9dr, o0kf0thn, val9knuy, z5zhuaq8; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 757.7,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 757.7 81.9 23.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f4f9okty",
   "source_url": "https://omlx.ai/benchmarks/performance/f4f9okty",
   "tg_tokens": null,
   "tps": 81.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-f4f9okty-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: bkb1tq86, ku53b9dr, o0kf0thn, val9knuy, z5zhuaq8; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 700.5,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 700.5 65.7 23.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f4f9okty",
   "source_url": "https://omlx.ai/benchmarks/performance/f4f9okty",
   "tg_tokens": null,
   "tps": 65.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-f4f9okty-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: bkb1tq86, ku53b9dr, o0kf0thn, val9knuy, z5zhuaq8; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 592.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 592.0 67.6 24.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f4f9okty",
   "source_url": "https://omlx.ai/benchmarks/performance/f4f9okty",
   "tg_tokens": null,
   "tps": 67.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-f4f9okty-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 25.7 GB at this context; duplicate traces of the same session: bkb1tq86, ku53b9dr, o0kf0thn, val9knuy, z5zhuaq8; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 428.1,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 32k 428.1 60.9 25.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f4f9okty",
   "source_url": "https://omlx.ai/benchmarks/performance/f4f9okty",
   "tg_tokens": null,
   "tps": 60.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-f4f9okty-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: bkb1tq86, ku53b9dr, o0kf0thn, val9knuy, z5zhuaq8; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 280.3,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 64k 280.3 40.2 26.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f4f9okty",
   "source_url": "https://omlx.ai/benchmarks/performance/f4f9okty",
   "tg_tokens": null,
   "tps": 40.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-3mi7zuz6-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 7glzwdb4, 9b56fh1m, dcqid01h, f1u12ll6, gkyeu7mi; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 113.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 1k 113.0 14.5 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3mi7zuz6",
   "source_url": "https://omlx.ai/benchmarks/performance/3mi7zuz6",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-3mi7zuz6-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: 7glzwdb4, 9b56fh1m, dcqid01h, f1u12ll6, gkyeu7mi; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 105.5 13.8 19.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3mi7zuz6",
   "source_url": "https://omlx.ai/benchmarks/performance/3mi7zuz6",
   "tg_tokens": null,
   "tps": 13.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-3mi7zuz6-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 7glzwdb4, 9b56fh1m, dcqid01h, f1u12ll6, gkyeu7mi; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 99.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 8k 99.5 12.7 20.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3mi7zuz6",
   "source_url": "https://omlx.ai/benchmarks/performance/3mi7zuz6",
   "tg_tokens": null,
   "tps": 12.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-3mi7zuz6-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: 7glzwdb4, 9b56fh1m, dcqid01h, f1u12ll6, gkyeu7mi; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 89.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 16k 89.3 10.8 20.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3mi7zuz6",
   "source_url": "https://omlx.ai/benchmarks/performance/3mi7zuz6",
   "tg_tokens": null,
   "tps": 10.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-3mi7zuz6-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: 7glzwdb4, 9b56fh1m, dcqid01h, f1u12ll6, gkyeu7mi; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 85.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 32k 85.6 11.5 22.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3mi7zuz6",
   "source_url": "https://omlx.ai/benchmarks/performance/3mi7zuz6",
   "tg_tokens": null,
   "tps": 11.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-3mi7zuz6-ctx65536",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: 7glzwdb4, 9b56fh1m, dcqid01h, f1u12ll6, gkyeu7mi; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 72.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 64k 72.8 9.9 24.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3mi7zuz6",
   "source_url": "https://omlx.ai/benchmarks/performance/3mi7zuz6",
   "tg_tokens": null,
   "tps": 9.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-y6w2mx22-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: 9qr0xscm; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 455.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 455.2 103.7 30.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y6w2mx22",
   "source_url": "https://omlx.ai/benchmarks/performance/y6w2mx22",
   "tg_tokens": null,
   "tps": 103.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-y6w2mx22-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.4 GB at this context; duplicate traces of the same session: 9qr0xscm; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 407.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 407.1 90.3 32.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y6w2mx22",
   "source_url": "https://omlx.ai/benchmarks/performance/y6w2mx22",
   "tg_tokens": null,
   "tps": 90.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-lc9rbmt1-ctx16384",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: 6mewm69t, p1o9p6ul; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 885.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 885.9 94.0 25.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lc9rbmt1",
   "source_url": "https://omlx.ai/benchmarks/performance/lc9rbmt1",
   "tg_tokens": null,
   "tps": 94.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-lc9rbmt1-ctx32768",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: 6mewm69t, p1o9p6ul; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 718.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 32k 718.0 89.5 26.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lc9rbmt1",
   "source_url": "https://omlx.ai/benchmarks/performance/lc9rbmt1",
   "tg_tokens": null,
   "tps": 89.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-lc9rbmt1-ctx65536",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.3 GB at this context; duplicate traces of the same session: 6mewm69t, p1o9p6ul; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 498.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 64k 498.4 64.7 27.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lc9rbmt1",
   "source_url": "https://omlx.ai/benchmarks/performance/lc9rbmt1",
   "tg_tokens": null,
   "tps": 64.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-cmcafodt-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 2ilkmlys, dq5bhbpw, nsz9lki0, p7sr43ot; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 829.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 829.2 60.6 23.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cmcafodt",
   "source_url": "https://omlx.ai/benchmarks/performance/cmcafodt",
   "tg_tokens": null,
   "tps": 60.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-cmcafodt-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: 2ilkmlys, dq5bhbpw, nsz9lki0, p7sr43ot; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 825.8,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 825.8 50.9 24.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cmcafodt",
   "source_url": "https://omlx.ai/benchmarks/performance/cmcafodt",
   "tg_tokens": null,
   "tps": 50.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-cmcafodt-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: 2ilkmlys, dq5bhbpw, nsz9lki0, p7sr43ot; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 710.3,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 710.3 56.5 24.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cmcafodt",
   "source_url": "https://omlx.ai/benchmarks/performance/cmcafodt",
   "tg_tokens": null,
   "tps": 56.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cmcafodt-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.2 GB at this context; duplicate traces of the same session: 2ilkmlys, dq5bhbpw, nsz9lki0, p7sr43ot; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 679.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 679.4 64.0 25.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cmcafodt",
   "source_url": "https://omlx.ai/benchmarks/performance/cmcafodt",
   "tg_tokens": null,
   "tps": 64.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cmcafodt-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: 2ilkmlys, dq5bhbpw, nsz9lki0, p7sr43ot; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 623.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 623.9 60.8 26.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cmcafodt",
   "source_url": "https://omlx.ai/benchmarks/performance/cmcafodt",
   "tg_tokens": null,
   "tps": 60.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fx9l0e44-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 861.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 861.1 27.6 19.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fx9l0e44",
   "source_url": "https://omlx.ai/benchmarks/performance/fx9l0e44",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fx9l0e44-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 863.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 863.7 26.8 20.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fx9l0e44",
   "source_url": "https://omlx.ai/benchmarks/performance/fx9l0e44",
   "tg_tokens": null,
   "tps": 26.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fx9l0e44-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 731.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 731.9 26.1 20.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fx9l0e44",
   "source_url": "https://omlx.ai/benchmarks/performance/fx9l0e44",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fx9l0e44-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 699.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 699.1 23.8 21.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fx9l0e44",
   "source_url": "https://omlx.ai/benchmarks/performance/fx9l0e44",
   "tg_tokens": null,
   "tps": 23.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fx9l0e44-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 653.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 653.2 21.4 22.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fx9l0e44",
   "source_url": "https://omlx.ai/benchmarks/performance/fx9l0e44",
   "tg_tokens": null,
   "tps": 21.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-cqdo86zn-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: 3bzogyyd, iillcdd5, u2hlsfpw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1424.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "gemma-4-26B-A4B-it-oQ5 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 5bit; 1k 1,424 93.6 17.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cqdo86zn",
   "source_url": "https://omlx.ai/benchmarks/performance/cqdo86zn",
   "tg_tokens": null,
   "tps": 93.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-cqdo86zn-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: 3bzogyyd, iillcdd5, u2hlsfpw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1450.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "gemma-4-26B-A4B-it-oQ5 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 5bit; 4k 1,450 90.5 17.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cqdo86zn",
   "source_url": "https://omlx.ai/benchmarks/performance/cqdo86zn",
   "tg_tokens": null,
   "tps": 90.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-cqdo86zn-ctx8192",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 17.9 GB at this context; duplicate traces of the same session: 3bzogyyd, iillcdd5, u2hlsfpw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1411.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "gemma-4-26B-A4B-it-oQ5 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 5bit; 8k 1,411 87.1 17.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cqdo86zn",
   "source_url": "https://omlx.ai/benchmarks/performance/cqdo86zn",
   "tg_tokens": null,
   "tps": 87.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-cqdo86zn-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: 3bzogyyd, iillcdd5, u2hlsfpw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1330.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "gemma-4-26B-A4B-it-oQ5 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 5bit; 16k 1,330 80.8 18.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cqdo86zn",
   "source_url": "https://omlx.ai/benchmarks/performance/cqdo86zn",
   "tg_tokens": null,
   "tps": 80.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ndup8owz-ctx4096",
   "model": "Laguna-S-2.1",
   "notes": "oMLX community benchmark trace; peak memory 49.3 GB at this context; duplicate traces of the same session: 0hszywc7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 179.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Laguna-S-2.1-oQ5e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 5bit; 4k 179.1 12.5 49.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ndup8owz",
   "source_url": "https://omlx.ai/benchmarks/performance/ndup8owz",
   "tg_tokens": null,
   "tps": 12.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ndup8owz-ctx16384",
   "model": "Laguna-S-2.1",
   "notes": "oMLX community benchmark trace; peak memory 49.9 GB at this context; duplicate traces of the same session: 0hszywc7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 190.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Laguna-S-2.1-oQ5e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 5bit; 16k 190.7 13.6 49.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ndup8owz",
   "source_url": "https://omlx.ai/benchmarks/performance/ndup8owz",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-7vnuxagt-ctx1024",
   "model": "Laguna-S-2.1",
   "notes": "oMLX community benchmark trace; peak memory 51.2 GB at this context; duplicate traces of the same session: w4kvgn1v; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 149.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Laguna-S-2.1-oQ5e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 5bit; 1k 149.2 13.2 51.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7vnuxagt",
   "source_url": "https://omlx.ai/benchmarks/performance/7vnuxagt",
   "tg_tokens": null,
   "tps": 13.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-7vnuxagt-ctx4096",
   "model": "Laguna-S-2.1",
   "notes": "oMLX community benchmark trace; peak memory 51.5 GB at this context; duplicate traces of the same session: w4kvgn1v; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 194.8,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Laguna-S-2.1-oQ5e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 5bit; 4k 194.8 13.0 51.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7vnuxagt",
   "source_url": "https://omlx.ai/benchmarks/performance/7vnuxagt",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-9jhuy410-ctx4096",
   "model": "Laguna-S-2.1",
   "notes": "oMLX community benchmark trace; peak memory 45.9 GB at this context; duplicate traces of the same session: 36nd0caa; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 117.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Laguna-S-2.1-oQ6e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 117.5 8.8 45.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9jhuy410",
   "source_url": "https://omlx.ai/benchmarks/performance/9jhuy410",
   "tg_tokens": null,
   "tps": 8.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-9jhuy410-ctx16384",
   "model": "Laguna-S-2.1",
   "notes": "oMLX community benchmark trace; peak memory 46.6 GB at this context; duplicate traces of the same session: 36nd0caa; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 119.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Laguna-S-2.1-oQ6e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 16k 119.3 9.4 46.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9jhuy410",
   "source_url": "https://omlx.ai/benchmarks/performance/9jhuy410",
   "tg_tokens": null,
   "tps": 9.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-pnkm0gwi-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: sedva8rp, urwk3ivn, y2uul85j; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 289.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 289.4 46.3 5.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pnkm0gwi",
   "source_url": "https://omlx.ai/benchmarks/performance/pnkm0gwi",
   "tg_tokens": null,
   "tps": 46.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-pnkm0gwi-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.5 GB at this context; duplicate traces of the same session: sedva8rp, urwk3ivn, y2uul85j; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 269.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 269.9 44.7 7.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pnkm0gwi",
   "source_url": "https://omlx.ai/benchmarks/performance/pnkm0gwi",
   "tg_tokens": null,
   "tps": 44.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-pnkm0gwi-ctx8192",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.0 GB at this context; duplicate traces of the same session: sedva8rp, urwk3ivn, y2uul85j; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 262.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 8k 262.1 42.4 8.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pnkm0gwi",
   "source_url": "https://omlx.ai/benchmarks/performance/pnkm0gwi",
   "tg_tokens": null,
   "tps": 42.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-pnkm0gwi-ctx16384",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 9.1 GB at this context; duplicate traces of the same session: sedva8rp, urwk3ivn, y2uul85j; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 247.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k 247.3 39.6 9.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pnkm0gwi",
   "source_url": "https://omlx.ai/benchmarks/performance/pnkm0gwi",
   "tg_tokens": null,
   "tps": 39.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-pnkm0gwi-ctx32768",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 9.9 GB at this context; duplicate traces of the same session: sedva8rp, urwk3ivn, y2uul85j; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 32k 220.3 34.3 9.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pnkm0gwi",
   "source_url": "https://omlx.ai/benchmarks/performance/pnkm0gwi",
   "tg_tokens": null,
   "tps": 34.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-pnkm0gwi-ctx65536",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.9 GB at this context; duplicate traces of the same session: sedva8rp, urwk3ivn, y2uul85j; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 180.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 64k 180.5 26.8 10.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pnkm0gwi",
   "source_url": "https://omlx.ai/benchmarks/performance/pnkm0gwi",
   "tg_tokens": null,
   "tps": 26.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-pnkm0gwi-ctx131072",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 12.9 GB at this context; duplicate traces of the same session: sedva8rp, urwk3ivn, y2uul85j; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 133.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 128k 133.7 18.8 12.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pnkm0gwi",
   "source_url": "https://omlx.ai/benchmarks/performance/pnkm0gwi",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-cjgs8cmr-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.8 GB at this context; duplicate traces of the same session: 1k4itb9j, 92khkep0, erue2fz6, hp4zjjf8, rw1s74s8, ww2spg41; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 284.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 6bit; 1k 284.8 35.2 7.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cjgs8cmr",
   "source_url": "https://omlx.ai/benchmarks/performance/cjgs8cmr",
   "tg_tokens": null,
   "tps": 35.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-cjgs8cmr-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 9.6 GB at this context; duplicate traces of the same session: 1k4itb9j, 92khkep0, erue2fz6, hp4zjjf8, rw1s74s8, ww2spg41; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 258.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 6bit; 4k 258.7 33.9 9.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cjgs8cmr",
   "source_url": "https://omlx.ai/benchmarks/performance/cjgs8cmr",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-cjgs8cmr-ctx8192",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.1 GB at this context; duplicate traces of the same session: 1k4itb9j, 92khkep0, erue2fz6, hp4zjjf8, rw1s74s8, ww2spg41; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 254.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 6bit; 8k 254.6 33.6 10.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cjgs8cmr",
   "source_url": "https://omlx.ai/benchmarks/performance/cjgs8cmr",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-cjgs8cmr-ctx16384",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.2 GB at this context; duplicate traces of the same session: 1k4itb9j, 92khkep0, erue2fz6, hp4zjjf8, rw1s74s8, ww2spg41; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 239.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 6bit; 16k 239.7 31.3 11.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cjgs8cmr",
   "source_url": "https://omlx.ai/benchmarks/performance/cjgs8cmr",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-cjgs8cmr-ctx32768",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 12.0 GB at this context; duplicate traces of the same session: 1k4itb9j, 92khkep0, erue2fz6, hp4zjjf8, rw1s74s8, ww2spg41; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 214.2,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 6bit; 32k 214.2 27.9 12.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cjgs8cmr",
   "source_url": "https://omlx.ai/benchmarks/performance/cjgs8cmr",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-cjgs8cmr-ctx65536",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 13.0 GB at this context; duplicate traces of the same session: 1k4itb9j, 92khkep0, erue2fz6, hp4zjjf8, rw1s74s8, ww2spg41; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 177.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 6bit; 64k 177.1 22.8 13.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cjgs8cmr",
   "source_url": "https://omlx.ai/benchmarks/performance/cjgs8cmr",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-cjgs8cmr-ctx131072",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 15.0 GB at this context; duplicate traces of the same session: 1k4itb9j, 92khkep0, erue2fz6, hp4zjjf8, rw1s74s8, ww2spg41; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 6bit; 128k 131.6 16.5 15.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cjgs8cmr",
   "source_url": "https://omlx.ai/benchmarks/performance/cjgs8cmr",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-bpz570ry-ctx1024",
   "model": "Qwen3-Coder-Next-REAM",
   "notes": "oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: 4phyyyq2, 8q00b70o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1251.0,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3-Coder-Next-REAM-mlx-3Bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 3bit; 1k 1,251 88.4 26.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bpz570ry",
   "source_url": "https://omlx.ai/benchmarks/performance/bpz570ry",
   "tg_tokens": null,
   "tps": 88.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-bpz570ry-ctx4096",
   "model": "Qwen3-Coder-Next-REAM",
   "notes": "oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: 4phyyyq2, 8q00b70o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1488.0,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3-Coder-Next-REAM-mlx-3Bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 3bit; 4k 1,488 80.7 26.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bpz570ry",
   "source_url": "https://omlx.ai/benchmarks/performance/bpz570ry",
   "tg_tokens": null,
   "tps": 80.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-bpz570ry-ctx8192",
   "model": "Qwen3-Coder-Next-REAM",
   "notes": "oMLX community benchmark trace; peak memory 26.9 GB at this context; duplicate traces of the same session: 4phyyyq2, 8q00b70o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1169.0,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3-Coder-Next-REAM-mlx-3Bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 3bit; 8k 1,169 84.7 26.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bpz570ry",
   "source_url": "https://omlx.ai/benchmarks/performance/bpz570ry",
   "tg_tokens": null,
   "tps": 84.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-yoek50ro-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 19fracor; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 265.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 265.5 27.4 20.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yoek50ro",
   "source_url": "https://omlx.ai/benchmarks/performance/yoek50ro",
   "tg_tokens": null,
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-yoek50ro-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 19fracor; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 246.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 246.2 25.5 21.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yoek50ro",
   "source_url": "https://omlx.ai/benchmarks/performance/yoek50ro",
   "tg_tokens": null,
   "tps": 25.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-yoek50ro-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: 19fracor; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 201.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 201.6 20.9 25.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yoek50ro",
   "source_url": "https://omlx.ai/benchmarks/performance/yoek50ro",
   "tg_tokens": null,
   "tps": 20.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mnv8i882-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: ima3s0in, kuy09rys; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 949.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 949.5 48.0 18.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mnv8i882",
   "source_url": "https://omlx.ai/benchmarks/performance/mnv8i882",
   "tg_tokens": null,
   "tps": 48.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mnv8i882-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: ima3s0in, kuy09rys; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 953.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 953.1 66.0 19.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mnv8i882",
   "source_url": "https://omlx.ai/benchmarks/performance/mnv8i882",
   "tg_tokens": null,
   "tps": 66.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mnv8i882-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: ima3s0in, kuy09rys; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 892.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 892.2 67.5 20.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mnv8i882",
   "source_url": "https://omlx.ai/benchmarks/performance/mnv8i882",
   "tg_tokens": null,
   "tps": 67.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-ff612ycp-ctx16384",
   "model": "Qwen3.8-9B-Distill-Uncensored-Heretic",
   "notes": "MTP build (mlx); oMLX community benchmark trace, single run at 16k context; peak memory 11.0 GB; TTFT 17576 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 932.2,
   "provenance": "community",
   "quant": "Q6G64",
   "quote": "Qwen3.8-9B-Distill-Uncensored-Heretic-MTPLX-Q6G64 on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 6bit; 16k tokens 932.2 PP tok/s 36.4 TG tok/s 17576 TTFT (ms) 11.0 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ff612ycp",
   "source_url": "https://omlx.ai/benchmarks/performance/ff612ycp",
   "tg_tokens": null,
   "tps": 36.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-2p2z974x-ctx16384",
   "model": "Qwen3.8-9B-Distill-Uncensored-Heretic",
   "notes": "MTP build (mlx); oMLX community benchmark trace, single run at 16k context; peak memory 8.6 GB; TTFT 16712 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 980.4,
   "provenance": "community",
   "quant": "Q6G64",
   "quote": "Qwen3.8-9B-Distill-Uncensored-Heretic-MTPLX-Q6G64 on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 6bit; 16k tokens 980.4 PP tok/s 19.3 TG tok/s 16712 TTFT (ms) 8.6 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2p2z974x",
   "source_url": "https://omlx.ai/benchmarks/performance/2p2z974x",
   "tg_tokens": null,
   "tps": 19.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-0rejuv1w-ctx16384",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace, single run at 16k context; peak memory 7.2 GB; TTFT 15952 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1027.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k tokens 1,027 PP tok/s 42.8 TG tok/s 15952 TTFT (ms) 7.2 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0rejuv1w",
   "source_url": "https://omlx.ai/benchmarks/performance/0rejuv1w",
   "tg_tokens": null,
   "tps": 42.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-z9qxz7pw-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.0 GB at this context; duplicate traces of the same session: rekiwqz1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 756.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 756.1 56.1 73.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z9qxz7pw",
   "source_url": "https://omlx.ai/benchmarks/performance/z9qxz7pw",
   "tg_tokens": null,
   "tps": 56.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-z9qxz7pw-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.8 GB at this context; duplicate traces of the same session: rekiwqz1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 719.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 719.5 55.5 78.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z9qxz7pw",
   "source_url": "https://omlx.ai/benchmarks/performance/z9qxz7pw",
   "tg_tokens": null,
   "tps": 55.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9qjrpose-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 71.0 GB; TTFT 2295 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1785.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 1,785 PP tok/s 71.2 TG tok/s 2295 TTFT (ms) 71.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9qjrpose",
   "source_url": "https://omlx.ai/benchmarks/performance/9qjrpose",
   "tg_tokens": null,
   "tps": 71.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mf1xu945-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: adzh54dq, i9i4tyay, p9rjgzxj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 861.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 861.1 27.6 19.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mf1xu945",
   "source_url": "https://omlx.ai/benchmarks/performance/mf1xu945",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mf1xu945-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: adzh54dq, i9i4tyay, p9rjgzxj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 863.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 863.7 26.8 20.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mf1xu945",
   "source_url": "https://omlx.ai/benchmarks/performance/mf1xu945",
   "tg_tokens": null,
   "tps": 26.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mf1xu945-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: adzh54dq, i9i4tyay, p9rjgzxj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 731.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 731.9 26.1 20.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mf1xu945",
   "source_url": "https://omlx.ai/benchmarks/performance/mf1xu945",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mf1xu945-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: adzh54dq, i9i4tyay, p9rjgzxj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 699.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 699.1 23.8 21.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mf1xu945",
   "source_url": "https://omlx.ai/benchmarks/performance/mf1xu945",
   "tg_tokens": null,
   "tps": 23.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mf1xu945-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: adzh54dq, i9i4tyay, p9rjgzxj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 653.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 653.2 21.4 22.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mf1xu945",
   "source_url": "https://omlx.ai/benchmarks/performance/mf1xu945",
   "tg_tokens": null,
   "tps": 21.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-1cytkeza-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: 230f8yl4, 5bfhjct6, 5cd7kw58, d4jzdgp8, htua5wc2, qylystdp, spli5m21; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1507.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "gemma-4-26B-A4B-it-oQ5 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 5bit; 1k 1,507 91.4 18.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1cytkeza",
   "source_url": "https://omlx.ai/benchmarks/performance/1cytkeza",
   "tg_tokens": null,
   "tps": 91.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-1cytkeza-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 230f8yl4, 5bfhjct6, 5cd7kw58, d4jzdgp8, htua5wc2, qylystdp, spli5m21; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1445.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "gemma-4-26B-A4B-it-oQ5 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 5bit; 4k 1,445 84.1 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1cytkeza",
   "source_url": "https://omlx.ai/benchmarks/performance/1cytkeza",
   "tg_tokens": null,
   "tps": 84.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-1cytkeza-ctx8192",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 230f8yl4, 5bfhjct6, 5cd7kw58, d4jzdgp8, htua5wc2, qylystdp, spli5m21; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1380.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "gemma-4-26B-A4B-it-oQ5 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 5bit; 8k 1,380 76.0 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1cytkeza",
   "source_url": "https://omlx.ai/benchmarks/performance/1cytkeza",
   "tg_tokens": null,
   "tps": 76.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-1cytkeza-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 230f8yl4, 5bfhjct6, 5cd7kw58, d4jzdgp8, htua5wc2, qylystdp, spli5m21; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1284.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "gemma-4-26B-A4B-it-oQ5 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 5bit; 16k 1,284 124.0 21.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1cytkeza",
   "source_url": "https://omlx.ai/benchmarks/performance/1cytkeza",
   "tg_tokens": null,
   "tps": 124.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-1cytkeza-ctx32768",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: 230f8yl4, 5bfhjct6, 5cd7kw58, d4jzdgp8, htua5wc2, qylystdp, spli5m21; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1133.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "gemma-4-26B-A4B-it-oQ5 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 5bit; 32k 1,133 74.2 23.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1cytkeza",
   "source_url": "https://omlx.ai/benchmarks/performance/1cytkeza",
   "tg_tokens": null,
   "tps": 74.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-1cytkeza-ctx65536",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: 230f8yl4, 5bfhjct6, 5cd7kw58, d4jzdgp8, htua5wc2, qylystdp, spli5m21; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 911.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "gemma-4-26B-A4B-it-oQ5 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 5bit; 64k 911.0 88.0 26.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1cytkeza",
   "source_url": "https://omlx.ai/benchmarks/performance/1cytkeza",
   "tg_tokens": null,
   "tps": 88.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-1cytkeza-ctx131072",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 32.2 GB at this context; duplicate traces of the same session: 230f8yl4, 5bfhjct6, 5cd7kw58, d4jzdgp8, htua5wc2, qylystdp, spli5m21; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 634.9,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "gemma-4-26B-A4B-it-oQ5 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 5bit; 128k 634.9 55.8 32.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1cytkeza",
   "source_url": "https://omlx.ai/benchmarks/performance/1cytkeza",
   "tg_tokens": null,
   "tps": 55.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-1cytkeza-ctx199680",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 36.6 GB at this context; duplicate traces of the same session: 230f8yl4, 5bfhjct6, 5cd7kw58, d4jzdgp8, htua5wc2, qylystdp, spli5m21; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 448.9,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "gemma-4-26B-A4B-it-oQ5 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 5bit; 195k 448.9 49.9 36.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1cytkeza",
   "source_url": "https://omlx.ai/benchmarks/performance/1cytkeza",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-6ztx57cx-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: flu8x61u, l3i9j147; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1506.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "gemma-4-26B-A4B-it-oQ5 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 5bit; 1k 1,506 85.4 18.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ztx57cx",
   "source_url": "https://omlx.ai/benchmarks/performance/6ztx57cx",
   "tg_tokens": null,
   "tps": 85.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-6ztx57cx-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: flu8x61u, l3i9j147; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1445.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "gemma-4-26B-A4B-it-oQ5 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 5bit; 4k 1,445 65.9 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ztx57cx",
   "source_url": "https://omlx.ai/benchmarks/performance/6ztx57cx",
   "tg_tokens": null,
   "tps": 65.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-6ztx57cx-ctx8192",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: flu8x61u, l3i9j147; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1380.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "gemma-4-26B-A4B-it-oQ5 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 5bit; 8k 1,380 125.1 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ztx57cx",
   "source_url": "https://omlx.ai/benchmarks/performance/6ztx57cx",
   "tg_tokens": null,
   "tps": 125.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-6ztx57cx-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: flu8x61u, l3i9j147; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1284.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "gemma-4-26B-A4B-it-oQ5 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 5bit; 16k 1,284 87.0 21.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ztx57cx",
   "source_url": "https://omlx.ai/benchmarks/performance/6ztx57cx",
   "tg_tokens": null,
   "tps": 87.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-v4wcy10n-ctx1024",
   "model": "Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored",
   "notes": "MTP speculative decoding enabled; DavidAU build; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: mzjvoyto, s4lgvakh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 816.2,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "DavidAU_Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored-oQ8-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 816.2 43.7 33.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v4wcy10n",
   "source_url": "https://omlx.ai/benchmarks/performance/v4wcy10n",
   "tg_tokens": null,
   "tps": 43.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-v4wcy10n-ctx4096",
   "model": "Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored",
   "notes": "MTP speculative decoding enabled; DavidAU build; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: mzjvoyto, s4lgvakh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 892.6,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "DavidAU_Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored-oQ8-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 892.6 45.6 34.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v4wcy10n",
   "source_url": "https://omlx.ai/benchmarks/performance/v4wcy10n",
   "tg_tokens": null,
   "tps": 45.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-v4wcy10n-ctx8192",
   "model": "Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored",
   "notes": "MTP speculative decoding enabled; DavidAU build; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: mzjvoyto, s4lgvakh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 838.4,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "DavidAU_Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored-oQ8-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 838.4 46.1 35.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v4wcy10n",
   "source_url": "https://omlx.ai/benchmarks/performance/v4wcy10n",
   "tg_tokens": null,
   "tps": 46.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ljh00eag-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.2 GB at this context; duplicate traces of the same session: jithrjqt, udq4c9bn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2339.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,339 108.4 173.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ljh00eag",
   "source_url": "https://omlx.ai/benchmarks/performance/ljh00eag",
   "tg_tokens": null,
   "tps": 108.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ljh00eag-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.7 GB at this context; duplicate traces of the same session: jithrjqt, udq4c9bn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2318.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,318 84.1 173.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ljh00eag",
   "source_url": "https://omlx.ai/benchmarks/performance/ljh00eag",
   "tg_tokens": null,
   "tps": 84.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ljh00eag-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.9 GB at this context; duplicate traces of the same session: jithrjqt, udq4c9bn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2266.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,266 124.0 174.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ljh00eag",
   "source_url": "https://omlx.ai/benchmarks/performance/ljh00eag",
   "tg_tokens": null,
   "tps": 124.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-i84rst80-ctx1024",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.5 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1305.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,305 63.5 177.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i84rst80",
   "source_url": "https://omlx.ai/benchmarks/performance/i84rst80",
   "tg_tokens": null,
   "tps": 63.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-i84rst80-ctx4096",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.3 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1929.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,929 55.4 178.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i84rst80",
   "source_url": "https://omlx.ai/benchmarks/performance/i84rst80",
   "tg_tokens": null,
   "tps": 55.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-qpq1ysrt-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.4 GB at this context; duplicate traces of the same session: wuxv2i3h; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1333.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,333 62.7 173.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qpq1ysrt",
   "source_url": "https://omlx.ai/benchmarks/performance/qpq1ysrt",
   "tg_tokens": null,
   "tps": 62.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-qpq1ysrt-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.3 GB at this context; duplicate traces of the same session: wuxv2i3h; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1969.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,969 57.2 174.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qpq1ysrt",
   "source_url": "https://omlx.ai/benchmarks/performance/qpq1ysrt",
   "tg_tokens": null,
   "tps": 57.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-512b67pm-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.6 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1311.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,311 60.6 175.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/512b67pm",
   "source_url": "https://omlx.ai/benchmarks/performance/512b67pm",
   "tg_tokens": null,
   "tps": 60.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-512b67pm-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.0 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2006.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 2,006 52.8 177.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/512b67pm",
   "source_url": "https://omlx.ai/benchmarks/performance/512b67pm",
   "tg_tokens": null,
   "tps": 52.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-jnl4h0kr-ctx1024",
   "model": "Qwen2.5-Coder-14B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 8.5 GB at this context; duplicate traces of the same session: m4cuo0u5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 146.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-14B-Instruct-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 146.3 18.9 8.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jnl4h0kr",
   "source_url": "https://omlx.ai/benchmarks/performance/jnl4h0kr",
   "tg_tokens": null,
   "tps": 18.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-jnl4h0kr-ctx4096",
   "model": "Qwen2.5-Coder-14B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 9.1 GB at this context; duplicate traces of the same session: m4cuo0u5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 139.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-14B-Instruct-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 139.6 17.9 9.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jnl4h0kr",
   "source_url": "https://omlx.ai/benchmarks/performance/jnl4h0kr",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-scdvesjb-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: 2vh4zifh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1449.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 1,449 83.0 27.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/scdvesjb",
   "source_url": "https://omlx.ai/benchmarks/performance/scdvesjb",
   "tg_tokens": null,
   "tps": 83.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-scdvesjb-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: 2vh4zifh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2344.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 2,344 80.8 28.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/scdvesjb",
   "source_url": "https://omlx.ai/benchmarks/performance/scdvesjb",
   "tg_tokens": null,
   "tps": 80.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-r09f4db7-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: dcu98dmo; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 469.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 469.9 18.3 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r09f4db7",
   "source_url": "https://omlx.ai/benchmarks/performance/r09f4db7",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-r09f4db7-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: dcu98dmo; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 512.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 512.6 17.9 17.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r09f4db7",
   "source_url": "https://omlx.ai/benchmarks/performance/r09f4db7",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-a2roqkhq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: hhuge2lo, zstfqp4k; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 436.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 436.0 27.2 17.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a2roqkhq",
   "source_url": "https://omlx.ai/benchmarks/performance/a2roqkhq",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-a2roqkhq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: hhuge2lo, zstfqp4k; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 462.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 462.8 38.9 18.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a2roqkhq",
   "source_url": "https://omlx.ai/benchmarks/performance/a2roqkhq",
   "tg_tokens": null,
   "tps": 38.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-a2roqkhq-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: hhuge2lo, zstfqp4k; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 397.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 397.8 43.0 20.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a2roqkhq",
   "source_url": "https://omlx.ai/benchmarks/performance/a2roqkhq",
   "tg_tokens": null,
   "tps": 43.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-qu3z4e5w-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 265.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 265.5 27.4 20.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qu3z4e5w",
   "source_url": "https://omlx.ai/benchmarks/performance/qu3z4e5w",
   "tg_tokens": null,
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-qu3z4e5w-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 246.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 246.2 25.5 21.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qu3z4e5w",
   "source_url": "https://omlx.ai/benchmarks/performance/qu3z4e5w",
   "tg_tokens": null,
   "tps": 25.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-qu3z4e5w-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 201.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 201.6 20.9 25.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qu3z4e5w",
   "source_url": "https://omlx.ai/benchmarks/performance/qu3z4e5w",
   "tg_tokens": null,
   "tps": 20.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-dejo9osr-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 52.0 GB; TTFT 20445 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 801.4,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 2bit; 16k tokens 801.4 PP tok/s 58.6 TG tok/s 20445 TTFT (ms) 52.0 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dejo9osr",
   "source_url": "https://omlx.ai/benchmarks/performance/dejo9osr",
   "tg_tokens": null,
   "tps": 58.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-gtokxicj-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 75.1 GB; TTFT 88793 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 738.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k tokens 738.1 PP tok/s 64.3 TG tok/s 88793 TTFT (ms) 75.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gtokxicj",
   "source_url": "https://omlx.ai/benchmarks/performance/gtokxicj",
   "tg_tokens": null,
   "tps": 64.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2bvnhav6-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 70.9 GB; TTFT 5783 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 708.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 708.2 PP tok/s 51.3 TG tok/s 5783 TTFT (ms) 70.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2bvnhav6",
   "source_url": "https://omlx.ai/benchmarks/performance/2bvnhav6",
   "tg_tokens": null,
   "tps": 51.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bng7qnn2-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.0 GB at this context; duplicate traces of the same session: 9sn1y45i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 695.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 695.0 79.9 99.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bng7qnn2",
   "source_url": "https://omlx.ai/benchmarks/performance/bng7qnn2",
   "tg_tokens": null,
   "tps": 79.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bng7qnn2-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.5 GB at this context; duplicate traces of the same session: 9sn1y45i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1132.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,132 60.0 99.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bng7qnn2",
   "source_url": "https://omlx.ai/benchmarks/performance/bng7qnn2",
   "tg_tokens": null,
   "tps": 60.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-nia1qvk9-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: 4nld8ygo, o048mb1w, ph4qfbmw, s64n8bxi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 789.0,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 789.0 51.3 33.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nia1qvk9",
   "source_url": "https://omlx.ai/benchmarks/performance/nia1qvk9",
   "tg_tokens": null,
   "tps": 51.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nia1qvk9-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: 4nld8ygo, o048mb1w, ph4qfbmw, s64n8bxi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 686.6,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 686.6 35.2 34.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nia1qvk9",
   "source_url": "https://omlx.ai/benchmarks/performance/nia1qvk9",
   "tg_tokens": null,
   "tps": 35.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-nia1qvk9-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: 4nld8ygo, o048mb1w, ph4qfbmw, s64n8bxi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 644.0,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 644.0 43.9 35.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nia1qvk9",
   "source_url": "https://omlx.ai/benchmarks/performance/nia1qvk9",
   "tg_tokens": null,
   "tps": 43.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-nia1qvk9-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: 4nld8ygo, o048mb1w, ph4qfbmw, s64n8bxi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 635.4,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 635.4 54.3 35.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nia1qvk9",
   "source_url": "https://omlx.ai/benchmarks/performance/nia1qvk9",
   "tg_tokens": null,
   "tps": 54.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-nia1qvk9-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: 4nld8ygo, o048mb1w, ph4qfbmw, s64n8bxi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 614.0,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 614.0 54.1 36.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nia1qvk9",
   "source_url": "https://omlx.ai/benchmarks/performance/nia1qvk9",
   "tg_tokens": null,
   "tps": 54.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-yq410fky-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1506.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "gemma-4-26B-A4B-it-oQ5 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 5bit; 1k 1,506 85.4 18.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yq410fky",
   "source_url": "https://omlx.ai/benchmarks/performance/yq410fky",
   "tg_tokens": null,
   "tps": 85.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-yq410fky-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1445.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "gemma-4-26B-A4B-it-oQ5 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 5bit; 4k 1,445 65.9 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yq410fky",
   "source_url": "https://omlx.ai/benchmarks/performance/yq410fky",
   "tg_tokens": null,
   "tps": 65.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-yq410fky-ctx8192",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1380.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "gemma-4-26B-A4B-it-oQ5 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 5bit; 8k 1,380 125.1 20.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yq410fky",
   "source_url": "https://omlx.ai/benchmarks/performance/yq410fky",
   "tg_tokens": null,
   "tps": 125.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-yq410fky-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1284.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "gemma-4-26B-A4B-it-oQ5 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 5bit; 16k 1,284 87.0 21.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yq410fky",
   "source_url": "https://omlx.ai/benchmarks/performance/yq410fky",
   "tg_tokens": null,
   "tps": 87.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-tq1j2ogh-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 2i8xhfje, 2kp3tiuk, 338k2eb4, 4h0myxvy, gisx4l78, gvdd6xk3, jg53yc5m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1502.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "gemma-4-26B-A4B-it-oQ5 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 5bit; 1k 1,502 83.5 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tq1j2ogh",
   "source_url": "https://omlx.ai/benchmarks/performance/tq1j2ogh",
   "tg_tokens": null,
   "tps": 83.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-tq1j2ogh-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: 2i8xhfje, 2kp3tiuk, 338k2eb4, 4h0myxvy, gisx4l78, gvdd6xk3, jg53yc5m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1383.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "gemma-4-26B-A4B-it-oQ5 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 5bit; 4k 1,383 103.3 18.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tq1j2ogh",
   "source_url": "https://omlx.ai/benchmarks/performance/tq1j2ogh",
   "tg_tokens": null,
   "tps": 103.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-tq1j2ogh-ctx8192",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 2i8xhfje, 2kp3tiuk, 338k2eb4, 4h0myxvy, gisx4l78, gvdd6xk3, jg53yc5m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1357.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "gemma-4-26B-A4B-it-oQ5 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 5bit; 8k 1,357 80.9 19.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tq1j2ogh",
   "source_url": "https://omlx.ai/benchmarks/performance/tq1j2ogh",
   "tg_tokens": null,
   "tps": 80.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-tq1j2ogh-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 2i8xhfje, 2kp3tiuk, 338k2eb4, 4h0myxvy, gisx4l78, gvdd6xk3, jg53yc5m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1269.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "gemma-4-26B-A4B-it-oQ5 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 5bit; 16k 1,269 75.1 20.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tq1j2ogh",
   "source_url": "https://omlx.ai/benchmarks/performance/tq1j2ogh",
   "tg_tokens": null,
   "tps": 75.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-tq1j2ogh-ctx32768",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: 2i8xhfje, 2kp3tiuk, 338k2eb4, 4h0myxvy, gisx4l78, gvdd6xk3, jg53yc5m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1131.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "gemma-4-26B-A4B-it-oQ5 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 5bit; 32k 1,131 81.8 22.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tq1j2ogh",
   "source_url": "https://omlx.ai/benchmarks/performance/tq1j2ogh",
   "tg_tokens": null,
   "tps": 81.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-tq1j2ogh-ctx65536",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: 2i8xhfje, 2kp3tiuk, 338k2eb4, 4h0myxvy, gisx4l78, gvdd6xk3, jg53yc5m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 910.9,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "gemma-4-26B-A4B-it-oQ5 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 5bit; 64k 910.9 85.5 25.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tq1j2ogh",
   "source_url": "https://omlx.ai/benchmarks/performance/tq1j2ogh",
   "tg_tokens": null,
   "tps": 85.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-tq1j2ogh-ctx131072",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 31.6 GB at this context; duplicate traces of the same session: 2i8xhfje, 2kp3tiuk, 338k2eb4, 4h0myxvy, gisx4l78, gvdd6xk3, jg53yc5m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 639.6,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "gemma-4-26B-A4B-it-oQ5 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 5bit; 128k 639.6 54.5 31.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tq1j2ogh",
   "source_url": "https://omlx.ai/benchmarks/performance/tq1j2ogh",
   "tg_tokens": null,
   "tps": 54.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-tq1j2ogh-ctx199680",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 36.0 GB at this context; duplicate traces of the same session: 2i8xhfje, 2kp3tiuk, 338k2eb4, 4h0myxvy, gisx4l78, gvdd6xk3, jg53yc5m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 475.3,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "gemma-4-26B-A4B-it-oQ5 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 5bit; 195k 475.3 51.6 36.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tq1j2ogh",
   "source_url": "https://omlx.ai/benchmarks/performance/tq1j2ogh",
   "tg_tokens": null,
   "tps": 51.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ivq2ldri-ctx1024",
   "model": "granite-4.2-30b",
   "notes": "oMLX community benchmark trace; peak memory 16.2 GB at this context; duplicate traces of the same session: ai8n30gm, eusipk7h, smuvayqi; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 97.3,
   "provenance": "community",
   "quant": "q4",
   "quote": "ibm-granite--granite-4.2-30b-q4-mlx on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 97.3 17.6 16.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ivq2ldri",
   "source_url": "https://omlx.ai/benchmarks/performance/ivq2ldri",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ivq2ldri-ctx4096",
   "model": "granite-4.2-30b",
   "notes": "oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: ai8n30gm, eusipk7h, smuvayqi; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 97.1,
   "provenance": "community",
   "quant": "q4",
   "quote": "ibm-granite--granite-4.2-30b-q4-mlx on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 97.1 16.2 17.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ivq2ldri",
   "source_url": "https://omlx.ai/benchmarks/performance/ivq2ldri",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ivq2ldri-ctx8192",
   "model": "granite-4.2-30b",
   "notes": "oMLX community benchmark trace; peak memory 18.2 GB at this context; duplicate traces of the same session: ai8n30gm, eusipk7h, smuvayqi; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 93.1,
   "provenance": "community",
   "quant": "q4",
   "quote": "ibm-granite--granite-4.2-30b-q4-mlx on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 93.1 15.2 18.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ivq2ldri",
   "source_url": "https://omlx.ai/benchmarks/performance/ivq2ldri",
   "tg_tokens": null,
   "tps": 15.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ivq2ldri-ctx16384",
   "model": "granite-4.2-30b",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: ai8n30gm, eusipk7h, smuvayqi; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 84.9,
   "provenance": "community",
   "quant": "q4",
   "quote": "ibm-granite--granite-4.2-30b-q4-mlx on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 84.9 12.5 20.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ivq2ldri",
   "source_url": "https://omlx.ai/benchmarks/performance/ivq2ldri",
   "tg_tokens": null,
   "tps": 12.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-r9jyatqd-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter build; oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: 2y3sxv6c, kz7pz4he; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 870.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "orcarouter_Qwen3.8-27B-Uncensored-MLX-8-bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 870.1 50.6 32.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r9jyatqd",
   "source_url": "https://omlx.ai/benchmarks/performance/r9jyatqd",
   "tg_tokens": null,
   "tps": 50.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-r9jyatqd-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter build; oMLX community benchmark trace; peak memory 33.7 GB at this context; duplicate traces of the same session: 2y3sxv6c, kz7pz4he; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 909.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "orcarouter_Qwen3.8-27B-Uncensored-MLX-8-bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 909.3 36.6 33.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r9jyatqd",
   "source_url": "https://omlx.ai/benchmarks/performance/r9jyatqd",
   "tg_tokens": null,
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-r9jyatqd-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter build; oMLX community benchmark trace; peak memory 34.6 GB at this context; duplicate traces of the same session: 2y3sxv6c, kz7pz4he; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 840.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "orcarouter_Qwen3.8-27B-Uncensored-MLX-8-bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 840.8 35.3 34.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r9jyatqd",
   "source_url": "https://omlx.ai/benchmarks/performance/r9jyatqd",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ujn61ekt-ctx1024",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: pnkil114, t3tvwpb3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1808.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,808 124.6 20.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ujn61ekt",
   "source_url": "https://omlx.ai/benchmarks/performance/ujn61ekt",
   "tg_tokens": null,
   "tps": 124.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ujn61ekt-ctx4096",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: pnkil114, t3tvwpb3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2124.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 2,124 103.8 21.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ujn61ekt",
   "source_url": "https://omlx.ai/benchmarks/performance/ujn61ekt",
   "tg_tokens": null,
   "tps": 103.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ujn61ekt-ctx16384",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: pnkil114, t3tvwpb3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1699.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,699 108.0 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ujn61ekt",
   "source_url": "https://omlx.ai/benchmarks/performance/ujn61ekt",
   "tg_tokens": null,
   "tps": 108.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-4k96bfqm-ctx1024",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: b7eh700d, id7z48ue; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1707.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ6e-MTP on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 1,707 92.0 27.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4k96bfqm",
   "source_url": "https://omlx.ai/benchmarks/performance/4k96bfqm",
   "tg_tokens": null,
   "tps": 92.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-4k96bfqm-ctx4096",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: b7eh700d, id7z48ue; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2044.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ6e-MTP on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 2,044 48.3 29.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4k96bfqm",
   "source_url": "https://omlx.ai/benchmarks/performance/4k96bfqm",
   "tg_tokens": null,
   "tps": 48.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-4k96bfqm-ctx16384",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.1 GB at this context; duplicate traces of the same session: b7eh700d, id7z48ue; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1624.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ6e-MTP on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k 1,624 89.0 30.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4k96bfqm",
   "source_url": "https://omlx.ai/benchmarks/performance/4k96bfqm",
   "tg_tokens": null,
   "tps": 89.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-bn8o4nwm-ctx1024",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.5 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1305.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,305 63.5 177.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bn8o4nwm",
   "source_url": "https://omlx.ai/benchmarks/performance/bn8o4nwm",
   "tg_tokens": null,
   "tps": 63.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-bn8o4nwm-ctx4096",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.3 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1929.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,929 55.4 178.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bn8o4nwm",
   "source_url": "https://omlx.ai/benchmarks/performance/bn8o4nwm",
   "tg_tokens": null,
   "tps": 55.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-sruzzz95-ctx1024",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.5 GB at this context; duplicate traces of the same session: 4brbndu8; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1254.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,254 45.5 177.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sruzzz95",
   "source_url": "https://omlx.ai/benchmarks/performance/sruzzz95",
   "tg_tokens": null,
   "tps": 45.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-sruzzz95-ctx4096",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.3 GB at this context; duplicate traces of the same session: 4brbndu8; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1933.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,933 57.9 178.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sruzzz95",
   "source_url": "https://omlx.ai/benchmarks/performance/sruzzz95",
   "tg_tokens": null,
   "tps": 57.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-u2kf97t0-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.6 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1311.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,311 60.6 175.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u2kf97t0",
   "source_url": "https://omlx.ai/benchmarks/performance/u2kf97t0",
   "tg_tokens": null,
   "tps": 60.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-u2kf97t0-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.0 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2006.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 2,006 52.8 177.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u2kf97t0",
   "source_url": "https://omlx.ai/benchmarks/performance/u2kf97t0",
   "tg_tokens": null,
   "tps": 52.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-xe3621qr-ctx1024",
   "model": "Qwen3.6-35B-A3B-Fable-Holo3.1-Qwopus-KAT-Coder-C-qx86-hi",
   "notes": "community merge; oMLX community benchmark trace; peak memory 37.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2042.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-Fable-Holo3.1-Qwopus-KAT-Coder-C-qx86-hi-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 2,042 71.4 37.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xe3621qr",
   "source_url": "https://omlx.ai/benchmarks/performance/xe3621qr",
   "tg_tokens": null,
   "tps": 71.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-xe3621qr-ctx4096",
   "model": "Qwen3.6-35B-A3B-Fable-Holo3.1-Qwopus-KAT-Coder-C-qx86-hi",
   "notes": "community merge; oMLX community benchmark trace; peak memory 37.9 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3050.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-Fable-Holo3.1-Qwopus-KAT-Coder-C-qx86-hi-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 3,050 69.4 37.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xe3621qr",
   "source_url": "https://omlx.ai/benchmarks/performance/xe3621qr",
   "tg_tokens": null,
   "tps": 69.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-hfos38tc-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 695efqg7, bhufkt7k, puxjouip, r14666o5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 928.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 928.0 78.3 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hfos38tc",
   "source_url": "https://omlx.ai/benchmarks/performance/hfos38tc",
   "tg_tokens": null,
   "tps": 78.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-hfos38tc-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 695efqg7, bhufkt7k, puxjouip, r14666o5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 980.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 980.1 75.3 20.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hfos38tc",
   "source_url": "https://omlx.ai/benchmarks/performance/hfos38tc",
   "tg_tokens": null,
   "tps": 75.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-hfos38tc-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 695efqg7, bhufkt7k, puxjouip, r14666o5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 935.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 935.7 72.9 21.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hfos38tc",
   "source_url": "https://omlx.ai/benchmarks/performance/hfos38tc",
   "tg_tokens": null,
   "tps": 72.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-hfos38tc-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 695efqg7, bhufkt7k, puxjouip, r14666o5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 830.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 830.3 67.6 21.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hfos38tc",
   "source_url": "https://omlx.ai/benchmarks/performance/hfos38tc",
   "tg_tokens": null,
   "tps": 67.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-hfos38tc-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 695efqg7, bhufkt7k, puxjouip, r14666o5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 690.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 690.8 59.4 21.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hfos38tc",
   "source_url": "https://omlx.ai/benchmarks/performance/hfos38tc",
   "tg_tokens": null,
   "tps": 59.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-isdn4hj7-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: j9szjktp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 693.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 693.8 76.2 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/isdn4hj7",
   "source_url": "https://omlx.ai/benchmarks/performance/isdn4hj7",
   "tg_tokens": null,
   "tps": 76.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-isdn4hj7-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: j9szjktp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 960.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 960.4 75.5 20.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/isdn4hj7",
   "source_url": "https://omlx.ai/benchmarks/performance/isdn4hj7",
   "tg_tokens": null,
   "tps": 75.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zvbf202m-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: l5k7ce4b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 701.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 701.4 142.0 20.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zvbf202m",
   "source_url": "https://omlx.ai/benchmarks/performance/zvbf202m",
   "tg_tokens": null,
   "tps": 142.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zvbf202m-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: l5k7ce4b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2419.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,419 134.4 20.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zvbf202m",
   "source_url": "https://omlx.ai/benchmarks/performance/zvbf202m",
   "tg_tokens": null,
   "tps": 134.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-qmqbcqad-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 4ivahl8l; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 109.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 109.8 19.5 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qmqbcqad",
   "source_url": "https://omlx.ai/benchmarks/performance/qmqbcqad",
   "tg_tokens": null,
   "tps": 19.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-qmqbcqad-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 4ivahl8l; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 103.9 18.8 19.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qmqbcqad",
   "source_url": "https://omlx.ai/benchmarks/performance/qmqbcqad",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-exmnptwd-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 208.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 208.7 21.8 16.0 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/exmnptwd",
   "source_url": "https://omlx.ai/benchmarks/performance/exmnptwd",
   "tg_tokens": null,
   "tps": 21.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-exmnptwd-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 204.6 21.6 17.0 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/exmnptwd",
   "source_url": "https://omlx.ai/benchmarks/performance/exmnptwd",
   "tg_tokens": null,
   "tps": 21.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-f3j4q0vm-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: zpoxbup3; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 107.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 107.4 16.3 20.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f3j4q0vm",
   "source_url": "https://omlx.ai/benchmarks/performance/f3j4q0vm",
   "tg_tokens": null,
   "tps": 16.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-f3j4q0vm-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: zpoxbup3; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 108.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 108.2 16.0 23.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f3j4q0vm",
   "source_url": "https://omlx.ai/benchmarks/performance/f3j4q0vm",
   "tg_tokens": null,
   "tps": 16.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-n3dgkahb-ctx1024",
   "model": "Qwen3.8-27B-OBLITERATED-Mythos-Class-Agentic",
   "notes": "oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 431.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-OBLITERATED-Mythos-Class-Agentic-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 431.5 12.6 22.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n3dgkahb",
   "source_url": "https://omlx.ai/benchmarks/performance/n3dgkahb",
   "tg_tokens": null,
   "tps": 12.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-n3dgkahb-ctx4096",
   "model": "Qwen3.8-27B-OBLITERATED-Mythos-Class-Agentic",
   "notes": "oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 480.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-OBLITERATED-Mythos-Class-Agentic-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 480.6 12.7 23.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n3dgkahb",
   "source_url": "https://omlx.ai/benchmarks/performance/n3dgkahb",
   "tg_tokens": null,
   "tps": 12.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-0faduooa-ctx1024",
   "model": "Qwen3.8-27B-UD-Q3_K_XL-DFlash2",
   "notes": "DFlash2 build; oMLX community benchmark trace; peak memory 13.3 GB at this context; duplicate traces of the same session: k4sb9201; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 111.3,
   "provenance": "community",
   "quant": "UD-Q3_K_XL",
   "quote": "Qwen3.8-27B-UD-Q3_K_XL-DFlash2-MLX on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 3bit; 1k 111.3 17.9 13.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0faduooa",
   "source_url": "https://omlx.ai/benchmarks/performance/0faduooa",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-0faduooa-ctx4096",
   "model": "Qwen3.8-27B-UD-Q3_K_XL-DFlash2",
   "notes": "DFlash2 build; oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: k4sb9201; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 110.1,
   "provenance": "community",
   "quant": "UD-Q3_K_XL",
   "quote": "Qwen3.8-27B-UD-Q3_K_XL-DFlash2-MLX on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 3bit; 4k 110.1 17.5 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0faduooa",
   "source_url": "https://omlx.ai/benchmarks/performance/0faduooa",
   "tg_tokens": null,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-dbzi5goy-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 23hdhu7w, 2i6wyqv0, vh9vw549; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 293.8,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 293.8 60.0 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dbzi5goy",
   "source_url": "https://omlx.ai/benchmarks/performance/dbzi5goy",
   "tg_tokens": null,
   "tps": 60.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-dbzi5goy-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: 23hdhu7w, 2i6wyqv0, vh9vw549; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 319.9,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 319.9 52.9 23.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dbzi5goy",
   "source_url": "https://omlx.ai/benchmarks/performance/dbzi5goy",
   "tg_tokens": null,
   "tps": 52.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-dbzi5goy-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: 23hdhu7w, 2i6wyqv0, vh9vw549; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 313.1,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 8k 313.1 63.5 23.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dbzi5goy",
   "source_url": "https://omlx.ai/benchmarks/performance/dbzi5goy",
   "tg_tokens": null,
   "tps": 63.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-dbzi5goy-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: 23hdhu7w, 2i6wyqv0, vh9vw549; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 298.4,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 16k 298.4 65.9 25.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dbzi5goy",
   "source_url": "https://omlx.ai/benchmarks/performance/dbzi5goy",
   "tg_tokens": null,
   "tps": 65.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-2rqlac08-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 18.5 GB; TTFT 11658 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1405.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k tokens 1,405 PP tok/s 46.6 TG tok/s 11658 TTFT (ms) 18.5 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2rqlac08",
   "source_url": "https://omlx.ai/benchmarks/performance/2rqlac08",
   "tg_tokens": null,
   "tps": 46.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-snbtmfay-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.2 GB at this context; duplicate traces of the same session: ifa97s3z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 580.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 580.2 62.0 69.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/snbtmfay",
   "source_url": "https://omlx.ai/benchmarks/performance/snbtmfay",
   "tg_tokens": null,
   "tps": 62.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-snbtmfay-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.3 GB at this context; duplicate traces of the same session: ifa97s3z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1276.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,276 50.9 69.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/snbtmfay",
   "source_url": "https://omlx.ai/benchmarks/performance/snbtmfay",
   "tg_tokens": null,
   "tps": 50.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0rakqbr7-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: 13zlwu03, 337p2khh, 56ca00jl, e1sbxwd6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 791.6,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 791.6 52.1 33.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0rakqbr7",
   "source_url": "https://omlx.ai/benchmarks/performance/0rakqbr7",
   "tg_tokens": null,
   "tps": 52.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0rakqbr7-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: 13zlwu03, 337p2khh, 56ca00jl, e1sbxwd6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 777.3,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 777.3 37.9 34.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0rakqbr7",
   "source_url": "https://omlx.ai/benchmarks/performance/0rakqbr7",
   "tg_tokens": null,
   "tps": 37.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0rakqbr7-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: 13zlwu03, 337p2khh, 56ca00jl, e1sbxwd6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 702.4,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 702.4 42.7 35.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0rakqbr7",
   "source_url": "https://omlx.ai/benchmarks/performance/0rakqbr7",
   "tg_tokens": null,
   "tps": 42.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0rakqbr7-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: 13zlwu03, 337p2khh, 56ca00jl, e1sbxwd6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 654.6,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 654.6 56.3 35.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0rakqbr7",
   "source_url": "https://omlx.ai/benchmarks/performance/0rakqbr7",
   "tg_tokens": null,
   "tps": 56.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0rakqbr7-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: 13zlwu03, 337p2khh, 56ca00jl, e1sbxwd6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 599.5,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 599.5 56.8 36.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0rakqbr7",
   "source_url": "https://omlx.ai/benchmarks/performance/0rakqbr7",
   "tg_tokens": null,
   "tps": 56.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-4jsp0nlo-ctx1024",
   "model": "Ternary-Bonsai-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 8.9 GB at this context; duplicate traces of the same session: e6u1x3zr; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 321.5,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-27B-mlx-2bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 2bit; 1k 321.5 18.5 8.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4jsp0nlo",
   "source_url": "https://omlx.ai/benchmarks/performance/4jsp0nlo",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-4jsp0nlo-ctx4096",
   "model": "Ternary-Bonsai-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 10.0 GB at this context; duplicate traces of the same session: e6u1x3zr; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 333.0,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-27B-mlx-2bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 2bit; 4k 333.0 18.0 10.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4jsp0nlo",
   "source_url": "https://omlx.ai/benchmarks/performance/4jsp0nlo",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-mh75zx2n-ctx1024",
   "model": "gemma-4-12B-it-qat",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 8.3 GB at this context; duplicate traces of the same session: 2lrjjstd; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 76.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-12B-it-qat-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 76.5 9.2 8.3 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mh75zx2n",
   "source_url": "https://omlx.ai/benchmarks/performance/mh75zx2n",
   "tg_tokens": null,
   "tps": 9.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-mh75zx2n-ctx4096",
   "model": "gemma-4-12B-it-qat",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 8.6 GB at this context; duplicate traces of the same session: 2lrjjstd; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 75.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-12B-it-qat-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 75.5 9.3 8.6 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mh75zx2n",
   "source_url": "https://omlx.ai/benchmarks/performance/mh75zx2n",
   "tg_tokens": null,
   "tps": 9.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-6zl84jii-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "Unsloth UD build; oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: bgawo62s; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 618.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "unsloth--gemma-4-26b-a4b-it-UD-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 618.7 53.4 14.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6zl84jii",
   "source_url": "https://omlx.ai/benchmarks/performance/6zl84jii",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-6zl84jii-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "Unsloth UD build; oMLX community benchmark trace; peak memory 15.1 GB at this context; duplicate traces of the same session: bgawo62s; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 632.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "unsloth--gemma-4-26b-a4b-it-UD-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 632.2 51.4 15.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6zl84jii",
   "source_url": "https://omlx.ai/benchmarks/performance/6zl84jii",
   "tg_tokens": null,
   "tps": 51.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-05vvfonk-ctx131072",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 198.7 GB; TTFT 967137 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 135.5,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-oQ3e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 128k tokens 135.5 PP tok/s 15.0 TG tok/s 967137 TTFT (ms) 198.7 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/05vvfonk",
   "source_url": "https://omlx.ai/benchmarks/performance/05vvfonk",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-17nvzmyf-ctx65536",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 180.1 GB; TTFT 34829 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1882.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k tokens 1,882 PP tok/s 66.9 TG tok/s 34829 TTFT (ms) 180.1 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/17nvzmyf",
   "source_url": "https://omlx.ai/benchmarks/performance/17nvzmyf",
   "tg_tokens": null,
   "tps": 66.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-u8m2sjc2-ctx32768",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 179.5 GB; TTFT 17230 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1902.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k tokens 1,902 PP tok/s 72.5 TG tok/s 17230 TTFT (ms) 179.5 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u8m2sjc2",
   "source_url": "https://omlx.ai/benchmarks/performance/u8m2sjc2",
   "tg_tokens": null,
   "tps": 72.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-2irh0u6x-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 175.5 GB; TTFT 34188 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1917.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k tokens 1,917 PP tok/s 55.8 TG tok/s 34188 TTFT (ms) 175.5 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2irh0u6x",
   "source_url": "https://omlx.ai/benchmarks/performance/2irh0u6x",
   "tg_tokens": null,
   "tps": 55.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-xtz35jaf-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 176.2 GB; TTFT 69975 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1873.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k tokens 1,873 PP tok/s 53.0 TG tok/s 69975 TTFT (ms) 176.2 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xtz35jaf",
   "source_url": "https://omlx.ai/benchmarks/performance/xtz35jaf",
   "tg_tokens": null,
   "tps": 53.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-b4gh11eo-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 175.2 GB; TTFT 16870 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1942.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k tokens 1,942 PP tok/s 56.6 TG tok/s 16870 TTFT (ms) 175.2 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b4gh11eo",
   "source_url": "https://omlx.ai/benchmarks/performance/b4gh11eo",
   "tg_tokens": null,
   "tps": 56.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-kld7z6i7-ctx1024",
   "model": "LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated-v3",
   "notes": "oMLX community benchmark trace; peak memory 9.4 GB at this context; duplicate traces of the same session: nkdkcn8w, zqn8befj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3255.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated-v3-8bit-MLX on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 3,255 178.0 9.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kld7z6i7",
   "source_url": "https://omlx.ai/benchmarks/performance/kld7z6i7",
   "tg_tokens": null,
   "tps": 178.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-kld7z6i7-ctx4096",
   "model": "LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated-v3",
   "notes": "oMLX community benchmark trace; peak memory 9.8 GB at this context; duplicate traces of the same session: nkdkcn8w, zqn8befj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3982.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated-v3-8bit-MLX on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 3,982 174.4 9.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kld7z6i7",
   "source_url": "https://omlx.ai/benchmarks/performance/kld7z6i7",
   "tg_tokens": null,
   "tps": 174.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-kld7z6i7-ctx16384",
   "model": "LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated-v3",
   "notes": "oMLX community benchmark trace; peak memory 9.9 GB at this context; duplicate traces of the same session: nkdkcn8w, zqn8befj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3450.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated-v3-8bit-MLX on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 16k 3,450 162.0 9.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kld7z6i7",
   "source_url": "https://omlx.ai/benchmarks/performance/kld7z6i7",
   "tg_tokens": null,
   "tps": 162.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-djeob09k-ctx1024",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: 35hs4gj2; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 632.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ6e on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 6bit; 1k 632.6 67.6 26.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/djeob09k",
   "source_url": "https://omlx.ai/benchmarks/performance/djeob09k",
   "tg_tokens": null,
   "tps": 67.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-djeob09k-ctx16384",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: 35hs4gj2; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 677.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ6e on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 6bit; 16k 677.2 61.2 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/djeob09k",
   "source_url": "https://omlx.ai/benchmarks/performance/djeob09k",
   "tg_tokens": null,
   "tps": 61.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qzmg08bu-ctx1024",
   "model": "Qwen3.6-35B-A3B-Fable-Holo3.1-Qwopus-KAT-Coder-C-qx86-hi",
   "notes": "community merge; oMLX community benchmark trace; peak memory 37.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2042.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-Fable-Holo3.1-Qwopus-KAT-Coder-C-qx86-hi-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 2,042 71.4 37.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qzmg08bu",
   "source_url": "https://omlx.ai/benchmarks/performance/qzmg08bu",
   "tg_tokens": null,
   "tps": 71.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qzmg08bu-ctx4096",
   "model": "Qwen3.6-35B-A3B-Fable-Holo3.1-Qwopus-KAT-Coder-C-qx86-hi",
   "notes": "community merge; oMLX community benchmark trace; peak memory 37.9 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3050.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-Fable-Holo3.1-Qwopus-KAT-Coder-C-qx86-hi-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 3,050 69.4 37.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qzmg08bu",
   "source_url": "https://omlx.ai/benchmarks/performance/qzmg08bu",
   "tg_tokens": null,
   "tps": 69.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-hxz98ryg-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: 19vjimtr; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 787.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 787.3 78.4 20.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hxz98ryg",
   "source_url": "https://omlx.ai/benchmarks/performance/hxz98ryg",
   "tg_tokens": null,
   "tps": 78.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-hxz98ryg-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 19vjimtr; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2102.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,102 76.5 21.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hxz98ryg",
   "source_url": "https://omlx.ai/benchmarks/performance/hxz98ryg",
   "tg_tokens": null,
   "tps": 76.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-vax2nbuc-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 208.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 208.7 21.8 16.0 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vax2nbuc",
   "source_url": "https://omlx.ai/benchmarks/performance/vax2nbuc",
   "tg_tokens": null,
   "tps": 21.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-vax2nbuc-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 204.6 21.6 17.0 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vax2nbuc",
   "source_url": "https://omlx.ai/benchmarks/performance/vax2nbuc",
   "tg_tokens": null,
   "tps": 21.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-aeawb4lb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 5z54z0ub; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 376.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 376.3 17.3 16.0 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aeawb4lb",
   "source_url": "https://omlx.ai/benchmarks/performance/aeawb4lb",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-aeawb4lb-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 5z54z0ub; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 405.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 405.3 17.0 17.0 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aeawb4lb",
   "source_url": "https://omlx.ai/benchmarks/performance/aeawb4lb",
   "tg_tokens": null,
   "tps": 17.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-tzr2pogv-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 28.3 GB at this context; duplicate traces of the same session: 0eqmi2ak, 3lzd0l2h, 5slxklpj, h0q5sztu, ifsdttj5, nv2g6rtj; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 84.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 1k 84.2 10.6 28.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tzr2pogv",
   "source_url": "https://omlx.ai/benchmarks/performance/tzr2pogv",
   "tg_tokens": null,
   "tps": 10.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-tzr2pogv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 32.4 GB at this context; duplicate traces of the same session: 0eqmi2ak, 3lzd0l2h, 5slxklpj, h0q5sztu, ifsdttj5, nv2g6rtj; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 83.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 4k 83.5 10.5 32.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tzr2pogv",
   "source_url": "https://omlx.ai/benchmarks/performance/tzr2pogv",
   "tg_tokens": null,
   "tps": 10.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-tzr2pogv-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 33.6 GB at this context; duplicate traces of the same session: 0eqmi2ak, 3lzd0l2h, 5slxklpj, h0q5sztu, ifsdttj5, nv2g6rtj; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 80.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 8k 80.9 10.1 33.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tzr2pogv",
   "source_url": "https://omlx.ai/benchmarks/performance/tzr2pogv",
   "tg_tokens": null,
   "tps": 10.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-tzr2pogv-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 34.2 GB at this context; duplicate traces of the same session: 0eqmi2ak, 3lzd0l2h, 5slxklpj, h0q5sztu, ifsdttj5, nv2g6rtj; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 76.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 16k 76.1 9.5 34.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tzr2pogv",
   "source_url": "https://omlx.ai/benchmarks/performance/tzr2pogv",
   "tg_tokens": null,
   "tps": 9.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-tzr2pogv-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: 0eqmi2ak, 3lzd0l2h, 5slxklpj, h0q5sztu, ifsdttj5, nv2g6rtj; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 68.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 32k 68.0 8.5 35.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tzr2pogv",
   "source_url": "https://omlx.ai/benchmarks/performance/tzr2pogv",
   "tg_tokens": null,
   "tps": 8.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-tzr2pogv-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: 0eqmi2ak, 3lzd0l2h, 5slxklpj, h0q5sztu, ifsdttj5, nv2g6rtj; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 56.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 64k 56.0 6.9 37.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tzr2pogv",
   "source_url": "https://omlx.ai/benchmarks/performance/tzr2pogv",
   "tg_tokens": null,
   "tps": 6.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-tzr2pogv-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 41.2 GB at this context; duplicate traces of the same session: 0eqmi2ak, 3lzd0l2h, 5slxklpj, h0q5sztu, ifsdttj5, nv2g6rtj; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 41.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 128k 41.6 5.2 41.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tzr2pogv",
   "source_url": "https://omlx.ai/benchmarks/performance/tzr2pogv",
   "tg_tokens": null,
   "tps": 5.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-mv2tdz7f-ctx1024",
   "model": "Qwen3.8-27B-OBLITERATED-Mythos-Class-Agentic",
   "notes": "oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 431.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-OBLITERATED-Mythos-Class-Agentic-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 431.5 12.6 22.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mv2tdz7f",
   "source_url": "https://omlx.ai/benchmarks/performance/mv2tdz7f",
   "tg_tokens": null,
   "tps": 12.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-mv2tdz7f-ctx4096",
   "model": "Qwen3.8-27B-OBLITERATED-Mythos-Class-Agentic",
   "notes": "oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 480.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-OBLITERATED-Mythos-Class-Agentic-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 480.6 12.7 23.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mv2tdz7f",
   "source_url": "https://omlx.ai/benchmarks/performance/mv2tdz7f",
   "tg_tokens": null,
   "tps": 12.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-mb366lao-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: hnbkqzrh; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 111.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 111.4 18.3 16.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mb366lao",
   "source_url": "https://omlx.ai/benchmarks/performance/mb366lao",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-mb366lao-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: hnbkqzrh; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 103.9 17.7 20.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mb366lao",
   "source_url": "https://omlx.ai/benchmarks/performance/mb366lao",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-98z29wqk-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: c966c0cc, o01hzuli, qzez8d2d, vxu0j3sz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 129.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 129.9 14.8 16.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/98z29wqk",
   "source_url": "https://omlx.ai/benchmarks/performance/98z29wqk",
   "tg_tokens": null,
   "tps": 14.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-98z29wqk-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: c966c0cc, o01hzuli, qzez8d2d, vxu0j3sz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 130.3 14.4 17.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/98z29wqk",
   "source_url": "https://omlx.ai/benchmarks/performance/98z29wqk",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-98z29wqk-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: c966c0cc, o01hzuli, qzez8d2d, vxu0j3sz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 128.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 128.6 13.7 18.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/98z29wqk",
   "source_url": "https://omlx.ai/benchmarks/performance/98z29wqk",
   "tg_tokens": null,
   "tps": 13.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-98z29wqk-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: c966c0cc, o01hzuli, qzez8d2d, vxu0j3sz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 115.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 115.7 13.6 18.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/98z29wqk",
   "source_url": "https://omlx.ai/benchmarks/performance/98z29wqk",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-98z29wqk-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: c966c0cc, o01hzuli, qzez8d2d, vxu0j3sz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 110.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 110.0 12.6 19.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/98z29wqk",
   "source_url": "https://omlx.ai/benchmarks/performance/98z29wqk",
   "tg_tokens": null,
   "tps": 12.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-tapzogtp-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 19.9 GB; TTFT 24924 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 657.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k tokens 657.4 PP tok/s 35.3 TG tok/s 24924 TTFT (ms) 19.9 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tapzogtp",
   "source_url": "https://omlx.ai/benchmarks/performance/tapzogtp",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-a93yeyg1-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "pi build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: dj9c251x, rth9vb7k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 261.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-pi-4bit-mtp-mlx on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 261.1 29.4 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a93yeyg1",
   "source_url": "https://omlx.ai/benchmarks/performance/a93yeyg1",
   "tg_tokens": null,
   "tps": 29.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-a93yeyg1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "pi build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: dj9c251x, rth9vb7k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 242.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-pi-4bit-mtp-mlx on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 242.1 27.2 19.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a93yeyg1",
   "source_url": "https://omlx.ai/benchmarks/performance/a93yeyg1",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-a93yeyg1-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "pi build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: dj9c251x, rth9vb7k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 215.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-pi-4bit-mtp-mlx on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 215.2 24.4 20.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a93yeyg1",
   "source_url": "https://omlx.ai/benchmarks/performance/a93yeyg1",
   "tg_tokens": null,
   "tps": 24.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-cxc1lr20-ctx16384",
   "model": "Qwen3.8-9B-Distill-Uncensored-Heretic",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 7.8 GB; TTFT 20190 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 811.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-9B-Distill-uncensored-heretic-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k tokens 811.5 PP tok/s 48.6 TG tok/s 20190 TTFT (ms) 7.8 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cxc1lr20",
   "source_url": "https://omlx.ai/benchmarks/performance/cxc1lr20",
   "tg_tokens": null,
   "tps": 48.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-clj6qxlt-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 94.4 GB; TTFT 37101 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3533.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 5bit; 128k tokens 3,533 PP tok/s 97.0 TG tok/s 37101 TTFT (ms) 94.4 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/clj6qxlt",
   "source_url": "https://omlx.ai/benchmarks/performance/clj6qxlt",
   "tg_tokens": null,
   "tps": 97.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-84yhaftv-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 8.0 GB at this context; duplicate traces of the same session: h6km07ot; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1451.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-e4b-8bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 8bit; 4k 1,451 22.9 8.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/84yhaftv",
   "source_url": "https://omlx.ai/benchmarks/performance/84yhaftv",
   "tg_tokens": null,
   "tps": 22.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-84yhaftv-ctx16384",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 8.2 GB at this context; duplicate traces of the same session: h6km07ot; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1715.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-e4b-8bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 8bit; 16k 1,715 20.7 8.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/84yhaftv",
   "source_url": "https://omlx.ai/benchmarks/performance/84yhaftv",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-bdu1szbr-ctx1024",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: 786k5g7p; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 643.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ6e on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 6bit; 1k 643.3 68.5 26.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bdu1szbr",
   "source_url": "https://omlx.ai/benchmarks/performance/bdu1szbr",
   "tg_tokens": null,
   "tps": 68.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-bdu1szbr-ctx16384",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: 786k5g7p; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 677.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ6e on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 6bit; 16k 677.1 62.3 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bdu1szbr",
   "source_url": "https://omlx.ai/benchmarks/performance/bdu1szbr",
   "tg_tokens": null,
   "tps": 62.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-7kxkvwq4-ctx1024",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: uui5d9ha; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 850.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ6e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 6bit; 1k 850.4 67.9 27.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7kxkvwq4",
   "source_url": "https://omlx.ai/benchmarks/performance/7kxkvwq4",
   "tg_tokens": null,
   "tps": 67.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-7kxkvwq4-ctx16384",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: uui5d9ha; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 883.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ6e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 6bit; 16k 883.5 62.0 28.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7kxkvwq4",
   "source_url": "https://omlx.ai/benchmarks/performance/7kxkvwq4",
   "tg_tokens": null,
   "tps": 62.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-iylrclfb-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 9md82gv9, svu8sbr4; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1367.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k 1,367 60.0 19.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iylrclfb",
   "source_url": "https://omlx.ai/benchmarks/performance/iylrclfb",
   "tg_tokens": null,
   "tps": 60.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-iylrclfb-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 9md82gv9, svu8sbr4; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1720.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 1,720 58.5 19.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iylrclfb",
   "source_url": "https://omlx.ai/benchmarks/performance/iylrclfb",
   "tg_tokens": null,
   "tps": 58.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-iylrclfb-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 9md82gv9, svu8sbr4; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1543.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 16k 1,543 50.7 20.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iylrclfb",
   "source_url": "https://omlx.ai/benchmarks/performance/iylrclfb",
   "tg_tokens": null,
   "tps": 50.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-xz6u50hb-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: b3woadtw, hwcfm0yt, ly97tb58; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1812.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 4k 1,812 105.7 21.3 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xz6u50hb",
   "source_url": "https://omlx.ai/benchmarks/performance/xz6u50hb",
   "tg_tokens": null,
   "tps": 105.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-xz6u50hb-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: b3woadtw, hwcfm0yt, ly97tb58; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1736.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 8k 1,736 96.7 21.4 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xz6u50hb",
   "source_url": "https://omlx.ai/benchmarks/performance/xz6u50hb",
   "tg_tokens": null,
   "tps": 96.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-xz6u50hb-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: b3woadtw, hwcfm0yt, ly97tb58; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1505.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 16k 1,505 92.5 21.9 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xz6u50hb",
   "source_url": "https://omlx.ai/benchmarks/performance/xz6u50hb",
   "tg_tokens": null,
   "tps": 92.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-xz6u50hb-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: b3woadtw, hwcfm0yt, ly97tb58; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1179.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 32k 1,179 84.3 22.7 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xz6u50hb",
   "source_url": "https://omlx.ai/benchmarks/performance/xz6u50hb",
   "tg_tokens": null,
   "tps": 84.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lpb2ddp7-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: 2kmzhu4l, gehuhf18, hhd3y682, ja7oc6q2, k9lqdd3w, m9wr93rr, q0xn5yj2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 311.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 311.2 43.5 34.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lpb2ddp7",
   "source_url": "https://omlx.ai/benchmarks/performance/lpb2ddp7",
   "tg_tokens": null,
   "tps": 43.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lpb2ddp7-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 38.6 GB at this context; duplicate traces of the same session: 2kmzhu4l, gehuhf18, hhd3y682, ja7oc6q2, k9lqdd3w, m9wr93rr, q0xn5yj2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 330.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 330.6 42.0 38.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lpb2ddp7",
   "source_url": "https://omlx.ai/benchmarks/performance/lpb2ddp7",
   "tg_tokens": null,
   "tps": 42.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-lpb2ddp7-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 39.7 GB at this context; duplicate traces of the same session: 2kmzhu4l, gehuhf18, hhd3y682, ja7oc6q2, k9lqdd3w, m9wr93rr, q0xn5yj2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 326.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 8k 326.3 63.9 39.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lpb2ddp7",
   "source_url": "https://omlx.ai/benchmarks/performance/lpb2ddp7",
   "tg_tokens": null,
   "tps": 63.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-lpb2ddp7-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 38.8 GB at this context; duplicate traces of the same session: 2kmzhu4l, gehuhf18, hhd3y682, ja7oc6q2, k9lqdd3w, m9wr93rr, q0xn5yj2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 314.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 16k 314.5 58.9 38.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lpb2ddp7",
   "source_url": "https://omlx.ai/benchmarks/performance/lpb2ddp7",
   "tg_tokens": null,
   "tps": 58.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lpb2ddp7-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 40.0 GB at this context; duplicate traces of the same session: 2kmzhu4l, gehuhf18, hhd3y682, ja7oc6q2, k9lqdd3w, m9wr93rr, q0xn5yj2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 291.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 32k 291.4 63.0 40.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lpb2ddp7",
   "source_url": "https://omlx.ai/benchmarks/performance/lpb2ddp7",
   "tg_tokens": null,
   "tps": 63.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-lpb2ddp7-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 42.3 GB at this context; duplicate traces of the same session: 2kmzhu4l, gehuhf18, hhd3y682, ja7oc6q2, k9lqdd3w, m9wr93rr, q0xn5yj2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 252.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 64k 252.7 47.5 42.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lpb2ddp7",
   "source_url": "https://omlx.ai/benchmarks/performance/lpb2ddp7",
   "tg_tokens": null,
   "tps": 47.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-lpb2ddp7-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 47.0 GB at this context; duplicate traces of the same session: 2kmzhu4l, gehuhf18, hhd3y682, ja7oc6q2, k9lqdd3w, m9wr93rr, q0xn5yj2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 198.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 128k 198.6 43.6 47.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lpb2ddp7",
   "source_url": "https://omlx.ai/benchmarks/performance/lpb2ddp7",
   "tg_tokens": null,
   "tps": 43.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-lpb2ddp7-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 52.2 GB at this context; duplicate traces of the same session: 2kmzhu4l, gehuhf18, hhd3y682, ja7oc6q2, k9lqdd3w, m9wr93rr, q0xn5yj2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 162.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 195k 162.1 30.0 52.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lpb2ddp7",
   "source_url": "https://omlx.ai/benchmarks/performance/lpb2ddp7",
   "tg_tokens": null,
   "tps": 30.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-1g76vzgk-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 105.7 GB at this context; duplicate traces of the same session: pjez1nz6; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 783.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 4k 783.8 31.0 105.7 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1g76vzgk",
   "source_url": "https://omlx.ai/benchmarks/performance/1g76vzgk",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-1g76vzgk-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 105.9 GB at this context; duplicate traces of the same session: pjez1nz6; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 782.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 16k 782.1 31.5 105.9 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1g76vzgk",
   "source_url": "https://omlx.ai/benchmarks/performance/1g76vzgk",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ekhw6dy8-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 5n0ujgbt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 463.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 463.6 19.3 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ekhw6dy8",
   "source_url": "https://omlx.ai/benchmarks/performance/ekhw6dy8",
   "tg_tokens": null,
   "tps": 19.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ekhw6dy8-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: 5n0ujgbt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 585.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 585.1 40.2 71.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ekhw6dy8",
   "source_url": "https://omlx.ai/benchmarks/performance/ekhw6dy8",
   "tg_tokens": null,
   "tps": 40.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ibg1u2yu-ctx4096",
   "model": "Gemma 4 E2B",
   "notes": "it-qat build; oMLX community benchmark trace; peak memory 3.8 GB at this context; duplicate traces of the same session: dho5m39j, mwa1bpx7, q5ia8qcm; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3509.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E2B-it-qat-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 3,509 84.9 3.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ibg1u2yu",
   "source_url": "https://omlx.ai/benchmarks/performance/ibg1u2yu",
   "tg_tokens": null,
   "tps": 84.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ibg1u2yu-ctx16384",
   "model": "Gemma 4 E2B",
   "notes": "it-qat build; oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: dho5m39j, mwa1bpx7, q5ia8qcm; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2389.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E2B-it-qat-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 2,389 74.0 3.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ibg1u2yu",
   "source_url": "https://omlx.ai/benchmarks/performance/ibg1u2yu",
   "tg_tokens": null,
   "tps": 74.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ibg1u2yu-ctx65536",
   "model": "Gemma 4 E2B",
   "notes": "it-qat build; oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: dho5m39j, mwa1bpx7, q5ia8qcm; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1496.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E2B-it-qat-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 64k 1,496 50.8 4.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ibg1u2yu",
   "source_url": "https://omlx.ai/benchmarks/performance/ibg1u2yu",
   "tg_tokens": null,
   "tps": 50.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ibg1u2yu-ctx131072",
   "model": "Gemma 4 E2B",
   "notes": "it-qat build; oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: dho5m39j, mwa1bpx7, q5ia8qcm; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1269.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E2B-it-qat-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 128k 1,269 32.4 6.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ibg1u2yu",
   "source_url": "https://omlx.ai/benchmarks/performance/ibg1u2yu",
   "tg_tokens": null,
   "tps": 32.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-gzve37n3-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: xdcsdoc8; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 541.3,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 1k 541.3 76.1 11.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gzve37n3",
   "source_url": "https://omlx.ai/benchmarks/performance/gzve37n3",
   "tg_tokens": null,
   "tps": 76.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-gzve37n3-ctx16384",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 12.1 GB at this context; duplicate traces of the same session: xdcsdoc8; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 567.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 16k 567.0 54.5 12.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gzve37n3",
   "source_url": "https://omlx.ai/benchmarks/performance/gzve37n3",
   "tg_tokens": null,
   "tps": 54.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-6r971o1c-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: b1myf165, uy2zptbw; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 240.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k 240.2 8.1 15.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6r971o1c",
   "source_url": "https://omlx.ai/benchmarks/performance/6r971o1c",
   "tg_tokens": null,
   "tps": 8.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-6r971o1c-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: b1myf165, uy2zptbw; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 247.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 247.8 8.2 17.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6r971o1c",
   "source_url": "https://omlx.ai/benchmarks/performance/6r971o1c",
   "tg_tokens": null,
   "tps": 8.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-6r971o1c-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: b1myf165, uy2zptbw; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 227.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 16k 227.3 7.7 17.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6r971o1c",
   "source_url": "https://omlx.ai/benchmarks/performance/6r971o1c",
   "tg_tokens": null,
   "tps": 7.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-clwrt38o-ctx1024",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "prism-ml MLX build; oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: 7a4dq9m7, xdyroi06; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 155.8,
   "provenance": "community",
   "quant": "2bit",
   "quote": "prism-ml--Ternary-Bonsai-2-27B-mlx-2bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 unknown; 1k 155.8 11.9 11.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/clwrt38o",
   "source_url": "https://omlx.ai/benchmarks/performance/clwrt38o",
   "tg_tokens": null,
   "tps": 11.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-clwrt38o-ctx4096",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "prism-ml MLX build; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: 7a4dq9m7, xdyroi06; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 155.5,
   "provenance": "community",
   "quant": "2bit",
   "quote": "prism-ml--Ternary-Bonsai-2-27B-mlx-2bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 unknown; 4k 155.5 11.7 13.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/clwrt38o",
   "source_url": "https://omlx.ai/benchmarks/performance/clwrt38o",
   "tg_tokens": null,
   "tps": 11.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-clwrt38o-ctx16384",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "prism-ml MLX build; oMLX community benchmark trace; peak memory 15.4 GB at this context; duplicate traces of the same session: 7a4dq9m7, xdyroi06; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 129.0,
   "provenance": "community",
   "quant": "2bit",
   "quote": "prism-ml--Ternary-Bonsai-2-27B-mlx-2bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 unknown; 16k 129.0 9.4 15.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/clwrt38o",
   "source_url": "https://omlx.ai/benchmarks/performance/clwrt38o",
   "tg_tokens": null,
   "tps": 9.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-ykv2av7d-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 106.3 GB; TTFT 42141 ms; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 777.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 32k tokens 777.6 PP tok/s 30.1 TG tok/s 42141 TTFT (ms) 106.3 Peak mem",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ykv2av7d",
   "source_url": "https://omlx.ai/benchmarks/performance/ykv2av7d",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-3wdfcbmc-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: y86xp7l0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 635.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 635.2 80.4 70.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3wdfcbmc",
   "source_url": "https://omlx.ai/benchmarks/performance/3wdfcbmc",
   "tg_tokens": null,
   "tps": 80.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-3wdfcbmc-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: y86xp7l0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 744.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 744.5 48.1 71.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3wdfcbmc",
   "source_url": "https://omlx.ai/benchmarks/performance/3wdfcbmc",
   "tg_tokens": null,
   "tps": 48.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-wqhc5k3r-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.2 GB at this context; duplicate traces of the same session: w2kz0qbx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 634.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 634.3 47.3 71.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wqhc5k3r",
   "source_url": "https://omlx.ai/benchmarks/performance/wqhc5k3r",
   "tg_tokens": null,
   "tps": 47.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-wqhc5k3r-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: w2kz0qbx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 734.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 734.3 63.6 71.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wqhc5k3r",
   "source_url": "https://omlx.ai/benchmarks/performance/wqhc5k3r",
   "tg_tokens": null,
   "tps": 63.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-yxfz9yjg-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 100.5 GB; TTFT 1388 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 737.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 737.5 PP tok/s 80.7 TG tok/s 1388 TTFT (ms) 100.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yxfz9yjg",
   "source_url": "https://omlx.ai/benchmarks/performance/yxfz9yjg",
   "tg_tokens": null,
   "tps": 80.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-fgmo14rm-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 100.7 GB; TTFT 1385 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 739.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 739.3 PP tok/s 80.2 TG tok/s 1385 TTFT (ms) 100.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fgmo14rm",
   "source_url": "https://omlx.ai/benchmarks/performance/fgmo14rm",
   "tg_tokens": null,
   "tps": 80.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wwc19dm7-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: eh6jrmxv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 585.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 585.0 25.7 21.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wwc19dm7",
   "source_url": "https://omlx.ai/benchmarks/performance/wwc19dm7",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wwc19dm7-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: eh6jrmxv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 905.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 905.1 25.5 22.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wwc19dm7",
   "source_url": "https://omlx.ai/benchmarks/performance/wwc19dm7",
   "tg_tokens": null,
   "tps": 25.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gmzhtwmv-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: 64dc7txv, 8a85uyik, 8ygxmvvs, et05sevv, lfdvr1gi, nus57bt4, ww0ocaa8; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 331.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 331.7 43.7 33.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gmzhtwmv",
   "source_url": "https://omlx.ai/benchmarks/performance/gmzhtwmv",
   "tg_tokens": null,
   "tps": 43.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gmzhtwmv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 37.5 GB at this context; duplicate traces of the same session: 64dc7txv, 8a85uyik, 8ygxmvvs, et05sevv, lfdvr1gi, nus57bt4, ww0ocaa8; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 337.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 337.7 35.7 37.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gmzhtwmv",
   "source_url": "https://omlx.ai/benchmarks/performance/gmzhtwmv",
   "tg_tokens": null,
   "tps": 35.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gmzhtwmv-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 38.5 GB at this context; duplicate traces of the same session: 64dc7txv, 8a85uyik, 8ygxmvvs, et05sevv, lfdvr1gi, nus57bt4, ww0ocaa8; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 333.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 8k 333.0 32.4 38.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gmzhtwmv",
   "source_url": "https://omlx.ai/benchmarks/performance/gmzhtwmv",
   "tg_tokens": null,
   "tps": 32.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gmzhtwmv-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 37.8 GB at this context; duplicate traces of the same session: 64dc7txv, 8a85uyik, 8ygxmvvs, et05sevv, lfdvr1gi, nus57bt4, ww0ocaa8; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 321.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 16k 321.3 41.1 37.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gmzhtwmv",
   "source_url": "https://omlx.ai/benchmarks/performance/gmzhtwmv",
   "tg_tokens": null,
   "tps": 41.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gmzhtwmv-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 38.8 GB at this context; duplicate traces of the same session: 64dc7txv, 8a85uyik, 8ygxmvvs, et05sevv, lfdvr1gi, nus57bt4, ww0ocaa8; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 298.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 32k 298.5 36.6 38.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gmzhtwmv",
   "source_url": "https://omlx.ai/benchmarks/performance/gmzhtwmv",
   "tg_tokens": null,
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gmzhtwmv-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 40.8 GB at this context; duplicate traces of the same session: 64dc7txv, 8a85uyik, 8ygxmvvs, et05sevv, lfdvr1gi, nus57bt4, ww0ocaa8; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 260.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 64k 260.2 33.6 40.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gmzhtwmv",
   "source_url": "https://omlx.ai/benchmarks/performance/gmzhtwmv",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gmzhtwmv-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 44.8 GB at this context; duplicate traces of the same session: 64dc7txv, 8a85uyik, 8ygxmvvs, et05sevv, lfdvr1gi, nus57bt4, ww0ocaa8; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 206.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 128k 206.2 30.4 44.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gmzhtwmv",
   "source_url": "https://omlx.ai/benchmarks/performance/gmzhtwmv",
   "tg_tokens": null,
   "tps": 30.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gmzhtwmv-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 49.0 GB at this context; duplicate traces of the same session: 64dc7txv, 8a85uyik, 8ygxmvvs, et05sevv, lfdvr1gi, nus57bt4, ww0ocaa8; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 169.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 195k 169.1 24.0 49.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gmzhtwmv",
   "source_url": "https://omlx.ai/benchmarks/performance/gmzhtwmv",
   "tg_tokens": null,
   "tps": 24.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-esildb8s-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: mzt5fwqw, qo868oh3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 637.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 637.1 61.7 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/esildb8s",
   "source_url": "https://omlx.ai/benchmarks/performance/esildb8s",
   "tg_tokens": null,
   "tps": 61.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-esildb8s-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: mzt5fwqw, qo868oh3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 733.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 733.8 45.0 71.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/esildb8s",
   "source_url": "https://omlx.ai/benchmarks/performance/esildb8s",
   "tg_tokens": null,
   "tps": 45.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-esildb8s-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: mzt5fwqw, qo868oh3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 824.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 824.5 52.6 71.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/esildb8s",
   "source_url": "https://omlx.ai/benchmarks/performance/esildb8s",
   "tg_tokens": null,
   "tps": 52.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-m13kvs96-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.8 GB; TTFT 1895 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 540.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 540.4 PP tok/s 44.6 TG tok/s 1895 TTFT (ms) 70.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m13kvs96",
   "source_url": "https://omlx.ai/benchmarks/performance/m13kvs96",
   "tg_tokens": null,
   "tps": 44.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-fl1c91wf-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 86.4 GB; TTFT 1961 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 522.3,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k tokens 522.3 PP tok/s 37.0 TG tok/s 1961 TTFT (ms) 86.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fl1c91wf",
   "source_url": "https://omlx.ai/benchmarks/performance/fl1c91wf",
   "tg_tokens": null,
   "tps": 37.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-0ea4dc62-ctx16384",
   "model": "Qwen3.5-4B-OptiQ",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 6.0 GB; TTFT 23840 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 687.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "community--Qwen3.5-4B-OptiQ on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k tokens 687.2 PP tok/s 65.6 TG tok/s 23840 TTFT (ms) 6.0 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ea4dc62",
   "source_url": "https://omlx.ai/benchmarks/performance/0ea4dc62",
   "tg_tokens": null,
   "tps": 65.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-aohrqit3-ctx1024",
   "model": "Muse Glimmer",
   "notes": "Abliterated MM build; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: jjtsgdq3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 36.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Muse-Glimmer-30B-Abliterated-MM-oQ5e on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 5bit; 1k 36.7 9.0 21.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aohrqit3",
   "source_url": "https://omlx.ai/benchmarks/performance/aohrqit3",
   "tg_tokens": null,
   "tps": 9.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-aohrqit3-ctx4096",
   "model": "Muse Glimmer",
   "notes": "Abliterated MM build; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: jjtsgdq3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 51.5,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Muse-Glimmer-30B-Abliterated-MM-oQ5e on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 5bit; 4k 51.5 7.4 21.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aohrqit3",
   "source_url": "https://omlx.ai/benchmarks/performance/aohrqit3",
   "tg_tokens": null,
   "tps": 7.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-lavvglkg-ctx1024",
   "model": "Muse Glimmer",
   "notes": "oMLX community benchmark trace; peak memory 15.2 GB at this context; duplicate traces of the same session: 27rxn0fc; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 395.7,
   "provenance": "community",
   "quant": "q4",
   "quote": "Muse-Glimmer-q4-MLX on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 395.7 17.8 15.2 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lavvglkg",
   "source_url": "https://omlx.ai/benchmarks/performance/lavvglkg",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-lavvglkg-ctx4096",
   "model": "Muse Glimmer",
   "notes": "oMLX community benchmark trace; peak memory 15.5 GB at this context; duplicate traces of the same session: 27rxn0fc; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 407.7,
   "provenance": "community",
   "quant": "q4",
   "quote": "Muse-Glimmer-q4-MLX on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 407.7 17.2 15.5 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lavvglkg",
   "source_url": "https://omlx.ai/benchmarks/performance/lavvglkg",
   "tg_tokens": null,
   "tps": 17.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-1l01r99u-ctx65536",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; coder q8 ngram build; oMLX community benchmark trace; peak memory 74.2 GB at this context; duplicate traces of the same session: jpmkzfk9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 567.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-coder-q8-ngram-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 567.7 40.1 74.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1l01r99u",
   "source_url": "https://omlx.ai/benchmarks/performance/1l01r99u",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-1l01r99u-ctx131072",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; coder q8 ngram build; oMLX community benchmark trace; peak memory 78.6 GB at this context; duplicate traces of the same session: jpmkzfk9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 552.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-coder-q8-ngram-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 552.9 35.8 78.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1l01r99u",
   "source_url": "https://omlx.ai/benchmarks/performance/1l01r99u",
   "tg_tokens": null,
   "tps": 35.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-tz9ktggs-ctx1024",
   "model": "Muse Glimmer",
   "notes": "Abliterated MM build; oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: t239rz72; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 153.5,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Muse-Glimmer-30B-Abliterated-MM-oQ5e on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 5bit; 1k 153.5 33.1 24.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tz9ktggs",
   "source_url": "https://omlx.ai/benchmarks/performance/tz9ktggs",
   "tg_tokens": null,
   "tps": 33.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-tz9ktggs-ctx4096",
   "model": "Muse Glimmer",
   "notes": "Abliterated MM build; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: t239rz72; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 184.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Muse-Glimmer-30B-Abliterated-MM-oQ5e on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 5bit; 4k 184.1 23.8 25.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tz9ktggs",
   "source_url": "https://omlx.ai/benchmarks/performance/tz9ktggs",
   "tg_tokens": null,
   "tps": 23.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-w8vuorqp-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: cdigoskn; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1516.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 4k 1,516 53.7 17.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w8vuorqp",
   "source_url": "https://omlx.ai/benchmarks/performance/w8vuorqp",
   "tg_tokens": null,
   "tps": 53.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-w8vuorqp-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: cdigoskn; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1476.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 16k 1,476 50.6 17.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w8vuorqp",
   "source_url": "https://omlx.ai/benchmarks/performance/w8vuorqp",
   "tg_tokens": null,
   "tps": 50.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-jeh7wmlj-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 4ofmbiaf, an350g7y, fvbtzeso, mpkwibuh, phba4y2c, ya3r46zp; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 81.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 81.0 13.9 19.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jeh7wmlj",
   "source_url": "https://omlx.ai/benchmarks/performance/jeh7wmlj",
   "tg_tokens": null,
   "tps": 13.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-jeh7wmlj-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: 4ofmbiaf, an350g7y, fvbtzeso, mpkwibuh, phba4y2c, ya3r46zp; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 80.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 80.8 13.6 24.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jeh7wmlj",
   "source_url": "https://omlx.ai/benchmarks/performance/jeh7wmlj",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-jeh7wmlj-ctx8192",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: 4ofmbiaf, an350g7y, fvbtzeso, mpkwibuh, phba4y2c, ya3r46zp; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 78.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 8k 78.6 13.2 25.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jeh7wmlj",
   "source_url": "https://omlx.ai/benchmarks/performance/jeh7wmlj",
   "tg_tokens": null,
   "tps": 13.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-jeh7wmlj-ctx16384",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: 4ofmbiaf, an350g7y, fvbtzeso, mpkwibuh, phba4y2c, ya3r46zp; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 73.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k 73.9 12.1 25.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jeh7wmlj",
   "source_url": "https://omlx.ai/benchmarks/performance/jeh7wmlj",
   "tg_tokens": null,
   "tps": 12.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-jeh7wmlj-ctx32768",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 26.9 GB at this context; duplicate traces of the same session: 4ofmbiaf, an350g7y, fvbtzeso, mpkwibuh, phba4y2c, ya3r46zp; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 66.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 32k 66.2 10.4 26.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jeh7wmlj",
   "source_url": "https://omlx.ai/benchmarks/performance/jeh7wmlj",
   "tg_tokens": null,
   "tps": 10.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-jeh7wmlj-ctx65536",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 28.8 GB at this context; duplicate traces of the same session: 4ofmbiaf, an350g7y, fvbtzeso, mpkwibuh, phba4y2c, ya3r46zp; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 55.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 64k 55.1 8.0 28.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jeh7wmlj",
   "source_url": "https://omlx.ai/benchmarks/performance/jeh7wmlj",
   "tg_tokens": null,
   "tps": 8.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-jeh7wmlj-ctx131072",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 32.8 GB at this context; duplicate traces of the same session: 4ofmbiaf, an350g7y, fvbtzeso, mpkwibuh, phba4y2c, ya3r46zp; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 41.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 128k 41.0 5.5 32.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jeh7wmlj",
   "source_url": "https://omlx.ai/benchmarks/performance/jeh7wmlj",
   "tg_tokens": null,
   "tps": 5.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-hou90kcx-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: 75wjtftr; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1309.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 4k 1,309 50.0 18.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hou90kcx",
   "source_url": "https://omlx.ai/benchmarks/performance/hou90kcx",
   "tg_tokens": null,
   "tps": 50.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-hou90kcx-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 75wjtftr; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1500.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 16k 1,500 108.3 19.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hou90kcx",
   "source_url": "https://omlx.ai/benchmarks/performance/hou90kcx",
   "tg_tokens": null,
   "tps": 108.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-mwuv1bl0-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 40.1 GB at this context; duplicate traces of the same session: 2pa4axnz, 8ijqvuvf, akmlejpb, c1ja4ylo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 250.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 250.6 42.2 40.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mwuv1bl0",
   "source_url": "https://omlx.ai/benchmarks/performance/mwuv1bl0",
   "tg_tokens": null,
   "tps": 42.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-mwuv1bl0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.4 GB at this context; duplicate traces of the same session: 2pa4axnz, 8ijqvuvf, akmlejpb, c1ja4ylo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 328.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 328.2 36.3 43.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mwuv1bl0",
   "source_url": "https://omlx.ai/benchmarks/performance/mwuv1bl0",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-mwuv1bl0-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.8 GB at this context; duplicate traces of the same session: 2pa4axnz, 8ijqvuvf, akmlejpb, c1ja4ylo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 331.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 331.7 39.8 43.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mwuv1bl0",
   "source_url": "https://omlx.ai/benchmarks/performance/mwuv1bl0",
   "tg_tokens": null,
   "tps": 39.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-mwuv1bl0-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 44.9 GB at this context; duplicate traces of the same session: 2pa4axnz, 8ijqvuvf, akmlejpb, c1ja4ylo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 316.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 316.7 47.8 44.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mwuv1bl0",
   "source_url": "https://omlx.ai/benchmarks/performance/mwuv1bl0",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-mwuv1bl0-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 46.7 GB at this context; duplicate traces of the same session: 2pa4axnz, 8ijqvuvf, akmlejpb, c1ja4ylo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 288.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 288.3 47.1 46.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mwuv1bl0",
   "source_url": "https://omlx.ai/benchmarks/performance/mwuv1bl0",
   "tg_tokens": null,
   "tps": 47.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-95dqsvcm-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.3 GB at this context; duplicate traces of the same session: fybenvsu, r5uehsaj, v099i8c9, w22kbohv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 603.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 603.0 68.8 69.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/95dqsvcm",
   "source_url": "https://omlx.ai/benchmarks/performance/95dqsvcm",
   "tg_tokens": null,
   "tps": 68.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-95dqsvcm-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.9 GB at this context; duplicate traces of the same session: fybenvsu, r5uehsaj, v099i8c9, w22kbohv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 735.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 735.9 60.5 69.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/95dqsvcm",
   "source_url": "https://omlx.ai/benchmarks/performance/95dqsvcm",
   "tg_tokens": null,
   "tps": 60.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-95dqsvcm-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.9 GB at this context; duplicate traces of the same session: fybenvsu, r5uehsaj, v099i8c9, w22kbohv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 844.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 844.7 61.5 69.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/95dqsvcm",
   "source_url": "https://omlx.ai/benchmarks/performance/95dqsvcm",
   "tg_tokens": null,
   "tps": 61.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-95dqsvcm-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.1 GB at this context; duplicate traces of the same session: fybenvsu, r5uehsaj, v099i8c9, w22kbohv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 840.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 840.5 58.7 70.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/95dqsvcm",
   "source_url": "https://omlx.ai/benchmarks/performance/95dqsvcm",
   "tg_tokens": null,
   "tps": 58.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-95dqsvcm-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: fybenvsu, r5uehsaj, v099i8c9, w22kbohv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 836.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 836.3 58.6 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/95dqsvcm",
   "source_url": "https://omlx.ai/benchmarks/performance/95dqsvcm",
   "tg_tokens": null,
   "tps": 58.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n7yoyipb-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: 7eu7a6ux, itv57k8f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 927.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 927.3 32.0 17.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n7yoyipb",
   "source_url": "https://omlx.ai/benchmarks/performance/n7yoyipb",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n7yoyipb-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 7eu7a6ux, itv57k8f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 982.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 982.2 31.1 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n7yoyipb",
   "source_url": "https://omlx.ai/benchmarks/performance/n7yoyipb",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n7yoyipb-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: 7eu7a6ux, itv57k8f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 631.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 631.6 21.3 24.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n7yoyipb",
   "source_url": "https://omlx.ai/benchmarks/performance/n7yoyipb",
   "tg_tokens": null,
   "tps": 21.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-75wjtftr-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1309.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 4k 1,309 50.0 18.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/75wjtftr",
   "source_url": "https://omlx.ai/benchmarks/performance/75wjtftr",
   "tg_tokens": null,
   "tps": 50.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-75wjtftr-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1500.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 16k 1,500 108.3 19.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/75wjtftr",
   "source_url": "https://omlx.ai/benchmarks/performance/75wjtftr",
   "tg_tokens": null,
   "tps": 108.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-c1ja4ylo-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 40.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 250.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 250.6 42.2 40.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c1ja4ylo",
   "source_url": "https://omlx.ai/benchmarks/performance/c1ja4ylo",
   "tg_tokens": null,
   "tps": 42.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-c1ja4ylo-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 328.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 328.2 36.3 43.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c1ja4ylo",
   "source_url": "https://omlx.ai/benchmarks/performance/c1ja4ylo",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-c1ja4ylo-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 331.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 331.7 39.8 43.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c1ja4ylo",
   "source_url": "https://omlx.ai/benchmarks/performance/c1ja4ylo",
   "tg_tokens": null,
   "tps": 39.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-c1ja4ylo-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 44.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 316.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 316.7 47.8 44.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c1ja4ylo",
   "source_url": "https://omlx.ai/benchmarks/performance/c1ja4ylo",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-c1ja4ylo-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 46.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 288.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 288.3 47.1 46.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c1ja4ylo",
   "source_url": "https://omlx.ai/benchmarks/performance/c1ja4ylo",
   "tg_tokens": null,
   "tps": 47.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-r5uehsaj-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.3 GB at this context; duplicate traces of the same session: v099i8c9, w22kbohv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 603.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 603.0 68.8 69.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r5uehsaj",
   "source_url": "https://omlx.ai/benchmarks/performance/r5uehsaj",
   "tg_tokens": null,
   "tps": 68.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-r5uehsaj-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.9 GB at this context; duplicate traces of the same session: v099i8c9, w22kbohv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 735.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 735.9 60.5 69.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r5uehsaj",
   "source_url": "https://omlx.ai/benchmarks/performance/r5uehsaj",
   "tg_tokens": null,
   "tps": 60.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-r5uehsaj-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.9 GB at this context; duplicate traces of the same session: v099i8c9, w22kbohv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 844.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 844.7 61.5 69.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r5uehsaj",
   "source_url": "https://omlx.ai/benchmarks/performance/r5uehsaj",
   "tg_tokens": null,
   "tps": 61.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-r5uehsaj-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.1 GB at this context; duplicate traces of the same session: v099i8c9, w22kbohv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 840.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 840.5 58.7 70.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r5uehsaj",
   "source_url": "https://omlx.ai/benchmarks/performance/r5uehsaj",
   "tg_tokens": null,
   "tps": 58.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-r5uehsaj-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: v099i8c9, w22kbohv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 836.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 836.3 58.6 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r5uehsaj",
   "source_url": "https://omlx.ai/benchmarks/performance/r5uehsaj",
   "tg_tokens": null,
   "tps": 58.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-zhbuafrf-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 10.8 GB at this context; duplicate traces of the same session: utlq3lc1; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 798.1,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "gpt-oss-20b-oQ4 on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 798.1 64.6 10.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zhbuafrf",
   "source_url": "https://omlx.ai/benchmarks/performance/zhbuafrf",
   "tg_tokens": null,
   "tps": 64.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-zhbuafrf-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 10.9 GB at this context; duplicate traces of the same session: utlq3lc1; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1422.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "gpt-oss-20b-oQ4 on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,422 61.5 10.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zhbuafrf",
   "source_url": "https://omlx.ai/benchmarks/performance/zhbuafrf",
   "tg_tokens": null,
   "tps": 61.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ryfhc5zf-ctx32768",
   "model": "Flash-Next",
   "notes": "MTP speculative decoding enabled; name as published, likely Qwen3.8-Flash-Next shorthand; oMLX community benchmark trace, single run at 32k context; peak memory 109.4 GB; TTFT 17457 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1877.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Flash-Next-4bit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k tokens 1,877 PP tok/s 35.4 TG tok/s 17457 TTFT (ms) 109.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ryfhc5zf",
   "source_url": "https://omlx.ai/benchmarks/performance/ryfhc5zf",
   "tg_tokens": null,
   "tps": 35.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ilkk5zy4-ctx32768",
   "model": "Flash-Next",
   "notes": "MTP speculative decoding enabled; name as published, likely Qwen3.8-Flash-Next shorthand; oMLX community benchmark trace, single run at 32k context; peak memory 106.1 GB; TTFT 20464 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1601.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Flash-Next-4bit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k tokens 1,601 PP tok/s 20.7 TG tok/s 20464 TTFT (ms) 106.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ilkk5zy4",
   "source_url": "https://omlx.ai/benchmarks/performance/ilkk5zy4",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-x1q1sfos-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: qczmtp95; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 480.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 480.2 30.2 18.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x1q1sfos",
   "source_url": "https://omlx.ai/benchmarks/performance/x1q1sfos",
   "tg_tokens": null,
   "tps": 30.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-x1q1sfos-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: qczmtp95; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 440.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 440.1 24.0 19.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x1q1sfos",
   "source_url": "https://omlx.ai/benchmarks/performance/x1q1sfos",
   "tg_tokens": null,
   "tps": 24.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-49kpoak1-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: 0nyzfy1c, 5r9nwbkx, 78qb1ud6, brw9hiz0, l17kfk39, rnkvblmk, ufxhiitx; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 330.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 330.9 52.7 32.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/49kpoak1",
   "source_url": "https://omlx.ai/benchmarks/performance/49kpoak1",
   "tg_tokens": null,
   "tps": 52.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-49kpoak1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.0 GB at this context; duplicate traces of the same session: 0nyzfy1c, 5r9nwbkx, 78qb1ud6, brw9hiz0, l17kfk39, rnkvblmk, ufxhiitx; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 335.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 335.3 38.0 34.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/49kpoak1",
   "source_url": "https://omlx.ai/benchmarks/performance/49kpoak1",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-49kpoak1-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 35.3 GB at this context; duplicate traces of the same session: 0nyzfy1c, 5r9nwbkx, 78qb1ud6, brw9hiz0, l17kfk39, rnkvblmk, ufxhiitx; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 330.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 8k 330.2 39.9 35.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/49kpoak1",
   "source_url": "https://omlx.ai/benchmarks/performance/49kpoak1",
   "tg_tokens": null,
   "tps": 39.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-49kpoak1-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 36.0 GB at this context; duplicate traces of the same session: 0nyzfy1c, 5r9nwbkx, 78qb1ud6, brw9hiz0, l17kfk39, rnkvblmk, ufxhiitx; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 315.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 16k 315.1 30.2 36.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/49kpoak1",
   "source_url": "https://omlx.ai/benchmarks/performance/49kpoak1",
   "tg_tokens": null,
   "tps": 30.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-49kpoak1-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 38.5 GB at this context; duplicate traces of the same session: 0nyzfy1c, 5r9nwbkx, 78qb1ud6, brw9hiz0, l17kfk39, rnkvblmk, ufxhiitx; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 290.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 32k 290.0 8.1 38.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/49kpoak1",
   "source_url": "https://omlx.ai/benchmarks/performance/49kpoak1",
   "tg_tokens": null,
   "tps": 8.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-49kpoak1-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 43.3 GB at this context; duplicate traces of the same session: 0nyzfy1c, 5r9nwbkx, 78qb1ud6, brw9hiz0, l17kfk39, rnkvblmk, ufxhiitx; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 249.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 64k 249.5 2.7 43.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/49kpoak1",
   "source_url": "https://omlx.ai/benchmarks/performance/49kpoak1",
   "tg_tokens": null,
   "tps": 2.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-49kpoak1-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 51.6 GB at this context; duplicate traces of the same session: 0nyzfy1c, 5r9nwbkx, 78qb1ud6, brw9hiz0, l17kfk39, rnkvblmk, ufxhiitx; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 194.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 128k 194.3 2.4 51.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/49kpoak1",
   "source_url": "https://omlx.ai/benchmarks/performance/49kpoak1",
   "tg_tokens": null,
   "tps": 2.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-08",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-49kpoak1-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 47.6 GB at this context; duplicate traces of the same session: 0nyzfy1c, 5r9nwbkx, 78qb1ud6, brw9hiz0, l17kfk39, rnkvblmk, ufxhiitx; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 157.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 195k 157.7 0.8 47.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/49kpoak1",
   "source_url": "https://omlx.ai/benchmarks/performance/49kpoak1",
   "tg_tokens": null,
   "tps": 0.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-fl0avuh3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: yqo3alg8; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 374.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 374.6 18.9 19.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fl0avuh3",
   "source_url": "https://omlx.ai/benchmarks/performance/fl0avuh3",
   "tg_tokens": null,
   "tps": 18.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-fl0avuh3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: yqo3alg8; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 416.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 416.8 19.7 20.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fl0avuh3",
   "source_url": "https://omlx.ai/benchmarks/performance/fl0avuh3",
   "tg_tokens": null,
   "tps": 19.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-vtioy0lj-ctx8192",
   "model": "K2-Horizon-7B",
   "notes": "oMLX community benchmark trace; peak memory 10.7 GB at this context; duplicate traces of the same session: cp1nob2o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 803.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "K2-Horizon-7B-oQ8e on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 803.8 45.4 10.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vtioy0lj",
   "source_url": "https://omlx.ai/benchmarks/performance/vtioy0lj",
   "tg_tokens": null,
   "tps": 45.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-vtioy0lj-ctx32768",
   "model": "K2-Horizon-7B",
   "notes": "oMLX community benchmark trace; peak memory 14.1 GB at this context; duplicate traces of the same session: cp1nob2o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 525.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "K2-Horizon-7B-oQ8e on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 525.4 34.5 14.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vtioy0lj",
   "source_url": "https://omlx.ai/benchmarks/performance/vtioy0lj",
   "tg_tokens": null,
   "tps": 34.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-s0mt2df8-ctx1024",
   "model": "Ling-3.0-tiny",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 30.4 GB; TTFT 266 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3854.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ling-3.0-tiny on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 unknown; 1k tokens 3,854 PP tok/s 107.2 TG tok/s 266 TTFT (ms) 30.4 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s0mt2df8",
   "source_url": "https://omlx.ai/benchmarks/performance/s0mt2df8",
   "tg_tokens": null,
   "tps": 107.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8u5xfw1b-ctx1024",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 30.8 GB at this context; duplicate traces of the same session: l6se5lr1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4730.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 4,730 92.2 30.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8u5xfw1b",
   "source_url": "https://omlx.ai/benchmarks/performance/8u5xfw1b",
   "tg_tokens": null,
   "tps": 92.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8u5xfw1b-ctx4096",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 31.3 GB at this context; duplicate traces of the same session: l6se5lr1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3781.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 3,781 87.9 31.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8u5xfw1b",
   "source_url": "https://omlx.ai/benchmarks/performance/8u5xfw1b",
   "tg_tokens": null,
   "tps": 87.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-uzuifqsw-ctx1024",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 30.8 GB; TTFT 208 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4923.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k tokens 4,923 PP tok/s 95.5 TG tok/s 208 TTFT (ms) 30.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uzuifqsw",
   "source_url": "https://omlx.ai/benchmarks/performance/uzuifqsw",
   "tg_tokens": null,
   "tps": 95.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-57tq3cry-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.5 GB at this context; duplicate traces of the same session: 60aazml7; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 287.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 287.8 9.1 16.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/57tq3cry",
   "source_url": "https://omlx.ai/benchmarks/performance/57tq3cry",
   "tg_tokens": null,
   "tps": 9.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-57tq3cry-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 60aazml7; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 300.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 300.8 9.0 17.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/57tq3cry",
   "source_url": "https://omlx.ai/benchmarks/performance/57tq3cry",
   "tg_tokens": null,
   "tps": 9.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-5b91ygfv-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: ibqoob1l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 436.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 436.3 59.0 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5b91ygfv",
   "source_url": "https://omlx.ai/benchmarks/performance/5b91ygfv",
   "tg_tokens": null,
   "tps": 59.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-5b91ygfv-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.5 GB at this context; duplicate traces of the same session: ibqoob1l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 595.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 595.9 37.3 78.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5b91ygfv",
   "source_url": "https://omlx.ai/benchmarks/performance/5b91ygfv",
   "tg_tokens": null,
   "tps": 37.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-snjlgfbw-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: w7cqytbw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 409.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 409.5 48.8 70.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/snjlgfbw",
   "source_url": "https://omlx.ai/benchmarks/performance/snjlgfbw",
   "tg_tokens": null,
   "tps": 48.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-snjlgfbw-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.2 GB at this context; duplicate traces of the same session: w7cqytbw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 620.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 620.3 36.4 74.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/snjlgfbw",
   "source_url": "https://omlx.ai/benchmarks/performance/snjlgfbw",
   "tg_tokens": null,
   "tps": 36.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-wdy77w1z-ctx4096",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 7.7 GB at this context; duplicate traces of the same session: iawenduo; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 224.5 31.5 7.7 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wdy77w1z",
   "source_url": "https://omlx.ai/benchmarks/performance/wdy77w1z",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-wdy77w1z-ctx16384",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 8.3 GB at this context; duplicate traces of the same session: iawenduo; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 213.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 16k 213.1 30.3 8.3 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wdy77w1z",
   "source_url": "https://omlx.ai/benchmarks/performance/wdy77w1z",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-0mujqbib-ctx1024",
   "model": "gemma-4-12B-it-qat",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.3 GB at this context; duplicate traces of the same session: j2ugmydx; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 560.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-12B-it-qat-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 560.9 19.6 7.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0mujqbib",
   "source_url": "https://omlx.ai/benchmarks/performance/0mujqbib",
   "tg_tokens": null,
   "tps": 19.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-0mujqbib-ctx4096",
   "model": "gemma-4-12B-it-qat",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.8 GB at this context; duplicate traces of the same session: j2ugmydx; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 625.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-12B-it-qat-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 625.4 19.3 7.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0mujqbib",
   "source_url": "https://omlx.ai/benchmarks/performance/0mujqbib",
   "tg_tokens": null,
   "tps": 19.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-oc06wbht-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 10.8 GB at this context; duplicate traces of the same session: azjdlpsd; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1438.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "gpt-oss-20b-oQ4 on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 1,438 65.7 10.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oc06wbht",
   "source_url": "https://omlx.ai/benchmarks/performance/oc06wbht",
   "tg_tokens": null,
   "tps": 65.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-oc06wbht-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 10.9 GB at this context; duplicate traces of the same session: azjdlpsd; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1490.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "gpt-oss-20b-oQ4 on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,490 62.3 10.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oc06wbht",
   "source_url": "https://omlx.ai/benchmarks/performance/oc06wbht",
   "tg_tokens": null,
   "tps": 62.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-goew35us-ctx199680",
   "model": "Flash-Next",
   "notes": "MTP speculative decoding enabled; name as published, likely Qwen3.8-Flash-Next shorthand; oMLX community benchmark trace, single run at 195k context; peak memory 119.4 GB; TTFT 146416 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1366.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Flash-Next-4bit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k tokens 1,366 PP tok/s 23.2 TG tok/s 146416 TTFT (ms) 119.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/goew35us",
   "source_url": "https://omlx.ai/benchmarks/performance/goew35us",
   "tg_tokens": null,
   "tps": 23.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q7v8l8py-ctx32768",
   "model": "Flash-Next",
   "notes": "MTP speculative decoding enabled; name as published, likely Qwen3.8-Flash-Next shorthand; oMLX community benchmark trace, single run at 32k context; peak memory 109.2 GB; TTFT 17876 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1833.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Flash-Next-4bit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k tokens 1,833 PP tok/s 64.5 TG tok/s 17876 TTFT (ms) 109.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q7v8l8py",
   "source_url": "https://omlx.ai/benchmarks/performance/q7v8l8py",
   "tg_tokens": null,
   "tps": 64.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-2tzufn9a-ctx8192",
   "model": "K2-Horizon-MoVA-36B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 39.5 GB at this context; duplicate traces of the same session: 121f6aqu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 831.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "K2-Horizon-MoVA-36B-A4B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 831.0 38.7 39.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2tzufn9a",
   "source_url": "https://omlx.ai/benchmarks/performance/2tzufn9a",
   "tg_tokens": null,
   "tps": 38.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-2tzufn9a-ctx32768",
   "model": "K2-Horizon-MoVA-36B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 43.9 GB at this context; duplicate traces of the same session: 121f6aqu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 470.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "K2-Horizon-MoVA-36B-A4B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 470.2 28.6 43.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2tzufn9a",
   "source_url": "https://omlx.ai/benchmarks/performance/2tzufn9a",
   "tg_tokens": null,
   "tps": 28.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lugb29q4-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: ttlcxr4i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 597.9,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 597.9 121.7 21.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lugb29q4",
   "source_url": "https://omlx.ai/benchmarks/performance/lugb29q4",
   "tg_tokens": null,
   "tps": 121.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lugb29q4-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: ttlcxr4i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1898.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,898 119.5 22.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lugb29q4",
   "source_url": "https://omlx.ai/benchmarks/performance/lugb29q4",
   "tg_tokens": null,
   "tps": 119.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-32fzunaa-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 20.5 GB; TTFT 213 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4800.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 4,800 PP tok/s 117.9 TG tok/s 213 TTFT (ms) 20.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/32fzunaa",
   "source_url": "https://omlx.ai/benchmarks/performance/32fzunaa",
   "tg_tokens": null,
   "tps": 117.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bqkssutz-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 20.1 GB; TTFT 406 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2522.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 2,522 PP tok/s 139.1 TG tok/s 406 TTFT (ms) 20.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bqkssutz",
   "source_url": "https://omlx.ai/benchmarks/performance/bqkssutz",
   "tg_tokens": null,
   "tps": 139.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mjkjue43-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 0z79v5ba, z6em9a6e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 921.1,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 921.1 32.2 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mjkjue43",
   "source_url": "https://omlx.ai/benchmarks/performance/mjkjue43",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mjkjue43-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 0z79v5ba, z6em9a6e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1012.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,012 31.9 17.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mjkjue43",
   "source_url": "https://omlx.ai/benchmarks/performance/mjkjue43",
   "tg_tokens": null,
   "tps": 31.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mjkjue43-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 0z79v5ba, z6em9a6e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1008.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,008 31.3 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mjkjue43",
   "source_url": "https://omlx.ai/benchmarks/performance/mjkjue43",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vrzcoltt-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 4qvrwwbd, r24y1mwu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 919.5,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 919.5 32.4 16.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vrzcoltt",
   "source_url": "https://omlx.ai/benchmarks/performance/vrzcoltt",
   "tg_tokens": null,
   "tps": 32.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vrzcoltt-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 4qvrwwbd, r24y1mwu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 978.2,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 978.2 31.7 17.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vrzcoltt",
   "source_url": "https://omlx.ai/benchmarks/performance/vrzcoltt",
   "tg_tokens": null,
   "tps": 31.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vrzcoltt-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 4qvrwwbd, r24y1mwu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 999.1,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 999.1 31.2 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vrzcoltt",
   "source_url": "https://omlx.ai/benchmarks/performance/vrzcoltt",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-z07cnpkw-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: tdlqq6oj, z8gqf1zw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 919.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 919.7 32.3 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z07cnpkw",
   "source_url": "https://omlx.ai/benchmarks/performance/z07cnpkw",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-z07cnpkw-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: tdlqq6oj, z8gqf1zw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1011.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,011 31.8 17.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z07cnpkw",
   "source_url": "https://omlx.ai/benchmarks/performance/z07cnpkw",
   "tg_tokens": null,
   "tps": 31.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-z07cnpkw-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: tdlqq6oj, z8gqf1zw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1004.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,004 30.9 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z07cnpkw",
   "source_url": "https://omlx.ai/benchmarks/performance/z07cnpkw",
   "tg_tokens": null,
   "tps": 30.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vco0qba3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: mg8fu2o9, t37oznxw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 858.8,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-27B-oQ6 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 858.8 24.1 22.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vco0qba3",
   "source_url": "https://omlx.ai/benchmarks/performance/vco0qba3",
   "tg_tokens": null,
   "tps": 24.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vco0qba3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: mg8fu2o9, t37oznxw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 950.6,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-27B-oQ6 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 950.6 23.9 23.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vco0qba3",
   "source_url": "https://omlx.ai/benchmarks/performance/vco0qba3",
   "tg_tokens": null,
   "tps": 23.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vco0qba3-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: mg8fu2o9, t37oznxw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 950.9,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-27B-oQ6 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 8k 950.9 23.5 24.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vco0qba3",
   "source_url": "https://omlx.ai/benchmarks/performance/vco0qba3",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ke3wnyr6-ctx1024",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: tzxhovg3, xelfgjbe; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 480.2,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 2bit; 1k 480.2 43.6 11.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ke3wnyr6",
   "source_url": "https://omlx.ai/benchmarks/performance/ke3wnyr6",
   "tg_tokens": null,
   "tps": 43.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ke3wnyr6-ctx4096",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 14.2 GB at this context; duplicate traces of the same session: tzxhovg3, xelfgjbe; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 645.2,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 2bit; 4k 645.2 41.3 14.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ke3wnyr6",
   "source_url": "https://omlx.ai/benchmarks/performance/ke3wnyr6",
   "tg_tokens": null,
   "tps": 41.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ke3wnyr6-ctx8192",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 14.7 GB at this context; duplicate traces of the same session: tzxhovg3, xelfgjbe; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 663.5,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 2bit; 8k 663.5 38.2 14.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ke3wnyr6",
   "source_url": "https://omlx.ai/benchmarks/performance/ke3wnyr6",
   "tg_tokens": null,
   "tps": 38.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-08",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jkdza3r8-ctx131072",
   "model": "Flash-Next",
   "notes": "MTP speculative decoding enabled; name as published, likely Qwen3.8-Flash-Next shorthand; oMLX community benchmark trace, single run at 128k context; peak memory 115.0 GB; TTFT 77267 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1696.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Flash-Next-4bit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k tokens 1,696 PP tok/s 38.4 TG tok/s 77267 TTFT (ms) 115.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jkdza3r8",
   "source_url": "https://omlx.ai/benchmarks/performance/jkdza3r8",
   "tg_tokens": null,
   "tps": 38.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-lsuzretv-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.2 GB at this context; duplicate traces of the same session: 1ebejaqr, 7akwwvwy, c32yojlx, jyf9vp8w, qw5zvp0n, wu9d9gws, z1csyx1s; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 331.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 331.0 51.3 32.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lsuzretv",
   "source_url": "https://omlx.ai/benchmarks/performance/lsuzretv",
   "tg_tokens": null,
   "tps": 51.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lsuzretv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.0 GB at this context; duplicate traces of the same session: 1ebejaqr, 7akwwvwy, c32yojlx, jyf9vp8w, qw5zvp0n, wu9d9gws, z1csyx1s; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 335.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 335.3 38.7 34.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lsuzretv",
   "source_url": "https://omlx.ai/benchmarks/performance/lsuzretv",
   "tg_tokens": null,
   "tps": 38.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lsuzretv-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 35.3 GB at this context; duplicate traces of the same session: 1ebejaqr, 7akwwvwy, c32yojlx, jyf9vp8w, qw5zvp0n, wu9d9gws, z1csyx1s; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 330.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 8k 330.3 39.4 35.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lsuzretv",
   "source_url": "https://omlx.ai/benchmarks/performance/lsuzretv",
   "tg_tokens": null,
   "tps": 39.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lsuzretv-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 36.0 GB at this context; duplicate traces of the same session: 1ebejaqr, 7akwwvwy, c32yojlx, jyf9vp8w, qw5zvp0n, wu9d9gws, z1csyx1s; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 315.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 16k 315.2 28.5 36.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lsuzretv",
   "source_url": "https://omlx.ai/benchmarks/performance/lsuzretv",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lsuzretv-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 38.5 GB at this context; duplicate traces of the same session: 1ebejaqr, 7akwwvwy, c32yojlx, jyf9vp8w, qw5zvp0n, wu9d9gws, z1csyx1s; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 290.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 32k 290.0 5.0 38.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lsuzretv",
   "source_url": "https://omlx.ai/benchmarks/performance/lsuzretv",
   "tg_tokens": null,
   "tps": 5.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lsuzretv-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 43.3 GB at this context; duplicate traces of the same session: 1ebejaqr, 7akwwvwy, c32yojlx, jyf9vp8w, qw5zvp0n, wu9d9gws, z1csyx1s; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 249.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 64k 249.5 2.6 43.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lsuzretv",
   "source_url": "https://omlx.ai/benchmarks/performance/lsuzretv",
   "tg_tokens": null,
   "tps": 2.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lsuzretv-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 51.6 GB at this context; duplicate traces of the same session: 1ebejaqr, 7akwwvwy, c32yojlx, jyf9vp8w, qw5zvp0n, wu9d9gws, z1csyx1s; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 194.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 128k 194.4 2.4 51.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lsuzretv",
   "source_url": "https://omlx.ai/benchmarks/performance/lsuzretv",
   "tg_tokens": null,
   "tps": 2.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lsuzretv-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 47.7 GB at this context; duplicate traces of the same session: 1ebejaqr, 7akwwvwy, c32yojlx, jyf9vp8w, qw5zvp0n, wu9d9gws, z1csyx1s; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 157.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 195k 157.8 0.8 47.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lsuzretv",
   "source_url": "https://omlx.ai/benchmarks/performance/lsuzretv",
   "tg_tokens": null,
   "tps": 0.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-08",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-yzk36erg-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 101.6 GB; TTFT 6796 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4822.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k tokens 4,822 PP tok/s 60.6 TG tok/s 6796 TTFT (ms) 101.6 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yzk36erg",
   "source_url": "https://omlx.ai/benchmarks/performance/yzk36erg",
   "tg_tokens": null,
   "tps": 60.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-aetbmwq3-ctx131072",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 47.4 GB; TTFT 2044191 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 64.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 128k tokens 64.1 PP tok/s 6.6 TG tok/s 2044191 TTFT (ms) 47.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aetbmwq3",
   "source_url": "https://omlx.ai/benchmarks/performance/aetbmwq3",
   "tg_tokens": null,
   "tps": 6.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2lnos4tp-ctx1024",
   "model": "Qwen3-4B-Instruct-2507",
   "notes": "oMLX community benchmark trace; peak memory 4.7 GB at this context; duplicate traces of the same session: 751etd6y; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4456.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-4B-Instruct-2507-MLX-8bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 8bit; 1k 4,456 83.2 4.7 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2lnos4tp",
   "source_url": "https://omlx.ai/benchmarks/performance/2lnos4tp",
   "tg_tokens": null,
   "tps": 83.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2lnos4tp-ctx4096",
   "model": "Qwen3-4B-Instruct-2507",
   "notes": "oMLX community benchmark trace; peak memory 5.2 GB at this context; duplicate traces of the same session: 751etd6y; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4451.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-4B-Instruct-2507-MLX-8bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 8bit; 4k 4,451 78.4 5.2 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2lnos4tp",
   "source_url": "https://omlx.ai/benchmarks/performance/2lnos4tp",
   "tg_tokens": null,
   "tps": 78.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-yrnfel5w-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 101.6 GB; TTFT 10529 ms; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 778.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 8k tokens 778.1 PP tok/s 71.6 TG tok/s 10529 TTFT (ms) 101.6 Peak mem",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yrnfel5w",
   "source_url": "https://omlx.ai/benchmarks/performance/yrnfel5w",
   "tg_tokens": null,
   "tps": 71.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-mtzo382r-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.8 GB at this context; duplicate traces of the same session: on9xglcp; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 778.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 8k 778.4 65.5 101.8 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mtzo382r",
   "source_url": "https://omlx.ai/benchmarks/performance/mtzo382r",
   "tg_tokens": null,
   "tps": 65.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-mtzo382r-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.7 GB at this context; duplicate traces of the same session: on9xglcp; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 772.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 32k 772.8 51.3 102.7 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mtzo382r",
   "source_url": "https://omlx.ai/benchmarks/performance/mtzo382r",
   "tg_tokens": null,
   "tps": 51.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-4reypxed-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: k49n1oqm; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 778.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 8k 778.8 39.6 101.3 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4reypxed",
   "source_url": "https://omlx.ai/benchmarks/performance/4reypxed",
   "tg_tokens": null,
   "tps": 39.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-4reypxed-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.2 GB at this context; duplicate traces of the same session: k49n1oqm; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 773.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 32k 773.2 37.8 102.2 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4reypxed",
   "source_url": "https://omlx.ai/benchmarks/performance/4reypxed",
   "tg_tokens": null,
   "tps": 37.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-iilye0jl-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 101.9 GB; TTFT 10537 ms; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 777.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 8k tokens 777.5 PP tok/s 47.7 TG tok/s 10537 TTFT (ms) 101.9 Peak mem",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iilye0jl",
   "source_url": "https://omlx.ai/benchmarks/performance/iilye0jl",
   "tg_tokens": null,
   "tps": 47.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-t0d3m4z7-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 101.3 GB; TTFT 3309 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1238.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k tokens 1,238 PP tok/s 91.7 TG tok/s 3309 TTFT (ms) 101.3 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t0d3m4z7",
   "source_url": "https://omlx.ai/benchmarks/performance/t0d3m4z7",
   "tg_tokens": null,
   "tps": 91.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ph57tus8-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: ni7ehgz6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 890.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 890.9 41.4 18.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ph57tus8",
   "source_url": "https://omlx.ai/benchmarks/performance/ph57tus8",
   "tg_tokens": null,
   "tps": 41.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ph57tus8-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: ni7ehgz6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 955.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 955.8 76.1 19.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ph57tus8",
   "source_url": "https://omlx.ai/benchmarks/performance/ph57tus8",
   "tg_tokens": null,
   "tps": 76.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-58dzsyfc-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: hh8nvo8d, o3t4izy8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 983.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 983.0 65.5 18.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/58dzsyfc",
   "source_url": "https://omlx.ai/benchmarks/performance/58dzsyfc",
   "tg_tokens": null,
   "tps": 65.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-58dzsyfc-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: hh8nvo8d, o3t4izy8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 944.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 944.0 78.0 19.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/58dzsyfc",
   "source_url": "https://omlx.ai/benchmarks/performance/58dzsyfc",
   "tg_tokens": null,
   "tps": 78.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-58dzsyfc-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: hh8nvo8d, o3t4izy8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 880.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 880.3 79.4 20.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/58dzsyfc",
   "source_url": "https://omlx.ai/benchmarks/performance/58dzsyfc",
   "tg_tokens": null,
   "tps": 79.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-5eahazuj-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: xdvlw0dc; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 104.3 15.2 16.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5eahazuj",
   "source_url": "https://omlx.ai/benchmarks/performance/5eahazuj",
   "tg_tokens": null,
   "tps": 15.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-5eahazuj-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: xdvlw0dc; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 116.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 116.2 17.4 19.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5eahazuj",
   "source_url": "https://omlx.ai/benchmarks/performance/5eahazuj",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-pvhp6tsq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 3f62cb3q; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 105.7 15.3 16.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pvhp6tsq",
   "source_url": "https://omlx.ai/benchmarks/performance/pvhp6tsq",
   "tg_tokens": null,
   "tps": 15.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-pvhp6tsq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 3f62cb3q; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 104.4 14.4 17.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pvhp6tsq",
   "source_url": "https://omlx.ai/benchmarks/performance/pvhp6tsq",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-svowkdap-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: 6oxb1uzc; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 290.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 290.8 9.7 15.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/svowkdap",
   "source_url": "https://omlx.ai/benchmarks/performance/svowkdap",
   "tg_tokens": null,
   "tps": 9.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-svowkdap-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 6oxb1uzc; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 313.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 313.2 9.6 17.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/svowkdap",
   "source_url": "https://omlx.ai/benchmarks/performance/svowkdap",
   "tg_tokens": null,
   "tps": 9.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-t12wyejr-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace, single run at 1k context; peak memory 34.6 GB; TTFT 1298 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 788.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k tokens 788.8 PP tok/s 50.8 TG tok/s 1298 TTFT (ms) 34.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t12wyejr",
   "source_url": "https://omlx.ai/benchmarks/performance/t12wyejr",
   "tg_tokens": null,
   "tps": 50.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hosng7vy-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.5 GB; TTFT 1072 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 955.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 955.5 PP tok/s 58.6 TG tok/s 1072 TTFT (ms) 70.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hosng7vy",
   "source_url": "https://omlx.ai/benchmarks/performance/hosng7vy",
   "tg_tokens": null,
   "tps": 58.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fd2ksvtb-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.5 GB; TTFT 921 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1112.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,112 PP tok/s 77.7 TG tok/s 921 TTFT (ms) 70.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fd2ksvtb",
   "source_url": "https://omlx.ai/benchmarks/performance/fd2ksvtb",
   "tg_tokens": null,
   "tps": 77.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-u0p4nomf-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.5 GB; TTFT 926 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1106.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,106 PP tok/s 76.0 TG tok/s 926 TTFT (ms) 70.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u0p4nomf",
   "source_url": "https://omlx.ai/benchmarks/performance/u0p4nomf",
   "tg_tokens": null,
   "tps": 76.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3ob6kuq9-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 214 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4788.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 4,788 PP tok/s 148.6 TG tok/s 214 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ob6kuq9",
   "source_url": "https://omlx.ai/benchmarks/performance/3ob6kuq9",
   "tg_tokens": null,
   "tps": 148.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-s3ipr8dk-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: lia4kcai, v8zp63tp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3401.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 3,401 93.7 6.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s3ipr8dk",
   "source_url": "https://omlx.ai/benchmarks/performance/s3ipr8dk",
   "tg_tokens": null,
   "tps": 93.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-s3ipr8dk-ctx32768",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.2 GB at this context; duplicate traces of the same session: lia4kcai, v8zp63tp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3167.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 3,167 86.9 7.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s3ipr8dk",
   "source_url": "https://omlx.ai/benchmarks/performance/s3ipr8dk",
   "tg_tokens": null,
   "tps": 86.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-s3ipr8dk-ctx65536",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.2 GB at this context; duplicate traces of the same session: lia4kcai, v8zp63tp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2760.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 2,760 73.7 8.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s3ipr8dk",
   "source_url": "https://omlx.ai/benchmarks/performance/s3ipr8dk",
   "tg_tokens": null,
   "tps": 73.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i4s6cvth-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: ak0857u2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4762.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 4,762 119.5 20.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i4s6cvth",
   "source_url": "https://omlx.ai/benchmarks/performance/i4s6cvth",
   "tg_tokens": null,
   "tps": 119.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i4s6cvth-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: ak0857u2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3726.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 3,726 101.2 21.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i4s6cvth",
   "source_url": "https://omlx.ai/benchmarks/performance/i4s6cvth",
   "tg_tokens": null,
   "tps": 101.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-y62dr33l-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4526.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,526 215.8 21.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y62dr33l",
   "source_url": "https://omlx.ai/benchmarks/performance/y62dr33l",
   "tg_tokens": null,
   "tps": 215.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-y62dr33l-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6357.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,357 173.6 22.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y62dr33l",
   "source_url": "https://omlx.ai/benchmarks/performance/y62dr33l",
   "tg_tokens": null,
   "tps": 173.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-z901nm66-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: irccjvnl; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 131.4 15.5 16.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z901nm66",
   "source_url": "https://omlx.ai/benchmarks/performance/z901nm66",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-z901nm66-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: irccjvnl; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 135.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 135.3 15.1 17.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z901nm66",
   "source_url": "https://omlx.ai/benchmarks/performance/z901nm66",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-70drwdxt-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 17.8 GB; TTFT 788 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1299.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k tokens 1,299 PP tok/s 146.3 TG tok/s 788 TTFT (ms) 17.8 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/70drwdxt",
   "source_url": "https://omlx.ai/benchmarks/performance/70drwdxt",
   "tg_tokens": null,
   "tps": 146.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ydm6u1cb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 18.0 GB; TTFT 784 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1306.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k tokens 1,306 PP tok/s 144.0 TG tok/s 784 TTFT (ms) 18.0 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ydm6u1cb",
   "source_url": "https://omlx.ai/benchmarks/performance/ydm6u1cb",
   "tg_tokens": null,
   "tps": 144.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gljodx9i-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 172.7 GB at this context; duplicate traces of the same session: ytt96zvx; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 253.9,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-27B-oQ8 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 253.9 4.7 172.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gljodx9i",
   "source_url": "https://omlx.ai/benchmarks/performance/gljodx9i",
   "tg_tokens": null,
   "tps": 4.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gljodx9i-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 177.6 GB at this context; duplicate traces of the same session: ytt96zvx; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 314.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-27B-oQ8 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 314.0 4.0 177.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gljodx9i",
   "source_url": "https://omlx.ai/benchmarks/performance/gljodx9i",
   "tg_tokens": null,
   "tps": 4.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kh7gk11s-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: 4xeu2k0d, 6jgwsp7o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1992.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,992 49.3 71.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kh7gk11s",
   "source_url": "https://omlx.ai/benchmarks/performance/kh7gk11s",
   "tg_tokens": null,
   "tps": 49.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kh7gk11s-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 4xeu2k0d, 6jgwsp7o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2589.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,589 68.9 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kh7gk11s",
   "source_url": "https://omlx.ai/benchmarks/performance/kh7gk11s",
   "tg_tokens": null,
   "tps": 68.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kh7gk11s-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.9 GB at this context; duplicate traces of the same session: 4xeu2k0d, 6jgwsp7o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2938.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,938 72.5 72.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kh7gk11s",
   "source_url": "https://omlx.ai/benchmarks/performance/kh7gk11s",
   "tg_tokens": null,
   "tps": 72.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ljo3h4dv-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 27.3 GB at this context; duplicate traces of the same session: 07avm17p, 220pbrs6, 3fl6r2oe, tiewd6ty, wpghuzof, x7l89is3; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 6bit; 1k 103.9 12.6 27.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ljo3h4dv",
   "source_url": "https://omlx.ai/benchmarks/performance/ljo3h4dv",
   "tg_tokens": null,
   "tps": 12.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-ljo3h4dv-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 31.7 GB at this context; duplicate traces of the same session: 07avm17p, 220pbrs6, 3fl6r2oe, tiewd6ty, wpghuzof, x7l89is3; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 114.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 6bit; 4k 114.6 11.4 31.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ljo3h4dv",
   "source_url": "https://omlx.ai/benchmarks/performance/ljo3h4dv",
   "tg_tokens": null,
   "tps": 11.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ljo3h4dv-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 32.9 GB at this context; duplicate traces of the same session: 07avm17p, 220pbrs6, 3fl6r2oe, tiewd6ty, wpghuzof, x7l89is3; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 112.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 6bit; 8k 112.4 11.6 32.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ljo3h4dv",
   "source_url": "https://omlx.ai/benchmarks/performance/ljo3h4dv",
   "tg_tokens": null,
   "tps": 11.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ljo3h4dv-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 33.5 GB at this context; duplicate traces of the same session: 07avm17p, 220pbrs6, 3fl6r2oe, tiewd6ty, wpghuzof, x7l89is3; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 107.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 6bit; 16k 107.3 12.1 33.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ljo3h4dv",
   "source_url": "https://omlx.ai/benchmarks/performance/ljo3h4dv",
   "tg_tokens": null,
   "tps": 12.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ljo3h4dv-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 34.6 GB at this context; duplicate traces of the same session: 07avm17p, 220pbrs6, 3fl6r2oe, tiewd6ty, wpghuzof, x7l89is3; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 98.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 6bit; 32k 98.8 10.6 34.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ljo3h4dv",
   "source_url": "https://omlx.ai/benchmarks/performance/ljo3h4dv",
   "tg_tokens": null,
   "tps": 10.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ljo3h4dv-ctx65536",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 36.8 GB at this context; duplicate traces of the same session: 07avm17p, 220pbrs6, 3fl6r2oe, tiewd6ty, wpghuzof, x7l89is3; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 85.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 6bit; 64k 85.9 8.4 36.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ljo3h4dv",
   "source_url": "https://omlx.ai/benchmarks/performance/ljo3h4dv",
   "tg_tokens": null,
   "tps": 8.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ljo3h4dv-ctx131072",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 41.3 GB at this context; duplicate traces of the same session: 07avm17p, 220pbrs6, 3fl6r2oe, tiewd6ty, wpghuzof, x7l89is3; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 66.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 6bit; 128k 66.3 5.9 41.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ljo3h4dv",
   "source_url": "https://omlx.ai/benchmarks/performance/ljo3h4dv",
   "tg_tokens": null,
   "tps": 5.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-x7i9w5aw-ctx1024",
   "model": "6-bit",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: gybkkcf8; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 139.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "6-bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 6bit; 1k 139.7 14.5 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x7i9w5aw",
   "source_url": "https://omlx.ai/benchmarks/performance/x7i9w5aw",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-x7i9w5aw-ctx4096",
   "model": "6-bit",
   "notes": "oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: gybkkcf8; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 129.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "6-bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 6bit; 4k 129.0 14.2 25.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x7i9w5aw",
   "source_url": "https://omlx.ai/benchmarks/performance/x7i9w5aw",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4df3tveb-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4526.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,526 215.8 21.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4df3tveb",
   "source_url": "https://omlx.ai/benchmarks/performance/4df3tveb",
   "tg_tokens": null,
   "tps": 215.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4df3tveb-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6357.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,357 173.6 22.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4df3tveb",
   "source_url": "https://omlx.ai/benchmarks/performance/4df3tveb",
   "tg_tokens": null,
   "tps": 173.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-bdnmciwl-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: 6q9cr12x, b8pncp15; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 299.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 5bit; 1k 299.9 35.1 20.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bdnmciwl",
   "source_url": "https://omlx.ai/benchmarks/performance/bdnmciwl",
   "tg_tokens": null,
   "tps": 35.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-bdnmciwl-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 6q9cr12x, b8pncp15; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 202.6,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 5bit; 4k 202.6 32.0 21.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bdnmciwl",
   "source_url": "https://omlx.ai/benchmarks/performance/bdnmciwl",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-bdnmciwl-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 6q9cr12x, b8pncp15; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 158.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 5bit; 8k 158.1 33.4 21.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bdnmciwl",
   "source_url": "https://omlx.ai/benchmarks/performance/bdnmciwl",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-sh3tlo6o-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: emx47h0e; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1329.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k 1,329 55.0 16.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sh3tlo6o",
   "source_url": "https://omlx.ai/benchmarks/performance/sh3tlo6o",
   "tg_tokens": null,
   "tps": 55.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-sh3tlo6o-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: emx47h0e; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1521.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 4k 1,521 53.0 17.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sh3tlo6o",
   "source_url": "https://omlx.ai/benchmarks/performance/sh3tlo6o",
   "tg_tokens": null,
   "tps": 53.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-goz016jk-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: 21603d26, j0vkhhug; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 394.9,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-27B-oQ3e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 3bit; 1k 394.9 20.5 14.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/goz016jk",
   "source_url": "https://omlx.ai/benchmarks/performance/goz016jk",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-goz016jk-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: 21603d26, j0vkhhug; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 411.6,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-27B-oQ3e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 3bit; 4k 411.6 20.9 16.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/goz016jk",
   "source_url": "https://omlx.ai/benchmarks/performance/goz016jk",
   "tg_tokens": null,
   "tps": 20.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-goz016jk-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 21603d26, j0vkhhug; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 339.0,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-27B-oQ3e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 3bit; 8k 339.0 17.5 17.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/goz016jk",
   "source_url": "https://omlx.ai/benchmarks/performance/goz016jk",
   "tg_tokens": null,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-2dv5st2s-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: p4r7k0si; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 137.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 1k 137.9 28.8 18.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2dv5st2s",
   "source_url": "https://omlx.ai/benchmarks/performance/2dv5st2s",
   "tg_tokens": null,
   "tps": 28.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-2dv5st2s-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: p4r7k0si; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 129.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k 129.2 27.4 22.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2dv5st2s",
   "source_url": "https://omlx.ai/benchmarks/performance/2dv5st2s",
   "tg_tokens": null,
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-jxm293x8-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 8xetyeub; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 104.9 14.7 16.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jxm293x8",
   "source_url": "https://omlx.ai/benchmarks/performance/jxm293x8",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-jxm293x8-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 8xetyeub; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 104.7 14.0 17.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jxm293x8",
   "source_url": "https://omlx.ai/benchmarks/performance/jxm293x8",
   "tg_tokens": null,
   "tps": 14.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-pybo3w97-ctx1024",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: 58csgv4l; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 888.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-MLX-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 1k 888.4 28.0 32.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pybo3w97",
   "source_url": "https://omlx.ai/benchmarks/performance/pybo3w97",
   "tg_tokens": null,
   "tps": 28.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-pybo3w97-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 33.1 GB at this context; duplicate traces of the same session: 58csgv4l; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 973.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-MLX-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 4k 973.8 27.5 33.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pybo3w97",
   "source_url": "https://omlx.ai/benchmarks/performance/pybo3w97",
   "tg_tokens": null,
   "tps": 27.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-dmfdrji0-ctx1024",
   "model": "Ling-3.0-flash",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 68.3 GB at this context; duplicate traces of the same session: 7xrsn2v6, ivkufr6d, khrxnhm0, w7c0xosm; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 555.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-flash-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 555.0 38.8 68.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dmfdrji0",
   "source_url": "https://omlx.ai/benchmarks/performance/dmfdrji0",
   "tg_tokens": null,
   "tps": 38.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-dmfdrji0-ctx4096",
   "model": "Ling-3.0-flash",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 69.5 GB at this context; duplicate traces of the same session: 7xrsn2v6, ivkufr6d, khrxnhm0, w7c0xosm; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 490.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-flash-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 490.6 37.7 69.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dmfdrji0",
   "source_url": "https://omlx.ai/benchmarks/performance/dmfdrji0",
   "tg_tokens": null,
   "tps": 37.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-dmfdrji0-ctx8192",
   "model": "Ling-3.0-flash",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 70.2 GB at this context; duplicate traces of the same session: 7xrsn2v6, ivkufr6d, khrxnhm0, w7c0xosm; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 491.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-flash-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 491.0 34.3 70.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dmfdrji0",
   "source_url": "https://omlx.ai/benchmarks/performance/dmfdrji0",
   "tg_tokens": null,
   "tps": 34.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-dmfdrji0-ctx16384",
   "model": "Ling-3.0-flash",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: 7xrsn2v6, ivkufr6d, khrxnhm0, w7c0xosm; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 462.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-flash-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 462.4 34.8 71.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dmfdrji0",
   "source_url": "https://omlx.ai/benchmarks/performance/dmfdrji0",
   "tg_tokens": null,
   "tps": 34.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-dmfdrji0-ctx32768",
   "model": "Ling-3.0-flash",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 74.7 GB at this context; duplicate traces of the same session: 7xrsn2v6, ivkufr6d, khrxnhm0, w7c0xosm; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 414.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-flash-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 32k 414.2 30.9 74.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dmfdrji0",
   "source_url": "https://omlx.ai/benchmarks/performance/dmfdrji0",
   "tg_tokens": null,
   "tps": 30.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fe2vauir-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: a82suxij; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2826.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,826 98.8 20.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fe2vauir",
   "source_url": "https://omlx.ai/benchmarks/performance/fe2vauir",
   "tg_tokens": null,
   "tps": 98.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fe2vauir-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: a82suxij; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2826.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 2,826 89.5 20.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fe2vauir",
   "source_url": "https://omlx.ai/benchmarks/performance/fe2vauir",
   "tg_tokens": null,
   "tps": 89.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-w9gi901v-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: 2zl9h23a, 3imvfdq0, 54ol8o11, h7x0e2wg, p5kkhxt9, q222xysp, vq3jcst2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2097.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 2,097 59.8 35.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w9gi901v",
   "source_url": "https://omlx.ai/benchmarks/performance/w9gi901v",
   "tg_tokens": null,
   "tps": 59.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-w9gi901v-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: 2zl9h23a, 3imvfdq0, 54ol8o11, h7x0e2wg, p5kkhxt9, q222xysp, vq3jcst2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2842.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 2,842 60.9 36.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w9gi901v",
   "source_url": "https://omlx.ai/benchmarks/performance/w9gi901v",
   "tg_tokens": null,
   "tps": 60.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-w9gi901v-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 36.2 GB at this context; duplicate traces of the same session: 2zl9h23a, 3imvfdq0, 54ol8o11, h7x0e2wg, p5kkhxt9, q222xysp, vq3jcst2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2802.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 2,802 58.2 36.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w9gi901v",
   "source_url": "https://omlx.ai/benchmarks/performance/w9gi901v",
   "tg_tokens": null,
   "tps": 58.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-w9gi901v-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 36.4 GB at this context; duplicate traces of the same session: 2zl9h23a, 3imvfdq0, 54ol8o11, h7x0e2wg, p5kkhxt9, q222xysp, vq3jcst2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2595.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 16k 2,595 56.2 36.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w9gi901v",
   "source_url": "https://omlx.ai/benchmarks/performance/w9gi901v",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-w9gi901v-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 36.7 GB at this context; duplicate traces of the same session: 2zl9h23a, 3imvfdq0, 54ol8o11, h7x0e2wg, p5kkhxt9, q222xysp, vq3jcst2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2220.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 32k 2,220 50.9 36.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w9gi901v",
   "source_url": "https://omlx.ai/benchmarks/performance/w9gi901v",
   "tg_tokens": null,
   "tps": 50.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-w9gi901v-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: 2zl9h23a, 3imvfdq0, 54ol8o11, h7x0e2wg, p5kkhxt9, q222xysp, vq3jcst2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1698.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 64k 1,698 42.8 37.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w9gi901v",
   "source_url": "https://omlx.ai/benchmarks/performance/w9gi901v",
   "tg_tokens": null,
   "tps": 42.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-w9gi901v-ctx131072",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 38.6 GB at this context; duplicate traces of the same session: 2zl9h23a, 3imvfdq0, 54ol8o11, h7x0e2wg, p5kkhxt9, q222xysp, vq3jcst2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1147.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 128k 1,147 33.3 38.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w9gi901v",
   "source_url": "https://omlx.ai/benchmarks/performance/w9gi901v",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-w9gi901v-ctx199680",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 39.9 GB at this context; duplicate traces of the same session: 2zl9h23a, 3imvfdq0, 54ol8o11, h7x0e2wg, p5kkhxt9, q222xysp, vq3jcst2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 850.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 195k 850.8 27.2 39.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w9gi901v",
   "source_url": "https://omlx.ai/benchmarks/performance/w9gi901v",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-nfon63wp-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: xatnh3ek; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 726.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 726.4 60.8 35.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nfon63wp",
   "source_url": "https://omlx.ai/benchmarks/performance/nfon63wp",
   "tg_tokens": null,
   "tps": 60.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-nfon63wp-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: xatnh3ek; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1812.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 1,812 60.1 36.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nfon63wp",
   "source_url": "https://omlx.ai/benchmarks/performance/nfon63wp",
   "tg_tokens": null,
   "tps": 60.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-f9tsds17-ctx4096",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 4rxj49fz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 356.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 356.1 38.2 19.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f9tsds17",
   "source_url": "https://omlx.ai/benchmarks/performance/f9tsds17",
   "tg_tokens": null,
   "tps": 38.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-f9tsds17-ctx16384",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: 4rxj49fz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 337.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 16k 337.1 35.4 20.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f9tsds17",
   "source_url": "https://omlx.ai/benchmarks/performance/f9tsds17",
   "tg_tokens": null,
   "tps": 35.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ge99ts31-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 6fmescb9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4433.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,433 204.1 21.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ge99ts31",
   "source_url": "https://omlx.ai/benchmarks/performance/ge99ts31",
   "tg_tokens": null,
   "tps": 204.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ge99ts31-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 6fmescb9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6403.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,403 182.6 22.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ge99ts31",
   "source_url": "https://omlx.ai/benchmarks/performance/ge99ts31",
   "tg_tokens": null,
   "tps": 182.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-fl49yz5v-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: otkitxnl, ww2ltbzl; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 76.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 76.7 14.2 20.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fl49yz5v",
   "source_url": "https://omlx.ai/benchmarks/performance/fl49yz5v",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-fl49yz5v-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: otkitxnl, ww2ltbzl; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 70.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k 70.0 13.0 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fl49yz5v",
   "source_url": "https://omlx.ai/benchmarks/performance/fl49yz5v",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-fl49yz5v-ctx65536",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: otkitxnl, ww2ltbzl; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 53.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 64k 53.4 9.7 25.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fl49yz5v",
   "source_url": "https://omlx.ai/benchmarks/performance/fl49yz5v",
   "tg_tokens": null,
   "tps": 9.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qifughkk-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: q622q9ze; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1142.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,142 74.0 17.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qifughkk",
   "source_url": "https://omlx.ai/benchmarks/performance/qifughkk",
   "tg_tokens": null,
   "tps": 74.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qifughkk-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: q622q9ze; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1385.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,385 62.4 19.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qifughkk",
   "source_url": "https://omlx.ai/benchmarks/performance/qifughkk",
   "tg_tokens": null,
   "tps": 62.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-qvme1taw-ctx8192",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: hclwyx8p; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 697.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 697.2 64.2 22.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qvme1taw",
   "source_url": "https://omlx.ai/benchmarks/performance/qvme1taw",
   "tg_tokens": null,
   "tps": 64.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-qvme1taw-ctx16384",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: hclwyx8p; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 598.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 598.0 59.6 23.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qvme1taw",
   "source_url": "https://omlx.ai/benchmarks/performance/qvme1taw",
   "tg_tokens": null,
   "tps": 59.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-hivgnh3q-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 22.1 GB; TTFT 5645 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 725.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k tokens 725.6 PP tok/s 67.8 TG tok/s 5645 TTFT (ms) 22.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hivgnh3q",
   "source_url": "https://omlx.ai/benchmarks/performance/hivgnh3q",
   "tg_tokens": null,
   "tps": 67.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-lqbzg6e7-ctx32768",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 24.4 GB; TTFT 70630 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 463.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 32k tokens 463.9 PP tok/s 53.8 TG tok/s 70630 TTFT (ms) 24.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lqbzg6e7",
   "source_url": "https://omlx.ai/benchmarks/performance/lqbzg6e7",
   "tg_tokens": null,
   "tps": 53.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0gxk39jo-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "True2456 community build; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: vg0hzef7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 285.3,
   "provenance": "community",
   "quant": "AWQ 5.0bpw",
   "quote": "True2456--Qwen3.8-27B-AWQ-5.0bpw on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 285.3 5.9 32.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0gxk39jo",
   "source_url": "https://omlx.ai/benchmarks/performance/0gxk39jo",
   "tg_tokens": null,
   "tps": 5.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0gxk39jo-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "True2456 community build; oMLX community benchmark trace; peak memory 34.6 GB at this context; duplicate traces of the same session: vg0hzef7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 275.3,
   "provenance": "community",
   "quant": "AWQ 5.0bpw",
   "quote": "True2456--Qwen3.8-27B-AWQ-5.0bpw on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 4k 275.3 5.8 34.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0gxk39jo",
   "source_url": "https://omlx.ai/benchmarks/performance/0gxk39jo",
   "tg_tokens": null,
   "tps": 5.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-52xo25p2-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: mq6027ki; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 101.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 101.1 23.1 18.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/52xo25p2",
   "source_url": "https://omlx.ai/benchmarks/performance/52xo25p2",
   "tg_tokens": null,
   "tps": 23.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-52xo25p2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: mq6027ki; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 102.7 23.3 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/52xo25p2",
   "source_url": "https://omlx.ai/benchmarks/performance/52xo25p2",
   "tg_tokens": null,
   "tps": 23.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-czywjqdj-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 17.8 GB; TTFT 788 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1300.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k tokens 1,300 PP tok/s 127.7 TG tok/s 788 TTFT (ms) 17.8 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/czywjqdj",
   "source_url": "https://omlx.ai/benchmarks/performance/czywjqdj",
   "tg_tokens": null,
   "tps": 127.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-n3kza7rn-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "MTP + PLE; oMLX community benchmark trace; peak memory 43.3 GB at this context; duplicate traces of the same session: 64wxnq2e, atbolv6b; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 341.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-REAP-288-oQ4e-BF16-MTP-PLE on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 341.9 35.3 43.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n3kza7rn",
   "source_url": "https://omlx.ai/benchmarks/performance/n3kza7rn",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-n3kza7rn-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "MTP + PLE; oMLX community benchmark trace; peak memory 43.3 GB at this context; duplicate traces of the same session: 64wxnq2e, atbolv6b; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 331.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-REAP-288-oQ4e-BF16-MTP-PLE on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 331.1 34.4 43.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n3kza7rn",
   "source_url": "https://omlx.ai/benchmarks/performance/n3kza7rn",
   "tg_tokens": null,
   "tps": 34.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-n3kza7rn-ctx8192",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "MTP + PLE; oMLX community benchmark trace; peak memory 43.5 GB at this context; duplicate traces of the same session: 64wxnq2e, atbolv6b; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 341.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-REAP-288-oQ4e-BF16-MTP-PLE on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 341.9 35.2 43.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n3kza7rn",
   "source_url": "https://omlx.ai/benchmarks/performance/n3kza7rn",
   "tg_tokens": null,
   "tps": 35.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-29teqrat-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 71.2 GB at this context; duplicate traces of the same session: bml7drc6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 959.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 959.0 60.7 71.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/29teqrat",
   "source_url": "https://omlx.ai/benchmarks/performance/29teqrat",
   "tg_tokens": null,
   "tps": 60.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-29teqrat-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 71.2 GB at this context; duplicate traces of the same session: bml7drc6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1905.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,905 57.3 71.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/29teqrat",
   "source_url": "https://omlx.ai/benchmarks/performance/29teqrat",
   "tg_tokens": null,
   "tps": 57.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-7l8p9q0s-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 167.5 GB; TTFT 568525 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 230.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 128k tokens 230.5 PP tok/s 16.5 TG tok/s 568525 TTFT (ms) 167.5 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7l8p9q0s",
   "source_url": "https://omlx.ai/benchmarks/performance/7l8p9q0s",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-cbk38h02-ctx32768",
   "model": "Ling-3.0-tiny",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 8.0 GB; TTFT 48893 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 670.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-tiny-oQ4e on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 4bit; 32k tokens 670.2 PP tok/s 47.1 TG tok/s 48893 TTFT (ms) 8.0 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cbk38h02",
   "source_url": "https://omlx.ai/benchmarks/performance/cbk38h02",
   "tg_tokens": null,
   "tps": 47.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-4m3mm51t-ctx32768",
   "model": "Ling-3.0-tiny",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 8.0 GB; TTFT 47445 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 690.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-tiny-oQ4e on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 4bit; 32k tokens 690.7 PP tok/s 48.1 TG tok/s 47445 TTFT (ms) 8.0 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4m3mm51t",
   "source_url": "https://omlx.ai/benchmarks/performance/4m3mm51t",
   "tg_tokens": null,
   "tps": 48.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-pb6op6ej-ctx4096",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.7 GB at this context; duplicate traces of the same session: pimis7uj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 256.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-27B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 256.6 16.3 31.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pb6op6ej",
   "source_url": "https://omlx.ai/benchmarks/performance/pb6op6ej",
   "tg_tokens": null,
   "tps": 16.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-pb6op6ej-ctx16384",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.0 GB at this context; duplicate traces of the same session: pimis7uj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 244.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-27B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 244.9 15.7 33.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pb6op6ej",
   "source_url": "https://omlx.ai/benchmarks/performance/pb6op6ej",
   "tg_tokens": null,
   "tps": 15.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-e2hqwa70-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: z4gl2v0x; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 126.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 126.2 15.5 17.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e2hqwa70",
   "source_url": "https://omlx.ai/benchmarks/performance/e2hqwa70",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-e2hqwa70-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: z4gl2v0x; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 120.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 120.7 14.5 19.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e2hqwa70",
   "source_url": "https://omlx.ai/benchmarks/performance/e2hqwa70",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-fl8pcwkz-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 4kxntvjb, 62s6i7zf, 7hqieh0n, guyw0sv9, hlhb4ots, v1ddzhwl, zcm0707o; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 430.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 430.3 12.8 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fl8pcwkz",
   "source_url": "https://omlx.ai/benchmarks/performance/fl8pcwkz",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-fl8pcwkz-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 4kxntvjb, 62s6i7zf, 7hqieh0n, guyw0sv9, hlhb4ots, v1ddzhwl, zcm0707o; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 459.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 459.4 12.5 21.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fl8pcwkz",
   "source_url": "https://omlx.ai/benchmarks/performance/fl8pcwkz",
   "tg_tokens": null,
   "tps": 12.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fl8pcwkz-ctx8192",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 4kxntvjb, 62s6i7zf, 7hqieh0n, guyw0sv9, hlhb4ots, v1ddzhwl, zcm0707o; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 449.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 449.0 12.4 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fl8pcwkz",
   "source_url": "https://omlx.ai/benchmarks/performance/fl8pcwkz",
   "tg_tokens": null,
   "tps": 12.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-fl8pcwkz-ctx16384",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 4kxntvjb, 62s6i7zf, 7hqieh0n, guyw0sv9, hlhb4ots, v1ddzhwl, zcm0707o; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 433.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 433.7 12.0 21.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fl8pcwkz",
   "source_url": "https://omlx.ai/benchmarks/performance/fl8pcwkz",
   "tg_tokens": null,
   "tps": 12.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-fl8pcwkz-ctx32768",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: 4kxntvjb, 62s6i7zf, 7hqieh0n, guyw0sv9, hlhb4ots, v1ddzhwl, zcm0707o; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 414.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 414.6 11.3 22.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fl8pcwkz",
   "source_url": "https://omlx.ai/benchmarks/performance/fl8pcwkz",
   "tg_tokens": null,
   "tps": 11.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-fl8pcwkz-ctx65536",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: 4kxntvjb, 62s6i7zf, 7hqieh0n, guyw0sv9, hlhb4ots, v1ddzhwl, zcm0707o; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 363.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 363.0 10.1 24.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fl8pcwkz",
   "source_url": "https://omlx.ai/benchmarks/performance/fl8pcwkz",
   "tg_tokens": null,
   "tps": 10.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-fl8pcwkz-ctx131072",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 28.9 GB at this context; duplicate traces of the same session: 4kxntvjb, 62s6i7zf, 7hqieh0n, guyw0sv9, hlhb4ots, v1ddzhwl, zcm0707o; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 288.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 128k 288.5 8.6 28.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fl8pcwkz",
   "source_url": "https://omlx.ai/benchmarks/performance/fl8pcwkz",
   "tg_tokens": null,
   "tps": 8.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fl8pcwkz-ctx199680",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 33.1 GB at this context; duplicate traces of the same session: 4kxntvjb, 62s6i7zf, 7hqieh0n, guyw0sv9, hlhb4ots, v1ddzhwl, zcm0707o; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 236.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 195k 236.6 7.4 33.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fl8pcwkz",
   "source_url": "https://omlx.ai/benchmarks/performance/fl8pcwkz",
   "tg_tokens": null,
   "tps": 7.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-pifj24qw-ctx1024",
   "model": "Qwen3.8-27B-OBLITERATED",
   "notes": "oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: 4ea1l4xw; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 138.2,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-OBLITERATED-OptiQ-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 1k 138.2 16.2 22.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pifj24qw",
   "source_url": "https://omlx.ai/benchmarks/performance/pifj24qw",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-pifj24qw-ctx4096",
   "model": "Qwen3.8-27B-OBLITERATED",
   "notes": "oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: 4ea1l4xw; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 133.1,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-OBLITERATED-OptiQ-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k 133.1 15.9 26.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pifj24qw",
   "source_url": "https://omlx.ai/benchmarks/performance/pifj24qw",
   "tg_tokens": null,
   "tps": 15.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-m7hsqkgd-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: m0y1yocg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 670.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 670.3 70.0 100.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m7hsqkgd",
   "source_url": "https://omlx.ai/benchmarks/performance/m7hsqkgd",
   "tg_tokens": null,
   "tps": 70.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-m7hsqkgd-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: m0y1yocg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 816.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 816.8 58.5 101.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m7hsqkgd",
   "source_url": "https://omlx.ai/benchmarks/performance/m7hsqkgd",
   "tg_tokens": null,
   "tps": 58.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-25ivzmtc-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 185.0 GB; TTFT 343081 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 382.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 128k tokens 382.0 PP tok/s 38.4 TG tok/s 343081 TTFT (ms) 185.0 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/25ivzmtc",
   "source_url": "https://omlx.ai/benchmarks/performance/25ivzmtc",
   "tg_tokens": null,
   "tps": 38.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xdf1tftx-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 175.8 GB; TTFT 518271 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 252.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 128k tokens 252.9 PP tok/s 22.0 TG tok/s 518271 TTFT (ms) 175.8 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xdf1tftx",
   "source_url": "https://omlx.ai/benchmarks/performance/xdf1tftx",
   "tg_tokens": null,
   "tps": 22.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-0utqiyhs-ctx131072",
   "model": "Muse Glimmer",
   "notes": "30B build; oMLX community benchmark trace, single run at 128k context; peak memory 21.0 GB; TTFT 436367 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 300.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Muse-Glimmer-30B-oQ5e on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 5bit; 128k tokens 300.4 PP tok/s 10.7 TG tok/s 436367 TTFT (ms) 21.0 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0utqiyhs",
   "source_url": "https://omlx.ai/benchmarks/performance/0utqiyhs",
   "tg_tokens": null,
   "tps": 10.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-7wpa5hx0-ctx1024",
   "model": "Qwen3.6-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: icqtv0hs; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 165.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-27B-MTP-4bit-MLX on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 1k 165.8 32.8 16.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7wpa5hx0",
   "source_url": "https://omlx.ai/benchmarks/performance/7wpa5hx0",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-7wpa5hx0-ctx4096",
   "model": "Qwen3.6-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: icqtv0hs; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 173.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-27B-MTP-4bit-MLX on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 4k 173.8 25.4 19.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7wpa5hx0",
   "source_url": "https://omlx.ai/benchmarks/performance/7wpa5hx0",
   "tg_tokens": null,
   "tps": 25.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xa2tg3l0-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: ejphvux9, n6o2ib6r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5399.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 5,399 43.4 21.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xa2tg3l0",
   "source_url": "https://omlx.ai/benchmarks/performance/xa2tg3l0",
   "tg_tokens": null,
   "tps": 43.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xa2tg3l0-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: ejphvux9, n6o2ib6r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4763.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 4,763 31.5 22.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xa2tg3l0",
   "source_url": "https://omlx.ai/benchmarks/performance/xa2tg3l0",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xa2tg3l0-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: ejphvux9, n6o2ib6r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3613.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 3,613 19.4 24.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xa2tg3l0",
   "source_url": "https://omlx.ai/benchmarks/performance/xa2tg3l0",
   "tg_tokens": null,
   "tps": 19.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-lmk8x4jy-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: z6i14d5i; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 488.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 488.9 14.3 21.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lmk8x4jy",
   "source_url": "https://omlx.ai/benchmarks/performance/lmk8x4jy",
   "tg_tokens": null,
   "tps": 14.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-lmk8x4jy-ctx16384",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: z6i14d5i; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 413.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 413.1 13.7 21.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lmk8x4jy",
   "source_url": "https://omlx.ai/benchmarks/performance/lmk8x4jy",
   "tg_tokens": null,
   "tps": 13.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-zjzr6xy0-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 920c1lx0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 127.9 33.3 20.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zjzr6xy0",
   "source_url": "https://omlx.ai/benchmarks/performance/zjzr6xy0",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-zjzr6xy0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: 920c1lx0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 190.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 190.1 25.2 24.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zjzr6xy0",
   "source_url": "https://omlx.ai/benchmarks/performance/zjzr6xy0",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-d27e55u7-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: weno7k8j; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 127.4 27.4 20.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d27e55u7",
   "source_url": "https://omlx.ai/benchmarks/performance/d27e55u7",
   "tg_tokens": null,
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-d27e55u7-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: weno7k8j; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 190.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 190.7 30.3 24.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d27e55u7",
   "source_url": "https://omlx.ai/benchmarks/performance/d27e55u7",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-jeosqzxr-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: ir3qgbmq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 127.8 27.6 20.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jeosqzxr",
   "source_url": "https://omlx.ai/benchmarks/performance/jeosqzxr",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-jeosqzxr-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: ir3qgbmq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 188.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 188.3 27.1 24.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jeosqzxr",
   "source_url": "https://omlx.ai/benchmarks/performance/jeosqzxr",
   "tg_tokens": null,
   "tps": 27.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-8iyt57sx-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 136.9 GB at this context; duplicate traces of the same session: 7gr1dfsf, nukawi11; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 651.7,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ8-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 651.7 33.2 136.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8iyt57sx",
   "source_url": "https://omlx.ai/benchmarks/performance/8iyt57sx",
   "tg_tokens": null,
   "tps": 33.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-8iyt57sx-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 140.3 GB at this context; duplicate traces of the same session: 7gr1dfsf, nukawi11; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1312.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ8-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 64k 1,312 43.2 140.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8iyt57sx",
   "source_url": "https://omlx.ai/benchmarks/performance/8iyt57sx",
   "tg_tokens": null,
   "tps": 43.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-8iyt57sx-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 145.2 GB at this context; duplicate traces of the same session: 7gr1dfsf, nukawi11; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1296.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ8-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 128k 1,296 51.5 145.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8iyt57sx",
   "source_url": "https://omlx.ai/benchmarks/performance/8iyt57sx",
   "tg_tokens": null,
   "tps": 51.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qdc392kl-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 37.5 GB; TTFT 9272 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 110.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k tokens 110.4 PP tok/s 11.8 TG tok/s 9272 TTFT (ms) 37.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qdc392kl",
   "source_url": "https://omlx.ai/benchmarks/performance/qdc392kl",
   "tg_tokens": null,
   "tps": 11.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-mkfern27-ctx1024",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 5.2 GB at this context; duplicate traces of the same session: svzot7mb; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1560.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-E2B-it-MLX-8bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 1k 1,560 78.9 5.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mkfern27",
   "source_url": "https://omlx.ai/benchmarks/performance/mkfern27",
   "tg_tokens": null,
   "tps": 78.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-mkfern27-ctx4096",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 5.3 GB at this context; duplicate traces of the same session: svzot7mb; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2661.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-E2B-it-MLX-8bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 4k 2,661 76.6 5.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mkfern27",
   "source_url": "https://omlx.ai/benchmarks/performance/mkfern27",
   "tg_tokens": null,
   "tps": 76.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4ztiy0tw-ctx1024",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 87.8 GB at this context; duplicate traces of the same session: 2m0de93y, kn3z1xn2, wdny0myu, xdn540n2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 399.8,
   "provenance": "community",
   "quant": "2.4bit mixed",
   "quote": "DeepSeek-V4-Flash-0731-2.4bit-mixed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 1k 399.8 40.0 87.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ztiy0tw",
   "source_url": "https://omlx.ai/benchmarks/performance/4ztiy0tw",
   "tg_tokens": null,
   "tps": 40.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4ztiy0tw-ctx4096",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 88.3 GB at this context; duplicate traces of the same session: 2m0de93y, kn3z1xn2, wdny0myu, xdn540n2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 605.8,
   "provenance": "community",
   "quant": "2.4bit mixed",
   "quote": "DeepSeek-V4-Flash-0731-2.4bit-mixed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 4k 605.8 40.0 88.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ztiy0tw",
   "source_url": "https://omlx.ai/benchmarks/performance/4ztiy0tw",
   "tg_tokens": null,
   "tps": 40.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4ztiy0tw-ctx8192",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 88.3 GB at this context; duplicate traces of the same session: 2m0de93y, kn3z1xn2, wdny0myu, xdn540n2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 613.5,
   "provenance": "community",
   "quant": "2.4bit mixed",
   "quote": "DeepSeek-V4-Flash-0731-2.4bit-mixed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 8k 613.5 38.5 88.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ztiy0tw",
   "source_url": "https://omlx.ai/benchmarks/performance/4ztiy0tw",
   "tg_tokens": null,
   "tps": 38.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4ztiy0tw-ctx16384",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 88.4 GB at this context; duplicate traces of the same session: 2m0de93y, kn3z1xn2, wdny0myu, xdn540n2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 548.8,
   "provenance": "community",
   "quant": "2.4bit mixed",
   "quote": "DeepSeek-V4-Flash-0731-2.4bit-mixed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 16k 548.8 42.8 88.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ztiy0tw",
   "source_url": "https://omlx.ai/benchmarks/performance/4ztiy0tw",
   "tg_tokens": null,
   "tps": 42.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4ztiy0tw-ctx32768",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 88.5 GB at this context; duplicate traces of the same session: 2m0de93y, kn3z1xn2, wdny0myu, xdn540n2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 509.4,
   "provenance": "community",
   "quant": "2.4bit mixed",
   "quote": "DeepSeek-V4-Flash-0731-2.4bit-mixed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 32k 509.4 44.2 88.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ztiy0tw",
   "source_url": "https://omlx.ai/benchmarks/performance/4ztiy0tw",
   "tg_tokens": null,
   "tps": 44.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-k2exszwe-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 288.7 GB at this context; duplicate traces of the same session: 3pl8tssk, c7i4j13m, l1oaheg3, w7tu0whq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 339.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 339.7 20.9 288.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k2exszwe",
   "source_url": "https://omlx.ai/benchmarks/performance/k2exszwe",
   "tg_tokens": null,
   "tps": 20.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-k2exszwe-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 289.3 GB at this context; duplicate traces of the same session: 3pl8tssk, c7i4j13m, l1oaheg3, w7tu0whq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 617.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 16k 617.7 35.2 289.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k2exszwe",
   "source_url": "https://omlx.ai/benchmarks/performance/k2exszwe",
   "tg_tokens": null,
   "tps": 35.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-k2exszwe-ctx32768",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 289.3 GB at this context; duplicate traces of the same session: 3pl8tssk, c7i4j13m, l1oaheg3, w7tu0whq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 690.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 32k 690.6 33.4 289.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k2exszwe",
   "source_url": "https://omlx.ai/benchmarks/performance/k2exszwe",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-k2exszwe-ctx65536",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 289.3 GB at this context; duplicate traces of the same session: 3pl8tssk, c7i4j13m, l1oaheg3, w7tu0whq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 687.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 64k 687.4 32.0 289.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k2exszwe",
   "source_url": "https://omlx.ai/benchmarks/performance/k2exszwe",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-k2exszwe-ctx131072",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 289.4 GB at this context; duplicate traces of the same session: 3pl8tssk, c7i4j13m, l1oaheg3, w7tu0whq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 665.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 128k 665.8 29.6 289.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k2exszwe",
   "source_url": "https://omlx.ai/benchmarks/performance/k2exszwe",
   "tg_tokens": null,
   "tps": 29.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-63imi2n6-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 0u906mlw, gtvjq10i, gyzvi3kg, po5whl5b, qobn0tet, zfipukby; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6312.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 4bit; 1k 6,312 150.0 21.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/63imi2n6",
   "source_url": "https://omlx.ai/benchmarks/performance/63imi2n6",
   "tg_tokens": null,
   "tps": 150.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-63imi2n6-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 0u906mlw, gtvjq10i, gyzvi3kg, po5whl5b, qobn0tet, zfipukby; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 9793.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 4bit; 4k 9,793 155.4 22.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/63imi2n6",
   "source_url": "https://omlx.ai/benchmarks/performance/63imi2n6",
   "tg_tokens": null,
   "tps": 155.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-63imi2n6-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 0u906mlw, gtvjq10i, gyzvi3kg, po5whl5b, qobn0tet, zfipukby; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 9095.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 4bit; 8k 9,095 161.6 22.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/63imi2n6",
   "source_url": "https://omlx.ai/benchmarks/performance/63imi2n6",
   "tg_tokens": null,
   "tps": 161.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-63imi2n6-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 0u906mlw, gtvjq10i, gyzvi3kg, po5whl5b, qobn0tet, zfipukby; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 8826.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 4bit; 16k 8,826 173.6 22.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/63imi2n6",
   "source_url": "https://omlx.ai/benchmarks/performance/63imi2n6",
   "tg_tokens": null,
   "tps": 173.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-63imi2n6-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 0u906mlw, gtvjq10i, gyzvi3kg, po5whl5b, qobn0tet, zfipukby; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 7405.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 4bit; 32k 7,405 178.5 22.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/63imi2n6",
   "source_url": "https://omlx.ai/benchmarks/performance/63imi2n6",
   "tg_tokens": null,
   "tps": 178.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-63imi2n6-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 0u906mlw, gtvjq10i, gyzvi3kg, po5whl5b, qobn0tet, zfipukby; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5739.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 4bit; 64k 5,739 129.1 23.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/63imi2n6",
   "source_url": "https://omlx.ai/benchmarks/performance/63imi2n6",
   "tg_tokens": null,
   "tps": 129.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-63imi2n6-ctx131072",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: 0u906mlw, gtvjq10i, gyzvi3kg, po5whl5b, qobn0tet, zfipukby; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3864.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 4bit; 128k 3,864 114.6 24.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/63imi2n6",
   "source_url": "https://omlx.ai/benchmarks/performance/63imi2n6",
   "tg_tokens": null,
   "tps": 114.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-63plvntt-ctx1024",
   "model": "Qwen3.6-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: qum4654n; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 166.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-27B-MTP-4bit-MLX on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 1k 166.9 36.8 16.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/63plvntt",
   "source_url": "https://omlx.ai/benchmarks/performance/63plvntt",
   "tg_tokens": null,
   "tps": 36.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-63plvntt-ctx4096",
   "model": "Qwen3.6-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: qum4654n; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 174.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-27B-MTP-4bit-MLX on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 4k 174.2 27.4 19.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/63plvntt",
   "source_url": "https://omlx.ai/benchmarks/performance/63plvntt",
   "tg_tokens": null,
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2gs4da90-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: h32ekezo, haod2xcv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 960.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 960.8 43.7 21.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2gs4da90",
   "source_url": "https://omlx.ai/benchmarks/performance/2gs4da90",
   "tg_tokens": null,
   "tps": 43.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2gs4da90-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: h32ekezo, haod2xcv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 890.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 890.9 9.3 24.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2gs4da90",
   "source_url": "https://omlx.ai/benchmarks/performance/2gs4da90",
   "tg_tokens": null,
   "tps": 9.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2gs4da90-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.4 GB at this context; duplicate traces of the same session: h32ekezo, haod2xcv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 775.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 775.9 4.1 29.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2gs4da90",
   "source_url": "https://omlx.ai/benchmarks/performance/2gs4da90",
   "tg_tokens": null,
   "tps": 4.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-h7o53tqs-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: e7bp045a; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 512.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 512.6 17.8 17.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h7o53tqs",
   "source_url": "https://omlx.ai/benchmarks/performance/h7o53tqs",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-h7o53tqs-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: e7bp045a; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 453.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 453.7 16.7 17.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h7o53tqs",
   "source_url": "https://omlx.ai/benchmarks/performance/h7o53tqs",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-a7wsvg5f-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: tmbnrey1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 506.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 506.3 17.8 17.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a7wsvg5f",
   "source_url": "https://omlx.ai/benchmarks/performance/a7wsvg5f",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-a7wsvg5f-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: tmbnrey1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 449.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 449.3 16.6 17.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a7wsvg5f",
   "source_url": "https://omlx.ai/benchmarks/performance/a7wsvg5f",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-49r4qlhl-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: mnq7drt6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 127.6 35.9 20.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/49r4qlhl",
   "source_url": "https://omlx.ai/benchmarks/performance/49r4qlhl",
   "tg_tokens": null,
   "tps": 35.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-49r4qlhl-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: mnq7drt6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 190.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 190.9 29.3 24.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/49r4qlhl",
   "source_url": "https://omlx.ai/benchmarks/performance/49r4qlhl",
   "tg_tokens": null,
   "tps": 29.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-9nqudftt-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 9t2li85b; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 127.5 25.7 20.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9nqudftt",
   "source_url": "https://omlx.ai/benchmarks/performance/9nqudftt",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-9nqudftt-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: 9t2li85b; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 181.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 181.8 28.5 24.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9nqudftt",
   "source_url": "https://omlx.ai/benchmarks/performance/9nqudftt",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-pqgx4bqi-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 7dvuusja; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 124.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 124.6 33.6 20.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pqgx4bqi",
   "source_url": "https://omlx.ai/benchmarks/performance/pqgx4bqi",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-pqgx4bqi-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: 7dvuusja; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 177.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 177.2 27.8 24.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pqgx4bqi",
   "source_url": "https://omlx.ai/benchmarks/performance/pqgx4bqi",
   "tg_tokens": null,
   "tps": 27.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-vcak9ybp-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: ydj82cnr; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 127.9 30.8 20.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vcak9ybp",
   "source_url": "https://omlx.ai/benchmarks/performance/vcak9ybp",
   "tg_tokens": null,
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-vcak9ybp-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: ydj82cnr; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 187.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 187.4 27.4 24.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vcak9ybp",
   "source_url": "https://omlx.ai/benchmarks/performance/vcak9ybp",
   "tg_tokens": null,
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-rbk9rsr1-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: c0bbgqou; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 127.9 35.0 20.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rbk9rsr1",
   "source_url": "https://omlx.ai/benchmarks/performance/rbk9rsr1",
   "tg_tokens": null,
   "tps": 35.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-rbk9rsr1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: c0bbgqou; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 184.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 184.3 28.2 24.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rbk9rsr1",
   "source_url": "https://omlx.ai/benchmarks/performance/rbk9rsr1",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-rcinab8h-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: qwn53yzp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 64.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 64.2 20.7 20.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rcinab8h",
   "source_url": "https://omlx.ai/benchmarks/performance/rcinab8h",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-rcinab8h-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: qwn53yzp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 163.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 163.2 28.2 24.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rcinab8h",
   "source_url": "https://omlx.ai/benchmarks/performance/rcinab8h",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-pskwlcw2-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: pyqgfhpy; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 130.6 14.6 19.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pskwlcw2",
   "source_url": "https://omlx.ai/benchmarks/performance/pskwlcw2",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-pskwlcw2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: pyqgfhpy; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 188.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 188.9 14.1 22.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pskwlcw2",
   "source_url": "https://omlx.ai/benchmarks/performance/pskwlcw2",
   "tg_tokens": null,
   "tps": 14.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-tbv6iqc3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: iwwo24wl; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 127.8 29.5 20.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tbv6iqc3",
   "source_url": "https://omlx.ai/benchmarks/performance/tbv6iqc3",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-tbv6iqc3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: iwwo24wl; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 186.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 186.9 27.3 24.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tbv6iqc3",
   "source_url": "https://omlx.ai/benchmarks/performance/tbv6iqc3",
   "tg_tokens": null,
   "tps": 27.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-sy3o0smp-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.6 GB at this context; duplicate traces of the same session: 2sxb23o4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 247.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-fp16-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 247.2 37.9 34.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sy3o0smp",
   "source_url": "https://omlx.ai/benchmarks/performance/sy3o0smp",
   "tg_tokens": null,
   "tps": 37.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-sy3o0smp-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.1 GB at this context; duplicate traces of the same session: 2sxb23o4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 252.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-fp16-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 252.2 27.9 38.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sy3o0smp",
   "source_url": "https://omlx.ai/benchmarks/performance/sy3o0smp",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-x16nqkh0-ctx1024",
   "model": "Ornith-1.5-35B-A3B-BigBang",
   "notes": "MTP speculative decoding enabled; zh build; oMLX community benchmark trace, single run at 1k context; peak memory 20.3 GB; TTFT 1398 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 732.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-BigBang-MTP-zh-mlx-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k tokens 732.4 PP tok/s 102.1 TG tok/s 1398 TTFT (ms) 20.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x16nqkh0",
   "source_url": "https://omlx.ai/benchmarks/performance/x16nqkh0",
   "tg_tokens": null,
   "tps": 102.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wf99c8bs-ctx1024",
   "model": "Qwen3.6-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: 1pzbzcyr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 897.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-27B-MTP-4bit-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 897.0 57.4 16.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wf99c8bs",
   "source_url": "https://omlx.ai/benchmarks/performance/wf99c8bs",
   "tg_tokens": null,
   "tps": 57.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wf99c8bs-ctx4096",
   "model": "Qwen3.6-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 1pzbzcyr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 990.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-27B-MTP-4bit-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 990.1 52.9 17.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wf99c8bs",
   "source_url": "https://omlx.ai/benchmarks/performance/wf99c8bs",
   "tg_tokens": null,
   "tps": 52.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-gxmc1f2d-ctx1024",
   "model": "Qwen3.6-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: qls5u6lb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 902.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-27B-MTP-4bit-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 902.4 66.2 16.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gxmc1f2d",
   "source_url": "https://omlx.ai/benchmarks/performance/gxmc1f2d",
   "tg_tokens": null,
   "tps": 66.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-gxmc1f2d-ctx4096",
   "model": "Qwen3.6-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: qls5u6lb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 993.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-27B-MTP-4bit-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 993.7 66.4 17.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gxmc1f2d",
   "source_url": "https://omlx.ai/benchmarks/performance/gxmc1f2d",
   "tg_tokens": null,
   "tps": 66.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-c5kgxrsp-ctx1024",
   "model": "Qwen3.6-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: wa29c98l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 831.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-27B-MTP-4bit-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 831.1 58.3 16.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c5kgxrsp",
   "source_url": "https://omlx.ai/benchmarks/performance/c5kgxrsp",
   "tg_tokens": null,
   "tps": 58.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-c5kgxrsp-ctx4096",
   "model": "Qwen3.6-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: wa29c98l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 942.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-27B-MTP-4bit-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 942.4 55.9 17.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c5kgxrsp",
   "source_url": "https://omlx.ai/benchmarks/performance/c5kgxrsp",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-dg872y0v-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: ugv73rv2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 123.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 123.4 29.5 20.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dg872y0v",
   "source_url": "https://omlx.ai/benchmarks/performance/dg872y0v",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-dg872y0v-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: ugv73rv2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 178.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 178.9 25.5 24.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dg872y0v",
   "source_url": "https://omlx.ai/benchmarks/performance/dg872y0v",
   "tg_tokens": null,
   "tps": 25.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ew6xflxb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: nq5w3oq5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 127.7 30.7 20.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ew6xflxb",
   "source_url": "https://omlx.ai/benchmarks/performance/ew6xflxb",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ew6xflxb-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: nq5w3oq5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 185.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 185.0 31.8 24.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ew6xflxb",
   "source_url": "https://omlx.ai/benchmarks/performance/ew6xflxb",
   "tg_tokens": null,
   "tps": 31.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-7375v1el-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: mmgzvjoj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 126.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 126.8 35.3 20.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7375v1el",
   "source_url": "https://omlx.ai/benchmarks/performance/7375v1el",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-7375v1el-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: mmgzvjoj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 184.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 184.2 27.2 24.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7375v1el",
   "source_url": "https://omlx.ai/benchmarks/performance/7375v1el",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-pxlrhi3m-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: mhxpqnzu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 127.7 34.4 20.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pxlrhi3m",
   "source_url": "https://omlx.ai/benchmarks/performance/pxlrhi3m",
   "tg_tokens": null,
   "tps": 34.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-pxlrhi3m-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: mhxpqnzu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 184.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 184.5 31.3 24.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pxlrhi3m",
   "source_url": "https://omlx.ai/benchmarks/performance/pxlrhi3m",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-hqjn8ek7-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: pgxt0ddw; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 127.1 28.5 20.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hqjn8ek7",
   "source_url": "https://omlx.ai/benchmarks/performance/hqjn8ek7",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-hqjn8ek7-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: pgxt0ddw; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 179.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 179.9 26.0 24.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hqjn8ek7",
   "source_url": "https://omlx.ai/benchmarks/performance/hqjn8ek7",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-jiwhwlhr-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: cm0u4wuu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 127.6 33.6 20.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jiwhwlhr",
   "source_url": "https://omlx.ai/benchmarks/performance/jiwhwlhr",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-jiwhwlhr-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: cm0u4wuu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 181.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 181.9 25.2 24.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jiwhwlhr",
   "source_url": "https://omlx.ai/benchmarks/performance/jiwhwlhr",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ti8l55y0-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: nazyil83; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 127.9 35.7 20.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ti8l55y0",
   "source_url": "https://omlx.ai/benchmarks/performance/ti8l55y0",
   "tg_tokens": null,
   "tps": 35.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ti8l55y0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: nazyil83; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 184.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 184.6 30.2 24.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ti8l55y0",
   "source_url": "https://omlx.ai/benchmarks/performance/ti8l55y0",
   "tg_tokens": null,
   "tps": 30.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-nlmw820v-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: 0bx0bolf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 820.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 820.9 62.4 26.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nlmw820v",
   "source_url": "https://omlx.ai/benchmarks/performance/nlmw820v",
   "tg_tokens": null,
   "tps": 62.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nlmw820v-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: 0bx0bolf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 903.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 903.8 52.8 27.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nlmw820v",
   "source_url": "https://omlx.ai/benchmarks/performance/nlmw820v",
   "tg_tokens": null,
   "tps": 52.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4yx6fdr7-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: s148bk9c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 717.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 717.4 41.4 26.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4yx6fdr7",
   "source_url": "https://omlx.ai/benchmarks/performance/4yx6fdr7",
   "tg_tokens": null,
   "tps": 41.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4yx6fdr7-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: s148bk9c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 898.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 898.6 43.9 27.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4yx6fdr7",
   "source_url": "https://omlx.ai/benchmarks/performance/4yx6fdr7",
   "tg_tokens": null,
   "tps": 43.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-yazhgsjt-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.3 GB at this context; duplicate traces of the same session: 3eqg9ox3, x9q878u6, y363ltxm; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1350.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 6bit; 8k 1,350 106.6 28.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yazhgsjt",
   "source_url": "https://omlx.ai/benchmarks/performance/yazhgsjt",
   "tg_tokens": null,
   "tps": 106.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-yazhgsjt-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: 3eqg9ox3, x9q878u6, y363ltxm; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1245.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 6bit; 32k 1,245 98.0 29.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yazhgsjt",
   "source_url": "https://omlx.ai/benchmarks/performance/yazhgsjt",
   "tg_tokens": null,
   "tps": 98.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-yazhgsjt-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: 3eqg9ox3, x9q878u6, y363ltxm; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1088.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 6bit; 64k 1,088 78.4 32.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yazhgsjt",
   "source_url": "https://omlx.ai/benchmarks/performance/yazhgsjt",
   "tg_tokens": null,
   "tps": 78.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-yazhgsjt-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.8 GB at this context; duplicate traces of the same session: 3eqg9ox3, x9q878u6, y363ltxm; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 868.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 6bit; 128k 868.7 71.6 36.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yazhgsjt",
   "source_url": "https://omlx.ai/benchmarks/performance/yazhgsjt",
   "tg_tokens": null,
   "tps": 71.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-hr49q86z-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 40.5 GB; TTFT 7741 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 529.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 529.1 PP tok/s 46.6 TG tok/s 7741 TTFT (ms) 40.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hr49q86z",
   "source_url": "https://omlx.ai/benchmarks/performance/hr49q86z",
   "tg_tokens": null,
   "tps": 46.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0tv6tifh-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 71.6 GB at this context; duplicate traces of the same session: hvtxfk4r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 190.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 190.1 9.2 71.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0tv6tifh",
   "source_url": "https://omlx.ai/benchmarks/performance/0tv6tifh",
   "tg_tokens": null,
   "tps": 9.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0tv6tifh-ctx16384",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 73.0 GB at this context; duplicate traces of the same session: hvtxfk4r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 466.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 466.6 15.8 73.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0tv6tifh",
   "source_url": "https://omlx.ai/benchmarks/performance/0tv6tifh",
   "tg_tokens": null,
   "tps": 15.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-j78hyam2-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: jo86rks9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1217.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,217 88.9 70.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j78hyam2",
   "source_url": "https://omlx.ai/benchmarks/performance/j78hyam2",
   "tg_tokens": null,
   "tps": 88.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-j78hyam2-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: jo86rks9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2079.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,079 65.4 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j78hyam2",
   "source_url": "https://omlx.ai/benchmarks/performance/j78hyam2",
   "tg_tokens": null,
   "tps": 65.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tw0uz98l-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.6 GB; TTFT 2832 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 361.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 361.6 PP tok/s 56.3 TG tok/s 2832 TTFT (ms) 70.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tw0uz98l",
   "source_url": "https://omlx.ai/benchmarks/performance/tw0uz98l",
   "tg_tokens": null,
   "tps": 56.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-r5i03wg0-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.2 GB at this context; duplicate traces of the same session: m0fx1xvi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 356.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 356.3 65.0 69.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r5i03wg0",
   "source_url": "https://omlx.ai/benchmarks/performance/r5i03wg0",
   "tg_tokens": null,
   "tps": 65.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-r5i03wg0-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.4 GB at this context; duplicate traces of the same session: m0fx1xvi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1087.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,087 52.7 69.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r5i03wg0",
   "source_url": "https://omlx.ai/benchmarks/performance/r5i03wg0",
   "tg_tokens": null,
   "tps": 52.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0k222mu6-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 108.4 GB; TTFT 1317 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 777.7,
   "provenance": "community",
   "quant": "2bit",
   "quote": "GLM-5.3-Flash-MLX-2bit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 1k tokens 777.7 PP tok/s 38.7 TG tok/s 1317 TTFT (ms) 108.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0k222mu6",
   "source_url": "https://omlx.ai/benchmarks/performance/0k222mu6",
   "tg_tokens": null,
   "tps": 38.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-a5n2j6dp-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 108.4 GB; TTFT 2011 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 509.3,
   "provenance": "community",
   "quant": "2bit",
   "quote": "GLM-5.3-Flash-MLX-2bit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 1k tokens 509.3 PP tok/s 37.9 TG tok/s 2011 TTFT (ms) 108.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a5n2j6dp",
   "source_url": "https://omlx.ai/benchmarks/performance/a5n2j6dp",
   "tg_tokens": null,
   "tps": 37.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-vc2ygl22-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 108.4 GB; TTFT 1315 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 778.8,
   "provenance": "community",
   "quant": "2bit",
   "quote": "GLM-5.3-Flash-MLX-2bit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 1k tokens 778.8 PP tok/s 38.6 TG tok/s 1315 TTFT (ms) 108.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vc2ygl22",
   "source_url": "https://omlx.ai/benchmarks/performance/vc2ygl22",
   "tg_tokens": null,
   "tps": 38.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wh6aekmb-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: maano3nl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3858.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 3,858 131.1 37.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wh6aekmb",
   "source_url": "https://omlx.ai/benchmarks/performance/wh6aekmb",
   "tg_tokens": null,
   "tps": 131.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wh6aekmb-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.0 GB at this context; duplicate traces of the same session: maano3nl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5659.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 5,659 130.9 38.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wh6aekmb",
   "source_url": "https://omlx.ai/benchmarks/performance/wh6aekmb",
   "tg_tokens": null,
   "tps": 130.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-b3j0bo09-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 987t0kal, mfacbhqs; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 690.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 690.8 66.8 19.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b3j0bo09",
   "source_url": "https://omlx.ai/benchmarks/performance/b3j0bo09",
   "tg_tokens": null,
   "tps": 66.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-b3j0bo09-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 987t0kal, mfacbhqs; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1719.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 8k 1,719 64.2 19.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b3j0bo09",
   "source_url": "https://omlx.ai/benchmarks/performance/b3j0bo09",
   "tg_tokens": null,
   "tps": 64.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-b3j0bo09-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 987t0kal, mfacbhqs; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1870.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 1,870 59.2 20.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b3j0bo09",
   "source_url": "https://omlx.ai/benchmarks/performance/b3j0bo09",
   "tg_tokens": null,
   "tps": 59.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-iytqnyef-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "DWQ build; oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: b5dq383l, i2mu56vh; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1532.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit-DWQ on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 8bit; 1k 1,532 47.2 20.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iytqnyef",
   "source_url": "https://omlx.ai/benchmarks/performance/iytqnyef",
   "tg_tokens": null,
   "tps": 47.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-iytqnyef-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "DWQ build; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: b5dq383l, i2mu56vh; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2010.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit-DWQ on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 8bit; 8k 2,010 47.0 21.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iytqnyef",
   "source_url": "https://omlx.ai/benchmarks/performance/iytqnyef",
   "tg_tokens": null,
   "tps": 47.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-iytqnyef-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "DWQ build; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: b5dq383l, i2mu56vh; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1833.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit-DWQ on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 8bit; 16k 1,833 42.8 21.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iytqnyef",
   "source_url": "https://omlx.ai/benchmarks/performance/iytqnyef",
   "tg_tokens": null,
   "tps": 42.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-28u1xgbj-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: 2hg7yub1, tjb6ezph; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 280.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 280.0 9.3 15.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/28u1xgbj",
   "source_url": "https://omlx.ai/benchmarks/performance/28u1xgbj",
   "tg_tokens": null,
   "tps": 9.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-28u1xgbj-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 2hg7yub1, tjb6ezph; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 278.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 8k 278.4 9.0 17.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/28u1xgbj",
   "source_url": "https://omlx.ai/benchmarks/performance/28u1xgbj",
   "tg_tokens": null,
   "tps": 9.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-28u1xgbj-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 2hg7yub1, tjb6ezph; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 269.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 269.4 8.7 17.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/28u1xgbj",
   "source_url": "https://omlx.ai/benchmarks/performance/28u1xgbj",
   "tg_tokens": null,
   "tps": 8.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-y4z5hj8x-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.3 GB at this context; duplicate traces of the same session: ot53t827; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 249.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 249.8 16.2 28.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y4z5hj8x",
   "source_url": "https://omlx.ai/benchmarks/performance/y4z5hj8x",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-y4z5hj8x-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.7 GB at this context; duplicate traces of the same session: ot53t827; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 253.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 253.3 16.0 31.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y4z5hj8x",
   "source_url": "https://omlx.ai/benchmarks/performance/y4z5hj8x",
   "tg_tokens": null,
   "tps": 16.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-1sxbhpgu-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.3 GB at this context; duplicate traces of the same session: 7xmcufjy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 249.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 249.5 16.1 28.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1sxbhpgu",
   "source_url": "https://omlx.ai/benchmarks/performance/1sxbhpgu",
   "tg_tokens": null,
   "tps": 16.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-1sxbhpgu-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.7 GB at this context; duplicate traces of the same session: 7xmcufjy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 253.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 253.3 16.0 31.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1sxbhpgu",
   "source_url": "https://omlx.ai/benchmarks/performance/1sxbhpgu",
   "tg_tokens": null,
   "tps": 16.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-s1zyw14f-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 23.8 GB; TTFT 72672 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 225.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k tokens 225.5 PP tok/s 39.7 TG tok/s 72672 TTFT (ms) 23.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s1zyw14f",
   "source_url": "https://omlx.ai/benchmarks/performance/s1zyw14f",
   "tg_tokens": null,
   "tps": 39.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-eg3wmfn3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 0ycb1oa6, evx8qgmo, t21bq8yu, vec4j3z8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 127.7 30.8 20.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eg3wmfn3",
   "source_url": "https://omlx.ai/benchmarks/performance/eg3wmfn3",
   "tg_tokens": null,
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-eg3wmfn3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: 0ycb1oa6, evx8qgmo, t21bq8yu, vec4j3z8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 184.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 184.7 26.5 24.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eg3wmfn3",
   "source_url": "https://omlx.ai/benchmarks/performance/eg3wmfn3",
   "tg_tokens": null,
   "tps": 26.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-eg3wmfn3-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: 0ycb1oa6, evx8qgmo, t21bq8yu, vec4j3z8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 175.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 175.2 38.0 25.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eg3wmfn3",
   "source_url": "https://omlx.ai/benchmarks/performance/eg3wmfn3",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-eg3wmfn3-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: 0ycb1oa6, evx8qgmo, t21bq8yu, vec4j3z8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 166.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 166.9 37.7 26.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eg3wmfn3",
   "source_url": "https://omlx.ai/benchmarks/performance/eg3wmfn3",
   "tg_tokens": null,
   "tps": 37.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-eg3wmfn3-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: 0ycb1oa6, evx8qgmo, t21bq8yu, vec4j3z8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 152.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 152.1 36.9 28.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eg3wmfn3",
   "source_url": "https://omlx.ai/benchmarks/performance/eg3wmfn3",
   "tg_tokens": null,
   "tps": 36.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-1m2m6p9u-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: dhka9c6z, n7948co4; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 270.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 270.6 22.5 18.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1m2m6p9u",
   "source_url": "https://omlx.ai/benchmarks/performance/1m2m6p9u",
   "tg_tokens": null,
   "tps": 22.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-1m2m6p9u-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: dhka9c6z, n7948co4; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 210.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 8k 210.9 22.2 19.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1m2m6p9u",
   "source_url": "https://omlx.ai/benchmarks/performance/1m2m6p9u",
   "tg_tokens": null,
   "tps": 22.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-1m2m6p9u-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: dhka9c6z, n7948co4; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 207.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 207.0 24.2 19.5 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1m2m6p9u",
   "source_url": "https://omlx.ai/benchmarks/performance/1m2m6p9u",
   "tg_tokens": null,
   "tps": 24.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lumypd2k-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: dfdldcnb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1217.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,217 70.0 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lumypd2k",
   "source_url": "https://omlx.ai/benchmarks/performance/lumypd2k",
   "tg_tokens": null,
   "tps": 70.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lumypd2k-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: dfdldcnb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2077.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,077 83.0 70.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lumypd2k",
   "source_url": "https://omlx.ai/benchmarks/performance/lumypd2k",
   "tg_tokens": null,
   "tps": 83.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gfbn3nwq-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.6 GB; TTFT 839 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1221.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,221 PP tok/s 71.2 TG tok/s 839 TTFT (ms) 70.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gfbn3nwq",
   "source_url": "https://omlx.ai/benchmarks/performance/gfbn3nwq",
   "tg_tokens": null,
   "tps": 71.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-grm9xh2o-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: gc8jukue; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1219.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,219 68.4 70.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/grm9xh2o",
   "source_url": "https://omlx.ai/benchmarks/performance/grm9xh2o",
   "tg_tokens": null,
   "tps": 68.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-grm9xh2o-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: gc8jukue; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2066.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,066 67.9 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/grm9xh2o",
   "source_url": "https://omlx.ai/benchmarks/performance/grm9xh2o",
   "tg_tokens": null,
   "tps": 67.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9d4sxp0d-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.6 GB; TTFT 839 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1220.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,220 PP tok/s 86.6 TG tok/s 839 TTFT (ms) 70.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9d4sxp0d",
   "source_url": "https://omlx.ai/benchmarks/performance/9d4sxp0d",
   "tg_tokens": null,
   "tps": 86.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-s1n3e5n1-ctx1024",
   "model": "AX-Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: 21s6958m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 912.4,
   "provenance": "community",
   "quant": "AXQ 6bit",
   "quote": "AX-Qwen3.6-35B-A3B-MLX-AXQ-6bit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 912.4 106.2 25.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s1n3e5n1",
   "source_url": "https://omlx.ai/benchmarks/performance/s1n3e5n1",
   "tg_tokens": null,
   "tps": 106.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-s1n3e5n1-ctx4096",
   "model": "AX-Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: 21s6958m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2748.0,
   "provenance": "community",
   "quant": "AXQ 6bit",
   "quote": "AX-Qwen3.6-35B-A3B-MLX-AXQ-6bit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,748 108.7 26.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s1n3e5n1",
   "source_url": "https://omlx.ai/benchmarks/performance/s1n3e5n1",
   "tg_tokens": null,
   "tps": 108.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-t94nugwj-ctx1024",
   "model": "Huihui-Qwen3.8-27B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: ijw9hf46; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 792.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Huihui-Qwen3.8-27B-abliterated-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 792.0 49.8 33.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t94nugwj",
   "source_url": "https://omlx.ai/benchmarks/performance/t94nugwj",
   "tg_tokens": null,
   "tps": 49.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-t94nugwj-ctx4096",
   "model": "Huihui-Qwen3.8-27B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: ijw9hf46; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 847.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Huihui-Qwen3.8-27B-abliterated-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 847.8 37.0 34.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t94nugwj",
   "source_url": "https://omlx.ai/benchmarks/performance/t94nugwj",
   "tg_tokens": null,
   "tps": 37.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-4uib8pbe-ctx1024",
   "model": "Huihui-gemma-4-26B-A4B-it-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: oita7f7m; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 642.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Huihui-gemma-4-26B-A4B-it-abliterated-oQ8e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 1k 642.2 57.3 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4uib8pbe",
   "source_url": "https://omlx.ai/benchmarks/performance/4uib8pbe",
   "tg_tokens": null,
   "tps": 57.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-4uib8pbe-ctx4096",
   "model": "Huihui-gemma-4-26B-A4B-it-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: oita7f7m; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 623.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Huihui-gemma-4-26B-A4B-it-abliterated-oQ8e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 4k 623.7 57.0 28.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4uib8pbe",
   "source_url": "https://omlx.ai/benchmarks/performance/4uib8pbe",
   "tg_tokens": null,
   "tps": 57.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-u3qg0r4m-ctx1024",
   "model": "Huihui-gemma-4-31B-it-abliterated-v2",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: bmu82dfm; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 85.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Huihui-gemma-4-31B-it-abliterated-v2-oQ8e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 1k 85.2 19.7 33.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u3qg0r4m",
   "source_url": "https://omlx.ai/benchmarks/performance/u3qg0r4m",
   "tg_tokens": null,
   "tps": 19.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-u3qg0r4m-ctx4096",
   "model": "Huihui-gemma-4-31B-it-abliterated-v2",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.0 GB at this context; duplicate traces of the same session: bmu82dfm; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 83.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Huihui-gemma-4-31B-it-abliterated-v2-oQ8e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 4k 83.5 15.0 35.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u3qg0r4m",
   "source_url": "https://omlx.ai/benchmarks/performance/u3qg0r4m",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ww7p1mzf-ctx1024",
   "model": "K2-Horizon-7B",
   "notes": "oMLX community benchmark trace; peak memory 9.7 GB at this context; duplicate traces of the same session: iwz1a3rk, rmqtt2hu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 910.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "K2-Horizon-7B-oQ8e on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 910.4 52.5 9.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ww7p1mzf",
   "source_url": "https://omlx.ai/benchmarks/performance/ww7p1mzf",
   "tg_tokens": null,
   "tps": 52.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ww7p1mzf-ctx8192",
   "model": "K2-Horizon-7B",
   "notes": "oMLX community benchmark trace; peak memory 10.7 GB at this context; duplicate traces of the same session: iwz1a3rk, rmqtt2hu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 822.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "K2-Horizon-7B-oQ8e on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 822.5 46.9 10.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ww7p1mzf",
   "source_url": "https://omlx.ai/benchmarks/performance/ww7p1mzf",
   "tg_tokens": null,
   "tps": 46.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ww7p1mzf-ctx32768",
   "model": "K2-Horizon-7B",
   "notes": "oMLX community benchmark trace; peak memory 14.1 GB at this context; duplicate traces of the same session: iwz1a3rk, rmqtt2hu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 507.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "K2-Horizon-7B-oQ8e on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 507.8 34.5 14.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ww7p1mzf",
   "source_url": "https://omlx.ai/benchmarks/performance/ww7p1mzf",
   "tg_tokens": null,
   "tps": 34.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-7u6gfq1r-ctx1024",
   "model": "Laguna-S-2.1",
   "notes": "oMLX community benchmark trace; peak memory 60.4 GB at this context; duplicate traces of the same session: 69lf0c09; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 665.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Laguna-S-2.1-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 665.6 61.9 60.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7u6gfq1r",
   "source_url": "https://omlx.ai/benchmarks/performance/7u6gfq1r",
   "tg_tokens": null,
   "tps": 61.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-7u6gfq1r-ctx4096",
   "model": "Laguna-S-2.1",
   "notes": "oMLX community benchmark trace; peak memory 60.5 GB at this context; duplicate traces of the same session: 69lf0c09; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1517.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Laguna-S-2.1-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,517 58.2 60.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7u6gfq1r",
   "source_url": "https://omlx.ai/benchmarks/performance/7u6gfq1r",
   "tg_tokens": null,
   "tps": 58.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-tnwl6gav-ctx8192",
   "model": "Laguna XS 2.1",
   "notes": "oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: n7llsult; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 973.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Laguna-XS-2.1-oQ5e on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 5bit; 8k 973.0 73.8 22.7 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tnwl6gav",
   "source_url": "https://omlx.ai/benchmarks/performance/tnwl6gav",
   "tg_tokens": null,
   "tps": 73.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-tnwl6gav-ctx32768",
   "model": "Laguna XS 2.1",
   "notes": "oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: n7llsult; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 761.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Laguna-XS-2.1-oQ5e on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 5bit; 32k 761.4 48.4 24.1 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tnwl6gav",
   "source_url": "https://omlx.ai/benchmarks/performance/tnwl6gav",
   "tg_tokens": null,
   "tps": 48.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-05kxmmfa-ctx1024",
   "model": "Muse Glimmer",
   "notes": "30B Abliterated MM build; fp16 weights; oMLX community benchmark trace; peak memory 17.9 GB at this context; duplicate traces of the same session: ghu4tre2, mbpctyrf, t4gide6e, zf6pdad0; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-Abliterated-MM-oQ4e-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 103.9 18.3 17.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/05kxmmfa",
   "source_url": "https://omlx.ai/benchmarks/performance/05kxmmfa",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-05kxmmfa-ctx4096",
   "model": "Muse Glimmer",
   "notes": "30B Abliterated MM build; fp16 weights; oMLX community benchmark trace; peak memory 18.2 GB at this context; duplicate traces of the same session: ghu4tre2, mbpctyrf, t4gide6e, zf6pdad0; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-Abliterated-MM-oQ4e-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 127.4 15.4 18.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/05kxmmfa",
   "source_url": "https://omlx.ai/benchmarks/performance/05kxmmfa",
   "tg_tokens": null,
   "tps": 15.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-05kxmmfa-ctx8192",
   "model": "Muse Glimmer",
   "notes": "30B Abliterated MM build; fp16 weights; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: ghu4tre2, mbpctyrf, t4gide6e, zf6pdad0; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 129.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-Abliterated-MM-oQ4e-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 129.9 15.9 18.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/05kxmmfa",
   "source_url": "https://omlx.ai/benchmarks/performance/05kxmmfa",
   "tg_tokens": null,
   "tps": 15.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-05kxmmfa-ctx16384",
   "model": "Muse Glimmer",
   "notes": "30B Abliterated MM build; fp16 weights; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: ghu4tre2, mbpctyrf, t4gide6e, zf6pdad0; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 128.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-Abliterated-MM-oQ4e-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 128.0 42.3 19.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/05kxmmfa",
   "source_url": "https://omlx.ai/benchmarks/performance/05kxmmfa",
   "tg_tokens": null,
   "tps": 42.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-05kxmmfa-ctx32768",
   "model": "Muse Glimmer",
   "notes": "30B Abliterated MM build; fp16 weights; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: ghu4tre2, mbpctyrf, t4gide6e, zf6pdad0; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 118.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-Abliterated-MM-oQ4e-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 32k 118.6 11.8 19.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/05kxmmfa",
   "source_url": "https://omlx.ai/benchmarks/performance/05kxmmfa",
   "tg_tokens": null,
   "tps": 11.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wv7cv2xl-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: hs1iocdb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3905.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 3,905 114.2 37.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wv7cv2xl",
   "source_url": "https://omlx.ai/benchmarks/performance/wv7cv2xl",
   "tg_tokens": null,
   "tps": 114.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wv7cv2xl-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.0 GB at this context; duplicate traces of the same session: hs1iocdb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5661.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 5,661 113.6 38.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wv7cv2xl",
   "source_url": "https://omlx.ai/benchmarks/performance/wv7cv2xl",
   "tg_tokens": null,
   "tps": 113.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4we11ry4-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: kkvwnq81; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3804.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 3,804 128.9 37.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4we11ry4",
   "source_url": "https://omlx.ai/benchmarks/performance/4we11ry4",
   "tg_tokens": null,
   "tps": 128.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4we11ry4-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.0 GB at this context; duplicate traces of the same session: kkvwnq81; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5662.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 5,662 113.0 38.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4we11ry4",
   "source_url": "https://omlx.ai/benchmarks/performance/4we11ry4",
   "tg_tokens": null,
   "tps": 113.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-639xj1ix-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: 6lihv81c; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 707.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 707.1 27.0 5.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/639xj1ix",
   "source_url": "https://omlx.ai/benchmarks/performance/639xj1ix",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-639xj1ix-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: 6lihv81c; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 765.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 765.0 26.3 6.2 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/639xj1ix",
   "source_url": "https://omlx.ai/benchmarks/performance/639xj1ix",
   "tg_tokens": null,
   "tps": 26.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-vw16ys7v-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 61.7 GB at this context; duplicate traces of the same session: 14wvpxkm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2321.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 2,321 75.8 61.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vw16ys7v",
   "source_url": "https://omlx.ai/benchmarks/performance/vw16ys7v",
   "tg_tokens": null,
   "tps": 75.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-vw16ys7v-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 62.5 GB at this context; duplicate traces of the same session: 14wvpxkm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3345.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Coder-Next-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 3,345 72.1 62.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vw16ys7v",
   "source_url": "https://omlx.ai/benchmarks/performance/vw16ys7v",
   "tg_tokens": null,
   "tps": 72.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-6zbcna7k-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 35.5 GB; TTFT 246 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4163.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MTP-mlx-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k tokens 4,163 PP tok/s 105.6 TG tok/s 246 TTFT (ms) 35.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6zbcna7k",
   "source_url": "https://omlx.ai/benchmarks/performance/6zbcna7k",
   "tg_tokens": null,
   "tps": 105.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-9cy8rlls-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.3 GB; TTFT 730 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1402.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k tokens 1,402 PP tok/s 70.9 TG tok/s 730 TTFT (ms) 19.3 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9cy8rlls",
   "source_url": "https://omlx.ai/benchmarks/performance/9cy8rlls",
   "tg_tokens": null,
   "tps": 70.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5h5b9953-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: a4ybm39g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 872.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 872.8 43.5 29.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5h5b9953",
   "source_url": "https://omlx.ai/benchmarks/performance/5h5b9953",
   "tg_tokens": null,
   "tps": 43.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5h5b9953-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.1 GB at this context; duplicate traces of the same session: a4ybm39g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 876.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 876.8 39.4 31.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5h5b9953",
   "source_url": "https://omlx.ai/benchmarks/performance/5h5b9953",
   "tg_tokens": null,
   "tps": 39.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-ffhcoxa1-ctx1024",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 17.6 GB; TTFT 3672 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 278.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Abliterated-MLX-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k tokens 278.9 PP tok/s 24.8 TG tok/s 3672 TTFT (ms) 17.6 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ffhcoxa1",
   "source_url": "https://omlx.ai/benchmarks/performance/ffhcoxa1",
   "tg_tokens": null,
   "tps": 24.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ymy2b6cr-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: ac0zx6qa; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 366.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 366.3 12.0 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ymy2b6cr",
   "source_url": "https://omlx.ai/benchmarks/performance/ymy2b6cr",
   "tg_tokens": null,
   "tps": 12.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ymy2b6cr-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: ac0zx6qa; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 463.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 463.4 12.5 23.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ymy2b6cr",
   "source_url": "https://omlx.ai/benchmarks/performance/ymy2b6cr",
   "tg_tokens": null,
   "tps": 12.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-ao1gtisv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: dezhezb9, n4oyx09u; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 124.3,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 124.3 17.8 20.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ao1gtisv",
   "source_url": "https://omlx.ai/benchmarks/performance/ao1gtisv",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ao1gtisv-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: dezhezb9, n4oyx09u; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 115.6,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k 115.6 14.7 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ao1gtisv",
   "source_url": "https://omlx.ai/benchmarks/performance/ao1gtisv",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ao1gtisv-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: dezhezb9, n4oyx09u; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 89.6,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 64k 89.6 11.3 25.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ao1gtisv",
   "source_url": "https://omlx.ai/benchmarks/performance/ao1gtisv",
   "tg_tokens": null,
   "tps": 11.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-g9695z8s-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 26.0 GB; TTFT 57212 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 286.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k tokens 286.4 PP tok/s 70.2 TG tok/s 57212 TTFT (ms) 26.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g9695z8s",
   "source_url": "https://omlx.ai/benchmarks/performance/g9695z8s",
   "tg_tokens": null,
   "tps": 70.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1yt5mt3y-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: 1twbwlyq, ajb6c5s2, ivzogq2u, myk83jpq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 684.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 684.1 45.4 17.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1yt5mt3y",
   "source_url": "https://omlx.ai/benchmarks/performance/1yt5mt3y",
   "tg_tokens": null,
   "tps": 45.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1yt5mt3y-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 1twbwlyq, ajb6c5s2, ivzogq2u, myk83jpq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 837.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 837.7 49.6 18.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1yt5mt3y",
   "source_url": "https://omlx.ai/benchmarks/performance/1yt5mt3y",
   "tg_tokens": null,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1yt5mt3y-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 1twbwlyq, ajb6c5s2, ivzogq2u, myk83jpq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 750.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 750.4 63.4 19.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1yt5mt3y",
   "source_url": "https://omlx.ai/benchmarks/performance/1yt5mt3y",
   "tg_tokens": null,
   "tps": 63.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1yt5mt3y-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 1twbwlyq, ajb6c5s2, ivzogq2u, myk83jpq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 718.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 718.9 66.4 19.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1yt5mt3y",
   "source_url": "https://omlx.ai/benchmarks/performance/1yt5mt3y",
   "tg_tokens": null,
   "tps": 66.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1yt5mt3y-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 1twbwlyq, ajb6c5s2, ivzogq2u, myk83jpq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 647.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 647.6 72.2 20.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1yt5mt3y",
   "source_url": "https://omlx.ai/benchmarks/performance/1yt5mt3y",
   "tg_tokens": null,
   "tps": 72.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1yt5mt3y-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: 1twbwlyq, ajb6c5s2, ivzogq2u, myk83jpq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 578.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 578.9 44.3 23.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1yt5mt3y",
   "source_url": "https://omlx.ai/benchmarks/performance/1yt5mt3y",
   "tg_tokens": null,
   "tps": 44.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yj2dkne2-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 39.8 GB at this context; duplicate traces of the same session: x246skxw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1056.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,056 62.0 39.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yj2dkne2",
   "source_url": "https://omlx.ai/benchmarks/performance/yj2dkne2",
   "tg_tokens": null,
   "tps": 62.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yj2dkne2-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 40.0 GB at this context; duplicate traces of the same session: x246skxw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1678.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,678 59.1 40.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yj2dkne2",
   "source_url": "https://omlx.ai/benchmarks/performance/yj2dkne2",
   "tg_tokens": null,
   "tps": 59.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-4i40i1i7-ctx1024",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 157.0 GB at this context; duplicate traces of the same session: 9vh8r5zz, arv5zbqk, udayysbc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 557.0,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "DeepSeek-V4-Flash-0731-MXFP4-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 557.0 31.7 157.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4i40i1i7",
   "source_url": "https://omlx.ai/benchmarks/performance/4i40i1i7",
   "tg_tokens": null,
   "tps": 31.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-4i40i1i7-ctx4096",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 157.4 GB at this context; duplicate traces of the same session: 9vh8r5zz, arv5zbqk, udayysbc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 618.1,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "DeepSeek-V4-Flash-0731-MXFP4-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 618.1 31.6 157.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4i40i1i7",
   "source_url": "https://omlx.ai/benchmarks/performance/4i40i1i7",
   "tg_tokens": null,
   "tps": 31.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4i40i1i7-ctx8192",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 165.6 GB at this context; duplicate traces of the same session: 9vh8r5zz, arv5zbqk, udayysbc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 610.4,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "DeepSeek-V4-Flash-0731-MXFP4-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 610.4 5.1 165.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4i40i1i7",
   "source_url": "https://omlx.ai/benchmarks/performance/4i40i1i7",
   "tg_tokens": null,
   "tps": 5.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-4i40i1i7-ctx16384",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 157.5 GB at this context; duplicate traces of the same session: 9vh8r5zz, arv5zbqk, udayysbc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 599.1,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "DeepSeek-V4-Flash-0731-MXFP4-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 599.1 26.7 157.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4i40i1i7",
   "source_url": "https://omlx.ai/benchmarks/performance/4i40i1i7",
   "tg_tokens": null,
   "tps": 26.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-rczqdija-ctx1024",
   "model": "Muse Glimmer",
   "notes": "30B Abliterated MM build; fp16 weights; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 66a0jgtt, 7jrfqdwj, ajc3xp2h, iepcskv5; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 86.3,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "Muse-Glimmer-30B-Abliterated-MM-oQ5-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 5bit; 1k 86.3 13.6 21.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rczqdija",
   "source_url": "https://omlx.ai/benchmarks/performance/rczqdija",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-rczqdija-ctx4096",
   "model": "Muse Glimmer",
   "notes": "30B Abliterated MM build; fp16 weights; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 66a0jgtt, 7jrfqdwj, ajc3xp2h, iepcskv5; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 115.8,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "Muse-Glimmer-30B-Abliterated-MM-oQ5-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 5bit; 4k 115.8 12.3 21.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rczqdija",
   "source_url": "https://omlx.ai/benchmarks/performance/rczqdija",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-rczqdija-ctx8192",
   "model": "Muse Glimmer",
   "notes": "30B Abliterated MM build; fp16 weights; oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: 66a0jgtt, 7jrfqdwj, ajc3xp2h, iepcskv5; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 122.8,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "Muse-Glimmer-30B-Abliterated-MM-oQ5-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 5bit; 8k 122.8 13.7 21.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rczqdija",
   "source_url": "https://omlx.ai/benchmarks/performance/rczqdija",
   "tg_tokens": null,
   "tps": 13.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-rczqdija-ctx16384",
   "model": "Muse Glimmer",
   "notes": "30B Abliterated MM build; fp16 weights; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: 66a0jgtt, 7jrfqdwj, ajc3xp2h, iepcskv5; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 113.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "Muse-Glimmer-30B-Abliterated-MM-oQ5-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 5bit; 16k 113.0 43.2 22.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rczqdija",
   "source_url": "https://omlx.ai/benchmarks/performance/rczqdija",
   "tg_tokens": null,
   "tps": 43.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-rczqdija-ctx32768",
   "model": "Muse Glimmer",
   "notes": "30B Abliterated MM build; fp16 weights; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: 66a0jgtt, 7jrfqdwj, ajc3xp2h, iepcskv5; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.6,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "Muse-Glimmer-30B-Abliterated-MM-oQ5-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 5bit; 32k 103.6 12.0 23.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rczqdija",
   "source_url": "https://omlx.ai/benchmarks/performance/rczqdija",
   "tg_tokens": null,
   "tps": 12.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-liotsgvp-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: ov12sk8r, tkzit3aw; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1400.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 1,400 70.7 19.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/liotsgvp",
   "source_url": "https://omlx.ai/benchmarks/performance/liotsgvp",
   "tg_tokens": null,
   "tps": 70.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-liotsgvp-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: ov12sk8r, tkzit3aw; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1885.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,885 69.2 19.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/liotsgvp",
   "source_url": "https://omlx.ai/benchmarks/performance/liotsgvp",
   "tg_tokens": null,
   "tps": 69.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-liotsgvp-ctx8192",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: ov12sk8r, tkzit3aw; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1828.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 8k 1,828 66.2 20.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/liotsgvp",
   "source_url": "https://omlx.ai/benchmarks/performance/liotsgvp",
   "tg_tokens": null,
   "tps": 66.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-phjbbz37-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: gtx5shet, sujmc0x4; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1388.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 1,388 70.3 19.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/phjbbz37",
   "source_url": "https://omlx.ai/benchmarks/performance/phjbbz37",
   "tg_tokens": null,
   "tps": 70.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-phjbbz37-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: gtx5shet, sujmc0x4; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1853.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,853 68.7 19.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/phjbbz37",
   "source_url": "https://omlx.ai/benchmarks/performance/phjbbz37",
   "tg_tokens": null,
   "tps": 68.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-phjbbz37-ctx8192",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: gtx5shet, sujmc0x4; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1815.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 8k 1,815 66.3 20.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/phjbbz37",
   "source_url": "https://omlx.ai/benchmarks/performance/phjbbz37",
   "tg_tokens": null,
   "tps": 66.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-rpjwy0oh-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: a6uzn4hz; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 67.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 67.7 6.6 17.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rpjwy0oh",
   "source_url": "https://omlx.ai/benchmarks/performance/rpjwy0oh",
   "tg_tokens": null,
   "tps": 6.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-rpjwy0oh-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: a6uzn4hz; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 64.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 16k 64.0 6.2 18.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rpjwy0oh",
   "source_url": "https://omlx.ai/benchmarks/performance/rpjwy0oh",
   "tg_tokens": null,
   "tps": 6.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ex5gc8z1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: 6ja2fflc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 734.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 734.8 15.8 29.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ex5gc8z1",
   "source_url": "https://omlx.ai/benchmarks/performance/ex5gc8z1",
   "tg_tokens": null,
   "tps": 15.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ex5gc8z1-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: 6ja2fflc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 522.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 522.3 14.1 30.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ex5gc8z1",
   "source_url": "https://omlx.ai/benchmarks/performance/ex5gc8z1",
   "tg_tokens": null,
   "tps": 14.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-rwdo2tih-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: 4tuzck22; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 472.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 472.3 28.3 18.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rwdo2tih",
   "source_url": "https://omlx.ai/benchmarks/performance/rwdo2tih",
   "tg_tokens": null,
   "tps": 28.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-rwdo2tih-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: 4tuzck22; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 455.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 455.9 28.4 18.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rwdo2tih",
   "source_url": "https://omlx.ai/benchmarks/performance/rwdo2tih",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-u0amhmjk-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 684.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 684.1 45.4 17.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u0amhmjk",
   "source_url": "https://omlx.ai/benchmarks/performance/u0amhmjk",
   "tg_tokens": null,
   "tps": 45.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-u0amhmjk-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 837.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 837.7 49.6 18.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u0amhmjk",
   "source_url": "https://omlx.ai/benchmarks/performance/u0amhmjk",
   "tg_tokens": null,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-u0amhmjk-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 750.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 750.4 63.4 19.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u0amhmjk",
   "source_url": "https://omlx.ai/benchmarks/performance/u0amhmjk",
   "tg_tokens": null,
   "tps": 63.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-u0amhmjk-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 718.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 718.9 66.4 19.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u0amhmjk",
   "source_url": "https://omlx.ai/benchmarks/performance/u0amhmjk",
   "tg_tokens": null,
   "tps": 66.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-u0amhmjk-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 647.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 647.6 72.2 20.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u0amhmjk",
   "source_url": "https://omlx.ai/benchmarks/performance/u0amhmjk",
   "tg_tokens": null,
   "tps": 72.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-u0amhmjk-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 578.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 578.9 44.3 23.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u0amhmjk",
   "source_url": "https://omlx.ai/benchmarks/performance/u0amhmjk",
   "tg_tokens": null,
   "tps": 44.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-a95sjnry-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.2 GB at this context; duplicate traces of the same session: 2okent1i, nbdezz1f, v94c6ek2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 392.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 392.7 22.9 39.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a95sjnry",
   "source_url": "https://omlx.ai/benchmarks/performance/a95sjnry",
   "tg_tokens": null,
   "tps": 22.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-a95sjnry-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 42.4 GB at this context; duplicate traces of the same session: 2okent1i, nbdezz1f, v94c6ek2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 465.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 465.8 32.0 42.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a95sjnry",
   "source_url": "https://omlx.ai/benchmarks/performance/a95sjnry",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-a95sjnry-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 42.9 GB at this context; duplicate traces of the same session: 2okent1i, nbdezz1f, v94c6ek2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 510.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 510.6 33.7 42.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a95sjnry",
   "source_url": "https://omlx.ai/benchmarks/performance/a95sjnry",
   "tg_tokens": null,
   "tps": 33.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-a95sjnry-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.8 GB at this context; duplicate traces of the same session: 2okent1i, nbdezz1f, v94c6ek2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 504.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k 504.3 30.5 43.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a95sjnry",
   "source_url": "https://omlx.ai/benchmarks/performance/a95sjnry",
   "tg_tokens": null,
   "tps": 30.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-9xgdymep-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: unmdl1zt; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 489.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27b-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 489.7 34.2 18.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9xgdymep",
   "source_url": "https://omlx.ai/benchmarks/performance/9xgdymep",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-9xgdymep-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: unmdl1zt; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 475.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27b-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 475.8 38.4 19.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9xgdymep",
   "source_url": "https://omlx.ai/benchmarks/performance/9xgdymep",
   "tg_tokens": null,
   "tps": 38.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-envsbguu-ctx32768",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: 4bx36mqd, 6zhwltzn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2069.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 2,069 51.8 23.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/envsbguu",
   "source_url": "https://omlx.ai/benchmarks/performance/envsbguu",
   "tg_tokens": null,
   "tps": 51.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-envsbguu-ctx65536",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: 4bx36mqd, 6zhwltzn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1507.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 1,507 41.2 24.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/envsbguu",
   "source_url": "https://omlx.ai/benchmarks/performance/envsbguu",
   "tg_tokens": null,
   "tps": 41.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-envsbguu-ctx131072",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: 4bx36mqd, 6zhwltzn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 939.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 128k 939.2 33.8 26.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/envsbguu",
   "source_url": "https://omlx.ai/benchmarks/performance/envsbguu",
   "tg_tokens": null,
   "tps": 33.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4z94o1d3-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.5 GB at this context; duplicate traces of the same session: 2xuhq255; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3303.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-9B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 8bit; 4k 3,303 57.5 10.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4z94o1d3",
   "source_url": "https://omlx.ai/benchmarks/performance/4z94o1d3",
   "tg_tokens": null,
   "tps": 57.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4z94o1d3-ctx16384",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.9 GB at this context; duplicate traces of the same session: 2xuhq255; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3188.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-9B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 8bit; 16k 3,188 55.6 10.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4z94o1d3",
   "source_url": "https://omlx.ai/benchmarks/performance/4z94o1d3",
   "tg_tokens": null,
   "tps": 55.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q06zkqs6-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: b593z3ry; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 783.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 783.3 40.2 17.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q06zkqs6",
   "source_url": "https://omlx.ai/benchmarks/performance/q06zkqs6",
   "tg_tokens": null,
   "tps": 40.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-q06zkqs6-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: b593z3ry; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1275.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,275 94.7 19.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q06zkqs6",
   "source_url": "https://omlx.ai/benchmarks/performance/q06zkqs6",
   "tg_tokens": null,
   "tps": 94.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-gzraz0s2-ctx32768",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: nw6pvd0v, tbnfsl6o; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2355.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 2,355 84.0 23.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gzraz0s2",
   "source_url": "https://omlx.ai/benchmarks/performance/gzraz0s2",
   "tg_tokens": null,
   "tps": 84.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-gzraz0s2-ctx65536",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: nw6pvd0v, tbnfsl6o; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1673.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 1,673 66.4 24.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gzraz0s2",
   "source_url": "https://omlx.ai/benchmarks/performance/gzraz0s2",
   "tg_tokens": null,
   "tps": 66.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-gzraz0s2-ctx131072",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: nw6pvd0v, tbnfsl6o; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1011.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 128k 1,011 52.7 26.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gzraz0s2",
   "source_url": "https://omlx.ai/benchmarks/performance/gzraz0s2",
   "tg_tokens": null,
   "tps": 52.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-eg69tes1-ctx32768",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: 4awtrjhh, m5te7jbn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2124.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 2,124 80.6 23.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eg69tes1",
   "source_url": "https://omlx.ai/benchmarks/performance/eg69tes1",
   "tg_tokens": null,
   "tps": 80.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-eg69tes1-ctx65536",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: 4awtrjhh, m5te7jbn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1420.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 1,420 62.2 24.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eg69tes1",
   "source_url": "https://omlx.ai/benchmarks/performance/eg69tes1",
   "tg_tokens": null,
   "tps": 62.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-eg69tes1-ctx131072",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: 4awtrjhh, m5te7jbn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 950.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 128k 950.7 58.3 26.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eg69tes1",
   "source_url": "https://omlx.ai/benchmarks/performance/eg69tes1",
   "tg_tokens": null,
   "tps": 58.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-bbinhusg-ctx4096",
   "model": "Qwen3-1.7B",
   "notes": "oMLX community benchmark trace; peak memory 2.0 GB at this context; duplicate traces of the same session: o8vdgo0p; config=19c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 955.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-1.7B-4bit on M2 Pro (19c); M2 Pro (19c) \u00b7 16 GB \u00b7 4bit; 4k 955.8 104.0 2.0 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bbinhusg",
   "source_url": "https://omlx.ai/benchmarks/performance/bbinhusg",
   "tg_tokens": null,
   "tps": 104.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-bbinhusg-ctx16384",
   "model": "Qwen3-1.7B",
   "notes": "oMLX community benchmark trace; peak memory 3.2 GB at this context; duplicate traces of the same session: o8vdgo0p; config=19c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 681.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-1.7B-4bit on M2 Pro (19c); M2 Pro (19c) \u00b7 16 GB \u00b7 4bit; 16k 681.1 54.8 3.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bbinhusg",
   "source_url": "https://omlx.ai/benchmarks/performance/bbinhusg",
   "tg_tokens": null,
   "tps": 54.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-e4uommve-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: 3xc3kjw4; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1066.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 16 GB \u00b7 4bit; 1k 1,066 28.8 5.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e4uommve",
   "source_url": "https://omlx.ai/benchmarks/performance/e4uommve",
   "tg_tokens": null,
   "tps": 28.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-e4uommve-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: 3xc3kjw4; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1113.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 16 GB \u00b7 4bit; 4k 1,113 28.0 6.2 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e4uommve",
   "source_url": "https://omlx.ai/benchmarks/performance/e4uommve",
   "tg_tokens": null,
   "tps": 28.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-7tcjqo34-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 3w4c0o13, 5q44q2gt, zpqyne9u; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 479.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 479.1 17.4 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7tcjqo34",
   "source_url": "https://omlx.ai/benchmarks/performance/7tcjqo34",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-7tcjqo34-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 3w4c0o13, 5q44q2gt, zpqyne9u; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 487.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 487.0 16.8 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7tcjqo34",
   "source_url": "https://omlx.ai/benchmarks/performance/7tcjqo34",
   "tg_tokens": null,
   "tps": 16.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-7tcjqo34-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 3w4c0o13, 5q44q2gt, zpqyne9u; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 421.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 421.1 16.3 17.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7tcjqo34",
   "source_url": "https://omlx.ai/benchmarks/performance/7tcjqo34",
   "tg_tokens": null,
   "tps": 16.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-7tcjqo34-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 3w4c0o13, 5q44q2gt, zpqyne9u; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 409.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 409.8 14.7 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7tcjqo34",
   "source_url": "https://omlx.ai/benchmarks/performance/7tcjqo34",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-hyww01qn-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: f83t07w3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 459.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 459.5 17.5 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hyww01qn",
   "source_url": "https://omlx.ai/benchmarks/performance/hyww01qn",
   "tg_tokens": null,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-hyww01qn-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: f83t07w3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 477.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 477.8 17.2 17.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hyww01qn",
   "source_url": "https://omlx.ai/benchmarks/performance/hyww01qn",
   "tg_tokens": null,
   "tps": 17.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-mkwaibqu-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: e439m16v; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 242.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 242.3 30.9 18.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mkwaibqu",
   "source_url": "https://omlx.ai/benchmarks/performance/mkwaibqu",
   "tg_tokens": null,
   "tps": 30.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-mkwaibqu-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: e439m16v; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 241.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 241.3 47.6 22.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mkwaibqu",
   "source_url": "https://omlx.ai/benchmarks/performance/mkwaibqu",
   "tg_tokens": null,
   "tps": 47.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qpmyoqc0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: om882vus; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 638.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 638.0 24.5 19.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qpmyoqc0",
   "source_url": "https://omlx.ai/benchmarks/performance/qpmyoqc0",
   "tg_tokens": null,
   "tps": 24.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qpmyoqc0-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: om882vus; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 670.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 670.7 48.6 19.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qpmyoqc0",
   "source_url": "https://omlx.ai/benchmarks/performance/qpmyoqc0",
   "tg_tokens": null,
   "tps": 48.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-17t5t91q-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: 39mhtxrq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 452.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 452.4 24.7 27.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/17t5t91q",
   "source_url": "https://omlx.ai/benchmarks/performance/17t5t91q",
   "tg_tokens": null,
   "tps": 24.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-17t5t91q-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: 39mhtxrq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 427.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 16k 427.6 34.4 28.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/17t5t91q",
   "source_url": "https://omlx.ai/benchmarks/performance/17t5t91q",
   "tg_tokens": null,
   "tps": 34.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-rxlppm21-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: gaix98tp, xp3z1pmb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 177.8,
   "provenance": "community",
   "quant": "GSQ-RCO-Q2",
   "quote": "Qwen3.8-Flash-Next-GSQ-RCO-Q2-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 177.8 30.3 33.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rxlppm21",
   "source_url": "https://omlx.ai/benchmarks/performance/rxlppm21",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-rxlppm21-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 35.4 GB at this context; duplicate traces of the same session: gaix98tp, xp3z1pmb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 194.3,
   "provenance": "community",
   "quant": "GSQ-RCO-Q2",
   "quote": "Qwen3.8-Flash-Next-GSQ-RCO-Q2-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 194.3 20.5 35.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rxlppm21",
   "source_url": "https://omlx.ai/benchmarks/performance/rxlppm21",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-rxlppm21-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 38.1 GB at this context; duplicate traces of the same session: gaix98tp, xp3z1pmb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 243.7,
   "provenance": "community",
   "quant": "GSQ-RCO-Q2",
   "quote": "Qwen3.8-Flash-Next-GSQ-RCO-Q2-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 243.7 22.9 38.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rxlppm21",
   "source_url": "https://omlx.ai/benchmarks/performance/rxlppm21",
   "tg_tokens": null,
   "tps": 22.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-hkzgadry-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: 4a6112lb, il4znrm2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 169.9,
   "provenance": "community",
   "quant": "GSQ-RCO-Q2",
   "quote": "Qwen3.8-Flash-Next-GSQ-RCO-Q2-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 169.9 19.3 24.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hkzgadry",
   "source_url": "https://omlx.ai/benchmarks/performance/hkzgadry",
   "tg_tokens": null,
   "tps": 19.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-hkzgadry-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: 4a6112lb, il4znrm2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 328.9,
   "provenance": "community",
   "quant": "GSQ-RCO-Q2",
   "quote": "Qwen3.8-Flash-Next-GSQ-RCO-Q2-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 328.9 14.1 26.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hkzgadry",
   "source_url": "https://omlx.ai/benchmarks/performance/hkzgadry",
   "tg_tokens": null,
   "tps": 14.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-hkzgadry-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: 4a6112lb, il4znrm2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 437.5,
   "provenance": "community",
   "quant": "GSQ-RCO-Q2",
   "quote": "Qwen3.8-Flash-Next-GSQ-RCO-Q2-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 437.5 14.7 28.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hkzgadry",
   "source_url": "https://omlx.ai/benchmarks/performance/hkzgadry",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-yci55trd-ctx1024",
   "model": "Qwen3.8-27B-aeon-ultimate-uncensored",
   "notes": "attention8 bf16recurrence vision mtplx build; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: t4boz2cv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 248.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "qwen3.8-27b-aeon-ultimate-uncensored-attention8-bf16recurrence-vision-mtplx on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 248.5 37.1 24.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yci55trd",
   "source_url": "https://omlx.ai/benchmarks/performance/yci55trd",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-yci55trd-ctx4096",
   "model": "Qwen3.8-27B-aeon-ultimate-uncensored",
   "notes": "attention8 bf16recurrence vision mtplx build; oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: t4boz2cv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 236.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "qwen3.8-27b-aeon-ultimate-uncensored-attention8-bf16recurrence-vision-mtplx on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 236.4 38.9 28.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yci55trd",
   "source_url": "https://omlx.ai/benchmarks/performance/yci55trd",
   "tg_tokens": null,
   "tps": 38.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-sm28t3qj-ctx1024",
   "model": "Agnes-3.0-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: oskd8ns0; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 67.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Agnes-3.0-Flash-MLX-oQ4e-MTP on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 67.9 14.5 22.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sm28t3qj",
   "source_url": "https://omlx.ai/benchmarks/performance/sm28t3qj",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-sm28t3qj-ctx8192",
   "model": "Agnes-3.0-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.2 GB at this context; duplicate traces of the same session: oskd8ns0; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 61.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Agnes-3.0-Flash-MLX-oQ4e-MTP on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 8k 61.7 14.9 28.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sm28t3qj",
   "source_url": "https://omlx.ai/benchmarks/performance/sm28t3qj",
   "tg_tokens": null,
   "tps": 14.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-de3xd0id-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: ghv3sfyt, qzmoph8q; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 315.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ4e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 315.3 24.8 25.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/de3xd0id",
   "source_url": "https://omlx.ai/benchmarks/performance/de3xd0id",
   "tg_tokens": null,
   "tps": 24.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-de3xd0id-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: ghv3sfyt, qzmoph8q; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 286.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ4e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 286.2 25.5 27.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/de3xd0id",
   "source_url": "https://omlx.ai/benchmarks/performance/de3xd0id",
   "tg_tokens": null,
   "tps": 25.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-de3xd0id-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace; peak memory 32.2 GB at this context; duplicate traces of the same session: ghv3sfyt, qzmoph8q; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 247.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ4e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 128k 247.5 25.0 32.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/de3xd0id",
   "source_url": "https://omlx.ai/benchmarks/performance/de3xd0id",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-9vp2rwmh-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 2befb0ly; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 731.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 1k 731.1 90.0 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9vp2rwmh",
   "source_url": "https://omlx.ai/benchmarks/performance/9vp2rwmh",
   "tg_tokens": null,
   "tps": 90.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-9vp2rwmh-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 2befb0ly; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 782.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 782.0 88.0 21.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9vp2rwmh",
   "source_url": "https://omlx.ai/benchmarks/performance/9vp2rwmh",
   "tg_tokens": null,
   "tps": 88.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-4isf7rm5-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: jpfkap8f; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1677.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 1,677 71.6 19.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4isf7rm5",
   "source_url": "https://omlx.ai/benchmarks/performance/4isf7rm5",
   "tg_tokens": null,
   "tps": 71.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-4isf7rm5-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: jpfkap8f; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2180.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 2,180 70.2 19.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4isf7rm5",
   "source_url": "https://omlx.ai/benchmarks/performance/4isf7rm5",
   "tg_tokens": null,
   "tps": 70.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-awnomr2d-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: lonbfzdw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4285.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,285 222.7 21.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/awnomr2d",
   "source_url": "https://omlx.ai/benchmarks/performance/awnomr2d",
   "tg_tokens": null,
   "tps": 222.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-awnomr2d-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: lonbfzdw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6288.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,288 199.7 22.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/awnomr2d",
   "source_url": "https://omlx.ai/benchmarks/performance/awnomr2d",
   "tg_tokens": null,
   "tps": 199.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0lzunq2t-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: udomboki; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 475.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 475.7 28.4 19.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0lzunq2t",
   "source_url": "https://omlx.ai/benchmarks/performance/0lzunq2t",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0lzunq2t-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: udomboki; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 469.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 469.8 30.5 21.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0lzunq2t",
   "source_url": "https://omlx.ai/benchmarks/performance/0lzunq2t",
   "tg_tokens": null,
   "tps": 30.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-38mba2zu-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: v9e8iuat; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 458.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 458.9 25.6 19.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/38mba2zu",
   "source_url": "https://omlx.ai/benchmarks/performance/38mba2zu",
   "tg_tokens": null,
   "tps": 25.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-38mba2zu-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: v9e8iuat; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 482.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 482.9 32.5 21.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/38mba2zu",
   "source_url": "https://omlx.ai/benchmarks/performance/38mba2zu",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-j2qgiiok-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: xfakshh7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 863.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 863.9 32.3 16.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j2qgiiok",
   "source_url": "https://omlx.ai/benchmarks/performance/j2qgiiok",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-j2qgiiok-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: xfakshh7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 982.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 982.8 31.6 17.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j2qgiiok",
   "source_url": "https://omlx.ai/benchmarks/performance/j2qgiiok",
   "tg_tokens": null,
   "tps": 31.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fdr4e48z-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: 5ckp35no, asx4i553, fuchrcds; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 486.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 486.6 33.1 17.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fdr4e48z",
   "source_url": "https://omlx.ai/benchmarks/performance/fdr4e48z",
   "tg_tokens": null,
   "tps": 33.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fdr4e48z-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: 5ckp35no, asx4i553, fuchrcds; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 497.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 497.5 35.6 18.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fdr4e48z",
   "source_url": "https://omlx.ai/benchmarks/performance/fdr4e48z",
   "tg_tokens": null,
   "tps": 35.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-fdr4e48z-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: 5ckp35no, asx4i553, fuchrcds; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 422.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 422.9 25.8 19.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fdr4e48z",
   "source_url": "https://omlx.ai/benchmarks/performance/fdr4e48z",
   "tg_tokens": null,
   "tps": 25.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-fdr4e48z-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 5ckp35no, asx4i553, fuchrcds; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 403.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 403.2 13.1 20.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fdr4e48z",
   "source_url": "https://omlx.ai/benchmarks/performance/fdr4e48z",
   "tg_tokens": null,
   "tps": 13.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-56dzw6vr-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 24.7 GB; TTFT 12348 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 331.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 4k tokens 331.7 PP tok/s 39.9 TG tok/s 12348 TTFT (ms) 24.7 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/56dzw6vr",
   "source_url": "https://omlx.ai/benchmarks/performance/56dzw6vr",
   "tg_tokens": null,
   "tps": 39.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-0744t5a8-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 29.8 GB; TTFT 255909 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 256.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 64k tokens 256.1 PP tok/s 2.4 TG tok/s 255909 TTFT (ms) 29.8 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0744t5a8",
   "source_url": "https://omlx.ai/benchmarks/performance/0744t5a8",
   "tg_tokens": null,
   "tps": 2.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fk2oylna-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: kky87lxe; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 415.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 415.9 26.5 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fk2oylna",
   "source_url": "https://omlx.ai/benchmarks/performance/fk2oylna",
   "tg_tokens": null,
   "tps": 26.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fk2oylna-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: kky87lxe; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 454.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 454.2 27.0 24.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fk2oylna",
   "source_url": "https://omlx.ai/benchmarks/performance/fk2oylna",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-pybva14r-ctx4096",
   "model": "Signal-3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: io4r8tc8; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 394.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Signal-3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 394.7 27.8 18.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pybva14r",
   "source_url": "https://omlx.ai/benchmarks/performance/pybva14r",
   "tg_tokens": null,
   "tps": 27.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-pybva14r-ctx16384",
   "model": "Signal-3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: io4r8tc8; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 381.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Signal-3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 381.6 33.2 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pybva14r",
   "source_url": "https://omlx.ai/benchmarks/performance/pybva14r",
   "tg_tokens": null,
   "tps": 33.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-989o81ml-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 8.1 GB at this context; duplicate traces of the same session: lz6k6ti9; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 351.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-9B-oQ4e-mtp on M3 Pro (14c); M3 Pro (14c) \u00b7 36 GB \u00b7 4bit; 4k 351.9 43.3 8.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/989o81ml",
   "source_url": "https://omlx.ai/benchmarks/performance/989o81ml",
   "tg_tokens": null,
   "tps": 43.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-989o81ml-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 8.7 GB at this context; duplicate traces of the same session: lz6k6ti9; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 339.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-9B-oQ4e-mtp on M3 Pro (14c); M3 Pro (14c) \u00b7 36 GB \u00b7 4bit; 16k 339.1 46.2 8.7 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/989o81ml",
   "source_url": "https://omlx.ai/benchmarks/performance/989o81ml",
   "tg_tokens": null,
   "tps": 46.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-u35rwo4y-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 39.8 GB; TTFT 50976 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1286.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k tokens 1,286 PP tok/s 50.1 TG tok/s 50976 TTFT (ms) 39.8 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u35rwo4y",
   "source_url": "https://omlx.ai/benchmarks/performance/u35rwo4y",
   "tg_tokens": null,
   "tps": 50.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-b71c4uwk-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 31.1 GB; TTFT 241480 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 271.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 64k tokens 271.4 PP tok/s 17.2 TG tok/s 241480 TTFT (ms) 31.1 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b71c4uwk",
   "source_url": "https://omlx.ai/benchmarks/performance/b71c4uwk",
   "tg_tokens": null,
   "tps": 17.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-fwx74mdj-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 35.3 GB at this context; duplicate traces of the same session: 52w0sngk, 86rx4f5b, af8golf0, u1lej8xq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 247.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 247.3 28.6 35.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fwx74mdj",
   "source_url": "https://omlx.ai/benchmarks/performance/fwx74mdj",
   "tg_tokens": null,
   "tps": 28.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-fwx74mdj-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 37.5 GB at this context; duplicate traces of the same session: 52w0sngk, 86rx4f5b, af8golf0, u1lej8xq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 239.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 239.9 39.4 37.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fwx74mdj",
   "source_url": "https://omlx.ai/benchmarks/performance/fwx74mdj",
   "tg_tokens": null,
   "tps": 39.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-fwx74mdj-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 41.5 GB at this context; duplicate traces of the same session: 52w0sngk, 86rx4f5b, af8golf0, u1lej8xq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 217.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 217.3 36.8 41.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fwx74mdj",
   "source_url": "https://omlx.ai/benchmarks/performance/fwx74mdj",
   "tg_tokens": null,
   "tps": 36.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-fwx74mdj-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 49.2 GB at this context; duplicate traces of the same session: 52w0sngk, 86rx4f5b, af8golf0, u1lej8xq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 177.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 177.6 26.2 49.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fwx74mdj",
   "source_url": "https://omlx.ai/benchmarks/performance/fwx74mdj",
   "tg_tokens": null,
   "tps": 26.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-fwx74mdj-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 63.7 GB at this context; duplicate traces of the same session: 52w0sngk, 86rx4f5b, af8golf0, u1lej8xq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 135.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 135.5 25.1 63.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fwx74mdj",
   "source_url": "https://omlx.ai/benchmarks/performance/fwx74mdj",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-x7vqxsea-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP320",
   "notes": "MTP Vision build; oMLX community benchmark trace; peak memory 37.8 GB at this context; duplicate traces of the same session: 02wntvo3, w63xvt4e; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 762.4,
   "provenance": "community",
   "quant": "oQ3e-DWQ",
   "quote": "Qwen3.8-Flash-Next-REAP320-oQ3e-DWQ-MTP-Vision-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 3bit; 1k 762.4 41.6 37.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x7vqxsea",
   "source_url": "https://omlx.ai/benchmarks/performance/x7vqxsea",
   "tg_tokens": null,
   "tps": 41.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-x7vqxsea-ctx8192",
   "model": "Qwen3.8-Flash-Next-REAP320",
   "notes": "MTP Vision build; oMLX community benchmark trace; peak memory 38.9 GB at this context; duplicate traces of the same session: 02wntvo3, w63xvt4e; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1097.0,
   "provenance": "community",
   "quant": "oQ3e-DWQ",
   "quote": "Qwen3.8-Flash-Next-REAP320-oQ3e-DWQ-MTP-Vision-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 3bit; 8k 1,097 42.1 38.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x7vqxsea",
   "source_url": "https://omlx.ai/benchmarks/performance/x7vqxsea",
   "tg_tokens": null,
   "tps": 42.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-x7vqxsea-ctx32768",
   "model": "Qwen3.8-Flash-Next-REAP320",
   "notes": "MTP Vision build; oMLX community benchmark trace; peak memory 40.5 GB at this context; duplicate traces of the same session: 02wntvo3, w63xvt4e; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1232.0,
   "provenance": "community",
   "quant": "oQ3e-DWQ",
   "quote": "Qwen3.8-Flash-Next-REAP320-oQ3e-DWQ-MTP-Vision-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 3bit; 32k 1,232 42.7 40.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x7vqxsea",
   "source_url": "https://omlx.ai/benchmarks/performance/x7vqxsea",
   "tg_tokens": null,
   "tps": 42.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-wmh0byrj-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP320",
   "notes": "MTP Vision build; oMLX community benchmark trace; peak memory 36.7 GB at this context; duplicate traces of the same session: fbxb1q27; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 782.6,
   "provenance": "community",
   "quant": "oQ3e-DWQ",
   "quote": "Qwen3.8-Flash-Next-REAP320-oQ3e-DWQ-MTP-Vision-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 3bit; 1k 782.6 40.2 36.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wmh0byrj",
   "source_url": "https://omlx.ai/benchmarks/performance/wmh0byrj",
   "tg_tokens": null,
   "tps": 40.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-wmh0byrj-ctx8192",
   "model": "Qwen3.8-Flash-Next-REAP320",
   "notes": "MTP Vision build; oMLX community benchmark trace; peak memory 37.9 GB at this context; duplicate traces of the same session: fbxb1q27; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1041.0,
   "provenance": "community",
   "quant": "oQ3e-DWQ",
   "quote": "Qwen3.8-Flash-Next-REAP320-oQ3e-DWQ-MTP-Vision-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 3bit; 8k 1,041 42.8 37.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wmh0byrj",
   "source_url": "https://omlx.ai/benchmarks/performance/wmh0byrj",
   "tg_tokens": null,
   "tps": 42.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-qn4e1c8r-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.2 GB at this context; duplicate traces of the same session: yqm0xppc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 417.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 417.2 56.7 69.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qn4e1c8r",
   "source_url": "https://omlx.ai/benchmarks/performance/qn4e1c8r",
   "tg_tokens": null,
   "tps": 56.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-qn4e1c8r-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.9 GB at this context; duplicate traces of the same session: yqm0xppc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 599.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 599.4 48.8 69.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qn4e1c8r",
   "source_url": "https://omlx.ai/benchmarks/performance/qn4e1c8r",
   "tg_tokens": null,
   "tps": 48.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2ecvoxpo-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: 1yzalxan, akzyfxpz, ckkqi47j, euokdxuk, k3zjsv3g, xecupaj5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 659.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 659.4 70.3 101.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ecvoxpo",
   "source_url": "https://omlx.ai/benchmarks/performance/2ecvoxpo",
   "tg_tokens": null,
   "tps": 70.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2ecvoxpo-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 103.8 GB at this context; duplicate traces of the same session: 1yzalxan, akzyfxpz, ckkqi47j, euokdxuk, k3zjsv3g, xecupaj5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 715.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 715.3 58.8 103.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ecvoxpo",
   "source_url": "https://omlx.ai/benchmarks/performance/2ecvoxpo",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-2ecvoxpo-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 103.9 GB at this context; duplicate traces of the same session: 1yzalxan, akzyfxpz, ckkqi47j, euokdxuk, k3zjsv3g, xecupaj5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 713.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 713.5 69.3 103.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ecvoxpo",
   "source_url": "https://omlx.ai/benchmarks/performance/2ecvoxpo",
   "tg_tokens": null,
   "tps": 69.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2ecvoxpo-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 104.2 GB at this context; duplicate traces of the same session: 1yzalxan, akzyfxpz, ckkqi47j, euokdxuk, k3zjsv3g, xecupaj5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 710.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 710.9 50.7 104.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ecvoxpo",
   "source_url": "https://omlx.ai/benchmarks/performance/2ecvoxpo",
   "tg_tokens": null,
   "tps": 50.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2ecvoxpo-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 104.8 GB at this context; duplicate traces of the same session: 1yzalxan, akzyfxpz, ckkqi47j, euokdxuk, k3zjsv3g, xecupaj5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 705.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 705.5 48.4 104.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ecvoxpo",
   "source_url": "https://omlx.ai/benchmarks/performance/2ecvoxpo",
   "tg_tokens": null,
   "tps": 48.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2ecvoxpo-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 106.0 GB at this context; duplicate traces of the same session: 1yzalxan, akzyfxpz, ckkqi47j, euokdxuk, k3zjsv3g, xecupaj5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 699.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 699.5 49.5 106.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ecvoxpo",
   "source_url": "https://omlx.ai/benchmarks/performance/2ecvoxpo",
   "tg_tokens": null,
   "tps": 49.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2ecvoxpo-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 108.8 GB at this context; duplicate traces of the same session: 1yzalxan, akzyfxpz, ckkqi47j, euokdxuk, k3zjsv3g, xecupaj5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 690.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 690.1 43.6 108.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ecvoxpo",
   "source_url": "https://omlx.ai/benchmarks/performance/2ecvoxpo",
   "tg_tokens": null,
   "tps": 43.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-r1bacwh7-ctx1024",
   "model": "Qwen2.5-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 4.8 GB at this context; duplicate traces of the same session: 890j49ns; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 615.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-7B-Instruct-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 1k 615.3 73.5 4.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r1bacwh7",
   "source_url": "https://omlx.ai/benchmarks/performance/r1bacwh7",
   "tg_tokens": null,
   "tps": 73.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-r1bacwh7-ctx4096",
   "model": "Qwen2.5-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: 890j49ns; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 600.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-7B-Instruct-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 600.3 68.0 4.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r1bacwh7",
   "source_url": "https://omlx.ai/benchmarks/performance/r1bacwh7",
   "tg_tokens": null,
   "tps": 68.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-25hulm7n-ctx1024",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 2argdl8t; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 627.6,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 627.6 70.2 19.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/25hulm7n",
   "source_url": "https://omlx.ai/benchmarks/performance/25hulm7n",
   "tg_tokens": null,
   "tps": 70.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-25hulm7n-ctx4096",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 2argdl8t; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 701.9,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 701.9 66.1 22.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/25hulm7n",
   "source_url": "https://omlx.ai/benchmarks/performance/25hulm7n",
   "tg_tokens": null,
   "tps": 66.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-oagtmnja-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 3jnztqly; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 730.2,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 1k 730.2 74.6 21.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oagtmnja",
   "source_url": "https://omlx.ai/benchmarks/performance/oagtmnja",
   "tg_tokens": null,
   "tps": 74.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-oagtmnja-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: 3jnztqly; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 802.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 802.0 72.7 23.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oagtmnja",
   "source_url": "https://omlx.ai/benchmarks/performance/oagtmnja",
   "tg_tokens": null,
   "tps": 72.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-zhjq882e-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 4hg6buwv; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 680.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 1k 680.6 78.7 20.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zhjq882e",
   "source_url": "https://omlx.ai/benchmarks/performance/zhjq882e",
   "tg_tokens": null,
   "tps": 78.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-zhjq882e-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 4hg6buwv; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 767.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 767.0 74.4 22.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zhjq882e",
   "source_url": "https://omlx.ai/benchmarks/performance/zhjq882e",
   "tg_tokens": null,
   "tps": 74.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-s8l3plpm-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 51qbe3h0; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 115.1,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 1k 115.1 16.7 20.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s8l3plpm",
   "source_url": "https://omlx.ai/benchmarks/performance/s8l3plpm",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-s8l3plpm-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 51qbe3h0; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 108.4,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 108.4 16.4 23.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s8l3plpm",
   "source_url": "https://omlx.ai/benchmarks/performance/s8l3plpm",
   "tg_tokens": null,
   "tps": 16.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-a90j2rg5-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 54.6 GB at this context; duplicate traces of the same session: 2zgi4st2, nuf4s2bg, w6dz6rjz, x103zjy3; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 294.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-mlx-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 1k 294.3 42.7 54.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a90j2rg5",
   "source_url": "https://omlx.ai/benchmarks/performance/a90j2rg5",
   "tg_tokens": null,
   "tps": 42.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-a90j2rg5-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 55.3 GB at this context; duplicate traces of the same session: 2zgi4st2, nuf4s2bg, w6dz6rjz, x103zjy3; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 318.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-mlx-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 4k 318.6 41.2 55.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a90j2rg5",
   "source_url": "https://omlx.ai/benchmarks/performance/a90j2rg5",
   "tg_tokens": null,
   "tps": 41.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-a90j2rg5-ctx8192",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 55.5 GB at this context; duplicate traces of the same session: 2zgi4st2, nuf4s2bg, w6dz6rjz, x103zjy3; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 542.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-mlx-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 8k 542.4 35.2 55.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a90j2rg5",
   "source_url": "https://omlx.ai/benchmarks/performance/a90j2rg5",
   "tg_tokens": null,
   "tps": 35.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-a90j2rg5-ctx16384",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 56.0 GB at this context; duplicate traces of the same session: 2zgi4st2, nuf4s2bg, w6dz6rjz, x103zjy3; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 514.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-mlx-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 16k 514.8 42.0 56.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a90j2rg5",
   "source_url": "https://omlx.ai/benchmarks/performance/a90j2rg5",
   "tg_tokens": null,
   "tps": 42.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-a90j2rg5-ctx32768",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 57.3 GB at this context; duplicate traces of the same session: 2zgi4st2, nuf4s2bg, w6dz6rjz, x103zjy3; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 529.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-mlx-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 32k 529.7 32.2 57.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a90j2rg5",
   "source_url": "https://omlx.ai/benchmarks/performance/a90j2rg5",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-l8mj8gdo-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: udvq656y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 551.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 551.7 59.0 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l8mj8gdo",
   "source_url": "https://omlx.ai/benchmarks/performance/l8mj8gdo",
   "tg_tokens": null,
   "tps": 59.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-l8mj8gdo-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: udvq656y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 805.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 805.9 51.2 71.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l8mj8gdo",
   "source_url": "https://omlx.ai/benchmarks/performance/l8mj8gdo",
   "tg_tokens": null,
   "tps": 51.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-2ld6cb01-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: izdsj58j; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 131.3 15.4 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ld6cb01",
   "source_url": "https://omlx.ai/benchmarks/performance/2ld6cb01",
   "tg_tokens": null,
   "tps": 15.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-2ld6cb01-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: izdsj58j; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 137.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 137.7 15.1 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ld6cb01",
   "source_url": "https://omlx.ai/benchmarks/performance/2ld6cb01",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-pg443ago-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 54.6 GB at this context; duplicate traces of the same session: fgk29m1f, fuc12gbb, jqr4aw2z, kzz4n5vz, s8zf5byz, vtf9u0ql; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 393.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-mlx-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 1k 393.1 47.8 54.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pg443ago",
   "source_url": "https://omlx.ai/benchmarks/performance/pg443ago",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-pg443ago-ctx8192",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 55.5 GB at this context; duplicate traces of the same session: fgk29m1f, fuc12gbb, jqr4aw2z, kzz4n5vz, s8zf5byz, vtf9u0ql; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 543.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-mlx-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 8k 543.7 38.7 55.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pg443ago",
   "source_url": "https://omlx.ai/benchmarks/performance/pg443ago",
   "tg_tokens": null,
   "tps": 38.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-pg443ago-ctx16384",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 56.0 GB at this context; duplicate traces of the same session: fgk29m1f, fuc12gbb, jqr4aw2z, kzz4n5vz, s8zf5byz, vtf9u0ql; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 513.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-mlx-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 16k 513.1 48.8 56.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pg443ago",
   "source_url": "https://omlx.ai/benchmarks/performance/pg443ago",
   "tg_tokens": null,
   "tps": 48.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-pg443ago-ctx32768",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 57.4 GB at this context; duplicate traces of the same session: fgk29m1f, fuc12gbb, jqr4aw2z, kzz4n5vz, s8zf5byz, vtf9u0ql; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 524.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-mlx-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 32k 524.6 50.6 57.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pg443ago",
   "source_url": "https://omlx.ai/benchmarks/performance/pg443ago",
   "tg_tokens": null,
   "tps": 50.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-pg443ago-ctx65536",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 60.1 GB at this context; duplicate traces of the same session: fgk29m1f, fuc12gbb, jqr4aw2z, kzz4n5vz, s8zf5byz, vtf9u0ql; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 520.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-mlx-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 64k 520.2 39.1 60.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pg443ago",
   "source_url": "https://omlx.ai/benchmarks/performance/pg443ago",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-pg443ago-ctx131072",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 66.3 GB at this context; duplicate traces of the same session: fgk29m1f, fuc12gbb, jqr4aw2z, kzz4n5vz, s8zf5byz, vtf9u0ql; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 479.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-mlx-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 128k 479.9 35.7 66.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pg443ago",
   "source_url": "https://omlx.ai/benchmarks/performance/pg443ago",
   "tg_tokens": null,
   "tps": 35.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-pg443ago-ctx199680",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.3 GB at this context; duplicate traces of the same session: fgk29m1f, fuc12gbb, jqr4aw2z, kzz4n5vz, s8zf5byz, vtf9u0ql; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 439.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-mlx-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 195k 439.5 27.7 69.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pg443ago",
   "source_url": "https://omlx.ai/benchmarks/performance/pg443ago",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ic0v8mcf-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 75.0 GB; TTFT 44408 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2952.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 128k tokens 2,952 PP tok/s 76.5 TG tok/s 44408 TTFT (ms) 75.0 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ic0v8mcf",
   "source_url": "https://omlx.ai/benchmarks/performance/ic0v8mcf",
   "tg_tokens": null,
   "tps": 76.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ic9gupg2-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: hcmtzou1, s81o7ja1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1198.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,198 78.5 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ic9gupg2",
   "source_url": "https://omlx.ai/benchmarks/performance/ic9gupg2",
   "tg_tokens": null,
   "tps": 78.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ic9gupg2-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: hcmtzou1, s81o7ja1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2100.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,100 92.0 70.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ic9gupg2",
   "source_url": "https://omlx.ai/benchmarks/performance/ic9gupg2",
   "tg_tokens": null,
   "tps": 92.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ic9gupg2-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: hcmtzou1, s81o7ja1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2674.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,674 86.8 73.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ic9gupg2",
   "source_url": "https://omlx.ai/benchmarks/performance/ic9gupg2",
   "tg_tokens": null,
   "tps": 86.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-d9fkdogu-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "QAT OptiQ build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: dz2ryoc6; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 679.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-OptiQ-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 1k 679.3 57.5 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d9fkdogu",
   "source_url": "https://omlx.ai/benchmarks/performance/d9fkdogu",
   "tg_tokens": null,
   "tps": 57.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-d9fkdogu-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "QAT OptiQ build; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: dz2ryoc6; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 654.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-OptiQ-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 654.6 58.0 20.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d9fkdogu",
   "source_url": "https://omlx.ai/benchmarks/performance/d9fkdogu",
   "tg_tokens": null,
   "tps": 58.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-yes5u53t-ctx1024",
   "model": "Ling-3.0-tiny",
   "notes": "oMLX community benchmark trace; peak memory 76.2 GB at this context; duplicate traces of the same session: 4i2elrfz, k7a5r8z5, o0c8qd0v, r36vjszt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 199.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-tiny-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 199.9 88.3 76.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yes5u53t",
   "source_url": "https://omlx.ai/benchmarks/performance/yes5u53t",
   "tg_tokens": null,
   "tps": 88.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-yes5u53t-ctx4096",
   "model": "Ling-3.0-tiny",
   "notes": "oMLX community benchmark trace; peak memory 77.4 GB at this context; duplicate traces of the same session: 4i2elrfz, k7a5r8z5, o0c8qd0v, r36vjszt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1024.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-tiny-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,024 183.8 77.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yes5u53t",
   "source_url": "https://omlx.ai/benchmarks/performance/yes5u53t",
   "tg_tokens": null,
   "tps": 183.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-yes5u53t-ctx8192",
   "model": "Ling-3.0-tiny",
   "notes": "oMLX community benchmark trace; peak memory 78.0 GB at this context; duplicate traces of the same session: 4i2elrfz, k7a5r8z5, o0c8qd0v, r36vjszt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1456.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-tiny-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,456 94.6 78.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yes5u53t",
   "source_url": "https://omlx.ai/benchmarks/performance/yes5u53t",
   "tg_tokens": null,
   "tps": 94.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-yes5u53t-ctx32768",
   "model": "Ling-3.0-tiny",
   "notes": "oMLX community benchmark trace; peak memory 80.2 GB at this context; duplicate traces of the same session: 4i2elrfz, k7a5r8z5, o0c8qd0v, r36vjszt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2792.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-tiny-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,792 121.0 80.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yes5u53t",
   "source_url": "https://omlx.ai/benchmarks/performance/yes5u53t",
   "tg_tokens": null,
   "tps": 121.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yes5u53t-ctx65536",
   "model": "Ling-3.0-tiny",
   "notes": "oMLX community benchmark trace; peak memory 83.8 GB at this context; duplicate traces of the same session: 4i2elrfz, k7a5r8z5, o0c8qd0v, r36vjszt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2160.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-tiny-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,160 43.2 83.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yes5u53t",
   "source_url": "https://omlx.ai/benchmarks/performance/yes5u53t",
   "tg_tokens": null,
   "tps": 43.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-f7s1f0dq-ctx65536",
   "model": "Ling-3.0-tiny",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 10.5 GB; TTFT 13609 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4816.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-tiny-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k tokens 4,816 PP tok/s 156.7 TG tok/s 13609 TTFT (ms) 10.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f7s1f0dq",
   "source_url": "https://omlx.ai/benchmarks/performance/f7s1f0dq",
   "tg_tokens": null,
   "tps": 156.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-gjjg2wqw-ctx199680",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 195k context; peak memory 68.6 GB; TTFT 428800 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 466.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-mlx-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 195k tokens 466.4 PP tok/s 30.7 TG tok/s 428800 TTFT (ms) 68.6 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gjjg2wqw",
   "source_url": "https://omlx.ai/benchmarks/performance/gjjg2wqw",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dpthanep-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: 356sz641, 8geizvzl, rbfhs1dy, v05wm163; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1025.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,025 62.5 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dpthanep",
   "source_url": "https://omlx.ai/benchmarks/performance/dpthanep",
   "tg_tokens": null,
   "tps": 62.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dpthanep-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: 356sz641, 8geizvzl, rbfhs1dy, v05wm163; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2072.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,072 74.5 71.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dpthanep",
   "source_url": "https://omlx.ai/benchmarks/performance/dpthanep",
   "tg_tokens": null,
   "tps": 74.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dpthanep-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 356sz641, 8geizvzl, rbfhs1dy, v05wm163; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2578.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,578 69.5 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dpthanep",
   "source_url": "https://omlx.ai/benchmarks/performance/dpthanep",
   "tg_tokens": null,
   "tps": 69.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dpthanep-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.9 GB at this context; duplicate traces of the same session: 356sz641, 8geizvzl, rbfhs1dy, v05wm163; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2903.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,903 70.4 72.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dpthanep",
   "source_url": "https://omlx.ai/benchmarks/performance/dpthanep",
   "tg_tokens": null,
   "tps": 70.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dpthanep-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.7 GB at this context; duplicate traces of the same session: 356sz641, 8geizvzl, rbfhs1dy, v05wm163; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2810.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,810 71.5 73.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dpthanep",
   "source_url": "https://omlx.ai/benchmarks/performance/dpthanep",
   "tg_tokens": null,
   "tps": 71.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-ag3af95n-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 105.2 GB at this context; duplicate traces of the same session: gmwngy9f; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 547.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 4bit; 1k 547.3 31.6 105.2 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ag3af95n",
   "source_url": "https://omlx.ai/benchmarks/performance/ag3af95n",
   "tg_tokens": null,
   "tps": 31.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-ag3af95n-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 105.6 GB at this context; duplicate traces of the same session: gmwngy9f; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 635.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 4bit; 4k 635.1 31.6 105.6 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ag3af95n",
   "source_url": "https://omlx.ai/benchmarks/performance/ag3af95n",
   "tg_tokens": null,
   "tps": 31.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qxzft5ux-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.2 GB at this context; duplicate traces of the same session: ozbgjiwc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2605.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 2,605 89.8 26.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qxzft5ux",
   "source_url": "https://omlx.ai/benchmarks/performance/qxzft5ux",
   "tg_tokens": null,
   "tps": 89.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qxzft5ux-ctx16384",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: ozbgjiwc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3602.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 16k 3,602 79.7 26.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qxzft5ux",
   "source_url": "https://omlx.ai/benchmarks/performance/qxzft5ux",
   "tg_tokens": null,
   "tps": 79.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-xq64l77z-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 14.5 GB at this context; duplicate traces of the same session: feih0xwu; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 673.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 1k 673.8 59.4 14.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xq64l77z",
   "source_url": "https://omlx.ai/benchmarks/performance/xq64l77z",
   "tg_tokens": null,
   "tps": 59.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-xq64l77z-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: feih0xwu; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 671.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 671.0 59.7 14.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xq64l77z",
   "source_url": "https://omlx.ai/benchmarks/performance/xq64l77z",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-oowqe8yt-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "QAT OptiQ build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: i11hh0qb; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 641.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-OptiQ-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 1k 641.9 58.4 20.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oowqe8yt",
   "source_url": "https://omlx.ai/benchmarks/performance/oowqe8yt",
   "tg_tokens": null,
   "tps": 58.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-oowqe8yt-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "QAT OptiQ build; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: i11hh0qb; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 669.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-OptiQ-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 669.3 58.0 20.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oowqe8yt",
   "source_url": "https://omlx.ai/benchmarks/performance/oowqe8yt",
   "tg_tokens": null,
   "tps": 58.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-5t3srjh9-ctx1024",
   "model": "FrogNano-4B-2609",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 5.2 GB at this context; duplicate traces of the same session: wgatlydg; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 394.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "FrogNano-4B-2609-oQ6e-mtp-MLX on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 6bit; 1k 394.5 43.4 5.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5t3srjh9",
   "source_url": "https://omlx.ai/benchmarks/performance/5t3srjh9",
   "tg_tokens": null,
   "tps": 43.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-5t3srjh9-ctx4096",
   "model": "FrogNano-4B-2609",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 5.5 GB at this context; duplicate traces of the same session: wgatlydg; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 540.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "FrogNano-4B-2609-oQ6e-mtp-MLX on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 6bit; 4k 540.1 40.4 5.5 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5t3srjh9",
   "source_url": "https://omlx.ai/benchmarks/performance/5t3srjh9",
   "tg_tokens": null,
   "tps": 40.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-beu021zq-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; Jundot build; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 64d0v1dm, qhshm8ya; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 826.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 826.6 63.2 20.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/beu021zq",
   "source_url": "https://omlx.ai/benchmarks/performance/beu021zq",
   "tg_tokens": null,
   "tps": 63.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-beu021zq-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; Jundot build; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 64d0v1dm, qhshm8ya; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 758.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 758.8 58.9 20.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/beu021zq",
   "source_url": "https://omlx.ai/benchmarks/performance/beu021zq",
   "tg_tokens": null,
   "tps": 58.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-beu021zq-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; Jundot build; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 64d0v1dm, qhshm8ya; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 685.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 685.7 48.8 21.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/beu021zq",
   "source_url": "https://omlx.ai/benchmarks/performance/beu021zq",
   "tg_tokens": null,
   "tps": 48.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fbq4koaw-ctx1024",
   "model": "LLama-3.1-128k-Darkest-Planet-Uncensored-16.5B",
   "notes": "oMLX community benchmark trace; peak memory 15.4 GB at this context; duplicate traces of the same session: nbhtipj4; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 498.7,
   "provenance": "community",
   "quant": "q5",
   "quote": "LLama-3.1-128k-Darkest-Planet-Uncensored-16.5B-q5-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 5bit; 1k 498.7 17.8 15.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fbq4koaw",
   "source_url": "https://omlx.ai/benchmarks/performance/fbq4koaw",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fbq4koaw-ctx4096",
   "model": "LLama-3.1-128k-Darkest-Planet-Uncensored-16.5B",
   "notes": "oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: nbhtipj4; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 494.8,
   "provenance": "community",
   "quant": "q5",
   "quote": "LLama-3.1-128k-Darkest-Planet-Uncensored-16.5B-q5-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 5bit; 4k 494.8 16.1 16.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fbq4koaw",
   "source_url": "https://omlx.ai/benchmarks/performance/fbq4koaw",
   "tg_tokens": null,
   "tps": 16.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-kqvasaku-ctx1024",
   "model": "Ornith-1.5-9B-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 5.5 GB at this context; duplicate traces of the same session: k0utbacp; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 711.1,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Ornith-1.5-9B-Abliterated-MLX-mxfp4 on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 711.1 29.0 5.5 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kqvasaku",
   "source_url": "https://omlx.ai/benchmarks/performance/kqvasaku",
   "tg_tokens": null,
   "tps": 29.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-kqvasaku-ctx4096",
   "model": "Ornith-1.5-9B-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: k0utbacp; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 780.9,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Ornith-1.5-9B-Abliterated-MLX-mxfp4 on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 780.9 27.9 5.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kqvasaku",
   "source_url": "https://omlx.ai/benchmarks/performance/kqvasaku",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-z6e26w1p-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: dtjzsnk2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1437.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 8bit; 1k 1,437 82.2 35.5 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z6e26w1p",
   "source_url": "https://omlx.ai/benchmarks/performance/z6e26w1p",
   "tg_tokens": null,
   "tps": 82.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-z6e26w1p-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: dtjzsnk2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1613.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 8bit; 4k 1,613 78.8 37.4 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z6e26w1p",
   "source_url": "https://omlx.ai/benchmarks/performance/z6e26w1p",
   "tg_tokens": null,
   "tps": 78.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-fdev1ie5-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: svmlv1hn; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 775.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 1k 775.9 87.4 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fdev1ie5",
   "source_url": "https://omlx.ai/benchmarks/performance/fdev1ie5",
   "tg_tokens": null,
   "tps": 87.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-fdev1ie5-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: svmlv1hn; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 779.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 779.3 84.8 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fdev1ie5",
   "source_url": "https://omlx.ai/benchmarks/performance/fdev1ie5",
   "tg_tokens": null,
   "tps": 84.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-rjj3ri8h-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: 1mdldl3b, 2ggergt0, abnyalq0; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 580.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 580.5 60.0 22.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rjj3ri8h",
   "source_url": "https://omlx.ai/benchmarks/performance/rjj3ri8h",
   "tg_tokens": null,
   "tps": 60.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-rjj3ri8h-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 1mdldl3b, 2ggergt0, abnyalq0; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 456.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k 456.0 51.6 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rjj3ri8h",
   "source_url": "https://omlx.ai/benchmarks/performance/rjj3ri8h",
   "tg_tokens": null,
   "tps": 51.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-rjj3ri8h-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: 1mdldl3b, 2ggergt0, abnyalq0; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 246.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 64k 246.0 37.0 24.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rjj3ri8h",
   "source_url": "https://omlx.ai/benchmarks/performance/rjj3ri8h",
   "tg_tokens": null,
   "tps": 37.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-rjj3ri8h-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: 1mdldl3b, 2ggergt0, abnyalq0; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 152.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 128k 152.6 27.0 26.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rjj3ri8h",
   "source_url": "https://omlx.ai/benchmarks/performance/rjj3ri8h",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-nmjevy7a-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: fknm1g78; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 753.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 753.8 101.6 21.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nmjevy7a",
   "source_url": "https://omlx.ai/benchmarks/performance/nmjevy7a",
   "tg_tokens": null,
   "tps": 101.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-nmjevy7a-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: fknm1g78; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 856.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 856.0 90.5 23.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nmjevy7a",
   "source_url": "https://omlx.ai/benchmarks/performance/nmjevy7a",
   "tg_tokens": null,
   "tps": 90.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-2otsao3v-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: ukzrgjw3; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 736.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 1k 736.9 89.2 37.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2otsao3v",
   "source_url": "https://omlx.ai/benchmarks/performance/2otsao3v",
   "tg_tokens": null,
   "tps": 89.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-2otsao3v-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.4 GB at this context; duplicate traces of the same session: ukzrgjw3; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 834.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 4k 834.6 87.8 39.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2otsao3v",
   "source_url": "https://omlx.ai/benchmarks/performance/2otsao3v",
   "tg_tokens": null,
   "tps": 87.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-w5dmpztc-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: rzxta3px; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 113.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 1k 113.4 14.7 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w5dmpztc",
   "source_url": "https://omlx.ai/benchmarks/performance/w5dmpztc",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-w5dmpztc-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: rzxta3px; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 105.8 13.4 19.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w5dmpztc",
   "source_url": "https://omlx.ai/benchmarks/performance/w5dmpztc",
   "tg_tokens": null,
   "tps": 13.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-d1xgu27g-ctx1024",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: h0cwuqnb; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 55.9,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 2bit; 1k 55.9 8.8 11.5 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d1xgu27g",
   "source_url": "https://omlx.ai/benchmarks/performance/d1xgu27g",
   "tg_tokens": null,
   "tps": 8.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-d1xgu27g-ctx4096",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 14.0 GB at this context; duplicate traces of the same session: h0cwuqnb; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 55.6,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 2bit; 4k 55.6 8.1 14.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d1xgu27g",
   "source_url": "https://omlx.ai/benchmarks/performance/d1xgu27g",
   "tg_tokens": null,
   "tps": 8.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-kw6yb15q-ctx1024",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: mq0my58s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 523.9,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 2bit; 1k 523.9 44.3 11.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kw6yb15q",
   "source_url": "https://omlx.ai/benchmarks/performance/kw6yb15q",
   "tg_tokens": null,
   "tps": 44.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-kw6yb15q-ctx4096",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 14.2 GB at this context; duplicate traces of the same session: mq0my58s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 691.2,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 2bit; 4k 691.2 42.8 14.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kw6yb15q",
   "source_url": "https://omlx.ai/benchmarks/performance/kw6yb15q",
   "tg_tokens": null,
   "tps": 42.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-80chektz-ctx1024",
   "model": "Mistral-Nemo-2407-12B-Thinking-Claude-Gemini-GPT5.2-Uncensored-HERETIC",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 4iwevpr1; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 544.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Mistral-Nemo-2407-12B-Thinking-Claude-Gemini-GPT5.2-Uncensored-HERETIC on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 unknown; 1k 544.0 22.2 23.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/80chektz",
   "source_url": "https://omlx.ai/benchmarks/performance/80chektz",
   "tg_tokens": null,
   "tps": 22.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-80chektz-ctx4096",
   "model": "Mistral-Nemo-2407-12B-Thinking-Claude-Gemini-GPT5.2-Uncensored-HERETIC",
   "notes": "oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: 4iwevpr1; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 553.5,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Mistral-Nemo-2407-12B-Thinking-Claude-Gemini-GPT5.2-Uncensored-HERETIC on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 unknown; 4k 553.5 21.6 24.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/80chektz",
   "source_url": "https://omlx.ai/benchmarks/performance/80chektz",
   "tg_tokens": null,
   "tps": 21.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-p2ezgdmq-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 6.7 GB; TTFT 11236 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 364.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k tokens 364.6 PP tok/s 44.6 TG tok/s 11236 TTFT (ms) 6.7 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p2ezgdmq",
   "source_url": "https://omlx.ai/benchmarks/performance/p2ezgdmq",
   "tg_tokens": null,
   "tps": 44.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-kvcnx2jb-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 8.3 GB; TTFT 14235 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 287.7,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.5-9B-OptiQ-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k tokens 287.7 PP tok/s 41.6 TG tok/s 14235 TTFT (ms) 8.3 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kvcnx2jb",
   "source_url": "https://omlx.ai/benchmarks/performance/kvcnx2jb",
   "tg_tokens": null,
   "tps": 41.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mxahjyuq-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-oQ5e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: ldrp10ls; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 847.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 5bit; 1k 847.2 49.4 23.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mxahjyuq",
   "source_url": "https://omlx.ai/benchmarks/performance/mxahjyuq",
   "tg_tokens": null,
   "tps": 49.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mxahjyuq-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-oQ5e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: ldrp10ls; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 935.5,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 5bit; 4k 935.5 48.9 24.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mxahjyuq",
   "source_url": "https://omlx.ai/benchmarks/performance/mxahjyuq",
   "tg_tokens": null,
   "tps": 48.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-0sdax319-ctx4096",
   "model": "Muse Glimmer",
   "notes": "30B build; oMLX community benchmark trace, single run at 4k context; peak memory 14.6 GB; TTFT 89564 ms; duplicate traces of the same session: none; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 45.7,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Muse-Glimmer-30B-MLX-MXFP4 on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 4bit; 4k tokens 45.7 PP tok/s 8.8 TG tok/s 89564 TTFT (ms) 14.6 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0sdax319",
   "source_url": "https://omlx.ai/benchmarks/performance/0sdax319",
   "tg_tokens": null,
   "tps": 8.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-dh6v91jj-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 6.7 GB; TTFT 11258 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 363.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k tokens 363.8 PP tok/s 44.2 TG tok/s 11258 TTFT (ms) 6.7 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dh6v91jj",
   "source_url": "https://omlx.ai/benchmarks/performance/dh6v91jj",
   "tg_tokens": null,
   "tps": 44.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-nfbsay56-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 7b99dj00; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 113.8,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 113.8 25.8 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nfbsay56",
   "source_url": "https://omlx.ai/benchmarks/performance/nfbsay56",
   "tg_tokens": null,
   "tps": 25.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-nfbsay56-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: 7b99dj00; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 32k 104.0 25.7 26.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nfbsay56",
   "source_url": "https://omlx.ai/benchmarks/performance/nfbsay56",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dit6i4gc-ctx1024",
   "model": "Ornith-1.5-35B-MLX-VL",
   "notes": "oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: 7ge8xq67, jfo3ghin, vg6cjryn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2009.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-MLX-VL-oQ6e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 2,009 107.6 27.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dit6i4gc",
   "source_url": "https://omlx.ai/benchmarks/performance/dit6i4gc",
   "tg_tokens": null,
   "tps": 107.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dit6i4gc-ctx8192",
   "model": "Ornith-1.5-35B-MLX-VL",
   "notes": "oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: 7ge8xq67, jfo3ghin, vg6cjryn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4847.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-MLX-VL-oQ6e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 4,847 104.4 28.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dit6i4gc",
   "source_url": "https://omlx.ai/benchmarks/performance/dit6i4gc",
   "tg_tokens": null,
   "tps": 104.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dit6i4gc-ctx16384",
   "model": "Ornith-1.5-35B-MLX-VL",
   "notes": "oMLX community benchmark trace; peak memory 28.7 GB at this context; duplicate traces of the same session: 7ge8xq67, jfo3ghin, vg6cjryn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4474.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-MLX-VL-oQ6e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 4,474 101.0 28.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dit6i4gc",
   "source_url": "https://omlx.ai/benchmarks/performance/dit6i4gc",
   "tg_tokens": null,
   "tps": 101.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dit6i4gc-ctx32768",
   "model": "Ornith-1.5-35B-MLX-VL",
   "notes": "oMLX community benchmark trace; peak memory 29.0 GB at this context; duplicate traces of the same session: 7ge8xq67, jfo3ghin, vg6cjryn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3762.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-MLX-VL-oQ6e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 3,762 92.5 29.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dit6i4gc",
   "source_url": "https://omlx.ai/benchmarks/performance/dit6i4gc",
   "tg_tokens": null,
   "tps": 92.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-deojwjf9-ctx1024",
   "model": "Ornith-1.5-35B-MLX-VL",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: n7a6atxg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4046.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-MLX-VL-oQ8e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 4,046 102.6 35.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/deojwjf9",
   "source_url": "https://omlx.ai/benchmarks/performance/deojwjf9",
   "tg_tokens": null,
   "tps": 102.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-deojwjf9-ctx32768",
   "model": "Ornith-1.5-35B-MLX-VL",
   "notes": "oMLX community benchmark trace; peak memory 36.7 GB at this context; duplicate traces of the same session: n7a6atxg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4322.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-MLX-VL-oQ8e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 4,322 87.8 36.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/deojwjf9",
   "source_url": "https://omlx.ai/benchmarks/performance/deojwjf9",
   "tg_tokens": null,
   "tps": 87.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-pxog0f5y-ctx1024",
   "model": "Ornith-1.5-35B-MLX-VL",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: t7u5ca6o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 720.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-MLX-VL-oQ8e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 720.8 104.3 35.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pxog0f5y",
   "source_url": "https://omlx.ai/benchmarks/performance/pxog0f5y",
   "tg_tokens": null,
   "tps": 104.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-pxog0f5y-ctx4096",
   "model": "Ornith-1.5-35B-MLX-VL",
   "notes": "oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: t7u5ca6o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2150.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-MLX-VL-oQ8e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 2,150 102.4 36.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pxog0f5y",
   "source_url": "https://omlx.ai/benchmarks/performance/pxog0f5y",
   "tg_tokens": null,
   "tps": 102.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-p7otop80-ctx1024",
   "model": "Ornith-Agents-A1-3.7-35B-A3B",
   "notes": "dare_ties_v4 build; fp16 weights; oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: cl8ck632, tu8wz7we; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 829.1,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Ornith-Agents-A1-3.7-35B-A3B-dare_ties_v4-oQ6-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 1k 829.1 47.3 28.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p7otop80",
   "source_url": "https://omlx.ai/benchmarks/performance/p7otop80",
   "tg_tokens": null,
   "tps": 47.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-p7otop80-ctx4096",
   "model": "Ornith-Agents-A1-3.7-35B-A3B",
   "notes": "dare_ties_v4 build; fp16 weights; oMLX community benchmark trace; peak memory 28.9 GB at this context; duplicate traces of the same session: cl8ck632, tu8wz7we; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 889.9,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Ornith-Agents-A1-3.7-35B-A3B-dare_ties_v4-oQ6-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 4k 889.9 52.0 28.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p7otop80",
   "source_url": "https://omlx.ai/benchmarks/performance/p7otop80",
   "tg_tokens": null,
   "tps": 52.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-p7otop80-ctx16384",
   "model": "Ornith-Agents-A1-3.7-35B-A3B",
   "notes": "dare_ties_v4 build; fp16 weights; oMLX community benchmark trace; peak memory 29.4 GB at this context; duplicate traces of the same session: cl8ck632, tu8wz7we; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 780.1,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Ornith-Agents-A1-3.7-35B-A3B-dare_ties_v4-oQ6-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 16k 780.1 28.3 29.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p7otop80",
   "source_url": "https://omlx.ai/benchmarks/performance/p7otop80",
   "tg_tokens": null,
   "tps": 28.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-jvxn0zdy-ctx4096",
   "model": "Ornith-Agents-A1-3.7-35B-A3B",
   "notes": "dare_ties_v4 build; fp16 weights; oMLX community benchmark trace; peak memory 28.9 GB at this context; duplicate traces of the same session: 07kmwnxh; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 927.6,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Ornith-Agents-A1-3.7-35B-A3B-dare_ties_v4-oQ6-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 4k 927.6 38.1 28.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jvxn0zdy",
   "source_url": "https://omlx.ai/benchmarks/performance/jvxn0zdy",
   "tg_tokens": null,
   "tps": 38.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-jvxn0zdy-ctx16384",
   "model": "Ornith-Agents-A1-3.7-35B-A3B",
   "notes": "dare_ties_v4 build; fp16 weights; oMLX community benchmark trace; peak memory 29.4 GB at this context; duplicate traces of the same session: 07kmwnxh; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 780.2,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Ornith-Agents-A1-3.7-35B-A3B-dare_ties_v4-oQ6-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 16k 780.2 32.2 29.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jvxn0zdy",
   "source_url": "https://omlx.ai/benchmarks/performance/jvxn0zdy",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-zz2e7cgq-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 6.7 GB; TTFT 11337 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 361.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k tokens 361.3 PP tok/s 44.2 TG tok/s 11337 TTFT (ms) 6.7 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zz2e7cgq",
   "source_url": "https://omlx.ai/benchmarks/performance/zz2e7cgq",
   "tg_tokens": null,
   "tps": 44.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-lyva7y7z-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: b3ayplay; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 404.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 1k 404.6 62.3 5.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lyva7y7z",
   "source_url": "https://omlx.ai/benchmarks/performance/lyva7y7z",
   "tg_tokens": null,
   "tps": 62.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-lyva7y7z-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: b3ayplay; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 388.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 4k 388.2 58.7 6.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lyva7y7z",
   "source_url": "https://omlx.ai/benchmarks/performance/lyva7y7z",
   "tg_tokens": null,
   "tps": 58.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-kw48u2y8-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 8.4 GB; TTFT 14159 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 289.3,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.5-9B-OptiQ-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k tokens 289.3 PP tok/s 41.7 TG tok/s 14159 TTFT (ms) 8.4 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kw48u2y8",
   "source_url": "https://omlx.ai/benchmarks/performance/kw48u2y8",
   "tg_tokens": null,
   "tps": 41.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-aq9zmwy1-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 6qo8tl56; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 462.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 462.9 54.7 21.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aq9zmwy1",
   "source_url": "https://omlx.ai/benchmarks/performance/aq9zmwy1",
   "tg_tokens": null,
   "tps": 54.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-aq9zmwy1-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: 6qo8tl56; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 403.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 403.3 46.9 22.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aq9zmwy1",
   "source_url": "https://omlx.ai/benchmarks/performance/aq9zmwy1",
   "tg_tokens": null,
   "tps": 46.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-q2g83y37-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 36.0 GB at this context; duplicate traces of the same session: 8d694l8g, faxx660a, sjxpswuz; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1311.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 16k 1,311 47.6 36.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q2g83y37",
   "source_url": "https://omlx.ai/benchmarks/performance/q2g83y37",
   "tg_tokens": null,
   "tps": 47.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-q2g83y37-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 36.8 GB at this context; duplicate traces of the same session: 8d694l8g, faxx660a, sjxpswuz; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1247.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 32k 1,247 43.7 36.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q2g83y37",
   "source_url": "https://omlx.ai/benchmarks/performance/q2g83y37",
   "tg_tokens": null,
   "tps": 43.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-q2g83y37-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 39.2 GB at this context; duplicate traces of the same session: 8d694l8g, faxx660a, sjxpswuz; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1072.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 64k 1,072 43.9 39.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q2g83y37",
   "source_url": "https://omlx.ai/benchmarks/performance/q2g83y37",
   "tg_tokens": null,
   "tps": 43.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-q2g83y37-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 58.3 GB at this context; duplicate traces of the same session: 8d694l8g, faxx660a, sjxpswuz; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 839.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 128k 839.8 10.3 58.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q2g83y37",
   "source_url": "https://omlx.ai/benchmarks/performance/q2g83y37",
   "tg_tokens": null,
   "tps": 10.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-al32fje6-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.5 GB at this context; duplicate traces of the same session: 5k2u70lv, hogszq4o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 882.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 882.8 51.9 17.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/al32fje6",
   "source_url": "https://omlx.ai/benchmarks/performance/al32fje6",
   "tg_tokens": null,
   "tps": 51.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-al32fje6-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 5k2u70lv, hogszq4o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 986.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 986.4 64.0 18.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/al32fje6",
   "source_url": "https://omlx.ai/benchmarks/performance/al32fje6",
   "tg_tokens": null,
   "tps": 64.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-al32fje6-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 5k2u70lv, hogszq4o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 936.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 936.9 75.0 19.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/al32fje6",
   "source_url": "https://omlx.ai/benchmarks/performance/al32fje6",
   "tg_tokens": null,
   "tps": 75.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-pk5sraxr-ctx1024",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: 3o97gsnc; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 85.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-31B-it-oQ8e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 1k 85.3 17.6 33.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pk5sraxr",
   "source_url": "https://omlx.ai/benchmarks/performance/pk5sraxr",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-pk5sraxr-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.0 GB at this context; duplicate traces of the same session: 3o97gsnc; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 83.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-31B-it-oQ8e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 4k 83.7 13.2 35.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pk5sraxr",
   "source_url": "https://omlx.ai/benchmarks/performance/pk5sraxr",
   "tg_tokens": null,
   "tps": 13.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8yl6c0en-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 78.5 GB at this context; duplicate traces of the same session: hcfsddpv, j8e2oqcc, w8i6k3j4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 56.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 56.7 5.8 78.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8yl6c0en",
   "source_url": "https://omlx.ai/benchmarks/performance/8yl6c0en",
   "tg_tokens": null,
   "tps": 5.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8yl6c0en-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 79.1 GB at this context; duplicate traces of the same session: hcfsddpv, j8e2oqcc, w8i6k3j4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 180.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 180.0 5.1 79.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8yl6c0en",
   "source_url": "https://omlx.ai/benchmarks/performance/8yl6c0en",
   "tg_tokens": null,
   "tps": 5.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8yl6c0en-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 79.5 GB at this context; duplicate traces of the same session: hcfsddpv, j8e2oqcc, w8i6k3j4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 173.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 173.7 4.6 79.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8yl6c0en",
   "source_url": "https://omlx.ai/benchmarks/performance/8yl6c0en",
   "tg_tokens": null,
   "tps": 4.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8yl6c0en-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 79.6 GB at this context; duplicate traces of the same session: hcfsddpv, j8e2oqcc, w8i6k3j4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 171.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 171.8 4.2 79.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8yl6c0en",
   "source_url": "https://omlx.ai/benchmarks/performance/8yl6c0en",
   "tg_tokens": null,
   "tps": 4.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-llr9zaxe-ctx1024",
   "model": "Llama-3-Groq-8B-Tool-Use",
   "notes": "oMLX community benchmark trace; peak memory 5.0 GB at this context; duplicate traces of the same session: qnkk88mt; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 376.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Llama-3-Groq-8B-Tool-Use-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 376.5 49.6 5.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/llr9zaxe",
   "source_url": "https://omlx.ai/benchmarks/performance/llr9zaxe",
   "tg_tokens": null,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-llr9zaxe-ctx4096",
   "model": "Llama-3-Groq-8B-Tool-Use",
   "notes": "oMLX community benchmark trace; peak memory 5.5 GB at this context; duplicate traces of the same session: qnkk88mt; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 352.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Llama-3-Groq-8B-Tool-Use-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 352.2 44.2 5.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/llr9zaxe",
   "source_url": "https://omlx.ai/benchmarks/performance/llr9zaxe",
   "tg_tokens": null,
   "tps": 44.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-223zq21n-ctx1024",
   "model": "Ornith-1.5-35B-A3B-Vision-MTPLX",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 385tbmui, ilzq1c2s, vgr6rmcm; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1876.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-Vision-MTPLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 1,876 198.9 21.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/223zq21n",
   "source_url": "https://omlx.ai/benchmarks/performance/223zq21n",
   "tg_tokens": null,
   "tps": 198.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-223zq21n-ctx4096",
   "model": "Ornith-1.5-35B-A3B-Vision-MTPLX",
   "notes": "oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: 385tbmui, ilzq1c2s, vgr6rmcm; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2580.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-Vision-MTPLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 2,580 170.3 23.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/223zq21n",
   "source_url": "https://omlx.ai/benchmarks/performance/223zq21n",
   "tg_tokens": null,
   "tps": 170.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-223zq21n-ctx8192",
   "model": "Ornith-1.5-35B-A3B-Vision-MTPLX",
   "notes": "oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: 385tbmui, ilzq1c2s, vgr6rmcm; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2485.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-Vision-MTPLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 2,485 182.7 23.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/223zq21n",
   "source_url": "https://omlx.ai/benchmarks/performance/223zq21n",
   "tg_tokens": null,
   "tps": 182.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-223zq21n-ctx16384",
   "model": "Ornith-1.5-35B-A3B-Vision-MTPLX",
   "notes": "oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 385tbmui, ilzq1c2s, vgr6rmcm; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2222.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-Vision-MTPLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 16k 2,222 193.2 23.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/223zq21n",
   "source_url": "https://omlx.ai/benchmarks/performance/223zq21n",
   "tg_tokens": null,
   "tps": 193.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-223zq21n-ctx32768",
   "model": "Ornith-1.5-35B-A3B-Vision-MTPLX",
   "notes": "oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: 385tbmui, ilzq1c2s, vgr6rmcm; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1808.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-Vision-MTPLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 1,808 175.3 24.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/223zq21n",
   "source_url": "https://omlx.ai/benchmarks/performance/223zq21n",
   "tg_tokens": null,
   "tps": 175.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-swrf00tm-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: ml1z817l; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 848.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 1k 848.1 54.2 28.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/swrf00tm",
   "source_url": "https://omlx.ai/benchmarks/performance/swrf00tm",
   "tg_tokens": null,
   "tps": 54.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-swrf00tm-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.4 GB at this context; duplicate traces of the same session: ml1z817l; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 780.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 16k 780.5 29.0 29.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/swrf00tm",
   "source_url": "https://omlx.ai/benchmarks/performance/swrf00tm",
   "tg_tokens": null,
   "tps": 29.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-4whhdgg5-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: 74twfeje, tsbkz5yx; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 839.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 1k 839.6 48.6 28.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4whhdgg5",
   "source_url": "https://omlx.ai/benchmarks/performance/4whhdgg5",
   "tg_tokens": null,
   "tps": 48.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-4whhdgg5-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.0 GB at this context; duplicate traces of the same session: 74twfeje, tsbkz5yx; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 877.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 4k 877.2 31.6 29.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4whhdgg5",
   "source_url": "https://omlx.ai/benchmarks/performance/4whhdgg5",
   "tg_tokens": null,
   "tps": 31.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-4whhdgg5-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.4 GB at this context; duplicate traces of the same session: 74twfeje, tsbkz5yx; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 780.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 16k 780.6 27.2 29.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4whhdgg5",
   "source_url": "https://omlx.ai/benchmarks/performance/4whhdgg5",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-p9imkf7c-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.8 GB at this context; duplicate traces of the same session: alrydw34, tojucvhs; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 818.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 1k 818.4 51.5 28.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p9imkf7c",
   "source_url": "https://omlx.ai/benchmarks/performance/p9imkf7c",
   "tg_tokens": null,
   "tps": 51.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-p9imkf7c-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: alrydw34, tojucvhs; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 889.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 4k 889.1 40.2 29.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p9imkf7c",
   "source_url": "https://omlx.ai/benchmarks/performance/p9imkf7c",
   "tg_tokens": null,
   "tps": 40.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-p9imkf7c-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.2 GB at this context; duplicate traces of the same session: alrydw34, tojucvhs; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 774.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 16k 774.9 23.6 30.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p9imkf7c",
   "source_url": "https://omlx.ai/benchmarks/performance/p9imkf7c",
   "tg_tokens": null,
   "tps": 23.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-5dkxmbb1-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: 44v7ro51, p21en6do; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 798.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 1k 798.5 43.8 28.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5dkxmbb1",
   "source_url": "https://omlx.ai/benchmarks/performance/5dkxmbb1",
   "tg_tokens": null,
   "tps": 43.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-5dkxmbb1-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.0 GB at this context; duplicate traces of the same session: 44v7ro51, p21en6do; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 870.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 4k 870.4 45.0 29.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5dkxmbb1",
   "source_url": "https://omlx.ai/benchmarks/performance/5dkxmbb1",
   "tg_tokens": null,
   "tps": 45.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-5dkxmbb1-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.1 GB at this context; duplicate traces of the same session: 44v7ro51, p21en6do; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 774.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 16k 774.7 33.9 30.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5dkxmbb1",
   "source_url": "https://omlx.ai/benchmarks/performance/5dkxmbb1",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-3lfpsh48-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: d0dyv06r, v10w65s7; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 812.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 1k 812.5 40.4 28.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3lfpsh48",
   "source_url": "https://omlx.ai/benchmarks/performance/3lfpsh48",
   "tg_tokens": null,
   "tps": 40.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-3lfpsh48-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.1 GB at this context; duplicate traces of the same session: d0dyv06r, v10w65s7; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 861.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 4k 861.5 43.9 29.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3lfpsh48",
   "source_url": "https://omlx.ai/benchmarks/performance/3lfpsh48",
   "tg_tokens": null,
   "tps": 43.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-3lfpsh48-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.1 GB at this context; duplicate traces of the same session: d0dyv06r, v10w65s7; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 771.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 16k 771.4 29.4 30.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3lfpsh48",
   "source_url": "https://omlx.ai/benchmarks/performance/3lfpsh48",
   "tg_tokens": null,
   "tps": 29.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-pcs50i3p-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: 71jyb39p, 9t9z8net; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 802.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 1k 802.8 46.0 28.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pcs50i3p",
   "source_url": "https://omlx.ai/benchmarks/performance/pcs50i3p",
   "tg_tokens": null,
   "tps": 46.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-pcs50i3p-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.0 GB at this context; duplicate traces of the same session: 71jyb39p, 9t9z8net; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 877.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 4k 877.1 48.8 29.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pcs50i3p",
   "source_url": "https://omlx.ai/benchmarks/performance/pcs50i3p",
   "tg_tokens": null,
   "tps": 48.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-pcs50i3p-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.1 GB at this context; duplicate traces of the same session: 71jyb39p, 9t9z8net; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 776.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 16k 776.6 33.4 30.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pcs50i3p",
   "source_url": "https://omlx.ai/benchmarks/performance/pcs50i3p",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-8feshiqp-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: arh0h392, zlrp5xpn; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 831.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 1k 831.2 44.9 28.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8feshiqp",
   "source_url": "https://omlx.ai/benchmarks/performance/8feshiqp",
   "tg_tokens": null,
   "tps": 44.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-8feshiqp-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.0 GB at this context; duplicate traces of the same session: arh0h392, zlrp5xpn; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 871.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 4k 871.8 70.5 29.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8feshiqp",
   "source_url": "https://omlx.ai/benchmarks/performance/8feshiqp",
   "tg_tokens": null,
   "tps": 70.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-8feshiqp-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.1 GB at this context; duplicate traces of the same session: arh0h392, zlrp5xpn; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 771.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 16k 771.9 24.2 30.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8feshiqp",
   "source_url": "https://omlx.ai/benchmarks/performance/8feshiqp",
   "tg_tokens": null,
   "tps": 24.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-24xx5bx7-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: jm4ugdyp, rsmiyhja; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 728.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 1k 728.6 44.0 28.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/24xx5bx7",
   "source_url": "https://omlx.ai/benchmarks/performance/24xx5bx7",
   "tg_tokens": null,
   "tps": 44.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-24xx5bx7-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.0 GB at this context; duplicate traces of the same session: jm4ugdyp, rsmiyhja; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 863.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 4k 863.9 35.2 29.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/24xx5bx7",
   "source_url": "https://omlx.ai/benchmarks/performance/24xx5bx7",
   "tg_tokens": null,
   "tps": 35.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-24xx5bx7-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.1 GB at this context; duplicate traces of the same session: jm4ugdyp, rsmiyhja; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 777.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 16k 777.4 24.4 30.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/24xx5bx7",
   "source_url": "https://omlx.ai/benchmarks/performance/24xx5bx7",
   "tg_tokens": null,
   "tps": 24.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-drtjxc3b-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 3h65zuie; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 437.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 437.7 37.2 16.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/drtjxc3b",
   "source_url": "https://omlx.ai/benchmarks/performance/drtjxc3b",
   "tg_tokens": null,
   "tps": 37.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-drtjxc3b-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 3h65zuie; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 460.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 460.8 36.5 19.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/drtjxc3b",
   "source_url": "https://omlx.ai/benchmarks/performance/drtjxc3b",
   "tg_tokens": null,
   "tps": 36.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-f5d2ltkb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.5 GB at this context; duplicate traces of the same session: 7h8rh1wq, w1i58o09; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 880.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 880.4 77.0 17.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f5d2ltkb",
   "source_url": "https://omlx.ai/benchmarks/performance/f5d2ltkb",
   "tg_tokens": null,
   "tps": 77.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-f5d2ltkb-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 7h8rh1wq, w1i58o09; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1002.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,002 62.9 18.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f5d2ltkb",
   "source_url": "https://omlx.ai/benchmarks/performance/f5d2ltkb",
   "tg_tokens": null,
   "tps": 62.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-f5d2ltkb-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 7h8rh1wq, w1i58o09; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 934.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 934.3 72.0 19.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f5d2ltkb",
   "source_url": "https://omlx.ai/benchmarks/performance/f5d2ltkb",
   "tg_tokens": null,
   "tps": 72.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-f5d2ltkb-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 7h8rh1wq, w1i58o09; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 886.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 886.9 76.2 19.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f5d2ltkb",
   "source_url": "https://omlx.ai/benchmarks/performance/f5d2ltkb",
   "tg_tokens": null,
   "tps": 76.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-2otv0tho-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "REAP288 MTP build; oMLX community benchmark trace, single run at 1k context; peak memory 32.1 GB; TTFT 1994 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 513.7,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-Flash-Next-Uncensored-REAP288-MTP-MLX-oQ3 on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 3bit; 1k tokens 513.7 PP tok/s 29.7 TG tok/s 1994 TTFT (ms) 32.1 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2otv0tho",
   "source_url": "https://omlx.ai/benchmarks/performance/2otv0tho",
   "tg_tokens": null,
   "tps": 29.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yixtf0pt-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.5 GB at this context; duplicate traces of the same session: 7bje7dbo, ajfaxohk, yis5ijiu, z9amczr2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1157.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,157 72.4 70.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yixtf0pt",
   "source_url": "https://omlx.ai/benchmarks/performance/yixtf0pt",
   "tg_tokens": null,
   "tps": 72.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yixtf0pt-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 7bje7dbo, ajfaxohk, yis5ijiu, z9amczr2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2049.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,049 67.1 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yixtf0pt",
   "source_url": "https://omlx.ai/benchmarks/performance/yixtf0pt",
   "tg_tokens": null,
   "tps": 67.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yixtf0pt-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 7bje7dbo, ajfaxohk, yis5ijiu, z9amczr2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2546.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,546 54.8 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yixtf0pt",
   "source_url": "https://omlx.ai/benchmarks/performance/yixtf0pt",
   "tg_tokens": null,
   "tps": 54.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yixtf0pt-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: 7bje7dbo, ajfaxohk, yis5ijiu, z9amczr2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2477.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,477 68.0 73.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yixtf0pt",
   "source_url": "https://omlx.ai/benchmarks/performance/yixtf0pt",
   "tg_tokens": null,
   "tps": 68.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-yixtf0pt-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.6 GB at this context; duplicate traces of the same session: 7bje7dbo, ajfaxohk, yis5ijiu, z9amczr2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2384.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,384 87.0 74.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yixtf0pt",
   "source_url": "https://omlx.ai/benchmarks/performance/yixtf0pt",
   "tg_tokens": null,
   "tps": 87.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7qjnookj-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.5 GB at this context; duplicate traces of the same session: 19sssku3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 579.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 579.6 36.6 70.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7qjnookj",
   "source_url": "https://omlx.ai/benchmarks/performance/7qjnookj",
   "tg_tokens": null,
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7qjnookj-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: 19sssku3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1938.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,938 63.4 71.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7qjnookj",
   "source_url": "https://omlx.ai/benchmarks/performance/7qjnookj",
   "tg_tokens": null,
   "tps": 63.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-taobtp50-ctx1024",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.7 GB at this context; duplicate traces of the same session: aylhjrn2, v6j370mp; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.4,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-OptiQ-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k 224.4 14.3 7.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/taobtp50",
   "source_url": "https://omlx.ai/benchmarks/performance/taobtp50",
   "tg_tokens": null,
   "tps": 14.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-taobtp50-ctx4096",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.1 GB at this context; duplicate traces of the same session: aylhjrn2, v6j370mp; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 229.3,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-OptiQ-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 229.3 14.5 8.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/taobtp50",
   "source_url": "https://omlx.ai/benchmarks/performance/taobtp50",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-taobtp50-ctx8192",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.3 GB at this context; duplicate traces of the same session: aylhjrn2, v6j370mp; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 219.4,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-OptiQ-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 8k 219.4 14.3 8.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/taobtp50",
   "source_url": "https://omlx.ai/benchmarks/performance/taobtp50",
   "tg_tokens": null,
   "tps": 14.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2v22b17z-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.3 GB; TTFT 242 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4239.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 4,239 PP tok/s 135.3 TG tok/s 242 TTFT (ms) 19.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2v22b17z",
   "source_url": "https://omlx.ai/benchmarks/performance/2v22b17z",
   "tg_tokens": null,
   "tps": 135.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-141zod0w-ctx1024",
   "model": "Ornith-1.5-35B-A3B-Vision-MTPLX",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1876.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-Vision-MTPLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 1,876 198.9 21.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/141zod0w",
   "source_url": "https://omlx.ai/benchmarks/performance/141zod0w",
   "tg_tokens": null,
   "tps": 198.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-141zod0w-ctx4096",
   "model": "Ornith-1.5-35B-A3B-Vision-MTPLX",
   "notes": "oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2580.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-Vision-MTPLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 2,580 170.3 23.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/141zod0w",
   "source_url": "https://omlx.ai/benchmarks/performance/141zod0w",
   "tg_tokens": null,
   "tps": 170.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-141zod0w-ctx8192",
   "model": "Ornith-1.5-35B-A3B-Vision-MTPLX",
   "notes": "oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2485.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-Vision-MTPLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 2,485 182.7 23.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/141zod0w",
   "source_url": "https://omlx.ai/benchmarks/performance/141zod0w",
   "tg_tokens": null,
   "tps": 182.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-141zod0w-ctx16384",
   "model": "Ornith-1.5-35B-A3B-Vision-MTPLX",
   "notes": "oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2222.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-Vision-MTPLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 16k 2,222 193.2 23.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/141zod0w",
   "source_url": "https://omlx.ai/benchmarks/performance/141zod0w",
   "tg_tokens": null,
   "tps": 193.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-141zod0w-ctx32768",
   "model": "Ornith-1.5-35B-A3B-Vision-MTPLX",
   "notes": "oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1808.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-Vision-MTPLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 1,808 175.3 24.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/141zod0w",
   "source_url": "https://omlx.ai/benchmarks/performance/141zod0w",
   "tg_tokens": null,
   "tps": 175.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-epvg1p52-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.8 GB at this context; duplicate traces of the same session: kagml18l, vsl0m39x; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 842.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 1k 842.8 70.9 30.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/epvg1p52",
   "source_url": "https://omlx.ai/benchmarks/performance/epvg1p52",
   "tg_tokens": null,
   "tps": 70.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-epvg1p52-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.9 GB at this context; duplicate traces of the same session: kagml18l, vsl0m39x; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 951.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 4k 951.3 68.8 32.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/epvg1p52",
   "source_url": "https://omlx.ai/benchmarks/performance/epvg1p52",
   "tg_tokens": null,
   "tps": 68.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-epvg1p52-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: kagml18l, vsl0m39x; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 822.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 16k 822.6 50.9 33.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/epvg1p52",
   "source_url": "https://omlx.ai/benchmarks/performance/epvg1p52",
   "tg_tokens": null,
   "tps": 50.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-1dld3fmq-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.8 GB at this context; duplicate traces of the same session: 009qoxam, r02wgo57; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 812.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 1k 812.3 36.4 28.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1dld3fmq",
   "source_url": "https://omlx.ai/benchmarks/performance/1dld3fmq",
   "tg_tokens": null,
   "tps": 36.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-1dld3fmq-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.8 GB at this context; duplicate traces of the same session: 009qoxam, r02wgo57; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 856.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 4k 856.3 47.8 29.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1dld3fmq",
   "source_url": "https://omlx.ai/benchmarks/performance/1dld3fmq",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-1dld3fmq-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.8 GB at this context; duplicate traces of the same session: 009qoxam, r02wgo57; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 767.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 16k 767.9 30.4 30.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1dld3fmq",
   "source_url": "https://omlx.ai/benchmarks/performance/1dld3fmq",
   "tg_tokens": null,
   "tps": 30.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-9djih2m2-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.8 GB at this context; duplicate traces of the same session: 9pgxy1kv, bmquzb4w; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 789.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 1k 789.3 39.4 28.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9djih2m2",
   "source_url": "https://omlx.ai/benchmarks/performance/9djih2m2",
   "tg_tokens": null,
   "tps": 39.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-9djih2m2-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.8 GB at this context; duplicate traces of the same session: 9pgxy1kv, bmquzb4w; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 855.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 4k 855.8 47.6 29.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9djih2m2",
   "source_url": "https://omlx.ai/benchmarks/performance/9djih2m2",
   "tg_tokens": null,
   "tps": 47.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-9djih2m2-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.8 GB at this context; duplicate traces of the same session: 9pgxy1kv, bmquzb4w; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 769.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 16k 769.7 29.6 30.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9djih2m2",
   "source_url": "https://omlx.ai/benchmarks/performance/9djih2m2",
   "tg_tokens": null,
   "tps": 29.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-y0hoqay1-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 880.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 880.4 77.0 17.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y0hoqay1",
   "source_url": "https://omlx.ai/benchmarks/performance/y0hoqay1",
   "tg_tokens": null,
   "tps": 77.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-y0hoqay1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1002.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,002 62.9 18.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y0hoqay1",
   "source_url": "https://omlx.ai/benchmarks/performance/y0hoqay1",
   "tg_tokens": null,
   "tps": 62.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-y0hoqay1-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 934.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 934.3 72.0 19.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y0hoqay1",
   "source_url": "https://omlx.ai/benchmarks/performance/y0hoqay1",
   "tg_tokens": null,
   "tps": 72.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-y0hoqay1-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 886.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 886.9 76.2 19.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y0hoqay1",
   "source_url": "https://omlx.ai/benchmarks/performance/y0hoqay1",
   "tg_tokens": null,
   "tps": 76.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1gffnd13-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 182.9 GB at this context; duplicate traces of the same session: enq0nv1r, m90wtr7s; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 846.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 846.8 86.8 182.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1gffnd13",
   "source_url": "https://omlx.ai/benchmarks/performance/1gffnd13",
   "tg_tokens": null,
   "tps": 86.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1gffnd13-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.7 GB at this context; duplicate traces of the same session: enq0nv1r, m90wtr7s; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1288.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 1,288 78.5 183.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1gffnd13",
   "source_url": "https://omlx.ai/benchmarks/performance/1gffnd13",
   "tg_tokens": null,
   "tps": 78.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1gffnd13-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 191.5 GB at this context; duplicate traces of the same session: enq0nv1r, m90wtr7s; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1261.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 128k 1,261 57.8 191.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1gffnd13",
   "source_url": "https://omlx.ai/benchmarks/performance/1gffnd13",
   "tg_tokens": null,
   "tps": 57.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4h1o40g3-ctx1024",
   "model": "Ornith-1.0-35B-Vision",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: j1vs2g3j, thfx98v6, tl5sqfmp, w1x4ph5l; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2212.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.0-35B-Vision on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 2,212 119.6 19.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4h1o40g3",
   "source_url": "https://omlx.ai/benchmarks/performance/4h1o40g3",
   "tg_tokens": null,
   "tps": 119.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4h1o40g3-ctx4096",
   "model": "Ornith-1.0-35B-Vision",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: j1vs2g3j, thfx98v6, tl5sqfmp, w1x4ph5l; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2745.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.0-35B-Vision on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 2,745 117.2 21.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4h1o40g3",
   "source_url": "https://omlx.ai/benchmarks/performance/4h1o40g3",
   "tg_tokens": null,
   "tps": 117.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4h1o40g3-ctx8192",
   "model": "Ornith-1.0-35B-Vision",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: j1vs2g3j, thfx98v6, tl5sqfmp, w1x4ph5l; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2598.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.0-35B-Vision on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 2,598 109.6 21.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4h1o40g3",
   "source_url": "https://omlx.ai/benchmarks/performance/4h1o40g3",
   "tg_tokens": null,
   "tps": 109.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4h1o40g3-ctx16384",
   "model": "Ornith-1.0-35B-Vision",
   "notes": "oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: j1vs2g3j, thfx98v6, tl5sqfmp, w1x4ph5l; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2325.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.0-35B-Vision on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 16k 2,325 102.4 21.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4h1o40g3",
   "source_url": "https://omlx.ai/benchmarks/performance/4h1o40g3",
   "tg_tokens": null,
   "tps": 102.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4h1o40g3-ctx32768",
   "model": "Ornith-1.0-35B-Vision",
   "notes": "oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: j1vs2g3j, thfx98v6, tl5sqfmp, w1x4ph5l; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1915.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.0-35B-Vision on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 1,915 94.0 22.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4h1o40g3",
   "source_url": "https://omlx.ai/benchmarks/performance/4h1o40g3",
   "tg_tokens": null,
   "tps": 94.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-hyx9aysv-ctx1024",
   "model": "Qwen3 14B",
   "notes": "oMLX community benchmark trace; peak memory 15.3 GB at this context; duplicate traces of the same session: s96dkw48; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 616.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-14B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 616.7 41.9 15.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hyx9aysv",
   "source_url": "https://omlx.ai/benchmarks/performance/hyx9aysv",
   "tg_tokens": null,
   "tps": 41.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-hyx9aysv-ctx4096",
   "model": "Qwen3 14B",
   "notes": "oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: s96dkw48; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 611.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-14B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 611.5 39.9 15.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hyx9aysv",
   "source_url": "https://omlx.ai/benchmarks/performance/hyx9aysv",
   "tg_tokens": null,
   "tps": 39.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fh4onyf5-ctx1024",
   "model": "Qwen3.6-35B-A3B-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 3w6bt860, 68z6jvdf, ct0qekps, pega2dz2, xuzvij52; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2447.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.6-35B-A3B-Heretic-MLX-mxfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 2,447 92.3 18.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fh4onyf5",
   "source_url": "https://omlx.ai/benchmarks/performance/fh4onyf5",
   "tg_tokens": null,
   "tps": 92.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fh4onyf5-ctx4096",
   "model": "Qwen3.6-35B-A3B-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 3w6bt860, 68z6jvdf, ct0qekps, pega2dz2, xuzvij52; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3117.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.6-35B-A3B-Heretic-MLX-mxfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 3,117 95.2 19.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fh4onyf5",
   "source_url": "https://omlx.ai/benchmarks/performance/fh4onyf5",
   "tg_tokens": null,
   "tps": 95.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fh4onyf5-ctx8192",
   "model": "Qwen3.6-35B-A3B-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 3w6bt860, 68z6jvdf, ct0qekps, pega2dz2, xuzvij52; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3143.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.6-35B-A3B-Heretic-MLX-mxfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 3,143 87.4 19.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fh4onyf5",
   "source_url": "https://omlx.ai/benchmarks/performance/fh4onyf5",
   "tg_tokens": null,
   "tps": 87.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fh4onyf5-ctx16384",
   "model": "Qwen3.6-35B-A3B-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 3w6bt860, 68z6jvdf, ct0qekps, pega2dz2, xuzvij52; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2875.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.6-35B-A3B-Heretic-MLX-mxfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 2,875 85.9 19.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fh4onyf5",
   "source_url": "https://omlx.ai/benchmarks/performance/fh4onyf5",
   "tg_tokens": null,
   "tps": 85.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fh4onyf5-ctx32768",
   "model": "Qwen3.6-35B-A3B-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 3w6bt860, 68z6jvdf, ct0qekps, pega2dz2, xuzvij52; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2388.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.6-35B-A3B-Heretic-MLX-mxfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 2,388 76.1 19.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fh4onyf5",
   "source_url": "https://omlx.ai/benchmarks/performance/fh4onyf5",
   "tg_tokens": null,
   "tps": 76.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fh4onyf5-ctx65536",
   "model": "Qwen3.6-35B-A3B-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 3w6bt860, 68z6jvdf, ct0qekps, pega2dz2, xuzvij52; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1831.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.6-35B-A3B-Heretic-MLX-mxfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 1,831 58.6 20.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fh4onyf5",
   "source_url": "https://omlx.ai/benchmarks/performance/fh4onyf5",
   "tg_tokens": null,
   "tps": 58.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-41ie19lc-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 1q9gtm0a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 158.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 158.7 12.8 19.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/41ie19lc",
   "source_url": "https://omlx.ai/benchmarks/performance/41ie19lc",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-41ie19lc-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 1q9gtm0a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 153.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 153.9 10.9 21.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/41ie19lc",
   "source_url": "https://omlx.ai/benchmarks/performance/41ie19lc",
   "tg_tokens": null,
   "tps": 10.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3u3qme0s-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 12hro8wm, q9snh7ru, uc4rfj00; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 589.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 589.9 40.0 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3u3qme0s",
   "source_url": "https://omlx.ai/benchmarks/performance/3u3qme0s",
   "tg_tokens": null,
   "tps": 40.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3u3qme0s-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 12hro8wm, q9snh7ru, uc4rfj00; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 692.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 692.5 35.8 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3u3qme0s",
   "source_url": "https://omlx.ai/benchmarks/performance/3u3qme0s",
   "tg_tokens": null,
   "tps": 35.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3u3qme0s-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 12hro8wm, q9snh7ru, uc4rfj00; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 716.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 716.4 34.6 19.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3u3qme0s",
   "source_url": "https://omlx.ai/benchmarks/performance/3u3qme0s",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3u3qme0s-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 12hro8wm, q9snh7ru, uc4rfj00; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 697.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 697.3 47.0 20.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3u3qme0s",
   "source_url": "https://omlx.ai/benchmarks/performance/3u3qme0s",
   "tg_tokens": null,
   "tps": 47.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3u3qme0s-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: 12hro8wm, q9snh7ru, uc4rfj00; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 635.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 635.9 44.1 23.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3u3qme0s",
   "source_url": "https://omlx.ai/benchmarks/performance/3u3qme0s",
   "tg_tokens": null,
   "tps": 44.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-rxdikqfc-ctx4096",
   "model": "gemma-4-12b-it",
   "notes": "OptiQ; oMLX community benchmark trace, single run at 4k context; peak memory 9.7 GB; TTFT 3061 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1338.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 1,338 PP tok/s 47.8 TG tok/s 3061 TTFT (ms) 9.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rxdikqfc",
   "source_url": "https://omlx.ai/benchmarks/performance/rxdikqfc",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-yx5v01hy-ctx1024",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 3.0 GB at this context; duplicate traces of the same session: 3gv9sa95, 3xbabtf2, 64wq5xlh, 82e9o6e4, vk2aizoh, w21q9fla; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 7929.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e2b-it-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 7,929 135.9 3.0 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yx5v01hy",
   "source_url": "https://omlx.ai/benchmarks/performance/yx5v01hy",
   "tg_tokens": null,
   "tps": 135.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-yx5v01hy-ctx4096",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 3.0 GB at this context; duplicate traces of the same session: 3gv9sa95, 3xbabtf2, 64wq5xlh, 82e9o6e4, vk2aizoh, w21q9fla; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 10452.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e2b-it-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 10,452 125.9 3.0 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yx5v01hy",
   "source_url": "https://omlx.ai/benchmarks/performance/yx5v01hy",
   "tg_tokens": null,
   "tps": 125.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-yx5v01hy-ctx8192",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 3.1 GB at this context; duplicate traces of the same session: 3gv9sa95, 3xbabtf2, 64wq5xlh, 82e9o6e4, vk2aizoh, w21q9fla; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 10141.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e2b-it-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 8k 10,141 123.5 3.1 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yx5v01hy",
   "source_url": "https://omlx.ai/benchmarks/performance/yx5v01hy",
   "tg_tokens": null,
   "tps": 123.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-yx5v01hy-ctx16384",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 3.2 GB at this context; duplicate traces of the same session: 3gv9sa95, 3xbabtf2, 64wq5xlh, 82e9o6e4, vk2aizoh, w21q9fla; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 9720.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e2b-it-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 16k 9,720 118.7 3.2 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yx5v01hy",
   "source_url": "https://omlx.ai/benchmarks/performance/yx5v01hy",
   "tg_tokens": null,
   "tps": 118.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-yx5v01hy-ctx32768",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 3.6 GB at this context; duplicate traces of the same session: 3gv9sa95, 3xbabtf2, 64wq5xlh, 82e9o6e4, vk2aizoh, w21q9fla; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 8281.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e2b-it-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 32k 8,281 103.6 3.6 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yx5v01hy",
   "source_url": "https://omlx.ai/benchmarks/performance/yx5v01hy",
   "tg_tokens": null,
   "tps": 103.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-yx5v01hy-ctx65536",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 4.4 GB at this context; duplicate traces of the same session: 3gv9sa95, 3xbabtf2, 64wq5xlh, 82e9o6e4, vk2aizoh, w21q9fla; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6228.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e2b-it-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 64k 6,228 84.8 4.4 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yx5v01hy",
   "source_url": "https://omlx.ai/benchmarks/performance/yx5v01hy",
   "tg_tokens": null,
   "tps": 84.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-yx5v01hy-ctx131072",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 6.0 GB at this context; duplicate traces of the same session: 3gv9sa95, 3xbabtf2, 64wq5xlh, 82e9o6e4, vk2aizoh, w21q9fla; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4153.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e2b-it-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 128k 4,153 61.9 6.0 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yx5v01hy",
   "source_url": "https://omlx.ai/benchmarks/performance/yx5v01hy",
   "tg_tokens": null,
   "tps": 61.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-prqnxxoh-ctx1024",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 3.0 GB at this context; duplicate traces of the same session: 4r35rl37, geudqug1; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 7858.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e2b-it-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 7,858 130.5 3.0 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/prqnxxoh",
   "source_url": "https://omlx.ai/benchmarks/performance/prqnxxoh",
   "tg_tokens": null,
   "tps": 130.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-prqnxxoh-ctx4096",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 3.0 GB at this context; duplicate traces of the same session: 4r35rl37, geudqug1; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 10439.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e2b-it-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 10,439 125.7 3.0 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/prqnxxoh",
   "source_url": "https://omlx.ai/benchmarks/performance/prqnxxoh",
   "tg_tokens": null,
   "tps": 125.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-prqnxxoh-ctx8192",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 3.1 GB at this context; duplicate traces of the same session: 4r35rl37, geudqug1; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 10203.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e2b-it-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 8k 10,203 121.5 3.1 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/prqnxxoh",
   "source_url": "https://omlx.ai/benchmarks/performance/prqnxxoh",
   "tg_tokens": null,
   "tps": 121.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xg99biei-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "lmstudio-community repack; oMLX community benchmark trace; peak memory 6.1 GB at this context; duplicate traces of the same session: 05gxs5gr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2798.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "lmstudio-community--gemma-4-E4B-it-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 2,798 89.3 6.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xg99biei",
   "source_url": "https://omlx.ai/benchmarks/performance/xg99biei",
   "tg_tokens": null,
   "tps": 89.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xg99biei-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "lmstudio-community repack; oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: 05gxs5gr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2949.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "lmstudio-community--gemma-4-E4B-it-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 2,949 85.6 6.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xg99biei",
   "source_url": "https://omlx.ai/benchmarks/performance/xg99biei",
   "tg_tokens": null,
   "tps": 85.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-d9406oav-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "lmstudio-community repack; oMLX community benchmark trace; peak memory 8.1 GB at this context; duplicate traces of the same session: 0irgksnp; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2795.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "lmstudio-community--gemma-4-E4B-it-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 2,795 83.7 8.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d9406oav",
   "source_url": "https://omlx.ai/benchmarks/performance/d9406oav",
   "tg_tokens": null,
   "tps": 83.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-d9406oav-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "lmstudio-community repack; oMLX community benchmark trace; peak memory 8.1 GB at this context; duplicate traces of the same session: 0irgksnp; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2722.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "lmstudio-community--gemma-4-E4B-it-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 2,722 80.6 8.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d9406oav",
   "source_url": "https://omlx.ai/benchmarks/performance/d9406oav",
   "tg_tokens": null,
   "tps": 80.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-agkl570r-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "mlx-community; oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: q4tvg0db; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2138.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 2,138 82.8 35.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/agkl570r",
   "source_url": "https://omlx.ai/benchmarks/performance/agkl570r",
   "tg_tokens": null,
   "tps": 82.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-agkl570r-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "mlx-community; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: q4tvg0db; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2654.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 2,654 80.5 37.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/agkl570r",
   "source_url": "https://omlx.ai/benchmarks/performance/agkl570r",
   "tg_tokens": null,
   "tps": 80.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-wpg5hujn-ctx1024",
   "model": "Occamy-1.0",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 0ztp6xlw, dfcvtnc1, tqc3ypgv, ul3yxhz4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2221.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "occamy-1.0-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 2,221 120.5 19.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wpg5hujn",
   "source_url": "https://omlx.ai/benchmarks/performance/wpg5hujn",
   "tg_tokens": null,
   "tps": 120.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-wpg5hujn-ctx4096",
   "model": "Occamy-1.0",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 0ztp6xlw, dfcvtnc1, tqc3ypgv, ul3yxhz4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2743.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "occamy-1.0-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 2,743 118.7 21.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wpg5hujn",
   "source_url": "https://omlx.ai/benchmarks/performance/wpg5hujn",
   "tg_tokens": null,
   "tps": 118.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-wpg5hujn-ctx8192",
   "model": "Occamy-1.0",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 0ztp6xlw, dfcvtnc1, tqc3ypgv, ul3yxhz4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2595.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "occamy-1.0-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 2,595 110.6 21.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wpg5hujn",
   "source_url": "https://omlx.ai/benchmarks/performance/wpg5hujn",
   "tg_tokens": null,
   "tps": 110.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-wpg5hujn-ctx16384",
   "model": "Occamy-1.0",
   "notes": "oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: 0ztp6xlw, dfcvtnc1, tqc3ypgv, ul3yxhz4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2326.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "occamy-1.0-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 16k 2,326 104.0 21.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wpg5hujn",
   "source_url": "https://omlx.ai/benchmarks/performance/wpg5hujn",
   "tg_tokens": null,
   "tps": 104.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-wpg5hujn-ctx32768",
   "model": "Occamy-1.0",
   "notes": "oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: 0ztp6xlw, dfcvtnc1, tqc3ypgv, ul3yxhz4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1915.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "occamy-1.0-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 1,915 94.7 22.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wpg5hujn",
   "source_url": "https://omlx.ai/benchmarks/performance/wpg5hujn",
   "tg_tokens": null,
   "tps": 94.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gffyvvl9-ctx1024",
   "model": "Occamy-1.0",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.0 GB at this context; duplicate traces of the same session: f5t3k2a8, jup5tyz8, my0ncje9, pgshkejg; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2028.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "occamy-1.0-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 1k 2,028 145.8 27.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gffyvvl9",
   "source_url": "https://omlx.ai/benchmarks/performance/gffyvvl9",
   "tg_tokens": null,
   "tps": 145.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gffyvvl9-ctx4096",
   "model": "Occamy-1.0",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.9 GB at this context; duplicate traces of the same session: f5t3k2a8, jup5tyz8, my0ncje9, pgshkejg; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2565.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "occamy-1.0-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 4k 2,565 150.8 28.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gffyvvl9",
   "source_url": "https://omlx.ai/benchmarks/performance/gffyvvl9",
   "tg_tokens": null,
   "tps": 150.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gffyvvl9-ctx8192",
   "model": "Occamy-1.0",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.1 GB at this context; duplicate traces of the same session: f5t3k2a8, jup5tyz8, my0ncje9, pgshkejg; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2424.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "occamy-1.0-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 8k 2,424 145.5 29.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gffyvvl9",
   "source_url": "https://omlx.ai/benchmarks/performance/gffyvvl9",
   "tg_tokens": null,
   "tps": 145.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gffyvvl9-ctx16384",
   "model": "Occamy-1.0",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.4 GB at this context; duplicate traces of the same session: f5t3k2a8, jup5tyz8, my0ncje9, pgshkejg; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2164.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "occamy-1.0-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 16k 2,164 185.9 29.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gffyvvl9",
   "source_url": "https://omlx.ai/benchmarks/performance/gffyvvl9",
   "tg_tokens": null,
   "tps": 185.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gffyvvl9-ctx32768",
   "model": "Occamy-1.0",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.1 GB at this context; duplicate traces of the same session: f5t3k2a8, jup5tyz8, my0ncje9, pgshkejg; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1772.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "occamy-1.0-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 32k 1,772 165.7 30.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gffyvvl9",
   "source_url": "https://omlx.ai/benchmarks/performance/gffyvvl9",
   "tg_tokens": null,
   "tps": 165.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rhy7dzvu-ctx1024",
   "model": "GLM-5.3-Flash-Abliterated",
   "notes": "MLX; oMLX community benchmark trace, single run at 1k context; peak memory 177.8 GB; TTFT 2230 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 459.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-Abliterated-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k tokens 459.2 PP tok/s 38.9 TG tok/s 2230 TTFT (ms) 177.8 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rhy7dzvu",
   "source_url": "https://omlx.ai/benchmarks/performance/rhy7dzvu",
   "tg_tokens": null,
   "tps": 38.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-0eh42oob-ctx4096",
   "model": "Mistral-7B-Instruct-v0.3",
   "notes": "oMLX community benchmark trace; peak memory 5.3 GB at this context; duplicate traces of the same session: 1m18yj2y; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 349.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Mistral-7B-Instruct-v0.3-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 4k 349.5 26.7 5.3 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0eh42oob",
   "source_url": "https://omlx.ai/benchmarks/performance/0eh42oob",
   "tg_tokens": null,
   "tps": 26.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-0eh42oob-ctx16384",
   "model": "Mistral-7B-Instruct-v0.3",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: 1m18yj2y; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 293.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Mistral-7B-Instruct-v0.3-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 16k 293.6 17.7 6.4 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0eh42oob",
   "source_url": "https://omlx.ai/benchmarks/performance/0eh42oob",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-rgd1qpki-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "Text build; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 2rrn0our, 3mic5r5o, 4w3cu9y3, w0sg9s1p; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 22.5,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.5-122B-A10B-Text-mxfp4-mlx on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 22.5 7.2 23.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rgd1qpki",
   "source_url": "https://omlx.ai/benchmarks/performance/rgd1qpki",
   "tg_tokens": null,
   "tps": 7.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-rgd1qpki-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "Text build; oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: 2rrn0our, 3mic5r5o, 4w3cu9y3, w0sg9s1p; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 154.8,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.5-122B-A10B-Text-mxfp4-mlx on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 154.8 6.6 27.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rgd1qpki",
   "source_url": "https://omlx.ai/benchmarks/performance/rgd1qpki",
   "tg_tokens": null,
   "tps": 6.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-rgd1qpki-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "Text build; oMLX community benchmark trace; peak memory 28.9 GB at this context; duplicate traces of the same session: 2rrn0our, 3mic5r5o, 4w3cu9y3, w0sg9s1p; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.8,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.5-122B-A10B-Text-mxfp4-mlx on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 125.8 6.6 28.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rgd1qpki",
   "source_url": "https://omlx.ai/benchmarks/performance/rgd1qpki",
   "tg_tokens": null,
   "tps": 6.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-rgd1qpki-ctx32768",
   "model": "Qwen3.5-122B-A10B",
   "notes": "Text build; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: 2rrn0our, 3mic5r5o, 4w3cu9y3, w0sg9s1p; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 120.5,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.5-122B-A10B-Text-mxfp4-mlx on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 32k 120.5 6.6 29.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rgd1qpki",
   "source_url": "https://omlx.ai/benchmarks/performance/rgd1qpki",
   "tg_tokens": null,
   "tps": 6.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-rgd1qpki-ctx65536",
   "model": "Qwen3.5-122B-A10B",
   "notes": "Text build; oMLX community benchmark trace; peak memory 31.8 GB at this context; duplicate traces of the same session: 2rrn0our, 3mic5r5o, 4w3cu9y3, w0sg9s1p; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 92.6,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.5-122B-A10B-Text-mxfp4-mlx on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 64k 92.6 5.6 31.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rgd1qpki",
   "source_url": "https://omlx.ai/benchmarks/performance/rgd1qpki",
   "tg_tokens": null,
   "tps": 5.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-c4d2um2r-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace, single run at 8k context; peak memory 7.5 GB; TTFT 8617 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 950.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k tokens 950.7 PP tok/s 84.2 TG tok/s 8617 TTFT (ms) 7.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c4d2um2r",
   "source_url": "https://omlx.ai/benchmarks/performance/c4d2um2r",
   "tg_tokens": null,
   "tps": 84.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-j3u2wxoh-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 1au2ae1b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2095.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 2,095 128.5 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j3u2wxoh",
   "source_url": "https://omlx.ai/benchmarks/performance/j3u2wxoh",
   "tg_tokens": null,
   "tps": 128.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-j3u2wxoh-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: 1au2ae1b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1785.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,785 112.6 21.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j3u2wxoh",
   "source_url": "https://omlx.ai/benchmarks/performance/j3u2wxoh",
   "tg_tokens": null,
   "tps": 112.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-qletxl7d-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 4qbrzcaw, e6eh6xaa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2126.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 2,126 130.6 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qletxl7d",
   "source_url": "https://omlx.ai/benchmarks/performance/qletxl7d",
   "tg_tokens": null,
   "tps": 130.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-qletxl7d-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 4qbrzcaw, e6eh6xaa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1447.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 1,447 103.2 22.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qletxl7d",
   "source_url": "https://omlx.ai/benchmarks/performance/qletxl7d",
   "tg_tokens": null,
   "tps": 103.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-qletxl7d-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: 4qbrzcaw, e6eh6xaa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 691.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 691.3 63.6 25.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qletxl7d",
   "source_url": "https://omlx.ai/benchmarks/performance/qletxl7d",
   "tg_tokens": null,
   "tps": 63.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-3k15y6ax-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: m5umhud8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1217.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,217 75.3 22.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3k15y6ax",
   "source_url": "https://omlx.ai/benchmarks/performance/3k15y6ax",
   "tg_tokens": null,
   "tps": 75.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-3k15y6ax-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: m5umhud8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1120.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,120 67.2 22.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3k15y6ax",
   "source_url": "https://omlx.ai/benchmarks/performance/3k15y6ax",
   "tg_tokens": null,
   "tps": 67.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-jfeejti8-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: opkt8gac; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 510.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 510.9 35.5 18.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jfeejti8",
   "source_url": "https://omlx.ai/benchmarks/performance/jfeejti8",
   "tg_tokens": null,
   "tps": 35.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-jfeejti8-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: opkt8gac; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 435.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 435.4 25.7 20.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jfeejti8",
   "source_url": "https://omlx.ai/benchmarks/performance/jfeejti8",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-ejs1jsky-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: qz8iw4a6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 212.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 212.5 20.7 25.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ejs1jsky",
   "source_url": "https://omlx.ai/benchmarks/performance/ejs1jsky",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-ejs1jsky-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: qz8iw4a6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 198.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 198.5 28.9 26.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ejs1jsky",
   "source_url": "https://omlx.ai/benchmarks/performance/ejs1jsky",
   "tg_tokens": null,
   "tps": 28.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-qc2ovhe1-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 0w37q6ou; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 130.0 23.5 17.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qc2ovhe1",
   "source_url": "https://omlx.ai/benchmarks/performance/qc2ovhe1",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-qc2ovhe1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 0w37q6ou; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 131.5 25.1 19.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qc2ovhe1",
   "source_url": "https://omlx.ai/benchmarks/performance/qc2ovhe1",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-nze518x9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 589.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 589.9 40.0 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nze518x9",
   "source_url": "https://omlx.ai/benchmarks/performance/nze518x9",
   "tg_tokens": null,
   "tps": 40.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-nze518x9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 692.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 692.5 35.8 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nze518x9",
   "source_url": "https://omlx.ai/benchmarks/performance/nze518x9",
   "tg_tokens": null,
   "tps": 35.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-nze518x9-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 716.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 716.4 34.6 19.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nze518x9",
   "source_url": "https://omlx.ai/benchmarks/performance/nze518x9",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-nze518x9-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 697.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 697.3 47.0 20.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nze518x9",
   "source_url": "https://omlx.ai/benchmarks/performance/nze518x9",
   "tg_tokens": null,
   "tps": 47.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-nze518x9-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 635.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 635.9 44.1 23.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nze518x9",
   "source_url": "https://omlx.ai/benchmarks/performance/nze518x9",
   "tg_tokens": null,
   "tps": 44.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-ds2l73gz-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 101.6 GB; TTFT 11006 ms; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 744.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 8k tokens 744.3 PP tok/s 48.5 TG tok/s 11006 TTFT (ms) 101.6 Peak mem",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ds2l73gz",
   "source_url": "https://omlx.ai/benchmarks/performance/ds2l73gz",
   "tg_tokens": null,
   "tps": 48.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-ii6kwkgg-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.7 GB at this context; duplicate traces of the same session: fa04tbl4; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 774.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 8k 774.6 58.1 101.7 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ii6kwkgg",
   "source_url": "https://omlx.ai/benchmarks/performance/ii6kwkgg",
   "tg_tokens": null,
   "tps": 58.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-ii6kwkgg-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.8 GB at this context; duplicate traces of the same session: fa04tbl4; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 773.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 32k 773.9 56.4 102.8 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ii6kwkgg",
   "source_url": "https://omlx.ai/benchmarks/performance/ii6kwkgg",
   "tg_tokens": null,
   "tps": 56.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ei1g3dfm-ctx1024",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: xa69zg8k; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 344.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-31b-it-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 344.3 38.0 34.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ei1g3dfm",
   "source_url": "https://omlx.ai/benchmarks/performance/ei1g3dfm",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ei1g3dfm-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.0 GB at this context; duplicate traces of the same session: xa69zg8k; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 346.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-31b-it-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 346.3 34.1 35.0 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ei1g3dfm",
   "source_url": "https://omlx.ai/benchmarks/performance/ei1g3dfm",
   "tg_tokens": null,
   "tps": 34.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-00zr3ujx-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "lmstudio-community repack; oMLX community benchmark trace; peak memory 5.4 GB at this context; duplicate traces of the same session: fukdavxl; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1976.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "lmstudio-community--Qwen3.5-4B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 1,976 111.5 5.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/00zr3ujx",
   "source_url": "https://omlx.ai/benchmarks/performance/00zr3ujx",
   "tg_tokens": null,
   "tps": 111.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-00zr3ujx-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "lmstudio-community repack; oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: fukdavxl; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2122.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "lmstudio-community--Qwen3.5-4B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 2,122 109.5 6.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/00zr3ujx",
   "source_url": "https://omlx.ai/benchmarks/performance/00zr3ujx",
   "tg_tokens": null,
   "tps": 109.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-155s8qwu-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: l0jhsh2z; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1977.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.5-4B-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 1,977 164.1 3.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/155s8qwu",
   "source_url": "https://omlx.ai/benchmarks/performance/155s8qwu",
   "tg_tokens": null,
   "tps": 164.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-155s8qwu-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.7 GB at this context; duplicate traces of the same session: l0jhsh2z; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2159.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.5-4B-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 2,159 158.4 4.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/155s8qwu",
   "source_url": "https://omlx.ai/benchmarks/performance/155s8qwu",
   "tg_tokens": null,
   "tps": 158.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-zrpq3m17-ctx65536",
   "model": "GLM-5.3-Flash-Abliterated",
   "notes": "MLX; oMLX community benchmark trace, single run at 64k context; peak memory 182.6 GB; TTFT 185827 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 352.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-Abliterated-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k tokens 352.7 PP tok/s 13.6 TG tok/s 185827 TTFT (ms) 182.6 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zrpq3m17",
   "source_url": "https://omlx.ai/benchmarks/performance/zrpq3m17",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yma3d9oz-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 21.0 GB; TTFT 255 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4012.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 4,012 PP tok/s 100.8 TG tok/s 255 TTFT (ms) 21.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yma3d9oz",
   "source_url": "https://omlx.ai/benchmarks/performance/yma3d9oz",
   "tg_tokens": null,
   "tps": 100.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jch1ws9i-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V13-dequantized",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.5 GB at this context; duplicate traces of the same session: dbgy25ld; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1947.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V13-dequantized-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 1,947 152.9 37.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jch1ws9i",
   "source_url": "https://omlx.ai/benchmarks/performance/jch1ws9i",
   "tg_tokens": null,
   "tps": 152.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jch1ws9i-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V13-dequantized",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.1 GB at this context; duplicate traces of the same session: dbgy25ld; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2382.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V13-dequantized-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 2,382 146.2 39.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jch1ws9i",
   "source_url": "https://omlx.ai/benchmarks/performance/jch1ws9i",
   "tg_tokens": null,
   "tps": 146.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-irnnc7ol-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 7fixzw2q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1245.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,245 86.9 23.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/irnnc7ol",
   "source_url": "https://omlx.ai/benchmarks/performance/irnnc7ol",
   "tg_tokens": null,
   "tps": 86.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-irnnc7ol-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 7fixzw2q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1064.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,064 90.4 23.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/irnnc7ol",
   "source_url": "https://omlx.ai/benchmarks/performance/irnnc7ol",
   "tg_tokens": null,
   "tps": 90.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-q5gs2tlj-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: p6i33pui; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1177.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,177 60.5 23.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q5gs2tlj",
   "source_url": "https://omlx.ai/benchmarks/performance/q5gs2tlj",
   "tg_tokens": null,
   "tps": 60.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-q5gs2tlj-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: p6i33pui; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1066.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,066 90.4 23.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q5gs2tlj",
   "source_url": "https://omlx.ai/benchmarks/performance/q5gs2tlj",
   "tg_tokens": null,
   "tps": 90.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-zn29v6lt-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 91iq451f, j2zgs20j, t0rq2ugq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 946.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 946.2 79.8 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zn29v6lt",
   "source_url": "https://omlx.ai/benchmarks/performance/zn29v6lt",
   "tg_tokens": null,
   "tps": 79.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-zn29v6lt-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 91iq451f, j2zgs20j, t0rq2ugq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1089.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,089 77.6 21.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zn29v6lt",
   "source_url": "https://omlx.ai/benchmarks/performance/zn29v6lt",
   "tg_tokens": null,
   "tps": 77.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-zn29v6lt-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 91iq451f, j2zgs20j, t0rq2ugq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1026.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 1,026 74.3 22.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zn29v6lt",
   "source_url": "https://omlx.ai/benchmarks/performance/zn29v6lt",
   "tg_tokens": null,
   "tps": 74.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-zn29v6lt-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: 91iq451f, j2zgs20j, t0rq2ugq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 918.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 918.3 68.6 22.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zn29v6lt",
   "source_url": "https://omlx.ai/benchmarks/performance/zn29v6lt",
   "tg_tokens": null,
   "tps": 68.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-dv33tqww-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-NEO-CODER",
   "notes": "1M context build; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: s1fk4bmc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 315.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-NEO-CODER-mlx-oQ4e-1M on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 315.4 64.4 19.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dv33tqww",
   "source_url": "https://omlx.ai/benchmarks/performance/dv33tqww",
   "tg_tokens": null,
   "tps": 64.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-dv33tqww-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-NEO-CODER",
   "notes": "1M context build; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: s1fk4bmc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 333.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-NEO-CODER-mlx-oQ4e-1M on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 333.6 63.4 22.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dv33tqww",
   "source_url": "https://omlx.ai/benchmarks/performance/dv33tqww",
   "tg_tokens": null,
   "tps": 63.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jwfs3ye4-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: nfbsd7hw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1035.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 1,035 56.0 17.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jwfs3ye4",
   "source_url": "https://omlx.ai/benchmarks/performance/jwfs3ye4",
   "tg_tokens": null,
   "tps": 56.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jwfs3ye4-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: nfbsd7hw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1351.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,351 54.0 19.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jwfs3ye4",
   "source_url": "https://omlx.ai/benchmarks/performance/jwfs3ye4",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-6tktkv2o-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.9 GB at this context; duplicate traces of the same session: lt3l3cro; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 773.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 8k 773.9 58.7 101.9 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6tktkv2o",
   "source_url": "https://omlx.ai/benchmarks/performance/6tktkv2o",
   "tg_tokens": null,
   "tps": 58.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-6tktkv2o-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.8 GB at this context; duplicate traces of the same session: lt3l3cro; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 771.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 32k 771.5 45.9 102.8 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6tktkv2o",
   "source_url": "https://omlx.ai/benchmarks/performance/6tktkv2o",
   "tg_tokens": null,
   "tps": 45.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-s11ur8rn-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 101.7 GB; TTFT 10571 ms; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 775.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 8k tokens 775.0 PP tok/s 51.8 TG tok/s 10571 TTFT (ms) 101.7 Peak mem",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s11ur8rn",
   "source_url": "https://omlx.ai/benchmarks/performance/s11ur8rn",
   "tg_tokens": null,
   "tps": 51.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-pwy8g6q9-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 101.7 GB; TTFT 10571 ms; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 775.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 8k tokens 775.0 PP tok/s 60.0 TG tok/s 10571 TTFT (ms) 101.7 Peak mem",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pwy8g6q9",
   "source_url": "https://omlx.ai/benchmarks/performance/pwy8g6q9",
   "tg_tokens": null,
   "tps": 60.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-kahnoc9b-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.7 GB at this context; duplicate traces of the same session: tmglnxj7; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 775.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 8k 775.2 60.1 101.7 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kahnoc9b",
   "source_url": "https://omlx.ai/benchmarks/performance/kahnoc9b",
   "tg_tokens": null,
   "tps": 60.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-kahnoc9b-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.8 GB at this context; duplicate traces of the same session: tmglnxj7; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 773.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 32k 773.8 49.1 102.8 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kahnoc9b",
   "source_url": "https://omlx.ai/benchmarks/performance/kahnoc9b",
   "tg_tokens": null,
   "tps": 49.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1v1g0x33-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: l73t8scl; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2310.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26b-a4b-it-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 2,310 96.2 27.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1v1g0x33",
   "source_url": "https://omlx.ai/benchmarks/performance/1v1g0x33",
   "tg_tokens": null,
   "tps": 96.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1v1g0x33-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: l73t8scl; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2316.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26b-a4b-it-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 2,316 99.9 28.0 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1v1g0x33",
   "source_url": "https://omlx.ai/benchmarks/performance/1v1g0x33",
   "tg_tokens": null,
   "tps": 99.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-pbdy088h-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "ornith-ai repo; oMLX community benchmark trace, single run at 1k context; peak memory 17.5 GB; TTFT 4109 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 249.2,
   "provenance": "community",
   "quant": "unknown",
   "quote": "ornith-ai--Ornith-1.5-9B-MLX on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 unknown; 1k tokens 249.2 PP tok/s 10.1 TG tok/s 4109 TTFT (ms) 17.5 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pbdy088h",
   "source_url": "https://omlx.ai/benchmarks/performance/pbdy088h",
   "tg_tokens": null,
   "tps": 10.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-og7fgqvj-ctx1024",
   "model": "GLM-4.7-Flash",
   "notes": "oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: 807yt0zt; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 572.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-4.7-Flash-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 572.0 66.8 16.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/og7fgqvj",
   "source_url": "https://omlx.ai/benchmarks/performance/og7fgqvj",
   "tg_tokens": null,
   "tps": 66.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-og7fgqvj-ctx4096",
   "model": "GLM-4.7-Flash",
   "notes": "oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 807yt0zt; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1101.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-4.7-Flash-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 1,101 55.7 17.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/og7fgqvj",
   "source_url": "https://omlx.ai/benchmarks/performance/og7fgqvj",
   "tg_tokens": null,
   "tps": 55.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-mumd4ume-ctx1024",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: c9i0i2no; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4153.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-mlx-mxfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 4,153 131.6 6.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mumd4ume",
   "source_url": "https://omlx.ai/benchmarks/performance/mumd4ume",
   "tg_tokens": null,
   "tps": 131.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-mumd4ume-ctx4096",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: c9i0i2no; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3827.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-mlx-mxfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,827 125.1 6.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mumd4ume",
   "source_url": "https://omlx.ai/benchmarks/performance/mumd4ume",
   "tg_tokens": null,
   "tps": 125.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-gkcqths0-ctx1024",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: 3qo0h5ko; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4163.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-mlx-mxfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 4,163 134.5 6.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gkcqths0",
   "source_url": "https://omlx.ai/benchmarks/performance/gkcqths0",
   "tg_tokens": null,
   "tps": 134.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-gkcqths0-ctx4096",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: 3qo0h5ko; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4044.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-mlx-mxfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 4,044 129.8 6.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gkcqths0",
   "source_url": "https://omlx.ai/benchmarks/performance/gkcqths0",
   "tg_tokens": null,
   "tps": 129.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-uz09rbv1-ctx1024",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: hblarq1s; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 730.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-mlx-mxfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 730.0 83.6 6.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uz09rbv1",
   "source_url": "https://omlx.ai/benchmarks/performance/uz09rbv1",
   "tg_tokens": null,
   "tps": 83.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-uz09rbv1-ctx4096",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: hblarq1s; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1745.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-mlx-mxfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,745 76.5 6.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uz09rbv1",
   "source_url": "https://omlx.ai/benchmarks/performance/uz09rbv1",
   "tg_tokens": null,
   "tps": 76.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-1h0odsrg-ctx1024",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "MLX; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 0yoseyxv; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2056.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 2,056 92.0 16.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1h0odsrg",
   "source_url": "https://omlx.ai/benchmarks/performance/1h0odsrg",
   "tg_tokens": null,
   "tps": 92.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-1h0odsrg-ctx4096",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "MLX; oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 0yoseyxv; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2136.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,136 79.6 17.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1h0odsrg",
   "source_url": "https://omlx.ai/benchmarks/performance/1h0odsrg",
   "tg_tokens": null,
   "tps": 79.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-020kfcbf-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: h4qoz3n1, rokbbip2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1560.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 1,560 131.0 29.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/020kfcbf",
   "source_url": "https://omlx.ai/benchmarks/performance/020kfcbf",
   "tg_tokens": null,
   "tps": 131.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-020kfcbf-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.6 GB at this context; duplicate traces of the same session: h4qoz3n1, rokbbip2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3619.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 3,619 187.2 30.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/020kfcbf",
   "source_url": "https://omlx.ai/benchmarks/performance/020kfcbf",
   "tg_tokens": null,
   "tps": 187.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-020kfcbf-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.9 GB at this context; duplicate traces of the same session: h4qoz3n1, rokbbip2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4573.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 4,573 113.7 30.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/020kfcbf",
   "source_url": "https://omlx.ai/benchmarks/performance/020kfcbf",
   "tg_tokens": null,
   "tps": 113.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-90bdbc8h-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 13.2 GB; TTFT 3340 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 306.6,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.8-27B-3bit-mlx on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 3bit; 1k tokens 306.6 PP tok/s 11.5 TG tok/s 3340 TTFT (ms) 13.2 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/90bdbc8h",
   "source_url": "https://omlx.ai/benchmarks/performance/90bdbc8h",
   "tg_tokens": null,
   "tps": 11.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3wlrgqn5-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 53r36lfz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 504.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 504.8 17.7 17.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3wlrgqn5",
   "source_url": "https://omlx.ai/benchmarks/performance/3wlrgqn5",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3wlrgqn5-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 53r36lfz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 491.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 491.9 16.8 17.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3wlrgqn5",
   "source_url": "https://omlx.ai/benchmarks/performance/3wlrgqn5",
   "tg_tokens": null,
   "tps": 16.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-t0rjmmhl-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP320",
   "notes": "MTP Vision build; oMLX community benchmark trace; peak memory 37.7 GB at this context; duplicate traces of the same session: bba77gs4, gnm9v3g3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 311.0,
   "provenance": "community",
   "quant": "oQ3e-DWQ",
   "quote": "Qwen3.8-Flash-Next-REAP320-oQ3e-DWQ-MTP-Vision-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 3bit; 1k 311.0 33.3 37.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t0rjmmhl",
   "source_url": "https://omlx.ai/benchmarks/performance/t0rjmmhl",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-t0rjmmhl-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP320",
   "notes": "MTP Vision build; oMLX community benchmark trace; peak memory 38.2 GB at this context; duplicate traces of the same session: bba77gs4, gnm9v3g3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 345.3,
   "provenance": "community",
   "quant": "oQ3e-DWQ",
   "quote": "Qwen3.8-Flash-Next-REAP320-oQ3e-DWQ-MTP-Vision-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 3bit; 4k 345.3 32.1 38.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t0rjmmhl",
   "source_url": "https://omlx.ai/benchmarks/performance/t0rjmmhl",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-t0rjmmhl-ctx8192",
   "model": "Qwen3.8-Flash-Next-REAP320",
   "notes": "MTP Vision build; oMLX community benchmark trace; peak memory 38.4 GB at this context; duplicate traces of the same session: bba77gs4, gnm9v3g3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 380.0,
   "provenance": "community",
   "quant": "oQ3e-DWQ",
   "quote": "Qwen3.8-Flash-Next-REAP320-oQ3e-DWQ-MTP-Vision-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 3bit; 8k 380.0 37.6 38.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t0rjmmhl",
   "source_url": "https://omlx.ai/benchmarks/performance/t0rjmmhl",
   "tg_tokens": null,
   "tps": 37.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-a5ni336p-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Serve build; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: l5x6w830; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1746.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3___6-35B-A3B-MLX-Serve-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,746 115.9 19.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a5ni336p",
   "source_url": "https://omlx.ai/benchmarks/performance/a5ni336p",
   "tg_tokens": null,
   "tps": 115.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-a5ni336p-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Serve build; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: l5x6w830; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1991.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3___6-35B-A3B-MLX-Serve-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,991 115.6 21.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a5ni336p",
   "source_url": "https://omlx.ai/benchmarks/performance/a5ni336p",
   "tg_tokens": null,
   "tps": 115.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-m67bgqqj-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 1riqoa4i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1546.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,546 110.4 20.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m67bgqqj",
   "source_url": "https://omlx.ai/benchmarks/performance/m67bgqqj",
   "tg_tokens": null,
   "tps": 110.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-m67bgqqj-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 1riqoa4i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1932.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,932 109.6 21.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m67bgqqj",
   "source_url": "https://omlx.ai/benchmarks/performance/m67bgqqj",
   "tg_tokens": null,
   "tps": 109.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-b4qxkc8d-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 9.9 GB at this context; duplicate traces of the same session: cybt990x, sc8s8xwd; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1218.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B-MLX-8bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 8bit; 1k 1,218 32.6 9.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b4qxkc8d",
   "source_url": "https://omlx.ai/benchmarks/performance/b4qxkc8d",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-b4qxkc8d-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 10.5 GB at this context; duplicate traces of the same session: cybt990x, sc8s8xwd; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1384.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B-MLX-8bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 8bit; 4k 1,384 32.2 10.5 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b4qxkc8d",
   "source_url": "https://omlx.ai/benchmarks/performance/b4qxkc8d",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-b4qxkc8d-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 10.6 GB at this context; duplicate traces of the same session: cybt990x, sc8s8xwd; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1412.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B-MLX-8bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 8bit; 8k 1,412 31.6 10.6 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b4qxkc8d",
   "source_url": "https://omlx.ai/benchmarks/performance/b4qxkc8d",
   "tg_tokens": null,
   "tps": 31.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-cjrantst-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: 0dxu48ba, 4viba3ha, 7s3vi9gm, aeutny77, gn6o9mky, ihgdew01, o1sjztjk; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 331.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 331.0 45.1 32.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cjrantst",
   "source_url": "https://omlx.ai/benchmarks/performance/cjrantst",
   "tg_tokens": null,
   "tps": 45.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-cjrantst-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.0 GB at this context; duplicate traces of the same session: 0dxu48ba, 4viba3ha, 7s3vi9gm, aeutny77, gn6o9mky, ihgdew01, o1sjztjk; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 335.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 335.4 38.8 34.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cjrantst",
   "source_url": "https://omlx.ai/benchmarks/performance/cjrantst",
   "tg_tokens": null,
   "tps": 38.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-cjrantst-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 35.3 GB at this context; duplicate traces of the same session: 0dxu48ba, 4viba3ha, 7s3vi9gm, aeutny77, gn6o9mky, ihgdew01, o1sjztjk; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 330.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 8k 330.3 34.1 35.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cjrantst",
   "source_url": "https://omlx.ai/benchmarks/performance/cjrantst",
   "tg_tokens": null,
   "tps": 34.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-cjrantst-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 36.0 GB at this context; duplicate traces of the same session: 0dxu48ba, 4viba3ha, 7s3vi9gm, aeutny77, gn6o9mky, ihgdew01, o1sjztjk; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 315.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 16k 315.1 13.0 36.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cjrantst",
   "source_url": "https://omlx.ai/benchmarks/performance/cjrantst",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-cjrantst-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 38.5 GB at this context; duplicate traces of the same session: 0dxu48ba, 4viba3ha, 7s3vi9gm, aeutny77, gn6o9mky, ihgdew01, o1sjztjk; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 290.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 32k 290.0 8.1 38.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cjrantst",
   "source_url": "https://omlx.ai/benchmarks/performance/cjrantst",
   "tg_tokens": null,
   "tps": 8.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-cjrantst-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 43.3 GB at this context; duplicate traces of the same session: 0dxu48ba, 4viba3ha, 7s3vi9gm, aeutny77, gn6o9mky, ihgdew01, o1sjztjk; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 249.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 64k 249.5 2.7 43.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cjrantst",
   "source_url": "https://omlx.ai/benchmarks/performance/cjrantst",
   "tg_tokens": null,
   "tps": 2.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-cjrantst-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 51.6 GB at this context; duplicate traces of the same session: 0dxu48ba, 4viba3ha, 7s3vi9gm, aeutny77, gn6o9mky, ihgdew01, o1sjztjk; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 194.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 128k 194.4 2.4 51.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cjrantst",
   "source_url": "https://omlx.ai/benchmarks/performance/cjrantst",
   "tg_tokens": null,
   "tps": 2.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-cjrantst-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 47.7 GB at this context; duplicate traces of the same session: 0dxu48ba, 4viba3ha, 7s3vi9gm, aeutny77, gn6o9mky, ihgdew01, o1sjztjk; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 157.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 195k 157.7 0.8 47.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cjrantst",
   "source_url": "https://omlx.ai/benchmarks/performance/cjrantst",
   "tg_tokens": null,
   "tps": 0.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-xh91gpdr-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: 7mqn9bic; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 477.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 477.4 28.0 18.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xh91gpdr",
   "source_url": "https://omlx.ai/benchmarks/performance/xh91gpdr",
   "tg_tokens": null,
   "tps": 28.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-xh91gpdr-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: 7mqn9bic; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 465.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 465.7 29.1 18.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xh91gpdr",
   "source_url": "https://omlx.ai/benchmarks/performance/xh91gpdr",
   "tg_tokens": null,
   "tps": 29.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-rmztnabk-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: tnf0r82c; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 64.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k 64.1 11.6 18.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rmztnabk",
   "source_url": "https://omlx.ai/benchmarks/performance/rmztnabk",
   "tg_tokens": null,
   "tps": 11.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-rmztnabk-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: tnf0r82c; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 65.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 65.8 12.3 19.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rmztnabk",
   "source_url": "https://omlx.ai/benchmarks/performance/rmztnabk",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bpjq7d85-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: 963kdhpz, ahqptz1u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 767.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 767.5 23.9 33.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bpjq7d85",
   "source_url": "https://omlx.ai/benchmarks/performance/bpjq7d85",
   "tg_tokens": null,
   "tps": 23.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bpjq7d85-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: 963kdhpz, ahqptz1u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 743.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 743.5 22.8 34.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bpjq7d85",
   "source_url": "https://omlx.ai/benchmarks/performance/bpjq7d85",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bpjq7d85-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.7 GB at this context; duplicate traces of the same session: 963kdhpz, ahqptz1u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 618.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 618.9 27.1 35.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bpjq7d85",
   "source_url": "https://omlx.ai/benchmarks/performance/bpjq7d85",
   "tg_tokens": null,
   "tps": 27.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-x5dtvwg0-ctx199680",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 195k context; peak memory 52.3 GB; TTFT 231890 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 862.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 195k tokens 862.5 PP tok/s 45.1 TG tok/s 231890 TTFT (ms) 52.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x5dtvwg0",
   "source_url": "https://omlx.ai/benchmarks/performance/x5dtvwg0",
   "tg_tokens": null,
   "tps": 45.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ft6wyc1z-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: f65br1sr; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 477.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 477.1 28.6 18.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ft6wyc1z",
   "source_url": "https://omlx.ai/benchmarks/performance/ft6wyc1z",
   "tg_tokens": null,
   "tps": 28.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ft6wyc1z-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: f65br1sr; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 463.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 463.4 27.8 18.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ft6wyc1z",
   "source_url": "https://omlx.ai/benchmarks/performance/ft6wyc1z",
   "tg_tokens": null,
   "tps": 27.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-vgs4ud15-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.2 GB at this context; duplicate traces of the same session: l4awv3ex, s2329pew; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 249.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 249.2 39.1 32.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vgs4ud15",
   "source_url": "https://omlx.ai/benchmarks/performance/vgs4ud15",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-vgs4ud15-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.7 GB at this context; duplicate traces of the same session: l4awv3ex, s2329pew; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 253.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 253.7 30.6 33.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vgs4ud15",
   "source_url": "https://omlx.ai/benchmarks/performance/vgs4ud15",
   "tg_tokens": null,
   "tps": 30.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-vgs4ud15-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: l4awv3ex, s2329pew; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 249.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 249.7 35.3 34.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vgs4ud15",
   "source_url": "https://omlx.ai/benchmarks/performance/vgs4ud15",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-0cqljcs3-ctx8192",
   "model": "GLM-5.3-Flash-CYBER",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.2 GB at this context; duplicate traces of the same session: w8omzgyj; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1189.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-CYBER-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 1,189 51.6 179.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0cqljcs3",
   "source_url": "https://omlx.ai/benchmarks/performance/0cqljcs3",
   "tg_tokens": null,
   "tps": 51.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-0cqljcs3-ctx32768",
   "model": "GLM-5.3-Flash-CYBER",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.6 GB at this context; duplicate traces of the same session: w8omzgyj; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1894.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-CYBER-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,894 54.9 179.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0cqljcs3",
   "source_url": "https://omlx.ai/benchmarks/performance/0cqljcs3",
   "tg_tokens": null,
   "tps": 54.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-rwvfv1iy-ctx1024",
   "model": "GLM-5.3-Flash-CYBER",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 177.5 GB; TTFT 1575 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 650.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-CYBER-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k tokens 650.2 PP tok/s 66.4 TG tok/s 1575 TTFT (ms) 177.5 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rwvfv1iy",
   "source_url": "https://omlx.ai/benchmarks/performance/rwvfv1iy",
   "tg_tokens": null,
   "tps": 66.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-a46g9eql-ctx16384",
   "model": "MiniCPM5-1B",
   "notes": "oMLX community benchmark trace; peak memory 1.4 GB at this context; duplicate traces of the same session: 06kc6cy2, koqhj0o4, o3jcjjkp; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4547.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM5-1B-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 16k 4,547 105.9 1.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a46g9eql",
   "source_url": "https://omlx.ai/benchmarks/performance/a46g9eql",
   "tg_tokens": null,
   "tps": 105.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-a46g9eql-ctx32768",
   "model": "MiniCPM5-1B",
   "notes": "oMLX community benchmark trace; peak memory 1.8 GB at this context; duplicate traces of the same session: 06kc6cy2, koqhj0o4, o3jcjjkp; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2972.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM5-1B-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 32k 2,972 88.7 1.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a46g9eql",
   "source_url": "https://omlx.ai/benchmarks/performance/a46g9eql",
   "tg_tokens": null,
   "tps": 88.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-a46g9eql-ctx65536",
   "model": "MiniCPM5-1B",
   "notes": "oMLX community benchmark trace; peak memory 2.6 GB at this context; duplicate traces of the same session: 06kc6cy2, koqhj0o4, o3jcjjkp; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1687.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM5-1B-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 64k 1,687 55.8 2.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a46g9eql",
   "source_url": "https://omlx.ai/benchmarks/performance/a46g9eql",
   "tg_tokens": null,
   "tps": 55.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-a46g9eql-ctx131072",
   "model": "MiniCPM5-1B",
   "notes": "oMLX community benchmark trace; peak memory 4.0 GB at this context; duplicate traces of the same session: 06kc6cy2, koqhj0o4, o3jcjjkp; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 886.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM5-1B-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 128k 886.8 32.5 4.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a46g9eql",
   "source_url": "https://omlx.ai/benchmarks/performance/a46g9eql",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-yvtxjii9-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: ro7jcju1; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 532.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 532.2 62.5 20.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yvtxjii9",
   "source_url": "https://omlx.ai/benchmarks/performance/yvtxjii9",
   "tg_tokens": null,
   "tps": 62.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-yvtxjii9-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: ro7jcju1; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 582.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 582.3 58.4 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yvtxjii9",
   "source_url": "https://omlx.ai/benchmarks/performance/yvtxjii9",
   "tg_tokens": null,
   "tps": 58.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-tkao2b4i-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: lnk5tm0x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 259.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 259.2 29.2 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tkao2b4i",
   "source_url": "https://omlx.ai/benchmarks/performance/tkao2b4i",
   "tg_tokens": null,
   "tps": 29.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-tkao2b4i-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: lnk5tm0x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 211.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 211.7 26.5 19.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tkao2b4i",
   "source_url": "https://omlx.ai/benchmarks/performance/tkao2b4i",
   "tg_tokens": null,
   "tps": 26.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ydc55cy9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: 5uz85ids, w3t2mk5l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 247.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 247.7 42.8 32.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ydc55cy9",
   "source_url": "https://omlx.ai/benchmarks/performance/ydc55cy9",
   "tg_tokens": null,
   "tps": 42.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ydc55cy9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.7 GB at this context; duplicate traces of the same session: 5uz85ids, w3t2mk5l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 253.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 253.2 28.2 33.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ydc55cy9",
   "source_url": "https://omlx.ai/benchmarks/performance/ydc55cy9",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ydc55cy9-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: 5uz85ids, w3t2mk5l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 249.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 249.5 37.8 34.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ydc55cy9",
   "source_url": "https://omlx.ai/benchmarks/performance/ydc55cy9",
   "tg_tokens": null,
   "tps": 37.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-dattq175-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: 7vm9bzox, bj775u7a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 247.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 247.6 35.8 29.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dattq175",
   "source_url": "https://omlx.ai/benchmarks/performance/dattq175",
   "tg_tokens": null,
   "tps": 35.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-dattq175-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.1 GB at this context; duplicate traces of the same session: 7vm9bzox, bj775u7a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 252.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 252.6 33.3 31.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dattq175",
   "source_url": "https://omlx.ai/benchmarks/performance/dattq175",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-dattq175-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: 7vm9bzox, bj775u7a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 248.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 248.8 40.2 32.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dattq175",
   "source_url": "https://omlx.ai/benchmarks/performance/dattq175",
   "tg_tokens": null,
   "tps": 40.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-eo4tsp87-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: 2a5xzo7d, 7mcduwrl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 248.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 248.6 38.6 29.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eo4tsp87",
   "source_url": "https://omlx.ai/benchmarks/performance/eo4tsp87",
   "tg_tokens": null,
   "tps": 38.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-eo4tsp87-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.1 GB at this context; duplicate traces of the same session: 2a5xzo7d, 7mcduwrl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 253.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 253.1 32.6 31.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eo4tsp87",
   "source_url": "https://omlx.ai/benchmarks/performance/eo4tsp87",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-eo4tsp87-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: 2a5xzo7d, 7mcduwrl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 249.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 249.5 33.6 32.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eo4tsp87",
   "source_url": "https://omlx.ai/benchmarks/performance/eo4tsp87",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-pyypdpw7-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: 2riiqchr, h2c7zzfa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 247.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 247.0 31.5 29.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pyypdpw7",
   "source_url": "https://omlx.ai/benchmarks/performance/pyypdpw7",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-pyypdpw7-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.1 GB at this context; duplicate traces of the same session: 2riiqchr, h2c7zzfa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 252.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 252.6 31.9 31.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pyypdpw7",
   "source_url": "https://omlx.ai/benchmarks/performance/pyypdpw7",
   "tg_tokens": null,
   "tps": 31.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-pyypdpw7-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: 2riiqchr, h2c7zzfa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 249.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 249.2 36.5 32.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pyypdpw7",
   "source_url": "https://omlx.ai/benchmarks/performance/pyypdpw7",
   "tg_tokens": null,
   "tps": 36.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-em2nmazd-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.3 GB at this context; duplicate traces of the same session: 8c53s3xc, sbpa04qn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 249.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 249.6 16.5 28.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/em2nmazd",
   "source_url": "https://omlx.ai/benchmarks/performance/em2nmazd",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-em2nmazd-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.7 GB at this context; duplicate traces of the same session: 8c53s3xc, sbpa04qn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 258.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 258.4 16.4 31.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/em2nmazd",
   "source_url": "https://omlx.ai/benchmarks/performance/em2nmazd",
   "tg_tokens": null,
   "tps": 16.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-em2nmazd-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: 8c53s3xc, sbpa04qn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 254.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 254.8 15.8 32.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/em2nmazd",
   "source_url": "https://omlx.ai/benchmarks/performance/em2nmazd",
   "tg_tokens": null,
   "tps": 15.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-61ybgxqb-ctx16384",
   "model": "gemma-4-31B-it",
   "notes": "mlx-community; oMLX community benchmark trace, single run at 16k context; peak memory 34.7 GB; TTFT 33005 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 496.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mlx-community--gemma-4-31b-it-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k tokens 496.4 PP tok/s 13.5 TG tok/s 33005 TTFT (ms) 34.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/61ybgxqb",
   "source_url": "https://omlx.ai/benchmarks/performance/61ybgxqb",
   "tg_tokens": null,
   "tps": 13.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0gggs06a-ctx16384",
   "model": "gemma-4-31B-it",
   "notes": "mlx-community; oMLX community benchmark trace, single run at 16k context; peak memory 36.6 GB; TTFT 32306 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 507.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mlx-community--gemma-4-31b-it-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k tokens 507.2 PP tok/s 28.6 TG tok/s 32306 TTFT (ms) 36.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0gggs06a",
   "source_url": "https://omlx.ai/benchmarks/performance/0gggs06a",
   "tg_tokens": null,
   "tps": 28.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-npirabpw-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: yekk78b9; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 535.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 535.4 31.7 21.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/npirabpw",
   "source_url": "https://omlx.ai/benchmarks/performance/npirabpw",
   "tg_tokens": null,
   "tps": 31.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-npirabpw-ctx131072",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: yekk78b9; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 349.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 128k 349.3 18.5 23.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/npirabpw",
   "source_url": "https://omlx.ai/benchmarks/performance/npirabpw",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-95c76uc1-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 2nfgiqir, 2pjc6kvw, r8i5nf9m, zb0pzpeo; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1760.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 1,760 100.4 21.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/95c76uc1",
   "source_url": "https://omlx.ai/benchmarks/performance/95c76uc1",
   "tg_tokens": null,
   "tps": 100.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-95c76uc1-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 2nfgiqir, 2pjc6kvw, r8i5nf9m, zb0pzpeo; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1985.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,985 79.6 22.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/95c76uc1",
   "source_url": "https://omlx.ai/benchmarks/performance/95c76uc1",
   "tg_tokens": null,
   "tps": 79.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-95c76uc1-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 2nfgiqir, 2pjc6kvw, r8i5nf9m, zb0pzpeo; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3242.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 3,242 96.7 22.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/95c76uc1",
   "source_url": "https://omlx.ai/benchmarks/performance/95c76uc1",
   "tg_tokens": null,
   "tps": 96.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-95c76uc1-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 2nfgiqir, 2pjc6kvw, r8i5nf9m, zb0pzpeo; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3022.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 3,022 91.8 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/95c76uc1",
   "source_url": "https://omlx.ai/benchmarks/performance/95c76uc1",
   "tg_tokens": null,
   "tps": 91.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-95c76uc1-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 2nfgiqir, 2pjc6kvw, r8i5nf9m, zb0pzpeo; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2462.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 2,462 57.4 22.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/95c76uc1",
   "source_url": "https://omlx.ai/benchmarks/performance/95c76uc1",
   "tg_tokens": null,
   "tps": 57.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-sv1zn3a2-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: zq2f34ee; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 127.3 15.3 15.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sv1zn3a2",
   "source_url": "https://omlx.ai/benchmarks/performance/sv1zn3a2",
   "tg_tokens": null,
   "tps": 15.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-sv1zn3a2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: zq2f34ee; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 129.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 129.2 14.2 17.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sv1zn3a2",
   "source_url": "https://omlx.ai/benchmarks/performance/sv1zn3a2",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-7s2oa45s-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.5 GB at this context; duplicate traces of the same session: 08m8271l, 8qi4qdor; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 239.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 239.9 40.2 34.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7s2oa45s",
   "source_url": "https://omlx.ai/benchmarks/performance/7s2oa45s",
   "tg_tokens": null,
   "tps": 40.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-7s2oa45s-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: 08m8271l, 8qi4qdor; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 222.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 222.8 37.7 37.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7s2oa45s",
   "source_url": "https://omlx.ai/benchmarks/performance/7s2oa45s",
   "tg_tokens": null,
   "tps": 37.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-7s2oa45s-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 42.7 GB at this context; duplicate traces of the same session: 08m8271l, 8qi4qdor; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 194.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 194.6 28.5 42.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7s2oa45s",
   "source_url": "https://omlx.ai/benchmarks/performance/7s2oa45s",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-cnww7g8j-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 21.1 GB; TTFT 582 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1760.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 1,760 PP tok/s 98.9 TG tok/s 582 TTFT (ms) 21.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cnww7g8j",
   "source_url": "https://omlx.ai/benchmarks/performance/cnww7g8j",
   "tg_tokens": null,
   "tps": 98.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-fsgt1m6d-ctx1024",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 157.0 GB at this context; duplicate traces of the same session: f50cps4z, qqaov9id, v6ypsfvr, wxaoq8dp, xgbklk2z; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 420.9,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "DeepSeek-V4-Flash-0731-MXFP4-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 420.9 39.2 157.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fsgt1m6d",
   "source_url": "https://omlx.ai/benchmarks/performance/fsgt1m6d",
   "tg_tokens": null,
   "tps": 39.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-fsgt1m6d-ctx4096",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 157.7 GB at this context; duplicate traces of the same session: f50cps4z, qqaov9id, v6ypsfvr, wxaoq8dp, xgbklk2z; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 561.7,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "DeepSeek-V4-Flash-0731-MXFP4-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 561.7 40.8 157.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fsgt1m6d",
   "source_url": "https://omlx.ai/benchmarks/performance/fsgt1m6d",
   "tg_tokens": null,
   "tps": 40.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-fsgt1m6d-ctx16384",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 159.1 GB at this context; duplicate traces of the same session: f50cps4z, qqaov9id, v6ypsfvr, wxaoq8dp, xgbklk2z; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 598.3,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "DeepSeek-V4-Flash-0731-MXFP4-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 598.3 42.6 159.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fsgt1m6d",
   "source_url": "https://omlx.ai/benchmarks/performance/fsgt1m6d",
   "tg_tokens": null,
   "tps": 42.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-fsgt1m6d-ctx65536",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 164.9 GB at this context; duplicate traces of the same session: f50cps4z, qqaov9id, v6ypsfvr, wxaoq8dp, xgbklk2z; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 552.3,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "DeepSeek-V4-Flash-0731-MXFP4-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 552.3 41.3 164.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fsgt1m6d",
   "source_url": "https://omlx.ai/benchmarks/performance/fsgt1m6d",
   "tg_tokens": null,
   "tps": 41.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-fsgt1m6d-ctx131072",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 172.4 GB at this context; duplicate traces of the same session: f50cps4z, qqaov9id, v6ypsfvr, wxaoq8dp, xgbklk2z; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 500.7,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "DeepSeek-V4-Flash-0731-MXFP4-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 500.7 35.7 172.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fsgt1m6d",
   "source_url": "https://omlx.ai/benchmarks/performance/fsgt1m6d",
   "tg_tokens": null,
   "tps": 35.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-fsgt1m6d-ctx199680",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 181.2 GB at this context; duplicate traces of the same session: f50cps4z, qqaov9id, v6ypsfvr, wxaoq8dp, xgbklk2z; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 455.9,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "DeepSeek-V4-Flash-0731-MXFP4-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 195k 455.9 31.7 181.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fsgt1m6d",
   "source_url": "https://omlx.ai/benchmarks/performance/fsgt1m6d",
   "tg_tokens": null,
   "tps": 31.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-7ti6wdly-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: 3i1xdsuy, cd7lz5rt, gr6baurb, qlbt2449, ufmn2n16, yiw1mxvg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 889.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 889.7 90.7 5.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ti6wdly",
   "source_url": "https://omlx.ai/benchmarks/performance/7ti6wdly",
   "tg_tokens": null,
   "tps": 90.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-7ti6wdly-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.2 GB at this context; duplicate traces of the same session: 3i1xdsuy, cd7lz5rt, gr6baurb, qlbt2449, ufmn2n16, yiw1mxvg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 982.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 982.0 88.8 7.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ti6wdly",
   "source_url": "https://omlx.ai/benchmarks/performance/7ti6wdly",
   "tg_tokens": null,
   "tps": 88.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-7ti6wdly-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.3 GB at this context; duplicate traces of the same session: 3i1xdsuy, cd7lz5rt, gr6baurb, qlbt2449, ufmn2n16, yiw1mxvg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 962.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 962.7 86.6 7.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ti6wdly",
   "source_url": "https://omlx.ai/benchmarks/performance/7ti6wdly",
   "tg_tokens": null,
   "tps": 86.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-7ti6wdly-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.8 GB at this context; duplicate traces of the same session: 3i1xdsuy, cd7lz5rt, gr6baurb, qlbt2449, ufmn2n16, yiw1mxvg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 923.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 923.7 81.5 7.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ti6wdly",
   "source_url": "https://omlx.ai/benchmarks/performance/7ti6wdly",
   "tg_tokens": null,
   "tps": 81.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-7ti6wdly-ctx32768",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.6 GB at this context; duplicate traces of the same session: 3i1xdsuy, cd7lz5rt, gr6baurb, qlbt2449, ufmn2n16, yiw1mxvg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 851.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 851.8 73.6 8.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ti6wdly",
   "source_url": "https://omlx.ai/benchmarks/performance/7ti6wdly",
   "tg_tokens": null,
   "tps": 73.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-7ti6wdly-ctx65536",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.3 GB at this context; duplicate traces of the same session: 3i1xdsuy, cd7lz5rt, gr6baurb, qlbt2449, ufmn2n16, yiw1mxvg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 735.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 735.7 62.1 10.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ti6wdly",
   "source_url": "https://omlx.ai/benchmarks/performance/7ti6wdly",
   "tg_tokens": null,
   "tps": 62.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-7ti6wdly-ctx131072",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 13.0 GB at this context; duplicate traces of the same session: 3i1xdsuy, cd7lz5rt, gr6baurb, qlbt2449, ufmn2n16, yiw1mxvg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 577.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 577.9 47.1 13.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ti6wdly",
   "source_url": "https://omlx.ai/benchmarks/performance/7ti6wdly",
   "tg_tokens": null,
   "tps": 47.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-zs8rxruq-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 34mjpg22; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 324.1,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 324.1 43.5 21.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zs8rxruq",
   "source_url": "https://omlx.ai/benchmarks/performance/zs8rxruq",
   "tg_tokens": null,
   "tps": 43.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-zs8rxruq-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: 34mjpg22; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 909.5,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 909.5 42.4 22.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zs8rxruq",
   "source_url": "https://omlx.ai/benchmarks/performance/zs8rxruq",
   "tg_tokens": null,
   "tps": 42.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qken10mf-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: c2sl8569; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1881.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 1,881 45.6 21.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qken10mf",
   "source_url": "https://omlx.ai/benchmarks/performance/qken10mf",
   "tg_tokens": null,
   "tps": 45.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qken10mf-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: c2sl8569; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1582.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 1,582 47.1 22.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qken10mf",
   "source_url": "https://omlx.ai/benchmarks/performance/qken10mf",
   "tg_tokens": null,
   "tps": 47.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-mfwcayp2-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "bf16-pt; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: ohncncee; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2056.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-bf16-pt-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 1k 2,056 155.5 25.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mfwcayp2",
   "source_url": "https://omlx.ai/benchmarks/performance/mfwcayp2",
   "tg_tokens": null,
   "tps": 155.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-mfwcayp2-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "bf16-pt; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: ohncncee; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2490.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-bf16-pt-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 4k 2,490 127.6 26.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mfwcayp2",
   "source_url": "https://omlx.ai/benchmarks/performance/mfwcayp2",
   "tg_tokens": null,
   "tps": 127.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-zlkmcspy-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: 2l7n1jvi, 59ij1783, 6167h35b, h4u141lg, qcvl020b; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 123.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 123.0 24.6 17.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zlkmcspy",
   "source_url": "https://omlx.ai/benchmarks/performance/zlkmcspy",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-zlkmcspy-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: 2l7n1jvi, 59ij1783, 6167h35b, h4u141lg, qcvl020b; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 114.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 114.9 22.0 18.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zlkmcspy",
   "source_url": "https://omlx.ai/benchmarks/performance/zlkmcspy",
   "tg_tokens": null,
   "tps": 22.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-zlkmcspy-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 2l7n1jvi, 59ij1783, 6167h35b, h4u141lg, qcvl020b; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 113.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 113.8 22.8 19.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zlkmcspy",
   "source_url": "https://omlx.ai/benchmarks/performance/zlkmcspy",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-zlkmcspy-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 2l7n1jvi, 59ij1783, 6167h35b, h4u141lg, qcvl020b; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 114.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 114.5 13.4 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zlkmcspy",
   "source_url": "https://omlx.ai/benchmarks/performance/zlkmcspy",
   "tg_tokens": null,
   "tps": 13.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-zlkmcspy-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: 2l7n1jvi, 59ij1783, 6167h35b, h4u141lg, qcvl020b; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 108.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 108.8 7.1 24.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zlkmcspy",
   "source_url": "https://omlx.ai/benchmarks/performance/zlkmcspy",
   "tg_tokens": null,
   "tps": 7.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-zlkmcspy-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: 2l7n1jvi, 59ij1783, 6167h35b, h4u141lg, qcvl020b; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 96.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 96.5 3.2 29.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zlkmcspy",
   "source_url": "https://omlx.ai/benchmarks/performance/zlkmcspy",
   "tg_tokens": null,
   "tps": 3.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-6i3ma787-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: b7zm47ua, g844xneu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 257.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 257.3 26.4 20.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6i3ma787",
   "source_url": "https://omlx.ai/benchmarks/performance/6i3ma787",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-6i3ma787-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: b7zm47ua, g844xneu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 239.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 239.8 25.0 22.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6i3ma787",
   "source_url": "https://omlx.ai/benchmarks/performance/6i3ma787",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-6i3ma787-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: b7zm47ua, g844xneu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 210.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 210.0 21.7 24.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6i3ma787",
   "source_url": "https://omlx.ai/benchmarks/performance/6i3ma787",
   "tg_tokens": null,
   "tps": 21.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-0cl24eqd-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: jvm0ry3c; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 77.9,
   "provenance": "community",
   "quant": "oQ2",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ2-MTP on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 2bit; 1k 77.9 8.7 17.8 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0cl24eqd",
   "source_url": "https://omlx.ai/benchmarks/performance/0cl24eqd",
   "tg_tokens": null,
   "tps": 8.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-0cl24eqd-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.2 GB at this context; duplicate traces of the same session: jvm0ry3c; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 129.1,
   "provenance": "community",
   "quant": "oQ2",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ2-MTP on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 2bit; 4k 129.1 7.1 18.2 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0cl24eqd",
   "source_url": "https://omlx.ai/benchmarks/performance/0cl24eqd",
   "tg_tokens": null,
   "tps": 7.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-hroe2bzq-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.2 GB at this context; duplicate traces of the same session: ujuwfhcz; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 70.7,
   "provenance": "community",
   "quant": "oQ2",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ2-MTP on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 2bit; 1k 70.7 7.6 16.2 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hroe2bzq",
   "source_url": "https://omlx.ai/benchmarks/performance/hroe2bzq",
   "tg_tokens": null,
   "tps": 7.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-hroe2bzq-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.5 GB at this context; duplicate traces of the same session: ujuwfhcz; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 126.1,
   "provenance": "community",
   "quant": "oQ2",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ2-MTP on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 2bit; 4k 126.1 8.3 16.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hroe2bzq",
   "source_url": "https://omlx.ai/benchmarks/performance/hroe2bzq",
   "tg_tokens": null,
   "tps": 8.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-hvgjwfzy-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 6.1 GB at this context; duplicate traces of the same session: aynagppt; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 680.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E4B-it-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 1k 680.4 47.1 6.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hvgjwfzy",
   "source_url": "https://omlx.ai/benchmarks/performance/hvgjwfzy",
   "tg_tokens": null,
   "tps": 47.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-hvgjwfzy-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: aynagppt; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 784.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E4B-it-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k 784.1 44.6 6.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hvgjwfzy",
   "source_url": "https://omlx.ai/benchmarks/performance/hvgjwfzy",
   "tg_tokens": null,
   "tps": 44.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-jtff237w-ctx1024",
   "model": "Qwen3-27B",
   "notes": "mlx-community; oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: tenk3avh; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 394.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "qwen3-27b-mlx on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 394.4 11.2 25.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jtff237w",
   "source_url": "https://omlx.ai/benchmarks/performance/jtff237w",
   "tg_tokens": null,
   "tps": 11.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-jtff237w-ctx4096",
   "model": "Qwen3-27B",
   "notes": "mlx-community; oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: tenk3avh; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 419.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "qwen3-27b-mlx on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 419.1 10.8 26.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jtff237w",
   "source_url": "https://omlx.ai/benchmarks/performance/jtff237w",
   "tg_tokens": null,
   "tps": 10.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-96c7eiqb-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: e40h88xi, krsqfy3f; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 797.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 1k 797.1 39.1 28.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/96c7eiqb",
   "source_url": "https://omlx.ai/benchmarks/performance/96c7eiqb",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-96c7eiqb-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.1 GB at this context; duplicate traces of the same session: e40h88xi, krsqfy3f; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 829.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 4k 829.5 31.1 29.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/96c7eiqb",
   "source_url": "https://omlx.ai/benchmarks/performance/96c7eiqb",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-96c7eiqb-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.2 GB at this context; duplicate traces of the same session: e40h88xi, krsqfy3f; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 729.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 16k 729.2 51.8 30.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/96c7eiqb",
   "source_url": "https://omlx.ai/benchmarks/performance/96c7eiqb",
   "tg_tokens": null,
   "tps": 51.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-40gki7hh-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "bf16-pt; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 25.0 GB; TTFT 502 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2041.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-bf16-pt-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 1k tokens 2,041 PP tok/s 155.7 TG tok/s 502 TTFT (ms) 25.0 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/40gki7hh",
   "source_url": "https://omlx.ai/benchmarks/performance/40gki7hh",
   "tg_tokens": null,
   "tps": 155.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-lc6uvdhr-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 6.1 GB at this context; duplicate traces of the same session: kdgdyydn; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 806.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E4B-it-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 1k 806.0 45.2 6.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lc6uvdhr",
   "source_url": "https://omlx.ai/benchmarks/performance/lc6uvdhr",
   "tg_tokens": null,
   "tps": 45.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-lc6uvdhr-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: kdgdyydn; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 767.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E4B-it-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k 767.9 43.6 6.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lc6uvdhr",
   "source_url": "https://omlx.ai/benchmarks/performance/lc6uvdhr",
   "tg_tokens": null,
   "tps": 43.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-k95m7mcp-ctx1024",
   "model": "Nanbeige4.2-3B",
   "notes": "oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: laoqmehw; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 387.2,
   "provenance": "community",
   "quant": "OptiQ 4bit",
   "quote": "Nanbeige4.2-3B-OptiQ-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 1k 387.2 24.3 3.9 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k95m7mcp",
   "source_url": "https://omlx.ai/benchmarks/performance/k95m7mcp",
   "tg_tokens": null,
   "tps": 24.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-k95m7mcp-ctx4096",
   "model": "Nanbeige4.2-3B",
   "notes": "oMLX community benchmark trace; peak memory 4.4 GB at this context; duplicate traces of the same session: laoqmehw; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 352.8,
   "provenance": "community",
   "quant": "OptiQ 4bit",
   "quote": "Nanbeige4.2-3B-OptiQ-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 4k 352.8 21.8 4.4 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k95m7mcp",
   "source_url": "https://omlx.ai/benchmarks/performance/k95m7mcp",
   "tg_tokens": null,
   "tps": 21.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-spumvlj1-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "bf16-pt; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: bctnzs00; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2052.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-bf16-pt-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 1k 2,052 155.3 25.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/spumvlj1",
   "source_url": "https://omlx.ai/benchmarks/performance/spumvlj1",
   "tg_tokens": null,
   "tps": 155.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-spumvlj1-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "bf16-pt; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: bctnzs00; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2484.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-bf16-pt-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 4k 2,484 128.6 26.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/spumvlj1",
   "source_url": "https://omlx.ai/benchmarks/performance/spumvlj1",
   "tg_tokens": null,
   "tps": 128.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1mlxowzv-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "bf16-pt; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: p52xw331; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2050.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-bf16-pt-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 1k 2,050 161.5 25.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1mlxowzv",
   "source_url": "https://omlx.ai/benchmarks/performance/1mlxowzv",
   "tg_tokens": null,
   "tps": 161.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1mlxowzv-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "bf16-pt; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: p52xw331; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2487.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-bf16-pt-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 4k 2,487 116.7 26.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1mlxowzv",
   "source_url": "https://omlx.ai/benchmarks/performance/1mlxowzv",
   "tg_tokens": null,
   "tps": 116.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-3pkyqn2y-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "bf16-pt; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: vjzjiwjc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2055.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-bf16-pt-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 1k 2,055 130.5 25.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3pkyqn2y",
   "source_url": "https://omlx.ai/benchmarks/performance/3pkyqn2y",
   "tg_tokens": null,
   "tps": 130.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3pkyqn2y-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "bf16-pt; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: vjzjiwjc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2489.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-bf16-pt-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 4k 2,489 126.6 26.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3pkyqn2y",
   "source_url": "https://omlx.ai/benchmarks/performance/3pkyqn2y",
   "tg_tokens": null,
   "tps": 126.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-tvrhjpd6-ctx1024",
   "model": "gemma-4-12b-it",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 8.3 GB at this context; duplicate traces of the same session: uhrwozeu; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 121.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-12B-it-oQ4e-fp16-mtp on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 121.5 20.7 8.3 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tvrhjpd6",
   "source_url": "https://omlx.ai/benchmarks/performance/tvrhjpd6",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-tvrhjpd6-ctx4096",
   "model": "gemma-4-12b-it",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 8.7 GB at this context; duplicate traces of the same session: uhrwozeu; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 116.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-12B-it-oQ4e-fp16-mtp on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 116.3 11.5 8.7 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tvrhjpd6",
   "source_url": "https://omlx.ai/benchmarks/performance/tvrhjpd6",
   "tg_tokens": null,
   "tps": 11.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-jt9t4z5b-ctx4096",
   "model": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 9.4 GB at this context; duplicate traces of the same session: ow1d2c7i; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 604.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-oQ6e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 6bit; 4k 604.5 36.6 9.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jt9t4z5b",
   "source_url": "https://omlx.ai/benchmarks/performance/jt9t4z5b",
   "tg_tokens": null,
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-jt9t4z5b-ctx16384",
   "model": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 10.1 GB at this context; duplicate traces of the same session: ow1d2c7i; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 558.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-oQ6e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 6bit; 16k 558.3 34.5 10.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jt9t4z5b",
   "source_url": "https://omlx.ai/benchmarks/performance/jt9t4z5b",
   "tg_tokens": null,
   "tps": 34.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-2bofjqxc-ctx4096",
   "model": "Qwen3.6-35B-A3B-MAYA",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: 4yrz4t3q, la083hq7, zlff3s6z; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 870.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-MAYA-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 870.3 69.2 22.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2bofjqxc",
   "source_url": "https://omlx.ai/benchmarks/performance/2bofjqxc",
   "tg_tokens": null,
   "tps": 69.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-2bofjqxc-ctx16384",
   "model": "Qwen3.6-35B-A3B-MAYA",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 4yrz4t3q, la083hq7, zlff3s6z; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 738.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-MAYA-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k 738.3 61.5 23.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2bofjqxc",
   "source_url": "https://omlx.ai/benchmarks/performance/2bofjqxc",
   "tg_tokens": null,
   "tps": 61.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-2bofjqxc-ctx65536",
   "model": "Qwen3.6-35B-A3B-MAYA",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: 4yrz4t3q, la083hq7, zlff3s6z; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 431.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-MAYA-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 64k 431.4 44.1 24.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2bofjqxc",
   "source_url": "https://omlx.ai/benchmarks/performance/2bofjqxc",
   "tg_tokens": null,
   "tps": 44.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-2bofjqxc-ctx131072",
   "model": "Qwen3.6-35B-A3B-MAYA",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: 4yrz4t3q, la083hq7, zlff3s6z; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 280.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-MAYA-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 128k 280.1 33.3 26.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2bofjqxc",
   "source_url": "https://omlx.ai/benchmarks/performance/2bofjqxc",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-ng7ehdmr-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 76.2 GB at this context; duplicate traces of the same session: 0bw8rter; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 341.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 16k 341.2 20.6 76.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ng7ehdmr",
   "source_url": "https://omlx.ai/benchmarks/performance/ng7ehdmr",
   "tg_tokens": null,
   "tps": 20.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-ng7ehdmr-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 76.5 GB at this context; duplicate traces of the same session: 0bw8rter; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 462.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 32k 462.6 12.9 76.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ng7ehdmr",
   "source_url": "https://omlx.ai/benchmarks/performance/ng7ehdmr",
   "tg_tokens": null,
   "tps": 12.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-0fl5yb45-ctx4096",
   "model": "gemma-4-12B-it-qat-uncensored-heretic",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 7.8 GB; TTFT 9234 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 443.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-qat-uncensored-heretic-mlx-4Bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k tokens 443.6 PP tok/s 44.2 TG tok/s 9234 TTFT (ms) 7.8 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0fl5yb45",
   "source_url": "https://omlx.ai/benchmarks/performance/0fl5yb45",
   "tg_tokens": null,
   "tps": 44.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ytglp1zi-ctx4096",
   "model": "nightmedia-Qwen3.8-27B-MindMeld",
   "notes": "oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: axli5ja0; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 161.9,
   "provenance": "community",
   "quant": "qx64 hi",
   "quote": "nightmedia-Qwen3.8-27B-MindMeld-qx64-hi-mlx on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 6bit; 4k 161.9 13.2 23.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ytglp1zi",
   "source_url": "https://omlx.ai/benchmarks/performance/ytglp1zi",
   "tg_tokens": null,
   "tps": 13.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ytglp1zi-ctx16384",
   "model": "nightmedia-Qwen3.8-27B-MindMeld",
   "notes": "oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: axli5ja0; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 156.3,
   "provenance": "community",
   "quant": "qx64 hi",
   "quote": "nightmedia-Qwen3.8-27B-MindMeld-qx64-hi-mlx on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 6bit; 16k 156.3 12.6 25.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ytglp1zi",
   "source_url": "https://omlx.ai/benchmarks/performance/ytglp1zi",
   "tg_tokens": null,
   "tps": 12.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-c8cv2adt-ctx1024",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: khi4wfp3, kohr15pk, rpz65k1g, uczivp2v; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 108.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 5bit; 1k 108.0 10.5 22.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c8cv2adt",
   "source_url": "https://omlx.ai/benchmarks/performance/c8cv2adt",
   "tg_tokens": null,
   "tps": 10.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-c8cv2adt-ctx4096",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: khi4wfp3, kohr15pk, rpz65k1g, uczivp2v; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.6,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 5bit; 4k 102.6 9.1 24.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c8cv2adt",
   "source_url": "https://omlx.ai/benchmarks/performance/c8cv2adt",
   "tg_tokens": null,
   "tps": 9.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-c8cv2adt-ctx8192",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 25.6 GB at this context; duplicate traces of the same session: khi4wfp3, kohr15pk, rpz65k1g, uczivp2v; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 98.1,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 5bit; 8k 98.1 11.5 25.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c8cv2adt",
   "source_url": "https://omlx.ai/benchmarks/performance/c8cv2adt",
   "tg_tokens": null,
   "tps": 11.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-c8cv2adt-ctx16384",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: khi4wfp3, kohr15pk, rpz65k1g, uczivp2v; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 91.7,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 5bit; 16k 91.7 7.2 27.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c8cv2adt",
   "source_url": "https://omlx.ai/benchmarks/performance/c8cv2adt",
   "tg_tokens": null,
   "tps": 7.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-c8cv2adt-ctx32768",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 30.6 GB at this context; duplicate traces of the same session: khi4wfp3, kohr15pk, rpz65k1g, uczivp2v; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 81.7,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 5bit; 32k 81.7 5.2 30.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c8cv2adt",
   "source_url": "https://omlx.ai/benchmarks/performance/c8cv2adt",
   "tg_tokens": null,
   "tps": 5.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-elhb6fqq-ctx131072",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 38.5 GB; TTFT 55133 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2377.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-oQ8-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 128k tokens 2,377 PP tok/s 61.4 TG tok/s 55133 TTFT (ms) 38.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/elhb6fqq",
   "source_url": "https://omlx.ai/benchmarks/performance/elhb6fqq",
   "tg_tokens": null,
   "tps": 61.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-bkqk5pyl-ctx32768",
   "model": "Q_pi_S",
   "notes": "oMLX community benchmark trace; peak memory 33.3 GB at this context; duplicate traces of the same session: s99igety, x2oy49ns; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 201.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Q_pi_S on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 201.5 39.9 33.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bkqk5pyl",
   "source_url": "https://omlx.ai/benchmarks/performance/bkqk5pyl",
   "tg_tokens": null,
   "tps": 39.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-bkqk5pyl-ctx65536",
   "model": "Q_pi_S",
   "notes": "oMLX community benchmark trace; peak memory 35.9 GB at this context; duplicate traces of the same session: s99igety, x2oy49ns; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 176.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Q_pi_S on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 64k 176.1 30.1 35.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bkqk5pyl",
   "source_url": "https://omlx.ai/benchmarks/performance/bkqk5pyl",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-bkqk5pyl-ctx131072",
   "model": "Q_pi_S",
   "notes": "oMLX community benchmark trace; peak memory 40.4 GB at this context; duplicate traces of the same session: s99igety, x2oy49ns; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 138.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Q_pi_S on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 128k 138.9 25.4 40.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bkqk5pyl",
   "source_url": "https://omlx.ai/benchmarks/performance/bkqk5pyl",
   "tg_tokens": null,
   "tps": 25.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-0euohv8t-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: 6uu0zerk, akem114h, lk5dcsky; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 894.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 894.7 70.2 22.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0euohv8t",
   "source_url": "https://omlx.ai/benchmarks/performance/0euohv8t",
   "tg_tokens": null,
   "tps": 70.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-0euohv8t-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 6uu0zerk, akem114h, lk5dcsky; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 738.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k 738.7 61.9 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0euohv8t",
   "source_url": "https://omlx.ai/benchmarks/performance/0euohv8t",
   "tg_tokens": null,
   "tps": 61.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-0euohv8t-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: 6uu0zerk, akem114h, lk5dcsky; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 431.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 64k 431.4 44.1 24.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0euohv8t",
   "source_url": "https://omlx.ai/benchmarks/performance/0euohv8t",
   "tg_tokens": null,
   "tps": 44.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-0euohv8t-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: 6uu0zerk, akem114h, lk5dcsky; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 280.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 128k 280.1 33.3 26.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0euohv8t",
   "source_url": "https://omlx.ai/benchmarks/performance/0euohv8t",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-w2k2mufg-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: cq3eul9a; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 87.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 4k 87.5 12.6 17.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w2k2mufg",
   "source_url": "https://omlx.ai/benchmarks/performance/w2k2mufg",
   "tg_tokens": null,
   "tps": 12.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-w2k2mufg-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: cq3eul9a; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 83.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 16k 83.4 10.7 18.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w2k2mufg",
   "source_url": "https://omlx.ai/benchmarks/performance/w2k2mufg",
   "tg_tokens": null,
   "tps": 10.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-workyfq6-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: sj0u2lqc; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 374.9,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-oQ4 on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 374.9 17.3 16.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/workyfq6",
   "source_url": "https://omlx.ai/benchmarks/performance/workyfq6",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-workyfq6-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: sj0u2lqc; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 411.8,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-oQ4 on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 411.8 17.1 17.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/workyfq6",
   "source_url": "https://omlx.ai/benchmarks/performance/workyfq6",
   "tg_tokens": null,
   "tps": 17.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-4ren3uf2-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 68.1 GB; TTFT 144209 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 227.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k tokens 227.2 PP tok/s 14.0 TG tok/s 144209 TTFT (ms) 68.1 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ren3uf2",
   "source_url": "https://omlx.ai/benchmarks/performance/4ren3uf2",
   "tg_tokens": null,
   "tps": 14.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-mrm5h9ul-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 38.3 GB; TTFT 145875 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k tokens 224.6 PP tok/s 14.8 TG tok/s 145875 TTFT (ms) 38.3 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mrm5h9ul",
   "source_url": "https://omlx.ai/benchmarks/performance/mrm5h9ul",
   "tg_tokens": null,
   "tps": 14.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-sny9yizj-ctx1024",
   "model": "Qwopus3.8-27B-Flash-V2",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.7 GB at this context; duplicate traces of the same session: qyw4s8fe, zqxl1yi1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 248.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwopus3.8-27B-Flash-V2-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 248.5 30.1 33.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sny9yizj",
   "source_url": "https://omlx.ai/benchmarks/performance/sny9yizj",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-sny9yizj-ctx4096",
   "model": "Qwopus3.8-27B-Flash-V2",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: qyw4s8fe, zqxl1yi1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 257.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwopus3.8-27B-Flash-V2-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 257.1 54.0 37.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sny9yizj",
   "source_url": "https://omlx.ai/benchmarks/performance/sny9yizj",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-sny9yizj-ctx8192",
   "model": "Qwopus3.8-27B-Flash-V2",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.8 GB at this context; duplicate traces of the same session: qyw4s8fe, zqxl1yi1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 253.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwopus3.8-27B-Flash-V2-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 253.3 28.0 37.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sny9yizj",
   "source_url": "https://omlx.ai/benchmarks/performance/sny9yizj",
   "tg_tokens": null,
   "tps": 28.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-9unvg036-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: ivwmqz2a, sdcnjh4y; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 275.5,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 5bit; 1k 275.5 49.9 24.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9unvg036",
   "source_url": "https://omlx.ai/benchmarks/performance/9unvg036",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-9unvg036-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: ivwmqz2a, sdcnjh4y; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 308.3,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 5bit; 4k 308.3 41.6 27.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9unvg036",
   "source_url": "https://omlx.ai/benchmarks/performance/9unvg036",
   "tg_tokens": null,
   "tps": 41.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-9unvg036-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: ivwmqz2a, sdcnjh4y; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 297.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 5bit; 16k 297.2 58.1 29.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9unvg036",
   "source_url": "https://omlx.ai/benchmarks/performance/9unvg036",
   "tg_tokens": null,
   "tps": 58.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-5gesvfi4-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.3 GB at this context; duplicate traces of the same session: 49scfg1e, 4pdbk0fq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 254.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 254.0 16.7 28.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5gesvfi4",
   "source_url": "https://omlx.ai/benchmarks/performance/5gesvfi4",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-5gesvfi4-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.7 GB at this context; duplicate traces of the same session: 49scfg1e, 4pdbk0fq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 262.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 262.7 16.5 31.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5gesvfi4",
   "source_url": "https://omlx.ai/benchmarks/performance/5gesvfi4",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-5gesvfi4-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: 49scfg1e, 4pdbk0fq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 258.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 258.2 16.2 32.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5gesvfi4",
   "source_url": "https://omlx.ai/benchmarks/performance/5gesvfi4",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-pp0uvyoj-ctx4096",
   "model": "qwen3.5-4b-agentic-coder-v4",
   "notes": "MLX; oMLX community benchmark trace; peak memory 3.8 GB at this context; duplicate traces of the same session: cqtje4za; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1900.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "qwen3.5-4b-agentic-coder-v4-4bit-mlx on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 4k 1,900 44.3 3.8 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pp0uvyoj",
   "source_url": "https://omlx.ai/benchmarks/performance/pp0uvyoj",
   "tg_tokens": null,
   "tps": 44.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-pp0uvyoj-ctx16384",
   "model": "qwen3.5-4b-agentic-coder-v4",
   "notes": "MLX; oMLX community benchmark trace; peak memory 4.1 GB at this context; duplicate traces of the same session: cqtje4za; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1757.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "qwen3.5-4b-agentic-coder-v4-4bit-mlx on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 16k 1,757 37.8 4.1 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pp0uvyoj",
   "source_url": "https://omlx.ai/benchmarks/performance/pp0uvyoj",
   "tg_tokens": null,
   "tps": 37.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ptvg5gtl-ctx1024",
   "model": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: kcuk7fht, winfxspt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1564.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated-mlx-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 1,564 94.4 35.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ptvg5gtl",
   "source_url": "https://omlx.ai/benchmarks/performance/ptvg5gtl",
   "tg_tokens": null,
   "tps": 94.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ptvg5gtl-ctx4096",
   "model": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated",
   "notes": "oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: kcuk7fht, winfxspt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2091.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated-mlx-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 2,091 92.4 37.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ptvg5gtl",
   "source_url": "https://omlx.ai/benchmarks/performance/ptvg5gtl",
   "tg_tokens": null,
   "tps": 92.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ptvg5gtl-ctx8192",
   "model": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated",
   "notes": "oMLX community benchmark trace; peak memory 37.5 GB at this context; duplicate traces of the same session: kcuk7fht, winfxspt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1946.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated-mlx-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 1,946 90.0 37.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ptvg5gtl",
   "source_url": "https://omlx.ai/benchmarks/performance/ptvg5gtl",
   "tg_tokens": null,
   "tps": 90.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-h09by0em-ctx65536",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 37.3 GB; TTFT 18678 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3509.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-oQ8-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 64k tokens 3,509 PP tok/s 74.8 TG tok/s 18678 TTFT (ms) 37.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h09by0em",
   "source_url": "https://omlx.ai/benchmarks/performance/h09by0em",
   "tg_tokens": null,
   "tps": 74.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-wiy2jx63-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.1 GB at this context; duplicate traces of the same session: fxvcbeq1, wehgk275; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 845.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-9B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 845.6 113.9 7.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wiy2jx63",
   "source_url": "https://omlx.ai/benchmarks/performance/wiy2jx63",
   "tg_tokens": null,
   "tps": 113.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-wiy2jx63-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 8.6 GB at this context; duplicate traces of the same session: fxvcbeq1, wehgk275; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 935.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-9B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 935.0 116.0 8.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wiy2jx63",
   "source_url": "https://omlx.ai/benchmarks/performance/wiy2jx63",
   "tg_tokens": null,
   "tps": 116.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-wiy2jx63-ctx8192",
   "model": "Ornith-1.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 8.8 GB at this context; duplicate traces of the same session: fxvcbeq1, wehgk275; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 905.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-9B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 905.9 109.0 8.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wiy2jx63",
   "source_url": "https://omlx.ai/benchmarks/performance/wiy2jx63",
   "tg_tokens": null,
   "tps": 109.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-w0uc35gd-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "bf16-pt; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: 2bd3up2j; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2043.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-bf16-pt-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 1k 2,043 166.5 25.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w0uc35gd",
   "source_url": "https://omlx.ai/benchmarks/performance/w0uc35gd",
   "tg_tokens": null,
   "tps": 166.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-w0uc35gd-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "bf16-pt; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: 2bd3up2j; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2489.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-bf16-pt-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 4k 2,489 124.5 26.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w0uc35gd",
   "source_url": "https://omlx.ai/benchmarks/performance/w0uc35gd",
   "tg_tokens": null,
   "tps": 124.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-pidaruuw-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: 6n0dbyt4, 7d84x31w, a0a9cady; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 590.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 590.5 61.4 22.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pidaruuw",
   "source_url": "https://omlx.ai/benchmarks/performance/pidaruuw",
   "tg_tokens": null,
   "tps": 61.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-pidaruuw-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 6n0dbyt4, 7d84x31w, a0a9cady; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 462.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k 462.9 53.1 23.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pidaruuw",
   "source_url": "https://omlx.ai/benchmarks/performance/pidaruuw",
   "tg_tokens": null,
   "tps": 53.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-pidaruuw-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: 6n0dbyt4, 7d84x31w, a0a9cady; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 247.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 64k 247.4 37.1 24.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pidaruuw",
   "source_url": "https://omlx.ai/benchmarks/performance/pidaruuw",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-pidaruuw-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: 6n0dbyt4, 7d84x31w, a0a9cady; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 153.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 128k 153.8 27.2 26.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pidaruuw",
   "source_url": "https://omlx.ai/benchmarks/performance/pidaruuw",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-uvezr0qx-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: 6023muw1; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 89.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 89.9 17.7 15.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uvezr0qx",
   "source_url": "https://omlx.ai/benchmarks/performance/uvezr0qx",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-uvezr0qx-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: 6023muw1; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 95.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 95.0 18.2 17.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uvezr0qx",
   "source_url": "https://omlx.ai/benchmarks/performance/uvezr0qx",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0zf8vbls-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: phny9ule; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 387.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 387.2 18.1 16.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0zf8vbls",
   "source_url": "https://omlx.ai/benchmarks/performance/0zf8vbls",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0zf8vbls-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: phny9ule; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 420.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 420.3 17.8 17.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0zf8vbls",
   "source_url": "https://omlx.ai/benchmarks/performance/0zf8vbls",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3d63ktuw-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: 8h9tne7f; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 675.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 6bit; 16k 675.4 16.4 24.2 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3d63ktuw",
   "source_url": "https://omlx.ai/benchmarks/performance/3d63ktuw",
   "tg_tokens": null,
   "tps": 16.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3d63ktuw-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: 8h9tne7f; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 529.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 6bit; 32k 529.1 15.1 25.0 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3d63ktuw",
   "source_url": "https://omlx.ai/benchmarks/performance/3d63ktuw",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-53nxidk5-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.0 GB at this context; duplicate traces of the same session: cl1da1su, ichqqkc2, kpuay2qi, lzblveuw; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 415.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-oQ4 on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 415.0 16.8 18.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/53nxidk5",
   "source_url": "https://omlx.ai/benchmarks/performance/53nxidk5",
   "tg_tokens": null,
   "tps": 16.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-53nxidk5-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: cl1da1su, ichqqkc2, kpuay2qi, lzblveuw; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 377.4,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-oQ4 on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 377.4 16.2 18.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/53nxidk5",
   "source_url": "https://omlx.ai/benchmarks/performance/53nxidk5",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-53nxidk5-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: cl1da1su, ichqqkc2, kpuay2qi, lzblveuw; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 363.6,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-oQ4 on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 363.6 15.0 19.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/53nxidk5",
   "source_url": "https://omlx.ai/benchmarks/performance/53nxidk5",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-53nxidk5-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: cl1da1su, ichqqkc2, kpuay2qi, lzblveuw; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 330.2,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-oQ4 on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 64k 330.2 13.1 21.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/53nxidk5",
   "source_url": "https://omlx.ai/benchmarks/performance/53nxidk5",
   "tg_tokens": null,
   "tps": 13.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-53nxidk5-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: cl1da1su, ichqqkc2, kpuay2qi, lzblveuw; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 265.8,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-oQ4 on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 128k 265.8 10.6 25.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/53nxidk5",
   "source_url": "https://omlx.ai/benchmarks/performance/53nxidk5",
   "tg_tokens": null,
   "tps": 10.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-jsncdlvb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.3 GB at this context; duplicate traces of the same session: 5jb2vth3, 6s5teza9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 246.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 246.5 44.1 32.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jsncdlvb",
   "source_url": "https://omlx.ai/benchmarks/performance/jsncdlvb",
   "tg_tokens": null,
   "tps": 44.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-jsncdlvb-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.7 GB at this context; duplicate traces of the same session: 5jb2vth3, 6s5teza9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 282.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 282.8 60.2 35.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jsncdlvb",
   "source_url": "https://omlx.ai/benchmarks/performance/jsncdlvb",
   "tg_tokens": null,
   "tps": 60.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-jsncdlvb-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.2 GB at this context; duplicate traces of the same session: 5jb2vth3, 6s5teza9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 295.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 295.6 33.2 36.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jsncdlvb",
   "source_url": "https://omlx.ai/benchmarks/performance/jsncdlvb",
   "tg_tokens": null,
   "tps": 33.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-rbr0b267-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: ck02h6hu; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 391.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 8bit; 4k 391.9 10.0 29.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rbr0b267",
   "source_url": "https://omlx.ai/benchmarks/performance/rbr0b267",
   "tg_tokens": null,
   "tps": 10.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-rbr0b267-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.3 GB at this context; duplicate traces of the same session: ck02h6hu; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 337.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 8bit; 32k 337.4 9.2 31.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rbr0b267",
   "source_url": "https://omlx.ai/benchmarks/performance/rbr0b267",
   "tg_tokens": null,
   "tps": 9.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-0rdwu2p7-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: ido4x5zz; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 383.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 8bit; 4k 383.7 29.7 34.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0rdwu2p7",
   "source_url": "https://omlx.ai/benchmarks/performance/0rdwu2p7",
   "tg_tokens": null,
   "tps": 29.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-0rdwu2p7-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.8 GB at this context; duplicate traces of the same session: ido4x5zz; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 333.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 8bit; 32k 333.1 20.7 36.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0rdwu2p7",
   "source_url": "https://omlx.ai/benchmarks/performance/0rdwu2p7",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ozge3oho-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.9 GB at this context; duplicate traces of the same session: 1ezkbbg5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 387.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 8bit; 4k 387.0 32.6 30.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ozge3oho",
   "source_url": "https://omlx.ai/benchmarks/performance/ozge3oho",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ozge3oho-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.5 GB at this context; duplicate traces of the same session: 1ezkbbg5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 334.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 8bit; 32k 334.3 8.5 33.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ozge3oho",
   "source_url": "https://omlx.ai/benchmarks/performance/ozge3oho",
   "tg_tokens": null,
   "tps": 8.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-iemmaiow-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: tiw3x6dm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 810.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 810.5 53.6 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iemmaiow",
   "source_url": "https://omlx.ai/benchmarks/performance/iemmaiow",
   "tg_tokens": null,
   "tps": 53.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-iemmaiow-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: tiw3x6dm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1435.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,435 46.4 70.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iemmaiow",
   "source_url": "https://omlx.ai/benchmarks/performance/iemmaiow",
   "tg_tokens": null,
   "tps": 46.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-8rv2lk9a-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.1 GB at this context; duplicate traces of the same session: rdxy1cdc, zzaor7qt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 998.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 998.1 81.4 183.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8rv2lk9a",
   "source_url": "https://omlx.ai/benchmarks/performance/8rv2lk9a",
   "tg_tokens": null,
   "tps": 81.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-8rv2lk9a-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.7 GB at this context; duplicate traces of the same session: rdxy1cdc, zzaor7qt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1172.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 1,172 81.0 183.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8rv2lk9a",
   "source_url": "https://omlx.ai/benchmarks/performance/8rv2lk9a",
   "tg_tokens": null,
   "tps": 81.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-8rv2lk9a-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 191.9 GB at this context; duplicate traces of the same session: rdxy1cdc, zzaor7qt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1251.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 195k 1,251 47.8 191.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8rv2lk9a",
   "source_url": "https://omlx.ai/benchmarks/performance/8rv2lk9a",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-bw7dlcbt-ctx4096",
   "model": "qwen3.5-4b-agentic-coder-v4",
   "notes": "MLX; oMLX community benchmark trace; peak memory 3.8 GB at this context; duplicate traces of the same session: 1surn3xe, 31iakly5; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1912.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "qwen3.5-4b-agentic-coder-v4-4bit-mlx on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 4k 1,912 43.8 3.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bw7dlcbt",
   "source_url": "https://omlx.ai/benchmarks/performance/bw7dlcbt",
   "tg_tokens": null,
   "tps": 43.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-bw7dlcbt-ctx16384",
   "model": "qwen3.5-4b-agentic-coder-v4",
   "notes": "MLX; oMLX community benchmark trace; peak memory 4.1 GB at this context; duplicate traces of the same session: 1surn3xe, 31iakly5; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1764.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "qwen3.5-4b-agentic-coder-v4-4bit-mlx on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 16k 1,764 39.1 4.1 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bw7dlcbt",
   "source_url": "https://omlx.ai/benchmarks/performance/bw7dlcbt",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-bw7dlcbt-ctx131072",
   "model": "qwen3.5-4b-agentic-coder-v4",
   "notes": "MLX; oMLX community benchmark trace; peak memory 7.6 GB at this context; duplicate traces of the same session: 1surn3xe, 31iakly5; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 806.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "qwen3.5-4b-agentic-coder-v4-4bit-mlx on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 128k 806.8 15.2 7.6 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bw7dlcbt",
   "source_url": "https://omlx.ai/benchmarks/performance/bw7dlcbt",
   "tg_tokens": null,
   "tps": 15.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-g9e11lm2-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: uwkeagm9; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2664.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 8bit; 4k 2,664 68.9 36.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g9e11lm2",
   "source_url": "https://omlx.ai/benchmarks/performance/g9e11lm2",
   "tg_tokens": null,
   "tps": 68.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-g9e11lm2-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 36.7 GB at this context; duplicate traces of the same session: uwkeagm9; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2081.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 8bit; 32k 2,081 56.5 36.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g9e11lm2",
   "source_url": "https://omlx.ai/benchmarks/performance/g9e11lm2",
   "tg_tokens": null,
   "tps": 56.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-cz6im0l6-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: 289gox18, 43b2dupq, byupmzlc, lmc51lkb, o0ovb3o8, r7bpy09b, rh09jeo6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 331.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 331.1 45.3 32.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cz6im0l6",
   "source_url": "https://omlx.ai/benchmarks/performance/cz6im0l6",
   "tg_tokens": null,
   "tps": 45.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-cz6im0l6-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.0 GB at this context; duplicate traces of the same session: 289gox18, 43b2dupq, byupmzlc, lmc51lkb, o0ovb3o8, r7bpy09b, rh09jeo6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 335.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 335.4 42.1 34.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cz6im0l6",
   "source_url": "https://omlx.ai/benchmarks/performance/cz6im0l6",
   "tg_tokens": null,
   "tps": 42.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-cz6im0l6-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 35.3 GB at this context; duplicate traces of the same session: 289gox18, 43b2dupq, byupmzlc, lmc51lkb, o0ovb3o8, r7bpy09b, rh09jeo6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 330.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 8k 330.3 42.0 35.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cz6im0l6",
   "source_url": "https://omlx.ai/benchmarks/performance/cz6im0l6",
   "tg_tokens": null,
   "tps": 42.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-cz6im0l6-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 37.5 GB at this context; duplicate traces of the same session: 289gox18, 43b2dupq, byupmzlc, lmc51lkb, o0ovb3o8, r7bpy09b, rh09jeo6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 319.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 16k 319.2 50.0 37.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cz6im0l6",
   "source_url": "https://omlx.ai/benchmarks/performance/cz6im0l6",
   "tg_tokens": null,
   "tps": 50.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-cz6im0l6-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 41.5 GB at this context; duplicate traces of the same session: 289gox18, 43b2dupq, byupmzlc, lmc51lkb, o0ovb3o8, r7bpy09b, rh09jeo6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 298.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 32k 298.4 42.5 41.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cz6im0l6",
   "source_url": "https://omlx.ai/benchmarks/performance/cz6im0l6",
   "tg_tokens": null,
   "tps": 42.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-cz6im0l6-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 49.2 GB at this context; duplicate traces of the same session: 289gox18, 43b2dupq, byupmzlc, lmc51lkb, o0ovb3o8, r7bpy09b, rh09jeo6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 259.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 64k 259.6 33.0 49.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cz6im0l6",
   "source_url": "https://omlx.ai/benchmarks/performance/cz6im0l6",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-cz6im0l6-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 63.7 GB at this context; duplicate traces of the same session: 289gox18, 43b2dupq, byupmzlc, lmc51lkb, o0ovb3o8, r7bpy09b, rh09jeo6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 202.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 128k 202.3 29.6 63.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cz6im0l6",
   "source_url": "https://omlx.ai/benchmarks/performance/cz6im0l6",
   "tg_tokens": null,
   "tps": 29.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-cz6im0l6-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 66.0 GB at this context; duplicate traces of the same session: 289gox18, 43b2dupq, byupmzlc, lmc51lkb, o0ovb3o8, r7bpy09b, rh09jeo6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 163.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 195k 163.6 20.2 66.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cz6im0l6",
   "source_url": "https://omlx.ai/benchmarks/performance/cz6im0l6",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-xyid244r-ctx65536",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace, single run at 64k context; peak memory 23.2 GB; TTFT 75239 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 871.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k tokens 871.0 PP tok/s 26.5 TG tok/s 75239 TTFT (ms) 23.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xyid244r",
   "source_url": "https://omlx.ai/benchmarks/performance/xyid244r",
   "tg_tokens": null,
   "tps": 26.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-sd0nmzj2-ctx65536",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace, single run at 64k context; peak memory 23.2 GB; TTFT 75722 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 865.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k tokens 865.5 PP tok/s 17.7 TG tok/s 75722 TTFT (ms) 23.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sd0nmzj2",
   "source_url": "https://omlx.ai/benchmarks/performance/sd0nmzj2",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1r8e3ppo-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: cwyz4ewf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 845.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 845.0 27.2 20.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1r8e3ppo",
   "source_url": "https://omlx.ai/benchmarks/performance/1r8e3ppo",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1r8e3ppo-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: cwyz4ewf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 998.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 998.4 27.0 21.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1r8e3ppo",
   "source_url": "https://omlx.ai/benchmarks/performance/1r8e3ppo",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-6lijevp2-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: ejtsw8zd, ii93ctqy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 398.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 398.3 49.5 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6lijevp2",
   "source_url": "https://omlx.ai/benchmarks/performance/6lijevp2",
   "tg_tokens": null,
   "tps": 49.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-6lijevp2-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: ejtsw8zd, ii93ctqy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 485.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 485.5 66.1 71.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6lijevp2",
   "source_url": "https://omlx.ai/benchmarks/performance/6lijevp2",
   "tg_tokens": null,
   "tps": 66.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-6lijevp2-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: ejtsw8zd, ii93ctqy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 796.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 796.8 39.9 71.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6lijevp2",
   "source_url": "https://omlx.ai/benchmarks/performance/6lijevp2",
   "tg_tokens": null,
   "tps": 39.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-8prk3ue8-ctx4096",
   "model": "qwen3.5-4b-agentic-coder-v4",
   "notes": "MLX; oMLX community benchmark trace; peak memory 3.8 GB at this context; duplicate traces of the same session: 2b8d4gpr, 6tzxfsi9, e6xoamgt, ty5kn1rv; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1913.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "qwen3.5-4b-agentic-coder-v4-4bit-mlx on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 4k 1,913 44.9 3.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8prk3ue8",
   "source_url": "https://omlx.ai/benchmarks/performance/8prk3ue8",
   "tg_tokens": null,
   "tps": 44.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-8prk3ue8-ctx16384",
   "model": "qwen3.5-4b-agentic-coder-v4",
   "notes": "MLX; oMLX community benchmark trace; peak memory 4.1 GB at this context; duplicate traces of the same session: 2b8d4gpr, 6tzxfsi9, e6xoamgt, ty5kn1rv; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1764.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "qwen3.5-4b-agentic-coder-v4-4bit-mlx on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 16k 1,764 38.6 4.1 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8prk3ue8",
   "source_url": "https://omlx.ai/benchmarks/performance/8prk3ue8",
   "tg_tokens": null,
   "tps": 38.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-8prk3ue8-ctx65536",
   "model": "qwen3.5-4b-agentic-coder-v4",
   "notes": "MLX; oMLX community benchmark trace; peak memory 5.6 GB at this context; duplicate traces of the same session: 2b8d4gpr, 6tzxfsi9, e6xoamgt, ty5kn1rv; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1185.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "qwen3.5-4b-agentic-coder-v4-4bit-mlx on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 64k 1,185 23.2 5.6 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8prk3ue8",
   "source_url": "https://omlx.ai/benchmarks/performance/8prk3ue8",
   "tg_tokens": null,
   "tps": 23.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-8prk3ue8-ctx131072",
   "model": "qwen3.5-4b-agentic-coder-v4",
   "notes": "MLX; oMLX community benchmark trace; peak memory 7.6 GB at this context; duplicate traces of the same session: 2b8d4gpr, 6tzxfsi9, e6xoamgt, ty5kn1rv; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 813.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "qwen3.5-4b-agentic-coder-v4-4bit-mlx on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 128k 813.1 15.5 7.6 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8prk3ue8",
   "source_url": "https://omlx.ai/benchmarks/performance/8prk3ue8",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-8prk3ue8-ctx199680",
   "model": "qwen3.5-4b-agentic-coder-v4",
   "notes": "MLX; oMLX community benchmark trace; peak memory 9.7 GB at this context; duplicate traces of the same session: 2b8d4gpr, 6tzxfsi9, e6xoamgt, ty5kn1rv; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 618.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "qwen3.5-4b-agentic-coder-v4-4bit-mlx on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 195k 618.8 11.0 9.7 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8prk3ue8",
   "source_url": "https://omlx.ai/benchmarks/performance/8prk3ue8",
   "tg_tokens": null,
   "tps": 11.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-go5hj4nd-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 28.7 GB; TTFT 18032 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1817.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-6bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 32k tokens 1,817 PP tok/s 64.3 TG tok/s 18032 TTFT (ms) 28.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/go5hj4nd",
   "source_url": "https://omlx.ai/benchmarks/performance/go5hj4nd",
   "tg_tokens": null,
   "tps": 64.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-77wckmm1-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: z2mjw21v; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 569.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 569.4 17.9 17.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/77wckmm1",
   "source_url": "https://omlx.ai/benchmarks/performance/77wckmm1",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-77wckmm1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: z2mjw21v; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 720.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 720.6 17.7 18.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/77wckmm1",
   "source_url": "https://omlx.ai/benchmarks/performance/77wckmm1",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-z5m2laab-ctx1024",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: 8zcy3ykn, ne9d12qz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1618.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 1,618 77.4 27.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z5m2laab",
   "source_url": "https://omlx.ai/benchmarks/performance/z5m2laab",
   "tg_tokens": null,
   "tps": 77.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-z5m2laab-ctx4096",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.4 GB at this context; duplicate traces of the same session: 8zcy3ykn, ne9d12qz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2438.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 2,438 74.6 28.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z5m2laab",
   "source_url": "https://omlx.ai/benchmarks/performance/z5m2laab",
   "tg_tokens": null,
   "tps": 74.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-z5m2laab-ctx16384",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.7 GB at this context; duplicate traces of the same session: 8zcy3ykn, ne9d12qz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2458.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 16k 2,458 68.9 28.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z5m2laab",
   "source_url": "https://omlx.ai/benchmarks/performance/z5m2laab",
   "tg_tokens": null,
   "tps": 68.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-55tonejf-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 1jkswp61, hrj7gond, n4hjiayq, rstauw6m, ymsuvh4r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 451.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 451.6 30.2 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/55tonejf",
   "source_url": "https://omlx.ai/benchmarks/performance/55tonejf",
   "tg_tokens": null,
   "tps": 30.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-55tonejf-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 1jkswp61, hrj7gond, n4hjiayq, rstauw6m, ymsuvh4r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1150.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,150 53.4 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/55tonejf",
   "source_url": "https://omlx.ai/benchmarks/performance/55tonejf",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-55tonejf-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: 1jkswp61, hrj7gond, n4hjiayq, rstauw6m, ymsuvh4r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1558.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,558 75.0 71.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/55tonejf",
   "source_url": "https://omlx.ai/benchmarks/performance/55tonejf",
   "tg_tokens": null,
   "tps": 75.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-55tonejf-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.7 GB at this context; duplicate traces of the same session: 1jkswp61, hrj7gond, n4hjiayq, rstauw6m, ymsuvh4r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1755.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,755 35.4 73.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/55tonejf",
   "source_url": "https://omlx.ai/benchmarks/performance/55tonejf",
   "tg_tokens": null,
   "tps": 35.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-55tonejf-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 75.4 GB at this context; duplicate traces of the same session: 1jkswp61, hrj7gond, n4hjiayq, rstauw6m, ymsuvh4r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1592.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 1,592 58.1 75.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/55tonejf",
   "source_url": "https://omlx.ai/benchmarks/performance/55tonejf",
   "tg_tokens": null,
   "tps": 58.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-55tonejf-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.2 GB at this context; duplicate traces of the same session: 1jkswp61, hrj7gond, n4hjiayq, rstauw6m, ymsuvh4r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1011.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 1,011 36.5 80.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/55tonejf",
   "source_url": "https://omlx.ai/benchmarks/performance/55tonejf",
   "tg_tokens": null,
   "tps": 36.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-6ltk5m3r-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "bf16-pt; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: h20t051e; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2233.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.6-35B-A3B-bf16-pt-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 1k 2,233 109.6 17.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ltk5m3r",
   "source_url": "https://omlx.ai/benchmarks/performance/6ltk5m3r",
   "tg_tokens": null,
   "tps": 109.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-6ltk5m3r-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "bf16-pt; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: h20t051e; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2707.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.6-35B-A3B-bf16-pt-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 4k 2,707 109.5 19.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ltk5m3r",
   "source_url": "https://omlx.ai/benchmarks/performance/6ltk5m3r",
   "tg_tokens": null,
   "tps": 109.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-aexfwc2y-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "bf16-pt; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: 17we112p; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1871.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.6-35B-A3B-bf16-pt-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 1k 1,871 94.1 17.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aexfwc2y",
   "source_url": "https://omlx.ai/benchmarks/performance/aexfwc2y",
   "tg_tokens": null,
   "tps": 94.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-aexfwc2y-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "bf16-pt; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 17we112p; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2589.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.6-35B-A3B-bf16-pt-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 4k 2,589 92.7 19.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aexfwc2y",
   "source_url": "https://omlx.ai/benchmarks/performance/aexfwc2y",
   "tg_tokens": null,
   "tps": 92.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-f20y3hev-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "bf16-pt; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 40.8 GB at this context; duplicate traces of the same session: ujdt2t7g; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1387.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.6-35B-A3B-bf16-pt-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 1k 1,387 84.8 40.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f20y3hev",
   "source_url": "https://omlx.ai/benchmarks/performance/f20y3hev",
   "tg_tokens": null,
   "tps": 84.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-f20y3hev-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "bf16-pt; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.0 GB at this context; duplicate traces of the same session: ujdt2t7g; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2606.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.6-35B-A3B-bf16-pt-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 4k 2,606 37.0 43.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f20y3hev",
   "source_url": "https://omlx.ai/benchmarks/performance/f20y3hev",
   "tg_tokens": null,
   "tps": 37.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-quu15gl4-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: a3s8rvzi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 874.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 874.2 26.3 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/quu15gl4",
   "source_url": "https://omlx.ai/benchmarks/performance/quu15gl4",
   "tg_tokens": null,
   "tps": 26.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-quu15gl4-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: a3s8rvzi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 965.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 965.1 26.1 21.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/quu15gl4",
   "source_url": "https://omlx.ai/benchmarks/performance/quu15gl4",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-408dx2gi-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: 0fq0wnve, 841if54w, wczu1zka; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.4,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 125.4 17.8 20.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/408dx2gi",
   "source_url": "https://omlx.ai/benchmarks/performance/408dx2gi",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-408dx2gi-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 0fq0wnve, 841if54w, wczu1zka; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 113.1,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k 113.1 12.7 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/408dx2gi",
   "source_url": "https://omlx.ai/benchmarks/performance/408dx2gi",
   "tg_tokens": null,
   "tps": 12.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-408dx2gi-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: 0fq0wnve, 841if54w, wczu1zka; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 91.2,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 64k 91.2 11.8 25.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/408dx2gi",
   "source_url": "https://omlx.ai/benchmarks/performance/408dx2gi",
   "tg_tokens": null,
   "tps": 11.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-408dx2gi-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.3 GB at this context; duplicate traces of the same session: 0fq0wnve, 841if54w, wczu1zka; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 71.5,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 128k 71.5 6.0 29.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/408dx2gi",
   "source_url": "https://omlx.ai/benchmarks/performance/408dx2gi",
   "tg_tokens": null,
   "tps": 6.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-qxi58gat-ctx1024",
   "model": "Qwen3.8-Flash-Next-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 76.4 GB at this context; duplicate traces of the same session: lag65pz1, o24qszey, rijghmwn; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 609.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-Abliterated-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 609.0 46.6 76.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qxi58gat",
   "source_url": "https://omlx.ai/benchmarks/performance/qxi58gat",
   "tg_tokens": null,
   "tps": 46.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-qxi58gat-ctx16384",
   "model": "Qwen3.8-Flash-Next-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 77.3 GB at this context; duplicate traces of the same session: lag65pz1, o24qszey, rijghmwn; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1032.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-Abliterated-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 1,032 43.4 77.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qxi58gat",
   "source_url": "https://omlx.ai/benchmarks/performance/qxi58gat",
   "tg_tokens": null,
   "tps": 43.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-qxi58gat-ctx32768",
   "model": "Qwen3.8-Flash-Next-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 77.8 GB at this context; duplicate traces of the same session: lag65pz1, o24qszey, rijghmwn; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1025.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-Abliterated-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 1,025 41.6 77.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qxi58gat",
   "source_url": "https://omlx.ai/benchmarks/performance/qxi58gat",
   "tg_tokens": null,
   "tps": 41.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qxi58gat-ctx65536",
   "model": "Qwen3.8-Flash-Next-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 78.9 GB at this context; duplicate traces of the same session: lag65pz1, o24qszey, rijghmwn; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1009.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-Abliterated-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 1,009 40.3 78.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qxi58gat",
   "source_url": "https://omlx.ai/benchmarks/performance/qxi58gat",
   "tg_tokens": null,
   "tps": 40.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-68okx20b-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 72.8 GB; TTFT 7124 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2300.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k tokens 2,300 PP tok/s 42.9 TG tok/s 7124 TTFT (ms) 72.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/68okx20b",
   "source_url": "https://omlx.ai/benchmarks/performance/68okx20b",
   "tg_tokens": null,
   "tps": 42.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3cmbpm0h-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: 1lcqjj5l, 4wyxj1pb, 8n7gr68v, e15licj9, shdstxil, ux9lrnt2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 956.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 956.2 71.6 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3cmbpm0h",
   "source_url": "https://omlx.ai/benchmarks/performance/3cmbpm0h",
   "tg_tokens": null,
   "tps": 71.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3cmbpm0h-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 1lcqjj5l, 4wyxj1pb, 8n7gr68v, e15licj9, shdstxil, ux9lrnt2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1935.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,935 79.9 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3cmbpm0h",
   "source_url": "https://omlx.ai/benchmarks/performance/3cmbpm0h",
   "tg_tokens": null,
   "tps": 79.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3cmbpm0h-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 1lcqjj5l, 4wyxj1pb, 8n7gr68v, e15licj9, shdstxil, ux9lrnt2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1930.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,930 79.9 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3cmbpm0h",
   "source_url": "https://omlx.ai/benchmarks/performance/3cmbpm0h",
   "tg_tokens": null,
   "tps": 79.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3cmbpm0h-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.7 GB at this context; duplicate traces of the same session: 1lcqjj5l, 4wyxj1pb, 8n7gr68v, e15licj9, shdstxil, ux9lrnt2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1983.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,983 76.2 73.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3cmbpm0h",
   "source_url": "https://omlx.ai/benchmarks/performance/3cmbpm0h",
   "tg_tokens": null,
   "tps": 76.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3cmbpm0h-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 75.4 GB at this context; duplicate traces of the same session: 1lcqjj5l, 4wyxj1pb, 8n7gr68v, e15licj9, shdstxil, ux9lrnt2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1935.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 1,935 72.8 75.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3cmbpm0h",
   "source_url": "https://omlx.ai/benchmarks/performance/3cmbpm0h",
   "tg_tokens": null,
   "tps": 72.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3cmbpm0h-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.3 GB at this context; duplicate traces of the same session: 1lcqjj5l, 4wyxj1pb, 8n7gr68v, e15licj9, shdstxil, ux9lrnt2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1745.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 1,745 55.8 80.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3cmbpm0h",
   "source_url": "https://omlx.ai/benchmarks/performance/3cmbpm0h",
   "tg_tokens": null,
   "tps": 55.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3cmbpm0h-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 83.4 GB at this context; duplicate traces of the same session: 1lcqjj5l, 4wyxj1pb, 8n7gr68v, e15licj9, shdstxil, ux9lrnt2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1240.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 1,240 46.1 83.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3cmbpm0h",
   "source_url": "https://omlx.ai/benchmarks/performance/3cmbpm0h",
   "tg_tokens": null,
   "tps": 46.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-7ky1bz1e-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-oQ5e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: fxrewum2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 835.3,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 5bit; 1k 835.3 45.8 23.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ky1bz1e",
   "source_url": "https://omlx.ai/benchmarks/performance/7ky1bz1e",
   "tg_tokens": null,
   "tps": 45.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-7ky1bz1e-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-oQ5e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: fxrewum2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 925.5,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 5bit; 4k 925.5 45.8 24.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ky1bz1e",
   "source_url": "https://omlx.ai/benchmarks/performance/7ky1bz1e",
   "tg_tokens": null,
   "tps": 45.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-95499dyc-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 8.1 GB at this context; duplicate traces of the same session: nq5pos3d; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1252.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "gemma-4-E4B-it-oQ8 on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 1k 1,252 42.8 8.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/95499dyc",
   "source_url": "https://omlx.ai/benchmarks/performance/95499dyc",
   "tg_tokens": null,
   "tps": 42.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-95499dyc-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 8.1 GB at this context; duplicate traces of the same session: nq5pos3d; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1289.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "gemma-4-E4B-it-oQ8 on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 4k 1,289 41.9 8.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/95499dyc",
   "source_url": "https://omlx.ai/benchmarks/performance/95499dyc",
   "tg_tokens": null,
   "tps": 41.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-yyqk98w0-ctx1024",
   "model": "K2-Horizon-7B",
   "notes": "oMLX community benchmark trace; peak memory 9.7 GB at this context; duplicate traces of the same session: i6sp9d3w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 910.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "K2-Horizon-7B-oQ8e on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 910.9 51.7 9.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yyqk98w0",
   "source_url": "https://omlx.ai/benchmarks/performance/yyqk98w0",
   "tg_tokens": null,
   "tps": 51.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-yyqk98w0-ctx4096",
   "model": "K2-Horizon-7B",
   "notes": "oMLX community benchmark trace; peak memory 10.2 GB at this context; duplicate traces of the same session: i6sp9d3w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 881.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "K2-Horizon-7B-oQ8e on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 881.6 49.2 10.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yyqk98w0",
   "source_url": "https://omlx.ai/benchmarks/performance/yyqk98w0",
   "tg_tokens": null,
   "tps": 49.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-p3tvjs3l-ctx1024",
   "model": "K2-Horizon-MoVA-36B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 38.0 GB at this context; duplicate traces of the same session: cztu45k0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 962.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "K2-Horizon-MoVA-36B-A4B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 962.0 45.4 38.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p3tvjs3l",
   "source_url": "https://omlx.ai/benchmarks/performance/p3tvjs3l",
   "tg_tokens": null,
   "tps": 45.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-p3tvjs3l-ctx4096",
   "model": "K2-Horizon-MoVA-36B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 38.8 GB at this context; duplicate traces of the same session: cztu45k0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 962.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "K2-Horizon-MoVA-36B-A4B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 962.9 43.6 38.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p3tvjs3l",
   "source_url": "https://omlx.ai/benchmarks/performance/p3tvjs3l",
   "tg_tokens": null,
   "tps": 43.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-esqe4chx-ctx1024",
   "model": "OLMoE-1B-7B-0125-Instruct",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 7.5 GB; TTFT 335 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3060.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "OLMoE-1B-7B-0125-Instruct-8bit on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 8bit; 1k tokens 3,060 PP tok/s 209.6 TG tok/s 335 TTFT (ms) 7.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/esqe4chx",
   "source_url": "https://omlx.ai/benchmarks/performance/esqe4chx",
   "tg_tokens": null,
   "tps": 209.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-8ek03s3q-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 16.1 GB; TTFT 15411 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 66.4,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 66.4 PP tok/s 4.8 TG tok/s 15411 TTFT (ms) 16.1 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8ek03s3q",
   "source_url": "https://omlx.ai/benchmarks/performance/8ek03s3q",
   "tg_tokens": null,
   "tps": 4.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-xb6m5n9x-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: pb84t546; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 227.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 227.6 8.0 17.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xb6m5n9x",
   "source_url": "https://omlx.ai/benchmarks/performance/xb6m5n9x",
   "tg_tokens": null,
   "tps": 8.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-xb6m5n9x-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: pb84t546; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 208.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 16k 208.3 7.5 19.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xb6m5n9x",
   "source_url": "https://omlx.ai/benchmarks/performance/xb6m5n9x",
   "tg_tokens": null,
   "tps": 7.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-khx3gq59-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.7 GB at this context; duplicate traces of the same session: 4qmk5fsr, ppph3sl2, ps4ko7ya, xo6b9zg9; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 420.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 1k 420.6 35.9 100.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/khx3gq59",
   "source_url": "https://omlx.ai/benchmarks/performance/khx3gq59",
   "tg_tokens": null,
   "tps": 35.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-khx3gq59-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: 4qmk5fsr, ppph3sl2, ps4ko7ya, xo6b9zg9; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 769.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 8k 769.6 26.4 101.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/khx3gq59",
   "source_url": "https://omlx.ai/benchmarks/performance/khx3gq59",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-khx3gq59-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.6 GB at this context; duplicate traces of the same session: 4qmk5fsr, ppph3sl2, ps4ko7ya, xo6b9zg9; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 774.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 16k 774.0 31.0 101.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/khx3gq59",
   "source_url": "https://omlx.ai/benchmarks/performance/khx3gq59",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-khx3gq59-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.2 GB at this context; duplicate traces of the same session: 4qmk5fsr, ppph3sl2, ps4ko7ya, xo6b9zg9; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 772.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 32k 772.6 16.5 102.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/khx3gq59",
   "source_url": "https://omlx.ai/benchmarks/performance/khx3gq59",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-khx3gq59-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.3 GB at this context; duplicate traces of the same session: 4qmk5fsr, ppph3sl2, ps4ko7ya, xo6b9zg9; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 755.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 64k 755.1 26.7 104.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/khx3gq59",
   "source_url": "https://omlx.ai/benchmarks/performance/khx3gq59",
   "tg_tokens": null,
   "tps": 26.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-qwvyuvmc-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: 4n3mzges, f5cvyyuz, fbye2z23, kryp2xou; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 245.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 245.0 41.1 26.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qwvyuvmc",
   "source_url": "https://omlx.ai/benchmarks/performance/qwvyuvmc",
   "tg_tokens": null,
   "tps": 41.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-qwvyuvmc-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.1 GB at this context; duplicate traces of the same session: 4n3mzges, f5cvyyuz, fbye2z23, kryp2xou; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 262.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 262.5 39.5 30.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qwvyuvmc",
   "source_url": "https://omlx.ai/benchmarks/performance/qwvyuvmc",
   "tg_tokens": null,
   "tps": 39.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-qwvyuvmc-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.6 GB at this context; duplicate traces of the same session: 4n3mzges, f5cvyyuz, fbye2z23, kryp2xou; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 258.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 8k 258.3 36.7 30.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qwvyuvmc",
   "source_url": "https://omlx.ai/benchmarks/performance/qwvyuvmc",
   "tg_tokens": null,
   "tps": 36.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-qwvyuvmc-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.5 GB at this context; duplicate traces of the same session: 4n3mzges, f5cvyyuz, fbye2z23, kryp2xou; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 249.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k 249.7 37.3 31.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qwvyuvmc",
   "source_url": "https://omlx.ai/benchmarks/performance/qwvyuvmc",
   "tg_tokens": null,
   "tps": 37.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-qwvyuvmc-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.3 GB at this context; duplicate traces of the same session: 4n3mzges, f5cvyyuz, fbye2z23, kryp2xou; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 231.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 231.7 31.8 33.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qwvyuvmc",
   "source_url": "https://omlx.ai/benchmarks/performance/qwvyuvmc",
   "tg_tokens": null,
   "tps": 31.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-npauevbs-ctx16384",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 41.8 GB; TTFT 98218 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 166.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "gemma-4-31B-it-oQ6e-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 6bit; 16k tokens 166.8 PP tok/s 14.3 TG tok/s 98218 TTFT (ms) 41.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/npauevbs",
   "source_url": "https://omlx.ai/benchmarks/performance/npauevbs",
   "tg_tokens": null,
   "tps": 14.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-4nh4afr3-ctx16384",
   "model": "gemma-4-31B-it",
   "notes": "qat; oMLX community benchmark trace, single run at 16k context; peak memory 41.5 GB; TTFT 98120 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 167.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "gemma-4-31B-it-qat-mxfp4 on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 16k tokens 167.0 PP tok/s 15.0 TG tok/s 98120 TTFT (ms) 41.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4nh4afr3",
   "source_url": "https://omlx.ai/benchmarks/performance/4nh4afr3",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1gjfur0b-ctx1024",
   "model": "DeepSeek-V4-Flash-0731-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 100.9 GB at this context; duplicate traces of the same session: 0ju2ycd8, 1j45l355, 6uuowdq9, lzwgxgyw, w9o7jmbj; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 422.6,
   "provenance": "community",
   "quant": "Mixed 2bit-3bit g64",
   "quote": "DeepSeek-V4-Flash-0731-Abliterated-MLX-Mixed-2bit-3bit-g64 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 1k 422.6 28.6 100.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1gjfur0b",
   "source_url": "https://omlx.ai/benchmarks/performance/1gjfur0b",
   "tg_tokens": null,
   "tps": 28.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1gjfur0b-ctx4096",
   "model": "DeepSeek-V4-Flash-0731-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: 0ju2ycd8, 1j45l355, 6uuowdq9, lzwgxgyw, w9o7jmbj; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 474.7,
   "provenance": "community",
   "quant": "Mixed 2bit-3bit g64",
   "quote": "DeepSeek-V4-Flash-0731-Abliterated-MLX-Mixed-2bit-3bit-g64 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 4k 474.7 26.3 101.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1gjfur0b",
   "source_url": "https://omlx.ai/benchmarks/performance/1gjfur0b",
   "tg_tokens": null,
   "tps": 26.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1gjfur0b-ctx8192",
   "model": "DeepSeek-V4-Flash-0731-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: 0ju2ycd8, 1j45l355, 6uuowdq9, lzwgxgyw, w9o7jmbj; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 477.6,
   "provenance": "community",
   "quant": "Mixed 2bit-3bit g64",
   "quote": "DeepSeek-V4-Flash-0731-Abliterated-MLX-Mixed-2bit-3bit-g64 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 8k 477.6 25.6 101.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1gjfur0b",
   "source_url": "https://omlx.ai/benchmarks/performance/1gjfur0b",
   "tg_tokens": null,
   "tps": 25.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1gjfur0b-ctx16384",
   "model": "DeepSeek-V4-Flash-0731-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 101.4 GB at this context; duplicate traces of the same session: 0ju2ycd8, 1j45l355, 6uuowdq9, lzwgxgyw, w9o7jmbj; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 468.5,
   "provenance": "community",
   "quant": "Mixed 2bit-3bit g64",
   "quote": "DeepSeek-V4-Flash-0731-Abliterated-MLX-Mixed-2bit-3bit-g64 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 16k 468.5 24.7 101.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1gjfur0b",
   "source_url": "https://omlx.ai/benchmarks/performance/1gjfur0b",
   "tg_tokens": null,
   "tps": 24.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1gjfur0b-ctx32768",
   "model": "DeepSeek-V4-Flash-0731-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: 0ju2ycd8, 1j45l355, 6uuowdq9, lzwgxgyw, w9o7jmbj; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 455.1,
   "provenance": "community",
   "quant": "Mixed 2bit-3bit g64",
   "quote": "DeepSeek-V4-Flash-0731-Abliterated-MLX-Mixed-2bit-3bit-g64 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 32k 455.1 24.0 101.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1gjfur0b",
   "source_url": "https://omlx.ai/benchmarks/performance/1gjfur0b",
   "tg_tokens": null,
   "tps": 24.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1gjfur0b-ctx65536",
   "model": "DeepSeek-V4-Flash-0731-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 101.8 GB at this context; duplicate traces of the same session: 0ju2ycd8, 1j45l355, 6uuowdq9, lzwgxgyw, w9o7jmbj; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 430.4,
   "provenance": "community",
   "quant": "Mixed 2bit-3bit g64",
   "quote": "DeepSeek-V4-Flash-0731-Abliterated-MLX-Mixed-2bit-3bit-g64 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 64k 430.4 23.1 101.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1gjfur0b",
   "source_url": "https://omlx.ai/benchmarks/performance/1gjfur0b",
   "tg_tokens": null,
   "tps": 23.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-e45rud90-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "fp16 weights; oMLX community benchmark trace; peak memory 7.7 GB at this context; duplicate traces of the same session: i558ptg6; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 461.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-9B-oQ4e-fp16 on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 1k 461.5 33.5 7.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e45rud90",
   "source_url": "https://omlx.ai/benchmarks/performance/e45rud90",
   "tg_tokens": null,
   "tps": 33.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-e45rud90-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "fp16 weights; oMLX community benchmark trace; peak memory 8.4 GB at this context; duplicate traces of the same session: i558ptg6; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 537.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-9B-oQ4e-fp16 on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 4k 537.0 36.8 8.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e45rud90",
   "source_url": "https://omlx.ai/benchmarks/performance/e45rud90",
   "tg_tokens": null,
   "tps": 36.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-l0a6z136-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: 3te05q0t; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 708.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k 708.1 25.9 6.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l0a6z136",
   "source_url": "https://omlx.ai/benchmarks/performance/l0a6z136",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-l0a6z136-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: 3te05q0t; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 916.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 916.5 25.6 6.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l0a6z136",
   "source_url": "https://omlx.ai/benchmarks/performance/l0a6z136",
   "tg_tokens": null,
   "tps": 25.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-uq3zmnvv-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 25.5 GB; TTFT 192478 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 681.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k tokens 681.0 PP tok/s 57.2 TG tok/s 192478 TTFT (ms) 25.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uq3zmnvv",
   "source_url": "https://omlx.ai/benchmarks/performance/uq3zmnvv",
   "tg_tokens": null,
   "tps": 57.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-39zjd2bx-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: 4jws2z9k; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 127.0 14.5 15.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/39zjd2bx",
   "source_url": "https://omlx.ai/benchmarks/performance/39zjd2bx",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-39zjd2bx-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 4jws2z9k; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 126.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 126.9 14.1 17.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/39zjd2bx",
   "source_url": "https://omlx.ai/benchmarks/performance/39zjd2bx",
   "tg_tokens": null,
   "tps": 14.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-62gpspan-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 0gu8ylqd; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 403.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 403.5 30.0 20.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/62gpspan",
   "source_url": "https://omlx.ai/benchmarks/performance/62gpspan",
   "tg_tokens": null,
   "tps": 30.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-62gpspan-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: 0gu8ylqd; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 387.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 387.1 14.3 21.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/62gpspan",
   "source_url": "https://omlx.ai/benchmarks/performance/62gpspan",
   "tg_tokens": null,
   "tps": 14.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-i53ndu8t-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 15.0 GB; TTFT 15513 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 66.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 66.0 PP tok/s 7.3 TG tok/s 15513 TTFT (ms) 15.0 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i53ndu8t",
   "source_url": "https://omlx.ai/benchmarks/performance/i53ndu8t",
   "tg_tokens": null,
   "tps": 7.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-sbejvhxs-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 15.0 GB; TTFT 15511 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 66.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 66.0 PP tok/s 7.2 TG tok/s 15511 TTFT (ms) 15.0 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sbejvhxs",
   "source_url": "https://omlx.ai/benchmarks/performance/sbejvhxs",
   "tg_tokens": null,
   "tps": 7.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-5yyra2ys-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 19.5 GB; TTFT 91159 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 359.5,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-oQ4 on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k tokens 359.5 PP tok/s 14.6 TG tok/s 91159 TTFT (ms) 19.5 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5yyra2ys",
   "source_url": "https://omlx.ai/benchmarks/performance/5yyra2ys",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wxylwp4v-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: ophi2twj, y6q2fgmr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 808.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 808.4 44.9 17.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wxylwp4v",
   "source_url": "https://omlx.ai/benchmarks/performance/wxylwp4v",
   "tg_tokens": null,
   "tps": 44.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wxylwp4v-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: ophi2twj, y6q2fgmr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 893.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 893.4 71.4 19.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wxylwp4v",
   "source_url": "https://omlx.ai/benchmarks/performance/wxylwp4v",
   "tg_tokens": null,
   "tps": 71.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wxylwp4v-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: ophi2twj, y6q2fgmr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 784.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 784.6 83.6 22.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wxylwp4v",
   "source_url": "https://omlx.ai/benchmarks/performance/wxylwp4v",
   "tg_tokens": null,
   "tps": 83.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-7eirgi8v-ctx1024",
   "model": "RavenX-CyberAgent-Qwen3.6-35B-A3B-Opus-4.7-OpenMythos-Pentester-BugHunter-RATH",
   "notes": "oMLX community benchmark trace; peak memory 65.8 GB at this context; duplicate traces of the same session: dzng4s83, q43m9wo1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1453.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "RavenX-CyberAgent-Qwen3.6-35B-A3B-Opus-4.7-OpenMythos-Pentester-BugHunter-RATH-mlx on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 1,453 63.6 65.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7eirgi8v",
   "source_url": "https://omlx.ai/benchmarks/performance/7eirgi8v",
   "tg_tokens": null,
   "tps": 63.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-7eirgi8v-ctx4096",
   "model": "RavenX-CyberAgent-Qwen3.6-35B-A3B-Opus-4.7-OpenMythos-Pentester-BugHunter-RATH",
   "notes": "oMLX community benchmark trace; peak memory 67.5 GB at this context; duplicate traces of the same session: dzng4s83, q43m9wo1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2178.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "RavenX-CyberAgent-Qwen3.6-35B-A3B-Opus-4.7-OpenMythos-Pentester-BugHunter-RATH-mlx on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 unknown; 4k 2,178 62.3 67.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7eirgi8v",
   "source_url": "https://omlx.ai/benchmarks/performance/7eirgi8v",
   "tg_tokens": null,
   "tps": 62.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-7eirgi8v-ctx8192",
   "model": "RavenX-CyberAgent-Qwen3.6-35B-A3B-Opus-4.7-OpenMythos-Pentester-BugHunter-RATH",
   "notes": "oMLX community benchmark trace; peak memory 67.6 GB at this context; duplicate traces of the same session: dzng4s83, q43m9wo1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1985.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "RavenX-CyberAgent-Qwen3.6-35B-A3B-Opus-4.7-OpenMythos-Pentester-BugHunter-RATH-mlx on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 1,985 61.1 67.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7eirgi8v",
   "source_url": "https://omlx.ai/benchmarks/performance/7eirgi8v",
   "tg_tokens": null,
   "tps": 61.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-v5t77ycj-ctx4096",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: csj1vog8, fp6nmk76, ljj5qyaq; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 125.4 17.8 20.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v5t77ycj",
   "source_url": "https://omlx.ai/benchmarks/performance/v5t77ycj",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-v5t77ycj-ctx16384",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: csj1vog8, fp6nmk76, ljj5qyaq; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 116.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k 116.9 16.3 22.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v5t77ycj",
   "source_url": "https://omlx.ai/benchmarks/performance/v5t77ycj",
   "tg_tokens": null,
   "tps": 16.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-v5t77ycj-ctx65536",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: csj1vog8, fp6nmk76, ljj5qyaq; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 91.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 64k 91.8 11.0 25.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v5t77ycj",
   "source_url": "https://omlx.ai/benchmarks/performance/v5t77ycj",
   "tg_tokens": null,
   "tps": 11.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-v5t77ycj-ctx131072",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 29.3 GB at this context; duplicate traces of the same session: csj1vog8, fp6nmk76, ljj5qyaq; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 71.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 128k 71.5 6.0 29.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v5t77ycj",
   "source_url": "https://omlx.ai/benchmarks/performance/v5t77ycj",
   "tg_tokens": null,
   "tps": 6.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-6xzlj0v9-ctx131072",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 20.9 GB; TTFT 238359 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 549.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k tokens 549.9 PP tok/s 30.0 TG tok/s 238359 TTFT (ms) 20.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6xzlj0v9",
   "source_url": "https://omlx.ai/benchmarks/performance/6xzlj0v9",
   "tg_tokens": null,
   "tps": 30.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kiyb1dsn-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: kleszojo, rgkkmv73; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 915.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mlx-community--Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 915.1 33.2 34.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kiyb1dsn",
   "source_url": "https://omlx.ai/benchmarks/performance/kiyb1dsn",
   "tg_tokens": null,
   "tps": 33.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kiyb1dsn-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 35.6 GB at this context; duplicate traces of the same session: kleszojo, rgkkmv73; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 846.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mlx-community--Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 846.8 30.3 35.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kiyb1dsn",
   "source_url": "https://omlx.ai/benchmarks/performance/kiyb1dsn",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kiyb1dsn-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 36.6 GB at this context; duplicate traces of the same session: kleszojo, rgkkmv73; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 771.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mlx-community--Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 771.1 31.7 36.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kiyb1dsn",
   "source_url": "https://omlx.ai/benchmarks/performance/kiyb1dsn",
   "tg_tokens": null,
   "tps": 31.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-m4plg5bf-ctx1024",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: 63omk5oz, 9oqcu1j8, hrev7o64; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 420.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "AREX-2-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 420.2 9.6 28.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m4plg5bf",
   "source_url": "https://omlx.ai/benchmarks/performance/m4plg5bf",
   "tg_tokens": null,
   "tps": 9.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-m4plg5bf-ctx8192",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 29.8 GB at this context; duplicate traces of the same session: 63omk5oz, 9oqcu1j8, hrev7o64; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 475.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "AREX-2-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 475.4 9.7 29.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m4plg5bf",
   "source_url": "https://omlx.ai/benchmarks/performance/m4plg5bf",
   "tg_tokens": null,
   "tps": 9.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-m4plg5bf-ctx32768",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 31.3 GB at this context; duplicate traces of the same session: 63omk5oz, 9oqcu1j8, hrev7o64; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 423.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "AREX-2-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 32k 423.4 9.2 31.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m4plg5bf",
   "source_url": "https://omlx.ai/benchmarks/performance/m4plg5bf",
   "tg_tokens": null,
   "tps": 9.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-m4plg5bf-ctx131072",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: 63omk5oz, 9oqcu1j8, hrev7o64; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 314.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "AREX-2-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 128k 314.8 7.0 37.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m4plg5bf",
   "source_url": "https://omlx.ai/benchmarks/performance/m4plg5bf",
   "tg_tokens": null,
   "tps": 7.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-t0j713hk-ctx1024",
   "model": "Qwen3.6-35B-A3B-Abliterated-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 8p8qwb96; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3712.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Abliterated-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 3,712 125.6 23.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t0j713hk",
   "source_url": "https://omlx.ai/benchmarks/performance/t0j713hk",
   "tg_tokens": null,
   "tps": 125.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-t0j713hk-ctx4096",
   "model": "Qwen3.6-35B-A3B-Abliterated-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: 8p8qwb96; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5777.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Abliterated-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 5,777 126.0 24.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t0j713hk",
   "source_url": "https://omlx.ai/benchmarks/performance/t0j713hk",
   "tg_tokens": null,
   "tps": 126.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-bgt93q0p-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 65.8 GB at this context; duplicate traces of the same session: eehncwoq, urz407lk, zq20q0n3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1586.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.6-35B-A3B-bf16 on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 bf16; 1k 1,586 62.7 65.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bgt93q0p",
   "source_url": "https://omlx.ai/benchmarks/performance/bgt93q0p",
   "tg_tokens": null,
   "tps": 62.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-bgt93q0p-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 67.5 GB at this context; duplicate traces of the same session: eehncwoq, urz407lk, zq20q0n3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2123.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.6-35B-A3B-bf16 on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 bf16; 4k 2,123 62.5 67.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bgt93q0p",
   "source_url": "https://omlx.ai/benchmarks/performance/bgt93q0p",
   "tg_tokens": null,
   "tps": 62.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-bgt93q0p-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 67.6 GB at this context; duplicate traces of the same session: eehncwoq, urz407lk, zq20q0n3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2014.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.6-35B-A3B-bf16 on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 bf16; 8k 2,014 63.4 67.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bgt93q0p",
   "source_url": "https://omlx.ai/benchmarks/performance/bgt93q0p",
   "tg_tokens": null,
   "tps": 63.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-bgt93q0p-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 67.9 GB at this context; duplicate traces of the same session: eehncwoq, urz407lk, zq20q0n3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1849.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.6-35B-A3B-bf16 on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 bf16; 16k 1,849 59.5 67.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bgt93q0p",
   "source_url": "https://omlx.ai/benchmarks/performance/bgt93q0p",
   "tg_tokens": null,
   "tps": 59.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-diqkk5of-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 14.2 GB; TTFT 3412 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 300.1,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.8-27B-3bit-mlx on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 3bit; 1k tokens 300.1 PP tok/s 22.1 TG tok/s 3412 TTFT (ms) 14.2 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/diqkk5of",
   "source_url": "https://omlx.ai/benchmarks/performance/diqkk5of",
   "tg_tokens": null,
   "tps": 22.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-o3wouedr-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 13.2 GB; TTFT 3334 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 307.2,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.8-27B-3bit-mlx on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 3bit; 1k tokens 307.2 PP tok/s 11.6 TG tok/s 3334 TTFT (ms) 13.2 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o3wouedr",
   "source_url": "https://omlx.ai/benchmarks/performance/o3wouedr",
   "tg_tokens": null,
   "tps": 11.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-2pp4976a-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 14.2 GB; TTFT 3521 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 290.8,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.8-27B-3bit-mlx on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 3bit; 1k tokens 290.8 PP tok/s 19.4 TG tok/s 3521 TTFT (ms) 14.2 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2pp4976a",
   "source_url": "https://omlx.ai/benchmarks/performance/2pp4976a",
   "tg_tokens": null,
   "tps": 19.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cuk08fko-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 2dlvpy60, a651dqe5, aypbqgcw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 980.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 980.0 32.9 17.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cuk08fko",
   "source_url": "https://omlx.ai/benchmarks/performance/cuk08fko",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cuk08fko-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 2dlvpy60, a651dqe5, aypbqgcw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 974.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 974.4 31.1 17.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cuk08fko",
   "source_url": "https://omlx.ai/benchmarks/performance/cuk08fko",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cuk08fko-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 2dlvpy60, a651dqe5, aypbqgcw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 908.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 908.2 29.4 18.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cuk08fko",
   "source_url": "https://omlx.ai/benchmarks/performance/cuk08fko",
   "tg_tokens": null,
   "tps": 29.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cuk08fko-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: 2dlvpy60, a651dqe5, aypbqgcw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 630.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 630.1 20.9 24.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cuk08fko",
   "source_url": "https://omlx.ai/benchmarks/performance/cuk08fko",
   "tg_tokens": null,
   "tps": 20.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-pf5ad8c1-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 18.9 GB; TTFT 90917 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 360.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k tokens 360.4 PP tok/s 15.0 TG tok/s 90917 TTFT (ms) 18.9 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pf5ad8c1",
   "source_url": "https://omlx.ai/benchmarks/performance/pf5ad8c1",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-04igcoo3-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 20.0 GB; TTFT 90363 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 362.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k tokens 362.6 PP tok/s 22.9 TG tok/s 90363 TTFT (ms) 20.0 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/04igcoo3",
   "source_url": "https://omlx.ai/benchmarks/performance/04igcoo3",
   "tg_tokens": null,
   "tps": 22.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-8ld98jpi-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 15.0 GB; TTFT 15517 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 66.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 66.0 PP tok/s 7.3 TG tok/s 15517 TTFT (ms) 15.0 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8ld98jpi",
   "source_url": "https://omlx.ai/benchmarks/performance/8ld98jpi",
   "tg_tokens": null,
   "tps": 7.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-u4z09kok-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 16.1 GB; TTFT 15410 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 66.5,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 66.5 PP tok/s 5.9 TG tok/s 15410 TTFT (ms) 16.1 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u4z09kok",
   "source_url": "https://omlx.ai/benchmarks/performance/u4z09kok",
   "tg_tokens": null,
   "tps": 5.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-1as1wyy8-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 15.0 GB; TTFT 15511 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 66.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 66.0 PP tok/s 7.3 TG tok/s 15511 TTFT (ms) 15.0 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1as1wyy8",
   "source_url": "https://omlx.ai/benchmarks/performance/1as1wyy8",
   "tg_tokens": null,
   "tps": 7.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-la2wmsa9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: eb8b5fjp, jtan9nr6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 868.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 868.7 71.3 17.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/la2wmsa9",
   "source_url": "https://omlx.ai/benchmarks/performance/la2wmsa9",
   "tg_tokens": null,
   "tps": 71.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-la2wmsa9-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: eb8b5fjp, jtan9nr6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 908.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 908.5 71.0 19.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/la2wmsa9",
   "source_url": "https://omlx.ai/benchmarks/performance/la2wmsa9",
   "tg_tokens": null,
   "tps": 71.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-la2wmsa9-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: eb8b5fjp, jtan9nr6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 779.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 779.7 140.7 22.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/la2wmsa9",
   "source_url": "https://omlx.ai/benchmarks/performance/la2wmsa9",
   "tg_tokens": null,
   "tps": 140.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-prj1zcnn-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 70.7 GB; TTFT 68938 ms; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 475.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 32k tokens 475.3 PP tok/s 40.2 TG tok/s 68938 TTFT (ms) 70.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/prj1zcnn",
   "source_url": "https://omlx.ai/benchmarks/performance/prj1zcnn",
   "tg_tokens": null,
   "tps": 40.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-v16w2f77-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 4tw708wf, et0xewck, f87ofa39, flc102cc, g4frpgak, q9ymhr2u, wxka2jne; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 703.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 703.1 48.1 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v16w2f77",
   "source_url": "https://omlx.ai/benchmarks/performance/v16w2f77",
   "tg_tokens": null,
   "tps": 48.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-v16w2f77-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 4tw708wf, et0xewck, f87ofa39, flc102cc, g4frpgak, q9ymhr2u, wxka2jne; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 856.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 856.0 60.6 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v16w2f77",
   "source_url": "https://omlx.ai/benchmarks/performance/v16w2f77",
   "tg_tokens": null,
   "tps": 60.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-v16w2f77-ctx8192",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: 4tw708wf, et0xewck, f87ofa39, flc102cc, g4frpgak, q9ymhr2u, wxka2jne; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 823.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 8k 823.0 64.4 24.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v16w2f77",
   "source_url": "https://omlx.ai/benchmarks/performance/v16w2f77",
   "tg_tokens": null,
   "tps": 64.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-v16w2f77-ctx16384",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 4tw708wf, et0xewck, f87ofa39, flc102cc, g4frpgak, q9ymhr2u, wxka2jne; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 717.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k 717.5 62.1 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v16w2f77",
   "source_url": "https://omlx.ai/benchmarks/performance/v16w2f77",
   "tg_tokens": null,
   "tps": 62.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-v16w2f77-ctx32768",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: 4tw708wf, et0xewck, f87ofa39, flc102cc, g4frpgak, q9ymhr2u, wxka2jne; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 586.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 32k 586.9 54.8 24.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v16w2f77",
   "source_url": "https://omlx.ai/benchmarks/performance/v16w2f77",
   "tg_tokens": null,
   "tps": 54.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-v16w2f77-ctx65536",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: 4tw708wf, et0xewck, f87ofa39, flc102cc, g4frpgak, q9ymhr2u, wxka2jne; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 425.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 64k 425.1 42.7 24.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v16w2f77",
   "source_url": "https://omlx.ai/benchmarks/performance/v16w2f77",
   "tg_tokens": null,
   "tps": 42.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-v16w2f77-ctx131072",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: 4tw708wf, et0xewck, f87ofa39, flc102cc, g4frpgak, q9ymhr2u, wxka2jne; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 271.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 128k 271.0 31.0 26.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v16w2f77",
   "source_url": "https://omlx.ai/benchmarks/performance/v16w2f77",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-v16w2f77-ctx199680",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: 4tw708wf, et0xewck, f87ofa39, flc102cc, g4frpgak, q9ymhr2u, wxka2jne; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 201.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 195k 201.1 24.6 28.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v16w2f77",
   "source_url": "https://omlx.ai/benchmarks/performance/v16w2f77",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ij5cbxtz-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: f03nvhh6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 814.3,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 814.3 72.2 11.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ij5cbxtz",
   "source_url": "https://omlx.ai/benchmarks/performance/ij5cbxtz",
   "tg_tokens": null,
   "tps": 72.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ij5cbxtz-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: f03nvhh6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 816.8,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 816.8 68.2 11.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ij5cbxtz",
   "source_url": "https://omlx.ai/benchmarks/performance/ij5cbxtz",
   "tg_tokens": null,
   "tps": 68.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-hkvld1lr-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: 7lqth0wb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1899.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,899 79.9 11.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hkvld1lr",
   "source_url": "https://omlx.ai/benchmarks/performance/hkvld1lr",
   "tg_tokens": null,
   "tps": 79.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-hkvld1lr-ctx16384",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 12.1 GB at this context; duplicate traces of the same session: 7lqth0wb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2627.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,627 66.2 12.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hkvld1lr",
   "source_url": "https://omlx.ai/benchmarks/performance/hkvld1lr",
   "tg_tokens": null,
   "tps": 66.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-sops9ymq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: 0zqv32s8, 6hd2r6fi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 955.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mlx-community--Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 955.5 33.3 34.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sops9ymq",
   "source_url": "https://omlx.ai/benchmarks/performance/sops9ymq",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-sops9ymq-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 35.6 GB at this context; duplicate traces of the same session: 0zqv32s8, 6hd2r6fi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 843.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mlx-community--Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 843.7 33.5 35.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sops9ymq",
   "source_url": "https://omlx.ai/benchmarks/performance/sops9ymq",
   "tg_tokens": null,
   "tps": 33.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-sops9ymq-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 36.6 GB at this context; duplicate traces of the same session: 0zqv32s8, 6hd2r6fi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 770.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mlx-community--Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 770.0 32.0 36.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sops9ymq",
   "source_url": "https://omlx.ai/benchmarks/performance/sops9ymq",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-vmk9znhj-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "mlx-community; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 02cdgnhp, ns93o952; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2771.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "mlx-community--gemma-4-26b-a4b-it-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 2,771 101.1 22.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vmk9znhj",
   "source_url": "https://omlx.ai/benchmarks/performance/vmk9znhj",
   "tg_tokens": null,
   "tps": 101.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-vmk9znhj-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "mlx-community; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 02cdgnhp, ns93o952; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3175.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "mlx-community--gemma-4-26b-a4b-it-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 3,175 78.2 22.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vmk9znhj",
   "source_url": "https://omlx.ai/benchmarks/performance/vmk9znhj",
   "tg_tokens": null,
   "tps": 78.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-vmk9znhj-ctx32768",
   "model": "gemma-4-26B-A4B-it",
   "notes": "mlx-community; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: 02cdgnhp, ns93o952; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2918.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "mlx-community--gemma-4-26b-a4b-it-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 2,918 62.8 23.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vmk9znhj",
   "source_url": "https://omlx.ai/benchmarks/performance/vmk9znhj",
   "tg_tokens": null,
   "tps": 62.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jtdiqkv7-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "mlx-community; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: jcbieam3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 659.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mlx-community--gemma-4-31b-it-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 659.2 26.1 37.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jtdiqkv7",
   "source_url": "https://omlx.ai/benchmarks/performance/jtdiqkv7",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jtdiqkv7-ctx16384",
   "model": "gemma-4-31B-it",
   "notes": "mlx-community; oMLX community benchmark trace; peak memory 41.0 GB at this context; duplicate traces of the same session: jcbieam3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 468.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mlx-community--gemma-4-31b-it-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 468.2 29.3 41.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jtdiqkv7",
   "source_url": "https://omlx.ai/benchmarks/performance/jtdiqkv7",
   "tg_tokens": null,
   "tps": 29.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-0m4xm6kr-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: e6mowaa4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 257.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 257.8 27.9 16.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0m4xm6kr",
   "source_url": "https://omlx.ai/benchmarks/performance/0m4xm6kr",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-0m4xm6kr-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: e6mowaa4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 213.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 213.6 24.0 21.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0m4xm6kr",
   "source_url": "https://omlx.ai/benchmarks/performance/0m4xm6kr",
   "tg_tokens": null,
   "tps": 24.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-wvy5h7o9-ctx8192",
   "model": "K2-Horizon-7B",
   "notes": "oMLX community benchmark trace; peak memory 10.7 GB at this context; duplicate traces of the same session: xlgre3lw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 830.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "K2-Horizon-7B-oQ8e on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 830.6 46.9 10.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wvy5h7o9",
   "source_url": "https://omlx.ai/benchmarks/performance/wvy5h7o9",
   "tg_tokens": null,
   "tps": 46.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-wvy5h7o9-ctx32768",
   "model": "K2-Horizon-7B",
   "notes": "oMLX community benchmark trace; peak memory 14.1 GB at this context; duplicate traces of the same session: xlgre3lw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 546.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "K2-Horizon-7B-oQ8e on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 546.7 34.9 14.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wvy5h7o9",
   "source_url": "https://omlx.ai/benchmarks/performance/wvy5h7o9",
   "tg_tokens": null,
   "tps": 34.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-2aqgs2z8-ctx1024",
   "model": "KAT-Coder-V2.5-Dev",
   "notes": "oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 712.8,
   "provenance": "community",
   "quant": "OptiQ 4bit",
   "quote": "KAT-Coder-V2.5-Dev-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 712.8 78.1 21.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2aqgs2z8",
   "source_url": "https://omlx.ai/benchmarks/performance/2aqgs2z8",
   "tg_tokens": null,
   "tps": 78.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-2aqgs2z8-ctx8192",
   "model": "KAT-Coder-V2.5-Dev",
   "notes": "oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3234.0,
   "provenance": "community",
   "quant": "OptiQ 4bit",
   "quote": "KAT-Coder-V2.5-Dev-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 3,234 67.7 22.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2aqgs2z8",
   "source_url": "https://omlx.ai/benchmarks/performance/2aqgs2z8",
   "tg_tokens": null,
   "tps": 67.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-2aqgs2z8-ctx32768",
   "model": "KAT-Coder-V2.5-Dev",
   "notes": "oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2489.0,
   "provenance": "community",
   "quant": "OptiQ 4bit",
   "quote": "KAT-Coder-V2.5-Dev-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 2,489 60.4 22.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2aqgs2z8",
   "source_url": "https://omlx.ai/benchmarks/performance/2aqgs2z8",
   "tg_tokens": null,
   "tps": 60.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-2aqgs2z8-ctx131072",
   "model": "KAT-Coder-V2.5-Dev",
   "notes": "oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1270.0,
   "provenance": "community",
   "quant": "OptiQ 4bit",
   "quote": "KAT-Coder-V2.5-Dev-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 128k 1,270 39.0 24.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2aqgs2z8",
   "source_url": "https://omlx.ai/benchmarks/performance/2aqgs2z8",
   "tg_tokens": null,
   "tps": 39.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-4l7xuar7-ctx1024",
   "model": "KAT-Coder-V2.5-Dev",
   "notes": "REAP-18B distilled; oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: de0zl4rv, ssq7ff08, zge3rgo6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2647.0,
   "provenance": "community",
   "quant": "OptiQ 4bit",
   "quote": "KAT-Coder-V2.5-Dev-OptiQ-4bit-REAP-18B on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,647 71.5 12.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4l7xuar7",
   "source_url": "https://omlx.ai/benchmarks/performance/4l7xuar7",
   "tg_tokens": null,
   "tps": 71.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-4l7xuar7-ctx8192",
   "model": "KAT-Coder-V2.5-Dev",
   "notes": "REAP-18B distilled; oMLX community benchmark trace; peak memory 13.3 GB at this context; duplicate traces of the same session: de0zl4rv, ssq7ff08, zge3rgo6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3235.0,
   "provenance": "community",
   "quant": "OptiQ 4bit",
   "quote": "KAT-Coder-V2.5-Dev-OptiQ-4bit-REAP-18B on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 3,235 68.0 13.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4l7xuar7",
   "source_url": "https://omlx.ai/benchmarks/performance/4l7xuar7",
   "tg_tokens": null,
   "tps": 68.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-4l7xuar7-ctx32768",
   "model": "KAT-Coder-V2.5-Dev",
   "notes": "REAP-18B distilled; oMLX community benchmark trace; peak memory 13.8 GB at this context; duplicate traces of the same session: de0zl4rv, ssq7ff08, zge3rgo6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2621.0,
   "provenance": "community",
   "quant": "OptiQ 4bit",
   "quote": "KAT-Coder-V2.5-Dev-OptiQ-4bit-REAP-18B on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 2,621 62.3 13.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4l7xuar7",
   "source_url": "https://omlx.ai/benchmarks/performance/4l7xuar7",
   "tg_tokens": null,
   "tps": 62.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-4l7xuar7-ctx131072",
   "model": "KAT-Coder-V2.5-Dev",
   "notes": "REAP-18B distilled; oMLX community benchmark trace; peak memory 15.7 GB at this context; duplicate traces of the same session: de0zl4rv, ssq7ff08, zge3rgo6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1291.0,
   "provenance": "community",
   "quant": "OptiQ 4bit",
   "quote": "KAT-Coder-V2.5-Dev-OptiQ-4bit-REAP-18B on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 128k 1,291 37.4 15.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4l7xuar7",
   "source_url": "https://omlx.ai/benchmarks/performance/4l7xuar7",
   "tg_tokens": null,
   "tps": 37.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-nwx2ocul-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: g3wtwh3s; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 219.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.5-4B-oQ8e-fp16-mtp on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 8bit; 1k 219.4 14.2 6.5 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nwx2ocul",
   "source_url": "https://omlx.ai/benchmarks/performance/nwx2ocul",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-nwx2ocul-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.4 GB at this context; duplicate traces of the same session: g3wtwh3s; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 231.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.5-4B-oQ8e-fp16-mtp on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 8bit; 4k 231.1 13.6 7.4 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nwx2ocul",
   "source_url": "https://omlx.ai/benchmarks/performance/nwx2ocul",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-m60y0jhf-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 65.8 GB at this context; duplicate traces of the same session: 4a4ro11i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1340.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.6-35B-A3B-bf16 on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 bf16; 1k 1,340 63.7 65.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m60y0jhf",
   "source_url": "https://omlx.ai/benchmarks/performance/m60y0jhf",
   "tg_tokens": null,
   "tps": 63.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-m60y0jhf-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 67.5 GB at this context; duplicate traces of the same session: 4a4ro11i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1918.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.6-35B-A3B-bf16 on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 bf16; 4k 1,918 62.6 67.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m60y0jhf",
   "source_url": "https://omlx.ai/benchmarks/performance/m60y0jhf",
   "tg_tokens": null,
   "tps": 62.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-f0prhmtf-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 20.7 GB; TTFT 1409 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 727.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k tokens 727.0 PP tok/s 77.4 TG tok/s 1409 TTFT (ms) 20.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f0prhmtf",
   "source_url": "https://omlx.ai/benchmarks/performance/f0prhmtf",
   "tg_tokens": null,
   "tps": 77.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-8f8kg1oc-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 14.2 GB; TTFT 3403 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 300.9,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.8-27B-3bit-mlx on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 3bit; 1k tokens 300.9 PP tok/s 19.8 TG tok/s 3403 TTFT (ms) 14.2 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8f8kg1oc",
   "source_url": "https://omlx.ai/benchmarks/performance/8f8kg1oc",
   "tg_tokens": null,
   "tps": 19.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-glh1qyw5-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 14.2 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 285.8,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.8-27B-3bit-mlx on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 3bit; 1k 285.8 23.0 14.2 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/glh1qyw5",
   "source_url": "https://omlx.ai/benchmarks/performance/glh1qyw5",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-glh1qyw5-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 315.5,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.8-27B-3bit-mlx on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 3bit; 4k 315.5 16.8 15.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/glh1qyw5",
   "source_url": "https://omlx.ai/benchmarks/performance/glh1qyw5",
   "tg_tokens": null,
   "tps": 16.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ugwrumpa-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 20.4 GB; TTFT 93661 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 349.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k tokens 349.9 PP tok/s 7.3 TG tok/s 93661 TTFT (ms) 20.4 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ugwrumpa",
   "source_url": "https://omlx.ai/benchmarks/performance/ugwrumpa",
   "tg_tokens": null,
   "tps": 7.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-uwf3brhi-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 446.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 446.0 17.3 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uwf3brhi",
   "source_url": "https://omlx.ai/benchmarks/performance/uwf3brhi",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-uwf3brhi-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 475.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 475.9 17.1 17.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uwf3brhi",
   "source_url": "https://omlx.ai/benchmarks/performance/uwf3brhi",
   "tg_tokens": null,
   "tps": 17.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5fbuwlfq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: 82hi2hfw, pn9emi8o, sdccjs4s, tt1dyb86, txaukvq6, y1o4vncp, zudlzjbb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 332.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 332.0 37.2 33.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5fbuwlfq",
   "source_url": "https://omlx.ai/benchmarks/performance/5fbuwlfq",
   "tg_tokens": null,
   "tps": 37.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5fbuwlfq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 37.5 GB at this context; duplicate traces of the same session: 82hi2hfw, pn9emi8o, sdccjs4s, tt1dyb86, txaukvq6, y1o4vncp, zudlzjbb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 337.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 337.8 38.4 37.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5fbuwlfq",
   "source_url": "https://omlx.ai/benchmarks/performance/5fbuwlfq",
   "tg_tokens": null,
   "tps": 38.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5fbuwlfq-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 38.5 GB at this context; duplicate traces of the same session: 82hi2hfw, pn9emi8o, sdccjs4s, tt1dyb86, txaukvq6, y1o4vncp, zudlzjbb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 333.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 8k 333.2 34.4 38.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5fbuwlfq",
   "source_url": "https://omlx.ai/benchmarks/performance/5fbuwlfq",
   "tg_tokens": null,
   "tps": 34.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5fbuwlfq-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 37.7 GB at this context; duplicate traces of the same session: 82hi2hfw, pn9emi8o, sdccjs4s, tt1dyb86, txaukvq6, y1o4vncp, zudlzjbb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 321.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 16k 321.3 38.8 37.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5fbuwlfq",
   "source_url": "https://omlx.ai/benchmarks/performance/5fbuwlfq",
   "tg_tokens": null,
   "tps": 38.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5fbuwlfq-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 38.7 GB at this context; duplicate traces of the same session: 82hi2hfw, pn9emi8o, sdccjs4s, tt1dyb86, txaukvq6, y1o4vncp, zudlzjbb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 298.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 32k 298.6 38.8 38.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5fbuwlfq",
   "source_url": "https://omlx.ai/benchmarks/performance/5fbuwlfq",
   "tg_tokens": null,
   "tps": 38.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5fbuwlfq-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 40.7 GB at this context; duplicate traces of the same session: 82hi2hfw, pn9emi8o, sdccjs4s, tt1dyb86, txaukvq6, y1o4vncp, zudlzjbb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 260.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 64k 260.2 33.8 40.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5fbuwlfq",
   "source_url": "https://omlx.ai/benchmarks/performance/5fbuwlfq",
   "tg_tokens": null,
   "tps": 33.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5fbuwlfq-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 44.7 GB at this context; duplicate traces of the same session: 82hi2hfw, pn9emi8o, sdccjs4s, tt1dyb86, txaukvq6, y1o4vncp, zudlzjbb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 128k 205.8 30.7 44.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5fbuwlfq",
   "source_url": "https://omlx.ai/benchmarks/performance/5fbuwlfq",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5fbuwlfq-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 49.0 GB at this context; duplicate traces of the same session: 82hi2hfw, pn9emi8o, sdccjs4s, tt1dyb86, txaukvq6, y1o4vncp, zudlzjbb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 168.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 195k 168.5 24.4 49.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5fbuwlfq",
   "source_url": "https://omlx.ai/benchmarks/performance/5fbuwlfq",
   "tg_tokens": null,
   "tps": 24.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-po0j95co-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: zc5zk4fa; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 437.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 437.3 14.0 20.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/po0j95co",
   "source_url": "https://omlx.ai/benchmarks/performance/po0j95co",
   "tg_tokens": null,
   "tps": 14.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-po0j95co-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: zc5zk4fa; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 486.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 486.4 13.9 21.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/po0j95co",
   "source_url": "https://omlx.ai/benchmarks/performance/po0j95co",
   "tg_tokens": null,
   "tps": 13.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-tkhj0b44-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 15.0 GB; TTFT 15519 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 66.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 66.0 PP tok/s 7.2 TG tok/s 15519 TTFT (ms) 15.0 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tkhj0b44",
   "source_url": "https://omlx.ai/benchmarks/performance/tkhj0b44",
   "tg_tokens": null,
   "tps": 7.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-l7szcwot-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 16.1 GB; TTFT 15406 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 66.5,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 66.5 PP tok/s 4.6 TG tok/s 15406 TTFT (ms) 16.1 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l7szcwot",
   "source_url": "https://omlx.ai/benchmarks/performance/l7szcwot",
   "tg_tokens": null,
   "tps": 4.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-45y1lb1u-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 15.0 GB; TTFT 15533 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 65.9,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 65.9 PP tok/s 7.2 TG tok/s 15533 TTFT (ms) 15.0 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/45y1lb1u",
   "source_url": "https://omlx.ai/benchmarks/performance/45y1lb1u",
   "tg_tokens": null,
   "tps": 7.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xdevprqm-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: gu1jg3ih, j6onrzlg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 878.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 878.1 58.5 17.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xdevprqm",
   "source_url": "https://omlx.ai/benchmarks/performance/xdevprqm",
   "tg_tokens": null,
   "tps": 58.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xdevprqm-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: gu1jg3ih, j6onrzlg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 906.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 906.9 72.2 19.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xdevprqm",
   "source_url": "https://omlx.ai/benchmarks/performance/xdevprqm",
   "tg_tokens": null,
   "tps": 72.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-xdevprqm-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: gu1jg3ih, j6onrzlg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 783.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 783.4 140.9 22.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xdevprqm",
   "source_url": "https://omlx.ai/benchmarks/performance/xdevprqm",
   "tg_tokens": null,
   "tps": 140.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-o1rgk6ph-ctx1024",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: 9b2audr0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 85.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-31B-it-oQ8e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 1k 85.2 20.3 33.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o1rgk6ph",
   "source_url": "https://omlx.ai/benchmarks/performance/o1rgk6ph",
   "tg_tokens": null,
   "tps": 20.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-o1rgk6ph-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.0 GB at this context; duplicate traces of the same session: 9b2audr0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 84.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-31B-it-oQ8e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 4k 84.1 18.0 35.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o1rgk6ph",
   "source_url": "https://omlx.ai/benchmarks/performance/o1rgk6ph",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-yo8tfj6e-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 197.5 GB at this context; duplicate traces of the same session: 226xpsk1, 5tlnkumd, 7h9itid1, b4wxtdmf, py4394db, wjm4u43a, zlye1owy; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 513.8,
   "provenance": "community",
   "quant": "oQ3e-2b27-CED",
   "quote": "keys-Mac-oMLX-0.7.0.dev2-DeepSeek-V4.1-Flash-oQ3e-2b27-CED-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 unknown; 1k 513.8 31.4 197.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yo8tfj6e",
   "source_url": "https://omlx.ai/benchmarks/performance/yo8tfj6e",
   "tg_tokens": null,
   "tps": 31.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-yo8tfj6e-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 198.0 GB at this context; duplicate traces of the same session: 226xpsk1, 5tlnkumd, 7h9itid1, b4wxtdmf, py4394db, wjm4u43a, zlye1owy; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 799.3,
   "provenance": "community",
   "quant": "oQ3e-2b27-CED",
   "quote": "keys-Mac-oMLX-0.7.0.dev2-DeepSeek-V4.1-Flash-oQ3e-2b27-CED-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 unknown; 4k 799.3 29.2 198.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yo8tfj6e",
   "source_url": "https://omlx.ai/benchmarks/performance/yo8tfj6e",
   "tg_tokens": null,
   "tps": 29.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-yo8tfj6e-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 198.0 GB at this context; duplicate traces of the same session: 226xpsk1, 5tlnkumd, 7h9itid1, b4wxtdmf, py4394db, wjm4u43a, zlye1owy; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 904.9,
   "provenance": "community",
   "quant": "oQ3e-2b27-CED",
   "quote": "keys-Mac-oMLX-0.7.0.dev2-DeepSeek-V4.1-Flash-oQ3e-2b27-CED-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 unknown; 8k 904.9 27.3 198.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yo8tfj6e",
   "source_url": "https://omlx.ai/benchmarks/performance/yo8tfj6e",
   "tg_tokens": null,
   "tps": 27.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-yo8tfj6e-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 198.1 GB at this context; duplicate traces of the same session: 226xpsk1, 5tlnkumd, 7h9itid1, b4wxtdmf, py4394db, wjm4u43a, zlye1owy; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 964.8,
   "provenance": "community",
   "quant": "oQ3e-2b27-CED",
   "quote": "keys-Mac-oMLX-0.7.0.dev2-DeepSeek-V4.1-Flash-oQ3e-2b27-CED-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 unknown; 16k 964.8 42.3 198.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yo8tfj6e",
   "source_url": "https://omlx.ai/benchmarks/performance/yo8tfj6e",
   "tg_tokens": null,
   "tps": 42.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-yo8tfj6e-ctx32768",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 198.1 GB at this context; duplicate traces of the same session: 226xpsk1, 5tlnkumd, 7h9itid1, b4wxtdmf, py4394db, wjm4u43a, zlye1owy; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1020.0,
   "provenance": "community",
   "quant": "oQ3e-2b27-CED",
   "quote": "keys-Mac-oMLX-0.7.0.dev2-DeepSeek-V4.1-Flash-oQ3e-2b27-CED-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 unknown; 32k 1,020 41.8 198.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yo8tfj6e",
   "source_url": "https://omlx.ai/benchmarks/performance/yo8tfj6e",
   "tg_tokens": null,
   "tps": 41.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-yo8tfj6e-ctx65536",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 198.1 GB at this context; duplicate traces of the same session: 226xpsk1, 5tlnkumd, 7h9itid1, b4wxtdmf, py4394db, wjm4u43a, zlye1owy; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 890.5,
   "provenance": "community",
   "quant": "oQ3e-2b27-CED",
   "quote": "keys-Mac-oMLX-0.7.0.dev2-DeepSeek-V4.1-Flash-oQ3e-2b27-CED-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 unknown; 64k 890.5 39.4 198.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yo8tfj6e",
   "source_url": "https://omlx.ai/benchmarks/performance/yo8tfj6e",
   "tg_tokens": null,
   "tps": 39.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-yo8tfj6e-ctx131072",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 198.2 GB at this context; duplicate traces of the same session: 226xpsk1, 5tlnkumd, 7h9itid1, b4wxtdmf, py4394db, wjm4u43a, zlye1owy; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 574.1,
   "provenance": "community",
   "quant": "oQ3e-2b27-CED",
   "quote": "keys-Mac-oMLX-0.7.0.dev2-DeepSeek-V4.1-Flash-oQ3e-2b27-CED-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 unknown; 128k 574.1 34.7 198.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yo8tfj6e",
   "source_url": "https://omlx.ai/benchmarks/performance/yo8tfj6e",
   "tg_tokens": null,
   "tps": 34.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-yo8tfj6e-ctx199680",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 198.3 GB at this context; duplicate traces of the same session: 226xpsk1, 5tlnkumd, 7h9itid1, b4wxtdmf, py4394db, wjm4u43a, zlye1owy; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 483.5,
   "provenance": "community",
   "quant": "oQ3e-2b27-CED",
   "quote": "keys-Mac-oMLX-0.7.0.dev2-DeepSeek-V4.1-Flash-oQ3e-2b27-CED-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 unknown; 195k 483.5 44.7 198.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yo8tfj6e",
   "source_url": "https://omlx.ai/benchmarks/performance/yo8tfj6e",
   "tg_tokens": null,
   "tps": 44.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-nzj8qp83-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "mlx-community; oMLX community benchmark trace; peak memory 35.0 GB at this context; duplicate traces of the same session: zgy9i3z4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 592.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mlx-community--gemma-4-31b-it-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 592.3 32.3 35.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nzj8qp83",
   "source_url": "https://omlx.ai/benchmarks/performance/nzj8qp83",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-nzj8qp83-ctx16384",
   "model": "gemma-4-31B-it",
   "notes": "mlx-community; oMLX community benchmark trace; peak memory 36.7 GB at this context; duplicate traces of the same session: zgy9i3z4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 476.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mlx-community--gemma-4-31b-it-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 476.6 29.7 36.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nzj8qp83",
   "source_url": "https://omlx.ai/benchmarks/performance/nzj8qp83",
   "tg_tokens": null,
   "tps": 29.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-5k87x8ma-ctx1024",
   "model": "KAT-Coder-V2.5-Dev",
   "notes": "oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: e72t433l, mbx4st7b; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 712.8,
   "provenance": "community",
   "quant": "OptiQ 4bit",
   "quote": "KAT-Coder-V2.5-Dev-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 712.8 78.1 21.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5k87x8ma",
   "source_url": "https://omlx.ai/benchmarks/performance/5k87x8ma",
   "tg_tokens": null,
   "tps": 78.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-5k87x8ma-ctx8192",
   "model": "KAT-Coder-V2.5-Dev",
   "notes": "oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: e72t433l, mbx4st7b; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3234.0,
   "provenance": "community",
   "quant": "OptiQ 4bit",
   "quote": "KAT-Coder-V2.5-Dev-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 3,234 67.7 22.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5k87x8ma",
   "source_url": "https://omlx.ai/benchmarks/performance/5k87x8ma",
   "tg_tokens": null,
   "tps": 67.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-5k87x8ma-ctx32768",
   "model": "KAT-Coder-V2.5-Dev",
   "notes": "oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: e72t433l, mbx4st7b; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2489.0,
   "provenance": "community",
   "quant": "OptiQ 4bit",
   "quote": "KAT-Coder-V2.5-Dev-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 2,489 60.4 22.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5k87x8ma",
   "source_url": "https://omlx.ai/benchmarks/performance/5k87x8ma",
   "tg_tokens": null,
   "tps": 60.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-5k87x8ma-ctx131072",
   "model": "KAT-Coder-V2.5-Dev",
   "notes": "oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: e72t433l, mbx4st7b; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1270.0,
   "provenance": "community",
   "quant": "OptiQ 4bit",
   "quote": "KAT-Coder-V2.5-Dev-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 128k 1,270 39.0 24.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5k87x8ma",
   "source_url": "https://omlx.ai/benchmarks/performance/5k87x8ma",
   "tg_tokens": null,
   "tps": 39.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-546px4b8-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: ag9fhlwb; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 770.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 4k 770.0 55.1 20.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/546px4b8",
   "source_url": "https://omlx.ai/benchmarks/performance/546px4b8",
   "tg_tokens": null,
   "tps": 55.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-546px4b8-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: ag9fhlwb; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 679.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 16k 679.4 49.9 20.4 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/546px4b8",
   "source_url": "https://omlx.ai/benchmarks/performance/546px4b8",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-ycw11c1u-ctx1024",
   "model": "Qwen2.5-Coder-1.5B",
   "notes": "oMLX community benchmark trace; peak memory 2.2 GB at this context; duplicate traces of the same session: d9difsz0, xly7r621; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1446.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen2.5-Coder-1.5B-8bit on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 8bit; 1k 1,446 85.5 2.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ycw11c1u",
   "source_url": "https://omlx.ai/benchmarks/performance/ycw11c1u",
   "tg_tokens": null,
   "tps": 85.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-ycw11c1u-ctx4096",
   "model": "Qwen2.5-Coder-1.5B",
   "notes": "oMLX community benchmark trace; peak memory 2.3 GB at this context; duplicate traces of the same session: d9difsz0, xly7r621; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1412.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen2.5-Coder-1.5B-8bit on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 8bit; 4k 1,412 87.8 2.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ycw11c1u",
   "source_url": "https://omlx.ai/benchmarks/performance/ycw11c1u",
   "tg_tokens": null,
   "tps": 87.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-ycw11c1u-ctx8192",
   "model": "Qwen2.5-Coder-1.5B",
   "notes": "oMLX community benchmark trace; peak memory 2.4 GB at this context; duplicate traces of the same session: d9difsz0, xly7r621; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1295.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen2.5-Coder-1.5B-8bit on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 8bit; 8k 1,295 78.9 2.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ycw11c1u",
   "source_url": "https://omlx.ai/benchmarks/performance/ycw11c1u",
   "tg_tokens": null,
   "tps": 78.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-tbcdadnx-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 67.5 GB at this context; duplicate traces of the same session: vdw2jhim; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2171.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.6-35B-A3B-bf16 on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 bf16; 4k 2,171 62.3 67.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tbcdadnx",
   "source_url": "https://omlx.ai/benchmarks/performance/tbcdadnx",
   "tg_tokens": null,
   "tps": 62.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-tbcdadnx-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 67.9 GB at this context; duplicate traces of the same session: vdw2jhim; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1851.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.6-35B-A3B-bf16 on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 bf16; 16k 1,851 58.6 67.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tbcdadnx",
   "source_url": "https://omlx.ai/benchmarks/performance/tbcdadnx",
   "tg_tokens": null,
   "tps": 58.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-t19gnx3y-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "bf16-pt; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.9 GB at this context; duplicate traces of the same session: 316hz4p1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2112.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.6-35B-A3B-bf16-pt-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 1k 2,112 146.4 17.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t19gnx3y",
   "source_url": "https://omlx.ai/benchmarks/performance/t19gnx3y",
   "tg_tokens": null,
   "tps": 146.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-t19gnx3y-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "bf16-pt; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 316hz4p1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2618.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.6-35B-A3B-bf16-pt-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 4k 2,618 123.3 19.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t19gnx3y",
   "source_url": "https://omlx.ai/benchmarks/performance/t19gnx3y",
   "tg_tokens": null,
   "tps": 123.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-zh48d480-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "bf16-pt; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.9 GB at this context; duplicate traces of the same session: oyj1pb91; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2103.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.6-35B-A3B-bf16-pt-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 1k 2,103 145.3 17.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zh48d480",
   "source_url": "https://omlx.ai/benchmarks/performance/zh48d480",
   "tg_tokens": null,
   "tps": 145.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-zh48d480-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "bf16-pt; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: oyj1pb91; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2609.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.6-35B-A3B-bf16-pt-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 4k 2,609 126.9 19.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zh48d480",
   "source_url": "https://omlx.ai/benchmarks/performance/zh48d480",
   "tg_tokens": null,
   "tps": 126.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-uegfr1c7-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 14.2 GB; TTFT 3565 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 287.3,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.8-27B-3bit-mlx on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 3bit; 1k tokens 287.3 PP tok/s 20.6 TG tok/s 3565 TTFT (ms) 14.2 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uegfr1c7",
   "source_url": "https://omlx.ai/benchmarks/performance/uegfr1c7",
   "tg_tokens": null,
   "tps": 20.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-v9xmk9rh-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 14.2 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 285.8,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.8-27B-3bit-mlx on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 3bit; 1k 285.8 23.0 14.2 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v9xmk9rh",
   "source_url": "https://omlx.ai/benchmarks/performance/v9xmk9rh",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-v9xmk9rh-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 315.5,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.8-27B-3bit-mlx on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 3bit; 4k 315.5 16.8 15.8 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v9xmk9rh",
   "source_url": "https://omlx.ai/benchmarks/performance/v9xmk9rh",
   "tg_tokens": null,
   "tps": 16.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-caz9ssl5-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 14.2 GB; TTFT 3444 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 297.4,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.8-27B-3bit-mlx on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 3bit; 1k tokens 297.4 PP tok/s 21.9 TG tok/s 3444 TTFT (ms) 14.2 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/caz9ssl5",
   "source_url": "https://omlx.ai/benchmarks/performance/caz9ssl5",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-onhw5v27-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 14.0 GB; TTFT 3378 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 303.2,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.8-27B-3bit-mlx on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 3bit; 1k tokens 303.2 PP tok/s 22.5 TG tok/s 3378 TTFT (ms) 14.0 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/onhw5v27",
   "source_url": "https://omlx.ai/benchmarks/performance/onhw5v27",
   "tg_tokens": null,
   "tps": 22.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-4prujg38-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 15.8 GB; TTFT 15551 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 65.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 65.8 PP tok/s 6.9 TG tok/s 15551 TTFT (ms) 15.8 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4prujg38",
   "source_url": "https://omlx.ai/benchmarks/performance/4prujg38",
   "tg_tokens": null,
   "tps": 6.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-yn38djxc-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 15.8 GB; TTFT 15576 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 65.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 65.7 PP tok/s 6.8 TG tok/s 15576 TTFT (ms) 15.8 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yn38djxc",
   "source_url": "https://omlx.ai/benchmarks/performance/yn38djxc",
   "tg_tokens": null,
   "tps": 6.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-dnmc6p1l-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 446.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 446.0 17.3 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dnmc6p1l",
   "source_url": "https://omlx.ai/benchmarks/performance/dnmc6p1l",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-dnmc6p1l-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 475.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 475.9 17.1 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dnmc6p1l",
   "source_url": "https://omlx.ai/benchmarks/performance/dnmc6p1l",
   "tg_tokens": null,
   "tps": 17.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-npul37j3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 15.0 GB; TTFT 15539 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 65.9,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 65.9 PP tok/s 7.2 TG tok/s 15539 TTFT (ms) 15.0 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/npul37j3",
   "source_url": "https://omlx.ai/benchmarks/performance/npul37j3",
   "tg_tokens": null,
   "tps": 7.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-xl9jini8-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 15.0 GB; TTFT 15512 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 66.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 66.0 PP tok/s 7.2 TG tok/s 15512 TTFT (ms) 15.0 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xl9jini8",
   "source_url": "https://omlx.ai/benchmarks/performance/xl9jini8",
   "tg_tokens": null,
   "tps": 7.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-yzncch2u-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 15.0 GB; TTFT 15512 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 66.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 66.0 PP tok/s 7.3 TG tok/s 15512 TTFT (ms) 15.0 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yzncch2u",
   "source_url": "https://omlx.ai/benchmarks/performance/yzncch2u",
   "tg_tokens": null,
   "tps": 7.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-t76vvrve-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: iacbaksl; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1461.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 8k 1,461 88.6 19.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t76vvrve",
   "source_url": "https://omlx.ai/benchmarks/performance/t76vvrve",
   "tg_tokens": null,
   "tps": 88.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-t76vvrve-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: iacbaksl; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1318.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 32k 1,318 118.0 20.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t76vvrve",
   "source_url": "https://omlx.ai/benchmarks/performance/t76vvrve",
   "tg_tokens": null,
   "tps": 118.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-t76vvrve-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: iacbaksl; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1141.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 64k 1,141 92.0 23.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t76vvrve",
   "source_url": "https://omlx.ai/benchmarks/performance/t76vvrve",
   "tg_tokens": null,
   "tps": 92.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-t76vvrve-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: iacbaksl; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 898.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 128k 898.3 81.7 27.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t76vvrve",
   "source_url": "https://omlx.ai/benchmarks/performance/t76vvrve",
   "tg_tokens": null,
   "tps": 81.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qudiyh4r-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.0 GB at this context; duplicate traces of the same session: pe6ng8gm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 698.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 698.6 48.0 29.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qudiyh4r",
   "source_url": "https://omlx.ai/benchmarks/performance/qudiyh4r",
   "tg_tokens": null,
   "tps": 48.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qudiyh4r-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: pe6ng8gm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 729.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 729.1 46.8 30.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qudiyh4r",
   "source_url": "https://omlx.ai/benchmarks/performance/qudiyh4r",
   "tg_tokens": null,
   "tps": 46.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wcsh3deb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.0 GB at this context; duplicate traces of the same session: 1s2y900v; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 764.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 764.4 47.6 29.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wcsh3deb",
   "source_url": "https://omlx.ai/benchmarks/performance/wcsh3deb",
   "tg_tokens": null,
   "tps": 47.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wcsh3deb-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: 1s2y900v; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 734.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 734.7 61.5 30.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wcsh3deb",
   "source_url": "https://omlx.ai/benchmarks/performance/wcsh3deb",
   "tg_tokens": null,
   "tps": 61.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-aultis7q-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.0 GB at this context; duplicate traces of the same session: syn6dg3k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 818.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 818.2 46.0 29.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aultis7q",
   "source_url": "https://omlx.ai/benchmarks/performance/aultis7q",
   "tg_tokens": null,
   "tps": 46.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-aultis7q-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: syn6dg3k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 793.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 793.2 50.3 30.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aultis7q",
   "source_url": "https://omlx.ai/benchmarks/performance/aultis7q",
   "tg_tokens": null,
   "tps": 50.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n499fiy0-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 796pbzad; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 807.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 807.4 18.9 28.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n499fiy0",
   "source_url": "https://omlx.ai/benchmarks/performance/n499fiy0",
   "tg_tokens": null,
   "tps": 18.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n499fiy0-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.8 GB at this context; duplicate traces of the same session: 796pbzad; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 841.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 841.8 18.4 29.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n499fiy0",
   "source_url": "https://omlx.ai/benchmarks/performance/n499fiy0",
   "tg_tokens": null,
   "tps": 18.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vekkmj74-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: oof5dh1l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 833.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 833.3 19.0 28.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vekkmj74",
   "source_url": "https://omlx.ai/benchmarks/performance/vekkmj74",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vekkmj74-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.8 GB at this context; duplicate traces of the same session: oof5dh1l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 906.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 906.1 18.5 29.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vekkmj74",
   "source_url": "https://omlx.ai/benchmarks/performance/vekkmj74",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-78wx7dcd-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: dhbpdcuu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 838.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 838.6 18.9 28.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/78wx7dcd",
   "source_url": "https://omlx.ai/benchmarks/performance/78wx7dcd",
   "tg_tokens": null,
   "tps": 18.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-78wx7dcd-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.8 GB at this context; duplicate traces of the same session: dhbpdcuu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 851.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 851.0 18.5 29.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/78wx7dcd",
   "source_url": "https://omlx.ai/benchmarks/performance/78wx7dcd",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-d3ix3gsw-ctx1024",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: 5qkqda2v; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1880.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 1,880 78.7 29.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d3ix3gsw",
   "source_url": "https://omlx.ai/benchmarks/performance/d3ix3gsw",
   "tg_tokens": null,
   "tps": 78.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-d3ix3gsw-ctx4096",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.5 GB at this context; duplicate traces of the same session: 5qkqda2v; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2587.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 2,587 90.5 30.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d3ix3gsw",
   "source_url": "https://omlx.ai/benchmarks/performance/d3ix3gsw",
   "tg_tokens": null,
   "tps": 90.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-d3ix3gsw-ctx16384",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.8 GB at this context; duplicate traces of the same session: 5qkqda2v; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2366.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 16k 2,366 103.9 30.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d3ix3gsw",
   "source_url": "https://omlx.ai/benchmarks/performance/d3ix3gsw",
   "tg_tokens": null,
   "tps": 103.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-hddhq486-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: d118pf94, d13nj2ox, x2r3c26z; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 592.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 1k 592.2 55.4 100.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hddhq486",
   "source_url": "https://omlx.ai/benchmarks/performance/hddhq486",
   "tg_tokens": null,
   "tps": 55.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-hddhq486-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: d118pf94, d13nj2ox, x2r3c26z; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 775.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 4k 775.6 43.8 101.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hddhq486",
   "source_url": "https://omlx.ai/benchmarks/performance/hddhq486",
   "tg_tokens": null,
   "tps": 43.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-hddhq486-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: d118pf94, d13nj2ox, x2r3c26z; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 778.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 8k 778.2 38.4 101.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hddhq486",
   "source_url": "https://omlx.ai/benchmarks/performance/hddhq486",
   "tg_tokens": null,
   "tps": 38.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-hddhq486-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.3 GB at this context; duplicate traces of the same session: d118pf94, d13nj2ox, x2r3c26z; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 759.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 64k 759.9 35.7 104.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hddhq486",
   "source_url": "https://omlx.ai/benchmarks/performance/hddhq486",
   "tg_tokens": null,
   "tps": 35.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-ms1sbffb-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.7 GB at this context; duplicate traces of the same session: o1uijmmc; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 571.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 4k 571.9 40.9 14.7 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ms1sbffb",
   "source_url": "https://omlx.ai/benchmarks/performance/ms1sbffb",
   "tg_tokens": null,
   "tps": 40.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Pro",
   "id": "omlx-ms1sbffb-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 15.2 GB at this context; duplicate traces of the same session: o1uijmmc; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 534.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 16k 534.6 37.5 15.2 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ms1sbffb",
   "source_url": "https://omlx.ai/benchmarks/performance/ms1sbffb",
   "tg_tokens": null,
   "tps": 37.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-hps4igw9-ctx4096",
   "model": "Devstral-Small-2-24B-Instruct-2512",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: w5qh62i3; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 111.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Devstral-Small-2-24B-Instruct-2512-6bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 6bit; 4k 111.4 6.4 20.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hps4igw9",
   "source_url": "https://omlx.ai/benchmarks/performance/hps4igw9",
   "tg_tokens": null,
   "tps": 6.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-hps4igw9-ctx16384",
   "model": "Devstral-Small-2-24B-Instruct-2512",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: w5qh62i3; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.2,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Devstral-Small-2-24B-Instruct-2512-6bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 6bit; 16k 103.2 5.7 22.0 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hps4igw9",
   "source_url": "https://omlx.ai/benchmarks/performance/hps4igw9",
   "tg_tokens": null,
   "tps": 5.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-u9wejmz3-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.2 GB at this context; duplicate traces of the same session: 7ufe173w; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 749.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-6bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 6bit; 4k 749.8 42.3 28.2 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u9wejmz3",
   "source_url": "https://omlx.ai/benchmarks/performance/u9wejmz3",
   "tg_tokens": null,
   "tps": 42.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-u9wejmz3-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 7ufe173w; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 662.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-6bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 6bit; 16k 662.1 38.1 28.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u9wejmz3",
   "source_url": "https://omlx.ai/benchmarks/performance/u9wejmz3",
   "tg_tokens": null,
   "tps": 38.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-4awab1ft-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 39.8 GB; TTFT 6138 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1335.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 8bit; 8k tokens 1,335 PP tok/s 62.9 TG tok/s 6138 TTFT (ms) 39.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4awab1ft",
   "source_url": "https://omlx.ai/benchmarks/performance/4awab1ft",
   "tg_tokens": null,
   "tps": 62.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-imydcqzm-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 39.8 GB; TTFT 6356 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1289.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 8bit; 8k tokens 1,289 PP tok/s 31.4 TG tok/s 6356 TTFT (ms) 39.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/imydcqzm",
   "source_url": "https://omlx.ai/benchmarks/performance/imydcqzm",
   "tg_tokens": null,
   "tps": 31.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-b2tg8sr3-ctx1024",
   "model": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 9.0 GB at this context; duplicate traces of the same session: 4e9hqj6d, o07jgfgr, t7rogtjt; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1403.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 1,403 52.4 9.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b2tg8sr3",
   "source_url": "https://omlx.ai/benchmarks/performance/b2tg8sr3",
   "tg_tokens": null,
   "tps": 52.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-b2tg8sr3-ctx4096",
   "model": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 9.9 GB at this context; duplicate traces of the same session: 4e9hqj6d, o07jgfgr, t7rogtjt; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1593.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 1,593 48.2 9.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b2tg8sr3",
   "source_url": "https://omlx.ai/benchmarks/performance/b2tg8sr3",
   "tg_tokens": null,
   "tps": 48.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-b2tg8sr3-ctx16384",
   "model": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 10.7 GB at this context; duplicate traces of the same session: 4e9hqj6d, o07jgfgr, t7rogtjt; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1456.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 16k 1,456 32.5 10.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b2tg8sr3",
   "source_url": "https://omlx.ai/benchmarks/performance/b2tg8sr3",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-b2tg8sr3-ctx65536",
   "model": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 14.7 GB at this context; duplicate traces of the same session: 4e9hqj6d, o07jgfgr, t7rogtjt; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1204.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 64k 1,204 6.5 14.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b2tg8sr3",
   "source_url": "https://omlx.ai/benchmarks/performance/b2tg8sr3",
   "tg_tokens": null,
   "tps": 6.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-usipttnc-ctx1024",
   "model": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 8.6 GB at this context; duplicate traces of the same session: 2hgf4bok, co1lj7mp, gze6pkoh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1571.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 1,571 49.9 8.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/usipttnc",
   "source_url": "https://omlx.ai/benchmarks/performance/usipttnc",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-usipttnc-ctx4096",
   "model": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 9.3 GB at this context; duplicate traces of the same session: 2hgf4bok, co1lj7mp, gze6pkoh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1617.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 1,617 51.7 9.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/usipttnc",
   "source_url": "https://omlx.ai/benchmarks/performance/usipttnc",
   "tg_tokens": null,
   "tps": 51.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-usipttnc-ctx16384",
   "model": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 10.1 GB at this context; duplicate traces of the same session: 2hgf4bok, co1lj7mp, gze6pkoh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1479.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 16k 1,479 31.5 10.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/usipttnc",
   "source_url": "https://omlx.ai/benchmarks/performance/usipttnc",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-usipttnc-ctx65536",
   "model": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 13.0 GB at this context; duplicate traces of the same session: 2hgf4bok, co1lj7mp, gze6pkoh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1220.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 64k 1,220 5.3 13.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/usipttnc",
   "source_url": "https://omlx.ai/benchmarks/performance/usipttnc",
   "tg_tokens": null,
   "tps": 5.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-gwpa8la2-ctx1024",
   "model": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: 2yl0jps0, ecj6mqld, f1x2xwli; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1321.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 1,321 46.1 11.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gwpa8la2",
   "source_url": "https://omlx.ai/benchmarks/performance/gwpa8la2",
   "tg_tokens": null,
   "tps": 46.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-gwpa8la2-ctx4096",
   "model": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: 2yl0jps0, ecj6mqld, f1x2xwli; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1710.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 1,710 66.8 11.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gwpa8la2",
   "source_url": "https://omlx.ai/benchmarks/performance/gwpa8la2",
   "tg_tokens": null,
   "tps": 66.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-gwpa8la2-ctx16384",
   "model": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 12.3 GB at this context; duplicate traces of the same session: 2yl0jps0, ecj6mqld, f1x2xwli; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1693.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 16k 1,693 74.8 12.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gwpa8la2",
   "source_url": "https://omlx.ai/benchmarks/performance/gwpa8la2",
   "tg_tokens": null,
   "tps": 74.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-gwpa8la2-ctx65536",
   "model": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 14.1 GB at this context; duplicate traces of the same session: 2yl0jps0, ecj6mqld, f1x2xwli; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1342.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 64k 1,342 53.4 14.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gwpa8la2",
   "source_url": "https://omlx.ai/benchmarks/performance/gwpa8la2",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-eo4cfc8d-ctx1024",
   "model": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 9.8 GB at this context; duplicate traces of the same session: 090yd2kt, 7i7c2mu8, uffmmpms; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1433.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 1,433 39.1 9.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eo4cfc8d",
   "source_url": "https://omlx.ai/benchmarks/performance/eo4cfc8d",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-eo4cfc8d-ctx4096",
   "model": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 10.1 GB at this context; duplicate traces of the same session: 090yd2kt, 7i7c2mu8, uffmmpms; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1846.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 1,846 38.9 10.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eo4cfc8d",
   "source_url": "https://omlx.ai/benchmarks/performance/eo4cfc8d",
   "tg_tokens": null,
   "tps": 38.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-eo4cfc8d-ctx16384",
   "model": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 10.5 GB at this context; duplicate traces of the same session: 090yd2kt, 7i7c2mu8, uffmmpms; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1802.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 16k 1,802 36.4 10.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eo4cfc8d",
   "source_url": "https://omlx.ai/benchmarks/performance/eo4cfc8d",
   "tg_tokens": null,
   "tps": 36.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-eo4cfc8d-ctx65536",
   "model": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 11.9 GB at this context; duplicate traces of the same session: 090yd2kt, 7i7c2mu8, uffmmpms; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1320.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 64k 1,320 28.1 11.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eo4cfc8d",
   "source_url": "https://omlx.ai/benchmarks/performance/eo4cfc8d",
   "tg_tokens": null,
   "tps": 28.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-b4ap8sg1-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 13.2 GB; TTFT 3339 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 306.7,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.8-27B-3bit-mlx on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 3bit; 1k tokens 306.7 PP tok/s 11.5 TG tok/s 3339 TTFT (ms) 13.2 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b4ap8sg1",
   "source_url": "https://omlx.ai/benchmarks/performance/b4ap8sg1",
   "tg_tokens": null,
   "tps": 11.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-h2tv53wr-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 15.5 GB; TTFT 13123 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 312.1,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.8-27B-3bit-mlx on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 3bit; 4k tokens 312.1 PP tok/s 19.6 TG tok/s 13123 TTFT (ms) 15.5 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h2tv53wr",
   "source_url": "https://omlx.ai/benchmarks/performance/h2tv53wr",
   "tg_tokens": null,
   "tps": 19.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-0ufnia8u-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 20.0 GB; TTFT 90695 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 361.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k tokens 361.3 PP tok/s 28.3 TG tok/s 90695 TTFT (ms) 20.0 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ufnia8u",
   "source_url": "https://omlx.ai/benchmarks/performance/0ufnia8u",
   "tg_tokens": null,
   "tps": 28.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-5lmjulop-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 20.2 GB; TTFT 92921 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 352.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k tokens 352.6 PP tok/s 7.2 TG tok/s 92921 TTFT (ms) 20.2 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5lmjulop",
   "source_url": "https://omlx.ai/benchmarks/performance/5lmjulop",
   "tg_tokens": null,
   "tps": 7.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-jnskehcf-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 16.1 GB; TTFT 15405 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 66.5,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 66.5 PP tok/s 5.0 TG tok/s 15405 TTFT (ms) 16.1 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jnskehcf",
   "source_url": "https://omlx.ai/benchmarks/performance/jnskehcf",
   "tg_tokens": null,
   "tps": 5.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-z06ig5cx-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: nzezn82w; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1461.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 8k 1,461 88.6 19.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z06ig5cx",
   "source_url": "https://omlx.ai/benchmarks/performance/z06ig5cx",
   "tg_tokens": null,
   "tps": 88.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-z06ig5cx-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: nzezn82w; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1318.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 32k 1,318 118.0 20.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z06ig5cx",
   "source_url": "https://omlx.ai/benchmarks/performance/z06ig5cx",
   "tg_tokens": null,
   "tps": 118.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-z06ig5cx-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: nzezn82w; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1141.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 64k 1,141 92.0 23.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z06ig5cx",
   "source_url": "https://omlx.ai/benchmarks/performance/z06ig5cx",
   "tg_tokens": null,
   "tps": 92.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-z06ig5cx-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: nzezn82w; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 898.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 128k 898.3 81.7 27.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z06ig5cx",
   "source_url": "https://omlx.ai/benchmarks/performance/z06ig5cx",
   "tg_tokens": null,
   "tps": 81.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qmybj9nu-ctx1024",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1880.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 1,880 78.7 29.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qmybj9nu",
   "source_url": "https://omlx.ai/benchmarks/performance/qmybj9nu",
   "tg_tokens": null,
   "tps": 78.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qmybj9nu-ctx4096",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.5 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2587.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 2,587 90.5 30.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qmybj9nu",
   "source_url": "https://omlx.ai/benchmarks/performance/qmybj9nu",
   "tg_tokens": null,
   "tps": 90.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qmybj9nu-ctx16384",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.8 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2366.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 16k 2,366 103.9 30.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qmybj9nu",
   "source_url": "https://omlx.ai/benchmarks/performance/qmybj9nu",
   "tg_tokens": null,
   "tps": 103.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-fczkqqi8-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-VL-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: 1l8w57io, 9v0juxeu, aklbfjgd, gid1u2hq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 243.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-VL-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 243.8 39.3 26.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fczkqqi8",
   "source_url": "https://omlx.ai/benchmarks/performance/fczkqqi8",
   "tg_tokens": null,
   "tps": 39.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-fczkqqi8-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-VL-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.1 GB at this context; duplicate traces of the same session: 1l8w57io, 9v0juxeu, aklbfjgd, gid1u2hq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 263.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-VL-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 263.4 37.4 30.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fczkqqi8",
   "source_url": "https://omlx.ai/benchmarks/performance/fczkqqi8",
   "tg_tokens": null,
   "tps": 37.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-fczkqqi8-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-VL-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.6 GB at this context; duplicate traces of the same session: 1l8w57io, 9v0juxeu, aklbfjgd, gid1u2hq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 260.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-VL-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 8k 260.1 39.8 30.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fczkqqi8",
   "source_url": "https://omlx.ai/benchmarks/performance/fczkqqi8",
   "tg_tokens": null,
   "tps": 39.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-fczkqqi8-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-VL-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.5 GB at this context; duplicate traces of the same session: 1l8w57io, 9v0juxeu, aklbfjgd, gid1u2hq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 250.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-VL-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k 250.5 37.2 31.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fczkqqi8",
   "source_url": "https://omlx.ai/benchmarks/performance/fczkqqi8",
   "tg_tokens": null,
   "tps": 37.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-fczkqqi8-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-VL-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.3 GB at this context; duplicate traces of the same session: 1l8w57io, 9v0juxeu, aklbfjgd, gid1u2hq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 232.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-VL-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 232.1 31.8 33.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fczkqqi8",
   "source_url": "https://omlx.ai/benchmarks/performance/fczkqqi8",
   "tg_tokens": null,
   "tps": 31.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mvyerbw1-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: jtwam7s9, uxf8rx5t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4341.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 4,341 144.9 19.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mvyerbw1",
   "source_url": "https://omlx.ai/benchmarks/performance/mvyerbw1",
   "tg_tokens": null,
   "tps": 144.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mvyerbw1-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: jtwam7s9, uxf8rx5t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4067.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 4,067 97.8 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mvyerbw1",
   "source_url": "https://omlx.ai/benchmarks/performance/mvyerbw1",
   "tg_tokens": null,
   "tps": 97.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mvyerbw1-ctx131072",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: jtwam7s9, uxf8rx5t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2764.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 2,764 79.3 22.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mvyerbw1",
   "source_url": "https://omlx.ai/benchmarks/performance/mvyerbw1",
   "tg_tokens": null,
   "tps": 79.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-tnrq2esk-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 19.8 GB; TTFT 40716 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 100.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k tokens 100.6 PP tok/s 18.3 TG tok/s 40716 TTFT (ms) 19.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tnrq2esk",
   "source_url": "https://omlx.ai/benchmarks/performance/tnrq2esk",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-cgz20qf6-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: z36lg8kn; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 151.6,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 151.6 21.5 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cgz20qf6",
   "source_url": "https://omlx.ai/benchmarks/performance/cgz20qf6",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-cgz20qf6-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: z36lg8kn; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 160.9,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 160.9 20.6 23.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cgz20qf6",
   "source_url": "https://omlx.ai/benchmarks/performance/cgz20qf6",
   "tg_tokens": null,
   "tps": 20.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vw7himiz-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: 0ou7fxe3, zyyulla8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 866.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 866.4 69.7 17.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vw7himiz",
   "source_url": "https://omlx.ai/benchmarks/performance/vw7himiz",
   "tg_tokens": null,
   "tps": 69.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vw7himiz-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: 0ou7fxe3, zyyulla8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 906.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 906.8 73.2 19.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vw7himiz",
   "source_url": "https://omlx.ai/benchmarks/performance/vw7himiz",
   "tg_tokens": null,
   "tps": 73.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-vw7himiz-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: 0ou7fxe3, zyyulla8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 771.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 771.2 135.9 22.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vw7himiz",
   "source_url": "https://omlx.ai/benchmarks/performance/vw7himiz",
   "tg_tokens": null,
   "tps": 135.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-heekvk85-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: coldg9dh, noa84mkw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 826.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 826.7 55.4 17.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/heekvk85",
   "source_url": "https://omlx.ai/benchmarks/performance/heekvk85",
   "tg_tokens": null,
   "tps": 55.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-heekvk85-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: coldg9dh, noa84mkw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 873.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 873.3 44.2 19.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/heekvk85",
   "source_url": "https://omlx.ai/benchmarks/performance/heekvk85",
   "tg_tokens": null,
   "tps": 44.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-heekvk85-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: coldg9dh, noa84mkw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 766.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 766.2 63.7 22.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/heekvk85",
   "source_url": "https://omlx.ai/benchmarks/performance/heekvk85",
   "tg_tokens": null,
   "tps": 63.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jww0zrdx-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: 2p1uvh64, c62cz6kw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 851.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 851.1 49.5 17.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jww0zrdx",
   "source_url": "https://omlx.ai/benchmarks/performance/jww0zrdx",
   "tg_tokens": null,
   "tps": 49.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jww0zrdx-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: 2p1uvh64, c62cz6kw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 897.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 897.3 69.1 19.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jww0zrdx",
   "source_url": "https://omlx.ai/benchmarks/performance/jww0zrdx",
   "tg_tokens": null,
   "tps": 69.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jww0zrdx-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: 2p1uvh64, c62cz6kw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 794.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 794.3 139.2 22.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jww0zrdx",
   "source_url": "https://omlx.ai/benchmarks/performance/jww0zrdx",
   "tg_tokens": null,
   "tps": 139.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-p88an2mm-ctx1024",
   "model": "Qwen3.8-35B-A3B-Distill-Heretic",
   "notes": "MTP speculative decoding enabled; fp16 build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: xk5zxz11; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 583.1,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-35B-A3B-Distill-Heretic-oQ4-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 1k 583.1 72.6 20.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p88an2mm",
   "source_url": "https://omlx.ai/benchmarks/performance/p88an2mm",
   "tg_tokens": null,
   "tps": 72.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-p88an2mm-ctx4096",
   "model": "Qwen3.8-35B-A3B-Distill-Heretic",
   "notes": "MTP speculative decoding enabled; fp16 build; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: xk5zxz11; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 798.2,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-35B-A3B-Distill-Heretic-oQ4-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k 798.2 67.5 20.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p88an2mm",
   "source_url": "https://omlx.ai/benchmarks/performance/p88an2mm",
   "tg_tokens": null,
   "tps": 67.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fpt9fep4-ctx4096",
   "model": "Qwen2.5-Coder-7B-Instruct",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 4.9 GB; TTFT 2247 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1823.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-7B-Instruct-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k tokens 1,823 PP tok/s 62.2 TG tok/s 2247 TTFT (ms) 4.9 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fpt9fep4",
   "source_url": "https://omlx.ai/benchmarks/performance/fpt9fep4",
   "tg_tokens": null,
   "tps": 62.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-zrq2lctu-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "bf16-pt; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.9 GB at this context; duplicate traces of the same session: koiontaa; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2110.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.6-35B-A3B-bf16-pt-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 1k 2,110 120.9 17.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zrq2lctu",
   "source_url": "https://omlx.ai/benchmarks/performance/zrq2lctu",
   "tg_tokens": null,
   "tps": 120.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-zrq2lctu-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "bf16-pt; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: koiontaa; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2592.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.6-35B-A3B-bf16-pt-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 4k 2,592 111.0 19.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zrq2lctu",
   "source_url": "https://omlx.ai/benchmarks/performance/zrq2lctu",
   "tg_tokens": null,
   "tps": 111.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-a7g1awzs-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 9z0jxqtn; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 104.0 23.1 21.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a7g1awzs",
   "source_url": "https://omlx.ai/benchmarks/performance/a7g1awzs",
   "tg_tokens": null,
   "tps": 23.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-a7g1awzs-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 9z0jxqtn; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 121.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 121.3 27.2 22.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a7g1awzs",
   "source_url": "https://omlx.ai/benchmarks/performance/a7g1awzs",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zrcuryum-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: 435glser, 8fh23qiq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 869.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 869.2 57.3 17.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zrcuryum",
   "source_url": "https://omlx.ai/benchmarks/performance/zrcuryum",
   "tg_tokens": null,
   "tps": 57.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-zrcuryum-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: 435glser, 8fh23qiq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 883.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 883.6 87.9 19.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zrcuryum",
   "source_url": "https://omlx.ai/benchmarks/performance/zrcuryum",
   "tg_tokens": null,
   "tps": 87.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-zrcuryum-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: 435glser, 8fh23qiq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 769.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 769.9 139.3 22.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zrcuryum",
   "source_url": "https://omlx.ai/benchmarks/performance/zrcuryum",
   "tg_tokens": null,
   "tps": 139.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2yr25kge-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: 0eo6fg5k, ruknwbw4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 853.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 853.7 57.1 17.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2yr25kge",
   "source_url": "https://omlx.ai/benchmarks/performance/2yr25kge",
   "tg_tokens": null,
   "tps": 57.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2yr25kge-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: 0eo6fg5k, ruknwbw4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 871.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 871.0 71.5 19.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2yr25kge",
   "source_url": "https://omlx.ai/benchmarks/performance/2yr25kge",
   "tg_tokens": null,
   "tps": 71.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2yr25kge-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: 0eo6fg5k, ruknwbw4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 757.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 757.2 138.5 22.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2yr25kge",
   "source_url": "https://omlx.ai/benchmarks/performance/2yr25kge",
   "tg_tokens": null,
   "tps": 138.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-6hmvsyi6-ctx1024",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: la7tkc4k; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 532.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-35B-A3B-Distill-mlx-4Bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 1k 532.7 72.2 19.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6hmvsyi6",
   "source_url": "https://omlx.ai/benchmarks/performance/6hmvsyi6",
   "tg_tokens": null,
   "tps": 72.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-6hmvsyi6-ctx4096",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: la7tkc4k; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 606.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-35B-A3B-Distill-mlx-4Bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k 606.6 69.4 20.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6hmvsyi6",
   "source_url": "https://omlx.ai/benchmarks/performance/6hmvsyi6",
   "tg_tokens": null,
   "tps": 69.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lt1kw8n4-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 5qxkvofi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 351.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 351.6 24.0 70.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lt1kw8n4",
   "source_url": "https://omlx.ai/benchmarks/performance/lt1kw8n4",
   "tg_tokens": null,
   "tps": 24.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lt1kw8n4-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: 5qxkvofi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1214.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,214 40.9 71.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lt1kw8n4",
   "source_url": "https://omlx.ai/benchmarks/performance/lt1kw8n4",
   "tg_tokens": null,
   "tps": 40.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-7d641uq3-ctx4096",
   "model": "Qwen3.8",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: yeiubdxx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 255.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "qwen38-oq4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 255.4 47.9 19.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7d641uq3",
   "source_url": "https://omlx.ai/benchmarks/performance/7d641uq3",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-7d641uq3-ctx32768",
   "model": "Qwen3.8",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: yeiubdxx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 230.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "qwen38-oq4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 230.0 34.6 21.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7d641uq3",
   "source_url": "https://omlx.ai/benchmarks/performance/7d641uq3",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-xj88hvgk-ctx1024",
   "model": "qwen35 27B",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: g2xbwkp0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 785.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 785.8 24.7 22.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xj88hvgk",
   "source_url": "https://omlx.ai/benchmarks/performance/xj88hvgk",
   "tg_tokens": null,
   "tps": 24.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-xj88hvgk-ctx4096",
   "model": "qwen35 27B",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: g2xbwkp0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 882.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-27B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 882.4 24.6 23.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xj88hvgk",
   "source_url": "https://omlx.ai/benchmarks/performance/xj88hvgk",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lxb99znm-ctx1024",
   "model": "qwen35 27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: so66j4nb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 784.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 784.7 19.1 28.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lxb99znm",
   "source_url": "https://omlx.ai/benchmarks/performance/lxb99znm",
   "tg_tokens": null,
   "tps": 19.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lxb99znm-ctx4096",
   "model": "qwen35 27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: so66j4nb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 937.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 937.0 18.8 29.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lxb99znm",
   "source_url": "https://omlx.ai/benchmarks/performance/lxb99znm",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-7ww57jeb-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "bf16-pt; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.9 GB at this context; duplicate traces of the same session: l90vgop0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2127.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.6-35B-A3B-bf16-pt-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 1k 2,127 140.1 17.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ww57jeb",
   "source_url": "https://omlx.ai/benchmarks/performance/7ww57jeb",
   "tg_tokens": null,
   "tps": 140.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-7ww57jeb-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "bf16-pt; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: l90vgop0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2610.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.6-35B-A3B-bf16-pt-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 4k 2,610 146.3 19.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ww57jeb",
   "source_url": "https://omlx.ai/benchmarks/performance/7ww57jeb",
   "tg_tokens": null,
   "tps": 146.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-l0zpx29r-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 464rdsms; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 751.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 751.4 92.9 21.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l0zpx29r",
   "source_url": "https://omlx.ai/benchmarks/performance/l0zpx29r",
   "tg_tokens": null,
   "tps": 92.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-l0zpx29r-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 464rdsms; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 817.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 817.7 87.4 22.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l0zpx29r",
   "source_url": "https://omlx.ai/benchmarks/performance/l0zpx29r",
   "tg_tokens": null,
   "tps": 87.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-6lj7vz6l-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: khskb0v6; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 745.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 745.7 90.5 21.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6lj7vz6l",
   "source_url": "https://omlx.ai/benchmarks/performance/6lj7vz6l",
   "tg_tokens": null,
   "tps": 90.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-6lj7vz6l-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: khskb0v6; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 849.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 849.6 63.9 21.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6lj7vz6l",
   "source_url": "https://omlx.ai/benchmarks/performance/6lj7vz6l",
   "tg_tokens": null,
   "tps": 63.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-he4f3tkd-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: snup91zr; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 435.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 435.5 13.1 22.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/he4f3tkd",
   "source_url": "https://omlx.ai/benchmarks/performance/he4f3tkd",
   "tg_tokens": null,
   "tps": 13.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-he4f3tkd-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: snup91zr; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 479.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 479.9 12.9 23.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/he4f3tkd",
   "source_url": "https://omlx.ai/benchmarks/performance/he4f3tkd",
   "tg_tokens": null,
   "tps": 12.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-7l6mf8pz-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 149.9 GB at this context; duplicate traces of the same session: dfa1bu7c, e9y7z3fk, j29cd0hy, sn63smzs, yyyfz3pq; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1664.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ6-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 1k 1,664 67.2 149.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7l6mf8pz",
   "source_url": "https://omlx.ai/benchmarks/performance/7l6mf8pz",
   "tg_tokens": null,
   "tps": 67.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-7l6mf8pz-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 150.3 GB at this context; duplicate traces of the same session: dfa1bu7c, e9y7z3fk, j29cd0hy, sn63smzs, yyyfz3pq; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2227.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ6-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 4k 2,227 70.8 150.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7l6mf8pz",
   "source_url": "https://omlx.ai/benchmarks/performance/7l6mf8pz",
   "tg_tokens": null,
   "tps": 70.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-7l6mf8pz-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 151.4 GB at this context; duplicate traces of the same session: dfa1bu7c, e9y7z3fk, j29cd0hy, sn63smzs, yyyfz3pq; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3133.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ6-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 8k 3,133 87.3 151.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7l6mf8pz",
   "source_url": "https://omlx.ai/benchmarks/performance/7l6mf8pz",
   "tg_tokens": null,
   "tps": 87.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-7l6mf8pz-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 154.5 GB at this context; duplicate traces of the same session: dfa1bu7c, e9y7z3fk, j29cd0hy, sn63smzs, yyyfz3pq; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3127.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ6-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 64k 3,127 49.6 154.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7l6mf8pz",
   "source_url": "https://omlx.ai/benchmarks/performance/7l6mf8pz",
   "tg_tokens": null,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-7l6mf8pz-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 159.5 GB at this context; duplicate traces of the same session: dfa1bu7c, e9y7z3fk, j29cd0hy, sn63smzs, yyyfz3pq; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3020.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ6-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 128k 3,020 70.8 159.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7l6mf8pz",
   "source_url": "https://omlx.ai/benchmarks/performance/7l6mf8pz",
   "tg_tokens": null,
   "tps": 70.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-7l6mf8pz-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 162.5 GB at this context; duplicate traces of the same session: dfa1bu7c, e9y7z3fk, j29cd0hy, sn63smzs, yyyfz3pq; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2962.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ6-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 195k 2,962 60.2 162.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7l6mf8pz",
   "source_url": "https://omlx.ai/benchmarks/performance/7l6mf8pz",
   "tg_tokens": null,
   "tps": 60.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-c2ry0a29-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "BAND-CODE58 HC8 KV8 test build; oMLX community benchmark trace; peak memory 60.3 GB at this context; duplicate traces of the same session: 3t91q9ng, 6vy5cxmu, nxv83aqp, z7ivcd0t; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 641.0,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.8-Flash-Next-UNCENSORED-BAND-CODE58-HC8-KV8-Test on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 3bit; 4k 641.0 56.0 60.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c2ry0a29",
   "source_url": "https://omlx.ai/benchmarks/performance/c2ry0a29",
   "tg_tokens": null,
   "tps": 56.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-c2ry0a29-ctx32768",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "BAND-CODE58 HC8 KV8 test build; oMLX community benchmark trace; peak memory 61.0 GB at this context; duplicate traces of the same session: 3t91q9ng, 6vy5cxmu, nxv83aqp, z7ivcd0t; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 933.0,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.8-Flash-Next-UNCENSORED-BAND-CODE58-HC8-KV8-Test on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 3bit; 32k 933.0 43.1 61.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c2ry0a29",
   "source_url": "https://omlx.ai/benchmarks/performance/c2ry0a29",
   "tg_tokens": null,
   "tps": 43.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-c2ry0a29-ctx65536",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "BAND-CODE58 HC8 KV8 test build; oMLX community benchmark trace; peak memory 62.8 GB at this context; duplicate traces of the same session: 3t91q9ng, 6vy5cxmu, nxv83aqp, z7ivcd0t; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 956.0,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.8-Flash-Next-UNCENSORED-BAND-CODE58-HC8-KV8-Test on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 3bit; 64k 956.0 36.8 62.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c2ry0a29",
   "source_url": "https://omlx.ai/benchmarks/performance/c2ry0a29",
   "tg_tokens": null,
   "tps": 36.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-c2ry0a29-ctx131072",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "BAND-CODE58 HC8 KV8 test build; oMLX community benchmark trace; peak memory 66.1 GB at this context; duplicate traces of the same session: 3t91q9ng, 6vy5cxmu, nxv83aqp, z7ivcd0t; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 929.2,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.8-Flash-Next-UNCENSORED-BAND-CODE58-HC8-KV8-Test on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 3bit; 128k 929.2 26.4 66.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c2ry0a29",
   "source_url": "https://omlx.ai/benchmarks/performance/c2ry0a29",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-c2ry0a29-ctx199680",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "BAND-CODE58 HC8 KV8 test build; oMLX community benchmark trace; peak memory 69.1 GB at this context; duplicate traces of the same session: 3t91q9ng, 6vy5cxmu, nxv83aqp, z7ivcd0t; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 900.9,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.8-Flash-Next-UNCENSORED-BAND-CODE58-HC8-KV8-Test on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 3bit; 195k 900.9 18.1 69.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c2ry0a29",
   "source_url": "https://omlx.ai/benchmarks/performance/c2ry0a29",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-crjo2fyv-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: 43eil1yb, umo13q18; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1000.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,000 57.0 71.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/crjo2fyv",
   "source_url": "https://omlx.ai/benchmarks/performance/crjo2fyv",
   "tg_tokens": null,
   "tps": 57.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-crjo2fyv-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: 43eil1yb, umo13q18; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2015.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,015 76.5 71.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/crjo2fyv",
   "source_url": "https://omlx.ai/benchmarks/performance/crjo2fyv",
   "tg_tokens": null,
   "tps": 76.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-crjo2fyv-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.7 GB at this context; duplicate traces of the same session: 43eil1yb, umo13q18; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2033.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,033 67.6 74.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/crjo2fyv",
   "source_url": "https://omlx.ai/benchmarks/performance/crjo2fyv",
   "tg_tokens": null,
   "tps": 67.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xgdde3x8-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: qpy0l4nl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1066.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,066 71.1 70.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xgdde3x8",
   "source_url": "https://omlx.ai/benchmarks/performance/xgdde3x8",
   "tg_tokens": null,
   "tps": 71.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xgdde3x8-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.8 GB at this context; duplicate traces of the same session: qpy0l4nl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2037.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,037 52.3 73.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xgdde3x8",
   "source_url": "https://omlx.ai/benchmarks/performance/xgdde3x8",
   "tg_tokens": null,
   "tps": 52.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-atvsy4a2-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: f5m2nri0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1057.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,057 66.3 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/atvsy4a2",
   "source_url": "https://omlx.ai/benchmarks/performance/atvsy4a2",
   "tg_tokens": null,
   "tps": 66.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-atvsy4a2-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.3 GB at this context; duplicate traces of the same session: f5m2nri0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2010.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,010 56.0 74.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/atvsy4a2",
   "source_url": "https://omlx.ai/benchmarks/performance/atvsy4a2",
   "tg_tokens": null,
   "tps": 56.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-4iqlay0c-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.2 GB at this context; duplicate traces of the same session: 794bgiyj, oharx0hs; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2302.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,302 128.6 173.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4iqlay0c",
   "source_url": "https://omlx.ai/benchmarks/performance/4iqlay0c",
   "tg_tokens": null,
   "tps": 128.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-4iqlay0c-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.7 GB at this context; duplicate traces of the same session: 794bgiyj, oharx0hs; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2308.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,308 85.9 173.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4iqlay0c",
   "source_url": "https://omlx.ai/benchmarks/performance/4iqlay0c",
   "tg_tokens": null,
   "tps": 85.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-4iqlay0c-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.9 GB at this context; duplicate traces of the same session: 794bgiyj, oharx0hs; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2261.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,261 133.2 174.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4iqlay0c",
   "source_url": "https://omlx.ai/benchmarks/performance/4iqlay0c",
   "tg_tokens": null,
   "tps": 133.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-humfjgh7-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.1 GB at this context; duplicate traces of the same session: 14q0xajh, it88ku1w; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2284.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,284 119.3 178.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/humfjgh7",
   "source_url": "https://omlx.ai/benchmarks/performance/humfjgh7",
   "tg_tokens": null,
   "tps": 119.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-humfjgh7-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.6 GB at this context; duplicate traces of the same session: 14q0xajh, it88ku1w; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2192.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,192 71.6 178.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/humfjgh7",
   "source_url": "https://omlx.ai/benchmarks/performance/humfjgh7",
   "tg_tokens": null,
   "tps": 71.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-humfjgh7-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 179.9 GB at this context; duplicate traces of the same session: 14q0xajh, it88ku1w; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2207.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,207 125.9 179.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/humfjgh7",
   "source_url": "https://omlx.ai/benchmarks/performance/humfjgh7",
   "tg_tokens": null,
   "tps": 125.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-64jpkwgq-ctx1024",
   "model": "Qwen2.5-Coder-1.5B",
   "notes": "oMLX community benchmark trace; peak memory 2.2 GB at this context; duplicate traces of the same session: 6jgdgchf, kfumnp04; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1549.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen2.5-Coder-1.5B-8bit on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 8bit; 1k 1,549 93.9 2.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/64jpkwgq",
   "source_url": "https://omlx.ai/benchmarks/performance/64jpkwgq",
   "tg_tokens": null,
   "tps": 93.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-64jpkwgq-ctx4096",
   "model": "Qwen2.5-Coder-1.5B",
   "notes": "oMLX community benchmark trace; peak memory 2.3 GB at this context; duplicate traces of the same session: 6jgdgchf, kfumnp04; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1467.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen2.5-Coder-1.5B-8bit on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 8bit; 4k 1,467 88.1 2.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/64jpkwgq",
   "source_url": "https://omlx.ai/benchmarks/performance/64jpkwgq",
   "tg_tokens": null,
   "tps": 88.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-64jpkwgq-ctx8192",
   "model": "Qwen2.5-Coder-1.5B",
   "notes": "oMLX community benchmark trace; peak memory 2.4 GB at this context; duplicate traces of the same session: 6jgdgchf, kfumnp04; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1278.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen2.5-Coder-1.5B-8bit on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 8bit; 8k 1,278 78.1 2.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/64jpkwgq",
   "source_url": "https://omlx.ai/benchmarks/performance/64jpkwgq",
   "tg_tokens": null,
   "tps": 78.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-5x2qgl7t-ctx1024",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "MLX; oMLX community benchmark trace; peak memory 30.8 GB at this context; duplicate traces of the same session: pfkkfyb5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1351.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 1,351 82.7 30.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5x2qgl7t",
   "source_url": "https://omlx.ai/benchmarks/performance/5x2qgl7t",
   "tg_tokens": null,
   "tps": 82.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-5x2qgl7t-ctx4096",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "MLX; oMLX community benchmark trace; peak memory 31.3 GB at this context; duplicate traces of the same session: pfkkfyb5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1586.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 1,586 74.5 31.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5x2qgl7t",
   "source_url": "https://omlx.ai/benchmarks/performance/5x2qgl7t",
   "tg_tokens": null,
   "tps": 74.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-o2p59kmr-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: o8ab6ma8; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 751.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 751.7 102.8 21.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o2p59kmr",
   "source_url": "https://omlx.ai/benchmarks/performance/o2p59kmr",
   "tg_tokens": null,
   "tps": 102.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-o2p59kmr-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: o8ab6ma8; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 817.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 817.0 87.0 22.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o2p59kmr",
   "source_url": "https://omlx.ai/benchmarks/performance/o2p59kmr",
   "tg_tokens": null,
   "tps": 87.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-hz2fyb28-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: h2oj8xfv; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 751.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 751.8 96.3 21.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hz2fyb28",
   "source_url": "https://omlx.ai/benchmarks/performance/hz2fyb28",
   "tg_tokens": null,
   "tps": 96.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-hz2fyb28-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: h2oj8xfv; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 864.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 864.8 87.2 21.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hz2fyb28",
   "source_url": "https://omlx.ai/benchmarks/performance/hz2fyb28",
   "tg_tokens": null,
   "tps": 87.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-80k5z4ym-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: hdmkhlgm; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 754.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 754.5 112.6 21.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/80k5z4ym",
   "source_url": "https://omlx.ai/benchmarks/performance/80k5z4ym",
   "tg_tokens": null,
   "tps": 112.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-80k5z4ym-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: hdmkhlgm; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 818.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 818.8 91.2 22.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/80k5z4ym",
   "source_url": "https://omlx.ai/benchmarks/performance/80k5z4ym",
   "tg_tokens": null,
   "tps": 91.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-jfb878o5-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 525avywp; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 83.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 1k 83.9 15.6 16.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jfb878o5",
   "source_url": "https://omlx.ai/benchmarks/performance/jfb878o5",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-jfb878o5-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: 525avywp; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 80.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k 80.4 15.1 17.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jfb878o5",
   "source_url": "https://omlx.ai/benchmarks/performance/jfb878o5",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-xkytrnd6-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 84.6 GB; TTFT 2359 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 434.1,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.8-Flash-Next-MLX-3bit-MTP on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 3bit; 1k tokens 434.1 PP tok/s 21.0 TG tok/s 2359 TTFT (ms) 84.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xkytrnd6",
   "source_url": "https://omlx.ai/benchmarks/performance/xkytrnd6",
   "tg_tokens": null,
   "tps": 21.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-w4xcofxv-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 149.9 GB at this context; duplicate traces of the same session: hxoycsja, jsny4uow, mo5m5nev, pfev094i, s8ttmas4, t7ols1st; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1652.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ6-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 1k 1,652 80.0 149.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w4xcofxv",
   "source_url": "https://omlx.ai/benchmarks/performance/w4xcofxv",
   "tg_tokens": null,
   "tps": 80.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-w4xcofxv-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 150.3 GB at this context; duplicate traces of the same session: hxoycsja, jsny4uow, mo5m5nev, pfev094i, s8ttmas4, t7ols1st; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2753.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ6-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 4k 2,753 64.1 150.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w4xcofxv",
   "source_url": "https://omlx.ai/benchmarks/performance/w4xcofxv",
   "tg_tokens": null,
   "tps": 64.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-w4xcofxv-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 151.7 GB at this context; duplicate traces of the same session: hxoycsja, jsny4uow, mo5m5nev, pfev094i, s8ttmas4, t7ols1st; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3074.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ6-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 8k 3,074 78.7 151.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w4xcofxv",
   "source_url": "https://omlx.ai/benchmarks/performance/w4xcofxv",
   "tg_tokens": null,
   "tps": 78.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-w4xcofxv-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 152.0 GB at this context; duplicate traces of the same session: hxoycsja, jsny4uow, mo5m5nev, pfev094i, s8ttmas4, t7ols1st; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3089.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ6-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 16k 3,089 64.1 152.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w4xcofxv",
   "source_url": "https://omlx.ai/benchmarks/performance/w4xcofxv",
   "tg_tokens": null,
   "tps": 64.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-w4xcofxv-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 154.5 GB at this context; duplicate traces of the same session: hxoycsja, jsny4uow, mo5m5nev, pfev094i, s8ttmas4, t7ols1st; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3119.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ6-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 64k 3,119 75.1 154.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w4xcofxv",
   "source_url": "https://omlx.ai/benchmarks/performance/w4xcofxv",
   "tg_tokens": null,
   "tps": 75.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-w4xcofxv-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 159.5 GB at this context; duplicate traces of the same session: hxoycsja, jsny4uow, mo5m5nev, pfev094i, s8ttmas4, t7ols1st; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3050.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ6-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 128k 3,050 72.0 159.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w4xcofxv",
   "source_url": "https://omlx.ai/benchmarks/performance/w4xcofxv",
   "tg_tokens": null,
   "tps": 72.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-w4xcofxv-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 162.5 GB at this context; duplicate traces of the same session: hxoycsja, jsny4uow, mo5m5nev, pfev094i, s8ttmas4, t7ols1st; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2966.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ6-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 195k 2,966 58.1 162.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w4xcofxv",
   "source_url": "https://omlx.ai/benchmarks/performance/w4xcofxv",
   "tg_tokens": null,
   "tps": 58.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-laddhe5y-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "XL; oMLX community benchmark trace, single run at 1k context; peak memory 23.0 GB; TTFT 486 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2107.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "gemma-4-26B-A4B-it-6bit-XL-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 1k tokens 2,107 PP tok/s 61.4 TG tok/s 486 TTFT (ms) 23.0 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/laddhe5y",
   "source_url": "https://omlx.ai/benchmarks/performance/laddhe5y",
   "tg_tokens": null,
   "tps": 61.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-dveld01c-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "XL; oMLX community benchmark trace, single run at 1k context; peak memory 21.0 GB; TTFT 500 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2049.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "gemma-4-26B-A4B-it-6bit-XL-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 1k tokens 2,049 PP tok/s 54.8 TG tok/s 500 TTFT (ms) 21.0 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dveld01c",
   "source_url": "https://omlx.ai/benchmarks/performance/dveld01c",
   "tg_tokens": null,
   "tps": 54.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-zpnao6jc-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "XL; oMLX community benchmark trace, single run at 1k context; peak memory 21.0 GB; TTFT 499 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2053.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "gemma-4-26B-A4B-it-6bit-XL-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 1k tokens 2,053 PP tok/s 54.0 TG tok/s 499 TTFT (ms) 21.0 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zpnao6jc",
   "source_url": "https://omlx.ai/benchmarks/performance/zpnao6jc",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-lm9kunz9-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "XL; oMLX community benchmark trace, single run at 1k context; peak memory 21.0 GB; TTFT 503 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2036.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "gemma-4-26B-A4B-it-6bit-XL-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 1k tokens 2,036 PP tok/s 54.8 TG tok/s 503 TTFT (ms) 21.0 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lm9kunz9",
   "source_url": "https://omlx.ai/benchmarks/performance/lm9kunz9",
   "tg_tokens": null,
   "tps": 54.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-vz76wfgt-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "XL; oMLX community benchmark trace, single run at 1k context; peak memory 21.0 GB; TTFT 514 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1991.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "gemma-4-26B-A4B-it-6bit-XL-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 1k tokens 1,991 PP tok/s 54.8 TG tok/s 514 TTFT (ms) 21.0 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vz76wfgt",
   "source_url": "https://omlx.ai/benchmarks/performance/vz76wfgt",
   "tg_tokens": null,
   "tps": 54.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-zvyel64g-ctx131072",
   "model": "GLM-OCR",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 10.6 GB; TTFT 104914 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1249.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "GLM-OCR-bf16 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 bf16; 128k tokens 1,249 PP tok/s 49.0 TG tok/s 104914 TTFT (ms) 10.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zvyel64g",
   "source_url": "https://omlx.ai/benchmarks/performance/zvyel64g",
   "tg_tokens": null,
   "tps": 49.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-414ur0yy-ctx1024",
   "model": "Hemmingway-1",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.7 GB at this context; duplicate traces of the same session: w9yjispu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 460.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Hemmingway-1-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 460.3 16.6 16.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/414ur0yy",
   "source_url": "https://omlx.ai/benchmarks/performance/414ur0yy",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-414ur0yy-ctx4096",
   "model": "Hemmingway-1",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: w9yjispu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 493.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Hemmingway-1-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 493.9 16.4 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/414ur0yy",
   "source_url": "https://omlx.ai/benchmarks/performance/414ur0yy",
   "tg_tokens": null,
   "tps": 16.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-uhg8j7vk-ctx131072",
   "model": "Ornith-1.0-35B",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 25.5 GB; TTFT 189754 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 690.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.0-35B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k tokens 690.7 PP tok/s 60.3 TG tok/s 189754 TTFT (ms) 25.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uhg8j7vk",
   "source_url": "https://omlx.ai/benchmarks/performance/uhg8j7vk",
   "tg_tokens": null,
   "tps": 60.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ckm2x8oc-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: t0umh1r4; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 690.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 690.1 65.4 21.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ckm2x8oc",
   "source_url": "https://omlx.ai/benchmarks/performance/ckm2x8oc",
   "tg_tokens": null,
   "tps": 65.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ckm2x8oc-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: t0umh1r4; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 463.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 32k 463.0 57.3 23.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ckm2x8oc",
   "source_url": "https://omlx.ai/benchmarks/performance/ckm2x8oc",
   "tg_tokens": null,
   "tps": 57.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-v2oosrr4-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 8o4plgck; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 680.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 680.7 75.3 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v2oosrr4",
   "source_url": "https://omlx.ai/benchmarks/performance/v2oosrr4",
   "tg_tokens": null,
   "tps": 75.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-v2oosrr4-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 8o4plgck; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 727.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 727.9 73.4 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v2oosrr4",
   "source_url": "https://omlx.ai/benchmarks/performance/v2oosrr4",
   "tg_tokens": null,
   "tps": 73.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-n6rxbe4c-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: v2bjwbyz; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 759.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 759.9 77.9 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n6rxbe4c",
   "source_url": "https://omlx.ai/benchmarks/performance/n6rxbe4c",
   "tg_tokens": null,
   "tps": 77.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-n6rxbe4c-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: v2bjwbyz; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 831.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 831.1 74.7 20.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n6rxbe4c",
   "source_url": "https://omlx.ai/benchmarks/performance/n6rxbe4c",
   "tg_tokens": null,
   "tps": 74.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-bdufg1te-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: k7d516d2; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 170.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 1k 170.7 17.8 16.0 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bdufg1te",
   "source_url": "https://omlx.ai/benchmarks/performance/bdufg1te",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-bdufg1te-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: k7d516d2; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 175.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 4k 175.8 17.4 17.0 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bdufg1te",
   "source_url": "https://omlx.ai/benchmarks/performance/bdufg1te",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q6n1r8mh-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: xwvh9j1g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 811.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 811.3 18.5 28.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q6n1r8mh",
   "source_url": "https://omlx.ai/benchmarks/performance/q6n1r8mh",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q6n1r8mh-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: xwvh9j1g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 699.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 699.8 18.4 29.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q6n1r8mh",
   "source_url": "https://omlx.ai/benchmarks/performance/q6n1r8mh",
   "tg_tokens": null,
   "tps": 18.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8c93ikph-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 30.9 GB at this context; duplicate traces of the same session: o5r26p4n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 900.6,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Qwen3.8-27B-mxfp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 900.6 14.2 30.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8c93ikph",
   "source_url": "https://omlx.ai/benchmarks/performance/8c93ikph",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8c93ikph-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.0 GB at this context; duplicate traces of the same session: o5r26p4n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 844.2,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Qwen3.8-27B-mxfp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 844.2 14.7 31.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8c93ikph",
   "source_url": "https://omlx.ai/benchmarks/performance/8c93ikph",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-as73iblk-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: nvh50gbb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 130.9 14.9 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/as73iblk",
   "source_url": "https://omlx.ai/benchmarks/performance/as73iblk",
   "tg_tokens": null,
   "tps": 14.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-as73iblk-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: nvh50gbb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 130.1 14.5 20.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/as73iblk",
   "source_url": "https://omlx.ai/benchmarks/performance/as73iblk",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-ss72r6i2-ctx1024",
   "model": "Qwen3.8-35B-A3B-Distill-Heretic",
   "notes": "MTP speculative decoding enabled; fp16 build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: bvr9neib; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 687.9,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-35B-A3B-Distill-Heretic-oQ4-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 1k 687.9 62.3 22.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ss72r6i2",
   "source_url": "https://omlx.ai/benchmarks/performance/ss72r6i2",
   "tg_tokens": null,
   "tps": 62.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ss72r6i2-ctx4096",
   "model": "Qwen3.8-35B-A3B-Distill-Heretic",
   "notes": "MTP speculative decoding enabled; fp16 build; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: bvr9neib; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 792.3,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-35B-A3B-Distill-Heretic-oQ4-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k 792.3 68.4 23.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ss72r6i2",
   "source_url": "https://omlx.ai/benchmarks/performance/ss72r6i2",
   "tg_tokens": null,
   "tps": 68.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-bhybo0uc-ctx1024",
   "model": "Qwen3.8-35B-A3B-Distill-Heretic",
   "notes": "MTP speculative decoding enabled; fp16 build; oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: nrjb7p7c; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 769.8,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-35B-A3B-Distill-Heretic-oQ4-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 1k 769.8 73.2 20.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhybo0uc",
   "source_url": "https://omlx.ai/benchmarks/performance/bhybo0uc",
   "tg_tokens": null,
   "tps": 73.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-bhybo0uc-ctx4096",
   "model": "Qwen3.8-35B-A3B-Distill-Heretic",
   "notes": "MTP speculative decoding enabled; fp16 build; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: nrjb7p7c; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 837.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-35B-A3B-Distill-Heretic-oQ4-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k 837.0 70.2 20.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhybo0uc",
   "source_url": "https://omlx.ai/benchmarks/performance/bhybo0uc",
   "tg_tokens": null,
   "tps": 70.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-4eowjsg6-ctx1024",
   "model": "Qwen3.8-Flash-Next-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 78.2 GB at this context; duplicate traces of the same session: attttvoq, g3d9hj66; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 621.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-Abliterated-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 621.6 31.0 78.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4eowjsg6",
   "source_url": "https://omlx.ai/benchmarks/performance/4eowjsg6",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4eowjsg6-ctx8192",
   "model": "Qwen3.8-Flash-Next-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 78.8 GB at this context; duplicate traces of the same session: attttvoq, g3d9hj66; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1006.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-Abliterated-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,006 52.3 78.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4eowjsg6",
   "source_url": "https://omlx.ai/benchmarks/performance/4eowjsg6",
   "tg_tokens": null,
   "tps": 52.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4eowjsg6-ctx65536",
   "model": "Qwen3.8-Flash-Next-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 81.9 GB at this context; duplicate traces of the same session: attttvoq, g3d9hj66; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 989.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-Abliterated-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 989.6 60.1 81.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4eowjsg6",
   "source_url": "https://omlx.ai/benchmarks/performance/4eowjsg6",
   "tg_tokens": null,
   "tps": 60.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-4gheivhw-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 86.9 GB; TTFT 2073 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 493.9,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.8-Flash-Next-MLX-3bit-MTP on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 3bit; 1k tokens 493.9 PP tok/s 24.2 TG tok/s 2073 TTFT (ms) 86.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4gheivhw",
   "source_url": "https://omlx.ai/benchmarks/performance/4gheivhw",
   "tg_tokens": null,
   "tps": 24.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-sc5x7u3a-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 86.6 GB; TTFT 2080 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 492.3,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.8-Flash-Next-MLX-3bit-MTP on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 3bit; 1k tokens 492.3 PP tok/s 21.9 TG tok/s 2080 TTFT (ms) 86.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sc5x7u3a",
   "source_url": "https://omlx.ai/benchmarks/performance/sc5x7u3a",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cy1up53d-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 39.8 GB at this context; duplicate traces of the same session: 1aekao1s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1018.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,018 66.7 39.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cy1up53d",
   "source_url": "https://omlx.ai/benchmarks/performance/cy1up53d",
   "tg_tokens": null,
   "tps": 66.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cy1up53d-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 40.0 GB at this context; duplicate traces of the same session: 1aekao1s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1299.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,299 64.0 40.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cy1up53d",
   "source_url": "https://omlx.ai/benchmarks/performance/cy1up53d",
   "tg_tokens": null,
   "tps": 64.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-x642xio4-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 35.3 GB; TTFT 34525 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 474.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k tokens 474.6 PP tok/s 8.2 TG tok/s 34525 TTFT (ms) 35.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x642xio4",
   "source_url": "https://omlx.ai/benchmarks/performance/x642xio4",
   "tg_tokens": null,
   "tps": 8.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-fp2kzmq3-ctx16384",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 28.6 GB; TTFT 252780 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 64.8,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Ternary-Bonsai-2-27B-MLX-oQ6-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 6bit; 16k tokens 64.8 PP tok/s 7.3 TG tok/s 252780 TTFT (ms) 28.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fp2kzmq3",
   "source_url": "https://omlx.ai/benchmarks/performance/fp2kzmq3",
   "tg_tokens": null,
   "tps": 7.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-7eg786yd-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "XL; oMLX community benchmark trace, single run at 1k context; peak memory 23.0 GB; TTFT 484 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2116.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "gemma-4-26B-A4B-it-6bit-XL-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 1k tokens 2,116 PP tok/s 60.6 TG tok/s 484 TTFT (ms) 23.0 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7eg786yd",
   "source_url": "https://omlx.ai/benchmarks/performance/7eg786yd",
   "tg_tokens": null,
   "tps": 60.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-bbwjy9k2-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "XL; oMLX community benchmark trace, single run at 1k context; peak memory 23.0 GB; TTFT 493 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2076.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "gemma-4-26B-A4B-it-6bit-XL-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 1k tokens 2,076 PP tok/s 49.0 TG tok/s 493 TTFT (ms) 23.0 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bbwjy9k2",
   "source_url": "https://omlx.ai/benchmarks/performance/bbwjy9k2",
   "tg_tokens": null,
   "tps": 49.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-d1q8x0eh-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "XL; oMLX community benchmark trace, single run at 1k context; peak memory 23.0 GB; TTFT 476 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2151.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "gemma-4-26B-A4B-it-6bit-XL-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 1k tokens 2,151 PP tok/s 55.2 TG tok/s 476 TTFT (ms) 23.0 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d1q8x0eh",
   "source_url": "https://omlx.ai/benchmarks/performance/d1q8x0eh",
   "tg_tokens": null,
   "tps": 55.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-08l5r8vo-ctx4096",
   "model": "qwen38 next",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 107.9 GB; TTFT 1407 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2911.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "qwen38_next on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k tokens 2,911 PP tok/s 98.8 TG tok/s 1407 TTFT (ms) 107.9 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/08l5r8vo",
   "source_url": "https://omlx.ai/benchmarks/performance/08l5r8vo",
   "tg_tokens": null,
   "tps": 98.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-72sgwh9e-ctx1024",
   "model": "qwen38 next",
   "notes": "oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: 87zq5v3j, kaxwns8b, qv8g3d88, sapiulwb, t0wox6io; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2060.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "qwen38_next_oq4e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 2,060 150.4 100.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/72sgwh9e",
   "source_url": "https://omlx.ai/benchmarks/performance/72sgwh9e",
   "tg_tokens": null,
   "tps": 150.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-72sgwh9e-ctx4096",
   "model": "qwen38 next",
   "notes": "oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: 87zq5v3j, kaxwns8b, qv8g3d88, sapiulwb, t0wox6io; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3495.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "qwen38_next_oq4e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 3,495 130.1 100.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/72sgwh9e",
   "source_url": "https://omlx.ai/benchmarks/performance/72sgwh9e",
   "tg_tokens": null,
   "tps": 130.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-72sgwh9e-ctx8192",
   "model": "qwen38 next",
   "notes": "oMLX community benchmark trace; peak memory 100.7 GB at this context; duplicate traces of the same session: 87zq5v3j, kaxwns8b, qv8g3d88, sapiulwb, t0wox6io; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3525.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "qwen38_next_oq4e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 3,525 127.0 100.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/72sgwh9e",
   "source_url": "https://omlx.ai/benchmarks/performance/72sgwh9e",
   "tg_tokens": null,
   "tps": 127.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-72sgwh9e-ctx65536",
   "model": "qwen38 next",
   "notes": "oMLX community benchmark trace; peak memory 104.3 GB at this context; duplicate traces of the same session: 87zq5v3j, kaxwns8b, qv8g3d88, sapiulwb, t0wox6io; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3396.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "qwen38_next_oq4e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 3,396 110.4 104.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/72sgwh9e",
   "source_url": "https://omlx.ai/benchmarks/performance/72sgwh9e",
   "tg_tokens": null,
   "tps": 110.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-72sgwh9e-ctx131072",
   "model": "qwen38 next",
   "notes": "oMLX community benchmark trace; peak memory 108.5 GB at this context; duplicate traces of the same session: 87zq5v3j, kaxwns8b, qv8g3d88, sapiulwb, t0wox6io; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3176.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "qwen38_next_oq4e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 3,176 97.3 108.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/72sgwh9e",
   "source_url": "https://omlx.ai/benchmarks/performance/72sgwh9e",
   "tg_tokens": null,
   "tps": 97.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-72sgwh9e-ctx199680",
   "model": "qwen38 next",
   "notes": "oMLX community benchmark trace; peak memory 110.8 GB at this context; duplicate traces of the same session: 87zq5v3j, kaxwns8b, qv8g3d88, sapiulwb, t0wox6io; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2950.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "qwen38_next_oq4e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 2,950 88.3 110.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/72sgwh9e",
   "source_url": "https://omlx.ai/benchmarks/performance/72sgwh9e",
   "tg_tokens": null,
   "tps": 88.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-71g890ou-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 44.0 GB at this context; duplicate traces of the same session: 9ntaj92t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1436.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,436 89.4 44.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/71g890ou",
   "source_url": "https://omlx.ai/benchmarks/performance/71g890ou",
   "tg_tokens": null,
   "tps": 89.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-71g890ou-ctx16384",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 44.2 GB at this context; duplicate traces of the same session: 9ntaj92t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1254.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,254 82.3 44.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/71g890ou",
   "source_url": "https://omlx.ai/benchmarks/performance/71g890ou",
   "tg_tokens": null,
   "tps": 82.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-zx48rt0g-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 22.4 GB; TTFT 28207 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 580.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k tokens 580.9 PP tok/s 53.2 TG tok/s 28207 TTFT (ms) 22.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zx48rt0g",
   "source_url": "https://omlx.ai/benchmarks/performance/zx48rt0g",
   "tg_tokens": null,
   "tps": 53.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cca4vqml-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: n54p404f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 924.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 924.8 33.6 16.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cca4vqml",
   "source_url": "https://omlx.ai/benchmarks/performance/cca4vqml",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cca4vqml-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: n54p404f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1025.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,025 32.9 17.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cca4vqml",
   "source_url": "https://omlx.ai/benchmarks/performance/cca4vqml",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-b2bw4bkw-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 2q8qb4nc, 9j4msfr4, b7yqv8r2; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 101.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 101.3 12.5 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b2bw4bkw",
   "source_url": "https://omlx.ai/benchmarks/performance/b2bw4bkw",
   "tg_tokens": null,
   "tps": 12.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-b2bw4bkw-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 2q8qb4nc, 9j4msfr4, b7yqv8r2; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 87.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 87.5 12.3 23.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b2bw4bkw",
   "source_url": "https://omlx.ai/benchmarks/performance/b2bw4bkw",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-b2bw4bkw-ctx8192",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: 2q8qb4nc, 9j4msfr4, b7yqv8r2; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 83.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 83.4 12.0 25.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b2bw4bkw",
   "source_url": "https://omlx.ai/benchmarks/performance/b2bw4bkw",
   "tg_tokens": null,
   "tps": 12.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-b2bw4bkw-ctx16384",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: 2q8qb4nc, 9j4msfr4, b7yqv8r2; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 80.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 80.7 11.5 26.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b2bw4bkw",
   "source_url": "https://omlx.ai/benchmarks/performance/b2bw4bkw",
   "tg_tokens": null,
   "tps": 11.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rwgxr44w-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: i1wcmzwr, y920btfv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 861.8,
   "provenance": "community",
   "quant": "OptiQ 4bit",
   "quote": "Qwen3.8-27B-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 861.8 27.4 20.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rwgxr44w",
   "source_url": "https://omlx.ai/benchmarks/performance/rwgxr44w",
   "tg_tokens": null,
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rwgxr44w-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: i1wcmzwr, y920btfv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 965.7,
   "provenance": "community",
   "quant": "OptiQ 4bit",
   "quote": "Qwen3.8-27B-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 965.7 27.3 21.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rwgxr44w",
   "source_url": "https://omlx.ai/benchmarks/performance/rwgxr44w",
   "tg_tokens": null,
   "tps": 27.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-rwgxr44w-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: i1wcmzwr, y920btfv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1009.0,
   "provenance": "community",
   "quant": "OptiQ 4bit",
   "quote": "Qwen3.8-27B-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,009 26.8 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rwgxr44w",
   "source_url": "https://omlx.ai/benchmarks/performance/rwgxr44w",
   "tg_tokens": null,
   "tps": 26.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-cwonjnsn-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 20.9 GB; TTFT 54197 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 604.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k tokens 604.6 PP tok/s 29.1 TG tok/s 54197 TTFT (ms) 20.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cwonjnsn",
   "source_url": "https://omlx.ai/benchmarks/performance/cwonjnsn",
   "tg_tokens": null,
   "tps": 29.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-bdpwx5ak-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 20.9 GB; TTFT 52895 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 619.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k tokens 619.5 PP tok/s 31.4 TG tok/s 52895 TTFT (ms) 20.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bdpwx5ak",
   "source_url": "https://omlx.ai/benchmarks/performance/bdpwx5ak",
   "tg_tokens": null,
   "tps": 31.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-4pelqwfb-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 20.9 GB; TTFT 58762 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 557.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k tokens 557.6 PP tok/s 42.7 TG tok/s 58762 TTFT (ms) 20.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4pelqwfb",
   "source_url": "https://omlx.ai/benchmarks/performance/4pelqwfb",
   "tg_tokens": null,
   "tps": 42.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-57cfsilo-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 29.6 GB; TTFT 89767 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 365.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 32k tokens 365.0 PP tok/s 24.1 TG tok/s 89767 TTFT (ms) 29.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/57cfsilo",
   "source_url": "https://omlx.ai/benchmarks/performance/57cfsilo",
   "tg_tokens": null,
   "tps": 24.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gmjk83za-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 122.4 GB at this context; duplicate traces of the same session: eprobf1j; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 849.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 1k 849.2 75.2 122.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gmjk83za",
   "source_url": "https://omlx.ai/benchmarks/performance/gmjk83za",
   "tg_tokens": null,
   "tps": 75.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gmjk83za-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 122.8 GB at this context; duplicate traces of the same session: eprobf1j; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1040.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 4k 1,040 82.6 122.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gmjk83za",
   "source_url": "https://omlx.ai/benchmarks/performance/gmjk83za",
   "tg_tokens": null,
   "tps": 82.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-9klyetyn-ctx32768",
   "model": "gemma-4-31B-it-scotoma-2",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 26.7 GB; TTFT 105885 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 309.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-scotoma-2-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k tokens 309.5 PP tok/s 8.8 TG tok/s 105885 TTFT (ms) 26.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9klyetyn",
   "source_url": "https://omlx.ai/benchmarks/performance/9klyetyn",
   "tg_tokens": null,
   "tps": 8.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-i909yakd-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "nightmedia repack; oMLX community benchmark trace; peak memory 77.9 GB at this context; duplicate traces of the same session: jnqv6nki; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2260.0,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "nightmedia--Qwen3-Coder-Next-mxfp8-mlx on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 2,260 78.2 77.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i909yakd",
   "source_url": "https://omlx.ai/benchmarks/performance/i909yakd",
   "tg_tokens": null,
   "tps": 78.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-i909yakd-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "nightmedia repack; oMLX community benchmark trace; peak memory 78.7 GB at this context; duplicate traces of the same session: jnqv6nki; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3278.0,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "nightmedia--Qwen3-Coder-Next-mxfp8-mlx on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 3,278 76.4 78.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i909yakd",
   "source_url": "https://omlx.ai/benchmarks/performance/i909yakd",
   "tg_tokens": null,
   "tps": 76.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-81dfaymq-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 323.6 GB at this context; duplicate traces of the same session: 0evirm48, zvkxxos9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 431.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 431.4 28.8 323.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/81dfaymq",
   "source_url": "https://omlx.ai/benchmarks/performance/81dfaymq",
   "tg_tokens": null,
   "tps": 28.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-81dfaymq-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 324.7 GB at this context; duplicate traces of the same session: 0evirm48, zvkxxos9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 554.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 554.0 32.8 324.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/81dfaymq",
   "source_url": "https://omlx.ai/benchmarks/performance/81dfaymq",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-81dfaymq-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 326.1 GB at this context; duplicate traces of the same session: 0evirm48, zvkxxos9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 517.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 64k 517.1 39.6 326.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/81dfaymq",
   "source_url": "https://omlx.ai/benchmarks/performance/81dfaymq",
   "tg_tokens": null,
   "tps": 39.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-puvzyw6b-ctx1024",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.8 GB at this context; duplicate traces of the same session: h00b7cv8; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 374.8,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-OptiQ-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 374.8 38.0 7.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/puvzyw6b",
   "source_url": "https://omlx.ai/benchmarks/performance/puvzyw6b",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-puvzyw6b-ctx4096",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.3 GB at this context; duplicate traces of the same session: h00b7cv8; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 380.8,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-OptiQ-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 380.8 37.0 8.3 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/puvzyw6b",
   "source_url": "https://omlx.ai/benchmarks/performance/puvzyw6b",
   "tg_tokens": null,
   "tps": 37.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-1h33ik9i-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 24.1 GB; TTFT 70608 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 464.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 32k tokens 464.1 PP tok/s 55.9 TG tok/s 70608 TTFT (ms) 24.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1h33ik9i",
   "source_url": "https://omlx.ai/benchmarks/performance/1h33ik9i",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-ojmpvdic-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: c04lbs1k; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 731.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 731.6 64.6 21.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ojmpvdic",
   "source_url": "https://omlx.ai/benchmarks/performance/ojmpvdic",
   "tg_tokens": null,
   "tps": 64.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-ojmpvdic-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: c04lbs1k; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 306.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 64k 306.9 35.4 24.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ojmpvdic",
   "source_url": "https://omlx.ai/benchmarks/performance/ojmpvdic",
   "tg_tokens": null,
   "tps": 35.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-tx0vkjyh-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: u33vnjjg; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1684.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 1,684 71.5 19.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tx0vkjyh",
   "source_url": "https://omlx.ai/benchmarks/performance/tx0vkjyh",
   "tg_tokens": null,
   "tps": 71.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-tx0vkjyh-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: u33vnjjg; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2040.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 2,040 69.6 19.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tx0vkjyh",
   "source_url": "https://omlx.ai/benchmarks/performance/tx0vkjyh",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3kpnyd46-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 22.6 GB; TTFT 13855 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2365.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k tokens 2,365 PP tok/s 67.6 TG tok/s 13855 TTFT (ms) 22.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3kpnyd46",
   "source_url": "https://omlx.ai/benchmarks/performance/3kpnyd46",
   "tg_tokens": null,
   "tps": 67.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-sxos55a1-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 1irn2nwz, 5vkfk87r, m18ww9hi; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 101.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 101.8 13.0 20.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sxos55a1",
   "source_url": "https://omlx.ai/benchmarks/performance/sxos55a1",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-sxos55a1-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 1irn2nwz, 5vkfk87r, m18ww9hi; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 86.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 86.8 12.1 23.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sxos55a1",
   "source_url": "https://omlx.ai/benchmarks/performance/sxos55a1",
   "tg_tokens": null,
   "tps": 12.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-sxos55a1-ctx8192",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: 1irn2nwz, 5vkfk87r, m18ww9hi; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 83.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 83.2 12.0 25.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sxos55a1",
   "source_url": "https://omlx.ai/benchmarks/performance/sxos55a1",
   "tg_tokens": null,
   "tps": 12.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-sxos55a1-ctx16384",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: 1irn2nwz, 5vkfk87r, m18ww9hi; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 80.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 80.6 11.4 26.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sxos55a1",
   "source_url": "https://omlx.ai/benchmarks/performance/sxos55a1",
   "tg_tokens": null,
   "tps": 11.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rhctn9l4-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 38.2 GB at this context; duplicate traces of the same session: abx3cg8r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 855.8,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Qwen3.8-27B-mxfp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 855.8 10.5 38.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rhctn9l4",
   "source_url": "https://omlx.ai/benchmarks/performance/rhctn9l4",
   "tg_tokens": null,
   "tps": 10.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rhctn9l4-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 38.1 GB at this context; duplicate traces of the same session: abx3cg8r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 458.0,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Qwen3.8-27B-mxfp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 458.0 14.4 38.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rhctn9l4",
   "source_url": "https://omlx.ai/benchmarks/performance/rhctn9l4",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-iwzpy05r-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 19.4 GB; TTFT 11561 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 708.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k tokens 708.6 PP tok/s 50.0 TG tok/s 11561 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iwzpy05r",
   "source_url": "https://omlx.ai/benchmarks/performance/iwzpy05r",
   "tg_tokens": null,
   "tps": 50.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-4134xqv6-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 19.0 GB; TTFT 17718 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 462.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k tokens 462.3 PP tok/s 38.7 TG tok/s 17718 TTFT (ms) 19.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4134xqv6",
   "source_url": "https://omlx.ai/benchmarks/performance/4134xqv6",
   "tg_tokens": null,
   "tps": 38.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-q9zkqlc0-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: 0oti28de; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1142.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,142 90.8 100.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q9zkqlc0",
   "source_url": "https://omlx.ai/benchmarks/performance/q9zkqlc0",
   "tg_tokens": null,
   "tps": 90.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-q9zkqlc0-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.5 GB at this context; duplicate traces of the same session: 0oti28de; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1367.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 1,367 73.5 104.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q9zkqlc0",
   "source_url": "https://omlx.ai/benchmarks/performance/q9zkqlc0",
   "tg_tokens": null,
   "tps": 73.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-evou0d43-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: 8lroixa4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 835.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 835.0 40.6 100.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/evou0d43",
   "source_url": "https://omlx.ai/benchmarks/performance/evou0d43",
   "tg_tokens": null,
   "tps": 40.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-evou0d43-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.0 GB at this context; duplicate traces of the same session: 8lroixa4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1356.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 1,356 47.3 104.0 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/evou0d43",
   "source_url": "https://omlx.ai/benchmarks/performance/evou0d43",
   "tg_tokens": null,
   "tps": 47.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-s6vowk5w-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 34.5 GB; TTFT 17806 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 460.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k tokens 460.1 PP tok/s 11.1 TG tok/s 17806 TTFT (ms) 34.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s6vowk5w",
   "source_url": "https://omlx.ai/benchmarks/performance/s6vowk5w",
   "tg_tokens": null,
   "tps": 11.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fdcjbetx-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 34.4 GB; TTFT 18474 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 443.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k tokens 443.4 PP tok/s 13.1 TG tok/s 18474 TTFT (ms) 34.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fdcjbetx",
   "source_url": "https://omlx.ai/benchmarks/performance/fdcjbetx",
   "tg_tokens": null,
   "tps": 13.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-2blu56l2-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 35.3 GB; TTFT 31951 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 512.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k tokens 512.8 PP tok/s 12.6 TG tok/s 31951 TTFT (ms) 35.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2blu56l2",
   "source_url": "https://omlx.ai/benchmarks/performance/2blu56l2",
   "tg_tokens": null,
   "tps": 12.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-loffd2sn-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 35.4 GB; TTFT 31736 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 516.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k tokens 516.3 PP tok/s 11.8 TG tok/s 31736 TTFT (ms) 35.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/loffd2sn",
   "source_url": "https://omlx.ai/benchmarks/performance/loffd2sn",
   "tg_tokens": null,
   "tps": 11.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-cjae61dz-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 35.3 GB; TTFT 32237 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 508.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k tokens 508.2 PP tok/s 13.0 TG tok/s 32237 TTFT (ms) 35.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cjae61dz",
   "source_url": "https://omlx.ai/benchmarks/performance/cjae61dz",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-4ck1oxy0-ctx32768",
   "model": "gemma-4-31B-it-scotoma-2",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 25.2 GB; TTFT 106047 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 309.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-scotoma-2-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k tokens 309.0 PP tok/s 12.2 TG tok/s 106047 TTFT (ms) 25.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ck1oxy0",
   "source_url": "https://omlx.ai/benchmarks/performance/4ck1oxy0",
   "tg_tokens": null,
   "tps": 12.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-tkwbkjec-ctx1024",
   "model": "Gemma-4-31B",
   "notes": "oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: sau43aax; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 230.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31b-it-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 230.7 7.9 17.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tkwbkjec",
   "source_url": "https://omlx.ai/benchmarks/performance/tkwbkjec",
   "tg_tokens": null,
   "tps": 7.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-tkwbkjec-ctx4096",
   "model": "Gemma-4-31B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: sau43aax; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 238.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31b-it-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 238.3 7.8 18.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tkwbkjec",
   "source_url": "https://omlx.ai/benchmarks/performance/tkwbkjec",
   "tg_tokens": null,
   "tps": 7.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-84x7rqdz-ctx1024",
   "model": "Devstral-Small-2-24B-Instruct-2512",
   "notes": "mistralai; oMLX community benchmark trace, single run at 1k context; peak memory 12.5 GB; TTFT 8508 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 120.4,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "mistralai-devstral-small-2-24b-instruct-2512-mlx-mxfp4 on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k tokens 120.4 PP tok/s 18.1 TG tok/s 8508 TTFT (ms) 12.5 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/84x7rqdz",
   "source_url": "https://omlx.ai/benchmarks/performance/84x7rqdz",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tcbnvswu-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 7pvwgdbf, i89069vd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 881.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 881.6 33.4 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tcbnvswu",
   "source_url": "https://omlx.ai/benchmarks/performance/tcbnvswu",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tcbnvswu-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 7pvwgdbf, i89069vd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1004.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,004 32.9 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tcbnvswu",
   "source_url": "https://omlx.ai/benchmarks/performance/tcbnvswu",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tcbnvswu-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 7pvwgdbf, i89069vd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1019.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,019 31.7 17.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tcbnvswu",
   "source_url": "https://omlx.ai/benchmarks/performance/tcbnvswu",
   "tg_tokens": null,
   "tps": 31.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-087upe5k-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "scottlowry MTP build; oMLX community benchmark trace; peak memory 30.0 GB at this context; duplicate traces of the same session: 7nlv2ny1; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 751.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "models--scottlowry--Ornith-1.5-35B-A3B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 751.1 65.7 30.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/087upe5k",
   "source_url": "https://omlx.ai/benchmarks/performance/087upe5k",
   "tg_tokens": null,
   "tps": 65.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-087upe5k-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "scottlowry MTP build; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: 7nlv2ny1; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 834.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "models--scottlowry--Ornith-1.5-35B-A3B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 834.3 68.9 32.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/087upe5k",
   "source_url": "https://omlx.ai/benchmarks/performance/087upe5k",
   "tg_tokens": null,
   "tps": 68.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ob5a8ltt-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "scottlowry MTP build; oMLX community benchmark trace; peak memory 30.0 GB at this context; duplicate traces of the same session: 53j4s5q8; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 752.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "models--scottlowry--Ornith-1.5-35B-A3B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 752.9 68.9 30.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ob5a8ltt",
   "source_url": "https://omlx.ai/benchmarks/performance/ob5a8ltt",
   "tg_tokens": null,
   "tps": 68.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ob5a8ltt-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "scottlowry MTP build; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: 53j4s5q8; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 824.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "models--scottlowry--Ornith-1.5-35B-A3B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 824.3 62.6 32.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ob5a8ltt",
   "source_url": "https://omlx.ai/benchmarks/performance/ob5a8ltt",
   "tg_tokens": null,
   "tps": 62.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-1ieedx2a-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "scottlowry MTP build; oMLX community benchmark trace; peak memory 30.0 GB at this context; duplicate traces of the same session: b4s26vye; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 752.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "models--scottlowry--Ornith-1.5-35B-A3B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 752.1 64.0 30.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ieedx2a",
   "source_url": "https://omlx.ai/benchmarks/performance/1ieedx2a",
   "tg_tokens": null,
   "tps": 64.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-1ieedx2a-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "scottlowry MTP build; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: b4s26vye; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 823.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "models--scottlowry--Ornith-1.5-35B-A3B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 823.9 74.4 32.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ieedx2a",
   "source_url": "https://omlx.ai/benchmarks/performance/1ieedx2a",
   "tg_tokens": null,
   "tps": 74.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-umn0ecaa-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "scottlowry MTP build; oMLX community benchmark trace; peak memory 30.0 GB at this context; duplicate traces of the same session: jjslzi84; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 751.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "models--scottlowry--Ornith-1.5-35B-A3B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 751.4 65.9 30.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/umn0ecaa",
   "source_url": "https://omlx.ai/benchmarks/performance/umn0ecaa",
   "tg_tokens": null,
   "tps": 65.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-umn0ecaa-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "scottlowry MTP build; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: jjslzi84; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 821.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "models--scottlowry--Ornith-1.5-35B-A3B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 821.8 66.5 32.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/umn0ecaa",
   "source_url": "https://omlx.ai/benchmarks/performance/umn0ecaa",
   "tg_tokens": null,
   "tps": 66.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ai76c3xo-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "scottlowry MTP build; oMLX community benchmark trace; peak memory 30.0 GB at this context; duplicate traces of the same session: j5wbubq7; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 600.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "models--scottlowry--Ornith-1.5-35B-A3B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 600.8 41.9 30.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ai76c3xo",
   "source_url": "https://omlx.ai/benchmarks/performance/ai76c3xo",
   "tg_tokens": null,
   "tps": 41.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ai76c3xo-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "scottlowry MTP build; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: j5wbubq7; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 806.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "models--scottlowry--Ornith-1.5-35B-A3B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 806.0 69.7 32.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ai76c3xo",
   "source_url": "https://omlx.ai/benchmarks/performance/ai76c3xo",
   "tg_tokens": null,
   "tps": 69.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-yjt5pli2-ctx1024",
   "model": "DeepSeek-V4-Flash-Vision-Exp",
   "notes": "oMLX community benchmark trace; peak memory 147.7 GB at this context; duplicate traces of the same session: njoxa1jk; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 621.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-V4-Flash-Vision-Exp-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 621.1 38.0 147.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yjt5pli2",
   "source_url": "https://omlx.ai/benchmarks/performance/yjt5pli2",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-yjt5pli2-ctx4096",
   "model": "DeepSeek-V4-Flash-Vision-Exp",
   "notes": "oMLX community benchmark trace; peak memory 149.9 GB at this context; duplicate traces of the same session: njoxa1jk; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 946.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-V4-Flash-Vision-Exp-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 946.5 33.9 149.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yjt5pli2",
   "source_url": "https://omlx.ai/benchmarks/performance/yjt5pli2",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-s89xp4ne-ctx4096",
   "model": "Devstral-Small-2-24B-Instruct-2512",
   "notes": "oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: b5cvrih1; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 124.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Devstral-Small-2-24B-Instruct-2512-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 124.2 21.1 14.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s89xp4ne",
   "source_url": "https://omlx.ai/benchmarks/performance/s89xp4ne",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-s89xp4ne-ctx16384",
   "model": "Devstral-Small-2-24B-Instruct-2512",
   "notes": "oMLX community benchmark trace; peak memory 16.7 GB at this context; duplicate traces of the same session: b5cvrih1; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 114.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Devstral-Small-2-24B-Instruct-2512-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 114.5 18.0 16.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s89xp4ne",
   "source_url": "https://omlx.ai/benchmarks/performance/s89xp4ne",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-953m25ib-ctx1024",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.8 GB at this context; duplicate traces of the same session: 50usjun1, 7gayn496; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 444.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 444.5 36.4 177.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/953m25ib",
   "source_url": "https://omlx.ai/benchmarks/performance/953m25ib",
   "tg_tokens": null,
   "tps": 36.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-953m25ib-ctx4096",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.9 GB at this context; duplicate traces of the same session: 50usjun1, 7gayn496; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 557.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 557.9 40.1 178.9 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/953m25ib",
   "source_url": "https://omlx.ai/benchmarks/performance/953m25ib",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-953m25ib-ctx8192",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.2 GB at this context; duplicate traces of the same session: 50usjun1, 7gayn496; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 511.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 511.3 53.3 179.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/953m25ib",
   "source_url": "https://omlx.ai/benchmarks/performance/953m25ib",
   "tg_tokens": null,
   "tps": 53.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-w60e2ubm-ctx1024",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.0 GB at this context; duplicate traces of the same session: 1ghyp5yk; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 470.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 470.6 40.3 178.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w60e2ubm",
   "source_url": "https://omlx.ai/benchmarks/performance/w60e2ubm",
   "tg_tokens": null,
   "tps": 40.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-w60e2ubm-ctx4096",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.6 GB at this context; duplicate traces of the same session: 1ghyp5yk; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 573.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 573.6 35.1 181.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w60e2ubm",
   "source_url": "https://omlx.ai/benchmarks/performance/w60e2ubm",
   "tg_tokens": null,
   "tps": 35.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-w60e2ubm-ctx8192",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.8 GB at this context; duplicate traces of the same session: 1ghyp5yk; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 537.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 537.7 41.7 181.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w60e2ubm",
   "source_url": "https://omlx.ai/benchmarks/performance/w60e2ubm",
   "tg_tokens": null,
   "tps": 41.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-bjcpca78-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace, single run at 8k context; peak memory 175.1 GB; TTFT 19400 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 422.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k tokens 422.3 PP tok/s 40.2 TG tok/s 19400 TTFT (ms) 175.1 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bjcpca78",
   "source_url": "https://omlx.ai/benchmarks/performance/bjcpca78",
   "tg_tokens": null,
   "tps": 40.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-umons4be-ctx1024",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 9.9 GB at this context; duplicate traces of the same session: 29ctdy4i, 5zd2d3dz, b932kn9b, cxytmv9j, plty2gtw, wmdjxu4e; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3165.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 3,165 102.6 9.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/umons4be",
   "source_url": "https://omlx.ai/benchmarks/performance/umons4be",
   "tg_tokens": null,
   "tps": 102.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-umons4be-ctx4096",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 9.9 GB at this context; duplicate traces of the same session: 29ctdy4i, 5zd2d3dz, b932kn9b, cxytmv9j, plty2gtw, wmdjxu4e; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3429.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 3,429 100.8 9.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/umons4be",
   "source_url": "https://omlx.ai/benchmarks/performance/umons4be",
   "tg_tokens": null,
   "tps": 100.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-umons4be-ctx8192",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 10.0 GB at this context; duplicate traces of the same session: 29ctdy4i, 5zd2d3dz, b932kn9b, cxytmv9j, plty2gtw, wmdjxu4e; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3333.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 8k 3,333 100.4 10.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/umons4be",
   "source_url": "https://omlx.ai/benchmarks/performance/umons4be",
   "tg_tokens": null,
   "tps": 100.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-umons4be-ctx16384",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 10.1 GB at this context; duplicate traces of the same session: 29ctdy4i, 5zd2d3dz, b932kn9b, cxytmv9j, plty2gtw, wmdjxu4e; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3167.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 16k 3,167 101.4 10.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/umons4be",
   "source_url": "https://omlx.ai/benchmarks/performance/umons4be",
   "tg_tokens": null,
   "tps": 101.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-umons4be-ctx32768",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 10.3 GB at this context; duplicate traces of the same session: 29ctdy4i, 5zd2d3dz, b932kn9b, cxytmv9j, plty2gtw, wmdjxu4e; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2873.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 32k 2,873 92.0 10.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/umons4be",
   "source_url": "https://omlx.ai/benchmarks/performance/umons4be",
   "tg_tokens": null,
   "tps": 92.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-umons4be-ctx65536",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 10.7 GB at this context; duplicate traces of the same session: 29ctdy4i, 5zd2d3dz, b932kn9b, cxytmv9j, plty2gtw, wmdjxu4e; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2436.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 64k 2,436 77.1 10.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/umons4be",
   "source_url": "https://omlx.ai/benchmarks/performance/umons4be",
   "tg_tokens": null,
   "tps": 77.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-umons4be-ctx131072",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 11.6 GB at this context; duplicate traces of the same session: 29ctdy4i, 5zd2d3dz, b932kn9b, cxytmv9j, plty2gtw, wmdjxu4e; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1866.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 128k 1,866 61.7 11.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/umons4be",
   "source_url": "https://omlx.ai/benchmarks/performance/umons4be",
   "tg_tokens": null,
   "tps": 61.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-75ahoci1-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: o1297wo9; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 694.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 694.0 72.0 19.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/75ahoci1",
   "source_url": "https://omlx.ai/benchmarks/performance/75ahoci1",
   "tg_tokens": null,
   "tps": 72.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-75ahoci1-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: o1297wo9; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1558.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,558 70.7 19.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/75ahoci1",
   "source_url": "https://omlx.ai/benchmarks/performance/75ahoci1",
   "tg_tokens": null,
   "tps": 70.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ge0xk88o-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: unaoaa1r, x4z3g1u8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2668.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 2,668 176.5 37.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ge0xk88o",
   "source_url": "https://omlx.ai/benchmarks/performance/ge0xk88o",
   "tg_tokens": null,
   "tps": 176.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ge0xk88o-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: unaoaa1r, x4z3g1u8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3311.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 3,311 141.1 39.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ge0xk88o",
   "source_url": "https://omlx.ai/benchmarks/performance/ge0xk88o",
   "tg_tokens": null,
   "tps": 141.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ge0xk88o-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.4 GB at this context; duplicate traces of the same session: unaoaa1r, x4z3g1u8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3146.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 3,146 150.6 39.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ge0xk88o",
   "source_url": "https://omlx.ai/benchmarks/performance/ge0xk88o",
   "tg_tokens": null,
   "tps": 150.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xzi6upj4-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: c070qeez, pm5rk9pm; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2235.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 2,235 94.5 37.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xzi6upj4",
   "source_url": "https://omlx.ai/benchmarks/performance/xzi6upj4",
   "tg_tokens": null,
   "tps": 94.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xzi6upj4-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: c070qeez, pm5rk9pm; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3315.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 3,315 134.1 39.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xzi6upj4",
   "source_url": "https://omlx.ai/benchmarks/performance/xzi6upj4",
   "tg_tokens": null,
   "tps": 134.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xzi6upj4-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.4 GB at this context; duplicate traces of the same session: c070qeez, pm5rk9pm; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3148.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 3,148 62.6 39.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xzi6upj4",
   "source_url": "https://omlx.ai/benchmarks/performance/xzi6upj4",
   "tg_tokens": null,
   "tps": 62.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-mgps5ag0-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: 5e7xujuf; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 294.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 294.4 9.9 15.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mgps5ag0",
   "source_url": "https://omlx.ai/benchmarks/performance/mgps5ag0",
   "tg_tokens": null,
   "tps": 9.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-mgps5ag0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 5e7xujuf; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 311.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 311.2 9.7 17.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mgps5ag0",
   "source_url": "https://omlx.ai/benchmarks/performance/mgps5ag0",
   "tg_tokens": null,
   "tps": 9.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-vmfzmzp9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: 5l6i4v5u, ll3pbgs0, qw1ofomt, u7400b3t; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.8,
   "provenance": "community",
   "quant": "AWQ-5.0bpw",
   "quote": "Qwen3.8-27B-AWQ-5.0bpw on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 unknown; 1k 127.8 30.7 23.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vmfzmzp9",
   "source_url": "https://omlx.ai/benchmarks/performance/vmfzmzp9",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-vmfzmzp9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: 5l6i4v5u, ll3pbgs0, qw1ofomt, u7400b3t; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 171.3,
   "provenance": "community",
   "quant": "AWQ-5.0bpw",
   "quote": "Qwen3.8-27B-AWQ-5.0bpw on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 unknown; 4k 171.3 25.0 26.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vmfzmzp9",
   "source_url": "https://omlx.ai/benchmarks/performance/vmfzmzp9",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-vmfzmzp9-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 27.2 GB at this context; duplicate traces of the same session: 5l6i4v5u, ll3pbgs0, qw1ofomt, u7400b3t; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 156.4,
   "provenance": "community",
   "quant": "AWQ-5.0bpw",
   "quote": "Qwen3.8-27B-AWQ-5.0bpw on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 unknown; 8k 156.4 25.4 27.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vmfzmzp9",
   "source_url": "https://omlx.ai/benchmarks/performance/vmfzmzp9",
   "tg_tokens": null,
   "tps": 25.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-vmfzmzp9-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: 5l6i4v5u, ll3pbgs0, qw1ofomt, u7400b3t; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 129.8,
   "provenance": "community",
   "quant": "AWQ-5.0bpw",
   "quote": "Qwen3.8-27B-AWQ-5.0bpw on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 unknown; 16k 129.8 31.1 27.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vmfzmzp9",
   "source_url": "https://omlx.ai/benchmarks/performance/vmfzmzp9",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-vmfzmzp9-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: 5l6i4v5u, ll3pbgs0, qw1ofomt, u7400b3t; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 109.7,
   "provenance": "community",
   "quant": "AWQ-5.0bpw",
   "quote": "Qwen3.8-27B-AWQ-5.0bpw on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 unknown; 32k 109.7 30.8 27.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vmfzmzp9",
   "source_url": "https://omlx.ai/benchmarks/performance/vmfzmzp9",
   "tg_tokens": null,
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-vmfzmzp9-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 30.0 GB at this context; duplicate traces of the same session: 5l6i4v5u, ll3pbgs0, qw1ofomt, u7400b3t; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 91.9,
   "provenance": "community",
   "quant": "AWQ-5.0bpw",
   "quote": "Qwen3.8-27B-AWQ-5.0bpw on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 unknown; 64k 91.9 20.6 30.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vmfzmzp9",
   "source_url": "https://omlx.ai/benchmarks/performance/vmfzmzp9",
   "tg_tokens": null,
   "tps": 20.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-vmfzmzp9-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.5 GB at this context; duplicate traces of the same session: 5l6i4v5u, ll3pbgs0, qw1ofomt, u7400b3t; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 80.0,
   "provenance": "community",
   "quant": "AWQ-5.0bpw",
   "quote": "Qwen3.8-27B-AWQ-5.0bpw on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 unknown; 128k 80.0 20.7 34.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vmfzmzp9",
   "source_url": "https://omlx.ai/benchmarks/performance/vmfzmzp9",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-vmfzmzp9-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 39.5 GB at this context; duplicate traces of the same session: 5l6i4v5u, ll3pbgs0, qw1ofomt, u7400b3t; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 66.9,
   "provenance": "community",
   "quant": "AWQ-5.0bpw",
   "quote": "Qwen3.8-27B-AWQ-5.0bpw on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 unknown; 195k 66.9 13.1 39.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vmfzmzp9",
   "source_url": "https://omlx.ai/benchmarks/performance/vmfzmzp9",
   "tg_tokens": null,
   "tps": 13.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7j4wb8b1-ctx1024",
   "model": "Qwen3.8-Flash-Next-heretic-2",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.8 GB; TTFT 2531 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 404.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-heretic-2-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 404.6 PP tok/s 53.5 TG tok/s 2531 TTFT (ms) 70.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7j4wb8b1",
   "source_url": "https://omlx.ai/benchmarks/performance/7j4wb8b1",
   "tg_tokens": null,
   "tps": 53.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-vpe3lsnu-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 71.5 GB; TTFT 9624 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 851.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k tokens 851.3 PP tok/s 50.9 TG tok/s 9624 TTFT (ms) 71.5 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vpe3lsnu",
   "source_url": "https://omlx.ai/benchmarks/performance/vpe3lsnu",
   "tg_tokens": null,
   "tps": 50.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ym87csd4-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: clxpyob2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 523.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 523.6 60.8 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ym87csd4",
   "source_url": "https://omlx.ai/benchmarks/performance/ym87csd4",
   "tg_tokens": null,
   "tps": 60.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ym87csd4-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: clxpyob2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 680.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 680.2 58.7 71.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ym87csd4",
   "source_url": "https://omlx.ai/benchmarks/performance/ym87csd4",
   "tg_tokens": null,
   "tps": 58.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-oyxnc90m-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.2 GB at this context; duplicate traces of the same session: 0stdid89, b86dy88h; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1098.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 1,098 82.2 183.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oyxnc90m",
   "source_url": "https://omlx.ai/benchmarks/performance/oyxnc90m",
   "tg_tokens": null,
   "tps": 82.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-oyxnc90m-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.6 GB at this context; duplicate traces of the same session: 0stdid89, b86dy88h; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1373.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 1,373 61.8 183.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oyxnc90m",
   "source_url": "https://omlx.ai/benchmarks/performance/oyxnc90m",
   "tg_tokens": null,
   "tps": 61.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-oyxnc90m-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.8 GB at this context; duplicate traces of the same session: 0stdid89, b86dy88h; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1361.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 1,361 74.8 183.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oyxnc90m",
   "source_url": "https://omlx.ai/benchmarks/performance/oyxnc90m",
   "tg_tokens": null,
   "tps": 74.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gddfgmp2-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.1 GB at this context; duplicate traces of the same session: tr172qzd; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1100.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 1,100 82.8 183.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gddfgmp2",
   "source_url": "https://omlx.ai/benchmarks/performance/gddfgmp2",
   "tg_tokens": null,
   "tps": 82.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gddfgmp2-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.6 GB at this context; duplicate traces of the same session: tr172qzd; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1373.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 1,373 82.5 183.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gddfgmp2",
   "source_url": "https://omlx.ai/benchmarks/performance/gddfgmp2",
   "tg_tokens": null,
   "tps": 82.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-fzpv8z18-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.2 GB at this context; duplicate traces of the same session: tgfsgo51, tp4xy2gb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 850.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 850.6 30.2 183.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fzpv8z18",
   "source_url": "https://omlx.ai/benchmarks/performance/fzpv8z18",
   "tg_tokens": null,
   "tps": 30.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-fzpv8z18-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.6 GB at this context; duplicate traces of the same session: tgfsgo51, tp4xy2gb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1117.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 1,117 76.2 183.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fzpv8z18",
   "source_url": "https://omlx.ai/benchmarks/performance/fzpv8z18",
   "tg_tokens": null,
   "tps": 76.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-fzpv8z18-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.8 GB at this context; duplicate traces of the same session: tgfsgo51, tp4xy2gb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1359.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 1,359 90.3 183.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fzpv8z18",
   "source_url": "https://omlx.ai/benchmarks/performance/fzpv8z18",
   "tg_tokens": null,
   "tps": 90.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-kqdlv42o-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 15.5 GB at this context; duplicate traces of the same session: nu8g9wek; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1196.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k 1,196 58.0 15.5 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kqdlv42o",
   "source_url": "https://omlx.ai/benchmarks/performance/kqdlv42o",
   "tg_tokens": null,
   "tps": 58.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-kqdlv42o-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: nu8g9wek; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1212.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 1,212 39.9 15.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kqdlv42o",
   "source_url": "https://omlx.ai/benchmarks/performance/kqdlv42o",
   "tg_tokens": null,
   "tps": 39.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-oqlxg7mt-ctx1024",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.0 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 470.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 470.6 40.3 178.0 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oqlxg7mt",
   "source_url": "https://omlx.ai/benchmarks/performance/oqlxg7mt",
   "tg_tokens": null,
   "tps": 40.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-oqlxg7mt-ctx4096",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.6 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 573.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 573.6 35.1 181.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oqlxg7mt",
   "source_url": "https://omlx.ai/benchmarks/performance/oqlxg7mt",
   "tg_tokens": null,
   "tps": 35.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-oqlxg7mt-ctx8192",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.8 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 537.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 537.7 41.7 181.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oqlxg7mt",
   "source_url": "https://omlx.ai/benchmarks/performance/oqlxg7mt",
   "tg_tokens": null,
   "tps": 41.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-e7iec7le-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 175.2 GB; TTFT 43121 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 380.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k tokens 380.0 PP tok/s 39.6 TG tok/s 43121 TTFT (ms) 175.2 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e7iec7le",
   "source_url": "https://omlx.ai/benchmarks/performance/e7iec7le",
   "tg_tokens": null,
   "tps": 39.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-6ea97pq3-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.9 GB at this context; duplicate traces of the same session: 979y3xp7; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 456.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 456.7 40.0 174.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ea97pq3",
   "source_url": "https://omlx.ai/benchmarks/performance/6ea97pq3",
   "tg_tokens": null,
   "tps": 40.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-6ea97pq3-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.1 GB at this context; duplicate traces of the same session: 979y3xp7; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 427.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 427.0 40.0 175.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ea97pq3",
   "source_url": "https://omlx.ai/benchmarks/performance/6ea97pq3",
   "tg_tokens": null,
   "tps": 40.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-fm4r6iqj-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 323.5 GB at this context; duplicate traces of the same session: 28jozh48, ieg02rxw; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 498.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 498.6 36.3 323.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fm4r6iqj",
   "source_url": "https://omlx.ai/benchmarks/performance/fm4r6iqj",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-fm4r6iqj-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 324.7 GB at this context; duplicate traces of the same session: 28jozh48, ieg02rxw; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 586.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 586.5 43.2 324.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fm4r6iqj",
   "source_url": "https://omlx.ai/benchmarks/performance/fm4r6iqj",
   "tg_tokens": null,
   "tps": 43.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-fm4r6iqj-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 326.1 GB at this context; duplicate traces of the same session: 28jozh48, ieg02rxw; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 521.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 64k 521.5 38.5 326.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fm4r6iqj",
   "source_url": "https://omlx.ai/benchmarks/performance/fm4r6iqj",
   "tg_tokens": null,
   "tps": 38.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-yg9kaxad-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: y9t2yigh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 818.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 818.4 109.3 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yg9kaxad",
   "source_url": "https://omlx.ai/benchmarks/performance/yg9kaxad",
   "tg_tokens": null,
   "tps": 109.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-yg9kaxad-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: y9t2yigh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1945.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,945 106.1 20.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yg9kaxad",
   "source_url": "https://omlx.ai/benchmarks/performance/yg9kaxad",
   "tg_tokens": null,
   "tps": 106.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-v0vjwn23-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 12.2 GB; TTFT 3276 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 312.5,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Qwen3.8-27B-2bit-mlx on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 2bit; 1k tokens 312.5 PP tok/s 32.6 TG tok/s 3276 TTFT (ms) 12.2 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v0vjwn23",
   "source_url": "https://omlx.ai/benchmarks/performance/v0vjwn23",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9vzivcbm-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: cdkj8s82; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 822.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 822.4 18.2 28.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9vzivcbm",
   "source_url": "https://omlx.ai/benchmarks/performance/9vzivcbm",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9vzivcbm-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: cdkj8s82; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 913.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 913.8 18.1 29.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9vzivcbm",
   "source_url": "https://omlx.ai/benchmarks/performance/9vzivcbm",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-taw24jip-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: q2reypyz, sce91oik; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.8,
   "provenance": "community",
   "quant": "AWQ-5.0bpw",
   "quote": "Qwen3.8-27B-AWQ-5.0bpw on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 unknown; 1k 127.8 30.7 23.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/taw24jip",
   "source_url": "https://omlx.ai/benchmarks/performance/taw24jip",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-taw24jip-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: q2reypyz, sce91oik; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 171.3,
   "provenance": "community",
   "quant": "AWQ-5.0bpw",
   "quote": "Qwen3.8-27B-AWQ-5.0bpw on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 unknown; 4k 171.3 25.0 26.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/taw24jip",
   "source_url": "https://omlx.ai/benchmarks/performance/taw24jip",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-taw24jip-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 27.2 GB at this context; duplicate traces of the same session: q2reypyz, sce91oik; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 156.4,
   "provenance": "community",
   "quant": "AWQ-5.0bpw",
   "quote": "Qwen3.8-27B-AWQ-5.0bpw on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 unknown; 8k 156.4 25.4 27.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/taw24jip",
   "source_url": "https://omlx.ai/benchmarks/performance/taw24jip",
   "tg_tokens": null,
   "tps": 25.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-taw24jip-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: q2reypyz, sce91oik; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 129.8,
   "provenance": "community",
   "quant": "AWQ-5.0bpw",
   "quote": "Qwen3.8-27B-AWQ-5.0bpw on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 unknown; 16k 129.8 31.1 27.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/taw24jip",
   "source_url": "https://omlx.ai/benchmarks/performance/taw24jip",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-taw24jip-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: q2reypyz, sce91oik; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 109.7,
   "provenance": "community",
   "quant": "AWQ-5.0bpw",
   "quote": "Qwen3.8-27B-AWQ-5.0bpw on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 unknown; 32k 109.7 30.8 27.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/taw24jip",
   "source_url": "https://omlx.ai/benchmarks/performance/taw24jip",
   "tg_tokens": null,
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-taw24jip-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 30.0 GB at this context; duplicate traces of the same session: q2reypyz, sce91oik; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 91.9,
   "provenance": "community",
   "quant": "AWQ-5.0bpw",
   "quote": "Qwen3.8-27B-AWQ-5.0bpw on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 unknown; 64k 91.9 20.6 30.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/taw24jip",
   "source_url": "https://omlx.ai/benchmarks/performance/taw24jip",
   "tg_tokens": null,
   "tps": 20.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-taw24jip-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.5 GB at this context; duplicate traces of the same session: q2reypyz, sce91oik; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 80.0,
   "provenance": "community",
   "quant": "AWQ-5.0bpw",
   "quote": "Qwen3.8-27B-AWQ-5.0bpw on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 unknown; 128k 80.0 20.7 34.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/taw24jip",
   "source_url": "https://omlx.ai/benchmarks/performance/taw24jip",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-taw24jip-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 39.5 GB at this context; duplicate traces of the same session: q2reypyz, sce91oik; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 66.9,
   "provenance": "community",
   "quant": "AWQ-5.0bpw",
   "quote": "Qwen3.8-27B-AWQ-5.0bpw on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 unknown; 195k 66.9 13.1 39.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/taw24jip",
   "source_url": "https://omlx.ai/benchmarks/performance/taw24jip",
   "tg_tokens": null,
   "tps": 13.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ko3tkfmr-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: j0tils9f; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 346.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 346.4 31.9 22.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ko3tkfmr",
   "source_url": "https://omlx.ai/benchmarks/performance/ko3tkfmr",
   "tg_tokens": null,
   "tps": 31.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ko3tkfmr-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: j0tils9f; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 341.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 341.3 48.2 22.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ko3tkfmr",
   "source_url": "https://omlx.ai/benchmarks/performance/ko3tkfmr",
   "tg_tokens": null,
   "tps": 48.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tvr30aaw-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 75.3 GB at this context; duplicate traces of the same session: 302hc8sb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 498.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 498.4 39.1 75.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tvr30aaw",
   "source_url": "https://omlx.ai/benchmarks/performance/tvr30aaw",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tvr30aaw-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 75.5 GB at this context; duplicate traces of the same session: 302hc8sb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1077.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,077 34.8 75.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tvr30aaw",
   "source_url": "https://omlx.ai/benchmarks/performance/tvr30aaw",
   "tg_tokens": null,
   "tps": 34.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-uzzzh8jf-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 71.7 GB; TTFT 17832 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 918.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k tokens 918.8 PP tok/s 38.7 TG tok/s 17832 TTFT (ms) 71.7 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uzzzh8jf",
   "source_url": "https://omlx.ai/benchmarks/performance/uzzzh8jf",
   "tg_tokens": null,
   "tps": 38.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-8zmonba5-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.3 GB at this context; duplicate traces of the same session: mobtlrcd, rk47fjw5, wewccdap; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 595.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 595.4 68.7 69.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8zmonba5",
   "source_url": "https://omlx.ai/benchmarks/performance/8zmonba5",
   "tg_tokens": null,
   "tps": 68.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-8zmonba5-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.8 GB at this context; duplicate traces of the same session: mobtlrcd, rk47fjw5, wewccdap; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 675.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 675.0 58.4 69.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8zmonba5",
   "source_url": "https://omlx.ai/benchmarks/performance/8zmonba5",
   "tg_tokens": null,
   "tps": 58.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-8zmonba5-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: mobtlrcd, rk47fjw5, wewccdap; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1089.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 1,089 65.0 70.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8zmonba5",
   "source_url": "https://omlx.ai/benchmarks/performance/8zmonba5",
   "tg_tokens": null,
   "tps": 65.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-8zmonba5-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: mobtlrcd, rk47fjw5, wewccdap; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1073.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k 1,073 56.7 71.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8zmonba5",
   "source_url": "https://omlx.ai/benchmarks/performance/8zmonba5",
   "tg_tokens": null,
   "tps": 56.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-8zmonba5-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.9 GB at this context; duplicate traces of the same session: mobtlrcd, rk47fjw5, wewccdap; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1055.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 128k 1,055 56.7 74.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8zmonba5",
   "source_url": "https://omlx.ai/benchmarks/performance/8zmonba5",
   "tg_tokens": null,
   "tps": 56.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-q61mvcli-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 186.4 GB at this context; duplicate traces of the same session: ik792019, jbgbze61; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 932.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 932.1 83.7 186.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q61mvcli",
   "source_url": "https://omlx.ai/benchmarks/performance/q61mvcli",
   "tg_tokens": null,
   "tps": 83.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-q61mvcli-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 188.6 GB at this context; duplicate traces of the same session: ik792019, jbgbze61; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1316.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 1,316 86.6 188.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q61mvcli",
   "source_url": "https://omlx.ai/benchmarks/performance/q61mvcli",
   "tg_tokens": null,
   "tps": 86.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-q61mvcli-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 188.6 GB at this context; duplicate traces of the same session: ik792019, jbgbze61; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1313.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 1,313 81.4 188.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q61mvcli",
   "source_url": "https://omlx.ai/benchmarks/performance/q61mvcli",
   "tg_tokens": null,
   "tps": 81.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-ayierl5l-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 0h4p02p1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 220.6 19.3 19.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ayierl5l",
   "source_url": "https://omlx.ai/benchmarks/performance/ayierl5l",
   "tg_tokens": null,
   "tps": 19.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-ayierl5l-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 0h4p02p1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 224.2 21.3 21.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ayierl5l",
   "source_url": "https://omlx.ai/benchmarks/performance/ayierl5l",
   "tg_tokens": null,
   "tps": 21.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-vjzfwef5-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: j9uramv5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 221.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 221.7 29.5 19.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vjzfwef5",
   "source_url": "https://omlx.ai/benchmarks/performance/vjzfwef5",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-vjzfwef5-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: j9uramv5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 226.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 226.0 35.4 21.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vjzfwef5",
   "source_url": "https://omlx.ai/benchmarks/performance/vjzfwef5",
   "tg_tokens": null,
   "tps": 35.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-p5vvpc50-ctx1024",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 4fbg5lpm, 808vtlzi, c8z2fuw5; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 331.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 8bit; 1k 331.4 34.8 16.6 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p5vvpc50",
   "source_url": "https://omlx.ai/benchmarks/performance/p5vvpc50",
   "tg_tokens": null,
   "tps": 34.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-p5vvpc50-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: 4fbg5lpm, 808vtlzi, c8z2fuw5; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 587.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 8bit; 4k 587.0 31.6 16.9 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p5vvpc50",
   "source_url": "https://omlx.ai/benchmarks/performance/p5vvpc50",
   "tg_tokens": null,
   "tps": 31.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-p5vvpc50-ctx8192",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: 4fbg5lpm, 808vtlzi, c8z2fuw5; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 567.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 8bit; 8k 567.6 32.2 17.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p5vvpc50",
   "source_url": "https://omlx.ai/benchmarks/performance/p5vvpc50",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-p5vvpc50-ctx16384",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 17.5 GB at this context; duplicate traces of the same session: 4fbg5lpm, 808vtlzi, c8z2fuw5; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 567.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 8bit; 16k 567.2 31.8 17.5 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p5vvpc50",
   "source_url": "https://omlx.ai/benchmarks/performance/p5vvpc50",
   "tg_tokens": null,
   "tps": 31.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ejsz64fh-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 861l52po, v6toexm4, xpcyiehg; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1479.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 1,479 115.7 19.4 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ejsz64fh",
   "source_url": "https://omlx.ai/benchmarks/performance/ejsz64fh",
   "tg_tokens": null,
   "tps": 115.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ejsz64fh-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 861l52po, v6toexm4, xpcyiehg; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1608.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 1,608 116.8 20.2 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ejsz64fh",
   "source_url": "https://omlx.ai/benchmarks/performance/ejsz64fh",
   "tg_tokens": null,
   "tps": 116.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ejsz64fh-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: 861l52po, v6toexm4, xpcyiehg; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1517.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 1,517 109.0 20.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ejsz64fh",
   "source_url": "https://omlx.ai/benchmarks/performance/ejsz64fh",
   "tg_tokens": null,
   "tps": 109.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ejsz64fh-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: 861l52po, v6toexm4, xpcyiehg; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1282.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 1,282 98.6 20.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ejsz64fh",
   "source_url": "https://omlx.ai/benchmarks/performance/ejsz64fh",
   "tg_tokens": null,
   "tps": 98.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-4mh175l8-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: 1t50ovq0, 6ihhw8w6, w8psiiqp; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1532.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 1,532 123.7 20.4 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4mh175l8",
   "source_url": "https://omlx.ai/benchmarks/performance/4mh175l8",
   "tg_tokens": null,
   "tps": 123.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-4mh175l8-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 1t50ovq0, 6ihhw8w6, w8psiiqp; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1555.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 1,555 98.1 21.4 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4mh175l8",
   "source_url": "https://omlx.ai/benchmarks/performance/4mh175l8",
   "tg_tokens": null,
   "tps": 98.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-4mh175l8-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 1t50ovq0, 6ihhw8w6, w8psiiqp; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1387.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 1,387 57.7 21.9 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4mh175l8",
   "source_url": "https://omlx.ai/benchmarks/performance/4mh175l8",
   "tg_tokens": null,
   "tps": 57.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-4mh175l8-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 1t50ovq0, 6ihhw8w6, w8psiiqp; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1217.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 1,217 75.4 21.9 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4mh175l8",
   "source_url": "https://omlx.ai/benchmarks/performance/4mh175l8",
   "tg_tokens": null,
   "tps": 75.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ahltoc0z-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 662g94ji, l6ybsb7g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1030.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,030 33.3 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ahltoc0z",
   "source_url": "https://omlx.ai/benchmarks/performance/ahltoc0z",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ahltoc0z-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 662g94ji, l6ybsb7g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1438.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,438 32.8 18.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ahltoc0z",
   "source_url": "https://omlx.ai/benchmarks/performance/ahltoc0z",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ahltoc0z-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: 662g94ji, l6ybsb7g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1497.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,497 31.9 18.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ahltoc0z",
   "source_url": "https://omlx.ai/benchmarks/performance/ahltoc0z",
   "tg_tokens": null,
   "tps": 31.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-kea9edp0-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace, single run at 8k context; peak memory 175.1 GB; TTFT 19280 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 424.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k tokens 424.9 PP tok/s 39.5 TG tok/s 19280 TTFT (ms) 175.1 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kea9edp0",
   "source_url": "https://omlx.ai/benchmarks/performance/kea9edp0",
   "tg_tokens": null,
   "tps": 39.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-c7lbvqjc-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: jjffo7gy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1431.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 1,431 101.6 19.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c7lbvqjc",
   "source_url": "https://omlx.ai/benchmarks/performance/c7lbvqjc",
   "tg_tokens": null,
   "tps": 101.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-c7lbvqjc-ctx8192",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: jjffo7gy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1575.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 8k 1,575 87.4 20.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c7lbvqjc",
   "source_url": "https://omlx.ai/benchmarks/performance/c7lbvqjc",
   "tg_tokens": null,
   "tps": 87.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-cz41zywr-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: e3defhdf, sl0rsleh, z7nb6eyp; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 220.4 31.7 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cz41zywr",
   "source_url": "https://omlx.ai/benchmarks/performance/cz41zywr",
   "tg_tokens": null,
   "tps": 31.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-cz41zywr-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: e3defhdf, sl0rsleh, z7nb6eyp; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 207.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 207.5 30.3 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cz41zywr",
   "source_url": "https://omlx.ai/benchmarks/performance/cz41zywr",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-cz41zywr-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: e3defhdf, sl0rsleh, z7nb6eyp; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 187.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 16k 187.4 25.6 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cz41zywr",
   "source_url": "https://omlx.ai/benchmarks/performance/cz41zywr",
   "tg_tokens": null,
   "tps": 25.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-cz41zywr-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: e3defhdf, sl0rsleh, z7nb6eyp; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 174.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 32k 174.7 21.7 22.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cz41zywr",
   "source_url": "https://omlx.ai/benchmarks/performance/cz41zywr",
   "tg_tokens": null,
   "tps": 21.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-nxv2xjwe-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 9faz2947, 9ofm5da3, jbj5c64u; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 222.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 222.1 31.4 18.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nxv2xjwe",
   "source_url": "https://omlx.ai/benchmarks/performance/nxv2xjwe",
   "tg_tokens": null,
   "tps": 31.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-nxv2xjwe-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 9faz2947, 9ofm5da3, jbj5c64u; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 240.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 240.5 30.2 22.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nxv2xjwe",
   "source_url": "https://omlx.ai/benchmarks/performance/nxv2xjwe",
   "tg_tokens": null,
   "tps": 30.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-nxv2xjwe-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: 9faz2947, 9ofm5da3, jbj5c64u; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 242.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 16k 242.5 25.5 24.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nxv2xjwe",
   "source_url": "https://omlx.ai/benchmarks/performance/nxv2xjwe",
   "tg_tokens": null,
   "tps": 25.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-nxv2xjwe-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: 9faz2947, 9ofm5da3, jbj5c64u; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 211.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 32k 211.1 21.6 25.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nxv2xjwe",
   "source_url": "https://omlx.ai/benchmarks/performance/nxv2xjwe",
   "tg_tokens": null,
   "tps": 21.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-j1z7e8px-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 7ar0qal1, artd7dnt, hat2j9j2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 222.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 222.1 31.3 18.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j1z7e8px",
   "source_url": "https://omlx.ai/benchmarks/performance/j1z7e8px",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-j1z7e8px-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 7ar0qal1, artd7dnt, hat2j9j2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 240.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 240.6 30.3 22.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j1z7e8px",
   "source_url": "https://omlx.ai/benchmarks/performance/j1z7e8px",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-j1z7e8px-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 7ar0qal1, artd7dnt, hat2j9j2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 203.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 16k 203.5 25.5 22.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j1z7e8px",
   "source_url": "https://omlx.ai/benchmarks/performance/j1z7e8px",
   "tg_tokens": null,
   "tps": 25.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-j1z7e8px-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: 7ar0qal1, artd7dnt, hat2j9j2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 174.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 32k 174.7 21.5 22.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j1z7e8px",
   "source_url": "https://omlx.ai/benchmarks/performance/j1z7e8px",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-r7elk7zu-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: 42e142ys; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 794.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 794.4 46.4 33.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r7elk7zu",
   "source_url": "https://omlx.ai/benchmarks/performance/r7elk7zu",
   "tg_tokens": null,
   "tps": 46.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-r7elk7zu-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: 42e142ys; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 882.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 882.1 40.4 34.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r7elk7zu",
   "source_url": "https://omlx.ai/benchmarks/performance/r7elk7zu",
   "tg_tokens": null,
   "tps": 40.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-6mdqo45v-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 25.1 GB; TTFT 18536 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 442.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k tokens 442.0 PP tok/s 24.4 TG tok/s 18536 TTFT (ms) 25.1 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6mdqo45v",
   "source_url": "https://omlx.ai/benchmarks/performance/6mdqo45v",
   "tg_tokens": null,
   "tps": 24.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hl20em3x-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 22.5 GB; TTFT 24054 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 340.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k tokens 340.6 PP tok/s 44.5 TG tok/s 24054 TTFT (ms) 22.5 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hl20em3x",
   "source_url": "https://omlx.ai/benchmarks/performance/hl20em3x",
   "tg_tokens": null,
   "tps": 44.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0em2sk0d-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: lyg10oyp, z8o09lzo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 906.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 906.0 61.6 17.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0em2sk0d",
   "source_url": "https://omlx.ai/benchmarks/performance/0em2sk0d",
   "tg_tokens": null,
   "tps": 61.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0em2sk0d-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: lyg10oyp, z8o09lzo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 982.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 982.4 68.2 18.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0em2sk0d",
   "source_url": "https://omlx.ai/benchmarks/performance/0em2sk0d",
   "tg_tokens": null,
   "tps": 68.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0em2sk0d-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: lyg10oyp, z8o09lzo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 902.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 902.1 64.5 19.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0em2sk0d",
   "source_url": "https://omlx.ai/benchmarks/performance/0em2sk0d",
   "tg_tokens": null,
   "tps": 64.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-6ij0dewx-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: jldrbz50; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 202.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 202.4 15.1 18.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ij0dewx",
   "source_url": "https://omlx.ai/benchmarks/performance/6ij0dewx",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-6ij0dewx-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: jldrbz50; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 200.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 200.3 22.0 19.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ij0dewx",
   "source_url": "https://omlx.ai/benchmarks/performance/6ij0dewx",
   "tg_tokens": null,
   "tps": 22.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-oh4t2g3h-ctx16384",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "vision; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: lr9z3852; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 844.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ5e-vision on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 5bit; 16k 844.0 59.8 24.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oh4t2g3h",
   "source_url": "https://omlx.ai/benchmarks/performance/oh4t2g3h",
   "tg_tokens": null,
   "tps": 59.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-oh4t2g3h-ctx65536",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "vision; oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: lr9z3852; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 512.8,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ5e-vision on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 5bit; 64k 512.8 43.0 26.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oh4t2g3h",
   "source_url": "https://omlx.ai/benchmarks/performance/oh4t2g3h",
   "tg_tokens": null,
   "tps": 43.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-wpe5359g-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.3 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 595.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 595.4 68.7 69.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wpe5359g",
   "source_url": "https://omlx.ai/benchmarks/performance/wpe5359g",
   "tg_tokens": null,
   "tps": 68.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-wpe5359g-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.8 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 675.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 675.0 58.4 69.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wpe5359g",
   "source_url": "https://omlx.ai/benchmarks/performance/wpe5359g",
   "tg_tokens": null,
   "tps": 58.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-wpe5359g-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1089.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 1,089 65.0 70.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wpe5359g",
   "source_url": "https://omlx.ai/benchmarks/performance/wpe5359g",
   "tg_tokens": null,
   "tps": 65.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-wpe5359g-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1073.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k 1,073 56.7 71.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wpe5359g",
   "source_url": "https://omlx.ai/benchmarks/performance/wpe5359g",
   "tg_tokens": null,
   "tps": 56.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-wpe5359g-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.9 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1055.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 128k 1,055 56.7 74.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wpe5359g",
   "source_url": "https://omlx.ai/benchmarks/performance/wpe5359g",
   "tg_tokens": null,
   "tps": 56.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-tgoi5jeo-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 86.4 GB; TTFT 1866 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 548.8,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 1k tokens 548.8 PP tok/s 44.0 TG tok/s 1866 TTFT (ms) 86.4 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tgoi5jeo",
   "source_url": "https://omlx.ai/benchmarks/performance/tgoi5jeo",
   "tg_tokens": null,
   "tps": 44.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-l6iukz7y-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: of475b7u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 217.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 217.4 37.2 19.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l6iukz7y",
   "source_url": "https://omlx.ai/benchmarks/performance/l6iukz7y",
   "tg_tokens": null,
   "tps": 37.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-l6iukz7y-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: of475b7u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 223.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 223.1 34.5 21.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l6iukz7y",
   "source_url": "https://omlx.ai/benchmarks/performance/l6iukz7y",
   "tg_tokens": null,
   "tps": 34.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-mgcscn9t-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: 0iq4138s, 0jhxb2en, 1y48e04s, 87r92y95, bc2huefm, dc5l5ekr, kx8fhtm6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 331.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 331.5 41.2 33.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mgcscn9t",
   "source_url": "https://omlx.ai/benchmarks/performance/mgcscn9t",
   "tg_tokens": null,
   "tps": 41.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-mgcscn9t-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 37.5 GB at this context; duplicate traces of the same session: 0iq4138s, 0jhxb2en, 1y48e04s, 87r92y95, bc2huefm, dc5l5ekr, kx8fhtm6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 337.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 337.9 38.5 37.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mgcscn9t",
   "source_url": "https://omlx.ai/benchmarks/performance/mgcscn9t",
   "tg_tokens": null,
   "tps": 38.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-mgcscn9t-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 38.5 GB at this context; duplicate traces of the same session: 0iq4138s, 0jhxb2en, 1y48e04s, 87r92y95, bc2huefm, dc5l5ekr, kx8fhtm6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 333.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 8k 333.2 36.1 38.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mgcscn9t",
   "source_url": "https://omlx.ai/benchmarks/performance/mgcscn9t",
   "tg_tokens": null,
   "tps": 36.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-mgcscn9t-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 37.7 GB at this context; duplicate traces of the same session: 0iq4138s, 0jhxb2en, 1y48e04s, 87r92y95, bc2huefm, dc5l5ekr, kx8fhtm6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 321.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 16k 321.3 42.1 37.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mgcscn9t",
   "source_url": "https://omlx.ai/benchmarks/performance/mgcscn9t",
   "tg_tokens": null,
   "tps": 42.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-mgcscn9t-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 38.7 GB at this context; duplicate traces of the same session: 0iq4138s, 0jhxb2en, 1y48e04s, 87r92y95, bc2huefm, dc5l5ekr, kx8fhtm6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 298.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 32k 298.6 36.8 38.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mgcscn9t",
   "source_url": "https://omlx.ai/benchmarks/performance/mgcscn9t",
   "tg_tokens": null,
   "tps": 36.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-mgcscn9t-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 40.7 GB at this context; duplicate traces of the same session: 0iq4138s, 0jhxb2en, 1y48e04s, 87r92y95, bc2huefm, dc5l5ekr, kx8fhtm6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 260.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 64k 260.2 33.7 40.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mgcscn9t",
   "source_url": "https://omlx.ai/benchmarks/performance/mgcscn9t",
   "tg_tokens": null,
   "tps": 33.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-mgcscn9t-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 44.7 GB at this context; duplicate traces of the same session: 0iq4138s, 0jhxb2en, 1y48e04s, 87r92y95, bc2huefm, dc5l5ekr, kx8fhtm6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 128k 205.8 29.8 44.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mgcscn9t",
   "source_url": "https://omlx.ai/benchmarks/performance/mgcscn9t",
   "tg_tokens": null,
   "tps": 29.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-mgcscn9t-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 49.0 GB at this context; duplicate traces of the same session: 0iq4138s, 0jhxb2en, 1y48e04s, 87r92y95, bc2huefm, dc5l5ekr, kx8fhtm6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 168.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 195k 168.5 23.1 49.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mgcscn9t",
   "source_url": "https://omlx.ai/benchmarks/performance/mgcscn9t",
   "tg_tokens": null,
   "tps": 23.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yjikgfyc-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: cvyqm7xk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 859.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 859.7 34.1 33.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yjikgfyc",
   "source_url": "https://omlx.ai/benchmarks/performance/yjikgfyc",
   "tg_tokens": null,
   "tps": 34.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yjikgfyc-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: cvyqm7xk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 927.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 927.8 32.4 34.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yjikgfyc",
   "source_url": "https://omlx.ai/benchmarks/performance/yjikgfyc",
   "tg_tokens": null,
   "tps": 32.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3vovzfcy-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: 8b6j9tcb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 861.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 861.3 34.7 33.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3vovzfcy",
   "source_url": "https://omlx.ai/benchmarks/performance/3vovzfcy",
   "tg_tokens": null,
   "tps": 34.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3vovzfcy-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: 8b6j9tcb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 956.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 956.6 33.1 34.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3vovzfcy",
   "source_url": "https://omlx.ai/benchmarks/performance/3vovzfcy",
   "tg_tokens": null,
   "tps": 33.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-rx8zd7iz-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 3ur3rbw3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1348.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 1k 1,348 32.2 28.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rx8zd7iz",
   "source_url": "https://omlx.ai/benchmarks/performance/rx8zd7iz",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-rx8zd7iz-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: 3ur3rbw3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1738.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 4k 1,738 32.3 29.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rx8zd7iz",
   "source_url": "https://omlx.ai/benchmarks/performance/rx8zd7iz",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-snx6ay10-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: qlm7z5f3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 808.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 808.9 51.0 33.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/snx6ay10",
   "source_url": "https://omlx.ai/benchmarks/performance/snx6ay10",
   "tg_tokens": null,
   "tps": 51.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-snx6ay10-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: qlm7z5f3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 892.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 892.7 45.8 34.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/snx6ay10",
   "source_url": "https://omlx.ai/benchmarks/performance/snx6ay10",
   "tg_tokens": null,
   "tps": 45.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-dwbicd28-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 54pqo4un; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 276.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 276.0 7.4 19.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dwbicd28",
   "source_url": "https://omlx.ai/benchmarks/performance/dwbicd28",
   "tg_tokens": null,
   "tps": 7.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-dwbicd28-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: 54pqo4un; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 279.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 279.0 7.2 21.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dwbicd28",
   "source_url": "https://omlx.ai/benchmarks/performance/dwbicd28",
   "tg_tokens": null,
   "tps": 7.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Pro",
   "id": "omlx-j4c7sirt-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: g2133pbu; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 98.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 1k 98.7 16.7 18.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j4c7sirt",
   "source_url": "https://omlx.ai/benchmarks/performance/j4c7sirt",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Pro",
   "id": "omlx-j4c7sirt-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: g2133pbu; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 4k 103.3 18.3 19.7 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j4c7sirt",
   "source_url": "https://omlx.ai/benchmarks/performance/j4c7sirt",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vwayz759-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 149.8 GB at this context; duplicate traces of the same session: pak7i0y5, pkz2j8tr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 751.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-Flash-Next-MLX-6bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 751.7 59.4 149.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vwayz759",
   "source_url": "https://omlx.ai/benchmarks/performance/vwayz759",
   "tg_tokens": null,
   "tps": 59.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-vwayz759-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 151.7 GB at this context; duplicate traces of the same session: pak7i0y5, pkz2j8tr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 796.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-Flash-Next-MLX-6bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 796.4 50.1 151.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vwayz759",
   "source_url": "https://omlx.ai/benchmarks/performance/vwayz759",
   "tg_tokens": null,
   "tps": 50.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-vwayz759-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 152.7 GB at this context; duplicate traces of the same session: pak7i0y5, pkz2j8tr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 872.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-Flash-Next-MLX-6bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 8k 872.1 48.3 152.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vwayz759",
   "source_url": "https://omlx.ai/benchmarks/performance/vwayz759",
   "tg_tokens": null,
   "tps": 48.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-a233t0ll-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.3 GB at this context; duplicate traces of the same session: 57ot0iub, blroj2he, kvfvukea, tmp38hue, yv7ofbhx; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 625.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 625.9 67.6 69.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a233t0ll",
   "source_url": "https://omlx.ai/benchmarks/performance/a233t0ll",
   "tg_tokens": null,
   "tps": 67.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-a233t0ll-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.8 GB at this context; duplicate traces of the same session: 57ot0iub, blroj2he, kvfvukea, tmp38hue, yv7ofbhx; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 756.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 756.0 63.5 69.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a233t0ll",
   "source_url": "https://omlx.ai/benchmarks/performance/a233t0ll",
   "tg_tokens": null,
   "tps": 63.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-a233t0ll-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.9 GB at this context; duplicate traces of the same session: 57ot0iub, blroj2he, kvfvukea, tmp38hue, yv7ofbhx; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1101.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 1,101 64.9 69.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a233t0ll",
   "source_url": "https://omlx.ai/benchmarks/performance/a233t0ll",
   "tg_tokens": null,
   "tps": 64.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-a233t0ll-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: 57ot0iub, blroj2he, kvfvukea, tmp38hue, yv7ofbhx; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1090.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 1,090 67.4 70.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a233t0ll",
   "source_url": "https://omlx.ai/benchmarks/performance/a233t0ll",
   "tg_tokens": null,
   "tps": 67.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-a233t0ll-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: 57ot0iub, blroj2he, kvfvukea, tmp38hue, yv7ofbhx; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1074.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k 1,074 62.5 71.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a233t0ll",
   "source_url": "https://omlx.ai/benchmarks/performance/a233t0ll",
   "tg_tokens": null,
   "tps": 62.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-a233t0ll-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.9 GB at this context; duplicate traces of the same session: 57ot0iub, blroj2he, kvfvukea, tmp38hue, yv7ofbhx; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1056.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 128k 1,056 56.5 74.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a233t0ll",
   "source_url": "https://omlx.ai/benchmarks/performance/a233t0ll",
   "tg_tokens": null,
   "tps": 56.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-5moikv3m-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 86.4 GB at this context; duplicate traces of the same session: xxvb09z8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 508.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 508.4 34.2 86.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5moikv3m",
   "source_url": "https://omlx.ai/benchmarks/performance/5moikv3m",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-5moikv3m-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 87.0 GB at this context; duplicate traces of the same session: xxvb09z8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 650.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 650.7 61.4 87.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5moikv3m",
   "source_url": "https://omlx.ai/benchmarks/performance/5moikv3m",
   "tg_tokens": null,
   "tps": 61.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-5moikv3m-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 87.2 GB at this context; duplicate traces of the same session: xxvb09z8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 812.8,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 812.8 44.3 87.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5moikv3m",
   "source_url": "https://omlx.ai/benchmarks/performance/5moikv3m",
   "tg_tokens": null,
   "tps": 44.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-5moikv3m-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 89.7 GB at this context; duplicate traces of the same session: xxvb09z8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 793.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 793.2 56.3 89.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5moikv3m",
   "source_url": "https://omlx.ai/benchmarks/performance/5moikv3m",
   "tg_tokens": null,
   "tps": 56.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-5moikv3m-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 94.0 GB at this context; duplicate traces of the same session: xxvb09z8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 778.6,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 778.6 45.6 94.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5moikv3m",
   "source_url": "https://omlx.ai/benchmarks/performance/5moikv3m",
   "tg_tokens": null,
   "tps": 45.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8zbuh29r-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: 9b015g0g, bn8bne3p, dldhji1s, n0xgrcze; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 773.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 773.0 35.8 24.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8zbuh29r",
   "source_url": "https://omlx.ai/benchmarks/performance/8zbuh29r",
   "tg_tokens": null,
   "tps": 35.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8zbuh29r-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.7 GB at this context; duplicate traces of the same session: 9b015g0g, bn8bne3p, dldhji1s, n0xgrcze; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1035.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 1,035 56.2 25.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8zbuh29r",
   "source_url": "https://omlx.ai/benchmarks/performance/8zbuh29r",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8zbuh29r-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: 9b015g0g, bn8bne3p, dldhji1s, n0xgrcze; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 979.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 979.4 49.5 26.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8zbuh29r",
   "source_url": "https://omlx.ai/benchmarks/performance/8zbuh29r",
   "tg_tokens": null,
   "tps": 49.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8zbuh29r-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: 9b015g0g, bn8bne3p, dldhji1s, n0xgrcze; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 923.6,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 923.6 72.0 26.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8zbuh29r",
   "source_url": "https://omlx.ai/benchmarks/performance/8zbuh29r",
   "tg_tokens": null,
   "tps": 72.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8zbuh29r-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: 9b015g0g, bn8bne3p, dldhji1s, n0xgrcze; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 841.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 841.0 108.0 27.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8zbuh29r",
   "source_url": "https://omlx.ai/benchmarks/performance/8zbuh29r",
   "tg_tokens": null,
   "tps": 108.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-0rwd10ht-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "bf16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: 7hh6x2xk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 213.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-bf16-mtp-MLX on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 6bit; 1k 213.4 36.1 26.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0rwd10ht",
   "source_url": "https://omlx.ai/benchmarks/performance/0rwd10ht",
   "tg_tokens": null,
   "tps": 36.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-0rwd10ht-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "bf16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: 7hh6x2xk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 223.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-bf16-mtp-MLX on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 6bit; 4k 223.4 30.8 27.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0rwd10ht",
   "source_url": "https://omlx.ai/benchmarks/performance/0rwd10ht",
   "tg_tokens": null,
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-0qphj0pn-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "bf16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 213.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-bf16-mtp-MLX on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 6bit; 1k 213.1 39.4 26.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0qphj0pn",
   "source_url": "https://omlx.ai/benchmarks/performance/0qphj0pn",
   "tg_tokens": null,
   "tps": 39.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-0qphj0pn-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "bf16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-bf16-mtp-MLX on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 6bit; 4k 220.8 34.5 27.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0qphj0pn",
   "source_url": "https://omlx.ai/benchmarks/performance/0qphj0pn",
   "tg_tokens": null,
   "tps": 34.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-jh2924yn-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 217.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 217.6 35.3 19.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jh2924yn",
   "source_url": "https://omlx.ai/benchmarks/performance/jh2924yn",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-jh2924yn-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 211.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 211.4 36.0 21.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jh2924yn",
   "source_url": "https://omlx.ai/benchmarks/performance/jh2924yn",
   "tg_tokens": null,
   "tps": 36.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-1ypryntx-ctx1024",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 10.2 GB; TTFT 3783 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 270.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "gemma-4-12B-it-MLX-6bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 6bit; 1k tokens 270.7 PP tok/s 22.7 TG tok/s 3783 TTFT (ms) 10.2 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ypryntx",
   "source_url": "https://omlx.ai/benchmarks/performance/1ypryntx",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ga25gwnu-ctx1024",
   "model": "gpt-oss 120B",
   "notes": "oMLX community benchmark trace; peak memory 116.1 GB at this context; duplicate traces of the same session: emp6bfog, iv0nljia, ootfqk7r, wdegjmse; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 876.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gpt-oss-120b-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 1k 876.2 74.3 116.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ga25gwnu",
   "source_url": "https://omlx.ai/benchmarks/performance/ga25gwnu",
   "tg_tokens": null,
   "tps": 74.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ga25gwnu-ctx4096",
   "model": "gpt-oss 120B",
   "notes": "oMLX community benchmark trace; peak memory 116.2 GB at this context; duplicate traces of the same session: emp6bfog, iv0nljia, ootfqk7r, wdegjmse; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1072.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gpt-oss-120b-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 4k 1,072 72.5 116.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ga25gwnu",
   "source_url": "https://omlx.ai/benchmarks/performance/ga25gwnu",
   "tg_tokens": null,
   "tps": 72.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ga25gwnu-ctx8192",
   "model": "gpt-oss 120B",
   "notes": "oMLX community benchmark trace; peak memory 116.3 GB at this context; duplicate traces of the same session: emp6bfog, iv0nljia, ootfqk7r, wdegjmse; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1068.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gpt-oss-120b-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 8k 1,068 69.6 116.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ga25gwnu",
   "source_url": "https://omlx.ai/benchmarks/performance/ga25gwnu",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ga25gwnu-ctx16384",
   "model": "gpt-oss 120B",
   "notes": "oMLX community benchmark trace; peak memory 116.6 GB at this context; duplicate traces of the same session: emp6bfog, iv0nljia, ootfqk7r, wdegjmse; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1011.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gpt-oss-120b-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 16k 1,011 65.5 116.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ga25gwnu",
   "source_url": "https://omlx.ai/benchmarks/performance/ga25gwnu",
   "tg_tokens": null,
   "tps": 65.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ga25gwnu-ctx32768",
   "model": "gpt-oss 120B",
   "notes": "oMLX community benchmark trace; peak memory 117.2 GB at this context; duplicate traces of the same session: emp6bfog, iv0nljia, ootfqk7r, wdegjmse; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 915.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gpt-oss-120b-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 32k 915.4 59.0 117.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ga25gwnu",
   "source_url": "https://omlx.ai/benchmarks/performance/ga25gwnu",
   "tg_tokens": null,
   "tps": 59.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-m5gz1vsy-ctx1024",
   "model": "Qwen2.5-14B",
   "notes": "Instruct; oMLX community benchmark trace; peak memory 8.4 GB at this context; duplicate traces of the same session: fu1phif8; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 123.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-14B-Instruct-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 123.0 12.2 8.4 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m5gz1vsy",
   "source_url": "https://omlx.ai/benchmarks/performance/m5gz1vsy",
   "tg_tokens": null,
   "tps": 12.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-m5gz1vsy-ctx4096",
   "model": "Qwen2.5-14B",
   "notes": "Instruct; oMLX community benchmark trace; peak memory 9.0 GB at this context; duplicate traces of the same session: fu1phif8; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 118.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-14B-Instruct-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 118.5 11.5 9.0 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m5gz1vsy",
   "source_url": "https://omlx.ai/benchmarks/performance/m5gz1vsy",
   "tg_tokens": null,
   "tps": 11.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-kf4v556j-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 38.2 GB at this context; duplicate traces of the same session: 3kf0kvrh, d0s1xjks, dboxugvu, e9gynkxu, eparwun8, lvlxucik, qosqs2tp; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2090.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 2,090 183.2 38.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kf4v556j",
   "source_url": "https://omlx.ai/benchmarks/performance/kf4v556j",
   "tg_tokens": null,
   "tps": 183.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-kf4v556j-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 40.3 GB at this context; duplicate traces of the same session: 3kf0kvrh, d0s1xjks, dboxugvu, e9gynkxu, eparwun8, lvlxucik, qosqs2tp; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2450.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 2,450 151.7 40.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kf4v556j",
   "source_url": "https://omlx.ai/benchmarks/performance/kf4v556j",
   "tg_tokens": null,
   "tps": 151.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-kf4v556j-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 41.0 GB at this context; duplicate traces of the same session: 3kf0kvrh, d0s1xjks, dboxugvu, e9gynkxu, eparwun8, lvlxucik, qosqs2tp; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2343.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 8k 2,343 144.9 41.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kf4v556j",
   "source_url": "https://omlx.ai/benchmarks/performance/kf4v556j",
   "tg_tokens": null,
   "tps": 144.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-kf4v556j-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 40.4 GB at this context; duplicate traces of the same session: 3kf0kvrh, d0s1xjks, dboxugvu, e9gynkxu, eparwun8, lvlxucik, qosqs2tp; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2097.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 16k 2,097 170.4 40.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kf4v556j",
   "source_url": "https://omlx.ai/benchmarks/performance/kf4v556j",
   "tg_tokens": null,
   "tps": 170.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-kf4v556j-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 40.7 GB at this context; duplicate traces of the same session: 3kf0kvrh, d0s1xjks, dboxugvu, e9gynkxu, eparwun8, lvlxucik, qosqs2tp; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1706.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 32k 1,706 152.8 40.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kf4v556j",
   "source_url": "https://omlx.ai/benchmarks/performance/kf4v556j",
   "tg_tokens": null,
   "tps": 152.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-kf4v556j-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 41.4 GB at this context; duplicate traces of the same session: 3kf0kvrh, d0s1xjks, dboxugvu, e9gynkxu, eparwun8, lvlxucik, qosqs2tp; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1236.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 64k 1,236 123.7 41.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kf4v556j",
   "source_url": "https://omlx.ai/benchmarks/performance/kf4v556j",
   "tg_tokens": null,
   "tps": 123.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-kf4v556j-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 43.0 GB at this context; duplicate traces of the same session: 3kf0kvrh, d0s1xjks, dboxugvu, e9gynkxu, eparwun8, lvlxucik, qosqs2tp; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 792.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 128k 792.2 108.1 43.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kf4v556j",
   "source_url": "https://omlx.ai/benchmarks/performance/kf4v556j",
   "tg_tokens": null,
   "tps": 108.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-kf4v556j-ctx199680",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 44.8 GB at this context; duplicate traces of the same session: 3kf0kvrh, d0s1xjks, dboxugvu, e9gynkxu, eparwun8, lvlxucik, qosqs2tp; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 573.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 195k 573.9 73.3 44.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kf4v556j",
   "source_url": "https://omlx.ai/benchmarks/performance/kf4v556j",
   "tg_tokens": null,
   "tps": 73.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-b2w9qx04-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: or5tntfg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 856.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 856.2 33.9 33.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b2w9qx04",
   "source_url": "https://omlx.ai/benchmarks/performance/b2w9qx04",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-b2w9qx04-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: or5tntfg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 892.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 892.9 34.2 34.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b2w9qx04",
   "source_url": "https://omlx.ai/benchmarks/performance/b2w9qx04",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nbuvx4se-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 42.6 GB; TTFT 266225 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 492.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k tokens 492.3 PP tok/s 29.9 TG tok/s 266225 TTFT (ms) 42.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nbuvx4se",
   "source_url": "https://omlx.ai/benchmarks/performance/nbuvx4se",
   "tg_tokens": null,
   "tps": 29.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lh87vl6t-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: 3gikmpdn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 777.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 777.0 38.0 33.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lh87vl6t",
   "source_url": "https://omlx.ai/benchmarks/performance/lh87vl6t",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lh87vl6t-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: 3gikmpdn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 727.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 727.0 33.6 34.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lh87vl6t",
   "source_url": "https://omlx.ai/benchmarks/performance/lh87vl6t",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0ywvkk9t-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 806.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 806.9 51.1 33.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ywvkk9t",
   "source_url": "https://omlx.ai/benchmarks/performance/0ywvkk9t",
   "tg_tokens": null,
   "tps": 51.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0ywvkk9t-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 907.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 907.1 47.1 34.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ywvkk9t",
   "source_url": "https://omlx.ai/benchmarks/performance/0ywvkk9t",
   "tg_tokens": null,
   "tps": 47.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-t0na23ap-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 55.3 GB at this context; duplicate traces of the same session: lgft5euv; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 338.4,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.8-27B-bf16 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 bf16; 4k 338.4 13.3 55.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t0na23ap",
   "source_url": "https://omlx.ai/benchmarks/performance/t0na23ap",
   "tg_tokens": null,
   "tps": 13.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-t0na23ap-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 56.4 GB at this context; duplicate traces of the same session: lgft5euv; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.2,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.8-27B-bf16 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 bf16; 16k 323.2 12.9 56.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t0na23ap",
   "source_url": "https://omlx.ai/benchmarks/performance/t0na23ap",
   "tg_tokens": null,
   "tps": 12.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-8y7ivxwy-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.2 GB at this context; duplicate traces of the same session: 4z93tv5c, 5gh82pd3, 8olnp8ky, kp4k5m0l, qko88ydh, ui4z99s5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 248.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 248.9 35.5 33.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8y7ivxwy",
   "source_url": "https://omlx.ai/benchmarks/performance/8y7ivxwy",
   "tg_tokens": null,
   "tps": 35.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-8y7ivxwy-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.4 GB at this context; duplicate traces of the same session: 4z93tv5c, 5gh82pd3, 8olnp8ky, kp4k5m0l, qko88ydh, ui4z99s5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 332.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 332.8 35.1 36.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8y7ivxwy",
   "source_url": "https://omlx.ai/benchmarks/performance/8y7ivxwy",
   "tg_tokens": null,
   "tps": 35.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-8y7ivxwy-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: 4z93tv5c, 5gh82pd3, 8olnp8ky, kp4k5m0l, qko88ydh, ui4z99s5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 337.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 337.2 44.6 36.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8y7ivxwy",
   "source_url": "https://omlx.ai/benchmarks/performance/8y7ivxwy",
   "tg_tokens": null,
   "tps": 44.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-8y7ivxwy-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.0 GB at this context; duplicate traces of the same session: 4z93tv5c, 5gh82pd3, 8olnp8ky, kp4k5m0l, qko88ydh, ui4z99s5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 321.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 321.2 44.6 38.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8y7ivxwy",
   "source_url": "https://omlx.ai/benchmarks/performance/8y7ivxwy",
   "tg_tokens": null,
   "tps": 44.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-8y7ivxwy-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.7 GB at this context; duplicate traces of the same session: 4z93tv5c, 5gh82pd3, 8olnp8ky, kp4k5m0l, qko88ydh, ui4z99s5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 291.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 291.8 61.0 39.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8y7ivxwy",
   "source_url": "https://omlx.ai/benchmarks/performance/8y7ivxwy",
   "tg_tokens": null,
   "tps": 61.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-8y7ivxwy-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 42.4 GB at this context; duplicate traces of the same session: 4z93tv5c, 5gh82pd3, 8olnp8ky, kp4k5m0l, qko88ydh, ui4z99s5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 246.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k 246.8 34.5 42.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8y7ivxwy",
   "source_url": "https://omlx.ai/benchmarks/performance/8y7ivxwy",
   "tg_tokens": null,
   "tps": 34.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-8y7ivxwy-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 46.9 GB at this context; duplicate traces of the same session: 4z93tv5c, 5gh82pd3, 8olnp8ky, kp4k5m0l, qko88ydh, ui4z99s5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 189.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 128k 189.3 32.1 46.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8y7ivxwy",
   "source_url": "https://omlx.ai/benchmarks/performance/8y7ivxwy",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-wh1b4mth-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; coder 8bit ngram build; oMLX community benchmark trace; peak memory 72.2 GB at this context; duplicate traces of the same session: yx8423le; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 551.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-coder-8bit-ngram-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 551.3 39.4 72.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wh1b4mth",
   "source_url": "https://omlx.ai/benchmarks/performance/wh1b4mth",
   "tg_tokens": null,
   "tps": 39.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-wh1b4mth-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; coder 8bit ngram build; oMLX community benchmark trace; peak memory 78.6 GB at this context; duplicate traces of the same session: yx8423le; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 584.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-coder-8bit-ngram-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 584.6 36.6 78.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wh1b4mth",
   "source_url": "https://omlx.ai/benchmarks/performance/wh1b4mth",
   "tg_tokens": null,
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-f1ry90fg-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 86.4 GB at this context; duplicate traces of the same session: 1hhwpsjw, xqkdtmsm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 508.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 508.4 34.2 86.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f1ry90fg",
   "source_url": "https://omlx.ai/benchmarks/performance/f1ry90fg",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-f1ry90fg-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 87.0 GB at this context; duplicate traces of the same session: 1hhwpsjw, xqkdtmsm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 650.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 650.7 61.4 87.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f1ry90fg",
   "source_url": "https://omlx.ai/benchmarks/performance/f1ry90fg",
   "tg_tokens": null,
   "tps": 61.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-f1ry90fg-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 87.2 GB at this context; duplicate traces of the same session: 1hhwpsjw, xqkdtmsm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 812.8,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 812.8 44.3 87.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f1ry90fg",
   "source_url": "https://omlx.ai/benchmarks/performance/f1ry90fg",
   "tg_tokens": null,
   "tps": 44.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-f1ry90fg-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 89.7 GB at this context; duplicate traces of the same session: 1hhwpsjw, xqkdtmsm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 793.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 793.2 56.3 89.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f1ry90fg",
   "source_url": "https://omlx.ai/benchmarks/performance/f1ry90fg",
   "tg_tokens": null,
   "tps": 56.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-f1ry90fg-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 94.0 GB at this context; duplicate traces of the same session: 1hhwpsjw, xqkdtmsm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 778.6,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 778.6 45.6 94.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f1ry90fg",
   "source_url": "https://omlx.ai/benchmarks/performance/f1ry90fg",
   "tg_tokens": null,
   "tps": 45.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-25134dmd-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "bf16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 213.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-bf16-mtp-MLX on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 6bit; 1k 213.1 39.4 26.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/25134dmd",
   "source_url": "https://omlx.ai/benchmarks/performance/25134dmd",
   "tg_tokens": null,
   "tps": 39.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-25134dmd-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "bf16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-bf16-mtp-MLX on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 6bit; 4k 220.8 34.5 27.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/25134dmd",
   "source_url": "https://omlx.ai/benchmarks/performance/25134dmd",
   "tg_tokens": null,
   "tps": 34.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-d3rf5xwb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: lbuzlgx8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 216.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 216.5 33.5 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d3rf5xwb",
   "source_url": "https://omlx.ai/benchmarks/performance/d3rf5xwb",
   "tg_tokens": null,
   "tps": 33.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-d3rf5xwb-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: lbuzlgx8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 212.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 212.1 32.8 21.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d3rf5xwb",
   "source_url": "https://omlx.ai/benchmarks/performance/d3rf5xwb",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-dzgy3niv-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 217.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 217.6 35.3 19.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dzgy3niv",
   "source_url": "https://omlx.ai/benchmarks/performance/dzgy3niv",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-dzgy3niv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 211.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 211.4 36.0 21.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dzgy3niv",
   "source_url": "https://omlx.ai/benchmarks/performance/dzgy3niv",
   "tg_tokens": null,
   "tps": 36.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-92ubx5bu-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: zbb5tgz0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 215.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 215.7 35.5 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/92ubx5bu",
   "source_url": "https://omlx.ai/benchmarks/performance/92ubx5bu",
   "tg_tokens": null,
   "tps": 35.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-92ubx5bu-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: zbb5tgz0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 196.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 196.8 37.3 21.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/92ubx5bu",
   "source_url": "https://omlx.ai/benchmarks/performance/92ubx5bu",
   "tg_tokens": null,
   "tps": 37.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-ywkuqy40-ctx1024",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: kcytnhju; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 85.4,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 2bit; 1k 85.4 11.0 11.8 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ywkuqy40",
   "source_url": "https://omlx.ai/benchmarks/performance/ywkuqy40",
   "tg_tokens": null,
   "tps": 11.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-ywkuqy40-ctx4096",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 14.5 GB at this context; duplicate traces of the same session: kcytnhju; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 86.0,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 2bit; 4k 86.0 11.0 14.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ywkuqy40",
   "source_url": "https://omlx.ai/benchmarks/performance/ywkuqy40",
   "tg_tokens": null,
   "tps": 11.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Pro",
   "id": "omlx-i1cic41o-ctx1024",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 12.1 GB at this context; duplicate traces of the same session: lykb5nxq; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 86.7,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 2bit; 1k 86.7 7.0 12.1 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i1cic41o",
   "source_url": "https://omlx.ai/benchmarks/performance/i1cic41o",
   "tg_tokens": null,
   "tps": 7.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Pro",
   "id": "omlx-i1cic41o-ctx4096",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 14.7 GB at this context; duplicate traces of the same session: lykb5nxq; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 86.2,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 2bit; 4k 86.2 6.3 14.7 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i1cic41o",
   "source_url": "https://omlx.ai/benchmarks/performance/i1cic41o",
   "tg_tokens": null,
   "tps": 6.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-s54so381-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: nil1vnp5; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1204.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 1,204 105.4 16.3 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s54so381",
   "source_url": "https://omlx.ai/benchmarks/performance/s54so381",
   "tg_tokens": null,
   "tps": 105.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-s54so381-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: nil1vnp5; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1221.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 1,221 75.5 16.6 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s54so381",
   "source_url": "https://omlx.ai/benchmarks/performance/s54so381",
   "tg_tokens": null,
   "tps": 75.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-moirf6re-ctx1024",
   "model": "Gemma 4 26B-A4B",
   "notes": "qmax; oMLX community benchmark trace; peak memory 14.9 GB at this context; duplicate traces of the same session: xs9vhnzl; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1243.0,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "gemma-4-26b-a4b-nvfp4-qmax-mlx on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 1,243 80.3 14.9 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/moirf6re",
   "source_url": "https://omlx.ai/benchmarks/performance/moirf6re",
   "tg_tokens": null,
   "tps": 80.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-moirf6re-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "qmax; oMLX community benchmark trace; peak memory 15.2 GB at this context; duplicate traces of the same session: xs9vhnzl; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1214.0,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "gemma-4-26b-a4b-nvfp4-qmax-mlx on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 1,214 78.3 15.2 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/moirf6re",
   "source_url": "https://omlx.ai/benchmarks/performance/moirf6re",
   "tg_tokens": null,
   "tps": 78.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0ecxe5p2-ctx65536",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: 4nyxl5je; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3252.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "gpt-oss-20b-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 3,252 59.4 14.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ecxe5p2",
   "source_url": "https://omlx.ai/benchmarks/performance/0ecxe5p2",
   "tg_tokens": null,
   "tps": 59.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0ecxe5p2-ctx131072",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 16.2 GB at this context; duplicate traces of the same session: 4nyxl5je; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2302.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "gpt-oss-20b-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,302 38.6 16.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ecxe5p2",
   "source_url": "https://omlx.ai/benchmarks/performance/0ecxe5p2",
   "tg_tokens": null,
   "tps": 38.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-pqjmtmht-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 403.7 GB at this context; duplicate traces of the same session: jy4hm949; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 609.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 609.6 38.2 403.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pqjmtmht",
   "source_url": "https://omlx.ai/benchmarks/performance/pqjmtmht",
   "tg_tokens": null,
   "tps": 38.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-pqjmtmht-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 403.7 GB at this context; duplicate traces of the same session: jy4hm949; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 739.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 16k 739.9 34.9 403.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pqjmtmht",
   "source_url": "https://omlx.ai/benchmarks/performance/pqjmtmht",
   "tg_tokens": null,
   "tps": 34.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-f2b8xom8-ctx1024",
   "model": "Qwable-v2",
   "notes": "DWQ; MTP speculative decoding enabled; Vision; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: fzjpsn2f; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 896.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwable-v2-oQ4e-DWQ-MTP-Vision-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 896.4 78.4 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f2b8xom8",
   "source_url": "https://omlx.ai/benchmarks/performance/f2b8xom8",
   "tg_tokens": null,
   "tps": 78.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-f2b8xom8-ctx4096",
   "model": "Qwable-v2",
   "notes": "DWQ; MTP speculative decoding enabled; Vision; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: fzjpsn2f; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1022.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwable-v2-oQ4e-DWQ-MTP-Vision-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 1,022 76.0 20.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f2b8xom8",
   "source_url": "https://omlx.ai/benchmarks/performance/f2b8xom8",
   "tg_tokens": null,
   "tps": 76.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-q2rhtayw-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: 9ma834z9; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 293.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M3 Pro (14c); M3 Pro (14c) \u00b7 18 GB \u00b7 4bit; 4k 293.2 27.0 6.4 GB view",
   "ram_gb": 18,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q2rhtayw",
   "source_url": "https://omlx.ai/benchmarks/performance/q2rhtayw",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-q2rhtayw-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: 9ma834z9; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 278.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M3 Pro (14c); M3 Pro (14c) \u00b7 18 GB \u00b7 4bit; 16k 278.2 24.1 6.9 GB current",
   "ram_gb": 18,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q2rhtayw",
   "source_url": "https://omlx.ai/benchmarks/performance/q2rhtayw",
   "tg_tokens": null,
   "tps": 24.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9zr3ddxu-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 38.0 GB at this context; duplicate traces of the same session: 0rdbavdr, 1jkeivgs, 36ex4dyu, 90gzj0sx, si58h41f, teo65s9i, yxpvitqw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2158.0,
   "provenance": "community",
   "quant": "mixed 9bit",
   "quote": "Qwen3.6-35B-A3B-MLX-mixed-9bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 2,158 93.9 38.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9zr3ddxu",
   "source_url": "https://omlx.ai/benchmarks/performance/9zr3ddxu",
   "tg_tokens": null,
   "tps": 93.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9zr3ddxu-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 40.0 GB at this context; duplicate traces of the same session: 0rdbavdr, 1jkeivgs, 36ex4dyu, 90gzj0sx, si58h41f, teo65s9i, yxpvitqw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2532.0,
   "provenance": "community",
   "quant": "mixed 9bit",
   "quote": "Qwen3.6-35B-A3B-MLX-mixed-9bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 2,532 92.6 40.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9zr3ddxu",
   "source_url": "https://omlx.ai/benchmarks/performance/9zr3ddxu",
   "tg_tokens": null,
   "tps": 92.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9zr3ddxu-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 40.6 GB at this context; duplicate traces of the same session: 0rdbavdr, 1jkeivgs, 36ex4dyu, 90gzj0sx, si58h41f, teo65s9i, yxpvitqw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2431.0,
   "provenance": "community",
   "quant": "mixed 9bit",
   "quote": "Qwen3.6-35B-A3B-MLX-mixed-9bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 8k 2,431 92.2 40.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9zr3ddxu",
   "source_url": "https://omlx.ai/benchmarks/performance/9zr3ddxu",
   "tg_tokens": null,
   "tps": 92.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9zr3ddxu-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 40.0 GB at this context; duplicate traces of the same session: 0rdbavdr, 1jkeivgs, 36ex4dyu, 90gzj0sx, si58h41f, teo65s9i, yxpvitqw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2192.0,
   "provenance": "community",
   "quant": "mixed 9bit",
   "quote": "Qwen3.6-35B-A3B-MLX-mixed-9bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 16k 2,192 85.3 40.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9zr3ddxu",
   "source_url": "https://omlx.ai/benchmarks/performance/9zr3ddxu",
   "tg_tokens": null,
   "tps": 85.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9zr3ddxu-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 40.3 GB at this context; duplicate traces of the same session: 0rdbavdr, 1jkeivgs, 36ex4dyu, 90gzj0sx, si58h41f, teo65s9i, yxpvitqw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1805.0,
   "provenance": "community",
   "quant": "mixed 9bit",
   "quote": "Qwen3.6-35B-A3B-MLX-mixed-9bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 32k 1,805 79.2 40.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9zr3ddxu",
   "source_url": "https://omlx.ai/benchmarks/performance/9zr3ddxu",
   "tg_tokens": null,
   "tps": 79.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9zr3ddxu-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 40.9 GB at this context; duplicate traces of the same session: 0rdbavdr, 1jkeivgs, 36ex4dyu, 90gzj0sx, si58h41f, teo65s9i, yxpvitqw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1324.0,
   "provenance": "community",
   "quant": "mixed 9bit",
   "quote": "Qwen3.6-35B-A3B-MLX-mixed-9bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 64k 1,324 68.6 40.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9zr3ddxu",
   "source_url": "https://omlx.ai/benchmarks/performance/9zr3ddxu",
   "tg_tokens": null,
   "tps": 68.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9zr3ddxu-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 42.2 GB at this context; duplicate traces of the same session: 0rdbavdr, 1jkeivgs, 36ex4dyu, 90gzj0sx, si58h41f, teo65s9i, yxpvitqw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 859.5,
   "provenance": "community",
   "quant": "mixed 9bit",
   "quote": "Qwen3.6-35B-A3B-MLX-mixed-9bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 128k 859.5 55.5 42.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9zr3ddxu",
   "source_url": "https://omlx.ai/benchmarks/performance/9zr3ddxu",
   "tg_tokens": null,
   "tps": 55.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9zr3ddxu-ctx199680",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 43.5 GB at this context; duplicate traces of the same session: 0rdbavdr, 1jkeivgs, 36ex4dyu, 90gzj0sx, si58h41f, teo65s9i, yxpvitqw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 627.0,
   "provenance": "community",
   "quant": "mixed 9bit",
   "quote": "Qwen3.6-35B-A3B-MLX-mixed-9bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 195k 627.0 46.3 43.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9zr3ddxu",
   "source_url": "https://omlx.ai/benchmarks/performance/9zr3ddxu",
   "tg_tokens": null,
   "tps": 46.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0eruj18v-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: z0hh8p1j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 862.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 862.8 41.2 33.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0eruj18v",
   "source_url": "https://omlx.ai/benchmarks/performance/0eruj18v",
   "tg_tokens": null,
   "tps": 41.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0eruj18v-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: z0hh8p1j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 925.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 925.1 32.3 34.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0eruj18v",
   "source_url": "https://omlx.ai/benchmarks/performance/0eruj18v",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5g3iox82-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: hsidu6zl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 861.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 861.1 35.4 33.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5g3iox82",
   "source_url": "https://omlx.ai/benchmarks/performance/5g3iox82",
   "tg_tokens": null,
   "tps": 35.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5g3iox82-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: hsidu6zl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 904.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 904.9 31.7 34.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5g3iox82",
   "source_url": "https://omlx.ai/benchmarks/performance/5g3iox82",
   "tg_tokens": null,
   "tps": 31.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n6013x5s-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: gswfibfm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 859.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 859.1 42.2 33.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n6013x5s",
   "source_url": "https://omlx.ai/benchmarks/performance/n6013x5s",
   "tg_tokens": null,
   "tps": 42.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n6013x5s-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: gswfibfm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 914.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 914.6 34.0 34.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n6013x5s",
   "source_url": "https://omlx.ai/benchmarks/performance/n6013x5s",
   "tg_tokens": null,
   "tps": 34.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-v1xnwkqr-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: z8zas8tg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 861.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 861.6 39.5 33.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v1xnwkqr",
   "source_url": "https://omlx.ai/benchmarks/performance/v1xnwkqr",
   "tg_tokens": null,
   "tps": 39.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-v1xnwkqr-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: z8zas8tg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 938.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 938.1 34.6 34.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v1xnwkqr",
   "source_url": "https://omlx.ai/benchmarks/performance/v1xnwkqr",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-01rmtc76-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 806.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 806.9 51.1 33.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/01rmtc76",
   "source_url": "https://omlx.ai/benchmarks/performance/01rmtc76",
   "tg_tokens": null,
   "tps": 51.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-01rmtc76-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 907.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 907.1 47.1 34.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/01rmtc76",
   "source_url": "https://omlx.ai/benchmarks/performance/01rmtc76",
   "tg_tokens": null,
   "tps": 47.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-w06guq0g-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: zwecy698; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 839.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 839.7 59.2 23.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w06guq0g",
   "source_url": "https://omlx.ai/benchmarks/performance/w06guq0g",
   "tg_tokens": null,
   "tps": 59.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-w06guq0g-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: zwecy698; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 885.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 885.4 51.8 24.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w06guq0g",
   "source_url": "https://omlx.ai/benchmarks/performance/w06guq0g",
   "tg_tokens": null,
   "tps": 51.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mjbwvuah-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: m8agodkj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 848.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 848.0 49.4 23.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mjbwvuah",
   "source_url": "https://omlx.ai/benchmarks/performance/mjbwvuah",
   "tg_tokens": null,
   "tps": 49.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mjbwvuah-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: m8agodkj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 949.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 949.5 51.1 24.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mjbwvuah",
   "source_url": "https://omlx.ai/benchmarks/performance/mjbwvuah",
   "tg_tokens": null,
   "tps": 51.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ktnt66ck-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: rz603xdd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 854.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 854.0 26.8 20.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ktnt66ck",
   "source_url": "https://omlx.ai/benchmarks/performance/ktnt66ck",
   "tg_tokens": null,
   "tps": 26.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ktnt66ck-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: rz603xdd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1005.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,005 26.3 21.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ktnt66ck",
   "source_url": "https://omlx.ai/benchmarks/performance/ktnt66ck",
   "tg_tokens": null,
   "tps": 26.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-u8sjnj2x-ctx8192",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "vision; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: 6me0ylv1, 6qmtprf6, 6y1xd89q, uv8z0lrj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 860.5,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ5e-vision on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 5bit; 8k 860.5 58.0 24.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u8sjnj2x",
   "source_url": "https://omlx.ai/benchmarks/performance/u8sjnj2x",
   "tg_tokens": null,
   "tps": 58.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-u8sjnj2x-ctx16384",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "vision; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: 6me0ylv1, 6qmtprf6, 6y1xd89q, uv8z0lrj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 777.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ5e-vision on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 5bit; 16k 777.2 54.3 24.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u8sjnj2x",
   "source_url": "https://omlx.ai/benchmarks/performance/u8sjnj2x",
   "tg_tokens": null,
   "tps": 54.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-u8sjnj2x-ctx32768",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "vision; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: 6me0ylv1, 6qmtprf6, 6y1xd89q, uv8z0lrj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 626.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ5e-vision on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 5bit; 32k 626.2 48.8 25.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u8sjnj2x",
   "source_url": "https://omlx.ai/benchmarks/performance/u8sjnj2x",
   "tg_tokens": null,
   "tps": 48.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-u8sjnj2x-ctx65536",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "vision; oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: 6me0ylv1, 6qmtprf6, 6y1xd89q, uv8z0lrj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 481.6,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ5e-vision on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 5bit; 64k 481.6 39.8 26.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u8sjnj2x",
   "source_url": "https://omlx.ai/benchmarks/performance/u8sjnj2x",
   "tg_tokens": null,
   "tps": 39.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-u8sjnj2x-ctx131072",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "vision; oMLX community benchmark trace; peak memory 28.3 GB at this context; duplicate traces of the same session: 6me0ylv1, 6qmtprf6, 6y1xd89q, uv8z0lrj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 329.3,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ5e-vision on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 5bit; 128k 329.3 30.4 28.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u8sjnj2x",
   "source_url": "https://omlx.ai/benchmarks/performance/u8sjnj2x",
   "tg_tokens": null,
   "tps": 30.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qygof4a3-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.7 GB; TTFT 1766 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 580.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k tokens 580.0 PP tok/s 56.2 TG tok/s 1766 TTFT (ms) 70.7 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qygof4a3",
   "source_url": "https://omlx.ai/benchmarks/performance/qygof4a3",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-c17v8wt4-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 71.1 GB; TTFT 2444 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1676.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 1,676 PP tok/s 79.8 TG tok/s 2444 TTFT (ms) 71.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c17v8wt4",
   "source_url": "https://omlx.ai/benchmarks/performance/c17v8wt4",
   "tg_tokens": null,
   "tps": 79.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-07jmmyai-ctx1024",
   "model": "Selene-1-Mini-Llama-3.1-8B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 4.9 GB; TTFT 1308 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 783.0,
   "provenance": "community",
   "quant": "q4",
   "quote": "Selene-1-Mini-Llama-3.1-8B-q4-mlx on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 783.0 PP tok/s 27.7 TG tok/s 1308 TTFT (ms) 4.9 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/07jmmyai",
   "source_url": "https://omlx.ai/benchmarks/performance/07jmmyai",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dq4sczwr-ctx65536",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.5 GB at this context; duplicate traces of the same session: wg5k6w0w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 411.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 411.6 28.3 38.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dq4sczwr",
   "source_url": "https://omlx.ai/benchmarks/performance/dq4sczwr",
   "tg_tokens": null,
   "tps": 28.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dq4sczwr-ctx131072",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 42.3 GB at this context; duplicate traces of the same session: wg5k6w0w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 322.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 322.8 24.5 42.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dq4sczwr",
   "source_url": "https://omlx.ai/benchmarks/performance/dq4sczwr",
   "tg_tokens": null,
   "tps": 24.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-r2sw400r-ctx1024",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: q7pd11vx; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 86.6,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 2bit; 1k 86.6 11.3 11.8 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r2sw400r",
   "source_url": "https://omlx.ai/benchmarks/performance/r2sw400r",
   "tg_tokens": null,
   "tps": 11.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-r2sw400r-ctx4096",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 14.5 GB at this context; duplicate traces of the same session: q7pd11vx; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 84.5,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 2bit; 4k 84.5 10.6 14.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r2sw400r",
   "source_url": "https://omlx.ai/benchmarks/performance/r2sw400r",
   "tg_tokens": null,
   "tps": 10.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-dusin764-ctx1024",
   "model": "Ternary-Bonsai-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 8.9 GB at this context; duplicate traces of the same session: aucfn9b2, eyv2jj31, jo3d3h85, ob260k8v; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 487.1,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-27B-mlx-2bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 2bit; 1k 487.1 31.3 8.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dusin764",
   "source_url": "https://omlx.ai/benchmarks/performance/dusin764",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-dusin764-ctx4096",
   "model": "Ternary-Bonsai-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 10.3 GB at this context; duplicate traces of the same session: aucfn9b2, eyv2jj31, jo3d3h85, ob260k8v; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 442.7,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-27B-mlx-2bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 2bit; 4k 442.7 30.4 10.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dusin764",
   "source_url": "https://omlx.ai/benchmarks/performance/dusin764",
   "tg_tokens": null,
   "tps": 30.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-dusin764-ctx8192",
   "model": "Ternary-Bonsai-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 10.9 GB at this context; duplicate traces of the same session: aucfn9b2, eyv2jj31, jo3d3h85, ob260k8v; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 385.6,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-27B-mlx-2bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 2bit; 8k 385.6 28.9 10.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dusin764",
   "source_url": "https://omlx.ai/benchmarks/performance/dusin764",
   "tg_tokens": null,
   "tps": 28.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-dusin764-ctx16384",
   "model": "Ternary-Bonsai-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 12.2 GB at this context; duplicate traces of the same session: aucfn9b2, eyv2jj31, jo3d3h85, ob260k8v; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 384.1,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-27B-mlx-2bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 2bit; 16k 384.1 26.5 12.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dusin764",
   "source_url": "https://omlx.ai/benchmarks/performance/dusin764",
   "tg_tokens": null,
   "tps": 26.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-dusin764-ctx32768",
   "model": "Ternary-Bonsai-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 14.7 GB at this context; duplicate traces of the same session: aucfn9b2, eyv2jj31, jo3d3h85, ob260k8v; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 372.9,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-27B-mlx-2bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 2bit; 32k 372.9 24.2 14.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dusin764",
   "source_url": "https://omlx.ai/benchmarks/performance/dusin764",
   "tg_tokens": null,
   "tps": 24.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vjplucbj-ctx1024",
   "model": "GLM5.3FH",
   "notes": "oMLX community benchmark trace; peak memory 249.0 GB at this context; duplicate traces of the same session: cng2y0d6, jj1x4994, mx2uvsrs, q3bh2wbj, r3ngsqlr, s965wbyz; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 405.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "GLM5.3FH-Q6 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 1k 405.0 42.2 249.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vjplucbj",
   "source_url": "https://omlx.ai/benchmarks/performance/vjplucbj",
   "tg_tokens": null,
   "tps": 42.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vjplucbj-ctx4096",
   "model": "GLM5.3FH",
   "notes": "oMLX community benchmark trace; peak memory 250.2 GB at this context; duplicate traces of the same session: cng2y0d6, jj1x4994, mx2uvsrs, q3bh2wbj, r3ngsqlr, s965wbyz; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 531.2,
   "provenance": "community",
   "quant": "6bit",
   "quote": "GLM5.3FH-Q6 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 4k 531.2 33.0 250.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vjplucbj",
   "source_url": "https://omlx.ai/benchmarks/performance/vjplucbj",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vjplucbj-ctx8192",
   "model": "GLM5.3FH",
   "notes": "oMLX community benchmark trace; peak memory 250.5 GB at this context; duplicate traces of the same session: cng2y0d6, jj1x4994, mx2uvsrs, q3bh2wbj, r3ngsqlr, s965wbyz; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 481.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "GLM5.3FH-Q6 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 8k 481.8 51.6 250.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vjplucbj",
   "source_url": "https://omlx.ai/benchmarks/performance/vjplucbj",
   "tg_tokens": null,
   "tps": 51.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vjplucbj-ctx16384",
   "model": "GLM5.3FH",
   "notes": "oMLX community benchmark trace; peak memory 250.7 GB at this context; duplicate traces of the same session: cng2y0d6, jj1x4994, mx2uvsrs, q3bh2wbj, r3ngsqlr, s965wbyz; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 522.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "GLM5.3FH-Q6 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 16k 522.6 40.3 250.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vjplucbj",
   "source_url": "https://omlx.ai/benchmarks/performance/vjplucbj",
   "tg_tokens": null,
   "tps": 40.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vjplucbj-ctx32768",
   "model": "GLM5.3FH",
   "notes": "oMLX community benchmark trace; peak memory 250.9 GB at this context; duplicate traces of the same session: cng2y0d6, jj1x4994, mx2uvsrs, q3bh2wbj, r3ngsqlr, s965wbyz; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 515.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "GLM5.3FH-Q6 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 32k 515.5 40.9 250.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vjplucbj",
   "source_url": "https://omlx.ai/benchmarks/performance/vjplucbj",
   "tg_tokens": null,
   "tps": 40.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vjplucbj-ctx65536",
   "model": "GLM5.3FH",
   "notes": "oMLX community benchmark trace; peak memory 251.6 GB at this context; duplicate traces of the same session: cng2y0d6, jj1x4994, mx2uvsrs, q3bh2wbj, r3ngsqlr, s965wbyz; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 508.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "GLM5.3FH-Q6 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 64k 508.3 31.4 251.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vjplucbj",
   "source_url": "https://omlx.ai/benchmarks/performance/vjplucbj",
   "tg_tokens": null,
   "tps": 31.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vjplucbj-ctx131072",
   "model": "GLM5.3FH",
   "notes": "oMLX community benchmark trace; peak memory 252.9 GB at this context; duplicate traces of the same session: cng2y0d6, jj1x4994, mx2uvsrs, q3bh2wbj, r3ngsqlr, s965wbyz; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 497.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "GLM5.3FH-Q6 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 128k 497.6 39.0 252.9 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vjplucbj",
   "source_url": "https://omlx.ai/benchmarks/performance/vjplucbj",
   "tg_tokens": null,
   "tps": 39.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-g9b6bygm-ctx1024",
   "model": "LFM2-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: 0m3jotg3, nesf8iuq, ue76qwfr; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 906.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2-24B-A2B-MLX-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 1k 906.9 67.4 24.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g9b6bygm",
   "source_url": "https://omlx.ai/benchmarks/performance/g9b6bygm",
   "tg_tokens": null,
   "tps": 67.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-g9b6bygm-ctx4096",
   "model": "LFM2-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: 0m3jotg3, nesf8iuq, ue76qwfr; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1043.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2-24B-A2B-MLX-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k 1,043 65.8 24.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g9b6bygm",
   "source_url": "https://omlx.ai/benchmarks/performance/g9b6bygm",
   "tg_tokens": null,
   "tps": 65.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-g9b6bygm-ctx8192",
   "model": "LFM2-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: 0m3jotg3, nesf8iuq, ue76qwfr; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 837.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2-24B-A2B-MLX-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 8k 837.4 62.9 25.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g9b6bygm",
   "source_url": "https://omlx.ai/benchmarks/performance/g9b6bygm",
   "tg_tokens": null,
   "tps": 62.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-g9b6bygm-ctx16384",
   "model": "LFM2-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: 0m3jotg3, nesf8iuq, ue76qwfr; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 780.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2-24B-A2B-MLX-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 16k 780.8 58.6 26.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g9b6bygm",
   "source_url": "https://omlx.ai/benchmarks/performance/g9b6bygm",
   "tg_tokens": null,
   "tps": 58.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ltkc89x5-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 36.3 GB at this context; duplicate traces of the same session: 6bdu80eu, f8mvbhyh, jumwn4i3, ksfx71p7, okzvti0o, ps64jydh, z1zefguh; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2225.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 2,225 144.6 36.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ltkc89x5",
   "source_url": "https://omlx.ai/benchmarks/performance/ltkc89x5",
   "tg_tokens": null,
   "tps": 144.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ltkc89x5-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: 6bdu80eu, f8mvbhyh, jumwn4i3, ksfx71p7, okzvti0o, ps64jydh, z1zefguh; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2397.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 2,397 117.2 37.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ltkc89x5",
   "source_url": "https://omlx.ai/benchmarks/performance/ltkc89x5",
   "tg_tokens": null,
   "tps": 117.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ltkc89x5-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.9 GB at this context; duplicate traces of the same session: 6bdu80eu, f8mvbhyh, jumwn4i3, ksfx71p7, okzvti0o, ps64jydh, z1zefguh; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2284.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 8k 2,284 128.7 37.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ltkc89x5",
   "source_url": "https://omlx.ai/benchmarks/performance/ltkc89x5",
   "tg_tokens": null,
   "tps": 128.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ltkc89x5-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 38.1 GB at this context; duplicate traces of the same session: 6bdu80eu, f8mvbhyh, jumwn4i3, ksfx71p7, okzvti0o, ps64jydh, z1zefguh; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1990.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 16k 1,990 119.2 38.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ltkc89x5",
   "source_url": "https://omlx.ai/benchmarks/performance/ltkc89x5",
   "tg_tokens": null,
   "tps": 119.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ltkc89x5-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 38.9 GB at this context; duplicate traces of the same session: 6bdu80eu, f8mvbhyh, jumwn4i3, ksfx71p7, okzvti0o, ps64jydh, z1zefguh; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1600.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 32k 1,600 102.9 38.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ltkc89x5",
   "source_url": "https://omlx.ai/benchmarks/performance/ltkc89x5",
   "tg_tokens": null,
   "tps": 102.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ltkc89x5-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 40.4 GB at this context; duplicate traces of the same session: 6bdu80eu, f8mvbhyh, jumwn4i3, ksfx71p7, okzvti0o, ps64jydh, z1zefguh; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1144.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 64k 1,144 90.9 40.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ltkc89x5",
   "source_url": "https://omlx.ai/benchmarks/performance/ltkc89x5",
   "tg_tokens": null,
   "tps": 90.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ltkc89x5-ctx131072",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 43.4 GB at this context; duplicate traces of the same session: 6bdu80eu, f8mvbhyh, jumwn4i3, ksfx71p7, okzvti0o, ps64jydh, z1zefguh; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 725.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 128k 725.8 78.7 43.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ltkc89x5",
   "source_url": "https://omlx.ai/benchmarks/performance/ltkc89x5",
   "tg_tokens": null,
   "tps": 78.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ltkc89x5-ctx199680",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 47.9 GB at this context; duplicate traces of the same session: 6bdu80eu, f8mvbhyh, jumwn4i3, ksfx71p7, okzvti0o, ps64jydh, z1zefguh; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 523.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 195k 523.0 70.4 47.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ltkc89x5",
   "source_url": "https://omlx.ai/benchmarks/performance/ltkc89x5",
   "tg_tokens": null,
   "tps": 70.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ivd16ook-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: insmzoqq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3503.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 3,503 86.9 26.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ivd16ook",
   "source_url": "https://omlx.ai/benchmarks/performance/ivd16ook",
   "tg_tokens": null,
   "tps": 86.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ivd16ook-ctx131072",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 27.3 GB at this context; duplicate traces of the same session: insmzoqq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2435.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,435 74.5 27.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ivd16ook",
   "source_url": "https://omlx.ai/benchmarks/performance/ivd16ook",
   "tg_tokens": null,
   "tps": 74.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-xoncmce7-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: ayhnoise; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 380.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 380.4 34.1 3.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xoncmce7",
   "source_url": "https://omlx.ai/benchmarks/performance/xoncmce7",
   "tg_tokens": null,
   "tps": 34.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-xoncmce7-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.8 GB at this context; duplicate traces of the same session: ayhnoise; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 402.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 402.3 32.6 3.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xoncmce7",
   "source_url": "https://omlx.ai/benchmarks/performance/xoncmce7",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qntnscg6-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: p1fko3cr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 909.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 909.8 31.5 17.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qntnscg6",
   "source_url": "https://omlx.ai/benchmarks/performance/qntnscg6",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qntnscg6-ctx16384",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: p1fko3cr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 854.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 854.3 29.7 18.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qntnscg6",
   "source_url": "https://omlx.ai/benchmarks/performance/qntnscg6",
   "tg_tokens": null,
   "tps": 29.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-p5fc3org-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 8caazwlo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 888.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 888.2 31.5 17.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p5fc3org",
   "source_url": "https://omlx.ai/benchmarks/performance/p5fc3org",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-p5fc3org-ctx16384",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 8caazwlo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 879.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 879.8 29.7 18.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p5fc3org",
   "source_url": "https://omlx.ai/benchmarks/performance/p5fc3org",
   "tg_tokens": null,
   "tps": 29.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-a0bwcbr1-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: xsokfhoa; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 98.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 1k 98.8 20.2 18.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a0bwcbr1",
   "source_url": "https://omlx.ai/benchmarks/performance/a0bwcbr1",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-a0bwcbr1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: xsokfhoa; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 4k 102.8 20.5 19.7 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a0bwcbr1",
   "source_url": "https://omlx.ai/benchmarks/performance/a0bwcbr1",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-qax1cmd4-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 1d21o2s8, aack6ml2, bm4qaszf, q57rj8zn, sryvno97; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 254.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 254.6 52.1 21.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qax1cmd4",
   "source_url": "https://omlx.ai/benchmarks/performance/qax1cmd4",
   "tg_tokens": null,
   "tps": 52.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-qax1cmd4-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: 1d21o2s8, aack6ml2, bm4qaszf, q57rj8zn, sryvno97; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 339.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 339.5 80.3 25.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qax1cmd4",
   "source_url": "https://omlx.ai/benchmarks/performance/qax1cmd4",
   "tg_tokens": null,
   "tps": 80.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-qax1cmd4-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.6 GB at this context; duplicate traces of the same session: 1d21o2s8, aack6ml2, bm4qaszf, q57rj8zn, sryvno97; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 340.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 340.5 46.1 25.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qax1cmd4",
   "source_url": "https://omlx.ai/benchmarks/performance/qax1cmd4",
   "tg_tokens": null,
   "tps": 46.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-qax1cmd4-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: 1d21o2s8, aack6ml2, bm4qaszf, q57rj8zn, sryvno97; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 323.6 58.0 26.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qax1cmd4",
   "source_url": "https://omlx.ai/benchmarks/performance/qax1cmd4",
   "tg_tokens": null,
   "tps": 58.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-qax1cmd4-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.3 GB at this context; duplicate traces of the same session: 1d21o2s8, aack6ml2, bm4qaszf, q57rj8zn, sryvno97; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 293.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 293.9 52.2 28.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qax1cmd4",
   "source_url": "https://omlx.ai/benchmarks/performance/qax1cmd4",
   "tg_tokens": null,
   "tps": 52.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-qax1cmd4-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.1 GB at this context; duplicate traces of the same session: 1d21o2s8, aack6ml2, bm4qaszf, q57rj8zn, sryvno97; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 245.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 245.0 35.6 31.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qax1cmd4",
   "source_url": "https://omlx.ai/benchmarks/performance/qax1cmd4",
   "tg_tokens": null,
   "tps": 35.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1pgqissc-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 149.8 GB at this context; duplicate traces of the same session: rxpx9gb4, wy7i6sz9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 936.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-Flash-Next-MLX-6bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 936.7 57.8 149.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1pgqissc",
   "source_url": "https://omlx.ai/benchmarks/performance/1pgqissc",
   "tg_tokens": null,
   "tps": 57.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1pgqissc-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 151.8 GB at this context; duplicate traces of the same session: rxpx9gb4, wy7i6sz9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 944.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-Flash-Next-MLX-6bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 944.4 70.4 151.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1pgqissc",
   "source_url": "https://omlx.ai/benchmarks/performance/1pgqissc",
   "tg_tokens": null,
   "tps": 70.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1pgqissc-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 153.0 GB at this context; duplicate traces of the same session: rxpx9gb4, wy7i6sz9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 877.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-Flash-Next-MLX-6bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 8k 877.8 69.3 153.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1pgqissc",
   "source_url": "https://omlx.ai/benchmarks/performance/1pgqissc",
   "tg_tokens": null,
   "tps": 69.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1pgqissc-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 158.1 GB at this context; duplicate traces of the same session: rxpx9gb4, wy7i6sz9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 820.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-Flash-Next-MLX-6bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 32k 820.7 48.9 158.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1pgqissc",
   "source_url": "https://omlx.ai/benchmarks/performance/1pgqissc",
   "tg_tokens": null,
   "tps": 48.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-22yt9sne-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.2 GB at this context; duplicate traces of the same session: zxhrq3of; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 325.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 325.4 66.9 69.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/22yt9sne",
   "source_url": "https://omlx.ai/benchmarks/performance/22yt9sne",
   "tg_tokens": null,
   "tps": 66.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-22yt9sne-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.4 GB at this context; duplicate traces of the same session: zxhrq3of; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1030.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,030 57.9 69.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/22yt9sne",
   "source_url": "https://omlx.ai/benchmarks/performance/22yt9sne",
   "tg_tokens": null,
   "tps": 57.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i1gsx2wh-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: jv3xp0wq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1770.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,770 69.6 71.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i1gsx2wh",
   "source_url": "https://omlx.ai/benchmarks/performance/i1gsx2wh",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i1gsx2wh-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 75.2 GB at this context; duplicate traces of the same session: jv3xp0wq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2467.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,467 71.3 75.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i1gsx2wh",
   "source_url": "https://omlx.ai/benchmarks/performance/i1gsx2wh",
   "tg_tokens": null,
   "tps": 71.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ckb35s6k-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 74.6 GB; TTFT 29176 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2246.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k tokens 2,246 PP tok/s 66.9 TG tok/s 29176 TTFT (ms) 74.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ckb35s6k",
   "source_url": "https://omlx.ai/benchmarks/performance/ckb35s6k",
   "tg_tokens": null,
   "tps": 66.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-i30f73fm-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 4.6 GB at this context; duplicate traces of the same session: gczfsq1r, rsy5raer; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2502.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-4bit on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 4k 2,502 40.7 4.6 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i30f73fm",
   "source_url": "https://omlx.ai/benchmarks/performance/i30f73fm",
   "tg_tokens": null,
   "tps": 40.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-i30f73fm-ctx16384",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 4.8 GB at this context; duplicate traces of the same session: gczfsq1r, rsy5raer; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2534.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-4bit on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 16k 2,534 35.5 4.8 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i30f73fm",
   "source_url": "https://omlx.ai/benchmarks/performance/i30f73fm",
   "tg_tokens": null,
   "tps": 35.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-i30f73fm-ctx65536",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: gczfsq1r, rsy5raer; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1894.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-4bit on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 64k 1,894 22.3 5.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i30f73fm",
   "source_url": "https://omlx.ai/benchmarks/performance/i30f73fm",
   "tg_tokens": null,
   "tps": 22.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-kts1271b-ctx1024",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: 0fe00xfg, 238d0jeu, b1rq16kn; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 264.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 264.7 33.0 20.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kts1271b",
   "source_url": "https://omlx.ai/benchmarks/performance/kts1271b",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-kts1271b-ctx4096",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 0fe00xfg, 238d0jeu, b1rq16kn; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 367.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 367.7 32.7 22.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kts1271b",
   "source_url": "https://omlx.ai/benchmarks/performance/kts1271b",
   "tg_tokens": null,
   "tps": 32.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-kts1271b-ctx8192",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: 0fe00xfg, 238d0jeu, b1rq16kn; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 348.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 348.7 29.4 23.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kts1271b",
   "source_url": "https://omlx.ai/benchmarks/performance/kts1271b",
   "tg_tokens": null,
   "tps": 29.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-kts1271b-ctx16384",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: 0fe00xfg, 238d0jeu, b1rq16kn; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 310.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 310.9 23.6 24.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kts1271b",
   "source_url": "https://omlx.ai/benchmarks/performance/kts1271b",
   "tg_tokens": null,
   "tps": 23.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-1cg5jgln-ctx1024",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: j5vbdgns, k9xvnqsb, qcfxe0h7; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 615.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 615.9 54.8 20.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1cg5jgln",
   "source_url": "https://omlx.ai/benchmarks/performance/1cg5jgln",
   "tg_tokens": null,
   "tps": 54.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-1cg5jgln-ctx4096",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: j5vbdgns, k9xvnqsb, qcfxe0h7; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 215.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 215.7 23.0 22.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1cg5jgln",
   "source_url": "https://omlx.ai/benchmarks/performance/1cg5jgln",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-1cg5jgln-ctx8192",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: j5vbdgns, k9xvnqsb, qcfxe0h7; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 240.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 240.1 21.8 23.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1cg5jgln",
   "source_url": "https://omlx.ai/benchmarks/performance/1cg5jgln",
   "tg_tokens": null,
   "tps": 21.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-1cg5jgln-ctx16384",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: j5vbdgns, k9xvnqsb, qcfxe0h7; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 219.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 219.0 15.5 23.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1cg5jgln",
   "source_url": "https://omlx.ai/benchmarks/performance/1cg5jgln",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-7ysg21a2-ctx4096",
   "model": "Laguna XS 2.1",
   "notes": "oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: c9domgql; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1188.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Laguna-XS-2.1-oQ4e on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,188 71.0 18.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ysg21a2",
   "source_url": "https://omlx.ai/benchmarks/performance/7ysg21a2",
   "tg_tokens": null,
   "tps": 71.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-7ysg21a2-ctx16384",
   "model": "Laguna XS 2.1",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: c9domgql; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1024.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Laguna-XS-2.1-oQ4e on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 1,024 56.8 19.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ysg21a2",
   "source_url": "https://omlx.ai/benchmarks/performance/7ysg21a2",
   "tg_tokens": null,
   "tps": 56.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-srpm3az7-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 2ioso62b, a7lrsztb, yf8l36aa; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1164.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 1,164 87.4 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/srpm3az7",
   "source_url": "https://omlx.ai/benchmarks/performance/srpm3az7",
   "tg_tokens": null,
   "tps": 87.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-srpm3az7-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 2ioso62b, a7lrsztb, yf8l36aa; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1489.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 1,489 84.9 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/srpm3az7",
   "source_url": "https://omlx.ai/benchmarks/performance/srpm3az7",
   "tg_tokens": null,
   "tps": 84.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-srpm3az7-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 2ioso62b, a7lrsztb, yf8l36aa; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1221.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 16k 1,221 75.1 22.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/srpm3az7",
   "source_url": "https://omlx.ai/benchmarks/performance/srpm3az7",
   "tg_tokens": null,
   "tps": 75.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-srpm3az7-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: 2ioso62b, a7lrsztb, yf8l36aa; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 960.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 32k 960.9 67.5 23.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/srpm3az7",
   "source_url": "https://omlx.ai/benchmarks/performance/srpm3az7",
   "tg_tokens": null,
   "tps": 67.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ukhyesaq-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 863.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 863.8 87.3 19.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ukhyesaq",
   "source_url": "https://omlx.ai/benchmarks/performance/ukhyesaq",
   "tg_tokens": null,
   "tps": 87.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ukhyesaq-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1007.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 1,007 82.7 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ukhyesaq",
   "source_url": "https://omlx.ai/benchmarks/performance/ukhyesaq",
   "tg_tokens": null,
   "tps": 82.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-h8wzm105-ctx4096",
   "model": "Qwen3.5-4B-OptiQ",
   "notes": "oMLX community benchmark trace; peak memory 4.8 GB at this context; duplicate traces of the same session: none; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 525.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-OptiQ-4bit on M3 Pro (14c); M3 Pro (14c) \u00b7 18 GB \u00b7 4bit; 4k 525.7 35.9 4.8 GB current",
   "ram_gb": 18,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h8wzm105",
   "source_url": "https://omlx.ai/benchmarks/performance/h8wzm105",
   "tg_tokens": null,
   "tps": 35.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-h8wzm105-ctx16384",
   "model": "Qwen3.5-4B-OptiQ",
   "notes": "oMLX community benchmark trace; peak memory 5.3 GB at this context; duplicate traces of the same session: none; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 474.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-OptiQ-4bit on M3 Pro (14c); M3 Pro (14c) \u00b7 18 GB \u00b7 4bit; 16k 474.7 31.2 5.3 GB view",
   "ram_gb": 18,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h8wzm105",
   "source_url": "https://omlx.ai/benchmarks/performance/h8wzm105",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-t2tldey7-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 8g55jvni, kz7kyt59, nuqz17zu, s22o9z0f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 249.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 249.6 52.1 21.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t2tldey7",
   "source_url": "https://omlx.ai/benchmarks/performance/t2tldey7",
   "tg_tokens": null,
   "tps": 52.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-t2tldey7-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: 8g55jvni, kz7kyt59, nuqz17zu, s22o9z0f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 338.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 338.7 46.7 25.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t2tldey7",
   "source_url": "https://omlx.ai/benchmarks/performance/t2tldey7",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-t2tldey7-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.6 GB at this context; duplicate traces of the same session: 8g55jvni, kz7kyt59, nuqz17zu, s22o9z0f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 340.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 340.0 38.2 25.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t2tldey7",
   "source_url": "https://omlx.ai/benchmarks/performance/t2tldey7",
   "tg_tokens": null,
   "tps": 38.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-t2tldey7-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: 8g55jvni, kz7kyt59, nuqz17zu, s22o9z0f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 319.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 319.4 56.1 26.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t2tldey7",
   "source_url": "https://omlx.ai/benchmarks/performance/t2tldey7",
   "tg_tokens": null,
   "tps": 56.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-t2tldey7-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.4 GB at this context; duplicate traces of the same session: 8g55jvni, kz7kyt59, nuqz17zu, s22o9z0f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 291.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 291.6 59.7 28.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t2tldey7",
   "source_url": "https://omlx.ai/benchmarks/performance/t2tldey7",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-tggov0vr-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 149.8 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 936.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-Flash-Next-MLX-6bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 936.7 57.8 149.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tggov0vr",
   "source_url": "https://omlx.ai/benchmarks/performance/tggov0vr",
   "tg_tokens": null,
   "tps": 57.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-tggov0vr-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 151.8 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 944.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-Flash-Next-MLX-6bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 944.4 70.4 151.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tggov0vr",
   "source_url": "https://omlx.ai/benchmarks/performance/tggov0vr",
   "tg_tokens": null,
   "tps": 70.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-tggov0vr-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 153.0 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 877.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-Flash-Next-MLX-6bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 8k 877.8 69.3 153.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tggov0vr",
   "source_url": "https://omlx.ai/benchmarks/performance/tggov0vr",
   "tg_tokens": null,
   "tps": 69.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-tggov0vr-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 158.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 820.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-Flash-Next-MLX-6bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 32k 820.7 48.9 158.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tggov0vr",
   "source_url": "https://omlx.ai/benchmarks/performance/tggov0vr",
   "tg_tokens": null,
   "tps": 48.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-82n2k15q-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 42.7 GB; TTFT 51472 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 159.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k tokens 159.2 PP tok/s 10.0 TG tok/s 51472 TTFT (ms) 42.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/82n2k15q",
   "source_url": "https://omlx.ai/benchmarks/performance/82n2k15q",
   "tg_tokens": null,
   "tps": 10.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-yimtp8xd-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 40.7 GB at this context; duplicate traces of the same session: 4gtlrkay; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 86.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 86.4 10.1 40.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yimtp8xd",
   "source_url": "https://omlx.ai/benchmarks/performance/yimtp8xd",
   "tg_tokens": null,
   "tps": 10.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-yimtp8xd-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 41.4 GB at this context; duplicate traces of the same session: 4gtlrkay; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 173.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 173.1 8.6 41.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yimtp8xd",
   "source_url": "https://omlx.ai/benchmarks/performance/yimtp8xd",
   "tg_tokens": null,
   "tps": 8.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ipwoogun-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: y8qsjbln; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1832.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,832 80.3 71.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ipwoogun",
   "source_url": "https://omlx.ai/benchmarks/performance/ipwoogun",
   "tg_tokens": null,
   "tps": 80.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ipwoogun-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.7 GB at this context; duplicate traces of the same session: y8qsjbln; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2188.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,188 63.5 74.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ipwoogun",
   "source_url": "https://omlx.ai/benchmarks/performance/ipwoogun",
   "tg_tokens": null,
   "tps": 63.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-xz2o34j0-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 86.4 GB at this context; duplicate traces of the same session: 0evnr4g6, brgz439w, eh2o0pjd, myae3qr0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 543.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 543.2 63.1 86.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xz2o34j0",
   "source_url": "https://omlx.ai/benchmarks/performance/xz2o34j0",
   "tg_tokens": null,
   "tps": 63.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-xz2o34j0-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 87.0 GB at this context; duplicate traces of the same session: 0evnr4g6, brgz439w, eh2o0pjd, myae3qr0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 688.5,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 688.5 60.9 87.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xz2o34j0",
   "source_url": "https://omlx.ai/benchmarks/performance/xz2o34j0",
   "tg_tokens": null,
   "tps": 60.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-xz2o34j0-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 87.5 GB at this context; duplicate traces of the same session: 0evnr4g6, brgz439w, eh2o0pjd, myae3qr0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 807.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 807.7 61.8 87.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xz2o34j0",
   "source_url": "https://omlx.ai/benchmarks/performance/xz2o34j0",
   "tg_tokens": null,
   "tps": 61.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-xz2o34j0-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 89.7 GB at this context; duplicate traces of the same session: 0evnr4g6, brgz439w, eh2o0pjd, myae3qr0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 792.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 792.7 60.7 89.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xz2o34j0",
   "source_url": "https://omlx.ai/benchmarks/performance/xz2o34j0",
   "tg_tokens": null,
   "tps": 60.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-xz2o34j0-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 94.0 GB at this context; duplicate traces of the same session: 0evnr4g6, brgz439w, eh2o0pjd, myae3qr0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 780.5,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 780.5 52.4 94.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xz2o34j0",
   "source_url": "https://omlx.ai/benchmarks/performance/xz2o34j0",
   "tg_tokens": null,
   "tps": 52.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-7nu17o31-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: xuzzygsu, yj1enfsx, zwq3t8a3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 286.5,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 5bit; 1k 286.5 50.6 24.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7nu17o31",
   "source_url": "https://omlx.ai/benchmarks/performance/7nu17o31",
   "tg_tokens": null,
   "tps": 50.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-7nu17o31-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: xuzzygsu, yj1enfsx, zwq3t8a3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 5bit; 4k 323.2 45.9 27.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7nu17o31",
   "source_url": "https://omlx.ai/benchmarks/performance/7nu17o31",
   "tg_tokens": null,
   "tps": 45.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-7nu17o31-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: xuzzygsu, yj1enfsx, zwq3t8a3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 317.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 5bit; 8k 317.0 61.1 28.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7nu17o31",
   "source_url": "https://omlx.ai/benchmarks/performance/7nu17o31",
   "tg_tokens": null,
   "tps": 61.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-7nu17o31-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: xuzzygsu, yj1enfsx, zwq3t8a3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 302.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 5bit; 16k 302.2 59.6 29.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7nu17o31",
   "source_url": "https://omlx.ai/benchmarks/performance/7nu17o31",
   "tg_tokens": null,
   "tps": 59.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-3yj1gvaw-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: idbnourl; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 286.5,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 5bit; 1k 286.5 45.1 24.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3yj1gvaw",
   "source_url": "https://omlx.ai/benchmarks/performance/3yj1gvaw",
   "tg_tokens": null,
   "tps": 45.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-3yj1gvaw-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: idbnourl; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 5bit; 4k 323.9 52.0 27.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3yj1gvaw",
   "source_url": "https://omlx.ai/benchmarks/performance/3yj1gvaw",
   "tg_tokens": null,
   "tps": 52.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-zn43wmq7-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: abmfl7iz, elenuogu, fibkax4r, hgs3vts3, lgpplx9l, qct5qyr0, tyvkpzdp; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 725.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 725.1 78.5 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zn43wmq7",
   "source_url": "https://omlx.ai/benchmarks/performance/zn43wmq7",
   "tg_tokens": null,
   "tps": 78.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-zn43wmq7-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: abmfl7iz, elenuogu, fibkax4r, hgs3vts3, lgpplx9l, qct5qyr0, tyvkpzdp; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 873.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 873.7 75.4 23.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zn43wmq7",
   "source_url": "https://omlx.ai/benchmarks/performance/zn43wmq7",
   "tg_tokens": null,
   "tps": 75.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-zn43wmq7-ctx8192",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: abmfl7iz, elenuogu, fibkax4r, hgs3vts3, lgpplx9l, qct5qyr0, tyvkpzdp; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 846.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 8k 846.7 72.3 24.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zn43wmq7",
   "source_url": "https://omlx.ai/benchmarks/performance/zn43wmq7",
   "tg_tokens": null,
   "tps": 72.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-zn43wmq7-ctx16384",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: abmfl7iz, elenuogu, fibkax4r, hgs3vts3, lgpplx9l, qct5qyr0, tyvkpzdp; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 741.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k 741.2 66.4 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zn43wmq7",
   "source_url": "https://omlx.ai/benchmarks/performance/zn43wmq7",
   "tg_tokens": null,
   "tps": 66.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-zn43wmq7-ctx32768",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: abmfl7iz, elenuogu, fibkax4r, hgs3vts3, lgpplx9l, qct5qyr0, tyvkpzdp; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 630.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 32k 630.6 59.1 24.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zn43wmq7",
   "source_url": "https://omlx.ai/benchmarks/performance/zn43wmq7",
   "tg_tokens": null,
   "tps": 59.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-zn43wmq7-ctx65536",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: abmfl7iz, elenuogu, fibkax4r, hgs3vts3, lgpplx9l, qct5qyr0, tyvkpzdp; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 480.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 64k 480.6 46.1 24.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zn43wmq7",
   "source_url": "https://omlx.ai/benchmarks/performance/zn43wmq7",
   "tg_tokens": null,
   "tps": 46.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-zn43wmq7-ctx131072",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: abmfl7iz, elenuogu, fibkax4r, hgs3vts3, lgpplx9l, qct5qyr0, tyvkpzdp; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 328.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 128k 328.0 33.2 26.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zn43wmq7",
   "source_url": "https://omlx.ai/benchmarks/performance/zn43wmq7",
   "tg_tokens": null,
   "tps": 33.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-zn43wmq7-ctx199680",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: abmfl7iz, elenuogu, fibkax4r, hgs3vts3, lgpplx9l, qct5qyr0, tyvkpzdp; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 246.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 195k 246.8 25.7 28.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zn43wmq7",
   "source_url": "https://omlx.ai/benchmarks/performance/zn43wmq7",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-x4siqoxh-ctx1024",
   "model": "LFM2-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: 57of3xkn, do9m0hls, hdwxy8u6, ogk18nic, t911m2kd, wsbqmhg4; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 923.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2-24B-A2B-MLX-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 1k 923.7 67.3 24.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x4siqoxh",
   "source_url": "https://omlx.ai/benchmarks/performance/x4siqoxh",
   "tg_tokens": null,
   "tps": 67.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-x4siqoxh-ctx4096",
   "model": "LFM2-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: 57of3xkn, do9m0hls, hdwxy8u6, ogk18nic, t911m2kd, wsbqmhg4; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1045.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2-24B-A2B-MLX-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k 1,045 65.8 24.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x4siqoxh",
   "source_url": "https://omlx.ai/benchmarks/performance/x4siqoxh",
   "tg_tokens": null,
   "tps": 65.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-x4siqoxh-ctx8192",
   "model": "LFM2-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: 57of3xkn, do9m0hls, hdwxy8u6, ogk18nic, t911m2kd, wsbqmhg4; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 830.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2-24B-A2B-MLX-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 8k 830.1 62.6 25.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x4siqoxh",
   "source_url": "https://omlx.ai/benchmarks/performance/x4siqoxh",
   "tg_tokens": null,
   "tps": 62.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-x4siqoxh-ctx16384",
   "model": "LFM2-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: 57of3xkn, do9m0hls, hdwxy8u6, ogk18nic, t911m2kd, wsbqmhg4; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 770.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2-24B-A2B-MLX-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 16k 770.8 58.6 26.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x4siqoxh",
   "source_url": "https://omlx.ai/benchmarks/performance/x4siqoxh",
   "tg_tokens": null,
   "tps": 58.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-x4siqoxh-ctx32768",
   "model": "LFM2-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: 57of3xkn, do9m0hls, hdwxy8u6, ogk18nic, t911m2kd, wsbqmhg4; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 696.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2-24B-A2B-MLX-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 32k 696.4 55.4 28.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x4siqoxh",
   "source_url": "https://omlx.ai/benchmarks/performance/x4siqoxh",
   "tg_tokens": null,
   "tps": 55.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-x4siqoxh-ctx65536",
   "model": "LFM2-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 33.0 GB at this context; duplicate traces of the same session: 57of3xkn, do9m0hls, hdwxy8u6, ogk18nic, t911m2kd, wsbqmhg4; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 603.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2-24B-A2B-MLX-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 64k 603.4 44.0 33.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x4siqoxh",
   "source_url": "https://omlx.ai/benchmarks/performance/x4siqoxh",
   "tg_tokens": null,
   "tps": 44.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-x4siqoxh-ctx131072",
   "model": "LFM2-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 41.8 GB at this context; duplicate traces of the same session: 57of3xkn, do9m0hls, hdwxy8u6, ogk18nic, t911m2kd, wsbqmhg4; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 477.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2-24B-A2B-MLX-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 128k 477.0 31.4 41.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x4siqoxh",
   "source_url": "https://omlx.ai/benchmarks/performance/x4siqoxh",
   "tg_tokens": null,
   "tps": 31.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-uwnxwn1h-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 863.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 863.8 87.3 19.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uwnxwn1h",
   "source_url": "https://omlx.ai/benchmarks/performance/uwnxwn1h",
   "tg_tokens": null,
   "tps": 87.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-uwnxwn1h-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1007.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 1,007 82.7 20.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uwnxwn1h",
   "source_url": "https://omlx.ai/benchmarks/performance/uwnxwn1h",
   "tg_tokens": null,
   "tps": 82.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-8qlhzqn3-ctx4096",
   "model": "Qwen3.5-4B-OptiQ",
   "notes": "oMLX community benchmark trace; peak memory 4.8 GB at this context; duplicate traces of the same session: none; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 525.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-OptiQ-4bit on M3 Pro (14c); M3 Pro (14c) \u00b7 18 GB \u00b7 4bit; 4k 525.7 35.9 4.8 GB view",
   "ram_gb": 18,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8qlhzqn3",
   "source_url": "https://omlx.ai/benchmarks/performance/8qlhzqn3",
   "tg_tokens": null,
   "tps": 35.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-8qlhzqn3-ctx16384",
   "model": "Qwen3.5-4B-OptiQ",
   "notes": "oMLX community benchmark trace; peak memory 5.3 GB at this context; duplicate traces of the same session: none; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 474.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-OptiQ-4bit on M3 Pro (14c); M3 Pro (14c) \u00b7 18 GB \u00b7 4bit; 16k 474.7 31.2 5.3 GB current",
   "ram_gb": 18,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8qlhzqn3",
   "source_url": "https://omlx.ai/benchmarks/performance/8qlhzqn3",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-tppttmgz-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 203.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 203.7 21.0 5.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tppttmgz",
   "source_url": "https://omlx.ai/benchmarks/performance/tppttmgz",
   "tg_tokens": null,
   "tps": 21.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-tppttmgz-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 215.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 215.2 20.1 6.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tppttmgz",
   "source_url": "https://omlx.ai/benchmarks/performance/tppttmgz",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-8jqkn4wn-ctx32768",
   "model": "Qwen3.6-35B-A3B-MLX-VL",
   "notes": "FP16 build; oMLX community benchmark trace, single run at 32k context; peak memory 24.6 GB; TTFT 55428 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 591.2,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-MLX-VL-oQ4-FP16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 32k tokens 591.2 PP tok/s 53.0 TG tok/s 55428 TTFT (ms) 24.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8jqkn4wn",
   "source_url": "https://omlx.ai/benchmarks/performance/8jqkn4wn",
   "tg_tokens": null,
   "tps": 53.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-oz8v31vb-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: pgqgxd55; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3456.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 3,456 133.8 21.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oz8v31vb",
   "source_url": "https://omlx.ai/benchmarks/performance/oz8v31vb",
   "tg_tokens": null,
   "tps": 133.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-oz8v31vb-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: pgqgxd55; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4982.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 4,982 121.1 22.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oz8v31vb",
   "source_url": "https://omlx.ai/benchmarks/performance/oz8v31vb",
   "tg_tokens": null,
   "tps": 121.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pvhrot4v-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 0uhbnmn5, 9ko44dw0, ls3bj1p5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3556.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 3,556 195.4 21.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pvhrot4v",
   "source_url": "https://omlx.ai/benchmarks/performance/pvhrot4v",
   "tg_tokens": null,
   "tps": 195.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pvhrot4v-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 0uhbnmn5, 9ko44dw0, ls3bj1p5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5655.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 5,655 156.8 22.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pvhrot4v",
   "source_url": "https://omlx.ai/benchmarks/performance/pvhrot4v",
   "tg_tokens": null,
   "tps": 156.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pvhrot4v-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 0uhbnmn5, 9ko44dw0, ls3bj1p5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5586.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 5,586 112.5 22.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pvhrot4v",
   "source_url": "https://omlx.ai/benchmarks/performance/pvhrot4v",
   "tg_tokens": null,
   "tps": 112.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pvhrot4v-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 0uhbnmn5, 9ko44dw0, ls3bj1p5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5285.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 5,285 189.1 22.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pvhrot4v",
   "source_url": "https://omlx.ai/benchmarks/performance/pvhrot4v",
   "tg_tokens": null,
   "tps": 189.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rbpsv7se-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 41rbyz92, qq595vnu, sgpxgn79; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1753.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,753 92.3 21.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rbpsv7se",
   "source_url": "https://omlx.ai/benchmarks/performance/rbpsv7se",
   "tg_tokens": null,
   "tps": 92.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rbpsv7se-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 41rbyz92, qq595vnu, sgpxgn79; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5009.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 5,009 137.1 22.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rbpsv7se",
   "source_url": "https://omlx.ai/benchmarks/performance/rbpsv7se",
   "tg_tokens": null,
   "tps": 137.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rbpsv7se-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 41rbyz92, qq595vnu, sgpxgn79; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5641.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 5,641 33.9 22.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rbpsv7se",
   "source_url": "https://omlx.ai/benchmarks/performance/rbpsv7se",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rbpsv7se-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 41rbyz92, qq595vnu, sgpxgn79; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5280.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 5,280 148.6 22.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rbpsv7se",
   "source_url": "https://omlx.ai/benchmarks/performance/rbpsv7se",
   "tg_tokens": null,
   "tps": 148.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-ewht439x-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 1cwwjotu, et9g10bh, kejgdy4d; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 222.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 222.4 31.5 18.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ewht439x",
   "source_url": "https://omlx.ai/benchmarks/performance/ewht439x",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-ewht439x-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 1cwwjotu, et9g10bh, kejgdy4d; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 243.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 243.1 30.0 22.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ewht439x",
   "source_url": "https://omlx.ai/benchmarks/performance/ewht439x",
   "tg_tokens": null,
   "tps": 30.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-ewht439x-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: 1cwwjotu, et9g10bh, kejgdy4d; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 243.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 16k 243.3 25.5 24.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ewht439x",
   "source_url": "https://omlx.ai/benchmarks/performance/ewht439x",
   "tg_tokens": null,
   "tps": 25.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-ewht439x-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: 1cwwjotu, et9g10bh, kejgdy4d; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 174.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 32k 174.4 21.4 22.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ewht439x",
   "source_url": "https://omlx.ai/benchmarks/performance/ewht439x",
   "tg_tokens": null,
   "tps": 21.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jy7db2kw-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.7 GB at this context; duplicate traces of the same session: 2rqngdiy, m22u04jm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 966.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 966.5 26.7 33.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jy7db2kw",
   "source_url": "https://omlx.ai/benchmarks/performance/jy7db2kw",
   "tg_tokens": null,
   "tps": 26.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jy7db2kw-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: 2rqngdiy, m22u04jm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1139.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 1,139 31.2 34.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jy7db2kw",
   "source_url": "https://omlx.ai/benchmarks/performance/jy7db2kw",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jy7db2kw-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: 2rqngdiy, m22u04jm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1061.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 1,061 46.8 35.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jy7db2kw",
   "source_url": "https://omlx.ai/benchmarks/performance/jy7db2kw",
   "tg_tokens": null,
   "tps": 46.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-kb8p12fg-ctx8192",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 8k context; peak memory 40.6 GB; TTFT 37220 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 8k tokens 220.1 PP tok/s 18.7 TG tok/s 37220 TTFT (ms) 40.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kb8p12fg",
   "source_url": "https://omlx.ai/benchmarks/performance/kb8p12fg",
   "tg_tokens": null,
   "tps": 18.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-7ea3fup7-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: re2kxsup, x4hp2i12; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 637.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 637.4 59.6 70.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ea3fup7",
   "source_url": "https://omlx.ai/benchmarks/performance/7ea3fup7",
   "tg_tokens": null,
   "tps": 59.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-7ea3fup7-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: re2kxsup, x4hp2i12; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 733.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 733.8 43.3 71.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ea3fup7",
   "source_url": "https://omlx.ai/benchmarks/performance/7ea3fup7",
   "tg_tokens": null,
   "tps": 43.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-7ea3fup7-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: re2kxsup, x4hp2i12; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 828.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 828.3 60.2 71.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ea3fup7",
   "source_url": "https://omlx.ai/benchmarks/performance/7ea3fup7",
   "tg_tokens": null,
   "tps": 60.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-8x5x99hk-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.4 GB at this context; duplicate traces of the same session: 34vl0e5n, a6loww9z, mdgurgoa; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 903.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 903.5 70.2 180.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8x5x99hk",
   "source_url": "https://omlx.ai/benchmarks/performance/8x5x99hk",
   "tg_tokens": null,
   "tps": 70.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-8x5x99hk-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.0 GB at this context; duplicate traces of the same session: 34vl0e5n, a6loww9z, mdgurgoa; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1090.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 1,090 65.2 181.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8x5x99hk",
   "source_url": "https://omlx.ai/benchmarks/performance/8x5x99hk",
   "tg_tokens": null,
   "tps": 65.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-8x5x99hk-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.2 GB at this context; duplicate traces of the same session: 34vl0e5n, a6loww9z, mdgurgoa; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1076.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 1,076 65.5 181.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8x5x99hk",
   "source_url": "https://omlx.ai/benchmarks/performance/8x5x99hk",
   "tg_tokens": null,
   "tps": 65.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-8x5x99hk-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 184.5 GB at this context; duplicate traces of the same session: 34vl0e5n, a6loww9z, mdgurgoa; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1043.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 64k 1,043 59.2 184.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8x5x99hk",
   "source_url": "https://omlx.ai/benchmarks/performance/8x5x99hk",
   "tg_tokens": null,
   "tps": 59.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ja9j043s-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.1 GB at this context; duplicate traces of the same session: 23mtoq6l, 3m4pswx1, agbd20eo, itaq2k15, v1sbvg7z; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 706.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 706.9 20.4 183.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ja9j043s",
   "source_url": "https://omlx.ai/benchmarks/performance/ja9j043s",
   "tg_tokens": null,
   "tps": 20.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ja9j043s-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.6 GB at this context; duplicate traces of the same session: 23mtoq6l, 3m4pswx1, agbd20eo, itaq2k15, v1sbvg7z; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 884.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 884.1 67.3 183.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ja9j043s",
   "source_url": "https://omlx.ai/benchmarks/performance/ja9j043s",
   "tg_tokens": null,
   "tps": 67.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ja9j043s-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.9 GB at this context; duplicate traces of the same session: 23mtoq6l, 3m4pswx1, agbd20eo, itaq2k15, v1sbvg7z; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1044.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 1,044 80.9 183.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ja9j043s",
   "source_url": "https://omlx.ai/benchmarks/performance/ja9j043s",
   "tg_tokens": null,
   "tps": 80.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ja9j043s-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 186.3 GB at this context; duplicate traces of the same session: 23mtoq6l, 3m4pswx1, agbd20eo, itaq2k15, v1sbvg7z; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1024.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 32k 1,024 32.4 186.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ja9j043s",
   "source_url": "https://omlx.ai/benchmarks/performance/ja9j043s",
   "tg_tokens": null,
   "tps": 32.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ja9j043s-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 189.5 GB at this context; duplicate traces of the same session: 23mtoq6l, 3m4pswx1, agbd20eo, itaq2k15, v1sbvg7z; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1010.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 64k 1,010 47.2 189.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ja9j043s",
   "source_url": "https://omlx.ai/benchmarks/performance/ja9j043s",
   "tg_tokens": null,
   "tps": 47.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ja9j043s-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 195.7 GB at this context; duplicate traces of the same session: 23mtoq6l, 3m4pswx1, agbd20eo, itaq2k15, v1sbvg7z; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 994.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 128k 994.7 53.7 195.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ja9j043s",
   "source_url": "https://omlx.ai/benchmarks/performance/ja9j043s",
   "tg_tokens": null,
   "tps": 53.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-e42w6sj5-ctx8192",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace, single run at 8k context; peak memory 21.0 GB; TTFT 84147 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 97.4,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 2bit; 8k tokens 97.4 PP tok/s 16.1 TG tok/s 84147 TTFT (ms) 21.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e42w6sj5",
   "source_url": "https://omlx.ai/benchmarks/performance/e42w6sj5",
   "tg_tokens": null,
   "tps": 16.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-p0a57t5i-ctx1024",
   "model": "gemma-4-26B-A4B-it-uncensored-heretic-ara",
   "notes": "oMLX community benchmark trace; peak memory 15.1 GB at this context; duplicate traces of the same session: 9c5uedvs, e5xewejt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1532.0,
   "provenance": "community",
   "quant": "mixed 4+6bit",
   "quote": "gemma-4-26B-A4B-it-uncensored-heretic-ara-MLX-4bit-mixed_4_6 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,532 106.6 15.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p0a57t5i",
   "source_url": "https://omlx.ai/benchmarks/performance/p0a57t5i",
   "tg_tokens": null,
   "tps": 106.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-p0a57t5i-ctx4096",
   "model": "gemma-4-26B-A4B-it-uncensored-heretic-ara",
   "notes": "oMLX community benchmark trace; peak memory 15.4 GB at this context; duplicate traces of the same session: 9c5uedvs, e5xewejt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1470.0,
   "provenance": "community",
   "quant": "mixed 4+6bit",
   "quote": "gemma-4-26B-A4B-it-uncensored-heretic-ara-MLX-4bit-mixed_4_6 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,470 106.5 15.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p0a57t5i",
   "source_url": "https://omlx.ai/benchmarks/performance/p0a57t5i",
   "tg_tokens": null,
   "tps": 106.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-p0a57t5i-ctx16384",
   "model": "gemma-4-26B-A4B-it-uncensored-heretic-ara",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 9c5uedvs, e5xewejt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1213.0,
   "provenance": "community",
   "quant": "mixed 4+6bit",
   "quote": "gemma-4-26B-A4B-it-uncensored-heretic-ara-MLX-4bit-mixed_4_6 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,213 91.1 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p0a57t5i",
   "source_url": "https://omlx.ai/benchmarks/performance/p0a57t5i",
   "tg_tokens": null,
   "tps": 91.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-urxe9th4-ctx1024",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: btqvgo5p, s59cubw7, ylc4qnur; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 625.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 625.8 53.2 20.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/urxe9th4",
   "source_url": "https://omlx.ai/benchmarks/performance/urxe9th4",
   "tg_tokens": null,
   "tps": 53.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-urxe9th4-ctx4096",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: btqvgo5p, s59cubw7, ylc4qnur; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 363.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 363.3 36.2 22.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/urxe9th4",
   "source_url": "https://omlx.ai/benchmarks/performance/urxe9th4",
   "tg_tokens": null,
   "tps": 36.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-urxe9th4-ctx8192",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: btqvgo5p, s59cubw7, ylc4qnur; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 323.8 32.3 23.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/urxe9th4",
   "source_url": "https://omlx.ai/benchmarks/performance/urxe9th4",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-urxe9th4-ctx16384",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: btqvgo5p, s59cubw7, ylc4qnur; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 286.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 286.6 22.9 24.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/urxe9th4",
   "source_url": "https://omlx.ai/benchmarks/performance/urxe9th4",
   "tg_tokens": null,
   "tps": 22.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-ifi3dzf3-ctx1024",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: 1zxfcx60, 54zn6sch, m1q1p4b9; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 252.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 252.5 32.0 20.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ifi3dzf3",
   "source_url": "https://omlx.ai/benchmarks/performance/ifi3dzf3",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-ifi3dzf3-ctx4096",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 1zxfcx60, 54zn6sch, m1q1p4b9; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 344.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 344.9 37.6 22.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ifi3dzf3",
   "source_url": "https://omlx.ai/benchmarks/performance/ifi3dzf3",
   "tg_tokens": null,
   "tps": 37.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-ifi3dzf3-ctx8192",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: 1zxfcx60, 54zn6sch, m1q1p4b9; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 338.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 338.0 30.2 23.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ifi3dzf3",
   "source_url": "https://omlx.ai/benchmarks/performance/ifi3dzf3",
   "tg_tokens": null,
   "tps": 30.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-ifi3dzf3-ctx16384",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: 1zxfcx60, 54zn6sch, m1q1p4b9; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 295.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 295.6 23.3 24.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ifi3dzf3",
   "source_url": "https://omlx.ai/benchmarks/performance/ifi3dzf3",
   "tg_tokens": null,
   "tps": 23.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-q9r3etth-ctx4096",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: 0k164nto; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 421.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 421.5 42.3 22.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q9r3etth",
   "source_url": "https://omlx.ai/benchmarks/performance/q9r3etth",
   "tg_tokens": null,
   "tps": 42.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-q9r3etth-ctx16384",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 0k164nto; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 344.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 344.2 31.4 23.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q9r3etth",
   "source_url": "https://omlx.ai/benchmarks/performance/q9r3etth",
   "tg_tokens": null,
   "tps": 31.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-8owt341m-ctx4096",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: wjn925ls; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 428.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 428.5 34.9 22.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8owt341m",
   "source_url": "https://omlx.ai/benchmarks/performance/8owt341m",
   "tg_tokens": null,
   "tps": 34.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-8owt341m-ctx16384",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: wjn925ls; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 342.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 342.2 29.8 23.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8owt341m",
   "source_url": "https://omlx.ai/benchmarks/performance/8owt341m",
   "tg_tokens": null,
   "tps": 29.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-o5ner1u5-ctx4096",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: jy0ec0zd; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 382.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 382.1 42.3 22.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o5ner1u5",
   "source_url": "https://omlx.ai/benchmarks/performance/o5ner1u5",
   "tg_tokens": null,
   "tps": 42.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-o5ner1u5-ctx16384",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: jy0ec0zd; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 323.8 24.8 23.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o5ner1u5",
   "source_url": "https://omlx.ai/benchmarks/performance/o5ner1u5",
   "tg_tokens": null,
   "tps": 24.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-798w8l1f-ctx1024",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 314.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 314.4 33.6 20.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/798w8l1f",
   "source_url": "https://omlx.ai/benchmarks/performance/798w8l1f",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-798w8l1f-ctx4096",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 376.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 376.4 34.2 22.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/798w8l1f",
   "source_url": "https://omlx.ai/benchmarks/performance/798w8l1f",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-798w8l1f-ctx8192",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 352.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 352.6 28.5 23.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/798w8l1f",
   "source_url": "https://omlx.ai/benchmarks/performance/798w8l1f",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-798w8l1f-ctx16384",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 298.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 298.6 21.4 23.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/798w8l1f",
   "source_url": "https://omlx.ai/benchmarks/performance/798w8l1f",
   "tg_tokens": null,
   "tps": 21.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-zfef54xl-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: oxtcko67; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 939.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 939.7 79.0 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zfef54xl",
   "source_url": "https://omlx.ai/benchmarks/performance/zfef54xl",
   "tg_tokens": null,
   "tps": 79.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-zfef54xl-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: oxtcko67; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 976.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 976.6 75.3 22.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zfef54xl",
   "source_url": "https://omlx.ai/benchmarks/performance/zfef54xl",
   "tg_tokens": null,
   "tps": 75.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-7s380v6m-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 203.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 203.7 21.0 5.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7s380v6m",
   "source_url": "https://omlx.ai/benchmarks/performance/7s380v6m",
   "tg_tokens": null,
   "tps": 21.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-7s380v6m-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 215.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 215.2 20.1 6.2 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7s380v6m",
   "source_url": "https://omlx.ai/benchmarks/performance/7s380v6m",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-aji04kci-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 6factqat; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 871.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 871.7 135.0 19.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aji04kci",
   "source_url": "https://omlx.ai/benchmarks/performance/aji04kci",
   "tg_tokens": null,
   "tps": 135.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-aji04kci-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 6factqat; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2382.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,382 131.2 20.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aji04kci",
   "source_url": "https://omlx.ai/benchmarks/performance/aji04kci",
   "tg_tokens": null,
   "tps": 131.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-g7k1wzc7-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 590.5,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-fp16-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 590.5 60.2 20.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g7k1wzc7",
   "source_url": "https://omlx.ai/benchmarks/performance/g7k1wzc7",
   "tg_tokens": null,
   "tps": 60.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-g7k1wzc7-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 643.1,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-fp16-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 643.1 57.7 20.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g7k1wzc7",
   "source_url": "https://omlx.ai/benchmarks/performance/g7k1wzc7",
   "tg_tokens": null,
   "tps": 57.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-rb40o4yu-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTPLX Optimized Quality FP16 build; oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: l5yw1qmg; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality-FP16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 1k 105.1 13.8 33.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rb40o4yu",
   "source_url": "https://omlx.ai/benchmarks/performance/rb40o4yu",
   "tg_tokens": null,
   "tps": 13.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-rb40o4yu-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTPLX Optimized Quality FP16 build; oMLX community benchmark trace; peak memory 38.0 GB at this context; duplicate traces of the same session: l5yw1qmg; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 91.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality-FP16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 4k 91.4 12.2 38.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rb40o4yu",
   "source_url": "https://omlx.ai/benchmarks/performance/rb40o4yu",
   "tg_tokens": null,
   "tps": 12.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-lygxea9k-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.9 GB at this context; duplicate traces of the same session: 3jxkuac2, l63zrfc9, o6c1w4z5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 465.9,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Qwen3.8-27B-mxfp8 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 465.9 9.7 28.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lygxea9k",
   "source_url": "https://omlx.ai/benchmarks/performance/lygxea9k",
   "tg_tokens": null,
   "tps": 9.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-lygxea9k-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: 3jxkuac2, l63zrfc9, o6c1w4z5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 435.5,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Qwen3.8-27B-mxfp8 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 16k 435.5 9.5 29.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lygxea9k",
   "source_url": "https://omlx.ai/benchmarks/performance/lygxea9k",
   "tg_tokens": null,
   "tps": 9.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-lygxea9k-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.5 GB at this context; duplicate traces of the same session: 3jxkuac2, l63zrfc9, o6c1w4z5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 350.9,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Qwen3.8-27B-mxfp8 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 64k 350.9 8.2 32.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lygxea9k",
   "source_url": "https://omlx.ai/benchmarks/performance/lygxea9k",
   "tg_tokens": null,
   "tps": 8.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-lygxea9k-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 36.5 GB at this context; duplicate traces of the same session: 3jxkuac2, l63zrfc9, o6c1w4z5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 285.3,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Qwen3.8-27B-mxfp8 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 128k 285.3 7.3 36.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lygxea9k",
   "source_url": "https://omlx.ai/benchmarks/performance/lygxea9k",
   "tg_tokens": null,
   "tps": 7.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-7ke3gc62-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: grlqcwbn, qev9byf6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 531.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 531.8 37.7 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ke3gc62",
   "source_url": "https://omlx.ai/benchmarks/performance/7ke3gc62",
   "tg_tokens": null,
   "tps": 37.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-7ke3gc62-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: grlqcwbn, qev9byf6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 680.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 680.4 47.0 71.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ke3gc62",
   "source_url": "https://omlx.ai/benchmarks/performance/7ke3gc62",
   "tg_tokens": null,
   "tps": 47.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-7ke3gc62-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: grlqcwbn, qev9byf6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 823.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 823.6 43.0 71.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ke3gc62",
   "source_url": "https://omlx.ai/benchmarks/performance/7ke3gc62",
   "tg_tokens": null,
   "tps": 43.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-57vcr588-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: 43ry7lzu, 45v8emfm, gcum10gr, ow56y3k4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 548.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 548.1 48.5 70.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/57vcr588",
   "source_url": "https://omlx.ai/benchmarks/performance/57vcr588",
   "tg_tokens": null,
   "tps": 48.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-57vcr588-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: 43ry7lzu, 45v8emfm, gcum10gr, ow56y3k4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 694.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 694.9 40.9 71.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/57vcr588",
   "source_url": "https://omlx.ai/benchmarks/performance/57vcr588",
   "tg_tokens": null,
   "tps": 40.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-57vcr588-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 43ry7lzu, 45v8emfm, gcum10gr, ow56y3k4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 822.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 822.1 37.0 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/57vcr588",
   "source_url": "https://omlx.ai/benchmarks/performance/57vcr588",
   "tg_tokens": null,
   "tps": 37.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-57vcr588-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.3 GB at this context; duplicate traces of the same session: 43ry7lzu, 45v8emfm, gcum10gr, ow56y3k4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 807.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 807.9 58.2 74.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/57vcr588",
   "source_url": "https://omlx.ai/benchmarks/performance/57vcr588",
   "tg_tokens": null,
   "tps": 58.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-57vcr588-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.6 GB at this context; duplicate traces of the same session: 43ry7lzu, 45v8emfm, gcum10gr, ow56y3k4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 794.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 794.2 57.1 78.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/57vcr588",
   "source_url": "https://omlx.ai/benchmarks/performance/57vcr588",
   "tg_tokens": null,
   "tps": 57.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-da2s4xdt-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.2 GB at this context; duplicate traces of the same session: bx53c03w, pvqkqk2k, r9yks0vd, xx69ce7s; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 845.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 845.5 74.1 183.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/da2s4xdt",
   "source_url": "https://omlx.ai/benchmarks/performance/da2s4xdt",
   "tg_tokens": null,
   "tps": 74.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-da2s4xdt-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.9 GB at this context; duplicate traces of the same session: bx53c03w, pvqkqk2k, r9yks0vd, xx69ce7s; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1061.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 1,061 60.7 183.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/da2s4xdt",
   "source_url": "https://omlx.ai/benchmarks/performance/da2s4xdt",
   "tg_tokens": null,
   "tps": 60.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-da2s4xdt-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.9 GB at this context; duplicate traces of the same session: bx53c03w, pvqkqk2k, r9yks0vd, xx69ce7s; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1050.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 1,050 66.5 183.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/da2s4xdt",
   "source_url": "https://omlx.ai/benchmarks/performance/da2s4xdt",
   "tg_tokens": null,
   "tps": 66.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-da2s4xdt-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 184.4 GB at this context; duplicate traces of the same session: bx53c03w, pvqkqk2k, r9yks0vd, xx69ce7s; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1039.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 16k 1,039 61.1 184.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/da2s4xdt",
   "source_url": "https://omlx.ai/benchmarks/performance/da2s4xdt",
   "tg_tokens": null,
   "tps": 61.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-da2s4xdt-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 186.1 GB at this context; duplicate traces of the same session: bx53c03w, pvqkqk2k, r9yks0vd, xx69ce7s; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1029.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 32k 1,029 64.7 186.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/da2s4xdt",
   "source_url": "https://omlx.ai/benchmarks/performance/da2s4xdt",
   "tg_tokens": null,
   "tps": 64.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-uy69et8w-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.1 GB at this context; duplicate traces of the same session: u70fzqsi; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 870.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 870.1 70.1 183.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uy69et8w",
   "source_url": "https://omlx.ai/benchmarks/performance/uy69et8w",
   "tg_tokens": null,
   "tps": 70.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-uy69et8w-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.6 GB at this context; duplicate traces of the same session: u70fzqsi; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1054.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 1,054 79.2 183.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uy69et8w",
   "source_url": "https://omlx.ai/benchmarks/performance/uy69et8w",
   "tg_tokens": null,
   "tps": 79.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-uy69et8w-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.9 GB at this context; duplicate traces of the same session: u70fzqsi; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1042.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 1,042 84.4 183.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uy69et8w",
   "source_url": "https://omlx.ai/benchmarks/performance/uy69et8w",
   "tg_tokens": null,
   "tps": 84.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-uy69et8w-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 184.4 GB at this context; duplicate traces of the same session: u70fzqsi; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1032.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 16k 1,032 66.2 184.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uy69et8w",
   "source_url": "https://omlx.ai/benchmarks/performance/uy69et8w",
   "tg_tokens": null,
   "tps": 66.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-xz0s9g3b-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: 37i3fpgm, 4wr8uhz0, egkhe7ts; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 285.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 5bit; 1k 285.9 40.8 24.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xz0s9g3b",
   "source_url": "https://omlx.ai/benchmarks/performance/xz0s9g3b",
   "tg_tokens": null,
   "tps": 40.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-xz0s9g3b-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: 37i3fpgm, 4wr8uhz0, egkhe7ts; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 5bit; 4k 323.7 47.6 27.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xz0s9g3b",
   "source_url": "https://omlx.ai/benchmarks/performance/xz0s9g3b",
   "tg_tokens": null,
   "tps": 47.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-xz0s9g3b-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: 37i3fpgm, 4wr8uhz0, egkhe7ts; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 316.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 5bit; 8k 316.7 57.5 28.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xz0s9g3b",
   "source_url": "https://omlx.ai/benchmarks/performance/xz0s9g3b",
   "tg_tokens": null,
   "tps": 57.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-xz0s9g3b-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: 37i3fpgm, 4wr8uhz0, egkhe7ts; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 302.3,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 5bit; 16k 302.3 57.7 29.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xz0s9g3b",
   "source_url": "https://omlx.ai/benchmarks/performance/xz0s9g3b",
   "tg_tokens": null,
   "tps": 57.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-8so0q0il-ctx1024",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: 8dllkrgy, uazfipx2; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 314.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 314.4 33.6 20.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8so0q0il",
   "source_url": "https://omlx.ai/benchmarks/performance/8so0q0il",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-8so0q0il-ctx4096",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: 8dllkrgy, uazfipx2; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 376.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 376.4 34.2 22.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8so0q0il",
   "source_url": "https://omlx.ai/benchmarks/performance/8so0q0il",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-8so0q0il-ctx8192",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: 8dllkrgy, uazfipx2; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 352.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 352.6 28.5 23.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8so0q0il",
   "source_url": "https://omlx.ai/benchmarks/performance/8so0q0il",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-8so0q0il-ctx16384",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 8dllkrgy, uazfipx2; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 298.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 298.6 21.4 23.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8so0q0il",
   "source_url": "https://omlx.ai/benchmarks/performance/8so0q0il",
   "tg_tokens": null,
   "tps": 21.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-w30ioxxp-ctx1024",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: n1vwc9jj, uw4lk3ji, xi8sto2p; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 636.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 636.0 53.1 20.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w30ioxxp",
   "source_url": "https://omlx.ai/benchmarks/performance/w30ioxxp",
   "tg_tokens": null,
   "tps": 53.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-w30ioxxp-ctx4096",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: n1vwc9jj, uw4lk3ji, xi8sto2p; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 643.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 643.6 53.9 22.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w30ioxxp",
   "source_url": "https://omlx.ai/benchmarks/performance/w30ioxxp",
   "tg_tokens": null,
   "tps": 53.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-w30ioxxp-ctx8192",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: n1vwc9jj, uw4lk3ji, xi8sto2p; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 537.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 537.3 43.1 23.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w30ioxxp",
   "source_url": "https://omlx.ai/benchmarks/performance/w30ioxxp",
   "tg_tokens": null,
   "tps": 43.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-w30ioxxp-ctx16384",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: n1vwc9jj, uw4lk3ji, xi8sto2p; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 445.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 445.3 33.8 23.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w30ioxxp",
   "source_url": "https://omlx.ai/benchmarks/performance/w30ioxxp",
   "tg_tokens": null,
   "tps": 33.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-a00ae942-ctx4096",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 22.4 GB; TTFT 6960 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 588.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k tokens 588.5 PP tok/s 59.4 TG tok/s 6960 TTFT (ms) 22.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a00ae942",
   "source_url": "https://omlx.ai/benchmarks/performance/a00ae942",
   "tg_tokens": null,
   "tps": 59.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-vn6i7ae1-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.2 GB at this context; duplicate traces of the same session: 4wtj0gk6, nsfevu9x; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2341.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,341 128.8 173.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vn6i7ae1",
   "source_url": "https://omlx.ai/benchmarks/performance/vn6i7ae1",
   "tg_tokens": null,
   "tps": 128.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-vn6i7ae1-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.7 GB at this context; duplicate traces of the same session: 4wtj0gk6, nsfevu9x; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2318.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,318 81.2 173.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vn6i7ae1",
   "source_url": "https://omlx.ai/benchmarks/performance/vn6i7ae1",
   "tg_tokens": null,
   "tps": 81.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-vn6i7ae1-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.9 GB at this context; duplicate traces of the same session: 4wtj0gk6, nsfevu9x; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2264.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,264 141.0 174.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vn6i7ae1",
   "source_url": "https://omlx.ai/benchmarks/performance/vn6i7ae1",
   "tg_tokens": null,
   "tps": 141.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-vm6grucm-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: sr21ar35; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 523.2,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 523.2 54.9 22.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vm6grucm",
   "source_url": "https://omlx.ai/benchmarks/performance/vm6grucm",
   "tg_tokens": null,
   "tps": 54.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-vm6grucm-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: sr21ar35; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 623.2,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 623.2 50.4 23.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vm6grucm",
   "source_url": "https://omlx.ai/benchmarks/performance/vm6grucm",
   "tg_tokens": null,
   "tps": 50.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-m1j1nyli-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 3dgoz1gt; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 884.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 884.6 80.9 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m1j1nyli",
   "source_url": "https://omlx.ai/benchmarks/performance/m1j1nyli",
   "tg_tokens": null,
   "tps": 80.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-m1j1nyli-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: 3dgoz1gt; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 951.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 951.3 84.7 23.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m1j1nyli",
   "source_url": "https://omlx.ai/benchmarks/performance/m1j1nyli",
   "tg_tokens": null,
   "tps": 84.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-8hgoasg0-ctx1024",
   "model": "Qwen3.6-27B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: vluiicgr; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 85.6,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-27B-oQ4-fp16-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 85.6 11.4 16.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8hgoasg0",
   "source_url": "https://omlx.ai/benchmarks/performance/8hgoasg0",
   "tg_tokens": null,
   "tps": 11.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-8hgoasg0-ctx4096",
   "model": "Qwen3.6-27B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: vluiicgr; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 88.5,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-27B-oQ4-fp16-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 88.5 11.0 17.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8hgoasg0",
   "source_url": "https://omlx.ai/benchmarks/performance/8hgoasg0",
   "tg_tokens": null,
   "tps": 11.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-f4dj1zx8-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 590.5,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-fp16-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 590.5 60.2 20.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f4dj1zx8",
   "source_url": "https://omlx.ai/benchmarks/performance/f4dj1zx8",
   "tg_tokens": null,
   "tps": 60.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-f4dj1zx8-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 643.1,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-fp16-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 643.1 57.7 20.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f4dj1zx8",
   "source_url": "https://omlx.ai/benchmarks/performance/f4dj1zx8",
   "tg_tokens": null,
   "tps": 57.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-v73ywvnz-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: f424r2rj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 59.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 59.0 10.8 18.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v73ywvnz",
   "source_url": "https://omlx.ai/benchmarks/performance/v73ywvnz",
   "tg_tokens": null,
   "tps": 10.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-v73ywvnz-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: f424r2rj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 57.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 57.4 9.8 19.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v73ywvnz",
   "source_url": "https://omlx.ai/benchmarks/performance/v73ywvnz",
   "tg_tokens": null,
   "tps": 9.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-vw0ytjce-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.5 GB at this context; duplicate traces of the same session: 8u976a31; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 152.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 152.3 19.8 32.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vw0ytjce",
   "source_url": "https://omlx.ai/benchmarks/performance/vw0ytjce",
   "tg_tokens": null,
   "tps": 19.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-vw0ytjce-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.7 GB at this context; duplicate traces of the same session: 8u976a31; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 138.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 138.7 17.6 35.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vw0ytjce",
   "source_url": "https://omlx.ai/benchmarks/performance/vw0ytjce",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-vw0ytjce-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: 8u976a31; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 133.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 8k 133.6 21.3 35.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vw0ytjce",
   "source_url": "https://omlx.ai/benchmarks/performance/vw0ytjce",
   "tg_tokens": null,
   "tps": 21.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-vw0ytjce-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.6 GB at this context; duplicate traces of the same session: 8u976a31; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k 130.5 23.8 31.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vw0ytjce",
   "source_url": "https://omlx.ai/benchmarks/performance/vw0ytjce",
   "tg_tokens": null,
   "tps": 23.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-jscsrjj6-ctx8192",
   "model": "Qwen3.8-Flash-Next-REAP320",
   "notes": "DWQ MTP Vision build; oMLX community benchmark trace, single run at 8k context; peak memory 49.3 GB; TTFT 43827 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 186.9,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-Flash-Next-REAP320-oQ3e-fp16-DWQ-MTP-Vision-MLX on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 3bit; 8k tokens 186.9 PP tok/s 9.7 TG tok/s 43827 TTFT (ms) 49.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jscsrjj6",
   "source_url": "https://omlx.ai/benchmarks/performance/jscsrjj6",
   "tg_tokens": null,
   "tps": 9.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-3js7lxxh-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.1 GB at this context; duplicate traces of the same session: 0wvyhwiz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 870.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 870.1 70.1 183.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3js7lxxh",
   "source_url": "https://omlx.ai/benchmarks/performance/3js7lxxh",
   "tg_tokens": null,
   "tps": 70.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3js7lxxh-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.6 GB at this context; duplicate traces of the same session: 0wvyhwiz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1054.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 1,054 79.2 183.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3js7lxxh",
   "source_url": "https://omlx.ai/benchmarks/performance/3js7lxxh",
   "tg_tokens": null,
   "tps": 79.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3js7lxxh-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.9 GB at this context; duplicate traces of the same session: 0wvyhwiz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1042.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 1,042 84.4 183.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3js7lxxh",
   "source_url": "https://omlx.ai/benchmarks/performance/3js7lxxh",
   "tg_tokens": null,
   "tps": 84.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-3js7lxxh-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 184.4 GB at this context; duplicate traces of the same session: 0wvyhwiz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1032.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 16k 1,032 66.2 184.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3js7lxxh",
   "source_url": "https://omlx.ai/benchmarks/performance/3js7lxxh",
   "tg_tokens": null,
   "tps": 66.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-reu54dfh-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.2 GB at this context; duplicate traces of the same session: e55enspk, m4r08pd2, u0asdvk1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 898.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 898.0 61.0 180.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/reu54dfh",
   "source_url": "https://omlx.ai/benchmarks/performance/reu54dfh",
   "tg_tokens": null,
   "tps": 61.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-reu54dfh-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.7 GB at this context; duplicate traces of the same session: e55enspk, m4r08pd2, u0asdvk1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1088.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 1,088 57.2 180.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/reu54dfh",
   "source_url": "https://omlx.ai/benchmarks/performance/reu54dfh",
   "tg_tokens": null,
   "tps": 57.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-reu54dfh-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.9 GB at this context; duplicate traces of the same session: e55enspk, m4r08pd2, u0asdvk1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1074.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 1,074 57.1 180.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/reu54dfh",
   "source_url": "https://omlx.ai/benchmarks/performance/reu54dfh",
   "tg_tokens": null,
   "tps": 57.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-reu54dfh-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 182.3 GB at this context; duplicate traces of the same session: e55enspk, m4r08pd2, u0asdvk1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1055.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 32k 1,055 56.7 182.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/reu54dfh",
   "source_url": "https://omlx.ai/benchmarks/performance/reu54dfh",
   "tg_tokens": null,
   "tps": 56.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-fqzc57bm-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.2 GB at this context; duplicate traces of the same session: 7t2c9ktl; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 901.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 901.2 61.4 180.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fqzc57bm",
   "source_url": "https://omlx.ai/benchmarks/performance/fqzc57bm",
   "tg_tokens": null,
   "tps": 61.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-fqzc57bm-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.7 GB at this context; duplicate traces of the same session: 7t2c9ktl; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1089.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 1,089 57.2 180.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fqzc57bm",
   "source_url": "https://omlx.ai/benchmarks/performance/fqzc57bm",
   "tg_tokens": null,
   "tps": 57.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1tzzoovs-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.2 GB at this context; duplicate traces of the same session: x0tk0ixb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 899.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 899.3 61.6 180.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1tzzoovs",
   "source_url": "https://omlx.ai/benchmarks/performance/1tzzoovs",
   "tg_tokens": null,
   "tps": 61.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1tzzoovs-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.7 GB at this context; duplicate traces of the same session: x0tk0ixb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1085.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 1,085 57.5 180.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1tzzoovs",
   "source_url": "https://omlx.ai/benchmarks/performance/1tzzoovs",
   "tg_tokens": null,
   "tps": 57.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-v2kqr3us-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.2 GB at this context; duplicate traces of the same session: 8b9656ca, j65if65u, u66wk5ze; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 862.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 862.0 74.1 183.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v2kqr3us",
   "source_url": "https://omlx.ai/benchmarks/performance/v2kqr3us",
   "tg_tokens": null,
   "tps": 74.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-v2kqr3us-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.9 GB at this context; duplicate traces of the same session: 8b9656ca, j65if65u, u66wk5ze; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1029.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 1,029 58.8 183.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v2kqr3us",
   "source_url": "https://omlx.ai/benchmarks/performance/v2kqr3us",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-v2kqr3us-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 184.4 GB at this context; duplicate traces of the same session: 8b9656ca, j65if65u, u66wk5ze; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1019.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 16k 1,019 72.1 184.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v2kqr3us",
   "source_url": "https://omlx.ai/benchmarks/performance/v2kqr3us",
   "tg_tokens": null,
   "tps": 72.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-v2kqr3us-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 185.8 GB at this context; duplicate traces of the same session: 8b9656ca, j65if65u, u66wk5ze; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1011.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 32k 1,011 67.1 185.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v2kqr3us",
   "source_url": "https://omlx.ai/benchmarks/performance/v2kqr3us",
   "tg_tokens": null,
   "tps": 67.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-u9mye0fz-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.9 GB at this context; duplicate traces of the same session: roga2siw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1029.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 1,029 63.8 183.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u9mye0fz",
   "source_url": "https://omlx.ai/benchmarks/performance/u9mye0fz",
   "tg_tokens": null,
   "tps": 63.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-u9mye0fz-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 184.4 GB at this context; duplicate traces of the same session: roga2siw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1019.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 16k 1,019 79.6 184.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u9mye0fz",
   "source_url": "https://omlx.ai/benchmarks/performance/u9mye0fz",
   "tg_tokens": null,
   "tps": 79.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-u9mye0fz-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 186.1 GB at this context; duplicate traces of the same session: roga2siw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1011.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 32k 1,011 68.5 186.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u9mye0fz",
   "source_url": "https://omlx.ai/benchmarks/performance/u9mye0fz",
   "tg_tokens": null,
   "tps": 68.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-wqjdlbsb-ctx1024",
   "model": "gemma-4-12b-coder-fable5-composer2.5",
   "notes": "oMLX community benchmark trace; peak memory 13.5 GB at this context; duplicate traces of the same session: ufepx1j7; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 151.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-12b-coder-fable5-composer2.5-8bit on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 8bit; 1k 151.4 15.7 13.5 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wqjdlbsb",
   "source_url": "https://omlx.ai/benchmarks/performance/wqjdlbsb",
   "tg_tokens": null,
   "tps": 15.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-wqjdlbsb-ctx4096",
   "model": "gemma-4-12b-coder-fable5-composer2.5",
   "notes": "oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: ufepx1j7; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 147.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-12b-coder-fable5-composer2.5-8bit on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 8bit; 4k 147.6 13.3 13.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wqjdlbsb",
   "source_url": "https://omlx.ai/benchmarks/performance/wqjdlbsb",
   "tg_tokens": null,
   "tps": 13.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-qttud81f-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 14.6 GB; TTFT 3379 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1212.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k tokens 1,212 PP tok/s 39.9 TG tok/s 3379 TTFT (ms) 14.6 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qttud81f",
   "source_url": "https://omlx.ai/benchmarks/performance/qttud81f",
   "tg_tokens": null,
   "tps": 39.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-k4h006qo-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: t018hzzc; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1175.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 1,175 37.8 14.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k4h006qo",
   "source_url": "https://omlx.ai/benchmarks/performance/k4h006qo",
   "tg_tokens": null,
   "tps": 37.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-k4h006qo-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 15.1 GB at this context; duplicate traces of the same session: t018hzzc; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1064.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 16k 1,064 33.4 15.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k4h006qo",
   "source_url": "https://omlx.ai/benchmarks/performance/k4h006qo",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-tqoouayi-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: q0kdhoqz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 787.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 787.0 66.1 16.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tqoouayi",
   "source_url": "https://omlx.ai/benchmarks/performance/tqoouayi",
   "tg_tokens": null,
   "tps": 66.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-tqoouayi-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: q0kdhoqz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 777.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 777.8 52.4 16.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tqoouayi",
   "source_url": "https://omlx.ai/benchmarks/performance/tqoouayi",
   "tg_tokens": null,
   "tps": 52.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-xu0cs19y-ctx4096",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 5bn0ex7z; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 398.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 398.3 40.0 22.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xu0cs19y",
   "source_url": "https://omlx.ai/benchmarks/performance/xu0cs19y",
   "tg_tokens": null,
   "tps": 40.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-xu0cs19y-ctx16384",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: 5bn0ex7z; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 317.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 317.6 24.2 24.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xu0cs19y",
   "source_url": "https://omlx.ai/benchmarks/performance/xu0cs19y",
   "tg_tokens": null,
   "tps": 24.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-6nyo4ai8-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: 0fp1tcf8; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 752.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 8bit; 1k 752.2 46.8 35.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6nyo4ai8",
   "source_url": "https://omlx.ai/benchmarks/performance/6nyo4ai8",
   "tg_tokens": null,
   "tps": 46.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-6nyo4ai8-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: 0fp1tcf8; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 895.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 8bit; 4k 895.0 45.3 37.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6nyo4ai8",
   "source_url": "https://omlx.ai/benchmarks/performance/6nyo4ai8",
   "tg_tokens": null,
   "tps": 45.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-4rp1cxx8-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 21.1 GB; TTFT 65702 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 997.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 64k tokens 997.5 PP tok/s 28.1 TG tok/s 65702 TTFT (ms) 21.1 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4rp1cxx8",
   "source_url": "https://omlx.ai/benchmarks/performance/4rp1cxx8",
   "tg_tokens": null,
   "tps": 28.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-qgytmjz5-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 19.9 GB; TTFT 3007 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1362.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k tokens 1,362 PP tok/s 57.9 TG tok/s 3007 TTFT (ms) 19.9 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qgytmjz5",
   "source_url": "https://omlx.ai/benchmarks/performance/qgytmjz5",
   "tg_tokens": null,
   "tps": 57.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-7xzeqj6y-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.5 GB at this context; duplicate traces of the same session: tyt3k71s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 152.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 152.3 19.8 32.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7xzeqj6y",
   "source_url": "https://omlx.ai/benchmarks/performance/7xzeqj6y",
   "tg_tokens": null,
   "tps": 19.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-7xzeqj6y-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.7 GB at this context; duplicate traces of the same session: tyt3k71s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 138.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 138.7 17.6 35.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7xzeqj6y",
   "source_url": "https://omlx.ai/benchmarks/performance/7xzeqj6y",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-7xzeqj6y-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: tyt3k71s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 133.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 8k 133.6 21.3 35.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7xzeqj6y",
   "source_url": "https://omlx.ai/benchmarks/performance/7xzeqj6y",
   "tg_tokens": null,
   "tps": 21.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-7xzeqj6y-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.6 GB at this context; duplicate traces of the same session: tyt3k71s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k 130.5 23.8 31.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7xzeqj6y",
   "source_url": "https://omlx.ai/benchmarks/performance/7xzeqj6y",
   "tg_tokens": null,
   "tps": 23.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-lplbp4n4-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: isfx526o, ry1h8pnx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 61.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 61.6 7.6 22.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lplbp4n4",
   "source_url": "https://omlx.ai/benchmarks/performance/lplbp4n4",
   "tg_tokens": null,
   "tps": 7.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-lplbp4n4-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: isfx526o, ry1h8pnx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 102.5 13.4 26.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lplbp4n4",
   "source_url": "https://omlx.ai/benchmarks/performance/lplbp4n4",
   "tg_tokens": null,
   "tps": 13.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-lplbp4n4-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: isfx526o, ry1h8pnx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 122.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 8k 122.9 10.0 26.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lplbp4n4",
   "source_url": "https://omlx.ai/benchmarks/performance/lplbp4n4",
   "tg_tokens": null,
   "tps": 10.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-01bdu2yq-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: jzau6xla, mef701pn; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 313.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 313.9 38.7 33.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/01bdu2yq",
   "source_url": "https://omlx.ai/benchmarks/performance/01bdu2yq",
   "tg_tokens": null,
   "tps": 38.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-01bdu2yq-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: jzau6xla, mef701pn; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 334.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 334.0 41.1 37.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/01bdu2yq",
   "source_url": "https://omlx.ai/benchmarks/performance/01bdu2yq",
   "tg_tokens": null,
   "tps": 41.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-01bdu2yq-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.8 GB at this context; duplicate traces of the same session: jzau6xla, mef701pn; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 329.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 8k 329.0 52.5 37.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/01bdu2yq",
   "source_url": "https://omlx.ai/benchmarks/performance/01bdu2yq",
   "tg_tokens": null,
   "tps": 52.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dhlhalzc-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.5 GB at this context; duplicate traces of the same session: 3lsue0qe, 4cewlwkf, a8337bak, d3n8gtou, ifvimglj, n0cx08xv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 890.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 890.7 75.2 17.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dhlhalzc",
   "source_url": "https://omlx.ai/benchmarks/performance/dhlhalzc",
   "tg_tokens": null,
   "tps": 75.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dhlhalzc-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 3lsue0qe, 4cewlwkf, a8337bak, d3n8gtou, ifvimglj, n0cx08xv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1009.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,009 90.5 18.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dhlhalzc",
   "source_url": "https://omlx.ai/benchmarks/performance/dhlhalzc",
   "tg_tokens": null,
   "tps": 90.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dhlhalzc-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 3lsue0qe, 4cewlwkf, a8337bak, d3n8gtou, ifvimglj, n0cx08xv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 957.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 8k 957.4 53.2 19.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dhlhalzc",
   "source_url": "https://omlx.ai/benchmarks/performance/dhlhalzc",
   "tg_tokens": null,
   "tps": 53.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dhlhalzc-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 3lsue0qe, 4cewlwkf, a8337bak, d3n8gtou, ifvimglj, n0cx08xv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 895.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 895.8 60.2 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dhlhalzc",
   "source_url": "https://omlx.ai/benchmarks/performance/dhlhalzc",
   "tg_tokens": null,
   "tps": 60.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dhlhalzc-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 3lsue0qe, 4cewlwkf, a8337bak, d3n8gtou, ifvimglj, n0cx08xv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 805.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 805.2 57.8 20.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dhlhalzc",
   "source_url": "https://omlx.ai/benchmarks/performance/dhlhalzc",
   "tg_tokens": null,
   "tps": 57.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dhlhalzc-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: 3lsue0qe, 4cewlwkf, a8337bak, d3n8gtou, ifvimglj, n0cx08xv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 734.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 64k 734.5 54.7 23.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dhlhalzc",
   "source_url": "https://omlx.ai/benchmarks/performance/dhlhalzc",
   "tg_tokens": null,
   "tps": 54.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dhlhalzc-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: 3lsue0qe, 4cewlwkf, a8337bak, d3n8gtou, ifvimglj, n0cx08xv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 594.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 128k 594.5 48.5 27.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dhlhalzc",
   "source_url": "https://omlx.ai/benchmarks/performance/dhlhalzc",
   "tg_tokens": null,
   "tps": 48.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-6ksyfj8b-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 6z8lny2k, 9vt32t9k, bopjix32, e82dv6yy, k1ezoxvg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 619.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 619.0 80.5 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ksyfj8b",
   "source_url": "https://omlx.ai/benchmarks/performance/6ksyfj8b",
   "tg_tokens": null,
   "tps": 80.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-6ksyfj8b-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: 6z8lny2k, 9vt32t9k, bopjix32, e82dv6yy, k1ezoxvg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 733.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 733.4 68.2 71.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ksyfj8b",
   "source_url": "https://omlx.ai/benchmarks/performance/6ksyfj8b",
   "tg_tokens": null,
   "tps": 68.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-6ksyfj8b-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: 6z8lny2k, 9vt32t9k, bopjix32, e82dv6yy, k1ezoxvg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 832.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 832.5 74.5 71.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ksyfj8b",
   "source_url": "https://omlx.ai/benchmarks/performance/6ksyfj8b",
   "tg_tokens": null,
   "tps": 74.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-6ksyfj8b-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.4 GB at this context; duplicate traces of the same session: 6z8lny2k, 9vt32t9k, bopjix32, e82dv6yy, k1ezoxvg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 820.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 820.1 65.1 72.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ksyfj8b",
   "source_url": "https://omlx.ai/benchmarks/performance/6ksyfj8b",
   "tg_tokens": null,
   "tps": 65.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-6ksyfj8b-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.8 GB at this context; duplicate traces of the same session: 6z8lny2k, 9vt32t9k, bopjix32, e82dv6yy, k1ezoxvg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 810.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 810.8 63.9 74.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ksyfj8b",
   "source_url": "https://omlx.ai/benchmarks/performance/6ksyfj8b",
   "tg_tokens": null,
   "tps": 63.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-6ksyfj8b-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.7 GB at this context; duplicate traces of the same session: 6z8lny2k, 9vt32t9k, bopjix32, e82dv6yy, k1ezoxvg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 798.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 798.7 59.0 78.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ksyfj8b",
   "source_url": "https://omlx.ai/benchmarks/performance/6ksyfj8b",
   "tg_tokens": null,
   "tps": 59.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-8wlcp9th-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 417.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 417.7 23.5 100.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8wlcp9th",
   "source_url": "https://omlx.ai/benchmarks/performance/8wlcp9th",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-8wlcp9th-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.0 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1133.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,133 62.7 101.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8wlcp9th",
   "source_url": "https://omlx.ai/benchmarks/performance/8wlcp9th",
   "tg_tokens": null,
   "tps": 62.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-8wlcp9th-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.1 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2751.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 2,751 137.8 102.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8wlcp9th",
   "source_url": "https://omlx.ai/benchmarks/performance/8wlcp9th",
   "tg_tokens": null,
   "tps": 137.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-8wlcp9th-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.6 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4420.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 4,420 102.3 102.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8wlcp9th",
   "source_url": "https://omlx.ai/benchmarks/performance/8wlcp9th",
   "tg_tokens": null,
   "tps": 102.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-8wlcp9th-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.6 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4579.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 4,579 99.3 104.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8wlcp9th",
   "source_url": "https://omlx.ai/benchmarks/performance/8wlcp9th",
   "tg_tokens": null,
   "tps": 99.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-8wlcp9th-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.1 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4468.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 4,468 100.1 109.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8wlcp9th",
   "source_url": "https://omlx.ai/benchmarks/performance/8wlcp9th",
   "tg_tokens": null,
   "tps": 100.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-8wlcp9th-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.7 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4340.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 4,340 87.2 110.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8wlcp9th",
   "source_url": "https://omlx.ai/benchmarks/performance/8wlcp9th",
   "tg_tokens": null,
   "tps": 87.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-74gwjsvx-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.6 GB at this context; duplicate traces of the same session: 0r3wo10z, 4lrsslcp, uilnyhze, ycv8hwh0, znncf9az; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1769.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 1k 1,769 100.2 142.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/74gwjsvx",
   "source_url": "https://omlx.ai/benchmarks/performance/74gwjsvx",
   "tg_tokens": null,
   "tps": 100.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-74gwjsvx-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 144.5 GB at this context; duplicate traces of the same session: 0r3wo10z, 4lrsslcp, uilnyhze, ycv8hwh0, znncf9az; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3944.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 8k 3,944 109.8 144.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/74gwjsvx",
   "source_url": "https://omlx.ai/benchmarks/performance/74gwjsvx",
   "tg_tokens": null,
   "tps": 109.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-74gwjsvx-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.7 GB at this context; duplicate traces of the same session: 0r3wo10z, 4lrsslcp, uilnyhze, ycv8hwh0, znncf9az; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4187.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 32k 4,187 80.7 145.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/74gwjsvx",
   "source_url": "https://omlx.ai/benchmarks/performance/74gwjsvx",
   "tg_tokens": null,
   "tps": 80.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-74gwjsvx-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 146.8 GB at this context; duplicate traces of the same session: 0r3wo10z, 4lrsslcp, uilnyhze, ycv8hwh0, znncf9az; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4107.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 64k 4,107 120.5 146.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/74gwjsvx",
   "source_url": "https://omlx.ai/benchmarks/performance/74gwjsvx",
   "tg_tokens": null,
   "tps": 120.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-74gwjsvx-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 150.6 GB at this context; duplicate traces of the same session: 0r3wo10z, 4lrsslcp, uilnyhze, ycv8hwh0, znncf9az; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3999.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 128k 3,999 99.0 150.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/74gwjsvx",
   "source_url": "https://omlx.ai/benchmarks/performance/74gwjsvx",
   "tg_tokens": null,
   "tps": 99.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-74gwjsvx-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 153.3 GB at this context; duplicate traces of the same session: 0r3wo10z, 4lrsslcp, uilnyhze, ycv8hwh0, znncf9az; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3882.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 195k 3,882 81.1 153.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/74gwjsvx",
   "source_url": "https://omlx.ai/benchmarks/performance/74gwjsvx",
   "tg_tokens": null,
   "tps": 81.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-i9l6hski-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.9 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1029.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 1,029 63.8 183.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i9l6hski",
   "source_url": "https://omlx.ai/benchmarks/performance/i9l6hski",
   "tg_tokens": null,
   "tps": 63.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-i9l6hski-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 184.4 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1019.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 16k 1,019 79.6 184.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i9l6hski",
   "source_url": "https://omlx.ai/benchmarks/performance/i9l6hski",
   "tg_tokens": null,
   "tps": 79.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-i9l6hski-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 186.1 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1011.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 32k 1,011 68.5 186.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i9l6hski",
   "source_url": "https://omlx.ai/benchmarks/performance/i9l6hski",
   "tg_tokens": null,
   "tps": 68.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-em6opwt9-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.1 GB at this context; duplicate traces of the same session: e8c288kl, wh4hmhwr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 863.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 863.3 73.4 183.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/em6opwt9",
   "source_url": "https://omlx.ai/benchmarks/performance/em6opwt9",
   "tg_tokens": null,
   "tps": 73.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-em6opwt9-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.6 GB at this context; duplicate traces of the same session: e8c288kl, wh4hmhwr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1041.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 1,041 72.2 183.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/em6opwt9",
   "source_url": "https://omlx.ai/benchmarks/performance/em6opwt9",
   "tg_tokens": null,
   "tps": 72.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-em6opwt9-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.9 GB at this context; duplicate traces of the same session: e8c288kl, wh4hmhwr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1029.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 1,029 73.7 183.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/em6opwt9",
   "source_url": "https://omlx.ai/benchmarks/performance/em6opwt9",
   "tg_tokens": null,
   "tps": 73.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-8obm9vqi-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.1 GB at this context; duplicate traces of the same session: 2hxxo45u; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 884.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 884.5 64.5 183.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8obm9vqi",
   "source_url": "https://omlx.ai/benchmarks/performance/8obm9vqi",
   "tg_tokens": null,
   "tps": 64.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-8obm9vqi-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.6 GB at this context; duplicate traces of the same session: 2hxxo45u; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1048.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 1,048 59.4 183.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8obm9vqi",
   "source_url": "https://omlx.ai/benchmarks/performance/8obm9vqi",
   "tg_tokens": null,
   "tps": 59.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-8obm9vqi-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.9 GB at this context; duplicate traces of the same session: 2hxxo45u; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1035.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 1,035 62.6 183.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8obm9vqi",
   "source_url": "https://omlx.ai/benchmarks/performance/8obm9vqi",
   "tg_tokens": null,
   "tps": 62.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-8obm9vqi-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 184.4 GB at this context; duplicate traces of the same session: 2hxxo45u; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1027.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 16k 1,027 56.3 184.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8obm9vqi",
   "source_url": "https://omlx.ai/benchmarks/performance/8obm9vqi",
   "tg_tokens": null,
   "tps": 56.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-8obm9vqi-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 185.5 GB at this context; duplicate traces of the same session: 2hxxo45u; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1016.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 32k 1,016 56.2 185.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8obm9vqi",
   "source_url": "https://omlx.ai/benchmarks/performance/8obm9vqi",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-v9ncljec-ctx4096",
   "model": "Underdog-27B-1.1",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: g90y4z9e, y471v46g; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 769.3,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Underdog-27B-1.1 on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 unknown; 4k 769.3 24.6 17.0 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v9ncljec",
   "source_url": "https://omlx.ai/benchmarks/performance/v9ncljec",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-v9ncljec-ctx16384",
   "model": "Underdog-27B-1.1",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: g90y4z9e, y471v46g; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 628.5,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Underdog-27B-1.1 on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 unknown; 16k 628.5 21.8 17.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v9ncljec",
   "source_url": "https://omlx.ai/benchmarks/performance/v9ncljec",
   "tg_tokens": null,
   "tps": 21.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-v9ncljec-ctx65536",
   "model": "Underdog-27B-1.1",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: g90y4z9e, y471v46g; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 386.6,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Underdog-27B-1.1 on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 unknown; 64k 386.6 9.7 20.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v9ncljec",
   "source_url": "https://omlx.ai/benchmarks/performance/v9ncljec",
   "tg_tokens": null,
   "tps": 9.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hrzf8db1-ctx1024",
   "model": "gemma-4-31B-it-uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: 22w6ozml, xnxsdtll; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 254.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-31B-it-uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 254.0 19.3 32.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hrzf8db1",
   "source_url": "https://omlx.ai/benchmarks/performance/hrzf8db1",
   "tg_tokens": null,
   "tps": 19.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hrzf8db1-ctx4096",
   "model": "gemma-4-31B-it-uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.0 GB at this context; duplicate traces of the same session: 22w6ozml, xnxsdtll; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 258.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-31B-it-uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 258.9 19.0 33.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hrzf8db1",
   "source_url": "https://omlx.ai/benchmarks/performance/hrzf8db1",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hrzf8db1-ctx8192",
   "model": "gemma-4-31B-it-uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.6 GB at this context; duplicate traces of the same session: 22w6ozml, xnxsdtll; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 256.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-31B-it-uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 8k 256.2 18.8 33.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hrzf8db1",
   "source_url": "https://omlx.ai/benchmarks/performance/hrzf8db1",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-v1p72jdc-ctx16384",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 24.2 GB; TTFT 48203 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 339.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k tokens 339.9 PP tok/s 27.9 TG tok/s 48203 TTFT (ms) 24.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v1p72jdc",
   "source_url": "https://omlx.ai/benchmarks/performance/v1p72jdc",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-04ngj0yx-ctx4096",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 9cc84qde; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 428.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 428.8 33.9 22.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/04ngj0yx",
   "source_url": "https://omlx.ai/benchmarks/performance/04ngj0yx",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-04ngj0yx-ctx16384",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: 9cc84qde; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 336.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 336.3 32.5 24.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/04ngj0yx",
   "source_url": "https://omlx.ai/benchmarks/performance/04ngj0yx",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-5z3kw9bw-ctx4096",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 1mwwx38l; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 429.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 429.2 51.5 22.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5z3kw9bw",
   "source_url": "https://omlx.ai/benchmarks/performance/5z3kw9bw",
   "tg_tokens": null,
   "tps": 51.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-5z3kw9bw-ctx16384",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: 1mwwx38l; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 327.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 327.9 27.5 24.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5z3kw9bw",
   "source_url": "https://omlx.ai/benchmarks/performance/5z3kw9bw",
   "tg_tokens": null,
   "tps": 27.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-plsv67a7-ctx4096",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 5hi7fv86; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 429.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 429.1 41.9 22.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/plsv67a7",
   "source_url": "https://omlx.ai/benchmarks/performance/plsv67a7",
   "tg_tokens": null,
   "tps": 41.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-plsv67a7-ctx16384",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: 5hi7fv86; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 329.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 329.2 34.3 24.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/plsv67a7",
   "source_url": "https://omlx.ai/benchmarks/performance/plsv67a7",
   "tg_tokens": null,
   "tps": 34.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-rioq3sxy-ctx4096",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: vabh9pqo; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 442.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 442.5 42.4 22.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rioq3sxy",
   "source_url": "https://omlx.ai/benchmarks/performance/rioq3sxy",
   "tg_tokens": null,
   "tps": 42.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-rioq3sxy-ctx16384",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: vabh9pqo; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 306.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 306.0 24.4 24.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rioq3sxy",
   "source_url": "https://omlx.ai/benchmarks/performance/rioq3sxy",
   "tg_tokens": null,
   "tps": 24.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-yc61c3i0-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: c3wkzn93, cdcma0bl, ec3sfqj8, expv17em, flc65vml, mmuyl8nk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3837.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 3,837 186.8 26.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yc61c3i0",
   "source_url": "https://omlx.ai/benchmarks/performance/yc61c3i0",
   "tg_tokens": null,
   "tps": 186.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yc61c3i0-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: c3wkzn93, cdcma0bl, ec3sfqj8, expv17em, flc65vml, mmuyl8nk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5079.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 5,079 167.0 26.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yc61c3i0",
   "source_url": "https://omlx.ai/benchmarks/performance/yc61c3i0",
   "tg_tokens": null,
   "tps": 167.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-yc61c3i0-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: c3wkzn93, cdcma0bl, ec3sfqj8, expv17em, flc65vml, mmuyl8nk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5119.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 5,119 174.0 26.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yc61c3i0",
   "source_url": "https://omlx.ai/benchmarks/performance/yc61c3i0",
   "tg_tokens": null,
   "tps": 174.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-yc61c3i0-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.0 GB at this context; duplicate traces of the same session: c3wkzn93, cdcma0bl, ec3sfqj8, expv17em, flc65vml, mmuyl8nk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4830.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 4,830 202.5 27.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yc61c3i0",
   "source_url": "https://omlx.ai/benchmarks/performance/yc61c3i0",
   "tg_tokens": null,
   "tps": 202.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-yc61c3i0-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: c3wkzn93, cdcma0bl, ec3sfqj8, expv17em, flc65vml, mmuyl8nk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4056.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 4,056 247.1 27.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yc61c3i0",
   "source_url": "https://omlx.ai/benchmarks/performance/yc61c3i0",
   "tg_tokens": null,
   "tps": 247.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-yc61c3i0-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: c3wkzn93, cdcma0bl, ec3sfqj8, expv17em, flc65vml, mmuyl8nk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3108.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 3,108 134.2 28.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yc61c3i0",
   "source_url": "https://omlx.ai/benchmarks/performance/yc61c3i0",
   "tg_tokens": null,
   "tps": 134.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-yc61c3i0-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: c3wkzn93, cdcma0bl, ec3sfqj8, expv17em, flc65vml, mmuyl8nk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2086.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 2,086 115.8 29.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yc61c3i0",
   "source_url": "https://omlx.ai/benchmarks/performance/yc61c3i0",
   "tg_tokens": null,
   "tps": 115.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cshmj64v-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: 42zg5wij, kk0rnp1x, wbozk4od, ysekts87, zkcwmapv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1341.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 1,341 100.4 26.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cshmj64v",
   "source_url": "https://omlx.ai/benchmarks/performance/cshmj64v",
   "tg_tokens": null,
   "tps": 100.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-cshmj64v-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: 42zg5wij, kk0rnp1x, wbozk4od, ysekts87, zkcwmapv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3399.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 3,399 189.4 26.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cshmj64v",
   "source_url": "https://omlx.ai/benchmarks/performance/cshmj64v",
   "tg_tokens": null,
   "tps": 189.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cshmj64v-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: 42zg5wij, kk0rnp1x, wbozk4od, ysekts87, zkcwmapv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5103.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 5,103 48.3 26.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cshmj64v",
   "source_url": "https://omlx.ai/benchmarks/performance/cshmj64v",
   "tg_tokens": null,
   "tps": 48.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cshmj64v-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.0 GB at this context; duplicate traces of the same session: 42zg5wij, kk0rnp1x, wbozk4od, ysekts87, zkcwmapv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4828.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 4,828 127.8 27.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cshmj64v",
   "source_url": "https://omlx.ai/benchmarks/performance/cshmj64v",
   "tg_tokens": null,
   "tps": 127.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cshmj64v-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: 42zg5wij, kk0rnp1x, wbozk4od, ysekts87, zkcwmapv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4104.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 4,104 210.5 27.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cshmj64v",
   "source_url": "https://omlx.ai/benchmarks/performance/cshmj64v",
   "tg_tokens": null,
   "tps": 210.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cshmj64v-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: 42zg5wij, kk0rnp1x, wbozk4od, ysekts87, zkcwmapv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3088.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 3,088 116.2 28.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cshmj64v",
   "source_url": "https://omlx.ai/benchmarks/performance/cshmj64v",
   "tg_tokens": null,
   "tps": 116.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-cshmj64v-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: 42zg5wij, kk0rnp1x, wbozk4od, ysekts87, zkcwmapv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2098.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 2,098 111.9 29.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cshmj64v",
   "source_url": "https://omlx.ai/benchmarks/performance/cshmj64v",
   "tg_tokens": null,
   "tps": 111.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ruc0m1t2-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 6r6zlotm, bfmiyewj, dh8xmoj1, ghx4z2g4, r3kg5tzq, si8mhf0j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 717.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 717.4 40.4 23.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ruc0m1t2",
   "source_url": "https://omlx.ai/benchmarks/performance/ruc0m1t2",
   "tg_tokens": null,
   "tps": 40.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ruc0m1t2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: 6r6zlotm, bfmiyewj, dh8xmoj1, ghx4z2g4, r3kg5tzq, si8mhf0j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 780.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 780.2 45.0 24.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ruc0m1t2",
   "source_url": "https://omlx.ai/benchmarks/performance/ruc0m1t2",
   "tg_tokens": null,
   "tps": 45.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ruc0m1t2-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: 6r6zlotm, bfmiyewj, dh8xmoj1, ghx4z2g4, r3kg5tzq, si8mhf0j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 803.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 803.1 56.6 24.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ruc0m1t2",
   "source_url": "https://omlx.ai/benchmarks/performance/ruc0m1t2",
   "tg_tokens": null,
   "tps": 56.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ruc0m1t2-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.2 GB at this context; duplicate traces of the same session: 6r6zlotm, bfmiyewj, dh8xmoj1, ghx4z2g4, r3kg5tzq, si8mhf0j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 797.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 797.2 62.5 25.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ruc0m1t2",
   "source_url": "https://omlx.ai/benchmarks/performance/ruc0m1t2",
   "tg_tokens": null,
   "tps": 62.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ruc0m1t2-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: 6r6zlotm, bfmiyewj, dh8xmoj1, ghx4z2g4, r3kg5tzq, si8mhf0j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 708.3,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 708.3 111.0 26.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ruc0m1t2",
   "source_url": "https://omlx.ai/benchmarks/performance/ruc0m1t2",
   "tg_tokens": null,
   "tps": 111.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ruc0m1t2-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.7 GB at this context; duplicate traces of the same session: 6r6zlotm, bfmiyewj, dh8xmoj1, ghx4z2g4, r3kg5tzq, si8mhf0j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 602.6,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 602.6 51.3 28.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ruc0m1t2",
   "source_url": "https://omlx.ai/benchmarks/performance/ruc0m1t2",
   "tg_tokens": null,
   "tps": 51.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ruc0m1t2-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.4 GB at this context; duplicate traces of the same session: 6r6zlotm, bfmiyewj, dh8xmoj1, ghx4z2g4, r3kg5tzq, si8mhf0j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 452.3,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 452.3 42.9 33.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ruc0m1t2",
   "source_url": "https://omlx.ai/benchmarks/performance/ruc0m1t2",
   "tg_tokens": null,
   "tps": 42.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-s4pdu94t-ctx65536",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 42.3 GB; TTFT 265340 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 247.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 64k tokens 247.0 PP tok/s 21.9 TG tok/s 265340 TTFT (ms) 42.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s4pdu94t",
   "source_url": "https://omlx.ai/benchmarks/performance/s4pdu94t",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-3wwpekln-ctx32768",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 41.3 GB; TTFT 143905 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 227.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 32k tokens 227.7 PP tok/s 18.7 TG tok/s 143905 TTFT (ms) 41.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3wwpekln",
   "source_url": "https://omlx.ai/benchmarks/performance/3wwpekln",
   "tg_tokens": null,
   "tps": 18.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-z2htwmae-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: ni48turp, qaw1xdu3, s12blyy2, scfgivhc, v1488b50; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 417.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 417.7 23.5 100.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z2htwmae",
   "source_url": "https://omlx.ai/benchmarks/performance/z2htwmae",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-z2htwmae-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.0 GB at this context; duplicate traces of the same session: ni48turp, qaw1xdu3, s12blyy2, scfgivhc, v1488b50; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1133.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,133 62.7 101.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z2htwmae",
   "source_url": "https://omlx.ai/benchmarks/performance/z2htwmae",
   "tg_tokens": null,
   "tps": 62.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-z2htwmae-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.1 GB at this context; duplicate traces of the same session: ni48turp, qaw1xdu3, s12blyy2, scfgivhc, v1488b50; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2751.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 2,751 137.8 102.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z2htwmae",
   "source_url": "https://omlx.ai/benchmarks/performance/z2htwmae",
   "tg_tokens": null,
   "tps": 137.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-z2htwmae-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.6 GB at this context; duplicate traces of the same session: ni48turp, qaw1xdu3, s12blyy2, scfgivhc, v1488b50; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4420.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 4,420 102.3 102.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z2htwmae",
   "source_url": "https://omlx.ai/benchmarks/performance/z2htwmae",
   "tg_tokens": null,
   "tps": 102.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-z2htwmae-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.6 GB at this context; duplicate traces of the same session: ni48turp, qaw1xdu3, s12blyy2, scfgivhc, v1488b50; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4579.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 4,579 99.3 104.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z2htwmae",
   "source_url": "https://omlx.ai/benchmarks/performance/z2htwmae",
   "tg_tokens": null,
   "tps": 99.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-z2htwmae-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.1 GB at this context; duplicate traces of the same session: ni48turp, qaw1xdu3, s12blyy2, scfgivhc, v1488b50; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4468.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 4,468 100.1 109.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z2htwmae",
   "source_url": "https://omlx.ai/benchmarks/performance/z2htwmae",
   "tg_tokens": null,
   "tps": 100.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-z2htwmae-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.7 GB at this context; duplicate traces of the same session: ni48turp, qaw1xdu3, s12blyy2, scfgivhc, v1488b50; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4340.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 4,340 87.2 110.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z2htwmae",
   "source_url": "https://omlx.ai/benchmarks/performance/z2htwmae",
   "tg_tokens": null,
   "tps": 87.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-twkatlef-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.1 GB at this context; duplicate traces of the same session: 8h17t9y3, ziimlz2g; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 884.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 884.5 64.5 183.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/twkatlef",
   "source_url": "https://omlx.ai/benchmarks/performance/twkatlef",
   "tg_tokens": null,
   "tps": 64.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-twkatlef-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.6 GB at this context; duplicate traces of the same session: 8h17t9y3, ziimlz2g; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1048.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 1,048 59.4 183.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/twkatlef",
   "source_url": "https://omlx.ai/benchmarks/performance/twkatlef",
   "tg_tokens": null,
   "tps": 59.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-twkatlef-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.9 GB at this context; duplicate traces of the same session: 8h17t9y3, ziimlz2g; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1035.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 1,035 62.6 183.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/twkatlef",
   "source_url": "https://omlx.ai/benchmarks/performance/twkatlef",
   "tg_tokens": null,
   "tps": 62.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-twkatlef-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 184.4 GB at this context; duplicate traces of the same session: 8h17t9y3, ziimlz2g; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1027.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 16k 1,027 56.3 184.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/twkatlef",
   "source_url": "https://omlx.ai/benchmarks/performance/twkatlef",
   "tg_tokens": null,
   "tps": 56.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-twkatlef-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 185.5 GB at this context; duplicate traces of the same session: 8h17t9y3, ziimlz2g; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1016.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 32k 1,016 56.2 185.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/twkatlef",
   "source_url": "https://omlx.ai/benchmarks/performance/twkatlef",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-dnsb7cbc-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.2 GB at this context; duplicate traces of the same session: 38ucm070, b896v7yh, p5085iqp, v1g1agig; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 900.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 900.6 65.2 183.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dnsb7cbc",
   "source_url": "https://omlx.ai/benchmarks/performance/dnsb7cbc",
   "tg_tokens": null,
   "tps": 65.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-dnsb7cbc-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.6 GB at this context; duplicate traces of the same session: 38ucm070, b896v7yh, p5085iqp, v1g1agig; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1065.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 1,065 62.0 183.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dnsb7cbc",
   "source_url": "https://omlx.ai/benchmarks/performance/dnsb7cbc",
   "tg_tokens": null,
   "tps": 62.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-dnsb7cbc-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.9 GB at this context; duplicate traces of the same session: 38ucm070, b896v7yh, p5085iqp, v1g1agig; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1051.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 1,051 62.4 183.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dnsb7cbc",
   "source_url": "https://omlx.ai/benchmarks/performance/dnsb7cbc",
   "tg_tokens": null,
   "tps": 62.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-dnsb7cbc-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 184.4 GB at this context; duplicate traces of the same session: 38ucm070, b896v7yh, p5085iqp, v1g1agig; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1040.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 16k 1,040 61.8 184.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dnsb7cbc",
   "source_url": "https://omlx.ai/benchmarks/performance/dnsb7cbc",
   "tg_tokens": null,
   "tps": 61.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-dnsb7cbc-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 185.5 GB at this context; duplicate traces of the same session: 38ucm070, b896v7yh, p5085iqp, v1g1agig; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1029.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 32k 1,029 61.1 185.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dnsb7cbc",
   "source_url": "https://omlx.ai/benchmarks/performance/dnsb7cbc",
   "tg_tokens": null,
   "tps": 61.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1iphvo9r-ctx1024",
   "model": "gemma-4-31B-it-uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 263.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-31B-it-uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 263.7 19.1 32.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1iphvo9r",
   "source_url": "https://omlx.ai/benchmarks/performance/1iphvo9r",
   "tg_tokens": null,
   "tps": 19.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1iphvo9r-ctx4096",
   "model": "gemma-4-31B-it-uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.0 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 262.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-31B-it-uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 262.3 18.9 33.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1iphvo9r",
   "source_url": "https://omlx.ai/benchmarks/performance/1iphvo9r",
   "tg_tokens": null,
   "tps": 18.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1iphvo9r-ctx8192",
   "model": "gemma-4-31B-it-uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.6 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 256.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-31B-it-uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 8k 256.6 18.6 33.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1iphvo9r",
   "source_url": "https://omlx.ai/benchmarks/performance/1iphvo9r",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-13y7uanr-ctx1024",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 28.3 GB at this context; duplicate traces of the same session: vgaaqcia, x43ljtuc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 787.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "AREX-2-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 787.2 18.1 28.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/13y7uanr",
   "source_url": "https://omlx.ai/benchmarks/performance/13y7uanr",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-13y7uanr-ctx4096",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: vgaaqcia, x43ljtuc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 825.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "AREX-2-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 825.5 17.5 29.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/13y7uanr",
   "source_url": "https://omlx.ai/benchmarks/performance/13y7uanr",
   "tg_tokens": null,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-13y7uanr-ctx8192",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 29.8 GB at this context; duplicate traces of the same session: vgaaqcia, x43ljtuc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 719.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "AREX-2-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 719.2 16.9 29.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/13y7uanr",
   "source_url": "https://omlx.ai/benchmarks/performance/13y7uanr",
   "tg_tokens": null,
   "tps": 16.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-kw81nq7n-ctx1024",
   "model": "DeepSeek-R1-0528-Qwen3-8B",
   "notes": "oMLX community benchmark trace; peak memory 8.9 GB at this context; duplicate traces of the same session: fc9abl6q; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1626.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "DeepSeek-R1-0528-Qwen3-8B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 8bit; 1k 1,626 33.2 8.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kw81nq7n",
   "source_url": "https://omlx.ai/benchmarks/performance/kw81nq7n",
   "tg_tokens": null,
   "tps": 33.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-kw81nq7n-ctx4096",
   "model": "DeepSeek-R1-0528-Qwen3-8B",
   "notes": "oMLX community benchmark trace; peak memory 9.3 GB at this context; duplicate traces of the same session: fc9abl6q; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1695.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "DeepSeek-R1-0528-Qwen3-8B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 8bit; 4k 1,695 31.3 9.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kw81nq7n",
   "source_url": "https://omlx.ai/benchmarks/performance/kw81nq7n",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-u2m5zix4-ctx1024",
   "model": "DeepSeek-R1-0528-Qwen3-8B",
   "notes": "oMLX community benchmark trace; peak memory 8.9 GB at this context; duplicate traces of the same session: d5aglwpj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1618.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "DeepSeek-R1-0528-Qwen3-8B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 8bit; 1k 1,618 33.2 8.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u2m5zix4",
   "source_url": "https://omlx.ai/benchmarks/performance/u2m5zix4",
   "tg_tokens": null,
   "tps": 33.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-u2m5zix4-ctx4096",
   "model": "DeepSeek-R1-0528-Qwen3-8B",
   "notes": "oMLX community benchmark trace; peak memory 9.3 GB at this context; duplicate traces of the same session: d5aglwpj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1685.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "DeepSeek-R1-0528-Qwen3-8B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 8bit; 4k 1,685 31.2 9.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u2m5zix4",
   "source_url": "https://omlx.ai/benchmarks/performance/u2m5zix4",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-7u60inpu-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.6 GB at this context; duplicate traces of the same session: frdsh24s, q03r61dr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2347.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,347 100.7 178.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7u60inpu",
   "source_url": "https://omlx.ai/benchmarks/performance/7u60inpu",
   "tg_tokens": null,
   "tps": 100.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-7u60inpu-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 179.9 GB at this context; duplicate traces of the same session: frdsh24s, q03r61dr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2322.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,322 101.9 179.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7u60inpu",
   "source_url": "https://omlx.ai/benchmarks/performance/7u60inpu",
   "tg_tokens": null,
   "tps": 101.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-7u60inpu-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 182.8 GB at this context; duplicate traces of the same session: frdsh24s, q03r61dr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2279.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,279 136.4 182.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7u60inpu",
   "source_url": "https://omlx.ai/benchmarks/performance/7u60inpu",
   "tg_tokens": null,
   "tps": 136.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-mksp3zon-ctx1024",
   "model": "Llama-3-Groq-8B-Tool-Use",
   "notes": "oMLX community benchmark trace; peak memory 5.0 GB at this context; duplicate traces of the same session: cwo62b63; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1749.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Llama-3-Groq-8B-Tool-Use-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 1,749 59.8 5.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mksp3zon",
   "source_url": "https://omlx.ai/benchmarks/performance/mksp3zon",
   "tg_tokens": null,
   "tps": 59.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-mksp3zon-ctx4096",
   "model": "Llama-3-Groq-8B-Tool-Use",
   "notes": "oMLX community benchmark trace; peak memory 5.4 GB at this context; duplicate traces of the same session: cwo62b63; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1727.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Llama-3-Groq-8B-Tool-Use-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 1,727 55.1 5.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mksp3zon",
   "source_url": "https://omlx.ai/benchmarks/performance/mksp3zon",
   "tg_tokens": null,
   "tps": 55.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-bscdtboy-ctx4096",
   "model": "Llama-3.2-3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 4.3 GB at this context; duplicate traces of the same session: 5ymp6f9v, dl23cmok; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1669.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Llama-3.2-3B-Instruct-8bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 8bit; 4k 1,669 68.2 4.3 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bscdtboy",
   "source_url": "https://omlx.ai/benchmarks/performance/bscdtboy",
   "tg_tokens": null,
   "tps": 68.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-bscdtboy-ctx8192",
   "model": "Llama-3.2-3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 4.7 GB at this context; duplicate traces of the same session: 5ymp6f9v, dl23cmok; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1562.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Llama-3.2-3B-Instruct-8bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 8bit; 8k 1,562 68.2 4.7 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bscdtboy",
   "source_url": "https://omlx.ai/benchmarks/performance/bscdtboy",
   "tg_tokens": null,
   "tps": 68.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-bscdtboy-ctx16384",
   "model": "Llama-3.2-3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 5.5 GB at this context; duplicate traces of the same session: 5ymp6f9v, dl23cmok; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1309.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Llama-3.2-3B-Instruct-8bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 8bit; 16k 1,309 64.1 5.5 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bscdtboy",
   "source_url": "https://omlx.ai/benchmarks/performance/bscdtboy",
   "tg_tokens": null,
   "tps": 64.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-37tt179i-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1341.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 1,341 100.4 26.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/37tt179i",
   "source_url": "https://omlx.ai/benchmarks/performance/37tt179i",
   "tg_tokens": null,
   "tps": 100.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-37tt179i-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3399.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 3,399 189.4 26.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/37tt179i",
   "source_url": "https://omlx.ai/benchmarks/performance/37tt179i",
   "tg_tokens": null,
   "tps": 189.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-37tt179i-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5103.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 5,103 48.3 26.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/37tt179i",
   "source_url": "https://omlx.ai/benchmarks/performance/37tt179i",
   "tg_tokens": null,
   "tps": 48.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-37tt179i-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4828.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 4,828 127.8 27.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/37tt179i",
   "source_url": "https://omlx.ai/benchmarks/performance/37tt179i",
   "tg_tokens": null,
   "tps": 127.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-37tt179i-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4104.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 4,104 210.5 27.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/37tt179i",
   "source_url": "https://omlx.ai/benchmarks/performance/37tt179i",
   "tg_tokens": null,
   "tps": 210.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-37tt179i-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3088.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 3,088 116.2 28.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/37tt179i",
   "source_url": "https://omlx.ai/benchmarks/performance/37tt179i",
   "tg_tokens": null,
   "tps": 116.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-37tt179i-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2098.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 2,098 111.9 29.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/37tt179i",
   "source_url": "https://omlx.ai/benchmarks/performance/37tt179i",
   "tg_tokens": null,
   "tps": 111.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ekjpbxuv-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: y5grcoeh; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 392.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 392.9 18.2 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ekjpbxuv",
   "source_url": "https://omlx.ai/benchmarks/performance/ekjpbxuv",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ekjpbxuv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: y5grcoeh; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 421.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 421.2 17.9 17.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ekjpbxuv",
   "source_url": "https://omlx.ai/benchmarks/performance/ekjpbxuv",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-qoh81xh5-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 20.9 GB; TTFT 273586 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 239.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 64k tokens 239.5 PP tok/s 7.1 TG tok/s 273586 TTFT (ms) 20.9 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qoh81xh5",
   "source_url": "https://omlx.ai/benchmarks/performance/qoh81xh5",
   "tg_tokens": null,
   "tps": 7.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-1ae3ufb1-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 0ltckjp7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 213.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 213.0 31.9 18.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ae3ufb1",
   "source_url": "https://omlx.ai/benchmarks/performance/1ae3ufb1",
   "tg_tokens": null,
   "tps": 31.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-1ae3ufb1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: 0ltckjp7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 209.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 209.7 43.6 22.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ae3ufb1",
   "source_url": "https://omlx.ai/benchmarks/performance/1ae3ufb1",
   "tg_tokens": null,
   "tps": 43.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-gfvs7959-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: ytl4iq1v; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 220.9 21.4 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gfvs7959",
   "source_url": "https://omlx.ai/benchmarks/performance/gfvs7959",
   "tg_tokens": null,
   "tps": 21.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-gfvs7959-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: ytl4iq1v; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 212.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 212.7 21.4 20.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gfvs7959",
   "source_url": "https://omlx.ai/benchmarks/performance/gfvs7959",
   "tg_tokens": null,
   "tps": 21.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-hdy6yrzw-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 18.9 GB; TTFT 8355 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 490.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k tokens 490.3 PP tok/s 47.8 TG tok/s 8355 TTFT (ms) 18.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hdy6yrzw",
   "source_url": "https://omlx.ai/benchmarks/performance/hdy6yrzw",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4rpj155g-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 01v8i9xn, 0g10mjii, 5ya7fpmn, h8wbrqea, tnf40zly, vj2cucs7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 840.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 840.4 66.3 23.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4rpj155g",
   "source_url": "https://omlx.ai/benchmarks/performance/4rpj155g",
   "tg_tokens": null,
   "tps": 66.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4rpj155g-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: 01v8i9xn, 0g10mjii, 5ya7fpmn, h8wbrqea, tnf40zly, vj2cucs7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 885.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 885.2 61.8 24.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4rpj155g",
   "source_url": "https://omlx.ai/benchmarks/performance/4rpj155g",
   "tg_tokens": null,
   "tps": 61.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4rpj155g-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: 01v8i9xn, 0g10mjii, 5ya7fpmn, h8wbrqea, tnf40zly, vj2cucs7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 825.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 825.9 53.7 24.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4rpj155g",
   "source_url": "https://omlx.ai/benchmarks/performance/4rpj155g",
   "tg_tokens": null,
   "tps": 53.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4rpj155g-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.2 GB at this context; duplicate traces of the same session: 01v8i9xn, 0g10mjii, 5ya7fpmn, h8wbrqea, tnf40zly, vj2cucs7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 779.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 779.7 81.1 25.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4rpj155g",
   "source_url": "https://omlx.ai/benchmarks/performance/4rpj155g",
   "tg_tokens": null,
   "tps": 81.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4rpj155g-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: 01v8i9xn, 0g10mjii, 5ya7fpmn, h8wbrqea, tnf40zly, vj2cucs7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 682.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 682.4 120.1 26.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4rpj155g",
   "source_url": "https://omlx.ai/benchmarks/performance/4rpj155g",
   "tg_tokens": null,
   "tps": 120.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4rpj155g-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.7 GB at this context; duplicate traces of the same session: 01v8i9xn, 0g10mjii, 5ya7fpmn, h8wbrqea, tnf40zly, vj2cucs7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 560.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 560.2 45.7 28.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4rpj155g",
   "source_url": "https://omlx.ai/benchmarks/performance/4rpj155g",
   "tg_tokens": null,
   "tps": 45.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4rpj155g-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.4 GB at this context; duplicate traces of the same session: 01v8i9xn, 0g10mjii, 5ya7fpmn, h8wbrqea, tnf40zly, vj2cucs7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 414.3,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 414.3 38.5 33.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4rpj155g",
   "source_url": "https://omlx.ai/benchmarks/performance/4rpj155g",
   "tg_tokens": null,
   "tps": 38.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-wulzi4nl-ctx131072",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 45.5 GB; TTFT 541990 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 241.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 128k tokens 241.8 PP tok/s 22.0 TG tok/s 541990 TTFT (ms) 45.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wulzi4nl",
   "source_url": "https://omlx.ai/benchmarks/performance/wulzi4nl",
   "tg_tokens": null,
   "tps": 22.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-d5xqjn5b-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: 9th77jkw; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 633.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 4k 633.6 46.5 71.5 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d5xqjn5b",
   "source_url": "https://omlx.ai/benchmarks/performance/d5xqjn5b",
   "tg_tokens": null,
   "tps": 46.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-d5xqjn5b-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: 9th77jkw; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 751.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 16k 751.3 46.0 71.7 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d5xqjn5b",
   "source_url": "https://omlx.ai/benchmarks/performance/d5xqjn5b",
   "tg_tokens": null,
   "tps": 46.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-75r6lebu-ctx1024",
   "model": "RavenXAiLabs-Chaos-Agent-Qwen3.8-27B-Frontier-Intelligence-Injected-OBLITERATED",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: ubgilmh2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 457.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "RavenXAiLabs-Chaos-Agent-Qwen3.8-27B-Frontier-Intelligence-Injected-OBLITERATED-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 457.8 17.4 16.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/75r6lebu",
   "source_url": "https://omlx.ai/benchmarks/performance/75r6lebu",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-75r6lebu-ctx4096",
   "model": "RavenXAiLabs-Chaos-Agent-Qwen3.8-27B-Frontier-Intelligence-Injected-OBLITERATED",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: ubgilmh2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 417.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "RavenXAiLabs-Chaos-Agent-Qwen3.8-27B-Frontier-Intelligence-Injected-OBLITERATED-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 417.7 16.6 17.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/75r6lebu",
   "source_url": "https://omlx.ai/benchmarks/performance/75r6lebu",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-3bzh45zj-ctx1024",
   "model": "Ternary-Bonsai-2-27B-Abliterated-v2",
   "notes": "PQ2_0; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 11.8 GB; TTFT 30988 ms; duplicate traces of the same session: none; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 33.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Ternary-Bonsai-2-27B-Abliterated-v2-PQ2_0-MTP-MLX on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 unknown; 1k tokens 33.0 PP tok/s 4.8 TG tok/s 30988 TTFT (ms) 11.8 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3bzh45zj",
   "source_url": "https://omlx.ai/benchmarks/performance/3bzh45zj",
   "tg_tokens": null,
   "tps": 4.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-4oet2i02-ctx1024",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: 7gcxue8e; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 279.4,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 2bit; 1k 279.4 23.5 11.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4oet2i02",
   "source_url": "https://omlx.ai/benchmarks/performance/4oet2i02",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-4oet2i02-ctx4096",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 14.2 GB at this context; duplicate traces of the same session: 7gcxue8e; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 336.2,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 2bit; 4k 336.2 22.8 14.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4oet2i02",
   "source_url": "https://omlx.ai/benchmarks/performance/4oet2i02",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-vb1tokaw-ctx1024",
   "model": "Ternary-Bonsai-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 9.1 GB at this context; duplicate traces of the same session: j0mtrrqx; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 311.3,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-27B-mlx-2bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 2bit; 1k 311.3 43.8 9.1 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vb1tokaw",
   "source_url": "https://omlx.ai/benchmarks/performance/vb1tokaw",
   "tg_tokens": null,
   "tps": 43.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-vb1tokaw-ctx4096",
   "model": "Ternary-Bonsai-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 12.4 GB at this context; duplicate traces of the same session: j0mtrrqx; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 338.6,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-27B-mlx-2bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 2bit; 4k 338.6 42.9 12.4 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vb1tokaw",
   "source_url": "https://omlx.ai/benchmarks/performance/vb1tokaw",
   "tg_tokens": null,
   "tps": 42.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-17pojrw6-ctx1024",
   "model": "gemma-4-31B-it-uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: 0qnu1z3c; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 263.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-31B-it-uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 263.7 19.1 32.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/17pojrw6",
   "source_url": "https://omlx.ai/benchmarks/performance/17pojrw6",
   "tg_tokens": null,
   "tps": 19.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-17pojrw6-ctx4096",
   "model": "gemma-4-31B-it-uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.0 GB at this context; duplicate traces of the same session: 0qnu1z3c; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 262.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-31B-it-uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 262.3 18.9 33.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/17pojrw6",
   "source_url": "https://omlx.ai/benchmarks/performance/17pojrw6",
   "tg_tokens": null,
   "tps": 18.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-17pojrw6-ctx8192",
   "model": "gemma-4-31B-it-uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.6 GB at this context; duplicate traces of the same session: 0qnu1z3c; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 256.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-31B-it-uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 8k 256.6 18.6 33.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/17pojrw6",
   "source_url": "https://omlx.ai/benchmarks/performance/17pojrw6",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7eb1t7yv-ctx1024",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 28.3 GB at this context; duplicate traces of the same session: dtom22ez; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 349.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "AREX-2-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 349.8 14.2 28.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7eb1t7yv",
   "source_url": "https://omlx.ai/benchmarks/performance/7eb1t7yv",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-7eb1t7yv-ctx65536",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 33.3 GB at this context; duplicate traces of the same session: dtom22ez; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 422.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "AREX-2-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 422.3 9.0 33.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7eb1t7yv",
   "source_url": "https://omlx.ai/benchmarks/performance/7eb1t7yv",
   "tg_tokens": null,
   "tps": 9.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-vmgnsker-ctx1024",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 6hhxz4q2, w10pamuz, zeij286e; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 729.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 729.0 72.5 20.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vmgnsker",
   "source_url": "https://omlx.ai/benchmarks/performance/vmgnsker",
   "tg_tokens": null,
   "tps": 72.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-vmgnsker-ctx4096",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 6hhxz4q2, w10pamuz, zeij286e; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 754.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 754.6 68.6 20.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vmgnsker",
   "source_url": "https://omlx.ai/benchmarks/performance/vmgnsker",
   "tg_tokens": null,
   "tps": 68.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-vmgnsker-ctx16384",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 6hhxz4q2, w10pamuz, zeij286e; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 664.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 664.3 61.5 21.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vmgnsker",
   "source_url": "https://omlx.ai/benchmarks/performance/vmgnsker",
   "tg_tokens": null,
   "tps": 61.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-vmgnsker-ctx32768",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: 6hhxz4q2, w10pamuz, zeij286e; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 569.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 569.2 53.3 23.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vmgnsker",
   "source_url": "https://omlx.ai/benchmarks/performance/vmgnsker",
   "tg_tokens": null,
   "tps": 53.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-z5q2x078-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: hlvq0c03, q4he5k5t, xa3y2cpl; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 726.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 726.5 70.1 20.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z5q2x078",
   "source_url": "https://omlx.ai/benchmarks/performance/z5q2x078",
   "tg_tokens": null,
   "tps": 70.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-z5q2x078-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: hlvq0c03, q4he5k5t, xa3y2cpl; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 750.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 750.8 66.8 20.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z5q2x078",
   "source_url": "https://omlx.ai/benchmarks/performance/z5q2x078",
   "tg_tokens": null,
   "tps": 66.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-z5q2x078-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: hlvq0c03, q4he5k5t, xa3y2cpl; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 661.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 661.3 55.1 21.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z5q2x078",
   "source_url": "https://omlx.ai/benchmarks/performance/z5q2x078",
   "tg_tokens": null,
   "tps": 55.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-z5q2x078-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: hlvq0c03, q4he5k5t, xa3y2cpl; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 565.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 565.7 45.9 23.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z5q2x078",
   "source_url": "https://omlx.ai/benchmarks/performance/z5q2x078",
   "tg_tokens": null,
   "tps": 45.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-lwcusewa-ctx1024",
   "model": "Qwen3.5-4B-OptiQ",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 4.1 GB; TTFT 5633 ms; duplicate traces of the same session: none; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 181.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-OptiQ-4bit on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 4bit; 1k tokens 181.8 PP tok/s 17.2 TG tok/s 5633 TTFT (ms) 4.1 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lwcusewa",
   "source_url": "https://omlx.ai/benchmarks/performance/lwcusewa",
   "tg_tokens": null,
   "tps": 17.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-ce5eoggg-ctx4096",
   "model": "Qwen3.5-4B-OptiQ",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 4.8 GB; TTFT 22756 ms; duplicate traces of the same session: none; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 180.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-OptiQ-4bit on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 4bit; 4k tokens 180.0 PP tok/s 16.6 TG tok/s 22756 TTFT (ms) 4.8 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ce5eoggg",
   "source_url": "https://omlx.ai/benchmarks/performance/ce5eoggg",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-79j71gb2-ctx1024",
   "model": "Qwen3.5-4B-OptiQ",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 4.1 GB; TTFT 5631 ms; duplicate traces of the same session: none; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 181.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-OptiQ-4bit on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 4bit; 1k tokens 181.8 PP tok/s 17.2 TG tok/s 5631 TTFT (ms) 4.1 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/79j71gb2",
   "source_url": "https://omlx.ai/benchmarks/performance/79j71gb2",
   "tg_tokens": null,
   "tps": 17.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-z1bgeqbb-ctx1024",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 8cs49jy9, jl56r7l0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1787.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,787 113.9 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z1bgeqbb",
   "source_url": "https://omlx.ai/benchmarks/performance/z1bgeqbb",
   "tg_tokens": null,
   "tps": 113.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-z1bgeqbb-ctx4096",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 8cs49jy9, jl56r7l0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2110.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 2,110 117.7 21.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z1bgeqbb",
   "source_url": "https://omlx.ai/benchmarks/performance/z1bgeqbb",
   "tg_tokens": null,
   "tps": 117.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-z1bgeqbb-ctx16384",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 8cs49jy9, jl56r7l0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1656.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,656 99.7 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z1bgeqbb",
   "source_url": "https://omlx.ai/benchmarks/performance/z1bgeqbb",
   "tg_tokens": null,
   "tps": 99.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-a55ekoeb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 0x24q0f7, 536hyrwz, 5xhl54h2, 9gnmi207, 9jff8sf4, rsx5uk8s, vj186xmj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 403.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 403.3 16.6 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a55ekoeb",
   "source_url": "https://omlx.ai/benchmarks/performance/a55ekoeb",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-a55ekoeb-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 0x24q0f7, 536hyrwz, 5xhl54h2, 9gnmi207, 9jff8sf4, rsx5uk8s, vj186xmj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 401.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 401.4 16.2 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a55ekoeb",
   "source_url": "https://omlx.ai/benchmarks/performance/a55ekoeb",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-a55ekoeb-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 0x24q0f7, 536hyrwz, 5xhl54h2, 9gnmi207, 9jff8sf4, rsx5uk8s, vj186xmj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 370.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 8k 370.5 15.7 17.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a55ekoeb",
   "source_url": "https://omlx.ai/benchmarks/performance/a55ekoeb",
   "tg_tokens": null,
   "tps": 15.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-a55ekoeb-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 0x24q0f7, 536hyrwz, 5xhl54h2, 9gnmi207, 9jff8sf4, rsx5uk8s, vj186xmj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 358.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 16k 358.6 15.3 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a55ekoeb",
   "source_url": "https://omlx.ai/benchmarks/performance/a55ekoeb",
   "tg_tokens": null,
   "tps": 15.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-a55ekoeb-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 0x24q0f7, 536hyrwz, 5xhl54h2, 9gnmi207, 9jff8sf4, rsx5uk8s, vj186xmj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 359.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 32k 359.0 14.2 18.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a55ekoeb",
   "source_url": "https://omlx.ai/benchmarks/performance/a55ekoeb",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-a55ekoeb-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 0x24q0f7, 536hyrwz, 5xhl54h2, 9gnmi207, 9jff8sf4, rsx5uk8s, vj186xmj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 333.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 64k 333.6 13.5 20.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a55ekoeb",
   "source_url": "https://omlx.ai/benchmarks/performance/a55ekoeb",
   "tg_tokens": null,
   "tps": 13.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-a55ekoeb-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: 0x24q0f7, 536hyrwz, 5xhl54h2, 9gnmi207, 9jff8sf4, rsx5uk8s, vj186xmj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 271.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 128k 271.0 10.1 24.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a55ekoeb",
   "source_url": "https://omlx.ai/benchmarks/performance/a55ekoeb",
   "tg_tokens": null,
   "tps": 10.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-a55ekoeb-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.0 GB at this context; duplicate traces of the same session: 0x24q0f7, 536hyrwz, 5xhl54h2, 9gnmi207, 9jff8sf4, rsx5uk8s, vj186xmj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 195k 224.7 8.8 29.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a55ekoeb",
   "source_url": "https://omlx.ai/benchmarks/performance/a55ekoeb",
   "tg_tokens": null,
   "tps": 8.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9drdkgg6-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.3 GB at this context; duplicate traces of the same session: wa245itz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 820.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 820.1 22.9 28.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9drdkgg6",
   "source_url": "https://omlx.ai/benchmarks/performance/9drdkgg6",
   "tg_tokens": null,
   "tps": 22.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9drdkgg6-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.8 GB at this context; duplicate traces of the same session: wa245itz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 676.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 676.7 22.8 28.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9drdkgg6",
   "source_url": "https://omlx.ai/benchmarks/performance/9drdkgg6",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-83xw42vb-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 36.8 GB; TTFT 35964 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 911.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k tokens 911.1 PP tok/s 68.2 TG tok/s 35964 TTFT (ms) 36.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/83xw42vb",
   "source_url": "https://omlx.ai/benchmarks/performance/83xw42vb",
   "tg_tokens": null,
   "tps": 68.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-t1lv3pu2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 34.9 GB; TTFT 3249 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1261.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k tokens 1,261 PP tok/s 33.0 TG tok/s 3249 TTFT (ms) 34.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t1lv3pu2",
   "source_url": "https://omlx.ai/benchmarks/performance/t1lv3pu2",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nupqr1ig-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: 4fpz56xe; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 760.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 760.1 38.5 33.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nupqr1ig",
   "source_url": "https://omlx.ai/benchmarks/performance/nupqr1ig",
   "tg_tokens": null,
   "tps": 38.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nupqr1ig-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: 4fpz56xe; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 940.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 940.2 40.1 34.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nupqr1ig",
   "source_url": "https://omlx.ai/benchmarks/performance/nupqr1ig",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gzzl1dog-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: hl1g0px3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 827.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 827.8 19.1 28.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gzzl1dog",
   "source_url": "https://omlx.ai/benchmarks/performance/gzzl1dog",
   "tg_tokens": null,
   "tps": 19.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gzzl1dog-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: hl1g0px3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 916.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 916.1 19.0 29.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gzzl1dog",
   "source_url": "https://omlx.ai/benchmarks/performance/gzzl1dog",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ntpbgbft-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 70.8 GB; TTFT 2253 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1818.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 1,818 PP tok/s 83.9 TG tok/s 2253 TTFT (ms) 70.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ntpbgbft",
   "source_url": "https://omlx.ai/benchmarks/performance/ntpbgbft",
   "tg_tokens": null,
   "tps": 83.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ekvoxjzh-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 70.8 GB; TTFT 2271 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1804.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 1,804 PP tok/s 68.8 TG tok/s 2271 TTFT (ms) 70.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ekvoxjzh",
   "source_url": "https://omlx.ai/benchmarks/performance/ekvoxjzh",
   "tg_tokens": null,
   "tps": 68.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xgme8dyz-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 70.9 GB; TTFT 2719 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1506.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 1,506 PP tok/s 46.0 TG tok/s 2719 TTFT (ms) 70.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xgme8dyz",
   "source_url": "https://omlx.ai/benchmarks/performance/xgme8dyz",
   "tg_tokens": null,
   "tps": 46.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9n5aamsk-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 78.5 GB; TTFT 68830 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1904.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k tokens 1,904 PP tok/s 52.6 TG tok/s 68830 TTFT (ms) 78.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9n5aamsk",
   "source_url": "https://omlx.ai/benchmarks/performance/9n5aamsk",
   "tg_tokens": null,
   "tps": 52.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vvlweumy-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 78.3 GB; TTFT 72770 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1801.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k tokens 1,801 PP tok/s 32.8 TG tok/s 72770 TTFT (ms) 78.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vvlweumy",
   "source_url": "https://omlx.ai/benchmarks/performance/vvlweumy",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-920hczsb-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "bf16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: sk9gnuxw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 213.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-bf16-mtp-MLX on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 6bit; 1k 213.6 40.4 26.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/920hczsb",
   "source_url": "https://omlx.ai/benchmarks/performance/920hczsb",
   "tg_tokens": null,
   "tps": 40.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-920hczsb-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "bf16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: sk9gnuxw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 223.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-bf16-mtp-MLX on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 6bit; 4k 223.6 30.6 27.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/920hczsb",
   "source_url": "https://omlx.ai/benchmarks/performance/920hczsb",
   "tg_tokens": null,
   "tps": 30.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-j2yp1wov-ctx1024",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 8k15v84n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 418.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "AREX-2-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 418.6 16.3 28.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j2yp1wov",
   "source_url": "https://omlx.ai/benchmarks/performance/j2yp1wov",
   "tg_tokens": null,
   "tps": 16.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-j2yp1wov-ctx131072",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 38.0 GB at this context; duplicate traces of the same session: 8k15v84n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 298.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "AREX-2-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 298.8 5.7 38.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j2yp1wov",
   "source_url": "https://omlx.ai/benchmarks/performance/j2yp1wov",
   "tg_tokens": null,
   "tps": 5.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-07zrkpql-ctx1024",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 004voij0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 789.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "AREX-2-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 789.6 18.0 28.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/07zrkpql",
   "source_url": "https://omlx.ai/benchmarks/performance/07zrkpql",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-07zrkpql-ctx8192",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 29.8 GB at this context; duplicate traces of the same session: 004voij0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 718.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "AREX-2-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 718.2 16.7 29.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/07zrkpql",
   "source_url": "https://omlx.ai/benchmarks/performance/07zrkpql",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cscp80fi-ctx1024",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 28.3 GB at this context; duplicate traces of the same session: xje9aeey; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 385.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "AREX-2-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 385.2 16.1 28.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cscp80fi",
   "source_url": "https://omlx.ai/benchmarks/performance/cscp80fi",
   "tg_tokens": null,
   "tps": 16.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-cscp80fi-ctx131072",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 38.0 GB at this context; duplicate traces of the same session: xje9aeey; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 324.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "AREX-2-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 324.3 6.2 38.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cscp80fi",
   "source_url": "https://omlx.ai/benchmarks/performance/cscp80fi",
   "tg_tokens": null,
   "tps": 6.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-eyifwdev-ctx1024",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: et9pdi0r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 803.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "AREX-2-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 803.8 17.9 28.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eyifwdev",
   "source_url": "https://omlx.ai/benchmarks/performance/eyifwdev",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-eyifwdev-ctx32768",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 31.3 GB at this context; duplicate traces of the same session: et9pdi0r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 621.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "AREX-2-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 621.0 13.9 31.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eyifwdev",
   "source_url": "https://omlx.ai/benchmarks/performance/eyifwdev",
   "tg_tokens": null,
   "tps": 13.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ac4hbwk0-ctx1024",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 9.9 GB at this context; duplicate traces of the same session: 0yg8438u, jfkd15zo; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3143.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 3,143 109.8 9.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ac4hbwk0",
   "source_url": "https://omlx.ai/benchmarks/performance/ac4hbwk0",
   "tg_tokens": null,
   "tps": 109.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ac4hbwk0-ctx4096",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 9.9 GB at this context; duplicate traces of the same session: 0yg8438u, jfkd15zo; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3410.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 3,410 103.5 9.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ac4hbwk0",
   "source_url": "https://omlx.ai/benchmarks/performance/ac4hbwk0",
   "tg_tokens": null,
   "tps": 103.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ac4hbwk0-ctx8192",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 10.0 GB at this context; duplicate traces of the same session: 0yg8438u, jfkd15zo; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3301.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 8k 3,301 104.2 10.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ac4hbwk0",
   "source_url": "https://omlx.ai/benchmarks/performance/ac4hbwk0",
   "tg_tokens": null,
   "tps": 104.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2aj864zx-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 43.5 GB; TTFT 234667 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 558.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k tokens 558.5 PP tok/s 33.6 TG tok/s 234667 TTFT (ms) 43.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2aj864zx",
   "source_url": "https://omlx.ai/benchmarks/performance/2aj864zx",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0d7trtp5-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.6 GB at this context; duplicate traces of the same session: svftv3ah; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 599.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 599.5 34.4 69.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0d7trtp5",
   "source_url": "https://omlx.ai/benchmarks/performance/0d7trtp5",
   "tg_tokens": null,
   "tps": 34.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0d7trtp5-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: svftv3ah; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1424.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,424 23.1 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0d7trtp5",
   "source_url": "https://omlx.ai/benchmarks/performance/0d7trtp5",
   "tg_tokens": null,
   "tps": 23.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-k06ziuxo-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 30.9 GB; TTFT 163127 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 803.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k tokens 803.5 PP tok/s 12.8 TG tok/s 163127 TTFT (ms) 30.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k06ziuxo",
   "source_url": "https://omlx.ai/benchmarks/performance/k06ziuxo",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ze6x64h1-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 59.8 GB; TTFT 13903 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 589.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k tokens 589.2 PP tok/s 25.3 TG tok/s 13903 TTFT (ms) 59.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ze6x64h1",
   "source_url": "https://omlx.ai/benchmarks/performance/ze6x64h1",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-063gsdg8-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 69.6 GB; TTFT 14126 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 579.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k tokens 579.9 PP tok/s 39.4 TG tok/s 14126 TTFT (ms) 69.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/063gsdg8",
   "source_url": "https://omlx.ai/benchmarks/performance/063gsdg8",
   "tg_tokens": null,
   "tps": 39.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-lzviva5s-ctx1024",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 3.0 GB; TTFT 1040 ms; duplicate traces of the same session: none; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 984.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e2b-it-4bit on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 4bit; 1k tokens 984.6 PP tok/s 38.6 TG tok/s 1040 TTFT (ms) 3.0 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lzviva5s",
   "source_url": "https://omlx.ai/benchmarks/performance/lzviva5s",
   "tg_tokens": null,
   "tps": 38.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-knga02le-ctx1024",
   "model": "GLM-5.3-Flash-Abliterated",
   "notes": "MLX; oMLX community benchmark trace; peak memory 177.5 GB at this context; duplicate traces of the same session: 8v76ab43, ce4ij6ff, i4w5qbyn, ofqapeed, ptf0mfwi, qszbeo21, suy3phrk; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1286.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-Abliterated-MLX-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,286 62.2 177.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/knga02le",
   "source_url": "https://omlx.ai/benchmarks/performance/knga02le",
   "tg_tokens": null,
   "tps": 62.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-knga02le-ctx4096",
   "model": "GLM-5.3-Flash-Abliterated",
   "notes": "MLX; oMLX community benchmark trace; peak memory 178.3 GB at this context; duplicate traces of the same session: 8v76ab43, ce4ij6ff, i4w5qbyn, ofqapeed, ptf0mfwi, qszbeo21, suy3phrk; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1900.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-Abliterated-MLX-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,900 56.6 178.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/knga02le",
   "source_url": "https://omlx.ai/benchmarks/performance/knga02le",
   "tg_tokens": null,
   "tps": 56.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-knga02le-ctx8192",
   "model": "GLM-5.3-Flash-Abliterated",
   "notes": "MLX; oMLX community benchmark trace; peak memory 179.0 GB at this context; duplicate traces of the same session: 8v76ab43, ce4ij6ff, i4w5qbyn, ofqapeed, ptf0mfwi, qszbeo21, suy3phrk; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1918.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-Abliterated-MLX-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 1,918 61.5 179.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/knga02le",
   "source_url": "https://omlx.ai/benchmarks/performance/knga02le",
   "tg_tokens": null,
   "tps": 61.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-knga02le-ctx16384",
   "model": "GLM-5.3-Flash-Abliterated",
   "notes": "MLX; oMLX community benchmark trace; peak memory 179.2 GB at this context; duplicate traces of the same session: 8v76ab43, ce4ij6ff, i4w5qbyn, ofqapeed, ptf0mfwi, qszbeo21, suy3phrk; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1912.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-Abliterated-MLX-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,912 81.2 179.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/knga02le",
   "source_url": "https://omlx.ai/benchmarks/performance/knga02le",
   "tg_tokens": null,
   "tps": 81.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-knga02le-ctx32768",
   "model": "GLM-5.3-Flash-Abliterated",
   "notes": "MLX; oMLX community benchmark trace; peak memory 179.6 GB at this context; duplicate traces of the same session: 8v76ab43, ce4ij6ff, i4w5qbyn, ofqapeed, ptf0mfwi, qszbeo21, suy3phrk; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1900.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-Abliterated-MLX-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,900 82.1 179.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/knga02le",
   "source_url": "https://omlx.ai/benchmarks/performance/knga02le",
   "tg_tokens": null,
   "tps": 82.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-knga02le-ctx65536",
   "model": "GLM-5.3-Flash-Abliterated",
   "notes": "MLX; oMLX community benchmark trace; peak memory 180.2 GB at this context; duplicate traces of the same session: 8v76ab43, ce4ij6ff, i4w5qbyn, ofqapeed, ptf0mfwi, qszbeo21, suy3phrk; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1871.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-Abliterated-MLX-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,871 72.1 180.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/knga02le",
   "source_url": "https://omlx.ai/benchmarks/performance/knga02le",
   "tg_tokens": null,
   "tps": 72.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-knga02le-ctx131072",
   "model": "GLM-5.3-Flash-Abliterated",
   "notes": "MLX; oMLX community benchmark trace; peak memory 181.4 GB at this context; duplicate traces of the same session: 8v76ab43, ce4ij6ff, i4w5qbyn, ofqapeed, ptf0mfwi, qszbeo21, suy3phrk; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1838.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-Abliterated-MLX-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 1,838 81.3 181.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/knga02le",
   "source_url": "https://omlx.ai/benchmarks/performance/knga02le",
   "tg_tokens": null,
   "tps": 81.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-knga02le-ctx199680",
   "model": "GLM-5.3-Flash-Abliterated",
   "notes": "MLX; oMLX community benchmark trace; peak memory 182.7 GB at this context; duplicate traces of the same session: 8v76ab43, ce4ij6ff, i4w5qbyn, ofqapeed, ptf0mfwi, qszbeo21, suy3phrk; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1796.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-Abliterated-MLX-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 1,796 59.5 182.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/knga02le",
   "source_url": "https://omlx.ai/benchmarks/performance/knga02le",
   "tg_tokens": null,
   "tps": 59.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-0x4pcpyv-ctx1024",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1328.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,328 97.8 22.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0x4pcpyv",
   "source_url": "https://omlx.ai/benchmarks/performance/0x4pcpyv",
   "tg_tokens": null,
   "tps": 97.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-0x4pcpyv-ctx4096",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1563.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,563 126.8 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0x4pcpyv",
   "source_url": "https://omlx.ai/benchmarks/performance/0x4pcpyv",
   "tg_tokens": null,
   "tps": 126.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-0x4pcpyv-ctx16384",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1308.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,308 118.9 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0x4pcpyv",
   "source_url": "https://omlx.ai/benchmarks/performance/0x4pcpyv",
   "tg_tokens": null,
   "tps": 118.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-lw0ahyl3-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: 41kv2xj5, ayu7213v, uo5rhpyb; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 666.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 666.7 78.1 3.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lw0ahyl3",
   "source_url": "https://omlx.ai/benchmarks/performance/lw0ahyl3",
   "tg_tokens": null,
   "tps": 78.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-lw0ahyl3-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.1 GB at this context; duplicate traces of the same session: 41kv2xj5, ayu7213v, uo5rhpyb; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 653.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 653.1 72.4 4.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lw0ahyl3",
   "source_url": "https://omlx.ai/benchmarks/performance/lw0ahyl3",
   "tg_tokens": null,
   "tps": 72.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-lw0ahyl3-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 5.2 GB at this context; duplicate traces of the same session: 41kv2xj5, ayu7213v, uo5rhpyb; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 597.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 597.7 64.5 5.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lw0ahyl3",
   "source_url": "https://omlx.ai/benchmarks/performance/lw0ahyl3",
   "tg_tokens": null,
   "tps": 64.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-lw0ahyl3-ctx32768",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: 41kv2xj5, ayu7213v, uo5rhpyb; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 534.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 534.2 53.7 6.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lw0ahyl3",
   "source_url": "https://omlx.ai/benchmarks/performance/lw0ahyl3",
   "tg_tokens": null,
   "tps": 53.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-tgso5c0r-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: 28a0emv4, le66plr7, sj8db1pw; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 363.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 363.9 47.0 5.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tgso5c0r",
   "source_url": "https://omlx.ai/benchmarks/performance/tgso5c0r",
   "tg_tokens": null,
   "tps": 47.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-tgso5c0r-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: 28a0emv4, le66plr7, sj8db1pw; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 361.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 361.4 46.2 6.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tgso5c0r",
   "source_url": "https://omlx.ai/benchmarks/performance/tgso5c0r",
   "tg_tokens": null,
   "tps": 46.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-tgso5c0r-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.6 GB at this context; duplicate traces of the same session: 28a0emv4, le66plr7, sj8db1pw; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 343.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 343.0 41.5 7.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tgso5c0r",
   "source_url": "https://omlx.ai/benchmarks/performance/tgso5c0r",
   "tg_tokens": null,
   "tps": 41.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-tgso5c0r-ctx32768",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 9.1 GB at this context; duplicate traces of the same session: 28a0emv4, le66plr7, sj8db1pw; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 322.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 322.6 37.2 9.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tgso5c0r",
   "source_url": "https://omlx.ai/benchmarks/performance/tgso5c0r",
   "tg_tokens": null,
   "tps": 37.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-shtnkpbh-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: qlw5rqdg, qn0vds2p, zkqmlfku; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 297.9,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 297.9 51.6 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/shtnkpbh",
   "source_url": "https://omlx.ai/benchmarks/performance/shtnkpbh",
   "tg_tokens": null,
   "tps": 51.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-shtnkpbh-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: qlw5rqdg, qn0vds2p, zkqmlfku; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 318.4,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 318.4 53.4 23.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/shtnkpbh",
   "source_url": "https://omlx.ai/benchmarks/performance/shtnkpbh",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-shtnkpbh-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: qlw5rqdg, qn0vds2p, zkqmlfku; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 316.6,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 8k 316.6 56.8 24.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/shtnkpbh",
   "source_url": "https://omlx.ai/benchmarks/performance/shtnkpbh",
   "tg_tokens": null,
   "tps": 56.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-shtnkpbh-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: qlw5rqdg, qn0vds2p, zkqmlfku; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 301.1,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 16k 301.1 63.5 25.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/shtnkpbh",
   "source_url": "https://omlx.ai/benchmarks/performance/shtnkpbh",
   "tg_tokens": null,
   "tps": 63.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ahdivkkd-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 5atueoom, msiewkzs, x3e7vrz6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 250.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 250.1 40.0 18.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ahdivkkd",
   "source_url": "https://omlx.ai/benchmarks/performance/ahdivkkd",
   "tg_tokens": null,
   "tps": 40.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ahdivkkd-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 5atueoom, msiewkzs, x3e7vrz6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 268.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 268.8 40.5 22.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ahdivkkd",
   "source_url": "https://omlx.ai/benchmarks/performance/ahdivkkd",
   "tg_tokens": null,
   "tps": 40.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ahdivkkd-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: 5atueoom, msiewkzs, x3e7vrz6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 264.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 264.4 37.4 22.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ahdivkkd",
   "source_url": "https://omlx.ai/benchmarks/performance/ahdivkkd",
   "tg_tokens": null,
   "tps": 37.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ahdivkkd-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 5atueoom, msiewkzs, x3e7vrz6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 254.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 254.5 35.0 23.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ahdivkkd",
   "source_url": "https://omlx.ai/benchmarks/performance/ahdivkkd",
   "tg_tokens": null,
   "tps": 35.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1qag4zpx-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 0at0302m, 359fzrie, bs3xacxt, ry3uay10, u3tzdpdf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 925.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 925.3 32.5 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1qag4zpx",
   "source_url": "https://omlx.ai/benchmarks/performance/1qag4zpx",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1qag4zpx-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 0at0302m, 359fzrie, bs3xacxt, ry3uay10, u3tzdpdf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1019.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,019 32.1 17.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1qag4zpx",
   "source_url": "https://omlx.ai/benchmarks/performance/1qag4zpx",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1qag4zpx-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 0at0302m, 359fzrie, bs3xacxt, ry3uay10, u3tzdpdf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1014.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,014 31.4 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1qag4zpx",
   "source_url": "https://omlx.ai/benchmarks/performance/1qag4zpx",
   "tg_tokens": null,
   "tps": 31.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1qag4zpx-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: 0at0302m, 359fzrie, bs3xacxt, ry3uay10, u3tzdpdf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 982.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 982.7 30.2 18.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1qag4zpx",
   "source_url": "https://omlx.ai/benchmarks/performance/1qag4zpx",
   "tg_tokens": null,
   "tps": 30.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1qag4zpx-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 0at0302m, 359fzrie, bs3xacxt, ry3uay10, u3tzdpdf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 918.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 918.1 28.7 19.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1qag4zpx",
   "source_url": "https://omlx.ai/benchmarks/performance/1qag4zpx",
   "tg_tokens": null,
   "tps": 28.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1qag4zpx-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 0at0302m, 359fzrie, bs3xacxt, ry3uay10, u3tzdpdf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 807.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 807.5 25.6 21.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1qag4zpx",
   "source_url": "https://omlx.ai/benchmarks/performance/1qag4zpx",
   "tg_tokens": null,
   "tps": 25.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bl8qi45t-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: v0rnn9ku; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 921.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 921.4 32.5 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bl8qi45t",
   "source_url": "https://omlx.ai/benchmarks/performance/bl8qi45t",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bl8qi45t-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: v0rnn9ku; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1019.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,019 32.1 17.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bl8qi45t",
   "source_url": "https://omlx.ai/benchmarks/performance/bl8qi45t",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-t4v2nfm3-ctx4096",
   "model": "Qwen3.8-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: adjh2hxv; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 240.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-mlx-4Bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 240.9 20.3 6.3 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t4v2nfm3",
   "source_url": "https://omlx.ai/benchmarks/performance/t4v2nfm3",
   "tg_tokens": null,
   "tps": 20.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-t4v2nfm3-ctx16384",
   "model": "Qwen3.8-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: adjh2hxv; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 229.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-mlx-4Bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 16k 229.8 18.2 6.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t4v2nfm3",
   "source_url": "https://omlx.ai/benchmarks/performance/t4v2nfm3",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-4axl0wha-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: 3i5c7tfm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 102.2 11.0 28.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4axl0wha",
   "source_url": "https://omlx.ai/benchmarks/performance/4axl0wha",
   "tg_tokens": null,
   "tps": 11.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-4axl0wha-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.0 GB at this context; duplicate traces of the same session: 3i5c7tfm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 433.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 433.2 12.9 30.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4axl0wha",
   "source_url": "https://omlx.ai/benchmarks/performance/4axl0wha",
   "tg_tokens": null,
   "tps": 12.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-40vnks7n-ctx4096",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 7.6 GB at this context; duplicate traces of the same session: 3lfu5wam; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 594.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 594.8 19.8 7.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/40vnks7n",
   "source_url": "https://omlx.ai/benchmarks/performance/40vnks7n",
   "tg_tokens": null,
   "tps": 19.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-40vnks7n-ctx16384",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 8.0 GB at this context; duplicate traces of the same session: 3lfu5wam; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 604.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 604.2 19.4 8.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/40vnks7n",
   "source_url": "https://omlx.ai/benchmarks/performance/40vnks7n",
   "tg_tokens": null,
   "tps": 19.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-78zsfjeb-ctx1024",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: viefdfo9; config=19c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 370.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e on M2 Pro (19c); M2 Pro (19c) \u00b7 32 GB \u00b7 4bit; 1k 370.6 52.3 20.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/78zsfjeb",
   "source_url": "https://omlx.ai/benchmarks/performance/78zsfjeb",
   "tg_tokens": null,
   "tps": 52.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-78zsfjeb-ctx4096",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: viefdfo9; config=19c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 408.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e on M2 Pro (19c); M2 Pro (19c) \u00b7 32 GB \u00b7 4bit; 4k 408.2 51.1 20.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/78zsfjeb",
   "source_url": "https://omlx.ai/benchmarks/performance/78zsfjeb",
   "tg_tokens": null,
   "tps": 51.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-kic8qncs-ctx4096",
   "model": "Llama-3.2-3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 4.3 GB at this context; duplicate traces of the same session: 5q1avfqy, r4vyten8; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1763.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Llama-3.2-3B-Instruct-8bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 8bit; 4k 1,763 84.4 4.3 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kic8qncs",
   "source_url": "https://omlx.ai/benchmarks/performance/kic8qncs",
   "tg_tokens": null,
   "tps": 84.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-kic8qncs-ctx8192",
   "model": "Llama-3.2-3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 4.7 GB at this context; duplicate traces of the same session: 5q1avfqy, r4vyten8; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1585.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Llama-3.2-3B-Instruct-8bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 8bit; 8k 1,585 76.7 4.7 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kic8qncs",
   "source_url": "https://omlx.ai/benchmarks/performance/kic8qncs",
   "tg_tokens": null,
   "tps": 76.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-kic8qncs-ctx16384",
   "model": "Llama-3.2-3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 5.5 GB at this context; duplicate traces of the same session: 5q1avfqy, r4vyten8; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1315.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Llama-3.2-3B-Instruct-8bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 8bit; 16k 1,315 64.1 5.5 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kic8qncs",
   "source_url": "https://omlx.ai/benchmarks/performance/kic8qncs",
   "tg_tokens": null,
   "tps": 64.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2",
   "id": "omlx-ul1vm0km-ctx1024",
   "model": "Qwen3.5-9B-Claude-4.6-Opus-Reasoning-Distilled-v2",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: pchw0io5; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 110.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MLX-Qwen3.5-9B-Claude-4.6-Opus-Reasoning-Distilled-v2-4bit on M2 (8c); M2 (8c) \u00b7 16 GB \u00b7 4bit; 1k 110.1 18.6 5.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ul1vm0km",
   "source_url": "https://omlx.ai/benchmarks/performance/ul1vm0km",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2",
   "id": "omlx-ul1vm0km-ctx4096",
   "model": "Qwen3.5-9B-Claude-4.6-Opus-Reasoning-Distilled-v2",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: pchw0io5; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MLX-Qwen3.5-9B-Claude-4.6-Opus-Reasoning-Distilled-v2-4bit on M2 (8c); M2 (8c) \u00b7 16 GB \u00b7 4bit; 4k 105.7 17.9 6.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ul1vm0km",
   "source_url": "https://omlx.ai/benchmarks/performance/ul1vm0km",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-o4f0c7gk-ctx4096",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "g64; oMLX community benchmark trace; peak memory 9.9 GB at this context; duplicate traces of the same session: ssfz4jfq, uuwchuuz; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1723.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-mlx-6bit-g64 on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 6bit; 4k 1,723 112.8 9.9 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o4f0c7gk",
   "source_url": "https://omlx.ai/benchmarks/performance/o4f0c7gk",
   "tg_tokens": null,
   "tps": 112.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-o4f0c7gk-ctx16384",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "g64; oMLX community benchmark trace; peak memory 10.1 GB at this context; duplicate traces of the same session: ssfz4jfq, uuwchuuz; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1537.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-mlx-6bit-g64 on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 6bit; 16k 1,537 105.5 10.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o4f0c7gk",
   "source_url": "https://omlx.ai/benchmarks/performance/o4f0c7gk",
   "tg_tokens": null,
   "tps": 105.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-o4f0c7gk-ctx131072",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "g64; oMLX community benchmark trace; peak memory 11.6 GB at this context; duplicate traces of the same session: ssfz4jfq, uuwchuuz; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 745.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-mlx-6bit-g64 on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 6bit; 128k 745.8 65.5 11.6 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o4f0c7gk",
   "source_url": "https://omlx.ai/benchmarks/performance/o4f0c7gk",
   "tg_tokens": null,
   "tps": 65.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-k00rwg60-ctx1024",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: kcigegug; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1328.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,328 97.8 22.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k00rwg60",
   "source_url": "https://omlx.ai/benchmarks/performance/k00rwg60",
   "tg_tokens": null,
   "tps": 97.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-k00rwg60-ctx4096",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: kcigegug; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1563.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,563 126.8 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k00rwg60",
   "source_url": "https://omlx.ai/benchmarks/performance/k00rwg60",
   "tg_tokens": null,
   "tps": 126.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-k00rwg60-ctx16384",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: kcigegug; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1308.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,308 118.9 23.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k00rwg60",
   "source_url": "https://omlx.ai/benchmarks/performance/k00rwg60",
   "tg_tokens": null,
   "tps": 118.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-b9ziwc3i-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: djco2tsg; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 892.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 892.2 89.4 19.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b9ziwc3i",
   "source_url": "https://omlx.ai/benchmarks/performance/b9ziwc3i",
   "tg_tokens": null,
   "tps": 89.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-b9ziwc3i-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: djco2tsg; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1025.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 1,025 87.4 20.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b9ziwc3i",
   "source_url": "https://omlx.ai/benchmarks/performance/b9ziwc3i",
   "tg_tokens": null,
   "tps": 87.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-2nhkbkdl-ctx1024",
   "model": "Qwen2.5-Coder-1.5B",
   "notes": "oMLX community benchmark trace; peak memory 2.2 GB at this context; duplicate traces of the same session: 3f44mubi, cmf7yxb9, y29xgg8g, yr1v753v; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1274.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen2.5-Coder-1.5B-8bit on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 8bit; 1k 1,274 94.6 2.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2nhkbkdl",
   "source_url": "https://omlx.ai/benchmarks/performance/2nhkbkdl",
   "tg_tokens": null,
   "tps": 94.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-2nhkbkdl-ctx4096",
   "model": "Qwen2.5-Coder-1.5B",
   "notes": "oMLX community benchmark trace; peak memory 2.3 GB at this context; duplicate traces of the same session: 3f44mubi, cmf7yxb9, y29xgg8g, yr1v753v; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1195.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen2.5-Coder-1.5B-8bit on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 8bit; 4k 1,195 86.8 2.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2nhkbkdl",
   "source_url": "https://omlx.ai/benchmarks/performance/2nhkbkdl",
   "tg_tokens": null,
   "tps": 86.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-2nhkbkdl-ctx8192",
   "model": "Qwen2.5-Coder-1.5B",
   "notes": "oMLX community benchmark trace; peak memory 2.4 GB at this context; duplicate traces of the same session: 3f44mubi, cmf7yxb9, y29xgg8g, yr1v753v; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1136.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen2.5-Coder-1.5B-8bit on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 8bit; 8k 1,136 78.8 2.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2nhkbkdl",
   "source_url": "https://omlx.ai/benchmarks/performance/2nhkbkdl",
   "tg_tokens": null,
   "tps": 78.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-2nhkbkdl-ctx16384",
   "model": "Qwen2.5-Coder-1.5B",
   "notes": "oMLX community benchmark trace; peak memory 2.6 GB at this context; duplicate traces of the same session: 3f44mubi, cmf7yxb9, y29xgg8g, yr1v753v; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 961.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen2.5-Coder-1.5B-8bit on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 8bit; 16k 961.3 66.0 2.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2nhkbkdl",
   "source_url": "https://omlx.ai/benchmarks/performance/2nhkbkdl",
   "tg_tokens": null,
   "tps": 66.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-2nhkbkdl-ctx32768",
   "model": "Qwen2.5-Coder-1.5B",
   "notes": "oMLX community benchmark trace; peak memory 3.0 GB at this context; duplicate traces of the same session: 3f44mubi, cmf7yxb9, y29xgg8g, yr1v753v; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 721.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen2.5-Coder-1.5B-8bit on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 8bit; 32k 721.9 47.4 3.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2nhkbkdl",
   "source_url": "https://omlx.ai/benchmarks/performance/2nhkbkdl",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-8kwe1qyj-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: i8c506jd, l292z9b6, luxqnhai, ujy2izpk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 511.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 511.2 17.7 17.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8kwe1qyj",
   "source_url": "https://omlx.ai/benchmarks/performance/8kwe1qyj",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-8kwe1qyj-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: i8c506jd, l292z9b6, luxqnhai, ujy2izpk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 474.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 474.7 16.7 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8kwe1qyj",
   "source_url": "https://omlx.ai/benchmarks/performance/8kwe1qyj",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-8kwe1qyj-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: i8c506jd, l292z9b6, luxqnhai, ujy2izpk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 382.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 382.0 13.7 20.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8kwe1qyj",
   "source_url": "https://omlx.ai/benchmarks/performance/8kwe1qyj",
   "tg_tokens": null,
   "tps": 13.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-8kwe1qyj-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: i8c506jd, l292z9b6, luxqnhai, ujy2izpk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 308.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 128k 308.0 10.5 24.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8kwe1qyj",
   "source_url": "https://omlx.ai/benchmarks/performance/8kwe1qyj",
   "tg_tokens": null,
   "tps": 10.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-8kwe1qyj-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.0 GB at this context; duplicate traces of the same session: i8c506jd, l292z9b6, luxqnhai, ujy2izpk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 256.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 195k 256.7 9.4 29.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8kwe1qyj",
   "source_url": "https://omlx.ai/benchmarks/performance/8kwe1qyj",
   "tg_tokens": null,
   "tps": 9.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-m4jyepgh-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 1h526vm2, hzhltupz, y47pyhsh; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 291.6,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 291.6 65.4 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m4jyepgh",
   "source_url": "https://omlx.ai/benchmarks/performance/m4jyepgh",
   "tg_tokens": null,
   "tps": 65.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-m4jyepgh-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: 1h526vm2, hzhltupz, y47pyhsh; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 324.9,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 324.9 53.4 23.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m4jyepgh",
   "source_url": "https://omlx.ai/benchmarks/performance/m4jyepgh",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-m4jyepgh-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 1h526vm2, hzhltupz, y47pyhsh; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 317.3,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 8k 317.3 59.0 23.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m4jyepgh",
   "source_url": "https://omlx.ai/benchmarks/performance/m4jyepgh",
   "tg_tokens": null,
   "tps": 59.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-m4jyepgh-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: 1h526vm2, hzhltupz, y47pyhsh; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 302.8,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 16k 302.8 72.1 25.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m4jyepgh",
   "source_url": "https://omlx.ai/benchmarks/performance/m4jyepgh",
   "tg_tokens": null,
   "tps": 72.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-e7cy7a9k-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 5trfhdye, 6ghr6c1m, ya1xf7nr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 283.6,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 283.6 31.9 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e7cy7a9k",
   "source_url": "https://omlx.ai/benchmarks/performance/e7cy7a9k",
   "tg_tokens": null,
   "tps": 31.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-e7cy7a9k-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: 5trfhdye, 6ghr6c1m, ya1xf7nr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 324.6,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 324.6 45.1 23.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e7cy7a9k",
   "source_url": "https://omlx.ai/benchmarks/performance/e7cy7a9k",
   "tg_tokens": null,
   "tps": 45.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-e7cy7a9k-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 5trfhdye, 6ghr6c1m, ya1xf7nr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 317.1,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 8k 317.1 40.2 23.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e7cy7a9k",
   "source_url": "https://omlx.ai/benchmarks/performance/e7cy7a9k",
   "tg_tokens": null,
   "tps": 40.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-e7cy7a9k-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: 5trfhdye, 6ghr6c1m, ya1xf7nr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 301.4,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 16k 301.4 58.8 25.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e7cy7a9k",
   "source_url": "https://omlx.ai/benchmarks/performance/e7cy7a9k",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-sjcwvp7k-ctx1024",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 155.8 GB at this context; duplicate traces of the same session: e70q07d7, exeajhuq, g8e6pld9, hvwbnbbo, p6t142uv, uhvy9u9s, xklnwbt0; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1126.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,126 42.4 155.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sjcwvp7k",
   "source_url": "https://omlx.ai/benchmarks/performance/sjcwvp7k",
   "tg_tokens": null,
   "tps": 42.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-sjcwvp7k-ctx4096",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.3 GB at this context; duplicate traces of the same session: e70q07d7, exeajhuq, g8e6pld9, hvwbnbbo, p6t142uv, uhvy9u9s, xklnwbt0; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1357.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,357 43.4 156.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sjcwvp7k",
   "source_url": "https://omlx.ai/benchmarks/performance/sjcwvp7k",
   "tg_tokens": null,
   "tps": 43.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-sjcwvp7k-ctx8192",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 164.5 GB at this context; duplicate traces of the same session: e70q07d7, exeajhuq, g8e6pld9, hvwbnbbo, p6t142uv, uhvy9u9s, xklnwbt0; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1312.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 1,312 10.7 164.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sjcwvp7k",
   "source_url": "https://omlx.ai/benchmarks/performance/sjcwvp7k",
   "tg_tokens": null,
   "tps": 10.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-sjcwvp7k-ctx16384",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 164.6 GB at this context; duplicate traces of the same session: e70q07d7, exeajhuq, g8e6pld9, hvwbnbbo, p6t142uv, uhvy9u9s, xklnwbt0; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1271.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,271 5.5 164.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sjcwvp7k",
   "source_url": "https://omlx.ai/benchmarks/performance/sjcwvp7k",
   "tg_tokens": null,
   "tps": 5.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-sjcwvp7k-ctx32768",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.5 GB at this context; duplicate traces of the same session: e70q07d7, exeajhuq, g8e6pld9, hvwbnbbo, p6t142uv, uhvy9u9s, xklnwbt0; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1217.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,217 38.6 156.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sjcwvp7k",
   "source_url": "https://omlx.ai/benchmarks/performance/sjcwvp7k",
   "tg_tokens": null,
   "tps": 38.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-sjcwvp7k-ctx65536",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 165.5 GB at this context; duplicate traces of the same session: e70q07d7, exeajhuq, g8e6pld9, hvwbnbbo, p6t142uv, uhvy9u9s, xklnwbt0; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1100.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,100 1.4 165.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sjcwvp7k",
   "source_url": "https://omlx.ai/benchmarks/performance/sjcwvp7k",
   "tg_tokens": null,
   "tps": 1.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-sjcwvp7k-ctx131072",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 157.1 GB at this context; duplicate traces of the same session: e70q07d7, exeajhuq, g8e6pld9, hvwbnbbo, p6t142uv, uhvy9u9s, xklnwbt0; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 893.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 893.1 30.8 157.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sjcwvp7k",
   "source_url": "https://omlx.ai/benchmarks/performance/sjcwvp7k",
   "tg_tokens": null,
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-sjcwvp7k-ctx199680",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 158.1 GB at this context; duplicate traces of the same session: e70q07d7, exeajhuq, g8e6pld9, hvwbnbbo, p6t142uv, uhvy9u9s, xklnwbt0; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 746.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 746.9 32.4 158.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sjcwvp7k",
   "source_url": "https://omlx.ai/benchmarks/performance/sjcwvp7k",
   "tg_tokens": null,
   "tps": 32.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-55gkkp5e-ctx16384",
   "model": "Laguna XS 2.1",
   "notes": "oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: lth0w5vj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2651.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Laguna-XS-2.1-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 2,651 127.3 18.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/55gkkp5e",
   "source_url": "https://omlx.ai/benchmarks/performance/55gkkp5e",
   "tg_tokens": null,
   "tps": 127.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-55gkkp5e-ctx32768",
   "model": "Laguna XS 2.1",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: lth0w5vj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2466.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Laguna-XS-2.1-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 2,466 108.3 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/55gkkp5e",
   "source_url": "https://omlx.ai/benchmarks/performance/55gkkp5e",
   "tg_tokens": null,
   "tps": 108.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-42lh3qwv-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: giv1wt5l; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 558.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 558.7 60.7 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/42lh3qwv",
   "source_url": "https://omlx.ai/benchmarks/performance/42lh3qwv",
   "tg_tokens": null,
   "tps": 60.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-42lh3qwv-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: giv1wt5l; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 589.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 589.6 58.2 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/42lh3qwv",
   "source_url": "https://omlx.ai/benchmarks/performance/42lh3qwv",
   "tg_tokens": null,
   "tps": 58.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-o3ahbd0z-ctx1024",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: 6yx6ussm, ac4tedam, rsplcfui; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 97.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-Abliterated-oQ6e on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 1k 97.9 20.4 26.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o3ahbd0z",
   "source_url": "https://omlx.ai/benchmarks/performance/o3ahbd0z",
   "tg_tokens": null,
   "tps": 20.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-o3ahbd0z-ctx4096",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: 6yx6ussm, ac4tedam, rsplcfui; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 97.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-Abliterated-oQ6e on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 4k 97.3 23.0 28.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o3ahbd0z",
   "source_url": "https://omlx.ai/benchmarks/performance/o3ahbd0z",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-o3ahbd0z-ctx16384",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: 6yx6ussm, ac4tedam, rsplcfui; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 94.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-Abliterated-oQ6e on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 16k 94.1 16.2 29.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o3ahbd0z",
   "source_url": "https://omlx.ai/benchmarks/performance/o3ahbd0z",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-o3ahbd0z-ctx32768",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 32.5 GB at this context; duplicate traces of the same session: 6yx6ussm, ac4tedam, rsplcfui; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 88.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-Abliterated-oQ6e on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 32k 88.8 17.0 32.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o3ahbd0z",
   "source_url": "https://omlx.ai/benchmarks/performance/o3ahbd0z",
   "tg_tokens": null,
   "tps": 17.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-okhzg2bv-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.0 GB at this context; duplicate traces of the same session: 1i6vny7d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 930.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 930.9 31.5 18.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/okhzg2bv",
   "source_url": "https://omlx.ai/benchmarks/performance/okhzg2bv",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-okhzg2bv-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 1i6vny7d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 787.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 787.0 25.4 18.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/okhzg2bv",
   "source_url": "https://omlx.ai/benchmarks/performance/okhzg2bv",
   "tg_tokens": null,
   "tps": 25.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7uwem00d-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.0 GB at this context; duplicate traces of the same session: co11riav; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 904.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 904.7 26.0 18.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7uwem00d",
   "source_url": "https://omlx.ai/benchmarks/performance/7uwem00d",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7uwem00d-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: co11riav; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 756.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 756.1 26.5 18.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7uwem00d",
   "source_url": "https://omlx.ai/benchmarks/performance/7uwem00d",
   "tg_tokens": null,
   "tps": 26.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-jc685o34-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: 0cqxunch, b7e1hqo7, ew8n4nj2, gbcg83ju, tao0j6k7, vk1p9e7y, z5l0z0z3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1552.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,552 60.7 177.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jc685o34",
   "source_url": "https://omlx.ai/benchmarks/performance/jc685o34",
   "tg_tokens": null,
   "tps": 60.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-jc685o34-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: 0cqxunch, b7e1hqo7, ew8n4nj2, gbcg83ju, tao0j6k7, vk1p9e7y, z5l0z0z3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2318.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 2,318 54.5 178.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jc685o34",
   "source_url": "https://omlx.ai/benchmarks/performance/jc685o34",
   "tg_tokens": null,
   "tps": 54.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-jc685o34-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.9 GB at this context; duplicate traces of the same session: 0cqxunch, b7e1hqo7, ew8n4nj2, gbcg83ju, tao0j6k7, vk1p9e7y, z5l0z0z3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2319.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 2,319 78.4 178.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jc685o34",
   "source_url": "https://omlx.ai/benchmarks/performance/jc685o34",
   "tg_tokens": null,
   "tps": 78.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-jc685o34-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.1 GB at this context; duplicate traces of the same session: 0cqxunch, b7e1hqo7, ew8n4nj2, gbcg83ju, tao0j6k7, vk1p9e7y, z5l0z0z3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2315.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 2,315 82.6 179.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jc685o34",
   "source_url": "https://omlx.ai/benchmarks/performance/jc685o34",
   "tg_tokens": null,
   "tps": 82.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-jc685o34-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.4 GB at this context; duplicate traces of the same session: 0cqxunch, b7e1hqo7, ew8n4nj2, gbcg83ju, tao0j6k7, vk1p9e7y, z5l0z0z3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2305.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,305 83.4 179.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jc685o34",
   "source_url": "https://omlx.ai/benchmarks/performance/jc685o34",
   "tg_tokens": null,
   "tps": 83.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-jc685o34-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.1 GB at this context; duplicate traces of the same session: 0cqxunch, b7e1hqo7, ew8n4nj2, gbcg83ju, tao0j6k7, vk1p9e7y, z5l0z0z3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2273.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,273 72.4 180.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jc685o34",
   "source_url": "https://omlx.ai/benchmarks/performance/jc685o34",
   "tg_tokens": null,
   "tps": 72.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-jc685o34-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.3 GB at this context; duplicate traces of the same session: 0cqxunch, b7e1hqo7, ew8n4nj2, gbcg83ju, tao0j6k7, vk1p9e7y, z5l0z0z3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2231.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,231 76.9 181.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jc685o34",
   "source_url": "https://omlx.ai/benchmarks/performance/jc685o34",
   "tg_tokens": null,
   "tps": 76.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-jc685o34-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 182.6 GB at this context; duplicate traces of the same session: 0cqxunch, b7e1hqo7, ew8n4nj2, gbcg83ju, tao0j6k7, vk1p9e7y, z5l0z0z3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2183.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 2,183 59.2 182.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jc685o34",
   "source_url": "https://omlx.ai/benchmarks/performance/jc685o34",
   "tg_tokens": null,
   "tps": 59.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-u98fj1t9-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.3 GB; TTFT 231 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4438.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 4,438 PP tok/s 148.2 TG tok/s 231 TTFT (ms) 19.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u98fj1t9",
   "source_url": "https://omlx.ai/benchmarks/performance/u98fj1t9",
   "tg_tokens": null,
   "tps": 148.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-sj3bo36e-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 0prc788t, lym69uo5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 850.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 850.7 146.5 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sj3bo36e",
   "source_url": "https://omlx.ai/benchmarks/performance/sj3bo36e",
   "tg_tokens": null,
   "tps": 146.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-sj3bo36e-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 0prc788t, lym69uo5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2879.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 2,879 144.8 20.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sj3bo36e",
   "source_url": "https://omlx.ai/benchmarks/performance/sj3bo36e",
   "tg_tokens": null,
   "tps": 144.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-sj3bo36e-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 0prc788t, lym69uo5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4966.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 4,966 121.0 20.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sj3bo36e",
   "source_url": "https://omlx.ai/benchmarks/performance/sj3bo36e",
   "tg_tokens": null,
   "tps": 121.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-scz00zuj-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: bys8r1jp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 546.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.5-122B-A10B-oQ4-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 546.0 70.9 70.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/scz00zuj",
   "source_url": "https://omlx.ai/benchmarks/performance/scz00zuj",
   "tg_tokens": null,
   "tps": 70.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-scz00zuj-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.9 GB at this context; duplicate traces of the same session: bys8r1jp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 637.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.5-122B-A10B-oQ4-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 637.0 63.3 73.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/scz00zuj",
   "source_url": "https://omlx.ai/benchmarks/performance/scz00zuj",
   "tg_tokens": null,
   "tps": 63.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-kbwzorho-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: pf3fy3r9; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 46.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 4bit; 1k 46.9 9.5 18.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kbwzorho",
   "source_url": "https://omlx.ai/benchmarks/performance/kbwzorho",
   "tg_tokens": null,
   "tps": 9.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-kbwzorho-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: pf3fy3r9; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 45.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 4bit; 4k 45.1 8.7 19.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kbwzorho",
   "source_url": "https://omlx.ai/benchmarks/performance/kbwzorho",
   "tg_tokens": null,
   "tps": 8.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-kx10xhgp-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.4 GB at this context; duplicate traces of the same session: c9dcbviu, oxe75dzx, rofvwlt7, shie6ase, zq1lcumc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2992.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 2,992 144.7 100.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kx10xhgp",
   "source_url": "https://omlx.ai/benchmarks/performance/kx10xhgp",
   "tg_tokens": null,
   "tps": 144.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-kx10xhgp-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.0 GB at this context; duplicate traces of the same session: c9dcbviu, oxe75dzx, rofvwlt7, shie6ase, zq1lcumc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5103.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 5,103 137.0 101.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kx10xhgp",
   "source_url": "https://omlx.ai/benchmarks/performance/kx10xhgp",
   "tg_tokens": null,
   "tps": 137.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-kx10xhgp-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.1 GB at this context; duplicate traces of the same session: c9dcbviu, oxe75dzx, rofvwlt7, shie6ase, zq1lcumc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5634.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 5,634 148.7 102.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kx10xhgp",
   "source_url": "https://omlx.ai/benchmarks/performance/kx10xhgp",
   "tg_tokens": null,
   "tps": 148.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-kx10xhgp-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.3 GB at this context; duplicate traces of the same session: c9dcbviu, oxe75dzx, rofvwlt7, shie6ase, zq1lcumc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5527.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 5,527 125.4 104.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kx10xhgp",
   "source_url": "https://omlx.ai/benchmarks/performance/kx10xhgp",
   "tg_tokens": null,
   "tps": 125.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-kx10xhgp-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 108.5 GB at this context; duplicate traces of the same session: c9dcbviu, oxe75dzx, rofvwlt7, shie6ase, zq1lcumc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5343.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,343 109.0 108.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kx10xhgp",
   "source_url": "https://omlx.ai/benchmarks/performance/kx10xhgp",
   "tg_tokens": null,
   "tps": 109.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-kx10xhgp-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.7 GB at this context; duplicate traces of the same session: c9dcbviu, oxe75dzx, rofvwlt7, shie6ase, zq1lcumc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5191.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 5,191 92.8 110.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kx10xhgp",
   "source_url": "https://omlx.ai/benchmarks/performance/kx10xhgp",
   "tg_tokens": null,
   "tps": 92.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-qa4ml5li-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.6 GB at this context; duplicate traces of the same session: 15nulue5, 36irswfi, c24unq7d, fxo8s50c, hivhcyxo; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2281.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 2,281 144.3 142.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qa4ml5li",
   "source_url": "https://omlx.ai/benchmarks/performance/qa4ml5li",
   "tg_tokens": null,
   "tps": 144.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-qa4ml5li-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 144.5 GB at this context; duplicate traces of the same session: 15nulue5, 36irswfi, c24unq7d, fxo8s50c, hivhcyxo; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4658.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 8k 4,658 133.7 144.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qa4ml5li",
   "source_url": "https://omlx.ai/benchmarks/performance/qa4ml5li",
   "tg_tokens": null,
   "tps": 133.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-qa4ml5li-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.1 GB at this context; duplicate traces of the same session: 15nulue5, 36irswfi, c24unq7d, fxo8s50c, hivhcyxo; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4696.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 16k 4,696 111.0 145.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qa4ml5li",
   "source_url": "https://omlx.ai/benchmarks/performance/qa4ml5li",
   "tg_tokens": null,
   "tps": 111.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-qa4ml5li-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 146.8 GB at this context; duplicate traces of the same session: 15nulue5, 36irswfi, c24unq7d, fxo8s50c, hivhcyxo; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4599.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 64k 4,599 123.1 146.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qa4ml5li",
   "source_url": "https://omlx.ai/benchmarks/performance/qa4ml5li",
   "tg_tokens": null,
   "tps": 123.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-qa4ml5li-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 151.2 GB at this context; duplicate traces of the same session: 15nulue5, 36irswfi, c24unq7d, fxo8s50c, hivhcyxo; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4469.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 128k 4,469 104.7 151.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qa4ml5li",
   "source_url": "https://omlx.ai/benchmarks/performance/qa4ml5li",
   "tg_tokens": null,
   "tps": 104.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-qa4ml5li-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 153.3 GB at this context; duplicate traces of the same session: 15nulue5, 36irswfi, c24unq7d, fxo8s50c, hivhcyxo; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4348.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 195k 4,348 85.5 153.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qa4ml5li",
   "source_url": "https://omlx.ai/benchmarks/performance/qa4ml5li",
   "tg_tokens": null,
   "tps": 85.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-bhgr1n4r-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: ej1dhu9n, gwlg35j7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 253.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 253.9 48.4 18.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhgr1n4r",
   "source_url": "https://omlx.ai/benchmarks/performance/bhgr1n4r",
   "tg_tokens": null,
   "tps": 48.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-bhgr1n4r-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: ej1dhu9n, gwlg35j7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 269.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 269.0 53.9 22.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhgr1n4r",
   "source_url": "https://omlx.ai/benchmarks/performance/bhgr1n4r",
   "tg_tokens": null,
   "tps": 53.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-bhgr1n4r-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: ej1dhu9n, gwlg35j7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 247.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 247.7 64.5 22.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhgr1n4r",
   "source_url": "https://omlx.ai/benchmarks/performance/bhgr1n4r",
   "tg_tokens": null,
   "tps": 64.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-c3vrtjp2-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "bf16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 023jh63j, 3zv2or3o, 87oxf0c6, m6c2o3jn, nx6xy87k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 788.2,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 788.2 24.0 22.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c3vrtjp2",
   "source_url": "https://omlx.ai/benchmarks/performance/c3vrtjp2",
   "tg_tokens": null,
   "tps": 24.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-c3vrtjp2-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "bf16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: 023jh63j, 3zv2or3o, 87oxf0c6, m6c2o3jn, nx6xy87k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 869.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 869.6 23.8 23.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c3vrtjp2",
   "source_url": "https://omlx.ai/benchmarks/performance/c3vrtjp2",
   "tg_tokens": null,
   "tps": 23.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-c3vrtjp2-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "bf16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: 023jh63j, 3zv2or3o, 87oxf0c6, m6c2o3jn, nx6xy87k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 856.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 8k 856.8 23.2 24.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c3vrtjp2",
   "source_url": "https://omlx.ai/benchmarks/performance/c3vrtjp2",
   "tg_tokens": null,
   "tps": 23.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-c3vrtjp2-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "bf16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: 023jh63j, 3zv2or3o, 87oxf0c6, m6c2o3jn, nx6xy87k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 756.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k 756.4 22.9 24.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c3vrtjp2",
   "source_url": "https://omlx.ai/benchmarks/performance/c3vrtjp2",
   "tg_tokens": null,
   "tps": 22.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-c3vrtjp2-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "bf16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: 023jh63j, 3zv2or3o, 87oxf0c6, m6c2o3jn, nx6xy87k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 733.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 733.4 21.7 25.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c3vrtjp2",
   "source_url": "https://omlx.ai/benchmarks/performance/c3vrtjp2",
   "tg_tokens": null,
   "tps": 21.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-c3vrtjp2-ctx65536",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "bf16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: 023jh63j, 3zv2or3o, 87oxf0c6, m6c2o3jn, nx6xy87k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 669.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 64k 669.6 19.1 27.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c3vrtjp2",
   "source_url": "https://omlx.ai/benchmarks/performance/c3vrtjp2",
   "tg_tokens": null,
   "tps": 19.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-in3mqh7a-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: 759bqh6r, dfmsjfqd, o2kwgd4x, r1slrh3m, tnz70eih, v4rffag0, wl4bmu12; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 828.6,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 828.6 58.3 177.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/in3mqh7a",
   "source_url": "https://omlx.ai/benchmarks/performance/in3mqh7a",
   "tg_tokens": null,
   "tps": 58.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-in3mqh7a-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: 759bqh6r, dfmsjfqd, o2kwgd4x, r1slrh3m, tnz70eih, v4rffag0, wl4bmu12; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2079.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 2,079 68.6 178.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/in3mqh7a",
   "source_url": "https://omlx.ai/benchmarks/performance/in3mqh7a",
   "tg_tokens": null,
   "tps": 68.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-in3mqh7a-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.9 GB at this context; duplicate traces of the same session: 759bqh6r, dfmsjfqd, o2kwgd4x, r1slrh3m, tnz70eih, v4rffag0, wl4bmu12; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2312.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 2,312 68.5 178.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/in3mqh7a",
   "source_url": "https://omlx.ai/benchmarks/performance/in3mqh7a",
   "tg_tokens": null,
   "tps": 68.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-in3mqh7a-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.1 GB at this context; duplicate traces of the same session: 759bqh6r, dfmsjfqd, o2kwgd4x, r1slrh3m, tnz70eih, v4rffag0, wl4bmu12; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2317.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 2,317 140.6 179.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/in3mqh7a",
   "source_url": "https://omlx.ai/benchmarks/performance/in3mqh7a",
   "tg_tokens": null,
   "tps": 140.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-in3mqh7a-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.4 GB at this context; duplicate traces of the same session: 759bqh6r, dfmsjfqd, o2kwgd4x, r1slrh3m, tnz70eih, v4rffag0, wl4bmu12; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2311.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,311 77.1 179.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/in3mqh7a",
   "source_url": "https://omlx.ai/benchmarks/performance/in3mqh7a",
   "tg_tokens": null,
   "tps": 77.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-in3mqh7a-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.1 GB at this context; duplicate traces of the same session: 759bqh6r, dfmsjfqd, o2kwgd4x, r1slrh3m, tnz70eih, v4rffag0, wl4bmu12; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2259.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,259 72.6 180.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/in3mqh7a",
   "source_url": "https://omlx.ai/benchmarks/performance/in3mqh7a",
   "tg_tokens": null,
   "tps": 72.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-in3mqh7a-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.3 GB at this context; duplicate traces of the same session: 759bqh6r, dfmsjfqd, o2kwgd4x, r1slrh3m, tnz70eih, v4rffag0, wl4bmu12; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2234.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,234 140.0 181.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/in3mqh7a",
   "source_url": "https://omlx.ai/benchmarks/performance/in3mqh7a",
   "tg_tokens": null,
   "tps": 140.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-in3mqh7a-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 182.6 GB at this context; duplicate traces of the same session: 759bqh6r, dfmsjfqd, o2kwgd4x, r1slrh3m, tnz70eih, v4rffag0, wl4bmu12; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2166.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 2,166 65.2 182.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/in3mqh7a",
   "source_url": "https://omlx.ai/benchmarks/performance/in3mqh7a",
   "tg_tokens": null,
   "tps": 65.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-v7k66pba-ctx16384",
   "model": "Laguna-S-2.1",
   "notes": "oMLX community benchmark trace; peak memory 49.6 GB at this context; duplicate traces of the same session: 4fzr8yer; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1159.0,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Laguna-S-2.1-MLX-3bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 3bit; 16k 1,159 73.6 49.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v7k66pba",
   "source_url": "https://omlx.ai/benchmarks/performance/v7k66pba",
   "tg_tokens": null,
   "tps": 73.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-v7k66pba-ctx32768",
   "model": "Laguna-S-2.1",
   "notes": "oMLX community benchmark trace; peak memory 50.3 GB at this context; duplicate traces of the same session: 4fzr8yer; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 862.2,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Laguna-S-2.1-MLX-3bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 3bit; 32k 862.2 50.1 50.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v7k66pba",
   "source_url": "https://omlx.ai/benchmarks/performance/v7k66pba",
   "tg_tokens": null,
   "tps": 50.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-x845hskl-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.8 GB; TTFT 1698 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 603.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 603.2 PP tok/s 63.7 TG tok/s 1698 TTFT (ms) 70.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x845hskl",
   "source_url": "https://omlx.ai/benchmarks/performance/x845hskl",
   "tg_tokens": null,
   "tps": 63.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-t02omoik-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: fe8zq7dd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 511.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 511.5 46.0 70.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t02omoik",
   "source_url": "https://omlx.ai/benchmarks/performance/t02omoik",
   "tg_tokens": null,
   "tps": 46.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-t02omoik-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: fe8zq7dd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 645.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 645.0 57.4 71.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t02omoik",
   "source_url": "https://omlx.ai/benchmarks/performance/t02omoik",
   "tg_tokens": null,
   "tps": 57.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-c33wkqcp-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.4 GB at this context; duplicate traces of the same session: ar0t58pb, hpr3x9k0, lg2kpp6u, pbfrvudi, woj04544, zvo9ffa9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2687.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 2,687 142.4 100.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c33wkqcp",
   "source_url": "https://omlx.ai/benchmarks/performance/c33wkqcp",
   "tg_tokens": null,
   "tps": 142.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-c33wkqcp-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.0 GB at this context; duplicate traces of the same session: ar0t58pb, hpr3x9k0, lg2kpp6u, pbfrvudi, woj04544, zvo9ffa9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5091.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 5,091 131.8 101.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c33wkqcp",
   "source_url": "https://omlx.ai/benchmarks/performance/c33wkqcp",
   "tg_tokens": null,
   "tps": 131.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-c33wkqcp-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.1 GB at this context; duplicate traces of the same session: ar0t58pb, hpr3x9k0, lg2kpp6u, pbfrvudi, woj04544, zvo9ffa9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5673.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 5,673 110.2 102.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c33wkqcp",
   "source_url": "https://omlx.ai/benchmarks/performance/c33wkqcp",
   "tg_tokens": null,
   "tps": 110.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-c33wkqcp-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.1 GB at this context; duplicate traces of the same session: ar0t58pb, hpr3x9k0, lg2kpp6u, pbfrvudi, woj04544, zvo9ffa9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5665.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 5,665 122.4 103.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c33wkqcp",
   "source_url": "https://omlx.ai/benchmarks/performance/c33wkqcp",
   "tg_tokens": null,
   "tps": 122.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-c33wkqcp-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.3 GB at this context; duplicate traces of the same session: ar0t58pb, hpr3x9k0, lg2kpp6u, pbfrvudi, woj04544, zvo9ffa9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5542.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 5,542 129.4 104.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c33wkqcp",
   "source_url": "https://omlx.ai/benchmarks/performance/c33wkqcp",
   "tg_tokens": null,
   "tps": 129.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-c33wkqcp-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 108.5 GB at this context; duplicate traces of the same session: ar0t58pb, hpr3x9k0, lg2kpp6u, pbfrvudi, woj04544, zvo9ffa9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5364.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,364 109.3 108.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c33wkqcp",
   "source_url": "https://omlx.ai/benchmarks/performance/c33wkqcp",
   "tg_tokens": null,
   "tps": 109.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-c33wkqcp-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.7 GB at this context; duplicate traces of the same session: ar0t58pb, hpr3x9k0, lg2kpp6u, pbfrvudi, woj04544, zvo9ffa9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5202.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 5,202 92.2 110.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c33wkqcp",
   "source_url": "https://omlx.ai/benchmarks/performance/c33wkqcp",
   "tg_tokens": null,
   "tps": 92.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-duhzcg72-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: 4b44dczb, 4iwfehwy, c417egt9, m6huuml0, mmbr63r9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1876.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,876 93.2 100.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/duhzcg72",
   "source_url": "https://omlx.ai/benchmarks/performance/duhzcg72",
   "tg_tokens": null,
   "tps": 93.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-duhzcg72-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.0 GB at this context; duplicate traces of the same session: 4b44dczb, 4iwfehwy, c417egt9, m6huuml0, mmbr63r9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2304.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 2,304 153.7 101.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/duhzcg72",
   "source_url": "https://omlx.ai/benchmarks/performance/duhzcg72",
   "tg_tokens": null,
   "tps": 153.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-duhzcg72-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.1 GB at this context; duplicate traces of the same session: 4b44dczb, 4iwfehwy, c417egt9, m6huuml0, mmbr63r9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5645.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 5,645 137.5 102.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/duhzcg72",
   "source_url": "https://omlx.ai/benchmarks/performance/duhzcg72",
   "tg_tokens": null,
   "tps": 137.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-duhzcg72-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.5 GB at this context; duplicate traces of the same session: 4b44dczb, 4iwfehwy, c417egt9, m6huuml0, mmbr63r9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5514.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 5,514 63.0 104.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/duhzcg72",
   "source_url": "https://omlx.ai/benchmarks/performance/duhzcg72",
   "tg_tokens": null,
   "tps": 63.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-duhzcg72-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.0 GB at this context; duplicate traces of the same session: 4b44dczb, 4iwfehwy, c417egt9, m6huuml0, mmbr63r9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5378.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,378 89.2 109.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/duhzcg72",
   "source_url": "https://omlx.ai/benchmarks/performance/duhzcg72",
   "tg_tokens": null,
   "tps": 89.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-duhzcg72-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.7 GB at this context; duplicate traces of the same session: 4b44dczb, 4iwfehwy, c417egt9, m6huuml0, mmbr63r9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5216.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 5,216 97.4 110.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/duhzcg72",
   "source_url": "https://omlx.ai/benchmarks/performance/duhzcg72",
   "tg_tokens": null,
   "tps": 97.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-46l5yhsb-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 86.4 GB at this context; duplicate traces of the same session: x5660z12; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 700.3,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 1k 700.3 85.9 86.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/46l5yhsb",
   "source_url": "https://omlx.ai/benchmarks/performance/46l5yhsb",
   "tg_tokens": null,
   "tps": 85.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-46l5yhsb-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 88.3 GB at this context; duplicate traces of the same session: x5660z12; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 746.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 4k 746.2 77.9 88.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/46l5yhsb",
   "source_url": "https://omlx.ai/benchmarks/performance/46l5yhsb",
   "tg_tokens": null,
   "tps": 77.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-o13ewlpg-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 86.4 GB at this context; duplicate traces of the same session: 1qk6ygmk; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 700.5,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 1k 700.5 82.6 86.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o13ewlpg",
   "source_url": "https://omlx.ai/benchmarks/performance/o13ewlpg",
   "tg_tokens": null,
   "tps": 82.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-o13ewlpg-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 88.3 GB at this context; duplicate traces of the same session: 1qk6ygmk; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 744.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 4k 744.4 72.6 88.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o13ewlpg",
   "source_url": "https://omlx.ai/benchmarks/performance/o13ewlpg",
   "tg_tokens": null,
   "tps": 72.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ognx1z8x-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.6 GB at this context; duplicate traces of the same session: 82t5j0ms, 8db0opiu, j37953km, lmga3m54, zdtihvt6; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2461.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 2,461 79.8 142.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ognx1z8x",
   "source_url": "https://omlx.ai/benchmarks/performance/ognx1z8x",
   "tg_tokens": null,
   "tps": 79.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ognx1z8x-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: 82t5j0ms, 8db0opiu, j37953km, lmga3m54, zdtihvt6; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4449.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 4,449 138.2 143.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ognx1z8x",
   "source_url": "https://omlx.ai/benchmarks/performance/ognx1z8x",
   "tg_tokens": null,
   "tps": 138.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ognx1z8x-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 144.5 GB at this context; duplicate traces of the same session: 82t5j0ms, 8db0opiu, j37953km, lmga3m54, zdtihvt6; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4904.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 8k 4,904 152.3 144.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ognx1z8x",
   "source_url": "https://omlx.ai/benchmarks/performance/ognx1z8x",
   "tg_tokens": null,
   "tps": 152.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ognx1z8x-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 146.8 GB at this context; duplicate traces of the same session: 82t5j0ms, 8db0opiu, j37953km, lmga3m54, zdtihvt6; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4832.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 64k 4,832 122.9 146.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ognx1z8x",
   "source_url": "https://omlx.ai/benchmarks/performance/ognx1z8x",
   "tg_tokens": null,
   "tps": 122.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ognx1z8x-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 150.6 GB at this context; duplicate traces of the same session: 82t5j0ms, 8db0opiu, j37953km, lmga3m54, zdtihvt6; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4685.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 128k 4,685 93.7 150.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ognx1z8x",
   "source_url": "https://omlx.ai/benchmarks/performance/ognx1z8x",
   "tg_tokens": null,
   "tps": 93.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ognx1z8x-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 153.3 GB at this context; duplicate traces of the same session: 82t5j0ms, 8db0opiu, j37953km, lmga3m54, zdtihvt6; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4558.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 195k 4,558 92.1 153.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ognx1z8x",
   "source_url": "https://omlx.ai/benchmarks/performance/ognx1z8x",
   "tg_tokens": null,
   "tps": 92.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-08e9zql1-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: 9b0quwpv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1185.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 8bit; 4k 1,185 60.4 26.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/08e9zql1",
   "source_url": "https://omlx.ai/benchmarks/performance/08e9zql1",
   "tg_tokens": null,
   "tps": 60.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-08e9zql1-ctx16384",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: 9b0quwpv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1022.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 8bit; 16k 1,022 51.4 26.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/08e9zql1",
   "source_url": "https://omlx.ai/benchmarks/performance/08e9zql1",
   "tg_tokens": null,
   "tps": 51.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-xm00pe4x-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: bqa4e6kl, cjiyei06, of4d5ucm, rofrbvrx, sh1ihcje, t78xi3f2, un1rotn5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1558.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,558 75.7 177.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xm00pe4x",
   "source_url": "https://omlx.ai/benchmarks/performance/xm00pe4x",
   "tg_tokens": null,
   "tps": 75.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-xm00pe4x-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: bqa4e6kl, cjiyei06, of4d5ucm, rofrbvrx, sh1ihcje, t78xi3f2, un1rotn5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2300.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 2,300 71.3 178.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xm00pe4x",
   "source_url": "https://omlx.ai/benchmarks/performance/xm00pe4x",
   "tg_tokens": null,
   "tps": 71.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-xm00pe4x-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.9 GB at this context; duplicate traces of the same session: bqa4e6kl, cjiyei06, of4d5ucm, rofrbvrx, sh1ihcje, t78xi3f2, un1rotn5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2302.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 2,302 90.1 178.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xm00pe4x",
   "source_url": "https://omlx.ai/benchmarks/performance/xm00pe4x",
   "tg_tokens": null,
   "tps": 90.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-xm00pe4x-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.1 GB at this context; duplicate traces of the same session: bqa4e6kl, cjiyei06, of4d5ucm, rofrbvrx, sh1ihcje, t78xi3f2, un1rotn5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2306.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 2,306 141.1 179.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xm00pe4x",
   "source_url": "https://omlx.ai/benchmarks/performance/xm00pe4x",
   "tg_tokens": null,
   "tps": 141.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-xm00pe4x-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.4 GB at this context; duplicate traces of the same session: bqa4e6kl, cjiyei06, of4d5ucm, rofrbvrx, sh1ihcje, t78xi3f2, un1rotn5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2296.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,296 97.6 179.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xm00pe4x",
   "source_url": "https://omlx.ai/benchmarks/performance/xm00pe4x",
   "tg_tokens": null,
   "tps": 97.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-xm00pe4x-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.1 GB at this context; duplicate traces of the same session: bqa4e6kl, cjiyei06, of4d5ucm, rofrbvrx, sh1ihcje, t78xi3f2, un1rotn5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2261.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,261 72.7 180.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xm00pe4x",
   "source_url": "https://omlx.ai/benchmarks/performance/xm00pe4x",
   "tg_tokens": null,
   "tps": 72.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-xm00pe4x-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.3 GB at this context; duplicate traces of the same session: bqa4e6kl, cjiyei06, of4d5ucm, rofrbvrx, sh1ihcje, t78xi3f2, un1rotn5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2216.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,216 139.0 181.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xm00pe4x",
   "source_url": "https://omlx.ai/benchmarks/performance/xm00pe4x",
   "tg_tokens": null,
   "tps": 139.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-xm00pe4x-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 182.6 GB at this context; duplicate traces of the same session: bqa4e6kl, cjiyei06, of4d5ucm, rofrbvrx, sh1ihcje, t78xi3f2, un1rotn5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2164.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 2,164 62.3 182.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xm00pe4x",
   "source_url": "https://omlx.ai/benchmarks/performance/xm00pe4x",
   "tg_tokens": null,
   "tps": 62.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-1ci0l3xd-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: um94c630, zauv8j0d; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1550.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,550 64.5 177.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ci0l3xd",
   "source_url": "https://omlx.ai/benchmarks/performance/1ci0l3xd",
   "tg_tokens": null,
   "tps": 64.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-1ci0l3xd-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: um94c630, zauv8j0d; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2292.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 2,292 54.5 178.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ci0l3xd",
   "source_url": "https://omlx.ai/benchmarks/performance/1ci0l3xd",
   "tg_tokens": null,
   "tps": 54.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-1ci0l3xd-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.9 GB at this context; duplicate traces of the same session: um94c630, zauv8j0d; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2308.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 2,308 54.8 178.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ci0l3xd",
   "source_url": "https://omlx.ai/benchmarks/performance/1ci0l3xd",
   "tg_tokens": null,
   "tps": 54.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-1ci0l3xd-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.1 GB at this context; duplicate traces of the same session: um94c630, zauv8j0d; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2310.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 2,310 77.3 179.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ci0l3xd",
   "source_url": "https://omlx.ai/benchmarks/performance/1ci0l3xd",
   "tg_tokens": null,
   "tps": 77.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-1ci0l3xd-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.4 GB at this context; duplicate traces of the same session: um94c630, zauv8j0d; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2298.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,298 78.8 179.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ci0l3xd",
   "source_url": "https://omlx.ai/benchmarks/performance/1ci0l3xd",
   "tg_tokens": null,
   "tps": 78.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-1ci0l3xd-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.1 GB at this context; duplicate traces of the same session: um94c630, zauv8j0d; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2263.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,263 63.9 180.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ci0l3xd",
   "source_url": "https://omlx.ai/benchmarks/performance/1ci0l3xd",
   "tg_tokens": null,
   "tps": 63.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-1ci0l3xd-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.3 GB at this context; duplicate traces of the same session: um94c630, zauv8j0d; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2221.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,221 78.8 181.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ci0l3xd",
   "source_url": "https://omlx.ai/benchmarks/performance/1ci0l3xd",
   "tg_tokens": null,
   "tps": 78.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-1ci0l3xd-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 182.6 GB at this context; duplicate traces of the same session: um94c630, zauv8j0d; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2171.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 2,171 54.0 182.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ci0l3xd",
   "source_url": "https://omlx.ai/benchmarks/performance/1ci0l3xd",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-hp30cc4u-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 1v86apps, 31v0cvsn, 9livoh0f, eou2zibj, h2xcw32s, k8u7dhc0, pkqsuo3i; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 558.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 558.8 60.1 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hp30cc4u",
   "source_url": "https://omlx.ai/benchmarks/performance/hp30cc4u",
   "tg_tokens": null,
   "tps": 60.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-hp30cc4u-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 1v86apps, 31v0cvsn, 9livoh0f, eou2zibj, h2xcw32s, k8u7dhc0, pkqsuo3i; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 590.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 590.2 57.5 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hp30cc4u",
   "source_url": "https://omlx.ai/benchmarks/performance/hp30cc4u",
   "tg_tokens": null,
   "tps": 57.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-hp30cc4u-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 1v86apps, 31v0cvsn, 9livoh0f, eou2zibj, h2xcw32s, k8u7dhc0, pkqsuo3i; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 542.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 8k 542.1 54.1 22.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hp30cc4u",
   "source_url": "https://omlx.ai/benchmarks/performance/hp30cc4u",
   "tg_tokens": null,
   "tps": 54.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-hp30cc4u-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: 1v86apps, 31v0cvsn, 9livoh0f, eou2zibj, h2xcw32s, k8u7dhc0, pkqsuo3i; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 461.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k 461.8 48.3 23.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hp30cc4u",
   "source_url": "https://omlx.ai/benchmarks/performance/hp30cc4u",
   "tg_tokens": null,
   "tps": 48.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-hp30cc4u-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 1v86apps, 31v0cvsn, 9livoh0f, eou2zibj, h2xcw32s, k8u7dhc0, pkqsuo3i; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 356.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 32k 356.8 41.1 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hp30cc4u",
   "source_url": "https://omlx.ai/benchmarks/performance/hp30cc4u",
   "tg_tokens": null,
   "tps": 41.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-hp30cc4u-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: 1v86apps, 31v0cvsn, 9livoh0f, eou2zibj, h2xcw32s, k8u7dhc0, pkqsuo3i; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 247.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 64k 247.1 30.9 24.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hp30cc4u",
   "source_url": "https://omlx.ai/benchmarks/performance/hp30cc4u",
   "tg_tokens": null,
   "tps": 30.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-hp30cc4u-ctx131072",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: 1v86apps, 31v0cvsn, 9livoh0f, eou2zibj, h2xcw32s, k8u7dhc0, pkqsuo3i; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 153.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 128k 153.9 19.3 25.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hp30cc4u",
   "source_url": "https://omlx.ai/benchmarks/performance/hp30cc4u",
   "tg_tokens": null,
   "tps": 19.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-hp30cc4u-ctx199680",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: 1v86apps, 31v0cvsn, 9livoh0f, eou2zibj, h2xcw32s, k8u7dhc0, pkqsuo3i; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 110.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 195k 110.3 4.8 26.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hp30cc4u",
   "source_url": "https://omlx.ai/benchmarks/performance/hp30cc4u",
   "tg_tokens": null,
   "tps": 4.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-jfmrikfs-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTPLX Bare Speed build; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: j3ctbq5z, nx3pd9aj, swiwuy5p; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 101.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Bare-Speed on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 101.5 21.0 19.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jfmrikfs",
   "source_url": "https://omlx.ai/benchmarks/performance/jfmrikfs",
   "tg_tokens": null,
   "tps": 21.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-jfmrikfs-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTPLX Bare Speed build; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: j3ctbq5z, nx3pd9aj, swiwuy5p; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 101.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Bare-Speed on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 101.1 27.7 20.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jfmrikfs",
   "source_url": "https://omlx.ai/benchmarks/performance/jfmrikfs",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-jfmrikfs-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTPLX Bare Speed build; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: j3ctbq5z, nx3pd9aj, swiwuy5p; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 96.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Bare-Speed on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 96.6 17.7 22.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jfmrikfs",
   "source_url": "https://omlx.ai/benchmarks/performance/jfmrikfs",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-jfmrikfs-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTPLX Bare Speed build; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: j3ctbq5z, nx3pd9aj, swiwuy5p; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 90.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Bare-Speed on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 90.9 18.6 25.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jfmrikfs",
   "source_url": "https://omlx.ai/benchmarks/performance/jfmrikfs",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-n5w72l7b-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "pi build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: jlig6kn3; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 107.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-pi-4bit-mtp-mlx on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 107.4 19.0 18.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n5w72l7b",
   "source_url": "https://omlx.ai/benchmarks/performance/n5w72l7b",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-n5w72l7b-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "pi build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: jlig6kn3; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-pi-4bit-mtp-mlx on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 102.4 26.2 23.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n5w72l7b",
   "source_url": "https://omlx.ai/benchmarks/performance/n5w72l7b",
   "tg_tokens": null,
   "tps": 26.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-n14jxgch-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.4 GB at this context; duplicate traces of the same session: 9ecattbb, c9x747xv, e5aq09q4, l91m2ncl, nyw9exv9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2519.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 2,519 150.4 100.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n14jxgch",
   "source_url": "https://omlx.ai/benchmarks/performance/n14jxgch",
   "tg_tokens": null,
   "tps": 150.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-n14jxgch-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.1 GB at this context; duplicate traces of the same session: 9ecattbb, c9x747xv, e5aq09q4, l91m2ncl, nyw9exv9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5673.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 5,673 152.7 102.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n14jxgch",
   "source_url": "https://omlx.ai/benchmarks/performance/n14jxgch",
   "tg_tokens": null,
   "tps": 152.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-n14jxgch-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.1 GB at this context; duplicate traces of the same session: 9ecattbb, c9x747xv, e5aq09q4, l91m2ncl, nyw9exv9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5664.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 5,664 131.5 103.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n14jxgch",
   "source_url": "https://omlx.ai/benchmarks/performance/n14jxgch",
   "tg_tokens": null,
   "tps": 131.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-n14jxgch-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.4 GB at this context; duplicate traces of the same session: 9ecattbb, c9x747xv, e5aq09q4, l91m2ncl, nyw9exv9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5537.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 5,537 139.4 104.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n14jxgch",
   "source_url": "https://omlx.ai/benchmarks/performance/n14jxgch",
   "tg_tokens": null,
   "tps": 139.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-n14jxgch-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 108.4 GB at this context; duplicate traces of the same session: 9ecattbb, c9x747xv, e5aq09q4, l91m2ncl, nyw9exv9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5364.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,364 114.1 108.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n14jxgch",
   "source_url": "https://omlx.ai/benchmarks/performance/n14jxgch",
   "tg_tokens": null,
   "tps": 114.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-n14jxgch-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.7 GB at this context; duplicate traces of the same session: 9ecattbb, c9x747xv, e5aq09q4, l91m2ncl, nyw9exv9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5181.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 5,181 101.3 110.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n14jxgch",
   "source_url": "https://omlx.ai/benchmarks/performance/n14jxgch",
   "tg_tokens": null,
   "tps": 101.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-rcasqm1p-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: 4qg8v5uq, btm9na3e, qi7ed3j3, r2l67eug; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2571.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 2,571 158.6 100.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rcasqm1p",
   "source_url": "https://omlx.ai/benchmarks/performance/rcasqm1p",
   "tg_tokens": null,
   "tps": 158.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-rcasqm1p-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.0 GB at this context; duplicate traces of the same session: 4qg8v5uq, btm9na3e, qi7ed3j3, r2l67eug; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5094.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 5,094 100.1 101.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rcasqm1p",
   "source_url": "https://omlx.ai/benchmarks/performance/rcasqm1p",
   "tg_tokens": null,
   "tps": 100.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-rcasqm1p-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.1 GB at this context; duplicate traces of the same session: 4qg8v5uq, btm9na3e, qi7ed3j3, r2l67eug; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5648.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 5,648 141.7 102.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rcasqm1p",
   "source_url": "https://omlx.ai/benchmarks/performance/rcasqm1p",
   "tg_tokens": null,
   "tps": 141.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-rcasqm1p-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.6 GB at this context; duplicate traces of the same session: 4qg8v5uq, btm9na3e, qi7ed3j3, r2l67eug; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5533.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 5,533 136.5 104.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rcasqm1p",
   "source_url": "https://omlx.ai/benchmarks/performance/rcasqm1p",
   "tg_tokens": null,
   "tps": 136.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-rcasqm1p-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.1 GB at this context; duplicate traces of the same session: 4qg8v5uq, btm9na3e, qi7ed3j3, r2l67eug; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5353.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,353 102.5 109.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rcasqm1p",
   "source_url": "https://omlx.ai/benchmarks/performance/rcasqm1p",
   "tg_tokens": null,
   "tps": 102.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-rcasqm1p-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.7 GB at this context; duplicate traces of the same session: 4qg8v5uq, btm9na3e, qi7ed3j3, r2l67eug; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5184.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 5,184 93.7 110.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rcasqm1p",
   "source_url": "https://omlx.ai/benchmarks/performance/rcasqm1p",
   "tg_tokens": null,
   "tps": 93.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-2qve3l6m-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.6 GB at this context; duplicate traces of the same session: 2w1du3ze, a32qeptm, dezd57y2, nl66w079, r9stysoj, sp886wxs; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2426.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 2,426 140.8 142.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2qve3l6m",
   "source_url": "https://omlx.ai/benchmarks/performance/2qve3l6m",
   "tg_tokens": null,
   "tps": 140.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-2qve3l6m-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: 2w1du3ze, a32qeptm, dezd57y2, nl66w079, r9stysoj, sp886wxs; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4474.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 4,474 135.1 143.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2qve3l6m",
   "source_url": "https://omlx.ai/benchmarks/performance/2qve3l6m",
   "tg_tokens": null,
   "tps": 135.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-2qve3l6m-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 144.4 GB at this context; duplicate traces of the same session: 2w1du3ze, a32qeptm, dezd57y2, nl66w079, r9stysoj, sp886wxs; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4916.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 8k 4,916 172.4 144.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2qve3l6m",
   "source_url": "https://omlx.ai/benchmarks/performance/2qve3l6m",
   "tg_tokens": null,
   "tps": 172.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-2qve3l6m-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.1 GB at this context; duplicate traces of the same session: 2w1du3ze, a32qeptm, dezd57y2, nl66w079, r9stysoj, sp886wxs; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4936.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 16k 4,936 130.4 145.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2qve3l6m",
   "source_url": "https://omlx.ai/benchmarks/performance/2qve3l6m",
   "tg_tokens": null,
   "tps": 130.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-2qve3l6m-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 146.8 GB at this context; duplicate traces of the same session: 2w1du3ze, a32qeptm, dezd57y2, nl66w079, r9stysoj, sp886wxs; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4833.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 64k 4,833 129.9 146.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2qve3l6m",
   "source_url": "https://omlx.ai/benchmarks/performance/2qve3l6m",
   "tg_tokens": null,
   "tps": 129.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-2qve3l6m-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 150.6 GB at this context; duplicate traces of the same session: 2w1du3ze, a32qeptm, dezd57y2, nl66w079, r9stysoj, sp886wxs; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4691.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 128k 4,691 106.9 150.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2qve3l6m",
   "source_url": "https://omlx.ai/benchmarks/performance/2qve3l6m",
   "tg_tokens": null,
   "tps": 106.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-2qve3l6m-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 153.3 GB at this context; duplicate traces of the same session: 2w1du3ze, a32qeptm, dezd57y2, nl66w079, r9stysoj, sp886wxs; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4563.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 195k 4,563 89.9 153.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2qve3l6m",
   "source_url": "https://omlx.ai/benchmarks/performance/2qve3l6m",
   "tg_tokens": null,
   "tps": 89.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-co0m8dtp-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.7 GB at this context; duplicate traces of the same session: 1cb8cxvt, 3p1vc6h6; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2283.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 2,283 147.7 142.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/co0m8dtp",
   "source_url": "https://omlx.ai/benchmarks/performance/co0m8dtp",
   "tg_tokens": null,
   "tps": 147.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-co0m8dtp-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: 1cb8cxvt, 3p1vc6h6; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4138.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 4,138 125.4 143.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/co0m8dtp",
   "source_url": "https://omlx.ai/benchmarks/performance/co0m8dtp",
   "tg_tokens": null,
   "tps": 125.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-co0m8dtp-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 144.5 GB at this context; duplicate traces of the same session: 1cb8cxvt, 3p1vc6h6; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4667.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 8k 4,667 141.6 144.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/co0m8dtp",
   "source_url": "https://omlx.ai/benchmarks/performance/co0m8dtp",
   "tg_tokens": null,
   "tps": 141.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-co0m8dtp-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 146.8 GB at this context; duplicate traces of the same session: 1cb8cxvt, 3p1vc6h6; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4583.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 64k 4,583 122.9 146.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/co0m8dtp",
   "source_url": "https://omlx.ai/benchmarks/performance/co0m8dtp",
   "tg_tokens": null,
   "tps": 122.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-co0m8dtp-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 150.5 GB at this context; duplicate traces of the same session: 1cb8cxvt, 3p1vc6h6; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4452.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 128k 4,452 100.5 150.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/co0m8dtp",
   "source_url": "https://omlx.ai/benchmarks/performance/co0m8dtp",
   "tg_tokens": null,
   "tps": 100.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-co0m8dtp-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 153.3 GB at this context; duplicate traces of the same session: 1cb8cxvt, 3p1vc6h6; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4340.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 195k 4,340 83.5 153.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/co0m8dtp",
   "source_url": "https://omlx.ai/benchmarks/performance/co0m8dtp",
   "tg_tokens": null,
   "tps": 83.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-7oppaa3x-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: 01ja8jtf, e7st9ru0, rjs9ovip, wkyn3osg; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1550.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,550 64.5 177.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7oppaa3x",
   "source_url": "https://omlx.ai/benchmarks/performance/7oppaa3x",
   "tg_tokens": null,
   "tps": 64.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-7oppaa3x-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: 01ja8jtf, e7st9ru0, rjs9ovip, wkyn3osg; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2292.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 2,292 54.5 178.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7oppaa3x",
   "source_url": "https://omlx.ai/benchmarks/performance/7oppaa3x",
   "tg_tokens": null,
   "tps": 54.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-7oppaa3x-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.9 GB at this context; duplicate traces of the same session: 01ja8jtf, e7st9ru0, rjs9ovip, wkyn3osg; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2308.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 2,308 54.8 178.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7oppaa3x",
   "source_url": "https://omlx.ai/benchmarks/performance/7oppaa3x",
   "tg_tokens": null,
   "tps": 54.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-7oppaa3x-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.1 GB at this context; duplicate traces of the same session: 01ja8jtf, e7st9ru0, rjs9ovip, wkyn3osg; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2310.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 2,310 77.3 179.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7oppaa3x",
   "source_url": "https://omlx.ai/benchmarks/performance/7oppaa3x",
   "tg_tokens": null,
   "tps": 77.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-7oppaa3x-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.4 GB at this context; duplicate traces of the same session: 01ja8jtf, e7st9ru0, rjs9ovip, wkyn3osg; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2298.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,298 78.8 179.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7oppaa3x",
   "source_url": "https://omlx.ai/benchmarks/performance/7oppaa3x",
   "tg_tokens": null,
   "tps": 78.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-7oppaa3x-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.1 GB at this context; duplicate traces of the same session: 01ja8jtf, e7st9ru0, rjs9ovip, wkyn3osg; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2263.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,263 63.9 180.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7oppaa3x",
   "source_url": "https://omlx.ai/benchmarks/performance/7oppaa3x",
   "tg_tokens": null,
   "tps": 63.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-7oppaa3x-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.3 GB at this context; duplicate traces of the same session: 01ja8jtf, e7st9ru0, rjs9ovip, wkyn3osg; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2221.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,221 78.8 181.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7oppaa3x",
   "source_url": "https://omlx.ai/benchmarks/performance/7oppaa3x",
   "tg_tokens": null,
   "tps": 78.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-7oppaa3x-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 182.6 GB at this context; duplicate traces of the same session: 01ja8jtf, e7st9ru0, rjs9ovip, wkyn3osg; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2171.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 2,171 54.0 182.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7oppaa3x",
   "source_url": "https://omlx.ai/benchmarks/performance/7oppaa3x",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-obbxpfs3-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: 44cnvi5o, 7ju96wrv, frvnsrpf, h2pj175b, js44hy8r, oig731si, tpg8whel; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1551.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,551 75.8 177.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/obbxpfs3",
   "source_url": "https://omlx.ai/benchmarks/performance/obbxpfs3",
   "tg_tokens": null,
   "tps": 75.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-obbxpfs3-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: 44cnvi5o, 7ju96wrv, frvnsrpf, h2pj175b, js44hy8r, oig731si, tpg8whel; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2299.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 2,299 64.1 178.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/obbxpfs3",
   "source_url": "https://omlx.ai/benchmarks/performance/obbxpfs3",
   "tg_tokens": null,
   "tps": 64.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-obbxpfs3-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.9 GB at this context; duplicate traces of the same session: 44cnvi5o, 7ju96wrv, frvnsrpf, h2pj175b, js44hy8r, oig731si, tpg8whel; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2305.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 2,305 90.7 178.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/obbxpfs3",
   "source_url": "https://omlx.ai/benchmarks/performance/obbxpfs3",
   "tg_tokens": null,
   "tps": 90.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-obbxpfs3-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.1 GB at this context; duplicate traces of the same session: 44cnvi5o, 7ju96wrv, frvnsrpf, h2pj175b, js44hy8r, oig731si, tpg8whel; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2306.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 2,306 140.6 179.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/obbxpfs3",
   "source_url": "https://omlx.ai/benchmarks/performance/obbxpfs3",
   "tg_tokens": null,
   "tps": 140.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-obbxpfs3-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.4 GB at this context; duplicate traces of the same session: 44cnvi5o, 7ju96wrv, frvnsrpf, h2pj175b, js44hy8r, oig731si, tpg8whel; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2296.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,296 103.0 179.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/obbxpfs3",
   "source_url": "https://omlx.ai/benchmarks/performance/obbxpfs3",
   "tg_tokens": null,
   "tps": 103.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-obbxpfs3-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.1 GB at this context; duplicate traces of the same session: 44cnvi5o, 7ju96wrv, frvnsrpf, h2pj175b, js44hy8r, oig731si, tpg8whel; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2262.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,262 76.3 180.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/obbxpfs3",
   "source_url": "https://omlx.ai/benchmarks/performance/obbxpfs3",
   "tg_tokens": null,
   "tps": 76.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-obbxpfs3-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.3 GB at this context; duplicate traces of the same session: 44cnvi5o, 7ju96wrv, frvnsrpf, h2pj175b, js44hy8r, oig731si, tpg8whel; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2216.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,216 140.3 181.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/obbxpfs3",
   "source_url": "https://omlx.ai/benchmarks/performance/obbxpfs3",
   "tg_tokens": null,
   "tps": 140.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-obbxpfs3-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 182.6 GB at this context; duplicate traces of the same session: 44cnvi5o, 7ju96wrv, frvnsrpf, h2pj175b, js44hy8r, oig731si, tpg8whel; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2168.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 2,168 67.7 182.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/obbxpfs3",
   "source_url": "https://omlx.ai/benchmarks/performance/obbxpfs3",
   "tg_tokens": null,
   "tps": 67.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-dg6r0ziq-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: 3a44l4a8, bsz9wjac, d5tl4rmy, rkigtbk3, zz5a53ue; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1546.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,546 65.9 177.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dg6r0ziq",
   "source_url": "https://omlx.ai/benchmarks/performance/dg6r0ziq",
   "tg_tokens": null,
   "tps": 65.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-dg6r0ziq-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: 3a44l4a8, bsz9wjac, d5tl4rmy, rkigtbk3, zz5a53ue; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2292.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 2,292 63.3 178.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dg6r0ziq",
   "source_url": "https://omlx.ai/benchmarks/performance/dg6r0ziq",
   "tg_tokens": null,
   "tps": 63.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-dg6r0ziq-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.9 GB at this context; duplicate traces of the same session: 3a44l4a8, bsz9wjac, d5tl4rmy, rkigtbk3, zz5a53ue; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2305.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 2,305 71.5 178.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dg6r0ziq",
   "source_url": "https://omlx.ai/benchmarks/performance/dg6r0ziq",
   "tg_tokens": null,
   "tps": 71.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-dg6r0ziq-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.1 GB at this context; duplicate traces of the same session: 3a44l4a8, bsz9wjac, d5tl4rmy, rkigtbk3, zz5a53ue; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2307.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 2,307 84.5 179.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dg6r0ziq",
   "source_url": "https://omlx.ai/benchmarks/performance/dg6r0ziq",
   "tg_tokens": null,
   "tps": 84.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-dg6r0ziq-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.4 GB at this context; duplicate traces of the same session: 3a44l4a8, bsz9wjac, d5tl4rmy, rkigtbk3, zz5a53ue; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2296.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,296 82.0 179.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dg6r0ziq",
   "source_url": "https://omlx.ai/benchmarks/performance/dg6r0ziq",
   "tg_tokens": null,
   "tps": 82.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-dg6r0ziq-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.1 GB at this context; duplicate traces of the same session: 3a44l4a8, bsz9wjac, d5tl4rmy, rkigtbk3, zz5a53ue; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2263.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,263 64.6 180.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dg6r0ziq",
   "source_url": "https://omlx.ai/benchmarks/performance/dg6r0ziq",
   "tg_tokens": null,
   "tps": 64.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-dg6r0ziq-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.3 GB at this context; duplicate traces of the same session: 3a44l4a8, bsz9wjac, d5tl4rmy, rkigtbk3, zz5a53ue; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2216.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,216 76.4 181.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dg6r0ziq",
   "source_url": "https://omlx.ai/benchmarks/performance/dg6r0ziq",
   "tg_tokens": null,
   "tps": 76.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-dg6r0ziq-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 182.6 GB at this context; duplicate traces of the same session: 3a44l4a8, bsz9wjac, d5tl4rmy, rkigtbk3, zz5a53ue; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2168.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 2,168 54.8 182.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dg6r0ziq",
   "source_url": "https://omlx.ai/benchmarks/performance/dg6r0ziq",
   "tg_tokens": null,
   "tps": 54.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-4noowxja-ctx1024",
   "model": "LFM2.5-8B-A1B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 5.3 GB; TTFT 186 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5520.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-8B-A1B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k tokens 5,520 PP tok/s 142.9 TG tok/s 186 TTFT (ms) 5.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4noowxja",
   "source_url": "https://omlx.ai/benchmarks/performance/4noowxja",
   "tg_tokens": null,
   "tps": 142.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3w5lzn40-ctx1024",
   "model": "NVIDIA-Nemotron-3-Nano-30B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 18.9 GB; TTFT 626 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1635.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3-Nano-30B-A3B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k tokens 1,635 PP tok/s 86.8 TG tok/s 626 TTFT (ms) 18.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3w5lzn40",
   "source_url": "https://omlx.ai/benchmarks/performance/3w5lzn40",
   "tg_tokens": null,
   "tps": 86.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-351b5hqj-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: uk7gx9an; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 798.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-6bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 798.3 58.2 27.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/351b5hqj",
   "source_url": "https://omlx.ai/benchmarks/performance/351b5hqj",
   "tg_tokens": null,
   "tps": 58.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-351b5hqj-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.4 GB at this context; duplicate traces of the same session: uk7gx9an; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1839.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-6bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 16k 1,839 64.6 28.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/351b5hqj",
   "source_url": "https://omlx.ai/benchmarks/performance/351b5hqj",
   "tg_tokens": null,
   "tps": 64.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-xnixw0d0-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 10.7 GB; TTFT 1093 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 936.7,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Ornith-1.5-9B-MXFP8 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 unknown; 1k tokens 936.7 PP tok/s 25.6 TG tok/s 1093 TTFT (ms) 10.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xnixw0d0",
   "source_url": "https://omlx.ai/benchmarks/performance/xnixw0d0",
   "tg_tokens": null,
   "tps": 25.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-4xycxzbc-ctx1024",
   "model": "Qwen3.5-2B",
   "notes": "SpecPrefill build; oMLX community benchmark trace, single run at 1k context; peak memory 2.0 GB; TTFT 273 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3745.0,
   "provenance": "community",
   "quant": "oQ3.5",
   "quote": "Qwen3.5-2B-SpecPrefill-oQ3.5 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 3bit; 1k tokens 3,745 PP tok/s 173.0 TG tok/s 273 TTFT (ms) 2.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4xycxzbc",
   "source_url": "https://omlx.ai/benchmarks/performance/4xycxzbc",
   "tg_tokens": null,
   "tps": 173.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-4fytbcxy-ctx1024",
   "model": "Qwen3.6-35B-A3B-Holo3-Qwopus-AgentWorld",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 29.2 GB; TTFT 547 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1872.0,
   "provenance": "community",
   "quant": "qx64 hi",
   "quote": "Qwen3.6-35B-A3B-Holo3-Qwopus-AgentWorld-qx64-hi-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k tokens 1,872 PP tok/s 68.1 TG tok/s 547 TTFT (ms) 29.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4fytbcxy",
   "source_url": "https://omlx.ai/benchmarks/performance/4fytbcxy",
   "tg_tokens": null,
   "tps": 68.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-nxvx0308-ctx1024",
   "model": "Qwen3.8-27B-Brainwaves-2M",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 20.5 GB; TTFT 2469 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 414.8,
   "provenance": "community",
   "quant": "qx64 hi",
   "quote": "Qwen3.8-27B-Brainwaves-2M-qx64-hi-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k tokens 414.8 PP tok/s 12.7 TG tok/s 2469 TTFT (ms) 20.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nxvx0308",
   "source_url": "https://omlx.ai/benchmarks/performance/nxvx0308",
   "tg_tokens": null,
   "tps": 12.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-9hj5z891-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: hlhznqdx; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 83.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 83.9 15.8 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9hj5z891",
   "source_url": "https://omlx.ai/benchmarks/performance/9hj5z891",
   "tg_tokens": null,
   "tps": 15.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-9hj5z891-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: hlhznqdx; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 78.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 78.7 15.0 20.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9hj5z891",
   "source_url": "https://omlx.ai/benchmarks/performance/9hj5z891",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-3akzh1x0-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2571.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 2,571 158.6 100.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3akzh1x0",
   "source_url": "https://omlx.ai/benchmarks/performance/3akzh1x0",
   "tg_tokens": null,
   "tps": 158.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-3akzh1x0-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.0 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5094.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 5,094 100.1 101.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3akzh1x0",
   "source_url": "https://omlx.ai/benchmarks/performance/3akzh1x0",
   "tg_tokens": null,
   "tps": 100.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-3akzh1x0-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5648.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 5,648 141.7 102.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3akzh1x0",
   "source_url": "https://omlx.ai/benchmarks/performance/3akzh1x0",
   "tg_tokens": null,
   "tps": 141.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-3akzh1x0-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.6 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5533.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 5,533 136.5 104.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3akzh1x0",
   "source_url": "https://omlx.ai/benchmarks/performance/3akzh1x0",
   "tg_tokens": null,
   "tps": 136.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-3akzh1x0-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5353.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,353 102.5 109.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3akzh1x0",
   "source_url": "https://omlx.ai/benchmarks/performance/3akzh1x0",
   "tg_tokens": null,
   "tps": 102.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-3akzh1x0-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5184.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 5,184 93.7 110.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3akzh1x0",
   "source_url": "https://omlx.ai/benchmarks/performance/3akzh1x0",
   "tg_tokens": null,
   "tps": 93.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-l0wacb37-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: 9ul8mrcv, new6rfat, rrysjfxo, tqae02oc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2651.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 2,651 154.6 100.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l0wacb37",
   "source_url": "https://omlx.ai/benchmarks/performance/l0wacb37",
   "tg_tokens": null,
   "tps": 154.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-l0wacb37-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.1 GB at this context; duplicate traces of the same session: 9ul8mrcv, new6rfat, rrysjfxo, tqae02oc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5655.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 5,655 161.6 102.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l0wacb37",
   "source_url": "https://omlx.ai/benchmarks/performance/l0wacb37",
   "tg_tokens": null,
   "tps": 161.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-l0wacb37-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.3 GB at this context; duplicate traces of the same session: 9ul8mrcv, new6rfat, rrysjfxo, tqae02oc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5539.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 5,539 128.6 104.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l0wacb37",
   "source_url": "https://omlx.ai/benchmarks/performance/l0wacb37",
   "tg_tokens": null,
   "tps": 128.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-l0wacb37-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 108.7 GB at this context; duplicate traces of the same session: 9ul8mrcv, new6rfat, rrysjfxo, tqae02oc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5356.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,356 111.9 108.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l0wacb37",
   "source_url": "https://omlx.ai/benchmarks/performance/l0wacb37",
   "tg_tokens": null,
   "tps": 111.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-l0wacb37-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.7 GB at this context; duplicate traces of the same session: 9ul8mrcv, new6rfat, rrysjfxo, tqae02oc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5185.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 5,185 97.0 110.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l0wacb37",
   "source_url": "https://omlx.ai/benchmarks/performance/l0wacb37",
   "tg_tokens": null,
   "tps": 97.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-zebbfaf2-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.7 GB at this context; duplicate traces of the same session: uvnh6j04, vvn1wxyc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2283.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 2,283 147.7 142.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zebbfaf2",
   "source_url": "https://omlx.ai/benchmarks/performance/zebbfaf2",
   "tg_tokens": null,
   "tps": 147.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-zebbfaf2-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: uvnh6j04, vvn1wxyc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4138.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 4,138 125.4 143.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zebbfaf2",
   "source_url": "https://omlx.ai/benchmarks/performance/zebbfaf2",
   "tg_tokens": null,
   "tps": 125.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-zebbfaf2-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 144.5 GB at this context; duplicate traces of the same session: uvnh6j04, vvn1wxyc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4667.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 8k 4,667 141.6 144.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zebbfaf2",
   "source_url": "https://omlx.ai/benchmarks/performance/zebbfaf2",
   "tg_tokens": null,
   "tps": 141.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-zebbfaf2-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 146.8 GB at this context; duplicate traces of the same session: uvnh6j04, vvn1wxyc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4583.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 64k 4,583 122.9 146.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zebbfaf2",
   "source_url": "https://omlx.ai/benchmarks/performance/zebbfaf2",
   "tg_tokens": null,
   "tps": 122.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-zebbfaf2-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 150.5 GB at this context; duplicate traces of the same session: uvnh6j04, vvn1wxyc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4452.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 128k 4,452 100.5 150.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zebbfaf2",
   "source_url": "https://omlx.ai/benchmarks/performance/zebbfaf2",
   "tg_tokens": null,
   "tps": 100.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-zebbfaf2-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 153.3 GB at this context; duplicate traces of the same session: uvnh6j04, vvn1wxyc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4340.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 195k 4,340 83.5 153.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zebbfaf2",
   "source_url": "https://omlx.ai/benchmarks/performance/zebbfaf2",
   "tg_tokens": null,
   "tps": 83.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-5cb5gl0c-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.7 GB at this context; duplicate traces of the same session: 1x6toblz, 7z5znzzh, je2k8n20, l29tqn6n, pge6puiv, setze2up; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2299.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 2,299 142.1 142.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5cb5gl0c",
   "source_url": "https://omlx.ai/benchmarks/performance/5cb5gl0c",
   "tg_tokens": null,
   "tps": 142.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-5cb5gl0c-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: 1x6toblz, 7z5znzzh, je2k8n20, l29tqn6n, pge6puiv, setze2up; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4141.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 4,141 122.4 143.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5cb5gl0c",
   "source_url": "https://omlx.ai/benchmarks/performance/5cb5gl0c",
   "tg_tokens": null,
   "tps": 122.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-5cb5gl0c-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 144.5 GB at this context; duplicate traces of the same session: 1x6toblz, 7z5znzzh, je2k8n20, l29tqn6n, pge6puiv, setze2up; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4673.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 8k 4,673 112.7 144.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5cb5gl0c",
   "source_url": "https://omlx.ai/benchmarks/performance/5cb5gl0c",
   "tg_tokens": null,
   "tps": 112.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-5cb5gl0c-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.1 GB at this context; duplicate traces of the same session: 1x6toblz, 7z5znzzh, je2k8n20, l29tqn6n, pge6puiv, setze2up; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4689.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 16k 4,689 125.7 145.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5cb5gl0c",
   "source_url": "https://omlx.ai/benchmarks/performance/5cb5gl0c",
   "tg_tokens": null,
   "tps": 125.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-5cb5gl0c-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 146.8 GB at this context; duplicate traces of the same session: 1x6toblz, 7z5znzzh, je2k8n20, l29tqn6n, pge6puiv, setze2up; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4585.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 64k 4,585 119.6 146.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5cb5gl0c",
   "source_url": "https://omlx.ai/benchmarks/performance/5cb5gl0c",
   "tg_tokens": null,
   "tps": 119.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-5cb5gl0c-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 150.5 GB at this context; duplicate traces of the same session: 1x6toblz, 7z5znzzh, je2k8n20, l29tqn6n, pge6puiv, setze2up; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4454.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 128k 4,454 100.1 150.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5cb5gl0c",
   "source_url": "https://omlx.ai/benchmarks/performance/5cb5gl0c",
   "tg_tokens": null,
   "tps": 100.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-5cb5gl0c-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 153.3 GB at this context; duplicate traces of the same session: 1x6toblz, 7z5znzzh, je2k8n20, l29tqn6n, pge6puiv, setze2up; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4334.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 195k 4,334 88.6 153.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5cb5gl0c",
   "source_url": "https://omlx.ai/benchmarks/performance/5cb5gl0c",
   "tg_tokens": null,
   "tps": 88.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-wvezotpg-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 11.8 GB; TTFT 469 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2184.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k tokens 2,184 PP tok/s 67.3 TG tok/s 469 TTFT (ms) 11.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wvezotpg",
   "source_url": "https://omlx.ai/benchmarks/performance/wvezotpg",
   "tg_tokens": null,
   "tps": 67.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-i5m5ulog-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: tck8vmu4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1546.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,546 65.9 177.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i5m5ulog",
   "source_url": "https://omlx.ai/benchmarks/performance/i5m5ulog",
   "tg_tokens": null,
   "tps": 65.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-i5m5ulog-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: tck8vmu4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2292.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 2,292 63.3 178.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i5m5ulog",
   "source_url": "https://omlx.ai/benchmarks/performance/i5m5ulog",
   "tg_tokens": null,
   "tps": 63.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-i5m5ulog-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.9 GB at this context; duplicate traces of the same session: tck8vmu4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2305.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 2,305 71.5 178.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i5m5ulog",
   "source_url": "https://omlx.ai/benchmarks/performance/i5m5ulog",
   "tg_tokens": null,
   "tps": 71.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-i5m5ulog-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.1 GB at this context; duplicate traces of the same session: tck8vmu4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2307.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 2,307 84.5 179.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i5m5ulog",
   "source_url": "https://omlx.ai/benchmarks/performance/i5m5ulog",
   "tg_tokens": null,
   "tps": 84.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-i5m5ulog-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.4 GB at this context; duplicate traces of the same session: tck8vmu4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2296.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,296 82.0 179.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i5m5ulog",
   "source_url": "https://omlx.ai/benchmarks/performance/i5m5ulog",
   "tg_tokens": null,
   "tps": 82.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-i5m5ulog-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.1 GB at this context; duplicate traces of the same session: tck8vmu4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2263.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,263 64.6 180.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i5m5ulog",
   "source_url": "https://omlx.ai/benchmarks/performance/i5m5ulog",
   "tg_tokens": null,
   "tps": 64.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-i5m5ulog-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.3 GB at this context; duplicate traces of the same session: tck8vmu4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2216.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,216 76.4 181.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i5m5ulog",
   "source_url": "https://omlx.ai/benchmarks/performance/i5m5ulog",
   "tg_tokens": null,
   "tps": 76.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-i5m5ulog-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 182.6 GB at this context; duplicate traces of the same session: tck8vmu4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2168.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 2,168 54.8 182.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i5m5ulog",
   "source_url": "https://omlx.ai/benchmarks/performance/i5m5ulog",
   "tg_tokens": null,
   "tps": 54.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-h83qewtb-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: 0rw2311j, 38192nb0, 79g11hkx, a3jxqfdy, fszcwru3, vun4ky30, xwmrofb1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1548.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,548 64.3 177.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h83qewtb",
   "source_url": "https://omlx.ai/benchmarks/performance/h83qewtb",
   "tg_tokens": null,
   "tps": 64.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-h83qewtb-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: 0rw2311j, 38192nb0, 79g11hkx, a3jxqfdy, fszcwru3, vun4ky30, xwmrofb1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2288.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 2,288 59.7 178.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h83qewtb",
   "source_url": "https://omlx.ai/benchmarks/performance/h83qewtb",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-h83qewtb-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.9 GB at this context; duplicate traces of the same session: 0rw2311j, 38192nb0, 79g11hkx, a3jxqfdy, fszcwru3, vun4ky30, xwmrofb1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2303.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 2,303 70.6 178.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h83qewtb",
   "source_url": "https://omlx.ai/benchmarks/performance/h83qewtb",
   "tg_tokens": null,
   "tps": 70.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-h83qewtb-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.1 GB at this context; duplicate traces of the same session: 0rw2311j, 38192nb0, 79g11hkx, a3jxqfdy, fszcwru3, vun4ky30, xwmrofb1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2304.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 2,304 85.2 179.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h83qewtb",
   "source_url": "https://omlx.ai/benchmarks/performance/h83qewtb",
   "tg_tokens": null,
   "tps": 85.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-h83qewtb-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.4 GB at this context; duplicate traces of the same session: 0rw2311j, 38192nb0, 79g11hkx, a3jxqfdy, fszcwru3, vun4ky30, xwmrofb1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2298.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,298 84.7 179.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h83qewtb",
   "source_url": "https://omlx.ai/benchmarks/performance/h83qewtb",
   "tg_tokens": null,
   "tps": 84.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-h83qewtb-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.1 GB at this context; duplicate traces of the same session: 0rw2311j, 38192nb0, 79g11hkx, a3jxqfdy, fszcwru3, vun4ky30, xwmrofb1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2261.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,261 71.8 180.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h83qewtb",
   "source_url": "https://omlx.ai/benchmarks/performance/h83qewtb",
   "tg_tokens": null,
   "tps": 71.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-h83qewtb-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.3 GB at this context; duplicate traces of the same session: 0rw2311j, 38192nb0, 79g11hkx, a3jxqfdy, fszcwru3, vun4ky30, xwmrofb1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2215.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,215 78.5 181.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h83qewtb",
   "source_url": "https://omlx.ai/benchmarks/performance/h83qewtb",
   "tg_tokens": null,
   "tps": 78.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-h83qewtb-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 182.6 GB at this context; duplicate traces of the same session: 0rw2311j, 38192nb0, 79g11hkx, a3jxqfdy, fszcwru3, vun4ky30, xwmrofb1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2167.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 2,167 54.8 182.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h83qewtb",
   "source_url": "https://omlx.ai/benchmarks/performance/h83qewtb",
   "tg_tokens": null,
   "tps": 54.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-94vaeaso-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.2 GB at this context; duplicate traces of the same session: fadhjhez, xzd31f7y; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 209.1,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 209.1 32.0 15.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/94vaeaso",
   "source_url": "https://omlx.ai/benchmarks/performance/94vaeaso",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-94vaeaso-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: fadhjhez, xzd31f7y; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 215.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 215.0 30.9 18.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/94vaeaso",
   "source_url": "https://omlx.ai/benchmarks/performance/94vaeaso",
   "tg_tokens": null,
   "tps": 30.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-94vaeaso-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: fadhjhez, xzd31f7y; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 206.9,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 8k 206.9 29.1 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/94vaeaso",
   "source_url": "https://omlx.ai/benchmarks/performance/94vaeaso",
   "tg_tokens": null,
   "tps": 29.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-53scoqgl-ctx1024",
   "model": "Qwen3.8-35B-A3B-Distill-Heretic",
   "notes": "MTP speculative decoding enabled; fp16 build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: z96b7fjw; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 724.4,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-35B-A3B-Distill-Heretic-oQ4-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 724.4 22.2 22.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/53scoqgl",
   "source_url": "https://omlx.ai/benchmarks/performance/53scoqgl",
   "tg_tokens": null,
   "tps": 22.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-53scoqgl-ctx4096",
   "model": "Qwen3.8-35B-A3B-Distill-Heretic",
   "notes": "MTP speculative decoding enabled; fp16 build; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: z96b7fjw; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1051.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-35B-A3B-Distill-Heretic-oQ4-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 1,051 63.1 24.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/53scoqgl",
   "source_url": "https://omlx.ai/benchmarks/performance/53scoqgl",
   "tg_tokens": null,
   "tps": 63.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-m3mrvpx9-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 77.8 GB at this context; duplicate traces of the same session: 2890mcrt, 2lc3xfbd, 63t294h4, 6dfqg0ex, njldjxxh, s4zlalbz, ta8z8xyp; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 287.9,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ4-MTP on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 1k 287.9 32.9 77.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m3mrvpx9",
   "source_url": "https://omlx.ai/benchmarks/performance/m3mrvpx9",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-m3mrvpx9-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 78.5 GB at this context; duplicate traces of the same session: 2890mcrt, 2lc3xfbd, 63t294h4, 6dfqg0ex, njldjxxh, s4zlalbz, ta8z8xyp; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 393.5,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ4-MTP on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 4k 393.5 54.7 78.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m3mrvpx9",
   "source_url": "https://omlx.ai/benchmarks/performance/m3mrvpx9",
   "tg_tokens": null,
   "tps": 54.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-m3mrvpx9-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 78.7 GB at this context; duplicate traces of the same session: 2890mcrt, 2lc3xfbd, 63t294h4, 6dfqg0ex, njldjxxh, s4zlalbz, ta8z8xyp; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 785.5,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ4-MTP on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 8k 785.5 54.3 78.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m3mrvpx9",
   "source_url": "https://omlx.ai/benchmarks/performance/m3mrvpx9",
   "tg_tokens": null,
   "tps": 54.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-m3mrvpx9-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 79.0 GB at this context; duplicate traces of the same session: 2890mcrt, 2lc3xfbd, 63t294h4, 6dfqg0ex, njldjxxh, s4zlalbz, ta8z8xyp; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 787.5,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ4-MTP on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 16k 787.5 41.5 79.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m3mrvpx9",
   "source_url": "https://omlx.ai/benchmarks/performance/m3mrvpx9",
   "tg_tokens": null,
   "tps": 41.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-m3mrvpx9-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 79.6 GB at this context; duplicate traces of the same session: 2890mcrt, 2lc3xfbd, 63t294h4, 6dfqg0ex, njldjxxh, s4zlalbz, ta8z8xyp; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 779.4,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ4-MTP on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 32k 779.4 51.5 79.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m3mrvpx9",
   "source_url": "https://omlx.ai/benchmarks/performance/m3mrvpx9",
   "tg_tokens": null,
   "tps": 51.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-m3mrvpx9-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 81.2 GB at this context; duplicate traces of the same session: 2890mcrt, 2lc3xfbd, 63t294h4, 6dfqg0ex, njldjxxh, s4zlalbz, ta8z8xyp; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 771.2,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ4-MTP on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 64k 771.2 51.9 81.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m3mrvpx9",
   "source_url": "https://omlx.ai/benchmarks/performance/m3mrvpx9",
   "tg_tokens": null,
   "tps": 51.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-m3mrvpx9-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 85.6 GB at this context; duplicate traces of the same session: 2890mcrt, 2lc3xfbd, 63t294h4, 6dfqg0ex, njldjxxh, s4zlalbz, ta8z8xyp; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 755.6,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ4-MTP on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 128k 755.6 46.6 85.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m3mrvpx9",
   "source_url": "https://omlx.ai/benchmarks/performance/m3mrvpx9",
   "tg_tokens": null,
   "tps": 46.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-m3mrvpx9-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 87.3 GB at this context; duplicate traces of the same session: 2890mcrt, 2lc3xfbd, 63t294h4, 6dfqg0ex, njldjxxh, s4zlalbz, ta8z8xyp; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 742.7,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ4-MTP on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 195k 742.7 40.4 87.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m3mrvpx9",
   "source_url": "https://omlx.ai/benchmarks/performance/m3mrvpx9",
   "tg_tokens": null,
   "tps": 40.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-y2khmdly-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.4 GB at this context; duplicate traces of the same session: 0ro1270r, 31xxzq5z, 3v9u9fa4, qb6hmzet, s044i6uk; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2667.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 2,667 114.5 100.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y2khmdly",
   "source_url": "https://omlx.ai/benchmarks/performance/y2khmdly",
   "tg_tokens": null,
   "tps": 114.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-y2khmdly-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.0 GB at this context; duplicate traces of the same session: 0ro1270r, 31xxzq5z, 3v9u9fa4, qb6hmzet, s044i6uk; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5096.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 5,096 136.4 101.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y2khmdly",
   "source_url": "https://omlx.ai/benchmarks/performance/y2khmdly",
   "tg_tokens": null,
   "tps": 136.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-y2khmdly-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.5 GB at this context; duplicate traces of the same session: 0ro1270r, 31xxzq5z, 3v9u9fa4, qb6hmzet, s044i6uk; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5676.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 5,676 150.1 102.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y2khmdly",
   "source_url": "https://omlx.ai/benchmarks/performance/y2khmdly",
   "tg_tokens": null,
   "tps": 150.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-y2khmdly-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.3 GB at this context; duplicate traces of the same session: 0ro1270r, 31xxzq5z, 3v9u9fa4, qb6hmzet, s044i6uk; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5541.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 5,541 134.1 104.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y2khmdly",
   "source_url": "https://omlx.ai/benchmarks/performance/y2khmdly",
   "tg_tokens": null,
   "tps": 134.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-y2khmdly-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.0 GB at this context; duplicate traces of the same session: 0ro1270r, 31xxzq5z, 3v9u9fa4, qb6hmzet, s044i6uk; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5356.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,356 110.2 109.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y2khmdly",
   "source_url": "https://omlx.ai/benchmarks/performance/y2khmdly",
   "tg_tokens": null,
   "tps": 110.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-y2khmdly-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.7 GB at this context; duplicate traces of the same session: 0ro1270r, 31xxzq5z, 3v9u9fa4, qb6hmzet, s044i6uk; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5186.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 5,186 91.0 110.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y2khmdly",
   "source_url": "https://omlx.ai/benchmarks/performance/y2khmdly",
   "tg_tokens": null,
   "tps": 91.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-40tk10ko-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: 1ruuvnds, 6c7r8h2d, hwcesnx9, hwedumql, liu6kclx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2627.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 2,627 159.0 100.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/40tk10ko",
   "source_url": "https://omlx.ai/benchmarks/performance/40tk10ko",
   "tg_tokens": null,
   "tps": 159.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-40tk10ko-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.0 GB at this context; duplicate traces of the same session: 1ruuvnds, 6c7r8h2d, hwcesnx9, hwedumql, liu6kclx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5076.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 5,076 132.1 101.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/40tk10ko",
   "source_url": "https://omlx.ai/benchmarks/performance/40tk10ko",
   "tg_tokens": null,
   "tps": 132.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-40tk10ko-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.1 GB at this context; duplicate traces of the same session: 1ruuvnds, 6c7r8h2d, hwcesnx9, hwedumql, liu6kclx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5664.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 5,664 142.4 102.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/40tk10ko",
   "source_url": "https://omlx.ai/benchmarks/performance/40tk10ko",
   "tg_tokens": null,
   "tps": 142.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-40tk10ko-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.1 GB at this context; duplicate traces of the same session: 1ruuvnds, 6c7r8h2d, hwcesnx9, hwedumql, liu6kclx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5657.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 5,657 95.0 103.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/40tk10ko",
   "source_url": "https://omlx.ai/benchmarks/performance/40tk10ko",
   "tg_tokens": null,
   "tps": 95.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-40tk10ko-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.4 GB at this context; duplicate traces of the same session: 1ruuvnds, 6c7r8h2d, hwcesnx9, hwedumql, liu6kclx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5542.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 5,542 118.4 104.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/40tk10ko",
   "source_url": "https://omlx.ai/benchmarks/performance/40tk10ko",
   "tg_tokens": null,
   "tps": 118.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-40tk10ko-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 108.9 GB at this context; duplicate traces of the same session: 1ruuvnds, 6c7r8h2d, hwcesnx9, hwedumql, liu6kclx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5364.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,364 108.6 108.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/40tk10ko",
   "source_url": "https://omlx.ai/benchmarks/performance/40tk10ko",
   "tg_tokens": null,
   "tps": 108.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-40tk10ko-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.7 GB at this context; duplicate traces of the same session: 1ruuvnds, 6c7r8h2d, hwcesnx9, hwedumql, liu6kclx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5185.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 5,185 101.0 110.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/40tk10ko",
   "source_url": "https://omlx.ai/benchmarks/performance/40tk10ko",
   "tg_tokens": null,
   "tps": 101.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-86czwhe7-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.3 GB at this context; duplicate traces of the same session: budb4rbf, gqab7o0h, ik9n4oh4, klc43aw5, w1bnuj3w; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2249.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 2,249 133.8 142.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/86czwhe7",
   "source_url": "https://omlx.ai/benchmarks/performance/86czwhe7",
   "tg_tokens": null,
   "tps": 133.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-86czwhe7-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: budb4rbf, gqab7o0h, ik9n4oh4, klc43aw5, w1bnuj3w; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4136.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 4,136 120.4 143.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/86czwhe7",
   "source_url": "https://omlx.ai/benchmarks/performance/86czwhe7",
   "tg_tokens": null,
   "tps": 120.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-86czwhe7-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 144.5 GB at this context; duplicate traces of the same session: budb4rbf, gqab7o0h, ik9n4oh4, klc43aw5, w1bnuj3w; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4670.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 8k 4,670 122.2 144.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/86czwhe7",
   "source_url": "https://omlx.ai/benchmarks/performance/86czwhe7",
   "tg_tokens": null,
   "tps": 122.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-86czwhe7-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 146.8 GB at this context; duplicate traces of the same session: budb4rbf, gqab7o0h, ik9n4oh4, klc43aw5, w1bnuj3w; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4580.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 64k 4,580 123.5 146.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/86czwhe7",
   "source_url": "https://omlx.ai/benchmarks/performance/86czwhe7",
   "tg_tokens": null,
   "tps": 123.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-86czwhe7-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 150.6 GB at this context; duplicate traces of the same session: budb4rbf, gqab7o0h, ik9n4oh4, klc43aw5, w1bnuj3w; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4453.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 128k 4,453 101.4 150.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/86czwhe7",
   "source_url": "https://omlx.ai/benchmarks/performance/86czwhe7",
   "tg_tokens": null,
   "tps": 101.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-86czwhe7-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 153.3 GB at this context; duplicate traces of the same session: budb4rbf, gqab7o0h, ik9n4oh4, klc43aw5, w1bnuj3w; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4336.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 195k 4,336 91.0 153.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/86czwhe7",
   "source_url": "https://omlx.ai/benchmarks/performance/86czwhe7",
   "tg_tokens": null,
   "tps": 91.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-4nkqxk61-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.6 GB at this context; duplicate traces of the same session: 305vqx8q, gjuy7kcm; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2472.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 2,472 135.2 142.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4nkqxk61",
   "source_url": "https://omlx.ai/benchmarks/performance/4nkqxk61",
   "tg_tokens": null,
   "tps": 135.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-4nkqxk61-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: 305vqx8q, gjuy7kcm; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4468.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 4,468 138.1 143.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4nkqxk61",
   "source_url": "https://omlx.ai/benchmarks/performance/4nkqxk61",
   "tg_tokens": null,
   "tps": 138.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-4nkqxk61-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 144.5 GB at this context; duplicate traces of the same session: 305vqx8q, gjuy7kcm; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4910.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 8k 4,910 146.5 144.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4nkqxk61",
   "source_url": "https://omlx.ai/benchmarks/performance/4nkqxk61",
   "tg_tokens": null,
   "tps": 146.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-4nkqxk61-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.7 GB at this context; duplicate traces of the same session: 305vqx8q, gjuy7kcm; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4916.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 32k 4,916 111.7 145.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4nkqxk61",
   "source_url": "https://omlx.ai/benchmarks/performance/4nkqxk61",
   "tg_tokens": null,
   "tps": 111.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-4nkqxk61-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 146.8 GB at this context; duplicate traces of the same session: 305vqx8q, gjuy7kcm; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4832.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 64k 4,832 133.3 146.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4nkqxk61",
   "source_url": "https://omlx.ai/benchmarks/performance/4nkqxk61",
   "tg_tokens": null,
   "tps": 133.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-4nkqxk61-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 150.6 GB at this context; duplicate traces of the same session: 305vqx8q, gjuy7kcm; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4690.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 128k 4,690 100.5 150.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4nkqxk61",
   "source_url": "https://omlx.ai/benchmarks/performance/4nkqxk61",
   "tg_tokens": null,
   "tps": 100.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-4nkqxk61-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 153.3 GB at this context; duplicate traces of the same session: 305vqx8q, gjuy7kcm; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4560.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 195k 4,560 90.5 153.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4nkqxk61",
   "source_url": "https://omlx.ai/benchmarks/performance/4nkqxk61",
   "tg_tokens": null,
   "tps": 90.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-wjj1odgo-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 25.1 GB; TTFT 73009 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 112.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 8k tokens 112.2 PP tok/s 24.2 TG tok/s 73009 TTFT (ms) 25.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wjj1odgo",
   "source_url": "https://omlx.ai/benchmarks/performance/wjj1odgo",
   "tg_tokens": null,
   "tps": 24.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-v6pa47pt-ctx8192",
   "model": "ThinkingCap-Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 32.3 GB; TTFT 120967 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 67.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "ThinkingCap-Qwen3.8-27B-oQ6e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 6bit; 8k tokens 67.7 PP tok/s 10.8 TG tok/s 120967 TTFT (ms) 32.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v6pa47pt",
   "source_url": "https://omlx.ai/benchmarks/performance/v6pa47pt",
   "tg_tokens": null,
   "tps": 10.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-cb7x0bkg-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: 0175m1s8, 06g0zto5, dilwydze, fmykboaq, iuwlxla7, ldt87sr4, zi5ov43x; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1561.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,561 76.9 177.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cb7x0bkg",
   "source_url": "https://omlx.ai/benchmarks/performance/cb7x0bkg",
   "tg_tokens": null,
   "tps": 76.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-cb7x0bkg-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: 0175m1s8, 06g0zto5, dilwydze, fmykboaq, iuwlxla7, ldt87sr4, zi5ov43x; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2309.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 2,309 71.8 178.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cb7x0bkg",
   "source_url": "https://omlx.ai/benchmarks/performance/cb7x0bkg",
   "tg_tokens": null,
   "tps": 71.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-cb7x0bkg-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.9 GB at this context; duplicate traces of the same session: 0175m1s8, 06g0zto5, dilwydze, fmykboaq, iuwlxla7, ldt87sr4, zi5ov43x; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2306.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 2,306 88.0 178.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cb7x0bkg",
   "source_url": "https://omlx.ai/benchmarks/performance/cb7x0bkg",
   "tg_tokens": null,
   "tps": 88.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-cb7x0bkg-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.1 GB at this context; duplicate traces of the same session: 0175m1s8, 06g0zto5, dilwydze, fmykboaq, iuwlxla7, ldt87sr4, zi5ov43x; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2305.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 2,305 141.8 179.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cb7x0bkg",
   "source_url": "https://omlx.ai/benchmarks/performance/cb7x0bkg",
   "tg_tokens": null,
   "tps": 141.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-cb7x0bkg-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.4 GB at this context; duplicate traces of the same session: 0175m1s8, 06g0zto5, dilwydze, fmykboaq, iuwlxla7, ldt87sr4, zi5ov43x; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2298.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,298 98.5 179.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cb7x0bkg",
   "source_url": "https://omlx.ai/benchmarks/performance/cb7x0bkg",
   "tg_tokens": null,
   "tps": 98.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-cb7x0bkg-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.1 GB at this context; duplicate traces of the same session: 0175m1s8, 06g0zto5, dilwydze, fmykboaq, iuwlxla7, ldt87sr4, zi5ov43x; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2260.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,260 81.2 180.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cb7x0bkg",
   "source_url": "https://omlx.ai/benchmarks/performance/cb7x0bkg",
   "tg_tokens": null,
   "tps": 81.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-cb7x0bkg-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.3 GB at this context; duplicate traces of the same session: 0175m1s8, 06g0zto5, dilwydze, fmykboaq, iuwlxla7, ldt87sr4, zi5ov43x; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2217.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,217 140.1 181.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cb7x0bkg",
   "source_url": "https://omlx.ai/benchmarks/performance/cb7x0bkg",
   "tg_tokens": null,
   "tps": 140.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-cb7x0bkg-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 182.6 GB at this context; duplicate traces of the same session: 0175m1s8, 06g0zto5, dilwydze, fmykboaq, iuwlxla7, ldt87sr4, zi5ov43x; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2164.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 2,164 69.5 182.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cb7x0bkg",
   "source_url": "https://omlx.ai/benchmarks/performance/cb7x0bkg",
   "tg_tokens": null,
   "tps": 69.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-yndjdvoy-ctx4096",
   "model": "Nail-Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 22.6 GB; TTFT 5266 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 777.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k tokens 777.9 PP tok/s 57.7 TG tok/s 5266 TTFT (ms) 22.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yndjdvoy",
   "source_url": "https://omlx.ai/benchmarks/performance/yndjdvoy",
   "tg_tokens": null,
   "tps": 57.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fvkyvrrc-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: x9f9en99, yxbj5y43; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5297.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 5,297 91.5 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fvkyvrrc",
   "source_url": "https://omlx.ai/benchmarks/performance/fvkyvrrc",
   "tg_tokens": null,
   "tps": 91.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fvkyvrrc-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: x9f9en99, yxbj5y43; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4405.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 4,405 174.3 22.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fvkyvrrc",
   "source_url": "https://omlx.ai/benchmarks/performance/fvkyvrrc",
   "tg_tokens": null,
   "tps": 174.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fvkyvrrc-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: x9f9en99, yxbj5y43; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3419.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 3,419 131.0 24.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fvkyvrrc",
   "source_url": "https://omlx.ai/benchmarks/performance/fvkyvrrc",
   "tg_tokens": null,
   "tps": 131.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-prtjj8kz-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 78.6 GB; TTFT 133809 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 979.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 128k tokens 979.5 PP tok/s 48.9 TG tok/s 133809 TTFT (ms) 78.6 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/prtjj8kz",
   "source_url": "https://omlx.ai/benchmarks/performance/prtjj8kz",
   "tg_tokens": null,
   "tps": 48.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-9viuyixj-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2627.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 2,627 159.0 100.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9viuyixj",
   "source_url": "https://omlx.ai/benchmarks/performance/9viuyixj",
   "tg_tokens": null,
   "tps": 159.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-9viuyixj-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.0 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5076.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 5,076 132.1 101.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9viuyixj",
   "source_url": "https://omlx.ai/benchmarks/performance/9viuyixj",
   "tg_tokens": null,
   "tps": 132.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-9viuyixj-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5664.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 5,664 142.4 102.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9viuyixj",
   "source_url": "https://omlx.ai/benchmarks/performance/9viuyixj",
   "tg_tokens": null,
   "tps": 142.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-9viuyixj-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5657.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 5,657 95.0 103.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9viuyixj",
   "source_url": "https://omlx.ai/benchmarks/performance/9viuyixj",
   "tg_tokens": null,
   "tps": 95.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-9viuyixj-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.4 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5542.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 5,542 118.4 104.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9viuyixj",
   "source_url": "https://omlx.ai/benchmarks/performance/9viuyixj",
   "tg_tokens": null,
   "tps": 118.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-9viuyixj-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 108.9 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5364.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,364 108.6 108.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9viuyixj",
   "source_url": "https://omlx.ai/benchmarks/performance/9viuyixj",
   "tg_tokens": null,
   "tps": 108.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-9viuyixj-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5185.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 5,185 101.0 110.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9viuyixj",
   "source_url": "https://omlx.ai/benchmarks/performance/9viuyixj",
   "tg_tokens": null,
   "tps": 101.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-1vdivt5d-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.6 GB at this context; duplicate traces of the same session: 8hnx7l7f, axypx6a4, innbk42l; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2472.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 2,472 135.2 142.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1vdivt5d",
   "source_url": "https://omlx.ai/benchmarks/performance/1vdivt5d",
   "tg_tokens": null,
   "tps": 135.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-1vdivt5d-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: 8hnx7l7f, axypx6a4, innbk42l; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4468.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 4,468 138.1 143.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1vdivt5d",
   "source_url": "https://omlx.ai/benchmarks/performance/1vdivt5d",
   "tg_tokens": null,
   "tps": 138.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-1vdivt5d-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 144.5 GB at this context; duplicate traces of the same session: 8hnx7l7f, axypx6a4, innbk42l; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4910.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 8k 4,910 146.5 144.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1vdivt5d",
   "source_url": "https://omlx.ai/benchmarks/performance/1vdivt5d",
   "tg_tokens": null,
   "tps": 146.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-1vdivt5d-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.7 GB at this context; duplicate traces of the same session: 8hnx7l7f, axypx6a4, innbk42l; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4916.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 32k 4,916 111.7 145.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1vdivt5d",
   "source_url": "https://omlx.ai/benchmarks/performance/1vdivt5d",
   "tg_tokens": null,
   "tps": 111.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-1vdivt5d-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 146.8 GB at this context; duplicate traces of the same session: 8hnx7l7f, axypx6a4, innbk42l; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4832.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 64k 4,832 133.3 146.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1vdivt5d",
   "source_url": "https://omlx.ai/benchmarks/performance/1vdivt5d",
   "tg_tokens": null,
   "tps": 133.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-1vdivt5d-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 150.6 GB at this context; duplicate traces of the same session: 8hnx7l7f, axypx6a4, innbk42l; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4690.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 128k 4,690 100.5 150.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1vdivt5d",
   "source_url": "https://omlx.ai/benchmarks/performance/1vdivt5d",
   "tg_tokens": null,
   "tps": 100.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-1vdivt5d-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 153.3 GB at this context; duplicate traces of the same session: 8hnx7l7f, axypx6a4, innbk42l; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4560.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 195k 4,560 90.5 153.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1vdivt5d",
   "source_url": "https://omlx.ai/benchmarks/performance/1vdivt5d",
   "tg_tokens": null,
   "tps": 90.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-wypft2wn-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.6 GB at this context; duplicate traces of the same session: 7scvo3do, 8hrgyp20, 9uo000hm, iku4z50n, jh2343ri, xk3lbt5s; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2458.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 2,458 132.5 142.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wypft2wn",
   "source_url": "https://omlx.ai/benchmarks/performance/wypft2wn",
   "tg_tokens": null,
   "tps": 132.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-wypft2wn-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: 7scvo3do, 8hrgyp20, 9uo000hm, iku4z50n, jh2343ri, xk3lbt5s; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4480.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 4,480 121.4 143.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wypft2wn",
   "source_url": "https://omlx.ai/benchmarks/performance/wypft2wn",
   "tg_tokens": null,
   "tps": 121.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-wypft2wn-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 144.5 GB at this context; duplicate traces of the same session: 7scvo3do, 8hrgyp20, 9uo000hm, iku4z50n, jh2343ri, xk3lbt5s; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4923.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 8k 4,923 146.1 144.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wypft2wn",
   "source_url": "https://omlx.ai/benchmarks/performance/wypft2wn",
   "tg_tokens": null,
   "tps": 146.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-wypft2wn-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.1 GB at this context; duplicate traces of the same session: 7scvo3do, 8hrgyp20, 9uo000hm, iku4z50n, jh2343ri, xk3lbt5s; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4939.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 16k 4,939 116.8 145.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wypft2wn",
   "source_url": "https://omlx.ai/benchmarks/performance/wypft2wn",
   "tg_tokens": null,
   "tps": 116.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-wypft2wn-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 146.8 GB at this context; duplicate traces of the same session: 7scvo3do, 8hrgyp20, 9uo000hm, iku4z50n, jh2343ri, xk3lbt5s; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4836.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 64k 4,836 128.7 146.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wypft2wn",
   "source_url": "https://omlx.ai/benchmarks/performance/wypft2wn",
   "tg_tokens": null,
   "tps": 128.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-wypft2wn-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 150.6 GB at this context; duplicate traces of the same session: 7scvo3do, 8hrgyp20, 9uo000hm, iku4z50n, jh2343ri, xk3lbt5s; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4689.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 128k 4,689 99.6 150.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wypft2wn",
   "source_url": "https://omlx.ai/benchmarks/performance/wypft2wn",
   "tg_tokens": null,
   "tps": 99.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-wypft2wn-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 153.3 GB at this context; duplicate traces of the same session: 7scvo3do, 8hrgyp20, 9uo000hm, iku4z50n, jh2343ri, xk3lbt5s; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4562.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 195k 4,562 88.5 153.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wypft2wn",
   "source_url": "https://omlx.ai/benchmarks/performance/wypft2wn",
   "tg_tokens": null,
   "tps": 88.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-zn4ig2u5-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.6 GB at this context; duplicate traces of the same session: fpranvqa, xshfpctd; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 320.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 320.4 25.8 173.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zn4ig2u5",
   "source_url": "https://omlx.ai/benchmarks/performance/zn4ig2u5",
   "tg_tokens": null,
   "tps": 25.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-zn4ig2u5-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.7 GB at this context; duplicate traces of the same session: fpranvqa, xshfpctd; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 506.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 506.7 23.9 174.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zn4ig2u5",
   "source_url": "https://omlx.ai/benchmarks/performance/zn4ig2u5",
   "tg_tokens": null,
   "tps": 23.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-zn4ig2u5-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.0 GB at this context; duplicate traces of the same session: fpranvqa, xshfpctd; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 427.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 427.7 23.0 175.0 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zn4ig2u5",
   "source_url": "https://omlx.ai/benchmarks/performance/zn4ig2u5",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8ke6ocfm-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: mg0bmn5c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 887.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 887.6 30.6 16.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8ke6ocfm",
   "source_url": "https://omlx.ai/benchmarks/performance/8ke6ocfm",
   "tg_tokens": null,
   "tps": 30.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8ke6ocfm-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: mg0bmn5c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1006.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,006 29.9 17.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8ke6ocfm",
   "source_url": "https://omlx.ai/benchmarks/performance/8ke6ocfm",
   "tg_tokens": null,
   "tps": 29.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-e9y0isfg-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 19.9 GB; TTFT 89094 ms; duplicate traces of the same session: none; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 46.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 4bit; 4k tokens 46.0 PP tok/s 9.1 TG tok/s 89094 TTFT (ms) 19.9 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e9y0isfg",
   "source_url": "https://omlx.ai/benchmarks/performance/e9y0isfg",
   "tg_tokens": null,
   "tps": 9.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1ic943bv-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.5 GB at this context; duplicate traces of the same session: j2y5oj4z; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 449.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 449.0 33.3 17.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ic943bv",
   "source_url": "https://omlx.ai/benchmarks/performance/1ic943bv",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1ic943bv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: j2y5oj4z; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 482.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 482.3 28.5 18.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ic943bv",
   "source_url": "https://omlx.ai/benchmarks/performance/1ic943bv",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1dny1ijm-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 69.6 GB at this context; duplicate traces of the same session: vglcmcbj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1687.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,687 24.6 69.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1dny1ijm",
   "source_url": "https://omlx.ai/benchmarks/performance/1dny1ijm",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1dny1ijm-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 69.8 GB at this context; duplicate traces of the same session: vglcmcbj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1941.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,941 23.7 69.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1dny1ijm",
   "source_url": "https://omlx.ai/benchmarks/performance/1dny1ijm",
   "tg_tokens": null,
   "tps": 23.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-amldcewf-ctx1024",
   "model": "Gemma 4 12B",
   "notes": "oMLX community benchmark trace; peak memory 12.8 GB at this context; duplicate traces of the same session: p2860ote; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1181.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-12B-it-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 1,181 38.6 12.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/amldcewf",
   "source_url": "https://omlx.ai/benchmarks/performance/amldcewf",
   "tg_tokens": null,
   "tps": 38.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-amldcewf-ctx4096",
   "model": "Gemma 4 12B",
   "notes": "oMLX community benchmark trace; peak memory 13.2 GB at this context; duplicate traces of the same session: p2860ote; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1589.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-12B-it-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 1,589 38.5 13.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/amldcewf",
   "source_url": "https://omlx.ai/benchmarks/performance/amldcewf",
   "tg_tokens": null,
   "tps": 38.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-k7mklqvh-ctx1024",
   "model": "gemma-4-26B-A4B-it-heretic",
   "notes": "oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: l3rdgwvz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2041.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-heretic-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 2,041 76.8 14.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k7mklqvh",
   "source_url": "https://omlx.ai/benchmarks/performance/k7mklqvh",
   "tg_tokens": null,
   "tps": 76.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-k7mklqvh-ctx4096",
   "model": "gemma-4-26B-A4B-it-heretic",
   "notes": "oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: l3rdgwvz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2350.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-heretic-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 2,350 71.6 14.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k7mklqvh",
   "source_url": "https://omlx.ai/benchmarks/performance/k7mklqvh",
   "tg_tokens": null,
   "tps": 71.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-3arm2mym-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: jcwohvh4; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2107.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 2,107 77.0 14.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3arm2mym",
   "source_url": "https://omlx.ai/benchmarks/performance/3arm2mym",
   "tg_tokens": null,
   "tps": 77.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-09",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-3arm2mym-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: jcwohvh4; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2325.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 2,325 71.9 14.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3arm2mym",
   "source_url": "https://omlx.ai/benchmarks/performance/3arm2mym",
   "tg_tokens": null,
   "tps": 71.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-0goneubc-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: 8qa1cbh7; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 429.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 429.2 38.7 3.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0goneubc",
   "source_url": "https://omlx.ai/benchmarks/performance/0goneubc",
   "tg_tokens": null,
   "tps": 38.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-0goneubc-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: 8qa1cbh7; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 423.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 423.8 37.2 3.9 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0goneubc",
   "source_url": "https://omlx.ai/benchmarks/performance/0goneubc",
   "tg_tokens": null,
   "tps": 37.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-b2udqxvv-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 42.4 GB; TTFT 60290 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2174.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k tokens 2,174 PP tok/s 62.9 TG tok/s 60290 TTFT (ms) 42.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b2udqxvv",
   "source_url": "https://omlx.ai/benchmarks/performance/b2udqxvv",
   "tg_tokens": null,
   "tps": 62.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-cjftthlo-ctx131072",
   "model": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 26.3 GB; TTFT 60136 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2180.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k tokens 2,180 PP tok/s 73.2 TG tok/s 60136 TTFT (ms) 26.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cjftthlo",
   "source_url": "https://omlx.ai/benchmarks/performance/cjftthlo",
   "tg_tokens": null,
   "tps": 73.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-xx9wmos0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: p1gul8mb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 453.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 453.8 28.5 18.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xx9wmos0",
   "source_url": "https://omlx.ai/benchmarks/performance/xx9wmos0",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-xx9wmos0-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: p1gul8mb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 423.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 423.8 30.1 18.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xx9wmos0",
   "source_url": "https://omlx.ai/benchmarks/performance/xx9wmos0",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tu0l0ag8-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 15w0n9hk, hdliq8br, kd3h104w, ngfjdbcf, vfbdju10, ywy4ce3n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 891.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 891.7 61.8 17.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tu0l0ag8",
   "source_url": "https://omlx.ai/benchmarks/performance/tu0l0ag8",
   "tg_tokens": null,
   "tps": 61.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tu0l0ag8-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 15w0n9hk, hdliq8br, kd3h104w, ngfjdbcf, vfbdju10, ywy4ce3n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1308.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,308 61.0 19.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tu0l0ag8",
   "source_url": "https://omlx.ai/benchmarks/performance/tu0l0ag8",
   "tg_tokens": null,
   "tps": 61.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-tu0l0ag8-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 15w0n9hk, hdliq8br, kd3h104w, ngfjdbcf, vfbdju10, ywy4ce3n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1286.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,286 86.3 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tu0l0ag8",
   "source_url": "https://omlx.ai/benchmarks/performance/tu0l0ag8",
   "tg_tokens": null,
   "tps": 86.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tu0l0ag8-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 15w0n9hk, hdliq8br, kd3h104w, ngfjdbcf, vfbdju10, ywy4ce3n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1198.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,198 67.3 19.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tu0l0ag8",
   "source_url": "https://omlx.ai/benchmarks/performance/tu0l0ag8",
   "tg_tokens": null,
   "tps": 67.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tu0l0ag8-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 15w0n9hk, hdliq8br, kd3h104w, ngfjdbcf, vfbdju10, ywy4ce3n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1020.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,020 123.3 21.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tu0l0ag8",
   "source_url": "https://omlx.ai/benchmarks/performance/tu0l0ag8",
   "tg_tokens": null,
   "tps": 123.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tu0l0ag8-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 15w0n9hk, hdliq8br, kd3h104w, ngfjdbcf, vfbdju10, ywy4ce3n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 792.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 792.9 36.3 23.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tu0l0ag8",
   "source_url": "https://omlx.ai/benchmarks/performance/tu0l0ag8",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tu0l0ag8-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: 15w0n9hk, hdliq8br, kd3h104w, ngfjdbcf, vfbdju10, ywy4ce3n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 544.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 544.3 41.8 28.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tu0l0ag8",
   "source_url": "https://omlx.ai/benchmarks/performance/tu0l0ag8",
   "tg_tokens": null,
   "tps": 41.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-o5qrpvjs-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: 3ky65lg4, 427l8vao, gjd6f2ay, lgdi1iln, oq8yy4mk, uxh2m2wt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 797.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 797.1 50.8 33.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o5qrpvjs",
   "source_url": "https://omlx.ai/benchmarks/performance/o5qrpvjs",
   "tg_tokens": null,
   "tps": 50.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-o5qrpvjs-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: 3ky65lg4, 427l8vao, gjd6f2ay, lgdi1iln, oq8yy4mk, uxh2m2wt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 890.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 890.3 49.0 34.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o5qrpvjs",
   "source_url": "https://omlx.ai/benchmarks/performance/o5qrpvjs",
   "tg_tokens": null,
   "tps": 49.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-o5qrpvjs-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: 3ky65lg4, 427l8vao, gjd6f2ay, lgdi1iln, oq8yy4mk, uxh2m2wt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 826.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 826.9 45.6 35.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o5qrpvjs",
   "source_url": "https://omlx.ai/benchmarks/performance/o5qrpvjs",
   "tg_tokens": null,
   "tps": 45.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-o5qrpvjs-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: 3ky65lg4, 427l8vao, gjd6f2ay, lgdi1iln, oq8yy4mk, uxh2m2wt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 739.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 739.2 59.1 35.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o5qrpvjs",
   "source_url": "https://omlx.ai/benchmarks/performance/o5qrpvjs",
   "tg_tokens": null,
   "tps": 59.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-o5qrpvjs-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: 3ky65lg4, 427l8vao, gjd6f2ay, lgdi1iln, oq8yy4mk, uxh2m2wt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 691.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 691.0 73.4 36.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o5qrpvjs",
   "source_url": "https://omlx.ai/benchmarks/performance/o5qrpvjs",
   "tg_tokens": null,
   "tps": 73.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-o5qrpvjs-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: 3ky65lg4, 427l8vao, gjd6f2ay, lgdi1iln, oq8yy4mk, uxh2m2wt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 578.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 578.4 36.2 39.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o5qrpvjs",
   "source_url": "https://omlx.ai/benchmarks/performance/o5qrpvjs",
   "tg_tokens": null,
   "tps": 36.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-o5qrpvjs-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 44.0 GB at this context; duplicate traces of the same session: 3ky65lg4, 427l8vao, gjd6f2ay, lgdi1iln, oq8yy4mk, uxh2m2wt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 434.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 434.9 33.6 44.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o5qrpvjs",
   "source_url": "https://omlx.ai/benchmarks/performance/o5qrpvjs",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-q3ij6q2d-ctx131072",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 130.0 GB at this context; duplicate traces of the same session: h1sfso15; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 969.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 128k 969.4 34.7 130.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q3ij6q2d",
   "source_url": "https://omlx.ai/benchmarks/performance/q3ij6q2d",
   "tg_tokens": null,
   "tps": 34.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-q3ij6q2d-ctx199680",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 131.3 GB at this context; duplicate traces of the same session: h1sfso15; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 944.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 195k 944.9 47.9 131.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q3ij6q2d",
   "source_url": "https://omlx.ai/benchmarks/performance/q3ij6q2d",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xeg693op-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.5 GB at this context; duplicate traces of the same session: 6ys5pysy, awjyurh6, dpnwzi7t; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 877.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 1k 877.6 96.8 142.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xeg693op",
   "source_url": "https://omlx.ai/benchmarks/performance/xeg693op",
   "tg_tokens": null,
   "tps": 96.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xeg693op-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.1 GB at this context; duplicate traces of the same session: 6ys5pysy, awjyurh6, dpnwzi7t; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1059.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 4k 1,059 78.6 143.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xeg693op",
   "source_url": "https://omlx.ai/benchmarks/performance/xeg693op",
   "tg_tokens": null,
   "tps": 78.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xeg693op-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.3 GB at this context; duplicate traces of the same session: 6ys5pysy, awjyurh6, dpnwzi7t; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1046.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 8k 1,046 88.2 143.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xeg693op",
   "source_url": "https://omlx.ai/benchmarks/performance/xeg693op",
   "tg_tokens": null,
   "tps": 88.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xeg693op-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.6 GB at this context; duplicate traces of the same session: 6ys5pysy, awjyurh6, dpnwzi7t; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1037.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 16k 1,037 82.8 143.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xeg693op",
   "source_url": "https://omlx.ai/benchmarks/performance/xeg693op",
   "tg_tokens": null,
   "tps": 82.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-ptrs9igl-ctx1024",
   "model": "Gemma 4 E2B",
   "notes": "it; oMLX community benchmark trace; peak memory 5.1 GB at this context; duplicate traces of the same session: t18nm33f; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2194.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-E2B-it-8bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 8bit; 1k 2,194 36.9 5.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ptrs9igl",
   "source_url": "https://omlx.ai/benchmarks/performance/ptrs9igl",
   "tg_tokens": null,
   "tps": 36.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-ptrs9igl-ctx4096",
   "model": "Gemma 4 E2B",
   "notes": "it; oMLX community benchmark trace; peak memory 5.2 GB at this context; duplicate traces of the same session: t18nm33f; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2261.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-E2B-it-8bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 8bit; 4k 2,261 36.1 5.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ptrs9igl",
   "source_url": "https://omlx.ai/benchmarks/performance/ptrs9igl",
   "tg_tokens": null,
   "tps": 36.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-oyeadut3-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 175.5 GB; TTFT 37206 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1761.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k tokens 1,761 PP tok/s 53.3 TG tok/s 37206 TTFT (ms) 175.5 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oyeadut3",
   "source_url": "https://omlx.ai/benchmarks/performance/oyeadut3",
   "tg_tokens": null,
   "tps": 53.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-aub08kxz-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 244.5 GB at this context; duplicate traces of the same session: 10pjcayk, g7sinfb7, ursrhulr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 549.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 4k 549.9 32.8 244.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aub08kxz",
   "source_url": "https://omlx.ai/benchmarks/performance/aub08kxz",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-aub08kxz-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 245.1 GB at this context; duplicate traces of the same session: 10pjcayk, g7sinfb7, ursrhulr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 492.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 8k 492.5 35.4 245.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aub08kxz",
   "source_url": "https://omlx.ai/benchmarks/performance/aub08kxz",
   "tg_tokens": null,
   "tps": 35.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-aub08kxz-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 245.2 GB at this context; duplicate traces of the same session: 10pjcayk, g7sinfb7, ursrhulr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 527.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 16k 527.2 35.1 245.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aub08kxz",
   "source_url": "https://omlx.ai/benchmarks/performance/aub08kxz",
   "tg_tokens": null,
   "tps": 35.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-aub08kxz-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 245.4 GB at this context; duplicate traces of the same session: 10pjcayk, g7sinfb7, ursrhulr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 520.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 32k 520.2 34.9 245.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aub08kxz",
   "source_url": "https://omlx.ai/benchmarks/performance/aub08kxz",
   "tg_tokens": null,
   "tps": 34.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-woscm1kv-ctx1024",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: 5ilxj8s9, l8xdgf1i; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 677.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 1k 677.8 81.1 18.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/woscm1kv",
   "source_url": "https://omlx.ai/benchmarks/performance/woscm1kv",
   "tg_tokens": null,
   "tps": 81.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-woscm1kv-ctx4096",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 5ilxj8s9, l8xdgf1i; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 720.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 720.0 87.1 19.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/woscm1kv",
   "source_url": "https://omlx.ai/benchmarks/performance/woscm1kv",
   "tg_tokens": null,
   "tps": 87.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-woscm1kv-ctx16384",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 5ilxj8s9, l8xdgf1i; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 675.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 16k 675.9 77.6 19.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/woscm1kv",
   "source_url": "https://omlx.ai/benchmarks/performance/woscm1kv",
   "tg_tokens": null,
   "tps": 77.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-viknro83-ctx1024",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: x3d1htjc, zz80a89r; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 870.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ6e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 6bit; 1k 870.4 68.8 27.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/viknro83",
   "source_url": "https://omlx.ai/benchmarks/performance/viknro83",
   "tg_tokens": null,
   "tps": 68.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-viknro83-ctx4096",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: x3d1htjc, zz80a89r; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 934.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ6e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 6bit; 4k 934.4 62.9 28.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/viknro83",
   "source_url": "https://omlx.ai/benchmarks/performance/viknro83",
   "tg_tokens": null,
   "tps": 62.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-viknro83-ctx16384",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: x3d1htjc, zz80a89r; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 884.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ6e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 6bit; 16k 884.7 59.6 28.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/viknro83",
   "source_url": "https://omlx.ai/benchmarks/performance/viknro83",
   "tg_tokens": null,
   "tps": 59.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-3f2x5ot2-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 23.7 GB; TTFT 52457 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 624.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 32k tokens 624.7 PP tok/s 70.7 TG tok/s 52457 TTFT (ms) 23.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3f2x5ot2",
   "source_url": "https://omlx.ai/benchmarks/performance/3f2x5ot2",
   "tg_tokens": null,
   "tps": 70.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-doq5du7d-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 27.4 GB; TTFT 531729 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 246.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 128k tokens 246.5 PP tok/s 17.6 TG tok/s 531729 TTFT (ms) 27.4 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/doq5du7d",
   "source_url": "https://omlx.ai/benchmarks/performance/doq5du7d",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rvmlet1s-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.7 GB at this context; duplicate traces of the same session: 7axvgcyy, k8o2gem2, vigw0e1k; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1256.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 1,256 31.0 183.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rvmlet1s",
   "source_url": "https://omlx.ai/benchmarks/performance/rvmlet1s",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rvmlet1s-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.9 GB at this context; duplicate traces of the same session: 7axvgcyy, k8o2gem2, vigw0e1k; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1354.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 1,354 73.8 183.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rvmlet1s",
   "source_url": "https://omlx.ai/benchmarks/performance/rvmlet1s",
   "tg_tokens": null,
   "tps": 73.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rvmlet1s-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 184.2 GB at this context; duplicate traces of the same session: 7axvgcyy, k8o2gem2, vigw0e1k; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1340.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k 1,340 78.1 184.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rvmlet1s",
   "source_url": "https://omlx.ai/benchmarks/performance/rvmlet1s",
   "tg_tokens": null,
   "tps": 78.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rvmlet1s-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 184.8 GB at this context; duplicate traces of the same session: 7axvgcyy, k8o2gem2, vigw0e1k; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1337.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 32k 1,337 66.1 184.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rvmlet1s",
   "source_url": "https://omlx.ai/benchmarks/performance/rvmlet1s",
   "tg_tokens": null,
   "tps": 66.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-epsp4nap-ctx65536",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 180.1 GB; TTFT 34987 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1873.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k tokens 1,873 PP tok/s 64.3 TG tok/s 34987 TTFT (ms) 180.1 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/epsp4nap",
   "source_url": "https://omlx.ai/benchmarks/performance/epsp4nap",
   "tg_tokens": null,
   "tps": 64.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-e0wx6n2b-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 182.8 GB; TTFT 34333 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1909.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k tokens 1,909 PP tok/s 46.1 TG tok/s 34333 TTFT (ms) 182.8 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e0wx6n2b",
   "source_url": "https://omlx.ai/benchmarks/performance/e0wx6n2b",
   "tg_tokens": null,
   "tps": 46.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-knq3jjqi-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.7 GB at this context; duplicate traces of the same session: c0q62gra, yr1p1smo; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 466.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 466.4 30.0 177.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/knq3jjqi",
   "source_url": "https://omlx.ai/benchmarks/performance/knq3jjqi",
   "tg_tokens": null,
   "tps": 30.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-knq3jjqi-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.0 GB at this context; duplicate traces of the same session: c0q62gra, yr1p1smo; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 486.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 486.0 28.6 179.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/knq3jjqi",
   "source_url": "https://omlx.ai/benchmarks/performance/knq3jjqi",
   "tg_tokens": null,
   "tps": 28.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-knq3jjqi-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.0 GB at this context; duplicate traces of the same session: c0q62gra, yr1p1smo; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 259.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 259.5 28.0 180.0 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/knq3jjqi",
   "source_url": "https://omlx.ai/benchmarks/performance/knq3jjqi",
   "tg_tokens": null,
   "tps": 28.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-80grrk39-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Jundot build; oMLX community benchmark trace; peak memory 99.1 GB at this context; duplicate traces of the same session: cy5d3kc7, xtqod2dr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 652.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 652.3 70.3 99.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/80grrk39",
   "source_url": "https://omlx.ai/benchmarks/performance/80grrk39",
   "tg_tokens": null,
   "tps": 70.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-80grrk39-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Jundot build; oMLX community benchmark trace; peak memory 99.6 GB at this context; duplicate traces of the same session: cy5d3kc7, xtqod2dr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 829.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 829.9 57.0 99.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/80grrk39",
   "source_url": "https://omlx.ai/benchmarks/performance/80grrk39",
   "tg_tokens": null,
   "tps": 57.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-80grrk39-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Jundot build; oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: cy5d3kc7, xtqod2dr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 836.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 836.5 54.0 101.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/80grrk39",
   "source_url": "https://omlx.ai/benchmarks/performance/80grrk39",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-r4ovi55b-ctx1024",
   "model": "Qwen3 8B",
   "notes": "oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: 3hscixtc; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1085.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-8B-4bit on M6 (12c); M6 (12c) \u00b7 16 GB \u00b7 4bit; 1k 1,085 29.6 4.9 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r4ovi55b",
   "source_url": "https://omlx.ai/benchmarks/performance/r4ovi55b",
   "tg_tokens": null,
   "tps": 29.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-r4ovi55b-ctx4096",
   "model": "Qwen3 8B",
   "notes": "oMLX community benchmark trace; peak memory 5.4 GB at this context; duplicate traces of the same session: 3hscixtc; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1041.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-8B-4bit on M6 (12c); M6 (12c) \u00b7 16 GB \u00b7 4bit; 4k 1,041 26.2 5.4 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r4ovi55b",
   "source_url": "https://omlx.ai/benchmarks/performance/r4ovi55b",
   "tg_tokens": null,
   "tps": 26.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-o1anc90p-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 1245 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 822.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k tokens 822.8 PP tok/s 85.9 TG tok/s 1245 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o1anc90p",
   "source_url": "https://omlx.ai/benchmarks/performance/o1anc90p",
   "tg_tokens": null,
   "tps": 85.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-zgl5byt5-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 20.7 GB; TTFT 10060 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3257.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 32k tokens 3,257 PP tok/s 105.4 TG tok/s 10060 TTFT (ms) 20.7 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zgl5byt5",
   "source_url": "https://omlx.ai/benchmarks/performance/zgl5byt5",
   "tg_tokens": null,
   "tps": 105.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-b6ca61dy-ctx1024",
   "model": "Qwen3.6-40B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 41.2 GB at this context; duplicate traces of the same session: 32m2oihs; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 69.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-40B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-Thinking-8bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 1k 69.4 5.9 41.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b6ca61dy",
   "source_url": "https://omlx.ai/benchmarks/performance/b6ca61dy",
   "tg_tokens": null,
   "tps": 5.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-b6ca61dy-ctx4096",
   "model": "Qwen3.6-40B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 46.0 GB at this context; duplicate traces of the same session: 32m2oihs; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 68.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-40B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-Thinking-8bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 4k 68.6 5.8 46.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b6ca61dy",
   "source_url": "https://omlx.ai/benchmarks/performance/b6ca61dy",
   "tg_tokens": null,
   "tps": 5.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-36gibv6r-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 4we1s5fi, ksp3e97e; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 424.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 424.6 70.2 18.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/36gibv6r",
   "source_url": "https://omlx.ai/benchmarks/performance/36gibv6r",
   "tg_tokens": null,
   "tps": 70.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-36gibv6r-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: 4we1s5fi, ksp3e97e; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 435.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 435.4 73.3 22.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/36gibv6r",
   "source_url": "https://omlx.ai/benchmarks/performance/36gibv6r",
   "tg_tokens": null,
   "tps": 73.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-36gibv6r-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 4we1s5fi, ksp3e97e; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 382.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 382.7 87.8 23.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/36gibv6r",
   "source_url": "https://omlx.ai/benchmarks/performance/36gibv6r",
   "tg_tokens": null,
   "tps": 87.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-6nvgc9nf-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 21.4 GB; TTFT 87958 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 372.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k tokens 372.5 PP tok/s 51.9 TG tok/s 87958 TTFT (ms) 21.4 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6nvgc9nf",
   "source_url": "https://omlx.ai/benchmarks/performance/6nvgc9nf",
   "tg_tokens": null,
   "tps": 51.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-s668fsaq-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 61.3 GB at this context; duplicate traces of the same session: j6pjg7bh, opznul54; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 622.8,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 3bit; 1k 622.8 62.7 61.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s668fsaq",
   "source_url": "https://omlx.ai/benchmarks/performance/s668fsaq",
   "tg_tokens": null,
   "tps": 62.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-s668fsaq-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 62.2 GB at this context; duplicate traces of the same session: j6pjg7bh, opznul54; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1041.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 3bit; 8k 1,041 58.3 62.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s668fsaq",
   "source_url": "https://omlx.ai/benchmarks/performance/s668fsaq",
   "tg_tokens": null,
   "tps": 58.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-s668fsaq-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 62.5 GB at this context; duplicate traces of the same session: j6pjg7bh, opznul54; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1041.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 3bit; 16k 1,041 68.8 62.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s668fsaq",
   "source_url": "https://omlx.ai/benchmarks/performance/s668fsaq",
   "tg_tokens": null,
   "tps": 68.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nqi6dltt-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: e7d9yu3j; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1085.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,085 87.2 100.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nqi6dltt",
   "source_url": "https://omlx.ai/benchmarks/performance/nqi6dltt",
   "tg_tokens": null,
   "tps": 87.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nqi6dltt-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 107.2 GB at this context; duplicate traces of the same session: e7d9yu3j; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 809.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 809.1 80.0 107.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nqi6dltt",
   "source_url": "https://omlx.ai/benchmarks/performance/nqi6dltt",
   "tg_tokens": null,
   "tps": 80.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-bm3xcwyi-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.7 GB at this context; duplicate traces of the same session: ptdgvtxq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1094.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,094 83.5 100.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bm3xcwyi",
   "source_url": "https://omlx.ai/benchmarks/performance/bm3xcwyi",
   "tg_tokens": null,
   "tps": 83.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-bm3xcwyi-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.3 GB at this context; duplicate traces of the same session: ptdgvtxq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 985.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 985.9 85.3 102.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bm3xcwyi",
   "source_url": "https://omlx.ai/benchmarks/performance/bm3xcwyi",
   "tg_tokens": null,
   "tps": 85.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-49pprbhi-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: 4n5xn7kg, bhebp7sd, kmq8omr1, nqmlrwbj, pbpotl4s, z3g1j63n; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 110.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 6bit; 1k 110.4 13.0 22.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/49pprbhi",
   "source_url": "https://omlx.ai/benchmarks/performance/49pprbhi",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-49pprbhi-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: 4n5xn7kg, bhebp7sd, kmq8omr1, nqmlrwbj, pbpotl4s, z3g1j63n; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 121.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 6bit; 4k 121.4 12.7 26.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/49pprbhi",
   "source_url": "https://omlx.ai/benchmarks/performance/49pprbhi",
   "tg_tokens": null,
   "tps": 12.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-49pprbhi-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: 4n5xn7kg, bhebp7sd, kmq8omr1, nqmlrwbj, pbpotl4s, z3g1j63n; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 118.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 6bit; 8k 118.7 12.2 27.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/49pprbhi",
   "source_url": "https://omlx.ai/benchmarks/performance/49pprbhi",
   "tg_tokens": null,
   "tps": 12.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-49pprbhi-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 4n5xn7kg, bhebp7sd, kmq8omr1, nqmlrwbj, pbpotl4s, z3g1j63n; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 114.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 6bit; 16k 114.1 11.2 28.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/49pprbhi",
   "source_url": "https://omlx.ai/benchmarks/performance/49pprbhi",
   "tg_tokens": null,
   "tps": 11.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-49pprbhi-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: 4n5xn7kg, bhebp7sd, kmq8omr1, nqmlrwbj, pbpotl4s, z3g1j63n; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 6bit; 32k 104.0 10.0 29.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/49pprbhi",
   "source_url": "https://omlx.ai/benchmarks/performance/49pprbhi",
   "tg_tokens": null,
   "tps": 10.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-49pprbhi-ctx65536",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 31.5 GB at this context; duplicate traces of the same session: 4n5xn7kg, bhebp7sd, kmq8omr1, nqmlrwbj, pbpotl4s, z3g1j63n; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 90.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 6bit; 64k 90.0 7.8 31.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/49pprbhi",
   "source_url": "https://omlx.ai/benchmarks/performance/49pprbhi",
   "tg_tokens": null,
   "tps": 7.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-49pprbhi-ctx131072",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: 4n5xn7kg, bhebp7sd, kmq8omr1, nqmlrwbj, pbpotl4s, z3g1j63n; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 69.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 6bit; 128k 69.6 5.7 35.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/49pprbhi",
   "source_url": "https://omlx.ai/benchmarks/performance/49pprbhi",
   "tg_tokens": null,
   "tps": 5.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-p23t7zqu-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "MTPLX Optimized Speed build; oMLX community benchmark trace; peak memory 3.8 GB at this context; duplicate traces of the same session: zw4gclll; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1253.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MTPLX-Optimized-Speed on M5 (10c); M5 (10c) \u00b7 16 GB \u00b7 4bit; 4k 1,253 43.4 3.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p23t7zqu",
   "source_url": "https://omlx.ai/benchmarks/performance/p23t7zqu",
   "tg_tokens": null,
   "tps": 43.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-p23t7zqu-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "MTPLX Optimized Speed build; oMLX community benchmark trace; peak memory 4.1 GB at this context; duplicate traces of the same session: zw4gclll; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1091.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MTPLX-Optimized-Speed on M5 (10c); M5 (10c) \u00b7 16 GB \u00b7 4bit; 16k 1,091 36.2 4.1 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p23t7zqu",
   "source_url": "https://omlx.ai/benchmarks/performance/p23t7zqu",
   "tg_tokens": null,
   "tps": 36.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-a84mbnzy-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace, single run at 128k context; peak memory 46.5 GB; TTFT 273188 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 479.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k tokens 479.8 PP tok/s 13.6 TG tok/s 273188 TTFT (ms) 46.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a84mbnzy",
   "source_url": "https://omlx.ai/benchmarks/performance/a84mbnzy",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0ede7ar4-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace, single run at 1k context; peak memory 34.6 GB; TTFT 1433 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 714.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k tokens 714.7 PP tok/s 38.2 TG tok/s 1433 TTFT (ms) 34.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ede7ar4",
   "source_url": "https://omlx.ai/benchmarks/performance/0ede7ar4",
   "tg_tokens": null,
   "tps": 38.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9yt5tpj3-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 63.1 GB at this context; duplicate traces of the same session: 41swze7w, yx39einf, zmq0t770; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1037.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 3bit; 32k 1,037 44.5 63.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9yt5tpj3",
   "source_url": "https://omlx.ai/benchmarks/performance/9yt5tpj3",
   "tg_tokens": null,
   "tps": 44.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9yt5tpj3-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 64.8 GB at this context; duplicate traces of the same session: 41swze7w, yx39einf, zmq0t770; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1021.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 3bit; 64k 1,021 45.9 64.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9yt5tpj3",
   "source_url": "https://omlx.ai/benchmarks/performance/9yt5tpj3",
   "tg_tokens": null,
   "tps": 45.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9yt5tpj3-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.1 GB at this context; duplicate traces of the same session: 41swze7w, yx39einf, zmq0t770; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 998.3,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 3bit; 128k 998.3 46.0 69.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9yt5tpj3",
   "source_url": "https://omlx.ai/benchmarks/performance/9yt5tpj3",
   "tg_tokens": null,
   "tps": 46.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9yt5tpj3-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 41swze7w, yx39einf, zmq0t770; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 972.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 3bit; 195k 972.0 48.9 70.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9yt5tpj3",
   "source_url": "https://omlx.ai/benchmarks/performance/9yt5tpj3",
   "tg_tokens": null,
   "tps": 48.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1im0zq22-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.5 GB; TTFT 935 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1095.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,095 PP tok/s 86.1 TG tok/s 935 TTFT (ms) 70.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1im0zq22",
   "source_url": "https://omlx.ai/benchmarks/performance/1im0zq22",
   "tg_tokens": null,
   "tps": 86.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-icw633y6-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.5 GB; TTFT 2683 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 381.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 381.7 PP tok/s 23.7 TG tok/s 2683 TTFT (ms) 70.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/icw633y6",
   "source_url": "https://omlx.ai/benchmarks/performance/icw633y6",
   "tg_tokens": null,
   "tps": 23.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-de7n0qap-ctx1024",
   "model": "VulnLLM-R-7B",
   "notes": "oMLX community benchmark trace; peak memory 8.1 GB at this context; duplicate traces of the same session: fpl3m66z; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "VulnLLM-R-7B-8bit on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 8bit; 1k 105.9 7.5 8.1 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/de7n0qap",
   "source_url": "https://omlx.ai/benchmarks/performance/de7n0qap",
   "tg_tokens": null,
   "tps": 7.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-de7n0qap-ctx4096",
   "model": "VulnLLM-R-7B",
   "notes": "oMLX community benchmark trace; peak memory 8.3 GB at this context; duplicate traces of the same session: fpl3m66z; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "VulnLLM-R-7B-8bit on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 8bit; 4k 103.6 7.1 8.3 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/de7n0qap",
   "source_url": "https://omlx.ai/benchmarks/performance/de7n0qap",
   "tg_tokens": null,
   "tps": 7.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xcq0no09-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 252.2 GB at this context; duplicate traces of the same session: o37pd7ua, pdhzkb2a, sfbfkt11; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 470.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-Q8mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 8k 470.0 28.3 252.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xcq0no09",
   "source_url": "https://omlx.ai/benchmarks/performance/xcq0no09",
   "tg_tokens": null,
   "tps": 28.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xcq0no09-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 252.4 GB at this context; duplicate traces of the same session: o37pd7ua, pdhzkb2a, sfbfkt11; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 499.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-Q8mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 16k 499.2 38.9 252.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xcq0no09",
   "source_url": "https://omlx.ai/benchmarks/performance/xcq0no09",
   "tg_tokens": null,
   "tps": 38.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xcq0no09-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 252.6 GB at this context; duplicate traces of the same session: o37pd7ua, pdhzkb2a, sfbfkt11; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 489.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-Q8mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 32k 489.9 39.1 252.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xcq0no09",
   "source_url": "https://omlx.ai/benchmarks/performance/xcq0no09",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xcq0no09-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 253.3 GB at this context; duplicate traces of the same session: o37pd7ua, pdhzkb2a, sfbfkt11; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 481.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-Q8mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 64k 481.5 42.9 253.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xcq0no09",
   "source_url": "https://omlx.ai/benchmarks/performance/xcq0no09",
   "tg_tokens": null,
   "tps": 42.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-haqxbawj-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 137.6 GB at this context; duplicate traces of the same session: dkrqcdoz, fu9lvith, iixav7ad; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1069.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ8-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 1,069 55.3 137.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/haqxbawj",
   "source_url": "https://omlx.ai/benchmarks/performance/haqxbawj",
   "tg_tokens": null,
   "tps": 55.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-haqxbawj-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 137.7 GB at this context; duplicate traces of the same session: dkrqcdoz, fu9lvith, iixav7ad; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1328.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ8-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k 1,328 69.0 137.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/haqxbawj",
   "source_url": "https://omlx.ai/benchmarks/performance/haqxbawj",
   "tg_tokens": null,
   "tps": 69.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-haqxbawj-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 138.4 GB at this context; duplicate traces of the same session: dkrqcdoz, fu9lvith, iixav7ad; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1323.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ8-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 32k 1,323 71.7 138.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/haqxbawj",
   "source_url": "https://omlx.ai/benchmarks/performance/haqxbawj",
   "tg_tokens": null,
   "tps": 71.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-haqxbawj-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 140.5 GB at this context; duplicate traces of the same session: dkrqcdoz, fu9lvith, iixav7ad; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1310.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ8-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 64k 1,310 63.0 140.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/haqxbawj",
   "source_url": "https://omlx.ai/benchmarks/performance/haqxbawj",
   "tg_tokens": null,
   "tps": 63.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9ieic1b6-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 136.9 GB at this context; duplicate traces of the same session: obb0c0f2, ys3ehi8d; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 682.6,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ8-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 682.6 27.5 136.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ieic1b6",
   "source_url": "https://omlx.ai/benchmarks/performance/9ieic1b6",
   "tg_tokens": null,
   "tps": 27.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9ieic1b6-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 140.7 GB at this context; duplicate traces of the same session: obb0c0f2, ys3ehi8d; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1306.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ8-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 64k 1,306 55.7 140.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ieic1b6",
   "source_url": "https://omlx.ai/benchmarks/performance/9ieic1b6",
   "tg_tokens": null,
   "tps": 55.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9ieic1b6-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 145.2 GB at this context; duplicate traces of the same session: obb0c0f2, ys3ehi8d; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1283.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ8-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 128k 1,283 51.5 145.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ieic1b6",
   "source_url": "https://omlx.ai/benchmarks/performance/9ieic1b6",
   "tg_tokens": null,
   "tps": 51.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-zvh24i8z-ctx131072",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 77.2 GB; TTFT 57510 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2279.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k tokens 2,279 PP tok/s 61.7 TG tok/s 57510 TTFT (ms) 77.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zvh24i8z",
   "source_url": "https://omlx.ai/benchmarks/performance/zvh24i8z",
   "tg_tokens": null,
   "tps": 61.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-zts8x875-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: 35vat8c0, dzo5g7ut, hkgqlfan, ztz6hhdy; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2650.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 2,650 120.8 100.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zts8x875",
   "source_url": "https://omlx.ai/benchmarks/performance/zts8x875",
   "tg_tokens": null,
   "tps": 120.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-zts8x875-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.1 GB at this context; duplicate traces of the same session: 35vat8c0, dzo5g7ut, hkgqlfan, ztz6hhdy; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4718.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 4,718 135.5 102.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zts8x875",
   "source_url": "https://omlx.ai/benchmarks/performance/zts8x875",
   "tg_tokens": null,
   "tps": 135.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-zts8x875-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.6 GB at this context; duplicate traces of the same session: 35vat8c0, dzo5g7ut, hkgqlfan, ztz6hhdy; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4658.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 4,658 132.0 104.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zts8x875",
   "source_url": "https://omlx.ai/benchmarks/performance/zts8x875",
   "tg_tokens": null,
   "tps": 132.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-zts8x875-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.0 GB at this context; duplicate traces of the same session: 35vat8c0, dzo5g7ut, hkgqlfan, ztz6hhdy; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4501.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 4,501 103.3 109.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zts8x875",
   "source_url": "https://omlx.ai/benchmarks/performance/zts8x875",
   "tg_tokens": null,
   "tps": 103.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-zts8x875-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.7 GB at this context; duplicate traces of the same session: 35vat8c0, dzo5g7ut, hkgqlfan, ztz6hhdy; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4368.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 4,368 87.9 110.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zts8x875",
   "source_url": "https://omlx.ai/benchmarks/performance/zts8x875",
   "tg_tokens": null,
   "tps": 87.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-gu0f78jd-ctx131072",
   "model": "Huihui-Qwen3.8-27B-abliterated",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 24.9 GB; TTFT 201512 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 650.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Huihui-Qwen3.8-27B-abliterated-mlx-4Bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k tokens 650.4 PP tok/s 21.7 TG tok/s 201512 TTFT (ms) 24.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gu0f78jd",
   "source_url": "https://omlx.ai/benchmarks/performance/gu0f78jd",
   "tg_tokens": null,
   "tps": 21.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-p6ngg1ym-ctx131072",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 38.6 GB; TTFT 51116 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2564.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 128k tokens 2,564 PP tok/s 66.1 TG tok/s 51116 TTFT (ms) 38.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p6ngg1ym",
   "source_url": "https://omlx.ai/benchmarks/performance/p6ngg1ym",
   "tg_tokens": null,
   "tps": 66.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1ez3fiuy-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace, single run at 4k context; peak memory 34.9 GB; TTFT 9071 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 451.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k tokens 451.5 PP tok/s 19.6 TG tok/s 9071 TTFT (ms) 34.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ez3fiuy",
   "source_url": "https://omlx.ai/benchmarks/performance/1ez3fiuy",
   "tg_tokens": null,
   "tps": 19.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-tswrzyzu-ctx32768",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 37.4 GB; TTFT 67120 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 488.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 32k tokens 488.2 PP tok/s 13.8 TG tok/s 67120 TTFT (ms) 37.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tswrzyzu",
   "source_url": "https://omlx.ai/benchmarks/performance/tswrzyzu",
   "tg_tokens": null,
   "tps": 13.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3m5a3jjo-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace, single run at 1k context; peak memory 34.6 GB; TTFT 1296 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 789.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k tokens 789.9 PP tok/s 42.5 TG tok/s 1296 TTFT (ms) 34.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3m5a3jjo",
   "source_url": "https://omlx.ai/benchmarks/performance/3m5a3jjo",
   "tg_tokens": null,
   "tps": 42.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vj5v8kv2-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace, single run at 1k context; peak memory 33.7 GB; TTFT 1260 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 812.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k tokens 812.5 PP tok/s 18.8 TG tok/s 1260 TTFT (ms) 33.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vj5v8kv2",
   "source_url": "https://omlx.ai/benchmarks/performance/vj5v8kv2",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-50vqd9me-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace, single run at 1k context; peak memory 33.7 GB; TTFT 1262 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 811.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k tokens 811.4 PP tok/s 18.8 TG tok/s 1262 TTFT (ms) 33.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/50vqd9me",
   "source_url": "https://omlx.ai/benchmarks/performance/50vqd9me",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ew8y65lv-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace, single run at 1k context; peak memory 39.9 GB; TTFT 1297 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 789.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k tokens 789.7 PP tok/s 49.7 TG tok/s 1297 TTFT (ms) 39.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ew8y65lv",
   "source_url": "https://omlx.ai/benchmarks/performance/ew8y65lv",
   "tg_tokens": null,
   "tps": 49.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-es953j3t-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace, single run at 1k context; peak memory 34.6 GB; TTFT 1294 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 791.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k tokens 791.3 PP tok/s 53.4 TG tok/s 1294 TTFT (ms) 34.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/es953j3t",
   "source_url": "https://omlx.ai/benchmarks/performance/es953j3t",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ev5yka8l-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace, single run at 1k context; peak memory 34.6 GB; TTFT 1295 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 790.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k tokens 790.5 PP tok/s 45.9 TG tok/s 1295 TTFT (ms) 34.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ev5yka8l",
   "source_url": "https://omlx.ai/benchmarks/performance/ev5yka8l",
   "tg_tokens": null,
   "tps": 45.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-o9u2gi0a-ctx1024",
   "model": "Apodex-1.1-mini",
   "notes": "baseline 20260915; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 17hhydpt, tx9g15t0, u145uz3h, w3mizepy; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 779.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Apodex-1.1-mini-oQ4e-baseline-20260915 on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 779.9 74.6 20.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9u2gi0a",
   "source_url": "https://omlx.ai/benchmarks/performance/o9u2gi0a",
   "tg_tokens": null,
   "tps": 74.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-o9u2gi0a-ctx4096",
   "model": "Apodex-1.1-mini",
   "notes": "baseline 20260915; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: 17hhydpt, tx9g15t0, u145uz3h, w3mizepy; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 827.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Apodex-1.1-mini-oQ4e-baseline-20260915 on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 827.1 71.5 20.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9u2gi0a",
   "source_url": "https://omlx.ai/benchmarks/performance/o9u2gi0a",
   "tg_tokens": null,
   "tps": 71.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-o9u2gi0a-ctx8192",
   "model": "Apodex-1.1-mini",
   "notes": "baseline 20260915; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 17hhydpt, tx9g15t0, u145uz3h, w3mizepy; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 784.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Apodex-1.1-mini-oQ4e-baseline-20260915 on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 784.8 68.3 20.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9u2gi0a",
   "source_url": "https://omlx.ai/benchmarks/performance/o9u2gi0a",
   "tg_tokens": null,
   "tps": 68.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-o9u2gi0a-ctx16384",
   "model": "Apodex-1.1-mini",
   "notes": "baseline 20260915; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 17hhydpt, tx9g15t0, u145uz3h, w3mizepy; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 686.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Apodex-1.1-mini-oQ4e-baseline-20260915 on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 686.4 62.4 21.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9u2gi0a",
   "source_url": "https://omlx.ai/benchmarks/performance/o9u2gi0a",
   "tg_tokens": null,
   "tps": 62.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-o9u2gi0a-ctx32768",
   "model": "Apodex-1.1-mini",
   "notes": "baseline 20260915; oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: 17hhydpt, tx9g15t0, u145uz3h, w3mizepy; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 565.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Apodex-1.1-mini-oQ4e-baseline-20260915 on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 565.4 54.1 21.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9u2gi0a",
   "source_url": "https://omlx.ai/benchmarks/performance/o9u2gi0a",
   "tg_tokens": null,
   "tps": 54.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-fhfga9rt-ctx1024",
   "model": "Apodex-1.1-mini",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 9l3lcdsv, f43pcub7, q97x10yd, sh8t2r92; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 729.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Apodex-1.1-mini-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 729.2 73.1 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fhfga9rt",
   "source_url": "https://omlx.ai/benchmarks/performance/fhfga9rt",
   "tg_tokens": null,
   "tps": 73.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-fhfga9rt-ctx4096",
   "model": "Apodex-1.1-mini",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: 9l3lcdsv, f43pcub7, q97x10yd, sh8t2r92; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 821.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Apodex-1.1-mini-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 821.8 70.8 20.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fhfga9rt",
   "source_url": "https://omlx.ai/benchmarks/performance/fhfga9rt",
   "tg_tokens": null,
   "tps": 70.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-fhfga9rt-ctx8192",
   "model": "Apodex-1.1-mini",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 9l3lcdsv, f43pcub7, q97x10yd, sh8t2r92; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 777.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Apodex-1.1-mini-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 777.3 66.5 20.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fhfga9rt",
   "source_url": "https://omlx.ai/benchmarks/performance/fhfga9rt",
   "tg_tokens": null,
   "tps": 66.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-fhfga9rt-ctx16384",
   "model": "Apodex-1.1-mini",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 9l3lcdsv, f43pcub7, q97x10yd, sh8t2r92; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 697.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Apodex-1.1-mini-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 697.3 61.3 21.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fhfga9rt",
   "source_url": "https://omlx.ai/benchmarks/performance/fhfga9rt",
   "tg_tokens": null,
   "tps": 61.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-fhfga9rt-ctx32768",
   "model": "Apodex-1.1-mini",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: 9l3lcdsv, f43pcub7, q97x10yd, sh8t2r92; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 570.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Apodex-1.1-mini-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 570.8 53.5 21.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fhfga9rt",
   "source_url": "https://omlx.ai/benchmarks/performance/fhfga9rt",
   "tg_tokens": null,
   "tps": 53.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-bw86iztt-ctx1024",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: jkxbny8i, y33t3gm3; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 673.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ4e-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 1k 673.7 71.5 22.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bw86iztt",
   "source_url": "https://omlx.ai/benchmarks/performance/bw86iztt",
   "tg_tokens": null,
   "tps": 71.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-bw86iztt-ctx8192",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: jkxbny8i, y33t3gm3; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 717.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ4e-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 8k 717.7 68.2 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bw86iztt",
   "source_url": "https://omlx.ai/benchmarks/performance/bw86iztt",
   "tg_tokens": null,
   "tps": 68.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-bw86iztt-ctx32768",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: jkxbny8i, y33t3gm3; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 627.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ4e-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 32k 627.7 53.7 23.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bw86iztt",
   "source_url": "https://omlx.ai/benchmarks/performance/bw86iztt",
   "tg_tokens": null,
   "tps": 53.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-fzyhaiug-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.6 GB at this context; duplicate traces of the same session: 7o49ncjo; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2614.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 2,614 87.7 35.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fzyhaiug",
   "source_url": "https://omlx.ai/benchmarks/performance/fzyhaiug",
   "tg_tokens": null,
   "tps": 87.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-fzyhaiug-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: 7o49ncjo; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3363.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 3,363 81.0 37.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fzyhaiug",
   "source_url": "https://omlx.ai/benchmarks/performance/fzyhaiug",
   "tg_tokens": null,
   "tps": 81.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-n5rszwyz-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: 6wfbt3nh; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1530.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,530 112.1 5.9 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n5rszwyz",
   "source_url": "https://omlx.ai/benchmarks/performance/n5rszwyz",
   "tg_tokens": null,
   "tps": 112.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-n5rszwyz-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.2 GB at this context; duplicate traces of the same session: 6wfbt3nh; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1653.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,653 109.6 7.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n5rszwyz",
   "source_url": "https://omlx.ai/benchmarks/performance/n5rszwyz",
   "tg_tokens": null,
   "tps": 109.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-yru4xqd3-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: aa8934dn; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 860.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 860.9 88.2 21.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yru4xqd3",
   "source_url": "https://omlx.ai/benchmarks/performance/yru4xqd3",
   "tg_tokens": null,
   "tps": 88.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-yru4xqd3-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: aa8934dn; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 746.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 16k 746.7 75.2 22.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yru4xqd3",
   "source_url": "https://omlx.ai/benchmarks/performance/yru4xqd3",
   "tg_tokens": null,
   "tps": 75.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-19ed7u4b-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 13.2 GB; TTFT 3336 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 306.9,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.8-27B-3bit-mlx on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 3bit; 1k tokens 306.9 PP tok/s 11.6 TG tok/s 3336 TTFT (ms) 13.2 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/19ed7u4b",
   "source_url": "https://omlx.ai/benchmarks/performance/19ed7u4b",
   "tg_tokens": null,
   "tps": 11.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-57vrzm77-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 18.4 GB; TTFT 3995 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1025.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 1,025 PP tok/s 56.2 TG tok/s 3995 TTFT (ms) 18.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/57vrzm77",
   "source_url": "https://omlx.ai/benchmarks/performance/57vrzm77",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8jmk5qvu-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 18.4 GB; TTFT 4280 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 956.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 956.9 PP tok/s 51.7 TG tok/s 4280 TTFT (ms) 18.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8jmk5qvu",
   "source_url": "https://omlx.ai/benchmarks/performance/8jmk5qvu",
   "tg_tokens": null,
   "tps": 51.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-52wp5i7f-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace, single run at 64k context; peak memory 39.0 GB; TTFT 182407 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 359.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 64k tokens 359.3 PP tok/s 21.9 TG tok/s 182407 TTFT (ms) 39.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/52wp5i7f",
   "source_url": "https://omlx.ai/benchmarks/performance/52wp5i7f",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-moot0ok3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: 7sym1sso; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 897.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 897.2 73.7 17.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/moot0ok3",
   "source_url": "https://omlx.ai/benchmarks/performance/moot0ok3",
   "tg_tokens": null,
   "tps": 73.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-moot0ok3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 7sym1sso; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1000.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,000 57.3 18.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/moot0ok3",
   "source_url": "https://omlx.ai/benchmarks/performance/moot0ok3",
   "tg_tokens": null,
   "tps": 57.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-b3ahntgj-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace, single run at 1k context; peak memory 41.2 GB; TTFT 1295 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 790.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k tokens 790.6 PP tok/s 48.3 TG tok/s 1295 TTFT (ms) 41.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b3ahntgj",
   "source_url": "https://omlx.ai/benchmarks/performance/b3ahntgj",
   "tg_tokens": null,
   "tps": 48.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-eg7rs57t-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace, single run at 1k context; peak memory 41.2 GB; TTFT 1296 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 790.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k tokens 790.0 PP tok/s 51.9 TG tok/s 1296 TTFT (ms) 41.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eg7rs57t",
   "source_url": "https://omlx.ai/benchmarks/performance/eg7rs57t",
   "tg_tokens": null,
   "tps": 51.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gsf4yxnj-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.1 GB at this context; duplicate traces of the same session: mbhdzo2z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 839.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 839.5 42.7 35.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gsf4yxnj",
   "source_url": "https://omlx.ai/benchmarks/performance/gsf4yxnj",
   "tg_tokens": null,
   "tps": 42.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-gsf4yxnj-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.0 GB at this context; duplicate traces of the same session: mbhdzo2z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 667.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 64k 667.3 40.1 39.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gsf4yxnj",
   "source_url": "https://omlx.ai/benchmarks/performance/gsf4yxnj",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-xf9kddvx-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 86.4 GB at this context; duplicate traces of the same session: 30qgxlq0, i86je3za, kx47lff5, l8v6qjfy, pxf4opyy, vizxp0fm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 543.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 543.9 47.1 86.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xf9kddvx",
   "source_url": "https://omlx.ai/benchmarks/performance/xf9kddvx",
   "tg_tokens": null,
   "tps": 47.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-xf9kddvx-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 87.0 GB at this context; duplicate traces of the same session: 30qgxlq0, i86je3za, kx47lff5, l8v6qjfy, pxf4opyy, vizxp0fm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 684.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 684.0 49.0 87.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xf9kddvx",
   "source_url": "https://omlx.ai/benchmarks/performance/xf9kddvx",
   "tg_tokens": null,
   "tps": 49.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-xf9kddvx-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 87.2 GB at this context; duplicate traces of the same session: 30qgxlq0, i86je3za, kx47lff5, l8v6qjfy, pxf4opyy, vizxp0fm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 811.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 811.7 50.2 87.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xf9kddvx",
   "source_url": "https://omlx.ai/benchmarks/performance/xf9kddvx",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-xf9kddvx-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 87.5 GB at this context; duplicate traces of the same session: 30qgxlq0, i86je3za, kx47lff5, l8v6qjfy, pxf4opyy, vizxp0fm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 807.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 807.7 67.7 87.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xf9kddvx",
   "source_url": "https://omlx.ai/benchmarks/performance/xf9kddvx",
   "tg_tokens": null,
   "tps": 67.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-xf9kddvx-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 89.7 GB at this context; duplicate traces of the same session: 30qgxlq0, i86je3za, kx47lff5, l8v6qjfy, pxf4opyy, vizxp0fm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 794.3,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 794.3 54.1 89.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xf9kddvx",
   "source_url": "https://omlx.ai/benchmarks/performance/xf9kddvx",
   "tg_tokens": null,
   "tps": 54.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-xf9kddvx-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 94.0 GB at this context; duplicate traces of the same session: 30qgxlq0, i86je3za, kx47lff5, l8v6qjfy, pxf4opyy, vizxp0fm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 781.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 781.4 51.3 94.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xf9kddvx",
   "source_url": "https://omlx.ai/benchmarks/performance/xf9kddvx",
   "tg_tokens": null,
   "tps": 51.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-xf9kddvx-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 95.8 GB at this context; duplicate traces of the same session: 30qgxlq0, i86je3za, kx47lff5, l8v6qjfy, pxf4opyy, vizxp0fm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 773.3,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 5bit; 195k 773.3 47.5 95.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xf9kddvx",
   "source_url": "https://omlx.ai/benchmarks/performance/xf9kddvx",
   "tg_tokens": null,
   "tps": 47.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-715q8rqf-ctx4096",
   "model": "gemma-4-26B-A4B-it-uncensored-heretic",
   "notes": "VLM; oMLX community benchmark trace, single run at 4k context; peak memory 20.3 GB; TTFT 1655 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2476.0,
   "provenance": "community",
   "quant": "6Bit",
   "quote": "gemma-4-26B-A4B-it-uncensored-heretic-mlx-vlm-6Bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k tokens 2,476 PP tok/s 100.0 TG tok/s 1655 TTFT (ms) 20.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/715q8rqf",
   "source_url": "https://omlx.ai/benchmarks/performance/715q8rqf",
   "tg_tokens": null,
   "tps": 100.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mod9nbpy-ctx4096",
   "model": "gemma-4-26B-A4B-it-uncensored-heretic",
   "notes": "VLM; oMLX community benchmark trace, single run at 4k context; peak memory 20.3 GB; TTFT 1061 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3861.0,
   "provenance": "community",
   "quant": "6Bit",
   "quote": "gemma-4-26B-A4B-it-uncensored-heretic-mlx-vlm-6Bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k tokens 3,861 PP tok/s 100.1 TG tok/s 1061 TTFT (ms) 20.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mod9nbpy",
   "source_url": "https://omlx.ai/benchmarks/performance/mod9nbpy",
   "tg_tokens": null,
   "tps": 100.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2myij2eo-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: r9l3whsv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1577.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,577 130.2 11.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2myij2eo",
   "source_url": "https://omlx.ai/benchmarks/performance/2myij2eo",
   "tg_tokens": null,
   "tps": 130.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2myij2eo-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: r9l3whsv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2472.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 2,472 124.9 11.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2myij2eo",
   "source_url": "https://omlx.ai/benchmarks/performance/2myij2eo",
   "tg_tokens": null,
   "tps": 124.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-vxmoily0-ctx4096",
   "model": "Llama-3.3-70B-Instruct-abliterated",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 39.0 GB; TTFT 11269 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 363.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Llama-3.3-70B-Instruct-abliterated-mlx-4Bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 363.5 PP tok/s 11.3 TG tok/s 11269 TTFT (ms) 39.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vxmoily0",
   "source_url": "https://omlx.ai/benchmarks/performance/vxmoily0",
   "tg_tokens": null,
   "tps": 11.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fwre1ned-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: miaza5x2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4020.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 4,020 95.7 35.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fwre1ned",
   "source_url": "https://omlx.ai/benchmarks/performance/fwre1ned",
   "tg_tokens": null,
   "tps": 95.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fwre1ned-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: miaza5x2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5777.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 5,777 92.2 36.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fwre1ned",
   "source_url": "https://omlx.ai/benchmarks/performance/fwre1ned",
   "tg_tokens": null,
   "tps": 92.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ziyze0ci-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: y82y8ioy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4305.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,305 108.3 21.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ziyze0ci",
   "source_url": "https://omlx.ai/benchmarks/performance/ziyze0ci",
   "tg_tokens": null,
   "tps": 108.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ziyze0ci-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: y82y8ioy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5752.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 5,752 101.1 22.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ziyze0ci",
   "source_url": "https://omlx.ai/benchmarks/performance/ziyze0ci",
   "tg_tokens": null,
   "tps": 101.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6qc66zki-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.1 GB at this context; duplicate traces of the same session: yp13b1tj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3141.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 3,141 85.2 43.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6qc66zki",
   "source_url": "https://omlx.ai/benchmarks/performance/6qc66zki",
   "tg_tokens": null,
   "tps": 85.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6qc66zki-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.9 GB at this context; duplicate traces of the same session: yp13b1tj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3984.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 3,984 80.6 43.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6qc66zki",
   "source_url": "https://omlx.ai/benchmarks/performance/6qc66zki",
   "tg_tokens": null,
   "tps": 80.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-quef5zn8-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 79v3rztz, g417dlnu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 543.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 543.7 49.6 21.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/quef5zn8",
   "source_url": "https://omlx.ai/benchmarks/performance/quef5zn8",
   "tg_tokens": null,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-quef5zn8-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: 79v3rztz, g417dlnu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 359.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 128k 359.1 37.2 23.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/quef5zn8",
   "source_url": "https://omlx.ai/benchmarks/performance/quef5zn8",
   "tg_tokens": null,
   "tps": 37.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-quef5zn8-ctx199680",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 25.7 GB at this context; duplicate traces of the same session: 79v3rztz, g417dlnu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 262.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 195k 262.6 28.2 25.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/quef5zn8",
   "source_url": "https://omlx.ai/benchmarks/performance/quef5zn8",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3g6jh2yr-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 21.8 GB; TTFT 219 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4682.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 4,682 PP tok/s 122.8 TG tok/s 219 TTFT (ms) 21.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3g6jh2yr",
   "source_url": "https://omlx.ai/benchmarks/performance/3g6jh2yr",
   "tg_tokens": null,
   "tps": 122.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nd1nbrm3-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 20.5 GB; TTFT 215 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4754.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 4,754 PP tok/s 119.2 TG tok/s 215 TTFT (ms) 20.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nd1nbrm3",
   "source_url": "https://omlx.ai/benchmarks/performance/nd1nbrm3",
   "tg_tokens": null,
   "tps": 119.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rrispcvz-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 20.1 GB; TTFT 218 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4698.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 4,698 PP tok/s 141.5 TG tok/s 218 TTFT (ms) 20.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rrispcvz",
   "source_url": "https://omlx.ai/benchmarks/performance/rrispcvz",
   "tg_tokens": null,
   "tps": 141.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-v5kjm8fo-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: clryogcn, rj7c5fof; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2852.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 2,852 113.4 27.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v5kjm8fo",
   "source_url": "https://omlx.ai/benchmarks/performance/v5kjm8fo",
   "tg_tokens": null,
   "tps": 113.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-v5kjm8fo-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: clryogcn, rj7c5fof; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4606.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 4,606 112.1 28.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v5kjm8fo",
   "source_url": "https://omlx.ai/benchmarks/performance/v5kjm8fo",
   "tg_tokens": null,
   "tps": 112.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-v5kjm8fo-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: clryogcn, rj7c5fof; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4986.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 8k 4,986 110.2 28.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v5kjm8fo",
   "source_url": "https://omlx.ai/benchmarks/performance/v5kjm8fo",
   "tg_tokens": null,
   "tps": 110.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-p09l1jpl-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 7sgw63a9, qbj1c7l9, swj5w3vo, yhtdor8i; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 98.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 98.7 21.1 18.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p09l1jpl",
   "source_url": "https://omlx.ai/benchmarks/performance/p09l1jpl",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-p09l1jpl-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 7sgw63a9, qbj1c7l9, swj5w3vo, yhtdor8i; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 98.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 98.3 23.2 19.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p09l1jpl",
   "source_url": "https://omlx.ai/benchmarks/performance/p09l1jpl",
   "tg_tokens": null,
   "tps": 23.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-p09l1jpl-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 7sgw63a9, qbj1c7l9, swj5w3vo, yhtdor8i; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 100.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 100.0 29.5 20.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p09l1jpl",
   "source_url": "https://omlx.ai/benchmarks/performance/p09l1jpl",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-p09l1jpl-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 7sgw63a9, qbj1c7l9, swj5w3vo, yhtdor8i; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 100.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 100.9 29.3 20.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p09l1jpl",
   "source_url": "https://omlx.ai/benchmarks/performance/p09l1jpl",
   "tg_tokens": null,
   "tps": 29.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "outlier"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-p09l1jpl-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: 7sgw63a9, qbj1c7l9, swj5w3vo, yhtdor8i; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 94.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 94.6 48.6 22.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p09l1jpl",
   "source_url": "https://omlx.ai/benchmarks/performance/p09l1jpl",
   "tg_tokens": null,
   "tps": 48.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-011ac7gs-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.5 GB; TTFT 950 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1078.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,078 PP tok/s 82.1 TG tok/s 950 TTFT (ms) 70.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/011ac7gs",
   "source_url": "https://omlx.ai/benchmarks/performance/011ac7gs",
   "tg_tokens": null,
   "tps": 82.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9tvs7fmh-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.7 GB; TTFT 950 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1078.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,078 PP tok/s 89.2 TG tok/s 950 TTFT (ms) 70.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9tvs7fmh",
   "source_url": "https://omlx.ai/benchmarks/performance/9tvs7fmh",
   "tg_tokens": null,
   "tps": 89.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ii50jtke-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.5 GB; TTFT 914 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1120.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,120 PP tok/s 67.2 TG tok/s 914 TTFT (ms) 70.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ii50jtke",
   "source_url": "https://omlx.ai/benchmarks/performance/ii50jtke",
   "tg_tokens": null,
   "tps": 67.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fzbcx0to-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.5 GB; TTFT 1130 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 906.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 906.3 PP tok/s 61.6 TG tok/s 1130 TTFT (ms) 70.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fzbcx0to",
   "source_url": "https://omlx.ai/benchmarks/performance/fzbcx0to",
   "tg_tokens": null,
   "tps": 61.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-m947o8n4-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 139.4 GB at this context; duplicate traces of the same session: hn2643km; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 870.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 1k 870.7 61.5 139.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m947o8n4",
   "source_url": "https://omlx.ai/benchmarks/performance/m947o8n4",
   "tg_tokens": null,
   "tps": 61.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-m947o8n4-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 140.0 GB at this context; duplicate traces of the same session: hn2643km; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1085.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 4k 1,085 58.5 140.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m947o8n4",
   "source_url": "https://omlx.ai/benchmarks/performance/m947o8n4",
   "tg_tokens": null,
   "tps": 58.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-u1gku4cp-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 139.4 GB at this context; duplicate traces of the same session: e4svyusl, e5b4mkip; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 904.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 1k 904.3 62.2 139.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u1gku4cp",
   "source_url": "https://omlx.ai/benchmarks/performance/u1gku4cp",
   "tg_tokens": null,
   "tps": 62.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-u1gku4cp-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 140.0 GB at this context; duplicate traces of the same session: e4svyusl, e5b4mkip; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1086.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 4k 1,086 59.3 140.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u1gku4cp",
   "source_url": "https://omlx.ai/benchmarks/performance/u1gku4cp",
   "tg_tokens": null,
   "tps": 59.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-u1gku4cp-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 140.1 GB at this context; duplicate traces of the same session: e4svyusl, e5b4mkip; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1074.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 8k 1,074 59.4 140.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u1gku4cp",
   "source_url": "https://omlx.ai/benchmarks/performance/u1gku4cp",
   "tg_tokens": null,
   "tps": 59.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-z1grri2f-ctx4096",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 10.6 GB; TTFT 2297 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1783.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "gemma-4-12B-it-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k tokens 1,783 PP tok/s 46.0 TG tok/s 2297 TTFT (ms) 10.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z1grri2f",
   "source_url": "https://omlx.ai/benchmarks/performance/z1grri2f",
   "tg_tokens": null,
   "tps": 46.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-iff6pl2k-ctx4096",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 10.6 GB; TTFT 2282 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1795.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "gemma-4-12B-it-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k tokens 1,795 PP tok/s 46.0 TG tok/s 2282 TTFT (ms) 10.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iff6pl2k",
   "source_url": "https://omlx.ai/benchmarks/performance/iff6pl2k",
   "tg_tokens": null,
   "tps": 46.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-65gt8si2-ctx4096",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 10.6 GB; TTFT 2740 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1495.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "gemma-4-12B-it-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k tokens 1,495 PP tok/s 45.9 TG tok/s 2740 TTFT (ms) 10.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/65gt8si2",
   "source_url": "https://omlx.ai/benchmarks/performance/65gt8si2",
   "tg_tokens": null,
   "tps": 45.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wt3zg6zo-ctx4096",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 10.6 GB; TTFT 2292 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1787.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "gemma-4-12B-it-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k tokens 1,787 PP tok/s 46.0 TG tok/s 2292 TTFT (ms) 10.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wt3zg6zo",
   "source_url": "https://omlx.ai/benchmarks/performance/wt3zg6zo",
   "tg_tokens": null,
   "tps": 46.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-6ph7qt1g-ctx4096",
   "model": "gemma-4-31B-it-heretic",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 23.2 GB; TTFT 5282 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 775.5,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "gemma-4-31B-it-heretic-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 775.5 PP tok/s 39.0 TG tok/s 5282 TTFT (ms) 23.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ph7qt1g",
   "source_url": "https://omlx.ai/benchmarks/performance/6ph7qt1g",
   "tg_tokens": null,
   "tps": 39.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zhjyn2pk-ctx4096",
   "model": "gemma-4-31B-it-heretic",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 23.2 GB; TTFT 5897 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 694.7,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "gemma-4-31B-it-heretic-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 694.7 PP tok/s 25.9 TG tok/s 5897 TTFT (ms) 23.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zhjyn2pk",
   "source_url": "https://omlx.ai/benchmarks/performance/zhjyn2pk",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5s15doi5-ctx1024",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 3.3 GB at this context; duplicate traces of the same session: cejxro0c, gv7abmni, r97p8fsj, yduyyx81; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2857.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "MiniCPM5-2B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 2,857 83.8 3.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5s15doi5",
   "source_url": "https://omlx.ai/benchmarks/performance/5s15doi5",
   "tg_tokens": null,
   "tps": 83.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5s15doi5-ctx4096",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 3.3 GB at this context; duplicate traces of the same session: cejxro0c, gv7abmni, r97p8fsj, yduyyx81; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2671.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "MiniCPM5-2B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 2,671 90.6 3.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5s15doi5",
   "source_url": "https://omlx.ai/benchmarks/performance/5s15doi5",
   "tg_tokens": null,
   "tps": 90.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5s15doi5-ctx8192",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 3.5 GB at this context; duplicate traces of the same session: cejxro0c, gv7abmni, r97p8fsj, yduyyx81; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2858.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "MiniCPM5-2B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 8k 2,858 101.6 3.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5s15doi5",
   "source_url": "https://omlx.ai/benchmarks/performance/5s15doi5",
   "tg_tokens": null,
   "tps": 101.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5s15doi5-ctx16384",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 3.8 GB at this context; duplicate traces of the same session: cejxro0c, gv7abmni, r97p8fsj, yduyyx81; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2753.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "MiniCPM5-2B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 16k 2,753 78.7 3.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5s15doi5",
   "source_url": "https://omlx.ai/benchmarks/performance/5s15doi5",
   "tg_tokens": null,
   "tps": 78.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5s15doi5-ctx32768",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 4.6 GB at this context; duplicate traces of the same session: cejxro0c, gv7abmni, r97p8fsj, yduyyx81; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2367.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "MiniCPM5-2B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 32k 2,367 63.8 4.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5s15doi5",
   "source_url": "https://omlx.ai/benchmarks/performance/5s15doi5",
   "tg_tokens": null,
   "tps": 63.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-yrgeeuio-ctx65536",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 5.8 GB; TTFT 20534 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3192.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "MiniCPM5-2B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 64k tokens 3,192 PP tok/s 80.4 TG tok/s 20534 TTFT (ms) 5.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yrgeeuio",
   "source_url": "https://omlx.ai/benchmarks/performance/yrgeeuio",
   "tg_tokens": null,
   "tps": 80.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-g1jeu7kk-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: l23qrjg7; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1519.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 1,519 78.7 21.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g1jeu7kk",
   "source_url": "https://omlx.ai/benchmarks/performance/g1jeu7kk",
   "tg_tokens": null,
   "tps": 78.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-g1jeu7kk-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: l23qrjg7; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2588.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 2,588 94.1 22.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g1jeu7kk",
   "source_url": "https://omlx.ai/benchmarks/performance/g1jeu7kk",
   "tg_tokens": null,
   "tps": 94.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wm03y9cy-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: p05qszh8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 992.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 992.8 47.7 18.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wm03y9cy",
   "source_url": "https://omlx.ai/benchmarks/performance/wm03y9cy",
   "tg_tokens": null,
   "tps": 47.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wm03y9cy-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: p05qszh8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 983.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 983.3 43.5 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wm03y9cy",
   "source_url": "https://omlx.ai/benchmarks/performance/wm03y9cy",
   "tg_tokens": null,
   "tps": 43.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-s5p45slx-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 17.8 GB; TTFT 781 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1311.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k tokens 1,311 PP tok/s 125.9 TG tok/s 781 TTFT (ms) 17.8 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s5p45slx",
   "source_url": "https://omlx.ai/benchmarks/performance/s5p45slx",
   "tg_tokens": null,
   "tps": 125.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-f5kegw8c-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; baseline 20260916; oMLX community benchmark trace; peak memory 18.0 GB at this context; duplicate traces of the same session: dvpsdjak, m6s1twt3, qvqcp64l, r3aqywb2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp-baseline-20260916 on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 103.9 27.1 18.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f5kegw8c",
   "source_url": "https://omlx.ai/benchmarks/performance/f5kegw8c",
   "tg_tokens": null,
   "tps": 27.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-f5kegw8c-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; baseline 20260916; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: dvpsdjak, m6s1twt3, qvqcp64l, r3aqywb2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp-baseline-20260916 on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 105.7 27.3 19.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f5kegw8c",
   "source_url": "https://omlx.ai/benchmarks/performance/f5kegw8c",
   "tg_tokens": null,
   "tps": 27.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-f5kegw8c-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; baseline 20260916; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: dvpsdjak, m6s1twt3, qvqcp64l, r3aqywb2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp-baseline-20260916 on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 105.0 27.5 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f5kegw8c",
   "source_url": "https://omlx.ai/benchmarks/performance/f5kegw8c",
   "tg_tokens": null,
   "tps": 27.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-f5kegw8c-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; baseline 20260916; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: dvpsdjak, m6s1twt3, qvqcp64l, r3aqywb2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp-baseline-20260916 on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 102.1 35.3 20.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f5kegw8c",
   "source_url": "https://omlx.ai/benchmarks/performance/f5kegw8c",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "outlier"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-f5kegw8c-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; baseline 20260916; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: dvpsdjak, m6s1twt3, qvqcp64l, r3aqywb2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 94.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp-baseline-20260916 on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 94.8 50.7 21.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f5kegw8c",
   "source_url": "https://omlx.ai/benchmarks/performance/f5kegw8c",
   "tg_tokens": null,
   "tps": 50.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-u8ui5xn3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: djfk9sdq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 303.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 1k 303.6 38.0 26.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u8ui5xn3",
   "source_url": "https://omlx.ai/benchmarks/performance/u8ui5xn3",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-u8ui5xn3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.0 GB at this context; duplicate traces of the same session: djfk9sdq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 338.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 4k 338.0 47.8 30.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u8ui5xn3",
   "source_url": "https://omlx.ai/benchmarks/performance/u8ui5xn3",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-8hxgzp8m-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 323.5 GB at this context; duplicate traces of the same session: aj1na2qf, bmyavqyk, btotr0az, l7je8gpr, pdy842yf, yme01a5r, z1x1unjf; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 425.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 425.2 30.1 323.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8hxgzp8m",
   "source_url": "https://omlx.ai/benchmarks/performance/8hxgzp8m",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-8hxgzp8m-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 324.7 GB at this context; duplicate traces of the same session: aj1na2qf, bmyavqyk, btotr0az, l7je8gpr, pdy842yf, yme01a5r, z1x1unjf; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 550.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 550.5 30.3 324.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8hxgzp8m",
   "source_url": "https://omlx.ai/benchmarks/performance/8hxgzp8m",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-8hxgzp8m-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 325.0 GB at this context; duplicate traces of the same session: aj1na2qf, bmyavqyk, btotr0az, l7je8gpr, pdy842yf, yme01a5r, z1x1unjf; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 506.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 506.1 36.9 325.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8hxgzp8m",
   "source_url": "https://omlx.ai/benchmarks/performance/8hxgzp8m",
   "tg_tokens": null,
   "tps": 36.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-8hxgzp8m-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 325.2 GB at this context; duplicate traces of the same session: aj1na2qf, bmyavqyk, btotr0az, l7je8gpr, pdy842yf, yme01a5r, z1x1unjf; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 535.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k 535.1 64.9 325.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8hxgzp8m",
   "source_url": "https://omlx.ai/benchmarks/performance/8hxgzp8m",
   "tg_tokens": null,
   "tps": 64.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-8hxgzp8m-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 325.4 GB at this context; duplicate traces of the same session: aj1na2qf, bmyavqyk, btotr0az, l7je8gpr, pdy842yf, yme01a5r, z1x1unjf; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 527.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 32k 527.5 44.3 325.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8hxgzp8m",
   "source_url": "https://omlx.ai/benchmarks/performance/8hxgzp8m",
   "tg_tokens": null,
   "tps": 44.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-8hxgzp8m-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 326.1 GB at this context; duplicate traces of the same session: aj1na2qf, bmyavqyk, btotr0az, l7je8gpr, pdy842yf, yme01a5r, z1x1unjf; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 520.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 64k 520.0 39.6 326.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8hxgzp8m",
   "source_url": "https://omlx.ai/benchmarks/performance/8hxgzp8m",
   "tg_tokens": null,
   "tps": 39.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-8hxgzp8m-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 327.4 GB at this context; duplicate traces of the same session: aj1na2qf, bmyavqyk, btotr0az, l7je8gpr, pdy842yf, yme01a5r, z1x1unjf; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 508.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 128k 508.5 58.9 327.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8hxgzp8m",
   "source_url": "https://omlx.ai/benchmarks/performance/8hxgzp8m",
   "tg_tokens": null,
   "tps": 58.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-8hxgzp8m-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 328.7 GB at this context; duplicate traces of the same session: aj1na2qf, bmyavqyk, btotr0az, l7je8gpr, pdy842yf, yme01a5r, z1x1unjf; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 498.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 195k 498.3 36.1 328.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8hxgzp8m",
   "source_url": "https://omlx.ai/benchmarks/performance/8hxgzp8m",
   "tg_tokens": null,
   "tps": 36.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-1xgoeoq3-ctx1024",
   "model": "LFM2.5-8B-A1B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 5.3 GB; TTFT 161 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6365.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-8B-A1B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k tokens 6,365 PP tok/s 179.7 TG tok/s 161 TTFT (ms) 5.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1xgoeoq3",
   "source_url": "https://omlx.ai/benchmarks/performance/1xgoeoq3",
   "tg_tokens": null,
   "tps": 179.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-yrs6ke1v-ctx1024",
   "model": "LFM2.5-8B-A1B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 5.3 GB; TTFT 162 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6307.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-8B-A1B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k tokens 6,307 PP tok/s 179.3 TG tok/s 162 TTFT (ms) 5.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yrs6ke1v",
   "source_url": "https://omlx.ai/benchmarks/performance/yrs6ke1v",
   "tg_tokens": null,
   "tps": 179.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-7moej305-ctx1024",
   "model": "LFM2.5-8B-A1B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 5.3 GB; TTFT 180 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5676.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-8B-A1B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k tokens 5,676 PP tok/s 165.7 TG tok/s 180 TTFT (ms) 5.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7moej305",
   "source_url": "https://omlx.ai/benchmarks/performance/7moej305",
   "tg_tokens": null,
   "tps": 165.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-m9yqhp7q-ctx1024",
   "model": "LFM2.5-8B-A1B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 5.3 GB; TTFT 165 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6212.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-8B-A1B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k tokens 6,212 PP tok/s 173.1 TG tok/s 165 TTFT (ms) 5.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m9yqhp7q",
   "source_url": "https://omlx.ai/benchmarks/performance/m9yqhp7q",
   "tg_tokens": null,
   "tps": 173.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-bwpvmb2n-ctx1024",
   "model": "NVIDIA-Nemotron-3-Nano-30B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 18.9 GB; TTFT 525 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1950.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3-Nano-30B-A3B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k tokens 1,950 PP tok/s 102.6 TG tok/s 525 TTFT (ms) 18.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bwpvmb2n",
   "source_url": "https://omlx.ai/benchmarks/performance/bwpvmb2n",
   "tg_tokens": null,
   "tps": 102.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-iku234xj-ctx1024",
   "model": "NVIDIA-Nemotron-3-Nano-30B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 18.9 GB; TTFT 539 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1901.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3-Nano-30B-A3B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k tokens 1,901 PP tok/s 98.6 TG tok/s 539 TTFT (ms) 18.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iku234xj",
   "source_url": "https://omlx.ai/benchmarks/performance/iku234xj",
   "tg_tokens": null,
   "tps": 98.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-cbwstgt9-ctx1024",
   "model": "NVIDIA-Nemotron-3-Nano-30B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 18.9 GB; TTFT 547 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1872.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3-Nano-30B-A3B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k tokens 1,872 PP tok/s 100.1 TG tok/s 547 TTFT (ms) 18.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cbwstgt9",
   "source_url": "https://omlx.ai/benchmarks/performance/cbwstgt9",
   "tg_tokens": null,
   "tps": 100.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-eiwzeoc6-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 10.7 GB; TTFT 914 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1120.0,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Ornith-1.5-9B-MXFP8 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 unknown; 1k tokens 1,120 PP tok/s 31.0 TG tok/s 914 TTFT (ms) 10.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eiwzeoc6",
   "source_url": "https://omlx.ai/benchmarks/performance/eiwzeoc6",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-xe4ytgpp-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 10.7 GB; TTFT 947 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1081.0,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Ornith-1.5-9B-MXFP8 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 unknown; 1k tokens 1,081 PP tok/s 29.6 TG tok/s 947 TTFT (ms) 10.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xe4ytgpp",
   "source_url": "https://omlx.ai/benchmarks/performance/xe4ytgpp",
   "tg_tokens": null,
   "tps": 29.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ncp9t4sc-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 10.7 GB; TTFT 949 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1079.0,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Ornith-1.5-9B-MXFP8 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 unknown; 1k tokens 1,079 PP tok/s 29.8 TG tok/s 949 TTFT (ms) 10.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ncp9t4sc",
   "source_url": "https://omlx.ai/benchmarks/performance/ncp9t4sc",
   "tg_tokens": null,
   "tps": 29.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-vek8p5ak-ctx1024",
   "model": "Qwen3.6-35B-A3B-Holo3-Qwopus-AgentWorld",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 29.3 GB; TTFT 517 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1980.0,
   "provenance": "community",
   "quant": "qx64 hi",
   "quote": "Qwen3.6-35B-A3B-Holo3-Qwopus-AgentWorld-qx64-hi-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k tokens 1,980 PP tok/s 78.4 TG tok/s 517 TTFT (ms) 29.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vek8p5ak",
   "source_url": "https://omlx.ai/benchmarks/performance/vek8p5ak",
   "tg_tokens": null,
   "tps": 78.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-xqy5078c-ctx1024",
   "model": "Qwen3.8-27B-Brainwaves-2M",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 20.6 GB; TTFT 2406 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 425.5,
   "provenance": "community",
   "quant": "qx64 hi",
   "quote": "Qwen3.8-27B-Brainwaves-2M-qx64-hi-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k tokens 425.5 PP tok/s 13.1 TG tok/s 2406 TTFT (ms) 20.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xqy5078c",
   "source_url": "https://omlx.ai/benchmarks/performance/xqy5078c",
   "tg_tokens": null,
   "tps": 13.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-spjh3of7-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: hsgenq3l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 249.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 249.7 38.8 32.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/spjh3of7",
   "source_url": "https://omlx.ai/benchmarks/performance/spjh3of7",
   "tg_tokens": null,
   "tps": 38.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-spjh3of7-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.7 GB at this context; duplicate traces of the same session: hsgenq3l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 257.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 257.1 30.5 33.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/spjh3of7",
   "source_url": "https://omlx.ai/benchmarks/performance/spjh3of7",
   "tg_tokens": null,
   "tps": 30.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-s5nhv70t-ctx1024",
   "model": "Ternary-Bonsai-27B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 1k context; peak memory 9.1 GB; TTFT 5638 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 181.6,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-27B-mlx-2bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 2bit; 1k tokens 181.6 PP tok/s 31.3 TG tok/s 5638 TTFT (ms) 9.1 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s5nhv70t",
   "source_url": "https://omlx.ai/benchmarks/performance/s5nhv70t",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-111umjmn-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 11.8 GB; TTFT 427 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2397.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k tokens 2,397 PP tok/s 81.6 TG tok/s 427 TTFT (ms) 11.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/111umjmn",
   "source_url": "https://omlx.ai/benchmarks/performance/111umjmn",
   "tg_tokens": null,
   "tps": 81.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-dshomzpl-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 11.8 GB; TTFT 431 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2377.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k tokens 2,377 PP tok/s 79.4 TG tok/s 431 TTFT (ms) 11.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dshomzpl",
   "source_url": "https://omlx.ai/benchmarks/performance/dshomzpl",
   "tg_tokens": null,
   "tps": 79.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-5b5er1p4-ctx1024",
   "model": "DeepSeek-R1-0528-Qwen3-8B",
   "notes": "oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: tj311gac; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1125.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-R1-0528-Qwen3-8B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 16 GB \u00b7 4bit; 1k 1,125 30.7 4.9 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5b5er1p4",
   "source_url": "https://omlx.ai/benchmarks/performance/5b5er1p4",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-5b5er1p4-ctx4096",
   "model": "DeepSeek-R1-0528-Qwen3-8B",
   "notes": "oMLX community benchmark trace; peak memory 5.4 GB at this context; duplicate traces of the same session: tj311gac; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1108.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-R1-0528-Qwen3-8B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 16 GB \u00b7 4bit; 4k 1,108 27.9 5.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5b5er1p4",
   "source_url": "https://omlx.ai/benchmarks/performance/5b5er1p4",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-tyda03r4-ctx1024",
   "model": "DeepSeek-R1-0528-Qwen3-8B",
   "notes": "oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: q8sedx96; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1118.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-R1-0528-Qwen3-8B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 16 GB \u00b7 4bit; 1k 1,118 30.4 4.9 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tyda03r4",
   "source_url": "https://omlx.ai/benchmarks/performance/tyda03r4",
   "tg_tokens": null,
   "tps": 30.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-tyda03r4-ctx4096",
   "model": "DeepSeek-R1-0528-Qwen3-8B",
   "notes": "oMLX community benchmark trace; peak memory 5.4 GB at this context; duplicate traces of the same session: q8sedx96; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1074.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-R1-0528-Qwen3-8B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 16 GB \u00b7 4bit; 4k 1,074 27.2 5.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tyda03r4",
   "source_url": "https://omlx.ai/benchmarks/performance/tyda03r4",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-sp3yj1ee-ctx1024",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 89.2 GB at this context; duplicate traces of the same session: xaom9heq, zexoivty; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 398.0,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4-Flash-0731-oQ2e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 1k 398.0 34.4 89.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sp3yj1ee",
   "source_url": "https://omlx.ai/benchmarks/performance/sp3yj1ee",
   "tg_tokens": null,
   "tps": 34.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-sp3yj1ee-ctx4096",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 89.7 GB at this context; duplicate traces of the same session: xaom9heq, zexoivty; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 705.2,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4-Flash-0731-oQ2e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 4k 705.2 31.3 89.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sp3yj1ee",
   "source_url": "https://omlx.ai/benchmarks/performance/sp3yj1ee",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-sp3yj1ee-ctx16384",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 89.8 GB at this context; duplicate traces of the same session: xaom9heq, zexoivty; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 781.9,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4-Flash-0731-oQ2e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 16k 781.9 31.5 89.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sp3yj1ee",
   "source_url": "https://omlx.ai/benchmarks/performance/sp3yj1ee",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bx7x0tfe-ctx1024",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.4 GB at this context; duplicate traces of the same session: c6314845, gmlom3sb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 756.8,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4-Flash-0731-oQ2e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 1k 756.8 55.2 99.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bx7x0tfe",
   "source_url": "https://omlx.ai/benchmarks/performance/bx7x0tfe",
   "tg_tokens": null,
   "tps": 55.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bx7x0tfe-ctx4096",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.8 GB at this context; duplicate traces of the same session: c6314845, gmlom3sb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 822.5,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4-Flash-0731-oQ2e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 4k 822.5 52.1 99.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bx7x0tfe",
   "source_url": "https://omlx.ai/benchmarks/performance/bx7x0tfe",
   "tg_tokens": null,
   "tps": 52.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bx7x0tfe-ctx16384",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.9 GB at this context; duplicate traces of the same session: c6314845, gmlom3sb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 758.6,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4-Flash-0731-oQ2e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 16k 758.6 55.3 99.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bx7x0tfe",
   "source_url": "https://omlx.ai/benchmarks/performance/bx7x0tfe",
   "tg_tokens": null,
   "tps": 55.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-lxzbokwu-ctx4096",
   "model": "Qwen2.5-Coder-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 5.0 GB at this context; duplicate traces of the same session: tc8tlxro; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 811.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-7B-Instruct-4bit on M5 (10c); M5 (10c) \u00b7 16 GB \u00b7 4bit; 4k 811.7 28.1 5.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lxzbokwu",
   "source_url": "https://omlx.ai/benchmarks/performance/lxzbokwu",
   "tg_tokens": null,
   "tps": 28.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-lxzbokwu-ctx16384",
   "model": "Qwen2.5-Coder-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 5.5 GB at this context; duplicate traces of the same session: tc8tlxro; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 723.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-7B-Instruct-4bit on M5 (10c); M5 (10c) \u00b7 16 GB \u00b7 4bit; 16k 723.8 23.9 5.5 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lxzbokwu",
   "source_url": "https://omlx.ai/benchmarks/performance/lxzbokwu",
   "tg_tokens": null,
   "tps": 23.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-58z7oj8h-ctx1024",
   "model": "Qwen3.6-40B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: ebg73p20; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 70.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-40B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-Thinking-mlx-4Bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 70.7 10.0 23.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/58z7oj8h",
   "source_url": "https://omlx.ai/benchmarks/performance/58z7oj8h",
   "tg_tokens": null,
   "tps": 10.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-58z7oj8h-ctx4096",
   "model": "Qwen3.6-40B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: ebg73p20; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 73.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-40B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-Thinking-mlx-4Bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 73.6 9.6 27.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/58z7oj8h",
   "source_url": "https://omlx.ai/benchmarks/performance/58z7oj8h",
   "tg_tokens": null,
   "tps": 9.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-rsw02bdb-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: ys612ktl; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 152.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 152.2 14.8 19.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rsw02bdb",
   "source_url": "https://omlx.ai/benchmarks/performance/rsw02bdb",
   "tg_tokens": null,
   "tps": 14.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-rsw02bdb-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: ys612ktl; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 128.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 128.9 14.0 18.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rsw02bdb",
   "source_url": "https://omlx.ai/benchmarks/performance/rsw02bdb",
   "tg_tokens": null,
   "tps": 14.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-hi90h7lg-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: xe7wnmd7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 487.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 487.0 30.9 19.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hi90h7lg",
   "source_url": "https://omlx.ai/benchmarks/performance/hi90h7lg",
   "tg_tokens": null,
   "tps": 30.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-hi90h7lg-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: xe7wnmd7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 503.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 503.5 31.5 21.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hi90h7lg",
   "source_url": "https://omlx.ai/benchmarks/performance/hi90h7lg",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-vrwz9udz-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: 2tipsmw7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 485.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 485.7 24.9 19.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vrwz9udz",
   "source_url": "https://omlx.ai/benchmarks/performance/vrwz9udz",
   "tg_tokens": null,
   "tps": 24.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-vrwz9udz-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 2tipsmw7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 505.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 505.0 31.0 21.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vrwz9udz",
   "source_url": "https://omlx.ai/benchmarks/performance/vrwz9udz",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-egobd30l-ctx32768",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 37.4 GB; TTFT 82545 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 397.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k tokens 397.0 PP tok/s 13.4 TG tok/s 82545 TTFT (ms) 37.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/egobd30l",
   "source_url": "https://omlx.ai/benchmarks/performance/egobd30l",
   "tg_tokens": null,
   "tps": 13.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ilhdp91e-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: tdpbldth; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 479.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 479.7 35.2 17.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ilhdp91e",
   "source_url": "https://omlx.ai/benchmarks/performance/ilhdp91e",
   "tg_tokens": null,
   "tps": 35.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ilhdp91e-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: tdpbldth; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 496.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 496.5 34.3 19.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ilhdp91e",
   "source_url": "https://omlx.ai/benchmarks/performance/ilhdp91e",
   "tg_tokens": null,
   "tps": 34.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ln99kph6-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: caqv5ayf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 465.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 465.5 34.9 17.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ln99kph6",
   "source_url": "https://omlx.ai/benchmarks/performance/ln99kph6",
   "tg_tokens": null,
   "tps": 34.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ln99kph6-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: caqv5ayf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 498.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 498.7 32.2 19.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ln99kph6",
   "source_url": "https://omlx.ai/benchmarks/performance/ln99kph6",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-n9lkh3fz-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 7y6jzct7, gg7q1e9p, i1fdrbh8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 609.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 609.3 81.0 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n9lkh3fz",
   "source_url": "https://omlx.ai/benchmarks/performance/n9lkh3fz",
   "tg_tokens": null,
   "tps": 81.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-n9lkh3fz-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: 7y6jzct7, gg7q1e9p, i1fdrbh8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 825.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 825.0 63.8 71.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n9lkh3fz",
   "source_url": "https://omlx.ai/benchmarks/performance/n9lkh3fz",
   "tg_tokens": null,
   "tps": 63.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-n9lkh3fz-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.5 GB at this context; duplicate traces of the same session: 7y6jzct7, gg7q1e9p, i1fdrbh8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 819.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 819.9 54.7 72.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n9lkh3fz",
   "source_url": "https://omlx.ai/benchmarks/performance/n9lkh3fz",
   "tg_tokens": null,
   "tps": 54.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-n9lkh3fz-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 75.1 GB at this context; duplicate traces of the same session: 7y6jzct7, gg7q1e9p, i1fdrbh8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 810.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 810.8 58.2 75.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n9lkh3fz",
   "source_url": "https://omlx.ai/benchmarks/performance/n9lkh3fz",
   "tg_tokens": null,
   "tps": 58.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-8gdas0yr-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: 0cfo2o0t, huyr4q3z, zq8ec4tk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 581.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 581.4 61.2 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8gdas0yr",
   "source_url": "https://omlx.ai/benchmarks/performance/8gdas0yr",
   "tg_tokens": null,
   "tps": 61.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-8gdas0yr-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: 0cfo2o0t, huyr4q3z, zq8ec4tk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 824.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 824.7 42.3 71.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8gdas0yr",
   "source_url": "https://omlx.ai/benchmarks/performance/8gdas0yr",
   "tg_tokens": null,
   "tps": 42.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-8gdas0yr-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.9 GB at this context; duplicate traces of the same session: 0cfo2o0t, huyr4q3z, zq8ec4tk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 819.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 819.5 66.1 72.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8gdas0yr",
   "source_url": "https://omlx.ai/benchmarks/performance/8gdas0yr",
   "tg_tokens": null,
   "tps": 66.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-8gdas0yr-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.8 GB at this context; duplicate traces of the same session: 0cfo2o0t, huyr4q3z, zq8ec4tk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 811.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 811.2 60.0 74.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8gdas0yr",
   "source_url": "https://omlx.ai/benchmarks/performance/8gdas0yr",
   "tg_tokens": null,
   "tps": 60.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-7avg4lgf-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 100.9 GB at this context; duplicate traces of the same session: 4xozy6hy, dade03n1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 742.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 1k 742.4 56.1 100.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7avg4lgf",
   "source_url": "https://omlx.ai/benchmarks/performance/7avg4lgf",
   "tg_tokens": null,
   "tps": 56.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-7avg4lgf-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 109.5 GB at this context; duplicate traces of the same session: 4xozy6hy, dade03n1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1293.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 128k 1,293 63.0 109.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7avg4lgf",
   "source_url": "https://omlx.ai/benchmarks/performance/7avg4lgf",
   "tg_tokens": null,
   "tps": 63.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-7avg4lgf-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 110.4 GB at this context; duplicate traces of the same session: 4xozy6hy, dade03n1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1211.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 195k 1,211 47.3 110.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7avg4lgf",
   "source_url": "https://omlx.ai/benchmarks/performance/7avg4lgf",
   "tg_tokens": null,
   "tps": 47.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-otob9lxr-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 100.9 GB at this context; duplicate traces of the same session: 4h8kffj0, jloksdpj, lpaog71e, zxdi5stl; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 751.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 1k 751.5 81.8 100.9 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/otob9lxr",
   "source_url": "https://omlx.ai/benchmarks/performance/otob9lxr",
   "tg_tokens": null,
   "tps": 81.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-otob9lxr-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 101.4 GB at this context; duplicate traces of the same session: 4h8kffj0, jloksdpj, lpaog71e, zxdi5stl; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 978.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 4k 978.2 76.4 101.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/otob9lxr",
   "source_url": "https://omlx.ai/benchmarks/performance/otob9lxr",
   "tg_tokens": null,
   "tps": 76.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-otob9lxr-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 102.5 GB at this context; duplicate traces of the same session: 4h8kffj0, jloksdpj, lpaog71e, zxdi5stl; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1336.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 32k 1,336 68.4 102.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/otob9lxr",
   "source_url": "https://omlx.ai/benchmarks/performance/otob9lxr",
   "tg_tokens": null,
   "tps": 68.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-otob9lxr-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 108.5 GB at this context; duplicate traces of the same session: 4h8kffj0, jloksdpj, lpaog71e, zxdi5stl; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1297.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 128k 1,297 56.7 108.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/otob9lxr",
   "source_url": "https://omlx.ai/benchmarks/performance/otob9lxr",
   "tg_tokens": null,
   "tps": 56.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-otob9lxr-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 110.1 GB at this context; duplicate traces of the same session: 4h8kffj0, jloksdpj, lpaog71e, zxdi5stl; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1278.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 195k 1,278 45.8 110.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/otob9lxr",
   "source_url": "https://omlx.ai/benchmarks/performance/otob9lxr",
   "tg_tokens": null,
   "tps": 45.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-31k8bl3c-ctx1024",
   "model": "DeepSeek-R1-0528-Qwen3-8B",
   "notes": "oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: vm2axa6h; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1124.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-R1-0528-Qwen3-8B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 16 GB \u00b7 4bit; 1k 1,124 30.7 4.9 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/31k8bl3c",
   "source_url": "https://omlx.ai/benchmarks/performance/31k8bl3c",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-31k8bl3c-ctx4096",
   "model": "DeepSeek-R1-0528-Qwen3-8B",
   "notes": "oMLX community benchmark trace; peak memory 5.4 GB at this context; duplicate traces of the same session: vm2axa6h; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1112.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-R1-0528-Qwen3-8B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 16 GB \u00b7 4bit; 4k 1,112 27.9 5.4 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/31k8bl3c",
   "source_url": "https://omlx.ai/benchmarks/performance/31k8bl3c",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fnrb1bkz-ctx16384",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 29.4 GB; TTFT 82659 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 198.2,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4-Flash-0731-oQ2e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 16k tokens 198.2 PP tok/s 10.5 TG tok/s 82659 TTFT (ms) 29.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fnrb1bkz",
   "source_url": "https://omlx.ai/benchmarks/performance/fnrb1bkz",
   "tg_tokens": null,
   "tps": 10.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jeyn7e69-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "EigenLabs; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 938.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "EigenLabs--Qwen3.8-27B-4bit-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 938.4 43.7 16.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jeyn7e69",
   "source_url": "https://omlx.ai/benchmarks/performance/jeyn7e69",
   "tg_tokens": null,
   "tps": 43.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jeyn7e69-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "EigenLabs; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1309.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "EigenLabs--Qwen3.8-27B-4bit-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,309 64.8 18.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jeyn7e69",
   "source_url": "https://omlx.ai/benchmarks/performance/jeyn7e69",
   "tg_tokens": null,
   "tps": 64.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jeyn7e69-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "EigenLabs; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1374.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "EigenLabs--Qwen3.8-27B-4bit-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,374 69.0 19.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jeyn7e69",
   "source_url": "https://omlx.ai/benchmarks/performance/jeyn7e69",
   "tg_tokens": null,
   "tps": 69.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-279dvhqy-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: brzqszew; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1059.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 16 GB \u00b7 4bit; 1k 1,059 29.9 5.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/279dvhqy",
   "source_url": "https://omlx.ai/benchmarks/performance/279dvhqy",
   "tg_tokens": null,
   "tps": 29.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-279dvhqy-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: brzqszew; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1119.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 16 GB \u00b7 4bit; 4k 1,119 29.3 6.2 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/279dvhqy",
   "source_url": "https://omlx.ai/benchmarks/performance/279dvhqy",
   "tg_tokens": null,
   "tps": 29.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-z1qmujro-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: gx4mm0f2; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1054.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 16 GB \u00b7 4bit; 1k 1,054 29.7 5.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z1qmujro",
   "source_url": "https://omlx.ai/benchmarks/performance/z1qmujro",
   "tg_tokens": null,
   "tps": 29.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-z1qmujro-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: gx4mm0f2; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1117.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 16 GB \u00b7 4bit; 4k 1,117 29.2 6.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z1qmujro",
   "source_url": "https://omlx.ai/benchmarks/performance/z1qmujro",
   "tg_tokens": null,
   "tps": 29.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-7goffx36-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "DWQ build; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 973.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit-DWQ on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 8bit; 1k 973.2 98.4 20.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7goffx36",
   "source_url": "https://omlx.ai/benchmarks/performance/7goffx36",
   "tg_tokens": null,
   "tps": 98.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-7goffx36-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "DWQ build; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1467.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit-DWQ on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 8bit; 8k 1,467 89.8 22.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7goffx36",
   "source_url": "https://omlx.ai/benchmarks/performance/7goffx36",
   "tg_tokens": null,
   "tps": 89.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-3ikwb7vh-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: da34fhnm; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 67.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 67.5 14.8 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ikwb7vh",
   "source_url": "https://omlx.ai/benchmarks/performance/3ikwb7vh",
   "tg_tokens": null,
   "tps": 14.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-3ikwb7vh-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: da34fhnm; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 75.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 75.3 13.7 20.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ikwb7vh",
   "source_url": "https://omlx.ai/benchmarks/performance/3ikwb7vh",
   "tg_tokens": null,
   "tps": 13.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-r24cpj61-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: z5mxwkwj; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 88.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 88.9 16.0 19.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r24cpj61",
   "source_url": "https://omlx.ai/benchmarks/performance/r24cpj61",
   "tg_tokens": null,
   "tps": 16.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-r24cpj61-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: z5mxwkwj; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 85.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 85.8 15.6 23.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r24cpj61",
   "source_url": "https://omlx.ai/benchmarks/performance/r24cpj61",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-x06oxd2g-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 58mgza4h, w6x76x67, wz2gd2hu, zer1qtkj; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 243.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k 243.6 15.5 17.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x06oxd2g",
   "source_url": "https://omlx.ai/benchmarks/performance/x06oxd2g",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-x06oxd2g-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 58mgza4h, w6x76x67, wz2gd2hu, zer1qtkj; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 291.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 291.0 14.8 19.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x06oxd2g",
   "source_url": "https://omlx.ai/benchmarks/performance/x06oxd2g",
   "tg_tokens": null,
   "tps": 14.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-x06oxd2g-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 58mgza4h, w6x76x67, wz2gd2hu, zer1qtkj; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 285.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 8k 285.4 14.6 19.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x06oxd2g",
   "source_url": "https://omlx.ai/benchmarks/performance/x06oxd2g",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-x06oxd2g-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 58mgza4h, w6x76x67, wz2gd2hu, zer1qtkj; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 260.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 16k 260.4 12.6 19.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x06oxd2g",
   "source_url": "https://omlx.ai/benchmarks/performance/x06oxd2g",
   "tg_tokens": null,
   "tps": 12.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-x06oxd2g-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: 58mgza4h, w6x76x67, wz2gd2hu, zer1qtkj; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 184.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 32k 184.9 6.9 21.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x06oxd2g",
   "source_url": "https://omlx.ai/benchmarks/performance/x06oxd2g",
   "tg_tokens": null,
   "tps": 6.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kyv8a6qn-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 86.5 GB at this context; duplicate traces of the same session: o504dley; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 807.5,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 807.5 36.4 86.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kyv8a6qn",
   "source_url": "https://omlx.ai/benchmarks/performance/kyv8a6qn",
   "tg_tokens": null,
   "tps": 36.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kyv8a6qn-ctx131072",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 93.9 GB at this context; duplicate traces of the same session: o504dley; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2033.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 2,033 53.0 93.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kyv8a6qn",
   "source_url": "https://omlx.ai/benchmarks/performance/kyv8a6qn",
   "tg_tokens": null,
   "tps": 53.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-k0ivmh2k-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: 20dl2z2r, bmiat3tr, pyy51rqp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 609.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 609.8 69.3 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k0ivmh2k",
   "source_url": "https://omlx.ai/benchmarks/performance/k0ivmh2k",
   "tg_tokens": null,
   "tps": 69.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-k0ivmh2k-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: 20dl2z2r, bmiat3tr, pyy51rqp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 718.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 718.6 75.5 71.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k0ivmh2k",
   "source_url": "https://omlx.ai/benchmarks/performance/k0ivmh2k",
   "tg_tokens": null,
   "tps": 75.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-k0ivmh2k-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: 20dl2z2r, bmiat3tr, pyy51rqp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 832.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 832.0 67.4 71.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k0ivmh2k",
   "source_url": "https://omlx.ai/benchmarks/performance/k0ivmh2k",
   "tg_tokens": null,
   "tps": 67.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-k0ivmh2k-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.4 GB at this context; duplicate traces of the same session: 20dl2z2r, bmiat3tr, pyy51rqp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 810.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 810.3 66.4 74.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k0ivmh2k",
   "source_url": "https://omlx.ai/benchmarks/performance/k0ivmh2k",
   "tg_tokens": null,
   "tps": 66.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-tcbe0x5t-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 854r5lsj, jr0uooic; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 610.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 610.7 80.7 70.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tcbe0x5t",
   "source_url": "https://omlx.ai/benchmarks/performance/tcbe0x5t",
   "tg_tokens": null,
   "tps": 80.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-tcbe0x5t-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.4 GB at this context; duplicate traces of the same session: 854r5lsj, jr0uooic; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 819.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 819.9 59.2 72.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tcbe0x5t",
   "source_url": "https://omlx.ai/benchmarks/performance/tcbe0x5t",
   "tg_tokens": null,
   "tps": 59.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-tcbe0x5t-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.8 GB at this context; duplicate traces of the same session: 854r5lsj, jr0uooic; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 811.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 811.2 65.9 74.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tcbe0x5t",
   "source_url": "https://omlx.ai/benchmarks/performance/tcbe0x5t",
   "tg_tokens": null,
   "tps": 65.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-71cga9nj-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 72.7 GB; TTFT 7257 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2258.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k tokens 2,258 PP tok/s 51.2 TG tok/s 7257 TTFT (ms) 72.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/71cga9nj",
   "source_url": "https://omlx.ai/benchmarks/performance/71cga9nj",
   "tg_tokens": null,
   "tps": 51.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-uok1cwr7-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: iszqxyd5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1168.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,168 57.5 70.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uok1cwr7",
   "source_url": "https://omlx.ai/benchmarks/performance/uok1cwr7",
   "tg_tokens": null,
   "tps": 57.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-uok1cwr7-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.9 GB at this context; duplicate traces of the same session: iszqxyd5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2302.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,302 45.7 78.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uok1cwr7",
   "source_url": "https://omlx.ai/benchmarks/performance/uok1cwr7",
   "tg_tokens": null,
   "tps": 45.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vm3tpavx-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.5 GB; TTFT 874 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1172.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,172 PP tok/s 73.1 TG tok/s 874 TTFT (ms) 70.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vm3tpavx",
   "source_url": "https://omlx.ai/benchmarks/performance/vm3tpavx",
   "tg_tokens": null,
   "tps": 73.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-902tb5uv-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 74.7 GB; TTFT 26389 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2484.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k tokens 2,484 PP tok/s 74.0 TG tok/s 26389 TTFT (ms) 74.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/902tb5uv",
   "source_url": "https://omlx.ai/benchmarks/performance/902tb5uv",
   "tg_tokens": null,
   "tps": 74.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-11y774pj-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 74.7 GB; TTFT 26083 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2513.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k tokens 2,513 PP tok/s 73.5 TG tok/s 26083 TTFT (ms) 74.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/11y774pj",
   "source_url": "https://omlx.ai/benchmarks/performance/11y774pj",
   "tg_tokens": null,
   "tps": 73.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5ibgcrx9-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.5 GB; TTFT 863 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1187.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,187 PP tok/s 94.3 TG tok/s 863 TTFT (ms) 70.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ibgcrx9",
   "source_url": "https://omlx.ai/benchmarks/performance/5ibgcrx9",
   "tg_tokens": null,
   "tps": 94.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-w7z76jh2-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: i5039pcd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1090.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,090 60.2 70.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w7z76jh2",
   "source_url": "https://omlx.ai/benchmarks/performance/w7z76jh2",
   "tg_tokens": null,
   "tps": 60.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-w7z76jh2-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.4 GB at this context; duplicate traces of the same session: i5039pcd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2093.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,093 46.6 78.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w7z76jh2",
   "source_url": "https://omlx.ai/benchmarks/performance/w7z76jh2",
   "tg_tokens": null,
   "tps": 46.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ddue8qev-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 101.6 GB at this context; duplicate traces of the same session: rbd3a6m1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1068.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 4k 1,068 80.1 101.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ddue8qev",
   "source_url": "https://omlx.ai/benchmarks/performance/ddue8qev",
   "tg_tokens": null,
   "tps": 80.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ddue8qev-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 104.4 GB at this context; duplicate traces of the same session: rbd3a6m1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1317.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 64k 1,317 76.5 104.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ddue8qev",
   "source_url": "https://omlx.ai/benchmarks/performance/ddue8qev",
   "tg_tokens": null,
   "tps": 76.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ddue8qev-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 108.6 GB at this context; duplicate traces of the same session: rbd3a6m1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1288.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 128k 1,288 57.1 108.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ddue8qev",
   "source_url": "https://omlx.ai/benchmarks/performance/ddue8qev",
   "tg_tokens": null,
   "tps": 57.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ddue8qev-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 110.2 GB at this context; duplicate traces of the same session: rbd3a6m1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1181.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 195k 1,181 46.5 110.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ddue8qev",
   "source_url": "https://omlx.ai/benchmarks/performance/ddue8qev",
   "tg_tokens": null,
   "tps": 46.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-q2wl1frk-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 127.3 GB; TTFT 9475 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 864.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k tokens 864.6 PP tok/s 49.6 TG tok/s 9475 TTFT (ms) 127.3 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q2wl1frk",
   "source_url": "https://omlx.ai/benchmarks/performance/q2wl1frk",
   "tg_tokens": null,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ggxjjuwy-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 126.5 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 578.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 578.8 44.9 126.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ggxjjuwy",
   "source_url": "https://omlx.ai/benchmarks/performance/ggxjjuwy",
   "tg_tokens": null,
   "tps": 44.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ggxjjuwy-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 127.1 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 767.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 767.2 48.3 127.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ggxjjuwy",
   "source_url": "https://omlx.ai/benchmarks/performance/ggxjjuwy",
   "tg_tokens": null,
   "tps": 48.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-jh6gexke-ctx1024",
   "model": "MiniCPM5-2B",
   "notes": "mlx-community; oMLX community benchmark trace; peak memory 2.0 GB at this context; duplicate traces of the same session: 45bhplpc, 5zawwaq5, g6lafzj5, nv4z5hnh; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 313.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--MiniCPM5-2B-mlx-4Bit on M1 (8c); M1 (8c) \u00b7 8 GB \u00b7 4bit; 1k 313.5 33.6 2.0 GB view",
   "ram_gb": 8,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jh6gexke",
   "source_url": "https://omlx.ai/benchmarks/performance/jh6gexke",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-jh6gexke-ctx4096",
   "model": "MiniCPM5-2B",
   "notes": "mlx-community; oMLX community benchmark trace; peak memory 2.1 GB at this context; duplicate traces of the same session: 45bhplpc, 5zawwaq5, g6lafzj5, nv4z5hnh; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 292.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--MiniCPM5-2B-mlx-4Bit on M1 (8c); M1 (8c) \u00b7 8 GB \u00b7 4bit; 4k 292.5 26.9 2.1 GB view",
   "ram_gb": 8,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jh6gexke",
   "source_url": "https://omlx.ai/benchmarks/performance/jh6gexke",
   "tg_tokens": null,
   "tps": 26.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-jh6gexke-ctx8192",
   "model": "MiniCPM5-2B",
   "notes": "mlx-community; oMLX community benchmark trace; peak memory 2.3 GB at this context; duplicate traces of the same session: 45bhplpc, 5zawwaq5, g6lafzj5, nv4z5hnh; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 247.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--MiniCPM5-2B-mlx-4Bit on M1 (8c); M1 (8c) \u00b7 8 GB \u00b7 4bit; 8k 247.1 19.4 2.3 GB view",
   "ram_gb": 8,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jh6gexke",
   "source_url": "https://omlx.ai/benchmarks/performance/jh6gexke",
   "tg_tokens": null,
   "tps": 19.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-jh6gexke-ctx16384",
   "model": "MiniCPM5-2B",
   "notes": "mlx-community; oMLX community benchmark trace; peak memory 2.6 GB at this context; duplicate traces of the same session: 45bhplpc, 5zawwaq5, g6lafzj5, nv4z5hnh; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 190.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--MiniCPM5-2B-mlx-4Bit on M1 (8c); M1 (8c) \u00b7 8 GB \u00b7 4bit; 16k 190.8 15.5 2.6 GB view",
   "ram_gb": 8,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jh6gexke",
   "source_url": "https://omlx.ai/benchmarks/performance/jh6gexke",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-jh6gexke-ctx32768",
   "model": "MiniCPM5-2B",
   "notes": "mlx-community; oMLX community benchmark trace; peak memory 3.1 GB at this context; duplicate traces of the same session: 45bhplpc, 5zawwaq5, g6lafzj5, nv4z5hnh; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 133.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--MiniCPM5-2B-mlx-4Bit on M1 (8c); M1 (8c) \u00b7 8 GB \u00b7 4bit; 32k 133.4 10.1 3.1 GB current",
   "ram_gb": 8,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jh6gexke",
   "source_url": "https://omlx.ai/benchmarks/performance/jh6gexke",
   "tg_tokens": null,
   "tps": 10.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-tsrte0mf-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 289.3 GB at this context; duplicate traces of the same session: d0j0ddf2, jtnaow01, klhl0m66, wbmtvvbg; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 592.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 8k 592.1 30.7 289.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tsrte0mf",
   "source_url": "https://omlx.ai/benchmarks/performance/tsrte0mf",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-tsrte0mf-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 289.3 GB at this context; duplicate traces of the same session: d0j0ddf2, jtnaow01, klhl0m66, wbmtvvbg; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 717.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 16k 717.4 35.8 289.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tsrte0mf",
   "source_url": "https://omlx.ai/benchmarks/performance/tsrte0mf",
   "tg_tokens": null,
   "tps": 35.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-tsrte0mf-ctx32768",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 289.3 GB at this context; duplicate traces of the same session: d0j0ddf2, jtnaow01, klhl0m66, wbmtvvbg; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 729.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 32k 729.9 41.7 289.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tsrte0mf",
   "source_url": "https://omlx.ai/benchmarks/performance/tsrte0mf",
   "tg_tokens": null,
   "tps": 41.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-tsrte0mf-ctx65536",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 289.3 GB at this context; duplicate traces of the same session: d0j0ddf2, jtnaow01, klhl0m66, wbmtvvbg; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 720.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 64k 720.5 35.8 289.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tsrte0mf",
   "source_url": "https://omlx.ai/benchmarks/performance/tsrte0mf",
   "tg_tokens": null,
   "tps": 35.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-tsrte0mf-ctx131072",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 289.4 GB at this context; duplicate traces of the same session: d0j0ddf2, jtnaow01, klhl0m66, wbmtvvbg; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 698.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 128k 698.2 30.8 289.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tsrte0mf",
   "source_url": "https://omlx.ai/benchmarks/performance/tsrte0mf",
   "tg_tokens": null,
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ne5oragw-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "EigenLabs; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: jwpze1sr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 938.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "EigenLabs--Qwen3.8-27B-4bit-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 938.4 43.7 16.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ne5oragw",
   "source_url": "https://omlx.ai/benchmarks/performance/ne5oragw",
   "tg_tokens": null,
   "tps": 43.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ne5oragw-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "EigenLabs; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: jwpze1sr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1309.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "EigenLabs--Qwen3.8-27B-4bit-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,309 64.8 18.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ne5oragw",
   "source_url": "https://omlx.ai/benchmarks/performance/ne5oragw",
   "tg_tokens": null,
   "tps": 64.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ne5oragw-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "EigenLabs; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: jwpze1sr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1374.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "EigenLabs--Qwen3.8-27B-4bit-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,374 69.0 19.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ne5oragw",
   "source_url": "https://omlx.ai/benchmarks/performance/ne5oragw",
   "tg_tokens": null,
   "tps": 69.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-t1df6urb-ctx1024",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: tmrzt1ec; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 649.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 649.7 94.7 18.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t1df6urb",
   "source_url": "https://omlx.ai/benchmarks/performance/t1df6urb",
   "tg_tokens": null,
   "tps": 94.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-t1df6urb-ctx4096",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: tmrzt1ec; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 682.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 682.2 32.2 19.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t1df6urb",
   "source_url": "https://omlx.ai/benchmarks/performance/t1df6urb",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-caiuxbm5-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: w5gcpmbd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 553.4,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 553.4 113.0 21.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/caiuxbm5",
   "source_url": "https://omlx.ai/benchmarks/performance/caiuxbm5",
   "tg_tokens": null,
   "tps": 113.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-caiuxbm5-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: w5gcpmbd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1833.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,833 106.8 22.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/caiuxbm5",
   "source_url": "https://omlx.ai/benchmarks/performance/caiuxbm5",
   "tg_tokens": null,
   "tps": 106.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-wf4r1a2g-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: dgnt505f; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 779.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 779.4 90.1 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wf4r1a2g",
   "source_url": "https://omlx.ai/benchmarks/performance/wf4r1a2g",
   "tg_tokens": null,
   "tps": 90.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-wf4r1a2g-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: dgnt505f; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 885.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 885.1 84.1 21.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wf4r1a2g",
   "source_url": "https://omlx.ai/benchmarks/performance/wf4r1a2g",
   "tg_tokens": null,
   "tps": 84.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-ocd2whcl-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "DWQ build; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 973.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit-DWQ on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 8bit; 1k 973.2 98.4 20.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ocd2whcl",
   "source_url": "https://omlx.ai/benchmarks/performance/ocd2whcl",
   "tg_tokens": null,
   "tps": 98.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-ocd2whcl-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "DWQ build; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1467.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit-DWQ on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 8bit; 8k 1,467 89.8 22.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ocd2whcl",
   "source_url": "https://omlx.ai/benchmarks/performance/ocd2whcl",
   "tg_tokens": null,
   "tps": 89.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-81806g3j-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: d4ehvk5n; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 109.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 109.0 16.2 20.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/81806g3j",
   "source_url": "https://omlx.ai/benchmarks/performance/81806g3j",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-81806g3j-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: d4ehvk5n; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 108.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 108.1 15.9 23.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/81806g3j",
   "source_url": "https://omlx.ai/benchmarks/performance/81806g3j",
   "tg_tokens": null,
   "tps": 15.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n8ggxuae-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 9z1169ax, hcgpbxwo, plxfmr3s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 857.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 857.8 33.6 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n8ggxuae",
   "source_url": "https://omlx.ai/benchmarks/performance/n8ggxuae",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n8ggxuae-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 9z1169ax, hcgpbxwo, plxfmr3s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 949.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 949.1 33.3 17.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n8ggxuae",
   "source_url": "https://omlx.ai/benchmarks/performance/n8ggxuae",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-n8ggxuae-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 9z1169ax, hcgpbxwo, plxfmr3s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1008.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,008 32.6 17.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n8ggxuae",
   "source_url": "https://omlx.ai/benchmarks/performance/n8ggxuae",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n8ggxuae-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 9z1169ax, hcgpbxwo, plxfmr3s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 989.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 989.0 31.4 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n8ggxuae",
   "source_url": "https://omlx.ai/benchmarks/performance/n8ggxuae",
   "tg_tokens": null,
   "tps": 31.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n8d60ytn-ctx131072",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace, single run at 128k context; peak memory 93.9 GB; TTFT 63514 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2064.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k tokens 2,064 PP tok/s 44.1 TG tok/s 63514 TTFT (ms) 93.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n8d60ytn",
   "source_url": "https://omlx.ai/benchmarks/performance/n8d60ytn",
   "tg_tokens": null,
   "tps": 44.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-r64a1e1m-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: go5g2ms0, q330v1t5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 612.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 612.3 57.1 71.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r64a1e1m",
   "source_url": "https://omlx.ai/benchmarks/performance/r64a1e1m",
   "tg_tokens": null,
   "tps": 57.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-r64a1e1m-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.0 GB at this context; duplicate traces of the same session: go5g2ms0, q330v1t5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 819.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 819.9 42.9 73.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r64a1e1m",
   "source_url": "https://omlx.ai/benchmarks/performance/r64a1e1m",
   "tg_tokens": null,
   "tps": 42.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-r64a1e1m-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 75.1 GB at this context; duplicate traces of the same session: go5g2ms0, q330v1t5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 811.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 811.4 58.0 75.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r64a1e1m",
   "source_url": "https://omlx.ai/benchmarks/performance/r64a1e1m",
   "tg_tokens": null,
   "tps": 58.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-czsssyif-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.5 GB; TTFT 2426 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 422.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 422.0 PP tok/s 24.6 TG tok/s 2426 TTFT (ms) 70.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/czsssyif",
   "source_url": "https://omlx.ai/benchmarks/performance/czsssyif",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wzuc8h9s-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: o5eah6o4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 997.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 997.7 51.3 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wzuc8h9s",
   "source_url": "https://omlx.ai/benchmarks/performance/wzuc8h9s",
   "tg_tokens": null,
   "tps": 51.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wzuc8h9s-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.4 GB at this context; duplicate traces of the same session: o5eah6o4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2264.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,264 53.1 78.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wzuc8h9s",
   "source_url": "https://omlx.ai/benchmarks/performance/wzuc8h9s",
   "tg_tokens": null,
   "tps": 53.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-jp3k9482-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 101.6 GB at this context; duplicate traces of the same session: hrsx6wxx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1068.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 4k 1,068 80.1 101.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jp3k9482",
   "source_url": "https://omlx.ai/benchmarks/performance/jp3k9482",
   "tg_tokens": null,
   "tps": 80.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jp3k9482-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 104.4 GB at this context; duplicate traces of the same session: hrsx6wxx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1317.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 64k 1,317 76.5 104.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jp3k9482",
   "source_url": "https://omlx.ai/benchmarks/performance/jp3k9482",
   "tg_tokens": null,
   "tps": 76.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-jp3k9482-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 108.6 GB at this context; duplicate traces of the same session: hrsx6wxx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1288.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 128k 1,288 57.1 108.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jp3k9482",
   "source_url": "https://omlx.ai/benchmarks/performance/jp3k9482",
   "tg_tokens": null,
   "tps": 57.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jp3k9482-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 110.2 GB at this context; duplicate traces of the same session: hrsx6wxx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1181.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 195k 1,181 46.5 110.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jp3k9482",
   "source_url": "https://omlx.ai/benchmarks/performance/jp3k9482",
   "tg_tokens": null,
   "tps": 46.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-mwcjskna-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 126.5 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 578.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 578.8 44.9 126.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mwcjskna",
   "source_url": "https://omlx.ai/benchmarks/performance/mwcjskna",
   "tg_tokens": null,
   "tps": 44.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-mwcjskna-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 127.1 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 767.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 767.2 48.3 127.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mwcjskna",
   "source_url": "https://omlx.ai/benchmarks/performance/mwcjskna",
   "tg_tokens": null,
   "tps": 48.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-em30n5lc-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 11.8 GB; TTFT 517 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1982.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k tokens 1,982 PP tok/s 64.6 TG tok/s 517 TTFT (ms) 11.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/em30n5lc",
   "source_url": "https://omlx.ai/benchmarks/performance/em30n5lc",
   "tg_tokens": null,
   "tps": 64.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qms78spk-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 11.8 GB; TTFT 540 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1895.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k tokens 1,895 PP tok/s 61.1 TG tok/s 540 TTFT (ms) 11.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qms78spk",
   "source_url": "https://omlx.ai/benchmarks/performance/qms78spk",
   "tg_tokens": null,
   "tps": 61.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-d24ycsyq-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 11.8 GB; TTFT 493 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2079.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k tokens 2,079 PP tok/s 69.9 TG tok/s 493 TTFT (ms) 11.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d24ycsyq",
   "source_url": "https://omlx.ai/benchmarks/performance/d24ycsyq",
   "tg_tokens": null,
   "tps": 69.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-mypgnfov-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 11.8 GB; TTFT 503 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2036.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k tokens 2,036 PP tok/s 66.3 TG tok/s 503 TTFT (ms) 11.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mypgnfov",
   "source_url": "https://omlx.ai/benchmarks/performance/mypgnfov",
   "tg_tokens": null,
   "tps": 66.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-t7fm8c80-ctx1024",
   "model": "KAT-Coder-V2.5-Dev",
   "notes": "oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: etvxwz8z; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 773.7,
   "provenance": "community",
   "quant": "OptiQ 4bit",
   "quote": "KAT-Coder-V2.5-Dev-OptiQ-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 773.7 68.4 21.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t7fm8c80",
   "source_url": "https://omlx.ai/benchmarks/performance/t7fm8c80",
   "tg_tokens": null,
   "tps": 68.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-t7fm8c80-ctx4096",
   "model": "KAT-Coder-V2.5-Dev",
   "notes": "oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: etvxwz8z; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 870.0,
   "provenance": "community",
   "quant": "OptiQ 4bit",
   "quote": "KAT-Coder-V2.5-Dev-OptiQ-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 870.0 65.7 23.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t7fm8c80",
   "source_url": "https://omlx.ai/benchmarks/performance/t7fm8c80",
   "tg_tokens": null,
   "tps": 65.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-vy7677eo-ctx1024",
   "model": "Qwen2.5-Coder-14B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 8.4 GB at this context; duplicate traces of the same session: 7r9jbr5i; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 15.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-14B-Instruct on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 4bit; 1k 15.3 1.9 8.4 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vy7677eo",
   "source_url": "https://omlx.ai/benchmarks/performance/vy7677eo",
   "tg_tokens": null,
   "tps": 1.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-vy7677eo-ctx4096",
   "model": "Qwen2.5-Coder-14B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 9.0 GB at this context; duplicate traces of the same session: 7r9jbr5i; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 10.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-14B-Instruct on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 4bit; 4k 10.3 1.7 9.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vy7677eo",
   "source_url": "https://omlx.ai/benchmarks/performance/vy7677eo",
   "tg_tokens": null,
   "tps": 1.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-57witigl-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.1 GB at this context; duplicate traces of the same session: cq19yohi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3090.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 3,090 88.4 43.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/57witigl",
   "source_url": "https://omlx.ai/benchmarks/performance/57witigl",
   "tg_tokens": null,
   "tps": 88.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-57witigl-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.9 GB at this context; duplicate traces of the same session: cq19yohi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3552.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 3,552 84.2 43.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/57witigl",
   "source_url": "https://omlx.ai/benchmarks/performance/57witigl",
   "tg_tokens": null,
   "tps": 84.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-o4vpgyft-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 20.1 GB; TTFT 1023 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1001.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k tokens 1,001 PP tok/s 81.6 TG tok/s 1023 TTFT (ms) 20.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o4vpgyft",
   "source_url": "https://omlx.ai/benchmarks/performance/o4vpgyft",
   "tg_tokens": null,
   "tps": 81.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-7a6xssx2-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 20.1 GB; TTFT 420 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2437.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k tokens 2,437 PP tok/s 81.5 TG tok/s 420 TTFT (ms) 20.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7a6xssx2",
   "source_url": "https://omlx.ai/benchmarks/performance/7a6xssx2",
   "tg_tokens": null,
   "tps": 81.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-pq5ddirl-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 20.1 GB; TTFT 427 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2398.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k tokens 2,398 PP tok/s 74.3 TG tok/s 427 TTFT (ms) 20.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pq5ddirl",
   "source_url": "https://omlx.ai/benchmarks/performance/pq5ddirl",
   "tg_tokens": null,
   "tps": 74.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-kachnm80-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: wvuv3b5b; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 423.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 423.9 27.4 17.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kachnm80",
   "source_url": "https://omlx.ai/benchmarks/performance/kachnm80",
   "tg_tokens": null,
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-kachnm80-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: wvuv3b5b; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 461.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 461.3 33.2 18.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kachnm80",
   "source_url": "https://omlx.ai/benchmarks/performance/kachnm80",
   "tg_tokens": null,
   "tps": 33.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-nkowcas3-ctx1024",
   "model": "Qwen3.8-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: f1kjmrvr; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1055.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-mlx-4Bit on M6 (12c); M6 (12c) \u00b7 16 GB \u00b7 4bit; 1k 1,055 30.0 5.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nkowcas3",
   "source_url": "https://omlx.ai/benchmarks/performance/nkowcas3",
   "tg_tokens": null,
   "tps": 30.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-nkowcas3-ctx4096",
   "model": "Qwen3.8-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: f1kjmrvr; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1121.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-mlx-4Bit on M6 (12c); M6 (12c) \u00b7 16 GB \u00b7 4bit; 4k 1,121 29.4 6.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nkowcas3",
   "source_url": "https://omlx.ai/benchmarks/performance/nkowcas3",
   "tg_tokens": null,
   "tps": 29.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ss3b1z63-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 3gabprkm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 387.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 387.5 34.0 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ss3b1z63",
   "source_url": "https://omlx.ai/benchmarks/performance/ss3b1z63",
   "tg_tokens": null,
   "tps": 34.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ss3b1z63-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.4 GB at this context; duplicate traces of the same session: 3gabprkm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2266.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,266 57.8 78.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ss3b1z63",
   "source_url": "https://omlx.ai/benchmarks/performance/ss3b1z63",
   "tg_tokens": null,
   "tps": 57.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-v312bitc-ctx4096",
   "model": "Qwen2.5-Coder-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: bpn4b8p6; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 321.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-7B-Instruct-4bit on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 321.6 38.1 4.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v312bitc",
   "source_url": "https://omlx.ai/benchmarks/performance/v312bitc",
   "tg_tokens": null,
   "tps": 38.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-v312bitc-ctx16384",
   "model": "Qwen2.5-Coder-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 5.5 GB at this context; duplicate traces of the same session: bpn4b8p6; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 271.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-7B-Instruct-4bit on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k 271.3 30.5 5.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v312bitc",
   "source_url": "https://omlx.ai/benchmarks/performance/v312bitc",
   "tg_tokens": null,
   "tps": 30.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-fdrs9nks-ctx4096",
   "model": "Qwen2.5-Coder-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: nz2r05p0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 319.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-7B-Instruct-4bit on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 319.6 38.9 4.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fdrs9nks",
   "source_url": "https://omlx.ai/benchmarks/performance/fdrs9nks",
   "tg_tokens": null,
   "tps": 38.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-fdrs9nks-ctx16384",
   "model": "Qwen2.5-Coder-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 5.5 GB at this context; duplicate traces of the same session: nz2r05p0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 271.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-7B-Instruct-4bit on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k 271.2 30.5 5.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fdrs9nks",
   "source_url": "https://omlx.ai/benchmarks/performance/fdrs9nks",
   "tg_tokens": null,
   "tps": 30.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-cq87v9p7-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.4 GB at this context; duplicate traces of the same session: yyrzprhh; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 425.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 425.4 69.4 100.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cq87v9p7",
   "source_url": "https://omlx.ai/benchmarks/performance/cq87v9p7",
   "tg_tokens": null,
   "tps": 69.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-cq87v9p7-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: yyrzprhh; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1357.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,357 81.1 101.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cq87v9p7",
   "source_url": "https://omlx.ai/benchmarks/performance/cq87v9p7",
   "tg_tokens": null,
   "tps": 81.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5urq2ke1-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: bczii21h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1188.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,188 65.8 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5urq2ke1",
   "source_url": "https://omlx.ai/benchmarks/performance/5urq2ke1",
   "tg_tokens": null,
   "tps": 65.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5urq2ke1-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.7 GB at this context; duplicate traces of the same session: bczii21h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2353.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,353 47.3 78.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5urq2ke1",
   "source_url": "https://omlx.ai/benchmarks/performance/5urq2ke1",
   "tg_tokens": null,
   "tps": 47.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rc3ik16y-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 9f92k4rf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1200.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,200 51.7 70.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rc3ik16y",
   "source_url": "https://omlx.ai/benchmarks/performance/rc3ik16y",
   "tg_tokens": null,
   "tps": 51.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rc3ik16y-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.4 GB at this context; duplicate traces of the same session: 9f92k4rf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2206.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,206 60.5 78.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rc3ik16y",
   "source_url": "https://omlx.ai/benchmarks/performance/rc3ik16y",
   "tg_tokens": null,
   "tps": 60.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2kf4wz4v-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: n3ciofq4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1068.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,068 47.6 70.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2kf4wz4v",
   "source_url": "https://omlx.ai/benchmarks/performance/2kf4wz4v",
   "tg_tokens": null,
   "tps": 47.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2kf4wz4v-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.7 GB at this context; duplicate traces of the same session: n3ciofq4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2299.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,299 54.0 78.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2kf4wz4v",
   "source_url": "https://omlx.ai/benchmarks/performance/2kf4wz4v",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-j2nv98t9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Vontra; oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 408ljcg0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 390.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Vontra--Qwen3.8-27B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 390.3 18.0 16.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j2nv98t9",
   "source_url": "https://omlx.ai/benchmarks/performance/j2nv98t9",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-j2nv98t9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "Vontra; oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 408ljcg0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 416.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Vontra--Qwen3.8-27B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 416.0 17.6 17.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j2nv98t9",
   "source_url": "https://omlx.ai/benchmarks/performance/j2nv98t9",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-rj3f9r29-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace, single run at 1k context; peak memory 14.5 GB; TTFT 2226 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 459.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k tokens 459.9 PP tok/s 20.5 TG tok/s 2226 TTFT (ms) 14.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rj3f9r29",
   "source_url": "https://omlx.ai/benchmarks/performance/rj3f9r29",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-bsml6sif-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace, single run at 1k context; peak memory 14.5 GB; TTFT 1844 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 555.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k tokens 555.4 PP tok/s 27.2 TG tok/s 1844 TTFT (ms) 14.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bsml6sif",
   "source_url": "https://omlx.ai/benchmarks/performance/bsml6sif",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-g4252ubf-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace, single run at 1k context; peak memory 14.5 GB; TTFT 1982 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 516.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k tokens 516.6 PP tok/s 23.3 TG tok/s 1982 TTFT (ms) 14.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g4252ubf",
   "source_url": "https://omlx.ai/benchmarks/performance/g4252ubf",
   "tg_tokens": null,
   "tps": 23.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-bxzn3tbu-ctx1024",
   "model": "Kimi-Linear-48B-A3B-Instruct",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 26.9 GB; TTFT 558 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1836.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Kimi-Linear-48B-A3B-Instruct-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k tokens 1,836 PP tok/s 74.6 TG tok/s 558 TTFT (ms) 26.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bxzn3tbu",
   "source_url": "https://omlx.ai/benchmarks/performance/bxzn3tbu",
   "tg_tokens": null,
   "tps": 74.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-9ob53jld-ctx1024",
   "model": "Kimi-Linear-48B-A3B-Instruct",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 26.9 GB; TTFT 553 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1851.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Kimi-Linear-48B-A3B-Instruct-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k tokens 1,851 PP tok/s 75.3 TG tok/s 553 TTFT (ms) 26.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ob53jld",
   "source_url": "https://omlx.ai/benchmarks/performance/9ob53jld",
   "tg_tokens": null,
   "tps": 75.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-8hov3oia-ctx1024",
   "model": "Kimi-Linear-48B-A3B-Instruct",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 26.9 GB; TTFT 928 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1104.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Kimi-Linear-48B-A3B-Instruct-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k tokens 1,104 PP tok/s 64.3 TG tok/s 928 TTFT (ms) 26.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8hov3oia",
   "source_url": "https://omlx.ai/benchmarks/performance/8hov3oia",
   "tg_tokens": null,
   "tps": 64.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3ymrsjqe-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: v8lq8fvb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4570.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,570 111.6 21.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ymrsjqe",
   "source_url": "https://omlx.ai/benchmarks/performance/3ymrsjqe",
   "tg_tokens": null,
   "tps": 111.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3ymrsjqe-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: v8lq8fvb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6204.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,204 105.7 22.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ymrsjqe",
   "source_url": "https://omlx.ai/benchmarks/performance/3ymrsjqe",
   "tg_tokens": null,
   "tps": 105.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7ymy01wy-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 2n15q3wc, s5s4gnhk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 922.8,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 922.8 44.7 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ymy01wy",
   "source_url": "https://omlx.ai/benchmarks/performance/7ymy01wy",
   "tg_tokens": null,
   "tps": 44.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7ymy01wy-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: 2n15q3wc, s5s4gnhk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 947.3,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 947.3 43.8 20.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ymy01wy",
   "source_url": "https://omlx.ai/benchmarks/performance/7ymy01wy",
   "tg_tokens": null,
   "tps": 43.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7ymy01wy-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 2n15q3wc, s5s4gnhk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 900.6,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 900.6 43.6 20.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ymy01wy",
   "source_url": "https://omlx.ai/benchmarks/performance/7ymy01wy",
   "tg_tokens": null,
   "tps": 43.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-kzwnwmvx-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: igh13gug, qw3guge0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 916.1,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 916.1 72.2 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kzwnwmvx",
   "source_url": "https://omlx.ai/benchmarks/performance/kzwnwmvx",
   "tg_tokens": null,
   "tps": 72.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kzwnwmvx-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: igh13gug, qw3guge0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 952.4,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 952.4 62.1 20.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kzwnwmvx",
   "source_url": "https://omlx.ai/benchmarks/performance/kzwnwmvx",
   "tg_tokens": null,
   "tps": 62.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-kzwnwmvx-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: igh13gug, qw3guge0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 907.9,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 907.9 74.3 20.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kzwnwmvx",
   "source_url": "https://omlx.ai/benchmarks/performance/kzwnwmvx",
   "tg_tokens": null,
   "tps": 74.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ds73rbes-ctx16384",
   "model": "Qwen3.8-35B-36.82G-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace, single run at 16k context; peak memory 37.7 GB; TTFT 10324 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1587.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-36.82G-A3B-Distill-oQ8-fp16-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k tokens 1,587 PP tok/s 86.0 TG tok/s 10324 TTFT (ms) 37.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ds73rbes",
   "source_url": "https://omlx.ai/benchmarks/performance/ds73rbes",
   "tg_tokens": null,
   "tps": 86.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2pn0djm2-ctx16384",
   "model": "Qwen3.8-35B-36.82G-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace, single run at 16k context; peak memory 37.7 GB; TTFT 10328 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1586.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-36.82G-A3B-Distill-oQ8-fp16-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k tokens 1,586 PP tok/s 86.1 TG tok/s 10328 TTFT (ms) 37.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2pn0djm2",
   "source_url": "https://omlx.ai/benchmarks/performance/2pn0djm2",
   "tg_tokens": null,
   "tps": 86.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-k8huv85t-ctx1024",
   "model": "K2-Horizon-MoVA-36B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: q4yjn06j; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 181.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "K2-Horizon-MoVA-36B-A4B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k 181.0 24.7 20.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k8huv85t",
   "source_url": "https://omlx.ai/benchmarks/performance/k8huv85t",
   "tg_tokens": null,
   "tps": 24.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-k8huv85t-ctx4096",
   "model": "K2-Horizon-MoVA-36B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: q4yjn06j; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 173.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "K2-Horizon-MoVA-36B-A4B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 173.5 20.3 21.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k8huv85t",
   "source_url": "https://omlx.ai/benchmarks/performance/k8huv85t",
   "tg_tokens": null,
   "tps": 20.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-8ifs93c6-ctx1024",
   "model": "KAT-Coder-V2.5-Dev",
   "notes": "oMLX community benchmark trace; peak memory 12.6 GB at this context; duplicate traces of the same session: 0nxfr6yt; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1075.0,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "KAT-Coder-V2.5-Dev-oQ2e on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 2bit; 1k 1,075 51.5 12.6 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8ifs93c6",
   "source_url": "https://omlx.ai/benchmarks/performance/8ifs93c6",
   "tg_tokens": null,
   "tps": 51.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-8ifs93c6-ctx4096",
   "model": "KAT-Coder-V2.5-Dev",
   "notes": "oMLX community benchmark trace; peak memory 13.2 GB at this context; duplicate traces of the same session: 0nxfr6yt; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1212.0,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "KAT-Coder-V2.5-Dev-oQ2e on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 2bit; 4k 1,212 50.1 13.2 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8ifs93c6",
   "source_url": "https://omlx.ai/benchmarks/performance/8ifs93c6",
   "tg_tokens": null,
   "tps": 50.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-mg9y893l-ctx1024",
   "model": "Ministral-3-8B-Instruct-2512",
   "notes": "oMLX community benchmark trace; peak memory 16.5 GB at this context; duplicate traces of the same session: osud0p40; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 754.2,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Ministral-3-8B-Instruct-2512 on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 unknown; 1k 754.2 7.7 16.5 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mg9y893l",
   "source_url": "https://omlx.ai/benchmarks/performance/mg9y893l",
   "tg_tokens": null,
   "tps": 7.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-mg9y893l-ctx4096",
   "model": "Ministral-3-8B-Instruct-2512",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: osud0p40; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 817.2,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Ministral-3-8B-Instruct-2512 on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 unknown; 4k 817.2 7.3 17.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mg9y893l",
   "source_url": "https://omlx.ai/benchmarks/performance/mg9y893l",
   "tg_tokens": null,
   "tps": 7.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lhdnn5fk-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 10fypdol, 2vc3pv4s, 30oxhvae, agjqrif8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4568.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 4,568 153.1 19.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lhdnn5fk",
   "source_url": "https://omlx.ai/benchmarks/performance/lhdnn5fk",
   "tg_tokens": null,
   "tps": 153.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lhdnn5fk-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 10fypdol, 2vc3pv4s, 30oxhvae, agjqrif8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4830.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 4,830 122.0 20.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lhdnn5fk",
   "source_url": "https://omlx.ai/benchmarks/performance/lhdnn5fk",
   "tg_tokens": null,
   "tps": 122.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lhdnn5fk-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 10fypdol, 2vc3pv4s, 30oxhvae, agjqrif8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3496.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 64k 3,496 101.4 21.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lhdnn5fk",
   "source_url": "https://omlx.ai/benchmarks/performance/lhdnn5fk",
   "tg_tokens": null,
   "tps": 101.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lhdnn5fk-ctx131072",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: 10fypdol, 2vc3pv4s, 30oxhvae, agjqrif8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2347.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 128k 2,347 78.5 22.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lhdnn5fk",
   "source_url": "https://omlx.ai/benchmarks/performance/lhdnn5fk",
   "tg_tokens": null,
   "tps": 78.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lhdnn5fk-ctx199680",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: 10fypdol, 2vc3pv4s, 30oxhvae, agjqrif8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1685.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 195k 1,685 62.1 23.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lhdnn5fk",
   "source_url": "https://omlx.ai/benchmarks/performance/lhdnn5fk",
   "tg_tokens": null,
   "tps": 62.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1dftv83b-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: k19yixhf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 765.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 765.7 150.7 19.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1dftv83b",
   "source_url": "https://omlx.ai/benchmarks/performance/1dftv83b",
   "tg_tokens": null,
   "tps": 150.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1dftv83b-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: k19yixhf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2259.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 2,259 148.4 20.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1dftv83b",
   "source_url": "https://omlx.ai/benchmarks/performance/1dftv83b",
   "tg_tokens": null,
   "tps": 148.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-d67qe5hg-ctx1024",
   "model": "Qwen3.5-9B-Uncensored-HauhauCS-Aggressive",
   "notes": "oMLX community benchmark trace; peak memory 5.6 GB at this context; duplicate traces of the same session: 1ic2sg2j, 6mxn6rt1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1661.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.5-9B-Uncensored-HauhauCS-Aggressive-MLX-mxfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 1,661 57.8 5.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d67qe5hg",
   "source_url": "https://omlx.ai/benchmarks/performance/d67qe5hg",
   "tg_tokens": null,
   "tps": 57.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-d67qe5hg-ctx4096",
   "model": "Qwen3.5-9B-Uncensored-HauhauCS-Aggressive",
   "notes": "oMLX community benchmark trace; peak memory 6.0 GB at this context; duplicate traces of the same session: 1ic2sg2j, 6mxn6rt1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1781.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.5-9B-Uncensored-HauhauCS-Aggressive-MLX-mxfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 1,781 56.5 6.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d67qe5hg",
   "source_url": "https://omlx.ai/benchmarks/performance/d67qe5hg",
   "tg_tokens": null,
   "tps": 56.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-d67qe5hg-ctx8192",
   "model": "Qwen3.5-9B-Uncensored-HauhauCS-Aggressive",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: 1ic2sg2j, 6mxn6rt1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1767.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.5-9B-Uncensored-HauhauCS-Aggressive-MLX-mxfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 1,767 54.5 6.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d67qe5hg",
   "source_url": "https://omlx.ai/benchmarks/performance/d67qe5hg",
   "tg_tokens": null,
   "tps": 54.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-x619wscr-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 23.7 GB; TTFT 82735 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 792.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k tokens 792.1 PP tok/s 57.2 TG tok/s 82735 TTFT (ms) 23.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x619wscr",
   "source_url": "https://omlx.ai/benchmarks/performance/x619wscr",
   "tg_tokens": null,
   "tps": 57.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-t80yt40u-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: qh5urmqz; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 460.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k 460.4 30.4 21.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t80yt40u",
   "source_url": "https://omlx.ai/benchmarks/performance/t80yt40u",
   "tg_tokens": null,
   "tps": 30.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-t80yt40u-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: qh5urmqz; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 508.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 508.4 39.4 22.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t80yt40u",
   "source_url": "https://omlx.ai/benchmarks/performance/t80yt40u",
   "tg_tokens": null,
   "tps": 39.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-n1w9nvud-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: y6ktto77; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 295.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 295.8 9.7 15.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n1w9nvud",
   "source_url": "https://omlx.ai/benchmarks/performance/n1w9nvud",
   "tg_tokens": null,
   "tps": 9.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-n1w9nvud-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: y6ktto77; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 310.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 310.9 9.6 17.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n1w9nvud",
   "source_url": "https://omlx.ai/benchmarks/performance/n1w9nvud",
   "tg_tokens": null,
   "tps": 9.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-nz7wi2hi-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 250.8 GB at this context; duplicate traces of the same session: 0j64bchv, 2jvo4udf, lqpreiq6, tx0830kk, wn401p8t, x3c2xvib; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 501.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-Q8mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 1k 501.7 45.1 250.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nz7wi2hi",
   "source_url": "https://omlx.ai/benchmarks/performance/nz7wi2hi",
   "tg_tokens": null,
   "tps": 45.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nz7wi2hi-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 252.2 GB at this context; duplicate traces of the same session: 0j64bchv, 2jvo4udf, lqpreiq6, tx0830kk, wn401p8t, x3c2xvib; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 531.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-Q8mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 8k 531.5 39.1 252.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nz7wi2hi",
   "source_url": "https://omlx.ai/benchmarks/performance/nz7wi2hi",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-nz7wi2hi-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 252.4 GB at this context; duplicate traces of the same session: 0j64bchv, 2jvo4udf, lqpreiq6, tx0830kk, wn401p8t, x3c2xvib; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 522.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-Q8mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 16k 522.0 76.6 252.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nz7wi2hi",
   "source_url": "https://omlx.ai/benchmarks/performance/nz7wi2hi",
   "tg_tokens": null,
   "tps": 76.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-nz7wi2hi-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 252.6 GB at this context; duplicate traces of the same session: 0j64bchv, 2jvo4udf, lqpreiq6, tx0830kk, wn401p8t, x3c2xvib; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 510.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-Q8mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 32k 510.6 54.0 252.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nz7wi2hi",
   "source_url": "https://omlx.ai/benchmarks/performance/nz7wi2hi",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nz7wi2hi-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 253.3 GB at this context; duplicate traces of the same session: 0j64bchv, 2jvo4udf, lqpreiq6, tx0830kk, wn401p8t, x3c2xvib; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 507.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-Q8mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 64k 507.6 42.6 253.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nz7wi2hi",
   "source_url": "https://omlx.ai/benchmarks/performance/nz7wi2hi",
   "tg_tokens": null,
   "tps": 42.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-nz7wi2hi-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 254.6 GB at this context; duplicate traces of the same session: 0j64bchv, 2jvo4udf, lqpreiq6, tx0830kk, wn401p8t, x3c2xvib; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 497.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-Q8mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 128k 497.4 76.8 254.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nz7wi2hi",
   "source_url": "https://omlx.ai/benchmarks/performance/nz7wi2hi",
   "tg_tokens": null,
   "tps": 76.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-nz7wi2hi-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 255.9 GB at this context; duplicate traces of the same session: 0j64bchv, 2jvo4udf, lqpreiq6, tx0830kk, wn401p8t, x3c2xvib; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 488.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-Q8mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 195k 488.4 41.6 255.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nz7wi2hi",
   "source_url": "https://omlx.ai/benchmarks/performance/nz7wi2hi",
   "tg_tokens": null,
   "tps": 41.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-r5ebbfs8-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 7xdmf21d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1774.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,774 108.0 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r5ebbfs8",
   "source_url": "https://omlx.ai/benchmarks/performance/r5ebbfs8",
   "tg_tokens": null,
   "tps": 108.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-r5ebbfs8-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 7xdmf21d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2098.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 2,098 103.3 21.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r5ebbfs8",
   "source_url": "https://omlx.ai/benchmarks/performance/r5ebbfs8",
   "tg_tokens": null,
   "tps": 103.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-5lbnod03-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: i3v16tc0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1742.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,742 128.5 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5lbnod03",
   "source_url": "https://omlx.ai/benchmarks/performance/5lbnod03",
   "tg_tokens": null,
   "tps": 128.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-5lbnod03-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: i3v16tc0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2036.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 2,036 116.6 23.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5lbnod03",
   "source_url": "https://omlx.ai/benchmarks/performance/5lbnod03",
   "tg_tokens": null,
   "tps": 116.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-c66yhwvt-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.0 GB at this context; duplicate traces of the same session: jaxdtvx5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1660.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 1,660 109.6 30.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c66yhwvt",
   "source_url": "https://omlx.ai/benchmarks/performance/c66yhwvt",
   "tg_tokens": null,
   "tps": 109.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-c66yhwvt-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.9 GB at this context; duplicate traces of the same session: jaxdtvx5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1968.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 1,968 106.8 31.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c66yhwvt",
   "source_url": "https://omlx.ai/benchmarks/performance/c66yhwvt",
   "tg_tokens": null,
   "tps": 106.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-zgywysl0-ctx1024",
   "model": "Qwable-v2",
   "notes": "DWQ; MTP speculative decoding enabled; Vision; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: ok0w3y98; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1106.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwable-v2-oQ4e-DWQ-MTP-Vision-MLX on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 1,106 69.7 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zgywysl0",
   "source_url": "https://omlx.ai/benchmarks/performance/zgywysl0",
   "tg_tokens": null,
   "tps": 69.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-zgywysl0-ctx8192",
   "model": "Qwable-v2",
   "notes": "DWQ; MTP speculative decoding enabled; Vision; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: ok0w3y98; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1347.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwable-v2-oQ4e-DWQ-MTP-Vision-MLX on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 8k 1,347 60.6 23.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zgywysl0",
   "source_url": "https://omlx.ai/benchmarks/performance/zgywysl0",
   "tg_tokens": null,
   "tps": 60.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ixjynrot-ctx1024",
   "model": "Qwen3.6-35B-A3B-Holo3-Qwopus-AgentWorld",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 29.5 GB; TTFT 582 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1761.0,
   "provenance": "community",
   "quant": "qx64 hi",
   "quote": "Qwen3.6-35B-A3B-Holo3-Qwopus-AgentWorld-qx64-hi-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k tokens 1,761 PP tok/s 73.9 TG tok/s 582 TTFT (ms) 29.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ixjynrot",
   "source_url": "https://omlx.ai/benchmarks/performance/ixjynrot",
   "tg_tokens": null,
   "tps": 73.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-8ijqln0j-ctx1024",
   "model": "Qwen3.6-35B-A3B-Holo3-Qwopus-AgentWorld",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 29.5 GB; TTFT 584 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1755.0,
   "provenance": "community",
   "quant": "qx64 hi",
   "quote": "Qwen3.6-35B-A3B-Holo3-Qwopus-AgentWorld-qx64-hi-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k tokens 1,755 PP tok/s 68.1 TG tok/s 584 TTFT (ms) 29.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8ijqln0j",
   "source_url": "https://omlx.ai/benchmarks/performance/8ijqln0j",
   "tg_tokens": null,
   "tps": 68.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-cbqp47zj-ctx1024",
   "model": "Qwen3.6-35B-A3B-Holo3-Qwopus-AgentWorld",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 29.5 GB; TTFT 585 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1750.0,
   "provenance": "community",
   "quant": "qx64 hi",
   "quote": "Qwen3.6-35B-A3B-Holo3-Qwopus-AgentWorld-qx64-hi-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k tokens 1,750 PP tok/s 70.2 TG tok/s 585 TTFT (ms) 29.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cbqp47zj",
   "source_url": "https://omlx.ai/benchmarks/performance/cbqp47zj",
   "tg_tokens": null,
   "tps": 70.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ecqv7b2k-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 23.5 GB; TTFT 87521 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 748.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k tokens 748.8 PP tok/s 54.8 TG tok/s 87521 TTFT (ms) 23.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ecqv7b2k",
   "source_url": "https://omlx.ai/benchmarks/performance/ecqv7b2k",
   "tg_tokens": null,
   "tps": 54.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-amrfxcqs-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: 3wokslru, 9txgbqlf, dyvas3bx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2463.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 2,463 89.4 35.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/amrfxcqs",
   "source_url": "https://omlx.ai/benchmarks/performance/amrfxcqs",
   "tg_tokens": null,
   "tps": 89.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-amrfxcqs-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: 3wokslru, 9txgbqlf, dyvas3bx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3349.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 3,349 88.0 37.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/amrfxcqs",
   "source_url": "https://omlx.ai/benchmarks/performance/amrfxcqs",
   "tg_tokens": null,
   "tps": 88.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-amrfxcqs-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.5 GB at this context; duplicate traces of the same session: 3wokslru, 9txgbqlf, dyvas3bx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3189.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 3,189 85.5 37.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/amrfxcqs",
   "source_url": "https://omlx.ai/benchmarks/performance/amrfxcqs",
   "tg_tokens": null,
   "tps": 85.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-amrfxcqs-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.7 GB at this context; duplicate traces of the same session: 3wokslru, 9txgbqlf, dyvas3bx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2902.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k 2,902 82.9 37.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/amrfxcqs",
   "source_url": "https://omlx.ai/benchmarks/performance/amrfxcqs",
   "tg_tokens": null,
   "tps": 82.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-kvy4am2z-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTPLX Optimized Quality FP16 build; oMLX community benchmark trace; peak memory 40.5 GB at this context; duplicate traces of the same session: d7aq7uq9; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality-FP16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 1k 104.6 14.3 40.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kvy4am2z",
   "source_url": "https://omlx.ai/benchmarks/performance/kvy4am2z",
   "tg_tokens": null,
   "tps": 14.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-kvy4am2z-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTPLX Optimized Quality FP16 build; oMLX community benchmark trace; peak memory 44.5 GB at this context; duplicate traces of the same session: d7aq7uq9; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality-FP16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 4k 105.0 11.3 44.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kvy4am2z",
   "source_url": "https://omlx.ai/benchmarks/performance/kvy4am2z",
   "tg_tokens": null,
   "tps": 11.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mr96pb6w-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 4zfmmzwp, djed42k0, jk9cxrjp, l031bgrj, s835p8o2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1127.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,127 61.2 20.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mr96pb6w",
   "source_url": "https://omlx.ai/benchmarks/performance/mr96pb6w",
   "tg_tokens": null,
   "tps": 61.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mr96pb6w-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 4zfmmzwp, djed42k0, jk9cxrjp, l031bgrj, s835p8o2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1341.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,341 58.8 21.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mr96pb6w",
   "source_url": "https://omlx.ai/benchmarks/performance/mr96pb6w",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mr96pb6w-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 4zfmmzwp, djed42k0, jk9cxrjp, l031bgrj, s835p8o2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1399.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,399 66.7 22.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mr96pb6w",
   "source_url": "https://omlx.ai/benchmarks/performance/mr96pb6w",
   "tg_tokens": null,
   "tps": 66.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mr96pb6w-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: 4zfmmzwp, djed42k0, jk9cxrjp, l031bgrj, s835p8o2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1317.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,317 83.5 22.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mr96pb6w",
   "source_url": "https://omlx.ai/benchmarks/performance/mr96pb6w",
   "tg_tokens": null,
   "tps": 83.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mr96pb6w-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: 4zfmmzwp, djed42k0, jk9cxrjp, l031bgrj, s835p8o2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1149.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 1,149 75.4 23.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mr96pb6w",
   "source_url": "https://omlx.ai/benchmarks/performance/mr96pb6w",
   "tg_tokens": null,
   "tps": 75.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mr96pb6w-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: 4zfmmzwp, djed42k0, jk9cxrjp, l031bgrj, s835p8o2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 953.5,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 953.5 53.7 25.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mr96pb6w",
   "source_url": "https://omlx.ai/benchmarks/performance/mr96pb6w",
   "tg_tokens": null,
   "tps": 53.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-t2okqqf4-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: dsyh9sp9, plwpbzpk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 955.5,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 955.5 59.6 20.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t2okqqf4",
   "source_url": "https://omlx.ai/benchmarks/performance/t2okqqf4",
   "tg_tokens": null,
   "tps": 59.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-t2okqqf4-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: dsyh9sp9, plwpbzpk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1363.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,363 58.1 21.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t2okqqf4",
   "source_url": "https://omlx.ai/benchmarks/performance/t2okqqf4",
   "tg_tokens": null,
   "tps": 58.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-t2okqqf4-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: dsyh9sp9, plwpbzpk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1390.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,390 69.5 22.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t2okqqf4",
   "source_url": "https://omlx.ai/benchmarks/performance/t2okqqf4",
   "tg_tokens": null,
   "tps": 69.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-xr3rwxay-ctx1024",
   "model": "supergemma4-26b-uncensored",
   "notes": "oMLX community benchmark trace; peak memory 14.1 GB at this context; duplicate traces of the same session: cc1taemg; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1223.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "supergemma4-26b-uncensored-mlx-4bit-v2 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 1,223 91.8 14.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xr3rwxay",
   "source_url": "https://omlx.ai/benchmarks/performance/xr3rwxay",
   "tg_tokens": null,
   "tps": 91.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-xr3rwxay-ctx8192",
   "model": "supergemma4-26b-uncensored",
   "notes": "oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: cc1taemg; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1177.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "supergemma4-26b-uncensored-mlx-4bit-v2 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 8k 1,177 85.3 14.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xr3rwxay",
   "source_url": "https://omlx.ai/benchmarks/performance/xr3rwxay",
   "tg_tokens": null,
   "tps": 85.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-6qmi029y-ctx1024",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 222 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4624.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 4,624 PP tok/s 149.7 TG tok/s 222 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6qmi029y",
   "source_url": "https://omlx.ai/benchmarks/performance/6qmi029y",
   "tg_tokens": null,
   "tps": 149.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7bsn9uif-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: toz2y59u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 834.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 834.5 33.1 15.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7bsn9uif",
   "source_url": "https://omlx.ai/benchmarks/performance/7bsn9uif",
   "tg_tokens": null,
   "tps": 33.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7bsn9uif-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: toz2y59u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 971.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 971.1 32.5 17.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7bsn9uif",
   "source_url": "https://omlx.ai/benchmarks/performance/7bsn9uif",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cxw9kdv9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: p0mf75w0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 891.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 891.8 33.2 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cxw9kdv9",
   "source_url": "https://omlx.ai/benchmarks/performance/cxw9kdv9",
   "tg_tokens": null,
   "tps": 33.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cxw9kdv9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: p0mf75w0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1018.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,018 32.9 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cxw9kdv9",
   "source_url": "https://omlx.ai/benchmarks/performance/cxw9kdv9",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7cuqe64e-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 9poo7di0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 935.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 935.5 33.4 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7cuqe64e",
   "source_url": "https://omlx.ai/benchmarks/performance/7cuqe64e",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7cuqe64e-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 9poo7di0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1012.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,012 32.8 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7cuqe64e",
   "source_url": "https://omlx.ai/benchmarks/performance/7cuqe64e",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-3ehfgqiz-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: ew3pag78, mpx9wpi5, rm6drzwn; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 100.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 100.6 22.8 23.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ehfgqiz",
   "source_url": "https://omlx.ai/benchmarks/performance/3ehfgqiz",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-3ehfgqiz-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: ew3pag78, mpx9wpi5, rm6drzwn; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 100.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 100.5 20.8 24.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ehfgqiz",
   "source_url": "https://omlx.ai/benchmarks/performance/3ehfgqiz",
   "tg_tokens": null,
   "tps": 20.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-3ehfgqiz-ctx16384",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: ew3pag78, mpx9wpi5, rm6drzwn; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 96.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 96.1 19.5 26.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ehfgqiz",
   "source_url": "https://omlx.ai/benchmarks/performance/3ehfgqiz",
   "tg_tokens": null,
   "tps": 19.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-3ehfgqiz-ctx32768",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: ew3pag78, mpx9wpi5, rm6drzwn; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 90.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 90.5 15.8 29.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ehfgqiz",
   "source_url": "https://omlx.ai/benchmarks/performance/3ehfgqiz",
   "tg_tokens": null,
   "tps": 15.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-yrpzio4l-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: cjm448an; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 686.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 686.5 37.5 19.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yrpzio4l",
   "source_url": "https://omlx.ai/benchmarks/performance/yrpzio4l",
   "tg_tokens": null,
   "tps": 37.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-yrpzio4l-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: cjm448an; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 585.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 585.8 28.9 21.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yrpzio4l",
   "source_url": "https://omlx.ai/benchmarks/performance/yrpzio4l",
   "tg_tokens": null,
   "tps": 28.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-2q8j83lt-ctx1024",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: 7qphm18l, tsnb33s3, u1c1ukef, y5njy76o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1291.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 1,291 75.2 35.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2q8j83lt",
   "source_url": "https://omlx.ai/benchmarks/performance/2q8j83lt",
   "tg_tokens": null,
   "tps": 75.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-2q8j83lt-ctx4096",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: 7qphm18l, tsnb33s3, u1c1ukef, y5njy76o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1661.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 1,661 74.3 37.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2q8j83lt",
   "source_url": "https://omlx.ai/benchmarks/performance/2q8j83lt",
   "tg_tokens": null,
   "tps": 74.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-2q8j83lt-ctx8192",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 37.5 GB at this context; duplicate traces of the same session: 7qphm18l, tsnb33s3, u1c1ukef, y5njy76o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1687.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 1,687 72.7 37.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2q8j83lt",
   "source_url": "https://omlx.ai/benchmarks/performance/2q8j83lt",
   "tg_tokens": null,
   "tps": 72.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-2q8j83lt-ctx16384",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 37.8 GB at this context; duplicate traces of the same session: 7qphm18l, tsnb33s3, u1c1ukef, y5njy76o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1365.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 1,365 68.9 37.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2q8j83lt",
   "source_url": "https://omlx.ai/benchmarks/performance/2q8j83lt",
   "tg_tokens": null,
   "tps": 68.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-2q8j83lt-ctx65536",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 39.9 GB at this context; duplicate traces of the same session: 7qphm18l, tsnb33s3, u1c1ukef, y5njy76o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 562.9,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 562.9 24.5 39.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2q8j83lt",
   "source_url": "https://omlx.ai/benchmarks/performance/2q8j83lt",
   "tg_tokens": null,
   "tps": 24.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0o5lsv7t-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 405.1 GB at this context; duplicate traces of the same session: 3a7bup3z; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 693.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 16k 693.3 38.4 405.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0o5lsv7t",
   "source_url": "https://omlx.ai/benchmarks/performance/0o5lsv7t",
   "tg_tokens": null,
   "tps": 38.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-0o5lsv7t-ctx32768",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 405.1 GB at this context; duplicate traces of the same session: 3a7bup3z; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 689.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 32k 689.0 35.1 405.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0o5lsv7t",
   "source_url": "https://omlx.ai/benchmarks/performance/0o5lsv7t",
   "tg_tokens": null,
   "tps": 35.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-63sf4kcb-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace, single run at 4k context; peak memory 405.1 GB; TTFT 6002 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 682.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k tokens 682.4 PP tok/s 33.5 TG tok/s 6002 TTFT (ms) 405.1 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/63sf4kcb",
   "source_url": "https://omlx.ai/benchmarks/performance/63sf4kcb",
   "tg_tokens": null,
   "tps": 33.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-piithfwm-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: tfn5l8vt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1807.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,807 121.6 20.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/piithfwm",
   "source_url": "https://omlx.ai/benchmarks/performance/piithfwm",
   "tg_tokens": null,
   "tps": 121.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-piithfwm-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: tfn5l8vt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2106.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 2,106 110.2 21.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/piithfwm",
   "source_url": "https://omlx.ai/benchmarks/performance/piithfwm",
   "tg_tokens": null,
   "tps": 110.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-lms574mr-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.0 GB at this context; duplicate traces of the same session: oxg0p04j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1620.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 1,620 103.6 30.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lms574mr",
   "source_url": "https://omlx.ai/benchmarks/performance/lms574mr",
   "tg_tokens": null,
   "tps": 103.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-lms574mr-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.7 GB at this context; duplicate traces of the same session: oxg0p04j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1877.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 1,877 114.4 31.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lms574mr",
   "source_url": "https://omlx.ai/benchmarks/performance/lms574mr",
   "tg_tokens": null,
   "tps": 114.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-gr02uhr6-ctx4096",
   "model": "Qwen3.6-35B-A3B-MAYA",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: l8p77hke, rwxt1u42, yci4cny8; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 891.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-MAYA-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 891.6 61.2 22.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gr02uhr6",
   "source_url": "https://omlx.ai/benchmarks/performance/gr02uhr6",
   "tg_tokens": null,
   "tps": 61.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-gr02uhr6-ctx16384",
   "model": "Qwen3.6-35B-A3B-MAYA",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: l8p77hke, rwxt1u42, yci4cny8; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 732.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-MAYA-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k 732.9 57.9 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gr02uhr6",
   "source_url": "https://omlx.ai/benchmarks/performance/gr02uhr6",
   "tg_tokens": null,
   "tps": 57.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-gr02uhr6-ctx65536",
   "model": "Qwen3.6-35B-A3B-MAYA",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: l8p77hke, rwxt1u42, yci4cny8; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 429.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-MAYA-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 64k 429.9 43.6 24.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gr02uhr6",
   "source_url": "https://omlx.ai/benchmarks/performance/gr02uhr6",
   "tg_tokens": null,
   "tps": 43.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-gr02uhr6-ctx131072",
   "model": "Qwen3.6-35B-A3B-MAYA",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: l8p77hke, rwxt1u42, yci4cny8; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 263.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-MAYA-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 128k 263.7 33.1 26.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gr02uhr6",
   "source_url": "https://omlx.ai/benchmarks/performance/gr02uhr6",
   "tg_tokens": null,
   "tps": 33.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-dlf2a2y0-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: 5kz4wcy3; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 40.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 40.1 8.2 15.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dlf2a2y0",
   "source_url": "https://omlx.ai/benchmarks/performance/dlf2a2y0",
   "tg_tokens": null,
   "tps": 8.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-dlf2a2y0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 5kz4wcy3; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 45.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 45.8 9.3 17.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dlf2a2y0",
   "source_url": "https://omlx.ai/benchmarks/performance/dlf2a2y0",
   "tg_tokens": null,
   "tps": 9.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-6ufv7j0s-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: wocnfygh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 432.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 432.7 23.7 23.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ufv7j0s",
   "source_url": "https://omlx.ai/benchmarks/performance/6ufv7j0s",
   "tg_tokens": null,
   "tps": 23.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-6ufv7j0s-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: wocnfygh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 318.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 128k 318.2 22.4 27.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ufv7j0s",
   "source_url": "https://omlx.ai/benchmarks/performance/6ufv7j0s",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-fm6zodl2-ctx1024",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 38.2 GB at this context; duplicate traces of the same session: bfgq6qv2, bkyzzkan, j4ixojc1, m4oz72w7, p1sseklt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 799.5,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 799.5 73.6 38.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fm6zodl2",
   "source_url": "https://omlx.ai/benchmarks/performance/fm6zodl2",
   "tg_tokens": null,
   "tps": 73.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-fm6zodl2-ctx4096",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 40.2 GB at this context; duplicate traces of the same session: bfgq6qv2, bkyzzkan, j4ixojc1, m4oz72w7, p1sseklt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 956.4,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 956.4 58.8 40.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fm6zodl2",
   "source_url": "https://omlx.ai/benchmarks/performance/fm6zodl2",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-fm6zodl2-ctx8192",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 40.4 GB at this context; duplicate traces of the same session: bfgq6qv2, bkyzzkan, j4ixojc1, m4oz72w7, p1sseklt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 906.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 906.0 68.5 40.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fm6zodl2",
   "source_url": "https://omlx.ai/benchmarks/performance/fm6zodl2",
   "tg_tokens": null,
   "tps": 68.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-fm6zodl2-ctx16384",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 40.8 GB at this context; duplicate traces of the same session: bfgq6qv2, bkyzzkan, j4ixojc1, m4oz72w7, p1sseklt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 798.9,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 798.9 75.1 40.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fm6zodl2",
   "source_url": "https://omlx.ai/benchmarks/performance/fm6zodl2",
   "tg_tokens": null,
   "tps": 75.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-fm6zodl2-ctx32768",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 41.5 GB at this context; duplicate traces of the same session: bfgq6qv2, bkyzzkan, j4ixojc1, m4oz72w7, p1sseklt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 644.1,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 644.1 70.8 41.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fm6zodl2",
   "source_url": "https://omlx.ai/benchmarks/performance/fm6zodl2",
   "tg_tokens": null,
   "tps": 70.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-fm6zodl2-ctx65536",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 43.0 GB at this context; duplicate traces of the same session: bfgq6qv2, bkyzzkan, j4ixojc1, m4oz72w7, p1sseklt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 453.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 453.0 38.1 43.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fm6zodl2",
   "source_url": "https://omlx.ai/benchmarks/performance/fm6zodl2",
   "tg_tokens": null,
   "tps": 38.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-hfgmboz5-ctx4096",
   "model": "Huihui-Qwen3.5-9B",
   "notes": "text; oMLX community benchmark trace; peak memory 10.6 GB at this context; duplicate traces of the same session: fh9fowda; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 283.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "huihui-qwen3.5-9b-8bit-text on M3 Pro (14c); M3 Pro (14c) \u00b7 18 GB \u00b7 8bit; 4k 283.1 15.2 10.6 GB current",
   "ram_gb": 18,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hfgmboz5",
   "source_url": "https://omlx.ai/benchmarks/performance/hfgmboz5",
   "tg_tokens": null,
   "tps": 15.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-hfgmboz5-ctx16384",
   "model": "Huihui-Qwen3.5-9B",
   "notes": "text; oMLX community benchmark trace; peak memory 11.4 GB at this context; duplicate traces of the same session: fh9fowda; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 271.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "huihui-qwen3.5-9b-8bit-text on M3 Pro (14c); M3 Pro (14c) \u00b7 18 GB \u00b7 8bit; 16k 271.7 14.5 11.4 GB view",
   "ram_gb": 18,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hfgmboz5",
   "source_url": "https://omlx.ai/benchmarks/performance/hfgmboz5",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-of6njcc1-ctx4096",
   "model": "Qwen3.6-35B-A3B-MAYA",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: alj8g6jh, bvag1ca1, wpzfive6; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 825.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-MAYA-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 825.2 31.3 24.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/of6njcc1",
   "source_url": "https://omlx.ai/benchmarks/performance/of6njcc1",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-of6njcc1-ctx16384",
   "model": "Qwen3.6-35B-A3B-MAYA",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: alj8g6jh, bvag1ca1, wpzfive6; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 591.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-MAYA-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k 591.5 41.8 26.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/of6njcc1",
   "source_url": "https://omlx.ai/benchmarks/performance/of6njcc1",
   "tg_tokens": null,
   "tps": 41.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-of6njcc1-ctx65536",
   "model": "Qwen3.6-35B-A3B-MAYA",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: alj8g6jh, bvag1ca1, wpzfive6; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 398.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-MAYA-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 64k 398.0 46.5 27.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/of6njcc1",
   "source_url": "https://omlx.ai/benchmarks/performance/of6njcc1",
   "tg_tokens": null,
   "tps": 46.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-of6njcc1-ctx131072",
   "model": "Qwen3.6-35B-A3B-MAYA",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.0 GB at this context; duplicate traces of the same session: alj8g6jh, bvag1ca1, wpzfive6; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 258.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-MAYA-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 128k 258.0 30.8 29.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/of6njcc1",
   "source_url": "https://omlx.ai/benchmarks/performance/of6njcc1",
   "tg_tokens": null,
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-021zf3cc-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 17.2 GB; TTFT 3481 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 294.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-mlx-4bit on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 1k tokens 294.2 PP tok/s 17.2 TG tok/s 3481 TTFT (ms) 17.2 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/021zf3cc",
   "source_url": "https://omlx.ai/benchmarks/performance/021zf3cc",
   "tg_tokens": null,
   "tps": 17.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-4qkcf71r-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 15.9 GB; TTFT 3496 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 292.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-mlx-4bit on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 1k tokens 292.9 PP tok/s 9.4 TG tok/s 3496 TTFT (ms) 15.9 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4qkcf71r",
   "source_url": "https://omlx.ai/benchmarks/performance/4qkcf71r",
   "tg_tokens": null,
   "tps": 9.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-62vlflv1-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: 1b07skdx, 1ohogfos, 3w56jti1, 60m5tugm, 7qekqk13; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 752.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 752.9 15.3 71.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/62vlflv1",
   "source_url": "https://omlx.ai/benchmarks/performance/62vlflv1",
   "tg_tokens": null,
   "tps": 15.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-62vlflv1-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: 1b07skdx, 1ohogfos, 3w56jti1, 60m5tugm, 7qekqk13; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2009.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,009 57.0 71.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/62vlflv1",
   "source_url": "https://omlx.ai/benchmarks/performance/62vlflv1",
   "tg_tokens": null,
   "tps": 57.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-62vlflv1-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 1b07skdx, 1ohogfos, 3w56jti1, 60m5tugm, 7qekqk13; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1807.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,807 67.6 71.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/62vlflv1",
   "source_url": "https://omlx.ai/benchmarks/performance/62vlflv1",
   "tg_tokens": null,
   "tps": 67.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-62vlflv1-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 76.7 GB at this context; duplicate traces of the same session: 1b07skdx, 1ohogfos, 3w56jti1, 60m5tugm, 7qekqk13; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2198.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,198 65.4 76.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/62vlflv1",
   "source_url": "https://omlx.ai/benchmarks/performance/62vlflv1",
   "tg_tokens": null,
   "tps": 65.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-62vlflv1-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 82.2 GB at this context; duplicate traces of the same session: 1b07skdx, 1ohogfos, 3w56jti1, 60m5tugm, 7qekqk13; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2215.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,215 66.9 82.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/62vlflv1",
   "source_url": "https://omlx.ai/benchmarks/performance/62vlflv1",
   "tg_tokens": null,
   "tps": 66.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-62vlflv1-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 84.5 GB at this context; duplicate traces of the same session: 1b07skdx, 1ohogfos, 3w56jti1, 60m5tugm, 7qekqk13; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2174.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,174 63.3 84.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/62vlflv1",
   "source_url": "https://omlx.ai/benchmarks/performance/62vlflv1",
   "tg_tokens": null,
   "tps": 63.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lnx19lwh-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.7 GB at this context; duplicate traces of the same session: pj3c0prc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1323.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 1,323 72.6 183.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lnx19lwh",
   "source_url": "https://omlx.ai/benchmarks/performance/lnx19lwh",
   "tg_tokens": null,
   "tps": 72.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-lnx19lwh-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.8 GB at this context; duplicate traces of the same session: pj3c0prc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1321.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 1,321 82.8 183.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lnx19lwh",
   "source_url": "https://omlx.ai/benchmarks/performance/lnx19lwh",
   "tg_tokens": null,
   "tps": 82.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ow485bia-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.0 GB at this context; duplicate traces of the same session: l2p0o2zj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1656.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 1,656 115.6 30.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ow485bia",
   "source_url": "https://omlx.ai/benchmarks/performance/ow485bia",
   "tg_tokens": null,
   "tps": 115.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ow485bia-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.9 GB at this context; duplicate traces of the same session: l2p0o2zj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1973.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 1,973 111.0 31.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ow485bia",
   "source_url": "https://omlx.ai/benchmarks/performance/ow485bia",
   "tg_tokens": null,
   "tps": 111.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-7mwp40sa-ctx4096",
   "model": "Qwen3.6-35B-A3B-MAYA",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 24.6 GB; TTFT 4741 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 864.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-MAYA-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k tokens 864.0 PP tok/s 70.7 TG tok/s 4741 TTFT (ms) 24.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7mwp40sa",
   "source_url": "https://omlx.ai/benchmarks/performance/7mwp40sa",
   "tg_tokens": null,
   "tps": 70.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-nippcp3q-ctx4096",
   "model": "Qwen3.6-35B-A3B-MAYA",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: e7qrzx9r, vv3h9aqn; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 869.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-MAYA-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 869.4 58.2 22.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nippcp3q",
   "source_url": "https://omlx.ai/benchmarks/performance/nippcp3q",
   "tg_tokens": null,
   "tps": 58.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-nippcp3q-ctx16384",
   "model": "Qwen3.6-35B-A3B-MAYA",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: e7qrzx9r, vv3h9aqn; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 735.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-MAYA-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k 735.1 49.7 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nippcp3q",
   "source_url": "https://omlx.ai/benchmarks/performance/nippcp3q",
   "tg_tokens": null,
   "tps": 49.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-nippcp3q-ctx65536",
   "model": "Qwen3.6-35B-A3B-MAYA",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: e7qrzx9r, vv3h9aqn; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 430.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-MAYA-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 64k 430.2 31.4 24.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nippcp3q",
   "source_url": "https://omlx.ai/benchmarks/performance/nippcp3q",
   "tg_tokens": null,
   "tps": 31.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vb303me8-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 15.8 GB; TTFT 1055 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 970.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 970.9 PP tok/s 33.9 TG tok/s 1055 TTFT (ms) 15.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vb303me8",
   "source_url": "https://omlx.ai/benchmarks/performance/vb303me8",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-dmb24qz2-ctx1024",
   "model": "DeepSeek-Coder-V2-Lite-Instruct",
   "notes": "MLX; oMLX community benchmark trace; peak memory 9.1 GB at this context; duplicate traces of the same session: xj89ga3f; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 894.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-Coder-V2-Lite-Instruct-4bit-mlx on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 1k 894.0 89.7 9.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dmb24qz2",
   "source_url": "https://omlx.ai/benchmarks/performance/dmb24qz2",
   "tg_tokens": null,
   "tps": 89.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-dmb24qz2-ctx4096",
   "model": "DeepSeek-Coder-V2-Lite-Instruct",
   "notes": "MLX; oMLX community benchmark trace; peak memory 10.2 GB at this context; duplicate traces of the same session: xj89ga3f; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 939.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-Coder-V2-Lite-Instruct-4bit-mlx on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k 939.4 70.4 10.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dmb24qz2",
   "source_url": "https://omlx.ai/benchmarks/performance/dmb24qz2",
   "tg_tokens": null,
   "tps": 70.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-1mg6rx39-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "refusal-ablated MTP build; MLX; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: f6avl2mm, jelo87m1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1615.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-refusal-ablated-MLX-oQ8-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 1,615 126.6 37.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1mg6rx39",
   "source_url": "https://omlx.ai/benchmarks/performance/1mg6rx39",
   "tg_tokens": null,
   "tps": 126.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-1mg6rx39-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "refusal-ablated MTP build; MLX; oMLX community benchmark trace; peak memory 39.5 GB at this context; duplicate traces of the same session: f6avl2mm, jelo87m1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1976.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-refusal-ablated-MLX-oQ8-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 1,976 129.3 39.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1mg6rx39",
   "source_url": "https://omlx.ai/benchmarks/performance/1mg6rx39",
   "tg_tokens": null,
   "tps": 129.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-1mg6rx39-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "refusal-ablated MTP build; MLX; oMLX community benchmark trace; peak memory 39.7 GB at this context; duplicate traces of the same session: f6avl2mm, jelo87m1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1798.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-refusal-ablated-MLX-oQ8-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 1,798 135.2 39.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1mg6rx39",
   "source_url": "https://omlx.ai/benchmarks/performance/1mg6rx39",
   "tg_tokens": null,
   "tps": 135.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-5a0qvegf-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "refusal-ablated MTP build; MLX; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: dsqh3qiu, sva32r3c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1617.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-refusal-ablated-MLX-oQ8-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 1,617 116.4 37.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5a0qvegf",
   "source_url": "https://omlx.ai/benchmarks/performance/5a0qvegf",
   "tg_tokens": null,
   "tps": 116.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-5a0qvegf-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "refusal-ablated MTP build; MLX; oMLX community benchmark trace; peak memory 39.5 GB at this context; duplicate traces of the same session: dsqh3qiu, sva32r3c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1972.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-refusal-ablated-MLX-oQ8-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 1,972 122.5 39.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5a0qvegf",
   "source_url": "https://omlx.ai/benchmarks/performance/5a0qvegf",
   "tg_tokens": null,
   "tps": 122.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-5a0qvegf-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "refusal-ablated MTP build; MLX; oMLX community benchmark trace; peak memory 39.7 GB at this context; duplicate traces of the same session: dsqh3qiu, sva32r3c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1836.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-refusal-ablated-MLX-oQ8-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 1,836 123.1 39.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5a0qvegf",
   "source_url": "https://omlx.ai/benchmarks/performance/5a0qvegf",
   "tg_tokens": null,
   "tps": 123.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-zb76djox-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "refusal-ablated MTP build; MLX; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: 70lnsxwt, pyolqc1j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1613.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-refusal-ablated-MLX-oQ8-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 1,613 97.4 37.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zb76djox",
   "source_url": "https://omlx.ai/benchmarks/performance/zb76djox",
   "tg_tokens": null,
   "tps": 97.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-zb76djox-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "refusal-ablated MTP build; MLX; oMLX community benchmark trace; peak memory 39.5 GB at this context; duplicate traces of the same session: 70lnsxwt, pyolqc1j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1967.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-refusal-ablated-MLX-oQ8-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 1,967 98.9 39.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zb76djox",
   "source_url": "https://omlx.ai/benchmarks/performance/zb76djox",
   "tg_tokens": null,
   "tps": 98.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-zb76djox-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "refusal-ablated MTP build; MLX; oMLX community benchmark trace; peak memory 39.7 GB at this context; duplicate traces of the same session: 70lnsxwt, pyolqc1j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1879.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-refusal-ablated-MLX-oQ8-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 1,879 113.7 39.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zb76djox",
   "source_url": "https://omlx.ai/benchmarks/performance/zb76djox",
   "tg_tokens": null,
   "tps": 113.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-lksy8fb9-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "refusal-ablated MTP build; MLX; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: jql763sz, wy5n8ytq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1624.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-refusal-ablated-MLX-oQ8-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 1,624 121.4 37.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lksy8fb9",
   "source_url": "https://omlx.ai/benchmarks/performance/lksy8fb9",
   "tg_tokens": null,
   "tps": 121.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-lksy8fb9-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "refusal-ablated MTP build; MLX; oMLX community benchmark trace; peak memory 39.5 GB at this context; duplicate traces of the same session: jql763sz, wy5n8ytq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1985.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-refusal-ablated-MLX-oQ8-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 1,985 124.1 39.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lksy8fb9",
   "source_url": "https://omlx.ai/benchmarks/performance/lksy8fb9",
   "tg_tokens": null,
   "tps": 124.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-lksy8fb9-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "refusal-ablated MTP build; MLX; oMLX community benchmark trace; peak memory 39.7 GB at this context; duplicate traces of the same session: jql763sz, wy5n8ytq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1882.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-refusal-ablated-MLX-oQ8-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 1,882 126.3 39.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lksy8fb9",
   "source_url": "https://omlx.ai/benchmarks/performance/lksy8fb9",
   "tg_tokens": null,
   "tps": 126.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-j3f51bma-ctx4096",
   "model": "Qwen3.6-35B-A3B-MAYA",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: 3xlqhg7g, dz2l0bxc; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 868.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-MAYA-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 868.7 70.3 24.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j3f51bma",
   "source_url": "https://omlx.ai/benchmarks/performance/j3f51bma",
   "tg_tokens": null,
   "tps": 70.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-j3f51bma-ctx16384",
   "model": "Qwen3.6-35B-A3B-MAYA",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: 3xlqhg7g, dz2l0bxc; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 706.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-MAYA-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k 706.3 89.6 26.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j3f51bma",
   "source_url": "https://omlx.ai/benchmarks/performance/j3f51bma",
   "tg_tokens": null,
   "tps": 89.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-j3f51bma-ctx65536",
   "model": "Qwen3.6-35B-A3B-MAYA",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: 3xlqhg7g, dz2l0bxc; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 404.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-MAYA-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 64k 404.8 50.4 27.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j3f51bma",
   "source_url": "https://omlx.ai/benchmarks/performance/j3f51bma",
   "tg_tokens": null,
   "tps": 50.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-bhi4d7fa-ctx4096",
   "model": "Qwen3.6-35B-A3B-MAYA",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: h5u0oldv; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 901.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-MAYA-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 901.3 64.9 22.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhi4d7fa",
   "source_url": "https://omlx.ai/benchmarks/performance/bhi4d7fa",
   "tg_tokens": null,
   "tps": 64.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-bhi4d7fa-ctx65536",
   "model": "Qwen3.6-35B-A3B-MAYA",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: h5u0oldv; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 430.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-MAYA-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 64k 430.9 43.5 24.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhi4d7fa",
   "source_url": "https://omlx.ai/benchmarks/performance/bhi4d7fa",
   "tg_tokens": null,
   "tps": 43.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-6yp1bfvm-ctx4096",
   "model": "Qwen3.6-35B-A3B-MAYA",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: mxce0zx8; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 775.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-MAYA-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 775.3 53.9 24.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6yp1bfvm",
   "source_url": "https://omlx.ai/benchmarks/performance/6yp1bfvm",
   "tg_tokens": null,
   "tps": 53.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-6yp1bfvm-ctx65536",
   "model": "Qwen3.6-35B-A3B-MAYA",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.2 GB at this context; duplicate traces of the same session: mxce0zx8; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 367.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-MAYA-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 64k 367.2 43.0 27.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6yp1bfvm",
   "source_url": "https://omlx.ai/benchmarks/performance/6yp1bfvm",
   "tg_tokens": null,
   "tps": 43.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-m7ag5s75-ctx4096",
   "model": "Qwen3.6-35B-A3B-MAYA",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 22.2 GB; TTFT 4578 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 894.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-MAYA-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k tokens 894.7 PP tok/s 64.4 TG tok/s 4578 TTFT (ms) 22.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m7ag5s75",
   "source_url": "https://omlx.ai/benchmarks/performance/m7ag5s75",
   "tg_tokens": null,
   "tps": 64.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-u3yer0ej-ctx4096",
   "model": "Qwen3.6-35B-A3B-MAYA",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 0naaybf0, 63qotx5q; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 663.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-MAYA-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 663.8 61.2 21.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u3yer0ej",
   "source_url": "https://omlx.ai/benchmarks/performance/u3yer0ej",
   "tg_tokens": null,
   "tps": 61.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-u3yer0ej-ctx16384",
   "model": "Qwen3.6-35B-A3B-MAYA",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 0naaybf0, 63qotx5q; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 602.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-MAYA-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k 602.3 56.3 23.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u3yer0ej",
   "source_url": "https://omlx.ai/benchmarks/performance/u3yer0ej",
   "tg_tokens": null,
   "tps": 56.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-u3yer0ej-ctx65536",
   "model": "Qwen3.6-35B-A3B-MAYA",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: 0naaybf0, 63qotx5q; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 382.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-MAYA-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 64k 382.0 43.6 24.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u3yer0ej",
   "source_url": "https://omlx.ai/benchmarks/performance/u3yer0ej",
   "tg_tokens": null,
   "tps": 43.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-t4ose3ec-ctx4096",
   "model": "Qwen3.6-35B-A3B-MAYA",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: a7vkehek; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 866.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-MAYA-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 866.8 55.7 24.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t4ose3ec",
   "source_url": "https://omlx.ai/benchmarks/performance/t4ose3ec",
   "tg_tokens": null,
   "tps": 55.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-t4ose3ec-ctx65536",
   "model": "Qwen3.6-35B-A3B-MAYA",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: a7vkehek; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 403.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-MAYA-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 64k 403.1 29.7 27.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t4ose3ec",
   "source_url": "https://omlx.ai/benchmarks/performance/t4ose3ec",
   "tg_tokens": null,
   "tps": 29.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-viwvc5no-ctx16384",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 38.6 GB at this context; duplicate traces of the same session: tu7z3kcw; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1405.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 8bit; 16k 1,405 54.7 38.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/viwvc5no",
   "source_url": "https://omlx.ai/benchmarks/performance/viwvc5no",
   "tg_tokens": null,
   "tps": 54.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-viwvc5no-ctx65536",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 41.2 GB at this context; duplicate traces of the same session: tu7z3kcw; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 804.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 8bit; 64k 804.0 48.2 41.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/viwvc5no",
   "source_url": "https://omlx.ai/benchmarks/performance/viwvc5no",
   "tg_tokens": null,
   "tps": 48.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-7npa6fz1-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 77.9 GB at this context; duplicate traces of the same session: 2yqf3zqe, 4n78a73d, 6cbt1028, hdib8t0t, nyj29vy7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 678.1,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ4-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 678.1 51.4 77.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7npa6fz1",
   "source_url": "https://omlx.ai/benchmarks/performance/7npa6fz1",
   "tg_tokens": null,
   "tps": 51.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-7npa6fz1-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 79.0 GB at this context; duplicate traces of the same session: 2yqf3zqe, 4n78a73d, 6cbt1028, hdib8t0t, nyj29vy7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1381.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ4-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 1,381 64.7 79.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7npa6fz1",
   "source_url": "https://omlx.ai/benchmarks/performance/7npa6fz1",
   "tg_tokens": null,
   "tps": 64.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-7npa6fz1-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 79.6 GB at this context; duplicate traces of the same session: 2yqf3zqe, 4n78a73d, 6cbt1028, hdib8t0t, nyj29vy7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1375.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ4-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 1,375 66.8 79.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7npa6fz1",
   "source_url": "https://omlx.ai/benchmarks/performance/7npa6fz1",
   "tg_tokens": null,
   "tps": 66.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-7npa6fz1-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 81.3 GB at this context; duplicate traces of the same session: 2yqf3zqe, 4n78a73d, 6cbt1028, hdib8t0t, nyj29vy7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1355.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ4-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 1,355 66.0 81.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7npa6fz1",
   "source_url": "https://omlx.ai/benchmarks/performance/7npa6fz1",
   "tg_tokens": null,
   "tps": 66.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-7npa6fz1-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 85.7 GB at this context; duplicate traces of the same session: 2yqf3zqe, 4n78a73d, 6cbt1028, hdib8t0t, nyj29vy7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1329.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ4-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 1,329 67.6 85.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7npa6fz1",
   "source_url": "https://omlx.ai/benchmarks/performance/7npa6fz1",
   "tg_tokens": null,
   "tps": 67.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-7npa6fz1-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 87.3 GB at this context; duplicate traces of the same session: 2yqf3zqe, 4n78a73d, 6cbt1028, hdib8t0t, nyj29vy7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1315.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ4-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 1,315 54.8 87.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7npa6fz1",
   "source_url": "https://omlx.ai/benchmarks/performance/7npa6fz1",
   "tg_tokens": null,
   "tps": 54.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ujpi0x4t-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace, single run at 4k context; peak memory 85.0 GB; TTFT 4515 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 907.3,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k tokens 907.3 PP tok/s 48.3 TG tok/s 4515 TTFT (ms) 85.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ujpi0x4t",
   "source_url": "https://omlx.ai/benchmarks/performance/ujpi0x4t",
   "tg_tokens": null,
   "tps": 48.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-cn4q9ky6-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "Alis build; oMLX community benchmark trace, single run at 16k context; peak memory 246.2 GB; TTFT 31230 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 524.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "GLM-5.3-Flash-Alis-MLX-6bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 16k tokens 524.6 PP tok/s 34.4 TG tok/s 31230 TTFT (ms) 246.2 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cn4q9ky6",
   "source_url": "https://omlx.ai/benchmarks/performance/cn4q9ky6",
   "tg_tokens": null,
   "tps": 34.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-zjqdji7p-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 175.2 GB; TTFT 32067 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 510.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k tokens 510.9 PP tok/s 38.4 TG tok/s 32067 TTFT (ms) 175.2 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zjqdji7p",
   "source_url": "https://omlx.ai/benchmarks/performance/zjqdji7p",
   "tg_tokens": null,
   "tps": 38.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-xh43t0sy-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 3dqyh00h, 89z6k6ga, cem8mgsy; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 901.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 901.1 64.7 20.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xh43t0sy",
   "source_url": "https://omlx.ai/benchmarks/performance/xh43t0sy",
   "tg_tokens": null,
   "tps": 64.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-xh43t0sy-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 3dqyh00h, 89z6k6ga, cem8mgsy; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 981.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 981.2 61.0 22.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xh43t0sy",
   "source_url": "https://omlx.ai/benchmarks/performance/xh43t0sy",
   "tg_tokens": null,
   "tps": 61.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-xh43t0sy-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: 3dqyh00h, 89z6k6ga, cem8mgsy; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 903.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 903.3 57.8 23.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xh43t0sy",
   "source_url": "https://omlx.ai/benchmarks/performance/xh43t0sy",
   "tg_tokens": null,
   "tps": 57.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-xh43t0sy-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: 3dqyh00h, 89z6k6ga, cem8mgsy; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 772.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 772.6 52.8 23.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xh43t0sy",
   "source_url": "https://omlx.ai/benchmarks/performance/xh43t0sy",
   "tg_tokens": null,
   "tps": 52.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-wh3d0ud1-ctx1024",
   "model": "Qwen3 8B",
   "notes": "oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: c7mu4xdn; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 233.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-8B-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 233.0 19.9 4.9 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wh3d0ud1",
   "source_url": "https://omlx.ai/benchmarks/performance/wh3d0ud1",
   "tg_tokens": null,
   "tps": 19.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-wh3d0ud1-ctx4096",
   "model": "Qwen3 8B",
   "notes": "oMLX community benchmark trace; peak memory 5.4 GB at this context; duplicate traces of the same session: c7mu4xdn; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 219.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-8B-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 219.2 17.7 5.4 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wh3d0ud1",
   "source_url": "https://omlx.ai/benchmarks/performance/wh3d0ud1",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ujt5zlbe-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: 2159m3t9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 890.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 890.7 82.1 6.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ujt5zlbe",
   "source_url": "https://omlx.ai/benchmarks/performance/ujt5zlbe",
   "tg_tokens": null,
   "tps": 82.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ujt5zlbe-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: 2159m3t9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 869.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 869.5 72.3 6.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ujt5zlbe",
   "source_url": "https://omlx.ai/benchmarks/performance/ujt5zlbe",
   "tg_tokens": null,
   "tps": 72.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-phsvdymd-ctx4096",
   "model": "Qwen3.6-35B-A3B-MAYA",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 24.6 GB; TTFT 4731 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 865.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-MAYA-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k tokens 865.9 PP tok/s 63.4 TG tok/s 4731 TTFT (ms) 24.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/phsvdymd",
   "source_url": "https://omlx.ai/benchmarks/performance/phsvdymd",
   "tg_tokens": null,
   "tps": 63.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-3719s830-ctx4096",
   "model": "Qwen3.6-35B-A3B-MAYA",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 24.6 GB; TTFT 4731 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 865.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-MAYA-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k tokens 865.8 PP tok/s 68.7 TG tok/s 4731 TTFT (ms) 24.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3719s830",
   "source_url": "https://omlx.ai/benchmarks/performance/3719s830",
   "tg_tokens": null,
   "tps": 68.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-lava5l5g-ctx65536",
   "model": "Qwen3.6-35B-A3B-MAYA",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 24.9 GB; TTFT 151707 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 432.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-MAYA-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 64k tokens 432.0 PP tok/s 31.4 TG tok/s 151707 TTFT (ms) 24.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lava5l5g",
   "source_url": "https://omlx.ai/benchmarks/performance/lava5l5g",
   "tg_tokens": null,
   "tps": 31.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-s8qnbucg-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: m3rkqgwt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 217.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 217.1 4.1 20.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s8qnbucg",
   "source_url": "https://omlx.ai/benchmarks/performance/s8qnbucg",
   "tg_tokens": null,
   "tps": 4.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-s8qnbucg-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: m3rkqgwt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 215.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 215.6 3.9 20.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s8qnbucg",
   "source_url": "https://omlx.ai/benchmarks/performance/s8qnbucg",
   "tg_tokens": null,
   "tps": 3.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ktjhghzc-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 1nh3ng4u, 4c7hlawh, ii4zy218, n98ibrs2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 215.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-fp16-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 215.4 46.6 22.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ktjhghzc",
   "source_url": "https://omlx.ai/benchmarks/performance/ktjhghzc",
   "tg_tokens": null,
   "tps": 46.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ktjhghzc-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: 1nh3ng4u, 4c7hlawh, ii4zy218, n98ibrs2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 250.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-fp16-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 250.6 39.1 25.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ktjhghzc",
   "source_url": "https://omlx.ai/benchmarks/performance/ktjhghzc",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ktjhghzc-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: 1nh3ng4u, 4c7hlawh, ii4zy218, n98ibrs2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 250.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-fp16-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 250.7 50.3 26.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ktjhghzc",
   "source_url": "https://omlx.ai/benchmarks/performance/ktjhghzc",
   "tg_tokens": null,
   "tps": 50.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ktjhghzc-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 29.2 GB at this context; duplicate traces of the same session: 1nh3ng4u, 4c7hlawh, ii4zy218, n98ibrs2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 247.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-fp16-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 247.1 51.0 29.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ktjhghzc",
   "source_url": "https://omlx.ai/benchmarks/performance/ktjhghzc",
   "tg_tokens": null,
   "tps": 51.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ktjhghzc-ctx65536",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 31.4 GB at this context; duplicate traces of the same session: 1nh3ng4u, 4c7hlawh, ii4zy218, n98ibrs2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 207.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-fp16-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 207.9 35.2 31.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ktjhghzc",
   "source_url": "https://omlx.ai/benchmarks/performance/ktjhghzc",
   "tg_tokens": null,
   "tps": 35.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-u9xzmvc5-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: rts42vew, u5wveind, wvjo3865; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 436.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 436.3 15.7 20.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u9xzmvc5",
   "source_url": "https://omlx.ai/benchmarks/performance/u9xzmvc5",
   "tg_tokens": null,
   "tps": 15.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-u9xzmvc5-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: rts42vew, u5wveind, wvjo3865; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 481.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 481.5 18.4 21.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u9xzmvc5",
   "source_url": "https://omlx.ai/benchmarks/performance/u9xzmvc5",
   "tg_tokens": null,
   "tps": 18.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-u9xzmvc5-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: rts42vew, u5wveind, wvjo3865; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 437.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 437.7 18.5 22.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u9xzmvc5",
   "source_url": "https://omlx.ai/benchmarks/performance/u9xzmvc5",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-u9xzmvc5-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: rts42vew, u5wveind, wvjo3865; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 420.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 420.4 5.5 24.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u9xzmvc5",
   "source_url": "https://omlx.ai/benchmarks/performance/u9xzmvc5",
   "tg_tokens": null,
   "tps": 5.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ot84pokk-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: hwov3rc0, k7bdexs8, r6odhnzp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 457.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 457.3 23.7 17.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ot84pokk",
   "source_url": "https://omlx.ai/benchmarks/performance/ot84pokk",
   "tg_tokens": null,
   "tps": 23.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ot84pokk-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: hwov3rc0, k7bdexs8, r6odhnzp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 466.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 466.5 24.3 18.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ot84pokk",
   "source_url": "https://omlx.ai/benchmarks/performance/ot84pokk",
   "tg_tokens": null,
   "tps": 24.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ot84pokk-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: hwov3rc0, k7bdexs8, r6odhnzp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 457.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 457.4 23.2 19.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ot84pokk",
   "source_url": "https://omlx.ai/benchmarks/performance/ot84pokk",
   "tg_tokens": null,
   "tps": 23.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ot84pokk-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: hwov3rc0, k7bdexs8, r6odhnzp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 426.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 426.1 22.3 19.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ot84pokk",
   "source_url": "https://omlx.ai/benchmarks/performance/ot84pokk",
   "tg_tokens": null,
   "tps": 22.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-2tvy81fh-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: 16sf82qd, cyrzvp7r; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 445.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 445.7 32.3 17.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2tvy81fh",
   "source_url": "https://omlx.ai/benchmarks/performance/2tvy81fh",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-2tvy81fh-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 16sf82qd, cyrzvp7r; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 462.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 462.3 36.9 18.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2tvy81fh",
   "source_url": "https://omlx.ai/benchmarks/performance/2tvy81fh",
   "tg_tokens": null,
   "tps": 36.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-2tvy81fh-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 16sf82qd, cyrzvp7r; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 463.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 463.6 40.0 19.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2tvy81fh",
   "source_url": "https://omlx.ai/benchmarks/performance/2tvy81fh",
   "tg_tokens": null,
   "tps": 40.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-2tvy81fh-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 16sf82qd, cyrzvp7r; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 437.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 437.5 42.7 19.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2tvy81fh",
   "source_url": "https://omlx.ai/benchmarks/performance/2tvy81fh",
   "tg_tokens": null,
   "tps": 42.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-a99it0vt-ctx16384",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 39.7 GB at this context; duplicate traces of the same session: sa4q530m; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1366.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 8bit; 16k 1,366 92.9 39.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a99it0vt",
   "source_url": "https://omlx.ai/benchmarks/performance/a99it0vt",
   "tg_tokens": null,
   "tps": 92.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-a99it0vt-ctx65536",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 42.4 GB at this context; duplicate traces of the same session: sa4q530m; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 748.4,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 8bit; 64k 748.4 51.1 42.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a99it0vt",
   "source_url": "https://omlx.ai/benchmarks/performance/a99it0vt",
   "tg_tokens": null,
   "tps": 51.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-j190juiy-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 108.3 GB at this context; duplicate traces of the same session: 7xws3h3p, a38qt63x, dccxp1jl; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1415.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ4-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 1,415 86.0 108.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j190juiy",
   "source_url": "https://omlx.ai/benchmarks/performance/j190juiy",
   "tg_tokens": null,
   "tps": 86.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-j190juiy-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 108.5 GB at this context; duplicate traces of the same session: 7xws3h3p, a38qt63x, dccxp1jl; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1406.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ4-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 1,406 100.7 108.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j190juiy",
   "source_url": "https://omlx.ai/benchmarks/performance/j190juiy",
   "tg_tokens": null,
   "tps": 100.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-j190juiy-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 115.7 GB at this context; duplicate traces of the same session: 7xws3h3p, a38qt63x, dccxp1jl; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1332.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ4-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 1,332 85.6 115.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j190juiy",
   "source_url": "https://omlx.ai/benchmarks/performance/j190juiy",
   "tg_tokens": null,
   "tps": 85.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-j190juiy-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 117.1 GB at this context; duplicate traces of the same session: 7xws3h3p, a38qt63x, dccxp1jl; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1315.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ4-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 1,315 53.5 117.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j190juiy",
   "source_url": "https://omlx.ai/benchmarks/performance/j190juiy",
   "tg_tokens": null,
   "tps": 53.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-it7226so-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 122.2 GB at this context; duplicate traces of the same session: 9ll855mk, jrrtcduh, nhimrnd1, ov9pkwx4, wl6bbavd; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1141.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 5bit; 1k 1,141 99.2 122.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/it7226so",
   "source_url": "https://omlx.ai/benchmarks/performance/it7226so",
   "tg_tokens": null,
   "tps": 99.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-it7226so-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 122.8 GB at this context; duplicate traces of the same session: 9ll855mk, jrrtcduh, nhimrnd1, ov9pkwx4, wl6bbavd; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1392.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 5bit; 4k 1,392 95.4 122.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/it7226so",
   "source_url": "https://omlx.ai/benchmarks/performance/it7226so",
   "tg_tokens": null,
   "tps": 95.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-it7226so-ctx8192",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 123.0 GB at this context; duplicate traces of the same session: 9ll855mk, jrrtcduh, nhimrnd1, ov9pkwx4, wl6bbavd; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1378.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 5bit; 8k 1,378 99.3 123.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/it7226so",
   "source_url": "https://omlx.ai/benchmarks/performance/it7226so",
   "tg_tokens": null,
   "tps": 99.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-it7226so-ctx65536",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 126.0 GB at this context; duplicate traces of the same session: 9ll855mk, jrrtcduh, nhimrnd1, ov9pkwx4, wl6bbavd; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1332.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 5bit; 64k 1,332 79.0 126.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/it7226so",
   "source_url": "https://omlx.ai/benchmarks/performance/it7226so",
   "tg_tokens": null,
   "tps": 79.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-it7226so-ctx131072",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 130.2 GB at this context; duplicate traces of the same session: 9ll855mk, jrrtcduh, nhimrnd1, ov9pkwx4, wl6bbavd; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1296.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 5bit; 128k 1,296 62.4 130.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/it7226so",
   "source_url": "https://omlx.ai/benchmarks/performance/it7226so",
   "tg_tokens": null,
   "tps": 62.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-it7226so-ctx199680",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 131.5 GB at this context; duplicate traces of the same session: 9ll855mk, jrrtcduh, nhimrnd1, ov9pkwx4, wl6bbavd; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1278.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 5bit; 195k 1,278 52.1 131.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/it7226so",
   "source_url": "https://omlx.ai/benchmarks/performance/it7226so",
   "tg_tokens": null,
   "tps": 52.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-kkf4ldgi-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.7 GB; TTFT 1737 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 589.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k tokens 589.4 PP tok/s 68.5 TG tok/s 1737 TTFT (ms) 70.7 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kkf4ldgi",
   "source_url": "https://omlx.ai/benchmarks/performance/kkf4ldgi",
   "tg_tokens": null,
   "tps": 68.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-4gju1wbp-ctx1024",
   "model": "medgemma-27b-text-it",
   "notes": "oMLX community benchmark trace; peak memory 29.3 GB at this context; duplicate traces of the same session: bbwiw2uu; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 124.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "medgemma-27b-text-it-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 124.0 9.6 29.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4gju1wbp",
   "source_url": "https://omlx.ai/benchmarks/performance/4gju1wbp",
   "tg_tokens": null,
   "tps": 9.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-4gju1wbp-ctx4096",
   "model": "medgemma-27b-text-it",
   "notes": "oMLX community benchmark trace; peak memory 29.8 GB at this context; duplicate traces of the same session: bbwiw2uu; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 110.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "medgemma-27b-text-it-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 110.6 9.2 29.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4gju1wbp",
   "source_url": "https://omlx.ai/benchmarks/performance/4gju1wbp",
   "tg_tokens": null,
   "tps": 9.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5mpdqhrh-ctx4096",
   "model": "GLM-5.3-Flash-Abliterated",
   "notes": "MLX; oMLX community benchmark trace; peak memory 179.1 GB at this context; duplicate traces of the same session: xxgfcuap; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 448.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-Abliterated-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 448.5 46.2 179.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5mpdqhrh",
   "source_url": "https://omlx.ai/benchmarks/performance/5mpdqhrh",
   "tg_tokens": null,
   "tps": 46.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5mpdqhrh-ctx16384",
   "model": "GLM-5.3-Flash-Abliterated",
   "notes": "MLX; oMLX community benchmark trace; peak memory 179.4 GB at this context; duplicate traces of the same session: xxgfcuap; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 410.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-Abliterated-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 410.4 83.7 179.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5mpdqhrh",
   "source_url": "https://omlx.ai/benchmarks/performance/5mpdqhrh",
   "tg_tokens": null,
   "tps": 83.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-7ahhscyt-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: 14e8i49o, eq3aoyi7, rblv9akk; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 927.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 927.6 69.0 20.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ahhscyt",
   "source_url": "https://omlx.ai/benchmarks/performance/7ahhscyt",
   "tg_tokens": null,
   "tps": 69.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-7ahhscyt-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 14e8i49o, eq3aoyi7, rblv9akk; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1026.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 1,026 68.4 22.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ahhscyt",
   "source_url": "https://omlx.ai/benchmarks/performance/7ahhscyt",
   "tg_tokens": null,
   "tps": 68.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-7ahhscyt-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 14e8i49o, eq3aoyi7, rblv9akk; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 944.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 944.2 64.4 22.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ahhscyt",
   "source_url": "https://omlx.ai/benchmarks/performance/7ahhscyt",
   "tg_tokens": null,
   "tps": 64.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-7ahhscyt-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: 14e8i49o, eq3aoyi7, rblv9akk; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 813.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 813.7 61.2 23.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ahhscyt",
   "source_url": "https://omlx.ai/benchmarks/performance/7ahhscyt",
   "tg_tokens": null,
   "tps": 61.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tyfttmgl-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 2w6vp652, ekk53x0s, k45w6z9e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4928.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 4,928 113.8 21.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tyfttmgl",
   "source_url": "https://omlx.ai/benchmarks/performance/tyfttmgl",
   "tg_tokens": null,
   "tps": 113.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tyfttmgl-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 2w6vp652, ekk53x0s, k45w6z9e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3561.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 3,561 95.3 21.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tyfttmgl",
   "source_url": "https://omlx.ai/benchmarks/performance/tyfttmgl",
   "tg_tokens": null,
   "tps": 95.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tyfttmgl-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: 2w6vp652, ekk53x0s, k45w6z9e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2411.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 2,411 72.4 23.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tyfttmgl",
   "source_url": "https://omlx.ai/benchmarks/performance/tyfttmgl",
   "tg_tokens": null,
   "tps": 72.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-tyfttmgl-ctx199680",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: 2w6vp652, ekk53x0s, k45w6z9e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1815.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 195k 1,815 57.9 24.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tyfttmgl",
   "source_url": "https://omlx.ai/benchmarks/performance/tyfttmgl",
   "tg_tokens": null,
   "tps": 57.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-x1easl0x-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: agq8j942, y35izkim; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 107.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 107.5 7.0 17.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x1easl0x",
   "source_url": "https://omlx.ai/benchmarks/performance/x1easl0x",
   "tg_tokens": null,
   "tps": 7.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-x1easl0x-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: agq8j942, y35izkim; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 105.4 6.7 18.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x1easl0x",
   "source_url": "https://omlx.ai/benchmarks/performance/x1easl0x",
   "tg_tokens": null,
   "tps": 6.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-x1easl0x-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: agq8j942, y35izkim; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 101.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 101.0 6.2 20.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x1easl0x",
   "source_url": "https://omlx.ai/benchmarks/performance/x1easl0x",
   "tg_tokens": null,
   "tps": 6.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-6y7ukj60-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: prrtryrf, tbog333e; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 107.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 107.3 6.9 17.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6y7ukj60",
   "source_url": "https://omlx.ai/benchmarks/performance/6y7ukj60",
   "tg_tokens": null,
   "tps": 6.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-6y7ukj60-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: prrtryrf, tbog333e; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 105.3 6.7 18.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6y7ukj60",
   "source_url": "https://omlx.ai/benchmarks/performance/6y7ukj60",
   "tg_tokens": null,
   "tps": 6.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-6y7ukj60-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: prrtryrf, tbog333e; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 100.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 100.7 6.1 20.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6y7ukj60",
   "source_url": "https://omlx.ai/benchmarks/performance/6y7ukj60",
   "tg_tokens": null,
   "tps": 6.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-k7crd31z-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 445.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 445.7 32.3 17.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k7crd31z",
   "source_url": "https://omlx.ai/benchmarks/performance/k7crd31z",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-k7crd31z-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 462.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 462.3 36.9 18.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k7crd31z",
   "source_url": "https://omlx.ai/benchmarks/performance/k7crd31z",
   "tg_tokens": null,
   "tps": 36.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-k7crd31z-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 463.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 463.6 40.0 19.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k7crd31z",
   "source_url": "https://omlx.ai/benchmarks/performance/k7crd31z",
   "tg_tokens": null,
   "tps": 40.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-k7crd31z-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 437.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 437.5 42.7 19.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k7crd31z",
   "source_url": "https://omlx.ai/benchmarks/performance/k7crd31z",
   "tg_tokens": null,
   "tps": 42.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-kdlw52l6-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: 1p0e6sdo, e8fllgq8, kcha8ksi, x5hjnmkp, xi9s2k2w; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 452.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 452.9 22.7 17.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kdlw52l6",
   "source_url": "https://omlx.ai/benchmarks/performance/kdlw52l6",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-kdlw52l6-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 1p0e6sdo, e8fllgq8, kcha8ksi, x5hjnmkp, xi9s2k2w; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 462.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 462.9 22.2 18.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kdlw52l6",
   "source_url": "https://omlx.ai/benchmarks/performance/kdlw52l6",
   "tg_tokens": null,
   "tps": 22.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-kdlw52l6-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 1p0e6sdo, e8fllgq8, kcha8ksi, x5hjnmkp, xi9s2k2w; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 441.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 441.9 24.2 19.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kdlw52l6",
   "source_url": "https://omlx.ai/benchmarks/performance/kdlw52l6",
   "tg_tokens": null,
   "tps": 24.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-kdlw52l6-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 1p0e6sdo, e8fllgq8, kcha8ksi, x5hjnmkp, xi9s2k2w; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 408.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 408.3 22.3 19.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kdlw52l6",
   "source_url": "https://omlx.ai/benchmarks/performance/kdlw52l6",
   "tg_tokens": null,
   "tps": 22.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-kdlw52l6-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 1p0e6sdo, e8fllgq8, kcha8ksi, x5hjnmkp, xi9s2k2w; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 398.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 398.8 20.4 20.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kdlw52l6",
   "source_url": "https://omlx.ai/benchmarks/performance/kdlw52l6",
   "tg_tokens": null,
   "tps": 20.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-kdlw52l6-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: 1p0e6sdo, e8fllgq8, kcha8ksi, x5hjnmkp, xi9s2k2w; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 355.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 355.1 22.1 23.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kdlw52l6",
   "source_url": "https://omlx.ai/benchmarks/performance/kdlw52l6",
   "tg_tokens": null,
   "tps": 22.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-d1yi9s74-ctx16384",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 38.6 GB at this context; duplicate traces of the same session: xjokzkzk; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1426.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 8bit; 16k 1,426 53.6 38.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d1yi9s74",
   "source_url": "https://omlx.ai/benchmarks/performance/d1yi9s74",
   "tg_tokens": null,
   "tps": 53.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-d1yi9s74-ctx65536",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 41.2 GB at this context; duplicate traces of the same session: xjokzkzk; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 793.4,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 8bit; 64k 793.4 47.5 41.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d1yi9s74",
   "source_url": "https://omlx.ai/benchmarks/performance/d1yi9s74",
   "tg_tokens": null,
   "tps": 47.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-sdzn8v9h-ctx4096",
   "model": "Qwen3.6-35B-A3B-MAYA",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: cnururx8, l8nska9k, seb6fkgp; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 870.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-MAYA-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 870.1 65.8 24.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sdzn8v9h",
   "source_url": "https://omlx.ai/benchmarks/performance/sdzn8v9h",
   "tg_tokens": null,
   "tps": 65.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-sdzn8v9h-ctx16384",
   "model": "Qwen3.6-35B-A3B-MAYA",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: cnururx8, l8nska9k, seb6fkgp; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 703.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-MAYA-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k 703.7 79.5 26.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sdzn8v9h",
   "source_url": "https://omlx.ai/benchmarks/performance/sdzn8v9h",
   "tg_tokens": null,
   "tps": 79.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-sdzn8v9h-ctx65536",
   "model": "Qwen3.6-35B-A3B-MAYA",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: cnururx8, l8nska9k, seb6fkgp; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 401.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-MAYA-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 64k 401.1 45.7 27.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sdzn8v9h",
   "source_url": "https://omlx.ai/benchmarks/performance/sdzn8v9h",
   "tg_tokens": null,
   "tps": 45.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-sdzn8v9h-ctx199680",
   "model": "Qwen3.6-35B-A3B-MAYA",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.6 GB at this context; duplicate traces of the same session: cnururx8, l8nska9k, seb6fkgp; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 184.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-MAYA-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 195k 184.8 24.7 30.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sdzn8v9h",
   "source_url": "https://omlx.ai/benchmarks/performance/sdzn8v9h",
   "tg_tokens": null,
   "tps": 24.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-b6yer8xu-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 1d8yay6s, 3t2jeadz, q7qp9le5, wxo653jo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4708.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 4,708 143.2 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b6yer8xu",
   "source_url": "https://omlx.ai/benchmarks/performance/b6yer8xu",
   "tg_tokens": null,
   "tps": 143.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-b6yer8xu-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 1d8yay6s, 3t2jeadz, q7qp9le5, wxo653jo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4688.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 4,688 112.9 21.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b6yer8xu",
   "source_url": "https://omlx.ai/benchmarks/performance/b6yer8xu",
   "tg_tokens": null,
   "tps": 112.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-b6yer8xu-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 1d8yay6s, 3t2jeadz, q7qp9le5, wxo653jo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3243.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 3,243 93.6 21.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b6yer8xu",
   "source_url": "https://omlx.ai/benchmarks/performance/b6yer8xu",
   "tg_tokens": null,
   "tps": 93.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-b6yer8xu-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: 1d8yay6s, 3t2jeadz, q7qp9le5, wxo653jo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2233.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 2,233 69.8 23.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b6yer8xu",
   "source_url": "https://omlx.ai/benchmarks/performance/b6yer8xu",
   "tg_tokens": null,
   "tps": 69.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-b6yer8xu-ctx199680",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: 1d8yay6s, 3t2jeadz, q7qp9le5, wxo653jo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1737.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 195k 1,737 56.2 24.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b6yer8xu",
   "source_url": "https://omlx.ai/benchmarks/performance/b6yer8xu",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-qduc8af7-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: 08ylkv7p, jkoc90qi; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 106.8 7.0 17.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qduc8af7",
   "source_url": "https://omlx.ai/benchmarks/performance/qduc8af7",
   "tg_tokens": null,
   "tps": 7.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-qduc8af7-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 08ylkv7p, jkoc90qi; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 104.8 6.7 18.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qduc8af7",
   "source_url": "https://omlx.ai/benchmarks/performance/qduc8af7",
   "tg_tokens": null,
   "tps": 6.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-qduc8af7-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 08ylkv7p, jkoc90qi; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 100.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 100.5 6.1 20.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qduc8af7",
   "source_url": "https://omlx.ai/benchmarks/performance/qduc8af7",
   "tg_tokens": null,
   "tps": 6.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-plh5xjgl-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 24.8 GB; TTFT 236725 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 69.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k tokens 69.2 PP tok/s 16.8 TG tok/s 236725 TTFT (ms) 24.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/plh5xjgl",
   "source_url": "https://omlx.ai/benchmarks/performance/plh5xjgl",
   "tg_tokens": null,
   "tps": 16.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-7n1wq2tk-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 21.9 GB; TTFT 228646 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 71.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k tokens 71.7 PP tok/s 13.7 TG tok/s 228646 TTFT (ms) 21.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7n1wq2tk",
   "source_url": "https://omlx.ai/benchmarks/performance/7n1wq2tk",
   "tg_tokens": null,
   "tps": 13.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ex2k6odd-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: rzkdaiaz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 313.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 313.4 49.2 33.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ex2k6odd",
   "source_url": "https://omlx.ai/benchmarks/performance/ex2k6odd",
   "tg_tokens": null,
   "tps": 49.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ex2k6odd-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: rzkdaiaz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 326.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 326.9 46.3 37.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ex2k6odd",
   "source_url": "https://omlx.ai/benchmarks/performance/ex2k6odd",
   "tg_tokens": null,
   "tps": 46.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-fvpyn5lb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: kynwxl3k; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 313.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 313.7 45.8 33.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fvpyn5lb",
   "source_url": "https://omlx.ai/benchmarks/performance/fvpyn5lb",
   "tg_tokens": null,
   "tps": 45.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-fvpyn5lb-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: kynwxl3k; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 327.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 327.9 43.3 37.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fvpyn5lb",
   "source_url": "https://omlx.ai/benchmarks/performance/fvpyn5lb",
   "tg_tokens": null,
   "tps": 43.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1535fojb-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: adszjthe; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 701.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 701.6 76.6 70.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1535fojb",
   "source_url": "https://omlx.ai/benchmarks/performance/1535fojb",
   "tg_tokens": null,
   "tps": 76.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1535fojb-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: adszjthe; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 861.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 861.1 53.4 71.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1535fojb",
   "source_url": "https://omlx.ai/benchmarks/performance/1535fojb",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-k6wnverc-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 20.6 GB; TTFT 78427 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 417.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k tokens 417.8 PP tok/s 22.7 TG tok/s 78427 TTFT (ms) 20.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k6wnverc",
   "source_url": "https://omlx.ai/benchmarks/performance/k6wnverc",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-qr7uf3j5-ctx1024",
   "model": "CoPaw-Flash-9B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 7.1 GB; TTFT 4941 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 207.2,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "CoPaw-Flash-9B-oQ4 on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k tokens 207.2 PP tok/s 26.6 TG tok/s 4941 TTFT (ms) 7.1 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qr7uf3j5",
   "source_url": "https://omlx.ai/benchmarks/performance/qr7uf3j5",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ohe8gxfm-ctx131072",
   "model": "Qwen3.6-35B-A3B-MAYA",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 28.9 GB; TTFT 508711 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 257.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-MAYA-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 128k tokens 257.7 PP tok/s 32.9 TG tok/s 508711 TTFT (ms) 28.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ohe8gxfm",
   "source_url": "https://omlx.ai/benchmarks/performance/ohe8gxfm",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-wx443spk-ctx4096",
   "model": "Qwen3.6-35B-A3B-MXFP8-CRACK",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 41.7 GB at this context; duplicate traces of the same session: 096rc0hv, f28jtmw3, sin6dh0h, wo6cr4ms; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1235.0,
   "provenance": "community",
   "quant": "MXFP8",
   "quote": "Qwen3.6-35B-A3B-MXFP8-CRACK-MTP on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 unknown; 4k 1,235 56.4 41.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wx443spk",
   "source_url": "https://omlx.ai/benchmarks/performance/wx443spk",
   "tg_tokens": null,
   "tps": 56.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-wx443spk-ctx8192",
   "model": "Qwen3.6-35B-A3B-MXFP8-CRACK",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 42.9 GB at this context; duplicate traces of the same session: 096rc0hv, f28jtmw3, sin6dh0h, wo6cr4ms; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1173.0,
   "provenance": "community",
   "quant": "MXFP8",
   "quote": "Qwen3.6-35B-A3B-MXFP8-CRACK-MTP on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 1,173 55.0 42.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wx443spk",
   "source_url": "https://omlx.ai/benchmarks/performance/wx443spk",
   "tg_tokens": null,
   "tps": 55.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-wx443spk-ctx16384",
   "model": "Qwen3.6-35B-A3B-MXFP8-CRACK",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 41.8 GB at this context; duplicate traces of the same session: 096rc0hv, f28jtmw3, sin6dh0h, wo6cr4ms; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 970.7,
   "provenance": "community",
   "quant": "MXFP8",
   "quote": "Qwen3.6-35B-A3B-MXFP8-CRACK-MTP on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 970.7 55.0 41.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wx443spk",
   "source_url": "https://omlx.ai/benchmarks/performance/wx443spk",
   "tg_tokens": null,
   "tps": 55.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-wx443spk-ctx32768",
   "model": "Qwen3.6-35B-A3B-MXFP8-CRACK",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 42.2 GB at this context; duplicate traces of the same session: 096rc0hv, f28jtmw3, sin6dh0h, wo6cr4ms; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 537.0,
   "provenance": "community",
   "quant": "MXFP8",
   "quote": "Qwen3.6-35B-A3B-MXFP8-CRACK-MTP on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 unknown; 32k 537.0 31.3 42.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wx443spk",
   "source_url": "https://omlx.ai/benchmarks/performance/wx443spk",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-wx443spk-ctx65536",
   "model": "Qwen3.6-35B-A3B-MXFP8-CRACK",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.7 GB at this context; duplicate traces of the same session: 096rc0hv, f28jtmw3, sin6dh0h, wo6cr4ms; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 212.1,
   "provenance": "community",
   "quant": "MXFP8",
   "quote": "Qwen3.6-35B-A3B-MXFP8-CRACK-MTP on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 unknown; 64k 212.1 14.8 43.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wx443spk",
   "source_url": "https://omlx.ai/benchmarks/performance/wx443spk",
   "tg_tokens": null,
   "tps": 14.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-8gsajfyn-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 84.7 GB at this context; duplicate traces of the same session: wupgb5g4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 608.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 608.2 63.6 84.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8gsajfyn",
   "source_url": "https://omlx.ai/benchmarks/performance/8gsajfyn",
   "tg_tokens": null,
   "tps": 63.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-8gsajfyn-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 85.2 GB at this context; duplicate traces of the same session: wupgb5g4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 725.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 725.4 51.1 85.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8gsajfyn",
   "source_url": "https://omlx.ai/benchmarks/performance/8gsajfyn",
   "tg_tokens": null,
   "tps": 51.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ojag8j39-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace, single run at 4k context; peak memory 85.3 GB; TTFT 7867 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 520.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k tokens 520.7 PP tok/s 47.5 TG tok/s 7867 TTFT (ms) 85.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ojag8j39",
   "source_url": "https://omlx.ai/benchmarks/performance/ojag8j39",
   "tg_tokens": null,
   "tps": 47.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-y0ydfsd2-ctx65536",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 22.9 GB; TTFT 181658 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 360.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k tokens 360.8 PP tok/s 18.8 TG tok/s 181658 TTFT (ms) 22.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y0ydfsd2",
   "source_url": "https://omlx.ai/benchmarks/performance/y0ydfsd2",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-0ggl3inj-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: 2qt4uprs, 3agadt1l; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 278.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 5bit; 1k 278.0 27.1 24.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ggl3inj",
   "source_url": "https://omlx.ai/benchmarks/performance/0ggl3inj",
   "tg_tokens": null,
   "tps": 27.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-0ggl3inj-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: 2qt4uprs, 3agadt1l; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 317.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 5bit; 4k 317.7 39.6 27.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ggl3inj",
   "source_url": "https://omlx.ai/benchmarks/performance/0ggl3inj",
   "tg_tokens": null,
   "tps": 39.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-0ggl3inj-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.7 GB at this context; duplicate traces of the same session: 2qt4uprs, 3agadt1l; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 313.6,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 5bit; 8k 313.6 44.2 28.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ggl3inj",
   "source_url": "https://omlx.ai/benchmarks/performance/0ggl3inj",
   "tg_tokens": null,
   "tps": 44.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-ebz6xclw-ctx1024",
   "model": "CoPaw-Flash-9B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 6.8 GB; TTFT 4799 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 213.4,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "CoPaw-Flash-9B-oQ4 on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k tokens 213.4 PP tok/s 32.4 TG tok/s 4799 TTFT (ms) 6.8 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ebz6xclw",
   "source_url": "https://omlx.ai/benchmarks/performance/ebz6xclw",
   "tg_tokens": null,
   "tps": 32.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-alzfgjga-ctx1024",
   "model": "CoPaw-Flash-9B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 6.8 GB; TTFT 4694 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 218.1,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "CoPaw-Flash-9B-oQ4 on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k tokens 218.1 PP tok/s 32.7 TG tok/s 4694 TTFT (ms) 6.8 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/alzfgjga",
   "source_url": "https://omlx.ai/benchmarks/performance/alzfgjga",
   "tg_tokens": null,
   "tps": 32.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-exw92err-ctx1024",
   "model": "CoPaw-Flash-9B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 6.9 GB; TTFT 4811 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 212.9,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "CoPaw-Flash-9B-oQ4 on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k tokens 212.9 PP tok/s 29.1 TG tok/s 4811 TTFT (ms) 6.9 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/exw92err",
   "source_url": "https://omlx.ai/benchmarks/performance/exw92err",
   "tg_tokens": null,
   "tps": 29.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-ahx88qku-ctx1024",
   "model": "CoPaw-Flash-9B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 6.9 GB; TTFT 4807 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 213.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "CoPaw-Flash-9B-oQ4 on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k tokens 213.0 PP tok/s 29.3 TG tok/s 4807 TTFT (ms) 6.9 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ahx88qku",
   "source_url": "https://omlx.ai/benchmarks/performance/ahx88qku",
   "tg_tokens": null,
   "tps": 29.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-f8t5g42x-ctx1024",
   "model": "CoPaw-Flash-9B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 7.1 GB; TTFT 4755 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 215.3,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "CoPaw-Flash-9B-oQ4 on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k tokens 215.3 PP tok/s 29.9 TG tok/s 4755 TTFT (ms) 7.1 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f8t5g42x",
   "source_url": "https://omlx.ai/benchmarks/performance/f8t5g42x",
   "tg_tokens": null,
   "tps": 29.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-0nyt1aa3-ctx1024",
   "model": "CoPaw-Flash-9B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 7.1 GB; TTFT 4705 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 217.6,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "CoPaw-Flash-9B-oQ4 on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k tokens 217.6 PP tok/s 30.5 TG tok/s 4705 TTFT (ms) 7.1 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0nyt1aa3",
   "source_url": "https://omlx.ai/benchmarks/performance/0nyt1aa3",
   "tg_tokens": null,
   "tps": 30.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-drefic02-ctx1024",
   "model": "CoPaw-Flash-9B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 6.8 GB; TTFT 4675 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 219.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "CoPaw-Flash-9B-oQ4 on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k tokens 219.0 PP tok/s 32.6 TG tok/s 4675 TTFT (ms) 6.8 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/drefic02",
   "source_url": "https://omlx.ai/benchmarks/performance/drefic02",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-gf77r6qv-ctx1024",
   "model": "CoPaw-Flash-9B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 6.8 GB; TTFT 4669 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 219.3,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "CoPaw-Flash-9B-oQ4 on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k tokens 219.3 PP tok/s 34.2 TG tok/s 4669 TTFT (ms) 6.8 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gf77r6qv",
   "source_url": "https://omlx.ai/benchmarks/performance/gf77r6qv",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-0jnaorfi-ctx1024",
   "model": "CoPaw-Flash-9B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 6.8 GB; TTFT 4670 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 219.3,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "CoPaw-Flash-9B-oQ4 on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k tokens 219.3 PP tok/s 33.5 TG tok/s 4670 TTFT (ms) 6.8 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0jnaorfi",
   "source_url": "https://omlx.ai/benchmarks/performance/0jnaorfi",
   "tg_tokens": null,
   "tps": 33.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-xu85dte9-ctx1024",
   "model": "CoPaw-Flash-9B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 7.1 GB; TTFT 4675 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 219.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "CoPaw-Flash-9B-oQ4 on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k tokens 219.0 PP tok/s 29.3 TG tok/s 4675 TTFT (ms) 7.1 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xu85dte9",
   "source_url": "https://omlx.ai/benchmarks/performance/xu85dte9",
   "tg_tokens": null,
   "tps": 29.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-qqykcrll-ctx1024",
   "model": "CoPaw-Flash-9B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 6.8 GB; TTFT 4648 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.3,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "CoPaw-Flash-9B-oQ4 on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k tokens 220.3 PP tok/s 34.9 TG tok/s 4648 TTFT (ms) 6.8 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qqykcrll",
   "source_url": "https://omlx.ai/benchmarks/performance/qqykcrll",
   "tg_tokens": null,
   "tps": 34.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-fqp78x92-ctx1024",
   "model": "CoPaw-Flash-9B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 6.9 GB; TTFT 4674 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 219.1,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "CoPaw-Flash-9B-oQ4 on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k tokens 219.1 PP tok/s 31.2 TG tok/s 4674 TTFT (ms) 6.9 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fqp78x92",
   "source_url": "https://omlx.ai/benchmarks/performance/fqp78x92",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-dsbqqttk-ctx1024",
   "model": "CoPaw-Flash-9B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 6.8 GB; TTFT 4659 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 219.8,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "CoPaw-Flash-9B-oQ4 on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k tokens 219.8 PP tok/s 36.9 TG tok/s 4659 TTFT (ms) 6.8 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dsbqqttk",
   "source_url": "https://omlx.ai/benchmarks/performance/dsbqqttk",
   "tg_tokens": null,
   "tps": 36.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-qqr5cvs8-ctx1024",
   "model": "CoPaw-Flash-9B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 7.1 GB; TTFT 4662 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 219.6,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "CoPaw-Flash-9B-oQ4 on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k tokens 219.6 PP tok/s 31.1 TG tok/s 4662 TTFT (ms) 7.1 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qqr5cvs8",
   "source_url": "https://omlx.ai/benchmarks/performance/qqr5cvs8",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-ou9oftjv-ctx1024",
   "model": "CoPaw-Flash-9B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 6.8 GB; TTFT 4641 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.6,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "CoPaw-Flash-9B-oQ4 on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k tokens 220.6 PP tok/s 33.0 TG tok/s 4641 TTFT (ms) 6.8 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ou9oftjv",
   "source_url": "https://omlx.ai/benchmarks/performance/ou9oftjv",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-1eq8zycm-ctx1024",
   "model": "CoPaw-Flash-9B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 7.1 GB; TTFT 4772 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 214.6,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "CoPaw-Flash-9B-oQ4 on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k tokens 214.6 PP tok/s 28.3 TG tok/s 4772 TTFT (ms) 7.1 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1eq8zycm",
   "source_url": "https://omlx.ai/benchmarks/performance/1eq8zycm",
   "tg_tokens": null,
   "tps": 28.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-wmk6wbjc-ctx1024",
   "model": "CoPaw-Flash-9B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 7.1 GB; TTFT 4663 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 219.6,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "CoPaw-Flash-9B-oQ4 on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k tokens 219.6 PP tok/s 31.0 TG tok/s 4663 TTFT (ms) 7.1 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wmk6wbjc",
   "source_url": "https://omlx.ai/benchmarks/performance/wmk6wbjc",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-vt1hid7d-ctx1024",
   "model": "CoPaw-Flash-9B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 6.9 GB; TTFT 4630 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 221.2,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "CoPaw-Flash-9B-oQ4 on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k tokens 221.2 PP tok/s 34.6 TG tok/s 4630 TTFT (ms) 6.9 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vt1hid7d",
   "source_url": "https://omlx.ai/benchmarks/performance/vt1hid7d",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-t44qmyt8-ctx1024",
   "model": "Qwen3.6-35B-A3B-Holo3-Qwopus-AgentWorld",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 29.5 GB; TTFT 743 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1378.0,
   "provenance": "community",
   "quant": "qx64 hi",
   "quote": "Qwen3.6-35B-A3B-Holo3-Qwopus-AgentWorld-qx64-hi-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k tokens 1,378 PP tok/s 74.1 TG tok/s 743 TTFT (ms) 29.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t44qmyt8",
   "source_url": "https://omlx.ai/benchmarks/performance/t44qmyt8",
   "tg_tokens": null,
   "tps": 74.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ipyksjwr-ctx1024",
   "model": "Qwen3.6-35B-A3B-Holo3-Qwopus-AgentWorld",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 29.5 GB; TTFT 499 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2054.0,
   "provenance": "community",
   "quant": "qx64 hi",
   "quote": "Qwen3.6-35B-A3B-Holo3-Qwopus-AgentWorld-qx64-hi-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k tokens 2,054 PP tok/s 80.6 TG tok/s 499 TTFT (ms) 29.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ipyksjwr",
   "source_url": "https://omlx.ai/benchmarks/performance/ipyksjwr",
   "tg_tokens": null,
   "tps": 80.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ju7e1999-ctx1024",
   "model": "Qwen3.6-35B-A3B-Holo3-Qwopus-AgentWorld",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 29.5 GB; TTFT 475 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2154.0,
   "provenance": "community",
   "quant": "qx64 hi",
   "quote": "Qwen3.6-35B-A3B-Holo3-Qwopus-AgentWorld-qx64-hi-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k tokens 2,154 PP tok/s 75.9 TG tok/s 475 TTFT (ms) 29.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ju7e1999",
   "source_url": "https://omlx.ai/benchmarks/performance/ju7e1999",
   "tg_tokens": null,
   "tps": 75.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-7rhftcbu-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: wygk3dmt; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 114.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 1k 114.9 13.0 34.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7rhftcbu",
   "source_url": "https://omlx.ai/benchmarks/performance/7rhftcbu",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-7rhftcbu-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 40.1 GB at this context; duplicate traces of the same session: wygk3dmt; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 114.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 8k 114.9 15.9 40.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7rhftcbu",
   "source_url": "https://omlx.ai/benchmarks/performance/7rhftcbu",
   "tg_tokens": null,
   "tps": 15.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-a2uj5die-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 86.4 GB at this context; duplicate traces of the same session: rkyhqwa5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 586.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 586.0 69.3 86.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a2uj5die",
   "source_url": "https://omlx.ai/benchmarks/performance/a2uj5die",
   "tg_tokens": null,
   "tps": 69.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-a2uj5die-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 87.0 GB at this context; duplicate traces of the same session: rkyhqwa5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 661.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 661.0 61.9 87.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a2uj5die",
   "source_url": "https://omlx.ai/benchmarks/performance/a2uj5die",
   "tg_tokens": null,
   "tps": 61.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-v03v3sdr-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 86.8 GB at this context; duplicate traces of the same session: diwmflcm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1391.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 1,391 66.0 86.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v03v3sdr",
   "source_url": "https://omlx.ai/benchmarks/performance/v03v3sdr",
   "tg_tokens": null,
   "tps": 66.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-v03v3sdr-ctx16384",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 88.9 GB at this context; duplicate traces of the same session: diwmflcm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1927.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 1,927 61.5 88.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v03v3sdr",
   "source_url": "https://omlx.ai/benchmarks/performance/v03v3sdr",
   "tg_tokens": null,
   "tps": 61.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-b69181kt-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: 4ewnz8vj, fnnjnkds, l7r102xd, szf8tc4m; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 915.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 915.7 79.8 100.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b69181kt",
   "source_url": "https://omlx.ai/benchmarks/performance/b69181kt",
   "tg_tokens": null,
   "tps": 79.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-b69181kt-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: 4ewnz8vj, fnnjnkds, l7r102xd, szf8tc4m; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1081.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,081 72.3 101.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b69181kt",
   "source_url": "https://omlx.ai/benchmarks/performance/b69181kt",
   "tg_tokens": null,
   "tps": 72.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-b69181kt-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.9 GB at this context; duplicate traces of the same session: 4ewnz8vj, fnnjnkds, l7r102xd, szf8tc4m; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1070.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 1,070 72.0 101.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b69181kt",
   "source_url": "https://omlx.ai/benchmarks/performance/b69181kt",
   "tg_tokens": null,
   "tps": 72.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-b69181kt-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.3 GB at this context; duplicate traces of the same session: 4ewnz8vj, fnnjnkds, l7r102xd, szf8tc4m; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1057.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 1,057 78.4 103.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b69181kt",
   "source_url": "https://omlx.ai/benchmarks/performance/b69181kt",
   "tg_tokens": null,
   "tps": 78.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-b69181kt-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 105.8 GB at this context; duplicate traces of the same session: 4ewnz8vj, fnnjnkds, l7r102xd, szf8tc4m; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1044.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 1,044 67.1 105.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b69181kt",
   "source_url": "https://omlx.ai/benchmarks/performance/b69181kt",
   "tg_tokens": null,
   "tps": 67.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-7juwc6fb-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.1 GB at this context; duplicate traces of the same session: 0rufjxhv, d2glp07k, xhqis8bc, y6fyim7w; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 905.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 905.7 82.4 99.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7juwc6fb",
   "source_url": "https://omlx.ai/benchmarks/performance/7juwc6fb",
   "tg_tokens": null,
   "tps": 82.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-7juwc6fb-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.6 GB at this context; duplicate traces of the same session: 0rufjxhv, d2glp07k, xhqis8bc, y6fyim7w; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1130.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 1,130 76.3 99.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7juwc6fb",
   "source_url": "https://omlx.ai/benchmarks/performance/7juwc6fb",
   "tg_tokens": null,
   "tps": 76.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-7juwc6fb-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.8 GB at this context; duplicate traces of the same session: 0rufjxhv, d2glp07k, xhqis8bc, y6fyim7w; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1114.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,114 76.0 99.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7juwc6fb",
   "source_url": "https://omlx.ai/benchmarks/performance/7juwc6fb",
   "tg_tokens": null,
   "tps": 76.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-7juwc6fb-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.3 GB at this context; duplicate traces of the same session: 0rufjxhv, d2glp07k, xhqis8bc, y6fyim7w; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1103.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 1,103 75.4 100.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7juwc6fb",
   "source_url": "https://omlx.ai/benchmarks/performance/7juwc6fb",
   "tg_tokens": null,
   "tps": 75.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-7juwc6fb-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.1 GB at this context; duplicate traces of the same session: 0rufjxhv, d2glp07k, xhqis8bc, y6fyim7w; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1078.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 1,078 69.1 103.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7juwc6fb",
   "source_url": "https://omlx.ai/benchmarks/performance/7juwc6fb",
   "tg_tokens": null,
   "tps": 69.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-9lvtb0xs-ctx65536",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 23.2 GB; TTFT 128012 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 512.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k tokens 512.0 PP tok/s 21.7 TG tok/s 128012 TTFT (ms) 23.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9lvtb0xs",
   "source_url": "https://omlx.ai/benchmarks/performance/9lvtb0xs",
   "tg_tokens": null,
   "tps": 21.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-f08sabt4-ctx16384",
   "model": "gemma-3-12b-it",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 8.6 GB; TTFT 1325835 ms; duplicate traces of the same session: none; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 12.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-3-12b-it on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 4bit; 16k tokens 12.4 PP tok/s 1.6 TG tok/s 1325835 TTFT (ms) 8.6 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f08sabt4",
   "source_url": "https://omlx.ai/benchmarks/performance/f08sabt4",
   "tg_tokens": null,
   "tps": 1.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-913i9tyf-ctx1024",
   "model": "CoPaw-Flash-9B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 6.8 GB; TTFT 4793 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 213.6,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "CoPaw-Flash-9B-oQ4 on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k tokens 213.6 PP tok/s 32.5 TG tok/s 4793 TTFT (ms) 6.8 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/913i9tyf",
   "source_url": "https://omlx.ai/benchmarks/performance/913i9tyf",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-fzhyjx8l-ctx1024",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: 7jrop3qg, lu6znoyk, tmd9usat, wpj4v8xw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1160.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-oQ8-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 1,160 51.6 37.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fzhyjx8l",
   "source_url": "https://omlx.ai/benchmarks/performance/fzhyjx8l",
   "tg_tokens": null,
   "tps": 51.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-fzhyjx8l-ctx4096",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.5 GB at this context; duplicate traces of the same session: 7jrop3qg, lu6znoyk, tmd9usat, wpj4v8xw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1682.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-oQ8-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 1,682 75.9 39.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fzhyjx8l",
   "source_url": "https://omlx.ai/benchmarks/performance/fzhyjx8l",
   "tg_tokens": null,
   "tps": 75.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-fzhyjx8l-ctx8192",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.7 GB at this context; duplicate traces of the same session: 7jrop3qg, lu6znoyk, tmd9usat, wpj4v8xw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1633.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-oQ8-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 1,633 99.5 39.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fzhyjx8l",
   "source_url": "https://omlx.ai/benchmarks/performance/fzhyjx8l",
   "tg_tokens": null,
   "tps": 99.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-fzhyjx8l-ctx16384",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 40.2 GB at this context; duplicate traces of the same session: 7jrop3qg, lu6znoyk, tmd9usat, wpj4v8xw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1337.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-oQ8-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 1,337 108.1 40.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fzhyjx8l",
   "source_url": "https://omlx.ai/benchmarks/performance/fzhyjx8l",
   "tg_tokens": null,
   "tps": 108.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-fzhyjx8l-ctx32768",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 41.0 GB at this context; duplicate traces of the same session: 7jrop3qg, lu6znoyk, tmd9usat, wpj4v8xw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1019.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-oQ8-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 1,019 101.8 41.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fzhyjx8l",
   "source_url": "https://omlx.ai/benchmarks/performance/fzhyjx8l",
   "tg_tokens": null,
   "tps": 101.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-fzhyjx8l-ctx65536",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 42.6 GB at this context; duplicate traces of the same session: 7jrop3qg, lu6znoyk, tmd9usat, wpj4v8xw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 385.7,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-oQ8-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 385.7 27.5 42.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fzhyjx8l",
   "source_url": "https://omlx.ai/benchmarks/performance/fzhyjx8l",
   "tg_tokens": null,
   "tps": 27.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-x4caw4nm-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 6l1srn18, 7nc29ea8, bhm2m42l; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1489.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 1,489 70.2 19.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x4caw4nm",
   "source_url": "https://omlx.ai/benchmarks/performance/x4caw4nm",
   "tg_tokens": null,
   "tps": 70.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-x4caw4nm-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 6l1srn18, 7nc29ea8, bhm2m42l; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1828.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,828 69.0 19.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x4caw4nm",
   "source_url": "https://omlx.ai/benchmarks/performance/x4caw4nm",
   "tg_tokens": null,
   "tps": 69.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-x4caw4nm-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 6l1srn18, 7nc29ea8, bhm2m42l; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1788.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 8k 1,788 66.4 20.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x4caw4nm",
   "source_url": "https://omlx.ai/benchmarks/performance/x4caw4nm",
   "tg_tokens": null,
   "tps": 66.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-x4caw4nm-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 6l1srn18, 7nc29ea8, bhm2m42l; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1667.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 1,667 61.0 20.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x4caw4nm",
   "source_url": "https://omlx.ai/benchmarks/performance/x4caw4nm",
   "tg_tokens": null,
   "tps": 61.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-oux5taui-ctx1024",
   "model": "Qwen3.8-9B-Distill-Uncensored-Heretic",
   "notes": "MTP build (mlx); oMLX community benchmark trace; peak memory 7.9 GB at this context; duplicate traces of the same session: m3e6fmfh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 449.3,
   "provenance": "community",
   "quant": "Q6G64",
   "quote": "Qwen3.8-9B-Distill-Uncensored-Heretic-MTPLX-Q6G64 on M4 Pro (20c); M4 Pro (20c) \u00b7 24 GB \u00b7 6bit; 1k 449.3 35.7 7.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oux5taui",
   "source_url": "https://omlx.ai/benchmarks/performance/oux5taui",
   "tg_tokens": null,
   "tps": 35.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-oux5taui-ctx4096",
   "model": "Qwen3.8-9B-Distill-Uncensored-Heretic",
   "notes": "MTP build (mlx); oMLX community benchmark trace; peak memory 8.4 GB at this context; duplicate traces of the same session: m3e6fmfh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 466.2,
   "provenance": "community",
   "quant": "Q6G64",
   "quote": "Qwen3.8-9B-Distill-Uncensored-Heretic-MTPLX-Q6G64 on M4 Pro (20c); M4 Pro (20c) \u00b7 24 GB \u00b7 6bit; 4k 466.2 35.0 8.4 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oux5taui",
   "source_url": "https://omlx.ai/benchmarks/performance/oux5taui",
   "tg_tokens": null,
   "tps": 35.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-j433h78i-ctx1024",
   "model": "Qwen3.8-9B-Distill-Uncensored-Heretic",
   "notes": "MTP build (mlx); oMLX community benchmark trace; peak memory 7.9 GB at this context; duplicate traces of the same session: azoppclz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 455.3,
   "provenance": "community",
   "quant": "Q6G64",
   "quote": "Qwen3.8-9B-Distill-Uncensored-Heretic-MTPLX-Q6G64 on M4 Pro (20c); M4 Pro (20c) \u00b7 24 GB \u00b7 6bit; 1k 455.3 34.7 7.9 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j433h78i",
   "source_url": "https://omlx.ai/benchmarks/performance/j433h78i",
   "tg_tokens": null,
   "tps": 34.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-j433h78i-ctx4096",
   "model": "Qwen3.8-9B-Distill-Uncensored-Heretic",
   "notes": "MTP build (mlx); oMLX community benchmark trace; peak memory 8.4 GB at this context; duplicate traces of the same session: azoppclz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 467.1,
   "provenance": "community",
   "quant": "Q6G64",
   "quote": "Qwen3.8-9B-Distill-Uncensored-Heretic-MTPLX-Q6G64 on M4 Pro (20c); M4 Pro (20c) \u00b7 24 GB \u00b7 6bit; 4k 467.1 34.6 8.4 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j433h78i",
   "source_url": "https://omlx.ai/benchmarks/performance/j433h78i",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-z7pwit1e-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.8 GB at this context; duplicate traces of the same session: 1jtgzgp8, 24ieacog, ewj12get, yvsqg36h, za4ic58m; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 914.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 914.9 93.4 100.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z7pwit1e",
   "source_url": "https://omlx.ai/benchmarks/performance/z7pwit1e",
   "tg_tokens": null,
   "tps": 93.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-z7pwit1e-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: 1jtgzgp8, 24ieacog, ewj12get, yvsqg36h, za4ic58m; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1081.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,081 74.8 101.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z7pwit1e",
   "source_url": "https://omlx.ai/benchmarks/performance/z7pwit1e",
   "tg_tokens": null,
   "tps": 74.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-z7pwit1e-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.9 GB at this context; duplicate traces of the same session: 1jtgzgp8, 24ieacog, ewj12get, yvsqg36h, za4ic58m; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1070.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 1,070 80.3 101.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z7pwit1e",
   "source_url": "https://omlx.ai/benchmarks/performance/z7pwit1e",
   "tg_tokens": null,
   "tps": 80.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-z7pwit1e-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.3 GB at this context; duplicate traces of the same session: 1jtgzgp8, 24ieacog, ewj12get, yvsqg36h, za4ic58m; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1060.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 1,060 82.1 103.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z7pwit1e",
   "source_url": "https://omlx.ai/benchmarks/performance/z7pwit1e",
   "tg_tokens": null,
   "tps": 82.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-z7pwit1e-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 107.1 GB at this context; duplicate traces of the same session: 1jtgzgp8, 24ieacog, ewj12get, yvsqg36h, za4ic58m; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1042.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 1,042 71.0 107.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z7pwit1e",
   "source_url": "https://omlx.ai/benchmarks/performance/z7pwit1e",
   "tg_tokens": null,
   "tps": 71.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-z7pwit1e-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 113.7 GB at this context; duplicate traces of the same session: 1jtgzgp8, 24ieacog, ewj12get, yvsqg36h, za4ic58m; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1027.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 128k 1,027 57.2 113.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z7pwit1e",
   "source_url": "https://omlx.ai/benchmarks/performance/z7pwit1e",
   "tg_tokens": null,
   "tps": 57.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gocbis8b-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: 1hjnq5fs, 28vyt2ly, 2rwd5pzi, 68d8uym7, 6rc85rmu, y0txqyt5; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 917.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 917.5 91.0 100.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gocbis8b",
   "source_url": "https://omlx.ai/benchmarks/performance/gocbis8b",
   "tg_tokens": null,
   "tps": 91.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gocbis8b-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: 1hjnq5fs, 28vyt2ly, 2rwd5pzi, 68d8uym7, 6rc85rmu, y0txqyt5; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1095.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 1,095 97.1 101.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gocbis8b",
   "source_url": "https://omlx.ai/benchmarks/performance/gocbis8b",
   "tg_tokens": null,
   "tps": 97.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gocbis8b-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: 1hjnq5fs, 28vyt2ly, 2rwd5pzi, 68d8uym7, 6rc85rmu, y0txqyt5; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1079.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,079 75.8 101.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gocbis8b",
   "source_url": "https://omlx.ai/benchmarks/performance/gocbis8b",
   "tg_tokens": null,
   "tps": 75.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gocbis8b-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.9 GB at this context; duplicate traces of the same session: 1hjnq5fs, 28vyt2ly, 2rwd5pzi, 68d8uym7, 6rc85rmu, y0txqyt5; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1069.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 1,069 80.8 101.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gocbis8b",
   "source_url": "https://omlx.ai/benchmarks/performance/gocbis8b",
   "tg_tokens": null,
   "tps": 80.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gocbis8b-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.7 GB at this context; duplicate traces of the same session: 1hjnq5fs, 28vyt2ly, 2rwd5pzi, 68d8uym7, 6rc85rmu, y0txqyt5; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1060.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 1,060 110.7 103.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gocbis8b",
   "source_url": "https://omlx.ai/benchmarks/performance/gocbis8b",
   "tg_tokens": null,
   "tps": 110.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gocbis8b-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 106.7 GB at this context; duplicate traces of the same session: 1hjnq5fs, 28vyt2ly, 2rwd5pzi, 68d8uym7, 6rc85rmu, y0txqyt5; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1045.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 1,045 68.5 106.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gocbis8b",
   "source_url": "https://omlx.ai/benchmarks/performance/gocbis8b",
   "tg_tokens": null,
   "tps": 68.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gocbis8b-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 113.1 GB at this context; duplicate traces of the same session: 1hjnq5fs, 28vyt2ly, 2rwd5pzi, 68d8uym7, 6rc85rmu, y0txqyt5; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1028.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 128k 1,028 55.9 113.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gocbis8b",
   "source_url": "https://omlx.ai/benchmarks/performance/gocbis8b",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-vljxaq27-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.7 GB at this context; duplicate traces of the same session: 0eeqzs6d, 35prkd58, dbujd7qv, elba2do6, rb358tgh; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 918.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 918.0 71.6 100.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vljxaq27",
   "source_url": "https://omlx.ai/benchmarks/performance/vljxaq27",
   "tg_tokens": null,
   "tps": 71.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-vljxaq27-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: 0eeqzs6d, 35prkd58, dbujd7qv, elba2do6, rb358tgh; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1093.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 1,093 68.7 101.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vljxaq27",
   "source_url": "https://omlx.ai/benchmarks/performance/vljxaq27",
   "tg_tokens": null,
   "tps": 68.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-vljxaq27-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: 0eeqzs6d, 35prkd58, dbujd7qv, elba2do6, rb358tgh; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1081.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,081 77.5 101.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vljxaq27",
   "source_url": "https://omlx.ai/benchmarks/performance/vljxaq27",
   "tg_tokens": null,
   "tps": 77.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-vljxaq27-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.9 GB at this context; duplicate traces of the same session: 0eeqzs6d, 35prkd58, dbujd7qv, elba2do6, rb358tgh; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1071.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 1,071 79.5 101.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vljxaq27",
   "source_url": "https://omlx.ai/benchmarks/performance/vljxaq27",
   "tg_tokens": null,
   "tps": 79.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-vljxaq27-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.0 GB at this context; duplicate traces of the same session: 0eeqzs6d, 35prkd58, dbujd7qv, elba2do6, rb358tgh; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1061.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 1,061 85.2 103.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vljxaq27",
   "source_url": "https://omlx.ai/benchmarks/performance/vljxaq27",
   "tg_tokens": null,
   "tps": 85.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vljxaq27-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 106.1 GB at this context; duplicate traces of the same session: 0eeqzs6d, 35prkd58, dbujd7qv, elba2do6, rb358tgh; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1045.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 1,045 71.1 106.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vljxaq27",
   "source_url": "https://omlx.ai/benchmarks/performance/vljxaq27",
   "tg_tokens": null,
   "tps": 71.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rs99bg1o-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 914.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 914.5 79.1 100.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rs99bg1o",
   "source_url": "https://omlx.ai/benchmarks/performance/rs99bg1o",
   "tg_tokens": null,
   "tps": 79.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rs99bg1o-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1080.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,080 65.7 101.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rs99bg1o",
   "source_url": "https://omlx.ai/benchmarks/performance/rs99bg1o",
   "tg_tokens": null,
   "tps": 65.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rs99bg1o-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.9 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1070.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 1,070 74.7 101.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rs99bg1o",
   "source_url": "https://omlx.ai/benchmarks/performance/rs99bg1o",
   "tg_tokens": null,
   "tps": 74.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rs99bg1o-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.9 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1060.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 1,060 87.2 102.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rs99bg1o",
   "source_url": "https://omlx.ai/benchmarks/performance/rs99bg1o",
   "tg_tokens": null,
   "tps": 87.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rs99bg1o-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 106.0 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1045.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 1,045 69.8 106.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rs99bg1o",
   "source_url": "https://omlx.ai/benchmarks/performance/rs99bg1o",
   "tg_tokens": null,
   "tps": 69.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0lbu4gs4-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.2 GB at this context; duplicate traces of the same session: s6xavyp1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2421.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,421 84.7 71.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0lbu4gs4",
   "source_url": "https://omlx.ai/benchmarks/performance/0lbu4gs4",
   "tg_tokens": null,
   "tps": 84.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0lbu4gs4-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.5 GB at this context; duplicate traces of the same session: s6xavyp1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2304.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,304 63.0 79.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0lbu4gs4",
   "source_url": "https://omlx.ai/benchmarks/performance/0lbu4gs4",
   "tg_tokens": null,
   "tps": 63.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6mng2o03-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.8 GB at this context; duplicate traces of the same session: scj36u63; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1755.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,755 72.7 100.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6mng2o03",
   "source_url": "https://omlx.ai/benchmarks/performance/6mng2o03",
   "tg_tokens": null,
   "tps": 72.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6mng2o03-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.7 GB at this context; duplicate traces of the same session: scj36u63; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2279.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,279 78.5 100.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6mng2o03",
   "source_url": "https://omlx.ai/benchmarks/performance/6mng2o03",
   "tg_tokens": null,
   "tps": 78.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-dl9ukkhs-ctx1024",
   "model": "Qwythos-9B-v3-Claude-Mythos-5-1M",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 12.2 GB at this context; duplicate traces of the same session: 3chzgn1e; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 368.1,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwythos-9B-v3-Claude-Mythos-5-1M-MLX-oQ8-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 1k 368.1 44.5 12.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dl9ukkhs",
   "source_url": "https://omlx.ai/benchmarks/performance/dl9ukkhs",
   "tg_tokens": null,
   "tps": 44.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-dl9ukkhs-ctx4096",
   "model": "Qwythos-9B-v3-Claude-Mythos-5-1M",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 12.9 GB at this context; duplicate traces of the same session: 3chzgn1e; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 372.6,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwythos-9B-v3-Claude-Mythos-5-1M-MLX-oQ8-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 4k 372.6 34.8 12.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dl9ukkhs",
   "source_url": "https://omlx.ai/benchmarks/performance/dl9ukkhs",
   "tg_tokens": null,
   "tps": 34.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-i9p9v8cs-ctx1024",
   "model": "Qwythos-9B-v3-Claude-Mythos-5-1M",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 12.2 GB at this context; duplicate traces of the same session: 19w1750k; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 416.8,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwythos-9B-v3-Claude-Mythos-5-1M-MLX-oQ8-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 1k 416.8 40.5 12.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i9p9v8cs",
   "source_url": "https://omlx.ai/benchmarks/performance/i9p9v8cs",
   "tg_tokens": null,
   "tps": 40.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-i9p9v8cs-ctx4096",
   "model": "Qwythos-9B-v3-Claude-Mythos-5-1M",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 12.9 GB at this context; duplicate traces of the same session: 19w1750k; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 384.9,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwythos-9B-v3-Claude-Mythos-5-1M-MLX-oQ8-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 4k 384.9 36.2 12.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i9p9v8cs",
   "source_url": "https://omlx.ai/benchmarks/performance/i9p9v8cs",
   "tg_tokens": null,
   "tps": 36.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mohcslm0-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "SamQuant C58; oMLX community benchmark trace; peak memory 62.3 GB at this context; duplicate traces of the same session: 0l1utdan; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 966.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "SamQuant-Qwen3.8-Flash-Next-Uncensored-oMLX-C58 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 966.5 56.1 62.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mohcslm0",
   "source_url": "https://omlx.ai/benchmarks/performance/mohcslm0",
   "tg_tokens": null,
   "tps": 56.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mohcslm0-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "SamQuant C58; oMLX community benchmark trace; peak memory 62.6 GB at this context; duplicate traces of the same session: 0l1utdan; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1524.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "SamQuant-Qwen3.8-Flash-Next-Uncensored-oMLX-C58 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,524 60.4 62.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mohcslm0",
   "source_url": "https://omlx.ai/benchmarks/performance/mohcslm0",
   "tg_tokens": null,
   "tps": 60.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0oqbb737-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "SamQuant C58; oMLX community benchmark trace, single run at 1k context; peak memory 62.3 GB; TTFT 921 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1112.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "SamQuant-Qwen3.8-Flash-Next-Uncensored-oMLX-C58 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,112 PP tok/s 52.1 TG tok/s 921 TTFT (ms) 62.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0oqbb737",
   "source_url": "https://omlx.ai/benchmarks/performance/0oqbb737",
   "tg_tokens": null,
   "tps": 52.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9qkp71b8-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "SamQuant C58; oMLX community benchmark trace; peak memory 59.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 561.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "SamQuant-Qwen3.8-Flash-Next-Uncensored-oMLX-C58 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 561.7 43.5 59.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9qkp71b8",
   "source_url": "https://omlx.ai/benchmarks/performance/9qkp71b8",
   "tg_tokens": null,
   "tps": 43.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9qkp71b8-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "SamQuant C58; oMLX community benchmark trace; peak memory 60.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1049.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "SamQuant-Qwen3.8-Flash-Next-Uncensored-oMLX-C58 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,049 38.6 60.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9qkp71b8",
   "source_url": "https://omlx.ai/benchmarks/performance/9qkp71b8",
   "tg_tokens": null,
   "tps": 38.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-bhbrkkwd-ctx1024",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1160.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-oQ8-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 1,160 51.6 37.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhbrkkwd",
   "source_url": "https://omlx.ai/benchmarks/performance/bhbrkkwd",
   "tg_tokens": null,
   "tps": 51.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-bhbrkkwd-ctx4096",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1682.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-oQ8-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 1,682 75.9 39.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhbrkkwd",
   "source_url": "https://omlx.ai/benchmarks/performance/bhbrkkwd",
   "tg_tokens": null,
   "tps": 75.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-bhbrkkwd-ctx8192",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1633.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-oQ8-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 1,633 99.5 39.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhbrkkwd",
   "source_url": "https://omlx.ai/benchmarks/performance/bhbrkkwd",
   "tg_tokens": null,
   "tps": 99.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-bhbrkkwd-ctx16384",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 40.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1337.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-oQ8-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 1,337 108.1 40.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhbrkkwd",
   "source_url": "https://omlx.ai/benchmarks/performance/bhbrkkwd",
   "tg_tokens": null,
   "tps": 108.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-bhbrkkwd-ctx32768",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 41.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1019.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-oQ8-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 1,019 101.8 41.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhbrkkwd",
   "source_url": "https://omlx.ai/benchmarks/performance/bhbrkkwd",
   "tg_tokens": null,
   "tps": 101.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-bhbrkkwd-ctx65536",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 42.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 385.7,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-oQ8-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 385.7 27.5 42.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhbrkkwd",
   "source_url": "https://omlx.ai/benchmarks/performance/bhbrkkwd",
   "tg_tokens": null,
   "tps": 27.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-y4nbc2zt-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: ajvquh89; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1528.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 1,528 72.5 19.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y4nbc2zt",
   "source_url": "https://omlx.ai/benchmarks/performance/y4nbc2zt",
   "tg_tokens": null,
   "tps": 72.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-y4nbc2zt-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: ajvquh89; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1906.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,906 71.1 19.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y4nbc2zt",
   "source_url": "https://omlx.ai/benchmarks/performance/y4nbc2zt",
   "tg_tokens": null,
   "tps": 71.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-7ds8wlab-ctx1024",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 13.5 GB; TTFT 9686 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.7,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.8-27B-ABLITERATED-mlx-q3km on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k tokens 105.7 PP tok/s 17.7 TG tok/s 9686 TTFT (ms) 13.5 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ds8wlab",
   "source_url": "https://omlx.ai/benchmarks/performance/7ds8wlab",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-tsbt77wv-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: daa346z5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 431.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 1k 431.9 12.7 22.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tsbt77wv",
   "source_url": "https://omlx.ai/benchmarks/performance/tsbt77wv",
   "tg_tokens": null,
   "tps": 12.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-tsbt77wv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: daa346z5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 471.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 4k 471.9 12.8 23.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tsbt77wv",
   "source_url": "https://omlx.ai/benchmarks/performance/tsbt77wv",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ll2qaulr-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: 5uam6qc9, i4g13yxl, kmjaxwvn, lbjm4owv, p11w8i6e; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 405.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 405.3 28.7 26.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ll2qaulr",
   "source_url": "https://omlx.ai/benchmarks/performance/ll2qaulr",
   "tg_tokens": null,
   "tps": 28.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ll2qaulr-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: 5uam6qc9, i4g13yxl, kmjaxwvn, lbjm4owv, p11w8i6e; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 451.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 451.3 22.8 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ll2qaulr",
   "source_url": "https://omlx.ai/benchmarks/performance/ll2qaulr",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ll2qaulr-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: 5uam6qc9, i4g13yxl, kmjaxwvn, lbjm4owv, p11w8i6e; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 431.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 8k 431.1 15.9 28.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ll2qaulr",
   "source_url": "https://omlx.ai/benchmarks/performance/ll2qaulr",
   "tg_tokens": null,
   "tps": 15.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ll2qaulr-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 5uam6qc9, i4g13yxl, kmjaxwvn, lbjm4owv, p11w8i6e; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 395.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 16k 395.4 21.4 28.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ll2qaulr",
   "source_url": "https://omlx.ai/benchmarks/performance/ll2qaulr",
   "tg_tokens": null,
   "tps": 21.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ll2qaulr-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: 5uam6qc9, i4g13yxl, kmjaxwvn, lbjm4owv, p11w8i6e; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 349.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 32k 349.4 14.0 29.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ll2qaulr",
   "source_url": "https://omlx.ai/benchmarks/performance/ll2qaulr",
   "tg_tokens": null,
   "tps": 14.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ll2qaulr-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: 5uam6qc9, i4g13yxl, kmjaxwvn, lbjm4owv, p11w8i6e; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 270.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 64k 270.8 10.7 32.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ll2qaulr",
   "source_url": "https://omlx.ai/benchmarks/performance/ll2qaulr",
   "tg_tokens": null,
   "tps": 10.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-jvzjcokv-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 143.1 GB at this context; duplicate traces of the same session: 9jx8d6jq, nlaknkv4, st7nizy9; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 838.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-fp16-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 6bit; 1k 838.6 53.3 143.1 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jvzjcokv",
   "source_url": "https://omlx.ai/benchmarks/performance/jvzjcokv",
   "tg_tokens": null,
   "tps": 53.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-jvzjcokv-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 145.6 GB at this context; duplicate traces of the same session: 9jx8d6jq, nlaknkv4, st7nizy9; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 945.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-fp16-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 6bit; 4k 945.2 51.7 145.6 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jvzjcokv",
   "source_url": "https://omlx.ai/benchmarks/performance/jvzjcokv",
   "tg_tokens": null,
   "tps": 51.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-jvzjcokv-ctx8192",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 145.7 GB at this context; duplicate traces of the same session: 9jx8d6jq, nlaknkv4, st7nizy9; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1038.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-fp16-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 6bit; 8k 1,038 55.6 145.7 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jvzjcokv",
   "source_url": "https://omlx.ai/benchmarks/performance/jvzjcokv",
   "tg_tokens": null,
   "tps": 55.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-jvzjcokv-ctx16384",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 146.0 GB at this context; duplicate traces of the same session: 9jx8d6jq, nlaknkv4, st7nizy9; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1030.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-fp16-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 6bit; 16k 1,030 37.9 146.0 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jvzjcokv",
   "source_url": "https://omlx.ai/benchmarks/performance/jvzjcokv",
   "tg_tokens": null,
   "tps": 37.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ru5a9bf4-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: 52y7ckb2, h1hf46dp, w2iniay2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 914.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 914.5 79.1 100.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ru5a9bf4",
   "source_url": "https://omlx.ai/benchmarks/performance/ru5a9bf4",
   "tg_tokens": null,
   "tps": 79.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ru5a9bf4-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: 52y7ckb2, h1hf46dp, w2iniay2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1080.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,080 65.7 101.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ru5a9bf4",
   "source_url": "https://omlx.ai/benchmarks/performance/ru5a9bf4",
   "tg_tokens": null,
   "tps": 65.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ru5a9bf4-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.9 GB at this context; duplicate traces of the same session: 52y7ckb2, h1hf46dp, w2iniay2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1070.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 1,070 74.7 101.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ru5a9bf4",
   "source_url": "https://omlx.ai/benchmarks/performance/ru5a9bf4",
   "tg_tokens": null,
   "tps": 74.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ru5a9bf4-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.9 GB at this context; duplicate traces of the same session: 52y7ckb2, h1hf46dp, w2iniay2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1060.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 1,060 87.2 102.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ru5a9bf4",
   "source_url": "https://omlx.ai/benchmarks/performance/ru5a9bf4",
   "tg_tokens": null,
   "tps": 87.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ru5a9bf4-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 106.0 GB at this context; duplicate traces of the same session: 52y7ckb2, h1hf46dp, w2iniay2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1045.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 1,045 69.8 106.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ru5a9bf4",
   "source_url": "https://omlx.ai/benchmarks/performance/ru5a9bf4",
   "tg_tokens": null,
   "tps": 69.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-m1239b5f-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "SamQuant C58; oMLX community benchmark trace; peak memory 59.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 561.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "SamQuant-Qwen3.8-Flash-Next-Uncensored-oMLX-C58 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 561.7 43.5 59.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m1239b5f",
   "source_url": "https://omlx.ai/benchmarks/performance/m1239b5f",
   "tg_tokens": null,
   "tps": 43.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-m1239b5f-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "SamQuant C58; oMLX community benchmark trace; peak memory 60.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1049.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "SamQuant-Qwen3.8-Flash-Next-Uncensored-oMLX-C58 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,049 38.6 60.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m1239b5f",
   "source_url": "https://omlx.ai/benchmarks/performance/m1239b5f",
   "tg_tokens": null,
   "tps": 38.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ybjfnu3i-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: 8tygr33a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 926.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 926.0 31.8 17.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ybjfnu3i",
   "source_url": "https://omlx.ai/benchmarks/performance/ybjfnu3i",
   "tg_tokens": null,
   "tps": 31.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ybjfnu3i-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 8tygr33a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 736.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 736.4 29.3 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ybjfnu3i",
   "source_url": "https://omlx.ai/benchmarks/performance/ybjfnu3i",
   "tg_tokens": null,
   "tps": 29.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4pesdzk6-ctx4096",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: 2y774va0, yhc101hu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1033.0,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Qwen3.8-27B-Abliterated-MLX-MXFP4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,033 35.2 16.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4pesdzk6",
   "source_url": "https://omlx.ai/benchmarks/performance/4pesdzk6",
   "tg_tokens": null,
   "tps": 35.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4pesdzk6-ctx16384",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 2y774va0, yhc101hu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 913.4,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Qwen3.8-27B-Abliterated-MLX-MXFP4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 913.4 32.9 17.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4pesdzk6",
   "source_url": "https://omlx.ai/benchmarks/performance/4pesdzk6",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4pesdzk6-ctx131072",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: 2y774va0, yhc101hu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 574.7,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Qwen3.8-27B-Abliterated-MLX-MXFP4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 128k 574.7 22.1 24.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4pesdzk6",
   "source_url": "https://omlx.ai/benchmarks/performance/4pesdzk6",
   "tg_tokens": null,
   "tps": 22.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1dn57899-ctx4096",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: e8fcue1c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1027.0,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Qwen3.8-27B-Abliterated-MLX-MXFP4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,027 35.1 16.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1dn57899",
   "source_url": "https://omlx.ai/benchmarks/performance/1dn57899",
   "tg_tokens": null,
   "tps": 35.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1dn57899-ctx16384",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: e8fcue1c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 882.7,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Qwen3.8-27B-Abliterated-MLX-MXFP4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 882.7 32.9 17.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1dn57899",
   "source_url": "https://omlx.ai/benchmarks/performance/1dn57899",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wpsmpe55-ctx4096",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: cz5smt4b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1028.0,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Qwen3.8-27B-Abliterated-MLX-MXFP4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,028 35.2 16.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wpsmpe55",
   "source_url": "https://omlx.ai/benchmarks/performance/wpsmpe55",
   "tg_tokens": null,
   "tps": 35.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wpsmpe55-ctx16384",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: cz5smt4b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 930.0,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Qwen3.8-27B-Abliterated-MLX-MXFP4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 930.0 33.0 17.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wpsmpe55",
   "source_url": "https://omlx.ai/benchmarks/performance/wpsmpe55",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-8tas9q5l-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 16.0 GB; TTFT 2098 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 488.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k tokens 488.2 PP tok/s 18.0 TG tok/s 2098 TTFT (ms) 16.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8tas9q5l",
   "source_url": "https://omlx.ai/benchmarks/performance/8tas9q5l",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-9p1za3pg-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: cg92iok2; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 99.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ8e-1M on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 99.4 7.9 28.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9p1za3pg",
   "source_url": "https://omlx.ai/benchmarks/performance/9p1za3pg",
   "tg_tokens": null,
   "tps": 7.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-9p1za3pg-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: cg92iok2; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 95.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ8e-1M on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 95.7 7.8 32.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9p1za3pg",
   "source_url": "https://omlx.ai/benchmarks/performance/9p1za3pg",
   "tg_tokens": null,
   "tps": 7.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-3682ag6g-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 13.6 GB at this context; duplicate traces of the same session: k5blmy6g; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 151.4,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-oQ3-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 3bit; 1k 151.4 16.7 13.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3682ag6g",
   "source_url": "https://omlx.ai/benchmarks/performance/3682ag6g",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-3682ag6g-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 14.7 GB at this context; duplicate traces of the same session: k5blmy6g; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 132.0,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-27B-oQ3-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 3bit; 4k 132.0 15.7 14.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3682ag6g",
   "source_url": "https://omlx.ai/benchmarks/performance/3682ag6g",
   "tg_tokens": null,
   "tps": 15.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mf3zzyja-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 73.0 GB; TTFT 601 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1703.0,
   "provenance": "community",
   "quant": "GSQ-RCO-Q2",
   "quote": "Qwen3.8-Flash-Next-GSQ-RCO-Q2-mlx on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,703 PP tok/s 54.4 TG tok/s 601 TTFT (ms) 73.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mf3zzyja",
   "source_url": "https://omlx.ai/benchmarks/performance/mf3zzyja",
   "tg_tokens": null,
   "tps": 54.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-357n22tc-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 76.1 GB; TTFT 30033 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2182.0,
   "provenance": "community",
   "quant": "GSQ-RCO-Q2",
   "quote": "Qwen3.8-Flash-Next-GSQ-RCO-Q2-mlx on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k tokens 2,182 PP tok/s 64.3 TG tok/s 30033 TTFT (ms) 76.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/357n22tc",
   "source_url": "https://omlx.ai/benchmarks/performance/357n22tc",
   "tg_tokens": null,
   "tps": 64.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-b761h3zx-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 72.7 GB; TTFT 597 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1715.0,
   "provenance": "community",
   "quant": "GSQ-RCO-Q2",
   "quote": "Qwen3.8-Flash-Next-GSQ-RCO-Q2-mlx on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,715 PP tok/s 96.4 TG tok/s 597 TTFT (ms) 72.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b761h3zx",
   "source_url": "https://omlx.ai/benchmarks/performance/b761h3zx",
   "tg_tokens": null,
   "tps": 96.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-zyvmjdso-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.8 GB at this context; duplicate traces of the same session: eapcvj7z; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1160.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,160 105.7 100.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zyvmjdso",
   "source_url": "https://omlx.ai/benchmarks/performance/zyvmjdso",
   "tg_tokens": null,
   "tps": 105.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-zyvmjdso-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: eapcvj7z; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1419.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,419 93.5 101.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zyvmjdso",
   "source_url": "https://omlx.ai/benchmarks/performance/zyvmjdso",
   "tg_tokens": null,
   "tps": 93.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-h6wkef55-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: 5gzmakj4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 831.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 831.1 26.7 100.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h6wkef55",
   "source_url": "https://omlx.ai/benchmarks/performance/h6wkef55",
   "tg_tokens": null,
   "tps": 26.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-h6wkef55-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: 5gzmakj4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1390.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 1,390 47.9 101.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h6wkef55",
   "source_url": "https://omlx.ai/benchmarks/performance/h6wkef55",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5u6wtrz8-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 87a7ibym, c0d4yact; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1511.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 1,511 89.4 19.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5u6wtrz8",
   "source_url": "https://omlx.ai/benchmarks/performance/5u6wtrz8",
   "tg_tokens": null,
   "tps": 89.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5u6wtrz8-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 87a7ibym, c0d4yact; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1888.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 1,888 88.0 20.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5u6wtrz8",
   "source_url": "https://omlx.ai/benchmarks/performance/5u6wtrz8",
   "tg_tokens": null,
   "tps": 88.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5u6wtrz8-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 87a7ibym, c0d4yact; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1459.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 1,459 78.0 22.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5u6wtrz8",
   "source_url": "https://omlx.ai/benchmarks/performance/5u6wtrz8",
   "tg_tokens": null,
   "tps": 78.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ig3k8d1b-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 1272 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 804.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k tokens 804.9 PP tok/s 108.5 TG tok/s 1272 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ig3k8d1b",
   "source_url": "https://omlx.ai/benchmarks/performance/ig3k8d1b",
   "tg_tokens": null,
   "tps": 108.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-619v9uyn-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 0br3r6cm, 1n5xlr8x, tkqh62fv; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 291.5,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 1k 291.5 44.4 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/619v9uyn",
   "source_url": "https://omlx.ai/benchmarks/performance/619v9uyn",
   "tg_tokens": null,
   "tps": 44.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-619v9uyn-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: 0br3r6cm, 1n5xlr8x, tkqh62fv; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 313.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 4k 313.0 43.3 23.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/619v9uyn",
   "source_url": "https://omlx.ai/benchmarks/performance/619v9uyn",
   "tg_tokens": null,
   "tps": 43.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-619v9uyn-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 0br3r6cm, 1n5xlr8x, tkqh62fv; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 306.1,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 8k 306.1 50.8 23.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/619v9uyn",
   "source_url": "https://omlx.ai/benchmarks/performance/619v9uyn",
   "tg_tokens": null,
   "tps": 50.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-619v9uyn-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: 0br3r6cm, 1n5xlr8x, tkqh62fv; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 291.6,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 16k 291.6 51.9 25.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/619v9uyn",
   "source_url": "https://omlx.ai/benchmarks/performance/619v9uyn",
   "tg_tokens": null,
   "tps": 51.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-yqno19oi-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 1mj1gjzi, 2rdcxt37, 89bjhyb6; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 290.9,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 1k 290.9 38.0 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yqno19oi",
   "source_url": "https://omlx.ai/benchmarks/performance/yqno19oi",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-yqno19oi-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: 1mj1gjzi, 2rdcxt37, 89bjhyb6; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 312.9,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 4k 312.9 49.3 23.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yqno19oi",
   "source_url": "https://omlx.ai/benchmarks/performance/yqno19oi",
   "tg_tokens": null,
   "tps": 49.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-yqno19oi-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 1mj1gjzi, 2rdcxt37, 89bjhyb6; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 306.1,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 8k 306.1 53.4 23.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yqno19oi",
   "source_url": "https://omlx.ai/benchmarks/performance/yqno19oi",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-yqno19oi-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: 1mj1gjzi, 2rdcxt37, 89bjhyb6; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 291.8,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 16k 291.8 60.4 25.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yqno19oi",
   "source_url": "https://omlx.ai/benchmarks/performance/yqno19oi",
   "tg_tokens": null,
   "tps": 60.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-fw0kzv1n-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 9h8wc16z, hjfqol1z, wa6xclv9; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 299.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 1k 299.0 30.4 16.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fw0kzv1n",
   "source_url": "https://omlx.ai/benchmarks/performance/fw0kzv1n",
   "tg_tokens": null,
   "tps": 30.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-fw0kzv1n-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 9h8wc16z, hjfqol1z, wa6xclv9; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 321.7,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 4k 321.7 30.5 20.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fw0kzv1n",
   "source_url": "https://omlx.ai/benchmarks/performance/fw0kzv1n",
   "tg_tokens": null,
   "tps": 30.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-fw0kzv1n-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: 9h8wc16z, hjfqol1z, wa6xclv9; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 314.5,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 8k 314.5 29.7 20.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fw0kzv1n",
   "source_url": "https://omlx.ai/benchmarks/performance/fw0kzv1n",
   "tg_tokens": null,
   "tps": 29.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-fw0kzv1n-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 9h8wc16z, hjfqol1z, wa6xclv9; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 300.4,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 16k 300.4 28.5 22.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fw0kzv1n",
   "source_url": "https://omlx.ai/benchmarks/performance/fw0kzv1n",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lqlegz52-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.8 GB at this context; duplicate traces of the same session: cwf80sqi, g10v1msh, hek44m0c, k8xqwwap, ollk2qd5, p8g64jbc, u3j5gcu2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1174.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,174 77.9 100.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lqlegz52",
   "source_url": "https://omlx.ai/benchmarks/performance/lqlegz52",
   "tg_tokens": null,
   "tps": 77.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lqlegz52-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: cwf80sqi, g10v1msh, hek44m0c, k8xqwwap, ollk2qd5, p8g64jbc, u3j5gcu2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1422.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,422 61.6 101.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lqlegz52",
   "source_url": "https://omlx.ai/benchmarks/performance/lqlegz52",
   "tg_tokens": null,
   "tps": 61.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lqlegz52-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: cwf80sqi, g10v1msh, hek44m0c, k8xqwwap, ollk2qd5, p8g64jbc, u3j5gcu2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1410.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 1,410 63.3 101.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lqlegz52",
   "source_url": "https://omlx.ai/benchmarks/performance/lqlegz52",
   "tg_tokens": null,
   "tps": 63.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lqlegz52-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.7 GB at this context; duplicate traces of the same session: cwf80sqi, g10v1msh, hek44m0c, k8xqwwap, ollk2qd5, p8g64jbc, u3j5gcu2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1399.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 1,399 74.2 101.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lqlegz52",
   "source_url": "https://omlx.ai/benchmarks/performance/lqlegz52",
   "tg_tokens": null,
   "tps": 74.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lqlegz52-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.9 GB at this context; duplicate traces of the same session: cwf80sqi, g10v1msh, hek44m0c, k8xqwwap, ollk2qd5, p8g64jbc, u3j5gcu2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1380.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 1,380 57.0 102.9 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lqlegz52",
   "source_url": "https://omlx.ai/benchmarks/performance/lqlegz52",
   "tg_tokens": null,
   "tps": 57.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lqlegz52-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 105.4 GB at this context; duplicate traces of the same session: cwf80sqi, g10v1msh, hek44m0c, k8xqwwap, ollk2qd5, p8g64jbc, u3j5gcu2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1358.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 1,358 47.1 105.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lqlegz52",
   "source_url": "https://omlx.ai/benchmarks/performance/lqlegz52",
   "tg_tokens": null,
   "tps": 47.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lqlegz52-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.8 GB at this context; duplicate traces of the same session: cwf80sqi, g10v1msh, hek44m0c, k8xqwwap, ollk2qd5, p8g64jbc, u3j5gcu2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1341.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 1,341 53.2 110.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lqlegz52",
   "source_url": "https://omlx.ai/benchmarks/performance/lqlegz52",
   "tg_tokens": null,
   "tps": 53.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lqlegz52-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 111.6 GB at this context; duplicate traces of the same session: cwf80sqi, g10v1msh, hek44m0c, k8xqwwap, ollk2qd5, p8g64jbc, u3j5gcu2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1323.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 195k 1,323 52.3 111.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lqlegz52",
   "source_url": "https://omlx.ai/benchmarks/performance/lqlegz52",
   "tg_tokens": null,
   "tps": 52.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-dtvhqlon-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 38.2 GB at this context; duplicate traces of the same session: 45tsjhr6, 6bafq5ql, cg9xutyf, uvvl0i3p; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1154.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,154 67.4 38.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dtvhqlon",
   "source_url": "https://omlx.ai/benchmarks/performance/dtvhqlon",
   "tg_tokens": null,
   "tps": 67.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-dtvhqlon-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 40.3 GB at this context; duplicate traces of the same session: 45tsjhr6, 6bafq5ql, cg9xutyf, uvvl0i3p; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1418.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 1,418 66.8 40.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dtvhqlon",
   "source_url": "https://omlx.ai/benchmarks/performance/dtvhqlon",
   "tg_tokens": null,
   "tps": 66.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-dtvhqlon-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 40.6 GB at this context; duplicate traces of the same session: 45tsjhr6, 6bafq5ql, cg9xutyf, uvvl0i3p; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1354.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 8k 1,354 65.5 40.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dtvhqlon",
   "source_url": "https://omlx.ai/benchmarks/performance/dtvhqlon",
   "tg_tokens": null,
   "tps": 65.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-dtvhqlon-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 41.3 GB at this context; duplicate traces of the same session: 45tsjhr6, 6bafq5ql, cg9xutyf, uvvl0i3p; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1191.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 16k 1,191 83.0 41.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dtvhqlon",
   "source_url": "https://omlx.ai/benchmarks/performance/dtvhqlon",
   "tg_tokens": null,
   "tps": 83.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-dtvhqlon-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 42.6 GB at this context; duplicate traces of the same session: 45tsjhr6, 6bafq5ql, cg9xutyf, uvvl0i3p; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 951.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 32k 951.0 95.5 42.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dtvhqlon",
   "source_url": "https://omlx.ai/benchmarks/performance/dtvhqlon",
   "tg_tokens": null,
   "tps": 95.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-vmqzkfc5-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 1247bnzi, a4l8b3nr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1547.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 1,547 88.1 19.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vmqzkfc5",
   "source_url": "https://omlx.ai/benchmarks/performance/vmqzkfc5",
   "tg_tokens": null,
   "tps": 88.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-vmqzkfc5-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 1247bnzi, a4l8b3nr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1943.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 1,943 81.5 20.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vmqzkfc5",
   "source_url": "https://omlx.ai/benchmarks/performance/vmqzkfc5",
   "tg_tokens": null,
   "tps": 81.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-vmqzkfc5-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 1247bnzi, a4l8b3nr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1481.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 1,481 76.9 22.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vmqzkfc5",
   "source_url": "https://omlx.ai/benchmarks/performance/vmqzkfc5",
   "tg_tokens": null,
   "tps": 76.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-scf3siht-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 3djctvzq; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 106.1 15.6 16.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/scf3siht",
   "source_url": "https://omlx.ai/benchmarks/performance/scf3siht",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-scf3siht-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 3djctvzq; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 108.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 108.8 15.2 17.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/scf3siht",
   "source_url": "https://omlx.ai/benchmarks/performance/scf3siht",
   "tg_tokens": null,
   "tps": 15.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-c3022x7a-ctx1024",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: v8u2jv0c, wvwx8gis; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 364.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Alis-MLX-6bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 364.0 43.4 27.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c3022x7a",
   "source_url": "https://omlx.ai/benchmarks/performance/c3022x7a",
   "tg_tokens": null,
   "tps": 43.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-c3022x7a-ctx4096",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 31.0 GB at this context; duplicate traces of the same session: v8u2jv0c, wvwx8gis; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 385.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Alis-MLX-6bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 385.3 35.6 31.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c3022x7a",
   "source_url": "https://omlx.ai/benchmarks/performance/c3022x7a",
   "tg_tokens": null,
   "tps": 35.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-c3022x7a-ctx32768",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 34.1 GB at this context; duplicate traces of the same session: v8u2jv0c, wvwx8gis; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 328.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Alis-MLX-6bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 328.4 46.5 34.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c3022x7a",
   "source_url": "https://omlx.ai/benchmarks/performance/c3022x7a",
   "tg_tokens": null,
   "tps": 46.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-k93q3he5-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: 4kt2dqn8, 85g6rnkc, khtkekby, u8u7xf64; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 186.1,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 186.1 26.7 30.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k93q3he5",
   "source_url": "https://omlx.ai/benchmarks/performance/k93q3he5",
   "tg_tokens": null,
   "tps": 26.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-k93q3he5-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.7 GB at this context; duplicate traces of the same session: 4kt2dqn8, 85g6rnkc, khtkekby, u8u7xf64; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 278.1,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 278.1 29.7 33.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k93q3he5",
   "source_url": "https://omlx.ai/benchmarks/performance/k93q3he5",
   "tg_tokens": null,
   "tps": 29.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-k93q3he5-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: 4kt2dqn8, 85g6rnkc, khtkekby, u8u7xf64; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 260.5,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 260.5 37.7 34.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k93q3he5",
   "source_url": "https://omlx.ai/benchmarks/performance/k93q3he5",
   "tg_tokens": null,
   "tps": 37.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-k93q3he5-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: 4kt2dqn8, 85g6rnkc, khtkekby, u8u7xf64; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 251.3,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 251.3 30.9 36.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k93q3he5",
   "source_url": "https://omlx.ai/benchmarks/performance/k93q3he5",
   "tg_tokens": null,
   "tps": 30.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-k93q3he5-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: 4kt2dqn8, 85g6rnkc, khtkekby, u8u7xf64; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.2,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 32k 220.2 53.7 37.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k93q3he5",
   "source_url": "https://omlx.ai/benchmarks/performance/k93q3he5",
   "tg_tokens": null,
   "tps": 53.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-wtbk48q9-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: xwc0hcyh; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 239.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 239.6 16.4 70.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wtbk48q9",
   "source_url": "https://omlx.ai/benchmarks/performance/wtbk48q9",
   "tg_tokens": null,
   "tps": 16.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-wtbk48q9-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: xwc0hcyh; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 293.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 293.2 33.8 71.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wtbk48q9",
   "source_url": "https://omlx.ai/benchmarks/performance/wtbk48q9",
   "tg_tokens": null,
   "tps": 33.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-wtbk48q9-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: xwc0hcyh; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 399.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 399.2 33.3 71.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wtbk48q9",
   "source_url": "https://omlx.ai/benchmarks/performance/wtbk48q9",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-wtbk48q9-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: xwc0hcyh; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 395.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 395.4 20.1 71.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wtbk48q9",
   "source_url": "https://omlx.ai/benchmarks/performance/wtbk48q9",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-wtbk48q9-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.5 GB at this context; duplicate traces of the same session: xwc0hcyh; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 395.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 32k 395.4 21.5 72.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wtbk48q9",
   "source_url": "https://omlx.ai/benchmarks/performance/wtbk48q9",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-cyo496m3-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1136.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,136 87.0 100.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cyo496m3",
   "source_url": "https://omlx.ai/benchmarks/performance/cyo496m3",
   "tg_tokens": null,
   "tps": 87.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-cyo496m3-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1384.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,384 81.9 101.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cyo496m3",
   "source_url": "https://omlx.ai/benchmarks/performance/cyo496m3",
   "tg_tokens": null,
   "tps": 81.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ud23dple-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.6 GB at this context; duplicate traces of the same session: zv7inv8m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2085.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,085 49.9 72.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ud23dple",
   "source_url": "https://omlx.ai/benchmarks/performance/ud23dple",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ud23dple-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.9 GB at this context; duplicate traces of the same session: zv7inv8m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2213.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,213 61.7 74.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ud23dple",
   "source_url": "https://omlx.ai/benchmarks/performance/ud23dple",
   "tg_tokens": null,
   "tps": 61.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-t8crp5l5-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: 18f1lr0g, b60w4gc7, bed7dkx4, h8iwbihi, injfhxla, pb6cn4ja, ts32kbo3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2979.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 2,979 122.7 100.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t8crp5l5",
   "source_url": "https://omlx.ai/benchmarks/performance/t8crp5l5",
   "tg_tokens": null,
   "tps": 122.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-t8crp5l5-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.0 GB at this context; duplicate traces of the same session: 18f1lr0g, b60w4gc7, bed7dkx4, h8iwbihi, injfhxla, pb6cn4ja, ts32kbo3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5111.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 5,111 130.9 101.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t8crp5l5",
   "source_url": "https://omlx.ai/benchmarks/performance/t8crp5l5",
   "tg_tokens": null,
   "tps": 130.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-t8crp5l5-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.1 GB at this context; duplicate traces of the same session: 18f1lr0g, b60w4gc7, bed7dkx4, h8iwbihi, injfhxla, pb6cn4ja, ts32kbo3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5558.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 5,558 148.0 102.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t8crp5l5",
   "source_url": "https://omlx.ai/benchmarks/performance/t8crp5l5",
   "tg_tokens": null,
   "tps": 148.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-t8crp5l5-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.5 GB at this context; duplicate traces of the same session: 18f1lr0g, b60w4gc7, bed7dkx4, h8iwbihi, injfhxla, pb6cn4ja, ts32kbo3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5638.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 5,638 136.5 102.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t8crp5l5",
   "source_url": "https://omlx.ai/benchmarks/performance/t8crp5l5",
   "tg_tokens": null,
   "tps": 136.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-t8crp5l5-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.1 GB at this context; duplicate traces of the same session: 18f1lr0g, b60w4gc7, bed7dkx4, h8iwbihi, injfhxla, pb6cn4ja, ts32kbo3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5624.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 5,624 109.6 103.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t8crp5l5",
   "source_url": "https://omlx.ai/benchmarks/performance/t8crp5l5",
   "tg_tokens": null,
   "tps": 109.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-t8crp5l5-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.7 GB at this context; duplicate traces of the same session: 18f1lr0g, b60w4gc7, bed7dkx4, h8iwbihi, injfhxla, pb6cn4ja, ts32kbo3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5523.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 5,523 141.2 104.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t8crp5l5",
   "source_url": "https://omlx.ai/benchmarks/performance/t8crp5l5",
   "tg_tokens": null,
   "tps": 141.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-t8crp5l5-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 108.7 GB at this context; duplicate traces of the same session: 18f1lr0g, b60w4gc7, bed7dkx4, h8iwbihi, injfhxla, pb6cn4ja, ts32kbo3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5354.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,354 109.3 108.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t8crp5l5",
   "source_url": "https://omlx.ai/benchmarks/performance/t8crp5l5",
   "tg_tokens": null,
   "tps": 109.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-t8crp5l5-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.7 GB at this context; duplicate traces of the same session: 18f1lr0g, b60w4gc7, bed7dkx4, h8iwbihi, injfhxla, pb6cn4ja, ts32kbo3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5207.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 5,207 89.3 110.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t8crp5l5",
   "source_url": "https://omlx.ai/benchmarks/performance/t8crp5l5",
   "tg_tokens": null,
   "tps": 89.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tm8wn0z6-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 84.9 GB at this context; duplicate traces of the same session: j5gtwjf3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1701.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 1,701 65.4 84.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tm8wn0z6",
   "source_url": "https://omlx.ai/benchmarks/performance/tm8wn0z6",
   "tg_tokens": null,
   "tps": 65.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tm8wn0z6-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 87.2 GB at this context; duplicate traces of the same session: j5gtwjf3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1668.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 1,668 40.8 87.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tm8wn0z6",
   "source_url": "https://omlx.ai/benchmarks/performance/tm8wn0z6",
   "tg_tokens": null,
   "tps": 40.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-r2t1czwv-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.6 GB at this context; duplicate traces of the same session: 4uk3ry71, cwrmkwuu, o9jxqh27, ocfpw4lh; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1889.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 1k 1,889 95.5 142.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r2t1czwv",
   "source_url": "https://omlx.ai/benchmarks/performance/r2t1czwv",
   "tg_tokens": null,
   "tps": 95.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-r2t1czwv-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 144.5 GB at this context; duplicate traces of the same session: 4uk3ry71, cwrmkwuu, o9jxqh27, ocfpw4lh; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3976.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 8k 3,976 127.7 144.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r2t1czwv",
   "source_url": "https://omlx.ai/benchmarks/performance/r2t1czwv",
   "tg_tokens": null,
   "tps": 127.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-r2t1czwv-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 146.8 GB at this context; duplicate traces of the same session: 4uk3ry71, cwrmkwuu, o9jxqh27, ocfpw4lh; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3917.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 64k 3,917 104.5 146.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r2t1czwv",
   "source_url": "https://omlx.ai/benchmarks/performance/r2t1czwv",
   "tg_tokens": null,
   "tps": 104.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-r2t1czwv-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 150.5 GB at this context; duplicate traces of the same session: 4uk3ry71, cwrmkwuu, o9jxqh27, ocfpw4lh; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3814.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 128k 3,814 93.4 150.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r2t1czwv",
   "source_url": "https://omlx.ai/benchmarks/performance/r2t1czwv",
   "tg_tokens": null,
   "tps": 93.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-r2t1czwv-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 153.3 GB at this context; duplicate traces of the same session: 4uk3ry71, cwrmkwuu, o9jxqh27, ocfpw4lh; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3711.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 195k 3,711 84.5 153.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r2t1czwv",
   "source_url": "https://omlx.ai/benchmarks/performance/r2t1czwv",
   "tg_tokens": null,
   "tps": 84.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-u3req7x5-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.6 GB at this context; duplicate traces of the same session: 41r8v92q, 82ez1k9c, ds0pfv6g, imiimm50; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2470.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 2,470 149.5 142.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u3req7x5",
   "source_url": "https://omlx.ai/benchmarks/performance/u3req7x5",
   "tg_tokens": null,
   "tps": 149.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-u3req7x5-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 144.3 GB at this context; duplicate traces of the same session: 41r8v92q, 82ez1k9c, ds0pfv6g, imiimm50; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4852.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 8k 4,852 132.7 144.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u3req7x5",
   "source_url": "https://omlx.ai/benchmarks/performance/u3req7x5",
   "tg_tokens": null,
   "tps": 132.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-u3req7x5-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 146.8 GB at this context; duplicate traces of the same session: 41r8v92q, 82ez1k9c, ds0pfv6g, imiimm50; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4784.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 64k 4,784 129.2 146.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u3req7x5",
   "source_url": "https://omlx.ai/benchmarks/performance/u3req7x5",
   "tg_tokens": null,
   "tps": 129.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-u3req7x5-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 150.6 GB at this context; duplicate traces of the same session: 41r8v92q, 82ez1k9c, ds0pfv6g, imiimm50; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4649.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 128k 4,649 102.9 150.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u3req7x5",
   "source_url": "https://omlx.ai/benchmarks/performance/u3req7x5",
   "tg_tokens": null,
   "tps": 102.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-u3req7x5-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 153.3 GB at this context; duplicate traces of the same session: 41r8v92q, 82ez1k9c, ds0pfv6g, imiimm50; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4529.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 195k 4,529 86.5 153.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u3req7x5",
   "source_url": "https://omlx.ai/benchmarks/performance/u3req7x5",
   "tg_tokens": null,
   "tps": 86.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-fgbpgcxb-ctx1024",
   "model": "gemma-4-31B-it-qat-q4_0-unquantized",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 237.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-qat-q4_0-unquantized-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 1k 237.9 21.7 18.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fgbpgcxb",
   "source_url": "https://omlx.ai/benchmarks/performance/fgbpgcxb",
   "tg_tokens": null,
   "tps": 21.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-fgbpgcxb-ctx4096",
   "model": "gemma-4-31B-it-qat-q4_0-unquantized",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 237.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-qat-q4_0-unquantized-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 4k 237.8 24.1 19.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fgbpgcxb",
   "source_url": "https://omlx.ai/benchmarks/performance/fgbpgcxb",
   "tg_tokens": null,
   "tps": 24.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-h81tdkw3-ctx4096",
   "model": "Llama-3.2-3B-Instruct",
   "notes": "mlx-community; oMLX community benchmark trace; peak memory 2.8 GB at this context; duplicate traces of the same session: kzmfbd4c; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1615.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Llama-3.2-3B-Instruct-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 1,615 135.4 2.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h81tdkw3",
   "source_url": "https://omlx.ai/benchmarks/performance/h81tdkw3",
   "tg_tokens": null,
   "tps": 135.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-h81tdkw3-ctx16384",
   "model": "Llama-3.2-3B-Instruct",
   "notes": "mlx-community; oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: kzmfbd4c; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1188.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Llama-3.2-3B-Instruct-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 1,188 88.7 3.9 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h81tdkw3",
   "source_url": "https://omlx.ai/benchmarks/performance/h81tdkw3",
   "tg_tokens": null,
   "tps": 88.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-1svl8w7b-ctx1024",
   "model": "CoPaw-Flash-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: hxr41m2y; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 211.1,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "CoPaw-Flash-9B-oQ4 on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 211.1 31.5 6.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1svl8w7b",
   "source_url": "https://omlx.ai/benchmarks/performance/1svl8w7b",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-1svl8w7b-ctx4096",
   "model": "CoPaw-Flash-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.4 GB at this context; duplicate traces of the same session: hxr41m2y; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.1,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "CoPaw-Flash-9B-oQ4 on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 224.1 35.2 7.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1svl8w7b",
   "source_url": "https://omlx.ai/benchmarks/performance/1svl8w7b",
   "tg_tokens": null,
   "tps": 35.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-1svl8w7b-ctx16384",
   "model": "CoPaw-Flash-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.4 GB at this context; duplicate traces of the same session: hxr41m2y; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 203.7,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "CoPaw-Flash-9B-oQ4 on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 16k 203.7 25.1 8.4 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1svl8w7b",
   "source_url": "https://omlx.ai/benchmarks/performance/1svl8w7b",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-7kjqj0ab-ctx1024",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "fp16 ccal build; oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: 9v586z6f, b3kl7p20, dv97dwdp, zw2lu5oq; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1197.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-oQ8e-fp16-ccal on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,197 68.9 35.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7kjqj0ab",
   "source_url": "https://omlx.ai/benchmarks/performance/7kjqj0ab",
   "tg_tokens": null,
   "tps": 68.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-7kjqj0ab-ctx4096",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "fp16 ccal build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: 9v586z6f, b3kl7p20, dv97dwdp, zw2lu5oq; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1485.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-oQ8e-fp16-ccal on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 1,485 66.6 37.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7kjqj0ab",
   "source_url": "https://omlx.ai/benchmarks/performance/7kjqj0ab",
   "tg_tokens": null,
   "tps": 66.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-7kjqj0ab-ctx8192",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "fp16 ccal build; oMLX community benchmark trace; peak memory 37.7 GB at this context; duplicate traces of the same session: 9v586z6f, b3kl7p20, dv97dwdp, zw2lu5oq; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1416.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-oQ8e-fp16-ccal on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 8k 1,416 62.1 37.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7kjqj0ab",
   "source_url": "https://omlx.ai/benchmarks/performance/7kjqj0ab",
   "tg_tokens": null,
   "tps": 62.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-7kjqj0ab-ctx16384",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "fp16 ccal build; oMLX community benchmark trace; peak memory 38.4 GB at this context; duplicate traces of the same session: 9v586z6f, b3kl7p20, dv97dwdp, zw2lu5oq; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1272.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-oQ8e-fp16-ccal on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 16k 1,272 63.4 38.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7kjqj0ab",
   "source_url": "https://omlx.ai/benchmarks/performance/7kjqj0ab",
   "tg_tokens": null,
   "tps": 63.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-7kjqj0ab-ctx32768",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "fp16 ccal build; oMLX community benchmark trace; peak memory 39.6 GB at this context; duplicate traces of the same session: 9v586z6f, b3kl7p20, dv97dwdp, zw2lu5oq; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1022.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-oQ8e-fp16-ccal on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 32k 1,022 55.2 39.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7kjqj0ab",
   "source_url": "https://omlx.ai/benchmarks/performance/7kjqj0ab",
   "tg_tokens": null,
   "tps": 55.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-o01m8bsv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 044novwg, 4d19ttel; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1019.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,019 59.2 18.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o01m8bsv",
   "source_url": "https://omlx.ai/benchmarks/performance/o01m8bsv",
   "tg_tokens": null,
   "tps": 59.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-o01m8bsv-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 044novwg, 4d19ttel; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 858.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 858.7 34.5 19.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o01m8bsv",
   "source_url": "https://omlx.ai/benchmarks/performance/o01m8bsv",
   "tg_tokens": null,
   "tps": 34.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-o01m8bsv-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: 044novwg, 4d19ttel; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 549.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 128k 549.0 21.4 24.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o01m8bsv",
   "source_url": "https://omlx.ai/benchmarks/performance/o01m8bsv",
   "tg_tokens": null,
   "tps": 21.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-eds3p3nc-ctx1024",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 3calai0p, b9bv6vsm; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 390.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Alis-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 390.4 51.3 19.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eds3p3nc",
   "source_url": "https://omlx.ai/benchmarks/performance/eds3p3nc",
   "tg_tokens": null,
   "tps": 51.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-eds3p3nc-ctx4096",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 3calai0p, b9bv6vsm; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 422.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Alis-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 422.9 50.3 23.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eds3p3nc",
   "source_url": "https://omlx.ai/benchmarks/performance/eds3p3nc",
   "tg_tokens": null,
   "tps": 50.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-eds3p3nc-ctx32768",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: 3calai0p, b9bv6vsm; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 358.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Alis-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 358.4 64.4 26.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eds3p3nc",
   "source_url": "https://omlx.ai/benchmarks/performance/eds3p3nc",
   "tg_tokens": null,
   "tps": 64.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-mvg9utql-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 108.2 GB; TTFT 2993 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1368.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-MLX-4bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k tokens 1,368 PP tok/s 90.3 TG tok/s 2993 TTFT (ms) 108.2 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mvg9utql",
   "source_url": "https://omlx.ai/benchmarks/performance/mvg9utql",
   "tg_tokens": null,
   "tps": 90.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-0k09dsz7-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 139.4 GB at this context; duplicate traces of the same session: pit8q7pr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1133.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 1k 1,133 55.2 139.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0k09dsz7",
   "source_url": "https://omlx.ai/benchmarks/performance/0k09dsz7",
   "tg_tokens": null,
   "tps": 55.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-0k09dsz7-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 140.0 GB at this context; duplicate traces of the same session: pit8q7pr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1387.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 4k 1,387 50.7 140.0 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0k09dsz7",
   "source_url": "https://omlx.ai/benchmarks/performance/0k09dsz7",
   "tg_tokens": null,
   "tps": 50.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-uwj19soh-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 4a2ju2oq, sm7x3oqq; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 239.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 239.6 16.4 70.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uwj19soh",
   "source_url": "https://omlx.ai/benchmarks/performance/uwj19soh",
   "tg_tokens": null,
   "tps": 16.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-uwj19soh-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: 4a2ju2oq, sm7x3oqq; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 293.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 293.2 33.8 71.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uwj19soh",
   "source_url": "https://omlx.ai/benchmarks/performance/uwj19soh",
   "tg_tokens": null,
   "tps": 33.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-uwj19soh-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: 4a2ju2oq, sm7x3oqq; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 399.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 399.2 33.3 71.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uwj19soh",
   "source_url": "https://omlx.ai/benchmarks/performance/uwj19soh",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-uwj19soh-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 4a2ju2oq, sm7x3oqq; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 395.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 395.4 20.1 71.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uwj19soh",
   "source_url": "https://omlx.ai/benchmarks/performance/uwj19soh",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-uwj19soh-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.5 GB at this context; duplicate traces of the same session: 4a2ju2oq, sm7x3oqq; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 395.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 32k 395.4 21.5 72.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uwj19soh",
   "source_url": "https://omlx.ai/benchmarks/performance/uwj19soh",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9rqfpgub-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1136.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,136 87.0 100.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9rqfpgub",
   "source_url": "https://omlx.ai/benchmarks/performance/9rqfpgub",
   "tg_tokens": null,
   "tps": 87.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9rqfpgub-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1384.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,384 81.9 101.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9rqfpgub",
   "source_url": "https://omlx.ai/benchmarks/performance/9rqfpgub",
   "tg_tokens": null,
   "tps": 81.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xonnoklu-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: 36zbox4k, pr32dkxe, qk2euhrd, vzo84cup, y36eaa4k; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1177.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,177 79.2 100.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xonnoklu",
   "source_url": "https://omlx.ai/benchmarks/performance/xonnoklu",
   "tg_tokens": null,
   "tps": 79.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xonnoklu-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: 36zbox4k, pr32dkxe, qk2euhrd, vzo84cup, y36eaa4k; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1423.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,423 83.2 101.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xonnoklu",
   "source_url": "https://omlx.ai/benchmarks/performance/xonnoklu",
   "tg_tokens": null,
   "tps": 83.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xonnoklu-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: 36zbox4k, pr32dkxe, qk2euhrd, vzo84cup, y36eaa4k; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1411.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 1,411 81.5 101.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xonnoklu",
   "source_url": "https://omlx.ai/benchmarks/performance/xonnoklu",
   "tg_tokens": null,
   "tps": 81.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xonnoklu-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.6 GB at this context; duplicate traces of the same session: 36zbox4k, pr32dkxe, qk2euhrd, vzo84cup, y36eaa4k; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1399.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 1,399 83.4 101.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xonnoklu",
   "source_url": "https://omlx.ai/benchmarks/performance/xonnoklu",
   "tg_tokens": null,
   "tps": 83.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xonnoklu-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.2 GB at this context; duplicate traces of the same session: 36zbox4k, pr32dkxe, qk2euhrd, vzo84cup, y36eaa4k; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1388.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 1,388 69.4 102.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xonnoklu",
   "source_url": "https://omlx.ai/benchmarks/performance/xonnoklu",
   "tg_tokens": null,
   "tps": 69.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xonnoklu-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.1 GB at this context; duplicate traces of the same session: 36zbox4k, pr32dkxe, qk2euhrd, vzo84cup, y36eaa4k; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1365.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 1,365 72.1 104.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xonnoklu",
   "source_url": "https://omlx.ai/benchmarks/performance/xonnoklu",
   "tg_tokens": null,
   "tps": 72.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xonnoklu-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 108.6 GB at this context; duplicate traces of the same session: 36zbox4k, pr32dkxe, qk2euhrd, vzo84cup, y36eaa4k; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1341.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 1,341 62.1 108.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xonnoklu",
   "source_url": "https://omlx.ai/benchmarks/performance/xonnoklu",
   "tg_tokens": null,
   "tps": 62.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xonnoklu-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.8 GB at this context; duplicate traces of the same session: 36zbox4k, pr32dkxe, qk2euhrd, vzo84cup, y36eaa4k; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1322.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 195k 1,322 57.6 109.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xonnoklu",
   "source_url": "https://omlx.ai/benchmarks/performance/xonnoklu",
   "tg_tokens": null,
   "tps": 57.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-rbcpr34w-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: d6c3mweu, jsxsgufa, pqfg3nvo, rd93ij7z; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5575.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 5,575 121.8 101.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rbcpr34w",
   "source_url": "https://omlx.ai/benchmarks/performance/rbcpr34w",
   "tg_tokens": null,
   "tps": 121.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-rbcpr34w-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.2 GB at this context; duplicate traces of the same session: d6c3mweu, jsxsgufa, pqfg3nvo, rd93ij7z; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6199.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 6,199 157.4 102.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rbcpr34w",
   "source_url": "https://omlx.ai/benchmarks/performance/rbcpr34w",
   "tg_tokens": null,
   "tps": 157.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-rbcpr34w-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.7 GB at this context; duplicate traces of the same session: d6c3mweu, jsxsgufa, pqfg3nvo, rd93ij7z; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6051.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 6,051 137.9 104.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rbcpr34w",
   "source_url": "https://omlx.ai/benchmarks/performance/rbcpr34w",
   "tg_tokens": null,
   "tps": 137.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-rbcpr34w-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 108.7 GB at this context; duplicate traces of the same session: d6c3mweu, jsxsgufa, pqfg3nvo, rd93ij7z; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5849.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,849 104.8 108.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rbcpr34w",
   "source_url": "https://omlx.ai/benchmarks/performance/rbcpr34w",
   "tg_tokens": null,
   "tps": 104.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-rbcpr34w-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.7 GB at this context; duplicate traces of the same session: d6c3mweu, jsxsgufa, pqfg3nvo, rd93ij7z; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5674.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 5,674 97.2 110.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rbcpr34w",
   "source_url": "https://omlx.ai/benchmarks/performance/rbcpr34w",
   "tg_tokens": null,
   "tps": 97.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nce5zb4k-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 84.7 GB at this context; duplicate traces of the same session: bhyqpbej, segyqe06; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 900.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 900.0 57.3 84.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nce5zb4k",
   "source_url": "https://omlx.ai/benchmarks/performance/nce5zb4k",
   "tg_tokens": null,
   "tps": 57.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nce5zb4k-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 84.8 GB at this context; duplicate traces of the same session: bhyqpbej, segyqe06; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1258.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 1,258 58.5 84.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nce5zb4k",
   "source_url": "https://omlx.ai/benchmarks/performance/nce5zb4k",
   "tg_tokens": null,
   "tps": 58.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nce5zb4k-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 84.8 GB at this context; duplicate traces of the same session: bhyqpbej, segyqe06; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1906.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 1,906 58.3 84.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nce5zb4k",
   "source_url": "https://omlx.ai/benchmarks/performance/nce5zb4k",
   "tg_tokens": null,
   "tps": 58.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-042tohqx-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 139.4 GB at this context; duplicate traces of the same session: 031lyxet, 368b8i7f, 4uaggw0n, 5b994emv, 8x8sl3fw, elyiyugv, yok2xxe4; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1006.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 1k 1,006 90.8 139.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/042tohqx",
   "source_url": "https://omlx.ai/benchmarks/performance/042tohqx",
   "tg_tokens": null,
   "tps": 90.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-042tohqx-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 140.0 GB at this context; duplicate traces of the same session: 031lyxet, 368b8i7f, 4uaggw0n, 5b994emv, 8x8sl3fw, elyiyugv, yok2xxe4; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2469.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 4k 2,469 86.2 140.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/042tohqx",
   "source_url": "https://omlx.ai/benchmarks/performance/042tohqx",
   "tg_tokens": null,
   "tps": 86.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-042tohqx-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 141.1 GB at this context; duplicate traces of the same session: 031lyxet, 368b8i7f, 4uaggw0n, 5b994emv, 8x8sl3fw, elyiyugv, yok2xxe4; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3544.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 8k 3,544 84.4 141.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/042tohqx",
   "source_url": "https://omlx.ai/benchmarks/performance/042tohqx",
   "tg_tokens": null,
   "tps": 84.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-042tohqx-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 141.3 GB at this context; duplicate traces of the same session: 031lyxet, 368b8i7f, 4uaggw0n, 5b994emv, 8x8sl3fw, elyiyugv, yok2xxe4; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4072.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 16k 4,072 85.0 141.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/042tohqx",
   "source_url": "https://omlx.ai/benchmarks/performance/042tohqx",
   "tg_tokens": null,
   "tps": 85.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-042tohqx-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 141.8 GB at this context; duplicate traces of the same session: 031lyxet, 368b8i7f, 4uaggw0n, 5b994emv, 8x8sl3fw, elyiyugv, yok2xxe4; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4106.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 32k 4,106 82.3 141.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/042tohqx",
   "source_url": "https://omlx.ai/benchmarks/performance/042tohqx",
   "tg_tokens": null,
   "tps": 82.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-042tohqx-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.8 GB at this context; duplicate traces of the same session: 031lyxet, 368b8i7f, 4uaggw0n, 5b994emv, 8x8sl3fw, elyiyugv, yok2xxe4; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4038.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 64k 4,038 76.5 142.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/042tohqx",
   "source_url": "https://omlx.ai/benchmarks/performance/042tohqx",
   "tg_tokens": null,
   "tps": 76.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-042tohqx-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.0 GB at this context; duplicate traces of the same session: 031lyxet, 368b8i7f, 4uaggw0n, 5b994emv, 8x8sl3fw, elyiyugv, yok2xxe4; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3941.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 128k 3,941 75.1 145.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/042tohqx",
   "source_url": "https://omlx.ai/benchmarks/performance/042tohqx",
   "tg_tokens": null,
   "tps": 75.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-042tohqx-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 148.7 GB at this context; duplicate traces of the same session: 031lyxet, 368b8i7f, 4uaggw0n, 5b994emv, 8x8sl3fw, elyiyugv, yok2xxe4; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3828.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 195k 3,828 74.7 148.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/042tohqx",
   "source_url": "https://omlx.ai/benchmarks/performance/042tohqx",
   "tg_tokens": null,
   "tps": 74.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-qbyhak3v-ctx1024",
   "model": "gemma-4-31B-it-qat-q4_0-unquantized",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 237.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-qat-q4_0-unquantized-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 1k 237.9 21.7 18.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qbyhak3v",
   "source_url": "https://omlx.ai/benchmarks/performance/qbyhak3v",
   "tg_tokens": null,
   "tps": 21.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-qbyhak3v-ctx4096",
   "model": "gemma-4-31B-it-qat-q4_0-unquantized",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 237.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-qat-q4_0-unquantized-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 4k 237.8 24.1 19.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qbyhak3v",
   "source_url": "https://omlx.ai/benchmarks/performance/qbyhak3v",
   "tg_tokens": null,
   "tps": 24.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-bhhj6r5s-ctx1024",
   "model": "gemma-4-31B-it-qat-q4_0-unquantized",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: ork17rui; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 237.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-qat-q4_0-unquantized-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 1k 237.9 21.7 18.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhhj6r5s",
   "source_url": "https://omlx.ai/benchmarks/performance/bhhj6r5s",
   "tg_tokens": null,
   "tps": 21.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-bhhj6r5s-ctx4096",
   "model": "gemma-4-31B-it-qat-q4_0-unquantized",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: ork17rui; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 236.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-qat-q4_0-unquantized-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 4k 236.3 24.1 19.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhhj6r5s",
   "source_url": "https://omlx.ai/benchmarks/performance/bhhj6r5s",
   "tg_tokens": null,
   "tps": 24.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-j3r10oe6-ctx1024",
   "model": "Gemma 4 E2B",
   "notes": "it, fp16; oMLX community benchmark trace, single run at 1k context; peak memory 4.2 GB; TTFT 488 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2098.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "gemma-4-E2B-it-oQ6e-fp16 on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 6bit; 1k tokens 2,098 PP tok/s 116.9 TG tok/s 488 TTFT (ms) 4.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j3r10oe6",
   "source_url": "https://omlx.ai/benchmarks/performance/j3r10oe6",
   "tg_tokens": null,
   "tps": 116.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-erbuf4n5-ctx1024",
   "model": "CoPaw-Flash-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 211.1,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "CoPaw-Flash-9B-oQ4 on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 211.1 31.5 6.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/erbuf4n5",
   "source_url": "https://omlx.ai/benchmarks/performance/erbuf4n5",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-erbuf4n5-ctx4096",
   "model": "CoPaw-Flash-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.4 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.1,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "CoPaw-Flash-9B-oQ4 on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 224.1 35.2 7.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/erbuf4n5",
   "source_url": "https://omlx.ai/benchmarks/performance/erbuf4n5",
   "tg_tokens": null,
   "tps": 35.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-erbuf4n5-ctx16384",
   "model": "CoPaw-Flash-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.4 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 203.7,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "CoPaw-Flash-9B-oQ4 on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 16k 203.7 25.1 8.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/erbuf4n5",
   "source_url": "https://omlx.ai/benchmarks/performance/erbuf4n5",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-230pv7d6-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: rtb1xm7h; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 106.5 15.5 16.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/230pv7d6",
   "source_url": "https://omlx.ai/benchmarks/performance/230pv7d6",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-230pv7d6-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: rtb1xm7h; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 108.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 108.3 15.1 17.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/230pv7d6",
   "source_url": "https://omlx.ai/benchmarks/performance/230pv7d6",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-6hwil0a7-ctx1024",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: vdtmk81y; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 394.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Alis-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 394.8 51.4 19.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6hwil0a7",
   "source_url": "https://omlx.ai/benchmarks/performance/6hwil0a7",
   "tg_tokens": null,
   "tps": 51.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-6hwil0a7-ctx4096",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: vdtmk81y; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 429.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Alis-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 429.8 41.9 23.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6hwil0a7",
   "source_url": "https://omlx.ai/benchmarks/performance/6hwil0a7",
   "tg_tokens": null,
   "tps": 41.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-6hwil0a7-ctx32768",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: vdtmk81y; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 362.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Alis-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 362.3 67.4 26.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6hwil0a7",
   "source_url": "https://omlx.ai/benchmarks/performance/6hwil0a7",
   "tg_tokens": null,
   "tps": 67.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-mwfhkgpn-ctx1024",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: tktr0q8c, uzqdbehz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 359.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Alis-MLX-6bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 359.8 41.8 27.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mwfhkgpn",
   "source_url": "https://omlx.ai/benchmarks/performance/mwfhkgpn",
   "tg_tokens": null,
   "tps": 41.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-mwfhkgpn-ctx4096",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 31.0 GB at this context; duplicate traces of the same session: tktr0q8c, uzqdbehz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 402.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Alis-MLX-6bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 402.0 39.4 31.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mwfhkgpn",
   "source_url": "https://omlx.ai/benchmarks/performance/mwfhkgpn",
   "tg_tokens": null,
   "tps": 39.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-mwfhkgpn-ctx32768",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 34.1 GB at this context; duplicate traces of the same session: tktr0q8c, uzqdbehz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 347.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Alis-MLX-6bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 347.3 50.0 34.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mwfhkgpn",
   "source_url": "https://omlx.ai/benchmarks/performance/mwfhkgpn",
   "tg_tokens": null,
   "tps": 50.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1jypkjmd-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: y5zaykoc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1177.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,177 79.2 100.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1jypkjmd",
   "source_url": "https://omlx.ai/benchmarks/performance/1jypkjmd",
   "tg_tokens": null,
   "tps": 79.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1jypkjmd-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: y5zaykoc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1423.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,423 83.2 101.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1jypkjmd",
   "source_url": "https://omlx.ai/benchmarks/performance/1jypkjmd",
   "tg_tokens": null,
   "tps": 83.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1jypkjmd-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: y5zaykoc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1411.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 1,411 81.5 101.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1jypkjmd",
   "source_url": "https://omlx.ai/benchmarks/performance/1jypkjmd",
   "tg_tokens": null,
   "tps": 81.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1jypkjmd-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.6 GB at this context; duplicate traces of the same session: y5zaykoc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1399.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 1,399 83.4 101.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1jypkjmd",
   "source_url": "https://omlx.ai/benchmarks/performance/1jypkjmd",
   "tg_tokens": null,
   "tps": 83.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1jypkjmd-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.2 GB at this context; duplicate traces of the same session: y5zaykoc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1388.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 1,388 69.4 102.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1jypkjmd",
   "source_url": "https://omlx.ai/benchmarks/performance/1jypkjmd",
   "tg_tokens": null,
   "tps": 69.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1jypkjmd-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.1 GB at this context; duplicate traces of the same session: y5zaykoc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1365.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 1,365 72.1 104.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1jypkjmd",
   "source_url": "https://omlx.ai/benchmarks/performance/1jypkjmd",
   "tg_tokens": null,
   "tps": 72.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1jypkjmd-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 108.6 GB at this context; duplicate traces of the same session: y5zaykoc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1341.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 1,341 62.1 108.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1jypkjmd",
   "source_url": "https://omlx.ai/benchmarks/performance/1jypkjmd",
   "tg_tokens": null,
   "tps": 62.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1jypkjmd-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.8 GB at this context; duplicate traces of the same session: y5zaykoc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1322.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 195k 1,322 57.6 109.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1jypkjmd",
   "source_url": "https://omlx.ai/benchmarks/performance/1jypkjmd",
   "tg_tokens": null,
   "tps": 57.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wfs3omua-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: t54inkd8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1621.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,621 61.3 71.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wfs3omua",
   "source_url": "https://omlx.ai/benchmarks/performance/wfs3omua",
   "tg_tokens": null,
   "tps": 61.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wfs3omua-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: t54inkd8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2692.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,692 79.1 72.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wfs3omua",
   "source_url": "https://omlx.ai/benchmarks/performance/wfs3omua",
   "tg_tokens": null,
   "tps": 79.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9yzz06qm-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: dyuhhyec; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1089.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,089 68.7 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9yzz06qm",
   "source_url": "https://omlx.ai/benchmarks/performance/9yzz06qm",
   "tg_tokens": null,
   "tps": 68.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9yzz06qm-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: dyuhhyec; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1989.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,989 69.0 70.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9yzz06qm",
   "source_url": "https://omlx.ai/benchmarks/performance/9yzz06qm",
   "tg_tokens": null,
   "tps": 69.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ijmnzmwl-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: 0w9wblom, 7sk0u25o, 8zkil1d9, d6640nve, i1llp18h, u9m6xln6; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2960.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 2,960 139.6 100.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ijmnzmwl",
   "source_url": "https://omlx.ai/benchmarks/performance/ijmnzmwl",
   "tg_tokens": null,
   "tps": 139.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ijmnzmwl-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.0 GB at this context; duplicate traces of the same session: 0w9wblom, 7sk0u25o, 8zkil1d9, d6640nve, i1llp18h, u9m6xln6; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5122.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 5,122 140.2 101.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ijmnzmwl",
   "source_url": "https://omlx.ai/benchmarks/performance/ijmnzmwl",
   "tg_tokens": null,
   "tps": 140.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ijmnzmwl-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.1 GB at this context; duplicate traces of the same session: 0w9wblom, 7sk0u25o, 8zkil1d9, d6640nve, i1llp18h, u9m6xln6; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5654.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 5,654 120.6 102.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ijmnzmwl",
   "source_url": "https://omlx.ai/benchmarks/performance/ijmnzmwl",
   "tg_tokens": null,
   "tps": 120.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ijmnzmwl-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.1 GB at this context; duplicate traces of the same session: 0w9wblom, 7sk0u25o, 8zkil1d9, d6640nve, i1llp18h, u9m6xln6; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5648.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 5,648 102.1 103.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ijmnzmwl",
   "source_url": "https://omlx.ai/benchmarks/performance/ijmnzmwl",
   "tg_tokens": null,
   "tps": 102.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ijmnzmwl-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.6 GB at this context; duplicate traces of the same session: 0w9wblom, 7sk0u25o, 8zkil1d9, d6640nve, i1llp18h, u9m6xln6; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5534.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 5,534 137.4 104.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ijmnzmwl",
   "source_url": "https://omlx.ai/benchmarks/performance/ijmnzmwl",
   "tg_tokens": null,
   "tps": 137.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ijmnzmwl-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.0 GB at this context; duplicate traces of the same session: 0w9wblom, 7sk0u25o, 8zkil1d9, d6640nve, i1llp18h, u9m6xln6; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5360.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,360 109.7 109.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ijmnzmwl",
   "source_url": "https://omlx.ai/benchmarks/performance/ijmnzmwl",
   "tg_tokens": null,
   "tps": 109.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ijmnzmwl-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.7 GB at this context; duplicate traces of the same session: 0w9wblom, 7sk0u25o, 8zkil1d9, d6640nve, i1llp18h, u9m6xln6; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5203.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 5,203 95.4 110.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ijmnzmwl",
   "source_url": "https://omlx.ai/benchmarks/performance/ijmnzmwl",
   "tg_tokens": null,
   "tps": 95.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-s92fzpwv-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.4 GB at this context; duplicate traces of the same session: 0djkh7lw, e0hajlv3, ib7ip7em, q2xxr069, r2wonn4c; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3121.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 3,121 139.0 100.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s92fzpwv",
   "source_url": "https://omlx.ai/benchmarks/performance/s92fzpwv",
   "tg_tokens": null,
   "tps": 139.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-s92fzpwv-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.2 GB at this context; duplicate traces of the same session: 0djkh7lw, e0hajlv3, ib7ip7em, q2xxr069, r2wonn4c; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6233.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 6,233 121.0 102.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s92fzpwv",
   "source_url": "https://omlx.ai/benchmarks/performance/s92fzpwv",
   "tg_tokens": null,
   "tps": 121.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-s92fzpwv-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.2 GB at this context; duplicate traces of the same session: 0djkh7lw, e0hajlv3, ib7ip7em, q2xxr069, r2wonn4c; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6229.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 6,229 126.0 103.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s92fzpwv",
   "source_url": "https://omlx.ai/benchmarks/performance/s92fzpwv",
   "tg_tokens": null,
   "tps": 126.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-s92fzpwv-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.7 GB at this context; duplicate traces of the same session: 0djkh7lw, e0hajlv3, ib7ip7em, q2xxr069, r2wonn4c; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6089.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 6,089 137.7 104.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s92fzpwv",
   "source_url": "https://omlx.ai/benchmarks/performance/s92fzpwv",
   "tg_tokens": null,
   "tps": 137.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-s92fzpwv-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.1 GB at this context; duplicate traces of the same session: 0djkh7lw, e0hajlv3, ib7ip7em, q2xxr069, r2wonn4c; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5891.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,891 108.8 109.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s92fzpwv",
   "source_url": "https://omlx.ai/benchmarks/performance/s92fzpwv",
   "tg_tokens": null,
   "tps": 108.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-s92fzpwv-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.7 GB at this context; duplicate traces of the same session: 0djkh7lw, e0hajlv3, ib7ip7em, q2xxr069, r2wonn4c; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5709.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 5,709 86.0 110.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s92fzpwv",
   "source_url": "https://omlx.ai/benchmarks/performance/s92fzpwv",
   "tg_tokens": null,
   "tps": 86.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-vfm01q20-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.6 GB at this context; duplicate traces of the same session: gwz9gylf, l2203weo, n0q98vml, nbmlod65; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1949.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 1k 1,949 125.5 142.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vfm01q20",
   "source_url": "https://omlx.ai/benchmarks/performance/vfm01q20",
   "tg_tokens": null,
   "tps": 125.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-vfm01q20-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: gwz9gylf, l2203weo, n0q98vml, nbmlod65; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3512.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 4k 3,512 118.4 143.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vfm01q20",
   "source_url": "https://omlx.ai/benchmarks/performance/vfm01q20",
   "tg_tokens": null,
   "tps": 118.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-vfm01q20-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 144.5 GB at this context; duplicate traces of the same session: gwz9gylf, l2203weo, n0q98vml, nbmlod65; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3970.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 8k 3,970 110.0 144.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vfm01q20",
   "source_url": "https://omlx.ai/benchmarks/performance/vfm01q20",
   "tg_tokens": null,
   "tps": 110.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-vfm01q20-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 146.8 GB at this context; duplicate traces of the same session: gwz9gylf, l2203weo, n0q98vml, nbmlod65; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3905.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 64k 3,905 110.8 146.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vfm01q20",
   "source_url": "https://omlx.ai/benchmarks/performance/vfm01q20",
   "tg_tokens": null,
   "tps": 110.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-vfm01q20-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 150.9 GB at this context; duplicate traces of the same session: gwz9gylf, l2203weo, n0q98vml, nbmlod65; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3808.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 128k 3,808 94.1 150.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vfm01q20",
   "source_url": "https://omlx.ai/benchmarks/performance/vfm01q20",
   "tg_tokens": null,
   "tps": 94.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-vfm01q20-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 153.3 GB at this context; duplicate traces of the same session: gwz9gylf, l2203weo, n0q98vml, nbmlod65; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3702.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 195k 3,702 81.7 153.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vfm01q20",
   "source_url": "https://omlx.ai/benchmarks/performance/vfm01q20",
   "tg_tokens": null,
   "tps": 81.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-18u6pe4m-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.6 GB at this context; duplicate traces of the same session: 01haat69, 0pc5veus, fnkji6qk, qk6v28sv, u72ko6df, w0oz4okn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2431.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 2,431 144.9 142.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/18u6pe4m",
   "source_url": "https://omlx.ai/benchmarks/performance/18u6pe4m",
   "tg_tokens": null,
   "tps": 144.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-18u6pe4m-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: 01haat69, 0pc5veus, fnkji6qk, qk6v28sv, u72ko6df, w0oz4okn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4445.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 4,445 131.4 143.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/18u6pe4m",
   "source_url": "https://omlx.ai/benchmarks/performance/18u6pe4m",
   "tg_tokens": null,
   "tps": 131.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-18u6pe4m-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 144.4 GB at this context; duplicate traces of the same session: 01haat69, 0pc5veus, fnkji6qk, qk6v28sv, u72ko6df, w0oz4okn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4864.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 8k 4,864 134.7 144.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/18u6pe4m",
   "source_url": "https://omlx.ai/benchmarks/performance/18u6pe4m",
   "tg_tokens": null,
   "tps": 134.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-18u6pe4m-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.1 GB at this context; duplicate traces of the same session: 01haat69, 0pc5veus, fnkji6qk, qk6v28sv, u72ko6df, w0oz4okn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4880.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 16k 4,880 128.0 145.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/18u6pe4m",
   "source_url": "https://omlx.ai/benchmarks/performance/18u6pe4m",
   "tg_tokens": null,
   "tps": 128.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-18u6pe4m-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 146.8 GB at this context; duplicate traces of the same session: 01haat69, 0pc5veus, fnkji6qk, qk6v28sv, u72ko6df, w0oz4okn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4781.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 64k 4,781 116.7 146.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/18u6pe4m",
   "source_url": "https://omlx.ai/benchmarks/performance/18u6pe4m",
   "tg_tokens": null,
   "tps": 116.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-18u6pe4m-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 150.6 GB at this context; duplicate traces of the same session: 01haat69, 0pc5veus, fnkji6qk, qk6v28sv, u72ko6df, w0oz4okn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4649.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 128k 4,649 102.5 150.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/18u6pe4m",
   "source_url": "https://omlx.ai/benchmarks/performance/18u6pe4m",
   "tg_tokens": null,
   "tps": 102.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-18u6pe4m-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 153.3 GB at this context; duplicate traces of the same session: 01haat69, 0pc5veus, fnkji6qk, qk6v28sv, u72ko6df, w0oz4okn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4528.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 195k 4,528 91.5 153.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/18u6pe4m",
   "source_url": "https://omlx.ai/benchmarks/performance/18u6pe4m",
   "tg_tokens": null,
   "tps": 91.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-fq1207kc-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.6 GB at this context; duplicate traces of the same session: 30g14l80, iscfu1am, nxlzhdt7, w993h1n1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2453.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 2,453 148.0 142.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fq1207kc",
   "source_url": "https://omlx.ai/benchmarks/performance/fq1207kc",
   "tg_tokens": null,
   "tps": 148.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-fq1207kc-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: 30g14l80, iscfu1am, nxlzhdt7, w993h1n1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4439.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 4,439 131.6 143.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fq1207kc",
   "source_url": "https://omlx.ai/benchmarks/performance/fq1207kc",
   "tg_tokens": null,
   "tps": 131.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-fq1207kc-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 146.8 GB at this context; duplicate traces of the same session: 30g14l80, iscfu1am, nxlzhdt7, w993h1n1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4788.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 64k 4,788 129.0 146.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fq1207kc",
   "source_url": "https://omlx.ai/benchmarks/performance/fq1207kc",
   "tg_tokens": null,
   "tps": 129.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-fq1207kc-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 150.6 GB at this context; duplicate traces of the same session: 30g14l80, iscfu1am, nxlzhdt7, w993h1n1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4654.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 128k 4,654 103.8 150.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fq1207kc",
   "source_url": "https://omlx.ai/benchmarks/performance/fq1207kc",
   "tg_tokens": null,
   "tps": 103.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-fq1207kc-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 153.3 GB at this context; duplicate traces of the same session: 30g14l80, iscfu1am, nxlzhdt7, w993h1n1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4539.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 195k 4,539 87.1 153.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fq1207kc",
   "source_url": "https://omlx.ai/benchmarks/performance/fq1207kc",
   "tg_tokens": null,
   "tps": 87.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ffl9vl3c-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; Swift 1.5 build; oMLX community benchmark trace; peak memory 75.9 GB at this context; duplicate traces of the same session: 815rud9x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 608.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 608.0 19.6 75.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ffl9vl3c",
   "source_url": "https://omlx.ai/benchmarks/performance/ffl9vl3c",
   "tg_tokens": null,
   "tps": 19.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ffl9vl3c-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; Swift 1.5 build; oMLX community benchmark trace; peak memory 78.2 GB at this context; duplicate traces of the same session: 815rud9x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 826.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 826.1 22.8 78.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ffl9vl3c",
   "source_url": "https://omlx.ai/benchmarks/performance/ffl9vl3c",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-rnv1ej50-ctx1024",
   "model": "Qwen3-VL-4B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 2.8 GB at this context; duplicate traces of the same session: none; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 326.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-VL-4B-Instruct-MLX-4bit on M4 (8c); M4 (8c) \u00b7 16 GB \u00b7 4bit; 1k 326.7 39.6 2.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rnv1ej50",
   "source_url": "https://omlx.ai/benchmarks/performance/rnv1ej50",
   "tg_tokens": null,
   "tps": 39.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-rnv1ej50-ctx4096",
   "model": "Qwen3-VL-4B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 3.7 GB at this context; duplicate traces of the same session: none; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 292.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-VL-4B-Instruct-MLX-4bit on M4 (8c); M4 (8c) \u00b7 16 GB \u00b7 4bit; 4k 292.4 26.0 3.7 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rnv1ej50",
   "source_url": "https://omlx.ai/benchmarks/performance/rnv1ej50",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-02e21zd4-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 7.1 GB; TTFT 49495 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 331.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 16k tokens 331.0 PP tok/s 50.7 TG tok/s 49495 TTFT (ms) 7.1 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/02e21zd4",
   "source_url": "https://omlx.ai/benchmarks/performance/02e21zd4",
   "tg_tokens": null,
   "tps": 50.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-svs9pchi-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: i2ami32w; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 106.2 15.5 16.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/svs9pchi",
   "source_url": "https://omlx.ai/benchmarks/performance/svs9pchi",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-svs9pchi-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: i2ami32w; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 108.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 108.8 15.2 17.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/svs9pchi",
   "source_url": "https://omlx.ai/benchmarks/performance/svs9pchi",
   "tg_tokens": null,
   "tps": 15.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-by039pzu-ctx1024",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 394.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Alis-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 394.8 51.4 19.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/by039pzu",
   "source_url": "https://omlx.ai/benchmarks/performance/by039pzu",
   "tg_tokens": null,
   "tps": 51.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-by039pzu-ctx4096",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 429.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Alis-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 429.8 41.9 23.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/by039pzu",
   "source_url": "https://omlx.ai/benchmarks/performance/by039pzu",
   "tg_tokens": null,
   "tps": 41.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-by039pzu-ctx32768",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 362.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Alis-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 362.3 67.4 26.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/by039pzu",
   "source_url": "https://omlx.ai/benchmarks/performance/by039pzu",
   "tg_tokens": null,
   "tps": 67.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-3vu886j2-ctx1024",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: 25tcofl5, 5hevl75t, 6trks8vf, ony35duu, vnahnlgj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 413.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 413.5 21.6 33.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3vu886j2",
   "source_url": "https://omlx.ai/benchmarks/performance/3vu886j2",
   "tg_tokens": null,
   "tps": 21.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-3vu886j2-ctx4096",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: 25tcofl5, 5hevl75t, 6trks8vf, ony35duu, vnahnlgj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 455.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 455.2 32.4 34.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3vu886j2",
   "source_url": "https://omlx.ai/benchmarks/performance/3vu886j2",
   "tg_tokens": null,
   "tps": 32.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3vu886j2-ctx8192",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: 25tcofl5, 5hevl75t, 6trks8vf, ony35duu, vnahnlgj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 409.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 409.8 15.6 35.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3vu886j2",
   "source_url": "https://omlx.ai/benchmarks/performance/3vu886j2",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3vu886j2-ctx16384",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: 25tcofl5, 5hevl75t, 6trks8vf, ony35duu, vnahnlgj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 378.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 16k 378.1 19.8 35.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3vu886j2",
   "source_url": "https://omlx.ai/benchmarks/performance/3vu886j2",
   "tg_tokens": null,
   "tps": 19.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3vu886j2-ctx32768",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: 25tcofl5, 5hevl75t, 6trks8vf, ony35duu, vnahnlgj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 309.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 32k 309.2 22.6 36.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3vu886j2",
   "source_url": "https://omlx.ai/benchmarks/performance/3vu886j2",
   "tg_tokens": null,
   "tps": 22.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3vu886j2-ctx65536",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: 25tcofl5, 5hevl75t, 6trks8vf, ony35duu, vnahnlgj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 260.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 64k 260.9 16.5 39.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3vu886j2",
   "source_url": "https://omlx.ai/benchmarks/performance/3vu886j2",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-nc7dju5p-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 38.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 227.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 227.6 35.8 38.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nc7dju5p",
   "source_url": "https://omlx.ai/benchmarks/performance/nc7dju5p",
   "tg_tokens": null,
   "tps": 35.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-nc7dju5p-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 46.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 180.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 180.8 25.0 46.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nc7dju5p",
   "source_url": "https://omlx.ai/benchmarks/performance/nc7dju5p",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-nc7dju5p-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 62.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 127.7 24.3 62.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nc7dju5p",
   "source_url": "https://omlx.ai/benchmarks/performance/nc7dju5p",
   "tg_tokens": null,
   "tps": 24.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-4qa2cv7g-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 77.9 GB; TTFT 167945 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 780.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k tokens 780.4 PP tok/s 36.8 TG tok/s 167945 TTFT (ms) 77.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4qa2cv7g",
   "source_url": "https://omlx.ai/benchmarks/performance/4qa2cv7g",
   "tg_tokens": null,
   "tps": 36.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-plkggfly-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: 88tz9y80, fihdrxxt, vwvot3y1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1009.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,009 56.6 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/plkggfly",
   "source_url": "https://omlx.ai/benchmarks/performance/plkggfly",
   "tg_tokens": null,
   "tps": 56.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-plkggfly-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: 88tz9y80, fihdrxxt, vwvot3y1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2044.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,044 59.5 71.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/plkggfly",
   "source_url": "https://omlx.ai/benchmarks/performance/plkggfly",
   "tg_tokens": null,
   "tps": 59.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-plkggfly-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 88tz9y80, fihdrxxt, vwvot3y1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2576.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,576 77.6 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/plkggfly",
   "source_url": "https://omlx.ai/benchmarks/performance/plkggfly",
   "tg_tokens": null,
   "tps": 77.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-plkggfly-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 76.0 GB at this context; duplicate traces of the same session: 88tz9y80, fihdrxxt, vwvot3y1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2549.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,549 69.3 76.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/plkggfly",
   "source_url": "https://omlx.ai/benchmarks/performance/plkggfly",
   "tg_tokens": null,
   "tps": 69.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-vr9k5ltg-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: 19wsgmu0, 6aivsjri, bebd9eqe, kci18jry, ttz0vd83, wtdnwgu0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3077.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 3,077 150.5 100.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vr9k5ltg",
   "source_url": "https://omlx.ai/benchmarks/performance/vr9k5ltg",
   "tg_tokens": null,
   "tps": 150.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-vr9k5ltg-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: 19wsgmu0, 6aivsjri, bebd9eqe, kci18jry, ttz0vd83, wtdnwgu0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5575.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 5,575 69.8 101.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vr9k5ltg",
   "source_url": "https://omlx.ai/benchmarks/performance/vr9k5ltg",
   "tg_tokens": null,
   "tps": 69.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-vr9k5ltg-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.2 GB at this context; duplicate traces of the same session: 19wsgmu0, 6aivsjri, bebd9eqe, kci18jry, ttz0vd83, wtdnwgu0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6224.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 6,224 140.1 102.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vr9k5ltg",
   "source_url": "https://omlx.ai/benchmarks/performance/vr9k5ltg",
   "tg_tokens": null,
   "tps": 140.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-vr9k5ltg-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.2 GB at this context; duplicate traces of the same session: 19wsgmu0, 6aivsjri, bebd9eqe, kci18jry, ttz0vd83, wtdnwgu0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6203.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 6,203 173.5 103.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vr9k5ltg",
   "source_url": "https://omlx.ai/benchmarks/performance/vr9k5ltg",
   "tg_tokens": null,
   "tps": 173.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-vr9k5ltg-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.7 GB at this context; duplicate traces of the same session: 19wsgmu0, 6aivsjri, bebd9eqe, kci18jry, ttz0vd83, wtdnwgu0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6049.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 6,049 135.0 104.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vr9k5ltg",
   "source_url": "https://omlx.ai/benchmarks/performance/vr9k5ltg",
   "tg_tokens": null,
   "tps": 135.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-vr9k5ltg-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.1 GB at this context; duplicate traces of the same session: 19wsgmu0, 6aivsjri, bebd9eqe, kci18jry, ttz0vd83, wtdnwgu0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5853.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,853 109.1 109.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vr9k5ltg",
   "source_url": "https://omlx.ai/benchmarks/performance/vr9k5ltg",
   "tg_tokens": null,
   "tps": 109.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-vr9k5ltg-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.7 GB at this context; duplicate traces of the same session: 19wsgmu0, 6aivsjri, bebd9eqe, kci18jry, ttz0vd83, wtdnwgu0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5674.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 5,674 96.3 110.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vr9k5ltg",
   "source_url": "https://omlx.ai/benchmarks/performance/vr9k5ltg",
   "tg_tokens": null,
   "tps": 96.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-t8as5tqz-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.6 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1949.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 1k 1,949 125.5 142.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t8as5tqz",
   "source_url": "https://omlx.ai/benchmarks/performance/t8as5tqz",
   "tg_tokens": null,
   "tps": 125.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-t8as5tqz-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3512.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 4k 3,512 118.4 143.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t8as5tqz",
   "source_url": "https://omlx.ai/benchmarks/performance/t8as5tqz",
   "tg_tokens": null,
   "tps": 118.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-t8as5tqz-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 144.5 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3970.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 8k 3,970 110.0 144.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t8as5tqz",
   "source_url": "https://omlx.ai/benchmarks/performance/t8as5tqz",
   "tg_tokens": null,
   "tps": 110.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-t8as5tqz-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 146.8 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3905.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 64k 3,905 110.8 146.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t8as5tqz",
   "source_url": "https://omlx.ai/benchmarks/performance/t8as5tqz",
   "tg_tokens": null,
   "tps": 110.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-t8as5tqz-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 150.9 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3808.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 128k 3,808 94.1 150.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t8as5tqz",
   "source_url": "https://omlx.ai/benchmarks/performance/t8as5tqz",
   "tg_tokens": null,
   "tps": 94.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-t8as5tqz-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 153.3 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3702.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 195k 3,702 81.7 153.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t8as5tqz",
   "source_url": "https://omlx.ai/benchmarks/performance/t8as5tqz",
   "tg_tokens": null,
   "tps": 81.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-yxia92aa-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.6 GB at this context; duplicate traces of the same session: go1ri7sk, ix7ihlwx, oe2oe1m5, uqq9week, v2c9tw5t; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2465.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 2,465 137.7 142.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yxia92aa",
   "source_url": "https://omlx.ai/benchmarks/performance/yxia92aa",
   "tg_tokens": null,
   "tps": 137.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-yxia92aa-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: go1ri7sk, ix7ihlwx, oe2oe1m5, uqq9week, v2c9tw5t; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4428.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 4,428 146.1 143.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yxia92aa",
   "source_url": "https://omlx.ai/benchmarks/performance/yxia92aa",
   "tg_tokens": null,
   "tps": 146.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-yxia92aa-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 144.3 GB at this context; duplicate traces of the same session: go1ri7sk, ix7ihlwx, oe2oe1m5, uqq9week, v2c9tw5t; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4836.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 8k 4,836 130.9 144.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yxia92aa",
   "source_url": "https://omlx.ai/benchmarks/performance/yxia92aa",
   "tg_tokens": null,
   "tps": 130.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-yxia92aa-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 146.8 GB at this context; duplicate traces of the same session: go1ri7sk, ix7ihlwx, oe2oe1m5, uqq9week, v2c9tw5t; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4761.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 64k 4,761 129.1 146.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yxia92aa",
   "source_url": "https://omlx.ai/benchmarks/performance/yxia92aa",
   "tg_tokens": null,
   "tps": 129.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-yxia92aa-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 150.6 GB at this context; duplicate traces of the same session: go1ri7sk, ix7ihlwx, oe2oe1m5, uqq9week, v2c9tw5t; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4627.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 128k 4,627 102.0 150.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yxia92aa",
   "source_url": "https://omlx.ai/benchmarks/performance/yxia92aa",
   "tg_tokens": null,
   "tps": 102.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-yxia92aa-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 153.3 GB at this context; duplicate traces of the same session: go1ri7sk, ix7ihlwx, oe2oe1m5, uqq9week, v2c9tw5t; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4517.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 195k 4,517 82.5 153.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yxia92aa",
   "source_url": "https://omlx.ai/benchmarks/performance/yxia92aa",
   "tg_tokens": null,
   "tps": 82.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-9beph7t6-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.6 GB at this context; duplicate traces of the same session: 0prir3zi, bc1bfc4q, fygv6jfl, nbyjlg1b, uiowe12d; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2431.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 2,431 150.1 142.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9beph7t6",
   "source_url": "https://omlx.ai/benchmarks/performance/9beph7t6",
   "tg_tokens": null,
   "tps": 150.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-9beph7t6-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: 0prir3zi, bc1bfc4q, fygv6jfl, nbyjlg1b, uiowe12d; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4422.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 4,422 126.8 143.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9beph7t6",
   "source_url": "https://omlx.ai/benchmarks/performance/9beph7t6",
   "tg_tokens": null,
   "tps": 126.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-9beph7t6-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 144.3 GB at this context; duplicate traces of the same session: 0prir3zi, bc1bfc4q, fygv6jfl, nbyjlg1b, uiowe12d; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4827.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 8k 4,827 133.9 144.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9beph7t6",
   "source_url": "https://omlx.ai/benchmarks/performance/9beph7t6",
   "tg_tokens": null,
   "tps": 133.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-9beph7t6-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 146.8 GB at this context; duplicate traces of the same session: 0prir3zi, bc1bfc4q, fygv6jfl, nbyjlg1b, uiowe12d; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4753.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 64k 4,753 124.1 146.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9beph7t6",
   "source_url": "https://omlx.ai/benchmarks/performance/9beph7t6",
   "tg_tokens": null,
   "tps": 124.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-9beph7t6-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 150.6 GB at this context; duplicate traces of the same session: 0prir3zi, bc1bfc4q, fygv6jfl, nbyjlg1b, uiowe12d; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4613.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 128k 4,613 103.3 150.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9beph7t6",
   "source_url": "https://omlx.ai/benchmarks/performance/9beph7t6",
   "tg_tokens": null,
   "tps": 103.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-9beph7t6-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 153.3 GB at this context; duplicate traces of the same session: 0prir3zi, bc1bfc4q, fygv6jfl, nbyjlg1b, uiowe12d; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4505.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 195k 4,505 89.8 153.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9beph7t6",
   "source_url": "https://omlx.ai/benchmarks/performance/9beph7t6",
   "tg_tokens": null,
   "tps": 89.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ugdh24yd-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: m2y1kvyc; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1864.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,864 120.9 26.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ugdh24yd",
   "source_url": "https://omlx.ai/benchmarks/performance/ugdh24yd",
   "tg_tokens": null,
   "tps": 120.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ugdh24yd-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace; peak memory 28.2 GB at this context; duplicate traces of the same session: m2y1kvyc; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1545.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,545 32.0 28.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ugdh24yd",
   "source_url": "https://omlx.ai/benchmarks/performance/ugdh24yd",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-flfoh7l4-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: vvth2er8; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1370.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,370 71.9 26.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/flfoh7l4",
   "source_url": "https://omlx.ai/benchmarks/performance/flfoh7l4",
   "tg_tokens": null,
   "tps": 71.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-flfoh7l4-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace; peak memory 28.2 GB at this context; duplicate traces of the same session: vvth2er8; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1190.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,190 32.1 28.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/flfoh7l4",
   "source_url": "https://omlx.ai/benchmarks/performance/flfoh7l4",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-9v0hj33g-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1296.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 1,296 82.2 25.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9v0hj33g",
   "source_url": "https://omlx.ai/benchmarks/performance/9v0hj33g",
   "tg_tokens": null,
   "tps": 82.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-9v0hj33g-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1415.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 1,415 80.0 26.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9v0hj33g",
   "source_url": "https://omlx.ai/benchmarks/performance/9v0hj33g",
   "tg_tokens": null,
   "tps": 80.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-zs5mhgiu-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: hrw0n68b; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2079.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 2,079 76.8 14.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zs5mhgiu",
   "source_url": "https://omlx.ai/benchmarks/performance/zs5mhgiu",
   "tg_tokens": null,
   "tps": 76.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-zs5mhgiu-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: hrw0n68b; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2342.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 2,342 71.8 14.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zs5mhgiu",
   "source_url": "https://omlx.ai/benchmarks/performance/zs5mhgiu",
   "tg_tokens": null,
   "tps": 71.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-pqxfp6z6-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: bdsecqhv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1046.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26b-a4b-it-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 8bit; 1k 1,046 54.6 25.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pqxfp6z6",
   "source_url": "https://omlx.ai/benchmarks/performance/pqxfp6z6",
   "tg_tokens": null,
   "tps": 54.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-pqxfp6z6-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: bdsecqhv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2190.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26b-a4b-it-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 8bit; 4k 2,190 51.7 26.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pqxfp6z6",
   "source_url": "https://omlx.ai/benchmarks/performance/pqxfp6z6",
   "tg_tokens": null,
   "tps": 51.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-pbmy1klr-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "fp16; oMLX community benchmark trace; peak memory 14.5 GB at this context; duplicate traces of the same session: ipg2ovdk, kwldee68, p0w1q9a5, pbtk10qh, qk0gy00y; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 677.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-9B-oQ8e-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 677.8 37.3 14.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pbmy1klr",
   "source_url": "https://omlx.ai/benchmarks/performance/pbmy1klr",
   "tg_tokens": null,
   "tps": 37.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-pbmy1klr-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "fp16; oMLX community benchmark trace; peak memory 15.6 GB at this context; duplicate traces of the same session: ipg2ovdk, kwldee68, p0w1q9a5, pbtk10qh, qk0gy00y; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 919.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-9B-oQ8e-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 919.8 36.4 15.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pbmy1klr",
   "source_url": "https://omlx.ai/benchmarks/performance/pbmy1klr",
   "tg_tokens": null,
   "tps": 36.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-pbmy1klr-ctx8192",
   "model": "Ornith-1.5-9B",
   "notes": "fp16; oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: ipg2ovdk, kwldee68, p0w1q9a5, pbtk10qh, qk0gy00y; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 929.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-9B-oQ8e-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 8k 929.9 35.1 16.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pbmy1klr",
   "source_url": "https://omlx.ai/benchmarks/performance/pbmy1klr",
   "tg_tokens": null,
   "tps": 35.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-pbmy1klr-ctx16384",
   "model": "Ornith-1.5-9B",
   "notes": "fp16; oMLX community benchmark trace; peak memory 16.8 GB at this context; duplicate traces of the same session: ipg2ovdk, kwldee68, p0w1q9a5, pbtk10qh, qk0gy00y; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 871.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-9B-oQ8e-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 16k 871.6 32.4 16.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pbmy1klr",
   "source_url": "https://omlx.ai/benchmarks/performance/pbmy1klr",
   "tg_tokens": null,
   "tps": 32.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-pbmy1klr-ctx32768",
   "model": "Ornith-1.5-9B",
   "notes": "fp16; oMLX community benchmark trace; peak memory 18.2 GB at this context; duplicate traces of the same session: ipg2ovdk, kwldee68, p0w1q9a5, pbtk10qh, qk0gy00y; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 761.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-9B-oQ8e-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 32k 761.7 28.7 18.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pbmy1klr",
   "source_url": "https://omlx.ai/benchmarks/performance/pbmy1klr",
   "tg_tokens": null,
   "tps": 28.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-pbmy1klr-ctx65536",
   "model": "Ornith-1.5-9B",
   "notes": "fp16; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: ipg2ovdk, kwldee68, p0w1q9a5, pbtk10qh, qk0gy00y; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 620.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-9B-oQ8e-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 64k 620.2 23.0 19.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pbmy1klr",
   "source_url": "https://omlx.ai/benchmarks/performance/pbmy1klr",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-9339fv22-ctx1024",
   "model": "Qwen3-VL-4B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 2.8 GB at this context; duplicate traces of the same session: none; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 326.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-VL-4B-Instruct-MLX-4bit on M4 (8c); M4 (8c) \u00b7 16 GB \u00b7 4bit; 1k 326.7 39.6 2.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9339fv22",
   "source_url": "https://omlx.ai/benchmarks/performance/9339fv22",
   "tg_tokens": null,
   "tps": 39.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-9339fv22-ctx4096",
   "model": "Qwen3-VL-4B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 3.7 GB at this context; duplicate traces of the same session: none; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 292.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-VL-4B-Instruct-MLX-4bit on M4 (8c); M4 (8c) \u00b7 16 GB \u00b7 4bit; 4k 292.4 26.0 3.7 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9339fv22",
   "source_url": "https://omlx.ai/benchmarks/performance/9339fv22",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-epa4bdmh-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.8 GB at this context; duplicate traces of the same session: 6cr3osji; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 413.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 413.9 31.3 3.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/epa4bdmh",
   "source_url": "https://omlx.ai/benchmarks/performance/epa4bdmh",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-epa4bdmh-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.4 GB at this context; duplicate traces of the same session: 6cr3osji; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 371.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 16k 371.1 27.6 4.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/epa4bdmh",
   "source_url": "https://omlx.ai/benchmarks/performance/epa4bdmh",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-v3up2nns-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 8k context; peak memory 24.8 GB; TTFT 1430 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5728.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k tokens 5,728 PP tok/s 102.5 TG tok/s 1430 TTFT (ms) 24.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v3up2nns",
   "source_url": "https://omlx.ai/benchmarks/performance/v3up2nns",
   "tg_tokens": null,
   "tps": 102.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mm7ic8v9-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 22.1 GB; TTFT 1352 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6058.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k tokens 6,058 PP tok/s 166.8 TG tok/s 1352 TTFT (ms) 22.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mm7ic8v9",
   "source_url": "https://omlx.ai/benchmarks/performance/mm7ic8v9",
   "tg_tokens": null,
   "tps": 166.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-26zzelzf-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 22.1 GB; TTFT 1418 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5777.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k tokens 5,777 PP tok/s 219.7 TG tok/s 1418 TTFT (ms) 22.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/26zzelzf",
   "source_url": "https://omlx.ai/benchmarks/performance/26zzelzf",
   "tg_tokens": null,
   "tps": 219.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2nrjydy4-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 22.1 GB; TTFT 1419 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5774.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k tokens 5,774 PP tok/s 208.5 TG tok/s 1419 TTFT (ms) 22.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2nrjydy4",
   "source_url": "https://omlx.ai/benchmarks/performance/2nrjydy4",
   "tg_tokens": null,
   "tps": 208.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-64fif55d-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 38.6 GB at this context; duplicate traces of the same session: 7imdusoi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 227.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 227.6 35.8 38.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/64fif55d",
   "source_url": "https://omlx.ai/benchmarks/performance/64fif55d",
   "tg_tokens": null,
   "tps": 35.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-64fif55d-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 46.6 GB at this context; duplicate traces of the same session: 7imdusoi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 180.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 180.8 25.0 46.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/64fif55d",
   "source_url": "https://omlx.ai/benchmarks/performance/64fif55d",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-64fif55d-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 62.5 GB at this context; duplicate traces of the same session: 7imdusoi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 127.7 24.3 62.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/64fif55d",
   "source_url": "https://omlx.ai/benchmarks/performance/64fif55d",
   "tg_tokens": null,
   "tps": 24.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-601c1lor-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.1 GB at this context; duplicate traces of the same session: 25tt9dcp, ci8u7wcz, fs4itnsm, qmcxgma7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5630.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 5,630 139.3 102.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/601c1lor",
   "source_url": "https://omlx.ai/benchmarks/performance/601c1lor",
   "tg_tokens": null,
   "tps": 139.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-601c1lor-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.1 GB at this context; duplicate traces of the same session: 25tt9dcp, ci8u7wcz, fs4itnsm, qmcxgma7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5614.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 5,614 137.6 103.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/601c1lor",
   "source_url": "https://omlx.ai/benchmarks/performance/601c1lor",
   "tg_tokens": null,
   "tps": 137.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-601c1lor-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 105.0 GB at this context; duplicate traces of the same session: 25tt9dcp, ci8u7wcz, fs4itnsm, qmcxgma7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5511.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 5,511 139.6 105.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/601c1lor",
   "source_url": "https://omlx.ai/benchmarks/performance/601c1lor",
   "tg_tokens": null,
   "tps": 139.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-601c1lor-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.0 GB at this context; duplicate traces of the same session: 25tt9dcp, ci8u7wcz, fs4itnsm, qmcxgma7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5332.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,332 113.1 109.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/601c1lor",
   "source_url": "https://omlx.ai/benchmarks/performance/601c1lor",
   "tg_tokens": null,
   "tps": 113.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-601c1lor-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.7 GB at this context; duplicate traces of the same session: 25tt9dcp, ci8u7wcz, fs4itnsm, qmcxgma7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5176.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 5,176 95.4 110.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/601c1lor",
   "source_url": "https://omlx.ai/benchmarks/performance/601c1lor",
   "tg_tokens": null,
   "tps": 95.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-wlw0n4eb-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.9 GB at this context; duplicate traces of the same session: fpa3rudc, h7i9be7l, w0vm73os; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6039.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 6,039 153.9 102.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wlw0n4eb",
   "source_url": "https://omlx.ai/benchmarks/performance/wlw0n4eb",
   "tg_tokens": null,
   "tps": 153.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-wlw0n4eb-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 105.0 GB at this context; duplicate traces of the same session: fpa3rudc, h7i9be7l, w0vm73os; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6064.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 6,064 131.9 105.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wlw0n4eb",
   "source_url": "https://omlx.ai/benchmarks/performance/wlw0n4eb",
   "tg_tokens": null,
   "tps": 131.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-wlw0n4eb-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.5 GB at this context; duplicate traces of the same session: fpa3rudc, h7i9be7l, w0vm73os; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5838.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,838 116.5 109.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wlw0n4eb",
   "source_url": "https://omlx.ai/benchmarks/performance/wlw0n4eb",
   "tg_tokens": null,
   "tps": 116.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-wlw0n4eb-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 111.1 GB at this context; duplicate traces of the same session: fpa3rudc, h7i9be7l, w0vm73os; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5653.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 5,653 92.7 111.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wlw0n4eb",
   "source_url": "https://omlx.ai/benchmarks/performance/wlw0n4eb",
   "tg_tokens": null,
   "tps": 92.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-xkkpqdze-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.9 GB at this context; duplicate traces of the same session: su77qgma; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5912.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 5,912 141.4 102.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xkkpqdze",
   "source_url": "https://omlx.ai/benchmarks/performance/xkkpqdze",
   "tg_tokens": null,
   "tps": 141.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-xkkpqdze-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 105.1 GB at this context; duplicate traces of the same session: su77qgma; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6056.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 6,056 144.2 105.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xkkpqdze",
   "source_url": "https://omlx.ai/benchmarks/performance/xkkpqdze",
   "tg_tokens": null,
   "tps": 144.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-xkkpqdze-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.4 GB at this context; duplicate traces of the same session: su77qgma; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5838.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,838 107.0 109.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xkkpqdze",
   "source_url": "https://omlx.ai/benchmarks/performance/xkkpqdze",
   "tg_tokens": null,
   "tps": 107.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-xkkpqdze-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 111.1 GB at this context; duplicate traces of the same session: su77qgma; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5657.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 5,657 94.4 111.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xkkpqdze",
   "source_url": "https://omlx.ai/benchmarks/performance/xkkpqdze",
   "tg_tokens": null,
   "tps": 94.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-7ulqm1c1-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 122.2 GB at this context; duplicate traces of the same session: 14pxnmby, 63qsq1zr, 6630l2jh, nat0wb8x, q357fb1w; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 862.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 1k 862.2 49.6 122.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ulqm1c1",
   "source_url": "https://omlx.ai/benchmarks/performance/7ulqm1c1",
   "tg_tokens": null,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-7ulqm1c1-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 123.0 GB at this context; duplicate traces of the same session: 14pxnmby, 63qsq1zr, 6630l2jh, nat0wb8x, q357fb1w; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1040.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 8k 1,040 76.7 123.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ulqm1c1",
   "source_url": "https://omlx.ai/benchmarks/performance/7ulqm1c1",
   "tg_tokens": null,
   "tps": 76.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-7ulqm1c1-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 123.6 GB at this context; duplicate traces of the same session: 14pxnmby, 63qsq1zr, 6630l2jh, nat0wb8x, q357fb1w; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1029.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 16k 1,029 58.7 123.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ulqm1c1",
   "source_url": "https://omlx.ai/benchmarks/performance/7ulqm1c1",
   "tg_tokens": null,
   "tps": 58.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-7ulqm1c1-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 128.3 GB at this context; duplicate traces of the same session: 14pxnmby, 63qsq1zr, 6630l2jh, nat0wb8x, q357fb1w; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1007.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 64k 1,007 69.0 128.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ulqm1c1",
   "source_url": "https://omlx.ai/benchmarks/performance/7ulqm1c1",
   "tg_tokens": null,
   "tps": 69.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-7ulqm1c1-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 134.1 GB at this context; duplicate traces of the same session: 14pxnmby, 63qsq1zr, 6630l2jh, nat0wb8x, q357fb1w; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 989.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 128k 989.2 56.6 134.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ulqm1c1",
   "source_url": "https://omlx.ai/benchmarks/performance/7ulqm1c1",
   "tg_tokens": null,
   "tps": 56.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-7ulqm1c1-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 136.9 GB at this context; duplicate traces of the same session: 14pxnmby, 63qsq1zr, 6630l2jh, nat0wb8x, q357fb1w; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 977.6,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 195k 977.6 47.6 136.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ulqm1c1",
   "source_url": "https://omlx.ai/benchmarks/performance/7ulqm1c1",
   "tg_tokens": null,
   "tps": 47.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-b0ba63t7-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.6 GB at this context; duplicate traces of the same session: 8180rxnz, d60m2azu, rhwolm84, rkkx4ioy, tr0rpnsi, w06box3p, wtc1th2g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2507.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 2,507 138.3 142.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b0ba63t7",
   "source_url": "https://omlx.ai/benchmarks/performance/b0ba63t7",
   "tg_tokens": null,
   "tps": 138.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-b0ba63t7-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: 8180rxnz, d60m2azu, rhwolm84, rkkx4ioy, tr0rpnsi, w06box3p, wtc1th2g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4424.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 4,424 134.0 143.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b0ba63t7",
   "source_url": "https://omlx.ai/benchmarks/performance/b0ba63t7",
   "tg_tokens": null,
   "tps": 134.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-b0ba63t7-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 144.4 GB at this context; duplicate traces of the same session: 8180rxnz, d60m2azu, rhwolm84, rkkx4ioy, tr0rpnsi, w06box3p, wtc1th2g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4836.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 8k 4,836 140.3 144.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b0ba63t7",
   "source_url": "https://omlx.ai/benchmarks/performance/b0ba63t7",
   "tg_tokens": null,
   "tps": 140.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-b0ba63t7-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.1 GB at this context; duplicate traces of the same session: 8180rxnz, d60m2azu, rhwolm84, rkkx4ioy, tr0rpnsi, w06box3p, wtc1th2g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4860.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 16k 4,860 132.4 145.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b0ba63t7",
   "source_url": "https://omlx.ai/benchmarks/performance/b0ba63t7",
   "tg_tokens": null,
   "tps": 132.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-b0ba63t7-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.7 GB at this context; duplicate traces of the same session: 8180rxnz, d60m2azu, rhwolm84, rkkx4ioy, tr0rpnsi, w06box3p, wtc1th2g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4846.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 32k 4,846 82.1 145.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b0ba63t7",
   "source_url": "https://omlx.ai/benchmarks/performance/b0ba63t7",
   "tg_tokens": null,
   "tps": 82.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-b0ba63t7-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 146.8 GB at this context; duplicate traces of the same session: 8180rxnz, d60m2azu, rhwolm84, rkkx4ioy, tr0rpnsi, w06box3p, wtc1th2g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4762.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 64k 4,762 126.1 146.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b0ba63t7",
   "source_url": "https://omlx.ai/benchmarks/performance/b0ba63t7",
   "tg_tokens": null,
   "tps": 126.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-b0ba63t7-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 150.6 GB at this context; duplicate traces of the same session: 8180rxnz, d60m2azu, rhwolm84, rkkx4ioy, tr0rpnsi, w06box3p, wtc1th2g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4622.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 128k 4,622 104.3 150.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b0ba63t7",
   "source_url": "https://omlx.ai/benchmarks/performance/b0ba63t7",
   "tg_tokens": null,
   "tps": 104.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-b0ba63t7-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 153.3 GB at this context; duplicate traces of the same session: 8180rxnz, d60m2azu, rhwolm84, rkkx4ioy, tr0rpnsi, w06box3p, wtc1th2g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4515.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 195k 4,515 89.6 153.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b0ba63t7",
   "source_url": "https://omlx.ai/benchmarks/performance/b0ba63t7",
   "tg_tokens": null,
   "tps": 89.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-lf5cf6wh-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 27.8 GB; TTFT 27503 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 148.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 5bit; 4k tokens 148.9 PP tok/s 13.4 TG tok/s 27503 TTFT (ms) 27.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lf5cf6wh",
   "source_url": "https://omlx.ai/benchmarks/performance/lf5cf6wh",
   "tg_tokens": null,
   "tps": 13.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-pbdjfkk2-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 27.8 GB; TTFT 12712 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 322.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 5bit; 4k tokens 322.2 PP tok/s 53.5 TG tok/s 12712 TTFT (ms) 27.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pbdjfkk2",
   "source_url": "https://omlx.ai/benchmarks/performance/pbdjfkk2",
   "tg_tokens": null,
   "tps": 53.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-sumqpfqy-ctx199680",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 195k context; peak memory 32.8 GB; TTFT 230421 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 868.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ6e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 195k tokens 868.0 PP tok/s 18.3 TG tok/s 230421 TTFT (ms) 32.8 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sumqpfqy",
   "source_url": "https://omlx.ai/benchmarks/performance/sumqpfqy",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-8obair8f-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1296.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 1,296 82.2 25.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8obair8f",
   "source_url": "https://omlx.ai/benchmarks/performance/8obair8f",
   "tg_tokens": null,
   "tps": 82.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-8obair8f-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1415.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 1,415 80.0 26.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8obair8f",
   "source_url": "https://omlx.ai/benchmarks/performance/8obair8f",
   "tg_tokens": null,
   "tps": 80.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-lemk703k-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.6 GB at this context; duplicate traces of the same session: sk0m5it9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1450.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 1,450 120.7 27.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lemk703k",
   "source_url": "https://omlx.ai/benchmarks/performance/lemk703k",
   "tg_tokens": null,
   "tps": 120.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-lemk703k-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: sk0m5it9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1427.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 1,427 66.9 28.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lemk703k",
   "source_url": "https://omlx.ai/benchmarks/performance/lemk703k",
   "tg_tokens": null,
   "tps": 66.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-p6j33s2u-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: 803ru0ay; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1148.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 1,148 91.5 24.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p6j33s2u",
   "source_url": "https://omlx.ai/benchmarks/performance/p6j33s2u",
   "tg_tokens": null,
   "tps": 91.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-p6j33s2u-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: 803ru0ay; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 946.7,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 16k 946.7 87.2 25.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p6j33s2u",
   "source_url": "https://omlx.ai/benchmarks/performance/p6j33s2u",
   "tg_tokens": null,
   "tps": 87.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-kraufon0-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 22.1 GB; TTFT 1417 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5780.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k tokens 5,780 PP tok/s 203.9 TG tok/s 1417 TTFT (ms) 22.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kraufon0",
   "source_url": "https://omlx.ai/benchmarks/performance/kraufon0",
   "tg_tokens": null,
   "tps": 203.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-frejlpls-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 30.7 GB; TTFT 1748 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4687.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k tokens 4,687 PP tok/s 131.9 TG tok/s 1748 TTFT (ms) 30.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/frejlpls",
   "source_url": "https://omlx.ai/benchmarks/performance/frejlpls",
   "tg_tokens": null,
   "tps": 131.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-tc8fj4xb-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 30.7 GB; TTFT 1745 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4694.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k tokens 4,694 PP tok/s 155.2 TG tok/s 1745 TTFT (ms) 30.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tc8fj4xb",
   "source_url": "https://omlx.ai/benchmarks/performance/tc8fj4xb",
   "tg_tokens": null,
   "tps": 155.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-56m7oxb3-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 30.7 GB; TTFT 1751 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4680.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k tokens 4,680 PP tok/s 143.7 TG tok/s 1751 TTFT (ms) 30.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/56m7oxb3",
   "source_url": "https://omlx.ai/benchmarks/performance/56m7oxb3",
   "tg_tokens": null,
   "tps": 143.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mha27pxe-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 30.7 GB; TTFT 1732 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4730.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k tokens 4,730 PP tok/s 167.0 TG tok/s 1732 TTFT (ms) 30.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mha27pxe",
   "source_url": "https://omlx.ai/benchmarks/performance/mha27pxe",
   "tg_tokens": null,
   "tps": 167.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-t1yonjn1-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 2c46s931; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 106.4 15.6 18.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t1yonjn1",
   "source_url": "https://omlx.ai/benchmarks/performance/t1yonjn1",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-t1yonjn1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 2c46s931; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 125.1 15.1 19.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t1yonjn1",
   "source_url": "https://omlx.ai/benchmarks/performance/t1yonjn1",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-8lant3rf-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: hijho9ox; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 88.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 88.4 10.4 16.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8lant3rf",
   "source_url": "https://omlx.ai/benchmarks/performance/8lant3rf",
   "tg_tokens": null,
   "tps": 10.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-8lant3rf-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: hijho9ox; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 93.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 93.0 4.0 20.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8lant3rf",
   "source_url": "https://omlx.ai/benchmarks/performance/8lant3rf",
   "tg_tokens": null,
   "tps": 4.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-mqz789ke-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: rgp71pev; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 115.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 115.2 13.9 16.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mqz789ke",
   "source_url": "https://omlx.ai/benchmarks/performance/mqz789ke",
   "tg_tokens": null,
   "tps": 13.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-mqz789ke-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: rgp71pev; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 117.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 117.2 13.1 19.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mqz789ke",
   "source_url": "https://omlx.ai/benchmarks/performance/mqz789ke",
   "tg_tokens": null,
   "tps": 13.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-2gkuy8i0-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 80t9n6yv; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 122.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 122.1 13.9 19.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2gkuy8i0",
   "source_url": "https://omlx.ai/benchmarks/performance/2gkuy8i0",
   "tg_tokens": null,
   "tps": 13.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-2gkuy8i0-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 80t9n6yv; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 102.2 11.3 21.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2gkuy8i0",
   "source_url": "https://omlx.ai/benchmarks/performance/2gkuy8i0",
   "tg_tokens": null,
   "tps": 11.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-labzz6jo-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.5 GB at this context; duplicate traces of the same session: ydckppjk; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 226.1,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 226.1 34.1 34.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/labzz6jo",
   "source_url": "https://omlx.ai/benchmarks/performance/labzz6jo",
   "tg_tokens": null,
   "tps": 34.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-labzz6jo-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.9 GB at this context; duplicate traces of the same session: ydckppjk; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 173.7,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 16k 173.7 32.7 35.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/labzz6jo",
   "source_url": "https://omlx.ai/benchmarks/performance/labzz6jo",
   "tg_tokens": null,
   "tps": 32.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2p0mzo53-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 190.3 GB at this context; duplicate traces of the same session: 89ykis1m, dl970o63, nom7dvud, tjujlf0y, tln4pvod, w5yznrql, y1skdw17; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 990.8,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ8-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 990.8 68.2 190.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2p0mzo53",
   "source_url": "https://omlx.ai/benchmarks/performance/2p0mzo53",
   "tg_tokens": null,
   "tps": 68.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2p0mzo53-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 190.9 GB at this context; duplicate traces of the same session: 89ykis1m, dl970o63, nom7dvud, tjujlf0y, tln4pvod, w5yznrql, y1skdw17; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1373.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ8-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 1,373 66.4 190.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2p0mzo53",
   "source_url": "https://omlx.ai/benchmarks/performance/2p0mzo53",
   "tg_tokens": null,
   "tps": 66.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2p0mzo53-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 191.1 GB at this context; duplicate traces of the same session: 89ykis1m, dl970o63, nom7dvud, tjujlf0y, tln4pvod, w5yznrql, y1skdw17; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1364.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ8-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 1,364 71.3 191.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2p0mzo53",
   "source_url": "https://omlx.ai/benchmarks/performance/2p0mzo53",
   "tg_tokens": null,
   "tps": 71.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2p0mzo53-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 191.4 GB at this context; duplicate traces of the same session: 89ykis1m, dl970o63, nom7dvud, tjujlf0y, tln4pvod, w5yznrql, y1skdw17; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1353.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ8-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k 1,353 66.6 191.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2p0mzo53",
   "source_url": "https://omlx.ai/benchmarks/performance/2p0mzo53",
   "tg_tokens": null,
   "tps": 66.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2p0mzo53-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 192.1 GB at this context; duplicate traces of the same session: 89ykis1m, dl970o63, nom7dvud, tjujlf0y, tln4pvod, w5yznrql, y1skdw17; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1342.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ8-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 32k 1,342 66.7 192.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2p0mzo53",
   "source_url": "https://omlx.ai/benchmarks/performance/2p0mzo53",
   "tg_tokens": null,
   "tps": 66.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2p0mzo53-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 194.6 GB at this context; duplicate traces of the same session: 89ykis1m, dl970o63, nom7dvud, tjujlf0y, tln4pvod, w5yznrql, y1skdw17; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1321.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ8-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 64k 1,321 61.7 194.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2p0mzo53",
   "source_url": "https://omlx.ai/benchmarks/performance/2p0mzo53",
   "tg_tokens": null,
   "tps": 61.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2p0mzo53-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 199.3 GB at this context; duplicate traces of the same session: 89ykis1m, dl970o63, nom7dvud, tjujlf0y, tln4pvod, w5yznrql, y1skdw17; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1300.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ8-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 128k 1,300 54.7 199.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2p0mzo53",
   "source_url": "https://omlx.ai/benchmarks/performance/2p0mzo53",
   "tg_tokens": null,
   "tps": 54.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2p0mzo53-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 199.9 GB at this context; duplicate traces of the same session: 89ykis1m, dl970o63, nom7dvud, tjujlf0y, tln4pvod, w5yznrql, y1skdw17; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1283.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ8-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 195k 1,283 50.8 199.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2p0mzo53",
   "source_url": "https://omlx.ai/benchmarks/performance/2p0mzo53",
   "tg_tokens": null,
   "tps": 50.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xeej4xom-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 100.5 GB; TTFT 1180 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 867.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k tokens 867.7 PP tok/s 94.2 TG tok/s 1180 TTFT (ms) 100.5 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xeej4xom",
   "source_url": "https://omlx.ai/benchmarks/performance/xeej4xom",
   "tg_tokens": null,
   "tps": 94.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-x8hbhngq-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 78.4 GB; TTFT 167952 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 780.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k tokens 780.4 PP tok/s 55.5 TG tok/s 167952 TTFT (ms) 78.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x8hbhngq",
   "source_url": "https://omlx.ai/benchmarks/performance/x8hbhngq",
   "tg_tokens": null,
   "tps": 55.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-5e2g3yuk-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.9 GB at this context; duplicate traces of the same session: 4nk7my5h; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5912.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 5,912 141.4 102.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5e2g3yuk",
   "source_url": "https://omlx.ai/benchmarks/performance/5e2g3yuk",
   "tg_tokens": null,
   "tps": 141.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-5e2g3yuk-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 105.1 GB at this context; duplicate traces of the same session: 4nk7my5h; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6056.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 6,056 144.2 105.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5e2g3yuk",
   "source_url": "https://omlx.ai/benchmarks/performance/5e2g3yuk",
   "tg_tokens": null,
   "tps": 144.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-5e2g3yuk-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.4 GB at this context; duplicate traces of the same session: 4nk7my5h; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5838.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,838 107.0 109.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5e2g3yuk",
   "source_url": "https://omlx.ai/benchmarks/performance/5e2g3yuk",
   "tg_tokens": null,
   "tps": 107.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-5e2g3yuk-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 111.1 GB at this context; duplicate traces of the same session: 4nk7my5h; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5657.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 5,657 94.4 111.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5e2g3yuk",
   "source_url": "https://omlx.ai/benchmarks/performance/5e2g3yuk",
   "tg_tokens": null,
   "tps": 94.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-4406j3hp-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.9 GB at this context; duplicate traces of the same session: 4qqjassw, 6plhzun1, s97or032; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6009.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 6,009 156.4 102.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4406j3hp",
   "source_url": "https://omlx.ai/benchmarks/performance/4406j3hp",
   "tg_tokens": null,
   "tps": 156.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-4406j3hp-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.7 GB at this context; duplicate traces of the same session: 4qqjassw, 6plhzun1, s97or032; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5843.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 5,843 142.7 104.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4406j3hp",
   "source_url": "https://omlx.ai/benchmarks/performance/4406j3hp",
   "tg_tokens": null,
   "tps": 142.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-4406j3hp-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.4 GB at this context; duplicate traces of the same session: 4qqjassw, 6plhzun1, s97or032; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5270.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,270 68.8 109.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4406j3hp",
   "source_url": "https://omlx.ai/benchmarks/performance/4406j3hp",
   "tg_tokens": null,
   "tps": 68.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-4406j3hp-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 111.1 GB at this context; duplicate traces of the same session: 4qqjassw, 6plhzun1, s97or032; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5228.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 5,228 90.9 111.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4406j3hp",
   "source_url": "https://omlx.ai/benchmarks/performance/4406j3hp",
   "tg_tokens": null,
   "tps": 90.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-gw0yzkpd-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.9 GB at this context; duplicate traces of the same session: 12hfadg2, 6wij44c7, ev2fjfm0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5941.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 5,941 123.9 102.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gw0yzkpd",
   "source_url": "https://omlx.ai/benchmarks/performance/gw0yzkpd",
   "tg_tokens": null,
   "tps": 123.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-gw0yzkpd-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 105.1 GB at this context; duplicate traces of the same session: 12hfadg2, 6wij44c7, ev2fjfm0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6015.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 6,015 133.1 105.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gw0yzkpd",
   "source_url": "https://omlx.ai/benchmarks/performance/gw0yzkpd",
   "tg_tokens": null,
   "tps": 133.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-gw0yzkpd-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.4 GB at this context; duplicate traces of the same session: 12hfadg2, 6wij44c7, ev2fjfm0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5827.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,827 110.1 109.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gw0yzkpd",
   "source_url": "https://omlx.ai/benchmarks/performance/gw0yzkpd",
   "tg_tokens": null,
   "tps": 110.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-gw0yzkpd-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 111.1 GB at this context; duplicate traces of the same session: 12hfadg2, 6wij44c7, ev2fjfm0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5651.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 5,651 90.3 111.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gw0yzkpd",
   "source_url": "https://omlx.ai/benchmarks/performance/gw0yzkpd",
   "tg_tokens": null,
   "tps": 90.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-oujwy1vl-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.9 GB at this context; duplicate traces of the same session: o95744nw, vrw3oksr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5983.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 5,983 149.6 102.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oujwy1vl",
   "source_url": "https://omlx.ai/benchmarks/performance/oujwy1vl",
   "tg_tokens": null,
   "tps": 149.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-oujwy1vl-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 105.4 GB at this context; duplicate traces of the same session: o95744nw, vrw3oksr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5998.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 5,998 131.7 105.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oujwy1vl",
   "source_url": "https://omlx.ai/benchmarks/performance/oujwy1vl",
   "tg_tokens": null,
   "tps": 131.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-oujwy1vl-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.5 GB at this context; duplicate traces of the same session: o95744nw, vrw3oksr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5767.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,767 104.5 109.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oujwy1vl",
   "source_url": "https://omlx.ai/benchmarks/performance/oujwy1vl",
   "tg_tokens": null,
   "tps": 104.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-oujwy1vl-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 126.7 GB at this context; duplicate traces of the same session: o95744nw, vrw3oksr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5628.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 5,628 23.8 126.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oujwy1vl",
   "source_url": "https://omlx.ai/benchmarks/performance/oujwy1vl",
   "tg_tokens": null,
   "tps": 23.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-4gmmdt4t-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 117.4 GB at this context; duplicate traces of the same session: hqu7kkmc, ks6ju0rn, vuhf0b0j; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 429.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 429.0 8.1 117.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4gmmdt4t",
   "source_url": "https://omlx.ai/benchmarks/performance/4gmmdt4t",
   "tg_tokens": null,
   "tps": 8.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-4gmmdt4t-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 117.8 GB at this context; duplicate traces of the same session: hqu7kkmc, ks6ju0rn, vuhf0b0j; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3226.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 3,226 15.2 117.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4gmmdt4t",
   "source_url": "https://omlx.ai/benchmarks/performance/4gmmdt4t",
   "tg_tokens": null,
   "tps": 15.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-4gmmdt4t-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 120.9 GB at this context; duplicate traces of the same session: hqu7kkmc, ks6ju0rn, vuhf0b0j; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3448.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 3,448 33.9 120.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4gmmdt4t",
   "source_url": "https://omlx.ai/benchmarks/performance/4gmmdt4t",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-4gmmdt4t-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 128.5 GB at this context; duplicate traces of the same session: hqu7kkmc, ks6ju0rn, vuhf0b0j; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3209.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 3,209 24.3 128.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4gmmdt4t",
   "source_url": "https://omlx.ai/benchmarks/performance/4gmmdt4t",
   "tg_tokens": null,
   "tps": 24.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-akmnxh60-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.9 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5937.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 5,937 155.5 102.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/akmnxh60",
   "source_url": "https://omlx.ai/benchmarks/performance/akmnxh60",
   "tg_tokens": null,
   "tps": 155.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-akmnxh60-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 105.0 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5998.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 5,998 134.3 105.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/akmnxh60",
   "source_url": "https://omlx.ai/benchmarks/performance/akmnxh60",
   "tg_tokens": null,
   "tps": 134.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-akmnxh60-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.5 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5780.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,780 107.6 109.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/akmnxh60",
   "source_url": "https://omlx.ai/benchmarks/performance/akmnxh60",
   "tg_tokens": null,
   "tps": 107.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-akmnxh60-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 111.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5457.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 5,457 88.1 111.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/akmnxh60",
   "source_url": "https://omlx.ai/benchmarks/performance/akmnxh60",
   "tg_tokens": null,
   "tps": 88.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-42jq8qhq-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 23.0 GB; TTFT 27554 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 148.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 5bit; 4k tokens 148.7 PP tok/s 10.7 TG tok/s 27554 TTFT (ms) 23.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/42jq8qhq",
   "source_url": "https://omlx.ai/benchmarks/performance/42jq8qhq",
   "tg_tokens": null,
   "tps": 10.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-719atimv-ctx1024",
   "model": "Qwen3.5-9B-DeepSeek-V4-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 6.0 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3405.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.5-9B-DeepSeek-V4-Flash-oQ4-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k 3,405 156.1 6.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/719atimv",
   "source_url": "https://omlx.ai/benchmarks/performance/719atimv",
   "tg_tokens": null,
   "tps": 156.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-719atimv-ctx4096",
   "model": "Qwen3.5-9B-DeepSeek-V4-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4895.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.5-9B-DeepSeek-V4-Flash-oQ4-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 4k 4,895 156.9 6.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/719atimv",
   "source_url": "https://omlx.ai/benchmarks/performance/719atimv",
   "tg_tokens": null,
   "tps": 156.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-r771esvw-ctx1024",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: affs0sta; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 768.6,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 768.6 72.8 19.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r771esvw",
   "source_url": "https://omlx.ai/benchmarks/performance/r771esvw",
   "tg_tokens": null,
   "tps": 72.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-r771esvw-ctx4096",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: affs0sta; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 787.7,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 787.7 66.7 20.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r771esvw",
   "source_url": "https://omlx.ai/benchmarks/performance/r771esvw",
   "tg_tokens": null,
   "tps": 66.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-8pldv148-ctx1024",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 669.8,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 669.8 69.8 19.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8pldv148",
   "source_url": "https://omlx.ai/benchmarks/performance/8pldv148",
   "tg_tokens": null,
   "tps": 69.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-8pldv148-ctx4096",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 765.9,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 765.9 67.0 20.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8pldv148",
   "source_url": "https://omlx.ai/benchmarks/performance/8pldv148",
   "tg_tokens": null,
   "tps": 67.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-nsf3kdmq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 7eitc6kg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 106.3 15.6 18.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nsf3kdmq",
   "source_url": "https://omlx.ai/benchmarks/performance/nsf3kdmq",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-nsf3kdmq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: 7eitc6kg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 123.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 123.4 15.2 19.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nsf3kdmq",
   "source_url": "https://omlx.ai/benchmarks/performance/nsf3kdmq",
   "tg_tokens": null,
   "tps": 15.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-lif246no-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: k0uwrhhc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 200.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 200.1 27.5 19.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lif246no",
   "source_url": "https://omlx.ai/benchmarks/performance/lif246no",
   "tg_tokens": null,
   "tps": 27.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-lif246no-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: k0uwrhhc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 207.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 207.9 42.3 23.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lif246no",
   "source_url": "https://omlx.ai/benchmarks/performance/lif246no",
   "tg_tokens": null,
   "tps": 42.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hpuaxklv-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.4 GB at this context; duplicate traces of the same session: 247o0cn5, 915j5maa, fjv2h8br, hl6kgtru, mie1tktq, qzl5lt9g, sjp1487v; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 410.1,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 410.1 28.7 30.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hpuaxklv",
   "source_url": "https://omlx.ai/benchmarks/performance/hpuaxklv",
   "tg_tokens": null,
   "tps": 28.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hpuaxklv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.5 GB at this context; duplicate traces of the same session: 247o0cn5, 915j5maa, fjv2h8br, hl6kgtru, mie1tktq, qzl5lt9g, sjp1487v; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 531.1,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 531.1 38.7 33.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hpuaxklv",
   "source_url": "https://omlx.ai/benchmarks/performance/hpuaxklv",
   "tg_tokens": null,
   "tps": 38.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hpuaxklv-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.0 GB at this context; duplicate traces of the same session: 247o0cn5, 915j5maa, fjv2h8br, hl6kgtru, mie1tktq, qzl5lt9g, sjp1487v; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 603.3,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 603.3 45.5 34.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hpuaxklv",
   "source_url": "https://omlx.ai/benchmarks/performance/hpuaxklv",
   "tg_tokens": null,
   "tps": 45.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hpuaxklv-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: 247o0cn5, 915j5maa, fjv2h8br, hl6kgtru, mie1tktq, qzl5lt9g, sjp1487v; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 568.8,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 568.8 50.7 34.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hpuaxklv",
   "source_url": "https://omlx.ai/benchmarks/performance/hpuaxklv",
   "tg_tokens": null,
   "tps": 50.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hpuaxklv-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.4 GB at this context; duplicate traces of the same session: 247o0cn5, 915j5maa, fjv2h8br, hl6kgtru, mie1tktq, qzl5lt9g, sjp1487v; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 510.3,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 510.3 45.2 36.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hpuaxklv",
   "source_url": "https://omlx.ai/benchmarks/performance/hpuaxklv",
   "tg_tokens": null,
   "tps": 45.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hpuaxklv-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.6 GB at this context; duplicate traces of the same session: 247o0cn5, 915j5maa, fjv2h8br, hl6kgtru, mie1tktq, qzl5lt9g, sjp1487v; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 422.8,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 422.8 32.9 39.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hpuaxklv",
   "source_url": "https://omlx.ai/benchmarks/performance/hpuaxklv",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-hpuaxklv-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 44.1 GB at this context; duplicate traces of the same session: 247o0cn5, 915j5maa, fjv2h8br, hl6kgtru, mie1tktq, qzl5lt9g, sjp1487v; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 318.3,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 318.3 35.3 44.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hpuaxklv",
   "source_url": "https://omlx.ai/benchmarks/performance/hpuaxklv",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-hpuaxklv-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 49.0 GB at this context; duplicate traces of the same session: 247o0cn5, 915j5maa, fjv2h8br, hl6kgtru, mie1tktq, qzl5lt9g, sjp1487v; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 253.2,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 195k 253.2 34.7 49.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hpuaxklv",
   "source_url": "https://omlx.ai/benchmarks/performance/hpuaxklv",
   "tg_tokens": null,
   "tps": 34.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-r409c20g-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: ozk6i79d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 251.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 251.4 27.0 16.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r409c20g",
   "source_url": "https://omlx.ai/benchmarks/performance/r409c20g",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-r409c20g-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: ozk6i79d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 88.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 88.3 4.9 18.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r409c20g",
   "source_url": "https://omlx.ai/benchmarks/performance/r409c20g",
   "tg_tokens": null,
   "tps": 4.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-kmji2ul8-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.8 GB at this context; duplicate traces of the same session: f5zwpaqg, od6nmxaa; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 846.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 846.3 91.4 100.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kmji2ul8",
   "source_url": "https://omlx.ai/benchmarks/performance/kmji2ul8",
   "tg_tokens": null,
   "tps": 91.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-kmji2ul8-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.0 GB at this context; duplicate traces of the same session: f5zwpaqg, od6nmxaa; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1056.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 1,056 95.6 102.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kmji2ul8",
   "source_url": "https://omlx.ai/benchmarks/performance/kmji2ul8",
   "tg_tokens": null,
   "tps": 95.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-kmji2ul8-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.5 GB at this context; duplicate traces of the same session: f5zwpaqg, od6nmxaa; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1046.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 1,046 79.5 103.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kmji2ul8",
   "source_url": "https://omlx.ai/benchmarks/performance/kmji2ul8",
   "tg_tokens": null,
   "tps": 79.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9kxk7ygu-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.8 GB at this context; duplicate traces of the same session: 48tb55y8, 73rts286, 7whotqmv, g97plyu5, htc6ydt5, jwn4zzkq, u4hzkk0d; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 905.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 905.5 95.0 100.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9kxk7ygu",
   "source_url": "https://omlx.ai/benchmarks/performance/9kxk7ygu",
   "tg_tokens": null,
   "tps": 95.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9kxk7ygu-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: 48tb55y8, 73rts286, 7whotqmv, g97plyu5, htc6ydt5, jwn4zzkq, u4hzkk0d; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1080.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 1,080 92.6 101.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9kxk7ygu",
   "source_url": "https://omlx.ai/benchmarks/performance/9kxk7ygu",
   "tg_tokens": null,
   "tps": 92.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9kxk7ygu-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: 48tb55y8, 73rts286, 7whotqmv, g97plyu5, htc6ydt5, jwn4zzkq, u4hzkk0d; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1064.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,064 85.6 101.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9kxk7ygu",
   "source_url": "https://omlx.ai/benchmarks/performance/9kxk7ygu",
   "tg_tokens": null,
   "tps": 85.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9kxk7ygu-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.9 GB at this context; duplicate traces of the same session: 48tb55y8, 73rts286, 7whotqmv, g97plyu5, htc6ydt5, jwn4zzkq, u4hzkk0d; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1055.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 1,055 82.8 101.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9kxk7ygu",
   "source_url": "https://omlx.ai/benchmarks/performance/9kxk7ygu",
   "tg_tokens": null,
   "tps": 82.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9kxk7ygu-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.3 GB at this context; duplicate traces of the same session: 48tb55y8, 73rts286, 7whotqmv, g97plyu5, htc6ydt5, jwn4zzkq, u4hzkk0d; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1046.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 1,046 84.6 103.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9kxk7ygu",
   "source_url": "https://omlx.ai/benchmarks/performance/9kxk7ygu",
   "tg_tokens": null,
   "tps": 84.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9kxk7ygu-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 106.5 GB at this context; duplicate traces of the same session: 48tb55y8, 73rts286, 7whotqmv, g97plyu5, htc6ydt5, jwn4zzkq, u4hzkk0d; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1032.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 1,032 76.8 106.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9kxk7ygu",
   "source_url": "https://omlx.ai/benchmarks/performance/9kxk7ygu",
   "tg_tokens": null,
   "tps": 76.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9kxk7ygu-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 113.1 GB at this context; duplicate traces of the same session: 48tb55y8, 73rts286, 7whotqmv, g97plyu5, htc6ydt5, jwn4zzkq, u4hzkk0d; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1014.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 128k 1,014 65.7 113.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9kxk7ygu",
   "source_url": "https://omlx.ai/benchmarks/performance/9kxk7ygu",
   "tg_tokens": null,
   "tps": 65.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9kxk7ygu-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 115.5 GB at this context; duplicate traces of the same session: 48tb55y8, 73rts286, 7whotqmv, g97plyu5, htc6ydt5, jwn4zzkq, u4hzkk0d; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 999.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 195k 999.5 52.6 115.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9kxk7ygu",
   "source_url": "https://omlx.ai/benchmarks/performance/9kxk7ygu",
   "tg_tokens": null,
   "tps": 52.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3enx8gfw-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: 4d746slg, eno6wvzx, yazt1f5a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1195.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,195 70.3 70.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3enx8gfw",
   "source_url": "https://omlx.ai/benchmarks/performance/3enx8gfw",
   "tg_tokens": null,
   "tps": 70.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3enx8gfw-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: 4d746slg, eno6wvzx, yazt1f5a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2301.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,301 85.2 71.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3enx8gfw",
   "source_url": "https://omlx.ai/benchmarks/performance/3enx8gfw",
   "tg_tokens": null,
   "tps": 85.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3enx8gfw-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 4d746slg, eno6wvzx, yazt1f5a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2870.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,870 86.6 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3enx8gfw",
   "source_url": "https://omlx.ai/benchmarks/performance/3enx8gfw",
   "tg_tokens": null,
   "tps": 86.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3enx8gfw-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 76.5 GB at this context; duplicate traces of the same session: 4d746slg, eno6wvzx, yazt1f5a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2616.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,616 73.7 76.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3enx8gfw",
   "source_url": "https://omlx.ai/benchmarks/performance/3enx8gfw",
   "tg_tokens": null,
   "tps": 73.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-qcty8xwf-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.9 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5983.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 5,983 149.6 102.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qcty8xwf",
   "source_url": "https://omlx.ai/benchmarks/performance/qcty8xwf",
   "tg_tokens": null,
   "tps": 149.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-qcty8xwf-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 105.4 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5998.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 5,998 131.7 105.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qcty8xwf",
   "source_url": "https://omlx.ai/benchmarks/performance/qcty8xwf",
   "tg_tokens": null,
   "tps": 131.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-qcty8xwf-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.5 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5767.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,767 104.5 109.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qcty8xwf",
   "source_url": "https://omlx.ai/benchmarks/performance/qcty8xwf",
   "tg_tokens": null,
   "tps": 104.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-qcty8xwf-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 126.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5628.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 5,628 23.8 126.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qcty8xwf",
   "source_url": "https://omlx.ai/benchmarks/performance/qcty8xwf",
   "tg_tokens": null,
   "tps": 23.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-v6mi7crl-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.9 GB at this context; duplicate traces of the same session: ebczl20j, ikdfkx0b; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5937.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 5,937 155.5 102.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v6mi7crl",
   "source_url": "https://omlx.ai/benchmarks/performance/v6mi7crl",
   "tg_tokens": null,
   "tps": 155.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-v6mi7crl-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 105.0 GB at this context; duplicate traces of the same session: ebczl20j, ikdfkx0b; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5998.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 5,998 134.3 105.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v6mi7crl",
   "source_url": "https://omlx.ai/benchmarks/performance/v6mi7crl",
   "tg_tokens": null,
   "tps": 134.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-v6mi7crl-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.5 GB at this context; duplicate traces of the same session: ebczl20j, ikdfkx0b; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5780.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,780 107.6 109.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v6mi7crl",
   "source_url": "https://omlx.ai/benchmarks/performance/v6mi7crl",
   "tg_tokens": null,
   "tps": 107.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-v6mi7crl-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 111.1 GB at this context; duplicate traces of the same session: ebczl20j, ikdfkx0b; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5457.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 5,457 88.1 111.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v6mi7crl",
   "source_url": "https://omlx.ai/benchmarks/performance/v6mi7crl",
   "tg_tokens": null,
   "tps": 88.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-zq5pigz3-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.0 GB at this context; duplicate traces of the same session: 8fzkifgs, ce63avj2, zpdsy3tz; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1128.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 1,128 106.1 102.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zq5pigz3",
   "source_url": "https://omlx.ai/benchmarks/performance/zq5pigz3",
   "tg_tokens": null,
   "tps": 106.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-zq5pigz3-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.6 GB at this context; duplicate traces of the same session: 8fzkifgs, ce63avj2, zpdsy3tz; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5690.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 5,690 73.6 104.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zq5pigz3",
   "source_url": "https://omlx.ai/benchmarks/performance/zq5pigz3",
   "tg_tokens": null,
   "tps": 73.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-zq5pigz3-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.3 GB at this context; duplicate traces of the same session: 8fzkifgs, ce63avj2, zpdsy3tz; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5706.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,706 29.9 110.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zq5pigz3",
   "source_url": "https://omlx.ai/benchmarks/performance/zq5pigz3",
   "tg_tokens": null,
   "tps": 29.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-zq5pigz3-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 129.8 GB at this context; duplicate traces of the same session: 8fzkifgs, ce63avj2, zpdsy3tz; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2782.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 2,782 5.8 129.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zq5pigz3",
   "source_url": "https://omlx.ai/benchmarks/performance/zq5pigz3",
   "tg_tokens": null,
   "tps": 5.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-69gw8nhk-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.9 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1030.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 1,030 64.9 183.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/69gw8nhk",
   "source_url": "https://omlx.ai/benchmarks/performance/69gw8nhk",
   "tg_tokens": null,
   "tps": 64.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-69gw8nhk-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 184.5 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1020.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 16k 1,020 75.5 184.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/69gw8nhk",
   "source_url": "https://omlx.ai/benchmarks/performance/69gw8nhk",
   "tg_tokens": null,
   "tps": 75.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-69gw8nhk-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 185.8 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1011.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 32k 1,011 77.2 185.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/69gw8nhk",
   "source_url": "https://omlx.ai/benchmarks/performance/69gw8nhk",
   "tg_tokens": null,
   "tps": 77.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-69gw8nhk-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 189.3 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 998.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 64k 998.9 62.3 189.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/69gw8nhk",
   "source_url": "https://omlx.ai/benchmarks/performance/69gw8nhk",
   "tg_tokens": null,
   "tps": 62.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ry3jaaxj-ctx8192",
   "model": "gemma-4-31B-it-scotoma-2",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 22.0 GB; TTFT 22025 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 371.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-scotoma-2-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k tokens 371.9 PP tok/s 34.8 TG tok/s 22025 TTFT (ms) 22.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ry3jaaxj",
   "source_url": "https://omlx.ai/benchmarks/performance/ry3jaaxj",
   "tg_tokens": null,
   "tps": 34.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-6yyfibzz-ctx8192",
   "model": "gemma-4-31B-it-scotoma-2",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 35.5 GB; TTFT 23594 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 347.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-31B-it-scotoma-2-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k tokens 347.2 PP tok/s 19.3 TG tok/s 23594 TTFT (ms) 35.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6yyfibzz",
   "source_url": "https://omlx.ai/benchmarks/performance/6yyfibzz",
   "tg_tokens": null,
   "tps": 19.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-ye105rxl-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "mlx-community; oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: sj2kxskz; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--gemma-4-26b-a4b-it-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 323.2 20.5 14.3 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ye105rxl",
   "source_url": "https://omlx.ai/benchmarks/performance/ye105rxl",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-ye105rxl-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "mlx-community; oMLX community benchmark trace; peak memory 14.9 GB at this context; duplicate traces of the same session: sj2kxskz; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 288.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--gemma-4-26b-a4b-it-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 16k 288.3 16.0 14.9 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ye105rxl",
   "source_url": "https://omlx.ai/benchmarks/performance/ye105rxl",
   "tg_tokens": null,
   "tps": 16.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-yiqrplle-ctx16384",
   "model": "Gemma4-26B-MoE",
   "notes": "oMLX community benchmark trace; peak memory 15.1 GB at this context; duplicate traces of the same session: 6l2ucmrq; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 343.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Gemma4-26B-MoE-Q4 on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 16k 343.4 29.2 15.1 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yiqrplle",
   "source_url": "https://omlx.ai/benchmarks/performance/yiqrplle",
   "tg_tokens": null,
   "tps": 29.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-yiqrplle-ctx32768",
   "model": "Gemma4-26B-MoE",
   "notes": "oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: 6l2ucmrq; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 302.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Gemma4-26B-MoE-Q4 on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 32k 302.8 24.6 15.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yiqrplle",
   "source_url": "https://omlx.ai/benchmarks/performance/yiqrplle",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-fqxyxfc9-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 9erx3l01; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 896.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 896.9 92.7 19.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fqxyxfc9",
   "source_url": "https://omlx.ai/benchmarks/performance/fqxyxfc9",
   "tg_tokens": null,
   "tps": 92.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-fqxyxfc9-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 9erx3l01; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1027.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 1,027 88.8 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fqxyxfc9",
   "source_url": "https://omlx.ai/benchmarks/performance/fqxyxfc9",
   "tg_tokens": null,
   "tps": 88.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-bwnupjk0-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: g522ovj9; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 898.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 898.8 88.9 19.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bwnupjk0",
   "source_url": "https://omlx.ai/benchmarks/performance/bwnupjk0",
   "tg_tokens": null,
   "tps": 88.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-bwnupjk0-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: g522ovj9; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 725.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 725.8 66.7 20.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bwnupjk0",
   "source_url": "https://omlx.ai/benchmarks/performance/bwnupjk0",
   "tg_tokens": null,
   "tps": 66.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-bwnupjk0-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: g522ovj9; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 530.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 530.0 49.9 21.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bwnupjk0",
   "source_url": "https://omlx.ai/benchmarks/performance/bwnupjk0",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-hz6rqsnl-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: witfog0k; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 896.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 896.1 95.2 19.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hz6rqsnl",
   "source_url": "https://omlx.ai/benchmarks/performance/hz6rqsnl",
   "tg_tokens": null,
   "tps": 95.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-hz6rqsnl-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: witfog0k; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1029.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 1,029 91.7 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hz6rqsnl",
   "source_url": "https://omlx.ai/benchmarks/performance/hz6rqsnl",
   "tg_tokens": null,
   "tps": 91.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-b388iwt7-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.2 GB at this context; duplicate traces of the same session: hwkimzzz, vjwna3no; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 828.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 1k 828.4 40.2 28.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b388iwt7",
   "source_url": "https://omlx.ai/benchmarks/performance/b388iwt7",
   "tg_tokens": null,
   "tps": 40.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-b388iwt7-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.1 GB at this context; duplicate traces of the same session: hwkimzzz, vjwna3no; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 882.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 4k 882.5 61.4 29.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b388iwt7",
   "source_url": "https://omlx.ai/benchmarks/performance/b388iwt7",
   "tg_tokens": null,
   "tps": 61.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-b388iwt7-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: hwkimzzz, vjwna3no; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 779.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 16k 779.2 15.6 29.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b388iwt7",
   "source_url": "https://omlx.ai/benchmarks/performance/b388iwt7",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-u58a1xul-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: 6cctjjh5, nlz4lyb0; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 815.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 1k 815.6 44.9 28.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u58a1xul",
   "source_url": "https://omlx.ai/benchmarks/performance/u58a1xul",
   "tg_tokens": null,
   "tps": 44.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-u58a1xul-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.0 GB at this context; duplicate traces of the same session: 6cctjjh5, nlz4lyb0; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 890.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 4k 890.8 32.0 29.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u58a1xul",
   "source_url": "https://omlx.ai/benchmarks/performance/u58a1xul",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-u58a1xul-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: 6cctjjh5, nlz4lyb0; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 767.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 16k 767.0 22.7 29.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u58a1xul",
   "source_url": "https://omlx.ai/benchmarks/performance/u58a1xul",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-9sa0mzi1-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: t6sikpme, urci3gv3; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 785.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 1k 785.6 43.7 28.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9sa0mzi1",
   "source_url": "https://omlx.ai/benchmarks/performance/9sa0mzi1",
   "tg_tokens": null,
   "tps": 43.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-9sa0mzi1-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.0 GB at this context; duplicate traces of the same session: t6sikpme, urci3gv3; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 868.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 4k 868.6 32.2 29.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9sa0mzi1",
   "source_url": "https://omlx.ai/benchmarks/performance/9sa0mzi1",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-9sa0mzi1-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: t6sikpme, urci3gv3; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 773.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 16k 773.2 19.9 29.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9sa0mzi1",
   "source_url": "https://omlx.ai/benchmarks/performance/9sa0mzi1",
   "tg_tokens": null,
   "tps": 19.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-x2mf2gbl-ctx16384",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "MLX; oMLX community benchmark trace, single run at 16k context; peak memory 18.1 GB; TTFT 62364 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 262.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 16k tokens 262.7 PP tok/s 27.1 TG tok/s 62364 TTFT (ms) 18.1 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x2mf2gbl",
   "source_url": "https://omlx.ai/benchmarks/performance/x2mf2gbl",
   "tg_tokens": null,
   "tps": 27.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-h418illo-ctx1024",
   "model": "Qwen3-VL-2B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 1.6 GB at this context; duplicate traces of the same session: 97dtt7eg, od9yzjl5, qdyy2y2l; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1120.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-VL-2B-Instruct-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 1,120 88.3 1.6 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h418illo",
   "source_url": "https://omlx.ai/benchmarks/performance/h418illo",
   "tg_tokens": null,
   "tps": 88.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-h418illo-ctx4096",
   "model": "Qwen3-VL-2B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 2.0 GB at this context; duplicate traces of the same session: 97dtt7eg, od9yzjl5, qdyy2y2l; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1016.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-VL-2B-Instruct-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 1,016 66.3 2.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h418illo",
   "source_url": "https://omlx.ai/benchmarks/performance/h418illo",
   "tg_tokens": null,
   "tps": 66.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-h418illo-ctx8192",
   "model": "Qwen3-VL-2B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 2.3 GB at this context; duplicate traces of the same session: 97dtt7eg, od9yzjl5, qdyy2y2l; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 895.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-VL-2B-Instruct-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 8k 895.3 50.5 2.3 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h418illo",
   "source_url": "https://omlx.ai/benchmarks/performance/h418illo",
   "tg_tokens": null,
   "tps": 50.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-h418illo-ctx16384",
   "model": "Qwen3-VL-2B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 3.0 GB at this context; duplicate traces of the same session: 97dtt7eg, od9yzjl5, qdyy2y2l; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 720.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-VL-2B-Instruct-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 16k 720.3 34.4 3.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h418illo",
   "source_url": "https://omlx.ai/benchmarks/performance/h418illo",
   "tg_tokens": null,
   "tps": 34.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-g5dni34x-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: rppas7ie; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 835.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 835.3 83.3 3.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g5dni34x",
   "source_url": "https://omlx.ai/benchmarks/performance/g5dni34x",
   "tg_tokens": null,
   "tps": 83.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-g5dni34x-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.1 GB at this context; duplicate traces of the same session: rppas7ie; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 843.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 843.0 80.1 4.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g5dni34x",
   "source_url": "https://omlx.ai/benchmarks/performance/g5dni34x",
   "tg_tokens": null,
   "tps": 80.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-jcszmwm6-ctx1024",
   "model": "Qwen3.5-9B-DeepSeek-V4-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 6.0 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3405.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.5-9B-DeepSeek-V4-Flash-oQ4-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k 3,405 156.1 6.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jcszmwm6",
   "source_url": "https://omlx.ai/benchmarks/performance/jcszmwm6",
   "tg_tokens": null,
   "tps": 156.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-jcszmwm6-ctx4096",
   "model": "Qwen3.5-9B-DeepSeek-V4-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4895.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.5-9B-DeepSeek-V4-Flash-oQ4-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 4k 4,895 156.9 6.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jcszmwm6",
   "source_url": "https://omlx.ai/benchmarks/performance/jcszmwm6",
   "tg_tokens": null,
   "tps": 156.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-j62s9xwc-ctx1024",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 669.8,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 669.8 69.8 19.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j62s9xwc",
   "source_url": "https://omlx.ai/benchmarks/performance/j62s9xwc",
   "tg_tokens": null,
   "tps": 69.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-j62s9xwc-ctx4096",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 765.9,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 765.9 67.0 20.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j62s9xwc",
   "source_url": "https://omlx.ai/benchmarks/performance/j62s9xwc",
   "tg_tokens": null,
   "tps": 67.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-odd5q4rd-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 18.3 GB; TTFT 2198 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 465.9,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.6-35B-A3B-mxfp4 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 465.9 PP tok/s 50.6 TG tok/s 2198 TTFT (ms) 18.3 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/odd5q4rd",
   "source_url": "https://omlx.ai/benchmarks/performance/odd5q4rd",
   "tg_tokens": null,
   "tps": 50.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-jl9977j8-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 4lfiar8j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 253.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 253.7 16.5 28.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jl9977j8",
   "source_url": "https://omlx.ai/benchmarks/performance/jl9977j8",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-jl9977j8-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.2 GB at this context; duplicate traces of the same session: 4lfiar8j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 258.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 258.9 16.3 32.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jl9977j8",
   "source_url": "https://omlx.ai/benchmarks/performance/jl9977j8",
   "tg_tokens": null,
   "tps": 16.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-55kivonw-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 15.0 GB; TTFT 15535 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 65.9,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 65.9 PP tok/s 7.2 TG tok/s 15535 TTFT (ms) 15.0 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/55kivonw",
   "source_url": "https://omlx.ai/benchmarks/performance/55kivonw",
   "tg_tokens": null,
   "tps": 7.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0xeanh6y-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1050.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 4k 1,050 78.1 143.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0xeanh6y",
   "source_url": "https://omlx.ai/benchmarks/performance/0xeanh6y",
   "tg_tokens": null,
   "tps": 78.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0xeanh6y-ctx16384",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 143.6 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1035.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 16k 1,035 70.4 143.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0xeanh6y",
   "source_url": "https://omlx.ai/benchmarks/performance/0xeanh6y",
   "tg_tokens": null,
   "tps": 70.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-li7fjtug-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: 0q6d64n8, 3coj2zv3, 4zde6hmo, hc2jemrc, pckhbo41, zrmlwq4e; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 905.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 905.3 86.9 100.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/li7fjtug",
   "source_url": "https://omlx.ai/benchmarks/performance/li7fjtug",
   "tg_tokens": null,
   "tps": 86.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-li7fjtug-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: 0q6d64n8, 3coj2zv3, 4zde6hmo, hc2jemrc, pckhbo41, zrmlwq4e; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1079.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 1,079 82.8 101.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/li7fjtug",
   "source_url": "https://omlx.ai/benchmarks/performance/li7fjtug",
   "tg_tokens": null,
   "tps": 82.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-li7fjtug-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: 0q6d64n8, 3coj2zv3, 4zde6hmo, hc2jemrc, pckhbo41, zrmlwq4e; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1065.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,065 82.1 101.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/li7fjtug",
   "source_url": "https://omlx.ai/benchmarks/performance/li7fjtug",
   "tg_tokens": null,
   "tps": 82.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-li7fjtug-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.9 GB at this context; duplicate traces of the same session: 0q6d64n8, 3coj2zv3, 4zde6hmo, hc2jemrc, pckhbo41, zrmlwq4e; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1055.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 1,055 79.9 101.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/li7fjtug",
   "source_url": "https://omlx.ai/benchmarks/performance/li7fjtug",
   "tg_tokens": null,
   "tps": 79.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-li7fjtug-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.3 GB at this context; duplicate traces of the same session: 0q6d64n8, 3coj2zv3, 4zde6hmo, hc2jemrc, pckhbo41, zrmlwq4e; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1046.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 1,046 80.8 103.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/li7fjtug",
   "source_url": "https://omlx.ai/benchmarks/performance/li7fjtug",
   "tg_tokens": null,
   "tps": 80.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-li7fjtug-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 106.7 GB at this context; duplicate traces of the same session: 0q6d64n8, 3coj2zv3, 4zde6hmo, hc2jemrc, pckhbo41, zrmlwq4e; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1033.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 1,033 73.0 106.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/li7fjtug",
   "source_url": "https://omlx.ai/benchmarks/performance/li7fjtug",
   "tg_tokens": null,
   "tps": 73.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-li7fjtug-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 113.1 GB at this context; duplicate traces of the same session: 0q6d64n8, 3coj2zv3, 4zde6hmo, hc2jemrc, pckhbo41, zrmlwq4e; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1014.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 128k 1,014 58.8 113.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/li7fjtug",
   "source_url": "https://omlx.ai/benchmarks/performance/li7fjtug",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9rpgsi8b-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.3 GB at this context; duplicate traces of the same session: 47vazxcl; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 519.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 519.4 13.0 72.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9rpgsi8b",
   "source_url": "https://omlx.ai/benchmarks/performance/9rpgsi8b",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9rpgsi8b-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.4 GB at this context; duplicate traces of the same session: 47vazxcl; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 926.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k 926.0 10.7 74.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9rpgsi8b",
   "source_url": "https://omlx.ai/benchmarks/performance/9rpgsi8b",
   "tg_tokens": null,
   "tps": 10.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-73hzj14r-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 124.5 GB; TTFT 6510 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2517.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 5bit; 16k tokens 2,517 PP tok/s 89.3 TG tok/s 6510 TTFT (ms) 124.5 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/73hzj14r",
   "source_url": "https://omlx.ai/benchmarks/performance/73hzj14r",
   "tg_tokens": null,
   "tps": 89.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-zl5r5bc1-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.9 GB at this context; duplicate traces of the same session: 1iy1gt44, zesgudu4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1030.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 1,030 64.9 183.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zl5r5bc1",
   "source_url": "https://omlx.ai/benchmarks/performance/zl5r5bc1",
   "tg_tokens": null,
   "tps": 64.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-zl5r5bc1-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 184.5 GB at this context; duplicate traces of the same session: 1iy1gt44, zesgudu4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1020.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 16k 1,020 75.5 184.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zl5r5bc1",
   "source_url": "https://omlx.ai/benchmarks/performance/zl5r5bc1",
   "tg_tokens": null,
   "tps": 75.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-zl5r5bc1-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 185.8 GB at this context; duplicate traces of the same session: 1iy1gt44, zesgudu4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1011.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 32k 1,011 77.2 185.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zl5r5bc1",
   "source_url": "https://omlx.ai/benchmarks/performance/zl5r5bc1",
   "tg_tokens": null,
   "tps": 77.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-zl5r5bc1-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 189.3 GB at this context; duplicate traces of the same session: 1iy1gt44, zesgudu4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 998.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 64k 998.9 62.3 189.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zl5r5bc1",
   "source_url": "https://omlx.ai/benchmarks/performance/zl5r5bc1",
   "tg_tokens": null,
   "tps": 62.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-5qcoiq1p-ctx199680",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 195k context; peak memory 27.0 GB; TTFT 234964 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 851.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 195k tokens 851.2 PP tok/s 18.0 TG tok/s 234964 TTFT (ms) 27.0 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5qcoiq1p",
   "source_url": "https://omlx.ai/benchmarks/performance/5qcoiq1p",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-absndbaf-ctx8192",
   "model": "gemma-4-31B-it-scotoma-2",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 22.0 GB; TTFT 21666 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 378.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-scotoma-2-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k tokens 378.1 PP tok/s 34.1 TG tok/s 21666 TTFT (ms) 22.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/absndbaf",
   "source_url": "https://omlx.ai/benchmarks/performance/absndbaf",
   "tg_tokens": null,
   "tps": 34.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lo84kay0-ctx4096",
   "model": "Ornith-1.5-35B-A3B-BigBang",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: ermmt4vz, f5kdoqv4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5700.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-BigBang-MTP-mlx-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 5,700 212.9 21.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lo84kay0",
   "source_url": "https://omlx.ai/benchmarks/performance/lo84kay0",
   "tg_tokens": null,
   "tps": 212.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lo84kay0-ctx16384",
   "model": "Ornith-1.5-35B-A3B-BigBang",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: ermmt4vz, f5kdoqv4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5487.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-BigBang-MTP-mlx-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 5,487 133.4 22.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lo84kay0",
   "source_url": "https://omlx.ai/benchmarks/performance/lo84kay0",
   "tg_tokens": null,
   "tps": 133.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lo84kay0-ctx65536",
   "model": "Ornith-1.5-35B-A3B-BigBang",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: ermmt4vz, f5kdoqv4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3524.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-BigBang-MTP-mlx-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 3,524 106.4 23.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lo84kay0",
   "source_url": "https://omlx.ai/benchmarks/performance/lo84kay0",
   "tg_tokens": null,
   "tps": 106.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-a5a0yw98-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 898.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 898.8 88.9 19.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a5a0yw98",
   "source_url": "https://omlx.ai/benchmarks/performance/a5a0yw98",
   "tg_tokens": null,
   "tps": 88.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-a5a0yw98-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 725.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 725.8 66.7 20.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a5a0yw98",
   "source_url": "https://omlx.ai/benchmarks/performance/a5a0yw98",
   "tg_tokens": null,
   "tps": 66.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-a5a0yw98-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 530.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 530.0 49.9 21.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a5a0yw98",
   "source_url": "https://omlx.ai/benchmarks/performance/a5a0yw98",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-yjuth4xa-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 32k141jv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 898.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 898.5 93.0 19.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yjuth4xa",
   "source_url": "https://omlx.ai/benchmarks/performance/yjuth4xa",
   "tg_tokens": null,
   "tps": 93.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-yjuth4xa-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 32k141jv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1029.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 1,029 87.6 20.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yjuth4xa",
   "source_url": "https://omlx.ai/benchmarks/performance/yjuth4xa",
   "tg_tokens": null,
   "tps": 87.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-lgn8f5o7-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: z008007l; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 898.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 898.6 92.0 19.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lgn8f5o7",
   "source_url": "https://omlx.ai/benchmarks/performance/lgn8f5o7",
   "tg_tokens": null,
   "tps": 92.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-lgn8f5o7-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: z008007l; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1021.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 1,021 88.6 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lgn8f5o7",
   "source_url": "https://omlx.ai/benchmarks/performance/lgn8f5o7",
   "tg_tokens": null,
   "tps": 88.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-q3xbg523-ctx1024",
   "model": "Ornith-1.5-35B-A3B-REAP192",
   "notes": "oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: 5pkybrvy, ldu2tmz1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 923.3,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Ornith-1.5-35B-A3B-REAP192-mxfp4-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 923.3 97.5 14.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q3xbg523",
   "source_url": "https://omlx.ai/benchmarks/performance/q3xbg523",
   "tg_tokens": null,
   "tps": 97.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-q3xbg523-ctx32768",
   "model": "Ornith-1.5-35B-A3B-REAP192",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: 5pkybrvy, ldu2tmz1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 737.9,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Ornith-1.5-35B-A3B-REAP192-mxfp4-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 737.9 69.4 15.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q3xbg523",
   "source_url": "https://omlx.ai/benchmarks/performance/q3xbg523",
   "tg_tokens": null,
   "tps": 69.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-q3xbg523-ctx65536",
   "model": "Ornith-1.5-35B-A3B-REAP192",
   "notes": "oMLX community benchmark trace; peak memory 16.8 GB at this context; duplicate traces of the same session: 5pkybrvy, ldu2tmz1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 545.3,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Ornith-1.5-35B-A3B-REAP192-mxfp4-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 545.3 54.1 16.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q3xbg523",
   "source_url": "https://omlx.ai/benchmarks/performance/q3xbg523",
   "tg_tokens": null,
   "tps": 54.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-ngr1zcj1-ctx4096",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "MLX; oMLX community benchmark trace, single run at 4k context; peak memory 17.0 GB; TTFT 9742 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 420.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k tokens 420.4 PP tok/s 38.4 TG tok/s 9742 TTFT (ms) 17.0 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ngr1zcj1",
   "source_url": "https://omlx.ai/benchmarks/performance/ngr1zcj1",
   "tg_tokens": null,
   "tps": 38.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ns0vl9ng-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 65.8 GB at this context; duplicate traces of the same session: z0w1qb6a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 560.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 560.2 59.1 65.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ns0vl9ng",
   "source_url": "https://omlx.ai/benchmarks/performance/ns0vl9ng",
   "tg_tokens": null,
   "tps": 59.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ns0vl9ng-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 69.5 GB at this context; duplicate traces of the same session: z0w1qb6a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 634.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 634.1 57.4 69.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ns0vl9ng",
   "source_url": "https://omlx.ai/benchmarks/performance/ns0vl9ng",
   "tg_tokens": null,
   "tps": 57.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1idszs20-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MLX Serve build; oMLX community benchmark trace, single run at 4k context; peak memory 20.1 GB; TTFT 2317 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1768.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-Serve-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 1,768 PP tok/s 136.5 TG tok/s 2317 TTFT (ms) 20.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1idszs20",
   "source_url": "https://omlx.ai/benchmarks/performance/1idszs20",
   "tg_tokens": null,
   "tps": 136.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-g3xqz6up-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: vwyp4nl1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 416.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 416.1 10.1 28.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g3xqz6up",
   "source_url": "https://omlx.ai/benchmarks/performance/g3xqz6up",
   "tg_tokens": null,
   "tps": 10.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-g3xqz6up-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: vwyp4nl1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 474.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 474.7 10.0 29.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g3xqz6up",
   "source_url": "https://omlx.ai/benchmarks/performance/g3xqz6up",
   "tg_tokens": null,
   "tps": 10.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-0cqgn1uy-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 13.5 GB at this context; duplicate traces of the same session: hfwcdo3i; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 65.7,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-27B-oQ3e-mtp on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 3bit; 1k 65.7 8.0 13.5 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0cqgn1uy",
   "source_url": "https://omlx.ai/benchmarks/performance/0cqgn1uy",
   "tg_tokens": null,
   "tps": 8.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-0cqgn1uy-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 14.7 GB at this context; duplicate traces of the same session: hfwcdo3i; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 65.6,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-27B-oQ3e-mtp on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 3bit; 4k 65.6 2.2 14.7 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0cqgn1uy",
   "source_url": "https://omlx.ai/benchmarks/performance/0cqgn1uy",
   "tg_tokens": null,
   "tps": 2.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-wiy1s9hy-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: w0gimcej; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 300.8,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 300.8 29.9 29.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wiy1s9hy",
   "source_url": "https://omlx.ai/benchmarks/performance/wiy1s9hy",
   "tg_tokens": null,
   "tps": 29.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-wiy1s9hy-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.4 GB at this context; duplicate traces of the same session: w0gimcej; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 263.2,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k 263.2 29.2 32.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wiy1s9hy",
   "source_url": "https://omlx.ai/benchmarks/performance/wiy1s9hy",
   "tg_tokens": null,
   "tps": 29.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-93p079ii-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 28.6 GB; TTFT 84142 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 389.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k tokens 389.4 PP tok/s 59.1 TG tok/s 84142 TTFT (ms) 28.6 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/93p079ii",
   "source_url": "https://omlx.ai/benchmarks/performance/93p079ii",
   "tg_tokens": null,
   "tps": 59.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rhgm17jk-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 28.6 GB; TTFT 83689 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 391.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k tokens 391.5 PP tok/s 63.2 TG tok/s 83689 TTFT (ms) 28.6 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rhgm17jk",
   "source_url": "https://omlx.ai/benchmarks/performance/rhgm17jk",
   "tg_tokens": null,
   "tps": 63.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-nlqw0c0w-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 28.6 GB; TTFT 83875 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 390.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k tokens 390.7 PP tok/s 134.3 TG tok/s 83875 TTFT (ms) 28.6 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nlqw0c0w",
   "source_url": "https://omlx.ai/benchmarks/performance/nlqw0c0w",
   "tg_tokens": null,
   "tps": 134.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1jf4cag6-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1050.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 4k 1,050 78.1 143.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1jf4cag6",
   "source_url": "https://omlx.ai/benchmarks/performance/1jf4cag6",
   "tg_tokens": null,
   "tps": 78.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1jf4cag6-ctx16384",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 143.6 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1035.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 16k 1,035 70.4 143.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1jf4cag6",
   "source_url": "https://omlx.ai/benchmarks/performance/1jf4cag6",
   "tg_tokens": null,
   "tps": 70.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-m12bv1c3-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: b9jsl8g8, er074vg8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 582.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 582.0 64.5 70.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m12bv1c3",
   "source_url": "https://omlx.ai/benchmarks/performance/m12bv1c3",
   "tg_tokens": null,
   "tps": 64.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-m12bv1c3-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: b9jsl8g8, er074vg8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 708.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 708.9 56.3 71.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m12bv1c3",
   "source_url": "https://omlx.ai/benchmarks/performance/m12bv1c3",
   "tg_tokens": null,
   "tps": 56.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-m12bv1c3-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: b9jsl8g8, er074vg8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 810.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 810.7 72.8 71.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m12bv1c3",
   "source_url": "https://omlx.ai/benchmarks/performance/m12bv1c3",
   "tg_tokens": null,
   "tps": 72.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ln4rc4mo-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 23bikf10; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 469.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 469.6 37.2 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ln4rc4mo",
   "source_url": "https://omlx.ai/benchmarks/performance/ln4rc4mo",
   "tg_tokens": null,
   "tps": 37.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ln4rc4mo-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: 23bikf10; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 581.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 581.0 31.8 71.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ln4rc4mo",
   "source_url": "https://omlx.ai/benchmarks/performance/ln4rc4mo",
   "tg_tokens": null,
   "tps": 31.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jt3g6g1c-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: 0uxfpkm3, 9b8611hn, ak082iqt, syfo5ndo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1832.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,832 98.7 100.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jt3g6g1c",
   "source_url": "https://omlx.ai/benchmarks/performance/jt3g6g1c",
   "tg_tokens": null,
   "tps": 98.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jt3g6g1c-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: 0uxfpkm3, 9b8611hn, ak082iqt, syfo5ndo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2882.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,882 95.8 101.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jt3g6g1c",
   "source_url": "https://omlx.ai/benchmarks/performance/jt3g6g1c",
   "tg_tokens": null,
   "tps": 95.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jt3g6g1c-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 102.2 GB at this context; duplicate traces of the same session: 0uxfpkm3, 9b8611hn, ak082iqt, syfo5ndo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2742.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,742 50.8 102.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jt3g6g1c",
   "source_url": "https://omlx.ai/benchmarks/performance/jt3g6g1c",
   "tg_tokens": null,
   "tps": 50.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jt3g6g1c-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 102.6 GB at this context; duplicate traces of the same session: 0uxfpkm3, 9b8611hn, ak082iqt, syfo5ndo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2794.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,794 74.5 102.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jt3g6g1c",
   "source_url": "https://omlx.ai/benchmarks/performance/jt3g6g1c",
   "tg_tokens": null,
   "tps": 74.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jt3g6g1c-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 103.2 GB at this context; duplicate traces of the same session: 0uxfpkm3, 9b8611hn, ak082iqt, syfo5ndo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2456.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,456 64.9 103.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jt3g6g1c",
   "source_url": "https://omlx.ai/benchmarks/performance/jt3g6g1c",
   "tg_tokens": null,
   "tps": 64.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-zo0bbak4-ctx1024",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: f85hxhj8; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 587.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31B-it-MLX-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 587.3 21.8 17.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zo0bbak4",
   "source_url": "https://omlx.ai/benchmarks/performance/zo0bbak4",
   "tg_tokens": null,
   "tps": 21.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-zo0bbak4-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: f85hxhj8; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 604.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31B-it-MLX-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 604.0 21.4 18.8 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zo0bbak4",
   "source_url": "https://omlx.ai/benchmarks/performance/zo0bbak4",
   "tg_tokens": null,
   "tps": 21.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-yidf23rq-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MLX Serve build; oMLX community benchmark trace, single run at 128k context; peak memory 22.5 GB; TTFT 50897 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2575.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-Serve-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k tokens 2,575 PP tok/s 75.6 TG tok/s 50897 TTFT (ms) 22.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yidf23rq",
   "source_url": "https://omlx.ai/benchmarks/performance/yidf23rq",
   "tg_tokens": null,
   "tps": 75.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-rlrtv4aa-ctx65536",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 28.9 GB at this context; duplicate traces of the same session: 1xb9cml0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 128.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 128.7 18.7 28.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rlrtv4aa",
   "source_url": "https://omlx.ai/benchmarks/performance/rlrtv4aa",
   "tg_tokens": null,
   "tps": 18.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-rlrtv4aa-ctx131072",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 33.4 GB at this context; duplicate traces of the same session: 1xb9cml0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 99.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 99.4 16.6 33.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rlrtv4aa",
   "source_url": "https://omlx.ai/benchmarks/performance/rlrtv4aa",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lwuyq3x7-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: 4kmpoga0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 340.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 340.4 51.9 22.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lwuyq3x7",
   "source_url": "https://omlx.ai/benchmarks/performance/lwuyq3x7",
   "tg_tokens": null,
   "tps": 51.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lwuyq3x7-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 4kmpoga0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 328.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 328.0 42.2 23.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lwuyq3x7",
   "source_url": "https://omlx.ai/benchmarks/performance/lwuyq3x7",
   "tg_tokens": null,
   "tps": 42.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-6mjmmrqa-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 71.5 GB; TTFT 8995 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 910.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k tokens 910.8 PP tok/s 65.8 TG tok/s 8995 TTFT (ms) 71.5 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6mjmmrqa",
   "source_url": "https://omlx.ai/benchmarks/performance/6mjmmrqa",
   "tg_tokens": null,
   "tps": 65.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-0nlizmgx-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 71.6 GB; TTFT 9521 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 860.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k tokens 860.4 PP tok/s 56.8 TG tok/s 9521 TTFT (ms) 71.6 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0nlizmgx",
   "source_url": "https://omlx.ai/benchmarks/performance/0nlizmgx",
   "tg_tokens": null,
   "tps": 56.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-za8zitzp-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.0 GB at this context; duplicate traces of the same session: xq7l1g0r; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2710.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 2,710 110.9 99.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/za8zitzp",
   "source_url": "https://omlx.ai/benchmarks/performance/za8zitzp",
   "tg_tokens": null,
   "tps": 110.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-za8zitzp-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.5 GB at this context; duplicate traces of the same session: xq7l1g0r; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4502.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 4,502 100.2 99.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/za8zitzp",
   "source_url": "https://omlx.ai/benchmarks/performance/za8zitzp",
   "tg_tokens": null,
   "tps": 100.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-f6803e2a-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 139.4 GB at this context; duplicate traces of the same session: d91vrwvd, mje01m0p, n5l2ttp8, ppkq6r35; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1755.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 1,755 85.3 139.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f6803e2a",
   "source_url": "https://omlx.ai/benchmarks/performance/f6803e2a",
   "tg_tokens": null,
   "tps": 85.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-f6803e2a-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 141.1 GB at this context; duplicate traces of the same session: d91vrwvd, mje01m0p, n5l2ttp8, ppkq6r35; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4649.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 8k 4,649 79.8 141.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f6803e2a",
   "source_url": "https://omlx.ai/benchmarks/performance/f6803e2a",
   "tg_tokens": null,
   "tps": 79.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-f6803e2a-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.8 GB at this context; duplicate traces of the same session: d91vrwvd, mje01m0p, n5l2ttp8, ppkq6r35; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4962.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 64k 4,962 77.3 142.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f6803e2a",
   "source_url": "https://omlx.ai/benchmarks/performance/f6803e2a",
   "tg_tokens": null,
   "tps": 77.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-f6803e2a-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.0 GB at this context; duplicate traces of the same session: d91vrwvd, mje01m0p, n5l2ttp8, ppkq6r35; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4827.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 128k 4,827 70.6 145.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f6803e2a",
   "source_url": "https://omlx.ai/benchmarks/performance/f6803e2a",
   "tg_tokens": null,
   "tps": 70.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-f6803e2a-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 148.7 GB at this context; duplicate traces of the same session: d91vrwvd, mje01m0p, n5l2ttp8, ppkq6r35; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4692.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 195k 4,692 70.0 148.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f6803e2a",
   "source_url": "https://omlx.ai/benchmarks/performance/f6803e2a",
   "tg_tokens": null,
   "tps": 70.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-4jxkxv0g-ctx1024",
   "model": "Bonsai 2 27B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: oxi489hc; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 248.1,
   "provenance": "community",
   "quant": "unknown",
   "quote": "bonsai2-27b-mlx on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 unknown; 1k 248.1 22.4 11.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4jxkxv0g",
   "source_url": "https://omlx.ai/benchmarks/performance/4jxkxv0g",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-4jxkxv0g-ctx4096",
   "model": "Bonsai 2 27B",
   "notes": "oMLX community benchmark trace; peak memory 14.1 GB at this context; duplicate traces of the same session: oxi489hc; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 289.3,
   "provenance": "community",
   "quant": "unknown",
   "quote": "bonsai2-27b-mlx on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 unknown; 4k 289.3 21.3 14.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4jxkxv0g",
   "source_url": "https://omlx.ai/benchmarks/performance/4jxkxv0g",
   "tg_tokens": null,
   "tps": 21.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mg2962t0-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: m5keqwfa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 823.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 823.1 18.5 28.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mg2962t0",
   "source_url": "https://omlx.ai/benchmarks/performance/mg2962t0",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mg2962t0-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: m5keqwfa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 950.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 950.9 18.4 29.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mg2962t0",
   "source_url": "https://omlx.ai/benchmarks/performance/mg2962t0",
   "tg_tokens": null,
   "tps": 18.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8pzv0nfo-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: jzjb9g7f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 830.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 830.8 18.5 28.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8pzv0nfo",
   "source_url": "https://omlx.ai/benchmarks/performance/8pzv0nfo",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8pzv0nfo-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: jzjb9g7f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 929.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 929.1 18.0 29.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8pzv0nfo",
   "source_url": "https://omlx.ai/benchmarks/performance/8pzv0nfo",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qodfe5be-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.6 GB at this context; duplicate traces of the same session: 2wdkpjyc, 7ooo9q7j, 8vgzatib, bpcdjloe; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 408.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 408.3 42.0 173.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qodfe5be",
   "source_url": "https://omlx.ai/benchmarks/performance/qodfe5be",
   "tg_tokens": null,
   "tps": 42.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qodfe5be-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.8 GB at this context; duplicate traces of the same session: 2wdkpjyc, 7ooo9q7j, 8vgzatib, bpcdjloe; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 462.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 462.4 40.0 174.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qodfe5be",
   "source_url": "https://omlx.ai/benchmarks/performance/qodfe5be",
   "tg_tokens": null,
   "tps": 40.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qodfe5be-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.1 GB at this context; duplicate traces of the same session: 2wdkpjyc, 7ooo9q7j, 8vgzatib, bpcdjloe; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 427.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 427.0 40.1 175.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qodfe5be",
   "source_url": "https://omlx.ai/benchmarks/performance/qodfe5be",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qodfe5be-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.2 GB at this context; duplicate traces of the same session: 2wdkpjyc, 7ooo9q7j, 8vgzatib, bpcdjloe; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 418.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 418.4 39.6 175.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qodfe5be",
   "source_url": "https://omlx.ai/benchmarks/performance/qodfe5be",
   "tg_tokens": null,
   "tps": 39.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qodfe5be-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.3 GB at this context; duplicate traces of the same session: 2wdkpjyc, 7ooo9q7j, 8vgzatib, bpcdjloe; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 412.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 412.3 38.9 175.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qodfe5be",
   "source_url": "https://omlx.ai/benchmarks/performance/qodfe5be",
   "tg_tokens": null,
   "tps": 38.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-6flcyslc-ctx1024",
   "model": "Ornith-1.5-35B-A3B-REAP192",
   "notes": "oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: j6a5x5ss; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 960.5,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Ornith-1.5-35B-A3B-REAP192-mxfp4-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 960.5 97.9 14.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6flcyslc",
   "source_url": "https://omlx.ai/benchmarks/performance/6flcyslc",
   "tg_tokens": null,
   "tps": 97.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-6flcyslc-ctx32768",
   "model": "Ornith-1.5-35B-A3B-REAP192",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: j6a5x5ss; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 741.4,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Ornith-1.5-35B-A3B-REAP192-mxfp4-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 741.4 70.5 15.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6flcyslc",
   "source_url": "https://omlx.ai/benchmarks/performance/6flcyslc",
   "tg_tokens": null,
   "tps": 70.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gaytq7c2-ctx1024",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "MLX; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 0jwthfvy; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2102.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 2,102 100.1 16.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gaytq7c2",
   "source_url": "https://omlx.ai/benchmarks/performance/gaytq7c2",
   "tg_tokens": null,
   "tps": 100.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gaytq7c2-ctx4096",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "MLX; oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: 0jwthfvy; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2037.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 2,037 92.0 17.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gaytq7c2",
   "source_url": "https://omlx.ai/benchmarks/performance/gaytq7c2",
   "tg_tokens": null,
   "tps": 92.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-zd6ymxxj-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 60woy1sp; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2260.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 2,260 99.3 20.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zd6ymxxj",
   "source_url": "https://omlx.ai/benchmarks/performance/zd6ymxxj",
   "tg_tokens": null,
   "tps": 99.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-zd6ymxxj-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 60woy1sp; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2537.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 2,537 98.3 22.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zd6ymxxj",
   "source_url": "https://omlx.ai/benchmarks/performance/zd6ymxxj",
   "tg_tokens": null,
   "tps": 98.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-invt5vuq-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 20.6 GB; TTFT 646 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1586.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k tokens 1,586 PP tok/s 98.8 TG tok/s 646 TTFT (ms) 20.6 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/invt5vuq",
   "source_url": "https://omlx.ai/benchmarks/performance/invt5vuq",
   "tg_tokens": null,
   "tps": 98.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-y77xu2fo-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 4ob3ic2h, 5qgjn6ob, bclaeknv, ecokcv1g; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2180.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 2,180 103.3 21.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y77xu2fo",
   "source_url": "https://omlx.ai/benchmarks/performance/y77xu2fo",
   "tg_tokens": null,
   "tps": 103.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-y77xu2fo-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 4ob3ic2h, 5qgjn6ob, bclaeknv, ecokcv1g; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2583.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 2,583 129.3 23.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y77xu2fo",
   "source_url": "https://omlx.ai/benchmarks/performance/y77xu2fo",
   "tg_tokens": null,
   "tps": 129.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-y77xu2fo-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: 4ob3ic2h, 5qgjn6ob, bclaeknv, ecokcv1g; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2469.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 2,469 75.9 23.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y77xu2fo",
   "source_url": "https://omlx.ai/benchmarks/performance/y77xu2fo",
   "tg_tokens": null,
   "tps": 75.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-y77xu2fo-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 4ob3ic2h, 5qgjn6ob, bclaeknv, ecokcv1g; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2224.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 2,224 118.5 23.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y77xu2fo",
   "source_url": "https://omlx.ai/benchmarks/performance/y77xu2fo",
   "tg_tokens": null,
   "tps": 118.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-y77xu2fo-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: 4ob3ic2h, 5qgjn6ob, bclaeknv, ecokcv1g; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1803.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 1,803 87.0 24.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y77xu2fo",
   "source_url": "https://omlx.ai/benchmarks/performance/y77xu2fo",
   "tg_tokens": null,
   "tps": 87.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-1k7knbgp-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: lr6is3vd; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 142.6,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 142.6 12.5 16.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1k7knbgp",
   "source_url": "https://omlx.ai/benchmarks/performance/1k7knbgp",
   "tg_tokens": null,
   "tps": 12.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-1k7knbgp-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: lr6is3vd; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 147.9,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 147.9 11.8 20.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1k7knbgp",
   "source_url": "https://omlx.ai/benchmarks/performance/1k7knbgp",
   "tg_tokens": null,
   "tps": 11.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-h6acb165-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 502z0ioc, 654j8071, au82jbmg; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 317.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 317.7 57.0 18.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h6acb165",
   "source_url": "https://omlx.ai/benchmarks/performance/h6acb165",
   "tg_tokens": null,
   "tps": 57.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-h6acb165-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 502z0ioc, 654j8071, au82jbmg; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 346.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 346.5 55.7 22.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h6acb165",
   "source_url": "https://omlx.ai/benchmarks/performance/h6acb165",
   "tg_tokens": null,
   "tps": 55.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-h6acb165-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: 502z0ioc, 654j8071, au82jbmg; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 340.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 340.9 52.2 22.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h6acb165",
   "source_url": "https://omlx.ai/benchmarks/performance/h6acb165",
   "tg_tokens": null,
   "tps": 52.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-h6acb165-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 502z0ioc, 654j8071, au82jbmg; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 328.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 328.0 48.4 23.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h6acb165",
   "source_url": "https://omlx.ai/benchmarks/performance/h6acb165",
   "tg_tokens": null,
   "tps": 48.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-z7zxxv2m-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: v8tkqvc8; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 316.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 316.2 43.3 18.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z7zxxv2m",
   "source_url": "https://omlx.ai/benchmarks/performance/z7zxxv2m",
   "tg_tokens": null,
   "tps": 43.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-z7zxxv2m-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: v8tkqvc8; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 336.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 336.6 72.3 21.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z7zxxv2m",
   "source_url": "https://omlx.ai/benchmarks/performance/z7zxxv2m",
   "tg_tokens": null,
   "tps": 72.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-xgkgjkzx-ctx1024",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: zktd1dfa; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 224.7 30.0 6.2 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xgkgjkzx",
   "source_url": "https://omlx.ai/benchmarks/performance/xgkgjkzx",
   "tg_tokens": null,
   "tps": 30.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-xgkgjkzx-ctx4096",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: zktd1dfa; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 227.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 227.0 31.8 6.7 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xgkgjkzx",
   "source_url": "https://omlx.ai/benchmarks/performance/xgkgjkzx",
   "tg_tokens": null,
   "tps": 31.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-pn4x3fbn-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: 1v13jl10, u4t6clb9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 693.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 693.7 63.4 70.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pn4x3fbn",
   "source_url": "https://omlx.ai/benchmarks/performance/pn4x3fbn",
   "tg_tokens": null,
   "tps": 63.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-pn4x3fbn-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: 1v13jl10, u4t6clb9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 863.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 863.6 71.9 71.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pn4x3fbn",
   "source_url": "https://omlx.ai/benchmarks/performance/pn4x3fbn",
   "tg_tokens": null,
   "tps": 71.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-pn4x3fbn-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: 1v13jl10, u4t6clb9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1063.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,063 80.1 71.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pn4x3fbn",
   "source_url": "https://omlx.ai/benchmarks/performance/pn4x3fbn",
   "tg_tokens": null,
   "tps": 80.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-oclyedlo-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: 4lir3eck, imo7m6vt; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1064.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,064 76.7 71.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oclyedlo",
   "source_url": "https://omlx.ai/benchmarks/performance/oclyedlo",
   "tg_tokens": null,
   "tps": 76.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-oclyedlo-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 4lir3eck, imo7m6vt; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1060.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 1,060 66.0 71.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oclyedlo",
   "source_url": "https://omlx.ai/benchmarks/performance/oclyedlo",
   "tg_tokens": null,
   "tps": 66.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-oclyedlo-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.5 GB at this context; duplicate traces of the same session: 4lir3eck, imo7m6vt; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1054.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 1,054 32.8 72.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oclyedlo",
   "source_url": "https://omlx.ai/benchmarks/performance/oclyedlo",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-a9d4wsie-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 71.5 GB; TTFT 9006 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 909.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k tokens 909.6 PP tok/s 70.7 TG tok/s 9006 TTFT (ms) 71.5 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a9d4wsie",
   "source_url": "https://omlx.ai/benchmarks/performance/a9d4wsie",
   "tg_tokens": null,
   "tps": 70.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-6fyhak7y-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 122.2 GB at this context; duplicate traces of the same session: eo3wat4j; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2066.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 5bit; 1k 2,066 105.3 122.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6fyhak7y",
   "source_url": "https://omlx.ai/benchmarks/performance/6fyhak7y",
   "tg_tokens": null,
   "tps": 105.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-6fyhak7y-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 123.9 GB at this context; duplicate traces of the same session: eo3wat4j; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2512.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 5bit; 8k 2,512 115.7 123.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6fyhak7y",
   "source_url": "https://omlx.ai/benchmarks/performance/6fyhak7y",
   "tg_tokens": null,
   "tps": 115.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-au3vzmyo-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "scottlowry ThinkingCap; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: jr9ltfrd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 904.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "scottlowry--ThinkingCap-Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 904.6 31.1 17.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/au3vzmyo",
   "source_url": "https://omlx.ai/benchmarks/performance/au3vzmyo",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-au3vzmyo-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "scottlowry ThinkingCap; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: jr9ltfrd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 874.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "scottlowry--ThinkingCap-Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 874.3 29.4 18.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/au3vzmyo",
   "source_url": "https://omlx.ai/benchmarks/performance/au3vzmyo",
   "tg_tokens": null,
   "tps": 29.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-zfdmfhwx-ctx1024",
   "model": "Huihui-Ornith-1.5-35B-A3B-abliterated",
   "notes": "oMLX community benchmark trace; peak memory 68.3 GB at this context; duplicate traces of the same session: 32u4cc0g, 4yt5ydch; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1097.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Huihui-Ornith-1.5-35B-A3B-abliterated on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 1,097 52.5 68.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zfdmfhwx",
   "source_url": "https://omlx.ai/benchmarks/performance/zfdmfhwx",
   "tg_tokens": null,
   "tps": 52.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-zfdmfhwx-ctx4096",
   "model": "Huihui-Ornith-1.5-35B-A3B-abliterated",
   "notes": "oMLX community benchmark trace; peak memory 70.1 GB at this context; duplicate traces of the same session: 32u4cc0g, 4yt5ydch; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1627.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Huihui-Ornith-1.5-35B-A3B-abliterated on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 unknown; 4k 1,627 54.9 70.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zfdmfhwx",
   "source_url": "https://omlx.ai/benchmarks/performance/zfdmfhwx",
   "tg_tokens": null,
   "tps": 54.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-zfdmfhwx-ctx16384",
   "model": "Huihui-Ornith-1.5-35B-A3B-abliterated",
   "notes": "oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: 32u4cc0g, 4yt5ydch; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1365.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Huihui-Ornith-1.5-35B-A3B-abliterated on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 1,365 63.5 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zfdmfhwx",
   "source_url": "https://omlx.ai/benchmarks/performance/zfdmfhwx",
   "tg_tokens": null,
   "tps": 63.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-85htyahw-ctx131072",
   "model": "Ornith-1.5-35B-A3B-REAP192",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 18.6 GB; TTFT 363795 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 360.3,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Ornith-1.5-35B-A3B-REAP192-mxfp4-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 128k tokens 360.3 PP tok/s 38.5 TG tok/s 363795 TTFT (ms) 18.6 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/85htyahw",
   "source_url": "https://omlx.ai/benchmarks/performance/85htyahw",
   "tg_tokens": null,
   "tps": 38.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-c1kh4pz5-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V10-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: 43vja1e5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3262.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V10-dequantized-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 3,262 141.5 37.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c1kh4pz5",
   "source_url": "https://omlx.ai/benchmarks/performance/c1kh4pz5",
   "tg_tokens": null,
   "tps": 141.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-c1kh4pz5-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V10-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.0 GB at this context; duplicate traces of the same session: 43vja1e5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5310.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V10-dequantized-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 5,310 137.8 38.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c1kh4pz5",
   "source_url": "https://omlx.ai/benchmarks/performance/c1kh4pz5",
   "tg_tokens": null,
   "tps": 137.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-c6qq0xir-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 239y6k0m; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1250.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 1,250 111.4 20.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c6qq0xir",
   "source_url": "https://omlx.ai/benchmarks/performance/c6qq0xir",
   "tg_tokens": null,
   "tps": 111.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-c6qq0xir-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: 239y6k0m; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1176.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 1,176 90.9 22.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c6qq0xir",
   "source_url": "https://omlx.ai/benchmarks/performance/c6qq0xir",
   "tg_tokens": null,
   "tps": 90.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-h9auz0v3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 33.7 GB at this context; duplicate traces of the same session: 8xsqh2ks; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 784.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 784.5 45.4 33.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h9auz0v3",
   "source_url": "https://omlx.ai/benchmarks/performance/h9auz0v3",
   "tg_tokens": null,
   "tps": 45.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-h9auz0v3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: 8xsqh2ks; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 833.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 833.2 38.8 34.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h9auz0v3",
   "source_url": "https://omlx.ai/benchmarks/performance/h9auz0v3",
   "tg_tokens": null,
   "tps": 38.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ivedqm6o-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 28.6 GB; TTFT 84436 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 388.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k tokens 388.1 PP tok/s 41.0 TG tok/s 84436 TTFT (ms) 28.6 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ivedqm6o",
   "source_url": "https://omlx.ai/benchmarks/performance/ivedqm6o",
   "tg_tokens": null,
   "tps": 41.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-m7s67ip1-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: lgd2j6rd; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 296.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 296.6 9.2 16.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m7s67ip1",
   "source_url": "https://omlx.ai/benchmarks/performance/m7s67ip1",
   "tg_tokens": null,
   "tps": 9.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-m7s67ip1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: lgd2j6rd; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 308.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 308.3 9.0 17.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m7s67ip1",
   "source_url": "https://omlx.ai/benchmarks/performance/m7s67ip1",
   "tg_tokens": null,
   "tps": 9.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-jd7py4x0-ctx1024",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: ph1vpj2w; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 224.2 34.2 6.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jd7py4x0",
   "source_url": "https://omlx.ai/benchmarks/performance/jd7py4x0",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-jd7py4x0-ctx4096",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: ph1vpj2w; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 232.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 232.2 30.1 6.7 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jd7py4x0",
   "source_url": "https://omlx.ai/benchmarks/performance/jd7py4x0",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2romjj4x-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 47.2 GB at this context; duplicate traces of the same session: yt5ge4rj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 124.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 124.3 14.2 47.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2romjj4x",
   "source_url": "https://omlx.ai/benchmarks/performance/2romjj4x",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2romjj4x-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 47.5 GB at this context; duplicate traces of the same session: yt5ge4rj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 121.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 121.3 13.8 47.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2romjj4x",
   "source_url": "https://omlx.ai/benchmarks/performance/2romjj4x",
   "tg_tokens": null,
   "tps": 13.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nucczs4k-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 61.6 GB at this context; duplicate traces of the same session: mno550ga; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 151.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 151.1 22.0 61.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nucczs4k",
   "source_url": "https://omlx.ai/benchmarks/performance/nucczs4k",
   "tg_tokens": null,
   "tps": 22.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nucczs4k-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 61.9 GB at this context; duplicate traces of the same session: mno550ga; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 119.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 119.7 20.3 61.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nucczs4k",
   "source_url": "https://omlx.ai/benchmarks/performance/nucczs4k",
   "tg_tokens": null,
   "tps": 20.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-u92lc83h-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 70.3 GB at this context; duplicate traces of the same session: olxf7zi1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1950.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,950 90.0 70.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u92lc83h",
   "source_url": "https://omlx.ai/benchmarks/performance/u92lc83h",
   "tg_tokens": null,
   "tps": 90.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-u92lc83h-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: olxf7zi1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2912.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,912 84.4 71.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u92lc83h",
   "source_url": "https://omlx.ai/benchmarks/performance/u92lc83h",
   "tg_tokens": null,
   "tps": 84.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fa6qhjg1-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: y9xu6p1a; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 766.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 766.8 109.2 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fa6qhjg1",
   "source_url": "https://omlx.ai/benchmarks/performance/fa6qhjg1",
   "tg_tokens": null,
   "tps": 109.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fa6qhjg1-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: y9xu6p1a; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2020.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,020 104.0 20.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fa6qhjg1",
   "source_url": "https://omlx.ai/benchmarks/performance/fa6qhjg1",
   "tg_tokens": null,
   "tps": 104.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ccj3tiws-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 33.4 GB; TTFT 18597 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1762.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 32k tokens 1,762 PP tok/s 145.2 TG tok/s 18597 TTFT (ms) 33.4 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ccj3tiws",
   "source_url": "https://omlx.ai/benchmarks/performance/ccj3tiws",
   "tg_tokens": null,
   "tps": 145.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ibp00b2b-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 33.5 GB; TTFT 19415 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1688.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 32k tokens 1,688 PP tok/s 133.4 TG tok/s 19415 TTFT (ms) 33.5 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ibp00b2b",
   "source_url": "https://omlx.ai/benchmarks/performance/ibp00b2b",
   "tg_tokens": null,
   "tps": 133.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-dpkohagz-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 33.4 GB; TTFT 18590 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1763.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 32k tokens 1,763 PP tok/s 144.3 TG tok/s 18590 TTFT (ms) 33.4 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dpkohagz",
   "source_url": "https://omlx.ai/benchmarks/performance/dpkohagz",
   "tg_tokens": null,
   "tps": 144.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-46lxhu5y-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 33.4 GB; TTFT 18594 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1762.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 32k tokens 1,762 PP tok/s 141.2 TG tok/s 18594 TTFT (ms) 33.4 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/46lxhu5y",
   "source_url": "https://omlx.ai/benchmarks/performance/46lxhu5y",
   "tg_tokens": null,
   "tps": 141.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-n5dhxbwb-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 33.4 GB; TTFT 18595 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1762.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 32k tokens 1,762 PP tok/s 182.2 TG tok/s 18595 TTFT (ms) 33.4 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n5dhxbwb",
   "source_url": "https://omlx.ai/benchmarks/performance/n5dhxbwb",
   "tg_tokens": null,
   "tps": 182.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-bxkdui81-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 33.4 GB; TTFT 18696 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1753.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 32k tokens 1,753 PP tok/s 120.0 TG tok/s 18696 TTFT (ms) 33.4 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bxkdui81",
   "source_url": "https://omlx.ai/benchmarks/performance/bxkdui81",
   "tg_tokens": null,
   "tps": 120.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gfnzs8zw-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 33.4 GB; TTFT 18593 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1762.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 32k tokens 1,762 PP tok/s 147.9 TG tok/s 18593 TTFT (ms) 33.4 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gfnzs8zw",
   "source_url": "https://omlx.ai/benchmarks/performance/gfnzs8zw",
   "tg_tokens": null,
   "tps": 147.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-c22zn1px-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 26.0 GB; TTFT 107604 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 304.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k tokens 304.5 PP tok/s 66.4 TG tok/s 107604 TTFT (ms) 26.0 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c22zn1px",
   "source_url": "https://omlx.ai/benchmarks/performance/c22zn1px",
   "tg_tokens": null,
   "tps": 66.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-8wjs907u-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 28.6 GB; TTFT 84570 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 387.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k tokens 387.5 PP tok/s 75.8 TG tok/s 84570 TTFT (ms) 28.6 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8wjs907u",
   "source_url": "https://omlx.ai/benchmarks/performance/8wjs907u",
   "tg_tokens": null,
   "tps": 75.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8ljge6re-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 65.1 GB at this context; duplicate traces of the same session: n4baow8n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 142.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 142.8 22.3 65.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8ljge6re",
   "source_url": "https://omlx.ai/benchmarks/performance/8ljge6re",
   "tg_tokens": null,
   "tps": 22.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8ljge6re-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 65.5 GB at this context; duplicate traces of the same session: n4baow8n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 160.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 160.5 23.5 65.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8ljge6re",
   "source_url": "https://omlx.ai/benchmarks/performance/8ljge6re",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-i4kdaak4-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace, single run at 64k context; peak memory 82.9 GB; TTFT 84092 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 779.3,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ4-MTP on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k tokens 779.3 PP tok/s 39.0 TG tok/s 84092 TTFT (ms) 82.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i4kdaak4",
   "source_url": "https://omlx.ai/benchmarks/performance/i4kdaak4",
   "tg_tokens": null,
   "tps": 39.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9wyj76h4-ctx131072",
   "model": "Swift 1.5 Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 150.3 GB; TTFT 131998 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 993.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 128k tokens 993.0 PP tok/s 50.8 TG tok/s 131998 TTFT (ms) 150.3 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9wyj76h4",
   "source_url": "https://omlx.ai/benchmarks/performance/9wyj76h4",
   "tg_tokens": null,
   "tps": 50.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-4j1avgx2-ctx131072",
   "model": "Swift 1.5 Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 150.4 GB; TTFT 133476 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 982.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 128k tokens 982.0 PP tok/s 30.6 TG tok/s 133476 TTFT (ms) 150.4 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4j1avgx2",
   "source_url": "https://omlx.ai/benchmarks/performance/4j1avgx2",
   "tg_tokens": null,
   "tps": 30.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-mc9vtjow-ctx1024",
   "model": "GLM-4.7-Flash",
   "notes": "oMLX community benchmark trace; peak memory 29.4 GB at this context; duplicate traces of the same session: 99ke3tg0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1213.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "GLM-4.7-Flash-8bit-mxfp8 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 1,213 59.8 29.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mc9vtjow",
   "source_url": "https://omlx.ai/benchmarks/performance/mc9vtjow",
   "tg_tokens": null,
   "tps": 59.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-mc9vtjow-ctx4096",
   "model": "GLM-4.7-Flash",
   "notes": "oMLX community benchmark trace; peak memory 30.4 GB at this context; duplicate traces of the same session: 99ke3tg0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1114.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "GLM-4.7-Flash-8bit-mxfp8 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 1,114 56.6 30.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mc9vtjow",
   "source_url": "https://omlx.ai/benchmarks/performance/mc9vtjow",
   "tg_tokens": null,
   "tps": 56.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ldwwu29j-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.7 GB at this context; duplicate traces of the same session: 0nnzd90n, bo2ojm0n, d2i9z85b, rqbqmpxa, tyllbk9e; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 446.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 446.0 40.4 173.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ldwwu29j",
   "source_url": "https://omlx.ai/benchmarks/performance/ldwwu29j",
   "tg_tokens": null,
   "tps": 40.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ldwwu29j-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.8 GB at this context; duplicate traces of the same session: 0nnzd90n, bo2ojm0n, d2i9z85b, rqbqmpxa, tyllbk9e; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 566.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 566.4 38.6 174.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ldwwu29j",
   "source_url": "https://omlx.ai/benchmarks/performance/ldwwu29j",
   "tg_tokens": null,
   "tps": 38.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ldwwu29j-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.4 GB at this context; duplicate traces of the same session: 0nnzd90n, bo2ojm0n, d2i9z85b, rqbqmpxa, tyllbk9e; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 509.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 509.3 38.6 175.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ldwwu29j",
   "source_url": "https://omlx.ai/benchmarks/performance/ldwwu29j",
   "tg_tokens": null,
   "tps": 38.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ldwwu29j-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.5 GB at this context; duplicate traces of the same session: 0nnzd90n, bo2ojm0n, d2i9z85b, rqbqmpxa, tyllbk9e; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 550.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 550.4 38.3 175.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ldwwu29j",
   "source_url": "https://omlx.ai/benchmarks/performance/ldwwu29j",
   "tg_tokens": null,
   "tps": 38.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ldwwu29j-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.7 GB at this context; duplicate traces of the same session: 0nnzd90n, bo2ojm0n, d2i9z85b, rqbqmpxa, tyllbk9e; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 542.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 542.8 37.7 175.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ldwwu29j",
   "source_url": "https://omlx.ai/benchmarks/performance/ldwwu29j",
   "tg_tokens": null,
   "tps": 37.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ldwwu29j-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 176.0 GB at this context; duplicate traces of the same session: 0nnzd90n, bo2ojm0n, d2i9z85b, rqbqmpxa, tyllbk9e; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 534.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 534.8 36.8 176.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ldwwu29j",
   "source_url": "https://omlx.ai/benchmarks/performance/ldwwu29j",
   "tg_tokens": null,
   "tps": 36.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-l4pknw0u-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 181.6 GB; TTFT 334826 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 391.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 128k tokens 391.5 PP tok/s 75.0 TG tok/s 334826 TTFT (ms) 181.6 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l4pknw0u",
   "source_url": "https://omlx.ai/benchmarks/performance/l4pknw0u",
   "tg_tokens": null,
   "tps": 75.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-8fzhub0q-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 181.6 GB; TTFT 339211 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 386.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 128k tokens 386.4 PP tok/s 56.2 TG tok/s 339211 TTFT (ms) 181.6 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8fzhub0q",
   "source_url": "https://omlx.ai/benchmarks/performance/8fzhub0q",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-o9fjubfr-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: 2bqgc01z, 35fzkc9g, ddcm2426, nvmo4bn9, qgbyh779, sx0epgus, w7ib4vxs; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3146.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 3,146 104.5 5.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9fjubfr",
   "source_url": "https://omlx.ai/benchmarks/performance/o9fjubfr",
   "tg_tokens": null,
   "tps": 104.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-o9fjubfr-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: 2bqgc01z, 35fzkc9g, ddcm2426, nvmo4bn9, qgbyh779, sx0epgus, w7ib4vxs; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3907.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 3,907 102.6 6.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9fjubfr",
   "source_url": "https://omlx.ai/benchmarks/performance/o9fjubfr",
   "tg_tokens": null,
   "tps": 102.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-o9fjubfr-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: 2bqgc01z, 35fzkc9g, ddcm2426, nvmo4bn9, qgbyh779, sx0epgus, w7ib4vxs; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4545.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 8k 4,545 98.9 6.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9fjubfr",
   "source_url": "https://omlx.ai/benchmarks/performance/o9fjubfr",
   "tg_tokens": null,
   "tps": 98.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-o9fjubfr-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.1 GB at this context; duplicate traces of the same session: 2bqgc01z, 35fzkc9g, ddcm2426, nvmo4bn9, qgbyh779, sx0epgus, w7ib4vxs; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4563.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 4,563 95.7 7.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9fjubfr",
   "source_url": "https://omlx.ai/benchmarks/performance/o9fjubfr",
   "tg_tokens": null,
   "tps": 95.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-o9fjubfr-ctx32768",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.6 GB at this context; duplicate traces of the same session: 2bqgc01z, 35fzkc9g, ddcm2426, nvmo4bn9, qgbyh779, sx0epgus, w7ib4vxs; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4193.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 4,193 86.7 7.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9fjubfr",
   "source_url": "https://omlx.ai/benchmarks/performance/o9fjubfr",
   "tg_tokens": null,
   "tps": 86.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-o9fjubfr-ctx65536",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.6 GB at this context; duplicate traces of the same session: 2bqgc01z, 35fzkc9g, ddcm2426, nvmo4bn9, qgbyh779, sx0epgus, w7ib4vxs; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3369.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 64k 3,369 73.4 8.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9fjubfr",
   "source_url": "https://omlx.ai/benchmarks/performance/o9fjubfr",
   "tg_tokens": null,
   "tps": 73.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-o9fjubfr-ctx131072",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.6 GB at this context; duplicate traces of the same session: 2bqgc01z, 35fzkc9g, ddcm2426, nvmo4bn9, qgbyh779, sx0epgus, w7ib4vxs; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2516.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 128k 2,516 56.7 10.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9fjubfr",
   "source_url": "https://omlx.ai/benchmarks/performance/o9fjubfr",
   "tg_tokens": null,
   "tps": 56.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-o9fjubfr-ctx199680",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: 2bqgc01z, 35fzkc9g, ddcm2426, nvmo4bn9, qgbyh779, sx0epgus, w7ib4vxs; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1930.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 195k 1,930 44.9 12.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9fjubfr",
   "source_url": "https://omlx.ai/benchmarks/performance/o9fjubfr",
   "tg_tokens": null,
   "tps": 44.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dh5qzzvc-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: 0mpne123, 3njcyo0n, 3tpmnfvl, ij4lrk1x, mhfap76i, r30t9wq5, xe5jpdpp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3129.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 3,129 104.2 5.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dh5qzzvc",
   "source_url": "https://omlx.ai/benchmarks/performance/dh5qzzvc",
   "tg_tokens": null,
   "tps": 104.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dh5qzzvc-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: 0mpne123, 3njcyo0n, 3tpmnfvl, ij4lrk1x, mhfap76i, r30t9wq5, xe5jpdpp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3500.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 3,500 99.1 6.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dh5qzzvc",
   "source_url": "https://omlx.ai/benchmarks/performance/dh5qzzvc",
   "tg_tokens": null,
   "tps": 99.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dh5qzzvc-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: 0mpne123, 3njcyo0n, 3tpmnfvl, ij4lrk1x, mhfap76i, r30t9wq5, xe5jpdpp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3472.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 8k 3,472 101.0 6.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dh5qzzvc",
   "source_url": "https://omlx.ai/benchmarks/performance/dh5qzzvc",
   "tg_tokens": null,
   "tps": 101.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dh5qzzvc-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: 0mpne123, 3njcyo0n, 3tpmnfvl, ij4lrk1x, mhfap76i, r30t9wq5, xe5jpdpp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3314.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 3,314 95.7 6.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dh5qzzvc",
   "source_url": "https://omlx.ai/benchmarks/performance/dh5qzzvc",
   "tg_tokens": null,
   "tps": 95.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dh5qzzvc-ctx32768",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.2 GB at this context; duplicate traces of the same session: 0mpne123, 3njcyo0n, 3tpmnfvl, ij4lrk1x, mhfap76i, r30t9wq5, xe5jpdpp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2951.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 2,951 86.9 7.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dh5qzzvc",
   "source_url": "https://omlx.ai/benchmarks/performance/dh5qzzvc",
   "tg_tokens": null,
   "tps": 86.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dh5qzzvc-ctx65536",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.2 GB at this context; duplicate traces of the same session: 0mpne123, 3njcyo0n, 3tpmnfvl, ij4lrk1x, mhfap76i, r30t9wq5, xe5jpdpp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2399.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 64k 2,399 72.3 8.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dh5qzzvc",
   "source_url": "https://omlx.ai/benchmarks/performance/dh5qzzvc",
   "tg_tokens": null,
   "tps": 72.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dh5qzzvc-ctx131072",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.2 GB at this context; duplicate traces of the same session: 0mpne123, 3njcyo0n, 3tpmnfvl, ij4lrk1x, mhfap76i, r30t9wq5, xe5jpdpp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1954.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 128k 1,954 53.5 10.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dh5qzzvc",
   "source_url": "https://omlx.ai/benchmarks/performance/dh5qzzvc",
   "tg_tokens": null,
   "tps": 53.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dh5qzzvc-ctx199680",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 12.3 GB at this context; duplicate traces of the same session: 0mpne123, 3njcyo0n, 3tpmnfvl, ij4lrk1x, mhfap76i, r30t9wq5, xe5jpdpp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1545.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 195k 1,545 45.5 12.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dh5qzzvc",
   "source_url": "https://omlx.ai/benchmarks/performance/dh5qzzvc",
   "tg_tokens": null,
   "tps": 45.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-9s5dhase-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 9wcsw5dp; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 562.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 562.7 69.6 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9s5dhase",
   "source_url": "https://omlx.ai/benchmarks/performance/9s5dhase",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-9s5dhase-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 9wcsw5dp; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 590.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 590.7 67.7 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9s5dhase",
   "source_url": "https://omlx.ai/benchmarks/performance/9s5dhase",
   "tg_tokens": null,
   "tps": 67.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-w07ncera-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 363o6k0i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1881.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,881 93.7 21.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w07ncera",
   "source_url": "https://omlx.ai/benchmarks/performance/w07ncera",
   "tg_tokens": null,
   "tps": 93.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-w07ncera-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 363o6k0i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1042.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 1,042 75.0 23.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w07ncera",
   "source_url": "https://omlx.ai/benchmarks/performance/w07ncera",
   "tg_tokens": null,
   "tps": 75.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-fi1h0t5s-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: ih0qgstf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1752.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,752 96.0 21.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fi1h0t5s",
   "source_url": "https://omlx.ai/benchmarks/performance/fi1h0t5s",
   "tg_tokens": null,
   "tps": 96.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-fi1h0t5s-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: ih0qgstf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 993.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 993.0 72.2 23.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fi1h0t5s",
   "source_url": "https://omlx.ai/benchmarks/performance/fi1h0t5s",
   "tg_tokens": null,
   "tps": 72.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-h6gc8wkr-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; FP16; MLX; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 8h2efjfx, m04csys1, tmmyja3b; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1111.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-FP16-MTP-MLX on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 1,111 71.8 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h6gc8wkr",
   "source_url": "https://omlx.ai/benchmarks/performance/h6gc8wkr",
   "tg_tokens": null,
   "tps": 71.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-h6gc8wkr-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; FP16; MLX; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 8h2efjfx, m04csys1, tmmyja3b; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2114.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-FP16-MTP-MLX on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 2,114 84.6 23.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h6gc8wkr",
   "source_url": "https://omlx.ai/benchmarks/performance/h6gc8wkr",
   "tg_tokens": null,
   "tps": 84.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-h6gc8wkr-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; FP16; MLX; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: 8h2efjfx, m04csys1, tmmyja3b; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2090.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-FP16-MTP-MLX on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 8k 2,090 58.9 23.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h6gc8wkr",
   "source_url": "https://omlx.ai/benchmarks/performance/h6gc8wkr",
   "tg_tokens": null,
   "tps": 58.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-h6gc8wkr-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; FP16; MLX; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: 8h2efjfx, m04csys1, tmmyja3b; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1850.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-FP16-MTP-MLX on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 16k 1,850 123.9 24.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h6gc8wkr",
   "source_url": "https://omlx.ai/benchmarks/performance/h6gc8wkr",
   "tg_tokens": null,
   "tps": 123.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3sar38qs-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 1f6lk1a0, 9j7iy5wu, fggj5inw, ivd259rw, m1arwkiv, tsbmf97e, xlsgf6ig; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 706.9,
   "provenance": "community",
   "quant": "AWQ-5.0bpw",
   "quote": "Qwen3.8-27B-AWQ-5.0bpw on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 706.9 44.7 20.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3sar38qs",
   "source_url": "https://omlx.ai/benchmarks/performance/3sar38qs",
   "tg_tokens": null,
   "tps": 44.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3sar38qs-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: 1f6lk1a0, 9j7iy5wu, fggj5inw, ivd259rw, m1arwkiv, tsbmf97e, xlsgf6ig; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 870.8,
   "provenance": "community",
   "quant": "AWQ-5.0bpw",
   "quote": "Qwen3.8-27B-AWQ-5.0bpw on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 4k 870.8 59.4 21.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3sar38qs",
   "source_url": "https://omlx.ai/benchmarks/performance/3sar38qs",
   "tg_tokens": null,
   "tps": 59.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3sar38qs-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 1f6lk1a0, 9j7iy5wu, fggj5inw, ivd259rw, m1arwkiv, tsbmf97e, xlsgf6ig; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 819.1,
   "provenance": "community",
   "quant": "AWQ-5.0bpw",
   "quote": "Qwen3.8-27B-AWQ-5.0bpw on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 819.1 67.5 22.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3sar38qs",
   "source_url": "https://omlx.ai/benchmarks/performance/3sar38qs",
   "tg_tokens": null,
   "tps": 67.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3sar38qs-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: 1f6lk1a0, 9j7iy5wu, fggj5inw, ivd259rw, m1arwkiv, tsbmf97e, xlsgf6ig; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 793.8,
   "provenance": "community",
   "quant": "AWQ-5.0bpw",
   "quote": "Qwen3.8-27B-AWQ-5.0bpw on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 793.8 64.5 22.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3sar38qs",
   "source_url": "https://omlx.ai/benchmarks/performance/3sar38qs",
   "tg_tokens": null,
   "tps": 64.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3sar38qs-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: 1f6lk1a0, 9j7iy5wu, fggj5inw, ivd259rw, m1arwkiv, tsbmf97e, xlsgf6ig; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 746.4,
   "provenance": "community",
   "quant": "AWQ-5.0bpw",
   "quote": "Qwen3.8-27B-AWQ-5.0bpw on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 32k 746.4 120.6 23.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3sar38qs",
   "source_url": "https://omlx.ai/benchmarks/performance/3sar38qs",
   "tg_tokens": null,
   "tps": 120.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3sar38qs-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: 1f6lk1a0, 9j7iy5wu, fggj5inw, ivd259rw, m1arwkiv, tsbmf97e, xlsgf6ig; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 652.6,
   "provenance": "community",
   "quant": "AWQ-5.0bpw",
   "quote": "Qwen3.8-27B-AWQ-5.0bpw on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 64k 652.6 53.9 26.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3sar38qs",
   "source_url": "https://omlx.ai/benchmarks/performance/3sar38qs",
   "tg_tokens": null,
   "tps": 53.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3sar38qs-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 30.8 GB at this context; duplicate traces of the same session: 1f6lk1a0, 9j7iy5wu, fggj5inw, ivd259rw, m1arwkiv, tsbmf97e, xlsgf6ig; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 497.7,
   "provenance": "community",
   "quant": "AWQ-5.0bpw",
   "quote": "Qwen3.8-27B-AWQ-5.0bpw on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 128k 497.7 45.3 30.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3sar38qs",
   "source_url": "https://omlx.ai/benchmarks/performance/3sar38qs",
   "tg_tokens": null,
   "tps": 45.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3sar38qs-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 36.0 GB at this context; duplicate traces of the same session: 1f6lk1a0, 9j7iy5wu, fggj5inw, ivd259rw, m1arwkiv, tsbmf97e, xlsgf6ig; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 354.8,
   "provenance": "community",
   "quant": "AWQ-5.0bpw",
   "quote": "Qwen3.8-27B-AWQ-5.0bpw on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 195k 354.8 26.8 36.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3sar38qs",
   "source_url": "https://omlx.ai/benchmarks/performance/3sar38qs",
   "tg_tokens": null,
   "tps": 26.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yjxaf2io-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 114p0f5y, ki7qeuy2, pi3i2ef4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4483.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 4,483 147.3 19.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yjxaf2io",
   "source_url": "https://omlx.ai/benchmarks/performance/yjxaf2io",
   "tg_tokens": null,
   "tps": 147.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-yjxaf2io-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 114p0f5y, ki7qeuy2, pi3i2ef4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3789.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 64k 3,789 92.8 21.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yjxaf2io",
   "source_url": "https://omlx.ai/benchmarks/performance/yjxaf2io",
   "tg_tokens": null,
   "tps": 92.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-yjxaf2io-ctx131072",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: 114p0f5y, ki7qeuy2, pi3i2ef4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2450.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 128k 2,450 76.3 22.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yjxaf2io",
   "source_url": "https://omlx.ai/benchmarks/performance/yjxaf2io",
   "tg_tokens": null,
   "tps": 76.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-yjxaf2io-ctx199680",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: 114p0f5y, ki7qeuy2, pi3i2ef4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1770.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 195k 1,770 60.2 23.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yjxaf2io",
   "source_url": "https://omlx.ai/benchmarks/performance/yjxaf2io",
   "tg_tokens": null,
   "tps": 60.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-wwgatp8k-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.7 GB at this context; duplicate traces of the same session: 8cb4hk1k; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1146.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,146 96.7 100.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wwgatp8k",
   "source_url": "https://omlx.ai/benchmarks/performance/wwgatp8k",
   "tg_tokens": null,
   "tps": 96.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-wwgatp8k-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: 8cb4hk1k; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1412.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,412 71.3 101.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wwgatp8k",
   "source_url": "https://omlx.ai/benchmarks/performance/wwgatp8k",
   "tg_tokens": null,
   "tps": 71.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-oaatk0i0-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: 34kmu4gr, qap5fmkk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1188.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,188 89.1 70.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oaatk0i0",
   "source_url": "https://omlx.ai/benchmarks/performance/oaatk0i0",
   "tg_tokens": null,
   "tps": 89.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-oaatk0i0-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 34kmu4gr, qap5fmkk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1991.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,991 88.6 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oaatk0i0",
   "source_url": "https://omlx.ai/benchmarks/performance/oaatk0i0",
   "tg_tokens": null,
   "tps": 88.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-oaatk0i0-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.7 GB at this context; duplicate traces of the same session: 34kmu4gr, qap5fmkk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2660.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,660 78.1 73.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oaatk0i0",
   "source_url": "https://omlx.ai/benchmarks/performance/oaatk0i0",
   "tg_tokens": null,
   "tps": 78.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-o691rl4q-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: nfqtxq4j, nqga4g98; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 936.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 936.7 57.3 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o691rl4q",
   "source_url": "https://omlx.ai/benchmarks/performance/o691rl4q",
   "tg_tokens": null,
   "tps": 57.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-o691rl4q-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: nfqtxq4j, nqga4g98; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1428.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,428 78.6 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o691rl4q",
   "source_url": "https://omlx.ai/benchmarks/performance/o691rl4q",
   "tg_tokens": null,
   "tps": 78.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-o691rl4q-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.7 GB at this context; duplicate traces of the same session: nfqtxq4j, nqga4g98; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2937.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,937 57.7 73.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o691rl4q",
   "source_url": "https://omlx.ai/benchmarks/performance/o691rl4q",
   "tg_tokens": null,
   "tps": 57.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lfdji9ko-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "QAT OptiQ build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: zzahmo1t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 720.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 720.0 99.0 20.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lfdji9ko",
   "source_url": "https://omlx.ai/benchmarks/performance/lfdji9ko",
   "tg_tokens": null,
   "tps": 99.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lfdji9ko-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "QAT OptiQ build; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: zzahmo1t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3211.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 3,211 95.9 20.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lfdji9ko",
   "source_url": "https://omlx.ai/benchmarks/performance/lfdji9ko",
   "tg_tokens": null,
   "tps": 95.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-p96adxnf-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 2g3qj6o7, 3pnk333k, 6ptddl7r, a2tnvnkk, dbds450s, pv2672xg, sqa7dky9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4570.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 4,570 145.8 19.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p96adxnf",
   "source_url": "https://omlx.ai/benchmarks/performance/p96adxnf",
   "tg_tokens": null,
   "tps": 145.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-p96adxnf-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 2g3qj6o7, 3pnk333k, 6ptddl7r, a2tnvnkk, dbds450s, pv2672xg, sqa7dky9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4470.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 4,470 144.1 20.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p96adxnf",
   "source_url": "https://omlx.ai/benchmarks/performance/p96adxnf",
   "tg_tokens": null,
   "tps": 144.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-p96adxnf-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 2g3qj6o7, 3pnk333k, 6ptddl7r, a2tnvnkk, dbds450s, pv2672xg, sqa7dky9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5178.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 8k 5,178 133.9 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p96adxnf",
   "source_url": "https://omlx.ai/benchmarks/performance/p96adxnf",
   "tg_tokens": null,
   "tps": 133.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-p96adxnf-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: 2g3qj6o7, 3pnk333k, 6ptddl7r, a2tnvnkk, dbds450s, pv2672xg, sqa7dky9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4775.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 4,775 127.4 20.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p96adxnf",
   "source_url": "https://omlx.ai/benchmarks/performance/p96adxnf",
   "tg_tokens": null,
   "tps": 127.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-p96adxnf-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 2g3qj6o7, 3pnk333k, 6ptddl7r, a2tnvnkk, dbds450s, pv2672xg, sqa7dky9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4234.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 4,234 119.7 20.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p96adxnf",
   "source_url": "https://omlx.ai/benchmarks/performance/p96adxnf",
   "tg_tokens": null,
   "tps": 119.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-p96adxnf-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 2g3qj6o7, 3pnk333k, 6ptddl7r, a2tnvnkk, dbds450s, pv2672xg, sqa7dky9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3264.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 64k 3,264 99.7 21.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p96adxnf",
   "source_url": "https://omlx.ai/benchmarks/performance/p96adxnf",
   "tg_tokens": null,
   "tps": 99.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-p96adxnf-ctx131072",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: 2g3qj6o7, 3pnk333k, 6ptddl7r, a2tnvnkk, dbds450s, pv2672xg, sqa7dky9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2271.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 128k 2,271 78.0 22.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p96adxnf",
   "source_url": "https://omlx.ai/benchmarks/performance/p96adxnf",
   "tg_tokens": null,
   "tps": 78.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-p96adxnf-ctx199680",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 2g3qj6o7, 3pnk333k, 6ptddl7r, a2tnvnkk, dbds450s, pv2672xg, sqa7dky9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1694.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 195k 1,694 61.7 23.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p96adxnf",
   "source_url": "https://omlx.ai/benchmarks/performance/p96adxnf",
   "tg_tokens": null,
   "tps": 61.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-x2ycmmxo-ctx1024",
   "model": "Qwen2.5-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 4.8 GB at this context; duplicate traces of the same session: buok8u47, cw2ntkow, km07so9i; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1695.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-7B-Instruct-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 1,695 65.5 4.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x2ycmmxo",
   "source_url": "https://omlx.ai/benchmarks/performance/x2ycmmxo",
   "tg_tokens": null,
   "tps": 65.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-x2ycmmxo-ctx4096",
   "model": "Qwen2.5-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: buok8u47, cw2ntkow, km07so9i; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1896.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-7B-Instruct-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,896 62.2 4.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x2ycmmxo",
   "source_url": "https://omlx.ai/benchmarks/performance/x2ycmmxo",
   "tg_tokens": null,
   "tps": 62.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-x2ycmmxo-ctx8192",
   "model": "Qwen2.5-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 5.1 GB at this context; duplicate traces of the same session: buok8u47, cw2ntkow, km07so9i; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1891.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-7B-Instruct-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 1,891 57.5 5.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x2ycmmxo",
   "source_url": "https://omlx.ai/benchmarks/performance/x2ycmmxo",
   "tg_tokens": null,
   "tps": 57.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-x2ycmmxo-ctx16384",
   "model": "Qwen2.5-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 5.5 GB at this context; duplicate traces of the same session: buok8u47, cw2ntkow, km07so9i; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1713.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-7B-Instruct-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 1,713 51.8 5.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x2ycmmxo",
   "source_url": "https://omlx.ai/benchmarks/performance/x2ycmmxo",
   "tg_tokens": null,
   "tps": 51.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-2o81opeu-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 8cv595pw; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 129.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 129.6 15.3 16.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2o81opeu",
   "source_url": "https://omlx.ai/benchmarks/performance/2o81opeu",
   "tg_tokens": null,
   "tps": 15.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-2o81opeu-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 8cv595pw; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 130.8 14.8 17.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2o81opeu",
   "source_url": "https://omlx.ai/benchmarks/performance/2o81opeu",
   "tg_tokens": null,
   "tps": 14.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3vv7up8p-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: p0mug9fu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 425.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 425.3 30.7 17.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3vv7up8p",
   "source_url": "https://omlx.ai/benchmarks/performance/3vv7up8p",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3vv7up8p-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: p0mug9fu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 516.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 516.4 44.7 18.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3vv7up8p",
   "source_url": "https://omlx.ai/benchmarks/performance/3vv7up8p",
   "tg_tokens": null,
   "tps": 44.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-8vyks55r-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 18.0 GB; TTFT 10006 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 409.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k tokens 409.3 PP tok/s 16.0 TG tok/s 10006 TTFT (ms) 18.0 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8vyks55r",
   "source_url": "https://omlx.ai/benchmarks/performance/8vyks55r",
   "tg_tokens": null,
   "tps": 16.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-8ac7qjlj-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 18.0 GB; TTFT 9992 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 409.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k tokens 409.9 PP tok/s 15.9 TG tok/s 9992 TTFT (ms) 18.0 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8ac7qjlj",
   "source_url": "https://omlx.ai/benchmarks/performance/8ac7qjlj",
   "tg_tokens": null,
   "tps": 15.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-kneed243-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 18.0 GB; TTFT 10330 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 396.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k tokens 396.5 PP tok/s 17.0 TG tok/s 10330 TTFT (ms) 18.0 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kneed243",
   "source_url": "https://omlx.ai/benchmarks/performance/kneed243",
   "tg_tokens": null,
   "tps": 17.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-z81t4dzs-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: s83x9k00; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 471.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 471.9 35.4 32.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z81t4dzs",
   "source_url": "https://omlx.ai/benchmarks/performance/z81t4dzs",
   "tg_tokens": null,
   "tps": 35.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-z81t4dzs-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.7 GB at this context; duplicate traces of the same session: s83x9k00; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 567.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 567.5 30.2 33.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z81t4dzs",
   "source_url": "https://omlx.ai/benchmarks/performance/z81t4dzs",
   "tg_tokens": null,
   "tps": 30.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-cce6h5vg-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: qrikaea7, tsfgsf12; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 109.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 1k 109.9 13.9 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cce6h5vg",
   "source_url": "https://omlx.ai/benchmarks/performance/cce6h5vg",
   "tg_tokens": null,
   "tps": 13.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-cce6h5vg-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: qrikaea7, tsfgsf12; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 97.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 97.9 11.2 22.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cce6h5vg",
   "source_url": "https://omlx.ai/benchmarks/performance/cce6h5vg",
   "tg_tokens": null,
   "tps": 11.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-cce6h5vg-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: qrikaea7, tsfgsf12; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 95.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 8k 95.7 11.2 23.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cce6h5vg",
   "source_url": "https://omlx.ai/benchmarks/performance/cce6h5vg",
   "tg_tokens": null,
   "tps": 11.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-ip3ck9ay-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 9b1ke1an, qqmpom0w; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 81.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 1k 81.8 9.6 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ip3ck9ay",
   "source_url": "https://omlx.ai/benchmarks/performance/ip3ck9ay",
   "tg_tokens": null,
   "tps": 9.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-ip3ck9ay-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: 9b1ke1an, qqmpom0w; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 90.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 90.3 12.8 22.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ip3ck9ay",
   "source_url": "https://omlx.ai/benchmarks/performance/ip3ck9ay",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-ip3ck9ay-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 9b1ke1an, qqmpom0w; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 97.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 8k 97.8 11.0 23.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ip3ck9ay",
   "source_url": "https://omlx.ai/benchmarks/performance/ip3ck9ay",
   "tg_tokens": null,
   "tps": 11.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-tg935tn6-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: loj6fvcb; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 411.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 411.6 26.2 18.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tg935tn6",
   "source_url": "https://omlx.ai/benchmarks/performance/tg935tn6",
   "tg_tokens": null,
   "tps": 26.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-tg935tn6-ctx65536",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: loj6fvcb; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 308.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 64k 308.0 22.6 23.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tg935tn6",
   "source_url": "https://omlx.ai/benchmarks/performance/tg935tn6",
   "tg_tokens": null,
   "tps": 22.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-if3a9d5q-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 26.0 GB; TTFT 107961 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 303.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k tokens 303.5 PP tok/s 68.4 TG tok/s 107961 TTFT (ms) 26.0 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/if3a9d5q",
   "source_url": "https://omlx.ai/benchmarks/performance/if3a9d5q",
   "tg_tokens": null,
   "tps": 68.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ceffs445-ctx1024",
   "model": "Swift-Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.2 GB at this context; duplicate traces of the same session: s26pz4x6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 782.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 782.3 40.8 39.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ceffs445",
   "source_url": "https://omlx.ai/benchmarks/performance/ceffs445",
   "tg_tokens": null,
   "tps": 40.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ceffs445-ctx4096",
   "model": "Swift-Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 40.2 GB at this context; duplicate traces of the same session: s26pz4x6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 690.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 690.6 40.4 40.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ceffs445",
   "source_url": "https://omlx.ai/benchmarks/performance/ceffs445",
   "tg_tokens": null,
   "tps": 40.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-vrzij9ro-ctx1024",
   "model": "Swift-Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.2 GB at this context; duplicate traces of the same session: m3n21e3t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 779.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 779.6 37.4 39.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vrzij9ro",
   "source_url": "https://omlx.ai/benchmarks/performance/vrzij9ro",
   "tg_tokens": null,
   "tps": 37.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vrzij9ro-ctx4096",
   "model": "Swift-Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 40.2 GB at this context; duplicate traces of the same session: m3n21e3t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 674.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 674.6 35.8 40.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vrzij9ro",
   "source_url": "https://omlx.ai/benchmarks/performance/vrzij9ro",
   "tg_tokens": null,
   "tps": 35.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9ld3l8wj-ctx1024",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: d6n7fu5y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 851.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 851.6 52.4 21.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ld3l8wj",
   "source_url": "https://omlx.ai/benchmarks/performance/9ld3l8wj",
   "tg_tokens": null,
   "tps": 52.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9ld3l8wj-ctx4096",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: d6n7fu5y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 698.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 698.9 41.4 22.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ld3l8wj",
   "source_url": "https://omlx.ai/benchmarks/performance/9ld3l8wj",
   "tg_tokens": null,
   "tps": 41.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-yvwdikru-ctx1024",
   "model": "clef-flash",
   "notes": "generic trace name; oMLX community benchmark trace; peak memory 5.7 GB at this context; duplicate traces of the same session: 83glgfd4; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1039.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "clef-flash-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 1,039 31.9 5.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yvwdikru",
   "source_url": "https://omlx.ai/benchmarks/performance/yvwdikru",
   "tg_tokens": null,
   "tps": 31.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-yvwdikru-ctx4096",
   "model": "clef-flash",
   "notes": "generic trace name; oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: 83glgfd4; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1154.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "clef-flash-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,154 31.2 6.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yvwdikru",
   "source_url": "https://omlx.ai/benchmarks/performance/yvwdikru",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-s8tgmat1-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: j7lt0351; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 917.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 917.8 90.3 19.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s8tgmat1",
   "source_url": "https://omlx.ai/benchmarks/performance/s8tgmat1",
   "tg_tokens": null,
   "tps": 90.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-s8tgmat1-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: j7lt0351; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1045.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 1,045 89.0 20.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s8tgmat1",
   "source_url": "https://omlx.ai/benchmarks/performance/s8tgmat1",
   "tg_tokens": null,
   "tps": 89.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gn9ktpb6-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 3ekv32sx, jbmrqk39, wieiapg3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4214.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 4,214 147.5 19.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gn9ktpb6",
   "source_url": "https://omlx.ai/benchmarks/performance/gn9ktpb6",
   "tg_tokens": null,
   "tps": 147.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gn9ktpb6-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 3ekv32sx, jbmrqk39, wieiapg3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3971.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 64k 3,971 99.9 21.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gn9ktpb6",
   "source_url": "https://omlx.ai/benchmarks/performance/gn9ktpb6",
   "tg_tokens": null,
   "tps": 99.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-gn9ktpb6-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: 3ekv32sx, jbmrqk39, wieiapg3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2531.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 128k 2,531 76.3 22.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gn9ktpb6",
   "source_url": "https://omlx.ai/benchmarks/performance/gn9ktpb6",
   "tg_tokens": null,
   "tps": 76.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-gn9ktpb6-ctx199680",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: 3ekv32sx, jbmrqk39, wieiapg3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1859.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 195k 1,859 61.3 23.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gn9ktpb6",
   "source_url": "https://omlx.ai/benchmarks/performance/gn9ktpb6",
   "tg_tokens": null,
   "tps": 61.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ir4isf4n-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "bf16-pt; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: ym1c2ueb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2060.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-bf16-pt-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 1k 2,060 146.7 25.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ir4isf4n",
   "source_url": "https://omlx.ai/benchmarks/performance/ir4isf4n",
   "tg_tokens": null,
   "tps": 146.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ir4isf4n-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "bf16-pt; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: ym1c2ueb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2492.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-bf16-pt-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 4k 2,492 137.8 26.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ir4isf4n",
   "source_url": "https://omlx.ai/benchmarks/performance/ir4isf4n",
   "tg_tokens": null,
   "tps": 137.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-njhvagde-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 26.0 GB; TTFT 107456 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 304.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k tokens 304.9 PP tok/s 96.6 TG tok/s 107456 TTFT (ms) 26.0 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/njhvagde",
   "source_url": "https://omlx.ai/benchmarks/performance/njhvagde",
   "tg_tokens": null,
   "tps": 96.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gbefhb8x-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: 5ehdb6ef, bag1cuzb, dcgz5nlm, g7hwis9w, nq3qa9il, u6cp700s, yvlmnyc1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 734.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 734.2 36.3 33.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gbefhb8x",
   "source_url": "https://omlx.ai/benchmarks/performance/gbefhb8x",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gbefhb8x-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.0 GB at this context; duplicate traces of the same session: 5ehdb6ef, bag1cuzb, dcgz5nlm, g7hwis9w, nq3qa9il, u6cp700s, yvlmnyc1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 651.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 651.5 33.7 35.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gbefhb8x",
   "source_url": "https://omlx.ai/benchmarks/performance/gbefhb8x",
   "tg_tokens": null,
   "tps": 33.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gbefhb8x-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.6 GB at this context; duplicate traces of the same session: 5ehdb6ef, bag1cuzb, dcgz5nlm, g7hwis9w, nq3qa9il, u6cp700s, yvlmnyc1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 630.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 630.1 44.3 35.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gbefhb8x",
   "source_url": "https://omlx.ai/benchmarks/performance/gbefhb8x",
   "tg_tokens": null,
   "tps": 44.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gbefhb8x-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.8 GB at this context; duplicate traces of the same session: 5ehdb6ef, bag1cuzb, dcgz5nlm, g7hwis9w, nq3qa9il, u6cp700s, yvlmnyc1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 577.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 577.5 49.1 36.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gbefhb8x",
   "source_url": "https://omlx.ai/benchmarks/performance/gbefhb8x",
   "tg_tokens": null,
   "tps": 49.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gbefhb8x-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.0 GB at this context; duplicate traces of the same session: 5ehdb6ef, bag1cuzb, dcgz5nlm, g7hwis9w, nq3qa9il, u6cp700s, yvlmnyc1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 557.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 557.1 43.9 39.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gbefhb8x",
   "source_url": "https://omlx.ai/benchmarks/performance/gbefhb8x",
   "tg_tokens": null,
   "tps": 43.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gbefhb8x-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.6 GB at this context; duplicate traces of the same session: 5ehdb6ef, bag1cuzb, dcgz5nlm, g7hwis9w, nq3qa9il, u6cp700s, yvlmnyc1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 435.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 435.8 27.2 43.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gbefhb8x",
   "source_url": "https://omlx.ai/benchmarks/performance/gbefhb8x",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gbefhb8x-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 53.0 GB at this context; duplicate traces of the same session: 5ehdb6ef, bag1cuzb, dcgz5nlm, g7hwis9w, nq3qa9il, u6cp700s, yvlmnyc1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 358.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 358.8 30.9 53.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gbefhb8x",
   "source_url": "https://omlx.ai/benchmarks/performance/gbefhb8x",
   "tg_tokens": null,
   "tps": 30.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gbefhb8x-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 63.0 GB at this context; duplicate traces of the same session: 5ehdb6ef, bag1cuzb, dcgz5nlm, g7hwis9w, nq3qa9il, u6cp700s, yvlmnyc1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 301.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 195k 301.9 18.5 63.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gbefhb8x",
   "source_url": "https://omlx.ai/benchmarks/performance/gbefhb8x",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4ug5xe36-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: b0n86p5n, cza3stsu, fv27114c, ld7j3pv5, phmr7gtj, u6oe68q4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 473.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 473.4 50.3 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ug5xe36",
   "source_url": "https://omlx.ai/benchmarks/performance/4ug5xe36",
   "tg_tokens": null,
   "tps": 50.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4ug5xe36-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: b0n86p5n, cza3stsu, fv27114c, ld7j3pv5, phmr7gtj, u6oe68q4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1202.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,202 30.2 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ug5xe36",
   "source_url": "https://omlx.ai/benchmarks/performance/4ug5xe36",
   "tg_tokens": null,
   "tps": 30.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4ug5xe36-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: b0n86p5n, cza3stsu, fv27114c, ld7j3pv5, phmr7gtj, u6oe68q4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1729.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,729 61.0 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ug5xe36",
   "source_url": "https://omlx.ai/benchmarks/performance/4ug5xe36",
   "tg_tokens": null,
   "tps": 61.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4ug5xe36-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.7 GB at this context; duplicate traces of the same session: b0n86p5n, cza3stsu, fv27114c, ld7j3pv5, phmr7gtj, u6oe68q4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2014.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,014 34.7 73.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ug5xe36",
   "source_url": "https://omlx.ai/benchmarks/performance/4ug5xe36",
   "tg_tokens": null,
   "tps": 34.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4ug5xe36-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 75.3 GB at this context; duplicate traces of the same session: b0n86p5n, cza3stsu, fv27114c, ld7j3pv5, phmr7gtj, u6oe68q4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1996.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 1,996 64.1 75.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ug5xe36",
   "source_url": "https://omlx.ai/benchmarks/performance/4ug5xe36",
   "tg_tokens": null,
   "tps": 64.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4ug5xe36-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.6 GB at this context; duplicate traces of the same session: b0n86p5n, cza3stsu, fv27114c, ld7j3pv5, phmr7gtj, u6oe68q4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1851.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 1,851 70.1 80.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ug5xe36",
   "source_url": "https://omlx.ai/benchmarks/performance/4ug5xe36",
   "tg_tokens": null,
   "tps": 70.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4ug5xe36-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 83.4 GB at this context; duplicate traces of the same session: b0n86p5n, cza3stsu, fv27114c, ld7j3pv5, phmr7gtj, u6oe68q4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1860.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 1,860 58.4 83.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ug5xe36",
   "source_url": "https://omlx.ai/benchmarks/performance/4ug5xe36",
   "tg_tokens": null,
   "tps": 58.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-tbyuq828-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 21.9 GB; TTFT 71900 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 113.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 8k tokens 113.9 PP tok/s 16.5 TG tok/s 71900 TTFT (ms) 21.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tbyuq828",
   "source_url": "https://omlx.ai/benchmarks/performance/tbyuq828",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-pb1844l3-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 21.9 GB; TTFT 70630 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 116.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 8k tokens 116.0 PP tok/s 16.6 TG tok/s 70630 TTFT (ms) 21.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pb1844l3",
   "source_url": "https://omlx.ai/benchmarks/performance/pb1844l3",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fgvz6bt2-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "mlx-community; oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: lsspieel; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1303.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 1,303 107.8 35.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fgvz6bt2",
   "source_url": "https://omlx.ai/benchmarks/performance/fgvz6bt2",
   "tg_tokens": null,
   "tps": 107.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fgvz6bt2-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "mlx-community; oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: lsspieel; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3421.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 3,421 107.8 36.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fgvz6bt2",
   "source_url": "https://omlx.ai/benchmarks/performance/fgvz6bt2",
   "tg_tokens": null,
   "tps": 107.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-usv1e2yz-ctx1024",
   "model": "Qwen3.6-27B",
   "notes": "ThinkingCap build; mlx-community; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 7mu05goy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 915.7,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "mlx-community--ThinkingCap-Qwen3.6-27B-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 915.7 28.2 19.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/usv1e2yz",
   "source_url": "https://omlx.ai/benchmarks/performance/usv1e2yz",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-usv1e2yz-ctx4096",
   "model": "Qwen3.6-27B",
   "notes": "ThinkingCap build; mlx-community; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: 7mu05goy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1009.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "mlx-community--ThinkingCap-Qwen3.6-27B-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,009 28.0 20.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/usv1e2yz",
   "source_url": "https://omlx.ai/benchmarks/performance/usv1e2yz",
   "tg_tokens": null,
   "tps": 28.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-56e9r4pl-ctx8192",
   "model": "gpt-oss 20B",
   "notes": "mlx-community; oMLX community benchmark trace, single run at 8k context; peak memory 11.9 GB; TTFT 14344 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 571.1,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "mlx-community--gpt-oss-20b-MXFP4-Q8 on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 8k tokens 571.1 PP tok/s 57.6 TG tok/s 14344 TTFT (ms) 11.9 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/56e9r4pl",
   "source_url": "https://omlx.ai/benchmarks/performance/56e9r4pl",
   "tg_tokens": null,
   "tps": 57.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nuauwds2-ctx1024",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 155.8 GB at this context; duplicate traces of the same session: 0vtxfpan, a6o07h55, d66r0z55, i9lmtcrp, qko9fdz2, ssxzmnz1, zvvgdqyy; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 548.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 548.7 38.1 155.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nuauwds2",
   "source_url": "https://omlx.ai/benchmarks/performance/nuauwds2",
   "tg_tokens": null,
   "tps": 38.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-nuauwds2-ctx4096",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.3 GB at this context; duplicate traces of the same session: 0vtxfpan, a6o07h55, d66r0z55, i9lmtcrp, qko9fdz2, ssxzmnz1, zvvgdqyy; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 602.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 602.0 40.9 156.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nuauwds2",
   "source_url": "https://omlx.ai/benchmarks/performance/nuauwds2",
   "tg_tokens": null,
   "tps": 40.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-nuauwds2-ctx8192",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.3 GB at this context; duplicate traces of the same session: 0vtxfpan, a6o07h55, d66r0z55, i9lmtcrp, qko9fdz2, ssxzmnz1, zvvgdqyy; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 593.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 593.3 52.1 156.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nuauwds2",
   "source_url": "https://omlx.ai/benchmarks/performance/nuauwds2",
   "tg_tokens": null,
   "tps": 52.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-nuauwds2-ctx16384",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.4 GB at this context; duplicate traces of the same session: 0vtxfpan, a6o07h55, d66r0z55, i9lmtcrp, qko9fdz2, ssxzmnz1, zvvgdqyy; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 583.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 583.0 48.0 156.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nuauwds2",
   "source_url": "https://omlx.ai/benchmarks/performance/nuauwds2",
   "tg_tokens": null,
   "tps": 48.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-nuauwds2-ctx32768",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.5 GB at this context; duplicate traces of the same session: 0vtxfpan, a6o07h55, d66r0z55, i9lmtcrp, qko9fdz2, ssxzmnz1, zvvgdqyy; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 566.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 566.5 51.7 156.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nuauwds2",
   "source_url": "https://omlx.ai/benchmarks/performance/nuauwds2",
   "tg_tokens": null,
   "tps": 51.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nuauwds2-ctx65536",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.8 GB at this context; duplicate traces of the same session: 0vtxfpan, a6o07h55, d66r0z55, i9lmtcrp, qko9fdz2, ssxzmnz1, zvvgdqyy; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 538.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 538.3 32.7 156.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nuauwds2",
   "source_url": "https://omlx.ai/benchmarks/performance/nuauwds2",
   "tg_tokens": null,
   "tps": 32.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-nuauwds2-ctx131072",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 157.1 GB at this context; duplicate traces of the same session: 0vtxfpan, a6o07h55, d66r0z55, i9lmtcrp, qko9fdz2, ssxzmnz1, zvvgdqyy; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 489.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 489.7 36.5 157.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nuauwds2",
   "source_url": "https://omlx.ai/benchmarks/performance/nuauwds2",
   "tg_tokens": null,
   "tps": 36.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-nuauwds2-ctx199680",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 158.1 GB at this context; duplicate traces of the same session: 0vtxfpan, a6o07h55, d66r0z55, i9lmtcrp, qko9fdz2, ssxzmnz1, zvvgdqyy; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 447.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 195k 447.7 32.8 158.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nuauwds2",
   "source_url": "https://omlx.ai/benchmarks/performance/nuauwds2",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-vo40ijls-ctx1024",
   "model": "Huihui-Qwen3.8-27B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 195.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.8-27B-abliterated-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 195.2 40.1 21.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vo40ijls",
   "source_url": "https://omlx.ai/benchmarks/performance/vo40ijls",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-vo40ijls-ctx4096",
   "model": "Huihui-Qwen3.8-27B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 231.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.8-27B-abliterated-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 231.4 31.8 24.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vo40ijls",
   "source_url": "https://omlx.ai/benchmarks/performance/vo40ijls",
   "tg_tokens": null,
   "tps": 31.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-vo40ijls-ctx16384",
   "model": "Huihui-Qwen3.8-27B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 240.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.8-27B-abliterated-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 240.5 41.4 26.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vo40ijls",
   "source_url": "https://omlx.ai/benchmarks/performance/vo40ijls",
   "tg_tokens": null,
   "tps": 41.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fvs8lfzm-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: kaninv53, lxsvx5cs, p6iker4o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4640.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 4,640 147.8 19.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fvs8lfzm",
   "source_url": "https://omlx.ai/benchmarks/performance/fvs8lfzm",
   "tg_tokens": null,
   "tps": 147.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fvs8lfzm-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: kaninv53, lxsvx5cs, p6iker4o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3898.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 64k 3,898 100.2 21.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fvs8lfzm",
   "source_url": "https://omlx.ai/benchmarks/performance/fvs8lfzm",
   "tg_tokens": null,
   "tps": 100.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fvs8lfzm-ctx131072",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: kaninv53, lxsvx5cs, p6iker4o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2520.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 128k 2,520 76.7 22.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fvs8lfzm",
   "source_url": "https://omlx.ai/benchmarks/performance/fvs8lfzm",
   "tg_tokens": null,
   "tps": 76.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fvs8lfzm-ctx199680",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: kaninv53, lxsvx5cs, p6iker4o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1849.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 195k 1,849 60.8 23.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fvs8lfzm",
   "source_url": "https://omlx.ai/benchmarks/performance/fvs8lfzm",
   "tg_tokens": null,
   "tps": 60.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-0abuwloi-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP fp16 build; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 24c1uwue, w53iq5vu; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 155.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-fp16-MLX on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 1k 155.0 15.9 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0abuwloi",
   "source_url": "https://omlx.ai/benchmarks/performance/0abuwloi",
   "tg_tokens": null,
   "tps": 15.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-0abuwloi-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP fp16 build; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: 24c1uwue, w53iq5vu; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 129.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-fp16-MLX on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 129.6 14.7 22.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0abuwloi",
   "source_url": "https://omlx.ai/benchmarks/performance/0abuwloi",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-0abuwloi-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP fp16 build; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 24c1uwue, w53iq5vu; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 123.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-fp16-MLX on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 8k 123.4 11.7 23.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0abuwloi",
   "source_url": "https://omlx.ai/benchmarks/performance/0abuwloi",
   "tg_tokens": null,
   "tps": 11.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vgeg2yxu-ctx1024",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 4y5u6idx, bgo9xlub; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 915.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 915.1 32.1 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vgeg2yxu",
   "source_url": "https://omlx.ai/benchmarks/performance/vgeg2yxu",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vgeg2yxu-ctx4096",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 4y5u6idx, bgo9xlub; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 992.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 992.9 31.5 17.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vgeg2yxu",
   "source_url": "https://omlx.ai/benchmarks/performance/vgeg2yxu",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vgeg2yxu-ctx8192",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 17.9 GB at this context; duplicate traces of the same session: 4y5u6idx, bgo9xlub; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 949.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 949.8 31.0 17.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vgeg2yxu",
   "source_url": "https://omlx.ai/benchmarks/performance/vgeg2yxu",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5klftvor-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: 22d0i9b0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 311.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 311.9 42.7 34.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5klftvor",
   "source_url": "https://omlx.ai/benchmarks/performance/5klftvor",
   "tg_tokens": null,
   "tps": 42.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5klftvor-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 38.1 GB at this context; duplicate traces of the same session: 22d0i9b0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 333.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 333.7 44.4 38.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5klftvor",
   "source_url": "https://omlx.ai/benchmarks/performance/5klftvor",
   "tg_tokens": null,
   "tps": 44.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1ls8ew59-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: 67ryzlsn; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 519.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 519.6 50.3 70.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ls8ew59",
   "source_url": "https://omlx.ai/benchmarks/performance/1ls8ew59",
   "tg_tokens": null,
   "tps": 50.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1ls8ew59-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: 67ryzlsn; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 693.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 693.7 37.9 71.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ls8ew59",
   "source_url": "https://omlx.ai/benchmarks/performance/1ls8ew59",
   "tg_tokens": null,
   "tps": 37.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-quya1f9t-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 8zup2409, 92vznu0g, c52wxwlh, j5qasxmc, n6i4p6jt, o7hxd11v, u09xggl2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 963.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 963.2 85.4 70.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/quya1f9t",
   "source_url": "https://omlx.ai/benchmarks/performance/quya1f9t",
   "tg_tokens": null,
   "tps": 85.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-quya1f9t-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: 8zup2409, 92vznu0g, c52wxwlh, j5qasxmc, n6i4p6jt, o7hxd11v, u09xggl2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1158.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,158 70.8 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/quya1f9t",
   "source_url": "https://omlx.ai/benchmarks/performance/quya1f9t",
   "tg_tokens": null,
   "tps": 70.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-quya1f9t-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 8zup2409, 92vznu0g, c52wxwlh, j5qasxmc, n6i4p6jt, o7hxd11v, u09xggl2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1344.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,344 44.8 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/quya1f9t",
   "source_url": "https://omlx.ai/benchmarks/performance/quya1f9t",
   "tg_tokens": null,
   "tps": 44.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-quya1f9t-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.9 GB at this context; duplicate traces of the same session: 8zup2409, 92vznu0g, c52wxwlh, j5qasxmc, n6i4p6jt, o7hxd11v, u09xggl2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1366.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,366 55.6 72.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/quya1f9t",
   "source_url": "https://omlx.ai/benchmarks/performance/quya1f9t",
   "tg_tokens": null,
   "tps": 55.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-quya1f9t-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.7 GB at this context; duplicate traces of the same session: 8zup2409, 92vznu0g, c52wxwlh, j5qasxmc, n6i4p6jt, o7hxd11v, u09xggl2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1383.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,383 70.6 73.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/quya1f9t",
   "source_url": "https://omlx.ai/benchmarks/performance/quya1f9t",
   "tg_tokens": null,
   "tps": 70.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-quya1f9t-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 75.4 GB at this context; duplicate traces of the same session: 8zup2409, 92vznu0g, c52wxwlh, j5qasxmc, n6i4p6jt, o7hxd11v, u09xggl2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1344.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 1,344 53.9 75.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/quya1f9t",
   "source_url": "https://omlx.ai/benchmarks/performance/quya1f9t",
   "tg_tokens": null,
   "tps": 53.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-quya1f9t-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.7 GB at this context; duplicate traces of the same session: 8zup2409, 92vznu0g, c52wxwlh, j5qasxmc, n6i4p6jt, o7hxd11v, u09xggl2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1285.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 1,285 45.6 80.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/quya1f9t",
   "source_url": "https://omlx.ai/benchmarks/performance/quya1f9t",
   "tg_tokens": null,
   "tps": 45.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-quya1f9t-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 83.4 GB at this context; duplicate traces of the same session: 8zup2409, 92vznu0g, c52wxwlh, j5qasxmc, n6i4p6jt, o7hxd11v, u09xggl2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1230.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 1,230 48.2 83.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/quya1f9t",
   "source_url": "https://omlx.ai/benchmarks/performance/quya1f9t",
   "tg_tokens": null,
   "tps": 48.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xy6pmgba-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1004.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,004 60.5 70.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xy6pmgba",
   "source_url": "https://omlx.ai/benchmarks/performance/xy6pmgba",
   "tg_tokens": null,
   "tps": 60.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xy6pmgba-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1995.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,995 67.7 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xy6pmgba",
   "source_url": "https://omlx.ai/benchmarks/performance/xy6pmgba",
   "tg_tokens": null,
   "tps": 67.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ukrfmhzg-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 86.5 GB at this context; duplicate traces of the same session: hww4p0za, lkey813z, wsm0atu9, xfz4fe34; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 672.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 672.7 36.2 86.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ukrfmhzg",
   "source_url": "https://omlx.ai/benchmarks/performance/ukrfmhzg",
   "tg_tokens": null,
   "tps": 36.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ukrfmhzg-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 87.8 GB at this context; duplicate traces of the same session: hww4p0za, lkey813z, wsm0atu9, xfz4fe34; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1996.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 1,996 38.5 87.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ukrfmhzg",
   "source_url": "https://omlx.ai/benchmarks/performance/ukrfmhzg",
   "tg_tokens": null,
   "tps": 38.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ukrfmhzg-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 90.7 GB at this context; duplicate traces of the same session: hww4p0za, lkey813z, wsm0atu9, xfz4fe34; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2201.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 2,201 67.9 90.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ukrfmhzg",
   "source_url": "https://omlx.ai/benchmarks/performance/ukrfmhzg",
   "tg_tokens": null,
   "tps": 67.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ukrfmhzg-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 94.2 GB at this context; duplicate traces of the same session: hww4p0za, lkey813z, wsm0atu9, xfz4fe34; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2061.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 2,061 52.5 94.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ukrfmhzg",
   "source_url": "https://omlx.ai/benchmarks/performance/ukrfmhzg",
   "tg_tokens": null,
   "tps": 52.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ukrfmhzg-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 97.1 GB at this context; duplicate traces of the same session: hww4p0za, lkey813z, wsm0atu9, xfz4fe34; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2158.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 195k 2,158 59.0 97.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ukrfmhzg",
   "source_url": "https://omlx.ai/benchmarks/performance/ukrfmhzg",
   "tg_tokens": null,
   "tps": 59.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-on9a3lmx-ctx1024",
   "model": "Qwopus3.8-27B-Flash-V2",
   "notes": "oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: ive9pvmb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 315.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwopus3.8-27B-Flash-V2-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 315.9 38.8 32.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/on9a3lmx",
   "source_url": "https://omlx.ai/benchmarks/performance/on9a3lmx",
   "tg_tokens": null,
   "tps": 38.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-on9a3lmx-ctx4096",
   "model": "Qwopus3.8-27B-Flash-V2",
   "notes": "oMLX community benchmark trace; peak memory 33.7 GB at this context; duplicate traces of the same session: ive9pvmb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 331.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwopus3.8-27B-Flash-V2-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 331.1 40.1 33.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/on9a3lmx",
   "source_url": "https://omlx.ai/benchmarks/performance/on9a3lmx",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-8r37zhwe-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 25.1 GB; TTFT 72913 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 112.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 8k tokens 112.4 PP tok/s 20.6 TG tok/s 72913 TTFT (ms) 25.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8r37zhwe",
   "source_url": "https://omlx.ai/benchmarks/performance/8r37zhwe",
   "tg_tokens": null,
   "tps": 20.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-csuv1jo0-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: xpib92fc, yxfxssu0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1085.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,085 72.8 19.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/csuv1jo0",
   "source_url": "https://omlx.ai/benchmarks/performance/csuv1jo0",
   "tg_tokens": null,
   "tps": 72.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-csuv1jo0-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: xpib92fc, yxfxssu0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1332.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,332 63.1 20.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/csuv1jo0",
   "source_url": "https://omlx.ai/benchmarks/performance/csuv1jo0",
   "tg_tokens": null,
   "tps": 63.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-csuv1jo0-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: xpib92fc, yxfxssu0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1338.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,338 71.6 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/csuv1jo0",
   "source_url": "https://omlx.ai/benchmarks/performance/csuv1jo0",
   "tg_tokens": null,
   "tps": 71.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-v29onteh-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1015.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,015 72.7 70.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v29onteh",
   "source_url": "https://omlx.ai/benchmarks/performance/v29onteh",
   "tg_tokens": null,
   "tps": 72.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-v29onteh-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1989.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,989 70.8 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v29onteh",
   "source_url": "https://omlx.ai/benchmarks/performance/v29onteh",
   "tg_tokens": null,
   "tps": 70.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0hutqeif-ctx1024",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.3 GB at this context; duplicate traces of the same session: b58xs7m5; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 252.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "gemma-4-31B-it-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 1k 252.5 35.1 28.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0hutqeif",
   "source_url": "https://omlx.ai/benchmarks/performance/0hutqeif",
   "tg_tokens": null,
   "tps": 35.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0hutqeif-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: b58xs7m5; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 254.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "gemma-4-31B-it-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 4k 254.9 42.4 28.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0hutqeif",
   "source_url": "https://omlx.ai/benchmarks/performance/0hutqeif",
   "tg_tokens": null,
   "tps": 42.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-uebx4vhz-ctx1024",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 155.8 GB at this context; duplicate traces of the same session: wuoihf7o; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 442.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 442.0 33.8 155.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uebx4vhz",
   "source_url": "https://omlx.ai/benchmarks/performance/uebx4vhz",
   "tg_tokens": null,
   "tps": 33.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-uebx4vhz-ctx4096",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.3 GB at this context; duplicate traces of the same session: wuoihf7o; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 472.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 472.1 36.5 156.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uebx4vhz",
   "source_url": "https://omlx.ai/benchmarks/performance/uebx4vhz",
   "tg_tokens": null,
   "tps": 36.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-cq4tabyr-ctx1024",
   "model": "Huihui-Qwen3.8-27B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: hg8cczxj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 195.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.8-27B-abliterated-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 195.2 40.1 21.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cq4tabyr",
   "source_url": "https://omlx.ai/benchmarks/performance/cq4tabyr",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-cq4tabyr-ctx4096",
   "model": "Huihui-Qwen3.8-27B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: hg8cczxj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 231.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.8-27B-abliterated-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 231.4 31.8 24.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cq4tabyr",
   "source_url": "https://omlx.ai/benchmarks/performance/cq4tabyr",
   "tg_tokens": null,
   "tps": 31.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-cq4tabyr-ctx16384",
   "model": "Huihui-Qwen3.8-27B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.2 GB at this context; duplicate traces of the same session: hg8cczxj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 240.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.8-27B-abliterated-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 240.5 41.4 26.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cq4tabyr",
   "source_url": "https://omlx.ai/benchmarks/performance/cq4tabyr",
   "tg_tokens": null,
   "tps": 41.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-0ret3zc2-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 247stpwv, 46aulvm3, cy4lqa9y; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1341.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 1,341 89.4 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ret3zc2",
   "source_url": "https://omlx.ai/benchmarks/performance/0ret3zc2",
   "tg_tokens": null,
   "tps": 89.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-0ret3zc2-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 247stpwv, 46aulvm3, cy4lqa9y; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1502.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 1,502 86.0 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ret3zc2",
   "source_url": "https://omlx.ai/benchmarks/performance/0ret3zc2",
   "tg_tokens": null,
   "tps": 86.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-0ret3zc2-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 247stpwv, 46aulvm3, cy4lqa9y; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1391.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 8k 1,391 83.5 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ret3zc2",
   "source_url": "https://omlx.ai/benchmarks/performance/0ret3zc2",
   "tg_tokens": null,
   "tps": 83.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-0ret3zc2-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 247stpwv, 46aulvm3, cy4lqa9y; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1199.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 16k 1,199 73.3 22.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ret3zc2",
   "source_url": "https://omlx.ai/benchmarks/performance/0ret3zc2",
   "tg_tokens": null,
   "tps": 73.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-2x9y361e-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 4hu721yy, 5wdawfbm, bywxww3t; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1255.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 1,255 89.0 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2x9y361e",
   "source_url": "https://omlx.ai/benchmarks/performance/2x9y361e",
   "tg_tokens": null,
   "tps": 89.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-2x9y361e-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 4hu721yy, 5wdawfbm, bywxww3t; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1325.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 1,325 87.9 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2x9y361e",
   "source_url": "https://omlx.ai/benchmarks/performance/2x9y361e",
   "tg_tokens": null,
   "tps": 87.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-2x9y361e-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 4hu721yy, 5wdawfbm, bywxww3t; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1372.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 8k 1,372 81.0 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2x9y361e",
   "source_url": "https://omlx.ai/benchmarks/performance/2x9y361e",
   "tg_tokens": null,
   "tps": 81.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-2x9y361e-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 4hu721yy, 5wdawfbm, bywxww3t; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1202.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 16k 1,202 76.5 22.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2x9y361e",
   "source_url": "https://omlx.ai/benchmarks/performance/2x9y361e",
   "tg_tokens": null,
   "tps": 76.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-4pc9rm8k-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; FP16; MLX; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: k19ebr98, lhuvlmqo, wp32b25t; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1727.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-FP16-MTP-MLX on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 1,727 133.3 21.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4pc9rm8k",
   "source_url": "https://omlx.ai/benchmarks/performance/4pc9rm8k",
   "tg_tokens": null,
   "tps": 133.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-4pc9rm8k-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; FP16; MLX; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: k19ebr98, lhuvlmqo, wp32b25t; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2183.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-FP16-MTP-MLX on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 2,183 91.5 23.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4pc9rm8k",
   "source_url": "https://omlx.ai/benchmarks/performance/4pc9rm8k",
   "tg_tokens": null,
   "tps": 91.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-4pc9rm8k-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; FP16; MLX; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: k19ebr98, lhuvlmqo, wp32b25t; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2026.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-FP16-MTP-MLX on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 8k 2,026 89.4 23.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4pc9rm8k",
   "source_url": "https://omlx.ai/benchmarks/performance/4pc9rm8k",
   "tg_tokens": null,
   "tps": 89.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-4pc9rm8k-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; FP16; MLX; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: k19ebr98, lhuvlmqo, wp32b25t; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1813.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-FP16-MTP-MLX on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 16k 1,813 106.8 24.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4pc9rm8k",
   "source_url": "https://omlx.ai/benchmarks/performance/4pc9rm8k",
   "tg_tokens": null,
   "tps": 106.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bjxdp1h1-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 0gwfxktd, 1bguj3b7, 3rx7auu2, kempvj2z, kuzlwm8e, qq8mx7mm, vkwpexih; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 909.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 909.0 33.0 16.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bjxdp1h1",
   "source_url": "https://omlx.ai/benchmarks/performance/bjxdp1h1",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bjxdp1h1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 0gwfxktd, 1bguj3b7, 3rx7auu2, kempvj2z, kuzlwm8e, qq8mx7mm, vkwpexih; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1185.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,185 33.1 18.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bjxdp1h1",
   "source_url": "https://omlx.ai/benchmarks/performance/bjxdp1h1",
   "tg_tokens": null,
   "tps": 33.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bjxdp1h1-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: 0gwfxktd, 1bguj3b7, 3rx7auu2, kempvj2z, kuzlwm8e, qq8mx7mm, vkwpexih; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1321.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 8k 1,321 32.2 18.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bjxdp1h1",
   "source_url": "https://omlx.ai/benchmarks/performance/bjxdp1h1",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bjxdp1h1-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 0gwfxktd, 1bguj3b7, 3rx7auu2, kempvj2z, kuzlwm8e, qq8mx7mm, vkwpexih; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1298.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 1,298 31.1 19.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bjxdp1h1",
   "source_url": "https://omlx.ai/benchmarks/performance/bjxdp1h1",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bjxdp1h1-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 0gwfxktd, 1bguj3b7, 3rx7auu2, kempvj2z, kuzlwm8e, qq8mx7mm, vkwpexih; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1170.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 1,170 29.5 20.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bjxdp1h1",
   "source_url": "https://omlx.ai/benchmarks/performance/bjxdp1h1",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bjxdp1h1-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: 0gwfxktd, 1bguj3b7, 3rx7auu2, kempvj2z, kuzlwm8e, qq8mx7mm, vkwpexih; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1001.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 64k 1,001 25.9 22.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bjxdp1h1",
   "source_url": "https://omlx.ai/benchmarks/performance/bjxdp1h1",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bjxdp1h1-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 26.2 GB at this context; duplicate traces of the same session: 0gwfxktd, 1bguj3b7, 3rx7auu2, kempvj2z, kuzlwm8e, qq8mx7mm, vkwpexih; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 690.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 128k 690.1 20.4 26.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bjxdp1h1",
   "source_url": "https://omlx.ai/benchmarks/performance/bjxdp1h1",
   "tg_tokens": null,
   "tps": 20.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bjxdp1h1-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 30.4 GB at this context; duplicate traces of the same session: 0gwfxktd, 1bguj3b7, 3rx7auu2, kempvj2z, kuzlwm8e, qq8mx7mm, vkwpexih; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 517.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 195k 517.1 15.7 30.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bjxdp1h1",
   "source_url": "https://omlx.ai/benchmarks/performance/bjxdp1h1",
   "tg_tokens": null,
   "tps": 15.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-johwm6dj-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: xf53bnjs; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 474.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 474.8 34.9 21.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/johwm6dj",
   "source_url": "https://omlx.ai/benchmarks/performance/johwm6dj",
   "tg_tokens": null,
   "tps": 34.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-johwm6dj-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: xf53bnjs; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 544.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 544.9 26.2 22.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/johwm6dj",
   "source_url": "https://omlx.ai/benchmarks/performance/johwm6dj",
   "tg_tokens": null,
   "tps": 26.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2jy5ko23-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: uby981dd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 878.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 878.4 67.7 21.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2jy5ko23",
   "source_url": "https://omlx.ai/benchmarks/performance/2jy5ko23",
   "tg_tokens": null,
   "tps": 67.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2jy5ko23-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: uby981dd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 721.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 721.1 63.8 22.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2jy5ko23",
   "source_url": "https://omlx.ai/benchmarks/performance/2jy5ko23",
   "tg_tokens": null,
   "tps": 63.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-s9wois1a-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 9hweii91, x036889g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1257.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,257 52.9 19.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s9wois1a",
   "source_url": "https://omlx.ai/benchmarks/performance/s9wois1a",
   "tg_tokens": null,
   "tps": 52.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-s9wois1a-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 9hweii91, x036889g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1283.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 1,283 74.8 19.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s9wois1a",
   "source_url": "https://omlx.ai/benchmarks/performance/s9wois1a",
   "tg_tokens": null,
   "tps": 74.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-s9wois1a-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: 9hweii91, x036889g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1076.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 1,076 130.2 21.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s9wois1a",
   "source_url": "https://omlx.ai/benchmarks/performance/s9wois1a",
   "tg_tokens": null,
   "tps": 130.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q71cdree-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1004.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,004 60.5 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q71cdree",
   "source_url": "https://omlx.ai/benchmarks/performance/q71cdree",
   "tg_tokens": null,
   "tps": 60.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q71cdree-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1995.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,995 67.7 70.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q71cdree",
   "source_url": "https://omlx.ai/benchmarks/performance/q71cdree",
   "tg_tokens": null,
   "tps": 67.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jolsorrd-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: 6st9izb0, 6zvznppz, spf8ait0, vjbc51xd, w4jjgpsr, w9lbwhap, yloq61ta; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 930.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 930.1 70.7 70.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jolsorrd",
   "source_url": "https://omlx.ai/benchmarks/performance/jolsorrd",
   "tg_tokens": null,
   "tps": 70.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jolsorrd-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: 6st9izb0, 6zvznppz, spf8ait0, vjbc51xd, w4jjgpsr, w9lbwhap, yloq61ta; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1447.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,447 68.1 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jolsorrd",
   "source_url": "https://omlx.ai/benchmarks/performance/jolsorrd",
   "tg_tokens": null,
   "tps": 68.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jolsorrd-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 6st9izb0, 6zvznppz, spf8ait0, vjbc51xd, w4jjgpsr, w9lbwhap, yloq61ta; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1671.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,671 71.2 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jolsorrd",
   "source_url": "https://omlx.ai/benchmarks/performance/jolsorrd",
   "tg_tokens": null,
   "tps": 71.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jolsorrd-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.9 GB at this context; duplicate traces of the same session: 6st9izb0, 6zvznppz, spf8ait0, vjbc51xd, w4jjgpsr, w9lbwhap, yloq61ta; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1889.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,889 67.7 72.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jolsorrd",
   "source_url": "https://omlx.ai/benchmarks/performance/jolsorrd",
   "tg_tokens": null,
   "tps": 67.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jolsorrd-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.7 GB at this context; duplicate traces of the same session: 6st9izb0, 6zvznppz, spf8ait0, vjbc51xd, w4jjgpsr, w9lbwhap, yloq61ta; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1929.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,929 64.6 73.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jolsorrd",
   "source_url": "https://omlx.ai/benchmarks/performance/jolsorrd",
   "tg_tokens": null,
   "tps": 64.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jolsorrd-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 75.3 GB at this context; duplicate traces of the same session: 6st9izb0, 6zvznppz, spf8ait0, vjbc51xd, w4jjgpsr, w9lbwhap, yloq61ta; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1887.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 1,887 59.7 75.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jolsorrd",
   "source_url": "https://omlx.ai/benchmarks/performance/jolsorrd",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jolsorrd-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.1 GB at this context; duplicate traces of the same session: 6st9izb0, 6zvznppz, spf8ait0, vjbc51xd, w4jjgpsr, w9lbwhap, yloq61ta; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1683.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 1,683 51.8 80.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jolsorrd",
   "source_url": "https://omlx.ai/benchmarks/performance/jolsorrd",
   "tg_tokens": null,
   "tps": 51.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jolsorrd-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 83.4 GB at this context; duplicate traces of the same session: 6st9izb0, 6zvznppz, spf8ait0, vjbc51xd, w4jjgpsr, w9lbwhap, yloq61ta; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1360.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 1,360 46.7 83.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jolsorrd",
   "source_url": "https://omlx.ai/benchmarks/performance/jolsorrd",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jfhdkwjd-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 86.5 GB at this context; duplicate traces of the same session: p3gcr6py; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1102.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 1,102 80.4 86.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jfhdkwjd",
   "source_url": "https://omlx.ai/benchmarks/performance/jfhdkwjd",
   "tg_tokens": null,
   "tps": 80.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jfhdkwjd-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 86.7 GB at this context; duplicate traces of the same session: p3gcr6py; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1830.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 1,830 81.1 86.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jfhdkwjd",
   "source_url": "https://omlx.ai/benchmarks/performance/jfhdkwjd",
   "tg_tokens": null,
   "tps": 81.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jfhdkwjd-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 87.8 GB at this context; duplicate traces of the same session: p3gcr6py; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2439.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 2,439 80.2 87.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jfhdkwjd",
   "source_url": "https://omlx.ai/benchmarks/performance/jfhdkwjd",
   "tg_tokens": null,
   "tps": 80.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jfhdkwjd-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 88.9 GB at this context; duplicate traces of the same session: p3gcr6py; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2415.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 2,415 79.5 88.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jfhdkwjd",
   "source_url": "https://omlx.ai/benchmarks/performance/jfhdkwjd",
   "tg_tokens": null,
   "tps": 79.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jfhdkwjd-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 89.6 GB at this context; duplicate traces of the same session: p3gcr6py; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2342.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 2,342 66.2 89.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jfhdkwjd",
   "source_url": "https://omlx.ai/benchmarks/performance/jfhdkwjd",
   "tg_tokens": null,
   "tps": 66.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jfhdkwjd-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 90.7 GB at this context; duplicate traces of the same session: p3gcr6py; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2175.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 2,175 81.7 90.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jfhdkwjd",
   "source_url": "https://omlx.ai/benchmarks/performance/jfhdkwjd",
   "tg_tokens": null,
   "tps": 81.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jfhdkwjd-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 94.5 GB at this context; duplicate traces of the same session: p3gcr6py; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2097.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 2,097 64.2 94.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jfhdkwjd",
   "source_url": "https://omlx.ai/benchmarks/performance/jfhdkwjd",
   "tg_tokens": null,
   "tps": 64.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jfhdkwjd-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 97.1 GB at this context; duplicate traces of the same session: p3gcr6py; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2135.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 195k 2,135 58.2 97.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jfhdkwjd",
   "source_url": "https://omlx.ai/benchmarks/performance/jfhdkwjd",
   "tg_tokens": null,
   "tps": 58.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bqxrjory-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace, single run at 1k context; peak memory 70.6 GB; TTFT 826 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1240.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,240 PP tok/s 80.1 TG tok/s 826 TTFT (ms) 70.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bqxrjory",
   "source_url": "https://omlx.ai/benchmarks/performance/bqxrjory",
   "tg_tokens": null,
   "tps": 80.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7a578n3l-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1015.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,015 72.7 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7a578n3l",
   "source_url": "https://omlx.ai/benchmarks/performance/7a578n3l",
   "tg_tokens": null,
   "tps": 72.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7a578n3l-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1989.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,989 70.8 70.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7a578n3l",
   "source_url": "https://omlx.ai/benchmarks/performance/7a578n3l",
   "tg_tokens": null,
   "tps": 70.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ha4tb8h7-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "bottlecapai ThinkingCap DWQ; oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: vx1oj729; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 632.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "bottlecapai--ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 632.2 25.3 21.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ha4tb8h7",
   "source_url": "https://omlx.ai/benchmarks/performance/ha4tb8h7",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ha4tb8h7-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "bottlecapai ThinkingCap DWQ; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: vx1oj729; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 913.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "bottlecapai--ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 913.3 25.2 22.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ha4tb8h7",
   "source_url": "https://omlx.ai/benchmarks/performance/ha4tb8h7",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-p9ewls1b-ctx1024",
   "model": "Bio-Medical-Llama-3-8B",
   "notes": "oMLX community benchmark trace; peak memory 5.0 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 704.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Bio-Medical-Llama-3-8B on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 704.5 49.4 5.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p9ewls1b",
   "source_url": "https://omlx.ai/benchmarks/performance/p9ewls1b",
   "tg_tokens": null,
   "tps": 49.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-p9ewls1b-ctx4096",
   "model": "Bio-Medical-Llama-3-8B",
   "notes": "oMLX community benchmark trace; peak memory 5.4 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 492.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Bio-Medical-Llama-3-8B on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 492.1 61.5 5.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p9ewls1b",
   "source_url": "https://omlx.ai/benchmarks/performance/p9ewls1b",
   "tg_tokens": null,
   "tps": 61.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xraphwsy-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 405.1 GB at this context; duplicate traces of the same session: 0hv5e1z0, 0ludqfri, 4qfs9605, 743eano2, 8b56j0r6, p8gy461x, siiblqyu; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 337.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 337.7 20.0 405.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xraphwsy",
   "source_url": "https://omlx.ai/benchmarks/performance/xraphwsy",
   "tg_tokens": null,
   "tps": 20.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xraphwsy-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 405.1 GB at this context; duplicate traces of the same session: 0hv5e1z0, 0ludqfri, 4qfs9605, 743eano2, 8b56j0r6, p8gy461x, siiblqyu; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 668.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 668.2 22.9 405.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xraphwsy",
   "source_url": "https://omlx.ai/benchmarks/performance/xraphwsy",
   "tg_tokens": null,
   "tps": 22.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xraphwsy-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 405.1 GB at this context; duplicate traces of the same session: 0hv5e1z0, 0ludqfri, 4qfs9605, 743eano2, 8b56j0r6, p8gy461x, siiblqyu; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 717.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 8k 717.8 20.0 405.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xraphwsy",
   "source_url": "https://omlx.ai/benchmarks/performance/xraphwsy",
   "tg_tokens": null,
   "tps": 20.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xraphwsy-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 405.1 GB at this context; duplicate traces of the same session: 0hv5e1z0, 0ludqfri, 4qfs9605, 743eano2, 8b56j0r6, p8gy461x, siiblqyu; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 719.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 16k 719.9 20.7 405.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xraphwsy",
   "source_url": "https://omlx.ai/benchmarks/performance/xraphwsy",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xraphwsy-ctx32768",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 405.1 GB at this context; duplicate traces of the same session: 0hv5e1z0, 0ludqfri, 4qfs9605, 743eano2, 8b56j0r6, p8gy461x, siiblqyu; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 725.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 32k 725.6 20.8 405.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xraphwsy",
   "source_url": "https://omlx.ai/benchmarks/performance/xraphwsy",
   "tg_tokens": null,
   "tps": 20.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xraphwsy-ctx65536",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 405.2 GB at this context; duplicate traces of the same session: 0hv5e1z0, 0ludqfri, 4qfs9605, 743eano2, 8b56j0r6, p8gy461x, siiblqyu; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 721.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 64k 721.3 20.6 405.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xraphwsy",
   "source_url": "https://omlx.ai/benchmarks/performance/xraphwsy",
   "tg_tokens": null,
   "tps": 20.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xraphwsy-ctx131072",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 405.3 GB at this context; duplicate traces of the same session: 0hv5e1z0, 0ludqfri, 4qfs9605, 743eano2, 8b56j0r6, p8gy461x, siiblqyu; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 702.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 128k 702.0 20.5 405.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xraphwsy",
   "source_url": "https://omlx.ai/benchmarks/performance/xraphwsy",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xraphwsy-ctx199680",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 405.4 GB at this context; duplicate traces of the same session: 0hv5e1z0, 0ludqfri, 4qfs9605, 743eano2, 8b56j0r6, p8gy461x, siiblqyu; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 684.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 195k 684.2 20.9 405.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xraphwsy",
   "source_url": "https://omlx.ai/benchmarks/performance/xraphwsy",
   "tg_tokens": null,
   "tps": 20.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-l6pq0son-ctx4096",
   "model": "Huihui-Ornith-1.5-35B-A3B-abliterated",
   "notes": "oMLX community benchmark trace; peak memory 70.1 GB at this context; duplicate traces of the same session: k8eqb8y3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1717.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Huihui-Ornith-1.5-35B-A3B-abliterated on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 unknown; 4k 1,717 58.4 70.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l6pq0son",
   "source_url": "https://omlx.ai/benchmarks/performance/l6pq0son",
   "tg_tokens": null,
   "tps": 58.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-l6pq0son-ctx16384",
   "model": "Huihui-Ornith-1.5-35B-A3B-abliterated",
   "notes": "oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: k8eqb8y3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1459.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Huihui-Ornith-1.5-35B-A3B-abliterated on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 1,459 74.5 70.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l6pq0son",
   "source_url": "https://omlx.ai/benchmarks/performance/l6pq0son",
   "tg_tokens": null,
   "tps": 74.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-wg8sefia-ctx1024",
   "model": "II-Medical-8B-1706",
   "notes": "oMLX community benchmark trace; peak memory 8.9 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 511.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "II-Medical-8B-1706-mlx-8Bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 511.6 37.9 8.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wg8sefia",
   "source_url": "https://omlx.ai/benchmarks/performance/wg8sefia",
   "tg_tokens": null,
   "tps": 37.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-wg8sefia-ctx4096",
   "model": "II-Medical-8B-1706",
   "notes": "oMLX community benchmark trace; peak memory 9.3 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 390.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "II-Medical-8B-1706-mlx-8Bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 390.2 29.7 9.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wg8sefia",
   "source_url": "https://omlx.ai/benchmarks/performance/wg8sefia",
   "tg_tokens": null,
   "tps": 29.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2rrtc90a-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: huecq0z1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 560.8,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.5-122B-A10B-oQ4-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 560.8 76.3 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2rrtc90a",
   "source_url": "https://omlx.ai/benchmarks/performance/2rrtc90a",
   "tg_tokens": null,
   "tps": 76.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2rrtc90a-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.9 GB at this context; duplicate traces of the same session: huecq0z1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 649.4,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.5-122B-A10B-oQ4-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 649.4 69.7 73.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2rrtc90a",
   "source_url": "https://omlx.ai/benchmarks/performance/2rrtc90a",
   "tg_tokens": null,
   "tps": 69.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-1yqr00lg-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: kmizr6yh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 560.9,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.5-122B-A10B-oQ4-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 560.9 87.7 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1yqr00lg",
   "source_url": "https://omlx.ai/benchmarks/performance/1yqr00lg",
   "tg_tokens": null,
   "tps": 87.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-1yqr00lg-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.9 GB at this context; duplicate traces of the same session: kmizr6yh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 648.4,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.5-122B-A10B-oQ4-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 648.4 80.9 73.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1yqr00lg",
   "source_url": "https://omlx.ai/benchmarks/performance/1yqr00lg",
   "tg_tokens": null,
   "tps": 80.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-0xl0jtna-ctx199680",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 195k context; peak memory 26.8 GB; TTFT 598112 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 334.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 195k tokens 334.4 PP tok/s 26.0 TG tok/s 598112 TTFT (ms) 26.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0xl0jtna",
   "source_url": "https://omlx.ai/benchmarks/performance/0xl0jtna",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-dvk9ekn1-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: fcxh82wt; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 761.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 761.1 74.5 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dvk9ekn1",
   "source_url": "https://omlx.ai/benchmarks/performance/dvk9ekn1",
   "tg_tokens": null,
   "tps": 74.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-dvk9ekn1-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: fcxh82wt; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 954.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 954.2 75.9 20.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dvk9ekn1",
   "source_url": "https://omlx.ai/benchmarks/performance/dvk9ekn1",
   "tg_tokens": null,
   "tps": 75.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-p3wk4izj-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: pgd9ckrx, txuw48dz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 246.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 246.1 31.1 22.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p3wk4izj",
   "source_url": "https://omlx.ai/benchmarks/performance/p3wk4izj",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-p3wk4izj-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: pgd9ckrx, txuw48dz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 283.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 283.9 31.8 25.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p3wk4izj",
   "source_url": "https://omlx.ai/benchmarks/performance/p3wk4izj",
   "tg_tokens": null,
   "tps": 31.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-p3wk4izj-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: pgd9ckrx, txuw48dz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 272.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 272.1 23.7 26.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p3wk4izj",
   "source_url": "https://omlx.ai/benchmarks/performance/p3wk4izj",
   "tg_tokens": null,
   "tps": 23.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xyqmpzlw-ctx1024",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: qks9zsem, urcbz8n5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1130.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,130 32.0 17.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xyqmpzlw",
   "source_url": "https://omlx.ai/benchmarks/performance/xyqmpzlw",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xyqmpzlw-ctx4096",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: qks9zsem, urcbz8n5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1387.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,387 31.5 19.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xyqmpzlw",
   "source_url": "https://omlx.ai/benchmarks/performance/xyqmpzlw",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xyqmpzlw-ctx8192",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: qks9zsem, urcbz8n5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1433.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,433 30.8 19.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xyqmpzlw",
   "source_url": "https://omlx.ai/benchmarks/performance/xyqmpzlw",
   "tg_tokens": null,
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mfsgmu92-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 22.9 GB; TTFT 36473 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 898.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k tokens 898.4 PP tok/s 61.1 TG tok/s 36473 TTFT (ms) 22.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mfsgmu92",
   "source_url": "https://omlx.ai/benchmarks/performance/mfsgmu92",
   "tg_tokens": null,
   "tps": 61.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8j7nrvak-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 32.1 GB; TTFT 87801 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 746.4,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-27B-oQ6 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 64k tokens 746.4 PP tok/s 33.6 TG tok/s 87801 TTFT (ms) 32.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8j7nrvak",
   "source_url": "https://omlx.ai/benchmarks/performance/8j7nrvak",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2c0jxvqs-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 27.7 GB; TTFT 4307 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 950.9,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-27B-oQ6 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k tokens 950.9 PP tok/s 41.5 TG tok/s 4307 TTFT (ms) 27.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2c0jxvqs",
   "source_url": "https://omlx.ai/benchmarks/performance/2c0jxvqs",
   "tg_tokens": null,
   "tps": 41.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n0up6nxy-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 27.7 GB; TTFT 4441 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 922.2,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-27B-oQ6 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k tokens 922.2 PP tok/s 35.1 TG tok/s 4441 TTFT (ms) 27.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n0up6nxy",
   "source_url": "https://omlx.ai/benchmarks/performance/n0up6nxy",
   "tg_tokens": null,
   "tps": 35.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-h2paydnd-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.7 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 854.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 854.8 69.1 100.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h2paydnd",
   "source_url": "https://omlx.ai/benchmarks/performance/h2paydnd",
   "tg_tokens": null,
   "tps": 69.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-h2paydnd-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1086.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 1,086 85.3 101.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h2paydnd",
   "source_url": "https://omlx.ai/benchmarks/performance/h2paydnd",
   "tg_tokens": null,
   "tps": 85.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-3taldkhs-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.8 GB; TTFT 2095 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 488.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 488.7 PP tok/s 64.6 TG tok/s 2095 TTFT (ms) 70.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3taldkhs",
   "source_url": "https://omlx.ai/benchmarks/performance/3taldkhs",
   "tg_tokens": null,
   "tps": 64.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-24eoin2f-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: f1bq55en, zkaltt4l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1214.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,214 68.7 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/24eoin2f",
   "source_url": "https://omlx.ai/benchmarks/performance/24eoin2f",
   "tg_tokens": null,
   "tps": 68.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-24eoin2f-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: f1bq55en, zkaltt4l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2125.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,125 81.5 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/24eoin2f",
   "source_url": "https://omlx.ai/benchmarks/performance/24eoin2f",
   "tg_tokens": null,
   "tps": 81.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-24eoin2f-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: f1bq55en, zkaltt4l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1735.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,735 76.2 71.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/24eoin2f",
   "source_url": "https://omlx.ai/benchmarks/performance/24eoin2f",
   "tg_tokens": null,
   "tps": 76.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-99eso2ld-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.6 GB; TTFT 932 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1099.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,099 PP tok/s 80.8 TG tok/s 932 TTFT (ms) 70.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/99eso2ld",
   "source_url": "https://omlx.ai/benchmarks/performance/99eso2ld",
   "tg_tokens": null,
   "tps": 80.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dj0dto2p-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: ikefvi51; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 731.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 731.4 55.4 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dj0dto2p",
   "source_url": "https://omlx.ai/benchmarks/performance/dj0dto2p",
   "tg_tokens": null,
   "tps": 55.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dj0dto2p-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: ikefvi51; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1136.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,136 60.5 70.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dj0dto2p",
   "source_url": "https://omlx.ai/benchmarks/performance/dj0dto2p",
   "tg_tokens": null,
   "tps": 60.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hu1kpien-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 86.5 GB at this context; duplicate traces of the same session: 1qqg3mvm, 2xq9ujvp, f0117gb3, mckr66av, usdoyyct; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1102.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 1,102 80.4 86.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hu1kpien",
   "source_url": "https://omlx.ai/benchmarks/performance/hu1kpien",
   "tg_tokens": null,
   "tps": 80.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hu1kpien-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 86.7 GB at this context; duplicate traces of the same session: 1qqg3mvm, 2xq9ujvp, f0117gb3, mckr66av, usdoyyct; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1830.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 1,830 81.1 86.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hu1kpien",
   "source_url": "https://omlx.ai/benchmarks/performance/hu1kpien",
   "tg_tokens": null,
   "tps": 81.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hu1kpien-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 87.8 GB at this context; duplicate traces of the same session: 1qqg3mvm, 2xq9ujvp, f0117gb3, mckr66av, usdoyyct; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2439.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 2,439 80.2 87.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hu1kpien",
   "source_url": "https://omlx.ai/benchmarks/performance/hu1kpien",
   "tg_tokens": null,
   "tps": 80.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hu1kpien-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 88.9 GB at this context; duplicate traces of the same session: 1qqg3mvm, 2xq9ujvp, f0117gb3, mckr66av, usdoyyct; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2415.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 2,415 79.5 88.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hu1kpien",
   "source_url": "https://omlx.ai/benchmarks/performance/hu1kpien",
   "tg_tokens": null,
   "tps": 79.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hu1kpien-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 89.6 GB at this context; duplicate traces of the same session: 1qqg3mvm, 2xq9ujvp, f0117gb3, mckr66av, usdoyyct; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2342.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 2,342 66.2 89.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hu1kpien",
   "source_url": "https://omlx.ai/benchmarks/performance/hu1kpien",
   "tg_tokens": null,
   "tps": 66.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hu1kpien-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 90.7 GB at this context; duplicate traces of the same session: 1qqg3mvm, 2xq9ujvp, f0117gb3, mckr66av, usdoyyct; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2175.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 2,175 81.7 90.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hu1kpien",
   "source_url": "https://omlx.ai/benchmarks/performance/hu1kpien",
   "tg_tokens": null,
   "tps": 81.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hu1kpien-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 94.5 GB at this context; duplicate traces of the same session: 1qqg3mvm, 2xq9ujvp, f0117gb3, mckr66av, usdoyyct; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2097.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 2,097 64.2 94.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hu1kpien",
   "source_url": "https://omlx.ai/benchmarks/performance/hu1kpien",
   "tg_tokens": null,
   "tps": 64.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hu1kpien-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 97.1 GB at this context; duplicate traces of the same session: 1qqg3mvm, 2xq9ujvp, f0117gb3, mckr66av, usdoyyct; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2135.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 195k 2,135 58.2 97.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hu1kpien",
   "source_url": "https://omlx.ai/benchmarks/performance/hu1kpien",
   "tg_tokens": null,
   "tps": 58.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-85n0xfqd-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: 3o5xwlyp, 5hm7rrp1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1091.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,091 68.1 19.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/85n0xfqd",
   "source_url": "https://omlx.ai/benchmarks/performance/85n0xfqd",
   "tg_tokens": null,
   "tps": 68.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-85n0xfqd-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 3o5xwlyp, 5hm7rrp1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1320.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,320 55.7 20.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/85n0xfqd",
   "source_url": "https://omlx.ai/benchmarks/performance/85n0xfqd",
   "tg_tokens": null,
   "tps": 55.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-85n0xfqd-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 3o5xwlyp, 5hm7rrp1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1335.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,335 61.7 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/85n0xfqd",
   "source_url": "https://omlx.ai/benchmarks/performance/85n0xfqd",
   "tg_tokens": null,
   "tps": 61.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-v2vltbci-ctx1024",
   "model": "Bio-Medical-Llama-3-8B",
   "notes": "oMLX community benchmark trace; peak memory 5.0 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 704.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Bio-Medical-Llama-3-8B on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 704.5 49.4 5.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v2vltbci",
   "source_url": "https://omlx.ai/benchmarks/performance/v2vltbci",
   "tg_tokens": null,
   "tps": 49.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-v2vltbci-ctx4096",
   "model": "Bio-Medical-Llama-3-8B",
   "notes": "oMLX community benchmark trace; peak memory 5.4 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 492.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Bio-Medical-Llama-3-8B on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 492.1 61.5 5.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v2vltbci",
   "source_url": "https://omlx.ai/benchmarks/performance/v2vltbci",
   "tg_tokens": null,
   "tps": 61.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ilkrcwg3-ctx1024",
   "model": "II-Medical-8B-1706",
   "notes": "oMLX community benchmark trace; peak memory 8.9 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 511.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "II-Medical-8B-1706-mlx-8Bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 511.6 37.9 8.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ilkrcwg3",
   "source_url": "https://omlx.ai/benchmarks/performance/ilkrcwg3",
   "tg_tokens": null,
   "tps": 37.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ilkrcwg3-ctx4096",
   "model": "II-Medical-8B-1706",
   "notes": "oMLX community benchmark trace; peak memory 9.3 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 390.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "II-Medical-8B-1706-mlx-8Bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 390.2 29.7 9.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ilkrcwg3",
   "source_url": "https://omlx.ai/benchmarks/performance/ilkrcwg3",
   "tg_tokens": null,
   "tps": 29.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3o9typh2-ctx1024",
   "model": "Jev-Style-Qwen3.5-2B-Decision-v2",
   "notes": "oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: 1nz84sz4, sla589l5, vc2f8177; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2691.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Jev-Style-Qwen3.5-2B-Decision-v2-MLX-bf16 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 bf16; 1k 2,691 35.7 6.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3o9typh2",
   "source_url": "https://omlx.ai/benchmarks/performance/3o9typh2",
   "tg_tokens": null,
   "tps": 35.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3o9typh2-ctx4096",
   "model": "Jev-Style-Qwen3.5-2B-Decision-v2",
   "notes": "oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: 1nz84sz4, sla589l5, vc2f8177; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3974.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Jev-Style-Qwen3.5-2B-Decision-v2-MLX-bf16 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 bf16; 4k 3,974 35.3 6.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3o9typh2",
   "source_url": "https://omlx.ai/benchmarks/performance/3o9typh2",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3o9typh2-ctx8192",
   "model": "Jev-Style-Qwen3.5-2B-Decision-v2",
   "notes": "oMLX community benchmark trace; peak memory 7.5 GB at this context; duplicate traces of the same session: 1nz84sz4, sla589l5, vc2f8177; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3891.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Jev-Style-Qwen3.5-2B-Decision-v2-MLX-bf16 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 bf16; 8k 3,891 35.3 7.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3o9typh2",
   "source_url": "https://omlx.ai/benchmarks/performance/3o9typh2",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3o9typh2-ctx16384",
   "model": "Jev-Style-Qwen3.5-2B-Decision-v2",
   "notes": "oMLX community benchmark trace; peak memory 7.2 GB at this context; duplicate traces of the same session: 1nz84sz4, sla589l5, vc2f8177; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3650.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Jev-Style-Qwen3.5-2B-Decision-v2-MLX-bf16 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 bf16; 16k 3,650 35.0 7.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3o9typh2",
   "source_url": "https://omlx.ai/benchmarks/performance/3o9typh2",
   "tg_tokens": null,
   "tps": 35.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3mq1bk19-ctx4096",
   "model": "Jev-Style-Qwen3.5-2B-Decision-v2",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 7.0 GB; TTFT 1286 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3185.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Jev-Style-Qwen3.5-2B-Decision-v2-MLX-bf16 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 bf16; 4k tokens 3,185 PP tok/s 35.4 TG tok/s 1286 TTFT (ms) 7.0 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3mq1bk19",
   "source_url": "https://omlx.ai/benchmarks/performance/3mq1bk19",
   "tg_tokens": null,
   "tps": 35.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5kln0d9l-ctx131072",
   "model": "Laguna XS 2.1",
   "notes": "oMLX community benchmark trace; peak memory 30.9 GB at this context; duplicate traces of the same session: 3b4wt2ou; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1367.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Laguna-XS-2.1-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 128k 1,367 54.5 30.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5kln0d9l",
   "source_url": "https://omlx.ai/benchmarks/performance/5kln0d9l",
   "tg_tokens": null,
   "tps": 54.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5kln0d9l-ctx199680",
   "model": "Laguna XS 2.1",
   "notes": "oMLX community benchmark trace; peak memory 33.5 GB at this context; duplicate traces of the same session: 3b4wt2ou; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1059.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Laguna-XS-2.1-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 195k 1,059 43.0 33.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5kln0d9l",
   "source_url": "https://omlx.ai/benchmarks/performance/5kln0d9l",
   "tg_tokens": null,
   "tps": 43.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bqxcupz0-ctx131072",
   "model": "Laguna XS 2.1",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 30.9 GB; TTFT 96594 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1357.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Laguna-XS-2.1-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 128k tokens 1,357 PP tok/s 54.0 TG tok/s 96594 TTFT (ms) 30.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bqxcupz0",
   "source_url": "https://omlx.ai/benchmarks/performance/bqxcupz0",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-2flhi5ba-ctx1024",
   "model": "Ministral-3-8B-Instruct-2512",
   "notes": "oMLX community benchmark trace; peak memory 5.3 GB at this context; duplicate traces of the same session: k1is6bhf; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 188.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ministral-3-8B-Instruct-2512-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 188.7 33.1 5.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2flhi5ba",
   "source_url": "https://omlx.ai/benchmarks/performance/2flhi5ba",
   "tg_tokens": null,
   "tps": 33.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-2flhi5ba-ctx4096",
   "model": "Ministral-3-8B-Instruct-2512",
   "notes": "oMLX community benchmark trace; peak memory 5.7 GB at this context; duplicate traces of the same session: k1is6bhf; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 180.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ministral-3-8B-Instruct-2512-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 180.9 29.0 5.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2flhi5ba",
   "source_url": "https://omlx.ai/benchmarks/performance/2flhi5ba",
   "tg_tokens": null,
   "tps": 29.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1udrz9d1-ctx1024",
   "model": "NeoHorse-Jev-4B-TmuxPocket",
   "notes": "oMLX community benchmark trace; peak memory 3.7 GB at this context; duplicate traces of the same session: c6845lxf, rzazp6q0, xospub4j; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1636.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "NeoHorse-Jev-4B-TmuxPocket-MLX-oQ4 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 1,636 153.9 3.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1udrz9d1",
   "source_url": "https://omlx.ai/benchmarks/performance/1udrz9d1",
   "tg_tokens": null,
   "tps": 153.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1udrz9d1-ctx4096",
   "model": "NeoHorse-Jev-4B-TmuxPocket",
   "notes": "oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: c6845lxf, rzazp6q0, xospub4j; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2216.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "NeoHorse-Jev-4B-TmuxPocket-MLX-oQ4 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 2,216 150.3 4.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1udrz9d1",
   "source_url": "https://omlx.ai/benchmarks/performance/1udrz9d1",
   "tg_tokens": null,
   "tps": 150.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1udrz9d1-ctx8192",
   "model": "NeoHorse-Jev-4B-TmuxPocket",
   "notes": "oMLX community benchmark trace; peak memory 5.1 GB at this context; duplicate traces of the same session: c6845lxf, rzazp6q0, xospub4j; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2137.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "NeoHorse-Jev-4B-TmuxPocket-MLX-oQ4 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 2,137 146.9 5.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1udrz9d1",
   "source_url": "https://omlx.ai/benchmarks/performance/1udrz9d1",
   "tg_tokens": null,
   "tps": 146.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1udrz9d1-ctx16384",
   "model": "NeoHorse-Jev-4B-TmuxPocket",
   "notes": "oMLX community benchmark trace; peak memory 5.5 GB at this context; duplicate traces of the same session: c6845lxf, rzazp6q0, xospub4j; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1979.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "NeoHorse-Jev-4B-TmuxPocket-MLX-oQ4 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 16k 1,979 135.1 5.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1udrz9d1",
   "source_url": "https://omlx.ai/benchmarks/performance/1udrz9d1",
   "tg_tokens": null,
   "tps": 135.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-pn29a2d6-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: micql9zm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 558.5,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.5-122B-A10B-oQ4-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 558.5 84.4 70.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pn29a2d6",
   "source_url": "https://omlx.ai/benchmarks/performance/pn29a2d6",
   "tg_tokens": null,
   "tps": 84.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-pn29a2d6-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.9 GB at this context; duplicate traces of the same session: micql9zm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 645.4,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.5-122B-A10B-oQ4-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 645.4 74.5 73.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pn29a2d6",
   "source_url": "https://omlx.ai/benchmarks/performance/pn29a2d6",
   "tg_tokens": null,
   "tps": 74.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-12brxmhm-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.4 GB at this context; duplicate traces of the same session: 20a9f2w9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1700.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 1,700 94.5 35.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/12brxmhm",
   "source_url": "https://omlx.ai/benchmarks/performance/12brxmhm",
   "tg_tokens": null,
   "tps": 94.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-12brxmhm-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: 20a9f2w9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1974.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 1,974 93.1 36.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/12brxmhm",
   "source_url": "https://omlx.ai/benchmarks/performance/12brxmhm",
   "tg_tokens": null,
   "tps": 93.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-r2xvzltx-ctx1024",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: h2c7gma1, h2ijuafx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1651.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,651 100.1 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r2xvzltx",
   "source_url": "https://omlx.ai/benchmarks/performance/r2xvzltx",
   "tg_tokens": null,
   "tps": 100.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-r2xvzltx-ctx4096",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: h2c7gma1, h2ijuafx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1955.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,955 100.3 21.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r2xvzltx",
   "source_url": "https://omlx.ai/benchmarks/performance/r2xvzltx",
   "tg_tokens": null,
   "tps": 100.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-r2xvzltx-ctx16384",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: h2c7gma1, h2ijuafx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1506.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,506 91.7 22.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r2xvzltx",
   "source_url": "https://omlx.ai/benchmarks/performance/r2xvzltx",
   "tg_tokens": null,
   "tps": 91.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-mztmgicz-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: szz61h5l; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2523.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 2,523 117.1 21.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mztmgicz",
   "source_url": "https://omlx.ai/benchmarks/performance/mztmgicz",
   "tg_tokens": null,
   "tps": 117.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-mztmgicz-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: szz61h5l; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3308.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 3,308 122.9 22.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mztmgicz",
   "source_url": "https://omlx.ai/benchmarks/performance/mztmgicz",
   "tg_tokens": null,
   "tps": 122.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-shspbob4-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 381 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2685.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k tokens 2,685 PP tok/s 105.6 TG tok/s 381 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/shspbob4",
   "source_url": "https://omlx.ai/benchmarks/performance/shspbob4",
   "tg_tokens": null,
   "tps": 105.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-9qq5ic6m-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 382 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2682.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k tokens 2,682 PP tok/s 102.8 TG tok/s 382 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9qq5ic6m",
   "source_url": "https://omlx.ai/benchmarks/performance/9qq5ic6m",
   "tg_tokens": null,
   "tps": 102.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5euc6a6v-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.3 GB at this context; duplicate traces of the same session: xpcf4bbf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 259.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 259.4 18.4 28.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5euc6a6v",
   "source_url": "https://omlx.ai/benchmarks/performance/5euc6a6v",
   "tg_tokens": null,
   "tps": 18.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5euc6a6v-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: xpcf4bbf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 265.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 265.5 18.2 29.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5euc6a6v",
   "source_url": "https://omlx.ai/benchmarks/performance/5euc6a6v",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-favthauf-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: deutlydo; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 130.1 15.3 16.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/favthauf",
   "source_url": "https://omlx.ai/benchmarks/performance/favthauf",
   "tg_tokens": null,
   "tps": 15.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-favthauf-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: deutlydo; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 133.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 133.4 15.1 17.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/favthauf",
   "source_url": "https://omlx.ai/benchmarks/performance/favthauf",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yo7y9q1i-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 23.0 GB; TTFT 36620 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 894.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k tokens 894.8 PP tok/s 52.4 TG tok/s 36620 TTFT (ms) 23.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yo7y9q1i",
   "source_url": "https://omlx.ai/benchmarks/performance/yo7y9q1i",
   "tg_tokens": null,
   "tps": 52.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-s6uivx5v-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 21.2 GB; TTFT 4156 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 985.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 985.5 PP tok/s 58.6 TG tok/s 4156 TTFT (ms) 21.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s6uivx5v",
   "source_url": "https://omlx.ai/benchmarks/performance/s6uivx5v",
   "tg_tokens": null,
   "tps": 58.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-yyud10yq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.2 GB at this context; duplicate traces of the same session: ptu27h58, wefx962m, x5shse6k; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 210.8,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 210.8 32.0 15.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yyud10yq",
   "source_url": "https://omlx.ai/benchmarks/performance/yyud10yq",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-yyud10yq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: ptu27h58, wefx962m, x5shse6k; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 214.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 214.0 30.9 18.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yyud10yq",
   "source_url": "https://omlx.ai/benchmarks/performance/yyud10yq",
   "tg_tokens": null,
   "tps": 30.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-yyud10yq-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: ptu27h58, wefx962m, x5shse6k; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 208.2,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 8k 208.2 29.8 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yyud10yq",
   "source_url": "https://omlx.ai/benchmarks/performance/yyud10yq",
   "tg_tokens": null,
   "tps": 29.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-yyud10yq-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: ptu27h58, wefx962m, x5shse6k; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 197.3,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 16k 197.3 28.5 20.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yyud10yq",
   "source_url": "https://omlx.ai/benchmarks/performance/yyud10yq",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-yyud10yq-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: ptu27h58, wefx962m, x5shse6k; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 178.4,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 32k 178.4 26.1 22.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yyud10yq",
   "source_url": "https://omlx.ai/benchmarks/performance/yyud10yq",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-yyud10yq-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: ptu27h58, wefx962m, x5shse6k; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 149.9,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 64k 149.9 22.5 23.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yyud10yq",
   "source_url": "https://omlx.ai/benchmarks/performance/yyud10yq",
   "tg_tokens": null,
   "tps": 22.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-yyud10yq-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: ptu27h58, wefx962m, x5shse6k; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 113.5,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 128k 113.5 17.8 27.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yyud10yq",
   "source_url": "https://omlx.ai/benchmarks/performance/yyud10yq",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-yyud10yq-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: ptu27h58, wefx962m, x5shse6k; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 89.8,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 195k 89.8 14.5 32.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yyud10yq",
   "source_url": "https://omlx.ai/benchmarks/performance/yyud10yq",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-vilg3pa2-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.7 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 854.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 854.8 69.1 100.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vilg3pa2",
   "source_url": "https://omlx.ai/benchmarks/performance/vilg3pa2",
   "tg_tokens": null,
   "tps": 69.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-vilg3pa2-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1086.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 1,086 85.3 101.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vilg3pa2",
   "source_url": "https://omlx.ai/benchmarks/performance/vilg3pa2",
   "tg_tokens": null,
   "tps": 85.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-rg6drkep-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5492.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 5,492 187.2 22.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rg6drkep",
   "source_url": "https://omlx.ai/benchmarks/performance/rg6drkep",
   "tg_tokens": null,
   "tps": 187.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-rg6drkep-ctx16384",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5450.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 5,450 204.6 22.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rg6drkep",
   "source_url": "https://omlx.ai/benchmarks/performance/rg6drkep",
   "tg_tokens": null,
   "tps": 204.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-72wca8t6-ctx1024",
   "model": "Gemma-4-31B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 0q43hte2, i8f0y1fv; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 91.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31b-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 1k 91.0 10.5 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/72wca8t6",
   "source_url": "https://omlx.ai/benchmarks/performance/72wca8t6",
   "tg_tokens": null,
   "tps": 10.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-72wca8t6-ctx4096",
   "model": "Gemma-4-31B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 0q43hte2, i8f0y1fv; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 81.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31b-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 81.3 9.6 18.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/72wca8t6",
   "source_url": "https://omlx.ai/benchmarks/performance/72wca8t6",
   "tg_tokens": null,
   "tps": 9.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-72wca8t6-ctx8192",
   "model": "Gemma-4-31B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 0q43hte2, i8f0y1fv; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 78.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31b-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 8k 78.1 9.0 19.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/72wca8t6",
   "source_url": "https://omlx.ai/benchmarks/performance/72wca8t6",
   "tg_tokens": null,
   "tps": 9.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-14wbujwi-ctx1024",
   "model": "gemma-4-31B-it-qat",
   "notes": "mlx-community OptiQ; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: s2pmouei, wetvt7ma, xknc2kl0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 694.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--gemma-4-31B-it-qat-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 694.8 21.8 22.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/14wbujwi",
   "source_url": "https://omlx.ai/benchmarks/performance/14wbujwi",
   "tg_tokens": null,
   "tps": 21.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-14wbujwi-ctx4096",
   "model": "gemma-4-31B-it-qat",
   "notes": "mlx-community OptiQ; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: s2pmouei, wetvt7ma, xknc2kl0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 708.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--gemma-4-31B-it-qat-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 708.1 21.1 23.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/14wbujwi",
   "source_url": "https://omlx.ai/benchmarks/performance/14wbujwi",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-14wbujwi-ctx8192",
   "model": "gemma-4-31B-it-qat",
   "notes": "mlx-community OptiQ; oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: s2pmouei, wetvt7ma, xknc2kl0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 638.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--gemma-4-31B-it-qat-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 638.7 19.8 24.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/14wbujwi",
   "source_url": "https://omlx.ai/benchmarks/performance/14wbujwi",
   "tg_tokens": null,
   "tps": 19.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-14wbujwi-ctx16384",
   "model": "gemma-4-31B-it-qat",
   "notes": "mlx-community OptiQ; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: s2pmouei, wetvt7ma, xknc2kl0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 556.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--gemma-4-31B-it-qat-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 556.4 18.3 25.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/14wbujwi",
   "source_url": "https://omlx.ai/benchmarks/performance/14wbujwi",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ibt32whi-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 108.3 GB; TTFT 4650 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 880.9,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "GLM-5.3-Flash-oQ2e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 4k tokens 880.9 PP tok/s 42.0 TG tok/s 4650 TTFT (ms) 108.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ibt32whi",
   "source_url": "https://omlx.ai/benchmarks/performance/ibt32whi",
   "tg_tokens": null,
   "tps": 42.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xv65dcfe-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 95.1 GB; TTFT 17546 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 233.4,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "GLM-5.3-Flash-oQ2e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 4k tokens 233.4 PP tok/s 22.0 TG tok/s 17546 TTFT (ms) 95.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xv65dcfe",
   "source_url": "https://omlx.ai/benchmarks/performance/xv65dcfe",
   "tg_tokens": null,
   "tps": 22.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3w2oor7e-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 41.5 GB; TTFT 26428 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 155.0,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "GLM-5.3-Flash-oQ2e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 4k tokens 155.0 PP tok/s 9.6 TG tok/s 26428 TTFT (ms) 41.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3w2oor7e",
   "source_url": "https://omlx.ai/benchmarks/performance/3w2oor7e",
   "tg_tokens": null,
   "tps": 9.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-tqbu2pt1-ctx4096",
   "model": "Huihui-Ornith-1.5-35B-A3B-abliterated",
   "notes": "oMLX community benchmark trace; peak memory 70.1 GB at this context; duplicate traces of the same session: uyjcnqk6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1707.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Huihui-Ornith-1.5-35B-A3B-abliterated on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 unknown; 4k 1,707 56.2 70.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tqbu2pt1",
   "source_url": "https://omlx.ai/benchmarks/performance/tqbu2pt1",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-tqbu2pt1-ctx16384",
   "model": "Huihui-Ornith-1.5-35B-A3B-abliterated",
   "notes": "oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: uyjcnqk6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1441.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Huihui-Ornith-1.5-35B-A3B-abliterated on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 1,441 68.3 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tqbu2pt1",
   "source_url": "https://omlx.ai/benchmarks/performance/tqbu2pt1",
   "tg_tokens": null,
   "tps": 68.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-60y5wtl0-ctx4096",
   "model": "Huihui-Ornith-1.5-35B-A3B-abliterated",
   "notes": "oMLX community benchmark trace; peak memory 70.1 GB at this context; duplicate traces of the same session: 7tgjngm2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1705.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Huihui-Ornith-1.5-35B-A3B-abliterated on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 unknown; 4k 1,705 57.2 70.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/60y5wtl0",
   "source_url": "https://omlx.ai/benchmarks/performance/60y5wtl0",
   "tg_tokens": null,
   "tps": 57.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-60y5wtl0-ctx16384",
   "model": "Huihui-Ornith-1.5-35B-A3B-abliterated",
   "notes": "oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: 7tgjngm2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1452.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Huihui-Ornith-1.5-35B-A3B-abliterated on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 1,452 73.5 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/60y5wtl0",
   "source_url": "https://omlx.ai/benchmarks/performance/60y5wtl0",
   "tg_tokens": null,
   "tps": 73.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-e1oi00nc-ctx4096",
   "model": "Huihui-Ornith-1.5-35B-A3B-abliterated",
   "notes": "oMLX community benchmark trace; peak memory 70.1 GB at this context; duplicate traces of the same session: co6z5zw1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1707.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Huihui-Ornith-1.5-35B-A3B-abliterated on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 unknown; 4k 1,707 58.2 70.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e1oi00nc",
   "source_url": "https://omlx.ai/benchmarks/performance/e1oi00nc",
   "tg_tokens": null,
   "tps": 58.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-e1oi00nc-ctx16384",
   "model": "Huihui-Ornith-1.5-35B-A3B-abliterated",
   "notes": "oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: co6z5zw1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1444.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Huihui-Ornith-1.5-35B-A3B-abliterated on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 1,444 74.3 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e1oi00nc",
   "source_url": "https://omlx.ai/benchmarks/performance/e1oi00nc",
   "tg_tokens": null,
   "tps": 74.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-k0mxirhf-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: mmwazcu4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1668.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 1,668 94.7 35.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k0mxirhf",
   "source_url": "https://omlx.ai/benchmarks/performance/k0mxirhf",
   "tg_tokens": null,
   "tps": 94.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-k0mxirhf-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: mmwazcu4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2129.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 2,129 92.6 37.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k0mxirhf",
   "source_url": "https://omlx.ai/benchmarks/performance/k0mxirhf",
   "tg_tokens": null,
   "tps": 92.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-uxml635v-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2685.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 2,685 93.2 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uxml635v",
   "source_url": "https://omlx.ai/benchmarks/performance/uxml635v",
   "tg_tokens": null,
   "tps": 93.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-uxml635v-ctx8192",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3325.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 3,325 87.1 20.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uxml635v",
   "source_url": "https://omlx.ai/benchmarks/performance/uxml635v",
   "tg_tokens": null,
   "tps": 87.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-cdk7wvoq-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: vht7boxc; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2575.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 2,575 122.6 21.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cdk7wvoq",
   "source_url": "https://omlx.ai/benchmarks/performance/cdk7wvoq",
   "tg_tokens": null,
   "tps": 122.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-cdk7wvoq-ctx8192",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: vht7boxc; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3220.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 3,220 95.0 22.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cdk7wvoq",
   "source_url": "https://omlx.ai/benchmarks/performance/cdk7wvoq",
   "tg_tokens": null,
   "tps": 95.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-tdm2s8y0-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 23.0 GB; TTFT 378 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2708.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k tokens 2,708 PP tok/s 47.3 TG tok/s 378 TTFT (ms) 23.0 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tdm2s8y0",
   "source_url": "https://omlx.ai/benchmarks/performance/tdm2s8y0",
   "tg_tokens": null,
   "tps": 47.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-2pp0h1al-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 14157kbb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2720.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 2,720 107.8 19.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2pp0h1al",
   "source_url": "https://omlx.ai/benchmarks/performance/2pp0h1al",
   "tg_tokens": null,
   "tps": 107.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-2pp0h1al-ctx8192",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: 14157kbb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3366.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 3,366 101.2 20.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2pp0h1al",
   "source_url": "https://omlx.ai/benchmarks/performance/2pp0h1al",
   "tg_tokens": null,
   "tps": 101.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1jn5sun9-ctx1024",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: getbqn3v, vxoz71v0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2552.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 2,552 120.9 21.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1jn5sun9",
   "source_url": "https://omlx.ai/benchmarks/performance/1jn5sun9",
   "tg_tokens": null,
   "tps": 120.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1jn5sun9-ctx8192",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: getbqn3v, vxoz71v0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3178.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 3,178 94.9 22.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1jn5sun9",
   "source_url": "https://omlx.ai/benchmarks/performance/1jn5sun9",
   "tg_tokens": null,
   "tps": 94.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1jn5sun9-ctx16384",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: getbqn3v, vxoz71v0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2920.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 2,920 95.5 22.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1jn5sun9",
   "source_url": "https://omlx.ai/benchmarks/performance/1jn5sun9",
   "tg_tokens": null,
   "tps": 95.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-nweyfnqg-ctx1024",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: pzmmi2fu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2711.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 2,711 93.0 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nweyfnqg",
   "source_url": "https://omlx.ai/benchmarks/performance/nweyfnqg",
   "tg_tokens": null,
   "tps": 93.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-nweyfnqg-ctx8192",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: pzmmi2fu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3324.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 3,324 86.2 20.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nweyfnqg",
   "source_url": "https://omlx.ai/benchmarks/performance/nweyfnqg",
   "tg_tokens": null,
   "tps": 86.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-793xbnwl-ctx1024",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 7k890hyz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2467.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 2,467 143.6 21.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/793xbnwl",
   "source_url": "https://omlx.ai/benchmarks/performance/793xbnwl",
   "tg_tokens": null,
   "tps": 143.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-793xbnwl-ctx8192",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 7k890hyz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3192.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 3,192 96.1 22.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/793xbnwl",
   "source_url": "https://omlx.ai/benchmarks/performance/793xbnwl",
   "tg_tokens": null,
   "tps": 96.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-hjj2oqji-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 22.9 GB; TTFT 393795 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 83.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 32k tokens 83.2 PP tok/s 6.1 TG tok/s 393795 TTFT (ms) 22.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hjj2oqji",
   "source_url": "https://omlx.ai/benchmarks/performance/hjj2oqji",
   "tg_tokens": null,
   "tps": 6.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-422jzk04-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: r8j3e9k4, x719a82j; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 1k 106.4 17.6 19.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/422jzk04",
   "source_url": "https://omlx.ai/benchmarks/performance/422jzk04",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-422jzk04-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: r8j3e9k4, x719a82j; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 94.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 94.5 15.3 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/422jzk04",
   "source_url": "https://omlx.ai/benchmarks/performance/422jzk04",
   "tg_tokens": null,
   "tps": 15.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-422jzk04-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: r8j3e9k4, x719a82j; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 91.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 8k 91.9 13.0 24.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/422jzk04",
   "source_url": "https://omlx.ai/benchmarks/performance/422jzk04",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-qug1w7jh-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.2 GB at this context; duplicate traces of the same session: ily5ukz6, lkc0lihq, ubrh0jdg; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 210.8,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 210.8 32.0 15.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qug1w7jh",
   "source_url": "https://omlx.ai/benchmarks/performance/qug1w7jh",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-qug1w7jh-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: ily5ukz6, lkc0lihq, ubrh0jdg; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 214.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 214.0 30.9 18.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qug1w7jh",
   "source_url": "https://omlx.ai/benchmarks/performance/qug1w7jh",
   "tg_tokens": null,
   "tps": 30.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-qug1w7jh-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: ily5ukz6, lkc0lihq, ubrh0jdg; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 208.2,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 8k 208.2 29.8 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qug1w7jh",
   "source_url": "https://omlx.ai/benchmarks/performance/qug1w7jh",
   "tg_tokens": null,
   "tps": 29.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-qug1w7jh-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: ily5ukz6, lkc0lihq, ubrh0jdg; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 197.3,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 16k 197.3 28.5 20.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qug1w7jh",
   "source_url": "https://omlx.ai/benchmarks/performance/qug1w7jh",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-qug1w7jh-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: ily5ukz6, lkc0lihq, ubrh0jdg; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 178.4,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 32k 178.4 26.1 22.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qug1w7jh",
   "source_url": "https://omlx.ai/benchmarks/performance/qug1w7jh",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-qug1w7jh-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: ily5ukz6, lkc0lihq, ubrh0jdg; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 149.9,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 64k 149.9 22.5 23.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qug1w7jh",
   "source_url": "https://omlx.ai/benchmarks/performance/qug1w7jh",
   "tg_tokens": null,
   "tps": 22.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-qug1w7jh-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: ily5ukz6, lkc0lihq, ubrh0jdg; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 113.5,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 128k 113.5 17.8 27.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qug1w7jh",
   "source_url": "https://omlx.ai/benchmarks/performance/qug1w7jh",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-qug1w7jh-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: ily5ukz6, lkc0lihq, ubrh0jdg; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 89.8,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 195k 89.8 14.5 32.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qug1w7jh",
   "source_url": "https://omlx.ai/benchmarks/performance/qug1w7jh",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-y71u35zm-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 37.0 GB; TTFT 219795 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 596.3,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-27B-oQ6 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 128k tokens 596.3 PP tok/s 33.3 TG tok/s 219795 TTFT (ms) 37.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y71u35zm",
   "source_url": "https://omlx.ai/benchmarks/performance/y71u35zm",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qwqkrsdk-ctx1024",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 38.1 GB at this context; duplicate traces of the same session: 0t4bz1xs, 9vokbiod, dlmz9451, gbi7ya35, nroz0s5c, q8djyqq0, tk3nz5hd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 896.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 896.0 57.4 38.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qwqkrsdk",
   "source_url": "https://omlx.ai/benchmarks/performance/qwqkrsdk",
   "tg_tokens": null,
   "tps": 57.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qwqkrsdk-ctx4096",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 38.8 GB at this context; duplicate traces of the same session: 0t4bz1xs, 9vokbiod, dlmz9451, gbi7ya35, nroz0s5c, q8djyqq0, tk3nz5hd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2585.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 2,585 83.3 38.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qwqkrsdk",
   "source_url": "https://omlx.ai/benchmarks/performance/qwqkrsdk",
   "tg_tokens": null,
   "tps": 83.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qwqkrsdk-ctx8192",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 38.9 GB at this context; duplicate traces of the same session: 0t4bz1xs, 9vokbiod, dlmz9451, gbi7ya35, nroz0s5c, q8djyqq0, tk3nz5hd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5641.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 5,641 78.5 38.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qwqkrsdk",
   "source_url": "https://omlx.ai/benchmarks/performance/qwqkrsdk",
   "tg_tokens": null,
   "tps": 78.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qwqkrsdk-ctx16384",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 39.1 GB at this context; duplicate traces of the same session: 0t4bz1xs, 9vokbiod, dlmz9451, gbi7ya35, nroz0s5c, q8djyqq0, tk3nz5hd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5487.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 5,487 83.2 39.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qwqkrsdk",
   "source_url": "https://omlx.ai/benchmarks/performance/qwqkrsdk",
   "tg_tokens": null,
   "tps": 83.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qwqkrsdk-ctx32768",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 39.5 GB at this context; duplicate traces of the same session: 0t4bz1xs, 9vokbiod, dlmz9451, gbi7ya35, nroz0s5c, q8djyqq0, tk3nz5hd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4371.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 4,371 75.9 39.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qwqkrsdk",
   "source_url": "https://omlx.ai/benchmarks/performance/qwqkrsdk",
   "tg_tokens": null,
   "tps": 75.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qwqkrsdk-ctx65536",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 40.2 GB at this context; duplicate traces of the same session: 0t4bz1xs, 9vokbiod, dlmz9451, gbi7ya35, nroz0s5c, q8djyqq0, tk3nz5hd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3107.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 3,107 52.5 40.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qwqkrsdk",
   "source_url": "https://omlx.ai/benchmarks/performance/qwqkrsdk",
   "tg_tokens": null,
   "tps": 52.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qwqkrsdk-ctx131072",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 41.7 GB at this context; duplicate traces of the same session: 0t4bz1xs, 9vokbiod, dlmz9451, gbi7ya35, nroz0s5c, q8djyqq0, tk3nz5hd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1999.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 1,999 41.1 41.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qwqkrsdk",
   "source_url": "https://omlx.ai/benchmarks/performance/qwqkrsdk",
   "tg_tokens": null,
   "tps": 41.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qwqkrsdk-ctx199680",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 44.2 GB at this context; duplicate traces of the same session: 0t4bz1xs, 9vokbiod, dlmz9451, gbi7ya35, nroz0s5c, q8djyqq0, tk3nz5hd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1207.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 195k 1,207 24.0 44.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qwqkrsdk",
   "source_url": "https://omlx.ai/benchmarks/performance/qwqkrsdk",
   "tg_tokens": null,
   "tps": 24.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-03w4sgts-ctx1024",
   "model": "Swift-Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.2 GB at this context; duplicate traces of the same session: xjk3v3yc; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.2,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Swift-Qwen3.8-27B-MLX-MXFP4 on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 106.2 16.4 15.2 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/03w4sgts",
   "source_url": "https://omlx.ai/benchmarks/performance/03w4sgts",
   "tg_tokens": null,
   "tps": 16.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-03w4sgts-ctx4096",
   "model": "Swift-Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.2 GB at this context; duplicate traces of the same session: xjk3v3yc; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.2,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Swift-Qwen3.8-27B-MLX-MXFP4 on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 106.2 16.1 16.2 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/03w4sgts",
   "source_url": "https://omlx.ai/benchmarks/performance/03w4sgts",
   "tg_tokens": null,
   "tps": 16.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-fl9gh94l-ctx1024",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MTPLX Optimized Speed build; oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: ht1a96ly; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 216.6,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Ternary-Bonsai-2-27B-MTPLX-Optimized-Speed on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 unknown; 1k 216.6 36.2 11.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fl9gh94l",
   "source_url": "https://omlx.ai/benchmarks/performance/fl9gh94l",
   "tg_tokens": null,
   "tps": 36.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-fl9gh94l-ctx4096",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MTPLX Optimized Speed build; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: ht1a96ly; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 173.1,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Ternary-Bonsai-2-27B-MTPLX-Optimized-Speed on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 unknown; 4k 173.1 31.5 19.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fl9gh94l",
   "source_url": "https://omlx.ai/benchmarks/performance/fl9gh94l",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-fl9gh94l-ctx16384",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MTPLX Optimized Speed build; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: ht1a96ly; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 155.5,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Ternary-Bonsai-2-27B-MTPLX-Optimized-Speed on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 unknown; 16k 155.5 27.4 19.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fl9gh94l",
   "source_url": "https://omlx.ai/benchmarks/performance/fl9gh94l",
   "tg_tokens": null,
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-vbdnxv01-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5492.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 5,492 187.2 22.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vbdnxv01",
   "source_url": "https://omlx.ai/benchmarks/performance/vbdnxv01",
   "tg_tokens": null,
   "tps": 187.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-vbdnxv01-ctx16384",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5450.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 5,450 204.6 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vbdnxv01",
   "source_url": "https://omlx.ai/benchmarks/performance/vbdnxv01",
   "tg_tokens": null,
   "tps": 204.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-b193slfi-ctx1024",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 156.7 GB at this context; duplicate traces of the same session: 20p72doq, 2wpy8wg6, ehez40hp, rf1dd8u0, te9lnqg8, xxiqriz1, zosel0jc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 606.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-V2.6-Flash-RL-MLX-4bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 606.1 50.9 156.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b193slfi",
   "source_url": "https://omlx.ai/benchmarks/performance/b193slfi",
   "tg_tokens": null,
   "tps": 50.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-b193slfi-ctx4096",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 157.8 GB at this context; duplicate traces of the same session: 20p72doq, 2wpy8wg6, ehez40hp, rf1dd8u0, te9lnqg8, xxiqriz1, zosel0jc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 673.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-V2.6-Flash-RL-MLX-4bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 673.1 51.7 157.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b193slfi",
   "source_url": "https://omlx.ai/benchmarks/performance/b193slfi",
   "tg_tokens": null,
   "tps": 51.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-b193slfi-ctx8192",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 158.9 GB at this context; duplicate traces of the same session: 20p72doq, 2wpy8wg6, ehez40hp, rf1dd8u0, te9lnqg8, xxiqriz1, zosel0jc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 655.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-V2.6-Flash-RL-MLX-4bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 655.5 48.7 158.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b193slfi",
   "source_url": "https://omlx.ai/benchmarks/performance/b193slfi",
   "tg_tokens": null,
   "tps": 48.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-b193slfi-ctx16384",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 161.1 GB at this context; duplicate traces of the same session: 20p72doq, 2wpy8wg6, ehez40hp, rf1dd8u0, te9lnqg8, xxiqriz1, zosel0jc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 618.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-V2.6-Flash-RL-MLX-4bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 618.3 45.6 161.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b193slfi",
   "source_url": "https://omlx.ai/benchmarks/performance/b193slfi",
   "tg_tokens": null,
   "tps": 45.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-b193slfi-ctx32768",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 165.4 GB at this context; duplicate traces of the same session: 20p72doq, 2wpy8wg6, ehez40hp, rf1dd8u0, te9lnqg8, xxiqriz1, zosel0jc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 547.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-V2.6-Flash-RL-MLX-4bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 547.6 41.4 165.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b193slfi",
   "source_url": "https://omlx.ai/benchmarks/performance/b193slfi",
   "tg_tokens": null,
   "tps": 41.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-b193slfi-ctx65536",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 174.1 GB at this context; duplicate traces of the same session: 20p72doq, 2wpy8wg6, ehez40hp, rf1dd8u0, te9lnqg8, xxiqriz1, zosel0jc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 439.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-V2.6-Flash-RL-MLX-4bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 439.8 34.4 174.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b193slfi",
   "source_url": "https://omlx.ai/benchmarks/performance/b193slfi",
   "tg_tokens": null,
   "tps": 34.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-b193slfi-ctx131072",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 191.6 GB at this context; duplicate traces of the same session: 20p72doq, 2wpy8wg6, ehez40hp, rf1dd8u0, te9lnqg8, xxiqriz1, zosel0jc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 309.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-V2.6-Flash-RL-MLX-4bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 309.8 25.4 191.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b193slfi",
   "source_url": "https://omlx.ai/benchmarks/performance/b193slfi",
   "tg_tokens": null,
   "tps": 25.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-b193slfi-ctx199680",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 209.7 GB at this context; duplicate traces of the same session: 20p72doq, 2wpy8wg6, ehez40hp, rf1dd8u0, te9lnqg8, xxiqriz1, zosel0jc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 233.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-V2.6-Flash-RL-MLX-4bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 195k 233.1 20.5 209.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b193slfi",
   "source_url": "https://omlx.ai/benchmarks/performance/b193slfi",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-z8182sc2-ctx1024",
   "model": "Qwen3.5-9B-DeepSeek-V4-Flash",
   "notes": "oMLX community benchmark trace; peak memory 9.8 GB at this context; duplicate traces of the same session: gergs0an; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 222.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B-DeepSeek-V4-Flash on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 8bit; 1k 222.4 11.3 9.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z8182sc2",
   "source_url": "https://omlx.ai/benchmarks/performance/z8182sc2",
   "tg_tokens": null,
   "tps": 11.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-z8182sc2-ctx4096",
   "model": "Qwen3.5-9B-DeepSeek-V4-Flash",
   "notes": "oMLX community benchmark trace; peak memory 9.9 GB at this context; duplicate traces of the same session: gergs0an; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 222.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B-DeepSeek-V4-Flash on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 8bit; 4k 222.1 10.9 9.9 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z8182sc2",
   "source_url": "https://omlx.ai/benchmarks/performance/z8182sc2",
   "tg_tokens": null,
   "tps": 10.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-smv913tx-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 7vby3gf2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 405.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 405.4 49.6 20.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/smv913tx",
   "source_url": "https://omlx.ai/benchmarks/performance/smv913tx",
   "tg_tokens": null,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-smv913tx-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: 7vby3gf2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 339.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k 339.9 37.1 20.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/smv913tx",
   "source_url": "https://omlx.ai/benchmarks/performance/smv913tx",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-uklel3ss-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1881.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 1,881 94.7 35.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uklel3ss",
   "source_url": "https://omlx.ai/benchmarks/performance/uklel3ss",
   "tg_tokens": null,
   "tps": 94.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-uklel3ss-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2130.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 2,130 92.8 37.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uklel3ss",
   "source_url": "https://omlx.ai/benchmarks/performance/uklel3ss",
   "tg_tokens": null,
   "tps": 92.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ql8egefe-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2685.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 2,685 93.2 20.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ql8egefe",
   "source_url": "https://omlx.ai/benchmarks/performance/ql8egefe",
   "tg_tokens": null,
   "tps": 93.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ql8egefe-ctx8192",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3325.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 3,325 87.1 20.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ql8egefe",
   "source_url": "https://omlx.ai/benchmarks/performance/ql8egefe",
   "tg_tokens": null,
   "tps": 87.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-wvvediqz-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: oef9jcpb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2422.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 2,422 115.9 21.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wvvediqz",
   "source_url": "https://omlx.ai/benchmarks/performance/wvvediqz",
   "tg_tokens": null,
   "tps": 115.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-wvvediqz-ctx8192",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: oef9jcpb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3221.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 3,221 100.5 22.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wvvediqz",
   "source_url": "https://omlx.ai/benchmarks/performance/wvvediqz",
   "tg_tokens": null,
   "tps": 100.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-z55qvtpg-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 2jq26umt, ygbhy4ky; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2577.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 2,577 93.3 20.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z55qvtpg",
   "source_url": "https://omlx.ai/benchmarks/performance/z55qvtpg",
   "tg_tokens": null,
   "tps": 93.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-z55qvtpg-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: 2jq26umt, ygbhy4ky; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3377.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 3,377 89.3 20.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z55qvtpg",
   "source_url": "https://omlx.ai/benchmarks/performance/z55qvtpg",
   "tg_tokens": null,
   "tps": 89.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-z55qvtpg-ctx8192",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 2jq26umt, ygbhy4ky; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3310.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 3,310 87.1 20.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z55qvtpg",
   "source_url": "https://omlx.ai/benchmarks/performance/z55qvtpg",
   "tg_tokens": null,
   "tps": 87.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-d2fkexkv-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: c08njqxl, n14pgknl; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2706.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 2,706 92.8 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d2fkexkv",
   "source_url": "https://omlx.ai/benchmarks/performance/d2fkexkv",
   "tg_tokens": null,
   "tps": 92.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-d2fkexkv-ctx8192",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: c08njqxl, n14pgknl; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3336.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 3,336 86.1 20.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d2fkexkv",
   "source_url": "https://omlx.ai/benchmarks/performance/d2fkexkv",
   "tg_tokens": null,
   "tps": 86.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-d2fkexkv-ctx16384",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: c08njqxl, n14pgknl; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3060.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 3,060 82.0 21.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d2fkexkv",
   "source_url": "https://omlx.ai/benchmarks/performance/d2fkexkv",
   "tg_tokens": null,
   "tps": 82.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-dex5j5jk-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: esulq3v3, xsqw2bk8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2683.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 2,683 106.1 19.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dex5j5jk",
   "source_url": "https://omlx.ai/benchmarks/performance/dex5j5jk",
   "tg_tokens": null,
   "tps": 106.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-dex5j5jk-ctx8192",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: esulq3v3, xsqw2bk8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3309.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 3,309 93.4 20.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dex5j5jk",
   "source_url": "https://omlx.ai/benchmarks/performance/dex5j5jk",
   "tg_tokens": null,
   "tps": 93.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-dex5j5jk-ctx16384",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: esulq3v3, xsqw2bk8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3058.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 3,058 89.6 20.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dex5j5jk",
   "source_url": "https://omlx.ai/benchmarks/performance/dex5j5jk",
   "tg_tokens": null,
   "tps": 89.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-1scidelz-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 29.9 GB; TTFT 2025 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 505.6,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 6bit; 1k tokens 505.6 PP tok/s 52.8 TG tok/s 2025 TTFT (ms) 29.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1scidelz",
   "source_url": "https://omlx.ai/benchmarks/performance/1scidelz",
   "tg_tokens": null,
   "tps": 52.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-42k6pbr0-ctx1024",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: dj7cdtkw, x5a3spbq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2717.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 2,717 93.1 20.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/42k6pbr0",
   "source_url": "https://omlx.ai/benchmarks/performance/42k6pbr0",
   "tg_tokens": null,
   "tps": 93.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-42k6pbr0-ctx8192",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: dj7cdtkw, x5a3spbq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3337.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 3,337 87.1 20.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/42k6pbr0",
   "source_url": "https://omlx.ai/benchmarks/performance/42k6pbr0",
   "tg_tokens": null,
   "tps": 87.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-42k6pbr0-ctx16384",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: dj7cdtkw, x5a3spbq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3071.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 3,071 81.0 21.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/42k6pbr0",
   "source_url": "https://omlx.ai/benchmarks/performance/42k6pbr0",
   "tg_tokens": null,
   "tps": 81.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jmj7h74m-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: 0qv7jquv, 8mzcg7hp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 679.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 679.3 28.7 18.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jmj7h74m",
   "source_url": "https://omlx.ai/benchmarks/performance/jmj7h74m",
   "tg_tokens": null,
   "tps": 28.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jmj7h74m-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 0qv7jquv, 8mzcg7hp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 549.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 549.8 30.7 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jmj7h74m",
   "source_url": "https://omlx.ai/benchmarks/performance/jmj7h74m",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jmj7h74m-ctx65536",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: 0qv7jquv, 8mzcg7hp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 298.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 298.9 20.9 22.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jmj7h74m",
   "source_url": "https://omlx.ai/benchmarks/performance/jmj7h74m",
   "tg_tokens": null,
   "tps": 20.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jmj7h74m-ctx131072",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: 0qv7jquv, 8mzcg7hp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 290.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 290.3 19.6 26.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jmj7h74m",
   "source_url": "https://omlx.ai/benchmarks/performance/jmj7h74m",
   "tg_tokens": null,
   "tps": 19.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-jwx541g5-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: fltv161j; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 91.9,
   "provenance": "community",
   "quant": "oQ6e-fp16",
   "quote": "Qwen3.8-27B-oQ6e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 6bit; 1k 91.9 9.9 22.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jwx541g5",
   "source_url": "https://omlx.ai/benchmarks/performance/jwx541g5",
   "tg_tokens": null,
   "tps": 9.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-jwx541g5-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: fltv161j; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 86.9,
   "provenance": "community",
   "quant": "oQ6e-fp16",
   "quote": "Qwen3.8-27B-oQ6e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 6bit; 4k 86.9 8.8 26.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jwx541g5",
   "source_url": "https://omlx.ai/benchmarks/performance/jwx541g5",
   "tg_tokens": null,
   "tps": 8.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-darmfrvj-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: xrjugumb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 543.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 543.6 52.5 70.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/darmfrvj",
   "source_url": "https://omlx.ai/benchmarks/performance/darmfrvj",
   "tg_tokens": null,
   "tps": 52.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-darmfrvj-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: xrjugumb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 588.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 588.2 54.5 71.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/darmfrvj",
   "source_url": "https://omlx.ai/benchmarks/performance/darmfrvj",
   "tg_tokens": null,
   "tps": 54.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-j2nl4ado-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: kgy0oe5b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 656.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 656.2 30.0 100.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j2nl4ado",
   "source_url": "https://omlx.ai/benchmarks/performance/j2nl4ado",
   "tg_tokens": null,
   "tps": 30.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-j2nl4ado-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: kgy0oe5b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 734.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 734.1 70.0 101.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j2nl4ado",
   "source_url": "https://omlx.ai/benchmarks/performance/j2nl4ado",
   "tg_tokens": null,
   "tps": 70.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-qhbhy61v-ctx1024",
   "model": "Sharp-MiniCPM5-2B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 2.6 GB; TTFT 1788 ms; duplicate traces of the same session: none; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 572.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Sharp-MiniCPM5-2B-MLX-oQ6e on M4 (8c); M4 (8c) \u00b7 16 GB \u00b7 6bit; 1k tokens 572.7 PP tok/s 50.4 TG tok/s 1788 TTFT (ms) 2.6 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qhbhy61v",
   "source_url": "https://omlx.ai/benchmarks/performance/qhbhy61v",
   "tg_tokens": null,
   "tps": 50.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-1mqkh7pt-ctx1024",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MTPLX Optimized Speed build; oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 216.6,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Ternary-Bonsai-2-27B-MTPLX-Optimized-Speed on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 unknown; 1k 216.6 36.2 11.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1mqkh7pt",
   "source_url": "https://omlx.ai/benchmarks/performance/1mqkh7pt",
   "tg_tokens": null,
   "tps": 36.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-1mqkh7pt-ctx4096",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MTPLX Optimized Speed build; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 173.1,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Ternary-Bonsai-2-27B-MTPLX-Optimized-Speed on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 unknown; 4k 173.1 31.5 19.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1mqkh7pt",
   "source_url": "https://omlx.ai/benchmarks/performance/1mqkh7pt",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-1mqkh7pt-ctx16384",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MTPLX Optimized Speed build; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 155.5,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Ternary-Bonsai-2-27B-MTPLX-Optimized-Speed on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 unknown; 16k 155.5 27.4 19.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1mqkh7pt",
   "source_url": "https://omlx.ai/benchmarks/performance/1mqkh7pt",
   "tg_tokens": null,
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-4ystfvka-ctx32768",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace, single run at 32k context; peak memory 22.1 GB; TTFT 313956 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.4,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 2bit; 32k tokens 104.4 PP tok/s 10.6 TG tok/s 313956 TTFT (ms) 22.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ystfvka",
   "source_url": "https://omlx.ai/benchmarks/performance/4ystfvka",
   "tg_tokens": null,
   "tps": 10.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ps09r0nq-ctx1024",
   "model": "gemma-4-26B-A4B-it-qat",
   "notes": "mlx-community OptiQ; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 6lw17cx6, 6snfkf4b, dhecu6eb, ewn77adx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3678.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--gemma-4-26B-A4B-it-qat-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 3,678 95.4 20.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ps09r0nq",
   "source_url": "https://omlx.ai/benchmarks/performance/ps09r0nq",
   "tg_tokens": null,
   "tps": 95.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ps09r0nq-ctx4096",
   "model": "gemma-4-26B-A4B-it-qat",
   "notes": "mlx-community OptiQ; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: 6lw17cx6, 6snfkf4b, dhecu6eb, ewn77adx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4107.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--gemma-4-26B-A4B-it-qat-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 4,107 95.9 20.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ps09r0nq",
   "source_url": "https://omlx.ai/benchmarks/performance/ps09r0nq",
   "tg_tokens": null,
   "tps": 95.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ps09r0nq-ctx8192",
   "model": "gemma-4-26B-A4B-it-qat",
   "notes": "mlx-community OptiQ; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 6lw17cx6, 6snfkf4b, dhecu6eb, ewn77adx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4035.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--gemma-4-26B-A4B-it-qat-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 4,035 91.3 20.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ps09r0nq",
   "source_url": "https://omlx.ai/benchmarks/performance/ps09r0nq",
   "tg_tokens": null,
   "tps": 91.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ps09r0nq-ctx16384",
   "model": "gemma-4-26B-A4B-it-qat",
   "notes": "mlx-community OptiQ; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: 6lw17cx6, 6snfkf4b, dhecu6eb, ewn77adx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3756.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--gemma-4-26B-A4B-it-qat-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 3,756 85.4 21.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ps09r0nq",
   "source_url": "https://omlx.ai/benchmarks/performance/ps09r0nq",
   "tg_tokens": null,
   "tps": 85.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ps09r0nq-ctx32768",
   "model": "gemma-4-26B-A4B-it-qat",
   "notes": "mlx-community OptiQ; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 6lw17cx6, 6snfkf4b, dhecu6eb, ewn77adx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3307.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--gemma-4-26B-A4B-it-qat-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 3,307 74.3 21.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ps09r0nq",
   "source_url": "https://omlx.ai/benchmarks/performance/ps09r0nq",
   "tg_tokens": null,
   "tps": 74.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-0ep1vn2j-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.5 GB at this context; duplicate traces of the same session: felcbyi0, fo5ca56v, iz1bppvc, yu82moxo; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1307.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-MLX-oQ4-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,307 69.1 177.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ep1vn2j",
   "source_url": "https://omlx.ai/benchmarks/performance/0ep1vn2j",
   "tg_tokens": null,
   "tps": 69.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-0ep1vn2j-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: felcbyi0, fo5ca56v, iz1bppvc, yu82moxo; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1902.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-MLX-oQ4-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,902 71.5 178.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ep1vn2j",
   "source_url": "https://omlx.ai/benchmarks/performance/0ep1vn2j",
   "tg_tokens": null,
   "tps": 71.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-0ep1vn2j-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.9 GB at this context; duplicate traces of the same session: felcbyi0, fo5ca56v, iz1bppvc, yu82moxo; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1893.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-MLX-oQ4-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 1,893 76.9 178.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ep1vn2j",
   "source_url": "https://omlx.ai/benchmarks/performance/0ep1vn2j",
   "tg_tokens": null,
   "tps": 76.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-0ep1vn2j-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.1 GB at this context; duplicate traces of the same session: felcbyi0, fo5ca56v, iz1bppvc, yu82moxo; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1882.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-MLX-oQ4-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,882 67.3 179.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ep1vn2j",
   "source_url": "https://omlx.ai/benchmarks/performance/0ep1vn2j",
   "tg_tokens": null,
   "tps": 67.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-0ep1vn2j-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.4 GB at this context; duplicate traces of the same session: felcbyi0, fo5ca56v, iz1bppvc, yu82moxo; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1901.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-MLX-oQ4-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,901 60.7 179.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ep1vn2j",
   "source_url": "https://omlx.ai/benchmarks/performance/0ep1vn2j",
   "tg_tokens": null,
   "tps": 60.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-16sgrdhx-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.5 GB at this context; duplicate traces of the same session: 597nr82v, jmiuqqob, lbau60or, z2t457ny; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1192.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-MLX-oQ4-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,192 70.4 177.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/16sgrdhx",
   "source_url": "https://omlx.ai/benchmarks/performance/16sgrdhx",
   "tg_tokens": null,
   "tps": 70.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-16sgrdhx-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: 597nr82v, jmiuqqob, lbau60or, z2t457ny; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1895.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-MLX-oQ4-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,895 58.7 178.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/16sgrdhx",
   "source_url": "https://omlx.ai/benchmarks/performance/16sgrdhx",
   "tg_tokens": null,
   "tps": 58.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-16sgrdhx-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.9 GB at this context; duplicate traces of the same session: 597nr82v, jmiuqqob, lbau60or, z2t457ny; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1883.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-MLX-oQ4-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 1,883 63.3 178.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/16sgrdhx",
   "source_url": "https://omlx.ai/benchmarks/performance/16sgrdhx",
   "tg_tokens": null,
   "tps": 63.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-16sgrdhx-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.1 GB at this context; duplicate traces of the same session: 597nr82v, jmiuqqob, lbau60or, z2t457ny; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1871.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-MLX-oQ4-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,871 57.8 179.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/16sgrdhx",
   "source_url": "https://omlx.ai/benchmarks/performance/16sgrdhx",
   "tg_tokens": null,
   "tps": 57.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-16sgrdhx-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.5 GB at this context; duplicate traces of the same session: 597nr82v, jmiuqqob, lbau60or, z2t457ny; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1862.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-MLX-oQ4-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,862 58.5 179.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/16sgrdhx",
   "source_url": "https://omlx.ai/benchmarks/performance/16sgrdhx",
   "tg_tokens": null,
   "tps": 58.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-n2jsnexy-ctx8192",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: gj13e8g9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3130.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 3,130 154.6 23.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n2jsnexy",
   "source_url": "https://omlx.ai/benchmarks/performance/n2jsnexy",
   "tg_tokens": null,
   "tps": 154.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-n2jsnexy-ctx32768",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: gj13e8g9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3896.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 3,896 160.0 23.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n2jsnexy",
   "source_url": "https://omlx.ai/benchmarks/performance/n2jsnexy",
   "tg_tokens": null,
   "tps": 160.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-qbbll59y-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: 5e33m7xy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 553.1,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.5-122B-A10B-oQ4-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 553.1 57.3 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qbbll59y",
   "source_url": "https://omlx.ai/benchmarks/performance/qbbll59y",
   "tg_tokens": null,
   "tps": 57.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-qbbll59y-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.9 GB at this context; duplicate traces of the same session: 5e33m7xy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 627.3,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.5-122B-A10B-oQ4-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 627.3 58.3 73.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qbbll59y",
   "source_url": "https://omlx.ai/benchmarks/performance/qbbll59y",
   "tg_tokens": null,
   "tps": 58.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-2qzt48zb-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1881.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 1,881 94.7 35.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2qzt48zb",
   "source_url": "https://omlx.ai/benchmarks/performance/2qzt48zb",
   "tg_tokens": null,
   "tps": 94.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-2qzt48zb-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2130.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 2,130 92.8 37.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2qzt48zb",
   "source_url": "https://omlx.ai/benchmarks/performance/2qzt48zb",
   "tg_tokens": null,
   "tps": 92.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-l97rgger-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: ovrop5dz, xybt92hh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 666.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 666.1 90.7 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l97rgger",
   "source_url": "https://omlx.ai/benchmarks/performance/l97rgger",
   "tg_tokens": null,
   "tps": 90.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-l97rgger-ctx8192",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: ovrop5dz, xybt92hh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2311.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 2,311 86.3 20.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l97rgger",
   "source_url": "https://omlx.ai/benchmarks/performance/l97rgger",
   "tg_tokens": null,
   "tps": 86.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-l97rgger-ctx16384",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: ovrop5dz, xybt92hh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3011.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 3,011 79.6 21.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l97rgger",
   "source_url": "https://omlx.ai/benchmarks/performance/l97rgger",
   "tg_tokens": null,
   "tps": 79.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-tx2933qv-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 5z1wwlux, vv30azh6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2734.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 2,734 107.6 19.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tx2933qv",
   "source_url": "https://omlx.ai/benchmarks/performance/tx2933qv",
   "tg_tokens": null,
   "tps": 107.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-tx2933qv-ctx8192",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: 5z1wwlux, vv30azh6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3364.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 3,364 102.4 20.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tx2933qv",
   "source_url": "https://omlx.ai/benchmarks/performance/tx2933qv",
   "tg_tokens": null,
   "tps": 102.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-tx2933qv-ctx16384",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: 5z1wwlux, vv30azh6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3033.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 3,033 92.2 20.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tx2933qv",
   "source_url": "https://omlx.ai/benchmarks/performance/tx2933qv",
   "tg_tokens": null,
   "tps": 92.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ywpenb83-ctx1024",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: msgggg0y, u8hf1m1l; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2524.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 2,524 108.4 21.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ywpenb83",
   "source_url": "https://omlx.ai/benchmarks/performance/ywpenb83",
   "tg_tokens": null,
   "tps": 108.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ywpenb83-ctx8192",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: msgggg0y, u8hf1m1l; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3182.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 3,182 111.1 22.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ywpenb83",
   "source_url": "https://omlx.ai/benchmarks/performance/ywpenb83",
   "tg_tokens": null,
   "tps": 111.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ywpenb83-ctx16384",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: msgggg0y, u8hf1m1l; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2936.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 2,936 144.9 22.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ywpenb83",
   "source_url": "https://omlx.ai/benchmarks/performance/ywpenb83",
   "tg_tokens": null,
   "tps": 144.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-az1j383x-ctx1024",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 04uiuwhx, mr5rivpo; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2736.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 2,736 92.3 20.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/az1j383x",
   "source_url": "https://omlx.ai/benchmarks/performance/az1j383x",
   "tg_tokens": null,
   "tps": 92.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-az1j383x-ctx8192",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 04uiuwhx, mr5rivpo; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3272.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 3,272 83.9 20.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/az1j383x",
   "source_url": "https://omlx.ai/benchmarks/performance/az1j383x",
   "tg_tokens": null,
   "tps": 83.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-az1j383x-ctx16384",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 04uiuwhx, mr5rivpo; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3020.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 3,020 80.6 21.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/az1j383x",
   "source_url": "https://omlx.ai/benchmarks/performance/az1j383x",
   "tg_tokens": null,
   "tps": 80.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-jtho24nd-ctx1024",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 6jemjalp, qo5ppuad; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2238.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 2,238 75.3 21.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jtho24nd",
   "source_url": "https://omlx.ai/benchmarks/performance/jtho24nd",
   "tg_tokens": null,
   "tps": 75.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-jtho24nd-ctx8192",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 6jemjalp, qo5ppuad; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3182.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 3,182 65.5 22.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jtho24nd",
   "source_url": "https://omlx.ai/benchmarks/performance/jtho24nd",
   "tg_tokens": null,
   "tps": 65.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-jtho24nd-ctx16384",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 6jemjalp, qo5ppuad; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2902.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 2,902 125.6 22.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jtho24nd",
   "source_url": "https://omlx.ai/benchmarks/performance/jtho24nd",
   "tg_tokens": null,
   "tps": 125.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-ml02pahl-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.2 GB at this context; duplicate traces of the same session: 5k5yylga; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 162.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 8bit; 4k 162.1 18.1 32.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ml02pahl",
   "source_url": "https://omlx.ai/benchmarks/performance/ml02pahl",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-ml02pahl-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.2 GB at this context; duplicate traces of the same session: 5k5yylga; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 147.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 8bit; 16k 147.7 16.7 34.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ml02pahl",
   "source_url": "https://omlx.ai/benchmarks/performance/ml02pahl",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-lcuxvgsm-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: btljuqum; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 252.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 252.0 38.7 25.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lcuxvgsm",
   "source_url": "https://omlx.ai/benchmarks/performance/lcuxvgsm",
   "tg_tokens": null,
   "tps": 38.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-lcuxvgsm-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: btljuqum; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 215.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 215.4 29.3 27.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lcuxvgsm",
   "source_url": "https://omlx.ai/benchmarks/performance/lcuxvgsm",
   "tg_tokens": null,
   "tps": 29.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-9o57t8rw-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.3 GB at this context; duplicate traces of the same session: p3sip173; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 250.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 250.0 14.6 33.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9o57t8rw",
   "source_url": "https://omlx.ai/benchmarks/performance/9o57t8rw",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-9o57t8rw-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 36.4 GB at this context; duplicate traces of the same session: p3sip173; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 303.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 303.7 13.6 36.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9o57t8rw",
   "source_url": "https://omlx.ai/benchmarks/performance/9o57t8rw",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7mzds81t-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 679.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 679.3 28.7 18.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7mzds81t",
   "source_url": "https://omlx.ai/benchmarks/performance/7mzds81t",
   "tg_tokens": null,
   "tps": 28.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7mzds81t-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 549.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 549.8 30.7 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7mzds81t",
   "source_url": "https://omlx.ai/benchmarks/performance/7mzds81t",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7mzds81t-ctx65536",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 298.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 298.9 20.9 22.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7mzds81t",
   "source_url": "https://omlx.ai/benchmarks/performance/7mzds81t",
   "tg_tokens": null,
   "tps": 20.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7mzds81t-ctx131072",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 290.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 290.3 19.6 26.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7mzds81t",
   "source_url": "https://omlx.ai/benchmarks/performance/7mzds81t",
   "tg_tokens": null,
   "tps": 19.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-cbdvcefo-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 475.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 475.7 34.2 18.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cbdvcefo",
   "source_url": "https://omlx.ai/benchmarks/performance/cbdvcefo",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-cbdvcefo-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 425.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 425.7 43.0 19.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cbdvcefo",
   "source_url": "https://omlx.ai/benchmarks/performance/cbdvcefo",
   "tg_tokens": null,
   "tps": 43.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ct3o3ex7-ctx1024",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 38.1 GB at this context; duplicate traces of the same session: elcygk0a, eoda1lgi, hphhauri, rwy60hnz, tcgh5kwr, wiybw8ky, z05e9zbv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3314.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 3,314 169.5 38.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ct3o3ex7",
   "source_url": "https://omlx.ai/benchmarks/performance/ct3o3ex7",
   "tg_tokens": null,
   "tps": 169.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ct3o3ex7-ctx4096",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 38.8 GB at this context; duplicate traces of the same session: elcygk0a, eoda1lgi, hphhauri, rwy60hnz, tcgh5kwr, wiybw8ky, z05e9zbv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5843.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 5,843 154.3 38.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ct3o3ex7",
   "source_url": "https://omlx.ai/benchmarks/performance/ct3o3ex7",
   "tg_tokens": null,
   "tps": 154.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ct3o3ex7-ctx8192",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 38.9 GB at this context; duplicate traces of the same session: elcygk0a, eoda1lgi, hphhauri, rwy60hnz, tcgh5kwr, wiybw8ky, z05e9zbv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5685.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 5,685 146.6 38.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ct3o3ex7",
   "source_url": "https://omlx.ai/benchmarks/performance/ct3o3ex7",
   "tg_tokens": null,
   "tps": 146.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ct3o3ex7-ctx16384",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 39.1 GB at this context; duplicate traces of the same session: elcygk0a, eoda1lgi, hphhauri, rwy60hnz, tcgh5kwr, wiybw8ky, z05e9zbv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5318.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 5,318 200.3 39.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ct3o3ex7",
   "source_url": "https://omlx.ai/benchmarks/performance/ct3o3ex7",
   "tg_tokens": null,
   "tps": 200.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ct3o3ex7-ctx32768",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 39.5 GB at this context; duplicate traces of the same session: elcygk0a, eoda1lgi, hphhauri, rwy60hnz, tcgh5kwr, wiybw8ky, z05e9zbv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4249.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 4,249 197.2 39.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ct3o3ex7",
   "source_url": "https://omlx.ai/benchmarks/performance/ct3o3ex7",
   "tg_tokens": null,
   "tps": 197.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ct3o3ex7-ctx65536",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 40.2 GB at this context; duplicate traces of the same session: elcygk0a, eoda1lgi, hphhauri, rwy60hnz, tcgh5kwr, wiybw8ky, z05e9zbv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2879.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 2,879 122.6 40.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ct3o3ex7",
   "source_url": "https://omlx.ai/benchmarks/performance/ct3o3ex7",
   "tg_tokens": null,
   "tps": 122.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ct3o3ex7-ctx131072",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 41.7 GB at this context; duplicate traces of the same session: elcygk0a, eoda1lgi, hphhauri, rwy60hnz, tcgh5kwr, wiybw8ky, z05e9zbv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1619.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 1,619 56.1 41.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ct3o3ex7",
   "source_url": "https://omlx.ai/benchmarks/performance/ct3o3ex7",
   "tg_tokens": null,
   "tps": 56.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ct3o3ex7-ctx199680",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 43.5 GB at this context; duplicate traces of the same session: elcygk0a, eoda1lgi, hphhauri, rwy60hnz, tcgh5kwr, wiybw8ky, z05e9zbv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 958.4,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 195k 958.4 47.1 43.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ct3o3ex7",
   "source_url": "https://omlx.ai/benchmarks/performance/ct3o3ex7",
   "tg_tokens": null,
   "tps": 47.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1cti56qt-ctx1024",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 177.7 GB at this context; duplicate traces of the same session: 2uq4tj9g, homyx3jc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 519.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 519.5 43.4 177.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1cti56qt",
   "source_url": "https://omlx.ai/benchmarks/performance/1cti56qt",
   "tg_tokens": null,
   "tps": 43.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1cti56qt-ctx4096",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 178.8 GB at this context; duplicate traces of the same session: 2uq4tj9g, homyx3jc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 591.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 591.4 35.4 178.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1cti56qt",
   "source_url": "https://omlx.ai/benchmarks/performance/1cti56qt",
   "tg_tokens": null,
   "tps": 35.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1cti56qt-ctx8192",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 179.1 GB at this context; duplicate traces of the same session: 2uq4tj9g, homyx3jc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 553.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 553.2 45.4 179.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1cti56qt",
   "source_url": "https://omlx.ai/benchmarks/performance/1cti56qt",
   "tg_tokens": null,
   "tps": 45.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1cti56qt-ctx16384",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 179.3 GB at this context; duplicate traces of the same session: 2uq4tj9g, homyx3jc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 542.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 542.2 42.1 179.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1cti56qt",
   "source_url": "https://omlx.ai/benchmarks/performance/1cti56qt",
   "tg_tokens": null,
   "tps": 42.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1cti56qt-ctx32768",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 179.6 GB at this context; duplicate traces of the same session: 2uq4tj9g, homyx3jc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 534.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 534.9 46.7 179.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1cti56qt",
   "source_url": "https://omlx.ai/benchmarks/performance/1cti56qt",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1cti56qt-ctx65536",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 180.1 GB at this context; duplicate traces of the same session: 2uq4tj9g, homyx3jc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 526.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 526.3 42.7 180.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1cti56qt",
   "source_url": "https://omlx.ai/benchmarks/performance/1cti56qt",
   "tg_tokens": null,
   "tps": 42.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1cti56qt-ctx131072",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 181.5 GB at this context; duplicate traces of the same session: 2uq4tj9g, homyx3jc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 515.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 515.3 34.6 181.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1cti56qt",
   "source_url": "https://omlx.ai/benchmarks/performance/1cti56qt",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1cti56qt-ctx199680",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 182.8 GB at this context; duplicate traces of the same session: 2uq4tj9g, homyx3jc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 505.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 195k 505.4 32.2 182.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1cti56qt",
   "source_url": "https://omlx.ai/benchmarks/performance/1cti56qt",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-t3ro6nj0-ctx4096",
   "model": "K2-Horizon-MoVA-36B-A4B",
   "notes": "fixed build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 3ift8uqm; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 252.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "K2-Horizon-MoVA-36B-A4B-oQ4e-fixed on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 252.9 21.2 22.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t3ro6nj0",
   "source_url": "https://omlx.ai/benchmarks/performance/t3ro6nj0",
   "tg_tokens": null,
   "tps": 21.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-t3ro6nj0-ctx8192",
   "model": "K2-Horizon-MoVA-36B-A4B",
   "notes": "fixed build; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: 3ift8uqm; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 233.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "K2-Horizon-MoVA-36B-A4B-oQ4e-fixed on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 8k 233.2 17.5 22.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t3ro6nj0",
   "source_url": "https://omlx.ai/benchmarks/performance/t3ro6nj0",
   "tg_tokens": null,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-mt1egb3b-ctx4096",
   "model": "KAT-Coder-V2.5-Dev",
   "notes": "oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: ssi0lc5e; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 608.5,
   "provenance": "community",
   "quant": "OptiQ 4bit",
   "quote": "KAT-Coder-V2.5-Dev-OptiQ-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 608.5 57.1 23.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mt1egb3b",
   "source_url": "https://omlx.ai/benchmarks/performance/mt1egb3b",
   "tg_tokens": null,
   "tps": 57.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-mt1egb3b-ctx8192",
   "model": "KAT-Coder-V2.5-Dev",
   "notes": "oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: ssi0lc5e; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 559.1,
   "provenance": "community",
   "quant": "OptiQ 4bit",
   "quote": "KAT-Coder-V2.5-Dev-OptiQ-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 8k 559.1 54.2 24.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mt1egb3b",
   "source_url": "https://omlx.ai/benchmarks/performance/mt1egb3b",
   "tg_tokens": null,
   "tps": 54.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1mnzmtct-ctx1024",
   "model": "Qwen3.5-0.8B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 2.0 GB at this context; duplicate traces of the same session: 3vxswwjv, acsn7iec, fldm76i1, vxlutbw0, z07gd4lz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3444.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.5-0.8B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 3,444 105.9 2.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1mnzmtct",
   "source_url": "https://omlx.ai/benchmarks/performance/1mnzmtct",
   "tg_tokens": null,
   "tps": 105.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1mnzmtct-ctx4096",
   "model": "Qwen3.5-0.8B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 2.4 GB at this context; duplicate traces of the same session: 3vxswwjv, acsn7iec, fldm76i1, vxlutbw0, z07gd4lz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 8503.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.5-0.8B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 8,503 237.6 2.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1mnzmtct",
   "source_url": "https://omlx.ai/benchmarks/performance/1mnzmtct",
   "tg_tokens": null,
   "tps": 237.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1mnzmtct-ctx32768",
   "model": "Qwen3.5-0.8B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 2.8 GB at this context; duplicate traces of the same session: 3vxswwjv, acsn7iec, fldm76i1, vxlutbw0, z07gd4lz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 17371.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.5-0.8B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 17,371 189.4 2.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1mnzmtct",
   "source_url": "https://omlx.ai/benchmarks/performance/1mnzmtct",
   "tg_tokens": null,
   "tps": 189.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1mnzmtct-ctx65536",
   "model": "Qwen3.5-0.8B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 3.3 GB at this context; duplicate traces of the same session: 3vxswwjv, acsn7iec, fldm76i1, vxlutbw0, z07gd4lz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 12581.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.5-0.8B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 12,581 148.8 3.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1mnzmtct",
   "source_url": "https://omlx.ai/benchmarks/performance/1mnzmtct",
   "tg_tokens": null,
   "tps": 148.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1mnzmtct-ctx131072",
   "model": "Qwen3.5-0.8B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.7 GB at this context; duplicate traces of the same session: 3vxswwjv, acsn7iec, fldm76i1, vxlutbw0, z07gd4lz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 7784.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.5-0.8B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 7,784 138.5 4.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1mnzmtct",
   "source_url": "https://omlx.ai/benchmarks/performance/1mnzmtct",
   "tg_tokens": null,
   "tps": 138.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1mnzmtct-ctx199680",
   "model": "Qwen3.5-0.8B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: 3vxswwjv, acsn7iec, fldm76i1, vxlutbw0, z07gd4lz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5163.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.5-0.8B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 195k 5,163 90.9 6.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1mnzmtct",
   "source_url": "https://omlx.ai/benchmarks/performance/1mnzmtct",
   "tg_tokens": null,
   "tps": 90.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-3m19mkdw-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: m3i0mhyd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 551.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.5-122B-A10B-oQ4-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 551.0 62.5 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3m19mkdw",
   "source_url": "https://omlx.ai/benchmarks/performance/3m19mkdw",
   "tg_tokens": null,
   "tps": 62.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-3m19mkdw-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.9 GB at this context; duplicate traces of the same session: m3i0mhyd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 632.1,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.5-122B-A10B-oQ4-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 632.1 70.9 73.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3m19mkdw",
   "source_url": "https://omlx.ai/benchmarks/performance/3m19mkdw",
   "tg_tokens": null,
   "tps": 70.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-cjdamunr-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 9.8 GB at this context; duplicate traces of the same session: t410je02; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 222.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 8bit; 1k 222.9 11.0 9.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cjdamunr",
   "source_url": "https://omlx.ai/benchmarks/performance/cjdamunr",
   "tg_tokens": null,
   "tps": 11.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-cjdamunr-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 9.8 GB at this context; duplicate traces of the same session: t410je02; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 218.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 8bit; 4k 218.0 10.4 9.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cjdamunr",
   "source_url": "https://omlx.ai/benchmarks/performance/cjdamunr",
   "tg_tokens": null,
   "tps": 10.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-vd5jaigc-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 9cm6v1md, xas371xz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2660.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 2,660 107.2 19.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vd5jaigc",
   "source_url": "https://omlx.ai/benchmarks/performance/vd5jaigc",
   "tg_tokens": null,
   "tps": 107.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-vd5jaigc-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: 9cm6v1md, xas371xz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3333.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 3,333 95.2 20.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vd5jaigc",
   "source_url": "https://omlx.ai/benchmarks/performance/vd5jaigc",
   "tg_tokens": null,
   "tps": 95.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-vd5jaigc-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: 9cm6v1md, xas371xz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3059.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 3,059 94.4 20.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vd5jaigc",
   "source_url": "https://omlx.ai/benchmarks/performance/vd5jaigc",
   "tg_tokens": null,
   "tps": 94.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-y2ax9g3a-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: arw9c369; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 729.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 1k 729.5 92.1 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y2ax9g3a",
   "source_url": "https://omlx.ai/benchmarks/performance/y2ax9g3a",
   "tg_tokens": null,
   "tps": 92.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-y2ax9g3a-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: arw9c369; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 762.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 762.1 89.4 21.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y2ax9g3a",
   "source_url": "https://omlx.ai/benchmarks/performance/y2ax9g3a",
   "tg_tokens": null,
   "tps": 89.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-es3tb8b0-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: vnojowrg; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 905.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 905.1 62.9 21.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/es3tb8b0",
   "source_url": "https://omlx.ai/benchmarks/performance/es3tb8b0",
   "tg_tokens": null,
   "tps": 62.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-es3tb8b0-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: vnojowrg; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 983.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 983.6 74.1 22.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/es3tb8b0",
   "source_url": "https://omlx.ai/benchmarks/performance/es3tb8b0",
   "tg_tokens": null,
   "tps": 74.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-pbg8hlno-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: fekwsp0b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 257.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 257.2 27.3 16.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pbg8hlno",
   "source_url": "https://omlx.ai/benchmarks/performance/pbg8hlno",
   "tg_tokens": null,
   "tps": 27.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-pbg8hlno-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: fekwsp0b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 259.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 259.1 27.1 19.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pbg8hlno",
   "source_url": "https://omlx.ai/benchmarks/performance/pbg8hlno",
   "tg_tokens": null,
   "tps": 27.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-phnyv1bv-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 475.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 475.7 34.2 18.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/phnyv1bv",
   "source_url": "https://omlx.ai/benchmarks/performance/phnyv1bv",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-phnyv1bv-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 425.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 425.7 43.0 19.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/phnyv1bv",
   "source_url": "https://omlx.ai/benchmarks/performance/phnyv1bv",
   "tg_tokens": null,
   "tps": 43.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-pycnq12g-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: vddewq6w; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 686.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 686.5 32.0 19.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pycnq12g",
   "source_url": "https://omlx.ai/benchmarks/performance/pycnq12g",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-pycnq12g-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: vddewq6w; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 673.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 673.3 40.6 19.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pycnq12g",
   "source_url": "https://omlx.ai/benchmarks/performance/pycnq12g",
   "tg_tokens": null,
   "tps": 40.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-z0po6a8x-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 27.7 GB; TTFT 4309 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 950.5,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-27B-oQ6 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k tokens 950.5 PP tok/s 41.8 TG tok/s 4309 TTFT (ms) 27.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z0po6a8x",
   "source_url": "https://omlx.ai/benchmarks/performance/z0po6a8x",
   "tg_tokens": null,
   "tps": 41.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-t0ydnv2i-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.2 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 342.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 4k 342.1 38.1 30.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t0ydnv2i",
   "source_url": "https://omlx.ai/benchmarks/performance/t0ydnv2i",
   "tg_tokens": null,
   "tps": 38.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-t0ydnv2i-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.4 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 16k 323.8 35.7 31.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t0ydnv2i",
   "source_url": "https://omlx.ai/benchmarks/performance/t0ydnv2i",
   "tg_tokens": null,
   "tps": 35.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-k03pw8hd-ctx1024",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 38.1 GB at this context; duplicate traces of the same session: 8takp8nt, c6j8j4c8, hf0x64ib, ijkjrsx3, jlan6nda, kkou52rn, lb5s448x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3816.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 3,816 113.5 38.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k03pw8hd",
   "source_url": "https://omlx.ai/benchmarks/performance/k03pw8hd",
   "tg_tokens": null,
   "tps": 113.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-k03pw8hd-ctx4096",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 38.8 GB at this context; duplicate traces of the same session: 8takp8nt, c6j8j4c8, hf0x64ib, ijkjrsx3, jlan6nda, kkou52rn, lb5s448x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5478.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 5,478 124.0 38.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k03pw8hd",
   "source_url": "https://omlx.ai/benchmarks/performance/k03pw8hd",
   "tg_tokens": null,
   "tps": 124.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-k03pw8hd-ctx8192",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 38.9 GB at this context; duplicate traces of the same session: 8takp8nt, c6j8j4c8, hf0x64ib, ijkjrsx3, jlan6nda, kkou52rn, lb5s448x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5595.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 5,595 133.3 38.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k03pw8hd",
   "source_url": "https://omlx.ai/benchmarks/performance/k03pw8hd",
   "tg_tokens": null,
   "tps": 133.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-k03pw8hd-ctx16384",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 39.1 GB at this context; duplicate traces of the same session: 8takp8nt, c6j8j4c8, hf0x64ib, ijkjrsx3, jlan6nda, kkou52rn, lb5s448x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5193.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 5,193 123.0 39.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k03pw8hd",
   "source_url": "https://omlx.ai/benchmarks/performance/k03pw8hd",
   "tg_tokens": null,
   "tps": 123.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-k03pw8hd-ctx32768",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 39.5 GB at this context; duplicate traces of the same session: 8takp8nt, c6j8j4c8, hf0x64ib, ijkjrsx3, jlan6nda, kkou52rn, lb5s448x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4125.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 4,125 95.9 39.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k03pw8hd",
   "source_url": "https://omlx.ai/benchmarks/performance/k03pw8hd",
   "tg_tokens": null,
   "tps": 95.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-k03pw8hd-ctx65536",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 40.2 GB at this context; duplicate traces of the same session: 8takp8nt, c6j8j4c8, hf0x64ib, ijkjrsx3, jlan6nda, kkou52rn, lb5s448x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2957.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 2,957 69.6 40.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k03pw8hd",
   "source_url": "https://omlx.ai/benchmarks/performance/k03pw8hd",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-k03pw8hd-ctx131072",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 41.7 GB at this context; duplicate traces of the same session: 8takp8nt, c6j8j4c8, hf0x64ib, ijkjrsx3, jlan6nda, kkou52rn, lb5s448x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1838.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 1,838 42.7 41.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k03pw8hd",
   "source_url": "https://omlx.ai/benchmarks/performance/k03pw8hd",
   "tg_tokens": null,
   "tps": 42.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-k03pw8hd-ctx199680",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 44.2 GB at this context; duplicate traces of the same session: 8takp8nt, c6j8j4c8, hf0x64ib, ijkjrsx3, jlan6nda, kkou52rn, lb5s448x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 997.4,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 195k 997.4 23.5 44.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k03pw8hd",
   "source_url": "https://omlx.ai/benchmarks/performance/k03pw8hd",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-gjy9a0ik-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: lunw35vl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 450.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 450.1 43.7 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gjy9a0ik",
   "source_url": "https://omlx.ai/benchmarks/performance/gjy9a0ik",
   "tg_tokens": null,
   "tps": 43.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-gjy9a0ik-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: lunw35vl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 512.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 512.6 41.4 71.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gjy9a0ik",
   "source_url": "https://omlx.ai/benchmarks/performance/gjy9a0ik",
   "tg_tokens": null,
   "tps": 41.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-oeluwoz5-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: 8jwhs3ag; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 370.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 370.7 51.5 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oeluwoz5",
   "source_url": "https://omlx.ai/benchmarks/performance/oeluwoz5",
   "tg_tokens": null,
   "tps": 51.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-oeluwoz5-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: 8jwhs3ag; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 475.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 475.7 55.9 71.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oeluwoz5",
   "source_url": "https://omlx.ai/benchmarks/performance/oeluwoz5",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-reeztgz7-ctx1024",
   "model": "gpt-oss-120b-Sonnet-Reasoning-Distilled",
   "notes": "oMLX community benchmark trace; peak memory 59.5 GB at this context; duplicate traces of the same session: rjf6zsct; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 617.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gpt-oss-120b-Sonnet-Reasoning-Distilled on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 617.7 77.7 59.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/reeztgz7",
   "source_url": "https://omlx.ai/benchmarks/performance/reeztgz7",
   "tg_tokens": null,
   "tps": 77.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-reeztgz7-ctx4096",
   "model": "gpt-oss-120b-Sonnet-Reasoning-Distilled",
   "notes": "oMLX community benchmark trace; peak memory 59.6 GB at this context; duplicate traces of the same session: rjf6zsct; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 593.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gpt-oss-120b-Sonnet-Reasoning-Distilled on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 593.6 76.2 59.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/reeztgz7",
   "source_url": "https://omlx.ai/benchmarks/performance/reeztgz7",
   "tg_tokens": null,
   "tps": 76.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-jehe7f3h-ctx4096",
   "model": "gemma-4-26B-A4B-it-qat",
   "notes": "unsloth; oMLX community benchmark trace, single run at 4k context; peak memory 14.7 GB; TTFT 13884 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 295.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "unsloth-gemma-4-26B-A4B-it-qat-oQ4 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k tokens 295.0 PP tok/s 27.8 TG tok/s 13884 TTFT (ms) 14.7 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jehe7f3h",
   "source_url": "https://omlx.ai/benchmarks/performance/jehe7f3h",
   "tg_tokens": null,
   "tps": 27.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-drwwwgc2-ctx1024",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 177.7 GB at this context; duplicate traces of the same session: 3k8edktb, i3eqsxks, zjjmq7bn, zl8jtm36; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 519.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 519.5 43.4 177.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/drwwwgc2",
   "source_url": "https://omlx.ai/benchmarks/performance/drwwwgc2",
   "tg_tokens": null,
   "tps": 43.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-drwwwgc2-ctx4096",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 178.8 GB at this context; duplicate traces of the same session: 3k8edktb, i3eqsxks, zjjmq7bn, zl8jtm36; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 591.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 591.4 35.4 178.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/drwwwgc2",
   "source_url": "https://omlx.ai/benchmarks/performance/drwwwgc2",
   "tg_tokens": null,
   "tps": 35.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-drwwwgc2-ctx8192",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 179.1 GB at this context; duplicate traces of the same session: 3k8edktb, i3eqsxks, zjjmq7bn, zl8jtm36; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 553.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 553.2 45.4 179.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/drwwwgc2",
   "source_url": "https://omlx.ai/benchmarks/performance/drwwwgc2",
   "tg_tokens": null,
   "tps": 45.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-drwwwgc2-ctx16384",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 179.3 GB at this context; duplicate traces of the same session: 3k8edktb, i3eqsxks, zjjmq7bn, zl8jtm36; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 542.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 542.2 42.1 179.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/drwwwgc2",
   "source_url": "https://omlx.ai/benchmarks/performance/drwwwgc2",
   "tg_tokens": null,
   "tps": 42.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-drwwwgc2-ctx32768",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 179.6 GB at this context; duplicate traces of the same session: 3k8edktb, i3eqsxks, zjjmq7bn, zl8jtm36; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 534.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 534.9 46.7 179.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/drwwwgc2",
   "source_url": "https://omlx.ai/benchmarks/performance/drwwwgc2",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-drwwwgc2-ctx65536",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 180.1 GB at this context; duplicate traces of the same session: 3k8edktb, i3eqsxks, zjjmq7bn, zl8jtm36; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 526.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 526.3 42.7 180.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/drwwwgc2",
   "source_url": "https://omlx.ai/benchmarks/performance/drwwwgc2",
   "tg_tokens": null,
   "tps": 42.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-drwwwgc2-ctx131072",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 181.5 GB at this context; duplicate traces of the same session: 3k8edktb, i3eqsxks, zjjmq7bn, zl8jtm36; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 515.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 515.3 34.6 181.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/drwwwgc2",
   "source_url": "https://omlx.ai/benchmarks/performance/drwwwgc2",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-drwwwgc2-ctx199680",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 182.8 GB at this context; duplicate traces of the same session: 3k8edktb, i3eqsxks, zjjmq7bn, zl8jtm36; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 505.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 195k 505.4 32.2 182.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/drwwwgc2",
   "source_url": "https://omlx.ai/benchmarks/performance/drwwwgc2",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-yzemlrqp-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.7 GB at this context; duplicate traces of the same session: 6q6jw1c1, aammf1t2, li5ok9py, p66xg60r, p9w5somu, pqmmq34f, yad2v1hr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 531.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 531.7 40.5 173.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yzemlrqp",
   "source_url": "https://omlx.ai/benchmarks/performance/yzemlrqp",
   "tg_tokens": null,
   "tps": 40.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-yzemlrqp-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.8 GB at this context; duplicate traces of the same session: 6q6jw1c1, aammf1t2, li5ok9py, p66xg60r, p9w5somu, pqmmq34f, yad2v1hr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 607.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 607.7 38.5 174.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yzemlrqp",
   "source_url": "https://omlx.ai/benchmarks/performance/yzemlrqp",
   "tg_tokens": null,
   "tps": 38.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-yzemlrqp-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.4 GB at this context; duplicate traces of the same session: 6q6jw1c1, aammf1t2, li5ok9py, p66xg60r, p9w5somu, pqmmq34f, yad2v1hr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 562.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 562.4 38.9 175.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yzemlrqp",
   "source_url": "https://omlx.ai/benchmarks/performance/yzemlrqp",
   "tg_tokens": null,
   "tps": 38.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-yzemlrqp-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.5 GB at this context; duplicate traces of the same session: 6q6jw1c1, aammf1t2, li5ok9py, p66xg60r, p9w5somu, pqmmq34f, yad2v1hr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 551.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 551.5 38.4 175.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yzemlrqp",
   "source_url": "https://omlx.ai/benchmarks/performance/yzemlrqp",
   "tg_tokens": null,
   "tps": 38.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-yzemlrqp-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.7 GB at this context; duplicate traces of the same session: 6q6jw1c1, aammf1t2, li5ok9py, p66xg60r, p9w5somu, pqmmq34f, yad2v1hr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 543.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 543.7 38.0 175.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yzemlrqp",
   "source_url": "https://omlx.ai/benchmarks/performance/yzemlrqp",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-yzemlrqp-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 176.0 GB at this context; duplicate traces of the same session: 6q6jw1c1, aammf1t2, li5ok9py, p66xg60r, p9w5somu, pqmmq34f, yad2v1hr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 535.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 535.2 37.0 176.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yzemlrqp",
   "source_url": "https://omlx.ai/benchmarks/performance/yzemlrqp",
   "tg_tokens": null,
   "tps": 37.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-yzemlrqp-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 176.8 GB at this context; duplicate traces of the same session: 6q6jw1c1, aammf1t2, li5ok9py, p66xg60r, p9w5somu, pqmmq34f, yad2v1hr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 523.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 523.8 35.5 176.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yzemlrqp",
   "source_url": "https://omlx.ai/benchmarks/performance/yzemlrqp",
   "tg_tokens": null,
   "tps": 35.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-yzemlrqp-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.5 GB at this context; duplicate traces of the same session: 6q6jw1c1, aammf1t2, li5ok9py, p66xg60r, p9w5somu, pqmmq34f, yad2v1hr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 513.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 195k 513.5 33.7 177.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yzemlrqp",
   "source_url": "https://omlx.ai/benchmarks/performance/yzemlrqp",
   "tg_tokens": null,
   "tps": 33.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-g0e5cfa9-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 244.6 GB at this context; duplicate traces of the same session: patjfund; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 485.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 4k 485.9 34.4 244.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g0e5cfa9",
   "source_url": "https://omlx.ai/benchmarks/performance/g0e5cfa9",
   "tg_tokens": null,
   "tps": 34.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-g0e5cfa9-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 245.2 GB at this context; duplicate traces of the same session: patjfund; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 494.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 16k 494.9 34.1 245.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g0e5cfa9",
   "source_url": "https://omlx.ai/benchmarks/performance/g0e5cfa9",
   "tg_tokens": null,
   "tps": 34.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-xl9jnbql-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: enmwv4ry; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 555.7,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.5-122B-A10B-oQ4-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 555.7 47.1 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xl9jnbql",
   "source_url": "https://omlx.ai/benchmarks/performance/xl9jnbql",
   "tg_tokens": null,
   "tps": 47.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-xl9jnbql-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.9 GB at this context; duplicate traces of the same session: enmwv4ry; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 637.7,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.5-122B-A10B-oQ4-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 637.7 64.2 73.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xl9jnbql",
   "source_url": "https://omlx.ai/benchmarks/performance/xl9jnbql",
   "tg_tokens": null,
   "tps": 64.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-rqsa9jzb-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: kra87v5a; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 422.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 422.7 50.3 20.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rqsa9jzb",
   "source_url": "https://omlx.ai/benchmarks/performance/rqsa9jzb",
   "tg_tokens": null,
   "tps": 50.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-rqsa9jzb-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: kra87v5a; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 343.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k 343.0 36.4 20.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rqsa9jzb",
   "source_url": "https://omlx.ai/benchmarks/performance/rqsa9jzb",
   "tg_tokens": null,
   "tps": 36.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-t98rlaio-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 9rctpwph, htbcni6v, njbzymat; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1032.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 1,032 93.3 22.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t98rlaio",
   "source_url": "https://omlx.ai/benchmarks/performance/t98rlaio",
   "tg_tokens": null,
   "tps": 93.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-t98rlaio-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: 9rctpwph, htbcni6v, njbzymat; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 874.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 874.8 103.3 22.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t98rlaio",
   "source_url": "https://omlx.ai/benchmarks/performance/t98rlaio",
   "tg_tokens": null,
   "tps": 103.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-t98rlaio-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: 9rctpwph, htbcni6v, njbzymat; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 515.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 515.5 61.2 24.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t98rlaio",
   "source_url": "https://omlx.ai/benchmarks/performance/t98rlaio",
   "tg_tokens": null,
   "tps": 61.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-t98rlaio-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: 9rctpwph, htbcni6v, njbzymat; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 334.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 128k 334.7 47.3 26.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t98rlaio",
   "source_url": "https://omlx.ai/benchmarks/performance/t98rlaio",
   "tg_tokens": null,
   "tps": 47.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-9m4zdm9c-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: dh3yhm4i; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 271.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M2 Ultra (76c); M2 Ultra (76c) \u00b7 64 GB \u00b7 4bit; 1k 271.8 9.8 17.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9m4zdm9c",
   "source_url": "https://omlx.ai/benchmarks/performance/9m4zdm9c",
   "tg_tokens": null,
   "tps": 9.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-9m4zdm9c-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: dh3yhm4i; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 258.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M2 Ultra (76c); M2 Ultra (76c) \u00b7 64 GB \u00b7 4bit; 4k 258.3 13.3 18.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9m4zdm9c",
   "source_url": "https://omlx.ai/benchmarks/performance/9m4zdm9c",
   "tg_tokens": null,
   "tps": 13.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-5saqe1bx-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: axtz0sxb; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 278.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M2 Ultra (76c); M2 Ultra (76c) \u00b7 64 GB \u00b7 4bit; 1k 278.6 9.5 17.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5saqe1bx",
   "source_url": "https://omlx.ai/benchmarks/performance/5saqe1bx",
   "tg_tokens": null,
   "tps": 9.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-5saqe1bx-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: axtz0sxb; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 254.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M2 Ultra (76c); M2 Ultra (76c) \u00b7 64 GB \u00b7 4bit; 4k 254.4 8.9 18.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5saqe1bx",
   "source_url": "https://omlx.ai/benchmarks/performance/5saqe1bx",
   "tg_tokens": null,
   "tps": 8.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pzzjqxm1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 27.7 GB; TTFT 4309 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 950.6,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-27B-oQ6 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k tokens 950.6 PP tok/s 34.0 TG tok/s 4309 TTFT (ms) 27.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pzzjqxm1",
   "source_url": "https://omlx.ai/benchmarks/performance/pzzjqxm1",
   "tg_tokens": null,
   "tps": 34.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-knbpa2t1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.2 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 342.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 4k 342.1 38.1 30.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/knbpa2t1",
   "source_url": "https://omlx.ai/benchmarks/performance/knbpa2t1",
   "tg_tokens": null,
   "tps": 38.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-knbpa2t1-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.4 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 16k 323.8 35.7 31.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/knbpa2t1",
   "source_url": "https://omlx.ai/benchmarks/performance/knbpa2t1",
   "tg_tokens": null,
   "tps": 35.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-n1hbqw1m-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: kntqgaf2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 443.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 443.1 54.1 70.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n1hbqw1m",
   "source_url": "https://omlx.ai/benchmarks/performance/n1hbqw1m",
   "tg_tokens": null,
   "tps": 54.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-n1hbqw1m-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: kntqgaf2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 503.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 503.0 47.3 71.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n1hbqw1m",
   "source_url": "https://omlx.ai/benchmarks/performance/n1hbqw1m",
   "tg_tokens": null,
   "tps": 47.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-zwtfmuju-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 71.5 GB; TTFT 12496 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 655.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k tokens 655.5 PP tok/s 44.6 TG tok/s 12496 TTFT (ms) 71.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zwtfmuju",
   "source_url": "https://omlx.ai/benchmarks/performance/zwtfmuju",
   "tg_tokens": null,
   "tps": 44.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-1nfc9rye-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: 03ssffc4, 37e1p4s3, 74wxvgzc, nl8u7i4i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 457.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 457.9 53.6 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1nfc9rye",
   "source_url": "https://omlx.ai/benchmarks/performance/1nfc9rye",
   "tg_tokens": null,
   "tps": 53.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-1nfc9rye-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: 03ssffc4, 37e1p4s3, 74wxvgzc, nl8u7i4i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 517.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 517.6 47.6 71.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1nfc9rye",
   "source_url": "https://omlx.ai/benchmarks/performance/1nfc9rye",
   "tg_tokens": null,
   "tps": 47.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-1nfc9rye-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: 03ssffc4, 37e1p4s3, 74wxvgzc, nl8u7i4i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 492.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 492.5 40.0 71.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1nfc9rye",
   "source_url": "https://omlx.ai/benchmarks/performance/1nfc9rye",
   "tg_tokens": null,
   "tps": 40.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-1nfc9rye-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 03ssffc4, 37e1p4s3, 74wxvgzc, nl8u7i4i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 471.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 471.2 36.9 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1nfc9rye",
   "source_url": "https://omlx.ai/benchmarks/performance/1nfc9rye",
   "tg_tokens": null,
   "tps": 36.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-1nfc9rye-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.4 GB at this context; duplicate traces of the same session: 03ssffc4, 37e1p4s3, 74wxvgzc, nl8u7i4i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 499.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 499.9 32.1 72.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1nfc9rye",
   "source_url": "https://omlx.ai/benchmarks/performance/1nfc9rye",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-t7jnaaob-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.7 GB; TTFT 2012 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 508.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 508.8 PP tok/s 54.3 TG tok/s 2012 TTFT (ms) 70.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t7jnaaob",
   "source_url": "https://omlx.ai/benchmarks/performance/t7jnaaob",
   "tg_tokens": null,
   "tps": 54.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5cw8rhmy-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1158.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,158 51.5 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5cw8rhmy",
   "source_url": "https://omlx.ai/benchmarks/performance/5cw8rhmy",
   "tg_tokens": null,
   "tps": 51.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5cw8rhmy-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2033.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,033 66.8 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5cw8rhmy",
   "source_url": "https://omlx.ai/benchmarks/performance/5cw8rhmy",
   "tg_tokens": null,
   "tps": 66.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5cw8rhmy-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2574.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,574 74.4 71.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5cw8rhmy",
   "source_url": "https://omlx.ai/benchmarks/performance/5cw8rhmy",
   "tg_tokens": null,
   "tps": 74.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5cw8rhmy-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2561.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,561 73.9 73.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5cw8rhmy",
   "source_url": "https://omlx.ai/benchmarks/performance/5cw8rhmy",
   "tg_tokens": null,
   "tps": 73.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-n0vsvjsg-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 14.5 GB at this context; duplicate traces of the same session: 2znghz94; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 661.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 661.2 63.4 14.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n0vsvjsg",
   "source_url": "https://omlx.ai/benchmarks/performance/n0vsvjsg",
   "tg_tokens": null,
   "tps": 63.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-n0vsvjsg-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: 2znghz94; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 628.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 628.0 61.5 14.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n0vsvjsg",
   "source_url": "https://omlx.ai/benchmarks/performance/n0vsvjsg",
   "tg_tokens": null,
   "tps": 61.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-uclqsv9g-ctx4096",
   "model": "gpt-oss-120b-Sonnet-Reasoning-Distilled",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 59.6 GB; TTFT 5319 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 770.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gpt-oss-120b-Sonnet-Reasoning-Distilled on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 770.1 PP tok/s 78.8 TG tok/s 5319 TTFT (ms) 59.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uclqsv9g",
   "source_url": "https://omlx.ai/benchmarks/performance/uclqsv9g",
   "tg_tokens": null,
   "tps": 78.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-48ohnpvj-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; rapid-mlx build; oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: o7trw73f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 985.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "rapid-mlx--Qwen3.8-27B-4bit-MTP-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 985.0 32.1 17.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/48ohnpvj",
   "source_url": "https://omlx.ai/benchmarks/performance/48ohnpvj",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-48ohnpvj-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; rapid-mlx build; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: o7trw73f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 981.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "rapid-mlx--Qwen3.8-27B-4bit-MTP-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 981.3 29.2 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/48ohnpvj",
   "source_url": "https://omlx.ai/benchmarks/performance/48ohnpvj",
   "tg_tokens": null,
   "tps": 29.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-284z0wod-ctx32768",
   "model": "gemma-4-26B-A4B-it-qat",
   "notes": "unsloth; oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: 51mf0rvm; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.9,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "unsloth-gemma-4-26B-A4B-it-qat-oQ4 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 32k 224.9 19.3 16.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/284z0wod",
   "source_url": "https://omlx.ai/benchmarks/performance/284z0wod",
   "tg_tokens": null,
   "tps": 19.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-284z0wod-ctx65536",
   "model": "gemma-4-26B-A4B-it-qat",
   "notes": "unsloth; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: 51mf0rvm; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 146.9,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "unsloth-gemma-4-26B-A4B-it-qat-oQ4 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 64k 146.9 13.1 17.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/284z0wod",
   "source_url": "https://omlx.ai/benchmarks/performance/284z0wod",
   "tg_tokens": null,
   "tps": 13.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-tuu4saos-ctx1024",
   "model": "Qwen3.5-4B-Uncensored",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 3.3 GB; TTFT 2423 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 422.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-Uncensored-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k tokens 422.7 PP tok/s 36.6 TG tok/s 2423 TTFT (ms) 3.3 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tuu4saos",
   "source_url": "https://omlx.ai/benchmarks/performance/tuu4saos",
   "tg_tokens": null,
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-il3z3e11-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.3 GB; TTFT 2593 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 395.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 395.0 PP tok/s 47.2 TG tok/s 2593 TTFT (ms) 19.3 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/il3z3e11",
   "source_url": "https://omlx.ai/benchmarks/performance/il3z3e11",
   "tg_tokens": null,
   "tps": 47.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-gtvuv5vq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: 2pi62npa, jf7t2ao8, mkvdsvv1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 488.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 488.8 33.0 17.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gtvuv5vq",
   "source_url": "https://omlx.ai/benchmarks/performance/gtvuv5vq",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-gtvuv5vq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: 2pi62npa, jf7t2ao8, mkvdsvv1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 498.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 498.1 34.0 18.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gtvuv5vq",
   "source_url": "https://omlx.ai/benchmarks/performance/gtvuv5vq",
   "tg_tokens": null,
   "tps": 34.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-gtvuv5vq-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: 2pi62npa, jf7t2ao8, mkvdsvv1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 455.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 455.7 30.3 19.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gtvuv5vq",
   "source_url": "https://omlx.ai/benchmarks/performance/gtvuv5vq",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-gtvuv5vq-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 2pi62npa, jf7t2ao8, mkvdsvv1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 431.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 431.0 35.2 20.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gtvuv5vq",
   "source_url": "https://omlx.ai/benchmarks/performance/gtvuv5vq",
   "tg_tokens": null,
   "tps": 35.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-g8jkzgb8-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: 45ohji2p; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 468.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 468.7 24.9 17.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g8jkzgb8",
   "source_url": "https://omlx.ai/benchmarks/performance/g8jkzgb8",
   "tg_tokens": null,
   "tps": 24.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-g8jkzgb8-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: 45ohji2p; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 491.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 491.4 33.8 18.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g8jkzgb8",
   "source_url": "https://omlx.ai/benchmarks/performance/g8jkzgb8",
   "tg_tokens": null,
   "tps": 33.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-uyqaae0k-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 35.4 GB at this context; duplicate traces of the same session: wt96lhvj; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 132.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 8bit; 32k 132.6 15.9 35.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uyqaae0k",
   "source_url": "https://omlx.ai/benchmarks/performance/uyqaae0k",
   "tg_tokens": null,
   "tps": 15.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-uyqaae0k-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 45.5 GB at this context; duplicate traces of the same session: wt96lhvj; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 63.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 8bit; 195k 63.4 9.8 45.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uyqaae0k",
   "source_url": "https://omlx.ai/benchmarks/performance/uyqaae0k",
   "tg_tokens": null,
   "tps": 9.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-phngfg5i-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 0a14v9ed; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 101.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 101.5 21.7 23.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/phngfg5i",
   "source_url": "https://omlx.ai/benchmarks/performance/phngfg5i",
   "tg_tokens": null,
   "tps": 21.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-phngfg5i-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: 0a14v9ed; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 103.5 30.3 24.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/phngfg5i",
   "source_url": "https://omlx.ai/benchmarks/performance/phngfg5i",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-pi4pkrbp-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: ndp15jtr; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 101.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 101.7 19.8 23.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pi4pkrbp",
   "source_url": "https://omlx.ai/benchmarks/performance/pi4pkrbp",
   "tg_tokens": null,
   "tps": 19.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-pi4pkrbp-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: ndp15jtr; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 94.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 94.7 25.2 24.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pi4pkrbp",
   "source_url": "https://omlx.ai/benchmarks/performance/pi4pkrbp",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-f8npzl8g-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: ds9frtxl; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 105.4 12.3 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f8npzl8g",
   "source_url": "https://omlx.ai/benchmarks/performance/f8npzl8g",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-f8npzl8g-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: ds9frtxl; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 105.6 12.2 21.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f8npzl8g",
   "source_url": "https://omlx.ai/benchmarks/performance/f8npzl8g",
   "tg_tokens": null,
   "tps": 12.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5w9ozh3a-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 38p3zbzu, gbix4fvr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 325.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 325.7 82.0 21.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5w9ozh3a",
   "source_url": "https://omlx.ai/benchmarks/performance/5w9ozh3a",
   "tg_tokens": null,
   "tps": 82.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5w9ozh3a-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: 38p3zbzu, gbix4fvr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 430.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 430.5 102.2 24.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5w9ozh3a",
   "source_url": "https://omlx.ai/benchmarks/performance/5w9ozh3a",
   "tg_tokens": null,
   "tps": 102.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5w9ozh3a-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: 38p3zbzu, gbix4fvr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 450.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 450.6 45.4 25.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5w9ozh3a",
   "source_url": "https://omlx.ai/benchmarks/performance/5w9ozh3a",
   "tg_tokens": null,
   "tps": 45.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5fgynwzh-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 25.1 GB; TTFT 4292 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 954.4,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-27B-oQ6 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k tokens 954.4 PP tok/s 49.5 TG tok/s 4292 TTFT (ms) 25.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5fgynwzh",
   "source_url": "https://omlx.ai/benchmarks/performance/5fgynwzh",
   "tg_tokens": null,
   "tps": 49.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-813t9bah-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 27.7 GB; TTFT 4312 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 949.9,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-27B-oQ6 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k tokens 949.9 PP tok/s 39.2 TG tok/s 4312 TTFT (ms) 27.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/813t9bah",
   "source_url": "https://omlx.ai/benchmarks/performance/813t9bah",
   "tg_tokens": null,
   "tps": 39.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-d9yh8drp-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 28.5 GB; TTFT 17954 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 912.6,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-27B-oQ6 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k tokens 912.6 PP tok/s 60.7 TG tok/s 17954 TTFT (ms) 28.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d9yh8drp",
   "source_url": "https://omlx.ai/benchmarks/performance/d9yh8drp",
   "tg_tokens": null,
   "tps": 60.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-psf64itg-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 28.6 GB; TTFT 17956 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 912.5,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-27B-oQ6 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k tokens 912.5 PP tok/s 59.3 TG tok/s 17956 TTFT (ms) 28.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/psf64itg",
   "source_url": "https://omlx.ai/benchmarks/performance/psf64itg",
   "tg_tokens": null,
   "tps": 59.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wc4tjwug-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 8k context; peak memory 28.0 GB; TTFT 8711 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 940.4,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-27B-oQ6 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 8k tokens 940.4 PP tok/s 39.3 TG tok/s 8711 TTFT (ms) 28.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wc4tjwug",
   "source_url": "https://omlx.ai/benchmarks/performance/wc4tjwug",
   "tg_tokens": null,
   "tps": 39.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-pre3ka8z-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.2 GB at this context; duplicate traces of the same session: dwrfqz7a; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 160.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 8bit; 4k 160.0 18.1 32.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pre3ka8z",
   "source_url": "https://omlx.ai/benchmarks/performance/pre3ka8z",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-pre3ka8z-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.2 GB at this context; duplicate traces of the same session: dwrfqz7a; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 146.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 8bit; 16k 146.5 16.5 34.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pre3ka8z",
   "source_url": "https://omlx.ai/benchmarks/performance/pre3ka8z",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-jq06uj2q-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: 7sx1p6dz, yixsediq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 572.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 572.5 46.2 71.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jq06uj2q",
   "source_url": "https://omlx.ai/benchmarks/performance/jq06uj2q",
   "tg_tokens": null,
   "tps": 46.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-jq06uj2q-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: 7sx1p6dz, yixsediq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 558.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 558.5 47.2 71.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jq06uj2q",
   "source_url": "https://omlx.ai/benchmarks/performance/jq06uj2q",
   "tg_tokens": null,
   "tps": 47.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-jq06uj2q-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.4 GB at this context; duplicate traces of the same session: 7sx1p6dz, yixsediq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 554.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 554.9 43.9 72.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jq06uj2q",
   "source_url": "https://omlx.ai/benchmarks/performance/jq06uj2q",
   "tg_tokens": null,
   "tps": 43.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qg59iv6i-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: gp8n3mpt, tyl6by9o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1158.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,158 51.5 70.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qg59iv6i",
   "source_url": "https://omlx.ai/benchmarks/performance/qg59iv6i",
   "tg_tokens": null,
   "tps": 51.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qg59iv6i-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: gp8n3mpt, tyl6by9o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2033.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,033 66.8 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qg59iv6i",
   "source_url": "https://omlx.ai/benchmarks/performance/qg59iv6i",
   "tg_tokens": null,
   "tps": 66.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qg59iv6i-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: gp8n3mpt, tyl6by9o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2574.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,574 74.4 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qg59iv6i",
   "source_url": "https://omlx.ai/benchmarks/performance/qg59iv6i",
   "tg_tokens": null,
   "tps": 74.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qg59iv6i-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: gp8n3mpt, tyl6by9o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2561.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,561 73.9 73.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qg59iv6i",
   "source_url": "https://omlx.ai/benchmarks/performance/qg59iv6i",
   "tg_tokens": null,
   "tps": 73.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-g6udd8ow-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: nt2bgi93; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 259.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 259.7 37.7 21.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g6udd8ow",
   "source_url": "https://omlx.ai/benchmarks/performance/g6udd8ow",
   "tg_tokens": null,
   "tps": 37.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-g6udd8ow-ctx16384",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: nt2bgi93; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 246.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 16k 246.6 32.4 23.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g6udd8ow",
   "source_url": "https://omlx.ai/benchmarks/performance/g6udd8ow",
   "tg_tokens": null,
   "tps": 32.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-v4lxnpjv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: tn67o4i6; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 194.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ4e on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 194.5 15.5 18.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v4lxnpjv",
   "source_url": "https://omlx.ai/benchmarks/performance/v4lxnpjv",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-v4lxnpjv-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: tn67o4i6; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 155.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ4e on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 16k 155.2 17.1 19.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v4lxnpjv",
   "source_url": "https://omlx.ai/benchmarks/performance/v4lxnpjv",
   "tg_tokens": null,
   "tps": 17.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-4bt0dk5k-ctx65536",
   "model": "Ornith-1.5-35B-A3B-BigBang",
   "notes": "KaedeTai MTP build; oMLX community benchmark trace, single run at 64k context; peak memory 27.0 GB; TTFT 116116 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 564.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "KaedeTai--Ornith-1.5-35B-A3B-BigBang-MTP-mlx-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 64k tokens 564.4 PP tok/s 60.5 TG tok/s 116116 TTFT (ms) 27.0 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4bt0dk5k",
   "source_url": "https://omlx.ai/benchmarks/performance/4bt0dk5k",
   "tg_tokens": null,
   "tps": 60.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-zoanyv7y-ctx4096",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: bymwlino, k5gta3k7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1027.0,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Qwen3.8-27B-Abliterated-MLX-MXFP4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,027 34.9 16.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zoanyv7y",
   "source_url": "https://omlx.ai/benchmarks/performance/zoanyv7y",
   "tg_tokens": null,
   "tps": 34.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-zoanyv7y-ctx16384",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: bymwlino, k5gta3k7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 896.7,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Qwen3.8-27B-Abliterated-MLX-MXFP4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 896.7 32.6 17.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zoanyv7y",
   "source_url": "https://omlx.ai/benchmarks/performance/zoanyv7y",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-zoanyv7y-ctx131072",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: bymwlino, k5gta3k7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 546.5,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Qwen3.8-27B-Abliterated-MLX-MXFP4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 128k 546.5 21.9 24.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zoanyv7y",
   "source_url": "https://omlx.ai/benchmarks/performance/zoanyv7y",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-e7dyx4rt-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: cjq2ngzl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 246.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 246.4 56.7 21.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e7dyx4rt",
   "source_url": "https://omlx.ai/benchmarks/performance/e7dyx4rt",
   "tg_tokens": null,
   "tps": 56.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-e7dyx4rt-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: cjq2ngzl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 330.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 330.3 51.0 25.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e7dyx4rt",
   "source_url": "https://omlx.ai/benchmarks/performance/e7dyx4rt",
   "tg_tokens": null,
   "tps": 51.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-e18nwd6n-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: dwdy2l7v; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 248.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 248.2 81.8 21.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e18nwd6n",
   "source_url": "https://omlx.ai/benchmarks/performance/e18nwd6n",
   "tg_tokens": null,
   "tps": 81.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-e18nwd6n-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: dwdy2l7v; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 332.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 332.7 82.4 25.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e18nwd6n",
   "source_url": "https://omlx.ai/benchmarks/performance/e18nwd6n",
   "tg_tokens": null,
   "tps": 82.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ip9qllqo-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "SAGE-PATHS MTP build; oMLX community benchmark trace, single run at 16k context; peak memory 29.7 GB; TTFT 41381 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 395.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-SAGE-PATHS-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 16k tokens 395.9 PP tok/s 22.5 TG tok/s 41381 TTFT (ms) 29.7 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ip9qllqo",
   "source_url": "https://omlx.ai/benchmarks/performance/ip9qllqo",
   "tg_tokens": null,
   "tps": 22.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xa57nzid-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "SAGE-PATHS MTP build; oMLX community benchmark trace, single run at 16k context; peak memory 29.7 GB; TTFT 41375 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 396.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-SAGE-PATHS-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 16k tokens 396.0 PP tok/s 36.4 TG tok/s 41375 TTFT (ms) 29.7 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xa57nzid",
   "source_url": "https://omlx.ai/benchmarks/performance/xa57nzid",
   "tg_tokens": null,
   "tps": 36.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-f41o1q99-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: f6r9nryj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 462.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 462.8 52.0 70.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f41o1q99",
   "source_url": "https://omlx.ai/benchmarks/performance/f41o1q99",
   "tg_tokens": null,
   "tps": 52.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-f41o1q99-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: f6r9nryj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 549.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 549.0 47.2 71.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f41o1q99",
   "source_url": "https://omlx.ai/benchmarks/performance/f41o1q99",
   "tg_tokens": null,
   "tps": 47.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-lktfb5oj-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: ibsn3pcw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 471.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 471.4 50.1 70.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lktfb5oj",
   "source_url": "https://omlx.ai/benchmarks/performance/lktfb5oj",
   "tg_tokens": null,
   "tps": 50.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-lktfb5oj-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: ibsn3pcw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 536.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 536.8 39.0 71.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lktfb5oj",
   "source_url": "https://omlx.ai/benchmarks/performance/lktfb5oj",
   "tg_tokens": null,
   "tps": 39.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-v06z4f32-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 5ua328y9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 458.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 458.7 52.5 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v06z4f32",
   "source_url": "https://omlx.ai/benchmarks/performance/v06z4f32",
   "tg_tokens": null,
   "tps": 52.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-v06z4f32-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: 5ua328y9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 495.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 495.8 38.0 71.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v06z4f32",
   "source_url": "https://omlx.ai/benchmarks/performance/v06z4f32",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-ylvnpc7y-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.8 GB; TTFT 2938 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 348.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 348.6 PP tok/s 30.4 TG tok/s 2938 TTFT (ms) 70.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ylvnpc7y",
   "source_url": "https://omlx.ai/benchmarks/performance/ylvnpc7y",
   "tg_tokens": null,
   "tps": 30.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-h06d2sc0-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 78.3 GB; TTFT 58520 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2240.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k tokens 2,240 PP tok/s 45.6 TG tok/s 58520 TTFT (ms) 78.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h06d2sc0",
   "source_url": "https://omlx.ai/benchmarks/performance/h06d2sc0",
   "tg_tokens": null,
   "tps": 45.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-imfaejzw-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 78.4 GB; TTFT 68181 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1922.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k tokens 1,922 PP tok/s 43.4 TG tok/s 68181 TTFT (ms) 78.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/imfaejzw",
   "source_url": "https://omlx.ai/benchmarks/performance/imfaejzw",
   "tg_tokens": null,
   "tps": 43.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-hgnyx10o-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "fp16 MTP build; oMLX community benchmark trace, single run at 4k context; peak memory 38.5 GB; TTFT 20876 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 196.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen:Qwen3.8-27B-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k tokens 196.2 PP tok/s 17.4 TG tok/s 20876 TTFT (ms) 38.5 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hgnyx10o",
   "source_url": "https://omlx.ai/benchmarks/performance/hgnyx10o",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-zogfaomq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace, single run at 4k context; peak memory 20.5 GB; TTFT 39372 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ4e on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k tokens 104.0 PP tok/s 17.7 TG tok/s 39372 TTFT (ms) 20.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zogfaomq",
   "source_url": "https://omlx.ai/benchmarks/performance/zogfaomq",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-8a0mxzwr-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace, single run at 16k context; peak memory 20.6 GB; TTFT 136767 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 119.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ4e on M2 Max (38c); M2 Max (38c) \u00b7 32 GB \u00b7 4bit; 16k tokens 119.8 PP tok/s 19.7 TG tok/s 136767 TTFT (ms) 20.6 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8a0mxzwr",
   "source_url": "https://omlx.ai/benchmarks/performance/8a0mxzwr",
   "tg_tokens": null,
   "tps": 19.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ikugi3ji-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace, single run at 16k context; peak memory 18.8 GB; TTFT 132209 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 123.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ4e on M2 Max (38c); M2 Max (38c) \u00b7 32 GB \u00b7 4bit; 16k tokens 123.9 PP tok/s 18.1 TG tok/s 132209 TTFT (ms) 18.8 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ikugi3ji",
   "source_url": "https://omlx.ai/benchmarks/performance/ikugi3ji",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-icv6a2r3-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace, single run at 64k context; peak memory 22.4 GB; TTFT 451129 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 145.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ4e on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 64k tokens 145.3 PP tok/s 13.4 TG tok/s 451129 TTFT (ms) 22.4 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/icv6a2r3",
   "source_url": "https://omlx.ai/benchmarks/performance/icv6a2r3",
   "tg_tokens": null,
   "tps": 13.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-12nb2npf-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: f9tv2zeu; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 183.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ4e on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 16k 183.4 18.0 19.5 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/12nb2npf",
   "source_url": "https://omlx.ai/benchmarks/performance/12nb2npf",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-12nb2npf-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: f9tv2zeu; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 162.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ4e on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 32k 162.3 25.0 20.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/12nb2npf",
   "source_url": "https://omlx.ai/benchmarks/performance/12nb2npf",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-6hrh4a3e-ctx4096",
   "model": "KAT-Coder-V2.5-Dev-VL",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: q0qpn3om; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 582.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "KAT-Coder-V2.5-Dev-VL-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 582.8 61.3 23.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6hrh4a3e",
   "source_url": "https://omlx.ai/benchmarks/performance/6hrh4a3e",
   "tg_tokens": null,
   "tps": 61.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-6hrh4a3e-ctx8192",
   "model": "KAT-Coder-V2.5-Dev-VL",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: q0qpn3om; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 532.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "KAT-Coder-V2.5-Dev-VL-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 8k 532.2 56.1 24.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6hrh4a3e",
   "source_url": "https://omlx.ai/benchmarks/performance/6hrh4a3e",
   "tg_tokens": null,
   "tps": 56.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-74mhkji0-ctx1024",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 12.6 GB at this context; duplicate traces of the same session: 7chu59nd, 8jilsg24, 9m3qk3on, ckihx6hv, mx0k00i4, vwzcjt36; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1781.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-oQ8e-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,781 98.9 12.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/74mhkji0",
   "source_url": "https://omlx.ai/benchmarks/performance/74mhkji0",
   "tg_tokens": null,
   "tps": 98.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-74mhkji0-ctx4096",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: 7chu59nd, 8jilsg24, 9m3qk3on, ckihx6hv, mx0k00i4, vwzcjt36; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1571.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-oQ8e-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 1,571 96.7 12.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/74mhkji0",
   "source_url": "https://omlx.ai/benchmarks/performance/74mhkji0",
   "tg_tokens": null,
   "tps": 96.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-74mhkji0-ctx8192",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 12.8 GB at this context; duplicate traces of the same session: 7chu59nd, 8jilsg24, 9m3qk3on, ckihx6hv, mx0k00i4, vwzcjt36; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1647.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-oQ8e-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 8k 1,647 93.8 12.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/74mhkji0",
   "source_url": "https://omlx.ai/benchmarks/performance/74mhkji0",
   "tg_tokens": null,
   "tps": 93.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-74mhkji0-ctx16384",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 12.9 GB at this context; duplicate traces of the same session: 7chu59nd, 8jilsg24, 9m3qk3on, ckihx6hv, mx0k00i4, vwzcjt36; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1544.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-oQ8e-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 16k 1,544 89.9 12.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/74mhkji0",
   "source_url": "https://omlx.ai/benchmarks/performance/74mhkji0",
   "tg_tokens": null,
   "tps": 89.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-74mhkji0-ctx32768",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 13.1 GB at this context; duplicate traces of the same session: 7chu59nd, 8jilsg24, 9m3qk3on, ckihx6hv, mx0k00i4, vwzcjt36; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1358.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-oQ8e-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 32k 1,358 82.3 13.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/74mhkji0",
   "source_url": "https://omlx.ai/benchmarks/performance/74mhkji0",
   "tg_tokens": null,
   "tps": 82.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-74mhkji0-ctx65536",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 13.6 GB at this context; duplicate traces of the same session: 7chu59nd, 8jilsg24, 9m3qk3on, ckihx6hv, mx0k00i4, vwzcjt36; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1101.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-oQ8e-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 64k 1,101 70.1 13.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/74mhkji0",
   "source_url": "https://omlx.ai/benchmarks/performance/74mhkji0",
   "tg_tokens": null,
   "tps": 70.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-74mhkji0-ctx131072",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: 7chu59nd, 8jilsg24, 9m3qk3on, ckihx6hv, mx0k00i4, vwzcjt36; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 798.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-oQ8e-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 128k 798.9 56.1 14.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/74mhkji0",
   "source_url": "https://omlx.ai/benchmarks/performance/74mhkji0",
   "tg_tokens": null,
   "tps": 56.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-sfn5ezw3-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.5 GB at this context; duplicate traces of the same session: jkpjr5qf; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2608.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 2,608 86.7 37.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sfn5ezw3",
   "source_url": "https://omlx.ai/benchmarks/performance/sfn5ezw3",
   "tg_tokens": null,
   "tps": 86.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-sfn5ezw3-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.7 GB at this context; duplicate traces of the same session: jkpjr5qf; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2280.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 16k 2,280 81.9 37.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sfn5ezw3",
   "source_url": "https://omlx.ai/benchmarks/performance/sfn5ezw3",
   "tg_tokens": null,
   "tps": 81.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-46oybovp-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.5 GB at this context; duplicate traces of the same session: m9ke14v2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2671.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 2,671 89.2 37.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/46oybovp",
   "source_url": "https://omlx.ai/benchmarks/performance/46oybovp",
   "tg_tokens": null,
   "tps": 89.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-46oybovp-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.7 GB at this context; duplicate traces of the same session: m9ke14v2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2291.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 16k 2,291 79.9 37.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/46oybovp",
   "source_url": "https://omlx.ai/benchmarks/performance/46oybovp",
   "tg_tokens": null,
   "tps": 79.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-6cyk2r8u-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 26.9 GB; TTFT 108125 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 606.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 64k tokens 606.1 PP tok/s 56.6 TG tok/s 108125 TTFT (ms) 26.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6cyk2r8u",
   "source_url": "https://omlx.ai/benchmarks/performance/6cyk2r8u",
   "tg_tokens": null,
   "tps": 56.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-ep26h13b-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: nffkzjc1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 172.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 172.0 12.4 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ep26h13b",
   "source_url": "https://omlx.ai/benchmarks/performance/ep26h13b",
   "tg_tokens": null,
   "tps": 12.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-ep26h13b-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: nffkzjc1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 166.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 166.4 11.4 20.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ep26h13b",
   "source_url": "https://omlx.ai/benchmarks/performance/ep26h13b",
   "tg_tokens": null,
   "tps": 11.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-mj5j6tgp-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.8 GB at this context; duplicate traces of the same session: 3rroew7p; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 337.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 337.9 22.5 31.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mj5j6tgp",
   "source_url": "https://omlx.ai/benchmarks/performance/mj5j6tgp",
   "tg_tokens": null,
   "tps": 22.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-mj5j6tgp-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.9 GB at this context; duplicate traces of the same session: 3rroew7p; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 319.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 16k 319.6 21.4 32.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mj5j6tgp",
   "source_url": "https://omlx.ai/benchmarks/performance/mj5j6tgp",
   "tg_tokens": null,
   "tps": 21.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-nfpu123k-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTPLX Optimized Quality FP16 build; oMLX community benchmark trace; peak memory 40.5 GB at this context; duplicate traces of the same session: k0xy5lvn; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality-FP16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 1k 103.2 12.2 40.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nfpu123k",
   "source_url": "https://omlx.ai/benchmarks/performance/nfpu123k",
   "tg_tokens": null,
   "tps": 12.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-nfpu123k-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTPLX Optimized Quality FP16 build; oMLX community benchmark trace; peak memory 44.5 GB at this context; duplicate traces of the same session: k0xy5lvn; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 100.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality-FP16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 4k 100.8 10.2 44.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nfpu123k",
   "source_url": "https://omlx.ai/benchmarks/performance/nfpu123k",
   "tg_tokens": null,
   "tps": 10.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-lz1o49n1-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.8 GB at this context; duplicate traces of the same session: i19qpkt3, ujg8usgd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 629.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 629.8 50.1 69.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lz1o49n1",
   "source_url": "https://omlx.ai/benchmarks/performance/lz1o49n1",
   "tg_tokens": null,
   "tps": 50.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-lz1o49n1-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.1 GB at this context; duplicate traces of the same session: i19qpkt3, ujg8usgd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 652.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 652.8 49.5 70.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lz1o49n1",
   "source_url": "https://omlx.ai/benchmarks/performance/lz1o49n1",
   "tg_tokens": null,
   "tps": 49.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-lz1o49n1-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: i19qpkt3, ujg8usgd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 631.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 631.2 42.9 70.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lz1o49n1",
   "source_url": "https://omlx.ai/benchmarks/performance/lz1o49n1",
   "tg_tokens": null,
   "tps": 42.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-e2l5cemk-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.8 GB at this context; duplicate traces of the same session: vawlvyzw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 618.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 618.3 48.0 69.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e2l5cemk",
   "source_url": "https://omlx.ai/benchmarks/performance/e2l5cemk",
   "tg_tokens": null,
   "tps": 48.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-e2l5cemk-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.9 GB at this context; duplicate traces of the same session: vawlvyzw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 655.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 655.9 47.1 69.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e2l5cemk",
   "source_url": "https://omlx.ai/benchmarks/performance/e2l5cemk",
   "tg_tokens": null,
   "tps": 47.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-lq7jm98s-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.3 GB at this context; duplicate traces of the same session: x99dkke6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 485.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 485.4 50.3 69.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lq7jm98s",
   "source_url": "https://omlx.ai/benchmarks/performance/lq7jm98s",
   "tg_tokens": null,
   "tps": 50.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-lq7jm98s-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.9 GB at this context; duplicate traces of the same session: x99dkke6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 646.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 646.2 51.6 69.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lq7jm98s",
   "source_url": "https://omlx.ai/benchmarks/performance/lq7jm98s",
   "tg_tokens": null,
   "tps": 51.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-bmf2f79d-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: 8b6rz3ub, i5jta4ay, lppmiomp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 461.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 461.6 48.3 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bmf2f79d",
   "source_url": "https://omlx.ai/benchmarks/performance/bmf2f79d",
   "tg_tokens": null,
   "tps": 48.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-bmf2f79d-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: 8b6rz3ub, i5jta4ay, lppmiomp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 616.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 616.2 45.1 71.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bmf2f79d",
   "source_url": "https://omlx.ai/benchmarks/performance/bmf2f79d",
   "tg_tokens": null,
   "tps": 45.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-bmf2f79d-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 8b6rz3ub, i5jta4ay, lppmiomp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 585.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 585.0 46.5 71.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bmf2f79d",
   "source_url": "https://omlx.ai/benchmarks/performance/bmf2f79d",
   "tg_tokens": null,
   "tps": 46.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-bmf2f79d-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.2 GB at this context; duplicate traces of the same session: 8b6rz3ub, i5jta4ay, lppmiomp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 576.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 576.7 40.4 73.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bmf2f79d",
   "source_url": "https://omlx.ai/benchmarks/performance/bmf2f79d",
   "tg_tokens": null,
   "tps": 40.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-ppkltxzi-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.7 GB at this context; duplicate traces of the same session: dmyamufo; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1568.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M5 (8c); M5 (8c) \u00b7 16 GB \u00b7 4bit; 1k 1,568 39.3 4.7 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ppkltxzi",
   "source_url": "https://omlx.ai/benchmarks/performance/ppkltxzi",
   "tg_tokens": null,
   "tps": 39.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-ppkltxzi-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.7 GB at this context; duplicate traces of the same session: dmyamufo; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1633.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M5 (8c); M5 (8c) \u00b7 16 GB \u00b7 4bit; 4k 1,633 38.2 4.7 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ppkltxzi",
   "source_url": "https://omlx.ai/benchmarks/performance/ppkltxzi",
   "tg_tokens": null,
   "tps": 38.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-zfxhxedt-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.7 GB at this context; duplicate traces of the same session: gbu0d4ap; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1586.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M5 (8c); M5 (8c) \u00b7 16 GB \u00b7 4bit; 1k 1,586 40.1 4.7 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zfxhxedt",
   "source_url": "https://omlx.ai/benchmarks/performance/zfxhxedt",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-zfxhxedt-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.7 GB at this context; duplicate traces of the same session: gbu0d4ap; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1661.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M5 (8c); M5 (8c) \u00b7 16 GB \u00b7 4bit; 4k 1,661 38.6 4.7 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zfxhxedt",
   "source_url": "https://omlx.ai/benchmarks/performance/zfxhxedt",
   "tg_tokens": null,
   "tps": 38.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-kd4p8ikf-ctx1024",
   "model": "Hemmingway-1",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.0 GB at this context; duplicate traces of the same session: 2xgz4pt5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 230.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Hemmingway-1-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 230.6 18.1 29.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kd4p8ikf",
   "source_url": "https://omlx.ai/benchmarks/performance/kd4p8ikf",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-kd4p8ikf-ctx4096",
   "model": "Hemmingway-1",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.3 GB at this context; duplicate traces of the same session: 2xgz4pt5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 234.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Hemmingway-1-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 234.8 19.7 32.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kd4p8ikf",
   "source_url": "https://omlx.ai/benchmarks/performance/kd4p8ikf",
   "tg_tokens": null,
   "tps": 19.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-dhm1gswm-ctx65536",
   "model": "Ornith-1.5-35B-A3B-BigBang",
   "notes": "MTP build; oMLX community benchmark trace, single run at 64k context; peak memory 26.8 GB; TTFT 105945 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 618.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-BigBang-MTP-mlx-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 64k tokens 618.6 PP tok/s 63.0 TG tok/s 105945 TTFT (ms) 26.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dhm1gswm",
   "source_url": "https://omlx.ai/benchmarks/performance/dhm1gswm",
   "tg_tokens": null,
   "tps": 63.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-8nj3xpr2-ctx65536",
   "model": "Qwen-AgentWorld-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 25.9 GB; TTFT 109141 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 600.5,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen-AgentWorld-35B-A3B-oQ4 on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 64k tokens 600.5 PP tok/s 47.0 TG tok/s 109141 TTFT (ms) 25.9 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8nj3xpr2",
   "source_url": "https://omlx.ai/benchmarks/performance/8nj3xpr2",
   "tg_tokens": null,
   "tps": 47.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xrhsniw1-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: 1fzrjsfd, 6l2xmbz4, 72z7w6jq, 8avvronb, m81l0tzi, njmsyehc, qmyt7osb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 931.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 931.6 81.5 37.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xrhsniw1",
   "source_url": "https://omlx.ai/benchmarks/performance/xrhsniw1",
   "tg_tokens": null,
   "tps": 81.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xrhsniw1-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 38.0 GB at this context; duplicate traces of the same session: 1fzrjsfd, 6l2xmbz4, 72z7w6jq, 8avvronb, m81l0tzi, njmsyehc, qmyt7osb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2668.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 2,668 129.8 38.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xrhsniw1",
   "source_url": "https://omlx.ai/benchmarks/performance/xrhsniw1",
   "tg_tokens": null,
   "tps": 129.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xrhsniw1-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 38.1 GB at this context; duplicate traces of the same session: 1fzrjsfd, 6l2xmbz4, 72z7w6jq, 8avvronb, m81l0tzi, njmsyehc, qmyt7osb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5414.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 5,414 125.1 38.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xrhsniw1",
   "source_url": "https://omlx.ai/benchmarks/performance/xrhsniw1",
   "tg_tokens": null,
   "tps": 125.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xrhsniw1-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 38.3 GB at this context; duplicate traces of the same session: 1fzrjsfd, 6l2xmbz4, 72z7w6jq, 8avvronb, m81l0tzi, njmsyehc, qmyt7osb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5179.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 5,179 128.6 38.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xrhsniw1",
   "source_url": "https://omlx.ai/benchmarks/performance/xrhsniw1",
   "tg_tokens": null,
   "tps": 128.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-xrhsniw1-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 38.7 GB at this context; duplicate traces of the same session: 1fzrjsfd, 6l2xmbz4, 72z7w6jq, 8avvronb, m81l0tzi, njmsyehc, qmyt7osb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4223.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 4,223 179.6 38.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xrhsniw1",
   "source_url": "https://omlx.ai/benchmarks/performance/xrhsniw1",
   "tg_tokens": null,
   "tps": 179.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-xrhsniw1-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.4 GB at this context; duplicate traces of the same session: 1fzrjsfd, 6l2xmbz4, 72z7w6jq, 8avvronb, m81l0tzi, njmsyehc, qmyt7osb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3137.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 3,137 120.8 39.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xrhsniw1",
   "source_url": "https://omlx.ai/benchmarks/performance/xrhsniw1",
   "tg_tokens": null,
   "tps": 120.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-xrhsniw1-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 40.9 GB at this context; duplicate traces of the same session: 1fzrjsfd, 6l2xmbz4, 72z7w6jq, 8avvronb, m81l0tzi, njmsyehc, qmyt7osb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2153.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 2,153 97.2 40.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xrhsniw1",
   "source_url": "https://omlx.ai/benchmarks/performance/xrhsniw1",
   "tg_tokens": null,
   "tps": 97.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-xrhsniw1-ctx199680",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 42.7 GB at this context; duplicate traces of the same session: 1fzrjsfd, 6l2xmbz4, 72z7w6jq, 8avvronb, m81l0tzi, njmsyehc, qmyt7osb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1533.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 195k 1,533 66.9 42.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xrhsniw1",
   "source_url": "https://omlx.ai/benchmarks/performance/xrhsniw1",
   "tg_tokens": null,
   "tps": 66.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-h585vyhh-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 32.1 GB; TTFT 2769 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 369.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 8bit; 1k tokens 369.8 PP tok/s 20.7 TG tok/s 2769 TTFT (ms) 32.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h585vyhh",
   "source_url": "https://omlx.ai/benchmarks/performance/h585vyhh",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-plaatfoc-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "SAGE-PATHS MTP build; oMLX community benchmark trace, single run at 16k context; peak memory 29.7 GB; TTFT 41799 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 392.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-SAGE-PATHS-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 16k tokens 392.0 PP tok/s 45.2 TG tok/s 41799 TTFT (ms) 29.7 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/plaatfoc",
   "source_url": "https://omlx.ai/benchmarks/performance/plaatfoc",
   "tg_tokens": null,
   "tps": 45.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-q4mm8nn5-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "SAGE-PATHS MTP build; oMLX community benchmark trace, single run at 16k context; peak memory 29.7 GB; TTFT 40822 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 401.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-SAGE-PATHS-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 16k tokens 401.4 PP tok/s 51.2 TG tok/s 40822 TTFT (ms) 29.7 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q4mm8nn5",
   "source_url": "https://omlx.ai/benchmarks/performance/q4mm8nn5",
   "tg_tokens": null,
   "tps": 51.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4gzfsi7x-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "SAGE-PATHS MTP build; oMLX community benchmark trace, single run at 16k context; peak memory 26.3 GB; TTFT 52286 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 313.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-SAGE-PATHS-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 16k tokens 313.4 PP tok/s 49.3 TG tok/s 52286 TTFT (ms) 26.3 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4gzfsi7x",
   "source_url": "https://omlx.ai/benchmarks/performance/4gzfsi7x",
   "tg_tokens": null,
   "tps": 49.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-h3wjznhy-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "SAGE-PATHS MTP build; oMLX community benchmark trace, single run at 16k context; peak memory 26.3 GB; TTFT 52630 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 311.3,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-SAGE-PATHS-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 16k tokens 311.3 PP tok/s 45.5 TG tok/s 52630 TTFT (ms) 26.3 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h3wjznhy",
   "source_url": "https://omlx.ai/benchmarks/performance/h3wjznhy",
   "tg_tokens": null,
   "tps": 45.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kv7f9sqq-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: 0i232ibg, 16h7exye, 2fyneo6z, bsw52ls9, mudkoufh, pk1dkgue; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 386.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 386.7 16.8 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kv7f9sqq",
   "source_url": "https://omlx.ai/benchmarks/performance/kv7f9sqq",
   "tg_tokens": null,
   "tps": 16.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kv7f9sqq-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 0i232ibg, 16h7exye, 2fyneo6z, bsw52ls9, mudkoufh, pk1dkgue; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1117.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,117 65.5 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kv7f9sqq",
   "source_url": "https://omlx.ai/benchmarks/performance/kv7f9sqq",
   "tg_tokens": null,
   "tps": 65.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kv7f9sqq-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 0i232ibg, 16h7exye, 2fyneo6z, bsw52ls9, mudkoufh, pk1dkgue; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1702.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,702 58.1 71.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kv7f9sqq",
   "source_url": "https://omlx.ai/benchmarks/performance/kv7f9sqq",
   "tg_tokens": null,
   "tps": 58.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kv7f9sqq-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: 0i232ibg, 16h7exye, 2fyneo6z, bsw52ls9, mudkoufh, pk1dkgue; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2485.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,485 38.8 73.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kv7f9sqq",
   "source_url": "https://omlx.ai/benchmarks/performance/kv7f9sqq",
   "tg_tokens": null,
   "tps": 38.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kv7f9sqq-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 75.4 GB at this context; duplicate traces of the same session: 0i232ibg, 16h7exye, 2fyneo6z, bsw52ls9, mudkoufh, pk1dkgue; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2317.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,317 48.3 75.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kv7f9sqq",
   "source_url": "https://omlx.ai/benchmarks/performance/kv7f9sqq",
   "tg_tokens": null,
   "tps": 48.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kv7f9sqq-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.8 GB at this context; duplicate traces of the same session: 0i232ibg, 16h7exye, 2fyneo6z, bsw52ls9, mudkoufh, pk1dkgue; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2210.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,210 60.3 79.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kv7f9sqq",
   "source_url": "https://omlx.ai/benchmarks/performance/kv7f9sqq",
   "tg_tokens": null,
   "tps": 60.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kv7f9sqq-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.9 GB at this context; duplicate traces of the same session: 0i232ibg, 16h7exye, 2fyneo6z, bsw52ls9, mudkoufh, pk1dkgue; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2051.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,051 57.4 80.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kv7f9sqq",
   "source_url": "https://omlx.ai/benchmarks/performance/kv7f9sqq",
   "tg_tokens": null,
   "tps": 57.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-a4lp4i93-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: 2uat8cni, 313qldl2, bbmurhis, kcyg1chw, l0o45jhx, qzn1y7g7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1164.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,164 68.0 71.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a4lp4i93",
   "source_url": "https://omlx.ai/benchmarks/performance/a4lp4i93",
   "tg_tokens": null,
   "tps": 68.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-a4lp4i93-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 71.2 GB at this context; duplicate traces of the same session: 2uat8cni, 313qldl2, bbmurhis, kcyg1chw, l0o45jhx, qzn1y7g7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1413.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,413 71.0 71.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a4lp4i93",
   "source_url": "https://omlx.ai/benchmarks/performance/a4lp4i93",
   "tg_tokens": null,
   "tps": 71.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-a4lp4i93-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 2uat8cni, 313qldl2, bbmurhis, kcyg1chw, l0o45jhx, qzn1y7g7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1776.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,776 74.4 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a4lp4i93",
   "source_url": "https://omlx.ai/benchmarks/performance/a4lp4i93",
   "tg_tokens": null,
   "tps": 74.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-a4lp4i93-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 73.8 GB at this context; duplicate traces of the same session: 2uat8cni, 313qldl2, bbmurhis, kcyg1chw, l0o45jhx, qzn1y7g7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2099.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,099 75.9 73.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a4lp4i93",
   "source_url": "https://omlx.ai/benchmarks/performance/a4lp4i93",
   "tg_tokens": null,
   "tps": 75.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-a4lp4i93-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 76.4 GB at this context; duplicate traces of the same session: 2uat8cni, 313qldl2, bbmurhis, kcyg1chw, l0o45jhx, qzn1y7g7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2086.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,086 69.0 76.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a4lp4i93",
   "source_url": "https://omlx.ai/benchmarks/performance/a4lp4i93",
   "tg_tokens": null,
   "tps": 69.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-a4lp4i93-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 81.8 GB at this context; duplicate traces of the same session: 2uat8cni, 313qldl2, bbmurhis, kcyg1chw, l0o45jhx, qzn1y7g7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2035.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,035 63.7 81.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a4lp4i93",
   "source_url": "https://omlx.ai/benchmarks/performance/a4lp4i93",
   "tg_tokens": null,
   "tps": 63.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-a4lp4i93-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 84.6 GB at this context; duplicate traces of the same session: 2uat8cni, 313qldl2, bbmurhis, kcyg1chw, l0o45jhx, qzn1y7g7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1908.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 1,908 54.0 84.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a4lp4i93",
   "source_url": "https://omlx.ai/benchmarks/performance/a4lp4i93",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gfgn2s0n-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "Youssofal MTPLX Optimized Speed; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: bnbjq1g0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 338.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Youssofal--Qwen3.8-27B-MTPLX-Optimized-Speed on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 338.0 29.9 23.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gfgn2s0n",
   "source_url": "https://omlx.ai/benchmarks/performance/gfgn2s0n",
   "tg_tokens": null,
   "tps": 29.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gfgn2s0n-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "Youssofal MTPLX Optimized Speed; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: bnbjq1g0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 322.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Youssofal--Qwen3.8-27B-MTPLX-Optimized-Speed on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 16k 322.1 29.7 24.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gfgn2s0n",
   "source_url": "https://omlx.ai/benchmarks/performance/gfgn2s0n",
   "tg_tokens": null,
   "tps": 29.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-zhvzdu1h-ctx1024",
   "model": "gemma-4-26b-a4b-it-UD",
   "notes": "oMLX community benchmark trace; peak memory 14.7 GB at this context; duplicate traces of the same session: azjoo0ob, nd5xuz3g, vnvz1n7f, zvpl0odd; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 56.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-UD-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 1k 56.9 14.0 14.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zhvzdu1h",
   "source_url": "https://omlx.ai/benchmarks/performance/zhvzdu1h",
   "tg_tokens": null,
   "tps": 14.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-zhvzdu1h-ctx4096",
   "model": "gemma-4-26b-a4b-it-UD",
   "notes": "oMLX community benchmark trace; peak memory 15.1 GB at this context; duplicate traces of the same session: azjoo0ob, nd5xuz3g, vnvz1n7f, zvpl0odd; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 446.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-UD-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 4k 446.7 34.9 15.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zhvzdu1h",
   "source_url": "https://omlx.ai/benchmarks/performance/zhvzdu1h",
   "tg_tokens": null,
   "tps": 34.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-zhvzdu1h-ctx8192",
   "model": "gemma-4-26b-a4b-it-UD",
   "notes": "oMLX community benchmark trace; peak memory 15.3 GB at this context; duplicate traces of the same session: azjoo0ob, nd5xuz3g, vnvz1n7f, zvpl0odd; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 126.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-UD-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 8k 126.7 5.7 15.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zhvzdu1h",
   "source_url": "https://omlx.ai/benchmarks/performance/zhvzdu1h",
   "tg_tokens": null,
   "tps": 5.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-zhvzdu1h-ctx16384",
   "model": "gemma-4-26b-a4b-it-UD",
   "notes": "oMLX community benchmark trace; peak memory 15.7 GB at this context; duplicate traces of the same session: azjoo0ob, nd5xuz3g, vnvz1n7f, zvpl0odd; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 145.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-UD-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 16k 145.4 24.6 15.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zhvzdu1h",
   "source_url": "https://omlx.ai/benchmarks/performance/zhvzdu1h",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-zhvzdu1h-ctx32768",
   "model": "gemma-4-26b-a4b-it-UD",
   "notes": "oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: azjoo0ob, nd5xuz3g, vnvz1n7f, zvpl0odd; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 195.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-UD-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 32k 195.2 28.3 16.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zhvzdu1h",
   "source_url": "https://omlx.ai/benchmarks/performance/zhvzdu1h",
   "tg_tokens": null,
   "tps": 28.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2",
   "id": "omlx-ur3gls1y-ctx1024",
   "model": "gemma-4-E2B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 4.1 GB; TTFT 881 ms; duplicate traces of the same session: none; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1162.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E2B-it-oQ4e-mtp on M2 (8c); M2 (8c) \u00b7 16 GB \u00b7 4bit; 1k tokens 1,162 PP tok/s 45.9 TG tok/s 881 TTFT (ms) 4.1 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ur3gls1y",
   "source_url": "https://omlx.ai/benchmarks/performance/ur3gls1y",
   "tg_tokens": null,
   "tps": 45.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-jpj2x5sy-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 5.7 GB at this context; duplicate traces of the same session: r69bndtt; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1539.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M5 (8c); M5 (8c) \u00b7 16 GB \u00b7 4bit; 1k 1,539 42.4 5.7 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jpj2x5sy",
   "source_url": "https://omlx.ai/benchmarks/performance/jpj2x5sy",
   "tg_tokens": null,
   "tps": 42.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-jpj2x5sy-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: r69bndtt; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1690.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M5 (8c); M5 (8c) \u00b7 16 GB \u00b7 4bit; 4k 1,690 72.1 5.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jpj2x5sy",
   "source_url": "https://omlx.ai/benchmarks/performance/jpj2x5sy",
   "tg_tokens": null,
   "tps": 72.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-8o7gx63x-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 5.7 GB at this context; duplicate traces of the same session: jq1vdqp9; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1485.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M5 (8c); M5 (8c) \u00b7 16 GB \u00b7 4bit; 1k 1,485 67.9 5.7 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8o7gx63x",
   "source_url": "https://omlx.ai/benchmarks/performance/8o7gx63x",
   "tg_tokens": null,
   "tps": 67.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-8o7gx63x-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: jq1vdqp9; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1682.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M5 (8c); M5 (8c) \u00b7 16 GB \u00b7 4bit; 4k 1,682 63.4 5.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8o7gx63x",
   "source_url": "https://omlx.ai/benchmarks/performance/8o7gx63x",
   "tg_tokens": null,
   "tps": 63.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-05ms7ss6-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Nail build; oMLX community benchmark trace; peak memory 30.7 GB at this context; duplicate traces of the same session: ynltnycq; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 521.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX-6bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 1k 521.5 51.4 30.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/05ms7ss6",
   "source_url": "https://omlx.ai/benchmarks/performance/05ms7ss6",
   "tg_tokens": null,
   "tps": 51.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-05ms7ss6-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Nail build; oMLX community benchmark trace; peak memory 32.5 GB at this context; duplicate traces of the same session: ynltnycq; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 577.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX-6bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 4k 577.0 49.9 32.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/05ms7ss6",
   "source_url": "https://omlx.ai/benchmarks/performance/05ms7ss6",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-grvsm4eq-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: uzjrfhw1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 894.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 894.9 91.6 19.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/grvsm4eq",
   "source_url": "https://omlx.ai/benchmarks/performance/grvsm4eq",
   "tg_tokens": null,
   "tps": 91.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-grvsm4eq-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: uzjrfhw1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 874.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 874.1 73.5 20.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/grvsm4eq",
   "source_url": "https://omlx.ai/benchmarks/performance/grvsm4eq",
   "tg_tokens": null,
   "tps": 73.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-knpzxi91-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 7sxoqk3m; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 885.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 885.6 84.4 19.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/knpzxi91",
   "source_url": "https://omlx.ai/benchmarks/performance/knpzxi91",
   "tg_tokens": null,
   "tps": 84.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-knpzxi91-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: 7sxoqk3m; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 872.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 872.4 77.2 20.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/knpzxi91",
   "source_url": "https://omlx.ai/benchmarks/performance/knpzxi91",
   "tg_tokens": null,
   "tps": 77.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-3t18wm88-ctx1024",
   "model": "Ornith-1.5-35B-A3B-REAP192",
   "notes": "oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: 8tdvpusp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 960.2,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Ornith-1.5-35B-A3B-REAP192-mxfp4-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 960.2 96.7 14.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3t18wm88",
   "source_url": "https://omlx.ai/benchmarks/performance/3t18wm88",
   "tg_tokens": null,
   "tps": 96.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-3t18wm88-ctx16384",
   "model": "Ornith-1.5-35B-A3B-REAP192",
   "notes": "oMLX community benchmark trace; peak memory 15.4 GB at this context; duplicate traces of the same session: 8tdvpusp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 893.6,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Ornith-1.5-35B-A3B-REAP192-mxfp4-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 893.6 80.2 15.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3t18wm88",
   "source_url": "https://omlx.ai/benchmarks/performance/3t18wm88",
   "tg_tokens": null,
   "tps": 80.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-m6p2qbhn-ctx1024",
   "model": "Ornith-1.5-35B-A3B-REAP192",
   "notes": "oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: kqn4p3lu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 958.6,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Ornith-1.5-35B-A3B-REAP192-mxfp4-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 958.6 96.7 14.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m6p2qbhn",
   "source_url": "https://omlx.ai/benchmarks/performance/m6p2qbhn",
   "tg_tokens": null,
   "tps": 96.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-m6p2qbhn-ctx16384",
   "model": "Ornith-1.5-35B-A3B-REAP192",
   "notes": "oMLX community benchmark trace; peak memory 15.4 GB at this context; duplicate traces of the same session: kqn4p3lu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 892.2,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Ornith-1.5-35B-A3B-REAP192-mxfp4-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 892.2 80.5 15.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m6p2qbhn",
   "source_url": "https://omlx.ai/benchmarks/performance/m6p2qbhn",
   "tg_tokens": null,
   "tps": 80.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-wct9dp1i-ctx1024",
   "model": "Ornith-1.5-35B-A3B-REAP192",
   "notes": "oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: zwhmp777; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 952.5,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Ornith-1.5-35B-A3B-REAP192-mxfp4-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 952.5 95.6 14.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wct9dp1i",
   "source_url": "https://omlx.ai/benchmarks/performance/wct9dp1i",
   "tg_tokens": null,
   "tps": 95.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-wct9dp1i-ctx16384",
   "model": "Ornith-1.5-35B-A3B-REAP192",
   "notes": "oMLX community benchmark trace; peak memory 15.4 GB at this context; duplicate traces of the same session: zwhmp777; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 889.9,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Ornith-1.5-35B-A3B-REAP192-mxfp4-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 889.9 80.4 15.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wct9dp1i",
   "source_url": "https://omlx.ai/benchmarks/performance/wct9dp1i",
   "tg_tokens": null,
   "tps": 80.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-pd7fu4wz-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 36.7 GB at this context; duplicate traces of the same session: 2avh0z84; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2493.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 2,493 97.1 36.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pd7fu4wz",
   "source_url": "https://omlx.ai/benchmarks/performance/pd7fu4wz",
   "tg_tokens": null,
   "tps": 97.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-pd7fu4wz-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: 2avh0z84; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2086.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 16k 2,086 57.2 37.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pd7fu4wz",
   "source_url": "https://omlx.ai/benchmarks/performance/pd7fu4wz",
   "tg_tokens": null,
   "tps": 57.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-n0lysu73-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.5 GB at this context; duplicate traces of the same session: ly8lrs76; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2628.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 2,628 85.0 37.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n0lysu73",
   "source_url": "https://omlx.ai/benchmarks/performance/n0lysu73",
   "tg_tokens": null,
   "tps": 85.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-n0lysu73-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.7 GB at this context; duplicate traces of the same session: ly8lrs76; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2289.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 16k 2,289 80.6 37.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n0lysu73",
   "source_url": "https://omlx.ai/benchmarks/performance/n0lysu73",
   "tg_tokens": null,
   "tps": 80.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-98zj2sej-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 3rbkzf01, v35jyupd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1137.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 1,137 63.6 28.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/98zj2sej",
   "source_url": "https://omlx.ai/benchmarks/performance/98zj2sej",
   "tg_tokens": null,
   "tps": 63.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-98zj2sej-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.7 GB at this context; duplicate traces of the same session: 3rbkzf01, v35jyupd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1088.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 1,088 78.0 30.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/98zj2sej",
   "source_url": "https://omlx.ai/benchmarks/performance/98zj2sej",
   "tg_tokens": null,
   "tps": 78.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-98zj2sej-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.1 GB at this context; duplicate traces of the same session: 3rbkzf01, v35jyupd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 951.2,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 951.2 76.5 31.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/98zj2sej",
   "source_url": "https://omlx.ai/benchmarks/performance/98zj2sej",
   "tg_tokens": null,
   "tps": 76.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-07jbbblz-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.3 GB at this context; duplicate traces of the same session: aia2uuza; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 330.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 330.9 38.9 33.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/07jbbblz",
   "source_url": "https://omlx.ai/benchmarks/performance/07jbbblz",
   "tg_tokens": null,
   "tps": 38.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-07jbbblz-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.3 GB at this context; duplicate traces of the same session: aia2uuza; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 313.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 16k 313.2 31.6 34.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/07jbbblz",
   "source_url": "https://omlx.ai/benchmarks/performance/07jbbblz",
   "tg_tokens": null,
   "tps": 31.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-02nrh2na-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: hq1j7cra, jh8bpxnp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1036.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,036 48.2 18.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/02nrh2na",
   "source_url": "https://omlx.ai/benchmarks/performance/02nrh2na",
   "tg_tokens": null,
   "tps": 48.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-02nrh2na-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: hq1j7cra, jh8bpxnp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 922.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 922.0 53.1 19.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/02nrh2na",
   "source_url": "https://omlx.ai/benchmarks/performance/02nrh2na",
   "tg_tokens": null,
   "tps": 53.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-02nrh2na-ctx131072",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: hq1j7cra, jh8bpxnp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 564.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 128k 564.0 43.8 26.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/02nrh2na",
   "source_url": "https://omlx.ai/benchmarks/performance/02nrh2na",
   "tg_tokens": null,
   "tps": 43.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-s760r0l8-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "SAGE-PATHS MTP build; oMLX community benchmark trace, single run at 16k context; peak memory 29.7 GB; TTFT 40710 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 402.5,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-SAGE-PATHS-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 16k tokens 402.5 PP tok/s 47.3 TG tok/s 40710 TTFT (ms) 29.7 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s760r0l8",
   "source_url": "https://omlx.ai/benchmarks/performance/s760r0l8",
   "tg_tokens": null,
   "tps": 47.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-vdlt9h7u-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "SAGE-PATHS MTP build; oMLX community benchmark trace, single run at 16k context; peak memory 29.7 GB; TTFT 40838 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 401.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-SAGE-PATHS-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 16k tokens 401.2 PP tok/s 22.6 TG tok/s 40838 TTFT (ms) 29.7 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vdlt9h7u",
   "source_url": "https://omlx.ai/benchmarks/performance/vdlt9h7u",
   "tg_tokens": null,
   "tps": 22.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-60u204rh-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "SAGE-PATHS MTP build; oMLX community benchmark trace, single run at 16k context; peak memory 29.7 GB; TTFT 40794 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 401.6,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-SAGE-PATHS-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 16k tokens 401.6 PP tok/s 48.7 TG tok/s 40794 TTFT (ms) 29.7 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/60u204rh",
   "source_url": "https://omlx.ai/benchmarks/performance/60u204rh",
   "tg_tokens": null,
   "tps": 48.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-zdkxx9u4-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "SAGE-PATHS MTP build; oMLX community benchmark trace, single run at 16k context; peak memory 26.3 GB; TTFT 50669 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-SAGE-PATHS-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 16k tokens 323.4 PP tok/s 48.1 TG tok/s 50669 TTFT (ms) 26.3 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zdkxx9u4",
   "source_url": "https://omlx.ai/benchmarks/performance/zdkxx9u4",
   "tg_tokens": null,
   "tps": 48.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-weapyv2s-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "SAGE-PATHS MTP build; oMLX community benchmark trace, single run at 16k context; peak memory 29.7 GB; TTFT 40688 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 402.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-SAGE-PATHS-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 16k tokens 402.7 PP tok/s 46.5 TG tok/s 40688 TTFT (ms) 29.7 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/weapyv2s",
   "source_url": "https://omlx.ai/benchmarks/performance/weapyv2s",
   "tg_tokens": null,
   "tps": 46.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-jcrc4yq6-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "SAGE-PATHS MTP build; oMLX community benchmark trace, single run at 16k context; peak memory 29.7 GB; TTFT 40734 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 402.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-SAGE-PATHS-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 16k tokens 402.2 PP tok/s 38.9 TG tok/s 40734 TTFT (ms) 29.7 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jcrc4yq6",
   "source_url": "https://omlx.ai/benchmarks/performance/jcrc4yq6",
   "tg_tokens": null,
   "tps": 38.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2",
   "id": "omlx-6s00s2jf-ctx1024",
   "model": "gemma-4-E2B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.1 GB at this context; duplicate traces of the same session: 1hv0gor0, 9fwidiio, h6acjlvo, sipkvz93, yzbb26eg; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1150.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E2B-it-oQ4e-mtp on M2 (8c); M2 (8c) \u00b7 16 GB \u00b7 4bit; 1k 1,150 49.3 4.1 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6s00s2jf",
   "source_url": "https://omlx.ai/benchmarks/performance/6s00s2jf",
   "tg_tokens": null,
   "tps": 49.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2",
   "id": "omlx-6s00s2jf-ctx4096",
   "model": "gemma-4-E2B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: 1hv0gor0, 9fwidiio, h6acjlvo, sipkvz93, yzbb26eg; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1168.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E2B-it-oQ4e-mtp on M2 (8c); M2 (8c) \u00b7 16 GB \u00b7 4bit; 4k 1,168 51.1 4.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6s00s2jf",
   "source_url": "https://omlx.ai/benchmarks/performance/6s00s2jf",
   "tg_tokens": null,
   "tps": 51.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2",
   "id": "omlx-6s00s2jf-ctx8192",
   "model": "gemma-4-E2B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.3 GB at this context; duplicate traces of the same session: 1hv0gor0, 9fwidiio, h6acjlvo, sipkvz93, yzbb26eg; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1137.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E2B-it-oQ4e-mtp on M2 (8c); M2 (8c) \u00b7 16 GB \u00b7 4bit; 8k 1,137 44.8 4.3 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6s00s2jf",
   "source_url": "https://omlx.ai/benchmarks/performance/6s00s2jf",
   "tg_tokens": null,
   "tps": 44.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2",
   "id": "omlx-6s00s2jf-ctx16384",
   "model": "gemma-4-E2B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.4 GB at this context; duplicate traces of the same session: 1hv0gor0, 9fwidiio, h6acjlvo, sipkvz93, yzbb26eg; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1057.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E2B-it-oQ4e-mtp on M2 (8c); M2 (8c) \u00b7 16 GB \u00b7 4bit; 16k 1,057 38.5 4.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6s00s2jf",
   "source_url": "https://omlx.ai/benchmarks/performance/6s00s2jf",
   "tg_tokens": null,
   "tps": 38.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2",
   "id": "omlx-6s00s2jf-ctx32768",
   "model": "gemma-4-E2B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 5.0 GB at this context; duplicate traces of the same session: 1hv0gor0, 9fwidiio, h6acjlvo, sipkvz93, yzbb26eg; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 751.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E2B-it-oQ4e-mtp on M2 (8c); M2 (8c) \u00b7 16 GB \u00b7 4bit; 32k 751.5 21.6 5.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6s00s2jf",
   "source_url": "https://omlx.ai/benchmarks/performance/6s00s2jf",
   "tg_tokens": null,
   "tps": 21.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2",
   "id": "omlx-6s00s2jf-ctx65536",
   "model": "gemma-4-E2B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 5.5 GB at this context; duplicate traces of the same session: 1hv0gor0, 9fwidiio, h6acjlvo, sipkvz93, yzbb26eg; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 596.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E2B-it-oQ4e-mtp on M2 (8c); M2 (8c) \u00b7 16 GB \u00b7 4bit; 64k 596.5 22.2 5.5 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6s00s2jf",
   "source_url": "https://omlx.ai/benchmarks/performance/6s00s2jf",
   "tg_tokens": null,
   "tps": 22.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-nn4qrmjx-ctx131072",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: 3qzt2oq0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 766.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Muse-Glimmer-30B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 128k 766.5 19.7 23.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nn4qrmjx",
   "source_url": "https://omlx.ai/benchmarks/performance/nn4qrmjx",
   "tg_tokens": null,
   "tps": 19.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-nn4qrmjx-ctx199680",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: 3qzt2oq0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 693.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Muse-Glimmer-30B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 195k 693.8 18.2 24.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nn4qrmjx",
   "source_url": "https://omlx.ai/benchmarks/performance/nn4qrmjx",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-91ob0c6m-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: id23dp7u, kq0es53z, m6o1vzdu, xuovgbes; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 719.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 719.3 37.3 5.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/91ob0c6m",
   "source_url": "https://omlx.ai/benchmarks/performance/91ob0c6m",
   "tg_tokens": null,
   "tps": 37.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-91ob0c6m-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: id23dp7u, kq0es53z, m6o1vzdu, xuovgbes; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 830.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 830.3 36.3 6.3 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/91ob0c6m",
   "source_url": "https://omlx.ai/benchmarks/performance/91ob0c6m",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-91ob0c6m-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: id23dp7u, kq0es53z, m6o1vzdu, xuovgbes; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 858.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 8k 858.6 35.1 6.4 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/91ob0c6m",
   "source_url": "https://omlx.ai/benchmarks/performance/91ob0c6m",
   "tg_tokens": null,
   "tps": 35.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-91ob0c6m-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: id23dp7u, kq0es53z, m6o1vzdu, xuovgbes; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 816.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 16k 816.1 33.2 6.7 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/91ob0c6m",
   "source_url": "https://omlx.ai/benchmarks/performance/91ob0c6m",
   "tg_tokens": null,
   "tps": 33.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-91ob0c6m-ctx32768",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.2 GB at this context; duplicate traces of the same session: id23dp7u, kq0es53z, m6o1vzdu, xuovgbes; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 774.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 32k 774.2 28.6 7.2 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/91ob0c6m",
   "source_url": "https://omlx.ai/benchmarks/performance/91ob0c6m",
   "tg_tokens": null,
   "tps": 28.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3q758ln2-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 39.1 GB at this context; duplicate traces of the same session: d9qk6jou; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 863.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 863.2 25.9 39.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3q758ln2",
   "source_url": "https://omlx.ai/benchmarks/performance/3q758ln2",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3q758ln2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 40.1 GB at this context; duplicate traces of the same session: d9qk6jou; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 677.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 677.1 32.1 40.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3q758ln2",
   "source_url": "https://omlx.ai/benchmarks/performance/3q758ln2",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-v53hd1hq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 39.1 GB at this context; duplicate traces of the same session: p4svv8fd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 863.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 863.8 34.3 39.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v53hd1hq",
   "source_url": "https://omlx.ai/benchmarks/performance/v53hd1hq",
   "tg_tokens": null,
   "tps": 34.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-v53hd1hq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 40.1 GB at this context; duplicate traces of the same session: p4svv8fd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 713.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 713.8 32.6 40.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v53hd1hq",
   "source_url": "https://omlx.ai/benchmarks/performance/v53hd1hq",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-14r85r9g-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 39.1 GB at this context; duplicate traces of the same session: kw0m72it; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 862.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 862.1 35.4 39.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/14r85r9g",
   "source_url": "https://omlx.ai/benchmarks/performance/14r85r9g",
   "tg_tokens": null,
   "tps": 35.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-14r85r9g-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 40.1 GB at this context; duplicate traces of the same session: kw0m72it; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 700.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 700.4 34.0 40.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/14r85r9g",
   "source_url": "https://omlx.ai/benchmarks/performance/14r85r9g",
   "tg_tokens": null,
   "tps": 34.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ycbg7yz7-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "SAGE-PATHS MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 58t4dp7x; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-SAGE-PATHS-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 5bit; 1k 105.9 12.5 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ycbg7yz7",
   "source_url": "https://omlx.ai/benchmarks/performance/ycbg7yz7",
   "tg_tokens": null,
   "tps": 12.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ycbg7yz7-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "SAGE-PATHS MTP build; oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: 58t4dp7x; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 98.8,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-SAGE-PATHS-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 5bit; 4k 98.8 12.3 25.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ycbg7yz7",
   "source_url": "https://omlx.ai/benchmarks/performance/ycbg7yz7",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-pd6bj3aq-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "SAGE-PATHS MTP build; oMLX community benchmark trace, single run at 16k context; peak memory 29.7 GB; TTFT 40715 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 402.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-SAGE-PATHS-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 16k tokens 402.4 PP tok/s 49.9 TG tok/s 40715 TTFT (ms) 29.7 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pd6bj3aq",
   "source_url": "https://omlx.ai/benchmarks/performance/pd6bj3aq",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-zfm6zs9h-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "SAGE-PATHS MTP build; oMLX community benchmark trace, single run at 16k context; peak memory 26.6 GB; TTFT 51701 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 316.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-SAGE-PATHS-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 16k tokens 316.9 PP tok/s 22.2 TG tok/s 51701 TTFT (ms) 26.6 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zfm6zs9h",
   "source_url": "https://omlx.ai/benchmarks/performance/zfm6zs9h",
   "tg_tokens": null,
   "tps": 22.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-muv09g6w-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "SAGE-PATHS MTP build; oMLX community benchmark trace, single run at 16k context; peak memory 29.7 GB; TTFT 41725 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 392.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-SAGE-PATHS-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 16k tokens 392.7 PP tok/s 49.1 TG tok/s 41725 TTFT (ms) 29.7 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/muv09g6w",
   "source_url": "https://omlx.ai/benchmarks/performance/muv09g6w",
   "tg_tokens": null,
   "tps": 49.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5v38uhsv-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "SAGE-PATHS MTP build; oMLX community benchmark trace, single run at 16k context; peak memory 29.7 GB; TTFT 41105 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 398.6,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-SAGE-PATHS-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 16k tokens 398.6 PP tok/s 34.4 TG tok/s 41105 TTFT (ms) 29.7 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5v38uhsv",
   "source_url": "https://omlx.ai/benchmarks/performance/5v38uhsv",
   "tg_tokens": null,
   "tps": 34.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5ymi2e83-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "SAGE-PATHS MTP build; oMLX community benchmark trace, single run at 16k context; peak memory 29.7 GB; TTFT 41205 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 397.6,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-SAGE-PATHS-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 16k tokens 397.6 PP tok/s 46.6 TG tok/s 41205 TTFT (ms) 29.7 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ymi2e83",
   "source_url": "https://omlx.ai/benchmarks/performance/5ymi2e83",
   "tg_tokens": null,
   "tps": 46.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-yng9i0c0-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "SAGE-PATHS MTP build; oMLX community benchmark trace, single run at 16k context; peak memory 29.7 GB; TTFT 41184 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 397.8,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-SAGE-PATHS-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 16k tokens 397.8 PP tok/s 43.8 TG tok/s 41184 TTFT (ms) 29.7 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yng9i0c0",
   "source_url": "https://omlx.ai/benchmarks/performance/yng9i0c0",
   "tg_tokens": null,
   "tps": 43.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-69e8ngmn-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "SAGE-PATHS MTP build; oMLX community benchmark trace, single run at 16k context; peak memory 29.7 GB; TTFT 40966 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 399.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-SAGE-PATHS-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 16k tokens 399.9 PP tok/s 39.6 TG tok/s 40966 TTFT (ms) 29.7 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/69e8ngmn",
   "source_url": "https://omlx.ai/benchmarks/performance/69e8ngmn",
   "tg_tokens": null,
   "tps": 39.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-nhvcurzt-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 73d461gx, 8e7hdqn5; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 216.1,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-27B-oQ6 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 6bit; 1k 216.1 23.0 22.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nhvcurzt",
   "source_url": "https://omlx.ai/benchmarks/performance/nhvcurzt",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-nhvcurzt-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 26.2 GB at this context; duplicate traces of the same session: 73d461gx, 8e7hdqn5; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 200.3,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-27B-oQ6 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 6bit; 4k 200.3 22.9 26.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nhvcurzt",
   "source_url": "https://omlx.ai/benchmarks/performance/nhvcurzt",
   "tg_tokens": null,
   "tps": 22.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-nhvcurzt-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 27.0 GB at this context; duplicate traces of the same session: 73d461gx, 8e7hdqn5; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 197.4,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-27B-oQ6 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 6bit; 8k 197.4 21.3 27.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nhvcurzt",
   "source_url": "https://omlx.ai/benchmarks/performance/nhvcurzt",
   "tg_tokens": null,
   "tps": 21.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-86xosl12-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 100.7 GB; TTFT 1396 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 733.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 733.6 PP tok/s 75.3 TG tok/s 1396 TTFT (ms) 100.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/86xosl12",
   "source_url": "https://omlx.ai/benchmarks/performance/86xosl12",
   "tg_tokens": null,
   "tps": 75.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-jgv5htpd-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: 9uh64wub; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 619.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 619.4 70.9 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jgv5htpd",
   "source_url": "https://omlx.ai/benchmarks/performance/jgv5htpd",
   "tg_tokens": null,
   "tps": 70.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-jgv5htpd-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: 9uh64wub; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 703.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 703.2 59.0 71.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jgv5htpd",
   "source_url": "https://omlx.ai/benchmarks/performance/jgv5htpd",
   "tg_tokens": null,
   "tps": 59.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-vyo7h3ik-ctx4096",
   "model": "gemma-4-12B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 9.8 GB at this context; duplicate traces of the same session: auy9lkq3; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-qat-OptiQ-4bit-MTP on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 127.4 16.3 9.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vyo7h3ik",
   "source_url": "https://omlx.ai/benchmarks/performance/vyo7h3ik",
   "tg_tokens": null,
   "tps": 16.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-vyo7h3ik-ctx16384",
   "model": "gemma-4-12B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 10.3 GB at this context; duplicate traces of the same session: auy9lkq3; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 124.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-qat-OptiQ-4bit-MTP on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k 124.1 16.0 10.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vyo7h3ik",
   "source_url": "https://omlx.ai/benchmarks/performance/vyo7h3ik",
   "tg_tokens": null,
   "tps": 16.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-lbal1mo8-ctx4096",
   "model": "Huihui-Ornith-1.5-9B-abliterated",
   "notes": "oMLX community benchmark trace; peak memory 7.3 GB at this context; duplicate traces of the same session: uh52ew5u; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 194.2,
   "provenance": "community",
   "quant": "5Bit",
   "quote": "Huihui-Ornith-1.5-9B-abliterated-mlx-5Bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 5bit; 4k 194.2 8.4 7.3 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lbal1mo8",
   "source_url": "https://omlx.ai/benchmarks/performance/lbal1mo8",
   "tg_tokens": null,
   "tps": 8.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-lbal1mo8-ctx16384",
   "model": "Huihui-Ornith-1.5-9B-abliterated",
   "notes": "oMLX community benchmark trace; peak memory 7.8 GB at this context; duplicate traces of the same session: uh52ew5u; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 126.8,
   "provenance": "community",
   "quant": "5Bit",
   "quote": "Huihui-Ornith-1.5-9B-abliterated-mlx-5Bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 5bit; 16k 126.8 8.4 7.8 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lbal1mo8",
   "source_url": "https://omlx.ai/benchmarks/performance/lbal1mo8",
   "tg_tokens": null,
   "tps": 8.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-i91132ln-ctx1024",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "oMLX community benchmark trace; peak memory 160.6 GB at this context; duplicate traces of the same session: 12ef2mqa, bm6d6z1i, i8xa4a5e, m9m7yn0o, mdp4bkzv, ojp5ph15, rl80m8hp; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 545.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "MiMo-V2.6-Flash-RL-mxfp4-q8 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 545.0 43.5 160.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i91132ln",
   "source_url": "https://omlx.ai/benchmarks/performance/i91132ln",
   "tg_tokens": null,
   "tps": 43.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-i91132ln-ctx4096",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "oMLX community benchmark trace; peak memory 161.9 GB at this context; duplicate traces of the same session: 12ef2mqa, bm6d6z1i, i8xa4a5e, m9m7yn0o, mdp4bkzv, ojp5ph15, rl80m8hp; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 632.5,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "MiMo-V2.6-Flash-RL-mxfp4-q8 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 632.5 50.1 161.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i91132ln",
   "source_url": "https://omlx.ai/benchmarks/performance/i91132ln",
   "tg_tokens": null,
   "tps": 50.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-i91132ln-ctx8192",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "oMLX community benchmark trace; peak memory 163.0 GB at this context; duplicate traces of the same session: 12ef2mqa, bm6d6z1i, i8xa4a5e, m9m7yn0o, mdp4bkzv, ojp5ph15, rl80m8hp; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 617.7,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "MiMo-V2.6-Flash-RL-mxfp4-q8 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 617.7 53.3 163.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i91132ln",
   "source_url": "https://omlx.ai/benchmarks/performance/i91132ln",
   "tg_tokens": null,
   "tps": 53.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-i91132ln-ctx16384",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "oMLX community benchmark trace; peak memory 165.2 GB at this context; duplicate traces of the same session: 12ef2mqa, bm6d6z1i, i8xa4a5e, m9m7yn0o, mdp4bkzv, ojp5ph15, rl80m8hp; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 585.5,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "MiMo-V2.6-Flash-RL-mxfp4-q8 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 585.5 51.0 165.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i91132ln",
   "source_url": "https://omlx.ai/benchmarks/performance/i91132ln",
   "tg_tokens": null,
   "tps": 51.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-i91132ln-ctx32768",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "oMLX community benchmark trace; peak memory 169.5 GB at this context; duplicate traces of the same session: 12ef2mqa, bm6d6z1i, i8xa4a5e, m9m7yn0o, mdp4bkzv, ojp5ph15, rl80m8hp; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 522.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "MiMo-V2.6-Flash-RL-mxfp4-q8 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 522.0 46.2 169.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i91132ln",
   "source_url": "https://omlx.ai/benchmarks/performance/i91132ln",
   "tg_tokens": null,
   "tps": 46.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-i91132ln-ctx65536",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: 12ef2mqa, bm6d6z1i, i8xa4a5e, m9m7yn0o, mdp4bkzv, ojp5ph15, rl80m8hp; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 424.2,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "MiMo-V2.6-Flash-RL-mxfp4-q8 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 424.2 31.4 178.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i91132ln",
   "source_url": "https://omlx.ai/benchmarks/performance/i91132ln",
   "tg_tokens": null,
   "tps": 31.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-i91132ln-ctx131072",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "oMLX community benchmark trace; peak memory 226.8 GB at this context; duplicate traces of the same session: 12ef2mqa, bm6d6z1i, i8xa4a5e, m9m7yn0o, mdp4bkzv, ojp5ph15, rl80m8hp; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 243.7,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "MiMo-V2.6-Flash-RL-mxfp4-q8 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 243.7 28.6 226.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i91132ln",
   "source_url": "https://omlx.ai/benchmarks/performance/i91132ln",
   "tg_tokens": null,
   "tps": 28.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-i91132ln-ctx199680",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "oMLX community benchmark trace; peak memory 248.1 GB at this context; duplicate traces of the same session: 12ef2mqa, bm6d6z1i, i8xa4a5e, m9m7yn0o, mdp4bkzv, ojp5ph15, rl80m8hp; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.3,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "MiMo-V2.6-Flash-RL-mxfp4-q8 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 195k 220.3 19.1 248.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i91132ln",
   "source_url": "https://omlx.ai/benchmarks/performance/i91132ln",
   "tg_tokens": null,
   "tps": 19.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-15b78j7i-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 0fhbzly6, ysa0r9q2; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 110.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 1k 110.7 16.1 19.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/15b78j7i",
   "source_url": "https://omlx.ai/benchmarks/performance/15b78j7i",
   "tg_tokens": null,
   "tps": 16.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-15b78j7i-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 0fhbzly6, ysa0r9q2; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 99.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 99.1 26.4 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/15b78j7i",
   "source_url": "https://omlx.ai/benchmarks/performance/15b78j7i",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-15b78j7i-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: 0fhbzly6, ysa0r9q2; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 96.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 8k 96.0 10.4 24.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/15b78j7i",
   "source_url": "https://omlx.ai/benchmarks/performance/15b78j7i",
   "tg_tokens": null,
   "tps": 10.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1u6n0lmg-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "SAGE-PATHS MTP build; oMLX community benchmark trace, single run at 16k context; peak memory 29.7 GB; TTFT 40605 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 403.5,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-SAGE-PATHS-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 16k tokens 403.5 PP tok/s 50.5 TG tok/s 40605 TTFT (ms) 29.7 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1u6n0lmg",
   "source_url": "https://omlx.ai/benchmarks/performance/1u6n0lmg",
   "tg_tokens": null,
   "tps": 50.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-6vxh8639-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "SAGE-PATHS MTP build; oMLX community benchmark trace, single run at 16k context; peak memory 29.7 GB; TTFT 41372 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 396.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-SAGE-PATHS-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 16k tokens 396.0 PP tok/s 50.3 TG tok/s 41372 TTFT (ms) 29.7 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6vxh8639",
   "source_url": "https://omlx.ai/benchmarks/performance/6vxh8639",
   "tg_tokens": null,
   "tps": 50.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-q3cxnbnh-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "SAGE-PATHS MTP build; oMLX community benchmark trace, single run at 16k context; peak memory 29.7 GB; TTFT 41259 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 397.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-SAGE-PATHS-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 16k tokens 397.1 PP tok/s 48.2 TG tok/s 41259 TTFT (ms) 29.7 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q3cxnbnh",
   "source_url": "https://omlx.ai/benchmarks/performance/q3cxnbnh",
   "tg_tokens": null,
   "tps": 48.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2f49chwn-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "SAGE-PATHS MTP build; oMLX community benchmark trace, single run at 16k context; peak memory 29.7 GB; TTFT 40942 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 400.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-SAGE-PATHS-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 16k tokens 400.2 PP tok/s 49.3 TG tok/s 40942 TTFT (ms) 29.7 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2f49chwn",
   "source_url": "https://omlx.ai/benchmarks/performance/2f49chwn",
   "tg_tokens": null,
   "tps": 49.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rtjuq1an-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "SAGE-PATHS MTP build; oMLX community benchmark trace, single run at 16k context; peak memory 29.7 GB; TTFT 41090 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 398.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-SAGE-PATHS-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 16k tokens 398.7 PP tok/s 48.4 TG tok/s 41090 TTFT (ms) 29.7 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rtjuq1an",
   "source_url": "https://omlx.ai/benchmarks/performance/rtjuq1an",
   "tg_tokens": null,
   "tps": 48.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ylftijjd-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "SAGE-PATHS MTP build; oMLX community benchmark trace, single run at 16k context; peak memory 26.6 GB; TTFT 51591 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 317.6,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-SAGE-PATHS-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 16k tokens 317.6 PP tok/s 21.2 TG tok/s 51591 TTFT (ms) 26.6 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ylftijjd",
   "source_url": "https://omlx.ai/benchmarks/performance/ylftijjd",
   "tg_tokens": null,
   "tps": 21.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-n3lis3wm-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "leonsarmiento pi MTP; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 8yqfaw2v; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "leonsarmiento--Qwen3.8-27B-pi-4bit-mtp-mlx on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 205.2 43.8 19.2 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n3lis3wm",
   "source_url": "https://omlx.ai/benchmarks/performance/n3lis3wm",
   "tg_tokens": null,
   "tps": 43.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-n3lis3wm-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "leonsarmiento pi MTP; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 8yqfaw2v; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 206.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "leonsarmiento--Qwen3.8-27B-pi-4bit-mtp-mlx on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 206.9 41.9 20.2 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n3lis3wm",
   "source_url": "https://omlx.ai/benchmarks/performance/n3lis3wm",
   "tg_tokens": null,
   "tps": 41.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hvgqg022-ctx4096",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 89.7 GB at this context; duplicate traces of the same session: 8honakma, xoq7sgp4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 765.5,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4-Flash-0731-oQ2e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 4k 765.5 31.9 89.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hvgqg022",
   "source_url": "https://omlx.ai/benchmarks/performance/hvgqg022",
   "tg_tokens": null,
   "tps": 31.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hvgqg022-ctx8192",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 89.7 GB at this context; duplicate traces of the same session: 8honakma, xoq7sgp4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 799.8,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4-Flash-0731-oQ2e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 8k 799.8 31.5 89.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hvgqg022",
   "source_url": "https://omlx.ai/benchmarks/performance/hvgqg022",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hvgqg022-ctx16384",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 89.8 GB at this context; duplicate traces of the same session: 8honakma, xoq7sgp4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 763.4,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4-Flash-0731-oQ2e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 16k 763.4 31.0 89.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hvgqg022",
   "source_url": "https://omlx.ai/benchmarks/performance/hvgqg022",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1pgmhx1k-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: kf22igdz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4248.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,248 132.8 20.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1pgmhx1k",
   "source_url": "https://omlx.ai/benchmarks/performance/1pgmhx1k",
   "tg_tokens": null,
   "tps": 132.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1pgmhx1k-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: kf22igdz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4485.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 4,485 112.3 22.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1pgmhx1k",
   "source_url": "https://omlx.ai/benchmarks/performance/1pgmhx1k",
   "tg_tokens": null,
   "tps": 112.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-r98l1w4i-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: gdxr5irl, qz13m0xz, tpcww1xu, udxphs7r; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 107.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 1k 107.6 16.1 19.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r98l1w4i",
   "source_url": "https://omlx.ai/benchmarks/performance/r98l1w4i",
   "tg_tokens": null,
   "tps": 16.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-r98l1w4i-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: gdxr5irl, qz13m0xz, tpcww1xu, udxphs7r; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 95.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 95.7 18.3 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r98l1w4i",
   "source_url": "https://omlx.ai/benchmarks/performance/r98l1w4i",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-r98l1w4i-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: gdxr5irl, qz13m0xz, tpcww1xu, udxphs7r; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 93.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 8k 93.7 11.0 24.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r98l1w4i",
   "source_url": "https://omlx.ai/benchmarks/performance/r98l1w4i",
   "tg_tokens": null,
   "tps": 11.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-r98l1w4i-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: gdxr5irl, qz13m0xz, tpcww1xu, udxphs7r; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 90.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 16k 90.6 11.5 24.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r98l1w4i",
   "source_url": "https://omlx.ai/benchmarks/performance/r98l1w4i",
   "tg_tokens": null,
   "tps": 11.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-r98l1w4i-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: gdxr5irl, qz13m0xz, tpcww1xu, udxphs7r; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 84.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 32k 84.0 10.8 25.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r98l1w4i",
   "source_url": "https://omlx.ai/benchmarks/performance/r98l1w4i",
   "tg_tokens": null,
   "tps": 10.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-z71zkas2-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: j2vohis8; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 152.5,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 4k 152.5 21.9 23.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z71zkas2",
   "source_url": "https://omlx.ai/benchmarks/performance/z71zkas2",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-z71zkas2-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 25.7 GB at this context; duplicate traces of the same session: j2vohis8; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 139.6,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 16k 139.6 20.0 25.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z71zkas2",
   "source_url": "https://omlx.ai/benchmarks/performance/z71zkas2",
   "tg_tokens": null,
   "tps": 20.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-n2tns4pv-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 122.3 GB at this context; duplicate traces of the same session: pcb99da1, q3gula9k, wioa99td, y1gq8vzq, z7frfhar; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1056.0,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-Flash-Next-NVFP4 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 nvfp4; 1k 1,056 35.3 122.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n2tns4pv",
   "source_url": "https://omlx.ai/benchmarks/performance/n2tns4pv",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-n2tns4pv-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 122.8 GB at this context; duplicate traces of the same session: pcb99da1, q3gula9k, wioa99td, y1gq8vzq, z7frfhar; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1276.0,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-Flash-Next-NVFP4 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 nvfp4; 8k 1,276 33.5 122.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n2tns4pv",
   "source_url": "https://omlx.ai/benchmarks/performance/n2tns4pv",
   "tg_tokens": null,
   "tps": 33.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-n2tns4pv-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 122.9 GB at this context; duplicate traces of the same session: pcb99da1, q3gula9k, wioa99td, y1gq8vzq, z7frfhar; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1271.0,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-Flash-Next-NVFP4 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 nvfp4; 16k 1,271 34.2 122.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n2tns4pv",
   "source_url": "https://omlx.ai/benchmarks/performance/n2tns4pv",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-n2tns4pv-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 123.4 GB at this context; duplicate traces of the same session: pcb99da1, q3gula9k, wioa99td, y1gq8vzq, z7frfhar; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1267.0,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-Flash-Next-NVFP4 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 nvfp4; 32k 1,267 32.8 123.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n2tns4pv",
   "source_url": "https://omlx.ai/benchmarks/performance/n2tns4pv",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-n2tns4pv-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 127.9 GB at this context; duplicate traces of the same session: pcb99da1, q3gula9k, wioa99td, y1gq8vzq, z7frfhar; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1235.0,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-Flash-Next-NVFP4 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 nvfp4; 128k 1,235 31.3 127.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n2tns4pv",
   "source_url": "https://omlx.ai/benchmarks/performance/n2tns4pv",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-n2tns4pv-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 168.6 GB at this context; duplicate traces of the same session: pcb99da1, q3gula9k, wioa99td, y1gq8vzq, z7frfhar; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1012.0,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-Flash-Next-NVFP4 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 nvfp4; 195k 1,012 31.1 168.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n2tns4pv",
   "source_url": "https://omlx.ai/benchmarks/performance/n2tns4pv",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3j20ahb8-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 122.3 GB at this context; duplicate traces of the same session: 11l4aq60, 2l3ep1gw, 5f6mgjiz, csl2tkee, sw66sjuv, zzr76lzb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 798.3,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-Flash-Next-NVFP4 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 nvfp4; 1k 798.3 35.4 122.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3j20ahb8",
   "source_url": "https://omlx.ai/benchmarks/performance/3j20ahb8",
   "tg_tokens": null,
   "tps": 35.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3j20ahb8-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 122.8 GB at this context; duplicate traces of the same session: 11l4aq60, 2l3ep1gw, 5f6mgjiz, csl2tkee, sw66sjuv, zzr76lzb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1059.0,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-Flash-Next-NVFP4 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 nvfp4; 4k 1,059 32.3 122.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3j20ahb8",
   "source_url": "https://omlx.ai/benchmarks/performance/3j20ahb8",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3j20ahb8-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 122.8 GB at this context; duplicate traces of the same session: 11l4aq60, 2l3ep1gw, 5f6mgjiz, csl2tkee, sw66sjuv, zzr76lzb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1260.0,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-Flash-Next-NVFP4 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 nvfp4; 8k 1,260 33.6 122.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3j20ahb8",
   "source_url": "https://omlx.ai/benchmarks/performance/3j20ahb8",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-3j20ahb8-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 123.4 GB at this context; duplicate traces of the same session: 11l4aq60, 2l3ep1gw, 5f6mgjiz, csl2tkee, sw66sjuv, zzr76lzb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1257.0,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-Flash-Next-NVFP4 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 nvfp4; 32k 1,257 20.5 123.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3j20ahb8",
   "source_url": "https://omlx.ai/benchmarks/performance/3j20ahb8",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3j20ahb8-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 124.5 GB at this context; duplicate traces of the same session: 11l4aq60, 2l3ep1gw, 5f6mgjiz, csl2tkee, sw66sjuv, zzr76lzb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1255.0,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-Flash-Next-NVFP4 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 nvfp4; 64k 1,255 29.5 124.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3j20ahb8",
   "source_url": "https://omlx.ai/benchmarks/performance/3j20ahb8",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3j20ahb8-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 127.9 GB at this context; duplicate traces of the same session: 11l4aq60, 2l3ep1gw, 5f6mgjiz, csl2tkee, sw66sjuv, zzr76lzb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1235.0,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-Flash-Next-NVFP4 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 nvfp4; 128k 1,235 31.0 127.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3j20ahb8",
   "source_url": "https://omlx.ai/benchmarks/performance/3j20ahb8",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3j20ahb8-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 131.8 GB at this context; duplicate traces of the same session: 11l4aq60, 2l3ep1gw, 5f6mgjiz, csl2tkee, sw66sjuv, zzr76lzb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1220.0,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-Flash-Next-NVFP4 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 nvfp4; 195k 1,220 30.9 131.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3j20ahb8",
   "source_url": "https://omlx.ai/benchmarks/performance/3j20ahb8",
   "tg_tokens": null,
   "tps": 30.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hnmr1a6x-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 71.1 GB; TTFT 1961 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2089.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 2,089 PP tok/s 58.9 TG tok/s 1961 TTFT (ms) 71.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hnmr1a6x",
   "source_url": "https://omlx.ai/benchmarks/performance/hnmr1a6x",
   "tg_tokens": null,
   "tps": 58.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-r0z8usm2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "Youssofal MTPLX Optimized Quality FP16; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: 1qh5remo; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 325.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Youssofal--Qwen3.8-27B-MTPLX-Optimized-Quality-FP16 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 325.9 34.8 37.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r0z8usm2",
   "source_url": "https://omlx.ai/benchmarks/performance/r0z8usm2",
   "tg_tokens": null,
   "tps": 34.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-r0z8usm2-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "Youssofal MTPLX Optimized Quality FP16; oMLX community benchmark trace; peak memory 38.6 GB at this context; duplicate traces of the same session: 1qh5remo; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 313.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Youssofal--Qwen3.8-27B-MTPLX-Optimized-Quality-FP16 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 16k 313.5 52.8 38.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r0z8usm2",
   "source_url": "https://omlx.ai/benchmarks/performance/r0z8usm2",
   "tg_tokens": null,
   "tps": 52.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-us0r55ds-ctx1024",
   "model": "Qwen3.6-35B-A3B-OptiQ",
   "notes": "dawncr0w; oMLX community benchmark trace, single run at 1k context; peak memory 19.7 GB; TTFT 620 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1653.0,
   "provenance": "community",
   "quant": "5bpw",
   "quote": "dawncr0w--Qwen3.6-35B-A3B-OptiQ-5bpw-MLX on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 1,653 PP tok/s 113.2 TG tok/s 620 TTFT (ms) 19.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/us0r55ds",
   "source_url": "https://omlx.ai/benchmarks/performance/us0r55ds",
   "tg_tokens": null,
   "tps": 113.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-nu921gtx-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: dloymqc7, pz403ox9, xyvvi3mo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1432.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "gemma-4-26B-A4B-it-oQ5 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 5bit; 1k 1,432 73.7 18.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nu921gtx",
   "source_url": "https://omlx.ai/benchmarks/performance/nu921gtx",
   "tg_tokens": null,
   "tps": 73.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-nu921gtx-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: dloymqc7, pz403ox9, xyvvi3mo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1458.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "gemma-4-26B-A4B-it-oQ5 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 5bit; 4k 1,458 79.9 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nu921gtx",
   "source_url": "https://omlx.ai/benchmarks/performance/nu921gtx",
   "tg_tokens": null,
   "tps": 79.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-nu921gtx-ctx131072",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 30.9 GB at this context; duplicate traces of the same session: dloymqc7, pz403ox9, xyvvi3mo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 635.7,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "gemma-4-26B-A4B-it-oQ5 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 5bit; 128k 635.7 52.7 30.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nu921gtx",
   "source_url": "https://omlx.ai/benchmarks/performance/nu921gtx",
   "tg_tokens": null,
   "tps": 52.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-nu921gtx-ctx199680",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 36.6 GB at this context; duplicate traces of the same session: dloymqc7, pz403ox9, xyvvi3mo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 484.2,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "gemma-4-26B-A4B-it-oQ5 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 5bit; 195k 484.2 52.9 36.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nu921gtx",
   "source_url": "https://omlx.ai/benchmarks/performance/nu921gtx",
   "tg_tokens": null,
   "tps": 52.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-v8skgzr8-ctx1024",
   "model": "Huihui-gemma-4-E4B-it-abliterated",
   "notes": "oMLX community benchmark trace; peak memory 6.0 GB at this context; duplicate traces of the same session: o1ufbafg, tomy9n0r; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 548.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Huihui-gemma-4-E4B-it-abliterated-4bit-mlx on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 548.4 21.3 6.0 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v8skgzr8",
   "source_url": "https://omlx.ai/benchmarks/performance/v8skgzr8",
   "tg_tokens": null,
   "tps": 21.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-v8skgzr8-ctx8192",
   "model": "Huihui-gemma-4-E4B-it-abliterated",
   "notes": "oMLX community benchmark trace; peak memory 6.1 GB at this context; duplicate traces of the same session: o1ufbafg, tomy9n0r; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 539.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Huihui-gemma-4-E4B-it-abliterated-4bit-mlx on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 8k 539.5 20.7 6.1 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v8skgzr8",
   "source_url": "https://omlx.ai/benchmarks/performance/v8skgzr8",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-v8skgzr8-ctx32768",
   "model": "Huihui-gemma-4-E4B-it-abliterated",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: o1ufbafg, tomy9n0r; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 462.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Huihui-gemma-4-E4B-it-abliterated-4bit-mlx on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 32k 462.2 15.9 6.7 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v8skgzr8",
   "source_url": "https://omlx.ai/benchmarks/performance/v8skgzr8",
   "tg_tokens": null,
   "tps": 15.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-d98cimlr-ctx1024",
   "model": "MiMo-7B-RL",
   "notes": "oMLX community benchmark trace; peak memory 4.7 GB at this context; duplicate traces of the same session: 4tfehd1u, rl6ndzf2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1261.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-7B-RL-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 1,261 120.3 4.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d98cimlr",
   "source_url": "https://omlx.ai/benchmarks/performance/d98cimlr",
   "tg_tokens": null,
   "tps": 120.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-d98cimlr-ctx4096",
   "model": "MiMo-7B-RL",
   "notes": "oMLX community benchmark trace; peak memory 5.2 GB at this context; duplicate traces of the same session: 4tfehd1u, rl6ndzf2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1221.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-7B-RL-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 1,221 107.6 5.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d98cimlr",
   "source_url": "https://omlx.ai/benchmarks/performance/d98cimlr",
   "tg_tokens": null,
   "tps": 107.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-d98cimlr-ctx8192",
   "model": "MiMo-7B-RL",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: 4tfehd1u, rl6ndzf2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1150.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-7B-RL-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 1,150 93.9 5.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d98cimlr",
   "source_url": "https://omlx.ai/benchmarks/performance/d98cimlr",
   "tg_tokens": null,
   "tps": 93.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-6fobfy1l-ctx1024",
   "model": "Qwen2.5-Coder-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 4.5 GB at this context; duplicate traces of the same session: abz2s672; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 148.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-7B-Instruct-4bit on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 4bit; 1k 148.9 14.6 4.5 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6fobfy1l",
   "source_url": "https://omlx.ai/benchmarks/performance/6fobfy1l",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-6fobfy1l-ctx4096",
   "model": "Qwen2.5-Coder-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: abz2s672; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 143.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-7B-Instruct-4bit on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 4bit; 4k 143.5 13.9 4.9 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6fobfy1l",
   "source_url": "https://omlx.ai/benchmarks/performance/6fobfy1l",
   "tg_tokens": null,
   "tps": 13.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-0i80k02q-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: 4obhm95h; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 181.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 4bit; 1k 181.2 22.8 3.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0i80k02q",
   "source_url": "https://omlx.ai/benchmarks/performance/0i80k02q",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-0i80k02q-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.8 GB at this context; duplicate traces of the same session: 4obhm95h; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 168.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 4bit; 4k 168.7 22.0 3.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0i80k02q",
   "source_url": "https://omlx.ai/benchmarks/performance/0i80k02q",
   "tg_tokens": null,
   "tps": 22.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wxhb52fe-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 1kxicrcf, ewsm22ks, faw54eja; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4470.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,470 131.7 20.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wxhb52fe",
   "source_url": "https://omlx.ai/benchmarks/performance/wxhb52fe",
   "tg_tokens": null,
   "tps": 131.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wxhb52fe-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 1kxicrcf, ewsm22ks, faw54eja; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4293.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 4,293 113.4 22.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wxhb52fe",
   "source_url": "https://omlx.ai/benchmarks/performance/wxhb52fe",
   "tg_tokens": null,
   "tps": 113.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wxhb52fe-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.9 GB at this context; duplicate traces of the same session: 1kxicrcf, ewsm22ks, faw54eja; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1978.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 1,978 63.2 26.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wxhb52fe",
   "source_url": "https://omlx.ai/benchmarks/performance/wxhb52fe",
   "tg_tokens": null,
   "tps": 63.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wxhb52fe-ctx199680",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.2 GB at this context; duplicate traces of the same session: 1kxicrcf, ewsm22ks, faw54eja; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1001.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 1,001 31.9 30.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wxhb52fe",
   "source_url": "https://omlx.ai/benchmarks/performance/wxhb52fe",
   "tg_tokens": null,
   "tps": 31.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ra1fkthq-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: 4h5vn2tb, 5h78fwlv, fr99n6i1, qyo3appj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3918.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 3,918 107.8 27.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ra1fkthq",
   "source_url": "https://omlx.ai/benchmarks/performance/ra1fkthq",
   "tg_tokens": null,
   "tps": 107.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ra1fkthq-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.1 GB at this context; duplicate traces of the same session: 4h5vn2tb, 5h78fwlv, fr99n6i1, qyo3appj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4882.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 4,882 99.4 29.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ra1fkthq",
   "source_url": "https://omlx.ai/benchmarks/performance/ra1fkthq",
   "tg_tokens": null,
   "tps": 99.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ra1fkthq-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: 4h5vn2tb, 5h78fwlv, fr99n6i1, qyo3appj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3731.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 3,731 93.0 29.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ra1fkthq",
   "source_url": "https://omlx.ai/benchmarks/performance/ra1fkthq",
   "tg_tokens": null,
   "tps": 93.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ra1fkthq-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: 4h5vn2tb, 5h78fwlv, fr99n6i1, qyo3appj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1892.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 128k 1,892 61.3 34.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ra1fkthq",
   "source_url": "https://omlx.ai/benchmarks/performance/ra1fkthq",
   "tg_tokens": null,
   "tps": 61.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ra1fkthq-ctx199680",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.0 GB at this context; duplicate traces of the same session: 4h5vn2tb, 5h78fwlv, fr99n6i1, qyo3appj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1058.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 195k 1,058 31.2 38.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ra1fkthq",
   "source_url": "https://omlx.ai/benchmarks/performance/ra1fkthq",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ser3x6tk-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: jhipx7gx; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 106.0 15.3 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ser3x6tk",
   "source_url": "https://omlx.ai/benchmarks/performance/ser3x6tk",
   "tg_tokens": null,
   "tps": 15.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ser3x6tk-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: jhipx7gx; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 111.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 111.1 14.9 19.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ser3x6tk",
   "source_url": "https://omlx.ai/benchmarks/performance/ser3x6tk",
   "tg_tokens": null,
   "tps": 14.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-jjby0tlh-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "Youssofal MTPLX Optimized Quality; oMLX community benchmark trace, single run at 4k context; peak memory 29.6 GB; TTFT 15714 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 260.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Youssofal--Qwen3.8-27B-MTPLX-Optimized-Quality on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 8bit; 4k tokens 260.7 PP tok/s 16.6 TG tok/s 15714 TTFT (ms) 29.6 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jjby0tlh",
   "source_url": "https://omlx.ai/benchmarks/performance/jjby0tlh",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-wxdspotv-ctx1024",
   "model": "gemma-4-26B-A4B-it-ultra-uncensored-heretic",
   "notes": "culturerevolt; oMLX community benchmark trace, single run at 1k context; peak memory 14.1 GB; TTFT 730 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1404.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "culturerevolt--gemma-4-26B-A4B-it-ultra-uncensored-heretic-mlx-4Bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 1,404 PP tok/s 100.9 TG tok/s 730 TTFT (ms) 14.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wxdspotv",
   "source_url": "https://omlx.ai/benchmarks/performance/wxdspotv",
   "tg_tokens": null,
   "tps": 100.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-41ephd2g-ctx65536",
   "model": "Qwen3.6-35B-A3B-OptiQ",
   "notes": "dawncr0w; oMLX community benchmark trace, single run at 64k context; peak memory 23.8 GB; TTFT 63962 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1025.0,
   "provenance": "community",
   "quant": "5bpw",
   "quote": "dawncr0w--Qwen3.6-35B-A3B-OptiQ-5bpw-MLX on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k tokens 1,025 PP tok/s 72.1 TG tok/s 63962 TTFT (ms) 23.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/41ephd2g",
   "source_url": "https://omlx.ai/benchmarks/performance/41ephd2g",
   "tg_tokens": null,
   "tps": 72.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-140pgd4m-ctx4096",
   "model": "Qwen3.6-35B-A3B-OptiQ",
   "notes": "dawncr0w; oMLX community benchmark trace, single run at 4k context; peak memory 21.6 GB; TTFT 2012 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2036.0,
   "provenance": "community",
   "quant": "5bpw",
   "quote": "dawncr0w--Qwen3.6-35B-A3B-OptiQ-5bpw-MLX on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 2,036 PP tok/s 111.6 TG tok/s 2012 TTFT (ms) 21.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/140pgd4m",
   "source_url": "https://omlx.ai/benchmarks/performance/140pgd4m",
   "tg_tokens": null,
   "tps": 111.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-l66ni98o-ctx8192",
   "model": "Qwen3.6-35B-A3B-OptiQ",
   "notes": "dawncr0w; oMLX community benchmark trace, single run at 8k context; peak memory 21.6 GB; TTFT 4251 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1927.0,
   "provenance": "community",
   "quant": "5bpw",
   "quote": "dawncr0w--Qwen3.6-35B-A3B-OptiQ-5bpw-MLX on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k tokens 1,927 PP tok/s 102.5 TG tok/s 4251 TTFT (ms) 21.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l66ni98o",
   "source_url": "https://omlx.ai/benchmarks/performance/l66ni98o",
   "tg_tokens": null,
   "tps": 102.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-iqcuxt9j-ctx16384",
   "model": "Qwen3.6-35B-A3B-OptiQ",
   "notes": "dawncr0w; oMLX community benchmark trace, single run at 16k context; peak memory 21.9 GB; TTFT 9528 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1720.0,
   "provenance": "community",
   "quant": "5bpw",
   "quote": "dawncr0w--Qwen3.6-35B-A3B-OptiQ-5bpw-MLX on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k tokens 1,720 PP tok/s 97.5 TG tok/s 9528 TTFT (ms) 21.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iqcuxt9j",
   "source_url": "https://omlx.ai/benchmarks/performance/iqcuxt9j",
   "tg_tokens": null,
   "tps": 97.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-gm4kk4va-ctx32768",
   "model": "Qwen3.6-35B-A3B-OptiQ",
   "notes": "dawncr0w; oMLX community benchmark trace, single run at 32k context; peak memory 22.5 GB; TTFT 23785 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1378.0,
   "provenance": "community",
   "quant": "5bpw",
   "quote": "dawncr0w--Qwen3.6-35B-A3B-OptiQ-5bpw-MLX on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k tokens 1,378 PP tok/s 86.7 TG tok/s 23785 TTFT (ms) 22.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gm4kk4va",
   "source_url": "https://omlx.ai/benchmarks/performance/gm4kk4va",
   "tg_tokens": null,
   "tps": 86.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-p1iukvji-ctx1024",
   "model": "Qwen3.6-35B-A3B-OptiQ",
   "notes": "dawncr0w; oMLX community benchmark trace, single run at 1k context; peak memory 19.7 GB; TTFT 597 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1714.0,
   "provenance": "community",
   "quant": "5bpw",
   "quote": "dawncr0w--Qwen3.6-35B-A3B-OptiQ-5bpw-MLX on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 1,714 PP tok/s 109.8 TG tok/s 597 TTFT (ms) 19.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p1iukvji",
   "source_url": "https://omlx.ai/benchmarks/performance/p1iukvji",
   "tg_tokens": null,
   "tps": 109.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-maeb4gd7-ctx1024",
   "model": "Qwen3.6-35B-A3B-OptiQ",
   "notes": "dawncr0w; oMLX community benchmark trace, single run at 1k context; peak memory 19.7 GB; TTFT 590 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1736.0,
   "provenance": "community",
   "quant": "5bpw",
   "quote": "dawncr0w--Qwen3.6-35B-A3B-OptiQ-5bpw-MLX on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 1,736 PP tok/s 112.8 TG tok/s 590 TTFT (ms) 19.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/maeb4gd7",
   "source_url": "https://omlx.ai/benchmarks/performance/maeb4gd7",
   "tg_tokens": null,
   "tps": 112.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-xg6nu6bd-ctx1024",
   "model": "Qwen3.6-35B-A3B-OptiQ",
   "notes": "dawncr0w; oMLX community benchmark trace, single run at 1k context; peak memory 19.7 GB; TTFT 584 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1754.0,
   "provenance": "community",
   "quant": "5bpw",
   "quote": "dawncr0w--Qwen3.6-35B-A3B-OptiQ-5bpw-MLX on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 1,754 PP tok/s 112.3 TG tok/s 584 TTFT (ms) 19.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xg6nu6bd",
   "source_url": "https://omlx.ai/benchmarks/performance/xg6nu6bd",
   "tg_tokens": null,
   "tps": 112.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-2lq1bimk-ctx1024",
   "model": "Qwen3.6-35B-A3B-OptiQ",
   "notes": "dawncr0w; oMLX community benchmark trace, single run at 1k context; peak memory 19.7 GB; TTFT 583 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1757.0,
   "provenance": "community",
   "quant": "5bpw",
   "quote": "dawncr0w--Qwen3.6-35B-A3B-OptiQ-5bpw-MLX on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 1,757 PP tok/s 113.1 TG tok/s 583 TTFT (ms) 19.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2lq1bimk",
   "source_url": "https://omlx.ai/benchmarks/performance/2lq1bimk",
   "tg_tokens": null,
   "tps": 113.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-xgwp9vyb-ctx1024",
   "model": "gemma-4-12B-it-qat",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 8.4 GB at this context; duplicate traces of the same session: beicj484; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-12B-it-qat-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 224.3 45.4 8.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xgwp9vyb",
   "source_url": "https://omlx.ai/benchmarks/performance/xgwp9vyb",
   "tg_tokens": null,
   "tps": 45.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-xgwp9vyb-ctx4096",
   "model": "gemma-4-12B-it-qat",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 8.8 GB at this context; duplicate traces of the same session: beicj484; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-12B-it-qat-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 220.5 29.3 8.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xgwp9vyb",
   "source_url": "https://omlx.ai/benchmarks/performance/xgwp9vyb",
   "tg_tokens": null,
   "tps": 29.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-m6x3yeyd-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Jundot build; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: gmuo4kdq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 749.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 749.4 49.1 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m6x3yeyd",
   "source_url": "https://omlx.ai/benchmarks/performance/m6x3yeyd",
   "tg_tokens": null,
   "tps": 49.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-m6x3yeyd-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Jundot build; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: gmuo4kdq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1935.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,935 57.5 72.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m6x3yeyd",
   "source_url": "https://omlx.ai/benchmarks/performance/m6x3yeyd",
   "tg_tokens": null,
   "tps": 57.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-665s8vhc-ctx1024",
   "model": "Ling-3.0-tiny",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 9.0 GB; TTFT 507 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2020.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ling-3.0-tiny-8bit-XL-mlx on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 8bit; 1k tokens 2,020 PP tok/s 109.8 TG tok/s 507 TTFT (ms) 9.0 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/665s8vhc",
   "source_url": "https://omlx.ai/benchmarks/performance/665s8vhc",
   "tg_tokens": null,
   "tps": 109.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-qxeig67d-ctx4096",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "g64; oMLX community benchmark trace; peak memory 9.8 GB at this context; duplicate traces of the same session: pvgphqod; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 449.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-mlx-6bit-g64 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 6bit; 4k 449.4 35.9 9.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qxeig67d",
   "source_url": "https://omlx.ai/benchmarks/performance/qxeig67d",
   "tg_tokens": null,
   "tps": 35.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-qxeig67d-ctx16384",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "g64; oMLX community benchmark trace; peak memory 9.9 GB at this context; duplicate traces of the same session: pvgphqod; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 371.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-mlx-6bit-g64 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 6bit; 16k 371.7 30.4 9.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qxeig67d",
   "source_url": "https://omlx.ai/benchmarks/performance/qxeig67d",
   "tg_tokens": null,
   "tps": 30.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-b48z9gcy-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.2 GB at this context; duplicate traces of the same session: 6tyfwzb8, vaf9ybuj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 856.3,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Ornith-1.5-35B-A3B-oQ3e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 3bit; 1k 856.3 83.9 16.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b48z9gcy",
   "source_url": "https://omlx.ai/benchmarks/performance/b48z9gcy",
   "tg_tokens": null,
   "tps": 83.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-b48z9gcy-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: 6tyfwzb8, vaf9ybuj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 859.3,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Ornith-1.5-35B-A3B-oQ3e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 3bit; 16k 859.3 71.6 17.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b48z9gcy",
   "source_url": "https://omlx.ai/benchmarks/performance/b48z9gcy",
   "tg_tokens": null,
   "tps": 71.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-b48z9gcy-ctx131072",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: 6tyfwzb8, vaf9ybuj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 353.4,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Ornith-1.5-35B-A3B-oQ3e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 3bit; 128k 353.4 35.5 20.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b48z9gcy",
   "source_url": "https://omlx.ai/benchmarks/performance/b48z9gcy",
   "tg_tokens": null,
   "tps": 35.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ia91x84t-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: k0zvu6j3, p4p4guy1; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 884.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k 884.6 58.8 17.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ia91x84t",
   "source_url": "https://omlx.ai/benchmarks/performance/ia91x84t",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ia91x84t-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: k0zvu6j3, p4p4guy1; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1436.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 8k 1,436 104.7 19.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ia91x84t",
   "source_url": "https://omlx.ai/benchmarks/performance/ia91x84t",
   "tg_tokens": null,
   "tps": 104.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ia91x84t-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: k0zvu6j3, p4p4guy1; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1381.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 32k 1,381 124.4 20.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ia91x84t",
   "source_url": "https://omlx.ai/benchmarks/performance/ia91x84t",
   "tg_tokens": null,
   "tps": 124.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-of49wgr5-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 42.9 GB; TTFT 16366 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2002.0,
   "provenance": "community",
   "quant": "GSQ-RCO-Q2",
   "quote": "Qwen3.8-Flash-Next-GSQ-RCO-Q2-mlx on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k tokens 2,002 PP tok/s 32.6 TG tok/s 16366 TTFT (ms) 42.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/of49wgr5",
   "source_url": "https://omlx.ai/benchmarks/performance/of49wgr5",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6c7tvs0q-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 40.5 GB; TTFT 1375 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 744.9,
   "provenance": "community",
   "quant": "GSQ-RCO-Q2",
   "quote": "Qwen3.8-Flash-Next-GSQ-RCO-Q2-mlx on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 744.9 PP tok/s 57.6 TG tok/s 1375 TTFT (ms) 40.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6c7tvs0q",
   "source_url": "https://omlx.ai/benchmarks/performance/6c7tvs0q",
   "tg_tokens": null,
   "tps": 57.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-mvcgbnyf-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.7 GB at this context; duplicate traces of the same session: 3lds94vb, 8zt5n1z1, dg0e3tio, xct9w7na; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2548.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 2,548 133.7 100.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mvcgbnyf",
   "source_url": "https://omlx.ai/benchmarks/performance/mvcgbnyf",
   "tg_tokens": null,
   "tps": 133.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-mvcgbnyf-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.2 GB at this context; duplicate traces of the same session: 3lds94vb, 8zt5n1z1, dg0e3tio, xct9w7na; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5062.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 5,062 118.2 102.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mvcgbnyf",
   "source_url": "https://omlx.ai/benchmarks/performance/mvcgbnyf",
   "tg_tokens": null,
   "tps": 118.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-mvcgbnyf-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.6 GB at this context; duplicate traces of the same session: 3lds94vb, 8zt5n1z1, dg0e3tio, xct9w7na; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5001.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 5,001 144.4 104.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mvcgbnyf",
   "source_url": "https://omlx.ai/benchmarks/performance/mvcgbnyf",
   "tg_tokens": null,
   "tps": 144.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-mvcgbnyf-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 108.8 GB at this context; duplicate traces of the same session: 3lds94vb, 8zt5n1z1, dg0e3tio, xct9w7na; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4859.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 4,859 119.9 108.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mvcgbnyf",
   "source_url": "https://omlx.ai/benchmarks/performance/mvcgbnyf",
   "tg_tokens": null,
   "tps": 119.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-mvcgbnyf-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 111.1 GB at this context; duplicate traces of the same session: 3lds94vb, 8zt5n1z1, dg0e3tio, xct9w7na; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4716.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 4,716 105.8 111.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mvcgbnyf",
   "source_url": "https://omlx.ai/benchmarks/performance/mvcgbnyf",
   "tg_tokens": null,
   "tps": 105.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ap7tulga-ctx1024",
   "model": "gemma-4-26B-A4B-it-heretic-msq",
   "notes": "oMLX community benchmark trace; peak memory 9.2 GB at this context; duplicate traces of the same session: 71m7gsxp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1849.0,
   "provenance": "community",
   "quant": "2.6bit",
   "quote": "gemma-4-26B-A4B-it-heretic-msq-2.6bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 1,849 71.4 9.2 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ap7tulga",
   "source_url": "https://omlx.ai/benchmarks/performance/ap7tulga",
   "tg_tokens": null,
   "tps": 71.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ap7tulga-ctx4096",
   "model": "gemma-4-26B-A4B-it-heretic-msq",
   "notes": "oMLX community benchmark trace; peak memory 9.5 GB at this context; duplicate traces of the same session: 71m7gsxp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1773.0,
   "provenance": "community",
   "quant": "2.6bit",
   "quote": "gemma-4-26B-A4B-it-heretic-msq-2.6bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 1,773 69.2 9.5 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ap7tulga",
   "source_url": "https://omlx.ai/benchmarks/performance/ap7tulga",
   "tg_tokens": null,
   "tps": 69.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-23czw72k-ctx1024",
   "model": "gemma-4-E4B-it",
   "notes": "oMLX community benchmark trace; peak memory 7.9 GB at this context; duplicate traces of the same session: o27wsseq; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 530.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-E4B-it on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 8bit; 1k 530.6 16.2 7.9 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/23czw72k",
   "source_url": "https://omlx.ai/benchmarks/performance/23czw72k",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-23czw72k-ctx4096",
   "model": "gemma-4-E4B-it",
   "notes": "oMLX community benchmark trace; peak memory 8.0 GB at this context; duplicate traces of the same session: o27wsseq; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 535.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-E4B-it on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 8bit; 4k 535.1 15.6 8.0 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/23czw72k",
   "source_url": "https://omlx.ai/benchmarks/performance/23czw72k",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-uywfbmzs-ctx1024",
   "model": "Ornith-1.5-35B-A3B-REAP192",
   "notes": "oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: d9pngdth, kbq5yc71; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 954.7,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Ornith-1.5-35B-A3B-REAP192-mxfp4-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 954.7 95.9 14.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uywfbmzs",
   "source_url": "https://omlx.ai/benchmarks/performance/uywfbmzs",
   "tg_tokens": null,
   "tps": 95.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-uywfbmzs-ctx16384",
   "model": "Ornith-1.5-35B-A3B-REAP192",
   "notes": "oMLX community benchmark trace; peak memory 15.4 GB at this context; duplicate traces of the same session: d9pngdth, kbq5yc71; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 890.6,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Ornith-1.5-35B-A3B-REAP192-mxfp4-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 890.6 80.0 15.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uywfbmzs",
   "source_url": "https://omlx.ai/benchmarks/performance/uywfbmzs",
   "tg_tokens": null,
   "tps": 80.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-uywfbmzs-ctx131072",
   "model": "Ornith-1.5-35B-A3B-REAP192",
   "notes": "oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: d9pngdth, kbq5yc71; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 352.7,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Ornith-1.5-35B-A3B-REAP192-mxfp4-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 128k 352.7 37.8 18.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uywfbmzs",
   "source_url": "https://omlx.ai/benchmarks/performance/uywfbmzs",
   "tg_tokens": null,
   "tps": 37.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-hnyqnw1m-ctx4096",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: 5zhc5fuv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 459.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 459.7 27.7 18.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hnyqnw1m",
   "source_url": "https://omlx.ai/benchmarks/performance/hnyqnw1m",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-hnyqnw1m-ctx32768",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: 5zhc5fuv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 388.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 388.6 22.5 20.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hnyqnw1m",
   "source_url": "https://omlx.ai/benchmarks/performance/hnyqnw1m",
   "tg_tokens": null,
   "tps": 22.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2y13msvz-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: 08adlfwb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 335.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 335.3 35.7 25.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2y13msvz",
   "source_url": "https://omlx.ai/benchmarks/performance/2y13msvz",
   "tg_tokens": null,
   "tps": 35.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2y13msvz-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: 08adlfwb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 268.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 268.5 46.1 26.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2y13msvz",
   "source_url": "https://omlx.ai/benchmarks/performance/2y13msvz",
   "tg_tokens": null,
   "tps": 46.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-rae2g2ad-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: itfq10s0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 473.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 473.4 42.2 17.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rae2g2ad",
   "source_url": "https://omlx.ai/benchmarks/performance/rae2g2ad",
   "tg_tokens": null,
   "tps": 42.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-rae2g2ad-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: itfq10s0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 479.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 479.8 36.0 18.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rae2g2ad",
   "source_url": "https://omlx.ai/benchmarks/performance/rae2g2ad",
   "tg_tokens": null,
   "tps": 36.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-tgs08iyz-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.2 GB at this context; duplicate traces of the same session: 69guzzi9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 826.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 826.3 46.7 101.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tgs08iyz",
   "source_url": "https://omlx.ai/benchmarks/performance/tgs08iyz",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-tgs08iyz-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: 69guzzi9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 778.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 778.4 58.2 101.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tgs08iyz",
   "source_url": "https://omlx.ai/benchmarks/performance/tgs08iyz",
   "tg_tokens": null,
   "tps": 58.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-my0kpyev-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: ecan2rzv, lb3bnt1h, uzydoypv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 254.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 254.4 53.2 21.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/my0kpyev",
   "source_url": "https://omlx.ai/benchmarks/performance/my0kpyev",
   "tg_tokens": null,
   "tps": 53.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-my0kpyev-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: ecan2rzv, lb3bnt1h, uzydoypv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 343.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 343.0 92.4 25.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/my0kpyev",
   "source_url": "https://omlx.ai/benchmarks/performance/my0kpyev",
   "tg_tokens": null,
   "tps": 92.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-my0kpyev-ctx131072",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.0 GB at this context; duplicate traces of the same session: ecan2rzv, lb3bnt1h, uzydoypv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 164.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 164.9 38.9 33.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/my0kpyev",
   "source_url": "https://omlx.ai/benchmarks/performance/my0kpyev",
   "tg_tokens": null,
   "tps": 38.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-my0kpyev-ctx199680",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: ecan2rzv, lb3bnt1h, uzydoypv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 135.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 195k 135.3 35.7 36.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/my0kpyev",
   "source_url": "https://omlx.ai/benchmarks/performance/my0kpyev",
   "tg_tokens": null,
   "tps": 35.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-n93cq8oy-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace; peak memory 25.2 GB at this context; duplicate traces of the same session: wv1kbgrl; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 150.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 4k 150.2 21.0 25.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n93cq8oy",
   "source_url": "https://omlx.ai/benchmarks/performance/n93cq8oy",
   "tg_tokens": null,
   "tps": 21.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-n93cq8oy-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace; peak memory 27.2 GB at this context; duplicate traces of the same session: wv1kbgrl; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 138.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 16k 138.1 19.3 27.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n93cq8oy",
   "source_url": "https://omlx.ai/benchmarks/performance/n93cq8oy",
   "tg_tokens": null,
   "tps": 19.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-yed9ivj5-ctx1024",
   "model": "gemma-4-12B-it-qat",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 8.4 GB at this context; duplicate traces of the same session: mb7spyfa, rk5xaf0p; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 209.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-12B-it-qat-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 209.4 27.7 8.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yed9ivj5",
   "source_url": "https://omlx.ai/benchmarks/performance/yed9ivj5",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-yed9ivj5-ctx4096",
   "model": "gemma-4-12B-it-qat",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 8.8 GB at this context; duplicate traces of the same session: mb7spyfa, rk5xaf0p; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 218.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-12B-it-qat-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 218.2 34.9 8.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yed9ivj5",
   "source_url": "https://omlx.ai/benchmarks/performance/yed9ivj5",
   "tg_tokens": null,
   "tps": 34.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-yed9ivj5-ctx8192",
   "model": "gemma-4-12B-it-qat",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 9.0 GB at this context; duplicate traces of the same session: mb7spyfa, rk5xaf0p; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 216.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-12B-it-qat-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 216.0 37.2 9.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yed9ivj5",
   "source_url": "https://omlx.ai/benchmarks/performance/yed9ivj5",
   "tg_tokens": null,
   "tps": 37.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-eoqeh014-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 7.9 GB at this context; duplicate traces of the same session: soombllg; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 519.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-E4B-it-MLX-8bit on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 8bit; 1k 519.1 15.6 7.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eoqeh014",
   "source_url": "https://omlx.ai/benchmarks/performance/eoqeh014",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-eoqeh014-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 7.9 GB at this context; duplicate traces of the same session: soombllg; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 542.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-E4B-it-MLX-8bit on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 8bit; 4k 542.5 15.4 7.9 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eoqeh014",
   "source_url": "https://omlx.ai/benchmarks/performance/eoqeh014",
   "tg_tokens": null,
   "tps": 15.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-tf8j18uz-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.2 GB at this context; duplicate traces of the same session: 0y7vsvrd, 58gk3of0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 876.8,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Ornith-1.5-35B-A3B-oQ3e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 3bit; 1k 876.8 80.4 16.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tf8j18uz",
   "source_url": "https://omlx.ai/benchmarks/performance/tf8j18uz",
   "tg_tokens": null,
   "tps": 80.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-tf8j18uz-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: 0y7vsvrd, 58gk3of0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 832.8,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Ornith-1.5-35B-A3B-oQ3e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 3bit; 16k 832.8 59.8 17.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tf8j18uz",
   "source_url": "https://omlx.ai/benchmarks/performance/tf8j18uz",
   "tg_tokens": null,
   "tps": 59.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-tf8j18uz-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: 0y7vsvrd, 58gk3of0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 510.6,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Ornith-1.5-35B-A3B-oQ3e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 3bit; 64k 510.6 39.6 18.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tf8j18uz",
   "source_url": "https://omlx.ai/benchmarks/performance/tf8j18uz",
   "tg_tokens": null,
   "tps": 39.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-qs6q1894-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: tyvqnh0s; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 148.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 4k 148.5 24.0 19.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qs6q1894",
   "source_url": "https://omlx.ai/benchmarks/performance/qs6q1894",
   "tg_tokens": null,
   "tps": 24.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-qs6q1894-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: tyvqnh0s; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 132.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 16k 132.8 21.5 21.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qs6q1894",
   "source_url": "https://omlx.ai/benchmarks/performance/qs6q1894",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-wmnlyjmf-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: b3uudtv3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 431.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 431.1 32.5 25.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wmnlyjmf",
   "source_url": "https://omlx.ai/benchmarks/performance/wmnlyjmf",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-wmnlyjmf-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: b3uudtv3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 381.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 381.9 45.9 27.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wmnlyjmf",
   "source_url": "https://omlx.ai/benchmarks/performance/wmnlyjmf",
   "tg_tokens": null,
   "tps": 45.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-2aly2rgy-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: w4o0jxiv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 598.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 598.4 42.3 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2aly2rgy",
   "source_url": "https://omlx.ai/benchmarks/performance/2aly2rgy",
   "tg_tokens": null,
   "tps": 42.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-2aly2rgy-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: w4o0jxiv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 695.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 695.2 35.7 19.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2aly2rgy",
   "source_url": "https://omlx.ai/benchmarks/performance/2aly2rgy",
   "tg_tokens": null,
   "tps": 35.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-24t7hqqk-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 2gln3izt, p4kf57hw; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1530.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 8k 1,530 90.0 19.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/24t7hqqk",
   "source_url": "https://omlx.ai/benchmarks/performance/24t7hqqk",
   "tg_tokens": null,
   "tps": 90.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-24t7hqqk-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: 2gln3izt, p4kf57hw; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1385.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 32k 1,385 199.4 20.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/24t7hqqk",
   "source_url": "https://omlx.ai/benchmarks/performance/24t7hqqk",
   "tg_tokens": null,
   "tps": 199.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-24t7hqqk-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.6 GB at this context; duplicate traces of the same session: 2gln3izt, p4kf57hw; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 931.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 128k 931.0 81.5 27.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/24t7hqqk",
   "source_url": "https://omlx.ai/benchmarks/performance/24t7hqqk",
   "tg_tokens": null,
   "tps": 81.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-cwwcit0a-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 182.7 GB at this context; duplicate traces of the same session: ks9d1w08, voj3fu7e; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2347.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,347 114.8 182.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cwwcit0a",
   "source_url": "https://omlx.ai/benchmarks/performance/cwwcit0a",
   "tg_tokens": null,
   "tps": 114.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-cwwcit0a-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 184.0 GB at this context; duplicate traces of the same session: ks9d1w08, voj3fu7e; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2323.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,323 84.7 184.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cwwcit0a",
   "source_url": "https://omlx.ai/benchmarks/performance/cwwcit0a",
   "tg_tokens": null,
   "tps": 84.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-cwwcit0a-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 186.9 GB at this context; duplicate traces of the same session: ks9d1w08, voj3fu7e; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2279.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,279 144.3 186.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cwwcit0a",
   "source_url": "https://omlx.ai/benchmarks/performance/cwwcit0a",
   "tg_tokens": null,
   "tps": 144.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-05tp4l43-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.6 GB at this context; duplicate traces of the same session: jn3zc001, nl4mngt2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2306.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,306 89.9 180.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/05tp4l43",
   "source_url": "https://omlx.ai/benchmarks/performance/05tp4l43",
   "tg_tokens": null,
   "tps": 89.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-05tp4l43-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 182.2 GB at this context; duplicate traces of the same session: jn3zc001, nl4mngt2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2287.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,287 82.4 182.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/05tp4l43",
   "source_url": "https://omlx.ai/benchmarks/performance/05tp4l43",
   "tg_tokens": null,
   "tps": 82.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-05tp4l43-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 185.7 GB at this context; duplicate traces of the same session: jn3zc001, nl4mngt2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2243.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,243 138.7 185.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/05tp4l43",
   "source_url": "https://omlx.ai/benchmarks/performance/05tp4l43",
   "tg_tokens": null,
   "tps": 138.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-288t40wn-ctx1024",
   "model": "GLM-5.3",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 316.4 GB at this context; duplicate traces of the same session: tn8uz75q; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 214.2,
   "provenance": "community",
   "quant": "mixed 3_6bit",
   "quote": "GLM-5.3-MLX-mixed-3_6bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 3bit; 1k 214.2 21.2 316.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/288t40wn",
   "source_url": "https://omlx.ai/benchmarks/performance/288t40wn",
   "tg_tokens": null,
   "tps": 21.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-288t40wn-ctx4096",
   "model": "GLM-5.3",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 318.3 GB at this context; duplicate traces of the same session: tn8uz75q; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 219.4,
   "provenance": "community",
   "quant": "mixed 3_6bit",
   "quote": "GLM-5.3-MLX-mixed-3_6bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 3bit; 4k 219.4 15.7 318.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/288t40wn",
   "source_url": "https://omlx.ai/benchmarks/performance/288t40wn",
   "tg_tokens": null,
   "tps": 15.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-288t40wn-ctx8192",
   "model": "GLM-5.3",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 319.6 GB at this context; duplicate traces of the same session: tn8uz75q; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 197.1,
   "provenance": "community",
   "quant": "mixed 3_6bit",
   "quote": "GLM-5.3-MLX-mixed-3_6bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 3bit; 8k 197.1 15.7 319.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/288t40wn",
   "source_url": "https://omlx.ai/benchmarks/performance/288t40wn",
   "tg_tokens": null,
   "tps": 15.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-df6ykp4p-ctx8192",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: gadyn0h7; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2441.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-35B-A3B-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 2,441 98.4 22.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/df6ykp4p",
   "source_url": "https://omlx.ai/benchmarks/performance/df6ykp4p",
   "tg_tokens": null,
   "tps": 98.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-df6ykp4p-ctx32768",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: gadyn0h7; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1883.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-35B-A3B-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 1,883 84.0 22.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/df6ykp4p",
   "source_url": "https://omlx.ai/benchmarks/performance/df6ykp4p",
   "tg_tokens": null,
   "tps": 84.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-o3sm6rp5-ctx4096",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: aizso3ww; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 635.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 635.0 30.5 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o3sm6rp5",
   "source_url": "https://omlx.ai/benchmarks/performance/o3sm6rp5",
   "tg_tokens": null,
   "tps": 30.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-o3sm6rp5-ctx32768",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: aizso3ww; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 585.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 585.5 24.8 20.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o3sm6rp5",
   "source_url": "https://omlx.ai/benchmarks/performance/o3sm6rp5",
   "tg_tokens": null,
   "tps": 24.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-7v61utjs-ctx4096",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 7of1ot9c; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 610.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 610.1 29.1 19.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7v61utjs",
   "source_url": "https://omlx.ai/benchmarks/performance/7v61utjs",
   "tg_tokens": null,
   "tps": 29.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-7v61utjs-ctx32768",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 7of1ot9c; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 575.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 575.6 23.5 20.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7v61utjs",
   "source_url": "https://omlx.ai/benchmarks/performance/7v61utjs",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-rqjett61-ctx4096",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: x7xkxa5i; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 620.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 620.6 30.1 19.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rqjett61",
   "source_url": "https://omlx.ai/benchmarks/performance/rqjett61",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-rqjett61-ctx32768",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: x7xkxa5i; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 570.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 570.9 22.8 20.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rqjett61",
   "source_url": "https://omlx.ai/benchmarks/performance/rqjett61",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1e4sj6g1-ctx4096",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: a13hd9x2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 451.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 451.7 23.9 21.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1e4sj6g1",
   "source_url": "https://omlx.ai/benchmarks/performance/1e4sj6g1",
   "tg_tokens": null,
   "tps": 23.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1e4sj6g1-ctx32768",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: a13hd9x2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 405.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 405.8 6.8 24.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1e4sj6g1",
   "source_url": "https://omlx.ai/benchmarks/performance/1e4sj6g1",
   "tg_tokens": null,
   "tps": 6.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3kyrx8lp-ctx4096",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: sa6si7gl; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 504.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 504.8 30.8 23.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3kyrx8lp",
   "source_url": "https://omlx.ai/benchmarks/performance/3kyrx8lp",
   "tg_tokens": null,
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3kyrx8lp-ctx32768",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: sa6si7gl; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 435.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 435.4 24.1 25.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3kyrx8lp",
   "source_url": "https://omlx.ai/benchmarks/performance/3kyrx8lp",
   "tg_tokens": null,
   "tps": 24.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-9jc1alqc-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: jf2uqxl5, q6dyktji; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1308.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k 1,308 106.3 17.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9jc1alqc",
   "source_url": "https://omlx.ai/benchmarks/performance/9jc1alqc",
   "tg_tokens": null,
   "tps": 106.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-9jc1alqc-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: jf2uqxl5, q6dyktji; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1540.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 8k 1,540 105.4 19.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9jc1alqc",
   "source_url": "https://omlx.ai/benchmarks/performance/9jc1alqc",
   "tg_tokens": null,
   "tps": 105.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-9jc1alqc-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: jf2uqxl5, q6dyktji; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1385.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 32k 1,385 201.1 20.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9jc1alqc",
   "source_url": "https://omlx.ai/benchmarks/performance/9jc1alqc",
   "tg_tokens": null,
   "tps": 201.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-asjlgjy8-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 19.2 GB; TTFT 5326 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1538.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 8k tokens 1,538 PP tok/s 111.2 TG tok/s 5326 TTFT (ms) 19.2 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/asjlgjy8",
   "source_url": "https://omlx.ai/benchmarks/performance/asjlgjy8",
   "tg_tokens": null,
   "tps": 111.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-56zfdjik-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: r01bfdzt, wsiev98v; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1169.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 6bit; 1k 1,169 84.2 26.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/56zfdjik",
   "source_url": "https://omlx.ai/benchmarks/performance/56zfdjik",
   "tg_tokens": null,
   "tps": 84.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-56zfdjik-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: r01bfdzt, wsiev98v; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1343.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 6bit; 8k 1,343 91.6 28.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/56zfdjik",
   "source_url": "https://omlx.ai/benchmarks/performance/56zfdjik",
   "tg_tokens": null,
   "tps": 91.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-56zfdjik-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: r01bfdzt, wsiev98v; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1237.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 6bit; 32k 1,237 126.6 29.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/56zfdjik",
   "source_url": "https://omlx.ai/benchmarks/performance/56zfdjik",
   "tg_tokens": null,
   "tps": 126.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-01odk7tp-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 35.3 GB; TTFT 5970 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1372.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 8k tokens 1,372 PP tok/s 75.0 TG tok/s 5970 TTFT (ms) 35.3 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/01odk7tp",
   "source_url": "https://omlx.ai/benchmarks/performance/01odk7tp",
   "tg_tokens": null,
   "tps": 75.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-0oj32f9h-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 33.9 GB; TTFT 907 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1129.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 1k tokens 1,129 PP tok/s 80.1 TG tok/s 907 TTFT (ms) 33.9 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0oj32f9h",
   "source_url": "https://omlx.ai/benchmarks/performance/0oj32f9h",
   "tg_tokens": null,
   "tps": 80.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-w7y6zejo-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 33.9 GB; TTFT 909 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1127.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 1k tokens 1,127 PP tok/s 75.8 TG tok/s 909 TTFT (ms) 33.9 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w7y6zejo",
   "source_url": "https://omlx.ai/benchmarks/performance/w7y6zejo",
   "tg_tokens": null,
   "tps": 75.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-3c5rim61-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 33.9 GB; TTFT 904 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1133.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 1k tokens 1,133 PP tok/s 86.8 TG tok/s 904 TTFT (ms) 33.9 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3c5rim61",
   "source_url": "https://omlx.ai/benchmarks/performance/3c5rim61",
   "tg_tokens": null,
   "tps": 86.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-emvad53z-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 35.4 GB; TTFT 5972 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1372.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 8k tokens 1,372 PP tok/s 88.3 TG tok/s 5972 TTFT (ms) 35.4 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/emvad53z",
   "source_url": "https://omlx.ai/benchmarks/performance/emvad53z",
   "tg_tokens": null,
   "tps": 88.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-0qss0evy-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 35.1 GB; TTFT 3005 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1363.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 4k tokens 1,363 PP tok/s 73.7 TG tok/s 3005 TTFT (ms) 35.1 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0qss0evy",
   "source_url": "https://omlx.ai/benchmarks/performance/0qss0evy",
   "tg_tokens": null,
   "tps": 73.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-m3xzhr5c-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: e78o0zsi, sbg7n6u5; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1142.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 1k 1,142 80.9 33.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m3xzhr5c",
   "source_url": "https://omlx.ai/benchmarks/performance/m3xzhr5c",
   "tg_tokens": null,
   "tps": 80.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-m3xzhr5c-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: e78o0zsi, sbg7n6u5; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1374.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 8k 1,374 78.3 35.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m3xzhr5c",
   "source_url": "https://omlx.ai/benchmarks/performance/m3xzhr5c",
   "tg_tokens": null,
   "tps": 78.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-m3xzhr5c-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.8 GB at this context; duplicate traces of the same session: e78o0zsi, sbg7n6u5; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1256.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 32k 1,256 116.6 36.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m3xzhr5c",
   "source_url": "https://omlx.ai/benchmarks/performance/m3xzhr5c",
   "tg_tokens": null,
   "tps": 116.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-y9glbpzv-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 37.0 GB; TTFT 26189 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1251.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 32k tokens 1,251 PP tok/s 120.4 TG tok/s 26189 TTFT (ms) 37.0 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y9glbpzv",
   "source_url": "https://omlx.ai/benchmarks/performance/y9glbpzv",
   "tg_tokens": null,
   "tps": 120.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-5jbgg31w-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: olz3yzjq; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 654.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 1k 654.4 74.5 100.6 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5jbgg31w",
   "source_url": "https://omlx.ai/benchmarks/performance/5jbgg31w",
   "tg_tokens": null,
   "tps": 74.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-5jbgg31w-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.3 GB at this context; duplicate traces of the same session: olz3yzjq; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 665.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 4k 665.6 71.4 102.3 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5jbgg31w",
   "source_url": "https://omlx.ai/benchmarks/performance/5jbgg31w",
   "tg_tokens": null,
   "tps": 71.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-abla1983-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 23.1 GB; TTFT 13394 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 305.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 4k tokens 305.8 PP tok/s 47.0 TG tok/s 13394 TTFT (ms) 23.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/abla1983",
   "source_url": "https://omlx.ai/benchmarks/performance/abla1983",
   "tg_tokens": null,
   "tps": 47.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-lb2k9557-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace, single run at 4k context; peak memory 31.2 GB; TTFT 13993 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 292.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 6bit; 4k tokens 292.7 PP tok/s 26.3 TG tok/s 13993 TTFT (ms) 31.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lb2k9557",
   "source_url": "https://omlx.ai/benchmarks/performance/lb2k9557",
   "tg_tokens": null,
   "tps": 26.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-la803o11-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace, single run at 4k context; peak memory 26.2 GB; TTFT 14202 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 288.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 6bit; 4k tokens 288.4 PP tok/s 23.3 TG tok/s 14202 TTFT (ms) 26.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/la803o11",
   "source_url": "https://omlx.ai/benchmarks/performance/la803o11",
   "tg_tokens": null,
   "tps": 23.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3",
   "id": "omlx-u27ioq1k-ctx1024",
   "model": "gemma-4-E4B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 7.9 GB at this context; duplicate traces of the same session: gzhsxyff; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 539.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-E4B-it-qat-8bit on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 8bit; 1k 539.3 16.9 7.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u27ioq1k",
   "source_url": "https://omlx.ai/benchmarks/performance/u27ioq1k",
   "tg_tokens": null,
   "tps": 16.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3",
   "id": "omlx-u27ioq1k-ctx4096",
   "model": "gemma-4-E4B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 7.9 GB at this context; duplicate traces of the same session: gzhsxyff; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 539.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-E4B-it-qat-8bit on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 8bit; 4k 539.8 16.7 7.9 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u27ioq1k",
   "source_url": "https://omlx.ai/benchmarks/performance/u27ioq1k",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1un0vsbt-ctx1024",
   "model": "GLM-5.3",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 316.4 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 214.2,
   "provenance": "community",
   "quant": "mixed 3_6bit",
   "quote": "GLM-5.3-MLX-mixed-3_6bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 3bit; 1k 214.2 21.2 316.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1un0vsbt",
   "source_url": "https://omlx.ai/benchmarks/performance/1un0vsbt",
   "tg_tokens": null,
   "tps": 21.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1un0vsbt-ctx4096",
   "model": "GLM-5.3",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 318.3 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 219.4,
   "provenance": "community",
   "quant": "mixed 3_6bit",
   "quote": "GLM-5.3-MLX-mixed-3_6bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 3bit; 4k 219.4 15.7 318.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1un0vsbt",
   "source_url": "https://omlx.ai/benchmarks/performance/1un0vsbt",
   "tg_tokens": null,
   "tps": 15.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1un0vsbt-ctx8192",
   "model": "GLM-5.3",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 319.6 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 197.1,
   "provenance": "community",
   "quant": "mixed 3_6bit",
   "quote": "GLM-5.3-MLX-mixed-3_6bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 3bit; 8k 197.1 15.7 319.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1un0vsbt",
   "source_url": "https://omlx.ai/benchmarks/performance/1un0vsbt",
   "tg_tokens": null,
   "tps": 15.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-lmarzvzu-ctx1024",
   "model": "LFM2.5-8B-A1B-Abliterated",
   "notes": "PocketAiHub; oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: udyranog; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1864.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "PocketAiHub--LFM2.5-8B-A1B-Abliterated on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 1,864 129.8 5.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lmarzvzu",
   "source_url": "https://omlx.ai/benchmarks/performance/lmarzvzu",
   "tg_tokens": null,
   "tps": 129.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-lmarzvzu-ctx4096",
   "model": "LFM2.5-8B-A1B-Abliterated",
   "notes": "PocketAiHub; oMLX community benchmark trace; peak memory 6.0 GB at this context; duplicate traces of the same session: udyranog; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1535.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "PocketAiHub--LFM2.5-8B-A1B-Abliterated on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 1,535 20.7 6.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lmarzvzu",
   "source_url": "https://omlx.ai/benchmarks/performance/lmarzvzu",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ospeqh81-ctx8192",
   "model": "LFM2.5-8B-A1B-Abliterated",
   "notes": "PocketAiHub; oMLX community benchmark trace; peak memory 6.1 GB at this context; duplicate traces of the same session: 907xmqc8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1996.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "PocketAiHub--LFM2.5-8B-A1B-Abliterated on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 1,996 139.0 6.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ospeqh81",
   "source_url": "https://omlx.ai/benchmarks/performance/ospeqh81",
   "tg_tokens": null,
   "tps": 139.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ospeqh81-ctx16384",
   "model": "LFM2.5-8B-A1B-Abliterated",
   "notes": "PocketAiHub; oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: 907xmqc8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1878.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "PocketAiHub--LFM2.5-8B-A1B-Abliterated on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 1,878 126.5 6.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ospeqh81",
   "source_url": "https://omlx.ai/benchmarks/performance/ospeqh81",
   "tg_tokens": null,
   "tps": 126.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-765tncvr-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 24.7 GB; TTFT 4261 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1923.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 8k tokens 1,923 PP tok/s 72.2 TG tok/s 4261 TTFT (ms) 24.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/765tncvr",
   "source_url": "https://omlx.ai/benchmarks/performance/765tncvr",
   "tg_tokens": null,
   "tps": 72.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-ofybyb7k-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 26.8 GB; TTFT 23435 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1398.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 32k tokens 1,398 PP tok/s 149.1 TG tok/s 23435 TTFT (ms) 26.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ofybyb7k",
   "source_url": "https://omlx.ai/benchmarks/performance/ofybyb7k",
   "tg_tokens": null,
   "tps": 149.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-g9u409m9-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 24.4 GB; TTFT 2070 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1979.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 4k tokens 1,979 PP tok/s 104.0 TG tok/s 2070 TTFT (ms) 24.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g9u409m9",
   "source_url": "https://omlx.ai/benchmarks/performance/g9u409m9",
   "tg_tokens": null,
   "tps": 104.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-xm1k9833-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace, single run at 4k context; peak memory 28.9 GB; TTFT 2169 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1889.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 5bit; 4k tokens 1,889 PP tok/s 93.9 TG tok/s 2169 TTFT (ms) 28.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xm1k9833",
   "source_url": "https://omlx.ai/benchmarks/performance/xm1k9833",
   "tg_tokens": null,
   "tps": 93.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-lhbn9sl2-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace, single run at 8k context; peak memory 33.4 GB; TTFT 4456 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1839.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 6bit; 8k tokens 1,839 PP tok/s 61.7 TG tok/s 4456 TTFT (ms) 33.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lhbn9sl2",
   "source_url": "https://omlx.ai/benchmarks/performance/lhbn9sl2",
   "tg_tokens": null,
   "tps": 61.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-nbaama99-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace, single run at 32k context; peak memory 35.5 GB; TTFT 24238 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1352.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 6bit; 32k tokens 1,352 PP tok/s 123.2 TG tok/s 24238 TTFT (ms) 35.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nbaama99",
   "source_url": "https://omlx.ai/benchmarks/performance/nbaama99",
   "tg_tokens": null,
   "tps": 123.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-fd7rz2t2-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace, single run at 4k context; peak memory 33.1 GB; TTFT 2174 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1884.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 6bit; 4k tokens 1,884 PP tok/s 94.4 TG tok/s 2174 TTFT (ms) 33.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fd7rz2t2",
   "source_url": "https://omlx.ai/benchmarks/performance/fd7rz2t2",
   "tg_tokens": null,
   "tps": 94.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-s5fillje-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: b6h8ny7h, bx66vphi; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 106.6 15.6 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s5fillje",
   "source_url": "https://omlx.ai/benchmarks/performance/s5fillje",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-s5fillje-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: b6h8ny7h, bx66vphi; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 111.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 111.4 15.1 19.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s5fillje",
   "source_url": "https://omlx.ai/benchmarks/performance/s5fillje",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-s5fillje-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: b6h8ny7h, bx66vphi; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 16k 105.0 14.0 21.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s5fillje",
   "source_url": "https://omlx.ai/benchmarks/performance/s5fillje",
   "tg_tokens": null,
   "tps": 14.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-rvknvu1z-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.4 GB at this context; duplicate traces of the same session: ru80l7yw, y13uqeq7; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 1k 103.6 8.7 29.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rvknvu1z",
   "source_url": "https://omlx.ai/benchmarks/performance/rvknvu1z",
   "tg_tokens": null,
   "tps": 8.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-rvknvu1z-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.8 GB at this context; duplicate traces of the same session: ru80l7yw, y13uqeq7; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 98.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 4k 98.6 8.6 32.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rvknvu1z",
   "source_url": "https://omlx.ai/benchmarks/performance/rvknvu1z",
   "tg_tokens": null,
   "tps": 8.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-rvknvu1z-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: ru80l7yw, y13uqeq7; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 99.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 16k 99.2 8.2 34.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rvknvu1z",
   "source_url": "https://omlx.ai/benchmarks/performance/rvknvu1z",
   "tg_tokens": null,
   "tps": 8.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-i1is7kvl-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: hte0e0i2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 474.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 474.0 34.0 17.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i1is7kvl",
   "source_url": "https://omlx.ai/benchmarks/performance/i1is7kvl",
   "tg_tokens": null,
   "tps": 34.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-i1is7kvl-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: hte0e0i2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 480.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 480.6 35.0 18.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i1is7kvl",
   "source_url": "https://omlx.ai/benchmarks/performance/i1is7kvl",
   "tg_tokens": null,
   "tps": 35.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-5xw0fpzl-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 9p6elhhv, ejfaul5g; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 106.2 14.9 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5xw0fpzl",
   "source_url": "https://omlx.ai/benchmarks/performance/5xw0fpzl",
   "tg_tokens": null,
   "tps": 14.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-5xw0fpzl-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 9p6elhhv, ejfaul5g; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 109.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 109.0 14.5 20.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5xw0fpzl",
   "source_url": "https://omlx.ai/benchmarks/performance/5xw0fpzl",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-5xw0fpzl-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 9p6elhhv, ejfaul5g; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 16k 104.8 13.5 22.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5xw0fpzl",
   "source_url": "https://omlx.ai/benchmarks/performance/5xw0fpzl",
   "tg_tokens": null,
   "tps": 13.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-f1f4hthj-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: o0x213ej; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 602.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 602.4 43.8 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f1f4hthj",
   "source_url": "https://omlx.ai/benchmarks/performance/f1f4hthj",
   "tg_tokens": null,
   "tps": 43.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-f1f4hthj-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: o0x213ej; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 694.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 694.3 34.8 19.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f1f4hthj",
   "source_url": "https://omlx.ai/benchmarks/performance/f1f4hthj",
   "tg_tokens": null,
   "tps": 34.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-qeasg7aq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: tpyh1wlc, vgs7zy4j; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1305.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k 1,305 61.5 17.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qeasg7aq",
   "source_url": "https://omlx.ai/benchmarks/performance/qeasg7aq",
   "tg_tokens": null,
   "tps": 61.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-qeasg7aq-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: tpyh1wlc, vgs7zy4j; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1553.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 8k 1,553 65.6 19.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qeasg7aq",
   "source_url": "https://omlx.ai/benchmarks/performance/qeasg7aq",
   "tg_tokens": null,
   "tps": 65.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-qeasg7aq-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: tpyh1wlc, vgs7zy4j; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1391.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 32k 1,391 51.6 20.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qeasg7aq",
   "source_url": "https://omlx.ai/benchmarks/performance/qeasg7aq",
   "tg_tokens": null,
   "tps": 51.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-k9r2woub-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1297.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k 1,297 92.0 17.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k9r2woub",
   "source_url": "https://omlx.ai/benchmarks/performance/k9r2woub",
   "tg_tokens": null,
   "tps": 92.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-k9r2woub-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1537.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 8k 1,537 114.5 19.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k9r2woub",
   "source_url": "https://omlx.ai/benchmarks/performance/k9r2woub",
   "tg_tokens": null,
   "tps": 114.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-k9r2woub-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1391.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 32k 1,391 200.3 20.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k9r2woub",
   "source_url": "https://omlx.ai/benchmarks/performance/k9r2woub",
   "tg_tokens": null,
   "tps": 200.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-153upxbw-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: isdix0lz, sc7u18u4; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1175.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 6bit; 1k 1,175 90.6 26.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/153upxbw",
   "source_url": "https://omlx.ai/benchmarks/performance/153upxbw",
   "tg_tokens": null,
   "tps": 90.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-153upxbw-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: isdix0lz, sc7u18u4; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1355.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 6bit; 8k 1,355 107.0 28.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/153upxbw",
   "source_url": "https://omlx.ai/benchmarks/performance/153upxbw",
   "tg_tokens": null,
   "tps": 107.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-153upxbw-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: isdix0lz, sc7u18u4; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1242.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 6bit; 32k 1,242 126.8 29.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/153upxbw",
   "source_url": "https://omlx.ai/benchmarks/performance/153upxbw",
   "tg_tokens": null,
   "tps": 126.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-iqi5vt7e-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: 4sqmunsi; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1178.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 6bit; 1k 1,178 93.7 26.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iqi5vt7e",
   "source_url": "https://omlx.ai/benchmarks/performance/iqi5vt7e",
   "tg_tokens": null,
   "tps": 93.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-iqi5vt7e-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: 4sqmunsi; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1356.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 6bit; 8k 1,356 93.3 28.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iqi5vt7e",
   "source_url": "https://omlx.ai/benchmarks/performance/iqi5vt7e",
   "tg_tokens": null,
   "tps": 93.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-iqi5vt7e-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: 4sqmunsi; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1240.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 6bit; 32k 1,240 124.8 29.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iqi5vt7e",
   "source_url": "https://omlx.ai/benchmarks/performance/iqi5vt7e",
   "tg_tokens": null,
   "tps": 124.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-t06extzp-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1141.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 1k 1,141 76.2 33.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t06extzp",
   "source_url": "https://omlx.ai/benchmarks/performance/t06extzp",
   "tg_tokens": null,
   "tps": 76.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-t06extzp-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1374.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 8k 1,374 66.3 35.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t06extzp",
   "source_url": "https://omlx.ai/benchmarks/performance/t06extzp",
   "tg_tokens": null,
   "tps": 66.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-t06extzp-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.8 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1257.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 32k 1,257 121.2 36.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t06extzp",
   "source_url": "https://omlx.ai/benchmarks/performance/t06extzp",
   "tg_tokens": null,
   "tps": 121.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-g2c4pf6t-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: notul366, z9q2w6f2; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1143.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 1k 1,143 78.5 33.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g2c4pf6t",
   "source_url": "https://omlx.ai/benchmarks/performance/g2c4pf6t",
   "tg_tokens": null,
   "tps": 78.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-g2c4pf6t-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: notul366, z9q2w6f2; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1375.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 8k 1,375 87.2 35.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g2c4pf6t",
   "source_url": "https://omlx.ai/benchmarks/performance/g2c4pf6t",
   "tg_tokens": null,
   "tps": 87.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-g2c4pf6t-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.8 GB at this context; duplicate traces of the same session: notul366, z9q2w6f2; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1258.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 32k 1,258 121.3 36.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g2c4pf6t",
   "source_url": "https://omlx.ai/benchmarks/performance/g2c4pf6t",
   "tg_tokens": null,
   "tps": 121.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-pfugvmxh-ctx4096",
   "model": "Swift-1.5-3bit",
   "notes": "TextOnly; oMLX community benchmark trace; peak memory 13.9 GB at this context; duplicate traces of the same session: 3uzjje0t; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 211.3,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Swift-1.5-3bit-MLX-TextOnly on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 3bit; 4k 211.3 9.6 13.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pfugvmxh",
   "source_url": "https://omlx.ai/benchmarks/performance/pfugvmxh",
   "tg_tokens": null,
   "tps": 9.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-pfugvmxh-ctx16384",
   "model": "Swift-1.5-3bit",
   "notes": "TextOnly; oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: 3uzjje0t; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 152.6,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Swift-1.5-3bit-MLX-TextOnly on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 3bit; 16k 152.6 5.4 14.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pfugvmxh",
   "source_url": "https://omlx.ai/benchmarks/performance/pfugvmxh",
   "tg_tokens": null,
   "tps": 5.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vtohh97i-ctx1024",
   "model": "gemma-4-12B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 7.7 GB at this context; duplicate traces of the same session: 50ekp8qp, 5yko47cg, lyf2jkf3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 645.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-12B-it-qat-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 645.6 65.9 7.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vtohh97i",
   "source_url": "https://omlx.ai/benchmarks/performance/vtohh97i",
   "tg_tokens": null,
   "tps": 65.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vtohh97i-ctx4096",
   "model": "gemma-4-12B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 8.0 GB at this context; duplicate traces of the same session: 50ekp8qp, 5yko47cg, lyf2jkf3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 693.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-12B-it-qat-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 693.0 65.1 8.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vtohh97i",
   "source_url": "https://omlx.ai/benchmarks/performance/vtohh97i",
   "tg_tokens": null,
   "tps": 65.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vtohh97i-ctx8192",
   "model": "gemma-4-12B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 8.2 GB at this context; duplicate traces of the same session: 50ekp8qp, 5yko47cg, lyf2jkf3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 676.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-12B-it-qat-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 676.4 64.5 8.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vtohh97i",
   "source_url": "https://omlx.ai/benchmarks/performance/vtohh97i",
   "tg_tokens": null,
   "tps": 64.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vtohh97i-ctx16384",
   "model": "gemma-4-12B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 8.5 GB at this context; duplicate traces of the same session: 50ekp8qp, 5yko47cg, lyf2jkf3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 648.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-12B-it-qat-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 16k 648.2 62.8 8.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vtohh97i",
   "source_url": "https://omlx.ai/benchmarks/performance/vtohh97i",
   "tg_tokens": null,
   "tps": 62.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-p7i4t3rw-ctx131072",
   "model": "Laguna XS 2.1",
   "notes": "oMLX community benchmark trace; peak memory 30.9 GB at this context; duplicate traces of the same session: ubu272cb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1365.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Laguna-XS-2.1-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 128k 1,365 54.6 30.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p7i4t3rw",
   "source_url": "https://omlx.ai/benchmarks/performance/p7i4t3rw",
   "tg_tokens": null,
   "tps": 54.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-p7i4t3rw-ctx199680",
   "model": "Laguna XS 2.1",
   "notes": "oMLX community benchmark trace; peak memory 33.5 GB at this context; duplicate traces of the same session: ubu272cb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1061.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Laguna-XS-2.1-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 195k 1,061 43.0 33.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p7i4t3rw",
   "source_url": "https://omlx.ai/benchmarks/performance/p7i4t3rw",
   "tg_tokens": null,
   "tps": 43.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-q4lbivgo-ctx131072",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.2 GB at this context; duplicate traces of the same session: kmmu2s65; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2597.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ6 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 128k 2,597 88.9 28.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q4lbivgo",
   "source_url": "https://omlx.ai/benchmarks/performance/q4lbivgo",
   "tg_tokens": null,
   "tps": 88.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-q4lbivgo-ctx199680",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: kmmu2s65; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2219.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ6 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 195k 2,219 77.7 28.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q4lbivgo",
   "source_url": "https://omlx.ai/benchmarks/performance/q4lbivgo",
   "tg_tokens": null,
   "tps": 77.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-8u150snq-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 6tyvrw0r; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 403.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 1k 403.3 89.2 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8u150snq",
   "source_url": "https://omlx.ai/benchmarks/performance/8u150snq",
   "tg_tokens": null,
   "tps": 89.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-8u150snq-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 6tyvrw0r; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 753.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 753.2 90.7 21.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8u150snq",
   "source_url": "https://omlx.ai/benchmarks/performance/8u150snq",
   "tg_tokens": null,
   "tps": 90.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-0ur9x8bx-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.2 GB at this context; duplicate traces of the same session: 7n5ndmaj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 879.4,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Ornith-1.5-35B-A3B-oQ3e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 3bit; 1k 879.4 79.6 16.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ur9x8bx",
   "source_url": "https://omlx.ai/benchmarks/performance/0ur9x8bx",
   "tg_tokens": null,
   "tps": 79.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-0ur9x8bx-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: 7n5ndmaj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 855.1,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Ornith-1.5-35B-A3B-oQ3e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 3bit; 16k 855.1 57.3 17.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ur9x8bx",
   "source_url": "https://omlx.ai/benchmarks/performance/0ur9x8bx",
   "tg_tokens": null,
   "tps": 57.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-gyd0s1ua-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.2 GB at this context; duplicate traces of the same session: 4o3l61zv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 879.1,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Ornith-1.5-35B-A3B-oQ3e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 3bit; 1k 879.1 81.3 16.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gyd0s1ua",
   "source_url": "https://omlx.ai/benchmarks/performance/gyd0s1ua",
   "tg_tokens": null,
   "tps": 81.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-gyd0s1ua-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: 4o3l61zv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 857.5,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Ornith-1.5-35B-A3B-oQ3e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 3bit; 16k 857.5 65.7 17.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gyd0s1ua",
   "source_url": "https://omlx.ai/benchmarks/performance/gyd0s1ua",
   "tg_tokens": null,
   "tps": 65.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-320okump-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: b5rt0umj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 828.8,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Ornith-1.5-35B-A3B-oQ3e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 3bit; 1k 828.8 69.7 16.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/320okump",
   "source_url": "https://omlx.ai/benchmarks/performance/320okump",
   "tg_tokens": null,
   "tps": 69.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-320okump-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: b5rt0umj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 785.6,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Ornith-1.5-35B-A3B-oQ3e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 3bit; 16k 785.6 53.3 17.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/320okump",
   "source_url": "https://omlx.ai/benchmarks/performance/320okump",
   "tg_tokens": null,
   "tps": 53.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Pro",
   "id": "omlx-wzur3tr2-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 5.8 GB; TTFT 4340 ms; duplicate traces of the same session: none; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 235.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 18 GB \u00b7 4bit; 1k tokens 235.9 PP tok/s 17.2 TG tok/s 4340 TTFT (ms) 5.8 Peak mem",
   "ram_gb": 18,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wzur3tr2",
   "source_url": "https://omlx.ai/benchmarks/performance/wzur3tr2",
   "tg_tokens": null,
   "tps": 17.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-653quxkb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized speed FP16 build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: ft0nu7h7, jmhwipz6; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed-FP16 on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 105.1 12.5 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/653quxkb",
   "source_url": "https://omlx.ai/benchmarks/performance/653quxkb",
   "tg_tokens": null,
   "tps": 12.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-653quxkb-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized speed FP16 build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: ft0nu7h7, jmhwipz6; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 97.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed-FP16 on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 97.2 12.2 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/653quxkb",
   "source_url": "https://omlx.ai/benchmarks/performance/653quxkb",
   "tg_tokens": null,
   "tps": 12.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-653quxkb-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized speed FP16 build; oMLX community benchmark trace; peak memory 25.7 GB at this context; duplicate traces of the same session: ft0nu7h7, jmhwipz6; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 96.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed-FP16 on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 16k 96.8 11.6 25.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/653quxkb",
   "source_url": "https://omlx.ai/benchmarks/performance/653quxkb",
   "tg_tokens": null,
   "tps": 11.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-46yug802-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 2f5xgdsl; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 603.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 603.1 41.8 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/46yug802",
   "source_url": "https://omlx.ai/benchmarks/performance/46yug802",
   "tg_tokens": null,
   "tps": 41.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-46yug802-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 2f5xgdsl; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 689.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 689.1 35.7 19.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/46yug802",
   "source_url": "https://omlx.ai/benchmarks/performance/46yug802",
   "tg_tokens": null,
   "tps": 35.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ex2nrykv-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 51.8 GB at this context; duplicate traces of the same session: 78zfxq7g, zf9w04en; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 113.3,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.8-27B-bf16 on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 bf16; 1k 113.3 4.7 51.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ex2nrykv",
   "source_url": "https://omlx.ai/benchmarks/performance/ex2nrykv",
   "tg_tokens": null,
   "tps": 4.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ex2nrykv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 55.4 GB at this context; duplicate traces of the same session: 78zfxq7g, zf9w04en; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 115.2,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.8-27B-bf16 on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 bf16; 4k 115.2 4.7 55.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ex2nrykv",
   "source_url": "https://omlx.ai/benchmarks/performance/ex2nrykv",
   "tg_tokens": null,
   "tps": 4.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ex2nrykv-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 56.6 GB at this context; duplicate traces of the same session: 78zfxq7g, zf9w04en; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 110.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.8-27B-bf16 on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 bf16; 16k 110.0 4.6 56.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ex2nrykv",
   "source_url": "https://omlx.ai/benchmarks/performance/ex2nrykv",
   "tg_tokens": null,
   "tps": 4.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-0sn2g3b4-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: okzs5w3y, uj71ojdx; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1118.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k 1,118 65.3 19.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0sn2g3b4",
   "source_url": "https://omlx.ai/benchmarks/performance/0sn2g3b4",
   "tg_tokens": null,
   "tps": 65.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-0sn2g3b4-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: okzs5w3y, uj71ojdx; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1408.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 8k 1,408 94.8 20.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0sn2g3b4",
   "source_url": "https://omlx.ai/benchmarks/performance/0sn2g3b4",
   "tg_tokens": null,
   "tps": 94.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-0sn2g3b4-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: okzs5w3y, uj71ojdx; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1309.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 32k 1,309 87.6 22.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0sn2g3b4",
   "source_url": "https://omlx.ai/benchmarks/performance/0sn2g3b4",
   "tg_tokens": null,
   "tps": 87.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-vxv2zjz4-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: w9vhxy0n; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1297.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k 1,297 92.0 17.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vxv2zjz4",
   "source_url": "https://omlx.ai/benchmarks/performance/vxv2zjz4",
   "tg_tokens": null,
   "tps": 92.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-vxv2zjz4-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: w9vhxy0n; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1537.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 8k 1,537 114.5 19.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vxv2zjz4",
   "source_url": "https://omlx.ai/benchmarks/performance/vxv2zjz4",
   "tg_tokens": null,
   "tps": 114.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-vxv2zjz4-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: w9vhxy0n; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1391.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 32k 1,391 200.3 20.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vxv2zjz4",
   "source_url": "https://omlx.ai/benchmarks/performance/vxv2zjz4",
   "tg_tokens": null,
   "tps": 200.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-tiw46m8r-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1178.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 6bit; 1k 1,178 93.7 26.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tiw46m8r",
   "source_url": "https://omlx.ai/benchmarks/performance/tiw46m8r",
   "tg_tokens": null,
   "tps": 93.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-tiw46m8r-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1356.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 6bit; 8k 1,356 93.3 28.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tiw46m8r",
   "source_url": "https://omlx.ai/benchmarks/performance/tiw46m8r",
   "tg_tokens": null,
   "tps": 93.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-tiw46m8r-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1240.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 6bit; 32k 1,240 124.8 29.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tiw46m8r",
   "source_url": "https://omlx.ai/benchmarks/performance/tiw46m8r",
   "tg_tokens": null,
   "tps": 124.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-zpliucwp-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: n9bbka42; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1141.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 1k 1,141 76.2 33.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zpliucwp",
   "source_url": "https://omlx.ai/benchmarks/performance/zpliucwp",
   "tg_tokens": null,
   "tps": 76.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-zpliucwp-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: n9bbka42; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1374.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 8k 1,374 66.3 35.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zpliucwp",
   "source_url": "https://omlx.ai/benchmarks/performance/zpliucwp",
   "tg_tokens": null,
   "tps": 66.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-zpliucwp-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.8 GB at this context; duplicate traces of the same session: n9bbka42; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1257.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 32k 1,257 121.2 36.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zpliucwp",
   "source_url": "https://omlx.ai/benchmarks/performance/zpliucwp",
   "tg_tokens": null,
   "tps": 121.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-v1tf34c3-ctx1024",
   "model": "gemma-4-12B-uncensored",
   "notes": "oMLX community benchmark trace; peak memory 12.8 GB at this context; duplicate traces of the same session: gik3q0au, qi7exze5, v1251pgx; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 692.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-12B-uncensored-oQ8e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 692.0 45.6 12.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v1tf34c3",
   "source_url": "https://omlx.ai/benchmarks/performance/v1tf34c3",
   "tg_tokens": null,
   "tps": 45.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-v1tf34c3-ctx4096",
   "model": "gemma-4-12B-uncensored",
   "notes": "oMLX community benchmark trace; peak memory 13.2 GB at this context; duplicate traces of the same session: gik3q0au, qi7exze5, v1251pgx; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 692.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-12B-uncensored-oQ8e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 692.6 45.6 13.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v1tf34c3",
   "source_url": "https://omlx.ai/benchmarks/performance/v1tf34c3",
   "tg_tokens": null,
   "tps": 45.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-v1tf34c3-ctx8192",
   "model": "gemma-4-12B-uncensored",
   "notes": "oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: gik3q0au, qi7exze5, v1251pgx; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 676.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-12B-uncensored-oQ8e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 8k 676.8 45.3 13.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v1tf34c3",
   "source_url": "https://omlx.ai/benchmarks/performance/v1tf34c3",
   "tg_tokens": null,
   "tps": 45.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-v1tf34c3-ctx16384",
   "model": "gemma-4-12B-uncensored",
   "notes": "oMLX community benchmark trace; peak memory 13.8 GB at this context; duplicate traces of the same session: gik3q0au, qi7exze5, v1251pgx; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 648.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-12B-uncensored-oQ8e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 16k 648.6 44.4 13.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v1tf34c3",
   "source_url": "https://omlx.ai/benchmarks/performance/v1tf34c3",
   "tg_tokens": null,
   "tps": 44.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-8uyf0ovd-ctx1024",
   "model": "Gemma 4 26B-A4B-it-uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: 0kk11z4s, 57fg9zsd, h6ws9t1c; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1879.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 1,879 90.4 25.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8uyf0ovd",
   "source_url": "https://omlx.ai/benchmarks/performance/8uyf0ovd",
   "tg_tokens": null,
   "tps": 90.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-8uyf0ovd-ctx4096",
   "model": "Gemma 4 26B-A4B-it-uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: 0kk11z4s, 57fg9zsd, h6ws9t1c; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1873.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 1,873 89.0 26.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8uyf0ovd",
   "source_url": "https://omlx.ai/benchmarks/performance/8uyf0ovd",
   "tg_tokens": null,
   "tps": 89.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-8uyf0ovd-ctx8192",
   "model": "Gemma 4 26B-A4B-it-uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: 0kk11z4s, 57fg9zsd, h6ws9t1c; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1811.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 8k 1,811 87.8 26.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8uyf0ovd",
   "source_url": "https://omlx.ai/benchmarks/performance/8uyf0ovd",
   "tg_tokens": null,
   "tps": 87.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-8uyf0ovd-ctx16384",
   "model": "Gemma 4 26B-A4B-it-uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: 0kk11z4s, 57fg9zsd, h6ws9t1c; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1703.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 16k 1,703 83.9 26.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8uyf0ovd",
   "source_url": "https://omlx.ai/benchmarks/performance/8uyf0ovd",
   "tg_tokens": null,
   "tps": 83.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-jy526gdn-ctx1024",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: um3jz23n; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 879.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-MLX-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 1k 879.2 28.2 32.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jy526gdn",
   "source_url": "https://omlx.ai/benchmarks/performance/jy526gdn",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-jy526gdn-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 33.1 GB at this context; duplicate traces of the same session: um3jz23n; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 864.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-MLX-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 4k 864.8 27.8 33.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jy526gdn",
   "source_url": "https://omlx.ai/benchmarks/performance/jy526gdn",
   "tg_tokens": null,
   "tps": 27.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-m8l4xlk7-ctx1024",
   "model": "gemma-4-E4B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 8.9 GB at this context; duplicate traces of the same session: jqs8t86f; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 551.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-E4B-it-qat-8bit on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 8bit; 1k 551.4 32.4 8.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m8l4xlk7",
   "source_url": "https://omlx.ai/benchmarks/performance/m8l4xlk7",
   "tg_tokens": null,
   "tps": 32.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-m8l4xlk7-ctx4096",
   "model": "gemma-4-E4B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 8.9 GB at this context; duplicate traces of the same session: jqs8t86f; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 557.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-E4B-it-qat-8bit on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 8bit; 4k 557.9 31.1 8.9 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m8l4xlk7",
   "source_url": "https://omlx.ai/benchmarks/performance/m8l4xlk7",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-kudubb4d-ctx199680",
   "model": "Flash-Next",
   "notes": "MTP speculative decoding enabled; name as published, likely Qwen3.8-Flash-Next shorthand; oMLX community benchmark trace, single run at 195k context; peak memory 87.7 GB; TTFT 124310 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1609.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Flash-Next-4bit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k tokens 1,609 PP tok/s 51.4 TG tok/s 124310 TTFT (ms) 87.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kudubb4d",
   "source_url": "https://omlx.ai/benchmarks/performance/kudubb4d",
   "tg_tokens": null,
   "tps": 51.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2710i568-ctx32768",
   "model": "Flash-Next",
   "notes": "MTP speculative decoding enabled; name as published, likely Qwen3.8-Flash-Next shorthand; oMLX community benchmark trace, single run at 32k context; peak memory 79.7 GB; TTFT 16850 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1945.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Flash-Next-4bit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k tokens 1,945 PP tok/s 88.6 TG tok/s 16850 TTFT (ms) 79.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2710i568",
   "source_url": "https://omlx.ai/benchmarks/performance/2710i568",
   "tg_tokens": null,
   "tps": 88.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ocp7ysiu-ctx4096",
   "model": "Flash-Next",
   "notes": "MTP speculative decoding enabled; name as published, likely Qwen3.8-Flash-Next shorthand; oMLX community benchmark trace; peak memory 78.0 GB at this context; duplicate traces of the same session: ur9ir05r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1392.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Flash-Next-4bit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,392 79.5 78.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ocp7ysiu",
   "source_url": "https://omlx.ai/benchmarks/performance/ocp7ysiu",
   "tg_tokens": null,
   "tps": 79.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ocp7ysiu-ctx32768",
   "model": "Flash-Next",
   "notes": "MTP speculative decoding enabled; name as published, likely Qwen3.8-Flash-Next shorthand; oMLX community benchmark trace; peak memory 79.7 GB at this context; duplicate traces of the same session: ur9ir05r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1939.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Flash-Next-4bit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,939 80.9 79.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ocp7ysiu",
   "source_url": "https://omlx.ai/benchmarks/performance/ocp7ysiu",
   "tg_tokens": null,
   "tps": 80.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-dfgazvn7-ctx16384",
   "model": "GLM-5.3",
   "notes": "oMLX community benchmark trace; peak memory 425.1 GB at this context; duplicate traces of the same session: mjz88plj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 160.4,
   "provenance": "community",
   "quant": "mixed 4_5bit",
   "quote": "GLM-5.3-mixed-4_5bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 160.4 15.9 425.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dfgazvn7",
   "source_url": "https://omlx.ai/benchmarks/performance/dfgazvn7",
   "tg_tokens": null,
   "tps": 15.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-dfgazvn7-ctx199680",
   "model": "GLM-5.3",
   "notes": "oMLX community benchmark trace; peak memory 443.6 GB at this context; duplicate traces of the same session: mjz88plj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 154.6,
   "provenance": "community",
   "quant": "mixed 4_5bit",
   "quote": "GLM-5.3-mixed-4_5bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 195k 154.6 12.4 443.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dfgazvn7",
   "source_url": "https://omlx.ai/benchmarks/performance/dfgazvn7",
   "tg_tokens": null,
   "tps": 12.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-wcq6wpcl-ctx4096",
   "model": "Qwen2.5-Coder-32B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 62.9 GB at this context; duplicate traces of the same session: ntshzijt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 216.4,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen2.5-Coder-32B-Instruct-bf16 on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 bf16; 4k 216.4 7.2 62.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wcq6wpcl",
   "source_url": "https://omlx.ai/benchmarks/performance/wcq6wpcl",
   "tg_tokens": null,
   "tps": 7.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-wcq6wpcl-ctx16384",
   "model": "Qwen2.5-Coder-32B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 65.7 GB at this context; duplicate traces of the same session: ntshzijt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 170.2,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen2.5-Coder-32B-Instruct-bf16 on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 bf16; 16k 170.2 6.7 65.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wcq6wpcl",
   "source_url": "https://omlx.ai/benchmarks/performance/wcq6wpcl",
   "tg_tokens": null,
   "tps": 6.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-d2j2v3qr-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.8 GB at this context; duplicate traces of the same session: hptygzfx; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 128.1,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.8-27B-3bit-mtp-mlx on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 3bit; 4k 128.1 8.9 16.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d2j2v3qr",
   "source_url": "https://omlx.ai/benchmarks/performance/d2j2v3qr",
   "tg_tokens": null,
   "tps": 8.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-d2j2v3qr-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.5 GB at this context; duplicate traces of the same session: hptygzfx; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.7,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.8-27B-3bit-mtp-mlx on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 3bit; 16k 125.7 12.2 17.5 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d2j2v3qr",
   "source_url": "https://omlx.ai/benchmarks/performance/d2j2v3qr",
   "tg_tokens": null,
   "tps": 12.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-rr3s7gu1-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: gjlhlcsl; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 177.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 4k 177.7 17.3 19.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rr3s7gu1",
   "source_url": "https://omlx.ai/benchmarks/performance/rr3s7gu1",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-rr3s7gu1-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: gjlhlcsl; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 169.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 16k 169.0 16.2 21.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rr3s7gu1",
   "source_url": "https://omlx.ai/benchmarks/performance/rr3s7gu1",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-efzicyqb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: dyilr9cf, eg4zihp1, mo33rz1d, owsgxi5q; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 435.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 435.3 37.3 16.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/efzicyqb",
   "source_url": "https://omlx.ai/benchmarks/performance/efzicyqb",
   "tg_tokens": null,
   "tps": 37.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-efzicyqb-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: dyilr9cf, eg4zihp1, mo33rz1d, owsgxi5q; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 445.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 445.1 37.1 20.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/efzicyqb",
   "source_url": "https://omlx.ai/benchmarks/performance/efzicyqb",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-efzicyqb-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: dyilr9cf, eg4zihp1, mo33rz1d, owsgxi5q; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 443.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 443.1 36.3 21.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/efzicyqb",
   "source_url": "https://omlx.ai/benchmarks/performance/efzicyqb",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-efzicyqb-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: dyilr9cf, eg4zihp1, mo33rz1d, owsgxi5q; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 428.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 428.3 34.8 20.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/efzicyqb",
   "source_url": "https://omlx.ai/benchmarks/performance/efzicyqb",
   "tg_tokens": null,
   "tps": 34.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-efzicyqb-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: dyilr9cf, eg4zihp1, mo33rz1d, owsgxi5q; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 397.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 397.9 31.7 22.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/efzicyqb",
   "source_url": "https://omlx.ai/benchmarks/performance/efzicyqb",
   "tg_tokens": null,
   "tps": 31.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3uup3mrr-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 33.9 GB; TTFT 3195 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 320.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k tokens 320.5 PP tok/s 58.0 TG tok/s 3195 TTFT (ms) 33.9 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3uup3mrr",
   "source_url": "https://omlx.ai/benchmarks/performance/3uup3mrr",
   "tg_tokens": null,
   "tps": 58.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-kf3sdz1f-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: ab5uuc8p, e58yckfp, se9wq9mn; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 320.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 320.5 53.0 33.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kf3sdz1f",
   "source_url": "https://omlx.ai/benchmarks/performance/kf3sdz1f",
   "tg_tokens": null,
   "tps": 53.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-kf3sdz1f-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: ab5uuc8p, e58yckfp, se9wq9mn; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 334.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 334.3 47.3 37.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kf3sdz1f",
   "source_url": "https://omlx.ai/benchmarks/performance/kf3sdz1f",
   "tg_tokens": null,
   "tps": 47.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-kf3sdz1f-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.8 GB at this context; duplicate traces of the same session: ab5uuc8p, e58yckfp, se9wq9mn; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 329.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 8k 329.1 50.3 37.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kf3sdz1f",
   "source_url": "https://omlx.ai/benchmarks/performance/kf3sdz1f",
   "tg_tokens": null,
   "tps": 50.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-kf3sdz1f-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.7 GB at this context; duplicate traces of the same session: ab5uuc8p, e58yckfp, se9wq9mn; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 317.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 16k 317.2 65.1 38.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kf3sdz1f",
   "source_url": "https://omlx.ai/benchmarks/performance/kf3sdz1f",
   "tg_tokens": null,
   "tps": 65.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-uvas21ya-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: im63rxan, wxr2ezq6; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 417.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 417.6 45.0 33.9 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uvas21ya",
   "source_url": "https://omlx.ai/benchmarks/performance/uvas21ya",
   "tg_tokens": null,
   "tps": 45.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-uvas21ya-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: im63rxan, wxr2ezq6; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 442.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 442.0 42.9 37.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uvas21ya",
   "source_url": "https://omlx.ai/benchmarks/performance/uvas21ya",
   "tg_tokens": null,
   "tps": 42.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-uvas21ya-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.6 GB at this context; duplicate traces of the same session: im63rxan, wxr2ezq6; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 436.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 436.0 51.0 37.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uvas21ya",
   "source_url": "https://omlx.ai/benchmarks/performance/uvas21ya",
   "tg_tokens": null,
   "tps": 51.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-p76b54jv-ctx1024",
   "model": "gemma-4-E4B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 8.9 GB at this context; duplicate traces of the same session: 01pqrjgr; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 551.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-E4B-it-qat-8bit on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 8bit; 1k 551.3 35.6 8.9 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p76b54jv",
   "source_url": "https://omlx.ai/benchmarks/performance/p76b54jv",
   "tg_tokens": null,
   "tps": 35.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-p76b54jv-ctx4096",
   "model": "gemma-4-E4B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 8.9 GB at this context; duplicate traces of the same session: 01pqrjgr; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 555.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-E4B-it-qat-8bit on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 8bit; 4k 555.7 30.3 8.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p76b54jv",
   "source_url": "https://omlx.ai/benchmarks/performance/p76b54jv",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5et226kx-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 212.5 GB at this context; duplicate traces of the same session: 4vtd21b9, 97v3bl2t, hevw5ocl, kgd9x9ov, kycvpqzp, nzuakanh, osike9ex; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 348.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "GLM-5.3-Flash-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 1k 348.7 34.8 212.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5et226kx",
   "source_url": "https://omlx.ai/benchmarks/performance/5et226kx",
   "tg_tokens": null,
   "tps": 34.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5et226kx-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 213.6 GB at this context; duplicate traces of the same session: 4vtd21b9, 97v3bl2t, hevw5ocl, kgd9x9ov, kycvpqzp, nzuakanh, osike9ex; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 418.5,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "GLM-5.3-Flash-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 4k 418.5 36.9 213.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5et226kx",
   "source_url": "https://omlx.ai/benchmarks/performance/5et226kx",
   "tg_tokens": null,
   "tps": 36.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5et226kx-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 214.1 GB at this context; duplicate traces of the same session: 4vtd21b9, 97v3bl2t, hevw5ocl, kgd9x9ov, kycvpqzp, nzuakanh, osike9ex; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 384.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "GLM-5.3-Flash-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 8k 384.9 47.5 214.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5et226kx",
   "source_url": "https://omlx.ai/benchmarks/performance/5et226kx",
   "tg_tokens": null,
   "tps": 47.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5et226kx-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 214.5 GB at this context; duplicate traces of the same session: 4vtd21b9, 97v3bl2t, hevw5ocl, kgd9x9ov, kycvpqzp, nzuakanh, osike9ex; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 399.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "GLM-5.3-Flash-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 16k 399.0 45.0 214.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5et226kx",
   "source_url": "https://omlx.ai/benchmarks/performance/5et226kx",
   "tg_tokens": null,
   "tps": 45.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5et226kx-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 215.3 GB at this context; duplicate traces of the same session: 4vtd21b9, 97v3bl2t, hevw5ocl, kgd9x9ov, kycvpqzp, nzuakanh, osike9ex; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 393.5,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "GLM-5.3-Flash-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 32k 393.5 43.3 215.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5et226kx",
   "source_url": "https://omlx.ai/benchmarks/performance/5et226kx",
   "tg_tokens": null,
   "tps": 43.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5et226kx-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 217.0 GB at this context; duplicate traces of the same session: 4vtd21b9, 97v3bl2t, hevw5ocl, kgd9x9ov, kycvpqzp, nzuakanh, osike9ex; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 387.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "GLM-5.3-Flash-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 64k 387.7 48.3 217.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5et226kx",
   "source_url": "https://omlx.ai/benchmarks/performance/5et226kx",
   "tg_tokens": null,
   "tps": 48.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5et226kx-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 220.6 GB at this context; duplicate traces of the same session: 4vtd21b9, 97v3bl2t, hevw5ocl, kgd9x9ov, kycvpqzp, nzuakanh, osike9ex; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 379.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "GLM-5.3-Flash-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 128k 379.7 51.2 220.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5et226kx",
   "source_url": "https://omlx.ai/benchmarks/performance/5et226kx",
   "tg_tokens": null,
   "tps": 51.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5et226kx-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 224.2 GB at this context; duplicate traces of the same session: 4vtd21b9, 97v3bl2t, hevw5ocl, kgd9x9ov, kycvpqzp, nzuakanh, osike9ex; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 372.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "GLM-5.3-Flash-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 195k 372.7 56.6 224.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5et226kx",
   "source_url": "https://omlx.ai/benchmarks/performance/5et226kx",
   "tg_tokens": null,
   "tps": 56.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-r0cad86b-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: wmap13el; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 790.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 1k 790.0 89.5 19.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r0cad86b",
   "source_url": "https://omlx.ai/benchmarks/performance/r0cad86b",
   "tg_tokens": null,
   "tps": 89.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-r0cad86b-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: wmap13el; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 725.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 4k 725.8 87.0 20.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r0cad86b",
   "source_url": "https://omlx.ai/benchmarks/performance/r0cad86b",
   "tg_tokens": null,
   "tps": 87.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-4eployi6-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 24.7 GB; TTFT 4278 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1915.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 8k tokens 1,915 PP tok/s 129.5 TG tok/s 4278 TTFT (ms) 24.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4eployi6",
   "source_url": "https://omlx.ai/benchmarks/performance/4eployi6",
   "tg_tokens": null,
   "tps": 129.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-1tux0kv1-ctx1024",
   "model": "4-bit",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: 9eudrsjy; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 144.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "4-bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 1k 144.7 21.0 15.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1tux0kv1",
   "source_url": "https://omlx.ai/benchmarks/performance/1tux0kv1",
   "tg_tokens": null,
   "tps": 21.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-1tux0kv1-ctx4096",
   "model": "4-bit",
   "notes": "oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: 9eudrsjy; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 132.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "4-bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k 132.8 20.4 19.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1tux0kv1",
   "source_url": "https://omlx.ai/benchmarks/performance/1tux0kv1",
   "tg_tokens": null,
   "tps": 20.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-qcxk09rg-ctx1024",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 157.0 GB at this context; duplicate traces of the same session: 5wgk1ove, 7y7eg0gi, i7j5psg6, kchgvnhy, n4lqok62, ufqms9ir, y2ipf2fc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 359.5,
   "provenance": "community",
   "quant": "unknown",
   "quote": "DeepSeek-V4-Flash-0731 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 1k 359.5 42.6 157.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qcxk09rg",
   "source_url": "https://omlx.ai/benchmarks/performance/qcxk09rg",
   "tg_tokens": null,
   "tps": 42.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-qcxk09rg-ctx4096",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 157.7 GB at this context; duplicate traces of the same session: 5wgk1ove, 7y7eg0gi, i7j5psg6, kchgvnhy, n4lqok62, ufqms9ir, y2ipf2fc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 440.9,
   "provenance": "community",
   "quant": "unknown",
   "quote": "DeepSeek-V4-Flash-0731 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 4k 440.9 45.9 157.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qcxk09rg",
   "source_url": "https://omlx.ai/benchmarks/performance/qcxk09rg",
   "tg_tokens": null,
   "tps": 45.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-qcxk09rg-ctx8192",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 158.1 GB at this context; duplicate traces of the same session: 5wgk1ove, 7y7eg0gi, i7j5psg6, kchgvnhy, n4lqok62, ufqms9ir, y2ipf2fc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 459.7,
   "provenance": "community",
   "quant": "unknown",
   "quote": "DeepSeek-V4-Flash-0731 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 8k 459.7 32.2 158.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qcxk09rg",
   "source_url": "https://omlx.ai/benchmarks/performance/qcxk09rg",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-qcxk09rg-ctx16384",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 159.1 GB at this context; duplicate traces of the same session: 5wgk1ove, 7y7eg0gi, i7j5psg6, kchgvnhy, n4lqok62, ufqms9ir, y2ipf2fc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 451.4,
   "provenance": "community",
   "quant": "unknown",
   "quote": "DeepSeek-V4-Flash-0731 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 16k 451.4 41.9 159.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qcxk09rg",
   "source_url": "https://omlx.ai/benchmarks/performance/qcxk09rg",
   "tg_tokens": null,
   "tps": 41.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-qcxk09rg-ctx32768",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 161.0 GB at this context; duplicate traces of the same session: 5wgk1ove, 7y7eg0gi, i7j5psg6, kchgvnhy, n4lqok62, ufqms9ir, y2ipf2fc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 438.4,
   "provenance": "community",
   "quant": "unknown",
   "quote": "DeepSeek-V4-Flash-0731 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 32k 438.4 46.5 161.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qcxk09rg",
   "source_url": "https://omlx.ai/benchmarks/performance/qcxk09rg",
   "tg_tokens": null,
   "tps": 46.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-qcxk09rg-ctx65536",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 164.9 GB at this context; duplicate traces of the same session: 5wgk1ove, 7y7eg0gi, i7j5psg6, kchgvnhy, n4lqok62, ufqms9ir, y2ipf2fc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 415.5,
   "provenance": "community",
   "quant": "unknown",
   "quote": "DeepSeek-V4-Flash-0731 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 64k 415.5 42.5 164.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qcxk09rg",
   "source_url": "https://omlx.ai/benchmarks/performance/qcxk09rg",
   "tg_tokens": null,
   "tps": 42.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-qcxk09rg-ctx131072",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 172.5 GB at this context; duplicate traces of the same session: 5wgk1ove, 7y7eg0gi, i7j5psg6, kchgvnhy, n4lqok62, ufqms9ir, y2ipf2fc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 377.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "DeepSeek-V4-Flash-0731 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 128k 377.0 39.3 172.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qcxk09rg",
   "source_url": "https://omlx.ai/benchmarks/performance/qcxk09rg",
   "tg_tokens": null,
   "tps": 39.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-qcxk09rg-ctx199680",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 181.3 GB at this context; duplicate traces of the same session: 5wgk1ove, 7y7eg0gi, i7j5psg6, kchgvnhy, n4lqok62, ufqms9ir, y2ipf2fc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 343.5,
   "provenance": "community",
   "quant": "unknown",
   "quote": "DeepSeek-V4-Flash-0731 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 195k 343.5 35.1 181.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qcxk09rg",
   "source_url": "https://omlx.ai/benchmarks/performance/qcxk09rg",
   "tg_tokens": null,
   "tps": 35.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-n0cv5dk7-ctx8192",
   "model": "Qwen3.5-0.8B",
   "notes": "fp16 MTP build; oMLX community benchmark trace, single run at 8k context; peak memory 2.6 GB; TTFT 965 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 8485.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-0.8B-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 8k tokens 8,485 PP tok/s 278.9 TG tok/s 965 TTFT (ms) 2.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n0cv5dk7",
   "source_url": "https://omlx.ai/benchmarks/performance/n0cv5dk7",
   "tg_tokens": null,
   "tps": 278.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-6080tesx-ctx32768",
   "model": "Qwen3.5-0.8B",
   "notes": "fp16 MTP build; oMLX community benchmark trace, single run at 32k context; peak memory 3.4 GB; TTFT 5847 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5604.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-0.8B-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 32k tokens 5,604 PP tok/s 31.1 TG tok/s 5847 TTFT (ms) 3.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6080tesx",
   "source_url": "https://omlx.ai/benchmarks/performance/6080tesx",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-nbnpitup-ctx8192",
   "model": "Qwen3.5-2B",
   "notes": "fp16 MTP build; oMLX community benchmark trace, single run at 8k context; peak memory 2.8 GB; TTFT 1743 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4700.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-2B-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 8k tokens 4,700 PP tok/s 204.6 TG tok/s 1743 TTFT (ms) 2.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nbnpitup",
   "source_url": "https://omlx.ai/benchmarks/performance/nbnpitup",
   "tg_tokens": null,
   "tps": 204.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-211shn0q-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 5.2 GB; TTFT 4336 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1889.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.5-4B-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 8k tokens 1,889 PP tok/s 121.0 TG tok/s 4336 TTFT (ms) 5.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/211shn0q",
   "source_url": "https://omlx.ai/benchmarks/performance/211shn0q",
   "tg_tokens": null,
   "tps": 121.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-lrds856u-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 24.7 GB; TTFT 4276 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1916.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 8k tokens 1,916 PP tok/s 142.0 TG tok/s 4276 TTFT (ms) 24.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lrds856u",
   "source_url": "https://omlx.ai/benchmarks/performance/lrds856u",
   "tg_tokens": null,
   "tps": 142.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-igws4if5-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 22.3 GB; TTFT 4039 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2028.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 8k tokens 2,028 PP tok/s 73.4 TG tok/s 4039 TTFT (ms) 22.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/igws4if5",
   "source_url": "https://omlx.ai/benchmarks/performance/igws4if5",
   "tg_tokens": null,
   "tps": 73.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-86nlqdgr-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 24.7 GB; TTFT 4284 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1912.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 8k tokens 1,912 PP tok/s 111.7 TG tok/s 4284 TTFT (ms) 24.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/86nlqdgr",
   "source_url": "https://omlx.ai/benchmarks/performance/86nlqdgr",
   "tg_tokens": null,
   "tps": 111.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-t8qpyzvu-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 20.0 GB; TTFT 783 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1308.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 1k tokens 1,308 PP tok/s 74.0 TG tok/s 783 TTFT (ms) 20.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t8qpyzvu",
   "source_url": "https://omlx.ai/benchmarks/performance/t8qpyzvu",
   "tg_tokens": null,
   "tps": 74.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-v45wbkd2-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 22.3 GB; TTFT 4126 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1986.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 8k tokens 1,986 PP tok/s 71.5 TG tok/s 4126 TTFT (ms) 22.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v45wbkd2",
   "source_url": "https://omlx.ai/benchmarks/performance/v45wbkd2",
   "tg_tokens": null,
   "tps": 71.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qa2yak2i-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "TensorFold MTP; oMLX community benchmark trace; peak memory 173.9 GB at this context; duplicate traces of the same session: 1t06hpp0, jqeq8my4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 410.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--GLM-5.3-Flash-MLX-4bit-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 410.6 37.0 173.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qa2yak2i",
   "source_url": "https://omlx.ai/benchmarks/performance/qa2yak2i",
   "tg_tokens": null,
   "tps": 37.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-qa2yak2i-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "TensorFold MTP; oMLX community benchmark trace; peak memory 174.2 GB at this context; duplicate traces of the same session: 1t06hpp0, jqeq8my4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 390.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--GLM-5.3-Flash-MLX-4bit-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 390.6 52.1 174.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qa2yak2i",
   "source_url": "https://omlx.ai/benchmarks/performance/qa2yak2i",
   "tg_tokens": null,
   "tps": 52.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-qa2yak2i-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "TensorFold MTP; oMLX community benchmark trace; peak memory 177.3 GB at this context; duplicate traces of the same session: 1t06hpp0, jqeq8my4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 390.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--GLM-5.3-Flash-MLX-4bit-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 390.2 5.4 177.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qa2yak2i",
   "source_url": "https://omlx.ai/benchmarks/performance/qa2yak2i",
   "tg_tokens": null,
   "tps": 5.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-o2jat6fi-ctx4096",
   "model": "Gemma-4-31B",
   "notes": "oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: yn1wfnpk; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 311.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31b-it-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 311.0 14.1 19.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o2jat6fi",
   "source_url": "https://omlx.ai/benchmarks/performance/o2jat6fi",
   "tg_tokens": null,
   "tps": 14.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-o2jat6fi-ctx16384",
   "model": "Gemma-4-31B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: yn1wfnpk; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 292.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31b-it-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 292.3 12.5 20.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o2jat6fi",
   "source_url": "https://omlx.ai/benchmarks/performance/o2jat6fi",
   "tg_tokens": null,
   "tps": 12.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-h7g9msf3-ctx8192",
   "model": "Qwen3.5-0.8B",
   "notes": "fp16 MTP build; oMLX community benchmark trace, single run at 8k context; peak memory 2.6 GB; TTFT 1027 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 7975.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-0.8B-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 8k tokens 7,975 PP tok/s 260.9 TG tok/s 1027 TTFT (ms) 2.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h7g9msf3",
   "source_url": "https://omlx.ai/benchmarks/performance/h7g9msf3",
   "tg_tokens": null,
   "tps": 260.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-is6fp5yj-ctx8192",
   "model": "Qwen3.5-2B",
   "notes": "fp16 MTP build; oMLX community benchmark trace, single run at 8k context; peak memory 2.8 GB; TTFT 1751 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4679.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-2B-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 8k tokens 4,679 PP tok/s 204.6 TG tok/s 1751 TTFT (ms) 2.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/is6fp5yj",
   "source_url": "https://omlx.ai/benchmarks/performance/is6fp5yj",
   "tg_tokens": null,
   "tps": 204.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-okti03bq-ctx8192",
   "model": "Qwen3.5-2B",
   "notes": "fp16 MTP build; oMLX community benchmark trace, single run at 8k context; peak memory 4.8 GB; TTFT 1892 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4331.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-2B-oQ6e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 6bit; 8k tokens 4,331 PP tok/s 165.5 TG tok/s 1892 TTFT (ms) 4.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/okti03bq",
   "source_url": "https://omlx.ai/benchmarks/performance/okti03bq",
   "tg_tokens": null,
   "tps": 165.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-34ouot7h-ctx8192",
   "model": "Qwen3.5-2B",
   "notes": "fp16 MTP build; oMLX community benchmark trace, single run at 8k context; peak memory 4.8 GB; TTFT 1874 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4371.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-2B-oQ6e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 6bit; 8k tokens 4,371 PP tok/s 147.4 TG tok/s 1874 TTFT (ms) 4.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/34ouot7h",
   "source_url": "https://omlx.ai/benchmarks/performance/34ouot7h",
   "tg_tokens": null,
   "tps": 147.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-mmlpwzu4-ctx8192",
   "model": "Qwen3.5-2B",
   "notes": "fp16 MTP build; oMLX community benchmark trace, single run at 8k context; peak memory 4.8 GB; TTFT 1880 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4357.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-2B-oQ6e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 6bit; 8k tokens 4,357 PP tok/s 169.5 TG tok/s 1880 TTFT (ms) 4.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mmlpwzu4",
   "source_url": "https://omlx.ai/benchmarks/performance/mmlpwzu4",
   "tg_tokens": null,
   "tps": 169.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-9hvebx8h-ctx8192",
   "model": "Qwen3.5-2B",
   "notes": "fp16 MTP build; oMLX community benchmark trace, single run at 8k context; peak memory 4.8 GB; TTFT 1881 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4354.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-2B-oQ6e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 6bit; 8k tokens 4,354 PP tok/s 184.3 TG tok/s 1881 TTFT (ms) 4.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9hvebx8h",
   "source_url": "https://omlx.ai/benchmarks/performance/9hvebx8h",
   "tg_tokens": null,
   "tps": 184.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-jaye3arm-ctx8192",
   "model": "Qwen3.5-2B",
   "notes": "fp16 MTP build; oMLX community benchmark trace, single run at 8k context; peak memory 4.8 GB; TTFT 1886 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4343.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-2B-oQ6e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 6bit; 8k tokens 4,343 PP tok/s 223.9 TG tok/s 1886 TTFT (ms) 4.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jaye3arm",
   "source_url": "https://omlx.ai/benchmarks/performance/jaye3arm",
   "tg_tokens": null,
   "tps": 223.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-935udg1t-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 0qf0f6tv; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 591.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 591.1 69.7 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/935udg1t",
   "source_url": "https://omlx.ai/benchmarks/performance/935udg1t",
   "tg_tokens": null,
   "tps": 69.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-935udg1t-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 0qf0f6tv; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 593.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 593.7 67.6 21.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/935udg1t",
   "source_url": "https://omlx.ai/benchmarks/performance/935udg1t",
   "tg_tokens": null,
   "tps": 67.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-playgtb3-ctx1024",
   "model": "FrogNano-4B-2609",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 5.4 GB at this context; duplicate traces of the same session: l7fhyfha; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 794.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "FrogNano-4B-2609-oQ8e-mtp-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 794.8 48.6 5.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/playgtb3",
   "source_url": "https://omlx.ai/benchmarks/performance/playgtb3",
   "tg_tokens": null,
   "tps": 48.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-playgtb3-ctx4096",
   "model": "FrogNano-4B-2609",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 6.6 GB at this context; duplicate traces of the same session: l7fhyfha; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 804.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "FrogNano-4B-2609-oQ8e-mtp-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 804.7 45.0 6.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/playgtb3",
   "source_url": "https://omlx.ai/benchmarks/performance/playgtb3",
   "tg_tokens": null,
   "tps": 45.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tc73gzts-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: 2a7cdqlw, 79s1v4nt, gw9ij22w, w3cy1lir, weenijy9, xng5dabv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1017.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,017 56.3 71.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tc73gzts",
   "source_url": "https://omlx.ai/benchmarks/performance/tc73gzts",
   "tg_tokens": null,
   "tps": 56.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tc73gzts-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: 2a7cdqlw, 79s1v4nt, gw9ij22w, w3cy1lir, weenijy9, xng5dabv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1863.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,863 71.3 71.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tc73gzts",
   "source_url": "https://omlx.ai/benchmarks/performance/tc73gzts",
   "tg_tokens": null,
   "tps": 71.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tc73gzts-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 2a7cdqlw, 79s1v4nt, gw9ij22w, w3cy1lir, weenijy9, xng5dabv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2067.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,067 47.5 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tc73gzts",
   "source_url": "https://omlx.ai/benchmarks/performance/tc73gzts",
   "tg_tokens": null,
   "tps": 47.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tc73gzts-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.9 GB at this context; duplicate traces of the same session: 2a7cdqlw, 79s1v4nt, gw9ij22w, w3cy1lir, weenijy9, xng5dabv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2196.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,196 69.5 72.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tc73gzts",
   "source_url": "https://omlx.ai/benchmarks/performance/tc73gzts",
   "tg_tokens": null,
   "tps": 69.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-tc73gzts-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.8 GB at this context; duplicate traces of the same session: 2a7cdqlw, 79s1v4nt, gw9ij22w, w3cy1lir, weenijy9, xng5dabv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2162.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,162 90.6 73.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tc73gzts",
   "source_url": "https://omlx.ai/benchmarks/performance/tc73gzts",
   "tg_tokens": null,
   "tps": 90.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tc73gzts-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 76.6 GB at this context; duplicate traces of the same session: 2a7cdqlw, 79s1v4nt, gw9ij22w, w3cy1lir, weenijy9, xng5dabv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2115.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,115 65.3 76.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tc73gzts",
   "source_url": "https://omlx.ai/benchmarks/performance/tc73gzts",
   "tg_tokens": null,
   "tps": 65.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-tc73gzts-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 81.8 GB at this context; duplicate traces of the same session: 2a7cdqlw, 79s1v4nt, gw9ij22w, w3cy1lir, weenijy9, xng5dabv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2058.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,058 71.7 81.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tc73gzts",
   "source_url": "https://omlx.ai/benchmarks/performance/tc73gzts",
   "tg_tokens": null,
   "tps": 71.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qynu7ujz-ctx4096",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.3 GB at this context; duplicate traces of the same session: ohg1ntpe, scrmj2kn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 504.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "DeepSeek-V4-Flash-0731-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 504.3 33.4 156.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qynu7ujz",
   "source_url": "https://omlx.ai/benchmarks/performance/qynu7ujz",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-qynu7ujz-ctx16384",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.4 GB at this context; duplicate traces of the same session: ohg1ntpe, scrmj2kn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 584.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "DeepSeek-V4-Flash-0731-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k 584.6 44.1 156.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qynu7ujz",
   "source_url": "https://omlx.ai/benchmarks/performance/qynu7ujz",
   "tg_tokens": null,
   "tps": 44.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-qynu7ujz-ctx65536",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.8 GB at this context; duplicate traces of the same session: ohg1ntpe, scrmj2kn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 539.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "DeepSeek-V4-Flash-0731-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 64k 539.2 43.8 156.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qynu7ujz",
   "source_url": "https://omlx.ai/benchmarks/performance/qynu7ujz",
   "tg_tokens": null,
   "tps": 43.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-shlagu4h-ctx131072",
   "model": "Laguna XS 2.1",
   "notes": "oMLX community benchmark trace; peak memory 30.9 GB at this context; duplicate traces of the same session: cvehcemi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1350.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Laguna-XS-2.1-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 128k 1,350 52.6 30.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/shlagu4h",
   "source_url": "https://omlx.ai/benchmarks/performance/shlagu4h",
   "tg_tokens": null,
   "tps": 52.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-shlagu4h-ctx199680",
   "model": "Laguna XS 2.1",
   "notes": "oMLX community benchmark trace; peak memory 33.5 GB at this context; duplicate traces of the same session: cvehcemi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1056.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Laguna-XS-2.1-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 195k 1,056 42.3 33.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/shlagu4h",
   "source_url": "https://omlx.ai/benchmarks/performance/shlagu4h",
   "tg_tokens": null,
   "tps": 42.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-vsmzsrrj-ctx1024",
   "model": "Qwen3.6-35B-A3B-MAYA",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: i9xnz5ee; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1156.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-MAYA-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 1,156 104.2 22.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vsmzsrrj",
   "source_url": "https://omlx.ai/benchmarks/performance/vsmzsrrj",
   "tg_tokens": null,
   "tps": 104.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-vsmzsrrj-ctx4096",
   "model": "Qwen3.6-35B-A3B-MAYA",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: i9xnz5ee; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1240.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-MAYA-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 1,240 77.3 24.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vsmzsrrj",
   "source_url": "https://omlx.ai/benchmarks/performance/vsmzsrrj",
   "tg_tokens": null,
   "tps": 77.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-do3chs51-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: vvjmgrm7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1774.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,774 133.1 23.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/do3chs51",
   "source_url": "https://omlx.ai/benchmarks/performance/do3chs51",
   "tg_tokens": null,
   "tps": 133.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-do3chs51-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: vvjmgrm7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1496.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,496 146.3 23.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/do3chs51",
   "source_url": "https://omlx.ai/benchmarks/performance/do3chs51",
   "tg_tokens": null,
   "tps": 146.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-tpqwrix5-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: mmee9f1w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1765.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,765 130.9 23.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tpqwrix5",
   "source_url": "https://omlx.ai/benchmarks/performance/tpqwrix5",
   "tg_tokens": null,
   "tps": 130.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-tpqwrix5-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: mmee9f1w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1495.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,495 136.3 23.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tpqwrix5",
   "source_url": "https://omlx.ai/benchmarks/performance/tpqwrix5",
   "tg_tokens": null,
   "tps": 136.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-j8apeyus-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 1smr5hho; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1773.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,773 130.0 23.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j8apeyus",
   "source_url": "https://omlx.ai/benchmarks/performance/j8apeyus",
   "tg_tokens": null,
   "tps": 130.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-j8apeyus-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 1smr5hho; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1497.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,497 139.8 23.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j8apeyus",
   "source_url": "https://omlx.ai/benchmarks/performance/j8apeyus",
   "tg_tokens": null,
   "tps": 139.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-q23q8hox-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 0buwg2hp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1655.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,655 51.0 23.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q23q8hox",
   "source_url": "https://omlx.ai/benchmarks/performance/q23q8hox",
   "tg_tokens": null,
   "tps": 51.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-q23q8hox-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 0buwg2hp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1495.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,495 79.5 23.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q23q8hox",
   "source_url": "https://omlx.ai/benchmarks/performance/q23q8hox",
   "tg_tokens": null,
   "tps": 79.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-x3wpu9tj-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: t7htlr05; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1737.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,737 118.1 23.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x3wpu9tj",
   "source_url": "https://omlx.ai/benchmarks/performance/x3wpu9tj",
   "tg_tokens": null,
   "tps": 118.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-x3wpu9tj-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: t7htlr05; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1485.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,485 143.2 23.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x3wpu9tj",
   "source_url": "https://omlx.ai/benchmarks/performance/x3wpu9tj",
   "tg_tokens": null,
   "tps": 143.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-s4a1g3e2-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: all9vr6m, blvnuexh, nui2v3pf; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 76.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 76.8 10.6 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s4a1g3e2",
   "source_url": "https://omlx.ai/benchmarks/performance/s4a1g3e2",
   "tg_tokens": null,
   "tps": 10.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-s4a1g3e2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: all9vr6m, blvnuexh, nui2v3pf; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 86.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 86.5 15.0 19.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s4a1g3e2",
   "source_url": "https://omlx.ai/benchmarks/performance/s4a1g3e2",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-s4a1g3e2-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: all9vr6m, blvnuexh, nui2v3pf; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 64.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 64.1 7.4 20.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s4a1g3e2",
   "source_url": "https://omlx.ai/benchmarks/performance/s4a1g3e2",
   "tg_tokens": null,
   "tps": 7.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-s4a1g3e2-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: all9vr6m, blvnuexh, nui2v3pf; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 51.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 51.5 11.1 21.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s4a1g3e2",
   "source_url": "https://omlx.ai/benchmarks/performance/s4a1g3e2",
   "tg_tokens": null,
   "tps": 11.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-86gcsqsk-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: vduf9xn0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 947.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 947.2 51.4 18.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/86gcsqsk",
   "source_url": "https://omlx.ai/benchmarks/performance/86gcsqsk",
   "tg_tokens": null,
   "tps": 51.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-86gcsqsk-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: vduf9xn0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 842.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 842.1 72.1 19.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/86gcsqsk",
   "source_url": "https://omlx.ai/benchmarks/performance/86gcsqsk",
   "tg_tokens": null,
   "tps": 72.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wl3zoeeb-ctx131072",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: r1zdk1hs; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 762.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Muse-Glimmer-30B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 128k 762.5 19.5 23.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wl3zoeeb",
   "source_url": "https://omlx.ai/benchmarks/performance/wl3zoeeb",
   "tg_tokens": null,
   "tps": 19.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wl3zoeeb-ctx199680",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: r1zdk1hs; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 692.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Muse-Glimmer-30B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 195k 692.3 18.1 24.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wl3zoeeb",
   "source_url": "https://omlx.ai/benchmarks/performance/wl3zoeeb",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wr8si4t9-ctx131072",
   "model": "North-Mini-Code-1.0",
   "notes": "oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: m5uz56l4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1840.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "North-Mini-Code-1.0-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 128k 1,840 56.2 28.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wr8si4t9",
   "source_url": "https://omlx.ai/benchmarks/performance/wr8si4t9",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wr8si4t9-ctx199680",
   "model": "North-Mini-Code-1.0",
   "notes": "oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: m5uz56l4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1459.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "North-Mini-Code-1.0-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 195k 1,459 47.9 30.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wr8si4t9",
   "source_url": "https://omlx.ai/benchmarks/performance/wr8si4t9",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-news4u4e-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 81.1 GB at this context; duplicate traces of the same session: 3f3pl2gl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1384.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 1,384 65.3 81.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/news4u4e",
   "source_url": "https://omlx.ai/benchmarks/performance/news4u4e",
   "tg_tokens": null,
   "tps": 65.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-news4u4e-ctx16384",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 81.3 GB at this context; duplicate traces of the same session: 3f3pl2gl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1157.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 1,157 61.2 81.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/news4u4e",
   "source_url": "https://omlx.ai/benchmarks/performance/news4u4e",
   "tg_tokens": null,
   "tps": 61.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-w3023xg3-ctx1024",
   "model": "Qwen3.6-35B-A3B-MAYA",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 4olawk3y, bz80x1ty, o5x06tsr, ptfvbiro, ydxaj18p; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1212.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-MAYA-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 1,212 103.5 22.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w3023xg3",
   "source_url": "https://omlx.ai/benchmarks/performance/w3023xg3",
   "tg_tokens": null,
   "tps": 103.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-w3023xg3-ctx4096",
   "model": "Qwen3.6-35B-A3B-MAYA",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: 4olawk3y, bz80x1ty, o5x06tsr, ptfvbiro, ydxaj18p; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1401.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-MAYA-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 1,401 72.1 24.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w3023xg3",
   "source_url": "https://omlx.ai/benchmarks/performance/w3023xg3",
   "tg_tokens": null,
   "tps": 72.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-w3023xg3-ctx8192",
   "model": "Qwen3.6-35B-A3B-MAYA",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: 4olawk3y, bz80x1ty, o5x06tsr, ptfvbiro, ydxaj18p; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1053.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-MAYA-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 1,053 81.6 24.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w3023xg3",
   "source_url": "https://omlx.ai/benchmarks/performance/w3023xg3",
   "tg_tokens": null,
   "tps": 81.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-w3023xg3-ctx16384",
   "model": "Qwen3.6-35B-A3B-MAYA",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: 4olawk3y, bz80x1ty, o5x06tsr, ptfvbiro, ydxaj18p; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 948.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-MAYA-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 948.7 86.9 25.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w3023xg3",
   "source_url": "https://omlx.ai/benchmarks/performance/w3023xg3",
   "tg_tokens": null,
   "tps": 86.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-w3023xg3-ctx32768",
   "model": "Qwen3.6-35B-A3B-MAYA",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: 4olawk3y, bz80x1ty, o5x06tsr, ptfvbiro, ydxaj18p; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 774.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-MAYA-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 32k 774.9 68.8 26.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w3023xg3",
   "source_url": "https://omlx.ai/benchmarks/performance/w3023xg3",
   "tg_tokens": null,
   "tps": 68.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-w3023xg3-ctx65536",
   "model": "Qwen3.6-35B-A3B-MAYA",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: 4olawk3y, bz80x1ty, o5x06tsr, ptfvbiro, ydxaj18p; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 569.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-MAYA-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 64k 569.3 49.0 27.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w3023xg3",
   "source_url": "https://omlx.ai/benchmarks/performance/w3023xg3",
   "tg_tokens": null,
   "tps": 49.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-8k7k8nsz-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 5g52ly4w, 6ser7xfh, dxrgrhjq, reoflzwh, zhe5jkwf; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1218.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 1,218 124.1 22.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8k7k8nsz",
   "source_url": "https://omlx.ai/benchmarks/performance/8k7k8nsz",
   "tg_tokens": null,
   "tps": 124.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-8k7k8nsz-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: 5g52ly4w, 6ser7xfh, dxrgrhjq, reoflzwh, zhe5jkwf; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1401.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 1,401 69.8 24.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8k7k8nsz",
   "source_url": "https://omlx.ai/benchmarks/performance/8k7k8nsz",
   "tg_tokens": null,
   "tps": 69.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-8k7k8nsz-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: 5g52ly4w, 6ser7xfh, dxrgrhjq, reoflzwh, zhe5jkwf; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1044.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 1,044 93.1 24.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8k7k8nsz",
   "source_url": "https://omlx.ai/benchmarks/performance/8k7k8nsz",
   "tg_tokens": null,
   "tps": 93.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-8k7k8nsz-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: 5g52ly4w, 6ser7xfh, dxrgrhjq, reoflzwh, zhe5jkwf; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 934.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 934.7 84.5 25.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8k7k8nsz",
   "source_url": "https://omlx.ai/benchmarks/performance/8k7k8nsz",
   "tg_tokens": null,
   "tps": 84.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-8k7k8nsz-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.9 GB at this context; duplicate traces of the same session: 5g52ly4w, 6ser7xfh, dxrgrhjq, reoflzwh, zhe5jkwf; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 755.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 32k 755.2 65.4 26.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8k7k8nsz",
   "source_url": "https://omlx.ai/benchmarks/performance/8k7k8nsz",
   "tg_tokens": null,
   "tps": 65.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-8k7k8nsz-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.6 GB at this context; duplicate traces of the same session: 5g52ly4w, 6ser7xfh, dxrgrhjq, reoflzwh, zhe5jkwf; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 556.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 64k 556.4 47.7 27.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8k7k8nsz",
   "source_url": "https://omlx.ai/benchmarks/performance/8k7k8nsz",
   "tg_tokens": null,
   "tps": 47.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-6eujj4r3-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: cxvplhhw; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1218.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 1,218 112.4 22.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6eujj4r3",
   "source_url": "https://omlx.ai/benchmarks/performance/6eujj4r3",
   "tg_tokens": null,
   "tps": 112.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-6eujj4r3-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: cxvplhhw; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1399.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 1,399 73.6 24.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6eujj4r3",
   "source_url": "https://omlx.ai/benchmarks/performance/6eujj4r3",
   "tg_tokens": null,
   "tps": 73.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-1z9pddiw-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: jkbgk31k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1769.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,769 131.1 23.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1z9pddiw",
   "source_url": "https://omlx.ai/benchmarks/performance/1z9pddiw",
   "tg_tokens": null,
   "tps": 131.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-1z9pddiw-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: jkbgk31k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1493.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,493 164.4 23.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1z9pddiw",
   "source_url": "https://omlx.ai/benchmarks/performance/1z9pddiw",
   "tg_tokens": null,
   "tps": 164.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-4akjnu9o-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: b8tjzpev; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1692.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,692 85.8 23.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4akjnu9o",
   "source_url": "https://omlx.ai/benchmarks/performance/4akjnu9o",
   "tg_tokens": null,
   "tps": 85.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-4akjnu9o-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: b8tjzpev; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1479.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,479 79.1 23.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4akjnu9o",
   "source_url": "https://omlx.ai/benchmarks/performance/4akjnu9o",
   "tg_tokens": null,
   "tps": 79.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-w9ojyvks-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: i33p7bcw; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 416.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 416.7 17.5 17.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w9ojyvks",
   "source_url": "https://omlx.ai/benchmarks/performance/w9ojyvks",
   "tg_tokens": null,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-w9ojyvks-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: i33p7bcw; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 402.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 16k 402.4 16.2 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w9ojyvks",
   "source_url": "https://omlx.ai/benchmarks/performance/w9ojyvks",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-df43tx00-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 4t9hw990, 7uf9wa8n, atrd1pxu; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 76.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 1k 76.4 9.5 19.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/df43tx00",
   "source_url": "https://omlx.ai/benchmarks/performance/df43tx00",
   "tg_tokens": null,
   "tps": 9.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-df43tx00-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 4t9hw990, 7uf9wa8n, atrd1pxu; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 72.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 72.3 10.6 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/df43tx00",
   "source_url": "https://omlx.ai/benchmarks/performance/df43tx00",
   "tg_tokens": null,
   "tps": 10.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-df43tx00-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: 4t9hw990, 7uf9wa8n, atrd1pxu; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 71.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 8k 71.7 8.3 24.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/df43tx00",
   "source_url": "https://omlx.ai/benchmarks/performance/df43tx00",
   "tg_tokens": null,
   "tps": 8.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-df43tx00-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: 4t9hw990, 7uf9wa8n, atrd1pxu; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 68.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 16k 68.3 7.5 24.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/df43tx00",
   "source_url": "https://omlx.ai/benchmarks/performance/df43tx00",
   "tg_tokens": null,
   "tps": 7.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-s1jhowwk-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.6 GB at this context; duplicate traces of the same session: 4ng7lt1d, c3ybo7yz, jcgcuhkj, la1cswvs, xvq29bnj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 144.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 144.3 19.3 39.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s1jhowwk",
   "source_url": "https://omlx.ai/benchmarks/performance/s1jhowwk",
   "tg_tokens": null,
   "tps": 19.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-s1jhowwk-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.6 GB at this context; duplicate traces of the same session: 4ng7lt1d, c3ybo7yz, jcgcuhkj, la1cswvs, xvq29bnj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 194.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 194.2 18.1 39.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s1jhowwk",
   "source_url": "https://omlx.ai/benchmarks/performance/s1jhowwk",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-s1jhowwk-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 40.4 GB at this context; duplicate traces of the same session: 4ng7lt1d, c3ybo7yz, jcgcuhkj, la1cswvs, xvq29bnj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 361.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 361.0 10.8 40.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s1jhowwk",
   "source_url": "https://omlx.ai/benchmarks/performance/s1jhowwk",
   "tg_tokens": null,
   "tps": 10.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-s1jhowwk-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 42.0 GB at this context; duplicate traces of the same session: 4ng7lt1d, c3ybo7yz, jcgcuhkj, la1cswvs, xvq29bnj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 438.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 438.0 17.7 42.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s1jhowwk",
   "source_url": "https://omlx.ai/benchmarks/performance/s1jhowwk",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-s1jhowwk-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 42.8 GB at this context; duplicate traces of the same session: 4ng7lt1d, c3ybo7yz, jcgcuhkj, la1cswvs, xvq29bnj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 494.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 494.5 22.5 42.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s1jhowwk",
   "source_url": "https://omlx.ai/benchmarks/performance/s1jhowwk",
   "tg_tokens": null,
   "tps": 22.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-s1jhowwk-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 44.9 GB at this context; duplicate traces of the same session: 4ng7lt1d, c3ybo7yz, jcgcuhkj, la1cswvs, xvq29bnj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 534.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 534.0 23.0 44.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s1jhowwk",
   "source_url": "https://omlx.ai/benchmarks/performance/s1jhowwk",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-mxnij07c-ctx32768",
   "model": "gemma-4-E4B-it-qat",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 9.6 GB; TTFT 68321 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 479.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-E4B-it-qat-8bit on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 8bit; 32k tokens 479.6 PP tok/s 15.5 TG tok/s 68321 TTFT (ms) 9.6 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mxnij07c",
   "source_url": "https://omlx.ai/benchmarks/performance/mxnij07c",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-elyapdex-ctx1024",
   "model": "GLM-4.6V-Flash",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: tlm3earz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1320.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "GLM-4.6V-Flash-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 1,320 28.3 11.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/elyapdex",
   "source_url": "https://omlx.ai/benchmarks/performance/elyapdex",
   "tg_tokens": null,
   "tps": 28.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-elyapdex-ctx4096",
   "model": "GLM-4.6V-Flash",
   "notes": "oMLX community benchmark trace; peak memory 11.9 GB at this context; duplicate traces of the same session: tlm3earz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1327.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "GLM-4.6V-Flash-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 1,327 25.9 11.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/elyapdex",
   "source_url": "https://omlx.ai/benchmarks/performance/elyapdex",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-dklyh4in-ctx1024",
   "model": "GLM-4.7-Flash",
   "notes": "oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: 6os50o7f; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1319.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "GLM-4.7-Flash-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 1,319 52.3 30.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dklyh4in",
   "source_url": "https://omlx.ai/benchmarks/performance/dklyh4in",
   "tg_tokens": null,
   "tps": 52.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-dklyh4in-ctx4096",
   "model": "GLM-4.7-Flash",
   "notes": "oMLX community benchmark trace; peak memory 31.3 GB at this context; duplicate traces of the same session: 6os50o7f; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1258.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "GLM-4.7-Flash-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 1,258 43.3 31.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dklyh4in",
   "source_url": "https://omlx.ai/benchmarks/performance/dklyh4in",
   "tg_tokens": null,
   "tps": 43.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-18xcizj2-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: diky08po; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 585.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 585.8 39.9 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/18xcizj2",
   "source_url": "https://omlx.ai/benchmarks/performance/18xcizj2",
   "tg_tokens": null,
   "tps": 39.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-18xcizj2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: diky08po; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 689.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 689.2 36.8 19.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/18xcizj2",
   "source_url": "https://omlx.ai/benchmarks/performance/18xcizj2",
   "tg_tokens": null,
   "tps": 36.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-tkf2j111-ctx1024",
   "model": "KAT-Coder-V2.5-Dev",
   "notes": "oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: wxxytd98; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1031.0,
   "provenance": "community",
   "quant": "OptiQ 4bit",
   "quote": "KAT-Coder-V2.5-Dev-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 1,031 79.5 21.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tkf2j111",
   "source_url": "https://omlx.ai/benchmarks/performance/tkf2j111",
   "tg_tokens": null,
   "tps": 79.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-tkf2j111-ctx4096",
   "model": "KAT-Coder-V2.5-Dev",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: wxxytd98; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2347.0,
   "provenance": "community",
   "quant": "OptiQ 4bit",
   "quote": "KAT-Coder-V2.5-Dev-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,347 77.1 22.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tkf2j111",
   "source_url": "https://omlx.ai/benchmarks/performance/tkf2j111",
   "tg_tokens": null,
   "tps": 77.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-5iwncx4e-ctx1024",
   "model": "KAT-Dev",
   "notes": "oMLX community benchmark trace; peak memory 33.3 GB at this context; duplicate traces of the same session: zca8wswh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 342.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "KAT-Dev-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 342.6 7.9 33.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5iwncx4e",
   "source_url": "https://omlx.ai/benchmarks/performance/5iwncx4e",
   "tg_tokens": null,
   "tps": 7.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-5iwncx4e-ctx4096",
   "model": "KAT-Dev",
   "notes": "oMLX community benchmark trace; peak memory 34.2 GB at this context; duplicate traces of the same session: zca8wswh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 338.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "KAT-Dev-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 338.5 7.9 34.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5iwncx4e",
   "source_url": "https://omlx.ai/benchmarks/performance/5iwncx4e",
   "tg_tokens": null,
   "tps": 7.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-9h4sl570-ctx1024",
   "model": "Llama-3.2-11B-Vision-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 11.3 GB at this context; duplicate traces of the same session: ibwio7js; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 618.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Llama-3.2-11B-Vision-Instruct-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 618.7 15.6 11.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9h4sl570",
   "source_url": "https://omlx.ai/benchmarks/performance/9h4sl570",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-9h4sl570-ctx4096",
   "model": "Llama-3.2-11B-Vision-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 11.6 GB at this context; duplicate traces of the same session: ibwio7js; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 589.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Llama-3.2-11B-Vision-Instruct-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 589.2 15.5 11.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9h4sl570",
   "source_url": "https://omlx.ai/benchmarks/performance/9h4sl570",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-horas8ar-ctx1024",
   "model": "Llama-3.2-11B-Vision-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 11.3 GB at this context; duplicate traces of the same session: zlwvrns8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1258.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Llama-3.2-11B-Vision-Instruct-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 1,258 28.0 11.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/horas8ar",
   "source_url": "https://omlx.ai/benchmarks/performance/horas8ar",
   "tg_tokens": null,
   "tps": 28.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-horas8ar-ctx4096",
   "model": "Llama-3.2-11B-Vision-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 11.6 GB at this context; duplicate traces of the same session: zlwvrns8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1335.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Llama-3.2-11B-Vision-Instruct-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 1,335 26.4 11.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/horas8ar",
   "source_url": "https://omlx.ai/benchmarks/performance/horas8ar",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ylkczjfd-ctx1024",
   "model": "Qwen2.5-VL-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 4.8 GB at this context; duplicate traces of the same session: nvamcqlf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 951.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-VL-7B-Instruct-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 951.1 36.9 4.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ylkczjfd",
   "source_url": "https://omlx.ai/benchmarks/performance/ylkczjfd",
   "tg_tokens": null,
   "tps": 36.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ylkczjfd-ctx4096",
   "model": "Qwen2.5-VL-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: nvamcqlf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 969.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-VL-7B-Instruct-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 969.4 35.8 4.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ylkczjfd",
   "source_url": "https://omlx.ai/benchmarks/performance/ylkczjfd",
   "tg_tokens": null,
   "tps": 35.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-6bsntqxt-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace, single run at 64k context; peak memory 141.8 GB; TTFT 48131 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1362.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 64k tokens 1,362 PP tok/s 39.0 TG tok/s 48131 TTFT (ms) 141.8 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6bsntqxt",
   "source_url": "https://omlx.ai/benchmarks/performance/6bsntqxt",
   "tg_tokens": null,
   "tps": 39.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vie8mui0-ctx32768",
   "model": "ThinkingCap-Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 42.5 GB; TTFT 109248 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 299.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "ThinkingCap-Qwen3.8-27B-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 32k tokens 299.9 PP tok/s 45.9 TG tok/s 109248 TTFT (ms) 42.5 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vie8mui0",
   "source_url": "https://omlx.ai/benchmarks/performance/vie8mui0",
   "tg_tokens": null,
   "tps": 45.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qlsdmawb-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "Vontra MTP; oMLX community benchmark trace; peak memory 172.8 GB at this context; duplicate traces of the same session: zcjj20jt; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 341.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Vontra--GLM-5.3-Flash-MLX-4bit-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 341.2 40.2 172.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qlsdmawb",
   "source_url": "https://omlx.ai/benchmarks/performance/qlsdmawb",
   "tg_tokens": null,
   "tps": 40.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qlsdmawb-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "Vontra MTP; oMLX community benchmark trace; peak memory 173.7 GB at this context; duplicate traces of the same session: zcjj20jt; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 422.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Vontra--GLM-5.3-Flash-MLX-4bit-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 422.0 34.5 173.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qlsdmawb",
   "source_url": "https://omlx.ai/benchmarks/performance/qlsdmawb",
   "tg_tokens": null,
   "tps": 34.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-0gpgirpb-ctx1024",
   "model": "Qwen2.5-Coder-3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 2.5 GB at this context; duplicate traces of the same session: u0qct3yk; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1458.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-3B-Instruct-MLX-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 1k 1,458 116.7 2.5 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0gpgirpb",
   "source_url": "https://omlx.ai/benchmarks/performance/0gpgirpb",
   "tg_tokens": null,
   "tps": 116.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-0gpgirpb-ctx4096",
   "model": "Qwen2.5-Coder-3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 2.6 GB at this context; duplicate traces of the same session: u0qct3yk; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1338.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-3B-Instruct-MLX-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 4k 1,338 107.3 2.6 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0gpgirpb",
   "source_url": "https://omlx.ai/benchmarks/performance/0gpgirpb",
   "tg_tokens": null,
   "tps": 107.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-clg8ivoz-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 20.1 GB; TTFT 1523 ms; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 672.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 1k tokens 672.3 PP tok/s 77.1 TG tok/s 1523 TTFT (ms) 20.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/clg8ivoz",
   "source_url": "https://omlx.ai/benchmarks/performance/clg8ivoz",
   "tg_tokens": null,
   "tps": 77.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-rwixxsw8-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 20.1 GB; TTFT 1012 ms; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1012.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 1k tokens 1,012 PP tok/s 77.3 TG tok/s 1012 TTFT (ms) 20.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rwixxsw8",
   "source_url": "https://omlx.ai/benchmarks/performance/rwixxsw8",
   "tg_tokens": null,
   "tps": 77.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ml4370k0-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 33.7 GB at this context; duplicate traces of the same session: nwhv7clc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 785.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ8e-1M on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 785.1 40.3 33.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ml4370k0",
   "source_url": "https://omlx.ai/benchmarks/performance/ml4370k0",
   "tg_tokens": null,
   "tps": 40.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ml4370k0-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: nwhv7clc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 766.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ8e-1M on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 766.2 41.8 34.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ml4370k0",
   "source_url": "https://omlx.ai/benchmarks/performance/ml4370k0",
   "tg_tokens": null,
   "tps": 41.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-25564jfo-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 70.3 GB; TTFT 532 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1925.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,925 PP tok/s 88.9 TG tok/s 532 TTFT (ms) 70.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/25564jfo",
   "source_url": "https://omlx.ai/benchmarks/performance/25564jfo",
   "tg_tokens": null,
   "tps": 88.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-sffqo1zy-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap MTP; oMLX community benchmark trace, single run at 32k context; peak memory 28.6 GB; TTFT 85940 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 381.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "ThinkingCap-Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k tokens 381.3 PP tok/s 60.0 TG tok/s 85940 TTFT (ms) 28.6 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sffqo1zy",
   "source_url": "https://omlx.ai/benchmarks/performance/sffqo1zy",
   "tg_tokens": null,
   "tps": 60.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rpxmyetb-ctx32768",
   "model": "ThinkingCap-Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 39.8 GB; TTFT 86494 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 378.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "ThinkingCap-Qwen3.8-27B-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 32k tokens 378.8 PP tok/s 44.9 TG tok/s 86494 TTFT (ms) 39.8 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rpxmyetb",
   "source_url": "https://omlx.ai/benchmarks/performance/rpxmyetb",
   "tg_tokens": null,
   "tps": 44.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0qohjagr-ctx32768",
   "model": "gemma-4-31B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: naji8teu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 620.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31B-it-qat-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 620.6 16.8 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0qohjagr",
   "source_url": "https://omlx.ai/benchmarks/performance/0qohjagr",
   "tg_tokens": null,
   "tps": 16.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0qohjagr-ctx65536",
   "model": "gemma-4-31B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 31.9 GB at this context; duplicate traces of the same session: naji8teu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 519.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31B-it-qat-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 519.8 13.5 31.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0qohjagr",
   "source_url": "https://omlx.ai/benchmarks/performance/0qohjagr",
   "tg_tokens": null,
   "tps": 13.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wsrk3qes-ctx1024",
   "model": "gemma-4-E4B-it-OBLITERATED",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 8.0 GB; TTFT 330 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3107.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-E4B-it-OBLITERATED-mlx-8Bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k tokens 3,107 PP tok/s 84.9 TG tok/s 330 TTFT (ms) 8.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wsrk3qes",
   "source_url": "https://omlx.ai/benchmarks/performance/wsrk3qes",
   "tg_tokens": null,
   "tps": 84.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-03pwzviq-ctx1024",
   "model": "gemma-4-E4B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: 9r911cu4, 9uo090cr, fy0rrj9r, oklcsr0c, wgmkpg24; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2825.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-qat-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 2,825 64.0 6.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/03pwzviq",
   "source_url": "https://omlx.ai/benchmarks/performance/03pwzviq",
   "tg_tokens": null,
   "tps": 64.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-03pwzviq-ctx4096",
   "model": "gemma-4-E4B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: 9r911cu4, 9uo090cr, fy0rrj9r, oklcsr0c, wgmkpg24; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3385.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-qat-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 3,385 61.3 6.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/03pwzviq",
   "source_url": "https://omlx.ai/benchmarks/performance/03pwzviq",
   "tg_tokens": null,
   "tps": 61.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-03pwzviq-ctx8192",
   "model": "gemma-4-E4B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: 9r911cu4, 9uo090cr, fy0rrj9r, oklcsr0c, wgmkpg24; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3378.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-qat-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 8k 3,378 58.8 6.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/03pwzviq",
   "source_url": "https://omlx.ai/benchmarks/performance/03pwzviq",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-03pwzviq-ctx16384",
   "model": "gemma-4-E4B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: 9r911cu4, 9uo090cr, fy0rrj9r, oklcsr0c, wgmkpg24; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3268.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-qat-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 16k 3,268 55.1 7.0 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/03pwzviq",
   "source_url": "https://omlx.ai/benchmarks/performance/03pwzviq",
   "tg_tokens": null,
   "tps": 55.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-03pwzviq-ctx32768",
   "model": "gemma-4-E4B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 7.5 GB at this context; duplicate traces of the same session: 9r911cu4, 9uo090cr, fy0rrj9r, oklcsr0c, wgmkpg24; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3006.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-qat-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 32k 3,006 48.4 7.5 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/03pwzviq",
   "source_url": "https://omlx.ai/benchmarks/performance/03pwzviq",
   "tg_tokens": null,
   "tps": 48.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-10",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-03pwzviq-ctx65536",
   "model": "gemma-4-E4B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 8.7 GB at this context; duplicate traces of the same session: 9r911cu4, 9uo090cr, fy0rrj9r, oklcsr0c, wgmkpg24; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2546.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-qat-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 64k 2,546 38.6 8.7 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/03pwzviq",
   "source_url": "https://omlx.ai/benchmarks/performance/03pwzviq",
   "tg_tokens": null,
   "tps": 38.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mpjxa2e6-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 11.8 GB; TTFT 206 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4964.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 4,964 PP tok/s 137.1 TG tok/s 206 TTFT (ms) 11.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mpjxa2e6",
   "source_url": "https://omlx.ai/benchmarks/performance/mpjxa2e6",
   "tg_tokens": null,
   "tps": 137.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-iq1fjaqw-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: faw4r2yr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1268.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,268 137.6 11.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iq1fjaqw",
   "source_url": "https://omlx.ai/benchmarks/performance/iq1fjaqw",
   "tg_tokens": null,
   "tps": 137.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-iq1fjaqw-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: faw4r2yr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4712.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 4,712 130.0 11.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iq1fjaqw",
   "source_url": "https://omlx.ai/benchmarks/performance/iq1fjaqw",
   "tg_tokens": null,
   "tps": 130.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ajcalbe9-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: 73z5thrg, 9erxdei0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4462.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 4,462 100.4 29.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ajcalbe9",
   "source_url": "https://omlx.ai/benchmarks/performance/ajcalbe9",
   "tg_tokens": null,
   "tps": 100.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ajcalbe9-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai; oMLX community benchmark trace; peak memory 31.2 GB at this context; duplicate traces of the same session: 73z5thrg, 9erxdei0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3511.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 64k 3,511 85.3 31.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ajcalbe9",
   "source_url": "https://omlx.ai/benchmarks/performance/ajcalbe9",
   "tg_tokens": null,
   "tps": 85.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ajcalbe9-ctx131072",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai; oMLX community benchmark trace; peak memory 34.4 GB at this context; duplicate traces of the same session: 73z5thrg, 9erxdei0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2421.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 128k 2,421 67.1 34.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ajcalbe9",
   "source_url": "https://omlx.ai/benchmarks/performance/ajcalbe9",
   "tg_tokens": null,
   "tps": 67.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-o3fa0e1k-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai; oMLX community benchmark trace, single run at 32k context; peak memory 29.6 GB; TTFT 7310 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4482.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k tokens 4,482 PP tok/s 100.0 TG tok/s 7310 TTFT (ms) 29.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o3fa0e1k",
   "source_url": "https://omlx.ai/benchmarks/performance/o3fa0e1k",
   "tg_tokens": null,
   "tps": 100.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-e154yair-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 1uy2ho83, 91llonsb, cix3r4xu; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1193.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 1k 1,193 95.7 19.4 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e154yair",
   "source_url": "https://omlx.ai/benchmarks/performance/e154yair",
   "tg_tokens": null,
   "tps": 95.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-e154yair-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 1uy2ho83, 91llonsb, cix3r4xu; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1267.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 4k 1,267 92.4 20.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e154yair",
   "source_url": "https://omlx.ai/benchmarks/performance/e154yair",
   "tg_tokens": null,
   "tps": 92.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-e154yair-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 1uy2ho83, 91llonsb, cix3r4xu; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1251.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 8k 1,251 86.8 20.2 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e154yair",
   "source_url": "https://omlx.ai/benchmarks/performance/e154yair",
   "tg_tokens": null,
   "tps": 86.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-e154yair-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: 1uy2ho83, 91llonsb, cix3r4xu; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1132.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 16k 1,132 78.1 20.4 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e154yair",
   "source_url": "https://omlx.ai/benchmarks/performance/e154yair",
   "tg_tokens": null,
   "tps": 78.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-r8nhgg34-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: 2trrl9l5; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1124.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 16k 1,124 80.9 20.4 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r8nhgg34",
   "source_url": "https://omlx.ai/benchmarks/performance/r8nhgg34",
   "tg_tokens": null,
   "tps": 80.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-r8nhgg34-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 2trrl9l5; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 932.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 32k 932.5 71.1 20.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r8nhgg34",
   "source_url": "https://omlx.ai/benchmarks/performance/r8nhgg34",
   "tg_tokens": null,
   "tps": 71.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-f78j9ghe-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 8k context; peak memory 20.2 GB; TTFT 6549 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1251.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 8k tokens 1,251 PP tok/s 84.2 TG tok/s 6549 TTFT (ms) 20.2 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f78j9ghe",
   "source_url": "https://omlx.ai/benchmarks/performance/f78j9ghe",
   "tg_tokens": null,
   "tps": 84.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-c2nds3y6-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: vj6kwdsx; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1183.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 1k 1,183 95.4 19.4 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c2nds3y6",
   "source_url": "https://omlx.ai/benchmarks/performance/c2nds3y6",
   "tg_tokens": null,
   "tps": 95.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-c2nds3y6-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: vj6kwdsx; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1261.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 4k 1,261 92.2 20.1 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c2nds3y6",
   "source_url": "https://omlx.ai/benchmarks/performance/c2nds3y6",
   "tg_tokens": null,
   "tps": 92.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-irjss7do-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: atkwpzk1; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 537.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 537.8 64.2 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/irjss7do",
   "source_url": "https://omlx.ai/benchmarks/performance/irjss7do",
   "tg_tokens": null,
   "tps": 64.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-irjss7do-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: atkwpzk1; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 613.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 613.9 60.3 22.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/irjss7do",
   "source_url": "https://omlx.ai/benchmarks/performance/irjss7do",
   "tg_tokens": null,
   "tps": 60.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-m63pp98d-ctx1024",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: 293tsbb0, 6o4o8vp5, aex8gunh, yt3tcsqk, zapfbt1v; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 397.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 397.1 19.6 33.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m63pp98d",
   "source_url": "https://omlx.ai/benchmarks/performance/m63pp98d",
   "tg_tokens": null,
   "tps": 19.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-m63pp98d-ctx4096",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: 293tsbb0, 6o4o8vp5, aex8gunh, yt3tcsqk, zapfbt1v; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 446.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 446.1 18.8 34.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m63pp98d",
   "source_url": "https://omlx.ai/benchmarks/performance/m63pp98d",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-m63pp98d-ctx8192",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: 293tsbb0, 6o4o8vp5, aex8gunh, yt3tcsqk, zapfbt1v; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 430.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 430.4 23.8 35.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m63pp98d",
   "source_url": "https://omlx.ai/benchmarks/performance/m63pp98d",
   "tg_tokens": null,
   "tps": 23.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-m63pp98d-ctx16384",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: 293tsbb0, 6o4o8vp5, aex8gunh, yt3tcsqk, zapfbt1v; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 392.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 16k 392.8 29.6 35.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m63pp98d",
   "source_url": "https://omlx.ai/benchmarks/performance/m63pp98d",
   "tg_tokens": null,
   "tps": 29.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-m63pp98d-ctx32768",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: 293tsbb0, 6o4o8vp5, aex8gunh, yt3tcsqk, zapfbt1v; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 375.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 32k 375.7 36.3 36.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m63pp98d",
   "source_url": "https://omlx.ai/benchmarks/performance/m63pp98d",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-m63pp98d-ctx65536",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: 293tsbb0, 6o4o8vp5, aex8gunh, yt3tcsqk, zapfbt1v; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 288.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 64k 288.2 21.4 39.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m63pp98d",
   "source_url": "https://omlx.ai/benchmarks/performance/m63pp98d",
   "tg_tokens": null,
   "tps": 21.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-q5wx2lda-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.2 GB at this context; duplicate traces of the same session: ct6qk15o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 608.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 128k 608.1 16.8 31.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q5wx2lda",
   "source_url": "https://omlx.ai/benchmarks/performance/q5wx2lda",
   "tg_tokens": null,
   "tps": 16.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-q5wx2lda-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 35.3 GB at this context; duplicate traces of the same session: ct6qk15o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 500.2,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 195k 500.2 14.6 35.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q5wx2lda",
   "source_url": "https://omlx.ai/benchmarks/performance/q5wx2lda",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ttlhi2d7-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 34.2 GB; TTFT 107875 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 303.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k tokens 303.8 PP tok/s 100.7 TG tok/s 107875 TTFT (ms) 34.2 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ttlhi2d7",
   "source_url": "https://omlx.ai/benchmarks/performance/ttlhi2d7",
   "tg_tokens": null,
   "tps": 100.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-dd20owoj-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: xytiu85v; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 309.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 309.8 47.0 34.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dd20owoj",
   "source_url": "https://omlx.ai/benchmarks/performance/dd20owoj",
   "tg_tokens": null,
   "tps": 47.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-dd20owoj-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 38.1 GB at this context; duplicate traces of the same session: xytiu85v; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 333.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 333.2 45.7 38.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dd20owoj",
   "source_url": "https://omlx.ai/benchmarks/performance/dd20owoj",
   "tg_tokens": null,
   "tps": 45.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2ssgqv36-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap MTP; oMLX community benchmark trace, single run at 32k context; peak memory 28.6 GB; TTFT 85505 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 383.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "ThinkingCap-Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k tokens 383.2 PP tok/s 58.6 TG tok/s 85505 TTFT (ms) 28.6 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ssgqv36",
   "source_url": "https://omlx.ai/benchmarks/performance/2ssgqv36",
   "tg_tokens": null,
   "tps": 58.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-8dz14um3-ctx4096",
   "model": "Hy-MT2-1.8B",
   "notes": "oMLX community benchmark trace; peak memory 2.7 GB at this context; duplicate traces of the same session: 0vj3psqp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 764.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Hy-MT2-1.8B-8bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 8bit; 4k 764.6 64.5 2.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8dz14um3",
   "source_url": "https://omlx.ai/benchmarks/performance/8dz14um3",
   "tg_tokens": null,
   "tps": 64.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-8dz14um3-ctx16384",
   "model": "Hy-MT2-1.8B",
   "notes": "oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: 0vj3psqp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 544.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Hy-MT2-1.8B-8bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 8bit; 16k 544.4 47.1 3.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8dz14um3",
   "source_url": "https://omlx.ai/benchmarks/performance/8dz14um3",
   "tg_tokens": null,
   "tps": 47.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-a3r37k7d-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: 6kesvj4p; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 8bit; 4k 127.9 9.5 32.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a3r37k7d",
   "source_url": "https://omlx.ai/benchmarks/performance/a3r37k7d",
   "tg_tokens": null,
   "tps": 9.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-a3r37k7d-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.2 GB at this context; duplicate traces of the same session: 6kesvj4p; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 119.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 8bit; 16k 119.4 9.2 34.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a3r37k7d",
   "source_url": "https://omlx.ai/benchmarks/performance/a3r37k7d",
   "tg_tokens": null,
   "tps": 9.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-jb7pcqic-ctx131072",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 43.7 GB; TTFT 471841 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 277.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 128k tokens 277.8 PP tok/s 21.5 TG tok/s 471841 TTFT (ms) 43.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jb7pcqic",
   "source_url": "https://omlx.ai/benchmarks/performance/jb7pcqic",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-84e15plr-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "MTP + PLE; oMLX community benchmark trace; peak memory 42.8 GB at this context; duplicate traces of the same session: 0pg6g8cg, jbhvqx41, uupc2bnt; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 313.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-REAP-288-oQ4e-BF16-MTP-PLE on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 313.6 23.6 42.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/84e15plr",
   "source_url": "https://omlx.ai/benchmarks/performance/84e15plr",
   "tg_tokens": null,
   "tps": 23.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-84e15plr-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "MTP + PLE; oMLX community benchmark trace; peak memory 43.3 GB at this context; duplicate traces of the same session: 0pg6g8cg, jbhvqx41, uupc2bnt; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-REAP-288-oQ4e-BF16-MTP-PLE on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 323.2 33.3 43.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/84e15plr",
   "source_url": "https://omlx.ai/benchmarks/performance/84e15plr",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-84e15plr-ctx8192",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "MTP + PLE; oMLX community benchmark trace; peak memory 43.5 GB at this context; duplicate traces of the same session: 0pg6g8cg, jbhvqx41, uupc2bnt; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 340.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-REAP-288-oQ4e-BF16-MTP-PLE on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 340.8 31.4 43.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/84e15plr",
   "source_url": "https://omlx.ai/benchmarks/performance/84e15plr",
   "tg_tokens": null,
   "tps": 31.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-84e15plr-ctx16384",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "MTP + PLE; oMLX community benchmark trace; peak memory 44.1 GB at this context; duplicate traces of the same session: 0pg6g8cg, jbhvqx41, uupc2bnt; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 341.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-REAP-288-oQ4e-BF16-MTP-PLE on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 341.3 28.4 44.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/84e15plr",
   "source_url": "https://omlx.ai/benchmarks/performance/84e15plr",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-wa6tul4o-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 195k context; peak memory 139.4 GB; TTFT 50884 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3931.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 195k tokens 3,931 PP tok/s 70.2 TG tok/s 50884 TTFT (ms) 139.4 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wa6tul4o",
   "source_url": "https://omlx.ai/benchmarks/performance/wa6tul4o",
   "tg_tokens": null,
   "tps": 70.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-l5bu9uyo-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 195k context; peak memory 139.4 GB; TTFT 53652 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3728.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 195k tokens 3,728 PP tok/s 44.3 TG tok/s 53652 TTFT (ms) 139.4 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l5bu9uyo",
   "source_url": "https://omlx.ai/benchmarks/performance/l5bu9uyo",
   "tg_tokens": null,
   "tps": 44.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-vqapzcoy-ctx1024",
   "model": "Qwopus3.8-27B-Flash-V2",
   "notes": "oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: zpz8ecx1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 313.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwopus3.8-27B-Flash-V2-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 313.7 33.7 32.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vqapzcoy",
   "source_url": "https://omlx.ai/benchmarks/performance/vqapzcoy",
   "tg_tokens": null,
   "tps": 33.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vqapzcoy-ctx4096",
   "model": "Qwopus3.8-27B-Flash-V2",
   "notes": "oMLX community benchmark trace; peak memory 33.7 GB at this context; duplicate traces of the same session: zpz8ecx1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 328.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwopus3.8-27B-Flash-V2-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 328.8 65.5 33.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vqapzcoy",
   "source_url": "https://omlx.ai/benchmarks/performance/vqapzcoy",
   "tg_tokens": null,
   "tps": 65.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-eyqtra57-ctx1024",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 7uxy9d36; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 441.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 441.1 54.5 20.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eyqtra57",
   "source_url": "https://omlx.ai/benchmarks/performance/eyqtra57",
   "tg_tokens": null,
   "tps": 54.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-eyqtra57-ctx4096",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 7uxy9d36; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 430.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 430.3 51.7 20.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eyqtra57",
   "source_url": "https://omlx.ai/benchmarks/performance/eyqtra57",
   "tg_tokens": null,
   "tps": 51.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-wlbfk0dg-ctx1024",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 87.5 GB at this context; duplicate traces of the same session: nk2td417; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 274.2,
   "provenance": "community",
   "quant": "OptiQ-2bit",
   "quote": "DeepSeek-V4-Flash-0731-OptiQ-2bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 2bit; 1k 274.2 28.1 87.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wlbfk0dg",
   "source_url": "https://omlx.ai/benchmarks/performance/wlbfk0dg",
   "tg_tokens": null,
   "tps": 28.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-wlbfk0dg-ctx4096",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 88.4 GB at this context; duplicate traces of the same session: nk2td417; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 275.0,
   "provenance": "community",
   "quant": "OptiQ-2bit",
   "quote": "DeepSeek-V4-Flash-0731-OptiQ-2bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 2bit; 4k 275.0 29.0 88.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wlbfk0dg",
   "source_url": "https://omlx.ai/benchmarks/performance/wlbfk0dg",
   "tg_tokens": null,
   "tps": 29.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-56cdcbwh-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Jundot build; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: elg9xs4t, k78bo95r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 912.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 912.8 65.3 71.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/56cdcbwh",
   "source_url": "https://omlx.ai/benchmarks/performance/56cdcbwh",
   "tg_tokens": null,
   "tps": 65.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-56cdcbwh-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Jundot build; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: elg9xs4t, k78bo95r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2400.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,400 46.2 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/56cdcbwh",
   "source_url": "https://omlx.ai/benchmarks/performance/56cdcbwh",
   "tg_tokens": null,
   "tps": 46.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-56cdcbwh-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Jundot build; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: elg9xs4t, k78bo95r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2931.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,931 72.2 72.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/56cdcbwh",
   "source_url": "https://omlx.ai/benchmarks/performance/56cdcbwh",
   "tg_tokens": null,
   "tps": 72.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-l0zimpig-ctx4096",
   "model": "Qwen3.6-27B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 28.7 GB at this context; duplicate traces of the same session: 6rimgliz, xlkejxtu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 928.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-27B-oQ6e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 928.5 44.8 28.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l0zimpig",
   "source_url": "https://omlx.ai/benchmarks/performance/l0zimpig",
   "tg_tokens": null,
   "tps": 44.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-l0zimpig-ctx32768",
   "model": "Qwen3.6-27B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 30.4 GB at this context; duplicate traces of the same session: 6rimgliz, xlkejxtu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 709.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-27B-oQ6e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 709.9 56.3 30.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l0zimpig",
   "source_url": "https://omlx.ai/benchmarks/performance/l0zimpig",
   "tg_tokens": null,
   "tps": 56.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-l0zimpig-ctx65536",
   "model": "Qwen3.6-27B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 32.8 GB at this context; duplicate traces of the same session: 6rimgliz, xlkejxtu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 561.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-27B-oQ6e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 64k 561.2 34.7 32.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l0zimpig",
   "source_url": "https://omlx.ai/benchmarks/performance/l0zimpig",
   "tg_tokens": null,
   "tps": 34.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-djaccvep-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: ugxi6alp; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1291.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k 1,291 56.0 20.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/djaccvep",
   "source_url": "https://omlx.ai/benchmarks/performance/djaccvep",
   "tg_tokens": null,
   "tps": 56.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-djaccvep-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: ugxi6alp; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1470.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 1,470 50.3 21.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/djaccvep",
   "source_url": "https://omlx.ai/benchmarks/performance/djaccvep",
   "tg_tokens": null,
   "tps": 50.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-mmu479qq-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: rj1i7fmr; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1288.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k 1,288 61.6 19.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mmu479qq",
   "source_url": "https://omlx.ai/benchmarks/performance/mmu479qq",
   "tg_tokens": null,
   "tps": 61.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-mmu479qq-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: rj1i7fmr; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1566.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 1,566 58.6 19.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mmu479qq",
   "source_url": "https://omlx.ai/benchmarks/performance/mmu479qq",
   "tg_tokens": null,
   "tps": 58.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fv703im4-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: cfg3unea, cxy16r1f, j20sw0wn, xgq6c4rk, ytsgzkn2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1165.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 1,165 42.5 33.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fv703im4",
   "source_url": "https://omlx.ai/benchmarks/performance/fv703im4",
   "tg_tokens": null,
   "tps": 42.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fv703im4-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: cfg3unea, cxy16r1f, j20sw0wn, xgq6c4rk, ytsgzkn2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1343.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 1,343 43.2 34.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fv703im4",
   "source_url": "https://omlx.ai/benchmarks/performance/fv703im4",
   "tg_tokens": null,
   "tps": 43.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fv703im4-ctx8192",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: cfg3unea, cxy16r1f, j20sw0wn, xgq6c4rk, ytsgzkn2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1176.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 1,176 51.3 35.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fv703im4",
   "source_url": "https://omlx.ai/benchmarks/performance/fv703im4",
   "tg_tokens": null,
   "tps": 51.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fv703im4-ctx16384",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: cfg3unea, cxy16r1f, j20sw0wn, xgq6c4rk, ytsgzkn2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1020.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 1,020 53.6 35.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fv703im4",
   "source_url": "https://omlx.ai/benchmarks/performance/fv703im4",
   "tg_tokens": null,
   "tps": 53.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fv703im4-ctx32768",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: cfg3unea, cxy16r1f, j20sw0wn, xgq6c4rk, ytsgzkn2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 866.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 866.8 73.7 36.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fv703im4",
   "source_url": "https://omlx.ai/benchmarks/performance/fv703im4",
   "tg_tokens": null,
   "tps": 73.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fv703im4-ctx65536",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.2 GB at this context; duplicate traces of the same session: cfg3unea, cxy16r1f, j20sw0wn, xgq6c4rk, ytsgzkn2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 764.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 764.9 40.9 39.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fv703im4",
   "source_url": "https://omlx.ai/benchmarks/performance/fv703im4",
   "tg_tokens": null,
   "tps": 40.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dc8kmer7-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 762.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 762.2 40.1 33.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dc8kmer7",
   "source_url": "https://omlx.ai/benchmarks/performance/dc8kmer7",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dc8kmer7-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 908.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 908.2 48.3 34.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dc8kmer7",
   "source_url": "https://omlx.ai/benchmarks/performance/dc8kmer7",
   "tg_tokens": null,
   "tps": 48.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dc8kmer7-ctx8192",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 832.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 832.7 49.8 35.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dc8kmer7",
   "source_url": "https://omlx.ai/benchmarks/performance/dc8kmer7",
   "tg_tokens": null,
   "tps": 49.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dc8kmer7-ctx16384",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 799.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 799.6 55.8 35.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dc8kmer7",
   "source_url": "https://omlx.ai/benchmarks/performance/dc8kmer7",
   "tg_tokens": null,
   "tps": 55.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dc8kmer7-ctx32768",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 749.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 749.1 82.9 36.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dc8kmer7",
   "source_url": "https://omlx.ai/benchmarks/performance/dc8kmer7",
   "tg_tokens": null,
   "tps": 82.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dc8kmer7-ctx65536",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 665.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 665.6 46.6 39.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dc8kmer7",
   "source_url": "https://omlx.ai/benchmarks/performance/dc8kmer7",
   "tg_tokens": null,
   "tps": 46.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-4ocfpcug-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 22.2 GB; TTFT 19015 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 215.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 215.4 PP tok/s 34.6 TG tok/s 19015 TTFT (ms) 22.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ocfpcug",
   "source_url": "https://omlx.ai/benchmarks/performance/4ocfpcug",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-r3v42wj2-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 101.5 GB; TTFT 15780 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1038.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k tokens 1,038 PP tok/s 69.1 TG tok/s 15780 TTFT (ms) 101.5 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r3v42wj2",
   "source_url": "https://omlx.ai/benchmarks/performance/r3v42wj2",
   "tg_tokens": null,
   "tps": 69.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gpmmyjw5-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: bvft10c1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1822.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,822 73.1 71.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gpmmyjw5",
   "source_url": "https://omlx.ai/benchmarks/performance/gpmmyjw5",
   "tg_tokens": null,
   "tps": 73.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gpmmyjw5-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.7 GB at this context; duplicate traces of the same session: bvft10c1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1958.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,958 70.8 72.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gpmmyjw5",
   "source_url": "https://omlx.ai/benchmarks/performance/gpmmyjw5",
   "tg_tokens": null,
   "tps": 70.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wnwpushu-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: 1mtc4w4s, fv7sorba, sbn5ehzz, uovfdqfn, wsxdwwcu, zycq41iq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1094.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,094 72.7 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wnwpushu",
   "source_url": "https://omlx.ai/benchmarks/performance/wnwpushu",
   "tg_tokens": null,
   "tps": 72.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wnwpushu-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 1mtc4w4s, fv7sorba, sbn5ehzz, uovfdqfn, wsxdwwcu, zycq41iq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1675.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,675 72.8 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wnwpushu",
   "source_url": "https://omlx.ai/benchmarks/performance/wnwpushu",
   "tg_tokens": null,
   "tps": 72.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wnwpushu-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 1mtc4w4s, fv7sorba, sbn5ehzz, uovfdqfn, wsxdwwcu, zycq41iq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2403.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,403 84.1 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wnwpushu",
   "source_url": "https://omlx.ai/benchmarks/performance/wnwpushu",
   "tg_tokens": null,
   "tps": 84.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wnwpushu-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.7 GB at this context; duplicate traces of the same session: 1mtc4w4s, fv7sorba, sbn5ehzz, uovfdqfn, wsxdwwcu, zycq41iq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2475.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,475 51.6 73.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wnwpushu",
   "source_url": "https://omlx.ai/benchmarks/performance/wnwpushu",
   "tg_tokens": null,
   "tps": 51.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wnwpushu-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 75.7 GB at this context; duplicate traces of the same session: 1mtc4w4s, fv7sorba, sbn5ehzz, uovfdqfn, wsxdwwcu, zycq41iq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2433.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,433 74.9 75.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wnwpushu",
   "source_url": "https://omlx.ai/benchmarks/performance/wnwpushu",
   "tg_tokens": null,
   "tps": 74.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wnwpushu-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 81.1 GB at this context; duplicate traces of the same session: 1mtc4w4s, fv7sorba, sbn5ehzz, uovfdqfn, wsxdwwcu, zycq41iq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2308.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,308 66.7 81.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wnwpushu",
   "source_url": "https://omlx.ai/benchmarks/performance/wnwpushu",
   "tg_tokens": null,
   "tps": 66.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wnwpushu-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 83.4 GB at this context; duplicate traces of the same session: 1mtc4w4s, fv7sorba, sbn5ehzz, uovfdqfn, wsxdwwcu, zycq41iq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2320.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,320 55.1 83.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wnwpushu",
   "source_url": "https://omlx.ai/benchmarks/performance/wnwpushu",
   "tg_tokens": null,
   "tps": 55.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xc3keif6-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 71.2 GB; TTFT 2314 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1770.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 1,770 PP tok/s 61.4 TG tok/s 2314 TTFT (ms) 71.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xc3keif6",
   "source_url": "https://omlx.ai/benchmarks/performance/xc3keif6",
   "tg_tokens": null,
   "tps": 61.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2xng4ne1-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.2 GB at this context; duplicate traces of the same session: 4f45iybu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1821.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,821 57.8 71.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2xng4ne1",
   "source_url": "https://omlx.ai/benchmarks/performance/2xng4ne1",
   "tg_tokens": null,
   "tps": 57.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2xng4ne1-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.7 GB at this context; duplicate traces of the same session: 4f45iybu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1762.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,762 60.9 72.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2xng4ne1",
   "source_url": "https://omlx.ai/benchmarks/performance/2xng4ne1",
   "tg_tokens": null,
   "tps": 60.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-r9st4m4z-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.2 GB at this context; duplicate traces of the same session: nupevcqg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1319.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,319 47.2 71.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r9st4m4z",
   "source_url": "https://omlx.ai/benchmarks/performance/r9st4m4z",
   "tg_tokens": null,
   "tps": 47.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-r9st4m4z-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.7 GB at this context; duplicate traces of the same session: nupevcqg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1592.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,592 56.3 72.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r9st4m4z",
   "source_url": "https://omlx.ai/benchmarks/performance/r9st4m4z",
   "tg_tokens": null,
   "tps": 56.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tt19r31f-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 195k context; peak memory 83.0 GB; TTFT 90316 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2215.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k tokens 2,215 PP tok/s 63.3 TG tok/s 90316 TTFT (ms) 83.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tt19r31f",
   "source_url": "https://omlx.ai/benchmarks/performance/tt19r31f",
   "tg_tokens": null,
   "tps": 63.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-b6n9wevu-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: zjq1xdkr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 800.2,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 800.2 51.2 33.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b6n9wevu",
   "source_url": "https://omlx.ai/benchmarks/performance/b6n9wevu",
   "tg_tokens": null,
   "tps": 51.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-b6n9wevu-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: zjq1xdkr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 928.0,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 928.0 50.9 34.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b6n9wevu",
   "source_url": "https://omlx.ai/benchmarks/performance/b6n9wevu",
   "tg_tokens": null,
   "tps": 50.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-2unhj5hb-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: an79uakv; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 442.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 442.4 53.8 20.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2unhj5hb",
   "source_url": "https://omlx.ai/benchmarks/performance/2unhj5hb",
   "tg_tokens": null,
   "tps": 53.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-2unhj5hb-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: an79uakv; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 428.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 428.3 51.8 20.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2unhj5hb",
   "source_url": "https://omlx.ai/benchmarks/performance/2unhj5hb",
   "tg_tokens": null,
   "tps": 51.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-xee4jojz-ctx1024",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 87.6 GB at this context; duplicate traces of the same session: b31mb1to; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 260.0,
   "provenance": "community",
   "quant": "OptiQ-2bit",
   "quote": "DeepSeek-V4-Flash-0731-OptiQ-2bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 2bit; 1k 260.0 13.8 87.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xee4jojz",
   "source_url": "https://omlx.ai/benchmarks/performance/xee4jojz",
   "tg_tokens": null,
   "tps": 13.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-xee4jojz-ctx4096",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 88.4 GB at this context; duplicate traces of the same session: b31mb1to; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 258.1,
   "provenance": "community",
   "quant": "OptiQ-2bit",
   "quote": "DeepSeek-V4-Flash-0731-OptiQ-2bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 2bit; 4k 258.1 27.6 88.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xee4jojz",
   "source_url": "https://omlx.ai/benchmarks/performance/xee4jojz",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-i01f55p1-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 172.8 GB; TTFT 2187 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 468.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k tokens 468.3 PP tok/s 50.9 TG tok/s 2187 TTFT (ms) 172.8 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i01f55p1",
   "source_url": "https://omlx.ai/benchmarks/performance/i01f55p1",
   "tg_tokens": null,
   "tps": 50.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-6rx9gon3-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 172.8 GB; TTFT 2184 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 468.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k tokens 468.9 PP tok/s 44.9 TG tok/s 2184 TTFT (ms) 172.8 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6rx9gon3",
   "source_url": "https://omlx.ai/benchmarks/performance/6rx9gon3",
   "tg_tokens": null,
   "tps": 44.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-r91k9mc1-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 174.9 GB; TTFT 45284 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 361.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k tokens 361.8 PP tok/s 54.9 TG tok/s 45284 TTFT (ms) 174.9 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r91k9mc1",
   "source_url": "https://omlx.ai/benchmarks/performance/r91k9mc1",
   "tg_tokens": null,
   "tps": 54.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2wncre7e-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 172.8 GB; TTFT 2203 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 464.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k tokens 464.7 PP tok/s 48.3 TG tok/s 2203 TTFT (ms) 172.8 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2wncre7e",
   "source_url": "https://omlx.ai/benchmarks/performance/2wncre7e",
   "tg_tokens": null,
   "tps": 48.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-u458a6n6-ctx4096",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 7.5 GB at this context; duplicate traces of the same session: k2x8fxbm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1211.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 1,211 109.0 7.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u458a6n6",
   "source_url": "https://omlx.ai/benchmarks/performance/u458a6n6",
   "tg_tokens": null,
   "tps": 109.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-u458a6n6-ctx16384",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 8.2 GB at this context; duplicate traces of the same session: k2x8fxbm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1087.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 1,087 100.0 8.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u458a6n6",
   "source_url": "https://omlx.ai/benchmarks/performance/u458a6n6",
   "tg_tokens": null,
   "tps": 100.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ukqoaar9-ctx1024",
   "model": "MiniMax-M2.7",
   "notes": "oMLX community benchmark trace; peak memory 94.1 GB at this context; duplicate traces of the same session: hgmwkueb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 922.3,
   "provenance": "community",
   "quant": "3bit",
   "quote": "MiniMax-M2.7-3bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 1k 922.3 60.0 94.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ukqoaar9",
   "source_url": "https://omlx.ai/benchmarks/performance/ukqoaar9",
   "tg_tokens": null,
   "tps": 60.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ukqoaar9-ctx4096",
   "model": "MiniMax-M2.7",
   "notes": "oMLX community benchmark trace; peak memory 95.1 GB at this context; duplicate traces of the same session: hgmwkueb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1257.0,
   "provenance": "community",
   "quant": "3bit",
   "quote": "MiniMax-M2.7-3bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 4k 1,257 55.1 95.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ukqoaar9",
   "source_url": "https://omlx.ai/benchmarks/performance/ukqoaar9",
   "tg_tokens": null,
   "tps": 55.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-hujj6zuv-ctx1024",
   "model": "Qwen3-VL-4B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 2.9 GB at this context; duplicate traces of the same session: n2iv7azg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 360.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-VL-4B-Instruct-MLX-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 360.4 54.1 2.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hujj6zuv",
   "source_url": "https://omlx.ai/benchmarks/performance/hujj6zuv",
   "tg_tokens": null,
   "tps": 54.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-hujj6zuv-ctx4096",
   "model": "Qwen3-VL-4B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: n2iv7azg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 328.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-VL-4B-Instruct-MLX-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 328.8 42.4 3.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hujj6zuv",
   "source_url": "https://omlx.ai/benchmarks/performance/hujj6zuv",
   "tg_tokens": null,
   "tps": 42.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-h922fcbe-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 1o87d9da, 5um5je3e, cocon1pw, k0kh1tar; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 831.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 831.8 120.1 21.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h922fcbe",
   "source_url": "https://omlx.ai/benchmarks/performance/h922fcbe",
   "tg_tokens": null,
   "tps": 120.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-h922fcbe-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 1o87d9da, 5um5je3e, cocon1pw, k0kh1tar; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3018.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 3,018 144.0 22.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h922fcbe",
   "source_url": "https://omlx.ai/benchmarks/performance/h922fcbe",
   "tg_tokens": null,
   "tps": 144.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-h922fcbe-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 1o87d9da, 5um5je3e, cocon1pw, k0kh1tar; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5988.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 5,988 47.4 22.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h922fcbe",
   "source_url": "https://omlx.ai/benchmarks/performance/h922fcbe",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-h922fcbe-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 1o87d9da, 5um5je3e, cocon1pw, k0kh1tar; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5530.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 5,530 176.4 22.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h922fcbe",
   "source_url": "https://omlx.ai/benchmarks/performance/h922fcbe",
   "tg_tokens": null,
   "tps": 176.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-h922fcbe-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 1o87d9da, 5um5je3e, cocon1pw, k0kh1tar; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4485.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 4,485 221.0 22.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h922fcbe",
   "source_url": "https://omlx.ai/benchmarks/performance/h922fcbe",
   "tg_tokens": null,
   "tps": 221.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-h922fcbe-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 1o87d9da, 5um5je3e, cocon1pw, k0kh1tar; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3333.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 3,333 142.7 23.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h922fcbe",
   "source_url": "https://omlx.ai/benchmarks/performance/h922fcbe",
   "tg_tokens": null,
   "tps": 142.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1zlv041c-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: ry7i5qkm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 878.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 878.3 33.5 16.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1zlv041c",
   "source_url": "https://omlx.ai/benchmarks/performance/1zlv041c",
   "tg_tokens": null,
   "tps": 33.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1zlv041c-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: ry7i5qkm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 985.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 985.2 32.5 17.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1zlv041c",
   "source_url": "https://omlx.ai/benchmarks/performance/1zlv041c",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3v7oipqz-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: elykoc9g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 775.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 775.6 18.8 28.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3v7oipqz",
   "source_url": "https://omlx.ai/benchmarks/performance/3v7oipqz",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3v7oipqz-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: elykoc9g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 905.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 905.2 18.7 29.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3v7oipqz",
   "source_url": "https://omlx.ai/benchmarks/performance/3v7oipqz",
   "tg_tokens": null,
   "tps": 18.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kpvy93qi-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 32zcmxma, 9ahtr8h9, b5nmsc9l, mev7hi8q, v8awm5uk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 821.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 821.6 41.4 17.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kpvy93qi",
   "source_url": "https://omlx.ai/benchmarks/performance/kpvy93qi",
   "tg_tokens": null,
   "tps": 41.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kpvy93qi-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 32zcmxma, 9ahtr8h9, b5nmsc9l, mev7hi8q, v8awm5uk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1304.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,304 53.1 19.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kpvy93qi",
   "source_url": "https://omlx.ai/benchmarks/performance/kpvy93qi",
   "tg_tokens": null,
   "tps": 53.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-kpvy93qi-ctx8192",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 32zcmxma, 9ahtr8h9, b5nmsc9l, mev7hi8q, v8awm5uk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1393.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,393 60.7 19.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kpvy93qi",
   "source_url": "https://omlx.ai/benchmarks/performance/kpvy93qi",
   "tg_tokens": null,
   "tps": 60.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-kpvy93qi-ctx16384",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 32zcmxma, 9ahtr8h9, b5nmsc9l, mev7hi8q, v8awm5uk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1300.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,300 77.9 19.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kpvy93qi",
   "source_url": "https://omlx.ai/benchmarks/performance/kpvy93qi",
   "tg_tokens": null,
   "tps": 77.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kpvy93qi-ctx32768",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: 32zcmxma, 9ahtr8h9, b5nmsc9l, mev7hi8q, v8awm5uk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1146.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,146 91.9 21.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kpvy93qi",
   "source_url": "https://omlx.ai/benchmarks/performance/kpvy93qi",
   "tg_tokens": null,
   "tps": 91.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-kpvy93qi-ctx65536",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 32zcmxma, 9ahtr8h9, b5nmsc9l, mev7hi8q, v8awm5uk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 963.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 963.9 48.8 23.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kpvy93qi",
   "source_url": "https://omlx.ai/benchmarks/performance/kpvy93qi",
   "tg_tokens": null,
   "tps": 48.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-m7tgyiah-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: 4xejwt5l, c5jr0o2s, da917huc, lpoajzbn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 762.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 762.2 40.1 33.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m7tgyiah",
   "source_url": "https://omlx.ai/benchmarks/performance/m7tgyiah",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-m7tgyiah-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: 4xejwt5l, c5jr0o2s, da917huc, lpoajzbn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 908.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 908.2 48.3 34.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m7tgyiah",
   "source_url": "https://omlx.ai/benchmarks/performance/m7tgyiah",
   "tg_tokens": null,
   "tps": 48.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-m7tgyiah-ctx8192",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: 4xejwt5l, c5jr0o2s, da917huc, lpoajzbn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 832.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 832.7 49.8 35.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m7tgyiah",
   "source_url": "https://omlx.ai/benchmarks/performance/m7tgyiah",
   "tg_tokens": null,
   "tps": 49.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-m7tgyiah-ctx16384",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: 4xejwt5l, c5jr0o2s, da917huc, lpoajzbn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 799.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 799.6 55.8 35.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m7tgyiah",
   "source_url": "https://omlx.ai/benchmarks/performance/m7tgyiah",
   "tg_tokens": null,
   "tps": 55.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-m7tgyiah-ctx32768",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: 4xejwt5l, c5jr0o2s, da917huc, lpoajzbn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 749.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 749.1 82.9 36.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m7tgyiah",
   "source_url": "https://omlx.ai/benchmarks/performance/m7tgyiah",
   "tg_tokens": null,
   "tps": 82.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-m7tgyiah-ctx65536",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: 4xejwt5l, c5jr0o2s, da917huc, lpoajzbn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 665.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 665.6 46.6 39.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m7tgyiah",
   "source_url": "https://omlx.ai/benchmarks/performance/m7tgyiah",
   "tg_tokens": null,
   "tps": 46.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-sq9otir0-ctx4096",
   "model": "Qwen3.8-3.6-27B-blend",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: wtd74y84; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 134.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-3.6-27B-blend-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 134.7 14.7 17.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sq9otir0",
   "source_url": "https://omlx.ai/benchmarks/performance/sq9otir0",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-sq9otir0-ctx16384",
   "model": "Qwen3.8-3.6-27B-blend",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: wtd74y84; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 129.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-3.6-27B-blend-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 129.6 13.7 19.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sq9otir0",
   "source_url": "https://omlx.ai/benchmarks/performance/sq9otir0",
   "tg_tokens": null,
   "tps": 13.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n72499pu-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 72.7 GB; TTFT 8241 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1988.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k tokens 1,988 PP tok/s 30.1 TG tok/s 8241 TTFT (ms) 72.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n72499pu",
   "source_url": "https://omlx.ai/benchmarks/performance/n72499pu",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-eklwx3xf-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 195k context; peak memory 83.1 GB; TTFT 95190 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2101.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k tokens 2,101 PP tok/s 42.5 TG tok/s 95190 TTFT (ms) 83.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eklwx3xf",
   "source_url": "https://omlx.ai/benchmarks/performance/eklwx3xf",
   "tg_tokens": null,
   "tps": 42.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-v5fbinkq-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace, single run at 8k context; peak memory 143.4 GB; TTFT 2099 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3903.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 8k tokens 3,903 PP tok/s 109.5 TG tok/s 2099 TTFT (ms) 143.4 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v5fbinkq",
   "source_url": "https://omlx.ai/benchmarks/performance/v5fbinkq",
   "tg_tokens": null,
   "tps": 109.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-74kae7mn-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.8 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2027.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 1k 2,027 114.2 142.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/74kae7mn",
   "source_url": "https://omlx.ai/benchmarks/performance/74kae7mn",
   "tg_tokens": null,
   "tps": 114.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-74kae7mn-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3350.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 4k 3,350 116.6 143.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/74kae7mn",
   "source_url": "https://omlx.ai/benchmarks/performance/74kae7mn",
   "tg_tokens": null,
   "tps": 116.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9ehhwibl-ctx131072",
   "model": "Gemma-4-31B",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 36.5 GB; TTFT 345448 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 379.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31b-it-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k tokens 379.4 PP tok/s 10.8 TG tok/s 345448 TTFT (ms) 36.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ehhwibl",
   "source_url": "https://omlx.ai/benchmarks/performance/9ehhwibl",
   "tg_tokens": null,
   "tps": 10.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-b8lx1kl5-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: eeot0vx3, td8b78jj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 941.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mlx-community--Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 941.6 18.7 29.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b8lx1kl5",
   "source_url": "https://omlx.ai/benchmarks/performance/b8lx1kl5",
   "tg_tokens": null,
   "tps": 18.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-b8lx1kl5-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: eeot0vx3, td8b78jj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 729.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mlx-community--Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 729.9 18.0 30.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b8lx1kl5",
   "source_url": "https://omlx.ai/benchmarks/performance/b8lx1kl5",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-b8lx1kl5-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 33.3 GB at this context; duplicate traces of the same session: eeot0vx3, td8b78jj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 582.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mlx-community--Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 582.7 15.5 33.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b8lx1kl5",
   "source_url": "https://omlx.ai/benchmarks/performance/b8lx1kl5",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-b31mb1to-ctx1024",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 87.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 260.0,
   "provenance": "community",
   "quant": "OptiQ-2bit",
   "quote": "DeepSeek-V4-Flash-0731-OptiQ-2bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 2bit; 1k 260.0 13.8 87.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b31mb1to",
   "source_url": "https://omlx.ai/benchmarks/performance/b31mb1to",
   "tg_tokens": null,
   "tps": 13.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-b31mb1to-ctx4096",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 88.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 258.1,
   "provenance": "community",
   "quant": "OptiQ-2bit",
   "quote": "DeepSeek-V4-Flash-0731-OptiQ-2bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 2bit; 4k 258.1 27.6 88.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b31mb1to",
   "source_url": "https://omlx.ai/benchmarks/performance/b31mb1to",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2iyrjqfx-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 175.1 GB; TTFT 43597 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 375.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k tokens 375.8 PP tok/s 36.2 TG tok/s 43597 TTFT (ms) 175.1 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2iyrjqfx",
   "source_url": "https://omlx.ai/benchmarks/performance/2iyrjqfx",
   "tg_tokens": null,
   "tps": 36.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-90dondc4-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Jundot build; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: fhddr4v7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 393.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 393.2 30.5 70.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/90dondc4",
   "source_url": "https://omlx.ai/benchmarks/performance/90dondc4",
   "tg_tokens": null,
   "tps": 30.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-90dondc4-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Jundot build; oMLX community benchmark trace; peak memory 80.4 GB at this context; duplicate traces of the same session: fhddr4v7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2176.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,176 43.4 80.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/90dondc4",
   "source_url": "https://omlx.ai/benchmarks/performance/90dondc4",
   "tg_tokens": null,
   "tps": 43.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-k2x8fxbm-ctx4096",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 7.5 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1211.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 1,211 109.0 7.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k2x8fxbm",
   "source_url": "https://omlx.ai/benchmarks/performance/k2x8fxbm",
   "tg_tokens": null,
   "tps": 109.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-k2x8fxbm-ctx16384",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 8.2 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1087.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 1,087 100.0 8.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k2x8fxbm",
   "source_url": "https://omlx.ai/benchmarks/performance/k2x8fxbm",
   "tg_tokens": null,
   "tps": 100.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hgmwkueb-ctx1024",
   "model": "MiniMax-M2.7",
   "notes": "oMLX community benchmark trace; peak memory 94.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 922.3,
   "provenance": "community",
   "quant": "3bit",
   "quote": "MiniMax-M2.7-3bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 1k 922.3 60.0 94.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hgmwkueb",
   "source_url": "https://omlx.ai/benchmarks/performance/hgmwkueb",
   "tg_tokens": null,
   "tps": 60.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hgmwkueb-ctx4096",
   "model": "MiniMax-M2.7",
   "notes": "oMLX community benchmark trace; peak memory 95.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1257.0,
   "provenance": "community",
   "quant": "3bit",
   "quote": "MiniMax-M2.7-3bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 4k 1,257 55.1 95.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hgmwkueb",
   "source_url": "https://omlx.ai/benchmarks/performance/hgmwkueb",
   "tg_tokens": null,
   "tps": 55.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-rswge00b-ctx1024",
   "model": "Occamy-1.0",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: 33wwatx3, 4561q73z, ey6akcji, khdbryfv, p8nn7fn3; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1386.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Occamy-1.0-oQ8-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,386 68.9 35.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rswge00b",
   "source_url": "https://omlx.ai/benchmarks/performance/rswge00b",
   "tg_tokens": null,
   "tps": 68.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-rswge00b-ctx4096",
   "model": "Occamy-1.0",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: 33wwatx3, 4561q73z, ey6akcji, khdbryfv, p8nn7fn3; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1296.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Occamy-1.0-oQ8-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 1,296 57.3 37.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rswge00b",
   "source_url": "https://omlx.ai/benchmarks/performance/rswge00b",
   "tg_tokens": null,
   "tps": 57.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-rswge00b-ctx8192",
   "model": "Occamy-1.0",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 37.9 GB at this context; duplicate traces of the same session: 33wwatx3, 4561q73z, ey6akcji, khdbryfv, p8nn7fn3; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1026.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Occamy-1.0-oQ8-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 8k 1,026 58.8 37.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rswge00b",
   "source_url": "https://omlx.ai/benchmarks/performance/rswge00b",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-rswge00b-ctx16384",
   "model": "Occamy-1.0",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 38.7 GB at this context; duplicate traces of the same session: 33wwatx3, 4561q73z, ey6akcji, khdbryfv, p8nn7fn3; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 960.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Occamy-1.0-oQ8-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 16k 960.5 50.3 38.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rswge00b",
   "source_url": "https://omlx.ai/benchmarks/performance/rswge00b",
   "tg_tokens": null,
   "tps": 50.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-rswge00b-ctx32768",
   "model": "Occamy-1.0",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 40.1 GB at this context; duplicate traces of the same session: 33wwatx3, 4561q73z, ey6akcji, khdbryfv, p8nn7fn3; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 792.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Occamy-1.0-oQ8-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 32k 792.0 49.6 40.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rswge00b",
   "source_url": "https://omlx.ai/benchmarks/performance/rswge00b",
   "tg_tokens": null,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-rswge00b-ctx65536",
   "model": "Occamy-1.0",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 41.2 GB at this context; duplicate traces of the same session: 33wwatx3, 4561q73z, ey6akcji, khdbryfv, p8nn7fn3; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 590.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Occamy-1.0-oQ8-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 64k 590.2 39.0 41.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rswge00b",
   "source_url": "https://omlx.ai/benchmarks/performance/rswge00b",
   "tg_tokens": null,
   "tps": 39.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zowtuyjy-ctx131072",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 30.6 GB at this context; duplicate traces of the same session: jc47sp5j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2313.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 128k 2,313 55.0 30.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zowtuyjy",
   "source_url": "https://omlx.ai/benchmarks/performance/zowtuyjy",
   "tg_tokens": null,
   "tps": 55.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-zowtuyjy-ctx199680",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 31.9 GB at this context; duplicate traces of the same session: jc47sp5j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1773.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 195k 1,773 52.9 31.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zowtuyjy",
   "source_url": "https://omlx.ai/benchmarks/performance/zowtuyjy",
   "tg_tokens": null,
   "tps": 52.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-a7sfim6h-ctx8192",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.3 GB at this context; duplicate traces of the same session: 21sqjsby, f4bo9r0d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 948.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 948.9 84.9 7.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a7sfim6h",
   "source_url": "https://omlx.ai/benchmarks/performance/a7sfim6h",
   "tg_tokens": null,
   "tps": 84.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-a7sfim6h-ctx65536",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.3 GB at this context; duplicate traces of the same session: 21sqjsby, f4bo9r0d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 731.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 731.5 61.6 10.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a7sfim6h",
   "source_url": "https://omlx.ai/benchmarks/performance/a7sfim6h",
   "tg_tokens": null,
   "tps": 61.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-a7sfim6h-ctx131072",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 12.9 GB at this context; duplicate traces of the same session: 21sqjsby, f4bo9r0d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 576.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 576.6 46.7 12.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a7sfim6h",
   "source_url": "https://omlx.ai/benchmarks/performance/a7sfim6h",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3p7zqn6v-ctx131072",
   "model": "Qwen-AgentWorld-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: n70dun8l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2525.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen-AgentWorld-35B-A3B-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 2,525 71.6 23.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3p7zqn6v",
   "source_url": "https://omlx.ai/benchmarks/performance/3p7zqn6v",
   "tg_tokens": null,
   "tps": 71.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3p7zqn6v-ctx199680",
   "model": "Qwen-AgentWorld-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: n70dun8l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1916.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen-AgentWorld-35B-A3B-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 195k 1,916 56.5 24.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3p7zqn6v",
   "source_url": "https://omlx.ai/benchmarks/performance/3p7zqn6v",
   "tg_tokens": null,
   "tps": 56.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-n2iv7azg-ctx1024",
   "model": "Qwen3-VL-4B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 2.9 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 360.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-VL-4B-Instruct-MLX-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 360.4 54.1 2.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n2iv7azg",
   "source_url": "https://omlx.ai/benchmarks/performance/n2iv7azg",
   "tg_tokens": null,
   "tps": 54.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-n2iv7azg-ctx4096",
   "model": "Qwen3-VL-4B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 328.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-VL-4B-Instruct-MLX-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 328.8 42.4 3.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n2iv7azg",
   "source_url": "https://omlx.ai/benchmarks/performance/n2iv7azg",
   "tg_tokens": null,
   "tps": 42.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-eru3gf0a-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1027.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 1,027 67.7 22.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eru3gf0a",
   "source_url": "https://omlx.ai/benchmarks/performance/eru3gf0a",
   "tg_tokens": null,
   "tps": 67.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-eru3gf0a-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 878.1,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 878.1 59.5 23.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eru3gf0a",
   "source_url": "https://omlx.ai/benchmarks/performance/eru3gf0a",
   "tg_tokens": null,
   "tps": 59.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-k0kh1tar-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 1o87d9da, 5um5je3e, cocon1pw, wqjtgpdr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 831.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 831.8 120.1 21.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k0kh1tar",
   "source_url": "https://omlx.ai/benchmarks/performance/k0kh1tar",
   "tg_tokens": null,
   "tps": 120.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-k0kh1tar-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 1o87d9da, 5um5je3e, cocon1pw, wqjtgpdr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3018.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 3,018 144.0 22.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k0kh1tar",
   "source_url": "https://omlx.ai/benchmarks/performance/k0kh1tar",
   "tg_tokens": null,
   "tps": 144.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-k0kh1tar-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 1o87d9da, 5um5je3e, cocon1pw, wqjtgpdr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5988.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 5,988 47.4 22.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k0kh1tar",
   "source_url": "https://omlx.ai/benchmarks/performance/k0kh1tar",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-k0kh1tar-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 1o87d9da, 5um5je3e, cocon1pw, wqjtgpdr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5530.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 5,530 176.4 22.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k0kh1tar",
   "source_url": "https://omlx.ai/benchmarks/performance/k0kh1tar",
   "tg_tokens": null,
   "tps": 176.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-k0kh1tar-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 1o87d9da, 5um5je3e, cocon1pw, wqjtgpdr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4485.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 4,485 221.0 22.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k0kh1tar",
   "source_url": "https://omlx.ai/benchmarks/performance/k0kh1tar",
   "tg_tokens": null,
   "tps": 221.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-k0kh1tar-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 1o87d9da, 5um5je3e, cocon1pw, wqjtgpdr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3333.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 3,333 142.7 23.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k0kh1tar",
   "source_url": "https://omlx.ai/benchmarks/performance/k0kh1tar",
   "tg_tokens": null,
   "tps": 142.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-m0y7q4gy-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: gjqyada2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 135.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 135.2 15.2 17.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m0y7q4gy",
   "source_url": "https://omlx.ai/benchmarks/performance/m0y7q4gy",
   "tg_tokens": null,
   "tps": 15.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-m0y7q4gy-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: gjqyada2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 129.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 129.2 14.0 19.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m0y7q4gy",
   "source_url": "https://omlx.ai/benchmarks/performance/m0y7q4gy",
   "tg_tokens": null,
   "tps": 14.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ry7i5qkm-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 878.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 878.3 33.5 16.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ry7i5qkm",
   "source_url": "https://omlx.ai/benchmarks/performance/ry7i5qkm",
   "tg_tokens": null,
   "tps": 33.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ry7i5qkm-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 985.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 985.2 32.5 17.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ry7i5qkm",
   "source_url": "https://omlx.ai/benchmarks/performance/ry7i5qkm",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-8molh42c-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: tnkkmlqe; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 8bit; 1k 224.1 21.1 28.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8molh42c",
   "source_url": "https://omlx.ai/benchmarks/performance/8molh42c",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-8molh42c-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.9 GB at this context; duplicate traces of the same session: tnkkmlqe; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 8bit; 4k 224.7 20.5 31.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8molh42c",
   "source_url": "https://omlx.ai/benchmarks/performance/8molh42c",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-elykoc9g-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 775.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 775.6 18.8 28.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/elykoc9g",
   "source_url": "https://omlx.ai/benchmarks/performance/elykoc9g",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-elykoc9g-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 905.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 905.2 18.7 29.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/elykoc9g",
   "source_url": "https://omlx.ai/benchmarks/performance/elykoc9g",
   "tg_tokens": null,
   "tps": 18.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-30bqj3s3-ctx199680",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 195k context; peak memory 49.2 GB; TTFT 842718 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 237.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 195k tokens 237.3 PP tok/s 6.5 TG tok/s 842718 TTFT (ms) 49.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/30bqj3s3",
   "source_url": "https://omlx.ai/benchmarks/performance/30bqj3s3",
   "tg_tokens": null,
   "tps": 6.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9ahtr8h9-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 32zcmxma, b5nmsc9l, mev7hi8q, v8awm5uk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 821.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 821.6 41.4 17.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ahtr8h9",
   "source_url": "https://omlx.ai/benchmarks/performance/9ahtr8h9",
   "tg_tokens": null,
   "tps": 41.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9ahtr8h9-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 32zcmxma, b5nmsc9l, mev7hi8q, v8awm5uk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1304.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,304 53.1 19.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ahtr8h9",
   "source_url": "https://omlx.ai/benchmarks/performance/9ahtr8h9",
   "tg_tokens": null,
   "tps": 53.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9ahtr8h9-ctx8192",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 32zcmxma, b5nmsc9l, mev7hi8q, v8awm5uk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1393.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,393 60.7 19.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ahtr8h9",
   "source_url": "https://omlx.ai/benchmarks/performance/9ahtr8h9",
   "tg_tokens": null,
   "tps": 60.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9ahtr8h9-ctx16384",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 32zcmxma, b5nmsc9l, mev7hi8q, v8awm5uk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1300.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,300 77.9 19.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ahtr8h9",
   "source_url": "https://omlx.ai/benchmarks/performance/9ahtr8h9",
   "tg_tokens": null,
   "tps": 77.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9ahtr8h9-ctx32768",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: 32zcmxma, b5nmsc9l, mev7hi8q, v8awm5uk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1146.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,146 91.9 21.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ahtr8h9",
   "source_url": "https://omlx.ai/benchmarks/performance/9ahtr8h9",
   "tg_tokens": null,
   "tps": 91.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9ahtr8h9-ctx65536",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 32zcmxma, b5nmsc9l, mev7hi8q, v8awm5uk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 963.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 963.9 48.8 23.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ahtr8h9",
   "source_url": "https://omlx.ai/benchmarks/performance/9ahtr8h9",
   "tg_tokens": null,
   "tps": 48.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-da917huc-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: 4xejwt5l, c5jr0o2s, lpoajzbn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 762.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 762.2 40.1 33.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/da917huc",
   "source_url": "https://omlx.ai/benchmarks/performance/da917huc",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-da917huc-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: 4xejwt5l, c5jr0o2s, lpoajzbn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 908.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 908.2 48.3 34.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/da917huc",
   "source_url": "https://omlx.ai/benchmarks/performance/da917huc",
   "tg_tokens": null,
   "tps": 48.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-da917huc-ctx8192",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: 4xejwt5l, c5jr0o2s, lpoajzbn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 832.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 832.7 49.8 35.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/da917huc",
   "source_url": "https://omlx.ai/benchmarks/performance/da917huc",
   "tg_tokens": null,
   "tps": 49.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-da917huc-ctx16384",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: 4xejwt5l, c5jr0o2s, lpoajzbn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 799.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 799.6 55.8 35.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/da917huc",
   "source_url": "https://omlx.ai/benchmarks/performance/da917huc",
   "tg_tokens": null,
   "tps": 55.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-da917huc-ctx32768",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: 4xejwt5l, c5jr0o2s, lpoajzbn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 749.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 749.1 82.9 36.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/da917huc",
   "source_url": "https://omlx.ai/benchmarks/performance/da917huc",
   "tg_tokens": null,
   "tps": 82.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-da917huc-ctx65536",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: 4xejwt5l, c5jr0o2s, lpoajzbn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 665.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 665.6 46.6 39.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/da917huc",
   "source_url": "https://omlx.ai/benchmarks/performance/da917huc",
   "tg_tokens": null,
   "tps": 46.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-x5p6n06b-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 23.2 GB; TTFT 39194 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 418.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k tokens 418.0 PP tok/s 99.4 TG tok/s 39194 TTFT (ms) 23.2 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x5p6n06b",
   "source_url": "https://omlx.ai/benchmarks/performance/x5p6n06b",
   "tg_tokens": null,
   "tps": 99.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-bw3g4k5q-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 23.2 GB; TTFT 39331 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 416.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k tokens 416.6 PP tok/s 33.6 TG tok/s 39331 TTFT (ms) 23.2 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bw3g4k5q",
   "source_url": "https://omlx.ai/benchmarks/performance/bw3g4k5q",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-0lnvz8c7-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 23.2 GB; TTFT 39203 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 417.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k tokens 417.9 PP tok/s 93.9 TG tok/s 39203 TTFT (ms) 23.2 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0lnvz8c7",
   "source_url": "https://omlx.ai/benchmarks/performance/0lnvz8c7",
   "tg_tokens": null,
   "tps": 93.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bsvc4w97-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 23.2 GB; TTFT 39284 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 417.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k tokens 417.1 PP tok/s 105.8 TG tok/s 39284 TTFT (ms) 23.2 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bsvc4w97",
   "source_url": "https://omlx.ai/benchmarks/performance/bsvc4w97",
   "tg_tokens": null,
   "tps": 105.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-s16phv7z-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 23.2 GB; TTFT 39184 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 418.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k tokens 418.1 PP tok/s 106.2 TG tok/s 39184 TTFT (ms) 23.2 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s16phv7z",
   "source_url": "https://omlx.ai/benchmarks/performance/s16phv7z",
   "tg_tokens": null,
   "tps": 106.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-u5eajlv4-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 16.6 GB; TTFT 2336 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 438.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k tokens 438.4 PP tok/s 37.9 TG tok/s 2336 TTFT (ms) 16.6 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u5eajlv4",
   "source_url": "https://omlx.ai/benchmarks/performance/u5eajlv4",
   "tg_tokens": null,
   "tps": 37.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9fqsovzq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 16.6 GB; TTFT 2339 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 437.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k tokens 437.9 PP tok/s 38.0 TG tok/s 2339 TTFT (ms) 16.6 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9fqsovzq",
   "source_url": "https://omlx.ai/benchmarks/performance/9fqsovzq",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-d56qm899-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 40.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 249.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 249.2 30.4 40.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d56qm899",
   "source_url": "https://omlx.ai/benchmarks/performance/d56qm899",
   "tg_tokens": null,
   "tps": 30.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-d56qm899-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 327.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 327.4 27.4 43.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d56qm899",
   "source_url": "https://omlx.ai/benchmarks/performance/d56qm899",
   "tg_tokens": null,
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-wtd74y84-ctx4096",
   "model": "Qwen3.8-3.6-27B-blend",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 134.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-3.6-27B-blend-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 134.7 14.7 17.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wtd74y84",
   "source_url": "https://omlx.ai/benchmarks/performance/wtd74y84",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-wtd74y84-ctx16384",
   "model": "Qwen3.8-3.6-27B-blend",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 129.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-3.6-27B-blend-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 129.6 13.7 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wtd74y84",
   "source_url": "https://omlx.ai/benchmarks/performance/wtd74y84",
   "tg_tokens": null,
   "tps": 13.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-wf0r7dhu-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.7 GB; TTFT 3166 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 323.4 PP tok/s 42.6 TG tok/s 3166 TTFT (ms) 70.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wf0r7dhu",
   "source_url": "https://omlx.ai/benchmarks/performance/wf0r7dhu",
   "tg_tokens": null,
   "tps": 42.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7b67w48l-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 72.7 GB; TTFT 9469 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1730.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k tokens 1,730 PP tok/s 19.0 TG tok/s 9469 TTFT (ms) 72.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7b67w48l",
   "source_url": "https://omlx.ai/benchmarks/performance/7b67w48l",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yy063icj-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: 8w5usze5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1768.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,768 72.9 71.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yy063icj",
   "source_url": "https://omlx.ai/benchmarks/performance/yy063icj",
   "tg_tokens": null,
   "tps": 72.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yy063icj-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.7 GB at this context; duplicate traces of the same session: 8w5usze5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2508.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,508 83.4 72.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yy063icj",
   "source_url": "https://omlx.ai/benchmarks/performance/yy063icj",
   "tg_tokens": null,
   "tps": 83.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-y6v2tk2f-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 71.0 GB; TTFT 2260 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1813.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 1,813 PP tok/s 71.3 TG tok/s 2260 TTFT (ms) 71.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y6v2tk2f",
   "source_url": "https://omlx.ai/benchmarks/performance/y6v2tk2f",
   "tg_tokens": null,
   "tps": 71.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-skzmkc4v-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 144.8 GB at this context; duplicate traces of the same session: llsr8qjq; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3918.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 8k 3,918 140.1 144.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/skzmkc4v",
   "source_url": "https://omlx.ai/benchmarks/performance/skzmkc4v",
   "tg_tokens": null,
   "tps": 140.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-skzmkc4v-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 146.8 GB at this context; duplicate traces of the same session: llsr8qjq; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3881.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 64k 3,881 114.0 146.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/skzmkc4v",
   "source_url": "https://omlx.ai/benchmarks/performance/skzmkc4v",
   "tg_tokens": null,
   "tps": 114.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-g2blp17z-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.8 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2027.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 1k 2,027 114.2 142.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g2blp17z",
   "source_url": "https://omlx.ai/benchmarks/performance/g2blp17z",
   "tg_tokens": null,
   "tps": 114.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-g2blp17z-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3350.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 4k 3,350 116.6 143.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g2blp17z",
   "source_url": "https://omlx.ai/benchmarks/performance/g2blp17z",
   "tg_tokens": null,
   "tps": 116.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-59mpspil-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: u3a5cp9s, wugry8vn; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 417.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 417.3 43.6 21.5 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/59mpspil",
   "source_url": "https://omlx.ai/benchmarks/performance/59mpspil",
   "tg_tokens": null,
   "tps": 43.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-59mpspil-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: u3a5cp9s, wugry8vn; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 405.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 405.3 46.3 22.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/59mpspil",
   "source_url": "https://omlx.ai/benchmarks/performance/59mpspil",
   "tg_tokens": null,
   "tps": 46.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-59mpspil-ctx8192",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: u3a5cp9s, wugry8vn; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 377.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 8k 377.1 40.0 22.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/59mpspil",
   "source_url": "https://omlx.ai/benchmarks/performance/59mpspil",
   "tg_tokens": null,
   "tps": 40.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-59mpspil-ctx16384",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: u3a5cp9s, wugry8vn; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 321.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k 321.1 34.2 22.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/59mpspil",
   "source_url": "https://omlx.ai/benchmarks/performance/59mpspil",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-td8b78jj-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: eeot0vx3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 941.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mlx-community--Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 941.6 18.7 29.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/td8b78jj",
   "source_url": "https://omlx.ai/benchmarks/performance/td8b78jj",
   "tg_tokens": null,
   "tps": 18.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-td8b78jj-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: eeot0vx3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 729.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mlx-community--Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 729.9 18.0 30.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/td8b78jj",
   "source_url": "https://omlx.ai/benchmarks/performance/td8b78jj",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-td8b78jj-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 33.3 GB at this context; duplicate traces of the same session: eeot0vx3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 582.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mlx-community--Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 582.7 15.5 33.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/td8b78jj",
   "source_url": "https://omlx.ai/benchmarks/performance/td8b78jj",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-0katq936-ctx199680",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace, single run at 195k context; peak memory 159.3 GB; TTFT 261778 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 764.0,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "DeepSeek-V4-Flash-0731-MXFP4-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k tokens 764.0 PP tok/s 40.2 TG tok/s 261778 TTFT (ms) 159.3 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0katq936",
   "source_url": "https://omlx.ai/benchmarks/performance/0katq936",
   "tg_tokens": null,
   "tps": 40.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-6rz5xgz3-ctx65536",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace, single run at 64k context; peak memory 180.0 GB; TTFT 29143 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2249.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k tokens 2,249 PP tok/s 76.5 TG tok/s 29143 TTFT (ms) 180.0 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6rz5xgz3",
   "source_url": "https://omlx.ai/benchmarks/performance/6rz5xgz3",
   "tg_tokens": null,
   "tps": 76.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-wvj3m0id-ctx65536",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 179.6 GB; TTFT 28603 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2291.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k tokens 2,291 PP tok/s 56.5 TG tok/s 28603 TTFT (ms) 179.6 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wvj3m0id",
   "source_url": "https://omlx.ai/benchmarks/performance/wvj3m0id",
   "tg_tokens": null,
   "tps": 56.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-sfe77eo8-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 175.5 GB; TTFT 28779 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2277.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k tokens 2,277 PP tok/s 56.4 TG tok/s 28779 TTFT (ms) 175.5 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sfe77eo8",
   "source_url": "https://omlx.ai/benchmarks/performance/sfe77eo8",
   "tg_tokens": null,
   "tps": 56.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fhddr4v7-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Jundot build; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 393.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 393.2 30.5 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fhddr4v7",
   "source_url": "https://omlx.ai/benchmarks/performance/fhddr4v7",
   "tg_tokens": null,
   "tps": 30.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fhddr4v7-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Jundot build; oMLX community benchmark trace; peak memory 80.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2176.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,176 43.4 80.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fhddr4v7",
   "source_url": "https://omlx.ai/benchmarks/performance/fhddr4v7",
   "tg_tokens": null,
   "tps": 43.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-bdurdgq8-ctx1024",
   "model": "Ling-3.0-flash",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 68.3 GB at this context; duplicate traces of the same session: 77fw40bb, 8lztys7d, i4413e2f, pqds84a8; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 594.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-flash-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 594.8 41.4 68.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bdurdgq8",
   "source_url": "https://omlx.ai/benchmarks/performance/bdurdgq8",
   "tg_tokens": null,
   "tps": 41.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-bdurdgq8-ctx4096",
   "model": "Ling-3.0-flash",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 69.5 GB at this context; duplicate traces of the same session: 77fw40bb, 8lztys7d, i4413e2f, pqds84a8; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 525.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-flash-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 525.3 37.9 69.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bdurdgq8",
   "source_url": "https://omlx.ai/benchmarks/performance/bdurdgq8",
   "tg_tokens": null,
   "tps": 37.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-bdurdgq8-ctx8192",
   "model": "Ling-3.0-flash",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 70.2 GB at this context; duplicate traces of the same session: 77fw40bb, 8lztys7d, i4413e2f, pqds84a8; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 505.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-flash-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 505.0 37.0 70.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bdurdgq8",
   "source_url": "https://omlx.ai/benchmarks/performance/bdurdgq8",
   "tg_tokens": null,
   "tps": 37.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-bdurdgq8-ctx16384",
   "model": "Ling-3.0-flash",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: 77fw40bb, 8lztys7d, i4413e2f, pqds84a8; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 476.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-flash-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 476.4 33.0 71.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bdurdgq8",
   "source_url": "https://omlx.ai/benchmarks/performance/bdurdgq8",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-bdurdgq8-ctx32768",
   "model": "Ling-3.0-flash",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 74.7 GB at this context; duplicate traces of the same session: 77fw40bb, 8lztys7d, i4413e2f, pqds84a8; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 427.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-flash-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 32k 427.5 32.6 74.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bdurdgq8",
   "source_url": "https://omlx.ai/benchmarks/performance/bdurdgq8",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-3ttk913m-ctx1024",
   "model": "llama-3-1-8b-instruct",
   "notes": "oMLX community benchmark trace; peak memory 15.5 GB at this context; duplicate traces of the same session: dioxvh5t; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 405.5,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Llama-3.1-8B-Instruct on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 unknown; 1k 405.5 16.2 15.5 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ttk913m",
   "source_url": "https://omlx.ai/benchmarks/performance/3ttk913m",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-3ttk913m-ctx4096",
   "model": "llama-3-1-8b-instruct",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: dioxvh5t; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 399.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Llama-3.1-8B-Instruct on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 unknown; 4k 399.0 15.8 16.0 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ttk913m",
   "source_url": "https://omlx.ai/benchmarks/performance/3ttk913m",
   "tg_tokens": null,
   "tps": 15.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-khdbryfv-ctx1024",
   "model": "Occamy-1.0",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: 33wwatx3, 4561q73z, ey6akcji, p8nn7fn3; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1386.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Occamy-1.0-oQ8-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,386 68.9 35.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/khdbryfv",
   "source_url": "https://omlx.ai/benchmarks/performance/khdbryfv",
   "tg_tokens": null,
   "tps": 68.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-khdbryfv-ctx4096",
   "model": "Occamy-1.0",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: 33wwatx3, 4561q73z, ey6akcji, p8nn7fn3; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1296.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Occamy-1.0-oQ8-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 1,296 57.3 37.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/khdbryfv",
   "source_url": "https://omlx.ai/benchmarks/performance/khdbryfv",
   "tg_tokens": null,
   "tps": 57.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-khdbryfv-ctx8192",
   "model": "Occamy-1.0",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 37.9 GB at this context; duplicate traces of the same session: 33wwatx3, 4561q73z, ey6akcji, p8nn7fn3; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1026.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Occamy-1.0-oQ8-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 8k 1,026 58.8 37.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/khdbryfv",
   "source_url": "https://omlx.ai/benchmarks/performance/khdbryfv",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-khdbryfv-ctx16384",
   "model": "Occamy-1.0",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 38.7 GB at this context; duplicate traces of the same session: 33wwatx3, 4561q73z, ey6akcji, p8nn7fn3; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 960.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Occamy-1.0-oQ8-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 16k 960.5 50.3 38.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/khdbryfv",
   "source_url": "https://omlx.ai/benchmarks/performance/khdbryfv",
   "tg_tokens": null,
   "tps": 50.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-khdbryfv-ctx32768",
   "model": "Occamy-1.0",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 40.1 GB at this context; duplicate traces of the same session: 33wwatx3, 4561q73z, ey6akcji, p8nn7fn3; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 792.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Occamy-1.0-oQ8-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 32k 792.0 49.6 40.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/khdbryfv",
   "source_url": "https://omlx.ai/benchmarks/performance/khdbryfv",
   "tg_tokens": null,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-khdbryfv-ctx65536",
   "model": "Occamy-1.0",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 41.2 GB at this context; duplicate traces of the same session: 33wwatx3, 4561q73z, ey6akcji, p8nn7fn3; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 590.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Occamy-1.0-oQ8-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 64k 590.2 39.0 41.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/khdbryfv",
   "source_url": "https://omlx.ai/benchmarks/performance/khdbryfv",
   "tg_tokens": null,
   "tps": 39.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-mve95bg6-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 6wayt66a, xfpwttad; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 919.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 919.0 90.1 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mve95bg6",
   "source_url": "https://omlx.ai/benchmarks/performance/mve95bg6",
   "tg_tokens": null,
   "tps": 90.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-mve95bg6-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 6wayt66a, xfpwttad; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1082.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,082 85.8 21.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mve95bg6",
   "source_url": "https://omlx.ai/benchmarks/performance/mve95bg6",
   "tg_tokens": null,
   "tps": 85.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-mve95bg6-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 6wayt66a, xfpwttad; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 998.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 998.5 80.2 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mve95bg6",
   "source_url": "https://omlx.ai/benchmarks/performance/mve95bg6",
   "tg_tokens": null,
   "tps": 80.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-mve95bg6-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 6wayt66a, xfpwttad; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 835.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 835.4 73.0 21.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mve95bg6",
   "source_url": "https://omlx.ai/benchmarks/performance/mve95bg6",
   "tg_tokens": null,
   "tps": 73.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-rjdxr9qd-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: zb1gaw3m; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 931.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 931.6 88.4 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rjdxr9qd",
   "source_url": "https://omlx.ai/benchmarks/performance/rjdxr9qd",
   "tg_tokens": null,
   "tps": 88.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-rjdxr9qd-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: zb1gaw3m; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1069.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,069 88.5 21.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rjdxr9qd",
   "source_url": "https://omlx.ai/benchmarks/performance/rjdxr9qd",
   "tg_tokens": null,
   "tps": 88.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-rjdxr9qd-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: zb1gaw3m; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1014.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 1,014 82.5 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rjdxr9qd",
   "source_url": "https://omlx.ai/benchmarks/performance/rjdxr9qd",
   "tg_tokens": null,
   "tps": 82.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-rjdxr9qd-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: zb1gaw3m; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 869.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 869.8 74.5 22.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rjdxr9qd",
   "source_url": "https://omlx.ai/benchmarks/performance/rjdxr9qd",
   "tg_tokens": null,
   "tps": 74.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jc47sp5j-ctx131072",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 30.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2313.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 128k 2,313 55.0 30.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jc47sp5j",
   "source_url": "https://omlx.ai/benchmarks/performance/jc47sp5j",
   "tg_tokens": null,
   "tps": 55.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jc47sp5j-ctx199680",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 31.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1773.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 195k 1,773 52.9 31.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jc47sp5j",
   "source_url": "https://omlx.ai/benchmarks/performance/jc47sp5j",
   "tg_tokens": null,
   "tps": 52.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-21sqjsby-ctx8192",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.3 GB at this context; duplicate traces of the same session: f4bo9r0d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 948.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 948.9 84.9 7.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/21sqjsby",
   "source_url": "https://omlx.ai/benchmarks/performance/21sqjsby",
   "tg_tokens": null,
   "tps": 84.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-21sqjsby-ctx65536",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.3 GB at this context; duplicate traces of the same session: f4bo9r0d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 731.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 731.5 61.6 10.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/21sqjsby",
   "source_url": "https://omlx.ai/benchmarks/performance/21sqjsby",
   "tg_tokens": null,
   "tps": 61.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-21sqjsby-ctx131072",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 12.9 GB at this context; duplicate traces of the same session: f4bo9r0d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 576.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 576.6 46.7 12.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/21sqjsby",
   "source_url": "https://omlx.ai/benchmarks/performance/21sqjsby",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-n70dun8l-ctx131072",
   "model": "Qwen-AgentWorld-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2525.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen-AgentWorld-35B-A3B-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 2,525 71.6 23.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n70dun8l",
   "source_url": "https://omlx.ai/benchmarks/performance/n70dun8l",
   "tg_tokens": null,
   "tps": 71.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-n70dun8l-ctx199680",
   "model": "Qwen-AgentWorld-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1916.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen-AgentWorld-35B-A3B-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 195k 1,916 56.5 24.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n70dun8l",
   "source_url": "https://omlx.ai/benchmarks/performance/n70dun8l",
   "tg_tokens": null,
   "tps": 56.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-0ew50dur-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: lw43b1wp; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 165.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (8c); M4 (8c) \u00b7 16 GB \u00b7 4bit; 1k 165.4 19.2 5.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ew50dur",
   "source_url": "https://omlx.ai/benchmarks/performance/0ew50dur",
   "tg_tokens": null,
   "tps": 19.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-0ew50dur-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: lw43b1wp; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 167.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (8c); M4 (8c) \u00b7 16 GB \u00b7 4bit; 4k 167.8 16.7 6.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ew50dur",
   "source_url": "https://omlx.ai/benchmarks/performance/0ew50dur",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-i8loe3o7-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 36.4 GB; TTFT 3304 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4959.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k tokens 4,959 PP tok/s 87.7 TG tok/s 3304 TTFT (ms) 36.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i8loe3o7",
   "source_url": "https://omlx.ai/benchmarks/performance/i8loe3o7",
   "tg_tokens": null,
   "tps": 87.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ort1ps83-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 36.2 GB; TTFT 876 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4675.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k tokens 4,675 PP tok/s 91.4 TG tok/s 876 TTFT (ms) 36.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ort1ps83",
   "source_url": "https://omlx.ai/benchmarks/performance/ort1ps83",
   "tg_tokens": null,
   "tps": 91.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-hd2hjb9a-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1027.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 1,027 67.7 22.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hd2hjb9a",
   "source_url": "https://omlx.ai/benchmarks/performance/hd2hjb9a",
   "tg_tokens": null,
   "tps": 67.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-hd2hjb9a-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 878.1,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 878.1 59.5 23.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hd2hjb9a",
   "source_url": "https://omlx.ai/benchmarks/performance/hd2hjb9a",
   "tg_tokens": null,
   "tps": 59.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lbmh5ybx-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: lpp4rwhe; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3188.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 3,188 94.9 21.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lbmh5ybx",
   "source_url": "https://omlx.ai/benchmarks/performance/lbmh5ybx",
   "tg_tokens": null,
   "tps": 94.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lbmh5ybx-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: lpp4rwhe; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2904.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,904 88.1 21.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lbmh5ybx",
   "source_url": "https://omlx.ai/benchmarks/performance/lbmh5ybx",
   "tg_tokens": null,
   "tps": 88.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cocon1pw-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 1o87d9da, 5um5je3e, wqjtgpdr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 831.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 831.8 120.1 21.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cocon1pw",
   "source_url": "https://omlx.ai/benchmarks/performance/cocon1pw",
   "tg_tokens": null,
   "tps": 120.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cocon1pw-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 1o87d9da, 5um5je3e, wqjtgpdr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3018.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 3,018 144.0 22.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cocon1pw",
   "source_url": "https://omlx.ai/benchmarks/performance/cocon1pw",
   "tg_tokens": null,
   "tps": 144.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cocon1pw-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 1o87d9da, 5um5je3e, wqjtgpdr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5988.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 5,988 47.4 22.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cocon1pw",
   "source_url": "https://omlx.ai/benchmarks/performance/cocon1pw",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cocon1pw-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 1o87d9da, 5um5je3e, wqjtgpdr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5530.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 5,530 176.4 22.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cocon1pw",
   "source_url": "https://omlx.ai/benchmarks/performance/cocon1pw",
   "tg_tokens": null,
   "tps": 176.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cocon1pw-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 1o87d9da, 5um5je3e, wqjtgpdr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4485.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 4,485 221.0 22.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cocon1pw",
   "source_url": "https://omlx.ai/benchmarks/performance/cocon1pw",
   "tg_tokens": null,
   "tps": 221.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-cocon1pw-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 1o87d9da, 5um5je3e, wqjtgpdr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3333.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 3,333 142.7 23.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cocon1pw",
   "source_url": "https://omlx.ai/benchmarks/performance/cocon1pw",
   "tg_tokens": null,
   "tps": 142.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-gjqyada2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 135.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 135.2 15.2 17.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gjqyada2",
   "source_url": "https://omlx.ai/benchmarks/performance/gjqyada2",
   "tg_tokens": null,
   "tps": 15.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-gjqyada2-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 129.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 129.2 14.0 19.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gjqyada2",
   "source_url": "https://omlx.ai/benchmarks/performance/gjqyada2",
   "tg_tokens": null,
   "tps": 14.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5l0u0s5g-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 1oxrncmf, fu2fk4sa, igrt6lcf, x8myzysw, zawmq8kw; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 767.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 767.5 26.2 16.0 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5l0u0s5g",
   "source_url": "https://omlx.ai/benchmarks/performance/5l0u0s5g",
   "tg_tokens": null,
   "tps": 26.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5l0u0s5g-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 1oxrncmf, fu2fk4sa, igrt6lcf, x8myzysw, zawmq8kw; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 822.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 822.2 25.6 17.0 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5l0u0s5g",
   "source_url": "https://omlx.ai/benchmarks/performance/5l0u0s5g",
   "tg_tokens": null,
   "tps": 25.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5l0u0s5g-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 1oxrncmf, fu2fk4sa, igrt6lcf, x8myzysw, zawmq8kw; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 824.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 824.3 25.2 17.3 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5l0u0s5g",
   "source_url": "https://omlx.ai/benchmarks/performance/5l0u0s5g",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5l0u0s5g-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 1oxrncmf, fu2fk4sa, igrt6lcf, x8myzysw, zawmq8kw; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 800.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 800.3 24.0 17.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5l0u0s5g",
   "source_url": "https://omlx.ai/benchmarks/performance/5l0u0s5g",
   "tg_tokens": null,
   "tps": 24.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5l0u0s5g-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 1oxrncmf, fu2fk4sa, igrt6lcf, x8myzysw, zawmq8kw; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 748.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 32k 748.0 22.0 18.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5l0u0s5g",
   "source_url": "https://omlx.ai/benchmarks/performance/5l0u0s5g",
   "tg_tokens": null,
   "tps": 22.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5l0u0s5g-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 1oxrncmf, fu2fk4sa, igrt6lcf, x8myzysw, zawmq8kw; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 662.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 64k 662.5 20.5 20.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5l0u0s5g",
   "source_url": "https://omlx.ai/benchmarks/performance/5l0u0s5g",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-tnkkmlqe-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 8bit; 1k 224.1 21.1 28.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tnkkmlqe",
   "source_url": "https://omlx.ai/benchmarks/performance/tnkkmlqe",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-tnkkmlqe-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.9 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 8bit; 4k 224.7 20.5 31.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tnkkmlqe",
   "source_url": "https://omlx.ai/benchmarks/performance/tnkkmlqe",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-8lnwhiqs-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 39.9 GB at this context; duplicate traces of the same session: vnoesubd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 248.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 248.6 32.1 39.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8lnwhiqs",
   "source_url": "https://omlx.ai/benchmarks/performance/8lnwhiqs",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-8lnwhiqs-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 43.4 GB at this context; duplicate traces of the same session: vnoesubd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 320.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 320.5 28.7 43.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8lnwhiqs",
   "source_url": "https://omlx.ai/benchmarks/performance/8lnwhiqs",
   "tg_tokens": null,
   "tps": 28.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-yqe65seb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 39.9 GB at this context; duplicate traces of the same session: 6ysw2xye; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 246.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 246.3 41.7 39.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yqe65seb",
   "source_url": "https://omlx.ai/benchmarks/performance/yqe65seb",
   "tg_tokens": null,
   "tps": 41.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-yqe65seb-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 43.4 GB at this context; duplicate traces of the same session: 6ysw2xye; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 324.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 324.4 35.9 43.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yqe65seb",
   "source_url": "https://omlx.ai/benchmarks/performance/yqe65seb",
   "tg_tokens": null,
   "tps": 35.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-cevaxcqb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 39.9 GB at this context; duplicate traces of the same session: 643xxb12; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 243.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 243.7 26.7 39.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cevaxcqb",
   "source_url": "https://omlx.ai/benchmarks/performance/cevaxcqb",
   "tg_tokens": null,
   "tps": 26.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-cevaxcqb-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 43.3 GB at this context; duplicate traces of the same session: 643xxb12; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 318.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 318.8 24.2 43.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cevaxcqb",
   "source_url": "https://omlx.ai/benchmarks/performance/cevaxcqb",
   "tg_tokens": null,
   "tps": 24.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-v8awm5uk-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 32zcmxma, b5nmsc9l, mev7hi8q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 821.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 821.6 41.4 17.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v8awm5uk",
   "source_url": "https://omlx.ai/benchmarks/performance/v8awm5uk",
   "tg_tokens": null,
   "tps": 41.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-v8awm5uk-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 32zcmxma, b5nmsc9l, mev7hi8q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1304.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,304 53.1 19.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v8awm5uk",
   "source_url": "https://omlx.ai/benchmarks/performance/v8awm5uk",
   "tg_tokens": null,
   "tps": 53.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-v8awm5uk-ctx8192",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 32zcmxma, b5nmsc9l, mev7hi8q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1393.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,393 60.7 19.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v8awm5uk",
   "source_url": "https://omlx.ai/benchmarks/performance/v8awm5uk",
   "tg_tokens": null,
   "tps": 60.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-v8awm5uk-ctx16384",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 32zcmxma, b5nmsc9l, mev7hi8q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1300.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,300 77.9 19.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v8awm5uk",
   "source_url": "https://omlx.ai/benchmarks/performance/v8awm5uk",
   "tg_tokens": null,
   "tps": 77.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-v8awm5uk-ctx32768",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: 32zcmxma, b5nmsc9l, mev7hi8q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1146.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,146 91.9 21.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v8awm5uk",
   "source_url": "https://omlx.ai/benchmarks/performance/v8awm5uk",
   "tg_tokens": null,
   "tps": 91.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-v8awm5uk-ctx65536",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 32zcmxma, b5nmsc9l, mev7hi8q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 963.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 963.9 48.8 23.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v8awm5uk",
   "source_url": "https://omlx.ai/benchmarks/performance/v8awm5uk",
   "tg_tokens": null,
   "tps": 48.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-c5jr0o2s-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: 4xejwt5l, lpoajzbn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 762.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 762.2 40.1 33.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c5jr0o2s",
   "source_url": "https://omlx.ai/benchmarks/performance/c5jr0o2s",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-c5jr0o2s-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: 4xejwt5l, lpoajzbn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 908.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 908.2 48.3 34.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c5jr0o2s",
   "source_url": "https://omlx.ai/benchmarks/performance/c5jr0o2s",
   "tg_tokens": null,
   "tps": 48.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-c5jr0o2s-ctx8192",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: 4xejwt5l, lpoajzbn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 832.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 832.7 49.8 35.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c5jr0o2s",
   "source_url": "https://omlx.ai/benchmarks/performance/c5jr0o2s",
   "tg_tokens": null,
   "tps": 49.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-c5jr0o2s-ctx16384",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: 4xejwt5l, lpoajzbn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 799.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 799.6 55.8 35.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c5jr0o2s",
   "source_url": "https://omlx.ai/benchmarks/performance/c5jr0o2s",
   "tg_tokens": null,
   "tps": 55.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-c5jr0o2s-ctx32768",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: 4xejwt5l, lpoajzbn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 749.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 749.1 82.9 36.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c5jr0o2s",
   "source_url": "https://omlx.ai/benchmarks/performance/c5jr0o2s",
   "tg_tokens": null,
   "tps": 82.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-c5jr0o2s-ctx65536",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: 4xejwt5l, lpoajzbn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 665.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 665.6 46.6 39.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c5jr0o2s",
   "source_url": "https://omlx.ai/benchmarks/performance/c5jr0o2s",
   "tg_tokens": null,
   "tps": 46.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ch32n90g-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 8066wjjq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 429.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 429.0 40.4 18.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ch32n90g",
   "source_url": "https://omlx.ai/benchmarks/performance/ch32n90g",
   "tg_tokens": null,
   "tps": 40.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ch32n90g-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: 8066wjjq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 438.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 438.8 67.4 22.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ch32n90g",
   "source_url": "https://omlx.ai/benchmarks/performance/ch32n90g",
   "tg_tokens": null,
   "tps": 67.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-lbjnglpj-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: xwq6tmwi; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 428.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 428.4 79.1 18.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lbjnglpj",
   "source_url": "https://omlx.ai/benchmarks/performance/lbjnglpj",
   "tg_tokens": null,
   "tps": 79.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lbjnglpj-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: xwq6tmwi; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 438.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 438.7 72.6 22.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lbjnglpj",
   "source_url": "https://omlx.ai/benchmarks/performance/lbjnglpj",
   "tg_tokens": null,
   "tps": 72.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-4c8m0c0p-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 40.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 249.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 249.2 30.4 40.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4c8m0c0p",
   "source_url": "https://omlx.ai/benchmarks/performance/4c8m0c0p",
   "tg_tokens": null,
   "tps": 30.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-4c8m0c0p-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 327.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 327.4 27.4 43.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4c8m0c0p",
   "source_url": "https://omlx.ai/benchmarks/performance/4c8m0c0p",
   "tg_tokens": null,
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8w5usze5-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1768.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,768 72.9 71.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8w5usze5",
   "source_url": "https://omlx.ai/benchmarks/performance/8w5usze5",
   "tg_tokens": null,
   "tps": 72.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8w5usze5-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2508.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,508 83.4 72.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8w5usze5",
   "source_url": "https://omlx.ai/benchmarks/performance/8w5usze5",
   "tg_tokens": null,
   "tps": 83.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-llsr8qjq-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 144.8 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3918.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 8k 3,918 140.1 144.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/llsr8qjq",
   "source_url": "https://omlx.ai/benchmarks/performance/llsr8qjq",
   "tg_tokens": null,
   "tps": 140.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-llsr8qjq-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 146.8 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3881.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 64k 3,881 114.0 146.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/llsr8qjq",
   "source_url": "https://omlx.ai/benchmarks/performance/llsr8qjq",
   "tg_tokens": null,
   "tps": 114.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-wugry8vn-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: e37y8l1l, u3a5cp9s; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 417.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 417.3 43.6 21.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wugry8vn",
   "source_url": "https://omlx.ai/benchmarks/performance/wugry8vn",
   "tg_tokens": null,
   "tps": 43.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-wugry8vn-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: e37y8l1l, u3a5cp9s; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 405.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 405.3 46.3 22.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wugry8vn",
   "source_url": "https://omlx.ai/benchmarks/performance/wugry8vn",
   "tg_tokens": null,
   "tps": 46.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-wugry8vn-ctx8192",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: e37y8l1l, u3a5cp9s; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 377.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 8k 377.1 40.0 22.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wugry8vn",
   "source_url": "https://omlx.ai/benchmarks/performance/wugry8vn",
   "tg_tokens": null,
   "tps": 40.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-wugry8vn-ctx16384",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: e37y8l1l, u3a5cp9s; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 321.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k 321.1 34.2 22.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wugry8vn",
   "source_url": "https://omlx.ai/benchmarks/performance/wugry8vn",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-eeot0vx3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 941.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mlx-community--Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 941.6 18.7 29.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eeot0vx3",
   "source_url": "https://omlx.ai/benchmarks/performance/eeot0vx3",
   "tg_tokens": null,
   "tps": 18.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-eeot0vx3-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 729.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mlx-community--Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 729.9 18.0 30.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eeot0vx3",
   "source_url": "https://omlx.ai/benchmarks/performance/eeot0vx3",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-eeot0vx3-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 33.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 582.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mlx-community--Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 582.7 15.5 33.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eeot0vx3",
   "source_url": "https://omlx.ai/benchmarks/performance/eeot0vx3",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-jb7ok8dk-ctx199680",
   "model": "GLM-5.3",
   "notes": "oMLX community benchmark trace, single run at 195k context; peak memory 177.5 GB; TTFT 108883 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1837.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "omlx-glm53-oq4e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k tokens 1,837 PP tok/s 50.5 TG tok/s 108883 TTFT (ms) 177.5 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jb7ok8dk",
   "source_url": "https://omlx.ai/benchmarks/performance/jb7ok8dk",
   "tg_tokens": null,
   "tps": 50.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-goehrfv7-ctx1024",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 2iynds49, 5ndc9lgm, adyhuux6; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 81.9,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 5bit; 1k 81.9 10.9 19.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/goehrfv7",
   "source_url": "https://omlx.ai/benchmarks/performance/goehrfv7",
   "tg_tokens": null,
   "tps": 10.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-goehrfv7-ctx4096",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: 2iynds49, 5ndc9lgm, adyhuux6; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 76.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 5bit; 4k 76.0 10.1 20.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/goehrfv7",
   "source_url": "https://omlx.ai/benchmarks/performance/goehrfv7",
   "tg_tokens": null,
   "tps": 10.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-goehrfv7-ctx8192",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: 2iynds49, 5ndc9lgm, adyhuux6; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 74.3,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 5bit; 8k 74.3 9.9 21.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/goehrfv7",
   "source_url": "https://omlx.ai/benchmarks/performance/goehrfv7",
   "tg_tokens": null,
   "tps": 9.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-goehrfv7-ctx16384",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 2iynds49, 5ndc9lgm, adyhuux6; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 70.4,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 5bit; 16k 70.4 9.1 21.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/goehrfv7",
   "source_url": "https://omlx.ai/benchmarks/performance/goehrfv7",
   "tg_tokens": null,
   "tps": 9.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-62juy64w-ctx1024",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 5ulvaqlf, pqg6m2tk; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 404.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 404.0 50.0 20.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/62juy64w",
   "source_url": "https://omlx.ai/benchmarks/performance/62juy64w",
   "tg_tokens": null,
   "tps": 50.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-62juy64w-ctx4096",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 5ulvaqlf, pqg6m2tk; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 415.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 415.1 46.6 20.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/62juy64w",
   "source_url": "https://omlx.ai/benchmarks/performance/62juy64w",
   "tg_tokens": null,
   "tps": 46.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-62juy64w-ctx8192",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: 5ulvaqlf, pqg6m2tk; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 380.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 8k 380.6 46.7 21.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/62juy64w",
   "source_url": "https://omlx.ai/benchmarks/performance/62juy64w",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-idfaq115-ctx1024",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: 2em8a3n1, lzdrrs9h; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 143.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 143.0 13.0 18.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/idfaq115",
   "source_url": "https://omlx.ai/benchmarks/performance/idfaq115",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-idfaq115-ctx4096",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 2em8a3n1, lzdrrs9h; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 348.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 348.8 12.3 19.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/idfaq115",
   "source_url": "https://omlx.ai/benchmarks/performance/idfaq115",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-idfaq115-ctx8192",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: 2em8a3n1, lzdrrs9h; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 330.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 8k 330.0 30.4 19.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/idfaq115",
   "source_url": "https://omlx.ai/benchmarks/performance/idfaq115",
   "tg_tokens": null,
   "tps": 30.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-0wu610he-ctx1024",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: r3j5yvwt; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 132.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 132.4 18.8 18.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0wu610he",
   "source_url": "https://omlx.ai/benchmarks/performance/0wu610he",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-0wu610he-ctx4096",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: r3j5yvwt; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 348.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 348.3 28.8 19.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0wu610he",
   "source_url": "https://omlx.ai/benchmarks/performance/0wu610he",
   "tg_tokens": null,
   "tps": 28.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-fkkwq6qb-ctx1024",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: n2989mma; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 151.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 151.9 16.7 13.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fkkwq6qb",
   "source_url": "https://omlx.ai/benchmarks/performance/fkkwq6qb",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-fkkwq6qb-ctx4096",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: n2989mma; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 287.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 287.6 18.5 14.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fkkwq6qb",
   "source_url": "https://omlx.ai/benchmarks/performance/fkkwq6qb",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-w4v8l1wv-ctx1024",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 87.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 272.1,
   "provenance": "community",
   "quant": "2.4bit mixed",
   "quote": "DeepSeek-V4-Flash-0731-2.4bit-mixed on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 2bit; 1k 272.1 29.3 87.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w4v8l1wv",
   "source_url": "https://omlx.ai/benchmarks/performance/w4v8l1wv",
   "tg_tokens": null,
   "tps": 29.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-w4v8l1wv-ctx4096",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 88.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 273.2,
   "provenance": "community",
   "quant": "2.4bit mixed",
   "quote": "DeepSeek-V4-Flash-0731-2.4bit-mixed on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 2bit; 4k 273.2 27.7 88.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w4v8l1wv",
   "source_url": "https://omlx.ai/benchmarks/performance/w4v8l1wv",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jhs9inq0-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 316.2 GB at this context; duplicate traces of the same session: 92q0notv, cn175l22; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 430.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "GLM-5.3-Flash-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 430.3 32.4 316.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jhs9inq0",
   "source_url": "https://omlx.ai/benchmarks/performance/jhs9inq0",
   "tg_tokens": null,
   "tps": 32.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jhs9inq0-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 317.3 GB at this context; duplicate traces of the same session: 92q0notv, cn175l22; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 551.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "GLM-5.3-Flash-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 551.2 31.1 317.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jhs9inq0",
   "source_url": "https://omlx.ai/benchmarks/performance/jhs9inq0",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jhs9inq0-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 317.6 GB at this context; duplicate traces of the same session: 92q0notv, cn175l22; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 513.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "GLM-5.3-Flash-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 513.3 31.2 317.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jhs9inq0",
   "source_url": "https://omlx.ai/benchmarks/performance/jhs9inq0",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jhs9inq0-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 317.7 GB at this context; duplicate traces of the same session: 92q0notv, cn175l22; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 545.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "GLM-5.3-Flash-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k 545.4 30.9 317.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jhs9inq0",
   "source_url": "https://omlx.ai/benchmarks/performance/jhs9inq0",
   "tg_tokens": null,
   "tps": 30.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-8lztys7d-ctx1024",
   "model": "Ling-3.0-flash",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 68.3 GB at this context; duplicate traces of the same session: 77fw40bb, i4413e2f, pqds84a8; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 594.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-flash-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 594.8 41.4 68.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8lztys7d",
   "source_url": "https://omlx.ai/benchmarks/performance/8lztys7d",
   "tg_tokens": null,
   "tps": 41.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-8lztys7d-ctx4096",
   "model": "Ling-3.0-flash",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 69.5 GB at this context; duplicate traces of the same session: 77fw40bb, i4413e2f, pqds84a8; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 525.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-flash-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 525.3 37.9 69.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8lztys7d",
   "source_url": "https://omlx.ai/benchmarks/performance/8lztys7d",
   "tg_tokens": null,
   "tps": 37.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-8lztys7d-ctx8192",
   "model": "Ling-3.0-flash",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 70.2 GB at this context; duplicate traces of the same session: 77fw40bb, i4413e2f, pqds84a8; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 505.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-flash-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 505.0 37.0 70.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8lztys7d",
   "source_url": "https://omlx.ai/benchmarks/performance/8lztys7d",
   "tg_tokens": null,
   "tps": 37.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-8lztys7d-ctx16384",
   "model": "Ling-3.0-flash",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: 77fw40bb, i4413e2f, pqds84a8; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 476.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-flash-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 476.4 33.0 71.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8lztys7d",
   "source_url": "https://omlx.ai/benchmarks/performance/8lztys7d",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-8lztys7d-ctx32768",
   "model": "Ling-3.0-flash",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 74.7 GB at this context; duplicate traces of the same session: 77fw40bb, i4413e2f, pqds84a8; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 427.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-flash-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 32k 427.5 32.6 74.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8lztys7d",
   "source_url": "https://omlx.ai/benchmarks/performance/8lztys7d",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-dioxvh5t-ctx1024",
   "model": "llama-3-1-8b-instruct",
   "notes": "oMLX community benchmark trace; peak memory 15.5 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 405.5,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Llama-3.1-8B-Instruct on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 unknown; 1k 405.5 16.2 15.5 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dioxvh5t",
   "source_url": "https://omlx.ai/benchmarks/performance/dioxvh5t",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-dioxvh5t-ctx4096",
   "model": "llama-3-1-8b-instruct",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 399.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Llama-3.1-8B-Instruct on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 unknown; 4k 399.0 15.8 16.0 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dioxvh5t",
   "source_url": "https://omlx.ai/benchmarks/performance/dioxvh5t",
   "tg_tokens": null,
   "tps": 15.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-p8nn7fn3-ctx1024",
   "model": "Occamy-1.0",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: 33wwatx3, 4561q73z, ey6akcji; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1386.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Occamy-1.0-oQ8-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,386 68.9 35.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p8nn7fn3",
   "source_url": "https://omlx.ai/benchmarks/performance/p8nn7fn3",
   "tg_tokens": null,
   "tps": 68.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-p8nn7fn3-ctx4096",
   "model": "Occamy-1.0",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: 33wwatx3, 4561q73z, ey6akcji; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1296.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Occamy-1.0-oQ8-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 1,296 57.3 37.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p8nn7fn3",
   "source_url": "https://omlx.ai/benchmarks/performance/p8nn7fn3",
   "tg_tokens": null,
   "tps": 57.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-p8nn7fn3-ctx8192",
   "model": "Occamy-1.0",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 37.9 GB at this context; duplicate traces of the same session: 33wwatx3, 4561q73z, ey6akcji; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1026.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Occamy-1.0-oQ8-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 8k 1,026 58.8 37.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p8nn7fn3",
   "source_url": "https://omlx.ai/benchmarks/performance/p8nn7fn3",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-p8nn7fn3-ctx16384",
   "model": "Occamy-1.0",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 38.7 GB at this context; duplicate traces of the same session: 33wwatx3, 4561q73z, ey6akcji; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 960.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Occamy-1.0-oQ8-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 16k 960.5 50.3 38.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p8nn7fn3",
   "source_url": "https://omlx.ai/benchmarks/performance/p8nn7fn3",
   "tg_tokens": null,
   "tps": 50.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-p8nn7fn3-ctx32768",
   "model": "Occamy-1.0",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 40.1 GB at this context; duplicate traces of the same session: 33wwatx3, 4561q73z, ey6akcji; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 792.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Occamy-1.0-oQ8-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 32k 792.0 49.6 40.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p8nn7fn3",
   "source_url": "https://omlx.ai/benchmarks/performance/p8nn7fn3",
   "tg_tokens": null,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-p8nn7fn3-ctx65536",
   "model": "Occamy-1.0",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 41.2 GB at this context; duplicate traces of the same session: 33wwatx3, 4561q73z, ey6akcji; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 590.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Occamy-1.0-oQ8-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 64k 590.2 39.0 41.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p8nn7fn3",
   "source_url": "https://omlx.ai/benchmarks/performance/p8nn7fn3",
   "tg_tokens": null,
   "tps": 39.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-vn0l7rh5-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: sez6hion; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 430.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 430.3 58.3 19.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vn0l7rh5",
   "source_url": "https://omlx.ai/benchmarks/performance/vn0l7rh5",
   "tg_tokens": null,
   "tps": 58.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-vn0l7rh5-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: sez6hion; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 409.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 409.6 56.7 20.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vn0l7rh5",
   "source_url": "https://omlx.ai/benchmarks/performance/vn0l7rh5",
   "tg_tokens": null,
   "tps": 56.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-l73ctic8-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 12.3 GB at this context; duplicate traces of the same session: 8r0n8dn5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 160.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 160.4 21.5 12.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l73ctic8",
   "source_url": "https://omlx.ai/benchmarks/performance/l73ctic8",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-l73ctic8-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 13.0 GB at this context; duplicate traces of the same session: 8r0n8dn5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 275.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 275.7 18.0 13.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l73ctic8",
   "source_url": "https://omlx.ai/benchmarks/performance/l73ctic8",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-ljb38wu3-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: y2mnfzw4; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 138.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 138.5 27.1 17.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ljb38wu3",
   "source_url": "https://omlx.ai/benchmarks/performance/ljb38wu3",
   "tg_tokens": null,
   "tps": 27.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-ljb38wu3-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 18.0 GB at this context; duplicate traces of the same session: y2mnfzw4; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 333.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 333.8 25.0 18.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ljb38wu3",
   "source_url": "https://omlx.ai/benchmarks/performance/ljb38wu3",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-xfpwttad-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 6wayt66a, fvoam9am; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 919.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 919.0 90.1 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xfpwttad",
   "source_url": "https://omlx.ai/benchmarks/performance/xfpwttad",
   "tg_tokens": null,
   "tps": 90.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-xfpwttad-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 6wayt66a, fvoam9am; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1082.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,082 85.8 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xfpwttad",
   "source_url": "https://omlx.ai/benchmarks/performance/xfpwttad",
   "tg_tokens": null,
   "tps": 85.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-xfpwttad-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 6wayt66a, fvoam9am; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 998.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 998.5 80.2 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xfpwttad",
   "source_url": "https://omlx.ai/benchmarks/performance/xfpwttad",
   "tg_tokens": null,
   "tps": 80.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-xfpwttad-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 6wayt66a, fvoam9am; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 835.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 835.4 73.0 21.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xfpwttad",
   "source_url": "https://omlx.ai/benchmarks/performance/xfpwttad",
   "tg_tokens": null,
   "tps": 73.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-zb1gaw3m-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 1dvrgy83, 61ppps46; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 931.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 931.6 88.4 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zb1gaw3m",
   "source_url": "https://omlx.ai/benchmarks/performance/zb1gaw3m",
   "tg_tokens": null,
   "tps": 88.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-zb1gaw3m-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 1dvrgy83, 61ppps46; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1069.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,069 88.5 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zb1gaw3m",
   "source_url": "https://omlx.ai/benchmarks/performance/zb1gaw3m",
   "tg_tokens": null,
   "tps": 88.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-zb1gaw3m-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 1dvrgy83, 61ppps46; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1014.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 1,014 82.5 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zb1gaw3m",
   "source_url": "https://omlx.ai/benchmarks/performance/zb1gaw3m",
   "tg_tokens": null,
   "tps": 82.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-zb1gaw3m-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 1dvrgy83, 61ppps46; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 869.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 869.8 74.5 22.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zb1gaw3m",
   "source_url": "https://omlx.ai/benchmarks/performance/zb1gaw3m",
   "tg_tokens": null,
   "tps": 74.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-236zsp9a-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 9tpknegx, ff9i7nzg, u5svoupk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 807.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 807.7 85.6 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/236zsp9a",
   "source_url": "https://omlx.ai/benchmarks/performance/236zsp9a",
   "tg_tokens": null,
   "tps": 85.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-236zsp9a-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 9tpknegx, ff9i7nzg, u5svoupk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1012.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,012 82.1 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/236zsp9a",
   "source_url": "https://omlx.ai/benchmarks/performance/236zsp9a",
   "tg_tokens": null,
   "tps": 82.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-236zsp9a-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 9tpknegx, ff9i7nzg, u5svoupk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 961.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 961.9 78.9 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/236zsp9a",
   "source_url": "https://omlx.ai/benchmarks/performance/236zsp9a",
   "tg_tokens": null,
   "tps": 78.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-236zsp9a-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 9tpknegx, ff9i7nzg, u5svoupk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 806.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 806.6 70.8 22.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/236zsp9a",
   "source_url": "https://omlx.ai/benchmarks/performance/236zsp9a",
   "tg_tokens": null,
   "tps": 70.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-f4bo9r0d-ctx8192",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 948.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 948.9 84.9 7.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f4bo9r0d",
   "source_url": "https://omlx.ai/benchmarks/performance/f4bo9r0d",
   "tg_tokens": null,
   "tps": 84.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-f4bo9r0d-ctx65536",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 731.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 731.5 61.6 10.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f4bo9r0d",
   "source_url": "https://omlx.ai/benchmarks/performance/f4bo9r0d",
   "tg_tokens": null,
   "tps": 61.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-f4bo9r0d-ctx131072",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 12.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 576.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 576.6 46.7 12.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f4bo9r0d",
   "source_url": "https://omlx.ai/benchmarks/performance/f4bo9r0d",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-l51daj83-ctx65536",
   "model": "Qwen2.5-Coder-7B-Instruct",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 11.7 GB; TTFT 396085 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 165.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen2.5-Coder-7B-Instruct-8bit on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 8bit; 64k tokens 165.5 PP tok/s 12.3 TG tok/s 396085 TTFT (ms) 11.7 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l51daj83",
   "source_url": "https://omlx.ai/benchmarks/performance/l51daj83",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-s3ry01io-ctx4096",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 2.2 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1929.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B-8bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 8bit; 4k 1,929 136.6 2.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s3ry01io",
   "source_url": "https://omlx.ai/benchmarks/performance/s3ry01io",
   "tg_tokens": null,
   "tps": 136.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-s3ry01io-ctx16384",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 2.3 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1416.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B-8bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 8bit; 16k 1,416 118.3 2.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s3ry01io",
   "source_url": "https://omlx.ai/benchmarks/performance/s3ry01io",
   "tg_tokens": null,
   "tps": 118.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-ty1cu4pv-ctx1024",
   "model": "Qwen3.5-9B-Distilled-OPUS-Heretic-MLX-VLM",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 572.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-Distilled-OPUS-Heretic-MLX-VLM-4bit on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 1k 572.7 23.0 5.8 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ty1cu4pv",
   "source_url": "https://omlx.ai/benchmarks/performance/ty1cu4pv",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-ty1cu4pv-ctx4096",
   "model": "Qwen3.5-9B-Distilled-OPUS-Heretic-MLX-VLM",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 603.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-Distilled-OPUS-Heretic-MLX-VLM-4bit on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 4k 603.9 22.5 6.2 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ty1cu4pv",
   "source_url": "https://omlx.ai/benchmarks/performance/ty1cu4pv",
   "tg_tokens": null,
   "tps": 22.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-lw43b1wp-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: none; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 165.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (8c); M4 (8c) \u00b7 16 GB \u00b7 4bit; 1k 165.4 19.2 5.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lw43b1wp",
   "source_url": "https://omlx.ai/benchmarks/performance/lw43b1wp",
   "tg_tokens": null,
   "tps": 19.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-lw43b1wp-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: none; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 167.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (8c); M4 (8c) \u00b7 16 GB \u00b7 4bit; 4k 167.8 16.7 6.2 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lw43b1wp",
   "source_url": "https://omlx.ai/benchmarks/performance/lw43b1wp",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lpp4rwhe-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3188.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 3,188 94.9 21.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lpp4rwhe",
   "source_url": "https://omlx.ai/benchmarks/performance/lpp4rwhe",
   "tg_tokens": null,
   "tps": 94.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lpp4rwhe-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2904.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,904 88.1 21.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lpp4rwhe",
   "source_url": "https://omlx.ai/benchmarks/performance/lpp4rwhe",
   "tg_tokens": null,
   "tps": 88.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5um5je3e-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 1o87d9da, wqjtgpdr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 831.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 831.8 120.1 21.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5um5je3e",
   "source_url": "https://omlx.ai/benchmarks/performance/5um5je3e",
   "tg_tokens": null,
   "tps": 120.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5um5je3e-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 1o87d9da, wqjtgpdr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3018.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 3,018 144.0 22.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5um5je3e",
   "source_url": "https://omlx.ai/benchmarks/performance/5um5je3e",
   "tg_tokens": null,
   "tps": 144.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5um5je3e-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 1o87d9da, wqjtgpdr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5988.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 5,988 47.4 22.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5um5je3e",
   "source_url": "https://omlx.ai/benchmarks/performance/5um5je3e",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5um5je3e-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 1o87d9da, wqjtgpdr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5530.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 5,530 176.4 22.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5um5je3e",
   "source_url": "https://omlx.ai/benchmarks/performance/5um5je3e",
   "tg_tokens": null,
   "tps": 176.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5um5je3e-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 1o87d9da, wqjtgpdr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4485.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 4,485 221.0 22.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5um5je3e",
   "source_url": "https://omlx.ai/benchmarks/performance/5um5je3e",
   "tg_tokens": null,
   "tps": 221.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5um5je3e-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 1o87d9da, wqjtgpdr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3333.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 3,333 142.7 23.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5um5je3e",
   "source_url": "https://omlx.ai/benchmarks/performance/5um5je3e",
   "tg_tokens": null,
   "tps": 142.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-q1iyeqd3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: oflhwc4r; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 87.2,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 1k 87.2 13.3 22.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q1iyeqd3",
   "source_url": "https://omlx.ai/benchmarks/performance/q1iyeqd3",
   "tg_tokens": null,
   "tps": 13.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-q1iyeqd3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: oflhwc4r; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 92.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 4k 92.6 13.0 23.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q1iyeqd3",
   "source_url": "https://omlx.ai/benchmarks/performance/q1iyeqd3",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-a33gakgw-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 18.4 GB; TTFT 18819 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 217.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k tokens 217.7 PP tok/s 15.9 TG tok/s 18819 TTFT (ms) 18.4 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a33gakgw",
   "source_url": "https://omlx.ai/benchmarks/performance/a33gakgw",
   "tg_tokens": null,
   "tps": 15.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fu2fk4sa-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 1oxrncmf, igrt6lcf, x8myzysw, zawmq8kw; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 767.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 767.5 26.2 16.0 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fu2fk4sa",
   "source_url": "https://omlx.ai/benchmarks/performance/fu2fk4sa",
   "tg_tokens": null,
   "tps": 26.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fu2fk4sa-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 1oxrncmf, igrt6lcf, x8myzysw, zawmq8kw; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 822.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 822.2 25.6 17.0 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fu2fk4sa",
   "source_url": "https://omlx.ai/benchmarks/performance/fu2fk4sa",
   "tg_tokens": null,
   "tps": 25.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fu2fk4sa-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 1oxrncmf, igrt6lcf, x8myzysw, zawmq8kw; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 824.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 824.3 25.2 17.3 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fu2fk4sa",
   "source_url": "https://omlx.ai/benchmarks/performance/fu2fk4sa",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fu2fk4sa-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 1oxrncmf, igrt6lcf, x8myzysw, zawmq8kw; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 800.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 800.3 24.0 17.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fu2fk4sa",
   "source_url": "https://omlx.ai/benchmarks/performance/fu2fk4sa",
   "tg_tokens": null,
   "tps": 24.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fu2fk4sa-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 1oxrncmf, igrt6lcf, x8myzysw, zawmq8kw; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 748.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 32k 748.0 22.0 18.8 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fu2fk4sa",
   "source_url": "https://omlx.ai/benchmarks/performance/fu2fk4sa",
   "tg_tokens": null,
   "tps": 22.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fu2fk4sa-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 1oxrncmf, igrt6lcf, x8myzysw, zawmq8kw; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 662.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 64k 662.5 20.5 20.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fu2fk4sa",
   "source_url": "https://omlx.ai/benchmarks/performance/fu2fk4sa",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-vnoesubd-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 39.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 248.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 248.6 32.1 39.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vnoesubd",
   "source_url": "https://omlx.ai/benchmarks/performance/vnoesubd",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-vnoesubd-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 43.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 320.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 320.5 28.7 43.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vnoesubd",
   "source_url": "https://omlx.ai/benchmarks/performance/vnoesubd",
   "tg_tokens": null,
   "tps": 28.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-6ysw2xye-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 39.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 246.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 246.3 41.7 39.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ysw2xye",
   "source_url": "https://omlx.ai/benchmarks/performance/6ysw2xye",
   "tg_tokens": null,
   "tps": 41.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-6ysw2xye-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 43.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 324.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 324.4 35.9 43.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ysw2xye",
   "source_url": "https://omlx.ai/benchmarks/performance/6ysw2xye",
   "tg_tokens": null,
   "tps": 35.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-643xxb12-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 39.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 243.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 243.7 26.7 39.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/643xxb12",
   "source_url": "https://omlx.ai/benchmarks/performance/643xxb12",
   "tg_tokens": null,
   "tps": 26.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-643xxb12-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 43.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 318.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 318.8 24.2 43.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/643xxb12",
   "source_url": "https://omlx.ai/benchmarks/performance/643xxb12",
   "tg_tokens": null,
   "tps": 24.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-b5nmsc9l-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 32zcmxma, mev7hi8q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 821.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 821.6 41.4 17.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b5nmsc9l",
   "source_url": "https://omlx.ai/benchmarks/performance/b5nmsc9l",
   "tg_tokens": null,
   "tps": 41.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-b5nmsc9l-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 32zcmxma, mev7hi8q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1304.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,304 53.1 19.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b5nmsc9l",
   "source_url": "https://omlx.ai/benchmarks/performance/b5nmsc9l",
   "tg_tokens": null,
   "tps": 53.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-b5nmsc9l-ctx8192",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 32zcmxma, mev7hi8q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1393.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,393 60.7 19.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b5nmsc9l",
   "source_url": "https://omlx.ai/benchmarks/performance/b5nmsc9l",
   "tg_tokens": null,
   "tps": 60.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-b5nmsc9l-ctx16384",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 32zcmxma, mev7hi8q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1300.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,300 77.9 19.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b5nmsc9l",
   "source_url": "https://omlx.ai/benchmarks/performance/b5nmsc9l",
   "tg_tokens": null,
   "tps": 77.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-b5nmsc9l-ctx32768",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: 32zcmxma, mev7hi8q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1146.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,146 91.9 21.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b5nmsc9l",
   "source_url": "https://omlx.ai/benchmarks/performance/b5nmsc9l",
   "tg_tokens": null,
   "tps": 91.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-b5nmsc9l-ctx65536",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 32zcmxma, mev7hi8q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 963.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 963.9 48.8 23.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b5nmsc9l",
   "source_url": "https://omlx.ai/benchmarks/performance/b5nmsc9l",
   "tg_tokens": null,
   "tps": 48.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lpoajzbn-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: 4xejwt5l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 762.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 762.2 40.1 33.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lpoajzbn",
   "source_url": "https://omlx.ai/benchmarks/performance/lpoajzbn",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lpoajzbn-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: 4xejwt5l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 908.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 908.2 48.3 34.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lpoajzbn",
   "source_url": "https://omlx.ai/benchmarks/performance/lpoajzbn",
   "tg_tokens": null,
   "tps": 48.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lpoajzbn-ctx8192",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: 4xejwt5l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 832.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 832.7 49.8 35.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lpoajzbn",
   "source_url": "https://omlx.ai/benchmarks/performance/lpoajzbn",
   "tg_tokens": null,
   "tps": 49.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lpoajzbn-ctx16384",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: 4xejwt5l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 799.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 799.6 55.8 35.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lpoajzbn",
   "source_url": "https://omlx.ai/benchmarks/performance/lpoajzbn",
   "tg_tokens": null,
   "tps": 55.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lpoajzbn-ctx32768",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: 4xejwt5l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 749.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 749.1 82.9 36.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lpoajzbn",
   "source_url": "https://omlx.ai/benchmarks/performance/lpoajzbn",
   "tg_tokens": null,
   "tps": 82.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lpoajzbn-ctx65536",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: 4xejwt5l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 665.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 665.6 46.6 39.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lpoajzbn",
   "source_url": "https://omlx.ai/benchmarks/performance/lpoajzbn",
   "tg_tokens": null,
   "tps": 46.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-mnt793dr-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 25.3 GB; TTFT 109441 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1198.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k tokens 1,198 PP tok/s 34.1 TG tok/s 109441 TTFT (ms) 25.3 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mnt793dr",
   "source_url": "https://omlx.ai/benchmarks/performance/mnt793dr",
   "tg_tokens": null,
   "tps": 34.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-8066wjjq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 429.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 429.0 40.4 18.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8066wjjq",
   "source_url": "https://omlx.ai/benchmarks/performance/8066wjjq",
   "tg_tokens": null,
   "tps": 40.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-8066wjjq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 438.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 438.8 67.4 22.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8066wjjq",
   "source_url": "https://omlx.ai/benchmarks/performance/8066wjjq",
   "tg_tokens": null,
   "tps": 67.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xwq6tmwi-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 428.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 428.4 79.1 18.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xwq6tmwi",
   "source_url": "https://omlx.ai/benchmarks/performance/xwq6tmwi",
   "tg_tokens": null,
   "tps": 79.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xwq6tmwi-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 438.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 438.7 72.6 22.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xwq6tmwi",
   "source_url": "https://omlx.ai/benchmarks/performance/xwq6tmwi",
   "tg_tokens": null,
   "tps": 72.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-fwn0cmcy-ctx4096",
   "model": "Qwen3.8-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: rpskl4f5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 186.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-mlx-4Bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 186.9 30.7 6.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fwn0cmcy",
   "source_url": "https://omlx.ai/benchmarks/performance/fwn0cmcy",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-fwn0cmcy-ctx16384",
   "model": "Qwen3.8-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.1 GB at this context; duplicate traces of the same session: rpskl4f5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 162.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-mlx-4Bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k 162.0 27.6 7.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fwn0cmcy",
   "source_url": "https://omlx.ai/benchmarks/performance/fwn0cmcy",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-k4h8l6zi-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.4 GB at this context; duplicate traces of the same session: l8jvfkkb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 924.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 924.2 81.9 100.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k4h8l6zi",
   "source_url": "https://omlx.ai/benchmarks/performance/k4h8l6zi",
   "tg_tokens": null,
   "tps": 81.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-k4h8l6zi-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: l8jvfkkb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1133.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,133 70.1 101.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k4h8l6zi",
   "source_url": "https://omlx.ai/benchmarks/performance/k4h8l6zi",
   "tg_tokens": null,
   "tps": 70.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-u3a5cp9s-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: e37y8l1l; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 417.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 417.3 43.6 21.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u3a5cp9s",
   "source_url": "https://omlx.ai/benchmarks/performance/u3a5cp9s",
   "tg_tokens": null,
   "tps": 43.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-u3a5cp9s-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: e37y8l1l; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 405.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 405.3 46.3 22.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u3a5cp9s",
   "source_url": "https://omlx.ai/benchmarks/performance/u3a5cp9s",
   "tg_tokens": null,
   "tps": 46.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-u3a5cp9s-ctx8192",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: e37y8l1l; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 377.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 8k 377.1 40.0 22.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u3a5cp9s",
   "source_url": "https://omlx.ai/benchmarks/performance/u3a5cp9s",
   "tg_tokens": null,
   "tps": 40.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-u3a5cp9s-ctx16384",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: e37y8l1l; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 321.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k 321.1 34.2 22.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u3a5cp9s",
   "source_url": "https://omlx.ai/benchmarks/performance/u3a5cp9s",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-5ndc9lgm-ctx1024",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 2iynds49, adyhuux6; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 81.9,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 5bit; 1k 81.9 10.9 19.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ndc9lgm",
   "source_url": "https://omlx.ai/benchmarks/performance/5ndc9lgm",
   "tg_tokens": null,
   "tps": 10.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-5ndc9lgm-ctx4096",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: 2iynds49, adyhuux6; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 76.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 5bit; 4k 76.0 10.1 20.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ndc9lgm",
   "source_url": "https://omlx.ai/benchmarks/performance/5ndc9lgm",
   "tg_tokens": null,
   "tps": 10.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-5ndc9lgm-ctx8192",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: 2iynds49, adyhuux6; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 74.3,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 5bit; 8k 74.3 9.9 21.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ndc9lgm",
   "source_url": "https://omlx.ai/benchmarks/performance/5ndc9lgm",
   "tg_tokens": null,
   "tps": 9.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-5ndc9lgm-ctx16384",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 2iynds49, adyhuux6; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 70.4,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 5bit; 16k 70.4 9.1 21.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ndc9lgm",
   "source_url": "https://omlx.ai/benchmarks/performance/5ndc9lgm",
   "tg_tokens": null,
   "tps": 9.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-pqg6m2tk-ctx1024",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 5ulvaqlf; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 404.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 404.0 50.0 20.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pqg6m2tk",
   "source_url": "https://omlx.ai/benchmarks/performance/pqg6m2tk",
   "tg_tokens": null,
   "tps": 50.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-pqg6m2tk-ctx4096",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 5ulvaqlf; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 415.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 415.1 46.6 20.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pqg6m2tk",
   "source_url": "https://omlx.ai/benchmarks/performance/pqg6m2tk",
   "tg_tokens": null,
   "tps": 46.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-pqg6m2tk-ctx8192",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: 5ulvaqlf; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 380.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 8k 380.6 46.7 21.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pqg6m2tk",
   "source_url": "https://omlx.ai/benchmarks/performance/pqg6m2tk",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-2em8a3n1-ctx1024",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: lzdrrs9h; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 143.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 143.0 13.0 18.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2em8a3n1",
   "source_url": "https://omlx.ai/benchmarks/performance/2em8a3n1",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-2em8a3n1-ctx4096",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: lzdrrs9h; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 348.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 348.8 12.3 19.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2em8a3n1",
   "source_url": "https://omlx.ai/benchmarks/performance/2em8a3n1",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-2em8a3n1-ctx8192",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: lzdrrs9h; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 330.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 8k 330.0 30.4 19.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2em8a3n1",
   "source_url": "https://omlx.ai/benchmarks/performance/2em8a3n1",
   "tg_tokens": null,
   "tps": 30.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-r3j5yvwt-ctx1024",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 132.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 132.4 18.8 18.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r3j5yvwt",
   "source_url": "https://omlx.ai/benchmarks/performance/r3j5yvwt",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-r3j5yvwt-ctx4096",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 348.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 348.3 28.8 19.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r3j5yvwt",
   "source_url": "https://omlx.ai/benchmarks/performance/r3j5yvwt",
   "tg_tokens": null,
   "tps": 28.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-n2989mma-ctx1024",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 151.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 151.9 16.7 13.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n2989mma",
   "source_url": "https://omlx.ai/benchmarks/performance/n2989mma",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-n2989mma-ctx4096",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 287.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 287.6 18.5 14.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n2989mma",
   "source_url": "https://omlx.ai/benchmarks/performance/n2989mma",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-f1dsp2wa-ctx1024",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 87.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 272.1,
   "provenance": "community",
   "quant": "2.4bit mixed",
   "quote": "DeepSeek-V4-Flash-0731-2.4bit-mixed on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 2bit; 1k 272.1 29.3 87.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f1dsp2wa",
   "source_url": "https://omlx.ai/benchmarks/performance/f1dsp2wa",
   "tg_tokens": null,
   "tps": 29.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-f1dsp2wa-ctx4096",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 88.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 273.2,
   "provenance": "community",
   "quant": "2.4bit mixed",
   "quote": "DeepSeek-V4-Flash-0731-2.4bit-mixed on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 2bit; 4k 273.2 27.7 88.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f1dsp2wa",
   "source_url": "https://omlx.ai/benchmarks/performance/f1dsp2wa",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-92q0notv-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 316.2 GB at this context; duplicate traces of the same session: cn175l22, k51hgfi7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 430.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "GLM-5.3-Flash-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 430.3 32.4 316.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/92q0notv",
   "source_url": "https://omlx.ai/benchmarks/performance/92q0notv",
   "tg_tokens": null,
   "tps": 32.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-92q0notv-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 317.3 GB at this context; duplicate traces of the same session: cn175l22, k51hgfi7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 551.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "GLM-5.3-Flash-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 551.2 31.1 317.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/92q0notv",
   "source_url": "https://omlx.ai/benchmarks/performance/92q0notv",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-92q0notv-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 317.6 GB at this context; duplicate traces of the same session: cn175l22, k51hgfi7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 513.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "GLM-5.3-Flash-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 513.3 31.2 317.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/92q0notv",
   "source_url": "https://omlx.ai/benchmarks/performance/92q0notv",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-92q0notv-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 317.7 GB at this context; duplicate traces of the same session: cn175l22, k51hgfi7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 545.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "GLM-5.3-Flash-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k 545.4 30.9 317.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/92q0notv",
   "source_url": "https://omlx.ai/benchmarks/performance/92q0notv",
   "tg_tokens": null,
   "tps": 30.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-77fw40bb-ctx1024",
   "model": "Ling-3.0-flash",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 68.3 GB at this context; duplicate traces of the same session: i4413e2f, pqds84a8; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 594.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-flash-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 594.8 41.4 68.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/77fw40bb",
   "source_url": "https://omlx.ai/benchmarks/performance/77fw40bb",
   "tg_tokens": null,
   "tps": 41.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-77fw40bb-ctx4096",
   "model": "Ling-3.0-flash",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 69.5 GB at this context; duplicate traces of the same session: i4413e2f, pqds84a8; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 525.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-flash-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 525.3 37.9 69.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/77fw40bb",
   "source_url": "https://omlx.ai/benchmarks/performance/77fw40bb",
   "tg_tokens": null,
   "tps": 37.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-77fw40bb-ctx8192",
   "model": "Ling-3.0-flash",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 70.2 GB at this context; duplicate traces of the same session: i4413e2f, pqds84a8; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 505.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-flash-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 505.0 37.0 70.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/77fw40bb",
   "source_url": "https://omlx.ai/benchmarks/performance/77fw40bb",
   "tg_tokens": null,
   "tps": 37.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-77fw40bb-ctx16384",
   "model": "Ling-3.0-flash",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: i4413e2f, pqds84a8; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 476.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-flash-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 476.4 33.0 71.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/77fw40bb",
   "source_url": "https://omlx.ai/benchmarks/performance/77fw40bb",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-77fw40bb-ctx32768",
   "model": "Ling-3.0-flash",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 74.7 GB at this context; duplicate traces of the same session: i4413e2f, pqds84a8; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 427.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-flash-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 32k 427.5 32.6 74.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/77fw40bb",
   "source_url": "https://omlx.ai/benchmarks/performance/77fw40bb",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-7zwg1v5n-ctx1024",
   "model": "llama-3-1-8b-instruct",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 15.7 GB; TTFT 386 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2651.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Llama-3.1-8B-Instruct on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 unknown; 1k tokens 2,651 PP tok/s 27.6 TG tok/s 386 TTFT (ms) 15.7 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7zwg1v5n",
   "source_url": "https://omlx.ai/benchmarks/performance/7zwg1v5n",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-q818xl8r-ctx1024",
   "model": "llama-3-1-8b-instruct",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 5.0 GB; TTFT 2681 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 381.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Llama-3.1-8B-Instruct-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k tokens 381.9 PP tok/s 54.2 TG tok/s 2681 TTFT (ms) 5.0 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q818xl8r",
   "source_url": "https://omlx.ai/benchmarks/performance/q818xl8r",
   "tg_tokens": null,
   "tps": 54.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-1kk30rob-ctx1024",
   "model": "llama-3-1-8b-instruct",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 5.0 GB; TTFT 2703 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 378.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Llama-3.1-8B-Instruct-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k tokens 378.9 PP tok/s 53.3 TG tok/s 2703 TTFT (ms) 5.0 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1kk30rob",
   "source_url": "https://omlx.ai/benchmarks/performance/1kk30rob",
   "tg_tokens": null,
   "tps": 53.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-nqwz82rj-ctx1024",
   "model": "llama-3-1-8b-instruct",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 5.0 GB; TTFT 358 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2860.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Llama-3.1-8B-Instruct-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k tokens 2,860 PP tok/s 87.1 TG tok/s 358 TTFT (ms) 5.0 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nqwz82rj",
   "source_url": "https://omlx.ai/benchmarks/performance/nqwz82rj",
   "tg_tokens": null,
   "tps": 87.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-nqiuwimv-ctx1024",
   "model": "Nemotron-Cascade-2-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: rlc9ja9s; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 634.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Nemotron-Cascade-2-30B-A3B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 634.1 93.6 18.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nqiuwimv",
   "source_url": "https://omlx.ai/benchmarks/performance/nqiuwimv",
   "tg_tokens": null,
   "tps": 93.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-nqiuwimv-ctx4096",
   "model": "Nemotron-Cascade-2-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: rlc9ja9s; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 629.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Nemotron-Cascade-2-30B-A3B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 629.0 88.2 19.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nqiuwimv",
   "source_url": "https://omlx.ai/benchmarks/performance/nqiuwimv",
   "tg_tokens": null,
   "tps": 88.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-4561q73z-ctx1024",
   "model": "Occamy-1.0",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: 33wwatx3, ey6akcji; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1386.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Occamy-1.0-oQ8-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,386 68.9 35.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4561q73z",
   "source_url": "https://omlx.ai/benchmarks/performance/4561q73z",
   "tg_tokens": null,
   "tps": 68.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-4561q73z-ctx4096",
   "model": "Occamy-1.0",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: 33wwatx3, ey6akcji; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1296.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Occamy-1.0-oQ8-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 1,296 57.3 37.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4561q73z",
   "source_url": "https://omlx.ai/benchmarks/performance/4561q73z",
   "tg_tokens": null,
   "tps": 57.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-4561q73z-ctx8192",
   "model": "Occamy-1.0",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 37.9 GB at this context; duplicate traces of the same session: 33wwatx3, ey6akcji; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1026.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Occamy-1.0-oQ8-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 8k 1,026 58.8 37.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4561q73z",
   "source_url": "https://omlx.ai/benchmarks/performance/4561q73z",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-4561q73z-ctx16384",
   "model": "Occamy-1.0",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 38.7 GB at this context; duplicate traces of the same session: 33wwatx3, ey6akcji; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 960.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Occamy-1.0-oQ8-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 16k 960.5 50.3 38.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4561q73z",
   "source_url": "https://omlx.ai/benchmarks/performance/4561q73z",
   "tg_tokens": null,
   "tps": 50.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-4561q73z-ctx32768",
   "model": "Occamy-1.0",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 40.1 GB at this context; duplicate traces of the same session: 33wwatx3, ey6akcji; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 792.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Occamy-1.0-oQ8-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 32k 792.0 49.6 40.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4561q73z",
   "source_url": "https://omlx.ai/benchmarks/performance/4561q73z",
   "tg_tokens": null,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-4561q73z-ctx65536",
   "model": "Occamy-1.0",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 41.2 GB at this context; duplicate traces of the same session: 33wwatx3, ey6akcji; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 590.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Occamy-1.0-oQ8-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 64k 590.2 39.0 41.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4561q73z",
   "source_url": "https://omlx.ai/benchmarks/performance/4561q73z",
   "tg_tokens": null,
   "tps": 39.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-sez6hion-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 430.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 430.3 58.3 19.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sez6hion",
   "source_url": "https://omlx.ai/benchmarks/performance/sez6hion",
   "tg_tokens": null,
   "tps": 58.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-sez6hion-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 409.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 409.6 56.7 20.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sez6hion",
   "source_url": "https://omlx.ai/benchmarks/performance/sez6hion",
   "tg_tokens": null,
   "tps": 56.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-y2mnfzw4-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 138.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 138.5 27.1 17.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y2mnfzw4",
   "source_url": "https://omlx.ai/benchmarks/performance/y2mnfzw4",
   "tg_tokens": null,
   "tps": 27.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-y2mnfzw4-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 18.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 333.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 333.8 25.0 18.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y2mnfzw4",
   "source_url": "https://omlx.ai/benchmarks/performance/y2mnfzw4",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-8r0n8dn5-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 12.3 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 160.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 160.4 21.5 12.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8r0n8dn5",
   "source_url": "https://omlx.ai/benchmarks/performance/8r0n8dn5",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-8r0n8dn5-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 13.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 275.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 275.7 18.0 13.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8r0n8dn5",
   "source_url": "https://omlx.ai/benchmarks/performance/8r0n8dn5",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-6wayt66a-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: fvoam9am; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 919.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 919.0 90.1 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6wayt66a",
   "source_url": "https://omlx.ai/benchmarks/performance/6wayt66a",
   "tg_tokens": null,
   "tps": 90.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-6wayt66a-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: fvoam9am; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1082.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,082 85.8 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6wayt66a",
   "source_url": "https://omlx.ai/benchmarks/performance/6wayt66a",
   "tg_tokens": null,
   "tps": 85.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-6wayt66a-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: fvoam9am; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 998.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 998.5 80.2 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6wayt66a",
   "source_url": "https://omlx.ai/benchmarks/performance/6wayt66a",
   "tg_tokens": null,
   "tps": 80.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-6wayt66a-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: fvoam9am; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 835.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 835.4 73.0 21.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6wayt66a",
   "source_url": "https://omlx.ai/benchmarks/performance/6wayt66a",
   "tg_tokens": null,
   "tps": 73.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-61ppps46-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 1dvrgy83; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 931.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 931.6 88.4 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/61ppps46",
   "source_url": "https://omlx.ai/benchmarks/performance/61ppps46",
   "tg_tokens": null,
   "tps": 88.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-61ppps46-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 1dvrgy83; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1069.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,069 88.5 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/61ppps46",
   "source_url": "https://omlx.ai/benchmarks/performance/61ppps46",
   "tg_tokens": null,
   "tps": 88.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-61ppps46-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 1dvrgy83; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1014.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 1,014 82.5 21.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/61ppps46",
   "source_url": "https://omlx.ai/benchmarks/performance/61ppps46",
   "tg_tokens": null,
   "tps": 82.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-61ppps46-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 1dvrgy83; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 869.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 869.8 74.5 22.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/61ppps46",
   "source_url": "https://omlx.ai/benchmarks/performance/61ppps46",
   "tg_tokens": null,
   "tps": 74.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ff9i7nzg-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 9tpknegx, u5svoupk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 807.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 807.7 85.6 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ff9i7nzg",
   "source_url": "https://omlx.ai/benchmarks/performance/ff9i7nzg",
   "tg_tokens": null,
   "tps": 85.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ff9i7nzg-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 9tpknegx, u5svoupk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1012.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,012 82.1 21.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ff9i7nzg",
   "source_url": "https://omlx.ai/benchmarks/performance/ff9i7nzg",
   "tg_tokens": null,
   "tps": 82.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ff9i7nzg-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 9tpknegx, u5svoupk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 961.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 961.9 78.9 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ff9i7nzg",
   "source_url": "https://omlx.ai/benchmarks/performance/ff9i7nzg",
   "tg_tokens": null,
   "tps": 78.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ff9i7nzg-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 9tpknegx, u5svoupk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 806.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 806.6 70.8 22.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ff9i7nzg",
   "source_url": "https://omlx.ai/benchmarks/performance/ff9i7nzg",
   "tg_tokens": null,
   "tps": 70.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-cu3bwhxq-ctx4096",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 2.2 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1929.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B-8bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 8bit; 4k 1,929 136.6 2.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cu3bwhxq",
   "source_url": "https://omlx.ai/benchmarks/performance/cu3bwhxq",
   "tg_tokens": null,
   "tps": 136.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-cu3bwhxq-ctx16384",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 2.3 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1416.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-0.8B-8bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 8bit; 16k 1,416 118.3 2.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cu3bwhxq",
   "source_url": "https://omlx.ai/benchmarks/performance/cu3bwhxq",
   "tg_tokens": null,
   "tps": 118.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-9krtj1ag-ctx32768",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 5.0 GB at this context; duplicate traces of the same session: eclofn7s; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1609.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 16 GB \u00b7 4bit; 32k 1,609 32.7 5.0 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9krtj1ag",
   "source_url": "https://omlx.ai/benchmarks/performance/9krtj1ag",
   "tg_tokens": null,
   "tps": 32.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-9krtj1ag-ctx131072",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 9.4 GB at this context; duplicate traces of the same session: eclofn7s; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 774.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 16 GB \u00b7 4bit; 128k 774.8 17.5 9.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9krtj1ag",
   "source_url": "https://omlx.ai/benchmarks/performance/9krtj1ag",
   "tg_tokens": null,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-kj3s8uuk-ctx1024",
   "model": "Qwen3.5-9B-Distilled-OPUS-Heretic-MLX-VLM",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 572.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-Distilled-OPUS-Heretic-MLX-VLM-4bit on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 1k 572.7 23.0 5.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kj3s8uuk",
   "source_url": "https://omlx.ai/benchmarks/performance/kj3s8uuk",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-kj3s8uuk-ctx4096",
   "model": "Qwen3.5-9B-Distilled-OPUS-Heretic-MLX-VLM",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 603.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-Distilled-OPUS-Heretic-MLX-VLM-4bit on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 4k 603.9 22.5 6.2 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kj3s8uuk",
   "source_url": "https://omlx.ai/benchmarks/performance/kj3s8uuk",
   "tg_tokens": null,
   "tps": 22.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-3dw8lbqt-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 4o60o389, 897mgrbk, y156cuxa; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 935.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 1k 935.7 82.6 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3dw8lbqt",
   "source_url": "https://omlx.ai/benchmarks/performance/3dw8lbqt",
   "tg_tokens": null,
   "tps": 82.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-3dw8lbqt-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 4o60o389, 897mgrbk, y156cuxa; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 973.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k 973.2 84.5 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3dw8lbqt",
   "source_url": "https://omlx.ai/benchmarks/performance/3dw8lbqt",
   "tg_tokens": null,
   "tps": 84.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-3dw8lbqt-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: 4o60o389, 897mgrbk, y156cuxa; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 800.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 8k 800.8 78.2 21.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3dw8lbqt",
   "source_url": "https://omlx.ai/benchmarks/performance/3dw8lbqt",
   "tg_tokens": null,
   "tps": 78.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-3dw8lbqt-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 4o60o389, 897mgrbk, y156cuxa; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 665.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 16k 665.2 68.9 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3dw8lbqt",
   "source_url": "https://omlx.ai/benchmarks/performance/3dw8lbqt",
   "tg_tokens": null,
   "tps": 68.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1o87d9da-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: wqjtgpdr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 831.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 831.8 120.1 21.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1o87d9da",
   "source_url": "https://omlx.ai/benchmarks/performance/1o87d9da",
   "tg_tokens": null,
   "tps": 120.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1o87d9da-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: wqjtgpdr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3018.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 3,018 144.0 22.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1o87d9da",
   "source_url": "https://omlx.ai/benchmarks/performance/1o87d9da",
   "tg_tokens": null,
   "tps": 144.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1o87d9da-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: wqjtgpdr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5988.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 5,988 47.4 22.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1o87d9da",
   "source_url": "https://omlx.ai/benchmarks/performance/1o87d9da",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1o87d9da-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: wqjtgpdr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5530.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 5,530 176.4 22.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1o87d9da",
   "source_url": "https://omlx.ai/benchmarks/performance/1o87d9da",
   "tg_tokens": null,
   "tps": 176.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1o87d9da-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: wqjtgpdr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4485.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 4,485 221.0 22.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1o87d9da",
   "source_url": "https://omlx.ai/benchmarks/performance/1o87d9da",
   "tg_tokens": null,
   "tps": 221.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1o87d9da-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: wqjtgpdr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3333.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 3,333 142.7 23.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1o87d9da",
   "source_url": "https://omlx.ai/benchmarks/performance/1o87d9da",
   "tg_tokens": null,
   "tps": 142.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-oflhwc4r-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 87.2,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 1k 87.2 13.3 22.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oflhwc4r",
   "source_url": "https://omlx.ai/benchmarks/performance/oflhwc4r",
   "tg_tokens": null,
   "tps": 13.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-oflhwc4r-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 92.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 4k 92.6 13.0 23.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oflhwc4r",
   "source_url": "https://omlx.ai/benchmarks/performance/oflhwc4r",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-o71yfvb2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 18.4 GB; TTFT 16996 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 241.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k tokens 241.0 PP tok/s 16.7 TG tok/s 16996 TTFT (ms) 18.4 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o71yfvb2",
   "source_url": "https://omlx.ai/benchmarks/performance/o71yfvb2",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-x8myzysw-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 1oxrncmf, igrt6lcf, zawmq8kw; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 767.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 767.5 26.2 16.0 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x8myzysw",
   "source_url": "https://omlx.ai/benchmarks/performance/x8myzysw",
   "tg_tokens": null,
   "tps": 26.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-x8myzysw-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 1oxrncmf, igrt6lcf, zawmq8kw; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 822.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 822.2 25.6 17.0 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x8myzysw",
   "source_url": "https://omlx.ai/benchmarks/performance/x8myzysw",
   "tg_tokens": null,
   "tps": 25.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-x8myzysw-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 1oxrncmf, igrt6lcf, zawmq8kw; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 824.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 824.3 25.2 17.3 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x8myzysw",
   "source_url": "https://omlx.ai/benchmarks/performance/x8myzysw",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-x8myzysw-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 1oxrncmf, igrt6lcf, zawmq8kw; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 800.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 800.3 24.0 17.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x8myzysw",
   "source_url": "https://omlx.ai/benchmarks/performance/x8myzysw",
   "tg_tokens": null,
   "tps": 24.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-x8myzysw-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 1oxrncmf, igrt6lcf, zawmq8kw; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 748.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 32k 748.0 22.0 18.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x8myzysw",
   "source_url": "https://omlx.ai/benchmarks/performance/x8myzysw",
   "tg_tokens": null,
   "tps": 22.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-x8myzysw-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 1oxrncmf, igrt6lcf, zawmq8kw; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 662.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 64k 662.5 20.5 20.8 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x8myzysw",
   "source_url": "https://omlx.ai/benchmarks/performance/x8myzysw",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-akxk3nwq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.3 GB at this context; duplicate traces of the same session: elxnvgn1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 1k 125.0 8.6 28.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/akxk3nwq",
   "source_url": "https://omlx.ai/benchmarks/performance/akxk3nwq",
   "tg_tokens": null,
   "tps": 8.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-akxk3nwq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: elxnvgn1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 126.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 4k 126.1 8.5 29.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/akxk3nwq",
   "source_url": "https://omlx.ai/benchmarks/performance/akxk3nwq",
   "tg_tokens": null,
   "tps": 8.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ahqdovwp-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: 2tis2g72, cfphn1v5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 895.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 895.0 59.6 23.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ahqdovwp",
   "source_url": "https://omlx.ai/benchmarks/performance/ahqdovwp",
   "tg_tokens": null,
   "tps": 59.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ahqdovwp-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: 2tis2g72, cfphn1v5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 779.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 779.5 45.9 28.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ahqdovwp",
   "source_url": "https://omlx.ai/benchmarks/performance/ahqdovwp",
   "tg_tokens": null,
   "tps": 45.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ahqdovwp-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 37.8 GB at this context; duplicate traces of the same session: 2tis2g72, cfphn1v5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 570.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 570.9 39.9 37.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ahqdovwp",
   "source_url": "https://omlx.ai/benchmarks/performance/ahqdovwp",
   "tg_tokens": null,
   "tps": 39.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-32zcmxma-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: mev7hi8q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 821.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 821.6 41.4 17.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/32zcmxma",
   "source_url": "https://omlx.ai/benchmarks/performance/32zcmxma",
   "tg_tokens": null,
   "tps": 41.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-32zcmxma-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: mev7hi8q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1304.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,304 53.1 19.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/32zcmxma",
   "source_url": "https://omlx.ai/benchmarks/performance/32zcmxma",
   "tg_tokens": null,
   "tps": 53.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-32zcmxma-ctx8192",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: mev7hi8q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1393.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,393 60.7 19.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/32zcmxma",
   "source_url": "https://omlx.ai/benchmarks/performance/32zcmxma",
   "tg_tokens": null,
   "tps": 60.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-32zcmxma-ctx16384",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: mev7hi8q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1300.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,300 77.9 19.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/32zcmxma",
   "source_url": "https://omlx.ai/benchmarks/performance/32zcmxma",
   "tg_tokens": null,
   "tps": 77.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-32zcmxma-ctx32768",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: mev7hi8q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1146.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,146 91.9 21.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/32zcmxma",
   "source_url": "https://omlx.ai/benchmarks/performance/32zcmxma",
   "tg_tokens": null,
   "tps": 91.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-32zcmxma-ctx65536",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: mev7hi8q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 963.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 963.9 48.8 23.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/32zcmxma",
   "source_url": "https://omlx.ai/benchmarks/performance/32zcmxma",
   "tg_tokens": null,
   "tps": 48.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4xejwt5l-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 762.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 762.2 40.1 33.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4xejwt5l",
   "source_url": "https://omlx.ai/benchmarks/performance/4xejwt5l",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4xejwt5l-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 908.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 908.2 48.3 34.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4xejwt5l",
   "source_url": "https://omlx.ai/benchmarks/performance/4xejwt5l",
   "tg_tokens": null,
   "tps": 48.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4xejwt5l-ctx8192",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 832.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 832.7 49.8 35.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4xejwt5l",
   "source_url": "https://omlx.ai/benchmarks/performance/4xejwt5l",
   "tg_tokens": null,
   "tps": 49.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4xejwt5l-ctx16384",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 799.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 799.6 55.8 35.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4xejwt5l",
   "source_url": "https://omlx.ai/benchmarks/performance/4xejwt5l",
   "tg_tokens": null,
   "tps": 55.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4xejwt5l-ctx32768",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 749.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 749.1 82.9 36.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4xejwt5l",
   "source_url": "https://omlx.ai/benchmarks/performance/4xejwt5l",
   "tg_tokens": null,
   "tps": 82.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4xejwt5l-ctx65536",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 665.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 665.6 46.6 39.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4xejwt5l",
   "source_url": "https://omlx.ai/benchmarks/performance/4xejwt5l",
   "tg_tokens": null,
   "tps": 46.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-0km9qt4w-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-NEO-CODER",
   "notes": "1M context build; oMLX community benchmark trace; peak memory 28.3 GB at this context; duplicate traces of the same session: ajbkm7xp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 255.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-NEO-CODER-mlx-oQ8e-1M on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 255.0 16.6 28.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0km9qt4w",
   "source_url": "https://omlx.ai/benchmarks/performance/0km9qt4w",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-0km9qt4w-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-NEO-CODER",
   "notes": "1M context build; oMLX community benchmark trace; peak memory 31.7 GB at this context; duplicate traces of the same session: ajbkm7xp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 262.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-NEO-CODER-mlx-oQ8e-1M on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 262.6 16.5 31.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0km9qt4w",
   "source_url": "https://omlx.ai/benchmarks/performance/0km9qt4w",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-elcuiri1-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: f6irs7xx; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 324.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 324.7 70.0 21.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/elcuiri1",
   "source_url": "https://omlx.ai/benchmarks/performance/elcuiri1",
   "tg_tokens": null,
   "tps": 70.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-elcuiri1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: f6irs7xx; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 431.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 431.7 66.8 24.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/elcuiri1",
   "source_url": "https://omlx.ai/benchmarks/performance/elcuiri1",
   "tg_tokens": null,
   "tps": 66.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qgoelvxa-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: vg160xma; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 323.5 40.3 21.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qgoelvxa",
   "source_url": "https://omlx.ai/benchmarks/performance/qgoelvxa",
   "tg_tokens": null,
   "tps": 40.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qgoelvxa-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: vg160xma; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 438.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 438.2 50.8 24.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qgoelvxa",
   "source_url": "https://omlx.ai/benchmarks/performance/qgoelvxa",
   "tg_tokens": null,
   "tps": 50.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-pwh16tun-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: coy0uff6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 333.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 333.2 36.9 16.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pwh16tun",
   "source_url": "https://omlx.ai/benchmarks/performance/pwh16tun",
   "tg_tokens": null,
   "tps": 36.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-pwh16tun-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: coy0uff6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 356.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 356.3 35.9 19.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pwh16tun",
   "source_url": "https://omlx.ai/benchmarks/performance/pwh16tun",
   "tg_tokens": null,
   "tps": 35.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-rpskl4f5-ctx4096",
   "model": "Qwen3.8-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 186.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-mlx-4Bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 186.9 30.7 6.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rpskl4f5",
   "source_url": "https://omlx.ai/benchmarks/performance/rpskl4f5",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-rpskl4f5-ctx16384",
   "model": "Qwen3.8-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.1 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 162.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-mlx-4Bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k 162.0 27.6 7.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rpskl4f5",
   "source_url": "https://omlx.ai/benchmarks/performance/rpskl4f5",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-6ar9h3sl-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 44.8 GB at this context; duplicate traces of the same session: hkwm4aw2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 618.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 618.0 49.8 44.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ar9h3sl",
   "source_url": "https://omlx.ai/benchmarks/performance/6ar9h3sl",
   "tg_tokens": null,
   "tps": 49.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-6ar9h3sl-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 45.4 GB at this context; duplicate traces of the same session: hkwm4aw2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 759.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 759.2 43.6 45.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ar9h3sl",
   "source_url": "https://omlx.ai/benchmarks/performance/6ar9h3sl",
   "tg_tokens": null,
   "tps": 43.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-h0gpsm4f-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 44.8 GB at this context; duplicate traces of the same session: x5fcn2gk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 507.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 507.6 40.1 44.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h0gpsm4f",
   "source_url": "https://omlx.ai/benchmarks/performance/h0gpsm4f",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-h0gpsm4f-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 45.4 GB at this context; duplicate traces of the same session: x5fcn2gk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 671.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 671.1 37.6 45.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h0gpsm4f",
   "source_url": "https://omlx.ai/benchmarks/performance/h0gpsm4f",
   "tg_tokens": null,
   "tps": 37.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-l8jvfkkb-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.4 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 924.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 924.2 81.9 100.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l8jvfkkb",
   "source_url": "https://omlx.ai/benchmarks/performance/l8jvfkkb",
   "tg_tokens": null,
   "tps": 81.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-l8jvfkkb-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1133.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,133 70.1 101.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l8jvfkkb",
   "source_url": "https://omlx.ai/benchmarks/performance/l8jvfkkb",
   "tg_tokens": null,
   "tps": 70.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-v8snseo1-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.4 GB at this context; duplicate traces of the same session: 1yp0hkd9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1073.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,073 104.1 100.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v8snseo1",
   "source_url": "https://omlx.ai/benchmarks/performance/v8snseo1",
   "tg_tokens": null,
   "tps": 104.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-v8snseo1-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: 1yp0hkd9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1379.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,379 95.7 101.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v8snseo1",
   "source_url": "https://omlx.ai/benchmarks/performance/v8snseo1",
   "tg_tokens": null,
   "tps": 95.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hdxtiqbi-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.4 GB at this context; duplicate traces of the same session: 1km0scoe; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1073.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,073 92.7 100.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hdxtiqbi",
   "source_url": "https://omlx.ai/benchmarks/performance/hdxtiqbi",
   "tg_tokens": null,
   "tps": 92.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hdxtiqbi-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: 1km0scoe; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1379.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,379 95.2 101.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hdxtiqbi",
   "source_url": "https://omlx.ai/benchmarks/performance/hdxtiqbi",
   "tg_tokens": null,
   "tps": 95.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-tckoggc3-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.1 GB at this context; duplicate traces of the same session: 2oo3v27w, k9li9eh0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1107.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 1,107 84.3 183.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tckoggc3",
   "source_url": "https://omlx.ai/benchmarks/performance/tckoggc3",
   "tg_tokens": null,
   "tps": 84.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-tckoggc3-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.6 GB at this context; duplicate traces of the same session: 2oo3v27w, k9li9eh0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1380.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 1,380 66.6 183.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tckoggc3",
   "source_url": "https://omlx.ai/benchmarks/performance/tckoggc3",
   "tg_tokens": null,
   "tps": 66.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-tckoggc3-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.8 GB at this context; duplicate traces of the same session: 2oo3v27w, k9li9eh0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1364.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 1,364 88.3 183.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tckoggc3",
   "source_url": "https://omlx.ai/benchmarks/performance/tckoggc3",
   "tg_tokens": null,
   "tps": 88.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-3i4pn0jr-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 183.1 GB; TTFT 1218 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 840.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k tokens 840.7 PP tok/s 48.0 TG tok/s 1218 TTFT (ms) 183.1 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3i4pn0jr",
   "source_url": "https://omlx.ai/benchmarks/performance/3i4pn0jr",
   "tg_tokens": null,
   "tps": 48.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-e37y8l1l-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 417.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 417.3 43.6 21.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e37y8l1l",
   "source_url": "https://omlx.ai/benchmarks/performance/e37y8l1l",
   "tg_tokens": null,
   "tps": 43.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-e37y8l1l-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 405.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 405.3 46.3 22.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e37y8l1l",
   "source_url": "https://omlx.ai/benchmarks/performance/e37y8l1l",
   "tg_tokens": null,
   "tps": 46.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-e37y8l1l-ctx8192",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 377.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 8k 377.1 40.0 22.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e37y8l1l",
   "source_url": "https://omlx.ai/benchmarks/performance/e37y8l1l",
   "tg_tokens": null,
   "tps": 40.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-e37y8l1l-ctx16384",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 321.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k 321.1 34.2 22.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e37y8l1l",
   "source_url": "https://omlx.ai/benchmarks/performance/e37y8l1l",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-2iynds49-ctx1024",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: adyhuux6; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 81.9,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 5bit; 1k 81.9 10.9 19.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2iynds49",
   "source_url": "https://omlx.ai/benchmarks/performance/2iynds49",
   "tg_tokens": null,
   "tps": 10.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-2iynds49-ctx4096",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: adyhuux6; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 76.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 5bit; 4k 76.0 10.1 20.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2iynds49",
   "source_url": "https://omlx.ai/benchmarks/performance/2iynds49",
   "tg_tokens": null,
   "tps": 10.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-2iynds49-ctx8192",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: adyhuux6; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 74.3,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 5bit; 8k 74.3 9.9 21.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2iynds49",
   "source_url": "https://omlx.ai/benchmarks/performance/2iynds49",
   "tg_tokens": null,
   "tps": 9.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-2iynds49-ctx16384",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: adyhuux6; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 70.4,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 5bit; 16k 70.4 9.1 21.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2iynds49",
   "source_url": "https://omlx.ai/benchmarks/performance/2iynds49",
   "tg_tokens": null,
   "tps": 9.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-5ulvaqlf-ctx1024",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 404.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 404.0 50.0 20.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ulvaqlf",
   "source_url": "https://omlx.ai/benchmarks/performance/5ulvaqlf",
   "tg_tokens": null,
   "tps": 50.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-5ulvaqlf-ctx4096",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 415.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 415.1 46.6 20.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ulvaqlf",
   "source_url": "https://omlx.ai/benchmarks/performance/5ulvaqlf",
   "tg_tokens": null,
   "tps": 46.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-5ulvaqlf-ctx8192",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 380.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 8k 380.6 46.7 21.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ulvaqlf",
   "source_url": "https://omlx.ai/benchmarks/performance/5ulvaqlf",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-lzdrrs9h-ctx1024",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 143.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 143.0 13.0 18.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lzdrrs9h",
   "source_url": "https://omlx.ai/benchmarks/performance/lzdrrs9h",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-lzdrrs9h-ctx4096",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 348.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 348.8 12.3 19.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lzdrrs9h",
   "source_url": "https://omlx.ai/benchmarks/performance/lzdrrs9h",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-lzdrrs9h-ctx8192",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 330.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 8k 330.0 30.4 19.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lzdrrs9h",
   "source_url": "https://omlx.ai/benchmarks/performance/lzdrrs9h",
   "tg_tokens": null,
   "tps": 30.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-cn175l22-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 316.2 GB at this context; duplicate traces of the same session: k51hgfi7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 430.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "GLM-5.3-Flash-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 430.3 32.4 316.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cn175l22",
   "source_url": "https://omlx.ai/benchmarks/performance/cn175l22",
   "tg_tokens": null,
   "tps": 32.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-cn175l22-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 317.3 GB at this context; duplicate traces of the same session: k51hgfi7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 551.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "GLM-5.3-Flash-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 551.2 31.1 317.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cn175l22",
   "source_url": "https://omlx.ai/benchmarks/performance/cn175l22",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-cn175l22-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 317.6 GB at this context; duplicate traces of the same session: k51hgfi7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 513.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "GLM-5.3-Flash-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 513.3 31.2 317.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cn175l22",
   "source_url": "https://omlx.ai/benchmarks/performance/cn175l22",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-cn175l22-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 317.7 GB at this context; duplicate traces of the same session: k51hgfi7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 545.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "GLM-5.3-Flash-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k 545.4 30.9 317.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cn175l22",
   "source_url": "https://omlx.ai/benchmarks/performance/cn175l22",
   "tg_tokens": null,
   "tps": 30.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-9wkowbnb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: g5h081v1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 240.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 240.8 65.4 18.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9wkowbnb",
   "source_url": "https://omlx.ai/benchmarks/performance/9wkowbnb",
   "tg_tokens": null,
   "tps": 65.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-9wkowbnb-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: g5h081v1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 240.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 240.4 76.8 23.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9wkowbnb",
   "source_url": "https://omlx.ai/benchmarks/performance/9wkowbnb",
   "tg_tokens": null,
   "tps": 76.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-9900mmaz-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 5zulu72e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 234.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 234.7 56.2 18.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9900mmaz",
   "source_url": "https://omlx.ai/benchmarks/performance/9900mmaz",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-9900mmaz-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 5zulu72e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 235.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 235.1 60.0 23.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9900mmaz",
   "source_url": "https://omlx.ai/benchmarks/performance/9900mmaz",
   "tg_tokens": null,
   "tps": 60.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-yzusuvtu-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: c45yoxsp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 244.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 244.2 25.0 16.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yzusuvtu",
   "source_url": "https://omlx.ai/benchmarks/performance/yzusuvtu",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-yzusuvtu-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: c45yoxsp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 240.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 240.6 21.1 21.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yzusuvtu",
   "source_url": "https://omlx.ai/benchmarks/performance/yzusuvtu",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-yiv0ytc2-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: f10yf2zw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 234.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 234.8 51.4 18.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yiv0ytc2",
   "source_url": "https://omlx.ai/benchmarks/performance/yiv0ytc2",
   "tg_tokens": null,
   "tps": 51.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-yiv0ytc2-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: f10yf2zw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 234.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 234.1 47.4 23.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yiv0ytc2",
   "source_url": "https://omlx.ai/benchmarks/performance/yiv0ytc2",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-cqteqm2s-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: uezdsol8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 238.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 238.6 65.0 18.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cqteqm2s",
   "source_url": "https://omlx.ai/benchmarks/performance/cqteqm2s",
   "tg_tokens": null,
   "tps": 65.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-cqteqm2s-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: uezdsol8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 234.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 234.8 67.2 23.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cqteqm2s",
   "source_url": "https://omlx.ai/benchmarks/performance/cqteqm2s",
   "tg_tokens": null,
   "tps": 67.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-pqds84a8-ctx1024",
   "model": "Ling-3.0-flash",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 68.3 GB at this context; duplicate traces of the same session: i4413e2f; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 594.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-flash-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 594.8 41.4 68.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pqds84a8",
   "source_url": "https://omlx.ai/benchmarks/performance/pqds84a8",
   "tg_tokens": null,
   "tps": 41.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-pqds84a8-ctx4096",
   "model": "Ling-3.0-flash",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 69.5 GB at this context; duplicate traces of the same session: i4413e2f; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 525.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-flash-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 525.3 37.9 69.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pqds84a8",
   "source_url": "https://omlx.ai/benchmarks/performance/pqds84a8",
   "tg_tokens": null,
   "tps": 37.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-pqds84a8-ctx8192",
   "model": "Ling-3.0-flash",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 70.2 GB at this context; duplicate traces of the same session: i4413e2f; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 505.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-flash-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 505.0 37.0 70.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pqds84a8",
   "source_url": "https://omlx.ai/benchmarks/performance/pqds84a8",
   "tg_tokens": null,
   "tps": 37.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-pqds84a8-ctx16384",
   "model": "Ling-3.0-flash",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: i4413e2f; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 476.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-flash-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 476.4 33.0 71.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pqds84a8",
   "source_url": "https://omlx.ai/benchmarks/performance/pqds84a8",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-pqds84a8-ctx32768",
   "model": "Ling-3.0-flash",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 74.7 GB at this context; duplicate traces of the same session: i4413e2f; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 427.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-flash-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 32k 427.5 32.6 74.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pqds84a8",
   "source_url": "https://omlx.ai/benchmarks/performance/pqds84a8",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-rlc9ja9s-ctx1024",
   "model": "Nemotron-Cascade-2-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 634.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Nemotron-Cascade-2-30B-A3B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 634.1 93.6 18.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rlc9ja9s",
   "source_url": "https://omlx.ai/benchmarks/performance/rlc9ja9s",
   "tg_tokens": null,
   "tps": 93.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-rlc9ja9s-ctx4096",
   "model": "Nemotron-Cascade-2-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 629.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Nemotron-Cascade-2-30B-A3B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 629.0 88.2 19.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rlc9ja9s",
   "source_url": "https://omlx.ai/benchmarks/performance/rlc9ja9s",
   "tg_tokens": null,
   "tps": 88.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ey6akcji-ctx1024",
   "model": "Occamy-1.0",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: 33wwatx3; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1386.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Occamy-1.0-oQ8-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,386 68.9 35.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ey6akcji",
   "source_url": "https://omlx.ai/benchmarks/performance/ey6akcji",
   "tg_tokens": null,
   "tps": 68.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ey6akcji-ctx4096",
   "model": "Occamy-1.0",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: 33wwatx3; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1296.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Occamy-1.0-oQ8-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 1,296 57.3 37.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ey6akcji",
   "source_url": "https://omlx.ai/benchmarks/performance/ey6akcji",
   "tg_tokens": null,
   "tps": 57.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ey6akcji-ctx8192",
   "model": "Occamy-1.0",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 37.9 GB at this context; duplicate traces of the same session: 33wwatx3; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1026.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Occamy-1.0-oQ8-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 8k 1,026 58.8 37.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ey6akcji",
   "source_url": "https://omlx.ai/benchmarks/performance/ey6akcji",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ey6akcji-ctx16384",
   "model": "Occamy-1.0",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 38.7 GB at this context; duplicate traces of the same session: 33wwatx3; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 960.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Occamy-1.0-oQ8-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 16k 960.5 50.3 38.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ey6akcji",
   "source_url": "https://omlx.ai/benchmarks/performance/ey6akcji",
   "tg_tokens": null,
   "tps": 50.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ey6akcji-ctx32768",
   "model": "Occamy-1.0",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 40.1 GB at this context; duplicate traces of the same session: 33wwatx3; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 792.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Occamy-1.0-oQ8-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 32k 792.0 49.6 40.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ey6akcji",
   "source_url": "https://omlx.ai/benchmarks/performance/ey6akcji",
   "tg_tokens": null,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ey6akcji-ctx65536",
   "model": "Occamy-1.0",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 41.2 GB at this context; duplicate traces of the same session: 33wwatx3; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 590.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Occamy-1.0-oQ8-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 64k 590.2 39.0 41.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ey6akcji",
   "source_url": "https://omlx.ai/benchmarks/performance/ey6akcji",
   "tg_tokens": null,
   "tps": 39.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-1dvrgy83-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 931.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 931.6 88.4 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1dvrgy83",
   "source_url": "https://omlx.ai/benchmarks/performance/1dvrgy83",
   "tg_tokens": null,
   "tps": 88.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-1dvrgy83-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1069.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,069 88.5 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1dvrgy83",
   "source_url": "https://omlx.ai/benchmarks/performance/1dvrgy83",
   "tg_tokens": null,
   "tps": 88.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-1dvrgy83-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1014.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 1,014 82.5 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1dvrgy83",
   "source_url": "https://omlx.ai/benchmarks/performance/1dvrgy83",
   "tg_tokens": null,
   "tps": 82.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-1dvrgy83-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 869.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 869.8 74.5 22.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1dvrgy83",
   "source_url": "https://omlx.ai/benchmarks/performance/1dvrgy83",
   "tg_tokens": null,
   "tps": 74.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-fvoam9am-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 919.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 919.0 90.1 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fvoam9am",
   "source_url": "https://omlx.ai/benchmarks/performance/fvoam9am",
   "tg_tokens": null,
   "tps": 90.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-fvoam9am-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1082.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,082 85.8 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fvoam9am",
   "source_url": "https://omlx.ai/benchmarks/performance/fvoam9am",
   "tg_tokens": null,
   "tps": 85.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-fvoam9am-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 998.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 998.5 80.2 21.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fvoam9am",
   "source_url": "https://omlx.ai/benchmarks/performance/fvoam9am",
   "tg_tokens": null,
   "tps": 80.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-fvoam9am-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 835.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 835.4 73.0 21.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fvoam9am",
   "source_url": "https://omlx.ai/benchmarks/performance/fvoam9am",
   "tg_tokens": null,
   "tps": 73.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-u5svoupk-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 9tpknegx; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 807.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 807.7 85.6 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u5svoupk",
   "source_url": "https://omlx.ai/benchmarks/performance/u5svoupk",
   "tg_tokens": null,
   "tps": 85.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-u5svoupk-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 9tpknegx; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1012.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,012 82.1 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u5svoupk",
   "source_url": "https://omlx.ai/benchmarks/performance/u5svoupk",
   "tg_tokens": null,
   "tps": 82.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-u5svoupk-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 9tpknegx; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 961.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 961.9 78.9 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u5svoupk",
   "source_url": "https://omlx.ai/benchmarks/performance/u5svoupk",
   "tg_tokens": null,
   "tps": 78.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-u5svoupk-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 9tpknegx; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 806.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 806.6 70.8 22.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u5svoupk",
   "source_url": "https://omlx.ai/benchmarks/performance/u5svoupk",
   "tg_tokens": null,
   "tps": 70.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-en4hnhjd-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1064.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 16 GB \u00b7 4bit; 4k 1,064 26.5 6.3 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/en4hnhjd",
   "source_url": "https://omlx.ai/benchmarks/performance/en4hnhjd",
   "tg_tokens": null,
   "tps": 26.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-en4hnhjd-ctx65536",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.1 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 697.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 16 GB \u00b7 4bit; 64k 697.1 16.8 8.1 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/en4hnhjd",
   "source_url": "https://omlx.ai/benchmarks/performance/en4hnhjd",
   "tg_tokens": null,
   "tps": 16.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-zov5piem-ctx4096",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 2.0 GB at this context; duplicate traces of the same session: p7al850j; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1958.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.5-0.8B-OptiQ-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 1,958 152.9 2.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zov5piem",
   "source_url": "https://omlx.ai/benchmarks/performance/zov5piem",
   "tg_tokens": null,
   "tps": 152.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-zov5piem-ctx16384",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 2.2 GB at this context; duplicate traces of the same session: p7al850j; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1406.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.5-0.8B-OptiQ-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k 1,406 126.9 2.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zov5piem",
   "source_url": "https://omlx.ai/benchmarks/performance/zov5piem",
   "tg_tokens": null,
   "tps": 126.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-io00o3c7-ctx4096",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: z2gx7sam; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2132.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 2,132 68.4 19.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/io00o3c7",
   "source_url": "https://omlx.ai/benchmarks/performance/io00o3c7",
   "tg_tokens": null,
   "tps": 68.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-io00o3c7-ctx16384",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: z2gx7sam; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1920.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 1,920 61.5 20.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/io00o3c7",
   "source_url": "https://omlx.ai/benchmarks/performance/io00o3c7",
   "tg_tokens": null,
   "tps": 61.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-xaz9surw-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.8 GB at this context; duplicate traces of the same session: 8pw774p6; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1832.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,832 47.4 3.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xaz9surw",
   "source_url": "https://omlx.ai/benchmarks/performance/xaz9surw",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-xaz9surw-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.1 GB at this context; duplicate traces of the same session: 8pw774p6; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1759.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 1,759 46.4 4.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xaz9surw",
   "source_url": "https://omlx.ai/benchmarks/performance/xaz9surw",
   "tg_tokens": null,
   "tps": 46.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-eclofn7s-ctx32768",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 5.0 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1609.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 16 GB \u00b7 4bit; 32k 1,609 32.7 5.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eclofn7s",
   "source_url": "https://omlx.ai/benchmarks/performance/eclofn7s",
   "tg_tokens": null,
   "tps": 32.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-eclofn7s-ctx131072",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 9.4 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 774.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 16 GB \u00b7 4bit; 128k 774.8 17.5 9.4 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eclofn7s",
   "source_url": "https://omlx.ai/benchmarks/performance/eclofn7s",
   "tg_tokens": null,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-l9kqnbju-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: t4t83rez; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1117.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,117 30.6 6.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l9kqnbju",
   "source_url": "https://omlx.ai/benchmarks/performance/l9kqnbju",
   "tg_tokens": null,
   "tps": 30.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-l9kqnbju-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.6 GB at this context; duplicate traces of the same session: t4t83rez; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1083.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 1,083 28.5 6.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l9kqnbju",
   "source_url": "https://omlx.ai/benchmarks/performance/l9kqnbju",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-pzd0catf-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: uc0kd8rf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2731.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,731 107.5 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pzd0catf",
   "source_url": "https://omlx.ai/benchmarks/performance/pzd0catf",
   "tg_tokens": null,
   "tps": 107.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-pzd0catf-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: uc0kd8rf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3412.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,412 104.9 20.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pzd0catf",
   "source_url": "https://omlx.ai/benchmarks/performance/pzd0catf",
   "tg_tokens": null,
   "tps": 104.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-pzd0catf-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: uc0kd8rf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3042.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 3,042 94.0 20.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pzd0catf",
   "source_url": "https://omlx.ai/benchmarks/performance/pzd0catf",
   "tg_tokens": null,
   "tps": 94.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-y156cuxa-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 4o60o389, 897mgrbk; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 935.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 1k 935.7 82.6 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y156cuxa",
   "source_url": "https://omlx.ai/benchmarks/performance/y156cuxa",
   "tg_tokens": null,
   "tps": 82.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-y156cuxa-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 4o60o389, 897mgrbk; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 973.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k 973.2 84.5 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y156cuxa",
   "source_url": "https://omlx.ai/benchmarks/performance/y156cuxa",
   "tg_tokens": null,
   "tps": 84.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-y156cuxa-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: 4o60o389, 897mgrbk; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 800.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 8k 800.8 78.2 21.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y156cuxa",
   "source_url": "https://omlx.ai/benchmarks/performance/y156cuxa",
   "tg_tokens": null,
   "tps": 78.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-y156cuxa-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 4o60o389, 897mgrbk; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 665.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 16k 665.2 68.9 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y156cuxa",
   "source_url": "https://omlx.ai/benchmarks/performance/y156cuxa",
   "tg_tokens": null,
   "tps": 68.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jgxn6kzp-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 6x5dd1b0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3830.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 3,830 103.4 21.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jgxn6kzp",
   "source_url": "https://omlx.ai/benchmarks/performance/jgxn6kzp",
   "tg_tokens": null,
   "tps": 103.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jgxn6kzp-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: 6x5dd1b0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2577.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,577 79.5 22.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jgxn6kzp",
   "source_url": "https://omlx.ai/benchmarks/performance/jgxn6kzp",
   "tg_tokens": null,
   "tps": 79.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mwuiy512-ctx1024",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: 1xjauexg, 42qc8u2n, 5cusv2p5, 98ogtog7, af67bxzg, hh6rzjmu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 422.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 422.5 110.9 12.7 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mwuiy512",
   "source_url": "https://omlx.ai/benchmarks/performance/mwuiy512",
   "tg_tokens": null,
   "tps": 110.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mwuiy512-ctx8192",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: 1xjauexg, 42qc8u2n, 5cusv2p5, 98ogtog7, af67bxzg, hh6rzjmu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4911.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 4,911 104.3 13.4 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mwuiy512",
   "source_url": "https://omlx.ai/benchmarks/performance/mwuiy512",
   "tg_tokens": null,
   "tps": 104.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mwuiy512-ctx16384",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.6 GB at this context; duplicate traces of the same session: 1xjauexg, 42qc8u2n, 5cusv2p5, 98ogtog7, af67bxzg, hh6rzjmu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4519.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 4,519 99.4 13.6 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mwuiy512",
   "source_url": "https://omlx.ai/benchmarks/performance/mwuiy512",
   "tg_tokens": null,
   "tps": 99.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mwuiy512-ctx32768",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.9 GB at this context; duplicate traces of the same session: 1xjauexg, 42qc8u2n, 5cusv2p5, 98ogtog7, af67bxzg, hh6rzjmu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3947.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 32k 3,947 89.5 13.9 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mwuiy512",
   "source_url": "https://omlx.ai/benchmarks/performance/mwuiy512",
   "tg_tokens": null,
   "tps": 89.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mwuiy512-ctx65536",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 14.5 GB at this context; duplicate traces of the same session: 1xjauexg, 42qc8u2n, 5cusv2p5, 98ogtog7, af67bxzg, hh6rzjmu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3024.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 64k 3,024 74.8 14.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mwuiy512",
   "source_url": "https://omlx.ai/benchmarks/performance/mwuiy512",
   "tg_tokens": null,
   "tps": 74.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mwuiy512-ctx131072",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: 1xjauexg, 42qc8u2n, 5cusv2p5, 98ogtog7, af67bxzg, hh6rzjmu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2041.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 128k 2,041 57.7 15.8 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mwuiy512",
   "source_url": "https://omlx.ai/benchmarks/performance/mwuiy512",
   "tg_tokens": null,
   "tps": 57.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mwuiy512-ctx199680",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: 1xjauexg, 42qc8u2n, 5cusv2p5, 98ogtog7, af67bxzg, hh6rzjmu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1519.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 195k 1,519 46.7 17.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mwuiy512",
   "source_url": "https://omlx.ai/benchmarks/performance/mwuiy512",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wqjtgpdr-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 831.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 831.8 120.1 21.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wqjtgpdr",
   "source_url": "https://omlx.ai/benchmarks/performance/wqjtgpdr",
   "tg_tokens": null,
   "tps": 120.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wqjtgpdr-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3018.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 3,018 144.0 22.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wqjtgpdr",
   "source_url": "https://omlx.ai/benchmarks/performance/wqjtgpdr",
   "tg_tokens": null,
   "tps": 144.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wqjtgpdr-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5988.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 5,988 47.4 22.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wqjtgpdr",
   "source_url": "https://omlx.ai/benchmarks/performance/wqjtgpdr",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wqjtgpdr-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5530.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 5,530 176.4 22.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wqjtgpdr",
   "source_url": "https://omlx.ai/benchmarks/performance/wqjtgpdr",
   "tg_tokens": null,
   "tps": 176.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wqjtgpdr-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4485.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 4,485 221.0 22.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wqjtgpdr",
   "source_url": "https://omlx.ai/benchmarks/performance/wqjtgpdr",
   "tg_tokens": null,
   "tps": 221.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wqjtgpdr-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3333.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 3,333 142.7 23.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wqjtgpdr",
   "source_url": "https://omlx.ai/benchmarks/performance/wqjtgpdr",
   "tg_tokens": null,
   "tps": 142.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-gbq8jlm9-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 15.3 GB; TTFT 163824 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 100.0,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.8-27B-3bit-mlx on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 3bit; 16k tokens 100.0 PP tok/s 15.4 TG tok/s 163824 TTFT (ms) 15.3 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gbq8jlm9",
   "source_url": "https://omlx.ai/benchmarks/performance/gbq8jlm9",
   "tg_tokens": null,
   "tps": 15.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-7dzm2nxs-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 15.3 GB; TTFT 163974 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 99.9,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.8-27B-3bit-mlx on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 3bit; 16k tokens 99.9 PP tok/s 15.4 TG tok/s 163974 TTFT (ms) 15.3 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7dzm2nxs",
   "source_url": "https://omlx.ai/benchmarks/performance/7dzm2nxs",
   "tg_tokens": null,
   "tps": 15.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-igrt6lcf-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 1oxrncmf, zawmq8kw; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 767.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 767.5 26.2 16.0 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/igrt6lcf",
   "source_url": "https://omlx.ai/benchmarks/performance/igrt6lcf",
   "tg_tokens": null,
   "tps": 26.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-igrt6lcf-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 1oxrncmf, zawmq8kw; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 822.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 822.2 25.6 17.0 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/igrt6lcf",
   "source_url": "https://omlx.ai/benchmarks/performance/igrt6lcf",
   "tg_tokens": null,
   "tps": 25.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-igrt6lcf-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 1oxrncmf, zawmq8kw; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 824.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 824.3 25.2 17.3 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/igrt6lcf",
   "source_url": "https://omlx.ai/benchmarks/performance/igrt6lcf",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-igrt6lcf-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 1oxrncmf, zawmq8kw; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 800.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 800.3 24.0 17.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/igrt6lcf",
   "source_url": "https://omlx.ai/benchmarks/performance/igrt6lcf",
   "tg_tokens": null,
   "tps": 24.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-igrt6lcf-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 1oxrncmf, zawmq8kw; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 748.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 32k 748.0 22.0 18.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/igrt6lcf",
   "source_url": "https://omlx.ai/benchmarks/performance/igrt6lcf",
   "tg_tokens": null,
   "tps": 22.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-igrt6lcf-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 1oxrncmf, zawmq8kw; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 662.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 64k 662.5 20.5 20.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/igrt6lcf",
   "source_url": "https://omlx.ai/benchmarks/performance/igrt6lcf",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-elxnvgn1-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 1k 125.0 8.6 28.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/elxnvgn1",
   "source_url": "https://omlx.ai/benchmarks/performance/elxnvgn1",
   "tg_tokens": null,
   "tps": 8.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-elxnvgn1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 126.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 4k 126.1 8.5 29.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/elxnvgn1",
   "source_url": "https://omlx.ai/benchmarks/performance/elxnvgn1",
   "tg_tokens": null,
   "tps": 8.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-o3wpq415-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 25.1 GB; TTFT 203255 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 644.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k tokens 644.9 PP tok/s 21.7 TG tok/s 203255 TTFT (ms) 25.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o3wpq415",
   "source_url": "https://omlx.ai/benchmarks/performance/o3wpq415",
   "tg_tokens": null,
   "tps": 21.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2tis2g72-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: cfphn1v5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 895.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 895.0 59.6 23.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2tis2g72",
   "source_url": "https://omlx.ai/benchmarks/performance/2tis2g72",
   "tg_tokens": null,
   "tps": 59.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2tis2g72-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: cfphn1v5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 779.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 779.5 45.9 28.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2tis2g72",
   "source_url": "https://omlx.ai/benchmarks/performance/2tis2g72",
   "tg_tokens": null,
   "tps": 45.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2tis2g72-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 37.8 GB at this context; duplicate traces of the same session: cfphn1v5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 570.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 570.9 39.9 37.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2tis2g72",
   "source_url": "https://omlx.ai/benchmarks/performance/2tis2g72",
   "tg_tokens": null,
   "tps": 39.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mev7hi8q-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 821.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 821.6 41.4 17.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mev7hi8q",
   "source_url": "https://omlx.ai/benchmarks/performance/mev7hi8q",
   "tg_tokens": null,
   "tps": 41.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mev7hi8q-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1304.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,304 53.1 19.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mev7hi8q",
   "source_url": "https://omlx.ai/benchmarks/performance/mev7hi8q",
   "tg_tokens": null,
   "tps": 53.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mev7hi8q-ctx8192",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1393.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,393 60.7 19.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mev7hi8q",
   "source_url": "https://omlx.ai/benchmarks/performance/mev7hi8q",
   "tg_tokens": null,
   "tps": 60.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mev7hi8q-ctx16384",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1300.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,300 77.9 19.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mev7hi8q",
   "source_url": "https://omlx.ai/benchmarks/performance/mev7hi8q",
   "tg_tokens": null,
   "tps": 77.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mev7hi8q-ctx32768",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1146.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,146 91.9 21.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mev7hi8q",
   "source_url": "https://omlx.ai/benchmarks/performance/mev7hi8q",
   "tg_tokens": null,
   "tps": 91.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mev7hi8q-ctx65536",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 963.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 963.9 48.8 23.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mev7hi8q",
   "source_url": "https://omlx.ai/benchmarks/performance/mev7hi8q",
   "tg_tokens": null,
   "tps": 48.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ajbkm7xp-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-NEO-CODER",
   "notes": "1M context build; oMLX community benchmark trace; peak memory 28.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 255.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-NEO-CODER-mlx-oQ8e-1M on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 255.0 16.6 28.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ajbkm7xp",
   "source_url": "https://omlx.ai/benchmarks/performance/ajbkm7xp",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ajbkm7xp-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-NEO-CODER",
   "notes": "1M context build; oMLX community benchmark trace; peak memory 31.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 262.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-NEO-CODER-mlx-oQ8e-1M on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 262.6 16.5 31.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ajbkm7xp",
   "source_url": "https://omlx.ai/benchmarks/performance/ajbkm7xp",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-f6irs7xx-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 324.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 324.7 70.0 21.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f6irs7xx",
   "source_url": "https://omlx.ai/benchmarks/performance/f6irs7xx",
   "tg_tokens": null,
   "tps": 70.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-f6irs7xx-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 431.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 431.7 66.8 24.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f6irs7xx",
   "source_url": "https://omlx.ai/benchmarks/performance/f6irs7xx",
   "tg_tokens": null,
   "tps": 66.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-vg160xma-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 323.5 40.3 21.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vg160xma",
   "source_url": "https://omlx.ai/benchmarks/performance/vg160xma",
   "tg_tokens": null,
   "tps": 40.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-vg160xma-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 438.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 438.2 50.8 24.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vg160xma",
   "source_url": "https://omlx.ai/benchmarks/performance/vg160xma",
   "tg_tokens": null,
   "tps": 50.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-coy0uff6-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 333.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 333.2 36.9 16.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/coy0uff6",
   "source_url": "https://omlx.ai/benchmarks/performance/coy0uff6",
   "tg_tokens": null,
   "tps": 36.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-coy0uff6-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 356.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 356.3 35.9 19.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/coy0uff6",
   "source_url": "https://omlx.ai/benchmarks/performance/coy0uff6",
   "tg_tokens": null,
   "tps": 35.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-hkwm4aw2-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 44.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 618.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 618.0 49.8 44.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hkwm4aw2",
   "source_url": "https://omlx.ai/benchmarks/performance/hkwm4aw2",
   "tg_tokens": null,
   "tps": 49.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-hkwm4aw2-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 45.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 759.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 759.2 43.6 45.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hkwm4aw2",
   "source_url": "https://omlx.ai/benchmarks/performance/hkwm4aw2",
   "tg_tokens": null,
   "tps": 43.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-x5fcn2gk-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 44.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 507.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 507.6 40.1 44.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x5fcn2gk",
   "source_url": "https://omlx.ai/benchmarks/performance/x5fcn2gk",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-x5fcn2gk-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 45.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 671.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 671.1 37.6 45.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x5fcn2gk",
   "source_url": "https://omlx.ai/benchmarks/performance/x5fcn2gk",
   "tg_tokens": null,
   "tps": 37.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1yp0hkd9-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.4 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1073.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,073 104.1 100.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1yp0hkd9",
   "source_url": "https://omlx.ai/benchmarks/performance/1yp0hkd9",
   "tg_tokens": null,
   "tps": 104.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1yp0hkd9-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1379.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,379 95.7 101.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1yp0hkd9",
   "source_url": "https://omlx.ai/benchmarks/performance/1yp0hkd9",
   "tg_tokens": null,
   "tps": 95.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1km0scoe-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.4 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1073.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,073 92.7 100.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1km0scoe",
   "source_url": "https://omlx.ai/benchmarks/performance/1km0scoe",
   "tg_tokens": null,
   "tps": 92.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1km0scoe-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1379.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,379 95.2 101.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1km0scoe",
   "source_url": "https://omlx.ai/benchmarks/performance/1km0scoe",
   "tg_tokens": null,
   "tps": 95.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2oo3v27w-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.1 GB at this context; duplicate traces of the same session: k9li9eh0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1107.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 1,107 84.3 183.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2oo3v27w",
   "source_url": "https://omlx.ai/benchmarks/performance/2oo3v27w",
   "tg_tokens": null,
   "tps": 84.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2oo3v27w-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.6 GB at this context; duplicate traces of the same session: k9li9eh0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1380.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 1,380 66.6 183.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2oo3v27w",
   "source_url": "https://omlx.ai/benchmarks/performance/2oo3v27w",
   "tg_tokens": null,
   "tps": 66.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2oo3v27w-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.8 GB at this context; duplicate traces of the same session: k9li9eh0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1364.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 1,364 88.3 183.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2oo3v27w",
   "source_url": "https://omlx.ai/benchmarks/performance/2oo3v27w",
   "tg_tokens": null,
   "tps": 88.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-73kxxib4-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 14.7 GB at this context; duplicate traces of the same session: 3lupzror; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1177.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 1,177 79.5 14.7 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/73kxxib4",
   "source_url": "https://omlx.ai/benchmarks/performance/73kxxib4",
   "tg_tokens": null,
   "tps": 79.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-73kxxib4-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: 3lupzror; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1109.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 1,109 72.2 15.8 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/73kxxib4",
   "source_url": "https://omlx.ai/benchmarks/performance/73kxxib4",
   "tg_tokens": null,
   "tps": 72.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-qr0q80y7-ctx4096",
   "model": "gemma-4-E4B-it",
   "notes": "oMLX community benchmark trace; peak memory 8.1 GB at this context; duplicate traces of the same session: xmn0lx83; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1989.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-e4b-it on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 8bit; 4k 1,989 56.3 8.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qr0q80y7",
   "source_url": "https://omlx.ai/benchmarks/performance/qr0q80y7",
   "tg_tokens": null,
   "tps": 56.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-qr0q80y7-ctx16384",
   "model": "gemma-4-E4B-it",
   "notes": "oMLX community benchmark trace; peak memory 8.5 GB at this context; duplicate traces of the same session: xmn0lx83; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1877.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-e4b-it on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 8bit; 16k 1,877 50.6 8.5 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qr0q80y7",
   "source_url": "https://omlx.ai/benchmarks/performance/qr0q80y7",
   "tg_tokens": null,
   "tps": 50.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0sgvr5et-ctx131072",
   "model": "gpt-oss 120B",
   "notes": "bf16 KV cache; oMLX community benchmark trace, single run at 128k context; peak memory 65.6 GB; TTFT 99415 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1318.0,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "gpt-oss-120b-mxfp4-bf16 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k tokens 1,318 PP tok/s 34.1 TG tok/s 99415 TTFT (ms) 65.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0sgvr5et",
   "source_url": "https://omlx.ai/benchmarks/performance/0sgvr5et",
   "tg_tokens": null,
   "tps": 34.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fi2gpfnk-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: elmfo951; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 892.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 892.8 47.4 19.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fi2gpfnk",
   "source_url": "https://omlx.ai/benchmarks/performance/fi2gpfnk",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fi2gpfnk-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: elmfo951; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1002.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,002 54.8 21.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fi2gpfnk",
   "source_url": "https://omlx.ai/benchmarks/performance/fi2gpfnk",
   "tg_tokens": null,
   "tps": 54.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9ccc7am9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: htb3nvcs; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 958.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 958.2 39.3 19.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ccc7am9",
   "source_url": "https://omlx.ai/benchmarks/performance/9ccc7am9",
   "tg_tokens": null,
   "tps": 39.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9ccc7am9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: htb3nvcs; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 999.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 999.0 31.2 21.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ccc7am9",
   "source_url": "https://omlx.ai/benchmarks/performance/9ccc7am9",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-3gt6ddkk-ctx4096",
   "model": "ALoDLM-8B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 17.0 GB; TTFT 2329 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1758.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "ALoDLM-8B on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 unknown; 4k tokens 1,758 PP tok/s 17.8 TG tok/s 2329 TTFT (ms) 17.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3gt6ddkk",
   "source_url": "https://omlx.ai/benchmarks/performance/3gt6ddkk",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-snr0oaze-ctx4096",
   "model": "ALoDLM-8B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 17.0 GB; TTFT 2542 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1612.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "ALoDLM-8B on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 unknown; 4k tokens 1,612 PP tok/s 18.0 TG tok/s 2542 TTFT (ms) 17.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/snr0oaze",
   "source_url": "https://omlx.ai/benchmarks/performance/snr0oaze",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-adyhuux6-ctx1024",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 81.9,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 5bit; 1k 81.9 10.9 19.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/adyhuux6",
   "source_url": "https://omlx.ai/benchmarks/performance/adyhuux6",
   "tg_tokens": null,
   "tps": 10.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-adyhuux6-ctx4096",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 76.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 5bit; 4k 76.0 10.1 20.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/adyhuux6",
   "source_url": "https://omlx.ai/benchmarks/performance/adyhuux6",
   "tg_tokens": null,
   "tps": 10.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-adyhuux6-ctx8192",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 74.3,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 5bit; 8k 74.3 9.9 21.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/adyhuux6",
   "source_url": "https://omlx.ai/benchmarks/performance/adyhuux6",
   "tg_tokens": null,
   "tps": 9.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-adyhuux6-ctx16384",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 70.4,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 5bit; 16k 70.4 9.1 21.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/adyhuux6",
   "source_url": "https://omlx.ai/benchmarks/performance/adyhuux6",
   "tg_tokens": null,
   "tps": 9.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-k51hgfi7-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 316.2 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 430.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "GLM-5.3-Flash-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 430.3 32.4 316.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k51hgfi7",
   "source_url": "https://omlx.ai/benchmarks/performance/k51hgfi7",
   "tg_tokens": null,
   "tps": 32.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-k51hgfi7-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 317.3 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 551.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "GLM-5.3-Flash-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 551.2 31.1 317.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k51hgfi7",
   "source_url": "https://omlx.ai/benchmarks/performance/k51hgfi7",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-k51hgfi7-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 317.6 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 513.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "GLM-5.3-Flash-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 513.3 31.2 317.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k51hgfi7",
   "source_url": "https://omlx.ai/benchmarks/performance/k51hgfi7",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-k51hgfi7-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 317.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 545.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "GLM-5.3-Flash-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k 545.4 30.9 317.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k51hgfi7",
   "source_url": "https://omlx.ai/benchmarks/performance/k51hgfi7",
   "tg_tokens": null,
   "tps": 30.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-g5h081v1-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 240.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 240.8 65.4 18.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g5h081v1",
   "source_url": "https://omlx.ai/benchmarks/performance/g5h081v1",
   "tg_tokens": null,
   "tps": 65.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-g5h081v1-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 240.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 240.4 76.8 23.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g5h081v1",
   "source_url": "https://omlx.ai/benchmarks/performance/g5h081v1",
   "tg_tokens": null,
   "tps": 76.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-c45yoxsp-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 244.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 244.2 25.0 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c45yoxsp",
   "source_url": "https://omlx.ai/benchmarks/performance/c45yoxsp",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-c45yoxsp-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 240.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 240.6 21.1 21.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c45yoxsp",
   "source_url": "https://omlx.ai/benchmarks/performance/c45yoxsp",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-5zulu72e-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 234.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 234.7 56.2 18.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5zulu72e",
   "source_url": "https://omlx.ai/benchmarks/performance/5zulu72e",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-5zulu72e-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 235.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 235.1 60.0 23.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5zulu72e",
   "source_url": "https://omlx.ai/benchmarks/performance/5zulu72e",
   "tg_tokens": null,
   "tps": 60.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-uezdsol8-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 238.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 238.6 65.0 18.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uezdsol8",
   "source_url": "https://omlx.ai/benchmarks/performance/uezdsol8",
   "tg_tokens": null,
   "tps": 65.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-uezdsol8-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 234.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 234.8 67.2 23.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uezdsol8",
   "source_url": "https://omlx.ai/benchmarks/performance/uezdsol8",
   "tg_tokens": null,
   "tps": 67.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-f10yf2zw-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 234.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 234.8 51.4 18.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f10yf2zw",
   "source_url": "https://omlx.ai/benchmarks/performance/f10yf2zw",
   "tg_tokens": null,
   "tps": 51.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-f10yf2zw-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 234.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 234.1 47.4 23.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f10yf2zw",
   "source_url": "https://omlx.ai/benchmarks/performance/f10yf2zw",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-hc6a8us9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: lfmqcryn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 240.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 240.9 26.0 16.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hc6a8us9",
   "source_url": "https://omlx.ai/benchmarks/performance/hc6a8us9",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-hc6a8us9-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: lfmqcryn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 240.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 240.1 23.8 21.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hc6a8us9",
   "source_url": "https://omlx.ai/benchmarks/performance/hc6a8us9",
   "tg_tokens": null,
   "tps": 23.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-6720ycsm-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: rxbelhkp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 225.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 225.9 30.6 18.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6720ycsm",
   "source_url": "https://omlx.ai/benchmarks/performance/6720ycsm",
   "tg_tokens": null,
   "tps": 30.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-6720ycsm-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: rxbelhkp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 234.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 234.8 58.0 23.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6720ycsm",
   "source_url": "https://omlx.ai/benchmarks/performance/6720ycsm",
   "tg_tokens": null,
   "tps": 58.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-qaxo3wns-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 16.6 GB; TTFT 4267 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 240.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 240.0 PP tok/s 26.4 TG tok/s 4267 TTFT (ms) 16.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qaxo3wns",
   "source_url": "https://omlx.ai/benchmarks/performance/qaxo3wns",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-wioz32s2-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 9m40oo1u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 251.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 251.9 26.6 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wioz32s2",
   "source_url": "https://omlx.ai/benchmarks/performance/wioz32s2",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-wioz32s2-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 9m40oo1u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 242.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 242.9 23.7 21.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wioz32s2",
   "source_url": "https://omlx.ai/benchmarks/performance/wioz32s2",
   "tg_tokens": null,
   "tps": 23.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-i4413e2f-ctx1024",
   "model": "Ling-3.0-flash",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 68.3 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 594.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-flash-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 594.8 41.4 68.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i4413e2f",
   "source_url": "https://omlx.ai/benchmarks/performance/i4413e2f",
   "tg_tokens": null,
   "tps": 41.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-i4413e2f-ctx4096",
   "model": "Ling-3.0-flash",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 69.5 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 525.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-flash-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 525.3 37.9 69.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i4413e2f",
   "source_url": "https://omlx.ai/benchmarks/performance/i4413e2f",
   "tg_tokens": null,
   "tps": 37.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-i4413e2f-ctx8192",
   "model": "Ling-3.0-flash",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 70.2 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 505.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-flash-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 505.0 37.0 70.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i4413e2f",
   "source_url": "https://omlx.ai/benchmarks/performance/i4413e2f",
   "tg_tokens": null,
   "tps": 37.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-i4413e2f-ctx16384",
   "model": "Ling-3.0-flash",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 476.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-flash-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 476.4 33.0 71.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i4413e2f",
   "source_url": "https://omlx.ai/benchmarks/performance/i4413e2f",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-i4413e2f-ctx32768",
   "model": "Ling-3.0-flash",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 74.7 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 427.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-flash-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 32k 427.5 32.6 74.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i4413e2f",
   "source_url": "https://omlx.ai/benchmarks/performance/i4413e2f",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-xcvqkety-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Nail build; oMLX community benchmark trace; peak memory 33.2 GB at this context; duplicate traces of the same session: 5ocvsi80; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 581.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX-6bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 1k 581.9 43.3 33.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xcvqkety",
   "source_url": "https://omlx.ai/benchmarks/performance/xcvqkety",
   "tg_tokens": null,
   "tps": 43.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-xcvqkety-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Nail build; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: 5ocvsi80; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 599.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX-6bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 4k 599.7 37.1 34.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xcvqkety",
   "source_url": "https://omlx.ai/benchmarks/performance/xcvqkety",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-z1lnzn98-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Nail build; oMLX community benchmark trace; peak memory 33.1 GB at this context; duplicate traces of the same session: 2ny51c0u; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 579.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX-6bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 1k 579.0 46.7 33.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z1lnzn98",
   "source_url": "https://omlx.ai/benchmarks/performance/z1lnzn98",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-z1lnzn98-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Nail build; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: 2ny51c0u; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 608.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX-6bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 4k 608.7 39.1 35.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z1lnzn98",
   "source_url": "https://omlx.ai/benchmarks/performance/z1lnzn98",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-2mzj2yfo-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Nail build; oMLX community benchmark trace; peak memory 30.9 GB at this context; duplicate traces of the same session: c80xamrd; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 578.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX-6bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 1k 578.9 34.4 30.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2mzj2yfo",
   "source_url": "https://omlx.ai/benchmarks/performance/2mzj2yfo",
   "tg_tokens": null,
   "tps": 34.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-2mzj2yfo-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Nail build; oMLX community benchmark trace; peak memory 31.9 GB at this context; duplicate traces of the same session: c80xamrd; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 568.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX-6bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 4k 568.1 34.6 31.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2mzj2yfo",
   "source_url": "https://omlx.ai/benchmarks/performance/2mzj2yfo",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-bi22wmb2-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Nail build; oMLX community benchmark trace; peak memory 33.2 GB at this context; duplicate traces of the same session: 3nbc4e5m; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 578.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX-6bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 1k 578.0 46.0 33.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bi22wmb2",
   "source_url": "https://omlx.ai/benchmarks/performance/bi22wmb2",
   "tg_tokens": null,
   "tps": 46.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-bi22wmb2-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Nail build; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: 3nbc4e5m; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 594.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX-6bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 4k 594.7 39.2 34.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bi22wmb2",
   "source_url": "https://omlx.ai/benchmarks/performance/bi22wmb2",
   "tg_tokens": null,
   "tps": 39.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-q5w3s21a-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Nail build; oMLX community benchmark trace; peak memory 33.1 GB at this context; duplicate traces of the same session: l1acqnhm; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 547.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX-6bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 1k 547.6 32.5 33.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q5w3s21a",
   "source_url": "https://omlx.ai/benchmarks/performance/q5w3s21a",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-q5w3s21a-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Nail build; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: l1acqnhm; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 611.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX-6bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 4k 611.7 38.5 35.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q5w3s21a",
   "source_url": "https://omlx.ai/benchmarks/performance/q5w3s21a",
   "tg_tokens": null,
   "tps": 38.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-33wwatx3-ctx1024",
   "model": "Occamy-1.0",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1386.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Occamy-1.0-oQ8-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,386 68.9 35.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/33wwatx3",
   "source_url": "https://omlx.ai/benchmarks/performance/33wwatx3",
   "tg_tokens": null,
   "tps": 68.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-33wwatx3-ctx4096",
   "model": "Occamy-1.0",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1296.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Occamy-1.0-oQ8-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 1,296 57.3 37.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/33wwatx3",
   "source_url": "https://omlx.ai/benchmarks/performance/33wwatx3",
   "tg_tokens": null,
   "tps": 57.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-33wwatx3-ctx8192",
   "model": "Occamy-1.0",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 37.9 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1026.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Occamy-1.0-oQ8-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 8k 1,026 58.8 37.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/33wwatx3",
   "source_url": "https://omlx.ai/benchmarks/performance/33wwatx3",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-33wwatx3-ctx16384",
   "model": "Occamy-1.0",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 38.7 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 960.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Occamy-1.0-oQ8-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 16k 960.5 50.3 38.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/33wwatx3",
   "source_url": "https://omlx.ai/benchmarks/performance/33wwatx3",
   "tg_tokens": null,
   "tps": 50.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-33wwatx3-ctx32768",
   "model": "Occamy-1.0",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 40.1 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 792.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Occamy-1.0-oQ8-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 32k 792.0 49.6 40.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/33wwatx3",
   "source_url": "https://omlx.ai/benchmarks/performance/33wwatx3",
   "tg_tokens": null,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-33wwatx3-ctx65536",
   "model": "Occamy-1.0",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 41.2 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 590.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Occamy-1.0-oQ8-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 64k 590.2 39.0 41.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/33wwatx3",
   "source_url": "https://omlx.ai/benchmarks/performance/33wwatx3",
   "tg_tokens": null,
   "tps": 39.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-9tpknegx-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 807.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 807.7 85.6 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9tpknegx",
   "source_url": "https://omlx.ai/benchmarks/performance/9tpknegx",
   "tg_tokens": null,
   "tps": 85.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-9tpknegx-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1012.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,012 82.1 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9tpknegx",
   "source_url": "https://omlx.ai/benchmarks/performance/9tpknegx",
   "tg_tokens": null,
   "tps": 82.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-9tpknegx-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 961.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 961.9 78.9 21.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9tpknegx",
   "source_url": "https://omlx.ai/benchmarks/performance/9tpknegx",
   "tg_tokens": null,
   "tps": 78.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-9tpknegx-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 806.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 806.6 70.8 22.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9tpknegx",
   "source_url": "https://omlx.ai/benchmarks/performance/9tpknegx",
   "tg_tokens": null,
   "tps": 70.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-f7w558be-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1064.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 16 GB \u00b7 4bit; 4k 1,064 26.5 6.3 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f7w558be",
   "source_url": "https://omlx.ai/benchmarks/performance/f7w558be",
   "tg_tokens": null,
   "tps": 26.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-f7w558be-ctx65536",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.1 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 697.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 16 GB \u00b7 4bit; 64k 697.1 16.8 8.1 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f7w558be",
   "source_url": "https://omlx.ai/benchmarks/performance/f7w558be",
   "tg_tokens": null,
   "tps": 16.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-vhmok9px-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.9 GB at this context; duplicate traces of the same session: vywlavbb; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 378.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 6bit; 1k 378.1 44.0 7.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vhmok9px",
   "source_url": "https://omlx.ai/benchmarks/performance/vhmok9px",
   "tg_tokens": null,
   "tps": 44.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-vhmok9px-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.4 GB at this context; duplicate traces of the same session: vywlavbb; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 368.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 6bit; 4k 368.9 42.6 8.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vhmok9px",
   "source_url": "https://omlx.ai/benchmarks/performance/vhmok9px",
   "tg_tokens": null,
   "tps": 42.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-z4mturt9-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.9 GB at this context; duplicate traces of the same session: 8js3mgdr; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 372.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 6bit; 1k 372.4 43.7 7.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z4mturt9",
   "source_url": "https://omlx.ai/benchmarks/performance/z4mturt9",
   "tg_tokens": null,
   "tps": 43.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-z4mturt9-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.4 GB at this context; duplicate traces of the same session: 8js3mgdr; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 365.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 6bit; 4k 365.0 42.4 8.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z4mturt9",
   "source_url": "https://omlx.ai/benchmarks/performance/z4mturt9",
   "tg_tokens": null,
   "tps": 42.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-p7al850j-ctx4096",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 2.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1958.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.5-0.8B-OptiQ-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 1,958 152.9 2.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p7al850j",
   "source_url": "https://omlx.ai/benchmarks/performance/p7al850j",
   "tg_tokens": null,
   "tps": 152.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-p7al850j-ctx16384",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 2.2 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1406.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.5-0.8B-OptiQ-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k 1,406 126.9 2.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p7al850j",
   "source_url": "https://omlx.ai/benchmarks/performance/p7al850j",
   "tg_tokens": null,
   "tps": 126.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-z2gx7sam-ctx4096",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2132.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 2,132 68.4 19.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z2gx7sam",
   "source_url": "https://omlx.ai/benchmarks/performance/z2gx7sam",
   "tg_tokens": null,
   "tps": 68.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-z2gx7sam-ctx16384",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1920.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 1,920 61.5 20.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z2gx7sam",
   "source_url": "https://omlx.ai/benchmarks/performance/z2gx7sam",
   "tg_tokens": null,
   "tps": 61.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-8pw774p6-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.8 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1832.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,832 47.4 3.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8pw774p6",
   "source_url": "https://omlx.ai/benchmarks/performance/8pw774p6",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-8pw774p6-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.1 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1759.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 1,759 46.4 4.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8pw774p6",
   "source_url": "https://omlx.ai/benchmarks/performance/8pw774p6",
   "tg_tokens": null,
   "tps": 46.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-t4t83rez-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1117.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,117 30.6 6.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t4t83rez",
   "source_url": "https://omlx.ai/benchmarks/performance/t4t83rez",
   "tg_tokens": null,
   "tps": 30.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-t4t83rez-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.6 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1083.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 1,083 28.5 6.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t4t83rez",
   "source_url": "https://omlx.ai/benchmarks/performance/t4t83rez",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-uc0kd8rf-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: ibvs1znv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2731.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,731 107.5 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uc0kd8rf",
   "source_url": "https://omlx.ai/benchmarks/performance/uc0kd8rf",
   "tg_tokens": null,
   "tps": 107.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-uc0kd8rf-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: ibvs1znv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3412.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,412 104.9 20.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uc0kd8rf",
   "source_url": "https://omlx.ai/benchmarks/performance/uc0kd8rf",
   "tg_tokens": null,
   "tps": 104.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-uc0kd8rf-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: ibvs1znv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3042.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 3,042 94.0 20.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uc0kd8rf",
   "source_url": "https://omlx.ai/benchmarks/performance/uc0kd8rf",
   "tg_tokens": null,
   "tps": 94.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-wkisz01b-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 2553hw4h, r65s1x3v; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 808.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 808.4 98.0 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wkisz01b",
   "source_url": "https://omlx.ai/benchmarks/performance/wkisz01b",
   "tg_tokens": null,
   "tps": 98.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-wkisz01b-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 2553hw4h, r65s1x3v; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1877.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,877 99.0 20.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wkisz01b",
   "source_url": "https://omlx.ai/benchmarks/performance/wkisz01b",
   "tg_tokens": null,
   "tps": 99.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-wkisz01b-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: 2553hw4h, r65s1x3v; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2970.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,970 78.9 20.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wkisz01b",
   "source_url": "https://omlx.ai/benchmarks/performance/wkisz01b",
   "tg_tokens": null,
   "tps": 78.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-4o60o389-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 897mgrbk; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 935.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 1k 935.7 82.6 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4o60o389",
   "source_url": "https://omlx.ai/benchmarks/performance/4o60o389",
   "tg_tokens": null,
   "tps": 82.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-4o60o389-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 897mgrbk; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 973.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k 973.2 84.5 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4o60o389",
   "source_url": "https://omlx.ai/benchmarks/performance/4o60o389",
   "tg_tokens": null,
   "tps": 84.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-4o60o389-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: 897mgrbk; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 800.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 8k 800.8 78.2 21.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4o60o389",
   "source_url": "https://omlx.ai/benchmarks/performance/4o60o389",
   "tg_tokens": null,
   "tps": 78.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-4o60o389-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 897mgrbk; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 665.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 16k 665.2 68.9 22.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4o60o389",
   "source_url": "https://omlx.ai/benchmarks/performance/4o60o389",
   "tg_tokens": null,
   "tps": 68.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6x5dd1b0-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3830.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 3,830 103.4 21.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6x5dd1b0",
   "source_url": "https://omlx.ai/benchmarks/performance/6x5dd1b0",
   "tg_tokens": null,
   "tps": 103.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-6x5dd1b0-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2577.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,577 79.5 22.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6x5dd1b0",
   "source_url": "https://omlx.ai/benchmarks/performance/6x5dd1b0",
   "tg_tokens": null,
   "tps": 79.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5cusv2p5-ctx1024",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: 1xjauexg, 42qc8u2n, 98ogtog7, af67bxzg, hh6rzjmu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 422.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 422.5 110.9 12.7 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5cusv2p5",
   "source_url": "https://omlx.ai/benchmarks/performance/5cusv2p5",
   "tg_tokens": null,
   "tps": 110.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5cusv2p5-ctx8192",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: 1xjauexg, 42qc8u2n, 98ogtog7, af67bxzg, hh6rzjmu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4911.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 4,911 104.3 13.4 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5cusv2p5",
   "source_url": "https://omlx.ai/benchmarks/performance/5cusv2p5",
   "tg_tokens": null,
   "tps": 104.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5cusv2p5-ctx16384",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.6 GB at this context; duplicate traces of the same session: 1xjauexg, 42qc8u2n, 98ogtog7, af67bxzg, hh6rzjmu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4519.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 4,519 99.4 13.6 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5cusv2p5",
   "source_url": "https://omlx.ai/benchmarks/performance/5cusv2p5",
   "tg_tokens": null,
   "tps": 99.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5cusv2p5-ctx32768",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.9 GB at this context; duplicate traces of the same session: 1xjauexg, 42qc8u2n, 98ogtog7, af67bxzg, hh6rzjmu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3947.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 32k 3,947 89.5 13.9 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5cusv2p5",
   "source_url": "https://omlx.ai/benchmarks/performance/5cusv2p5",
   "tg_tokens": null,
   "tps": 89.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5cusv2p5-ctx65536",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 14.5 GB at this context; duplicate traces of the same session: 1xjauexg, 42qc8u2n, 98ogtog7, af67bxzg, hh6rzjmu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3024.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 64k 3,024 74.8 14.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5cusv2p5",
   "source_url": "https://omlx.ai/benchmarks/performance/5cusv2p5",
   "tg_tokens": null,
   "tps": 74.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5cusv2p5-ctx131072",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: 1xjauexg, 42qc8u2n, 98ogtog7, af67bxzg, hh6rzjmu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2041.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 128k 2,041 57.7 15.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5cusv2p5",
   "source_url": "https://omlx.ai/benchmarks/performance/5cusv2p5",
   "tg_tokens": null,
   "tps": 57.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5cusv2p5-ctx199680",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: 1xjauexg, 42qc8u2n, 98ogtog7, af67bxzg, hh6rzjmu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1519.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 195k 1,519 46.7 17.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5cusv2p5",
   "source_url": "https://omlx.ai/benchmarks/performance/5cusv2p5",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-u8ikj7j6-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 14.2 GB; TTFT 38450 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.5,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.8-27B-3bit-mlx on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 3bit; 4k tokens 106.5 PP tok/s 17.0 TG tok/s 38450 TTFT (ms) 14.2 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u8ikj7j6",
   "source_url": "https://omlx.ai/benchmarks/performance/u8ikj7j6",
   "tg_tokens": null,
   "tps": 17.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ig395rhy-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 14.2 GB; TTFT 38441 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.6,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.8-27B-3bit-mlx on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 3bit; 4k tokens 106.6 PP tok/s 17.0 TG tok/s 38441 TTFT (ms) 14.2 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ig395rhy",
   "source_url": "https://omlx.ai/benchmarks/performance/ig395rhy",
   "tg_tokens": null,
   "tps": 17.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1oxrncmf-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: zawmq8kw; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 767.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 767.5 26.2 16.0 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1oxrncmf",
   "source_url": "https://omlx.ai/benchmarks/performance/1oxrncmf",
   "tg_tokens": null,
   "tps": 26.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1oxrncmf-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: zawmq8kw; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 822.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 822.2 25.6 17.0 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1oxrncmf",
   "source_url": "https://omlx.ai/benchmarks/performance/1oxrncmf",
   "tg_tokens": null,
   "tps": 25.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1oxrncmf-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: zawmq8kw; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 824.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 824.3 25.2 17.3 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1oxrncmf",
   "source_url": "https://omlx.ai/benchmarks/performance/1oxrncmf",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1oxrncmf-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: zawmq8kw; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 800.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 800.3 24.0 17.8 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1oxrncmf",
   "source_url": "https://omlx.ai/benchmarks/performance/1oxrncmf",
   "tg_tokens": null,
   "tps": 24.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1oxrncmf-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: zawmq8kw; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 748.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 32k 748.0 22.0 18.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1oxrncmf",
   "source_url": "https://omlx.ai/benchmarks/performance/1oxrncmf",
   "tg_tokens": null,
   "tps": 22.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1oxrncmf-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: zawmq8kw; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 662.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 64k 662.5 20.5 20.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1oxrncmf",
   "source_url": "https://omlx.ai/benchmarks/performance/1oxrncmf",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cfphn1v5-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 895.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 895.0 59.6 23.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cfphn1v5",
   "source_url": "https://omlx.ai/benchmarks/performance/cfphn1v5",
   "tg_tokens": null,
   "tps": 59.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cfphn1v5-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 779.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 779.5 45.9 28.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cfphn1v5",
   "source_url": "https://omlx.ai/benchmarks/performance/cfphn1v5",
   "tg_tokens": null,
   "tps": 45.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cfphn1v5-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 37.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 570.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 570.9 39.9 37.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cfphn1v5",
   "source_url": "https://omlx.ai/benchmarks/performance/cfphn1v5",
   "tg_tokens": null,
   "tps": 39.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-bsgl5oh1-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 4a33t77w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 219.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 219.4 12.7 28.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bsgl5oh1",
   "source_url": "https://omlx.ai/benchmarks/performance/bsgl5oh1",
   "tg_tokens": null,
   "tps": 12.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-bsgl5oh1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.7 GB at this context; duplicate traces of the same session: 4a33t77w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 168.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 168.6 11.9 31.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bsgl5oh1",
   "source_url": "https://omlx.ai/benchmarks/performance/bsgl5oh1",
   "tg_tokens": null,
   "tps": 11.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-gg4vjowj-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 57.2 GB at this context; duplicate traces of the same session: 468m3y2m; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 300.5,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.8-27B-bf16 on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 bf16; 16k 300.5 11.5 57.2 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gg4vjowj",
   "source_url": "https://omlx.ai/benchmarks/performance/gg4vjowj",
   "tg_tokens": null,
   "tps": 11.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-gg4vjowj-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 64.7 GB at this context; duplicate traces of the same session: 468m3y2m; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 159.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.8-27B-bf16 on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 bf16; 128k 159.0 9.7 64.7 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gg4vjowj",
   "source_url": "https://omlx.ai/benchmarks/performance/gg4vjowj",
   "tg_tokens": null,
   "tps": 9.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-myvqox5m-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 23.5 GB; TTFT 41794 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 784.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k tokens 784.0 PP tok/s 42.6 TG tok/s 41794 TTFT (ms) 23.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/myvqox5m",
   "source_url": "https://omlx.ai/benchmarks/performance/myvqox5m",
   "tg_tokens": null,
   "tps": 42.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-uqvmfc6j-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 812.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 812.1 53.6 20.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uqvmfc6j",
   "source_url": "https://omlx.ai/benchmarks/performance/uqvmfc6j",
   "tg_tokens": null,
   "tps": 53.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-uqvmfc6j-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 910.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 910.1 50.7 21.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uqvmfc6j",
   "source_url": "https://omlx.ai/benchmarks/performance/uqvmfc6j",
   "tg_tokens": null,
   "tps": 50.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-k9li9eh0-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1107.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 1,107 84.3 183.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k9li9eh0",
   "source_url": "https://omlx.ai/benchmarks/performance/k9li9eh0",
   "tg_tokens": null,
   "tps": 84.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-k9li9eh0-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.6 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1380.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 1,380 66.6 183.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k9li9eh0",
   "source_url": "https://omlx.ai/benchmarks/performance/k9li9eh0",
   "tg_tokens": null,
   "tps": 66.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-k9li9eh0-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.8 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1364.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 1,364 88.3 183.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k9li9eh0",
   "source_url": "https://omlx.ai/benchmarks/performance/k9li9eh0",
   "tg_tokens": null,
   "tps": 88.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-bihmpoeo-ctx4096",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 13.8 GB at this context; duplicate traces of the same session: gb80nyce; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 194.3,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 2bit; 4k 194.3 13.5 13.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bihmpoeo",
   "source_url": "https://omlx.ai/benchmarks/performance/bihmpoeo",
   "tg_tokens": null,
   "tps": 13.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-bihmpoeo-ctx16384",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 15.2 GB at this context; duplicate traces of the same session: gb80nyce; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 159.7,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 2bit; 16k 159.7 10.8 15.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bihmpoeo",
   "source_url": "https://omlx.ai/benchmarks/performance/bihmpoeo",
   "tg_tokens": null,
   "tps": 10.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-3lupzror-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 14.7 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1177.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 1,177 79.5 14.7 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3lupzror",
   "source_url": "https://omlx.ai/benchmarks/performance/3lupzror",
   "tg_tokens": null,
   "tps": 79.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-3lupzror-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1109.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 1,109 72.2 15.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3lupzror",
   "source_url": "https://omlx.ai/benchmarks/performance/3lupzror",
   "tg_tokens": null,
   "tps": 72.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-xmn0lx83-ctx4096",
   "model": "gemma-4-E4B-it",
   "notes": "oMLX community benchmark trace; peak memory 8.1 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1989.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-e4b-it on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 8bit; 4k 1,989 56.3 8.1 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xmn0lx83",
   "source_url": "https://omlx.ai/benchmarks/performance/xmn0lx83",
   "tg_tokens": null,
   "tps": 56.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-xmn0lx83-ctx16384",
   "model": "gemma-4-E4B-it",
   "notes": "oMLX community benchmark trace; peak memory 8.5 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1877.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-e4b-it on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 8bit; 16k 1,877 50.6 8.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xmn0lx83",
   "source_url": "https://omlx.ai/benchmarks/performance/xmn0lx83",
   "tg_tokens": null,
   "tps": 50.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8x6o9m84-ctx1024",
   "model": "gpt-oss 120B",
   "notes": "bf16 KV cache; oMLX community benchmark trace; peak memory 61.2 GB at this context; duplicate traces of the same session: isojhkdv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 727.7,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "gpt-oss-120b-mxfp4-bf16 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 727.7 81.2 61.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8x6o9m84",
   "source_url": "https://omlx.ai/benchmarks/performance/8x6o9m84",
   "tg_tokens": null,
   "tps": 81.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8x6o9m84-ctx4096",
   "model": "gpt-oss 120B",
   "notes": "bf16 KV cache; oMLX community benchmark trace; peak memory 61.3 GB at this context; duplicate traces of the same session: isojhkdv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2110.0,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "gpt-oss-120b-mxfp4-bf16 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,110 78.5 61.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8x6o9m84",
   "source_url": "https://omlx.ai/benchmarks/performance/8x6o9m84",
   "tg_tokens": null,
   "tps": 78.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-elmfo951-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 892.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 892.8 47.4 19.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/elmfo951",
   "source_url": "https://omlx.ai/benchmarks/performance/elmfo951",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-elmfo951-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1002.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,002 54.8 21.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/elmfo951",
   "source_url": "https://omlx.ai/benchmarks/performance/elmfo951",
   "tg_tokens": null,
   "tps": 54.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-htb3nvcs-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 958.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 958.2 39.3 19.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/htb3nvcs",
   "source_url": "https://omlx.ai/benchmarks/performance/htb3nvcs",
   "tg_tokens": null,
   "tps": 39.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-htb3nvcs-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 999.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 999.0 31.2 21.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/htb3nvcs",
   "source_url": "https://omlx.ai/benchmarks/performance/htb3nvcs",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8gfdt77n-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTPLX MTP build; oMLX community benchmark trace; peak memory 27.0 GB at this context; duplicate traces of the same session: 76gji899; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 697.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit-MTPLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 697.6 40.6 27.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8gfdt77n",
   "source_url": "https://omlx.ai/benchmarks/performance/8gfdt77n",
   "tg_tokens": null,
   "tps": 40.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8gfdt77n-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTPLX MTP build; oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: 76gji899; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 858.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit-MTPLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 858.9 51.7 28.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8gfdt77n",
   "source_url": "https://omlx.ai/benchmarks/performance/8gfdt77n",
   "tg_tokens": null,
   "tps": 51.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ncyl5lx0-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 331.7 GB at this context; duplicate traces of the same session: 4w96uy6o; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 255.8,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-oQ3e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 255.8 27.2 331.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ncyl5lx0",
   "source_url": "https://omlx.ai/benchmarks/performance/ncyl5lx0",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ncyl5lx0-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 332.4 GB at this context; duplicate traces of the same session: 4w96uy6o; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 322.8,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-oQ3e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 322.8 25.7 332.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ncyl5lx0",
   "source_url": "https://omlx.ai/benchmarks/performance/ncyl5lx0",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ncyl5lx0-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 332.7 GB at this context; duplicate traces of the same session: 4w96uy6o; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 348.4,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-oQ3e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 8k 348.4 28.3 332.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ncyl5lx0",
   "source_url": "https://omlx.ai/benchmarks/performance/ncyl5lx0",
   "tg_tokens": null,
   "tps": 28.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ncyl5lx0-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 333.2 GB at this context; duplicate traces of the same session: 4w96uy6o; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.2,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-oQ3e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 16k 323.2 35.7 333.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ncyl5lx0",
   "source_url": "https://omlx.ai/benchmarks/performance/ncyl5lx0",
   "tg_tokens": null,
   "tps": 35.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-rxbelhkp-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 225.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 225.9 30.6 18.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rxbelhkp",
   "source_url": "https://omlx.ai/benchmarks/performance/rxbelhkp",
   "tg_tokens": null,
   "tps": 30.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-rxbelhkp-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 234.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 234.8 58.0 23.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rxbelhkp",
   "source_url": "https://omlx.ai/benchmarks/performance/rxbelhkp",
   "tg_tokens": null,
   "tps": 58.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-lfmqcryn-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 240.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 240.9 26.0 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lfmqcryn",
   "source_url": "https://omlx.ai/benchmarks/performance/lfmqcryn",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-lfmqcryn-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 240.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 240.1 23.8 21.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lfmqcryn",
   "source_url": "https://omlx.ai/benchmarks/performance/lfmqcryn",
   "tg_tokens": null,
   "tps": 23.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-9m40oo1u-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 251.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 251.9 26.6 16.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9m40oo1u",
   "source_url": "https://omlx.ai/benchmarks/performance/9m40oo1u",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-9m40oo1u-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 242.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 242.9 23.7 21.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9m40oo1u",
   "source_url": "https://omlx.ai/benchmarks/performance/9m40oo1u",
   "tg_tokens": null,
   "tps": 23.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-qohypqrw-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 18.5 GB; TTFT 4241 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 241.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 241.5 PP tok/s 53.6 TG tok/s 4241 TTFT (ms) 18.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qohypqrw",
   "source_url": "https://omlx.ai/benchmarks/performance/qohypqrw",
   "tg_tokens": null,
   "tps": 53.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-iwdwq3kl-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: 4lz8o2rf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 232.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 232.7 23.5 22.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iwdwq3kl",
   "source_url": "https://omlx.ai/benchmarks/performance/iwdwq3kl",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-iwdwq3kl-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: 4lz8o2rf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 202.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 202.9 20.1 25.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iwdwq3kl",
   "source_url": "https://omlx.ai/benchmarks/performance/iwdwq3kl",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-g7nvyyue-ctx1024",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: 18p09w6b, coa1ozft, ewtcva3h, thlx72wc, v116sjvq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1273.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,273 113.6 22.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g7nvyyue",
   "source_url": "https://omlx.ai/benchmarks/performance/g7nvyyue",
   "tg_tokens": null,
   "tps": 113.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-g7nvyyue-ctx4096",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: 18p09w6b, coa1ozft, ewtcva3h, thlx72wc, v116sjvq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1556.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,556 109.6 23.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g7nvyyue",
   "source_url": "https://omlx.ai/benchmarks/performance/g7nvyyue",
   "tg_tokens": null,
   "tps": 109.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-g7nvyyue-ctx8192",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 18p09w6b, coa1ozft, ewtcva3h, thlx72wc, v116sjvq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1407.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,407 107.5 23.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g7nvyyue",
   "source_url": "https://omlx.ai/benchmarks/performance/g7nvyyue",
   "tg_tokens": null,
   "tps": 107.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-g7nvyyue-ctx16384",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 18p09w6b, coa1ozft, ewtcva3h, thlx72wc, v116sjvq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1333.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,333 98.2 23.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g7nvyyue",
   "source_url": "https://omlx.ai/benchmarks/performance/g7nvyyue",
   "tg_tokens": null,
   "tps": 98.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-g7nvyyue-ctx32768",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: 18p09w6b, coa1ozft, ewtcva3h, thlx72wc, v116sjvq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1160.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 1,160 89.9 24.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g7nvyyue",
   "source_url": "https://omlx.ai/benchmarks/performance/g7nvyyue",
   "tg_tokens": null,
   "tps": 89.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-g7nvyyue-ctx65536",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: 18p09w6b, coa1ozft, ewtcva3h, thlx72wc, v116sjvq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 977.1,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 977.1 77.4 24.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g7nvyyue",
   "source_url": "https://omlx.ai/benchmarks/performance/g7nvyyue",
   "tg_tokens": null,
   "tps": 77.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-5ocvsi80-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Nail build; oMLX community benchmark trace; peak memory 33.2 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 581.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX-6bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 1k 581.9 43.3 33.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ocvsi80",
   "source_url": "https://omlx.ai/benchmarks/performance/5ocvsi80",
   "tg_tokens": null,
   "tps": 43.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-5ocvsi80-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Nail build; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 599.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX-6bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 4k 599.7 37.1 34.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ocvsi80",
   "source_url": "https://omlx.ai/benchmarks/performance/5ocvsi80",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-2ny51c0u-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Nail build; oMLX community benchmark trace; peak memory 33.1 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 579.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX-6bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 1k 579.0 46.7 33.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ny51c0u",
   "source_url": "https://omlx.ai/benchmarks/performance/2ny51c0u",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-2ny51c0u-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Nail build; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 608.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX-6bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 4k 608.7 39.1 35.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ny51c0u",
   "source_url": "https://omlx.ai/benchmarks/performance/2ny51c0u",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-c80xamrd-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Nail build; oMLX community benchmark trace; peak memory 30.9 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 578.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX-6bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 1k 578.9 34.4 30.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c80xamrd",
   "source_url": "https://omlx.ai/benchmarks/performance/c80xamrd",
   "tg_tokens": null,
   "tps": 34.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-c80xamrd-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Nail build; oMLX community benchmark trace; peak memory 31.9 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 568.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX-6bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 4k 568.1 34.6 31.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c80xamrd",
   "source_url": "https://omlx.ai/benchmarks/performance/c80xamrd",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-3nbc4e5m-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Nail build; oMLX community benchmark trace; peak memory 33.2 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 578.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX-6bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 1k 578.0 46.0 33.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3nbc4e5m",
   "source_url": "https://omlx.ai/benchmarks/performance/3nbc4e5m",
   "tg_tokens": null,
   "tps": 46.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-3nbc4e5m-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Nail build; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 594.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX-6bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 4k 594.7 39.2 34.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3nbc4e5m",
   "source_url": "https://omlx.ai/benchmarks/performance/3nbc4e5m",
   "tg_tokens": null,
   "tps": 39.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-l1acqnhm-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Nail build; oMLX community benchmark trace; peak memory 33.1 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 547.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX-6bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 1k 547.6 32.5 33.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l1acqnhm",
   "source_url": "https://omlx.ai/benchmarks/performance/l1acqnhm",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-l1acqnhm-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Nail build; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 611.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX-6bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 4k 611.7 38.5 35.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l1acqnhm",
   "source_url": "https://omlx.ai/benchmarks/performance/l1acqnhm",
   "tg_tokens": null,
   "tps": 38.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-vywlavbb-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.9 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 378.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 6bit; 1k 378.1 44.0 7.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vywlavbb",
   "source_url": "https://omlx.ai/benchmarks/performance/vywlavbb",
   "tg_tokens": null,
   "tps": 44.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-vywlavbb-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.4 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 368.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 6bit; 4k 368.9 42.6 8.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vywlavbb",
   "source_url": "https://omlx.ai/benchmarks/performance/vywlavbb",
   "tg_tokens": null,
   "tps": 42.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-8js3mgdr-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.9 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 372.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 6bit; 1k 372.4 43.7 7.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8js3mgdr",
   "source_url": "https://omlx.ai/benchmarks/performance/8js3mgdr",
   "tg_tokens": null,
   "tps": 43.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-8js3mgdr-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.4 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 365.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 6bit; 4k 365.0 42.4 8.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8js3mgdr",
   "source_url": "https://omlx.ai/benchmarks/performance/8js3mgdr",
   "tg_tokens": null,
   "tps": 42.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-pn31s3mb-ctx4096",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: dhcdf3tq; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1761.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,761 52.5 17.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pn31s3mb",
   "source_url": "https://omlx.ai/benchmarks/performance/pn31s3mb",
   "tg_tokens": null,
   "tps": 52.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-pn31s3mb-ctx16384",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: dhcdf3tq; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1338.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 1,338 40.9 18.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pn31s3mb",
   "source_url": "https://omlx.ai/benchmarks/performance/pn31s3mb",
   "tg_tokens": null,
   "tps": 40.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ibvs1znv-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2731.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,731 107.5 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ibvs1znv",
   "source_url": "https://omlx.ai/benchmarks/performance/ibvs1znv",
   "tg_tokens": null,
   "tps": 107.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ibvs1znv-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3412.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,412 104.9 20.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ibvs1znv",
   "source_url": "https://omlx.ai/benchmarks/performance/ibvs1znv",
   "tg_tokens": null,
   "tps": 104.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ibvs1znv-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3042.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 3,042 94.0 20.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ibvs1znv",
   "source_url": "https://omlx.ai/benchmarks/performance/ibvs1znv",
   "tg_tokens": null,
   "tps": 94.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-2553hw4h-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: r65s1x3v; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 808.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 808.4 98.0 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2553hw4h",
   "source_url": "https://omlx.ai/benchmarks/performance/2553hw4h",
   "tg_tokens": null,
   "tps": 98.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-2553hw4h-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: r65s1x3v; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1877.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,877 99.0 20.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2553hw4h",
   "source_url": "https://omlx.ai/benchmarks/performance/2553hw4h",
   "tg_tokens": null,
   "tps": 99.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-2553hw4h-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: r65s1x3v; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2970.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,970 78.9 20.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2553hw4h",
   "source_url": "https://omlx.ai/benchmarks/performance/2553hw4h",
   "tg_tokens": null,
   "tps": 78.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-897mgrbk-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 935.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 1k 935.7 82.6 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/897mgrbk",
   "source_url": "https://omlx.ai/benchmarks/performance/897mgrbk",
   "tg_tokens": null,
   "tps": 82.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-897mgrbk-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 973.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k 973.2 84.5 21.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/897mgrbk",
   "source_url": "https://omlx.ai/benchmarks/performance/897mgrbk",
   "tg_tokens": null,
   "tps": 84.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-897mgrbk-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 800.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 8k 800.8 78.2 21.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/897mgrbk",
   "source_url": "https://omlx.ai/benchmarks/performance/897mgrbk",
   "tg_tokens": null,
   "tps": 78.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-897mgrbk-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 665.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 16k 665.2 68.9 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/897mgrbk",
   "source_url": "https://omlx.ai/benchmarks/performance/897mgrbk",
   "tg_tokens": null,
   "tps": 68.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-42qc8u2n-ctx1024",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: 1xjauexg, 98ogtog7, af67bxzg, hh6rzjmu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 422.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 422.5 110.9 12.7 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/42qc8u2n",
   "source_url": "https://omlx.ai/benchmarks/performance/42qc8u2n",
   "tg_tokens": null,
   "tps": 110.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-42qc8u2n-ctx8192",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: 1xjauexg, 98ogtog7, af67bxzg, hh6rzjmu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4911.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 4,911 104.3 13.4 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/42qc8u2n",
   "source_url": "https://omlx.ai/benchmarks/performance/42qc8u2n",
   "tg_tokens": null,
   "tps": 104.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-42qc8u2n-ctx16384",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.6 GB at this context; duplicate traces of the same session: 1xjauexg, 98ogtog7, af67bxzg, hh6rzjmu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4519.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 4,519 99.4 13.6 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/42qc8u2n",
   "source_url": "https://omlx.ai/benchmarks/performance/42qc8u2n",
   "tg_tokens": null,
   "tps": 99.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-42qc8u2n-ctx32768",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.9 GB at this context; duplicate traces of the same session: 1xjauexg, 98ogtog7, af67bxzg, hh6rzjmu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3947.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 32k 3,947 89.5 13.9 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/42qc8u2n",
   "source_url": "https://omlx.ai/benchmarks/performance/42qc8u2n",
   "tg_tokens": null,
   "tps": 89.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-42qc8u2n-ctx65536",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 14.5 GB at this context; duplicate traces of the same session: 1xjauexg, 98ogtog7, af67bxzg, hh6rzjmu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3024.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 64k 3,024 74.8 14.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/42qc8u2n",
   "source_url": "https://omlx.ai/benchmarks/performance/42qc8u2n",
   "tg_tokens": null,
   "tps": 74.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-42qc8u2n-ctx131072",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: 1xjauexg, 98ogtog7, af67bxzg, hh6rzjmu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2041.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 128k 2,041 57.7 15.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/42qc8u2n",
   "source_url": "https://omlx.ai/benchmarks/performance/42qc8u2n",
   "tg_tokens": null,
   "tps": 57.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-42qc8u2n-ctx199680",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: 1xjauexg, 98ogtog7, af67bxzg, hh6rzjmu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1519.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 195k 1,519 46.7 17.1 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/42qc8u2n",
   "source_url": "https://omlx.ai/benchmarks/performance/42qc8u2n",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-f4byuoui-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final",
   "notes": "dequantized fp16 MTP build; oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: omewudpv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 890.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized-oQ8e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 890.5 88.3 35.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f4byuoui",
   "source_url": "https://omlx.ai/benchmarks/performance/f4byuoui",
   "tg_tokens": null,
   "tps": 88.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-f4byuoui-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final",
   "notes": "dequantized fp16 MTP build; oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: omewudpv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2759.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized-oQ8e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 2,759 89.4 36.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f4byuoui",
   "source_url": "https://omlx.ai/benchmarks/performance/f4byuoui",
   "tg_tokens": null,
   "tps": 89.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-n3u1v1p0-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V7-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.4 GB at this context; duplicate traces of the same session: jg40irzo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3779.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V7-dequantized-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 3,779 104.9 36.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n3u1v1p0",
   "source_url": "https://omlx.ai/benchmarks/performance/n3u1v1p0",
   "tg_tokens": null,
   "tps": 104.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-n3u1v1p0-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V7-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.0 GB at this context; duplicate traces of the same session: jg40irzo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4972.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V7-dequantized-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 4,972 108.5 37.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n3u1v1p0",
   "source_url": "https://omlx.ai/benchmarks/performance/n3u1v1p0",
   "tg_tokens": null,
   "tps": 108.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zawmq8kw-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 767.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 767.5 26.2 16.0 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zawmq8kw",
   "source_url": "https://omlx.ai/benchmarks/performance/zawmq8kw",
   "tg_tokens": null,
   "tps": 26.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zawmq8kw-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 822.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 822.2 25.6 17.0 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zawmq8kw",
   "source_url": "https://omlx.ai/benchmarks/performance/zawmq8kw",
   "tg_tokens": null,
   "tps": 25.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-zawmq8kw-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 824.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 824.3 25.2 17.3 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zawmq8kw",
   "source_url": "https://omlx.ai/benchmarks/performance/zawmq8kw",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-zawmq8kw-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 800.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 800.3 24.0 17.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zawmq8kw",
   "source_url": "https://omlx.ai/benchmarks/performance/zawmq8kw",
   "tg_tokens": null,
   "tps": 24.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-zawmq8kw-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 748.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 32k 748.0 22.0 18.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zawmq8kw",
   "source_url": "https://omlx.ai/benchmarks/performance/zawmq8kw",
   "tg_tokens": null,
   "tps": 22.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-zawmq8kw-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 662.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 64k 662.5 20.5 20.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zawmq8kw",
   "source_url": "https://omlx.ai/benchmarks/performance/zawmq8kw",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-76gji899-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTPLX MTP build; oMLX community benchmark trace; peak memory 27.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 697.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit-MTPLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 697.6 40.6 27.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/76gji899",
   "source_url": "https://omlx.ai/benchmarks/performance/76gji899",
   "tg_tokens": null,
   "tps": 40.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-76gji899-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTPLX MTP build; oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 858.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit-MTPLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 858.9 51.7 28.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/76gji899",
   "source_url": "https://omlx.ai/benchmarks/performance/76gji899",
   "tg_tokens": null,
   "tps": 51.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-4a33t77w-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 219.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 219.4 12.7 28.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4a33t77w",
   "source_url": "https://omlx.ai/benchmarks/performance/4a33t77w",
   "tg_tokens": null,
   "tps": 12.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-4a33t77w-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 168.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 168.6 11.9 31.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4a33t77w",
   "source_url": "https://omlx.ai/benchmarks/performance/4a33t77w",
   "tg_tokens": null,
   "tps": 11.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-468m3y2m-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 57.2 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 300.5,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.8-27B-bf16 on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 bf16; 16k 300.5 11.5 57.2 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/468m3y2m",
   "source_url": "https://omlx.ai/benchmarks/performance/468m3y2m",
   "tg_tokens": null,
   "tps": 11.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-468m3y2m-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 64.7 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 159.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.8-27B-bf16 on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 bf16; 128k 159.0 9.7 64.7 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/468m3y2m",
   "source_url": "https://omlx.ai/benchmarks/performance/468m3y2m",
   "tg_tokens": null,
   "tps": 9.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-pbywa13n-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: y8wemls1; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 106.7 18.2 18.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pbywa13n",
   "source_url": "https://omlx.ai/benchmarks/performance/pbywa13n",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-pbywa13n-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: y8wemls1; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 102.2 21.3 22.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pbywa13n",
   "source_url": "https://omlx.ai/benchmarks/performance/pbywa13n",
   "tg_tokens": null,
   "tps": 21.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-qyhkw3ja-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: xn5wp2cm; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 111.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 111.2 18.1 16.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qyhkw3ja",
   "source_url": "https://omlx.ai/benchmarks/performance/qyhkw3ja",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-qyhkw3ja-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: xn5wp2cm; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 103.8 17.5 20.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qyhkw3ja",
   "source_url": "https://omlx.ai/benchmarks/performance/qyhkw3ja",
   "tg_tokens": null,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qe2f82jh-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 812.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 812.1 53.6 20.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qe2f82jh",
   "source_url": "https://omlx.ai/benchmarks/performance/qe2f82jh",
   "tg_tokens": null,
   "tps": 53.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qe2f82jh-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 910.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 910.1 50.7 21.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qe2f82jh",
   "source_url": "https://omlx.ai/benchmarks/performance/qe2f82jh",
   "tg_tokens": null,
   "tps": 50.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1mytfttx-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 192.0 GB at this context; duplicate traces of the same session: 3m17flw8, nyhn0nqo; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 892.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-Flash-Next-MLX-8bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 892.2 57.6 192.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1mytfttx",
   "source_url": "https://omlx.ai/benchmarks/performance/1mytfttx",
   "tg_tokens": null,
   "tps": 57.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1mytfttx-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 192.4 GB at this context; duplicate traces of the same session: 3m17flw8, nyhn0nqo; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1194.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-Flash-Next-MLX-8bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 1,194 57.5 192.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1mytfttx",
   "source_url": "https://omlx.ai/benchmarks/performance/1mytfttx",
   "tg_tokens": null,
   "tps": 57.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1mytfttx-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 192.8 GB at this context; duplicate traces of the same session: 3m17flw8, nyhn0nqo; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1206.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-Flash-Next-MLX-8bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k 1,206 53.9 192.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1mytfttx",
   "source_url": "https://omlx.ai/benchmarks/performance/1mytfttx",
   "tg_tokens": null,
   "tps": 53.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-y9cw9wjl-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 190.5 GB at this context; duplicate traces of the same session: 5gxcmi0i, zw7kugfv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 906.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ8-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 906.0 46.8 190.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y9cw9wjl",
   "source_url": "https://omlx.ai/benchmarks/performance/y9cw9wjl",
   "tg_tokens": null,
   "tps": 46.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-y9cw9wjl-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 190.8 GB at this context; duplicate traces of the same session: 5gxcmi0i, zw7kugfv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1204.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ8-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 1,204 56.7 190.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y9cw9wjl",
   "source_url": "https://omlx.ai/benchmarks/performance/y9cw9wjl",
   "tg_tokens": null,
   "tps": 56.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-y9cw9wjl-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 191.0 GB at this context; duplicate traces of the same session: 5gxcmi0i, zw7kugfv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1211.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ8-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 1,211 55.7 191.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y9cw9wjl",
   "source_url": "https://omlx.ai/benchmarks/performance/y9cw9wjl",
   "tg_tokens": null,
   "tps": 55.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-nn0tfhg8-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 66.2 GB at this context; duplicate traces of the same session: mjvnpstt, w6m2d2au, ztfkb8pj; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 78.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 1k 78.8 17.2 66.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nn0tfhg8",
   "source_url": "https://omlx.ai/benchmarks/performance/nn0tfhg8",
   "tg_tokens": null,
   "tps": 17.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-nn0tfhg8-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 66.6 GB at this context; duplicate traces of the same session: mjvnpstt, w6m2d2au, ztfkb8pj; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 216.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 4k 216.3 18.2 66.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nn0tfhg8",
   "source_url": "https://omlx.ai/benchmarks/performance/nn0tfhg8",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-nn0tfhg8-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 67.2 GB at this context; duplicate traces of the same session: mjvnpstt, w6m2d2au, ztfkb8pj; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 138.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 16k 138.6 14.2 67.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nn0tfhg8",
   "source_url": "https://omlx.ai/benchmarks/performance/nn0tfhg8",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-nn0tfhg8-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 68.3 GB at this context; duplicate traces of the same session: mjvnpstt, w6m2d2au, ztfkb8pj; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 64k 104.8 17.4 68.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nn0tfhg8",
   "source_url": "https://omlx.ai/benchmarks/performance/nn0tfhg8",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ni1rq5iy-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: b6b8on7r, mkiv2sfa; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1103.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,103 104.5 100.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ni1rq5iy",
   "source_url": "https://omlx.ai/benchmarks/performance/ni1rq5iy",
   "tg_tokens": null,
   "tps": 104.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ni1rq5iy-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: b6b8on7r, mkiv2sfa; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1369.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,369 82.6 101.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ni1rq5iy",
   "source_url": "https://omlx.ai/benchmarks/performance/ni1rq5iy",
   "tg_tokens": null,
   "tps": 82.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ni1rq5iy-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: b6b8on7r, mkiv2sfa; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1376.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 1,376 83.8 101.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ni1rq5iy",
   "source_url": "https://omlx.ai/benchmarks/performance/ni1rq5iy",
   "tg_tokens": null,
   "tps": 83.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ud5mcllj-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-VL-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.5 GB at this context; duplicate traces of the same session: 8u7kea1m, au77fpxx, lxhjezbb, pcvxzkgc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 244.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-VL-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 244.8 40.9 32.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ud5mcllj",
   "source_url": "https://omlx.ai/benchmarks/performance/ud5mcllj",
   "tg_tokens": null,
   "tps": 40.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ud5mcllj-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-VL-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.7 GB at this context; duplicate traces of the same session: 8u7kea1m, au77fpxx, lxhjezbb, pcvxzkgc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 276.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-VL-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 276.1 43.8 35.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ud5mcllj",
   "source_url": "https://omlx.ai/benchmarks/performance/ud5mcllj",
   "tg_tokens": null,
   "tps": 43.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ud5mcllj-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-VL-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: 8u7kea1m, au77fpxx, lxhjezbb, pcvxzkgc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 275.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-VL-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 8k 275.7 37.6 35.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ud5mcllj",
   "source_url": "https://omlx.ai/benchmarks/performance/ud5mcllj",
   "tg_tokens": null,
   "tps": 37.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ud5mcllj-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-VL-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.5 GB at this context; duplicate traces of the same session: 8u7kea1m, au77fpxx, lxhjezbb, pcvxzkgc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 250.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-VL-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k 250.5 32.6 31.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ud5mcllj",
   "source_url": "https://omlx.ai/benchmarks/performance/ud5mcllj",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ud5mcllj-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-VL-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.3 GB at this context; duplicate traces of the same session: 8u7kea1m, au77fpxx, lxhjezbb, pcvxzkgc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 232.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-VL-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 232.4 30.3 33.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ud5mcllj",
   "source_url": "https://omlx.ai/benchmarks/performance/ud5mcllj",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-u1xjksc2-ctx1024",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: 4uo2462c; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 150.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 150.9 29.5 18.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u1xjksc2",
   "source_url": "https://omlx.ai/benchmarks/performance/u1xjksc2",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-u1xjksc2-ctx4096",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: 4uo2462c; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 161.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 161.3 23.9 23.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u1xjksc2",
   "source_url": "https://omlx.ai/benchmarks/performance/u1xjksc2",
   "tg_tokens": null,
   "tps": 23.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-gb80nyce-ctx4096",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 13.8 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 194.3,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 2bit; 4k 194.3 13.5 13.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gb80nyce",
   "source_url": "https://omlx.ai/benchmarks/performance/gb80nyce",
   "tg_tokens": null,
   "tps": 13.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-gb80nyce-ctx16384",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 15.2 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 159.7,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 2bit; 16k 159.7 10.8 15.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gb80nyce",
   "source_url": "https://omlx.ai/benchmarks/performance/gb80nyce",
   "tg_tokens": null,
   "tps": 10.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-isojhkdv-ctx1024",
   "model": "gpt-oss 120B",
   "notes": "bf16 KV cache; oMLX community benchmark trace; peak memory 61.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 727.7,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "gpt-oss-120b-mxfp4-bf16 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 727.7 81.2 61.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/isojhkdv",
   "source_url": "https://omlx.ai/benchmarks/performance/isojhkdv",
   "tg_tokens": null,
   "tps": 81.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-isojhkdv-ctx4096",
   "model": "gpt-oss 120B",
   "notes": "bf16 KV cache; oMLX community benchmark trace; peak memory 61.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2110.0,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "gpt-oss-120b-mxfp4-bf16 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,110 78.5 61.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/isojhkdv",
   "source_url": "https://omlx.ai/benchmarks/performance/isojhkdv",
   "tg_tokens": null,
   "tps": 78.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4w96uy6o-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 331.7 GB at this context; duplicate traces of the same session: 7uqj4zmt, ljzy79y4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 255.8,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-oQ3e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 255.8 27.2 331.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4w96uy6o",
   "source_url": "https://omlx.ai/benchmarks/performance/4w96uy6o",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4w96uy6o-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 332.4 GB at this context; duplicate traces of the same session: 7uqj4zmt, ljzy79y4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 322.8,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-oQ3e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 322.8 25.7 332.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4w96uy6o",
   "source_url": "https://omlx.ai/benchmarks/performance/4w96uy6o",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4w96uy6o-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 332.7 GB at this context; duplicate traces of the same session: 7uqj4zmt, ljzy79y4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 348.4,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-oQ3e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 8k 348.4 28.3 332.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4w96uy6o",
   "source_url": "https://omlx.ai/benchmarks/performance/4w96uy6o",
   "tg_tokens": null,
   "tps": 28.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4w96uy6o-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 333.2 GB at this context; duplicate traces of the same session: 7uqj4zmt, ljzy79y4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.2,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-oQ3e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 16k 323.2 35.7 333.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4w96uy6o",
   "source_url": "https://omlx.ai/benchmarks/performance/4w96uy6o",
   "tg_tokens": null,
   "tps": 35.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-4lz8o2rf-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 232.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 232.7 23.5 22.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4lz8o2rf",
   "source_url": "https://omlx.ai/benchmarks/performance/4lz8o2rf",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-4lz8o2rf-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 202.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 202.9 20.1 25.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4lz8o2rf",
   "source_url": "https://omlx.ai/benchmarks/performance/4lz8o2rf",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-sz0h8y3o-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 18.5 GB; TTFT 4245 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 241.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 241.2 PP tok/s 63.9 TG tok/s 4245 TTFT (ms) 18.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sz0h8y3o",
   "source_url": "https://omlx.ai/benchmarks/performance/sz0h8y3o",
   "tg_tokens": null,
   "tps": 63.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ahr8v8gq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 3mysjztu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 247.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 247.7 26.3 16.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ahr8v8gq",
   "source_url": "https://omlx.ai/benchmarks/performance/ahr8v8gq",
   "tg_tokens": null,
   "tps": 26.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ahr8v8gq-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 3mysjztu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 246.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 246.2 24.0 21.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ahr8v8gq",
   "source_url": "https://omlx.ai/benchmarks/performance/ahr8v8gq",
   "tg_tokens": null,
   "tps": 24.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-7duq89gx-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: oktt2kft; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 221.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 221.9 80.2 25.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7duq89gx",
   "source_url": "https://omlx.ai/benchmarks/performance/7duq89gx",
   "tg_tokens": null,
   "tps": 80.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-7duq89gx-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: oktt2kft; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 194.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 194.1 46.7 27.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7duq89gx",
   "source_url": "https://omlx.ai/benchmarks/performance/7duq89gx",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-thlx72wc-ctx1024",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: 18p09w6b, coa1ozft, ewtcva3h, v116sjvq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1273.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,273 113.6 22.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/thlx72wc",
   "source_url": "https://omlx.ai/benchmarks/performance/thlx72wc",
   "tg_tokens": null,
   "tps": 113.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-thlx72wc-ctx4096",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: 18p09w6b, coa1ozft, ewtcva3h, v116sjvq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1556.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,556 109.6 23.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/thlx72wc",
   "source_url": "https://omlx.ai/benchmarks/performance/thlx72wc",
   "tg_tokens": null,
   "tps": 109.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-thlx72wc-ctx8192",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 18p09w6b, coa1ozft, ewtcva3h, v116sjvq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1407.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,407 107.5 23.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/thlx72wc",
   "source_url": "https://omlx.ai/benchmarks/performance/thlx72wc",
   "tg_tokens": null,
   "tps": 107.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-thlx72wc-ctx16384",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 18p09w6b, coa1ozft, ewtcva3h, v116sjvq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1333.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,333 98.2 23.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/thlx72wc",
   "source_url": "https://omlx.ai/benchmarks/performance/thlx72wc",
   "tg_tokens": null,
   "tps": 98.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-thlx72wc-ctx32768",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: 18p09w6b, coa1ozft, ewtcva3h, v116sjvq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1160.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 1,160 89.9 24.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/thlx72wc",
   "source_url": "https://omlx.ai/benchmarks/performance/thlx72wc",
   "tg_tokens": null,
   "tps": 89.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-thlx72wc-ctx65536",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: 18p09w6b, coa1ozft, ewtcva3h, v116sjvq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 977.1,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 977.1 77.4 24.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/thlx72wc",
   "source_url": "https://omlx.ai/benchmarks/performance/thlx72wc",
   "tg_tokens": null,
   "tps": 77.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-j0dcke5i-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 957.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen-3.8-27B-Uncensored-8-Bit-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 957.1 18.6 29.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j0dcke5i",
   "source_url": "https://omlx.ai/benchmarks/performance/j0dcke5i",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-j0dcke5i-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 933.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen-3.8-27B-Uncensored-8-Bit-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 16k 933.3 18.1 30.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j0dcke5i",
   "source_url": "https://omlx.ai/benchmarks/performance/j0dcke5i",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-dhcdf3tq-ctx4096",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1761.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,761 52.5 17.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dhcdf3tq",
   "source_url": "https://omlx.ai/benchmarks/performance/dhcdf3tq",
   "tg_tokens": null,
   "tps": 52.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-dhcdf3tq-ctx16384",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1338.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 1,338 40.9 18.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dhcdf3tq",
   "source_url": "https://omlx.ai/benchmarks/performance/dhcdf3tq",
   "tg_tokens": null,
   "tps": 40.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-r65s1x3v-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 808.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 808.4 98.0 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r65s1x3v",
   "source_url": "https://omlx.ai/benchmarks/performance/r65s1x3v",
   "tg_tokens": null,
   "tps": 98.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-r65s1x3v-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1877.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,877 99.0 20.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r65s1x3v",
   "source_url": "https://omlx.ai/benchmarks/performance/r65s1x3v",
   "tg_tokens": null,
   "tps": 99.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-r65s1x3v-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2970.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,970 78.9 20.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r65s1x3v",
   "source_url": "https://omlx.ai/benchmarks/performance/r65s1x3v",
   "tg_tokens": null,
   "tps": 78.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1xjauexg-ctx1024",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: 98ogtog7, af67bxzg, hh6rzjmu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 422.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 422.5 110.9 12.7 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1xjauexg",
   "source_url": "https://omlx.ai/benchmarks/performance/1xjauexg",
   "tg_tokens": null,
   "tps": 110.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1xjauexg-ctx8192",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: 98ogtog7, af67bxzg, hh6rzjmu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4911.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 4,911 104.3 13.4 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1xjauexg",
   "source_url": "https://omlx.ai/benchmarks/performance/1xjauexg",
   "tg_tokens": null,
   "tps": 104.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1xjauexg-ctx16384",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.6 GB at this context; duplicate traces of the same session: 98ogtog7, af67bxzg, hh6rzjmu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4519.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 4,519 99.4 13.6 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1xjauexg",
   "source_url": "https://omlx.ai/benchmarks/performance/1xjauexg",
   "tg_tokens": null,
   "tps": 99.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1xjauexg-ctx32768",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.9 GB at this context; duplicate traces of the same session: 98ogtog7, af67bxzg, hh6rzjmu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3947.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 32k 3,947 89.5 13.9 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1xjauexg",
   "source_url": "https://omlx.ai/benchmarks/performance/1xjauexg",
   "tg_tokens": null,
   "tps": 89.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1xjauexg-ctx65536",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 14.5 GB at this context; duplicate traces of the same session: 98ogtog7, af67bxzg, hh6rzjmu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3024.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 64k 3,024 74.8 14.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1xjauexg",
   "source_url": "https://omlx.ai/benchmarks/performance/1xjauexg",
   "tg_tokens": null,
   "tps": 74.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1xjauexg-ctx131072",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: 98ogtog7, af67bxzg, hh6rzjmu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2041.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 128k 2,041 57.7 15.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1xjauexg",
   "source_url": "https://omlx.ai/benchmarks/performance/1xjauexg",
   "tg_tokens": null,
   "tps": 57.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1xjauexg-ctx199680",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: 98ogtog7, af67bxzg, hh6rzjmu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1519.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 195k 1,519 46.7 17.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1xjauexg",
   "source_url": "https://omlx.ai/benchmarks/performance/1xjauexg",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-omewudpv-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final",
   "notes": "dequantized fp16 MTP build; oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 890.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized-oQ8e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 890.5 88.3 35.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/omewudpv",
   "source_url": "https://omlx.ai/benchmarks/performance/omewudpv",
   "tg_tokens": null,
   "tps": 88.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-omewudpv-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final",
   "notes": "dequantized fp16 MTP build; oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2759.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized-oQ8e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 2,759 89.4 36.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/omewudpv",
   "source_url": "https://omlx.ai/benchmarks/performance/omewudpv",
   "tg_tokens": null,
   "tps": 89.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jg40irzo-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V7-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3779.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V7-dequantized-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 3,779 104.9 36.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jg40irzo",
   "source_url": "https://omlx.ai/benchmarks/performance/jg40irzo",
   "tg_tokens": null,
   "tps": 104.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jg40irzo-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V7-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4972.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V7-dequantized-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 4,972 108.5 37.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jg40irzo",
   "source_url": "https://omlx.ai/benchmarks/performance/jg40irzo",
   "tg_tokens": null,
   "tps": 108.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gn2j4f6g-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 17.8 GB; TTFT 17484 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 937.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k tokens 937.1 PP tok/s 31.6 TG tok/s 17484 TTFT (ms) 17.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gn2j4f6g",
   "source_url": "https://omlx.ai/benchmarks/performance/gn2j4f6g",
   "tg_tokens": null,
   "tps": 31.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-o1uvx3tf-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: ftbo9vuq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 843.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 843.0 17.0 29.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o1uvx3tf",
   "source_url": "https://omlx.ai/benchmarks/performance/o1uvx3tf",
   "tg_tokens": null,
   "tps": 17.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-o1uvx3tf-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: ftbo9vuq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 835.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 16k 835.9 16.1 30.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o1uvx3tf",
   "source_url": "https://omlx.ai/benchmarks/performance/o1uvx3tf",
   "tg_tokens": null,
   "tps": 16.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-y8wemls1-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 106.7 18.2 18.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y8wemls1",
   "source_url": "https://omlx.ai/benchmarks/performance/y8wemls1",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-y8wemls1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 102.2 21.3 22.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y8wemls1",
   "source_url": "https://omlx.ai/benchmarks/performance/y8wemls1",
   "tg_tokens": null,
   "tps": 21.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-xn5wp2cm-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 111.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 111.2 18.1 16.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xn5wp2cm",
   "source_url": "https://omlx.ai/benchmarks/performance/xn5wp2cm",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-xn5wp2cm-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 103.8 17.5 20.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xn5wp2cm",
   "source_url": "https://omlx.ai/benchmarks/performance/xn5wp2cm",
   "tg_tokens": null,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-y2sef2xy-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace, single run at 195k context; peak memory 76.5 GB; TTFT 589420 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 339.3,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.8-Flash-Next-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 3bit; 195k tokens 339.3 PP tok/s 37.6 TG tok/s 589420 TTFT (ms) 76.5 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y2sef2xy",
   "source_url": "https://omlx.ai/benchmarks/performance/y2sef2xy",
   "tg_tokens": null,
   "tps": 37.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3m17flw8-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 192.0 GB at this context; duplicate traces of the same session: nyhn0nqo; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 892.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-Flash-Next-MLX-8bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 892.2 57.6 192.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3m17flw8",
   "source_url": "https://omlx.ai/benchmarks/performance/3m17flw8",
   "tg_tokens": null,
   "tps": 57.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-3m17flw8-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 192.4 GB at this context; duplicate traces of the same session: nyhn0nqo; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1194.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-Flash-Next-MLX-8bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 1,194 57.5 192.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3m17flw8",
   "source_url": "https://omlx.ai/benchmarks/performance/3m17flw8",
   "tg_tokens": null,
   "tps": 57.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3m17flw8-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 192.8 GB at this context; duplicate traces of the same session: nyhn0nqo; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1206.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-Flash-Next-MLX-8bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k 1,206 53.9 192.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3m17flw8",
   "source_url": "https://omlx.ai/benchmarks/performance/3m17flw8",
   "tg_tokens": null,
   "tps": 53.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-pv2ww6l8-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 191.9 GB at this context; duplicate traces of the same session: y857coyt, ylr8ejhh; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 906.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-Flash-Next-MLX-8bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 906.9 56.4 191.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pv2ww6l8",
   "source_url": "https://omlx.ai/benchmarks/performance/pv2ww6l8",
   "tg_tokens": null,
   "tps": 56.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-pv2ww6l8-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 192.4 GB at this context; duplicate traces of the same session: y857coyt, ylr8ejhh; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1194.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-Flash-Next-MLX-8bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 1,194 56.4 192.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pv2ww6l8",
   "source_url": "https://omlx.ai/benchmarks/performance/pv2ww6l8",
   "tg_tokens": null,
   "tps": 56.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-pv2ww6l8-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 192.5 GB at this context; duplicate traces of the same session: y857coyt, ylr8ejhh; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1207.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-Flash-Next-MLX-8bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 1,207 60.9 192.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pv2ww6l8",
   "source_url": "https://omlx.ai/benchmarks/performance/pv2ww6l8",
   "tg_tokens": null,
   "tps": 60.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-zw7kugfv-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 190.5 GB at this context; duplicate traces of the same session: 5gxcmi0i; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 906.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ8-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 906.0 46.8 190.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zw7kugfv",
   "source_url": "https://omlx.ai/benchmarks/performance/zw7kugfv",
   "tg_tokens": null,
   "tps": 46.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-zw7kugfv-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 190.8 GB at this context; duplicate traces of the same session: 5gxcmi0i; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1204.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ8-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 1,204 56.7 190.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zw7kugfv",
   "source_url": "https://omlx.ai/benchmarks/performance/zw7kugfv",
   "tg_tokens": null,
   "tps": 56.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-zw7kugfv-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 191.0 GB at this context; duplicate traces of the same session: 5gxcmi0i; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1211.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ8-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 1,211 55.7 191.0 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zw7kugfv",
   "source_url": "https://omlx.ai/benchmarks/performance/zw7kugfv",
   "tg_tokens": null,
   "tps": 55.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bxmnuhme-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 122.5 GB at this context; duplicate traces of the same session: gma4bv9e; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 818.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 1k 818.4 81.8 122.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bxmnuhme",
   "source_url": "https://omlx.ai/benchmarks/performance/bxmnuhme",
   "tg_tokens": null,
   "tps": 81.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bxmnuhme-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 122.8 GB at this context; duplicate traces of the same session: gma4bv9e; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 997.8,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 4k 997.8 74.5 122.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bxmnuhme",
   "source_url": "https://omlx.ai/benchmarks/performance/bxmnuhme",
   "tg_tokens": null,
   "tps": 74.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-ztfkb8pj-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 66.2 GB at this context; duplicate traces of the same session: mjvnpstt, w6m2d2au; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 78.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 1k 78.8 17.2 66.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ztfkb8pj",
   "source_url": "https://omlx.ai/benchmarks/performance/ztfkb8pj",
   "tg_tokens": null,
   "tps": 17.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-ztfkb8pj-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 66.6 GB at this context; duplicate traces of the same session: mjvnpstt, w6m2d2au; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 216.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 4k 216.3 18.2 66.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ztfkb8pj",
   "source_url": "https://omlx.ai/benchmarks/performance/ztfkb8pj",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-ztfkb8pj-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 67.2 GB at this context; duplicate traces of the same session: mjvnpstt, w6m2d2au; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 138.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 16k 138.6 14.2 67.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ztfkb8pj",
   "source_url": "https://omlx.ai/benchmarks/performance/ztfkb8pj",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-ztfkb8pj-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 68.3 GB at this context; duplicate traces of the same session: mjvnpstt, w6m2d2au; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 64k 104.8 17.4 68.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ztfkb8pj",
   "source_url": "https://omlx.ai/benchmarks/performance/ztfkb8pj",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-b6b8on7r-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: mkiv2sfa; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1103.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,103 104.5 100.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b6b8on7r",
   "source_url": "https://omlx.ai/benchmarks/performance/b6b8on7r",
   "tg_tokens": null,
   "tps": 104.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-b6b8on7r-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: mkiv2sfa; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1369.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,369 82.6 101.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b6b8on7r",
   "source_url": "https://omlx.ai/benchmarks/performance/b6b8on7r",
   "tg_tokens": null,
   "tps": 82.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-b6b8on7r-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: mkiv2sfa; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1376.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 1,376 83.8 101.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b6b8on7r",
   "source_url": "https://omlx.ai/benchmarks/performance/b6b8on7r",
   "tg_tokens": null,
   "tps": 83.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-8u7kea1m-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-VL-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.5 GB at this context; duplicate traces of the same session: au77fpxx, lxhjezbb, pcvxzkgc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 244.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-VL-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 244.8 40.9 32.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8u7kea1m",
   "source_url": "https://omlx.ai/benchmarks/performance/8u7kea1m",
   "tg_tokens": null,
   "tps": 40.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-8u7kea1m-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-VL-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.7 GB at this context; duplicate traces of the same session: au77fpxx, lxhjezbb, pcvxzkgc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 276.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-VL-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 276.1 43.8 35.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8u7kea1m",
   "source_url": "https://omlx.ai/benchmarks/performance/8u7kea1m",
   "tg_tokens": null,
   "tps": 43.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-8u7kea1m-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-VL-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: au77fpxx, lxhjezbb, pcvxzkgc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 275.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-VL-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 8k 275.7 37.6 35.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8u7kea1m",
   "source_url": "https://omlx.ai/benchmarks/performance/8u7kea1m",
   "tg_tokens": null,
   "tps": 37.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-8u7kea1m-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-VL-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.5 GB at this context; duplicate traces of the same session: au77fpxx, lxhjezbb, pcvxzkgc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 250.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-VL-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k 250.5 32.6 31.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8u7kea1m",
   "source_url": "https://omlx.ai/benchmarks/performance/8u7kea1m",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-8u7kea1m-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-VL-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.3 GB at this context; duplicate traces of the same session: au77fpxx, lxhjezbb, pcvxzkgc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 232.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-VL-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 232.4 30.3 33.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8u7kea1m",
   "source_url": "https://omlx.ai/benchmarks/performance/8u7kea1m",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-4uo2462c-ctx1024",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 150.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 150.9 29.5 18.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4uo2462c",
   "source_url": "https://omlx.ai/benchmarks/performance/4uo2462c",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-4uo2462c-ctx4096",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 161.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 161.3 23.9 23.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4uo2462c",
   "source_url": "https://omlx.ai/benchmarks/performance/4uo2462c",
   "tg_tokens": null,
   "tps": 23.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-esby3umz-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Youssofal MTPLX Optimized Speed; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: 1z399iss, 4wmrz3x7, q9lh2wd2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 422.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Youssofal--Qwen3.8-27B-MTPLX-Optimized-Speed on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 422.9 68.9 24.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/esby3umz",
   "source_url": "https://omlx.ai/benchmarks/performance/esby3umz",
   "tg_tokens": null,
   "tps": 68.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-esby3umz-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "Youssofal MTPLX Optimized Speed; oMLX community benchmark trace; peak memory 27.6 GB at this context; duplicate traces of the same session: 1z399iss, 4wmrz3x7, q9lh2wd2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 442.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Youssofal--Qwen3.8-27B-MTPLX-Optimized-Speed on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 442.2 58.8 27.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/esby3umz",
   "source_url": "https://omlx.ai/benchmarks/performance/esby3umz",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-esby3umz-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "Youssofal MTPLX Optimized Speed; oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: 1z399iss, 4wmrz3x7, q9lh2wd2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 436.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Youssofal--Qwen3.8-27B-MTPLX-Optimized-Speed on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 436.4 80.2 28.0 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/esby3umz",
   "source_url": "https://omlx.ai/benchmarks/performance/esby3umz",
   "tg_tokens": null,
   "tps": 80.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-esby3umz-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "Youssofal MTPLX Optimized Speed; oMLX community benchmark trace; peak memory 28.8 GB at this context; duplicate traces of the same session: 1z399iss, 4wmrz3x7, q9lh2wd2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 420.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Youssofal--Qwen3.8-27B-MTPLX-Optimized-Speed on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 420.9 74.9 28.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/esby3umz",
   "source_url": "https://omlx.ai/benchmarks/performance/esby3umz",
   "tg_tokens": null,
   "tps": 74.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-6fqfk7zw-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "chriswessels ThinkingCap; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 96naazzq; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "chriswessels--ThinkingCap-Qwen3.8-27B-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 204.0 45.0 18.5 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6fqfk7zw",
   "source_url": "https://omlx.ai/benchmarks/performance/6fqfk7zw",
   "tg_tokens": null,
   "tps": 45.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-6fqfk7zw-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "chriswessels ThinkingCap; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 96naazzq; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "chriswessels--ThinkingCap-Qwen3.8-27B-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 205.9 39.2 19.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6fqfk7zw",
   "source_url": "https://omlx.ai/benchmarks/performance/6fqfk7zw",
   "tg_tokens": null,
   "tps": 39.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2f3qj11b-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "chriswessels ThinkingCap; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 3uv4sl8m, 4wve2rv9; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "chriswessels--ThinkingCap-Qwen3.8-27B-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 204.4 37.5 18.5 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2f3qj11b",
   "source_url": "https://omlx.ai/benchmarks/performance/2f3qj11b",
   "tg_tokens": null,
   "tps": 37.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-2f3qj11b-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "chriswessels ThinkingCap; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 3uv4sl8m, 4wve2rv9; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 206.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "chriswessels--ThinkingCap-Qwen3.8-27B-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 206.0 38.6 19.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2f3qj11b",
   "source_url": "https://omlx.ai/benchmarks/performance/2f3qj11b",
   "tg_tokens": null,
   "tps": 38.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-2f3qj11b-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "chriswessels ThinkingCap; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: 3uv4sl8m, 4wve2rv9; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 202.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "chriswessels--ThinkingCap-Qwen3.8-27B-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 202.2 36.7 20.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2f3qj11b",
   "source_url": "https://omlx.ai/benchmarks/performance/2f3qj11b",
   "tg_tokens": null,
   "tps": 36.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-6sc2dijo-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "fcmeyer MTP build; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: p73j72ib; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "fcmeyer--Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 204.7 33.4 18.9 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6sc2dijo",
   "source_url": "https://omlx.ai/benchmarks/performance/6sc2dijo",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-6sc2dijo-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "fcmeyer MTP build; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: p73j72ib; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "fcmeyer--Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 205.9 38.0 19.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6sc2dijo",
   "source_url": "https://omlx.ai/benchmarks/performance/6sc2dijo",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ie5cypa2-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 14.7 GB at this context; duplicate traces of the same session: 2meabal7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3941.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 3,941 105.6 14.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ie5cypa2",
   "source_url": "https://omlx.ai/benchmarks/performance/ie5cypa2",
   "tg_tokens": null,
   "tps": 105.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ie5cypa2-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 15.2 GB at this context; duplicate traces of the same session: 2meabal7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3630.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 3,630 94.0 15.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ie5cypa2",
   "source_url": "https://omlx.ai/benchmarks/performance/ie5cypa2",
   "tg_tokens": null,
   "tps": 94.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-pi38fw5d-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.2 GB at this context; duplicate traces of the same session: 1wmeknir; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2539.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26b-a4b-it-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 2,539 83.6 26.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pi38fw5d",
   "source_url": "https://omlx.ai/benchmarks/performance/pi38fw5d",
   "tg_tokens": null,
   "tps": 83.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pi38fw5d-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: 1wmeknir; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3476.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26b-a4b-it-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 16k 3,476 75.3 26.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pi38fw5d",
   "source_url": "https://omlx.ai/benchmarks/performance/pi38fw5d",
   "tg_tokens": null,
   "tps": 75.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-pyoqh5uj-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "leonsarmiento pi MTP; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: xmmirmxn; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "leonsarmiento--Qwen3.8-27B-pi-4bit-mtp-mlx on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 205.4 36.8 19.2 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pyoqh5uj",
   "source_url": "https://omlx.ai/benchmarks/performance/pyoqh5uj",
   "tg_tokens": null,
   "tps": 36.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-pyoqh5uj-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "leonsarmiento pi MTP; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: xmmirmxn; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 206.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "leonsarmiento--Qwen3.8-27B-pi-4bit-mtp-mlx on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 206.8 42.0 20.2 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pyoqh5uj",
   "source_url": "https://omlx.ai/benchmarks/performance/pyoqh5uj",
   "tg_tokens": null,
   "tps": 42.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-kc9d147a-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "leonsarmiento pi MTP; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "leonsarmiento--Qwen3.8-27B-pi-4bit-mtp-mlx on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 205.5 35.3 19.3 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kc9d147a",
   "source_url": "https://omlx.ai/benchmarks/performance/kc9d147a",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-kc9d147a-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "leonsarmiento pi MTP; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 206.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "leonsarmiento--Qwen3.8-27B-pi-4bit-mtp-mlx on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 206.7 40.3 20.2 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kc9d147a",
   "source_url": "https://omlx.ai/benchmarks/performance/kc9d147a",
   "tg_tokens": null,
   "tps": 40.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-7uqj4zmt-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 331.7 GB at this context; duplicate traces of the same session: ljzy79y4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 255.8,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-oQ3e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 255.8 27.2 331.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7uqj4zmt",
   "source_url": "https://omlx.ai/benchmarks/performance/7uqj4zmt",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-7uqj4zmt-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 332.4 GB at this context; duplicate traces of the same session: ljzy79y4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 322.8,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-oQ3e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 322.8 25.7 332.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7uqj4zmt",
   "source_url": "https://omlx.ai/benchmarks/performance/7uqj4zmt",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-7uqj4zmt-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 332.7 GB at this context; duplicate traces of the same session: ljzy79y4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 348.4,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-oQ3e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 8k 348.4 28.3 332.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7uqj4zmt",
   "source_url": "https://omlx.ai/benchmarks/performance/7uqj4zmt",
   "tg_tokens": null,
   "tps": 28.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-7uqj4zmt-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 333.2 GB at this context; duplicate traces of the same session: ljzy79y4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.2,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-oQ3e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 16k 323.2 35.7 333.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7uqj4zmt",
   "source_url": "https://omlx.ai/benchmarks/performance/7uqj4zmt",
   "tg_tokens": null,
   "tps": 35.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-81qsshvs-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.8 GB at this context; duplicate traces of the same session: lyb5gcby; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 499.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 499.8 35.6 175.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/81qsshvs",
   "source_url": "https://omlx.ai/benchmarks/performance/81qsshvs",
   "tg_tokens": null,
   "tps": 35.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-81qsshvs-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: lyb5gcby; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 544.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 544.3 70.4 177.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/81qsshvs",
   "source_url": "https://omlx.ai/benchmarks/performance/81qsshvs",
   "tg_tokens": null,
   "tps": 70.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-81qsshvs-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.0 GB at this context; duplicate traces of the same session: lyb5gcby; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 526.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 526.1 37.5 178.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/81qsshvs",
   "source_url": "https://omlx.ai/benchmarks/performance/81qsshvs",
   "tg_tokens": null,
   "tps": 37.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-81qsshvs-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.7 GB at this context; duplicate traces of the same session: lyb5gcby; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 514.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 514.9 26.8 178.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/81qsshvs",
   "source_url": "https://omlx.ai/benchmarks/performance/81qsshvs",
   "tg_tokens": null,
   "tps": 26.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-3mysjztu-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 247.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 247.7 26.3 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3mysjztu",
   "source_url": "https://omlx.ai/benchmarks/performance/3mysjztu",
   "tg_tokens": null,
   "tps": 26.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-3mysjztu-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 246.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 246.2 24.0 21.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3mysjztu",
   "source_url": "https://omlx.ai/benchmarks/performance/3mysjztu",
   "tg_tokens": null,
   "tps": 24.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-oktt2kft-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 221.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 221.9 80.2 25.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oktt2kft",
   "source_url": "https://omlx.ai/benchmarks/performance/oktt2kft",
   "tg_tokens": null,
   "tps": 80.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-oktt2kft-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 194.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 194.1 46.7 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oktt2kft",
   "source_url": "https://omlx.ai/benchmarks/performance/oktt2kft",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-gzhimbdy-ctx32768",
   "model": "Laguna XS 2.1",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 8ehvpcox; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2553.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Laguna-XS-2.1-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 2,553 106.3 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gzhimbdy",
   "source_url": "https://omlx.ai/benchmarks/performance/gzhimbdy",
   "tg_tokens": null,
   "tps": 106.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-gzhimbdy-ctx65536",
   "model": "Laguna XS 2.1",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 8ehvpcox; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1923.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Laguna-XS-2.1-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 1,923 80.8 20.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gzhimbdy",
   "source_url": "https://omlx.ai/benchmarks/performance/gzhimbdy",
   "tg_tokens": null,
   "tps": 80.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-moo2x24g-ctx4096",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 12.7 GB; TTFT 1787 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2292.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k tokens 2,292 PP tok/s 125.6 TG tok/s 1787 TTFT (ms) 12.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/moo2x24g",
   "source_url": "https://omlx.ai/benchmarks/performance/moo2x24g",
   "tg_tokens": null,
   "tps": 125.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-tvyox5gm-ctx8192",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace, single run at 8k context; peak memory 12.8 GB; TTFT 3725 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2199.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k tokens 2,199 PP tok/s 122.4 TG tok/s 3725 TTFT (ms) 12.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tvyox5gm",
   "source_url": "https://omlx.ai/benchmarks/performance/tvyox5gm",
   "tg_tokens": null,
   "tps": 122.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-7x3yhatq-ctx1024",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 12.6 GB; TTFT 651 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1574.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k tokens 1,574 PP tok/s 131.9 TG tok/s 651 TTFT (ms) 12.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7x3yhatq",
   "source_url": "https://omlx.ai/benchmarks/performance/7x3yhatq",
   "tg_tokens": null,
   "tps": 131.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-17e7y7g9-ctx8192",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace, single run at 8k context; peak memory 12.8 GB; TTFT 3724 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2200.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k tokens 2,200 PP tok/s 123.1 TG tok/s 3724 TTFT (ms) 12.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/17e7y7g9",
   "source_url": "https://omlx.ai/benchmarks/performance/17e7y7g9",
   "tg_tokens": null,
   "tps": 123.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-18p09w6b-ctx1024",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: coa1ozft, ewtcva3h, v116sjvq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1273.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,273 113.6 22.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/18p09w6b",
   "source_url": "https://omlx.ai/benchmarks/performance/18p09w6b",
   "tg_tokens": null,
   "tps": 113.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-18p09w6b-ctx4096",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: coa1ozft, ewtcva3h, v116sjvq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1556.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,556 109.6 23.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/18p09w6b",
   "source_url": "https://omlx.ai/benchmarks/performance/18p09w6b",
   "tg_tokens": null,
   "tps": 109.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-18p09w6b-ctx8192",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: coa1ozft, ewtcva3h, v116sjvq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1407.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,407 107.5 23.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/18p09w6b",
   "source_url": "https://omlx.ai/benchmarks/performance/18p09w6b",
   "tg_tokens": null,
   "tps": 107.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-18p09w6b-ctx16384",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: coa1ozft, ewtcva3h, v116sjvq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1333.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,333 98.2 23.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/18p09w6b",
   "source_url": "https://omlx.ai/benchmarks/performance/18p09w6b",
   "tg_tokens": null,
   "tps": 98.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-18p09w6b-ctx32768",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: coa1ozft, ewtcva3h, v116sjvq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1160.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 1,160 89.9 24.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/18p09w6b",
   "source_url": "https://omlx.ai/benchmarks/performance/18p09w6b",
   "tg_tokens": null,
   "tps": 89.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-18p09w6b-ctx65536",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: coa1ozft, ewtcva3h, v116sjvq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 977.1,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 977.1 77.4 24.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/18p09w6b",
   "source_url": "https://omlx.ai/benchmarks/performance/18p09w6b",
   "tg_tokens": null,
   "tps": 77.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-y3n2ywlx-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 957.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen-3.8-27B-Uncensored-8-Bit-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 957.1 18.6 29.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y3n2ywlx",
   "source_url": "https://omlx.ai/benchmarks/performance/y3n2ywlx",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-y3n2ywlx-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 933.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen-3.8-27B-Uncensored-8-Bit-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 16k 933.3 18.1 30.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y3n2ywlx",
   "source_url": "https://omlx.ai/benchmarks/performance/y3n2ywlx",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-98ogtog7-ctx1024",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: af67bxzg, hh6rzjmu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 422.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 422.5 110.9 12.7 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/98ogtog7",
   "source_url": "https://omlx.ai/benchmarks/performance/98ogtog7",
   "tg_tokens": null,
   "tps": 110.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-98ogtog7-ctx8192",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: af67bxzg, hh6rzjmu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4911.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 4,911 104.3 13.4 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/98ogtog7",
   "source_url": "https://omlx.ai/benchmarks/performance/98ogtog7",
   "tg_tokens": null,
   "tps": 104.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-98ogtog7-ctx16384",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.6 GB at this context; duplicate traces of the same session: af67bxzg, hh6rzjmu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4519.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 4,519 99.4 13.6 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/98ogtog7",
   "source_url": "https://omlx.ai/benchmarks/performance/98ogtog7",
   "tg_tokens": null,
   "tps": 99.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-98ogtog7-ctx32768",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.9 GB at this context; duplicate traces of the same session: af67bxzg, hh6rzjmu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3947.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 32k 3,947 89.5 13.9 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/98ogtog7",
   "source_url": "https://omlx.ai/benchmarks/performance/98ogtog7",
   "tg_tokens": null,
   "tps": 89.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-98ogtog7-ctx65536",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 14.5 GB at this context; duplicate traces of the same session: af67bxzg, hh6rzjmu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3024.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 64k 3,024 74.8 14.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/98ogtog7",
   "source_url": "https://omlx.ai/benchmarks/performance/98ogtog7",
   "tg_tokens": null,
   "tps": 74.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-98ogtog7-ctx131072",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: af67bxzg, hh6rzjmu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2041.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 128k 2,041 57.7 15.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/98ogtog7",
   "source_url": "https://omlx.ai/benchmarks/performance/98ogtog7",
   "tg_tokens": null,
   "tps": 57.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-98ogtog7-ctx199680",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: af67bxzg, hh6rzjmu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1519.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 195k 1,519 46.7 17.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/98ogtog7",
   "source_url": "https://omlx.ai/benchmarks/performance/98ogtog7",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-80yk5q3o-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 65.8 GB at this context; duplicate traces of the same session: v7cjlwq6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1951.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.6-35B-A3B-bf16-pt on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 bf16; 1k 1,951 70.9 65.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/80yk5q3o",
   "source_url": "https://omlx.ai/benchmarks/performance/80yk5q3o",
   "tg_tokens": null,
   "tps": 70.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-80yk5q3o-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 67.5 GB at this context; duplicate traces of the same session: v7cjlwq6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2610.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.6-35B-A3B-bf16-pt on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 bf16; 4k 2,610 68.5 67.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/80yk5q3o",
   "source_url": "https://omlx.ai/benchmarks/performance/80yk5q3o",
   "tg_tokens": null,
   "tps": 68.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-cbwk4xyl-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: txxfrvo7; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 412.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 412.7 28.1 18.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cbwk4xyl",
   "source_url": "https://omlx.ai/benchmarks/performance/cbwk4xyl",
   "tg_tokens": null,
   "tps": 28.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-cbwk4xyl-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: txxfrvo7; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 381.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 381.6 21.1 19.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cbwk4xyl",
   "source_url": "https://omlx.ai/benchmarks/performance/cbwk4xyl",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-a0ehukq5-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: ckuko7vg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 411.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 411.3 22.7 18.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a0ehukq5",
   "source_url": "https://omlx.ai/benchmarks/performance/a0ehukq5",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-a0ehukq5-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: ckuko7vg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 374.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 374.6 25.1 19.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a0ehukq5",
   "source_url": "https://omlx.ai/benchmarks/performance/a0ehukq5",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ftbo9vuq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 843.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 843.0 17.0 29.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ftbo9vuq",
   "source_url": "https://omlx.ai/benchmarks/performance/ftbo9vuq",
   "tg_tokens": null,
   "tps": 17.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ftbo9vuq-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 835.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 16k 835.9 16.1 30.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ftbo9vuq",
   "source_url": "https://omlx.ai/benchmarks/performance/ftbo9vuq",
   "tg_tokens": null,
   "tps": 16.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-cg9qj4nj-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 33.9 GB; TTFT 3370 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 303.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k tokens 303.9 PP tok/s 36.5 TG tok/s 3370 TTFT (ms) 33.9 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cg9qj4nj",
   "source_url": "https://omlx.ai/benchmarks/performance/cg9qj4nj",
   "tg_tokens": null,
   "tps": 36.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-iexvmf1z-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 44.4 GB at this context; duplicate traces of the same session: bsvojzdu, oj8833vv, ps1zbl25; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 629.9,
   "provenance": "community",
   "quant": "GSQ-RCO-Q2",
   "quote": "Qwen3.8-Flash-Next-GSQ-RCO-Q2-mlx on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 32k 629.9 47.7 44.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iexvmf1z",
   "source_url": "https://omlx.ai/benchmarks/performance/iexvmf1z",
   "tg_tokens": null,
   "tps": 47.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-iexvmf1z-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 46.3 GB at this context; duplicate traces of the same session: bsvojzdu, oj8833vv, ps1zbl25; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 626.9,
   "provenance": "community",
   "quant": "GSQ-RCO-Q2",
   "quote": "Qwen3.8-Flash-Next-GSQ-RCO-Q2-mlx on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 64k 626.9 41.1 46.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iexvmf1z",
   "source_url": "https://omlx.ai/benchmarks/performance/iexvmf1z",
   "tg_tokens": null,
   "tps": 41.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-iexvmf1z-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 50.7 GB at this context; duplicate traces of the same session: bsvojzdu, oj8833vv, ps1zbl25; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 613.4,
   "provenance": "community",
   "quant": "GSQ-RCO-Q2",
   "quote": "Qwen3.8-Flash-Next-GSQ-RCO-Q2-mlx on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 128k 613.4 40.7 50.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iexvmf1z",
   "source_url": "https://omlx.ai/benchmarks/performance/iexvmf1z",
   "tg_tokens": null,
   "tps": 40.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-iexvmf1z-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 52.1 GB at this context; duplicate traces of the same session: bsvojzdu, oj8833vv, ps1zbl25; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 604.3,
   "provenance": "community",
   "quant": "GSQ-RCO-Q2",
   "quote": "Qwen3.8-Flash-Next-GSQ-RCO-Q2-mlx on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 195k 604.3 42.8 52.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iexvmf1z",
   "source_url": "https://omlx.ai/benchmarks/performance/iexvmf1z",
   "tg_tokens": null,
   "tps": 42.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-nyhn0nqo-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 192.0 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 892.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-Flash-Next-MLX-8bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 892.2 57.6 192.0 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nyhn0nqo",
   "source_url": "https://omlx.ai/benchmarks/performance/nyhn0nqo",
   "tg_tokens": null,
   "tps": 57.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nyhn0nqo-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 192.4 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1194.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-Flash-Next-MLX-8bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 1,194 57.5 192.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nyhn0nqo",
   "source_url": "https://omlx.ai/benchmarks/performance/nyhn0nqo",
   "tg_tokens": null,
   "tps": 57.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-nyhn0nqo-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 192.8 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1206.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-Flash-Next-MLX-8bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k 1,206 53.9 192.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nyhn0nqo",
   "source_url": "https://omlx.ai/benchmarks/performance/nyhn0nqo",
   "tg_tokens": null,
   "tps": 53.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ylr8ejhh-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 191.9 GB at this context; duplicate traces of the same session: y857coyt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 906.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-Flash-Next-MLX-8bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 906.9 56.4 191.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ylr8ejhh",
   "source_url": "https://omlx.ai/benchmarks/performance/ylr8ejhh",
   "tg_tokens": null,
   "tps": 56.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ylr8ejhh-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 192.4 GB at this context; duplicate traces of the same session: y857coyt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1194.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-Flash-Next-MLX-8bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 1,194 56.4 192.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ylr8ejhh",
   "source_url": "https://omlx.ai/benchmarks/performance/ylr8ejhh",
   "tg_tokens": null,
   "tps": 56.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ylr8ejhh-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 192.5 GB at this context; duplicate traces of the same session: y857coyt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1207.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-Flash-Next-MLX-8bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 1,207 60.9 192.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ylr8ejhh",
   "source_url": "https://omlx.ai/benchmarks/performance/ylr8ejhh",
   "tg_tokens": null,
   "tps": 60.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5gxcmi0i-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 190.5 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 906.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ8-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 906.0 46.8 190.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5gxcmi0i",
   "source_url": "https://omlx.ai/benchmarks/performance/5gxcmi0i",
   "tg_tokens": null,
   "tps": 46.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5gxcmi0i-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 190.8 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1204.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ8-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 1,204 56.7 190.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5gxcmi0i",
   "source_url": "https://omlx.ai/benchmarks/performance/5gxcmi0i",
   "tg_tokens": null,
   "tps": 56.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5gxcmi0i-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 191.0 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1211.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ8-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 1,211 55.7 191.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5gxcmi0i",
   "source_url": "https://omlx.ai/benchmarks/performance/5gxcmi0i",
   "tg_tokens": null,
   "tps": 55.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gma4bv9e-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 122.5 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 818.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 1k 818.4 81.8 122.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gma4bv9e",
   "source_url": "https://omlx.ai/benchmarks/performance/gma4bv9e",
   "tg_tokens": null,
   "tps": 81.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gma4bv9e-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 122.8 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 997.8,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 4k 997.8 74.5 122.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gma4bv9e",
   "source_url": "https://omlx.ai/benchmarks/performance/gma4bv9e",
   "tg_tokens": null,
   "tps": 74.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-b0qeppql-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 128.1 GB at this context; duplicate traces of the same session: w52h1jx7; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 109.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 1k 109.0 18.9 128.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b0qeppql",
   "source_url": "https://omlx.ai/benchmarks/performance/b0qeppql",
   "tg_tokens": null,
   "tps": 18.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-b0qeppql-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 128.1 GB at this context; duplicate traces of the same session: w52h1jx7; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 445.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 4k 445.4 28.3 128.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b0qeppql",
   "source_url": "https://omlx.ai/benchmarks/performance/b0qeppql",
   "tg_tokens": null,
   "tps": 28.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-w6m2d2au-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 66.2 GB at this context; duplicate traces of the same session: mjvnpstt; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 78.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 1k 78.8 17.2 66.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w6m2d2au",
   "source_url": "https://omlx.ai/benchmarks/performance/w6m2d2au",
   "tg_tokens": null,
   "tps": 17.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-w6m2d2au-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 66.6 GB at this context; duplicate traces of the same session: mjvnpstt; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 216.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 4k 216.3 18.2 66.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w6m2d2au",
   "source_url": "https://omlx.ai/benchmarks/performance/w6m2d2au",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-w6m2d2au-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 67.2 GB at this context; duplicate traces of the same session: mjvnpstt; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 138.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 16k 138.6 14.2 67.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w6m2d2au",
   "source_url": "https://omlx.ai/benchmarks/performance/w6m2d2au",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-w6m2d2au-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 68.3 GB at this context; duplicate traces of the same session: mjvnpstt; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 64k 104.8 17.4 68.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w6m2d2au",
   "source_url": "https://omlx.ai/benchmarks/performance/w6m2d2au",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-mkiv2sfa-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1103.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,103 104.5 100.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mkiv2sfa",
   "source_url": "https://omlx.ai/benchmarks/performance/mkiv2sfa",
   "tg_tokens": null,
   "tps": 104.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-mkiv2sfa-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1369.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,369 82.6 101.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mkiv2sfa",
   "source_url": "https://omlx.ai/benchmarks/performance/mkiv2sfa",
   "tg_tokens": null,
   "tps": 82.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-mkiv2sfa-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1376.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 1,376 83.8 101.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mkiv2sfa",
   "source_url": "https://omlx.ai/benchmarks/performance/mkiv2sfa",
   "tg_tokens": null,
   "tps": 83.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-hbr7gshi-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1104.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,104 122.3 100.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hbr7gshi",
   "source_url": "https://omlx.ai/benchmarks/performance/hbr7gshi",
   "tg_tokens": null,
   "tps": 122.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hbr7gshi-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1377.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,377 97.1 101.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hbr7gshi",
   "source_url": "https://omlx.ai/benchmarks/performance/hbr7gshi",
   "tg_tokens": null,
   "tps": 97.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hbr7gshi-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1375.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 1,375 88.9 101.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hbr7gshi",
   "source_url": "https://omlx.ai/benchmarks/performance/hbr7gshi",
   "tg_tokens": null,
   "tps": 88.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hbr7gshi-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.6 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1369.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 1,369 73.1 101.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hbr7gshi",
   "source_url": "https://omlx.ai/benchmarks/performance/hbr7gshi",
   "tg_tokens": null,
   "tps": 73.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-pcvxzkgc-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-VL-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.5 GB at this context; duplicate traces of the same session: au77fpxx, lxhjezbb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 244.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-VL-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 244.8 40.9 32.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pcvxzkgc",
   "source_url": "https://omlx.ai/benchmarks/performance/pcvxzkgc",
   "tg_tokens": null,
   "tps": 40.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-pcvxzkgc-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-VL-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.7 GB at this context; duplicate traces of the same session: au77fpxx, lxhjezbb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 276.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-VL-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 276.1 43.8 35.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pcvxzkgc",
   "source_url": "https://omlx.ai/benchmarks/performance/pcvxzkgc",
   "tg_tokens": null,
   "tps": 43.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-pcvxzkgc-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-VL-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: au77fpxx, lxhjezbb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 275.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-VL-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 8k 275.7 37.6 35.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pcvxzkgc",
   "source_url": "https://omlx.ai/benchmarks/performance/pcvxzkgc",
   "tg_tokens": null,
   "tps": 37.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-pcvxzkgc-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-VL-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.5 GB at this context; duplicate traces of the same session: au77fpxx, lxhjezbb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 250.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-VL-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k 250.5 32.6 31.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pcvxzkgc",
   "source_url": "https://omlx.ai/benchmarks/performance/pcvxzkgc",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-pcvxzkgc-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-VL-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.3 GB at this context; duplicate traces of the same session: au77fpxx, lxhjezbb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 232.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-VL-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 232.4 30.3 33.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pcvxzkgc",
   "source_url": "https://omlx.ai/benchmarks/performance/pcvxzkgc",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-l780nzep-ctx1024",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "YCF-AI MTP FP16 build; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: bt8tukfx; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 202.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "YCF-AI--Qwen3.8-27B-Huihui-abliterated-oQ4e-MTP-FP16-MLX on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 202.7 28.1 18.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l780nzep",
   "source_url": "https://omlx.ai/benchmarks/performance/l780nzep",
   "tg_tokens": null,
   "tps": 28.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-l780nzep-ctx4096",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "YCF-AI MTP FP16 build; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: bt8tukfx; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "YCF-AI--Qwen3.8-27B-Huihui-abliterated-oQ4e-MTP-FP16-MLX on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 205.5 30.9 19.8 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l780nzep",
   "source_url": "https://omlx.ai/benchmarks/performance/l780nzep",
   "tg_tokens": null,
   "tps": 30.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4wmrz3x7-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Youssofal MTPLX Optimized Speed; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: 1z399iss, q9lh2wd2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 422.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Youssofal--Qwen3.8-27B-MTPLX-Optimized-Speed on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 422.9 68.9 24.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4wmrz3x7",
   "source_url": "https://omlx.ai/benchmarks/performance/4wmrz3x7",
   "tg_tokens": null,
   "tps": 68.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4wmrz3x7-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "Youssofal MTPLX Optimized Speed; oMLX community benchmark trace; peak memory 27.6 GB at this context; duplicate traces of the same session: 1z399iss, q9lh2wd2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 442.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Youssofal--Qwen3.8-27B-MTPLX-Optimized-Speed on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 442.2 58.8 27.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4wmrz3x7",
   "source_url": "https://omlx.ai/benchmarks/performance/4wmrz3x7",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4wmrz3x7-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "Youssofal MTPLX Optimized Speed; oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: 1z399iss, q9lh2wd2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 436.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Youssofal--Qwen3.8-27B-MTPLX-Optimized-Speed on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 436.4 80.2 28.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4wmrz3x7",
   "source_url": "https://omlx.ai/benchmarks/performance/4wmrz3x7",
   "tg_tokens": null,
   "tps": 80.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4wmrz3x7-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "Youssofal MTPLX Optimized Speed; oMLX community benchmark trace; peak memory 28.8 GB at this context; duplicate traces of the same session: 1z399iss, q9lh2wd2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 420.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Youssofal--Qwen3.8-27B-MTPLX-Optimized-Speed on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 420.9 74.9 28.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4wmrz3x7",
   "source_url": "https://omlx.ai/benchmarks/performance/4wmrz3x7",
   "tg_tokens": null,
   "tps": 74.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-3uv4sl8m-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "chriswessels ThinkingCap; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 4wve2rv9; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "chriswessels--ThinkingCap-Qwen3.8-27B-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 204.4 37.5 18.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3uv4sl8m",
   "source_url": "https://omlx.ai/benchmarks/performance/3uv4sl8m",
   "tg_tokens": null,
   "tps": 37.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-3uv4sl8m-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "chriswessels ThinkingCap; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 4wve2rv9; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 206.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "chriswessels--ThinkingCap-Qwen3.8-27B-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 206.0 38.6 19.8 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3uv4sl8m",
   "source_url": "https://omlx.ai/benchmarks/performance/3uv4sl8m",
   "tg_tokens": null,
   "tps": 38.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-3uv4sl8m-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "chriswessels ThinkingCap; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: 4wve2rv9; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 202.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "chriswessels--ThinkingCap-Qwen3.8-27B-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 202.2 36.7 20.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3uv4sl8m",
   "source_url": "https://omlx.ai/benchmarks/performance/3uv4sl8m",
   "tg_tokens": null,
   "tps": 36.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-96naazzq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "chriswessels ThinkingCap; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "chriswessels--ThinkingCap-Qwen3.8-27B-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 204.0 45.0 18.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/96naazzq",
   "source_url": "https://omlx.ai/benchmarks/performance/96naazzq",
   "tg_tokens": null,
   "tps": 45.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-96naazzq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "chriswessels ThinkingCap; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "chriswessels--ThinkingCap-Qwen3.8-27B-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 205.9 39.2 19.8 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/96naazzq",
   "source_url": "https://omlx.ai/benchmarks/performance/96naazzq",
   "tg_tokens": null,
   "tps": 39.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-4zmecmt3-ctx65536",
   "model": "Qwen3.6-35B-A3B-OptiQ",
   "notes": "dawncr0w; oMLX community benchmark trace, single run at 64k context; peak memory 23.2 GB; TTFT 402775 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 162.7,
   "provenance": "community",
   "quant": "5bpw",
   "quote": "dawncr0w--Qwen3.6-35B-A3B-OptiQ-5bpw-MLX on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 64k tokens 162.7 PP tok/s 28.0 TG tok/s 402775 TTFT (ms) 23.2 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4zmecmt3",
   "source_url": "https://omlx.ai/benchmarks/performance/4zmecmt3",
   "tg_tokens": null,
   "tps": 28.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-fs9riooi-ctx16384",
   "model": "Qwen3.6-35B-A3B-OptiQ",
   "notes": "dawncr0w; oMLX community benchmark trace, single run at 16k context; peak memory 21.0 GB; TTFT 53787 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 304.6,
   "provenance": "community",
   "quant": "5bpw",
   "quote": "dawncr0w--Qwen3.6-35B-A3B-OptiQ-5bpw-MLX on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k tokens 304.6 PP tok/s 42.3 TG tok/s 53787 TTFT (ms) 21.0 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fs9riooi",
   "source_url": "https://omlx.ai/benchmarks/performance/fs9riooi",
   "tg_tokens": null,
   "tps": 42.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-p73j72ib-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "fcmeyer MTP build; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "fcmeyer--Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 204.7 33.4 18.9 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p73j72ib",
   "source_url": "https://omlx.ai/benchmarks/performance/p73j72ib",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-p73j72ib-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "fcmeyer MTP build; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "fcmeyer--Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 205.9 38.0 19.8 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p73j72ib",
   "source_url": "https://omlx.ai/benchmarks/performance/p73j72ib",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2meabal7-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 14.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3941.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 3,941 105.6 14.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2meabal7",
   "source_url": "https://omlx.ai/benchmarks/performance/2meabal7",
   "tg_tokens": null,
   "tps": 105.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2meabal7-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 15.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3630.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 3,630 94.0 15.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2meabal7",
   "source_url": "https://omlx.ai/benchmarks/performance/2meabal7",
   "tg_tokens": null,
   "tps": 94.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1wmeknir-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2539.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26b-a4b-it-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 2,539 83.6 26.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1wmeknir",
   "source_url": "https://omlx.ai/benchmarks/performance/1wmeknir",
   "tg_tokens": null,
   "tps": 83.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1wmeknir-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3476.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26b-a4b-it-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 16k 3,476 75.3 26.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1wmeknir",
   "source_url": "https://omlx.ai/benchmarks/performance/1wmeknir",
   "tg_tokens": null,
   "tps": 75.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-xmmirmxn-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "leonsarmiento pi MTP; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "leonsarmiento--Qwen3.8-27B-pi-4bit-mtp-mlx on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 205.4 36.8 19.2 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xmmirmxn",
   "source_url": "https://omlx.ai/benchmarks/performance/xmmirmxn",
   "tg_tokens": null,
   "tps": 36.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-xmmirmxn-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "leonsarmiento pi MTP; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 206.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "leonsarmiento--Qwen3.8-27B-pi-4bit-mtp-mlx on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 206.8 42.0 20.2 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xmmirmxn",
   "source_url": "https://omlx.ai/benchmarks/performance/xmmirmxn",
   "tg_tokens": null,
   "tps": 42.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-7yw10f2h-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "leonsarmiento pi MTP; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "leonsarmiento--Qwen3.8-27B-pi-4bit-mtp-mlx on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 205.5 35.3 19.3 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7yw10f2h",
   "source_url": "https://omlx.ai/benchmarks/performance/7yw10f2h",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-7yw10f2h-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "leonsarmiento pi MTP; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 206.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "leonsarmiento--Qwen3.8-27B-pi-4bit-mtp-mlx on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 206.7 40.3 20.2 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7yw10f2h",
   "source_url": "https://omlx.ai/benchmarks/performance/7yw10f2h",
   "tg_tokens": null,
   "tps": 40.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-b5gftjry-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "mlx-community repack; oMLX community benchmark trace; peak memory 77.9 GB at this context; duplicate traces of the same session: j90q9zm9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1687.0,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "mlx-community--Qwen3-Coder-Next-mxfp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 1,687 83.2 77.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b5gftjry",
   "source_url": "https://omlx.ai/benchmarks/performance/b5gftjry",
   "tg_tokens": null,
   "tps": 83.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-b5gftjry-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "mlx-community repack; oMLX community benchmark trace; peak memory 78.7 GB at this context; duplicate traces of the same session: j90q9zm9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3189.0,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "mlx-community--Qwen3-Coder-Next-mxfp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 3,189 81.9 78.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b5gftjry",
   "source_url": "https://omlx.ai/benchmarks/performance/b5gftjry",
   "tg_tokens": null,
   "tps": 81.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-b5gftjry-ctx8192",
   "model": "Qwen3-Coder-Next",
   "notes": "mlx-community repack; oMLX community benchmark trace; peak memory 78.8 GB at this context; duplicate traces of the same session: j90q9zm9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3260.0,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "mlx-community--Qwen3-Coder-Next-mxfp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 3,260 80.0 78.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b5gftjry",
   "source_url": "https://omlx.ai/benchmarks/performance/b5gftjry",
   "tg_tokens": null,
   "tps": 80.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-b5gftjry-ctx16384",
   "model": "Qwen3-Coder-Next",
   "notes": "mlx-community repack; oMLX community benchmark trace; peak memory 79.0 GB at this context; duplicate traces of the same session: j90q9zm9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3037.0,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "mlx-community--Qwen3-Coder-Next-mxfp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 3,037 76.5 79.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b5gftjry",
   "source_url": "https://omlx.ai/benchmarks/performance/b5gftjry",
   "tg_tokens": null,
   "tps": 76.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ljzy79y4-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 331.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 255.8,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-oQ3e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 255.8 27.2 331.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ljzy79y4",
   "source_url": "https://omlx.ai/benchmarks/performance/ljzy79y4",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ljzy79y4-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 332.4 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 322.8,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-oQ3e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 322.8 25.7 332.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ljzy79y4",
   "source_url": "https://omlx.ai/benchmarks/performance/ljzy79y4",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ljzy79y4-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 332.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 348.4,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-oQ3e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 8k 348.4 28.3 332.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ljzy79y4",
   "source_url": "https://omlx.ai/benchmarks/performance/ljzy79y4",
   "tg_tokens": null,
   "tps": 28.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ljzy79y4-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 333.2 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.2,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-oQ3e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 16k 323.2 35.7 333.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ljzy79y4",
   "source_url": "https://omlx.ai/benchmarks/performance/ljzy79y4",
   "tg_tokens": null,
   "tps": 35.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-beyvecux-ctx1024",
   "model": "Devstral-Small-2-24B-Instruct-2512",
   "notes": "oMLX community benchmark trace; peak memory 14.1 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 457.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Devstral-Small-2-24B-Instruct-2512-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 457.8 19.9 14.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/beyvecux",
   "source_url": "https://omlx.ai/benchmarks/performance/beyvecux",
   "tg_tokens": null,
   "tps": 19.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-beyvecux-ctx4096",
   "model": "Devstral-Small-2-24B-Instruct-2512",
   "notes": "oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 478.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Devstral-Small-2-24B-Instruct-2512-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 478.4 19.1 14.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/beyvecux",
   "source_url": "https://omlx.ai/benchmarks/performance/beyvecux",
   "tg_tokens": null,
   "tps": 19.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lyb5gcby-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.8 GB at this context; duplicate traces of the same session: ac3vecvr, rn2lsr5a; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 499.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 499.8 35.6 175.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lyb5gcby",
   "source_url": "https://omlx.ai/benchmarks/performance/lyb5gcby",
   "tg_tokens": null,
   "tps": 35.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-lyb5gcby-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: ac3vecvr, rn2lsr5a; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 544.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 544.3 70.4 177.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lyb5gcby",
   "source_url": "https://omlx.ai/benchmarks/performance/lyb5gcby",
   "tg_tokens": null,
   "tps": 70.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lyb5gcby-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.0 GB at this context; duplicate traces of the same session: ac3vecvr, rn2lsr5a; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 526.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 526.1 37.5 178.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lyb5gcby",
   "source_url": "https://omlx.ai/benchmarks/performance/lyb5gcby",
   "tg_tokens": null,
   "tps": 37.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lyb5gcby-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.7 GB at this context; duplicate traces of the same session: ac3vecvr, rn2lsr5a; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 514.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 514.9 26.8 178.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lyb5gcby",
   "source_url": "https://omlx.ai/benchmarks/performance/lyb5gcby",
   "tg_tokens": null,
   "tps": 26.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-a135ad75-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.8 GB at this context; duplicate traces of the same session: 1lczbwsm, y4issq84; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 500.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 500.5 38.5 175.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a135ad75",
   "source_url": "https://omlx.ai/benchmarks/performance/a135ad75",
   "tg_tokens": null,
   "tps": 38.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-a135ad75-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: 1lczbwsm, y4issq84; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 545.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 545.0 25.4 177.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a135ad75",
   "source_url": "https://omlx.ai/benchmarks/performance/a135ad75",
   "tg_tokens": null,
   "tps": 25.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-a135ad75-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.7 GB at this context; duplicate traces of the same session: 1lczbwsm, y4issq84; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 514.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 514.4 64.4 178.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a135ad75",
   "source_url": "https://omlx.ai/benchmarks/performance/a135ad75",
   "tg_tokens": null,
   "tps": 64.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8ehvpcox-ctx32768",
   "model": "Laguna XS 2.1",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2553.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Laguna-XS-2.1-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 2,553 106.3 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8ehvpcox",
   "source_url": "https://omlx.ai/benchmarks/performance/8ehvpcox",
   "tg_tokens": null,
   "tps": 106.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8ehvpcox-ctx65536",
   "model": "Laguna XS 2.1",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1923.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Laguna-XS-2.1-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 1,923 80.8 20.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8ehvpcox",
   "source_url": "https://omlx.ai/benchmarks/performance/8ehvpcox",
   "tg_tokens": null,
   "tps": 80.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ewtcva3h-ctx1024",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: coa1ozft, v116sjvq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1273.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,273 113.6 22.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ewtcva3h",
   "source_url": "https://omlx.ai/benchmarks/performance/ewtcva3h",
   "tg_tokens": null,
   "tps": 113.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ewtcva3h-ctx4096",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: coa1ozft, v116sjvq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1556.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,556 109.6 23.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ewtcva3h",
   "source_url": "https://omlx.ai/benchmarks/performance/ewtcva3h",
   "tg_tokens": null,
   "tps": 109.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ewtcva3h-ctx8192",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: coa1ozft, v116sjvq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1407.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,407 107.5 23.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ewtcva3h",
   "source_url": "https://omlx.ai/benchmarks/performance/ewtcva3h",
   "tg_tokens": null,
   "tps": 107.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ewtcva3h-ctx16384",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: coa1ozft, v116sjvq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1333.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,333 98.2 23.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ewtcva3h",
   "source_url": "https://omlx.ai/benchmarks/performance/ewtcva3h",
   "tg_tokens": null,
   "tps": 98.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ewtcva3h-ctx32768",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: coa1ozft, v116sjvq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1160.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 1,160 89.9 24.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ewtcva3h",
   "source_url": "https://omlx.ai/benchmarks/performance/ewtcva3h",
   "tg_tokens": null,
   "tps": 89.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ewtcva3h-ctx65536",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: coa1ozft, v116sjvq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 977.1,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 977.1 77.4 24.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ewtcva3h",
   "source_url": "https://omlx.ai/benchmarks/performance/ewtcva3h",
   "tg_tokens": null,
   "tps": 77.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2",
   "id": "omlx-5y9i38og-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: 3m4ve5qg, 8lzw8hav; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 110.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M2 (8c); M2 (8c) \u00b7 16 GB \u00b7 4bit; 1k 110.4 18.8 5.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5y9i38og",
   "source_url": "https://omlx.ai/benchmarks/performance/5y9i38og",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2",
   "id": "omlx-5y9i38og-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: 3m4ve5qg, 8lzw8hav; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M2 (8c); M2 (8c) \u00b7 16 GB \u00b7 4bit; 4k 105.8 18.6 6.2 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5y9i38og",
   "source_url": "https://omlx.ai/benchmarks/performance/5y9i38og",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2",
   "id": "omlx-5y9i38og-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: 3m4ve5qg, 8lzw8hav; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 101.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M2 (8c); M2 (8c) \u00b7 16 GB \u00b7 4bit; 8k 101.5 17.8 6.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5y9i38og",
   "source_url": "https://omlx.ai/benchmarks/performance/5y9i38og",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-af67bxzg-ctx1024",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: hh6rzjmu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 422.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 422.5 110.9 12.7 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/af67bxzg",
   "source_url": "https://omlx.ai/benchmarks/performance/af67bxzg",
   "tg_tokens": null,
   "tps": 110.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-af67bxzg-ctx8192",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: hh6rzjmu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4911.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 4,911 104.3 13.4 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/af67bxzg",
   "source_url": "https://omlx.ai/benchmarks/performance/af67bxzg",
   "tg_tokens": null,
   "tps": 104.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-af67bxzg-ctx16384",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.6 GB at this context; duplicate traces of the same session: hh6rzjmu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4519.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 4,519 99.4 13.6 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/af67bxzg",
   "source_url": "https://omlx.ai/benchmarks/performance/af67bxzg",
   "tg_tokens": null,
   "tps": 99.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-af67bxzg-ctx32768",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.9 GB at this context; duplicate traces of the same session: hh6rzjmu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3947.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 32k 3,947 89.5 13.9 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/af67bxzg",
   "source_url": "https://omlx.ai/benchmarks/performance/af67bxzg",
   "tg_tokens": null,
   "tps": 89.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-af67bxzg-ctx65536",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 14.5 GB at this context; duplicate traces of the same session: hh6rzjmu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3024.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 64k 3,024 74.8 14.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/af67bxzg",
   "source_url": "https://omlx.ai/benchmarks/performance/af67bxzg",
   "tg_tokens": null,
   "tps": 74.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-af67bxzg-ctx131072",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: hh6rzjmu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2041.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 128k 2,041 57.7 15.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/af67bxzg",
   "source_url": "https://omlx.ai/benchmarks/performance/af67bxzg",
   "tg_tokens": null,
   "tps": 57.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-af67bxzg-ctx199680",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: hh6rzjmu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1519.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 195k 1,519 46.7 17.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/af67bxzg",
   "source_url": "https://omlx.ai/benchmarks/performance/af67bxzg",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-v7cjlwq6-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 65.8 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1951.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.6-35B-A3B-bf16-pt on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 bf16; 1k 1,951 70.9 65.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v7cjlwq6",
   "source_url": "https://omlx.ai/benchmarks/performance/v7cjlwq6",
   "tg_tokens": null,
   "tps": 70.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-v7cjlwq6-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 67.5 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2610.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.6-35B-A3B-bf16-pt on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 bf16; 4k 2,610 68.5 67.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v7cjlwq6",
   "source_url": "https://omlx.ai/benchmarks/performance/v7cjlwq6",
   "tg_tokens": null,
   "tps": 68.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-txxfrvo7-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 412.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 412.7 28.1 18.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/txxfrvo7",
   "source_url": "https://omlx.ai/benchmarks/performance/txxfrvo7",
   "tg_tokens": null,
   "tps": 28.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-txxfrvo7-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 381.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 381.6 21.1 19.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/txxfrvo7",
   "source_url": "https://omlx.ai/benchmarks/performance/txxfrvo7",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ckuko7vg-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 411.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 411.3 22.7 18.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ckuko7vg",
   "source_url": "https://omlx.ai/benchmarks/performance/ckuko7vg",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ckuko7vg-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 374.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 374.6 25.1 19.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ckuko7vg",
   "source_url": "https://omlx.ai/benchmarks/performance/ckuko7vg",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-bsvojzdu-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 44.4 GB at this context; duplicate traces of the same session: oj8833vv, ps1zbl25; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 629.9,
   "provenance": "community",
   "quant": "GSQ-RCO-Q2",
   "quote": "Qwen3.8-Flash-Next-GSQ-RCO-Q2-mlx on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 32k 629.9 47.7 44.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bsvojzdu",
   "source_url": "https://omlx.ai/benchmarks/performance/bsvojzdu",
   "tg_tokens": null,
   "tps": 47.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-bsvojzdu-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 46.3 GB at this context; duplicate traces of the same session: oj8833vv, ps1zbl25; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 626.9,
   "provenance": "community",
   "quant": "GSQ-RCO-Q2",
   "quote": "Qwen3.8-Flash-Next-GSQ-RCO-Q2-mlx on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 64k 626.9 41.1 46.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bsvojzdu",
   "source_url": "https://omlx.ai/benchmarks/performance/bsvojzdu",
   "tg_tokens": null,
   "tps": 41.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-bsvojzdu-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 50.7 GB at this context; duplicate traces of the same session: oj8833vv, ps1zbl25; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 613.4,
   "provenance": "community",
   "quant": "GSQ-RCO-Q2",
   "quote": "Qwen3.8-Flash-Next-GSQ-RCO-Q2-mlx on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 128k 613.4 40.7 50.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bsvojzdu",
   "source_url": "https://omlx.ai/benchmarks/performance/bsvojzdu",
   "tg_tokens": null,
   "tps": 40.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-bsvojzdu-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 52.1 GB at this context; duplicate traces of the same session: oj8833vv, ps1zbl25; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 604.3,
   "provenance": "community",
   "quant": "GSQ-RCO-Q2",
   "quote": "Qwen3.8-Flash-Next-GSQ-RCO-Q2-mlx on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 195k 604.3 42.8 52.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bsvojzdu",
   "source_url": "https://omlx.ai/benchmarks/performance/bsvojzdu",
   "tg_tokens": null,
   "tps": 42.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-y857coyt-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 191.9 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 906.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-Flash-Next-MLX-8bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 906.9 56.4 191.9 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y857coyt",
   "source_url": "https://omlx.ai/benchmarks/performance/y857coyt",
   "tg_tokens": null,
   "tps": 56.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-y857coyt-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 192.4 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1194.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-Flash-Next-MLX-8bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 1,194 56.4 192.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y857coyt",
   "source_url": "https://omlx.ai/benchmarks/performance/y857coyt",
   "tg_tokens": null,
   "tps": 56.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-y857coyt-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 192.5 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1207.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-Flash-Next-MLX-8bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 1,207 60.9 192.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y857coyt",
   "source_url": "https://omlx.ai/benchmarks/performance/y857coyt",
   "tg_tokens": null,
   "tps": 60.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rsayaueg-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 191.9 GB at this context; duplicate traces of the same session: qbk92re7, uvudo0ti, wy63q01k; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 852.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-Flash-Next-MLX-8bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 852.2 63.2 191.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rsayaueg",
   "source_url": "https://omlx.ai/benchmarks/performance/rsayaueg",
   "tg_tokens": null,
   "tps": 63.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rsayaueg-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 192.4 GB at this context; duplicate traces of the same session: qbk92re7, uvudo0ti, wy63q01k; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1197.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-Flash-Next-MLX-8bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 1,197 54.7 192.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rsayaueg",
   "source_url": "https://omlx.ai/benchmarks/performance/rsayaueg",
   "tg_tokens": null,
   "tps": 54.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rsayaueg-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 192.5 GB at this context; duplicate traces of the same session: qbk92re7, uvudo0ti, wy63q01k; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1208.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-Flash-Next-MLX-8bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 1,208 46.8 192.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rsayaueg",
   "source_url": "https://omlx.ai/benchmarks/performance/rsayaueg",
   "tg_tokens": null,
   "tps": 46.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rsayaueg-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 192.8 GB at this context; duplicate traces of the same session: qbk92re7, uvudo0ti, wy63q01k; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1209.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-Flash-Next-MLX-8bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k 1,209 59.8 192.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rsayaueg",
   "source_url": "https://omlx.ai/benchmarks/performance/rsayaueg",
   "tg_tokens": null,
   "tps": 59.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-tk0j9ndo-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 190.5 GB at this context; duplicate traces of the same session: 3n3ygdi3, ubmuhqpy; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 911.9,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ8-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 911.9 50.3 190.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tk0j9ndo",
   "source_url": "https://omlx.ai/benchmarks/performance/tk0j9ndo",
   "tg_tokens": null,
   "tps": 50.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-tk0j9ndo-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 190.8 GB at this context; duplicate traces of the same session: 3n3ygdi3, ubmuhqpy; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1202.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ8-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 1,202 56.2 190.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tk0j9ndo",
   "source_url": "https://omlx.ai/benchmarks/performance/tk0j9ndo",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-tk0j9ndo-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 191.0 GB at this context; duplicate traces of the same session: 3n3ygdi3, ubmuhqpy; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1197.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ8-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 1,197 49.2 191.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tk0j9ndo",
   "source_url": "https://omlx.ai/benchmarks/performance/tk0j9ndo",
   "tg_tokens": null,
   "tps": 49.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-w52h1jx7-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 128.1 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 109.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 1k 109.0 18.9 128.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w52h1jx7",
   "source_url": "https://omlx.ai/benchmarks/performance/w52h1jx7",
   "tg_tokens": null,
   "tps": 18.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-w52h1jx7-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 128.1 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 445.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 4k 445.4 28.3 128.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w52h1jx7",
   "source_url": "https://omlx.ai/benchmarks/performance/w52h1jx7",
   "tg_tokens": null,
   "tps": 28.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-mjvnpstt-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 66.2 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 78.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 1k 78.8 17.2 66.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mjvnpstt",
   "source_url": "https://omlx.ai/benchmarks/performance/mjvnpstt",
   "tg_tokens": null,
   "tps": 17.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-mjvnpstt-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 66.6 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 216.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 4k 216.3 18.2 66.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mjvnpstt",
   "source_url": "https://omlx.ai/benchmarks/performance/mjvnpstt",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-mjvnpstt-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 67.2 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 138.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 16k 138.6 14.2 67.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mjvnpstt",
   "source_url": "https://omlx.ai/benchmarks/performance/mjvnpstt",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-mjvnpstt-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 68.3 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 64k 104.8 17.4 68.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mjvnpstt",
   "source_url": "https://omlx.ai/benchmarks/performance/mjvnpstt",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-85s3s8pf-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: 9hsyd4z2, i2v21nmv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1104.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,104 122.3 100.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/85s3s8pf",
   "source_url": "https://omlx.ai/benchmarks/performance/85s3s8pf",
   "tg_tokens": null,
   "tps": 122.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-85s3s8pf-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: 9hsyd4z2, i2v21nmv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1377.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,377 97.1 101.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/85s3s8pf",
   "source_url": "https://omlx.ai/benchmarks/performance/85s3s8pf",
   "tg_tokens": null,
   "tps": 97.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-85s3s8pf-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: 9hsyd4z2, i2v21nmv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1375.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 1,375 88.9 101.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/85s3s8pf",
   "source_url": "https://omlx.ai/benchmarks/performance/85s3s8pf",
   "tg_tokens": null,
   "tps": 88.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-85s3s8pf-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.6 GB at this context; duplicate traces of the same session: 9hsyd4z2, i2v21nmv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1369.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 1,369 73.1 101.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/85s3s8pf",
   "source_url": "https://omlx.ai/benchmarks/performance/85s3s8pf",
   "tg_tokens": null,
   "tps": 73.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-igwouq4g-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 100.7 GB; TTFT 923 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1110.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k tokens 1,110 PP tok/s 104.9 TG tok/s 923 TTFT (ms) 100.7 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/igwouq4g",
   "source_url": "https://omlx.ai/benchmarks/performance/igwouq4g",
   "tg_tokens": null,
   "tps": 104.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9p070b21-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: h7zv7la6, n82bayk6, wjw2fhh6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1123.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,123 87.7 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9p070b21",
   "source_url": "https://omlx.ai/benchmarks/performance/9p070b21",
   "tg_tokens": null,
   "tps": 87.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9p070b21-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: h7zv7la6, n82bayk6, wjw2fhh6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2475.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,475 65.9 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9p070b21",
   "source_url": "https://omlx.ai/benchmarks/performance/9p070b21",
   "tg_tokens": null,
   "tps": 65.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9p070b21-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: h7zv7la6, n82bayk6, wjw2fhh6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2660.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,660 71.3 72.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9p070b21",
   "source_url": "https://omlx.ai/benchmarks/performance/9p070b21",
   "tg_tokens": null,
   "tps": 71.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9p070b21-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.5 GB at this context; duplicate traces of the same session: h7zv7la6, n82bayk6, wjw2fhh6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2263.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,263 63.1 80.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9p070b21",
   "source_url": "https://omlx.ai/benchmarks/performance/9p070b21",
   "tg_tokens": null,
   "tps": 63.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-lxhjezbb-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-VL-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.5 GB at this context; duplicate traces of the same session: au77fpxx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 244.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-VL-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 244.8 40.9 32.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lxhjezbb",
   "source_url": "https://omlx.ai/benchmarks/performance/lxhjezbb",
   "tg_tokens": null,
   "tps": 40.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-lxhjezbb-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-VL-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.7 GB at this context; duplicate traces of the same session: au77fpxx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 276.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-VL-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 276.1 43.8 35.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lxhjezbb",
   "source_url": "https://omlx.ai/benchmarks/performance/lxhjezbb",
   "tg_tokens": null,
   "tps": 43.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-lxhjezbb-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-VL-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: au77fpxx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 275.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-VL-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 8k 275.7 37.6 35.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lxhjezbb",
   "source_url": "https://omlx.ai/benchmarks/performance/lxhjezbb",
   "tg_tokens": null,
   "tps": 37.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-lxhjezbb-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-VL-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.5 GB at this context; duplicate traces of the same session: au77fpxx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 250.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-VL-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k 250.5 32.6 31.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lxhjezbb",
   "source_url": "https://omlx.ai/benchmarks/performance/lxhjezbb",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-lxhjezbb-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-VL-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.3 GB at this context; duplicate traces of the same session: au77fpxx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 232.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-VL-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 232.4 30.3 33.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lxhjezbb",
   "source_url": "https://omlx.ai/benchmarks/performance/lxhjezbb",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-bt8tukfx-ctx1024",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "YCF-AI MTP FP16 build; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 202.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "YCF-AI--Qwen3.8-27B-Huihui-abliterated-oQ4e-MTP-FP16-MLX on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 202.7 28.1 18.8 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bt8tukfx",
   "source_url": "https://omlx.ai/benchmarks/performance/bt8tukfx",
   "tg_tokens": null,
   "tps": 28.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-bt8tukfx-ctx4096",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "YCF-AI MTP FP16 build; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "YCF-AI--Qwen3.8-27B-Huihui-abliterated-oQ4e-MTP-FP16-MLX on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 205.5 30.9 19.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bt8tukfx",
   "source_url": "https://omlx.ai/benchmarks/performance/bt8tukfx",
   "tg_tokens": null,
   "tps": 30.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-q9lh2wd2-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Youssofal MTPLX Optimized Speed; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: 1z399iss; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 422.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Youssofal--Qwen3.8-27B-MTPLX-Optimized-Speed on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 422.9 68.9 24.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q9lh2wd2",
   "source_url": "https://omlx.ai/benchmarks/performance/q9lh2wd2",
   "tg_tokens": null,
   "tps": 68.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-q9lh2wd2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "Youssofal MTPLX Optimized Speed; oMLX community benchmark trace; peak memory 27.6 GB at this context; duplicate traces of the same session: 1z399iss; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 442.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Youssofal--Qwen3.8-27B-MTPLX-Optimized-Speed on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 442.2 58.8 27.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q9lh2wd2",
   "source_url": "https://omlx.ai/benchmarks/performance/q9lh2wd2",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-q9lh2wd2-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "Youssofal MTPLX Optimized Speed; oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: 1z399iss; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 436.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Youssofal--Qwen3.8-27B-MTPLX-Optimized-Speed on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 436.4 80.2 28.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q9lh2wd2",
   "source_url": "https://omlx.ai/benchmarks/performance/q9lh2wd2",
   "tg_tokens": null,
   "tps": 80.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-q9lh2wd2-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "Youssofal MTPLX Optimized Speed; oMLX community benchmark trace; peak memory 28.8 GB at this context; duplicate traces of the same session: 1z399iss; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 420.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Youssofal--Qwen3.8-27B-MTPLX-Optimized-Speed on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 420.9 74.9 28.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q9lh2wd2",
   "source_url": "https://omlx.ai/benchmarks/performance/q9lh2wd2",
   "tg_tokens": null,
   "tps": 74.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-4wve2rv9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "chriswessels ThinkingCap; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "chriswessels--ThinkingCap-Qwen3.8-27B-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 204.4 37.5 18.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4wve2rv9",
   "source_url": "https://omlx.ai/benchmarks/performance/4wve2rv9",
   "tg_tokens": null,
   "tps": 37.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-4wve2rv9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "chriswessels ThinkingCap; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 206.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "chriswessels--ThinkingCap-Qwen3.8-27B-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 206.0 38.6 19.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4wve2rv9",
   "source_url": "https://omlx.ai/benchmarks/performance/4wve2rv9",
   "tg_tokens": null,
   "tps": 38.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-4wve2rv9-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "chriswessels ThinkingCap; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 202.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "chriswessels--ThinkingCap-Qwen3.8-27B-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 202.2 36.7 20.5 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4wve2rv9",
   "source_url": "https://omlx.ai/benchmarks/performance/4wve2rv9",
   "tg_tokens": null,
   "tps": 36.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-arlv2jez-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "leonsarmiento pi MTP; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: hg14hnmd, i62jp0kk, l7osqd9p; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "leonsarmiento--Qwen3.8-27B-pi-4bit-mtp-mlx on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 204.8 31.1 19.3 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/arlv2jez",
   "source_url": "https://omlx.ai/benchmarks/performance/arlv2jez",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-arlv2jez-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "leonsarmiento pi MTP; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: hg14hnmd, i62jp0kk, l7osqd9p; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 206.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "leonsarmiento--Qwen3.8-27B-pi-4bit-mtp-mlx on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 206.2 36.6 20.2 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/arlv2jez",
   "source_url": "https://omlx.ai/benchmarks/performance/arlv2jez",
   "tg_tokens": null,
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-arlv2jez-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "leonsarmiento pi MTP; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: hg14hnmd, i62jp0kk, l7osqd9p; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 203.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "leonsarmiento--Qwen3.8-27B-pi-4bit-mtp-mlx on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 203.5 34.1 20.9 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/arlv2jez",
   "source_url": "https://omlx.ai/benchmarks/performance/arlv2jez",
   "tg_tokens": null,
   "tps": 34.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-arlv2jez-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "leonsarmiento pi MTP; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: hg14hnmd, i62jp0kk, l7osqd9p; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 195.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "leonsarmiento--Qwen3.8-27B-pi-4bit-mtp-mlx on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 195.0 30.6 20.7 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/arlv2jez",
   "source_url": "https://omlx.ai/benchmarks/performance/arlv2jez",
   "tg_tokens": null,
   "tps": 30.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-j90q9zm9-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "mlx-community repack; oMLX community benchmark trace; peak memory 77.9 GB at this context; duplicate traces of the same session: 92zcatgs, 9guweh5o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1687.0,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "mlx-community--Qwen3-Coder-Next-mxfp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 1,687 83.2 77.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j90q9zm9",
   "source_url": "https://omlx.ai/benchmarks/performance/j90q9zm9",
   "tg_tokens": null,
   "tps": 83.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-j90q9zm9-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "mlx-community repack; oMLX community benchmark trace; peak memory 78.7 GB at this context; duplicate traces of the same session: 92zcatgs, 9guweh5o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3189.0,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "mlx-community--Qwen3-Coder-Next-mxfp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 3,189 81.9 78.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j90q9zm9",
   "source_url": "https://omlx.ai/benchmarks/performance/j90q9zm9",
   "tg_tokens": null,
   "tps": 81.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-j90q9zm9-ctx8192",
   "model": "Qwen3-Coder-Next",
   "notes": "mlx-community repack; oMLX community benchmark trace; peak memory 78.8 GB at this context; duplicate traces of the same session: 92zcatgs, 9guweh5o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3260.0,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "mlx-community--Qwen3-Coder-Next-mxfp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 3,260 80.0 78.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j90q9zm9",
   "source_url": "https://omlx.ai/benchmarks/performance/j90q9zm9",
   "tg_tokens": null,
   "tps": 80.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-j90q9zm9-ctx16384",
   "model": "Qwen3-Coder-Next",
   "notes": "mlx-community repack; oMLX community benchmark trace; peak memory 79.0 GB at this context; duplicate traces of the same session: 92zcatgs, 9guweh5o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3037.0,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "mlx-community--Qwen3-Coder-Next-mxfp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 3,037 76.5 79.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j90q9zm9",
   "source_url": "https://omlx.ai/benchmarks/performance/j90q9zm9",
   "tg_tokens": null,
   "tps": 76.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-mbc245nf-ctx1024",
   "model": "Devstral-Small-2-24B-Instruct-2512",
   "notes": "oMLX community benchmark trace; peak memory 14.1 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 457.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Devstral-Small-2-24B-Instruct-2512-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 457.8 19.9 14.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mbc245nf",
   "source_url": "https://omlx.ai/benchmarks/performance/mbc245nf",
   "tg_tokens": null,
   "tps": 19.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-mbc245nf-ctx4096",
   "model": "Devstral-Small-2-24B-Instruct-2512",
   "notes": "oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 478.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Devstral-Small-2-24B-Instruct-2512-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 478.4 19.1 14.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mbc245nf",
   "source_url": "https://omlx.ai/benchmarks/performance/mbc245nf",
   "tg_tokens": null,
   "tps": 19.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rn2lsr5a-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.8 GB at this context; duplicate traces of the same session: ac3vecvr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 499.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 499.8 35.6 175.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rn2lsr5a",
   "source_url": "https://omlx.ai/benchmarks/performance/rn2lsr5a",
   "tg_tokens": null,
   "tps": 35.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rn2lsr5a-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: ac3vecvr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 544.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 544.3 70.4 177.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rn2lsr5a",
   "source_url": "https://omlx.ai/benchmarks/performance/rn2lsr5a",
   "tg_tokens": null,
   "tps": 70.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rn2lsr5a-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.0 GB at this context; duplicate traces of the same session: ac3vecvr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 526.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 526.1 37.5 178.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rn2lsr5a",
   "source_url": "https://omlx.ai/benchmarks/performance/rn2lsr5a",
   "tg_tokens": null,
   "tps": 37.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rn2lsr5a-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.7 GB at this context; duplicate traces of the same session: ac3vecvr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 514.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 514.9 26.8 178.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rn2lsr5a",
   "source_url": "https://omlx.ai/benchmarks/performance/rn2lsr5a",
   "tg_tokens": null,
   "tps": 26.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1lczbwsm-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.8 GB at this context; duplicate traces of the same session: y4issq84; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 500.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 500.5 38.5 175.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1lczbwsm",
   "source_url": "https://omlx.ai/benchmarks/performance/1lczbwsm",
   "tg_tokens": null,
   "tps": 38.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1lczbwsm-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: y4issq84; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 545.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 545.0 25.4 177.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1lczbwsm",
   "source_url": "https://omlx.ai/benchmarks/performance/1lczbwsm",
   "tg_tokens": null,
   "tps": 25.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1lczbwsm-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.7 GB at this context; duplicate traces of the same session: y4issq84; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 514.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 514.4 64.4 178.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1lczbwsm",
   "source_url": "https://omlx.ai/benchmarks/performance/1lczbwsm",
   "tg_tokens": null,
   "tps": 64.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-kxx2qmon-ctx16384",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: 23y21y40; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1457.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 1,457 93.9 34.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kxx2qmon",
   "source_url": "https://omlx.ai/benchmarks/performance/kxx2qmon",
   "tg_tokens": null,
   "tps": 93.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-kxx2qmon-ctx32768",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: 23y21y40; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1293.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 1,293 90.4 34.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kxx2qmon",
   "source_url": "https://omlx.ai/benchmarks/performance/kxx2qmon",
   "tg_tokens": null,
   "tps": 90.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-3cmu4wcx-ctx16384",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: b8zjrei1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1459.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 1,459 94.2 34.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3cmu4wcx",
   "source_url": "https://omlx.ai/benchmarks/performance/3cmu4wcx",
   "tg_tokens": null,
   "tps": 94.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-3cmu4wcx-ctx32768",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: b8zjrei1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1246.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 1,246 88.9 34.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3cmu4wcx",
   "source_url": "https://omlx.ai/benchmarks/performance/3cmu4wcx",
   "tg_tokens": null,
   "tps": 88.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-coa1ozft-ctx1024",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: v116sjvq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1273.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,273 113.6 22.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/coa1ozft",
   "source_url": "https://omlx.ai/benchmarks/performance/coa1ozft",
   "tg_tokens": null,
   "tps": 113.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-coa1ozft-ctx4096",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: v116sjvq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1556.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,556 109.6 23.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/coa1ozft",
   "source_url": "https://omlx.ai/benchmarks/performance/coa1ozft",
   "tg_tokens": null,
   "tps": 109.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-coa1ozft-ctx8192",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: v116sjvq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1407.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,407 107.5 23.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/coa1ozft",
   "source_url": "https://omlx.ai/benchmarks/performance/coa1ozft",
   "tg_tokens": null,
   "tps": 107.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-coa1ozft-ctx16384",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: v116sjvq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1333.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,333 98.2 23.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/coa1ozft",
   "source_url": "https://omlx.ai/benchmarks/performance/coa1ozft",
   "tg_tokens": null,
   "tps": 98.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-coa1ozft-ctx32768",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: v116sjvq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1160.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 1,160 89.9 24.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/coa1ozft",
   "source_url": "https://omlx.ai/benchmarks/performance/coa1ozft",
   "tg_tokens": null,
   "tps": 89.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-coa1ozft-ctx65536",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: v116sjvq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 977.1,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 977.1 77.4 24.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/coa1ozft",
   "source_url": "https://omlx.ai/benchmarks/performance/coa1ozft",
   "tg_tokens": null,
   "tps": 77.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2",
   "id": "omlx-8lzw8hav-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: 3m4ve5qg; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 110.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M2 (8c); M2 (8c) \u00b7 16 GB \u00b7 4bit; 1k 110.4 18.8 5.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8lzw8hav",
   "source_url": "https://omlx.ai/benchmarks/performance/8lzw8hav",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2",
   "id": "omlx-8lzw8hav-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: 3m4ve5qg; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M2 (8c); M2 (8c) \u00b7 16 GB \u00b7 4bit; 4k 105.8 18.6 6.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8lzw8hav",
   "source_url": "https://omlx.ai/benchmarks/performance/8lzw8hav",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2",
   "id": "omlx-8lzw8hav-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: 3m4ve5qg; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 101.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M2 (8c); M2 (8c) \u00b7 16 GB \u00b7 4bit; 8k 101.5 17.8 6.4 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8lzw8hav",
   "source_url": "https://omlx.ai/benchmarks/performance/8lzw8hav",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-qld2bshe-ctx1024",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: ojsbxllo; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 518.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 518.7 34.1 12.7 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qld2bshe",
   "source_url": "https://omlx.ai/benchmarks/performance/qld2bshe",
   "tg_tokens": null,
   "tps": 34.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-qld2bshe-ctx4096",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.3 GB at this context; duplicate traces of the same session: ojsbxllo; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 532.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 532.5 33.3 13.3 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qld2bshe",
   "source_url": "https://omlx.ai/benchmarks/performance/qld2bshe",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hh6rzjmu-ctx1024",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 422.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 422.5 110.9 12.7 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hh6rzjmu",
   "source_url": "https://omlx.ai/benchmarks/performance/hh6rzjmu",
   "tg_tokens": null,
   "tps": 110.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hh6rzjmu-ctx8192",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4911.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 4,911 104.3 13.4 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hh6rzjmu",
   "source_url": "https://omlx.ai/benchmarks/performance/hh6rzjmu",
   "tg_tokens": null,
   "tps": 104.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hh6rzjmu-ctx16384",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.6 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4519.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 4,519 99.4 13.6 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hh6rzjmu",
   "source_url": "https://omlx.ai/benchmarks/performance/hh6rzjmu",
   "tg_tokens": null,
   "tps": 99.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hh6rzjmu-ctx32768",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.9 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3947.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 32k 3,947 89.5 13.9 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hh6rzjmu",
   "source_url": "https://omlx.ai/benchmarks/performance/hh6rzjmu",
   "tg_tokens": null,
   "tps": 89.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hh6rzjmu-ctx65536",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 14.5 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3024.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 64k 3,024 74.8 14.5 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hh6rzjmu",
   "source_url": "https://omlx.ai/benchmarks/performance/hh6rzjmu",
   "tg_tokens": null,
   "tps": 74.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hh6rzjmu-ctx131072",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2041.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 128k 2,041 57.7 15.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hh6rzjmu",
   "source_url": "https://omlx.ai/benchmarks/performance/hh6rzjmu",
   "tg_tokens": null,
   "tps": 57.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hh6rzjmu-ctx199680",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1519.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 195k 1,519 46.7 17.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hh6rzjmu",
   "source_url": "https://omlx.ai/benchmarks/performance/hh6rzjmu",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-g9u4efup-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: cks3y95n; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 429.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 429.5 10.1 28.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g9u4efup",
   "source_url": "https://omlx.ai/benchmarks/performance/g9u4efup",
   "tg_tokens": null,
   "tps": 10.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-g9u4efup-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: cks3y95n; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 470.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 470.6 10.0 29.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g9u4efup",
   "source_url": "https://omlx.ai/benchmarks/performance/g9u4efup",
   "tg_tokens": null,
   "tps": 10.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-w08l115u-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 33.1 GB at this context; duplicate traces of the same session: nqyd2uyx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 216.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 216.7 15.3 33.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w08l115u",
   "source_url": "https://omlx.ai/benchmarks/performance/w08l115u",
   "tg_tokens": null,
   "tps": 15.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-w08l115u-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: nqyd2uyx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 209.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 209.2 14.9 34.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w08l115u",
   "source_url": "https://omlx.ai/benchmarks/performance/w08l115u",
   "tg_tokens": null,
   "tps": 14.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-ps1zbl25-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 44.4 GB at this context; duplicate traces of the same session: oj8833vv; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 629.9,
   "provenance": "community",
   "quant": "GSQ-RCO-Q2",
   "quote": "Qwen3.8-Flash-Next-GSQ-RCO-Q2-mlx on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 32k 629.9 47.7 44.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ps1zbl25",
   "source_url": "https://omlx.ai/benchmarks/performance/ps1zbl25",
   "tg_tokens": null,
   "tps": 47.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-ps1zbl25-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 46.3 GB at this context; duplicate traces of the same session: oj8833vv; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 626.9,
   "provenance": "community",
   "quant": "GSQ-RCO-Q2",
   "quote": "Qwen3.8-Flash-Next-GSQ-RCO-Q2-mlx on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 64k 626.9 41.1 46.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ps1zbl25",
   "source_url": "https://omlx.ai/benchmarks/performance/ps1zbl25",
   "tg_tokens": null,
   "tps": 41.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-ps1zbl25-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 50.7 GB at this context; duplicate traces of the same session: oj8833vv; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 613.4,
   "provenance": "community",
   "quant": "GSQ-RCO-Q2",
   "quote": "Qwen3.8-Flash-Next-GSQ-RCO-Q2-mlx on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 128k 613.4 40.7 50.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ps1zbl25",
   "source_url": "https://omlx.ai/benchmarks/performance/ps1zbl25",
   "tg_tokens": null,
   "tps": 40.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-ps1zbl25-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 52.1 GB at this context; duplicate traces of the same session: oj8833vv; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 604.3,
   "provenance": "community",
   "quant": "GSQ-RCO-Q2",
   "quote": "Qwen3.8-Flash-Next-GSQ-RCO-Q2-mlx on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 195k 604.3 42.8 52.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ps1zbl25",
   "source_url": "https://omlx.ai/benchmarks/performance/ps1zbl25",
   "tg_tokens": null,
   "tps": 42.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-uvudo0ti-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 191.9 GB at this context; duplicate traces of the same session: qbk92re7, wy63q01k; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 852.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-Flash-Next-MLX-8bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 852.2 63.2 191.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uvudo0ti",
   "source_url": "https://omlx.ai/benchmarks/performance/uvudo0ti",
   "tg_tokens": null,
   "tps": 63.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-uvudo0ti-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 192.4 GB at this context; duplicate traces of the same session: qbk92re7, wy63q01k; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1197.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-Flash-Next-MLX-8bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 1,197 54.7 192.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uvudo0ti",
   "source_url": "https://omlx.ai/benchmarks/performance/uvudo0ti",
   "tg_tokens": null,
   "tps": 54.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-uvudo0ti-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 192.5 GB at this context; duplicate traces of the same session: qbk92re7, wy63q01k; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1208.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-Flash-Next-MLX-8bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 1,208 46.8 192.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uvudo0ti",
   "source_url": "https://omlx.ai/benchmarks/performance/uvudo0ti",
   "tg_tokens": null,
   "tps": 46.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-uvudo0ti-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 192.8 GB at this context; duplicate traces of the same session: qbk92re7, wy63q01k; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1209.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-Flash-Next-MLX-8bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k 1,209 59.8 192.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uvudo0ti",
   "source_url": "https://omlx.ai/benchmarks/performance/uvudo0ti",
   "tg_tokens": null,
   "tps": 59.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ubmuhqpy-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 190.5 GB at this context; duplicate traces of the same session: 3n3ygdi3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 911.9,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ8-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 911.9 50.3 190.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ubmuhqpy",
   "source_url": "https://omlx.ai/benchmarks/performance/ubmuhqpy",
   "tg_tokens": null,
   "tps": 50.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ubmuhqpy-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 190.8 GB at this context; duplicate traces of the same session: 3n3ygdi3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1202.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ8-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 1,202 56.2 190.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ubmuhqpy",
   "source_url": "https://omlx.ai/benchmarks/performance/ubmuhqpy",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ubmuhqpy-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 191.0 GB at this context; duplicate traces of the same session: 3n3ygdi3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1197.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ8-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 1,197 49.2 191.0 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ubmuhqpy",
   "source_url": "https://omlx.ai/benchmarks/performance/ubmuhqpy",
   "tg_tokens": null,
   "tps": 49.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-czfq9ti0-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 122.2 GB at this context; duplicate traces of the same session: 3me7kpv9, pkspuzop; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1063.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 5bit; 1k 1,063 99.1 122.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/czfq9ti0",
   "source_url": "https://omlx.ai/benchmarks/performance/czfq9ti0",
   "tg_tokens": null,
   "tps": 99.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-czfq9ti0-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 122.8 GB at this context; duplicate traces of the same session: 3me7kpv9, pkspuzop; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1348.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 5bit; 4k 1,348 91.3 122.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/czfq9ti0",
   "source_url": "https://omlx.ai/benchmarks/performance/czfq9ti0",
   "tg_tokens": null,
   "tps": 91.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-czfq9ti0-ctx8192",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 123.0 GB at this context; duplicate traces of the same session: 3me7kpv9, pkspuzop; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1343.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 5bit; 8k 1,343 86.1 123.0 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/czfq9ti0",
   "source_url": "https://omlx.ai/benchmarks/performance/czfq9ti0",
   "tg_tokens": null,
   "tps": 86.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-d81a54yx-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 122.8 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1344.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 5bit; 4k 1,344 91.5 122.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d81a54yx",
   "source_url": "https://omlx.ai/benchmarks/performance/d81a54yx",
   "tg_tokens": null,
   "tps": 91.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-d81a54yx-ctx8192",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 123.0 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1342.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 5bit; 8k 1,342 88.1 123.0 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d81a54yx",
   "source_url": "https://omlx.ai/benchmarks/performance/d81a54yx",
   "tg_tokens": null,
   "tps": 88.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-i2v21nmv-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: 9hsyd4z2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1104.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,104 122.3 100.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i2v21nmv",
   "source_url": "https://omlx.ai/benchmarks/performance/i2v21nmv",
   "tg_tokens": null,
   "tps": 122.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-i2v21nmv-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: 9hsyd4z2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1377.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,377 97.1 101.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i2v21nmv",
   "source_url": "https://omlx.ai/benchmarks/performance/i2v21nmv",
   "tg_tokens": null,
   "tps": 97.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-i2v21nmv-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: 9hsyd4z2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1375.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 1,375 88.9 101.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i2v21nmv",
   "source_url": "https://omlx.ai/benchmarks/performance/i2v21nmv",
   "tg_tokens": null,
   "tps": 88.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-i2v21nmv-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.6 GB at this context; duplicate traces of the same session: 9hsyd4z2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1369.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 1,369 73.1 101.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i2v21nmv",
   "source_url": "https://omlx.ai/benchmarks/performance/i2v21nmv",
   "tg_tokens": null,
   "tps": 73.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-h7zv7la6-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: n82bayk6, wjw2fhh6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1123.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,123 87.7 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h7zv7la6",
   "source_url": "https://omlx.ai/benchmarks/performance/h7zv7la6",
   "tg_tokens": null,
   "tps": 87.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-h7zv7la6-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: n82bayk6, wjw2fhh6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2475.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,475 65.9 71.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h7zv7la6",
   "source_url": "https://omlx.ai/benchmarks/performance/h7zv7la6",
   "tg_tokens": null,
   "tps": 65.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-h7zv7la6-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: n82bayk6, wjw2fhh6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2660.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,660 71.3 72.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h7zv7la6",
   "source_url": "https://omlx.ai/benchmarks/performance/h7zv7la6",
   "tg_tokens": null,
   "tps": 71.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-h7zv7la6-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.5 GB at this context; duplicate traces of the same session: n82bayk6, wjw2fhh6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2263.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,263 63.1 80.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h7zv7la6",
   "source_url": "https://omlx.ai/benchmarks/performance/h7zv7la6",
   "tg_tokens": null,
   "tps": 63.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-00rdocz2-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: 1ctk4ngb, iy9x080v; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1967.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,967 35.0 71.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/00rdocz2",
   "source_url": "https://omlx.ai/benchmarks/performance/00rdocz2",
   "tg_tokens": null,
   "tps": 35.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-00rdocz2-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.4 GB at this context; duplicate traces of the same session: 1ctk4ngb, iy9x080v; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2418.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,418 60.5 78.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/00rdocz2",
   "source_url": "https://omlx.ai/benchmarks/performance/00rdocz2",
   "tg_tokens": null,
   "tps": 60.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-00rdocz2-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.8 GB at this context; duplicate traces of the same session: 1ctk4ngb, iy9x080v; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2252.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,252 63.8 80.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/00rdocz2",
   "source_url": "https://omlx.ai/benchmarks/performance/00rdocz2",
   "tg_tokens": null,
   "tps": 63.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-au77fpxx-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-VL-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 244.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-VL-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 244.8 40.9 32.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/au77fpxx",
   "source_url": "https://omlx.ai/benchmarks/performance/au77fpxx",
   "tg_tokens": null,
   "tps": 40.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-au77fpxx-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-VL-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 276.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-VL-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 276.1 43.8 35.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/au77fpxx",
   "source_url": "https://omlx.ai/benchmarks/performance/au77fpxx",
   "tg_tokens": null,
   "tps": 43.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-au77fpxx-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-VL-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 275.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-VL-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 8k 275.7 37.6 35.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/au77fpxx",
   "source_url": "https://omlx.ai/benchmarks/performance/au77fpxx",
   "tg_tokens": null,
   "tps": 37.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-au77fpxx-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-VL-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 250.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-VL-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k 250.5 32.6 31.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/au77fpxx",
   "source_url": "https://omlx.ai/benchmarks/performance/au77fpxx",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-au77fpxx-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-VL-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 232.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-VL-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 232.4 30.3 33.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/au77fpxx",
   "source_url": "https://omlx.ai/benchmarks/performance/au77fpxx",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-rgfpe187-ctx1024",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "YCF-AI MTP FP16 build; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: m9er6sfx, s2f23ysa, to3e0k7u; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 203.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "YCF-AI--Qwen3.8-27B-Huihui-abliterated-oQ4e-MTP-FP16-MLX on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 203.6 29.6 18.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rgfpe187",
   "source_url": "https://omlx.ai/benchmarks/performance/rgfpe187",
   "tg_tokens": null,
   "tps": 29.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-rgfpe187-ctx4096",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "YCF-AI MTP FP16 build; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: m9er6sfx, s2f23ysa, to3e0k7u; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "YCF-AI--Qwen3.8-27B-Huihui-abliterated-oQ4e-MTP-FP16-MLX on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 204.0 32.0 19.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rgfpe187",
   "source_url": "https://omlx.ai/benchmarks/performance/rgfpe187",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-rgfpe187-ctx8192",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "YCF-AI MTP FP16 build; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: m9er6sfx, s2f23ysa, to3e0k7u; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 202.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "YCF-AI--Qwen3.8-27B-Huihui-abliterated-oQ4e-MTP-FP16-MLX on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 202.6 29.2 20.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rgfpe187",
   "source_url": "https://omlx.ai/benchmarks/performance/rgfpe187",
   "tg_tokens": null,
   "tps": 29.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-rgfpe187-ctx16384",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "YCF-AI MTP FP16 build; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: m9er6sfx, s2f23ysa, to3e0k7u; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 194.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "YCF-AI--Qwen3.8-27B-Huihui-abliterated-oQ4e-MTP-FP16-MLX on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 194.5 29.7 20.4 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rgfpe187",
   "source_url": "https://omlx.ai/benchmarks/performance/rgfpe187",
   "tg_tokens": null,
   "tps": 29.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1z399iss-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Youssofal MTPLX Optimized Speed; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 422.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Youssofal--Qwen3.8-27B-MTPLX-Optimized-Speed on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 422.9 68.9 24.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1z399iss",
   "source_url": "https://omlx.ai/benchmarks/performance/1z399iss",
   "tg_tokens": null,
   "tps": 68.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1z399iss-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "Youssofal MTPLX Optimized Speed; oMLX community benchmark trace; peak memory 27.6 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 442.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Youssofal--Qwen3.8-27B-MTPLX-Optimized-Speed on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 442.2 58.8 27.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1z399iss",
   "source_url": "https://omlx.ai/benchmarks/performance/1z399iss",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1z399iss-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "Youssofal MTPLX Optimized Speed; oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 436.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Youssofal--Qwen3.8-27B-MTPLX-Optimized-Speed on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 436.4 80.2 28.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1z399iss",
   "source_url": "https://omlx.ai/benchmarks/performance/1z399iss",
   "tg_tokens": null,
   "tps": 80.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1z399iss-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "Youssofal MTPLX Optimized Speed; oMLX community benchmark trace; peak memory 28.8 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 420.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Youssofal--Qwen3.8-27B-MTPLX-Optimized-Speed on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 420.9 74.9 28.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1z399iss",
   "source_url": "https://omlx.ai/benchmarks/performance/1z399iss",
   "tg_tokens": null,
   "tps": 74.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-f43ujcpz-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "fcmeyer MTP build; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 5f89m5z7, e2ahkv6x, pzkq01al; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "fcmeyer--Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 204.0 30.3 18.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f43ujcpz",
   "source_url": "https://omlx.ai/benchmarks/performance/f43ujcpz",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-f43ujcpz-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "fcmeyer MTP build; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 5f89m5z7, e2ahkv6x, pzkq01al; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "fcmeyer--Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 204.9 32.3 19.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f43ujcpz",
   "source_url": "https://omlx.ai/benchmarks/performance/f43ujcpz",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-f43ujcpz-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "fcmeyer MTP build; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: 5f89m5z7, e2ahkv6x, pzkq01al; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 202.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "fcmeyer--Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 202.6 33.7 20.5 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f43ujcpz",
   "source_url": "https://omlx.ai/benchmarks/performance/f43ujcpz",
   "tg_tokens": null,
   "tps": 33.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-f43ujcpz-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "fcmeyer MTP build; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: 5f89m5z7, e2ahkv6x, pzkq01al; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 194.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "fcmeyer--Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 194.3 28.2 20.4 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f43ujcpz",
   "source_url": "https://omlx.ai/benchmarks/performance/f43ujcpz",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-38zxeyt5-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: td0y1afs; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1297.0,
   "provenance": "community",
   "quant": "qat-8bit",
   "quote": "gemma-4-26B-A4B-it-qat-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 1,297 69.7 26.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/38zxeyt5",
   "source_url": "https://omlx.ai/benchmarks/performance/38zxeyt5",
   "tg_tokens": null,
   "tps": 69.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-38zxeyt5-ctx32768",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: td0y1afs; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1051.0,
   "provenance": "community",
   "quant": "qat-8bit",
   "quote": "gemma-4-26B-A4B-it-qat-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 1,051 60.8 27.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/38zxeyt5",
   "source_url": "https://omlx.ai/benchmarks/performance/38zxeyt5",
   "tg_tokens": null,
   "tps": 60.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-hg14hnmd-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "leonsarmiento pi MTP; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: i62jp0kk, l7osqd9p; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "leonsarmiento--Qwen3.8-27B-pi-4bit-mtp-mlx on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 204.8 31.1 19.3 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hg14hnmd",
   "source_url": "https://omlx.ai/benchmarks/performance/hg14hnmd",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-hg14hnmd-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "leonsarmiento pi MTP; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: i62jp0kk, l7osqd9p; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 206.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "leonsarmiento--Qwen3.8-27B-pi-4bit-mtp-mlx on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 206.2 36.6 20.2 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hg14hnmd",
   "source_url": "https://omlx.ai/benchmarks/performance/hg14hnmd",
   "tg_tokens": null,
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-hg14hnmd-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "leonsarmiento pi MTP; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: i62jp0kk, l7osqd9p; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 203.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "leonsarmiento--Qwen3.8-27B-pi-4bit-mtp-mlx on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 203.5 34.1 20.9 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hg14hnmd",
   "source_url": "https://omlx.ai/benchmarks/performance/hg14hnmd",
   "tg_tokens": null,
   "tps": 34.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-hg14hnmd-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "leonsarmiento pi MTP; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: i62jp0kk, l7osqd9p; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 195.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "leonsarmiento--Qwen3.8-27B-pi-4bit-mtp-mlx on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 195.0 30.6 20.7 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hg14hnmd",
   "source_url": "https://omlx.ai/benchmarks/performance/hg14hnmd",
   "tg_tokens": null,
   "tps": 30.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-92zcatgs-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "mlx-community repack; oMLX community benchmark trace; peak memory 77.9 GB at this context; duplicate traces of the same session: 9guweh5o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1687.0,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "mlx-community--Qwen3-Coder-Next-mxfp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 1,687 83.2 77.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/92zcatgs",
   "source_url": "https://omlx.ai/benchmarks/performance/92zcatgs",
   "tg_tokens": null,
   "tps": 83.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-92zcatgs-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "mlx-community repack; oMLX community benchmark trace; peak memory 78.7 GB at this context; duplicate traces of the same session: 9guweh5o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3189.0,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "mlx-community--Qwen3-Coder-Next-mxfp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 3,189 81.9 78.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/92zcatgs",
   "source_url": "https://omlx.ai/benchmarks/performance/92zcatgs",
   "tg_tokens": null,
   "tps": 81.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-92zcatgs-ctx8192",
   "model": "Qwen3-Coder-Next",
   "notes": "mlx-community repack; oMLX community benchmark trace; peak memory 78.8 GB at this context; duplicate traces of the same session: 9guweh5o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3260.0,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "mlx-community--Qwen3-Coder-Next-mxfp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 3,260 80.0 78.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/92zcatgs",
   "source_url": "https://omlx.ai/benchmarks/performance/92zcatgs",
   "tg_tokens": null,
   "tps": 80.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-92zcatgs-ctx16384",
   "model": "Qwen3-Coder-Next",
   "notes": "mlx-community repack; oMLX community benchmark trace; peak memory 79.0 GB at this context; duplicate traces of the same session: 9guweh5o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3037.0,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "mlx-community--Qwen3-Coder-Next-mxfp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 3,037 76.5 79.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/92zcatgs",
   "source_url": "https://omlx.ai/benchmarks/performance/92zcatgs",
   "tg_tokens": null,
   "tps": 76.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-ftii7j1l-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: h3ieee67; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 218.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 218.2 16.7 19.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ftii7j1l",
   "source_url": "https://omlx.ai/benchmarks/performance/ftii7j1l",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-ftii7j1l-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: h3ieee67; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 306.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 306.9 18.2 21.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ftii7j1l",
   "source_url": "https://omlx.ai/benchmarks/performance/ftii7j1l",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-ftii7j1l-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: h3ieee67; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 298.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 8k 298.6 25.6 22.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ftii7j1l",
   "source_url": "https://omlx.ai/benchmarks/performance/ftii7j1l",
   "tg_tokens": null,
   "tps": 25.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-ftii7j1l-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: h3ieee67; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 275.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 275.4 20.6 23.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ftii7j1l",
   "source_url": "https://omlx.ai/benchmarks/performance/ftii7j1l",
   "tg_tokens": null,
   "tps": 20.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-ftii7j1l-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: h3ieee67; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 248.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 32k 248.0 6.9 26.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ftii7j1l",
   "source_url": "https://omlx.ai/benchmarks/performance/ftii7j1l",
   "tg_tokens": null,
   "tps": 6.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-ftii7j1l-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 30.5 GB at this context; duplicate traces of the same session: h3ieee67; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 206.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 64k 206.8 4.3 30.5 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ftii7j1l",
   "source_url": "https://omlx.ai/benchmarks/performance/ftii7j1l",
   "tg_tokens": null,
   "tps": 4.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ac3vecvr-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.8 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 499.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 499.8 35.6 175.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ac3vecvr",
   "source_url": "https://omlx.ai/benchmarks/performance/ac3vecvr",
   "tg_tokens": null,
   "tps": 35.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ac3vecvr-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 544.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 544.3 70.4 177.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ac3vecvr",
   "source_url": "https://omlx.ai/benchmarks/performance/ac3vecvr",
   "tg_tokens": null,
   "tps": 70.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ac3vecvr-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.0 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 526.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 526.1 37.5 178.0 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ac3vecvr",
   "source_url": "https://omlx.ai/benchmarks/performance/ac3vecvr",
   "tg_tokens": null,
   "tps": 37.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ac3vecvr-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 514.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 514.9 26.8 178.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ac3vecvr",
   "source_url": "https://omlx.ai/benchmarks/performance/ac3vecvr",
   "tg_tokens": null,
   "tps": 26.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-y4issq84-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.8 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 500.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 500.5 38.5 175.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y4issq84",
   "source_url": "https://omlx.ai/benchmarks/performance/y4issq84",
   "tg_tokens": null,
   "tps": 38.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-y4issq84-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 545.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 545.0 25.4 177.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y4issq84",
   "source_url": "https://omlx.ai/benchmarks/performance/y4issq84",
   "tg_tokens": null,
   "tps": 25.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-y4issq84-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 514.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 514.4 64.4 178.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y4issq84",
   "source_url": "https://omlx.ai/benchmarks/performance/y4issq84",
   "tg_tokens": null,
   "tps": 64.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ypk0of8l-ctx16384",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 12.9 GB at this context; duplicate traces of the same session: fq0p5mdm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2054.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 2,054 118.8 12.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ypk0of8l",
   "source_url": "https://omlx.ai/benchmarks/performance/ypk0of8l",
   "tg_tokens": null,
   "tps": 118.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ypk0of8l-ctx32768",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 13.1 GB at this context; duplicate traces of the same session: fq0p5mdm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1679.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 1,679 109.0 13.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ypk0of8l",
   "source_url": "https://omlx.ai/benchmarks/performance/ypk0of8l",
   "tg_tokens": null,
   "tps": 109.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-63nxlwrl-ctx65536",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 13.6 GB; TTFT 48155 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1361.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k tokens 1,361 PP tok/s 94.7 TG tok/s 48155 TTFT (ms) 13.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/63nxlwrl",
   "source_url": "https://omlx.ai/benchmarks/performance/63nxlwrl",
   "tg_tokens": null,
   "tps": 94.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-1t0oj560-ctx16384",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 0ycul5zf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2050.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-bf16 on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 bf16; 16k 2,050 80.9 23.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1t0oj560",
   "source_url": "https://omlx.ai/benchmarks/performance/1t0oj560",
   "tg_tokens": null,
   "tps": 80.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-1t0oj560-ctx32768",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 0ycul5zf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1644.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-bf16 on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 bf16; 32k 1,644 76.3 23.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1t0oj560",
   "source_url": "https://omlx.ai/benchmarks/performance/1t0oj560",
   "tg_tokens": null,
   "tps": 76.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-0bwo8trg-ctx16384",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: ribhuqzt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2048.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-bf16 on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 bf16; 16k 2,048 81.1 23.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0bwo8trg",
   "source_url": "https://omlx.ai/benchmarks/performance/0bwo8trg",
   "tg_tokens": null,
   "tps": 81.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-0bwo8trg-ctx32768",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: ribhuqzt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1622.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-bf16 on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 bf16; 32k 1,622 76.7 23.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0bwo8trg",
   "source_url": "https://omlx.ai/benchmarks/performance/0bwo8trg",
   "tg_tokens": null,
   "tps": 76.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-nkjstbl3-ctx8192",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace, single run at 8k context; peak memory 23.3 GB; TTFT 3737 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2192.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-bf16 on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 bf16; 8k tokens 2,192 PP tok/s 82.6 TG tok/s 3737 TTFT (ms) 23.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nkjstbl3",
   "source_url": "https://omlx.ai/benchmarks/performance/nkjstbl3",
   "tg_tokens": null,
   "tps": 82.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-4nizg8dp-ctx4096",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 23.3 GB; TTFT 1839 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2228.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-bf16 on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 bf16; 4k tokens 2,228 PP tok/s 83.4 TG tok/s 1839 TTFT (ms) 23.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4nizg8dp",
   "source_url": "https://omlx.ai/benchmarks/performance/4nizg8dp",
   "tg_tokens": null,
   "tps": 83.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-b8zjrei1-ctx16384",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1459.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 1,459 94.2 34.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b8zjrei1",
   "source_url": "https://omlx.ai/benchmarks/performance/b8zjrei1",
   "tg_tokens": null,
   "tps": 94.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-b8zjrei1-ctx32768",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1246.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 1,246 88.9 34.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b8zjrei1",
   "source_url": "https://omlx.ai/benchmarks/performance/b8zjrei1",
   "tg_tokens": null,
   "tps": 88.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-23y21y40-ctx16384",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1457.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 1,457 93.9 34.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/23y21y40",
   "source_url": "https://omlx.ai/benchmarks/performance/23y21y40",
   "tg_tokens": null,
   "tps": 93.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-23y21y40-ctx32768",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1293.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 1,293 90.4 34.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/23y21y40",
   "source_url": "https://omlx.ai/benchmarks/performance/23y21y40",
   "tg_tokens": null,
   "tps": 90.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-v116sjvq-ctx1024",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1273.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,273 113.6 22.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v116sjvq",
   "source_url": "https://omlx.ai/benchmarks/performance/v116sjvq",
   "tg_tokens": null,
   "tps": 113.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-v116sjvq-ctx4096",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1556.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,556 109.6 23.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v116sjvq",
   "source_url": "https://omlx.ai/benchmarks/performance/v116sjvq",
   "tg_tokens": null,
   "tps": 109.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-v116sjvq-ctx8192",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1407.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,407 107.5 23.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v116sjvq",
   "source_url": "https://omlx.ai/benchmarks/performance/v116sjvq",
   "tg_tokens": null,
   "tps": 107.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-v116sjvq-ctx16384",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1333.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,333 98.2 23.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v116sjvq",
   "source_url": "https://omlx.ai/benchmarks/performance/v116sjvq",
   "tg_tokens": null,
   "tps": 98.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-v116sjvq-ctx32768",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1160.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 1,160 89.9 24.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v116sjvq",
   "source_url": "https://omlx.ai/benchmarks/performance/v116sjvq",
   "tg_tokens": null,
   "tps": 89.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-v116sjvq-ctx65536",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 977.1,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 977.1 77.4 24.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v116sjvq",
   "source_url": "https://omlx.ai/benchmarks/performance/v116sjvq",
   "tg_tokens": null,
   "tps": 77.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2",
   "id": "omlx-3m4ve5qg-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: none; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 110.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M2 (8c); M2 (8c) \u00b7 16 GB \u00b7 4bit; 1k 110.4 18.8 5.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3m4ve5qg",
   "source_url": "https://omlx.ai/benchmarks/performance/3m4ve5qg",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2",
   "id": "omlx-3m4ve5qg-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: none; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M2 (8c); M2 (8c) \u00b7 16 GB \u00b7 4bit; 4k 105.8 18.6 6.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3m4ve5qg",
   "source_url": "https://omlx.ai/benchmarks/performance/3m4ve5qg",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2",
   "id": "omlx-3m4ve5qg-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: none; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 101.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M2 (8c); M2 (8c) \u00b7 16 GB \u00b7 4bit; 8k 101.5 17.8 6.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3m4ve5qg",
   "source_url": "https://omlx.ai/benchmarks/performance/3m4ve5qg",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-ojsbxllo-ctx1024",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 518.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 518.7 34.1 12.7 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ojsbxllo",
   "source_url": "https://omlx.ai/benchmarks/performance/ojsbxllo",
   "tg_tokens": null,
   "tps": 34.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-ojsbxllo-ctx4096",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.3 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 532.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 532.5 33.3 13.3 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ojsbxllo",
   "source_url": "https://omlx.ai/benchmarks/performance/ojsbxllo",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-cks3y95n-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 429.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 429.5 10.1 28.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cks3y95n",
   "source_url": "https://omlx.ai/benchmarks/performance/cks3y95n",
   "tg_tokens": null,
   "tps": 10.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-cks3y95n-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 470.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 470.6 10.0 29.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cks3y95n",
   "source_url": "https://omlx.ai/benchmarks/performance/cks3y95n",
   "tg_tokens": null,
   "tps": 10.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-nqyd2uyx-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 33.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 216.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 216.7 15.3 33.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nqyd2uyx",
   "source_url": "https://omlx.ai/benchmarks/performance/nqyd2uyx",
   "tg_tokens": null,
   "tps": 15.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-nqyd2uyx-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 209.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 209.2 14.9 34.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nqyd2uyx",
   "source_url": "https://omlx.ai/benchmarks/performance/nqyd2uyx",
   "tg_tokens": null,
   "tps": 14.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-oj8833vv-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 44.4 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 629.9,
   "provenance": "community",
   "quant": "GSQ-RCO-Q2",
   "quote": "Qwen3.8-Flash-Next-GSQ-RCO-Q2-mlx on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 32k 629.9 47.7 44.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oj8833vv",
   "source_url": "https://omlx.ai/benchmarks/performance/oj8833vv",
   "tg_tokens": null,
   "tps": 47.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-oj8833vv-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 46.3 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 626.9,
   "provenance": "community",
   "quant": "GSQ-RCO-Q2",
   "quote": "Qwen3.8-Flash-Next-GSQ-RCO-Q2-mlx on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 64k 626.9 41.1 46.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oj8833vv",
   "source_url": "https://omlx.ai/benchmarks/performance/oj8833vv",
   "tg_tokens": null,
   "tps": 41.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-oj8833vv-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 50.7 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 613.4,
   "provenance": "community",
   "quant": "GSQ-RCO-Q2",
   "quote": "Qwen3.8-Flash-Next-GSQ-RCO-Q2-mlx on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 128k 613.4 40.7 50.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oj8833vv",
   "source_url": "https://omlx.ai/benchmarks/performance/oj8833vv",
   "tg_tokens": null,
   "tps": 40.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-oj8833vv-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 52.1 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 604.3,
   "provenance": "community",
   "quant": "GSQ-RCO-Q2",
   "quote": "Qwen3.8-Flash-Next-GSQ-RCO-Q2-mlx on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 195k 604.3 42.8 52.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oj8833vv",
   "source_url": "https://omlx.ai/benchmarks/performance/oj8833vv",
   "tg_tokens": null,
   "tps": 42.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-wy63q01k-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 191.9 GB at this context; duplicate traces of the same session: qbk92re7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 852.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-Flash-Next-MLX-8bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 852.2 63.2 191.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wy63q01k",
   "source_url": "https://omlx.ai/benchmarks/performance/wy63q01k",
   "tg_tokens": null,
   "tps": 63.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-wy63q01k-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 192.4 GB at this context; duplicate traces of the same session: qbk92re7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1197.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-Flash-Next-MLX-8bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 1,197 54.7 192.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wy63q01k",
   "source_url": "https://omlx.ai/benchmarks/performance/wy63q01k",
   "tg_tokens": null,
   "tps": 54.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-wy63q01k-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 192.5 GB at this context; duplicate traces of the same session: qbk92re7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1208.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-Flash-Next-MLX-8bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 1,208 46.8 192.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wy63q01k",
   "source_url": "https://omlx.ai/benchmarks/performance/wy63q01k",
   "tg_tokens": null,
   "tps": 46.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-wy63q01k-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 192.8 GB at this context; duplicate traces of the same session: qbk92re7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1209.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-Flash-Next-MLX-8bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k 1,209 59.8 192.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wy63q01k",
   "source_url": "https://omlx.ai/benchmarks/performance/wy63q01k",
   "tg_tokens": null,
   "tps": 59.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-3n3ygdi3-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 190.5 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 911.9,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ8-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 911.9 50.3 190.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3n3ygdi3",
   "source_url": "https://omlx.ai/benchmarks/performance/3n3ygdi3",
   "tg_tokens": null,
   "tps": 50.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-3n3ygdi3-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 190.8 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1202.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ8-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 1,202 56.2 190.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3n3ygdi3",
   "source_url": "https://omlx.ai/benchmarks/performance/3n3ygdi3",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-3n3ygdi3-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 191.0 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1197.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ8-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 1,197 49.2 191.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3n3ygdi3",
   "source_url": "https://omlx.ai/benchmarks/performance/3n3ygdi3",
   "tg_tokens": null,
   "tps": 49.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-pkspuzop-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 122.2 GB at this context; duplicate traces of the same session: 3me7kpv9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1063.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 5bit; 1k 1,063 99.1 122.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pkspuzop",
   "source_url": "https://omlx.ai/benchmarks/performance/pkspuzop",
   "tg_tokens": null,
   "tps": 99.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-pkspuzop-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 122.8 GB at this context; duplicate traces of the same session: 3me7kpv9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1348.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 5bit; 4k 1,348 91.3 122.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pkspuzop",
   "source_url": "https://omlx.ai/benchmarks/performance/pkspuzop",
   "tg_tokens": null,
   "tps": 91.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-pkspuzop-ctx8192",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 123.0 GB at this context; duplicate traces of the same session: 3me7kpv9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1343.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 5bit; 8k 1,343 86.1 123.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pkspuzop",
   "source_url": "https://omlx.ai/benchmarks/performance/pkspuzop",
   "tg_tokens": null,
   "tps": 86.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-mw0wy1qz-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 122.8 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1344.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 5bit; 4k 1,344 91.5 122.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mw0wy1qz",
   "source_url": "https://omlx.ai/benchmarks/performance/mw0wy1qz",
   "tg_tokens": null,
   "tps": 91.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-mw0wy1qz-ctx8192",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 123.0 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1342.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 5bit; 8k 1,342 88.1 123.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mw0wy1qz",
   "source_url": "https://omlx.ai/benchmarks/performance/mw0wy1qz",
   "tg_tokens": null,
   "tps": 88.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9hsyd4z2-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1104.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,104 122.3 100.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9hsyd4z2",
   "source_url": "https://omlx.ai/benchmarks/performance/9hsyd4z2",
   "tg_tokens": null,
   "tps": 122.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9hsyd4z2-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1377.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,377 97.1 101.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9hsyd4z2",
   "source_url": "https://omlx.ai/benchmarks/performance/9hsyd4z2",
   "tg_tokens": null,
   "tps": 97.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9hsyd4z2-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1375.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 1,375 88.9 101.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9hsyd4z2",
   "source_url": "https://omlx.ai/benchmarks/performance/9hsyd4z2",
   "tg_tokens": null,
   "tps": 88.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9hsyd4z2-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.6 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1369.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 1,369 73.1 101.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9hsyd4z2",
   "source_url": "https://omlx.ai/benchmarks/performance/9hsyd4z2",
   "tg_tokens": null,
   "tps": 73.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n82bayk6-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: wjw2fhh6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1123.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,123 87.7 70.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n82bayk6",
   "source_url": "https://omlx.ai/benchmarks/performance/n82bayk6",
   "tg_tokens": null,
   "tps": 87.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n82bayk6-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: wjw2fhh6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2475.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,475 65.9 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n82bayk6",
   "source_url": "https://omlx.ai/benchmarks/performance/n82bayk6",
   "tg_tokens": null,
   "tps": 65.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n82bayk6-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: wjw2fhh6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2660.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,660 71.3 72.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n82bayk6",
   "source_url": "https://omlx.ai/benchmarks/performance/n82bayk6",
   "tg_tokens": null,
   "tps": 71.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n82bayk6-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.5 GB at this context; duplicate traces of the same session: wjw2fhh6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2263.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,263 63.1 80.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n82bayk6",
   "source_url": "https://omlx.ai/benchmarks/performance/n82bayk6",
   "tg_tokens": null,
   "tps": 63.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-iy9x080v-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: 1ctk4ngb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1967.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,967 35.0 71.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iy9x080v",
   "source_url": "https://omlx.ai/benchmarks/performance/iy9x080v",
   "tg_tokens": null,
   "tps": 35.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-iy9x080v-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.4 GB at this context; duplicate traces of the same session: 1ctk4ngb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2418.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,418 60.5 78.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iy9x080v",
   "source_url": "https://omlx.ai/benchmarks/performance/iy9x080v",
   "tg_tokens": null,
   "tps": 60.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-iy9x080v-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.8 GB at this context; duplicate traces of the same session: 1ctk4ngb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2252.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,252 63.8 80.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iy9x080v",
   "source_url": "https://omlx.ai/benchmarks/performance/iy9x080v",
   "tg_tokens": null,
   "tps": 63.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fz7l4l26-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: rfl4hdug, rms21oyz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2401.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,401 70.7 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fz7l4l26",
   "source_url": "https://omlx.ai/benchmarks/performance/fz7l4l26",
   "tg_tokens": null,
   "tps": 70.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fz7l4l26-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: rfl4hdug, rms21oyz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2619.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,619 72.9 73.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fz7l4l26",
   "source_url": "https://omlx.ai/benchmarks/performance/fz7l4l26",
   "tg_tokens": null,
   "tps": 72.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fz7l4l26-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.6 GB at this context; duplicate traces of the same session: rfl4hdug, rms21oyz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2524.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,524 58.1 74.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fz7l4l26",
   "source_url": "https://omlx.ai/benchmarks/performance/fz7l4l26",
   "tg_tokens": null,
   "tps": 58.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-x3w8zcyb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Steven10429 UD build; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 36xx1viz, wcmudw50; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 210.1,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "Steven10429--Qwen3.8-27B-UD-IQ4_XS-MLX on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 210.1 29.1 17.6 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x3w8zcyb",
   "source_url": "https://omlx.ai/benchmarks/performance/x3w8zcyb",
   "tg_tokens": null,
   "tps": 29.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-x3w8zcyb-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "Steven10429 UD build; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 36xx1viz, wcmudw50; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 210.8,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "Steven10429--Qwen3.8-27B-UD-IQ4_XS-MLX on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 210.8 23.3 18.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x3w8zcyb",
   "source_url": "https://omlx.ai/benchmarks/performance/x3w8zcyb",
   "tg_tokens": null,
   "tps": 23.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-x3w8zcyb-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "Steven10429 UD build; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 36xx1viz, wcmudw50; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 207.9,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "Steven10429--Qwen3.8-27B-UD-IQ4_XS-MLX on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 207.9 21.1 19.4 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x3w8zcyb",
   "source_url": "https://omlx.ai/benchmarks/performance/x3w8zcyb",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-s1eqxc0t-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Steven10429 UD build; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 33dwtg7z, 5jwi932s; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 210.6,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "Steven10429--Qwen3.8-27B-UD-IQ4_XS-MLX on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 210.6 29.5 17.6 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s1eqxc0t",
   "source_url": "https://omlx.ai/benchmarks/performance/s1eqxc0t",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-s1eqxc0t-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "Steven10429 UD build; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 33dwtg7z, 5jwi932s; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 209.9,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "Steven10429--Qwen3.8-27B-UD-IQ4_XS-MLX on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 209.9 25.4 18.8 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s1eqxc0t",
   "source_url": "https://omlx.ai/benchmarks/performance/s1eqxc0t",
   "tg_tokens": null,
   "tps": 25.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-s1eqxc0t-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "Steven10429 UD build; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 33dwtg7z, 5jwi932s; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 207.9,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "Steven10429--Qwen3.8-27B-UD-IQ4_XS-MLX on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 207.9 24.3 19.4 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s1eqxc0t",
   "source_url": "https://omlx.ai/benchmarks/performance/s1eqxc0t",
   "tg_tokens": null,
   "tps": 24.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-m9er6sfx-ctx1024",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "YCF-AI MTP FP16 build; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: s2f23ysa, to3e0k7u; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 203.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "YCF-AI--Qwen3.8-27B-Huihui-abliterated-oQ4e-MTP-FP16-MLX on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 203.6 29.6 18.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m9er6sfx",
   "source_url": "https://omlx.ai/benchmarks/performance/m9er6sfx",
   "tg_tokens": null,
   "tps": 29.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-m9er6sfx-ctx4096",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "YCF-AI MTP FP16 build; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: s2f23ysa, to3e0k7u; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "YCF-AI--Qwen3.8-27B-Huihui-abliterated-oQ4e-MTP-FP16-MLX on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 204.0 32.0 19.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m9er6sfx",
   "source_url": "https://omlx.ai/benchmarks/performance/m9er6sfx",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-m9er6sfx-ctx8192",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "YCF-AI MTP FP16 build; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: s2f23ysa, to3e0k7u; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 202.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "YCF-AI--Qwen3.8-27B-Huihui-abliterated-oQ4e-MTP-FP16-MLX on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 202.6 29.2 20.5 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m9er6sfx",
   "source_url": "https://omlx.ai/benchmarks/performance/m9er6sfx",
   "tg_tokens": null,
   "tps": 29.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-m9er6sfx-ctx16384",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "YCF-AI MTP FP16 build; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: s2f23ysa, to3e0k7u; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 194.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "YCF-AI--Qwen3.8-27B-Huihui-abliterated-oQ4e-MTP-FP16-MLX on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 194.5 29.7 20.4 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m9er6sfx",
   "source_url": "https://omlx.ai/benchmarks/performance/m9er6sfx",
   "tg_tokens": null,
   "tps": 29.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-7mzu5vi3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "chriswessels ThinkingCap; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 203.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "chriswessels--ThinkingCap-Qwen3.8-27B-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 203.1 32.9 18.5 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7mzu5vi3",
   "source_url": "https://omlx.ai/benchmarks/performance/7mzu5vi3",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-7mzu5vi3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "chriswessels ThinkingCap; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "chriswessels--ThinkingCap-Qwen3.8-27B-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 205.4 34.9 19.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7mzu5vi3",
   "source_url": "https://omlx.ai/benchmarks/performance/7mzu5vi3",
   "tg_tokens": null,
   "tps": 34.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-5f89m5z7-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "fcmeyer MTP build; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: e2ahkv6x, pzkq01al; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "fcmeyer--Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 204.0 30.3 18.8 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5f89m5z7",
   "source_url": "https://omlx.ai/benchmarks/performance/5f89m5z7",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-5f89m5z7-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "fcmeyer MTP build; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: e2ahkv6x, pzkq01al; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "fcmeyer--Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 204.9 32.3 19.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5f89m5z7",
   "source_url": "https://omlx.ai/benchmarks/performance/5f89m5z7",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-5f89m5z7-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "fcmeyer MTP build; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: e2ahkv6x, pzkq01al; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 202.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "fcmeyer--Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 202.6 33.7 20.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5f89m5z7",
   "source_url": "https://omlx.ai/benchmarks/performance/5f89m5z7",
   "tg_tokens": null,
   "tps": 33.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-5f89m5z7-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "fcmeyer MTP build; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: e2ahkv6x, pzkq01al; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 194.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "fcmeyer--Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 194.3 28.2 20.4 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5f89m5z7",
   "source_url": "https://omlx.ai/benchmarks/performance/5f89m5z7",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-njdw13qj-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "fcmeyer MTP build; oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: 31llko7u, 47d7yd3r; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 203.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "fcmeyer--Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 203.5 34.5 18.6 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/njdw13qj",
   "source_url": "https://omlx.ai/benchmarks/performance/njdw13qj",
   "tg_tokens": null,
   "tps": 34.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-njdw13qj-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "fcmeyer MTP build; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 31llko7u, 47d7yd3r; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "fcmeyer--Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 204.2 34.6 19.8 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/njdw13qj",
   "source_url": "https://omlx.ai/benchmarks/performance/njdw13qj",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-njdw13qj-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "fcmeyer MTP build; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: 31llko7u, 47d7yd3r; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 202.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "fcmeyer--Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 202.7 31.8 20.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/njdw13qj",
   "source_url": "https://omlx.ai/benchmarks/performance/njdw13qj",
   "tg_tokens": null,
   "tps": 31.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-td0y1afs-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1297.0,
   "provenance": "community",
   "quant": "qat-8bit",
   "quote": "gemma-4-26B-A4B-it-qat-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 1,297 69.7 26.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/td0y1afs",
   "source_url": "https://omlx.ai/benchmarks/performance/td0y1afs",
   "tg_tokens": null,
   "tps": 69.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-td0y1afs-ctx32768",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1051.0,
   "provenance": "community",
   "quant": "qat-8bit",
   "quote": "gemma-4-26B-A4B-it-qat-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 1,051 60.8 27.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/td0y1afs",
   "source_url": "https://omlx.ai/benchmarks/performance/td0y1afs",
   "tg_tokens": null,
   "tps": 60.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-l7osqd9p-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "leonsarmiento pi MTP; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: i62jp0kk; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "leonsarmiento--Qwen3.8-27B-pi-4bit-mtp-mlx on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 204.8 31.1 19.3 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l7osqd9p",
   "source_url": "https://omlx.ai/benchmarks/performance/l7osqd9p",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-l7osqd9p-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "leonsarmiento pi MTP; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: i62jp0kk; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 206.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "leonsarmiento--Qwen3.8-27B-pi-4bit-mtp-mlx on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 206.2 36.6 20.2 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l7osqd9p",
   "source_url": "https://omlx.ai/benchmarks/performance/l7osqd9p",
   "tg_tokens": null,
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-l7osqd9p-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "leonsarmiento pi MTP; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: i62jp0kk; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 203.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "leonsarmiento--Qwen3.8-27B-pi-4bit-mtp-mlx on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 203.5 34.1 20.9 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l7osqd9p",
   "source_url": "https://omlx.ai/benchmarks/performance/l7osqd9p",
   "tg_tokens": null,
   "tps": 34.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-l7osqd9p-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "leonsarmiento pi MTP; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: i62jp0kk; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 195.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "leonsarmiento--Qwen3.8-27B-pi-4bit-mtp-mlx on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 195.0 30.6 20.7 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l7osqd9p",
   "source_url": "https://omlx.ai/benchmarks/performance/l7osqd9p",
   "tg_tokens": null,
   "tps": 30.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-c0pq0qr5-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "leonsarmiento pi MTP; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: trx6cgey, uk3tp1ep; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "leonsarmiento--Qwen3.8-27B-pi-4bit-mtp-mlx on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 204.4 34.6 18.9 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c0pq0qr5",
   "source_url": "https://omlx.ai/benchmarks/performance/c0pq0qr5",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-c0pq0qr5-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "leonsarmiento pi MTP; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: trx6cgey, uk3tp1ep; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 206.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "leonsarmiento--Qwen3.8-27B-pi-4bit-mtp-mlx on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 206.2 38.0 20.2 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c0pq0qr5",
   "source_url": "https://omlx.ai/benchmarks/performance/c0pq0qr5",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-c0pq0qr5-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "leonsarmiento pi MTP; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: trx6cgey, uk3tp1ep; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 203.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "leonsarmiento--Qwen3.8-27B-pi-4bit-mtp-mlx on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 203.5 28.1 20.9 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c0pq0qr5",
   "source_url": "https://omlx.ai/benchmarks/performance/c0pq0qr5",
   "tg_tokens": null,
   "tps": 28.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9guweh5o-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "mlx-community repack; oMLX community benchmark trace; peak memory 77.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1687.0,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "mlx-community--Qwen3-Coder-Next-mxfp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 1,687 83.2 77.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9guweh5o",
   "source_url": "https://omlx.ai/benchmarks/performance/9guweh5o",
   "tg_tokens": null,
   "tps": 83.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9guweh5o-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "mlx-community repack; oMLX community benchmark trace; peak memory 78.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3189.0,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "mlx-community--Qwen3-Coder-Next-mxfp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 3,189 81.9 78.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9guweh5o",
   "source_url": "https://omlx.ai/benchmarks/performance/9guweh5o",
   "tg_tokens": null,
   "tps": 81.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9guweh5o-ctx8192",
   "model": "Qwen3-Coder-Next",
   "notes": "mlx-community repack; oMLX community benchmark trace; peak memory 78.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3260.0,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "mlx-community--Qwen3-Coder-Next-mxfp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 3,260 80.0 78.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9guweh5o",
   "source_url": "https://omlx.ai/benchmarks/performance/9guweh5o",
   "tg_tokens": null,
   "tps": 80.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9guweh5o-ctx16384",
   "model": "Qwen3-Coder-Next",
   "notes": "mlx-community repack; oMLX community benchmark trace; peak memory 79.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3037.0,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "mlx-community--Qwen3-Coder-Next-mxfp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 3,037 76.5 79.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9guweh5o",
   "source_url": "https://omlx.ai/benchmarks/performance/9guweh5o",
   "tg_tokens": null,
   "tps": 76.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-h3ieee67-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 7iz2k1h4, vywrywjp, yusay2oj, zwi7jdgz; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 218.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 218.2 16.7 19.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h3ieee67",
   "source_url": "https://omlx.ai/benchmarks/performance/h3ieee67",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-h3ieee67-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 7iz2k1h4, vywrywjp, yusay2oj, zwi7jdgz; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 306.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 306.9 18.2 21.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h3ieee67",
   "source_url": "https://omlx.ai/benchmarks/performance/h3ieee67",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-h3ieee67-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 7iz2k1h4, vywrywjp, yusay2oj, zwi7jdgz; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 298.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 8k 298.6 25.6 22.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h3ieee67",
   "source_url": "https://omlx.ai/benchmarks/performance/h3ieee67",
   "tg_tokens": null,
   "tps": 25.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-h3ieee67-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: 7iz2k1h4, vywrywjp, yusay2oj, zwi7jdgz; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 275.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 275.4 20.6 23.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h3ieee67",
   "source_url": "https://omlx.ai/benchmarks/performance/h3ieee67",
   "tg_tokens": null,
   "tps": 20.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-h3ieee67-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: 7iz2k1h4, vywrywjp, yusay2oj, zwi7jdgz; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 248.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 32k 248.0 6.9 26.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h3ieee67",
   "source_url": "https://omlx.ai/benchmarks/performance/h3ieee67",
   "tg_tokens": null,
   "tps": 6.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-h3ieee67-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 30.5 GB at this context; duplicate traces of the same session: 7iz2k1h4, vywrywjp, yusay2oj, zwi7jdgz; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 206.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 64k 206.8 4.3 30.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h3ieee67",
   "source_url": "https://omlx.ai/benchmarks/performance/h3ieee67",
   "tg_tokens": null,
   "tps": 4.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-7mp6fjr3-ctx1024",
   "model": "Bonsai-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 7.1 GB; TTFT 26097 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 39.2,
   "provenance": "community",
   "quant": "1bit",
   "quote": "Bonsai-27B-mlx-1bit on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 unknown; 1k tokens 39.2 PP tok/s 12.4 TG tok/s 26097 TTFT (ms) 7.1 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7mp6fjr3",
   "source_url": "https://omlx.ai/benchmarks/performance/7mp6fjr3",
   "tg_tokens": null,
   "tps": 12.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-7rucgs0p-ctx1024",
   "model": "Bonsai-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 7.1 GB; TTFT 26661 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 38.4,
   "provenance": "community",
   "quant": "1bit",
   "quote": "Bonsai-27B-mlx-1bit on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 unknown; 1k tokens 38.4 PP tok/s 11.7 TG tok/s 26661 TTFT (ms) 7.1 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7rucgs0p",
   "source_url": "https://omlx.ai/benchmarks/performance/7rucgs0p",
   "tg_tokens": null,
   "tps": 11.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-fq0p5mdm-ctx16384",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 12.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2054.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 2,054 118.8 12.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fq0p5mdm",
   "source_url": "https://omlx.ai/benchmarks/performance/fq0p5mdm",
   "tg_tokens": null,
   "tps": 118.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-fq0p5mdm-ctx32768",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 13.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1679.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 1,679 109.0 13.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fq0p5mdm",
   "source_url": "https://omlx.ai/benchmarks/performance/fq0p5mdm",
   "tg_tokens": null,
   "tps": 109.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-cho10ddt-ctx4096",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 12.7 GB; TTFT 1791 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2287.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k tokens 2,287 PP tok/s 124.6 TG tok/s 1791 TTFT (ms) 12.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cho10ddt",
   "source_url": "https://omlx.ai/benchmarks/performance/cho10ddt",
   "tg_tokens": null,
   "tps": 124.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-h9v0hl5c-ctx32768",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 13.1 GB; TTFT 18331 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1788.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k tokens 1,788 PP tok/s 109.6 TG tok/s 18331 TTFT (ms) 13.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h9v0hl5c",
   "source_url": "https://omlx.ai/benchmarks/performance/h9v0hl5c",
   "tg_tokens": null,
   "tps": 109.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-sl7iivor-ctx32768",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 13.1 GB; TTFT 19480 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1682.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k tokens 1,682 PP tok/s 111.6 TG tok/s 19480 TTFT (ms) 13.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sl7iivor",
   "source_url": "https://omlx.ai/benchmarks/performance/sl7iivor",
   "tg_tokens": null,
   "tps": 111.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-0ycul5zf-ctx16384",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2050.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-bf16 on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 bf16; 16k 2,050 80.9 23.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ycul5zf",
   "source_url": "https://omlx.ai/benchmarks/performance/0ycul5zf",
   "tg_tokens": null,
   "tps": 80.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-0ycul5zf-ctx32768",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1644.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-bf16 on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 bf16; 32k 1,644 76.3 23.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ycul5zf",
   "source_url": "https://omlx.ai/benchmarks/performance/0ycul5zf",
   "tg_tokens": null,
   "tps": 76.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ribhuqzt-ctx16384",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2048.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-bf16 on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 bf16; 16k 2,048 81.1 23.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ribhuqzt",
   "source_url": "https://omlx.ai/benchmarks/performance/ribhuqzt",
   "tg_tokens": null,
   "tps": 81.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ribhuqzt-ctx32768",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1622.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-bf16 on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 bf16; 32k 1,622 76.7 23.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ribhuqzt",
   "source_url": "https://omlx.ai/benchmarks/performance/ribhuqzt",
   "tg_tokens": null,
   "tps": 76.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-n5r8vg6d-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: b4r2e9aa; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1645.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 1,645 71.2 19.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n5r8vg6d",
   "source_url": "https://omlx.ai/benchmarks/performance/n5r8vg6d",
   "tg_tokens": null,
   "tps": 71.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-n5r8vg6d-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: b4r2e9aa; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2023.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 2,023 69.7 19.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n5r8vg6d",
   "source_url": "https://omlx.ai/benchmarks/performance/n5r8vg6d",
   "tg_tokens": null,
   "tps": 69.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-9qbk86b9-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: 7roup3ay; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1630.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 1,630 48.4 20.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9qbk86b9",
   "source_url": "https://omlx.ai/benchmarks/performance/9qbk86b9",
   "tg_tokens": null,
   "tps": 48.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-9qbk86b9-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: 7roup3ay; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1970.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,970 47.5 21.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9qbk86b9",
   "source_url": "https://omlx.ai/benchmarks/performance/9qbk86b9",
   "tg_tokens": null,
   "tps": 47.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2zrdsm6b-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.1 GB at this context; duplicate traces of the same session: 3gbiggpl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 225.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 225.5 15.5 33.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2zrdsm6b",
   "source_url": "https://omlx.ai/benchmarks/performance/2zrdsm6b",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2zrdsm6b-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: 3gbiggpl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 211.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 211.6 15.0 34.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2zrdsm6b",
   "source_url": "https://omlx.ai/benchmarks/performance/2zrdsm6b",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-uhamx5s3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: kipqnfd0; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 96.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 96.6 11.8 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uhamx5s3",
   "source_url": "https://omlx.ai/benchmarks/performance/uhamx5s3",
   "tg_tokens": null,
   "tps": 11.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-uhamx5s3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: kipqnfd0; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 90.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 90.7 13.2 19.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uhamx5s3",
   "source_url": "https://omlx.ai/benchmarks/performance/uhamx5s3",
   "tg_tokens": null,
   "tps": 13.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9o37vfut-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 34.2 GB at this context; duplicate traces of the same session: yx3wstqx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 259.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 259.5 29.0 34.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9o37vfut",
   "source_url": "https://omlx.ai/benchmarks/performance/9o37vfut",
   "tg_tokens": null,
   "tps": 29.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9o37vfut-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: yx3wstqx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 218.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 218.0 26.8 37.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9o37vfut",
   "source_url": "https://omlx.ai/benchmarks/performance/9o37vfut",
   "tg_tokens": null,
   "tps": 26.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9o37vfut-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 38.0 GB at this context; duplicate traces of the same session: yx3wstqx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 216.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 216.1 24.3 38.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9o37vfut",
   "source_url": "https://omlx.ai/benchmarks/performance/9o37vfut",
   "tg_tokens": null,
   "tps": 24.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-natyt23p-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: auc59til; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 162.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 162.9 16.9 20.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/natyt23p",
   "source_url": "https://omlx.ai/benchmarks/performance/natyt23p",
   "tg_tokens": null,
   "tps": 16.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-natyt23p-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: auc59til; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 289.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 289.4 16.6 21.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/natyt23p",
   "source_url": "https://omlx.ai/benchmarks/performance/natyt23p",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-6ud94zqn-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 105.8 GB; TTFT 4723 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 867.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k tokens 867.3 PP tok/s 53.6 TG tok/s 4723 TTFT (ms) 105.8 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ud94zqn",
   "source_url": "https://omlx.ai/benchmarks/performance/6ud94zqn",
   "tg_tokens": null,
   "tps": 53.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-qbk92re7-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 191.9 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 852.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-Flash-Next-MLX-8bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 852.2 63.2 191.9 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qbk92re7",
   "source_url": "https://omlx.ai/benchmarks/performance/qbk92re7",
   "tg_tokens": null,
   "tps": 63.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qbk92re7-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 192.4 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1197.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-Flash-Next-MLX-8bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 1,197 54.7 192.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qbk92re7",
   "source_url": "https://omlx.ai/benchmarks/performance/qbk92re7",
   "tg_tokens": null,
   "tps": 54.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qbk92re7-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 192.5 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1208.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-Flash-Next-MLX-8bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 1,208 46.8 192.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qbk92re7",
   "source_url": "https://omlx.ai/benchmarks/performance/qbk92re7",
   "tg_tokens": null,
   "tps": 46.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-qbk92re7-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 192.8 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1209.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-Flash-Next-MLX-8bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k 1,209 59.8 192.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qbk92re7",
   "source_url": "https://omlx.ai/benchmarks/performance/qbk92re7",
   "tg_tokens": null,
   "tps": 59.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3me7kpv9-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 122.2 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1063.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 5bit; 1k 1,063 99.1 122.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3me7kpv9",
   "source_url": "https://omlx.ai/benchmarks/performance/3me7kpv9",
   "tg_tokens": null,
   "tps": 99.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3me7kpv9-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 122.8 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1348.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 5bit; 4k 1,348 91.3 122.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3me7kpv9",
   "source_url": "https://omlx.ai/benchmarks/performance/3me7kpv9",
   "tg_tokens": null,
   "tps": 91.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-3me7kpv9-ctx8192",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 123.0 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1343.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 5bit; 8k 1,343 86.1 123.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3me7kpv9",
   "source_url": "https://omlx.ai/benchmarks/performance/3me7kpv9",
   "tg_tokens": null,
   "tps": 86.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-avj1hsd9-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: pv5k3t80; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 737.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 737.1 46.2 100.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/avj1hsd9",
   "source_url": "https://omlx.ai/benchmarks/performance/avj1hsd9",
   "tg_tokens": null,
   "tps": 46.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-avj1hsd9-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.5 GB at this context; duplicate traces of the same session: pv5k3t80; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 661.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 661.5 66.3 102.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/avj1hsd9",
   "source_url": "https://omlx.ai/benchmarks/performance/avj1hsd9",
   "tg_tokens": null,
   "tps": 66.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wjw2fhh6-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1123.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,123 87.7 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wjw2fhh6",
   "source_url": "https://omlx.ai/benchmarks/performance/wjw2fhh6",
   "tg_tokens": null,
   "tps": 87.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wjw2fhh6-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2475.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,475 65.9 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wjw2fhh6",
   "source_url": "https://omlx.ai/benchmarks/performance/wjw2fhh6",
   "tg_tokens": null,
   "tps": 65.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wjw2fhh6-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2660.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,660 71.3 72.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wjw2fhh6",
   "source_url": "https://omlx.ai/benchmarks/performance/wjw2fhh6",
   "tg_tokens": null,
   "tps": 71.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wjw2fhh6-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2263.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,263 63.1 80.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wjw2fhh6",
   "source_url": "https://omlx.ai/benchmarks/performance/wjw2fhh6",
   "tg_tokens": null,
   "tps": 63.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1ctk4ngb-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1967.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,967 35.0 71.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ctk4ngb",
   "source_url": "https://omlx.ai/benchmarks/performance/1ctk4ngb",
   "tg_tokens": null,
   "tps": 35.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1ctk4ngb-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2418.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,418 60.5 78.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ctk4ngb",
   "source_url": "https://omlx.ai/benchmarks/performance/1ctk4ngb",
   "tg_tokens": null,
   "tps": 60.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1ctk4ngb-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2252.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,252 63.8 80.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ctk4ngb",
   "source_url": "https://omlx.ai/benchmarks/performance/1ctk4ngb",
   "tg_tokens": null,
   "tps": 63.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rms21oyz-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: rfl4hdug; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2401.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,401 70.7 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rms21oyz",
   "source_url": "https://omlx.ai/benchmarks/performance/rms21oyz",
   "tg_tokens": null,
   "tps": 70.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rms21oyz-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: rfl4hdug; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2619.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,619 72.9 73.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rms21oyz",
   "source_url": "https://omlx.ai/benchmarks/performance/rms21oyz",
   "tg_tokens": null,
   "tps": 72.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rms21oyz-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.6 GB at this context; duplicate traces of the same session: rfl4hdug; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2524.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,524 58.1 74.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rms21oyz",
   "source_url": "https://omlx.ai/benchmarks/performance/rms21oyz",
   "tg_tokens": null,
   "tps": 58.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-o73ub4km-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 70.8 GB; TTFT 1743 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2351.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 2,351 PP tok/s 76.8 TG tok/s 1743 TTFT (ms) 70.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o73ub4km",
   "source_url": "https://omlx.ai/benchmarks/performance/o73ub4km",
   "tg_tokens": null,
   "tps": 76.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-33dwtg7z-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Steven10429 UD build; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 5jwi932s; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 210.6,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "Steven10429--Qwen3.8-27B-UD-IQ4_XS-MLX on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 210.6 29.5 17.6 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/33dwtg7z",
   "source_url": "https://omlx.ai/benchmarks/performance/33dwtg7z",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-33dwtg7z-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "Steven10429 UD build; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 5jwi932s; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 209.9,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "Steven10429--Qwen3.8-27B-UD-IQ4_XS-MLX on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 209.9 25.4 18.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/33dwtg7z",
   "source_url": "https://omlx.ai/benchmarks/performance/33dwtg7z",
   "tg_tokens": null,
   "tps": 25.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-33dwtg7z-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "Steven10429 UD build; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 5jwi932s; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 207.9,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "Steven10429--Qwen3.8-27B-UD-IQ4_XS-MLX on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 207.9 24.3 19.4 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/33dwtg7z",
   "source_url": "https://omlx.ai/benchmarks/performance/33dwtg7z",
   "tg_tokens": null,
   "tps": 24.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-wcmudw50-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Steven10429 UD build; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 36xx1viz; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 210.1,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "Steven10429--Qwen3.8-27B-UD-IQ4_XS-MLX on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 210.1 29.1 17.6 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wcmudw50",
   "source_url": "https://omlx.ai/benchmarks/performance/wcmudw50",
   "tg_tokens": null,
   "tps": 29.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-wcmudw50-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "Steven10429 UD build; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 36xx1viz; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 210.8,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "Steven10429--Qwen3.8-27B-UD-IQ4_XS-MLX on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 210.8 23.3 18.8 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wcmudw50",
   "source_url": "https://omlx.ai/benchmarks/performance/wcmudw50",
   "tg_tokens": null,
   "tps": 23.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-wcmudw50-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "Steven10429 UD build; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 36xx1viz; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 207.9,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "Steven10429--Qwen3.8-27B-UD-IQ4_XS-MLX on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 207.9 21.1 19.4 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wcmudw50",
   "source_url": "https://omlx.ai/benchmarks/performance/wcmudw50",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-quvvpo0m-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 3grss6p1, wq1rwsr4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1124.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,124 75.3 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/quvvpo0m",
   "source_url": "https://omlx.ai/benchmarks/performance/quvvpo0m",
   "tg_tokens": null,
   "tps": 75.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-quvvpo0m-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 3grss6p1, wq1rwsr4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1336.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,336 64.6 19.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/quvvpo0m",
   "source_url": "https://omlx.ai/benchmarks/performance/quvvpo0m",
   "tg_tokens": null,
   "tps": 64.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-quvvpo0m-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 3grss6p1, wq1rwsr4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1349.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,349 69.9 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/quvvpo0m",
   "source_url": "https://omlx.ai/benchmarks/performance/quvvpo0m",
   "tg_tokens": null,
   "tps": 69.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-k5ql0j09-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 0b3q8doi, ersxe269; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1126.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,126 62.4 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k5ql0j09",
   "source_url": "https://omlx.ai/benchmarks/performance/k5ql0j09",
   "tg_tokens": null,
   "tps": 62.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-k5ql0j09-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 0b3q8doi, ersxe269; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1334.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,334 66.1 19.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k5ql0j09",
   "source_url": "https://omlx.ai/benchmarks/performance/k5ql0j09",
   "tg_tokens": null,
   "tps": 66.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-k5ql0j09-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 0b3q8doi, ersxe269; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1337.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,337 65.1 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k5ql0j09",
   "source_url": "https://omlx.ai/benchmarks/performance/k5ql0j09",
   "tg_tokens": null,
   "tps": 65.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-to3e0k7u-ctx1024",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "YCF-AI MTP FP16 build; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: s2f23ysa; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 203.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "YCF-AI--Qwen3.8-27B-Huihui-abliterated-oQ4e-MTP-FP16-MLX on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 203.6 29.6 18.5 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/to3e0k7u",
   "source_url": "https://omlx.ai/benchmarks/performance/to3e0k7u",
   "tg_tokens": null,
   "tps": 29.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-to3e0k7u-ctx4096",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "YCF-AI MTP FP16 build; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: s2f23ysa; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "YCF-AI--Qwen3.8-27B-Huihui-abliterated-oQ4e-MTP-FP16-MLX on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 204.0 32.0 19.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/to3e0k7u",
   "source_url": "https://omlx.ai/benchmarks/performance/to3e0k7u",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-to3e0k7u-ctx8192",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "YCF-AI MTP FP16 build; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: s2f23ysa; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 202.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "YCF-AI--Qwen3.8-27B-Huihui-abliterated-oQ4e-MTP-FP16-MLX on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 202.6 29.2 20.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/to3e0k7u",
   "source_url": "https://omlx.ai/benchmarks/performance/to3e0k7u",
   "tg_tokens": null,
   "tps": 29.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-to3e0k7u-ctx16384",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "YCF-AI MTP FP16 build; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: s2f23ysa; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 194.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "YCF-AI--Qwen3.8-27B-Huihui-abliterated-oQ4e-MTP-FP16-MLX on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 194.5 29.7 20.4 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/to3e0k7u",
   "source_url": "https://omlx.ai/benchmarks/performance/to3e0k7u",
   "tg_tokens": null,
   "tps": 29.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-o12ava61-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "chriswessels ThinkingCap; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 203.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "chriswessels--ThinkingCap-Qwen3.8-27B-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 203.1 32.9 18.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o12ava61",
   "source_url": "https://omlx.ai/benchmarks/performance/o12ava61",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-o12ava61-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "chriswessels ThinkingCap; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "chriswessels--ThinkingCap-Qwen3.8-27B-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 205.4 34.9 19.8 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o12ava61",
   "source_url": "https://omlx.ai/benchmarks/performance/o12ava61",
   "tg_tokens": null,
   "tps": 34.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-pzkq01al-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "fcmeyer MTP build; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: e2ahkv6x; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "fcmeyer--Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 204.0 30.3 18.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pzkq01al",
   "source_url": "https://omlx.ai/benchmarks/performance/pzkq01al",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-pzkq01al-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "fcmeyer MTP build; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: e2ahkv6x; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "fcmeyer--Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 204.9 32.3 19.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pzkq01al",
   "source_url": "https://omlx.ai/benchmarks/performance/pzkq01al",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-pzkq01al-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "fcmeyer MTP build; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: e2ahkv6x; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 202.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "fcmeyer--Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 202.6 33.7 20.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pzkq01al",
   "source_url": "https://omlx.ai/benchmarks/performance/pzkq01al",
   "tg_tokens": null,
   "tps": 33.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-pzkq01al-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "fcmeyer MTP build; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: e2ahkv6x; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 194.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "fcmeyer--Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 194.3 28.2 20.4 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pzkq01al",
   "source_url": "https://omlx.ai/benchmarks/performance/pzkq01al",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-31llko7u-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "fcmeyer MTP build; oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: 47d7yd3r; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 203.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "fcmeyer--Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 203.5 34.5 18.6 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/31llko7u",
   "source_url": "https://omlx.ai/benchmarks/performance/31llko7u",
   "tg_tokens": null,
   "tps": 34.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-31llko7u-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "fcmeyer MTP build; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 47d7yd3r; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "fcmeyer--Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 204.2 34.6 19.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/31llko7u",
   "source_url": "https://omlx.ai/benchmarks/performance/31llko7u",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-31llko7u-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "fcmeyer MTP build; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: 47d7yd3r; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 202.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "fcmeyer--Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 202.7 31.8 20.5 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/31llko7u",
   "source_url": "https://omlx.ai/benchmarks/performance/31llko7u",
   "tg_tokens": null,
   "tps": 31.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-i62jp0kk-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "leonsarmiento pi MTP; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "leonsarmiento--Qwen3.8-27B-pi-4bit-mtp-mlx on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 204.8 31.1 19.3 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i62jp0kk",
   "source_url": "https://omlx.ai/benchmarks/performance/i62jp0kk",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-i62jp0kk-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "leonsarmiento pi MTP; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 206.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "leonsarmiento--Qwen3.8-27B-pi-4bit-mtp-mlx on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 206.2 36.6 20.2 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i62jp0kk",
   "source_url": "https://omlx.ai/benchmarks/performance/i62jp0kk",
   "tg_tokens": null,
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-i62jp0kk-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "leonsarmiento pi MTP; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 203.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "leonsarmiento--Qwen3.8-27B-pi-4bit-mtp-mlx on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 203.5 34.1 20.9 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i62jp0kk",
   "source_url": "https://omlx.ai/benchmarks/performance/i62jp0kk",
   "tg_tokens": null,
   "tps": 34.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-i62jp0kk-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "leonsarmiento pi MTP; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 195.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "leonsarmiento--Qwen3.8-27B-pi-4bit-mtp-mlx on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 195.0 30.6 20.7 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i62jp0kk",
   "source_url": "https://omlx.ai/benchmarks/performance/i62jp0kk",
   "tg_tokens": null,
   "tps": 30.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-trx6cgey-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "leonsarmiento pi MTP; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: uk3tp1ep; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "leonsarmiento--Qwen3.8-27B-pi-4bit-mtp-mlx on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 204.4 34.6 18.9 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/trx6cgey",
   "source_url": "https://omlx.ai/benchmarks/performance/trx6cgey",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-trx6cgey-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "leonsarmiento pi MTP; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: uk3tp1ep; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 206.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "leonsarmiento--Qwen3.8-27B-pi-4bit-mtp-mlx on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 206.2 38.0 20.2 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/trx6cgey",
   "source_url": "https://omlx.ai/benchmarks/performance/trx6cgey",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-trx6cgey-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "leonsarmiento pi MTP; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: uk3tp1ep; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 203.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "leonsarmiento--Qwen3.8-27B-pi-4bit-mtp-mlx on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 203.5 28.1 20.9 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/trx6cgey",
   "source_url": "https://omlx.ai/benchmarks/performance/trx6cgey",
   "tg_tokens": null,
   "tps": 28.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-7iz2k1h4-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: vywrywjp, yusay2oj, zwi7jdgz; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 218.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 218.2 16.7 19.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7iz2k1h4",
   "source_url": "https://omlx.ai/benchmarks/performance/7iz2k1h4",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-7iz2k1h4-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: vywrywjp, yusay2oj, zwi7jdgz; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 306.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 306.9 18.2 21.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7iz2k1h4",
   "source_url": "https://omlx.ai/benchmarks/performance/7iz2k1h4",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-7iz2k1h4-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: vywrywjp, yusay2oj, zwi7jdgz; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 298.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 8k 298.6 25.6 22.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7iz2k1h4",
   "source_url": "https://omlx.ai/benchmarks/performance/7iz2k1h4",
   "tg_tokens": null,
   "tps": 25.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-7iz2k1h4-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: vywrywjp, yusay2oj, zwi7jdgz; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 275.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 275.4 20.6 23.5 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7iz2k1h4",
   "source_url": "https://omlx.ai/benchmarks/performance/7iz2k1h4",
   "tg_tokens": null,
   "tps": 20.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-7iz2k1h4-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: vywrywjp, yusay2oj, zwi7jdgz; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 248.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 32k 248.0 6.9 26.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7iz2k1h4",
   "source_url": "https://omlx.ai/benchmarks/performance/7iz2k1h4",
   "tg_tokens": null,
   "tps": 6.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-7iz2k1h4-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 30.5 GB at this context; duplicate traces of the same session: vywrywjp, yusay2oj, zwi7jdgz; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 206.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 64k 206.8 4.3 30.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7iz2k1h4",
   "source_url": "https://omlx.ai/benchmarks/performance/7iz2k1h4",
   "tg_tokens": null,
   "tps": 4.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-wc50svnb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: sdzba6rj; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 213.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 213.6 17.2 19.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wc50svnb",
   "source_url": "https://omlx.ai/benchmarks/performance/wc50svnb",
   "tg_tokens": null,
   "tps": 17.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-wc50svnb-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: sdzba6rj; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 306.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 306.7 18.5 21.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wc50svnb",
   "source_url": "https://omlx.ai/benchmarks/performance/wc50svnb",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-rmx6sxu7-ctx1024",
   "model": "Bonsai-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 7.1 GB; TTFT 26778 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 38.2,
   "provenance": "community",
   "quant": "1bit",
   "quote": "Bonsai-27B-mlx-1bit on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 unknown; 1k tokens 38.2 PP tok/s 12.1 TG tok/s 26778 TTFT (ms) 7.1 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rmx6sxu7",
   "source_url": "https://omlx.ai/benchmarks/performance/rmx6sxu7",
   "tg_tokens": null,
   "tps": 12.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-b4r2e9aa-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1645.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 1,645 71.2 19.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b4r2e9aa",
   "source_url": "https://omlx.ai/benchmarks/performance/b4r2e9aa",
   "tg_tokens": null,
   "tps": 71.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-b4r2e9aa-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2023.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 2,023 69.7 19.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b4r2e9aa",
   "source_url": "https://omlx.ai/benchmarks/performance/b4r2e9aa",
   "tg_tokens": null,
   "tps": 69.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-usilvm3u-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 5lwsrfa7, 5m8n590y, 6xveliwu, grnsxxa2, o2ks9cjv; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1646.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 1,646 69.5 19.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/usilvm3u",
   "source_url": "https://omlx.ai/benchmarks/performance/usilvm3u",
   "tg_tokens": null,
   "tps": 69.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-usilvm3u-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 5lwsrfa7, 5m8n590y, 6xveliwu, grnsxxa2, o2ks9cjv; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1966.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,966 70.0 19.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/usilvm3u",
   "source_url": "https://omlx.ai/benchmarks/performance/usilvm3u",
   "tg_tokens": null,
   "tps": 70.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-usilvm3u-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 5lwsrfa7, 5m8n590y, 6xveliwu, grnsxxa2, o2ks9cjv; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1937.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 8k 1,937 66.7 19.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/usilvm3u",
   "source_url": "https://omlx.ai/benchmarks/performance/usilvm3u",
   "tg_tokens": null,
   "tps": 66.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-usilvm3u-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 5lwsrfa7, 5m8n590y, 6xveliwu, grnsxxa2, o2ks9cjv; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1802.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 1,802 61.5 20.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/usilvm3u",
   "source_url": "https://omlx.ai/benchmarks/performance/usilvm3u",
   "tg_tokens": null,
   "tps": 61.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-usilvm3u-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: 5lwsrfa7, 5m8n590y, 6xveliwu, grnsxxa2, o2ks9cjv; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1520.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 32k 1,520 51.8 20.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/usilvm3u",
   "source_url": "https://omlx.ai/benchmarks/performance/usilvm3u",
   "tg_tokens": null,
   "tps": 51.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-usilvm3u-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: 5lwsrfa7, 5m8n590y, 6xveliwu, grnsxxa2, o2ks9cjv; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1133.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 64k 1,133 38.0 21.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/usilvm3u",
   "source_url": "https://omlx.ai/benchmarks/performance/usilvm3u",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-7roup3ay-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1630.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 1,630 48.4 20.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7roup3ay",
   "source_url": "https://omlx.ai/benchmarks/performance/7roup3ay",
   "tg_tokens": null,
   "tps": 48.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-7roup3ay-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1970.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,970 47.5 21.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7roup3ay",
   "source_url": "https://omlx.ai/benchmarks/performance/7roup3ay",
   "tg_tokens": null,
   "tps": 47.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-3gbiggpl-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 225.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 225.5 15.5 33.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3gbiggpl",
   "source_url": "https://omlx.ai/benchmarks/performance/3gbiggpl",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-3gbiggpl-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 211.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 211.6 15.0 34.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3gbiggpl",
   "source_url": "https://omlx.ai/benchmarks/performance/3gbiggpl",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-kipqnfd0-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 96.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 96.6 11.8 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kipqnfd0",
   "source_url": "https://omlx.ai/benchmarks/performance/kipqnfd0",
   "tg_tokens": null,
   "tps": 11.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-kipqnfd0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 90.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 90.7 13.2 19.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kipqnfd0",
   "source_url": "https://omlx.ai/benchmarks/performance/kipqnfd0",
   "tg_tokens": null,
   "tps": 13.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ragy1fqe-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: zl7vo72a; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 498.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 498.5 17.4 17.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ragy1fqe",
   "source_url": "https://omlx.ai/benchmarks/performance/ragy1fqe",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ragy1fqe-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: zl7vo72a; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 465.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 465.0 16.3 18.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ragy1fqe",
   "source_url": "https://omlx.ai/benchmarks/performance/ragy1fqe",
   "tg_tokens": null,
   "tps": 16.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-8lq7ynkr-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: f4d96pyl; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 477.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 477.6 22.5 16.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8lq7ynkr",
   "source_url": "https://omlx.ai/benchmarks/performance/8lq7ynkr",
   "tg_tokens": null,
   "tps": 22.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-8lq7ynkr-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.2 GB at this context; duplicate traces of the same session: f4d96pyl; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 475.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 475.8 24.2 18.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8lq7ynkr",
   "source_url": "https://omlx.ai/benchmarks/performance/8lq7ynkr",
   "tg_tokens": null,
   "tps": 24.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yx3wstqx-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 34.2 GB at this context; duplicate traces of the same session: 7qgayjzv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 259.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 259.5 29.0 34.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yx3wstqx",
   "source_url": "https://omlx.ai/benchmarks/performance/yx3wstqx",
   "tg_tokens": null,
   "tps": 29.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yx3wstqx-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: 7qgayjzv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 218.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 218.0 26.8 37.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yx3wstqx",
   "source_url": "https://omlx.ai/benchmarks/performance/yx3wstqx",
   "tg_tokens": null,
   "tps": 26.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yx3wstqx-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 38.0 GB at this context; duplicate traces of the same session: 7qgayjzv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 216.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 216.1 24.3 38.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yx3wstqx",
   "source_url": "https://omlx.ai/benchmarks/performance/yx3wstqx",
   "tg_tokens": null,
   "tps": 24.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-auc59til-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 162.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 162.9 16.9 20.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/auc59til",
   "source_url": "https://omlx.ai/benchmarks/performance/auc59til",
   "tg_tokens": null,
   "tps": 16.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-auc59til-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 289.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 289.4 16.6 21.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/auc59til",
   "source_url": "https://omlx.ai/benchmarks/performance/auc59til",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-aa1kfatl-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: gl6rj7dv, jbd4ntm0, n716phs3, qklxi5h3, x5fw8qpx; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 277.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 277.0 21.9 17.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aa1kfatl",
   "source_url": "https://omlx.ai/benchmarks/performance/aa1kfatl",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-aa1kfatl-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: gl6rj7dv, jbd4ntm0, n716phs3, qklxi5h3, x5fw8qpx; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 233.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 233.1 20.5 18.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aa1kfatl",
   "source_url": "https://omlx.ai/benchmarks/performance/aa1kfatl",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-aa1kfatl-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: gl6rj7dv, jbd4ntm0, n716phs3, qklxi5h3, x5fw8qpx; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 235.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 8k 235.4 26.1 18.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aa1kfatl",
   "source_url": "https://omlx.ai/benchmarks/performance/aa1kfatl",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-aa1kfatl-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: gl6rj7dv, jbd4ntm0, n716phs3, qklxi5h3, x5fw8qpx; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 222.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 222.4 23.0 19.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aa1kfatl",
   "source_url": "https://omlx.ai/benchmarks/performance/aa1kfatl",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-aa1kfatl-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: gl6rj7dv, jbd4ntm0, n716phs3, qklxi5h3, x5fw8qpx; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 214.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 32k 214.9 25.3 20.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aa1kfatl",
   "source_url": "https://omlx.ai/benchmarks/performance/aa1kfatl",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-aa1kfatl-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: gl6rj7dv, jbd4ntm0, n716phs3, qklxi5h3, x5fw8qpx; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 193.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 64k 193.3 17.9 23.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aa1kfatl",
   "source_url": "https://omlx.ai/benchmarks/performance/aa1kfatl",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-pv5k3t80-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 737.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 737.1 46.2 100.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pv5k3t80",
   "source_url": "https://omlx.ai/benchmarks/performance/pv5k3t80",
   "tg_tokens": null,
   "tps": 46.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-pv5k3t80-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.5 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 661.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 661.5 66.3 102.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pv5k3t80",
   "source_url": "https://omlx.ai/benchmarks/performance/pv5k3t80",
   "tg_tokens": null,
   "tps": 66.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rfl4hdug-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2401.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,401 70.7 70.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rfl4hdug",
   "source_url": "https://omlx.ai/benchmarks/performance/rfl4hdug",
   "tg_tokens": null,
   "tps": 70.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rfl4hdug-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2619.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,619 72.9 73.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rfl4hdug",
   "source_url": "https://omlx.ai/benchmarks/performance/rfl4hdug",
   "tg_tokens": null,
   "tps": 72.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rfl4hdug-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2524.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,524 58.1 74.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rfl4hdug",
   "source_url": "https://omlx.ai/benchmarks/performance/rfl4hdug",
   "tg_tokens": null,
   "tps": 58.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mhjb3t8v-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: 8fze8ktg, ytj1j6kv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2453.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,453 84.2 71.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mhjb3t8v",
   "source_url": "https://omlx.ai/benchmarks/performance/mhjb3t8v",
   "tg_tokens": null,
   "tps": 84.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mhjb3t8v-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.7 GB at this context; duplicate traces of the same session: 8fze8ktg, ytj1j6kv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2586.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,586 58.2 72.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mhjb3t8v",
   "source_url": "https://omlx.ai/benchmarks/performance/mhjb3t8v",
   "tg_tokens": null,
   "tps": 58.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mhjb3t8v-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.0 GB at this context; duplicate traces of the same session: 8fze8ktg, ytj1j6kv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2422.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,422 72.3 79.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mhjb3t8v",
   "source_url": "https://omlx.ai/benchmarks/performance/mhjb3t8v",
   "tg_tokens": null,
   "tps": 72.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rq4esfmt-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 71.1 GB; TTFT 1665 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2460.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 2,460 PP tok/s 79.6 TG tok/s 1665 TTFT (ms) 71.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rq4esfmt",
   "source_url": "https://omlx.ai/benchmarks/performance/rq4esfmt",
   "tg_tokens": null,
   "tps": 79.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ce8ksjgx-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: h0936z6p; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1128.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,128 75.0 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ce8ksjgx",
   "source_url": "https://omlx.ai/benchmarks/performance/ce8ksjgx",
   "tg_tokens": null,
   "tps": 75.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ce8ksjgx-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: h0936z6p; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1763.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,763 71.4 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ce8ksjgx",
   "source_url": "https://omlx.ai/benchmarks/performance/ce8ksjgx",
   "tg_tokens": null,
   "tps": 71.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ce8ksjgx-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: h0936z6p; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2155.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,155 88.9 71.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ce8ksjgx",
   "source_url": "https://omlx.ai/benchmarks/performance/ce8ksjgx",
   "tg_tokens": null,
   "tps": 88.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ce8ksjgx-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.7 GB at this context; duplicate traces of the same session: h0936z6p; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2327.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,327 78.7 72.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ce8ksjgx",
   "source_url": "https://omlx.ai/benchmarks/performance/ce8ksjgx",
   "tg_tokens": null,
   "tps": 78.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ce8ksjgx-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.5 GB at this context; duplicate traces of the same session: h0936z6p; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2213.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,213 63.2 80.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ce8ksjgx",
   "source_url": "https://omlx.ai/benchmarks/performance/ce8ksjgx",
   "tg_tokens": null,
   "tps": 63.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-5jwi932s-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Steven10429 UD build; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 210.6,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "Steven10429--Qwen3.8-27B-UD-IQ4_XS-MLX on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 210.6 29.5 17.6 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5jwi932s",
   "source_url": "https://omlx.ai/benchmarks/performance/5jwi932s",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-5jwi932s-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "Steven10429 UD build; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 209.9,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "Steven10429--Qwen3.8-27B-UD-IQ4_XS-MLX on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 209.9 25.4 18.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5jwi932s",
   "source_url": "https://omlx.ai/benchmarks/performance/5jwi932s",
   "tg_tokens": null,
   "tps": 25.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-5jwi932s-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "Steven10429 UD build; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 207.9,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "Steven10429--Qwen3.8-27B-UD-IQ4_XS-MLX on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 207.9 24.3 19.4 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5jwi932s",
   "source_url": "https://omlx.ai/benchmarks/performance/5jwi932s",
   "tg_tokens": null,
   "tps": 24.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-36xx1viz-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Steven10429 UD build; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 210.1,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "Steven10429--Qwen3.8-27B-UD-IQ4_XS-MLX on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 210.1 29.1 17.6 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/36xx1viz",
   "source_url": "https://omlx.ai/benchmarks/performance/36xx1viz",
   "tg_tokens": null,
   "tps": 29.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-36xx1viz-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "Steven10429 UD build; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 210.8,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "Steven10429--Qwen3.8-27B-UD-IQ4_XS-MLX on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 210.8 23.3 18.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/36xx1viz",
   "source_url": "https://omlx.ai/benchmarks/performance/36xx1viz",
   "tg_tokens": null,
   "tps": 23.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-36xx1viz-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "Steven10429 UD build; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 207.9,
   "provenance": "community",
   "quant": "IQ4_XS",
   "quote": "Steven10429--Qwen3.8-27B-UD-IQ4_XS-MLX on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 207.9 21.1 19.4 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/36xx1viz",
   "source_url": "https://omlx.ai/benchmarks/performance/36xx1viz",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3grss6p1-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: wq1rwsr4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1124.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,124 75.3 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3grss6p1",
   "source_url": "https://omlx.ai/benchmarks/performance/3grss6p1",
   "tg_tokens": null,
   "tps": 75.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3grss6p1-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: wq1rwsr4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1336.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,336 64.6 19.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3grss6p1",
   "source_url": "https://omlx.ai/benchmarks/performance/3grss6p1",
   "tg_tokens": null,
   "tps": 64.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3grss6p1-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: wq1rwsr4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1349.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,349 69.9 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3grss6p1",
   "source_url": "https://omlx.ai/benchmarks/performance/3grss6p1",
   "tg_tokens": null,
   "tps": 69.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0b3q8doi-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: ersxe269; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1126.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,126 62.4 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0b3q8doi",
   "source_url": "https://omlx.ai/benchmarks/performance/0b3q8doi",
   "tg_tokens": null,
   "tps": 62.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0b3q8doi-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: ersxe269; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1334.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,334 66.1 19.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0b3q8doi",
   "source_url": "https://omlx.ai/benchmarks/performance/0b3q8doi",
   "tg_tokens": null,
   "tps": 66.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0b3q8doi-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: ersxe269; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1337.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,337 65.1 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0b3q8doi",
   "source_url": "https://omlx.ai/benchmarks/performance/0b3q8doi",
   "tg_tokens": null,
   "tps": 65.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-s2f23ysa-ctx1024",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "YCF-AI MTP FP16 build; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 203.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "YCF-AI--Qwen3.8-27B-Huihui-abliterated-oQ4e-MTP-FP16-MLX on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 203.6 29.6 18.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s2f23ysa",
   "source_url": "https://omlx.ai/benchmarks/performance/s2f23ysa",
   "tg_tokens": null,
   "tps": 29.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-s2f23ysa-ctx4096",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "YCF-AI MTP FP16 build; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "YCF-AI--Qwen3.8-27B-Huihui-abliterated-oQ4e-MTP-FP16-MLX on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 204.0 32.0 19.8 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s2f23ysa",
   "source_url": "https://omlx.ai/benchmarks/performance/s2f23ysa",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-s2f23ysa-ctx8192",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "YCF-AI MTP FP16 build; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 202.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "YCF-AI--Qwen3.8-27B-Huihui-abliterated-oQ4e-MTP-FP16-MLX on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 202.6 29.2 20.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s2f23ysa",
   "source_url": "https://omlx.ai/benchmarks/performance/s2f23ysa",
   "tg_tokens": null,
   "tps": 29.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-s2f23ysa-ctx16384",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "YCF-AI MTP FP16 build; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 194.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "YCF-AI--Qwen3.8-27B-Huihui-abliterated-oQ4e-MTP-FP16-MLX on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 194.5 29.7 20.4 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s2f23ysa",
   "source_url": "https://omlx.ai/benchmarks/performance/s2f23ysa",
   "tg_tokens": null,
   "tps": 29.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ms1klt64-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "chriswessels ThinkingCap; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: c250f0py; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 203.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "chriswessels--ThinkingCap-Qwen3.8-27B-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 203.7 36.5 18.6 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ms1klt64",
   "source_url": "https://omlx.ai/benchmarks/performance/ms1klt64",
   "tg_tokens": null,
   "tps": 36.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ms1klt64-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "chriswessels ThinkingCap; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: c250f0py; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "chriswessels--ThinkingCap-Qwen3.8-27B-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 205.2 30.4 19.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ms1klt64",
   "source_url": "https://omlx.ai/benchmarks/performance/ms1klt64",
   "tg_tokens": null,
   "tps": 30.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ms1klt64-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "chriswessels ThinkingCap; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: c250f0py; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 202.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "chriswessels--ThinkingCap-Qwen3.8-27B-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 202.7 34.3 20.5 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ms1klt64",
   "source_url": "https://omlx.ai/benchmarks/performance/ms1klt64",
   "tg_tokens": null,
   "tps": 34.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ms1klt64-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "chriswessels ThinkingCap; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: c250f0py; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 194.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "chriswessels--ThinkingCap-Qwen3.8-27B-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 194.3 29.0 20.4 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ms1klt64",
   "source_url": "https://omlx.ai/benchmarks/performance/ms1klt64",
   "tg_tokens": null,
   "tps": 29.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-e2ahkv6x-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "fcmeyer MTP build; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "fcmeyer--Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 204.0 30.3 18.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e2ahkv6x",
   "source_url": "https://omlx.ai/benchmarks/performance/e2ahkv6x",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-e2ahkv6x-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "fcmeyer MTP build; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "fcmeyer--Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 204.9 32.3 19.8 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e2ahkv6x",
   "source_url": "https://omlx.ai/benchmarks/performance/e2ahkv6x",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-e2ahkv6x-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "fcmeyer MTP build; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 202.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "fcmeyer--Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 202.6 33.7 20.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e2ahkv6x",
   "source_url": "https://omlx.ai/benchmarks/performance/e2ahkv6x",
   "tg_tokens": null,
   "tps": 33.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-e2ahkv6x-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "fcmeyer MTP build; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 194.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "fcmeyer--Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 194.3 28.2 20.4 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e2ahkv6x",
   "source_url": "https://omlx.ai/benchmarks/performance/e2ahkv6x",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-47d7yd3r-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "fcmeyer MTP build; oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 203.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "fcmeyer--Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 203.5 34.5 18.6 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/47d7yd3r",
   "source_url": "https://omlx.ai/benchmarks/performance/47d7yd3r",
   "tg_tokens": null,
   "tps": 34.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-47d7yd3r-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "fcmeyer MTP build; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "fcmeyer--Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 204.2 34.6 19.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/47d7yd3r",
   "source_url": "https://omlx.ai/benchmarks/performance/47d7yd3r",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-47d7yd3r-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "fcmeyer MTP build; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 202.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "fcmeyer--Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 202.7 31.8 20.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/47d7yd3r",
   "source_url": "https://omlx.ai/benchmarks/performance/47d7yd3r",
   "tg_tokens": null,
   "tps": 31.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-po6wv3n0-ctx16384",
   "model": "gemma-4-31B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: vjbasx6l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 175.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-qat-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 175.0 12.7 34.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/po6wv3n0",
   "source_url": "https://omlx.ai/benchmarks/performance/po6wv3n0",
   "tg_tokens": null,
   "tps": 12.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-po6wv3n0-ctx32768",
   "model": "gemma-4-31B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 37.0 GB at this context; duplicate traces of the same session: vjbasx6l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 168.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-qat-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 168.9 11.9 37.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/po6wv3n0",
   "source_url": "https://omlx.ai/benchmarks/performance/po6wv3n0",
   "tg_tokens": null,
   "tps": 11.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ra0rfkt4-ctx16384",
   "model": "gemma-4-31B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: z6nc7s0o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 175.8,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "gemma-4-31B-it-qat-oQ8 on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 175.8 12.7 34.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ra0rfkt4",
   "source_url": "https://omlx.ai/benchmarks/performance/ra0rfkt4",
   "tg_tokens": null,
   "tps": 12.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ra0rfkt4-ctx32768",
   "model": "gemma-4-31B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 37.0 GB at this context; duplicate traces of the same session: z6nc7s0o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 169.7,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "gemma-4-31B-it-qat-oQ8 on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 169.7 12.0 37.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ra0rfkt4",
   "source_url": "https://omlx.ai/benchmarks/performance/ra0rfkt4",
   "tg_tokens": null,
   "tps": 12.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-uk3tp1ep-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "leonsarmiento pi MTP; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "leonsarmiento--Qwen3.8-27B-pi-4bit-mtp-mlx on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 204.4 34.6 18.9 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uk3tp1ep",
   "source_url": "https://omlx.ai/benchmarks/performance/uk3tp1ep",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-uk3tp1ep-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "leonsarmiento pi MTP; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 206.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "leonsarmiento--Qwen3.8-27B-pi-4bit-mtp-mlx on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 206.2 38.0 20.2 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uk3tp1ep",
   "source_url": "https://omlx.ai/benchmarks/performance/uk3tp1ep",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-uk3tp1ep-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "leonsarmiento pi MTP; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 203.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "leonsarmiento--Qwen3.8-27B-pi-4bit-mtp-mlx on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 203.5 28.1 20.9 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uk3tp1ep",
   "source_url": "https://omlx.ai/benchmarks/performance/uk3tp1ep",
   "tg_tokens": null,
   "tps": 28.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-zwi7jdgz-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: vywrywjp, yusay2oj; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 218.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 218.2 16.7 19.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zwi7jdgz",
   "source_url": "https://omlx.ai/benchmarks/performance/zwi7jdgz",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-zwi7jdgz-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: vywrywjp, yusay2oj; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 306.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 306.9 18.2 21.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zwi7jdgz",
   "source_url": "https://omlx.ai/benchmarks/performance/zwi7jdgz",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-zwi7jdgz-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: vywrywjp, yusay2oj; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 298.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 8k 298.6 25.6 22.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zwi7jdgz",
   "source_url": "https://omlx.ai/benchmarks/performance/zwi7jdgz",
   "tg_tokens": null,
   "tps": 25.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-zwi7jdgz-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: vywrywjp, yusay2oj; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 275.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 275.4 20.6 23.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zwi7jdgz",
   "source_url": "https://omlx.ai/benchmarks/performance/zwi7jdgz",
   "tg_tokens": null,
   "tps": 20.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-zwi7jdgz-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: vywrywjp, yusay2oj; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 248.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 32k 248.0 6.9 26.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zwi7jdgz",
   "source_url": "https://omlx.ai/benchmarks/performance/zwi7jdgz",
   "tg_tokens": null,
   "tps": 6.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-zwi7jdgz-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 30.5 GB at this context; duplicate traces of the same session: vywrywjp, yusay2oj; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 206.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 64k 206.8 4.3 30.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zwi7jdgz",
   "source_url": "https://omlx.ai/benchmarks/performance/zwi7jdgz",
   "tg_tokens": null,
   "tps": 4.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-sdzba6rj-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 213.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 213.6 17.2 19.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sdzba6rj",
   "source_url": "https://omlx.ai/benchmarks/performance/sdzba6rj",
   "tg_tokens": null,
   "tps": 17.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-sdzba6rj-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 306.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 306.7 18.5 21.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sdzba6rj",
   "source_url": "https://omlx.ai/benchmarks/performance/sdzba6rj",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-lig5toef-ctx1024",
   "model": "MiMo-7B-RL",
   "notes": "oMLX community benchmark trace; peak memory 4.7 GB at this context; duplicate traces of the same session: u88mqp4k, vgbixf8g; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1290.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-7B-RL-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 1,290 121.0 4.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lig5toef",
   "source_url": "https://omlx.ai/benchmarks/performance/lig5toef",
   "tg_tokens": null,
   "tps": 121.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-lig5toef-ctx4096",
   "model": "MiMo-7B-RL",
   "notes": "oMLX community benchmark trace; peak memory 5.2 GB at this context; duplicate traces of the same session: u88mqp4k, vgbixf8g; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1246.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-7B-RL-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 1,246 107.3 5.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lig5toef",
   "source_url": "https://omlx.ai/benchmarks/performance/lig5toef",
   "tg_tokens": null,
   "tps": 107.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-lig5toef-ctx8192",
   "model": "MiMo-7B-RL",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: u88mqp4k, vgbixf8g; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1150.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-7B-RL-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 1,150 93.6 5.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lig5toef",
   "source_url": "https://omlx.ai/benchmarks/performance/lig5toef",
   "tg_tokens": null,
   "tps": 93.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-5m8n590y-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 5lwsrfa7, 6xveliwu, grnsxxa2, o2ks9cjv; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1646.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 1,646 69.5 19.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5m8n590y",
   "source_url": "https://omlx.ai/benchmarks/performance/5m8n590y",
   "tg_tokens": null,
   "tps": 69.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-5m8n590y-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 5lwsrfa7, 6xveliwu, grnsxxa2, o2ks9cjv; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1966.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,966 70.0 19.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5m8n590y",
   "source_url": "https://omlx.ai/benchmarks/performance/5m8n590y",
   "tg_tokens": null,
   "tps": 70.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-5m8n590y-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 5lwsrfa7, 6xveliwu, grnsxxa2, o2ks9cjv; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1937.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 8k 1,937 66.7 19.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5m8n590y",
   "source_url": "https://omlx.ai/benchmarks/performance/5m8n590y",
   "tg_tokens": null,
   "tps": 66.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-5m8n590y-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 5lwsrfa7, 6xveliwu, grnsxxa2, o2ks9cjv; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1802.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 1,802 61.5 20.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5m8n590y",
   "source_url": "https://omlx.ai/benchmarks/performance/5m8n590y",
   "tg_tokens": null,
   "tps": 61.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-5m8n590y-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: 5lwsrfa7, 6xveliwu, grnsxxa2, o2ks9cjv; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1520.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 32k 1,520 51.8 20.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5m8n590y",
   "source_url": "https://omlx.ai/benchmarks/performance/5m8n590y",
   "tg_tokens": null,
   "tps": 51.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-5m8n590y-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: 5lwsrfa7, 6xveliwu, grnsxxa2, o2ks9cjv; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1133.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 64k 1,133 38.0 21.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5m8n590y",
   "source_url": "https://omlx.ai/benchmarks/performance/5m8n590y",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-1uwcfice-ctx1024",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: 1swtucti; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 291.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-27B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 291.2 9.7 15.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1uwcfice",
   "source_url": "https://omlx.ai/benchmarks/performance/1uwcfice",
   "tg_tokens": null,
   "tps": 9.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-1uwcfice-ctx4096",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 1swtucti; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 296.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-27B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 296.1 9.3 17.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1uwcfice",
   "source_url": "https://omlx.ai/benchmarks/performance/1uwcfice",
   "tg_tokens": null,
   "tps": 9.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3ry85sf1-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: giayrly2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3787.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 3,787 118.1 26.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ry85sf1",
   "source_url": "https://omlx.ai/benchmarks/performance/3ry85sf1",
   "tg_tokens": null,
   "tps": 118.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3ry85sf1-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: giayrly2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5216.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 5,216 173.4 26.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ry85sf1",
   "source_url": "https://omlx.ai/benchmarks/performance/3ry85sf1",
   "tg_tokens": null,
   "tps": 173.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3ry85sf1-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: giayrly2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5215.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 5,215 169.6 26.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ry85sf1",
   "source_url": "https://omlx.ai/benchmarks/performance/3ry85sf1",
   "tg_tokens": null,
   "tps": 169.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3ry85sf1-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.0 GB at this context; duplicate traces of the same session: giayrly2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4892.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 4,892 183.3 27.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ry85sf1",
   "source_url": "https://omlx.ai/benchmarks/performance/3ry85sf1",
   "tg_tokens": null,
   "tps": 183.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3ry85sf1-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: giayrly2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4084.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 4,084 239.0 27.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ry85sf1",
   "source_url": "https://omlx.ai/benchmarks/performance/3ry85sf1",
   "tg_tokens": null,
   "tps": 239.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3ry85sf1-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: giayrly2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2991.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 2,991 130.2 28.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ry85sf1",
   "source_url": "https://omlx.ai/benchmarks/performance/3ry85sf1",
   "tg_tokens": null,
   "tps": 130.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3ry85sf1-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: giayrly2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1992.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 1,992 101.7 29.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ry85sf1",
   "source_url": "https://omlx.ai/benchmarks/performance/3ry85sf1",
   "tg_tokens": null,
   "tps": 101.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-q6uyzxbh-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: b75r98kz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 129.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 129.3 15.3 16.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q6uyzxbh",
   "source_url": "https://omlx.ai/benchmarks/performance/q6uyzxbh",
   "tg_tokens": null,
   "tps": 15.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-q6uyzxbh-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: b75r98kz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 130.4 14.9 17.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q6uyzxbh",
   "source_url": "https://omlx.ai/benchmarks/performance/q6uyzxbh",
   "tg_tokens": null,
   "tps": 14.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-zl7vo72a-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 498.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 498.5 17.4 17.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zl7vo72a",
   "source_url": "https://omlx.ai/benchmarks/performance/zl7vo72a",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-zl7vo72a-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 465.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 465.0 16.3 18.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zl7vo72a",
   "source_url": "https://omlx.ai/benchmarks/performance/zl7vo72a",
   "tg_tokens": null,
   "tps": 16.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-f4d96pyl-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 477.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 477.6 22.5 16.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f4d96pyl",
   "source_url": "https://omlx.ai/benchmarks/performance/f4d96pyl",
   "tg_tokens": null,
   "tps": 22.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-f4d96pyl-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.2 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 475.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 475.8 24.2 18.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f4d96pyl",
   "source_url": "https://omlx.ai/benchmarks/performance/f4d96pyl",
   "tg_tokens": null,
   "tps": 24.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-5g8za4wf-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 23.4 GB; TTFT 36689 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 446.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k tokens 446.6 PP tok/s 24.6 TG tok/s 36689 TTFT (ms) 23.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5g8za4wf",
   "source_url": "https://omlx.ai/benchmarks/performance/5g8za4wf",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7qgayjzv-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 34.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 259.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 259.5 29.0 34.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7qgayjzv",
   "source_url": "https://omlx.ai/benchmarks/performance/7qgayjzv",
   "tg_tokens": null,
   "tps": 29.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7qgayjzv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 218.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 218.0 26.8 37.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7qgayjzv",
   "source_url": "https://omlx.ai/benchmarks/performance/7qgayjzv",
   "tg_tokens": null,
   "tps": 26.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7qgayjzv-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 38.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 216.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 216.1 24.3 38.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7qgayjzv",
   "source_url": "https://omlx.ai/benchmarks/performance/7qgayjzv",
   "tg_tokens": null,
   "tps": 24.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-jbd4ntm0-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: gl6rj7dv, n716phs3, qklxi5h3, x5fw8qpx; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 277.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 277.0 21.9 17.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jbd4ntm0",
   "source_url": "https://omlx.ai/benchmarks/performance/jbd4ntm0",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-jbd4ntm0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: gl6rj7dv, n716phs3, qklxi5h3, x5fw8qpx; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 233.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 233.1 20.5 18.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jbd4ntm0",
   "source_url": "https://omlx.ai/benchmarks/performance/jbd4ntm0",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-jbd4ntm0-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: gl6rj7dv, n716phs3, qklxi5h3, x5fw8qpx; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 235.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 8k 235.4 26.1 18.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jbd4ntm0",
   "source_url": "https://omlx.ai/benchmarks/performance/jbd4ntm0",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-jbd4ntm0-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: gl6rj7dv, n716phs3, qklxi5h3, x5fw8qpx; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 222.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 222.4 23.0 19.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jbd4ntm0",
   "source_url": "https://omlx.ai/benchmarks/performance/jbd4ntm0",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-jbd4ntm0-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: gl6rj7dv, n716phs3, qklxi5h3, x5fw8qpx; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 214.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 32k 214.9 25.3 20.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jbd4ntm0",
   "source_url": "https://omlx.ai/benchmarks/performance/jbd4ntm0",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-jbd4ntm0-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: gl6rj7dv, n716phs3, qklxi5h3, x5fw8qpx; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 193.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 64k 193.3 17.9 23.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jbd4ntm0",
   "source_url": "https://omlx.ai/benchmarks/performance/jbd4ntm0",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8fze8ktg-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: ytj1j6kv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2453.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,453 84.2 71.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8fze8ktg",
   "source_url": "https://omlx.ai/benchmarks/performance/8fze8ktg",
   "tg_tokens": null,
   "tps": 84.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8fze8ktg-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.7 GB at this context; duplicate traces of the same session: ytj1j6kv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2586.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,586 58.2 72.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8fze8ktg",
   "source_url": "https://omlx.ai/benchmarks/performance/8fze8ktg",
   "tg_tokens": null,
   "tps": 58.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8fze8ktg-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.0 GB at this context; duplicate traces of the same session: ytj1j6kv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2422.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,422 72.3 79.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8fze8ktg",
   "source_url": "https://omlx.ai/benchmarks/performance/8fze8ktg",
   "tg_tokens": null,
   "tps": 72.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-h0936z6p-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: 4riwznwc, aarafb5b, mfqdarc4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1128.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,128 75.0 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h0936z6p",
   "source_url": "https://omlx.ai/benchmarks/performance/h0936z6p",
   "tg_tokens": null,
   "tps": 75.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-h0936z6p-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: 4riwznwc, aarafb5b, mfqdarc4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1763.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,763 71.4 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h0936z6p",
   "source_url": "https://omlx.ai/benchmarks/performance/h0936z6p",
   "tg_tokens": null,
   "tps": 71.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-h0936z6p-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 4riwznwc, aarafb5b, mfqdarc4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2155.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,155 88.9 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h0936z6p",
   "source_url": "https://omlx.ai/benchmarks/performance/h0936z6p",
   "tg_tokens": null,
   "tps": 88.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-h0936z6p-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.7 GB at this context; duplicate traces of the same session: 4riwznwc, aarafb5b, mfqdarc4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2327.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,327 78.7 72.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h0936z6p",
   "source_url": "https://omlx.ai/benchmarks/performance/h0936z6p",
   "tg_tokens": null,
   "tps": 78.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-h0936z6p-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.5 GB at this context; duplicate traces of the same session: 4riwznwc, aarafb5b, mfqdarc4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2213.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,213 63.2 80.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h0936z6p",
   "source_url": "https://omlx.ai/benchmarks/performance/h0936z6p",
   "tg_tokens": null,
   "tps": 63.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-j554j5bh-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: 4nrfgsd1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 708.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 708.2 60.3 70.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j554j5bh",
   "source_url": "https://omlx.ai/benchmarks/performance/j554j5bh",
   "tg_tokens": null,
   "tps": 60.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-j554j5bh-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: 4nrfgsd1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1988.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,988 56.5 72.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j554j5bh",
   "source_url": "https://omlx.ai/benchmarks/performance/j554j5bh",
   "tg_tokens": null,
   "tps": 56.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-snwwdnpf-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.7 GB at this context; duplicate traces of the same session: 5muyj0sp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2416.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,416 65.8 78.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/snwwdnpf",
   "source_url": "https://omlx.ai/benchmarks/performance/snwwdnpf",
   "tg_tokens": null,
   "tps": 65.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-snwwdnpf-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.8 GB at this context; duplicate traces of the same session: 5muyj0sp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2292.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,292 53.5 80.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/snwwdnpf",
   "source_url": "https://omlx.ai/benchmarks/performance/snwwdnpf",
   "tg_tokens": null,
   "tps": 53.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6sxt8nhy-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: fd13w0ls, m4jbgzax, qtatf586; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1161.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,161 62.4 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6sxt8nhy",
   "source_url": "https://omlx.ai/benchmarks/performance/6sxt8nhy",
   "tg_tokens": null,
   "tps": 62.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6sxt8nhy-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: fd13w0ls, m4jbgzax, qtatf586; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2027.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,027 71.7 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6sxt8nhy",
   "source_url": "https://omlx.ai/benchmarks/performance/6sxt8nhy",
   "tg_tokens": null,
   "tps": 71.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-6sxt8nhy-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: fd13w0ls, m4jbgzax, qtatf586; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2465.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,465 82.0 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6sxt8nhy",
   "source_url": "https://omlx.ai/benchmarks/performance/6sxt8nhy",
   "tg_tokens": null,
   "tps": 82.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6sxt8nhy-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.5 GB at this context; duplicate traces of the same session: fd13w0ls, m4jbgzax, qtatf586; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2243.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,243 62.7 80.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6sxt8nhy",
   "source_url": "https://omlx.ai/benchmarks/performance/6sxt8nhy",
   "tg_tokens": null,
   "tps": 62.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wq1rwsr4-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1124.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,124 75.3 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wq1rwsr4",
   "source_url": "https://omlx.ai/benchmarks/performance/wq1rwsr4",
   "tg_tokens": null,
   "tps": 75.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wq1rwsr4-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1336.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,336 64.6 19.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wq1rwsr4",
   "source_url": "https://omlx.ai/benchmarks/performance/wq1rwsr4",
   "tg_tokens": null,
   "tps": 64.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wq1rwsr4-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1349.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,349 69.9 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wq1rwsr4",
   "source_url": "https://omlx.ai/benchmarks/performance/wq1rwsr4",
   "tg_tokens": null,
   "tps": 69.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ersxe269-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1126.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,126 62.4 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ersxe269",
   "source_url": "https://omlx.ai/benchmarks/performance/ersxe269",
   "tg_tokens": null,
   "tps": 62.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ersxe269-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1334.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,334 66.1 19.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ersxe269",
   "source_url": "https://omlx.ai/benchmarks/performance/ersxe269",
   "tg_tokens": null,
   "tps": 66.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ersxe269-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1337.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,337 65.1 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ersxe269",
   "source_url": "https://omlx.ai/benchmarks/performance/ersxe269",
   "tg_tokens": null,
   "tps": 65.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-na55li7o-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.5 GB at this context; duplicate traces of the same session: 0ltnb0m3, 6ba4comj, 83u5c4mm, wg1z3mi4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 245.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 245.7 41.2 32.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/na55li7o",
   "source_url": "https://omlx.ai/benchmarks/performance/na55li7o",
   "tg_tokens": null,
   "tps": 41.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-na55li7o-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.7 GB at this context; duplicate traces of the same session: 0ltnb0m3, 6ba4comj, 83u5c4mm, wg1z3mi4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 274.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 274.1 41.6 35.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/na55li7o",
   "source_url": "https://omlx.ai/benchmarks/performance/na55li7o",
   "tg_tokens": null,
   "tps": 41.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-na55li7o-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: 0ltnb0m3, 6ba4comj, 83u5c4mm, wg1z3mi4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 277.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 8k 277.3 37.8 35.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/na55li7o",
   "source_url": "https://omlx.ai/benchmarks/performance/na55li7o",
   "tg_tokens": null,
   "tps": 37.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-na55li7o-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.5 GB at this context; duplicate traces of the same session: 0ltnb0m3, 6ba4comj, 83u5c4mm, wg1z3mi4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 251.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k 251.3 37.0 31.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/na55li7o",
   "source_url": "https://omlx.ai/benchmarks/performance/na55li7o",
   "tg_tokens": null,
   "tps": 37.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-na55li7o-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.3 GB at this context; duplicate traces of the same session: 0ltnb0m3, 6ba4comj, 83u5c4mm, wg1z3mi4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 233.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 233.1 29.8 33.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/na55li7o",
   "source_url": "https://omlx.ai/benchmarks/performance/na55li7o",
   "tg_tokens": null,
   "tps": 29.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-c250f0py-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "chriswessels ThinkingCap; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: jjdj9i8o, v5cutx1e; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 203.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "chriswessels--ThinkingCap-Qwen3.8-27B-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 203.7 36.5 18.6 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c250f0py",
   "source_url": "https://omlx.ai/benchmarks/performance/c250f0py",
   "tg_tokens": null,
   "tps": 36.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-c250f0py-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "chriswessels ThinkingCap; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: jjdj9i8o, v5cutx1e; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "chriswessels--ThinkingCap-Qwen3.8-27B-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 205.2 30.4 19.8 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c250f0py",
   "source_url": "https://omlx.ai/benchmarks/performance/c250f0py",
   "tg_tokens": null,
   "tps": 30.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-c250f0py-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "chriswessels ThinkingCap; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: jjdj9i8o, v5cutx1e; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 202.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "chriswessels--ThinkingCap-Qwen3.8-27B-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 202.7 34.3 20.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c250f0py",
   "source_url": "https://omlx.ai/benchmarks/performance/c250f0py",
   "tg_tokens": null,
   "tps": 34.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-c250f0py-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "chriswessels ThinkingCap; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: jjdj9i8o, v5cutx1e; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 194.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "chriswessels--ThinkingCap-Qwen3.8-27B-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 194.3 29.0 20.4 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c250f0py",
   "source_url": "https://omlx.ai/benchmarks/performance/c250f0py",
   "tg_tokens": null,
   "tps": 29.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-lyrm0gwj-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.0 GB at this context; duplicate traces of the same session: n0nbdhln; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 396.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 396.7 38.4 14.0 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lyrm0gwj",
   "source_url": "https://omlx.ai/benchmarks/performance/lyrm0gwj",
   "tg_tokens": null,
   "tps": 38.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-lyrm0gwj-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: n0nbdhln; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 392.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 392.5 37.6 14.3 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lyrm0gwj",
   "source_url": "https://omlx.ai/benchmarks/performance/lyrm0gwj",
   "tg_tokens": null,
   "tps": 37.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-vjbasx6l-ctx16384",
   "model": "gemma-4-31B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 175.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-qat-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 175.0 12.7 34.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vjbasx6l",
   "source_url": "https://omlx.ai/benchmarks/performance/vjbasx6l",
   "tg_tokens": null,
   "tps": 12.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-vjbasx6l-ctx32768",
   "model": "gemma-4-31B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 37.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 168.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-qat-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 168.9 11.9 37.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vjbasx6l",
   "source_url": "https://omlx.ai/benchmarks/performance/vjbasx6l",
   "tg_tokens": null,
   "tps": 11.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-z6nc7s0o-ctx16384",
   "model": "gemma-4-31B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 175.8,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "gemma-4-31B-it-qat-oQ8 on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 175.8 12.7 34.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z6nc7s0o",
   "source_url": "https://omlx.ai/benchmarks/performance/z6nc7s0o",
   "tg_tokens": null,
   "tps": 12.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-z6nc7s0o-ctx32768",
   "model": "gemma-4-31B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 37.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 169.7,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "gemma-4-31B-it-qat-oQ8 on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 169.7 12.0 37.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z6nc7s0o",
   "source_url": "https://omlx.ai/benchmarks/performance/z6nc7s0o",
   "tg_tokens": null,
   "tps": 12.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-yusay2oj-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: vywrywjp; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 218.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 218.2 16.7 19.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yusay2oj",
   "source_url": "https://omlx.ai/benchmarks/performance/yusay2oj",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-yusay2oj-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: vywrywjp; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 306.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 306.9 18.2 21.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yusay2oj",
   "source_url": "https://omlx.ai/benchmarks/performance/yusay2oj",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-yusay2oj-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: vywrywjp; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 298.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 8k 298.6 25.6 22.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yusay2oj",
   "source_url": "https://omlx.ai/benchmarks/performance/yusay2oj",
   "tg_tokens": null,
   "tps": 25.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-yusay2oj-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: vywrywjp; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 275.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 275.4 20.6 23.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yusay2oj",
   "source_url": "https://omlx.ai/benchmarks/performance/yusay2oj",
   "tg_tokens": null,
   "tps": 20.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-yusay2oj-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: vywrywjp; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 248.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 32k 248.0 6.9 26.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yusay2oj",
   "source_url": "https://omlx.ai/benchmarks/performance/yusay2oj",
   "tg_tokens": null,
   "tps": 6.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-yusay2oj-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 30.5 GB at this context; duplicate traces of the same session: vywrywjp; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 206.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 64k 206.8 4.3 30.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yusay2oj",
   "source_url": "https://omlx.ai/benchmarks/performance/yusay2oj",
   "tg_tokens": null,
   "tps": 4.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-u88mqp4k-ctx1024",
   "model": "MiMo-7B-RL",
   "notes": "oMLX community benchmark trace; peak memory 4.7 GB at this context; duplicate traces of the same session: vgbixf8g; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1290.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-7B-RL-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 1,290 121.0 4.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u88mqp4k",
   "source_url": "https://omlx.ai/benchmarks/performance/u88mqp4k",
   "tg_tokens": null,
   "tps": 121.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-u88mqp4k-ctx4096",
   "model": "MiMo-7B-RL",
   "notes": "oMLX community benchmark trace; peak memory 5.2 GB at this context; duplicate traces of the same session: vgbixf8g; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1246.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-7B-RL-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 1,246 107.3 5.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u88mqp4k",
   "source_url": "https://omlx.ai/benchmarks/performance/u88mqp4k",
   "tg_tokens": null,
   "tps": 107.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-u88mqp4k-ctx8192",
   "model": "MiMo-7B-RL",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: vgbixf8g; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1150.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-7B-RL-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 1,150 93.6 5.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u88mqp4k",
   "source_url": "https://omlx.ai/benchmarks/performance/u88mqp4k",
   "tg_tokens": null,
   "tps": 93.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-o2ks9cjv-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 5lwsrfa7, 6xveliwu, grnsxxa2; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1646.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 1,646 69.5 19.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o2ks9cjv",
   "source_url": "https://omlx.ai/benchmarks/performance/o2ks9cjv",
   "tg_tokens": null,
   "tps": 69.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-o2ks9cjv-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 5lwsrfa7, 6xveliwu, grnsxxa2; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1966.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,966 70.0 19.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o2ks9cjv",
   "source_url": "https://omlx.ai/benchmarks/performance/o2ks9cjv",
   "tg_tokens": null,
   "tps": 70.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-o2ks9cjv-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 5lwsrfa7, 6xveliwu, grnsxxa2; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1937.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 8k 1,937 66.7 19.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o2ks9cjv",
   "source_url": "https://omlx.ai/benchmarks/performance/o2ks9cjv",
   "tg_tokens": null,
   "tps": 66.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-o2ks9cjv-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 5lwsrfa7, 6xveliwu, grnsxxa2; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1802.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 1,802 61.5 20.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o2ks9cjv",
   "source_url": "https://omlx.ai/benchmarks/performance/o2ks9cjv",
   "tg_tokens": null,
   "tps": 61.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-o2ks9cjv-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: 5lwsrfa7, 6xveliwu, grnsxxa2; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1520.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 32k 1,520 51.8 20.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o2ks9cjv",
   "source_url": "https://omlx.ai/benchmarks/performance/o2ks9cjv",
   "tg_tokens": null,
   "tps": 51.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-o2ks9cjv-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: 5lwsrfa7, 6xveliwu, grnsxxa2; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1133.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 64k 1,133 38.0 21.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o2ks9cjv",
   "source_url": "https://omlx.ai/benchmarks/performance/o2ks9cjv",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-1swtucti-ctx1024",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 291.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-27B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 291.2 9.7 15.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1swtucti",
   "source_url": "https://omlx.ai/benchmarks/performance/1swtucti",
   "tg_tokens": null,
   "tps": 9.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-1swtucti-ctx4096",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 296.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-27B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 296.1 9.3 17.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1swtucti",
   "source_url": "https://omlx.ai/benchmarks/performance/1swtucti",
   "tg_tokens": null,
   "tps": 9.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-giayrly2-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: dssdp0o7, eucfvguy, igbn28oo, l9le3ol1, lr1gbf61; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3787.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 3,787 118.1 26.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/giayrly2",
   "source_url": "https://omlx.ai/benchmarks/performance/giayrly2",
   "tg_tokens": null,
   "tps": 118.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-giayrly2-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: dssdp0o7, eucfvguy, igbn28oo, l9le3ol1, lr1gbf61; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5216.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 5,216 173.4 26.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/giayrly2",
   "source_url": "https://omlx.ai/benchmarks/performance/giayrly2",
   "tg_tokens": null,
   "tps": 173.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-giayrly2-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: dssdp0o7, eucfvguy, igbn28oo, l9le3ol1, lr1gbf61; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5215.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 5,215 169.6 26.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/giayrly2",
   "source_url": "https://omlx.ai/benchmarks/performance/giayrly2",
   "tg_tokens": null,
   "tps": 169.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-giayrly2-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.0 GB at this context; duplicate traces of the same session: dssdp0o7, eucfvguy, igbn28oo, l9le3ol1, lr1gbf61; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4892.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 4,892 183.3 27.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/giayrly2",
   "source_url": "https://omlx.ai/benchmarks/performance/giayrly2",
   "tg_tokens": null,
   "tps": 183.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-giayrly2-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: dssdp0o7, eucfvguy, igbn28oo, l9le3ol1, lr1gbf61; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4084.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 4,084 239.0 27.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/giayrly2",
   "source_url": "https://omlx.ai/benchmarks/performance/giayrly2",
   "tg_tokens": null,
   "tps": 239.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-giayrly2-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: dssdp0o7, eucfvguy, igbn28oo, l9le3ol1, lr1gbf61; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2991.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 2,991 130.2 28.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/giayrly2",
   "source_url": "https://omlx.ai/benchmarks/performance/giayrly2",
   "tg_tokens": null,
   "tps": 130.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-giayrly2-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: dssdp0o7, eucfvguy, igbn28oo, l9le3ol1, lr1gbf61; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1992.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 1,992 101.7 29.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/giayrly2",
   "source_url": "https://omlx.ai/benchmarks/performance/giayrly2",
   "tg_tokens": null,
   "tps": 101.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-ub0g5x72-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: f9h20840; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 768.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 1k 768.4 56.4 37.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ub0g5x72",
   "source_url": "https://omlx.ai/benchmarks/performance/ub0g5x72",
   "tg_tokens": null,
   "tps": 56.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ub0g5x72-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 39.5 GB at this context; duplicate traces of the same session: f9h20840; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 609.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 32k 609.9 47.6 39.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ub0g5x72",
   "source_url": "https://omlx.ai/benchmarks/performance/ub0g5x72",
   "tg_tokens": null,
   "tps": 47.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-b75r98kz-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 129.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 129.3 15.3 16.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b75r98kz",
   "source_url": "https://omlx.ai/benchmarks/performance/b75r98kz",
   "tg_tokens": null,
   "tps": 15.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-b75r98kz-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 130.4 14.9 17.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b75r98kz",
   "source_url": "https://omlx.ai/benchmarks/performance/b75r98kz",
   "tg_tokens": null,
   "tps": 14.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-zepplcxy-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "unsloth build; oMLX community benchmark trace, single run at 4k context; peak memory 27.4 GB; TTFT 32501 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 126.0,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-27B-NVFP4-unsloth on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 nvfp4; 4k tokens 126.0 PP tok/s 18.0 TG tok/s 32501 TTFT (ms) 27.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zepplcxy",
   "source_url": "https://omlx.ai/benchmarks/performance/zepplcxy",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-qklxi5h3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: gl6rj7dv, n716phs3, x5fw8qpx; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 277.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 277.0 21.9 17.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qklxi5h3",
   "source_url": "https://omlx.ai/benchmarks/performance/qklxi5h3",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-qklxi5h3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: gl6rj7dv, n716phs3, x5fw8qpx; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 233.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 233.1 20.5 18.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qklxi5h3",
   "source_url": "https://omlx.ai/benchmarks/performance/qklxi5h3",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-qklxi5h3-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: gl6rj7dv, n716phs3, x5fw8qpx; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 235.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 8k 235.4 26.1 18.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qklxi5h3",
   "source_url": "https://omlx.ai/benchmarks/performance/qklxi5h3",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-qklxi5h3-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: gl6rj7dv, n716phs3, x5fw8qpx; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 222.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 222.4 23.0 19.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qklxi5h3",
   "source_url": "https://omlx.ai/benchmarks/performance/qklxi5h3",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-qklxi5h3-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: gl6rj7dv, n716phs3, x5fw8qpx; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 214.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 32k 214.9 25.3 20.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qklxi5h3",
   "source_url": "https://omlx.ai/benchmarks/performance/qklxi5h3",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-qklxi5h3-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: gl6rj7dv, n716phs3, x5fw8qpx; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 193.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 64k 193.3 17.9 23.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qklxi5h3",
   "source_url": "https://omlx.ai/benchmarks/performance/qklxi5h3",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-piuw4270-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: 0kaewohk; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 119.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 1k 119.4 11.1 30.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/piuw4270",
   "source_url": "https://omlx.ai/benchmarks/performance/piuw4270",
   "tg_tokens": null,
   "tps": 11.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-piuw4270-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 35.3 GB at this context; duplicate traces of the same session: 0kaewohk; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 32k 104.8 9.5 35.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/piuw4270",
   "source_url": "https://omlx.ai/benchmarks/performance/piuw4270",
   "tg_tokens": null,
   "tps": 9.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-6t6inme7-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 122.5 GB at this context; duplicate traces of the same session: oby062k4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 852.5,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 1k 852.5 89.2 122.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6t6inme7",
   "source_url": "https://omlx.ai/benchmarks/performance/6t6inme7",
   "tg_tokens": null,
   "tps": 89.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-6t6inme7-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 122.8 GB at this context; duplicate traces of the same session: oby062k4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1043.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 4k 1,043 82.3 122.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6t6inme7",
   "source_url": "https://omlx.ai/benchmarks/performance/6t6inme7",
   "tg_tokens": null,
   "tps": 82.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-avb7who1-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace, single run at 1k context; peak memory 122.4 GB; TTFT 1205 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 849.8,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 1k tokens 849.8 PP tok/s 72.1 TG tok/s 1205 TTFT (ms) 122.4 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/avb7who1",
   "source_url": "https://omlx.ai/benchmarks/performance/avb7who1",
   "tg_tokens": null,
   "tps": 72.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ytj1j6kv-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2453.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,453 84.2 71.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ytj1j6kv",
   "source_url": "https://omlx.ai/benchmarks/performance/ytj1j6kv",
   "tg_tokens": null,
   "tps": 84.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ytj1j6kv-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2586.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,586 58.2 72.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ytj1j6kv",
   "source_url": "https://omlx.ai/benchmarks/performance/ytj1j6kv",
   "tg_tokens": null,
   "tps": 58.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ytj1j6kv-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2422.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,422 72.3 79.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ytj1j6kv",
   "source_url": "https://omlx.ai/benchmarks/performance/ytj1j6kv",
   "tg_tokens": null,
   "tps": 72.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-aarafb5b-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: 4riwznwc, mfqdarc4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1128.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,128 75.0 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aarafb5b",
   "source_url": "https://omlx.ai/benchmarks/performance/aarafb5b",
   "tg_tokens": null,
   "tps": 75.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-aarafb5b-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: 4riwznwc, mfqdarc4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1763.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,763 71.4 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aarafb5b",
   "source_url": "https://omlx.ai/benchmarks/performance/aarafb5b",
   "tg_tokens": null,
   "tps": 71.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-aarafb5b-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 4riwznwc, mfqdarc4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2155.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,155 88.9 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aarafb5b",
   "source_url": "https://omlx.ai/benchmarks/performance/aarafb5b",
   "tg_tokens": null,
   "tps": 88.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-aarafb5b-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.7 GB at this context; duplicate traces of the same session: 4riwznwc, mfqdarc4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2327.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,327 78.7 72.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aarafb5b",
   "source_url": "https://omlx.ai/benchmarks/performance/aarafb5b",
   "tg_tokens": null,
   "tps": 78.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-aarafb5b-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.5 GB at this context; duplicate traces of the same session: 4riwznwc, mfqdarc4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2213.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,213 63.2 80.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aarafb5b",
   "source_url": "https://omlx.ai/benchmarks/performance/aarafb5b",
   "tg_tokens": null,
   "tps": 63.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4nrfgsd1-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 708.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 708.2 60.3 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4nrfgsd1",
   "source_url": "https://omlx.ai/benchmarks/performance/4nrfgsd1",
   "tg_tokens": null,
   "tps": 60.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4nrfgsd1-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1988.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,988 56.5 72.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4nrfgsd1",
   "source_url": "https://omlx.ai/benchmarks/performance/4nrfgsd1",
   "tg_tokens": null,
   "tps": 56.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qtatf586-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: fd13w0ls, m4jbgzax; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1161.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,161 62.4 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qtatf586",
   "source_url": "https://omlx.ai/benchmarks/performance/qtatf586",
   "tg_tokens": null,
   "tps": 62.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qtatf586-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: fd13w0ls, m4jbgzax; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2027.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,027 71.7 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qtatf586",
   "source_url": "https://omlx.ai/benchmarks/performance/qtatf586",
   "tg_tokens": null,
   "tps": 71.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qtatf586-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: fd13w0ls, m4jbgzax; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2465.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,465 82.0 71.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qtatf586",
   "source_url": "https://omlx.ai/benchmarks/performance/qtatf586",
   "tg_tokens": null,
   "tps": 82.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qtatf586-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.5 GB at this context; duplicate traces of the same session: fd13w0ls, m4jbgzax; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2243.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,243 62.7 80.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qtatf586",
   "source_url": "https://omlx.ai/benchmarks/performance/qtatf586",
   "tg_tokens": null,
   "tps": 62.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5muyj0sp-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2416.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,416 65.8 78.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5muyj0sp",
   "source_url": "https://omlx.ai/benchmarks/performance/5muyj0sp",
   "tg_tokens": null,
   "tps": 65.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5muyj0sp-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2292.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,292 53.5 80.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5muyj0sp",
   "source_url": "https://omlx.ai/benchmarks/performance/5muyj0sp",
   "tg_tokens": null,
   "tps": 53.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hggqq9p9-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 141.9 GB at this context; duplicate traces of the same session: hzlneami, svxr6uce; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1047.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 64k 1,047 56.2 141.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hggqq9p9",
   "source_url": "https://omlx.ai/benchmarks/performance/hggqq9p9",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-hggqq9p9-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.1 GB at this context; duplicate traces of the same session: hzlneami, svxr6uce; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1030.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 128k 1,030 51.8 145.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hggqq9p9",
   "source_url": "https://omlx.ai/benchmarks/performance/hggqq9p9",
   "tg_tokens": null,
   "tps": 51.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-hggqq9p9-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 149.0 GB at this context; duplicate traces of the same session: hzlneami, svxr6uce; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1018.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 195k 1,018 50.1 149.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hggqq9p9",
   "source_url": "https://omlx.ai/benchmarks/performance/hggqq9p9",
   "tg_tokens": null,
   "tps": 50.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-22y02ucf-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: c4um5x2a, yke0u1ug; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1066.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,066 58.7 19.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/22y02ucf",
   "source_url": "https://omlx.ai/benchmarks/performance/22y02ucf",
   "tg_tokens": null,
   "tps": 58.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-22y02ucf-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: c4um5x2a, yke0u1ug; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1313.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,313 63.0 20.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/22y02ucf",
   "source_url": "https://omlx.ai/benchmarks/performance/22y02ucf",
   "tg_tokens": null,
   "tps": 63.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-22y02ucf-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: c4um5x2a, yke0u1ug; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1334.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,334 57.5 21.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/22y02ucf",
   "source_url": "https://omlx.ai/benchmarks/performance/22y02ucf",
   "tg_tokens": null,
   "tps": 57.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-r83gy705-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 57insz4w, psztr93l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1118.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,118 58.2 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r83gy705",
   "source_url": "https://omlx.ai/benchmarks/performance/r83gy705",
   "tg_tokens": null,
   "tps": 58.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-r83gy705-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 57insz4w, psztr93l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1331.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,331 64.4 19.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r83gy705",
   "source_url": "https://omlx.ai/benchmarks/performance/r83gy705",
   "tg_tokens": null,
   "tps": 64.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-r83gy705-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 57insz4w, psztr93l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1321.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,321 69.8 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r83gy705",
   "source_url": "https://omlx.ai/benchmarks/performance/r83gy705",
   "tg_tokens": null,
   "tps": 69.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-6ba4comj-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.5 GB at this context; duplicate traces of the same session: 0ltnb0m3, 83u5c4mm, wg1z3mi4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 245.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 245.7 41.2 32.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ba4comj",
   "source_url": "https://omlx.ai/benchmarks/performance/6ba4comj",
   "tg_tokens": null,
   "tps": 41.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-6ba4comj-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.7 GB at this context; duplicate traces of the same session: 0ltnb0m3, 83u5c4mm, wg1z3mi4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 274.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 274.1 41.6 35.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ba4comj",
   "source_url": "https://omlx.ai/benchmarks/performance/6ba4comj",
   "tg_tokens": null,
   "tps": 41.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-6ba4comj-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: 0ltnb0m3, 83u5c4mm, wg1z3mi4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 277.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 8k 277.3 37.8 35.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ba4comj",
   "source_url": "https://omlx.ai/benchmarks/performance/6ba4comj",
   "tg_tokens": null,
   "tps": 37.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-6ba4comj-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.5 GB at this context; duplicate traces of the same session: 0ltnb0m3, 83u5c4mm, wg1z3mi4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 251.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k 251.3 37.0 31.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ba4comj",
   "source_url": "https://omlx.ai/benchmarks/performance/6ba4comj",
   "tg_tokens": null,
   "tps": 37.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-6ba4comj-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.3 GB at this context; duplicate traces of the same session: 0ltnb0m3, 83u5c4mm, wg1z3mi4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 233.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 233.1 29.8 33.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ba4comj",
   "source_url": "https://omlx.ai/benchmarks/performance/6ba4comj",
   "tg_tokens": null,
   "tps": 29.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-jjdj9i8o-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "chriswessels ThinkingCap; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: v5cutx1e; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 203.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "chriswessels--ThinkingCap-Qwen3.8-27B-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 203.7 36.5 18.6 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jjdj9i8o",
   "source_url": "https://omlx.ai/benchmarks/performance/jjdj9i8o",
   "tg_tokens": null,
   "tps": 36.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-jjdj9i8o-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "chriswessels ThinkingCap; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: v5cutx1e; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "chriswessels--ThinkingCap-Qwen3.8-27B-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 205.2 30.4 19.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jjdj9i8o",
   "source_url": "https://omlx.ai/benchmarks/performance/jjdj9i8o",
   "tg_tokens": null,
   "tps": 30.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-jjdj9i8o-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "chriswessels ThinkingCap; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: v5cutx1e; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 202.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "chriswessels--ThinkingCap-Qwen3.8-27B-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 202.7 34.3 20.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jjdj9i8o",
   "source_url": "https://omlx.ai/benchmarks/performance/jjdj9i8o",
   "tg_tokens": null,
   "tps": 34.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-jjdj9i8o-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "chriswessels ThinkingCap; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: v5cutx1e; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 194.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "chriswessels--ThinkingCap-Qwen3.8-27B-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 194.3 29.0 20.4 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jjdj9i8o",
   "source_url": "https://omlx.ai/benchmarks/performance/jjdj9i8o",
   "tg_tokens": null,
   "tps": 29.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-0siag7ec-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: 7jbivu99, 8ezii5v6, e76si6sh, fzoc69ia, wrlvh6ig; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 676.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 1k 676.0 70.3 14.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0siag7ec",
   "source_url": "https://omlx.ai/benchmarks/performance/0siag7ec",
   "tg_tokens": null,
   "tps": 70.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-0siag7ec-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: 7jbivu99, 8ezii5v6, e76si6sh, fzoc69ia, wrlvh6ig; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 672.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 4k 672.4 67.1 14.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0siag7ec",
   "source_url": "https://omlx.ai/benchmarks/performance/0siag7ec",
   "tg_tokens": null,
   "tps": 67.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-0siag7ec-ctx8192",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: 7jbivu99, 8ezii5v6, e76si6sh, fzoc69ia, wrlvh6ig; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 662.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 8k 662.4 63.1 14.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0siag7ec",
   "source_url": "https://omlx.ai/benchmarks/performance/0siag7ec",
   "tg_tokens": null,
   "tps": 63.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-0siag7ec-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 15.2 GB at this context; duplicate traces of the same session: 7jbivu99, 8ezii5v6, e76si6sh, fzoc69ia, wrlvh6ig; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 632.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 16k 632.0 56.2 15.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0siag7ec",
   "source_url": "https://omlx.ai/benchmarks/performance/0siag7ec",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-0siag7ec-ctx32768",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 7jbivu99, 8ezii5v6, e76si6sh, fzoc69ia, wrlvh6ig; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 578.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 32k 578.4 46.9 16.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0siag7ec",
   "source_url": "https://omlx.ai/benchmarks/performance/0siag7ec",
   "tg_tokens": null,
   "tps": 46.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-0siag7ec-ctx65536",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: 7jbivu99, 8ezii5v6, e76si6sh, fzoc69ia, wrlvh6ig; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 493.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 64k 493.5 34.6 17.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0siag7ec",
   "source_url": "https://omlx.ai/benchmarks/performance/0siag7ec",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-n0nbdhln-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.0 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 396.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 396.7 38.4 14.0 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n0nbdhln",
   "source_url": "https://omlx.ai/benchmarks/performance/n0nbdhln",
   "tg_tokens": null,
   "tps": 38.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-n0nbdhln-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 392.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 392.5 37.6 14.3 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n0nbdhln",
   "source_url": "https://omlx.ai/benchmarks/performance/n0nbdhln",
   "tg_tokens": null,
   "tps": 37.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-v220hgma-ctx1024",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: ay300quz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 208.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 208.4 20.5 18.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v220hgma",
   "source_url": "https://omlx.ai/benchmarks/performance/v220hgma",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-v220hgma-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: ay300quz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 205.8 23.0 19.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v220hgma",
   "source_url": "https://omlx.ai/benchmarks/performance/v220hgma",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-v220hgma-ctx8192",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: ay300quz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 197.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 197.3 22.4 19.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v220hgma",
   "source_url": "https://omlx.ai/benchmarks/performance/v220hgma",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-v220hgma-ctx16384",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: ay300quz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 193.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 193.0 21.5 21.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v220hgma",
   "source_url": "https://omlx.ai/benchmarks/performance/v220hgma",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-v220hgma-ctx32768",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: ay300quz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 178.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 178.0 19.8 23.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v220hgma",
   "source_url": "https://omlx.ai/benchmarks/performance/v220hgma",
   "tg_tokens": null,
   "tps": 19.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-v220hgma-ctx65536",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: ay300quz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 155.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 155.1 17.0 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v220hgma",
   "source_url": "https://omlx.ai/benchmarks/performance/v220hgma",
   "tg_tokens": null,
   "tps": 17.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-vywrywjp-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 218.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 218.2 16.7 19.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vywrywjp",
   "source_url": "https://omlx.ai/benchmarks/performance/vywrywjp",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-vywrywjp-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 306.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 306.9 18.2 21.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vywrywjp",
   "source_url": "https://omlx.ai/benchmarks/performance/vywrywjp",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-vywrywjp-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 298.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 8k 298.6 25.6 22.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vywrywjp",
   "source_url": "https://omlx.ai/benchmarks/performance/vywrywjp",
   "tg_tokens": null,
   "tps": 25.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-vywrywjp-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 275.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 275.4 20.6 23.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vywrywjp",
   "source_url": "https://omlx.ai/benchmarks/performance/vywrywjp",
   "tg_tokens": null,
   "tps": 20.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-vywrywjp-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 248.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 32k 248.0 6.9 26.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vywrywjp",
   "source_url": "https://omlx.ai/benchmarks/performance/vywrywjp",
   "tg_tokens": null,
   "tps": 6.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-vywrywjp-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 30.5 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 206.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 64k 206.8 4.3 30.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vywrywjp",
   "source_url": "https://omlx.ai/benchmarks/performance/vywrywjp",
   "tg_tokens": null,
   "tps": 4.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vgbixf8g-ctx1024",
   "model": "MiMo-7B-RL",
   "notes": "oMLX community benchmark trace; peak memory 4.7 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1290.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-7B-RL-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 1,290 121.0 4.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vgbixf8g",
   "source_url": "https://omlx.ai/benchmarks/performance/vgbixf8g",
   "tg_tokens": null,
   "tps": 121.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vgbixf8g-ctx4096",
   "model": "MiMo-7B-RL",
   "notes": "oMLX community benchmark trace; peak memory 5.2 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1246.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-7B-RL-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 1,246 107.3 5.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vgbixf8g",
   "source_url": "https://omlx.ai/benchmarks/performance/vgbixf8g",
   "tg_tokens": null,
   "tps": 107.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vgbixf8g-ctx8192",
   "model": "MiMo-7B-RL",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1150.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-7B-RL-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 1,150 93.6 5.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vgbixf8g",
   "source_url": "https://omlx.ai/benchmarks/performance/vgbixf8g",
   "tg_tokens": null,
   "tps": 93.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-f5hs855u-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 7plb09ou, f542iagc, tiyz9aj8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 941.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 941.0 91.7 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f5hs855u",
   "source_url": "https://omlx.ai/benchmarks/performance/f5hs855u",
   "tg_tokens": null,
   "tps": 91.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-f5hs855u-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 7plb09ou, f542iagc, tiyz9aj8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1077.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,077 87.2 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f5hs855u",
   "source_url": "https://omlx.ai/benchmarks/performance/f5hs855u",
   "tg_tokens": null,
   "tps": 87.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-f5hs855u-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 7plb09ou, f542iagc, tiyz9aj8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 997.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 997.9 82.7 21.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f5hs855u",
   "source_url": "https://omlx.ai/benchmarks/performance/f5hs855u",
   "tg_tokens": null,
   "tps": 82.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-f5hs855u-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 7plb09ou, f542iagc, tiyz9aj8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 904.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 904.2 75.8 21.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f5hs855u",
   "source_url": "https://omlx.ai/benchmarks/performance/f5hs855u",
   "tg_tokens": null,
   "tps": 75.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-6xveliwu-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 5lwsrfa7, grnsxxa2; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1646.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 1,646 69.5 19.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6xveliwu",
   "source_url": "https://omlx.ai/benchmarks/performance/6xveliwu",
   "tg_tokens": null,
   "tps": 69.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-6xveliwu-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 5lwsrfa7, grnsxxa2; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1966.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,966 70.0 19.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6xveliwu",
   "source_url": "https://omlx.ai/benchmarks/performance/6xveliwu",
   "tg_tokens": null,
   "tps": 70.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-6xveliwu-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 5lwsrfa7, grnsxxa2; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1937.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 8k 1,937 66.7 19.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6xveliwu",
   "source_url": "https://omlx.ai/benchmarks/performance/6xveliwu",
   "tg_tokens": null,
   "tps": 66.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-6xveliwu-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 5lwsrfa7, grnsxxa2; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1802.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 1,802 61.5 20.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6xveliwu",
   "source_url": "https://omlx.ai/benchmarks/performance/6xveliwu",
   "tg_tokens": null,
   "tps": 61.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-6xveliwu-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: 5lwsrfa7, grnsxxa2; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1520.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 32k 1,520 51.8 20.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6xveliwu",
   "source_url": "https://omlx.ai/benchmarks/performance/6xveliwu",
   "tg_tokens": null,
   "tps": 51.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-6xveliwu-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: 5lwsrfa7, grnsxxa2; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1133.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 64k 1,133 38.0 21.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6xveliwu",
   "source_url": "https://omlx.ai/benchmarks/performance/6xveliwu",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-br31rm7o-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 7ienoz4v, 8jcvz5bh, knzhcmw3, m4niz9nb, mbbiknnp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 893.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 893.3 78.2 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/br31rm7o",
   "source_url": "https://omlx.ai/benchmarks/performance/br31rm7o",
   "tg_tokens": null,
   "tps": 78.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-br31rm7o-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: 7ienoz4v, 8jcvz5bh, knzhcmw3, m4niz9nb, mbbiknnp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1024.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,024 79.4 23.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/br31rm7o",
   "source_url": "https://omlx.ai/benchmarks/performance/br31rm7o",
   "tg_tokens": null,
   "tps": 79.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-br31rm7o-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: 7ienoz4v, 8jcvz5bh, knzhcmw3, m4niz9nb, mbbiknnp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 954.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 954.7 85.3 23.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/br31rm7o",
   "source_url": "https://omlx.ai/benchmarks/performance/br31rm7o",
   "tg_tokens": null,
   "tps": 85.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-br31rm7o-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: 7ienoz4v, 8jcvz5bh, knzhcmw3, m4niz9nb, mbbiknnp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 837.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 837.7 77.1 24.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/br31rm7o",
   "source_url": "https://omlx.ai/benchmarks/performance/br31rm7o",
   "tg_tokens": null,
   "tps": 77.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-br31rm7o-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: 7ienoz4v, 8jcvz5bh, knzhcmw3, m4niz9nb, mbbiknnp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 678.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 678.6 67.6 25.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/br31rm7o",
   "source_url": "https://omlx.ai/benchmarks/performance/br31rm7o",
   "tg_tokens": null,
   "tps": 67.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-br31rm7o-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: 7ienoz4v, 8jcvz5bh, knzhcmw3, m4niz9nb, mbbiknnp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 491.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 491.9 43.5 26.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/br31rm7o",
   "source_url": "https://omlx.ai/benchmarks/performance/br31rm7o",
   "tg_tokens": null,
   "tps": 43.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-okvgxatt-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "DWQ build; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: onc7jv8r; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1489.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit-DWQ on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 8bit; 1k 1,489 94.6 20.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/okvgxatt",
   "source_url": "https://omlx.ai/benchmarks/performance/okvgxatt",
   "tg_tokens": null,
   "tps": 94.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-okvgxatt-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "DWQ build; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: onc7jv8r; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1468.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit-DWQ on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 8bit; 8k 1,468 86.7 22.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/okvgxatt",
   "source_url": "https://omlx.ai/benchmarks/performance/okvgxatt",
   "tg_tokens": null,
   "tps": 86.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-eucfvguy-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: dssdp0o7, igbn28oo, l9le3ol1, lr1gbf61; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3787.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 3,787 118.1 26.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eucfvguy",
   "source_url": "https://omlx.ai/benchmarks/performance/eucfvguy",
   "tg_tokens": null,
   "tps": 118.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-eucfvguy-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: dssdp0o7, igbn28oo, l9le3ol1, lr1gbf61; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5216.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 5,216 173.4 26.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eucfvguy",
   "source_url": "https://omlx.ai/benchmarks/performance/eucfvguy",
   "tg_tokens": null,
   "tps": 173.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-eucfvguy-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: dssdp0o7, igbn28oo, l9le3ol1, lr1gbf61; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5215.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 5,215 169.6 26.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eucfvguy",
   "source_url": "https://omlx.ai/benchmarks/performance/eucfvguy",
   "tg_tokens": null,
   "tps": 169.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-eucfvguy-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.0 GB at this context; duplicate traces of the same session: dssdp0o7, igbn28oo, l9le3ol1, lr1gbf61; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4892.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 4,892 183.3 27.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eucfvguy",
   "source_url": "https://omlx.ai/benchmarks/performance/eucfvguy",
   "tg_tokens": null,
   "tps": 183.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-eucfvguy-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: dssdp0o7, igbn28oo, l9le3ol1, lr1gbf61; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4084.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 4,084 239.0 27.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eucfvguy",
   "source_url": "https://omlx.ai/benchmarks/performance/eucfvguy",
   "tg_tokens": null,
   "tps": 239.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-eucfvguy-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: dssdp0o7, igbn28oo, l9le3ol1, lr1gbf61; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2991.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 2,991 130.2 28.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eucfvguy",
   "source_url": "https://omlx.ai/benchmarks/performance/eucfvguy",
   "tg_tokens": null,
   "tps": 130.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-eucfvguy-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: dssdp0o7, igbn28oo, l9le3ol1, lr1gbf61; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1992.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 1,992 101.7 29.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eucfvguy",
   "source_url": "https://omlx.ai/benchmarks/performance/eucfvguy",
   "tg_tokens": null,
   "tps": 101.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-f9h20840-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 768.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 1k 768.4 56.4 37.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f9h20840",
   "source_url": "https://omlx.ai/benchmarks/performance/f9h20840",
   "tg_tokens": null,
   "tps": 56.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-f9h20840-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 39.5 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 609.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 32k 609.9 47.6 39.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f9h20840",
   "source_url": "https://omlx.ai/benchmarks/performance/f9h20840",
   "tg_tokens": null,
   "tps": 47.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-v67fc5w0-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: o07mxj9a; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1360.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,360 77.3 17.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v67fc5w0",
   "source_url": "https://omlx.ai/benchmarks/performance/v67fc5w0",
   "tg_tokens": null,
   "tps": 77.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-v67fc5w0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: o07mxj9a; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1628.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,628 97.3 18.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v67fc5w0",
   "source_url": "https://omlx.ai/benchmarks/performance/v67fc5w0",
   "tg_tokens": null,
   "tps": 97.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4oo3c6vx-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: 19vgdd18, f055p843, js5lvb7m, wghm80x6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 939.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 939.2 32.8 17.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4oo3c6vx",
   "source_url": "https://omlx.ai/benchmarks/performance/4oo3c6vx",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4oo3c6vx-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 19vgdd18, f055p843, js5lvb7m, wghm80x6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 973.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 973.1 31.0 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4oo3c6vx",
   "source_url": "https://omlx.ai/benchmarks/performance/4oo3c6vx",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4oo3c6vx-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 19vgdd18, f055p843, js5lvb7m, wghm80x6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 798.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 798.8 25.7 20.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4oo3c6vx",
   "source_url": "https://omlx.ai/benchmarks/performance/4oo3c6vx",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4oo3c6vx-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: 19vgdd18, f055p843, js5lvb7m, wghm80x6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 637.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 637.0 21.6 24.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4oo3c6vx",
   "source_url": "https://omlx.ai/benchmarks/performance/4oo3c6vx",
   "tg_tokens": null,
   "tps": 21.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4oo3c6vx-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.0 GB at this context; duplicate traces of the same session: 19vgdd18, f055p843, js5lvb7m, wghm80x6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 515.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 195k 515.5 18.1 29.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4oo3c6vx",
   "source_url": "https://omlx.ai/benchmarks/performance/4oo3c6vx",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-x5fw8qpx-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: gl6rj7dv, n716phs3; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 277.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 277.0 21.9 17.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x5fw8qpx",
   "source_url": "https://omlx.ai/benchmarks/performance/x5fw8qpx",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-x5fw8qpx-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: gl6rj7dv, n716phs3; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 233.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 233.1 20.5 18.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x5fw8qpx",
   "source_url": "https://omlx.ai/benchmarks/performance/x5fw8qpx",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-x5fw8qpx-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: gl6rj7dv, n716phs3; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 235.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 8k 235.4 26.1 18.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x5fw8qpx",
   "source_url": "https://omlx.ai/benchmarks/performance/x5fw8qpx",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-x5fw8qpx-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: gl6rj7dv, n716phs3; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 222.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 222.4 23.0 19.5 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x5fw8qpx",
   "source_url": "https://omlx.ai/benchmarks/performance/x5fw8qpx",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-x5fw8qpx-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: gl6rj7dv, n716phs3; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 214.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 32k 214.9 25.3 20.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x5fw8qpx",
   "source_url": "https://omlx.ai/benchmarks/performance/x5fw8qpx",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-x5fw8qpx-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: gl6rj7dv, n716phs3; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 193.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 64k 193.3 17.9 23.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x5fw8qpx",
   "source_url": "https://omlx.ai/benchmarks/performance/x5fw8qpx",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-r5byxc05-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: 36asr4pa, p8461p1d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 899.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 899.1 67.5 24.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r5byxc05",
   "source_url": "https://omlx.ai/benchmarks/performance/r5byxc05",
   "tg_tokens": null,
   "tps": 67.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-r5byxc05-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.7 GB at this context; duplicate traces of the same session: 36asr4pa, p8461p1d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1157.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 1,157 58.7 25.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r5byxc05",
   "source_url": "https://omlx.ai/benchmarks/performance/r5byxc05",
   "tg_tokens": null,
   "tps": 58.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-r5byxc05-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: 36asr4pa, p8461p1d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1125.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 1,125 60.4 26.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r5byxc05",
   "source_url": "https://omlx.ai/benchmarks/performance/r5byxc05",
   "tg_tokens": null,
   "tps": 60.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-r5byxc05-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: 36asr4pa, p8461p1d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1048.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 1,048 78.8 26.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r5byxc05",
   "source_url": "https://omlx.ai/benchmarks/performance/r5byxc05",
   "tg_tokens": null,
   "tps": 78.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-r5byxc05-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: 36asr4pa, p8461p1d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 919.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 919.9 113.4 27.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r5byxc05",
   "source_url": "https://omlx.ai/benchmarks/performance/r5byxc05",
   "tg_tokens": null,
   "tps": 113.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-r5byxc05-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.0 GB at this context; duplicate traces of the same session: 36asr4pa, p8461p1d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 754.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 754.1 39.1 30.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r5byxc05",
   "source_url": "https://omlx.ai/benchmarks/performance/r5byxc05",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-r5byxc05-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: 36asr4pa, p8461p1d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 538.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 538.9 39.8 34.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r5byxc05",
   "source_url": "https://omlx.ai/benchmarks/performance/r5byxc05",
   "tg_tokens": null,
   "tps": 39.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ytavq1i9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: kvg8ovtl, uadv12ta; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 831.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 831.7 55.9 23.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ytavq1i9",
   "source_url": "https://omlx.ai/benchmarks/performance/ytavq1i9",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ytavq1i9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: kvg8ovtl, uadv12ta; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 886.6,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 886.6 53.0 24.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ytavq1i9",
   "source_url": "https://omlx.ai/benchmarks/performance/ytavq1i9",
   "tg_tokens": null,
   "tps": 53.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ytavq1i9-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: kvg8ovtl, uadv12ta; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 816.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 816.0 66.0 24.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ytavq1i9",
   "source_url": "https://omlx.ai/benchmarks/performance/ytavq1i9",
   "tg_tokens": null,
   "tps": 66.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ytavq1i9-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.2 GB at this context; duplicate traces of the same session: kvg8ovtl, uadv12ta; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 783.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 783.9 75.3 25.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ytavq1i9",
   "source_url": "https://omlx.ai/benchmarks/performance/ytavq1i9",
   "tg_tokens": null,
   "tps": 75.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ytavq1i9-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: kvg8ovtl, uadv12ta; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 676.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 676.4 113.5 26.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ytavq1i9",
   "source_url": "https://omlx.ai/benchmarks/performance/ytavq1i9",
   "tg_tokens": null,
   "tps": 113.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ytavq1i9-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.7 GB at this context; duplicate traces of the same session: kvg8ovtl, uadv12ta; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 566.3,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 566.3 44.5 28.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ytavq1i9",
   "source_url": "https://omlx.ai/benchmarks/performance/ytavq1i9",
   "tg_tokens": null,
   "tps": 44.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ytavq1i9-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.4 GB at this context; duplicate traces of the same session: kvg8ovtl, uadv12ta; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 442.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 442.4 40.1 33.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ytavq1i9",
   "source_url": "https://omlx.ai/benchmarks/performance/ytavq1i9",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-0kaewohk-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 119.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 1k 119.4 11.1 30.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0kaewohk",
   "source_url": "https://omlx.ai/benchmarks/performance/0kaewohk",
   "tg_tokens": null,
   "tps": 11.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-0kaewohk-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 35.3 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 32k 104.8 9.5 35.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0kaewohk",
   "source_url": "https://omlx.ai/benchmarks/performance/0kaewohk",
   "tg_tokens": null,
   "tps": 9.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-oby062k4-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 122.5 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 852.5,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 1k 852.5 89.2 122.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oby062k4",
   "source_url": "https://omlx.ai/benchmarks/performance/oby062k4",
   "tg_tokens": null,
   "tps": 89.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-oby062k4-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 122.8 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1043.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 4k 1,043 82.3 122.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oby062k4",
   "source_url": "https://omlx.ai/benchmarks/performance/oby062k4",
   "tg_tokens": null,
   "tps": 82.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mfqdarc4-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: 4riwznwc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1128.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,128 75.0 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mfqdarc4",
   "source_url": "https://omlx.ai/benchmarks/performance/mfqdarc4",
   "tg_tokens": null,
   "tps": 75.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mfqdarc4-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: 4riwznwc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1763.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,763 71.4 70.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mfqdarc4",
   "source_url": "https://omlx.ai/benchmarks/performance/mfqdarc4",
   "tg_tokens": null,
   "tps": 71.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mfqdarc4-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 4riwznwc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2155.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,155 88.9 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mfqdarc4",
   "source_url": "https://omlx.ai/benchmarks/performance/mfqdarc4",
   "tg_tokens": null,
   "tps": 88.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mfqdarc4-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.7 GB at this context; duplicate traces of the same session: 4riwznwc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2327.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,327 78.7 72.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mfqdarc4",
   "source_url": "https://omlx.ai/benchmarks/performance/mfqdarc4",
   "tg_tokens": null,
   "tps": 78.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mfqdarc4-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.5 GB at this context; duplicate traces of the same session: 4riwznwc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2213.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,213 63.2 80.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mfqdarc4",
   "source_url": "https://omlx.ai/benchmarks/performance/mfqdarc4",
   "tg_tokens": null,
   "tps": 63.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fd13w0ls-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: m4jbgzax; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1161.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,161 62.4 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fd13w0ls",
   "source_url": "https://omlx.ai/benchmarks/performance/fd13w0ls",
   "tg_tokens": null,
   "tps": 62.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fd13w0ls-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: m4jbgzax; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2027.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,027 71.7 70.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fd13w0ls",
   "source_url": "https://omlx.ai/benchmarks/performance/fd13w0ls",
   "tg_tokens": null,
   "tps": 71.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fd13w0ls-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: m4jbgzax; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2465.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,465 82.0 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fd13w0ls",
   "source_url": "https://omlx.ai/benchmarks/performance/fd13w0ls",
   "tg_tokens": null,
   "tps": 82.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fd13w0ls-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.5 GB at this context; duplicate traces of the same session: m4jbgzax; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2243.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,243 62.7 80.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fd13w0ls",
   "source_url": "https://omlx.ai/benchmarks/performance/fd13w0ls",
   "tg_tokens": null,
   "tps": 62.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-svxr6uce-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 141.9 GB at this context; duplicate traces of the same session: hzlneami; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1047.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 64k 1,047 56.2 141.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/svxr6uce",
   "source_url": "https://omlx.ai/benchmarks/performance/svxr6uce",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-svxr6uce-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.1 GB at this context; duplicate traces of the same session: hzlneami; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1030.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 128k 1,030 51.8 145.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/svxr6uce",
   "source_url": "https://omlx.ai/benchmarks/performance/svxr6uce",
   "tg_tokens": null,
   "tps": 51.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-svxr6uce-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 149.0 GB at this context; duplicate traces of the same session: hzlneami; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1018.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 195k 1,018 50.1 149.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/svxr6uce",
   "source_url": "https://omlx.ai/benchmarks/performance/svxr6uce",
   "tg_tokens": null,
   "tps": 50.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-29oq4idn-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 183.1 GB; TTFT 1275 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 802.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k tokens 802.9 PP tok/s 83.1 TG tok/s 1275 TTFT (ms) 183.1 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/29oq4idn",
   "source_url": "https://omlx.ai/benchmarks/performance/29oq4idn",
   "tg_tokens": null,
   "tps": 83.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-c4um5x2a-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: yke0u1ug; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1066.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,066 58.7 19.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c4um5x2a",
   "source_url": "https://omlx.ai/benchmarks/performance/c4um5x2a",
   "tg_tokens": null,
   "tps": 58.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-c4um5x2a-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: yke0u1ug; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1313.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,313 63.0 20.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c4um5x2a",
   "source_url": "https://omlx.ai/benchmarks/performance/c4um5x2a",
   "tg_tokens": null,
   "tps": 63.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-c4um5x2a-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: yke0u1ug; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1334.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,334 57.5 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c4um5x2a",
   "source_url": "https://omlx.ai/benchmarks/performance/c4um5x2a",
   "tg_tokens": null,
   "tps": 57.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-57insz4w-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: psztr93l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1118.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,118 58.2 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/57insz4w",
   "source_url": "https://omlx.ai/benchmarks/performance/57insz4w",
   "tg_tokens": null,
   "tps": 58.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-57insz4w-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: psztr93l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1331.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,331 64.4 19.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/57insz4w",
   "source_url": "https://omlx.ai/benchmarks/performance/57insz4w",
   "tg_tokens": null,
   "tps": 64.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-57insz4w-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: psztr93l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1321.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,321 69.8 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/57insz4w",
   "source_url": "https://omlx.ai/benchmarks/performance/57insz4w",
   "tg_tokens": null,
   "tps": 69.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-83u5c4mm-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.5 GB at this context; duplicate traces of the same session: 0ltnb0m3, wg1z3mi4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 245.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 245.7 41.2 32.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/83u5c4mm",
   "source_url": "https://omlx.ai/benchmarks/performance/83u5c4mm",
   "tg_tokens": null,
   "tps": 41.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-83u5c4mm-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.7 GB at this context; duplicate traces of the same session: 0ltnb0m3, wg1z3mi4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 274.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 274.1 41.6 35.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/83u5c4mm",
   "source_url": "https://omlx.ai/benchmarks/performance/83u5c4mm",
   "tg_tokens": null,
   "tps": 41.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-83u5c4mm-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: 0ltnb0m3, wg1z3mi4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 277.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 8k 277.3 37.8 35.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/83u5c4mm",
   "source_url": "https://omlx.ai/benchmarks/performance/83u5c4mm",
   "tg_tokens": null,
   "tps": 37.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-83u5c4mm-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.5 GB at this context; duplicate traces of the same session: 0ltnb0m3, wg1z3mi4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 251.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k 251.3 37.0 31.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/83u5c4mm",
   "source_url": "https://omlx.ai/benchmarks/performance/83u5c4mm",
   "tg_tokens": null,
   "tps": 37.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-83u5c4mm-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.3 GB at this context; duplicate traces of the same session: 0ltnb0m3, wg1z3mi4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 233.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 233.1 29.8 33.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/83u5c4mm",
   "source_url": "https://omlx.ai/benchmarks/performance/83u5c4mm",
   "tg_tokens": null,
   "tps": 29.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-v5cutx1e-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "chriswessels ThinkingCap; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 203.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "chriswessels--ThinkingCap-Qwen3.8-27B-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 203.7 36.5 18.6 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v5cutx1e",
   "source_url": "https://omlx.ai/benchmarks/performance/v5cutx1e",
   "tg_tokens": null,
   "tps": 36.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-v5cutx1e-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "chriswessels ThinkingCap; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "chriswessels--ThinkingCap-Qwen3.8-27B-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 205.2 30.4 19.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v5cutx1e",
   "source_url": "https://omlx.ai/benchmarks/performance/v5cutx1e",
   "tg_tokens": null,
   "tps": 30.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-v5cutx1e-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "chriswessels ThinkingCap; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 202.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "chriswessels--ThinkingCap-Qwen3.8-27B-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 202.7 34.3 20.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v5cutx1e",
   "source_url": "https://omlx.ai/benchmarks/performance/v5cutx1e",
   "tg_tokens": null,
   "tps": 34.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-v5cutx1e-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "chriswessels ThinkingCap; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 194.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "chriswessels--ThinkingCap-Qwen3.8-27B-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 194.3 29.0 20.4 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v5cutx1e",
   "source_url": "https://omlx.ai/benchmarks/performance/v5cutx1e",
   "tg_tokens": null,
   "tps": 29.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-wrlvh6ig-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: 7jbivu99, 8ezii5v6, e76si6sh, fzoc69ia; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 676.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 1k 676.0 70.3 14.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wrlvh6ig",
   "source_url": "https://omlx.ai/benchmarks/performance/wrlvh6ig",
   "tg_tokens": null,
   "tps": 70.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-wrlvh6ig-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: 7jbivu99, 8ezii5v6, e76si6sh, fzoc69ia; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 672.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 4k 672.4 67.1 14.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wrlvh6ig",
   "source_url": "https://omlx.ai/benchmarks/performance/wrlvh6ig",
   "tg_tokens": null,
   "tps": 67.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-wrlvh6ig-ctx8192",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: 7jbivu99, 8ezii5v6, e76si6sh, fzoc69ia; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 662.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 8k 662.4 63.1 14.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wrlvh6ig",
   "source_url": "https://omlx.ai/benchmarks/performance/wrlvh6ig",
   "tg_tokens": null,
   "tps": 63.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-wrlvh6ig-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 15.2 GB at this context; duplicate traces of the same session: 7jbivu99, 8ezii5v6, e76si6sh, fzoc69ia; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 632.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 16k 632.0 56.2 15.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wrlvh6ig",
   "source_url": "https://omlx.ai/benchmarks/performance/wrlvh6ig",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-wrlvh6ig-ctx32768",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 7jbivu99, 8ezii5v6, e76si6sh, fzoc69ia; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 578.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 32k 578.4 46.9 16.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wrlvh6ig",
   "source_url": "https://omlx.ai/benchmarks/performance/wrlvh6ig",
   "tg_tokens": null,
   "tps": 46.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-wrlvh6ig-ctx65536",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: 7jbivu99, 8ezii5v6, e76si6sh, fzoc69ia; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 493.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 64k 493.5 34.6 17.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wrlvh6ig",
   "source_url": "https://omlx.ai/benchmarks/performance/wrlvh6ig",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ay300quz-ctx1024",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 28cxtguv, abfhn6u1, c1veqwq1, qfxswvss; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 208.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 208.4 20.5 18.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ay300quz",
   "source_url": "https://omlx.ai/benchmarks/performance/ay300quz",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ay300quz-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 28cxtguv, abfhn6u1, c1veqwq1, qfxswvss; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 205.8 23.0 19.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ay300quz",
   "source_url": "https://omlx.ai/benchmarks/performance/ay300quz",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ay300quz-ctx8192",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 28cxtguv, abfhn6u1, c1veqwq1, qfxswvss; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 197.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 197.3 22.4 19.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ay300quz",
   "source_url": "https://omlx.ai/benchmarks/performance/ay300quz",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ay300quz-ctx16384",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: 28cxtguv, abfhn6u1, c1veqwq1, qfxswvss; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 193.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 193.0 21.5 21.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ay300quz",
   "source_url": "https://omlx.ai/benchmarks/performance/ay300quz",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ay300quz-ctx32768",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 28cxtguv, abfhn6u1, c1veqwq1, qfxswvss; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 178.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 178.0 19.8 23.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ay300quz",
   "source_url": "https://omlx.ai/benchmarks/performance/ay300quz",
   "tg_tokens": null,
   "tps": 19.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ay300quz-ctx65536",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: 28cxtguv, abfhn6u1, c1veqwq1, qfxswvss; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 155.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 155.1 17.0 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ay300quz",
   "source_url": "https://omlx.ai/benchmarks/performance/ay300quz",
   "tg_tokens": null,
   "tps": 17.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ym5f7df9-ctx1024",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 87.8 GB at this context; duplicate traces of the same session: 8n6mka8s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 435.5,
   "provenance": "community",
   "quant": "2.4bit mixed",
   "quote": "DeepSeek-V4-Flash-0731-2.4bit-mixed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 1k 435.5 39.9 87.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ym5f7df9",
   "source_url": "https://omlx.ai/benchmarks/performance/ym5f7df9",
   "tg_tokens": null,
   "tps": 39.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ym5f7df9-ctx4096",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 88.3 GB at this context; duplicate traces of the same session: 8n6mka8s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 714.9,
   "provenance": "community",
   "quant": "2.4bit mixed",
   "quote": "DeepSeek-V4-Flash-0731-2.4bit-mixed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 4k 714.9 40.9 88.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ym5f7df9",
   "source_url": "https://omlx.ai/benchmarks/performance/ym5f7df9",
   "tg_tokens": null,
   "tps": 40.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ri66hfuh-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Huihui abliterated build; oMLX community benchmark trace, single run at 1k context; peak memory 23.1 GB; TTFT 5192 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 197.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.8-Flash-Next-abliterated-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k tokens 197.2 PP tok/s 25.5 TG tok/s 5192 TTFT (ms) 23.1 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ri66hfuh",
   "source_url": "https://omlx.ai/benchmarks/performance/ri66hfuh",
   "tg_tokens": null,
   "tps": 25.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-v7e5wwky-ctx16384",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: n577x7av; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1452.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 1,452 91.7 34.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v7e5wwky",
   "source_url": "https://omlx.ai/benchmarks/performance/v7e5wwky",
   "tg_tokens": null,
   "tps": 91.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-v7e5wwky-ctx32768",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: n577x7av; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1254.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 1,254 88.1 34.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v7e5wwky",
   "source_url": "https://omlx.ai/benchmarks/performance/v7e5wwky",
   "tg_tokens": null,
   "tps": 88.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-tiyz9aj8-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 7plb09ou, f542iagc; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 941.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 941.0 91.7 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tiyz9aj8",
   "source_url": "https://omlx.ai/benchmarks/performance/tiyz9aj8",
   "tg_tokens": null,
   "tps": 91.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-tiyz9aj8-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 7plb09ou, f542iagc; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1077.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,077 87.2 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tiyz9aj8",
   "source_url": "https://omlx.ai/benchmarks/performance/tiyz9aj8",
   "tg_tokens": null,
   "tps": 87.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-tiyz9aj8-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 7plb09ou, f542iagc; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 997.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 997.9 82.7 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tiyz9aj8",
   "source_url": "https://omlx.ai/benchmarks/performance/tiyz9aj8",
   "tg_tokens": null,
   "tps": 82.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-tiyz9aj8-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 7plb09ou, f542iagc; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 904.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 904.2 75.8 21.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tiyz9aj8",
   "source_url": "https://omlx.ai/benchmarks/performance/tiyz9aj8",
   "tg_tokens": null,
   "tps": 75.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-5lwsrfa7-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: grnsxxa2; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1646.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 1,646 69.5 19.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5lwsrfa7",
   "source_url": "https://omlx.ai/benchmarks/performance/5lwsrfa7",
   "tg_tokens": null,
   "tps": 69.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-5lwsrfa7-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: grnsxxa2; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1966.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,966 70.0 19.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5lwsrfa7",
   "source_url": "https://omlx.ai/benchmarks/performance/5lwsrfa7",
   "tg_tokens": null,
   "tps": 70.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-5lwsrfa7-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: grnsxxa2; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1937.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 8k 1,937 66.7 19.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5lwsrfa7",
   "source_url": "https://omlx.ai/benchmarks/performance/5lwsrfa7",
   "tg_tokens": null,
   "tps": 66.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-5lwsrfa7-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: grnsxxa2; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1802.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 1,802 61.5 20.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5lwsrfa7",
   "source_url": "https://omlx.ai/benchmarks/performance/5lwsrfa7",
   "tg_tokens": null,
   "tps": 61.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-5lwsrfa7-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: grnsxxa2; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1520.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 32k 1,520 51.8 20.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5lwsrfa7",
   "source_url": "https://omlx.ai/benchmarks/performance/5lwsrfa7",
   "tg_tokens": null,
   "tps": 51.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-5lwsrfa7-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: grnsxxa2; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1133.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 64k 1,133 38.0 21.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5lwsrfa7",
   "source_url": "https://omlx.ai/benchmarks/performance/5lwsrfa7",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-mbbiknnp-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 7ienoz4v, 8jcvz5bh, knzhcmw3, m4niz9nb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 893.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 893.3 78.2 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mbbiknnp",
   "source_url": "https://omlx.ai/benchmarks/performance/mbbiknnp",
   "tg_tokens": null,
   "tps": 78.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-mbbiknnp-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: 7ienoz4v, 8jcvz5bh, knzhcmw3, m4niz9nb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1024.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,024 79.4 23.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mbbiknnp",
   "source_url": "https://omlx.ai/benchmarks/performance/mbbiknnp",
   "tg_tokens": null,
   "tps": 79.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-mbbiknnp-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: 7ienoz4v, 8jcvz5bh, knzhcmw3, m4niz9nb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 954.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 954.7 85.3 23.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mbbiknnp",
   "source_url": "https://omlx.ai/benchmarks/performance/mbbiknnp",
   "tg_tokens": null,
   "tps": 85.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-mbbiknnp-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: 7ienoz4v, 8jcvz5bh, knzhcmw3, m4niz9nb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 837.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 837.7 77.1 24.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mbbiknnp",
   "source_url": "https://omlx.ai/benchmarks/performance/mbbiknnp",
   "tg_tokens": null,
   "tps": 77.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-mbbiknnp-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: 7ienoz4v, 8jcvz5bh, knzhcmw3, m4niz9nb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 678.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 678.6 67.6 25.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mbbiknnp",
   "source_url": "https://omlx.ai/benchmarks/performance/mbbiknnp",
   "tg_tokens": null,
   "tps": 67.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-mbbiknnp-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: 7ienoz4v, 8jcvz5bh, knzhcmw3, m4niz9nb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 491.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 491.9 43.5 26.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mbbiknnp",
   "source_url": "https://omlx.ai/benchmarks/performance/mbbiknnp",
   "tg_tokens": null,
   "tps": 43.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-ysixsbuv-ctx4096",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "MLX; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: qepei3pf; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1178.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-MLX-5bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 5bit; 4k 1,178 41.0 20.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ysixsbuv",
   "source_url": "https://omlx.ai/benchmarks/performance/ysixsbuv",
   "tg_tokens": null,
   "tps": 41.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-ysixsbuv-ctx16384",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "MLX; oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: qepei3pf; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 797.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-MLX-5bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 5bit; 16k 797.0 27.9 21.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ysixsbuv",
   "source_url": "https://omlx.ai/benchmarks/performance/ysixsbuv",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-onc7jv8r-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "DWQ build; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1489.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit-DWQ on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 8bit; 1k 1,489 94.6 20.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/onc7jv8r",
   "source_url": "https://omlx.ai/benchmarks/performance/onc7jv8r",
   "tg_tokens": null,
   "tps": 94.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-onc7jv8r-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "DWQ build; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1468.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit-DWQ on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 8bit; 8k 1,468 86.7 22.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/onc7jv8r",
   "source_url": "https://omlx.ai/benchmarks/performance/onc7jv8r",
   "tg_tokens": null,
   "tps": 86.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bs30lxn1-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: duri0ypz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 761.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 761.1 107.2 35.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bs30lxn1",
   "source_url": "https://omlx.ai/benchmarks/performance/bs30lxn1",
   "tg_tokens": null,
   "tps": 107.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bs30lxn1-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: duri0ypz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2396.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 2,396 105.6 36.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bs30lxn1",
   "source_url": "https://omlx.ai/benchmarks/performance/bs30lxn1",
   "tg_tokens": null,
   "tps": 105.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lr1gbf61-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: dssdp0o7, igbn28oo, l9le3ol1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3787.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 3,787 118.1 26.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lr1gbf61",
   "source_url": "https://omlx.ai/benchmarks/performance/lr1gbf61",
   "tg_tokens": null,
   "tps": 118.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lr1gbf61-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: dssdp0o7, igbn28oo, l9le3ol1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5216.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 5,216 173.4 26.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lr1gbf61",
   "source_url": "https://omlx.ai/benchmarks/performance/lr1gbf61",
   "tg_tokens": null,
   "tps": 173.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lr1gbf61-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: dssdp0o7, igbn28oo, l9le3ol1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5215.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 5,215 169.6 26.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lr1gbf61",
   "source_url": "https://omlx.ai/benchmarks/performance/lr1gbf61",
   "tg_tokens": null,
   "tps": 169.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lr1gbf61-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.0 GB at this context; duplicate traces of the same session: dssdp0o7, igbn28oo, l9le3ol1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4892.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 4,892 183.3 27.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lr1gbf61",
   "source_url": "https://omlx.ai/benchmarks/performance/lr1gbf61",
   "tg_tokens": null,
   "tps": 183.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lr1gbf61-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: dssdp0o7, igbn28oo, l9le3ol1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4084.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 4,084 239.0 27.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lr1gbf61",
   "source_url": "https://omlx.ai/benchmarks/performance/lr1gbf61",
   "tg_tokens": null,
   "tps": 239.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lr1gbf61-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: dssdp0o7, igbn28oo, l9le3ol1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2991.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 2,991 130.2 28.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lr1gbf61",
   "source_url": "https://omlx.ai/benchmarks/performance/lr1gbf61",
   "tg_tokens": null,
   "tps": 130.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lr1gbf61-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: dssdp0o7, igbn28oo, l9le3ol1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1992.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 1,992 101.7 29.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lr1gbf61",
   "source_url": "https://omlx.ai/benchmarks/performance/lr1gbf61",
   "tg_tokens": null,
   "tps": 101.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-879p8cnh-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 25.5 GB; TTFT 15602 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 262.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 262.5 PP tok/s 30.4 TG tok/s 15602 TTFT (ms) 25.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/879p8cnh",
   "source_url": "https://omlx.ai/benchmarks/performance/879p8cnh",
   "tg_tokens": null,
   "tps": 30.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-o07mxj9a-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1360.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,360 77.3 17.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o07mxj9a",
   "source_url": "https://omlx.ai/benchmarks/performance/o07mxj9a",
   "tg_tokens": null,
   "tps": 77.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-o07mxj9a-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1628.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,628 97.3 18.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o07mxj9a",
   "source_url": "https://omlx.ai/benchmarks/performance/o07mxj9a",
   "tg_tokens": null,
   "tps": 97.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-e8gyij6x-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1375.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,375 84.7 17.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e8gyij6x",
   "source_url": "https://omlx.ai/benchmarks/performance/e8gyij6x",
   "tg_tokens": null,
   "tps": 84.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-e8gyij6x-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1624.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,624 137.9 18.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e8gyij6x",
   "source_url": "https://omlx.ai/benchmarks/performance/e8gyij6x",
   "tg_tokens": null,
   "tps": 137.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-e8gyij6x-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1619.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 1,619 61.7 19.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e8gyij6x",
   "source_url": "https://omlx.ai/benchmarks/performance/e8gyij6x",
   "tg_tokens": null,
   "tps": 61.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-e8gyij6x-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1566.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,566 73.8 19.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e8gyij6x",
   "source_url": "https://omlx.ai/benchmarks/performance/e8gyij6x",
   "tg_tokens": null,
   "tps": 73.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-e8gyij6x-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1453.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,453 116.5 20.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e8gyij6x",
   "source_url": "https://omlx.ai/benchmarks/performance/e8gyij6x",
   "tg_tokens": null,
   "tps": 116.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-e8gyij6x-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1255.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,255 77.1 23.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e8gyij6x",
   "source_url": "https://omlx.ai/benchmarks/performance/e8gyij6x",
   "tg_tokens": null,
   "tps": 77.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-e8gyij6x-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 978.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 978.9 77.6 27.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e8gyij6x",
   "source_url": "https://omlx.ai/benchmarks/performance/e8gyij6x",
   "tg_tokens": null,
   "tps": 77.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-e8gyij6x-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 33.0 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 795.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 795.4 69.2 33.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e8gyij6x",
   "source_url": "https://omlx.ai/benchmarks/performance/e8gyij6x",
   "tg_tokens": null,
   "tps": 69.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-f055p843-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: 19vgdd18, js5lvb7m, wghm80x6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 939.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 939.2 32.8 17.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f055p843",
   "source_url": "https://omlx.ai/benchmarks/performance/f055p843",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-f055p843-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 19vgdd18, js5lvb7m, wghm80x6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 973.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 973.1 31.0 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f055p843",
   "source_url": "https://omlx.ai/benchmarks/performance/f055p843",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-f055p843-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 19vgdd18, js5lvb7m, wghm80x6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 798.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 798.8 25.7 20.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f055p843",
   "source_url": "https://omlx.ai/benchmarks/performance/f055p843",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-f055p843-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: 19vgdd18, js5lvb7m, wghm80x6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 637.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 637.0 21.6 24.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f055p843",
   "source_url": "https://omlx.ai/benchmarks/performance/f055p843",
   "tg_tokens": null,
   "tps": 21.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-f055p843-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.0 GB at this context; duplicate traces of the same session: 19vgdd18, js5lvb7m, wghm80x6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 515.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 195k 515.5 18.1 29.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f055p843",
   "source_url": "https://omlx.ai/benchmarks/performance/f055p843",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bkusuzqt-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: x3prmi86; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 820.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 820.7 18.8 28.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bkusuzqt",
   "source_url": "https://omlx.ai/benchmarks/performance/bkusuzqt",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bkusuzqt-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: x3prmi86; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 836.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 836.7 18.7 29.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bkusuzqt",
   "source_url": "https://omlx.ai/benchmarks/performance/bkusuzqt",
   "tg_tokens": null,
   "tps": 18.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-gl6rj7dv-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: n716phs3; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 277.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 277.0 21.9 17.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gl6rj7dv",
   "source_url": "https://omlx.ai/benchmarks/performance/gl6rj7dv",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-gl6rj7dv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: n716phs3; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 233.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 233.1 20.5 18.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gl6rj7dv",
   "source_url": "https://omlx.ai/benchmarks/performance/gl6rj7dv",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-gl6rj7dv-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: n716phs3; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 235.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 8k 235.4 26.1 18.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gl6rj7dv",
   "source_url": "https://omlx.ai/benchmarks/performance/gl6rj7dv",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-gl6rj7dv-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: n716phs3; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 222.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 222.4 23.0 19.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gl6rj7dv",
   "source_url": "https://omlx.ai/benchmarks/performance/gl6rj7dv",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-gl6rj7dv-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: n716phs3; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 214.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 32k 214.9 25.3 20.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gl6rj7dv",
   "source_url": "https://omlx.ai/benchmarks/performance/gl6rj7dv",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-gl6rj7dv-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: n716phs3; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 193.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 64k 193.3 17.9 23.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gl6rj7dv",
   "source_url": "https://omlx.ai/benchmarks/performance/gl6rj7dv",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-36asr4pa-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: f05iixoj, p8461p1d, sfck1s2e, t5e5p84z, v4z1qwxa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 899.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 899.1 67.5 24.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/36asr4pa",
   "source_url": "https://omlx.ai/benchmarks/performance/36asr4pa",
   "tg_tokens": null,
   "tps": 67.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-36asr4pa-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.7 GB at this context; duplicate traces of the same session: f05iixoj, p8461p1d, sfck1s2e, t5e5p84z, v4z1qwxa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1157.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 1,157 58.7 25.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/36asr4pa",
   "source_url": "https://omlx.ai/benchmarks/performance/36asr4pa",
   "tg_tokens": null,
   "tps": 58.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-36asr4pa-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: f05iixoj, p8461p1d, sfck1s2e, t5e5p84z, v4z1qwxa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1125.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 1,125 60.4 26.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/36asr4pa",
   "source_url": "https://omlx.ai/benchmarks/performance/36asr4pa",
   "tg_tokens": null,
   "tps": 60.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-36asr4pa-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: f05iixoj, p8461p1d, sfck1s2e, t5e5p84z, v4z1qwxa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1048.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 1,048 78.8 26.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/36asr4pa",
   "source_url": "https://omlx.ai/benchmarks/performance/36asr4pa",
   "tg_tokens": null,
   "tps": 78.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-36asr4pa-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: f05iixoj, p8461p1d, sfck1s2e, t5e5p84z, v4z1qwxa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 919.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 919.9 113.4 27.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/36asr4pa",
   "source_url": "https://omlx.ai/benchmarks/performance/36asr4pa",
   "tg_tokens": null,
   "tps": 113.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-36asr4pa-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.0 GB at this context; duplicate traces of the same session: f05iixoj, p8461p1d, sfck1s2e, t5e5p84z, v4z1qwxa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 754.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 754.1 39.1 30.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/36asr4pa",
   "source_url": "https://omlx.ai/benchmarks/performance/36asr4pa",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-36asr4pa-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: f05iixoj, p8461p1d, sfck1s2e, t5e5p84z, v4z1qwxa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 538.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 538.9 39.8 34.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/36asr4pa",
   "source_url": "https://omlx.ai/benchmarks/performance/36asr4pa",
   "tg_tokens": null,
   "tps": 39.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-uadv12ta-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: i8ref11h, kvg8ovtl, p1os044e, pfhzn4ht, qsmyfdcx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 831.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 831.7 55.9 23.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uadv12ta",
   "source_url": "https://omlx.ai/benchmarks/performance/uadv12ta",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-uadv12ta-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: i8ref11h, kvg8ovtl, p1os044e, pfhzn4ht, qsmyfdcx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 886.6,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 886.6 53.0 24.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uadv12ta",
   "source_url": "https://omlx.ai/benchmarks/performance/uadv12ta",
   "tg_tokens": null,
   "tps": 53.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-uadv12ta-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: i8ref11h, kvg8ovtl, p1os044e, pfhzn4ht, qsmyfdcx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 816.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 816.0 66.0 24.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uadv12ta",
   "source_url": "https://omlx.ai/benchmarks/performance/uadv12ta",
   "tg_tokens": null,
   "tps": 66.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-uadv12ta-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.2 GB at this context; duplicate traces of the same session: i8ref11h, kvg8ovtl, p1os044e, pfhzn4ht, qsmyfdcx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 783.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 783.9 75.3 25.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uadv12ta",
   "source_url": "https://omlx.ai/benchmarks/performance/uadv12ta",
   "tg_tokens": null,
   "tps": 75.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-uadv12ta-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: i8ref11h, kvg8ovtl, p1os044e, pfhzn4ht, qsmyfdcx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 676.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 676.4 113.5 26.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uadv12ta",
   "source_url": "https://omlx.ai/benchmarks/performance/uadv12ta",
   "tg_tokens": null,
   "tps": 113.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-uadv12ta-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.7 GB at this context; duplicate traces of the same session: i8ref11h, kvg8ovtl, p1os044e, pfhzn4ht, qsmyfdcx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 566.3,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 566.3 44.5 28.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uadv12ta",
   "source_url": "https://omlx.ai/benchmarks/performance/uadv12ta",
   "tg_tokens": null,
   "tps": 44.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-uadv12ta-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.4 GB at this context; duplicate traces of the same session: i8ref11h, kvg8ovtl, p1os044e, pfhzn4ht, qsmyfdcx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 442.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 442.4 40.1 33.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uadv12ta",
   "source_url": "https://omlx.ai/benchmarks/performance/uadv12ta",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4riwznwc-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1128.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,128 75.0 70.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4riwznwc",
   "source_url": "https://omlx.ai/benchmarks/performance/4riwznwc",
   "tg_tokens": null,
   "tps": 75.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4riwznwc-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1763.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,763 71.4 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4riwznwc",
   "source_url": "https://omlx.ai/benchmarks/performance/4riwznwc",
   "tg_tokens": null,
   "tps": 71.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4riwznwc-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2155.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,155 88.9 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4riwznwc",
   "source_url": "https://omlx.ai/benchmarks/performance/4riwznwc",
   "tg_tokens": null,
   "tps": 88.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4riwznwc-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2327.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,327 78.7 72.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4riwznwc",
   "source_url": "https://omlx.ai/benchmarks/performance/4riwznwc",
   "tg_tokens": null,
   "tps": 78.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4riwznwc-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2213.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,213 63.2 80.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4riwznwc",
   "source_url": "https://omlx.ai/benchmarks/performance/4riwznwc",
   "tg_tokens": null,
   "tps": 63.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-m4jbgzax-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1161.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,161 62.4 70.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m4jbgzax",
   "source_url": "https://omlx.ai/benchmarks/performance/m4jbgzax",
   "tg_tokens": null,
   "tps": 62.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-m4jbgzax-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2027.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,027 71.7 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m4jbgzax",
   "source_url": "https://omlx.ai/benchmarks/performance/m4jbgzax",
   "tg_tokens": null,
   "tps": 71.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-m4jbgzax-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2465.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,465 82.0 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m4jbgzax",
   "source_url": "https://omlx.ai/benchmarks/performance/m4jbgzax",
   "tg_tokens": null,
   "tps": 82.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-m4jbgzax-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2243.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,243 62.7 80.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m4jbgzax",
   "source_url": "https://omlx.ai/benchmarks/performance/m4jbgzax",
   "tg_tokens": null,
   "tps": 62.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-f5iosw3g-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: 05isnu7u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1002.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,002 65.1 71.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f5iosw3g",
   "source_url": "https://omlx.ai/benchmarks/performance/f5iosw3g",
   "tg_tokens": null,
   "tps": 65.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-f5iosw3g-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.2 GB at this context; duplicate traces of the same session: 05isnu7u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1962.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,962 65.2 71.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f5iosw3g",
   "source_url": "https://omlx.ai/benchmarks/performance/f5iosw3g",
   "tg_tokens": null,
   "tps": 65.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hzlneami-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 141.9 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1047.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 64k 1,047 56.2 141.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hzlneami",
   "source_url": "https://omlx.ai/benchmarks/performance/hzlneami",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-hzlneami-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.1 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1030.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 128k 1,030 51.8 145.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hzlneami",
   "source_url": "https://omlx.ai/benchmarks/performance/hzlneami",
   "tg_tokens": null,
   "tps": 51.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-hzlneami-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 149.0 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1018.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 195k 1,018 50.1 149.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hzlneami",
   "source_url": "https://omlx.ai/benchmarks/performance/hzlneami",
   "tg_tokens": null,
   "tps": 50.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-q4w3s8j8-ctx1024",
   "model": "SuperQwen3.8-27b-abliterated",
   "notes": "MLX; oMLX community benchmark trace, single run at 1k context; peak memory 16.0 GB; TTFT 777 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1318.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "SuperQwen3.8-27b-abliterated-MLX-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k tokens 1,318 PP tok/s 55.1 TG tok/s 777 TTFT (ms) 16.0 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q4w3s8j8",
   "source_url": "https://omlx.ai/benchmarks/performance/q4w3s8j8",
   "tg_tokens": null,
   "tps": 55.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yke0u1ug-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1066.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,066 58.7 19.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yke0u1ug",
   "source_url": "https://omlx.ai/benchmarks/performance/yke0u1ug",
   "tg_tokens": null,
   "tps": 58.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-yke0u1ug-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1313.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,313 63.0 20.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yke0u1ug",
   "source_url": "https://omlx.ai/benchmarks/performance/yke0u1ug",
   "tg_tokens": null,
   "tps": 63.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yke0u1ug-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1334.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,334 57.5 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yke0u1ug",
   "source_url": "https://omlx.ai/benchmarks/performance/yke0u1ug",
   "tg_tokens": null,
   "tps": 57.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-psztr93l-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1118.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,118 58.2 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/psztr93l",
   "source_url": "https://omlx.ai/benchmarks/performance/psztr93l",
   "tg_tokens": null,
   "tps": 58.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-psztr93l-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1331.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,331 64.4 19.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/psztr93l",
   "source_url": "https://omlx.ai/benchmarks/performance/psztr93l",
   "tg_tokens": null,
   "tps": 64.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-psztr93l-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1321.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,321 69.8 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/psztr93l",
   "source_url": "https://omlx.ai/benchmarks/performance/psztr93l",
   "tg_tokens": null,
   "tps": 69.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-wg1z3mi4-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.5 GB at this context; duplicate traces of the same session: 0ltnb0m3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 245.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 245.7 41.2 32.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wg1z3mi4",
   "source_url": "https://omlx.ai/benchmarks/performance/wg1z3mi4",
   "tg_tokens": null,
   "tps": 41.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-wg1z3mi4-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.7 GB at this context; duplicate traces of the same session: 0ltnb0m3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 274.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 274.1 41.6 35.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wg1z3mi4",
   "source_url": "https://omlx.ai/benchmarks/performance/wg1z3mi4",
   "tg_tokens": null,
   "tps": 41.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-wg1z3mi4-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: 0ltnb0m3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 277.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 8k 277.3 37.8 35.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wg1z3mi4",
   "source_url": "https://omlx.ai/benchmarks/performance/wg1z3mi4",
   "tg_tokens": null,
   "tps": 37.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-wg1z3mi4-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.5 GB at this context; duplicate traces of the same session: 0ltnb0m3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 251.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k 251.3 37.0 31.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wg1z3mi4",
   "source_url": "https://omlx.ai/benchmarks/performance/wg1z3mi4",
   "tg_tokens": null,
   "tps": 37.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-wg1z3mi4-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.3 GB at this context; duplicate traces of the same session: 0ltnb0m3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 233.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 233.1 29.8 33.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wg1z3mi4",
   "source_url": "https://omlx.ai/benchmarks/performance/wg1z3mi4",
   "tg_tokens": null,
   "tps": 29.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tyw5830v-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: 18clxc08; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1004.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,004 58.1 71.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tyw5830v",
   "source_url": "https://omlx.ai/benchmarks/performance/tyw5830v",
   "tg_tokens": null,
   "tps": 58.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tyw5830v-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 71.2 GB at this context; duplicate traces of the same session: 18clxc08; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1924.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,924 62.5 71.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tyw5830v",
   "source_url": "https://omlx.ai/benchmarks/performance/tyw5830v",
   "tg_tokens": null,
   "tps": 62.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-fzoc69ia-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: 7jbivu99, 8ezii5v6, e76si6sh; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 676.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 1k 676.0 70.3 14.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fzoc69ia",
   "source_url": "https://omlx.ai/benchmarks/performance/fzoc69ia",
   "tg_tokens": null,
   "tps": 70.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-fzoc69ia-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: 7jbivu99, 8ezii5v6, e76si6sh; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 672.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 4k 672.4 67.1 14.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fzoc69ia",
   "source_url": "https://omlx.ai/benchmarks/performance/fzoc69ia",
   "tg_tokens": null,
   "tps": 67.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-fzoc69ia-ctx8192",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: 7jbivu99, 8ezii5v6, e76si6sh; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 662.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 8k 662.4 63.1 14.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fzoc69ia",
   "source_url": "https://omlx.ai/benchmarks/performance/fzoc69ia",
   "tg_tokens": null,
   "tps": 63.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-fzoc69ia-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 15.2 GB at this context; duplicate traces of the same session: 7jbivu99, 8ezii5v6, e76si6sh; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 632.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 16k 632.0 56.2 15.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fzoc69ia",
   "source_url": "https://omlx.ai/benchmarks/performance/fzoc69ia",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-fzoc69ia-ctx32768",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 7jbivu99, 8ezii5v6, e76si6sh; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 578.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 32k 578.4 46.9 16.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fzoc69ia",
   "source_url": "https://omlx.ai/benchmarks/performance/fzoc69ia",
   "tg_tokens": null,
   "tps": 46.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-fzoc69ia-ctx65536",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: 7jbivu99, 8ezii5v6, e76si6sh; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 493.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 64k 493.5 34.6 17.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fzoc69ia",
   "source_url": "https://omlx.ai/benchmarks/performance/fzoc69ia",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-dyshvu0n-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 5.3 GB at this context; duplicate traces of the same session: 407nk4ed; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 212.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 212.4 29.4 5.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dyshvu0n",
   "source_url": "https://omlx.ai/benchmarks/performance/dyshvu0n",
   "tg_tokens": null,
   "tps": 29.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-dyshvu0n-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 5.6 GB at this context; duplicate traces of the same session: 407nk4ed; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 334.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 334.9 26.9 5.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dyshvu0n",
   "source_url": "https://omlx.ai/benchmarks/performance/dyshvu0n",
   "tg_tokens": null,
   "tps": 26.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-qfxswvss-ctx1024",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 28cxtguv, abfhn6u1, c1veqwq1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 208.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 208.4 20.5 18.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qfxswvss",
   "source_url": "https://omlx.ai/benchmarks/performance/qfxswvss",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-qfxswvss-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 28cxtguv, abfhn6u1, c1veqwq1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 205.8 23.0 19.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qfxswvss",
   "source_url": "https://omlx.ai/benchmarks/performance/qfxswvss",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-qfxswvss-ctx8192",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 28cxtguv, abfhn6u1, c1veqwq1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 197.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 197.3 22.4 19.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qfxswvss",
   "source_url": "https://omlx.ai/benchmarks/performance/qfxswvss",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-qfxswvss-ctx16384",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: 28cxtguv, abfhn6u1, c1veqwq1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 193.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 193.0 21.5 21.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qfxswvss",
   "source_url": "https://omlx.ai/benchmarks/performance/qfxswvss",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-qfxswvss-ctx32768",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 28cxtguv, abfhn6u1, c1veqwq1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 178.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 178.0 19.8 23.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qfxswvss",
   "source_url": "https://omlx.ai/benchmarks/performance/qfxswvss",
   "tg_tokens": null,
   "tps": 19.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-qfxswvss-ctx65536",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: 28cxtguv, abfhn6u1, c1veqwq1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 155.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 155.1 17.0 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qfxswvss",
   "source_url": "https://omlx.ai/benchmarks/performance/qfxswvss",
   "tg_tokens": null,
   "tps": 17.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-dm5rvs5z-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: qm16i3ct; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3395.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,395 104.8 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dm5rvs5z",
   "source_url": "https://omlx.ai/benchmarks/performance/dm5rvs5z",
   "tg_tokens": null,
   "tps": 104.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-dm5rvs5z-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: qm16i3ct; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3023.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 3,023 92.4 20.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dm5rvs5z",
   "source_url": "https://omlx.ai/benchmarks/performance/dm5rvs5z",
   "tg_tokens": null,
   "tps": 92.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8n6mka8s-ctx1024",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 87.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 435.5,
   "provenance": "community",
   "quant": "2.4bit mixed",
   "quote": "DeepSeek-V4-Flash-0731-2.4bit-mixed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 1k 435.5 39.9 87.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8n6mka8s",
   "source_url": "https://omlx.ai/benchmarks/performance/8n6mka8s",
   "tg_tokens": null,
   "tps": 39.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8n6mka8s-ctx4096",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 88.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 714.9,
   "provenance": "community",
   "quant": "2.4bit mixed",
   "quote": "DeepSeek-V4-Flash-0731-2.4bit-mixed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 4k 714.9 40.9 88.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8n6mka8s",
   "source_url": "https://omlx.ai/benchmarks/performance/8n6mka8s",
   "tg_tokens": null,
   "tps": 40.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ffoh2978-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Huihui abliterated build; oMLX community benchmark trace, single run at 1k context; peak memory 23.8 GB; TTFT 5462 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 187.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.8-Flash-Next-abliterated-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k tokens 187.5 PP tok/s 31.1 TG tok/s 5462 TTFT (ms) 23.8 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ffoh2978",
   "source_url": "https://omlx.ai/benchmarks/performance/ffoh2978",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-n577x7av-ctx16384",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1452.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 1,452 91.7 34.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n577x7av",
   "source_url": "https://omlx.ai/benchmarks/performance/n577x7av",
   "tg_tokens": null,
   "tps": 91.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-n577x7av-ctx32768",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1254.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 1,254 88.1 34.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n577x7av",
   "source_url": "https://omlx.ai/benchmarks/performance/n577x7av",
   "tg_tokens": null,
   "tps": 88.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-is690f4l-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Nail build; oMLX community benchmark trace; peak memory 33.2 GB at this context; duplicate traces of the same session: gu3bbah9; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 592.2,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX-6bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 1k 592.2 44.6 33.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/is690f4l",
   "source_url": "https://omlx.ai/benchmarks/performance/is690f4l",
   "tg_tokens": null,
   "tps": 44.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-is690f4l-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Nail build; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: gu3bbah9; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 607.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX-6bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 4k 607.1 36.8 34.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/is690f4l",
   "source_url": "https://omlx.ai/benchmarks/performance/is690f4l",
   "tg_tokens": null,
   "tps": 36.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-7plb09ou-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: f542iagc; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 941.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 941.0 91.7 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7plb09ou",
   "source_url": "https://omlx.ai/benchmarks/performance/7plb09ou",
   "tg_tokens": null,
   "tps": 91.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-7plb09ou-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: f542iagc; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1077.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,077 87.2 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7plb09ou",
   "source_url": "https://omlx.ai/benchmarks/performance/7plb09ou",
   "tg_tokens": null,
   "tps": 87.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-7plb09ou-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: f542iagc; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 997.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 997.9 82.7 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7plb09ou",
   "source_url": "https://omlx.ai/benchmarks/performance/7plb09ou",
   "tg_tokens": null,
   "tps": 82.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-7plb09ou-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: f542iagc; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 904.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 904.2 75.8 21.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7plb09ou",
   "source_url": "https://omlx.ai/benchmarks/performance/7plb09ou",
   "tg_tokens": null,
   "tps": 75.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-grnsxxa2-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1646.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 1,646 69.5 19.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/grnsxxa2",
   "source_url": "https://omlx.ai/benchmarks/performance/grnsxxa2",
   "tg_tokens": null,
   "tps": 69.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-grnsxxa2-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1966.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,966 70.0 19.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/grnsxxa2",
   "source_url": "https://omlx.ai/benchmarks/performance/grnsxxa2",
   "tg_tokens": null,
   "tps": 70.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-grnsxxa2-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1937.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 8k 1,937 66.7 19.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/grnsxxa2",
   "source_url": "https://omlx.ai/benchmarks/performance/grnsxxa2",
   "tg_tokens": null,
   "tps": 66.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-grnsxxa2-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1802.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 1,802 61.5 20.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/grnsxxa2",
   "source_url": "https://omlx.ai/benchmarks/performance/grnsxxa2",
   "tg_tokens": null,
   "tps": 61.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-grnsxxa2-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1520.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 32k 1,520 51.8 20.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/grnsxxa2",
   "source_url": "https://omlx.ai/benchmarks/performance/grnsxxa2",
   "tg_tokens": null,
   "tps": 51.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-grnsxxa2-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1133.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 64k 1,133 38.0 21.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/grnsxxa2",
   "source_url": "https://omlx.ai/benchmarks/performance/grnsxxa2",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-knzhcmw3-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 7ienoz4v, 8jcvz5bh, m4niz9nb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 893.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 893.3 78.2 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/knzhcmw3",
   "source_url": "https://omlx.ai/benchmarks/performance/knzhcmw3",
   "tg_tokens": null,
   "tps": 78.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-knzhcmw3-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: 7ienoz4v, 8jcvz5bh, m4niz9nb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1024.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,024 79.4 23.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/knzhcmw3",
   "source_url": "https://omlx.ai/benchmarks/performance/knzhcmw3",
   "tg_tokens": null,
   "tps": 79.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-knzhcmw3-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: 7ienoz4v, 8jcvz5bh, m4niz9nb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 954.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 954.7 85.3 23.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/knzhcmw3",
   "source_url": "https://omlx.ai/benchmarks/performance/knzhcmw3",
   "tg_tokens": null,
   "tps": 85.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-knzhcmw3-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: 7ienoz4v, 8jcvz5bh, m4niz9nb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 837.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 837.7 77.1 24.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/knzhcmw3",
   "source_url": "https://omlx.ai/benchmarks/performance/knzhcmw3",
   "tg_tokens": null,
   "tps": 77.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-knzhcmw3-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: 7ienoz4v, 8jcvz5bh, m4niz9nb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 678.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 678.6 67.6 25.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/knzhcmw3",
   "source_url": "https://omlx.ai/benchmarks/performance/knzhcmw3",
   "tg_tokens": null,
   "tps": 67.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-knzhcmw3-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: 7ienoz4v, 8jcvz5bh, m4niz9nb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 491.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 491.9 43.5 26.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/knzhcmw3",
   "source_url": "https://omlx.ai/benchmarks/performance/knzhcmw3",
   "tg_tokens": null,
   "tps": 43.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-qepei3pf-ctx4096",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "MLX; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1178.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-MLX-5bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 5bit; 4k 1,178 41.0 20.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qepei3pf",
   "source_url": "https://omlx.ai/benchmarks/performance/qepei3pf",
   "tg_tokens": null,
   "tps": 41.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-qepei3pf-ctx16384",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "MLX; oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 797.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-MLX-5bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 5bit; 16k 797.0 27.9 21.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qepei3pf",
   "source_url": "https://omlx.ai/benchmarks/performance/qepei3pf",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-duri0ypz-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 761.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 761.1 107.2 35.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/duri0ypz",
   "source_url": "https://omlx.ai/benchmarks/performance/duri0ypz",
   "tg_tokens": null,
   "tps": 107.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-duri0ypz-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2396.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 2,396 105.6 36.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/duri0ypz",
   "source_url": "https://omlx.ai/benchmarks/performance/duri0ypz",
   "tg_tokens": null,
   "tps": 105.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-l9le3ol1-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: dssdp0o7, igbn28oo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3787.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 3,787 118.1 26.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l9le3ol1",
   "source_url": "https://omlx.ai/benchmarks/performance/l9le3ol1",
   "tg_tokens": null,
   "tps": 118.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-l9le3ol1-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: dssdp0o7, igbn28oo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5216.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 5,216 173.4 26.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l9le3ol1",
   "source_url": "https://omlx.ai/benchmarks/performance/l9le3ol1",
   "tg_tokens": null,
   "tps": 173.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-l9le3ol1-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: dssdp0o7, igbn28oo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5215.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 5,215 169.6 26.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l9le3ol1",
   "source_url": "https://omlx.ai/benchmarks/performance/l9le3ol1",
   "tg_tokens": null,
   "tps": 169.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-l9le3ol1-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.0 GB at this context; duplicate traces of the same session: dssdp0o7, igbn28oo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4892.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 4,892 183.3 27.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l9le3ol1",
   "source_url": "https://omlx.ai/benchmarks/performance/l9le3ol1",
   "tg_tokens": null,
   "tps": 183.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-l9le3ol1-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: dssdp0o7, igbn28oo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4084.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 4,084 239.0 27.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l9le3ol1",
   "source_url": "https://omlx.ai/benchmarks/performance/l9le3ol1",
   "tg_tokens": null,
   "tps": 239.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-l9le3ol1-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: dssdp0o7, igbn28oo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2991.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 2,991 130.2 28.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l9le3ol1",
   "source_url": "https://omlx.ai/benchmarks/performance/l9le3ol1",
   "tg_tokens": null,
   "tps": 130.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-l9le3ol1-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: dssdp0o7, igbn28oo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1992.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 1,992 101.7 29.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l9le3ol1",
   "source_url": "https://omlx.ai/benchmarks/performance/l9le3ol1",
   "tg_tokens": null,
   "tps": 101.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-re29w5v7-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 6yyueta2, bnqpqo0x, fqyyhltx, mxe2lqp5, uxjubqh3, yzxy2fw0; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1375.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,375 84.7 17.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/re29w5v7",
   "source_url": "https://omlx.ai/benchmarks/performance/re29w5v7",
   "tg_tokens": null,
   "tps": 84.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-re29w5v7-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: 6yyueta2, bnqpqo0x, fqyyhltx, mxe2lqp5, uxjubqh3, yzxy2fw0; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1624.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,624 137.9 18.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/re29w5v7",
   "source_url": "https://omlx.ai/benchmarks/performance/re29w5v7",
   "tg_tokens": null,
   "tps": 137.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-re29w5v7-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 6yyueta2, bnqpqo0x, fqyyhltx, mxe2lqp5, uxjubqh3, yzxy2fw0; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1619.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 1,619 61.7 19.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/re29w5v7",
   "source_url": "https://omlx.ai/benchmarks/performance/re29w5v7",
   "tg_tokens": null,
   "tps": 61.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-re29w5v7-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 6yyueta2, bnqpqo0x, fqyyhltx, mxe2lqp5, uxjubqh3, yzxy2fw0; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1566.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,566 73.8 19.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/re29w5v7",
   "source_url": "https://omlx.ai/benchmarks/performance/re29w5v7",
   "tg_tokens": null,
   "tps": 73.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-re29w5v7-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: 6yyueta2, bnqpqo0x, fqyyhltx, mxe2lqp5, uxjubqh3, yzxy2fw0; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1453.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,453 116.5 20.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/re29w5v7",
   "source_url": "https://omlx.ai/benchmarks/performance/re29w5v7",
   "tg_tokens": null,
   "tps": 116.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-re29w5v7-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: 6yyueta2, bnqpqo0x, fqyyhltx, mxe2lqp5, uxjubqh3, yzxy2fw0; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1255.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,255 77.1 23.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/re29w5v7",
   "source_url": "https://omlx.ai/benchmarks/performance/re29w5v7",
   "tg_tokens": null,
   "tps": 77.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-re29w5v7-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: 6yyueta2, bnqpqo0x, fqyyhltx, mxe2lqp5, uxjubqh3, yzxy2fw0; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 978.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 978.9 77.6 27.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/re29w5v7",
   "source_url": "https://omlx.ai/benchmarks/performance/re29w5v7",
   "tg_tokens": null,
   "tps": 77.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-re29w5v7-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 33.0 GB at this context; duplicate traces of the same session: 6yyueta2, bnqpqo0x, fqyyhltx, mxe2lqp5, uxjubqh3, yzxy2fw0; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 795.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 795.4 69.2 33.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/re29w5v7",
   "source_url": "https://omlx.ai/benchmarks/performance/re29w5v7",
   "tg_tokens": null,
   "tps": 69.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-js5lvb7m-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: 19vgdd18, wghm80x6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 939.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 939.2 32.8 17.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/js5lvb7m",
   "source_url": "https://omlx.ai/benchmarks/performance/js5lvb7m",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-js5lvb7m-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 19vgdd18, wghm80x6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 973.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 973.1 31.0 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/js5lvb7m",
   "source_url": "https://omlx.ai/benchmarks/performance/js5lvb7m",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-js5lvb7m-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 19vgdd18, wghm80x6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 798.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 798.8 25.7 20.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/js5lvb7m",
   "source_url": "https://omlx.ai/benchmarks/performance/js5lvb7m",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-js5lvb7m-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: 19vgdd18, wghm80x6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 637.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 637.0 21.6 24.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/js5lvb7m",
   "source_url": "https://omlx.ai/benchmarks/performance/js5lvb7m",
   "tg_tokens": null,
   "tps": 21.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-js5lvb7m-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.0 GB at this context; duplicate traces of the same session: 19vgdd18, wghm80x6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 515.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 195k 515.5 18.1 29.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/js5lvb7m",
   "source_url": "https://omlx.ai/benchmarks/performance/js5lvb7m",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-x3prmi86-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 820.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 820.7 18.8 28.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x3prmi86",
   "source_url": "https://omlx.ai/benchmarks/performance/x3prmi86",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-x3prmi86-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 836.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 836.7 18.7 29.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x3prmi86",
   "source_url": "https://omlx.ai/benchmarks/performance/x3prmi86",
   "tg_tokens": null,
   "tps": 18.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-cs39n1ht-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: ck8d3h72, gi320am8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 420.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 420.5 73.1 21.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cs39n1ht",
   "source_url": "https://omlx.ai/benchmarks/performance/cs39n1ht",
   "tg_tokens": null,
   "tps": 73.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-cs39n1ht-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: ck8d3h72, gi320am8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 559.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 559.9 111.1 25.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cs39n1ht",
   "source_url": "https://omlx.ai/benchmarks/performance/cs39n1ht",
   "tg_tokens": null,
   "tps": 111.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-cs39n1ht-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: ck8d3h72, gi320am8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 551.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 551.5 54.4 25.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cs39n1ht",
   "source_url": "https://omlx.ai/benchmarks/performance/cs39n1ht",
   "tg_tokens": null,
   "tps": 54.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-cs39n1ht-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: ck8d3h72, gi320am8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 526.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 526.8 76.3 26.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cs39n1ht",
   "source_url": "https://omlx.ai/benchmarks/performance/cs39n1ht",
   "tg_tokens": null,
   "tps": 76.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-n716phs3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 277.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 277.0 21.9 17.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n716phs3",
   "source_url": "https://omlx.ai/benchmarks/performance/n716phs3",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-n716phs3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 233.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 233.1 20.5 18.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n716phs3",
   "source_url": "https://omlx.ai/benchmarks/performance/n716phs3",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-n716phs3-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 235.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 8k 235.4 26.1 18.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n716phs3",
   "source_url": "https://omlx.ai/benchmarks/performance/n716phs3",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-n716phs3-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 222.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 222.4 23.0 19.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n716phs3",
   "source_url": "https://omlx.ai/benchmarks/performance/n716phs3",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-n716phs3-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 214.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 32k 214.9 25.3 20.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n716phs3",
   "source_url": "https://omlx.ai/benchmarks/performance/n716phs3",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-n716phs3-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 193.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 64k 193.3 17.9 23.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n716phs3",
   "source_url": "https://omlx.ai/benchmarks/performance/n716phs3",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kvg8ovtl-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: i8ref11h, p1os044e, pfhzn4ht, qsmyfdcx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 831.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 831.7 55.9 23.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kvg8ovtl",
   "source_url": "https://omlx.ai/benchmarks/performance/kvg8ovtl",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kvg8ovtl-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: i8ref11h, p1os044e, pfhzn4ht, qsmyfdcx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 886.6,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 886.6 53.0 24.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kvg8ovtl",
   "source_url": "https://omlx.ai/benchmarks/performance/kvg8ovtl",
   "tg_tokens": null,
   "tps": 53.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-kvg8ovtl-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: i8ref11h, p1os044e, pfhzn4ht, qsmyfdcx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 816.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 816.0 66.0 24.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kvg8ovtl",
   "source_url": "https://omlx.ai/benchmarks/performance/kvg8ovtl",
   "tg_tokens": null,
   "tps": 66.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kvg8ovtl-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.2 GB at this context; duplicate traces of the same session: i8ref11h, p1os044e, pfhzn4ht, qsmyfdcx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 783.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 783.9 75.3 25.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kvg8ovtl",
   "source_url": "https://omlx.ai/benchmarks/performance/kvg8ovtl",
   "tg_tokens": null,
   "tps": 75.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kvg8ovtl-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: i8ref11h, p1os044e, pfhzn4ht, qsmyfdcx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 676.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 676.4 113.5 26.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kvg8ovtl",
   "source_url": "https://omlx.ai/benchmarks/performance/kvg8ovtl",
   "tg_tokens": null,
   "tps": 113.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kvg8ovtl-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.7 GB at this context; duplicate traces of the same session: i8ref11h, p1os044e, pfhzn4ht, qsmyfdcx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 566.3,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 566.3 44.5 28.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kvg8ovtl",
   "source_url": "https://omlx.ai/benchmarks/performance/kvg8ovtl",
   "tg_tokens": null,
   "tps": 44.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-kvg8ovtl-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.4 GB at this context; duplicate traces of the same session: i8ref11h, p1os044e, pfhzn4ht, qsmyfdcx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 442.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 442.4 40.1 33.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kvg8ovtl",
   "source_url": "https://omlx.ai/benchmarks/performance/kvg8ovtl",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-p8461p1d-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: f05iixoj, sfck1s2e, t5e5p84z, v4z1qwxa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 899.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 899.1 67.5 24.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p8461p1d",
   "source_url": "https://omlx.ai/benchmarks/performance/p8461p1d",
   "tg_tokens": null,
   "tps": 67.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-p8461p1d-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.7 GB at this context; duplicate traces of the same session: f05iixoj, sfck1s2e, t5e5p84z, v4z1qwxa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1157.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 1,157 58.7 25.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p8461p1d",
   "source_url": "https://omlx.ai/benchmarks/performance/p8461p1d",
   "tg_tokens": null,
   "tps": 58.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-p8461p1d-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: f05iixoj, sfck1s2e, t5e5p84z, v4z1qwxa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1125.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 1,125 60.4 26.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p8461p1d",
   "source_url": "https://omlx.ai/benchmarks/performance/p8461p1d",
   "tg_tokens": null,
   "tps": 60.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-p8461p1d-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: f05iixoj, sfck1s2e, t5e5p84z, v4z1qwxa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1048.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 1,048 78.8 26.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p8461p1d",
   "source_url": "https://omlx.ai/benchmarks/performance/p8461p1d",
   "tg_tokens": null,
   "tps": 78.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-p8461p1d-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: f05iixoj, sfck1s2e, t5e5p84z, v4z1qwxa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 919.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 919.9 113.4 27.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p8461p1d",
   "source_url": "https://omlx.ai/benchmarks/performance/p8461p1d",
   "tg_tokens": null,
   "tps": 113.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-p8461p1d-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.0 GB at this context; duplicate traces of the same session: f05iixoj, sfck1s2e, t5e5p84z, v4z1qwxa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 754.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 754.1 39.1 30.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p8461p1d",
   "source_url": "https://omlx.ai/benchmarks/performance/p8461p1d",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-p8461p1d-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: f05iixoj, sfck1s2e, t5e5p84z, v4z1qwxa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 538.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 538.9 39.8 34.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p8461p1d",
   "source_url": "https://omlx.ai/benchmarks/performance/p8461p1d",
   "tg_tokens": null,
   "tps": 39.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pc3ghho4-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: 0d5mozkt, 5kitbe2f, dzuk1zp6, q6df6psy, td20xvnd, xznrsudm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 816.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 816.7 55.8 24.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pc3ghho4",
   "source_url": "https://omlx.ai/benchmarks/performance/pc3ghho4",
   "tg_tokens": null,
   "tps": 55.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pc3ghho4-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.7 GB at this context; duplicate traces of the same session: 0d5mozkt, 5kitbe2f, dzuk1zp6, q6df6psy, td20xvnd, xznrsudm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 868.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 868.4 47.8 25.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pc3ghho4",
   "source_url": "https://omlx.ai/benchmarks/performance/pc3ghho4",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pc3ghho4-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: 0d5mozkt, 5kitbe2f, dzuk1zp6, q6df6psy, td20xvnd, xznrsudm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 951.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 951.9 53.4 26.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pc3ghho4",
   "source_url": "https://omlx.ai/benchmarks/performance/pc3ghho4",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-pc3ghho4-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: 0d5mozkt, 5kitbe2f, dzuk1zp6, q6df6psy, td20xvnd, xznrsudm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 961.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 961.9 84.9 26.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pc3ghho4",
   "source_url": "https://omlx.ai/benchmarks/performance/pc3ghho4",
   "tg_tokens": null,
   "tps": 84.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pc3ghho4-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: 0d5mozkt, 5kitbe2f, dzuk1zp6, q6df6psy, td20xvnd, xznrsudm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 891.8,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 891.8 105.7 27.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pc3ghho4",
   "source_url": "https://omlx.ai/benchmarks/performance/pc3ghho4",
   "tg_tokens": null,
   "tps": 105.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pc3ghho4-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.0 GB at this context; duplicate traces of the same session: 0d5mozkt, 5kitbe2f, dzuk1zp6, q6df6psy, td20xvnd, xznrsudm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 740.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 740.2 43.7 30.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pc3ghho4",
   "source_url": "https://omlx.ai/benchmarks/performance/pc3ghho4",
   "tg_tokens": null,
   "tps": 43.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-pc3ghho4-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: 0d5mozkt, 5kitbe2f, dzuk1zp6, q6df6psy, td20xvnd, xznrsudm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 554.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 554.4 41.8 34.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pc3ghho4",
   "source_url": "https://omlx.ai/benchmarks/performance/pc3ghho4",
   "tg_tokens": null,
   "tps": 41.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-uy9qxpgy-ctx16384",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 41.5 GB; TTFT 13649 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1200.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k tokens 1,200 PP tok/s 39.7 TG tok/s 13649 TTFT (ms) 41.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uy9qxpgy",
   "source_url": "https://omlx.ai/benchmarks/performance/uy9qxpgy",
   "tg_tokens": null,
   "tps": 39.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-bq3ku0vf-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 403.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 1k 403.8 37.5 70.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bq3ku0vf",
   "source_url": "https://omlx.ai/benchmarks/performance/bq3ku0vf",
   "tg_tokens": null,
   "tps": 37.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-bq3ku0vf-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 483.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 4k 483.2 38.0 71.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bq3ku0vf",
   "source_url": "https://omlx.ai/benchmarks/performance/bq3ku0vf",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-bq3ku0vf-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.1 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 502.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 16k 502.5 32.1 72.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bq3ku0vf",
   "source_url": "https://omlx.ai/benchmarks/performance/bq3ku0vf",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-bq3ku0vf-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.1 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 474.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 64k 474.2 30.1 74.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bq3ku0vf",
   "source_url": "https://omlx.ai/benchmarks/performance/bq3ku0vf",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-bq3ku0vf-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.8 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 376.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 128k 376.4 27.8 78.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bq3ku0vf",
   "source_url": "https://omlx.ai/benchmarks/performance/bq3ku0vf",
   "tg_tokens": null,
   "tps": 27.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-05isnu7u-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1002.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,002 65.1 71.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/05isnu7u",
   "source_url": "https://omlx.ai/benchmarks/performance/05isnu7u",
   "tg_tokens": null,
   "tps": 65.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-05isnu7u-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1962.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,962 65.2 71.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/05isnu7u",
   "source_url": "https://omlx.ai/benchmarks/performance/05isnu7u",
   "tg_tokens": null,
   "tps": 65.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-0ltnb0m3-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 245.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 245.7 41.2 32.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ltnb0m3",
   "source_url": "https://omlx.ai/benchmarks/performance/0ltnb0m3",
   "tg_tokens": null,
   "tps": 41.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-0ltnb0m3-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 274.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 274.1 41.6 35.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ltnb0m3",
   "source_url": "https://omlx.ai/benchmarks/performance/0ltnb0m3",
   "tg_tokens": null,
   "tps": 41.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-0ltnb0m3-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 277.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 8k 277.3 37.8 35.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ltnb0m3",
   "source_url": "https://omlx.ai/benchmarks/performance/0ltnb0m3",
   "tg_tokens": null,
   "tps": 37.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-0ltnb0m3-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 251.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k 251.3 37.0 31.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ltnb0m3",
   "source_url": "https://omlx.ai/benchmarks/performance/0ltnb0m3",
   "tg_tokens": null,
   "tps": 37.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-0ltnb0m3-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 233.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 233.1 29.8 33.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ltnb0m3",
   "source_url": "https://omlx.ai/benchmarks/performance/0ltnb0m3",
   "tg_tokens": null,
   "tps": 29.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-18clxc08-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1004.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,004 58.1 71.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/18clxc08",
   "source_url": "https://omlx.ai/benchmarks/performance/18clxc08",
   "tg_tokens": null,
   "tps": 58.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-18clxc08-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 71.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1924.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,924 62.5 71.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/18clxc08",
   "source_url": "https://omlx.ai/benchmarks/performance/18clxc08",
   "tg_tokens": null,
   "tps": 62.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-8ezii5v6-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: 7jbivu99, e76si6sh; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 676.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 1k 676.0 70.3 14.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8ezii5v6",
   "source_url": "https://omlx.ai/benchmarks/performance/8ezii5v6",
   "tg_tokens": null,
   "tps": 70.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-8ezii5v6-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: 7jbivu99, e76si6sh; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 672.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 4k 672.4 67.1 14.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8ezii5v6",
   "source_url": "https://omlx.ai/benchmarks/performance/8ezii5v6",
   "tg_tokens": null,
   "tps": 67.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-8ezii5v6-ctx8192",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: 7jbivu99, e76si6sh; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 662.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 8k 662.4 63.1 14.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8ezii5v6",
   "source_url": "https://omlx.ai/benchmarks/performance/8ezii5v6",
   "tg_tokens": null,
   "tps": 63.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-8ezii5v6-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 15.2 GB at this context; duplicate traces of the same session: 7jbivu99, e76si6sh; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 632.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 16k 632.0 56.2 15.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8ezii5v6",
   "source_url": "https://omlx.ai/benchmarks/performance/8ezii5v6",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-8ezii5v6-ctx32768",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 7jbivu99, e76si6sh; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 578.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 32k 578.4 46.9 16.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8ezii5v6",
   "source_url": "https://omlx.ai/benchmarks/performance/8ezii5v6",
   "tg_tokens": null,
   "tps": 46.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-8ezii5v6-ctx65536",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: 7jbivu99, e76si6sh; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 493.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 64k 493.5 34.6 17.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8ezii5v6",
   "source_url": "https://omlx.ai/benchmarks/performance/8ezii5v6",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-407nk4ed-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 5.3 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 212.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 212.4 29.4 5.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/407nk4ed",
   "source_url": "https://omlx.ai/benchmarks/performance/407nk4ed",
   "tg_tokens": null,
   "tps": 29.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-407nk4ed-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 5.6 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 334.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 334.9 26.9 5.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/407nk4ed",
   "source_url": "https://omlx.ai/benchmarks/performance/407nk4ed",
   "tg_tokens": null,
   "tps": 26.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-c1veqwq1-ctx1024",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 28cxtguv, abfhn6u1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 208.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 208.4 20.5 18.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c1veqwq1",
   "source_url": "https://omlx.ai/benchmarks/performance/c1veqwq1",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-c1veqwq1-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 28cxtguv, abfhn6u1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 205.8 23.0 19.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c1veqwq1",
   "source_url": "https://omlx.ai/benchmarks/performance/c1veqwq1",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-c1veqwq1-ctx8192",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 28cxtguv, abfhn6u1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 197.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 197.3 22.4 19.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c1veqwq1",
   "source_url": "https://omlx.ai/benchmarks/performance/c1veqwq1",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-c1veqwq1-ctx16384",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: 28cxtguv, abfhn6u1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 193.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 193.0 21.5 21.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c1veqwq1",
   "source_url": "https://omlx.ai/benchmarks/performance/c1veqwq1",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-c1veqwq1-ctx32768",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 28cxtguv, abfhn6u1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 178.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 178.0 19.8 23.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c1veqwq1",
   "source_url": "https://omlx.ai/benchmarks/performance/c1veqwq1",
   "tg_tokens": null,
   "tps": 19.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-c1veqwq1-ctx65536",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: 28cxtguv, abfhn6u1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 155.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 155.1 17.0 27.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c1veqwq1",
   "source_url": "https://omlx.ai/benchmarks/performance/c1veqwq1",
   "tg_tokens": null,
   "tps": 17.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-m6ek9coc-ctx1024",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 46sqmiqt, edojcgp7, g9qikzc8, ifea3ajl, tsoiobp2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 207.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 207.0 28.5 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m6ek9coc",
   "source_url": "https://omlx.ai/benchmarks/performance/m6ek9coc",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-m6ek9coc-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 46sqmiqt, edojcgp7, g9qikzc8, ifea3ajl, tsoiobp2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 204.4 46.7 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m6ek9coc",
   "source_url": "https://omlx.ai/benchmarks/performance/m6ek9coc",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-m6ek9coc-ctx8192",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 46sqmiqt, edojcgp7, g9qikzc8, ifea3ajl, tsoiobp2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 193.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 193.4 30.7 21.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m6ek9coc",
   "source_url": "https://omlx.ai/benchmarks/performance/m6ek9coc",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-m6ek9coc-ctx16384",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: 46sqmiqt, edojcgp7, g9qikzc8, ifea3ajl, tsoiobp2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 193.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 193.4 39.1 23.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m6ek9coc",
   "source_url": "https://omlx.ai/benchmarks/performance/m6ek9coc",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-m6ek9coc-ctx32768",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: 46sqmiqt, edojcgp7, g9qikzc8, ifea3ajl, tsoiobp2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 179.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 179.7 26.2 25.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m6ek9coc",
   "source_url": "https://omlx.ai/benchmarks/performance/m6ek9coc",
   "tg_tokens": null,
   "tps": 26.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-m6ek9coc-ctx65536",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: 46sqmiqt, edojcgp7, g9qikzc8, ifea3ajl, tsoiobp2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 158.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 158.0 20.1 29.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m6ek9coc",
   "source_url": "https://omlx.ai/benchmarks/performance/m6ek9coc",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-czy3zruv-ctx4096",
   "model": "Gemma-4-31B",
   "notes": "oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: zufieqr8; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 162.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31b-it-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 162.4 17.7 19.0 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/czy3zruv",
   "source_url": "https://omlx.ai/benchmarks/performance/czy3zruv",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-czy3zruv-ctx16384",
   "model": "Gemma-4-31B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: zufieqr8; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 153.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31b-it-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 153.5 16.7 20.4 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/czy3zruv",
   "source_url": "https://omlx.ai/benchmarks/performance/czy3zruv",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qm16i3ct-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3395.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,395 104.8 20.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qm16i3ct",
   "source_url": "https://omlx.ai/benchmarks/performance/qm16i3ct",
   "tg_tokens": null,
   "tps": 104.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qm16i3ct-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3023.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 3,023 92.4 20.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qm16i3ct",
   "source_url": "https://omlx.ai/benchmarks/performance/qm16i3ct",
   "tg_tokens": null,
   "tps": 92.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-gu3bbah9-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Nail build; oMLX community benchmark trace; peak memory 33.2 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 592.2,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX-6bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 1k 592.2 44.6 33.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gu3bbah9",
   "source_url": "https://omlx.ai/benchmarks/performance/gu3bbah9",
   "tg_tokens": null,
   "tps": 44.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-gu3bbah9-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Nail build; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 607.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX-6bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 4k 607.1 36.8 34.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gu3bbah9",
   "source_url": "https://omlx.ai/benchmarks/performance/gu3bbah9",
   "tg_tokens": null,
   "tps": 36.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-f542iagc-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 941.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 941.0 91.7 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f542iagc",
   "source_url": "https://omlx.ai/benchmarks/performance/f542iagc",
   "tg_tokens": null,
   "tps": 91.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-f542iagc-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1077.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,077 87.2 21.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f542iagc",
   "source_url": "https://omlx.ai/benchmarks/performance/f542iagc",
   "tg_tokens": null,
   "tps": 87.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-f542iagc-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 997.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 997.9 82.7 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f542iagc",
   "source_url": "https://omlx.ai/benchmarks/performance/f542iagc",
   "tg_tokens": null,
   "tps": 82.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-f542iagc-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 904.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 904.2 75.8 21.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f542iagc",
   "source_url": "https://omlx.ai/benchmarks/performance/f542iagc",
   "tg_tokens": null,
   "tps": 75.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-m4niz9nb-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 7ienoz4v, 8jcvz5bh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 893.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 893.3 78.2 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m4niz9nb",
   "source_url": "https://omlx.ai/benchmarks/performance/m4niz9nb",
   "tg_tokens": null,
   "tps": 78.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-m4niz9nb-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: 7ienoz4v, 8jcvz5bh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1024.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,024 79.4 23.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m4niz9nb",
   "source_url": "https://omlx.ai/benchmarks/performance/m4niz9nb",
   "tg_tokens": null,
   "tps": 79.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-m4niz9nb-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: 7ienoz4v, 8jcvz5bh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 954.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 954.7 85.3 23.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m4niz9nb",
   "source_url": "https://omlx.ai/benchmarks/performance/m4niz9nb",
   "tg_tokens": null,
   "tps": 85.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-m4niz9nb-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: 7ienoz4v, 8jcvz5bh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 837.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 837.7 77.1 24.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m4niz9nb",
   "source_url": "https://omlx.ai/benchmarks/performance/m4niz9nb",
   "tg_tokens": null,
   "tps": 77.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-m4niz9nb-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: 7ienoz4v, 8jcvz5bh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 678.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 678.6 67.6 25.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m4niz9nb",
   "source_url": "https://omlx.ai/benchmarks/performance/m4niz9nb",
   "tg_tokens": null,
   "tps": 67.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-m4niz9nb-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: 7ienoz4v, 8jcvz5bh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 491.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 491.9 43.5 26.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m4niz9nb",
   "source_url": "https://omlx.ai/benchmarks/performance/m4niz9nb",
   "tg_tokens": null,
   "tps": 43.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1ef6b69d-ctx1024",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 9brouuzo, 9r3ze2f9, pga4c31c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4438.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 4,438 136.5 19.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ef6b69d",
   "source_url": "https://omlx.ai/benchmarks/performance/1ef6b69d",
   "tg_tokens": null,
   "tps": 136.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1ef6b69d-ctx32768",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 9brouuzo, 9r3ze2f9, pga4c31c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4546.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 4,546 113.3 20.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ef6b69d",
   "source_url": "https://omlx.ai/benchmarks/performance/1ef6b69d",
   "tg_tokens": null,
   "tps": 113.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1ef6b69d-ctx65536",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 9brouuzo, 9r3ze2f9, pga4c31c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3366.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 64k 3,366 99.7 21.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ef6b69d",
   "source_url": "https://omlx.ai/benchmarks/performance/1ef6b69d",
   "tg_tokens": null,
   "tps": 99.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1ef6b69d-ctx131072",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: 9brouuzo, 9r3ze2f9, pga4c31c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2252.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 128k 2,252 76.4 22.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ef6b69d",
   "source_url": "https://omlx.ai/benchmarks/performance/1ef6b69d",
   "tg_tokens": null,
   "tps": 76.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1ef6b69d-ctx199680",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: 9brouuzo, 9r3ze2f9, pga4c31c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1576.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 195k 1,576 58.2 23.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ef6b69d",
   "source_url": "https://omlx.ai/benchmarks/performance/1ef6b69d",
   "tg_tokens": null,
   "tps": 58.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-oa0dugz0-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 4hkxtv25; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 927.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 927.6 94.3 19.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oa0dugz0",
   "source_url": "https://omlx.ai/benchmarks/performance/oa0dugz0",
   "tg_tokens": null,
   "tps": 94.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-oa0dugz0-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 4hkxtv25; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1054.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 1,054 90.7 20.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oa0dugz0",
   "source_url": "https://omlx.ai/benchmarks/performance/oa0dugz0",
   "tg_tokens": null,
   "tps": 90.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-igbn28oo-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: dssdp0o7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3787.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 3,787 118.1 26.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/igbn28oo",
   "source_url": "https://omlx.ai/benchmarks/performance/igbn28oo",
   "tg_tokens": null,
   "tps": 118.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-igbn28oo-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: dssdp0o7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5216.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 5,216 173.4 26.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/igbn28oo",
   "source_url": "https://omlx.ai/benchmarks/performance/igbn28oo",
   "tg_tokens": null,
   "tps": 173.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-igbn28oo-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: dssdp0o7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5215.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 5,215 169.6 26.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/igbn28oo",
   "source_url": "https://omlx.ai/benchmarks/performance/igbn28oo",
   "tg_tokens": null,
   "tps": 169.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-igbn28oo-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.0 GB at this context; duplicate traces of the same session: dssdp0o7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4892.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 4,892 183.3 27.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/igbn28oo",
   "source_url": "https://omlx.ai/benchmarks/performance/igbn28oo",
   "tg_tokens": null,
   "tps": 183.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-igbn28oo-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: dssdp0o7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4084.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 4,084 239.0 27.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/igbn28oo",
   "source_url": "https://omlx.ai/benchmarks/performance/igbn28oo",
   "tg_tokens": null,
   "tps": 239.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-igbn28oo-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: dssdp0o7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2991.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 2,991 130.2 28.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/igbn28oo",
   "source_url": "https://omlx.ai/benchmarks/performance/igbn28oo",
   "tg_tokens": null,
   "tps": 130.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-igbn28oo-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: dssdp0o7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1992.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 1,992 101.7 29.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/igbn28oo",
   "source_url": "https://omlx.ai/benchmarks/performance/igbn28oo",
   "tg_tokens": null,
   "tps": 101.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-oxu23guc-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: 2w4uiyks, 3sqfr2pw, c8x51isy, dfnf3lrx, ewz4alio, lesilos7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2967.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 2,967 202.6 26.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oxu23guc",
   "source_url": "https://omlx.ai/benchmarks/performance/oxu23guc",
   "tg_tokens": null,
   "tps": 202.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-oxu23guc-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: 2w4uiyks, 3sqfr2pw, c8x51isy, dfnf3lrx, ewz4alio, lesilos7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4950.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 4,950 158.3 26.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oxu23guc",
   "source_url": "https://omlx.ai/benchmarks/performance/oxu23guc",
   "tg_tokens": null,
   "tps": 158.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-oxu23guc-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: 2w4uiyks, 3sqfr2pw, c8x51isy, dfnf3lrx, ewz4alio, lesilos7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5350.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 5,350 180.0 26.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oxu23guc",
   "source_url": "https://omlx.ai/benchmarks/performance/oxu23guc",
   "tg_tokens": null,
   "tps": 180.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-oxu23guc-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.0 GB at this context; duplicate traces of the same session: 2w4uiyks, 3sqfr2pw, c8x51isy, dfnf3lrx, ewz4alio, lesilos7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5012.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 5,012 211.3 27.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oxu23guc",
   "source_url": "https://omlx.ai/benchmarks/performance/oxu23guc",
   "tg_tokens": null,
   "tps": 211.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-oxu23guc-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: 2w4uiyks, 3sqfr2pw, c8x51isy, dfnf3lrx, ewz4alio, lesilos7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4144.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 4,144 229.4 27.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oxu23guc",
   "source_url": "https://omlx.ai/benchmarks/performance/oxu23guc",
   "tg_tokens": null,
   "tps": 229.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-oxu23guc-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: 2w4uiyks, 3sqfr2pw, c8x51isy, dfnf3lrx, ewz4alio, lesilos7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3088.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 3,088 144.9 28.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oxu23guc",
   "source_url": "https://omlx.ai/benchmarks/performance/oxu23guc",
   "tg_tokens": null,
   "tps": 144.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-oxu23guc-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: 2w4uiyks, 3sqfr2pw, c8x51isy, dfnf3lrx, ewz4alio, lesilos7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2110.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 2,110 109.8 29.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oxu23guc",
   "source_url": "https://omlx.ai/benchmarks/performance/oxu23guc",
   "tg_tokens": null,
   "tps": 109.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1owtu29m-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: f31gykzy, mfdxrpgr, tspkjmul; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 432.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 432.1 20.2 18.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1owtu29m",
   "source_url": "https://omlx.ai/benchmarks/performance/1owtu29m",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1owtu29m-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: f31gykzy, mfdxrpgr, tspkjmul; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 428.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 428.3 26.1 19.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1owtu29m",
   "source_url": "https://omlx.ai/benchmarks/performance/1owtu29m",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1owtu29m-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: f31gykzy, mfdxrpgr, tspkjmul; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 404.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 404.3 24.0 22.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1owtu29m",
   "source_url": "https://omlx.ai/benchmarks/performance/1owtu29m",
   "tg_tokens": null,
   "tps": 24.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1owtu29m-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: f31gykzy, mfdxrpgr, tspkjmul; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 360.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 360.5 23.4 26.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1owtu29m",
   "source_url": "https://omlx.ai/benchmarks/performance/1owtu29m",
   "tg_tokens": null,
   "tps": 23.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-mxe2lqp5-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 6yyueta2, bnqpqo0x, fqyyhltx, uxjubqh3, yzxy2fw0; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1375.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,375 84.7 17.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mxe2lqp5",
   "source_url": "https://omlx.ai/benchmarks/performance/mxe2lqp5",
   "tg_tokens": null,
   "tps": 84.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-mxe2lqp5-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: 6yyueta2, bnqpqo0x, fqyyhltx, uxjubqh3, yzxy2fw0; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1624.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,624 137.9 18.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mxe2lqp5",
   "source_url": "https://omlx.ai/benchmarks/performance/mxe2lqp5",
   "tg_tokens": null,
   "tps": 137.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-mxe2lqp5-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 6yyueta2, bnqpqo0x, fqyyhltx, uxjubqh3, yzxy2fw0; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1619.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 1,619 61.7 19.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mxe2lqp5",
   "source_url": "https://omlx.ai/benchmarks/performance/mxe2lqp5",
   "tg_tokens": null,
   "tps": 61.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-mxe2lqp5-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 6yyueta2, bnqpqo0x, fqyyhltx, uxjubqh3, yzxy2fw0; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1566.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,566 73.8 19.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mxe2lqp5",
   "source_url": "https://omlx.ai/benchmarks/performance/mxe2lqp5",
   "tg_tokens": null,
   "tps": 73.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-mxe2lqp5-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: 6yyueta2, bnqpqo0x, fqyyhltx, uxjubqh3, yzxy2fw0; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1453.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,453 116.5 20.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mxe2lqp5",
   "source_url": "https://omlx.ai/benchmarks/performance/mxe2lqp5",
   "tg_tokens": null,
   "tps": 116.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-mxe2lqp5-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: 6yyueta2, bnqpqo0x, fqyyhltx, uxjubqh3, yzxy2fw0; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1255.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,255 77.1 23.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mxe2lqp5",
   "source_url": "https://omlx.ai/benchmarks/performance/mxe2lqp5",
   "tg_tokens": null,
   "tps": 77.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-mxe2lqp5-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: 6yyueta2, bnqpqo0x, fqyyhltx, uxjubqh3, yzxy2fw0; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 978.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 978.9 77.6 27.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mxe2lqp5",
   "source_url": "https://omlx.ai/benchmarks/performance/mxe2lqp5",
   "tg_tokens": null,
   "tps": 77.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-mxe2lqp5-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 33.0 GB at this context; duplicate traces of the same session: 6yyueta2, bnqpqo0x, fqyyhltx, uxjubqh3, yzxy2fw0; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 795.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 795.4 69.2 33.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mxe2lqp5",
   "source_url": "https://omlx.ai/benchmarks/performance/mxe2lqp5",
   "tg_tokens": null,
   "tps": 69.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-7izwnpuf-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: q987p750; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1377.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,377 79.3 17.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7izwnpuf",
   "source_url": "https://omlx.ai/benchmarks/performance/7izwnpuf",
   "tg_tokens": null,
   "tps": 79.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-7izwnpuf-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: q987p750; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1621.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,621 100.6 18.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7izwnpuf",
   "source_url": "https://omlx.ai/benchmarks/performance/7izwnpuf",
   "tg_tokens": null,
   "tps": 100.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wghm80x6-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: 19vgdd18; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 939.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 939.2 32.8 17.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wghm80x6",
   "source_url": "https://omlx.ai/benchmarks/performance/wghm80x6",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wghm80x6-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 19vgdd18; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 973.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 973.1 31.0 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wghm80x6",
   "source_url": "https://omlx.ai/benchmarks/performance/wghm80x6",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wghm80x6-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 19vgdd18; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 798.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 798.8 25.7 20.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wghm80x6",
   "source_url": "https://omlx.ai/benchmarks/performance/wghm80x6",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wghm80x6-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: 19vgdd18; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 637.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 637.0 21.6 24.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wghm80x6",
   "source_url": "https://omlx.ai/benchmarks/performance/wghm80x6",
   "tg_tokens": null,
   "tps": 21.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wghm80x6-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.0 GB at this context; duplicate traces of the same session: 19vgdd18; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 515.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 195k 515.5 18.1 29.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wghm80x6",
   "source_url": "https://omlx.ai/benchmarks/performance/wghm80x6",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-6ihtl45n-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "unsloth build; oMLX community benchmark trace, single run at 4k context; peak memory 27.4 GB; TTFT 18797 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 217.9,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-27B-NVFP4-unsloth on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 nvfp4; 4k tokens 217.9 PP tok/s 27.5 TG tok/s 18797 TTFT (ms) 27.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ihtl45n",
   "source_url": "https://omlx.ai/benchmarks/performance/6ihtl45n",
   "tg_tokens": null,
   "tps": 27.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-rfbso3y8-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "unsloth build; oMLX community benchmark trace, single run at 4k context; peak memory 27.4 GB; TTFT 21793 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 188.0,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-27B-NVFP4-unsloth on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 nvfp4; 4k tokens 188.0 PP tok/s 36.5 TG tok/s 21793 TTFT (ms) 27.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rfbso3y8",
   "source_url": "https://omlx.ai/benchmarks/performance/rfbso3y8",
   "tg_tokens": null,
   "tps": 36.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-mn4uy0ei-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "unsloth build; oMLX community benchmark trace, single run at 4k context; peak memory 27.4 GB; TTFT 20468 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 200.1,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-27B-NVFP4-unsloth on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 nvfp4; 4k tokens 200.1 PP tok/s 33.4 TG tok/s 20468 TTFT (ms) 27.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mn4uy0ei",
   "source_url": "https://omlx.ai/benchmarks/performance/mn4uy0ei",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ck8d3h72-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: e7ax1svx, gi320am8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 420.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 420.5 73.1 21.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ck8d3h72",
   "source_url": "https://omlx.ai/benchmarks/performance/ck8d3h72",
   "tg_tokens": null,
   "tps": 73.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ck8d3h72-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: e7ax1svx, gi320am8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 559.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 559.9 111.1 25.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ck8d3h72",
   "source_url": "https://omlx.ai/benchmarks/performance/ck8d3h72",
   "tg_tokens": null,
   "tps": 111.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ck8d3h72-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: e7ax1svx, gi320am8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 551.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 551.5 54.4 25.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ck8d3h72",
   "source_url": "https://omlx.ai/benchmarks/performance/ck8d3h72",
   "tg_tokens": null,
   "tps": 54.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ck8d3h72-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: e7ax1svx, gi320am8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 526.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 526.8 76.3 26.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ck8d3h72",
   "source_url": "https://omlx.ai/benchmarks/performance/ck8d3h72",
   "tg_tokens": null,
   "tps": 76.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-sfck1s2e-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: f05iixoj, t5e5p84z, v4z1qwxa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 899.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 899.1 67.5 24.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sfck1s2e",
   "source_url": "https://omlx.ai/benchmarks/performance/sfck1s2e",
   "tg_tokens": null,
   "tps": 67.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-sfck1s2e-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.7 GB at this context; duplicate traces of the same session: f05iixoj, t5e5p84z, v4z1qwxa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1157.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 1,157 58.7 25.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sfck1s2e",
   "source_url": "https://omlx.ai/benchmarks/performance/sfck1s2e",
   "tg_tokens": null,
   "tps": 58.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-sfck1s2e-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: f05iixoj, t5e5p84z, v4z1qwxa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1125.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 1,125 60.4 26.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sfck1s2e",
   "source_url": "https://omlx.ai/benchmarks/performance/sfck1s2e",
   "tg_tokens": null,
   "tps": 60.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-sfck1s2e-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: f05iixoj, t5e5p84z, v4z1qwxa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1048.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 1,048 78.8 26.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sfck1s2e",
   "source_url": "https://omlx.ai/benchmarks/performance/sfck1s2e",
   "tg_tokens": null,
   "tps": 78.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-sfck1s2e-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: f05iixoj, t5e5p84z, v4z1qwxa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 919.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 919.9 113.4 27.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sfck1s2e",
   "source_url": "https://omlx.ai/benchmarks/performance/sfck1s2e",
   "tg_tokens": null,
   "tps": 113.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-sfck1s2e-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.0 GB at this context; duplicate traces of the same session: f05iixoj, t5e5p84z, v4z1qwxa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 754.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 754.1 39.1 30.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sfck1s2e",
   "source_url": "https://omlx.ai/benchmarks/performance/sfck1s2e",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-sfck1s2e-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: f05iixoj, t5e5p84z, v4z1qwxa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 538.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 538.9 39.8 34.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sfck1s2e",
   "source_url": "https://omlx.ai/benchmarks/performance/sfck1s2e",
   "tg_tokens": null,
   "tps": 39.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qsmyfdcx-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: i8ref11h, p1os044e, pfhzn4ht; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 831.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 831.7 55.9 23.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qsmyfdcx",
   "source_url": "https://omlx.ai/benchmarks/performance/qsmyfdcx",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qsmyfdcx-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: i8ref11h, p1os044e, pfhzn4ht; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 886.6,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 886.6 53.0 24.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qsmyfdcx",
   "source_url": "https://omlx.ai/benchmarks/performance/qsmyfdcx",
   "tg_tokens": null,
   "tps": 53.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qsmyfdcx-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: i8ref11h, p1os044e, pfhzn4ht; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 816.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 816.0 66.0 24.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qsmyfdcx",
   "source_url": "https://omlx.ai/benchmarks/performance/qsmyfdcx",
   "tg_tokens": null,
   "tps": 66.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qsmyfdcx-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.2 GB at this context; duplicate traces of the same session: i8ref11h, p1os044e, pfhzn4ht; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 783.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 783.9 75.3 25.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qsmyfdcx",
   "source_url": "https://omlx.ai/benchmarks/performance/qsmyfdcx",
   "tg_tokens": null,
   "tps": 75.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qsmyfdcx-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: i8ref11h, p1os044e, pfhzn4ht; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 676.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 676.4 113.5 26.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qsmyfdcx",
   "source_url": "https://omlx.ai/benchmarks/performance/qsmyfdcx",
   "tg_tokens": null,
   "tps": 113.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qsmyfdcx-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.7 GB at this context; duplicate traces of the same session: i8ref11h, p1os044e, pfhzn4ht; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 566.3,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 566.3 44.5 28.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qsmyfdcx",
   "source_url": "https://omlx.ai/benchmarks/performance/qsmyfdcx",
   "tg_tokens": null,
   "tps": 44.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qsmyfdcx-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.4 GB at this context; duplicate traces of the same session: i8ref11h, p1os044e, pfhzn4ht; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 442.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 442.4 40.1 33.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qsmyfdcx",
   "source_url": "https://omlx.ai/benchmarks/performance/qsmyfdcx",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5kitbe2f-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: 0d5mozkt, dzuk1zp6, q6df6psy, td20xvnd, xznrsudm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 816.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 816.7 55.8 24.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5kitbe2f",
   "source_url": "https://omlx.ai/benchmarks/performance/5kitbe2f",
   "tg_tokens": null,
   "tps": 55.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5kitbe2f-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.7 GB at this context; duplicate traces of the same session: 0d5mozkt, dzuk1zp6, q6df6psy, td20xvnd, xznrsudm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 868.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 868.4 47.8 25.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5kitbe2f",
   "source_url": "https://omlx.ai/benchmarks/performance/5kitbe2f",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5kitbe2f-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: 0d5mozkt, dzuk1zp6, q6df6psy, td20xvnd, xznrsudm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 951.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 951.9 53.4 26.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5kitbe2f",
   "source_url": "https://omlx.ai/benchmarks/performance/5kitbe2f",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5kitbe2f-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: 0d5mozkt, dzuk1zp6, q6df6psy, td20xvnd, xznrsudm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 961.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 961.9 84.9 26.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5kitbe2f",
   "source_url": "https://omlx.ai/benchmarks/performance/5kitbe2f",
   "tg_tokens": null,
   "tps": 84.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5kitbe2f-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: 0d5mozkt, dzuk1zp6, q6df6psy, td20xvnd, xznrsudm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 891.8,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 891.8 105.7 27.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5kitbe2f",
   "source_url": "https://omlx.ai/benchmarks/performance/5kitbe2f",
   "tg_tokens": null,
   "tps": 105.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5kitbe2f-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.0 GB at this context; duplicate traces of the same session: 0d5mozkt, dzuk1zp6, q6df6psy, td20xvnd, xznrsudm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 740.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 740.2 43.7 30.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5kitbe2f",
   "source_url": "https://omlx.ai/benchmarks/performance/5kitbe2f",
   "tg_tokens": null,
   "tps": 43.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5kitbe2f-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: 0d5mozkt, dzuk1zp6, q6df6psy, td20xvnd, xznrsudm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 554.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 554.4 41.8 34.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5kitbe2f",
   "source_url": "https://omlx.ai/benchmarks/performance/5kitbe2f",
   "tg_tokens": null,
   "tps": 41.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-9zrpqloq-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: eh5pdjbl, huise6v0, ksz9bfj4; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 403.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 1k 403.8 37.5 70.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9zrpqloq",
   "source_url": "https://omlx.ai/benchmarks/performance/9zrpqloq",
   "tg_tokens": null,
   "tps": 37.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-9zrpqloq-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: eh5pdjbl, huise6v0, ksz9bfj4; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 483.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 4k 483.2 38.0 71.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9zrpqloq",
   "source_url": "https://omlx.ai/benchmarks/performance/9zrpqloq",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-9zrpqloq-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.1 GB at this context; duplicate traces of the same session: eh5pdjbl, huise6v0, ksz9bfj4; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 502.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 16k 502.5 32.1 72.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9zrpqloq",
   "source_url": "https://omlx.ai/benchmarks/performance/9zrpqloq",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-9zrpqloq-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.1 GB at this context; duplicate traces of the same session: eh5pdjbl, huise6v0, ksz9bfj4; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 474.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 64k 474.2 30.1 74.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9zrpqloq",
   "source_url": "https://omlx.ai/benchmarks/performance/9zrpqloq",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-9zrpqloq-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.8 GB at this context; duplicate traces of the same session: eh5pdjbl, huise6v0, ksz9bfj4; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 376.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 128k 376.4 27.8 78.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9zrpqloq",
   "source_url": "https://omlx.ai/benchmarks/performance/9zrpqloq",
   "tg_tokens": null,
   "tps": 27.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-u8cy17p7-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace, single run at 64k context; peak memory 145.6 GB; TTFT 65353 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1003.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 64k tokens 1,003 PP tok/s 65.6 TG tok/s 65353 TTFT (ms) 145.6 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u8cy17p7",
   "source_url": "https://omlx.ai/benchmarks/performance/u8cy17p7",
   "tg_tokens": null,
   "tps": 65.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-e76si6sh-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: 7jbivu99; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 676.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 1k 676.0 70.3 14.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e76si6sh",
   "source_url": "https://omlx.ai/benchmarks/performance/e76si6sh",
   "tg_tokens": null,
   "tps": 70.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-e76si6sh-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: 7jbivu99; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 672.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 4k 672.4 67.1 14.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e76si6sh",
   "source_url": "https://omlx.ai/benchmarks/performance/e76si6sh",
   "tg_tokens": null,
   "tps": 67.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-e76si6sh-ctx8192",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: 7jbivu99; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 662.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 8k 662.4 63.1 14.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e76si6sh",
   "source_url": "https://omlx.ai/benchmarks/performance/e76si6sh",
   "tg_tokens": null,
   "tps": 63.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-e76si6sh-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 15.2 GB at this context; duplicate traces of the same session: 7jbivu99; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 632.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 16k 632.0 56.2 15.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e76si6sh",
   "source_url": "https://omlx.ai/benchmarks/performance/e76si6sh",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-e76si6sh-ctx32768",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 7jbivu99; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 578.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 32k 578.4 46.9 16.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e76si6sh",
   "source_url": "https://omlx.ai/benchmarks/performance/e76si6sh",
   "tg_tokens": null,
   "tps": 46.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-e76si6sh-ctx65536",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: 7jbivu99; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 493.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 64k 493.5 34.6 17.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e76si6sh",
   "source_url": "https://omlx.ai/benchmarks/performance/e76si6sh",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-28cxtguv-ctx1024",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: abfhn6u1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 208.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 208.4 20.5 18.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/28cxtguv",
   "source_url": "https://omlx.ai/benchmarks/performance/28cxtguv",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-28cxtguv-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: abfhn6u1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 205.8 23.0 19.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/28cxtguv",
   "source_url": "https://omlx.ai/benchmarks/performance/28cxtguv",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-28cxtguv-ctx8192",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: abfhn6u1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 197.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 197.3 22.4 19.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/28cxtguv",
   "source_url": "https://omlx.ai/benchmarks/performance/28cxtguv",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-28cxtguv-ctx16384",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: abfhn6u1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 193.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 193.0 21.5 21.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/28cxtguv",
   "source_url": "https://omlx.ai/benchmarks/performance/28cxtguv",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-28cxtguv-ctx32768",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: abfhn6u1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 178.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 178.0 19.8 23.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/28cxtguv",
   "source_url": "https://omlx.ai/benchmarks/performance/28cxtguv",
   "tg_tokens": null,
   "tps": 19.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-28cxtguv-ctx65536",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: abfhn6u1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 155.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 155.1 17.0 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/28cxtguv",
   "source_url": "https://omlx.ai/benchmarks/performance/28cxtguv",
   "tg_tokens": null,
   "tps": 17.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ifea3ajl-ctx1024",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 46sqmiqt, edojcgp7, g9qikzc8, tsoiobp2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 207.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 207.0 28.5 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ifea3ajl",
   "source_url": "https://omlx.ai/benchmarks/performance/ifea3ajl",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ifea3ajl-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 46sqmiqt, edojcgp7, g9qikzc8, tsoiobp2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 204.4 46.7 21.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ifea3ajl",
   "source_url": "https://omlx.ai/benchmarks/performance/ifea3ajl",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ifea3ajl-ctx8192",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 46sqmiqt, edojcgp7, g9qikzc8, tsoiobp2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 193.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 193.4 30.7 21.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ifea3ajl",
   "source_url": "https://omlx.ai/benchmarks/performance/ifea3ajl",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ifea3ajl-ctx16384",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: 46sqmiqt, edojcgp7, g9qikzc8, tsoiobp2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 193.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 193.4 39.1 23.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ifea3ajl",
   "source_url": "https://omlx.ai/benchmarks/performance/ifea3ajl",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ifea3ajl-ctx32768",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: 46sqmiqt, edojcgp7, g9qikzc8, tsoiobp2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 179.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 179.7 26.2 25.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ifea3ajl",
   "source_url": "https://omlx.ai/benchmarks/performance/ifea3ajl",
   "tg_tokens": null,
   "tps": 26.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ifea3ajl-ctx65536",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: 46sqmiqt, edojcgp7, g9qikzc8, tsoiobp2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 158.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 158.0 20.1 29.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ifea3ajl",
   "source_url": "https://omlx.ai/benchmarks/performance/ifea3ajl",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-zufieqr8-ctx4096",
   "model": "Gemma-4-31B",
   "notes": "oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 162.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31b-it-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 162.4 17.7 19.0 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zufieqr8",
   "source_url": "https://omlx.ai/benchmarks/performance/zufieqr8",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-zufieqr8-ctx16384",
   "model": "Gemma-4-31B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 153.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31b-it-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 153.5 16.7 20.4 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zufieqr8",
   "source_url": "https://omlx.ai/benchmarks/performance/zufieqr8",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-aby8t9gs-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 6.0 GB at this context; duplicate traces of the same session: cd2nlpsm; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1559.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E4B-it-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 1,559 26.8 6.0 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aby8t9gs",
   "source_url": "https://omlx.ai/benchmarks/performance/aby8t9gs",
   "tg_tokens": null,
   "tps": 26.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-aby8t9gs-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 6.1 GB at this context; duplicate traces of the same session: cd2nlpsm; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1760.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E4B-it-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 1,760 26.0 6.1 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aby8t9gs",
   "source_url": "https://omlx.ai/benchmarks/performance/aby8t9gs",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-7ienoz4v-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 8jcvz5bh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 893.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 893.3 78.2 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ienoz4v",
   "source_url": "https://omlx.ai/benchmarks/performance/7ienoz4v",
   "tg_tokens": null,
   "tps": 78.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-7ienoz4v-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: 8jcvz5bh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1024.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,024 79.4 23.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ienoz4v",
   "source_url": "https://omlx.ai/benchmarks/performance/7ienoz4v",
   "tg_tokens": null,
   "tps": 79.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-7ienoz4v-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: 8jcvz5bh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 954.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 954.7 85.3 23.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ienoz4v",
   "source_url": "https://omlx.ai/benchmarks/performance/7ienoz4v",
   "tg_tokens": null,
   "tps": 85.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-7ienoz4v-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: 8jcvz5bh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 837.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 837.7 77.1 24.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ienoz4v",
   "source_url": "https://omlx.ai/benchmarks/performance/7ienoz4v",
   "tg_tokens": null,
   "tps": 77.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-7ienoz4v-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: 8jcvz5bh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 678.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 678.6 67.6 25.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ienoz4v",
   "source_url": "https://omlx.ai/benchmarks/performance/7ienoz4v",
   "tg_tokens": null,
   "tps": 67.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-7ienoz4v-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: 8jcvz5bh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 491.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 491.9 43.5 26.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ienoz4v",
   "source_url": "https://omlx.ai/benchmarks/performance/7ienoz4v",
   "tg_tokens": null,
   "tps": 43.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-q3aw5is6-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 36.8 GB; TTFT 8159 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4016.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k tokens 4,016 PP tok/s 89.4 TG tok/s 8159 TTFT (ms) 36.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q3aw5is6",
   "source_url": "https://omlx.ai/benchmarks/performance/q3aw5is6",
   "tg_tokens": null,
   "tps": 89.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pga4c31c-ctx1024",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 9brouuzo, 9r3ze2f9, mqw7bjz2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4438.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 4,438 136.5 19.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pga4c31c",
   "source_url": "https://omlx.ai/benchmarks/performance/pga4c31c",
   "tg_tokens": null,
   "tps": 136.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pga4c31c-ctx32768",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 9brouuzo, 9r3ze2f9, mqw7bjz2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4546.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 4,546 113.3 20.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pga4c31c",
   "source_url": "https://omlx.ai/benchmarks/performance/pga4c31c",
   "tg_tokens": null,
   "tps": 113.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pga4c31c-ctx65536",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 9brouuzo, 9r3ze2f9, mqw7bjz2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3366.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 64k 3,366 99.7 21.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pga4c31c",
   "source_url": "https://omlx.ai/benchmarks/performance/pga4c31c",
   "tg_tokens": null,
   "tps": 99.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pga4c31c-ctx131072",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: 9brouuzo, 9r3ze2f9, mqw7bjz2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2252.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 128k 2,252 76.4 22.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pga4c31c",
   "source_url": "https://omlx.ai/benchmarks/performance/pga4c31c",
   "tg_tokens": null,
   "tps": 76.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-pga4c31c-ctx199680",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: 9brouuzo, 9r3ze2f9, mqw7bjz2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1576.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 195k 1,576 58.2 23.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pga4c31c",
   "source_url": "https://omlx.ai/benchmarks/performance/pga4c31c",
   "tg_tokens": null,
   "tps": 58.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lcm78unk-ctx1024",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: ae4gu2lz, ebfmytxb, llmaeu4o, o9z0xrl4, tj22e7jo, xzacwcec, z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4163.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 4,163 192.7 19.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lcm78unk",
   "source_url": "https://omlx.ai/benchmarks/performance/lcm78unk",
   "tg_tokens": null,
   "tps": 192.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lcm78unk-ctx4096",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: ae4gu2lz, ebfmytxb, llmaeu4o, o9z0xrl4, tj22e7jo, xzacwcec, z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5736.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 5,736 159.3 20.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lcm78unk",
   "source_url": "https://omlx.ai/benchmarks/performance/lcm78unk",
   "tg_tokens": null,
   "tps": 159.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lcm78unk-ctx8192",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: ae4gu2lz, ebfmytxb, llmaeu4o, o9z0xrl4, tj22e7jo, xzacwcec, z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6112.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 8k 6,112 203.0 20.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lcm78unk",
   "source_url": "https://omlx.ai/benchmarks/performance/lcm78unk",
   "tg_tokens": null,
   "tps": 203.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lcm78unk-ctx16384",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: ae4gu2lz, ebfmytxb, llmaeu4o, o9z0xrl4, tj22e7jo, xzacwcec, z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5607.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 5,607 195.0 20.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lcm78unk",
   "source_url": "https://omlx.ai/benchmarks/performance/lcm78unk",
   "tg_tokens": null,
   "tps": 195.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lcm78unk-ctx32768",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: ae4gu2lz, ebfmytxb, llmaeu4o, o9z0xrl4, tj22e7jo, xzacwcec, z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4807.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 4,807 170.8 21.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lcm78unk",
   "source_url": "https://omlx.ai/benchmarks/performance/lcm78unk",
   "tg_tokens": null,
   "tps": 170.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lcm78unk-ctx65536",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: ae4gu2lz, ebfmytxb, llmaeu4o, o9z0xrl4, tj22e7jo, xzacwcec, z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3580.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 64k 3,580 126.1 22.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lcm78unk",
   "source_url": "https://omlx.ai/benchmarks/performance/lcm78unk",
   "tg_tokens": null,
   "tps": 126.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lcm78unk-ctx131072",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: ae4gu2lz, ebfmytxb, llmaeu4o, o9z0xrl4, tj22e7jo, xzacwcec, z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2310.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 128k 2,310 87.1 23.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lcm78unk",
   "source_url": "https://omlx.ai/benchmarks/performance/lcm78unk",
   "tg_tokens": null,
   "tps": 87.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lcm78unk-ctx199680",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: ae4gu2lz, ebfmytxb, llmaeu4o, o9z0xrl4, tj22e7jo, xzacwcec, z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1579.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 195k 1,579 59.7 25.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lcm78unk",
   "source_url": "https://omlx.ai/benchmarks/performance/lcm78unk",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-4hkxtv25-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 927.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 927.6 94.3 19.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4hkxtv25",
   "source_url": "https://omlx.ai/benchmarks/performance/4hkxtv25",
   "tg_tokens": null,
   "tps": 94.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-4hkxtv25-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1054.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 1,054 90.7 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4hkxtv25",
   "source_url": "https://omlx.ai/benchmarks/performance/4hkxtv25",
   "tg_tokens": null,
   "tps": 90.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dssdp0o7-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3787.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 3,787 118.1 26.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dssdp0o7",
   "source_url": "https://omlx.ai/benchmarks/performance/dssdp0o7",
   "tg_tokens": null,
   "tps": 118.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dssdp0o7-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5216.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 5,216 173.4 26.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dssdp0o7",
   "source_url": "https://omlx.ai/benchmarks/performance/dssdp0o7",
   "tg_tokens": null,
   "tps": 173.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dssdp0o7-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5215.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 5,215 169.6 26.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dssdp0o7",
   "source_url": "https://omlx.ai/benchmarks/performance/dssdp0o7",
   "tg_tokens": null,
   "tps": 169.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dssdp0o7-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4892.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 4,892 183.3 27.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dssdp0o7",
   "source_url": "https://omlx.ai/benchmarks/performance/dssdp0o7",
   "tg_tokens": null,
   "tps": 183.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dssdp0o7-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4084.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 4,084 239.0 27.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dssdp0o7",
   "source_url": "https://omlx.ai/benchmarks/performance/dssdp0o7",
   "tg_tokens": null,
   "tps": 239.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dssdp0o7-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2991.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 2,991 130.2 28.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dssdp0o7",
   "source_url": "https://omlx.ai/benchmarks/performance/dssdp0o7",
   "tg_tokens": null,
   "tps": 130.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dssdp0o7-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1992.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 1,992 101.7 29.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dssdp0o7",
   "source_url": "https://omlx.ai/benchmarks/performance/dssdp0o7",
   "tg_tokens": null,
   "tps": 101.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ewz4alio-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: 2w4uiyks, 3sqfr2pw, c8x51isy, dfnf3lrx, lesilos7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2967.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 2,967 202.6 26.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ewz4alio",
   "source_url": "https://omlx.ai/benchmarks/performance/ewz4alio",
   "tg_tokens": null,
   "tps": 202.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ewz4alio-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: 2w4uiyks, 3sqfr2pw, c8x51isy, dfnf3lrx, lesilos7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4950.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 4,950 158.3 26.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ewz4alio",
   "source_url": "https://omlx.ai/benchmarks/performance/ewz4alio",
   "tg_tokens": null,
   "tps": 158.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ewz4alio-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: 2w4uiyks, 3sqfr2pw, c8x51isy, dfnf3lrx, lesilos7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5350.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 5,350 180.0 26.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ewz4alio",
   "source_url": "https://omlx.ai/benchmarks/performance/ewz4alio",
   "tg_tokens": null,
   "tps": 180.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ewz4alio-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.0 GB at this context; duplicate traces of the same session: 2w4uiyks, 3sqfr2pw, c8x51isy, dfnf3lrx, lesilos7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5012.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 5,012 211.3 27.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ewz4alio",
   "source_url": "https://omlx.ai/benchmarks/performance/ewz4alio",
   "tg_tokens": null,
   "tps": 211.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ewz4alio-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: 2w4uiyks, 3sqfr2pw, c8x51isy, dfnf3lrx, lesilos7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4144.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 4,144 229.4 27.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ewz4alio",
   "source_url": "https://omlx.ai/benchmarks/performance/ewz4alio",
   "tg_tokens": null,
   "tps": 229.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ewz4alio-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: 2w4uiyks, 3sqfr2pw, c8x51isy, dfnf3lrx, lesilos7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3088.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 3,088 144.9 28.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ewz4alio",
   "source_url": "https://omlx.ai/benchmarks/performance/ewz4alio",
   "tg_tokens": null,
   "tps": 144.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ewz4alio-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: 2w4uiyks, 3sqfr2pw, c8x51isy, dfnf3lrx, lesilos7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2110.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 2,110 109.8 29.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ewz4alio",
   "source_url": "https://omlx.ai/benchmarks/performance/ewz4alio",
   "tg_tokens": null,
   "tps": 109.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-mfdxrpgr-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: f31gykzy, tspkjmul; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 432.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 432.1 20.2 18.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mfdxrpgr",
   "source_url": "https://omlx.ai/benchmarks/performance/mfdxrpgr",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-mfdxrpgr-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: f31gykzy, tspkjmul; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 428.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 428.3 26.1 19.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mfdxrpgr",
   "source_url": "https://omlx.ai/benchmarks/performance/mfdxrpgr",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-mfdxrpgr-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: f31gykzy, tspkjmul; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 404.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 404.3 24.0 22.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mfdxrpgr",
   "source_url": "https://omlx.ai/benchmarks/performance/mfdxrpgr",
   "tg_tokens": null,
   "tps": 24.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-mfdxrpgr-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: f31gykzy, tspkjmul; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 360.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 360.5 23.4 26.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mfdxrpgr",
   "source_url": "https://omlx.ai/benchmarks/performance/mfdxrpgr",
   "tg_tokens": null,
   "tps": 23.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-yzxy2fw0-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 6yyueta2, bnqpqo0x, fqyyhltx, uxjubqh3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1375.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,375 84.7 17.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yzxy2fw0",
   "source_url": "https://omlx.ai/benchmarks/performance/yzxy2fw0",
   "tg_tokens": null,
   "tps": 84.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-yzxy2fw0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: 6yyueta2, bnqpqo0x, fqyyhltx, uxjubqh3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1624.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,624 137.9 18.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yzxy2fw0",
   "source_url": "https://omlx.ai/benchmarks/performance/yzxy2fw0",
   "tg_tokens": null,
   "tps": 137.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-yzxy2fw0-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 6yyueta2, bnqpqo0x, fqyyhltx, uxjubqh3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1619.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 1,619 61.7 19.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yzxy2fw0",
   "source_url": "https://omlx.ai/benchmarks/performance/yzxy2fw0",
   "tg_tokens": null,
   "tps": 61.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-yzxy2fw0-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 6yyueta2, bnqpqo0x, fqyyhltx, uxjubqh3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1566.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,566 73.8 19.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yzxy2fw0",
   "source_url": "https://omlx.ai/benchmarks/performance/yzxy2fw0",
   "tg_tokens": null,
   "tps": 73.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-yzxy2fw0-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: 6yyueta2, bnqpqo0x, fqyyhltx, uxjubqh3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1453.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,453 116.5 20.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yzxy2fw0",
   "source_url": "https://omlx.ai/benchmarks/performance/yzxy2fw0",
   "tg_tokens": null,
   "tps": 116.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-yzxy2fw0-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: 6yyueta2, bnqpqo0x, fqyyhltx, uxjubqh3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1255.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,255 77.1 23.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yzxy2fw0",
   "source_url": "https://omlx.ai/benchmarks/performance/yzxy2fw0",
   "tg_tokens": null,
   "tps": 77.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-yzxy2fw0-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: 6yyueta2, bnqpqo0x, fqyyhltx, uxjubqh3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 978.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 978.9 77.6 27.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yzxy2fw0",
   "source_url": "https://omlx.ai/benchmarks/performance/yzxy2fw0",
   "tg_tokens": null,
   "tps": 77.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-yzxy2fw0-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 33.0 GB at this context; duplicate traces of the same session: 6yyueta2, bnqpqo0x, fqyyhltx, uxjubqh3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 795.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 795.4 69.2 33.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yzxy2fw0",
   "source_url": "https://omlx.ai/benchmarks/performance/yzxy2fw0",
   "tg_tokens": null,
   "tps": 69.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-q987p750-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1377.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,377 79.3 17.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q987p750",
   "source_url": "https://omlx.ai/benchmarks/performance/q987p750",
   "tg_tokens": null,
   "tps": 79.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-q987p750-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1621.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,621 100.6 18.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q987p750",
   "source_url": "https://omlx.ai/benchmarks/performance/q987p750",
   "tg_tokens": null,
   "tps": 100.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-19vgdd18-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 939.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 939.2 32.8 17.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/19vgdd18",
   "source_url": "https://omlx.ai/benchmarks/performance/19vgdd18",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-19vgdd18-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 973.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 973.1 31.0 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/19vgdd18",
   "source_url": "https://omlx.ai/benchmarks/performance/19vgdd18",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-19vgdd18-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 798.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 798.8 25.7 20.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/19vgdd18",
   "source_url": "https://omlx.ai/benchmarks/performance/19vgdd18",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-19vgdd18-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 637.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 637.0 21.6 24.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/19vgdd18",
   "source_url": "https://omlx.ai/benchmarks/performance/19vgdd18",
   "tg_tokens": null,
   "tps": 21.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-19vgdd18-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 515.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 195k 515.5 18.1 29.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/19vgdd18",
   "source_url": "https://omlx.ai/benchmarks/performance/19vgdd18",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-htajawiv-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 24.9 GB; TTFT 110132 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 595.1,
   "provenance": "community",
   "quant": "OptiQ 4bit",
   "quote": "Qwen3.8-27B-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k tokens 595.1 PP tok/s 18.1 TG tok/s 110132 TTFT (ms) 24.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/htajawiv",
   "source_url": "https://omlx.ai/benchmarks/performance/htajawiv",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gi320am8-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: e7ax1svx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 420.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 420.5 73.1 21.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gi320am8",
   "source_url": "https://omlx.ai/benchmarks/performance/gi320am8",
   "tg_tokens": null,
   "tps": 73.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gi320am8-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: e7ax1svx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 559.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 559.9 111.1 25.0 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gi320am8",
   "source_url": "https://omlx.ai/benchmarks/performance/gi320am8",
   "tg_tokens": null,
   "tps": 111.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gi320am8-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: e7ax1svx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 551.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 551.5 54.4 25.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gi320am8",
   "source_url": "https://omlx.ai/benchmarks/performance/gi320am8",
   "tg_tokens": null,
   "tps": 54.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gi320am8-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: e7ax1svx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 526.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 526.8 76.3 26.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gi320am8",
   "source_url": "https://omlx.ai/benchmarks/performance/gi320am8",
   "tg_tokens": null,
   "tps": 76.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-v4z1qwxa-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: f05iixoj, t5e5p84z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 899.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 899.1 67.5 24.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v4z1qwxa",
   "source_url": "https://omlx.ai/benchmarks/performance/v4z1qwxa",
   "tg_tokens": null,
   "tps": 67.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-v4z1qwxa-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.7 GB at this context; duplicate traces of the same session: f05iixoj, t5e5p84z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1157.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 1,157 58.7 25.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v4z1qwxa",
   "source_url": "https://omlx.ai/benchmarks/performance/v4z1qwxa",
   "tg_tokens": null,
   "tps": 58.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-v4z1qwxa-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: f05iixoj, t5e5p84z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1125.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 1,125 60.4 26.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v4z1qwxa",
   "source_url": "https://omlx.ai/benchmarks/performance/v4z1qwxa",
   "tg_tokens": null,
   "tps": 60.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-v4z1qwxa-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: f05iixoj, t5e5p84z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1048.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 1,048 78.8 26.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v4z1qwxa",
   "source_url": "https://omlx.ai/benchmarks/performance/v4z1qwxa",
   "tg_tokens": null,
   "tps": 78.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-v4z1qwxa-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: f05iixoj, t5e5p84z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 919.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 919.9 113.4 27.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v4z1qwxa",
   "source_url": "https://omlx.ai/benchmarks/performance/v4z1qwxa",
   "tg_tokens": null,
   "tps": 113.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-v4z1qwxa-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.0 GB at this context; duplicate traces of the same session: f05iixoj, t5e5p84z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 754.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 754.1 39.1 30.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v4z1qwxa",
   "source_url": "https://omlx.ai/benchmarks/performance/v4z1qwxa",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-v4z1qwxa-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: f05iixoj, t5e5p84z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 538.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 538.9 39.8 34.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v4z1qwxa",
   "source_url": "https://omlx.ai/benchmarks/performance/v4z1qwxa",
   "tg_tokens": null,
   "tps": 39.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-p1os044e-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: i8ref11h, pfhzn4ht; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 831.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 831.7 55.9 23.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p1os044e",
   "source_url": "https://omlx.ai/benchmarks/performance/p1os044e",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-p1os044e-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: i8ref11h, pfhzn4ht; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 886.6,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 886.6 53.0 24.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p1os044e",
   "source_url": "https://omlx.ai/benchmarks/performance/p1os044e",
   "tg_tokens": null,
   "tps": 53.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-p1os044e-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: i8ref11h, pfhzn4ht; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 816.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 816.0 66.0 24.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p1os044e",
   "source_url": "https://omlx.ai/benchmarks/performance/p1os044e",
   "tg_tokens": null,
   "tps": 66.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-p1os044e-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.2 GB at this context; duplicate traces of the same session: i8ref11h, pfhzn4ht; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 783.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 783.9 75.3 25.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p1os044e",
   "source_url": "https://omlx.ai/benchmarks/performance/p1os044e",
   "tg_tokens": null,
   "tps": 75.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-p1os044e-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: i8ref11h, pfhzn4ht; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 676.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 676.4 113.5 26.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p1os044e",
   "source_url": "https://omlx.ai/benchmarks/performance/p1os044e",
   "tg_tokens": null,
   "tps": 113.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-p1os044e-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.7 GB at this context; duplicate traces of the same session: i8ref11h, pfhzn4ht; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 566.3,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 566.3 44.5 28.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p1os044e",
   "source_url": "https://omlx.ai/benchmarks/performance/p1os044e",
   "tg_tokens": null,
   "tps": 44.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-p1os044e-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.4 GB at this context; duplicate traces of the same session: i8ref11h, pfhzn4ht; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 442.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 442.4 40.1 33.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p1os044e",
   "source_url": "https://omlx.ai/benchmarks/performance/p1os044e",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q6df6psy-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: 0d5mozkt, dzuk1zp6, td20xvnd, xznrsudm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 816.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 816.7 55.8 24.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q6df6psy",
   "source_url": "https://omlx.ai/benchmarks/performance/q6df6psy",
   "tg_tokens": null,
   "tps": 55.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q6df6psy-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.7 GB at this context; duplicate traces of the same session: 0d5mozkt, dzuk1zp6, td20xvnd, xznrsudm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 868.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 868.4 47.8 25.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q6df6psy",
   "source_url": "https://omlx.ai/benchmarks/performance/q6df6psy",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q6df6psy-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: 0d5mozkt, dzuk1zp6, td20xvnd, xznrsudm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 951.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 951.9 53.4 26.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q6df6psy",
   "source_url": "https://omlx.ai/benchmarks/performance/q6df6psy",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-q6df6psy-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: 0d5mozkt, dzuk1zp6, td20xvnd, xznrsudm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 961.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 961.9 84.9 26.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q6df6psy",
   "source_url": "https://omlx.ai/benchmarks/performance/q6df6psy",
   "tg_tokens": null,
   "tps": 84.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q6df6psy-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: 0d5mozkt, dzuk1zp6, td20xvnd, xznrsudm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 891.8,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 891.8 105.7 27.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q6df6psy",
   "source_url": "https://omlx.ai/benchmarks/performance/q6df6psy",
   "tg_tokens": null,
   "tps": 105.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q6df6psy-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.0 GB at this context; duplicate traces of the same session: 0d5mozkt, dzuk1zp6, td20xvnd, xznrsudm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 740.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 740.2 43.7 30.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q6df6psy",
   "source_url": "https://omlx.ai/benchmarks/performance/q6df6psy",
   "tg_tokens": null,
   "tps": 43.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-q6df6psy-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: 0d5mozkt, dzuk1zp6, td20xvnd, xznrsudm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 554.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 554.4 41.8 34.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q6df6psy",
   "source_url": "https://omlx.ai/benchmarks/performance/q6df6psy",
   "tg_tokens": null,
   "tps": 41.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-eh5pdjbl-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: huise6v0, ksz9bfj4; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 403.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 1k 403.8 37.5 70.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eh5pdjbl",
   "source_url": "https://omlx.ai/benchmarks/performance/eh5pdjbl",
   "tg_tokens": null,
   "tps": 37.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-eh5pdjbl-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: huise6v0, ksz9bfj4; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 483.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 4k 483.2 38.0 71.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eh5pdjbl",
   "source_url": "https://omlx.ai/benchmarks/performance/eh5pdjbl",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-eh5pdjbl-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.1 GB at this context; duplicate traces of the same session: huise6v0, ksz9bfj4; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 502.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 16k 502.5 32.1 72.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eh5pdjbl",
   "source_url": "https://omlx.ai/benchmarks/performance/eh5pdjbl",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-eh5pdjbl-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.1 GB at this context; duplicate traces of the same session: huise6v0, ksz9bfj4; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 474.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 64k 474.2 30.1 74.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eh5pdjbl",
   "source_url": "https://omlx.ai/benchmarks/performance/eh5pdjbl",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-eh5pdjbl-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.8 GB at this context; duplicate traces of the same session: huise6v0, ksz9bfj4; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 376.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 128k 376.4 27.8 78.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eh5pdjbl",
   "source_url": "https://omlx.ai/benchmarks/performance/eh5pdjbl",
   "tg_tokens": null,
   "tps": 27.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-igls3ssm-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.2 GB at this context; duplicate traces of the same session: jbsn8l4d, n68n803o, qj4vilpx, vcn5b7c9; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 811.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 1k 811.9 45.3 183.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/igls3ssm",
   "source_url": "https://omlx.ai/benchmarks/performance/igls3ssm",
   "tg_tokens": null,
   "tps": 45.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-igls3ssm-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.6 GB at this context; duplicate traces of the same session: jbsn8l4d, n68n803o, qj4vilpx, vcn5b7c9; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2420.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 4k 2,420 117.1 183.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/igls3ssm",
   "source_url": "https://omlx.ai/benchmarks/performance/igls3ssm",
   "tg_tokens": null,
   "tps": 117.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-igls3ssm-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 184.7 GB at this context; duplicate traces of the same session: jbsn8l4d, n68n803o, qj4vilpx, vcn5b7c9; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2716.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 8k 2,716 111.0 184.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/igls3ssm",
   "source_url": "https://omlx.ai/benchmarks/performance/igls3ssm",
   "tg_tokens": null,
   "tps": 111.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-igls3ssm-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 186.9 GB at this context; duplicate traces of the same session: jbsn8l4d, n68n803o, qj4vilpx, vcn5b7c9; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4444.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 64k 4,444 114.8 186.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/igls3ssm",
   "source_url": "https://omlx.ai/benchmarks/performance/igls3ssm",
   "tg_tokens": null,
   "tps": 114.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-igls3ssm-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 191.6 GB at this context; duplicate traces of the same session: jbsn8l4d, n68n803o, qj4vilpx, vcn5b7c9; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4314.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 128k 4,314 90.2 191.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/igls3ssm",
   "source_url": "https://omlx.ai/benchmarks/performance/igls3ssm",
   "tg_tokens": null,
   "tps": 90.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-igls3ssm-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 193.3 GB at this context; duplicate traces of the same session: jbsn8l4d, n68n803o, qj4vilpx, vcn5b7c9; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4189.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 195k 4,189 78.1 193.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/igls3ssm",
   "source_url": "https://omlx.ai/benchmarks/performance/igls3ssm",
   "tg_tokens": null,
   "tps": 78.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-3c9hmohk-ctx8192",
   "model": "Qwopus3.8-27B-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: 6xmd1ie5, iuoch7w4; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 492.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwopus3.8-27B-Flash-oQ4e-mtp-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 492.7 16.1 18.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3c9hmohk",
   "source_url": "https://omlx.ai/benchmarks/performance/3c9hmohk",
   "tg_tokens": null,
   "tps": 16.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-3c9hmohk-ctx16384",
   "model": "Qwopus3.8-27B-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 6xmd1ie5, iuoch7w4; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 493.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwopus3.8-27B-Flash-oQ4e-mtp-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 493.4 16.4 19.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3c9hmohk",
   "source_url": "https://omlx.ai/benchmarks/performance/3c9hmohk",
   "tg_tokens": null,
   "tps": 16.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-3c9hmohk-ctx32768",
   "model": "Qwopus3.8-27B-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 6xmd1ie5, iuoch7w4; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 446.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwopus3.8-27B-Flash-oQ4e-mtp-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 446.5 14.1 21.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3c9hmohk",
   "source_url": "https://omlx.ai/benchmarks/performance/3c9hmohk",
   "tg_tokens": null,
   "tps": 14.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ia4qoc4d-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: 8xct0lop, zaerwmel; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 455.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 455.1 25.6 19.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ia4qoc4d",
   "source_url": "https://omlx.ai/benchmarks/performance/ia4qoc4d",
   "tg_tokens": null,
   "tps": 25.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ia4qoc4d-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 8xct0lop, zaerwmel; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 441.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 441.9 33.4 20.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ia4qoc4d",
   "source_url": "https://omlx.ai/benchmarks/performance/ia4qoc4d",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ia4qoc4d-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: 8xct0lop, zaerwmel; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 402.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 402.5 30.6 22.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ia4qoc4d",
   "source_url": "https://omlx.ai/benchmarks/performance/ia4qoc4d",
   "tg_tokens": null,
   "tps": 30.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-0bv02ptn-ctx16384",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "fp16 MLX build; oMLX community benchmark trace, single run at 16k context; peak memory 13.4 GB; TTFT 318677 ms; duplicate traces of the same session: none; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 51.4,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 2bit; 16k tokens 51.4 PP tok/s 7.2 TG tok/s 318677 TTFT (ms) 13.4 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0bv02ptn",
   "source_url": "https://omlx.ai/benchmarks/performance/0bv02ptn",
   "tg_tokens": null,
   "tps": 7.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-nzo68lbp-ctx8192",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "fp16 MLX build; oMLX community benchmark trace, single run at 8k context; peak memory 13.1 GB; TTFT 143579 ms; duplicate traces of the same session: none; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 57.1,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 2bit; 8k tokens 57.1 PP tok/s 9.5 TG tok/s 143579 TTFT (ms) 13.1 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nzo68lbp",
   "source_url": "https://omlx.ai/benchmarks/performance/nzo68lbp",
   "tg_tokens": null,
   "tps": 9.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-8z63rlpm-ctx8192",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: ogbtfgrc, qy3kieah; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3098.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 3,098 91.2 22.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8z63rlpm",
   "source_url": "https://omlx.ai/benchmarks/performance/8z63rlpm",
   "tg_tokens": null,
   "tps": 91.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-8z63rlpm-ctx16384",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: ogbtfgrc, qy3kieah; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2805.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 2,805 87.5 22.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8z63rlpm",
   "source_url": "https://omlx.ai/benchmarks/performance/8z63rlpm",
   "tg_tokens": null,
   "tps": 87.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-8z63rlpm-ctx32768",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: ogbtfgrc, qy3kieah; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2321.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 2,321 87.9 23.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8z63rlpm",
   "source_url": "https://omlx.ai/benchmarks/performance/8z63rlpm",
   "tg_tokens": null,
   "tps": 87.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0vv1ycya-ctx8192",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 121sd4zt, 1zi1ag5u; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1923.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 1,923 47.4 22.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0vv1ycya",
   "source_url": "https://omlx.ai/benchmarks/performance/0vv1ycya",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-0vv1ycya-ctx16384",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 121sd4zt, 1zi1ag5u; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2790.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 2,790 81.8 22.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0vv1ycya",
   "source_url": "https://omlx.ai/benchmarks/performance/0vv1ycya",
   "tg_tokens": null,
   "tps": 81.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0vv1ycya-ctx32768",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 121sd4zt, 1zi1ag5u; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2398.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 2,398 102.6 23.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0vv1ycya",
   "source_url": "https://omlx.ai/benchmarks/performance/0vv1ycya",
   "tg_tokens": null,
   "tps": 102.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-0vv1ycya-ctx199680",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 32.2 GB at this context; duplicate traces of the same session: 121sd4zt, 1zi1ag5u; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 913.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 195k 913.1 54.3 32.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0vv1ycya",
   "source_url": "https://omlx.ai/benchmarks/performance/0vv1ycya",
   "tg_tokens": null,
   "tps": 54.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-7jbivu99-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 676.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 1k 676.0 70.3 14.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7jbivu99",
   "source_url": "https://omlx.ai/benchmarks/performance/7jbivu99",
   "tg_tokens": null,
   "tps": 70.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-7jbivu99-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 672.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 4k 672.4 67.1 14.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7jbivu99",
   "source_url": "https://omlx.ai/benchmarks/performance/7jbivu99",
   "tg_tokens": null,
   "tps": 67.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-7jbivu99-ctx8192",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 662.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 8k 662.4 63.1 14.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7jbivu99",
   "source_url": "https://omlx.ai/benchmarks/performance/7jbivu99",
   "tg_tokens": null,
   "tps": 63.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-7jbivu99-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 15.2 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 632.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 16k 632.0 56.2 15.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7jbivu99",
   "source_url": "https://omlx.ai/benchmarks/performance/7jbivu99",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-7jbivu99-ctx32768",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 578.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 32k 578.4 46.9 16.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7jbivu99",
   "source_url": "https://omlx.ai/benchmarks/performance/7jbivu99",
   "tg_tokens": null,
   "tps": 46.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-7jbivu99-ctx65536",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 493.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 64k 493.5 34.6 17.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7jbivu99",
   "source_url": "https://omlx.ai/benchmarks/performance/7jbivu99",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-abfhn6u1-ctx1024",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 208.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 208.4 20.5 18.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/abfhn6u1",
   "source_url": "https://omlx.ai/benchmarks/performance/abfhn6u1",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-abfhn6u1-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 205.8 23.0 19.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/abfhn6u1",
   "source_url": "https://omlx.ai/benchmarks/performance/abfhn6u1",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-abfhn6u1-ctx8192",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 197.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 197.3 22.4 19.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/abfhn6u1",
   "source_url": "https://omlx.ai/benchmarks/performance/abfhn6u1",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-abfhn6u1-ctx16384",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 193.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 193.0 21.5 21.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/abfhn6u1",
   "source_url": "https://omlx.ai/benchmarks/performance/abfhn6u1",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-abfhn6u1-ctx32768",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 178.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 178.0 19.8 23.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/abfhn6u1",
   "source_url": "https://omlx.ai/benchmarks/performance/abfhn6u1",
   "tg_tokens": null,
   "tps": 19.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-abfhn6u1-ctx65536",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 155.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 155.1 17.0 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/abfhn6u1",
   "source_url": "https://omlx.ai/benchmarks/performance/abfhn6u1",
   "tg_tokens": null,
   "tps": 17.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-edojcgp7-ctx1024",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 46sqmiqt, g9qikzc8, tsoiobp2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 207.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 207.0 28.5 20.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/edojcgp7",
   "source_url": "https://omlx.ai/benchmarks/performance/edojcgp7",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-edojcgp7-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 46sqmiqt, g9qikzc8, tsoiobp2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 204.4 46.7 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/edojcgp7",
   "source_url": "https://omlx.ai/benchmarks/performance/edojcgp7",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-edojcgp7-ctx8192",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 46sqmiqt, g9qikzc8, tsoiobp2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 193.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 193.4 30.7 21.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/edojcgp7",
   "source_url": "https://omlx.ai/benchmarks/performance/edojcgp7",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-edojcgp7-ctx16384",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: 46sqmiqt, g9qikzc8, tsoiobp2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 193.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 193.4 39.1 23.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/edojcgp7",
   "source_url": "https://omlx.ai/benchmarks/performance/edojcgp7",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-edojcgp7-ctx32768",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: 46sqmiqt, g9qikzc8, tsoiobp2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 179.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 179.7 26.2 25.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/edojcgp7",
   "source_url": "https://omlx.ai/benchmarks/performance/edojcgp7",
   "tg_tokens": null,
   "tps": 26.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-edojcgp7-ctx65536",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: 46sqmiqt, g9qikzc8, tsoiobp2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 158.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 158.0 20.1 29.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/edojcgp7",
   "source_url": "https://omlx.ai/benchmarks/performance/edojcgp7",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-cd2nlpsm-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 6.0 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1559.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E4B-it-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 1,559 26.8 6.0 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cd2nlpsm",
   "source_url": "https://omlx.ai/benchmarks/performance/cd2nlpsm",
   "tg_tokens": null,
   "tps": 26.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-cd2nlpsm-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 6.1 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1760.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E4B-it-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 1,760 26.0 6.1 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cd2nlpsm",
   "source_url": "https://omlx.ai/benchmarks/performance/cd2nlpsm",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-cm2lhkes-ctx1024",
   "model": "DeepSeek-V4-Flash",
   "notes": "oMLX community benchmark trace; peak memory 26.9 GB at this context; duplicate traces of the same session: x0a3uvid; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 109.1,
   "provenance": "community",
   "quant": "2bit DQ",
   "quote": "DeepSeek-V4-Flash-2bit-DQ on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 109.1 13.6 26.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cm2lhkes",
   "source_url": "https://omlx.ai/benchmarks/performance/cm2lhkes",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-cm2lhkes-ctx4096",
   "model": "DeepSeek-V4-Flash",
   "notes": "oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: x0a3uvid; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 180.0,
   "provenance": "community",
   "quant": "2bit DQ",
   "quote": "DeepSeek-V4-Flash-2bit-DQ on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 180.0 9.1 27.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cm2lhkes",
   "source_url": "https://omlx.ai/benchmarks/performance/cm2lhkes",
   "tg_tokens": null,
   "tps": 9.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-st00lkvv-ctx1024",
   "model": "Ling-3.0-tiny",
   "notes": "oMLX community benchmark trace; peak memory 5.1 GB at this context; duplicate traces of the same session: cixry28q; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 830.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-tiny-oQ4e on M4 (8c); M4 (8c) \u00b7 16 GB \u00b7 4bit; 1k 830.1 95.3 5.1 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/st00lkvv",
   "source_url": "https://omlx.ai/benchmarks/performance/st00lkvv",
   "tg_tokens": null,
   "tps": 95.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-st00lkvv-ctx4096",
   "model": "Ling-3.0-tiny",
   "notes": "oMLX community benchmark trace; peak memory 5.5 GB at this context; duplicate traces of the same session: cixry28q; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 957.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-tiny-oQ4e on M4 (8c); M4 (8c) \u00b7 16 GB \u00b7 4bit; 4k 957.3 88.1 5.5 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/st00lkvv",
   "source_url": "https://omlx.ai/benchmarks/performance/st00lkvv",
   "tg_tokens": null,
   "tps": 88.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ruzzi4tz-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Noctalin build; fp16 MTP; oMLX community benchmark trace, single run at 64k context; peak memory 25.2 GB; TTFT 69197 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 947.1,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Noctalin--Qwen3.6-35B-A3B-oQ4-fp16-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k tokens 947.1 PP tok/s 88.1 TG tok/s 69197 TTFT (ms) 25.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ruzzi4tz",
   "source_url": "https://omlx.ai/benchmarks/performance/ruzzi4tz",
   "tg_tokens": null,
   "tps": 88.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-8jcvz5bh-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 893.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 893.3 78.2 21.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8jcvz5bh",
   "source_url": "https://omlx.ai/benchmarks/performance/8jcvz5bh",
   "tg_tokens": null,
   "tps": 78.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-8jcvz5bh-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1024.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,024 79.4 23.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8jcvz5bh",
   "source_url": "https://omlx.ai/benchmarks/performance/8jcvz5bh",
   "tg_tokens": null,
   "tps": 79.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-8jcvz5bh-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 954.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 954.7 85.3 23.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8jcvz5bh",
   "source_url": "https://omlx.ai/benchmarks/performance/8jcvz5bh",
   "tg_tokens": null,
   "tps": 85.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-8jcvz5bh-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 837.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 837.7 77.1 24.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8jcvz5bh",
   "source_url": "https://omlx.ai/benchmarks/performance/8jcvz5bh",
   "tg_tokens": null,
   "tps": 77.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-8jcvz5bh-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 678.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 678.6 67.6 25.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8jcvz5bh",
   "source_url": "https://omlx.ai/benchmarks/performance/8jcvz5bh",
   "tg_tokens": null,
   "tps": 67.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-8jcvz5bh-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 491.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 491.9 43.5 26.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8jcvz5bh",
   "source_url": "https://omlx.ai/benchmarks/performance/8jcvz5bh",
   "tg_tokens": null,
   "tps": 43.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9brouuzo-ctx1024",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 9r3ze2f9, mqw7bjz2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4438.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 4,438 136.5 19.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9brouuzo",
   "source_url": "https://omlx.ai/benchmarks/performance/9brouuzo",
   "tg_tokens": null,
   "tps": 136.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9brouuzo-ctx32768",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 9r3ze2f9, mqw7bjz2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4546.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 4,546 113.3 20.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9brouuzo",
   "source_url": "https://omlx.ai/benchmarks/performance/9brouuzo",
   "tg_tokens": null,
   "tps": 113.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9brouuzo-ctx65536",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 9r3ze2f9, mqw7bjz2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3366.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 64k 3,366 99.7 21.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9brouuzo",
   "source_url": "https://omlx.ai/benchmarks/performance/9brouuzo",
   "tg_tokens": null,
   "tps": 99.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9brouuzo-ctx131072",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: 9r3ze2f9, mqw7bjz2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2252.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 128k 2,252 76.4 22.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9brouuzo",
   "source_url": "https://omlx.ai/benchmarks/performance/9brouuzo",
   "tg_tokens": null,
   "tps": 76.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9brouuzo-ctx199680",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: 9r3ze2f9, mqw7bjz2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1576.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 195k 1,576 58.2 23.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9brouuzo",
   "source_url": "https://omlx.ai/benchmarks/performance/9brouuzo",
   "tg_tokens": null,
   "tps": 58.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-o9z0xrl4-ctx1024",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: ae4gu2lz, ebfmytxb, llmaeu4o, tj22e7jo, xzacwcec, z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4163.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 4,163 192.7 19.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9z0xrl4",
   "source_url": "https://omlx.ai/benchmarks/performance/o9z0xrl4",
   "tg_tokens": null,
   "tps": 192.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-o9z0xrl4-ctx4096",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: ae4gu2lz, ebfmytxb, llmaeu4o, tj22e7jo, xzacwcec, z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5736.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 5,736 159.3 20.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9z0xrl4",
   "source_url": "https://omlx.ai/benchmarks/performance/o9z0xrl4",
   "tg_tokens": null,
   "tps": 159.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-o9z0xrl4-ctx8192",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: ae4gu2lz, ebfmytxb, llmaeu4o, tj22e7jo, xzacwcec, z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6112.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 8k 6,112 203.0 20.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9z0xrl4",
   "source_url": "https://omlx.ai/benchmarks/performance/o9z0xrl4",
   "tg_tokens": null,
   "tps": 203.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-o9z0xrl4-ctx16384",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: ae4gu2lz, ebfmytxb, llmaeu4o, tj22e7jo, xzacwcec, z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5607.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 5,607 195.0 20.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9z0xrl4",
   "source_url": "https://omlx.ai/benchmarks/performance/o9z0xrl4",
   "tg_tokens": null,
   "tps": 195.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-o9z0xrl4-ctx32768",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: ae4gu2lz, ebfmytxb, llmaeu4o, tj22e7jo, xzacwcec, z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4807.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 4,807 170.8 21.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9z0xrl4",
   "source_url": "https://omlx.ai/benchmarks/performance/o9z0xrl4",
   "tg_tokens": null,
   "tps": 170.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-o9z0xrl4-ctx65536",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: ae4gu2lz, ebfmytxb, llmaeu4o, tj22e7jo, xzacwcec, z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3580.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 64k 3,580 126.1 22.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9z0xrl4",
   "source_url": "https://omlx.ai/benchmarks/performance/o9z0xrl4",
   "tg_tokens": null,
   "tps": 126.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-o9z0xrl4-ctx131072",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: ae4gu2lz, ebfmytxb, llmaeu4o, tj22e7jo, xzacwcec, z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2310.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 128k 2,310 87.1 23.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9z0xrl4",
   "source_url": "https://omlx.ai/benchmarks/performance/o9z0xrl4",
   "tg_tokens": null,
   "tps": 87.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-o9z0xrl4-ctx199680",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: ae4gu2lz, ebfmytxb, llmaeu4o, tj22e7jo, xzacwcec, z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1579.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 195k 1,579 59.7 25.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9z0xrl4",
   "source_url": "https://omlx.ai/benchmarks/performance/o9z0xrl4",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n3s01r8j-ctx1024",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 007jy9ug, 3cy8ezo9, 8smq47fl, 90yw4img, esbg3qoh, yu64gvip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2615.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 2,615 151.1 19.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n3s01r8j",
   "source_url": "https://omlx.ai/benchmarks/performance/n3s01r8j",
   "tg_tokens": null,
   "tps": 151.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-n3s01r8j-ctx4096",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: 007jy9ug, 3cy8ezo9, 8smq47fl, 90yw4img, esbg3qoh, yu64gvip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2684.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 2,684 173.6 20.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n3s01r8j",
   "source_url": "https://omlx.ai/benchmarks/performance/n3s01r8j",
   "tg_tokens": null,
   "tps": 173.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-n3s01r8j-ctx8192",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 007jy9ug, 3cy8ezo9, 8smq47fl, 90yw4img, esbg3qoh, yu64gvip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4567.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 8k 4,567 190.8 20.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n3s01r8j",
   "source_url": "https://omlx.ai/benchmarks/performance/n3s01r8j",
   "tg_tokens": null,
   "tps": 190.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-n3s01r8j-ctx16384",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 007jy9ug, 3cy8ezo9, 8smq47fl, 90yw4img, esbg3qoh, yu64gvip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4610.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 4,610 177.6 20.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n3s01r8j",
   "source_url": "https://omlx.ai/benchmarks/performance/n3s01r8j",
   "tg_tokens": null,
   "tps": 177.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-n3s01r8j-ctx32768",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 007jy9ug, 3cy8ezo9, 8smq47fl, 90yw4img, esbg3qoh, yu64gvip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4019.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 4,019 164.2 21.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n3s01r8j",
   "source_url": "https://omlx.ai/benchmarks/performance/n3s01r8j",
   "tg_tokens": null,
   "tps": 164.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-n3s01r8j-ctx65536",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 007jy9ug, 3cy8ezo9, 8smq47fl, 90yw4img, esbg3qoh, yu64gvip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2973.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 64k 2,973 121.2 21.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n3s01r8j",
   "source_url": "https://omlx.ai/benchmarks/performance/n3s01r8j",
   "tg_tokens": null,
   "tps": 121.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-n3s01r8j-ctx131072",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 007jy9ug, 3cy8ezo9, 8smq47fl, 90yw4img, esbg3qoh, yu64gvip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2042.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 128k 2,042 91.8 23.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n3s01r8j",
   "source_url": "https://omlx.ai/benchmarks/performance/n3s01r8j",
   "tg_tokens": null,
   "tps": 91.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-n3s01r8j-ctx199680",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: 007jy9ug, 3cy8ezo9, 8smq47fl, 90yw4img, esbg3qoh, yu64gvip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1432.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 195k 1,432 62.6 24.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n3s01r8j",
   "source_url": "https://omlx.ai/benchmarks/performance/n3s01r8j",
   "tg_tokens": null,
   "tps": 62.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-rlamn7ox-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: woqc1w7p, z9qsd51n; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 822.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 822.4 86.2 20.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rlamn7ox",
   "source_url": "https://omlx.ai/benchmarks/performance/rlamn7ox",
   "tg_tokens": null,
   "tps": 86.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-rlamn7ox-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: woqc1w7p, z9qsd51n; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 703.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 703.5 73.2 20.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rlamn7ox",
   "source_url": "https://omlx.ai/benchmarks/performance/rlamn7ox",
   "tg_tokens": null,
   "tps": 73.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-rlamn7ox-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: woqc1w7p, z9qsd51n; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 286.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 128k 286.4 33.4 24.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rlamn7ox",
   "source_url": "https://omlx.ai/benchmarks/performance/rlamn7ox",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2w4uiyks-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: 3sqfr2pw, c8x51isy, dfnf3lrx, lesilos7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2967.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 2,967 202.6 26.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2w4uiyks",
   "source_url": "https://omlx.ai/benchmarks/performance/2w4uiyks",
   "tg_tokens": null,
   "tps": 202.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2w4uiyks-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: 3sqfr2pw, c8x51isy, dfnf3lrx, lesilos7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4950.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 4,950 158.3 26.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2w4uiyks",
   "source_url": "https://omlx.ai/benchmarks/performance/2w4uiyks",
   "tg_tokens": null,
   "tps": 158.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2w4uiyks-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: 3sqfr2pw, c8x51isy, dfnf3lrx, lesilos7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5350.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 5,350 180.0 26.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2w4uiyks",
   "source_url": "https://omlx.ai/benchmarks/performance/2w4uiyks",
   "tg_tokens": null,
   "tps": 180.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2w4uiyks-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.0 GB at this context; duplicate traces of the same session: 3sqfr2pw, c8x51isy, dfnf3lrx, lesilos7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5012.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 5,012 211.3 27.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2w4uiyks",
   "source_url": "https://omlx.ai/benchmarks/performance/2w4uiyks",
   "tg_tokens": null,
   "tps": 211.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2w4uiyks-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: 3sqfr2pw, c8x51isy, dfnf3lrx, lesilos7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4144.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 4,144 229.4 27.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2w4uiyks",
   "source_url": "https://omlx.ai/benchmarks/performance/2w4uiyks",
   "tg_tokens": null,
   "tps": 229.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2w4uiyks-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: 3sqfr2pw, c8x51isy, dfnf3lrx, lesilos7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3088.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 3,088 144.9 28.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2w4uiyks",
   "source_url": "https://omlx.ai/benchmarks/performance/2w4uiyks",
   "tg_tokens": null,
   "tps": 144.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2w4uiyks-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: 3sqfr2pw, c8x51isy, dfnf3lrx, lesilos7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2110.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 2,110 109.8 29.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2w4uiyks",
   "source_url": "https://omlx.ai/benchmarks/performance/2w4uiyks",
   "tg_tokens": null,
   "tps": 109.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ydfahqwc-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: m9gw1nld; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 217.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 217.7 35.2 17.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ydfahqwc",
   "source_url": "https://omlx.ai/benchmarks/performance/ydfahqwc",
   "tg_tokens": null,
   "tps": 35.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ydfahqwc-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: m9gw1nld; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 203.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 203.7 38.6 18.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ydfahqwc",
   "source_url": "https://omlx.ai/benchmarks/performance/ydfahqwc",
   "tg_tokens": null,
   "tps": 38.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-tspkjmul-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: f31gykzy; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 432.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 432.1 20.2 18.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tspkjmul",
   "source_url": "https://omlx.ai/benchmarks/performance/tspkjmul",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-tspkjmul-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: f31gykzy; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 428.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 428.3 26.1 19.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tspkjmul",
   "source_url": "https://omlx.ai/benchmarks/performance/tspkjmul",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-tspkjmul-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: f31gykzy; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 404.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 404.3 24.0 22.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tspkjmul",
   "source_url": "https://omlx.ai/benchmarks/performance/tspkjmul",
   "tg_tokens": null,
   "tps": 24.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-tspkjmul-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: f31gykzy; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 360.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 360.5 23.4 26.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tspkjmul",
   "source_url": "https://omlx.ai/benchmarks/performance/tspkjmul",
   "tg_tokens": null,
   "tps": 23.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-bnqpqo0x-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 6yyueta2, fqyyhltx, uxjubqh3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1375.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,375 84.7 17.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bnqpqo0x",
   "source_url": "https://omlx.ai/benchmarks/performance/bnqpqo0x",
   "tg_tokens": null,
   "tps": 84.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-bnqpqo0x-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: 6yyueta2, fqyyhltx, uxjubqh3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1624.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,624 137.9 18.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bnqpqo0x",
   "source_url": "https://omlx.ai/benchmarks/performance/bnqpqo0x",
   "tg_tokens": null,
   "tps": 137.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-bnqpqo0x-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 6yyueta2, fqyyhltx, uxjubqh3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1619.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 1,619 61.7 19.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bnqpqo0x",
   "source_url": "https://omlx.ai/benchmarks/performance/bnqpqo0x",
   "tg_tokens": null,
   "tps": 61.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-bnqpqo0x-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 6yyueta2, fqyyhltx, uxjubqh3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1566.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,566 73.8 19.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bnqpqo0x",
   "source_url": "https://omlx.ai/benchmarks/performance/bnqpqo0x",
   "tg_tokens": null,
   "tps": 73.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-bnqpqo0x-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: 6yyueta2, fqyyhltx, uxjubqh3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1453.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,453 116.5 20.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bnqpqo0x",
   "source_url": "https://omlx.ai/benchmarks/performance/bnqpqo0x",
   "tg_tokens": null,
   "tps": 116.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-bnqpqo0x-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: 6yyueta2, fqyyhltx, uxjubqh3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1255.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,255 77.1 23.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bnqpqo0x",
   "source_url": "https://omlx.ai/benchmarks/performance/bnqpqo0x",
   "tg_tokens": null,
   "tps": 77.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-bnqpqo0x-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: 6yyueta2, fqyyhltx, uxjubqh3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 978.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 978.9 77.6 27.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bnqpqo0x",
   "source_url": "https://omlx.ai/benchmarks/performance/bnqpqo0x",
   "tg_tokens": null,
   "tps": 77.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-bnqpqo0x-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 33.0 GB at this context; duplicate traces of the same session: 6yyueta2, fqyyhltx, uxjubqh3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 795.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 795.4 69.2 33.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bnqpqo0x",
   "source_url": "https://omlx.ai/benchmarks/performance/bnqpqo0x",
   "tg_tokens": null,
   "tps": 69.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jqtdxdfd-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 20.9 GB; TTFT 104776 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 625.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k tokens 625.5 PP tok/s 22.4 TG tok/s 104776 TTFT (ms) 20.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jqtdxdfd",
   "source_url": "https://omlx.ai/benchmarks/performance/jqtdxdfd",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-m1ykaep0-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 17.9 GB; TTFT 20275 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 808.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k tokens 808.1 PP tok/s 28.3 TG tok/s 20275 TTFT (ms) 17.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m1ykaep0",
   "source_url": "https://omlx.ai/benchmarks/performance/m1ykaep0",
   "tg_tokens": null,
   "tps": 28.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-e7ax1svx-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 420.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 420.5 73.1 21.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e7ax1svx",
   "source_url": "https://omlx.ai/benchmarks/performance/e7ax1svx",
   "tg_tokens": null,
   "tps": 73.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-e7ax1svx-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 559.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 559.9 111.1 25.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e7ax1svx",
   "source_url": "https://omlx.ai/benchmarks/performance/e7ax1svx",
   "tg_tokens": null,
   "tps": 111.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-e7ax1svx-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 551.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 551.5 54.4 25.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e7ax1svx",
   "source_url": "https://omlx.ai/benchmarks/performance/e7ax1svx",
   "tg_tokens": null,
   "tps": 54.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-e7ax1svx-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 526.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 526.8 76.3 26.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e7ax1svx",
   "source_url": "https://omlx.ai/benchmarks/performance/e7ax1svx",
   "tg_tokens": null,
   "tps": 76.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-t5e5p84z-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: f05iixoj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 899.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 899.1 67.5 24.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t5e5p84z",
   "source_url": "https://omlx.ai/benchmarks/performance/t5e5p84z",
   "tg_tokens": null,
   "tps": 67.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-t5e5p84z-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.7 GB at this context; duplicate traces of the same session: f05iixoj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1157.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 1,157 58.7 25.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t5e5p84z",
   "source_url": "https://omlx.ai/benchmarks/performance/t5e5p84z",
   "tg_tokens": null,
   "tps": 58.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-t5e5p84z-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: f05iixoj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1125.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 1,125 60.4 26.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t5e5p84z",
   "source_url": "https://omlx.ai/benchmarks/performance/t5e5p84z",
   "tg_tokens": null,
   "tps": 60.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-t5e5p84z-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: f05iixoj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1048.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 1,048 78.8 26.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t5e5p84z",
   "source_url": "https://omlx.ai/benchmarks/performance/t5e5p84z",
   "tg_tokens": null,
   "tps": 78.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-t5e5p84z-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: f05iixoj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 919.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 919.9 113.4 27.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t5e5p84z",
   "source_url": "https://omlx.ai/benchmarks/performance/t5e5p84z",
   "tg_tokens": null,
   "tps": 113.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-t5e5p84z-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.0 GB at this context; duplicate traces of the same session: f05iixoj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 754.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 754.1 39.1 30.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t5e5p84z",
   "source_url": "https://omlx.ai/benchmarks/performance/t5e5p84z",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-t5e5p84z-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: f05iixoj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 538.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 538.9 39.8 34.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t5e5p84z",
   "source_url": "https://omlx.ai/benchmarks/performance/t5e5p84z",
   "tg_tokens": null,
   "tps": 39.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pfhzn4ht-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: i8ref11h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 831.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 831.7 55.9 23.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pfhzn4ht",
   "source_url": "https://omlx.ai/benchmarks/performance/pfhzn4ht",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pfhzn4ht-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: i8ref11h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 886.6,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 886.6 53.0 24.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pfhzn4ht",
   "source_url": "https://omlx.ai/benchmarks/performance/pfhzn4ht",
   "tg_tokens": null,
   "tps": 53.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-pfhzn4ht-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: i8ref11h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 816.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 816.0 66.0 24.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pfhzn4ht",
   "source_url": "https://omlx.ai/benchmarks/performance/pfhzn4ht",
   "tg_tokens": null,
   "tps": 66.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pfhzn4ht-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.2 GB at this context; duplicate traces of the same session: i8ref11h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 783.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 783.9 75.3 25.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pfhzn4ht",
   "source_url": "https://omlx.ai/benchmarks/performance/pfhzn4ht",
   "tg_tokens": null,
   "tps": 75.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pfhzn4ht-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: i8ref11h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 676.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 676.4 113.5 26.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pfhzn4ht",
   "source_url": "https://omlx.ai/benchmarks/performance/pfhzn4ht",
   "tg_tokens": null,
   "tps": 113.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pfhzn4ht-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.7 GB at this context; duplicate traces of the same session: i8ref11h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 566.3,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 566.3 44.5 28.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pfhzn4ht",
   "source_url": "https://omlx.ai/benchmarks/performance/pfhzn4ht",
   "tg_tokens": null,
   "tps": 44.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-pfhzn4ht-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.4 GB at this context; duplicate traces of the same session: i8ref11h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 442.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 442.4 40.1 33.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pfhzn4ht",
   "source_url": "https://omlx.ai/benchmarks/performance/pfhzn4ht",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0d5mozkt-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: dzuk1zp6, td20xvnd, xznrsudm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 816.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 816.7 55.8 24.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0d5mozkt",
   "source_url": "https://omlx.ai/benchmarks/performance/0d5mozkt",
   "tg_tokens": null,
   "tps": 55.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0d5mozkt-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.7 GB at this context; duplicate traces of the same session: dzuk1zp6, td20xvnd, xznrsudm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 868.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 868.4 47.8 25.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0d5mozkt",
   "source_url": "https://omlx.ai/benchmarks/performance/0d5mozkt",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0d5mozkt-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: dzuk1zp6, td20xvnd, xznrsudm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 951.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 951.9 53.4 26.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0d5mozkt",
   "source_url": "https://omlx.ai/benchmarks/performance/0d5mozkt",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0d5mozkt-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: dzuk1zp6, td20xvnd, xznrsudm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 961.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 961.9 84.9 26.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0d5mozkt",
   "source_url": "https://omlx.ai/benchmarks/performance/0d5mozkt",
   "tg_tokens": null,
   "tps": 84.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0d5mozkt-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: dzuk1zp6, td20xvnd, xznrsudm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 891.8,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 891.8 105.7 27.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0d5mozkt",
   "source_url": "https://omlx.ai/benchmarks/performance/0d5mozkt",
   "tg_tokens": null,
   "tps": 105.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0d5mozkt-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.0 GB at this context; duplicate traces of the same session: dzuk1zp6, td20xvnd, xznrsudm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 740.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 740.2 43.7 30.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0d5mozkt",
   "source_url": "https://omlx.ai/benchmarks/performance/0d5mozkt",
   "tg_tokens": null,
   "tps": 43.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0d5mozkt-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: dzuk1zp6, td20xvnd, xznrsudm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 554.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 554.4 41.8 34.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0d5mozkt",
   "source_url": "https://omlx.ai/benchmarks/performance/0d5mozkt",
   "tg_tokens": null,
   "tps": 41.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-vri45koj-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 33.9 GB; TTFT 3197 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 320.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k tokens 320.3 PP tok/s 50.2 TG tok/s 3197 TTFT (ms) 33.9 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vri45koj",
   "source_url": "https://omlx.ai/benchmarks/performance/vri45koj",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-oig9lb20-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 86.9 GB at this context; duplicate traces of the same session: ju90orxd; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 486.9,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.8-Flash-Next-MLX-3bit-MTP on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 3bit; 1k 486.9 26.3 86.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oig9lb20",
   "source_url": "https://omlx.ai/benchmarks/performance/oig9lb20",
   "tg_tokens": null,
   "tps": 26.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-oig9lb20-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 87.0 GB at this context; duplicate traces of the same session: ju90orxd; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 607.0,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.8-Flash-Next-MLX-3bit-MTP on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 3bit; 4k 607.0 20.3 87.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oig9lb20",
   "source_url": "https://omlx.ai/benchmarks/performance/oig9lb20",
   "tg_tokens": null,
   "tps": 20.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-vcl2u96e-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 63.0 GB; TTFT 4004 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 255.7,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.8-Flash-Next-MLX-3bit-MTP on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 3bit; 1k tokens 255.7 PP tok/s 23.4 TG tok/s 4004 TTFT (ms) 63.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vcl2u96e",
   "source_url": "https://omlx.ai/benchmarks/performance/vcl2u96e",
   "tg_tokens": null,
   "tps": 23.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ksz9bfj4-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: huise6v0; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 403.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 1k 403.8 37.5 70.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ksz9bfj4",
   "source_url": "https://omlx.ai/benchmarks/performance/ksz9bfj4",
   "tg_tokens": null,
   "tps": 37.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ksz9bfj4-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: huise6v0; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 483.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 4k 483.2 38.0 71.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ksz9bfj4",
   "source_url": "https://omlx.ai/benchmarks/performance/ksz9bfj4",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ksz9bfj4-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.1 GB at this context; duplicate traces of the same session: huise6v0; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 502.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 16k 502.5 32.1 72.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ksz9bfj4",
   "source_url": "https://omlx.ai/benchmarks/performance/ksz9bfj4",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ksz9bfj4-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.1 GB at this context; duplicate traces of the same session: huise6v0; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 474.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 64k 474.2 30.1 74.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ksz9bfj4",
   "source_url": "https://omlx.ai/benchmarks/performance/ksz9bfj4",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ksz9bfj4-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.8 GB at this context; duplicate traces of the same session: huise6v0; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 376.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 128k 376.4 27.8 78.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ksz9bfj4",
   "source_url": "https://omlx.ai/benchmarks/performance/ksz9bfj4",
   "tg_tokens": null,
   "tps": 27.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-p0ds1dsz-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2443.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,443 79.6 71.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p0ds1dsz",
   "source_url": "https://omlx.ai/benchmarks/performance/p0ds1dsz",
   "tg_tokens": null,
   "tps": 79.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-p0ds1dsz-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2429.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,429 67.2 78.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p0ds1dsz",
   "source_url": "https://omlx.ai/benchmarks/performance/p0ds1dsz",
   "tg_tokens": null,
   "tps": 67.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-vcn5b7c9-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.2 GB at this context; duplicate traces of the same session: 4ssye3nk, jbsn8l4d, n68n803o, qj4vilpx; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 811.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 1k 811.9 45.3 183.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vcn5b7c9",
   "source_url": "https://omlx.ai/benchmarks/performance/vcn5b7c9",
   "tg_tokens": null,
   "tps": 45.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-vcn5b7c9-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.6 GB at this context; duplicate traces of the same session: 4ssye3nk, jbsn8l4d, n68n803o, qj4vilpx; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2420.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 4k 2,420 117.1 183.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vcn5b7c9",
   "source_url": "https://omlx.ai/benchmarks/performance/vcn5b7c9",
   "tg_tokens": null,
   "tps": 117.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-vcn5b7c9-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 184.7 GB at this context; duplicate traces of the same session: 4ssye3nk, jbsn8l4d, n68n803o, qj4vilpx; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2716.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 8k 2,716 111.0 184.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vcn5b7c9",
   "source_url": "https://omlx.ai/benchmarks/performance/vcn5b7c9",
   "tg_tokens": null,
   "tps": 111.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-vcn5b7c9-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 186.9 GB at this context; duplicate traces of the same session: 4ssye3nk, jbsn8l4d, n68n803o, qj4vilpx; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4444.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 64k 4,444 114.8 186.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vcn5b7c9",
   "source_url": "https://omlx.ai/benchmarks/performance/vcn5b7c9",
   "tg_tokens": null,
   "tps": 114.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-vcn5b7c9-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 191.6 GB at this context; duplicate traces of the same session: 4ssye3nk, jbsn8l4d, n68n803o, qj4vilpx; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4314.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 128k 4,314 90.2 191.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vcn5b7c9",
   "source_url": "https://omlx.ai/benchmarks/performance/vcn5b7c9",
   "tg_tokens": null,
   "tps": 90.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-vcn5b7c9-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 193.3 GB at this context; duplicate traces of the same session: 4ssye3nk, jbsn8l4d, n68n803o, qj4vilpx; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4189.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 195k 4,189 78.1 193.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vcn5b7c9",
   "source_url": "https://omlx.ai/benchmarks/performance/vcn5b7c9",
   "tg_tokens": null,
   "tps": 78.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-6xmd1ie5-ctx8192",
   "model": "Qwopus3.8-27B-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: iuoch7w4; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 492.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwopus3.8-27B-Flash-oQ4e-mtp-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 492.7 16.1 18.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6xmd1ie5",
   "source_url": "https://omlx.ai/benchmarks/performance/6xmd1ie5",
   "tg_tokens": null,
   "tps": 16.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-6xmd1ie5-ctx16384",
   "model": "Qwopus3.8-27B-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: iuoch7w4; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 493.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwopus3.8-27B-Flash-oQ4e-mtp-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 493.4 16.4 19.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6xmd1ie5",
   "source_url": "https://omlx.ai/benchmarks/performance/6xmd1ie5",
   "tg_tokens": null,
   "tps": 16.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-6xmd1ie5-ctx32768",
   "model": "Qwopus3.8-27B-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: iuoch7w4; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 446.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwopus3.8-27B-Flash-oQ4e-mtp-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 446.5 14.1 21.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6xmd1ie5",
   "source_url": "https://omlx.ai/benchmarks/performance/6xmd1ie5",
   "tg_tokens": null,
   "tps": 14.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-zaerwmel-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: 8xct0lop; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 455.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 455.1 25.6 19.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zaerwmel",
   "source_url": "https://omlx.ai/benchmarks/performance/zaerwmel",
   "tg_tokens": null,
   "tps": 25.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-zaerwmel-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 8xct0lop; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 441.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 441.9 33.4 20.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zaerwmel",
   "source_url": "https://omlx.ai/benchmarks/performance/zaerwmel",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-zaerwmel-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: 8xct0lop; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 402.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 402.5 30.6 22.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zaerwmel",
   "source_url": "https://omlx.ai/benchmarks/performance/zaerwmel",
   "tg_tokens": null,
   "tps": 30.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qy3kieah-ctx8192",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: ogbtfgrc; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3098.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 3,098 91.2 22.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qy3kieah",
   "source_url": "https://omlx.ai/benchmarks/performance/qy3kieah",
   "tg_tokens": null,
   "tps": 91.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qy3kieah-ctx16384",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: ogbtfgrc; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2805.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 2,805 87.5 22.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qy3kieah",
   "source_url": "https://omlx.ai/benchmarks/performance/qy3kieah",
   "tg_tokens": null,
   "tps": 87.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qy3kieah-ctx32768",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: ogbtfgrc; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2321.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 2,321 87.9 23.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qy3kieah",
   "source_url": "https://omlx.ai/benchmarks/performance/qy3kieah",
   "tg_tokens": null,
   "tps": 87.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-121sd4zt-ctx8192",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 1zi1ag5u, oqmh7fhy; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1923.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 1,923 47.4 22.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/121sd4zt",
   "source_url": "https://omlx.ai/benchmarks/performance/121sd4zt",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-121sd4zt-ctx16384",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 1zi1ag5u, oqmh7fhy; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2790.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 2,790 81.8 22.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/121sd4zt",
   "source_url": "https://omlx.ai/benchmarks/performance/121sd4zt",
   "tg_tokens": null,
   "tps": 81.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-121sd4zt-ctx32768",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 1zi1ag5u, oqmh7fhy; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2398.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 2,398 102.6 23.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/121sd4zt",
   "source_url": "https://omlx.ai/benchmarks/performance/121sd4zt",
   "tg_tokens": null,
   "tps": 102.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-121sd4zt-ctx199680",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 32.2 GB at this context; duplicate traces of the same session: 1zi1ag5u, oqmh7fhy; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 913.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 195k 913.1 54.3 32.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/121sd4zt",
   "source_url": "https://omlx.ai/benchmarks/performance/121sd4zt",
   "tg_tokens": null,
   "tps": 54.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-tsoiobp2-ctx1024",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 46sqmiqt, g9qikzc8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 207.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 207.0 28.5 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tsoiobp2",
   "source_url": "https://omlx.ai/benchmarks/performance/tsoiobp2",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-tsoiobp2-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 46sqmiqt, g9qikzc8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 204.4 46.7 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tsoiobp2",
   "source_url": "https://omlx.ai/benchmarks/performance/tsoiobp2",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-tsoiobp2-ctx8192",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 46sqmiqt, g9qikzc8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 193.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 193.4 30.7 21.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tsoiobp2",
   "source_url": "https://omlx.ai/benchmarks/performance/tsoiobp2",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-tsoiobp2-ctx16384",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: 46sqmiqt, g9qikzc8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 193.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 193.4 39.1 23.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tsoiobp2",
   "source_url": "https://omlx.ai/benchmarks/performance/tsoiobp2",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-tsoiobp2-ctx32768",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: 46sqmiqt, g9qikzc8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 179.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 179.7 26.2 25.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tsoiobp2",
   "source_url": "https://omlx.ai/benchmarks/performance/tsoiobp2",
   "tg_tokens": null,
   "tps": 26.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-tsoiobp2-ctx65536",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: 46sqmiqt, g9qikzc8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 158.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 158.0 20.1 29.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tsoiobp2",
   "source_url": "https://omlx.ai/benchmarks/performance/tsoiobp2",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-010cslxi-ctx1024",
   "model": "gemma-4-31B-it-uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 32.0 GB; TTFT 3886 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 263.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-31B-it-uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k tokens 263.5 PP tok/s 19.2 TG tok/s 3886 TTFT (ms) 32.0 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/010cslxi",
   "source_url": "https://omlx.ai/benchmarks/performance/010cslxi",
   "tg_tokens": null,
   "tps": 19.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-zfao0k0u-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Noctalin build; fp16 MTP; oMLX community benchmark trace, single run at 64k context; peak memory 24.4 GB; TTFT 253846 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 258.2,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "models--Noctalin--Qwen3.6-35B-A3B-oQ4-fp16-mtp on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 64k tokens 258.2 PP tok/s 28.2 TG tok/s 253846 TTFT (ms) 24.4 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zfao0k0u",
   "source_url": "https://omlx.ai/benchmarks/performance/zfao0k0u",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-gruub484-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: qjrpvtv1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3808.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 3,808 115.7 27.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gruub484",
   "source_url": "https://omlx.ai/benchmarks/performance/gruub484",
   "tg_tokens": null,
   "tps": 115.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-gruub484-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: qjrpvtv1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4655.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 4,655 114.4 28.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gruub484",
   "source_url": "https://omlx.ai/benchmarks/performance/gruub484",
   "tg_tokens": null,
   "tps": 114.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-x0a3uvid-ctx1024",
   "model": "DeepSeek-V4-Flash",
   "notes": "oMLX community benchmark trace; peak memory 26.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 109.1,
   "provenance": "community",
   "quant": "2bit DQ",
   "quote": "DeepSeek-V4-Flash-2bit-DQ on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 109.1 13.6 26.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x0a3uvid",
   "source_url": "https://omlx.ai/benchmarks/performance/x0a3uvid",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-x0a3uvid-ctx4096",
   "model": "DeepSeek-V4-Flash",
   "notes": "oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 180.0,
   "provenance": "community",
   "quant": "2bit DQ",
   "quote": "DeepSeek-V4-Flash-2bit-DQ on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 180.0 9.1 27.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x0a3uvid",
   "source_url": "https://omlx.ai/benchmarks/performance/x0a3uvid",
   "tg_tokens": null,
   "tps": 9.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ndkc0khk-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.6 GB at this context; duplicate traces of the same session: 261k5swf, kim3l18j; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 395.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 395.6 31.5 173.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ndkc0khk",
   "source_url": "https://omlx.ai/benchmarks/performance/ndkc0khk",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ndkc0khk-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.7 GB at this context; duplicate traces of the same session: 261k5swf, kim3l18j; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 557.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 557.8 27.3 174.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ndkc0khk",
   "source_url": "https://omlx.ai/benchmarks/performance/ndkc0khk",
   "tg_tokens": null,
   "tps": 27.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ndkc0khk-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.0 GB at this context; duplicate traces of the same session: 261k5swf, kim3l18j; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 506.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 506.7 27.3 175.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ndkc0khk",
   "source_url": "https://omlx.ai/benchmarks/performance/ndkc0khk",
   "tg_tokens": null,
   "tps": 27.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ndkc0khk-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.1 GB at this context; duplicate traces of the same session: 261k5swf, kim3l18j; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 537.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 537.0 27.0 175.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ndkc0khk",
   "source_url": "https://omlx.ai/benchmarks/performance/ndkc0khk",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ndkc0khk-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.2 GB at this context; duplicate traces of the same session: 261k5swf, kim3l18j; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 528.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 528.1 26.9 175.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ndkc0khk",
   "source_url": "https://omlx.ai/benchmarks/performance/ndkc0khk",
   "tg_tokens": null,
   "tps": 26.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ndkc0khk-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.6 GB at this context; duplicate traces of the same session: 261k5swf, kim3l18j; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 518.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 518.7 25.2 175.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ndkc0khk",
   "source_url": "https://omlx.ai/benchmarks/performance/ndkc0khk",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ndkc0khk-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 176.3 GB at this context; duplicate traces of the same session: 261k5swf, kim3l18j; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 509.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 509.0 21.5 176.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ndkc0khk",
   "source_url": "https://omlx.ai/benchmarks/performance/ndkc0khk",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-cixry28q-ctx1024",
   "model": "Ling-3.0-tiny",
   "notes": "oMLX community benchmark trace; peak memory 5.1 GB at this context; duplicate traces of the same session: none; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 830.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-tiny-oQ4e on M4 (8c); M4 (8c) \u00b7 16 GB \u00b7 4bit; 1k 830.1 95.3 5.1 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cixry28q",
   "source_url": "https://omlx.ai/benchmarks/performance/cixry28q",
   "tg_tokens": null,
   "tps": 95.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-cixry28q-ctx4096",
   "model": "Ling-3.0-tiny",
   "notes": "oMLX community benchmark trace; peak memory 5.5 GB at this context; duplicate traces of the same session: none; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 957.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-tiny-oQ4e on M4 (8c); M4 (8c) \u00b7 16 GB \u00b7 4bit; 4k 957.3 88.1 5.5 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cixry28q",
   "source_url": "https://omlx.ai/benchmarks/performance/cixry28q",
   "tg_tokens": null,
   "tps": 88.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-n5h53nac-ctx16384",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: jaqk3sjr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1441.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 1,441 96.7 34.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n5h53nac",
   "source_url": "https://omlx.ai/benchmarks/performance/n5h53nac",
   "tg_tokens": null,
   "tps": 96.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-n5h53nac-ctx32768",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: jaqk3sjr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1204.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 1,204 90.9 34.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n5h53nac",
   "source_url": "https://omlx.ai/benchmarks/performance/n5h53nac",
   "tg_tokens": null,
   "tps": 90.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-rsevx220-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Noctalin build; fp16 MTP; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: bsx5dptd, udvp547j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1312.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Noctalin--Qwen3.6-35B-A3B-oQ4-fp16-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,312 99.8 20.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rsevx220",
   "source_url": "https://omlx.ai/benchmarks/performance/rsevx220",
   "tg_tokens": null,
   "tps": 99.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-rsevx220-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Noctalin build; fp16 MTP; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: bsx5dptd, udvp547j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1621.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Noctalin--Qwen3.6-35B-A3B-oQ4-fp16-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,621 86.6 22.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rsevx220",
   "source_url": "https://omlx.ai/benchmarks/performance/rsevx220",
   "tg_tokens": null,
   "tps": 86.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-rsevx220-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Noctalin build; fp16 MTP; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: bsx5dptd, udvp547j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1387.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Noctalin--Qwen3.6-35B-A3B-oQ4-fp16-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 1,387 79.4 23.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rsevx220",
   "source_url": "https://omlx.ai/benchmarks/performance/rsevx220",
   "tg_tokens": null,
   "tps": 79.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9r3ze2f9-ctx1024",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: mqw7bjz2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4438.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 4,438 136.5 19.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9r3ze2f9",
   "source_url": "https://omlx.ai/benchmarks/performance/9r3ze2f9",
   "tg_tokens": null,
   "tps": 136.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9r3ze2f9-ctx32768",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: mqw7bjz2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4546.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 4,546 113.3 20.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9r3ze2f9",
   "source_url": "https://omlx.ai/benchmarks/performance/9r3ze2f9",
   "tg_tokens": null,
   "tps": 113.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9r3ze2f9-ctx65536",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: mqw7bjz2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3366.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 64k 3,366 99.7 21.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9r3ze2f9",
   "source_url": "https://omlx.ai/benchmarks/performance/9r3ze2f9",
   "tg_tokens": null,
   "tps": 99.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9r3ze2f9-ctx131072",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: mqw7bjz2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2252.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 128k 2,252 76.4 22.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9r3ze2f9",
   "source_url": "https://omlx.ai/benchmarks/performance/9r3ze2f9",
   "tg_tokens": null,
   "tps": 76.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9r3ze2f9-ctx199680",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: mqw7bjz2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1576.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 195k 1,576 58.2 23.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9r3ze2f9",
   "source_url": "https://omlx.ai/benchmarks/performance/9r3ze2f9",
   "tg_tokens": null,
   "tps": 58.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-tj22e7jo-ctx1024",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: ae4gu2lz, ebfmytxb, llmaeu4o, xzacwcec, z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4163.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 4,163 192.7 19.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tj22e7jo",
   "source_url": "https://omlx.ai/benchmarks/performance/tj22e7jo",
   "tg_tokens": null,
   "tps": 192.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-tj22e7jo-ctx4096",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: ae4gu2lz, ebfmytxb, llmaeu4o, xzacwcec, z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5736.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 5,736 159.3 20.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tj22e7jo",
   "source_url": "https://omlx.ai/benchmarks/performance/tj22e7jo",
   "tg_tokens": null,
   "tps": 159.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-tj22e7jo-ctx8192",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: ae4gu2lz, ebfmytxb, llmaeu4o, xzacwcec, z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6112.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 8k 6,112 203.0 20.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tj22e7jo",
   "source_url": "https://omlx.ai/benchmarks/performance/tj22e7jo",
   "tg_tokens": null,
   "tps": 203.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-tj22e7jo-ctx16384",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: ae4gu2lz, ebfmytxb, llmaeu4o, xzacwcec, z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5607.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 5,607 195.0 20.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tj22e7jo",
   "source_url": "https://omlx.ai/benchmarks/performance/tj22e7jo",
   "tg_tokens": null,
   "tps": 195.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-tj22e7jo-ctx32768",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: ae4gu2lz, ebfmytxb, llmaeu4o, xzacwcec, z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4807.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 4,807 170.8 21.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tj22e7jo",
   "source_url": "https://omlx.ai/benchmarks/performance/tj22e7jo",
   "tg_tokens": null,
   "tps": 170.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-tj22e7jo-ctx65536",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: ae4gu2lz, ebfmytxb, llmaeu4o, xzacwcec, z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3580.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 64k 3,580 126.1 22.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tj22e7jo",
   "source_url": "https://omlx.ai/benchmarks/performance/tj22e7jo",
   "tg_tokens": null,
   "tps": 126.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-tj22e7jo-ctx131072",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: ae4gu2lz, ebfmytxb, llmaeu4o, xzacwcec, z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2310.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 128k 2,310 87.1 23.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tj22e7jo",
   "source_url": "https://omlx.ai/benchmarks/performance/tj22e7jo",
   "tg_tokens": null,
   "tps": 87.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-tj22e7jo-ctx199680",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: ae4gu2lz, ebfmytxb, llmaeu4o, xzacwcec, z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1579.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 195k 1,579 59.7 25.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tj22e7jo",
   "source_url": "https://omlx.ai/benchmarks/performance/tj22e7jo",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8smq47fl-ctx1024",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 007jy9ug, 3cy8ezo9, 90yw4img, esbg3qoh, ur3j8yno, yu64gvip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2615.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 2,615 151.1 19.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8smq47fl",
   "source_url": "https://omlx.ai/benchmarks/performance/8smq47fl",
   "tg_tokens": null,
   "tps": 151.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8smq47fl-ctx4096",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: 007jy9ug, 3cy8ezo9, 90yw4img, esbg3qoh, ur3j8yno, yu64gvip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2684.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 2,684 173.6 20.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8smq47fl",
   "source_url": "https://omlx.ai/benchmarks/performance/8smq47fl",
   "tg_tokens": null,
   "tps": 173.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8smq47fl-ctx8192",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 007jy9ug, 3cy8ezo9, 90yw4img, esbg3qoh, ur3j8yno, yu64gvip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4567.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 8k 4,567 190.8 20.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8smq47fl",
   "source_url": "https://omlx.ai/benchmarks/performance/8smq47fl",
   "tg_tokens": null,
   "tps": 190.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8smq47fl-ctx16384",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 007jy9ug, 3cy8ezo9, 90yw4img, esbg3qoh, ur3j8yno, yu64gvip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4610.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 4,610 177.6 20.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8smq47fl",
   "source_url": "https://omlx.ai/benchmarks/performance/8smq47fl",
   "tg_tokens": null,
   "tps": 177.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8smq47fl-ctx32768",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 007jy9ug, 3cy8ezo9, 90yw4img, esbg3qoh, ur3j8yno, yu64gvip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4019.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 4,019 164.2 21.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8smq47fl",
   "source_url": "https://omlx.ai/benchmarks/performance/8smq47fl",
   "tg_tokens": null,
   "tps": 164.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8smq47fl-ctx65536",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 007jy9ug, 3cy8ezo9, 90yw4img, esbg3qoh, ur3j8yno, yu64gvip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2973.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 64k 2,973 121.2 21.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8smq47fl",
   "source_url": "https://omlx.ai/benchmarks/performance/8smq47fl",
   "tg_tokens": null,
   "tps": 121.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8smq47fl-ctx131072",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 007jy9ug, 3cy8ezo9, 90yw4img, esbg3qoh, ur3j8yno, yu64gvip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2042.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 128k 2,042 91.8 23.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8smq47fl",
   "source_url": "https://omlx.ai/benchmarks/performance/8smq47fl",
   "tg_tokens": null,
   "tps": 91.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8smq47fl-ctx199680",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: 007jy9ug, 3cy8ezo9, 90yw4img, esbg3qoh, ur3j8yno, yu64gvip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1432.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 195k 1,432 62.6 24.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8smq47fl",
   "source_url": "https://omlx.ai/benchmarks/performance/8smq47fl",
   "tg_tokens": null,
   "tps": 62.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-woqc1w7p-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: z9qsd51n; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 822.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 822.4 86.2 20.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/woqc1w7p",
   "source_url": "https://omlx.ai/benchmarks/performance/woqc1w7p",
   "tg_tokens": null,
   "tps": 86.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-woqc1w7p-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: z9qsd51n; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 703.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 703.5 73.2 20.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/woqc1w7p",
   "source_url": "https://omlx.ai/benchmarks/performance/woqc1w7p",
   "tg_tokens": null,
   "tps": 73.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-woqc1w7p-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: z9qsd51n; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 286.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 128k 286.4 33.4 24.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/woqc1w7p",
   "source_url": "https://omlx.ai/benchmarks/performance/woqc1w7p",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3sqfr2pw-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: c8x51isy, dfnf3lrx, lesilos7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2967.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 2,967 202.6 26.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3sqfr2pw",
   "source_url": "https://omlx.ai/benchmarks/performance/3sqfr2pw",
   "tg_tokens": null,
   "tps": 202.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3sqfr2pw-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: c8x51isy, dfnf3lrx, lesilos7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4950.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 4,950 158.3 26.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3sqfr2pw",
   "source_url": "https://omlx.ai/benchmarks/performance/3sqfr2pw",
   "tg_tokens": null,
   "tps": 158.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3sqfr2pw-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: c8x51isy, dfnf3lrx, lesilos7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5350.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 5,350 180.0 26.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3sqfr2pw",
   "source_url": "https://omlx.ai/benchmarks/performance/3sqfr2pw",
   "tg_tokens": null,
   "tps": 180.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3sqfr2pw-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.0 GB at this context; duplicate traces of the same session: c8x51isy, dfnf3lrx, lesilos7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5012.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 5,012 211.3 27.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3sqfr2pw",
   "source_url": "https://omlx.ai/benchmarks/performance/3sqfr2pw",
   "tg_tokens": null,
   "tps": 211.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3sqfr2pw-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: c8x51isy, dfnf3lrx, lesilos7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4144.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 4,144 229.4 27.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3sqfr2pw",
   "source_url": "https://omlx.ai/benchmarks/performance/3sqfr2pw",
   "tg_tokens": null,
   "tps": 229.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3sqfr2pw-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: c8x51isy, dfnf3lrx, lesilos7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3088.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 3,088 144.9 28.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3sqfr2pw",
   "source_url": "https://omlx.ai/benchmarks/performance/3sqfr2pw",
   "tg_tokens": null,
   "tps": 144.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3sqfr2pw-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: c8x51isy, dfnf3lrx, lesilos7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2110.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 2,110 109.8 29.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3sqfr2pw",
   "source_url": "https://omlx.ai/benchmarks/performance/3sqfr2pw",
   "tg_tokens": null,
   "tps": 109.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-m9gw1nld-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 217.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 217.7 35.2 17.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m9gw1nld",
   "source_url": "https://omlx.ai/benchmarks/performance/m9gw1nld",
   "tg_tokens": null,
   "tps": 35.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-m9gw1nld-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 203.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 203.7 38.6 18.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m9gw1nld",
   "source_url": "https://omlx.ai/benchmarks/performance/m9gw1nld",
   "tg_tokens": null,
   "tps": 38.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-30fi9e5e-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 18.4 GB; TTFT 20753 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 197.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k tokens 197.4 PP tok/s 16.3 TG tok/s 20753 TTFT (ms) 18.4 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/30fi9e5e",
   "source_url": "https://omlx.ai/benchmarks/performance/30fi9e5e",
   "tg_tokens": null,
   "tps": 16.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-f31gykzy-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 432.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 432.1 20.2 18.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f31gykzy",
   "source_url": "https://omlx.ai/benchmarks/performance/f31gykzy",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-f31gykzy-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 428.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 428.3 26.1 19.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f31gykzy",
   "source_url": "https://omlx.ai/benchmarks/performance/f31gykzy",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-f31gykzy-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 404.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 404.3 24.0 22.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f31gykzy",
   "source_url": "https://omlx.ai/benchmarks/performance/f31gykzy",
   "tg_tokens": null,
   "tps": 24.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-f31gykzy-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 360.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 360.5 23.4 26.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f31gykzy",
   "source_url": "https://omlx.ai/benchmarks/performance/f31gykzy",
   "tg_tokens": null,
   "tps": 23.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-6yyueta2-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: fqyyhltx, uxjubqh3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1375.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,375 84.7 17.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6yyueta2",
   "source_url": "https://omlx.ai/benchmarks/performance/6yyueta2",
   "tg_tokens": null,
   "tps": 84.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-6yyueta2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: fqyyhltx, uxjubqh3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1624.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,624 137.9 18.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6yyueta2",
   "source_url": "https://omlx.ai/benchmarks/performance/6yyueta2",
   "tg_tokens": null,
   "tps": 137.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-6yyueta2-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: fqyyhltx, uxjubqh3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1619.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 1,619 61.7 19.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6yyueta2",
   "source_url": "https://omlx.ai/benchmarks/performance/6yyueta2",
   "tg_tokens": null,
   "tps": 61.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-6yyueta2-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: fqyyhltx, uxjubqh3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1566.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,566 73.8 19.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6yyueta2",
   "source_url": "https://omlx.ai/benchmarks/performance/6yyueta2",
   "tg_tokens": null,
   "tps": 73.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-6yyueta2-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: fqyyhltx, uxjubqh3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1453.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,453 116.5 20.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6yyueta2",
   "source_url": "https://omlx.ai/benchmarks/performance/6yyueta2",
   "tg_tokens": null,
   "tps": 116.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-6yyueta2-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: fqyyhltx, uxjubqh3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1255.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,255 77.1 23.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6yyueta2",
   "source_url": "https://omlx.ai/benchmarks/performance/6yyueta2",
   "tg_tokens": null,
   "tps": 77.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-6yyueta2-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: fqyyhltx, uxjubqh3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 978.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 978.9 77.6 27.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6yyueta2",
   "source_url": "https://omlx.ai/benchmarks/performance/6yyueta2",
   "tg_tokens": null,
   "tps": 77.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-6yyueta2-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 33.0 GB at this context; duplicate traces of the same session: fqyyhltx, uxjubqh3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 795.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 795.4 69.2 33.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6yyueta2",
   "source_url": "https://omlx.ai/benchmarks/performance/6yyueta2",
   "tg_tokens": null,
   "tps": 69.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-oepsk0sh-ctx4096",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: 21ezyi7d; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 497.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 497.1 76.0 25.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oepsk0sh",
   "source_url": "https://omlx.ai/benchmarks/performance/oepsk0sh",
   "tg_tokens": null,
   "tps": 76.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-oepsk0sh-ctx8192",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: 21ezyi7d; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 522.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 522.7 55.4 25.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oepsk0sh",
   "source_url": "https://omlx.ai/benchmarks/performance/oepsk0sh",
   "tg_tokens": null,
   "tps": 55.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ugy8cr2j-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 17.2 GB; TTFT 4169 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 982.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 982.6 PP tok/s 32.1 TG tok/s 4169 TTFT (ms) 17.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ugy8cr2j",
   "source_url": "https://omlx.ai/benchmarks/performance/ugy8cr2j",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1upl5onq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 825.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 8bit; 1k 825.5 18.5 28.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1upl5onq",
   "source_url": "https://omlx.ai/benchmarks/performance/1upl5onq",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1upl5onq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 946.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 8bit; 4k 946.0 18.4 29.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1upl5onq",
   "source_url": "https://omlx.ai/benchmarks/performance/1upl5onq",
   "tg_tokens": null,
   "tps": 18.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-31t90ejo-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: 36m1zzf6, 3sexbqow, hubzwapt; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 389.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 389.0 20.1 22.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/31t90ejo",
   "source_url": "https://omlx.ai/benchmarks/performance/31t90ejo",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-31t90ejo-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 36m1zzf6, 3sexbqow, hubzwapt; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 351.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 351.3 23.8 23.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/31t90ejo",
   "source_url": "https://omlx.ai/benchmarks/performance/31t90ejo",
   "tg_tokens": null,
   "tps": 23.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-31t90ejo-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: 36m1zzf6, 3sexbqow, hubzwapt; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 367.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 367.7 22.7 19.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/31t90ejo",
   "source_url": "https://omlx.ai/benchmarks/performance/31t90ejo",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-31t90ejo-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 36m1zzf6, 3sexbqow, hubzwapt; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 306.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 306.5 14.5 20.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/31t90ejo",
   "source_url": "https://omlx.ai/benchmarks/performance/31t90ejo",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5lyrjnlw-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 24.9 GB; TTFT 106234 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 616.9,
   "provenance": "community",
   "quant": "OptiQ 4bit",
   "quote": "Qwen3.8-27B-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k tokens 616.9 PP tok/s 16.6 TG tok/s 106234 TTFT (ms) 24.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5lyrjnlw",
   "source_url": "https://omlx.ai/benchmarks/performance/5lyrjnlw",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-bo371v3k-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 24.8 GB; TTFT 15827 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 258.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 258.8 PP tok/s 12.6 TG tok/s 15827 TTFT (ms) 24.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bo371v3k",
   "source_url": "https://omlx.ai/benchmarks/performance/bo371v3k",
   "tg_tokens": null,
   "tps": 12.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-f05iixoj-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 899.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 899.1 67.5 24.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f05iixoj",
   "source_url": "https://omlx.ai/benchmarks/performance/f05iixoj",
   "tg_tokens": null,
   "tps": 67.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-f05iixoj-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1157.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 1,157 58.7 25.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f05iixoj",
   "source_url": "https://omlx.ai/benchmarks/performance/f05iixoj",
   "tg_tokens": null,
   "tps": 58.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-f05iixoj-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1125.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 1,125 60.4 26.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f05iixoj",
   "source_url": "https://omlx.ai/benchmarks/performance/f05iixoj",
   "tg_tokens": null,
   "tps": 60.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-f05iixoj-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1048.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 1,048 78.8 26.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f05iixoj",
   "source_url": "https://omlx.ai/benchmarks/performance/f05iixoj",
   "tg_tokens": null,
   "tps": 78.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-f05iixoj-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 919.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 919.9 113.4 27.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f05iixoj",
   "source_url": "https://omlx.ai/benchmarks/performance/f05iixoj",
   "tg_tokens": null,
   "tps": 113.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-f05iixoj-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 754.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 754.1 39.1 30.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f05iixoj",
   "source_url": "https://omlx.ai/benchmarks/performance/f05iixoj",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-f05iixoj-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 538.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 538.9 39.8 34.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f05iixoj",
   "source_url": "https://omlx.ai/benchmarks/performance/f05iixoj",
   "tg_tokens": null,
   "tps": 39.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i8ref11h-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 831.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 831.7 55.9 23.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i8ref11h",
   "source_url": "https://omlx.ai/benchmarks/performance/i8ref11h",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i8ref11h-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 886.6,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 886.6 53.0 24.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i8ref11h",
   "source_url": "https://omlx.ai/benchmarks/performance/i8ref11h",
   "tg_tokens": null,
   "tps": 53.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-i8ref11h-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 816.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 816.0 66.0 24.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i8ref11h",
   "source_url": "https://omlx.ai/benchmarks/performance/i8ref11h",
   "tg_tokens": null,
   "tps": 66.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i8ref11h-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 783.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 783.9 75.3 25.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i8ref11h",
   "source_url": "https://omlx.ai/benchmarks/performance/i8ref11h",
   "tg_tokens": null,
   "tps": 75.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i8ref11h-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 676.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 676.4 113.5 26.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i8ref11h",
   "source_url": "https://omlx.ai/benchmarks/performance/i8ref11h",
   "tg_tokens": null,
   "tps": 113.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i8ref11h-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 566.3,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 566.3 44.5 28.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i8ref11h",
   "source_url": "https://omlx.ai/benchmarks/performance/i8ref11h",
   "tg_tokens": null,
   "tps": 44.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-i8ref11h-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 442.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 442.4 40.1 33.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i8ref11h",
   "source_url": "https://omlx.ai/benchmarks/performance/i8ref11h",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dzuk1zp6-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: td20xvnd, xznrsudm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 816.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 816.7 55.8 24.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dzuk1zp6",
   "source_url": "https://omlx.ai/benchmarks/performance/dzuk1zp6",
   "tg_tokens": null,
   "tps": 55.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dzuk1zp6-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.7 GB at this context; duplicate traces of the same session: td20xvnd, xznrsudm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 868.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 868.4 47.8 25.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dzuk1zp6",
   "source_url": "https://omlx.ai/benchmarks/performance/dzuk1zp6",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dzuk1zp6-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: td20xvnd, xznrsudm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 951.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 951.9 53.4 26.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dzuk1zp6",
   "source_url": "https://omlx.ai/benchmarks/performance/dzuk1zp6",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dzuk1zp6-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: td20xvnd, xznrsudm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 961.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 961.9 84.9 26.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dzuk1zp6",
   "source_url": "https://omlx.ai/benchmarks/performance/dzuk1zp6",
   "tg_tokens": null,
   "tps": 84.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dzuk1zp6-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: td20xvnd, xznrsudm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 891.8,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 891.8 105.7 27.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dzuk1zp6",
   "source_url": "https://omlx.ai/benchmarks/performance/dzuk1zp6",
   "tg_tokens": null,
   "tps": 105.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dzuk1zp6-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.0 GB at this context; duplicate traces of the same session: td20xvnd, xznrsudm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 740.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 740.2 43.7 30.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dzuk1zp6",
   "source_url": "https://omlx.ai/benchmarks/performance/dzuk1zp6",
   "tg_tokens": null,
   "tps": 43.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dzuk1zp6-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: td20xvnd, xznrsudm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 554.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 554.4 41.8 34.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dzuk1zp6",
   "source_url": "https://omlx.ai/benchmarks/performance/dzuk1zp6",
   "tg_tokens": null,
   "tps": 41.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-ju90orxd-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 86.9 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 486.9,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.8-Flash-Next-MLX-3bit-MTP on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 3bit; 1k 486.9 26.3 86.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ju90orxd",
   "source_url": "https://omlx.ai/benchmarks/performance/ju90orxd",
   "tg_tokens": null,
   "tps": 26.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-ju90orxd-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 87.0 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 607.0,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.8-Flash-Next-MLX-3bit-MTP on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 3bit; 4k 607.0 20.3 87.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ju90orxd",
   "source_url": "https://omlx.ai/benchmarks/performance/ju90orxd",
   "tg_tokens": null,
   "tps": 20.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-huise6v0-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 403.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 1k 403.8 37.5 70.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/huise6v0",
   "source_url": "https://omlx.ai/benchmarks/performance/huise6v0",
   "tg_tokens": null,
   "tps": 37.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-huise6v0-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 483.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 4k 483.2 38.0 71.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/huise6v0",
   "source_url": "https://omlx.ai/benchmarks/performance/huise6v0",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-huise6v0-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.1 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 502.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 16k 502.5 32.1 72.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/huise6v0",
   "source_url": "https://omlx.ai/benchmarks/performance/huise6v0",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-huise6v0-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.1 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 474.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 64k 474.2 30.1 74.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/huise6v0",
   "source_url": "https://omlx.ai/benchmarks/performance/huise6v0",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-huise6v0-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.8 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 376.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 128k 376.4 27.8 78.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/huise6v0",
   "source_url": "https://omlx.ai/benchmarks/performance/huise6v0",
   "tg_tokens": null,
   "tps": 27.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2jrhxdiy-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: j9p60nra; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2352.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,352 70.6 71.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2jrhxdiy",
   "source_url": "https://omlx.ai/benchmarks/performance/2jrhxdiy",
   "tg_tokens": null,
   "tps": 70.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2jrhxdiy-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.6 GB at this context; duplicate traces of the same session: j9p60nra; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2309.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,309 65.7 78.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2jrhxdiy",
   "source_url": "https://omlx.ai/benchmarks/performance/2jrhxdiy",
   "tg_tokens": null,
   "tps": 65.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-or90bng9-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2443.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,443 79.6 71.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/or90bng9",
   "source_url": "https://omlx.ai/benchmarks/performance/or90bng9",
   "tg_tokens": null,
   "tps": 79.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-or90bng9-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2429.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,429 67.2 78.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/or90bng9",
   "source_url": "https://omlx.ai/benchmarks/performance/or90bng9",
   "tg_tokens": null,
   "tps": 67.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-koc8jhb3-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.4 GB at this context; duplicate traces of the same session: kg6kfvix; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 144.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 144.2 13.6 33.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/koc8jhb3",
   "source_url": "https://omlx.ai/benchmarks/performance/koc8jhb3",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-koc8jhb3-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.1 GB at this context; duplicate traces of the same session: kg6kfvix; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 415.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 415.2 11.1 35.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/koc8jhb3",
   "source_url": "https://omlx.ai/benchmarks/performance/koc8jhb3",
   "tg_tokens": null,
   "tps": 11.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-n68n803o-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.2 GB at this context; duplicate traces of the same session: 4ssye3nk, jbsn8l4d, qj4vilpx; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 811.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 1k 811.9 45.3 183.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n68n803o",
   "source_url": "https://omlx.ai/benchmarks/performance/n68n803o",
   "tg_tokens": null,
   "tps": 45.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-n68n803o-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.6 GB at this context; duplicate traces of the same session: 4ssye3nk, jbsn8l4d, qj4vilpx; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2420.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 4k 2,420 117.1 183.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n68n803o",
   "source_url": "https://omlx.ai/benchmarks/performance/n68n803o",
   "tg_tokens": null,
   "tps": 117.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-n68n803o-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 184.7 GB at this context; duplicate traces of the same session: 4ssye3nk, jbsn8l4d, qj4vilpx; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2716.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 8k 2,716 111.0 184.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n68n803o",
   "source_url": "https://omlx.ai/benchmarks/performance/n68n803o",
   "tg_tokens": null,
   "tps": 111.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-n68n803o-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 186.9 GB at this context; duplicate traces of the same session: 4ssye3nk, jbsn8l4d, qj4vilpx; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4444.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 64k 4,444 114.8 186.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n68n803o",
   "source_url": "https://omlx.ai/benchmarks/performance/n68n803o",
   "tg_tokens": null,
   "tps": 114.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-n68n803o-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 191.6 GB at this context; duplicate traces of the same session: 4ssye3nk, jbsn8l4d, qj4vilpx; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4314.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 128k 4,314 90.2 191.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n68n803o",
   "source_url": "https://omlx.ai/benchmarks/performance/n68n803o",
   "tg_tokens": null,
   "tps": 90.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-n68n803o-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 193.3 GB at this context; duplicate traces of the same session: 4ssye3nk, jbsn8l4d, qj4vilpx; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4189.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 195k 4,189 78.1 193.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n68n803o",
   "source_url": "https://omlx.ai/benchmarks/performance/n68n803o",
   "tg_tokens": null,
   "tps": 78.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-iuoch7w4-ctx8192",
   "model": "Qwopus3.8-27B-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 492.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwopus3.8-27B-Flash-oQ4e-mtp-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 492.7 16.1 18.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iuoch7w4",
   "source_url": "https://omlx.ai/benchmarks/performance/iuoch7w4",
   "tg_tokens": null,
   "tps": 16.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-iuoch7w4-ctx16384",
   "model": "Qwopus3.8-27B-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 493.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwopus3.8-27B-Flash-oQ4e-mtp-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 493.4 16.4 19.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iuoch7w4",
   "source_url": "https://omlx.ai/benchmarks/performance/iuoch7w4",
   "tg_tokens": null,
   "tps": 16.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-iuoch7w4-ctx32768",
   "model": "Qwopus3.8-27B-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 446.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwopus3.8-27B-Flash-oQ4e-mtp-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 446.5 14.1 21.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iuoch7w4",
   "source_url": "https://omlx.ai/benchmarks/performance/iuoch7w4",
   "tg_tokens": null,
   "tps": 14.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-8xct0lop-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 455.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 455.1 25.6 19.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8xct0lop",
   "source_url": "https://omlx.ai/benchmarks/performance/8xct0lop",
   "tg_tokens": null,
   "tps": 25.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-8xct0lop-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 441.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 441.9 33.4 20.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8xct0lop",
   "source_url": "https://omlx.ai/benchmarks/performance/8xct0lop",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-8xct0lop-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 402.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 402.5 30.6 22.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8xct0lop",
   "source_url": "https://omlx.ai/benchmarks/performance/8xct0lop",
   "tg_tokens": null,
   "tps": 30.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ogbtfgrc-ctx8192",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3098.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 3,098 91.2 22.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ogbtfgrc",
   "source_url": "https://omlx.ai/benchmarks/performance/ogbtfgrc",
   "tg_tokens": null,
   "tps": 91.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ogbtfgrc-ctx16384",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2805.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 2,805 87.5 22.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ogbtfgrc",
   "source_url": "https://omlx.ai/benchmarks/performance/ogbtfgrc",
   "tg_tokens": null,
   "tps": 87.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ogbtfgrc-ctx32768",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2321.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 2,321 87.9 23.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ogbtfgrc",
   "source_url": "https://omlx.ai/benchmarks/performance/ogbtfgrc",
   "tg_tokens": null,
   "tps": 87.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1zi1ag5u-ctx8192",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: oqmh7fhy; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1923.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 1,923 47.4 22.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1zi1ag5u",
   "source_url": "https://omlx.ai/benchmarks/performance/1zi1ag5u",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-1zi1ag5u-ctx16384",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: oqmh7fhy; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2790.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 2,790 81.8 22.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1zi1ag5u",
   "source_url": "https://omlx.ai/benchmarks/performance/1zi1ag5u",
   "tg_tokens": null,
   "tps": 81.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1zi1ag5u-ctx32768",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: oqmh7fhy; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2398.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 2,398 102.6 23.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1zi1ag5u",
   "source_url": "https://omlx.ai/benchmarks/performance/1zi1ag5u",
   "tg_tokens": null,
   "tps": 102.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-1zi1ag5u-ctx199680",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 32.2 GB at this context; duplicate traces of the same session: oqmh7fhy; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 913.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 195k 913.1 54.3 32.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1zi1ag5u",
   "source_url": "https://omlx.ai/benchmarks/performance/1zi1ag5u",
   "tg_tokens": null,
   "tps": 54.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-g9qikzc8-ctx1024",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 46sqmiqt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 207.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 207.0 28.5 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g9qikzc8",
   "source_url": "https://omlx.ai/benchmarks/performance/g9qikzc8",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-g9qikzc8-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 46sqmiqt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 204.4 46.7 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g9qikzc8",
   "source_url": "https://omlx.ai/benchmarks/performance/g9qikzc8",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-g9qikzc8-ctx8192",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 46sqmiqt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 193.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 193.4 30.7 21.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g9qikzc8",
   "source_url": "https://omlx.ai/benchmarks/performance/g9qikzc8",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-g9qikzc8-ctx16384",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: 46sqmiqt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 193.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 193.4 39.1 23.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g9qikzc8",
   "source_url": "https://omlx.ai/benchmarks/performance/g9qikzc8",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-g9qikzc8-ctx32768",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: 46sqmiqt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 179.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 179.7 26.2 25.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g9qikzc8",
   "source_url": "https://omlx.ai/benchmarks/performance/g9qikzc8",
   "tg_tokens": null,
   "tps": 26.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-g9qikzc8-ctx65536",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: 46sqmiqt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 158.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 158.0 20.1 29.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g9qikzc8",
   "source_url": "https://omlx.ai/benchmarks/performance/g9qikzc8",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-ahypeav5-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Noctalin build; fp16 MTP; oMLX community benchmark trace, single run at 16k context; peak memory 22.1 GB; TTFT 36880 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 444.3,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "models--Noctalin--Qwen3.6-35B-A3B-oQ4-fp16-mtp on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k tokens 444.3 PP tok/s 28.1 TG tok/s 36880 TTFT (ms) 22.1 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ahypeav5",
   "source_url": "https://omlx.ai/benchmarks/performance/ahypeav5",
   "tg_tokens": null,
   "tps": 28.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-dm2wh1u1-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Noctalin build; fp16 MTP; oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: 4w55o9gk, 9fqq9ewh, v6qth67l; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 462.9,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "models--Noctalin--Qwen3.6-35B-A3B-oQ4-fp16-mtp on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 462.9 49.4 20.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dm2wh1u1",
   "source_url": "https://omlx.ai/benchmarks/performance/dm2wh1u1",
   "tg_tokens": null,
   "tps": 49.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-dm2wh1u1-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Noctalin build; fp16 MTP; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 4w55o9gk, 9fqq9ewh, v6qth67l; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 463.5,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "models--Noctalin--Qwen3.6-35B-A3B-oQ4-fp16-mtp on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k 463.5 42.4 21.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dm2wh1u1",
   "source_url": "https://omlx.ai/benchmarks/performance/dm2wh1u1",
   "tg_tokens": null,
   "tps": 42.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-dm2wh1u1-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Noctalin build; fp16 MTP; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: 4w55o9gk, 9fqq9ewh, v6qth67l; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 379.4,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "models--Noctalin--Qwen3.6-35B-A3B-oQ4-fp16-mtp on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 32k 379.4 37.6 22.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dm2wh1u1",
   "source_url": "https://omlx.ai/benchmarks/performance/dm2wh1u1",
   "tg_tokens": null,
   "tps": 37.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-dm2wh1u1-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Noctalin build; fp16 MTP; oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: 4w55o9gk, 9fqq9ewh, v6qth67l; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 275.4,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "models--Noctalin--Qwen3.6-35B-A3B-oQ4-fp16-mtp on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 64k 275.4 29.5 23.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dm2wh1u1",
   "source_url": "https://omlx.ai/benchmarks/performance/dm2wh1u1",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qjrpvtv1-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3808.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 3,808 115.7 27.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qjrpvtv1",
   "source_url": "https://omlx.ai/benchmarks/performance/qjrpvtv1",
   "tg_tokens": null,
   "tps": 115.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qjrpvtv1-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4655.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 4,655 114.4 28.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qjrpvtv1",
   "source_url": "https://omlx.ai/benchmarks/performance/qjrpvtv1",
   "tg_tokens": null,
   "tps": 114.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-rqzqu3c3-ctx1024",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: x0idxl8l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 247.2,
   "provenance": "community",
   "quant": "6bit",
   "quote": "AREX-2-6bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 247.2 20.2 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rqzqu3c3",
   "source_url": "https://omlx.ai/benchmarks/performance/rqzqu3c3",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-rqzqu3c3-ctx4096",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: x0idxl8l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 213.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "AREX-2-6bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 213.8 19.2 25.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rqzqu3c3",
   "source_url": "https://omlx.ai/benchmarks/performance/rqzqu3c3",
   "tg_tokens": null,
   "tps": 19.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-kim3l18j-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.6 GB at this context; duplicate traces of the same session: 261k5swf, 368tvpbg, hf9dlog8, qvarwslx, vyn1tdk5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 395.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 395.6 31.5 173.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kim3l18j",
   "source_url": "https://omlx.ai/benchmarks/performance/kim3l18j",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-kim3l18j-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.7 GB at this context; duplicate traces of the same session: 261k5swf, 368tvpbg, hf9dlog8, qvarwslx, vyn1tdk5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 557.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 557.8 27.3 174.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kim3l18j",
   "source_url": "https://omlx.ai/benchmarks/performance/kim3l18j",
   "tg_tokens": null,
   "tps": 27.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-kim3l18j-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.0 GB at this context; duplicate traces of the same session: 261k5swf, 368tvpbg, hf9dlog8, qvarwslx, vyn1tdk5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 506.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 506.7 27.3 175.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kim3l18j",
   "source_url": "https://omlx.ai/benchmarks/performance/kim3l18j",
   "tg_tokens": null,
   "tps": 27.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-kim3l18j-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.1 GB at this context; duplicate traces of the same session: 261k5swf, 368tvpbg, hf9dlog8, qvarwslx, vyn1tdk5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 537.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 537.0 27.0 175.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kim3l18j",
   "source_url": "https://omlx.ai/benchmarks/performance/kim3l18j",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-kim3l18j-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.2 GB at this context; duplicate traces of the same session: 261k5swf, 368tvpbg, hf9dlog8, qvarwslx, vyn1tdk5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 528.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 528.1 26.9 175.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kim3l18j",
   "source_url": "https://omlx.ai/benchmarks/performance/kim3l18j",
   "tg_tokens": null,
   "tps": 26.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-kim3l18j-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.6 GB at this context; duplicate traces of the same session: 261k5swf, 368tvpbg, hf9dlog8, qvarwslx, vyn1tdk5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 518.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 518.7 25.2 175.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kim3l18j",
   "source_url": "https://omlx.ai/benchmarks/performance/kim3l18j",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-kim3l18j-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 176.3 GB at this context; duplicate traces of the same session: 261k5swf, 368tvpbg, hf9dlog8, qvarwslx, vyn1tdk5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 509.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 509.0 21.5 176.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kim3l18j",
   "source_url": "https://omlx.ai/benchmarks/performance/kim3l18j",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-jaqk3sjr-ctx16384",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1441.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 1,441 96.7 34.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jaqk3sjr",
   "source_url": "https://omlx.ai/benchmarks/performance/jaqk3sjr",
   "tg_tokens": null,
   "tps": 96.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-jaqk3sjr-ctx32768",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1204.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 1,204 90.9 34.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jaqk3sjr",
   "source_url": "https://omlx.ai/benchmarks/performance/jaqk3sjr",
   "tg_tokens": null,
   "tps": 90.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-bsx5dptd-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Noctalin build; fp16 MTP; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: udvp547j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1312.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Noctalin--Qwen3.6-35B-A3B-oQ4-fp16-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,312 99.8 20.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bsx5dptd",
   "source_url": "https://omlx.ai/benchmarks/performance/bsx5dptd",
   "tg_tokens": null,
   "tps": 99.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-bsx5dptd-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Noctalin build; fp16 MTP; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: udvp547j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1621.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Noctalin--Qwen3.6-35B-A3B-oQ4-fp16-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,621 86.6 22.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bsx5dptd",
   "source_url": "https://omlx.ai/benchmarks/performance/bsx5dptd",
   "tg_tokens": null,
   "tps": 86.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-bsx5dptd-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Noctalin build; fp16 MTP; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: udvp547j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1387.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Noctalin--Qwen3.6-35B-A3B-oQ4-fp16-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 1,387 79.4 23.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bsx5dptd",
   "source_url": "https://omlx.ai/benchmarks/performance/bsx5dptd",
   "tg_tokens": null,
   "tps": 79.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-xcyitn13-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "fp16 weights; oMLX community benchmark trace, single run at 1k context; peak memory 5.9 GB; TTFT 4269 ms; duplicate traces of the same session: none; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 239.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-9B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 1k tokens 239.9 PP tok/s 34.4 TG tok/s 4269 TTFT (ms) 5.9 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xcyitn13",
   "source_url": "https://omlx.ai/benchmarks/performance/xcyitn13",
   "tg_tokens": null,
   "tps": 34.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-w4jp01mn-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "fp16 weights; oMLX community benchmark trace, single run at 1k context; peak memory 9.3 GB; TTFT 4240 ms; duplicate traces of the same session: none; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 241.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-9B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 1k tokens 241.5 PP tok/s 34.5 TG tok/s 4240 TTFT (ms) 9.3 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w4jp01mn",
   "source_url": "https://omlx.ai/benchmarks/performance/w4jp01mn",
   "tg_tokens": null,
   "tps": 34.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-ssgmdp3z-ctx4096",
   "model": "Qwen2.5-Coder-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: cafua26u; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1106.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-7B-Instruct-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,106 30.9 4.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ssgmdp3z",
   "source_url": "https://omlx.ai/benchmarks/performance/ssgmdp3z",
   "tg_tokens": null,
   "tps": 30.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-ssgmdp3z-ctx16384",
   "model": "Qwen2.5-Coder-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 5.5 GB at this context; duplicate traces of the same session: cafua26u; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 957.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-7B-Instruct-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 957.2 26.6 5.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ssgmdp3z",
   "source_url": "https://omlx.ai/benchmarks/performance/ssgmdp3z",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hdxarvex-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 38.6 GB; TTFT 71913 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1823.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k tokens 1,823 PP tok/s 53.3 TG tok/s 71913 TTFT (ms) 38.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hdxarvex",
   "source_url": "https://omlx.ai/benchmarks/performance/hdxarvex",
   "tg_tokens": null,
   "tps": 53.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-s6d8zh5p-ctx199680",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 195k context; peak memory 40.0 GB; TTFT 127404 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1570.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 195k tokens 1,570 PP tok/s 42.5 TG tok/s 127404 TTFT (ms) 40.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s6d8zh5p",
   "source_url": "https://omlx.ai/benchmarks/performance/s6d8zh5p",
   "tg_tokens": null,
   "tps": 42.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mqw7bjz2-ctx1024",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4438.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 4,438 136.5 19.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mqw7bjz2",
   "source_url": "https://omlx.ai/benchmarks/performance/mqw7bjz2",
   "tg_tokens": null,
   "tps": 136.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mqw7bjz2-ctx32768",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4546.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 4,546 113.3 20.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mqw7bjz2",
   "source_url": "https://omlx.ai/benchmarks/performance/mqw7bjz2",
   "tg_tokens": null,
   "tps": 113.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mqw7bjz2-ctx65536",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3366.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 64k 3,366 99.7 21.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mqw7bjz2",
   "source_url": "https://omlx.ai/benchmarks/performance/mqw7bjz2",
   "tg_tokens": null,
   "tps": 99.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mqw7bjz2-ctx131072",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2252.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 128k 2,252 76.4 22.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mqw7bjz2",
   "source_url": "https://omlx.ai/benchmarks/performance/mqw7bjz2",
   "tg_tokens": null,
   "tps": 76.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mqw7bjz2-ctx199680",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1576.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 195k 1,576 58.2 23.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mqw7bjz2",
   "source_url": "https://omlx.ai/benchmarks/performance/mqw7bjz2",
   "tg_tokens": null,
   "tps": 58.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ae4gu2lz-ctx1024",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: ebfmytxb, llmaeu4o, xzacwcec, z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4163.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 4,163 192.7 19.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ae4gu2lz",
   "source_url": "https://omlx.ai/benchmarks/performance/ae4gu2lz",
   "tg_tokens": null,
   "tps": 192.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ae4gu2lz-ctx4096",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: ebfmytxb, llmaeu4o, xzacwcec, z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5736.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 5,736 159.3 20.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ae4gu2lz",
   "source_url": "https://omlx.ai/benchmarks/performance/ae4gu2lz",
   "tg_tokens": null,
   "tps": 159.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ae4gu2lz-ctx8192",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: ebfmytxb, llmaeu4o, xzacwcec, z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6112.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 8k 6,112 203.0 20.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ae4gu2lz",
   "source_url": "https://omlx.ai/benchmarks/performance/ae4gu2lz",
   "tg_tokens": null,
   "tps": 203.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ae4gu2lz-ctx16384",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: ebfmytxb, llmaeu4o, xzacwcec, z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5607.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 5,607 195.0 20.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ae4gu2lz",
   "source_url": "https://omlx.ai/benchmarks/performance/ae4gu2lz",
   "tg_tokens": null,
   "tps": 195.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ae4gu2lz-ctx32768",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: ebfmytxb, llmaeu4o, xzacwcec, z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4807.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 4,807 170.8 21.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ae4gu2lz",
   "source_url": "https://omlx.ai/benchmarks/performance/ae4gu2lz",
   "tg_tokens": null,
   "tps": 170.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ae4gu2lz-ctx65536",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: ebfmytxb, llmaeu4o, xzacwcec, z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3580.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 64k 3,580 126.1 22.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ae4gu2lz",
   "source_url": "https://omlx.ai/benchmarks/performance/ae4gu2lz",
   "tg_tokens": null,
   "tps": 126.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ae4gu2lz-ctx131072",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: ebfmytxb, llmaeu4o, xzacwcec, z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2310.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 128k 2,310 87.1 23.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ae4gu2lz",
   "source_url": "https://omlx.ai/benchmarks/performance/ae4gu2lz",
   "tg_tokens": null,
   "tps": 87.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ae4gu2lz-ctx199680",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: ebfmytxb, llmaeu4o, xzacwcec, z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1579.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 195k 1,579 59.7 25.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ae4gu2lz",
   "source_url": "https://omlx.ai/benchmarks/performance/ae4gu2lz",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-90yw4img-ctx1024",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 007jy9ug, 3cy8ezo9, esbg3qoh, ur3j8yno, yu64gvip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2615.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 2,615 151.1 19.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/90yw4img",
   "source_url": "https://omlx.ai/benchmarks/performance/90yw4img",
   "tg_tokens": null,
   "tps": 151.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-90yw4img-ctx4096",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: 007jy9ug, 3cy8ezo9, esbg3qoh, ur3j8yno, yu64gvip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2684.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 2,684 173.6 20.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/90yw4img",
   "source_url": "https://omlx.ai/benchmarks/performance/90yw4img",
   "tg_tokens": null,
   "tps": 173.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-90yw4img-ctx8192",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 007jy9ug, 3cy8ezo9, esbg3qoh, ur3j8yno, yu64gvip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4567.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 8k 4,567 190.8 20.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/90yw4img",
   "source_url": "https://omlx.ai/benchmarks/performance/90yw4img",
   "tg_tokens": null,
   "tps": 190.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-90yw4img-ctx16384",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 007jy9ug, 3cy8ezo9, esbg3qoh, ur3j8yno, yu64gvip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4610.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 4,610 177.6 20.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/90yw4img",
   "source_url": "https://omlx.ai/benchmarks/performance/90yw4img",
   "tg_tokens": null,
   "tps": 177.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-90yw4img-ctx32768",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 007jy9ug, 3cy8ezo9, esbg3qoh, ur3j8yno, yu64gvip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4019.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 4,019 164.2 21.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/90yw4img",
   "source_url": "https://omlx.ai/benchmarks/performance/90yw4img",
   "tg_tokens": null,
   "tps": 164.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-90yw4img-ctx65536",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 007jy9ug, 3cy8ezo9, esbg3qoh, ur3j8yno, yu64gvip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2973.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 64k 2,973 121.2 21.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/90yw4img",
   "source_url": "https://omlx.ai/benchmarks/performance/90yw4img",
   "tg_tokens": null,
   "tps": 121.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-90yw4img-ctx131072",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 007jy9ug, 3cy8ezo9, esbg3qoh, ur3j8yno, yu64gvip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2042.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 128k 2,042 91.8 23.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/90yw4img",
   "source_url": "https://omlx.ai/benchmarks/performance/90yw4img",
   "tg_tokens": null,
   "tps": 91.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-90yw4img-ctx199680",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: 007jy9ug, 3cy8ezo9, esbg3qoh, ur3j8yno, yu64gvip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1432.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 195k 1,432 62.6 24.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/90yw4img",
   "source_url": "https://omlx.ai/benchmarks/performance/90yw4img",
   "tg_tokens": null,
   "tps": 62.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-z9qsd51n-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 822.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 822.4 86.2 20.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z9qsd51n",
   "source_url": "https://omlx.ai/benchmarks/performance/z9qsd51n",
   "tg_tokens": null,
   "tps": 86.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-z9qsd51n-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 703.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 703.5 73.2 20.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z9qsd51n",
   "source_url": "https://omlx.ai/benchmarks/performance/z9qsd51n",
   "tg_tokens": null,
   "tps": 73.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-z9qsd51n-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 286.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 128k 286.4 33.4 24.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z9qsd51n",
   "source_url": "https://omlx.ai/benchmarks/performance/z9qsd51n",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lesilos7-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: c8x51isy, dfnf3lrx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2967.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 2,967 202.6 26.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lesilos7",
   "source_url": "https://omlx.ai/benchmarks/performance/lesilos7",
   "tg_tokens": null,
   "tps": 202.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lesilos7-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: c8x51isy, dfnf3lrx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4950.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 4,950 158.3 26.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lesilos7",
   "source_url": "https://omlx.ai/benchmarks/performance/lesilos7",
   "tg_tokens": null,
   "tps": 158.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lesilos7-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: c8x51isy, dfnf3lrx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5350.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 5,350 180.0 26.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lesilos7",
   "source_url": "https://omlx.ai/benchmarks/performance/lesilos7",
   "tg_tokens": null,
   "tps": 180.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lesilos7-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.0 GB at this context; duplicate traces of the same session: c8x51isy, dfnf3lrx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5012.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 5,012 211.3 27.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lesilos7",
   "source_url": "https://omlx.ai/benchmarks/performance/lesilos7",
   "tg_tokens": null,
   "tps": 211.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lesilos7-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: c8x51isy, dfnf3lrx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4144.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 4,144 229.4 27.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lesilos7",
   "source_url": "https://omlx.ai/benchmarks/performance/lesilos7",
   "tg_tokens": null,
   "tps": 229.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lesilos7-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: c8x51isy, dfnf3lrx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3088.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 3,088 144.9 28.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lesilos7",
   "source_url": "https://omlx.ai/benchmarks/performance/lesilos7",
   "tg_tokens": null,
   "tps": 144.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lesilos7-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: c8x51isy, dfnf3lrx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2110.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 2,110 109.8 29.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lesilos7",
   "source_url": "https://omlx.ai/benchmarks/performance/lesilos7",
   "tg_tokens": null,
   "tps": 109.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-xlm9gujr-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 25.5 GB; TTFT 15473 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 264.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 264.7 PP tok/s 32.8 TG tok/s 15473 TTFT (ms) 25.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xlm9gujr",
   "source_url": "https://omlx.ai/benchmarks/performance/xlm9gujr",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-cfiolg0t-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 22.8 GB; TTFT 16057 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 255.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 255.1 PP tok/s 34.1 TG tok/s 16057 TTFT (ms) 22.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cfiolg0t",
   "source_url": "https://omlx.ai/benchmarks/performance/cfiolg0t",
   "tg_tokens": null,
   "tps": 34.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-svb2ond4-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: pwenazo1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 880.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 880.0 33.6 16.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/svb2ond4",
   "source_url": "https://omlx.ai/benchmarks/performance/svb2ond4",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-svb2ond4-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: pwenazo1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 994.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 994.3 33.0 17.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/svb2ond4",
   "source_url": "https://omlx.ai/benchmarks/performance/svb2ond4",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-fqyyhltx-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: uxjubqh3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1375.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,375 84.7 17.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fqyyhltx",
   "source_url": "https://omlx.ai/benchmarks/performance/fqyyhltx",
   "tg_tokens": null,
   "tps": 84.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-fqyyhltx-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: uxjubqh3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1624.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,624 137.9 18.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fqyyhltx",
   "source_url": "https://omlx.ai/benchmarks/performance/fqyyhltx",
   "tg_tokens": null,
   "tps": 137.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-fqyyhltx-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: uxjubqh3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1619.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 1,619 61.7 19.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fqyyhltx",
   "source_url": "https://omlx.ai/benchmarks/performance/fqyyhltx",
   "tg_tokens": null,
   "tps": 61.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-fqyyhltx-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: uxjubqh3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1566.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,566 73.8 19.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fqyyhltx",
   "source_url": "https://omlx.ai/benchmarks/performance/fqyyhltx",
   "tg_tokens": null,
   "tps": 73.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-fqyyhltx-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: uxjubqh3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1453.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,453 116.5 20.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fqyyhltx",
   "source_url": "https://omlx.ai/benchmarks/performance/fqyyhltx",
   "tg_tokens": null,
   "tps": 116.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-fqyyhltx-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: uxjubqh3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1255.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,255 77.1 23.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fqyyhltx",
   "source_url": "https://omlx.ai/benchmarks/performance/fqyyhltx",
   "tg_tokens": null,
   "tps": 77.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-fqyyhltx-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: uxjubqh3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 978.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 978.9 77.6 27.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fqyyhltx",
   "source_url": "https://omlx.ai/benchmarks/performance/fqyyhltx",
   "tg_tokens": null,
   "tps": 77.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-fqyyhltx-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 33.0 GB at this context; duplicate traces of the same session: uxjubqh3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 795.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 795.4 69.2 33.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fqyyhltx",
   "source_url": "https://omlx.ai/benchmarks/performance/fqyyhltx",
   "tg_tokens": null,
   "tps": 69.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-21ezyi7d-ctx4096",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 497.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 497.1 76.0 25.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/21ezyi7d",
   "source_url": "https://omlx.ai/benchmarks/performance/21ezyi7d",
   "tg_tokens": null,
   "tps": 76.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-21ezyi7d-ctx8192",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 522.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 522.7 55.4 25.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/21ezyi7d",
   "source_url": "https://omlx.ai/benchmarks/performance/21ezyi7d",
   "tg_tokens": null,
   "tps": 55.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dn2cb05l-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 825.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 8bit; 1k 825.5 18.5 28.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dn2cb05l",
   "source_url": "https://omlx.ai/benchmarks/performance/dn2cb05l",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dn2cb05l-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 946.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 8bit; 4k 946.0 18.4 29.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dn2cb05l",
   "source_url": "https://omlx.ai/benchmarks/performance/dn2cb05l",
   "tg_tokens": null,
   "tps": 18.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xok6cvge-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: 3j6fjwv9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 949.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 8bit; 4k 949.4 18.3 29.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xok6cvge",
   "source_url": "https://omlx.ai/benchmarks/performance/xok6cvge",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xok6cvge-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: 3j6fjwv9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 850.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 8bit; 16k 850.9 17.6 30.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xok6cvge",
   "source_url": "https://omlx.ai/benchmarks/performance/xok6cvge",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-hubzwapt-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: 36m1zzf6, 3sexbqow; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 389.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 389.0 20.1 22.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hubzwapt",
   "source_url": "https://omlx.ai/benchmarks/performance/hubzwapt",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-hubzwapt-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 36m1zzf6, 3sexbqow; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 351.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 351.3 23.8 23.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hubzwapt",
   "source_url": "https://omlx.ai/benchmarks/performance/hubzwapt",
   "tg_tokens": null,
   "tps": 23.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-hubzwapt-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: 36m1zzf6, 3sexbqow; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 367.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 367.7 22.7 19.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hubzwapt",
   "source_url": "https://omlx.ai/benchmarks/performance/hubzwapt",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-hubzwapt-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 36m1zzf6, 3sexbqow; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 306.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 306.5 14.5 20.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hubzwapt",
   "source_url": "https://omlx.ai/benchmarks/performance/hubzwapt",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-voesobf6-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace, single run at 64k context; peak memory 39.0 GB; TTFT 96526 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 678.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k tokens 678.9 PP tok/s 32.7 TG tok/s 96526 TTFT (ms) 39.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/voesobf6",
   "source_url": "https://omlx.ai/benchmarks/performance/voesobf6",
   "tg_tokens": null,
   "tps": 32.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-yqrijq0b-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "unsloth build; oMLX community benchmark trace, single run at 4k context; peak memory 27.4 GB; TTFT 18603 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.2,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-27B-NVFP4-unsloth on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 nvfp4; 4k tokens 220.2 PP tok/s 31.0 TG tok/s 18603 TTFT (ms) 27.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yqrijq0b",
   "source_url": "https://omlx.ai/benchmarks/performance/yqrijq0b",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-djfmmw44-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "unsloth build; oMLX community benchmark trace, single run at 4k context; peak memory 27.8 GB; TTFT 18027 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 227.2,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-27B-NVFP4-unsloth on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 nvfp4; 4k tokens 227.2 PP tok/s 25.4 TG tok/s 18027 TTFT (ms) 27.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/djfmmw44",
   "source_url": "https://omlx.ai/benchmarks/performance/djfmmw44",
   "tg_tokens": null,
   "tps": 25.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xznrsudm-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: td20xvnd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 816.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 816.7 55.8 24.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xznrsudm",
   "source_url": "https://omlx.ai/benchmarks/performance/xznrsudm",
   "tg_tokens": null,
   "tps": 55.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xznrsudm-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.7 GB at this context; duplicate traces of the same session: td20xvnd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 868.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 868.4 47.8 25.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xznrsudm",
   "source_url": "https://omlx.ai/benchmarks/performance/xznrsudm",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xznrsudm-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: td20xvnd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 951.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 951.9 53.4 26.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xznrsudm",
   "source_url": "https://omlx.ai/benchmarks/performance/xznrsudm",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-xznrsudm-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: td20xvnd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 961.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 961.9 84.9 26.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xznrsudm",
   "source_url": "https://omlx.ai/benchmarks/performance/xznrsudm",
   "tg_tokens": null,
   "tps": 84.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xznrsudm-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: td20xvnd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 891.8,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 891.8 105.7 27.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xznrsudm",
   "source_url": "https://omlx.ai/benchmarks/performance/xznrsudm",
   "tg_tokens": null,
   "tps": 105.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xznrsudm-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.0 GB at this context; duplicate traces of the same session: td20xvnd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 740.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 740.2 43.7 30.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xznrsudm",
   "source_url": "https://omlx.ai/benchmarks/performance/xznrsudm",
   "tg_tokens": null,
   "tps": 43.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-xznrsudm-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: td20xvnd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 554.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 554.4 41.8 34.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xznrsudm",
   "source_url": "https://omlx.ai/benchmarks/performance/xznrsudm",
   "tg_tokens": null,
   "tps": 41.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-j9p60nra-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2352.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,352 70.6 71.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j9p60nra",
   "source_url": "https://omlx.ai/benchmarks/performance/j9p60nra",
   "tg_tokens": null,
   "tps": 70.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-j9p60nra-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2309.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,309 65.7 78.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j9p60nra",
   "source_url": "https://omlx.ai/benchmarks/performance/j9p60nra",
   "tg_tokens": null,
   "tps": 65.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-kg6kfvix-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.4 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 144.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 144.2 13.6 33.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kg6kfvix",
   "source_url": "https://omlx.ai/benchmarks/performance/kg6kfvix",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-kg6kfvix-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 415.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 415.2 11.1 35.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kg6kfvix",
   "source_url": "https://omlx.ai/benchmarks/performance/kg6kfvix",
   "tg_tokens": null,
   "tps": 11.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-jbsn8l4d-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.2 GB at this context; duplicate traces of the same session: 4ssye3nk, qj4vilpx; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 811.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 1k 811.9 45.3 183.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jbsn8l4d",
   "source_url": "https://omlx.ai/benchmarks/performance/jbsn8l4d",
   "tg_tokens": null,
   "tps": 45.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-jbsn8l4d-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.6 GB at this context; duplicate traces of the same session: 4ssye3nk, qj4vilpx; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2420.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 4k 2,420 117.1 183.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jbsn8l4d",
   "source_url": "https://omlx.ai/benchmarks/performance/jbsn8l4d",
   "tg_tokens": null,
   "tps": 117.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-jbsn8l4d-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 184.7 GB at this context; duplicate traces of the same session: 4ssye3nk, qj4vilpx; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2716.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 8k 2,716 111.0 184.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jbsn8l4d",
   "source_url": "https://omlx.ai/benchmarks/performance/jbsn8l4d",
   "tg_tokens": null,
   "tps": 111.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-jbsn8l4d-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 186.9 GB at this context; duplicate traces of the same session: 4ssye3nk, qj4vilpx; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4444.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 64k 4,444 114.8 186.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jbsn8l4d",
   "source_url": "https://omlx.ai/benchmarks/performance/jbsn8l4d",
   "tg_tokens": null,
   "tps": 114.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-jbsn8l4d-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 191.6 GB at this context; duplicate traces of the same session: 4ssye3nk, qj4vilpx; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4314.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 128k 4,314 90.2 191.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jbsn8l4d",
   "source_url": "https://omlx.ai/benchmarks/performance/jbsn8l4d",
   "tg_tokens": null,
   "tps": 90.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-jbsn8l4d-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 193.3 GB at this context; duplicate traces of the same session: 4ssye3nk, qj4vilpx; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4189.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 195k 4,189 78.1 193.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jbsn8l4d",
   "source_url": "https://omlx.ai/benchmarks/performance/jbsn8l4d",
   "tg_tokens": null,
   "tps": 78.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-rb4triwk-ctx4096",
   "model": "Qwopus3.6-27B-Coder",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: mryg9io2; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 296.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwopus3.6-27B-Coder-oQ4-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 296.0 8.8 17.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rb4triwk",
   "source_url": "https://omlx.ai/benchmarks/performance/rb4triwk",
   "tg_tokens": null,
   "tps": 8.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-rb4triwk-ctx16384",
   "model": "Qwopus3.6-27B-Coder",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: mryg9io2; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 288.7,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwopus3.6-27B-Coder-oQ4-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 288.7 8.5 18.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rb4triwk",
   "source_url": "https://omlx.ai/benchmarks/performance/rb4triwk",
   "tg_tokens": null,
   "tps": 8.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-oqmh7fhy-ctx8192",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1923.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 1,923 47.4 22.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oqmh7fhy",
   "source_url": "https://omlx.ai/benchmarks/performance/oqmh7fhy",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-oqmh7fhy-ctx16384",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2790.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 2,790 81.8 22.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oqmh7fhy",
   "source_url": "https://omlx.ai/benchmarks/performance/oqmh7fhy",
   "tg_tokens": null,
   "tps": 81.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-oqmh7fhy-ctx32768",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2398.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 2,398 102.6 23.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oqmh7fhy",
   "source_url": "https://omlx.ai/benchmarks/performance/oqmh7fhy",
   "tg_tokens": null,
   "tps": 102.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-oqmh7fhy-ctx199680",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 32.2 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 913.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 195k 913.1 54.3 32.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oqmh7fhy",
   "source_url": "https://omlx.ai/benchmarks/performance/oqmh7fhy",
   "tg_tokens": null,
   "tps": 54.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-dy91zrgz-ctx32768",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 32k context; peak memory 23.6 GB; TTFT 14116 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2321.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k tokens 2,321 PP tok/s 127.4 TG tok/s 14116 TTFT (ms) 23.6 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dy91zrgz",
   "source_url": "https://omlx.ai/benchmarks/performance/dy91zrgz",
   "tg_tokens": null,
   "tps": 127.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-new89s20-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: taaj911u; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1688.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 1,688 112.4 16.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/new89s20",
   "source_url": "https://omlx.ai/benchmarks/performance/new89s20",
   "tg_tokens": null,
   "tps": 112.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-new89s20-ctx8192",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: taaj911u; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1835.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,835 114.5 16.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/new89s20",
   "source_url": "https://omlx.ai/benchmarks/performance/new89s20",
   "tg_tokens": null,
   "tps": 114.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-new89s20-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: taaj911u; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1724.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 1,724 110.2 17.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/new89s20",
   "source_url": "https://omlx.ai/benchmarks/performance/new89s20",
   "tg_tokens": null,
   "tps": 110.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-athv5woy-ctx1024",
   "model": "Gemma 4 26B-A4B-it-uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 25.8 GB; TTFT 545 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1880.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k tokens 1,880 PP tok/s 88.7 TG tok/s 545 TTFT (ms) 25.8 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/athv5woy",
   "source_url": "https://omlx.ai/benchmarks/performance/athv5woy",
   "tg_tokens": null,
   "tps": 88.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-46sqmiqt-ctx1024",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 207.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 207.0 28.5 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/46sqmiqt",
   "source_url": "https://omlx.ai/benchmarks/performance/46sqmiqt",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-46sqmiqt-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 204.4 46.7 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/46sqmiqt",
   "source_url": "https://omlx.ai/benchmarks/performance/46sqmiqt",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-46sqmiqt-ctx8192",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 193.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 193.4 30.7 21.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/46sqmiqt",
   "source_url": "https://omlx.ai/benchmarks/performance/46sqmiqt",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-46sqmiqt-ctx16384",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 193.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 193.4 39.1 23.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/46sqmiqt",
   "source_url": "https://omlx.ai/benchmarks/performance/46sqmiqt",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-46sqmiqt-ctx32768",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 179.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 179.7 26.2 25.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/46sqmiqt",
   "source_url": "https://omlx.ai/benchmarks/performance/46sqmiqt",
   "tg_tokens": null,
   "tps": 26.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-46sqmiqt-ctx65536",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 158.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 158.0 20.1 29.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/46sqmiqt",
   "source_url": "https://omlx.ai/benchmarks/performance/46sqmiqt",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-lyov1sur-ctx1024",
   "model": "gemma-4-31B-it-uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 18.4 GB; TTFT 3834 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 267.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-uncensored-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k tokens 267.1 PP tok/s 29.8 TG tok/s 3834 TTFT (ms) 18.4 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lyov1sur",
   "source_url": "https://omlx.ai/benchmarks/performance/lyov1sur",
   "tg_tokens": null,
   "tps": 29.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-v6qth67l-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Noctalin build; fp16 MTP; oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: 4w55o9gk, 9fqq9ewh; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 462.9,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "models--Noctalin--Qwen3.6-35B-A3B-oQ4-fp16-mtp on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 462.9 49.4 20.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v6qth67l",
   "source_url": "https://omlx.ai/benchmarks/performance/v6qth67l",
   "tg_tokens": null,
   "tps": 49.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-v6qth67l-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Noctalin build; fp16 MTP; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 4w55o9gk, 9fqq9ewh; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 463.5,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "models--Noctalin--Qwen3.6-35B-A3B-oQ4-fp16-mtp on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k 463.5 42.4 21.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v6qth67l",
   "source_url": "https://omlx.ai/benchmarks/performance/v6qth67l",
   "tg_tokens": null,
   "tps": 42.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-v6qth67l-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Noctalin build; fp16 MTP; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: 4w55o9gk, 9fqq9ewh; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 379.4,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "models--Noctalin--Qwen3.6-35B-A3B-oQ4-fp16-mtp on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 32k 379.4 37.6 22.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v6qth67l",
   "source_url": "https://omlx.ai/benchmarks/performance/v6qth67l",
   "tg_tokens": null,
   "tps": 37.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-v6qth67l-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Noctalin build; fp16 MTP; oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: 4w55o9gk, 9fqq9ewh; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 275.4,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "models--Noctalin--Qwen3.6-35B-A3B-oQ4-fp16-mtp on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 64k 275.4 29.5 23.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v6qth67l",
   "source_url": "https://omlx.ai/benchmarks/performance/v6qth67l",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-qggohs22-ctx16384",
   "model": "qwen3.5-4B-super-coder",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 4.4 GB; TTFT 41072 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 398.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "qwen3.5-4B-super-coder-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 16k tokens 398.9 PP tok/s 31.6 TG tok/s 41072 TTFT (ms) 4.4 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qggohs22",
   "source_url": "https://omlx.ai/benchmarks/performance/qggohs22",
   "tg_tokens": null,
   "tps": 31.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-x0idxl8l-ctx1024",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 247.2,
   "provenance": "community",
   "quant": "6bit",
   "quote": "AREX-2-6bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 247.2 20.2 22.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x0idxl8l",
   "source_url": "https://omlx.ai/benchmarks/performance/x0idxl8l",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-x0idxl8l-ctx4096",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 213.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "AREX-2-6bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 213.8 19.2 25.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x0idxl8l",
   "source_url": "https://omlx.ai/benchmarks/performance/x0idxl8l",
   "tg_tokens": null,
   "tps": 19.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-261k5swf-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.6 GB at this context; duplicate traces of the same session: 368tvpbg, hf9dlog8, qvarwslx, vyn1tdk5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 395.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 395.6 31.5 173.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/261k5swf",
   "source_url": "https://omlx.ai/benchmarks/performance/261k5swf",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-261k5swf-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.7 GB at this context; duplicate traces of the same session: 368tvpbg, hf9dlog8, qvarwslx, vyn1tdk5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 557.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 557.8 27.3 174.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/261k5swf",
   "source_url": "https://omlx.ai/benchmarks/performance/261k5swf",
   "tg_tokens": null,
   "tps": 27.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-261k5swf-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.0 GB at this context; duplicate traces of the same session: 368tvpbg, hf9dlog8, qvarwslx, vyn1tdk5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 506.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 506.7 27.3 175.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/261k5swf",
   "source_url": "https://omlx.ai/benchmarks/performance/261k5swf",
   "tg_tokens": null,
   "tps": 27.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-261k5swf-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.1 GB at this context; duplicate traces of the same session: 368tvpbg, hf9dlog8, qvarwslx, vyn1tdk5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 537.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 537.0 27.0 175.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/261k5swf",
   "source_url": "https://omlx.ai/benchmarks/performance/261k5swf",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-261k5swf-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.2 GB at this context; duplicate traces of the same session: 368tvpbg, hf9dlog8, qvarwslx, vyn1tdk5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 528.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 528.1 26.9 175.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/261k5swf",
   "source_url": "https://omlx.ai/benchmarks/performance/261k5swf",
   "tg_tokens": null,
   "tps": 26.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-261k5swf-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.6 GB at this context; duplicate traces of the same session: 368tvpbg, hf9dlog8, qvarwslx, vyn1tdk5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 518.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 518.7 25.2 175.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/261k5swf",
   "source_url": "https://omlx.ai/benchmarks/performance/261k5swf",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-261k5swf-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 176.3 GB at this context; duplicate traces of the same session: 368tvpbg, hf9dlog8, qvarwslx, vyn1tdk5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 509.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 509.0 21.5 176.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/261k5swf",
   "source_url": "https://omlx.ai/benchmarks/performance/261k5swf",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-yrd7axkn-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.6 GB at this context; duplicate traces of the same session: c5uo3kfi, ofamc6be, ttaipvsd; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 435.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 435.2 31.3 173.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yrd7axkn",
   "source_url": "https://omlx.ai/benchmarks/performance/yrd7axkn",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-yrd7axkn-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.7 GB at this context; duplicate traces of the same session: c5uo3kfi, ofamc6be, ttaipvsd; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 562.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 562.6 27.2 174.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yrd7axkn",
   "source_url": "https://omlx.ai/benchmarks/performance/yrd7axkn",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-yrd7axkn-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.0 GB at this context; duplicate traces of the same session: c5uo3kfi, ofamc6be, ttaipvsd; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 545.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 545.3 27.0 175.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yrd7axkn",
   "source_url": "https://omlx.ai/benchmarks/performance/yrd7axkn",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-yrd7axkn-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.1 GB at this context; duplicate traces of the same session: c5uo3kfi, ofamc6be, ttaipvsd; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 536.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 536.9 27.0 175.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yrd7axkn",
   "source_url": "https://omlx.ai/benchmarks/performance/yrd7axkn",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ls56ru6l-ctx16384",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 10.1 GB at this context; duplicate traces of the same session: hxfb4k0v; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1028.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-6bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 6bit; 16k 1,028 85.7 10.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ls56ru6l",
   "source_url": "https://omlx.ai/benchmarks/performance/ls56ru6l",
   "tg_tokens": null,
   "tps": 85.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ls56ru6l-ctx32768",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 10.3 GB at this context; duplicate traces of the same session: hxfb4k0v; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 911.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-6bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 6bit; 32k 911.7 77.5 10.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ls56ru6l",
   "source_url": "https://omlx.ai/benchmarks/performance/ls56ru6l",
   "tg_tokens": null,
   "tps": 77.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-xqw7yegy-ctx1024",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: 348ysugm; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1253.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-mlx-mxfp4 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 1,253 119.5 6.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xqw7yegy",
   "source_url": "https://omlx.ai/benchmarks/performance/xqw7yegy",
   "tg_tokens": null,
   "tps": 119.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-xqw7yegy-ctx4096",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: 348ysugm; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1182.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-mlx-mxfp4 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 1,182 114.9 6.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xqw7yegy",
   "source_url": "https://omlx.ai/benchmarks/performance/xqw7yegy",
   "tg_tokens": null,
   "tps": 114.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-fctxdjb1-ctx1024",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: 5c6ab4n6; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 818.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 818.9 100.6 18.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fctxdjb1",
   "source_url": "https://omlx.ai/benchmarks/performance/fctxdjb1",
   "tg_tokens": null,
   "tps": 100.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-fctxdjb1-ctx4096",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 5c6ab4n6; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 784.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 784.7 82.7 19.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fctxdjb1",
   "source_url": "https://omlx.ai/benchmarks/performance/fctxdjb1",
   "tg_tokens": null,
   "tps": 82.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-udvp547j-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Noctalin build; fp16 MTP; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1312.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Noctalin--Qwen3.6-35B-A3B-oQ4-fp16-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,312 99.8 20.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/udvp547j",
   "source_url": "https://omlx.ai/benchmarks/performance/udvp547j",
   "tg_tokens": null,
   "tps": 99.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-udvp547j-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Noctalin build; fp16 MTP; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1621.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Noctalin--Qwen3.6-35B-A3B-oQ4-fp16-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,621 86.6 22.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/udvp547j",
   "source_url": "https://omlx.ai/benchmarks/performance/udvp547j",
   "tg_tokens": null,
   "tps": 86.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-udvp547j-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Noctalin build; fp16 MTP; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1387.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Noctalin--Qwen3.6-35B-A3B-oQ4-fp16-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 1,387 79.4 23.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/udvp547j",
   "source_url": "https://omlx.ai/benchmarks/performance/udvp547j",
   "tg_tokens": null,
   "tps": 79.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-cafua26u-ctx4096",
   "model": "Qwen2.5-Coder-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1106.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-7B-Instruct-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,106 30.9 4.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cafua26u",
   "source_url": "https://omlx.ai/benchmarks/performance/cafua26u",
   "tg_tokens": null,
   "tps": 30.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-cafua26u-ctx16384",
   "model": "Qwen2.5-Coder-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 5.5 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 957.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-7B-Instruct-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 957.2 26.6 5.5 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cafua26u",
   "source_url": "https://omlx.ai/benchmarks/performance/cafua26u",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-vuqkdl48-ctx4096",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "MLX; oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: jmuciv9y; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2437.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,437 62.7 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vuqkdl48",
   "source_url": "https://omlx.ai/benchmarks/performance/vuqkdl48",
   "tg_tokens": null,
   "tps": 62.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-vuqkdl48-ctx16384",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "MLX; oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: jmuciv9y; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1757.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 1,757 45.5 18.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vuqkdl48",
   "source_url": "https://omlx.ai/benchmarks/performance/vuqkdl48",
   "tg_tokens": null,
   "tps": 45.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ft8dr8ej-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: p8bfia2u; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 905.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 905.8 107.1 3.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ft8dr8ej",
   "source_url": "https://omlx.ai/benchmarks/performance/ft8dr8ej",
   "tg_tokens": null,
   "tps": 107.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-ft8dr8ej-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.8 GB at this context; duplicate traces of the same session: p8bfia2u; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 842.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 842.5 89.3 4.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ft8dr8ej",
   "source_url": "https://omlx.ai/benchmarks/performance/ft8dr8ej",
   "tg_tokens": null,
   "tps": 89.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-qeosr92g-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 234.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 234.9 20.3 5.8 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qeosr92g",
   "source_url": "https://omlx.ai/benchmarks/performance/qeosr92g",
   "tg_tokens": null,
   "tps": 20.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-qeosr92g-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 237.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 237.5 20.1 6.2 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qeosr92g",
   "source_url": "https://omlx.ai/benchmarks/performance/qeosr92g",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ss2wxzfb-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: zgts131l; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1008.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,008 66.8 35.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ss2wxzfb",
   "source_url": "https://omlx.ai/benchmarks/performance/ss2wxzfb",
   "tg_tokens": null,
   "tps": 66.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ss2wxzfb-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: zgts131l; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 941.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 941.7 58.9 37.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ss2wxzfb",
   "source_url": "https://omlx.ai/benchmarks/performance/ss2wxzfb",
   "tg_tokens": null,
   "tps": 58.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ebfmytxb-ctx1024",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: llmaeu4o, xzacwcec, z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4163.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 4,163 192.7 19.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ebfmytxb",
   "source_url": "https://omlx.ai/benchmarks/performance/ebfmytxb",
   "tg_tokens": null,
   "tps": 192.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ebfmytxb-ctx4096",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: llmaeu4o, xzacwcec, z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5736.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 5,736 159.3 20.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ebfmytxb",
   "source_url": "https://omlx.ai/benchmarks/performance/ebfmytxb",
   "tg_tokens": null,
   "tps": 159.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ebfmytxb-ctx8192",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: llmaeu4o, xzacwcec, z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6112.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 8k 6,112 203.0 20.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ebfmytxb",
   "source_url": "https://omlx.ai/benchmarks/performance/ebfmytxb",
   "tg_tokens": null,
   "tps": 203.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ebfmytxb-ctx16384",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: llmaeu4o, xzacwcec, z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5607.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 5,607 195.0 20.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ebfmytxb",
   "source_url": "https://omlx.ai/benchmarks/performance/ebfmytxb",
   "tg_tokens": null,
   "tps": 195.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ebfmytxb-ctx32768",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: llmaeu4o, xzacwcec, z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4807.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 4,807 170.8 21.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ebfmytxb",
   "source_url": "https://omlx.ai/benchmarks/performance/ebfmytxb",
   "tg_tokens": null,
   "tps": 170.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ebfmytxb-ctx65536",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: llmaeu4o, xzacwcec, z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3580.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 64k 3,580 126.1 22.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ebfmytxb",
   "source_url": "https://omlx.ai/benchmarks/performance/ebfmytxb",
   "tg_tokens": null,
   "tps": 126.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ebfmytxb-ctx131072",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: llmaeu4o, xzacwcec, z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2310.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 128k 2,310 87.1 23.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ebfmytxb",
   "source_url": "https://omlx.ai/benchmarks/performance/ebfmytxb",
   "tg_tokens": null,
   "tps": 87.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ebfmytxb-ctx199680",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: llmaeu4o, xzacwcec, z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1579.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 195k 1,579 59.7 25.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ebfmytxb",
   "source_url": "https://omlx.ai/benchmarks/performance/ebfmytxb",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-007jy9ug-ctx1024",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 3cy8ezo9, esbg3qoh, ur3j8yno, yu64gvip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2615.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 2,615 151.1 19.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/007jy9ug",
   "source_url": "https://omlx.ai/benchmarks/performance/007jy9ug",
   "tg_tokens": null,
   "tps": 151.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-007jy9ug-ctx4096",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: 3cy8ezo9, esbg3qoh, ur3j8yno, yu64gvip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2684.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 2,684 173.6 20.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/007jy9ug",
   "source_url": "https://omlx.ai/benchmarks/performance/007jy9ug",
   "tg_tokens": null,
   "tps": 173.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-007jy9ug-ctx8192",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 3cy8ezo9, esbg3qoh, ur3j8yno, yu64gvip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4567.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 8k 4,567 190.8 20.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/007jy9ug",
   "source_url": "https://omlx.ai/benchmarks/performance/007jy9ug",
   "tg_tokens": null,
   "tps": 190.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-007jy9ug-ctx16384",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 3cy8ezo9, esbg3qoh, ur3j8yno, yu64gvip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4610.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 4,610 177.6 20.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/007jy9ug",
   "source_url": "https://omlx.ai/benchmarks/performance/007jy9ug",
   "tg_tokens": null,
   "tps": 177.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-007jy9ug-ctx32768",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 3cy8ezo9, esbg3qoh, ur3j8yno, yu64gvip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4019.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 4,019 164.2 21.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/007jy9ug",
   "source_url": "https://omlx.ai/benchmarks/performance/007jy9ug",
   "tg_tokens": null,
   "tps": 164.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-007jy9ug-ctx65536",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 3cy8ezo9, esbg3qoh, ur3j8yno, yu64gvip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2973.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 64k 2,973 121.2 21.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/007jy9ug",
   "source_url": "https://omlx.ai/benchmarks/performance/007jy9ug",
   "tg_tokens": null,
   "tps": 121.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-007jy9ug-ctx131072",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 3cy8ezo9, esbg3qoh, ur3j8yno, yu64gvip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2042.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 128k 2,042 91.8 23.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/007jy9ug",
   "source_url": "https://omlx.ai/benchmarks/performance/007jy9ug",
   "tg_tokens": null,
   "tps": 91.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-007jy9ug-ctx199680",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: 3cy8ezo9, esbg3qoh, ur3j8yno, yu64gvip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1432.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 195k 1,432 62.6 24.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/007jy9ug",
   "source_url": "https://omlx.ai/benchmarks/performance/007jy9ug",
   "tg_tokens": null,
   "tps": 62.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-r6ztv7vf-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: 101rpa4k; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1078.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 5bit; 1k 1,078 72.6 24.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r6ztv7vf",
   "source_url": "https://omlx.ai/benchmarks/performance/r6ztv7vf",
   "tg_tokens": null,
   "tps": 72.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-r6ztv7vf-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: 101rpa4k; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1244.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 5bit; 4k 1,244 59.8 25.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r6ztv7vf",
   "source_url": "https://omlx.ai/benchmarks/performance/r6ztv7vf",
   "tg_tokens": null,
   "tps": 59.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dfnf3lrx-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: c8x51isy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2967.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 2,967 202.6 26.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dfnf3lrx",
   "source_url": "https://omlx.ai/benchmarks/performance/dfnf3lrx",
   "tg_tokens": null,
   "tps": 202.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dfnf3lrx-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: c8x51isy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4950.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 4,950 158.3 26.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dfnf3lrx",
   "source_url": "https://omlx.ai/benchmarks/performance/dfnf3lrx",
   "tg_tokens": null,
   "tps": 158.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dfnf3lrx-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: c8x51isy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5350.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 5,350 180.0 26.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dfnf3lrx",
   "source_url": "https://omlx.ai/benchmarks/performance/dfnf3lrx",
   "tg_tokens": null,
   "tps": 180.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dfnf3lrx-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.0 GB at this context; duplicate traces of the same session: c8x51isy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5012.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 5,012 211.3 27.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dfnf3lrx",
   "source_url": "https://omlx.ai/benchmarks/performance/dfnf3lrx",
   "tg_tokens": null,
   "tps": 211.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dfnf3lrx-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: c8x51isy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4144.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 4,144 229.4 27.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dfnf3lrx",
   "source_url": "https://omlx.ai/benchmarks/performance/dfnf3lrx",
   "tg_tokens": null,
   "tps": 229.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dfnf3lrx-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: c8x51isy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3088.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 3,088 144.9 28.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dfnf3lrx",
   "source_url": "https://omlx.ai/benchmarks/performance/dfnf3lrx",
   "tg_tokens": null,
   "tps": 144.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dfnf3lrx-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: c8x51isy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2110.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 2,110 109.8 29.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dfnf3lrx",
   "source_url": "https://omlx.ai/benchmarks/performance/dfnf3lrx",
   "tg_tokens": null,
   "tps": 109.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pwenazo1-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 880.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 880.0 33.6 16.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pwenazo1",
   "source_url": "https://omlx.ai/benchmarks/performance/pwenazo1",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pwenazo1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 994.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 994.3 33.0 17.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pwenazo1",
   "source_url": "https://omlx.ai/benchmarks/performance/pwenazo1",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-uxjubqh3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1375.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,375 84.7 17.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uxjubqh3",
   "source_url": "https://omlx.ai/benchmarks/performance/uxjubqh3",
   "tg_tokens": null,
   "tps": 84.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-uxjubqh3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1624.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,624 137.9 18.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uxjubqh3",
   "source_url": "https://omlx.ai/benchmarks/performance/uxjubqh3",
   "tg_tokens": null,
   "tps": 137.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-uxjubqh3-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1619.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 1,619 61.7 19.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uxjubqh3",
   "source_url": "https://omlx.ai/benchmarks/performance/uxjubqh3",
   "tg_tokens": null,
   "tps": 61.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-uxjubqh3-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1566.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,566 73.8 19.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uxjubqh3",
   "source_url": "https://omlx.ai/benchmarks/performance/uxjubqh3",
   "tg_tokens": null,
   "tps": 73.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-uxjubqh3-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1453.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,453 116.5 20.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uxjubqh3",
   "source_url": "https://omlx.ai/benchmarks/performance/uxjubqh3",
   "tg_tokens": null,
   "tps": 116.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-uxjubqh3-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1255.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,255 77.1 23.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uxjubqh3",
   "source_url": "https://omlx.ai/benchmarks/performance/uxjubqh3",
   "tg_tokens": null,
   "tps": 77.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-uxjubqh3-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 978.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 978.9 77.6 27.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uxjubqh3",
   "source_url": "https://omlx.ai/benchmarks/performance/uxjubqh3",
   "tg_tokens": null,
   "tps": 77.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-uxjubqh3-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 33.0 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 795.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 795.4 69.2 33.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uxjubqh3",
   "source_url": "https://omlx.ai/benchmarks/performance/uxjubqh3",
   "tg_tokens": null,
   "tps": 69.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-286j8net-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: ur5pbo0r; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 6bit; 1k 204.6 15.9 22.3 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/286j8net",
   "source_url": "https://omlx.ai/benchmarks/performance/286j8net",
   "tg_tokens": null,
   "tps": 15.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-286j8net-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: ur5pbo0r; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 212.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 6bit; 4k 212.8 15.6 23.3 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/286j8net",
   "source_url": "https://omlx.ai/benchmarks/performance/286j8net",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3j6fjwv9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 949.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 8bit; 4k 949.4 18.3 29.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3j6fjwv9",
   "source_url": "https://omlx.ai/benchmarks/performance/3j6fjwv9",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3j6fjwv9-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 850.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 8bit; 16k 850.9 17.6 30.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3j6fjwv9",
   "source_url": "https://omlx.ai/benchmarks/performance/3j6fjwv9",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-36m1zzf6-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: 3sexbqow; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 389.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 389.0 20.1 22.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/36m1zzf6",
   "source_url": "https://omlx.ai/benchmarks/performance/36m1zzf6",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-36m1zzf6-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 3sexbqow; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 351.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 351.3 23.8 23.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/36m1zzf6",
   "source_url": "https://omlx.ai/benchmarks/performance/36m1zzf6",
   "tg_tokens": null,
   "tps": 23.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-36m1zzf6-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: 3sexbqow; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 367.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 367.7 22.7 19.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/36m1zzf6",
   "source_url": "https://omlx.ai/benchmarks/performance/36m1zzf6",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-36m1zzf6-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 3sexbqow; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 306.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 306.5 14.5 20.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/36m1zzf6",
   "source_url": "https://omlx.ai/benchmarks/performance/36m1zzf6",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ok7lmu16-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: h79k9vue; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 744.7,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-27B-NVFP4 on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 nvfp4; 4k 744.7 39.1 25.0 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ok7lmu16",
   "source_url": "https://omlx.ai/benchmarks/performance/ok7lmu16",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ok7lmu16-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: h79k9vue; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 730.6,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-27B-NVFP4 on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 nvfp4; 16k 730.6 39.3 25.8 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ok7lmu16",
   "source_url": "https://omlx.ai/benchmarks/performance/ok7lmu16",
   "tg_tokens": null,
   "tps": 39.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-7ml6p0t5-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 16.1 GB; TTFT 15883 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 64.5,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 64.5 PP tok/s 11.5 TG tok/s 15883 TTFT (ms) 16.1 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ml6p0t5",
   "source_url": "https://omlx.ai/benchmarks/performance/7ml6p0t5",
   "tg_tokens": null,
   "tps": 11.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-88j44k4z-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "SAGE-ARCS MTP build; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: y496b92g; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 122.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-SAGE-ARCS-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 5bit; 1k 122.4 16.2 26.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/88j44k4z",
   "source_url": "https://omlx.ai/benchmarks/performance/88j44k4z",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-88j44k4z-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "SAGE-ARCS MTP build; oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: y496b92g; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 121.3,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-SAGE-ARCS-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 5bit; 4k 121.3 10.9 30.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/88j44k4z",
   "source_url": "https://omlx.ai/benchmarks/performance/88j44k4z",
   "tg_tokens": null,
   "tps": 10.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-td20xvnd-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 816.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 816.7 55.8 24.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/td20xvnd",
   "source_url": "https://omlx.ai/benchmarks/performance/td20xvnd",
   "tg_tokens": null,
   "tps": 55.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-td20xvnd-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 868.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 868.4 47.8 25.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/td20xvnd",
   "source_url": "https://omlx.ai/benchmarks/performance/td20xvnd",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-td20xvnd-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 951.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 951.9 53.4 26.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/td20xvnd",
   "source_url": "https://omlx.ai/benchmarks/performance/td20xvnd",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-td20xvnd-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 961.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 961.9 84.9 26.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/td20xvnd",
   "source_url": "https://omlx.ai/benchmarks/performance/td20xvnd",
   "tg_tokens": null,
   "tps": 84.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-td20xvnd-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 891.8,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 891.8 105.7 27.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/td20xvnd",
   "source_url": "https://omlx.ai/benchmarks/performance/td20xvnd",
   "tg_tokens": null,
   "tps": 105.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-td20xvnd-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 740.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 740.2 43.7 30.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/td20xvnd",
   "source_url": "https://omlx.ai/benchmarks/performance/td20xvnd",
   "tg_tokens": null,
   "tps": 43.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-td20xvnd-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 554.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 554.4 41.8 34.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/td20xvnd",
   "source_url": "https://omlx.ai/benchmarks/performance/td20xvnd",
   "tg_tokens": null,
   "tps": 41.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-9s3z3i4v-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.8 GB; TTFT 5601 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 182.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k tokens 182.8 PP tok/s 17.1 TG tok/s 5601 TTFT (ms) 70.8 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9s3z3i4v",
   "source_url": "https://omlx.ai/benchmarks/performance/9s3z3i4v",
   "tg_tokens": null,
   "tps": 17.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-0s1adhml-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.3 GB at this context; duplicate traces of the same session: vuo2kyuk; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 535.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 535.4 15.5 72.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0s1adhml",
   "source_url": "https://omlx.ai/benchmarks/performance/0s1adhml",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-0s1adhml-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.5 GB at this context; duplicate traces of the same session: vuo2kyuk; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 936.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k 936.2 15.1 74.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0s1adhml",
   "source_url": "https://omlx.ai/benchmarks/performance/0s1adhml",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-qj4vilpx-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.2 GB at this context; duplicate traces of the same session: 4ssye3nk; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 811.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 1k 811.9 45.3 183.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qj4vilpx",
   "source_url": "https://omlx.ai/benchmarks/performance/qj4vilpx",
   "tg_tokens": null,
   "tps": 45.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-qj4vilpx-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.6 GB at this context; duplicate traces of the same session: 4ssye3nk; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2420.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 4k 2,420 117.1 183.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qj4vilpx",
   "source_url": "https://omlx.ai/benchmarks/performance/qj4vilpx",
   "tg_tokens": null,
   "tps": 117.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-qj4vilpx-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 184.7 GB at this context; duplicate traces of the same session: 4ssye3nk; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2716.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 8k 2,716 111.0 184.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qj4vilpx",
   "source_url": "https://omlx.ai/benchmarks/performance/qj4vilpx",
   "tg_tokens": null,
   "tps": 111.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-qj4vilpx-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 186.9 GB at this context; duplicate traces of the same session: 4ssye3nk; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4444.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 64k 4,444 114.8 186.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qj4vilpx",
   "source_url": "https://omlx.ai/benchmarks/performance/qj4vilpx",
   "tg_tokens": null,
   "tps": 114.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-qj4vilpx-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 191.6 GB at this context; duplicate traces of the same session: 4ssye3nk; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4314.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 128k 4,314 90.2 191.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qj4vilpx",
   "source_url": "https://omlx.ai/benchmarks/performance/qj4vilpx",
   "tg_tokens": null,
   "tps": 90.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-qj4vilpx-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 193.3 GB at this context; duplicate traces of the same session: 4ssye3nk; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4189.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 195k 4,189 78.1 193.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qj4vilpx",
   "source_url": "https://omlx.ai/benchmarks/performance/qj4vilpx",
   "tg_tokens": null,
   "tps": 78.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-mryg9io2-ctx4096",
   "model": "Qwopus3.6-27B-Coder",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 296.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwopus3.6-27B-Coder-oQ4-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 296.0 8.8 17.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mryg9io2",
   "source_url": "https://omlx.ai/benchmarks/performance/mryg9io2",
   "tg_tokens": null,
   "tps": 8.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-mryg9io2-ctx16384",
   "model": "Qwopus3.6-27B-Coder",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 288.7,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwopus3.6-27B-Coder-oQ4-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 288.7 8.5 18.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mryg9io2",
   "source_url": "https://omlx.ai/benchmarks/performance/mryg9io2",
   "tg_tokens": null,
   "tps": 8.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-taaj911u-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: fvf70j5y; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1688.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 1,688 112.4 16.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/taaj911u",
   "source_url": "https://omlx.ai/benchmarks/performance/taaj911u",
   "tg_tokens": null,
   "tps": 112.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-taaj911u-ctx8192",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: fvf70j5y; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1835.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,835 114.5 16.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/taaj911u",
   "source_url": "https://omlx.ai/benchmarks/performance/taaj911u",
   "tg_tokens": null,
   "tps": 114.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-taaj911u-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: fvf70j5y; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1724.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 1,724 110.2 17.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/taaj911u",
   "source_url": "https://omlx.ai/benchmarks/performance/taaj911u",
   "tg_tokens": null,
   "tps": 110.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-9fqq9ewh-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Noctalin build; fp16 MTP; oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: 4w55o9gk; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 462.9,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "models--Noctalin--Qwen3.6-35B-A3B-oQ4-fp16-mtp on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 462.9 49.4 20.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9fqq9ewh",
   "source_url": "https://omlx.ai/benchmarks/performance/9fqq9ewh",
   "tg_tokens": null,
   "tps": 49.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-9fqq9ewh-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Noctalin build; fp16 MTP; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 4w55o9gk; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 463.5,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "models--Noctalin--Qwen3.6-35B-A3B-oQ4-fp16-mtp on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k 463.5 42.4 21.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9fqq9ewh",
   "source_url": "https://omlx.ai/benchmarks/performance/9fqq9ewh",
   "tg_tokens": null,
   "tps": 42.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-9fqq9ewh-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Noctalin build; fp16 MTP; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: 4w55o9gk; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 379.4,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "models--Noctalin--Qwen3.6-35B-A3B-oQ4-fp16-mtp on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 32k 379.4 37.6 22.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9fqq9ewh",
   "source_url": "https://omlx.ai/benchmarks/performance/9fqq9ewh",
   "tg_tokens": null,
   "tps": 37.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-9fqq9ewh-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Noctalin build; fp16 MTP; oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: 4w55o9gk; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 275.4,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "models--Noctalin--Qwen3.6-35B-A3B-oQ4-fp16-mtp on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 64k 275.4 29.5 23.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9fqq9ewh",
   "source_url": "https://omlx.ai/benchmarks/performance/9fqq9ewh",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-vyn1tdk5-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.6 GB at this context; duplicate traces of the same session: 368tvpbg, hf9dlog8, qvarwslx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 395.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 395.6 31.5 173.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vyn1tdk5",
   "source_url": "https://omlx.ai/benchmarks/performance/vyn1tdk5",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-vyn1tdk5-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.7 GB at this context; duplicate traces of the same session: 368tvpbg, hf9dlog8, qvarwslx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 557.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 557.8 27.3 174.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vyn1tdk5",
   "source_url": "https://omlx.ai/benchmarks/performance/vyn1tdk5",
   "tg_tokens": null,
   "tps": 27.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-vyn1tdk5-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.0 GB at this context; duplicate traces of the same session: 368tvpbg, hf9dlog8, qvarwslx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 506.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 506.7 27.3 175.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vyn1tdk5",
   "source_url": "https://omlx.ai/benchmarks/performance/vyn1tdk5",
   "tg_tokens": null,
   "tps": 27.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-vyn1tdk5-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.1 GB at this context; duplicate traces of the same session: 368tvpbg, hf9dlog8, qvarwslx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 537.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 537.0 27.0 175.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vyn1tdk5",
   "source_url": "https://omlx.ai/benchmarks/performance/vyn1tdk5",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-vyn1tdk5-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.2 GB at this context; duplicate traces of the same session: 368tvpbg, hf9dlog8, qvarwslx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 528.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 528.1 26.9 175.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vyn1tdk5",
   "source_url": "https://omlx.ai/benchmarks/performance/vyn1tdk5",
   "tg_tokens": null,
   "tps": 26.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-vyn1tdk5-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.6 GB at this context; duplicate traces of the same session: 368tvpbg, hf9dlog8, qvarwslx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 518.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 518.7 25.2 175.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vyn1tdk5",
   "source_url": "https://omlx.ai/benchmarks/performance/vyn1tdk5",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-vyn1tdk5-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 176.3 GB at this context; duplicate traces of the same session: 368tvpbg, hf9dlog8, qvarwslx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 509.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 509.0 21.5 176.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vyn1tdk5",
   "source_url": "https://omlx.ai/benchmarks/performance/vyn1tdk5",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ttaipvsd-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.6 GB at this context; duplicate traces of the same session: c5uo3kfi, ofamc6be; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 435.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 435.2 31.3 173.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ttaipvsd",
   "source_url": "https://omlx.ai/benchmarks/performance/ttaipvsd",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ttaipvsd-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.7 GB at this context; duplicate traces of the same session: c5uo3kfi, ofamc6be; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 562.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 562.6 27.2 174.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ttaipvsd",
   "source_url": "https://omlx.ai/benchmarks/performance/ttaipvsd",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ttaipvsd-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.0 GB at this context; duplicate traces of the same session: c5uo3kfi, ofamc6be; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 545.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 545.3 27.0 175.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ttaipvsd",
   "source_url": "https://omlx.ai/benchmarks/performance/ttaipvsd",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ttaipvsd-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.1 GB at this context; duplicate traces of the same session: c5uo3kfi, ofamc6be; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 536.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 536.9 27.0 175.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ttaipvsd",
   "source_url": "https://omlx.ai/benchmarks/performance/ttaipvsd",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hs1yaxkr-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.7 GB at this context; duplicate traces of the same session: 2x0c20ae, 71soa0o2, bqg5sfd1, fzwjewvb, o0wgl2q0, xqtnr6be; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 440.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 440.1 40.8 173.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hs1yaxkr",
   "source_url": "https://omlx.ai/benchmarks/performance/hs1yaxkr",
   "tg_tokens": null,
   "tps": 40.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hs1yaxkr-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.8 GB at this context; duplicate traces of the same session: 2x0c20ae, 71soa0o2, bqg5sfd1, fzwjewvb, o0wgl2q0, xqtnr6be; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 566.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 566.5 39.0 174.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hs1yaxkr",
   "source_url": "https://omlx.ai/benchmarks/performance/hs1yaxkr",
   "tg_tokens": null,
   "tps": 39.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hs1yaxkr-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.1 GB at this context; duplicate traces of the same session: 2x0c20ae, 71soa0o2, bqg5sfd1, fzwjewvb, o0wgl2q0, xqtnr6be; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 518.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 518.4 38.8 175.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hs1yaxkr",
   "source_url": "https://omlx.ai/benchmarks/performance/hs1yaxkr",
   "tg_tokens": null,
   "tps": 38.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hs1yaxkr-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.2 GB at this context; duplicate traces of the same session: 2x0c20ae, 71soa0o2, bqg5sfd1, fzwjewvb, o0wgl2q0, xqtnr6be; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 550.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 550.4 38.6 175.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hs1yaxkr",
   "source_url": "https://omlx.ai/benchmarks/performance/hs1yaxkr",
   "tg_tokens": null,
   "tps": 38.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hs1yaxkr-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.3 GB at this context; duplicate traces of the same session: 2x0c20ae, 71soa0o2, bqg5sfd1, fzwjewvb, o0wgl2q0, xqtnr6be; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 542.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 542.7 38.3 175.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hs1yaxkr",
   "source_url": "https://omlx.ai/benchmarks/performance/hs1yaxkr",
   "tg_tokens": null,
   "tps": 38.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hs1yaxkr-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.6 GB at this context; duplicate traces of the same session: 2x0c20ae, 71soa0o2, bqg5sfd1, fzwjewvb, o0wgl2q0, xqtnr6be; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 534.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 534.4 37.4 175.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hs1yaxkr",
   "source_url": "https://omlx.ai/benchmarks/performance/hs1yaxkr",
   "tg_tokens": null,
   "tps": 37.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hs1yaxkr-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 176.4 GB at this context; duplicate traces of the same session: 2x0c20ae, 71soa0o2, bqg5sfd1, fzwjewvb, o0wgl2q0, xqtnr6be; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 510.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 510.7 35.9 176.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hs1yaxkr",
   "source_url": "https://omlx.ai/benchmarks/performance/hs1yaxkr",
   "tg_tokens": null,
   "tps": 35.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-hxfb4k0v-ctx16384",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 10.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1028.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-6bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 6bit; 16k 1,028 85.7 10.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hxfb4k0v",
   "source_url": "https://omlx.ai/benchmarks/performance/hxfb4k0v",
   "tg_tokens": null,
   "tps": 85.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-hxfb4k0v-ctx32768",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 10.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 911.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-6bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 6bit; 32k 911.7 77.5 10.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hxfb4k0v",
   "source_url": "https://omlx.ai/benchmarks/performance/hxfb4k0v",
   "tg_tokens": null,
   "tps": 77.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-348ysugm-ctx1024",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1253.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-mlx-mxfp4 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 1,253 119.5 6.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/348ysugm",
   "source_url": "https://omlx.ai/benchmarks/performance/348ysugm",
   "tg_tokens": null,
   "tps": 119.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-348ysugm-ctx4096",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1182.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-mlx-mxfp4 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 1,182 114.9 6.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/348ysugm",
   "source_url": "https://omlx.ai/benchmarks/performance/348ysugm",
   "tg_tokens": null,
   "tps": 114.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-5c6ab4n6-ctx1024",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 818.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 818.9 100.6 18.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5c6ab4n6",
   "source_url": "https://omlx.ai/benchmarks/performance/5c6ab4n6",
   "tg_tokens": null,
   "tps": 100.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-5c6ab4n6-ctx4096",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 784.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 784.7 82.7 19.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5c6ab4n6",
   "source_url": "https://omlx.ai/benchmarks/performance/5c6ab4n6",
   "tg_tokens": null,
   "tps": 82.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-b3489dbx-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "OsaurusAI build; oMLX community benchmark trace, single run at 1k context; peak memory 19.7 GB; TTFT 592 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1731.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "OsaurusAI-Qwen3.6-35B-A3B-mxfp4 on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,731 PP tok/s 183.6 TG tok/s 592 TTFT (ms) 19.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b3489dbx",
   "source_url": "https://omlx.ai/benchmarks/performance/b3489dbx",
   "tg_tokens": null,
   "tps": 183.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-jmuciv9y-ctx4096",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "MLX; oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2437.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,437 62.7 17.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jmuciv9y",
   "source_url": "https://omlx.ai/benchmarks/performance/jmuciv9y",
   "tg_tokens": null,
   "tps": 62.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-jmuciv9y-ctx16384",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "MLX; oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1757.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 1,757 45.5 18.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jmuciv9y",
   "source_url": "https://omlx.ai/benchmarks/performance/jmuciv9y",
   "tg_tokens": null,
   "tps": 45.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-p8bfia2u-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 905.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 905.8 107.1 3.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p8bfia2u",
   "source_url": "https://omlx.ai/benchmarks/performance/p8bfia2u",
   "tg_tokens": null,
   "tps": 107.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-p8bfia2u-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.8 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 842.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 842.5 89.3 4.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p8bfia2u",
   "source_url": "https://omlx.ai/benchmarks/performance/p8bfia2u",
   "tg_tokens": null,
   "tps": 89.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-0dc5yioh-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: 2p6ffc9v; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 899.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 899.5 106.7 3.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0dc5yioh",
   "source_url": "https://omlx.ai/benchmarks/performance/0dc5yioh",
   "tg_tokens": null,
   "tps": 106.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-0dc5yioh-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.8 GB at this context; duplicate traces of the same session: 2p6ffc9v; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 849.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 849.0 86.9 4.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0dc5yioh",
   "source_url": "https://omlx.ai/benchmarks/performance/0dc5yioh",
   "tg_tokens": null,
   "tps": 86.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-zexy05o3-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 234.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 234.9 20.3 5.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zexy05o3",
   "source_url": "https://omlx.ai/benchmarks/performance/zexy05o3",
   "tg_tokens": null,
   "tps": 20.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-zexy05o3-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 237.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 237.5 20.1 6.2 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zexy05o3",
   "source_url": "https://omlx.ai/benchmarks/performance/zexy05o3",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-zgts131l-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1008.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,008 66.8 35.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zgts131l",
   "source_url": "https://omlx.ai/benchmarks/performance/zgts131l",
   "tg_tokens": null,
   "tps": 66.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-zgts131l-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 941.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 941.7 58.9 37.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zgts131l",
   "source_url": "https://omlx.ai/benchmarks/performance/zgts131l",
   "tg_tokens": null,
   "tps": 58.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-5cu0lekz-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.6 GB at this context; duplicate traces of the same session: eej7y6hx; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1013.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,013 71.3 35.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5cu0lekz",
   "source_url": "https://omlx.ai/benchmarks/performance/5cu0lekz",
   "tg_tokens": null,
   "tps": 71.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-5cu0lekz-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: eej7y6hx; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1018.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 1,018 63.7 37.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5cu0lekz",
   "source_url": "https://omlx.ai/benchmarks/performance/5cu0lekz",
   "tg_tokens": null,
   "tps": 63.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-llmaeu4o-ctx1024",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: xzacwcec, z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4163.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 4,163 192.7 19.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/llmaeu4o",
   "source_url": "https://omlx.ai/benchmarks/performance/llmaeu4o",
   "tg_tokens": null,
   "tps": 192.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-llmaeu4o-ctx4096",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: xzacwcec, z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5736.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 5,736 159.3 20.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/llmaeu4o",
   "source_url": "https://omlx.ai/benchmarks/performance/llmaeu4o",
   "tg_tokens": null,
   "tps": 159.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-llmaeu4o-ctx8192",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: xzacwcec, z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6112.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 8k 6,112 203.0 20.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/llmaeu4o",
   "source_url": "https://omlx.ai/benchmarks/performance/llmaeu4o",
   "tg_tokens": null,
   "tps": 203.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-llmaeu4o-ctx16384",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: xzacwcec, z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5607.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 5,607 195.0 20.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/llmaeu4o",
   "source_url": "https://omlx.ai/benchmarks/performance/llmaeu4o",
   "tg_tokens": null,
   "tps": 195.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-llmaeu4o-ctx32768",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: xzacwcec, z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4807.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 4,807 170.8 21.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/llmaeu4o",
   "source_url": "https://omlx.ai/benchmarks/performance/llmaeu4o",
   "tg_tokens": null,
   "tps": 170.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-llmaeu4o-ctx65536",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: xzacwcec, z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3580.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 64k 3,580 126.1 22.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/llmaeu4o",
   "source_url": "https://omlx.ai/benchmarks/performance/llmaeu4o",
   "tg_tokens": null,
   "tps": 126.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-llmaeu4o-ctx131072",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: xzacwcec, z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2310.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 128k 2,310 87.1 23.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/llmaeu4o",
   "source_url": "https://omlx.ai/benchmarks/performance/llmaeu4o",
   "tg_tokens": null,
   "tps": 87.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-llmaeu4o-ctx199680",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: xzacwcec, z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1579.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 195k 1,579 59.7 25.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/llmaeu4o",
   "source_url": "https://omlx.ai/benchmarks/performance/llmaeu4o",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3cy8ezo9-ctx1024",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: esbg3qoh, ur3j8yno, yu64gvip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2615.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 2,615 151.1 19.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3cy8ezo9",
   "source_url": "https://omlx.ai/benchmarks/performance/3cy8ezo9",
   "tg_tokens": null,
   "tps": 151.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3cy8ezo9-ctx4096",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: esbg3qoh, ur3j8yno, yu64gvip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2684.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 2,684 173.6 20.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3cy8ezo9",
   "source_url": "https://omlx.ai/benchmarks/performance/3cy8ezo9",
   "tg_tokens": null,
   "tps": 173.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3cy8ezo9-ctx8192",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: esbg3qoh, ur3j8yno, yu64gvip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4567.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 8k 4,567 190.8 20.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3cy8ezo9",
   "source_url": "https://omlx.ai/benchmarks/performance/3cy8ezo9",
   "tg_tokens": null,
   "tps": 190.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3cy8ezo9-ctx16384",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: esbg3qoh, ur3j8yno, yu64gvip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4610.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 4,610 177.6 20.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3cy8ezo9",
   "source_url": "https://omlx.ai/benchmarks/performance/3cy8ezo9",
   "tg_tokens": null,
   "tps": 177.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3cy8ezo9-ctx32768",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: esbg3qoh, ur3j8yno, yu64gvip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4019.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 4,019 164.2 21.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3cy8ezo9",
   "source_url": "https://omlx.ai/benchmarks/performance/3cy8ezo9",
   "tg_tokens": null,
   "tps": 164.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3cy8ezo9-ctx65536",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: esbg3qoh, ur3j8yno, yu64gvip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2973.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 64k 2,973 121.2 21.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3cy8ezo9",
   "source_url": "https://omlx.ai/benchmarks/performance/3cy8ezo9",
   "tg_tokens": null,
   "tps": 121.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3cy8ezo9-ctx131072",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: esbg3qoh, ur3j8yno, yu64gvip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2042.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 128k 2,042 91.8 23.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3cy8ezo9",
   "source_url": "https://omlx.ai/benchmarks/performance/3cy8ezo9",
   "tg_tokens": null,
   "tps": 91.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3cy8ezo9-ctx199680",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: esbg3qoh, ur3j8yno, yu64gvip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1432.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 195k 1,432 62.6 24.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3cy8ezo9",
   "source_url": "https://omlx.ai/benchmarks/performance/3cy8ezo9",
   "tg_tokens": null,
   "tps": 62.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-4hh9zlv4-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 18.3 GB; TTFT 2291 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 447.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.6-35B-A3B-mxfp4 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 447.0 PP tok/s 48.6 TG tok/s 2291 TTFT (ms) 18.3 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4hh9zlv4",
   "source_url": "https://omlx.ai/benchmarks/performance/4hh9zlv4",
   "tg_tokens": null,
   "tps": 48.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-wgux8rrs-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 18.3 GB; TTFT 2292 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 446.8,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.6-35B-A3B-mxfp4 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 446.8 PP tok/s 48.7 TG tok/s 2292 TTFT (ms) 18.3 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wgux8rrs",
   "source_url": "https://omlx.ai/benchmarks/performance/wgux8rrs",
   "tg_tokens": null,
   "tps": 48.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-101rpa4k-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1078.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 5bit; 1k 1,078 72.6 24.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/101rpa4k",
   "source_url": "https://omlx.ai/benchmarks/performance/101rpa4k",
   "tg_tokens": null,
   "tps": 72.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-101rpa4k-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1244.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 5bit; 4k 1,244 59.8 25.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/101rpa4k",
   "source_url": "https://omlx.ai/benchmarks/performance/101rpa4k",
   "tg_tokens": null,
   "tps": 59.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-c8x51isy-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2967.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 2,967 202.6 26.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c8x51isy",
   "source_url": "https://omlx.ai/benchmarks/performance/c8x51isy",
   "tg_tokens": null,
   "tps": 202.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-c8x51isy-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4950.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 4,950 158.3 26.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c8x51isy",
   "source_url": "https://omlx.ai/benchmarks/performance/c8x51isy",
   "tg_tokens": null,
   "tps": 158.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-c8x51isy-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5350.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 5,350 180.0 26.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c8x51isy",
   "source_url": "https://omlx.ai/benchmarks/performance/c8x51isy",
   "tg_tokens": null,
   "tps": 180.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-c8x51isy-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5012.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 5,012 211.3 27.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c8x51isy",
   "source_url": "https://omlx.ai/benchmarks/performance/c8x51isy",
   "tg_tokens": null,
   "tps": 211.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-c8x51isy-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4144.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 4,144 229.4 27.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c8x51isy",
   "source_url": "https://omlx.ai/benchmarks/performance/c8x51isy",
   "tg_tokens": null,
   "tps": 229.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-c8x51isy-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3088.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 3,088 144.9 28.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c8x51isy",
   "source_url": "https://omlx.ai/benchmarks/performance/c8x51isy",
   "tg_tokens": null,
   "tps": 144.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-c8x51isy-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2110.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 2,110 109.8 29.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c8x51isy",
   "source_url": "https://omlx.ai/benchmarks/performance/c8x51isy",
   "tg_tokens": null,
   "tps": 109.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-4rnoyw4d-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: kkmq8qd4; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 232.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 232.4 18.4 18.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4rnoyw4d",
   "source_url": "https://omlx.ai/benchmarks/performance/4rnoyw4d",
   "tg_tokens": null,
   "tps": 18.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "outlier"
   ],
   "hardware": "M5",
   "id": "omlx-4rnoyw4d-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: kkmq8qd4; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 199.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 16k 199.7 24.6 19.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4rnoyw4d",
   "source_url": "https://omlx.ai/benchmarks/performance/4rnoyw4d",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-8wpply12-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 2s7lrp0f, jwywy3s2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 127.5 8.5 28.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8wpply12",
   "source_url": "https://omlx.ai/benchmarks/performance/8wpply12",
   "tg_tokens": null,
   "tps": 8.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-8wpply12-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.9 GB at this context; duplicate traces of the same session: 2s7lrp0f, jwywy3s2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 126.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 126.7 8.5 31.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8wpply12",
   "source_url": "https://omlx.ai/benchmarks/performance/8wpply12",
   "tg_tokens": null,
   "tps": 8.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-8wpply12-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.5 GB at this context; duplicate traces of the same session: 2s7lrp0f, jwywy3s2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 112.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 112.7 8.2 32.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8wpply12",
   "source_url": "https://omlx.ai/benchmarks/performance/8wpply12",
   "tg_tokens": null,
   "tps": 8.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ur5pbo0r-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 6bit; 1k 204.6 15.9 22.3 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ur5pbo0r",
   "source_url": "https://omlx.ai/benchmarks/performance/ur5pbo0r",
   "tg_tokens": null,
   "tps": 15.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ur5pbo0r-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 212.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 6bit; 4k 212.8 15.6 23.3 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ur5pbo0r",
   "source_url": "https://omlx.ai/benchmarks/performance/ur5pbo0r",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3sexbqow-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 389.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 389.0 20.1 22.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3sexbqow",
   "source_url": "https://omlx.ai/benchmarks/performance/3sexbqow",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3sexbqow-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 351.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 351.3 23.8 23.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3sexbqow",
   "source_url": "https://omlx.ai/benchmarks/performance/3sexbqow",
   "tg_tokens": null,
   "tps": 23.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3sexbqow-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 367.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 367.7 22.7 19.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3sexbqow",
   "source_url": "https://omlx.ai/benchmarks/performance/3sexbqow",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3sexbqow-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 306.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 306.5 14.5 20.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3sexbqow",
   "source_url": "https://omlx.ai/benchmarks/performance/3sexbqow",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-h79k9vue-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 744.7,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-27B-NVFP4 on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 nvfp4; 4k 744.7 39.1 25.0 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h79k9vue",
   "source_url": "https://omlx.ai/benchmarks/performance/h79k9vue",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-h79k9vue-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 730.6,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-27B-NVFP4 on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 nvfp4; 16k 730.6 39.3 25.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h79k9vue",
   "source_url": "https://omlx.ai/benchmarks/performance/h79k9vue",
   "tg_tokens": null,
   "tps": 39.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-8pyc66bg-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 16.1 GB; TTFT 15887 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 64.5,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 64.5 PP tok/s 11.8 TG tok/s 15887 TTFT (ms) 16.1 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8pyc66bg",
   "source_url": "https://omlx.ai/benchmarks/performance/8pyc66bg",
   "tg_tokens": null,
   "tps": 11.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-m1ysht8k-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 15.0 GB; TTFT 16152 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 63.4,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 63.4 PP tok/s 7.2 TG tok/s 16152 TTFT (ms) 15.0 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m1ysht8k",
   "source_url": "https://omlx.ai/benchmarks/performance/m1ysht8k",
   "tg_tokens": null,
   "tps": 7.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-svotomtt-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 24.8 GB; TTFT 15359 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 266.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 266.7 PP tok/s 54.0 TG tok/s 15359 TTFT (ms) 24.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/svotomtt",
   "source_url": "https://omlx.ai/benchmarks/performance/svotomtt",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ctm38cv4-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 22.1 GB; TTFT 16514 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 248.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 248.0 PP tok/s 43.2 TG tok/s 16514 TTFT (ms) 22.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ctm38cv4",
   "source_url": "https://omlx.ai/benchmarks/performance/ctm38cv4",
   "tg_tokens": null,
   "tps": 43.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-if6ofj93-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 24.7 GB; TTFT 16035 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 255.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 255.4 PP tok/s 54.6 TG tok/s 16035 TTFT (ms) 24.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/if6ofj93",
   "source_url": "https://omlx.ai/benchmarks/performance/if6ofj93",
   "tg_tokens": null,
   "tps": 54.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-y496b92g-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "SAGE-ARCS MTP build; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 122.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-SAGE-ARCS-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 5bit; 1k 122.4 16.2 26.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y496b92g",
   "source_url": "https://omlx.ai/benchmarks/performance/y496b92g",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-y496b92g-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "SAGE-ARCS MTP build; oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 121.3,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-SAGE-ARCS-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 5bit; 4k 121.3 10.9 30.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y496b92g",
   "source_url": "https://omlx.ai/benchmarks/performance/y496b92g",
   "tg_tokens": null,
   "tps": 10.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-vuo2kyuk-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.3 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 535.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 535.4 15.5 72.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vuo2kyuk",
   "source_url": "https://omlx.ai/benchmarks/performance/vuo2kyuk",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-vuo2kyuk-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.5 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 936.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k 936.2 15.1 74.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vuo2kyuk",
   "source_url": "https://omlx.ai/benchmarks/performance/vuo2kyuk",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-f3unl8w2-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: 7nno64su, 81s1szm8, z6m2f172; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1110.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,110 91.4 100.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f3unl8w2",
   "source_url": "https://omlx.ai/benchmarks/performance/f3unl8w2",
   "tg_tokens": null,
   "tps": 91.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-f3unl8w2-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: 7nno64su, 81s1szm8, z6m2f172; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1331.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,331 91.5 101.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f3unl8w2",
   "source_url": "https://omlx.ai/benchmarks/performance/f3unl8w2",
   "tg_tokens": null,
   "tps": 91.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-f3unl8w2-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: 7nno64su, 81s1szm8, z6m2f172; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1373.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 1,373 93.4 101.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f3unl8w2",
   "source_url": "https://omlx.ai/benchmarks/performance/f3unl8w2",
   "tg_tokens": null,
   "tps": 93.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-f3unl8w2-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.6 GB at this context; duplicate traces of the same session: 7nno64su, 81s1szm8, z6m2f172; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1364.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 1,364 68.7 101.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f3unl8w2",
   "source_url": "https://omlx.ai/benchmarks/performance/f3unl8w2",
   "tg_tokens": null,
   "tps": 68.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-4ssye3nk-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.2 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 811.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 1k 811.9 45.3 183.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ssye3nk",
   "source_url": "https://omlx.ai/benchmarks/performance/4ssye3nk",
   "tg_tokens": null,
   "tps": 45.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-4ssye3nk-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.6 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2420.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 4k 2,420 117.1 183.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ssye3nk",
   "source_url": "https://omlx.ai/benchmarks/performance/4ssye3nk",
   "tg_tokens": null,
   "tps": 117.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-4ssye3nk-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 184.7 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2716.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 8k 2,716 111.0 184.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ssye3nk",
   "source_url": "https://omlx.ai/benchmarks/performance/4ssye3nk",
   "tg_tokens": null,
   "tps": 111.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-4ssye3nk-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 186.9 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4444.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 64k 4,444 114.8 186.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ssye3nk",
   "source_url": "https://omlx.ai/benchmarks/performance/4ssye3nk",
   "tg_tokens": null,
   "tps": 114.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-4ssye3nk-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 191.6 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4314.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 128k 4,314 90.2 191.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ssye3nk",
   "source_url": "https://omlx.ai/benchmarks/performance/4ssye3nk",
   "tg_tokens": null,
   "tps": 90.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-4ssye3nk-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 193.3 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4189.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 195k 4,189 78.1 193.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ssye3nk",
   "source_url": "https://omlx.ai/benchmarks/performance/4ssye3nk",
   "tg_tokens": null,
   "tps": 78.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-nz0d1iqf-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1375.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,375 111.7 14.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nz0d1iqf",
   "source_url": "https://omlx.ai/benchmarks/performance/nz0d1iqf",
   "tg_tokens": null,
   "tps": 111.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-nz0d1iqf-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 15.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1232.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,232 101.1 15.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nz0d1iqf",
   "source_url": "https://omlx.ai/benchmarks/performance/nz0d1iqf",
   "tg_tokens": null,
   "tps": 101.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-fvf70j5y-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1688.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 1,688 112.4 16.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fvf70j5y",
   "source_url": "https://omlx.ai/benchmarks/performance/fvf70j5y",
   "tg_tokens": null,
   "tps": 112.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-fvf70j5y-ctx8192",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1835.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,835 114.5 16.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fvf70j5y",
   "source_url": "https://omlx.ai/benchmarks/performance/fvf70j5y",
   "tg_tokens": null,
   "tps": 114.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-fvf70j5y-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1724.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 1,724 110.2 17.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fvf70j5y",
   "source_url": "https://omlx.ai/benchmarks/performance/fvf70j5y",
   "tg_tokens": null,
   "tps": 110.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-4w55o9gk-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Noctalin build; fp16 MTP; oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 462.9,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "models--Noctalin--Qwen3.6-35B-A3B-oQ4-fp16-mtp on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 462.9 49.4 20.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4w55o9gk",
   "source_url": "https://omlx.ai/benchmarks/performance/4w55o9gk",
   "tg_tokens": null,
   "tps": 49.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-4w55o9gk-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Noctalin build; fp16 MTP; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 463.5,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "models--Noctalin--Qwen3.6-35B-A3B-oQ4-fp16-mtp on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k 463.5 42.4 21.5 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4w55o9gk",
   "source_url": "https://omlx.ai/benchmarks/performance/4w55o9gk",
   "tg_tokens": null,
   "tps": 42.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-4w55o9gk-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Noctalin build; fp16 MTP; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 379.4,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "models--Noctalin--Qwen3.6-35B-A3B-oQ4-fp16-mtp on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 32k 379.4 37.6 22.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4w55o9gk",
   "source_url": "https://omlx.ai/benchmarks/performance/4w55o9gk",
   "tg_tokens": null,
   "tps": 37.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-4w55o9gk-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Noctalin build; fp16 MTP; oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 275.4,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "models--Noctalin--Qwen3.6-35B-A3B-oQ4-fp16-mtp on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 64k 275.4 29.5 23.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4w55o9gk",
   "source_url": "https://omlx.ai/benchmarks/performance/4w55o9gk",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ryq225h3-ctx16384",
   "model": "Ornith-1.5-397B",
   "notes": "AX build; AXQ MTP; oMLX community benchmark trace, single run at 16k context; peak memory 205.0 GB; TTFT 29670 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 552.2,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "AX-Ornith-1.5-397B-MLX-AXQ-MXFP4-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k tokens 552.2 PP tok/s 35.8 TG tok/s 29670 TTFT (ms) 205.0 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ryq225h3",
   "source_url": "https://omlx.ai/benchmarks/performance/ryq225h3",
   "tg_tokens": null,
   "tps": 35.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qvarwslx-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.6 GB at this context; duplicate traces of the same session: 368tvpbg, hf9dlog8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 395.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 395.6 31.5 173.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qvarwslx",
   "source_url": "https://omlx.ai/benchmarks/performance/qvarwslx",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qvarwslx-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.7 GB at this context; duplicate traces of the same session: 368tvpbg, hf9dlog8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 557.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 557.8 27.3 174.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qvarwslx",
   "source_url": "https://omlx.ai/benchmarks/performance/qvarwslx",
   "tg_tokens": null,
   "tps": 27.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qvarwslx-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.0 GB at this context; duplicate traces of the same session: 368tvpbg, hf9dlog8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 506.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 506.7 27.3 175.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qvarwslx",
   "source_url": "https://omlx.ai/benchmarks/performance/qvarwslx",
   "tg_tokens": null,
   "tps": 27.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qvarwslx-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.1 GB at this context; duplicate traces of the same session: 368tvpbg, hf9dlog8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 537.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 537.0 27.0 175.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qvarwslx",
   "source_url": "https://omlx.ai/benchmarks/performance/qvarwslx",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qvarwslx-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.2 GB at this context; duplicate traces of the same session: 368tvpbg, hf9dlog8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 528.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 528.1 26.9 175.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qvarwslx",
   "source_url": "https://omlx.ai/benchmarks/performance/qvarwslx",
   "tg_tokens": null,
   "tps": 26.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qvarwslx-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.6 GB at this context; duplicate traces of the same session: 368tvpbg, hf9dlog8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 518.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 518.7 25.2 175.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qvarwslx",
   "source_url": "https://omlx.ai/benchmarks/performance/qvarwslx",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qvarwslx-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 176.3 GB at this context; duplicate traces of the same session: 368tvpbg, hf9dlog8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 509.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 509.0 21.5 176.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qvarwslx",
   "source_url": "https://omlx.ai/benchmarks/performance/qvarwslx",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ofamc6be-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.6 GB at this context; duplicate traces of the same session: c5uo3kfi; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 435.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 435.2 31.3 173.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ofamc6be",
   "source_url": "https://omlx.ai/benchmarks/performance/ofamc6be",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ofamc6be-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.7 GB at this context; duplicate traces of the same session: c5uo3kfi; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 562.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 562.6 27.2 174.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ofamc6be",
   "source_url": "https://omlx.ai/benchmarks/performance/ofamc6be",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ofamc6be-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.0 GB at this context; duplicate traces of the same session: c5uo3kfi; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 545.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 545.3 27.0 175.0 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ofamc6be",
   "source_url": "https://omlx.ai/benchmarks/performance/ofamc6be",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ofamc6be-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.1 GB at this context; duplicate traces of the same session: c5uo3kfi; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 536.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 536.9 27.0 175.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ofamc6be",
   "source_url": "https://omlx.ai/benchmarks/performance/ofamc6be",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xqtnr6be-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.7 GB at this context; duplicate traces of the same session: 2x0c20ae, 71soa0o2, bqg5sfd1, fzwjewvb, o0wgl2q0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 440.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 440.1 40.8 173.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xqtnr6be",
   "source_url": "https://omlx.ai/benchmarks/performance/xqtnr6be",
   "tg_tokens": null,
   "tps": 40.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xqtnr6be-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.8 GB at this context; duplicate traces of the same session: 2x0c20ae, 71soa0o2, bqg5sfd1, fzwjewvb, o0wgl2q0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 566.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 566.5 39.0 174.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xqtnr6be",
   "source_url": "https://omlx.ai/benchmarks/performance/xqtnr6be",
   "tg_tokens": null,
   "tps": 39.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xqtnr6be-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.1 GB at this context; duplicate traces of the same session: 2x0c20ae, 71soa0o2, bqg5sfd1, fzwjewvb, o0wgl2q0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 518.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 518.4 38.8 175.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xqtnr6be",
   "source_url": "https://omlx.ai/benchmarks/performance/xqtnr6be",
   "tg_tokens": null,
   "tps": 38.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xqtnr6be-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.2 GB at this context; duplicate traces of the same session: 2x0c20ae, 71soa0o2, bqg5sfd1, fzwjewvb, o0wgl2q0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 550.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 550.4 38.6 175.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xqtnr6be",
   "source_url": "https://omlx.ai/benchmarks/performance/xqtnr6be",
   "tg_tokens": null,
   "tps": 38.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xqtnr6be-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.3 GB at this context; duplicate traces of the same session: 2x0c20ae, 71soa0o2, bqg5sfd1, fzwjewvb, o0wgl2q0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 542.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 542.7 38.3 175.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xqtnr6be",
   "source_url": "https://omlx.ai/benchmarks/performance/xqtnr6be",
   "tg_tokens": null,
   "tps": 38.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xqtnr6be-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.6 GB at this context; duplicate traces of the same session: 2x0c20ae, 71soa0o2, bqg5sfd1, fzwjewvb, o0wgl2q0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 534.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 534.4 37.4 175.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xqtnr6be",
   "source_url": "https://omlx.ai/benchmarks/performance/xqtnr6be",
   "tg_tokens": null,
   "tps": 37.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xqtnr6be-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 176.4 GB at this context; duplicate traces of the same session: 2x0c20ae, 71soa0o2, bqg5sfd1, fzwjewvb, o0wgl2q0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 510.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 510.7 35.9 176.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xqtnr6be",
   "source_url": "https://omlx.ai/benchmarks/performance/xqtnr6be",
   "tg_tokens": null,
   "tps": 35.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-2p6ffc9v-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 899.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 899.5 106.7 3.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2p6ffc9v",
   "source_url": "https://omlx.ai/benchmarks/performance/2p6ffc9v",
   "tg_tokens": null,
   "tps": 106.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-2p6ffc9v-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.8 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 849.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 849.0 86.9 4.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2p6ffc9v",
   "source_url": "https://omlx.ai/benchmarks/performance/2p6ffc9v",
   "tg_tokens": null,
   "tps": 86.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-itup949y-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: c0q5rnzx; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 909.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 909.9 107.3 3.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/itup949y",
   "source_url": "https://omlx.ai/benchmarks/performance/itup949y",
   "tg_tokens": null,
   "tps": 107.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-itup949y-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.8 GB at this context; duplicate traces of the same session: c0q5rnzx; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 842.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 842.7 99.9 4.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/itup949y",
   "source_url": "https://omlx.ai/benchmarks/performance/itup949y",
   "tg_tokens": null,
   "tps": 99.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-eej7y6hx-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.6 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1013.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,013 71.3 35.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eej7y6hx",
   "source_url": "https://omlx.ai/benchmarks/performance/eej7y6hx",
   "tg_tokens": null,
   "tps": 71.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-eej7y6hx-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1018.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 1,018 63.7 37.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eej7y6hx",
   "source_url": "https://omlx.ai/benchmarks/performance/eej7y6hx",
   "tg_tokens": null,
   "tps": 63.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-pldp09om-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.6 GB at this context; duplicate traces of the same session: hdqs336a; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1011.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,011 70.4 35.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pldp09om",
   "source_url": "https://omlx.ai/benchmarks/performance/pldp09om",
   "tg_tokens": null,
   "tps": 70.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-pldp09om-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: hdqs336a; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1010.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 1,010 63.1 37.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pldp09om",
   "source_url": "https://omlx.ai/benchmarks/performance/pldp09om",
   "tg_tokens": null,
   "tps": 63.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-v2i7tmmx-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.6 GB at this context; duplicate traces of the same session: lm719gom; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1012.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,012 71.0 35.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v2i7tmmx",
   "source_url": "https://omlx.ai/benchmarks/performance/v2i7tmmx",
   "tg_tokens": null,
   "tps": 71.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-v2i7tmmx-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: lm719gom; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1010.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 1,010 63.3 37.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v2i7tmmx",
   "source_url": "https://omlx.ai/benchmarks/performance/v2i7tmmx",
   "tg_tokens": null,
   "tps": 63.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-rdl3hgbn-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.6 GB at this context; duplicate traces of the same session: vlb2tzum; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1004.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,004 69.8 35.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rdl3hgbn",
   "source_url": "https://omlx.ai/benchmarks/performance/rdl3hgbn",
   "tg_tokens": null,
   "tps": 69.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-rdl3hgbn-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: vlb2tzum; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 983.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 983.9 62.0 37.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rdl3hgbn",
   "source_url": "https://omlx.ai/benchmarks/performance/rdl3hgbn",
   "tg_tokens": null,
   "tps": 62.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-xzacwcec-ctx1024",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4163.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 4,163 192.7 19.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xzacwcec",
   "source_url": "https://omlx.ai/benchmarks/performance/xzacwcec",
   "tg_tokens": null,
   "tps": 192.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-xzacwcec-ctx4096",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5736.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 5,736 159.3 20.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xzacwcec",
   "source_url": "https://omlx.ai/benchmarks/performance/xzacwcec",
   "tg_tokens": null,
   "tps": 159.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-xzacwcec-ctx8192",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6112.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 8k 6,112 203.0 20.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xzacwcec",
   "source_url": "https://omlx.ai/benchmarks/performance/xzacwcec",
   "tg_tokens": null,
   "tps": 203.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-xzacwcec-ctx16384",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5607.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 5,607 195.0 20.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xzacwcec",
   "source_url": "https://omlx.ai/benchmarks/performance/xzacwcec",
   "tg_tokens": null,
   "tps": 195.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-xzacwcec-ctx32768",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4807.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 4,807 170.8 21.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xzacwcec",
   "source_url": "https://omlx.ai/benchmarks/performance/xzacwcec",
   "tg_tokens": null,
   "tps": 170.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-xzacwcec-ctx65536",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3580.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 64k 3,580 126.1 22.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xzacwcec",
   "source_url": "https://omlx.ai/benchmarks/performance/xzacwcec",
   "tg_tokens": null,
   "tps": 126.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-xzacwcec-ctx131072",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2310.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 128k 2,310 87.1 23.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xzacwcec",
   "source_url": "https://omlx.ai/benchmarks/performance/xzacwcec",
   "tg_tokens": null,
   "tps": 87.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-xzacwcec-ctx199680",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: z2v7hqio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1579.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 195k 1,579 59.7 25.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xzacwcec",
   "source_url": "https://omlx.ai/benchmarks/performance/xzacwcec",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-esbg3qoh-ctx1024",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: ur3j8yno, yu64gvip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2615.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 2,615 151.1 19.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/esbg3qoh",
   "source_url": "https://omlx.ai/benchmarks/performance/esbg3qoh",
   "tg_tokens": null,
   "tps": 151.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-esbg3qoh-ctx4096",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: ur3j8yno, yu64gvip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2684.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 2,684 173.6 20.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/esbg3qoh",
   "source_url": "https://omlx.ai/benchmarks/performance/esbg3qoh",
   "tg_tokens": null,
   "tps": 173.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-esbg3qoh-ctx8192",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: ur3j8yno, yu64gvip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4567.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 8k 4,567 190.8 20.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/esbg3qoh",
   "source_url": "https://omlx.ai/benchmarks/performance/esbg3qoh",
   "tg_tokens": null,
   "tps": 190.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-esbg3qoh-ctx16384",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: ur3j8yno, yu64gvip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4610.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 4,610 177.6 20.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/esbg3qoh",
   "source_url": "https://omlx.ai/benchmarks/performance/esbg3qoh",
   "tg_tokens": null,
   "tps": 177.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-esbg3qoh-ctx32768",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: ur3j8yno, yu64gvip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4019.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 4,019 164.2 21.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/esbg3qoh",
   "source_url": "https://omlx.ai/benchmarks/performance/esbg3qoh",
   "tg_tokens": null,
   "tps": 164.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-esbg3qoh-ctx65536",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: ur3j8yno, yu64gvip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2973.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 64k 2,973 121.2 21.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/esbg3qoh",
   "source_url": "https://omlx.ai/benchmarks/performance/esbg3qoh",
   "tg_tokens": null,
   "tps": 121.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-esbg3qoh-ctx131072",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: ur3j8yno, yu64gvip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2042.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 128k 2,042 91.8 23.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/esbg3qoh",
   "source_url": "https://omlx.ai/benchmarks/performance/esbg3qoh",
   "tg_tokens": null,
   "tps": 91.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-esbg3qoh-ctx199680",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: ur3j8yno, yu64gvip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1432.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 195k 1,432 62.6 24.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/esbg3qoh",
   "source_url": "https://omlx.ai/benchmarks/performance/esbg3qoh",
   "tg_tokens": null,
   "tps": 62.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-pydh0gyi-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 26.9 GB at this context; duplicate traces of the same session: 7ua722c9; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1223.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 5bit; 1k 1,223 107.6 26.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pydh0gyi",
   "source_url": "https://omlx.ai/benchmarks/performance/pydh0gyi",
   "tg_tokens": null,
   "tps": 107.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-pydh0gyi-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 29.0 GB at this context; duplicate traces of the same session: 7ua722c9; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1349.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 5bit; 4k 1,349 73.4 29.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pydh0gyi",
   "source_url": "https://omlx.ai/benchmarks/performance/pydh0gyi",
   "tg_tokens": null,
   "tps": 73.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-kkmq8qd4-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 232.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 232.4 18.4 18.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kkmq8qd4",
   "source_url": "https://omlx.ai/benchmarks/performance/kkmq8qd4",
   "tg_tokens": null,
   "tps": 18.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "outlier"
   ],
   "hardware": "M5",
   "id": "omlx-kkmq8qd4-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 199.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 16k 199.7 24.6 19.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kkmq8qd4",
   "source_url": "https://omlx.ai/benchmarks/performance/kkmq8qd4",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-jwywy3s2-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 2s7lrp0f; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 127.5 8.5 28.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jwywy3s2",
   "source_url": "https://omlx.ai/benchmarks/performance/jwywy3s2",
   "tg_tokens": null,
   "tps": 8.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-jwywy3s2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.9 GB at this context; duplicate traces of the same session: 2s7lrp0f; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 126.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 126.7 8.5 31.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jwywy3s2",
   "source_url": "https://omlx.ai/benchmarks/performance/jwywy3s2",
   "tg_tokens": null,
   "tps": 8.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-jwywy3s2-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.5 GB at this context; duplicate traces of the same session: 2s7lrp0f; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 112.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 112.7 8.2 32.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jwywy3s2",
   "source_url": "https://omlx.ai/benchmarks/performance/jwywy3s2",
   "tg_tokens": null,
   "tps": 8.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-c3j0tz10-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 25.0 GB; TTFT 47663 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 687.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 6bit; 32k tokens 687.5 PP tok/s 15.5 TG tok/s 47663 TTFT (ms) 25.0 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c3j0tz10",
   "source_url": "https://omlx.ai/benchmarks/performance/c3j0tz10",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-406z7cw8-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 24.0 GB; TTFT 22311 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 734.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 6bit; 16k tokens 734.3 PP tok/s 16.6 TG tok/s 22311 TTFT (ms) 24.0 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/406z7cw8",
   "source_url": "https://omlx.ai/benchmarks/performance/406z7cw8",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-fg72lazs-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: z1k5reyj; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 46.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 46.9 3.3 33.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fg72lazs",
   "source_url": "https://omlx.ai/benchmarks/performance/fg72lazs",
   "tg_tokens": null,
   "tps": 3.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-fg72lazs-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 40.5 GB at this context; duplicate traces of the same session: z1k5reyj; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 59.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 59.7 3.5 40.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fg72lazs",
   "source_url": "https://omlx.ai/benchmarks/performance/fg72lazs",
   "tg_tokens": null,
   "tps": 3.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-z6m2f172-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: 7nno64su, 81s1szm8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1110.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,110 91.4 100.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z6m2f172",
   "source_url": "https://omlx.ai/benchmarks/performance/z6m2f172",
   "tg_tokens": null,
   "tps": 91.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-z6m2f172-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: 7nno64su, 81s1szm8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1331.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,331 91.5 101.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z6m2f172",
   "source_url": "https://omlx.ai/benchmarks/performance/z6m2f172",
   "tg_tokens": null,
   "tps": 91.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-z6m2f172-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: 7nno64su, 81s1szm8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1373.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 1,373 93.4 101.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z6m2f172",
   "source_url": "https://omlx.ai/benchmarks/performance/z6m2f172",
   "tg_tokens": null,
   "tps": 93.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-z6m2f172-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.6 GB at this context; duplicate traces of the same session: 7nno64su, 81s1szm8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1364.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 1,364 68.7 101.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z6m2f172",
   "source_url": "https://omlx.ai/benchmarks/performance/z6m2f172",
   "tg_tokens": null,
   "tps": 68.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-judfqkg7-ctx65536",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace, single run at 64k context; peak memory 21.9 GB; TTFT 231361 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 283.3,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 2bit; 64k tokens 283.3 PP tok/s 19.5 TG tok/s 231361 TTFT (ms) 21.9 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/judfqkg7",
   "source_url": "https://omlx.ai/benchmarks/performance/judfqkg7",
   "tg_tokens": null,
   "tps": 19.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-xonqequ1-ctx1024",
   "model": "Ternary-Bonsai-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 10.2 GB at this context; duplicate traces of the same session: zrk3x3m0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 310.9,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-27B-mlx-2bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 2bit; 1k 310.9 11.5 10.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xonqequ1",
   "source_url": "https://omlx.ai/benchmarks/performance/xonqequ1",
   "tg_tokens": null,
   "tps": 11.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-xonqequ1-ctx4096",
   "model": "Ternary-Bonsai-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: zrk3x3m0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 317.9,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-27B-mlx-2bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 2bit; 4k 317.9 13.4 11.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xonqequ1",
   "source_url": "https://omlx.ai/benchmarks/performance/xonqequ1",
   "tg_tokens": null,
   "tps": 13.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-xet1fdcf-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 7.0 GB; TTFT 6275 ms; duplicate traces of the same session: none; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 163.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-fp16-mtp on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 1k tokens 163.2 PP tok/s 9.9 TG tok/s 6275 TTFT (ms) 7.0 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xet1fdcf",
   "source_url": "https://omlx.ai/benchmarks/performance/xet1fdcf",
   "tg_tokens": null,
   "tps": 9.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-kpjxwdbx-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 11.2 GB; TTFT 6110 ms; duplicate traces of the same session: none; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 167.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-fp16-mtp on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 1k tokens 167.6 PP tok/s 9.3 TG tok/s 6110 TTFT (ms) 11.2 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kpjxwdbx",
   "source_url": "https://omlx.ai/benchmarks/performance/kpjxwdbx",
   "tg_tokens": null,
   "tps": 9.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vgr4rvru-ctx1024",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "it variant; oMLX community benchmark trace; peak memory 15.4 GB at this context; duplicate traces of the same session: 5ajwahti, 7tbob2la, tghs39f7, u3zummpr, vla6c6wz; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2796.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "diffusiongemma-26B-A4B-it-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 2,796 27.7 15.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vgr4rvru",
   "source_url": "https://omlx.ai/benchmarks/performance/vgr4rvru",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vgr4rvru-ctx4096",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "it variant; oMLX community benchmark trace; peak memory 16.2 GB at this context; duplicate traces of the same session: 5ajwahti, 7tbob2la, tghs39f7, u3zummpr, vla6c6wz; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2551.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "diffusiongemma-26B-A4B-it-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 2,551 20.8 16.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vgr4rvru",
   "source_url": "https://omlx.ai/benchmarks/performance/vgr4rvru",
   "tg_tokens": null,
   "tps": 20.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vgr4rvru-ctx8192",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "it variant; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 5ajwahti, 7tbob2la, tghs39f7, u3zummpr, vla6c6wz; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2437.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "diffusiongemma-26B-A4B-it-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 2,437 23.5 16.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vgr4rvru",
   "source_url": "https://omlx.ai/benchmarks/performance/vgr4rvru",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vgr4rvru-ctx16384",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "it variant; oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 5ajwahti, 7tbob2la, tghs39f7, u3zummpr, vla6c6wz; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2323.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "diffusiongemma-26B-A4B-it-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 2,323 47.6 17.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vgr4rvru",
   "source_url": "https://omlx.ai/benchmarks/performance/vgr4rvru",
   "tg_tokens": null,
   "tps": 47.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vgr4rvru-ctx32768",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "it variant; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 5ajwahti, 7tbob2la, tghs39f7, u3zummpr, vla6c6wz; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2053.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "diffusiongemma-26B-A4B-it-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 2,053 18.8 18.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vgr4rvru",
   "source_url": "https://omlx.ai/benchmarks/performance/vgr4rvru",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vgr4rvru-ctx65536",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "it variant; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 5ajwahti, 7tbob2la, tghs39f7, u3zummpr, vla6c6wz; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1621.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "diffusiongemma-26B-A4B-it-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 1,621 22.2 21.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vgr4rvru",
   "source_url": "https://omlx.ai/benchmarks/performance/vgr4rvru",
   "tg_tokens": null,
   "tps": 22.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vgr4rvru-ctx131072",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "it variant; oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: 5ajwahti, 7tbob2la, tghs39f7, u3zummpr, vla6c6wz; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1097.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "diffusiongemma-26B-A4B-it-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 1,097 10.0 26.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vgr4rvru",
   "source_url": "https://omlx.ai/benchmarks/performance/vgr4rvru",
   "tg_tokens": null,
   "tps": 10.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-9pl17zfj-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: mvmvlcb1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1292.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,292 47.8 14.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9pl17zfj",
   "source_url": "https://omlx.ai/benchmarks/performance/9pl17zfj",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-9pl17zfj-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 15.7 GB at this context; duplicate traces of the same session: mvmvlcb1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1193.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,193 133.4 15.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9pl17zfj",
   "source_url": "https://omlx.ai/benchmarks/performance/9pl17zfj",
   "tg_tokens": null,
   "tps": 133.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-l99bzm0i-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1375.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,375 111.7 14.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l99bzm0i",
   "source_url": "https://omlx.ai/benchmarks/performance/l99bzm0i",
   "tg_tokens": null,
   "tps": 111.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-l99bzm0i-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 15.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1232.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,232 101.1 15.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l99bzm0i",
   "source_url": "https://omlx.ai/benchmarks/performance/l99bzm0i",
   "tg_tokens": null,
   "tps": 101.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-y3177o4s-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 6.2 GB; TTFT 2135 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1919.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E4B-it-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 1,919 PP tok/s 61.9 TG tok/s 2135 TTFT (ms) 6.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y3177o4s",
   "source_url": "https://omlx.ai/benchmarks/performance/y3177o4s",
   "tg_tokens": null,
   "tps": 61.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-368tvpbg-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.6 GB at this context; duplicate traces of the same session: hf9dlog8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 395.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 395.6 31.5 173.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/368tvpbg",
   "source_url": "https://omlx.ai/benchmarks/performance/368tvpbg",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-368tvpbg-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.7 GB at this context; duplicate traces of the same session: hf9dlog8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 557.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 557.8 27.3 174.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/368tvpbg",
   "source_url": "https://omlx.ai/benchmarks/performance/368tvpbg",
   "tg_tokens": null,
   "tps": 27.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-368tvpbg-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.0 GB at this context; duplicate traces of the same session: hf9dlog8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 506.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 506.7 27.3 175.0 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/368tvpbg",
   "source_url": "https://omlx.ai/benchmarks/performance/368tvpbg",
   "tg_tokens": null,
   "tps": 27.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-368tvpbg-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.1 GB at this context; duplicate traces of the same session: hf9dlog8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 537.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 537.0 27.0 175.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/368tvpbg",
   "source_url": "https://omlx.ai/benchmarks/performance/368tvpbg",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-368tvpbg-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.2 GB at this context; duplicate traces of the same session: hf9dlog8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 528.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 528.1 26.9 175.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/368tvpbg",
   "source_url": "https://omlx.ai/benchmarks/performance/368tvpbg",
   "tg_tokens": null,
   "tps": 26.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-368tvpbg-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.6 GB at this context; duplicate traces of the same session: hf9dlog8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 518.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 518.7 25.2 175.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/368tvpbg",
   "source_url": "https://omlx.ai/benchmarks/performance/368tvpbg",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-368tvpbg-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 176.3 GB at this context; duplicate traces of the same session: hf9dlog8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 509.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 509.0 21.5 176.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/368tvpbg",
   "source_url": "https://omlx.ai/benchmarks/performance/368tvpbg",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-c5uo3kfi-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.6 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 435.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 435.2 31.3 173.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c5uo3kfi",
   "source_url": "https://omlx.ai/benchmarks/performance/c5uo3kfi",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-c5uo3kfi-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 562.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 562.6 27.2 174.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c5uo3kfi",
   "source_url": "https://omlx.ai/benchmarks/performance/c5uo3kfi",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-c5uo3kfi-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.0 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 545.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 545.3 27.0 175.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c5uo3kfi",
   "source_url": "https://omlx.ai/benchmarks/performance/c5uo3kfi",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-c5uo3kfi-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 536.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 536.9 27.0 175.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c5uo3kfi",
   "source_url": "https://omlx.ai/benchmarks/performance/c5uo3kfi",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-71soa0o2-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.7 GB at this context; duplicate traces of the same session: 2x0c20ae, bqg5sfd1, fzwjewvb, o0wgl2q0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 440.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 440.1 40.8 173.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/71soa0o2",
   "source_url": "https://omlx.ai/benchmarks/performance/71soa0o2",
   "tg_tokens": null,
   "tps": 40.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-71soa0o2-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.8 GB at this context; duplicate traces of the same session: 2x0c20ae, bqg5sfd1, fzwjewvb, o0wgl2q0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 566.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 566.5 39.0 174.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/71soa0o2",
   "source_url": "https://omlx.ai/benchmarks/performance/71soa0o2",
   "tg_tokens": null,
   "tps": 39.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-71soa0o2-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.1 GB at this context; duplicate traces of the same session: 2x0c20ae, bqg5sfd1, fzwjewvb, o0wgl2q0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 518.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 518.4 38.8 175.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/71soa0o2",
   "source_url": "https://omlx.ai/benchmarks/performance/71soa0o2",
   "tg_tokens": null,
   "tps": 38.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-71soa0o2-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.2 GB at this context; duplicate traces of the same session: 2x0c20ae, bqg5sfd1, fzwjewvb, o0wgl2q0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 550.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 550.4 38.6 175.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/71soa0o2",
   "source_url": "https://omlx.ai/benchmarks/performance/71soa0o2",
   "tg_tokens": null,
   "tps": 38.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-71soa0o2-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.3 GB at this context; duplicate traces of the same session: 2x0c20ae, bqg5sfd1, fzwjewvb, o0wgl2q0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 542.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 542.7 38.3 175.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/71soa0o2",
   "source_url": "https://omlx.ai/benchmarks/performance/71soa0o2",
   "tg_tokens": null,
   "tps": 38.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-71soa0o2-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.6 GB at this context; duplicate traces of the same session: 2x0c20ae, bqg5sfd1, fzwjewvb, o0wgl2q0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 534.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 534.4 37.4 175.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/71soa0o2",
   "source_url": "https://omlx.ai/benchmarks/performance/71soa0o2",
   "tg_tokens": null,
   "tps": 37.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-71soa0o2-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 176.4 GB at this context; duplicate traces of the same session: 2x0c20ae, bqg5sfd1, fzwjewvb, o0wgl2q0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 510.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 510.7 35.9 176.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/71soa0o2",
   "source_url": "https://omlx.ai/benchmarks/performance/71soa0o2",
   "tg_tokens": null,
   "tps": 35.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-c0q5rnzx-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 909.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 909.9 107.3 3.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c0q5rnzx",
   "source_url": "https://omlx.ai/benchmarks/performance/c0q5rnzx",
   "tg_tokens": null,
   "tps": 107.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-c0q5rnzx-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.8 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 842.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 842.7 99.9 4.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c0q5rnzx",
   "source_url": "https://omlx.ai/benchmarks/performance/c0q5rnzx",
   "tg_tokens": null,
   "tps": 99.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-lm719gom-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.6 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1012.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,012 71.0 35.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lm719gom",
   "source_url": "https://omlx.ai/benchmarks/performance/lm719gom",
   "tg_tokens": null,
   "tps": 71.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-lm719gom-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1010.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 1,010 63.3 37.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lm719gom",
   "source_url": "https://omlx.ai/benchmarks/performance/lm719gom",
   "tg_tokens": null,
   "tps": 63.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-hdqs336a-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.6 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1011.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,011 70.4 35.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hdqs336a",
   "source_url": "https://omlx.ai/benchmarks/performance/hdqs336a",
   "tg_tokens": null,
   "tps": 70.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-hdqs336a-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1010.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 1,010 63.1 37.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hdqs336a",
   "source_url": "https://omlx.ai/benchmarks/performance/hdqs336a",
   "tg_tokens": null,
   "tps": 63.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-vlb2tzum-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.6 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1004.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,004 69.8 35.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vlb2tzum",
   "source_url": "https://omlx.ai/benchmarks/performance/vlb2tzum",
   "tg_tokens": null,
   "tps": 69.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-vlb2tzum-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 983.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 983.9 62.0 37.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vlb2tzum",
   "source_url": "https://omlx.ai/benchmarks/performance/vlb2tzum",
   "tg_tokens": null,
   "tps": 62.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-18p1mt3z-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.6 GB at this context; duplicate traces of the same session: smt0ywe9; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1014.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,014 70.7 35.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/18p1mt3z",
   "source_url": "https://omlx.ai/benchmarks/performance/18p1mt3z",
   "tg_tokens": null,
   "tps": 70.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-18p1mt3z-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: smt0ywe9; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 997.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 997.7 63.7 37.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/18p1mt3z",
   "source_url": "https://omlx.ai/benchmarks/performance/18p1mt3z",
   "tg_tokens": null,
   "tps": 63.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-z2v7hqio-ctx1024",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4163.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 4,163 192.7 19.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z2v7hqio",
   "source_url": "https://omlx.ai/benchmarks/performance/z2v7hqio",
   "tg_tokens": null,
   "tps": 192.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-z2v7hqio-ctx4096",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5736.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 5,736 159.3 20.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z2v7hqio",
   "source_url": "https://omlx.ai/benchmarks/performance/z2v7hqio",
   "tg_tokens": null,
   "tps": 159.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-z2v7hqio-ctx8192",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6112.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 8k 6,112 203.0 20.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z2v7hqio",
   "source_url": "https://omlx.ai/benchmarks/performance/z2v7hqio",
   "tg_tokens": null,
   "tps": 203.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-z2v7hqio-ctx16384",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5607.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 5,607 195.0 20.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z2v7hqio",
   "source_url": "https://omlx.ai/benchmarks/performance/z2v7hqio",
   "tg_tokens": null,
   "tps": 195.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-z2v7hqio-ctx32768",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4807.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 4,807 170.8 21.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z2v7hqio",
   "source_url": "https://omlx.ai/benchmarks/performance/z2v7hqio",
   "tg_tokens": null,
   "tps": 170.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-z2v7hqio-ctx65536",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3580.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 64k 3,580 126.1 22.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z2v7hqio",
   "source_url": "https://omlx.ai/benchmarks/performance/z2v7hqio",
   "tg_tokens": null,
   "tps": 126.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-z2v7hqio-ctx131072",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2310.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 128k 2,310 87.1 23.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z2v7hqio",
   "source_url": "https://omlx.ai/benchmarks/performance/z2v7hqio",
   "tg_tokens": null,
   "tps": 87.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-z2v7hqio-ctx199680",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1579.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 195k 1,579 59.7 25.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z2v7hqio",
   "source_url": "https://omlx.ai/benchmarks/performance/z2v7hqio",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yu64gvip-ctx1024",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: ur3j8yno; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2615.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 2,615 151.1 19.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yu64gvip",
   "source_url": "https://omlx.ai/benchmarks/performance/yu64gvip",
   "tg_tokens": null,
   "tps": 151.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-yu64gvip-ctx4096",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: ur3j8yno; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2684.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 2,684 173.6 20.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yu64gvip",
   "source_url": "https://omlx.ai/benchmarks/performance/yu64gvip",
   "tg_tokens": null,
   "tps": 173.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-yu64gvip-ctx8192",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: ur3j8yno; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4567.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 8k 4,567 190.8 20.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yu64gvip",
   "source_url": "https://omlx.ai/benchmarks/performance/yu64gvip",
   "tg_tokens": null,
   "tps": 190.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-yu64gvip-ctx16384",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: ur3j8yno; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4610.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 4,610 177.6 20.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yu64gvip",
   "source_url": "https://omlx.ai/benchmarks/performance/yu64gvip",
   "tg_tokens": null,
   "tps": 177.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-yu64gvip-ctx32768",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: ur3j8yno; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4019.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 4,019 164.2 21.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yu64gvip",
   "source_url": "https://omlx.ai/benchmarks/performance/yu64gvip",
   "tg_tokens": null,
   "tps": 164.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-yu64gvip-ctx65536",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: ur3j8yno; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2973.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 64k 2,973 121.2 21.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yu64gvip",
   "source_url": "https://omlx.ai/benchmarks/performance/yu64gvip",
   "tg_tokens": null,
   "tps": 121.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-yu64gvip-ctx131072",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: ur3j8yno; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2042.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 128k 2,042 91.8 23.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yu64gvip",
   "source_url": "https://omlx.ai/benchmarks/performance/yu64gvip",
   "tg_tokens": null,
   "tps": 91.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-yu64gvip-ctx199680",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: ur3j8yno; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1432.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 195k 1,432 62.6 24.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yu64gvip",
   "source_url": "https://omlx.ai/benchmarks/performance/yu64gvip",
   "tg_tokens": null,
   "tps": 62.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-7ua722c9-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 26.9 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1223.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 5bit; 1k 1,223 107.6 26.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ua722c9",
   "source_url": "https://omlx.ai/benchmarks/performance/7ua722c9",
   "tg_tokens": null,
   "tps": 107.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-7ua722c9-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 29.0 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1349.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 5bit; 4k 1,349 73.4 29.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ua722c9",
   "source_url": "https://omlx.ai/benchmarks/performance/7ua722c9",
   "tg_tokens": null,
   "tps": 73.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-mneqkdkl-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 26.9 GB at this context; duplicate traces of the same session: 0ek42bka; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1239.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 5bit; 1k 1,239 111.2 26.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mneqkdkl",
   "source_url": "https://omlx.ai/benchmarks/performance/mneqkdkl",
   "tg_tokens": null,
   "tps": 111.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-mneqkdkl-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 29.0 GB at this context; duplicate traces of the same session: 0ek42bka; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1320.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 5bit; 4k 1,320 67.7 29.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mneqkdkl",
   "source_url": "https://omlx.ai/benchmarks/performance/mneqkdkl",
   "tg_tokens": null,
   "tps": 67.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-wx81v0y0-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 26.9 GB at this context; duplicate traces of the same session: ell1qhlt; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1044.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 5bit; 1k 1,044 65.6 26.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wx81v0y0",
   "source_url": "https://omlx.ai/benchmarks/performance/wx81v0y0",
   "tg_tokens": null,
   "tps": 65.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-wx81v0y0-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 29.0 GB at this context; duplicate traces of the same session: ell1qhlt; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1353.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 5bit; 4k 1,353 65.8 29.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wx81v0y0",
   "source_url": "https://omlx.ai/benchmarks/performance/wx81v0y0",
   "tg_tokens": null,
   "tps": 65.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2jpgm1iy-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 38.8 GB at this context; duplicate traces of the same session: 8g81cet6, n3vt3t67; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 320.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 320.3 38.5 38.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2jpgm1iy",
   "source_url": "https://omlx.ai/benchmarks/performance/2jpgm1iy",
   "tg_tokens": null,
   "tps": 38.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2jpgm1iy-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 44.4 GB at this context; duplicate traces of the same session: 8g81cet6, n3vt3t67; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 335.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 335.2 32.8 44.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2jpgm1iy",
   "source_url": "https://omlx.ai/benchmarks/performance/2jpgm1iy",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2jpgm1iy-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 44.9 GB at this context; duplicate traces of the same session: 8g81cet6, n3vt3t67; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 329.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 329.9 34.2 44.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2jpgm1iy",
   "source_url": "https://omlx.ai/benchmarks/performance/2jpgm1iy",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-2s7lrp0f-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 127.5 8.5 28.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2s7lrp0f",
   "source_url": "https://omlx.ai/benchmarks/performance/2s7lrp0f",
   "tg_tokens": null,
   "tps": 8.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-2s7lrp0f-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.9 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 126.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 126.7 8.5 31.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2s7lrp0f",
   "source_url": "https://omlx.ai/benchmarks/performance/2s7lrp0f",
   "tg_tokens": null,
   "tps": 8.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-2s7lrp0f-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.5 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 112.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 112.7 8.2 32.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2s7lrp0f",
   "source_url": "https://omlx.ai/benchmarks/performance/2s7lrp0f",
   "tg_tokens": null,
   "tps": 8.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-i3rpb752-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 26.9 GB; TTFT 48719 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 672.6,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-27B-NVFP4 on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 nvfp4; 32k tokens 672.6 PP tok/s 28.1 TG tok/s 48719 TTFT (ms) 26.9 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i3rpb752",
   "source_url": "https://omlx.ai/benchmarks/performance/i3rpb752",
   "tg_tokens": null,
   "tps": 28.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-kmfk1g6h-ctx16384",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 40.3 GB; TTFT 12123 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1351.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k tokens 1,351 PP tok/s 45.8 TG tok/s 12123 TTFT (ms) 40.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kmfk1g6h",
   "source_url": "https://omlx.ai/benchmarks/performance/kmfk1g6h",
   "tg_tokens": null,
   "tps": 45.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-g7syk1p1-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 39.8 GB; TTFT 1498 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 683.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k tokens 683.5 PP tok/s 45.0 TG tok/s 1498 TTFT (ms) 39.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g7syk1p1",
   "source_url": "https://omlx.ai/benchmarks/performance/g7syk1p1",
   "tg_tokens": null,
   "tps": 45.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-z1k5reyj-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 46.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 46.9 3.3 33.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z1k5reyj",
   "source_url": "https://omlx.ai/benchmarks/performance/z1k5reyj",
   "tg_tokens": null,
   "tps": 3.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-z1k5reyj-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 40.5 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 59.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 59.7 3.5 40.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z1k5reyj",
   "source_url": "https://omlx.ai/benchmarks/performance/z1k5reyj",
   "tg_tokens": null,
   "tps": 3.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-1ug4zn1h-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: i2btl46w; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 46.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 46.5 3.5 24.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ug4zn1h",
   "source_url": "https://omlx.ai/benchmarks/performance/1ug4zn1h",
   "tg_tokens": null,
   "tps": 3.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-1ug4zn1h-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 32.4 GB at this context; duplicate traces of the same session: i2btl46w; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 79.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 79.3 4.0 32.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ug4zn1h",
   "source_url": "https://omlx.ai/benchmarks/performance/1ug4zn1h",
   "tg_tokens": null,
   "tps": 4.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-81s1szm8-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: 7nno64su; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1110.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,110 91.4 100.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/81s1szm8",
   "source_url": "https://omlx.ai/benchmarks/performance/81s1szm8",
   "tg_tokens": null,
   "tps": 91.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-81s1szm8-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: 7nno64su; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1331.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,331 91.5 101.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/81s1szm8",
   "source_url": "https://omlx.ai/benchmarks/performance/81s1szm8",
   "tg_tokens": null,
   "tps": 91.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-81s1szm8-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: 7nno64su; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1373.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 1,373 93.4 101.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/81s1szm8",
   "source_url": "https://omlx.ai/benchmarks/performance/81s1szm8",
   "tg_tokens": null,
   "tps": 93.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-81s1szm8-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.6 GB at this context; duplicate traces of the same session: 7nno64su; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1364.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 1,364 68.7 101.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/81s1szm8",
   "source_url": "https://omlx.ai/benchmarks/performance/81s1szm8",
   "tg_tokens": null,
   "tps": 68.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rpmo4ree-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.1 GB at this context; duplicate traces of the same session: 0nr9oj75, 7wgv4sru, ilmo0jya, td1w1309, yxwvdyh8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1143.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,143 87.6 99.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rpmo4ree",
   "source_url": "https://omlx.ai/benchmarks/performance/rpmo4ree",
   "tg_tokens": null,
   "tps": 87.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rpmo4ree-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.7 GB at this context; duplicate traces of the same session: 0nr9oj75, 7wgv4sru, ilmo0jya, td1w1309, yxwvdyh8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1423.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 1,423 78.7 99.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rpmo4ree",
   "source_url": "https://omlx.ai/benchmarks/performance/rpmo4ree",
   "tg_tokens": null,
   "tps": 78.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rpmo4ree-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.0 GB at this context; duplicate traces of the same session: 0nr9oj75, 7wgv4sru, ilmo0jya, td1w1309, yxwvdyh8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1416.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 1,416 76.7 100.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rpmo4ree",
   "source_url": "https://omlx.ai/benchmarks/performance/rpmo4ree",
   "tg_tokens": null,
   "tps": 76.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rpmo4ree-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: 0nr9oj75, 7wgv4sru, ilmo0jya, td1w1309, yxwvdyh8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1397.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 1,397 50.8 100.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rpmo4ree",
   "source_url": "https://omlx.ai/benchmarks/performance/rpmo4ree",
   "tg_tokens": null,
   "tps": 50.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rpmo4ree-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: 0nr9oj75, 7wgv4sru, ilmo0jya, td1w1309, yxwvdyh8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1393.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 1,393 58.7 101.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rpmo4ree",
   "source_url": "https://omlx.ai/benchmarks/performance/rpmo4ree",
   "tg_tokens": null,
   "tps": 58.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rpmo4ree-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.7 GB at this context; duplicate traces of the same session: 0nr9oj75, 7wgv4sru, ilmo0jya, td1w1309, yxwvdyh8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1351.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 1,351 66.1 104.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rpmo4ree",
   "source_url": "https://omlx.ai/benchmarks/performance/rpmo4ree",
   "tg_tokens": null,
   "tps": 66.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dfqcx7y0-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.6 GB; TTFT 2652 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 386.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 386.2 PP tok/s 68.9 TG tok/s 2652 TTFT (ms) 70.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dfqcx7y0",
   "source_url": "https://omlx.ai/benchmarks/performance/dfqcx7y0",
   "tg_tokens": null,
   "tps": 68.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-zrk3x3m0-ctx1024",
   "model": "Ternary-Bonsai-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 10.2 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 310.9,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-27B-mlx-2bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 2bit; 1k 310.9 11.5 10.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zrk3x3m0",
   "source_url": "https://omlx.ai/benchmarks/performance/zrk3x3m0",
   "tg_tokens": null,
   "tps": 11.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-zrk3x3m0-ctx4096",
   "model": "Ternary-Bonsai-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 317.9,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-27B-mlx-2bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 2bit; 4k 317.9 13.4 11.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zrk3x3m0",
   "source_url": "https://omlx.ai/benchmarks/performance/zrk3x3m0",
   "tg_tokens": null,
   "tps": 13.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vla6c6wz-ctx1024",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "it variant; oMLX community benchmark trace; peak memory 15.4 GB at this context; duplicate traces of the same session: 5ajwahti, 7tbob2la, tghs39f7, u3zummpr, yxu8xe7g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2796.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "diffusiongemma-26B-A4B-it-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 2,796 27.7 15.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vla6c6wz",
   "source_url": "https://omlx.ai/benchmarks/performance/vla6c6wz",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vla6c6wz-ctx4096",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "it variant; oMLX community benchmark trace; peak memory 16.2 GB at this context; duplicate traces of the same session: 5ajwahti, 7tbob2la, tghs39f7, u3zummpr, yxu8xe7g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2551.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "diffusiongemma-26B-A4B-it-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 2,551 20.8 16.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vla6c6wz",
   "source_url": "https://omlx.ai/benchmarks/performance/vla6c6wz",
   "tg_tokens": null,
   "tps": 20.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vla6c6wz-ctx8192",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "it variant; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 5ajwahti, 7tbob2la, tghs39f7, u3zummpr, yxu8xe7g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2437.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "diffusiongemma-26B-A4B-it-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 2,437 23.5 16.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vla6c6wz",
   "source_url": "https://omlx.ai/benchmarks/performance/vla6c6wz",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vla6c6wz-ctx16384",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "it variant; oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 5ajwahti, 7tbob2la, tghs39f7, u3zummpr, yxu8xe7g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2323.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "diffusiongemma-26B-A4B-it-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 2,323 47.6 17.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vla6c6wz",
   "source_url": "https://omlx.ai/benchmarks/performance/vla6c6wz",
   "tg_tokens": null,
   "tps": 47.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vla6c6wz-ctx32768",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "it variant; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 5ajwahti, 7tbob2la, tghs39f7, u3zummpr, yxu8xe7g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2053.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "diffusiongemma-26B-A4B-it-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 2,053 18.8 18.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vla6c6wz",
   "source_url": "https://omlx.ai/benchmarks/performance/vla6c6wz",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vla6c6wz-ctx65536",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "it variant; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 5ajwahti, 7tbob2la, tghs39f7, u3zummpr, yxu8xe7g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1621.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "diffusiongemma-26B-A4B-it-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 1,621 22.2 21.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vla6c6wz",
   "source_url": "https://omlx.ai/benchmarks/performance/vla6c6wz",
   "tg_tokens": null,
   "tps": 22.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vla6c6wz-ctx131072",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "it variant; oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: 5ajwahti, 7tbob2la, tghs39f7, u3zummpr, yxu8xe7g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1097.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "diffusiongemma-26B-A4B-it-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 1,097 10.0 26.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vla6c6wz",
   "source_url": "https://omlx.ai/benchmarks/performance/vla6c6wz",
   "tg_tokens": null,
   "tps": 10.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-g92bg1r7-ctx1024",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 7.3 GB at this context; duplicate traces of the same session: nvlfvd52; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 383.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 383.4 34.0 7.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g92bg1r7",
   "source_url": "https://omlx.ai/benchmarks/performance/g92bg1r7",
   "tg_tokens": null,
   "tps": 34.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-g92bg1r7-ctx4096",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 7.7 GB at this context; duplicate traces of the same session: nvlfvd52; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 400.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 400.4 35.8 7.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g92bg1r7",
   "source_url": "https://omlx.ai/benchmarks/performance/g92bg1r7",
   "tg_tokens": null,
   "tps": 35.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-mvmvlcb1-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1292.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,292 47.8 14.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mvmvlcb1",
   "source_url": "https://omlx.ai/benchmarks/performance/mvmvlcb1",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-mvmvlcb1-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 15.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1193.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,193 133.4 15.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mvmvlcb1",
   "source_url": "https://omlx.ai/benchmarks/performance/mvmvlcb1",
   "tg_tokens": null,
   "tps": 133.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-v747xhjc-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: 258dc6ib; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 981.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 981.2 85.1 14.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v747xhjc",
   "source_url": "https://omlx.ai/benchmarks/performance/v747xhjc",
   "tg_tokens": null,
   "tps": 85.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-v747xhjc-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: 258dc6ib; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1109.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,109 78.6 14.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v747xhjc",
   "source_url": "https://omlx.ai/benchmarks/performance/v747xhjc",
   "tg_tokens": null,
   "tps": 78.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-lbrvq4q9-ctx1024",
   "model": "gemma-4-31B-it",
   "notes": "Mystery-Fine-Tune-HERETIC-UNCENSORED-Thinking-qx86-hi build; oMLX community benchmark trace; peak memory 28.4 GB at this context; duplicate traces of the same session: q6oi51tt; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 72.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-Mystery-Fine-Tune-HERETIC-UNCENSORED-Thinking-qx86-hi-mlx on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 1k 72.2 7.9 28.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lbrvq4q9",
   "source_url": "https://omlx.ai/benchmarks/performance/lbrvq4q9",
   "tg_tokens": null,
   "tps": 7.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-lbrvq4q9-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "Mystery-Fine-Tune-HERETIC-UNCENSORED-Thinking-qx86-hi build; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: q6oi51tt; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 67.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-Mystery-Fine-Tune-HERETIC-UNCENSORED-Thinking-qx86-hi-mlx on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k 67.1 7.3 29.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lbrvq4q9",
   "source_url": "https://omlx.ai/benchmarks/performance/lbrvq4q9",
   "tg_tokens": null,
   "tps": 7.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-sw7ujvv2-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 6.2 GB; TTFT 2421 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1692.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E4B-it-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 1,692 PP tok/s 59.1 TG tok/s 2421 TTFT (ms) 6.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sw7ujvv2",
   "source_url": "https://omlx.ai/benchmarks/performance/sw7ujvv2",
   "tg_tokens": null,
   "tps": 59.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-oszjta3n-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "ornith-ai repo; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2972.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "ornith-ai--Ornith-1.5-9B-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 unknown; 1k 2,972 33.7 17.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oszjta3n",
   "source_url": "https://omlx.ai/benchmarks/performance/oszjta3n",
   "tg_tokens": null,
   "tps": 33.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-oszjta3n-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "ornith-ai repo; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2601.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "ornith-ai--Ornith-1.5-9B-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 unknown; 4k 2,601 33.6 18.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oszjta3n",
   "source_url": "https://omlx.ai/benchmarks/performance/oszjta3n",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-oszjta3n-ctx8192",
   "model": "Ornith-1.5-9B",
   "notes": "ornith-ai repo; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3675.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "ornith-ai--Ornith-1.5-9B-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 unknown; 8k 3,675 33.3 18.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oszjta3n",
   "source_url": "https://omlx.ai/benchmarks/performance/oszjta3n",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-oszjta3n-ctx16384",
   "model": "Ornith-1.5-9B",
   "notes": "ornith-ai repo; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3576.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "ornith-ai--Ornith-1.5-9B-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 unknown; 16k 3,576 32.7 18.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oszjta3n",
   "source_url": "https://omlx.ai/benchmarks/performance/oszjta3n",
   "tg_tokens": null,
   "tps": 32.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-oszjta3n-ctx32768",
   "model": "Ornith-1.5-9B",
   "notes": "ornith-ai repo; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3333.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "ornith-ai--Ornith-1.5-9B-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 unknown; 32k 3,333 31.5 19.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oszjta3n",
   "source_url": "https://omlx.ai/benchmarks/performance/oszjta3n",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-w242c687-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 403.3 GB at this context; duplicate traces of the same session: 24512zra, 2puocypl; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 360.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 360.6 27.4 403.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w242c687",
   "source_url": "https://omlx.ai/benchmarks/performance/w242c687",
   "tg_tokens": null,
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-w242c687-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 403.7 GB at this context; duplicate traces of the same session: 24512zra, 2puocypl; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 700.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 700.0 38.7 403.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w242c687",
   "source_url": "https://omlx.ai/benchmarks/performance/w242c687",
   "tg_tokens": null,
   "tps": 38.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-w242c687-ctx65536",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 403.8 GB at this context; duplicate traces of the same session: 24512zra, 2puocypl; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 721.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 64k 721.0 40.2 403.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w242c687",
   "source_url": "https://omlx.ai/benchmarks/performance/w242c687",
   "tg_tokens": null,
   "tps": 40.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hf9dlog8-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.6 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 395.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 395.6 31.5 173.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hf9dlog8",
   "source_url": "https://omlx.ai/benchmarks/performance/hf9dlog8",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hf9dlog8-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 557.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 557.8 27.3 174.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hf9dlog8",
   "source_url": "https://omlx.ai/benchmarks/performance/hf9dlog8",
   "tg_tokens": null,
   "tps": 27.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hf9dlog8-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.0 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 506.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 506.7 27.3 175.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hf9dlog8",
   "source_url": "https://omlx.ai/benchmarks/performance/hf9dlog8",
   "tg_tokens": null,
   "tps": 27.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hf9dlog8-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 537.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 537.0 27.0 175.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hf9dlog8",
   "source_url": "https://omlx.ai/benchmarks/performance/hf9dlog8",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hf9dlog8-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.2 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 528.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 528.1 26.9 175.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hf9dlog8",
   "source_url": "https://omlx.ai/benchmarks/performance/hf9dlog8",
   "tg_tokens": null,
   "tps": 26.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hf9dlog8-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.6 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 518.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 518.7 25.2 175.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hf9dlog8",
   "source_url": "https://omlx.ai/benchmarks/performance/hf9dlog8",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hf9dlog8-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 176.3 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 509.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 509.0 21.5 176.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hf9dlog8",
   "source_url": "https://omlx.ai/benchmarks/performance/hf9dlog8",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-bqg5sfd1-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.7 GB at this context; duplicate traces of the same session: 2x0c20ae, fzwjewvb, o0wgl2q0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 440.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 440.1 40.8 173.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bqg5sfd1",
   "source_url": "https://omlx.ai/benchmarks/performance/bqg5sfd1",
   "tg_tokens": null,
   "tps": 40.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-bqg5sfd1-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.8 GB at this context; duplicate traces of the same session: 2x0c20ae, fzwjewvb, o0wgl2q0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 566.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 566.5 39.0 174.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bqg5sfd1",
   "source_url": "https://omlx.ai/benchmarks/performance/bqg5sfd1",
   "tg_tokens": null,
   "tps": 39.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-bqg5sfd1-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.1 GB at this context; duplicate traces of the same session: 2x0c20ae, fzwjewvb, o0wgl2q0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 518.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 518.4 38.8 175.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bqg5sfd1",
   "source_url": "https://omlx.ai/benchmarks/performance/bqg5sfd1",
   "tg_tokens": null,
   "tps": 38.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-bqg5sfd1-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.2 GB at this context; duplicate traces of the same session: 2x0c20ae, fzwjewvb, o0wgl2q0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 550.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 550.4 38.6 175.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bqg5sfd1",
   "source_url": "https://omlx.ai/benchmarks/performance/bqg5sfd1",
   "tg_tokens": null,
   "tps": 38.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-bqg5sfd1-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.3 GB at this context; duplicate traces of the same session: 2x0c20ae, fzwjewvb, o0wgl2q0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 542.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 542.7 38.3 175.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bqg5sfd1",
   "source_url": "https://omlx.ai/benchmarks/performance/bqg5sfd1",
   "tg_tokens": null,
   "tps": 38.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-bqg5sfd1-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.6 GB at this context; duplicate traces of the same session: 2x0c20ae, fzwjewvb, o0wgl2q0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 534.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 534.4 37.4 175.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bqg5sfd1",
   "source_url": "https://omlx.ai/benchmarks/performance/bqg5sfd1",
   "tg_tokens": null,
   "tps": 37.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-bqg5sfd1-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 176.4 GB at this context; duplicate traces of the same session: 2x0c20ae, fzwjewvb, o0wgl2q0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 510.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 510.7 35.9 176.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bqg5sfd1",
   "source_url": "https://omlx.ai/benchmarks/performance/bqg5sfd1",
   "tg_tokens": null,
   "tps": 35.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-fna9jfwl-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: vxov0l91; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1005.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,005 69.5 35.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fna9jfwl",
   "source_url": "https://omlx.ai/benchmarks/performance/fna9jfwl",
   "tg_tokens": null,
   "tps": 69.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-fna9jfwl-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: vxov0l91; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 974.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 974.5 61.9 37.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fna9jfwl",
   "source_url": "https://omlx.ai/benchmarks/performance/fna9jfwl",
   "tg_tokens": null,
   "tps": 61.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-92xj4l2z-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4526.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 4,526 111.6 28.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/92xj4l2z",
   "source_url": "https://omlx.ai/benchmarks/performance/92xj4l2z",
   "tg_tokens": null,
   "tps": 111.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-92xj4l2z-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4429.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 4,429 97.7 29.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/92xj4l2z",
   "source_url": "https://omlx.ai/benchmarks/performance/92xj4l2z",
   "tg_tokens": null,
   "tps": 97.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-wpas0va5-ctx1024",
   "model": "Qwen 3 30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: xcvumdyu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1276.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-30B-A3B-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 1,276 108.4 16.6 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wpas0va5",
   "source_url": "https://omlx.ai/benchmarks/performance/wpas0va5",
   "tg_tokens": null,
   "tps": 108.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-wpas0va5-ctx4096",
   "model": "Qwen 3 30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: xcvumdyu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1331.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-30B-A3B-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 1,331 94.5 17.0 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wpas0va5",
   "source_url": "https://omlx.ai/benchmarks/performance/wpas0va5",
   "tg_tokens": null,
   "tps": 94.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-smt0ywe9-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.6 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1014.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,014 70.7 35.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/smt0ywe9",
   "source_url": "https://omlx.ai/benchmarks/performance/smt0ywe9",
   "tg_tokens": null,
   "tps": 70.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-smt0ywe9-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 997.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 997.7 63.7 37.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/smt0ywe9",
   "source_url": "https://omlx.ai/benchmarks/performance/smt0ywe9",
   "tg_tokens": null,
   "tps": 63.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-h3lx0w46-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: e1luerc3; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1011.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,011 70.1 35.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h3lx0w46",
   "source_url": "https://omlx.ai/benchmarks/performance/h3lx0w46",
   "tg_tokens": null,
   "tps": 70.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-h3lx0w46-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: e1luerc3; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 966.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 966.7 61.1 37.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h3lx0w46",
   "source_url": "https://omlx.ai/benchmarks/performance/h3lx0w46",
   "tg_tokens": null,
   "tps": 61.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ur3j8yno-ctx1024",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2615.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 2,615 151.1 19.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ur3j8yno",
   "source_url": "https://omlx.ai/benchmarks/performance/ur3j8yno",
   "tg_tokens": null,
   "tps": 151.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ur3j8yno-ctx4096",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2684.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 2,684 173.6 20.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ur3j8yno",
   "source_url": "https://omlx.ai/benchmarks/performance/ur3j8yno",
   "tg_tokens": null,
   "tps": 173.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ur3j8yno-ctx8192",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4567.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 8k 4,567 190.8 20.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ur3j8yno",
   "source_url": "https://omlx.ai/benchmarks/performance/ur3j8yno",
   "tg_tokens": null,
   "tps": 190.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ur3j8yno-ctx16384",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4610.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 4,610 177.6 20.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ur3j8yno",
   "source_url": "https://omlx.ai/benchmarks/performance/ur3j8yno",
   "tg_tokens": null,
   "tps": 177.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ur3j8yno-ctx32768",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4019.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 4,019 164.2 21.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ur3j8yno",
   "source_url": "https://omlx.ai/benchmarks/performance/ur3j8yno",
   "tg_tokens": null,
   "tps": 164.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ur3j8yno-ctx65536",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2973.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 64k 2,973 121.2 21.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ur3j8yno",
   "source_url": "https://omlx.ai/benchmarks/performance/ur3j8yno",
   "tg_tokens": null,
   "tps": 121.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ur3j8yno-ctx131072",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2042.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 128k 2,042 91.8 23.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ur3j8yno",
   "source_url": "https://omlx.ai/benchmarks/performance/ur3j8yno",
   "tg_tokens": null,
   "tps": 91.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ur3j8yno-ctx199680",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1432.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 195k 1,432 62.6 24.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ur3j8yno",
   "source_url": "https://omlx.ai/benchmarks/performance/ur3j8yno",
   "tg_tokens": null,
   "tps": 62.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-0ek42bka-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 26.9 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1239.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 5bit; 1k 1,239 111.2 26.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ek42bka",
   "source_url": "https://omlx.ai/benchmarks/performance/0ek42bka",
   "tg_tokens": null,
   "tps": 111.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-0ek42bka-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 29.0 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1320.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 5bit; 4k 1,320 67.7 29.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ek42bka",
   "source_url": "https://omlx.ai/benchmarks/performance/0ek42bka",
   "tg_tokens": null,
   "tps": 67.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-ell1qhlt-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 26.9 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1044.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 5bit; 1k 1,044 65.6 26.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ell1qhlt",
   "source_url": "https://omlx.ai/benchmarks/performance/ell1qhlt",
   "tg_tokens": null,
   "tps": 65.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-ell1qhlt-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 29.0 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1353.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 5bit; 4k 1,353 65.8 29.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ell1qhlt",
   "source_url": "https://omlx.ai/benchmarks/performance/ell1qhlt",
   "tg_tokens": null,
   "tps": 65.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-l4hxwqvb-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: fw6fy7j7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 431.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 431.0 13.6 17.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l4hxwqvb",
   "source_url": "https://omlx.ai/benchmarks/performance/l4hxwqvb",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-l4hxwqvb-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: fw6fy7j7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 362.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 362.8 12.2 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l4hxwqvb",
   "source_url": "https://omlx.ai/benchmarks/performance/l4hxwqvb",
   "tg_tokens": null,
   "tps": 12.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-n3vt3t67-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 38.8 GB at this context; duplicate traces of the same session: 8g81cet6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 320.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 320.3 38.5 38.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n3vt3t67",
   "source_url": "https://omlx.ai/benchmarks/performance/n3vt3t67",
   "tg_tokens": null,
   "tps": 38.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-n3vt3t67-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 44.4 GB at this context; duplicate traces of the same session: 8g81cet6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 335.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 335.2 32.8 44.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n3vt3t67",
   "source_url": "https://omlx.ai/benchmarks/performance/n3vt3t67",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-n3vt3t67-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 44.9 GB at this context; duplicate traces of the same session: 8g81cet6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 329.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 329.9 34.2 44.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n3vt3t67",
   "source_url": "https://omlx.ai/benchmarks/performance/n3vt3t67",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-74viq9p6-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 2pj5zttg; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 210.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 210.1 22.6 16.0 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/74viq9p6",
   "source_url": "https://omlx.ai/benchmarks/performance/74viq9p6",
   "tg_tokens": null,
   "tps": 22.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-74viq9p6-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 2pj5zttg; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 216.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 216.5 22.1 17.0 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/74viq9p6",
   "source_url": "https://omlx.ai/benchmarks/performance/74viq9p6",
   "tg_tokens": null,
   "tps": 22.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-i9jcmndk-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 24.8 GB; TTFT 13432 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 304.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 304.9 PP tok/s 76.0 TG tok/s 13432 TTFT (ms) 24.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i9jcmndk",
   "source_url": "https://omlx.ai/benchmarks/performance/i9jcmndk",
   "tg_tokens": null,
   "tps": 76.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2jwhqoyw-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 25.2 GB; TTFT 12693 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 322.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 322.7 PP tok/s 75.5 TG tok/s 12693 TTFT (ms) 25.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2jwhqoyw",
   "source_url": "https://omlx.ai/benchmarks/performance/2jwhqoyw",
   "tg_tokens": null,
   "tps": 75.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-8cwgz6ui-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 25.3 GB; TTFT 13179 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 310.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 310.8 PP tok/s 48.3 TG tok/s 13179 TTFT (ms) 25.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8cwgz6ui",
   "source_url": "https://omlx.ai/benchmarks/performance/8cwgz6ui",
   "tg_tokens": null,
   "tps": 48.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-5gdpn9lf-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 25.2 GB; TTFT 12601 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 325.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 325.1 PP tok/s 68.0 TG tok/s 12601 TTFT (ms) 25.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5gdpn9lf",
   "source_url": "https://omlx.ai/benchmarks/performance/5gdpn9lf",
   "tg_tokens": null,
   "tps": 68.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-i2btl46w-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 46.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 46.5 3.5 24.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i2btl46w",
   "source_url": "https://omlx.ai/benchmarks/performance/i2btl46w",
   "tg_tokens": null,
   "tps": 3.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-i2btl46w-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 32.4 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 79.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 79.3 4.0 32.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i2btl46w",
   "source_url": "https://omlx.ai/benchmarks/performance/i2btl46w",
   "tg_tokens": null,
   "tps": 4.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-7nno64su-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1110.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,110 91.4 100.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7nno64su",
   "source_url": "https://omlx.ai/benchmarks/performance/7nno64su",
   "tg_tokens": null,
   "tps": 91.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-7nno64su-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1331.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,331 91.5 101.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7nno64su",
   "source_url": "https://omlx.ai/benchmarks/performance/7nno64su",
   "tg_tokens": null,
   "tps": 91.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-7nno64su-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1373.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 1,373 93.4 101.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7nno64su",
   "source_url": "https://omlx.ai/benchmarks/performance/7nno64su",
   "tg_tokens": null,
   "tps": 93.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-7nno64su-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.6 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1364.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 1,364 68.7 101.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7nno64su",
   "source_url": "https://omlx.ai/benchmarks/performance/7nno64su",
   "tg_tokens": null,
   "tps": 68.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-7wgv4sru-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.1 GB at this context; duplicate traces of the same session: 0nr9oj75, ilmo0jya, td1w1309, yxwvdyh8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1143.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,143 87.6 99.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7wgv4sru",
   "source_url": "https://omlx.ai/benchmarks/performance/7wgv4sru",
   "tg_tokens": null,
   "tps": 87.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-7wgv4sru-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.7 GB at this context; duplicate traces of the same session: 0nr9oj75, ilmo0jya, td1w1309, yxwvdyh8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1423.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 1,423 78.7 99.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7wgv4sru",
   "source_url": "https://omlx.ai/benchmarks/performance/7wgv4sru",
   "tg_tokens": null,
   "tps": 78.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-7wgv4sru-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.0 GB at this context; duplicate traces of the same session: 0nr9oj75, ilmo0jya, td1w1309, yxwvdyh8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1416.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 1,416 76.7 100.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7wgv4sru",
   "source_url": "https://omlx.ai/benchmarks/performance/7wgv4sru",
   "tg_tokens": null,
   "tps": 76.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-7wgv4sru-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: 0nr9oj75, ilmo0jya, td1w1309, yxwvdyh8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1397.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 1,397 50.8 100.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7wgv4sru",
   "source_url": "https://omlx.ai/benchmarks/performance/7wgv4sru",
   "tg_tokens": null,
   "tps": 50.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-7wgv4sru-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: 0nr9oj75, ilmo0jya, td1w1309, yxwvdyh8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1393.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 1,393 58.7 101.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7wgv4sru",
   "source_url": "https://omlx.ai/benchmarks/performance/7wgv4sru",
   "tg_tokens": null,
   "tps": 58.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-7wgv4sru-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.7 GB at this context; duplicate traces of the same session: 0nr9oj75, ilmo0jya, td1w1309, yxwvdyh8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1351.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 1,351 66.1 104.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7wgv4sru",
   "source_url": "https://omlx.ai/benchmarks/performance/7wgv4sru",
   "tg_tokens": null,
   "tps": 66.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5ajwahti-ctx1024",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "it variant; oMLX community benchmark trace; peak memory 15.4 GB at this context; duplicate traces of the same session: 7tbob2la, tghs39f7, u3zummpr, yxu8xe7g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2796.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "diffusiongemma-26B-A4B-it-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 2,796 27.7 15.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ajwahti",
   "source_url": "https://omlx.ai/benchmarks/performance/5ajwahti",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5ajwahti-ctx4096",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "it variant; oMLX community benchmark trace; peak memory 16.2 GB at this context; duplicate traces of the same session: 7tbob2la, tghs39f7, u3zummpr, yxu8xe7g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2551.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "diffusiongemma-26B-A4B-it-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 2,551 20.8 16.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ajwahti",
   "source_url": "https://omlx.ai/benchmarks/performance/5ajwahti",
   "tg_tokens": null,
   "tps": 20.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5ajwahti-ctx8192",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "it variant; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 7tbob2la, tghs39f7, u3zummpr, yxu8xe7g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2437.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "diffusiongemma-26B-A4B-it-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 2,437 23.5 16.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ajwahti",
   "source_url": "https://omlx.ai/benchmarks/performance/5ajwahti",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5ajwahti-ctx16384",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "it variant; oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 7tbob2la, tghs39f7, u3zummpr, yxu8xe7g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2323.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "diffusiongemma-26B-A4B-it-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 2,323 47.6 17.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ajwahti",
   "source_url": "https://omlx.ai/benchmarks/performance/5ajwahti",
   "tg_tokens": null,
   "tps": 47.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5ajwahti-ctx32768",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "it variant; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 7tbob2la, tghs39f7, u3zummpr, yxu8xe7g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2053.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "diffusiongemma-26B-A4B-it-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 2,053 18.8 18.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ajwahti",
   "source_url": "https://omlx.ai/benchmarks/performance/5ajwahti",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5ajwahti-ctx65536",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "it variant; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 7tbob2la, tghs39f7, u3zummpr, yxu8xe7g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1621.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "diffusiongemma-26B-A4B-it-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 1,621 22.2 21.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ajwahti",
   "source_url": "https://omlx.ai/benchmarks/performance/5ajwahti",
   "tg_tokens": null,
   "tps": 22.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5ajwahti-ctx131072",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "it variant; oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: 7tbob2la, tghs39f7, u3zummpr, yxu8xe7g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1097.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "diffusiongemma-26B-A4B-it-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 1,097 10.0 26.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ajwahti",
   "source_url": "https://omlx.ai/benchmarks/performance/5ajwahti",
   "tg_tokens": null,
   "tps": 10.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-zd0qnr5g-ctx16384",
   "model": "Gemma 4 12B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 13.8 GB; TTFT 39727 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 412.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-12B-it-8bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 8bit; 16k tokens 412.4 PP tok/s 25.1 TG tok/s 39727 TTFT (ms) 13.8 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zd0qnr5g",
   "source_url": "https://omlx.ai/benchmarks/performance/zd0qnr5g",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-nvlfvd52-ctx1024",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 7.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 383.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 383.4 34.0 7.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nvlfvd52",
   "source_url": "https://omlx.ai/benchmarks/performance/nvlfvd52",
   "tg_tokens": null,
   "tps": 34.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-nvlfvd52-ctx4096",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 7.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 400.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 400.4 35.8 7.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nvlfvd52",
   "source_url": "https://omlx.ai/benchmarks/performance/nvlfvd52",
   "tg_tokens": null,
   "tps": 35.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-258dc6ib-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 981.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 981.2 85.1 14.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/258dc6ib",
   "source_url": "https://omlx.ai/benchmarks/performance/258dc6ib",
   "tg_tokens": null,
   "tps": 85.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-258dc6ib-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1109.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,109 78.6 14.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/258dc6ib",
   "source_url": "https://omlx.ai/benchmarks/performance/258dc6ib",
   "tg_tokens": null,
   "tps": 78.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-w7wv9lod-ctx131072",
   "model": "gemma-4-26B-A4B-it-uncensored-heretic",
   "notes": "VLM; oMLX community benchmark trace, single run at 128k context; peak memory 26.5 GB; TTFT 72540 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1807.0,
   "provenance": "community",
   "quant": "6Bit",
   "quote": "gemma-4-26B-A4B-it-uncensored-heretic-mlx-vlm-6Bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 128k tokens 1,807 PP tok/s 61.7 TG tok/s 72540 TTFT (ms) 26.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w7wv9lod",
   "source_url": "https://omlx.ai/benchmarks/performance/w7wv9lod",
   "tg_tokens": null,
   "tps": 61.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-4ss4yvvq-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 14.9 GB; TTFT 15087 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1086.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k tokens 1,086 PP tok/s 82.2 TG tok/s 15087 TTFT (ms) 14.9 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ss4yvvq",
   "source_url": "https://omlx.ai/benchmarks/performance/4ss4yvvq",
   "tg_tokens": null,
   "tps": 82.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-q6oi51tt-ctx1024",
   "model": "gemma-4-31B-it",
   "notes": "Mystery-Fine-Tune-HERETIC-UNCENSORED-Thinking-qx86-hi build; oMLX community benchmark trace; peak memory 28.4 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 72.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-Mystery-Fine-Tune-HERETIC-UNCENSORED-Thinking-qx86-hi-mlx on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 1k 72.2 7.9 28.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q6oi51tt",
   "source_url": "https://omlx.ai/benchmarks/performance/q6oi51tt",
   "tg_tokens": null,
   "tps": 7.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-q6oi51tt-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "Mystery-Fine-Tune-HERETIC-UNCENSORED-Thinking-qx86-hi build; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 67.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-Mystery-Fine-Tune-HERETIC-UNCENSORED-Thinking-qx86-hi-mlx on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k 67.1 7.3 29.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q6oi51tt",
   "source_url": "https://omlx.ai/benchmarks/performance/q6oi51tt",
   "tg_tokens": null,
   "tps": 7.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2z4d5208-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: 8f2mb1cr, f8uhhlry, w87kyy53, yxmo2xpk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3803.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 3,803 115.1 27.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2z4d5208",
   "source_url": "https://omlx.ai/benchmarks/performance/2z4d5208",
   "tg_tokens": null,
   "tps": 115.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2z4d5208-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: 8f2mb1cr, f8uhhlry, w87kyy53, yxmo2xpk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4720.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 4,720 114.5 28.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2z4d5208",
   "source_url": "https://omlx.ai/benchmarks/performance/2z4d5208",
   "tg_tokens": null,
   "tps": 114.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2z4d5208-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: 8f2mb1cr, f8uhhlry, w87kyy53, yxmo2xpk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4731.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 8k 4,731 114.0 28.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2z4d5208",
   "source_url": "https://omlx.ai/benchmarks/performance/2z4d5208",
   "tg_tokens": null,
   "tps": 114.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2z4d5208-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai; oMLX community benchmark trace; peak memory 28.8 GB at this context; duplicate traces of the same session: 8f2mb1cr, f8uhhlry, w87kyy53, yxmo2xpk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4451.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k 4,451 109.5 28.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2z4d5208",
   "source_url": "https://omlx.ai/benchmarks/performance/2z4d5208",
   "tg_tokens": null,
   "tps": 109.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2z4d5208-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai; oMLX community benchmark trace; peak memory 29.1 GB at this context; duplicate traces of the same session: 8f2mb1cr, f8uhhlry, w87kyy53, yxmo2xpk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3947.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 3,947 101.3 29.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2z4d5208",
   "source_url": "https://omlx.ai/benchmarks/performance/2z4d5208",
   "tg_tokens": null,
   "tps": 101.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-s6hvdqn7-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "ornith-ai repo; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: d6rtk67p, frpvynyn, h22tvhs5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2972.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "ornith-ai--Ornith-1.5-9B-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 unknown; 1k 2,972 33.7 17.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s6hvdqn7",
   "source_url": "https://omlx.ai/benchmarks/performance/s6hvdqn7",
   "tg_tokens": null,
   "tps": 33.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-s6hvdqn7-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "ornith-ai repo; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: d6rtk67p, frpvynyn, h22tvhs5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2601.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "ornith-ai--Ornith-1.5-9B-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 unknown; 4k 2,601 33.6 18.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s6hvdqn7",
   "source_url": "https://omlx.ai/benchmarks/performance/s6hvdqn7",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-s6hvdqn7-ctx8192",
   "model": "Ornith-1.5-9B",
   "notes": "ornith-ai repo; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: d6rtk67p, frpvynyn, h22tvhs5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3675.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "ornith-ai--Ornith-1.5-9B-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 unknown; 8k 3,675 33.3 18.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s6hvdqn7",
   "source_url": "https://omlx.ai/benchmarks/performance/s6hvdqn7",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-s6hvdqn7-ctx16384",
   "model": "Ornith-1.5-9B",
   "notes": "ornith-ai repo; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: d6rtk67p, frpvynyn, h22tvhs5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3576.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "ornith-ai--Ornith-1.5-9B-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 unknown; 16k 3,576 32.7 18.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s6hvdqn7",
   "source_url": "https://omlx.ai/benchmarks/performance/s6hvdqn7",
   "tg_tokens": null,
   "tps": 32.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-s6hvdqn7-ctx32768",
   "model": "Ornith-1.5-9B",
   "notes": "ornith-ai repo; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: d6rtk67p, frpvynyn, h22tvhs5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3333.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "ornith-ai--Ornith-1.5-9B-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 unknown; 32k 3,333 31.5 19.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s6hvdqn7",
   "source_url": "https://omlx.ai/benchmarks/performance/s6hvdqn7",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-hzcz99to-ctx4096",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: fvskdm51; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 193.8,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 5bit; 4k 193.8 24.2 21.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hzcz99to",
   "source_url": "https://omlx.ai/benchmarks/performance/hzcz99to",
   "tg_tokens": null,
   "tps": 24.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-hzcz99to-ctx16384",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: fvskdm51; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 91.3,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 5bit; 16k 91.3 9.1 23.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hzcz99to",
   "source_url": "https://omlx.ai/benchmarks/performance/hzcz99to",
   "tg_tokens": null,
   "tps": 9.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-24512zra-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 403.3 GB at this context; duplicate traces of the same session: 2puocypl; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 360.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 360.6 27.4 403.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/24512zra",
   "source_url": "https://omlx.ai/benchmarks/performance/24512zra",
   "tg_tokens": null,
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-24512zra-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 403.7 GB at this context; duplicate traces of the same session: 2puocypl; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 700.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 700.0 38.7 403.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/24512zra",
   "source_url": "https://omlx.ai/benchmarks/performance/24512zra",
   "tg_tokens": null,
   "tps": 38.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-24512zra-ctx65536",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 403.8 GB at this context; duplicate traces of the same session: 2puocypl; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 721.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 64k 721.0 40.2 403.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/24512zra",
   "source_url": "https://omlx.ai/benchmarks/performance/24512zra",
   "tg_tokens": null,
   "tps": 40.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-o0wgl2q0-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.7 GB at this context; duplicate traces of the same session: 2x0c20ae, fzwjewvb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 440.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 440.1 40.8 173.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o0wgl2q0",
   "source_url": "https://omlx.ai/benchmarks/performance/o0wgl2q0",
   "tg_tokens": null,
   "tps": 40.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-o0wgl2q0-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.8 GB at this context; duplicate traces of the same session: 2x0c20ae, fzwjewvb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 566.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 566.5 39.0 174.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o0wgl2q0",
   "source_url": "https://omlx.ai/benchmarks/performance/o0wgl2q0",
   "tg_tokens": null,
   "tps": 39.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-o0wgl2q0-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.1 GB at this context; duplicate traces of the same session: 2x0c20ae, fzwjewvb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 518.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 518.4 38.8 175.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o0wgl2q0",
   "source_url": "https://omlx.ai/benchmarks/performance/o0wgl2q0",
   "tg_tokens": null,
   "tps": 38.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-o0wgl2q0-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.2 GB at this context; duplicate traces of the same session: 2x0c20ae, fzwjewvb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 550.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 550.4 38.6 175.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o0wgl2q0",
   "source_url": "https://omlx.ai/benchmarks/performance/o0wgl2q0",
   "tg_tokens": null,
   "tps": 38.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-o0wgl2q0-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.3 GB at this context; duplicate traces of the same session: 2x0c20ae, fzwjewvb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 542.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 542.7 38.3 175.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o0wgl2q0",
   "source_url": "https://omlx.ai/benchmarks/performance/o0wgl2q0",
   "tg_tokens": null,
   "tps": 38.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-o0wgl2q0-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.6 GB at this context; duplicate traces of the same session: 2x0c20ae, fzwjewvb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 534.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 534.4 37.4 175.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o0wgl2q0",
   "source_url": "https://omlx.ai/benchmarks/performance/o0wgl2q0",
   "tg_tokens": null,
   "tps": 37.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-o0wgl2q0-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 176.4 GB at this context; duplicate traces of the same session: 2x0c20ae, fzwjewvb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 510.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 510.7 35.9 176.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o0wgl2q0",
   "source_url": "https://omlx.ai/benchmarks/performance/o0wgl2q0",
   "tg_tokens": null,
   "tps": 35.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-zkybddyg-ctx1024",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: bcqa7xp0; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 856.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 856.5 100.1 18.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zkybddyg",
   "source_url": "https://omlx.ai/benchmarks/performance/zkybddyg",
   "tg_tokens": null,
   "tps": 100.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-zkybddyg-ctx4096",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: bcqa7xp0; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 825.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 825.1 85.3 19.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zkybddyg",
   "source_url": "https://omlx.ai/benchmarks/performance/zkybddyg",
   "tg_tokens": null,
   "tps": 85.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-vxov0l91-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1005.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,005 69.5 35.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vxov0l91",
   "source_url": "https://omlx.ai/benchmarks/performance/vxov0l91",
   "tg_tokens": null,
   "tps": 69.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-vxov0l91-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 974.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 974.5 61.9 37.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vxov0l91",
   "source_url": "https://omlx.ai/benchmarks/performance/vxov0l91",
   "tg_tokens": null,
   "tps": 61.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-331j1375-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4526.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 4,526 111.6 28.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/331j1375",
   "source_url": "https://omlx.ai/benchmarks/performance/331j1375",
   "tg_tokens": null,
   "tps": 111.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-331j1375-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4429.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 4,429 97.7 29.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/331j1375",
   "source_url": "https://omlx.ai/benchmarks/performance/331j1375",
   "tg_tokens": null,
   "tps": 97.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-xcvumdyu-ctx1024",
   "model": "Qwen 3 30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1276.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-30B-A3B-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 1,276 108.4 16.6 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xcvumdyu",
   "source_url": "https://omlx.ai/benchmarks/performance/xcvumdyu",
   "tg_tokens": null,
   "tps": 108.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-xcvumdyu-ctx4096",
   "model": "Qwen 3 30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1331.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-30B-A3B-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 1,331 94.5 17.0 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xcvumdyu",
   "source_url": "https://omlx.ai/benchmarks/performance/xcvumdyu",
   "tg_tokens": null,
   "tps": 94.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-kjgthcp8-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: xr0mokwc; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 911.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 911.0 106.7 3.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kjgthcp8",
   "source_url": "https://omlx.ai/benchmarks/performance/kjgthcp8",
   "tg_tokens": null,
   "tps": 106.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-kjgthcp8-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.8 GB at this context; duplicate traces of the same session: xr0mokwc; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 836.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 836.1 79.8 4.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kjgthcp8",
   "source_url": "https://omlx.ai/benchmarks/performance/kjgthcp8",
   "tg_tokens": null,
   "tps": 79.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-e1luerc3-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1011.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,011 70.1 35.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e1luerc3",
   "source_url": "https://omlx.ai/benchmarks/performance/e1luerc3",
   "tg_tokens": null,
   "tps": 70.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-e1luerc3-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 966.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 966.7 61.1 37.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e1luerc3",
   "source_url": "https://omlx.ai/benchmarks/performance/e1luerc3",
   "tg_tokens": null,
   "tps": 61.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dgwwsfvr-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 29.0 GB; TTFT 7398 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4429.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k tokens 4,429 PP tok/s 99.0 TG tok/s 7398 TTFT (ms) 29.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dgwwsfvr",
   "source_url": "https://omlx.ai/benchmarks/performance/dgwwsfvr",
   "tg_tokens": null,
   "tps": 99.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-24le6iwu-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: bnt6izsq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 268.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 268.3 27.9 19.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/24le6iwu",
   "source_url": "https://omlx.ai/benchmarks/performance/24le6iwu",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-24le6iwu-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: bnt6izsq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 211.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 211.4 21.6 24.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/24le6iwu",
   "source_url": "https://omlx.ai/benchmarks/performance/24le6iwu",
   "tg_tokens": null,
   "tps": 21.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-24le6iwu-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.7 GB at this context; duplicate traces of the same session: bnt6izsq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 169.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 169.8 17.6 28.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/24le6iwu",
   "source_url": "https://omlx.ai/benchmarks/performance/24le6iwu",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-fw6fy7j7-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 431.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 431.0 13.6 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fw6fy7j7",
   "source_url": "https://omlx.ai/benchmarks/performance/fw6fy7j7",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-fw6fy7j7-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 362.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 362.8 12.2 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fw6fy7j7",
   "source_url": "https://omlx.ai/benchmarks/performance/fw6fy7j7",
   "tg_tokens": null,
   "tps": 12.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-8g81cet6-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 38.8 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 320.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 320.3 38.5 38.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8g81cet6",
   "source_url": "https://omlx.ai/benchmarks/performance/8g81cet6",
   "tg_tokens": null,
   "tps": 38.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-8g81cet6-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 44.4 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 335.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 335.2 32.8 44.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8g81cet6",
   "source_url": "https://omlx.ai/benchmarks/performance/8g81cet6",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-8g81cet6-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 44.9 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 329.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 329.9 34.2 44.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8g81cet6",
   "source_url": "https://omlx.ai/benchmarks/performance/8g81cet6",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2pj5zttg-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 210.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 210.1 22.6 16.0 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2pj5zttg",
   "source_url": "https://omlx.ai/benchmarks/performance/2pj5zttg",
   "tg_tokens": null,
   "tps": 22.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2pj5zttg-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 216.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 216.5 22.1 17.0 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2pj5zttg",
   "source_url": "https://omlx.ai/benchmarks/performance/2pj5zttg",
   "tg_tokens": null,
   "tps": 22.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-n31tppqo-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 24.7 GB; TTFT 15525 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 263.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 263.8 PP tok/s 80.6 TG tok/s 15525 TTFT (ms) 24.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n31tppqo",
   "source_url": "https://omlx.ai/benchmarks/performance/n31tppqo",
   "tg_tokens": null,
   "tps": 80.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ektr22cd-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 24.8 GB; TTFT 16274 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 251.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 251.7 PP tok/s 53.6 TG tok/s 16274 TTFT (ms) 24.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ektr22cd",
   "source_url": "https://omlx.ai/benchmarks/performance/ektr22cd",
   "tg_tokens": null,
   "tps": 53.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-mtusntao-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 25.6 GB; TTFT 14163 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 289.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 289.2 PP tok/s 80.2 TG tok/s 14163 TTFT (ms) 25.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mtusntao",
   "source_url": "https://omlx.ai/benchmarks/performance/mtusntao",
   "tg_tokens": null,
   "tps": 80.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-a6oyobrj-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "SAGE-ARCS MTP build; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: xkqlrzkf; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 123.6,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-SAGE-ARCS-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 5bit; 1k 123.6 17.1 26.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a6oyobrj",
   "source_url": "https://omlx.ai/benchmarks/performance/a6oyobrj",
   "tg_tokens": null,
   "tps": 17.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-a6oyobrj-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "SAGE-ARCS MTP build; oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: xkqlrzkf; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 122.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-SAGE-ARCS-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 5bit; 4k 122.1 13.8 30.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a6oyobrj",
   "source_url": "https://omlx.ai/benchmarks/performance/a6oyobrj",
   "tg_tokens": null,
   "tps": 13.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ok9hwsgw-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 69.5 GB at this context; duplicate traces of the same session: dgtrkp0r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1986.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,986 81.0 69.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ok9hwsgw",
   "source_url": "https://omlx.ai/benchmarks/performance/ok9hwsgw",
   "tg_tokens": null,
   "tps": 81.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ok9hwsgw-ctx8192",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: dgtrkp0r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2954.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,954 39.1 71.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ok9hwsgw",
   "source_url": "https://omlx.ai/benchmarks/performance/ok9hwsgw",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-4vcyh5v7-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.3 GB at this context; duplicate traces of the same session: v92l18si; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 435.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 1k 435.0 50.3 69.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4vcyh5v7",
   "source_url": "https://omlx.ai/benchmarks/performance/4vcyh5v7",
   "tg_tokens": null,
   "tps": 50.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-4vcyh5v7-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.9 GB at this context; duplicate traces of the same session: v92l18si; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 532.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 4k 532.9 47.0 69.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4vcyh5v7",
   "source_url": "https://omlx.ai/benchmarks/performance/4vcyh5v7",
   "tg_tokens": null,
   "tps": 47.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-4vcyh5v7-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.9 GB at this context; duplicate traces of the same session: v92l18si; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 624.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 8k 624.8 49.4 69.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4vcyh5v7",
   "source_url": "https://omlx.ai/benchmarks/performance/4vcyh5v7",
   "tg_tokens": null,
   "tps": 49.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-4vcyh5v7-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.1 GB at this context; duplicate traces of the same session: v92l18si; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 625.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 16k 625.8 47.0 70.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4vcyh5v7",
   "source_url": "https://omlx.ai/benchmarks/performance/4vcyh5v7",
   "tg_tokens": null,
   "tps": 47.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-r1baup4e-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: 7shczqjh; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 417.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 1k 417.6 43.2 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r1baup4e",
   "source_url": "https://omlx.ai/benchmarks/performance/r1baup4e",
   "tg_tokens": null,
   "tps": 43.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-r1baup4e-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: 7shczqjh; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 513.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 4k 513.2 41.6 71.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r1baup4e",
   "source_url": "https://omlx.ai/benchmarks/performance/r1baup4e",
   "tg_tokens": null,
   "tps": 41.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ilmo0jya-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.1 GB at this context; duplicate traces of the same session: 0nr9oj75, td1w1309, yxwvdyh8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1143.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,143 87.6 99.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ilmo0jya",
   "source_url": "https://omlx.ai/benchmarks/performance/ilmo0jya",
   "tg_tokens": null,
   "tps": 87.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ilmo0jya-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.7 GB at this context; duplicate traces of the same session: 0nr9oj75, td1w1309, yxwvdyh8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1423.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 1,423 78.7 99.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ilmo0jya",
   "source_url": "https://omlx.ai/benchmarks/performance/ilmo0jya",
   "tg_tokens": null,
   "tps": 78.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ilmo0jya-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.0 GB at this context; duplicate traces of the same session: 0nr9oj75, td1w1309, yxwvdyh8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1416.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 1,416 76.7 100.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ilmo0jya",
   "source_url": "https://omlx.ai/benchmarks/performance/ilmo0jya",
   "tg_tokens": null,
   "tps": 76.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ilmo0jya-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: 0nr9oj75, td1w1309, yxwvdyh8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1397.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 1,397 50.8 100.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ilmo0jya",
   "source_url": "https://omlx.ai/benchmarks/performance/ilmo0jya",
   "tg_tokens": null,
   "tps": 50.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ilmo0jya-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: 0nr9oj75, td1w1309, yxwvdyh8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1393.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 1,393 58.7 101.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ilmo0jya",
   "source_url": "https://omlx.ai/benchmarks/performance/ilmo0jya",
   "tg_tokens": null,
   "tps": 58.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ilmo0jya-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.7 GB at this context; duplicate traces of the same session: 0nr9oj75, td1w1309, yxwvdyh8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1351.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 1,351 66.1 104.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ilmo0jya",
   "source_url": "https://omlx.ai/benchmarks/performance/ilmo0jya",
   "tg_tokens": null,
   "tps": 66.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4ger7s5c-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: 27fsc2l4, 72bnn1yl, jqhag980, yed6b0bn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 919.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 919.3 53.9 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ger7s5c",
   "source_url": "https://omlx.ai/benchmarks/performance/4ger7s5c",
   "tg_tokens": null,
   "tps": 53.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4ger7s5c-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 27fsc2l4, 72bnn1yl, jqhag980, yed6b0bn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1819.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,819 76.3 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ger7s5c",
   "source_url": "https://omlx.ai/benchmarks/performance/4ger7s5c",
   "tg_tokens": null,
   "tps": 76.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4ger7s5c-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.9 GB at this context; duplicate traces of the same session: 27fsc2l4, 72bnn1yl, jqhag980, yed6b0bn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2646.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,646 70.9 74.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ger7s5c",
   "source_url": "https://omlx.ai/benchmarks/performance/4ger7s5c",
   "tg_tokens": null,
   "tps": 70.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4ger7s5c-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.4 GB at this context; duplicate traces of the same session: 27fsc2l4, 72bnn1yl, jqhag980, yed6b0bn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2547.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,547 71.1 79.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ger7s5c",
   "source_url": "https://omlx.ai/benchmarks/performance/4ger7s5c",
   "tg_tokens": null,
   "tps": 71.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4ger7s5c-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.9 GB at this context; duplicate traces of the same session: 27fsc2l4, 72bnn1yl, jqhag980, yed6b0bn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2300.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,300 56.5 80.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ger7s5c",
   "source_url": "https://omlx.ai/benchmarks/performance/4ger7s5c",
   "tg_tokens": null,
   "tps": 56.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-6p2iqclh-ctx1024",
   "model": "RAZOR-19B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 8.2 GB at this context; duplicate traces of the same session: y5u9iz6g; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 498.2,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "RAZOR-19B-A3B-oQ2e-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 2bit; 1k 498.2 45.9 8.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6p2iqclh",
   "source_url": "https://omlx.ai/benchmarks/performance/6p2iqclh",
   "tg_tokens": null,
   "tps": 45.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-6p2iqclh-ctx4096",
   "model": "RAZOR-19B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 8.8 GB at this context; duplicate traces of the same session: y5u9iz6g; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 518.3,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "RAZOR-19B-A3B-oQ2e-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 2bit; 4k 518.3 45.8 8.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6p2iqclh",
   "source_url": "https://omlx.ai/benchmarks/performance/6p2iqclh",
   "tg_tokens": null,
   "tps": 45.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-tghs39f7-ctx1024",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "it variant; oMLX community benchmark trace; peak memory 15.4 GB at this context; duplicate traces of the same session: 7tbob2la, u3zummpr, yxu8xe7g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2796.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "diffusiongemma-26B-A4B-it-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 2,796 27.7 15.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tghs39f7",
   "source_url": "https://omlx.ai/benchmarks/performance/tghs39f7",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-tghs39f7-ctx4096",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "it variant; oMLX community benchmark trace; peak memory 16.2 GB at this context; duplicate traces of the same session: 7tbob2la, u3zummpr, yxu8xe7g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2551.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "diffusiongemma-26B-A4B-it-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 2,551 20.8 16.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tghs39f7",
   "source_url": "https://omlx.ai/benchmarks/performance/tghs39f7",
   "tg_tokens": null,
   "tps": 20.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-tghs39f7-ctx8192",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "it variant; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 7tbob2la, u3zummpr, yxu8xe7g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2437.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "diffusiongemma-26B-A4B-it-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 2,437 23.5 16.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tghs39f7",
   "source_url": "https://omlx.ai/benchmarks/performance/tghs39f7",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-tghs39f7-ctx16384",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "it variant; oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 7tbob2la, u3zummpr, yxu8xe7g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2323.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "diffusiongemma-26B-A4B-it-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 2,323 47.6 17.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tghs39f7",
   "source_url": "https://omlx.ai/benchmarks/performance/tghs39f7",
   "tg_tokens": null,
   "tps": 47.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-tghs39f7-ctx32768",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "it variant; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 7tbob2la, u3zummpr, yxu8xe7g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2053.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "diffusiongemma-26B-A4B-it-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 2,053 18.8 18.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tghs39f7",
   "source_url": "https://omlx.ai/benchmarks/performance/tghs39f7",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-tghs39f7-ctx65536",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "it variant; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 7tbob2la, u3zummpr, yxu8xe7g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1621.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "diffusiongemma-26B-A4B-it-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 1,621 22.2 21.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tghs39f7",
   "source_url": "https://omlx.ai/benchmarks/performance/tghs39f7",
   "tg_tokens": null,
   "tps": 22.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-tghs39f7-ctx131072",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "it variant; oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: 7tbob2la, u3zummpr, yxu8xe7g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1097.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "diffusiongemma-26B-A4B-it-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 1,097 10.0 26.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tghs39f7",
   "source_url": "https://omlx.ai/benchmarks/performance/tghs39f7",
   "tg_tokens": null,
   "tps": 10.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-f9eyve9k-ctx1024",
   "model": "Gemma 4 E2B",
   "notes": "fp16 weights; oMLX community benchmark trace; peak memory 5.2 GB at this context; duplicate traces of the same session: xq8bo5w6; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2936.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "gemma-4-E2B-it-oQ8-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 2,936 101.7 5.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f9eyve9k",
   "source_url": "https://omlx.ai/benchmarks/performance/f9eyve9k",
   "tg_tokens": null,
   "tps": 101.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-f9eyve9k-ctx4096",
   "model": "Gemma 4 E2B",
   "notes": "fp16 weights; oMLX community benchmark trace; peak memory 5.2 GB at this context; duplicate traces of the same session: xq8bo5w6; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4988.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "gemma-4-E2B-it-oQ8-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 4,988 100.2 5.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f9eyve9k",
   "source_url": "https://omlx.ai/benchmarks/performance/f9eyve9k",
   "tg_tokens": null,
   "tps": 100.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-yxmo2xpk-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: 8f2mb1cr, f8uhhlry, w87kyy53; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3803.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 3,803 115.1 27.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yxmo2xpk",
   "source_url": "https://omlx.ai/benchmarks/performance/yxmo2xpk",
   "tg_tokens": null,
   "tps": 115.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-yxmo2xpk-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: 8f2mb1cr, f8uhhlry, w87kyy53; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4720.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 4,720 114.5 28.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yxmo2xpk",
   "source_url": "https://omlx.ai/benchmarks/performance/yxmo2xpk",
   "tg_tokens": null,
   "tps": 114.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-yxmo2xpk-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: 8f2mb1cr, f8uhhlry, w87kyy53; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4731.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 8k 4,731 114.0 28.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yxmo2xpk",
   "source_url": "https://omlx.ai/benchmarks/performance/yxmo2xpk",
   "tg_tokens": null,
   "tps": 114.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-yxmo2xpk-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai; oMLX community benchmark trace; peak memory 28.8 GB at this context; duplicate traces of the same session: 8f2mb1cr, f8uhhlry, w87kyy53; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4451.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k 4,451 109.5 28.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yxmo2xpk",
   "source_url": "https://omlx.ai/benchmarks/performance/yxmo2xpk",
   "tg_tokens": null,
   "tps": 109.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-yxmo2xpk-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai; oMLX community benchmark trace; peak memory 29.1 GB at this context; duplicate traces of the same session: 8f2mb1cr, f8uhhlry, w87kyy53; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3947.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 3,947 101.3 29.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yxmo2xpk",
   "source_url": "https://omlx.ai/benchmarks/performance/yxmo2xpk",
   "tg_tokens": null,
   "tps": 101.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-frpvynyn-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "ornith-ai repo; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: d6rtk67p, h22tvhs5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2972.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "ornith-ai--Ornith-1.5-9B-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 unknown; 1k 2,972 33.7 17.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/frpvynyn",
   "source_url": "https://omlx.ai/benchmarks/performance/frpvynyn",
   "tg_tokens": null,
   "tps": 33.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-frpvynyn-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "ornith-ai repo; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: d6rtk67p, h22tvhs5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2601.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "ornith-ai--Ornith-1.5-9B-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 unknown; 4k 2,601 33.6 18.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/frpvynyn",
   "source_url": "https://omlx.ai/benchmarks/performance/frpvynyn",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-frpvynyn-ctx8192",
   "model": "Ornith-1.5-9B",
   "notes": "ornith-ai repo; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: d6rtk67p, h22tvhs5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3675.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "ornith-ai--Ornith-1.5-9B-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 unknown; 8k 3,675 33.3 18.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/frpvynyn",
   "source_url": "https://omlx.ai/benchmarks/performance/frpvynyn",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-frpvynyn-ctx16384",
   "model": "Ornith-1.5-9B",
   "notes": "ornith-ai repo; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: d6rtk67p, h22tvhs5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3576.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "ornith-ai--Ornith-1.5-9B-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 unknown; 16k 3,576 32.7 18.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/frpvynyn",
   "source_url": "https://omlx.ai/benchmarks/performance/frpvynyn",
   "tg_tokens": null,
   "tps": 32.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-frpvynyn-ctx32768",
   "model": "Ornith-1.5-9B",
   "notes": "ornith-ai repo; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: d6rtk67p, h22tvhs5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3333.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "ornith-ai--Ornith-1.5-9B-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 unknown; 32k 3,333 31.5 19.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/frpvynyn",
   "source_url": "https://omlx.ai/benchmarks/performance/frpvynyn",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-fvskdm51-ctx4096",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 193.8,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 5bit; 4k 193.8 24.2 21.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fvskdm51",
   "source_url": "https://omlx.ai/benchmarks/performance/fvskdm51",
   "tg_tokens": null,
   "tps": 24.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-fvskdm51-ctx16384",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 91.3,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 5bit; 16k 91.3 9.1 23.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fvskdm51",
   "source_url": "https://omlx.ai/benchmarks/performance/fvskdm51",
   "tg_tokens": null,
   "tps": 9.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-6he95vwa-ctx1024",
   "model": "DeepSeek-V4-Flash-0731-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 100.9 GB at this context; duplicate traces of the same session: 3i9dra25, y1m2pnbu; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 469.1,
   "provenance": "community",
   "quant": "Mixed 2bit-3bit g64",
   "quote": "DeepSeek-V4-Flash-0731-Abliterated-MLX-Mixed-2bit-3bit-g64 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 1k 469.1 28.0 100.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6he95vwa",
   "source_url": "https://omlx.ai/benchmarks/performance/6he95vwa",
   "tg_tokens": null,
   "tps": 28.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-6he95vwa-ctx4096",
   "model": "DeepSeek-V4-Flash-0731-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: 3i9dra25, y1m2pnbu; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 476.8,
   "provenance": "community",
   "quant": "Mixed 2bit-3bit g64",
   "quote": "DeepSeek-V4-Flash-0731-Abliterated-MLX-Mixed-2bit-3bit-g64 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 4k 476.8 25.8 101.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6he95vwa",
   "source_url": "https://omlx.ai/benchmarks/performance/6he95vwa",
   "tg_tokens": null,
   "tps": 25.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-6he95vwa-ctx8192",
   "model": "DeepSeek-V4-Flash-0731-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: 3i9dra25, y1m2pnbu; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 470.7,
   "provenance": "community",
   "quant": "Mixed 2bit-3bit g64",
   "quote": "DeepSeek-V4-Flash-0731-Abliterated-MLX-Mixed-2bit-3bit-g64 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 8k 470.7 25.1 101.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6he95vwa",
   "source_url": "https://omlx.ai/benchmarks/performance/6he95vwa",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-6he95vwa-ctx16384",
   "model": "DeepSeek-V4-Flash-0731-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 101.4 GB at this context; duplicate traces of the same session: 3i9dra25, y1m2pnbu; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 462.1,
   "provenance": "community",
   "quant": "Mixed 2bit-3bit g64",
   "quote": "DeepSeek-V4-Flash-0731-Abliterated-MLX-Mixed-2bit-3bit-g64 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 16k 462.1 24.3 101.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6he95vwa",
   "source_url": "https://omlx.ai/benchmarks/performance/6he95vwa",
   "tg_tokens": null,
   "tps": 24.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-6he95vwa-ctx32768",
   "model": "DeepSeek-V4-Flash-0731-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: 3i9dra25, y1m2pnbu; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 448.6,
   "provenance": "community",
   "quant": "Mixed 2bit-3bit g64",
   "quote": "DeepSeek-V4-Flash-0731-Abliterated-MLX-Mixed-2bit-3bit-g64 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 32k 448.6 23.0 101.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6he95vwa",
   "source_url": "https://omlx.ai/benchmarks/performance/6he95vwa",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-6he95vwa-ctx65536",
   "model": "DeepSeek-V4-Flash-0731-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 101.8 GB at this context; duplicate traces of the same session: 3i9dra25, y1m2pnbu; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 424.9,
   "provenance": "community",
   "quant": "Mixed 2bit-3bit g64",
   "quote": "DeepSeek-V4-Flash-0731-Abliterated-MLX-Mixed-2bit-3bit-g64 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 64k 424.9 22.8 101.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6he95vwa",
   "source_url": "https://omlx.ai/benchmarks/performance/6he95vwa",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2puocypl-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 403.3 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 360.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 360.6 27.4 403.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2puocypl",
   "source_url": "https://omlx.ai/benchmarks/performance/2puocypl",
   "tg_tokens": null,
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2puocypl-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 403.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 700.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 700.0 38.7 403.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2puocypl",
   "source_url": "https://omlx.ai/benchmarks/performance/2puocypl",
   "tg_tokens": null,
   "tps": 38.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2puocypl-ctx65536",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 403.8 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 721.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 64k 721.0 40.2 403.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2puocypl",
   "source_url": "https://omlx.ai/benchmarks/performance/2puocypl",
   "tg_tokens": null,
   "tps": 40.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2x0c20ae-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.7 GB at this context; duplicate traces of the same session: fzwjewvb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 440.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 440.1 40.8 173.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2x0c20ae",
   "source_url": "https://omlx.ai/benchmarks/performance/2x0c20ae",
   "tg_tokens": null,
   "tps": 40.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2x0c20ae-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.8 GB at this context; duplicate traces of the same session: fzwjewvb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 566.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 566.5 39.0 174.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2x0c20ae",
   "source_url": "https://omlx.ai/benchmarks/performance/2x0c20ae",
   "tg_tokens": null,
   "tps": 39.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2x0c20ae-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.1 GB at this context; duplicate traces of the same session: fzwjewvb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 518.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 518.4 38.8 175.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2x0c20ae",
   "source_url": "https://omlx.ai/benchmarks/performance/2x0c20ae",
   "tg_tokens": null,
   "tps": 38.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2x0c20ae-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.2 GB at this context; duplicate traces of the same session: fzwjewvb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 550.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 550.4 38.6 175.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2x0c20ae",
   "source_url": "https://omlx.ai/benchmarks/performance/2x0c20ae",
   "tg_tokens": null,
   "tps": 38.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2x0c20ae-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.3 GB at this context; duplicate traces of the same session: fzwjewvb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 542.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 542.7 38.3 175.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2x0c20ae",
   "source_url": "https://omlx.ai/benchmarks/performance/2x0c20ae",
   "tg_tokens": null,
   "tps": 38.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2x0c20ae-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.6 GB at this context; duplicate traces of the same session: fzwjewvb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 534.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 534.4 37.4 175.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2x0c20ae",
   "source_url": "https://omlx.ai/benchmarks/performance/2x0c20ae",
   "tg_tokens": null,
   "tps": 37.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2x0c20ae-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 176.4 GB at this context; duplicate traces of the same session: fzwjewvb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 510.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 510.7 35.9 176.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2x0c20ae",
   "source_url": "https://omlx.ai/benchmarks/performance/2x0c20ae",
   "tg_tokens": null,
   "tps": 35.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ld7pw1gl-ctx4096",
   "model": "Gemma 4 E2B",
   "notes": "Jundot build; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 4.3 GB; TTFT 769 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5325.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--gemma-4-E2B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 5,325 PP tok/s 93.4 TG tok/s 769 TTFT (ms) 4.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ld7pw1gl",
   "source_url": "https://omlx.ai/benchmarks/performance/ld7pw1gl",
   "tg_tokens": null,
   "tps": 93.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-3bitgla8-ctx4096",
   "model": "Gemma 4 E2B",
   "notes": "Jundot build; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 4.3 GB; TTFT 510 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 8029.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--gemma-4-E2B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 8,029 PP tok/s 90.2 TG tok/s 510 TTFT (ms) 4.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3bitgla8",
   "source_url": "https://omlx.ai/benchmarks/performance/3bitgla8",
   "tg_tokens": null,
   "tps": 90.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-x7en24d2-ctx1024",
   "model": "MiMo-7B-RL",
   "notes": "oMLX community benchmark trace; peak memory 4.7 GB at this context; duplicate traces of the same session: 3r118bef, 5srv21v4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1289.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-7B-RL-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 1,289 120.3 4.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x7en24d2",
   "source_url": "https://omlx.ai/benchmarks/performance/x7en24d2",
   "tg_tokens": null,
   "tps": 120.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-x7en24d2-ctx4096",
   "model": "MiMo-7B-RL",
   "notes": "oMLX community benchmark trace; peak memory 5.2 GB at this context; duplicate traces of the same session: 3r118bef, 5srv21v4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1247.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-7B-RL-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 1,247 108.0 5.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x7en24d2",
   "source_url": "https://omlx.ai/benchmarks/performance/x7en24d2",
   "tg_tokens": null,
   "tps": 108.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-x7en24d2-ctx8192",
   "model": "MiMo-7B-RL",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: 3r118bef, 5srv21v4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1150.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-7B-RL-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 1,150 94.2 5.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x7en24d2",
   "source_url": "https://omlx.ai/benchmarks/performance/x7en24d2",
   "tg_tokens": null,
   "tps": 94.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-bcqa7xp0-ctx1024",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 856.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 856.5 100.1 18.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bcqa7xp0",
   "source_url": "https://omlx.ai/benchmarks/performance/bcqa7xp0",
   "tg_tokens": null,
   "tps": 100.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-bcqa7xp0-ctx4096",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 825.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 825.1 85.3 19.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bcqa7xp0",
   "source_url": "https://omlx.ai/benchmarks/performance/bcqa7xp0",
   "tg_tokens": null,
   "tps": 85.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-grunip58-ctx16384",
   "model": "Qwen3 8B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 7.1 GB; TTFT 28561 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 573.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-8B-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k tokens 573.6 PP tok/s 44.2 TG tok/s 28561 TTFT (ms) 7.1 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/grunip58",
   "source_url": "https://omlx.ai/benchmarks/performance/grunip58",
   "tg_tokens": null,
   "tps": 44.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-xr0mokwc-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 911.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 911.0 106.7 3.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xr0mokwc",
   "source_url": "https://omlx.ai/benchmarks/performance/xr0mokwc",
   "tg_tokens": null,
   "tps": 106.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-xr0mokwc-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.8 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 836.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 836.1 79.8 4.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xr0mokwc",
   "source_url": "https://omlx.ai/benchmarks/performance/xr0mokwc",
   "tg_tokens": null,
   "tps": 79.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-7lxqj7dv-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: 2hl6jmsh, z52faqck; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 392.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 1k 392.7 59.9 5.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7lxqj7dv",
   "source_url": "https://omlx.ai/benchmarks/performance/7lxqj7dv",
   "tg_tokens": null,
   "tps": 59.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-7lxqj7dv-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: 2hl6jmsh, z52faqck; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 341.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 4k 341.0 39.6 6.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7lxqj7dv",
   "source_url": "https://omlx.ai/benchmarks/performance/7lxqj7dv",
   "tg_tokens": null,
   "tps": 39.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-7lxqj7dv-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.6 GB at this context; duplicate traces of the same session: 2hl6jmsh, z52faqck; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 328.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 8k 328.0 39.1 6.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7lxqj7dv",
   "source_url": "https://omlx.ai/benchmarks/performance/7lxqj7dv",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-u8t25pev-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: lcxkb850, yt8g68ra; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 398.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 1k 398.0 59.7 5.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u8t25pev",
   "source_url": "https://omlx.ai/benchmarks/performance/u8t25pev",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-u8t25pev-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: lcxkb850, yt8g68ra; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 366.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 4k 366.9 45.6 6.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u8t25pev",
   "source_url": "https://omlx.ai/benchmarks/performance/u8t25pev",
   "tg_tokens": null,
   "tps": 45.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-u8t25pev-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.6 GB at this context; duplicate traces of the same session: lcxkb850, yt8g68ra; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 346.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 8k 346.0 41.6 6.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u8t25pev",
   "source_url": "https://omlx.ai/benchmarks/performance/u8t25pev",
   "tg_tokens": null,
   "tps": 41.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2nrl7suk-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 0zvbb5wx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 257.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 257.5 29.3 16.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2nrl7suk",
   "source_url": "https://omlx.ai/benchmarks/performance/2nrl7suk",
   "tg_tokens": null,
   "tps": 29.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2nrl7suk-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 0zvbb5wx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 270.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 270.0 28.5 19.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2nrl7suk",
   "source_url": "https://omlx.ai/benchmarks/performance/2nrl7suk",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-bnt6izsq-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: 5sbbc3sa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 268.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 268.3 27.9 19.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bnt6izsq",
   "source_url": "https://omlx.ai/benchmarks/performance/bnt6izsq",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-bnt6izsq-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: 5sbbc3sa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 211.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 211.4 21.6 24.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bnt6izsq",
   "source_url": "https://omlx.ai/benchmarks/performance/bnt6izsq",
   "tg_tokens": null,
   "tps": 21.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-bnt6izsq-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.7 GB at this context; duplicate traces of the same session: 5sbbc3sa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 169.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 169.8 17.6 28.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bnt6izsq",
   "source_url": "https://omlx.ai/benchmarks/performance/bnt6izsq",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ujr55g2p-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: b59w22cx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 821.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 8bit; 1k 821.8 18.5 28.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ujr55g2p",
   "source_url": "https://omlx.ai/benchmarks/performance/ujr55g2p",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ujr55g2p-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: b59w22cx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 954.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 8bit; 4k 954.9 18.3 29.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ujr55g2p",
   "source_url": "https://omlx.ai/benchmarks/performance/ujr55g2p",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-gv7i2gmf-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 22.1 GB; TTFT 17258 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 237.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 237.3 PP tok/s 76.3 TG tok/s 17258 TTFT (ms) 22.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gv7i2gmf",
   "source_url": "https://omlx.ai/benchmarks/performance/gv7i2gmf",
   "tg_tokens": null,
   "tps": 76.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-aeca11wc-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 25.3 GB; TTFT 16075 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 254.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 254.8 PP tok/s 75.4 TG tok/s 16075 TTFT (ms) 25.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aeca11wc",
   "source_url": "https://omlx.ai/benchmarks/performance/aeca11wc",
   "tg_tokens": null,
   "tps": 75.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-xnmbgrns-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 25.3 GB; TTFT 12999 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 315.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 315.1 PP tok/s 76.6 TG tok/s 12999 TTFT (ms) 25.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xnmbgrns",
   "source_url": "https://omlx.ai/benchmarks/performance/xnmbgrns",
   "tg_tokens": null,
   "tps": 76.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-xkqlrzkf-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "SAGE-ARCS MTP build; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 123.6,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-SAGE-ARCS-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 5bit; 1k 123.6 17.1 26.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xkqlrzkf",
   "source_url": "https://omlx.ai/benchmarks/performance/xkqlrzkf",
   "tg_tokens": null,
   "tps": 17.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-xkqlrzkf-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "SAGE-ARCS MTP build; oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 122.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-SAGE-ARCS-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 5bit; 4k 122.1 13.8 30.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xkqlrzkf",
   "source_url": "https://omlx.ai/benchmarks/performance/xkqlrzkf",
   "tg_tokens": null,
   "tps": 13.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dgtrkp0r-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 69.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1986.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,986 81.0 69.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dgtrkp0r",
   "source_url": "https://omlx.ai/benchmarks/performance/dgtrkp0r",
   "tg_tokens": null,
   "tps": 81.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dgtrkp0r-ctx8192",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2954.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,954 39.1 71.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dgtrkp0r",
   "source_url": "https://omlx.ai/benchmarks/performance/dgtrkp0r",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-7shczqjh-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 417.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 1k 417.6 43.2 70.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7shczqjh",
   "source_url": "https://omlx.ai/benchmarks/performance/7shczqjh",
   "tg_tokens": null,
   "tps": 43.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-7shczqjh-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 513.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 4k 513.2 41.6 71.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7shczqjh",
   "source_url": "https://omlx.ai/benchmarks/performance/7shczqjh",
   "tg_tokens": null,
   "tps": 41.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-v92l18si-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.3 GB at this context; duplicate traces of the same session: i8cpft67, lbov168k; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 435.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 1k 435.0 50.3 69.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v92l18si",
   "source_url": "https://omlx.ai/benchmarks/performance/v92l18si",
   "tg_tokens": null,
   "tps": 50.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-v92l18si-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.9 GB at this context; duplicate traces of the same session: i8cpft67, lbov168k; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 532.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 4k 532.9 47.0 69.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v92l18si",
   "source_url": "https://omlx.ai/benchmarks/performance/v92l18si",
   "tg_tokens": null,
   "tps": 47.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-v92l18si-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.9 GB at this context; duplicate traces of the same session: i8cpft67, lbov168k; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 624.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 8k 624.8 49.4 69.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v92l18si",
   "source_url": "https://omlx.ai/benchmarks/performance/v92l18si",
   "tg_tokens": null,
   "tps": 49.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-v92l18si-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.1 GB at this context; duplicate traces of the same session: i8cpft67, lbov168k; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 625.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 16k 625.8 47.0 70.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v92l18si",
   "source_url": "https://omlx.ai/benchmarks/performance/v92l18si",
   "tg_tokens": null,
   "tps": 47.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-xuoyko5v-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.3 GB at this context; duplicate traces of the same session: fthdh1wz, vor87sya; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 362.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 1k 362.6 44.0 69.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xuoyko5v",
   "source_url": "https://omlx.ai/benchmarks/performance/xuoyko5v",
   "tg_tokens": null,
   "tps": 44.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-xuoyko5v-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.9 GB at this context; duplicate traces of the same session: fthdh1wz, vor87sya; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 628.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 8k 628.6 45.9 69.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xuoyko5v",
   "source_url": "https://omlx.ai/benchmarks/performance/xuoyko5v",
   "tg_tokens": null,
   "tps": 45.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-xuoyko5v-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.1 GB at this context; duplicate traces of the same session: fthdh1wz, vor87sya; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 630.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 16k 630.2 48.2 70.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xuoyko5v",
   "source_url": "https://omlx.ai/benchmarks/performance/xuoyko5v",
   "tg_tokens": null,
   "tps": 48.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-td1w1309-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.1 GB at this context; duplicate traces of the same session: 0nr9oj75, yxwvdyh8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1143.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,143 87.6 99.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/td1w1309",
   "source_url": "https://omlx.ai/benchmarks/performance/td1w1309",
   "tg_tokens": null,
   "tps": 87.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-td1w1309-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.7 GB at this context; duplicate traces of the same session: 0nr9oj75, yxwvdyh8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1423.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 1,423 78.7 99.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/td1w1309",
   "source_url": "https://omlx.ai/benchmarks/performance/td1w1309",
   "tg_tokens": null,
   "tps": 78.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-td1w1309-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.0 GB at this context; duplicate traces of the same session: 0nr9oj75, yxwvdyh8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1416.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 1,416 76.7 100.0 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/td1w1309",
   "source_url": "https://omlx.ai/benchmarks/performance/td1w1309",
   "tg_tokens": null,
   "tps": 76.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-td1w1309-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: 0nr9oj75, yxwvdyh8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1397.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 1,397 50.8 100.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/td1w1309",
   "source_url": "https://omlx.ai/benchmarks/performance/td1w1309",
   "tg_tokens": null,
   "tps": 50.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-td1w1309-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: 0nr9oj75, yxwvdyh8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1393.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 1,393 58.7 101.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/td1w1309",
   "source_url": "https://omlx.ai/benchmarks/performance/td1w1309",
   "tg_tokens": null,
   "tps": 58.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-td1w1309-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.7 GB at this context; duplicate traces of the same session: 0nr9oj75, yxwvdyh8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1351.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 1,351 66.1 104.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/td1w1309",
   "source_url": "https://omlx.ai/benchmarks/performance/td1w1309",
   "tg_tokens": null,
   "tps": 66.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-72bnn1yl-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: 27fsc2l4, jqhag980, yed6b0bn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 919.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 919.3 53.9 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/72bnn1yl",
   "source_url": "https://omlx.ai/benchmarks/performance/72bnn1yl",
   "tg_tokens": null,
   "tps": 53.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-72bnn1yl-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 27fsc2l4, jqhag980, yed6b0bn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1819.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,819 76.3 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/72bnn1yl",
   "source_url": "https://omlx.ai/benchmarks/performance/72bnn1yl",
   "tg_tokens": null,
   "tps": 76.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-72bnn1yl-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.9 GB at this context; duplicate traces of the same session: 27fsc2l4, jqhag980, yed6b0bn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2646.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,646 70.9 74.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/72bnn1yl",
   "source_url": "https://omlx.ai/benchmarks/performance/72bnn1yl",
   "tg_tokens": null,
   "tps": 70.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-72bnn1yl-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.4 GB at this context; duplicate traces of the same session: 27fsc2l4, jqhag980, yed6b0bn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2547.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,547 71.1 79.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/72bnn1yl",
   "source_url": "https://omlx.ai/benchmarks/performance/72bnn1yl",
   "tg_tokens": null,
   "tps": 71.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-72bnn1yl-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.9 GB at this context; duplicate traces of the same session: 27fsc2l4, jqhag980, yed6b0bn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2300.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,300 56.5 80.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/72bnn1yl",
   "source_url": "https://omlx.ai/benchmarks/performance/72bnn1yl",
   "tg_tokens": null,
   "tps": 56.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-y5u9iz6g-ctx1024",
   "model": "RAZOR-19B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 8.2 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 498.2,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "RAZOR-19B-A3B-oQ2e-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 2bit; 1k 498.2 45.9 8.2 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y5u9iz6g",
   "source_url": "https://omlx.ai/benchmarks/performance/y5u9iz6g",
   "tg_tokens": null,
   "tps": 45.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-y5u9iz6g-ctx4096",
   "model": "RAZOR-19B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 8.8 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 518.3,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "RAZOR-19B-A3B-oQ2e-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 2bit; 4k 518.3 45.8 8.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y5u9iz6g",
   "source_url": "https://omlx.ai/benchmarks/performance/y5u9iz6g",
   "tg_tokens": null,
   "tps": 45.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-u3zummpr-ctx1024",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "it variant; oMLX community benchmark trace; peak memory 15.4 GB at this context; duplicate traces of the same session: 7tbob2la, yxu8xe7g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2796.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "diffusiongemma-26B-A4B-it-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 2,796 27.7 15.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u3zummpr",
   "source_url": "https://omlx.ai/benchmarks/performance/u3zummpr",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-u3zummpr-ctx4096",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "it variant; oMLX community benchmark trace; peak memory 16.2 GB at this context; duplicate traces of the same session: 7tbob2la, yxu8xe7g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2551.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "diffusiongemma-26B-A4B-it-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 2,551 20.8 16.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u3zummpr",
   "source_url": "https://omlx.ai/benchmarks/performance/u3zummpr",
   "tg_tokens": null,
   "tps": 20.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-u3zummpr-ctx8192",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "it variant; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 7tbob2la, yxu8xe7g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2437.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "diffusiongemma-26B-A4B-it-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 2,437 23.5 16.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u3zummpr",
   "source_url": "https://omlx.ai/benchmarks/performance/u3zummpr",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-u3zummpr-ctx16384",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "it variant; oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 7tbob2la, yxu8xe7g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2323.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "diffusiongemma-26B-A4B-it-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 2,323 47.6 17.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u3zummpr",
   "source_url": "https://omlx.ai/benchmarks/performance/u3zummpr",
   "tg_tokens": null,
   "tps": 47.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-u3zummpr-ctx32768",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "it variant; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 7tbob2la, yxu8xe7g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2053.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "diffusiongemma-26B-A4B-it-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 2,053 18.8 18.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u3zummpr",
   "source_url": "https://omlx.ai/benchmarks/performance/u3zummpr",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-u3zummpr-ctx65536",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "it variant; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 7tbob2la, yxu8xe7g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1621.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "diffusiongemma-26B-A4B-it-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 1,621 22.2 21.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u3zummpr",
   "source_url": "https://omlx.ai/benchmarks/performance/u3zummpr",
   "tg_tokens": null,
   "tps": 22.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-u3zummpr-ctx131072",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "it variant; oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: 7tbob2la, yxu8xe7g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1097.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "diffusiongemma-26B-A4B-it-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 1,097 10.0 26.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u3zummpr",
   "source_url": "https://omlx.ai/benchmarks/performance/u3zummpr",
   "tg_tokens": null,
   "tps": 10.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-z5qh2l0f-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: pxos6zsf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 766.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 766.7 62.3 14.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z5qh2l0f",
   "source_url": "https://omlx.ai/benchmarks/performance/z5qh2l0f",
   "tg_tokens": null,
   "tps": 62.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-z5qh2l0f-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: pxos6zsf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 719.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 719.9 63.5 14.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z5qh2l0f",
   "source_url": "https://omlx.ai/benchmarks/performance/z5qh2l0f",
   "tg_tokens": null,
   "tps": 63.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-xq8bo5w6-ctx1024",
   "model": "Gemma 4 E2B",
   "notes": "fp16 weights; oMLX community benchmark trace; peak memory 5.2 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2936.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "gemma-4-E2B-it-oQ8-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 2,936 101.7 5.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xq8bo5w6",
   "source_url": "https://omlx.ai/benchmarks/performance/xq8bo5w6",
   "tg_tokens": null,
   "tps": 101.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-xq8bo5w6-ctx4096",
   "model": "Gemma 4 E2B",
   "notes": "fp16 weights; oMLX community benchmark trace; peak memory 5.2 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4988.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "gemma-4-E2B-it-oQ8-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 4,988 100.2 5.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xq8bo5w6",
   "source_url": "https://omlx.ai/benchmarks/performance/xq8bo5w6",
   "tg_tokens": null,
   "tps": 100.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8f2mb1cr-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: f8uhhlry, w87kyy53; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3803.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 3,803 115.1 27.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8f2mb1cr",
   "source_url": "https://omlx.ai/benchmarks/performance/8f2mb1cr",
   "tg_tokens": null,
   "tps": 115.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8f2mb1cr-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: f8uhhlry, w87kyy53; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4720.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 4,720 114.5 28.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8f2mb1cr",
   "source_url": "https://omlx.ai/benchmarks/performance/8f2mb1cr",
   "tg_tokens": null,
   "tps": 114.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8f2mb1cr-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: f8uhhlry, w87kyy53; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4731.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 8k 4,731 114.0 28.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8f2mb1cr",
   "source_url": "https://omlx.ai/benchmarks/performance/8f2mb1cr",
   "tg_tokens": null,
   "tps": 114.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8f2mb1cr-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai; oMLX community benchmark trace; peak memory 28.8 GB at this context; duplicate traces of the same session: f8uhhlry, w87kyy53; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4451.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k 4,451 109.5 28.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8f2mb1cr",
   "source_url": "https://omlx.ai/benchmarks/performance/8f2mb1cr",
   "tg_tokens": null,
   "tps": 109.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8f2mb1cr-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai; oMLX community benchmark trace; peak memory 29.1 GB at this context; duplicate traces of the same session: f8uhhlry, w87kyy53; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3947.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 3,947 101.3 29.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8f2mb1cr",
   "source_url": "https://omlx.ai/benchmarks/performance/8f2mb1cr",
   "tg_tokens": null,
   "tps": 101.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-d6rtk67p-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "ornith-ai repo; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: h22tvhs5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2972.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "ornith-ai--Ornith-1.5-9B-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 unknown; 1k 2,972 33.7 17.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d6rtk67p",
   "source_url": "https://omlx.ai/benchmarks/performance/d6rtk67p",
   "tg_tokens": null,
   "tps": 33.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-d6rtk67p-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "ornith-ai repo; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: h22tvhs5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2601.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "ornith-ai--Ornith-1.5-9B-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 unknown; 4k 2,601 33.6 18.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d6rtk67p",
   "source_url": "https://omlx.ai/benchmarks/performance/d6rtk67p",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-d6rtk67p-ctx8192",
   "model": "Ornith-1.5-9B",
   "notes": "ornith-ai repo; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: h22tvhs5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3675.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "ornith-ai--Ornith-1.5-9B-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 unknown; 8k 3,675 33.3 18.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d6rtk67p",
   "source_url": "https://omlx.ai/benchmarks/performance/d6rtk67p",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-d6rtk67p-ctx16384",
   "model": "Ornith-1.5-9B",
   "notes": "ornith-ai repo; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: h22tvhs5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3576.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "ornith-ai--Ornith-1.5-9B-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 unknown; 16k 3,576 32.7 18.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d6rtk67p",
   "source_url": "https://omlx.ai/benchmarks/performance/d6rtk67p",
   "tg_tokens": null,
   "tps": 32.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-d6rtk67p-ctx32768",
   "model": "Ornith-1.5-9B",
   "notes": "ornith-ai repo; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: h22tvhs5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3333.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "ornith-ai--Ornith-1.5-9B-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 unknown; 32k 3,333 31.5 19.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d6rtk67p",
   "source_url": "https://omlx.ai/benchmarks/performance/d6rtk67p",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-5mjh7lcm-ctx1024",
   "model": "6-bit",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 6gxrtp1y; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 437.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "6-bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 1k 437.5 12.8 22.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5mjh7lcm",
   "source_url": "https://omlx.ai/benchmarks/performance/5mjh7lcm",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-5mjh7lcm-ctx4096",
   "model": "6-bit",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 6gxrtp1y; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 459.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "6-bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 4k 459.3 12.6 23.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5mjh7lcm",
   "source_url": "https://omlx.ai/benchmarks/performance/5mjh7lcm",
   "tg_tokens": null,
   "tps": 12.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-egt5q661-ctx1024",
   "model": "Agents-A1",
   "notes": "REAP-19B; oMLX community benchmark trace, single run at 1k context; peak memory 12.9 GB; TTFT 1231 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 832.2,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Agents-A1-OptiQ-4bit-REAP-19B on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k tokens 832.2 PP tok/s 65.8 TG tok/s 1231 TTFT (ms) 12.9 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/egt5q661",
   "source_url": "https://omlx.ai/benchmarks/performance/egt5q661",
   "tg_tokens": null,
   "tps": 65.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3i9dra25-ctx1024",
   "model": "DeepSeek-V4-Flash-0731-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 100.9 GB at this context; duplicate traces of the same session: 0fk3zwwr, 251ru7y9, uogwy2cq, y1m2pnbu; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 469.1,
   "provenance": "community",
   "quant": "Mixed 2bit-3bit g64",
   "quote": "DeepSeek-V4-Flash-0731-Abliterated-MLX-Mixed-2bit-3bit-g64 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 1k 469.1 28.0 100.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3i9dra25",
   "source_url": "https://omlx.ai/benchmarks/performance/3i9dra25",
   "tg_tokens": null,
   "tps": 28.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3i9dra25-ctx4096",
   "model": "DeepSeek-V4-Flash-0731-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: 0fk3zwwr, 251ru7y9, uogwy2cq, y1m2pnbu; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 476.8,
   "provenance": "community",
   "quant": "Mixed 2bit-3bit g64",
   "quote": "DeepSeek-V4-Flash-0731-Abliterated-MLX-Mixed-2bit-3bit-g64 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 4k 476.8 25.8 101.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3i9dra25",
   "source_url": "https://omlx.ai/benchmarks/performance/3i9dra25",
   "tg_tokens": null,
   "tps": 25.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3i9dra25-ctx8192",
   "model": "DeepSeek-V4-Flash-0731-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: 0fk3zwwr, 251ru7y9, uogwy2cq, y1m2pnbu; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 470.7,
   "provenance": "community",
   "quant": "Mixed 2bit-3bit g64",
   "quote": "DeepSeek-V4-Flash-0731-Abliterated-MLX-Mixed-2bit-3bit-g64 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 8k 470.7 25.1 101.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3i9dra25",
   "source_url": "https://omlx.ai/benchmarks/performance/3i9dra25",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3i9dra25-ctx16384",
   "model": "DeepSeek-V4-Flash-0731-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 101.4 GB at this context; duplicate traces of the same session: 0fk3zwwr, 251ru7y9, uogwy2cq, y1m2pnbu; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 462.1,
   "provenance": "community",
   "quant": "Mixed 2bit-3bit g64",
   "quote": "DeepSeek-V4-Flash-0731-Abliterated-MLX-Mixed-2bit-3bit-g64 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 16k 462.1 24.3 101.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3i9dra25",
   "source_url": "https://omlx.ai/benchmarks/performance/3i9dra25",
   "tg_tokens": null,
   "tps": 24.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3i9dra25-ctx32768",
   "model": "DeepSeek-V4-Flash-0731-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: 0fk3zwwr, 251ru7y9, uogwy2cq, y1m2pnbu; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 448.6,
   "provenance": "community",
   "quant": "Mixed 2bit-3bit g64",
   "quote": "DeepSeek-V4-Flash-0731-Abliterated-MLX-Mixed-2bit-3bit-g64 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 32k 448.6 23.0 101.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3i9dra25",
   "source_url": "https://omlx.ai/benchmarks/performance/3i9dra25",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3i9dra25-ctx65536",
   "model": "DeepSeek-V4-Flash-0731-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 101.8 GB at this context; duplicate traces of the same session: 0fk3zwwr, 251ru7y9, uogwy2cq, y1m2pnbu; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 424.9,
   "provenance": "community",
   "quant": "Mixed 2bit-3bit g64",
   "quote": "DeepSeek-V4-Flash-0731-Abliterated-MLX-Mixed-2bit-3bit-g64 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 64k 424.9 22.8 101.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3i9dra25",
   "source_url": "https://omlx.ai/benchmarks/performance/3i9dra25",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-fzwjewvb-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 440.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 440.1 40.8 173.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fzwjewvb",
   "source_url": "https://omlx.ai/benchmarks/performance/fzwjewvb",
   "tg_tokens": null,
   "tps": 40.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-fzwjewvb-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.8 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 566.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 566.5 39.0 174.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fzwjewvb",
   "source_url": "https://omlx.ai/benchmarks/performance/fzwjewvb",
   "tg_tokens": null,
   "tps": 39.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-fzwjewvb-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 518.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 518.4 38.8 175.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fzwjewvb",
   "source_url": "https://omlx.ai/benchmarks/performance/fzwjewvb",
   "tg_tokens": null,
   "tps": 38.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-fzwjewvb-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.2 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 550.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 550.4 38.6 175.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fzwjewvb",
   "source_url": "https://omlx.ai/benchmarks/performance/fzwjewvb",
   "tg_tokens": null,
   "tps": 38.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-fzwjewvb-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.3 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 542.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 542.7 38.3 175.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fzwjewvb",
   "source_url": "https://omlx.ai/benchmarks/performance/fzwjewvb",
   "tg_tokens": null,
   "tps": 38.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-fzwjewvb-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.6 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 534.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 534.4 37.4 175.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fzwjewvb",
   "source_url": "https://omlx.ai/benchmarks/performance/fzwjewvb",
   "tg_tokens": null,
   "tps": 37.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-fzwjewvb-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 176.4 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 510.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 510.7 35.9 176.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fzwjewvb",
   "source_url": "https://omlx.ai/benchmarks/performance/fzwjewvb",
   "tg_tokens": null,
   "tps": 35.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-u9hd0q3x-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "Jundot build; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 21.8 GB; TTFT 1783 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2298.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--gemma-4-E4B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 2,298 PP tok/s 25.9 TG tok/s 1783 TTFT (ms) 21.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u9hd0q3x",
   "source_url": "https://omlx.ai/benchmarks/performance/u9hd0q3x",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-r3zn923r-ctx1024",
   "model": "Ling-3.0-tiny",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 9.0 GB; TTFT 505 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2028.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ling-3.0-tiny-8bit-XL-mlx on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 8bit; 1k tokens 2,028 PP tok/s 103.6 TG tok/s 505 TTFT (ms) 9.0 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r3zn923r",
   "source_url": "https://omlx.ai/benchmarks/performance/r3zn923r",
   "tg_tokens": null,
   "tps": 103.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3r118bef-ctx1024",
   "model": "MiMo-7B-RL",
   "notes": "oMLX community benchmark trace; peak memory 4.7 GB at this context; duplicate traces of the same session: 5srv21v4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1289.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-7B-RL-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 1,289 120.3 4.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3r118bef",
   "source_url": "https://omlx.ai/benchmarks/performance/3r118bef",
   "tg_tokens": null,
   "tps": 120.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3r118bef-ctx4096",
   "model": "MiMo-7B-RL",
   "notes": "oMLX community benchmark trace; peak memory 5.2 GB at this context; duplicate traces of the same session: 5srv21v4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1247.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-7B-RL-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 1,247 108.0 5.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3r118bef",
   "source_url": "https://omlx.ai/benchmarks/performance/3r118bef",
   "tg_tokens": null,
   "tps": 108.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3r118bef-ctx8192",
   "model": "MiMo-7B-RL",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: 5srv21v4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1150.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-7B-RL-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 1,150 94.2 5.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3r118bef",
   "source_url": "https://omlx.ai/benchmarks/performance/3r118bef",
   "tg_tokens": null,
   "tps": 94.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-a5z3irwy-ctx16384",
   "model": "Nemotron-3.5-Lightning-30B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: p2xjxwad; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1456.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Nemotron-3.5-Lightning-30B-A3B-8bit-MTP on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 1,456 97.3 34.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a5z3irwy",
   "source_url": "https://omlx.ai/benchmarks/performance/a5z3irwy",
   "tg_tokens": null,
   "tps": 97.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-a5z3irwy-ctx32768",
   "model": "Nemotron-3.5-Lightning-30B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: p2xjxwad; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1274.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Nemotron-3.5-Lightning-30B-A3B-8bit-MTP on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 1,274 91.5 34.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a5z3irwy",
   "source_url": "https://omlx.ai/benchmarks/performance/a5z3irwy",
   "tg_tokens": null,
   "tps": 91.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-rd02a28n-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 5.9 GB; TTFT 2698 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 379.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k tokens 379.5 PP tok/s 48.8 TG tok/s 2698 TTFT (ms) 5.9 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rd02a28n",
   "source_url": "https://omlx.ai/benchmarks/performance/rd02a28n",
   "tg_tokens": null,
   "tps": 48.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-q9mrb9wj-ctx1024",
   "model": "Qwen3 8B",
   "notes": "oMLX community benchmark trace; peak memory 5.0 GB at this context; duplicate traces of the same session: wolk723l; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1497.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-8B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 1,497 60.2 5.0 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q9mrb9wj",
   "source_url": "https://omlx.ai/benchmarks/performance/q9mrb9wj",
   "tg_tokens": null,
   "tps": 60.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-q9mrb9wj-ctx4096",
   "model": "Qwen3 8B",
   "notes": "oMLX community benchmark trace; peak memory 5.5 GB at this context; duplicate traces of the same session: wolk723l; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1426.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-8B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 1,426 55.1 5.5 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q9mrb9wj",
   "source_url": "https://omlx.ai/benchmarks/performance/q9mrb9wj",
   "tg_tokens": null,
   "tps": 55.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-z52faqck-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: 2hl6jmsh; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 392.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 1k 392.7 59.9 5.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z52faqck",
   "source_url": "https://omlx.ai/benchmarks/performance/z52faqck",
   "tg_tokens": null,
   "tps": 59.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-z52faqck-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: 2hl6jmsh; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 341.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 4k 341.0 39.6 6.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z52faqck",
   "source_url": "https://omlx.ai/benchmarks/performance/z52faqck",
   "tg_tokens": null,
   "tps": 39.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-z52faqck-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.6 GB at this context; duplicate traces of the same session: 2hl6jmsh; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 328.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 8k 328.0 39.1 6.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z52faqck",
   "source_url": "https://omlx.ai/benchmarks/performance/z52faqck",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-lcxkb850-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: yt8g68ra; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 398.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 1k 398.0 59.7 5.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lcxkb850",
   "source_url": "https://omlx.ai/benchmarks/performance/lcxkb850",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-lcxkb850-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: yt8g68ra; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 366.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 4k 366.9 45.6 6.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lcxkb850",
   "source_url": "https://omlx.ai/benchmarks/performance/lcxkb850",
   "tg_tokens": null,
   "tps": 45.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-lcxkb850-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.6 GB at this context; duplicate traces of the same session: yt8g68ra; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 346.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 8k 346.0 41.6 6.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lcxkb850",
   "source_url": "https://omlx.ai/benchmarks/performance/lcxkb850",
   "tg_tokens": null,
   "tps": 41.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-1f2d9gkb-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "DWQ build; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: angse6ff; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1492.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit-DWQ on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 8bit; 1k 1,492 97.2 20.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1f2d9gkb",
   "source_url": "https://omlx.ai/benchmarks/performance/1f2d9gkb",
   "tg_tokens": null,
   "tps": 97.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-1f2d9gkb-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "DWQ build; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: angse6ff; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1469.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit-DWQ on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 8bit; 8k 1,469 87.7 22.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1f2d9gkb",
   "source_url": "https://omlx.ai/benchmarks/performance/1f2d9gkb",
   "tg_tokens": null,
   "tps": 87.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-zvsfhsz0-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "DWQ build; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1488.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit-DWQ on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 8bit; 1k 1,488 97.4 20.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zvsfhsz0",
   "source_url": "https://omlx.ai/benchmarks/performance/zvsfhsz0",
   "tg_tokens": null,
   "tps": 97.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-zvsfhsz0-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "DWQ build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1594.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit-DWQ on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 8bit; 4k 1,594 90.7 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zvsfhsz0",
   "source_url": "https://omlx.ai/benchmarks/performance/zvsfhsz0",
   "tg_tokens": null,
   "tps": 90.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-b4d43gqr-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "DWQ build; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: fp1yhywl; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1487.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit-DWQ on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 8bit; 1k 1,487 97.2 20.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b4d43gqr",
   "source_url": "https://omlx.ai/benchmarks/performance/b4d43gqr",
   "tg_tokens": null,
   "tps": 97.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-b4d43gqr-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "DWQ build; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: fp1yhywl; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1468.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit-DWQ on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 8bit; 8k 1,468 88.2 22.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b4d43gqr",
   "source_url": "https://omlx.ai/benchmarks/performance/b4d43gqr",
   "tg_tokens": null,
   "tps": 88.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-mbl9wtol-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: 9q4k4lb3; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2756.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,756 77.1 22.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mbl9wtol",
   "source_url": "https://omlx.ai/benchmarks/performance/mbl9wtol",
   "tg_tokens": null,
   "tps": 77.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-mbl9wtol-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 9q4k4lb3; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2474.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 2,474 69.6 22.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mbl9wtol",
   "source_url": "https://omlx.ai/benchmarks/performance/mbl9wtol",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-l0npz6be-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: tqkzs913; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2772.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,772 76.9 22.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l0npz6be",
   "source_url": "https://omlx.ai/benchmarks/performance/l0npz6be",
   "tg_tokens": null,
   "tps": 76.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-l0npz6be-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: tqkzs913; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2497.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 2,497 69.6 22.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l0npz6be",
   "source_url": "https://omlx.ai/benchmarks/performance/l0npz6be",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-0zvbb5wx-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 257.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 257.5 29.3 16.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0zvbb5wx",
   "source_url": "https://omlx.ai/benchmarks/performance/0zvbb5wx",
   "tg_tokens": null,
   "tps": 29.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-0zvbb5wx-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 270.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 270.0 28.5 19.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0zvbb5wx",
   "source_url": "https://omlx.ai/benchmarks/performance/0zvbb5wx",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-5sbbc3sa-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 268.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 268.3 27.9 19.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5sbbc3sa",
   "source_url": "https://omlx.ai/benchmarks/performance/5sbbc3sa",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-5sbbc3sa-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 211.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 211.4 21.6 24.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5sbbc3sa",
   "source_url": "https://omlx.ai/benchmarks/performance/5sbbc3sa",
   "tg_tokens": null,
   "tps": 21.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-5sbbc3sa-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 169.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 169.8 17.6 28.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5sbbc3sa",
   "source_url": "https://omlx.ai/benchmarks/performance/5sbbc3sa",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ro2gp9fw-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: e6abao79; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 457.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 457.3 38.3 17.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ro2gp9fw",
   "source_url": "https://omlx.ai/benchmarks/performance/ro2gp9fw",
   "tg_tokens": null,
   "tps": 38.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ro2gp9fw-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: e6abao79; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 559.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 559.1 39.2 18.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ro2gp9fw",
   "source_url": "https://omlx.ai/benchmarks/performance/ro2gp9fw",
   "tg_tokens": null,
   "tps": 39.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-p2mpdwyx-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: ijb447vc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 927.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 927.8 32.9 17.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p2mpdwyx",
   "source_url": "https://omlx.ai/benchmarks/performance/p2mpdwyx",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-p2mpdwyx-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: ijb447vc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 781.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 781.6 30.9 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p2mpdwyx",
   "source_url": "https://omlx.ai/benchmarks/performance/p2mpdwyx",
   "tg_tokens": null,
   "tps": 30.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-b59w22cx-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 821.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 8bit; 1k 821.8 18.5 28.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b59w22cx",
   "source_url": "https://omlx.ai/benchmarks/performance/b59w22cx",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-b59w22cx-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 954.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 8bit; 4k 954.9 18.3 29.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b59w22cx",
   "source_url": "https://omlx.ai/benchmarks/performance/b59w22cx",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vdx6drb2-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MLX; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.0 GB at this context; duplicate traces of the same session: g7r3eg8u, x87fld60; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 394.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 394.5 69.5 99.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vdx6drb2",
   "source_url": "https://omlx.ai/benchmarks/performance/vdx6drb2",
   "tg_tokens": null,
   "tps": 69.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vdx6drb2-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MLX; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.5 GB at this context; duplicate traces of the same session: g7r3eg8u, x87fld60; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1114.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,114 63.2 99.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vdx6drb2",
   "source_url": "https://omlx.ai/benchmarks/performance/vdx6drb2",
   "tg_tokens": null,
   "tps": 63.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vdx6drb2-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MLX; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.4 GB at this context; duplicate traces of the same session: g7r3eg8u, x87fld60; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2010.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,010 64.9 100.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vdx6drb2",
   "source_url": "https://omlx.ai/benchmarks/performance/vdx6drb2",
   "tg_tokens": null,
   "tps": 64.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-i8cpft67-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.3 GB at this context; duplicate traces of the same session: lbov168k; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 435.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 1k 435.0 50.3 69.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i8cpft67",
   "source_url": "https://omlx.ai/benchmarks/performance/i8cpft67",
   "tg_tokens": null,
   "tps": 50.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-i8cpft67-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.9 GB at this context; duplicate traces of the same session: lbov168k; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 532.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 4k 532.9 47.0 69.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i8cpft67",
   "source_url": "https://omlx.ai/benchmarks/performance/i8cpft67",
   "tg_tokens": null,
   "tps": 47.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-i8cpft67-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.9 GB at this context; duplicate traces of the same session: lbov168k; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 624.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 8k 624.8 49.4 69.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i8cpft67",
   "source_url": "https://omlx.ai/benchmarks/performance/i8cpft67",
   "tg_tokens": null,
   "tps": 49.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-i8cpft67-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.1 GB at this context; duplicate traces of the same session: lbov168k; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 625.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 16k 625.8 47.0 70.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i8cpft67",
   "source_url": "https://omlx.ai/benchmarks/performance/i8cpft67",
   "tg_tokens": null,
   "tps": 47.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-vor87sya-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.3 GB at this context; duplicate traces of the same session: fthdh1wz; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 362.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 1k 362.6 44.0 69.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vor87sya",
   "source_url": "https://omlx.ai/benchmarks/performance/vor87sya",
   "tg_tokens": null,
   "tps": 44.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-vor87sya-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.9 GB at this context; duplicate traces of the same session: fthdh1wz; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 628.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 8k 628.6 45.9 69.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vor87sya",
   "source_url": "https://omlx.ai/benchmarks/performance/vor87sya",
   "tg_tokens": null,
   "tps": 45.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-vor87sya-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.1 GB at this context; duplicate traces of the same session: fthdh1wz; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 630.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 16k 630.2 48.2 70.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vor87sya",
   "source_url": "https://omlx.ai/benchmarks/performance/vor87sya",
   "tg_tokens": null,
   "tps": 48.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-yxwvdyh8-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.1 GB at this context; duplicate traces of the same session: 0nr9oj75; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1143.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,143 87.6 99.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yxwvdyh8",
   "source_url": "https://omlx.ai/benchmarks/performance/yxwvdyh8",
   "tg_tokens": null,
   "tps": 87.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-yxwvdyh8-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.7 GB at this context; duplicate traces of the same session: 0nr9oj75; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1423.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 1,423 78.7 99.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yxwvdyh8",
   "source_url": "https://omlx.ai/benchmarks/performance/yxwvdyh8",
   "tg_tokens": null,
   "tps": 78.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-yxwvdyh8-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.0 GB at this context; duplicate traces of the same session: 0nr9oj75; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1416.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 1,416 76.7 100.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yxwvdyh8",
   "source_url": "https://omlx.ai/benchmarks/performance/yxwvdyh8",
   "tg_tokens": null,
   "tps": 76.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-yxwvdyh8-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: 0nr9oj75; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1397.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 1,397 50.8 100.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yxwvdyh8",
   "source_url": "https://omlx.ai/benchmarks/performance/yxwvdyh8",
   "tg_tokens": null,
   "tps": 50.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-yxwvdyh8-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: 0nr9oj75; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1393.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 1,393 58.7 101.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yxwvdyh8",
   "source_url": "https://omlx.ai/benchmarks/performance/yxwvdyh8",
   "tg_tokens": null,
   "tps": 58.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-yxwvdyh8-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.7 GB at this context; duplicate traces of the same session: 0nr9oj75; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1351.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 1,351 66.1 104.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yxwvdyh8",
   "source_url": "https://omlx.ai/benchmarks/performance/yxwvdyh8",
   "tg_tokens": null,
   "tps": 66.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-27fsc2l4-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: jqhag980, yed6b0bn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 919.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 919.3 53.9 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/27fsc2l4",
   "source_url": "https://omlx.ai/benchmarks/performance/27fsc2l4",
   "tg_tokens": null,
   "tps": 53.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-27fsc2l4-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: jqhag980, yed6b0bn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1819.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,819 76.3 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/27fsc2l4",
   "source_url": "https://omlx.ai/benchmarks/performance/27fsc2l4",
   "tg_tokens": null,
   "tps": 76.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-27fsc2l4-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.9 GB at this context; duplicate traces of the same session: jqhag980, yed6b0bn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2646.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,646 70.9 74.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/27fsc2l4",
   "source_url": "https://omlx.ai/benchmarks/performance/27fsc2l4",
   "tg_tokens": null,
   "tps": 70.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-27fsc2l4-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.4 GB at this context; duplicate traces of the same session: jqhag980, yed6b0bn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2547.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,547 71.1 79.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/27fsc2l4",
   "source_url": "https://omlx.ai/benchmarks/performance/27fsc2l4",
   "tg_tokens": null,
   "tps": 71.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-27fsc2l4-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.9 GB at this context; duplicate traces of the same session: jqhag980, yed6b0bn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2300.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,300 56.5 80.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/27fsc2l4",
   "source_url": "https://omlx.ai/benchmarks/performance/27fsc2l4",
   "tg_tokens": null,
   "tps": 56.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-7tbob2la-ctx1024",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "it variant; oMLX community benchmark trace; peak memory 15.4 GB at this context; duplicate traces of the same session: yxu8xe7g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2796.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "diffusiongemma-26B-A4B-it-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 2,796 27.7 15.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7tbob2la",
   "source_url": "https://omlx.ai/benchmarks/performance/7tbob2la",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-7tbob2la-ctx4096",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "it variant; oMLX community benchmark trace; peak memory 16.2 GB at this context; duplicate traces of the same session: yxu8xe7g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2551.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "diffusiongemma-26B-A4B-it-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 2,551 20.8 16.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7tbob2la",
   "source_url": "https://omlx.ai/benchmarks/performance/7tbob2la",
   "tg_tokens": null,
   "tps": 20.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-7tbob2la-ctx8192",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "it variant; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: yxu8xe7g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2437.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "diffusiongemma-26B-A4B-it-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 2,437 23.5 16.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7tbob2la",
   "source_url": "https://omlx.ai/benchmarks/performance/7tbob2la",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-7tbob2la-ctx16384",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "it variant; oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: yxu8xe7g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2323.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "diffusiongemma-26B-A4B-it-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 2,323 47.6 17.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7tbob2la",
   "source_url": "https://omlx.ai/benchmarks/performance/7tbob2la",
   "tg_tokens": null,
   "tps": 47.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-7tbob2la-ctx32768",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "it variant; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: yxu8xe7g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2053.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "diffusiongemma-26B-A4B-it-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 2,053 18.8 18.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7tbob2la",
   "source_url": "https://omlx.ai/benchmarks/performance/7tbob2la",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-7tbob2la-ctx65536",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "it variant; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: yxu8xe7g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1621.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "diffusiongemma-26B-A4B-it-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 1,621 22.2 21.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7tbob2la",
   "source_url": "https://omlx.ai/benchmarks/performance/7tbob2la",
   "tg_tokens": null,
   "tps": 22.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-7tbob2la-ctx131072",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "it variant; oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: yxu8xe7g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1097.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "diffusiongemma-26B-A4B-it-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 1,097 10.0 26.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7tbob2la",
   "source_url": "https://omlx.ai/benchmarks/performance/7tbob2la",
   "tg_tokens": null,
   "tps": 10.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-pxos6zsf-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 766.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 766.7 62.3 14.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pxos6zsf",
   "source_url": "https://omlx.ai/benchmarks/performance/pxos6zsf",
   "tg_tokens": null,
   "tps": 62.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-pxos6zsf-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 719.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 719.9 63.5 14.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pxos6zsf",
   "source_url": "https://omlx.ai/benchmarks/performance/pxos6zsf",
   "tg_tokens": null,
   "tps": 63.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-6dc4bg7o-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 13.3 GB; TTFT 1648 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 621.4,
   "provenance": "community",
   "quant": "unknown",
   "quote": "gpt-oss-20b on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 unknown; 1k tokens 621.4 PP tok/s 54.5 TG tok/s 1648 TTFT (ms) 13.3 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6dc4bg7o",
   "source_url": "https://omlx.ai/benchmarks/performance/6dc4bg7o",
   "tg_tokens": null,
   "tps": 54.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-f8uhhlry-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: w87kyy53; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3803.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 3,803 115.1 27.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f8uhhlry",
   "source_url": "https://omlx.ai/benchmarks/performance/f8uhhlry",
   "tg_tokens": null,
   "tps": 115.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-f8uhhlry-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: w87kyy53; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4720.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 4,720 114.5 28.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f8uhhlry",
   "source_url": "https://omlx.ai/benchmarks/performance/f8uhhlry",
   "tg_tokens": null,
   "tps": 114.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-f8uhhlry-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: w87kyy53; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4731.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 8k 4,731 114.0 28.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f8uhhlry",
   "source_url": "https://omlx.ai/benchmarks/performance/f8uhhlry",
   "tg_tokens": null,
   "tps": 114.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-f8uhhlry-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai; oMLX community benchmark trace; peak memory 28.8 GB at this context; duplicate traces of the same session: w87kyy53; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4451.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k 4,451 109.5 28.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f8uhhlry",
   "source_url": "https://omlx.ai/benchmarks/performance/f8uhhlry",
   "tg_tokens": null,
   "tps": 109.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-f8uhhlry-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai; oMLX community benchmark trace; peak memory 29.1 GB at this context; duplicate traces of the same session: w87kyy53; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3947.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 3,947 101.3 29.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f8uhhlry",
   "source_url": "https://omlx.ai/benchmarks/performance/f8uhhlry",
   "tg_tokens": null,
   "tps": 101.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-h22tvhs5-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "ornith-ai repo; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2972.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "ornith-ai--Ornith-1.5-9B-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 unknown; 1k 2,972 33.7 17.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h22tvhs5",
   "source_url": "https://omlx.ai/benchmarks/performance/h22tvhs5",
   "tg_tokens": null,
   "tps": 33.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-h22tvhs5-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "ornith-ai repo; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2601.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "ornith-ai--Ornith-1.5-9B-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 unknown; 4k 2,601 33.6 18.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h22tvhs5",
   "source_url": "https://omlx.ai/benchmarks/performance/h22tvhs5",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-h22tvhs5-ctx8192",
   "model": "Ornith-1.5-9B",
   "notes": "ornith-ai repo; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3675.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "ornith-ai--Ornith-1.5-9B-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 unknown; 8k 3,675 33.3 18.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h22tvhs5",
   "source_url": "https://omlx.ai/benchmarks/performance/h22tvhs5",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-h22tvhs5-ctx16384",
   "model": "Ornith-1.5-9B",
   "notes": "ornith-ai repo; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3576.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "ornith-ai--Ornith-1.5-9B-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 unknown; 16k 3,576 32.7 18.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h22tvhs5",
   "source_url": "https://omlx.ai/benchmarks/performance/h22tvhs5",
   "tg_tokens": null,
   "tps": 32.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-h22tvhs5-ctx32768",
   "model": "Ornith-1.5-9B",
   "notes": "ornith-ai repo; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3333.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "ornith-ai--Ornith-1.5-9B-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 unknown; 32k 3,333 31.5 19.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h22tvhs5",
   "source_url": "https://omlx.ai/benchmarks/performance/h22tvhs5",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-6gxrtp1y-ctx1024",
   "model": "6-bit",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 437.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "6-bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 1k 437.5 12.8 22.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6gxrtp1y",
   "source_url": "https://omlx.ai/benchmarks/performance/6gxrtp1y",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-6gxrtp1y-ctx4096",
   "model": "6-bit",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 459.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "6-bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 4k 459.3 12.6 23.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6gxrtp1y",
   "source_url": "https://omlx.ai/benchmarks/performance/6gxrtp1y",
   "tg_tokens": null,
   "tps": 12.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-y1m2pnbu-ctx1024",
   "model": "DeepSeek-V4-Flash-0731-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 100.9 GB at this context; duplicate traces of the same session: 0fk3zwwr, 251ru7y9, uogwy2cq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 469.1,
   "provenance": "community",
   "quant": "Mixed 2bit-3bit g64",
   "quote": "DeepSeek-V4-Flash-0731-Abliterated-MLX-Mixed-2bit-3bit-g64 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 1k 469.1 28.0 100.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y1m2pnbu",
   "source_url": "https://omlx.ai/benchmarks/performance/y1m2pnbu",
   "tg_tokens": null,
   "tps": 28.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-y1m2pnbu-ctx4096",
   "model": "DeepSeek-V4-Flash-0731-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: 0fk3zwwr, 251ru7y9, uogwy2cq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 476.8,
   "provenance": "community",
   "quant": "Mixed 2bit-3bit g64",
   "quote": "DeepSeek-V4-Flash-0731-Abliterated-MLX-Mixed-2bit-3bit-g64 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 4k 476.8 25.8 101.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y1m2pnbu",
   "source_url": "https://omlx.ai/benchmarks/performance/y1m2pnbu",
   "tg_tokens": null,
   "tps": 25.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-y1m2pnbu-ctx8192",
   "model": "DeepSeek-V4-Flash-0731-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: 0fk3zwwr, 251ru7y9, uogwy2cq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 470.7,
   "provenance": "community",
   "quant": "Mixed 2bit-3bit g64",
   "quote": "DeepSeek-V4-Flash-0731-Abliterated-MLX-Mixed-2bit-3bit-g64 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 8k 470.7 25.1 101.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y1m2pnbu",
   "source_url": "https://omlx.ai/benchmarks/performance/y1m2pnbu",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-y1m2pnbu-ctx16384",
   "model": "DeepSeek-V4-Flash-0731-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 101.4 GB at this context; duplicate traces of the same session: 0fk3zwwr, 251ru7y9, uogwy2cq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 462.1,
   "provenance": "community",
   "quant": "Mixed 2bit-3bit g64",
   "quote": "DeepSeek-V4-Flash-0731-Abliterated-MLX-Mixed-2bit-3bit-g64 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 16k 462.1 24.3 101.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y1m2pnbu",
   "source_url": "https://omlx.ai/benchmarks/performance/y1m2pnbu",
   "tg_tokens": null,
   "tps": 24.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-y1m2pnbu-ctx32768",
   "model": "DeepSeek-V4-Flash-0731-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: 0fk3zwwr, 251ru7y9, uogwy2cq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 448.6,
   "provenance": "community",
   "quant": "Mixed 2bit-3bit g64",
   "quote": "DeepSeek-V4-Flash-0731-Abliterated-MLX-Mixed-2bit-3bit-g64 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 32k 448.6 23.0 101.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y1m2pnbu",
   "source_url": "https://omlx.ai/benchmarks/performance/y1m2pnbu",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-y1m2pnbu-ctx65536",
   "model": "DeepSeek-V4-Flash-0731-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 101.8 GB at this context; duplicate traces of the same session: 0fk3zwwr, 251ru7y9, uogwy2cq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 424.9,
   "provenance": "community",
   "quant": "Mixed 2bit-3bit g64",
   "quote": "DeepSeek-V4-Flash-0731-Abliterated-MLX-Mixed-2bit-3bit-g64 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 64k 424.9 22.8 101.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y1m2pnbu",
   "source_url": "https://omlx.ai/benchmarks/performance/y1m2pnbu",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-yplwbprq-ctx1024",
   "model": "Huihui-Qwen3-VL-4B-Instruct-abliterated",
   "notes": "MLX; oMLX community benchmark trace; peak memory 2.9 GB at this context; duplicate traces of the same session: sikzqtlw; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1319.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Huihui-Qwen3-VL-4B-Instruct-abliterated-4bit-mlx on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 1,319 115.2 2.9 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yplwbprq",
   "source_url": "https://omlx.ai/benchmarks/performance/yplwbprq",
   "tg_tokens": null,
   "tps": 115.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-yplwbprq-ctx4096",
   "model": "Huihui-Qwen3-VL-4B-Instruct-abliterated",
   "notes": "MLX; oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: sikzqtlw; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1212.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Huihui-Qwen3-VL-4B-Instruct-abliterated-4bit-mlx on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 1,212 95.8 3.4 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yplwbprq",
   "source_url": "https://omlx.ai/benchmarks/performance/yplwbprq",
   "tg_tokens": null,
   "tps": 95.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-q5rxsf9q-ctx1024",
   "model": "Huihui-Qwen3-VL-4B-Instruct-abliterated",
   "notes": "MLX; oMLX community benchmark trace; peak memory 2.9 GB at this context; duplicate traces of the same session: js7bsbby; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1323.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Huihui-Qwen3-VL-4B-Instruct-abliterated-4bit-mlx on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 1,323 115.0 2.9 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q5rxsf9q",
   "source_url": "https://omlx.ai/benchmarks/performance/q5rxsf9q",
   "tg_tokens": null,
   "tps": 115.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-q5rxsf9q-ctx4096",
   "model": "Huihui-Qwen3-VL-4B-Instruct-abliterated",
   "notes": "MLX; oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: js7bsbby; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1224.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Huihui-Qwen3-VL-4B-Instruct-abliterated-4bit-mlx on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 1,224 96.3 3.4 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q5rxsf9q",
   "source_url": "https://omlx.ai/benchmarks/performance/q5rxsf9q",
   "tg_tokens": null,
   "tps": 96.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-fzx0u28j-ctx4096",
   "model": "Gemma 4 E2B",
   "notes": "Jundot build; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 4.3 GB; TTFT 517 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 7923.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--gemma-4-E2B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 7,923 PP tok/s 376.3 TG tok/s 517 TTFT (ms) 4.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fzx0u28j",
   "source_url": "https://omlx.ai/benchmarks/performance/fzx0u28j",
   "tg_tokens": null,
   "tps": 376.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-cap8m0iy-ctx4096",
   "model": "Gemma 4 E2B",
   "notes": "Jundot build; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 4.3 GB; TTFT 515 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 7956.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--gemma-4-E2B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 7,956 PP tok/s 144.6 TG tok/s 515 TTFT (ms) 4.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cap8m0iy",
   "source_url": "https://omlx.ai/benchmarks/performance/cap8m0iy",
   "tg_tokens": null,
   "tps": 144.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-37iek1z1-ctx4096",
   "model": "Gemma 4 E2B",
   "notes": "Jundot build; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 4.2 GB; TTFT 516 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 7935.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--gemma-4-E2B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 7,935 PP tok/s 379.0 TG tok/s 516 TTFT (ms) 4.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/37iek1z1",
   "source_url": "https://omlx.ai/benchmarks/performance/37iek1z1",
   "tg_tokens": null,
   "tps": 379.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-hg902we3-ctx4096",
   "model": "Gemma 4 E2B",
   "notes": "Jundot build; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 4.2 GB; TTFT 520 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 7881.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--gemma-4-E2B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 7,881 PP tok/s 347.2 TG tok/s 520 TTFT (ms) 4.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hg902we3",
   "source_url": "https://omlx.ai/benchmarks/performance/hg902we3",
   "tg_tokens": null,
   "tps": 347.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5srv21v4-ctx1024",
   "model": "MiMo-7B-RL",
   "notes": "oMLX community benchmark trace; peak memory 4.7 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1289.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-7B-RL-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 1,289 120.3 4.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5srv21v4",
   "source_url": "https://omlx.ai/benchmarks/performance/5srv21v4",
   "tg_tokens": null,
   "tps": 120.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5srv21v4-ctx4096",
   "model": "MiMo-7B-RL",
   "notes": "oMLX community benchmark trace; peak memory 5.2 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1247.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-7B-RL-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 1,247 108.0 5.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5srv21v4",
   "source_url": "https://omlx.ai/benchmarks/performance/5srv21v4",
   "tg_tokens": null,
   "tps": 108.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5srv21v4-ctx8192",
   "model": "MiMo-7B-RL",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1150.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-7B-RL-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 1,150 94.2 5.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5srv21v4",
   "source_url": "https://omlx.ai/benchmarks/performance/5srv21v4",
   "tg_tokens": null,
   "tps": 94.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-p2xjxwad-ctx16384",
   "model": "Nemotron-3.5-Lightning-30B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1456.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Nemotron-3.5-Lightning-30B-A3B-8bit-MTP on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 1,456 97.3 34.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p2xjxwad",
   "source_url": "https://omlx.ai/benchmarks/performance/p2xjxwad",
   "tg_tokens": null,
   "tps": 97.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-p2xjxwad-ctx32768",
   "model": "Nemotron-3.5-Lightning-30B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1274.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Nemotron-3.5-Lightning-30B-A3B-8bit-MTP on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 1,274 91.5 34.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p2xjxwad",
   "source_url": "https://omlx.ai/benchmarks/performance/p2xjxwad",
   "tg_tokens": null,
   "tps": 91.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-wolk723l-ctx1024",
   "model": "Qwen3 8B",
   "notes": "oMLX community benchmark trace; peak memory 5.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1497.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-8B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 1,497 60.2 5.0 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wolk723l",
   "source_url": "https://omlx.ai/benchmarks/performance/wolk723l",
   "tg_tokens": null,
   "tps": 60.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-wolk723l-ctx4096",
   "model": "Qwen3 8B",
   "notes": "oMLX community benchmark trace; peak memory 5.5 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1426.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-8B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 1,426 55.1 5.5 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wolk723l",
   "source_url": "https://omlx.ai/benchmarks/performance/wolk723l",
   "tg_tokens": null,
   "tps": 55.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-2hl6jmsh-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 392.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 1k 392.7 59.9 5.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2hl6jmsh",
   "source_url": "https://omlx.ai/benchmarks/performance/2hl6jmsh",
   "tg_tokens": null,
   "tps": 59.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-2hl6jmsh-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 341.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 4k 341.0 39.6 6.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2hl6jmsh",
   "source_url": "https://omlx.ai/benchmarks/performance/2hl6jmsh",
   "tg_tokens": null,
   "tps": 39.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-2hl6jmsh-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.6 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 328.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 8k 328.0 39.1 6.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2hl6jmsh",
   "source_url": "https://omlx.ai/benchmarks/performance/2hl6jmsh",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-yt8g68ra-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 398.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 1k 398.0 59.7 5.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yt8g68ra",
   "source_url": "https://omlx.ai/benchmarks/performance/yt8g68ra",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-yt8g68ra-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 366.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 4k 366.9 45.6 6.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yt8g68ra",
   "source_url": "https://omlx.ai/benchmarks/performance/yt8g68ra",
   "tg_tokens": null,
   "tps": 45.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-yt8g68ra-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.6 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 346.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 8k 346.0 41.6 6.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yt8g68ra",
   "source_url": "https://omlx.ai/benchmarks/performance/yt8g68ra",
   "tg_tokens": null,
   "tps": 41.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-fp1yhywl-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "DWQ build; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1487.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit-DWQ on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 8bit; 1k 1,487 97.2 20.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fp1yhywl",
   "source_url": "https://omlx.ai/benchmarks/performance/fp1yhywl",
   "tg_tokens": null,
   "tps": 97.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-fp1yhywl-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "DWQ build; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1468.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit-DWQ on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 8bit; 8k 1,468 88.2 22.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fp1yhywl",
   "source_url": "https://omlx.ai/benchmarks/performance/fp1yhywl",
   "tg_tokens": null,
   "tps": 88.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-angse6ff-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "DWQ build; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1492.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit-DWQ on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 8bit; 1k 1,492 97.2 20.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/angse6ff",
   "source_url": "https://omlx.ai/benchmarks/performance/angse6ff",
   "tg_tokens": null,
   "tps": 97.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-angse6ff-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "DWQ build; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1469.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit-DWQ on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 8bit; 8k 1,469 87.7 22.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/angse6ff",
   "source_url": "https://omlx.ai/benchmarks/performance/angse6ff",
   "tg_tokens": null,
   "tps": 87.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-apqq6ig2-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "DWQ build; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1488.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit-DWQ on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 8bit; 1k 1,488 97.4 20.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/apqq6ig2",
   "source_url": "https://omlx.ai/benchmarks/performance/apqq6ig2",
   "tg_tokens": null,
   "tps": 97.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-apqq6ig2-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "DWQ build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1594.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit-DWQ on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 8bit; 4k 1,594 90.7 22.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/apqq6ig2",
   "source_url": "https://omlx.ai/benchmarks/performance/apqq6ig2",
   "tg_tokens": null,
   "tps": 90.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-9q4k4lb3-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2756.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,756 77.1 22.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9q4k4lb3",
   "source_url": "https://omlx.ai/benchmarks/performance/9q4k4lb3",
   "tg_tokens": null,
   "tps": 77.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-9q4k4lb3-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2474.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 2,474 69.6 22.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9q4k4lb3",
   "source_url": "https://omlx.ai/benchmarks/performance/9q4k4lb3",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-tqkzs913-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2772.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,772 76.9 22.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tqkzs913",
   "source_url": "https://omlx.ai/benchmarks/performance/tqkzs913",
   "tg_tokens": null,
   "tps": 76.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-tqkzs913-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2497.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 2,497 69.6 22.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tqkzs913",
   "source_url": "https://omlx.ai/benchmarks/performance/tqkzs913",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-2hs522fa-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: l8i1uy64; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1379.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.6-35B-A3B-mxfp4 on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 1,379 89.5 19.6 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2hs522fa",
   "source_url": "https://omlx.ai/benchmarks/performance/2hs522fa",
   "tg_tokens": null,
   "tps": 89.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-2hs522fa-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: l8i1uy64; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1610.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.6-35B-A3B-mxfp4 on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 1,610 82.1 20.3 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2hs522fa",
   "source_url": "https://omlx.ai/benchmarks/performance/2hs522fa",
   "tg_tokens": null,
   "tps": 82.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-cph4anjk-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: yhf027pg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 256.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 256.9 28.6 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cph4anjk",
   "source_url": "https://omlx.ai/benchmarks/performance/cph4anjk",
   "tg_tokens": null,
   "tps": 28.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-cph4anjk-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: yhf027pg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 270.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 270.4 28.2 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cph4anjk",
   "source_url": "https://omlx.ai/benchmarks/performance/cph4anjk",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-e6abao79-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 457.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 457.3 38.3 17.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e6abao79",
   "source_url": "https://omlx.ai/benchmarks/performance/e6abao79",
   "tg_tokens": null,
   "tps": 38.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-e6abao79-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 559.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 559.1 39.2 18.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e6abao79",
   "source_url": "https://omlx.ai/benchmarks/performance/e6abao79",
   "tg_tokens": null,
   "tps": 39.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ijb447vc-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 927.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 927.8 32.9 17.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ijb447vc",
   "source_url": "https://omlx.ai/benchmarks/performance/ijb447vc",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ijb447vc-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 781.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 781.6 30.9 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ijb447vc",
   "source_url": "https://omlx.ai/benchmarks/performance/ijb447vc",
   "tg_tokens": null,
   "tps": 30.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kqaubgd7-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 26.9 GB; TTFT 49937 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 656.2,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-27B-NVFP4 on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 nvfp4; 32k tokens 656.2 PP tok/s 24.4 TG tok/s 49937 TTFT (ms) 26.9 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kqaubgd7",
   "source_url": "https://omlx.ai/benchmarks/performance/kqaubgd7",
   "tg_tokens": null,
   "tps": 24.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-glhki8qt-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 27.7 GB; TTFT 11972 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 342.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-Uncensored-oQ6e-mtp on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 6bit; 4k tokens 342.1 PP tok/s 36.7 TG tok/s 11972 TTFT (ms) 27.7 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/glhki8qt",
   "source_url": "https://omlx.ai/benchmarks/performance/glhki8qt",
   "tg_tokens": null,
   "tps": 36.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-nh3438tf-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.2 GB at this context; duplicate traces of the same session: goy771u3; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 209.4,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 209.4 15.6 16.2 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nh3438tf",
   "source_url": "https://omlx.ai/benchmarks/performance/nh3438tf",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-nh3438tf-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: goy771u3; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 212.3,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 212.3 18.9 17.3 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nh3438tf",
   "source_url": "https://omlx.ai/benchmarks/performance/nh3438tf",
   "tg_tokens": null,
   "tps": 18.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2t32b5d9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 747.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 747.2 43.9 17.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2t32b5d9",
   "source_url": "https://omlx.ai/benchmarks/performance/2t32b5d9",
   "tg_tokens": null,
   "tps": 43.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2t32b5d9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 926.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 926.6 50.6 18.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2t32b5d9",
   "source_url": "https://omlx.ai/benchmarks/performance/2t32b5d9",
   "tg_tokens": null,
   "tps": 50.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-x87fld60-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MLX; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.0 GB at this context; duplicate traces of the same session: g7r3eg8u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 394.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 394.5 69.5 99.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x87fld60",
   "source_url": "https://omlx.ai/benchmarks/performance/x87fld60",
   "tg_tokens": null,
   "tps": 69.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-x87fld60-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MLX; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.5 GB at this context; duplicate traces of the same session: g7r3eg8u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1114.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,114 63.2 99.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x87fld60",
   "source_url": "https://omlx.ai/benchmarks/performance/x87fld60",
   "tg_tokens": null,
   "tps": 63.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-x87fld60-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MLX; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.4 GB at this context; duplicate traces of the same session: g7r3eg8u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2010.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,010 64.9 100.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x87fld60",
   "source_url": "https://omlx.ai/benchmarks/performance/x87fld60",
   "tg_tokens": null,
   "tps": 64.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-lbov168k-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.3 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 435.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 1k 435.0 50.3 69.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lbov168k",
   "source_url": "https://omlx.ai/benchmarks/performance/lbov168k",
   "tg_tokens": null,
   "tps": 50.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-lbov168k-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.9 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 532.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 4k 532.9 47.0 69.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lbov168k",
   "source_url": "https://omlx.ai/benchmarks/performance/lbov168k",
   "tg_tokens": null,
   "tps": 47.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-lbov168k-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.9 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 624.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 8k 624.8 49.4 69.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lbov168k",
   "source_url": "https://omlx.ai/benchmarks/performance/lbov168k",
   "tg_tokens": null,
   "tps": 49.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-lbov168k-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.1 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 625.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 16k 625.8 47.0 70.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lbov168k",
   "source_url": "https://omlx.ai/benchmarks/performance/lbov168k",
   "tg_tokens": null,
   "tps": 47.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-fthdh1wz-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.3 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 362.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 1k 362.6 44.0 69.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fthdh1wz",
   "source_url": "https://omlx.ai/benchmarks/performance/fthdh1wz",
   "tg_tokens": null,
   "tps": 44.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-fthdh1wz-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.9 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 628.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 8k 628.6 45.9 69.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fthdh1wz",
   "source_url": "https://omlx.ai/benchmarks/performance/fthdh1wz",
   "tg_tokens": null,
   "tps": 45.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-fthdh1wz-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.1 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 630.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 16k 630.2 48.2 70.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fthdh1wz",
   "source_url": "https://omlx.ai/benchmarks/performance/fthdh1wz",
   "tg_tokens": null,
   "tps": 48.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-0nr9oj75-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1143.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,143 87.6 99.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0nr9oj75",
   "source_url": "https://omlx.ai/benchmarks/performance/0nr9oj75",
   "tg_tokens": null,
   "tps": 87.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-0nr9oj75-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1423.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 1,423 78.7 99.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0nr9oj75",
   "source_url": "https://omlx.ai/benchmarks/performance/0nr9oj75",
   "tg_tokens": null,
   "tps": 78.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-0nr9oj75-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.0 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1416.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 1,416 76.7 100.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0nr9oj75",
   "source_url": "https://omlx.ai/benchmarks/performance/0nr9oj75",
   "tg_tokens": null,
   "tps": 76.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-0nr9oj75-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1397.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 1,397 50.8 100.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0nr9oj75",
   "source_url": "https://omlx.ai/benchmarks/performance/0nr9oj75",
   "tg_tokens": null,
   "tps": 50.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-0nr9oj75-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1393.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 1,393 58.7 101.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0nr9oj75",
   "source_url": "https://omlx.ai/benchmarks/performance/0nr9oj75",
   "tg_tokens": null,
   "tps": 58.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0nr9oj75-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1351.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 1,351 66.1 104.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0nr9oj75",
   "source_url": "https://omlx.ai/benchmarks/performance/0nr9oj75",
   "tg_tokens": null,
   "tps": 66.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yed6b0bn-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: jqhag980; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 919.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 919.3 53.9 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yed6b0bn",
   "source_url": "https://omlx.ai/benchmarks/performance/yed6b0bn",
   "tg_tokens": null,
   "tps": 53.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yed6b0bn-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: jqhag980; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1819.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,819 76.3 71.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yed6b0bn",
   "source_url": "https://omlx.ai/benchmarks/performance/yed6b0bn",
   "tg_tokens": null,
   "tps": 76.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yed6b0bn-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.9 GB at this context; duplicate traces of the same session: jqhag980; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2646.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,646 70.9 74.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yed6b0bn",
   "source_url": "https://omlx.ai/benchmarks/performance/yed6b0bn",
   "tg_tokens": null,
   "tps": 70.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-yed6b0bn-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.4 GB at this context; duplicate traces of the same session: jqhag980; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2547.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,547 71.1 79.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yed6b0bn",
   "source_url": "https://omlx.ai/benchmarks/performance/yed6b0bn",
   "tg_tokens": null,
   "tps": 71.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yed6b0bn-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.9 GB at this context; duplicate traces of the same session: jqhag980; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2300.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,300 56.5 80.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yed6b0bn",
   "source_url": "https://omlx.ai/benchmarks/performance/yed6b0bn",
   "tg_tokens": null,
   "tps": 56.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-e4vivjra-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.8 GB; TTFT 2818 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 363.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 363.4 PP tok/s 44.2 TG tok/s 2818 TTFT (ms) 70.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e4vivjra",
   "source_url": "https://omlx.ai/benchmarks/performance/e4vivjra",
   "tg_tokens": null,
   "tps": 44.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dhugw513-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: zngxxnho; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1158.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,158 63.6 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dhugw513",
   "source_url": "https://omlx.ai/benchmarks/performance/dhugw513",
   "tg_tokens": null,
   "tps": 63.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dhugw513-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: zngxxnho; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1981.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,981 52.9 70.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dhugw513",
   "source_url": "https://omlx.ai/benchmarks/performance/dhugw513",
   "tg_tokens": null,
   "tps": 52.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n3nmy13p-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 74.4 GB; TTFT 29961 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2187.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k tokens 2,187 PP tok/s 54.9 TG tok/s 29961 TTFT (ms) 74.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n3nmy13p",
   "source_url": "https://omlx.ai/benchmarks/performance/n3nmy13p",
   "tg_tokens": null,
   "tps": 54.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-yq5dz1vt-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.4 GB at this context; duplicate traces of the same session: 2gu1ip7s, a57hc7rx, sgkzlmux; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 246.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 246.3 38.7 32.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yq5dz1vt",
   "source_url": "https://omlx.ai/benchmarks/performance/yq5dz1vt",
   "tg_tokens": null,
   "tps": 38.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-yq5dz1vt-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.6 GB at this context; duplicate traces of the same session: 2gu1ip7s, a57hc7rx, sgkzlmux; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 276.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 276.6 43.3 35.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yq5dz1vt",
   "source_url": "https://omlx.ai/benchmarks/performance/yq5dz1vt",
   "tg_tokens": null,
   "tps": 43.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-yq5dz1vt-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: 2gu1ip7s, a57hc7rx, sgkzlmux; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 296.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 8k 296.5 41.0 36.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yq5dz1vt",
   "source_url": "https://omlx.ai/benchmarks/performance/yq5dz1vt",
   "tg_tokens": null,
   "tps": 41.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-yq5dz1vt-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.0 GB at this context; duplicate traces of the same session: 2gu1ip7s, a57hc7rx, sgkzlmux; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 292.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k 292.6 35.0 37.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yq5dz1vt",
   "source_url": "https://omlx.ai/benchmarks/performance/yq5dz1vt",
   "tg_tokens": null,
   "tps": 35.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-yxu8xe7g-ctx1024",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "it variant; oMLX community benchmark trace; peak memory 15.4 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2796.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "diffusiongemma-26B-A4B-it-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 2,796 27.7 15.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yxu8xe7g",
   "source_url": "https://omlx.ai/benchmarks/performance/yxu8xe7g",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-yxu8xe7g-ctx4096",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "it variant; oMLX community benchmark trace; peak memory 16.2 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2551.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "diffusiongemma-26B-A4B-it-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 2,551 20.8 16.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yxu8xe7g",
   "source_url": "https://omlx.ai/benchmarks/performance/yxu8xe7g",
   "tg_tokens": null,
   "tps": 20.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-yxu8xe7g-ctx8192",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "it variant; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2437.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "diffusiongemma-26B-A4B-it-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 2,437 23.5 16.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yxu8xe7g",
   "source_url": "https://omlx.ai/benchmarks/performance/yxu8xe7g",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-yxu8xe7g-ctx16384",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "it variant; oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2323.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "diffusiongemma-26B-A4B-it-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 2,323 47.6 17.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yxu8xe7g",
   "source_url": "https://omlx.ai/benchmarks/performance/yxu8xe7g",
   "tg_tokens": null,
   "tps": 47.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-yxu8xe7g-ctx32768",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "it variant; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2053.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "diffusiongemma-26B-A4B-it-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 2,053 18.8 18.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yxu8xe7g",
   "source_url": "https://omlx.ai/benchmarks/performance/yxu8xe7g",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-yxu8xe7g-ctx65536",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "it variant; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1621.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "diffusiongemma-26B-A4B-it-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 1,621 22.2 21.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yxu8xe7g",
   "source_url": "https://omlx.ai/benchmarks/performance/yxu8xe7g",
   "tg_tokens": null,
   "tps": 22.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-yxu8xe7g-ctx131072",
   "model": "DiffusionGemma-26B-A4B",
   "notes": "it variant; oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1097.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "diffusiongemma-26B-A4B-it-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 1,097 10.0 26.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yxu8xe7g",
   "source_url": "https://omlx.ai/benchmarks/performance/yxu8xe7g",
   "tg_tokens": null,
   "tps": 10.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ahcaydjx-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 652.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 1k 652.2 70.6 14.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ahcaydjx",
   "source_url": "https://omlx.ai/benchmarks/performance/ahcaydjx",
   "tg_tokens": null,
   "tps": 70.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-ahcaydjx-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 588.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 4k 588.7 60.0 14.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ahcaydjx",
   "source_url": "https://omlx.ai/benchmarks/performance/ahcaydjx",
   "tg_tokens": null,
   "tps": 60.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ahcaydjx-ctx8192",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 485.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 8k 485.0 58.1 14.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ahcaydjx",
   "source_url": "https://omlx.ai/benchmarks/performance/ahcaydjx",
   "tg_tokens": null,
   "tps": 58.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-f1qg4mqo-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.1 GB at this context; duplicate traces of the same session: 2lz5604k, 9wtp68d1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1712.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 1,712 112.0 14.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f1qg4mqo",
   "source_url": "https://omlx.ai/benchmarks/performance/f1qg4mqo",
   "tg_tokens": null,
   "tps": 112.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-f1qg4mqo-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: 2lz5604k, 9wtp68d1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1727.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 1,727 110.8 14.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f1qg4mqo",
   "source_url": "https://omlx.ai/benchmarks/performance/f1qg4mqo",
   "tg_tokens": null,
   "tps": 110.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-f1qg4mqo-ctx8192",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: 2lz5604k, 9wtp68d1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1806.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 1,806 103.8 14.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f1qg4mqo",
   "source_url": "https://omlx.ai/benchmarks/performance/f1qg4mqo",
   "tg_tokens": null,
   "tps": 103.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-w87kyy53-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3803.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 3,803 115.1 27.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w87kyy53",
   "source_url": "https://omlx.ai/benchmarks/performance/w87kyy53",
   "tg_tokens": null,
   "tps": 115.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-w87kyy53-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4720.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 4,720 114.5 28.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w87kyy53",
   "source_url": "https://omlx.ai/benchmarks/performance/w87kyy53",
   "tg_tokens": null,
   "tps": 114.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-w87kyy53-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4731.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 8k 4,731 114.0 28.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w87kyy53",
   "source_url": "https://omlx.ai/benchmarks/performance/w87kyy53",
   "tg_tokens": null,
   "tps": 114.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-w87kyy53-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai; oMLX community benchmark trace; peak memory 28.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4451.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k 4,451 109.5 28.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w87kyy53",
   "source_url": "https://omlx.ai/benchmarks/performance/w87kyy53",
   "tg_tokens": null,
   "tps": 109.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-w87kyy53-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai; oMLX community benchmark trace; peak memory 29.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3947.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 3,947 101.3 29.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w87kyy53",
   "source_url": "https://omlx.ai/benchmarks/performance/w87kyy53",
   "tg_tokens": null,
   "tps": 101.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-251ru7y9-ctx1024",
   "model": "DeepSeek-V4-Flash-0731-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 100.9 GB at this context; duplicate traces of the same session: 0fk3zwwr, uogwy2cq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 469.1,
   "provenance": "community",
   "quant": "Mixed 2bit-3bit g64",
   "quote": "DeepSeek-V4-Flash-0731-Abliterated-MLX-Mixed-2bit-3bit-g64 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 1k 469.1 28.0 100.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/251ru7y9",
   "source_url": "https://omlx.ai/benchmarks/performance/251ru7y9",
   "tg_tokens": null,
   "tps": 28.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-251ru7y9-ctx4096",
   "model": "DeepSeek-V4-Flash-0731-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: 0fk3zwwr, uogwy2cq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 476.8,
   "provenance": "community",
   "quant": "Mixed 2bit-3bit g64",
   "quote": "DeepSeek-V4-Flash-0731-Abliterated-MLX-Mixed-2bit-3bit-g64 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 4k 476.8 25.8 101.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/251ru7y9",
   "source_url": "https://omlx.ai/benchmarks/performance/251ru7y9",
   "tg_tokens": null,
   "tps": 25.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-251ru7y9-ctx8192",
   "model": "DeepSeek-V4-Flash-0731-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: 0fk3zwwr, uogwy2cq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 470.7,
   "provenance": "community",
   "quant": "Mixed 2bit-3bit g64",
   "quote": "DeepSeek-V4-Flash-0731-Abliterated-MLX-Mixed-2bit-3bit-g64 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 8k 470.7 25.1 101.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/251ru7y9",
   "source_url": "https://omlx.ai/benchmarks/performance/251ru7y9",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-251ru7y9-ctx16384",
   "model": "DeepSeek-V4-Flash-0731-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 101.4 GB at this context; duplicate traces of the same session: 0fk3zwwr, uogwy2cq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 462.1,
   "provenance": "community",
   "quant": "Mixed 2bit-3bit g64",
   "quote": "DeepSeek-V4-Flash-0731-Abliterated-MLX-Mixed-2bit-3bit-g64 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 16k 462.1 24.3 101.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/251ru7y9",
   "source_url": "https://omlx.ai/benchmarks/performance/251ru7y9",
   "tg_tokens": null,
   "tps": 24.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-251ru7y9-ctx32768",
   "model": "DeepSeek-V4-Flash-0731-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: 0fk3zwwr, uogwy2cq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 448.6,
   "provenance": "community",
   "quant": "Mixed 2bit-3bit g64",
   "quote": "DeepSeek-V4-Flash-0731-Abliterated-MLX-Mixed-2bit-3bit-g64 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 32k 448.6 23.0 101.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/251ru7y9",
   "source_url": "https://omlx.ai/benchmarks/performance/251ru7y9",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-251ru7y9-ctx65536",
   "model": "DeepSeek-V4-Flash-0731-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 101.8 GB at this context; duplicate traces of the same session: 0fk3zwwr, uogwy2cq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 424.9,
   "provenance": "community",
   "quant": "Mixed 2bit-3bit g64",
   "quote": "DeepSeek-V4-Flash-0731-Abliterated-MLX-Mixed-2bit-3bit-g64 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 64k 424.9 22.8 101.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/251ru7y9",
   "source_url": "https://omlx.ai/benchmarks/performance/251ru7y9",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-foa9be91-ctx1024",
   "model": "Gemma-4-Dark-Gemistry-31B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 33.7 GB at this context; duplicate traces of the same session: 6snphjc1, 78ms2a10; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 92.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Gemma-4-Dark-Gemistry-31B-mlx-8Bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 1k 92.3 8.5 33.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/foa9be91",
   "source_url": "https://omlx.ai/benchmarks/performance/foa9be91",
   "tg_tokens": null,
   "tps": 8.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-foa9be91-ctx4096",
   "model": "Gemma-4-Dark-Gemistry-31B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 36.8 GB at this context; duplicate traces of the same session: 6snphjc1, 78ms2a10; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 82.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Gemma-4-Dark-Gemistry-31B-mlx-8Bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 4k 82.8 7.7 36.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/foa9be91",
   "source_url": "https://omlx.ai/benchmarks/performance/foa9be91",
   "tg_tokens": null,
   "tps": 7.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-foa9be91-ctx8192",
   "model": "Gemma-4-Dark-Gemistry-31B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 40.4 GB at this context; duplicate traces of the same session: 6snphjc1, 78ms2a10; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 79.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Gemma-4-Dark-Gemistry-31B-mlx-8Bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 8k 79.9 7.4 40.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/foa9be91",
   "source_url": "https://omlx.ai/benchmarks/performance/foa9be91",
   "tg_tokens": null,
   "tps": 7.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-sikzqtlw-ctx1024",
   "model": "Huihui-Qwen3-VL-4B-Instruct-abliterated",
   "notes": "MLX; oMLX community benchmark trace; peak memory 2.9 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1319.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Huihui-Qwen3-VL-4B-Instruct-abliterated-4bit-mlx on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 1,319 115.2 2.9 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sikzqtlw",
   "source_url": "https://omlx.ai/benchmarks/performance/sikzqtlw",
   "tg_tokens": null,
   "tps": 115.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-sikzqtlw-ctx4096",
   "model": "Huihui-Qwen3-VL-4B-Instruct-abliterated",
   "notes": "MLX; oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1212.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Huihui-Qwen3-VL-4B-Instruct-abliterated-4bit-mlx on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 1,212 95.8 3.4 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sikzqtlw",
   "source_url": "https://omlx.ai/benchmarks/performance/sikzqtlw",
   "tg_tokens": null,
   "tps": 95.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-js7bsbby-ctx1024",
   "model": "Huihui-Qwen3-VL-4B-Instruct-abliterated",
   "notes": "MLX; oMLX community benchmark trace; peak memory 2.9 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1323.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Huihui-Qwen3-VL-4B-Instruct-abliterated-4bit-mlx on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 1,323 115.0 2.9 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/js7bsbby",
   "source_url": "https://omlx.ai/benchmarks/performance/js7bsbby",
   "tg_tokens": null,
   "tps": 115.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-js7bsbby-ctx4096",
   "model": "Huihui-Qwen3-VL-4B-Instruct-abliterated",
   "notes": "MLX; oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1224.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Huihui-Qwen3-VL-4B-Instruct-abliterated-4bit-mlx on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 1,224 96.3 3.4 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/js7bsbby",
   "source_url": "https://omlx.ai/benchmarks/performance/js7bsbby",
   "tg_tokens": null,
   "tps": 96.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-szqhnyj4-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "Jundot build; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 5.9 GB; TTFT 1677 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2443.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--gemma-4-E4B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 2,443 PP tok/s 232.6 TG tok/s 1677 TTFT (ms) 5.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/szqhnyj4",
   "source_url": "https://omlx.ai/benchmarks/performance/szqhnyj4",
   "tg_tokens": null,
   "tps": 232.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-upg55afw-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "Jundot build; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 5.9 GB; TTFT 1679 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2440.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--gemma-4-E4B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 2,440 PP tok/s 92.7 TG tok/s 1679 TTFT (ms) 5.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/upg55afw",
   "source_url": "https://omlx.ai/benchmarks/performance/upg55afw",
   "tg_tokens": null,
   "tps": 92.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-z6vd5o82-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "Jundot build; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 5.9 GB; TTFT 1692 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2421.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--gemma-4-E4B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 2,421 PP tok/s 183.1 TG tok/s 1692 TTFT (ms) 5.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z6vd5o82",
   "source_url": "https://omlx.ai/benchmarks/performance/z6vd5o82",
   "tg_tokens": null,
   "tps": 183.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-38tr6752-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "Jundot build; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 4.9 GB; TTFT 1689 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2425.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--gemma-4-E4B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 2,425 PP tok/s 100.0 TG tok/s 1689 TTFT (ms) 4.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/38tr6752",
   "source_url": "https://omlx.ai/benchmarks/performance/38tr6752",
   "tg_tokens": null,
   "tps": 100.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-3vnl2kvu-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "Jundot build; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 4.9 GB; TTFT 1690 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2424.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--gemma-4-E4B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 2,424 PP tok/s 100.2 TG tok/s 1690 TTFT (ms) 4.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3vnl2kvu",
   "source_url": "https://omlx.ai/benchmarks/performance/3vnl2kvu",
   "tg_tokens": null,
   "tps": 100.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-yarbtli2-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "Jundot build; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 5.9 GB; TTFT 1656 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2473.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--gemma-4-E4B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 2,473 PP tok/s 238.1 TG tok/s 1656 TTFT (ms) 5.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yarbtli2",
   "source_url": "https://omlx.ai/benchmarks/performance/yarbtli2",
   "tg_tokens": null,
   "tps": 238.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-gixnnl8l-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "Jundot build; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 4.9 GB; TTFT 1650 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2483.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--gemma-4-E4B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 2,483 PP tok/s 102.5 TG tok/s 1650 TTFT (ms) 4.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gixnnl8l",
   "source_url": "https://omlx.ai/benchmarks/performance/gixnnl8l",
   "tg_tokens": null,
   "tps": 102.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ijhptvlw-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "Jundot build; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 4.9 GB; TTFT 1672 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2449.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--gemma-4-E4B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 2,449 PP tok/s 100.3 TG tok/s 1672 TTFT (ms) 4.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ijhptvlw",
   "source_url": "https://omlx.ai/benchmarks/performance/ijhptvlw",
   "tg_tokens": null,
   "tps": 100.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-pomc78tk-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 10.3 GB at this context; duplicate traces of the same session: s858ejmx; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 695.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B-MLX-8bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 8bit; 4k 695.7 13.9 10.3 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pomc78tk",
   "source_url": "https://omlx.ai/benchmarks/performance/pomc78tk",
   "tg_tokens": null,
   "tps": 13.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-pomc78tk-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 10.7 GB at this context; duplicate traces of the same session: s858ejmx; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 690.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B-MLX-8bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 8bit; 16k 690.9 13.2 10.7 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pomc78tk",
   "source_url": "https://omlx.ai/benchmarks/performance/pomc78tk",
   "tg_tokens": null,
   "tps": 13.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-qpc9rgot-ctx16384",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 13.8 GB; TTFT 42971 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 381.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 16k tokens 381.3 PP tok/s 29.1 TG tok/s 42971 TTFT (ms) 13.8 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qpc9rgot",
   "source_url": "https://omlx.ai/benchmarks/performance/qpc9rgot",
   "tg_tokens": null,
   "tps": 29.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-l8i1uy64-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1379.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.6-35B-A3B-mxfp4 on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 1,379 89.5 19.6 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l8i1uy64",
   "source_url": "https://omlx.ai/benchmarks/performance/l8i1uy64",
   "tg_tokens": null,
   "tps": 89.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-l8i1uy64-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1610.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.6-35B-A3B-mxfp4 on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 1,610 82.1 20.3 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l8i1uy64",
   "source_url": "https://omlx.ai/benchmarks/performance/l8i1uy64",
   "tg_tokens": null,
   "tps": 82.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-illnf07b-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 16.0 GB; TTFT 6238 ms; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 164.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 1k tokens 164.1 PP tok/s 25.2 TG tok/s 6238 TTFT (ms) 16.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/illnf07b",
   "source_url": "https://omlx.ai/benchmarks/performance/illnf07b",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-yhf027pg-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 256.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 256.9 28.6 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yhf027pg",
   "source_url": "https://omlx.ai/benchmarks/performance/yhf027pg",
   "tg_tokens": null,
   "tps": 28.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-yhf027pg-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 270.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 270.4 28.2 19.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yhf027pg",
   "source_url": "https://omlx.ai/benchmarks/performance/yhf027pg",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-goy771u3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.2 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 209.4,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 209.4 15.6 16.2 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/goy771u3",
   "source_url": "https://omlx.ai/benchmarks/performance/goy771u3",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-goy771u3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 212.3,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 212.3 18.9 17.3 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/goy771u3",
   "source_url": "https://omlx.ai/benchmarks/performance/goy771u3",
   "tg_tokens": null,
   "tps": 18.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-1ktytvl9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 19.8 GB; TTFT 17910 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 228.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k tokens 228.7 PP tok/s 27.6 TG tok/s 17910 TTFT (ms) 19.8 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ktytvl9",
   "source_url": "https://omlx.ai/benchmarks/performance/1ktytvl9",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hw7xpbof-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 747.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 747.2 43.9 17.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hw7xpbof",
   "source_url": "https://omlx.ai/benchmarks/performance/hw7xpbof",
   "tg_tokens": null,
   "tps": 43.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hw7xpbof-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 926.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 926.6 50.6 18.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hw7xpbof",
   "source_url": "https://omlx.ai/benchmarks/performance/hw7xpbof",
   "tg_tokens": null,
   "tps": 50.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-g7r3eg8u-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MLX; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 394.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 394.5 69.5 99.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g7r3eg8u",
   "source_url": "https://omlx.ai/benchmarks/performance/g7r3eg8u",
   "tg_tokens": null,
   "tps": 69.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-g7r3eg8u-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MLX; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1114.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,114 63.2 99.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g7r3eg8u",
   "source_url": "https://omlx.ai/benchmarks/performance/g7r3eg8u",
   "tg_tokens": null,
   "tps": 63.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-g7r3eg8u-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MLX; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2010.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,010 64.9 100.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g7r3eg8u",
   "source_url": "https://omlx.ai/benchmarks/performance/g7r3eg8u",
   "tg_tokens": null,
   "tps": 64.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ws8udebd-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MLX; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 56.2 GB; TTFT 43065 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 760.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k tokens 760.9 PP tok/s 9.0 TG tok/s 43065 TTFT (ms) 56.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ws8udebd",
   "source_url": "https://omlx.ai/benchmarks/performance/ws8udebd",
   "tg_tokens": null,
   "tps": 9.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ztj5ikiu-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: edwvgzdj, lxymgxiz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 598.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 598.1 42.9 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ztj5ikiu",
   "source_url": "https://omlx.ai/benchmarks/performance/ztj5ikiu",
   "tg_tokens": null,
   "tps": 42.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ztj5ikiu-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: edwvgzdj, lxymgxiz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 707.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 707.7 56.5 71.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ztj5ikiu",
   "source_url": "https://omlx.ai/benchmarks/performance/ztj5ikiu",
   "tg_tokens": null,
   "tps": 56.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ztj5ikiu-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.4 GB at this context; duplicate traces of the same session: edwvgzdj, lxymgxiz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 809.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 809.4 57.1 72.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ztj5ikiu",
   "source_url": "https://omlx.ai/benchmarks/performance/ztj5ikiu",
   "tg_tokens": null,
   "tps": 57.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-70s69s0v-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.3 GB at this context; duplicate traces of the same session: t51zula0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 498.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 498.2 56.3 69.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/70s69s0v",
   "source_url": "https://omlx.ai/benchmarks/performance/70s69s0v",
   "tg_tokens": null,
   "tps": 56.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-70s69s0v-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.9 GB at this context; duplicate traces of the same session: t51zula0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 621.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 621.6 52.4 69.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/70s69s0v",
   "source_url": "https://omlx.ai/benchmarks/performance/70s69s0v",
   "tg_tokens": null,
   "tps": 52.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-70s69s0v-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.9 GB at this context; duplicate traces of the same session: t51zula0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 842.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 842.1 54.8 69.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/70s69s0v",
   "source_url": "https://omlx.ai/benchmarks/performance/70s69s0v",
   "tg_tokens": null,
   "tps": 54.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jqhag980-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 919.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 919.3 53.9 70.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jqhag980",
   "source_url": "https://omlx.ai/benchmarks/performance/jqhag980",
   "tg_tokens": null,
   "tps": 53.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jqhag980-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1819.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,819 76.3 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jqhag980",
   "source_url": "https://omlx.ai/benchmarks/performance/jqhag980",
   "tg_tokens": null,
   "tps": 76.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jqhag980-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2646.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,646 70.9 74.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jqhag980",
   "source_url": "https://omlx.ai/benchmarks/performance/jqhag980",
   "tg_tokens": null,
   "tps": 70.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jqhag980-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2547.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,547 71.1 79.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jqhag980",
   "source_url": "https://omlx.ai/benchmarks/performance/jqhag980",
   "tg_tokens": null,
   "tps": 71.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jqhag980-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2300.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,300 56.5 80.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jqhag980",
   "source_url": "https://omlx.ai/benchmarks/performance/jqhag980",
   "tg_tokens": null,
   "tps": 56.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zngxxnho-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1158.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,158 63.6 70.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zngxxnho",
   "source_url": "https://omlx.ai/benchmarks/performance/zngxxnho",
   "tg_tokens": null,
   "tps": 63.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zngxxnho-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1981.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,981 52.9 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zngxxnho",
   "source_url": "https://omlx.ai/benchmarks/performance/zngxxnho",
   "tg_tokens": null,
   "tps": 52.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-a57hc7rx-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.4 GB at this context; duplicate traces of the same session: 2gu1ip7s, sgkzlmux; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 246.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 246.3 38.7 32.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a57hc7rx",
   "source_url": "https://omlx.ai/benchmarks/performance/a57hc7rx",
   "tg_tokens": null,
   "tps": 38.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-a57hc7rx-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.6 GB at this context; duplicate traces of the same session: 2gu1ip7s, sgkzlmux; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 276.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 276.6 43.3 35.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a57hc7rx",
   "source_url": "https://omlx.ai/benchmarks/performance/a57hc7rx",
   "tg_tokens": null,
   "tps": 43.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-a57hc7rx-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: 2gu1ip7s, sgkzlmux; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 296.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 8k 296.5 41.0 36.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a57hc7rx",
   "source_url": "https://omlx.ai/benchmarks/performance/a57hc7rx",
   "tg_tokens": null,
   "tps": 41.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-a57hc7rx-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.0 GB at this context; duplicate traces of the same session: 2gu1ip7s, sgkzlmux; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 292.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k 292.6 35.0 37.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a57hc7rx",
   "source_url": "https://omlx.ai/benchmarks/performance/a57hc7rx",
   "tg_tokens": null,
   "tps": 35.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-2ex1jhzq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace, single run at 4k context; peak memory 17.2 GB; TTFT 18152 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 225.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k tokens 225.6 PP tok/s 20.5 TG tok/s 18152 TTFT (ms) 17.2 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ex1jhzq",
   "source_url": "https://omlx.ai/benchmarks/performance/2ex1jhzq",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-efwmhmw0-ctx8192",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 14.8 GB; TTFT 36434 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.8,
   "provenance": "community",
   "quant": "oQ2",
   "quote": "Ternary-Bonsai-2-27B-MLX-oQ2-mtp on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 2bit; 8k tokens 224.8 PP tok/s 25.5 TG tok/s 36434 TTFT (ms) 14.8 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/efwmhmw0",
   "source_url": "https://omlx.ai/benchmarks/performance/efwmhmw0",
   "tg_tokens": null,
   "tps": 25.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-a8v1d6fp-ctx8192",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 13.5 GB; TTFT 37590 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 217.9,
   "provenance": "community",
   "quant": "oQ2",
   "quote": "Ternary-Bonsai-2-27B-MLX-oQ2-mtp on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 2bit; 8k tokens 217.9 PP tok/s 26.4 TG tok/s 37590 TTFT (ms) 13.5 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a8v1d6fp",
   "source_url": "https://omlx.ai/benchmarks/performance/a8v1d6fp",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-04i25da0-ctx8192",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 13.1 GB; TTFT 35210 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 232.7,
   "provenance": "community",
   "quant": "oQ2",
   "quote": "Ternary-Bonsai-2-27B-MLX-oQ2-mtp on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 2bit; 8k tokens 232.7 PP tok/s 29.1 TG tok/s 35210 TTFT (ms) 13.1 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/04i25da0",
   "source_url": "https://omlx.ai/benchmarks/performance/04i25da0",
   "tg_tokens": null,
   "tps": 29.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-11iod53h-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: vh51p7pw; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 652.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 1k 652.2 70.6 14.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/11iod53h",
   "source_url": "https://omlx.ai/benchmarks/performance/11iod53h",
   "tg_tokens": null,
   "tps": 70.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-11iod53h-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: vh51p7pw; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 588.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 4k 588.7 60.0 14.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/11iod53h",
   "source_url": "https://omlx.ai/benchmarks/performance/11iod53h",
   "tg_tokens": null,
   "tps": 60.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-11iod53h-ctx8192",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: vh51p7pw; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 485.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 8k 485.0 58.1 14.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/11iod53h",
   "source_url": "https://omlx.ai/benchmarks/performance/11iod53h",
   "tg_tokens": null,
   "tps": 58.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2lz5604k-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.1 GB at this context; duplicate traces of the same session: 9wtp68d1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1712.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 1,712 112.0 14.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2lz5604k",
   "source_url": "https://omlx.ai/benchmarks/performance/2lz5604k",
   "tg_tokens": null,
   "tps": 112.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2lz5604k-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: 9wtp68d1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1727.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 1,727 110.8 14.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2lz5604k",
   "source_url": "https://omlx.ai/benchmarks/performance/2lz5604k",
   "tg_tokens": null,
   "tps": 110.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2lz5604k-ctx8192",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: 9wtp68d1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1806.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 1,806 103.8 14.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2lz5604k",
   "source_url": "https://omlx.ai/benchmarks/performance/2lz5604k",
   "tg_tokens": null,
   "tps": 103.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-g01ep0et-ctx4096",
   "model": "gemma-4-26B-A4B-it-qat",
   "notes": "unsloth; oMLX community benchmark trace, single run at 4k context; peak memory 14.7 GB; TTFT 12744 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 321.4,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "unsloth-gemma-4-26B-A4B-it-qat-oQ4 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k tokens 321.4 PP tok/s 30.8 TG tok/s 12744 TTFT (ms) 14.7 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g01ep0et",
   "source_url": "https://omlx.ai/benchmarks/performance/g01ep0et",
   "tg_tokens": null,
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0fk3zwwr-ctx1024",
   "model": "DeepSeek-V4-Flash-0731-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 100.9 GB at this context; duplicate traces of the same session: uogwy2cq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 469.1,
   "provenance": "community",
   "quant": "Mixed 2bit-3bit g64",
   "quote": "DeepSeek-V4-Flash-0731-Abliterated-MLX-Mixed-2bit-3bit-g64 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 1k 469.1 28.0 100.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0fk3zwwr",
   "source_url": "https://omlx.ai/benchmarks/performance/0fk3zwwr",
   "tg_tokens": null,
   "tps": 28.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0fk3zwwr-ctx4096",
   "model": "DeepSeek-V4-Flash-0731-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: uogwy2cq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 476.8,
   "provenance": "community",
   "quant": "Mixed 2bit-3bit g64",
   "quote": "DeepSeek-V4-Flash-0731-Abliterated-MLX-Mixed-2bit-3bit-g64 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 4k 476.8 25.8 101.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0fk3zwwr",
   "source_url": "https://omlx.ai/benchmarks/performance/0fk3zwwr",
   "tg_tokens": null,
   "tps": 25.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0fk3zwwr-ctx8192",
   "model": "DeepSeek-V4-Flash-0731-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: uogwy2cq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 470.7,
   "provenance": "community",
   "quant": "Mixed 2bit-3bit g64",
   "quote": "DeepSeek-V4-Flash-0731-Abliterated-MLX-Mixed-2bit-3bit-g64 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 8k 470.7 25.1 101.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0fk3zwwr",
   "source_url": "https://omlx.ai/benchmarks/performance/0fk3zwwr",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0fk3zwwr-ctx16384",
   "model": "DeepSeek-V4-Flash-0731-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 101.4 GB at this context; duplicate traces of the same session: uogwy2cq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 462.1,
   "provenance": "community",
   "quant": "Mixed 2bit-3bit g64",
   "quote": "DeepSeek-V4-Flash-0731-Abliterated-MLX-Mixed-2bit-3bit-g64 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 16k 462.1 24.3 101.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0fk3zwwr",
   "source_url": "https://omlx.ai/benchmarks/performance/0fk3zwwr",
   "tg_tokens": null,
   "tps": 24.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0fk3zwwr-ctx32768",
   "model": "DeepSeek-V4-Flash-0731-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: uogwy2cq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 448.6,
   "provenance": "community",
   "quant": "Mixed 2bit-3bit g64",
   "quote": "DeepSeek-V4-Flash-0731-Abliterated-MLX-Mixed-2bit-3bit-g64 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 32k 448.6 23.0 101.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0fk3zwwr",
   "source_url": "https://omlx.ai/benchmarks/performance/0fk3zwwr",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0fk3zwwr-ctx65536",
   "model": "DeepSeek-V4-Flash-0731-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 101.8 GB at this context; duplicate traces of the same session: uogwy2cq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 424.9,
   "provenance": "community",
   "quant": "Mixed 2bit-3bit g64",
   "quote": "DeepSeek-V4-Flash-0731-Abliterated-MLX-Mixed-2bit-3bit-g64 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 64k 424.9 22.8 101.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0fk3zwwr",
   "source_url": "https://omlx.ai/benchmarks/performance/0fk3zwwr",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wxee6jca-ctx1024",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 89.2 GB at this context; duplicate traces of the same session: hnmbbuyl, n02wdcqs, nlana977, wggfow5b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 383.1,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4-Flash-0731-oQ2e-mtp-OMLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 1k 383.1 31.9 89.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wxee6jca",
   "source_url": "https://omlx.ai/benchmarks/performance/wxee6jca",
   "tg_tokens": null,
   "tps": 31.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wxee6jca-ctx4096",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 89.7 GB at this context; duplicate traces of the same session: hnmbbuyl, n02wdcqs, nlana977, wggfow5b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 670.9,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4-Flash-0731-oQ2e-mtp-OMLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 4k 670.9 28.9 89.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wxee6jca",
   "source_url": "https://omlx.ai/benchmarks/performance/wxee6jca",
   "tg_tokens": null,
   "tps": 28.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wxee6jca-ctx8192",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 89.7 GB at this context; duplicate traces of the same session: hnmbbuyl, n02wdcqs, nlana977, wggfow5b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 719.8,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4-Flash-0731-oQ2e-mtp-OMLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 8k 719.8 26.2 89.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wxee6jca",
   "source_url": "https://omlx.ai/benchmarks/performance/wxee6jca",
   "tg_tokens": null,
   "tps": 26.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wxee6jca-ctx16384",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 89.8 GB at this context; duplicate traces of the same session: hnmbbuyl, n02wdcqs, nlana977, wggfow5b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 527.2,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4-Flash-0731-oQ2e-mtp-OMLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 16k 527.2 23.8 89.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wxee6jca",
   "source_url": "https://omlx.ai/benchmarks/performance/wxee6jca",
   "tg_tokens": null,
   "tps": 23.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wxee6jca-ctx32768",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 89.9 GB at this context; duplicate traces of the same session: hnmbbuyl, n02wdcqs, nlana977, wggfow5b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 327.9,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4-Flash-0731-oQ2e-mtp-OMLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 32k 327.9 24.0 89.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wxee6jca",
   "source_url": "https://omlx.ai/benchmarks/performance/wxee6jca",
   "tg_tokens": null,
   "tps": 24.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-6snphjc1-ctx1024",
   "model": "Gemma-4-Dark-Gemistry-31B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 33.7 GB at this context; duplicate traces of the same session: 78ms2a10; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 92.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Gemma-4-Dark-Gemistry-31B-mlx-8Bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 1k 92.3 8.5 33.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6snphjc1",
   "source_url": "https://omlx.ai/benchmarks/performance/6snphjc1",
   "tg_tokens": null,
   "tps": 8.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-6snphjc1-ctx4096",
   "model": "Gemma-4-Dark-Gemistry-31B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 36.8 GB at this context; duplicate traces of the same session: 78ms2a10; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 82.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Gemma-4-Dark-Gemistry-31B-mlx-8Bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 4k 82.8 7.7 36.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6snphjc1",
   "source_url": "https://omlx.ai/benchmarks/performance/6snphjc1",
   "tg_tokens": null,
   "tps": 7.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-6snphjc1-ctx8192",
   "model": "Gemma-4-Dark-Gemistry-31B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 40.4 GB at this context; duplicate traces of the same session: 78ms2a10; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 79.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Gemma-4-Dark-Gemistry-31B-mlx-8Bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 8k 79.9 7.4 40.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6snphjc1",
   "source_url": "https://omlx.ai/benchmarks/performance/6snphjc1",
   "tg_tokens": null,
   "tps": 7.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-uupj9sbk-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "Jundot build; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 4.9 GB; TTFT 1673 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2448.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--gemma-4-E4B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 2,448 PP tok/s 102.8 TG tok/s 1673 TTFT (ms) 4.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uupj9sbk",
   "source_url": "https://omlx.ai/benchmarks/performance/uupj9sbk",
   "tg_tokens": null,
   "tps": 102.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-a1en90k8-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "Jundot build; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 5.9 GB; TTFT 1666 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2459.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--gemma-4-E4B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 2,459 PP tok/s 41.6 TG tok/s 1666 TTFT (ms) 5.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a1en90k8",
   "source_url": "https://omlx.ai/benchmarks/performance/a1en90k8",
   "tg_tokens": null,
   "tps": 41.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-y18rgjg0-ctx1024",
   "model": "LFM2.5-8B-A1B",
   "notes": "gptq jang q4 MLX; oMLX community benchmark trace, single run at 1k context; peak memory 5.4 GB; TTFT 634 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1614.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-8B-A1B-gptq-mlx-jang-q4 on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k tokens 1,614 PP tok/s 146.4 TG tok/s 634 TTFT (ms) 5.4 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y18rgjg0",
   "source_url": "https://omlx.ai/benchmarks/performance/y18rgjg0",
   "tg_tokens": null,
   "tps": 146.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mbjm7agm-ctx4096",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: sx2prtez; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4315.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 4,315 173.1 19.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mbjm7agm",
   "source_url": "https://omlx.ai/benchmarks/performance/mbjm7agm",
   "tg_tokens": null,
   "tps": 173.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mbjm7agm-ctx16384",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: sx2prtez; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4106.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 4,106 154.9 19.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mbjm7agm",
   "source_url": "https://omlx.ai/benchmarks/performance/mbjm7agm",
   "tg_tokens": null,
   "tps": 154.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-s858ejmx-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 10.3 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 695.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B-MLX-8bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 8bit; 4k 695.7 13.9 10.3 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s858ejmx",
   "source_url": "https://omlx.ai/benchmarks/performance/s858ejmx",
   "tg_tokens": null,
   "tps": 13.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-s858ejmx-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 10.7 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 690.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B-MLX-8bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 8bit; 16k 690.9 13.2 10.7 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s858ejmx",
   "source_url": "https://omlx.ai/benchmarks/performance/s858ejmx",
   "tg_tokens": null,
   "tps": 13.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-3d1g4yky-ctx4096",
   "model": "Qwen3.6-14B-A3B-FableVibes",
   "notes": "MLX; oMLX community benchmark trace; peak memory 9.0 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 540.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-14B-A3B-FableVibes-mlx-q4 on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 540.2 45.2 9.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3d1g4yky",
   "source_url": "https://omlx.ai/benchmarks/performance/3d1g4yky",
   "tg_tokens": null,
   "tps": 45.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-3d1g4yky-ctx16384",
   "model": "Qwen3.6-14B-A3B-FableVibes",
   "notes": "MLX; oMLX community benchmark trace; peak memory 9.5 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 455.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-14B-A3B-FableVibes-mlx-q4 on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 16k 455.1 38.9 9.5 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3d1g4yky",
   "source_url": "https://omlx.ai/benchmarks/performance/3d1g4yky",
   "tg_tokens": null,
   "tps": 38.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-oz7nh4l0-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: i3i9x9xh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 924.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 924.9 33.4 18.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oz7nh4l0",
   "source_url": "https://omlx.ai/benchmarks/performance/oz7nh4l0",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-oz7nh4l0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: i3i9x9xh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 727.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 727.7 32.9 19.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oz7nh4l0",
   "source_url": "https://omlx.ai/benchmarks/performance/oz7nh4l0",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-o8tqkmyh-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 2owmwbut; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1088.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,088 33.5 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o8tqkmyh",
   "source_url": "https://omlx.ai/benchmarks/performance/o8tqkmyh",
   "tg_tokens": null,
   "tps": 33.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-o8tqkmyh-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 2owmwbut; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1453.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,453 32.9 18.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o8tqkmyh",
   "source_url": "https://omlx.ai/benchmarks/performance/o8tqkmyh",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dsatseyb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: lbjx0w6z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 871.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 871.9 33.5 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dsatseyb",
   "source_url": "https://omlx.ai/benchmarks/performance/dsatseyb",
   "tg_tokens": null,
   "tps": 33.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dsatseyb-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: lbjx0w6z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1002.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,002 32.9 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dsatseyb",
   "source_url": "https://omlx.ai/benchmarks/performance/dsatseyb",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5u9feoh8-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 33.7 GB at this context; duplicate traces of the same session: gvq291nh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 784.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ8e-1M on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 784.6 38.1 33.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5u9feoh8",
   "source_url": "https://omlx.ai/benchmarks/performance/5u9feoh8",
   "tg_tokens": null,
   "tps": 38.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5u9feoh8-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: gvq291nh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 757.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ8e-1M on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 757.4 36.3 34.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5u9feoh8",
   "source_url": "https://omlx.ai/benchmarks/performance/5u9feoh8",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-5zg3pq3x-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 0nebxx31, 2qjmtv25, 5efcsk45, k05d1bd9; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 102.5 11.0 18.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5zg3pq3x",
   "source_url": "https://omlx.ai/benchmarks/performance/5zg3pq3x",
   "tg_tokens": null,
   "tps": 11.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-5zg3pq3x-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 0nebxx31, 2qjmtv25, 5efcsk45, k05d1bd9; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 102.2 20.6 19.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5zg3pq3x",
   "source_url": "https://omlx.ai/benchmarks/performance/5zg3pq3x",
   "tg_tokens": null,
   "tps": 20.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-5zg3pq3x-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: 0nebxx31, 2qjmtv25, 5efcsk45, k05d1bd9; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 93.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 93.1 11.7 20.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5zg3pq3x",
   "source_url": "https://omlx.ai/benchmarks/performance/5zg3pq3x",
   "tg_tokens": null,
   "tps": 11.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-5zg3pq3x-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 0nebxx31, 2qjmtv25, 5efcsk45, k05d1bd9; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 89.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 89.5 18.1 20.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5zg3pq3x",
   "source_url": "https://omlx.ai/benchmarks/performance/5zg3pq3x",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-5zg3pq3x-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 0nebxx31, 2qjmtv25, 5efcsk45, k05d1bd9; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 86.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 86.4 14.4 21.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5zg3pq3x",
   "source_url": "https://omlx.ai/benchmarks/performance/5zg3pq3x",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-mr347kh4-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 33.8 GB; TTFT 703508 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 186.3,
   "provenance": "community",
   "quant": "oQ6e-fp16",
   "quote": "Qwen3.8-27B-oQ6e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 6bit; 128k tokens 186.3 PP tok/s 17.9 TG tok/s 703508 TTFT (ms) 33.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mr347kh4",
   "source_url": "https://omlx.ai/benchmarks/performance/mr347kh4",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-079bo9po-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 142.6 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1022.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 1k 1,022 84.7 142.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/079bo9po",
   "source_url": "https://omlx.ai/benchmarks/performance/079bo9po",
   "tg_tokens": null,
   "tps": 84.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-079bo9po-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1320.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 4k 1,320 79.6 143.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/079bo9po",
   "source_url": "https://omlx.ai/benchmarks/performance/079bo9po",
   "tg_tokens": null,
   "tps": 79.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-079bo9po-ctx199680",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 151.5 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1275.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 195k 1,275 64.6 151.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/079bo9po",
   "source_url": "https://omlx.ai/benchmarks/performance/079bo9po",
   "tg_tokens": null,
   "tps": 64.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-emjrvgqo-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: 189fxyou; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 244.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 1k 244.1 11.9 71.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/emjrvgqo",
   "source_url": "https://omlx.ai/benchmarks/performance/emjrvgqo",
   "tg_tokens": null,
   "tps": 11.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-emjrvgqo-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 81.6 GB at this context; duplicate traces of the same session: 189fxyou; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 510.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 16k 510.0 12.1 81.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/emjrvgqo",
   "source_url": "https://omlx.ai/benchmarks/performance/emjrvgqo",
   "tg_tokens": null,
   "tps": 12.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-edwvgzdj-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: lxymgxiz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 598.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 598.1 42.9 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/edwvgzdj",
   "source_url": "https://omlx.ai/benchmarks/performance/edwvgzdj",
   "tg_tokens": null,
   "tps": 42.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-edwvgzdj-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: lxymgxiz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 707.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 707.7 56.5 71.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/edwvgzdj",
   "source_url": "https://omlx.ai/benchmarks/performance/edwvgzdj",
   "tg_tokens": null,
   "tps": 56.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-edwvgzdj-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.4 GB at this context; duplicate traces of the same session: lxymgxiz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 809.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 809.4 57.1 72.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/edwvgzdj",
   "source_url": "https://omlx.ai/benchmarks/performance/edwvgzdj",
   "tg_tokens": null,
   "tps": 57.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-t51zula0-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.3 GB at this context; duplicate traces of the same session: dflgpon4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 498.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 498.2 56.3 69.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t51zula0",
   "source_url": "https://omlx.ai/benchmarks/performance/t51zula0",
   "tg_tokens": null,
   "tps": 56.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-t51zula0-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.9 GB at this context; duplicate traces of the same session: dflgpon4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 621.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 621.6 52.4 69.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t51zula0",
   "source_url": "https://omlx.ai/benchmarks/performance/t51zula0",
   "tg_tokens": null,
   "tps": 52.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-t51zula0-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.9 GB at this context; duplicate traces of the same session: dflgpon4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 842.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 842.1 54.8 69.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t51zula0",
   "source_url": "https://omlx.ai/benchmarks/performance/t51zula0",
   "tg_tokens": null,
   "tps": 54.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rmjeehip-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.6 GB; TTFT 853 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1200.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,200 PP tok/s 71.4 TG tok/s 853 TTFT (ms) 70.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rmjeehip",
   "source_url": "https://omlx.ai/benchmarks/performance/rmjeehip",
   "tg_tokens": null,
   "tps": 71.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-2gu1ip7s-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.4 GB at this context; duplicate traces of the same session: sgkzlmux; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 246.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 246.3 38.7 32.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2gu1ip7s",
   "source_url": "https://omlx.ai/benchmarks/performance/2gu1ip7s",
   "tg_tokens": null,
   "tps": 38.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-2gu1ip7s-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.6 GB at this context; duplicate traces of the same session: sgkzlmux; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 276.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 276.6 43.3 35.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2gu1ip7s",
   "source_url": "https://omlx.ai/benchmarks/performance/2gu1ip7s",
   "tg_tokens": null,
   "tps": 43.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-2gu1ip7s-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: sgkzlmux; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 296.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 8k 296.5 41.0 36.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2gu1ip7s",
   "source_url": "https://omlx.ai/benchmarks/performance/2gu1ip7s",
   "tg_tokens": null,
   "tps": 41.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-2gu1ip7s-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.0 GB at this context; duplicate traces of the same session: sgkzlmux; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 292.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k 292.6 35.0 37.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2gu1ip7s",
   "source_url": "https://omlx.ai/benchmarks/performance/2gu1ip7s",
   "tg_tokens": null,
   "tps": 35.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-vh51p7pw-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 652.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 1k 652.2 70.6 14.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vh51p7pw",
   "source_url": "https://omlx.ai/benchmarks/performance/vh51p7pw",
   "tg_tokens": null,
   "tps": 70.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-vh51p7pw-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 588.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 4k 588.7 60.0 14.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vh51p7pw",
   "source_url": "https://omlx.ai/benchmarks/performance/vh51p7pw",
   "tg_tokens": null,
   "tps": 60.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-vh51p7pw-ctx8192",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 485.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 8k 485.0 58.1 14.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vh51p7pw",
   "source_url": "https://omlx.ai/benchmarks/performance/vh51p7pw",
   "tg_tokens": null,
   "tps": 58.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9wtp68d1-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.1 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1712.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 1,712 112.0 14.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9wtp68d1",
   "source_url": "https://omlx.ai/benchmarks/performance/9wtp68d1",
   "tg_tokens": null,
   "tps": 112.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9wtp68d1-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1727.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 1,727 110.8 14.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9wtp68d1",
   "source_url": "https://omlx.ai/benchmarks/performance/9wtp68d1",
   "tg_tokens": null,
   "tps": 110.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9wtp68d1-ctx8192",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1806.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 1,806 103.8 14.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9wtp68d1",
   "source_url": "https://omlx.ai/benchmarks/performance/9wtp68d1",
   "tg_tokens": null,
   "tps": 103.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-uogwy2cq-ctx1024",
   "model": "DeepSeek-V4-Flash-0731-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 100.9 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 469.1,
   "provenance": "community",
   "quant": "Mixed 2bit-3bit g64",
   "quote": "DeepSeek-V4-Flash-0731-Abliterated-MLX-Mixed-2bit-3bit-g64 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 1k 469.1 28.0 100.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uogwy2cq",
   "source_url": "https://omlx.ai/benchmarks/performance/uogwy2cq",
   "tg_tokens": null,
   "tps": 28.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-uogwy2cq-ctx4096",
   "model": "DeepSeek-V4-Flash-0731-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 476.8,
   "provenance": "community",
   "quant": "Mixed 2bit-3bit g64",
   "quote": "DeepSeek-V4-Flash-0731-Abliterated-MLX-Mixed-2bit-3bit-g64 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 4k 476.8 25.8 101.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uogwy2cq",
   "source_url": "https://omlx.ai/benchmarks/performance/uogwy2cq",
   "tg_tokens": null,
   "tps": 25.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-uogwy2cq-ctx8192",
   "model": "DeepSeek-V4-Flash-0731-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 470.7,
   "provenance": "community",
   "quant": "Mixed 2bit-3bit g64",
   "quote": "DeepSeek-V4-Flash-0731-Abliterated-MLX-Mixed-2bit-3bit-g64 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 8k 470.7 25.1 101.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uogwy2cq",
   "source_url": "https://omlx.ai/benchmarks/performance/uogwy2cq",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-uogwy2cq-ctx16384",
   "model": "DeepSeek-V4-Flash-0731-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 101.4 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 462.1,
   "provenance": "community",
   "quant": "Mixed 2bit-3bit g64",
   "quote": "DeepSeek-V4-Flash-0731-Abliterated-MLX-Mixed-2bit-3bit-g64 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 16k 462.1 24.3 101.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uogwy2cq",
   "source_url": "https://omlx.ai/benchmarks/performance/uogwy2cq",
   "tg_tokens": null,
   "tps": 24.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-uogwy2cq-ctx32768",
   "model": "DeepSeek-V4-Flash-0731-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 448.6,
   "provenance": "community",
   "quant": "Mixed 2bit-3bit g64",
   "quote": "DeepSeek-V4-Flash-0731-Abliterated-MLX-Mixed-2bit-3bit-g64 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 32k 448.6 23.0 101.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uogwy2cq",
   "source_url": "https://omlx.ai/benchmarks/performance/uogwy2cq",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-uogwy2cq-ctx65536",
   "model": "DeepSeek-V4-Flash-0731-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 101.8 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 424.9,
   "provenance": "community",
   "quant": "Mixed 2bit-3bit g64",
   "quote": "DeepSeek-V4-Flash-0731-Abliterated-MLX-Mixed-2bit-3bit-g64 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 64k 424.9 22.8 101.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uogwy2cq",
   "source_url": "https://omlx.ai/benchmarks/performance/uogwy2cq",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n02wdcqs-ctx1024",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 89.2 GB at this context; duplicate traces of the same session: hnmbbuyl, nlana977, wggfow5b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 383.1,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4-Flash-0731-oQ2e-mtp-OMLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 1k 383.1 31.9 89.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n02wdcqs",
   "source_url": "https://omlx.ai/benchmarks/performance/n02wdcqs",
   "tg_tokens": null,
   "tps": 31.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n02wdcqs-ctx4096",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 89.7 GB at this context; duplicate traces of the same session: hnmbbuyl, nlana977, wggfow5b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 670.9,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4-Flash-0731-oQ2e-mtp-OMLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 4k 670.9 28.9 89.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n02wdcqs",
   "source_url": "https://omlx.ai/benchmarks/performance/n02wdcqs",
   "tg_tokens": null,
   "tps": 28.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n02wdcqs-ctx8192",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 89.7 GB at this context; duplicate traces of the same session: hnmbbuyl, nlana977, wggfow5b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 719.8,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4-Flash-0731-oQ2e-mtp-OMLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 8k 719.8 26.2 89.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n02wdcqs",
   "source_url": "https://omlx.ai/benchmarks/performance/n02wdcqs",
   "tg_tokens": null,
   "tps": 26.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n02wdcqs-ctx16384",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 89.8 GB at this context; duplicate traces of the same session: hnmbbuyl, nlana977, wggfow5b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 527.2,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4-Flash-0731-oQ2e-mtp-OMLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 16k 527.2 23.8 89.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n02wdcqs",
   "source_url": "https://omlx.ai/benchmarks/performance/n02wdcqs",
   "tg_tokens": null,
   "tps": 23.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-n02wdcqs-ctx32768",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 89.9 GB at this context; duplicate traces of the same session: hnmbbuyl, nlana977, wggfow5b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 327.9,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4-Flash-0731-oQ2e-mtp-OMLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 32k 327.9 24.0 89.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n02wdcqs",
   "source_url": "https://omlx.ai/benchmarks/performance/n02wdcqs",
   "tg_tokens": null,
   "tps": 24.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-78ms2a10-ctx1024",
   "model": "Gemma-4-Dark-Gemistry-31B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 33.7 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 92.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Gemma-4-Dark-Gemistry-31B-mlx-8Bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 1k 92.3 8.5 33.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/78ms2a10",
   "source_url": "https://omlx.ai/benchmarks/performance/78ms2a10",
   "tg_tokens": null,
   "tps": 8.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-78ms2a10-ctx4096",
   "model": "Gemma-4-Dark-Gemistry-31B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 36.8 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 82.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Gemma-4-Dark-Gemistry-31B-mlx-8Bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 4k 82.8 7.7 36.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/78ms2a10",
   "source_url": "https://omlx.ai/benchmarks/performance/78ms2a10",
   "tg_tokens": null,
   "tps": 7.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-78ms2a10-ctx8192",
   "model": "Gemma-4-Dark-Gemistry-31B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 40.4 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 79.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Gemma-4-Dark-Gemistry-31B-mlx-8Bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 8k 79.9 7.4 40.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/78ms2a10",
   "source_url": "https://omlx.ai/benchmarks/performance/78ms2a10",
   "tg_tokens": null,
   "tps": 7.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-mqlxk09q-ctx1024",
   "model": "Ling-3.0-tiny",
   "notes": "bf16; MLX; oMLX community benchmark trace, single run at 1k context; peak memory 8.7 GB; TTFT 530 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1931.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ling-3.0-tiny-oQ8e-bf16-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 8bit; 1k tokens 1,931 PP tok/s 116.9 TG tok/s 530 TTFT (ms) 8.7 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mqlxk09q",
   "source_url": "https://omlx.ai/benchmarks/performance/mqlxk09q",
   "tg_tokens": null,
   "tps": 116.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-sx2prtez-ctx4096",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4315.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 4,315 173.1 19.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sx2prtez",
   "source_url": "https://omlx.ai/benchmarks/performance/sx2prtez",
   "tg_tokens": null,
   "tps": 173.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-sx2prtez-ctx16384",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4106.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 4,106 154.9 19.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sx2prtez",
   "source_url": "https://omlx.ai/benchmarks/performance/sx2prtez",
   "tg_tokens": null,
   "tps": 154.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lff46mox-ctx4096",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 9bw94oip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1800.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,800 172.9 19.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lff46mox",
   "source_url": "https://omlx.ai/benchmarks/performance/lff46mox",
   "tg_tokens": null,
   "tps": 172.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lff46mox-ctx16384",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 9bw94oip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4080.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 4,080 159.7 19.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lff46mox",
   "source_url": "https://omlx.ai/benchmarks/performance/lff46mox",
   "tg_tokens": null,
   "tps": 159.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-0lsg509l-ctx4096",
   "model": "Qwen3.6-14B-A3B-FableVibes",
   "notes": "MLX; oMLX community benchmark trace; peak memory 9.0 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 540.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-14B-A3B-FableVibes-mlx-q4 on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 540.2 45.2 9.0 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0lsg509l",
   "source_url": "https://omlx.ai/benchmarks/performance/0lsg509l",
   "tg_tokens": null,
   "tps": 45.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-0lsg509l-ctx16384",
   "model": "Qwen3.6-14B-A3B-FableVibes",
   "notes": "MLX; oMLX community benchmark trace; peak memory 9.5 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 455.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-14B-A3B-FableVibes-mlx-q4 on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 16k 455.1 38.9 9.5 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0lsg509l",
   "source_url": "https://omlx.ai/benchmarks/performance/0lsg509l",
   "tg_tokens": null,
   "tps": 38.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-wwrj60z0-ctx16384",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 15.7 GB; TTFT 43294 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 378.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 16k tokens 378.4 PP tok/s 33.9 TG tok/s 43294 TTFT (ms) 15.7 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wwrj60z0",
   "source_url": "https://omlx.ai/benchmarks/performance/wwrj60z0",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-zfsxzb68-ctx16384",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 13.8 GB; TTFT 43860 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 373.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 16k tokens 373.6 PP tok/s 28.6 TG tok/s 43860 TTFT (ms) 13.8 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zfsxzb68",
   "source_url": "https://omlx.ai/benchmarks/performance/zfsxzb68",
   "tg_tokens": null,
   "tps": 28.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-eekqyujz-ctx16384",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 13.8 GB; TTFT 43392 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 377.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 16k tokens 377.6 PP tok/s 28.7 TG tok/s 43392 TTFT (ms) 13.8 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eekqyujz",
   "source_url": "https://omlx.ai/benchmarks/performance/eekqyujz",
   "tg_tokens": null,
   "tps": 28.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-r4kden5t-ctx1024",
   "model": "Qwen3.6-35B-MoE",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 7dls4ln6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2626.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-MoE-Q4 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,626 101.2 19.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r4kden5t",
   "source_url": "https://omlx.ai/benchmarks/performance/r4kden5t",
   "tg_tokens": null,
   "tps": 101.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-r4kden5t-ctx4096",
   "model": "Qwen3.6-35B-MoE",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: 7dls4ln6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3216.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-MoE-Q4 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,216 99.9 20.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r4kden5t",
   "source_url": "https://omlx.ai/benchmarks/performance/r4kden5t",
   "tg_tokens": null,
   "tps": 99.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i3i9x9xh-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 924.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 924.9 33.4 18.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i3i9x9xh",
   "source_url": "https://omlx.ai/benchmarks/performance/i3i9x9xh",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i3i9x9xh-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 727.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 727.7 32.9 19.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i3i9x9xh",
   "source_url": "https://omlx.ai/benchmarks/performance/i3i9x9xh",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2owmwbut-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1088.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,088 33.5 17.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2owmwbut",
   "source_url": "https://omlx.ai/benchmarks/performance/2owmwbut",
   "tg_tokens": null,
   "tps": 33.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2owmwbut-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1453.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,453 32.9 18.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2owmwbut",
   "source_url": "https://omlx.ai/benchmarks/performance/2owmwbut",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lbjx0w6z-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 871.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 871.9 33.5 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lbjx0w6z",
   "source_url": "https://omlx.ai/benchmarks/performance/lbjx0w6z",
   "tg_tokens": null,
   "tps": 33.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lbjx0w6z-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1002.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,002 32.9 17.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lbjx0w6z",
   "source_url": "https://omlx.ai/benchmarks/performance/lbjx0w6z",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-gvq291nh-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 33.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 784.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ8e-1M on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 784.6 38.1 33.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gvq291nh",
   "source_url": "https://omlx.ai/benchmarks/performance/gvq291nh",
   "tg_tokens": null,
   "tps": 38.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gvq291nh-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 757.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ8e-1M on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 757.4 36.3 34.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gvq291nh",
   "source_url": "https://omlx.ai/benchmarks/performance/gvq291nh",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-2m4w4sou-ctx1024",
   "model": "Qwen3.8-27B-heretic-ara",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: ezq70joo, n0y5aj87, qhm4h1a6; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 291.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-heretic-ara-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 1k 291.7 41.8 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2m4w4sou",
   "source_url": "https://omlx.ai/benchmarks/performance/2m4w4sou",
   "tg_tokens": null,
   "tps": 41.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-2m4w4sou-ctx4096",
   "model": "Qwen3.8-27B-heretic-ara",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: ezq70joo, n0y5aj87, qhm4h1a6; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 311.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-heretic-ara-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 4k 311.7 42.9 23.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2m4w4sou",
   "source_url": "https://omlx.ai/benchmarks/performance/2m4w4sou",
   "tg_tokens": null,
   "tps": 42.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-2m4w4sou-ctx8192",
   "model": "Qwen3.8-27B-heretic-ara",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: ezq70joo, n0y5aj87, qhm4h1a6; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 304.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-heretic-ara-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 8k 304.8 44.0 23.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2m4w4sou",
   "source_url": "https://omlx.ai/benchmarks/performance/2m4w4sou",
   "tg_tokens": null,
   "tps": 44.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-2m4w4sou-ctx16384",
   "model": "Qwen3.8-27B-heretic-ara",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: ezq70joo, n0y5aj87, qhm4h1a6; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 290.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-heretic-ara-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 16k 290.5 55.5 25.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2m4w4sou",
   "source_url": "https://omlx.ai/benchmarks/performance/2m4w4sou",
   "tg_tokens": null,
   "tps": 55.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-9w27hva6-ctx1024",
   "model": "Qwen3.8-27B-heretic-ara",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 08fy6xlj; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 294.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-heretic-ara-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 1k 294.5 30.4 16.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9w27hva6",
   "source_url": "https://omlx.ai/benchmarks/performance/9w27hva6",
   "tg_tokens": null,
   "tps": 30.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-9w27hva6-ctx4096",
   "model": "Qwen3.8-27B-heretic-ara",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 08fy6xlj; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-heretic-ara-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 4k 323.2 30.5 20.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9w27hva6",
   "source_url": "https://omlx.ai/benchmarks/performance/9w27hva6",
   "tg_tokens": null,
   "tps": 30.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-9w27hva6-ctx8192",
   "model": "Qwen3.8-27B-heretic-ara",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 08fy6xlj; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 316.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-heretic-ara-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 8k 316.2 29.8 20.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9w27hva6",
   "source_url": "https://omlx.ai/benchmarks/performance/9w27hva6",
   "tg_tokens": null,
   "tps": 29.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-9w27hva6-ctx16384",
   "model": "Qwen3.8-27B-heretic-ara",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: 08fy6xlj; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 302.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-heretic-ara-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 16k 302.0 28.5 22.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9w27hva6",
   "source_url": "https://omlx.ai/benchmarks/performance/9w27hva6",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-k05d1bd9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 0nebxx31, 2qjmtv25, 5efcsk45; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 102.5 11.0 18.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k05d1bd9",
   "source_url": "https://omlx.ai/benchmarks/performance/k05d1bd9",
   "tg_tokens": null,
   "tps": 11.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-k05d1bd9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 0nebxx31, 2qjmtv25, 5efcsk45; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 102.2 20.6 19.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k05d1bd9",
   "source_url": "https://omlx.ai/benchmarks/performance/k05d1bd9",
   "tg_tokens": null,
   "tps": 20.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-k05d1bd9-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: 0nebxx31, 2qjmtv25, 5efcsk45; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 93.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 93.1 11.7 20.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k05d1bd9",
   "source_url": "https://omlx.ai/benchmarks/performance/k05d1bd9",
   "tg_tokens": null,
   "tps": 11.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-k05d1bd9-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 0nebxx31, 2qjmtv25, 5efcsk45; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 89.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 89.5 18.1 20.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k05d1bd9",
   "source_url": "https://omlx.ai/benchmarks/performance/k05d1bd9",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-k05d1bd9-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 0nebxx31, 2qjmtv25, 5efcsk45; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 86.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 86.4 14.4 21.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k05d1bd9",
   "source_url": "https://omlx.ai/benchmarks/performance/k05d1bd9",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xfqusv1r-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: x5o921l3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 804.6,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-27B-oQ6 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 804.6 19.2 23.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xfqusv1r",
   "source_url": "https://omlx.ai/benchmarks/performance/xfqusv1r",
   "tg_tokens": null,
   "tps": 19.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xfqusv1r-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: x5o921l3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 552.3,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-27B-oQ6 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 552.3 17.4 24.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xfqusv1r",
   "source_url": "https://omlx.ai/benchmarks/performance/xfqusv1r",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-kosa0ibr-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 39.6 GB; TTFT 735704 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 178.2,
   "provenance": "community",
   "quant": "oQ6e-fp16",
   "quote": "Qwen3.8-27B-oQ6e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 6bit; 128k tokens 178.2 PP tok/s 17.7 TG tok/s 735704 TTFT (ms) 39.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kosa0ibr",
   "source_url": "https://omlx.ai/benchmarks/performance/kosa0ibr",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-sszl8hga-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 32.8 GB; TTFT 127664 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 256.7,
   "provenance": "community",
   "quant": "oQ6e-fp16",
   "quote": "Qwen3.8-27B-oQ6e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 6bit; 32k tokens 256.7 PP tok/s 29.0 TG tok/s 127664 TTFT (ms) 32.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sszl8hga",
   "source_url": "https://omlx.ai/benchmarks/performance/sszl8hga",
   "tg_tokens": null,
   "tps": 29.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ryle5fcb-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 142.6 GB at this context; duplicate traces of the same session: pqopb8gv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1022.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 1k 1,022 84.7 142.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ryle5fcb",
   "source_url": "https://omlx.ai/benchmarks/performance/ryle5fcb",
   "tg_tokens": null,
   "tps": 84.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ryle5fcb-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: pqopb8gv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1320.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 4k 1,320 79.6 143.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ryle5fcb",
   "source_url": "https://omlx.ai/benchmarks/performance/ryle5fcb",
   "tg_tokens": null,
   "tps": 79.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ryle5fcb-ctx199680",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 151.5 GB at this context; duplicate traces of the same session: pqopb8gv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1275.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 195k 1,275 64.6 151.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ryle5fcb",
   "source_url": "https://omlx.ai/benchmarks/performance/ryle5fcb",
   "tg_tokens": null,
   "tps": 64.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-189fxyou-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 244.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 1k 244.1 11.9 71.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/189fxyou",
   "source_url": "https://omlx.ai/benchmarks/performance/189fxyou",
   "tg_tokens": null,
   "tps": 11.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-189fxyou-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 81.6 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 510.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 16k 510.0 12.1 81.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/189fxyou",
   "source_url": "https://omlx.ai/benchmarks/performance/189fxyou",
   "tg_tokens": null,
   "tps": 12.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-lxymgxiz-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 598.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 598.1 42.9 70.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lxymgxiz",
   "source_url": "https://omlx.ai/benchmarks/performance/lxymgxiz",
   "tg_tokens": null,
   "tps": 42.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-lxymgxiz-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 707.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 707.7 56.5 71.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lxymgxiz",
   "source_url": "https://omlx.ai/benchmarks/performance/lxymgxiz",
   "tg_tokens": null,
   "tps": 56.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-lxymgxiz-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 809.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 809.4 57.1 72.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lxymgxiz",
   "source_url": "https://omlx.ai/benchmarks/performance/lxymgxiz",
   "tg_tokens": null,
   "tps": 57.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-dflgpon4-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 498.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 498.2 56.3 69.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dflgpon4",
   "source_url": "https://omlx.ai/benchmarks/performance/dflgpon4",
   "tg_tokens": null,
   "tps": 56.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-dflgpon4-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 621.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 621.6 52.4 69.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dflgpon4",
   "source_url": "https://omlx.ai/benchmarks/performance/dflgpon4",
   "tg_tokens": null,
   "tps": 52.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-dflgpon4-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 842.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 842.1 54.8 69.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dflgpon4",
   "source_url": "https://omlx.ai/benchmarks/performance/dflgpon4",
   "tg_tokens": null,
   "tps": 54.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-dsmecp9f-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: scbbltqj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 600.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 600.0 60.3 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dsmecp9f",
   "source_url": "https://omlx.ai/benchmarks/performance/dsmecp9f",
   "tg_tokens": null,
   "tps": 60.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-dsmecp9f-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: scbbltqj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 716.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 716.3 59.1 71.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dsmecp9f",
   "source_url": "https://omlx.ai/benchmarks/performance/dsmecp9f",
   "tg_tokens": null,
   "tps": 59.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-dsmecp9f-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: scbbltqj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 819.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 819.3 71.7 71.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dsmecp9f",
   "source_url": "https://omlx.ai/benchmarks/performance/dsmecp9f",
   "tg_tokens": null,
   "tps": 71.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-g5h1pb96-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.2 GB at this context; duplicate traces of the same session: xzns351a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1198.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,198 77.7 69.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g5h1pb96",
   "source_url": "https://omlx.ai/benchmarks/performance/g5h1pb96",
   "tg_tokens": null,
   "tps": 77.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-g5h1pb96-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.3 GB at this context; duplicate traces of the same session: xzns351a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2145.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,145 70.2 69.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g5h1pb96",
   "source_url": "https://omlx.ai/benchmarks/performance/g5h1pb96",
   "tg_tokens": null,
   "tps": 70.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-63l3b8n3-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 39.1 GB; TTFT 6994 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 146.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 146.4 PP tok/s 19.6 TG tok/s 6994 TTFT (ms) 39.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/63l3b8n3",
   "source_url": "https://omlx.ai/benchmarks/performance/63l3b8n3",
   "tg_tokens": null,
   "tps": 19.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wu5mp26h-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: n71bcawo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 480.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 480.0 33.5 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wu5mp26h",
   "source_url": "https://omlx.ai/benchmarks/performance/wu5mp26h",
   "tg_tokens": null,
   "tps": 33.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wu5mp26h-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: n71bcawo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1139.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,139 55.9 70.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wu5mp26h",
   "source_url": "https://omlx.ai/benchmarks/performance/wu5mp26h",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-sgkzlmux-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 246.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 246.3 38.7 32.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sgkzlmux",
   "source_url": "https://omlx.ai/benchmarks/performance/sgkzlmux",
   "tg_tokens": null,
   "tps": 38.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-sgkzlmux-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 276.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 276.6 43.3 35.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sgkzlmux",
   "source_url": "https://omlx.ai/benchmarks/performance/sgkzlmux",
   "tg_tokens": null,
   "tps": 43.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-sgkzlmux-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 296.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 8k 296.5 41.0 36.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sgkzlmux",
   "source_url": "https://omlx.ai/benchmarks/performance/sgkzlmux",
   "tg_tokens": null,
   "tps": 41.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-sgkzlmux-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 292.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k 292.6 35.0 37.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sgkzlmux",
   "source_url": "https://omlx.ai/benchmarks/performance/sgkzlmux",
   "tg_tokens": null,
   "tps": 35.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-msursqrw-ctx4096",
   "model": "pplx-computer-qwen-3-6-35b-a3b",
   "notes": "MLX build 20260709; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: s64baypc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1725.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "perplexity-ai--pplx-computer-qwen-3-6-35b-a3b-mlx-20260709 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,725 100.8 20.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/msursqrw",
   "source_url": "https://omlx.ai/benchmarks/performance/msursqrw",
   "tg_tokens": null,
   "tps": 100.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-msursqrw-ctx16384",
   "model": "pplx-computer-qwen-3-6-35b-a3b",
   "notes": "MLX build 20260709; oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: s64baypc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2829.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "perplexity-ai--pplx-computer-qwen-3-6-35b-a3b-mlx-20260709 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,829 104.2 20.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/msursqrw",
   "source_url": "https://omlx.ai/benchmarks/performance/msursqrw",
   "tg_tokens": null,
   "tps": 104.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nlana977-ctx1024",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 89.2 GB at this context; duplicate traces of the same session: hnmbbuyl, wggfow5b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 383.1,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4-Flash-0731-oQ2e-mtp-OMLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 1k 383.1 31.9 89.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nlana977",
   "source_url": "https://omlx.ai/benchmarks/performance/nlana977",
   "tg_tokens": null,
   "tps": 31.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nlana977-ctx4096",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 89.7 GB at this context; duplicate traces of the same session: hnmbbuyl, wggfow5b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 670.9,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4-Flash-0731-oQ2e-mtp-OMLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 4k 670.9 28.9 89.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nlana977",
   "source_url": "https://omlx.ai/benchmarks/performance/nlana977",
   "tg_tokens": null,
   "tps": 28.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nlana977-ctx8192",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 89.7 GB at this context; duplicate traces of the same session: hnmbbuyl, wggfow5b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 719.8,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4-Flash-0731-oQ2e-mtp-OMLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 8k 719.8 26.2 89.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nlana977",
   "source_url": "https://omlx.ai/benchmarks/performance/nlana977",
   "tg_tokens": null,
   "tps": 26.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nlana977-ctx16384",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 89.8 GB at this context; duplicate traces of the same session: hnmbbuyl, wggfow5b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 527.2,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4-Flash-0731-oQ2e-mtp-OMLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 16k 527.2 23.8 89.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nlana977",
   "source_url": "https://omlx.ai/benchmarks/performance/nlana977",
   "tg_tokens": null,
   "tps": 23.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-nlana977-ctx32768",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 89.9 GB at this context; duplicate traces of the same session: hnmbbuyl, wggfow5b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 327.9,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4-Flash-0731-oQ2e-mtp-OMLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 32k 327.9 24.0 89.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nlana977",
   "source_url": "https://omlx.ai/benchmarks/performance/nlana977",
   "tg_tokens": null,
   "tps": 24.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ztpbp1uw-ctx65536",
   "model": "Muse Glimmer",
   "notes": "oMLX community benchmark trace; peak memory 16.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 896.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-30B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 896.6 28.4 16.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ztpbp1uw",
   "source_url": "https://omlx.ai/benchmarks/performance/ztpbp1uw",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ztpbp1uw-ctx131072",
   "model": "Muse Glimmer",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 807.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-30B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 807.4 25.9 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ztpbp1uw",
   "source_url": "https://omlx.ai/benchmarks/performance/ztpbp1uw",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-gn0gp2q9-ctx1024",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: bc1amj33; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 645.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 645.8 93.3 18.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gn0gp2q9",
   "source_url": "https://omlx.ai/benchmarks/performance/gn0gp2q9",
   "tg_tokens": null,
   "tps": 93.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-gn0gp2q9-ctx4096",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: bc1amj33; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 680.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 680.3 91.5 19.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gn0gp2q9",
   "source_url": "https://omlx.ai/benchmarks/performance/gn0gp2q9",
   "tg_tokens": null,
   "tps": 91.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9bw94oip-ctx4096",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1800.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,800 172.9 19.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9bw94oip",
   "source_url": "https://omlx.ai/benchmarks/performance/9bw94oip",
   "tg_tokens": null,
   "tps": 172.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9bw94oip-ctx16384",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4080.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 4,080 159.7 19.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9bw94oip",
   "source_url": "https://omlx.ai/benchmarks/performance/9bw94oip",
   "tg_tokens": null,
   "tps": 159.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-veryy8tq-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "REAP-19B build; oMLX community benchmark trace; peak memory 13.8 GB at this context; duplicate traces of the same session: 5t5kbiyk, abitgt94; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 383.7,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit-REAP-19B on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 16k 383.7 29.5 13.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/veryy8tq",
   "source_url": "https://omlx.ai/benchmarks/performance/veryy8tq",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-veryy8tq-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "REAP-19B build; oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: 5t5kbiyk, abitgt94; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 309.4,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit-REAP-19B on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 32k 309.4 25.1 14.4 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/veryy8tq",
   "source_url": "https://omlx.ai/benchmarks/performance/veryy8tq",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-veryy8tq-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "REAP-19B build; oMLX community benchmark trace; peak memory 15.7 GB at this context; duplicate traces of the same session: 5t5kbiyk, abitgt94; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 203.4,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit-REAP-19B on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 64k 203.4 15.4 15.7 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/veryy8tq",
   "source_url": "https://omlx.ai/benchmarks/performance/veryy8tq",
   "tg_tokens": null,
   "tps": 15.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-207vhvkp-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: 1ycsoptk, 55o29pv1, bdb18he1, zzpzx7gg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 660.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 1k 660.6 72.5 29.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/207vhvkp",
   "source_url": "https://omlx.ai/benchmarks/performance/207vhvkp",
   "tg_tokens": null,
   "tps": 72.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-207vhvkp-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.8 GB at this context; duplicate traces of the same session: 1ycsoptk, 55o29pv1, bdb18he1, zzpzx7gg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 745.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 4k 745.6 76.5 30.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/207vhvkp",
   "source_url": "https://omlx.ai/benchmarks/performance/207vhvkp",
   "tg_tokens": null,
   "tps": 76.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-207vhvkp-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.1 GB at this context; duplicate traces of the same session: 1ycsoptk, 55o29pv1, bdb18he1, zzpzx7gg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 718.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 8k 718.9 68.5 31.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/207vhvkp",
   "source_url": "https://omlx.ai/benchmarks/performance/207vhvkp",
   "tg_tokens": null,
   "tps": 68.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-207vhvkp-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.2 GB at this context; duplicate traces of the same session: 1ycsoptk, 55o29pv1, bdb18he1, zzpzx7gg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 636.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 16k 636.8 55.0 31.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/207vhvkp",
   "source_url": "https://omlx.ai/benchmarks/performance/207vhvkp",
   "tg_tokens": null,
   "tps": 55.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-207vhvkp-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.5 GB at this context; duplicate traces of the same session: 1ycsoptk, 55o29pv1, bdb18he1, zzpzx7gg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 509.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 32k 509.3 46.6 31.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/207vhvkp",
   "source_url": "https://omlx.ai/benchmarks/performance/207vhvkp",
   "tg_tokens": null,
   "tps": 46.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-1447i6m8-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: bycz4r7t, cn2jreel, kqegxiqf, m81ahvhn; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 631.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 1k 631.6 46.6 29.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1447i6m8",
   "source_url": "https://omlx.ai/benchmarks/performance/1447i6m8",
   "tg_tokens": null,
   "tps": 46.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-1447i6m8-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.8 GB at this context; duplicate traces of the same session: bycz4r7t, cn2jreel, kqegxiqf, m81ahvhn; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 736.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 4k 736.0 72.6 30.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1447i6m8",
   "source_url": "https://omlx.ai/benchmarks/performance/1447i6m8",
   "tg_tokens": null,
   "tps": 72.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-1447i6m8-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.1 GB at this context; duplicate traces of the same session: bycz4r7t, cn2jreel, kqegxiqf, m81ahvhn; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 723.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 8k 723.1 62.5 31.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1447i6m8",
   "source_url": "https://omlx.ai/benchmarks/performance/1447i6m8",
   "tg_tokens": null,
   "tps": 62.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-1447i6m8-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.2 GB at this context; duplicate traces of the same session: bycz4r7t, cn2jreel, kqegxiqf, m81ahvhn; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 637.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 16k 637.4 55.7 31.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1447i6m8",
   "source_url": "https://omlx.ai/benchmarks/performance/1447i6m8",
   "tg_tokens": null,
   "tps": 55.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-1447i6m8-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.5 GB at this context; duplicate traces of the same session: bycz4r7t, cn2jreel, kqegxiqf, m81ahvhn; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 516.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 32k 516.0 34.3 31.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1447i6m8",
   "source_url": "https://omlx.ai/benchmarks/performance/1447i6m8",
   "tg_tokens": null,
   "tps": 34.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-7dls4ln6-ctx1024",
   "model": "Qwen3.6-35B-MoE",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2626.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-MoE-Q4 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,626 101.2 19.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7dls4ln6",
   "source_url": "https://omlx.ai/benchmarks/performance/7dls4ln6",
   "tg_tokens": null,
   "tps": 101.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-7dls4ln6-ctx4096",
   "model": "Qwen3.6-35B-MoE",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3216.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-MoE-Q4 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,216 99.9 20.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7dls4ln6",
   "source_url": "https://omlx.ai/benchmarks/performance/7dls4ln6",
   "tg_tokens": null,
   "tps": 99.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-cl07o1oq-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 27.7 GB; TTFT 11577 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 353.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-Uncensored-oQ6e-mtp on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 6bit; 4k tokens 353.8 PP tok/s 39.9 TG tok/s 11577 TTFT (ms) 27.7 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cl07o1oq",
   "source_url": "https://omlx.ai/benchmarks/performance/cl07o1oq",
   "tg_tokens": null,
   "tps": 39.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-qhm4h1a6-ctx1024",
   "model": "Qwen3.8-27B-heretic-ara",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: ezq70joo, n0y5aj87; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 291.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-heretic-ara-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 1k 291.7 41.8 19.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qhm4h1a6",
   "source_url": "https://omlx.ai/benchmarks/performance/qhm4h1a6",
   "tg_tokens": null,
   "tps": 41.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-qhm4h1a6-ctx4096",
   "model": "Qwen3.8-27B-heretic-ara",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: ezq70joo, n0y5aj87; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 311.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-heretic-ara-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 4k 311.7 42.9 23.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qhm4h1a6",
   "source_url": "https://omlx.ai/benchmarks/performance/qhm4h1a6",
   "tg_tokens": null,
   "tps": 42.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-qhm4h1a6-ctx8192",
   "model": "Qwen3.8-27B-heretic-ara",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: ezq70joo, n0y5aj87; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 304.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-heretic-ara-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 8k 304.8 44.0 23.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qhm4h1a6",
   "source_url": "https://omlx.ai/benchmarks/performance/qhm4h1a6",
   "tg_tokens": null,
   "tps": 44.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-qhm4h1a6-ctx16384",
   "model": "Qwen3.8-27B-heretic-ara",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: ezq70joo, n0y5aj87; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 290.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-heretic-ara-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 16k 290.5 55.5 25.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qhm4h1a6",
   "source_url": "https://omlx.ai/benchmarks/performance/qhm4h1a6",
   "tg_tokens": null,
   "tps": 55.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-08fy6xlj-ctx1024",
   "model": "Qwen3.8-27B-heretic-ara",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 9e0uth8r, a8ode2x7; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 294.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-heretic-ara-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 1k 294.5 30.4 16.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/08fy6xlj",
   "source_url": "https://omlx.ai/benchmarks/performance/08fy6xlj",
   "tg_tokens": null,
   "tps": 30.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-08fy6xlj-ctx4096",
   "model": "Qwen3.8-27B-heretic-ara",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 9e0uth8r, a8ode2x7; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-heretic-ara-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 4k 323.2 30.5 20.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/08fy6xlj",
   "source_url": "https://omlx.ai/benchmarks/performance/08fy6xlj",
   "tg_tokens": null,
   "tps": 30.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-08fy6xlj-ctx8192",
   "model": "Qwen3.8-27B-heretic-ara",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 9e0uth8r, a8ode2x7; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 316.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-heretic-ara-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 8k 316.2 29.8 20.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/08fy6xlj",
   "source_url": "https://omlx.ai/benchmarks/performance/08fy6xlj",
   "tg_tokens": null,
   "tps": 29.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-08fy6xlj-ctx16384",
   "model": "Qwen3.8-27B-heretic-ara",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: 9e0uth8r, a8ode2x7; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 302.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-heretic-ara-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 16k 302.0 28.5 22.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/08fy6xlj",
   "source_url": "https://omlx.ai/benchmarks/performance/08fy6xlj",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-l8bn2inv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: p08fe7e6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 386.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 386.1 40.6 24.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l8bn2inv",
   "source_url": "https://omlx.ai/benchmarks/performance/l8bn2inv",
   "tg_tokens": null,
   "tps": 40.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-l8bn2inv-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: p08fe7e6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 408.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 408.0 81.2 25.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l8bn2inv",
   "source_url": "https://omlx.ai/benchmarks/performance/l8bn2inv",
   "tg_tokens": null,
   "tps": 81.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-5efcsk45-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 0nebxx31, 2qjmtv25; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 102.5 11.0 18.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5efcsk45",
   "source_url": "https://omlx.ai/benchmarks/performance/5efcsk45",
   "tg_tokens": null,
   "tps": 11.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-5efcsk45-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 0nebxx31, 2qjmtv25; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 102.2 20.6 19.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5efcsk45",
   "source_url": "https://omlx.ai/benchmarks/performance/5efcsk45",
   "tg_tokens": null,
   "tps": 20.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-5efcsk45-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: 0nebxx31, 2qjmtv25; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 93.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 93.1 11.7 20.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5efcsk45",
   "source_url": "https://omlx.ai/benchmarks/performance/5efcsk45",
   "tg_tokens": null,
   "tps": 11.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-5efcsk45-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 0nebxx31, 2qjmtv25; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 89.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 89.5 18.1 20.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5efcsk45",
   "source_url": "https://omlx.ai/benchmarks/performance/5efcsk45",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-5efcsk45-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 0nebxx31, 2qjmtv25; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 86.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 86.4 14.4 21.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5efcsk45",
   "source_url": "https://omlx.ai/benchmarks/performance/5efcsk45",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1kv8i3tw-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 7go46kdw, e6myc3g5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 992.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 992.5 66.1 19.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1kv8i3tw",
   "source_url": "https://omlx.ai/benchmarks/performance/1kv8i3tw",
   "tg_tokens": null,
   "tps": 66.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1kv8i3tw-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 7go46kdw, e6myc3g5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 966.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 966.4 83.6 19.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1kv8i3tw",
   "source_url": "https://omlx.ai/benchmarks/performance/1kv8i3tw",
   "tg_tokens": null,
   "tps": 83.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1kv8i3tw-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: 7go46kdw, e6myc3g5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 618.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 618.8 47.6 27.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1kv8i3tw",
   "source_url": "https://omlx.ai/benchmarks/performance/1kv8i3tw",
   "tg_tokens": null,
   "tps": 47.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-x5o921l3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 804.6,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-27B-oQ6 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 804.6 19.2 23.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x5o921l3",
   "source_url": "https://omlx.ai/benchmarks/performance/x5o921l3",
   "tg_tokens": null,
   "tps": 19.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-x5o921l3-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 552.3,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-27B-oQ6 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 552.3 17.4 24.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x5o921l3",
   "source_url": "https://omlx.ai/benchmarks/performance/x5o921l3",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-pqopb8gv-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 142.6 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1022.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 1k 1,022 84.7 142.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pqopb8gv",
   "source_url": "https://omlx.ai/benchmarks/performance/pqopb8gv",
   "tg_tokens": null,
   "tps": 84.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-pqopb8gv-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1320.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 4k 1,320 79.6 143.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pqopb8gv",
   "source_url": "https://omlx.ai/benchmarks/performance/pqopb8gv",
   "tg_tokens": null,
   "tps": 79.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-pqopb8gv-ctx199680",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 151.5 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1275.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 195k 1,275 64.6 151.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pqopb8gv",
   "source_url": "https://omlx.ai/benchmarks/performance/pqopb8gv",
   "tg_tokens": null,
   "tps": 64.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-scbbltqj-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: zdmo1rn4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 600.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 600.0 60.3 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/scbbltqj",
   "source_url": "https://omlx.ai/benchmarks/performance/scbbltqj",
   "tg_tokens": null,
   "tps": 60.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-scbbltqj-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: zdmo1rn4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 716.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 716.3 59.1 71.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/scbbltqj",
   "source_url": "https://omlx.ai/benchmarks/performance/scbbltqj",
   "tg_tokens": null,
   "tps": 59.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-scbbltqj-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: zdmo1rn4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 819.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 819.3 71.7 71.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/scbbltqj",
   "source_url": "https://omlx.ai/benchmarks/performance/scbbltqj",
   "tg_tokens": null,
   "tps": 71.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xzns351a-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1198.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,198 77.7 69.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xzns351a",
   "source_url": "https://omlx.ai/benchmarks/performance/xzns351a",
   "tg_tokens": null,
   "tps": 77.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xzns351a-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2145.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,145 70.2 69.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xzns351a",
   "source_url": "https://omlx.ai/benchmarks/performance/xzns351a",
   "tg_tokens": null,
   "tps": 70.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n71bcawo-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 480.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 480.0 33.5 70.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n71bcawo",
   "source_url": "https://omlx.ai/benchmarks/performance/n71bcawo",
   "tg_tokens": null,
   "tps": 33.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n71bcawo-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1139.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,139 55.9 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n71bcawo",
   "source_url": "https://omlx.ai/benchmarks/performance/n71bcawo",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3bynaf7q-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace, single run at 4k context; peak memory 143.4 GB; TTFT 4018 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1019.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 4k tokens 1,019 PP tok/s 87.9 TG tok/s 4018 TTFT (ms) 143.4 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3bynaf7q",
   "source_url": "https://omlx.ai/benchmarks/performance/3bynaf7q",
   "tg_tokens": null,
   "tps": 87.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-wg8iz6jh-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap build; oMLX community benchmark trace; peak memory 57.0 GB at this context; duplicate traces of the same session: fjo34ake; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1470.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "ThinkingCap-Qwen3.8-27B on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 unknown; 32k 1,470 63.5 57.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wg8iz6jh",
   "source_url": "https://omlx.ai/benchmarks/performance/wg8iz6jh",
   "tg_tokens": null,
   "tps": 63.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-wg8iz6jh-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap build; oMLX community benchmark trace; peak memory 59.2 GB at this context; duplicate traces of the same session: fjo34ake; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1272.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "ThinkingCap-Qwen3.8-27B on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 unknown; 64k 1,272 43.6 59.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wg8iz6jh",
   "source_url": "https://omlx.ai/benchmarks/performance/wg8iz6jh",
   "tg_tokens": null,
   "tps": 43.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-2yjrsg40-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.0 GB at this context; duplicate traces of the same session: k3stvlim; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2153.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,153 83.0 14.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2yjrsg40",
   "source_url": "https://omlx.ai/benchmarks/performance/2yjrsg40",
   "tg_tokens": null,
   "tps": 83.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-2yjrsg40-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.9 GB at this context; duplicate traces of the same session: k3stvlim; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2511.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,511 77.8 14.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2yjrsg40",
   "source_url": "https://omlx.ai/benchmarks/performance/2yjrsg40",
   "tg_tokens": null,
   "tps": 77.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-s64baypc-ctx4096",
   "model": "pplx-computer-qwen-3-6-35b-a3b",
   "notes": "MLX build 20260709; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1725.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "perplexity-ai--pplx-computer-qwen-3-6-35b-a3b-mlx-20260709 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,725 100.8 20.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s64baypc",
   "source_url": "https://omlx.ai/benchmarks/performance/s64baypc",
   "tg_tokens": null,
   "tps": 100.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-s64baypc-ctx16384",
   "model": "pplx-computer-qwen-3-6-35b-a3b",
   "notes": "MLX build 20260709; oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2829.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "perplexity-ai--pplx-computer-qwen-3-6-35b-a3b-mlx-20260709 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,829 104.2 20.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s64baypc",
   "source_url": "https://omlx.ai/benchmarks/performance/s64baypc",
   "tg_tokens": null,
   "tps": 104.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-5ykucys7-ctx4096",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "IsValorum APEX MiniPlus Abliterated MLX build; oMLX community benchmark trace, single run at 4k context; peak memory 22.7 GB; TTFT 9380 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 436.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "IsValorum--Qwen3.8-35B-A3B-Distill-MLX-APEX-MiniPlus-Abliterated on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k tokens 436.7 PP tok/s 16.5 TG tok/s 9380 TTFT (ms) 22.7 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ykucys7",
   "source_url": "https://omlx.ai/benchmarks/performance/5ykucys7",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-ih7emyw7-ctx32768",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "IsValorum APEX MiniPlus Abliterated MLX build; oMLX community benchmark trace, single run at 32k context; peak memory 24.0 GB; TTFT 149540 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 219.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "IsValorum--Qwen3.8-35B-A3B-Distill-MLX-APEX-MiniPlus-Abliterated on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 32k tokens 219.1 PP tok/s 13.8 TG tok/s 149540 TTFT (ms) 24.0 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ih7emyw7",
   "source_url": "https://omlx.ai/benchmarks/performance/ih7emyw7",
   "tg_tokens": null,
   "tps": 13.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-nljvhc18-ctx4096",
   "model": "LFM2.5-8B-A1B",
   "notes": "LiquidAI MLX build; oMLX community benchmark trace, single run at 4k context; peak memory 5.4 GB; TTFT 4958 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 826.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LiquidAI--LFM2.5-8B-A1B-MLX-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k tokens 826.2 PP tok/s 93.7 TG tok/s 4958 TTFT (ms) 5.4 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nljvhc18",
   "source_url": "https://omlx.ai/benchmarks/performance/nljvhc18",
   "tg_tokens": null,
   "tps": 93.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-jtrnk5ix-ctx1024",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 1k context; peak memory 5.8 GB; TTFT 1630 ms; duplicate traces of the same session: none; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 628.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-oQ4e-mtp on M5 (8c); M5 (8c) \u00b7 16 GB \u00b7 4bit; 1k tokens 628.1 PP tok/s 26.6 TG tok/s 1630 TTFT (ms) 5.8 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jtrnk5ix",
   "source_url": "https://omlx.ai/benchmarks/performance/jtrnk5ix",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jxig8tik-ctx65536",
   "model": "Muse Glimmer",
   "notes": "oMLX community benchmark trace; peak memory 16.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 896.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-30B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 896.6 28.4 16.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jxig8tik",
   "source_url": "https://omlx.ai/benchmarks/performance/jxig8tik",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jxig8tik-ctx131072",
   "model": "Muse Glimmer",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 807.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-30B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 807.4 25.9 17.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jxig8tik",
   "source_url": "https://omlx.ai/benchmarks/performance/jxig8tik",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-3vlj25oe-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: 9z1tu7lf; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 237.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-9B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 237.9 20.4 6.3 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3vlj25oe",
   "source_url": "https://omlx.ai/benchmarks/performance/3vlj25oe",
   "tg_tokens": null,
   "tps": 20.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-3vlj25oe-ctx8192",
   "model": "Ornith-1.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: 9z1tu7lf; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 235.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-9B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 8k 235.2 19.3 6.5 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3vlj25oe",
   "source_url": "https://omlx.ai/benchmarks/performance/3vlj25oe",
   "tg_tokens": null,
   "tps": 19.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-jw543wz2-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: b3qp2bt9; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 589.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 589.6 71.6 3.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jw543wz2",
   "source_url": "https://omlx.ai/benchmarks/performance/jw543wz2",
   "tg_tokens": null,
   "tps": 71.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-jw543wz2-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.5 GB at this context; duplicate traces of the same session: b3qp2bt9; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 521.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 16k 521.5 56.0 4.5 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jw543wz2",
   "source_url": "https://omlx.ai/benchmarks/performance/jw543wz2",
   "tg_tokens": null,
   "tps": 56.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-3hnfh033-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 21.2 GB; TTFT 3358 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1220.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k tokens 1,220 PP tok/s 40.9 TG tok/s 3358 TTFT (ms) 21.2 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3hnfh033",
   "source_url": "https://omlx.ai/benchmarks/performance/3hnfh033",
   "tg_tokens": null,
   "tps": 40.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-q32kghv9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 17bm1vv4, 1n91irih, 1vja03y3, lcem6hdz; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 97.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 97.9 22.4 20.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q32kghv9",
   "source_url": "https://omlx.ai/benchmarks/performance/q32kghv9",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-q32kghv9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 17bm1vv4, 1n91irih, 1vja03y3, lcem6hdz; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 103.1 26.8 22.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q32kghv9",
   "source_url": "https://omlx.ai/benchmarks/performance/q32kghv9",
   "tg_tokens": null,
   "tps": 26.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-q32kghv9-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 17bm1vv4, 1n91irih, 1vja03y3, lcem6hdz; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 101.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 101.8 15.1 23.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q32kghv9",
   "source_url": "https://omlx.ai/benchmarks/performance/q32kghv9",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-q32kghv9-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: 17bm1vv4, 1n91irih, 1vja03y3, lcem6hdz; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 94.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 94.4 8.3 24.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q32kghv9",
   "source_url": "https://omlx.ai/benchmarks/performance/q32kghv9",
   "tg_tokens": null,
   "tps": 8.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-q32kghv9-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: 17bm1vv4, 1n91irih, 1vja03y3, lcem6hdz; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 82.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 82.6 4.4 26.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q32kghv9",
   "source_url": "https://omlx.ai/benchmarks/performance/q32kghv9",
   "tg_tokens": null,
   "tps": 4.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ohoelv4g-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: 3v5jkdee; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 126.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 126.5 14.9 16.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ohoelv4g",
   "source_url": "https://omlx.ai/benchmarks/performance/ohoelv4g",
   "tg_tokens": null,
   "tps": 14.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ohoelv4g-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: 3v5jkdee; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 131.2 14.5 17.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ohoelv4g",
   "source_url": "https://omlx.ai/benchmarks/performance/ohoelv4g",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-kg8utul5-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: n60u1svj, t8j00cy5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 403.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 403.7 37.8 22.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kg8utul5",
   "source_url": "https://omlx.ai/benchmarks/performance/kg8utul5",
   "tg_tokens": null,
   "tps": 37.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-kg8utul5-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: n60u1svj, t8j00cy5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 399.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 16k 399.0 23.2 23.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kg8utul5",
   "source_url": "https://omlx.ai/benchmarks/performance/kg8utul5",
   "tg_tokens": null,
   "tps": 23.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-kg8utul5-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.2 GB at this context; duplicate traces of the same session: n60u1svj, t8j00cy5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 242.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 128k 242.0 9.5 28.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kg8utul5",
   "source_url": "https://omlx.ai/benchmarks/performance/kg8utul5",
   "tg_tokens": null,
   "tps": 9.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-whn3h6pd-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 39.0 GB; TTFT 97444 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 672.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k tokens 672.5 PP tok/s 23.3 TG tok/s 97444 TTFT (ms) 39.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/whn3h6pd",
   "source_url": "https://omlx.ai/benchmarks/performance/whn3h6pd",
   "tg_tokens": null,
   "tps": 23.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2pdwvhdi-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 87.9 GB at this context; duplicate traces of the same session: v3pcbetg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 501.9,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ3-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 1k 501.9 52.4 87.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2pdwvhdi",
   "source_url": "https://omlx.ai/benchmarks/performance/2pdwvhdi",
   "tg_tokens": null,
   "tps": 52.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2pdwvhdi-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 88.2 GB at this context; duplicate traces of the same session: v3pcbetg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 871.0,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ3-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 4k 871.0 61.2 88.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2pdwvhdi",
   "source_url": "https://omlx.ai/benchmarks/performance/2pdwvhdi",
   "tg_tokens": null,
   "tps": 61.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2pdwvhdi-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 89.4 GB at this context; duplicate traces of the same session: v3pcbetg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1401.0,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ3-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 8k 1,401 61.6 89.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2pdwvhdi",
   "source_url": "https://omlx.ai/benchmarks/performance/2pdwvhdi",
   "tg_tokens": null,
   "tps": 61.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2pdwvhdi-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 89.8 GB at this context; duplicate traces of the same session: v3pcbetg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1715.0,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ3-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 16k 1,715 73.8 89.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2pdwvhdi",
   "source_url": "https://omlx.ai/benchmarks/performance/2pdwvhdi",
   "tg_tokens": null,
   "tps": 73.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ctm0kjae-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: gi9bzkvz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 586.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 586.4 57.5 101.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ctm0kjae",
   "source_url": "https://omlx.ai/benchmarks/performance/ctm0kjae",
   "tg_tokens": null,
   "tps": 57.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ctm0kjae-ctx16384",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.9 GB at this context; duplicate traces of the same session: gi9bzkvz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 762.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 762.4 46.3 101.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ctm0kjae",
   "source_url": "https://omlx.ai/benchmarks/performance/ctm0kjae",
   "tg_tokens": null,
   "tps": 46.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1rwimmms-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace, single run at 128k context; peak memory 150.5 GB; TTFT 132021 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 992.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 128k tokens 992.8 PP tok/s 57.5 TG tok/s 132021 TTFT (ms) 150.5 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1rwimmms",
   "source_url": "https://omlx.ai/benchmarks/performance/1rwimmms",
   "tg_tokens": null,
   "tps": 57.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-x45yxm82-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 74.1 GB; TTFT 164843 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 397.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k tokens 397.6 PP tok/s 11.3 TG tok/s 164843 TTFT (ms) 74.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x45yxm82",
   "source_url": "https://omlx.ai/benchmarks/performance/x45yxm82",
   "tg_tokens": null,
   "tps": 11.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0b3u9vgu-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 44.2 GB; TTFT 162057 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 404.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k tokens 404.4 PP tok/s 7.5 TG tok/s 162057 TTFT (ms) 44.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0b3u9vgu",
   "source_url": "https://omlx.ai/benchmarks/performance/0b3u9vgu",
   "tg_tokens": null,
   "tps": 7.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-8ci1cit2-ctx32768",
   "model": "Ornith-1.5-35B-A3B-REAP192",
   "notes": "mchen04 MLX build; oMLX community benchmark trace, single run at 32k context; peak memory 16.4 GB; TTFT 138770 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 236.1,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "mchen04--Ornith-1.5-35B-A3B-REAP192-mxfp4-MLX on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 32k tokens 236.1 PP tok/s 37.7 TG tok/s 138770 TTFT (ms) 16.4 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8ci1cit2",
   "source_url": "https://omlx.ai/benchmarks/performance/8ci1cit2",
   "tg_tokens": null,
   "tps": 37.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-n5mwcv3v-ctx4096",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "prism-ml MLX build; oMLX community benchmark trace, single run at 4k context; peak memory 14.7 GB; TTFT 62720 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 65.3,
   "provenance": "community",
   "quant": "2bit",
   "quote": "prism-ml--Ternary-Bonsai-2-27B-mlx-2bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 unknown; 4k tokens 65.3 PP tok/s 10.9 TG tok/s 62720 TTFT (ms) 14.7 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n5mwcv3v",
   "source_url": "https://omlx.ai/benchmarks/performance/n5mwcv3v",
   "tg_tokens": null,
   "tps": 10.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-5k9mjket-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: d6x8ld2t; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 379.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 379.4 47.7 20.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5k9mjket",
   "source_url": "https://omlx.ai/benchmarks/performance/5k9mjket",
   "tg_tokens": null,
   "tps": 47.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-5k9mjket-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: d6x8ld2t; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 302.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k 302.7 39.4 20.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5k9mjket",
   "source_url": "https://omlx.ai/benchmarks/performance/5k9mjket",
   "tg_tokens": null,
   "tps": 39.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-3xl0ytqr-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: 1cdjd2lk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 126.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 126.9 15.4 15.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3xl0ytqr",
   "source_url": "https://omlx.ai/benchmarks/performance/3xl0ytqr",
   "tg_tokens": null,
   "tps": 15.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-3xl0ytqr-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: 1cdjd2lk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 131.8 15.0 17.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3xl0ytqr",
   "source_url": "https://omlx.ai/benchmarks/performance/3xl0ytqr",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-7olvtcge-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 9k6lqyyy; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 38.4,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 38.4 2.3 20.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7olvtcge",
   "source_url": "https://omlx.ai/benchmarks/performance/7olvtcge",
   "tg_tokens": null,
   "tps": 2.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-7olvtcge-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: 9k6lqyyy; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 49.4,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 49.4 2.1 21.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7olvtcge",
   "source_url": "https://omlx.ai/benchmarks/performance/7olvtcge",
   "tg_tokens": null,
   "tps": 2.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bg0ibz5t-ctx1024",
   "model": "Qwen3.8-27B-heretic-ara",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: a10yygdf, ome9bpi1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 728.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-heretic-ara-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 728.3 48.1 17.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bg0ibz5t",
   "source_url": "https://omlx.ai/benchmarks/performance/bg0ibz5t",
   "tg_tokens": null,
   "tps": 48.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bg0ibz5t-ctx4096",
   "model": "Qwen3.8-27B-heretic-ara",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: a10yygdf, ome9bpi1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 879.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-heretic-ara-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 879.3 54.1 18.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bg0ibz5t",
   "source_url": "https://omlx.ai/benchmarks/performance/bg0ibz5t",
   "tg_tokens": null,
   "tps": 54.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bg0ibz5t-ctx8192",
   "model": "Qwen3.8-27B-heretic-ara",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: a10yygdf, ome9bpi1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 803.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-heretic-ara-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 803.2 62.6 19.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bg0ibz5t",
   "source_url": "https://omlx.ai/benchmarks/performance/bg0ibz5t",
   "tg_tokens": null,
   "tps": 62.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bg0ibz5t-ctx16384",
   "model": "Qwen3.8-27B-heretic-ara",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: a10yygdf, ome9bpi1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 754.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-heretic-ara-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 754.6 73.0 19.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bg0ibz5t",
   "source_url": "https://omlx.ai/benchmarks/performance/bg0ibz5t",
   "tg_tokens": null,
   "tps": 73.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-cw2gvd09-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 87.9 GB at this context; duplicate traces of the same session: khyu115b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 501.9,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ3-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 1k 501.9 52.4 87.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cw2gvd09",
   "source_url": "https://omlx.ai/benchmarks/performance/cw2gvd09",
   "tg_tokens": null,
   "tps": 52.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-cw2gvd09-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 88.2 GB at this context; duplicate traces of the same session: khyu115b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 871.0,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ3-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 4k 871.0 61.2 88.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cw2gvd09",
   "source_url": "https://omlx.ai/benchmarks/performance/cw2gvd09",
   "tg_tokens": null,
   "tps": 61.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-cw2gvd09-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 89.4 GB at this context; duplicate traces of the same session: khyu115b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1401.0,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ3-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 8k 1,401 61.6 89.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cw2gvd09",
   "source_url": "https://omlx.ai/benchmarks/performance/cw2gvd09",
   "tg_tokens": null,
   "tps": 61.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-cw2gvd09-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 89.8 GB at this context; duplicate traces of the same session: khyu115b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1715.0,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ3-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 16k 1,715 73.8 89.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cw2gvd09",
   "source_url": "https://omlx.ai/benchmarks/performance/cw2gvd09",
   "tg_tokens": null,
   "tps": 73.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3vax8t0j-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 2wht77kt, 495mkp2o, by2vkzm2, c08szlwg, ghwmbfuc, lmqeuhmu, wm0nrfhx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1130.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,130 76.7 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3vax8t0j",
   "source_url": "https://omlx.ai/benchmarks/performance/3vax8t0j",
   "tg_tokens": null,
   "tps": 76.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3vax8t0j-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: 2wht77kt, 495mkp2o, by2vkzm2, c08szlwg, ghwmbfuc, lmqeuhmu, wm0nrfhx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2022.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,022 82.0 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3vax8t0j",
   "source_url": "https://omlx.ai/benchmarks/performance/3vax8t0j",
   "tg_tokens": null,
   "tps": 82.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3vax8t0j-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 2wht77kt, 495mkp2o, by2vkzm2, c08szlwg, ghwmbfuc, lmqeuhmu, wm0nrfhx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2308.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,308 57.9 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3vax8t0j",
   "source_url": "https://omlx.ai/benchmarks/performance/3vax8t0j",
   "tg_tokens": null,
   "tps": 57.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3vax8t0j-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.9 GB at this context; duplicate traces of the same session: 2wht77kt, 495mkp2o, by2vkzm2, c08szlwg, ghwmbfuc, lmqeuhmu, wm0nrfhx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2633.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,633 84.5 72.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3vax8t0j",
   "source_url": "https://omlx.ai/benchmarks/performance/3vax8t0j",
   "tg_tokens": null,
   "tps": 84.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3vax8t0j-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.7 GB at this context; duplicate traces of the same session: 2wht77kt, 495mkp2o, by2vkzm2, c08szlwg, ghwmbfuc, lmqeuhmu, wm0nrfhx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2551.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,551 78.7 73.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3vax8t0j",
   "source_url": "https://omlx.ai/benchmarks/performance/3vax8t0j",
   "tg_tokens": null,
   "tps": 78.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3vax8t0j-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 75.3 GB at this context; duplicate traces of the same session: 2wht77kt, 495mkp2o, by2vkzm2, c08szlwg, ghwmbfuc, lmqeuhmu, wm0nrfhx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2406.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,406 70.1 75.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3vax8t0j",
   "source_url": "https://omlx.ai/benchmarks/performance/3vax8t0j",
   "tg_tokens": null,
   "tps": 70.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3vax8t0j-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 81.0 GB at this context; duplicate traces of the same session: 2wht77kt, 495mkp2o, by2vkzm2, c08szlwg, ghwmbfuc, lmqeuhmu, wm0nrfhx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2311.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,311 62.0 81.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3vax8t0j",
   "source_url": "https://omlx.ai/benchmarks/performance/3vax8t0j",
   "tg_tokens": null,
   "tps": 62.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3vax8t0j-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 84.0 GB at this context; duplicate traces of the same session: 2wht77kt, 495mkp2o, by2vkzm2, c08szlwg, ghwmbfuc, lmqeuhmu, wm0nrfhx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2313.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,313 73.5 84.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3vax8t0j",
   "source_url": "https://omlx.ai/benchmarks/performance/3vax8t0j",
   "tg_tokens": null,
   "tps": 73.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-w3vkj8gv-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 722.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 722.6 92.8 21.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w3vkj8gv",
   "source_url": "https://omlx.ai/benchmarks/performance/w3vkj8gv",
   "tg_tokens": null,
   "tps": 92.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-w3vkj8gv-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 835.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 835.1 85.2 21.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w3vkj8gv",
   "source_url": "https://omlx.ai/benchmarks/performance/w3vkj8gv",
   "tg_tokens": null,
   "tps": 85.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-w3vkj8gv-ctx8192",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 791.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 791.5 74.7 22.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w3vkj8gv",
   "source_url": "https://omlx.ai/benchmarks/performance/w3vkj8gv",
   "tg_tokens": null,
   "tps": 74.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ip7xvyra-ctx1024",
   "model": "Youtu-Parsing-Omni",
   "notes": "MLX; oMLX community benchmark trace; peak memory 7.3 GB at this context; duplicate traces of the same session: c1ylb76o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3702.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Youtu-Parsing-Omni-mlx-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 3,702 83.7 7.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ip7xvyra",
   "source_url": "https://omlx.ai/benchmarks/performance/ip7xvyra",
   "tg_tokens": null,
   "tps": 83.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ip7xvyra-ctx4096",
   "model": "Youtu-Parsing-Omni",
   "notes": "MLX; oMLX community benchmark trace; peak memory 8.9 GB at this context; duplicate traces of the same session: c1ylb76o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3217.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Youtu-Parsing-Omni-mlx-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 3,217 75.3 8.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ip7xvyra",
   "source_url": "https://omlx.ai/benchmarks/performance/ip7xvyra",
   "tg_tokens": null,
   "tps": 75.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-skww8rg5-ctx1024",
   "model": "Youtu-Parsing-Omni",
   "notes": "MLX; oMLX community benchmark trace; peak memory 7.3 GB at this context; duplicate traces of the same session: kh2b1ic3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3764.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Youtu-Parsing-Omni-mlx-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 3,764 84.1 7.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/skww8rg5",
   "source_url": "https://omlx.ai/benchmarks/performance/skww8rg5",
   "tg_tokens": null,
   "tps": 84.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-skww8rg5-ctx4096",
   "model": "Youtu-Parsing-Omni",
   "notes": "MLX; oMLX community benchmark trace; peak memory 8.9 GB at this context; duplicate traces of the same session: kh2b1ic3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2471.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Youtu-Parsing-Omni-mlx-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 2,471 75.0 8.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/skww8rg5",
   "source_url": "https://omlx.ai/benchmarks/performance/skww8rg5",
   "tg_tokens": null,
   "tps": 75.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-b4hc1z3j-ctx1024",
   "model": "Youtu-Parsing-Omni",
   "notes": "MLX; oMLX community benchmark trace; peak memory 7.3 GB at this context; duplicate traces of the same session: o4sc8wlc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4080.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Youtu-Parsing-Omni-mlx-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 4,080 84.2 7.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b4hc1z3j",
   "source_url": "https://omlx.ai/benchmarks/performance/b4hc1z3j",
   "tg_tokens": null,
   "tps": 84.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-b4hc1z3j-ctx4096",
   "model": "Youtu-Parsing-Omni",
   "notes": "MLX; oMLX community benchmark trace; peak memory 8.9 GB at this context; duplicate traces of the same session: o4sc8wlc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3276.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Youtu-Parsing-Omni-mlx-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 3,276 73.6 8.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b4hc1z3j",
   "source_url": "https://omlx.ai/benchmarks/performance/b4hc1z3j",
   "tg_tokens": null,
   "tps": 73.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0uefptwl-ctx1024",
   "model": "Youtu-Parsing-Omni",
   "notes": "MLX; oMLX community benchmark trace; peak memory 7.3 GB at this context; duplicate traces of the same session: xw8jd5iq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2044.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Youtu-Parsing-Omni-mlx-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 2,044 42.3 7.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0uefptwl",
   "source_url": "https://omlx.ai/benchmarks/performance/0uefptwl",
   "tg_tokens": null,
   "tps": 42.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0uefptwl-ctx4096",
   "model": "Youtu-Parsing-Omni",
   "notes": "MLX; oMLX community benchmark trace; peak memory 8.9 GB at this context; duplicate traces of the same session: xw8jd5iq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1626.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Youtu-Parsing-Omni-mlx-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 1,626 35.1 8.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0uefptwl",
   "source_url": "https://omlx.ai/benchmarks/performance/0uefptwl",
   "tg_tokens": null,
   "tps": 35.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-akgdskeg-ctx1024",
   "model": "qwen38-obliteratus",
   "notes": "oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 2crd013a; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "qwen38-obliteratus on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 1k 103.8 11.1 22.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/akgdskeg",
   "source_url": "https://omlx.ai/benchmarks/performance/akgdskeg",
   "tg_tokens": null,
   "tps": 11.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-akgdskeg-ctx4096",
   "model": "qwen38-obliteratus",
   "notes": "oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: 2crd013a; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "qwen38-obliteratus on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 4k 104.1 10.9 24.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/akgdskeg",
   "source_url": "https://omlx.ai/benchmarks/performance/akgdskeg",
   "tg_tokens": null,
   "tps": 10.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fcl75g2e-ctx65536",
   "model": "muse-glimmer 30B",
   "notes": "OptiQ; oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: 0jtvvbcz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 872.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-30B-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 872.3 24.0 20.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fcl75g2e",
   "source_url": "https://omlx.ai/benchmarks/performance/fcl75g2e",
   "tg_tokens": null,
   "tps": 24.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fcl75g2e-ctx131072",
   "model": "muse-glimmer 30B",
   "notes": "OptiQ; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 0jtvvbcz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 785.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-30B-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 785.3 22.2 21.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fcl75g2e",
   "source_url": "https://omlx.ai/benchmarks/performance/fcl75g2e",
   "tg_tokens": null,
   "tps": 22.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-un36o0hr-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: 215cq04x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1007.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,007 32.0 17.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/un36o0hr",
   "source_url": "https://omlx.ai/benchmarks/performance/un36o0hr",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-un36o0hr-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 215cq04x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 885.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 885.9 30.6 17.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/un36o0hr",
   "source_url": "https://omlx.ai/benchmarks/performance/un36o0hr",
   "tg_tokens": null,
   "tps": 30.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-pgboqx4d-ctx32768",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace, single run at 32k context; peak memory 22.9 GB; TTFT 79299 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 413.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k tokens 413.2 PP tok/s 10.6 TG tok/s 79299 TTFT (ms) 22.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pgboqx4d",
   "source_url": "https://omlx.ai/benchmarks/performance/pgboqx4d",
   "tg_tokens": null,
   "tps": 10.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-kbvllyy2-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 21.0 GB; TTFT 19013 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 215.4,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k tokens 215.4 PP tok/s 8.1 TG tok/s 19013 TTFT (ms) 21.0 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kbvllyy2",
   "source_url": "https://omlx.ai/benchmarks/performance/kbvllyy2",
   "tg_tokens": null,
   "tps": 8.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-thstrch0-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 20.0 GB; TTFT 5398 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 189.7,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k tokens 189.7 PP tok/s 8.4 TG tok/s 5398 TTFT (ms) 20.0 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/thstrch0",
   "source_url": "https://omlx.ai/benchmarks/performance/thstrch0",
   "tg_tokens": null,
   "tps": 8.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-l4roih7n-ctx1024",
   "model": "Qwen3.8-27B-heretic-ara",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 728.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-heretic-ara-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 728.3 48.1 17.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l4roih7n",
   "source_url": "https://omlx.ai/benchmarks/performance/l4roih7n",
   "tg_tokens": null,
   "tps": 48.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-l4roih7n-ctx4096",
   "model": "Qwen3.8-27B-heretic-ara",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 879.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-heretic-ara-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 879.3 54.1 18.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l4roih7n",
   "source_url": "https://omlx.ai/benchmarks/performance/l4roih7n",
   "tg_tokens": null,
   "tps": 54.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-l4roih7n-ctx8192",
   "model": "Qwen3.8-27B-heretic-ara",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 803.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-heretic-ara-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 803.2 62.6 19.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l4roih7n",
   "source_url": "https://omlx.ai/benchmarks/performance/l4roih7n",
   "tg_tokens": null,
   "tps": 62.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-l4roih7n-ctx16384",
   "model": "Qwen3.8-27B-heretic-ara",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 754.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-heretic-ara-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 754.6 73.0 19.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l4roih7n",
   "source_url": "https://omlx.ai/benchmarks/performance/l4roih7n",
   "tg_tokens": null,
   "tps": 73.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-t93ml2y3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: ontjboie; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1392.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 4bit; 1k 1,392 52.2 16.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t93ml2y3",
   "source_url": "https://omlx.ai/benchmarks/performance/t93ml2y3",
   "tg_tokens": null,
   "tps": 52.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-t93ml2y3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: ontjboie; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1559.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 4bit; 4k 1,559 50.8 17.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t93ml2y3",
   "source_url": "https://omlx.ai/benchmarks/performance/t93ml2y3",
   "tg_tokens": null,
   "tps": 50.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-lbwvbpat-ctx16384",
   "model": "Qwen3.8-35B-36.82G-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace, single run at 16k context; peak memory 40.7 GB; TTFT 10578 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1549.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-36.82G-A3B-Distill-oQ8-fp16-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k tokens 1,549 PP tok/s 119.9 TG tok/s 10578 TTFT (ms) 40.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lbwvbpat",
   "source_url": "https://omlx.ai/benchmarks/performance/lbwvbpat",
   "tg_tokens": null,
   "tps": 119.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-q0jmc4c2-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 453.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 453.9 47.4 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q0jmc4c2",
   "source_url": "https://omlx.ai/benchmarks/performance/q0jmc4c2",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-q0jmc4c2-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 567.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 567.5 32.2 71.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q0jmc4c2",
   "source_url": "https://omlx.ai/benchmarks/performance/q0jmc4c2",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-slf2gyf7-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 0lc931c8; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 722.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 722.6 92.8 21.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/slf2gyf7",
   "source_url": "https://omlx.ai/benchmarks/performance/slf2gyf7",
   "tg_tokens": null,
   "tps": 92.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-slf2gyf7-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 0lc931c8; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 835.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 835.1 85.2 21.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/slf2gyf7",
   "source_url": "https://omlx.ai/benchmarks/performance/slf2gyf7",
   "tg_tokens": null,
   "tps": 85.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-slf2gyf7-ctx8192",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 0lc931c8; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 791.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 791.5 74.7 22.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/slf2gyf7",
   "source_url": "https://omlx.ai/benchmarks/performance/slf2gyf7",
   "tg_tokens": null,
   "tps": 74.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-l1byreua-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 5lbcshgt, np1gbfss; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 763.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 763.4 74.8 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l1byreua",
   "source_url": "https://omlx.ai/benchmarks/performance/l1byreua",
   "tg_tokens": null,
   "tps": 74.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-l1byreua-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: 5lbcshgt, np1gbfss; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 861.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 861.4 74.2 20.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l1byreua",
   "source_url": "https://omlx.ai/benchmarks/performance/l1byreua",
   "tg_tokens": null,
   "tps": 74.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-l1byreua-ctx8192",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 5lbcshgt, np1gbfss; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 828.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 828.7 70.7 20.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l1byreua",
   "source_url": "https://omlx.ai/benchmarks/performance/l1byreua",
   "tg_tokens": null,
   "tps": 70.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-retpwf3o-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: g2pb7x3a, m9tl3no7; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 739.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 739.7 91.5 21.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/retpwf3o",
   "source_url": "https://omlx.ai/benchmarks/performance/retpwf3o",
   "tg_tokens": null,
   "tps": 91.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-retpwf3o-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: g2pb7x3a, m9tl3no7; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 811.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 811.4 93.4 22.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/retpwf3o",
   "source_url": "https://omlx.ai/benchmarks/performance/retpwf3o",
   "tg_tokens": null,
   "tps": 93.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-retpwf3o-ctx8192",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: g2pb7x3a, m9tl3no7; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 767.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 767.5 71.0 22.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/retpwf3o",
   "source_url": "https://omlx.ai/benchmarks/performance/retpwf3o",
   "tg_tokens": null,
   "tps": 71.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-2kd3qebt-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 2lmq3y7r, tx3dobga; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 761.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 761.8 70.7 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2kd3qebt",
   "source_url": "https://omlx.ai/benchmarks/performance/2kd3qebt",
   "tg_tokens": null,
   "tps": 70.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-2kd3qebt-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 2lmq3y7r, tx3dobga; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 835.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 835.8 74.1 20.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2kd3qebt",
   "source_url": "https://omlx.ai/benchmarks/performance/2kd3qebt",
   "tg_tokens": null,
   "tps": 74.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-2kd3qebt-ctx8192",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 2lmq3y7r, tx3dobga; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 797.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 797.0 69.8 21.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2kd3qebt",
   "source_url": "https://omlx.ai/benchmarks/performance/2kd3qebt",
   "tg_tokens": null,
   "tps": 69.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-3uun9x6f-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: n0jd715h, sqidnhcj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 736.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 736.6 94.1 21.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3uun9x6f",
   "source_url": "https://omlx.ai/benchmarks/performance/3uun9x6f",
   "tg_tokens": null,
   "tps": 94.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-3uun9x6f-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: n0jd715h, sqidnhcj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 807.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 807.8 101.2 22.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3uun9x6f",
   "source_url": "https://omlx.ai/benchmarks/performance/3uun9x6f",
   "tg_tokens": null,
   "tps": 101.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-3uun9x6f-ctx8192",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: n0jd715h, sqidnhcj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 764.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 764.8 83.9 22.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3uun9x6f",
   "source_url": "https://omlx.ai/benchmarks/performance/3uun9x6f",
   "tg_tokens": null,
   "tps": 83.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0u423y49-ctx1024",
   "model": "Youtu-Parsing-Omni",
   "notes": "MLX; oMLX community benchmark trace; peak memory 7.3 GB at this context; duplicate traces of the same session: tm5s6eun; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1896.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Youtu-Parsing-Omni-mlx-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 1,896 42.1 7.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0u423y49",
   "source_url": "https://omlx.ai/benchmarks/performance/0u423y49",
   "tg_tokens": null,
   "tps": 42.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0u423y49-ctx4096",
   "model": "Youtu-Parsing-Omni",
   "notes": "MLX; oMLX community benchmark trace; peak memory 8.9 GB at this context; duplicate traces of the same session: tm5s6eun; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1501.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Youtu-Parsing-Omni-mlx-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 1,501 35.7 8.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0u423y49",
   "source_url": "https://omlx.ai/benchmarks/performance/0u423y49",
   "tg_tokens": null,
   "tps": 35.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-cs59vqi2-ctx32768",
   "model": "Bonsai-27B",
   "notes": "prism-ml MLX build; oMLX community benchmark trace, single run at 32k context; peak memory 11.2 GB; TTFT 576784 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 56.8,
   "provenance": "community",
   "quant": "1bit",
   "quote": "prism-ml--Bonsai-27B-mlx-1bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 unknown; 32k tokens 56.8 PP tok/s 9.5 TG tok/s 576784 TTFT (ms) 11.2 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cs59vqi2",
   "source_url": "https://omlx.ai/benchmarks/performance/cs59vqi2",
   "tg_tokens": null,
   "tps": 9.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-nvki0i39-ctx16384",
   "model": "Qwen3.6-35B",
   "notes": "A3B; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 39.8 GB; TTFT 10491 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1562.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.6-35B-35.98G-A3B-oQ8-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k tokens 1,562 PP tok/s 103.7 TG tok/s 10491 TTFT (ms) 39.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nvki0i39",
   "source_url": "https://omlx.ai/benchmarks/performance/nvki0i39",
   "tg_tokens": null,
   "tps": 103.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-cie7q19a-ctx1024",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 1k context; peak memory 6.9 GB; TTFT 1605 ms; duplicate traces of the same session: none; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 638.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-oQ4e-mtp on M5 (8c); M5 (8c) \u00b7 16 GB \u00b7 4bit; 1k tokens 638.0 PP tok/s 24.4 TG tok/s 1605 TTFT (ms) 6.9 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cie7q19a",
   "source_url": "https://omlx.ai/benchmarks/performance/cie7q19a",
   "tg_tokens": null,
   "tps": 24.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-wn97eyqe-ctx1024",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 1k context; peak memory 6.9 GB; TTFT 1692 ms; duplicate traces of the same session: none; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 605.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-oQ4e-mtp on M5 (8c); M5 (8c) \u00b7 16 GB \u00b7 4bit; 1k tokens 605.2 PP tok/s 23.5 TG tok/s 1692 TTFT (ms) 6.9 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wn97eyqe",
   "source_url": "https://omlx.ai/benchmarks/performance/wn97eyqe",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-er5n49ln-ctx1024",
   "model": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: v9igynw7; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 223.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 223.9 30.3 7.0 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/er5n49ln",
   "source_url": "https://omlx.ai/benchmarks/performance/er5n49ln",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-er5n49ln-ctx8192",
   "model": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.7 GB at this context; duplicate traces of the same session: v9igynw7; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 229.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 8k 229.5 24.8 7.7 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/er5n49ln",
   "source_url": "https://omlx.ai/benchmarks/performance/er5n49ln",
   "tg_tokens": null,
   "tps": 24.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-zuq83mwo-ctx1024",
   "model": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: 72qui22p; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 224.1 25.8 7.0 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zuq83mwo",
   "source_url": "https://omlx.ai/benchmarks/performance/zuq83mwo",
   "tg_tokens": null,
   "tps": 25.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-zuq83mwo-ctx8192",
   "model": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.7 GB at this context; duplicate traces of the same session: 72qui22p; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 229.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 8k 229.5 25.2 7.7 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zuq83mwo",
   "source_url": "https://omlx.ai/benchmarks/performance/zuq83mwo",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-1nporz8m-ctx1024",
   "model": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: vkj1yf90; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 223.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 223.7 40.5 7.0 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1nporz8m",
   "source_url": "https://omlx.ai/benchmarks/performance/1nporz8m",
   "tg_tokens": null,
   "tps": 40.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-1nporz8m-ctx8192",
   "model": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.7 GB at this context; duplicate traces of the same session: vkj1yf90; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 227.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 8k 227.3 21.1 7.7 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1nporz8m",
   "source_url": "https://omlx.ai/benchmarks/performance/1nporz8m",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-z8bxqvxj-ctx1024",
   "model": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 218.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 218.4 34.1 7.0 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z8bxqvxj",
   "source_url": "https://omlx.ai/benchmarks/performance/z8bxqvxj",
   "tg_tokens": null,
   "tps": 34.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-z8bxqvxj-ctx8192",
   "model": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.7 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 229.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 8k 229.3 22.2 7.7 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z8bxqvxj",
   "source_url": "https://omlx.ai/benchmarks/performance/z8bxqvxj",
   "tg_tokens": null,
   "tps": 22.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dno5rhx6-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: j6mrdbjz, qsuqm11a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 784.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 784.5 25.8 20.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dno5rhx6",
   "source_url": "https://omlx.ai/benchmarks/performance/dno5rhx6",
   "tg_tokens": null,
   "tps": 25.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dno5rhx6-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: j6mrdbjz, qsuqm11a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 627.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 627.0 21.4 24.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dno5rhx6",
   "source_url": "https://omlx.ai/benchmarks/performance/dno5rhx6",
   "tg_tokens": null,
   "tps": 21.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dno5rhx6-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.0 GB at this context; duplicate traces of the same session: j6mrdbjz, qsuqm11a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 512.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 195k 512.0 16.9 29.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dno5rhx6",
   "source_url": "https://omlx.ai/benchmarks/performance/dno5rhx6",
   "tg_tokens": null,
   "tps": 16.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-2fv3rfzq-ctx1024",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "MLX; oMLX community benchmark trace, single run at 1k context; peak memory 21.5 GB; TTFT 3647 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 280.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k tokens 280.8 PP tok/s 19.2 TG tok/s 3647 TTFT (ms) 21.5 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2fv3rfzq",
   "source_url": "https://omlx.ai/benchmarks/performance/2fv3rfzq",
   "tg_tokens": null,
   "tps": 19.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-13ezhrv1-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 20.7 GB; TTFT 78899 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 415.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k tokens 415.3 PP tok/s 22.8 TG tok/s 78899 TTFT (ms) 20.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/13ezhrv1",
   "source_url": "https://omlx.ai/benchmarks/performance/13ezhrv1",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-tguwj5j7-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "SAGE-ARCS MTP build; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: s9tt8m0z; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 122.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-SAGE-ARCS-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 5bit; 1k 122.4 17.3 26.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tguwj5j7",
   "source_url": "https://omlx.ai/benchmarks/performance/tguwj5j7",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-tguwj5j7-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "SAGE-ARCS MTP build; oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: s9tt8m0z; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 122.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-SAGE-ARCS-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 5bit; 4k 122.1 11.6 30.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tguwj5j7",
   "source_url": "https://omlx.ai/benchmarks/performance/tguwj5j7",
   "tg_tokens": null,
   "tps": 11.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-c38jmd6o-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: hqdqosbg; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 430.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 430.5 25.6 26.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c38jmd6o",
   "source_url": "https://omlx.ai/benchmarks/performance/c38jmd6o",
   "tg_tokens": null,
   "tps": 25.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-c38jmd6o-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: hqdqosbg; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 453.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 453.8 27.2 27.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c38jmd6o",
   "source_url": "https://omlx.ai/benchmarks/performance/c38jmd6o",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-lnbdop6h-ctx16384",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 607.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 16k 607.8 40.5 71.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lnbdop6h",
   "source_url": "https://omlx.ai/benchmarks/performance/lnbdop6h",
   "tg_tokens": null,
   "tps": 40.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-lnbdop6h-ctx199680",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.6 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 570.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 195k 570.1 12.1 79.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lnbdop6h",
   "source_url": "https://omlx.ai/benchmarks/performance/lnbdop6h",
   "tg_tokens": null,
   "tps": 12.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wh0lemk4-ctx8192",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: c2rk38u6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2337.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,337 74.6 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wh0lemk4",
   "source_url": "https://omlx.ai/benchmarks/performance/wh0lemk4",
   "tg_tokens": null,
   "tps": 74.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wh0lemk4-ctx32768",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: c2rk38u6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2262.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,262 83.1 73.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wh0lemk4",
   "source_url": "https://omlx.ai/benchmarks/performance/wh0lemk4",
   "tg_tokens": null,
   "tps": 83.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2d14ulee-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 77.4 GB at this context; duplicate traces of the same session: 1andybhx, e6zoundy, ipdy1c4p, tkj234lz, v43lxb09; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 543.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 543.0 40.8 77.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2d14ulee",
   "source_url": "https://omlx.ai/benchmarks/performance/2d14ulee",
   "tg_tokens": null,
   "tps": 40.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2d14ulee-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.0 GB at this context; duplicate traces of the same session: 1andybhx, e6zoundy, ipdy1c4p, tkj234lz, v43lxb09; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 615.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 615.7 49.0 78.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2d14ulee",
   "source_url": "https://omlx.ai/benchmarks/performance/2d14ulee",
   "tg_tokens": null,
   "tps": 49.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-2d14ulee-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.2 GB at this context; duplicate traces of the same session: 1andybhx, e6zoundy, ipdy1c4p, tkj234lz, v43lxb09; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 822.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 822.0 70.5 78.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2d14ulee",
   "source_url": "https://omlx.ai/benchmarks/performance/2d14ulee",
   "tg_tokens": null,
   "tps": 70.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2d14ulee-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.1 GB at this context; duplicate traces of the same session: 1andybhx, e6zoundy, ipdy1c4p, tkj234lz, v43lxb09; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 814.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 814.2 45.7 79.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2d14ulee",
   "source_url": "https://omlx.ai/benchmarks/performance/2d14ulee",
   "tg_tokens": null,
   "tps": 45.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2d14ulee-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.7 GB at this context; duplicate traces of the same session: 1andybhx, e6zoundy, ipdy1c4p, tkj234lz, v43lxb09; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 805.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 805.7 60.5 80.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2d14ulee",
   "source_url": "https://omlx.ai/benchmarks/performance/2d14ulee",
   "tg_tokens": null,
   "tps": 60.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2d14ulee-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 85.0 GB at this context; duplicate traces of the same session: 1andybhx, e6zoundy, ipdy1c4p, tkj234lz, v43lxb09; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 793.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 793.7 49.4 85.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2d14ulee",
   "source_url": "https://omlx.ai/benchmarks/performance/2d14ulee",
   "tg_tokens": null,
   "tps": 49.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-5neljbgk-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 453.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 453.9 47.4 70.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5neljbgk",
   "source_url": "https://omlx.ai/benchmarks/performance/5neljbgk",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-5neljbgk-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 567.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 567.5 32.2 71.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5neljbgk",
   "source_url": "https://omlx.ai/benchmarks/performance/5neljbgk",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-crgjtitm-ctx4096",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 218.2,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 5bit; 4k 218.2 13.5 23.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/crgjtitm",
   "source_url": "https://omlx.ai/benchmarks/performance/crgjtitm",
   "tg_tokens": null,
   "tps": 13.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-crgjtitm-ctx16384",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 170.6,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 5bit; 16k 170.6 5.1 25.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/crgjtitm",
   "source_url": "https://omlx.ai/benchmarks/performance/crgjtitm",
   "tg_tokens": null,
   "tps": 5.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-8fj5rabv-ctx1024",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: plylpi3h, wae8bxwc; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 100.3,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 5bit; 1k 100.3 13.7 22.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8fj5rabv",
   "source_url": "https://omlx.ai/benchmarks/performance/8fj5rabv",
   "tg_tokens": null,
   "tps": 13.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-8fj5rabv-ctx4096",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: plylpi3h, wae8bxwc; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 101.6,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 5bit; 4k 101.6 24.2 24.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8fj5rabv",
   "source_url": "https://omlx.ai/benchmarks/performance/8fj5rabv",
   "tg_tokens": null,
   "tps": 24.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-8fj5rabv-ctx8192",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: plylpi3h, wae8bxwc; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 101.3,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 5bit; 8k 101.3 9.9 25.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8fj5rabv",
   "source_url": "https://omlx.ai/benchmarks/performance/8fj5rabv",
   "tg_tokens": null,
   "tps": 9.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-quscf1bb-ctx1024",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 1k context; peak memory 5.8 GB; TTFT 1501 ms; duplicate traces of the same session: none; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 682.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-oQ4e-mtp on M5 (8c); M5 (8c) \u00b7 16 GB \u00b7 4bit; 1k tokens 682.3 PP tok/s 27.1 TG tok/s 1501 TTFT (ms) 5.8 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/quscf1bb",
   "source_url": "https://omlx.ai/benchmarks/performance/quscf1bb",
   "tg_tokens": null,
   "tps": 27.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-d5nzr0ec-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: 4cstffmg; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 437.5,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.5-4B-oQ4e-fp16-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 437.5 34.7 3.9 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d5nzr0ec",
   "source_url": "https://omlx.ai/benchmarks/performance/d5nzr0ec",
   "tg_tokens": null,
   "tps": 34.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-d5nzr0ec-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.1 GB at this context; duplicate traces of the same session: 4cstffmg; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 428.7,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.5-4B-oQ4e-fp16-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 8k 428.7 29.3 4.1 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d5nzr0ec",
   "source_url": "https://omlx.ai/benchmarks/performance/d5nzr0ec",
   "tg_tokens": null,
   "tps": 29.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-afcq0fdi-ctx1024",
   "model": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 218.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 218.4 34.1 7.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/afcq0fdi",
   "source_url": "https://omlx.ai/benchmarks/performance/afcq0fdi",
   "tg_tokens": null,
   "tps": 34.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-afcq0fdi-ctx8192",
   "model": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.7 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 229.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 8k 229.3 22.2 7.7 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/afcq0fdi",
   "source_url": "https://omlx.ai/benchmarks/performance/afcq0fdi",
   "tg_tokens": null,
   "tps": 22.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gf2esjas-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 16g1hjec, 5ldhyddk, laje3oht, quuryi9g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 908.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 908.8 68.6 22.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gf2esjas",
   "source_url": "https://omlx.ai/benchmarks/performance/gf2esjas",
   "tg_tokens": null,
   "tps": 68.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gf2esjas-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: 16g1hjec, 5ldhyddk, laje3oht, quuryi9g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 983.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 983.3 63.7 24.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gf2esjas",
   "source_url": "https://omlx.ai/benchmarks/performance/gf2esjas",
   "tg_tokens": null,
   "tps": 63.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gf2esjas-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: 16g1hjec, 5ldhyddk, laje3oht, quuryi9g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 919.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 919.3 72.1 24.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gf2esjas",
   "source_url": "https://omlx.ai/benchmarks/performance/gf2esjas",
   "tg_tokens": null,
   "tps": 72.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gf2esjas-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: 16g1hjec, 5ldhyddk, laje3oht, quuryi9g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 844.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 844.9 77.2 25.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gf2esjas",
   "source_url": "https://omlx.ai/benchmarks/performance/gf2esjas",
   "tg_tokens": null,
   "tps": 77.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gf2esjas-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: 16g1hjec, 5ldhyddk, laje3oht, quuryi9g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 752.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 752.9 79.1 25.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gf2esjas",
   "source_url": "https://omlx.ai/benchmarks/performance/gf2esjas",
   "tg_tokens": null,
   "tps": 79.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pc7h0gbj-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.9 GB at this context; duplicate traces of the same session: 0526hfwf, 6lfabmsn, nn8xbg0d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 820.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 820.5 52.5 38.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pc7h0gbj",
   "source_url": "https://omlx.ai/benchmarks/performance/pc7h0gbj",
   "tg_tokens": null,
   "tps": 52.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pc7h0gbj-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 41.1 GB at this context; duplicate traces of the same session: 0526hfwf, 6lfabmsn, nn8xbg0d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 913.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 913.0 43.6 41.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pc7h0gbj",
   "source_url": "https://omlx.ai/benchmarks/performance/pc7h0gbj",
   "tg_tokens": null,
   "tps": 43.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pc7h0gbj-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 41.4 GB at this context; duplicate traces of the same session: 0526hfwf, 6lfabmsn, nn8xbg0d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 843.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 843.7 51.2 41.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pc7h0gbj",
   "source_url": "https://omlx.ai/benchmarks/performance/pc7h0gbj",
   "tg_tokens": null,
   "tps": 51.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-pc7h0gbj-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 41.9 GB at this context; duplicate traces of the same session: 0526hfwf, 6lfabmsn, nn8xbg0d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 783.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 783.6 54.3 41.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pc7h0gbj",
   "source_url": "https://omlx.ai/benchmarks/performance/pc7h0gbj",
   "tg_tokens": null,
   "tps": 54.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pc7h0gbj-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 42.8 GB at this context; duplicate traces of the same session: 0526hfwf, 6lfabmsn, nn8xbg0d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 711.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 711.8 54.5 42.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pc7h0gbj",
   "source_url": "https://omlx.ai/benchmarks/performance/pc7h0gbj",
   "tg_tokens": null,
   "tps": 54.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-8umvjvn7-ctx16384",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 607.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 16k 607.8 40.5 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8umvjvn7",
   "source_url": "https://omlx.ai/benchmarks/performance/8umvjvn7",
   "tg_tokens": null,
   "tps": 40.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-8umvjvn7-ctx199680",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.6 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 570.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 195k 570.1 12.1 79.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8umvjvn7",
   "source_url": "https://omlx.ai/benchmarks/performance/8umvjvn7",
   "tg_tokens": null,
   "tps": 12.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-068jc81o-ctx8192",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.0 GB at this context; duplicate traces of the same session: r0le3rdq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 827.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 827.4 18.2 72.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/068jc81o",
   "source_url": "https://omlx.ai/benchmarks/performance/068jc81o",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-068jc81o-ctx32768",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: r0le3rdq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2256.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,256 53.0 73.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/068jc81o",
   "source_url": "https://omlx.ai/benchmarks/performance/068jc81o",
   "tg_tokens": null,
   "tps": 53.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-vq31il44-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 3jny6vyz, 9owjvrv6, eyao8sw5; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 290.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 1k 290.2 40.6 19.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vq31il44",
   "source_url": "https://omlx.ai/benchmarks/performance/vq31il44",
   "tg_tokens": null,
   "tps": 40.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-vq31il44-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: 3jny6vyz, 9owjvrv6, eyao8sw5; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 312.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 4k 312.1 44.3 23.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vq31il44",
   "source_url": "https://omlx.ai/benchmarks/performance/vq31il44",
   "tg_tokens": null,
   "tps": 44.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-vq31il44-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 3jny6vyz, 9owjvrv6, eyao8sw5; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 305.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 8k 305.4 50.3 23.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vq31il44",
   "source_url": "https://omlx.ai/benchmarks/performance/vq31il44",
   "tg_tokens": null,
   "tps": 50.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-vq31il44-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: 3jny6vyz, 9owjvrv6, eyao8sw5; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 289.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 16k 289.1 54.6 25.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vq31il44",
   "source_url": "https://omlx.ai/benchmarks/performance/vq31il44",
   "tg_tokens": null,
   "tps": 54.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0elisp81-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: 2bzbacjw, bpbcgbxk, g60deb20, iagpakae, mnhnjzt6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 942.3,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 942.3 54.8 24.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0elisp81",
   "source_url": "https://omlx.ai/benchmarks/performance/0elisp81",
   "tg_tokens": null,
   "tps": 54.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0elisp81-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.7 GB at this context; duplicate traces of the same session: 2bzbacjw, bpbcgbxk, g60deb20, iagpakae, mnhnjzt6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1053.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 1,053 47.6 25.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0elisp81",
   "source_url": "https://omlx.ai/benchmarks/performance/0elisp81",
   "tg_tokens": null,
   "tps": 47.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0elisp81-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: 2bzbacjw, bpbcgbxk, g60deb20, iagpakae, mnhnjzt6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 978.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 978.1 58.1 26.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0elisp81",
   "source_url": "https://omlx.ai/benchmarks/performance/0elisp81",
   "tg_tokens": null,
   "tps": 58.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0elisp81-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: 2bzbacjw, bpbcgbxk, g60deb20, iagpakae, mnhnjzt6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 917.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 917.7 75.9 26.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0elisp81",
   "source_url": "https://omlx.ai/benchmarks/performance/0elisp81",
   "tg_tokens": null,
   "tps": 75.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0elisp81-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: 2bzbacjw, bpbcgbxk, g60deb20, iagpakae, mnhnjzt6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 823.3,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 823.3 109.4 27.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0elisp81",
   "source_url": "https://omlx.ai/benchmarks/performance/0elisp81",
   "tg_tokens": null,
   "tps": 109.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0elisp81-ctx65536",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.1 GB at this context; duplicate traces of the same session: 2bzbacjw, bpbcgbxk, g60deb20, iagpakae, mnhnjzt6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 713.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 713.2 44.1 30.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0elisp81",
   "source_url": "https://omlx.ai/benchmarks/performance/0elisp81",
   "tg_tokens": null,
   "tps": 44.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-8exclqno-ctx4096",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 218.2,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 5bit; 4k 218.2 13.5 23.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8exclqno",
   "source_url": "https://omlx.ai/benchmarks/performance/8exclqno",
   "tg_tokens": null,
   "tps": 13.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-8exclqno-ctx16384",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 170.6,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 5bit; 16k 170.6 5.1 25.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8exclqno",
   "source_url": "https://omlx.ai/benchmarks/performance/8exclqno",
   "tg_tokens": null,
   "tps": 5.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-e472574g-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: h1kiezl3, n1d7zin6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3685.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 3,685 100.6 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e472574g",
   "source_url": "https://omlx.ai/benchmarks/performance/e472574g",
   "tg_tokens": null,
   "tps": 100.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-e472574g-ctx131072",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: h1kiezl3, n1d7zin6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2513.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 2,513 73.9 22.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e472574g",
   "source_url": "https://omlx.ai/benchmarks/performance/e472574g",
   "tg_tokens": null,
   "tps": 73.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-e472574g-ctx199680",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: h1kiezl3, n1d7zin6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1874.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 195k 1,874 61.6 23.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e472574g",
   "source_url": "https://omlx.ai/benchmarks/performance/e472574g",
   "tg_tokens": null,
   "tps": 61.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cebos38l-ctx4096",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: h2wmv4nd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 992.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 992.3 33.0 17.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cebos38l",
   "source_url": "https://omlx.ai/benchmarks/performance/cebos38l",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-cebos38l-ctx16384",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: h2wmv4nd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 892.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 892.1 31.0 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cebos38l",
   "source_url": "https://omlx.ai/benchmarks/performance/cebos38l",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-6tdyorwk-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 8n3du6e1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 521.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 521.7 47.8 21.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6tdyorwk",
   "source_url": "https://omlx.ai/benchmarks/performance/6tdyorwk",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-6tdyorwk-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 8n3du6e1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2897.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,897 129.0 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6tdyorwk",
   "source_url": "https://omlx.ai/benchmarks/performance/6tdyorwk",
   "tg_tokens": null,
   "tps": 129.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-6tdyorwk-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 8n3du6e1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1589.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 1,589 77.4 23.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6tdyorwk",
   "source_url": "https://omlx.ai/benchmarks/performance/6tdyorwk",
   "tg_tokens": null,
   "tps": 77.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-66kylqn2-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 1i0m9g75, gf64883c; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 605.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 605.9 45.3 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/66kylqn2",
   "source_url": "https://omlx.ai/benchmarks/performance/66kylqn2",
   "tg_tokens": null,
   "tps": 45.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-66kylqn2-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 1i0m9g75, gf64883c; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2897.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,897 112.6 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/66kylqn2",
   "source_url": "https://omlx.ai/benchmarks/performance/66kylqn2",
   "tg_tokens": null,
   "tps": 112.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-66kylqn2-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 1i0m9g75, gf64883c; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1593.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 1,593 73.8 23.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/66kylqn2",
   "source_url": "https://omlx.ai/benchmarks/performance/66kylqn2",
   "tg_tokens": null,
   "tps": 73.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ky1o4jz5-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: f5x6zfi2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 725.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 725.7 47.7 21.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ky1o4jz5",
   "source_url": "https://omlx.ai/benchmarks/performance/ky1o4jz5",
   "tg_tokens": null,
   "tps": 47.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ky1o4jz5-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: f5x6zfi2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2879.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,879 118.0 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ky1o4jz5",
   "source_url": "https://omlx.ai/benchmarks/performance/ky1o4jz5",
   "tg_tokens": null,
   "tps": 118.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ky1o4jz5-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: f5x6zfi2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1592.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 1,592 75.1 23.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ky1o4jz5",
   "source_url": "https://omlx.ai/benchmarks/performance/ky1o4jz5",
   "tg_tokens": null,
   "tps": 75.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bdz6od6g-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 4ts0dpej; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 862.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 862.2 33.6 16.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bdz6od6g",
   "source_url": "https://omlx.ai/benchmarks/performance/bdz6od6g",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bdz6od6g-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 4ts0dpej; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 969.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 969.9 33.2 17.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bdz6od6g",
   "source_url": "https://omlx.ai/benchmarks/performance/bdz6od6g",
   "tg_tokens": null,
   "tps": 33.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-bd829vq2-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: s2laz03v; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 402.6,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 402.6 14.7 19.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bd829vq2",
   "source_url": "https://omlx.ai/benchmarks/performance/bd829vq2",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-bd829vq2-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: s2laz03v; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 422.2,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 422.2 14.5 21.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bd829vq2",
   "source_url": "https://omlx.ai/benchmarks/performance/bd829vq2",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gyx6j9j4-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 820.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 820.5 52.5 38.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gyx6j9j4",
   "source_url": "https://omlx.ai/benchmarks/performance/gyx6j9j4",
   "tg_tokens": null,
   "tps": 52.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gyx6j9j4-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 41.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 913.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 913.0 43.6 41.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gyx6j9j4",
   "source_url": "https://omlx.ai/benchmarks/performance/gyx6j9j4",
   "tg_tokens": null,
   "tps": 43.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gyx6j9j4-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 41.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 843.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 843.7 51.2 41.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gyx6j9j4",
   "source_url": "https://omlx.ai/benchmarks/performance/gyx6j9j4",
   "tg_tokens": null,
   "tps": 51.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-gyx6j9j4-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 41.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 783.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 783.6 54.3 41.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gyx6j9j4",
   "source_url": "https://omlx.ai/benchmarks/performance/gyx6j9j4",
   "tg_tokens": null,
   "tps": 54.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gyx6j9j4-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 42.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 711.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 711.8 54.5 42.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gyx6j9j4",
   "source_url": "https://omlx.ai/benchmarks/performance/gyx6j9j4",
   "tg_tokens": null,
   "tps": 54.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-5pnksdzn-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: ks8t5j47, p1ryvn88, uvoog950; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 298.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 1k 298.1 30.3 16.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5pnksdzn",
   "source_url": "https://omlx.ai/benchmarks/performance/5pnksdzn",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-5pnksdzn-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: ks8t5j47, p1ryvn88, uvoog950; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 321.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 4k 321.6 30.5 20.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5pnksdzn",
   "source_url": "https://omlx.ai/benchmarks/performance/5pnksdzn",
   "tg_tokens": null,
   "tps": 30.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-5pnksdzn-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: ks8t5j47, p1ryvn88, uvoog950; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 314.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 8k 314.0 29.5 20.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5pnksdzn",
   "source_url": "https://omlx.ai/benchmarks/performance/5pnksdzn",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-5pnksdzn-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: ks8t5j47, p1ryvn88, uvoog950; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 300.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 16k 300.1 28.5 22.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5pnksdzn",
   "source_url": "https://omlx.ai/benchmarks/performance/5pnksdzn",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-58ge0t1r-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 120lhgqu, fi266zu9, g5zd89ic, rvvb1l14, veg1ycuj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 908.5,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 908.5 48.2 23.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/58ge0t1r",
   "source_url": "https://omlx.ai/benchmarks/performance/58ge0t1r",
   "tg_tokens": null,
   "tps": 48.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-58ge0t1r-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: 120lhgqu, fi266zu9, g5zd89ic, rvvb1l14, veg1ycuj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 796.5,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 796.5 37.7 24.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/58ge0t1r",
   "source_url": "https://omlx.ai/benchmarks/performance/58ge0t1r",
   "tg_tokens": null,
   "tps": 37.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-58ge0t1r-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.7 GB at this context; duplicate traces of the same session: 120lhgqu, fi266zu9, g5zd89ic, rvvb1l14, veg1ycuj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 711.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 711.4 42.7 25.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/58ge0t1r",
   "source_url": "https://omlx.ai/benchmarks/performance/58ge0t1r",
   "tg_tokens": null,
   "tps": 42.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-58ge0t1r-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.1 GB at this context; duplicate traces of the same session: 120lhgqu, fi266zu9, g5zd89ic, rvvb1l14, veg1ycuj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 682.5,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 682.5 33.1 27.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/58ge0t1r",
   "source_url": "https://omlx.ai/benchmarks/performance/58ge0t1r",
   "tg_tokens": null,
   "tps": 33.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-58ge0t1r-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: 120lhgqu, fi266zu9, g5zd89ic, rvvb1l14, veg1ycuj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 618.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 618.7 23.4 29.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/58ge0t1r",
   "source_url": "https://omlx.ai/benchmarks/performance/58ge0t1r",
   "tg_tokens": null,
   "tps": 23.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-58ge0t1r-ctx65536",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.4 GB at this context; duplicate traces of the same session: 120lhgqu, fi266zu9, g5zd89ic, rvvb1l14, veg1ycuj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 566.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 566.9 13.6 34.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/58ge0t1r",
   "source_url": "https://omlx.ai/benchmarks/performance/58ge0t1r",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-giljj1i6-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.7 GB at this context; duplicate traces of the same session: 9reiywmh; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 650.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 650.7 31.9 4.7 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/giljj1i6",
   "source_url": "https://omlx.ai/benchmarks/performance/giljj1i6",
   "tg_tokens": null,
   "tps": 31.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-giljj1i6-ctx8192",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.8 GB at this context; duplicate traces of the same session: 9reiywmh; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 648.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 8k 648.8 30.7 4.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/giljj1i6",
   "source_url": "https://omlx.ai/benchmarks/performance/giljj1i6",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-4hj35msa-ctx4096",
   "model": "Huihui-Qwen3.8-27B-abliterated",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 1t08i21i; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 52.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Huihui-Qwen3.8-27B-abliterated-mlx-4Bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 52.5 9.7 17.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4hj35msa",
   "source_url": "https://omlx.ai/benchmarks/performance/4hj35msa",
   "tg_tokens": null,
   "tps": 9.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-4hj35msa-ctx16384",
   "model": "Huihui-Qwen3.8-27B-abliterated",
   "notes": "oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: 1t08i21i; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 46.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Huihui-Qwen3.8-27B-abliterated-mlx-4Bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k 46.1 8.1 18.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4hj35msa",
   "source_url": "https://omlx.ai/benchmarks/performance/4hj35msa",
   "tg_tokens": null,
   "tps": 8.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-kpx70e2t-ctx4096",
   "model": "Llama-3.2-3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 2.8 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4203.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Llama-3.2-3B-Instruct-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 4,203 109.6 2.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kpx70e2t",
   "source_url": "https://omlx.ai/benchmarks/performance/kpx70e2t",
   "tg_tokens": null,
   "tps": 109.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-kpx70e2t-ctx16384",
   "model": "Llama-3.2-3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3278.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Llama-3.2-3B-Instruct-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 3,278 68.7 3.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kpx70e2t",
   "source_url": "https://omlx.ai/benchmarks/performance/kpx70e2t",
   "tg_tokens": null,
   "tps": 68.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ldr8ls8l-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V13-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 0085ztuo, jixeen7n, rl2a3z16, zxlxqzf6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 970.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V13-dequantized-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 970.7 109.5 21.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ldr8ls8l",
   "source_url": "https://omlx.ai/benchmarks/performance/ldr8ls8l",
   "tg_tokens": null,
   "tps": 109.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ldr8ls8l-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V13-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: 0085ztuo, jixeen7n, rl2a3z16, zxlxqzf6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1664.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V13-dequantized-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,664 106.7 23.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ldr8ls8l",
   "source_url": "https://omlx.ai/benchmarks/performance/ldr8ls8l",
   "tg_tokens": null,
   "tps": 106.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ldr8ls8l-ctx8192",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V13-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: 0085ztuo, jixeen7n, rl2a3z16, zxlxqzf6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1547.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V13-dequantized-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,547 104.5 24.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ldr8ls8l",
   "source_url": "https://omlx.ai/benchmarks/performance/ldr8ls8l",
   "tg_tokens": null,
   "tps": 104.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ldr8ls8l-ctx16384",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V13-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 0085ztuo, jixeen7n, rl2a3z16, zxlxqzf6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1405.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V13-dequantized-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,405 106.7 23.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ldr8ls8l",
   "source_url": "https://omlx.ai/benchmarks/performance/ldr8ls8l",
   "tg_tokens": null,
   "tps": 106.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ldr8ls8l-ctx32768",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V13-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 0085ztuo, jixeen7n, rl2a3z16, zxlxqzf6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1160.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V13-dequantized-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,160 67.1 23.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ldr8ls8l",
   "source_url": "https://omlx.ai/benchmarks/performance/ldr8ls8l",
   "tg_tokens": null,
   "tps": 67.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-lc905mce-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 6.3 GB; TTFT 4380 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 233.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k tokens 233.8 PP tok/s 25.7 TG tok/s 4380 TTFT (ms) 6.3 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lc905mce",
   "source_url": "https://omlx.ai/benchmarks/performance/lc905mce",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-b91332qf-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 725.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 725.7 47.7 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b91332qf",
   "source_url": "https://omlx.ai/benchmarks/performance/b91332qf",
   "tg_tokens": null,
   "tps": 47.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-b91332qf-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2879.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,879 118.0 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b91332qf",
   "source_url": "https://omlx.ai/benchmarks/performance/b91332qf",
   "tg_tokens": null,
   "tps": 118.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-b91332qf-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1592.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 1,592 75.1 23.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b91332qf",
   "source_url": "https://omlx.ai/benchmarks/performance/b91332qf",
   "tg_tokens": null,
   "tps": 75.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-iddv4pb3-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 521.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 521.7 47.8 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iddv4pb3",
   "source_url": "https://omlx.ai/benchmarks/performance/iddv4pb3",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-iddv4pb3-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2897.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,897 129.0 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iddv4pb3",
   "source_url": "https://omlx.ai/benchmarks/performance/iddv4pb3",
   "tg_tokens": null,
   "tps": 129.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-iddv4pb3-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1589.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 1,589 77.4 23.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iddv4pb3",
   "source_url": "https://omlx.ai/benchmarks/performance/iddv4pb3",
   "tg_tokens": null,
   "tps": 77.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-vezyompx-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: o5ze29f6, pbtzzd5l; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 462.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 462.1 45.7 21.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vezyompx",
   "source_url": "https://omlx.ai/benchmarks/performance/vezyompx",
   "tg_tokens": null,
   "tps": 45.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-vezyompx-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: o5ze29f6, pbtzzd5l; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2552.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,552 103.8 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vezyompx",
   "source_url": "https://omlx.ai/benchmarks/performance/vezyompx",
   "tg_tokens": null,
   "tps": 103.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-vezyompx-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: o5ze29f6, pbtzzd5l; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1579.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 1,579 78.6 23.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vezyompx",
   "source_url": "https://omlx.ai/benchmarks/performance/vezyompx",
   "tg_tokens": null,
   "tps": 78.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-bubtbrw3-ctx4096",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "MLX; oMLX community benchmark trace, single run at 4k context; peak memory 25.7 GB; TTFT 14888 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 275.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k tokens 275.1 PP tok/s 18.5 TG tok/s 14888 TTFT (ms) 25.7 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bubtbrw3",
   "source_url": "https://omlx.ai/benchmarks/performance/bubtbrw3",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-pomnad65-ctx1024",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "MLX; oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: yjcwz702; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 278.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 278.9 19.0 21.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pomnad65",
   "source_url": "https://omlx.ai/benchmarks/performance/pomnad65",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-pomnad65-ctx4096",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "MLX; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: yjcwz702; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 288.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 288.2 19.2 22.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pomnad65",
   "source_url": "https://omlx.ai/benchmarks/performance/pomnad65",
   "tg_tokens": null,
   "tps": 19.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-jqzj1la4-ctx1024",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "MLX; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 279.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 279.3 20.8 20.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jqzj1la4",
   "source_url": "https://omlx.ai/benchmarks/performance/jqzj1la4",
   "tg_tokens": null,
   "tps": 20.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-jqzj1la4-ctx4096",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "MLX; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 288.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 288.5 19.4 22.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jqzj1la4",
   "source_url": "https://omlx.ai/benchmarks/performance/jqzj1la4",
   "tg_tokens": null,
   "tps": 19.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-svz7dkav-ctx65536",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 2e02uoid, hzgazg1n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2744.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 2,744 66.9 17.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/svz7dkav",
   "source_url": "https://omlx.ai/benchmarks/performance/svz7dkav",
   "tg_tokens": null,
   "tps": 66.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-svz7dkav-ctx131072",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 2e02uoid, hzgazg1n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1886.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 1,886 45.5 20.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/svz7dkav",
   "source_url": "https://omlx.ai/benchmarks/performance/svz7dkav",
   "tg_tokens": null,
   "tps": 45.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-svz7dkav-ctx199680",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: 2e02uoid, hzgazg1n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1389.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 195k 1,389 34.6 24.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/svz7dkav",
   "source_url": "https://omlx.ai/benchmarks/performance/svz7dkav",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qjk7xeia-ctx65536",
   "model": "gemma-4-26B-A4B-it-qat",
   "notes": "QAT 6bit; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: wg60geum; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2416.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "gemma-4-26B-A4B-it-qat-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 64k 2,416 61.0 23.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qjk7xeia",
   "source_url": "https://omlx.ai/benchmarks/performance/qjk7xeia",
   "tg_tokens": null,
   "tps": 61.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qjk7xeia-ctx131072",
   "model": "gemma-4-26B-A4B-it-qat",
   "notes": "QAT 6bit; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: wg60geum; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1726.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "gemma-4-26B-A4B-it-qat-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 128k 1,726 43.3 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qjk7xeia",
   "source_url": "https://omlx.ai/benchmarks/performance/qjk7xeia",
   "tg_tokens": null,
   "tps": 43.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qjk7xeia-ctx199680",
   "model": "gemma-4-26B-A4B-it-qat",
   "notes": "QAT 6bit; oMLX community benchmark trace; peak memory 30.1 GB at this context; duplicate traces of the same session: wg60geum; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1300.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "gemma-4-26B-A4B-it-qat-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 195k 1,300 33.0 30.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qjk7xeia",
   "source_url": "https://omlx.ai/benchmarks/performance/qjk7xeia",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-kgh3bnll-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: r2pfg73o; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1337.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,337 48.3 14.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kgh3bnll",
   "source_url": "https://omlx.ai/benchmarks/performance/kgh3bnll",
   "tg_tokens": null,
   "tps": 48.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-kgh3bnll-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.9 GB at this context; duplicate traces of the same session: r2pfg73o; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1286.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 1,286 42.8 14.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kgh3bnll",
   "source_url": "https://omlx.ai/benchmarks/performance/kgh3bnll",
   "tg_tokens": null,
   "tps": 42.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-cy6u9pfy-ctx1024",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: 0gfcmgcg; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 341.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 341.5 18.5 32.0 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cy6u9pfy",
   "source_url": "https://omlx.ai/benchmarks/performance/cy6u9pfy",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-cy6u9pfy-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 33.1 GB at this context; duplicate traces of the same session: 0gfcmgcg; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 339.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 339.6 18.3 33.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cy6u9pfy",
   "source_url": "https://omlx.ai/benchmarks/performance/cy6u9pfy",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-cy6u9pfy-ctx8192",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 33.6 GB at this context; duplicate traces of the same session: 0gfcmgcg; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 336.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 336.2 17.8 33.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cy6u9pfy",
   "source_url": "https://omlx.ai/benchmarks/performance/cy6u9pfy",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-3ekbrobi-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 76.0 GB at this context; duplicate traces of the same session: huo8elc3, i2op5u2k; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-Flash-Next-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 4k 323.8 39.9 76.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ekbrobi",
   "source_url": "https://omlx.ai/benchmarks/performance/3ekbrobi",
   "tg_tokens": null,
   "tps": 39.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-3ekbrobi-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 76.2 GB at this context; duplicate traces of the same session: huo8elc3, i2op5u2k; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 649.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-Flash-Next-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 16k 649.1 43.3 76.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ekbrobi",
   "source_url": "https://omlx.ai/benchmarks/performance/3ekbrobi",
   "tg_tokens": null,
   "tps": 43.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-3ekbrobi-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 77.7 GB at this context; duplicate traces of the same session: huo8elc3, i2op5u2k; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 638.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-Flash-Next-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 64k 638.4 25.2 77.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ekbrobi",
   "source_url": "https://omlx.ai/benchmarks/performance/3ekbrobi",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-3ci2vvao-ctx1024",
   "model": "LFM2.5-8B-A1B",
   "notes": "gptq jang q4 MLX; oMLX community benchmark trace, single run at 1k context; peak memory 5.4 GB; TTFT 627 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1633.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-8B-A1B-gptq-mlx-jang-q4 on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k tokens 1,633 PP tok/s 146.3 TG tok/s 627 TTFT (ms) 5.4 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ci2vvao",
   "source_url": "https://omlx.ai/benchmarks/performance/3ci2vvao",
   "tg_tokens": null,
   "tps": 146.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-cmqa58rv-ctx4096",
   "model": "Llama-3.2-3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 2.8 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4203.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Llama-3.2-3B-Instruct-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 4,203 109.6 2.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cmqa58rv",
   "source_url": "https://omlx.ai/benchmarks/performance/cmqa58rv",
   "tg_tokens": null,
   "tps": 109.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-cmqa58rv-ctx16384",
   "model": "Llama-3.2-3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3278.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Llama-3.2-3B-Instruct-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 3,278 68.7 3.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cmqa58rv",
   "source_url": "https://omlx.ai/benchmarks/performance/cmqa58rv",
   "tg_tokens": null,
   "tps": 68.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-nfeeezfd-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: qa6cqo82, wvmh6fjj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 692.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 1k 692.8 70.1 29.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nfeeezfd",
   "source_url": "https://omlx.ai/benchmarks/performance/nfeeezfd",
   "tg_tokens": null,
   "tps": 70.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-nfeeezfd-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.8 GB at this context; duplicate traces of the same session: qa6cqo82, wvmh6fjj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 754.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 4k 754.5 76.7 30.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nfeeezfd",
   "source_url": "https://omlx.ai/benchmarks/performance/nfeeezfd",
   "tg_tokens": null,
   "tps": 76.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-nfeeezfd-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.1 GB at this context; duplicate traces of the same session: qa6cqo82, wvmh6fjj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 723.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 8k 723.7 67.3 31.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nfeeezfd",
   "source_url": "https://omlx.ai/benchmarks/performance/nfeeezfd",
   "tg_tokens": null,
   "tps": 67.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-nfeeezfd-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.2 GB at this context; duplicate traces of the same session: qa6cqo82, wvmh6fjj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 634.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 16k 634.3 52.6 31.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nfeeezfd",
   "source_url": "https://omlx.ai/benchmarks/performance/nfeeezfd",
   "tg_tokens": null,
   "tps": 52.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-nfeeezfd-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.5 GB at this context; duplicate traces of the same session: qa6cqo82, wvmh6fjj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 517.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 32k 517.0 48.2 31.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nfeeezfd",
   "source_url": "https://omlx.ai/benchmarks/performance/nfeeezfd",
   "tg_tokens": null,
   "tps": 48.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bszsujyv-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 30.7 GB; TTFT 1581 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5182.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k tokens 5,182 PP tok/s 168.6 TG tok/s 1581 TTFT (ms) 30.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bszsujyv",
   "source_url": "https://omlx.ai/benchmarks/performance/bszsujyv",
   "tg_tokens": null,
   "tps": 168.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-905rlzx7-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: lv9k3ywp; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 84.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 1k 84.8 15.5 16.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/905rlzx7",
   "source_url": "https://omlx.ai/benchmarks/performance/905rlzx7",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-905rlzx7-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: lv9k3ywp; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 81.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k 81.3 15.3 17.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/905rlzx7",
   "source_url": "https://omlx.ai/benchmarks/performance/905rlzx7",
   "tg_tokens": null,
   "tps": 15.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-4a2m2mj7-ctx4096",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "MLX; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 6vil8yus; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 188.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Abliterated-MLX-4-bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 188.0 25.1 18.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4a2m2mj7",
   "source_url": "https://omlx.ai/benchmarks/performance/4a2m2mj7",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-4a2m2mj7-ctx16384",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "MLX; oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 6vil8yus; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 138.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Abliterated-MLX-4-bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 138.7 16.5 20.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4a2m2mj7",
   "source_url": "https://omlx.ai/benchmarks/performance/4a2m2mj7",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qt03vmxb-ctx1024",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "MLX; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 279.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 279.3 20.8 20.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qt03vmxb",
   "source_url": "https://omlx.ai/benchmarks/performance/qt03vmxb",
   "tg_tokens": null,
   "tps": 20.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qt03vmxb-ctx4096",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "MLX; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 288.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 288.5 19.4 22.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qt03vmxb",
   "source_url": "https://omlx.ai/benchmarks/performance/qt03vmxb",
   "tg_tokens": null,
   "tps": 19.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-lbquhmf2-ctx1024",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "MLX; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: flcv4mln; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 148.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 148.8 9.2 22.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lbquhmf2",
   "source_url": "https://omlx.ai/benchmarks/performance/lbquhmf2",
   "tg_tokens": null,
   "tps": 9.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-lbquhmf2-ctx4096",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "MLX; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: flcv4mln; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 150.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 150.5 10.8 22.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lbquhmf2",
   "source_url": "https://omlx.ai/benchmarks/performance/lbquhmf2",
   "tg_tokens": null,
   "tps": 10.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-kmwupe7h-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: lkcjwb8v, z6brh6no; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 173.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 173.8 24.7 17.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kmwupe7h",
   "source_url": "https://omlx.ai/benchmarks/performance/kmwupe7h",
   "tg_tokens": null,
   "tps": 24.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-kmwupe7h-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: lkcjwb8v, z6brh6no; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 451.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 451.3 47.4 19.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kmwupe7h",
   "source_url": "https://omlx.ai/benchmarks/performance/kmwupe7h",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-kmwupe7h-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: lkcjwb8v, z6brh6no; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 376.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 376.2 33.3 23.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kmwupe7h",
   "source_url": "https://omlx.ai/benchmarks/performance/kmwupe7h",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5au20c2q-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 56.9 GB at this context; duplicate traces of the same session: 2a18q1kn, usmlaelw, zogwl55j; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 668.5,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-Flash-Next-oQ3e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 3bit; 8k 668.5 33.9 56.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5au20c2q",
   "source_url": "https://omlx.ai/benchmarks/performance/5au20c2q",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5au20c2q-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 57.4 GB at this context; duplicate traces of the same session: 2a18q1kn, usmlaelw, zogwl55j; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 970.9,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-Flash-Next-oQ3e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 3bit; 16k 970.9 37.0 57.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5au20c2q",
   "source_url": "https://omlx.ai/benchmarks/performance/5au20c2q",
   "tg_tokens": null,
   "tps": 37.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5au20c2q-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 61.9 GB at this context; duplicate traces of the same session: 2a18q1kn, usmlaelw, zogwl55j; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 959.5,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-Flash-Next-oQ3e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 3bit; 64k 959.5 34.3 61.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5au20c2q",
   "source_url": "https://omlx.ai/benchmarks/performance/5au20c2q",
   "tg_tokens": null,
   "tps": 34.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5au20c2q-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 68.0 GB at this context; duplicate traces of the same session: 2a18q1kn, usmlaelw, zogwl55j; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 925.2,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-Flash-Next-oQ3e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 3bit; 128k 925.2 31.2 68.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5au20c2q",
   "source_url": "https://omlx.ai/benchmarks/performance/5au20c2q",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-odax7p50-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.2 GB at this context; duplicate traces of the same session: w6iv3p22; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 486.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 486.0 52.0 71.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/odax7p50",
   "source_url": "https://omlx.ai/benchmarks/performance/odax7p50",
   "tg_tokens": null,
   "tps": 52.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-odax7p50-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: w6iv3p22; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 550.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 550.7 41.5 71.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/odax7p50",
   "source_url": "https://omlx.ai/benchmarks/performance/odax7p50",
   "tg_tokens": null,
   "tps": 41.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-w7ei8fd5-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: aywszqw9; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 736.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 736.6 87.7 21.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w7ei8fd5",
   "source_url": "https://omlx.ai/benchmarks/performance/w7ei8fd5",
   "tg_tokens": null,
   "tps": 87.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-w7ei8fd5-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: aywszqw9; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 855.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 855.8 101.6 21.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w7ei8fd5",
   "source_url": "https://omlx.ai/benchmarks/performance/w7ei8fd5",
   "tg_tokens": null,
   "tps": 101.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-w7ei8fd5-ctx8192",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: aywszqw9; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 806.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 806.1 89.4 22.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w7ei8fd5",
   "source_url": "https://omlx.ai/benchmarks/performance/w7ei8fd5",
   "tg_tokens": null,
   "tps": 89.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-w7ei8fd5-ctx16384",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: aywszqw9; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 714.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 714.2 68.1 22.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w7ei8fd5",
   "source_url": "https://omlx.ai/benchmarks/performance/w7ei8fd5",
   "tg_tokens": null,
   "tps": 68.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-w7ei8fd5-ctx32768",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: aywszqw9; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 577.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 577.9 67.3 23.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w7ei8fd5",
   "source_url": "https://omlx.ai/benchmarks/performance/w7ei8fd5",
   "tg_tokens": null,
   "tps": 67.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-w7ei8fd5-ctx65536",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: aywszqw9; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 419.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 64k 419.5 52.4 25.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w7ei8fd5",
   "source_url": "https://omlx.ai/benchmarks/performance/w7ei8fd5",
   "tg_tokens": null,
   "tps": 52.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-5ez4rba1-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: wxjii4bf, wykpmdm0, zxjt9rm3; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 163.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-fp16-mtp on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 1k 163.1 9.6 7.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ez4rba1",
   "source_url": "https://omlx.ai/benchmarks/performance/5ez4rba1",
   "tg_tokens": null,
   "tps": 9.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-5ez4rba1-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 8.0 GB at this context; duplicate traces of the same session: wxjii4bf, wykpmdm0, zxjt9rm3; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 264.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-fp16-mtp on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 4k 264.8 8.9 8.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ez4rba1",
   "source_url": "https://omlx.ai/benchmarks/performance/5ez4rba1",
   "tg_tokens": null,
   "tps": 8.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-5ez4rba1-ctx8192",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 8.3 GB at this context; duplicate traces of the same session: wxjii4bf, wykpmdm0, zxjt9rm3; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 261.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-fp16-mtp on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 8k 261.2 10.0 8.3 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ez4rba1",
   "source_url": "https://omlx.ai/benchmarks/performance/5ez4rba1",
   "tg_tokens": null,
   "tps": 10.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-5ez4rba1-ctx16384",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 8.3 GB at this context; duplicate traces of the same session: wxjii4bf, wykpmdm0, zxjt9rm3; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 246.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-fp16-mtp on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 16k 246.0 10.9 8.3 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ez4rba1",
   "source_url": "https://omlx.ai/benchmarks/performance/5ez4rba1",
   "tg_tokens": null,
   "tps": 10.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-5ez4rba1-ctx32768",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 8.7 GB at this context; duplicate traces of the same session: wxjii4bf, wykpmdm0, zxjt9rm3; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 222.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-fp16-mtp on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 32k 222.8 8.3 8.7 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ez4rba1",
   "source_url": "https://omlx.ai/benchmarks/performance/5ez4rba1",
   "tg_tokens": null,
   "tps": 8.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-5ez4rba1-ctx65536",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 9.3 GB at this context; duplicate traces of the same session: wxjii4bf, wykpmdm0, zxjt9rm3; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 186.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-fp16-mtp on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 64k 186.9 7.7 9.3 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ez4rba1",
   "source_url": "https://omlx.ai/benchmarks/performance/5ez4rba1",
   "tg_tokens": null,
   "tps": 7.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-5ez4rba1-ctx131072",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 10.5 GB at this context; duplicate traces of the same session: wxjii4bf, wykpmdm0, zxjt9rm3; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 140.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-fp16-mtp on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 128k 140.7 7.1 10.5 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ez4rba1",
   "source_url": "https://omlx.ai/benchmarks/performance/5ez4rba1",
   "tg_tokens": null,
   "tps": 7.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-33hak2fo-ctx65536",
   "model": "gemma-4-26B-A4B-it-qat",
   "notes": "QAT 6bit; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2416.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "gemma-4-26B-A4B-it-qat-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 64k 2,416 61.0 23.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/33hak2fo",
   "source_url": "https://omlx.ai/benchmarks/performance/33hak2fo",
   "tg_tokens": null,
   "tps": 61.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-33hak2fo-ctx131072",
   "model": "gemma-4-26B-A4B-it-qat",
   "notes": "QAT 6bit; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1726.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "gemma-4-26B-A4B-it-qat-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 128k 1,726 43.3 26.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/33hak2fo",
   "source_url": "https://omlx.ai/benchmarks/performance/33hak2fo",
   "tg_tokens": null,
   "tps": 43.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-33hak2fo-ctx199680",
   "model": "gemma-4-26B-A4B-it-qat",
   "notes": "QAT 6bit; oMLX community benchmark trace; peak memory 30.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1300.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "gemma-4-26B-A4B-it-qat-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 195k 1,300 33.0 30.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/33hak2fo",
   "source_url": "https://omlx.ai/benchmarks/performance/33hak2fo",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-m60efqzu-ctx1024",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 341.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 341.5 18.5 32.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m60efqzu",
   "source_url": "https://omlx.ai/benchmarks/performance/m60efqzu",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-m60efqzu-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 33.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 339.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 339.6 18.3 33.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m60efqzu",
   "source_url": "https://omlx.ai/benchmarks/performance/m60efqzu",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-m60efqzu-ctx8192",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 33.6 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 336.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 336.2 17.8 33.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m60efqzu",
   "source_url": "https://omlx.ai/benchmarks/performance/m60efqzu",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ksm931l8-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace, single run at 1k context; peak memory 403.3 GB; TTFT 2176 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 470.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k tokens 470.7 PP tok/s 30.6 TG tok/s 2176 TTFT (ms) 403.3 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ksm931l8",
   "source_url": "https://omlx.ai/benchmarks/performance/ksm931l8",
   "tg_tokens": null,
   "tps": 30.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qkj7in6n-ctx1024",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 177.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 530.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 530.5 40.6 177.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qkj7in6n",
   "source_url": "https://omlx.ai/benchmarks/performance/qkj7in6n",
   "tg_tokens": null,
   "tps": 40.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qkj7in6n-ctx4096",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 178.8 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 604.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 604.6 38.7 178.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qkj7in6n",
   "source_url": "https://omlx.ai/benchmarks/performance/qkj7in6n",
   "tg_tokens": null,
   "tps": 38.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qkj7in6n-ctx8192",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 179.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 558.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 558.0 38.8 179.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qkj7in6n",
   "source_url": "https://omlx.ai/benchmarks/performance/qkj7in6n",
   "tg_tokens": null,
   "tps": 38.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qkj7in6n-ctx16384",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 179.2 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 546.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 546.5 38.4 179.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qkj7in6n",
   "source_url": "https://omlx.ai/benchmarks/performance/qkj7in6n",
   "tg_tokens": null,
   "tps": 38.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qkj7in6n-ctx32768",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 179.3 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 539.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 539.0 37.9 179.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qkj7in6n",
   "source_url": "https://omlx.ai/benchmarks/performance/qkj7in6n",
   "tg_tokens": null,
   "tps": 37.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-wk6mcjna-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: o9ad9gi3; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 744.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 1k 744.2 93.2 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wk6mcjna",
   "source_url": "https://omlx.ai/benchmarks/performance/wk6mcjna",
   "tg_tokens": null,
   "tps": 93.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-wk6mcjna-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: o9ad9gi3; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 774.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 774.6 89.5 21.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wk6mcjna",
   "source_url": "https://omlx.ai/benchmarks/performance/wk6mcjna",
   "tg_tokens": null,
   "tps": 89.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-dtfnrttx-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: uwe4d2ff; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 692.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 1k 692.8 70.1 29.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dtfnrttx",
   "source_url": "https://omlx.ai/benchmarks/performance/dtfnrttx",
   "tg_tokens": null,
   "tps": 70.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-dtfnrttx-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.8 GB at this context; duplicate traces of the same session: uwe4d2ff; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 754.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 4k 754.5 76.7 30.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dtfnrttx",
   "source_url": "https://omlx.ai/benchmarks/performance/dtfnrttx",
   "tg_tokens": null,
   "tps": 76.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-dtfnrttx-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.1 GB at this context; duplicate traces of the same session: uwe4d2ff; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 723.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 8k 723.7 67.3 31.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dtfnrttx",
   "source_url": "https://omlx.ai/benchmarks/performance/dtfnrttx",
   "tg_tokens": null,
   "tps": 67.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-dtfnrttx-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.2 GB at this context; duplicate traces of the same session: uwe4d2ff; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 634.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 16k 634.3 52.6 31.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dtfnrttx",
   "source_url": "https://omlx.ai/benchmarks/performance/dtfnrttx",
   "tg_tokens": null,
   "tps": 52.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-dtfnrttx-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.5 GB at this context; duplicate traces of the same session: uwe4d2ff; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 517.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 32k 517.0 48.2 31.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dtfnrttx",
   "source_url": "https://omlx.ai/benchmarks/performance/dtfnrttx",
   "tg_tokens": null,
   "tps": 48.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-62euczk3-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 30.7 GB; TTFT 1601 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5116.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k tokens 5,116 PP tok/s 160.8 TG tok/s 1601 TTFT (ms) 30.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/62euczk3",
   "source_url": "https://omlx.ai/benchmarks/performance/62euczk3",
   "tg_tokens": null,
   "tps": 160.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-hztdmv3c-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 5vj2funr; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 441.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 441.8 20.2 16.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hztdmv3c",
   "source_url": "https://omlx.ai/benchmarks/performance/hztdmv3c",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-hztdmv3c-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 5vj2funr; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 545.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 545.6 19.9 17.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hztdmv3c",
   "source_url": "https://omlx.ai/benchmarks/performance/hztdmv3c",
   "tg_tokens": null,
   "tps": 19.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-1cjd1k0w-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: qk0m3zrx; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 377.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 377.2 19.9 16.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1cjd1k0w",
   "source_url": "https://omlx.ai/benchmarks/performance/1cjd1k0w",
   "tg_tokens": null,
   "tps": 19.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-1cjd1k0w-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: qk0m3zrx; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 474.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 474.4 19.9 17.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1cjd1k0w",
   "source_url": "https://omlx.ai/benchmarks/performance/1cjd1k0w",
   "tg_tokens": null,
   "tps": 19.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-sfsz2vgl-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: d3ioatc9, eywdsp0f, lj4z0qtn, vu05qxjd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 217.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 217.6 29.9 18.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sfsz2vgl",
   "source_url": "https://omlx.ai/benchmarks/performance/sfsz2vgl",
   "tg_tokens": null,
   "tps": 29.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-sfsz2vgl-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: d3ioatc9, eywdsp0f, lj4z0qtn, vu05qxjd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 222.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 222.6 42.1 22.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sfsz2vgl",
   "source_url": "https://omlx.ai/benchmarks/performance/sfsz2vgl",
   "tg_tokens": null,
   "tps": 42.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-sfsz2vgl-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: d3ioatc9, eywdsp0f, lj4z0qtn, vu05qxjd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 220.4 49.1 23.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sfsz2vgl",
   "source_url": "https://omlx.ai/benchmarks/performance/sfsz2vgl",
   "tg_tokens": null,
   "tps": 49.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-sfsz2vgl-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: d3ioatc9, eywdsp0f, lj4z0qtn, vu05qxjd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 212.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 212.3 60.7 22.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sfsz2vgl",
   "source_url": "https://omlx.ai/benchmarks/performance/sfsz2vgl",
   "tg_tokens": null,
   "tps": 60.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-sfsz2vgl-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: d3ioatc9, eywdsp0f, lj4z0qtn, vu05qxjd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 196.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 196.5 61.2 23.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sfsz2vgl",
   "source_url": "https://omlx.ai/benchmarks/performance/sfsz2vgl",
   "tg_tokens": null,
   "tps": 61.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-4lciq5hi-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: dzh8j6ll, kfrh1956; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 171.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 171.5 25.9 17.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4lciq5hi",
   "source_url": "https://omlx.ai/benchmarks/performance/4lciq5hi",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-4lciq5hi-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: dzh8j6ll, kfrh1956; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 462.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 462.3 55.4 19.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4lciq5hi",
   "source_url": "https://omlx.ai/benchmarks/performance/4lciq5hi",
   "tg_tokens": null,
   "tps": 55.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-4lciq5hi-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: dzh8j6ll, kfrh1956; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 381.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 381.7 30.0 23.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4lciq5hi",
   "source_url": "https://omlx.ai/benchmarks/performance/4lciq5hi",
   "tg_tokens": null,
   "tps": 30.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-e3hcz5f0-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: 9ojaknxt, asojiji0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 172.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 172.2 21.7 17.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e3hcz5f0",
   "source_url": "https://omlx.ai/benchmarks/performance/e3hcz5f0",
   "tg_tokens": null,
   "tps": 21.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-e3hcz5f0-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 9ojaknxt, asojiji0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 418.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 418.0 41.4 19.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e3hcz5f0",
   "source_url": "https://omlx.ai/benchmarks/performance/e3hcz5f0",
   "tg_tokens": null,
   "tps": 41.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-e3hcz5f0-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: 9ojaknxt, asojiji0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 376.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 376.1 30.6 23.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e3hcz5f0",
   "source_url": "https://omlx.ai/benchmarks/performance/e3hcz5f0",
   "tg_tokens": null,
   "tps": 30.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-aw1152eu-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: lah0e7d9, lcf3ol9y; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 163.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-fp16-mtp on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 1k 163.1 9.6 7.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aw1152eu",
   "source_url": "https://omlx.ai/benchmarks/performance/aw1152eu",
   "tg_tokens": null,
   "tps": 9.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-aw1152eu-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 8.0 GB at this context; duplicate traces of the same session: lah0e7d9, lcf3ol9y; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 264.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-fp16-mtp on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 4k 264.8 8.9 8.0 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aw1152eu",
   "source_url": "https://omlx.ai/benchmarks/performance/aw1152eu",
   "tg_tokens": null,
   "tps": 8.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-aw1152eu-ctx8192",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 8.3 GB at this context; duplicate traces of the same session: lah0e7d9, lcf3ol9y; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 261.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-fp16-mtp on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 8k 261.2 10.0 8.3 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aw1152eu",
   "source_url": "https://omlx.ai/benchmarks/performance/aw1152eu",
   "tg_tokens": null,
   "tps": 10.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-aw1152eu-ctx16384",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 8.3 GB at this context; duplicate traces of the same session: lah0e7d9, lcf3ol9y; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 246.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-fp16-mtp on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 16k 246.0 10.9 8.3 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aw1152eu",
   "source_url": "https://omlx.ai/benchmarks/performance/aw1152eu",
   "tg_tokens": null,
   "tps": 10.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-aw1152eu-ctx32768",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 8.7 GB at this context; duplicate traces of the same session: lah0e7d9, lcf3ol9y; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 222.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-fp16-mtp on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 32k 222.8 8.3 8.7 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aw1152eu",
   "source_url": "https://omlx.ai/benchmarks/performance/aw1152eu",
   "tg_tokens": null,
   "tps": 8.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-aw1152eu-ctx65536",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 9.3 GB at this context; duplicate traces of the same session: lah0e7d9, lcf3ol9y; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 186.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-fp16-mtp on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 64k 186.9 7.7 9.3 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aw1152eu",
   "source_url": "https://omlx.ai/benchmarks/performance/aw1152eu",
   "tg_tokens": null,
   "tps": 7.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-aw1152eu-ctx131072",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 10.5 GB at this context; duplicate traces of the same session: lah0e7d9, lcf3ol9y; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 140.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-fp16-mtp on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 128k 140.7 7.1 10.5 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aw1152eu",
   "source_url": "https://omlx.ai/benchmarks/performance/aw1152eu",
   "tg_tokens": null,
   "tps": 7.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-cd4xrvkx-ctx1024",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: 3jj23vyu, 8st0yb13; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2074.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 2,074 85.5 25.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cd4xrvkx",
   "source_url": "https://omlx.ai/benchmarks/performance/cd4xrvkx",
   "tg_tokens": null,
   "tps": 85.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-cd4xrvkx-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: 3jj23vyu, 8st0yb13; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2253.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 2,253 83.0 26.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cd4xrvkx",
   "source_url": "https://omlx.ai/benchmarks/performance/cd4xrvkx",
   "tg_tokens": null,
   "tps": 83.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-cd4xrvkx-ctx8192",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: 3jj23vyu, 8st0yb13; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2213.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 2,213 79.5 26.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cd4xrvkx",
   "source_url": "https://omlx.ai/benchmarks/performance/cd4xrvkx",
   "tg_tokens": null,
   "tps": 79.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-tsbxgrlv-ctx4096",
   "model": "AREX-Turbo",
   "notes": "oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: xesnc8qg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1097.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "AREX-Turbo-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,097 89.4 3.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tsbxgrlv",
   "source_url": "https://omlx.ai/benchmarks/performance/tsbxgrlv",
   "tg_tokens": null,
   "tps": 89.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-tsbxgrlv-ctx16384",
   "model": "AREX-Turbo",
   "notes": "oMLX community benchmark trace; peak memory 4.4 GB at this context; duplicate traces of the same session: xesnc8qg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1055.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "AREX-Turbo-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 1,055 79.0 4.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tsbxgrlv",
   "source_url": "https://omlx.ai/benchmarks/performance/tsbxgrlv",
   "tg_tokens": null,
   "tps": 79.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-kpdewe5q-ctx1024",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 177.7 GB at this context; duplicate traces of the same session: 745euup1, f650ur3n, o3m4prkq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 530.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 530.5 40.6 177.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kpdewe5q",
   "source_url": "https://omlx.ai/benchmarks/performance/kpdewe5q",
   "tg_tokens": null,
   "tps": 40.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-kpdewe5q-ctx4096",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 178.8 GB at this context; duplicate traces of the same session: 745euup1, f650ur3n, o3m4prkq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 604.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 604.6 38.7 178.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kpdewe5q",
   "source_url": "https://omlx.ai/benchmarks/performance/kpdewe5q",
   "tg_tokens": null,
   "tps": 38.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-kpdewe5q-ctx8192",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 179.1 GB at this context; duplicate traces of the same session: 745euup1, f650ur3n, o3m4prkq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 558.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 558.0 38.8 179.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kpdewe5q",
   "source_url": "https://omlx.ai/benchmarks/performance/kpdewe5q",
   "tg_tokens": null,
   "tps": 38.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-kpdewe5q-ctx16384",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 179.2 GB at this context; duplicate traces of the same session: 745euup1, f650ur3n, o3m4prkq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 546.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 546.5 38.4 179.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kpdewe5q",
   "source_url": "https://omlx.ai/benchmarks/performance/kpdewe5q",
   "tg_tokens": null,
   "tps": 38.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-kpdewe5q-ctx32768",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 179.3 GB at this context; duplicate traces of the same session: 745euup1, f650ur3n, o3m4prkq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 539.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 539.0 37.9 179.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kpdewe5q",
   "source_url": "https://omlx.ai/benchmarks/performance/kpdewe5q",
   "tg_tokens": null,
   "tps": 37.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-msuwpfet-ctx1024",
   "model": "Nanbeige4.2-3B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: xbsdk9jg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 190.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Nanbeige4.2-3B-mlx-8bit on M1 Pro (16c); M1 Pro (16c) \u00b7 16 GB \u00b7 8bit; 1k 190.7 16.9 4.9 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/msuwpfet",
   "source_url": "https://omlx.ai/benchmarks/performance/msuwpfet",
   "tg_tokens": null,
   "tps": 16.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-msuwpfet-ctx4096",
   "model": "Nanbeige4.2-3B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 5.5 GB at this context; duplicate traces of the same session: xbsdk9jg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 179.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Nanbeige4.2-3B-mlx-8bit on M1 Pro (16c); M1 Pro (16c) \u00b7 16 GB \u00b7 8bit; 4k 179.7 14.4 5.5 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/msuwpfet",
   "source_url": "https://omlx.ai/benchmarks/performance/msuwpfet",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-bc2dbgon-ctx4096",
   "model": "Qwen3 8B",
   "notes": "oMLX community benchmark trace; peak memory 5.5 GB at this context; duplicate traces of the same session: hras25o6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 450.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-8B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 24 GB \u00b7 4bit; 4k 450.0 46.0 5.5 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bc2dbgon",
   "source_url": "https://omlx.ai/benchmarks/performance/bc2dbgon",
   "tg_tokens": null,
   "tps": 46.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-bc2dbgon-ctx16384",
   "model": "Qwen3 8B",
   "notes": "oMLX community benchmark trace; peak memory 7.1 GB at this context; duplicate traces of the same session: hras25o6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 362.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-8B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 24 GB \u00b7 4bit; 16k 362.6 31.1 7.1 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bc2dbgon",
   "source_url": "https://omlx.ai/benchmarks/performance/bc2dbgon",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-iz8cggz7-ctx1024",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 9.6 GB; TTFT 6393 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 160.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 16 GB \u00b7 4bit; 1k tokens 160.2 PP tok/s 8.1 TG tok/s 6393 TTFT (ms) 9.6 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iz8cggz7",
   "source_url": "https://omlx.ai/benchmarks/performance/iz8cggz7",
   "tg_tokens": null,
   "tps": 8.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-jccsx0z0-ctx1024",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 8.7 GB; TTFT 5555 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 184.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 16 GB \u00b7 4bit; 1k tokens 184.3 PP tok/s 12.2 TG tok/s 5555 TTFT (ms) 8.7 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jccsx0z0",
   "source_url": "https://omlx.ai/benchmarks/performance/jccsx0z0",
   "tg_tokens": null,
   "tps": 12.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-aom33chl-ctx1024",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 9.6 GB; TTFT 6375 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 160.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 16 GB \u00b7 4bit; 1k tokens 160.6 PP tok/s 12.6 TG tok/s 6375 TTFT (ms) 9.6 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aom33chl",
   "source_url": "https://omlx.ai/benchmarks/performance/aom33chl",
   "tg_tokens": null,
   "tps": 12.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-nwcr82jw-ctx1024",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 9.6 GB; TTFT 6502 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 157.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 16 GB \u00b7 4bit; 1k tokens 157.5 PP tok/s 10.4 TG tok/s 6502 TTFT (ms) 9.6 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nwcr82jw",
   "source_url": "https://omlx.ai/benchmarks/performance/nwcr82jw",
   "tg_tokens": null,
   "tps": 10.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-oguu6uf1-ctx1024",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 9.6 GB; TTFT 6448 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 158.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 16 GB \u00b7 4bit; 1k tokens 158.8 PP tok/s 13.3 TG tok/s 6448 TTFT (ms) 9.6 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oguu6uf1",
   "source_url": "https://omlx.ai/benchmarks/performance/oguu6uf1",
   "tg_tokens": null,
   "tps": 13.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-iddg2ryq-ctx1024",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 7.9 GB; TTFT 5613 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 182.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 16 GB \u00b7 4bit; 1k tokens 182.4 PP tok/s 11.2 TG tok/s 5613 TTFT (ms) 7.9 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iddg2ryq",
   "source_url": "https://omlx.ai/benchmarks/performance/iddg2ryq",
   "tg_tokens": null,
   "tps": 11.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-hadt4br8-ctx1024",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 9.6 GB; TTFT 6537 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 156.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 16 GB \u00b7 4bit; 1k tokens 156.7 PP tok/s 10.4 TG tok/s 6537 TTFT (ms) 9.6 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hadt4br8",
   "source_url": "https://omlx.ai/benchmarks/performance/hadt4br8",
   "tg_tokens": null,
   "tps": 10.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-5crfz2cp-ctx1024",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 7.9 GB; TTFT 5714 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 179.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 16 GB \u00b7 4bit; 1k tokens 179.2 PP tok/s 11.6 TG tok/s 5714 TTFT (ms) 7.9 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5crfz2cp",
   "source_url": "https://omlx.ai/benchmarks/performance/5crfz2cp",
   "tg_tokens": null,
   "tps": 11.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-m1i2cqis-ctx1024",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 7.9 GB; TTFT 5737 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 178.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 16 GB \u00b7 4bit; 1k tokens 178.5 PP tok/s 11.4 TG tok/s 5737 TTFT (ms) 7.9 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m1i2cqis",
   "source_url": "https://omlx.ai/benchmarks/performance/m1i2cqis",
   "tg_tokens": null,
   "tps": 11.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-g3liw97j-ctx1024",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 8.7 GB; TTFT 6255 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 163.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 16 GB \u00b7 4bit; 1k tokens 163.7 PP tok/s 10.6 TG tok/s 6255 TTFT (ms) 8.7 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g3liw97j",
   "source_url": "https://omlx.ai/benchmarks/performance/g3liw97j",
   "tg_tokens": null,
   "tps": 10.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-gdzqtsco-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: af1z8r03; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 110.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 24 GB \u00b7 4bit; 1k 110.9 22.4 17.1 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gdzqtsco",
   "source_url": "https://omlx.ai/benchmarks/performance/gdzqtsco",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-gdzqtsco-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: af1z8r03; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 129.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 24 GB \u00b7 4bit; 4k 129.9 25.8 18.7 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gdzqtsco",
   "source_url": "https://omlx.ai/benchmarks/performance/gdzqtsco",
   "tg_tokens": null,
   "tps": 25.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-d1aveahw-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: muzvojje, qf5pr47n, v6kfe3p8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 214.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 214.8 50.8 18.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d1aveahw",
   "source_url": "https://omlx.ai/benchmarks/performance/d1aveahw",
   "tg_tokens": null,
   "tps": 50.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-d1aveahw-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: muzvojje, qf5pr47n, v6kfe3p8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 227.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 227.0 46.3 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d1aveahw",
   "source_url": "https://omlx.ai/benchmarks/performance/d1aveahw",
   "tg_tokens": null,
   "tps": 46.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-d1aveahw-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: muzvojje, qf5pr47n, v6kfe3p8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 222.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 222.5 39.8 20.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d1aveahw",
   "source_url": "https://omlx.ai/benchmarks/performance/d1aveahw",
   "tg_tokens": null,
   "tps": 39.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-d1aveahw-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: muzvojje, qf5pr47n, v6kfe3p8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 213.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 213.9 100.6 21.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d1aveahw",
   "source_url": "https://omlx.ai/benchmarks/performance/d1aveahw",
   "tg_tokens": null,
   "tps": 100.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-h60z35o6-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: hly52y3h, urkaucme, xzyrr7fi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 234.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 234.5 27.4 16.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h60z35o6",
   "source_url": "https://omlx.ai/benchmarks/performance/h60z35o6",
   "tg_tokens": null,
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-h60z35o6-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: hly52y3h, urkaucme, xzyrr7fi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 244.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 244.3 26.3 17.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h60z35o6",
   "source_url": "https://omlx.ai/benchmarks/performance/h60z35o6",
   "tg_tokens": null,
   "tps": 26.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-h60z35o6-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: hly52y3h, urkaucme, xzyrr7fi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 244.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 244.8 24.9 18.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h60z35o6",
   "source_url": "https://omlx.ai/benchmarks/performance/h60z35o6",
   "tg_tokens": null,
   "tps": 24.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-h60z35o6-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: hly52y3h, urkaucme, xzyrr7fi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 220.5 22.8 19.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h60z35o6",
   "source_url": "https://omlx.ai/benchmarks/performance/h60z35o6",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-fl9nn305-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 179.5 GB; TTFT 28711 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2283.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k tokens 2,283 PP tok/s 56.6 TG tok/s 28711 TTFT (ms) 179.5 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fl9nn305",
   "source_url": "https://omlx.ai/benchmarks/performance/fl9nn305",
   "tg_tokens": null,
   "tps": 56.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-lb0davtm-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 694.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 1k 694.3 72.3 29.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lb0davtm",
   "source_url": "https://omlx.ai/benchmarks/performance/lb0davtm",
   "tg_tokens": null,
   "tps": 72.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-lb0davtm-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.8 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 756.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 4k 756.0 72.7 30.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lb0davtm",
   "source_url": "https://omlx.ai/benchmarks/performance/lb0davtm",
   "tg_tokens": null,
   "tps": 72.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-lb0davtm-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.1 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 724.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 8k 724.1 69.6 31.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lb0davtm",
   "source_url": "https://omlx.ai/benchmarks/performance/lb0davtm",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-lb0davtm-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.2 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 638.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 16k 638.9 57.3 31.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lb0davtm",
   "source_url": "https://omlx.ai/benchmarks/performance/lb0davtm",
   "tg_tokens": null,
   "tps": 57.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-lb0davtm-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.5 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 517.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 32k 517.7 48.7 31.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lb0davtm",
   "source_url": "https://omlx.ai/benchmarks/performance/lb0davtm",
   "tg_tokens": null,
   "tps": 48.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-lb0davtm-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.2 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 375.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 64k 375.0 23.1 32.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lb0davtm",
   "source_url": "https://omlx.ai/benchmarks/performance/lb0davtm",
   "tg_tokens": null,
   "tps": 23.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-o15m8q1c-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 20.8 GB; TTFT 2118 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3869.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k tokens 3,869 PP tok/s 126.8 TG tok/s 2118 TTFT (ms) 20.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o15m8q1c",
   "source_url": "https://omlx.ai/benchmarks/performance/o15m8q1c",
   "tg_tokens": null,
   "tps": 126.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vokzfxfm-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 22.1 GB; TTFT 1387 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5906.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k tokens 5,906 PP tok/s 123.4 TG tok/s 1387 TTFT (ms) 22.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vokzfxfm",
   "source_url": "https://omlx.ai/benchmarks/performance/vokzfxfm",
   "tg_tokens": null,
   "tps": 123.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9xnlr2x7-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 22.1 GB; TTFT 1387 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5908.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k tokens 5,908 PP tok/s 194.9 TG tok/s 1387 TTFT (ms) 22.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9xnlr2x7",
   "source_url": "https://omlx.ai/benchmarks/performance/9xnlr2x7",
   "tg_tokens": null,
   "tps": 194.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-eou81o29-ctx1024",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 7.0 GB; TTFT 5633 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 181.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 16 GB \u00b7 4bit; 1k tokens 181.8 PP tok/s 7.9 TG tok/s 5633 TTFT (ms) 7.0 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eou81o29",
   "source_url": "https://omlx.ai/benchmarks/performance/eou81o29",
   "tg_tokens": null,
   "tps": 7.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-7izezyuq-ctx1024",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 9.6 GB; TTFT 6151 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 166.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 16 GB \u00b7 4bit; 1k tokens 166.5 PP tok/s 10.1 TG tok/s 6151 TTFT (ms) 9.6 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7izezyuq",
   "source_url": "https://omlx.ai/benchmarks/performance/7izezyuq",
   "tg_tokens": null,
   "tps": 10.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-h3uuxxl0-ctx1024",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 9.6 GB; TTFT 7143 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 143.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 16 GB \u00b7 4bit; 1k tokens 143.4 PP tok/s 9.7 TG tok/s 7143 TTFT (ms) 9.6 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h3uuxxl0",
   "source_url": "https://omlx.ai/benchmarks/performance/h3uuxxl0",
   "tg_tokens": null,
   "tps": 9.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-0jdebxmc-ctx1024",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 9.6 GB; TTFT 6490 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 157.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 16 GB \u00b7 4bit; 1k tokens 157.8 PP tok/s 10.0 TG tok/s 6490 TTFT (ms) 9.6 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0jdebxmc",
   "source_url": "https://omlx.ai/benchmarks/performance/0jdebxmc",
   "tg_tokens": null,
   "tps": 10.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-w8yovbe4-ctx1024",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 9.6 GB; TTFT 7629 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 134.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 16 GB \u00b7 4bit; 1k tokens 134.2 PP tok/s 11.5 TG tok/s 7629 TTFT (ms) 9.6 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w8yovbe4",
   "source_url": "https://omlx.ai/benchmarks/performance/w8yovbe4",
   "tg_tokens": null,
   "tps": 11.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-g3y0qouu-ctx1024",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 7.0 GB; TTFT 5776 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 177.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 16 GB \u00b7 4bit; 1k tokens 177.3 PP tok/s 8.1 TG tok/s 5776 TTFT (ms) 7.0 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g3y0qouu",
   "source_url": "https://omlx.ai/benchmarks/performance/g3y0qouu",
   "tg_tokens": null,
   "tps": 8.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-o440lede-ctx1024",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 7.0 GB; TTFT 6048 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 169.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 16 GB \u00b7 4bit; 1k tokens 169.3 PP tok/s 12.9 TG tok/s 6048 TTFT (ms) 7.0 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o440lede",
   "source_url": "https://omlx.ai/benchmarks/performance/o440lede",
   "tg_tokens": null,
   "tps": 12.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-04m7bi50-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 00ns9qpy, 0c3j1ek8, 6lielgog, rtb6zr9y, yv69ut4p; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 377.2,
   "provenance": "community",
   "quant": "OptiQ 4bit",
   "quote": "Qwen3.8-27B-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 377.2 14.5 23.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/04m7bi50",
   "source_url": "https://omlx.ai/benchmarks/performance/04m7bi50",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-04m7bi50-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: 00ns9qpy, 0c3j1ek8, 6lielgog, rtb6zr9y, yv69ut4p; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 434.3,
   "provenance": "community",
   "quant": "OptiQ 4bit",
   "quote": "Qwen3.8-27B-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 434.3 14.4 24.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/04m7bi50",
   "source_url": "https://omlx.ai/benchmarks/performance/04m7bi50",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-04m7bi50-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: 00ns9qpy, 0c3j1ek8, 6lielgog, rtb6zr9y, yv69ut4p; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 427.9,
   "provenance": "community",
   "quant": "OptiQ 4bit",
   "quote": "Qwen3.8-27B-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 427.9 14.1 24.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/04m7bi50",
   "source_url": "https://omlx.ai/benchmarks/performance/04m7bi50",
   "tg_tokens": null,
   "tps": 14.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-04m7bi50-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: 00ns9qpy, 0c3j1ek8, 6lielgog, rtb6zr9y, yv69ut4p; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 366.6,
   "provenance": "community",
   "quant": "OptiQ 4bit",
   "quote": "Qwen3.8-27B-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 366.6 13.7 25.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/04m7bi50",
   "source_url": "https://omlx.ai/benchmarks/performance/04m7bi50",
   "tg_tokens": null,
   "tps": 13.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-04m7bi50-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: 00ns9qpy, 0c3j1ek8, 6lielgog, rtb6zr9y, yv69ut4p; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 348.3,
   "provenance": "community",
   "quant": "OptiQ 4bit",
   "quote": "Qwen3.8-27B-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 348.3 12.9 22.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/04m7bi50",
   "source_url": "https://omlx.ai/benchmarks/performance/04m7bi50",
   "tg_tokens": null,
   "tps": 12.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-04m7bi50-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: 00ns9qpy, 0c3j1ek8, 6lielgog, rtb6zr9y, yv69ut4p; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 299.8,
   "provenance": "community",
   "quant": "OptiQ 4bit",
   "quote": "Qwen3.8-27B-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 64k 299.8 11.4 24.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/04m7bi50",
   "source_url": "https://omlx.ai/benchmarks/performance/04m7bi50",
   "tg_tokens": null,
   "tps": 11.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-tcoohq6i-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: bdx9i3uj, bs3cijr4, hnab1ru5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 228.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 228.1 27.4 16.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tcoohq6i",
   "source_url": "https://omlx.ai/benchmarks/performance/tcoohq6i",
   "tg_tokens": null,
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-tcoohq6i-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: bdx9i3uj, bs3cijr4, hnab1ru5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 241.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 241.0 26.4 17.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tcoohq6i",
   "source_url": "https://omlx.ai/benchmarks/performance/tcoohq6i",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-tcoohq6i-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: bdx9i3uj, bs3cijr4, hnab1ru5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 240.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 240.3 25.0 18.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tcoohq6i",
   "source_url": "https://omlx.ai/benchmarks/performance/tcoohq6i",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-tcoohq6i-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: bdx9i3uj, bs3cijr4, hnab1ru5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 221.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 221.7 23.2 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tcoohq6i",
   "source_url": "https://omlx.ai/benchmarks/performance/tcoohq6i",
   "tg_tokens": null,
   "tps": 23.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-yqf40om0-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "QAT OptiQ build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2289.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,289 60.0 20.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yqf40om0",
   "source_url": "https://omlx.ai/benchmarks/performance/yqf40om0",
   "tg_tokens": null,
   "tps": 60.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-yqf40om0-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "QAT OptiQ build; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2388.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,388 58.4 20.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yqf40om0",
   "source_url": "https://omlx.ai/benchmarks/performance/yqf40om0",
   "tg_tokens": null,
   "tps": 58.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-yqf40om0-ctx8192",
   "model": "gemma4 26B.A4B",
   "notes": "QAT OptiQ build; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2322.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 2,322 56.2 20.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yqf40om0",
   "source_url": "https://omlx.ai/benchmarks/performance/yqf40om0",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-e09egwcs-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.1 GB at this context; duplicate traces of the same session: 0rrghqxu, sa79fojk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2461.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,461 85.1 14.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e09egwcs",
   "source_url": "https://omlx.ai/benchmarks/performance/e09egwcs",
   "tg_tokens": null,
   "tps": 85.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-e09egwcs-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: 0rrghqxu, sa79fojk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2542.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,542 81.5 14.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e09egwcs",
   "source_url": "https://omlx.ai/benchmarks/performance/e09egwcs",
   "tg_tokens": null,
   "tps": 81.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-e09egwcs-ctx8192",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: 0rrghqxu, sa79fojk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2450.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 2,450 76.2 14.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e09egwcs",
   "source_url": "https://omlx.ai/benchmarks/performance/e09egwcs",
   "tg_tokens": null,
   "tps": 76.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-zp3c1mmu-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: cw7m20lz, ec1x781e, wt85y4j0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 259.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community-Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 259.1 28.5 15.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zp3c1mmu",
   "source_url": "https://omlx.ai/benchmarks/performance/zp3c1mmu",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-zp3c1mmu-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: cw7m20lz, ec1x781e, wt85y4j0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 251.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community-Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 251.4 27.4 17.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zp3c1mmu",
   "source_url": "https://omlx.ai/benchmarks/performance/zp3c1mmu",
   "tg_tokens": null,
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-zp3c1mmu-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 17.9 GB at this context; duplicate traces of the same session: cw7m20lz, ec1x781e, wt85y4j0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 230.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community-Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 230.0 25.6 17.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zp3c1mmu",
   "source_url": "https://omlx.ai/benchmarks/performance/zp3c1mmu",
   "tg_tokens": null,
   "tps": 25.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-zp3c1mmu-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: cw7m20lz, ec1x781e, wt85y4j0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 213.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community-Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 213.2 24.3 19.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zp3c1mmu",
   "source_url": "https://omlx.ai/benchmarks/performance/zp3c1mmu",
   "tg_tokens": null,
   "tps": 24.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ljpgnidw-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: 2rt4ubl8, 67qhwyg3, mi776zf7, y2vb7rmd; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 694.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 1k 694.3 72.3 29.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ljpgnidw",
   "source_url": "https://omlx.ai/benchmarks/performance/ljpgnidw",
   "tg_tokens": null,
   "tps": 72.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ljpgnidw-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.8 GB at this context; duplicate traces of the same session: 2rt4ubl8, 67qhwyg3, mi776zf7, y2vb7rmd; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 756.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 4k 756.0 72.7 30.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ljpgnidw",
   "source_url": "https://omlx.ai/benchmarks/performance/ljpgnidw",
   "tg_tokens": null,
   "tps": 72.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ljpgnidw-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.1 GB at this context; duplicate traces of the same session: 2rt4ubl8, 67qhwyg3, mi776zf7, y2vb7rmd; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 724.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 8k 724.1 69.6 31.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ljpgnidw",
   "source_url": "https://omlx.ai/benchmarks/performance/ljpgnidw",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ljpgnidw-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.2 GB at this context; duplicate traces of the same session: 2rt4ubl8, 67qhwyg3, mi776zf7, y2vb7rmd; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 638.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 16k 638.9 57.3 31.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ljpgnidw",
   "source_url": "https://omlx.ai/benchmarks/performance/ljpgnidw",
   "tg_tokens": null,
   "tps": 57.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ljpgnidw-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.5 GB at this context; duplicate traces of the same session: 2rt4ubl8, 67qhwyg3, mi776zf7, y2vb7rmd; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 517.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 32k 517.7 48.7 31.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ljpgnidw",
   "source_url": "https://omlx.ai/benchmarks/performance/ljpgnidw",
   "tg_tokens": null,
   "tps": 48.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ljpgnidw-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.2 GB at this context; duplicate traces of the same session: 2rt4ubl8, 67qhwyg3, mi776zf7, y2vb7rmd; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 375.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 64k 375.0 23.1 32.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ljpgnidw",
   "source_url": "https://omlx.ai/benchmarks/performance/ljpgnidw",
   "tg_tokens": null,
   "tps": 23.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-merodj3j-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 103.6 24.1 23.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/merodj3j",
   "source_url": "https://omlx.ai/benchmarks/performance/merodj3j",
   "tg_tokens": null,
   "tps": 24.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-merodj3j-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 104.2 21.0 24.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/merodj3j",
   "source_url": "https://omlx.ai/benchmarks/performance/merodj3j",
   "tg_tokens": null,
   "tps": 21.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-rtv27vlp-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: abh8p5es; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 242.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 242.1 25.4 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rtv27vlp",
   "source_url": "https://omlx.ai/benchmarks/performance/rtv27vlp",
   "tg_tokens": null,
   "tps": 25.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-rtv27vlp-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: abh8p5es; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 301.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 301.2 28.0 24.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rtv27vlp",
   "source_url": "https://omlx.ai/benchmarks/performance/rtv27vlp",
   "tg_tokens": null,
   "tps": 28.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-3f5bd1l4-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 39.0 GB; TTFT 11027 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 92.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 92.9 PP tok/s 11.0 TG tok/s 11027 TTFT (ms) 39.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3f5bd1l4",
   "source_url": "https://omlx.ai/benchmarks/performance/3f5bd1l4",
   "tg_tokens": null,
   "tps": 11.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-f5q66kij-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 0uu2ah34, 2tdvpq3t, 488yjasc; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 736.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 736.6 87.7 21.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f5q66kij",
   "source_url": "https://omlx.ai/benchmarks/performance/f5q66kij",
   "tg_tokens": null,
   "tps": 87.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-f5q66kij-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 0uu2ah34, 2tdvpq3t, 488yjasc; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 855.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 855.8 101.6 21.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f5q66kij",
   "source_url": "https://omlx.ai/benchmarks/performance/f5q66kij",
   "tg_tokens": null,
   "tps": 101.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-f5q66kij-ctx8192",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 0uu2ah34, 2tdvpq3t, 488yjasc; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 806.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 806.1 89.4 22.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f5q66kij",
   "source_url": "https://omlx.ai/benchmarks/performance/f5q66kij",
   "tg_tokens": null,
   "tps": 89.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-f5q66kij-ctx16384",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: 0uu2ah34, 2tdvpq3t, 488yjasc; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 714.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 714.2 68.1 22.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f5q66kij",
   "source_url": "https://omlx.ai/benchmarks/performance/f5q66kij",
   "tg_tokens": null,
   "tps": 68.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-f5q66kij-ctx32768",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: 0uu2ah34, 2tdvpq3t, 488yjasc; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 577.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 577.9 67.3 23.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f5q66kij",
   "source_url": "https://omlx.ai/benchmarks/performance/f5q66kij",
   "tg_tokens": null,
   "tps": 67.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-f5q66kij-ctx65536",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: 0uu2ah34, 2tdvpq3t, 488yjasc; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 419.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 64k 419.5 52.4 25.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f5q66kij",
   "source_url": "https://omlx.ai/benchmarks/performance/f5q66kij",
   "tg_tokens": null,
   "tps": 52.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-oboije07-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "QAT OptiQ build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: eu1opdj6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2289.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,289 60.0 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oboije07",
   "source_url": "https://omlx.ai/benchmarks/performance/oboije07",
   "tg_tokens": null,
   "tps": 60.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-oboije07-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "QAT OptiQ build; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: eu1opdj6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2388.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,388 58.4 20.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oboije07",
   "source_url": "https://omlx.ai/benchmarks/performance/oboije07",
   "tg_tokens": null,
   "tps": 58.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-oboije07-ctx8192",
   "model": "gemma4 26B.A4B",
   "notes": "QAT OptiQ build; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: eu1opdj6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2322.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 2,322 56.2 20.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oboije07",
   "source_url": "https://omlx.ai/benchmarks/performance/oboije07",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1snzrdfq-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "QAT OptiQ build; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: 0xsoyajx; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2288.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,288 88.8 21.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1snzrdfq",
   "source_url": "https://omlx.ai/benchmarks/performance/1snzrdfq",
   "tg_tokens": null,
   "tps": 88.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1snzrdfq-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "QAT OptiQ build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 0xsoyajx; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1416.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,416 29.1 22.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1snzrdfq",
   "source_url": "https://omlx.ai/benchmarks/performance/1snzrdfq",
   "tg_tokens": null,
   "tps": 29.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-9v3eoi7l-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "QAT OptiQ build; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: 2tb1wun0, dupruzqc; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2456.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,456 49.7 21.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9v3eoi7l",
   "source_url": "https://omlx.ai/benchmarks/performance/9v3eoi7l",
   "tg_tokens": null,
   "tps": 49.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-9v3eoi7l-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "QAT OptiQ build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 2tb1wun0, dupruzqc; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2406.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,406 59.4 22.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9v3eoi7l",
   "source_url": "https://omlx.ai/benchmarks/performance/9v3eoi7l",
   "tg_tokens": null,
   "tps": 59.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-9v3eoi7l-ctx8192",
   "model": "gemma4 26B.A4B",
   "notes": "QAT OptiQ build; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: 2tb1wun0, dupruzqc; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2164.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 2,164 39.0 22.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9v3eoi7l",
   "source_url": "https://omlx.ai/benchmarks/performance/9v3eoi7l",
   "tg_tokens": null,
   "tps": 39.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-n2etttau-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: bdj39zxx, iaht1wzf, o63j1nbi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 226.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community-Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 226.1 28.5 15.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n2etttau",
   "source_url": "https://omlx.ai/benchmarks/performance/n2etttau",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-n2etttau-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: bdj39zxx, iaht1wzf, o63j1nbi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 242.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community-Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 242.6 26.8 17.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n2etttau",
   "source_url": "https://omlx.ai/benchmarks/performance/n2etttau",
   "tg_tokens": null,
   "tps": 26.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-n2etttau-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 17.9 GB at this context; duplicate traces of the same session: bdj39zxx, iaht1wzf, o63j1nbi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 243.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community-Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 243.9 25.9 17.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n2etttau",
   "source_url": "https://omlx.ai/benchmarks/performance/n2etttau",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-n2etttau-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: bdj39zxx, iaht1wzf, o63j1nbi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 222.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community-Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 222.6 24.2 19.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n2etttau",
   "source_url": "https://omlx.ai/benchmarks/performance/n2etttau",
   "tg_tokens": null,
   "tps": 24.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-d9xiskrw-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: 1viwl6ps, aohitfzv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 261.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community-Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 261.0 28.4 15.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d9xiskrw",
   "source_url": "https://omlx.ai/benchmarks/performance/d9xiskrw",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-d9xiskrw-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 1viwl6ps, aohitfzv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 264.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community-Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 264.0 28.0 17.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d9xiskrw",
   "source_url": "https://omlx.ai/benchmarks/performance/d9xiskrw",
   "tg_tokens": null,
   "tps": 28.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-d9xiskrw-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 17.9 GB at this context; duplicate traces of the same session: 1viwl6ps, aohitfzv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 246.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community-Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 246.6 26.0 17.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d9xiskrw",
   "source_url": "https://omlx.ai/benchmarks/performance/d9xiskrw",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-d9xiskrw-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 1viwl6ps, aohitfzv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community-Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 224.7 24.7 19.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d9xiskrw",
   "source_url": "https://omlx.ai/benchmarks/performance/d9xiskrw",
   "tg_tokens": null,
   "tps": 24.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-lgpd67ai-ctx1024",
   "model": "DeepSeek-Coder-V2-Lite-Instruct",
   "notes": "MLX; oMLX community benchmark trace; peak memory 9.1 GB at this context; duplicate traces of the same session: uz23eg6s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 533.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-Coder-V2-Lite-Instruct-4bit-mlx on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 533.1 18.5 9.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lgpd67ai",
   "source_url": "https://omlx.ai/benchmarks/performance/lgpd67ai",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-lgpd67ai-ctx4096",
   "model": "DeepSeek-Coder-V2-Lite-Instruct",
   "notes": "MLX; oMLX community benchmark trace; peak memory 10.2 GB at this context; duplicate traces of the same session: uz23eg6s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 183.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-Coder-V2-Lite-Instruct-4bit-mlx on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 183.9 15.5 10.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lgpd67ai",
   "source_url": "https://omlx.ai/benchmarks/performance/lgpd67ai",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-eb7622t7-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: kp7cued0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 854.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 854.5 93.5 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eb7622t7",
   "source_url": "https://omlx.ai/benchmarks/performance/eb7622t7",
   "tg_tokens": null,
   "tps": 93.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-eb7622t7-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: kp7cued0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1025.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,025 89.7 21.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eb7622t7",
   "source_url": "https://omlx.ai/benchmarks/performance/eb7622t7",
   "tg_tokens": null,
   "tps": 89.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kytxoee0-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 16.4 GB; TTFT 1369 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5984.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.6-35B-A3B-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 8k tokens 5,984 PP tok/s 113.6 TG tok/s 1369 TTFT (ms) 16.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kytxoee0",
   "source_url": "https://omlx.ai/benchmarks/performance/kytxoee0",
   "tg_tokens": null,
   "tps": 113.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-pf5uc430-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 16.4 GB; TTFT 1359 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6028.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.6-35B-A3B-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 8k tokens 6,028 PP tok/s 187.8 TG tok/s 1359 TTFT (ms) 16.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pf5uc430",
   "source_url": "https://omlx.ai/benchmarks/performance/pf5uc430",
   "tg_tokens": null,
   "tps": 187.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-30nm6n38-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 16.4 GB; TTFT 1357 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6039.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.6-35B-A3B-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 8k tokens 6,039 PP tok/s 182.2 TG tok/s 1357 TTFT (ms) 16.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/30nm6n38",
   "source_url": "https://omlx.ai/benchmarks/performance/30nm6n38",
   "tg_tokens": null,
   "tps": 182.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-r2knvcbo-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 16.4 GB; TTFT 1578 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5191.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.6-35B-A3B-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 8k tokens 5,191 PP tok/s 147.3 TG tok/s 1578 TTFT (ms) 16.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r2knvcbo",
   "source_url": "https://omlx.ai/benchmarks/performance/r2knvcbo",
   "tg_tokens": null,
   "tps": 147.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-b84c35d8-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 16.4 GB; TTFT 1375 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5957.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.6-35B-A3B-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 8k tokens 5,957 PP tok/s 141.5 TG tok/s 1375 TTFT (ms) 16.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b84c35d8",
   "source_url": "https://omlx.ai/benchmarks/performance/b84c35d8",
   "tg_tokens": null,
   "tps": 141.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-y8ef3pk8-ctx8192",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "W4A16; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: ziysg5bz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 228.9,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-27B-Coder390-NVFP4-W4A16 on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 nvfp4; 8k 228.9 28.8 25.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y8ef3pk8",
   "source_url": "https://omlx.ai/benchmarks/performance/y8ef3pk8",
   "tg_tokens": null,
   "tps": 28.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-y8ef3pk8-ctx32768",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "W4A16; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: ziysg5bz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 150.7,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-27B-Coder390-NVFP4-W4A16 on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 nvfp4; 32k 150.7 20.6 27.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y8ef3pk8",
   "source_url": "https://omlx.ai/benchmarks/performance/y8ef3pk8",
   "tg_tokens": null,
   "tps": 20.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-jam4szf5-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 103.6 24.1 23.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jam4szf5",
   "source_url": "https://omlx.ai/benchmarks/performance/jam4szf5",
   "tg_tokens": null,
   "tps": 24.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-jam4szf5-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 104.2 21.0 24.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jam4szf5",
   "source_url": "https://omlx.ai/benchmarks/performance/jam4szf5",
   "tg_tokens": null,
   "tps": 21.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-pjd08tia-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 8j6fws6v, 9mjj1yem, dnsg4acu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 236.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 236.1 55.6 18.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pjd08tia",
   "source_url": "https://omlx.ai/benchmarks/performance/pjd08tia",
   "tg_tokens": null,
   "tps": 55.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-pjd08tia-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 8j6fws6v, 9mjj1yem, dnsg4acu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 252.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 252.9 49.7 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pjd08tia",
   "source_url": "https://omlx.ai/benchmarks/performance/pjd08tia",
   "tg_tokens": null,
   "tps": 49.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-pjd08tia-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: 8j6fws6v, 9mjj1yem, dnsg4acu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 243.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 243.5 82.1 20.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pjd08tia",
   "source_url": "https://omlx.ai/benchmarks/performance/pjd08tia",
   "tg_tokens": null,
   "tps": 82.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-pjd08tia-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: 8j6fws6v, 9mjj1yem, dnsg4acu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 226.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 226.9 105.8 21.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pjd08tia",
   "source_url": "https://omlx.ai/benchmarks/performance/pjd08tia",
   "tg_tokens": null,
   "tps": 105.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-woe38lnr-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 213.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 213.8 26.8 16.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/woe38lnr",
   "source_url": "https://omlx.ai/benchmarks/performance/woe38lnr",
   "tg_tokens": null,
   "tps": 26.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-woe38lnr-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 223.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 223.0 26.8 17.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/woe38lnr",
   "source_url": "https://omlx.ai/benchmarks/performance/woe38lnr",
   "tg_tokens": null,
   "tps": 26.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-woe38lnr-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 243.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 243.1 25.0 18.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/woe38lnr",
   "source_url": "https://omlx.ai/benchmarks/performance/woe38lnr",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-woe38lnr-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 230.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 230.2 23.6 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/woe38lnr",
   "source_url": "https://omlx.ai/benchmarks/performance/woe38lnr",
   "tg_tokens": null,
   "tps": 23.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-d4s2gz91-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: f9cqzjgc; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 142.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 5bit; 4k 142.2 10.8 23.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d4s2gz91",
   "source_url": "https://omlx.ai/benchmarks/performance/d4s2gz91",
   "tg_tokens": null,
   "tps": 10.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-d4s2gz91-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: f9cqzjgc; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.5,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 5bit; 16k 131.5 10.0 25.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d4s2gz91",
   "source_url": "https://omlx.ai/benchmarks/performance/d4s2gz91",
   "tg_tokens": null,
   "tps": 10.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0btwen8p-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Youssofal MTPLX Optimized Speed; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: mh63zx8m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 839.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Youssofal--Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 839.0 52.9 23.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0btwen8p",
   "source_url": "https://omlx.ai/benchmarks/performance/0btwen8p",
   "tg_tokens": null,
   "tps": 52.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0btwen8p-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "Youssofal MTPLX Optimized Speed; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: mh63zx8m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 957.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Youssofal--Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 957.2 46.5 24.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0btwen8p",
   "source_url": "https://omlx.ai/benchmarks/performance/0btwen8p",
   "tg_tokens": null,
   "tps": 46.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-js64yqa0-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 261.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community-Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 261.0 28.4 15.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/js64yqa0",
   "source_url": "https://omlx.ai/benchmarks/performance/js64yqa0",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-js64yqa0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 264.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community-Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 264.0 28.0 17.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/js64yqa0",
   "source_url": "https://omlx.ai/benchmarks/performance/js64yqa0",
   "tg_tokens": null,
   "tps": 28.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-js64yqa0-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 17.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 246.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community-Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 246.6 26.0 17.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/js64yqa0",
   "source_url": "https://omlx.ai/benchmarks/performance/js64yqa0",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-js64yqa0-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community-Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 224.7 24.7 19.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/js64yqa0",
   "source_url": "https://omlx.ai/benchmarks/performance/js64yqa0",
   "tg_tokens": null,
   "tps": 24.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-5nbiv18p-ctx4096",
   "model": "AREX-Turbo",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: m12mymj6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 590.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "AREX-Turbo-8bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 8bit; 4k 590.8 42.6 5.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5nbiv18p",
   "source_url": "https://omlx.ai/benchmarks/performance/5nbiv18p",
   "tg_tokens": null,
   "tps": 42.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-5nbiv18p-ctx16384",
   "model": "AREX-Turbo",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: m12mymj6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 559.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "AREX-Turbo-8bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 8bit; 16k 559.5 35.1 6.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5nbiv18p",
   "source_url": "https://omlx.ai/benchmarks/performance/5nbiv18p",
   "tg_tokens": null,
   "tps": 35.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2e9ycgi4-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 16.4 GB; TTFT 1468 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5582.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.6-35B-A3B-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 8k tokens 5,582 PP tok/s 175.6 TG tok/s 1468 TTFT (ms) 16.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2e9ycgi4",
   "source_url": "https://omlx.ai/benchmarks/performance/2e9ycgi4",
   "tg_tokens": null,
   "tps": 175.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jn96p1z8-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 16.4 GB; TTFT 1904 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4303.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.6-35B-A3B-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 8k tokens 4,303 PP tok/s 186.2 TG tok/s 1904 TTFT (ms) 16.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jn96p1z8",
   "source_url": "https://omlx.ai/benchmarks/performance/jn96p1z8",
   "tg_tokens": null,
   "tps": 186.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-0t13fv0b-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: ixdxjad6, o1e0ryex; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 213.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 213.8 26.8 16.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0t13fv0b",
   "source_url": "https://omlx.ai/benchmarks/performance/0t13fv0b",
   "tg_tokens": null,
   "tps": 26.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-0t13fv0b-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: ixdxjad6, o1e0ryex; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 223.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 223.0 26.8 17.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0t13fv0b",
   "source_url": "https://omlx.ai/benchmarks/performance/0t13fv0b",
   "tg_tokens": null,
   "tps": 26.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-0t13fv0b-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: ixdxjad6, o1e0ryex; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 243.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 243.1 25.0 18.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0t13fv0b",
   "source_url": "https://omlx.ai/benchmarks/performance/0t13fv0b",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-0t13fv0b-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: ixdxjad6, o1e0ryex; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 230.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 230.2 23.6 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0t13fv0b",
   "source_url": "https://omlx.ai/benchmarks/performance/0t13fv0b",
   "tg_tokens": null,
   "tps": 23.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-dcr71y42-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: kqcusyn0, v6fat1pj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 260.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 260.1 27.6 16.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dcr71y42",
   "source_url": "https://omlx.ai/benchmarks/performance/dcr71y42",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-dcr71y42-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: kqcusyn0, v6fat1pj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 261.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 261.1 26.7 17.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dcr71y42",
   "source_url": "https://omlx.ai/benchmarks/performance/dcr71y42",
   "tg_tokens": null,
   "tps": 26.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-dcr71y42-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: kqcusyn0, v6fat1pj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 250.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 250.1 24.6 18.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dcr71y42",
   "source_url": "https://omlx.ai/benchmarks/performance/dcr71y42",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-dcr71y42-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: kqcusyn0, v6fat1pj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 232.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 232.9 23.5 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dcr71y42",
   "source_url": "https://omlx.ai/benchmarks/performance/dcr71y42",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ybwnkt0t-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 254.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 254.3 27.5 16.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ybwnkt0t",
   "source_url": "https://omlx.ai/benchmarks/performance/ybwnkt0t",
   "tg_tokens": null,
   "tps": 27.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ybwnkt0t-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 273.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 273.0 26.9 19.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ybwnkt0t",
   "source_url": "https://omlx.ai/benchmarks/performance/ybwnkt0t",
   "tg_tokens": null,
   "tps": 26.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8zalxig1-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 43.0 GB; TTFT 943 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1086.0,
   "provenance": "community",
   "quant": "GSQ-RCO-Q2",
   "quote": "Qwen3.8-Flash-Next-GSQ-RCO-Q2-mlx on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,086 PP tok/s 65.2 TG tok/s 943 TTFT (ms) 43.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8zalxig1",
   "source_url": "https://omlx.ai/benchmarks/performance/8zalxig1",
   "tg_tokens": null,
   "tps": 65.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-eqm76245-ctx1024",
   "model": "gemma-4-12b-coder-fable5-composer2.5",
   "notes": "community coder fine-tune; oMLX community benchmark trace; peak memory 7.4 GB at this context; duplicate traces of the same session: xdneymkw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 464.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12b-coder-fable5-composer2.5-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 464.7 45.5 7.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eqm76245",
   "source_url": "https://omlx.ai/benchmarks/performance/eqm76245",
   "tg_tokens": null,
   "tps": 45.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-eqm76245-ctx4096",
   "model": "gemma-4-12b-coder-fable5-composer2.5",
   "notes": "community coder fine-tune; oMLX community benchmark trace; peak memory 7.7 GB at this context; duplicate traces of the same session: xdneymkw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 462.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12b-coder-fable5-composer2.5-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 462.8 38.9 7.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eqm76245",
   "source_url": "https://omlx.ai/benchmarks/performance/eqm76245",
   "tg_tokens": null,
   "tps": 38.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-p8au8t0v-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: tyj2s700; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 675.9,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 675.9 73.8 11.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p8au8t0v",
   "source_url": "https://omlx.ai/benchmarks/performance/p8au8t0v",
   "tg_tokens": null,
   "tps": 73.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-p8au8t0v-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: tyj2s700; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 782.3,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 782.3 65.9 11.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p8au8t0v",
   "source_url": "https://omlx.ai/benchmarks/performance/p8au8t0v",
   "tg_tokens": null,
   "tps": 65.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-2fq0ddqi-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: 2wtigqsg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 853.1,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 853.1 83.4 11.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2fq0ddqi",
   "source_url": "https://omlx.ai/benchmarks/performance/2fq0ddqi",
   "tg_tokens": null,
   "tps": 83.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-2fq0ddqi-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: 2wtigqsg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2245.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 2,245 75.3 11.8 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2fq0ddqi",
   "source_url": "https://omlx.ai/benchmarks/performance/2fq0ddqi",
   "tg_tokens": null,
   "tps": 75.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-8mqspoon-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: mdyhq80m, q893pzki, v3j1k83m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 260.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community-Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 260.1 28.2 15.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8mqspoon",
   "source_url": "https://omlx.ai/benchmarks/performance/8mqspoon",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-8mqspoon-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: mdyhq80m, q893pzki, v3j1k83m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 257.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community-Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 257.5 28.1 17.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8mqspoon",
   "source_url": "https://omlx.ai/benchmarks/performance/8mqspoon",
   "tg_tokens": null,
   "tps": 28.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-8mqspoon-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 17.9 GB at this context; duplicate traces of the same session: mdyhq80m, q893pzki, v3j1k83m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 240.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community-Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 240.4 25.9 17.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8mqspoon",
   "source_url": "https://omlx.ai/benchmarks/performance/8mqspoon",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-8mqspoon-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: mdyhq80m, q893pzki, v3j1k83m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 229.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community-Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 229.1 24.4 19.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8mqspoon",
   "source_url": "https://omlx.ai/benchmarks/performance/8mqspoon",
   "tg_tokens": null,
   "tps": 24.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-fcms0ktb-ctx4096",
   "model": "AREX-Turbo",
   "notes": "oMLX community benchmark trace; peak memory 4.8 GB at this context; duplicate traces of the same session: r2lfvqx1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1255.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "AREX-Turbo-6bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 6bit; 4k 1,255 77.3 4.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fcms0ktb",
   "source_url": "https://omlx.ai/benchmarks/performance/fcms0ktb",
   "tg_tokens": null,
   "tps": 77.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-fcms0ktb-ctx16384",
   "model": "AREX-Turbo",
   "notes": "oMLX community benchmark trace; peak memory 5.3 GB at this context; duplicate traces of the same session: r2lfvqx1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1156.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "AREX-Turbo-6bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 6bit; 16k 1,156 65.0 5.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fcms0ktb",
   "source_url": "https://omlx.ai/benchmarks/performance/fcms0ktb",
   "tg_tokens": null,
   "tps": 65.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2kvv90s3-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 19.6 GB; TTFT 1476 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5550.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.6-35B-A3B-oQ3.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 8k tokens 5,550 PP tok/s 176.7 TG tok/s 1476 TTFT (ms) 19.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2kvv90s3",
   "source_url": "https://omlx.ai/benchmarks/performance/2kvv90s3",
   "tg_tokens": null,
   "tps": 176.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-8pkrh3j0-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 185.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 185.2 38.3 19.6 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8pkrh3j0",
   "source_url": "https://omlx.ai/benchmarks/performance/8pkrh3j0",
   "tg_tokens": null,
   "tps": 38.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-8pkrh3j0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 202.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 202.9 34.8 21.5 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8pkrh3j0",
   "source_url": "https://omlx.ai/benchmarks/performance/8pkrh3j0",
   "tg_tokens": null,
   "tps": 34.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-zgpaqr0t-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 8k context; peak memory 18.8 GB; TTFT 63514 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 129.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 24 GB \u00b7 4bit; 8k tokens 129.0 PP tok/s 16.2 TG tok/s 63514 TTFT (ms) 18.8 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zgpaqr0t",
   "source_url": "https://omlx.ai/benchmarks/performance/zgpaqr0t",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-2adad97u-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 18.4 GB; TTFT 31103 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 24 GB \u00b7 4bit; 4k tokens 131.7 PP tok/s 22.6 TG tok/s 31103 TTFT (ms) 18.4 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2adad97u",
   "source_url": "https://omlx.ai/benchmarks/performance/2adad97u",
   "tg_tokens": null,
   "tps": 22.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ctorgzjd-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.7 GB at this context; duplicate traces of the same session: ct9v585h, q9voj3ds; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 842.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 842.8 32.2 33.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ctorgzjd",
   "source_url": "https://omlx.ai/benchmarks/performance/ctorgzjd",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ctorgzjd-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: ct9v585h, q9voj3ds; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 928.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 928.0 31.5 34.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ctorgzjd",
   "source_url": "https://omlx.ai/benchmarks/performance/ctorgzjd",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ctorgzjd-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 35.1 GB at this context; duplicate traces of the same session: ct9v585h, q9voj3ds; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 851.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 851.3 25.1 35.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ctorgzjd",
   "source_url": "https://omlx.ai/benchmarks/performance/ctorgzjd",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-bhty2d71-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 260.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 260.1 27.6 16.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhty2d71",
   "source_url": "https://omlx.ai/benchmarks/performance/bhty2d71",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-bhty2d71-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 261.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 261.1 26.7 17.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhty2d71",
   "source_url": "https://omlx.ai/benchmarks/performance/bhty2d71",
   "tg_tokens": null,
   "tps": 26.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-bhty2d71-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 250.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 250.1 24.6 18.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhty2d71",
   "source_url": "https://omlx.ai/benchmarks/performance/bhty2d71",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-bhty2d71-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 232.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 232.9 23.5 19.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhty2d71",
   "source_url": "https://omlx.ai/benchmarks/performance/bhty2d71",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-nuow50h9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 254.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 254.3 27.5 16.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nuow50h9",
   "source_url": "https://omlx.ai/benchmarks/performance/nuow50h9",
   "tg_tokens": null,
   "tps": 27.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-nuow50h9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 273.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 273.0 26.9 19.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nuow50h9",
   "source_url": "https://omlx.ai/benchmarks/performance/nuow50h9",
   "tg_tokens": null,
   "tps": 26.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-yo4wjibe-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 40.5 GB; TTFT 1320 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 776.0,
   "provenance": "community",
   "quant": "GSQ-RCO-Q2",
   "quote": "Qwen3.8-Flash-Next-GSQ-RCO-Q2-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k tokens 776.0 PP tok/s 45.0 TG tok/s 1320 TTFT (ms) 40.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yo4wjibe",
   "source_url": "https://omlx.ai/benchmarks/performance/yo4wjibe",
   "tg_tokens": null,
   "tps": 45.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-z0cxf6lo-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 69.5 GB at this context; duplicate traces of the same session: i7jkfemq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1948.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,948 76.7 69.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z0cxf6lo",
   "source_url": "https://omlx.ai/benchmarks/performance/z0cxf6lo",
   "tg_tokens": null,
   "tps": 76.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-z0cxf6lo-ctx8192",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: i7jkfemq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2464.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,464 73.6 71.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z0cxf6lo",
   "source_url": "https://omlx.ai/benchmarks/performance/z0cxf6lo",
   "tg_tokens": null,
   "tps": 73.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-woghmqji-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 85.1 GB; TTFT 199135 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 658.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k tokens 658.2 PP tok/s 41.7 TG tok/s 199135 TTFT (ms) 85.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/woghmqji",
   "source_url": "https://omlx.ai/benchmarks/performance/woghmqji",
   "tg_tokens": null,
   "tps": 41.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-pwl8e4mc-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 84.8 GB; TTFT 196120 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 668.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k tokens 668.3 PP tok/s 42.9 TG tok/s 196120 TTFT (ms) 84.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pwl8e4mc",
   "source_url": "https://omlx.ai/benchmarks/performance/pwl8e4mc",
   "tg_tokens": null,
   "tps": 42.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-bglpoujf-ctx1024",
   "model": "gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau2-uncensored-heretic",
   "notes": "oMLX community benchmark trace; peak memory 7.4 GB at this context; duplicate traces of the same session: ztthe3mu; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 141.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau2-uncensored-heretic-oQ4e on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 141.3 13.0 7.4 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bglpoujf",
   "source_url": "https://omlx.ai/benchmarks/performance/bglpoujf",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-bglpoujf-ctx4096",
   "model": "gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau2-uncensored-heretic",
   "notes": "oMLX community benchmark trace; peak memory 7.8 GB at this context; duplicate traces of the same session: ztthe3mu; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 138.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau2-uncensored-heretic-oQ4e on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 138.1 13.1 7.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bglpoujf",
   "source_url": "https://omlx.ai/benchmarks/performance/bglpoujf",
   "tg_tokens": null,
   "tps": 13.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-8qu1qkvf-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: 25oca0v6, hxvvhn2s, iztwa74d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 245.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community-Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 245.9 28.2 15.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8qu1qkvf",
   "source_url": "https://omlx.ai/benchmarks/performance/8qu1qkvf",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-8qu1qkvf-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 25oca0v6, hxvvhn2s, iztwa74d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 260.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community-Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 260.2 28.0 17.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8qu1qkvf",
   "source_url": "https://omlx.ai/benchmarks/performance/8qu1qkvf",
   "tg_tokens": null,
   "tps": 28.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-8qu1qkvf-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 17.9 GB at this context; duplicate traces of the same session: 25oca0v6, hxvvhn2s, iztwa74d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 248.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community-Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 248.2 25.9 17.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8qu1qkvf",
   "source_url": "https://omlx.ai/benchmarks/performance/8qu1qkvf",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-8qu1qkvf-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 25oca0v6, hxvvhn2s, iztwa74d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 233.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community-Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 233.8 24.3 19.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8qu1qkvf",
   "source_url": "https://omlx.ai/benchmarks/performance/8qu1qkvf",
   "tg_tokens": null,
   "tps": 24.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-nzt8y2mt-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: 4mt2ir66; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 241.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 241.3 20.9 6.3 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nzt8y2mt",
   "source_url": "https://omlx.ai/benchmarks/performance/nzt8y2mt",
   "tg_tokens": null,
   "tps": 20.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-nzt8y2mt-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: 4mt2ir66; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 226.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 16k 226.9 18.3 6.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nzt8y2mt",
   "source_url": "https://omlx.ai/benchmarks/performance/nzt8y2mt",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-6en83ho9-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 6ea7wwr0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 915.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 915.3 93.8 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6en83ho9",
   "source_url": "https://omlx.ai/benchmarks/performance/6en83ho9",
   "tg_tokens": null,
   "tps": 93.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-6en83ho9-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 6ea7wwr0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1080.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,080 89.9 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6en83ho9",
   "source_url": "https://omlx.ai/benchmarks/performance/6en83ho9",
   "tg_tokens": null,
   "tps": 89.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ijddqxau-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 19.6 GB; TTFT 1545 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5301.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.6-35B-A3B-oQ3.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 8k tokens 5,301 PP tok/s 172.5 TG tok/s 1545 TTFT (ms) 19.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ijddqxau",
   "source_url": "https://omlx.ai/benchmarks/performance/ijddqxau",
   "tg_tokens": null,
   "tps": 172.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-8mk3zr7r-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: t6vwvpjo; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 150.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 4k 150.6 24.5 19.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8mk3zr7r",
   "source_url": "https://omlx.ai/benchmarks/performance/8mk3zr7r",
   "tg_tokens": null,
   "tps": 24.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-8mk3zr7r-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: t6vwvpjo; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 138.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 16k 138.4 22.2 21.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8mk3zr7r",
   "source_url": "https://omlx.ai/benchmarks/performance/8mk3zr7r",
   "tg_tokens": null,
   "tps": 22.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-m27ydhm2-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 8kzh7xjm; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 207.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 207.1 22.3 16.0 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m27ydhm2",
   "source_url": "https://omlx.ai/benchmarks/performance/m27ydhm2",
   "tg_tokens": null,
   "tps": 22.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-m27ydhm2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 8kzh7xjm; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 200.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 200.6 21.6 17.3 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m27ydhm2",
   "source_url": "https://omlx.ai/benchmarks/performance/m27ydhm2",
   "tg_tokens": null,
   "tps": 21.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-sq5gq0vg-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 04tf584n; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 190.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 190.7 35.2 19.6 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sq5gq0vg",
   "source_url": "https://omlx.ai/benchmarks/performance/sq5gq0vg",
   "tg_tokens": null,
   "tps": 35.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-sq5gq0vg-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 04tf584n; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 201.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 201.7 37.5 21.5 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sq5gq0vg",
   "source_url": "https://omlx.ai/benchmarks/performance/sq5gq0vg",
   "tg_tokens": null,
   "tps": 37.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-x0rdhl1b-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 185.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 185.2 38.3 19.6 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x0rdhl1b",
   "source_url": "https://omlx.ai/benchmarks/performance/x0rdhl1b",
   "tg_tokens": null,
   "tps": 38.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-x0rdhl1b-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 202.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 202.9 34.8 21.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x0rdhl1b",
   "source_url": "https://omlx.ai/benchmarks/performance/x0rdhl1b",
   "tg_tokens": null,
   "tps": 34.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-61yr6yn2-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: g3nw7259; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 103.6 26.0 24.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/61yr6yn2",
   "source_url": "https://omlx.ai/benchmarks/performance/61yr6yn2",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-61yr6yn2-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 25.2 GB at this context; duplicate traces of the same session: g3nw7259; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 107.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 107.6 19.9 25.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/61yr6yn2",
   "source_url": "https://omlx.ai/benchmarks/performance/61yr6yn2",
   "tg_tokens": null,
   "tps": 19.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2z9tqk3d-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 5w2za8ew; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 243.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 243.6 36.2 18.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2z9tqk3d",
   "source_url": "https://omlx.ai/benchmarks/performance/2z9tqk3d",
   "tg_tokens": null,
   "tps": 36.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2z9tqk3d-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 5w2za8ew; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 254.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 254.8 35.6 19.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2z9tqk3d",
   "source_url": "https://omlx.ai/benchmarks/performance/2z9tqk3d",
   "tg_tokens": null,
   "tps": 35.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2z9tqk3d-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: 5w2za8ew; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 246.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 246.7 38.8 20.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2z9tqk3d",
   "source_url": "https://omlx.ai/benchmarks/performance/2z9tqk3d",
   "tg_tokens": null,
   "tps": 38.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2z9tqk3d-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: 5w2za8ew; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 229.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 229.3 29.3 21.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2z9tqk3d",
   "source_url": "https://omlx.ai/benchmarks/performance/2z9tqk3d",
   "tg_tokens": null,
   "tps": 29.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-0ss6fowf-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: pxwt00my, tpr23oeg, xt0zrnni; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 240.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 240.4 27.9 16.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ss6fowf",
   "source_url": "https://omlx.ai/benchmarks/performance/0ss6fowf",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-0ss6fowf-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: pxwt00my, tpr23oeg, xt0zrnni; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 252.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 252.2 26.0 17.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ss6fowf",
   "source_url": "https://omlx.ai/benchmarks/performance/0ss6fowf",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-0ss6fowf-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: pxwt00my, tpr23oeg, xt0zrnni; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 249.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 249.3 25.6 18.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ss6fowf",
   "source_url": "https://omlx.ai/benchmarks/performance/0ss6fowf",
   "tg_tokens": null,
   "tps": 25.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-0ss6fowf-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: pxwt00my, tpr23oeg, xt0zrnni; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 232.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 232.6 24.0 19.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ss6fowf",
   "source_url": "https://omlx.ai/benchmarks/performance/0ss6fowf",
   "tg_tokens": null,
   "tps": 24.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-9bwowbfj-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 1k context; peak memory 22.6 GB; TTFT 425 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2407.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k tokens 2,407 PP tok/s 101.0 TG tok/s 425 TTFT (ms) 22.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9bwowbfj",
   "source_url": "https://omlx.ai/benchmarks/performance/9bwowbfj",
   "tg_tokens": null,
   "tps": 101.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-pj0ediy7-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 1k context; peak memory 20.1 GB; TTFT 388 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2643.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k tokens 2,643 PP tok/s 87.7 TG tok/s 388 TTFT (ms) 20.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pj0ediy7",
   "source_url": "https://omlx.ai/benchmarks/performance/pj0ediy7",
   "tg_tokens": null,
   "tps": 87.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-n13u558n-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 25.8 GB; TTFT 50266 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1304.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k tokens 1,304 PP tok/s 46.4 TG tok/s 50266 TTFT (ms) 25.8 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n13u558n",
   "source_url": "https://omlx.ai/benchmarks/performance/n13u558n",
   "tg_tokens": null,
   "tps": 46.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-tnb3fv45-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 23.9 GB; TTFT 7490 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2188.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 16k tokens 2,188 PP tok/s 121.2 TG tok/s 7490 TTFT (ms) 23.9 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tnb3fv45",
   "source_url": "https://omlx.ai/benchmarks/performance/tnb3fv45",
   "tg_tokens": null,
   "tps": 121.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-6oh5qtvg-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 25.8 GB; TTFT 50155 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1307.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k tokens 1,307 PP tok/s 67.4 TG tok/s 50155 TTFT (ms) 25.8 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6oh5qtvg",
   "source_url": "https://omlx.ai/benchmarks/performance/6oh5qtvg",
   "tg_tokens": null,
   "tps": 67.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-okf9gimg-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 24.6 GB; TTFT 18314 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1789.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k tokens 1,789 PP tok/s 105.2 TG tok/s 18314 TTFT (ms) 24.6 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/okf9gimg",
   "source_url": "https://omlx.ai/benchmarks/performance/okf9gimg",
   "tg_tokens": null,
   "tps": 105.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nbcq9lcg-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 23.4 GB; TTFT 1594 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2571.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k tokens 2,571 PP tok/s 128.2 TG tok/s 1594 TTFT (ms) 23.4 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nbcq9lcg",
   "source_url": "https://omlx.ai/benchmarks/performance/nbcq9lcg",
   "tg_tokens": null,
   "tps": 128.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nolq8hla-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 22.1 GB; TTFT 1384 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5919.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k tokens 5,919 PP tok/s 193.6 TG tok/s 1384 TTFT (ms) 22.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nolq8hla",
   "source_url": "https://omlx.ai/benchmarks/performance/nolq8hla",
   "tg_tokens": null,
   "tps": 193.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-27sqoef4-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 30.7 GB; TTFT 1581 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5182.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k tokens 5,182 PP tok/s 161.1 TG tok/s 1581 TTFT (ms) 30.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/27sqoef4",
   "source_url": "https://omlx.ai/benchmarks/performance/27sqoef4",
   "tg_tokens": null,
   "tps": 161.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1kj5f04p-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 30.7 GB; TTFT 1574 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5206.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k tokens 5,206 PP tok/s 177.4 TG tok/s 1574 TTFT (ms) 30.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1kj5f04p",
   "source_url": "https://omlx.ai/benchmarks/performance/1kj5f04p",
   "tg_tokens": null,
   "tps": 177.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9er0k82o-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 1vokcmrm, rr5liv8t, v1o4ct99; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 915.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 915.8 47.6 17.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9er0k82o",
   "source_url": "https://omlx.ai/benchmarks/performance/9er0k82o",
   "tg_tokens": null,
   "tps": 47.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9er0k82o-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: 1vokcmrm, rr5liv8t, v1o4ct99; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1192.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,192 64.0 18.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9er0k82o",
   "source_url": "https://omlx.ai/benchmarks/performance/9er0k82o",
   "tg_tokens": null,
   "tps": 64.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9er0k82o-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 1vokcmrm, rr5liv8t, v1o4ct99; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1284.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 8k 1,284 77.3 18.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9er0k82o",
   "source_url": "https://omlx.ai/benchmarks/performance/9er0k82o",
   "tg_tokens": null,
   "tps": 77.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9er0k82o-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 1vokcmrm, rr5liv8t, v1o4ct99; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1235.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 1,235 78.4 19.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9er0k82o",
   "source_url": "https://omlx.ai/benchmarks/performance/9er0k82o",
   "tg_tokens": null,
   "tps": 78.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9er0k82o-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 1vokcmrm, rr5liv8t, v1o4ct99; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1105.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 1,105 80.8 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9er0k82o",
   "source_url": "https://omlx.ai/benchmarks/performance/9er0k82o",
   "tg_tokens": null,
   "tps": 80.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9er0k82o-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: 1vokcmrm, rr5liv8t, v1o4ct99; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 953.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 64k 953.5 54.6 22.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9er0k82o",
   "source_url": "https://omlx.ai/benchmarks/performance/9er0k82o",
   "tg_tokens": null,
   "tps": 54.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9er0k82o-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 27.2 GB at this context; duplicate traces of the same session: 1vokcmrm, rr5liv8t, v1o4ct99; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 615.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 128k 615.7 48.8 27.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9er0k82o",
   "source_url": "https://omlx.ai/benchmarks/performance/9er0k82o",
   "tg_tokens": null,
   "tps": 48.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-blpmx3dp-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: meymuqhn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 514.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 514.3 17.8 17.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/blpmx3dp",
   "source_url": "https://omlx.ai/benchmarks/performance/blpmx3dp",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-blpmx3dp-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: meymuqhn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 453.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 453.9 16.7 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/blpmx3dp",
   "source_url": "https://omlx.ai/benchmarks/performance/blpmx3dp",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-096yo9vg-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 19.1 GB; TTFT 75865 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 431.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k tokens 431.9 PP tok/s 15.7 TG tok/s 75865 TTFT (ms) 19.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/096yo9vg",
   "source_url": "https://omlx.ai/benchmarks/performance/096yo9vg",
   "tg_tokens": null,
   "tps": 15.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-dd5r2ohd-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: slcqaiy2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 243.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 243.6 36.2 18.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dd5r2ohd",
   "source_url": "https://omlx.ai/benchmarks/performance/dd5r2ohd",
   "tg_tokens": null,
   "tps": 36.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-dd5r2ohd-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: slcqaiy2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 254.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 254.8 35.6 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dd5r2ohd",
   "source_url": "https://omlx.ai/benchmarks/performance/dd5r2ohd",
   "tg_tokens": null,
   "tps": 35.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-dd5r2ohd-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: slcqaiy2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 246.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 246.7 38.8 20.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dd5r2ohd",
   "source_url": "https://omlx.ai/benchmarks/performance/dd5r2ohd",
   "tg_tokens": null,
   "tps": 38.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-dd5r2ohd-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: slcqaiy2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 229.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 229.3 29.3 21.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dd5r2ohd",
   "source_url": "https://omlx.ai/benchmarks/performance/dd5r2ohd",
   "tg_tokens": null,
   "tps": 29.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-jdhjwtoa-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 84.8 GB; TTFT 195507 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 670.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k tokens 670.4 PP tok/s 41.3 TG tok/s 195507 TTFT (ms) 84.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jdhjwtoa",
   "source_url": "https://omlx.ai/benchmarks/performance/jdhjwtoa",
   "tg_tokens": null,
   "tps": 41.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-x4bjya0i-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 4uwl3pug, j1wlsgzo, o8s5wz2s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 254.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 254.3 70.2 18.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x4bjya0i",
   "source_url": "https://omlx.ai/benchmarks/performance/x4bjya0i",
   "tg_tokens": null,
   "tps": 70.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-x4bjya0i-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 4uwl3pug, j1wlsgzo, o8s5wz2s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 268.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 268.9 61.3 22.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x4bjya0i",
   "source_url": "https://omlx.ai/benchmarks/performance/x4bjya0i",
   "tg_tokens": null,
   "tps": 61.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-x4bjya0i-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: 4uwl3pug, j1wlsgzo, o8s5wz2s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 264.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 264.1 42.3 22.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x4bjya0i",
   "source_url": "https://omlx.ai/benchmarks/performance/x4bjya0i",
   "tg_tokens": null,
   "tps": 42.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-x4bjya0i-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 4uwl3pug, j1wlsgzo, o8s5wz2s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 253.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 253.9 59.6 23.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x4bjya0i",
   "source_url": "https://omlx.ai/benchmarks/performance/x4bjya0i",
   "tg_tokens": null,
   "tps": 59.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-xsb0wemp-ctx4096",
   "model": "Devstral-Small-2505",
   "notes": "oMLX community benchmark trace; peak memory 13.9 GB at this context; duplicate traces of the same session: wlud0qov; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 298.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Devstral-Small-2505-4bit on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 4k 298.0 8.9 13.9 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xsb0wemp",
   "source_url": "https://omlx.ai/benchmarks/performance/xsb0wemp",
   "tg_tokens": null,
   "tps": 8.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-xsb0wemp-ctx16384",
   "model": "Devstral-Small-2505",
   "notes": "oMLX community benchmark trace; peak memory 15.6 GB at this context; duplicate traces of the same session: wlud0qov; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 277.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Devstral-Small-2505-4bit on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 16k 277.7 7.8 15.6 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xsb0wemp",
   "source_url": "https://omlx.ai/benchmarks/performance/xsb0wemp",
   "tg_tokens": null,
   "tps": 7.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-we6i1jl5-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: 01872wv7, 45tz38a7, aln2cf3m, c8vb1iep, m61jff3d; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1306.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 6bit; 1k 1,306 86.0 29.9 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/we6i1jl5",
   "source_url": "https://omlx.ai/benchmarks/performance/we6i1jl5",
   "tg_tokens": null,
   "tps": 86.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-we6i1jl5-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.9 GB at this context; duplicate traces of the same session: 01872wv7, 45tz38a7, aln2cf3m, c8vb1iep, m61jff3d; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2658.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 6bit; 4k 2,658 110.6 31.9 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/we6i1jl5",
   "source_url": "https://omlx.ai/benchmarks/performance/we6i1jl5",
   "tg_tokens": null,
   "tps": 110.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-we6i1jl5-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.2 GB at this context; duplicate traces of the same session: 01872wv7, 45tz38a7, aln2cf3m, c8vb1iep, m61jff3d; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2509.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 6bit; 8k 2,509 45.1 32.2 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/we6i1jl5",
   "source_url": "https://omlx.ai/benchmarks/performance/we6i1jl5",
   "tg_tokens": null,
   "tps": 45.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-we6i1jl5-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.7 GB at this context; duplicate traces of the same session: 01872wv7, 45tz38a7, aln2cf3m, c8vb1iep, m61jff3d; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2228.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 6bit; 16k 2,228 143.4 32.7 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/we6i1jl5",
   "source_url": "https://omlx.ai/benchmarks/performance/we6i1jl5",
   "tg_tokens": null,
   "tps": 143.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "outlier"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-we6i1jl5-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.6 GB at this context; duplicate traces of the same session: 01872wv7, 45tz38a7, aln2cf3m, c8vb1iep, m61jff3d; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1805.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 6bit; 32k 1,805 156.4 33.6 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/we6i1jl5",
   "source_url": "https://omlx.ai/benchmarks/performance/we6i1jl5",
   "tg_tokens": null,
   "tps": 156.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "outlier"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-we6i1jl5-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: 01872wv7, 45tz38a7, aln2cf3m, c8vb1iep, m61jff3d; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1294.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 6bit; 64k 1,294 99.6 35.2 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/we6i1jl5",
   "source_url": "https://omlx.ai/benchmarks/performance/we6i1jl5",
   "tg_tokens": null,
   "tps": 99.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2c6xarb9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 94lei1c0, v36clpun; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 915.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 915.8 47.6 17.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2c6xarb9",
   "source_url": "https://omlx.ai/benchmarks/performance/2c6xarb9",
   "tg_tokens": null,
   "tps": 47.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2c6xarb9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: 94lei1c0, v36clpun; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1192.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,192 64.0 18.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2c6xarb9",
   "source_url": "https://omlx.ai/benchmarks/performance/2c6xarb9",
   "tg_tokens": null,
   "tps": 64.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2c6xarb9-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 94lei1c0, v36clpun; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1284.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 8k 1,284 77.3 18.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2c6xarb9",
   "source_url": "https://omlx.ai/benchmarks/performance/2c6xarb9",
   "tg_tokens": null,
   "tps": 77.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2c6xarb9-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 94lei1c0, v36clpun; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1235.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 1,235 78.4 19.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2c6xarb9",
   "source_url": "https://omlx.ai/benchmarks/performance/2c6xarb9",
   "tg_tokens": null,
   "tps": 78.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2c6xarb9-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 94lei1c0, v36clpun; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1105.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 1,105 80.8 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2c6xarb9",
   "source_url": "https://omlx.ai/benchmarks/performance/2c6xarb9",
   "tg_tokens": null,
   "tps": 80.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2c6xarb9-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: 94lei1c0, v36clpun; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 953.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 64k 953.5 54.6 22.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2c6xarb9",
   "source_url": "https://omlx.ai/benchmarks/performance/2c6xarb9",
   "tg_tokens": null,
   "tps": 54.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2c6xarb9-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 27.2 GB at this context; duplicate traces of the same session: 94lei1c0, v36clpun; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 615.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 128k 615.7 48.8 27.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2c6xarb9",
   "source_url": "https://omlx.ai/benchmarks/performance/2c6xarb9",
   "tg_tokens": null,
   "tps": 48.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-halqel6o-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: it7kgmxn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 510.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 510.3 17.9 17.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/halqel6o",
   "source_url": "https://omlx.ai/benchmarks/performance/halqel6o",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-halqel6o-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: it7kgmxn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 489.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 489.4 16.7 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/halqel6o",
   "source_url": "https://omlx.ai/benchmarks/performance/halqel6o",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nzpuf44n-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.2 GB at this context; duplicate traces of the same session: 09dem841, rdlc7z4g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 881.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 881.0 59.7 180.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nzpuf44n",
   "source_url": "https://omlx.ai/benchmarks/performance/nzpuf44n",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nzpuf44n-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.7 GB at this context; duplicate traces of the same session: 09dem841, rdlc7z4g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1218.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 1,218 52.7 180.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nzpuf44n",
   "source_url": "https://omlx.ai/benchmarks/performance/nzpuf44n",
   "tg_tokens": null,
   "tps": 52.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nzpuf44n-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.8 GB at this context; duplicate traces of the same session: 09dem841, rdlc7z4g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1409.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 1,409 56.2 180.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nzpuf44n",
   "source_url": "https://omlx.ai/benchmarks/performance/nzpuf44n",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-gc5crqs0-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 373t77ij, possm4bw, r30015f1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 254.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 254.2 49.6 18.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gc5crqs0",
   "source_url": "https://omlx.ai/benchmarks/performance/gc5crqs0",
   "tg_tokens": null,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-gc5crqs0-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 373t77ij, possm4bw, r30015f1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 269.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 269.3 67.4 22.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gc5crqs0",
   "source_url": "https://omlx.ai/benchmarks/performance/gc5crqs0",
   "tg_tokens": null,
   "tps": 67.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-gc5crqs0-ctx131072",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.9 GB at this context; duplicate traces of the same session: 373t77ij, possm4bw, r30015f1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 164.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 164.0 38.9 31.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gc5crqs0",
   "source_url": "https://omlx.ai/benchmarks/performance/gc5crqs0",
   "tg_tokens": null,
   "tps": 38.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-gc5crqs0-ctx199680",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: 373t77ij, possm4bw, r30015f1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 135.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 195k 135.3 34.3 36.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gc5crqs0",
   "source_url": "https://omlx.ai/benchmarks/performance/gc5crqs0",
   "tg_tokens": null,
   "tps": 34.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-egwx21er-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; Swift 1.5 build; oMLX community benchmark trace; peak memory 74.1 GB at this context; duplicate traces of the same session: pqhmopsj; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 276.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 1k 276.2 10.0 74.1 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/egwx21er",
   "source_url": "https://omlx.ai/benchmarks/performance/egwx21er",
   "tg_tokens": null,
   "tps": 10.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-egwx21er-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; Swift 1.5 build; oMLX community benchmark trace; peak memory 81.1 GB at this context; duplicate traces of the same session: pqhmopsj; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 341.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 4k 341.4 14.6 81.1 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/egwx21er",
   "source_url": "https://omlx.ai/benchmarks/performance/egwx21er",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-egwx21er-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; Swift 1.5 build; oMLX community benchmark trace; peak memory 84.1 GB at this context; duplicate traces of the same session: pqhmopsj; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 621.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 8k 621.0 16.8 84.1 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/egwx21er",
   "source_url": "https://omlx.ai/benchmarks/performance/egwx21er",
   "tg_tokens": null,
   "tps": 16.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-egwx21er-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; Swift 1.5 build; oMLX community benchmark trace; peak memory 84.8 GB at this context; duplicate traces of the same session: pqhmopsj; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 616.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 16k 616.7 19.9 84.8 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/egwx21er",
   "source_url": "https://omlx.ai/benchmarks/performance/egwx21er",
   "tg_tokens": null,
   "tps": 19.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-keo0vtsy-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 22.6 GB; TTFT 7636 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1073.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k tokens 1,073 PP tok/s 65.7 TG tok/s 7636 TTFT (ms) 22.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/keo0vtsy",
   "source_url": "https://omlx.ai/benchmarks/performance/keo0vtsy",
   "tg_tokens": null,
   "tps": 65.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-fq3axz69-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 20.1 GB; TTFT 1292 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 792.8,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k tokens 792.8 PP tok/s 74.0 TG tok/s 1292 TTFT (ms) 20.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fq3axz69",
   "source_url": "https://omlx.ai/benchmarks/performance/fq3axz69",
   "tg_tokens": null,
   "tps": 74.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-3zqld31n-ctx1024",
   "model": "Qwen3.5-9B-Claude-4.6-HighIQ-INSTRUCT-HERETIC-UNCENSORED",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: 4hgi1b8h; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 234.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-9B-Claude-4.6-HighIQ-INSTRUCT-HERETIC-UNCENSORED-oQ4e on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 234.2 21.2 5.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3zqld31n",
   "source_url": "https://omlx.ai/benchmarks/performance/3zqld31n",
   "tg_tokens": null,
   "tps": 21.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-3zqld31n-ctx4096",
   "model": "Qwen3.5-9B-Claude-4.6-HighIQ-INSTRUCT-HERETIC-UNCENSORED",
   "notes": "oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: 4hgi1b8h; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 241.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-9B-Claude-4.6-HighIQ-INSTRUCT-HERETIC-UNCENSORED-oQ4e on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 241.5 20.4 6.3 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3zqld31n",
   "source_url": "https://omlx.ai/benchmarks/performance/3zqld31n",
   "tg_tokens": null,
   "tps": 20.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-w5iqms2k-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: 9oxxp7f5; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 313.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 313.4 9.8 17.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w5iqms2k",
   "source_url": "https://omlx.ai/benchmarks/performance/w5iqms2k",
   "tg_tokens": null,
   "tps": 9.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-w5iqms2k-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 9oxxp7f5; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 302.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 302.4 9.2 17.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w5iqms2k",
   "source_url": "https://omlx.ai/benchmarks/performance/w5iqms2k",
   "tg_tokens": null,
   "tps": 9.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-yl1s15x7-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized speed FP16 build; oMLX community benchmark trace, single run at 1k context; peak memory 23.6 GB; TTFT 8731 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 117.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed-FP16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k tokens 117.3 PP tok/s 17.3 TG tok/s 8731 TTFT (ms) 23.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yl1s15x7",
   "source_url": "https://omlx.ai/benchmarks/performance/yl1s15x7",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-krd47hja-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized speed FP16 build; oMLX community benchmark trace, single run at 8k context; peak memory 28.8 GB; TTFT 56401 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 145.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed-FP16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k tokens 145.2 PP tok/s 21.9 TG tok/s 56401 TTFT (ms) 28.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/krd47hja",
   "source_url": "https://omlx.ai/benchmarks/performance/krd47hja",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-6zj6g2ct-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 25.2 GB; TTFT 18436 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 444.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k tokens 444.3 PP tok/s 32.6 TG tok/s 18436 TTFT (ms) 25.2 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6zj6g2ct",
   "source_url": "https://omlx.ai/benchmarks/performance/6zj6g2ct",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5uf6nn07-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 26.0 GB; TTFT 38632 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 424.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 16k tokens 424.1 PP tok/s 49.1 TG tok/s 38632 TTFT (ms) 26.0 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5uf6nn07",
   "source_url": "https://omlx.ai/benchmarks/performance/5uf6nn07",
   "tg_tokens": null,
   "tps": 49.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ihtmjlnz-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 21.2 GB; TTFT 3158 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 324.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k tokens 324.3 PP tok/s 47.1 TG tok/s 3158 TTFT (ms) 21.2 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ihtmjlnz",
   "source_url": "https://omlx.ai/benchmarks/performance/ihtmjlnz",
   "tg_tokens": null,
   "tps": 47.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-iyz9q2tt-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 18.5 GB; TTFT 3170 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k tokens 323.0 PP tok/s 30.4 TG tok/s 3170 TTFT (ms) 18.5 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iyz9q2tt",
   "source_url": "https://omlx.ai/benchmarks/performance/iyz9q2tt",
   "tg_tokens": null,
   "tps": 30.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-yhm2ber1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 22.1 GB; TTFT 11865 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 345.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k tokens 345.2 PP tok/s 34.4 TG tok/s 11865 TTFT (ms) 22.1 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yhm2ber1",
   "source_url": "https://omlx.ai/benchmarks/performance/yhm2ber1",
   "tg_tokens": null,
   "tps": 34.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-lbcis16x-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 71.8 GB; TTFT 12742 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 642.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k tokens 642.9 PP tok/s 62.8 TG tok/s 12742 TTFT (ms) 71.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lbcis16x",
   "source_url": "https://omlx.ai/benchmarks/performance/lbcis16x",
   "tg_tokens": null,
   "tps": 62.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-nsibfqmi-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: hulj8gfi; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 108.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 108.7 26.5 18.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nsibfqmi",
   "source_url": "https://omlx.ai/benchmarks/performance/nsibfqmi",
   "tg_tokens": null,
   "tps": 26.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-nsibfqmi-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: hulj8gfi; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 102.2 21.1 22.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nsibfqmi",
   "source_url": "https://omlx.ai/benchmarks/performance/nsibfqmi",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-j7uu9jl5-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 108.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 108.9 25.2 18.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j7uu9jl5",
   "source_url": "https://omlx.ai/benchmarks/performance/j7uu9jl5",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-j7uu9jl5-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 102.2 22.8 22.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j7uu9jl5",
   "source_url": "https://omlx.ai/benchmarks/performance/j7uu9jl5",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-viarwncy-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 111.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 111.3 18.1 16.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/viarwncy",
   "source_url": "https://omlx.ai/benchmarks/performance/viarwncy",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-viarwncy-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 104.0 17.4 20.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/viarwncy",
   "source_url": "https://omlx.ai/benchmarks/performance/viarwncy",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-2f1v3l2w-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; Swift 1.5 build; oMLX community benchmark trace; peak memory 74.1 GB at this context; duplicate traces of the same session: uq3ctiaq; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 276.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 1k 276.2 10.0 74.1 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2f1v3l2w",
   "source_url": "https://omlx.ai/benchmarks/performance/2f1v3l2w",
   "tg_tokens": null,
   "tps": 10.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-2f1v3l2w-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; Swift 1.5 build; oMLX community benchmark trace; peak memory 81.1 GB at this context; duplicate traces of the same session: uq3ctiaq; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 341.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 4k 341.4 14.6 81.1 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2f1v3l2w",
   "source_url": "https://omlx.ai/benchmarks/performance/2f1v3l2w",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-2f1v3l2w-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; Swift 1.5 build; oMLX community benchmark trace; peak memory 84.1 GB at this context; duplicate traces of the same session: uq3ctiaq; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 621.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 8k 621.0 16.8 84.1 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2f1v3l2w",
   "source_url": "https://omlx.ai/benchmarks/performance/2f1v3l2w",
   "tg_tokens": null,
   "tps": 16.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-2f1v3l2w-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; Swift 1.5 build; oMLX community benchmark trace; peak memory 84.8 GB at this context; duplicate traces of the same session: uq3ctiaq; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 616.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 16k 616.7 19.9 84.8 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2f1v3l2w",
   "source_url": "https://omlx.ai/benchmarks/performance/2f1v3l2w",
   "tg_tokens": null,
   "tps": 19.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-fuwy3r0b-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: r2a2g4cw; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 788.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 788.4 37.1 14.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fuwy3r0b",
   "source_url": "https://omlx.ai/benchmarks/performance/fuwy3r0b",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-fuwy3r0b-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.9 GB at this context; duplicate traces of the same session: r2a2g4cw; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 756.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 16k 756.5 31.4 14.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fuwy3r0b",
   "source_url": "https://omlx.ai/benchmarks/performance/fuwy3r0b",
   "tg_tokens": null,
   "tps": 31.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-u6b0cj16-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 9pl06b8q; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 81.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k 81.4 15.4 17.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u6b0cj16",
   "source_url": "https://omlx.ai/benchmarks/performance/u6b0cj16",
   "tg_tokens": null,
   "tps": 15.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-u6b0cj16-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: 9pl06b8q; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 72.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 16k 72.7 13.9 18.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u6b0cj16",
   "source_url": "https://omlx.ai/benchmarks/performance/u6b0cj16",
   "tg_tokens": null,
   "tps": 13.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-gxbkpj9m-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 39.8 GB at this context; duplicate traces of the same session: al4iobg6; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 440.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k 440.3 35.6 39.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gxbkpj9m",
   "source_url": "https://omlx.ai/benchmarks/performance/gxbkpj9m",
   "tg_tokens": null,
   "tps": 35.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-gxbkpj9m-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 40.0 GB at this context; duplicate traces of the same session: al4iobg6; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 933.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 4k 933.3 33.3 40.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gxbkpj9m",
   "source_url": "https://omlx.ai/benchmarks/performance/gxbkpj9m",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-e4c54yld-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 108.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 108.9 25.2 18.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e4c54yld",
   "source_url": "https://omlx.ai/benchmarks/performance/e4c54yld",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-e4c54yld-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 102.2 22.8 22.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e4c54yld",
   "source_url": "https://omlx.ai/benchmarks/performance/e4c54yld",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-bg40n9se-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 111.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 111.3 18.1 16.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bg40n9se",
   "source_url": "https://omlx.ai/benchmarks/performance/bg40n9se",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-bg40n9se-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 104.0 17.4 20.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bg40n9se",
   "source_url": "https://omlx.ai/benchmarks/performance/bg40n9se",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-qcl8w693-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: ne8fowwk, p1pgrsvo, s78yqp7a, y89c1mj7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 254.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 254.0 68.1 18.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qcl8w693",
   "source_url": "https://omlx.ai/benchmarks/performance/qcl8w693",
   "tg_tokens": null,
   "tps": 68.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-qcl8w693-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: ne8fowwk, p1pgrsvo, s78yqp7a, y89c1mj7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 268.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 268.7 93.1 22.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qcl8w693",
   "source_url": "https://omlx.ai/benchmarks/performance/qcl8w693",
   "tg_tokens": null,
   "tps": 93.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-qcl8w693-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: ne8fowwk, p1pgrsvo, s78yqp7a, y89c1mj7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 264.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 264.1 39.9 22.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qcl8w693",
   "source_url": "https://omlx.ai/benchmarks/performance/qcl8w693",
   "tg_tokens": null,
   "tps": 39.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-qcl8w693-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: ne8fowwk, p1pgrsvo, s78yqp7a, y89c1mj7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 253.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 253.9 41.1 23.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qcl8w693",
   "source_url": "https://omlx.ai/benchmarks/performance/qcl8w693",
   "tg_tokens": null,
   "tps": 41.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-qcl8w693-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.2 GB at this context; duplicate traces of the same session: ne8fowwk, p1pgrsvo, s78yqp7a, y89c1mj7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 235.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 235.2 52.4 25.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qcl8w693",
   "source_url": "https://omlx.ai/benchmarks/performance/qcl8w693",
   "tg_tokens": null,
   "tps": 52.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-cjufv2ox-ctx1024",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 4.0 GB at this context; duplicate traces of the same session: 2a9olu06, 5prhwf0e, 6knzy753, ks0e4528; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 9789.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e2b-it-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 9,789 146.3 4.0 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cjufv2ox",
   "source_url": "https://omlx.ai/benchmarks/performance/cjufv2ox",
   "tg_tokens": null,
   "tps": 146.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-cjufv2ox-ctx4096",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 4.0 GB at this context; duplicate traces of the same session: 2a9olu06, 5prhwf0e, 6knzy753, ks0e4528; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 11106.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e2b-it-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 11,106 201.9 4.0 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cjufv2ox",
   "source_url": "https://omlx.ai/benchmarks/performance/cjufv2ox",
   "tg_tokens": null,
   "tps": 201.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-cjufv2ox-ctx8192",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 4.1 GB at this context; duplicate traces of the same session: 2a9olu06, 5prhwf0e, 6knzy753, ks0e4528; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 10696.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e2b-it-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 8k 10,696 182.8 4.1 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cjufv2ox",
   "source_url": "https://omlx.ai/benchmarks/performance/cjufv2ox",
   "tg_tokens": null,
   "tps": 182.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-cjufv2ox-ctx16384",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 4.3 GB at this context; duplicate traces of the same session: 2a9olu06, 5prhwf0e, 6knzy753, ks0e4528; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 9855.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e2b-it-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 16k 9,855 172.7 4.3 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cjufv2ox",
   "source_url": "https://omlx.ai/benchmarks/performance/cjufv2ox",
   "tg_tokens": null,
   "tps": 172.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-cjufv2ox-ctx32768",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 4.7 GB at this context; duplicate traces of the same session: 2a9olu06, 5prhwf0e, 6knzy753, ks0e4528; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 8580.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e2b-it-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 32k 8,580 134.5 4.7 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cjufv2ox",
   "source_url": "https://omlx.ai/benchmarks/performance/cjufv2ox",
   "tg_tokens": null,
   "tps": 134.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-sp0we8o7-ctx1024",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 3.0 GB at this context; duplicate traces of the same session: 47jy2tmj, 6m144l4x, og95fy1u, yjs13tdo; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 9249.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e2b-it-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 9,249 141.3 3.0 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sp0we8o7",
   "source_url": "https://omlx.ai/benchmarks/performance/sp0we8o7",
   "tg_tokens": null,
   "tps": 141.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-sp0we8o7-ctx4096",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 3.0 GB at this context; duplicate traces of the same session: 47jy2tmj, 6m144l4x, og95fy1u, yjs13tdo; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 10894.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e2b-it-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 10,894 130.6 3.0 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sp0we8o7",
   "source_url": "https://omlx.ai/benchmarks/performance/sp0we8o7",
   "tg_tokens": null,
   "tps": 130.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-sp0we8o7-ctx8192",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 3.1 GB at this context; duplicate traces of the same session: 47jy2tmj, 6m144l4x, og95fy1u, yjs13tdo; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 10542.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e2b-it-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 8k 10,542 126.7 3.1 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sp0we8o7",
   "source_url": "https://omlx.ai/benchmarks/performance/sp0we8o7",
   "tg_tokens": null,
   "tps": 126.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-sp0we8o7-ctx16384",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 3.2 GB at this context; duplicate traces of the same session: 47jy2tmj, 6m144l4x, og95fy1u, yjs13tdo; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 9865.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e2b-it-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 16k 9,865 119.1 3.2 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sp0we8o7",
   "source_url": "https://omlx.ai/benchmarks/performance/sp0we8o7",
   "tg_tokens": null,
   "tps": 119.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-sp0we8o7-ctx32768",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 3.6 GB at this context; duplicate traces of the same session: 47jy2tmj, 6m144l4x, og95fy1u, yjs13tdo; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 8604.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e2b-it-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 32k 8,604 107.1 3.6 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sp0we8o7",
   "source_url": "https://omlx.ai/benchmarks/performance/sp0we8o7",
   "tg_tokens": null,
   "tps": 107.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-p38280gm-ctx1024",
   "model": "gemma-4-E4B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3145.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-qat-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 3,145 64.4 6.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p38280gm",
   "source_url": "https://omlx.ai/benchmarks/performance/p38280gm",
   "tg_tokens": null,
   "tps": 64.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-p38280gm-ctx4096",
   "model": "gemma-4-E4B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3413.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-qat-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 3,413 62.5 6.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p38280gm",
   "source_url": "https://omlx.ai/benchmarks/performance/p38280gm",
   "tg_tokens": null,
   "tps": 62.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-p38280gm-ctx8192",
   "model": "gemma-4-E4B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3388.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-qat-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 8k 3,388 59.1 6.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p38280gm",
   "source_url": "https://omlx.ai/benchmarks/performance/p38280gm",
   "tg_tokens": null,
   "tps": 59.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-p38280gm-ctx16384",
   "model": "gemma-4-E4B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3266.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-qat-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 16k 3,266 55.5 7.0 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p38280gm",
   "source_url": "https://omlx.ai/benchmarks/performance/p38280gm",
   "tg_tokens": null,
   "tps": 55.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-p38280gm-ctx32768",
   "model": "gemma-4-E4B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 7.5 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3000.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-qat-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 32k 3,000 47.9 7.5 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p38280gm",
   "source_url": "https://omlx.ai/benchmarks/performance/p38280gm",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bstms34w-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "UNCENSORED MTP build; oMLX community benchmark trace; peak memory 288.7 GB at this context; duplicate traces of the same session: hofznow2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 231.3,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "DeepSeek-V4.1-Flash-UNCENSORED-oQ4-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 231.3 18.6 288.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bstms34w",
   "source_url": "https://omlx.ai/benchmarks/performance/bstms34w",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bstms34w-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "UNCENSORED MTP build; oMLX community benchmark trace; peak memory 289.4 GB at this context; duplicate traces of the same session: hofznow2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 328.8,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "DeepSeek-V4.1-Flash-UNCENSORED-oQ4-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 328.8 20.8 289.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bstms34w",
   "source_url": "https://omlx.ai/benchmarks/performance/bstms34w",
   "tg_tokens": null,
   "tps": 20.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bstms34w-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "UNCENSORED MTP build; oMLX community benchmark trace; peak memory 289.4 GB at this context; duplicate traces of the same session: hofznow2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 348.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "DeepSeek-V4.1-Flash-UNCENSORED-oQ4-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 8k 348.0 22.8 289.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bstms34w",
   "source_url": "https://omlx.ai/benchmarks/performance/bstms34w",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-il8d5viq-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 5.2 GB at this context; duplicate traces of the same session: xrg7owb7; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 425.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-4B-MLX-8bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 8bit; 1k 425.5 21.8 5.2 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/il8d5viq",
   "source_url": "https://omlx.ai/benchmarks/performance/il8d5viq",
   "tg_tokens": null,
   "tps": 21.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-il8d5viq-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 5.7 GB at this context; duplicate traces of the same session: xrg7owb7; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 428.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-4B-MLX-8bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 8bit; 4k 428.1 21.2 5.7 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/il8d5viq",
   "source_url": "https://omlx.ai/benchmarks/performance/il8d5viq",
   "tg_tokens": null,
   "tps": 21.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-evewhfly-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: j4dslo54; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 235.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 235.4 21.4 5.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/evewhfly",
   "source_url": "https://omlx.ai/benchmarks/performance/evewhfly",
   "tg_tokens": null,
   "tps": 21.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-evewhfly-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: j4dslo54; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 242.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 242.0 20.9 6.2 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/evewhfly",
   "source_url": "https://omlx.ai/benchmarks/performance/evewhfly",
   "tg_tokens": null,
   "tps": 20.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-3l5ruj91-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 29.1 GB at this context; duplicate traces of the same session: jgmozxn8; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1131.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 5bit; 4k 1,131 80.4 29.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3l5ruj91",
   "source_url": "https://omlx.ai/benchmarks/performance/3l5ruj91",
   "tg_tokens": null,
   "tps": 80.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-3l5ruj91-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: jgmozxn8; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 933.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 5bit; 16k 933.9 90.2 30.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3l5ruj91",
   "source_url": "https://omlx.ai/benchmarks/performance/3l5ruj91",
   "tg_tokens": null,
   "tps": 90.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-80bvxhkl-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.3 GB at this context; duplicate traces of the same session: 9ryqyvty; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1006.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,006 51.6 29.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/80bvxhkl",
   "source_url": "https://omlx.ai/benchmarks/performance/80bvxhkl",
   "tg_tokens": null,
   "tps": 51.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-80bvxhkl-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 30.4 GB at this context; duplicate traces of the same session: 9ryqyvty; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 960.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 960.6 42.8 30.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/80bvxhkl",
   "source_url": "https://omlx.ai/benchmarks/performance/80bvxhkl",
   "tg_tokens": null,
   "tps": 42.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-g1q750pm-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 30.9 GB; TTFT 156270 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 419.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 96 GB \u00b7 4bit; 64k tokens 419.4 PP tok/s 20.9 TG tok/s 156270 TTFT (ms) 30.9 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g1q750pm",
   "source_url": "https://omlx.ai/benchmarks/performance/g1q750pm",
   "tg_tokens": null,
   "tps": 20.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-0epdnifc-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; build 20261011-c7e8cd38; oMLX community benchmark trace; peak memory 69.3 GB at this context; duplicate traces of the same session: 6flxot3i, ayyp2vva, gh0zf21w; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 555.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-20261011-c7e8cd38 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 555.3 64.1 69.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0epdnifc",
   "source_url": "https://omlx.ai/benchmarks/performance/0epdnifc",
   "tg_tokens": null,
   "tps": 64.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-0epdnifc-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; build 20261011-c7e8cd38; oMLX community benchmark trace; peak memory 69.9 GB at this context; duplicate traces of the same session: 6flxot3i, ayyp2vva, gh0zf21w; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1089.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-20261011-c7e8cd38 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 1,089 64.3 69.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0epdnifc",
   "source_url": "https://omlx.ai/benchmarks/performance/0epdnifc",
   "tg_tokens": null,
   "tps": 64.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-0epdnifc-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; build 20261011-c7e8cd38; oMLX community benchmark trace; peak memory 70.1 GB at this context; duplicate traces of the same session: 6flxot3i, ayyp2vva, gh0zf21w; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1095.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-20261011-c7e8cd38 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 16k 1,095 72.0 70.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0epdnifc",
   "source_url": "https://omlx.ai/benchmarks/performance/0epdnifc",
   "tg_tokens": null,
   "tps": 72.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-0epdnifc-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; build 20261011-c7e8cd38; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: 6flxot3i, ayyp2vva, gh0zf21w; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1075.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-20261011-c7e8cd38 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k 1,075 62.5 71.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0epdnifc",
   "source_url": "https://omlx.ai/benchmarks/performance/0epdnifc",
   "tg_tokens": null,
   "tps": 62.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-0epdnifc-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; build 20261011-c7e8cd38; oMLX community benchmark trace; peak memory 74.9 GB at this context; duplicate traces of the same session: 6flxot3i, ayyp2vva, gh0zf21w; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1058.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-20261011-c7e8cd38 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 128k 1,058 54.9 74.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0epdnifc",
   "source_url": "https://omlx.ai/benchmarks/performance/0epdnifc",
   "tg_tokens": null,
   "tps": 54.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-uk4xv054-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: pkn2jqhz; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 844.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 844.0 39.5 14.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uk4xv054",
   "source_url": "https://omlx.ai/benchmarks/performance/uk4xv054",
   "tg_tokens": null,
   "tps": 39.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-uk4xv054-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.9 GB at this context; duplicate traces of the same session: pkn2jqhz; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 801.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 16k 801.2 32.5 14.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uk4xv054",
   "source_url": "https://omlx.ai/benchmarks/performance/uk4xv054",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-3l5g6ozl-ctx1024",
   "model": "gemma-4-E4B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: 0n6m67i7, jw1g318b, vu8jocvk; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3145.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-qat-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 3,145 64.4 6.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3l5g6ozl",
   "source_url": "https://omlx.ai/benchmarks/performance/3l5g6ozl",
   "tg_tokens": null,
   "tps": 64.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-3l5g6ozl-ctx4096",
   "model": "gemma-4-E4B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: 0n6m67i7, jw1g318b, vu8jocvk; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3413.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-qat-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 3,413 62.5 6.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3l5g6ozl",
   "source_url": "https://omlx.ai/benchmarks/performance/3l5g6ozl",
   "tg_tokens": null,
   "tps": 62.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-3l5g6ozl-ctx8192",
   "model": "gemma-4-E4B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: 0n6m67i7, jw1g318b, vu8jocvk; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3388.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-qat-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 8k 3,388 59.1 6.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3l5g6ozl",
   "source_url": "https://omlx.ai/benchmarks/performance/3l5g6ozl",
   "tg_tokens": null,
   "tps": 59.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-3l5g6ozl-ctx16384",
   "model": "gemma-4-E4B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: 0n6m67i7, jw1g318b, vu8jocvk; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3266.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-qat-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 16k 3,266 55.5 7.0 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3l5g6ozl",
   "source_url": "https://omlx.ai/benchmarks/performance/3l5g6ozl",
   "tg_tokens": null,
   "tps": 55.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-3l5g6ozl-ctx32768",
   "model": "gemma-4-E4B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 7.5 GB at this context; duplicate traces of the same session: 0n6m67i7, jw1g318b, vu8jocvk; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3000.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-qat-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 32k 3,000 47.9 7.5 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3l5g6ozl",
   "source_url": "https://omlx.ai/benchmarks/performance/3l5g6ozl",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lueb9n0n-ctx32768",
   "model": "Qwen3 4B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 9.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2643.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-4B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 32k 2,643 52.0 9.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lueb9n0n",
   "source_url": "https://omlx.ai/benchmarks/performance/lueb9n0n",
   "tg_tokens": null,
   "tps": 52.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lueb9n0n-ctx65536",
   "model": "Qwen3 4B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1352.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-4B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 64k 1,352 35.2 13.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lueb9n0n",
   "source_url": "https://omlx.ai/benchmarks/performance/lueb9n0n",
   "tg_tokens": null,
   "tps": 35.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-u245e1gq-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "UNCENSORED MTP build; oMLX community benchmark trace; peak memory 288.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 231.3,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "DeepSeek-V4.1-Flash-UNCENSORED-oQ4-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 231.3 18.6 288.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u245e1gq",
   "source_url": "https://omlx.ai/benchmarks/performance/u245e1gq",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-u245e1gq-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "UNCENSORED MTP build; oMLX community benchmark trace; peak memory 289.4 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 328.8,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "DeepSeek-V4.1-Flash-UNCENSORED-oQ4-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 328.8 20.8 289.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u245e1gq",
   "source_url": "https://omlx.ai/benchmarks/performance/u245e1gq",
   "tg_tokens": null,
   "tps": 20.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-u245e1gq-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "UNCENSORED MTP build; oMLX community benchmark trace; peak memory 289.4 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 348.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "DeepSeek-V4.1-Flash-UNCENSORED-oQ4-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 8k 348.0 22.8 289.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u245e1gq",
   "source_url": "https://omlx.ai/benchmarks/performance/u245e1gq",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gkhj9mqo-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 181.6 GB; TTFT 335059 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 391.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 128k tokens 391.2 PP tok/s 81.9 TG tok/s 335059 TTFT (ms) 181.6 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gkhj9mqo",
   "source_url": "https://omlx.ai/benchmarks/performance/gkhj9mqo",
   "tg_tokens": null,
   "tps": 81.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-n46zllsr-ctx8192",
   "model": "Qwen3 4B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: cxsydmku; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4528.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-4B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 8k 4,528 90.0 5.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n46zllsr",
   "source_url": "https://omlx.ai/benchmarks/performance/n46zllsr",
   "tg_tokens": null,
   "tps": 90.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-n46zllsr-ctx16384",
   "model": "Qwen3 4B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: cxsydmku; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3844.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-4B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 16k 3,844 74.5 6.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n46zllsr",
   "source_url": "https://omlx.ai/benchmarks/performance/n46zllsr",
   "tg_tokens": null,
   "tps": 74.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-19nr2kos-ctx32768",
   "model": "Qwen3 4B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 9.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2643.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-4B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 32k 2,643 52.0 9.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/19nr2kos",
   "source_url": "https://omlx.ai/benchmarks/performance/19nr2kos",
   "tg_tokens": null,
   "tps": 52.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-19nr2kos-ctx65536",
   "model": "Qwen3 4B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1352.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-4B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 64k 1,352 35.2 13.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/19nr2kos",
   "source_url": "https://omlx.ai/benchmarks/performance/19nr2kos",
   "tg_tokens": null,
   "tps": 35.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-u1fqupyt-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 30.9 GB; TTFT 156229 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 419.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 96 GB \u00b7 4bit; 64k tokens 419.5 PP tok/s 33.6 TG tok/s 156229 TTFT (ms) 30.9 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u1fqupyt",
   "source_url": "https://omlx.ai/benchmarks/performance/u1fqupyt",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-puvxwk7x-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: 3dhi3513, lo1ze90l, yw1u0gho; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.7,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 104.7 14.1 29.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/puvxwk7x",
   "source_url": "https://omlx.ai/benchmarks/performance/puvxwk7x",
   "tg_tokens": null,
   "tps": 14.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-puvxwk7x-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.4 GB at this context; duplicate traces of the same session: 3dhi3513, lo1ze90l, yw1u0gho; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 129.6,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 129.6 13.4 33.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/puvxwk7x",
   "source_url": "https://omlx.ai/benchmarks/performance/puvxwk7x",
   "tg_tokens": null,
   "tps": 13.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-puvxwk7x-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.6 GB at this context; duplicate traces of the same session: 3dhi3513, lo1ze90l, yw1u0gho; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.3,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 130.3 12.9 34.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/puvxwk7x",
   "source_url": "https://omlx.ai/benchmarks/performance/puvxwk7x",
   "tg_tokens": null,
   "tps": 12.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-puvxwk7x-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.6 GB at this context; duplicate traces of the same session: 3dhi3513, lo1ze90l, yw1u0gho; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.5,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 125.5 17.3 35.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/puvxwk7x",
   "source_url": "https://omlx.ai/benchmarks/performance/puvxwk7x",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-oy4s2n8g-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: 4io3zuj6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 322.9,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 322.9 54.3 23.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oy4s2n8g",
   "source_url": "https://omlx.ai/benchmarks/performance/oy4s2n8g",
   "tg_tokens": null,
   "tps": 54.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-oy4s2n8g-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: 4io3zuj6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 303.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 16k 303.0 68.2 25.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oy4s2n8g",
   "source_url": "https://omlx.ai/benchmarks/performance/oy4s2n8g",
   "tg_tokens": null,
   "tps": 68.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lifd0t5i-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; build 20261011-c7e8cd38; oMLX community benchmark trace; peak memory 69.3 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 555.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-20261011-c7e8cd38 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 555.3 64.1 69.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lifd0t5i",
   "source_url": "https://omlx.ai/benchmarks/performance/lifd0t5i",
   "tg_tokens": null,
   "tps": 64.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lifd0t5i-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; build 20261011-c7e8cd38; oMLX community benchmark trace; peak memory 69.9 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1089.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-20261011-c7e8cd38 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 1,089 64.3 69.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lifd0t5i",
   "source_url": "https://omlx.ai/benchmarks/performance/lifd0t5i",
   "tg_tokens": null,
   "tps": 64.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lifd0t5i-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; build 20261011-c7e8cd38; oMLX community benchmark trace; peak memory 70.1 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1095.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-20261011-c7e8cd38 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 16k 1,095 72.0 70.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lifd0t5i",
   "source_url": "https://omlx.ai/benchmarks/performance/lifd0t5i",
   "tg_tokens": null,
   "tps": 72.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lifd0t5i-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; build 20261011-c7e8cd38; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1075.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-20261011-c7e8cd38 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k 1,075 62.5 71.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lifd0t5i",
   "source_url": "https://omlx.ai/benchmarks/performance/lifd0t5i",
   "tg_tokens": null,
   "tps": 62.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lifd0t5i-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; build 20261011-c7e8cd38; oMLX community benchmark trace; peak memory 74.9 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1058.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-20261011-c7e8cd38 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 128k 1,058 54.9 74.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lifd0t5i",
   "source_url": "https://omlx.ai/benchmarks/performance/lifd0t5i",
   "tg_tokens": null,
   "tps": 54.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-u9gd4hr3-ctx1024",
   "model": "Spark-X2.5-4B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: d2reh2wc, p0vyn6z1, u2v23c0d, v2poxxyt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1308.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Spark-X2.5-4B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 8bit; 1k 1,308 91.0 4.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u9gd4hr3",
   "source_url": "https://omlx.ai/benchmarks/performance/u9gd4hr3",
   "tg_tokens": null,
   "tps": 91.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-u9gd4hr3-ctx4096",
   "model": "Spark-X2.5-4B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: d2reh2wc, p0vyn6z1, u2v23c0d, v2poxxyt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1408.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Spark-X2.5-4B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 8bit; 4k 1,408 89.6 4.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u9gd4hr3",
   "source_url": "https://omlx.ai/benchmarks/performance/u9gd4hr3",
   "tg_tokens": null,
   "tps": 89.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-u9gd4hr3-ctx8192",
   "model": "Spark-X2.5-4B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 5.0 GB at this context; duplicate traces of the same session: d2reh2wc, p0vyn6z1, u2v23c0d, v2poxxyt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1384.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Spark-X2.5-4B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 8bit; 8k 1,384 86.2 5.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u9gd4hr3",
   "source_url": "https://omlx.ai/benchmarks/performance/u9gd4hr3",
   "tg_tokens": null,
   "tps": 86.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-u9gd4hr3-ctx16384",
   "model": "Spark-X2.5-4B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 5.3 GB at this context; duplicate traces of the same session: d2reh2wc, p0vyn6z1, u2v23c0d, v2poxxyt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1142.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Spark-X2.5-4B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 8bit; 16k 1,142 81.1 5.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u9gd4hr3",
   "source_url": "https://omlx.ai/benchmarks/performance/u9gd4hr3",
   "tg_tokens": null,
   "tps": 81.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-u9gd4hr3-ctx32768",
   "model": "Spark-X2.5-4B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: d2reh2wc, p0vyn6z1, u2v23c0d, v2poxxyt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 971.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Spark-X2.5-4B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 8bit; 32k 971.4 72.0 5.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u9gd4hr3",
   "source_url": "https://omlx.ai/benchmarks/performance/u9gd4hr3",
   "tg_tokens": null,
   "tps": 72.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-tmrcopnm-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai build; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 82ibqxlk, bcjdyrao, yc90bz3h; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2829.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 2,829 116.8 19.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tmrcopnm",
   "source_url": "https://omlx.ai/benchmarks/performance/tmrcopnm",
   "tg_tokens": null,
   "tps": 116.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-tmrcopnm-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai build; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 82ibqxlk, bcjdyrao, yc90bz3h; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3523.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 3,523 115.2 21.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tmrcopnm",
   "source_url": "https://omlx.ai/benchmarks/performance/tmrcopnm",
   "tg_tokens": null,
   "tps": 115.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-tmrcopnm-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai build; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 82ibqxlk, bcjdyrao, yc90bz3h; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3335.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 3,335 107.1 21.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tmrcopnm",
   "source_url": "https://omlx.ai/benchmarks/performance/tmrcopnm",
   "tg_tokens": null,
   "tps": 107.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-tmrcopnm-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 82ibqxlk, bcjdyrao, yc90bz3h; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3003.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 3,003 99.8 21.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tmrcopnm",
   "source_url": "https://omlx.ai/benchmarks/performance/tmrcopnm",
   "tg_tokens": null,
   "tps": 99.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-tmrcopnm-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 82ibqxlk, bcjdyrao, yc90bz3h; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2501.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 2,501 95.5 22.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tmrcopnm",
   "source_url": "https://omlx.ai/benchmarks/performance/tmrcopnm",
   "tg_tokens": null,
   "tps": 95.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-tmrcopnm-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai build; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: 82ibqxlk, bcjdyrao, yc90bz3h; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1861.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 1,861 80.8 23.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tmrcopnm",
   "source_url": "https://omlx.ai/benchmarks/performance/tmrcopnm",
   "tg_tokens": null,
   "tps": 80.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-b8qa7j5g-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "EigenLabs; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: s4pkjt23; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1121.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "EigenLabs--Qwen3.8-27B-4bit-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,121 70.8 16.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b8qa7j5g",
   "source_url": "https://omlx.ai/benchmarks/performance/b8qa7j5g",
   "tg_tokens": null,
   "tps": 70.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-b8qa7j5g-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "EigenLabs; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: s4pkjt23; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1446.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "EigenLabs--Qwen3.8-27B-4bit-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,446 92.2 18.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b8qa7j5g",
   "source_url": "https://omlx.ai/benchmarks/performance/b8qa7j5g",
   "tg_tokens": null,
   "tps": 92.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rxjznetn-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "EigenLabs; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: iimf2oxg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1400.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "EigenLabs--Qwen3.8-27B-4bit-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,400 67.0 18.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rxjznetn",
   "source_url": "https://omlx.ai/benchmarks/performance/rxjznetn",
   "tg_tokens": null,
   "tps": 67.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rxjznetn-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "EigenLabs; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: iimf2oxg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1201.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "EigenLabs--Qwen3.8-27B-4bit-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,201 144.6 20.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rxjznetn",
   "source_url": "https://omlx.ai/benchmarks/performance/rxjznetn",
   "tg_tokens": null,
   "tps": 144.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-4xav2klz-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: 15dep3h7; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 385.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 385.7 49.1 6.3 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4xav2klz",
   "source_url": "https://omlx.ai/benchmarks/performance/4xav2klz",
   "tg_tokens": null,
   "tps": 49.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-4xav2klz-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: 15dep3h7; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 367.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 16k 367.8 44.3 6.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4xav2klz",
   "source_url": "https://omlx.ai/benchmarks/performance/4xav2klz",
   "tg_tokens": null,
   "tps": 44.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-855kh14e-ctx1024",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: k3h953tu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2588.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 2,588 80.3 5.9 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/855kh14e",
   "source_url": "https://omlx.ai/benchmarks/performance/855kh14e",
   "tg_tokens": null,
   "tps": 80.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-855kh14e-ctx4096",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: k3h953tu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2772.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 2,772 79.9 6.3 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/855kh14e",
   "source_url": "https://omlx.ai/benchmarks/performance/855kh14e",
   "tg_tokens": null,
   "tps": 79.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-kk0tys51-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.8 GB at this context; duplicate traces of the same session: 02bwl08b, jgc2lcmk, lkxe30db, ntspj7u2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1149.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,149 63.9 100.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kk0tys51",
   "source_url": "https://omlx.ai/benchmarks/performance/kk0tys51",
   "tg_tokens": null,
   "tps": 63.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-kk0tys51-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: 02bwl08b, jgc2lcmk, lkxe30db, ntspj7u2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1392.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 1,392 75.9 101.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kk0tys51",
   "source_url": "https://omlx.ai/benchmarks/performance/kk0tys51",
   "tg_tokens": null,
   "tps": 75.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-kk0tys51-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.9 GB at this context; duplicate traces of the same session: 02bwl08b, jgc2lcmk, lkxe30db, ntspj7u2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1378.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 1,378 106.8 101.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kk0tys51",
   "source_url": "https://omlx.ai/benchmarks/performance/kk0tys51",
   "tg_tokens": null,
   "tps": 106.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-kk0tys51-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.1 GB at this context; duplicate traces of the same session: 02bwl08b, jgc2lcmk, lkxe30db, ntspj7u2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1367.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 1,367 76.4 103.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kk0tys51",
   "source_url": "https://omlx.ai/benchmarks/performance/kk0tys51",
   "tg_tokens": null,
   "tps": 76.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-kk0tys51-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 105.6 GB at this context; duplicate traces of the same session: 02bwl08b, jgc2lcmk, lkxe30db, ntspj7u2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1347.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 1,347 70.5 105.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kk0tys51",
   "source_url": "https://omlx.ai/benchmarks/performance/kk0tys51",
   "tg_tokens": null,
   "tps": 70.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-eg8eex01-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 69.9 GB; TTFT 9556 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 857.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k tokens 857.2 PP tok/s 60.8 TG tok/s 9556 TTFT (ms) 69.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eg8eex01",
   "source_url": "https://omlx.ai/benchmarks/performance/eg8eex01",
   "tg_tokens": null,
   "tps": 60.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-1ah13wq8-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 40.1 GB at this context; duplicate traces of the same session: 6dz3l3l3, s5kih4lc, ui8124kb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 252.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 252.2 44.6 40.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ah13wq8",
   "source_url": "https://omlx.ai/benchmarks/performance/1ah13wq8",
   "tg_tokens": null,
   "tps": 44.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-1ah13wq8-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.4 GB at this context; duplicate traces of the same session: 6dz3l3l3, s5kih4lc, ui8124kb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 332.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 332.4 33.1 43.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ah13wq8",
   "source_url": "https://omlx.ai/benchmarks/performance/1ah13wq8",
   "tg_tokens": null,
   "tps": 33.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-1ah13wq8-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.9 GB at this context; duplicate traces of the same session: 6dz3l3l3, s5kih4lc, ui8124kb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 334.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 334.4 43.0 43.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ah13wq8",
   "source_url": "https://omlx.ai/benchmarks/performance/1ah13wq8",
   "tg_tokens": null,
   "tps": 43.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-1ah13wq8-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 44.9 GB at this context; duplicate traces of the same session: 6dz3l3l3, s5kih4lc, ui8124kb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 318.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 318.7 46.8 44.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ah13wq8",
   "source_url": "https://omlx.ai/benchmarks/performance/1ah13wq8",
   "tg_tokens": null,
   "tps": 46.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-u245cvzj-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai build; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: p52db8iz; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2829.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 2,829 116.8 19.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u245cvzj",
   "source_url": "https://omlx.ai/benchmarks/performance/u245cvzj",
   "tg_tokens": null,
   "tps": 116.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-u245cvzj-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai build; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: p52db8iz; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3523.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 3,523 115.2 21.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u245cvzj",
   "source_url": "https://omlx.ai/benchmarks/performance/u245cvzj",
   "tg_tokens": null,
   "tps": 115.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-u245cvzj-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai build; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: p52db8iz; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3335.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 3,335 107.1 21.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u245cvzj",
   "source_url": "https://omlx.ai/benchmarks/performance/u245cvzj",
   "tg_tokens": null,
   "tps": 107.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-u245cvzj-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: p52db8iz; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3003.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 3,003 99.8 21.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u245cvzj",
   "source_url": "https://omlx.ai/benchmarks/performance/u245cvzj",
   "tg_tokens": null,
   "tps": 99.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-u245cvzj-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: p52db8iz; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2501.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 2,501 95.5 22.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u245cvzj",
   "source_url": "https://omlx.ai/benchmarks/performance/u245cvzj",
   "tg_tokens": null,
   "tps": 95.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-u245cvzj-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai build; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: p52db8iz; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1861.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 1,861 80.8 23.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u245cvzj",
   "source_url": "https://omlx.ai/benchmarks/performance/u245cvzj",
   "tg_tokens": null,
   "tps": 80.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-orr8afj9-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 71920jjj; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 339.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 16k 339.6 38.8 22.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/orr8afj9",
   "source_url": "https://omlx.ai/benchmarks/performance/orr8afj9",
   "tg_tokens": null,
   "tps": 38.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-orr8afj9-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai build; oMLX community benchmark trace; peak memory 26.9 GB at this context; duplicate traces of the same session: 71920jjj; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 155.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 64k 155.7 17.4 26.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/orr8afj9",
   "source_url": "https://omlx.ai/benchmarks/performance/orr8afj9",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-pvo0ckg6-ctx4096",
   "model": "Qwen2.5-Coder-3B",
   "notes": "oMLX community benchmark trace; peak memory 2.6 GB at this context; duplicate traces of the same session: v9jxy7gz; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 868.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-3B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 868.5 99.8 2.6 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pvo0ckg6",
   "source_url": "https://omlx.ai/benchmarks/performance/pvo0ckg6",
   "tg_tokens": null,
   "tps": 99.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-pvo0ckg6-ctx16384",
   "model": "Qwen2.5-Coder-3B",
   "notes": "oMLX community benchmark trace; peak memory 3.0 GB at this context; duplicate traces of the same session: v9jxy7gz; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 685.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-3B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 16k 685.1 79.4 3.0 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pvo0ckg6",
   "source_url": "https://omlx.ai/benchmarks/performance/pvo0ckg6",
   "tg_tokens": null,
   "tps": 79.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-drg13jlu-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 1wuj6vh4, 24cfq5qr, q5zxjz1v, rj5r3gs2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 917.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 917.4 50.5 21.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/drg13jlu",
   "source_url": "https://omlx.ai/benchmarks/performance/drg13jlu",
   "tg_tokens": null,
   "tps": 50.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-drg13jlu-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 25.2 GB at this context; duplicate traces of the same session: 1wuj6vh4, 24cfq5qr, q5zxjz1v, rj5r3gs2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1177.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,177 62.5 25.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/drg13jlu",
   "source_url": "https://omlx.ai/benchmarks/performance/drg13jlu",
   "tg_tokens": null,
   "tps": 62.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-drg13jlu-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: 1wuj6vh4, 24cfq5qr, q5zxjz1v, rj5r3gs2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1331.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,331 72.9 25.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/drg13jlu",
   "source_url": "https://omlx.ai/benchmarks/performance/drg13jlu",
   "tg_tokens": null,
   "tps": 72.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-drg13jlu-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 26.2 GB at this context; duplicate traces of the same session: 1wuj6vh4, 24cfq5qr, q5zxjz1v, rj5r3gs2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1359.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,359 77.8 26.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/drg13jlu",
   "source_url": "https://omlx.ai/benchmarks/performance/drg13jlu",
   "tg_tokens": null,
   "tps": 77.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-drg13jlu-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 27.3 GB at this context; duplicate traces of the same session: 1wuj6vh4, 24cfq5qr, q5zxjz1v, rj5r3gs2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1265.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 1,265 57.1 27.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/drg13jlu",
   "source_url": "https://omlx.ai/benchmarks/performance/drg13jlu",
   "tg_tokens": null,
   "tps": 57.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-a4wf9v2z-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 27.7 GB; TTFT 185143 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 354.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 96 GB \u00b7 4bit; 64k tokens 354.0 PP tok/s 74.7 TG tok/s 185143 TTFT (ms) 27.7 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a4wf9v2z",
   "source_url": "https://omlx.ai/benchmarks/performance/a4wf9v2z",
   "tg_tokens": null,
   "tps": 74.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-idsd5kma-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: nayszl20, ydl7kidp; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1135.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,135 76.5 100.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/idsd5kma",
   "source_url": "https://omlx.ai/benchmarks/performance/idsd5kma",
   "tg_tokens": null,
   "tps": 76.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-idsd5kma-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: nayszl20, ydl7kidp; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1392.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 1,392 67.0 101.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/idsd5kma",
   "source_url": "https://omlx.ai/benchmarks/performance/idsd5kma",
   "tg_tokens": null,
   "tps": 67.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-idsd5kma-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.6 GB at this context; duplicate traces of the same session: nayszl20, ydl7kidp; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1379.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 1,379 51.6 101.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/idsd5kma",
   "source_url": "https://omlx.ai/benchmarks/performance/idsd5kma",
   "tg_tokens": null,
   "tps": 51.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-idsd5kma-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.3 GB at this context; duplicate traces of the same session: nayszl20, ydl7kidp; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1368.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 1,368 51.9 102.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/idsd5kma",
   "source_url": "https://omlx.ai/benchmarks/performance/idsd5kma",
   "tg_tokens": null,
   "tps": 51.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-idsd5kma-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.3 GB at this context; duplicate traces of the same session: nayszl20, ydl7kidp; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1347.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 1,347 63.5 104.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/idsd5kma",
   "source_url": "https://omlx.ai/benchmarks/performance/idsd5kma",
   "tg_tokens": null,
   "tps": 63.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-35jauv3b-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: 5tt5r87q, bjmivpuz, bq8s2jcy, drhj01ia, k2lfwlip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 592.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 592.1 71.6 71.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/35jauv3b",
   "source_url": "https://omlx.ai/benchmarks/performance/35jauv3b",
   "tg_tokens": null,
   "tps": 71.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-35jauv3b-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: 5tt5r87q, bjmivpuz, bq8s2jcy, drhj01ia, k2lfwlip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 710.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 710.9 66.1 71.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/35jauv3b",
   "source_url": "https://omlx.ai/benchmarks/performance/35jauv3b",
   "tg_tokens": null,
   "tps": 66.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-35jauv3b-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: 5tt5r87q, bjmivpuz, bq8s2jcy, drhj01ia, k2lfwlip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 818.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 818.2 66.2 71.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/35jauv3b",
   "source_url": "https://omlx.ai/benchmarks/performance/35jauv3b",
   "tg_tokens": null,
   "tps": 66.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-35jauv3b-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 5tt5r87q, bjmivpuz, bq8s2jcy, drhj01ia, k2lfwlip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 814.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 814.4 76.2 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/35jauv3b",
   "source_url": "https://omlx.ai/benchmarks/performance/35jauv3b",
   "tg_tokens": null,
   "tps": 76.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-35jauv3b-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.0 GB at this context; duplicate traces of the same session: 5tt5r87q, bjmivpuz, bq8s2jcy, drhj01ia, k2lfwlip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 808.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 808.6 64.9 73.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/35jauv3b",
   "source_url": "https://omlx.ai/benchmarks/performance/35jauv3b",
   "tg_tokens": null,
   "tps": 64.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-35jauv3b-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 75.2 GB at this context; duplicate traces of the same session: 5tt5r87q, bjmivpuz, bq8s2jcy, drhj01ia, k2lfwlip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 802.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 802.4 65.6 75.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/35jauv3b",
   "source_url": "https://omlx.ai/benchmarks/performance/35jauv3b",
   "tg_tokens": null,
   "tps": 65.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-zsho3u21-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 40.1 GB at this context; duplicate traces of the same session: b3t7utn1, z8042x8m, zl9xqz85; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 245.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 245.1 29.5 40.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zsho3u21",
   "source_url": "https://omlx.ai/benchmarks/performance/zsho3u21",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-zsho3u21-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.4 GB at this context; duplicate traces of the same session: b3t7utn1, z8042x8m, zl9xqz85; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 329.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 329.3 36.9 43.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zsho3u21",
   "source_url": "https://omlx.ai/benchmarks/performance/zsho3u21",
   "tg_tokens": null,
   "tps": 36.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-zsho3u21-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.9 GB at this context; duplicate traces of the same session: b3t7utn1, z8042x8m, zl9xqz85; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 334.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 334.6 43.0 43.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zsho3u21",
   "source_url": "https://omlx.ai/benchmarks/performance/zsho3u21",
   "tg_tokens": null,
   "tps": 43.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-zsho3u21-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 44.9 GB at this context; duplicate traces of the same session: b3t7utn1, z8042x8m, zl9xqz85; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 318.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 318.8 43.1 44.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zsho3u21",
   "source_url": "https://omlx.ai/benchmarks/performance/zsho3u21",
   "tg_tokens": null,
   "tps": 43.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-cjfvsl2d-ctx1024",
   "model": "Qwen3-VL-8B-Instruct",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 8.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3108.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mlx-community--Qwen3-VL-8B-Instruct-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 3,108 57.0 8.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cjfvsl2d",
   "source_url": "https://omlx.ai/benchmarks/performance/cjfvsl2d",
   "tg_tokens": null,
   "tps": 57.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-cjfvsl2d-ctx131072",
   "model": "Qwen3-VL-8B-Instruct",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 815.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mlx-community--Qwen3-VL-8B-Instruct-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 815.1 18.5 27.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cjfvsl2d",
   "source_url": "https://omlx.ai/benchmarks/performance/cjfvsl2d",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-acpb8tz3-ctx1024",
   "model": "FuseO1-DeepSeekR1-Qwen2.5-Coder-7B-Preview",
   "notes": "oMLX community benchmark trace; peak memory 4.5 GB at this context; duplicate traces of the same session: k6w0peuk; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 734.1,
   "provenance": "community",
   "quant": "q4",
   "quote": "FuseO1-DeepSeekR1-Qwen2.5-Coder-7B-Preview-q4 on M5 (8c); M5 (8c) \u00b7 16 GB \u00b7 4bit; 1k 734.1 30.1 4.5 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/acpb8tz3",
   "source_url": "https://omlx.ai/benchmarks/performance/acpb8tz3",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-acpb8tz3-ctx4096",
   "model": "FuseO1-DeepSeekR1-Qwen2.5-Coder-7B-Preview",
   "notes": "oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: k6w0peuk; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 739.1,
   "provenance": "community",
   "quant": "q4",
   "quote": "FuseO1-DeepSeekR1-Qwen2.5-Coder-7B-Preview-q4 on M5 (8c); M5 (8c) \u00b7 16 GB \u00b7 4bit; 4k 739.1 28.5 4.9 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/acpb8tz3",
   "source_url": "https://omlx.ai/benchmarks/performance/acpb8tz3",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-wo6tlfsv-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 30.9 GB; TTFT 191912 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 341.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 96 GB \u00b7 4bit; 64k tokens 341.5 PP tok/s 72.7 TG tok/s 191912 TTFT (ms) 30.9 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wo6tlfsv",
   "source_url": "https://omlx.ai/benchmarks/performance/wo6tlfsv",
   "tg_tokens": null,
   "tps": 72.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-n0t0tse2-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 30.9 GB; TTFT 190892 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 343.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 96 GB \u00b7 4bit; 64k tokens 343.3 PP tok/s 72.7 TG tok/s 190892 TTFT (ms) 30.9 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n0t0tse2",
   "source_url": "https://omlx.ai/benchmarks/performance/n0t0tse2",
   "tg_tokens": null,
   "tps": 72.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0pq11dqt-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 30.9 GB; TTFT 155932 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 420.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 96 GB \u00b7 4bit; 64k tokens 420.3 PP tok/s 72.7 TG tok/s 155932 TTFT (ms) 30.9 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0pq11dqt",
   "source_url": "https://omlx.ai/benchmarks/performance/0pq11dqt",
   "tg_tokens": null,
   "tps": 72.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2vr1ic4m-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: ungxalpe; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1135.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,135 76.5 100.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2vr1ic4m",
   "source_url": "https://omlx.ai/benchmarks/performance/2vr1ic4m",
   "tg_tokens": null,
   "tps": 76.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2vr1ic4m-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: ungxalpe; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1392.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 1,392 67.0 101.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2vr1ic4m",
   "source_url": "https://omlx.ai/benchmarks/performance/2vr1ic4m",
   "tg_tokens": null,
   "tps": 67.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2vr1ic4m-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.6 GB at this context; duplicate traces of the same session: ungxalpe; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1379.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 1,379 51.6 101.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2vr1ic4m",
   "source_url": "https://omlx.ai/benchmarks/performance/2vr1ic4m",
   "tg_tokens": null,
   "tps": 51.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2vr1ic4m-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.3 GB at this context; duplicate traces of the same session: ungxalpe; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1368.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 1,368 51.9 102.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2vr1ic4m",
   "source_url": "https://omlx.ai/benchmarks/performance/2vr1ic4m",
   "tg_tokens": null,
   "tps": 51.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2vr1ic4m-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.3 GB at this context; duplicate traces of the same session: ungxalpe; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1347.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 1,347 63.5 104.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2vr1ic4m",
   "source_url": "https://omlx.ai/benchmarks/performance/2vr1ic4m",
   "tg_tokens": null,
   "tps": 63.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-z0j0lt1d-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.1 GB at this context; duplicate traces of the same session: eliedebd, ohmgjopg, sceautfe; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1173.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,173 79.2 99.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z0j0lt1d",
   "source_url": "https://omlx.ai/benchmarks/performance/z0j0lt1d",
   "tg_tokens": null,
   "tps": 79.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-z0j0lt1d-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.7 GB at this context; duplicate traces of the same session: eliedebd, ohmgjopg, sceautfe; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1446.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 1,446 75.5 99.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z0j0lt1d",
   "source_url": "https://omlx.ai/benchmarks/performance/z0j0lt1d",
   "tg_tokens": null,
   "tps": 75.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-z0j0lt1d-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.0 GB at this context; duplicate traces of the same session: eliedebd, ohmgjopg, sceautfe; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1435.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 1,435 75.3 100.0 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z0j0lt1d",
   "source_url": "https://omlx.ai/benchmarks/performance/z0j0lt1d",
   "tg_tokens": null,
   "tps": 75.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-z0j0lt1d-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: eliedebd, ohmgjopg, sceautfe; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1400.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 1,400 49.3 101.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z0j0lt1d",
   "source_url": "https://omlx.ai/benchmarks/performance/z0j0lt1d",
   "tg_tokens": null,
   "tps": 49.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-xu53xbc4-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 78.3 GB; TTFT 200515 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 653.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k tokens 653.7 PP tok/s 56.0 TG tok/s 200515 TTFT (ms) 78.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xu53xbc4",
   "source_url": "https://omlx.ai/benchmarks/performance/xu53xbc4",
   "tg_tokens": null,
   "tps": 56.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-kag06gd1-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.2 GB at this context; duplicate traces of the same session: 1iyftm9y, ofpv9c4b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 511.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 511.8 52.0 71.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kag06gd1",
   "source_url": "https://omlx.ai/benchmarks/performance/kag06gd1",
   "tg_tokens": null,
   "tps": 52.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-kag06gd1-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: 1iyftm9y, ofpv9c4b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 658.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 658.0 47.0 71.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kag06gd1",
   "source_url": "https://omlx.ai/benchmarks/performance/kag06gd1",
   "tg_tokens": null,
   "tps": 47.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-kag06gd1-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 75.0 GB at this context; duplicate traces of the same session: 1iyftm9y, ofpv9c4b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 802.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 802.0 60.2 75.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kag06gd1",
   "source_url": "https://omlx.ai/benchmarks/performance/kag06gd1",
   "tg_tokens": null,
   "tps": 60.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-1lfl3dog-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: a2rge7bz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 616.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 616.6 77.7 71.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1lfl3dog",
   "source_url": "https://omlx.ai/benchmarks/performance/1lfl3dog",
   "tg_tokens": null,
   "tps": 77.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-1lfl3dog-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.0 GB at this context; duplicate traces of the same session: a2rge7bz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 824.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 824.2 80.2 72.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1lfl3dog",
   "source_url": "https://omlx.ai/benchmarks/performance/1lfl3dog",
   "tg_tokens": null,
   "tps": 80.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-1lfl3dog-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 75.1 GB at this context; duplicate traces of the same session: a2rge7bz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 811.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 811.5 57.5 75.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1lfl3dog",
   "source_url": "https://omlx.ai/benchmarks/performance/1lfl3dog",
   "tg_tokens": null,
   "tps": 57.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-fwljvnoc-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: b3i13qhx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 590.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 590.3 71.4 71.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fwljvnoc",
   "source_url": "https://omlx.ai/benchmarks/performance/fwljvnoc",
   "tg_tokens": null,
   "tps": 71.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-fwljvnoc-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: b3i13qhx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 717.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 717.4 68.9 71.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fwljvnoc",
   "source_url": "https://omlx.ai/benchmarks/performance/fwljvnoc",
   "tg_tokens": null,
   "tps": 68.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-fwljvnoc-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.0 GB at this context; duplicate traces of the same session: b3i13qhx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 819.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 819.6 53.0 73.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fwljvnoc",
   "source_url": "https://omlx.ai/benchmarks/performance/fwljvnoc",
   "tg_tokens": null,
   "tps": 53.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-fwljvnoc-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 75.0 GB at this context; duplicate traces of the same session: b3i13qhx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 811.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 811.7 62.6 75.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fwljvnoc",
   "source_url": "https://omlx.ai/benchmarks/performance/fwljvnoc",
   "tg_tokens": null,
   "tps": 62.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-xg1bqouo-ctx1024",
   "model": "Qwen3-VL-8B-Instruct",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 8.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3108.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mlx-community--Qwen3-VL-8B-Instruct-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 3,108 57.0 8.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xg1bqouo",
   "source_url": "https://omlx.ai/benchmarks/performance/xg1bqouo",
   "tg_tokens": null,
   "tps": 57.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-xg1bqouo-ctx131072",
   "model": "Qwen3-VL-8B-Instruct",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 815.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mlx-community--Qwen3-VL-8B-Instruct-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 815.1 18.5 27.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xg1bqouo",
   "source_url": "https://omlx.ai/benchmarks/performance/xg1bqouo",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-niqzie0c-ctx1024",
   "model": "Qwen3-VL-8B-Instruct",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 8.9 GB at this context; duplicate traces of the same session: bdw2bhlp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2876.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mlx-community--Qwen3-VL-8B-Instruct-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 2,876 62.8 8.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/niqzie0c",
   "source_url": "https://omlx.ai/benchmarks/performance/niqzie0c",
   "tg_tokens": null,
   "tps": 62.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-niqzie0c-ctx4096",
   "model": "Qwen3-VL-8B-Instruct",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 9.3 GB at this context; duplicate traces of the same session: bdw2bhlp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3059.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mlx-community--Qwen3-VL-8B-Instruct-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 3,059 60.4 9.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/niqzie0c",
   "source_url": "https://omlx.ai/benchmarks/performance/niqzie0c",
   "tg_tokens": null,
   "tps": 60.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ln2wh2x4-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "AutomatosX AXQ build, MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: j24e1nx7, nro6nqyw, p1ehg19n, sc3j4izf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 913.2,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "AutomatosX--AX-Tiel-Coder-35B-A3B-MLX-AXQ-MXFP4-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 913.2 74.5 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ln2wh2x4",
   "source_url": "https://omlx.ai/benchmarks/performance/ln2wh2x4",
   "tg_tokens": null,
   "tps": 74.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ln2wh2x4-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "AutomatosX AXQ build, MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: j24e1nx7, nro6nqyw, p1ehg19n, sc3j4izf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1048.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "AutomatosX--AX-Tiel-Coder-35B-A3B-MLX-AXQ-MXFP4-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,048 71.8 21.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ln2wh2x4",
   "source_url": "https://omlx.ai/benchmarks/performance/ln2wh2x4",
   "tg_tokens": null,
   "tps": 71.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ln2wh2x4-ctx8192",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "AutomatosX AXQ build, MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: j24e1nx7, nro6nqyw, p1ehg19n, sc3j4izf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 991.2,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "AutomatosX--AX-Tiel-Coder-35B-A3B-MLX-AXQ-MXFP4-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 991.2 68.8 21.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ln2wh2x4",
   "source_url": "https://omlx.ai/benchmarks/performance/ln2wh2x4",
   "tg_tokens": null,
   "tps": 68.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ln2wh2x4-ctx16384",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "AutomatosX AXQ build, MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: j24e1nx7, nro6nqyw, p1ehg19n, sc3j4izf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 885.9,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "AutomatosX--AX-Tiel-Coder-35B-A3B-MLX-AXQ-MXFP4-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 885.9 64.3 21.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ln2wh2x4",
   "source_url": "https://omlx.ai/benchmarks/performance/ln2wh2x4",
   "tg_tokens": null,
   "tps": 64.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ln2wh2x4-ctx32768",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "AutomatosX AXQ build, MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: j24e1nx7, nro6nqyw, p1ehg19n, sc3j4izf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 689.1,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "AutomatosX--AX-Tiel-Coder-35B-A3B-MLX-AXQ-MXFP4-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 689.1 56.2 22.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ln2wh2x4",
   "source_url": "https://omlx.ai/benchmarks/performance/ln2wh2x4",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-1kdgm1uj-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "AutomatosX AXQ build, MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 9rutpc6h; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 918.2,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "AutomatosX--AX-Tiel-Coder-35B-A3B-MLX-AXQ-MXFP4-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 918.2 74.5 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1kdgm1uj",
   "source_url": "https://omlx.ai/benchmarks/performance/1kdgm1uj",
   "tg_tokens": null,
   "tps": 74.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-1kdgm1uj-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "AutomatosX AXQ build, MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 9rutpc6h; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1050.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "AutomatosX--AX-Tiel-Coder-35B-A3B-MLX-AXQ-MXFP4-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,050 72.2 21.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1kdgm1uj",
   "source_url": "https://omlx.ai/benchmarks/performance/1kdgm1uj",
   "tg_tokens": null,
   "tps": 72.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-lq97kk1r-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 30.9 GB; TTFT 155864 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 420.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 96 GB \u00b7 4bit; 64k tokens 420.5 PP tok/s 72.6 TG tok/s 155864 TTFT (ms) 30.9 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lq97kk1r",
   "source_url": "https://omlx.ai/benchmarks/performance/lq97kk1r",
   "tg_tokens": null,
   "tps": 72.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-5dng3b1x-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.9 GB; TTFT 2197 ms; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 466.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 1k tokens 466.1 PP tok/s 62.0 TG tok/s 2197 TTFT (ms) 70.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5dng3b1x",
   "source_url": "https://omlx.ai/benchmarks/performance/5dng3b1x",
   "tg_tokens": null,
   "tps": 62.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ulzi0t0f-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: g5s0tc4f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 590.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 590.3 71.4 71.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ulzi0t0f",
   "source_url": "https://omlx.ai/benchmarks/performance/ulzi0t0f",
   "tg_tokens": null,
   "tps": 71.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ulzi0t0f-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: g5s0tc4f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 717.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 717.4 68.9 71.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ulzi0t0f",
   "source_url": "https://omlx.ai/benchmarks/performance/ulzi0t0f",
   "tg_tokens": null,
   "tps": 68.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ulzi0t0f-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.0 GB at this context; duplicate traces of the same session: g5s0tc4f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 819.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 819.6 53.0 73.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ulzi0t0f",
   "source_url": "https://omlx.ai/benchmarks/performance/ulzi0t0f",
   "tg_tokens": null,
   "tps": 53.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ulzi0t0f-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 75.0 GB at this context; duplicate traces of the same session: g5s0tc4f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 811.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 811.7 62.6 75.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ulzi0t0f",
   "source_url": "https://omlx.ai/benchmarks/performance/ulzi0t0f",
   "tg_tokens": null,
   "tps": 62.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-onmjx32y-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 616.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 616.6 77.7 71.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/onmjx32y",
   "source_url": "https://omlx.ai/benchmarks/performance/onmjx32y",
   "tg_tokens": null,
   "tps": 77.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-onmjx32y-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 824.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 824.2 80.2 72.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/onmjx32y",
   "source_url": "https://omlx.ai/benchmarks/performance/onmjx32y",
   "tg_tokens": null,
   "tps": 80.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-onmjx32y-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 75.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 811.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 811.5 57.5 75.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/onmjx32y",
   "source_url": "https://omlx.ai/benchmarks/performance/onmjx32y",
   "tg_tokens": null,
   "tps": 57.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-4bla2qnb-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.3 GB at this context; duplicate traces of the same session: ojljebfj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 560.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 560.3 53.6 69.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4bla2qnb",
   "source_url": "https://omlx.ai/benchmarks/performance/4bla2qnb",
   "tg_tokens": null,
   "tps": 53.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-4bla2qnb-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.9 GB at this context; duplicate traces of the same session: ojljebfj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 858.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 858.8 57.5 69.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4bla2qnb",
   "source_url": "https://omlx.ai/benchmarks/performance/4bla2qnb",
   "tg_tokens": null,
   "tps": 57.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-bqh4myg2-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: a2nvkurk, z5ou9k6v; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 556.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 556.4 35.0 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bqh4myg2",
   "source_url": "https://omlx.ai/benchmarks/performance/bqh4myg2",
   "tg_tokens": null,
   "tps": 35.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-bqh4myg2-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: a2nvkurk, z5ou9k6v; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 605.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 605.0 50.2 71.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bqh4myg2",
   "source_url": "https://omlx.ai/benchmarks/performance/bqh4myg2",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-bqh4myg2-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: a2nvkurk, z5ou9k6v; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 833.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 833.0 64.0 71.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bqh4myg2",
   "source_url": "https://omlx.ai/benchmarks/performance/bqh4myg2",
   "tg_tokens": null,
   "tps": 64.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-xvavpghw-ctx1024",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2397.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 2,397 103.0 26.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xvavpghw",
   "source_url": "https://omlx.ai/benchmarks/performance/xvavpghw",
   "tg_tokens": null,
   "tps": 103.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-xvavpghw-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 27.2 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2369.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 2,369 53.5 27.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xvavpghw",
   "source_url": "https://omlx.ai/benchmarks/performance/xvavpghw",
   "tg_tokens": null,
   "tps": 53.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-xvavpghw-ctx8192",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2260.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 2,260 60.7 27.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xvavpghw",
   "source_url": "https://omlx.ai/benchmarks/performance/xvavpghw",
   "tg_tokens": null,
   "tps": 60.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-6xx1knjw-ctx1024",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2411.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 2,411 72.4 26.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6xx1knjw",
   "source_url": "https://omlx.ai/benchmarks/performance/6xx1knjw",
   "tg_tokens": null,
   "tps": 72.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-6xx1knjw-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 27.2 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2373.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 2,373 110.4 27.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6xx1knjw",
   "source_url": "https://omlx.ai/benchmarks/performance/6xx1knjw",
   "tg_tokens": null,
   "tps": 110.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-6xx1knjw-ctx8192",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2266.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 2,266 72.5 27.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6xx1knjw",
   "source_url": "https://omlx.ai/benchmarks/performance/6xx1knjw",
   "tg_tokens": null,
   "tps": 72.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-jams61ed-ctx1024",
   "model": "gemma-4-26B-A4B-it-heretic-msq",
   "notes": "oMLX community benchmark trace; peak memory 9.2 GB at this context; duplicate traces of the same session: 9dt3m670; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1865.0,
   "provenance": "community",
   "quant": "2.6bit",
   "quote": "gemma-4-26B-A4B-it-heretic-msq-2.6bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 1,865 70.2 9.2 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jams61ed",
   "source_url": "https://omlx.ai/benchmarks/performance/jams61ed",
   "tg_tokens": null,
   "tps": 70.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-jams61ed-ctx4096",
   "model": "gemma-4-26B-A4B-it-heretic-msq",
   "notes": "oMLX community benchmark trace; peak memory 9.5 GB at this context; duplicate traces of the same session: 9dt3m670; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1922.0,
   "provenance": "community",
   "quant": "2.6bit",
   "quote": "gemma-4-26B-A4B-it-heretic-msq-2.6bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 1,922 68.6 9.5 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jams61ed",
   "source_url": "https://omlx.ai/benchmarks/performance/jams61ed",
   "tg_tokens": null,
   "tps": 68.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ookk6897-ctx65536",
   "model": "Occamy-1.0",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 23.2 GB; TTFT 35029 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1871.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "occamy-1.0-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 96 GB \u00b7 4bit; 64k tokens 1,871 PP tok/s 85.4 TG tok/s 35029 TTFT (ms) 23.2 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ookk6897",
   "source_url": "https://omlx.ai/benchmarks/performance/ookk6897",
   "tg_tokens": null,
   "tps": 85.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5l4h25ll-ctx1024",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 177.7 GB at this context; duplicate traces of the same session: 3sq2tze9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 515.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 515.0 48.6 177.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5l4h25ll",
   "source_url": "https://omlx.ai/benchmarks/performance/5l4h25ll",
   "tg_tokens": null,
   "tps": 48.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5l4h25ll-ctx4096",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 178.8 GB at this context; duplicate traces of the same session: 3sq2tze9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 587.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 587.0 51.8 178.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5l4h25ll",
   "source_url": "https://omlx.ai/benchmarks/performance/5l4h25ll",
   "tg_tokens": null,
   "tps": 51.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-a960wr0z-ctx1024",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 177.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 511.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 511.7 50.2 177.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a960wr0z",
   "source_url": "https://omlx.ai/benchmarks/performance/a960wr0z",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-a960wr0z-ctx4096",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 178.8 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 589.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 589.0 45.7 178.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a960wr0z",
   "source_url": "https://omlx.ai/benchmarks/performance/a960wr0z",
   "tg_tokens": null,
   "tps": 45.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qsxqry3k-ctx1024",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 177.7 GB at this context; duplicate traces of the same session: ah25yyjh; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 514.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 514.9 44.6 177.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qsxqry3k",
   "source_url": "https://omlx.ai/benchmarks/performance/qsxqry3k",
   "tg_tokens": null,
   "tps": 44.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qsxqry3k-ctx4096",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 178.8 GB at this context; duplicate traces of the same session: ah25yyjh; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 586.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 586.2 44.8 178.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qsxqry3k",
   "source_url": "https://omlx.ai/benchmarks/performance/qsxqry3k",
   "tg_tokens": null,
   "tps": 44.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-xpm14uyb-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 27.0 GB at this context; duplicate traces of the same session: 3jstpqpp, 3n4viujt; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2365.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 2,365 54.9 27.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xpm14uyb",
   "source_url": "https://omlx.ai/benchmarks/performance/xpm14uyb",
   "tg_tokens": null,
   "tps": 54.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-xpm14uyb-ctx8192",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: 3jstpqpp, 3n4viujt; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2263.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 2,263 64.9 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xpm14uyb",
   "source_url": "https://omlx.ai/benchmarks/performance/xpm14uyb",
   "tg_tokens": null,
   "tps": 64.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-xpm14uyb-ctx16384",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 28.8 GB at this context; duplicate traces of the same session: 3jstpqpp, 3n4viujt; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2088.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 16k 2,088 71.4 28.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xpm14uyb",
   "source_url": "https://omlx.ai/benchmarks/performance/xpm14uyb",
   "tg_tokens": null,
   "tps": 71.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-lnyozvkp-ctx1024",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: h5ml1nqq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2411.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 2,411 72.4 26.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lnyozvkp",
   "source_url": "https://omlx.ai/benchmarks/performance/lnyozvkp",
   "tg_tokens": null,
   "tps": 72.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-lnyozvkp-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 27.2 GB at this context; duplicate traces of the same session: h5ml1nqq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2373.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 2,373 110.4 27.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lnyozvkp",
   "source_url": "https://omlx.ai/benchmarks/performance/lnyozvkp",
   "tg_tokens": null,
   "tps": 110.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-lnyozvkp-ctx8192",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: h5ml1nqq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2266.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 2,266 72.5 27.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lnyozvkp",
   "source_url": "https://omlx.ai/benchmarks/performance/lnyozvkp",
   "tg_tokens": null,
   "tps": 72.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-6y6cc39l-ctx1024",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: ogjfa45o; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2397.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 2,397 103.0 26.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6y6cc39l",
   "source_url": "https://omlx.ai/benchmarks/performance/6y6cc39l",
   "tg_tokens": null,
   "tps": 103.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-6y6cc39l-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 27.2 GB at this context; duplicate traces of the same session: ogjfa45o; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2369.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 2,369 53.5 27.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6y6cc39l",
   "source_url": "https://omlx.ai/benchmarks/performance/6y6cc39l",
   "tg_tokens": null,
   "tps": 53.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-6y6cc39l-ctx8192",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: ogjfa45o; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2260.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 2,260 60.7 27.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6y6cc39l",
   "source_url": "https://omlx.ai/benchmarks/performance/6y6cc39l",
   "tg_tokens": null,
   "tps": 60.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-p249sjkt-ctx1024",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: 16t7dg00, 833xzfjd; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2267.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 2,267 56.2 25.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p249sjkt",
   "source_url": "https://omlx.ai/benchmarks/performance/p249sjkt",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-p249sjkt-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: 16t7dg00, 833xzfjd; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2373.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 2,373 54.6 26.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p249sjkt",
   "source_url": "https://omlx.ai/benchmarks/performance/p249sjkt",
   "tg_tokens": null,
   "tps": 54.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-p249sjkt-ctx8192",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: 16t7dg00, 833xzfjd; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2292.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 2,292 52.3 26.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p249sjkt",
   "source_url": "https://omlx.ai/benchmarks/performance/p249sjkt",
   "tg_tokens": null,
   "tps": 52.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-4vow9pdp-ctx1024",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: 3v67u485, n81lujqf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2218.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 2,218 56.3 25.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4vow9pdp",
   "source_url": "https://omlx.ai/benchmarks/performance/4vow9pdp",
   "tg_tokens": null,
   "tps": 56.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-4vow9pdp-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: 3v67u485, n81lujqf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2337.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 2,337 54.5 26.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4vow9pdp",
   "source_url": "https://omlx.ai/benchmarks/performance/4vow9pdp",
   "tg_tokens": null,
   "tps": 54.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-4vow9pdp-ctx8192",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: 3v67u485, n81lujqf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2281.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 2,281 52.4 26.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4vow9pdp",
   "source_url": "https://omlx.ai/benchmarks/performance/4vow9pdp",
   "tg_tokens": null,
   "tps": 52.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-g52fwknh-ctx1024",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: 3stj270d, wiqxb5kt; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2406.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 2,406 63.9 26.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g52fwknh",
   "source_url": "https://omlx.ai/benchmarks/performance/g52fwknh",
   "tg_tokens": null,
   "tps": 63.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-g52fwknh-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: 3stj270d, wiqxb5kt; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2366.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 2,366 47.9 27.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g52fwknh",
   "source_url": "https://omlx.ai/benchmarks/performance/g52fwknh",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-g52fwknh-ctx8192",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 3stj270d, wiqxb5kt; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2261.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 2,261 58.3 28.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g52fwknh",
   "source_url": "https://omlx.ai/benchmarks/performance/g52fwknh",
   "tg_tokens": null,
   "tps": 58.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-rksom3mb-ctx1024",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2418.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 2,418 77.8 26.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rksom3mb",
   "source_url": "https://omlx.ai/benchmarks/performance/rksom3mb",
   "tg_tokens": null,
   "tps": 77.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-rksom3mb-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2366.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 2,366 53.9 27.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rksom3mb",
   "source_url": "https://omlx.ai/benchmarks/performance/rksom3mb",
   "tg_tokens": null,
   "tps": 53.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-rksom3mb-ctx8192",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2260.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 2,260 63.4 28.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rksom3mb",
   "source_url": "https://omlx.ai/benchmarks/performance/rksom3mb",
   "tg_tokens": null,
   "tps": 63.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-o9cerikq-ctx1024",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 177.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 511.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 511.7 50.2 177.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9cerikq",
   "source_url": "https://omlx.ai/benchmarks/performance/o9cerikq",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-o9cerikq-ctx4096",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 178.8 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 589.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 589.0 45.7 178.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9cerikq",
   "source_url": "https://omlx.ai/benchmarks/performance/o9cerikq",
   "tg_tokens": null,
   "tps": 45.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-eg0jie2w-ctx1024",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 177.7 GB at this context; duplicate traces of the same session: mao7pj9w; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 527.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 527.0 40.1 177.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eg0jie2w",
   "source_url": "https://omlx.ai/benchmarks/performance/eg0jie2w",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-eg0jie2w-ctx4096",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 178.8 GB at this context; duplicate traces of the same session: mao7pj9w; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 605.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 605.4 38.2 178.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eg0jie2w",
   "source_url": "https://omlx.ai/benchmarks/performance/eg0jie2w",
   "tg_tokens": null,
   "tps": 38.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-i0533yol-ctx1024",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 177.7 GB at this context; duplicate traces of the same session: inql7lpv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 519.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 519.7 39.1 177.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i0533yol",
   "source_url": "https://omlx.ai/benchmarks/performance/i0533yol",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-i0533yol-ctx4096",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 178.8 GB at this context; duplicate traces of the same session: inql7lpv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 588.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 588.1 43.8 178.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i0533yol",
   "source_url": "https://omlx.ai/benchmarks/performance/i0533yol",
   "tg_tokens": null,
   "tps": 43.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hd23e4zy-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.7 GB at this context; duplicate traces of the same session: 1h90a3u4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 532.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 532.0 36.0 173.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hd23e4zy",
   "source_url": "https://omlx.ai/benchmarks/performance/hd23e4zy",
   "tg_tokens": null,
   "tps": 36.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hd23e4zy-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.8 GB at this context; duplicate traces of the same session: 1h90a3u4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 604.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 604.5 38.5 174.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hd23e4zy",
   "source_url": "https://omlx.ai/benchmarks/performance/hd23e4zy",
   "tg_tokens": null,
   "tps": 38.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hd23e4zy-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.1 GB at this context; duplicate traces of the same session: 1h90a3u4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 559.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 559.7 38.7 175.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hd23e4zy",
   "source_url": "https://omlx.ai/benchmarks/performance/hd23e4zy",
   "tg_tokens": null,
   "tps": 38.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hd23e4zy-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.2 GB at this context; duplicate traces of the same session: 1h90a3u4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 548.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 548.2 38.3 175.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hd23e4zy",
   "source_url": "https://omlx.ai/benchmarks/performance/hd23e4zy",
   "tg_tokens": null,
   "tps": 38.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hd23e4zy-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.3 GB at this context; duplicate traces of the same session: 1h90a3u4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 539.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 539.8 38.1 175.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hd23e4zy",
   "source_url": "https://omlx.ai/benchmarks/performance/hd23e4zy",
   "tg_tokens": null,
   "tps": 38.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hd23e4zy-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.6 GB at this context; duplicate traces of the same session: 1h90a3u4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 530.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 530.3 37.0 175.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hd23e4zy",
   "source_url": "https://omlx.ai/benchmarks/performance/hd23e4zy",
   "tg_tokens": null,
   "tps": 37.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hd23e4zy-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 176.4 GB at this context; duplicate traces of the same session: 1h90a3u4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 508.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 508.3 35.5 176.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hd23e4zy",
   "source_url": "https://omlx.ai/benchmarks/performance/hd23e4zy",
   "tg_tokens": null,
   "tps": 35.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-sknwo3w6-ctx1024",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: oqokuzeo, xf37qcsb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2397.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 2,397 62.2 26.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sknwo3w6",
   "source_url": "https://omlx.ai/benchmarks/performance/sknwo3w6",
   "tg_tokens": null,
   "tps": 62.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-sknwo3w6-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: oqokuzeo, xf37qcsb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2364.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 2,364 61.7 27.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sknwo3w6",
   "source_url": "https://omlx.ai/benchmarks/performance/sknwo3w6",
   "tg_tokens": null,
   "tps": 61.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-sknwo3w6-ctx8192",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 28.4 GB at this context; duplicate traces of the same session: oqokuzeo, xf37qcsb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2258.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 2,258 47.2 28.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sknwo3w6",
   "source_url": "https://omlx.ai/benchmarks/performance/sknwo3w6",
   "tg_tokens": null,
   "tps": 47.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ce54fgf2-ctx1024",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: rx4z99xs; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2418.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 2,418 77.8 26.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ce54fgf2",
   "source_url": "https://omlx.ai/benchmarks/performance/ce54fgf2",
   "tg_tokens": null,
   "tps": 77.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ce54fgf2-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: rx4z99xs; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2366.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 2,366 53.9 27.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ce54fgf2",
   "source_url": "https://omlx.ai/benchmarks/performance/ce54fgf2",
   "tg_tokens": null,
   "tps": 53.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ce54fgf2-ctx8192",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: rx4z99xs; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2260.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 2,260 63.4 28.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ce54fgf2",
   "source_url": "https://omlx.ai/benchmarks/performance/ce54fgf2",
   "tg_tokens": null,
   "tps": 63.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3e0sycwc-ctx1024",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: fxwn62jy, tui4c6xm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2306.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 2,306 60.6 27.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3e0sycwc",
   "source_url": "https://omlx.ai/benchmarks/performance/3e0sycwc",
   "tg_tokens": null,
   "tps": 60.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3e0sycwc-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: fxwn62jy, tui4c6xm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2353.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 2,353 45.0 28.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3e0sycwc",
   "source_url": "https://omlx.ai/benchmarks/performance/3e0sycwc",
   "tg_tokens": null,
   "tps": 45.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3e0sycwc-ctx8192",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 28.2 GB at this context; duplicate traces of the same session: fxwn62jy, tui4c6xm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2280.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 2,280 51.7 28.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3e0sycwc",
   "source_url": "https://omlx.ai/benchmarks/performance/3e0sycwc",
   "tg_tokens": null,
   "tps": 51.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-jfbbm77o-ctx1024",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: 7s2g9dwn, dtv92015; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2285.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 2,285 61.9 27.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jfbbm77o",
   "source_url": "https://omlx.ai/benchmarks/performance/jfbbm77o",
   "tg_tokens": null,
   "tps": 61.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-jfbbm77o-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: 7s2g9dwn, dtv92015; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2375.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 2,375 45.9 28.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jfbbm77o",
   "source_url": "https://omlx.ai/benchmarks/performance/jfbbm77o",
   "tg_tokens": null,
   "tps": 45.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-jfbbm77o-ctx8192",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 28.2 GB at this context; duplicate traces of the same session: 7s2g9dwn, dtv92015; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2295.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 2,295 53.6 28.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jfbbm77o",
   "source_url": "https://omlx.ai/benchmarks/performance/jfbbm77o",
   "tg_tokens": null,
   "tps": 53.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-zgwfwp5i-ctx1024",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: 5dz8dwf3, gvptm0jg; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2403.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 2,403 94.1 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zgwfwp5i",
   "source_url": "https://omlx.ai/benchmarks/performance/zgwfwp5i",
   "tg_tokens": null,
   "tps": 94.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-zgwfwp5i-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: 5dz8dwf3, gvptm0jg; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2367.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 2,367 53.2 27.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zgwfwp5i",
   "source_url": "https://omlx.ai/benchmarks/performance/zgwfwp5i",
   "tg_tokens": null,
   "tps": 53.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-zgwfwp5i-ctx8192",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 28.4 GB at this context; duplicate traces of the same session: 5dz8dwf3, gvptm0jg; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2260.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 2,260 58.1 28.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zgwfwp5i",
   "source_url": "https://omlx.ai/benchmarks/performance/zgwfwp5i",
   "tg_tokens": null,
   "tps": 58.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-tc9qw9qb-ctx1024",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2403.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 2,403 96.1 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tc9qw9qb",
   "source_url": "https://omlx.ai/benchmarks/performance/tc9qw9qb",
   "tg_tokens": null,
   "tps": 96.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-tc9qw9qb-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2364.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 2,364 88.1 27.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tc9qw9qb",
   "source_url": "https://omlx.ai/benchmarks/performance/tc9qw9qb",
   "tg_tokens": null,
   "tps": 88.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-tc9qw9qb-ctx8192",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 28.4 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2264.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 2,264 65.7 28.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tc9qw9qb",
   "source_url": "https://omlx.ai/benchmarks/performance/tc9qw9qb",
   "tg_tokens": null,
   "tps": 65.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qwji1z9v-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.7 GB at this context; duplicate traces of the same session: au8elnph, jybqq0ul, mke80sbx, rh7lo6f2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 532.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 532.0 36.0 173.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qwji1z9v",
   "source_url": "https://omlx.ai/benchmarks/performance/qwji1z9v",
   "tg_tokens": null,
   "tps": 36.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qwji1z9v-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.8 GB at this context; duplicate traces of the same session: au8elnph, jybqq0ul, mke80sbx, rh7lo6f2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 604.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 604.5 38.5 174.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qwji1z9v",
   "source_url": "https://omlx.ai/benchmarks/performance/qwji1z9v",
   "tg_tokens": null,
   "tps": 38.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qwji1z9v-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.1 GB at this context; duplicate traces of the same session: au8elnph, jybqq0ul, mke80sbx, rh7lo6f2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 559.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 559.7 38.7 175.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qwji1z9v",
   "source_url": "https://omlx.ai/benchmarks/performance/qwji1z9v",
   "tg_tokens": null,
   "tps": 38.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qwji1z9v-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.2 GB at this context; duplicate traces of the same session: au8elnph, jybqq0ul, mke80sbx, rh7lo6f2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 548.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 548.2 38.3 175.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qwji1z9v",
   "source_url": "https://omlx.ai/benchmarks/performance/qwji1z9v",
   "tg_tokens": null,
   "tps": 38.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qwji1z9v-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.3 GB at this context; duplicate traces of the same session: au8elnph, jybqq0ul, mke80sbx, rh7lo6f2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 539.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 539.8 38.1 175.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qwji1z9v",
   "source_url": "https://omlx.ai/benchmarks/performance/qwji1z9v",
   "tg_tokens": null,
   "tps": 38.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qwji1z9v-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.6 GB at this context; duplicate traces of the same session: au8elnph, jybqq0ul, mke80sbx, rh7lo6f2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 530.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 530.3 37.0 175.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qwji1z9v",
   "source_url": "https://omlx.ai/benchmarks/performance/qwji1z9v",
   "tg_tokens": null,
   "tps": 37.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qwji1z9v-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 176.4 GB at this context; duplicate traces of the same session: au8elnph, jybqq0ul, mke80sbx, rh7lo6f2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 508.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 508.3 35.5 176.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qwji1z9v",
   "source_url": "https://omlx.ai/benchmarks/performance/qwji1z9v",
   "tg_tokens": null,
   "tps": 35.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-n2haksyd-ctx1024",
   "model": "Qwen3 8B",
   "notes": "oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 845.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-8B-4bit on M5 (10c); M5 (10c) \u00b7 16 GB \u00b7 4bit; 1k 845.3 27.1 4.9 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n2haksyd",
   "source_url": "https://omlx.ai/benchmarks/performance/n2haksyd",
   "tg_tokens": null,
   "tps": 27.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-n2haksyd-ctx4096",
   "model": "Qwen3 8B",
   "notes": "oMLX community benchmark trace; peak memory 5.4 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 847.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-8B-4bit on M5 (10c); M5 (10c) \u00b7 16 GB \u00b7 4bit; 4k 847.5 24.7 5.4 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n2haksyd",
   "source_url": "https://omlx.ai/benchmarks/performance/n2haksyd",
   "tg_tokens": null,
   "tps": 24.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-sxfsev1i-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: v7xhuv0s; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1847.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,847 44.1 26.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sxfsev1i",
   "source_url": "https://omlx.ai/benchmarks/performance/sxfsev1i",
   "tg_tokens": null,
   "tps": 44.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-sxfsev1i-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace; peak memory 28.2 GB at this context; duplicate traces of the same session: v7xhuv0s; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1534.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,534 21.9 28.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sxfsev1i",
   "source_url": "https://omlx.ai/benchmarks/performance/sxfsev1i",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-yojhwjt6-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace, single run at 32k context; peak memory 26.0 GB; TTFT 17541 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1868.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k tokens 1,868 PP tok/s 49.3 TG tok/s 17541 TTFT (ms) 26.0 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yojhwjt6",
   "source_url": "https://omlx.ai/benchmarks/performance/yojhwjt6",
   "tg_tokens": null,
   "tps": 49.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wsen8jua-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "csolares2 build; MTP; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: 2l9hdx3d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 833.8,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "csolares2--Qwen3.8-27B-oQ8-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 833.8 30.8 34.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wsen8jua",
   "source_url": "https://omlx.ai/benchmarks/performance/wsen8jua",
   "tg_tokens": null,
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wsen8jua-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "csolares2 build; MTP; oMLX community benchmark trace; peak memory 35.7 GB at this context; duplicate traces of the same session: 2l9hdx3d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 592.2,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "csolares2--Qwen3.8-27B-oQ8-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 592.2 46.2 35.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wsen8jua",
   "source_url": "https://omlx.ai/benchmarks/performance/wsen8jua",
   "tg_tokens": null,
   "tps": 46.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-1ws0e0s1-ctx1024",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: zrx0qvgn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2403.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 2,403 96.1 26.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ws0e0s1",
   "source_url": "https://omlx.ai/benchmarks/performance/1ws0e0s1",
   "tg_tokens": null,
   "tps": 96.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1ws0e0s1-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: zrx0qvgn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2364.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 2,364 88.1 27.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ws0e0s1",
   "source_url": "https://omlx.ai/benchmarks/performance/1ws0e0s1",
   "tg_tokens": null,
   "tps": 88.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-1ws0e0s1-ctx8192",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 28.4 GB at this context; duplicate traces of the same session: zrx0qvgn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2264.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 2,264 65.7 28.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ws0e0s1",
   "source_url": "https://omlx.ai/benchmarks/performance/1ws0e0s1",
   "tg_tokens": null,
   "tps": 65.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-kt02bkwz-ctx1024",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: ebot2xqw, mk5abbt3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2392.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 2,392 63.5 26.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kt02bkwz",
   "source_url": "https://omlx.ai/benchmarks/performance/kt02bkwz",
   "tg_tokens": null,
   "tps": 63.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-kt02bkwz-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: ebot2xqw, mk5abbt3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2357.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 2,357 46.0 27.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kt02bkwz",
   "source_url": "https://omlx.ai/benchmarks/performance/kt02bkwz",
   "tg_tokens": null,
   "tps": 46.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-kt02bkwz-ctx8192",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: ebot2xqw, mk5abbt3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2257.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 2,257 62.2 28.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kt02bkwz",
   "source_url": "https://omlx.ai/benchmarks/performance/kt02bkwz",
   "tg_tokens": null,
   "tps": 62.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-tz1o3m42-ctx1024",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: e333h3hq, uku40dn3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2253.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 2,253 56.2 25.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tz1o3m42",
   "source_url": "https://omlx.ai/benchmarks/performance/tz1o3m42",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-tz1o3m42-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: e333h3hq, uku40dn3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2363.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 2,363 54.1 26.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tz1o3m42",
   "source_url": "https://omlx.ai/benchmarks/performance/tz1o3m42",
   "tg_tokens": null,
   "tps": 54.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-tz1o3m42-ctx8192",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: e333h3hq, uku40dn3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2288.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 2,288 52.4 26.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tz1o3m42",
   "source_url": "https://omlx.ai/benchmarks/performance/tz1o3m42",
   "tg_tokens": null,
   "tps": 52.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-90315gxy-ctx1024",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2254.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 2,254 56.2 25.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/90315gxy",
   "source_url": "https://omlx.ai/benchmarks/performance/90315gxy",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-90315gxy-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2355.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 2,355 54.5 26.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/90315gxy",
   "source_url": "https://omlx.ai/benchmarks/performance/90315gxy",
   "tg_tokens": null,
   "tps": 54.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-bk9a9mkl-ctx1024",
   "model": "gemma-4-26B-A4B-it-heretic-msq",
   "notes": "oMLX community benchmark trace; peak memory 9.2 GB at this context; duplicate traces of the same session: vorea6gk; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1927.0,
   "provenance": "community",
   "quant": "2.6bit",
   "quote": "gemma-4-26B-A4B-it-heretic-msq-2.6bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 1,927 71.3 9.2 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bk9a9mkl",
   "source_url": "https://omlx.ai/benchmarks/performance/bk9a9mkl",
   "tg_tokens": null,
   "tps": 71.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-bk9a9mkl-ctx4096",
   "model": "gemma-4-26B-A4B-it-heretic-msq",
   "notes": "oMLX community benchmark trace; peak memory 9.5 GB at this context; duplicate traces of the same session: vorea6gk; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1904.0,
   "provenance": "community",
   "quant": "2.6bit",
   "quote": "gemma-4-26B-A4B-it-heretic-msq-2.6bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 1,904 69.0 9.5 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bk9a9mkl",
   "source_url": "https://omlx.ai/benchmarks/performance/bk9a9mkl",
   "tg_tokens": null,
   "tps": 69.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-fwj100g5-ctx4096",
   "model": "GLM-4.5-Air",
   "notes": "oMLX community benchmark trace; peak memory 58.0 GB at this context; duplicate traces of the same session: z912ch9v; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1739.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-4.5-Air-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 4k 1,739 64.8 58.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fwj100g5",
   "source_url": "https://omlx.ai/benchmarks/performance/fwj100g5",
   "tg_tokens": null,
   "tps": 64.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-fwj100g5-ctx16384",
   "model": "GLM-4.5-Air",
   "notes": "oMLX community benchmark trace; peak memory 59.9 GB at this context; duplicate traces of the same session: z912ch9v; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1728.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-4.5-Air-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 16k 1,728 48.7 59.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fwj100g5",
   "source_url": "https://omlx.ai/benchmarks/performance/fwj100g5",
   "tg_tokens": null,
   "tps": 48.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-jgicq7x2-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 150.8 GB at this context; duplicate traces of the same session: 252ojpp2, 2wvatp1h, 3zejkdwd, aded017j, f4i1zhqm, ibem65c4, ouf18ef2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 372.3,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "GLM-5.3-Flash-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 1k 372.3 42.1 150.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jgicq7x2",
   "source_url": "https://omlx.ai/benchmarks/performance/jgicq7x2",
   "tg_tokens": null,
   "tps": 42.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-jgicq7x2-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 151.8 GB at this context; duplicate traces of the same session: 252ojpp2, 2wvatp1h, 3zejkdwd, aded017j, f4i1zhqm, ibem65c4, ouf18ef2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 442.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "GLM-5.3-Flash-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 4k 442.0 37.3 151.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jgicq7x2",
   "source_url": "https://omlx.ai/benchmarks/performance/jgicq7x2",
   "tg_tokens": null,
   "tps": 37.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-jgicq7x2-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 152.1 GB at this context; duplicate traces of the same session: 252ojpp2, 2wvatp1h, 3zejkdwd, aded017j, f4i1zhqm, ibem65c4, ouf18ef2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 415.3,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "GLM-5.3-Flash-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 8k 415.3 43.8 152.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jgicq7x2",
   "source_url": "https://omlx.ai/benchmarks/performance/jgicq7x2",
   "tg_tokens": null,
   "tps": 43.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-jgicq7x2-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 152.2 GB at this context; duplicate traces of the same session: 252ojpp2, 2wvatp1h, 3zejkdwd, aded017j, f4i1zhqm, ibem65c4, ouf18ef2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 407.6,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "GLM-5.3-Flash-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 16k 407.6 51.7 152.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jgicq7x2",
   "source_url": "https://omlx.ai/benchmarks/performance/jgicq7x2",
   "tg_tokens": null,
   "tps": 51.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jgicq7x2-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 152.4 GB at this context; duplicate traces of the same session: 252ojpp2, 2wvatp1h, 3zejkdwd, aded017j, f4i1zhqm, ibem65c4, ouf18ef2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 402.3,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "GLM-5.3-Flash-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 32k 402.3 48.5 152.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jgicq7x2",
   "source_url": "https://omlx.ai/benchmarks/performance/jgicq7x2",
   "tg_tokens": null,
   "tps": 48.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jgicq7x2-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 153.1 GB at this context; duplicate traces of the same session: 252ojpp2, 2wvatp1h, 3zejkdwd, aded017j, f4i1zhqm, ibem65c4, ouf18ef2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 396.5,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "GLM-5.3-Flash-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 64k 396.5 35.2 153.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jgicq7x2",
   "source_url": "https://omlx.ai/benchmarks/performance/jgicq7x2",
   "tg_tokens": null,
   "tps": 35.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jgicq7x2-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 154.4 GB at this context; duplicate traces of the same session: 252ojpp2, 2wvatp1h, 3zejkdwd, aded017j, f4i1zhqm, ibem65c4, ouf18ef2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 388.3,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "GLM-5.3-Flash-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 128k 388.3 41.1 154.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jgicq7x2",
   "source_url": "https://omlx.ai/benchmarks/performance/jgicq7x2",
   "tg_tokens": null,
   "tps": 41.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jgicq7x2-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 155.7 GB at this context; duplicate traces of the same session: 252ojpp2, 2wvatp1h, 3zejkdwd, aded017j, f4i1zhqm, ibem65c4, ouf18ef2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 380.9,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "GLM-5.3-Flash-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 195k 380.9 27.6 155.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jgicq7x2",
   "source_url": "https://omlx.ai/benchmarks/performance/jgicq7x2",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-r9ubvh9p-ctx1024",
   "model": "Qwen3 8B",
   "notes": "oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 845.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-8B-4bit on M5 (10c); M5 (10c) \u00b7 16 GB \u00b7 4bit; 1k 845.3 27.1 4.9 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r9ubvh9p",
   "source_url": "https://omlx.ai/benchmarks/performance/r9ubvh9p",
   "tg_tokens": null,
   "tps": 27.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-r9ubvh9p-ctx4096",
   "model": "Qwen3 8B",
   "notes": "oMLX community benchmark trace; peak memory 5.4 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 847.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-8B-4bit on M5 (10c); M5 (10c) \u00b7 16 GB \u00b7 4bit; 4k 847.5 24.7 5.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r9ubvh9p",
   "source_url": "https://omlx.ai/benchmarks/performance/r9ubvh9p",
   "tg_tokens": null,
   "tps": 24.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ug40j2it-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: fys9klfj, km6uwlre, v5kuybl2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 929.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 929.7 80.0 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ug40j2it",
   "source_url": "https://omlx.ai/benchmarks/performance/ug40j2it",
   "tg_tokens": null,
   "tps": 80.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ug40j2it-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: fys9klfj, km6uwlre, v5kuybl2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1077.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,077 77.2 21.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ug40j2it",
   "source_url": "https://omlx.ai/benchmarks/performance/ug40j2it",
   "tg_tokens": null,
   "tps": 77.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ug40j2it-ctx8192",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: fys9klfj, km6uwlre, v5kuybl2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1016.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 1,016 73.9 22.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ug40j2it",
   "source_url": "https://omlx.ai/benchmarks/performance/ug40j2it",
   "tg_tokens": null,
   "tps": 73.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ug40j2it-ctx16384",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: fys9klfj, km6uwlre, v5kuybl2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 883.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 883.6 68.5 22.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ug40j2it",
   "source_url": "https://omlx.ai/benchmarks/performance/ug40j2it",
   "tg_tokens": null,
   "tps": 68.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-i774kjjw-ctx1024",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2254.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 2,254 56.2 25.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i774kjjw",
   "source_url": "https://omlx.ai/benchmarks/performance/i774kjjw",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-i774kjjw-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2355.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 2,355 54.5 26.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i774kjjw",
   "source_url": "https://omlx.ai/benchmarks/performance/i774kjjw",
   "tg_tokens": null,
   "tps": 54.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-z7eee7qs-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 27.6 GB at this context; duplicate traces of the same session: 810m8rwr, lyetxr1n; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2351.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 2,351 51.6 27.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z7eee7qs",
   "source_url": "https://omlx.ai/benchmarks/performance/z7eee7qs",
   "tg_tokens": null,
   "tps": 51.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-z7eee7qs-ctx8192",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 28.3 GB at this context; duplicate traces of the same session: 810m8rwr, lyetxr1n; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2251.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 2,251 55.7 28.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z7eee7qs",
   "source_url": "https://omlx.ai/benchmarks/performance/z7eee7qs",
   "tg_tokens": null,
   "tps": 55.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-z7eee7qs-ctx16384",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 29.3 GB at this context; duplicate traces of the same session: 810m8rwr, lyetxr1n; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2082.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 16k 2,082 71.2 29.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z7eee7qs",
   "source_url": "https://omlx.ai/benchmarks/performance/z7eee7qs",
   "tg_tokens": null,
   "tps": 71.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ye54sjgq-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: bvk27gx7, xa467r0f; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2328.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 2,328 59.6 27.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ye54sjgq",
   "source_url": "https://omlx.ai/benchmarks/performance/ye54sjgq",
   "tg_tokens": null,
   "tps": 59.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ye54sjgq-ctx8192",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 28.2 GB at this context; duplicate traces of the same session: bvk27gx7, xa467r0f; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2257.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 2,257 62.3 28.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ye54sjgq",
   "source_url": "https://omlx.ai/benchmarks/performance/ye54sjgq",
   "tg_tokens": null,
   "tps": 62.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ye54sjgq-ctx16384",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 29.3 GB at this context; duplicate traces of the same session: bvk27gx7, xa467r0f; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2083.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 16k 2,083 71.5 29.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ye54sjgq",
   "source_url": "https://omlx.ai/benchmarks/performance/ye54sjgq",
   "tg_tokens": null,
   "tps": 71.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-u8630pfx-ctx8192",
   "model": "DeepSeek-R1-Distill-Qwen-7B",
   "notes": "oMLX community benchmark trace, single run at 8k context; peak memory 5.0 GB; TTFT 268934 ms; duplicate traces of the same session: none; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 30.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-R1-Distill-Qwen-7B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 4bit; 8k tokens 30.5 PP tok/s 3.2 TG tok/s 268934 TTFT (ms) 5.0 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u8630pfx",
   "source_url": "https://omlx.ai/benchmarks/performance/u8630pfx",
   "tg_tokens": null,
   "tps": 3.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-0dan65cq-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.3 GB at this context; duplicate traces of the same session: 15w5wcpd, 9tlyhfxo, t2h2zr2c, xoxxez8u; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1087.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 1,087 79.1 183.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0dan65cq",
   "source_url": "https://omlx.ai/benchmarks/performance/0dan65cq",
   "tg_tokens": null,
   "tps": 79.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-0dan65cq-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.7 GB at this context; duplicate traces of the same session: 15w5wcpd, 9tlyhfxo, t2h2zr2c, xoxxez8u; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1365.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 1,365 73.1 183.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0dan65cq",
   "source_url": "https://omlx.ai/benchmarks/performance/0dan65cq",
   "tg_tokens": null,
   "tps": 73.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-0dan65cq-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.9 GB at this context; duplicate traces of the same session: 15w5wcpd, 9tlyhfxo, t2h2zr2c, xoxxez8u; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1356.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 1,356 77.5 183.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0dan65cq",
   "source_url": "https://omlx.ai/benchmarks/performance/0dan65cq",
   "tg_tokens": null,
   "tps": 77.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0dan65cq-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 187.8 GB at this context; duplicate traces of the same session: 15w5wcpd, 9tlyhfxo, t2h2zr2c, xoxxez8u; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1315.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 64k 1,315 69.6 187.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0dan65cq",
   "source_url": "https://omlx.ai/benchmarks/performance/0dan65cq",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0dan65cq-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 192.7 GB at this context; duplicate traces of the same session: 15w5wcpd, 9tlyhfxo, t2h2zr2c, xoxxez8u; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1289.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 128k 1,289 65.4 192.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0dan65cq",
   "source_url": "https://omlx.ai/benchmarks/performance/0dan65cq",
   "tg_tokens": null,
   "tps": 65.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-19tb7w0v-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.4 GB at this context; duplicate traces of the same session: sy3nadrp; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1099.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 1,099 69.6 183.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/19tb7w0v",
   "source_url": "https://omlx.ai/benchmarks/performance/19tb7w0v",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-19tb7w0v-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.7 GB at this context; duplicate traces of the same session: sy3nadrp; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1368.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 1,368 81.8 183.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/19tb7w0v",
   "source_url": "https://omlx.ai/benchmarks/performance/19tb7w0v",
   "tg_tokens": null,
   "tps": 81.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-6b9qjq0b-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace, single run at 32k context; peak memory 26.0 GB; TTFT 17493 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1873.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k tokens 1,873 PP tok/s 119.1 TG tok/s 17493 TTFT (ms) 26.0 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6b9qjq0b",
   "source_url": "https://omlx.ai/benchmarks/performance/6b9qjq0b",
   "tg_tokens": null,
   "tps": 119.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-n48lvxef-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace, single run at 32k context; peak memory 26.0 GB; TTFT 17359 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1888.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k tokens 1,888 PP tok/s 130.8 TG tok/s 17359 TTFT (ms) 26.0 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n48lvxef",
   "source_url": "https://omlx.ai/benchmarks/performance/n48lvxef",
   "tg_tokens": null,
   "tps": 130.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-uiuln8y2-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace, single run at 32k context; peak memory 26.0 GB; TTFT 17481 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1875.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k tokens 1,875 PP tok/s 126.2 TG tok/s 17481 TTFT (ms) 26.0 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uiuln8y2",
   "source_url": "https://omlx.ai/benchmarks/performance/uiuln8y2",
   "tg_tokens": null,
   "tps": 126.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-yrmuycht-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: 50i6qzyh, liqkk794; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2359.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 2,359 35.5 27.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yrmuycht",
   "source_url": "https://omlx.ai/benchmarks/performance/yrmuycht",
   "tg_tokens": null,
   "tps": 35.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-yrmuycht-ctx8192",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 28.2 GB at this context; duplicate traces of the same session: 50i6qzyh, liqkk794; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2241.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 2,241 68.0 28.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yrmuycht",
   "source_url": "https://omlx.ai/benchmarks/performance/yrmuycht",
   "tg_tokens": null,
   "tps": 68.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-yrmuycht-ctx16384",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 29.3 GB at this context; duplicate traces of the same session: 50i6qzyh, liqkk794; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1985.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 16k 1,985 74.4 29.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yrmuycht",
   "source_url": "https://omlx.ai/benchmarks/performance/yrmuycht",
   "tg_tokens": null,
   "tps": 74.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qw5eu0dx-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 27.5 GB; TTFT 1741 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2353.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k tokens 2,353 PP tok/s 61.4 TG tok/s 1741 TTFT (ms) 27.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qw5eu0dx",
   "source_url": "https://omlx.ai/benchmarks/performance/qw5eu0dx",
   "tg_tokens": null,
   "tps": 61.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-op12ixsu-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 27.0 GB at this context; duplicate traces of the same session: eikqmqpz, p1qdkvkf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2354.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 2,354 45.3 27.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/op12ixsu",
   "source_url": "https://omlx.ai/benchmarks/performance/op12ixsu",
   "tg_tokens": null,
   "tps": 45.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-op12ixsu-ctx8192",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 27.6 GB at this context; duplicate traces of the same session: eikqmqpz, p1qdkvkf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2182.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 2,182 59.1 27.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/op12ixsu",
   "source_url": "https://omlx.ai/benchmarks/performance/op12ixsu",
   "tg_tokens": null,
   "tps": 59.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-op12ixsu-ctx16384",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 28.7 GB at this context; duplicate traces of the same session: eikqmqpz, p1qdkvkf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1936.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 16k 1,936 69.1 28.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/op12ixsu",
   "source_url": "https://omlx.ai/benchmarks/performance/op12ixsu",
   "tg_tokens": null,
   "tps": 69.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-dnr95kby-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 27.0 GB at this context; duplicate traces of the same session: led14p2s, xb4877eb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2365.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 2,365 58.6 27.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dnr95kby",
   "source_url": "https://omlx.ai/benchmarks/performance/dnr95kby",
   "tg_tokens": null,
   "tps": 58.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-dnr95kby-ctx8192",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 27.6 GB at this context; duplicate traces of the same session: led14p2s, xb4877eb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2256.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 2,256 57.2 27.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dnr95kby",
   "source_url": "https://omlx.ai/benchmarks/performance/dnr95kby",
   "tg_tokens": null,
   "tps": 57.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-dnr95kby-ctx16384",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 28.7 GB at this context; duplicate traces of the same session: led14p2s, xb4877eb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2059.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 16k 2,059 73.1 28.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dnr95kby",
   "source_url": "https://omlx.ai/benchmarks/performance/dnr95kby",
   "tg_tokens": null,
   "tps": 73.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1xdyotsn-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "QAT OptiQ build; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2339.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,339 49.9 21.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1xdyotsn",
   "source_url": "https://omlx.ai/benchmarks/performance/1xdyotsn",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1xdyotsn-ctx8192",
   "model": "gemma4 26B.A4B",
   "notes": "QAT OptiQ build; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2130.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 2,130 58.9 22.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1xdyotsn",
   "source_url": "https://omlx.ai/benchmarks/performance/1xdyotsn",
   "tg_tokens": null,
   "tps": 58.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1xdyotsn-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "QAT OptiQ build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1914.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 1,914 74.5 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1xdyotsn",
   "source_url": "https://omlx.ai/benchmarks/performance/1xdyotsn",
   "tg_tokens": null,
   "tps": 74.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-xz93jcyx-ctx1024",
   "model": "DeepSeek-Coder-V2-Lite-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 9.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2300.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-Coder-V2-Lite-Instruct-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 2,300 148.6 9.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xz93jcyx",
   "source_url": "https://omlx.ai/benchmarks/performance/xz93jcyx",
   "tg_tokens": null,
   "tps": 148.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-xz93jcyx-ctx4096",
   "model": "DeepSeek-Coder-V2-Lite-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 10.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2157.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-Coder-V2-Lite-Instruct-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 2,157 119.3 10.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xz93jcyx",
   "source_url": "https://omlx.ai/benchmarks/performance/xz93jcyx",
   "tg_tokens": null,
   "tps": 119.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-xz93jcyx-ctx8192",
   "model": "DeepSeek-Coder-V2-Lite-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 12.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1814.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-Coder-V2-Lite-Instruct-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,814 89.0 12.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xz93jcyx",
   "source_url": "https://omlx.ai/benchmarks/performance/xz93jcyx",
   "tg_tokens": null,
   "tps": 89.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-s2mj5d1w-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 183.3 GB; TTFT 933 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1097.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k tokens 1,097 PP tok/s 76.4 TG tok/s 933 TTFT (ms) 183.3 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s2mj5d1w",
   "source_url": "https://omlx.ai/benchmarks/performance/s2mj5d1w",
   "tg_tokens": null,
   "tps": 76.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9clbyr92-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.5 GB at this context; duplicate traces of the same session: uxg3zq10; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1098.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 1,098 28.4 183.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9clbyr92",
   "source_url": "https://omlx.ai/benchmarks/performance/9clbyr92",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9clbyr92-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.7 GB at this context; duplicate traces of the same session: uxg3zq10; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1362.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 1,362 73.4 183.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9clbyr92",
   "source_url": "https://omlx.ai/benchmarks/performance/9clbyr92",
   "tg_tokens": null,
   "tps": 73.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-7hjzmhjf-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.2 GB at this context; duplicate traces of the same session: jk9hdwgn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1140.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 1,140 60.2 180.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7hjzmhjf",
   "source_url": "https://omlx.ai/benchmarks/performance/7hjzmhjf",
   "tg_tokens": null,
   "tps": 60.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-7hjzmhjf-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.7 GB at this context; duplicate traces of the same session: jk9hdwgn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1412.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 1,412 55.5 180.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7hjzmhjf",
   "source_url": "https://omlx.ai/benchmarks/performance/7hjzmhjf",
   "tg_tokens": null,
   "tps": 55.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-s7dcjbpw-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 180.2 GB; TTFT 900 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1138.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k tokens 1,138 PP tok/s 60.0 TG tok/s 900 TTFT (ms) 180.2 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s7dcjbpw",
   "source_url": "https://omlx.ai/benchmarks/performance/s7dcjbpw",
   "tg_tokens": null,
   "tps": 60.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-r3cqdfme-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "QAT OptiQ build; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 26ng8ywk, 2hh0ry16; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2389.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,389 106.2 21.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r3cqdfme",
   "source_url": "https://omlx.ai/benchmarks/performance/r3cqdfme",
   "tg_tokens": null,
   "tps": 106.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-r3cqdfme-ctx8192",
   "model": "gemma4 26B.A4B",
   "notes": "QAT OptiQ build; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: 26ng8ywk, 2hh0ry16; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2228.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 2,228 74.9 22.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r3cqdfme",
   "source_url": "https://omlx.ai/benchmarks/performance/r3cqdfme",
   "tg_tokens": null,
   "tps": 74.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-r3cqdfme-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "QAT OptiQ build; oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: 26ng8ywk, 2hh0ry16; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1968.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 1,968 73.7 23.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r3cqdfme",
   "source_url": "https://omlx.ai/benchmarks/performance/r3cqdfme",
   "tg_tokens": null,
   "tps": 73.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-no3e869v-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "QAT OptiQ build; oMLX community benchmark trace, single run at 4k context; peak memory 21.3 GB; TTFT 1706 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2401.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k tokens 2,401 PP tok/s 49.8 TG tok/s 1706 TTFT (ms) 21.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/no3e869v",
   "source_url": "https://omlx.ai/benchmarks/performance/no3e869v",
   "tg_tokens": null,
   "tps": 49.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-yr7wnsgw-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "QAT OptiQ build; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 8jfpys6f, epd1rp4e; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2401.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,401 35.3 21.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yr7wnsgw",
   "source_url": "https://omlx.ai/benchmarks/performance/yr7wnsgw",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-yr7wnsgw-ctx8192",
   "model": "gemma4 26B.A4B",
   "notes": "QAT OptiQ build; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: 8jfpys6f, epd1rp4e; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2293.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 2,293 64.5 22.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yr7wnsgw",
   "source_url": "https://omlx.ai/benchmarks/performance/yr7wnsgw",
   "tg_tokens": null,
   "tps": 64.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-yr7wnsgw-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "QAT OptiQ build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 8jfpys6f, epd1rp4e; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2088.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,088 71.6 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yr7wnsgw",
   "source_url": "https://omlx.ai/benchmarks/performance/yr7wnsgw",
   "tg_tokens": null,
   "tps": 71.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-zpzlnwop-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "QAT OptiQ build; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 4wo0ixvh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2339.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,339 49.9 21.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zpzlnwop",
   "source_url": "https://omlx.ai/benchmarks/performance/zpzlnwop",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-zpzlnwop-ctx8192",
   "model": "gemma4 26B.A4B",
   "notes": "QAT OptiQ build; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: 4wo0ixvh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2130.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 2,130 58.9 22.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zpzlnwop",
   "source_url": "https://omlx.ai/benchmarks/performance/zpzlnwop",
   "tg_tokens": null,
   "tps": 58.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-zpzlnwop-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "QAT OptiQ build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 4wo0ixvh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1914.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 1,914 74.5 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zpzlnwop",
   "source_url": "https://omlx.ai/benchmarks/performance/zpzlnwop",
   "tg_tokens": null,
   "tps": 74.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-o9p0whcq-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "QAT OptiQ build; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: jvz2i8w5, ozdmixxj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2361.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,361 58.2 20.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9p0whcq",
   "source_url": "https://omlx.ai/benchmarks/performance/o9p0whcq",
   "tg_tokens": null,
   "tps": 58.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-o9p0whcq-ctx8192",
   "model": "gemma4 26B.A4B",
   "notes": "QAT OptiQ build; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: jvz2i8w5, ozdmixxj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2312.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 2,312 56.0 20.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9p0whcq",
   "source_url": "https://omlx.ai/benchmarks/performance/o9p0whcq",
   "tg_tokens": null,
   "tps": 56.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-o9p0whcq-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "QAT OptiQ build; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: jvz2i8w5, ozdmixxj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2161.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,161 50.8 21.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9p0whcq",
   "source_url": "https://omlx.ai/benchmarks/performance/o9p0whcq",
   "tg_tokens": null,
   "tps": 50.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-26uxs0jl-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 15.3 GB at this context; duplicate traces of the same session: nh1hzx44, r7bkchwp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2515.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,515 75.2 15.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/26uxs0jl",
   "source_url": "https://omlx.ai/benchmarks/performance/26uxs0jl",
   "tg_tokens": null,
   "tps": 75.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-26uxs0jl-ctx8192",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: nh1hzx44, r7bkchwp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2380.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 2,380 90.5 15.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/26uxs0jl",
   "source_url": "https://omlx.ai/benchmarks/performance/26uxs0jl",
   "tg_tokens": null,
   "tps": 90.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-26uxs0jl-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: nh1hzx44, r7bkchwp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2137.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,137 91.7 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/26uxs0jl",
   "source_url": "https://omlx.ai/benchmarks/performance/26uxs0jl",
   "tg_tokens": null,
   "tps": 91.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-z4e6pk80-ctx1024",
   "model": "DeepSeek-Coder-V2-Lite-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 9.6 GB at this context; duplicate traces of the same session: db9fi1u3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2300.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-Coder-V2-Lite-Instruct-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 2,300 148.6 9.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z4e6pk80",
   "source_url": "https://omlx.ai/benchmarks/performance/z4e6pk80",
   "tg_tokens": null,
   "tps": 148.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-z4e6pk80-ctx4096",
   "model": "DeepSeek-Coder-V2-Lite-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 10.6 GB at this context; duplicate traces of the same session: db9fi1u3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2157.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-Coder-V2-Lite-Instruct-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 2,157 119.3 10.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z4e6pk80",
   "source_url": "https://omlx.ai/benchmarks/performance/z4e6pk80",
   "tg_tokens": null,
   "tps": 119.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-z4e6pk80-ctx8192",
   "model": "DeepSeek-Coder-V2-Lite-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 12.2 GB at this context; duplicate traces of the same session: db9fi1u3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1814.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-Coder-V2-Lite-Instruct-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,814 89.0 12.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z4e6pk80",
   "source_url": "https://omlx.ai/benchmarks/performance/z4e6pk80",
   "tg_tokens": null,
   "tps": 89.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-s9ohpfla-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1701.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,701 139.5 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s9ohpfla",
   "source_url": "https://omlx.ai/benchmarks/performance/s9ohpfla",
   "tg_tokens": null,
   "tps": 139.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-s9ohpfla-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 598.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 598.2 44.8 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s9ohpfla",
   "source_url": "https://omlx.ai/benchmarks/performance/s9ohpfla",
   "tg_tokens": null,
   "tps": 44.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ohrpdk1a-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.2 GB at this context; duplicate traces of the same session: ve0fggfb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1137.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 1,137 59.9 180.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ohrpdk1a",
   "source_url": "https://omlx.ai/benchmarks/performance/ohrpdk1a",
   "tg_tokens": null,
   "tps": 59.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ohrpdk1a-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.7 GB at this context; duplicate traces of the same session: ve0fggfb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1408.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 1,408 55.7 180.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ohrpdk1a",
   "source_url": "https://omlx.ai/benchmarks/performance/ohrpdk1a",
   "tg_tokens": null,
   "tps": 55.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ctuk7shx-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "oMLX build; oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 433.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Swift-1.5-Qwen3.8-27B-MLX-oMLX on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 433.8 25.0 19.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ctuk7shx",
   "source_url": "https://omlx.ai/benchmarks/performance/ctuk7shx",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ctuk7shx-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "oMLX build; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 446.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Swift-1.5-Qwen3.8-27B-MLX-oMLX on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 446.5 30.1 21.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ctuk7shx",
   "source_url": "https://omlx.ai/benchmarks/performance/ctuk7shx",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-d61idl78-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace, single run at 32k context; peak memory 26.0 GB; TTFT 17595 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1862.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k tokens 1,862 PP tok/s 47.8 TG tok/s 17595 TTFT (ms) 26.0 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d61idl78",
   "source_url": "https://omlx.ai/benchmarks/performance/d61idl78",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-li2mvxwo-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Youssofal MTPLX Optimized Quality FP16; oMLX community benchmark trace, single run at 1k context; peak memory 33.8 GB; TTFT 11924 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 85.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Youssofal--Qwen3.8-27B-MTPLX-Optimized-Quality-FP16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 1k tokens 85.9 PP tok/s 9.8 TG tok/s 11924 TTFT (ms) 33.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/li2mvxwo",
   "source_url": "https://omlx.ai/benchmarks/performance/li2mvxwo",
   "tg_tokens": null,
   "tps": 9.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-ltk3dbf3-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1090.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,090 46.9 14.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ltk3dbf3",
   "source_url": "https://omlx.ai/benchmarks/performance/ltk3dbf3",
   "tg_tokens": null,
   "tps": 46.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-ltk3dbf3-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 15.1 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1361.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 1,361 40.7 15.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ltk3dbf3",
   "source_url": "https://omlx.ai/benchmarks/performance/ltk3dbf3",
   "tg_tokens": null,
   "tps": 40.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-g41z9twn-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 15.2 GB at this context; duplicate traces of the same session: hf0wgbf9, u8x4qzrh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2512.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,512 70.0 15.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g41z9twn",
   "source_url": "https://omlx.ai/benchmarks/performance/g41z9twn",
   "tg_tokens": null,
   "tps": 70.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-g41z9twn-ctx8192",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: hf0wgbf9, u8x4qzrh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2386.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 2,386 99.2 15.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g41z9twn",
   "source_url": "https://omlx.ai/benchmarks/performance/g41z9twn",
   "tg_tokens": null,
   "tps": 99.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-g41z9twn-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: hf0wgbf9, u8x4qzrh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2188.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,188 98.3 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g41z9twn",
   "source_url": "https://omlx.ai/benchmarks/performance/g41z9twn",
   "tg_tokens": null,
   "tps": 98.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-v396rgn2-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: kw9cm2xn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2509.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,509 81.5 14.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v396rgn2",
   "source_url": "https://omlx.ai/benchmarks/performance/v396rgn2",
   "tg_tokens": null,
   "tps": 81.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-v396rgn2-ctx8192",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: kw9cm2xn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2441.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 2,441 76.6 14.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v396rgn2",
   "source_url": "https://omlx.ai/benchmarks/performance/v396rgn2",
   "tg_tokens": null,
   "tps": 76.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-v396rgn2-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.9 GB at this context; duplicate traces of the same session: kw9cm2xn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2283.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,283 67.6 14.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v396rgn2",
   "source_url": "https://omlx.ai/benchmarks/performance/v396rgn2",
   "tg_tokens": null,
   "tps": 67.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-745zpq61-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: fdmscgb7, ul65y7aw; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2509.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,509 79.9 15.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/745zpq61",
   "source_url": "https://omlx.ai/benchmarks/performance/745zpq61",
   "tg_tokens": null,
   "tps": 79.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-745zpq61-ctx8192",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 16.5 GB at this context; duplicate traces of the same session: fdmscgb7, ul65y7aw; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2382.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 2,382 85.6 16.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/745zpq61",
   "source_url": "https://omlx.ai/benchmarks/performance/745zpq61",
   "tg_tokens": null,
   "tps": 85.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-745zpq61-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 17.5 GB at this context; duplicate traces of the same session: fdmscgb7, ul65y7aw; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2175.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,175 94.5 17.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/745zpq61",
   "source_url": "https://omlx.ai/benchmarks/performance/745zpq61",
   "tg_tokens": null,
   "tps": 94.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-aaapdn8z-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: ie1wz22a, le80gw79; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2505.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,505 56.2 15.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aaapdn8z",
   "source_url": "https://omlx.ai/benchmarks/performance/aaapdn8z",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-aaapdn8z-ctx8192",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 16.5 GB at this context; duplicate traces of the same session: ie1wz22a, le80gw79; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2383.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 2,383 94.0 16.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aaapdn8z",
   "source_url": "https://omlx.ai/benchmarks/performance/aaapdn8z",
   "tg_tokens": null,
   "tps": 94.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-aaapdn8z-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 17.5 GB at this context; duplicate traces of the same session: ie1wz22a, le80gw79; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2187.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,187 86.6 17.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aaapdn8z",
   "source_url": "https://omlx.ai/benchmarks/performance/aaapdn8z",
   "tg_tokens": null,
   "tps": 86.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-p6468uqx-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 15.3 GB at this context; duplicate traces of the same session: i59aoj19; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2514.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,514 60.1 15.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p6468uqx",
   "source_url": "https://omlx.ai/benchmarks/performance/p6468uqx",
   "tg_tokens": null,
   "tps": 60.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-p6468uqx-ctx8192",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: i59aoj19; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2387.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 2,387 96.5 15.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p6468uqx",
   "source_url": "https://omlx.ai/benchmarks/performance/p6468uqx",
   "tg_tokens": null,
   "tps": 96.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-p6468uqx-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: i59aoj19; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2190.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,190 101.6 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p6468uqx",
   "source_url": "https://omlx.ai/benchmarks/performance/p6468uqx",
   "tg_tokens": null,
   "tps": 101.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-aipf8t3y-ctx1024",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "MLX; oMLX community benchmark trace; peak memory 30.8 GB at this context; duplicate traces of the same session: q515bi15; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2430.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 2,430 65.2 30.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aipf8t3y",
   "source_url": "https://omlx.ai/benchmarks/performance/aipf8t3y",
   "tg_tokens": null,
   "tps": 65.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-aipf8t3y-ctx4096",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "MLX; oMLX community benchmark trace; peak memory 31.3 GB at this context; duplicate traces of the same session: q515bi15; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2451.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 2,451 59.5 31.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aipf8t3y",
   "source_url": "https://omlx.ai/benchmarks/performance/aipf8t3y",
   "tg_tokens": null,
   "tps": 59.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-9m2ft47x-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1701.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,701 139.5 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9m2ft47x",
   "source_url": "https://omlx.ai/benchmarks/performance/9m2ft47x",
   "tg_tokens": null,
   "tps": 139.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-9m2ft47x-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 598.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 598.2 44.8 21.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9m2ft47x",
   "source_url": "https://omlx.ai/benchmarks/performance/9m2ft47x",
   "tg_tokens": null,
   "tps": 44.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4aimpyeb-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 870.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 870.7 32.0 16.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4aimpyeb",
   "source_url": "https://omlx.ai/benchmarks/performance/4aimpyeb",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4aimpyeb-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 448.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 448.4 31.4 17.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4aimpyeb",
   "source_url": "https://omlx.ai/benchmarks/performance/4aimpyeb",
   "tg_tokens": null,
   "tps": 31.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-btdqfz59-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 256.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 256.9 27.9 16.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/btdqfz59",
   "source_url": "https://omlx.ai/benchmarks/performance/btdqfz59",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-btdqfz59-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 244.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 244.1 24.6 19.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/btdqfz59",
   "source_url": "https://omlx.ai/benchmarks/performance/btdqfz59",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-btdqfz59-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 230.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 230.2 23.2 20.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/btdqfz59",
   "source_url": "https://omlx.ai/benchmarks/performance/btdqfz59",
   "tg_tokens": null,
   "tps": 23.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-btdqfz59-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 218.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 218.4 21.7 21.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/btdqfz59",
   "source_url": "https://omlx.ai/benchmarks/performance/btdqfz59",
   "tg_tokens": null,
   "tps": 21.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-btdqfz59-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 189.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 189.7 16.6 22.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/btdqfz59",
   "source_url": "https://omlx.ai/benchmarks/performance/btdqfz59",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-btdqfz59-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 155.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 155.3 14.6 25.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/btdqfz59",
   "source_url": "https://omlx.ai/benchmarks/performance/btdqfz59",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-btdqfz59-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 125.8 11.5 29.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/btdqfz59",
   "source_url": "https://omlx.ai/benchmarks/performance/btdqfz59",
   "tg_tokens": null,
   "tps": 11.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-btdqfz59-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 100.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 100.1 8.0 33.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/btdqfz59",
   "source_url": "https://omlx.ai/benchmarks/performance/btdqfz59",
   "tg_tokens": null,
   "tps": 8.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-oxfi0701-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "oMLX build; oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 433.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Swift-1.5-Qwen3.8-27B-MLX-oMLX on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 433.8 25.0 19.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oxfi0701",
   "source_url": "https://omlx.ai/benchmarks/performance/oxfi0701",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-oxfi0701-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "oMLX build; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 446.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Swift-1.5-Qwen3.8-27B-MLX-oMLX on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 446.5 30.1 21.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oxfi0701",
   "source_url": "https://omlx.ai/benchmarks/performance/oxfi0701",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-sqvmyyvm-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace, single run at 32k context; peak memory 26.0 GB; TTFT 17681 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1853.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k tokens 1,853 PP tok/s 44.6 TG tok/s 17681 TTFT (ms) 26.0 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sqvmyyvm",
   "source_url": "https://omlx.ai/benchmarks/performance/sqvmyyvm",
   "tg_tokens": null,
   "tps": 44.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-pfid2e8f-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1090.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,090 46.9 14.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pfid2e8f",
   "source_url": "https://omlx.ai/benchmarks/performance/pfid2e8f",
   "tg_tokens": null,
   "tps": 46.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-pfid2e8f-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 15.1 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1361.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 1,361 40.7 15.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pfid2e8f",
   "source_url": "https://omlx.ai/benchmarks/performance/pfid2e8f",
   "tg_tokens": null,
   "tps": 40.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-lm0qie0e-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2509.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,509 81.5 14.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lm0qie0e",
   "source_url": "https://omlx.ai/benchmarks/performance/lm0qie0e",
   "tg_tokens": null,
   "tps": 81.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-lm0qie0e-ctx8192",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2441.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 2,441 76.6 14.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lm0qie0e",
   "source_url": "https://omlx.ai/benchmarks/performance/lm0qie0e",
   "tg_tokens": null,
   "tps": 76.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-lm0qie0e-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.9 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2283.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,283 67.6 14.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lm0qie0e",
   "source_url": "https://omlx.ai/benchmarks/performance/lm0qie0e",
   "tg_tokens": null,
   "tps": 67.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-junjzbqn-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 15.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2514.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,514 60.1 15.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/junjzbqn",
   "source_url": "https://omlx.ai/benchmarks/performance/junjzbqn",
   "tg_tokens": null,
   "tps": 60.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-junjzbqn-ctx8192",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2387.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 2,387 96.5 15.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/junjzbqn",
   "source_url": "https://omlx.ai/benchmarks/performance/junjzbqn",
   "tg_tokens": null,
   "tps": 96.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-junjzbqn-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2190.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,190 101.6 17.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/junjzbqn",
   "source_url": "https://omlx.ai/benchmarks/performance/junjzbqn",
   "tg_tokens": null,
   "tps": 101.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fvopq0db-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 15.2 GB at this context; duplicate traces of the same session: 6gof8pwa, powel4j1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2514.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,514 123.0 15.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fvopq0db",
   "source_url": "https://omlx.ai/benchmarks/performance/fvopq0db",
   "tg_tokens": null,
   "tps": 123.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-fvopq0db-ctx8192",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: 6gof8pwa, powel4j1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2389.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 2,389 86.8 15.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fvopq0db",
   "source_url": "https://omlx.ai/benchmarks/performance/fvopq0db",
   "tg_tokens": null,
   "tps": 86.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-fvopq0db-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 6gof8pwa, powel4j1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2183.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,183 89.0 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fvopq0db",
   "source_url": "https://omlx.ai/benchmarks/performance/fvopq0db",
   "tg_tokens": null,
   "tps": 89.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1wjto62x-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: fmvzbqea, phppf6cv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2520.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,520 81.4 14.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1wjto62x",
   "source_url": "https://omlx.ai/benchmarks/performance/1wjto62x",
   "tg_tokens": null,
   "tps": 81.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1wjto62x-ctx8192",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: fmvzbqea, phppf6cv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2448.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 2,448 76.4 14.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1wjto62x",
   "source_url": "https://omlx.ai/benchmarks/performance/1wjto62x",
   "tg_tokens": null,
   "tps": 76.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1wjto62x-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.9 GB at this context; duplicate traces of the same session: fmvzbqea, phppf6cv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2281.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,281 67.4 14.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1wjto62x",
   "source_url": "https://omlx.ai/benchmarks/performance/1wjto62x",
   "tg_tokens": null,
   "tps": 67.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-lewo0fxr-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 15.2 GB at this context; duplicate traces of the same session: 9ncej3w7, gc9ixsfy; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2513.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,513 56.4 15.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lewo0fxr",
   "source_url": "https://omlx.ai/benchmarks/performance/lewo0fxr",
   "tg_tokens": null,
   "tps": 56.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-lewo0fxr-ctx8192",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: 9ncej3w7, gc9ixsfy; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2387.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 2,387 89.4 15.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lewo0fxr",
   "source_url": "https://omlx.ai/benchmarks/performance/lewo0fxr",
   "tg_tokens": null,
   "tps": 89.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-lewo0fxr-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 9ncej3w7, gc9ixsfy; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2190.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,190 103.9 17.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lewo0fxr",
   "source_url": "https://omlx.ai/benchmarks/performance/lewo0fxr",
   "tg_tokens": null,
   "tps": 103.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-g6nobejk-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: bousip8b, tv6elgub; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2512.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,512 81.4 14.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g6nobejk",
   "source_url": "https://omlx.ai/benchmarks/performance/g6nobejk",
   "tg_tokens": null,
   "tps": 81.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-g6nobejk-ctx8192",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: bousip8b, tv6elgub; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2440.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 2,440 76.6 14.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g6nobejk",
   "source_url": "https://omlx.ai/benchmarks/performance/g6nobejk",
   "tg_tokens": null,
   "tps": 76.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-g6nobejk-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.9 GB at this context; duplicate traces of the same session: bousip8b, tv6elgub; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2281.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,281 67.4 14.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g6nobejk",
   "source_url": "https://omlx.ai/benchmarks/performance/g6nobejk",
   "tg_tokens": null,
   "tps": 67.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1",
   "id": "omlx-3trn15l0-ctx32768",
   "model": "DeepSeek-R1-Distill-Qwen-7B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 6.3 GB; TTFT 1528812 ms; duplicate traces of the same session: none; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 21.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-R1-Distill-Qwen-7B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 4bit; 32k tokens 21.4 PP tok/s 2.1 TG tok/s 1528812 TTFT (ms) 6.3 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3trn15l0",
   "source_url": "https://omlx.ai/benchmarks/performance/3trn15l0",
   "tg_tokens": null,
   "tps": 2.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7n4xdrgb-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 30.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4831.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,831 126.7 30.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7n4xdrgb",
   "source_url": "https://omlx.ai/benchmarks/performance/7n4xdrgb",
   "tg_tokens": null,
   "tps": 126.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-7n4xdrgb-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 31.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6060.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,060 137.2 31.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7n4xdrgb",
   "source_url": "https://omlx.ai/benchmarks/performance/7n4xdrgb",
   "tg_tokens": null,
   "tps": 137.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fi0uyt43-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: u624vjje; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 907.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 907.3 94.2 35.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fi0uyt43",
   "source_url": "https://omlx.ai/benchmarks/performance/fi0uyt43",
   "tg_tokens": null,
   "tps": 94.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fi0uyt43-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: u624vjje; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2734.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 2,734 92.3 36.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fi0uyt43",
   "source_url": "https://omlx.ai/benchmarks/performance/fi0uyt43",
   "tg_tokens": null,
   "tps": 92.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-74h4h47g-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: 0kacz3xq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2117.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 2,117 115.7 27.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/74h4h47g",
   "source_url": "https://omlx.ai/benchmarks/performance/74h4h47g",
   "tg_tokens": null,
   "tps": 115.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-74h4h47g-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: 0kacz3xq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3228.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 3,228 114.3 28.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/74h4h47g",
   "source_url": "https://omlx.ai/benchmarks/performance/74h4h47g",
   "tg_tokens": null,
   "tps": 114.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-d4bmbn3p-ctx1024",
   "model": "Qwen3.8-27B-OBLITERATED-Mythos-Class-Agentic",
   "notes": "oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: hgaima4u; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 437.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-OBLITERATED-Mythos-Class-Agentic-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 437.4 12.4 22.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d4bmbn3p",
   "source_url": "https://omlx.ai/benchmarks/performance/d4bmbn3p",
   "tg_tokens": null,
   "tps": 12.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-d4bmbn3p-ctx4096",
   "model": "Qwen3.8-27B-OBLITERATED-Mythos-Class-Agentic",
   "notes": "oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: hgaima4u; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 477.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-OBLITERATED-Mythos-Class-Agentic-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 477.3 12.5 23.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d4bmbn3p",
   "source_url": "https://omlx.ai/benchmarks/performance/d4bmbn3p",
   "tg_tokens": null,
   "tps": 12.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-73iea9vt-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 870.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 870.7 32.0 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/73iea9vt",
   "source_url": "https://omlx.ai/benchmarks/performance/73iea9vt",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-73iea9vt-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 448.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 448.4 31.4 17.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/73iea9vt",
   "source_url": "https://omlx.ai/benchmarks/performance/73iea9vt",
   "tg_tokens": null,
   "tps": 31.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ljuarwtl-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 11.7 GB at this context; duplicate traces of the same session: vzlb3m2z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 881.8,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ2e on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 2bit; 1k 881.8 40.0 11.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ljuarwtl",
   "source_url": "https://omlx.ai/benchmarks/performance/ljuarwtl",
   "tg_tokens": null,
   "tps": 40.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ljuarwtl-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 13.0 GB at this context; duplicate traces of the same session: vzlb3m2z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 432.5,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ2e on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 2bit; 4k 432.5 40.8 13.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ljuarwtl",
   "source_url": "https://omlx.ai/benchmarks/performance/ljuarwtl",
   "tg_tokens": null,
   "tps": 40.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-nzzqtca7-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 11.9 GB at this context; duplicate traces of the same session: soi8thkw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 835.6,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ2e on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 2bit; 1k 835.6 44.2 11.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nzzqtca7",
   "source_url": "https://omlx.ai/benchmarks/performance/nzzqtca7",
   "tg_tokens": null,
   "tps": 44.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-nzzqtca7-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "community merge; oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: soi8thkw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 449.0,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ2e on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 2bit; 4k 449.0 43.3 12.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nzzqtca7",
   "source_url": "https://omlx.ai/benchmarks/performance/nzzqtca7",
   "tg_tokens": null,
   "tps": 43.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-o21wrk2z-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 67hf29m0, 8il75exn, e0jj4cp4, feu1qoov, jt3cxcv0, x6vzxqbn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 256.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 256.9 27.9 16.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o21wrk2z",
   "source_url": "https://omlx.ai/benchmarks/performance/o21wrk2z",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-o21wrk2z-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 67hf29m0, 8il75exn, e0jj4cp4, feu1qoov, jt3cxcv0, x6vzxqbn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 244.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 244.1 24.6 19.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o21wrk2z",
   "source_url": "https://omlx.ai/benchmarks/performance/o21wrk2z",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-o21wrk2z-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 67hf29m0, 8il75exn, e0jj4cp4, feu1qoov, jt3cxcv0, x6vzxqbn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 230.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 230.2 23.2 20.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o21wrk2z",
   "source_url": "https://omlx.ai/benchmarks/performance/o21wrk2z",
   "tg_tokens": null,
   "tps": 23.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-o21wrk2z-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 67hf29m0, 8il75exn, e0jj4cp4, feu1qoov, jt3cxcv0, x6vzxqbn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 218.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 218.4 21.7 21.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o21wrk2z",
   "source_url": "https://omlx.ai/benchmarks/performance/o21wrk2z",
   "tg_tokens": null,
   "tps": 21.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-o21wrk2z-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: 67hf29m0, 8il75exn, e0jj4cp4, feu1qoov, jt3cxcv0, x6vzxqbn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 189.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 189.7 16.6 22.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o21wrk2z",
   "source_url": "https://omlx.ai/benchmarks/performance/o21wrk2z",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-o21wrk2z-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: 67hf29m0, 8il75exn, e0jj4cp4, feu1qoov, jt3cxcv0, x6vzxqbn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 155.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 155.3 14.6 25.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o21wrk2z",
   "source_url": "https://omlx.ai/benchmarks/performance/o21wrk2z",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-11",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-o21wrk2z-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.3 GB at this context; duplicate traces of the same session: 67hf29m0, 8il75exn, e0jj4cp4, feu1qoov, jt3cxcv0, x6vzxqbn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 125.8 11.5 29.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o21wrk2z",
   "source_url": "https://omlx.ai/benchmarks/performance/o21wrk2z",
   "tg_tokens": null,
   "tps": 11.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-11",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-o21wrk2z-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.5 GB at this context; duplicate traces of the same session: 67hf29m0, 8il75exn, e0jj4cp4, feu1qoov, jt3cxcv0, x6vzxqbn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 100.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 100.1 8.0 33.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o21wrk2z",
   "source_url": "https://omlx.ai/benchmarks/performance/o21wrk2z",
   "tg_tokens": null,
   "tps": 8.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ztlifein-ctx1024",
   "model": "Gemma 4 12B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: 7w51w4nq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 907.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-12B-it-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 907.3 38.2 12.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ztlifein",
   "source_url": "https://omlx.ai/benchmarks/performance/ztlifein",
   "tg_tokens": null,
   "tps": 38.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ztlifein-ctx4096",
   "model": "Gemma 4 12B",
   "notes": "oMLX community benchmark trace; peak memory 13.2 GB at this context; duplicate traces of the same session: 7w51w4nq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1505.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-12B-it-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 1,505 38.3 13.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ztlifein",
   "source_url": "https://omlx.ai/benchmarks/performance/ztlifein",
   "tg_tokens": null,
   "tps": 38.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-88m4imqs-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 87.1 GB; TTFT 99971 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 327.8,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "GLM-5.3-Flash-oQ2e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 32k tokens 327.8 PP tok/s 17.1 TG tok/s 99971 TTFT (ms) 87.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/88m4imqs",
   "source_url": "https://omlx.ai/benchmarks/performance/88m4imqs",
   "tg_tokens": null,
   "tps": 17.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-m1x2ac51-ctx65536",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 21.0 GB; TTFT 56432 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1161.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 64k tokens 1,161 PP tok/s 38.0 TG tok/s 56432 TTFT (ms) 21.0 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m1x2ac51",
   "source_url": "https://omlx.ai/benchmarks/performance/m1x2ac51",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-uaqe9yxh-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 229.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 229.9 20.7 5.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uaqe9yxh",
   "source_url": "https://omlx.ai/benchmarks/performance/uaqe9yxh",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-uaqe9yxh-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 237.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 237.3 20.7 6.2 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uaqe9yxh",
   "source_url": "https://omlx.ai/benchmarks/performance/uaqe9yxh",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-g0aromhv-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: kftv3wja; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 727.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 727.8 126.9 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g0aromhv",
   "source_url": "https://omlx.ai/benchmarks/performance/g0aromhv",
   "tg_tokens": null,
   "tps": 126.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-g0aromhv-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: kftv3wja; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2291.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,291 120.9 20.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g0aromhv",
   "source_url": "https://omlx.ai/benchmarks/performance/g0aromhv",
   "tg_tokens": null,
   "tps": 120.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-v67cr6uj-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 30.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4831.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,831 126.7 30.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v67cr6uj",
   "source_url": "https://omlx.ai/benchmarks/performance/v67cr6uj",
   "tg_tokens": null,
   "tps": 126.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-v67cr6uj-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 31.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6060.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,060 137.2 31.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v67cr6uj",
   "source_url": "https://omlx.ai/benchmarks/performance/v67cr6uj",
   "tg_tokens": null,
   "tps": 137.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xc4rye1y-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: j5mrv9gx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1484.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 1,484 82.2 27.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xc4rye1y",
   "source_url": "https://omlx.ai/benchmarks/performance/xc4rye1y",
   "tg_tokens": null,
   "tps": 82.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xc4rye1y-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.3 GB at this context; duplicate traces of the same session: j5mrv9gx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3240.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 3,240 99.8 28.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xc4rye1y",
   "source_url": "https://omlx.ai/benchmarks/performance/xc4rye1y",
   "tg_tokens": null,
   "tps": 99.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qq8ocdaa-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 930.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 930.5 33.5 16.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qq8ocdaa",
   "source_url": "https://omlx.ai/benchmarks/performance/qq8ocdaa",
   "tg_tokens": null,
   "tps": 33.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qq8ocdaa-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1010.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,010 32.9 17.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qq8ocdaa",
   "source_url": "https://omlx.ai/benchmarks/performance/qq8ocdaa",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qq8ocdaa-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 860.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 860.2 31.0 17.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qq8ocdaa",
   "source_url": "https://omlx.ai/benchmarks/performance/qq8ocdaa",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-gjx97irr-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 28.5 GB; TTFT 2908 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 352.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 8bit; 1k tokens 352.1 PP tok/s 9.9 TG tok/s 2908 TTFT (ms) 28.5 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gjx97irr",
   "source_url": "https://omlx.ai/benchmarks/performance/gjx97irr",
   "tg_tokens": null,
   "tps": 9.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-7s33d1rh-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 136cdo64, hw73c5l3, lolbfv08, wk5tfz3r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 613.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 613.5 69.6 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7s33d1rh",
   "source_url": "https://omlx.ai/benchmarks/performance/7s33d1rh",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-7s33d1rh-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: 136cdo64, hw73c5l3, lolbfv08, wk5tfz3r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 726.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 726.7 57.5 71.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7s33d1rh",
   "source_url": "https://omlx.ai/benchmarks/performance/7s33d1rh",
   "tg_tokens": null,
   "tps": 57.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-7s33d1rh-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: 136cdo64, hw73c5l3, lolbfv08, wk5tfz3r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 835.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 835.7 62.5 71.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7s33d1rh",
   "source_url": "https://omlx.ai/benchmarks/performance/7s33d1rh",
   "tg_tokens": null,
   "tps": 62.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-7s33d1rh-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 136cdo64, hw73c5l3, lolbfv08, wk5tfz3r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 830.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 830.2 78.3 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7s33d1rh",
   "source_url": "https://omlx.ai/benchmarks/performance/7s33d1rh",
   "tg_tokens": null,
   "tps": 78.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-7s33d1rh-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 72.6 GB at this context; duplicate traces of the same session: 136cdo64, hw73c5l3, lolbfv08, wk5tfz3r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 823.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 823.6 44.7 72.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7s33d1rh",
   "source_url": "https://omlx.ai/benchmarks/performance/7s33d1rh",
   "tg_tokens": null,
   "tps": 44.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-jn28msud-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 8v1uontn, 9d86ht5o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 544.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 544.8 59.6 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jn28msud",
   "source_url": "https://omlx.ai/benchmarks/performance/jn28msud",
   "tg_tokens": null,
   "tps": 59.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-jn28msud-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: 8v1uontn, 9d86ht5o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 684.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 684.8 61.8 71.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jn28msud",
   "source_url": "https://omlx.ai/benchmarks/performance/jn28msud",
   "tg_tokens": null,
   "tps": 61.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-jn28msud-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: 8v1uontn, 9d86ht5o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 833.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 833.3 75.1 71.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jn28msud",
   "source_url": "https://omlx.ai/benchmarks/performance/jn28msud",
   "tg_tokens": null,
   "tps": 75.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-x3gvp2nq-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace, single run at 32k context; peak memory 26.0 GB; TTFT 23741 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1380.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k tokens 1,380 PP tok/s 115.4 TG tok/s 23741 TTFT (ms) 26.0 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x3gvp2nq",
   "source_url": "https://omlx.ai/benchmarks/performance/x3gvp2nq",
   "tg_tokens": null,
   "tps": 115.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-doq6msi7-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace, single run at 32k context; peak memory 26.0 GB; TTFT 17515 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1871.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k tokens 1,871 PP tok/s 117.2 TG tok/s 17515 TTFT (ms) 26.0 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/doq6msi7",
   "source_url": "https://omlx.ai/benchmarks/performance/doq6msi7",
   "tg_tokens": null,
   "tps": 117.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-z0z00iwt-ctx65536",
   "model": "Occamy-1.0",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 23.2 GB; TTFT 35081 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1868.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "occamy-1.0-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 96 GB \u00b7 4bit; 64k tokens 1,868 PP tok/s 73.6 TG tok/s 35081 TTFT (ms) 23.2 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z0z00iwt",
   "source_url": "https://omlx.ai/benchmarks/performance/z0z00iwt",
   "tg_tokens": null,
   "tps": 73.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-agb4s0wm-ctx65536",
   "model": "Occamy-1.0",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 23.2 GB; TTFT 34946 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1875.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "occamy-1.0-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 96 GB \u00b7 4bit; 64k tokens 1,875 PP tok/s 85.3 TG tok/s 34946 TTFT (ms) 23.2 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/agb4s0wm",
   "source_url": "https://omlx.ai/benchmarks/performance/agb4s0wm",
   "tg_tokens": null,
   "tps": 85.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-abdrca0a-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.8 GB at this context; duplicate traces of the same session: vhso8k8a, z4trt1g3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 466.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 466.3 31.0 173.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/abdrca0a",
   "source_url": "https://omlx.ai/benchmarks/performance/abdrca0a",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-abdrca0a-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.5 GB at this context; duplicate traces of the same session: vhso8k8a, z4trt1g3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 578.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 578.2 27.1 177.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/abdrca0a",
   "source_url": "https://omlx.ai/benchmarks/performance/abdrca0a",
   "tg_tokens": null,
   "tps": 27.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-abdrca0a-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.6 GB at this context; duplicate traces of the same session: vhso8k8a, z4trt1g3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 533.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 533.9 26.9 177.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/abdrca0a",
   "source_url": "https://omlx.ai/benchmarks/performance/abdrca0a",
   "tg_tokens": null,
   "tps": 26.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-abdrca0a-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.7 GB at this context; duplicate traces of the same session: vhso8k8a, z4trt1g3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 524.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 524.6 19.8 177.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/abdrca0a",
   "source_url": "https://omlx.ai/benchmarks/performance/abdrca0a",
   "tg_tokens": null,
   "tps": 19.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-abdrca0a-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.9 GB at this context; duplicate traces of the same session: vhso8k8a, z4trt1g3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 516.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 516.4 26.4 177.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/abdrca0a",
   "source_url": "https://omlx.ai/benchmarks/performance/abdrca0a",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-abdrca0a-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.3 GB at this context; duplicate traces of the same session: vhso8k8a, z4trt1g3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 504.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 504.8 25.4 178.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/abdrca0a",
   "source_url": "https://omlx.ai/benchmarks/performance/abdrca0a",
   "tg_tokens": null,
   "tps": 25.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-abdrca0a-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 179.8 GB at this context; duplicate traces of the same session: vhso8k8a, z4trt1g3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 259.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 259.1 7.1 179.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/abdrca0a",
   "source_url": "https://omlx.ai/benchmarks/performance/abdrca0a",
   "tg_tokens": null,
   "tps": 7.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-3b45h0ig-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 229.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 229.9 20.7 5.8 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3b45h0ig",
   "source_url": "https://omlx.ai/benchmarks/performance/3b45h0ig",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-3b45h0ig-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 237.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 237.3 20.7 6.2 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3b45h0ig",
   "source_url": "https://omlx.ai/benchmarks/performance/3b45h0ig",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-019bm1tf-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 68r1k9ac; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 53.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 53.9 10.3 17.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/019bm1tf",
   "source_url": "https://omlx.ai/benchmarks/performance/019bm1tf",
   "tg_tokens": null,
   "tps": 10.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-019bm1tf-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: 68r1k9ac; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 48.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k 48.4 9.4 18.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/019bm1tf",
   "source_url": "https://omlx.ai/benchmarks/performance/019bm1tf",
   "tg_tokens": null,
   "tps": 9.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yhva1p1n-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: 4svv630v, ouwt1rni, xmdbvod2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 933.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 933.4 32.7 17.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yhva1p1n",
   "source_url": "https://omlx.ai/benchmarks/performance/yhva1p1n",
   "tg_tokens": null,
   "tps": 32.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yhva1p1n-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 4svv630v, ouwt1rni, xmdbvod2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 873.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 873.3 30.7 17.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yhva1p1n",
   "source_url": "https://omlx.ai/benchmarks/performance/yhva1p1n",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yhva1p1n-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 4svv630v, ouwt1rni, xmdbvod2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 674.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 674.3 25.8 20.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yhva1p1n",
   "source_url": "https://omlx.ai/benchmarks/performance/yhva1p1n",
   "tg_tokens": null,
   "tps": 25.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yhva1p1n-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: 4svv630v, ouwt1rni, xmdbvod2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 536.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 536.0 20.8 24.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yhva1p1n",
   "source_url": "https://omlx.ai/benchmarks/performance/yhva1p1n",
   "tg_tokens": null,
   "tps": 20.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qj9t0v6m-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: osxs1ueq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 930.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 930.5 33.5 16.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qj9t0v6m",
   "source_url": "https://omlx.ai/benchmarks/performance/qj9t0v6m",
   "tg_tokens": null,
   "tps": 33.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qj9t0v6m-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: osxs1ueq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1010.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,010 32.9 17.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qj9t0v6m",
   "source_url": "https://omlx.ai/benchmarks/performance/qj9t0v6m",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qj9t0v6m-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: osxs1ueq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 860.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 860.2 31.0 17.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qj9t0v6m",
   "source_url": "https://omlx.ai/benchmarks/performance/qj9t0v6m",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kib5rvtf-ctx1024",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "ET-O5.5-G6A-G4.7-DSV4-K3-RLOO MTP DFlash2 build; oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: w7czvsw3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1075.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-ET-O5.5-G6A-G4.7-DSV4-K3-RLOO-MTP-DFlash2-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,075 52.4 18.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kib5rvtf",
   "source_url": "https://omlx.ai/benchmarks/performance/kib5rvtf",
   "tg_tokens": null,
   "tps": 52.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-kib5rvtf-ctx4096",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "ET-O5.5-G6A-G4.7-DSV4-K3-RLOO MTP DFlash2 build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: w7czvsw3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1136.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-ET-O5.5-G6A-G4.7-DSV4-K3-RLOO-MTP-DFlash2-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,136 57.6 19.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kib5rvtf",
   "source_url": "https://omlx.ai/benchmarks/performance/kib5rvtf",
   "tg_tokens": null,
   "tps": 57.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ew4fkg22-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4Bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 106.9 20.7 19.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ew4fkg22",
   "source_url": "https://omlx.ai/benchmarks/performance/ew4fkg22",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ew4fkg22-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 108.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4Bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 108.9 19.2 20.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ew4fkg22",
   "source_url": "https://omlx.ai/benchmarks/performance/ew4fkg22",
   "tg_tokens": null,
   "tps": 19.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-vun3p57x-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 37.7 GB; TTFT 352276 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 93.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-8Bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 8bit; 32k tokens 93.0 PP tok/s 14.6 TG tok/s 352276 TTFT (ms) 37.7 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vun3p57x",
   "source_url": "https://omlx.ai/benchmarks/performance/vun3p57x",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-83xxe6iv-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 34.2 GB at this context; duplicate traces of the same session: 5bg8w7qx; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-8Bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 8bit; 1k 105.8 15.9 34.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/83xxe6iv",
   "source_url": "https://omlx.ai/benchmarks/performance/83xxe6iv",
   "tg_tokens": null,
   "tps": 15.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-83xxe6iv-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: 5bg8w7qx; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-8Bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 8bit; 4k 106.0 14.6 35.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/83xxe6iv",
   "source_url": "https://omlx.ai/benchmarks/performance/83xxe6iv",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tkuk09m9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: mf2r9uaw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1051.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,051 56.6 17.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tkuk09m9",
   "source_url": "https://omlx.ai/benchmarks/performance/tkuk09m9",
   "tg_tokens": null,
   "tps": 56.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tkuk09m9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: mf2r9uaw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1067.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,067 61.8 19.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tkuk09m9",
   "source_url": "https://omlx.ai/benchmarks/performance/tkuk09m9",
   "tg_tokens": null,
   "tps": 61.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-uxrfn0y7-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: 1pf35w7t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 935.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 935.4 63.2 70.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uxrfn0y7",
   "source_url": "https://omlx.ai/benchmarks/performance/uxrfn0y7",
   "tg_tokens": null,
   "tps": 63.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-uxrfn0y7-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: 1pf35w7t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1881.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,881 68.5 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uxrfn0y7",
   "source_url": "https://omlx.ai/benchmarks/performance/uxrfn0y7",
   "tg_tokens": null,
   "tps": 68.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-uiqhv1pt-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.8 GB at this context; duplicate traces of the same session: l6lj0gve, qh7gnvfr, r4nfqtfl; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 466.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 466.3 31.0 173.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uiqhv1pt",
   "source_url": "https://omlx.ai/benchmarks/performance/uiqhv1pt",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-uiqhv1pt-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.5 GB at this context; duplicate traces of the same session: l6lj0gve, qh7gnvfr, r4nfqtfl; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 578.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 578.2 27.1 177.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uiqhv1pt",
   "source_url": "https://omlx.ai/benchmarks/performance/uiqhv1pt",
   "tg_tokens": null,
   "tps": 27.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-uiqhv1pt-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.6 GB at this context; duplicate traces of the same session: l6lj0gve, qh7gnvfr, r4nfqtfl; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 533.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 533.9 26.9 177.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uiqhv1pt",
   "source_url": "https://omlx.ai/benchmarks/performance/uiqhv1pt",
   "tg_tokens": null,
   "tps": 26.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-uiqhv1pt-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.7 GB at this context; duplicate traces of the same session: l6lj0gve, qh7gnvfr, r4nfqtfl; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 524.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 524.6 19.8 177.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uiqhv1pt",
   "source_url": "https://omlx.ai/benchmarks/performance/uiqhv1pt",
   "tg_tokens": null,
   "tps": 19.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-uiqhv1pt-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.9 GB at this context; duplicate traces of the same session: l6lj0gve, qh7gnvfr, r4nfqtfl; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 516.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 516.4 26.4 177.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uiqhv1pt",
   "source_url": "https://omlx.ai/benchmarks/performance/uiqhv1pt",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-uiqhv1pt-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.3 GB at this context; duplicate traces of the same session: l6lj0gve, qh7gnvfr, r4nfqtfl; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 504.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 504.8 25.4 178.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uiqhv1pt",
   "source_url": "https://omlx.ai/benchmarks/performance/uiqhv1pt",
   "tg_tokens": null,
   "tps": 25.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-uiqhv1pt-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 179.8 GB at this context; duplicate traces of the same session: l6lj0gve, qh7gnvfr, r4nfqtfl; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 259.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 259.1 7.1 179.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uiqhv1pt",
   "source_url": "https://omlx.ai/benchmarks/performance/uiqhv1pt",
   "tg_tokens": null,
   "tps": 7.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-4yszwcs2-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: pnitoaz1; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 233.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 233.8 21.3 5.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4yszwcs2",
   "source_url": "https://omlx.ai/benchmarks/performance/4yszwcs2",
   "tg_tokens": null,
   "tps": 21.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-4yszwcs2-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: pnitoaz1; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 240.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 240.6 20.9 6.2 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4yszwcs2",
   "source_url": "https://omlx.ai/benchmarks/performance/4yszwcs2",
   "tg_tokens": null,
   "tps": 20.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pfqhh5jh-ctx1024",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "ET-O5.5-G6A-G4.7-DSV4-K3-RLOO MTP DFlash2 build; oMLX community benchmark trace; peak memory 16.8 GB at this context; duplicate traces of the same session: c1nj5gww; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 918.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-ET-O5.5-G6A-G4.7-DSV4-K3-RLOO-MTP-DFlash2-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 918.6 31.3 16.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pfqhh5jh",
   "source_url": "https://omlx.ai/benchmarks/performance/pfqhh5jh",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pfqhh5jh-ctx4096",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "ET-O5.5-G6A-G4.7-DSV4-K3-RLOO MTP DFlash2 build; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: c1nj5gww; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 987.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-ET-O5.5-G6A-G4.7-DSV4-K3-RLOO-MTP-DFlash2-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 987.2 30.9 17.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pfqhh5jh",
   "source_url": "https://omlx.ai/benchmarks/performance/pfqhh5jh",
   "tg_tokens": null,
   "tps": 30.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-j910yjne-ctx1024",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "ET-O5.5-G6A-G4.7-DSV4-K3-RLOO MTP DFlash2 build; oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: 1to2vabe; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 892.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-ET-O5.5-G6A-G4.7-DSV4-K3-RLOO-MTP-DFlash2-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 892.6 53.9 17.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j910yjne",
   "source_url": "https://omlx.ai/benchmarks/performance/j910yjne",
   "tg_tokens": null,
   "tps": 53.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-j910yjne-ctx4096",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "ET-O5.5-G6A-G4.7-DSV4-K3-RLOO MTP DFlash2 build; oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: 1to2vabe; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 965.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-ET-O5.5-G6A-G4.7-DSV4-K3-RLOO-MTP-DFlash2-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 965.1 52.1 18.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j910yjne",
   "source_url": "https://omlx.ai/benchmarks/performance/j910yjne",
   "tg_tokens": null,
   "tps": 52.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zq1ty176-ctx1024",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "ET-O5.5-G6A-G4.7-DSV4-K3-RLOO MTP DFlash2 build; oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: 2zi96ihy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 699.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-ET-O5.5-G6A-G4.7-DSV4-K3-RLOO-MTP-DFlash2-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 699.5 50.7 17.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zq1ty176",
   "source_url": "https://omlx.ai/benchmarks/performance/zq1ty176",
   "tg_tokens": null,
   "tps": 50.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-zq1ty176-ctx4096",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "ET-O5.5-G6A-G4.7-DSV4-K3-RLOO MTP DFlash2 build; oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: 2zi96ihy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 724.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-ET-O5.5-G6A-G4.7-DSV4-K3-RLOO-MTP-DFlash2-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 724.5 52.3 18.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zq1ty176",
   "source_url": "https://omlx.ai/benchmarks/performance/zq1ty176",
   "tg_tokens": null,
   "tps": 52.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fzm2l2us-ctx1024",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "ET-O5.5-G6A-G4.7-DSV4-K3-RLOO MTP DFlash2 build; oMLX community benchmark trace; peak memory 16.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 865.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-ET-O5.5-G6A-G4.7-DSV4-K3-RLOO-MTP-DFlash2-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 865.9 30.1 16.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fzm2l2us",
   "source_url": "https://omlx.ai/benchmarks/performance/fzm2l2us",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fzm2l2us-ctx4096",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "ET-O5.5-G6A-G4.7-DSV4-K3-RLOO MTP DFlash2 build; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 980.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-ET-O5.5-G6A-G4.7-DSV4-K3-RLOO-MTP-DFlash2-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 980.6 30.8 17.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fzm2l2us",
   "source_url": "https://omlx.ai/benchmarks/performance/fzm2l2us",
   "tg_tokens": null,
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1whravdu-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 872.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 872.2 32.9 33.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1whravdu",
   "source_url": "https://omlx.ai/benchmarks/performance/1whravdu",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1whravdu-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 902.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 902.6 16.5 34.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1whravdu",
   "source_url": "https://omlx.ai/benchmarks/performance/1whravdu",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-f6cuyxw3-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4Bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 106.9 20.7 19.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f6cuyxw3",
   "source_url": "https://omlx.ai/benchmarks/performance/f6cuyxw3",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-f6cuyxw3-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 108.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4Bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 108.9 19.2 20.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f6cuyxw3",
   "source_url": "https://omlx.ai/benchmarks/performance/f6cuyxw3",
   "tg_tokens": null,
   "tps": 19.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-2liccfuu-ctx1024",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: wf5hgln5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1357.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 1,357 73.9 6.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2liccfuu",
   "source_url": "https://omlx.ai/benchmarks/performance/2liccfuu",
   "tg_tokens": null,
   "tps": 73.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-2liccfuu-ctx4096",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: wf5hgln5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1441.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 1,441 81.0 6.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2liccfuu",
   "source_url": "https://omlx.ai/benchmarks/performance/2liccfuu",
   "tg_tokens": null,
   "tps": 81.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-glb0d4w5-ctx65536",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "384 experts oQ4e + 128 experts oQ2 mixed MoE quant; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 39.0 GB; TTFT 346307 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 189.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-384e-oQ4e-128e-oQ2-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k tokens 189.2 PP tok/s 6.3 TG tok/s 346307 TTFT (ms) 39.0 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/glb0d4w5",
   "source_url": "https://omlx.ai/benchmarks/performance/glb0d4w5",
   "tg_tokens": null,
   "tps": 6.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-bz1lvy50-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: u20a8kdc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 472.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 472.7 64.9 70.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bz1lvy50",
   "source_url": "https://omlx.ai/benchmarks/performance/bz1lvy50",
   "tg_tokens": null,
   "tps": 64.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-bz1lvy50-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: u20a8kdc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 114.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 114.8 58.3 71.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bz1lvy50",
   "source_url": "https://omlx.ai/benchmarks/performance/bz1lvy50",
   "tg_tokens": null,
   "tps": 58.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-j9l8cjm3-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.0 GB at this context; duplicate traces of the same session: btd653fw; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 705.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k 705.8 46.9 14.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j9l8cjm3",
   "source_url": "https://omlx.ai/benchmarks/performance/j9l8cjm3",
   "tg_tokens": null,
   "tps": 46.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-j9l8cjm3-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: btd653fw; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1067.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 1,067 46.2 14.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j9l8cjm3",
   "source_url": "https://omlx.ai/benchmarks/performance/j9l8cjm3",
   "tg_tokens": null,
   "tps": 46.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-q3vz982j-ctx1024",
   "model": "Qwen2.5-Coder-14B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 8.5 GB at this context; duplicate traces of the same session: 8ygo7q19, w697nq98; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 781.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-14B-Instruct-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 781.1 32.4 8.5 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q3vz982j",
   "source_url": "https://omlx.ai/benchmarks/performance/q3vz982j",
   "tg_tokens": null,
   "tps": 32.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-q3vz982j-ctx4096",
   "model": "Qwen2.5-Coder-14B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 9.1 GB at this context; duplicate traces of the same session: 8ygo7q19, w697nq98; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 779.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-14B-Instruct-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 779.0 30.2 9.1 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q3vz982j",
   "source_url": "https://omlx.ai/benchmarks/performance/q3vz982j",
   "tg_tokens": null,
   "tps": 30.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-q3vz982j-ctx8192",
   "model": "Qwen2.5-Coder-14B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 9.9 GB at this context; duplicate traces of the same session: 8ygo7q19, w697nq98; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 684.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-14B-Instruct-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 8k 684.9 27.2 9.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q3vz982j",
   "source_url": "https://omlx.ai/benchmarks/performance/q3vz982j",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-37kh91c3-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: ah3x7092; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2473.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 2,473 104.9 19.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/37kh91c3",
   "source_url": "https://omlx.ai/benchmarks/performance/37kh91c3",
   "tg_tokens": null,
   "tps": 104.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-37kh91c3-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: ah3x7092; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2948.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,948 102.2 20.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/37kh91c3",
   "source_url": "https://omlx.ai/benchmarks/performance/37kh91c3",
   "tg_tokens": null,
   "tps": 102.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-46y14ysf-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 19.2 GB; TTFT 28865 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 567.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k tokens 567.6 PP tok/s 15.1 TG tok/s 28865 TTFT (ms) 19.2 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/46y14ysf",
   "source_url": "https://omlx.ai/benchmarks/performance/46y14ysf",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-amwb6buv-ctx1024",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "ET-O5.5-G6A-G4.7-DSV4-K3-RLOO MTP DFlash2 build; oMLX community benchmark trace; peak memory 16.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 865.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-ET-O5.5-G6A-G4.7-DSV4-K3-RLOO-MTP-DFlash2-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 865.9 30.1 16.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/amwb6buv",
   "source_url": "https://omlx.ai/benchmarks/performance/amwb6buv",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-amwb6buv-ctx4096",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "ET-O5.5-G6A-G4.7-DSV4-K3-RLOO MTP DFlash2 build; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 980.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-ET-O5.5-G6A-G4.7-DSV4-K3-RLOO-MTP-DFlash2-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 980.6 30.8 17.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/amwb6buv",
   "source_url": "https://omlx.ai/benchmarks/performance/amwb6buv",
   "tg_tokens": null,
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-v5bf67wv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 872.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 872.2 32.9 33.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v5bf67wv",
   "source_url": "https://omlx.ai/benchmarks/performance/v5bf67wv",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-v5bf67wv-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 902.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 902.6 16.5 34.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v5bf67wv",
   "source_url": "https://omlx.ai/benchmarks/performance/v5bf67wv",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-gmbvf8u7-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: p9woe32w; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 386.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 386.5 20.4 23.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gmbvf8u7",
   "source_url": "https://omlx.ai/benchmarks/performance/gmbvf8u7",
   "tg_tokens": null,
   "tps": 20.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-gmbvf8u7-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: p9woe32w; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 386.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 386.8 23.1 24.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gmbvf8u7",
   "source_url": "https://omlx.ai/benchmarks/performance/gmbvf8u7",
   "tg_tokens": null,
   "tps": 23.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0m67ae2a-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace, single run at 1k context; peak memory 19.9 GB; TTFT 2645 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 387.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k tokens 387.2 PP tok/s 14.5 TG tok/s 2645 TTFT (ms) 19.9 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0m67ae2a",
   "source_url": "https://omlx.ai/benchmarks/performance/0m67ae2a",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-7jm38492-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: ytyok4jf; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 413.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 413.5 18.1 15.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7jm38492",
   "source_url": "https://omlx.ai/benchmarks/performance/7jm38492",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-7jm38492-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: ytyok4jf; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 426.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 426.9 17.9 17.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7jm38492",
   "source_url": "https://omlx.ai/benchmarks/performance/7jm38492",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-jfg7tzsb-ctx1024",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "MLX; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: ipd0im0y; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 277.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 277.6 19.1 21.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jfg7tzsb",
   "source_url": "https://omlx.ai/benchmarks/performance/jfg7tzsb",
   "tg_tokens": null,
   "tps": 19.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-jfg7tzsb-ctx4096",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "MLX; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: ipd0im0y; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 278.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 278.8 18.6 22.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jfg7tzsb",
   "source_url": "https://omlx.ai/benchmarks/performance/jfg7tzsb",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-exl4z6ow-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1347.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,347 68.5 19.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/exl4z6ow",
   "source_url": "https://omlx.ai/benchmarks/performance/exl4z6ow",
   "tg_tokens": null,
   "tps": 68.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-exl4z6ow-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1166.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,166 79.2 19.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/exl4z6ow",
   "source_url": "https://omlx.ai/benchmarks/performance/exl4z6ow",
   "tg_tokens": null,
   "tps": 79.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-exl4z6ow-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 969.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 969.5 74.9 21.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/exl4z6ow",
   "source_url": "https://omlx.ai/benchmarks/performance/exl4z6ow",
   "tg_tokens": null,
   "tps": 74.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-exl4z6ow-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 779.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 779.5 44.3 23.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/exl4z6ow",
   "source_url": "https://omlx.ai/benchmarks/performance/exl4z6ow",
   "tg_tokens": null,
   "tps": 44.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-exl4z6ow-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 485.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 485.1 28.6 28.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/exl4z6ow",
   "source_url": "https://omlx.ai/benchmarks/performance/exl4z6ow",
   "tg_tokens": null,
   "tps": 28.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-exl4z6ow-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 369.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 369.6 13.6 70.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/exl4z6ow",
   "source_url": "https://omlx.ai/benchmarks/performance/exl4z6ow",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ntvqv678-ctx1024",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: dw2c1kbs; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1364.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 1,364 51.0 6.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ntvqv678",
   "source_url": "https://omlx.ai/benchmarks/performance/ntvqv678",
   "tg_tokens": null,
   "tps": 51.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ntvqv678-ctx4096",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: dw2c1kbs; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1438.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 1,438 75.5 6.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ntvqv678",
   "source_url": "https://omlx.ai/benchmarks/performance/ntvqv678",
   "tg_tokens": null,
   "tps": 75.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-5m5v1f6m-ctx32768",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "384 experts oQ4e + 128 experts oQ2 mixed MoE quant; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 37.8 GB; TTFT 159109 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-384e-oQ4e-128e-oQ2-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k tokens 205.9 PP tok/s 5.6 TG tok/s 159109 TTFT (ms) 37.8 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5m5v1f6m",
   "source_url": "https://omlx.ai/benchmarks/performance/5m5v1f6m",
   "tg_tokens": null,
   "tps": 5.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-k1bxbkso-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 101.5 GB; TTFT 12222 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1341.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k tokens 1,341 PP tok/s 43.0 TG tok/s 12222 TTFT (ms) 101.5 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k1bxbkso",
   "source_url": "https://omlx.ai/benchmarks/performance/k1bxbkso",
   "tg_tokens": null,
   "tps": 43.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-m5oo4l17-ctx1024",
   "model": "Ternary-Bonsai-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 9.1 GB at this context; duplicate traces of the same session: 1z53t6l2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.6,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-27B-mlx-2bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 2bit; 1k 131.6 25.4 9.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m5oo4l17",
   "source_url": "https://omlx.ai/benchmarks/performance/m5oo4l17",
   "tg_tokens": null,
   "tps": 25.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-m5oo4l17-ctx4096",
   "model": "Ternary-Bonsai-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 10.1 GB at this context; duplicate traces of the same session: 1z53t6l2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 135.8,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-27B-mlx-2bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 2bit; 4k 135.8 24.7 10.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m5oo4l17",
   "source_url": "https://omlx.ai/benchmarks/performance/m5oo4l17",
   "tg_tokens": null,
   "tps": 24.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-m9nemqei-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: du6xg8f8; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 333.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 333.3 39.5 15.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m9nemqei",
   "source_url": "https://omlx.ai/benchmarks/performance/m9nemqei",
   "tg_tokens": null,
   "tps": 39.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-m9nemqei-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: du6xg8f8; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 358.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 358.0 38.1 19.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m9nemqei",
   "source_url": "https://omlx.ai/benchmarks/performance/m9nemqei",
   "tg_tokens": null,
   "tps": 38.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-q7wt51th-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 329.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 329.3 50.4 17.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q7wt51th",
   "source_url": "https://omlx.ai/benchmarks/performance/q7wt51th",
   "tg_tokens": null,
   "tps": 50.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-q7wt51th-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 340.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 340.3 51.3 18.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q7wt51th",
   "source_url": "https://omlx.ai/benchmarks/performance/q7wt51th",
   "tg_tokens": null,
   "tps": 51.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-sgq8n9cq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 330.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 330.0 44.1 19.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sgq8n9cq",
   "source_url": "https://omlx.ai/benchmarks/performance/sgq8n9cq",
   "tg_tokens": null,
   "tps": 44.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-sgq8n9cq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 340.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 340.5 49.1 21.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sgq8n9cq",
   "source_url": "https://omlx.ai/benchmarks/performance/sgq8n9cq",
   "tg_tokens": null,
   "tps": 49.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-g6cqh8d3-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 17.0 GB; TTFT 22323 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 183.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k tokens 183.5 PP tok/s 9.9 TG tok/s 22323 TTFT (ms) 17.0 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g6cqh8d3",
   "source_url": "https://omlx.ai/benchmarks/performance/g6cqh8d3",
   "tg_tokens": null,
   "tps": 9.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-r0w10i8c-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: i4l1dyev, oqb8x4k5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 482.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 482.8 73.6 24.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r0w10i8c",
   "source_url": "https://omlx.ai/benchmarks/performance/r0w10i8c",
   "tg_tokens": null,
   "tps": 73.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-r0w10i8c-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: i4l1dyev, oqb8x4k5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 520.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 520.0 64.2 25.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r0w10i8c",
   "source_url": "https://omlx.ai/benchmarks/performance/r0w10i8c",
   "tg_tokens": null,
   "tps": 64.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-r0w10i8c-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.2 GB at this context; duplicate traces of the same session: i4l1dyev, oqb8x4k5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 516.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 516.3 59.8 26.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r0w10i8c",
   "source_url": "https://omlx.ai/benchmarks/performance/r0w10i8c",
   "tg_tokens": null,
   "tps": 59.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ptlpz2fa-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 21.3 GB; TTFT 2404 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 426.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k tokens 426.0 PP tok/s 55.3 TG tok/s 2404 TTFT (ms) 21.3 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ptlpz2fa",
   "source_url": "https://omlx.ai/benchmarks/performance/ptlpz2fa",
   "tg_tokens": null,
   "tps": 55.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bkf32q9b-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: kgyczbfk, uhwqxcbv, vm9n9ufe, xdpywo55; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1347.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,347 68.5 19.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bkf32q9b",
   "source_url": "https://omlx.ai/benchmarks/performance/bkf32q9b",
   "tg_tokens": null,
   "tps": 68.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bkf32q9b-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: kgyczbfk, uhwqxcbv, vm9n9ufe, xdpywo55; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1166.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,166 79.2 19.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bkf32q9b",
   "source_url": "https://omlx.ai/benchmarks/performance/bkf32q9b",
   "tg_tokens": null,
   "tps": 79.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bkf32q9b-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: kgyczbfk, uhwqxcbv, vm9n9ufe, xdpywo55; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 969.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 969.5 74.9 21.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bkf32q9b",
   "source_url": "https://omlx.ai/benchmarks/performance/bkf32q9b",
   "tg_tokens": null,
   "tps": 74.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bkf32q9b-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: kgyczbfk, uhwqxcbv, vm9n9ufe, xdpywo55; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 779.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 779.5 44.3 23.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bkf32q9b",
   "source_url": "https://omlx.ai/benchmarks/performance/bkf32q9b",
   "tg_tokens": null,
   "tps": 44.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bkf32q9b-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: kgyczbfk, uhwqxcbv, vm9n9ufe, xdpywo55; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 485.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 485.1 28.6 28.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bkf32q9b",
   "source_url": "https://omlx.ai/benchmarks/performance/bkf32q9b",
   "tg_tokens": null,
   "tps": 28.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bkf32q9b-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: kgyczbfk, uhwqxcbv, vm9n9ufe, xdpywo55; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 369.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 369.6 13.6 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bkf32q9b",
   "source_url": "https://omlx.ai/benchmarks/performance/bkf32q9b",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1u29oc1e-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: bp9d4wdf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 439.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 439.2 15.7 18.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1u29oc1e",
   "source_url": "https://omlx.ai/benchmarks/performance/1u29oc1e",
   "tg_tokens": null,
   "tps": 15.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1u29oc1e-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: bp9d4wdf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 365.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 365.7 14.0 19.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1u29oc1e",
   "source_url": "https://omlx.ai/benchmarks/performance/1u29oc1e",
   "tg_tokens": null,
   "tps": 14.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-yom003v7-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.7 GB at this context; duplicate traces of the same session: 3gpckn8w; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 316.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 316.4 48.8 33.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yom003v7",
   "source_url": "https://omlx.ai/benchmarks/performance/yom003v7",
   "tg_tokens": null,
   "tps": 48.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-yom003v7-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: 3gpckn8w; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 332.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 332.8 45.3 37.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yom003v7",
   "source_url": "https://omlx.ai/benchmarks/performance/yom003v7",
   "tg_tokens": null,
   "tps": 45.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ysip8as6-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.8 GB; TTFT 2129 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 481.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k tokens 481.1 PP tok/s 43.8 TG tok/s 2129 TTFT (ms) 70.8 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ysip8as6",
   "source_url": "https://omlx.ai/benchmarks/performance/ysip8as6",
   "tg_tokens": null,
   "tps": 43.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-bod7j594-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.7 GB; TTFT 1546 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 662.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k tokens 662.6 PP tok/s 68.6 TG tok/s 1546 TTFT (ms) 70.7 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bod7j594",
   "source_url": "https://omlx.ai/benchmarks/performance/bod7j594",
   "tg_tokens": null,
   "tps": 68.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-fk4v8d7d-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.3 GB at this context; duplicate traces of the same session: lmo0puug; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 556.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 556.7 54.4 69.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fk4v8d7d",
   "source_url": "https://omlx.ai/benchmarks/performance/fk4v8d7d",
   "tg_tokens": null,
   "tps": 54.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-fk4v8d7d-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.9 GB at this context; duplicate traces of the same session: lmo0puug; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 892.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 892.2 60.7 69.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fk4v8d7d",
   "source_url": "https://omlx.ai/benchmarks/performance/fk4v8d7d",
   "tg_tokens": null,
   "tps": 60.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qh5uh2ae-ctx4096",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: rq6gscs1, zi06xkwa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 877.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Muse-Glimmer-30B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 877.7 22.5 21.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qh5uh2ae",
   "source_url": "https://omlx.ai/benchmarks/performance/qh5uh2ae",
   "tg_tokens": null,
   "tps": 22.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qh5uh2ae-ctx16384",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: rq6gscs1, zi06xkwa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 785.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Muse-Glimmer-30B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 785.8 21.9 22.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qh5uh2ae",
   "source_url": "https://omlx.ai/benchmarks/performance/qh5uh2ae",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qh5uh2ae-ctx131072",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: rq6gscs1, zi06xkwa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 632.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Muse-Glimmer-30B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 128k 632.7 19.1 23.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qh5uh2ae",
   "source_url": "https://omlx.ai/benchmarks/performance/qh5uh2ae",
   "tg_tokens": null,
   "tps": 19.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-05rlmgp8-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: r38jnmdf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5785.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 5,785 181.3 22.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/05rlmgp8",
   "source_url": "https://omlx.ai/benchmarks/performance/05rlmgp8",
   "tg_tokens": null,
   "tps": 181.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-05rlmgp8-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: r38jnmdf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5820.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 5,820 194.6 22.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/05rlmgp8",
   "source_url": "https://omlx.ai/benchmarks/performance/05rlmgp8",
   "tg_tokens": null,
   "tps": 194.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-05rlmgp8-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: r38jnmdf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4361.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 4,361 192.2 22.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/05rlmgp8",
   "source_url": "https://omlx.ai/benchmarks/performance/05rlmgp8",
   "tg_tokens": null,
   "tps": 192.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-05rlmgp8-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: r38jnmdf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3013.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 3,013 146.1 23.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/05rlmgp8",
   "source_url": "https://omlx.ai/benchmarks/performance/05rlmgp8",
   "tg_tokens": null,
   "tps": 146.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-05rlmgp8-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: r38jnmdf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1847.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 1,847 110.1 24.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/05rlmgp8",
   "source_url": "https://omlx.ai/benchmarks/performance/05rlmgp8",
   "tg_tokens": null,
   "tps": 110.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-05rlmgp8-ctx199680",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: r38jnmdf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1171.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 1,171 53.4 26.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/05rlmgp8",
   "source_url": "https://omlx.ai/benchmarks/performance/05rlmgp8",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xz8dvqae-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 329.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 329.3 50.4 17.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xz8dvqae",
   "source_url": "https://omlx.ai/benchmarks/performance/xz8dvqae",
   "tg_tokens": null,
   "tps": 50.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xz8dvqae-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 340.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 340.3 51.3 18.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xz8dvqae",
   "source_url": "https://omlx.ai/benchmarks/performance/xz8dvqae",
   "tg_tokens": null,
   "tps": 51.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ieg6qghg-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 330.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 330.0 44.1 19.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ieg6qghg",
   "source_url": "https://omlx.ai/benchmarks/performance/ieg6qghg",
   "tg_tokens": null,
   "tps": 44.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ieg6qghg-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 340.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 340.5 49.1 21.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ieg6qghg",
   "source_url": "https://omlx.ai/benchmarks/performance/ieg6qghg",
   "tg_tokens": null,
   "tps": 49.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-yy779lw9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: 2xbd5d74; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 329.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 329.9 42.6 19.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yy779lw9",
   "source_url": "https://omlx.ai/benchmarks/performance/yy779lw9",
   "tg_tokens": null,
   "tps": 42.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-yy779lw9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 2xbd5d74; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 339.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 339.9 50.1 21.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yy779lw9",
   "source_url": "https://omlx.ai/benchmarks/performance/yy779lw9",
   "tg_tokens": null,
   "tps": 50.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-7mr3olwa-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: tf4pjfwb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 331.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 331.6 54.9 17.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7mr3olwa",
   "source_url": "https://omlx.ai/benchmarks/performance/7mr3olwa",
   "tg_tokens": null,
   "tps": 54.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-7mr3olwa-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: tf4pjfwb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 340.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 340.5 51.6 18.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7mr3olwa",
   "source_url": "https://omlx.ai/benchmarks/performance/7mr3olwa",
   "tg_tokens": null,
   "tps": 51.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ysfkf8j7-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: 8o6u8w6w; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 332.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 332.8 39.4 15.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ysfkf8j7",
   "source_url": "https://omlx.ai/benchmarks/performance/ysfkf8j7",
   "tg_tokens": null,
   "tps": 39.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ysfkf8j7-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 8o6u8w6w; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 357.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 357.2 38.1 19.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ysfkf8j7",
   "source_url": "https://omlx.ai/benchmarks/performance/ysfkf8j7",
   "tg_tokens": null,
   "tps": 38.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-f9ndazcs-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 4yd1emsg, odkrm0dr, qgzsy4ut; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4Bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 106.1 15.5 16.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f9ndazcs",
   "source_url": "https://omlx.ai/benchmarks/performance/f9ndazcs",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-f9ndazcs-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 4yd1emsg, odkrm0dr, qgzsy4ut; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 107.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4Bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 107.7 15.1 17.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f9ndazcs",
   "source_url": "https://omlx.ai/benchmarks/performance/f9ndazcs",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-f9ndazcs-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: 4yd1emsg, odkrm0dr, qgzsy4ut; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 107.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4Bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 107.6 14.8 17.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f9ndazcs",
   "source_url": "https://omlx.ai/benchmarks/performance/f9ndazcs",
   "tg_tokens": null,
   "tps": 14.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-f9ndazcs-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: 4yd1emsg, odkrm0dr, qgzsy4ut; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4Bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 103.9 14.0 18.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f9ndazcs",
   "source_url": "https://omlx.ai/benchmarks/performance/f9ndazcs",
   "tg_tokens": null,
   "tps": 14.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ylri1o4f-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 33.7 GB; TTFT 1297 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 789.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k tokens 789.4 PP tok/s 47.9 TG tok/s 1297 TTFT (ms) 33.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ylri1o4f",
   "source_url": "https://omlx.ai/benchmarks/performance/ylri1o4f",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gyn7bkxz-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 33.7 GB; TTFT 1296 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 790.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k tokens 790.2 PP tok/s 43.9 TG tok/s 1296 TTFT (ms) 33.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gyn7bkxz",
   "source_url": "https://omlx.ai/benchmarks/performance/gyn7bkxz",
   "tg_tokens": null,
   "tps": 43.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1ju23anq-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP320",
   "notes": "MTP Vision build; oMLX community benchmark trace; peak memory 67.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1441.0,
   "provenance": "community",
   "quant": "oQ3e-DWQ",
   "quote": "Qwen3.8-Flash-Next-REAP320-oQ3e-DWQ-MTP-Vision-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 1k 1,441 36.8 67.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ju23anq",
   "source_url": "https://omlx.ai/benchmarks/performance/1ju23anq",
   "tg_tokens": null,
   "tps": 36.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1ju23anq-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP320",
   "notes": "MTP Vision build; oMLX community benchmark trace; peak memory 67.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1733.0,
   "provenance": "community",
   "quant": "oQ3e-DWQ",
   "quote": "Qwen3.8-Flash-Next-REAP320-oQ3e-DWQ-MTP-Vision-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 4k 1,733 40.5 67.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ju23anq",
   "source_url": "https://omlx.ai/benchmarks/performance/1ju23anq",
   "tg_tokens": null,
   "tps": 40.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-bea171ka-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1116.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 4k 1,116 53.7 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bea171ka",
   "source_url": "https://omlx.ai/benchmarks/performance/bea171ka",
   "tg_tokens": null,
   "tps": 53.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-bea171ka-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 899.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 16k 899.3 50.8 22.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bea171ka",
   "source_url": "https://omlx.ai/benchmarks/performance/bea171ka",
   "tg_tokens": null,
   "tps": 50.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-bea171ka-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 481.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 64k 481.3 39.9 24.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bea171ka",
   "source_url": "https://omlx.ai/benchmarks/performance/bea171ka",
   "tg_tokens": null,
   "tps": 39.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-1rxt7uxv-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 20.0 GB; TTFT 1381 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2966.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k tokens 2,966 PP tok/s 103.7 TG tok/s 1381 TTFT (ms) 20.0 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1rxt7uxv",
   "source_url": "https://omlx.ai/benchmarks/performance/1rxt7uxv",
   "tg_tokens": null,
   "tps": 103.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-c7jrmf6u-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 554uzz1p, v58z3b5z, xo40k684; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5785.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 5,785 181.3 22.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c7jrmf6u",
   "source_url": "https://omlx.ai/benchmarks/performance/c7jrmf6u",
   "tg_tokens": null,
   "tps": 181.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-c7jrmf6u-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 554uzz1p, v58z3b5z, xo40k684; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5820.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 5,820 194.6 22.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c7jrmf6u",
   "source_url": "https://omlx.ai/benchmarks/performance/c7jrmf6u",
   "tg_tokens": null,
   "tps": 194.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-c7jrmf6u-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 554uzz1p, v58z3b5z, xo40k684; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4361.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 4,361 192.2 22.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c7jrmf6u",
   "source_url": "https://omlx.ai/benchmarks/performance/c7jrmf6u",
   "tg_tokens": null,
   "tps": 192.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-c7jrmf6u-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 554uzz1p, v58z3b5z, xo40k684; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3013.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 3,013 146.1 23.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c7jrmf6u",
   "source_url": "https://omlx.ai/benchmarks/performance/c7jrmf6u",
   "tg_tokens": null,
   "tps": 146.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-c7jrmf6u-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: 554uzz1p, v58z3b5z, xo40k684; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1847.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 1,847 110.1 24.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c7jrmf6u",
   "source_url": "https://omlx.ai/benchmarks/performance/c7jrmf6u",
   "tg_tokens": null,
   "tps": 110.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-c7jrmf6u-ctx199680",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: 554uzz1p, v58z3b5z, xo40k684; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1171.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 1,171 53.4 26.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c7jrmf6u",
   "source_url": "https://omlx.ai/benchmarks/performance/c7jrmf6u",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-6ka5lzkg-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: etgw3cad; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 413.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 413.3 18.2 15.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ka5lzkg",
   "source_url": "https://omlx.ai/benchmarks/performance/6ka5lzkg",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-6ka5lzkg-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: etgw3cad; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 398.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 398.8 17.9 17.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ka5lzkg",
   "source_url": "https://omlx.ai/benchmarks/performance/6ka5lzkg",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-yvsnf7v6-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 35i8c11h, d9pj3sd3, ru4czif4; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-6Bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 1k 103.4 11.2 22.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yvsnf7v6",
   "source_url": "https://omlx.ai/benchmarks/performance/yvsnf7v6",
   "tg_tokens": null,
   "tps": 11.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-yvsnf7v6-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 35i8c11h, d9pj3sd3, ru4czif4; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-6Bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 4k 106.0 10.9 23.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yvsnf7v6",
   "source_url": "https://omlx.ai/benchmarks/performance/yvsnf7v6",
   "tg_tokens": null,
   "tps": 10.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-yvsnf7v6-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: 35i8c11h, d9pj3sd3, ru4czif4; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-6Bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 8k 105.7 10.6 23.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yvsnf7v6",
   "source_url": "https://omlx.ai/benchmarks/performance/yvsnf7v6",
   "tg_tokens": null,
   "tps": 10.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-yvsnf7v6-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: 35i8c11h, d9pj3sd3, ru4czif4; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-6Bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 16k 102.4 10.4 24.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yvsnf7v6",
   "source_url": "https://omlx.ai/benchmarks/performance/yvsnf7v6",
   "tg_tokens": null,
   "tps": 10.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-6urpgrp8-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-8Bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 8bit; 1k 104.4 8.7 28.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6urpgrp8",
   "source_url": "https://omlx.ai/benchmarks/performance/6urpgrp8",
   "tg_tokens": null,
   "tps": 8.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-6urpgrp8-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-8Bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 8bit; 4k 105.5 8.6 29.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6urpgrp8",
   "source_url": "https://omlx.ai/benchmarks/performance/6urpgrp8",
   "tg_tokens": null,
   "tps": 8.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-6urpgrp8-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-8Bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 8bit; 8k 103.8 8.4 29.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6urpgrp8",
   "source_url": "https://omlx.ai/benchmarks/performance/6urpgrp8",
   "tg_tokens": null,
   "tps": 8.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-6urpgrp8-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 30.7 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 100.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-8Bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 8bit; 16k 100.1 8.1 30.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6urpgrp8",
   "source_url": "https://omlx.ai/benchmarks/performance/6urpgrp8",
   "tg_tokens": null,
   "tps": 8.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gqc2hmmp-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 43.3 GB; TTFT 1301 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 786.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k tokens 786.9 PP tok/s 40.8 TG tok/s 1301 TTFT (ms) 43.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gqc2hmmp",
   "source_url": "https://omlx.ai/benchmarks/performance/gqc2hmmp",
   "tg_tokens": null,
   "tps": 40.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wjeodk0s-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP320",
   "notes": "MTP Vision build; oMLX community benchmark trace; peak memory 67.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1441.0,
   "provenance": "community",
   "quant": "oQ3e-DWQ",
   "quote": "Qwen3.8-Flash-Next-REAP320-oQ3e-DWQ-MTP-Vision-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 1k 1,441 36.8 67.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wjeodk0s",
   "source_url": "https://omlx.ai/benchmarks/performance/wjeodk0s",
   "tg_tokens": null,
   "tps": 36.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wjeodk0s-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP320",
   "notes": "MTP Vision build; oMLX community benchmark trace; peak memory 67.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1733.0,
   "provenance": "community",
   "quant": "oQ3e-DWQ",
   "quote": "Qwen3.8-Flash-Next-REAP320-oQ3e-DWQ-MTP-Vision-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 4k 1,733 40.5 67.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wjeodk0s",
   "source_url": "https://omlx.ai/benchmarks/performance/wjeodk0s",
   "tg_tokens": null,
   "tps": 40.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-jexguzzu-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.2 GB at this context; duplicate traces of the same session: 9nuf37bg, ept6d0lu, kau86rew, wzhho585; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1376.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,376 81.2 101.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jexguzzu",
   "source_url": "https://omlx.ai/benchmarks/performance/jexguzzu",
   "tg_tokens": null,
   "tps": 81.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-jexguzzu-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: 9nuf37bg, ept6d0lu, kau86rew, wzhho585; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1385.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 1,385 100.9 101.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jexguzzu",
   "source_url": "https://omlx.ai/benchmarks/performance/jexguzzu",
   "tg_tokens": null,
   "tps": 100.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jexguzzu-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.4 GB at this context; duplicate traces of the same session: 9nuf37bg, ept6d0lu, kau86rew, wzhho585; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1350.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 1,350 88.9 104.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jexguzzu",
   "source_url": "https://omlx.ai/benchmarks/performance/jexguzzu",
   "tg_tokens": null,
   "tps": 88.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jexguzzu-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.1 GB at this context; duplicate traces of the same session: 9nuf37bg, ept6d0lu, kau86rew, wzhho585; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1327.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 1,327 65.1 109.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jexguzzu",
   "source_url": "https://omlx.ai/benchmarks/performance/jexguzzu",
   "tg_tokens": null,
   "tps": 65.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jexguzzu-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.9 GB at this context; duplicate traces of the same session: 9nuf37bg, ept6d0lu, kau86rew, wzhho585; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1308.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 195k 1,308 61.7 109.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jexguzzu",
   "source_url": "https://omlx.ai/benchmarks/performance/jexguzzu",
   "tg_tokens": null,
   "tps": 61.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-khx3etrc-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 25.5 GB; TTFT 362088 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 362.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 128k tokens 362.0 PP tok/s 36.3 TG tok/s 362088 TTFT (ms) 25.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/khx3etrc",
   "source_url": "https://omlx.ai/benchmarks/performance/khx3etrc",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-k1oz7kcr-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: vz5lxebv; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1116.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 4k 1,116 53.7 21.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k1oz7kcr",
   "source_url": "https://omlx.ai/benchmarks/performance/k1oz7kcr",
   "tg_tokens": null,
   "tps": 53.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-k1oz7kcr-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: vz5lxebv; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 899.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 16k 899.3 50.8 22.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k1oz7kcr",
   "source_url": "https://omlx.ai/benchmarks/performance/k1oz7kcr",
   "tg_tokens": null,
   "tps": 50.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-k1oz7kcr-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: vz5lxebv; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 481.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 64k 481.3 39.9 24.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k1oz7kcr",
   "source_url": "https://omlx.ai/benchmarks/performance/k1oz7kcr",
   "tg_tokens": null,
   "tps": 39.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-yalnd5s7-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: hg84c3gz; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2407.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 2,407 108.0 19.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yalnd5s7",
   "source_url": "https://omlx.ai/benchmarks/performance/yalnd5s7",
   "tg_tokens": null,
   "tps": 108.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-yalnd5s7-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: hg84c3gz; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2960.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,960 103.8 20.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yalnd5s7",
   "source_url": "https://omlx.ai/benchmarks/performance/yalnd5s7",
   "tg_tokens": null,
   "tps": 103.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-zgbul6b7-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 15.8 GB; TTFT 2488 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 411.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k tokens 411.5 PP tok/s 18.2 TG tok/s 2488 TTFT (ms) 15.8 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zgbul6b7",
   "source_url": "https://omlx.ai/benchmarks/performance/zgbul6b7",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-5rold5fm-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: pzbrp9rp, tpvndckk; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-8Bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 8bit; 1k 104.4 8.7 28.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5rold5fm",
   "source_url": "https://omlx.ai/benchmarks/performance/5rold5fm",
   "tg_tokens": null,
   "tps": 8.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-5rold5fm-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: pzbrp9rp, tpvndckk; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-8Bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 8bit; 4k 105.5 8.6 29.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5rold5fm",
   "source_url": "https://omlx.ai/benchmarks/performance/5rold5fm",
   "tg_tokens": null,
   "tps": 8.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-5rold5fm-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: pzbrp9rp, tpvndckk; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-8Bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 8bit; 8k 103.8 8.4 29.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5rold5fm",
   "source_url": "https://omlx.ai/benchmarks/performance/5rold5fm",
   "tg_tokens": null,
   "tps": 8.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-5rold5fm-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 30.7 GB at this context; duplicate traces of the same session: pzbrp9rp, tpvndckk; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 100.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-8Bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 8bit; 16k 100.1 8.1 30.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5rold5fm",
   "source_url": "https://omlx.ai/benchmarks/performance/5rold5fm",
   "tg_tokens": null,
   "tps": 8.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-7vp51bkh-ctx1024",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "MLX; oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: ucbb8ew9; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 279.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 279.4 19.6 21.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7vp51bkh",
   "source_url": "https://omlx.ai/benchmarks/performance/7vp51bkh",
   "tg_tokens": null,
   "tps": 19.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-7vp51bkh-ctx4096",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "MLX; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: ucbb8ew9; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 288.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 288.1 20.1 22.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7vp51bkh",
   "source_url": "https://omlx.ai/benchmarks/performance/7vp51bkh",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-yqpy9vbh-ctx1024",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "MLX; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: mner2s18; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 133.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 133.6 7.6 19.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yqpy9vbh",
   "source_url": "https://omlx.ai/benchmarks/performance/yqpy9vbh",
   "tg_tokens": null,
   "tps": 7.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-yqpy9vbh-ctx4096",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "MLX; oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: mner2s18; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 152.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 152.8 7.5 21.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yqpy9vbh",
   "source_url": "https://omlx.ai/benchmarks/performance/yqpy9vbh",
   "tg_tokens": null,
   "tps": 7.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-iocf0urd-ctx32768",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "384 experts oQ4e + 128 experts oQ2 mixed MoE quant; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 35.3 GB; TTFT 178157 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 183.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-384e-oQ4e-128e-oQ2-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k tokens 183.9 PP tok/s 6.0 TG tok/s 178157 TTFT (ms) 35.3 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iocf0urd",
   "source_url": "https://omlx.ai/benchmarks/performance/iocf0urd",
   "tg_tokens": null,
   "tps": 6.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-tihhlv4s-ctx1024",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 208.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 208.0 31.6 34.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tihhlv4s",
   "source_url": "https://omlx.ai/benchmarks/performance/tihhlv4s",
   "tg_tokens": null,
   "tps": 31.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-tihhlv4s-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 35.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 185.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 185.1 28.4 35.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tihhlv4s",
   "source_url": "https://omlx.ai/benchmarks/performance/tihhlv4s",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-vqrtssfx-ctx1024",
   "model": "gemma-4-E4B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: 14umkhxu, ogzibzo1, yot5ldfz; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2763.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-qat-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 2,763 64.5 6.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vqrtssfx",
   "source_url": "https://omlx.ai/benchmarks/performance/vqrtssfx",
   "tg_tokens": null,
   "tps": 64.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-vqrtssfx-ctx4096",
   "model": "gemma-4-E4B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: 14umkhxu, ogzibzo1, yot5ldfz; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3448.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-qat-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 3,448 63.0 6.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vqrtssfx",
   "source_url": "https://omlx.ai/benchmarks/performance/vqrtssfx",
   "tg_tokens": null,
   "tps": 63.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-vqrtssfx-ctx8192",
   "model": "gemma-4-E4B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: 14umkhxu, ogzibzo1, yot5ldfz; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3420.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-qat-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 3,420 59.8 6.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vqrtssfx",
   "source_url": "https://omlx.ai/benchmarks/performance/vqrtssfx",
   "tg_tokens": null,
   "tps": 59.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-vqrtssfx-ctx16384",
   "model": "gemma-4-E4B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: 14umkhxu, ogzibzo1, yot5ldfz; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3131.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-qat-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 3,131 55.0 7.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vqrtssfx",
   "source_url": "https://omlx.ai/benchmarks/performance/vqrtssfx",
   "tg_tokens": null,
   "tps": 55.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-vqrtssfx-ctx32768",
   "model": "gemma-4-E4B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 7.5 GB at this context; duplicate traces of the same session: 14umkhxu, ogzibzo1, yot5ldfz; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3035.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-qat-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 3,035 47.9 7.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vqrtssfx",
   "source_url": "https://omlx.ai/benchmarks/performance/vqrtssfx",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-vqrtssfx-ctx65536",
   "model": "gemma-4-E4B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 8.5 GB at this context; duplicate traces of the same session: 14umkhxu, ogzibzo1, yot5ldfz; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2550.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-qat-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 64k 2,550 37.4 8.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vqrtssfx",
   "source_url": "https://omlx.ai/benchmarks/performance/vqrtssfx",
   "tg_tokens": null,
   "tps": 37.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-4ty3l7hw-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MLX build; oMLX community benchmark trace, single run at 16k context; peak memory 213.7 GB; TTFT 42116 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 389.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "glm-5.3-flash-4bit-mlx on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k tokens 389.0 PP tok/s 26.6 TG tok/s 42116 TTFT (ms) 213.7 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ty3l7hw",
   "source_url": "https://omlx.ai/benchmarks/performance/4ty3l7hw",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-95pfhhve-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1119.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 4k 1,119 53.4 21.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/95pfhhve",
   "source_url": "https://omlx.ai/benchmarks/performance/95pfhhve",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-95pfhhve-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 886.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 16k 886.7 50.4 22.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/95pfhhve",
   "source_url": "https://omlx.ai/benchmarks/performance/95pfhhve",
   "tg_tokens": null,
   "tps": 50.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-95pfhhve-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 473.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 64k 473.2 39.4 24.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/95pfhhve",
   "source_url": "https://omlx.ai/benchmarks/performance/95pfhhve",
   "tg_tokens": null,
   "tps": 39.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-kp1d8es7-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 20.0 GB; TTFT 3030 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1352.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k tokens 1,352 PP tok/s 83.6 TG tok/s 3030 TTFT (ms) 20.0 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kp1d8es7",
   "source_url": "https://omlx.ai/benchmarks/performance/kp1d8es7",
   "tg_tokens": null,
   "tps": 83.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-xj4b6tzc-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: jq0dfgpo; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 673.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 1k 673.9 80.8 20.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xj4b6tzc",
   "source_url": "https://omlx.ai/benchmarks/performance/xj4b6tzc",
   "tg_tokens": null,
   "tps": 80.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-xj4b6tzc-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: jq0dfgpo; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 712.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 4k 712.8 76.4 20.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xj4b6tzc",
   "source_url": "https://omlx.ai/benchmarks/performance/xj4b6tzc",
   "tg_tokens": null,
   "tps": 76.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-xj4b6tzc-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: jq0dfgpo; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 713.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 8k 713.6 74.1 20.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xj4b6tzc",
   "source_url": "https://omlx.ai/benchmarks/performance/xj4b6tzc",
   "tg_tokens": null,
   "tps": 74.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-it8q7kpn-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: zqqdtss0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 413.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 413.8 18.1 15.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/it8q7kpn",
   "source_url": "https://omlx.ai/benchmarks/performance/it8q7kpn",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-it8q7kpn-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: zqqdtss0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 219.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 219.1 9.6 17.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/it8q7kpn",
   "source_url": "https://omlx.ai/benchmarks/performance/it8q7kpn",
   "tg_tokens": null,
   "tps": 9.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-7saz332b-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.1 GB at this context; duplicate traces of the same session: reff9w9d; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 458.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 458.9 22.6 35.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7saz332b",
   "source_url": "https://omlx.ai/benchmarks/performance/7saz332b",
   "tg_tokens": null,
   "tps": 22.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-7saz332b-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.7 GB at this context; duplicate traces of the same session: reff9w9d; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 386.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 16k 386.3 31.4 35.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7saz332b",
   "source_url": "https://omlx.ai/benchmarks/performance/7saz332b",
   "tg_tokens": null,
   "tps": 31.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-d5vklpf1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: x0rpbbte; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 478.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 478.0 10.0 29.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d5vklpf1",
   "source_url": "https://omlx.ai/benchmarks/performance/d5vklpf1",
   "tg_tokens": null,
   "tps": 10.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-d5vklpf1-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: x0rpbbte; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 436.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 16k 436.1 9.5 30.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d5vklpf1",
   "source_url": "https://omlx.ai/benchmarks/performance/d5vklpf1",
   "tg_tokens": null,
   "tps": 9.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-83yessya-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.0 GB at this context; duplicate traces of the same session: h6gylkp5, la1h3b3l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 94.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 94.3 14.7 39.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/83yessya",
   "source_url": "https://omlx.ai/benchmarks/performance/83yessya",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-83yessya-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 40.9 GB at this context; duplicate traces of the same session: h6gylkp5, la1h3b3l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 455.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 455.5 11.7 40.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/83yessya",
   "source_url": "https://omlx.ai/benchmarks/performance/83yessya",
   "tg_tokens": null,
   "tps": 11.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-83yessya-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 41.7 GB at this context; duplicate traces of the same session: h6gylkp5, la1h3b3l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 356.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 356.6 14.3 41.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/83yessya",
   "source_url": "https://omlx.ai/benchmarks/performance/83yessya",
   "tg_tokens": null,
   "tps": 14.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-gfpwgayg-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.1 GB at this context; duplicate traces of the same session: 1pu0e0b7, du60e8uj, snfc4hse; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 93.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 93.9 14.4 39.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gfpwgayg",
   "source_url": "https://omlx.ai/benchmarks/performance/gfpwgayg",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-gfpwgayg-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 40.9 GB at this context; duplicate traces of the same session: 1pu0e0b7, du60e8uj, snfc4hse; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 455.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 455.8 14.5 40.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gfpwgayg",
   "source_url": "https://omlx.ai/benchmarks/performance/gfpwgayg",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-gfpwgayg-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 41.7 GB at this context; duplicate traces of the same session: 1pu0e0b7, du60e8uj, snfc4hse; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 386.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 386.9 14.5 41.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gfpwgayg",
   "source_url": "https://omlx.ai/benchmarks/performance/gfpwgayg",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-gfpwgayg-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 42.3 GB at this context; duplicate traces of the same session: 1pu0e0b7, du60e8uj, snfc4hse; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 375.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 375.1 10.0 42.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gfpwgayg",
   "source_url": "https://omlx.ai/benchmarks/performance/gfpwgayg",
   "tg_tokens": null,
   "tps": 10.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-q9qs2qbq-ctx1024",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 208.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 208.0 31.6 34.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q9qs2qbq",
   "source_url": "https://omlx.ai/benchmarks/performance/q9qs2qbq",
   "tg_tokens": null,
   "tps": 31.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-q9qs2qbq-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 35.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 185.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 185.1 28.4 35.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q9qs2qbq",
   "source_url": "https://omlx.ai/benchmarks/performance/q9qs2qbq",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-q9b9pdk3-ctx1024",
   "model": "gemma-4-E4B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: 40z0l37p; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2763.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-qat-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 2,763 64.5 6.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q9b9pdk3",
   "source_url": "https://omlx.ai/benchmarks/performance/q9b9pdk3",
   "tg_tokens": null,
   "tps": 64.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-q9b9pdk3-ctx4096",
   "model": "gemma-4-E4B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: 40z0l37p; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3448.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-qat-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 3,448 63.0 6.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q9b9pdk3",
   "source_url": "https://omlx.ai/benchmarks/performance/q9b9pdk3",
   "tg_tokens": null,
   "tps": 63.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-q9b9pdk3-ctx8192",
   "model": "gemma-4-E4B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: 40z0l37p; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3420.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-qat-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 3,420 59.8 6.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q9b9pdk3",
   "source_url": "https://omlx.ai/benchmarks/performance/q9b9pdk3",
   "tg_tokens": null,
   "tps": 59.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-q9b9pdk3-ctx16384",
   "model": "gemma-4-E4B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: 40z0l37p; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3131.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-qat-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 3,131 55.0 7.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q9b9pdk3",
   "source_url": "https://omlx.ai/benchmarks/performance/q9b9pdk3",
   "tg_tokens": null,
   "tps": 55.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-q9b9pdk3-ctx32768",
   "model": "gemma-4-E4B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 7.5 GB at this context; duplicate traces of the same session: 40z0l37p; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3035.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-qat-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 3,035 47.9 7.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q9b9pdk3",
   "source_url": "https://omlx.ai/benchmarks/performance/q9b9pdk3",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-q9b9pdk3-ctx65536",
   "model": "gemma-4-E4B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 8.5 GB at this context; duplicate traces of the same session: 40z0l37p; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2550.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-qat-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 64k 2,550 37.4 8.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q9b9pdk3",
   "source_url": "https://omlx.ai/benchmarks/performance/q9b9pdk3",
   "tg_tokens": null,
   "tps": 37.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3gy8kfmw-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTPLX Optimized Speed build; oMLX community benchmark trace; peak memory 36.2 GB at this context; duplicate traces of the same session: tmr0no42; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1080.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 1,080 42.9 36.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3gy8kfmw",
   "source_url": "https://omlx.ai/benchmarks/performance/3gy8kfmw",
   "tg_tokens": null,
   "tps": 42.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3gy8kfmw-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTPLX Optimized Speed build; oMLX community benchmark trace; peak memory 36.4 GB at this context; duplicate traces of the same session: tmr0no42; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1476.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 16k 1,476 38.3 36.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3gy8kfmw",
   "source_url": "https://omlx.ai/benchmarks/performance/3gy8kfmw",
   "tg_tokens": null,
   "tps": 38.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-yrcklmpd-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: y78jiffg; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1119.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 4k 1,119 53.4 21.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yrcklmpd",
   "source_url": "https://omlx.ai/benchmarks/performance/yrcklmpd",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-yrcklmpd-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: y78jiffg; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 886.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 16k 886.7 50.4 22.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yrcklmpd",
   "source_url": "https://omlx.ai/benchmarks/performance/yrcklmpd",
   "tg_tokens": null,
   "tps": 50.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-yrcklmpd-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: y78jiffg; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 473.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 64k 473.2 39.4 24.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yrcklmpd",
   "source_url": "https://omlx.ai/benchmarks/performance/yrcklmpd",
   "tg_tokens": null,
   "tps": 39.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-f9upu6ea-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2267.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,267 104.4 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f9upu6ea",
   "source_url": "https://omlx.ai/benchmarks/performance/f9upu6ea",
   "tg_tokens": null,
   "tps": 104.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-f9upu6ea-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2949.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,949 92.7 20.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f9upu6ea",
   "source_url": "https://omlx.ai/benchmarks/performance/f9upu6ea",
   "tg_tokens": null,
   "tps": 92.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-9of9a5sb-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 36.2 GB at this context; duplicate traces of the same session: osbz6l26; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1689.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 1,689 67.7 36.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9of9a5sb",
   "source_url": "https://omlx.ai/benchmarks/performance/9of9a5sb",
   "tg_tokens": null,
   "tps": 67.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-9of9a5sb-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 36.4 GB at this context; duplicate traces of the same session: osbz6l26; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2703.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 16k 2,703 63.1 36.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9of9a5sb",
   "source_url": "https://omlx.ai/benchmarks/performance/9of9a5sb",
   "tg_tokens": null,
   "tps": 63.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-shg73t4i-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 673.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 1k 673.9 80.8 20.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/shg73t4i",
   "source_url": "https://omlx.ai/benchmarks/performance/shg73t4i",
   "tg_tokens": null,
   "tps": 80.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-shg73t4i-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 712.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 4k 712.8 76.4 20.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/shg73t4i",
   "source_url": "https://omlx.ai/benchmarks/performance/shg73t4i",
   "tg_tokens": null,
   "tps": 76.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-shg73t4i-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 713.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 8k 713.6 74.1 20.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/shg73t4i",
   "source_url": "https://omlx.ai/benchmarks/performance/shg73t4i",
   "tg_tokens": null,
   "tps": 74.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-j3ihnndj-ctx4096",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "MTPLX Optimized Speed build; oMLX community benchmark trace; peak memory 29.4 GB at this context; duplicate traces of the same session: m6vhf5hl; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 215.4,
   "provenance": "community",
   "quant": "fp16",
   "quote": "Qwen3.8-27B-Abliterated-MTPLX-Optimized-Speed-fp16 on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 4k 215.4 19.4 29.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j3ihnndj",
   "source_url": "https://omlx.ai/benchmarks/performance/j3ihnndj",
   "tg_tokens": null,
   "tps": 19.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-j3ihnndj-ctx32768",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "MTPLX Optimized Speed build; oMLX community benchmark trace; peak memory 31.2 GB at this context; duplicate traces of the same session: m6vhf5hl; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 182.8,
   "provenance": "community",
   "quant": "fp16",
   "quote": "Qwen3.8-27B-Abliterated-MTPLX-Optimized-Speed-fp16 on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 32k 182.8 20.2 31.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j3ihnndj",
   "source_url": "https://omlx.ai/benchmarks/performance/j3ihnndj",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-h8ehuogb-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 17.0 GB; TTFT 18147 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 225.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k tokens 225.7 PP tok/s 9.8 TG tok/s 18147 TTFT (ms) 17.0 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h8ehuogb",
   "source_url": "https://omlx.ai/benchmarks/performance/h8ehuogb",
   "tg_tokens": null,
   "tps": 9.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-eyvoq5x4-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: xbf1jt3d; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 193.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 193.3 10.1 15.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eyvoq5x4",
   "source_url": "https://omlx.ai/benchmarks/performance/eyvoq5x4",
   "tg_tokens": null,
   "tps": 10.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-eyvoq5x4-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: xbf1jt3d; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 227.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 227.4 9.8 17.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eyvoq5x4",
   "source_url": "https://omlx.ai/benchmarks/performance/eyvoq5x4",
   "tg_tokens": null,
   "tps": 9.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-g07ustje-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 16.8 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 229.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 229.6 10.5 16.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g07ustje",
   "source_url": "https://omlx.ai/benchmarks/performance/g07ustje",
   "tg_tokens": null,
   "tps": 10.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-g07ustje-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 354.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 354.3 9.8 18.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g07ustje",
   "source_url": "https://omlx.ai/benchmarks/performance/g07ustje",
   "tg_tokens": null,
   "tps": 9.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-9bxdtvwa-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 17.0 GB; TTFT 9592 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 427.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k tokens 427.0 PP tok/s 17.8 TG tok/s 9592 TTFT (ms) 17.0 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9bxdtvwa",
   "source_url": "https://omlx.ai/benchmarks/performance/9bxdtvwa",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-scymd05z-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 475.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 475.0 26.7 18.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/scymd05z",
   "source_url": "https://omlx.ai/benchmarks/performance/scymd05z",
   "tg_tokens": null,
   "tps": 26.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-scymd05z-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 429.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 429.5 38.4 19.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/scymd05z",
   "source_url": "https://omlx.ai/benchmarks/performance/scymd05z",
   "tg_tokens": null,
   "tps": 38.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-iifwzvwv-ctx32768",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "384 experts oQ4e + 128 experts oQ2 mixed MoE quant; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 26.8 GB; TTFT 208402 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 157.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-384e-oQ4e-128e-oQ2-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k tokens 157.2 PP tok/s 6.3 TG tok/s 208402 TTFT (ms) 26.8 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iifwzvwv",
   "source_url": "https://omlx.ai/benchmarks/performance/iifwzvwv",
   "tg_tokens": null,
   "tps": 6.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ljiqe52p-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace, single run at 1k context; peak memory 100.9 GB; TTFT 1413 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 724.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k tokens 724.5 PP tok/s 37.1 TG tok/s 1413 TTFT (ms) 100.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ljiqe52p",
   "source_url": "https://omlx.ai/benchmarks/performance/ljiqe52p",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-02f0dsbg-ctx1024",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 8.6 GB at this context; duplicate traces of the same session: jun33t0k; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 143.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-12B-it-oQ4e on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 143.1 24.6 8.6 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/02f0dsbg",
   "source_url": "https://omlx.ai/benchmarks/performance/02f0dsbg",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-02f0dsbg-ctx4096",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 8.6 GB at this context; duplicate traces of the same session: jun33t0k; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 139.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-12B-it-oQ4e on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 139.2 15.4 8.6 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/02f0dsbg",
   "source_url": "https://omlx.ai/benchmarks/performance/02f0dsbg",
   "tg_tokens": null,
   "tps": 15.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-z3w35uae-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 403.3 GB at this context; duplicate traces of the same session: 605honrq, 619tfieh, si1e3s3w; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 266.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 266.0 20.2 403.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z3w35uae",
   "source_url": "https://omlx.ai/benchmarks/performance/z3w35uae",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-z3w35uae-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 403.7 GB at this context; duplicate traces of the same session: 605honrq, 619tfieh, si1e3s3w; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 416.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 416.8 20.0 403.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z3w35uae",
   "source_url": "https://omlx.ai/benchmarks/performance/z3w35uae",
   "tg_tokens": null,
   "tps": 20.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-z3w35uae-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 404.1 GB at this context; duplicate traces of the same session: 605honrq, 619tfieh, si1e3s3w; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 429.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 8k 429.7 19.9 404.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z3w35uae",
   "source_url": "https://omlx.ai/benchmarks/performance/z3w35uae",
   "tg_tokens": null,
   "tps": 19.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-z3w35uae-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 404.6 GB at this context; duplicate traces of the same session: 605honrq, 619tfieh, si1e3s3w; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 426.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 16k 426.9 19.6 404.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z3w35uae",
   "source_url": "https://omlx.ai/benchmarks/performance/z3w35uae",
   "tg_tokens": null,
   "tps": 19.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-gfr0879u-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2267.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,267 104.4 20.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gfr0879u",
   "source_url": "https://omlx.ai/benchmarks/performance/gfr0879u",
   "tg_tokens": null,
   "tps": 104.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-gfr0879u-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2949.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,949 92.7 20.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gfr0879u",
   "source_url": "https://omlx.ai/benchmarks/performance/gfr0879u",
   "tg_tokens": null,
   "tps": 92.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-3idxu4ie-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 9fsj8jhf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 259.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 259.2 29.5 16.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3idxu4ie",
   "source_url": "https://omlx.ai/benchmarks/performance/3idxu4ie",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-3idxu4ie-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 9fsj8jhf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 271.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 271.6 28.5 19.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3idxu4ie",
   "source_url": "https://omlx.ai/benchmarks/performance/3idxu4ie",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-zoqsc57y-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 16.8 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 229.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 229.6 10.5 16.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zoqsc57y",
   "source_url": "https://omlx.ai/benchmarks/performance/zoqsc57y",
   "tg_tokens": null,
   "tps": 10.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-zoqsc57y-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 354.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 354.3 9.8 18.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zoqsc57y",
   "source_url": "https://omlx.ai/benchmarks/performance/zoqsc57y",
   "tg_tokens": null,
   "tps": 9.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-9sladlo8-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: h2dg1vw3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 494.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 494.2 16.9 17.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9sladlo8",
   "source_url": "https://omlx.ai/benchmarks/performance/9sladlo8",
   "tg_tokens": null,
   "tps": 16.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-9sladlo8-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: h2dg1vw3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 471.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 471.9 16.0 18.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9sladlo8",
   "source_url": "https://omlx.ai/benchmarks/performance/9sladlo8",
   "tg_tokens": null,
   "tps": 16.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-whyvpyvm-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 475.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 475.0 26.7 18.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/whyvpyvm",
   "source_url": "https://omlx.ai/benchmarks/performance/whyvpyvm",
   "tg_tokens": null,
   "tps": 26.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-whyvpyvm-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 429.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 429.5 38.4 19.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/whyvpyvm",
   "source_url": "https://omlx.ai/benchmarks/performance/whyvpyvm",
   "tg_tokens": null,
   "tps": 38.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-c56ww5ee-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: 2s6mp4bd; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 464.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 464.1 28.0 27.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c56ww5ee",
   "source_url": "https://omlx.ai/benchmarks/performance/c56ww5ee",
   "tg_tokens": null,
   "tps": 28.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-c56ww5ee-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: 2s6mp4bd; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 414.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 16k 414.2 36.1 28.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c56ww5ee",
   "source_url": "https://omlx.ai/benchmarks/performance/c56ww5ee",
   "tg_tokens": null,
   "tps": 36.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-o81c9jvz-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.1 GB at this context; duplicate traces of the same session: polczpew; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1421.0,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-Flash-Next-oQ3e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 1k 1,421 52.9 80.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o81c9jvz",
   "source_url": "https://omlx.ai/benchmarks/performance/o81c9jvz",
   "tg_tokens": null,
   "tps": 52.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-o81c9jvz-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 82.3 GB at this context; duplicate traces of the same session: polczpew; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1975.0,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-Flash-Next-oQ3e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 16k 1,975 51.1 82.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o81c9jvz",
   "source_url": "https://omlx.ai/benchmarks/performance/o81c9jvz",
   "tg_tokens": null,
   "tps": 51.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qcwsdntd-ctx16384",
   "model": "dsv4-0731",
   "notes": "author abbreviation for DeepSeek-V4-Flash-0731; oMLX community benchmark trace, single run at 16k context; peak memory 157.6 GB; TTFT 28253 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 579.9,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "dsv4-0731-mxfp4 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k tokens 579.9 PP tok/s 21.4 TG tok/s 28253 TTFT (ms) 157.6 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qcwsdntd",
   "source_url": "https://omlx.ai/benchmarks/performance/qcwsdntd",
   "tg_tokens": null,
   "tps": 21.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-fvqlqkxc-ctx199680",
   "model": "dsv4-0731",
   "notes": "author abbreviation for DeepSeek-V4-Flash-0731; oMLX community benchmark trace, single run at 195k context; peak memory 148.4 GB; TTFT 435837 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 458.9,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "dsv4-0731-mxfp4 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 195k tokens 458.9 PP tok/s 21.2 TG tok/s 435837 TTFT (ms) 148.4 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fvqlqkxc",
   "source_url": "https://omlx.ai/benchmarks/performance/fvqlqkxc",
   "tg_tokens": null,
   "tps": 21.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-20ba53w1-ctx4096",
   "model": "dsv4-0731",
   "notes": "author abbreviation for DeepSeek-V4-Flash-0731; oMLX community benchmark trace; peak memory 147.3 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 593.3,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "dsv4-0731-mxfp4 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 593.3 25.6 147.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/20ba53w1",
   "source_url": "https://omlx.ai/benchmarks/performance/20ba53w1",
   "tg_tokens": null,
   "tps": 25.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-20ba53w1-ctx16384",
   "model": "dsv4-0731",
   "notes": "author abbreviation for DeepSeek-V4-Flash-0731; oMLX community benchmark trace; peak memory 147.4 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 598.2,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "dsv4-0731-mxfp4 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 598.2 24.5 147.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/20ba53w1",
   "source_url": "https://omlx.ai/benchmarks/performance/20ba53w1",
   "tg_tokens": null,
   "tps": 24.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-x5ip24yy-ctx1024",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: ezxaglbl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 541.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 541.7 28.1 34.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x5ip24yy",
   "source_url": "https://omlx.ai/benchmarks/performance/x5ip24yy",
   "tg_tokens": null,
   "tps": 28.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-x5ip24yy-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 35.0 GB at this context; duplicate traces of the same session: ezxaglbl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 623.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 623.4 30.5 35.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x5ip24yy",
   "source_url": "https://omlx.ai/benchmarks/performance/x5ip24yy",
   "tg_tokens": null,
   "tps": 30.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-93fdywlq-ctx16384",
   "model": "gemma-4-12b-it",
   "notes": "AEON Abliterated MLX build; oMLX community benchmark trace, single run at 16k context; peak memory 14.1 GB; TTFT 37581 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 436.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Gemma-4-12B-it-AEON-Abliterated-MLX-8bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 8bit; 16k tokens 436.0 PP tok/s 8.8 TG tok/s 37581 TTFT (ms) 14.1 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/93fdywlq",
   "source_url": "https://omlx.ai/benchmarks/performance/93fdywlq",
   "tg_tokens": null,
   "tps": 8.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-8v77h9ne-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "UD MLX FP16 build; oMLX community benchmark trace, single run at 8k context; peak memory 17.5 GB; TTFT 18668 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 438.8,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-3bit-FP16 on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 3bit; 8k tokens 438.8 PP tok/s 37.0 TG tok/s 18668 TTFT (ms) 17.5 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8v77h9ne",
   "source_url": "https://omlx.ai/benchmarks/performance/8v77h9ne",
   "tg_tokens": null,
   "tps": 37.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-6cba28r0-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: b2n1ukec; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 259.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 259.1 29.4 16.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6cba28r0",
   "source_url": "https://omlx.ai/benchmarks/performance/6cba28r0",
   "tg_tokens": null,
   "tps": 29.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-6cba28r0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: b2n1ukec; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 274.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 274.2 28.5 19.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6cba28r0",
   "source_url": "https://omlx.ai/benchmarks/performance/6cba28r0",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1wu88mw9-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 82.9 GB; TTFT 16508 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1985.0,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-Flash-Next-oQ3e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 32k tokens 1,985 PP tok/s 63.8 TG tok/s 16508 TTFT (ms) 82.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1wu88mw9",
   "source_url": "https://omlx.ai/benchmarks/performance/1wu88mw9",
   "tg_tokens": null,
   "tps": 63.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-azadx3at-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: 93uectcs, c6qsj27k, mks7vwel; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 608.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 608.3 78.5 71.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/azadx3at",
   "source_url": "https://omlx.ai/benchmarks/performance/azadx3at",
   "tg_tokens": null,
   "tps": 78.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-azadx3at-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: 93uectcs, c6qsj27k, mks7vwel; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 832.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 832.2 59.1 71.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/azadx3at",
   "source_url": "https://omlx.ai/benchmarks/performance/azadx3at",
   "tg_tokens": null,
   "tps": 59.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-azadx3at-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.0 GB at this context; duplicate traces of the same session: 93uectcs, c6qsj27k, mks7vwel; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 825.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 825.3 70.4 72.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/azadx3at",
   "source_url": "https://omlx.ai/benchmarks/performance/azadx3at",
   "tg_tokens": null,
   "tps": 70.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-azadx3at-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 75.1 GB at this context; duplicate traces of the same session: 93uectcs, c6qsj27k, mks7vwel; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 811.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 811.9 59.8 75.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/azadx3at",
   "source_url": "https://omlx.ai/benchmarks/performance/azadx3at",
   "tg_tokens": null,
   "tps": 59.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bkyhke7l-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: 0od5v0u2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 172.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 172.5 14.1 23.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bkyhke7l",
   "source_url": "https://omlx.ai/benchmarks/performance/bkyhke7l",
   "tg_tokens": null,
   "tps": 14.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bkyhke7l-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 0od5v0u2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 467.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 467.2 10.0 23.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bkyhke7l",
   "source_url": "https://omlx.ai/benchmarks/performance/bkyhke7l",
   "tg_tokens": null,
   "tps": 10.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-s37tw5i4-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2081.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,081 67.4 70.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s37tw5i4",
   "source_url": "https://omlx.ai/benchmarks/performance/s37tw5i4",
   "tg_tokens": null,
   "tps": 67.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-s37tw5i4-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2474.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,474 65.7 71.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s37tw5i4",
   "source_url": "https://omlx.ai/benchmarks/performance/s37tw5i4",
   "tg_tokens": null,
   "tps": 65.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-s37tw5i4-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2397.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,397 64.3 72.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s37tw5i4",
   "source_url": "https://omlx.ai/benchmarks/performance/s37tw5i4",
   "tg_tokens": null,
   "tps": 64.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-s37tw5i4-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2343.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,343 62.6 74.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s37tw5i4",
   "source_url": "https://omlx.ai/benchmarks/performance/s37tw5i4",
   "tg_tokens": null,
   "tps": 62.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9a683p9v-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.3 GB at this context; duplicate traces of the same session: gpxr4l33; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1047.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,047 40.0 70.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9a683p9v",
   "source_url": "https://omlx.ai/benchmarks/performance/9a683p9v",
   "tg_tokens": null,
   "tps": 40.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9a683p9v-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.4 GB at this context; duplicate traces of the same session: gpxr4l33; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2411.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,411 36.1 71.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9a683p9v",
   "source_url": "https://omlx.ai/benchmarks/performance/9a683p9v",
   "tg_tokens": null,
   "tps": 36.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9a683p9v-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.4 GB at this context; duplicate traces of the same session: gpxr4l33; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2412.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,412 54.4 72.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9a683p9v",
   "source_url": "https://omlx.ai/benchmarks/performance/9a683p9v",
   "tg_tokens": null,
   "tps": 54.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9a683p9v-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.9 GB at this context; duplicate traces of the same session: gpxr4l33; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2340.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,340 58.8 74.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9a683p9v",
   "source_url": "https://omlx.ai/benchmarks/performance/9a683p9v",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-gsd4o8cn-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 51.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-oQ4e-mtp on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 51.7 9.3 18.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gsd4o8cn",
   "source_url": "https://omlx.ai/benchmarks/performance/gsd4o8cn",
   "tg_tokens": null,
   "tps": 9.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-gsd4o8cn-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 41.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-oQ4e-mtp on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 32k 41.2 6.8 21.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gsd4o8cn",
   "source_url": "https://omlx.ai/benchmarks/performance/gsd4o8cn",
   "tg_tokens": null,
   "tps": 6.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-r9f14pn9-ctx4096",
   "model": "dsv4-0731",
   "notes": "author abbreviation for DeepSeek-V4-Flash-0731; oMLX community benchmark trace; peak memory 147.3 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 593.3,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "dsv4-0731-mxfp4 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 593.3 25.6 147.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r9f14pn9",
   "source_url": "https://omlx.ai/benchmarks/performance/r9f14pn9",
   "tg_tokens": null,
   "tps": 25.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-r9f14pn9-ctx16384",
   "model": "dsv4-0731",
   "notes": "author abbreviation for DeepSeek-V4-Flash-0731; oMLX community benchmark trace; peak memory 147.4 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 598.2,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "dsv4-0731-mxfp4 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 598.2 24.5 147.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r9f14pn9",
   "source_url": "https://omlx.ai/benchmarks/performance/r9f14pn9",
   "tg_tokens": null,
   "tps": 24.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-d9w8go0v-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.7 GB at this context; duplicate traces of the same session: 29zxfp0l; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2385.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,385 78.4 14.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d9w8go0v",
   "source_url": "https://omlx.ai/benchmarks/performance/d9w8go0v",
   "tg_tokens": null,
   "tps": 78.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-d9w8go0v-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 15.2 GB at this context; duplicate traces of the same session: 29zxfp0l; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2207.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,207 70.9 15.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d9w8go0v",
   "source_url": "https://omlx.ai/benchmarks/performance/d9w8go0v",
   "tg_tokens": null,
   "tps": 70.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pt2764i6-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: 0advqibp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 959.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 959.3 45.8 19.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pt2764i6",
   "source_url": "https://omlx.ai/benchmarks/performance/pt2764i6",
   "tg_tokens": null,
   "tps": 45.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pt2764i6-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 0advqibp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 991.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 991.3 49.5 21.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pt2764i6",
   "source_url": "https://omlx.ai/benchmarks/performance/pt2764i6",
   "tg_tokens": null,
   "tps": 49.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-972qr6c5-ctx1024",
   "model": "Ornith-1.5-35B-A3B-BigBang",
   "notes": "KaedeTai MTP build; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 2mukyo80; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1130.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "KaedeTai--Ornith-1.5-35B-A3B-BigBang-MTP-mlx-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 1k 1,130 80.4 19.8 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/972qr6c5",
   "source_url": "https://omlx.ai/benchmarks/performance/972qr6c5",
   "tg_tokens": null,
   "tps": 80.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-972qr6c5-ctx4096",
   "model": "Ornith-1.5-35B-A3B-BigBang",
   "notes": "KaedeTai MTP build; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: 2mukyo80; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1114.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "KaedeTai--Ornith-1.5-35B-A3B-BigBang-MTP-mlx-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 4k 1,114 75.8 20.7 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/972qr6c5",
   "source_url": "https://omlx.ai/benchmarks/performance/972qr6c5",
   "tg_tokens": null,
   "tps": 75.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-htjx77hi-ctx1024",
   "model": "Qwen-AgentWorld-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: vuw2lgfr; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1163.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen-AgentWorld-35B-A3B-oQ4 on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 1k 1,163 73.0 20.2 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/htjx77hi",
   "source_url": "https://omlx.ai/benchmarks/performance/htjx77hi",
   "tg_tokens": null,
   "tps": 73.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-htjx77hi-ctx4096",
   "model": "Qwen-AgentWorld-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: vuw2lgfr; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1116.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen-AgentWorld-35B-A3B-oQ4 on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 4k 1,116 70.7 21.0 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/htjx77hi",
   "source_url": "https://omlx.ai/benchmarks/performance/htjx77hi",
   "tg_tokens": null,
   "tps": 70.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wy2jjwt8-ctx1024",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 2.4 GB at this context; duplicate traces of the same session: lg5yo06f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 17404.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.5-0.8B-bf16 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 bf16; 1k 17,404 247.9 2.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wy2jjwt8",
   "source_url": "https://omlx.ai/benchmarks/performance/wy2jjwt8",
   "tg_tokens": null,
   "tps": 247.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wy2jjwt8-ctx4096",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 2.6 GB at this context; duplicate traces of the same session: lg5yo06f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 19054.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.5-0.8B-bf16 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 bf16; 4k 19,054 241.4 2.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wy2jjwt8",
   "source_url": "https://omlx.ai/benchmarks/performance/wy2jjwt8",
   "tg_tokens": null,
   "tps": 241.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wy2jjwt8-ctx32768",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 2.9 GB at this context; duplicate traces of the same session: lg5yo06f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 17501.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.5-0.8B-bf16 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 bf16; 32k 17,501 178.9 2.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wy2jjwt8",
   "source_url": "https://omlx.ai/benchmarks/performance/wy2jjwt8",
   "tg_tokens": null,
   "tps": 178.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wy2jjwt8-ctx65536",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 3.2 GB at this context; duplicate traces of the same session: lg5yo06f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 13575.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.5-0.8B-bf16 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 bf16; 64k 13,575 138.4 3.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wy2jjwt8",
   "source_url": "https://omlx.ai/benchmarks/performance/wy2jjwt8",
   "tg_tokens": null,
   "tps": 138.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wy2jjwt8-ctx131072",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 4.5 GB at this context; duplicate traces of the same session: lg5yo06f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 9059.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.5-0.8B-bf16 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 bf16; 128k 9,059 92.1 4.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wy2jjwt8",
   "source_url": "https://omlx.ai/benchmarks/performance/wy2jjwt8",
   "tg_tokens": null,
   "tps": 92.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wy2jjwt8-ctx199680",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: lg5yo06f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6308.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.5-0.8B-bf16 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 bf16; 195k 6,308 69.8 6.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wy2jjwt8",
   "source_url": "https://omlx.ai/benchmarks/performance/wy2jjwt8",
   "tg_tokens": null,
   "tps": 69.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tf1vfr6e-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 07wauoo2, 3t4keih5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5644.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 5,644 127.3 20.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tf1vfr6e",
   "source_url": "https://omlx.ai/benchmarks/performance/tf1vfr6e",
   "tg_tokens": null,
   "tps": 127.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tf1vfr6e-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: 07wauoo2, 3t4keih5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4238.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 4,238 115.1 20.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tf1vfr6e",
   "source_url": "https://omlx.ai/benchmarks/performance/tf1vfr6e",
   "tg_tokens": null,
   "tps": 115.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tf1vfr6e-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 07wauoo2, 3t4keih5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3050.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 3,050 82.2 21.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tf1vfr6e",
   "source_url": "https://omlx.ai/benchmarks/performance/tf1vfr6e",
   "tg_tokens": null,
   "tps": 82.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xnwk8a9c-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 86.8 GB at this context; duplicate traces of the same session: 14tpucqm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1400.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 1,400 56.1 86.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xnwk8a9c",
   "source_url": "https://omlx.ai/benchmarks/performance/xnwk8a9c",
   "tg_tokens": null,
   "tps": 56.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xnwk8a9c-ctx65536",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 90.6 GB at this context; duplicate traces of the same session: 14tpucqm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1586.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 1,586 59.3 90.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xnwk8a9c",
   "source_url": "https://omlx.ai/benchmarks/performance/xnwk8a9c",
   "tg_tokens": null,
   "tps": 59.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-93ysicsy-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 86.8 GB at this context; duplicate traces of the same session: yrxqktpy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1612.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 1,612 62.7 86.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/93ysicsy",
   "source_url": "https://omlx.ai/benchmarks/performance/93ysicsy",
   "tg_tokens": null,
   "tps": 62.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-93ysicsy-ctx65536",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 90.6 GB at this context; duplicate traces of the same session: yrxqktpy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1876.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 1,876 74.8 90.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/93ysicsy",
   "source_url": "https://omlx.ai/benchmarks/performance/93ysicsy",
   "tg_tokens": null,
   "tps": 74.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-t4gx2yhv-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.3 GB at this context; duplicate traces of the same session: rjoxaqmz, vzp7qcdg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2081.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,081 67.4 70.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t4gx2yhv",
   "source_url": "https://omlx.ai/benchmarks/performance/t4gx2yhv",
   "tg_tokens": null,
   "tps": 67.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-t4gx2yhv-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.4 GB at this context; duplicate traces of the same session: rjoxaqmz, vzp7qcdg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2474.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,474 65.7 71.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t4gx2yhv",
   "source_url": "https://omlx.ai/benchmarks/performance/t4gx2yhv",
   "tg_tokens": null,
   "tps": 65.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-t4gx2yhv-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.4 GB at this context; duplicate traces of the same session: rjoxaqmz, vzp7qcdg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2397.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,397 64.3 72.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t4gx2yhv",
   "source_url": "https://omlx.ai/benchmarks/performance/t4gx2yhv",
   "tg_tokens": null,
   "tps": 64.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-t4gx2yhv-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.9 GB at this context; duplicate traces of the same session: rjoxaqmz, vzp7qcdg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2343.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,343 62.6 74.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t4gx2yhv",
   "source_url": "https://omlx.ai/benchmarks/performance/t4gx2yhv",
   "tg_tokens": null,
   "tps": 62.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-j7083fix-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.3 GB at this context; duplicate traces of the same session: ueqywlll; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1047.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,047 40.0 70.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j7083fix",
   "source_url": "https://omlx.ai/benchmarks/performance/j7083fix",
   "tg_tokens": null,
   "tps": 40.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-j7083fix-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.4 GB at this context; duplicate traces of the same session: ueqywlll; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2411.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,411 36.1 71.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j7083fix",
   "source_url": "https://omlx.ai/benchmarks/performance/j7083fix",
   "tg_tokens": null,
   "tps": 36.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-j7083fix-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.4 GB at this context; duplicate traces of the same session: ueqywlll; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2412.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,412 54.4 72.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j7083fix",
   "source_url": "https://omlx.ai/benchmarks/performance/j7083fix",
   "tg_tokens": null,
   "tps": 54.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-j7083fix-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.9 GB at this context; duplicate traces of the same session: ueqywlll; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2340.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,340 58.8 74.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j7083fix",
   "source_url": "https://omlx.ai/benchmarks/performance/j7083fix",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-ookgxeo9-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 51.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-oQ4e-mtp on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 51.7 9.3 18.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ookgxeo9",
   "source_url": "https://omlx.ai/benchmarks/performance/ookgxeo9",
   "tg_tokens": null,
   "tps": 9.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-ookgxeo9-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 41.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-oQ4e-mtp on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 32k 41.2 6.8 21.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ookgxeo9",
   "source_url": "https://omlx.ai/benchmarks/performance/ookgxeo9",
   "tg_tokens": null,
   "tps": 6.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-30l7392n-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 19.9 GB; TTFT 79574 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 51.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-oQ4e-mtp on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k tokens 51.5 PP tok/s 11.0 TG tok/s 79574 TTFT (ms) 19.9 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/30l7392n",
   "source_url": "https://omlx.ai/benchmarks/performance/30l7392n",
   "tg_tokens": null,
   "tps": 11.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-7lfk0m1x-ctx1024",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 912.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 912.3 118.9 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7lfk0m1x",
   "source_url": "https://omlx.ai/benchmarks/performance/7lfk0m1x",
   "tg_tokens": null,
   "tps": 118.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-7lfk0m1x-ctx4096",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1053.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,053 84.4 23.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7lfk0m1x",
   "source_url": "https://omlx.ai/benchmarks/performance/7lfk0m1x",
   "tg_tokens": null,
   "tps": 84.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-7lfk0m1x-ctx8192",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 988.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 988.9 107.7 23.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7lfk0m1x",
   "source_url": "https://omlx.ai/benchmarks/performance/7lfk0m1x",
   "tg_tokens": null,
   "tps": 107.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-7lfk0m1x-ctx16384",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 873.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 873.6 104.2 24.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7lfk0m1x",
   "source_url": "https://omlx.ai/benchmarks/performance/7lfk0m1x",
   "tg_tokens": null,
   "tps": 104.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-84f3cpy8-ctx1024",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.2 GB at this context; duplicate traces of the same session: m5gckx6q; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 203.1,
   "provenance": "community",
   "quant": "OptiQ",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-OptiQ on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 203.1 16.2 7.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/84f3cpy8",
   "source_url": "https://omlx.ai/benchmarks/performance/84f3cpy8",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-84f3cpy8-ctx4096",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.6 GB at this context; duplicate traces of the same session: m5gckx6q; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 216.6,
   "provenance": "community",
   "quant": "OptiQ",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-OptiQ on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 216.6 15.9 7.6 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/84f3cpy8",
   "source_url": "https://omlx.ai/benchmarks/performance/84f3cpy8",
   "tg_tokens": null,
   "tps": 15.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-7miic193-ctx1024",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 2.4 GB at this context; duplicate traces of the same session: 04mq6qel, gnkmg0go, wag8p4un; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 17404.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.5-0.8B-bf16 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 bf16; 1k 17,404 247.9 2.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7miic193",
   "source_url": "https://omlx.ai/benchmarks/performance/7miic193",
   "tg_tokens": null,
   "tps": 247.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-7miic193-ctx4096",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 2.6 GB at this context; duplicate traces of the same session: 04mq6qel, gnkmg0go, wag8p4un; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 19054.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.5-0.8B-bf16 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 bf16; 4k 19,054 241.4 2.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7miic193",
   "source_url": "https://omlx.ai/benchmarks/performance/7miic193",
   "tg_tokens": null,
   "tps": 241.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-7miic193-ctx32768",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 2.9 GB at this context; duplicate traces of the same session: 04mq6qel, gnkmg0go, wag8p4un; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 17501.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.5-0.8B-bf16 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 bf16; 32k 17,501 178.9 2.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7miic193",
   "source_url": "https://omlx.ai/benchmarks/performance/7miic193",
   "tg_tokens": null,
   "tps": 178.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-7miic193-ctx65536",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 3.2 GB at this context; duplicate traces of the same session: 04mq6qel, gnkmg0go, wag8p4un; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 13575.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.5-0.8B-bf16 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 bf16; 64k 13,575 138.4 3.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7miic193",
   "source_url": "https://omlx.ai/benchmarks/performance/7miic193",
   "tg_tokens": null,
   "tps": 138.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-7miic193-ctx131072",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 4.5 GB at this context; duplicate traces of the same session: 04mq6qel, gnkmg0go, wag8p4un; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 9059.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.5-0.8B-bf16 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 bf16; 128k 9,059 92.1 4.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7miic193",
   "source_url": "https://omlx.ai/benchmarks/performance/7miic193",
   "tg_tokens": null,
   "tps": 92.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-7miic193-ctx199680",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: 04mq6qel, gnkmg0go, wag8p4un; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6308.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.5-0.8B-bf16 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 bf16; 195k 6,308 69.8 6.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7miic193",
   "source_url": "https://omlx.ai/benchmarks/performance/7miic193",
   "tg_tokens": null,
   "tps": 69.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-gy9l619e-ctx1024",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 2.4 GB at this context; duplicate traces of the same session: sut2whux; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 17192.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.5-0.8B-bf16 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 bf16; 1k 17,192 229.3 2.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gy9l619e",
   "source_url": "https://omlx.ai/benchmarks/performance/gy9l619e",
   "tg_tokens": null,
   "tps": 229.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-gy9l619e-ctx4096",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 2.6 GB at this context; duplicate traces of the same session: sut2whux; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 15550.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.5-0.8B-bf16 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 bf16; 4k 15,550 240.8 2.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gy9l619e",
   "source_url": "https://omlx.ai/benchmarks/performance/gy9l619e",
   "tg_tokens": null,
   "tps": 240.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zuyfxkpj-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: ub6rjxc6, wxbafkq4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1443.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,443 33.0 18.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zuyfxkpj",
   "source_url": "https://omlx.ai/benchmarks/performance/zuyfxkpj",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zuyfxkpj-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: ub6rjxc6, wxbafkq4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1245.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,245 29.8 19.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zuyfxkpj",
   "source_url": "https://omlx.ai/benchmarks/performance/zuyfxkpj",
   "tg_tokens": null,
   "tps": 29.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zuyfxkpj-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: ub6rjxc6, wxbafkq4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 854.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 854.9 22.5 22.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zuyfxkpj",
   "source_url": "https://omlx.ai/benchmarks/performance/zuyfxkpj",
   "tg_tokens": null,
   "tps": 22.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-azywpnr5-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 355.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 355.1 38.2 19.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/azywpnr5",
   "source_url": "https://omlx.ai/benchmarks/performance/azywpnr5",
   "tg_tokens": null,
   "tps": 38.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-azywpnr5-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 337.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 16k 337.5 35.4 20.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/azywpnr5",
   "source_url": "https://omlx.ai/benchmarks/performance/azywpnr5",
   "tg_tokens": null,
   "tps": 35.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-azywpnr5-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 312.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 312.6 32.5 22.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/azywpnr5",
   "source_url": "https://omlx.ai/benchmarks/performance/azywpnr5",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fxp5j1fc-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP320",
   "notes": "MTP Vision build; oMLX community benchmark trace; peak memory 36.6 GB at this context; duplicate traces of the same session: cjb5eroj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 922.1,
   "provenance": "community",
   "quant": "oQ3e-DWQ",
   "quote": "Qwen3.8-Flash-Next-REAP320-oQ3e-DWQ-MTP-Vision-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 3bit; 1k 922.1 59.8 36.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fxp5j1fc",
   "source_url": "https://omlx.ai/benchmarks/performance/fxp5j1fc",
   "tg_tokens": null,
   "tps": 59.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fxp5j1fc-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP320",
   "notes": "MTP Vision build; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: cjb5eroj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1878.0,
   "provenance": "community",
   "quant": "oQ3e-DWQ",
   "quote": "Qwen3.8-Flash-Next-REAP320-oQ3e-DWQ-MTP-Vision-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 3bit; 4k 1,878 58.5 36.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fxp5j1fc",
   "source_url": "https://omlx.ai/benchmarks/performance/fxp5j1fc",
   "tg_tokens": null,
   "tps": 58.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2sj34k5z-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: evir1n0t, m4cukyn1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 555.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 555.3 61.8 70.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2sj34k5z",
   "source_url": "https://omlx.ai/benchmarks/performance/2sj34k5z",
   "tg_tokens": null,
   "tps": 61.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2sj34k5z-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.6 GB at this context; duplicate traces of the same session: evir1n0t, m4cukyn1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 821.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 821.0 59.6 72.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2sj34k5z",
   "source_url": "https://omlx.ai/benchmarks/performance/2sj34k5z",
   "tg_tokens": null,
   "tps": 59.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2sj34k5z-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.4 GB at this context; duplicate traces of the same session: evir1n0t, m4cukyn1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 811.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 811.8 57.5 74.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2sj34k5z",
   "source_url": "https://omlx.ai/benchmarks/performance/2sj34k5z",
   "tg_tokens": null,
   "tps": 57.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-w88jbwu6-ctx1024",
   "model": "gemma-4-26B-A4B-it-heretic-msq",
   "notes": "oMLX community benchmark trace; peak memory 9.2 GB at this context; duplicate traces of the same session: lznx0bo5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1920.0,
   "provenance": "community",
   "quant": "2.6bit",
   "quote": "gemma-4-26B-A4B-it-heretic-msq-2.6bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 1,920 71.0 9.2 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w88jbwu6",
   "source_url": "https://omlx.ai/benchmarks/performance/w88jbwu6",
   "tg_tokens": null,
   "tps": 71.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-w88jbwu6-ctx4096",
   "model": "gemma-4-26B-A4B-it-heretic-msq",
   "notes": "oMLX community benchmark trace; peak memory 9.5 GB at this context; duplicate traces of the same session: lznx0bo5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1906.0,
   "provenance": "community",
   "quant": "2.6bit",
   "quote": "gemma-4-26B-A4B-it-heretic-msq-2.6bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 1,906 69.1 9.5 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w88jbwu6",
   "source_url": "https://omlx.ai/benchmarks/performance/w88jbwu6",
   "tg_tokens": null,
   "tps": 69.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-hhko08us-ctx1024",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 9dax4ydk, 9zej0sg0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 912.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 912.3 118.9 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hhko08us",
   "source_url": "https://omlx.ai/benchmarks/performance/hhko08us",
   "tg_tokens": null,
   "tps": 118.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-hhko08us-ctx4096",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: 9dax4ydk, 9zej0sg0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1053.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,053 84.4 23.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hhko08us",
   "source_url": "https://omlx.ai/benchmarks/performance/hhko08us",
   "tg_tokens": null,
   "tps": 84.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-hhko08us-ctx8192",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: 9dax4ydk, 9zej0sg0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 988.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 988.9 107.7 23.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hhko08us",
   "source_url": "https://omlx.ai/benchmarks/performance/hhko08us",
   "tg_tokens": null,
   "tps": 107.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-hhko08us-ctx16384",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: 9dax4ydk, 9zej0sg0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 873.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 873.6 104.2 24.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hhko08us",
   "source_url": "https://omlx.ai/benchmarks/performance/hhko08us",
   "tg_tokens": null,
   "tps": 104.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-xlehdbxv-ctx1024",
   "model": "DeepSeek-Coder-V2-Lite-Instruct",
   "notes": "MLX; oMLX community benchmark trace; peak memory 9.1 GB at this context; duplicate traces of the same session: cmpthp1p; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1869.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-Coder-V2-Lite-Instruct-4bit-mlx on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 1,869 122.3 9.1 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xlehdbxv",
   "source_url": "https://omlx.ai/benchmarks/performance/xlehdbxv",
   "tg_tokens": null,
   "tps": 122.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-xlehdbxv-ctx4096",
   "model": "DeepSeek-Coder-V2-Lite-Instruct",
   "notes": "MLX; oMLX community benchmark trace; peak memory 10.2 GB at this context; duplicate traces of the same session: cmpthp1p; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2026.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-Coder-V2-Lite-Instruct-4bit-mlx on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 2,026 89.5 10.2 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xlehdbxv",
   "source_url": "https://omlx.ai/benchmarks/performance/xlehdbxv",
   "tg_tokens": null,
   "tps": 89.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-93uo6aoq-ctx1024",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 112.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 112.4 12.7 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/93uo6aoq",
   "source_url": "https://omlx.ai/benchmarks/performance/93uo6aoq",
   "tg_tokens": null,
   "tps": 12.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-93uo6aoq-ctx4096",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 142.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 142.5 12.7 19.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/93uo6aoq",
   "source_url": "https://omlx.ai/benchmarks/performance/93uo6aoq",
   "tg_tokens": null,
   "tps": 12.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-iabihpok-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 33.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 247.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 247.1 48.1 33.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iabihpok",
   "source_url": "https://omlx.ai/benchmarks/performance/iabihpok",
   "tg_tokens": null,
   "tps": 48.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-iabihpok-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 232.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 232.4 36.9 37.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iabihpok",
   "source_url": "https://omlx.ai/benchmarks/performance/iabihpok",
   "tg_tokens": null,
   "tps": 36.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-wzhj67rf-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: kw407o8o; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 355.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 355.1 38.2 19.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wzhj67rf",
   "source_url": "https://omlx.ai/benchmarks/performance/wzhj67rf",
   "tg_tokens": null,
   "tps": 38.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-wzhj67rf-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: kw407o8o; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 337.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 16k 337.5 35.4 20.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wzhj67rf",
   "source_url": "https://omlx.ai/benchmarks/performance/wzhj67rf",
   "tg_tokens": null,
   "tps": 35.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-wzhj67rf-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: kw407o8o; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 312.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 312.6 32.5 22.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wzhj67rf",
   "source_url": "https://omlx.ai/benchmarks/performance/wzhj67rf",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-k7b5n6e3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: kbq83gey, vwokos9k; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 668.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 668.6 29.6 19.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k7b5n6e3",
   "source_url": "https://omlx.ai/benchmarks/performance/k7b5n6e3",
   "tg_tokens": null,
   "tps": 29.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-k7b5n6e3-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: kbq83gey, vwokos9k; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 687.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 687.9 41.0 19.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k7b5n6e3",
   "source_url": "https://omlx.ai/benchmarks/performance/k7b5n6e3",
   "tg_tokens": null,
   "tps": 41.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-k7b5n6e3-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: kbq83gey, vwokos9k; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 494.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 494.9 31.4 23.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k7b5n6e3",
   "source_url": "https://omlx.ai/benchmarks/performance/k7b5n6e3",
   "tg_tokens": null,
   "tps": 31.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-y3cgzvs1-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: a1h5kvdo, fjq7jkkq, tfgny73h; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 906.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 906.0 83.7 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y3cgzvs1",
   "source_url": "https://omlx.ai/benchmarks/performance/y3cgzvs1",
   "tg_tokens": null,
   "tps": 83.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-y3cgzvs1-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: a1h5kvdo, fjq7jkkq, tfgny73h; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1049.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,049 90.6 23.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y3cgzvs1",
   "source_url": "https://omlx.ai/benchmarks/performance/y3cgzvs1",
   "tg_tokens": null,
   "tps": 90.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-y3cgzvs1-ctx8192",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: a1h5kvdo, fjq7jkkq, tfgny73h; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 947.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 947.7 103.5 23.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y3cgzvs1",
   "source_url": "https://omlx.ai/benchmarks/performance/y3cgzvs1",
   "tg_tokens": null,
   "tps": 103.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-y3cgzvs1-ctx16384",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: a1h5kvdo, fjq7jkkq, tfgny73h; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 793.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 793.3 100.6 24.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y3cgzvs1",
   "source_url": "https://omlx.ai/benchmarks/performance/y3cgzvs1",
   "tg_tokens": null,
   "tps": 100.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-j3qkxreo-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: au9cix27, m5tmyfvv, t009ut32; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 908.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 908.5 91.8 21.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j3qkxreo",
   "source_url": "https://omlx.ai/benchmarks/performance/j3qkxreo",
   "tg_tokens": null,
   "tps": 91.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-j3qkxreo-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: au9cix27, m5tmyfvv, t009ut32; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1052.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,052 89.6 23.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j3qkxreo",
   "source_url": "https://omlx.ai/benchmarks/performance/j3qkxreo",
   "tg_tokens": null,
   "tps": 89.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-j3qkxreo-ctx8192",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: au9cix27, m5tmyfvv, t009ut32; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 988.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 988.8 106.7 23.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j3qkxreo",
   "source_url": "https://omlx.ai/benchmarks/performance/j3qkxreo",
   "tg_tokens": null,
   "tps": 106.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-j3qkxreo-ctx16384",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: au9cix27, m5tmyfvv, t009ut32; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 874.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 874.5 104.3 24.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j3qkxreo",
   "source_url": "https://omlx.ai/benchmarks/performance/j3qkxreo",
   "tg_tokens": null,
   "tps": 104.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-nyhmacvb-ctx16384",
   "model": "DeepSeek-R1-Distill-Qwen-7B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 9.0 GB; TTFT 653211 ms; duplicate traces of the same session: none; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 25.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-R1-Distill-Qwen-7B on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 8bit; 16k tokens 25.1 PP tok/s 1.9 TG tok/s 653211 TTFT (ms) 9.0 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nyhmacvb",
   "source_url": "https://omlx.ai/benchmarks/performance/nyhmacvb",
   "tg_tokens": null,
   "tps": 1.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-b4ceqn3f-ctx1024",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 112.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 112.4 12.7 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b4ceqn3f",
   "source_url": "https://omlx.ai/benchmarks/performance/b4ceqn3f",
   "tg_tokens": null,
   "tps": 12.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-b4ceqn3f-ctx4096",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 142.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 142.5 12.7 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b4ceqn3f",
   "source_url": "https://omlx.ai/benchmarks/performance/b4ceqn3f",
   "tg_tokens": null,
   "tps": 12.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-pyh8iez3-ctx4096",
   "model": "Qwen3.5-9B-Uncensored-HauhauCS-Aggressive",
   "notes": "oMLX community benchmark trace; peak memory 6.1 GB at this context; duplicate traces of the same session: twiutaea; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 380.4,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.5-9B-Uncensored-HauhauCS-Aggressive-MLX-mxfp4 on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 380.4 51.6 6.1 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pyh8iez3",
   "source_url": "https://omlx.ai/benchmarks/performance/pyh8iez3",
   "tg_tokens": null,
   "tps": 51.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-pyh8iez3-ctx16384",
   "model": "Qwen3.5-9B-Uncensored-HauhauCS-Aggressive",
   "notes": "oMLX community benchmark trace; peak memory 6.6 GB at this context; duplicate traces of the same session: twiutaea; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 358.4,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.5-9B-Uncensored-HauhauCS-Aggressive-MLX-mxfp4 on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 16k 358.4 46.7 6.6 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pyh8iez3",
   "source_url": "https://omlx.ai/benchmarks/performance/pyh8iez3",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-yh9ivcx1-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 33.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 247.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 247.1 48.1 33.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yh9ivcx1",
   "source_url": "https://omlx.ai/benchmarks/performance/yh9ivcx1",
   "tg_tokens": null,
   "tps": 48.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-yh9ivcx1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 232.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 232.4 36.9 37.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yh9ivcx1",
   "source_url": "https://omlx.ai/benchmarks/performance/yh9ivcx1",
   "tg_tokens": null,
   "tps": 36.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-7q3lv8jc-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 16.6 GB; TTFT 6471 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 158.2,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 1k tokens 158.2 PP tok/s 20.2 TG tok/s 6471 TTFT (ms) 16.6 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7q3lv8jc",
   "source_url": "https://omlx.ai/benchmarks/performance/7q3lv8jc",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3v5vzmlz-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: q7egioof; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 820.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 820.2 64.8 19.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3v5vzmlz",
   "source_url": "https://omlx.ai/benchmarks/performance/3v5vzmlz",
   "tg_tokens": null,
   "tps": 64.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3v5vzmlz-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: q7egioof; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 781.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 781.6 92.9 19.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3v5vzmlz",
   "source_url": "https://omlx.ai/benchmarks/performance/3v5vzmlz",
   "tg_tokens": null,
   "tps": 92.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-atll39wr-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: mlwp7ev3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 948.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 948.3 23.3 18.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/atll39wr",
   "source_url": "https://omlx.ai/benchmarks/performance/atll39wr",
   "tg_tokens": null,
   "tps": 23.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-atll39wr-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: mlwp7ev3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 814.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 814.0 52.6 19.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/atll39wr",
   "source_url": "https://omlx.ai/benchmarks/performance/atll39wr",
   "tg_tokens": null,
   "tps": 52.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-r8jtvf9x-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: mkltfq2r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 885.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 885.3 86.9 19.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r8jtvf9x",
   "source_url": "https://omlx.ai/benchmarks/performance/r8jtvf9x",
   "tg_tokens": null,
   "tps": 86.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-r8jtvf9x-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: mkltfq2r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 818.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 818.0 66.9 19.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r8jtvf9x",
   "source_url": "https://omlx.ai/benchmarks/performance/r8jtvf9x",
   "tg_tokens": null,
   "tps": 66.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0tp40l2v-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 17.7 GB; TTFT 2626 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 390.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k tokens 390.0 PP tok/s 32.9 TG tok/s 2626 TTFT (ms) 17.7 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0tp40l2v",
   "source_url": "https://omlx.ai/benchmarks/performance/0tp40l2v",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8btltphw-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 99.0 GB; TTFT 2936 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 348.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 348.8 PP tok/s 82.6 TG tok/s 2936 TTFT (ms) 99.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8btltphw",
   "source_url": "https://omlx.ai/benchmarks/performance/8btltphw",
   "tg_tokens": null,
   "tps": 82.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-oilxulm5-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 99.0 GB; TTFT 500 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2048.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 2,048 PP tok/s 82.6 TG tok/s 500 TTFT (ms) 99.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oilxulm5",
   "source_url": "https://omlx.ai/benchmarks/performance/oilxulm5",
   "tg_tokens": null,
   "tps": 82.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-38mkd0e4-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 101.0 GB at this context; duplicate traces of the same session: jpdi1het; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 318.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 318.9 23.4 101.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/38mkd0e4",
   "source_url": "https://omlx.ai/benchmarks/performance/38mkd0e4",
   "tg_tokens": null,
   "tps": 23.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-11",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-38mkd0e4-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 101.6 GB at this context; duplicate traces of the same session: jpdi1het; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 360.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 360.9 42.9 101.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/38mkd0e4",
   "source_url": "https://omlx.ai/benchmarks/performance/38mkd0e4",
   "tg_tokens": null,
   "tps": 42.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-4mjkspmg-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 139.4 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2021.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 1k 2,021 91.6 139.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4mjkspmg",
   "source_url": "https://omlx.ai/benchmarks/performance/4mjkspmg",
   "tg_tokens": null,
   "tps": 91.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-4mjkspmg-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 140.0 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3632.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 4k 3,632 86.0 140.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4mjkspmg",
   "source_url": "https://omlx.ai/benchmarks/performance/4mjkspmg",
   "tg_tokens": null,
   "tps": 86.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-4mjkspmg-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 141.1 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4106.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 8k 4,106 86.6 141.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4mjkspmg",
   "source_url": "https://omlx.ai/benchmarks/performance/4mjkspmg",
   "tg_tokens": null,
   "tps": 86.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-4mjkspmg-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 141.8 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4099.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 32k 4,099 80.0 141.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4mjkspmg",
   "source_url": "https://omlx.ai/benchmarks/performance/4mjkspmg",
   "tg_tokens": null,
   "tps": 80.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-4mjkspmg-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.8 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4042.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 64k 4,042 76.6 142.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4mjkspmg",
   "source_url": "https://omlx.ai/benchmarks/performance/4mjkspmg",
   "tg_tokens": null,
   "tps": 76.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-4mjkspmg-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.0 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3905.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 128k 3,905 75.2 145.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4mjkspmg",
   "source_url": "https://omlx.ai/benchmarks/performance/4mjkspmg",
   "tg_tokens": null,
   "tps": 75.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-4mjkspmg-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 148.7 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3834.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 195k 3,834 74.7 148.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4mjkspmg",
   "source_url": "https://omlx.ai/benchmarks/performance/4mjkspmg",
   "tg_tokens": null,
   "tps": 74.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-imrzn38h-ctx32768",
   "model": "ThinkingCap-Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 42.2 GB; TTFT 33751 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 970.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "ThinkingCap-Qwen3.8-27B-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 32k tokens 970.9 PP tok/s 91.1 TG tok/s 33751 TTFT (ms) 42.2 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/imrzn38h",
   "source_url": "https://omlx.ai/benchmarks/performance/imrzn38h",
   "tg_tokens": null,
   "tps": 91.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-11",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-9dhf2u44-ctx32768",
   "model": "ThinkingCap-Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 32.3 GB; TTFT 19088 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1717.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "ThinkingCap-Qwen3.8-27B-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 32k tokens 1,717 PP tok/s 30.6 TG tok/s 19088 TTFT (ms) 32.3 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9dhf2u44",
   "source_url": "https://omlx.ai/benchmarks/performance/9dhf2u44",
   "tg_tokens": null,
   "tps": 30.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-8ghgxm4j-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: vgs1i95d; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 432.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 432.5 17.5 17.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8ghgxm4j",
   "source_url": "https://omlx.ai/benchmarks/performance/8ghgxm4j",
   "tg_tokens": null,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-8ghgxm4j-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: vgs1i95d; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 401.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 401.5 16.2 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8ghgxm4j",
   "source_url": "https://omlx.ai/benchmarks/performance/8ghgxm4j",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-mx4hj25l-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 17.1 GB; TTFT 17922 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 228.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k tokens 228.5 PP tok/s 9.8 TG tok/s 17922 TTFT (ms) 17.1 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mx4hj25l",
   "source_url": "https://omlx.ai/benchmarks/performance/mx4hj25l",
   "tg_tokens": null,
   "tps": 9.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-0pbjmero-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: ejj85lr0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 405.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 405.3 18.1 16.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0pbjmero",
   "source_url": "https://omlx.ai/benchmarks/performance/0pbjmero",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0pbjmero-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: ejj85lr0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 427.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 427.0 17.9 17.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0pbjmero",
   "source_url": "https://omlx.ai/benchmarks/performance/0pbjmero",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-bq4bmp89-ctx1024",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "MLX; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 277.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 277.0 20.6 21.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bq4bmp89",
   "source_url": "https://omlx.ai/benchmarks/performance/bq4bmp89",
   "tg_tokens": null,
   "tps": 20.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-bq4bmp89-ctx4096",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "MLX; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 287.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 287.3 17.4 23.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bq4bmp89",
   "source_url": "https://omlx.ai/benchmarks/performance/bq4bmp89",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-90v8g7yr-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 71.1 GB; TTFT 851 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1204.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,204 PP tok/s 72.6 TG tok/s 851 TTFT (ms) 71.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/90v8g7yr",
   "source_url": "https://omlx.ai/benchmarks/performance/90v8g7yr",
   "tg_tokens": null,
   "tps": 72.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-p4wi0zgl-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 71.1 GB; TTFT 1024 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1000.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,000 PP tok/s 49.8 TG tok/s 1024 TTFT (ms) 71.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p4wi0zgl",
   "source_url": "https://omlx.ai/benchmarks/performance/p4wi0zgl",
   "tg_tokens": null,
   "tps": 49.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xjnvdh9s-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 71.1 GB; TTFT 852 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1202.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,202 PP tok/s 72.9 TG tok/s 852 TTFT (ms) 71.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xjnvdh9s",
   "source_url": "https://omlx.ai/benchmarks/performance/xjnvdh9s",
   "tg_tokens": null,
   "tps": 72.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-p4t8bno3-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 139.4 GB at this context; duplicate traces of the same session: 0q14h0k4, 970fr6w5, gqf6e206, lw748mce, w4arj90u; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2021.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 1k 2,021 91.6 139.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p4t8bno3",
   "source_url": "https://omlx.ai/benchmarks/performance/p4t8bno3",
   "tg_tokens": null,
   "tps": 91.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-p4t8bno3-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 140.0 GB at this context; duplicate traces of the same session: 0q14h0k4, 970fr6w5, gqf6e206, lw748mce, w4arj90u; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3632.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 4k 3,632 86.0 140.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p4t8bno3",
   "source_url": "https://omlx.ai/benchmarks/performance/p4t8bno3",
   "tg_tokens": null,
   "tps": 86.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-p4t8bno3-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 141.1 GB at this context; duplicate traces of the same session: 0q14h0k4, 970fr6w5, gqf6e206, lw748mce, w4arj90u; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4106.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 8k 4,106 86.6 141.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p4t8bno3",
   "source_url": "https://omlx.ai/benchmarks/performance/p4t8bno3",
   "tg_tokens": null,
   "tps": 86.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-p4t8bno3-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 141.8 GB at this context; duplicate traces of the same session: 0q14h0k4, 970fr6w5, gqf6e206, lw748mce, w4arj90u; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4099.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 32k 4,099 80.0 141.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p4t8bno3",
   "source_url": "https://omlx.ai/benchmarks/performance/p4t8bno3",
   "tg_tokens": null,
   "tps": 80.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-p4t8bno3-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.8 GB at this context; duplicate traces of the same session: 0q14h0k4, 970fr6w5, gqf6e206, lw748mce, w4arj90u; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4042.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 64k 4,042 76.6 142.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p4t8bno3",
   "source_url": "https://omlx.ai/benchmarks/performance/p4t8bno3",
   "tg_tokens": null,
   "tps": 76.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-p4t8bno3-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.0 GB at this context; duplicate traces of the same session: 0q14h0k4, 970fr6w5, gqf6e206, lw748mce, w4arj90u; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3905.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 128k 3,905 75.2 145.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p4t8bno3",
   "source_url": "https://omlx.ai/benchmarks/performance/p4t8bno3",
   "tg_tokens": null,
   "tps": 75.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-p4t8bno3-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 148.7 GB at this context; duplicate traces of the same session: 0q14h0k4, 970fr6w5, gqf6e206, lw748mce, w4arj90u; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3834.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 195k 3,834 74.7 148.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p4t8bno3",
   "source_url": "https://omlx.ai/benchmarks/performance/p4t8bno3",
   "tg_tokens": null,
   "tps": 74.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-96dugvvw-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.8 GB at this context; duplicate traces of the same session: 0ajadb00, 2lrv0fne, l7wp49er, nnlik470; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1890.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 1k 1,890 127.3 142.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/96dugvvw",
   "source_url": "https://omlx.ai/benchmarks/performance/96dugvvw",
   "tg_tokens": null,
   "tps": 127.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-96dugvvw-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: 0ajadb00, 2lrv0fne, l7wp49er, nnlik470; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3516.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 4k 3,516 109.8 143.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/96dugvvw",
   "source_url": "https://omlx.ai/benchmarks/performance/96dugvvw",
   "tg_tokens": null,
   "tps": 109.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-96dugvvw-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 144.5 GB at this context; duplicate traces of the same session: 0ajadb00, 2lrv0fne, l7wp49er, nnlik470; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3966.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 8k 3,966 112.0 144.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/96dugvvw",
   "source_url": "https://omlx.ai/benchmarks/performance/96dugvvw",
   "tg_tokens": null,
   "tps": 112.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-96dugvvw-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.7 GB at this context; duplicate traces of the same session: 0ajadb00, 2lrv0fne, l7wp49er, nnlik470; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3979.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 32k 3,979 77.2 145.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/96dugvvw",
   "source_url": "https://omlx.ai/benchmarks/performance/96dugvvw",
   "tg_tokens": null,
   "tps": 77.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-96dugvvw-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 146.8 GB at this context; duplicate traces of the same session: 0ajadb00, 2lrv0fne, l7wp49er, nnlik470; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3899.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 64k 3,899 110.6 146.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/96dugvvw",
   "source_url": "https://omlx.ai/benchmarks/performance/96dugvvw",
   "tg_tokens": null,
   "tps": 110.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-96dugvvw-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 150.7 GB at this context; duplicate traces of the same session: 0ajadb00, 2lrv0fne, l7wp49er, nnlik470; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3809.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 128k 3,809 96.0 150.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/96dugvvw",
   "source_url": "https://omlx.ai/benchmarks/performance/96dugvvw",
   "tg_tokens": null,
   "tps": 96.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-96dugvvw-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 153.3 GB at this context; duplicate traces of the same session: 0ajadb00, 2lrv0fne, l7wp49er, nnlik470; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3710.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 195k 3,710 79.4 153.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/96dugvvw",
   "source_url": "https://omlx.ai/benchmarks/performance/96dugvvw",
   "tg_tokens": null,
   "tps": 79.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-ho1a0jwn-ctx1024",
   "model": "Ling-3.0-tiny",
   "notes": "oMLX community benchmark trace; peak memory 5.1 GB at this context; duplicate traces of the same session: 5vdfqkzq; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1210.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-tiny-oQ4e on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 1,210 101.3 5.1 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ho1a0jwn",
   "source_url": "https://omlx.ai/benchmarks/performance/ho1a0jwn",
   "tg_tokens": null,
   "tps": 101.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-ho1a0jwn-ctx4096",
   "model": "Ling-3.0-tiny",
   "notes": "oMLX community benchmark trace; peak memory 5.5 GB at this context; duplicate traces of the same session: 5vdfqkzq; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1245.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-tiny-oQ4e on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 1,245 95.3 5.5 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ho1a0jwn",
   "source_url": "https://omlx.ai/benchmarks/performance/ho1a0jwn",
   "tg_tokens": null,
   "tps": 95.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-13j1venm-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 406.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 406.2 18.2 16.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/13j1venm",
   "source_url": "https://omlx.ai/benchmarks/performance/13j1venm",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-13j1venm-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 425.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 425.4 17.5 17.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/13j1venm",
   "source_url": "https://omlx.ai/benchmarks/performance/13j1venm",
   "tg_tokens": null,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ecozd79e-ctx1024",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "MLX; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 277.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 277.0 20.6 21.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ecozd79e",
   "source_url": "https://omlx.ai/benchmarks/performance/ecozd79e",
   "tg_tokens": null,
   "tps": 20.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ecozd79e-ctx4096",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "MLX; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 287.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 287.3 17.4 23.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ecozd79e",
   "source_url": "https://omlx.ai/benchmarks/performance/ecozd79e",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-t1t1v2mw-ctx1024",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "MLX; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 8u3jl56b; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 276.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 276.6 21.0 21.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t1t1v2mw",
   "source_url": "https://omlx.ai/benchmarks/performance/t1t1v2mw",
   "tg_tokens": null,
   "tps": 21.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-t1t1v2mw-ctx4096",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "MLX; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: 8u3jl56b; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 286.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 286.6 21.0 22.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t1t1v2mw",
   "source_url": "https://omlx.ai/benchmarks/performance/t1t1v2mw",
   "tg_tokens": null,
   "tps": 21.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-hu060mxl-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.7 GB at this context; duplicate traces of the same session: 90wt5psn, gz4zcwy3, m5uzwq5j, mml2rg17, zphm3e9r; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1909.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 1k 1,909 128.9 142.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hu060mxl",
   "source_url": "https://omlx.ai/benchmarks/performance/hu060mxl",
   "tg_tokens": null,
   "tps": 128.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-hu060mxl-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: 90wt5psn, gz4zcwy3, m5uzwq5j, mml2rg17, zphm3e9r; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3526.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 4k 3,526 116.5 143.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hu060mxl",
   "source_url": "https://omlx.ai/benchmarks/performance/hu060mxl",
   "tg_tokens": null,
   "tps": 116.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-hu060mxl-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 144.5 GB at this context; duplicate traces of the same session: 90wt5psn, gz4zcwy3, m5uzwq5j, mml2rg17, zphm3e9r; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3980.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 8k 3,980 111.7 144.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hu060mxl",
   "source_url": "https://omlx.ai/benchmarks/performance/hu060mxl",
   "tg_tokens": null,
   "tps": 111.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-hu060mxl-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 146.8 GB at this context; duplicate traces of the same session: 90wt5psn, gz4zcwy3, m5uzwq5j, mml2rg17, zphm3e9r; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3909.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 64k 3,909 114.0 146.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hu060mxl",
   "source_url": "https://omlx.ai/benchmarks/performance/hu060mxl",
   "tg_tokens": null,
   "tps": 114.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-hu060mxl-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 151.1 GB at this context; duplicate traces of the same session: 90wt5psn, gz4zcwy3, m5uzwq5j, mml2rg17, zphm3e9r; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3810.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 128k 3,810 98.7 151.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hu060mxl",
   "source_url": "https://omlx.ai/benchmarks/performance/hu060mxl",
   "tg_tokens": null,
   "tps": 98.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-hu060mxl-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 153.3 GB at this context; duplicate traces of the same session: 90wt5psn, gz4zcwy3, m5uzwq5j, mml2rg17, zphm3e9r; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3710.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 195k 3,710 83.9 153.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hu060mxl",
   "source_url": "https://omlx.ai/benchmarks/performance/hu060mxl",
   "tg_tokens": null,
   "tps": 83.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-kr4xtuao-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.8 GB at this context; duplicate traces of the same session: 5j2qw53z; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1890.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 1k 1,890 127.3 142.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kr4xtuao",
   "source_url": "https://omlx.ai/benchmarks/performance/kr4xtuao",
   "tg_tokens": null,
   "tps": 127.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-kr4xtuao-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: 5j2qw53z; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3516.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 4k 3,516 109.8 143.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kr4xtuao",
   "source_url": "https://omlx.ai/benchmarks/performance/kr4xtuao",
   "tg_tokens": null,
   "tps": 109.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-kr4xtuao-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 144.5 GB at this context; duplicate traces of the same session: 5j2qw53z; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3966.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 8k 3,966 112.0 144.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kr4xtuao",
   "source_url": "https://omlx.ai/benchmarks/performance/kr4xtuao",
   "tg_tokens": null,
   "tps": 112.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-kr4xtuao-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.7 GB at this context; duplicate traces of the same session: 5j2qw53z; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3979.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 32k 3,979 77.2 145.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kr4xtuao",
   "source_url": "https://omlx.ai/benchmarks/performance/kr4xtuao",
   "tg_tokens": null,
   "tps": 77.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-kr4xtuao-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 146.8 GB at this context; duplicate traces of the same session: 5j2qw53z; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3899.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 64k 3,899 110.6 146.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kr4xtuao",
   "source_url": "https://omlx.ai/benchmarks/performance/kr4xtuao",
   "tg_tokens": null,
   "tps": 110.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-kr4xtuao-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 150.7 GB at this context; duplicate traces of the same session: 5j2qw53z; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3809.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 128k 3,809 96.0 150.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kr4xtuao",
   "source_url": "https://omlx.ai/benchmarks/performance/kr4xtuao",
   "tg_tokens": null,
   "tps": 96.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-kr4xtuao-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 153.3 GB at this context; duplicate traces of the same session: 5j2qw53z; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3710.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 195k 3,710 79.4 153.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kr4xtuao",
   "source_url": "https://omlx.ai/benchmarks/performance/kr4xtuao",
   "tg_tokens": null,
   "tps": 79.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-agagpz4v-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.7 GB at this context; duplicate traces of the same session: 14po7m9o, 5k287lai, a7txur4t, d8wtln9v, zo4hz6qb; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1851.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 1k 1,851 119.1 142.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/agagpz4v",
   "source_url": "https://omlx.ai/benchmarks/performance/agagpz4v",
   "tg_tokens": null,
   "tps": 119.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-agagpz4v-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: 14po7m9o, 5k287lai, a7txur4t, d8wtln9v, zo4hz6qb; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3389.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 4k 3,389 120.1 143.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/agagpz4v",
   "source_url": "https://omlx.ai/benchmarks/performance/agagpz4v",
   "tg_tokens": null,
   "tps": 120.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-agagpz4v-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 144.5 GB at this context; duplicate traces of the same session: 14po7m9o, 5k287lai, a7txur4t, d8wtln9v, zo4hz6qb; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3970.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 8k 3,970 128.1 144.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/agagpz4v",
   "source_url": "https://omlx.ai/benchmarks/performance/agagpz4v",
   "tg_tokens": null,
   "tps": 128.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-agagpz4v-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.7 GB at this context; duplicate traces of the same session: 14po7m9o, 5k287lai, a7txur4t, d8wtln9v, zo4hz6qb; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3984.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 32k 3,984 91.1 145.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/agagpz4v",
   "source_url": "https://omlx.ai/benchmarks/performance/agagpz4v",
   "tg_tokens": null,
   "tps": 91.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-agagpz4v-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 146.8 GB at this context; duplicate traces of the same session: 14po7m9o, 5k287lai, a7txur4t, d8wtln9v, zo4hz6qb; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3918.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 64k 3,918 115.5 146.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/agagpz4v",
   "source_url": "https://omlx.ai/benchmarks/performance/agagpz4v",
   "tg_tokens": null,
   "tps": 115.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-agagpz4v-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 150.7 GB at this context; duplicate traces of the same session: 14po7m9o, 5k287lai, a7txur4t, d8wtln9v, zo4hz6qb; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3821.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 128k 3,821 96.2 150.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/agagpz4v",
   "source_url": "https://omlx.ai/benchmarks/performance/agagpz4v",
   "tg_tokens": null,
   "tps": 96.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-agagpz4v-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 153.3 GB at this context; duplicate traces of the same session: 14po7m9o, 5k287lai, a7txur4t, d8wtln9v, zo4hz6qb; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3716.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 195k 3,716 79.0 153.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/agagpz4v",
   "source_url": "https://omlx.ai/benchmarks/performance/agagpz4v",
   "tg_tokens": null,
   "tps": 79.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0apbon1s-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 249.1 GB at this context; duplicate traces of the same session: 5w2ytpra, 9fqj3axl, frs5g3x8, gm2s2bdr, n2x2fa0w, w889be0f; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 490.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 1k 490.7 42.3 249.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0apbon1s",
   "source_url": "https://omlx.ai/benchmarks/performance/0apbon1s",
   "tg_tokens": null,
   "tps": 42.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0apbon1s-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 250.6 GB at this context; duplicate traces of the same session: 5w2ytpra, 9fqj3axl, frs5g3x8, gm2s2bdr, n2x2fa0w, w889be0f; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 529.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 8k 529.0 44.0 250.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0apbon1s",
   "source_url": "https://omlx.ai/benchmarks/performance/0apbon1s",
   "tg_tokens": null,
   "tps": 44.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0apbon1s-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 250.7 GB at this context; duplicate traces of the same session: 5w2ytpra, 9fqj3axl, frs5g3x8, gm2s2bdr, n2x2fa0w, w889be0f; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 520.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 16k 520.1 73.0 250.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0apbon1s",
   "source_url": "https://omlx.ai/benchmarks/performance/0apbon1s",
   "tg_tokens": null,
   "tps": 73.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0apbon1s-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 251.0 GB at this context; duplicate traces of the same session: 5w2ytpra, 9fqj3axl, frs5g3x8, gm2s2bdr, n2x2fa0w, w889be0f; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 513.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 32k 513.2 54.0 251.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0apbon1s",
   "source_url": "https://omlx.ai/benchmarks/performance/0apbon1s",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0apbon1s-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 251.7 GB at this context; duplicate traces of the same session: 5w2ytpra, 9fqj3axl, frs5g3x8, gm2s2bdr, n2x2fa0w, w889be0f; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 505.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 64k 505.9 46.6 251.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0apbon1s",
   "source_url": "https://omlx.ai/benchmarks/performance/0apbon1s",
   "tg_tokens": null,
   "tps": 46.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0apbon1s-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 252.9 GB at this context; duplicate traces of the same session: 5w2ytpra, 9fqj3axl, frs5g3x8, gm2s2bdr, n2x2fa0w, w889be0f; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 495.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 128k 495.7 72.8 252.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0apbon1s",
   "source_url": "https://omlx.ai/benchmarks/performance/0apbon1s",
   "tg_tokens": null,
   "tps": 72.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0apbon1s-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 254.3 GB at this context; duplicate traces of the same session: 5w2ytpra, 9fqj3axl, frs5g3x8, gm2s2bdr, n2x2fa0w, w889be0f; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 488.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 195k 488.1 40.7 254.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0apbon1s",
   "source_url": "https://omlx.ai/benchmarks/performance/0apbon1s",
   "tg_tokens": null,
   "tps": 40.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-91e847yf-ctx1024",
   "model": "NeoHorse-1-4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 3.8 GB at this context; duplicate traces of the same session: mhljobyb; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 410.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NeoHorse-1-4B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 410.8 53.9 3.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/91e847yf",
   "source_url": "https://omlx.ai/benchmarks/performance/91e847yf",
   "tg_tokens": null,
   "tps": 53.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-91e847yf-ctx4096",
   "model": "NeoHorse-1-4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.0 GB at this context; duplicate traces of the same session: mhljobyb; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 542.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NeoHorse-1-4B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 542.9 57.2 4.0 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/91e847yf",
   "source_url": "https://omlx.ai/benchmarks/performance/91e847yf",
   "tg_tokens": null,
   "tps": 57.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-iazaccaq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: lrsxeqmz, quesks85, usiivc12; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 928.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 928.2 33.5 16.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iazaccaq",
   "source_url": "https://omlx.ai/benchmarks/performance/iazaccaq",
   "tg_tokens": null,
   "tps": 33.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-iazaccaq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: lrsxeqmz, quesks85, usiivc12; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1015.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,015 32.9 17.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iazaccaq",
   "source_url": "https://omlx.ai/benchmarks/performance/iazaccaq",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-iazaccaq-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: lrsxeqmz, quesks85, usiivc12; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 873.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 873.8 31.0 17.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iazaccaq",
   "source_url": "https://omlx.ai/benchmarks/performance/iazaccaq",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-iazaccaq-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: lrsxeqmz, quesks85, usiivc12; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 695.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 695.0 25.8 20.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iazaccaq",
   "source_url": "https://omlx.ai/benchmarks/performance/iazaccaq",
   "tg_tokens": null,
   "tps": 25.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-3f8ecs9z-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 406.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 406.2 18.2 16.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3f8ecs9z",
   "source_url": "https://omlx.ai/benchmarks/performance/3f8ecs9z",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3f8ecs9z-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 425.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 425.4 17.5 17.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3f8ecs9z",
   "source_url": "https://omlx.ai/benchmarks/performance/3f8ecs9z",
   "tg_tokens": null,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3pph0e3v-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 44.9 GB; TTFT 413308 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 317.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k tokens 317.1 PP tok/s 64.3 TG tok/s 413308 TTFT (ms) 44.9 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3pph0e3v",
   "source_url": "https://omlx.ai/benchmarks/performance/3pph0e3v",
   "tg_tokens": null,
   "tps": 64.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wh85hjej-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 195k context; peak memory 29.6 GB; TTFT 944369 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 211.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k tokens 211.8 PP tok/s 9.7 TG tok/s 944369 TTFT (ms) 29.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wh85hjej",
   "source_url": "https://omlx.ai/benchmarks/performance/wh85hjej",
   "tg_tokens": null,
   "tps": 9.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-edrjneif-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 88.0 GB; TTFT 31258 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2097.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k tokens 2,097 PP tok/s 58.3 TG tok/s 31258 TTFT (ms) 88.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/edrjneif",
   "source_url": "https://omlx.ai/benchmarks/performance/edrjneif",
   "tg_tokens": null,
   "tps": 58.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-kjfckepo-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.7 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1851.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 1k 1,851 119.1 142.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kjfckepo",
   "source_url": "https://omlx.ai/benchmarks/performance/kjfckepo",
   "tg_tokens": null,
   "tps": 119.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-kjfckepo-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3389.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 4k 3,389 120.1 143.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kjfckepo",
   "source_url": "https://omlx.ai/benchmarks/performance/kjfckepo",
   "tg_tokens": null,
   "tps": 120.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-kjfckepo-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 144.5 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3970.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 8k 3,970 128.1 144.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kjfckepo",
   "source_url": "https://omlx.ai/benchmarks/performance/kjfckepo",
   "tg_tokens": null,
   "tps": 128.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-kjfckepo-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.7 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3984.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 32k 3,984 91.1 145.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kjfckepo",
   "source_url": "https://omlx.ai/benchmarks/performance/kjfckepo",
   "tg_tokens": null,
   "tps": 91.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-kjfckepo-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 146.8 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3918.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 64k 3,918 115.5 146.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kjfckepo",
   "source_url": "https://omlx.ai/benchmarks/performance/kjfckepo",
   "tg_tokens": null,
   "tps": 115.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-kjfckepo-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 150.7 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3821.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 128k 3,821 96.2 150.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kjfckepo",
   "source_url": "https://omlx.ai/benchmarks/performance/kjfckepo",
   "tg_tokens": null,
   "tps": 96.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-kjfckepo-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 153.3 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3716.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 195k 3,716 79.0 153.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kjfckepo",
   "source_url": "https://omlx.ai/benchmarks/performance/kjfckepo",
   "tg_tokens": null,
   "tps": 79.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-230mhlsw-ctx1024",
   "model": "RAZOR-19B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 9.2 GB at this context; duplicate traces of the same session: v4hccaqg; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 484.5,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "RAZOR-19B-A3B-oQ2e-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 2bit; 1k 484.5 54.4 9.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/230mhlsw",
   "source_url": "https://omlx.ai/benchmarks/performance/230mhlsw",
   "tg_tokens": null,
   "tps": 54.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-230mhlsw-ctx4096",
   "model": "RAZOR-19B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 9.8 GB at this context; duplicate traces of the same session: v4hccaqg; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 519.3,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "RAZOR-19B-A3B-oQ2e-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 2bit; 4k 519.3 56.3 9.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/230mhlsw",
   "source_url": "https://omlx.ai/benchmarks/performance/230mhlsw",
   "tg_tokens": null,
   "tps": 56.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-030wxl2p-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 20.6 GB; TTFT 6766 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4843.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k tokens 4,843 PP tok/s 118.7 TG tok/s 6766 TTFT (ms) 20.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/030wxl2p",
   "source_url": "https://omlx.ai/benchmarks/performance/030wxl2p",
   "tg_tokens": null,
   "tps": 118.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n8807z3z-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 56dbgx3n, ghlga12g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 934.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 934.3 33.7 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n8807z3z",
   "source_url": "https://omlx.ai/benchmarks/performance/n8807z3z",
   "tg_tokens": null,
   "tps": 33.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n8807z3z-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 56dbgx3n, ghlga12g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1009.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,009 33.1 17.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n8807z3z",
   "source_url": "https://omlx.ai/benchmarks/performance/n8807z3z",
   "tg_tokens": null,
   "tps": 33.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n8807z3z-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 56dbgx3n, ghlga12g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1017.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,017 32.2 17.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n8807z3z",
   "source_url": "https://omlx.ai/benchmarks/performance/n8807z3z",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vqbui75r-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: ciuv4p1g, lpbnybml; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 868.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 868.0 33.5 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vqbui75r",
   "source_url": "https://omlx.ai/benchmarks/performance/vqbui75r",
   "tg_tokens": null,
   "tps": 33.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vqbui75r-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: ciuv4p1g, lpbnybml; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 948.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 948.8 33.0 17.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vqbui75r",
   "source_url": "https://omlx.ai/benchmarks/performance/vqbui75r",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vqbui75r-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: ciuv4p1g, lpbnybml; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1010.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,010 32.0 17.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vqbui75r",
   "source_url": "https://omlx.ai/benchmarks/performance/vqbui75r",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-uio96th9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: er5d5ahr, q4duyq4k; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 401.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 401.3 46.8 18.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uio96th9",
   "source_url": "https://omlx.ai/benchmarks/performance/uio96th9",
   "tg_tokens": null,
   "tps": 46.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-uio96th9-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: er5d5ahr, q4duyq4k; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 385.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 385.5 34.9 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uio96th9",
   "source_url": "https://omlx.ai/benchmarks/performance/uio96th9",
   "tg_tokens": null,
   "tps": 34.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-uio96th9-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: er5d5ahr, q4duyq4k; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 353.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 353.6 31.5 20.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uio96th9",
   "source_url": "https://omlx.ai/benchmarks/performance/uio96th9",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-0c6wvm16-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.3 GB at this context; duplicate traces of the same session: 0vmxpf1w, 8608f73m, csa72pne, prl9oevw, w6ut050m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 432.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 432.2 53.6 69.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0c6wvm16",
   "source_url": "https://omlx.ai/benchmarks/performance/0c6wvm16",
   "tg_tokens": null,
   "tps": 53.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-0c6wvm16-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.9 GB at this context; duplicate traces of the same session: 0vmxpf1w, 8608f73m, csa72pne, prl9oevw, w6ut050m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 680.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 680.0 55.0 69.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0c6wvm16",
   "source_url": "https://omlx.ai/benchmarks/performance/0c6wvm16",
   "tg_tokens": null,
   "tps": 55.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-0c6wvm16-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: 0vmxpf1w, 8608f73m, csa72pne, prl9oevw, w6ut050m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 614.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 614.4 39.3 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0c6wvm16",
   "source_url": "https://omlx.ai/benchmarks/performance/0c6wvm16",
   "tg_tokens": null,
   "tps": 39.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-0c6wvm16-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: 0vmxpf1w, 8608f73m, csa72pne, prl9oevw, w6ut050m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 425.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 425.1 30.3 71.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0c6wvm16",
   "source_url": "https://omlx.ai/benchmarks/performance/0c6wvm16",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-0c6wvm16-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.9 GB at this context; duplicate traces of the same session: 0vmxpf1w, 8608f73m, csa72pne, prl9oevw, w6ut050m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 493.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 493.7 38.3 74.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0c6wvm16",
   "source_url": "https://omlx.ai/benchmarks/performance/0c6wvm16",
   "tg_tokens": null,
   "tps": 38.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-0c6wvm16-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.9 GB at this context; duplicate traces of the same session: 0vmxpf1w, 8608f73m, csa72pne, prl9oevw, w6ut050m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 522.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 522.0 42.2 78.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0c6wvm16",
   "source_url": "https://omlx.ai/benchmarks/performance/0c6wvm16",
   "tg_tokens": null,
   "tps": 42.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-z63qyyk2-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 139.4 GB at this context; duplicate traces of the same session: lxp25ouq; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2007.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 1k 2,007 91.9 139.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z63qyyk2",
   "source_url": "https://omlx.ai/benchmarks/performance/z63qyyk2",
   "tg_tokens": null,
   "tps": 91.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-z63qyyk2-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 140.0 GB at this context; duplicate traces of the same session: lxp25ouq; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3620.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 4k 3,620 86.0 140.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z63qyyk2",
   "source_url": "https://omlx.ai/benchmarks/performance/z63qyyk2",
   "tg_tokens": null,
   "tps": 86.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-z63qyyk2-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 141.1 GB at this context; duplicate traces of the same session: lxp25ouq; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4100.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 8k 4,100 86.0 141.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z63qyyk2",
   "source_url": "https://omlx.ai/benchmarks/performance/z63qyyk2",
   "tg_tokens": null,
   "tps": 86.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-z63qyyk2-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 141.3 GB at this context; duplicate traces of the same session: lxp25ouq; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4108.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 16k 4,108 84.3 141.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z63qyyk2",
   "source_url": "https://omlx.ai/benchmarks/performance/z63qyyk2",
   "tg_tokens": null,
   "tps": 84.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-z63qyyk2-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.8 GB at this context; duplicate traces of the same session: lxp25ouq; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4037.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 64k 4,037 76.5 142.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z63qyyk2",
   "source_url": "https://omlx.ai/benchmarks/performance/z63qyyk2",
   "tg_tokens": null,
   "tps": 76.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-z63qyyk2-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.0 GB at this context; duplicate traces of the same session: lxp25ouq; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3933.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 128k 3,933 75.2 145.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z63qyyk2",
   "source_url": "https://omlx.ai/benchmarks/performance/z63qyyk2",
   "tg_tokens": null,
   "tps": 75.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-z63qyyk2-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 148.7 GB at this context; duplicate traces of the same session: lxp25ouq; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3834.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 195k 3,834 74.6 148.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z63qyyk2",
   "source_url": "https://omlx.ai/benchmarks/performance/z63qyyk2",
   "tg_tokens": null,
   "tps": 74.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-imt0zaqh-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: newf1z8l; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 140.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 32 GB \u00b7 4bit; 1k 140.6 13.5 19.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/imt0zaqh",
   "source_url": "https://omlx.ai/benchmarks/performance/imt0zaqh",
   "tg_tokens": null,
   "tps": 13.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-imt0zaqh-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: newf1z8l; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 141.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 32 GB \u00b7 4bit; 4k 141.3 13.7 20.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/imt0zaqh",
   "source_url": "https://omlx.ai/benchmarks/performance/imt0zaqh",
   "tg_tokens": null,
   "tps": 13.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-t7m1b28o-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 14.9 GB at this context; duplicate traces of the same session: 4b51j87o; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 406.4,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-27B-oQ3e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 3bit; 4k 406.4 31.4 14.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t7m1b28o",
   "source_url": "https://omlx.ai/benchmarks/performance/t7m1b28o",
   "tg_tokens": null,
   "tps": 31.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-t7m1b28o-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 15.7 GB at this context; duplicate traces of the same session: 4b51j87o; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 390.3,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-27B-oQ3e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 3bit; 16k 390.3 27.5 15.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t7m1b28o",
   "source_url": "https://omlx.ai/benchmarks/performance/t7m1b28o",
   "tg_tokens": null,
   "tps": 27.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gm7g7tj5-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: 3u63bq7t, pni1e2ez, sp9yf5tb, v091ikse, zzs2c761; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 928.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 928.9 73.1 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gm7g7tj5",
   "source_url": "https://omlx.ai/benchmarks/performance/gm7g7tj5",
   "tg_tokens": null,
   "tps": 73.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gm7g7tj5-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 3u63bq7t, pni1e2ez, sp9yf5tb, v091ikse, zzs2c761; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1689.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,689 94.9 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gm7g7tj5",
   "source_url": "https://omlx.ai/benchmarks/performance/gm7g7tj5",
   "tg_tokens": null,
   "tps": 94.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gm7g7tj5-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: 3u63bq7t, pni1e2ez, sp9yf5tb, v091ikse, zzs2c761; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2078.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,078 75.8 72.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gm7g7tj5",
   "source_url": "https://omlx.ai/benchmarks/performance/gm7g7tj5",
   "tg_tokens": null,
   "tps": 75.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gm7g7tj5-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: 3u63bq7t, pni1e2ez, sp9yf5tb, v091ikse, zzs2c761; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2104.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,104 59.2 73.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gm7g7tj5",
   "source_url": "https://omlx.ai/benchmarks/performance/gm7g7tj5",
   "tg_tokens": null,
   "tps": 59.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gm7g7tj5-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.5 GB at this context; duplicate traces of the same session: 3u63bq7t, pni1e2ez, sp9yf5tb, v091ikse, zzs2c761; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1963.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 1,963 66.0 74.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gm7g7tj5",
   "source_url": "https://omlx.ai/benchmarks/performance/gm7g7tj5",
   "tg_tokens": null,
   "tps": 66.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gm7g7tj5-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.9 GB at this context; duplicate traces of the same session: 3u63bq7t, pni1e2ez, sp9yf5tb, v091ikse, zzs2c761; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1904.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 1,904 60.2 78.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gm7g7tj5",
   "source_url": "https://omlx.ai/benchmarks/performance/gm7g7tj5",
   "tg_tokens": null,
   "tps": 60.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fxkv1pi1-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 71.0 GB; TTFT 944 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1085.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,085 PP tok/s 65.2 TG tok/s 944 TTFT (ms) 71.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fxkv1pi1",
   "source_url": "https://omlx.ai/benchmarks/performance/fxkv1pi1",
   "tg_tokens": null,
   "tps": 65.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-va370htm-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 139.4 GB at this context; duplicate traces of the same session: 05s7qxqv, 694a5eph, vwkcb2yt, x5zw8ciz; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2007.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 1k 2,007 91.9 139.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/va370htm",
   "source_url": "https://omlx.ai/benchmarks/performance/va370htm",
   "tg_tokens": null,
   "tps": 91.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-va370htm-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 140.0 GB at this context; duplicate traces of the same session: 05s7qxqv, 694a5eph, vwkcb2yt, x5zw8ciz; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3620.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 4k 3,620 86.0 140.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/va370htm",
   "source_url": "https://omlx.ai/benchmarks/performance/va370htm",
   "tg_tokens": null,
   "tps": 86.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-va370htm-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 141.1 GB at this context; duplicate traces of the same session: 05s7qxqv, 694a5eph, vwkcb2yt, x5zw8ciz; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4100.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 8k 4,100 86.0 141.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/va370htm",
   "source_url": "https://omlx.ai/benchmarks/performance/va370htm",
   "tg_tokens": null,
   "tps": 86.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-va370htm-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 141.3 GB at this context; duplicate traces of the same session: 05s7qxqv, 694a5eph, vwkcb2yt, x5zw8ciz; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4108.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 16k 4,108 84.3 141.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/va370htm",
   "source_url": "https://omlx.ai/benchmarks/performance/va370htm",
   "tg_tokens": null,
   "tps": 84.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-va370htm-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.8 GB at this context; duplicate traces of the same session: 05s7qxqv, 694a5eph, vwkcb2yt, x5zw8ciz; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4037.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 64k 4,037 76.5 142.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/va370htm",
   "source_url": "https://omlx.ai/benchmarks/performance/va370htm",
   "tg_tokens": null,
   "tps": 76.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-va370htm-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.0 GB at this context; duplicate traces of the same session: 05s7qxqv, 694a5eph, vwkcb2yt, x5zw8ciz; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3933.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 128k 3,933 75.2 145.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/va370htm",
   "source_url": "https://omlx.ai/benchmarks/performance/va370htm",
   "tg_tokens": null,
   "tps": 75.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-va370htm-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 148.7 GB at this context; duplicate traces of the same session: 05s7qxqv, 694a5eph, vwkcb2yt, x5zw8ciz; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3834.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 195k 3,834 74.6 148.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/va370htm",
   "source_url": "https://omlx.ai/benchmarks/performance/va370htm",
   "tg_tokens": null,
   "tps": 74.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-uyuh6bhw-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 139.4 GB at this context; duplicate traces of the same session: chklg687, eeu5nmfb, h0wi166s, vxm2om44, xney811q; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2002.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 1k 2,002 92.1 139.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uyuh6bhw",
   "source_url": "https://omlx.ai/benchmarks/performance/uyuh6bhw",
   "tg_tokens": null,
   "tps": 92.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-uyuh6bhw-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 140.0 GB at this context; duplicate traces of the same session: chklg687, eeu5nmfb, h0wi166s, vxm2om44, xney811q; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3617.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 4k 3,617 86.0 140.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uyuh6bhw",
   "source_url": "https://omlx.ai/benchmarks/performance/uyuh6bhw",
   "tg_tokens": null,
   "tps": 86.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-uyuh6bhw-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 141.1 GB at this context; duplicate traces of the same session: chklg687, eeu5nmfb, h0wi166s, vxm2om44, xney811q; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4090.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 8k 4,090 85.1 141.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uyuh6bhw",
   "source_url": "https://omlx.ai/benchmarks/performance/uyuh6bhw",
   "tg_tokens": null,
   "tps": 85.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-uyuh6bhw-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 141.3 GB at this context; duplicate traces of the same session: chklg687, eeu5nmfb, h0wi166s, vxm2om44, xney811q; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4095.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 16k 4,095 83.7 141.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uyuh6bhw",
   "source_url": "https://omlx.ai/benchmarks/performance/uyuh6bhw",
   "tg_tokens": null,
   "tps": 83.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-uyuh6bhw-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 141.8 GB at this context; duplicate traces of the same session: chklg687, eeu5nmfb, h0wi166s, vxm2om44, xney811q; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4098.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 32k 4,098 82.2 141.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uyuh6bhw",
   "source_url": "https://omlx.ai/benchmarks/performance/uyuh6bhw",
   "tg_tokens": null,
   "tps": 82.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-uyuh6bhw-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.8 GB at this context; duplicate traces of the same session: chklg687, eeu5nmfb, h0wi166s, vxm2om44, xney811q; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4031.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 64k 4,031 76.4 142.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uyuh6bhw",
   "source_url": "https://omlx.ai/benchmarks/performance/uyuh6bhw",
   "tg_tokens": null,
   "tps": 76.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-uyuh6bhw-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.0 GB at this context; duplicate traces of the same session: chklg687, eeu5nmfb, h0wi166s, vxm2om44, xney811q; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3931.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 128k 3,931 75.0 145.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uyuh6bhw",
   "source_url": "https://omlx.ai/benchmarks/performance/uyuh6bhw",
   "tg_tokens": null,
   "tps": 75.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-uyuh6bhw-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 148.7 GB at this context; duplicate traces of the same session: chklg687, eeu5nmfb, h0wi166s, vxm2om44, xney811q; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3810.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 195k 3,810 74.3 148.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uyuh6bhw",
   "source_url": "https://omlx.ai/benchmarks/performance/uyuh6bhw",
   "tg_tokens": null,
   "tps": 74.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-2jnmbqme-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.4 GB at this context; duplicate traces of the same session: ej309kmu; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 454.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 454.8 62.4 173.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2jnmbqme",
   "source_url": "https://omlx.ai/benchmarks/performance/2jnmbqme",
   "tg_tokens": null,
   "tps": 62.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-2jnmbqme-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.3 GB at this context; duplicate traces of the same session: ej309kmu; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1021.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 1,021 57.0 174.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2jnmbqme",
   "source_url": "https://omlx.ai/benchmarks/performance/2jnmbqme",
   "tg_tokens": null,
   "tps": 57.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-8au0comi-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: 0n9e828l, r9nmu85s; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1321.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k 1,321 88.4 17.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8au0comi",
   "source_url": "https://omlx.ai/benchmarks/performance/8au0comi",
   "tg_tokens": null,
   "tps": 88.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-8au0comi-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 0n9e828l, r9nmu85s; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1562.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 4k 1,562 101.4 18.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8au0comi",
   "source_url": "https://omlx.ai/benchmarks/performance/8au0comi",
   "tg_tokens": null,
   "tps": 101.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-8au0comi-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 0n9e828l, r9nmu85s; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1561.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 8k 1,561 102.1 19.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8au0comi",
   "source_url": "https://omlx.ai/benchmarks/performance/8au0comi",
   "tg_tokens": null,
   "tps": 102.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-jd2f8sbv-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: 12apussk, dtjug1aw; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1312.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k 1,312 89.7 17.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jd2f8sbv",
   "source_url": "https://omlx.ai/benchmarks/performance/jd2f8sbv",
   "tg_tokens": null,
   "tps": 89.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-jd2f8sbv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 12apussk, dtjug1aw; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1566.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 4k 1,566 96.7 18.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jd2f8sbv",
   "source_url": "https://omlx.ai/benchmarks/performance/jd2f8sbv",
   "tg_tokens": null,
   "tps": 96.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-jd2f8sbv-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 12apussk, dtjug1aw; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1561.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 8k 1,561 124.1 19.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jd2f8sbv",
   "source_url": "https://omlx.ai/benchmarks/performance/jd2f8sbv",
   "tg_tokens": null,
   "tps": 124.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-u3yg4l1b-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 71.1 GB; TTFT 2372 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 431.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 431.8 PP tok/s 29.8 TG tok/s 2372 TTFT (ms) 71.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u3yg4l1b",
   "source_url": "https://omlx.ai/benchmarks/performance/u3yg4l1b",
   "tg_tokens": null,
   "tps": 29.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-754qzec8-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 139.4 GB at this context; duplicate traces of the same session: 6ghzbbpr; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2002.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 1k 2,002 92.1 139.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/754qzec8",
   "source_url": "https://omlx.ai/benchmarks/performance/754qzec8",
   "tg_tokens": null,
   "tps": 92.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-754qzec8-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 140.0 GB at this context; duplicate traces of the same session: 6ghzbbpr; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3617.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 4k 3,617 86.0 140.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/754qzec8",
   "source_url": "https://omlx.ai/benchmarks/performance/754qzec8",
   "tg_tokens": null,
   "tps": 86.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-754qzec8-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 141.1 GB at this context; duplicate traces of the same session: 6ghzbbpr; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4090.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 8k 4,090 85.1 141.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/754qzec8",
   "source_url": "https://omlx.ai/benchmarks/performance/754qzec8",
   "tg_tokens": null,
   "tps": 85.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-754qzec8-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 141.3 GB at this context; duplicate traces of the same session: 6ghzbbpr; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4095.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 16k 4,095 83.7 141.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/754qzec8",
   "source_url": "https://omlx.ai/benchmarks/performance/754qzec8",
   "tg_tokens": null,
   "tps": 83.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-754qzec8-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 141.8 GB at this context; duplicate traces of the same session: 6ghzbbpr; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4098.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 32k 4,098 82.2 141.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/754qzec8",
   "source_url": "https://omlx.ai/benchmarks/performance/754qzec8",
   "tg_tokens": null,
   "tps": 82.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-754qzec8-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.8 GB at this context; duplicate traces of the same session: 6ghzbbpr; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4031.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 64k 4,031 76.4 142.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/754qzec8",
   "source_url": "https://omlx.ai/benchmarks/performance/754qzec8",
   "tg_tokens": null,
   "tps": 76.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-754qzec8-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.0 GB at this context; duplicate traces of the same session: 6ghzbbpr; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3931.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 128k 3,931 75.0 145.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/754qzec8",
   "source_url": "https://omlx.ai/benchmarks/performance/754qzec8",
   "tg_tokens": null,
   "tps": 75.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-754qzec8-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 148.7 GB at this context; duplicate traces of the same session: 6ghzbbpr; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3810.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 195k 3,810 74.3 148.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/754qzec8",
   "source_url": "https://omlx.ai/benchmarks/performance/754qzec8",
   "tg_tokens": null,
   "tps": 74.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-q4zqq2a3-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.6 GB at this context; duplicate traces of the same session: 73f4bpz3, j3b6dcc5, rjhv3lyc, snanlwmo, y717wjij; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2013.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 1k 2,013 125.5 142.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q4zqq2a3",
   "source_url": "https://omlx.ai/benchmarks/performance/q4zqq2a3",
   "tg_tokens": null,
   "tps": 125.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-q4zqq2a3-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: 73f4bpz3, j3b6dcc5, rjhv3lyc, snanlwmo, y717wjij; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3755.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 4k 3,755 123.6 143.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q4zqq2a3",
   "source_url": "https://omlx.ai/benchmarks/performance/q4zqq2a3",
   "tg_tokens": null,
   "tps": 123.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-q4zqq2a3-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 144.5 GB at this context; duplicate traces of the same session: 73f4bpz3, j3b6dcc5, rjhv3lyc, snanlwmo, y717wjij; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4114.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 8k 4,114 133.4 144.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q4zqq2a3",
   "source_url": "https://omlx.ai/benchmarks/performance/q4zqq2a3",
   "tg_tokens": null,
   "tps": 133.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-q4zqq2a3-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 146.8 GB at this context; duplicate traces of the same session: 73f4bpz3, j3b6dcc5, rjhv3lyc, snanlwmo, y717wjij; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4081.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 64k 4,081 112.3 146.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q4zqq2a3",
   "source_url": "https://omlx.ai/benchmarks/performance/q4zqq2a3",
   "tg_tokens": null,
   "tps": 112.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-q4zqq2a3-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 150.6 GB at this context; duplicate traces of the same session: 73f4bpz3, j3b6dcc5, rjhv3lyc, snanlwmo, y717wjij; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3989.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 128k 3,989 95.6 150.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q4zqq2a3",
   "source_url": "https://omlx.ai/benchmarks/performance/q4zqq2a3",
   "tg_tokens": null,
   "tps": 95.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-q4zqq2a3-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 153.3 GB at this context; duplicate traces of the same session: 73f4bpz3, j3b6dcc5, rjhv3lyc, snanlwmo, y717wjij; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3867.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 195k 3,867 80.0 153.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q4zqq2a3",
   "source_url": "https://omlx.ai/benchmarks/performance/q4zqq2a3",
   "tg_tokens": null,
   "tps": 80.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-095kq6lt-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 141.9 GB at this context; duplicate traces of the same session: mc7m4cyi, vys9c8ev; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 733.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 1k 733.4 32.9 141.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/095kq6lt",
   "source_url": "https://omlx.ai/benchmarks/performance/095kq6lt",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-095kq6lt-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: mc7m4cyi, vys9c8ev; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1451.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 4k 1,451 111.7 143.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/095kq6lt",
   "source_url": "https://omlx.ai/benchmarks/performance/095kq6lt",
   "tg_tokens": null,
   "tps": 111.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-095kq6lt-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 144.5 GB at this context; duplicate traces of the same session: mc7m4cyi, vys9c8ev; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3661.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 8k 3,661 140.2 144.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/095kq6lt",
   "source_url": "https://omlx.ai/benchmarks/performance/095kq6lt",
   "tg_tokens": null,
   "tps": 140.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-095kq6lt-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.1 GB at this context; duplicate traces of the same session: mc7m4cyi, vys9c8ev; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3964.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 16k 3,964 47.8 145.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/095kq6lt",
   "source_url": "https://omlx.ai/benchmarks/performance/095kq6lt",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-095kq6lt-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.7 GB at this context; duplicate traces of the same session: mc7m4cyi, vys9c8ev; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4169.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 32k 4,169 67.1 145.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/095kq6lt",
   "source_url": "https://omlx.ai/benchmarks/performance/095kq6lt",
   "tg_tokens": null,
   "tps": 67.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-095kq6lt-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 146.8 GB at this context; duplicate traces of the same session: mc7m4cyi, vys9c8ev; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4096.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 64k 4,096 92.3 146.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/095kq6lt",
   "source_url": "https://omlx.ai/benchmarks/performance/095kq6lt",
   "tg_tokens": null,
   "tps": 92.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-095kq6lt-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 150.6 GB at this context; duplicate traces of the same session: mc7m4cyi, vys9c8ev; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4000.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 128k 4,000 96.0 150.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/095kq6lt",
   "source_url": "https://omlx.ai/benchmarks/performance/095kq6lt",
   "tg_tokens": null,
   "tps": 96.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-095kq6lt-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 153.3 GB at this context; duplicate traces of the same session: mc7m4cyi, vys9c8ev; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3891.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 195k 3,891 83.0 153.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/095kq6lt",
   "source_url": "https://omlx.ai/benchmarks/performance/095kq6lt",
   "tg_tokens": null,
   "tps": 83.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jthn0y5d-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: fjuxups0, lhz03n2d, yttodvmz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4690.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,690 228.5 19.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jthn0y5d",
   "source_url": "https://omlx.ai/benchmarks/performance/jthn0y5d",
   "tg_tokens": null,
   "tps": 228.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jthn0y5d-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: fjuxups0, lhz03n2d, yttodvmz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3579.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 3,579 142.7 20.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jthn0y5d",
   "source_url": "https://omlx.ai/benchmarks/performance/jthn0y5d",
   "tg_tokens": null,
   "tps": 142.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jthn0y5d-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: fjuxups0, lhz03n2d, yttodvmz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3748.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 3,748 132.4 21.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jthn0y5d",
   "source_url": "https://omlx.ai/benchmarks/performance/jthn0y5d",
   "tg_tokens": null,
   "tps": 132.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jthn0y5d-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: fjuxups0, lhz03n2d, yttodvmz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3076.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 3,076 48.9 22.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jthn0y5d",
   "source_url": "https://omlx.ai/benchmarks/performance/jthn0y5d",
   "tg_tokens": null,
   "tps": 48.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-k8i46xyh-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 04a0awli; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4154.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,154 127.1 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k8i46xyh",
   "source_url": "https://omlx.ai/benchmarks/performance/k8i46xyh",
   "tg_tokens": null,
   "tps": 127.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-k8i46xyh-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: 04a0awli; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3178.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 3,178 110.4 20.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k8i46xyh",
   "source_url": "https://omlx.ai/benchmarks/performance/k8i46xyh",
   "tg_tokens": null,
   "tps": 110.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-c9o839ir-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4258.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,258 130.6 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c9o839ir",
   "source_url": "https://omlx.ai/benchmarks/performance/c9o839ir",
   "tg_tokens": null,
   "tps": 130.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-c9o839ir-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4184.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 4,184 105.3 20.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c9o839ir",
   "source_url": "https://omlx.ai/benchmarks/performance/c9o839ir",
   "tg_tokens": null,
   "tps": 105.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lbti28ax-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3486.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 3,486 117.7 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lbti28ax",
   "source_url": "https://omlx.ai/benchmarks/performance/lbti28ax",
   "tg_tokens": null,
   "tps": 117.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lbti28ax-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3636.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 3,636 110.1 20.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lbti28ax",
   "source_url": "https://omlx.ai/benchmarks/performance/lbti28ax",
   "tg_tokens": null,
   "tps": 110.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ka1bv4n1-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: nfq8q4uo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4275.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,275 128.8 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ka1bv4n1",
   "source_url": "https://omlx.ai/benchmarks/performance/ka1bv4n1",
   "tg_tokens": null,
   "tps": 128.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ka1bv4n1-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: nfq8q4uo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4816.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 4,816 109.2 20.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ka1bv4n1",
   "source_url": "https://omlx.ai/benchmarks/performance/ka1bv4n1",
   "tg_tokens": null,
   "tps": 109.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-puoh8rcw-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 865.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 865.2 32.0 16.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/puoh8rcw",
   "source_url": "https://omlx.ai/benchmarks/performance/puoh8rcw",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-puoh8rcw-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 920.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 920.6 31.7 17.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/puoh8rcw",
   "source_url": "https://omlx.ai/benchmarks/performance/puoh8rcw",
   "tg_tokens": null,
   "tps": 31.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-7e1a0q6l-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: 7onhfibw, imsegx21; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1326.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k 1,326 107.9 17.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7e1a0q6l",
   "source_url": "https://omlx.ai/benchmarks/performance/7e1a0q6l",
   "tg_tokens": null,
   "tps": 107.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-7e1a0q6l-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 7onhfibw, imsegx21; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1564.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 4k 1,564 88.7 18.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7e1a0q6l",
   "source_url": "https://omlx.ai/benchmarks/performance/7e1a0q6l",
   "tg_tokens": null,
   "tps": 88.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-7e1a0q6l-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 7onhfibw, imsegx21; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1562.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 8k 1,562 110.5 19.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7e1a0q6l",
   "source_url": "https://omlx.ai/benchmarks/performance/7e1a0q6l",
   "tg_tokens": null,
   "tps": 110.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-yrxjfmh9-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 52.3 GB; TTFT 15500 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1057.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 2bit; 16k tokens 1,057 PP tok/s 30.2 TG tok/s 15500 TTFT (ms) 52.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yrxjfmh9",
   "source_url": "https://omlx.ai/benchmarks/performance/yrxjfmh9",
   "tg_tokens": null,
   "tps": 30.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-j1lxpuxc-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 141.9 GB at this context; duplicate traces of the same session: 2d01k13a, je0n2k3h, ltr867kz, tf7dd75z; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 733.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 1k 733.4 32.9 141.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j1lxpuxc",
   "source_url": "https://omlx.ai/benchmarks/performance/j1lxpuxc",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-j1lxpuxc-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: 2d01k13a, je0n2k3h, ltr867kz, tf7dd75z; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1451.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 4k 1,451 111.7 143.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j1lxpuxc",
   "source_url": "https://omlx.ai/benchmarks/performance/j1lxpuxc",
   "tg_tokens": null,
   "tps": 111.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-j1lxpuxc-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 144.5 GB at this context; duplicate traces of the same session: 2d01k13a, je0n2k3h, ltr867kz, tf7dd75z; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3661.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 8k 3,661 140.2 144.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j1lxpuxc",
   "source_url": "https://omlx.ai/benchmarks/performance/j1lxpuxc",
   "tg_tokens": null,
   "tps": 140.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-j1lxpuxc-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.1 GB at this context; duplicate traces of the same session: 2d01k13a, je0n2k3h, ltr867kz, tf7dd75z; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3964.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 16k 3,964 47.8 145.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j1lxpuxc",
   "source_url": "https://omlx.ai/benchmarks/performance/j1lxpuxc",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-j1lxpuxc-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.7 GB at this context; duplicate traces of the same session: 2d01k13a, je0n2k3h, ltr867kz, tf7dd75z; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4169.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 32k 4,169 67.1 145.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j1lxpuxc",
   "source_url": "https://omlx.ai/benchmarks/performance/j1lxpuxc",
   "tg_tokens": null,
   "tps": 67.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-j1lxpuxc-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 146.8 GB at this context; duplicate traces of the same session: 2d01k13a, je0n2k3h, ltr867kz, tf7dd75z; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4096.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 64k 4,096 92.3 146.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j1lxpuxc",
   "source_url": "https://omlx.ai/benchmarks/performance/j1lxpuxc",
   "tg_tokens": null,
   "tps": 92.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-j1lxpuxc-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 150.6 GB at this context; duplicate traces of the same session: 2d01k13a, je0n2k3h, ltr867kz, tf7dd75z; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4000.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 128k 4,000 96.0 150.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j1lxpuxc",
   "source_url": "https://omlx.ai/benchmarks/performance/j1lxpuxc",
   "tg_tokens": null,
   "tps": 96.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-j1lxpuxc-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 153.3 GB at this context; duplicate traces of the same session: 2d01k13a, je0n2k3h, ltr867kz, tf7dd75z; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3891.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 195k 3,891 83.0 153.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j1lxpuxc",
   "source_url": "https://omlx.ai/benchmarks/performance/j1lxpuxc",
   "tg_tokens": null,
   "tps": 83.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8q7b2vnb-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4258.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,258 130.6 19.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8q7b2vnb",
   "source_url": "https://omlx.ai/benchmarks/performance/8q7b2vnb",
   "tg_tokens": null,
   "tps": 130.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8q7b2vnb-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4184.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 4,184 105.3 20.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8q7b2vnb",
   "source_url": "https://omlx.ai/benchmarks/performance/8q7b2vnb",
   "tg_tokens": null,
   "tps": 105.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9402zbbw-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3486.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 3,486 117.7 19.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9402zbbw",
   "source_url": "https://omlx.ai/benchmarks/performance/9402zbbw",
   "tg_tokens": null,
   "tps": 117.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9402zbbw-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3636.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 3,636 110.1 20.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9402zbbw",
   "source_url": "https://omlx.ai/benchmarks/performance/9402zbbw",
   "tg_tokens": null,
   "tps": 110.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-99ppecfe-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: klcfvdbl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3347.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 3,347 91.9 20.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/99ppecfe",
   "source_url": "https://omlx.ai/benchmarks/performance/99ppecfe",
   "tg_tokens": null,
   "tps": 91.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-99ppecfe-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: klcfvdbl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3436.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 3,436 100.2 20.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/99ppecfe",
   "source_url": "https://omlx.ai/benchmarks/performance/99ppecfe",
   "tg_tokens": null,
   "tps": 100.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-3tr8lwgy-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: e9lo6j25, xw9ps8k3; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1636.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 1,636 103.6 20.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3tr8lwgy",
   "source_url": "https://omlx.ai/benchmarks/performance/3tr8lwgy",
   "tg_tokens": null,
   "tps": 103.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-3tr8lwgy-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: e9lo6j25, xw9ps8k3; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2554.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 2,554 89.9 20.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3tr8lwgy",
   "source_url": "https://omlx.ai/benchmarks/performance/3tr8lwgy",
   "tg_tokens": null,
   "tps": 89.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-3tr8lwgy-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: e9lo6j25, xw9ps8k3; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2149.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 2,149 77.3 20.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3tr8lwgy",
   "source_url": "https://omlx.ai/benchmarks/performance/3tr8lwgy",
   "tg_tokens": null,
   "tps": 77.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-oy3b4lls-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 8uaisl45; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2301.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 2,301 108.0 19.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oy3b4lls",
   "source_url": "https://omlx.ai/benchmarks/performance/oy3b4lls",
   "tg_tokens": null,
   "tps": 108.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-oy3b4lls-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 8uaisl45; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2950.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,950 104.0 20.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oy3b4lls",
   "source_url": "https://omlx.ai/benchmarks/performance/oy3b4lls",
   "tg_tokens": null,
   "tps": 104.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-9ruhh2g6-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 16.0 GB; TTFT 2517 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 406.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k tokens 406.9 PP tok/s 18.2 TG tok/s 2517 TTFT (ms) 16.0 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ruhh2g6",
   "source_url": "https://omlx.ai/benchmarks/performance/9ruhh2g6",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-fiwcac1m-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 14.9 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 406.3,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-27B-oQ3e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 3bit; 4k 406.3 34.5 14.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fiwcac1m",
   "source_url": "https://omlx.ai/benchmarks/performance/fiwcac1m",
   "tg_tokens": null,
   "tps": 34.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fiwcac1m-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 15.7 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 389.7,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-27B-oQ3e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 3bit; 16k 389.7 29.5 15.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fiwcac1m",
   "source_url": "https://omlx.ai/benchmarks/performance/fiwcac1m",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fiwcac1m-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 356.5,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-27B-oQ3e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 3bit; 32k 356.5 23.1 16.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fiwcac1m",
   "source_url": "https://omlx.ai/benchmarks/performance/fiwcac1m",
   "tg_tokens": null,
   "tps": 23.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-1z1th5n3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: 1y55tnvb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 136.3,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 136.3 14.8 17.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1z1th5n3",
   "source_url": "https://omlx.ai/benchmarks/performance/1z1th5n3",
   "tg_tokens": null,
   "tps": 14.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-1z1th5n3-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 1y55tnvb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 131.0 13.7 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1z1th5n3",
   "source_url": "https://omlx.ai/benchmarks/performance/1z1th5n3",
   "tg_tokens": null,
   "tps": 13.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-eju9f80y-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 865.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 865.2 32.0 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eju9f80y",
   "source_url": "https://omlx.ai/benchmarks/performance/eju9f80y",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-eju9f80y-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 920.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 920.6 31.7 17.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eju9f80y",
   "source_url": "https://omlx.ai/benchmarks/performance/eju9f80y",
   "tg_tokens": null,
   "tps": 31.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-feuk1hvk-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 47.4 GB at this context; duplicate traces of the same session: e49le6pf, tm7yxd3c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 91.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 91.5 18.6 47.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/feuk1hvk",
   "source_url": "https://omlx.ai/benchmarks/performance/feuk1hvk",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-feuk1hvk-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 49.2 GB at this context; duplicate traces of the same session: e49le6pf, tm7yxd3c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 320.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 320.5 19.2 49.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/feuk1hvk",
   "source_url": "https://omlx.ai/benchmarks/performance/feuk1hvk",
   "tg_tokens": null,
   "tps": 19.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-feuk1hvk-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 49.3 GB at this context; duplicate traces of the same session: e49le6pf, tm7yxd3c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 154.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 154.0 5.8 49.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/feuk1hvk",
   "source_url": "https://omlx.ai/benchmarks/performance/feuk1hvk",
   "tg_tokens": null,
   "tps": 5.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-sta7qp5u-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 35.4 GB; TTFT 41265 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 397.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k tokens 397.0 PP tok/s 10.6 TG tok/s 41265 TTFT (ms) 35.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sta7qp5u",
   "source_url": "https://omlx.ai/benchmarks/performance/sta7qp5u",
   "tg_tokens": null,
   "tps": 10.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-y67s16pk-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Youssofal MTPLX Optimized Speed; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 6q7tsmcc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 832.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Youssofal--Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 832.1 51.2 23.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y67s16pk",
   "source_url": "https://omlx.ai/benchmarks/performance/y67s16pk",
   "tg_tokens": null,
   "tps": 51.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-y67s16pk-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "Youssofal MTPLX Optimized Speed; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: 6q7tsmcc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 944.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Youssofal--Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 944.2 42.6 24.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y67s16pk",
   "source_url": "https://omlx.ai/benchmarks/performance/y67s16pk",
   "tg_tokens": null,
   "tps": 42.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-oywxa096-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.4 GB at this context; duplicate traces of the same session: 8a463zg3, kcohn72l; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1359.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-9B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 1,359 60.7 7.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oywxa096",
   "source_url": "https://omlx.ai/benchmarks/performance/oywxa096",
   "tg_tokens": null,
   "tps": 60.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-oywxa096-ctx16384",
   "model": "Ornith-1.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.8 GB at this context; duplicate traces of the same session: 8a463zg3, kcohn72l; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1342.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-9B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 1,342 55.4 7.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oywxa096",
   "source_url": "https://omlx.ai/benchmarks/performance/oywxa096",
   "tg_tokens": null,
   "tps": 55.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-oywxa096-ctx32768",
   "model": "Ornith-1.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 8.4 GB at this context; duplicate traces of the same session: 8a463zg3, kcohn72l; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1228.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-9B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 1,228 69.5 8.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oywxa096",
   "source_url": "https://omlx.ai/benchmarks/performance/oywxa096",
   "tg_tokens": null,
   "tps": 69.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-3wvggbdz-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 22.5 GB; TTFT 39496 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 414.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 16k tokens 414.8 PP tok/s 38.2 TG tok/s 39496 TTFT (ms) 22.5 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3wvggbdz",
   "source_url": "https://omlx.ai/benchmarks/performance/3wvggbdz",
   "tg_tokens": null,
   "tps": 38.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-a84wn2n6-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: hy76adyl; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 506.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 506.6 17.3 17.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a84wn2n6",
   "source_url": "https://omlx.ai/benchmarks/performance/a84wn2n6",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-a84wn2n6-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: hy76adyl; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 460.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 460.9 16.1 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a84wn2n6",
   "source_url": "https://omlx.ai/benchmarks/performance/a84wn2n6",
   "tg_tokens": null,
   "tps": 16.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-gmhrgm6g-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 14.9 GB at this context; duplicate traces of the same session: ca1pfmn8; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 406.3,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-27B-oQ3e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 3bit; 4k 406.3 34.5 14.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gmhrgm6g",
   "source_url": "https://omlx.ai/benchmarks/performance/gmhrgm6g",
   "tg_tokens": null,
   "tps": 34.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-gmhrgm6g-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 15.7 GB at this context; duplicate traces of the same session: ca1pfmn8; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 389.7,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-27B-oQ3e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 3bit; 16k 389.7 29.5 15.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gmhrgm6g",
   "source_url": "https://omlx.ai/benchmarks/performance/gmhrgm6g",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-gmhrgm6g-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: ca1pfmn8; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 356.5,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-27B-oQ3e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 3bit; 32k 356.5 23.1 16.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gmhrgm6g",
   "source_url": "https://omlx.ai/benchmarks/performance/gmhrgm6g",
   "tg_tokens": null,
   "tps": 23.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-jk4xgrvq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 14.9 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 405.9,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-27B-oQ3e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 3bit; 4k 405.9 42.5 14.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jk4xgrvq",
   "source_url": "https://omlx.ai/benchmarks/performance/jk4xgrvq",
   "tg_tokens": null,
   "tps": 42.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-jk4xgrvq-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 15.7 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 385.6,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-27B-oQ3e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 3bit; 16k 385.6 26.3 15.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jk4xgrvq",
   "source_url": "https://omlx.ai/benchmarks/performance/jk4xgrvq",
   "tg_tokens": null,
   "tps": 26.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-jk4xgrvq-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 354.7,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-27B-oQ3e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 3bit; 32k 354.7 21.0 16.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jk4xgrvq",
   "source_url": "https://omlx.ai/benchmarks/performance/jk4xgrvq",
   "tg_tokens": null,
   "tps": 21.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-rnkcjn2d-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 3o8k2nw5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 601.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 601.3 29.1 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rnkcjn2d",
   "source_url": "https://omlx.ai/benchmarks/performance/rnkcjn2d",
   "tg_tokens": null,
   "tps": 29.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-rnkcjn2d-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 3o8k2nw5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 688.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 688.1 32.6 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rnkcjn2d",
   "source_url": "https://omlx.ai/benchmarks/performance/rnkcjn2d",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-rnkcjn2d-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 3o8k2nw5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 713.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 713.8 41.5 19.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rnkcjn2d",
   "source_url": "https://omlx.ai/benchmarks/performance/rnkcjn2d",
   "tg_tokens": null,
   "tps": 41.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-rnkcjn2d-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 3o8k2nw5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 692.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 692.3 47.3 20.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rnkcjn2d",
   "source_url": "https://omlx.ai/benchmarks/performance/rnkcjn2d",
   "tg_tokens": null,
   "tps": 47.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-x83hvwz2-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.0 GB at this context; duplicate traces of the same session: 99j6b5e5, irw0f72u, zyic74n7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 93.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 93.5 14.6 39.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x83hvwz2",
   "source_url": "https://omlx.ai/benchmarks/performance/x83hvwz2",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-x83hvwz2-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.7 GB at this context; duplicate traces of the same session: 99j6b5e5, irw0f72u, zyic74n7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 326.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 326.3 15.4 39.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x83hvwz2",
   "source_url": "https://omlx.ai/benchmarks/performance/x83hvwz2",
   "tg_tokens": null,
   "tps": 15.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-x83hvwz2-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 41.7 GB at this context; duplicate traces of the same session: 99j6b5e5, irw0f72u, zyic74n7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 451.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 451.2 10.3 41.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x83hvwz2",
   "source_url": "https://omlx.ai/benchmarks/performance/x83hvwz2",
   "tg_tokens": null,
   "tps": 10.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-x83hvwz2-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 42.3 GB at this context; duplicate traces of the same session: 99j6b5e5, irw0f72u, zyic74n7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 424.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 424.0 11.4 42.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x83hvwz2",
   "source_url": "https://omlx.ai/benchmarks/performance/x83hvwz2",
   "tg_tokens": null,
   "tps": 11.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-8bhxixif-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.7 GB at this context; duplicate traces of the same session: 3usx90ny, w28yo0oh; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1805.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 1k 1,805 132.5 142.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8bhxixif",
   "source_url": "https://omlx.ai/benchmarks/performance/8bhxixif",
   "tg_tokens": null,
   "tps": 132.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-8bhxixif-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: 3usx90ny, w28yo0oh; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3439.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 4k 3,439 115.0 143.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8bhxixif",
   "source_url": "https://omlx.ai/benchmarks/performance/8bhxixif",
   "tg_tokens": null,
   "tps": 115.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-8bhxixif-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 144.5 GB at this context; duplicate traces of the same session: 3usx90ny, w28yo0oh; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3889.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 8k 3,889 122.5 144.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8bhxixif",
   "source_url": "https://omlx.ai/benchmarks/performance/8bhxixif",
   "tg_tokens": null,
   "tps": 122.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-3nsz6hzv-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: wj37lzx3; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 108.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 108.8 25.0 18.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3nsz6hzv",
   "source_url": "https://omlx.ai/benchmarks/performance/3nsz6hzv",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-3nsz6hzv-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: wj37lzx3; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 102.2 23.4 22.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3nsz6hzv",
   "source_url": "https://omlx.ai/benchmarks/performance/3nsz6hzv",
   "tg_tokens": null,
   "tps": 23.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-j5ejg8wn-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: sqir9nl2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 952.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 952.4 18.8 29.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j5ejg8wn",
   "source_url": "https://omlx.ai/benchmarks/performance/j5ejg8wn",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-j5ejg8wn-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: sqir9nl2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 933.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 16k 933.0 18.1 30.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j5ejg8wn",
   "source_url": "https://omlx.ai/benchmarks/performance/j5ejg8wn",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kfzq9e11-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: jm919ifb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 806.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 806.3 19.0 28.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kfzq9e11",
   "source_url": "https://omlx.ai/benchmarks/performance/kfzq9e11",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kfzq9e11-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: jm919ifb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 918.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 918.6 18.8 29.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kfzq9e11",
   "source_url": "https://omlx.ai/benchmarks/performance/kfzq9e11",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-5pynqspu-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 14.9 GB at this context; duplicate traces of the same session: e12i9izh; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 405.9,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-27B-oQ3e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 3bit; 4k 405.9 42.5 14.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5pynqspu",
   "source_url": "https://omlx.ai/benchmarks/performance/5pynqspu",
   "tg_tokens": null,
   "tps": 42.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-5pynqspu-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 15.7 GB at this context; duplicate traces of the same session: e12i9izh; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 385.6,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-27B-oQ3e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 3bit; 16k 385.6 26.3 15.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5pynqspu",
   "source_url": "https://omlx.ai/benchmarks/performance/5pynqspu",
   "tg_tokens": null,
   "tps": 26.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-5pynqspu-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: e12i9izh; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 354.7,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-27B-oQ3e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 3bit; 32k 354.7 21.0 16.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5pynqspu",
   "source_url": "https://omlx.ai/benchmarks/performance/5pynqspu",
   "tg_tokens": null,
   "tps": 21.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-d0l8ybpr-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: a0208m2e; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 601.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 601.3 29.1 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d0l8ybpr",
   "source_url": "https://omlx.ai/benchmarks/performance/d0l8ybpr",
   "tg_tokens": null,
   "tps": 29.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-d0l8ybpr-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: a0208m2e; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 688.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 688.1 32.6 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d0l8ybpr",
   "source_url": "https://omlx.ai/benchmarks/performance/d0l8ybpr",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-d0l8ybpr-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: a0208m2e; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 713.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 713.8 41.5 19.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d0l8ybpr",
   "source_url": "https://omlx.ai/benchmarks/performance/d0l8ybpr",
   "tg_tokens": null,
   "tps": 41.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-d0l8ybpr-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: a0208m2e; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 692.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 692.3 47.3 20.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d0l8ybpr",
   "source_url": "https://omlx.ai/benchmarks/performance/d0l8ybpr",
   "tg_tokens": null,
   "tps": 47.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qejvrb5e-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 3fy5s226, hxl6y2qv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 515.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 515.5 27.5 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qejvrb5e",
   "source_url": "https://omlx.ai/benchmarks/performance/qejvrb5e",
   "tg_tokens": null,
   "tps": 27.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qejvrb5e-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 3fy5s226, hxl6y2qv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 674.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 674.9 34.4 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qejvrb5e",
   "source_url": "https://omlx.ai/benchmarks/performance/qejvrb5e",
   "tg_tokens": null,
   "tps": 34.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qejvrb5e-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 3fy5s226, hxl6y2qv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 713.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 713.6 30.0 19.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qejvrb5e",
   "source_url": "https://omlx.ai/benchmarks/performance/qejvrb5e",
   "tg_tokens": null,
   "tps": 30.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qejvrb5e-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 3fy5s226, hxl6y2qv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 693.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 693.4 45.2 20.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qejvrb5e",
   "source_url": "https://omlx.ai/benchmarks/performance/qejvrb5e",
   "tg_tokens": null,
   "tps": 45.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-0i678630-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 139.4 GB at this context; duplicate traces of the same session: 4z4czal3, bdme32tt, g15wa8v3, klx6jfcu, lxr4sabu; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2110.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 1k 2,110 84.0 139.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0i678630",
   "source_url": "https://omlx.ai/benchmarks/performance/0i678630",
   "tg_tokens": null,
   "tps": 84.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-0i678630-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 140.0 GB at this context; duplicate traces of the same session: 4z4czal3, bdme32tt, g15wa8v3, klx6jfcu, lxr4sabu; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3889.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 4k 3,889 78.4 140.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0i678630",
   "source_url": "https://omlx.ai/benchmarks/performance/0i678630",
   "tg_tokens": null,
   "tps": 78.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-0i678630-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 141.1 GB at this context; duplicate traces of the same session: 4z4czal3, bdme32tt, g15wa8v3, klx6jfcu, lxr4sabu; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4247.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 8k 4,247 77.4 141.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0i678630",
   "source_url": "https://omlx.ai/benchmarks/performance/0i678630",
   "tg_tokens": null,
   "tps": 77.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-0i678630-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.8 GB at this context; duplicate traces of the same session: 4z4czal3, bdme32tt, g15wa8v3, klx6jfcu, lxr4sabu; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4204.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 64k 4,204 73.0 142.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0i678630",
   "source_url": "https://omlx.ai/benchmarks/performance/0i678630",
   "tg_tokens": null,
   "tps": 73.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-0i678630-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.1 GB at this context; duplicate traces of the same session: 4z4czal3, bdme32tt, g15wa8v3, klx6jfcu, lxr4sabu; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4117.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 128k 4,117 67.5 145.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0i678630",
   "source_url": "https://omlx.ai/benchmarks/performance/0i678630",
   "tg_tokens": null,
   "tps": 67.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-0i678630-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 148.7 GB at this context; duplicate traces of the same session: 4z4czal3, bdme32tt, g15wa8v3, klx6jfcu, lxr4sabu; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4014.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 195k 4,014 67.2 148.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0i678630",
   "source_url": "https://omlx.ai/benchmarks/performance/0i678630",
   "tg_tokens": null,
   "tps": 67.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qfdz8xmr-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 22.8 GB; TTFT 92309 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 355.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k tokens 355.0 PP tok/s 69.7 TG tok/s 92309 TTFT (ms) 22.8 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qfdz8xmr",
   "source_url": "https://omlx.ai/benchmarks/performance/qfdz8xmr",
   "tg_tokens": null,
   "tps": 69.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ghyki8sq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: p70r3kth; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1552.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,552 54.5 16.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ghyki8sq",
   "source_url": "https://omlx.ai/benchmarks/performance/ghyki8sq",
   "tg_tokens": null,
   "tps": 54.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ghyki8sq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: p70r3kth; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1854.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 1,854 53.2 17.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ghyki8sq",
   "source_url": "https://omlx.ai/benchmarks/performance/ghyki8sq",
   "tg_tokens": null,
   "tps": 53.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ghyki8sq-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: p70r3kth; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1860.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 1,860 52.5 17.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ghyki8sq",
   "source_url": "https://omlx.ai/benchmarks/performance/ghyki8sq",
   "tg_tokens": null,
   "tps": 52.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ghyki8sq-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: p70r3kth; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1805.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 1,805 50.2 17.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ghyki8sq",
   "source_url": "https://omlx.ai/benchmarks/performance/ghyki8sq",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ghyki8sq-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: p70r3kth; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1699.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 1,699 46.3 18.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ghyki8sq",
   "source_url": "https://omlx.ai/benchmarks/performance/ghyki8sq",
   "tg_tokens": null,
   "tps": 46.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ghyki8sq-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: p70r3kth; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1498.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 1,498 41.8 20.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ghyki8sq",
   "source_url": "https://omlx.ai/benchmarks/performance/ghyki8sq",
   "tg_tokens": null,
   "tps": 41.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ghyki8sq-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 123.2 GB at this context; duplicate traces of the same session: p70r3kth; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1174.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 1,174 32.0 123.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ghyki8sq",
   "source_url": "https://omlx.ai/benchmarks/performance/ghyki8sq",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ghyki8sq-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 129.8 GB at this context; duplicate traces of the same session: p70r3kth; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 689.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 689.1 28.1 129.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ghyki8sq",
   "source_url": "https://omlx.ai/benchmarks/performance/ghyki8sq",
   "tg_tokens": null,
   "tps": 28.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wfvav2l6-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: doqda2ra; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 879.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 879.2 53.8 17.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wfvav2l6",
   "source_url": "https://omlx.ai/benchmarks/performance/wfvav2l6",
   "tg_tokens": null,
   "tps": 53.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wfvav2l6-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: doqda2ra; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 924.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 924.3 59.0 18.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wfvav2l6",
   "source_url": "https://omlx.ai/benchmarks/performance/wfvav2l6",
   "tg_tokens": null,
   "tps": 59.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wfvav2l6-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: doqda2ra; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 770.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 770.8 55.0 19.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wfvav2l6",
   "source_url": "https://omlx.ai/benchmarks/performance/wfvav2l6",
   "tg_tokens": null,
   "tps": 55.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wfvav2l6-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: doqda2ra; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 722.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 722.0 58.3 19.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wfvav2l6",
   "source_url": "https://omlx.ai/benchmarks/performance/wfvav2l6",
   "tg_tokens": null,
   "tps": 58.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wfvav2l6-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: doqda2ra; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 673.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 673.8 120.8 20.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wfvav2l6",
   "source_url": "https://omlx.ai/benchmarks/performance/wfvav2l6",
   "tg_tokens": null,
   "tps": 120.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wfvav2l6-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: doqda2ra; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 544.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 544.8 33.5 23.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wfvav2l6",
   "source_url": "https://omlx.ai/benchmarks/performance/wfvav2l6",
   "tg_tokens": null,
   "tps": 33.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-6h1p3sc6-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 515.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 515.5 27.5 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6h1p3sc6",
   "source_url": "https://omlx.ai/benchmarks/performance/6h1p3sc6",
   "tg_tokens": null,
   "tps": 27.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-6h1p3sc6-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 674.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 674.9 34.4 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6h1p3sc6",
   "source_url": "https://omlx.ai/benchmarks/performance/6h1p3sc6",
   "tg_tokens": null,
   "tps": 34.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-6h1p3sc6-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 713.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 713.6 30.0 19.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6h1p3sc6",
   "source_url": "https://omlx.ai/benchmarks/performance/6h1p3sc6",
   "tg_tokens": null,
   "tps": 30.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-6h1p3sc6-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 693.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 693.4 45.2 20.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6h1p3sc6",
   "source_url": "https://omlx.ai/benchmarks/performance/6h1p3sc6",
   "tg_tokens": null,
   "tps": 45.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2vlpspj9-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.8 GB at this context; duplicate traces of the same session: w2xhkcww; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 94.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 94.1 13.2 32.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2vlpspj9",
   "source_url": "https://omlx.ai/benchmarks/performance/2vlpspj9",
   "tg_tokens": null,
   "tps": 13.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2vlpspj9-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.6 GB at this context; duplicate traces of the same session: w2xhkcww; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 431.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 431.3 13.6 34.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2vlpspj9",
   "source_url": "https://omlx.ai/benchmarks/performance/2vlpspj9",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2vlpspj9-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.4 GB at this context; duplicate traces of the same session: w2xhkcww; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 463.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 463.8 14.2 35.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2vlpspj9",
   "source_url": "https://omlx.ai/benchmarks/performance/2vlpspj9",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2vlpspj9-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.0 GB at this context; duplicate traces of the same session: w2xhkcww; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 460.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 460.7 11.1 36.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2vlpspj9",
   "source_url": "https://omlx.ai/benchmarks/performance/2vlpspj9",
   "tg_tokens": null,
   "tps": 11.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-5m7h4s40-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 139.4 GB at this context; duplicate traces of the same session: 0mjz2wax, 9ybgo0im, itwk8hpi, scojlrz2, u03t6y44, v0nnkryw; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2124.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 1k 2,124 85.1 139.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5m7h4s40",
   "source_url": "https://omlx.ai/benchmarks/performance/5m7h4s40",
   "tg_tokens": null,
   "tps": 85.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-5m7h4s40-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 140.0 GB at this context; duplicate traces of the same session: 0mjz2wax, 9ybgo0im, itwk8hpi, scojlrz2, u03t6y44, v0nnkryw; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3903.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 4k 3,903 79.8 140.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5m7h4s40",
   "source_url": "https://omlx.ai/benchmarks/performance/5m7h4s40",
   "tg_tokens": null,
   "tps": 79.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-5m7h4s40-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 141.1 GB at this context; duplicate traces of the same session: 0mjz2wax, 9ybgo0im, itwk8hpi, scojlrz2, u03t6y44, v0nnkryw; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4292.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 8k 4,292 77.8 141.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5m7h4s40",
   "source_url": "https://omlx.ai/benchmarks/performance/5m7h4s40",
   "tg_tokens": null,
   "tps": 77.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-5m7h4s40-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 141.3 GB at this context; duplicate traces of the same session: 0mjz2wax, 9ybgo0im, itwk8hpi, scojlrz2, u03t6y44, v0nnkryw; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4303.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 16k 4,303 75.2 141.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5m7h4s40",
   "source_url": "https://omlx.ai/benchmarks/performance/5m7h4s40",
   "tg_tokens": null,
   "tps": 75.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-5m7h4s40-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.8 GB at this context; duplicate traces of the same session: 0mjz2wax, 9ybgo0im, itwk8hpi, scojlrz2, u03t6y44, v0nnkryw; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4238.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 64k 4,238 74.4 142.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5m7h4s40",
   "source_url": "https://omlx.ai/benchmarks/performance/5m7h4s40",
   "tg_tokens": null,
   "tps": 74.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-5m7h4s40-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.1 GB at this context; duplicate traces of the same session: 0mjz2wax, 9ybgo0im, itwk8hpi, scojlrz2, u03t6y44, v0nnkryw; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4128.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 128k 4,128 68.5 145.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5m7h4s40",
   "source_url": "https://omlx.ai/benchmarks/performance/5m7h4s40",
   "tg_tokens": null,
   "tps": 68.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-5m7h4s40-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 148.7 GB at this context; duplicate traces of the same session: 0mjz2wax, 9ybgo0im, itwk8hpi, scojlrz2, u03t6y44, v0nnkryw; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4012.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 195k 4,012 64.7 148.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5m7h4s40",
   "source_url": "https://omlx.ai/benchmarks/performance/5m7h4s40",
   "tg_tokens": null,
   "tps": 64.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-tcluavvr-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 22.8 GB; TTFT 100108 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 327.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k tokens 327.3 PP tok/s 69.5 TG tok/s 100108 TTFT (ms) 22.8 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tcluavvr",
   "source_url": "https://omlx.ai/benchmarks/performance/tcluavvr",
   "tg_tokens": null,
   "tps": 69.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-yy7dcmh4-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 3sxyvkqn, 49l4mp1w, fc7nj5in, huwt74xh, iokm9a2l; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1552.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,552 54.5 16.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yy7dcmh4",
   "source_url": "https://omlx.ai/benchmarks/performance/yy7dcmh4",
   "tg_tokens": null,
   "tps": 54.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-yy7dcmh4-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 3sxyvkqn, 49l4mp1w, fc7nj5in, huwt74xh, iokm9a2l; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1854.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 1,854 53.2 17.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yy7dcmh4",
   "source_url": "https://omlx.ai/benchmarks/performance/yy7dcmh4",
   "tg_tokens": null,
   "tps": 53.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-yy7dcmh4-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 3sxyvkqn, 49l4mp1w, fc7nj5in, huwt74xh, iokm9a2l; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1860.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 1,860 52.5 17.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yy7dcmh4",
   "source_url": "https://omlx.ai/benchmarks/performance/yy7dcmh4",
   "tg_tokens": null,
   "tps": 52.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-yy7dcmh4-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 3sxyvkqn, 49l4mp1w, fc7nj5in, huwt74xh, iokm9a2l; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1805.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 1,805 50.2 17.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yy7dcmh4",
   "source_url": "https://omlx.ai/benchmarks/performance/yy7dcmh4",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-yy7dcmh4-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 3sxyvkqn, 49l4mp1w, fc7nj5in, huwt74xh, iokm9a2l; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1699.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 1,699 46.3 18.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yy7dcmh4",
   "source_url": "https://omlx.ai/benchmarks/performance/yy7dcmh4",
   "tg_tokens": null,
   "tps": 46.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-yy7dcmh4-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 3sxyvkqn, 49l4mp1w, fc7nj5in, huwt74xh, iokm9a2l; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1498.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 1,498 41.8 20.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yy7dcmh4",
   "source_url": "https://omlx.ai/benchmarks/performance/yy7dcmh4",
   "tg_tokens": null,
   "tps": 41.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-yy7dcmh4-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 123.2 GB at this context; duplicate traces of the same session: 3sxyvkqn, 49l4mp1w, fc7nj5in, huwt74xh, iokm9a2l; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1174.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 1,174 32.0 123.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yy7dcmh4",
   "source_url": "https://omlx.ai/benchmarks/performance/yy7dcmh4",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-yy7dcmh4-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 129.8 GB at this context; duplicate traces of the same session: 3sxyvkqn, 49l4mp1w, fc7nj5in, huwt74xh, iokm9a2l; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 689.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 689.1 28.1 129.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yy7dcmh4",
   "source_url": "https://omlx.ai/benchmarks/performance/yy7dcmh4",
   "tg_tokens": null,
   "tps": 28.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nbw2u1lb-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 954.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "qwen3.8-flash-next-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 954.8 82.9 101.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nbw2u1lb",
   "source_url": "https://omlx.ai/benchmarks/performance/nbw2u1lb",
   "tg_tokens": null,
   "tps": 82.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nbw2u1lb-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1376.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "qwen3.8-flash-next-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 1,376 22.0 101.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nbw2u1lb",
   "source_url": "https://omlx.ai/benchmarks/performance/nbw2u1lb",
   "tg_tokens": null,
   "tps": 22.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jippy6v6-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: 0xfhevqa, 9vw968j9, u45iqp5e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2649.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 2,649 101.6 29.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jippy6v6",
   "source_url": "https://omlx.ai/benchmarks/performance/jippy6v6",
   "tg_tokens": null,
   "tps": 101.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jippy6v6-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.6 GB at this context; duplicate traces of the same session: 0xfhevqa, 9vw968j9, u45iqp5e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5150.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 5,150 136.9 30.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jippy6v6",
   "source_url": "https://omlx.ai/benchmarks/performance/jippy6v6",
   "tg_tokens": null,
   "tps": 136.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jippy6v6-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.7 GB at this context; duplicate traces of the same session: 0xfhevqa, 9vw968j9, u45iqp5e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4987.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 4,987 117.6 30.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jippy6v6",
   "source_url": "https://omlx.ai/benchmarks/performance/jippy6v6",
   "tg_tokens": null,
   "tps": 117.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jippy6v6-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.9 GB at this context; duplicate traces of the same session: 0xfhevqa, 9vw968j9, u45iqp5e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4322.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 4,322 143.3 30.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jippy6v6",
   "source_url": "https://omlx.ai/benchmarks/performance/jippy6v6",
   "tg_tokens": null,
   "tps": 143.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jippy6v6-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.2 GB at this context; duplicate traces of the same session: 0xfhevqa, 9vw968j9, u45iqp5e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3247.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 3,247 183.5 31.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jippy6v6",
   "source_url": "https://omlx.ai/benchmarks/performance/jippy6v6",
   "tg_tokens": null,
   "tps": 183.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jippy6v6-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: 0xfhevqa, 9vw968j9, u45iqp5e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2239.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k 2,239 109.9 32.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jippy6v6",
   "source_url": "https://omlx.ai/benchmarks/performance/jippy6v6",
   "tg_tokens": null,
   "tps": 109.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4cqkxvpd-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: 680i8xka, hedzlz4o, jc40sdto; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 879.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 879.2 53.8 17.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4cqkxvpd",
   "source_url": "https://omlx.ai/benchmarks/performance/4cqkxvpd",
   "tg_tokens": null,
   "tps": 53.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4cqkxvpd-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 680i8xka, hedzlz4o, jc40sdto; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 924.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 924.3 59.0 18.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4cqkxvpd",
   "source_url": "https://omlx.ai/benchmarks/performance/4cqkxvpd",
   "tg_tokens": null,
   "tps": 59.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4cqkxvpd-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 680i8xka, hedzlz4o, jc40sdto; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 770.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 770.8 55.0 19.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4cqkxvpd",
   "source_url": "https://omlx.ai/benchmarks/performance/4cqkxvpd",
   "tg_tokens": null,
   "tps": 55.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4cqkxvpd-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 680i8xka, hedzlz4o, jc40sdto; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 722.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 722.0 58.3 19.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4cqkxvpd",
   "source_url": "https://omlx.ai/benchmarks/performance/4cqkxvpd",
   "tg_tokens": null,
   "tps": 58.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4cqkxvpd-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 680i8xka, hedzlz4o, jc40sdto; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 673.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 673.8 120.8 20.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4cqkxvpd",
   "source_url": "https://omlx.ai/benchmarks/performance/4cqkxvpd",
   "tg_tokens": null,
   "tps": 120.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4cqkxvpd-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: 680i8xka, hedzlz4o, jc40sdto; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 544.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 544.8 33.5 23.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4cqkxvpd",
   "source_url": "https://omlx.ai/benchmarks/performance/4cqkxvpd",
   "tg_tokens": null,
   "tps": 33.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-6p7ydxpy-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 32.7 GB; TTFT 51812 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 632.4,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 2bit; 32k tokens 632.4 PP tok/s 16.2 TG tok/s 51812 TTFT (ms) 32.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6p7ydxpy",
   "source_url": "https://omlx.ai/benchmarks/performance/6p7ydxpy",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ov3zsc0b-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 32.1 GB; TTFT 28542 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 574.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 2bit; 16k tokens 574.0 PP tok/s 18.8 TG tok/s 28542 TTFT (ms) 32.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ov3zsc0b",
   "source_url": "https://omlx.ai/benchmarks/performance/ov3zsc0b",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-0xzw795l-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.8 GB at this context; duplicate traces of the same session: vnruehph; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 94.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 94.1 13.2 32.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0xzw795l",
   "source_url": "https://omlx.ai/benchmarks/performance/0xzw795l",
   "tg_tokens": null,
   "tps": 13.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-0xzw795l-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.6 GB at this context; duplicate traces of the same session: vnruehph; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 431.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 431.3 13.6 34.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0xzw795l",
   "source_url": "https://omlx.ai/benchmarks/performance/0xzw795l",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-0xzw795l-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.4 GB at this context; duplicate traces of the same session: vnruehph; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 463.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 463.8 14.2 35.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0xzw795l",
   "source_url": "https://omlx.ai/benchmarks/performance/0xzw795l",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-0xzw795l-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.0 GB at this context; duplicate traces of the same session: vnruehph; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 460.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 460.7 11.1 36.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0xzw795l",
   "source_url": "https://omlx.ai/benchmarks/performance/0xzw795l",
   "tg_tokens": null,
   "tps": 11.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-i2jyjm0l-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 139.4 GB at this context; duplicate traces of the same session: 2o5zucr9, 2qvkrz48, 96pmd20n, eb2ing9j, m2wdlh03; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2093.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 1k 2,093 83.8 139.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i2jyjm0l",
   "source_url": "https://omlx.ai/benchmarks/performance/i2jyjm0l",
   "tg_tokens": null,
   "tps": 83.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-i2jyjm0l-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 140.0 GB at this context; duplicate traces of the same session: 2o5zucr9, 2qvkrz48, 96pmd20n, eb2ing9j, m2wdlh03; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3877.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 4k 3,877 79.4 140.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i2jyjm0l",
   "source_url": "https://omlx.ai/benchmarks/performance/i2jyjm0l",
   "tg_tokens": null,
   "tps": 79.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-i2jyjm0l-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 141.1 GB at this context; duplicate traces of the same session: 2o5zucr9, 2qvkrz48, 96pmd20n, eb2ing9j, m2wdlh03; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4296.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 8k 4,296 79.9 141.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i2jyjm0l",
   "source_url": "https://omlx.ai/benchmarks/performance/i2jyjm0l",
   "tg_tokens": null,
   "tps": 79.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-i2jyjm0l-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.8 GB at this context; duplicate traces of the same session: 2o5zucr9, 2qvkrz48, 96pmd20n, eb2ing9j, m2wdlh03; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4232.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 64k 4,232 72.3 142.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i2jyjm0l",
   "source_url": "https://omlx.ai/benchmarks/performance/i2jyjm0l",
   "tg_tokens": null,
   "tps": 72.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-i2jyjm0l-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.1 GB at this context; duplicate traces of the same session: 2o5zucr9, 2qvkrz48, 96pmd20n, eb2ing9j, m2wdlh03; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4115.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 128k 4,115 66.3 145.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i2jyjm0l",
   "source_url": "https://omlx.ai/benchmarks/performance/i2jyjm0l",
   "tg_tokens": null,
   "tps": 66.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-i2jyjm0l-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 148.7 GB at this context; duplicate traces of the same session: 2o5zucr9, 2qvkrz48, 96pmd20n, eb2ing9j, m2wdlh03; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4008.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 195k 4,008 66.7 148.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i2jyjm0l",
   "source_url": "https://omlx.ai/benchmarks/performance/i2jyjm0l",
   "tg_tokens": null,
   "tps": 66.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-y3dpcnsm-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.6 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2008.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 1k 2,008 131.3 142.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y3dpcnsm",
   "source_url": "https://omlx.ai/benchmarks/performance/y3dpcnsm",
   "tg_tokens": null,
   "tps": 131.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-y3dpcnsm-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3776.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 4k 3,776 111.2 143.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y3dpcnsm",
   "source_url": "https://omlx.ai/benchmarks/performance/y3dpcnsm",
   "tg_tokens": null,
   "tps": 111.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-y3dpcnsm-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 144.5 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4161.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 8k 4,161 119.6 144.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y3dpcnsm",
   "source_url": "https://omlx.ai/benchmarks/performance/y3dpcnsm",
   "tg_tokens": null,
   "tps": 119.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-y3dpcnsm-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.1 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4178.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 16k 4,178 101.3 145.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y3dpcnsm",
   "source_url": "https://omlx.ai/benchmarks/performance/y3dpcnsm",
   "tg_tokens": null,
   "tps": 101.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-y3dpcnsm-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.7 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4164.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 32k 4,164 85.7 145.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y3dpcnsm",
   "source_url": "https://omlx.ai/benchmarks/performance/y3dpcnsm",
   "tg_tokens": null,
   "tps": 85.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-y3dpcnsm-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 146.8 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4092.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 64k 4,092 118.3 146.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y3dpcnsm",
   "source_url": "https://omlx.ai/benchmarks/performance/y3dpcnsm",
   "tg_tokens": null,
   "tps": 118.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-y3dpcnsm-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 150.6 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3985.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 128k 3,985 95.1 150.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y3dpcnsm",
   "source_url": "https://omlx.ai/benchmarks/performance/y3dpcnsm",
   "tg_tokens": null,
   "tps": 95.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-y3dpcnsm-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 153.3 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3871.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 195k 3,871 76.8 153.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y3dpcnsm",
   "source_url": "https://omlx.ai/benchmarks/performance/y3dpcnsm",
   "tg_tokens": null,
   "tps": 76.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-publ7j1s-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 954.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "qwen3.8-flash-next-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 954.8 82.9 101.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/publ7j1s",
   "source_url": "https://omlx.ai/benchmarks/performance/publ7j1s",
   "tg_tokens": null,
   "tps": 82.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-publ7j1s-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1376.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "qwen3.8-flash-next-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 1,376 22.0 101.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/publ7j1s",
   "source_url": "https://omlx.ai/benchmarks/performance/publ7j1s",
   "tg_tokens": null,
   "tps": 22.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "outlier"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-04thsaya-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: 5v140v3c, uf30iv0m; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2195.0,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Ornith-1.5-35B-A3B-oQ3e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 3bit; 4k 2,195 94.2 16.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/04thsaya",
   "source_url": "https://omlx.ai/benchmarks/performance/04thsaya",
   "tg_tokens": null,
   "tps": 94.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "outlier"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-04thsaya-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: 5v140v3c, uf30iv0m; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2343.0,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Ornith-1.5-35B-A3B-oQ3e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 3bit; 16k 2,343 83.0 17.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/04thsaya",
   "source_url": "https://omlx.ai/benchmarks/performance/04thsaya",
   "tg_tokens": null,
   "tps": 83.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "outlier"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-04thsaya-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: 5v140v3c, uf30iv0m; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2001.0,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Ornith-1.5-35B-A3B-oQ3e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 3bit; 32k 2,001 72.1 17.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/04thsaya",
   "source_url": "https://omlx.ai/benchmarks/performance/04thsaya",
   "tg_tokens": null,
   "tps": 72.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3b94a37c-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: yhdis7g3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2649.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 2,649 101.6 29.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3b94a37c",
   "source_url": "https://omlx.ai/benchmarks/performance/3b94a37c",
   "tg_tokens": null,
   "tps": 101.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3b94a37c-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.6 GB at this context; duplicate traces of the same session: yhdis7g3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5150.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 5,150 136.9 30.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3b94a37c",
   "source_url": "https://omlx.ai/benchmarks/performance/3b94a37c",
   "tg_tokens": null,
   "tps": 136.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3b94a37c-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.7 GB at this context; duplicate traces of the same session: yhdis7g3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4987.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 4,987 117.6 30.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3b94a37c",
   "source_url": "https://omlx.ai/benchmarks/performance/3b94a37c",
   "tg_tokens": null,
   "tps": 117.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3b94a37c-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.9 GB at this context; duplicate traces of the same session: yhdis7g3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4322.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 4,322 143.3 30.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3b94a37c",
   "source_url": "https://omlx.ai/benchmarks/performance/3b94a37c",
   "tg_tokens": null,
   "tps": 143.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3b94a37c-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.2 GB at this context; duplicate traces of the same session: yhdis7g3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3247.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 3,247 183.5 31.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3b94a37c",
   "source_url": "https://omlx.ai/benchmarks/performance/3b94a37c",
   "tg_tokens": null,
   "tps": 183.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3b94a37c-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: yhdis7g3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2239.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k 2,239 109.9 32.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3b94a37c",
   "source_url": "https://omlx.ai/benchmarks/performance/3b94a37c",
   "tg_tokens": null,
   "tps": 109.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-sx01y45q-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: 24etunj5, mqq1cr88, o8i20f4f; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1318.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k 1,318 89.6 17.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sx01y45q",
   "source_url": "https://omlx.ai/benchmarks/performance/sx01y45q",
   "tg_tokens": null,
   "tps": 89.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-sx01y45q-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 24etunj5, mqq1cr88, o8i20f4f; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1563.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 4k 1,563 87.0 18.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sx01y45q",
   "source_url": "https://omlx.ai/benchmarks/performance/sx01y45q",
   "tg_tokens": null,
   "tps": 87.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-sx01y45q-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 24etunj5, mqq1cr88, o8i20f4f; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1561.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 8k 1,561 110.5 19.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sx01y45q",
   "source_url": "https://omlx.ai/benchmarks/performance/sx01y45q",
   "tg_tokens": null,
   "tps": 110.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-sx01y45q-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 24etunj5, mqq1cr88, o8i20f4f; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1512.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 16k 1,512 108.7 19.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sx01y45q",
   "source_url": "https://omlx.ai/benchmarks/performance/sx01y45q",
   "tg_tokens": null,
   "tps": 108.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-sx01y45q-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 24etunj5, mqq1cr88, o8i20f4f; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1404.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 32k 1,404 123.0 20.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sx01y45q",
   "source_url": "https://omlx.ai/benchmarks/performance/sx01y45q",
   "tg_tokens": null,
   "tps": 123.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-4mtxpb7f-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 45.0 GB; TTFT 49707 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 659.2,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 2bit; 32k tokens 659.2 PP tok/s 21.3 TG tok/s 49707 TTFT (ms) 45.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4mtxpb7f",
   "source_url": "https://omlx.ai/benchmarks/performance/4mtxpb7f",
   "tg_tokens": null,
   "tps": 21.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-r9tp2lwy-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.6 GB at this context; duplicate traces of the same session: 0qiyq42o, 87fpz9gc, cl7zzpv0, fwbarzsz, mhbl6m97, s3688lmk; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2008.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 1k 2,008 131.3 142.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r9tp2lwy",
   "source_url": "https://omlx.ai/benchmarks/performance/r9tp2lwy",
   "tg_tokens": null,
   "tps": 131.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-r9tp2lwy-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: 0qiyq42o, 87fpz9gc, cl7zzpv0, fwbarzsz, mhbl6m97, s3688lmk; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3776.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 4k 3,776 111.2 143.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r9tp2lwy",
   "source_url": "https://omlx.ai/benchmarks/performance/r9tp2lwy",
   "tg_tokens": null,
   "tps": 111.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-r9tp2lwy-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 144.5 GB at this context; duplicate traces of the same session: 0qiyq42o, 87fpz9gc, cl7zzpv0, fwbarzsz, mhbl6m97, s3688lmk; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4161.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 8k 4,161 119.6 144.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r9tp2lwy",
   "source_url": "https://omlx.ai/benchmarks/performance/r9tp2lwy",
   "tg_tokens": null,
   "tps": 119.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-r9tp2lwy-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.1 GB at this context; duplicate traces of the same session: 0qiyq42o, 87fpz9gc, cl7zzpv0, fwbarzsz, mhbl6m97, s3688lmk; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4178.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 16k 4,178 101.3 145.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r9tp2lwy",
   "source_url": "https://omlx.ai/benchmarks/performance/r9tp2lwy",
   "tg_tokens": null,
   "tps": 101.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-r9tp2lwy-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.7 GB at this context; duplicate traces of the same session: 0qiyq42o, 87fpz9gc, cl7zzpv0, fwbarzsz, mhbl6m97, s3688lmk; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4164.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 32k 4,164 85.7 145.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r9tp2lwy",
   "source_url": "https://omlx.ai/benchmarks/performance/r9tp2lwy",
   "tg_tokens": null,
   "tps": 85.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-r9tp2lwy-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 146.8 GB at this context; duplicate traces of the same session: 0qiyq42o, 87fpz9gc, cl7zzpv0, fwbarzsz, mhbl6m97, s3688lmk; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4092.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 64k 4,092 118.3 146.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r9tp2lwy",
   "source_url": "https://omlx.ai/benchmarks/performance/r9tp2lwy",
   "tg_tokens": null,
   "tps": 118.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-r9tp2lwy-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 150.6 GB at this context; duplicate traces of the same session: 0qiyq42o, 87fpz9gc, cl7zzpv0, fwbarzsz, mhbl6m97, s3688lmk; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3985.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 128k 3,985 95.1 150.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r9tp2lwy",
   "source_url": "https://omlx.ai/benchmarks/performance/r9tp2lwy",
   "tg_tokens": null,
   "tps": 95.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-r9tp2lwy-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 153.3 GB at this context; duplicate traces of the same session: 0qiyq42o, 87fpz9gc, cl7zzpv0, fwbarzsz, mhbl6m97, s3688lmk; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3871.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 195k 3,871 76.8 153.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r9tp2lwy",
   "source_url": "https://omlx.ai/benchmarks/performance/r9tp2lwy",
   "tg_tokens": null,
   "tps": 76.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-0dom8yeb-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.6 GB at this context; duplicate traces of the same session: 0p1z4lkd, 6oc2bbgg; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1995.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 1k 1,995 129.8 142.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0dom8yeb",
   "source_url": "https://omlx.ai/benchmarks/performance/0dom8yeb",
   "tg_tokens": null,
   "tps": 129.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-0dom8yeb-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 144.5 GB at this context; duplicate traces of the same session: 0p1z4lkd, 6oc2bbgg; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4166.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 8k 4,166 108.6 144.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0dom8yeb",
   "source_url": "https://omlx.ai/benchmarks/performance/0dom8yeb",
   "tg_tokens": null,
   "tps": 108.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-0dom8yeb-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.7 GB at this context; duplicate traces of the same session: 0p1z4lkd, 6oc2bbgg; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4167.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 32k 4,167 86.0 145.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0dom8yeb",
   "source_url": "https://omlx.ai/benchmarks/performance/0dom8yeb",
   "tg_tokens": null,
   "tps": 86.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-0dom8yeb-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 146.8 GB at this context; duplicate traces of the same session: 0p1z4lkd, 6oc2bbgg; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4096.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 64k 4,096 112.7 146.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0dom8yeb",
   "source_url": "https://omlx.ai/benchmarks/performance/0dom8yeb",
   "tg_tokens": null,
   "tps": 112.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-0dom8yeb-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 150.6 GB at this context; duplicate traces of the same session: 0p1z4lkd, 6oc2bbgg; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3991.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 128k 3,991 96.3 150.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0dom8yeb",
   "source_url": "https://omlx.ai/benchmarks/performance/0dom8yeb",
   "tg_tokens": null,
   "tps": 96.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-0dom8yeb-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 153.3 GB at this context; duplicate traces of the same session: 0p1z4lkd, 6oc2bbgg; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3834.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 195k 3,834 78.1 153.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0dom8yeb",
   "source_url": "https://omlx.ai/benchmarks/performance/0dom8yeb",
   "tg_tokens": null,
   "tps": 78.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ctvincib-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1532.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,532 97.8 20.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ctvincib",
   "source_url": "https://omlx.ai/benchmarks/performance/ctvincib",
   "tg_tokens": null,
   "tps": 97.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ctvincib-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 626.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 626.2 95.3 21.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ctvincib",
   "source_url": "https://omlx.ai/benchmarks/performance/ctvincib",
   "tg_tokens": null,
   "tps": 95.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ctvincib-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1629.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,629 86.2 22.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ctvincib",
   "source_url": "https://omlx.ai/benchmarks/performance/ctvincib",
   "tg_tokens": null,
   "tps": 86.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ctvincib-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1475.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,475 75.6 21.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ctvincib",
   "source_url": "https://omlx.ai/benchmarks/performance/ctvincib",
   "tg_tokens": null,
   "tps": 75.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ctvincib-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1203.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,203 61.6 21.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ctvincib",
   "source_url": "https://omlx.ai/benchmarks/performance/ctvincib",
   "tg_tokens": null,
   "tps": 61.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ctvincib-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 890.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 890.5 42.9 22.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ctvincib",
   "source_url": "https://omlx.ai/benchmarks/performance/ctvincib",
   "tg_tokens": null,
   "tps": 42.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ctvincib-ctx131072",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 580.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 580.1 26.4 23.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ctvincib",
   "source_url": "https://omlx.ai/benchmarks/performance/ctvincib",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ye4zldgp-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 13.3 GB at this context; duplicate traces of the same session: l73h2m8r, ptscctcp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1969.0,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "Ornith-1.5-35B-A3B-oQ2e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 2bit; 4k 1,969 93.1 13.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ye4zldgp",
   "source_url": "https://omlx.ai/benchmarks/performance/ye4zldgp",
   "tg_tokens": null,
   "tps": 93.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ye4zldgp-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 13.5 GB at this context; duplicate traces of the same session: l73h2m8r, ptscctcp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2409.0,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "Ornith-1.5-35B-A3B-oQ2e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 2bit; 16k 2,409 82.5 13.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ye4zldgp",
   "source_url": "https://omlx.ai/benchmarks/performance/ye4zldgp",
   "tg_tokens": null,
   "tps": 82.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ye4zldgp-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 13.8 GB at this context; duplicate traces of the same session: l73h2m8r, ptscctcp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2051.0,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "Ornith-1.5-35B-A3B-oQ2e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 2bit; 32k 2,051 71.5 13.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ye4zldgp",
   "source_url": "https://omlx.ai/benchmarks/performance/ye4zldgp",
   "tg_tokens": null,
   "tps": 71.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-1uah6jnl-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 30.1 GB; TTFT 15094 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1086.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-6bit on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 6bit; 16k tokens 1,086 PP tok/s 59.3 TG tok/s 15094 TTFT (ms) 30.1 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1uah6jnl",
   "source_url": "https://omlx.ai/benchmarks/performance/1uah6jnl",
   "tg_tokens": null,
   "tps": 59.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-2p5x23kn-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.8 GB; TTFT 9563 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 107.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4Bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k tokens 107.1 PP tok/s 19.0 TG tok/s 9563 TTFT (ms) 19.8 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2p5x23kn",
   "source_url": "https://omlx.ai/benchmarks/performance/2p5x23kn",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-jjhes2u0-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 34.2 GB; TTFT 9676 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-8Bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 8bit; 1k tokens 105.8 PP tok/s 15.2 TG tok/s 9676 TTFT (ms) 34.2 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jjhes2u0",
   "source_url": "https://omlx.ai/benchmarks/performance/jjhes2u0",
   "tg_tokens": null,
   "tps": 15.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-r2jmtac6-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 2sauk8f2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 565.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 565.9 35.2 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r2jmtac6",
   "source_url": "https://omlx.ai/benchmarks/performance/r2jmtac6",
   "tg_tokens": null,
   "tps": 35.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-r2jmtac6-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: 2sauk8f2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 599.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 599.1 64.2 23.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r2jmtac6",
   "source_url": "https://omlx.ai/benchmarks/performance/r2jmtac6",
   "tg_tokens": null,
   "tps": 64.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-js1msji8-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1318.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k 1,318 89.6 17.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/js1msji8",
   "source_url": "https://omlx.ai/benchmarks/performance/js1msji8",
   "tg_tokens": null,
   "tps": 89.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-js1msji8-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1563.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 4k 1,563 87.0 18.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/js1msji8",
   "source_url": "https://omlx.ai/benchmarks/performance/js1msji8",
   "tg_tokens": null,
   "tps": 87.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-js1msji8-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1561.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 8k 1,561 110.5 19.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/js1msji8",
   "source_url": "https://omlx.ai/benchmarks/performance/js1msji8",
   "tg_tokens": null,
   "tps": 110.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-js1msji8-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1512.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 16k 1,512 108.7 19.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/js1msji8",
   "source_url": "https://omlx.ai/benchmarks/performance/js1msji8",
   "tg_tokens": null,
   "tps": 108.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-js1msji8-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1404.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 32k 1,404 123.0 20.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/js1msji8",
   "source_url": "https://omlx.ai/benchmarks/performance/js1msji8",
   "tg_tokens": null,
   "tps": 123.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-1faxyi8j-ctx1024",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "APEX MiniPlus Abliterated MLX build; oMLX community benchmark trace; peak memory 7.5 GB at this context; duplicate traces of the same session: kfxd5kje; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 152.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-35B-A3B-Distill-MLX-APEX-MiniPlus-Abliterated on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 152.9 5.3 7.5 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1faxyi8j",
   "source_url": "https://omlx.ai/benchmarks/performance/1faxyi8j",
   "tg_tokens": null,
   "tps": 5.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-1faxyi8j-ctx4096",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "APEX MiniPlus Abliterated MLX build; oMLX community benchmark trace; peak memory 9.3 GB at this context; duplicate traces of the same session: kfxd5kje; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 211.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-35B-A3B-Distill-MLX-APEX-MiniPlus-Abliterated on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 211.3 5.3 9.3 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1faxyi8j",
   "source_url": "https://omlx.ai/benchmarks/performance/1faxyi8j",
   "tg_tokens": null,
   "tps": 5.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-rg63yd18-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.6 GB at this context; duplicate traces of the same session: 4d02c7nv, 6nytund7; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1995.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 1k 1,995 129.8 142.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rg63yd18",
   "source_url": "https://omlx.ai/benchmarks/performance/rg63yd18",
   "tg_tokens": null,
   "tps": 129.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-rg63yd18-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 144.5 GB at this context; duplicate traces of the same session: 4d02c7nv, 6nytund7; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4166.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 8k 4,166 108.6 144.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rg63yd18",
   "source_url": "https://omlx.ai/benchmarks/performance/rg63yd18",
   "tg_tokens": null,
   "tps": 108.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-rg63yd18-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.7 GB at this context; duplicate traces of the same session: 4d02c7nv, 6nytund7; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4167.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 32k 4,167 86.0 145.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rg63yd18",
   "source_url": "https://omlx.ai/benchmarks/performance/rg63yd18",
   "tg_tokens": null,
   "tps": 86.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-rg63yd18-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 146.8 GB at this context; duplicate traces of the same session: 4d02c7nv, 6nytund7; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4096.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 64k 4,096 112.7 146.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rg63yd18",
   "source_url": "https://omlx.ai/benchmarks/performance/rg63yd18",
   "tg_tokens": null,
   "tps": 112.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-rg63yd18-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 150.6 GB at this context; duplicate traces of the same session: 4d02c7nv, 6nytund7; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3991.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 128k 3,991 96.3 150.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rg63yd18",
   "source_url": "https://omlx.ai/benchmarks/performance/rg63yd18",
   "tg_tokens": null,
   "tps": 96.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-rg63yd18-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 153.3 GB at this context; duplicate traces of the same session: 4d02c7nv, 6nytund7; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3834.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 195k 3,834 78.1 153.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rg63yd18",
   "source_url": "https://omlx.ai/benchmarks/performance/rg63yd18",
   "tg_tokens": null,
   "tps": 78.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-5vojqn4c-ctx1024",
   "model": "Qwythos-9B-Claude-Mythos-5-1M-BF16-MLX-MTP-Visual",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 17.6 GB; TTFT 1080 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 947.9,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwythos-9B-Claude-Mythos-5-1M-BF16-MLX-MTP-Visual on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 bf16; 1k tokens 947.9 PP tok/s 28.8 TG tok/s 1080 TTFT (ms) 17.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5vojqn4c",
   "source_url": "https://omlx.ai/benchmarks/performance/5vojqn4c",
   "tg_tokens": null,
   "tps": 28.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-2u8k17to-ctx1024",
   "model": "Qwythos-9B-Claude-Mythos-5-1M-BF16-MLX-MTP-Visual",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 17.6 GB; TTFT 1078 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 949.5,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwythos-9B-Claude-Mythos-5-1M-BF16-MLX-MTP-Visual on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 bf16; 1k tokens 949.5 PP tok/s 28.8 TG tok/s 1078 TTFT (ms) 17.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2u8k17to",
   "source_url": "https://omlx.ai/benchmarks/performance/2u8k17to",
   "tg_tokens": null,
   "tps": 28.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-tkgeh54d-ctx1024",
   "model": "Qwythos-9B-Claude-Mythos-5-1M-BF16-MLX-MTP-Visual",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 17.6 GB; TTFT 1116 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 917.7,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwythos-9B-Claude-Mythos-5-1M-BF16-MLX-MTP-Visual on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 bf16; 1k tokens 917.7 PP tok/s 28.8 TG tok/s 1116 TTFT (ms) 17.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tkgeh54d",
   "source_url": "https://omlx.ai/benchmarks/performance/tkgeh54d",
   "tg_tokens": null,
   "tps": 28.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-esu220ui-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Youssofal MTPLX Optimized Speed; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: n3tg8ikv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 719.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Youssofal--Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 719.8 34.7 23.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/esu220ui",
   "source_url": "https://omlx.ai/benchmarks/performance/esu220ui",
   "tg_tokens": null,
   "tps": 34.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-esu220ui-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "Youssofal MTPLX Optimized Speed; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: n3tg8ikv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 926.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Youssofal--Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 926.2 36.9 24.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/esu220ui",
   "source_url": "https://omlx.ai/benchmarks/performance/esu220ui",
   "tg_tokens": null,
   "tps": 36.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-yczpdsti-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 80ohh5l9, az2r4gnf, bvjguheh, gic23oqd, u257rd13; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1532.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,532 97.8 20.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yczpdsti",
   "source_url": "https://omlx.ai/benchmarks/performance/yczpdsti",
   "tg_tokens": null,
   "tps": 97.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-yczpdsti-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 80ohh5l9, az2r4gnf, bvjguheh, gic23oqd, u257rd13; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 626.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 626.2 95.3 21.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yczpdsti",
   "source_url": "https://omlx.ai/benchmarks/performance/yczpdsti",
   "tg_tokens": null,
   "tps": 95.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-yczpdsti-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 80ohh5l9, az2r4gnf, bvjguheh, gic23oqd, u257rd13; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1629.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,629 86.2 22.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yczpdsti",
   "source_url": "https://omlx.ai/benchmarks/performance/yczpdsti",
   "tg_tokens": null,
   "tps": 86.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-yczpdsti-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 80ohh5l9, az2r4gnf, bvjguheh, gic23oqd, u257rd13; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1475.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,475 75.6 21.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yczpdsti",
   "source_url": "https://omlx.ai/benchmarks/performance/yczpdsti",
   "tg_tokens": null,
   "tps": 75.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-yczpdsti-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: 80ohh5l9, az2r4gnf, bvjguheh, gic23oqd, u257rd13; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1203.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,203 61.6 21.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yczpdsti",
   "source_url": "https://omlx.ai/benchmarks/performance/yczpdsti",
   "tg_tokens": null,
   "tps": 61.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-yczpdsti-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 80ohh5l9, az2r4gnf, bvjguheh, gic23oqd, u257rd13; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 890.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 890.5 42.9 22.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yczpdsti",
   "source_url": "https://omlx.ai/benchmarks/performance/yczpdsti",
   "tg_tokens": null,
   "tps": 42.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-yczpdsti-ctx131072",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 80ohh5l9, az2r4gnf, bvjguheh, gic23oqd, u257rd13; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 580.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 580.1 26.4 23.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yczpdsti",
   "source_url": "https://omlx.ai/benchmarks/performance/yczpdsti",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-adcts9rs-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 30.1 GB; TTFT 14231 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1151.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-6bit on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 6bit; 16k tokens 1,151 PP tok/s 60.0 TG tok/s 14231 TTFT (ms) 30.1 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/adcts9rs",
   "source_url": "https://omlx.ai/benchmarks/performance/adcts9rs",
   "tg_tokens": null,
   "tps": 60.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-6zf3h8sf-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: 2pal9fh0; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1145.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-6bit on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 6bit; 1k 1,145 69.0 27.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6zf3h8sf",
   "source_url": "https://omlx.ai/benchmarks/performance/6zf3h8sf",
   "tg_tokens": null,
   "tps": 69.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-6zf3h8sf-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 29.3 GB at this context; duplicate traces of the same session: 2pal9fh0; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1356.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-6bit on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 6bit; 4k 1,356 66.8 29.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6zf3h8sf",
   "source_url": "https://omlx.ai/benchmarks/performance/6zf3h8sf",
   "tg_tokens": null,
   "tps": 66.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-4zmst14u-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.8 GB; TTFT 9596 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4Bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k tokens 106.7 PP tok/s 19.0 TG tok/s 9596 TTFT (ms) 19.8 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4zmst14u",
   "source_url": "https://omlx.ai/benchmarks/performance/4zmst14u",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-dw33uzcd-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: afjvls53; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 548.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 548.2 38.3 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dw33uzcd",
   "source_url": "https://omlx.ai/benchmarks/performance/dw33uzcd",
   "tg_tokens": null,
   "tps": 38.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-dw33uzcd-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: afjvls53; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 619.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 619.5 73.5 23.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dw33uzcd",
   "source_url": "https://omlx.ai/benchmarks/performance/dw33uzcd",
   "tg_tokens": null,
   "tps": 73.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-r20v78k3-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 45.3 GB; TTFT 10842 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 94.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 94.5 PP tok/s 15.6 TG tok/s 10842 TTFT (ms) 45.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r20v78k3",
   "source_url": "https://omlx.ai/benchmarks/performance/r20v78k3",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-7ve7scza-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 42.3 GB at this context; duplicate traces of the same session: y1z0szhl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 94.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 94.6 14.8 42.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ve7scza",
   "source_url": "https://omlx.ai/benchmarks/performance/7ve7scza",
   "tg_tokens": null,
   "tps": 14.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-7ve7scza-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 42.9 GB at this context; duplicate traces of the same session: y1z0szhl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 343.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 343.9 14.7 42.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ve7scza",
   "source_url": "https://omlx.ai/benchmarks/performance/7ve7scza",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-7ve7scza-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 44.1 GB at this context; duplicate traces of the same session: y1z0szhl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 460.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 460.4 13.2 44.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ve7scza",
   "source_url": "https://omlx.ai/benchmarks/performance/7ve7scza",
   "tg_tokens": null,
   "tps": 13.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-7ve7scza-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 44.9 GB at this context; duplicate traces of the same session: y1z0szhl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 379.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 379.0 13.4 44.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ve7scza",
   "source_url": "https://omlx.ai/benchmarks/performance/7ve7scza",
   "tg_tokens": null,
   "tps": 13.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8ixmx5v5-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 71.0 GB; TTFT 950 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1078.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,078 PP tok/s 71.8 TG tok/s 950 TTFT (ms) 71.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8ixmx5v5",
   "source_url": "https://omlx.ai/benchmarks/performance/8ixmx5v5",
   "tg_tokens": null,
   "tps": 71.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-77w23ibf-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; Swift 1.5 build; oMLX community benchmark trace, single run at 4k context; peak memory 92.0 GB; TTFT 20746 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 197.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 197.4 PP tok/s 21.5 TG tok/s 20746 TTFT (ms) 92.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/77w23ibf",
   "source_url": "https://omlx.ai/benchmarks/performance/77w23ibf",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5q96eh75-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.2 GB at this context; duplicate traces of the same session: jp3ivzea; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3997.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 3,997 90.3 26.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5q96eh75",
   "source_url": "https://omlx.ai/benchmarks/performance/5q96eh75",
   "tg_tokens": null,
   "tps": 90.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5q96eh75-ctx16384",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: jp3ivzea; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3728.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 16k 3,728 80.2 26.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5q96eh75",
   "source_url": "https://omlx.ai/benchmarks/performance/5q96eh75",
   "tg_tokens": null,
   "tps": 80.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-an6pzf2o-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4043.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 4,043 90.1 26.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/an6pzf2o",
   "source_url": "https://omlx.ai/benchmarks/performance/an6pzf2o",
   "tg_tokens": null,
   "tps": 90.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-an6pzf2o-ctx16384",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3756.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 16k 3,756 80.9 26.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/an6pzf2o",
   "source_url": "https://omlx.ai/benchmarks/performance/an6pzf2o",
   "tg_tokens": null,
   "tps": 80.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-q0amf18q-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: axvtmeec, mwzlhnlu, na7lyl8m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1958.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,958 141.0 22.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q0amf18q",
   "source_url": "https://omlx.ai/benchmarks/performance/q0amf18q",
   "tg_tokens": null,
   "tps": 141.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-q0amf18q-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: axvtmeec, mwzlhnlu, na7lyl8m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1514.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 1,514 200.4 22.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q0amf18q",
   "source_url": "https://omlx.ai/benchmarks/performance/q0amf18q",
   "tg_tokens": null,
   "tps": 200.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-q0amf18q-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: axvtmeec, mwzlhnlu, na7lyl8m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1169.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 1,169 201.7 22.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q0amf18q",
   "source_url": "https://omlx.ai/benchmarks/performance/q0amf18q",
   "tg_tokens": null,
   "tps": 201.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-q0amf18q-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: axvtmeec, mwzlhnlu, na7lyl8m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 838.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 64k 838.1 98.5 23.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q0amf18q",
   "source_url": "https://omlx.ai/benchmarks/performance/q0amf18q",
   "tg_tokens": null,
   "tps": 98.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-cmd4p3w7-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: 0tezzy54; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1841.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,841 157.7 22.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cmd4p3w7",
   "source_url": "https://omlx.ai/benchmarks/performance/cmd4p3w7",
   "tg_tokens": null,
   "tps": 157.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-cmd4p3w7-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: 0tezzy54; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1611.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 1,611 142.9 22.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cmd4p3w7",
   "source_url": "https://omlx.ai/benchmarks/performance/cmd4p3w7",
   "tg_tokens": null,
   "tps": 142.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-h45lirby-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 25.4 GB; TTFT 14001 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 292.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 292.6 PP tok/s 68.4 TG tok/s 14001 TTFT (ms) 25.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h45lirby",
   "source_url": "https://omlx.ai/benchmarks/performance/h45lirby",
   "tg_tokens": null,
   "tps": 68.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-k4ra67lk-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 42.3 GB at this context; duplicate traces of the same session: tcyssnqg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 94.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 94.6 14.8 42.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k4ra67lk",
   "source_url": "https://omlx.ai/benchmarks/performance/k4ra67lk",
   "tg_tokens": null,
   "tps": 14.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-k4ra67lk-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 42.9 GB at this context; duplicate traces of the same session: tcyssnqg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 343.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 343.9 14.7 42.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k4ra67lk",
   "source_url": "https://omlx.ai/benchmarks/performance/k4ra67lk",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-k4ra67lk-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 44.1 GB at this context; duplicate traces of the same session: tcyssnqg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 460.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 460.4 13.2 44.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k4ra67lk",
   "source_url": "https://omlx.ai/benchmarks/performance/k4ra67lk",
   "tg_tokens": null,
   "tps": 13.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-k4ra67lk-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 44.9 GB at this context; duplicate traces of the same session: tcyssnqg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 379.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 379.0 13.4 44.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k4ra67lk",
   "source_url": "https://omlx.ai/benchmarks/performance/k4ra67lk",
   "tg_tokens": null,
   "tps": 13.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-eqwvlehf-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.5 GB at this context; duplicate traces of the same session: 386bch9o, dv1rn766, n0uuou4f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1127.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,127 79.1 70.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eqwvlehf",
   "source_url": "https://omlx.ai/benchmarks/performance/eqwvlehf",
   "tg_tokens": null,
   "tps": 79.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-eqwvlehf-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 386bch9o, dv1rn766, n0uuou4f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2015.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,015 60.8 71.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eqwvlehf",
   "source_url": "https://omlx.ai/benchmarks/performance/eqwvlehf",
   "tg_tokens": null,
   "tps": 60.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-eqwvlehf-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: 386bch9o, dv1rn766, n0uuou4f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2132.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,132 92.6 72.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eqwvlehf",
   "source_url": "https://omlx.ai/benchmarks/performance/eqwvlehf",
   "tg_tokens": null,
   "tps": 92.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-eqwvlehf-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: 386bch9o, dv1rn766, n0uuou4f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2035.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,035 65.0 73.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eqwvlehf",
   "source_url": "https://omlx.ai/benchmarks/performance/eqwvlehf",
   "tg_tokens": null,
   "tps": 65.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-eqwvlehf-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.6 GB at this context; duplicate traces of the same session: 386bch9o, dv1rn766, n0uuou4f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1887.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 1,887 71.5 74.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eqwvlehf",
   "source_url": "https://omlx.ai/benchmarks/performance/eqwvlehf",
   "tg_tokens": null,
   "tps": 71.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-eqwvlehf-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.0 GB at this context; duplicate traces of the same session: 386bch9o, dv1rn766, n0uuou4f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1734.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 1,734 55.2 79.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eqwvlehf",
   "source_url": "https://omlx.ai/benchmarks/performance/eqwvlehf",
   "tg_tokens": null,
   "tps": 55.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-eqwvlehf-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.9 GB at this context; duplicate traces of the same session: 386bch9o, dv1rn766, n0uuou4f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1274.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 1,274 42.5 80.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eqwvlehf",
   "source_url": "https://omlx.ai/benchmarks/performance/eqwvlehf",
   "tg_tokens": null,
   "tps": 42.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-tqq8w45g-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4043.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 4,043 90.1 26.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tqq8w45g",
   "source_url": "https://omlx.ai/benchmarks/performance/tqq8w45g",
   "tg_tokens": null,
   "tps": 90.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-tqq8w45g-ctx16384",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3756.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 16k 3,756 80.9 26.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tqq8w45g",
   "source_url": "https://omlx.ai/benchmarks/performance/tqq8w45g",
   "tg_tokens": null,
   "tps": 80.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3lulk51r-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 101.8 GB; TTFT 11814 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1387.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "qwen3.8-flash-next-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k tokens 1,387 PP tok/s 136.8 TG tok/s 11814 TTFT (ms) 101.8 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3lulk51r",
   "source_url": "https://omlx.ai/benchmarks/performance/3lulk51r",
   "tg_tokens": null,
   "tps": 136.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-7mcldnpq-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.3 GB at this context; duplicate traces of the same session: 8iwibn0v; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 759.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 759.6 71.1 10.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7mcldnpq",
   "source_url": "https://omlx.ai/benchmarks/performance/7mcldnpq",
   "tg_tokens": null,
   "tps": 71.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-7mcldnpq-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 9.5 GB at this context; duplicate traces of the same session: 8iwibn0v; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 571.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 571.8 48.7 9.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7mcldnpq",
   "source_url": "https://omlx.ai/benchmarks/performance/7mcldnpq",
   "tg_tokens": null,
   "tps": 48.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-7mcldnpq-ctx8192",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.2 GB at this context; duplicate traces of the same session: 8iwibn0v; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 762.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 762.8 43.1 10.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7mcldnpq",
   "source_url": "https://omlx.ai/benchmarks/performance/7mcldnpq",
   "tg_tokens": null,
   "tps": 43.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-7mcldnpq-ctx16384",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 9.6 GB at this context; duplicate traces of the same session: 8iwibn0v; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 652.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 652.6 43.6 9.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7mcldnpq",
   "source_url": "https://omlx.ai/benchmarks/performance/7mcldnpq",
   "tg_tokens": null,
   "tps": 43.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-7mcldnpq-ctx32768",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.1 GB at this context; duplicate traces of the same session: 8iwibn0v; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 606.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 606.1 35.3 10.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7mcldnpq",
   "source_url": "https://omlx.ai/benchmarks/performance/7mcldnpq",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-7mcldnpq-ctx65536",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.1 GB at this context; duplicate traces of the same session: 8iwibn0v; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 572.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k 572.3 29.9 11.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7mcldnpq",
   "source_url": "https://omlx.ai/benchmarks/performance/7mcldnpq",
   "tg_tokens": null,
   "tps": 29.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-7mcldnpq-ctx131072",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 14.1 GB at this context; duplicate traces of the same session: 8iwibn0v; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 458.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 128k 458.7 22.0 14.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7mcldnpq",
   "source_url": "https://omlx.ai/benchmarks/performance/7mcldnpq",
   "tg_tokens": null,
   "tps": 22.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-8kch9w38-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized speed FP16 build; oMLX community benchmark trace, single run at 8k context; peak memory 28.8 GB; TTFT 55469 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 147.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed-FP16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k tokens 147.7 PP tok/s 24.9 TG tok/s 55469 TTFT (ms) 28.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8kch9w38",
   "source_url": "https://omlx.ai/benchmarks/performance/8kch9w38",
   "tg_tokens": null,
   "tps": 24.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-7pbnclj1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 25.3 GB; TTFT 14356 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 285.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 285.3 PP tok/s 42.3 TG tok/s 14356 TTFT (ms) 25.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7pbnclj1",
   "source_url": "https://omlx.ai/benchmarks/performance/7pbnclj1",
   "tg_tokens": null,
   "tps": 42.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-v0wm5emm-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 4yil7z0r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 915.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 915.7 32.0 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v0wm5emm",
   "source_url": "https://omlx.ai/benchmarks/performance/v0wm5emm",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-v0wm5emm-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 4yil7z0r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 996.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 996.4 31.7 17.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v0wm5emm",
   "source_url": "https://omlx.ai/benchmarks/performance/v0wm5emm",
   "tg_tokens": null,
   "tps": 31.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-whw6e4dw-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.5 GB at this context; duplicate traces of the same session: zbnoavok, ztj60gpd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1127.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,127 79.1 70.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/whw6e4dw",
   "source_url": "https://omlx.ai/benchmarks/performance/whw6e4dw",
   "tg_tokens": null,
   "tps": 79.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-whw6e4dw-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: zbnoavok, ztj60gpd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2015.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,015 60.8 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/whw6e4dw",
   "source_url": "https://omlx.ai/benchmarks/performance/whw6e4dw",
   "tg_tokens": null,
   "tps": 60.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-whw6e4dw-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: zbnoavok, ztj60gpd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2132.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,132 92.6 72.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/whw6e4dw",
   "source_url": "https://omlx.ai/benchmarks/performance/whw6e4dw",
   "tg_tokens": null,
   "tps": 92.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-whw6e4dw-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: zbnoavok, ztj60gpd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2035.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,035 65.0 73.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/whw6e4dw",
   "source_url": "https://omlx.ai/benchmarks/performance/whw6e4dw",
   "tg_tokens": null,
   "tps": 65.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-whw6e4dw-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.6 GB at this context; duplicate traces of the same session: zbnoavok, ztj60gpd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1887.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 1,887 71.5 74.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/whw6e4dw",
   "source_url": "https://omlx.ai/benchmarks/performance/whw6e4dw",
   "tg_tokens": null,
   "tps": 71.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-whw6e4dw-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.0 GB at this context; duplicate traces of the same session: zbnoavok, ztj60gpd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1734.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 1,734 55.2 79.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/whw6e4dw",
   "source_url": "https://omlx.ai/benchmarks/performance/whw6e4dw",
   "tg_tokens": null,
   "tps": 55.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-whw6e4dw-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.9 GB at this context; duplicate traces of the same session: zbnoavok, ztj60gpd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1274.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 1,274 42.5 80.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/whw6e4dw",
   "source_url": "https://omlx.ai/benchmarks/performance/whw6e4dw",
   "tg_tokens": null,
   "tps": 42.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-sxch256b-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.5 GB at this context; duplicate traces of the same session: 626jnslk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 897.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 897.9 57.8 70.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sxch256b",
   "source_url": "https://omlx.ai/benchmarks/performance/sxch256b",
   "tg_tokens": null,
   "tps": 57.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-sxch256b-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 626jnslk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1874.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,874 67.8 70.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sxch256b",
   "source_url": "https://omlx.ai/benchmarks/performance/sxch256b",
   "tg_tokens": null,
   "tps": 67.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ri1r2oo9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Youssofal MTPLX Optimized Speed; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 0maw5yd1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 893.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Youssofal--Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 893.7 59.3 23.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ri1r2oo9",
   "source_url": "https://omlx.ai/benchmarks/performance/ri1r2oo9",
   "tg_tokens": null,
   "tps": 59.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ri1r2oo9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "Youssofal MTPLX Optimized Speed; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: 0maw5yd1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1020.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Youssofal--Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,020 49.8 24.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ri1r2oo9",
   "source_url": "https://omlx.ai/benchmarks/performance/ri1r2oo9",
   "tg_tokens": null,
   "tps": 49.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-im90q4qc-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.3 GB at this context; duplicate traces of the same session: g234054d, kh0zfppz, lmfu91no, vrjwpfn8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 759.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 759.6 71.1 10.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/im90q4qc",
   "source_url": "https://omlx.ai/benchmarks/performance/im90q4qc",
   "tg_tokens": null,
   "tps": 71.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-im90q4qc-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 9.5 GB at this context; duplicate traces of the same session: g234054d, kh0zfppz, lmfu91no, vrjwpfn8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 571.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 571.8 48.7 9.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/im90q4qc",
   "source_url": "https://omlx.ai/benchmarks/performance/im90q4qc",
   "tg_tokens": null,
   "tps": 48.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-im90q4qc-ctx8192",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.2 GB at this context; duplicate traces of the same session: g234054d, kh0zfppz, lmfu91no, vrjwpfn8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 762.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 762.8 43.1 10.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/im90q4qc",
   "source_url": "https://omlx.ai/benchmarks/performance/im90q4qc",
   "tg_tokens": null,
   "tps": 43.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-im90q4qc-ctx16384",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 9.6 GB at this context; duplicate traces of the same session: g234054d, kh0zfppz, lmfu91no, vrjwpfn8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 652.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 652.6 43.6 9.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/im90q4qc",
   "source_url": "https://omlx.ai/benchmarks/performance/im90q4qc",
   "tg_tokens": null,
   "tps": 43.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-im90q4qc-ctx32768",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.1 GB at this context; duplicate traces of the same session: g234054d, kh0zfppz, lmfu91no, vrjwpfn8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 606.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 606.1 35.3 10.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/im90q4qc",
   "source_url": "https://omlx.ai/benchmarks/performance/im90q4qc",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-im90q4qc-ctx65536",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.1 GB at this context; duplicate traces of the same session: g234054d, kh0zfppz, lmfu91no, vrjwpfn8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 572.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k 572.3 29.9 11.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/im90q4qc",
   "source_url": "https://omlx.ai/benchmarks/performance/im90q4qc",
   "tg_tokens": null,
   "tps": 29.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-im90q4qc-ctx131072",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 14.1 GB at this context; duplicate traces of the same session: g234054d, kh0zfppz, lmfu91no, vrjwpfn8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 458.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 128k 458.7 22.0 14.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/im90q4qc",
   "source_url": "https://omlx.ai/benchmarks/performance/im90q4qc",
   "tg_tokens": null,
   "tps": 22.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-98s7sd8r-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: mvvpb5o0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 852.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 852.4 37.0 32.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/98s7sd8r",
   "source_url": "https://omlx.ai/benchmarks/performance/98s7sd8r",
   "tg_tokens": null,
   "tps": 37.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-98s7sd8r-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.7 GB at this context; duplicate traces of the same session: mvvpb5o0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 893.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 893.3 35.6 33.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/98s7sd8r",
   "source_url": "https://omlx.ai/benchmarks/performance/98s7sd8r",
   "tg_tokens": null,
   "tps": 35.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8ssmc4ye-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 91urdlzd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 901.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 901.7 32.0 16.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8ssmc4ye",
   "source_url": "https://omlx.ai/benchmarks/performance/8ssmc4ye",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8ssmc4ye-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 91urdlzd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1003.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,003 31.7 17.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8ssmc4ye",
   "source_url": "https://omlx.ai/benchmarks/performance/8ssmc4ye",
   "tg_tokens": null,
   "tps": 31.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ofwv7ue0-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 106.1 15.0 16.6 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ofwv7ue0",
   "source_url": "https://omlx.ai/benchmarks/performance/ofwv7ue0",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ofwv7ue0-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 108.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 108.4 14.5 17.2 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ofwv7ue0",
   "source_url": "https://omlx.ai/benchmarks/performance/ofwv7ue0",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-c51n85ld-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Youssofal MTPLX Optimized Speed; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: mqv9kwln; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 882.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Youssofal--Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 882.4 50.7 23.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c51n85ld",
   "source_url": "https://omlx.ai/benchmarks/performance/c51n85ld",
   "tg_tokens": null,
   "tps": 50.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-c51n85ld-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "Youssofal MTPLX Optimized Speed; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: mqv9kwln; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 997.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Youssofal--Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 997.0 48.0 24.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c51n85ld",
   "source_url": "https://omlx.ai/benchmarks/performance/c51n85ld",
   "tg_tokens": null,
   "tps": 48.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-wgf145c8-ctx4096",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 3.0 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2018.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e2b-it-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 2,018 62.9 3.0 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wgf145c8",
   "source_url": "https://omlx.ai/benchmarks/performance/wgf145c8",
   "tg_tokens": null,
   "tps": 62.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-wgf145c8-ctx16384",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 3.1 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2038.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e2b-it-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 16k 2,038 56.0 3.1 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wgf145c8",
   "source_url": "https://omlx.ai/benchmarks/performance/wgf145c8",
   "tg_tokens": null,
   "tps": 56.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-q4ra18or-ctx4096",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1879.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,879 105.7 20.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q4ra18or",
   "source_url": "https://omlx.ai/benchmarks/performance/q4ra18or",
   "tg_tokens": null,
   "tps": 105.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-q4ra18or-ctx32768",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2900.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 2,900 81.0 21.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q4ra18or",
   "source_url": "https://omlx.ai/benchmarks/performance/q4ra18or",
   "tg_tokens": null,
   "tps": 81.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-gvpvgfoj-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 3ga9rbfi; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1681.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 1,681 68.6 19.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gvpvgfoj",
   "source_url": "https://omlx.ai/benchmarks/performance/gvpvgfoj",
   "tg_tokens": null,
   "tps": 68.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-gvpvgfoj-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 3ga9rbfi; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2127.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 2,127 65.8 19.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gvpvgfoj",
   "source_url": "https://omlx.ai/benchmarks/performance/gvpvgfoj",
   "tg_tokens": null,
   "tps": 65.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-gvpvgfoj-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: 3ga9rbfi; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1883.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 1,883 54.8 20.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gvpvgfoj",
   "source_url": "https://omlx.ai/benchmarks/performance/gvpvgfoj",
   "tg_tokens": null,
   "tps": 54.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-gvpvgfoj-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 3ga9rbfi; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1563.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 32k 1,563 45.6 21.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gvpvgfoj",
   "source_url": "https://omlx.ai/benchmarks/performance/gvpvgfoj",
   "tg_tokens": null,
   "tps": 45.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-kvq3oe3v-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: cu8ctvpk, fujc0w1y, wshulkec; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1695.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 1,695 69.8 19.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kvq3oe3v",
   "source_url": "https://omlx.ai/benchmarks/performance/kvq3oe3v",
   "tg_tokens": null,
   "tps": 69.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-kvq3oe3v-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: cu8ctvpk, fujc0w1y, wshulkec; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2183.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 2,183 68.0 19.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kvq3oe3v",
   "source_url": "https://omlx.ai/benchmarks/performance/kvq3oe3v",
   "tg_tokens": null,
   "tps": 68.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-kvq3oe3v-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: cu8ctvpk, fujc0w1y, wshulkec; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1904.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 1,904 60.0 20.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kvq3oe3v",
   "source_url": "https://omlx.ai/benchmarks/performance/kvq3oe3v",
   "tg_tokens": null,
   "tps": 60.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-kvq3oe3v-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: cu8ctvpk, fujc0w1y, wshulkec; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1572.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 32k 1,572 50.7 21.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kvq3oe3v",
   "source_url": "https://omlx.ai/benchmarks/performance/kvq3oe3v",
   "tg_tokens": null,
   "tps": 50.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-81mev0q2-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: wa0iyehp; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2183.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 2,183 68.3 19.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/81mev0q2",
   "source_url": "https://omlx.ai/benchmarks/performance/81mev0q2",
   "tg_tokens": null,
   "tps": 68.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-81mev0q2-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: wa0iyehp; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1907.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 1,907 60.3 20.5 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/81mev0q2",
   "source_url": "https://omlx.ai/benchmarks/performance/81mev0q2",
   "tg_tokens": null,
   "tps": 60.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hnaewtcu-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 8wv0t1rq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 887.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 887.7 30.5 16.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hnaewtcu",
   "source_url": "https://omlx.ai/benchmarks/performance/hnaewtcu",
   "tg_tokens": null,
   "tps": 30.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hnaewtcu-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 8wv0t1rq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 986.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 986.2 30.6 17.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hnaewtcu",
   "source_url": "https://omlx.ai/benchmarks/performance/hnaewtcu",
   "tg_tokens": null,
   "tps": 30.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-cm3mdqmo-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 106.1 15.0 16.6 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cm3mdqmo",
   "source_url": "https://omlx.ai/benchmarks/performance/cm3mdqmo",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-cm3mdqmo-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 108.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 108.4 14.5 17.2 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cm3mdqmo",
   "source_url": "https://omlx.ai/benchmarks/performance/cm3mdqmo",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-8f0g1ehm-ctx4096",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 3.0 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2018.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e2b-it-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 2,018 62.9 3.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8f0g1ehm",
   "source_url": "https://omlx.ai/benchmarks/performance/8f0g1ehm",
   "tg_tokens": null,
   "tps": 62.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-8f0g1ehm-ctx16384",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 3.1 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2038.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e2b-it-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 16k 2,038 56.0 3.1 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8f0g1ehm",
   "source_url": "https://omlx.ai/benchmarks/performance/8f0g1ehm",
   "tg_tokens": null,
   "tps": 56.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-knd3jqst-ctx4096",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1879.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,879 105.7 20.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/knd3jqst",
   "source_url": "https://omlx.ai/benchmarks/performance/knd3jqst",
   "tg_tokens": null,
   "tps": 105.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-knd3jqst-ctx32768",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2900.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 2,900 81.0 21.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/knd3jqst",
   "source_url": "https://omlx.ai/benchmarks/performance/knd3jqst",
   "tg_tokens": null,
   "tps": 81.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-04uvxp0m-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 81hkdkvm, g8kja6s2, ncj65992; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1674.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 1,674 68.5 19.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/04uvxp0m",
   "source_url": "https://omlx.ai/benchmarks/performance/04uvxp0m",
   "tg_tokens": null,
   "tps": 68.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-04uvxp0m-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 81hkdkvm, g8kja6s2, ncj65992; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2118.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 2,118 63.8 19.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/04uvxp0m",
   "source_url": "https://omlx.ai/benchmarks/performance/04uvxp0m",
   "tg_tokens": null,
   "tps": 63.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-04uvxp0m-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: 81hkdkvm, g8kja6s2, ncj65992; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1908.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 1,908 47.6 20.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/04uvxp0m",
   "source_url": "https://omlx.ai/benchmarks/performance/04uvxp0m",
   "tg_tokens": null,
   "tps": 47.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-04uvxp0m-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 81hkdkvm, g8kja6s2, ncj65992; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1584.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 32k 1,584 35.7 21.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/04uvxp0m",
   "source_url": "https://omlx.ai/benchmarks/performance/04uvxp0m",
   "tg_tokens": null,
   "tps": 35.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-qfbeudq8-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 0p30ub5z; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1681.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 1,681 68.6 19.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qfbeudq8",
   "source_url": "https://omlx.ai/benchmarks/performance/qfbeudq8",
   "tg_tokens": null,
   "tps": 68.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-qfbeudq8-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 0p30ub5z; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2127.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 2,127 65.8 19.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qfbeudq8",
   "source_url": "https://omlx.ai/benchmarks/performance/qfbeudq8",
   "tg_tokens": null,
   "tps": 65.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-qfbeudq8-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: 0p30ub5z; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1883.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 1,883 54.8 20.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qfbeudq8",
   "source_url": "https://omlx.ai/benchmarks/performance/qfbeudq8",
   "tg_tokens": null,
   "tps": 54.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-qfbeudq8-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 0p30ub5z; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1563.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 32k 1,563 45.6 21.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qfbeudq8",
   "source_url": "https://omlx.ai/benchmarks/performance/qfbeudq8",
   "tg_tokens": null,
   "tps": 45.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-oc4brjkn-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 638.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 638.0 35.5 19.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oc4brjkn",
   "source_url": "https://omlx.ai/benchmarks/performance/oc4brjkn",
   "tg_tokens": null,
   "tps": 35.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-oc4brjkn-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 611.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 611.8 71.2 23.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oc4brjkn",
   "source_url": "https://omlx.ai/benchmarks/performance/oc4brjkn",
   "tg_tokens": null,
   "tps": 71.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-oj31mkad-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 77.5 GB at this context; duplicate traces of the same session: 1j6wmg0w, 81qh6e9o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 539.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 539.4 59.9 77.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oj31mkad",
   "source_url": "https://omlx.ai/benchmarks/performance/oj31mkad",
   "tg_tokens": null,
   "tps": 59.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-oj31mkad-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.6 GB at this context; duplicate traces of the same session: 1j6wmg0w, 81qh6e9o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 819.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 819.0 65.8 78.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oj31mkad",
   "source_url": "https://omlx.ai/benchmarks/performance/oj31mkad",
   "tg_tokens": null,
   "tps": 65.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-oj31mkad-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.2 GB at this context; duplicate traces of the same session: 1j6wmg0w, 81qh6e9o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 813.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 813.4 61.9 79.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oj31mkad",
   "source_url": "https://omlx.ai/benchmarks/performance/oj31mkad",
   "tg_tokens": null,
   "tps": 61.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-pmxga5ka-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 345g56ab; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 594.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 594.3 65.5 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pmxga5ka",
   "source_url": "https://omlx.ai/benchmarks/performance/pmxga5ka",
   "tg_tokens": null,
   "tps": 65.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-pmxga5ka-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: 345g56ab; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 726.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 726.0 62.0 71.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pmxga5ka",
   "source_url": "https://omlx.ai/benchmarks/performance/pmxga5ka",
   "tg_tokens": null,
   "tps": 62.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-pmxga5ka-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: 345g56ab; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 836.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 836.3 73.2 71.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pmxga5ka",
   "source_url": "https://omlx.ai/benchmarks/performance/pmxga5ka",
   "tg_tokens": null,
   "tps": 73.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-pmxga5ka-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 345g56ab; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 829.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 829.9 79.0 71.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pmxga5ka",
   "source_url": "https://omlx.ai/benchmarks/performance/pmxga5ka",
   "tg_tokens": null,
   "tps": 79.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-pmxga5ka-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 72.5 GB at this context; duplicate traces of the same session: 345g56ab; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 824.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 824.4 73.2 72.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pmxga5ka",
   "source_url": "https://omlx.ai/benchmarks/performance/pmxga5ka",
   "tg_tokens": null,
   "tps": 73.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-pmxga5ka-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 74.8 GB at this context; duplicate traces of the same session: 345g56ab; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 813.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 813.0 53.3 74.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pmxga5ka",
   "source_url": "https://omlx.ai/benchmarks/performance/pmxga5ka",
   "tg_tokens": null,
   "tps": 53.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-pmxga5ka-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 78.7 GB at this context; duplicate traces of the same session: 345g56ab; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 802.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 802.3 54.5 78.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pmxga5ka",
   "source_url": "https://omlx.ai/benchmarks/performance/pmxga5ka",
   "tg_tokens": null,
   "tps": 54.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-pmxga5ka-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 80.0 GB at this context; duplicate traces of the same session: 345g56ab; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 792.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 792.0 50.1 80.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pmxga5ka",
   "source_url": "https://omlx.ai/benchmarks/performance/pmxga5ka",
   "tg_tokens": null,
   "tps": 50.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-p1cmk834-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: fymo1aky, io2is9j6, nvmcac4s; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1653.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 1,653 69.3 19.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p1cmk834",
   "source_url": "https://omlx.ai/benchmarks/performance/p1cmk834",
   "tg_tokens": null,
   "tps": 69.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-p1cmk834-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: fymo1aky, io2is9j6, nvmcac4s; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2175.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 2,175 67.6 19.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p1cmk834",
   "source_url": "https://omlx.ai/benchmarks/performance/p1cmk834",
   "tg_tokens": null,
   "tps": 67.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-p1cmk834-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: fymo1aky, io2is9j6, nvmcac4s; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1903.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 1,903 60.5 20.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p1cmk834",
   "source_url": "https://omlx.ai/benchmarks/performance/p1cmk834",
   "tg_tokens": null,
   "tps": 60.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-p1cmk834-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: fymo1aky, io2is9j6, nvmcac4s; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1576.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 32k 1,576 50.6 21.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p1cmk834",
   "source_url": "https://omlx.ai/benchmarks/performance/p1cmk834",
   "tg_tokens": null,
   "tps": 50.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-a9tebvxg-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.3 GB at this context; duplicate traces of the same session: aeuzx9rv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 506.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 506.8 18.8 28.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a9tebvxg",
   "source_url": "https://omlx.ai/benchmarks/performance/a9tebvxg",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-a9tebvxg-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.8 GB at this context; duplicate traces of the same session: aeuzx9rv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 609.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 609.3 18.7 29.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a9tebvxg",
   "source_url": "https://omlx.ai/benchmarks/performance/a9tebvxg",
   "tg_tokens": null,
   "tps": 18.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-c88c9f04-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 638.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 638.0 35.5 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c88c9f04",
   "source_url": "https://omlx.ai/benchmarks/performance/c88c9f04",
   "tg_tokens": null,
   "tps": 35.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-c88c9f04-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 611.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 611.8 71.2 23.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c88c9f04",
   "source_url": "https://omlx.ai/benchmarks/performance/c88c9f04",
   "tg_tokens": null,
   "tps": 71.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-eo6tjbqh-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 3pft7l0r, 8czf4ym4, hj7cvdt6, kvauvb2m, sbu3jeml; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 594.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 594.3 65.5 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eo6tjbqh",
   "source_url": "https://omlx.ai/benchmarks/performance/eo6tjbqh",
   "tg_tokens": null,
   "tps": 65.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-eo6tjbqh-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: 3pft7l0r, 8czf4ym4, hj7cvdt6, kvauvb2m, sbu3jeml; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 726.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 726.0 62.0 71.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eo6tjbqh",
   "source_url": "https://omlx.ai/benchmarks/performance/eo6tjbqh",
   "tg_tokens": null,
   "tps": 62.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-eo6tjbqh-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: 3pft7l0r, 8czf4ym4, hj7cvdt6, kvauvb2m, sbu3jeml; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 836.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 836.3 73.2 71.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eo6tjbqh",
   "source_url": "https://omlx.ai/benchmarks/performance/eo6tjbqh",
   "tg_tokens": null,
   "tps": 73.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-eo6tjbqh-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 3pft7l0r, 8czf4ym4, hj7cvdt6, kvauvb2m, sbu3jeml; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 829.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 829.9 79.0 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eo6tjbqh",
   "source_url": "https://omlx.ai/benchmarks/performance/eo6tjbqh",
   "tg_tokens": null,
   "tps": 79.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-eo6tjbqh-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 72.5 GB at this context; duplicate traces of the same session: 3pft7l0r, 8czf4ym4, hj7cvdt6, kvauvb2m, sbu3jeml; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 824.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 824.4 73.2 72.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eo6tjbqh",
   "source_url": "https://omlx.ai/benchmarks/performance/eo6tjbqh",
   "tg_tokens": null,
   "tps": 73.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-eo6tjbqh-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 74.8 GB at this context; duplicate traces of the same session: 3pft7l0r, 8czf4ym4, hj7cvdt6, kvauvb2m, sbu3jeml; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 813.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 813.0 53.3 74.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eo6tjbqh",
   "source_url": "https://omlx.ai/benchmarks/performance/eo6tjbqh",
   "tg_tokens": null,
   "tps": 53.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-eo6tjbqh-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 78.7 GB at this context; duplicate traces of the same session: 3pft7l0r, 8czf4ym4, hj7cvdt6, kvauvb2m, sbu3jeml; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 802.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 802.3 54.5 78.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eo6tjbqh",
   "source_url": "https://omlx.ai/benchmarks/performance/eo6tjbqh",
   "tg_tokens": null,
   "tps": 54.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-eo6tjbqh-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 80.0 GB at this context; duplicate traces of the same session: 3pft7l0r, 8czf4ym4, hj7cvdt6, kvauvb2m, sbu3jeml; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 792.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 792.0 50.1 80.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eo6tjbqh",
   "source_url": "https://omlx.ai/benchmarks/performance/eo6tjbqh",
   "tg_tokens": null,
   "tps": 50.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-bjuqm4ce-ctx1024",
   "model": "Edge0-35B-A3B-preview",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 813 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1260.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Edge0-35B-A3B-preview on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k tokens 1,260 PP tok/s 119.9 TG tok/s 813 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bjuqm4ce",
   "source_url": "https://omlx.ai/benchmarks/performance/bjuqm4ce",
   "tg_tokens": null,
   "tps": 119.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-jm05wsvv-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 211.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 211.5 32.3 18.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jm05wsvv",
   "source_url": "https://omlx.ai/benchmarks/performance/jm05wsvv",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-jm05wsvv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 221.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 221.8 40.6 22.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jm05wsvv",
   "source_url": "https://omlx.ai/benchmarks/performance/jm05wsvv",
   "tg_tokens": null,
   "tps": 40.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-os9gdwsf-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: n06es20z; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3035.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,035 87.6 22.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/os9gdwsf",
   "source_url": "https://omlx.ai/benchmarks/performance/os9gdwsf",
   "tg_tokens": null,
   "tps": 87.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-os9gdwsf-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: n06es20z; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2292.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 2,292 233.1 23.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/os9gdwsf",
   "source_url": "https://omlx.ai/benchmarks/performance/os9gdwsf",
   "tg_tokens": null,
   "tps": 233.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-vy36zmal-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: q3e1wils; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2943.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,943 83.0 22.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vy36zmal",
   "source_url": "https://omlx.ai/benchmarks/performance/vy36zmal",
   "tg_tokens": null,
   "tps": 83.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-vy36zmal-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: q3e1wils; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2148.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 2,148 128.0 23.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vy36zmal",
   "source_url": "https://omlx.ai/benchmarks/performance/vy36zmal",
   "tg_tokens": null,
   "tps": 128.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-v06ns1pi-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 27.6 GB at this context; duplicate traces of the same session: hcud3tjs, iya8pidg, u4j0u2fj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 811.1,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Qwen3.8-27B-mxfp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 811.1 19.0 27.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v06ns1pi",
   "source_url": "https://omlx.ai/benchmarks/performance/v06ns1pi",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-v06ns1pi-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.0 GB at this context; duplicate traces of the same session: hcud3tjs, iya8pidg, u4j0u2fj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 849.8,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Qwen3.8-27B-mxfp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 849.8 18.8 29.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v06ns1pi",
   "source_url": "https://omlx.ai/benchmarks/performance/v06ns1pi",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-v06ns1pi-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: hcud3tjs, iya8pidg, u4j0u2fj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 783.1,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Qwen3.8-27B-mxfp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 783.1 18.4 29.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v06ns1pi",
   "source_url": "https://omlx.ai/benchmarks/performance/v06ns1pi",
   "tg_tokens": null,
   "tps": 18.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-v06ns1pi-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.4 GB at this context; duplicate traces of the same session: hcud3tjs, iya8pidg, u4j0u2fj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 630.7,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Qwen3.8-27B-mxfp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 630.7 17.3 33.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v06ns1pi",
   "source_url": "https://omlx.ai/benchmarks/performance/v06ns1pi",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-14jgdtu4-ctx32768",
   "model": "ThinkingCap-Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 37.0 GB; TTFT 19695 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1664.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "ThinkingCap-Qwen3.8-27B-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 32k tokens 1,664 PP tok/s 101.7 TG tok/s 19695 TTFT (ms) 37.0 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/14jgdtu4",
   "source_url": "https://omlx.ai/benchmarks/performance/14jgdtu4",
   "tg_tokens": null,
   "tps": 101.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-tme3iibg-ctx4096",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 3.0 GB at this context; duplicate traces of the same session: bplz75al; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2343.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e2b-it-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 2,343 63.5 3.0 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tme3iibg",
   "source_url": "https://omlx.ai/benchmarks/performance/tme3iibg",
   "tg_tokens": null,
   "tps": 63.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-tme3iibg-ctx16384",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 3.1 GB at this context; duplicate traces of the same session: bplz75al; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2048.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e2b-it-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 16k 2,048 57.1 3.1 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tme3iibg",
   "source_url": "https://omlx.ai/benchmarks/performance/tme3iibg",
   "tg_tokens": null,
   "tps": 57.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-hendvrvx-ctx4096",
   "model": "Huihui-Qwen3.8-27B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: v1n5t2rt; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 632.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.8-27B-abliterated-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 632.1 30.1 19.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hendvrvx",
   "source_url": "https://omlx.ai/benchmarks/performance/hendvrvx",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-hendvrvx-ctx32768",
   "model": "Huihui-Qwen3.8-27B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: v1n5t2rt; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 600.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.8-27B-abliterated-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 600.4 62.6 23.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hendvrvx",
   "source_url": "https://omlx.ai/benchmarks/performance/hendvrvx",
   "tg_tokens": null,
   "tps": 62.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-f5irhy7s-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Huihui abliterated build; oMLX community benchmark trace; peak memory 99.1 GB at this context; duplicate traces of the same session: 15ojp0p0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1182.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.8-Flash-Next-abliterated-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,182 82.9 99.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f5irhy7s",
   "source_url": "https://omlx.ai/benchmarks/performance/f5irhy7s",
   "tg_tokens": null,
   "tps": 82.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-f5irhy7s-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Huihui abliterated build; oMLX community benchmark trace; peak memory 99.6 GB at this context; duplicate traces of the same session: 15ojp0p0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1468.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.8-Flash-Next-abliterated-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,468 75.1 99.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f5irhy7s",
   "source_url": "https://omlx.ai/benchmarks/performance/f5irhy7s",
   "tg_tokens": null,
   "tps": 75.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-f7pl514z-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Huihui abliterated build; oMLX community benchmark trace, single run at 1k context; peak memory 23.1 GB; TTFT 7377 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 138.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.8-Flash-Next-abliterated-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k tokens 138.8 PP tok/s 9.0 TG tok/s 7377 TTFT (ms) 23.1 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f7pl514z",
   "source_url": "https://omlx.ai/benchmarks/performance/f7pl514z",
   "tg_tokens": null,
   "tps": 9.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-g1egzhn9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 211.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 211.5 32.3 18.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g1egzhn9",
   "source_url": "https://omlx.ai/benchmarks/performance/g1egzhn9",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-g1egzhn9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 221.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 221.8 40.6 22.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g1egzhn9",
   "source_url": "https://omlx.ai/benchmarks/performance/g1egzhn9",
   "tg_tokens": null,
   "tps": 40.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1o43utq0-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5746.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 5,746 191.2 22.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1o43utq0",
   "source_url": "https://omlx.ai/benchmarks/performance/1o43utq0",
   "tg_tokens": null,
   "tps": 191.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1o43utq0-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5374.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 5,374 204.1 22.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1o43utq0",
   "source_url": "https://omlx.ai/benchmarks/performance/1o43utq0",
   "tg_tokens": null,
   "tps": 204.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1o43utq0-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2398.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 2,398 110.1 24.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1o43utq0",
   "source_url": "https://omlx.ai/benchmarks/performance/1o43utq0",
   "tg_tokens": null,
   "tps": 110.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9twz5yek-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: 2jwrw49g, yehmfwze; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 889.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 889.0 57.8 27.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9twz5yek",
   "source_url": "https://omlx.ai/benchmarks/performance/9twz5yek",
   "tg_tokens": null,
   "tps": 57.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9twz5yek-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: 2jwrw49g, yehmfwze; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 909.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 8k 909.9 66.6 28.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9twz5yek",
   "source_url": "https://omlx.ai/benchmarks/performance/9twz5yek",
   "tg_tokens": null,
   "tps": 66.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9twz5yek-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: 2jwrw49g, yehmfwze; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 878.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k 878.5 81.1 28.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9twz5yek",
   "source_url": "https://omlx.ai/benchmarks/performance/9twz5yek",
   "tg_tokens": null,
   "tps": 81.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9twz5yek-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: 2jwrw49g, yehmfwze; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 824.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 824.8 63.2 29.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9twz5yek",
   "source_url": "https://omlx.ai/benchmarks/performance/9twz5yek",
   "tg_tokens": null,
   "tps": 63.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9twz5yek-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: 2jwrw49g, yehmfwze; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 732.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 64k 732.6 48.7 32.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9twz5yek",
   "source_url": "https://omlx.ai/benchmarks/performance/9twz5yek",
   "tg_tokens": null,
   "tps": 48.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ghfzv6eu-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MLX; oMLX community benchmark trace; peak memory 75.3 GB at this context; duplicate traces of the same session: i683twpl, xymhimt3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 882.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 882.8 70.4 75.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ghfzv6eu",
   "source_url": "https://omlx.ai/benchmarks/performance/ghfzv6eu",
   "tg_tokens": null,
   "tps": 70.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ghfzv6eu-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MLX; oMLX community benchmark trace; peak memory 75.5 GB at this context; duplicate traces of the same session: i683twpl, xymhimt3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2071.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,071 53.9 75.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ghfzv6eu",
   "source_url": "https://omlx.ai/benchmarks/performance/ghfzv6eu",
   "tg_tokens": null,
   "tps": 53.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ghfzv6eu-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MLX; oMLX community benchmark trace; peak memory 78.6 GB at this context; duplicate traces of the same session: i683twpl, xymhimt3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2320.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,320 32.6 78.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ghfzv6eu",
   "source_url": "https://omlx.ai/benchmarks/performance/ghfzv6eu",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vwd9idsp-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: v5yh6on2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5746.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 5,746 191.2 22.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vwd9idsp",
   "source_url": "https://omlx.ai/benchmarks/performance/vwd9idsp",
   "tg_tokens": null,
   "tps": 191.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-vwd9idsp-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: v5yh6on2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5374.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 5,374 204.1 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vwd9idsp",
   "source_url": "https://omlx.ai/benchmarks/performance/vwd9idsp",
   "tg_tokens": null,
   "tps": 204.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-vwd9idsp-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: v5yh6on2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2398.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 2,398 110.1 24.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vwd9idsp",
   "source_url": "https://omlx.ai/benchmarks/performance/vwd9idsp",
   "tg_tokens": null,
   "tps": 110.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-90dcop97-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 18.8 GB; TTFT 4294 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 953.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 953.9 PP tok/s 46.5 TG tok/s 4294 TTFT (ms) 18.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/90dcop97",
   "source_url": "https://omlx.ai/benchmarks/performance/90dcop97",
   "tg_tokens": null,
   "tps": 46.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8v6vkjkt-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: tc84aipc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 889.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 889.0 57.8 27.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8v6vkjkt",
   "source_url": "https://omlx.ai/benchmarks/performance/8v6vkjkt",
   "tg_tokens": null,
   "tps": 57.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8v6vkjkt-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: tc84aipc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 909.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 8k 909.9 66.6 28.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8v6vkjkt",
   "source_url": "https://omlx.ai/benchmarks/performance/8v6vkjkt",
   "tg_tokens": null,
   "tps": 66.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8v6vkjkt-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: tc84aipc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 878.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k 878.5 81.1 28.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8v6vkjkt",
   "source_url": "https://omlx.ai/benchmarks/performance/8v6vkjkt",
   "tg_tokens": null,
   "tps": 81.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8v6vkjkt-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: tc84aipc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 824.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 824.8 63.2 29.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8v6vkjkt",
   "source_url": "https://omlx.ai/benchmarks/performance/8v6vkjkt",
   "tg_tokens": null,
   "tps": 63.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8v6vkjkt-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: tc84aipc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 732.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 64k 732.6 48.7 32.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8v6vkjkt",
   "source_url": "https://omlx.ai/benchmarks/performance/8v6vkjkt",
   "tg_tokens": null,
   "tps": 48.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-v80e0m48-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: 262jo6ho, 7dwa29zu, l2zkbva2, lfl7de5p, ou5pbl44, ru4ng0tl, ystfdtry; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 695.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 695.9 48.4 33.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v80e0m48",
   "source_url": "https://omlx.ai/benchmarks/performance/v80e0m48",
   "tg_tokens": null,
   "tps": 48.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-v80e0m48-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: 262jo6ho, 7dwa29zu, l2zkbva2, lfl7de5p, ou5pbl44, ru4ng0tl, ystfdtry; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 739.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 739.4 45.9 34.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v80e0m48",
   "source_url": "https://omlx.ai/benchmarks/performance/v80e0m48",
   "tg_tokens": null,
   "tps": 45.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-v80e0m48-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: 262jo6ho, 7dwa29zu, l2zkbva2, lfl7de5p, ou5pbl44, ru4ng0tl, ystfdtry; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 757.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 757.0 51.0 35.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v80e0m48",
   "source_url": "https://omlx.ai/benchmarks/performance/v80e0m48",
   "tg_tokens": null,
   "tps": 51.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-v80e0m48-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: 262jo6ho, 7dwa29zu, l2zkbva2, lfl7de5p, ou5pbl44, ru4ng0tl, ystfdtry; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 774.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 774.8 53.0 35.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v80e0m48",
   "source_url": "https://omlx.ai/benchmarks/performance/v80e0m48",
   "tg_tokens": null,
   "tps": 53.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-v80e0m48-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: 262jo6ho, 7dwa29zu, l2zkbva2, lfl7de5p, ou5pbl44, ru4ng0tl, ystfdtry; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 715.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 715.0 58.7 36.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v80e0m48",
   "source_url": "https://omlx.ai/benchmarks/performance/v80e0m48",
   "tg_tokens": null,
   "tps": 58.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-v80e0m48-ctx65536",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: 262jo6ho, 7dwa29zu, l2zkbva2, lfl7de5p, ou5pbl44, ru4ng0tl, ystfdtry; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 630.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 630.8 42.9 39.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v80e0m48",
   "source_url": "https://omlx.ai/benchmarks/performance/v80e0m48",
   "tg_tokens": null,
   "tps": 42.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-v80e0m48-ctx131072",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 44.0 GB at this context; duplicate traces of the same session: 262jo6ho, 7dwa29zu, l2zkbva2, lfl7de5p, ou5pbl44, ru4ng0tl, ystfdtry; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 482.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 482.9 40.5 44.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v80e0m48",
   "source_url": "https://omlx.ai/benchmarks/performance/v80e0m48",
   "tg_tokens": null,
   "tps": 40.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-v80e0m48-ctx199680",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 49.2 GB at this context; duplicate traces of the same session: 262jo6ho, 7dwa29zu, l2zkbva2, lfl7de5p, ou5pbl44, ru4ng0tl, ystfdtry; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 383.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 195k 383.7 25.8 49.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v80e0m48",
   "source_url": "https://omlx.ai/benchmarks/performance/v80e0m48",
   "tg_tokens": null,
   "tps": 25.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-xl86v80j-ctx1024",
   "model": "Edge0-35B-A3B-preview",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 86oq55n3, jw5qmcpi; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1546.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Edge0-35B-A3B-preview on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 1,546 119.8 19.4 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xl86v80j",
   "source_url": "https://omlx.ai/benchmarks/performance/xl86v80j",
   "tg_tokens": null,
   "tps": 119.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-xl86v80j-ctx8192",
   "model": "Edge0-35B-A3B-preview",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 86oq55n3, jw5qmcpi; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1565.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Edge0-35B-A3B-preview on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 1,565 106.1 20.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xl86v80j",
   "source_url": "https://omlx.ai/benchmarks/performance/xl86v80j",
   "tg_tokens": null,
   "tps": 106.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-xl86v80j-ctx16384",
   "model": "Edge0-35B-A3B-preview",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: 86oq55n3, jw5qmcpi; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1337.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Edge0-35B-A3B-preview on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 1,337 90.8 20.4 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xl86v80j",
   "source_url": "https://omlx.ai/benchmarks/performance/xl86v80j",
   "tg_tokens": null,
   "tps": 90.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-q9oojkd5-ctx1024",
   "model": "Qwen2.5-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 4.8 GB at this context; duplicate traces of the same session: 5evewkh5, lhbvbigb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 479.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-7B-Instruct-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 479.5 53.4 4.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q9oojkd5",
   "source_url": "https://omlx.ai/benchmarks/performance/q9oojkd5",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-q9oojkd5-ctx4096",
   "model": "Qwen2.5-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: 5evewkh5, lhbvbigb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 459.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-7B-Instruct-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 459.1 50.4 4.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q9oojkd5",
   "source_url": "https://omlx.ai/benchmarks/performance/q9oojkd5",
   "tg_tokens": null,
   "tps": 50.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-q9oojkd5-ctx8192",
   "model": "Qwen2.5-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 5.1 GB at this context; duplicate traces of the same session: 5evewkh5, lhbvbigb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 378.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-7B-Instruct-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 378.5 44.1 5.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q9oojkd5",
   "source_url": "https://omlx.ai/benchmarks/performance/q9oojkd5",
   "tg_tokens": null,
   "tps": 44.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-4sy3gc5n-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: aa0k54hw; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 510.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 510.5 17.0 17.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4sy3gc5n",
   "source_url": "https://omlx.ai/benchmarks/performance/4sy3gc5n",
   "tg_tokens": null,
   "tps": 17.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-4sy3gc5n-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: aa0k54hw; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 444.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 444.5 16.0 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4sy3gc5n",
   "source_url": "https://omlx.ai/benchmarks/performance/4sy3gc5n",
   "tg_tokens": null,
   "tps": 16.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dmzz7asv-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 1mwg4fsp, 22dt6uh4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 830.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 830.7 18.2 28.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dmzz7asv",
   "source_url": "https://omlx.ai/benchmarks/performance/dmzz7asv",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dmzz7asv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: 1mwg4fsp, 22dt6uh4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 913.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 913.3 18.1 29.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dmzz7asv",
   "source_url": "https://omlx.ai/benchmarks/performance/dmzz7asv",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dmzz7asv-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.8 GB at this context; duplicate traces of the same session: 1mwg4fsp, 22dt6uh4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 854.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 854.2 18.4 29.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dmzz7asv",
   "source_url": "https://omlx.ai/benchmarks/performance/dmzz7asv",
   "tg_tokens": null,
   "tps": 18.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dmzz7asv-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.3 GB at this context; duplicate traces of the same session: 1mwg4fsp, 22dt6uh4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 718.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 718.5 17.3 31.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dmzz7asv",
   "source_url": "https://omlx.ai/benchmarks/performance/dmzz7asv",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dmzz7asv-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.3 GB at this context; duplicate traces of the same session: 1mwg4fsp, 22dt6uh4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 619.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 619.9 16.0 33.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dmzz7asv",
   "source_url": "https://omlx.ai/benchmarks/performance/dmzz7asv",
   "tg_tokens": null,
   "tps": 16.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ow00kug7-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.6 GB at this context; duplicate traces of the same session: z76ykesk; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1015.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 1,015 99.4 142.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ow00kug7",
   "source_url": "https://omlx.ai/benchmarks/performance/ow00kug7",
   "tg_tokens": null,
   "tps": 99.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ow00kug7-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: z76ykesk; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1355.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 1,355 94.0 143.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ow00kug7",
   "source_url": "https://omlx.ai/benchmarks/performance/ow00kug7",
   "tg_tokens": null,
   "tps": 94.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ow00kug7-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.7 GB at this context; duplicate traces of the same session: z76ykesk; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1335.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 16k 1,335 33.4 143.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ow00kug7",
   "source_url": "https://omlx.ai/benchmarks/performance/ow00kug7",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ow00kug7-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 144.3 GB at this context; duplicate traces of the same session: z76ykesk; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1327.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 32k 1,327 81.2 144.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ow00kug7",
   "source_url": "https://omlx.ai/benchmarks/performance/ow00kug7",
   "tg_tokens": null,
   "tps": 81.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ow00kug7-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 146.4 GB at this context; duplicate traces of the same session: z76ykesk; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1311.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 64k 1,311 65.9 146.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ow00kug7",
   "source_url": "https://omlx.ai/benchmarks/performance/ow00kug7",
   "tg_tokens": null,
   "tps": 65.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ow00kug7-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 150.6 GB at this context; duplicate traces of the same session: z76ykesk; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1298.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 128k 1,298 55.0 150.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ow00kug7",
   "source_url": "https://omlx.ai/benchmarks/performance/ow00kug7",
   "tg_tokens": null,
   "tps": 55.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-083cosbn-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: fxcjz1xf, sqqz1mb1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2499.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,499 97.5 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/083cosbn",
   "source_url": "https://omlx.ai/benchmarks/performance/083cosbn",
   "tg_tokens": null,
   "tps": 97.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-083cosbn-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: fxcjz1xf, sqqz1mb1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3271.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,271 92.3 22.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/083cosbn",
   "source_url": "https://omlx.ai/benchmarks/performance/083cosbn",
   "tg_tokens": null,
   "tps": 92.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-083cosbn-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: fxcjz1xf, sqqz1mb1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3324.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 3,324 124.4 22.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/083cosbn",
   "source_url": "https://omlx.ai/benchmarks/performance/083cosbn",
   "tg_tokens": null,
   "tps": 124.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-083cosbn-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: fxcjz1xf, sqqz1mb1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2975.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,975 101.6 22.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/083cosbn",
   "source_url": "https://omlx.ai/benchmarks/performance/083cosbn",
   "tg_tokens": null,
   "tps": 101.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-083cosbn-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: fxcjz1xf, sqqz1mb1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2566.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 2,566 129.2 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/083cosbn",
   "source_url": "https://omlx.ai/benchmarks/performance/083cosbn",
   "tg_tokens": null,
   "tps": 129.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wr2z13to-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: duw8682f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 830.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 830.7 18.2 28.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wr2z13to",
   "source_url": "https://omlx.ai/benchmarks/performance/wr2z13to",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wr2z13to-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: duw8682f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 913.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 913.3 18.1 29.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wr2z13to",
   "source_url": "https://omlx.ai/benchmarks/performance/wr2z13to",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wr2z13to-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.8 GB at this context; duplicate traces of the same session: duw8682f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 854.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 854.2 18.4 29.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wr2z13to",
   "source_url": "https://omlx.ai/benchmarks/performance/wr2z13to",
   "tg_tokens": null,
   "tps": 18.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wr2z13to-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.3 GB at this context; duplicate traces of the same session: duw8682f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 718.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 718.5 17.3 31.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wr2z13to",
   "source_url": "https://omlx.ai/benchmarks/performance/wr2z13to",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wr2z13to-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.3 GB at this context; duplicate traces of the same session: duw8682f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 619.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 619.9 16.0 33.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wr2z13to",
   "source_url": "https://omlx.ai/benchmarks/performance/wr2z13to",
   "tg_tokens": null,
   "tps": 16.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rmejmuge-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 833.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 833.9 18.5 28.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rmejmuge",
   "source_url": "https://omlx.ai/benchmarks/performance/rmejmuge",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rmejmuge-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 936.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 936.7 18.6 29.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rmejmuge",
   "source_url": "https://omlx.ai/benchmarks/performance/rmejmuge",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rmejmuge-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 29.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 871.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 871.7 18.4 29.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rmejmuge",
   "source_url": "https://omlx.ai/benchmarks/performance/rmejmuge",
   "tg_tokens": null,
   "tps": 18.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rmejmuge-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 31.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 698.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 698.2 17.3 31.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rmejmuge",
   "source_url": "https://omlx.ai/benchmarks/performance/rmejmuge",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rmejmuge-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 33.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 615.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 615.2 16.0 33.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rmejmuge",
   "source_url": "https://omlx.ai/benchmarks/performance/rmejmuge",
   "tg_tokens": null,
   "tps": 16.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-9z9gb5js-ctx4096",
   "model": "Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-uncher",
   "notes": "bf16 last4 8bit vision MTP build; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: vohma9km; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 148.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-uncher-oQ5e-6target-bf16-last4_8bit-vision-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 5bit; 4k 148.7 18.4 25.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9z9gb5js",
   "source_url": "https://omlx.ai/benchmarks/performance/9z9gb5js",
   "tg_tokens": null,
   "tps": 18.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-9z9gb5js-ctx16384",
   "model": "Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-uncher",
   "notes": "bf16 last4 8bit vision MTP build; oMLX community benchmark trace; peak memory 27.2 GB at this context; duplicate traces of the same session: vohma9km; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 135.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-uncher-oQ5e-6target-bf16-last4_8bit-vision-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 5bit; 16k 135.9 16.2 27.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9z9gb5js",
   "source_url": "https://omlx.ai/benchmarks/performance/9z9gb5js",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-4l64j8f8-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 102.1 GB; TTFT 24238 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1352.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k tokens 1,352 PP tok/s 70.8 TG tok/s 24238 TTFT (ms) 102.1 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4l64j8f8",
   "source_url": "https://omlx.ai/benchmarks/performance/4l64j8f8",
   "tg_tokens": null,
   "tps": 70.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2ftrlvow-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.6 GB at this context; duplicate traces of the same session: kr6e891h, o3ahmsc2, oftg9ozb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1015.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 1,015 99.4 142.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ftrlvow",
   "source_url": "https://omlx.ai/benchmarks/performance/2ftrlvow",
   "tg_tokens": null,
   "tps": 99.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2ftrlvow-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: kr6e891h, o3ahmsc2, oftg9ozb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1355.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 1,355 94.0 143.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ftrlvow",
   "source_url": "https://omlx.ai/benchmarks/performance/2ftrlvow",
   "tg_tokens": null,
   "tps": 94.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2ftrlvow-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.7 GB at this context; duplicate traces of the same session: kr6e891h, o3ahmsc2, oftg9ozb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1335.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 16k 1,335 33.4 143.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ftrlvow",
   "source_url": "https://omlx.ai/benchmarks/performance/2ftrlvow",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2ftrlvow-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 144.3 GB at this context; duplicate traces of the same session: kr6e891h, o3ahmsc2, oftg9ozb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1327.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 32k 1,327 81.2 144.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ftrlvow",
   "source_url": "https://omlx.ai/benchmarks/performance/2ftrlvow",
   "tg_tokens": null,
   "tps": 81.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2ftrlvow-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 146.4 GB at this context; duplicate traces of the same session: kr6e891h, o3ahmsc2, oftg9ozb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1311.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 64k 1,311 65.9 146.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ftrlvow",
   "source_url": "https://omlx.ai/benchmarks/performance/2ftrlvow",
   "tg_tokens": null,
   "tps": 65.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2ftrlvow-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 150.6 GB at this context; duplicate traces of the same session: kr6e891h, o3ahmsc2, oftg9ozb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1298.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 128k 1,298 55.0 150.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ftrlvow",
   "source_url": "https://omlx.ai/benchmarks/performance/2ftrlvow",
   "tg_tokens": null,
   "tps": 55.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-f9oohody-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3391.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 3,391 146.8 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f9oohody",
   "source_url": "https://omlx.ai/benchmarks/performance/f9oohody",
   "tg_tokens": null,
   "tps": 146.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-f9oohody-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2605.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 2,605 174.8 23.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f9oohody",
   "source_url": "https://omlx.ai/benchmarks/performance/f9oohody",
   "tg_tokens": null,
   "tps": 174.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-7c8h0kqv-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 1xmu29kk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2499.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,499 97.5 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7c8h0kqv",
   "source_url": "https://omlx.ai/benchmarks/performance/7c8h0kqv",
   "tg_tokens": null,
   "tps": 97.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-7c8h0kqv-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 1xmu29kk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3271.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,271 92.3 22.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7c8h0kqv",
   "source_url": "https://omlx.ai/benchmarks/performance/7c8h0kqv",
   "tg_tokens": null,
   "tps": 92.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-7c8h0kqv-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 1xmu29kk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3324.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 3,324 124.4 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7c8h0kqv",
   "source_url": "https://omlx.ai/benchmarks/performance/7c8h0kqv",
   "tg_tokens": null,
   "tps": 124.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-7c8h0kqv-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: 1xmu29kk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2975.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,975 101.6 22.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7c8h0kqv",
   "source_url": "https://omlx.ai/benchmarks/performance/7c8h0kqv",
   "tg_tokens": null,
   "tps": 101.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-7c8h0kqv-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 1xmu29kk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2566.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 2,566 129.2 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7c8h0kqv",
   "source_url": "https://omlx.ai/benchmarks/performance/7c8h0kqv",
   "tg_tokens": null,
   "tps": 129.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-mfys63gy-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 4d3dcznr; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3360.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 3,360 119.1 22.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mfys63gy",
   "source_url": "https://omlx.ai/benchmarks/performance/mfys63gy",
   "tg_tokens": null,
   "tps": 119.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-mfys63gy-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 4d3dcznr; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2553.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 2,553 133.2 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mfys63gy",
   "source_url": "https://omlx.ai/benchmarks/performance/mfys63gy",
   "tg_tokens": null,
   "tps": 133.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-1nmqexxe-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 16aysild; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2607.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,607 108.2 21.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1nmqexxe",
   "source_url": "https://omlx.ai/benchmarks/performance/1nmqexxe",
   "tg_tokens": null,
   "tps": 108.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1nmqexxe-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 16aysild; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2619.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 2,619 139.4 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1nmqexxe",
   "source_url": "https://omlx.ai/benchmarks/performance/1nmqexxe",
   "tg_tokens": null,
   "tps": 139.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-4qsvpiu7-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2570.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,570 96.7 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4qsvpiu7",
   "source_url": "https://omlx.ai/benchmarks/performance/4qsvpiu7",
   "tg_tokens": null,
   "tps": 96.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-4qsvpiu7-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3441.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,441 96.7 22.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4qsvpiu7",
   "source_url": "https://omlx.ai/benchmarks/performance/4qsvpiu7",
   "tg_tokens": null,
   "tps": 96.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-4qsvpiu7-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3399.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 3,399 124.6 22.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4qsvpiu7",
   "source_url": "https://omlx.ai/benchmarks/performance/4qsvpiu7",
   "tg_tokens": null,
   "tps": 124.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bsu6iarz-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: m5yux4ms, qw1mmh5e, v306w4ee; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 833.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 833.9 18.5 28.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bsu6iarz",
   "source_url": "https://omlx.ai/benchmarks/performance/bsu6iarz",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bsu6iarz-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: m5yux4ms, qw1mmh5e, v306w4ee; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 936.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 936.7 18.6 29.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bsu6iarz",
   "source_url": "https://omlx.ai/benchmarks/performance/bsu6iarz",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bsu6iarz-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 29.8 GB at this context; duplicate traces of the same session: m5yux4ms, qw1mmh5e, v306w4ee; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 871.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 871.7 18.4 29.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bsu6iarz",
   "source_url": "https://omlx.ai/benchmarks/performance/bsu6iarz",
   "tg_tokens": null,
   "tps": 18.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bsu6iarz-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 31.3 GB at this context; duplicate traces of the same session: m5yux4ms, qw1mmh5e, v306w4ee; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 698.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 698.2 17.3 31.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bsu6iarz",
   "source_url": "https://omlx.ai/benchmarks/performance/bsu6iarz",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bsu6iarz-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 33.3 GB at this context; duplicate traces of the same session: m5yux4ms, qw1mmh5e, v306w4ee; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 615.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 615.2 16.0 33.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bsu6iarz",
   "source_url": "https://omlx.ai/benchmarks/performance/bsu6iarz",
   "tg_tokens": null,
   "tps": 16.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2vbkjsb7-ctx8192",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 69.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 921.8,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 921.8 20.6 69.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2vbkjsb7",
   "source_url": "https://omlx.ai/benchmarks/performance/2vbkjsb7",
   "tg_tokens": null,
   "tps": 20.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2vbkjsb7-ctx16384",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 69.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1206.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 1,206 23.6 69.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2vbkjsb7",
   "source_url": "https://omlx.ai/benchmarks/performance/2vbkjsb7",
   "tg_tokens": null,
   "tps": 23.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-w06pagsg-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3391.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 3,391 146.8 22.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w06pagsg",
   "source_url": "https://omlx.ai/benchmarks/performance/w06pagsg",
   "tg_tokens": null,
   "tps": 146.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-w06pagsg-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2605.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 2,605 174.8 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w06pagsg",
   "source_url": "https://omlx.ai/benchmarks/performance/w06pagsg",
   "tg_tokens": null,
   "tps": 174.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ipk7wbzy-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 2tm5oh1m, 5bovxosk, mrpxjfeh, ptwdale1, vbeefkyp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2410.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,410 101.2 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ipk7wbzy",
   "source_url": "https://omlx.ai/benchmarks/performance/ipk7wbzy",
   "tg_tokens": null,
   "tps": 101.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ipk7wbzy-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 2tm5oh1m, 5bovxosk, mrpxjfeh, ptwdale1, vbeefkyp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3167.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,167 102.8 22.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ipk7wbzy",
   "source_url": "https://omlx.ai/benchmarks/performance/ipk7wbzy",
   "tg_tokens": null,
   "tps": 102.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ipk7wbzy-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 2tm5oh1m, 5bovxosk, mrpxjfeh, ptwdale1, vbeefkyp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3064.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 3,064 100.2 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ipk7wbzy",
   "source_url": "https://omlx.ai/benchmarks/performance/ipk7wbzy",
   "tg_tokens": null,
   "tps": 100.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ipk7wbzy-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 2tm5oh1m, 5bovxosk, mrpxjfeh, ptwdale1, vbeefkyp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2812.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,812 112.4 22.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ipk7wbzy",
   "source_url": "https://omlx.ai/benchmarks/performance/ipk7wbzy",
   "tg_tokens": null,
   "tps": 112.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ipk7wbzy-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 2tm5oh1m, 5bovxosk, mrpxjfeh, ptwdale1, vbeefkyp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2374.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 2,374 149.2 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ipk7wbzy",
   "source_url": "https://omlx.ai/benchmarks/performance/ipk7wbzy",
   "tg_tokens": null,
   "tps": 149.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ipk7wbzy-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: 2tm5oh1m, 5bovxosk, mrpxjfeh, ptwdale1, vbeefkyp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1830.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 1,830 67.9 25.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ipk7wbzy",
   "source_url": "https://omlx.ai/benchmarks/performance/ipk7wbzy",
   "tg_tokens": null,
   "tps": 67.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-i2xxpx6q-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 9s59a7rd; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2570.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,570 96.7 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i2xxpx6q",
   "source_url": "https://omlx.ai/benchmarks/performance/i2xxpx6q",
   "tg_tokens": null,
   "tps": 96.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-i2xxpx6q-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 9s59a7rd; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3441.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,441 96.7 22.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i2xxpx6q",
   "source_url": "https://omlx.ai/benchmarks/performance/i2xxpx6q",
   "tg_tokens": null,
   "tps": 96.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-i2xxpx6q-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 9s59a7rd; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3399.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 3,399 124.6 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i2xxpx6q",
   "source_url": "https://omlx.ai/benchmarks/performance/i2xxpx6q",
   "tg_tokens": null,
   "tps": 124.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-uhmeaifl-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: nizfo3ye, xq12kfzh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2245.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,245 90.3 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uhmeaifl",
   "source_url": "https://omlx.ai/benchmarks/performance/uhmeaifl",
   "tg_tokens": null,
   "tps": 90.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-uhmeaifl-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: nizfo3ye, xq12kfzh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3072.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,072 89.2 22.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uhmeaifl",
   "source_url": "https://omlx.ai/benchmarks/performance/uhmeaifl",
   "tg_tokens": null,
   "tps": 89.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-uhmeaifl-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: nizfo3ye, xq12kfzh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2985.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 2,985 84.7 22.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uhmeaifl",
   "source_url": "https://omlx.ai/benchmarks/performance/uhmeaifl",
   "tg_tokens": null,
   "tps": 84.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-uhmeaifl-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: nizfo3ye, xq12kfzh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2775.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,775 126.0 22.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uhmeaifl",
   "source_url": "https://omlx.ai/benchmarks/performance/uhmeaifl",
   "tg_tokens": null,
   "tps": 126.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-uhmeaifl-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: nizfo3ye, xq12kfzh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2328.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 2,328 145.7 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uhmeaifl",
   "source_url": "https://omlx.ai/benchmarks/performance/uhmeaifl",
   "tg_tokens": null,
   "tps": 145.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pg6emlp3-ctx8192",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 69.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 921.8,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 921.8 20.6 69.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pg6emlp3",
   "source_url": "https://omlx.ai/benchmarks/performance/pg6emlp3",
   "tg_tokens": null,
   "tps": 20.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pg6emlp3-ctx16384",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 69.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1206.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 1,206 23.6 69.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pg6emlp3",
   "source_url": "https://omlx.ai/benchmarks/performance/pg6emlp3",
   "tg_tokens": null,
   "tps": 23.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-lu75ih3a-ctx1024",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2560.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,560 116.8 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lu75ih3a",
   "source_url": "https://omlx.ai/benchmarks/performance/lu75ih3a",
   "tg_tokens": null,
   "tps": 116.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-lu75ih3a-ctx4096",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3302.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,302 102.5 22.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lu75ih3a",
   "source_url": "https://omlx.ai/benchmarks/performance/lu75ih3a",
   "tg_tokens": null,
   "tps": 102.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-lu75ih3a-ctx8192",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3205.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 3,205 127.1 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lu75ih3a",
   "source_url": "https://omlx.ai/benchmarks/performance/lu75ih3a",
   "tg_tokens": null,
   "tps": 127.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-lu75ih3a-ctx16384",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2930.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,930 142.5 22.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lu75ih3a",
   "source_url": "https://omlx.ai/benchmarks/performance/lu75ih3a",
   "tg_tokens": null,
   "tps": 142.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-lu75ih3a-ctx32768",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2409.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 2,409 157.9 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lu75ih3a",
   "source_url": "https://omlx.ai/benchmarks/performance/lu75ih3a",
   "tg_tokens": null,
   "tps": 157.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0kcbsa0z-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 08ajn620; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2245.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,245 90.3 21.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0kcbsa0z",
   "source_url": "https://omlx.ai/benchmarks/performance/0kcbsa0z",
   "tg_tokens": null,
   "tps": 90.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0kcbsa0z-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 08ajn620; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3072.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,072 89.2 22.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0kcbsa0z",
   "source_url": "https://omlx.ai/benchmarks/performance/0kcbsa0z",
   "tg_tokens": null,
   "tps": 89.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0kcbsa0z-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 08ajn620; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2985.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 2,985 84.7 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0kcbsa0z",
   "source_url": "https://omlx.ai/benchmarks/performance/0kcbsa0z",
   "tg_tokens": null,
   "tps": 84.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-0kcbsa0z-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 08ajn620; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2775.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,775 126.0 22.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0kcbsa0z",
   "source_url": "https://omlx.ai/benchmarks/performance/0kcbsa0z",
   "tg_tokens": null,
   "tps": 126.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0kcbsa0z-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 08ajn620; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2328.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 2,328 145.7 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0kcbsa0z",
   "source_url": "https://omlx.ai/benchmarks/performance/0kcbsa0z",
   "tg_tokens": null,
   "tps": 145.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-q8gkcihg-ctx1024",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 12.8 GB at this context; duplicate traces of the same session: 5z4yx1cz, lc6knu3b; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1561.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 1,561 91.6 12.8 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q8gkcihg",
   "source_url": "https://omlx.ai/benchmarks/performance/q8gkcihg",
   "tg_tokens": null,
   "tps": 91.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-q8gkcihg-ctx8192",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.5 GB at this context; duplicate traces of the same session: 5z4yx1cz, lc6knu3b; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1581.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 1,581 86.4 13.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q8gkcihg",
   "source_url": "https://omlx.ai/benchmarks/performance/q8gkcihg",
   "tg_tokens": null,
   "tps": 86.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-q8gkcihg-ctx16384",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: 5z4yx1cz, lc6knu3b; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1375.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 1,375 81.1 13.7 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q8gkcihg",
   "source_url": "https://omlx.ai/benchmarks/performance/q8gkcihg",
   "tg_tokens": null,
   "tps": 81.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-gbb6sv35-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: 0e9iqmdm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 256.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 256.3 43.8 19.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gbb6sv35",
   "source_url": "https://omlx.ai/benchmarks/performance/gbb6sv35",
   "tg_tokens": null,
   "tps": 43.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-gbb6sv35-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 0e9iqmdm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 259.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 259.2 40.5 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gbb6sv35",
   "source_url": "https://omlx.ai/benchmarks/performance/gbb6sv35",
   "tg_tokens": null,
   "tps": 40.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-ost30iqv-ctx1024",
   "model": "Qwen3.8-27B-heretic-ara",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 215.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-heretic-ara-oQ4e-mtp-fp16 on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 1k 215.7 32.5 18.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ost30iqv",
   "source_url": "https://omlx.ai/benchmarks/performance/ost30iqv",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-ost30iqv-ctx32768",
   "model": "Qwen3.8-27B-heretic-ara",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 25.6 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 201.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-heretic-ara-oQ4e-mtp-fp16 on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 32k 201.3 23.1 25.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ost30iqv",
   "source_url": "https://omlx.ai/benchmarks/performance/ost30iqv",
   "tg_tokens": null,
   "tps": 23.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-q9s97fri-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 9brz0qkm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 254.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 254.2 52.8 18.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q9s97fri",
   "source_url": "https://omlx.ai/benchmarks/performance/q9s97fri",
   "tg_tokens": null,
   "tps": 52.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-q9s97fri-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 9brz0qkm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 257.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 257.8 65.2 21.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q9s97fri",
   "source_url": "https://omlx.ai/benchmarks/performance/q9s97fri",
   "tg_tokens": null,
   "tps": 65.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-h88y2eai-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 56.8 GB; TTFT 6411 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 638.9,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-Flash-Next-oQ3e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 3bit; 4k tokens 638.9 PP tok/s 42.1 TG tok/s 6411 TTFT (ms) 56.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h88y2eai",
   "source_url": "https://omlx.ai/benchmarks/performance/h88y2eai",
   "tg_tokens": null,
   "tps": 42.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-gxof6m2w-ctx1024",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace, single run at 1k context; peak memory 11.5 GB; TTFT 21981 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 46.6,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 2bit; 1k tokens 46.6 PP tok/s 7.4 TG tok/s 21981 TTFT (ms) 11.5 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gxof6m2w",
   "source_url": "https://omlx.ai/benchmarks/performance/gxof6m2w",
   "tg_tokens": null,
   "tps": 7.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qt6eqp2u-ctx1024",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: no3mv4tb, sms8guj4, xz2oux4r; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2560.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,560 116.8 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qt6eqp2u",
   "source_url": "https://omlx.ai/benchmarks/performance/qt6eqp2u",
   "tg_tokens": null,
   "tps": 116.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qt6eqp2u-ctx4096",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: no3mv4tb, sms8guj4, xz2oux4r; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3302.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,302 102.5 22.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qt6eqp2u",
   "source_url": "https://omlx.ai/benchmarks/performance/qt6eqp2u",
   "tg_tokens": null,
   "tps": 102.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qt6eqp2u-ctx8192",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: no3mv4tb, sms8guj4, xz2oux4r; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3205.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 3,205 127.1 22.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qt6eqp2u",
   "source_url": "https://omlx.ai/benchmarks/performance/qt6eqp2u",
   "tg_tokens": null,
   "tps": 127.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qt6eqp2u-ctx16384",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: no3mv4tb, sms8guj4, xz2oux4r; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2930.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,930 142.5 22.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qt6eqp2u",
   "source_url": "https://omlx.ai/benchmarks/performance/qt6eqp2u",
   "tg_tokens": null,
   "tps": 142.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qt6eqp2u-ctx32768",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: no3mv4tb, sms8guj4, xz2oux4r; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2409.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 2,409 157.9 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qt6eqp2u",
   "source_url": "https://omlx.ai/benchmarks/performance/qt6eqp2u",
   "tg_tokens": null,
   "tps": 157.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-hkhxh6u3-ctx1024",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: jb47ecnr; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2395.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,395 112.1 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hkhxh6u3",
   "source_url": "https://omlx.ai/benchmarks/performance/hkhxh6u3",
   "tg_tokens": null,
   "tps": 112.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-hkhxh6u3-ctx4096",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: jb47ecnr; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3121.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,121 95.1 22.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hkhxh6u3",
   "source_url": "https://omlx.ai/benchmarks/performance/hkhxh6u3",
   "tg_tokens": null,
   "tps": 95.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-hkhxh6u3-ctx8192",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: jb47ecnr; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3267.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 3,267 124.6 22.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hkhxh6u3",
   "source_url": "https://omlx.ai/benchmarks/performance/hkhxh6u3",
   "tg_tokens": null,
   "tps": 124.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-hkhxh6u3-ctx16384",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: jb47ecnr; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3094.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 3,094 149.7 22.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hkhxh6u3",
   "source_url": "https://omlx.ai/benchmarks/performance/hkhxh6u3",
   "tg_tokens": null,
   "tps": 149.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-hkhxh6u3-ctx32768",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: jb47ecnr; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2476.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 2,476 149.5 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hkhxh6u3",
   "source_url": "https://omlx.ai/benchmarks/performance/hkhxh6u3",
   "tg_tokens": null,
   "tps": 149.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-v9ghcgox-ctx1024",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: zrk4ae79; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 394.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 4bit; 1k 394.3 27.6 12.7 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v9ghcgox",
   "source_url": "https://omlx.ai/benchmarks/performance/v9ghcgox",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-v9ghcgox-ctx4096",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.3 GB at this context; duplicate traces of the same session: zrk4ae79; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 428.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 4bit; 4k 428.5 26.8 13.3 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v9ghcgox",
   "source_url": "https://omlx.ai/benchmarks/performance/v9ghcgox",
   "tg_tokens": null,
   "tps": 26.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-tpxefmf5-ctx1024",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: gf0ovth5; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 386.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 4bit; 1k 386.9 28.4 12.7 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tpxefmf5",
   "source_url": "https://omlx.ai/benchmarks/performance/tpxefmf5",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-tpxefmf5-ctx4096",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.3 GB at this context; duplicate traces of the same session: gf0ovth5; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 418.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 4bit; 4k 418.2 27.5 13.3 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tpxefmf5",
   "source_url": "https://omlx.ai/benchmarks/performance/tpxefmf5",
   "tg_tokens": null,
   "tps": 27.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-t04ed527-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: xszb4xt7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 256.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 256.2 39.6 19.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t04ed527",
   "source_url": "https://omlx.ai/benchmarks/performance/t04ed527",
   "tg_tokens": null,
   "tps": 39.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-t04ed527-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: xszb4xt7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 259.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 259.2 46.2 21.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t04ed527",
   "source_url": "https://omlx.ai/benchmarks/performance/t04ed527",
   "tg_tokens": null,
   "tps": 46.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-s0lhm5a5-ctx1024",
   "model": "Qwen3.8-27B-heretic-ara",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 215.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-heretic-ara-oQ4e-mtp-fp16 on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 1k 215.7 32.5 18.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s0lhm5a5",
   "source_url": "https://omlx.ai/benchmarks/performance/s0lhm5a5",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-s0lhm5a5-ctx32768",
   "model": "Qwen3.8-27B-heretic-ara",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 25.6 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 201.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-heretic-ara-oQ4e-mtp-fp16 on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 32k 201.3 23.1 25.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s0lhm5a5",
   "source_url": "https://omlx.ai/benchmarks/performance/s0lhm5a5",
   "tg_tokens": null,
   "tps": 23.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-39iepkky-ctx1024",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: upt7f85q, y2a2hrll; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2395.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,395 112.1 21.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/39iepkky",
   "source_url": "https://omlx.ai/benchmarks/performance/39iepkky",
   "tg_tokens": null,
   "tps": 112.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-39iepkky-ctx4096",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: upt7f85q, y2a2hrll; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3121.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,121 95.1 22.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/39iepkky",
   "source_url": "https://omlx.ai/benchmarks/performance/39iepkky",
   "tg_tokens": null,
   "tps": 95.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-39iepkky-ctx8192",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: upt7f85q, y2a2hrll; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3267.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 3,267 124.6 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/39iepkky",
   "source_url": "https://omlx.ai/benchmarks/performance/39iepkky",
   "tg_tokens": null,
   "tps": 124.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-39iepkky-ctx16384",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: upt7f85q, y2a2hrll; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3094.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 3,094 149.7 22.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/39iepkky",
   "source_url": "https://omlx.ai/benchmarks/performance/39iepkky",
   "tg_tokens": null,
   "tps": 149.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-39iepkky-ctx32768",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: upt7f85q, y2a2hrll; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2476.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 2,476 149.5 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/39iepkky",
   "source_url": "https://omlx.ai/benchmarks/performance/39iepkky",
   "tg_tokens": null,
   "tps": 149.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ouius09l-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 1dwleykd, ydx1avkw; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2473.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,473 101.1 21.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ouius09l",
   "source_url": "https://omlx.ai/benchmarks/performance/ouius09l",
   "tg_tokens": null,
   "tps": 101.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ouius09l-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 1dwleykd, ydx1avkw; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3336.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,336 96.5 22.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ouius09l",
   "source_url": "https://omlx.ai/benchmarks/performance/ouius09l",
   "tg_tokens": null,
   "tps": 96.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ouius09l-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 1dwleykd, ydx1avkw; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3359.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 3,359 114.4 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ouius09l",
   "source_url": "https://omlx.ai/benchmarks/performance/ouius09l",
   "tg_tokens": null,
   "tps": 114.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ouius09l-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: 1dwleykd, ydx1avkw; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3030.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 3,030 129.7 22.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ouius09l",
   "source_url": "https://omlx.ai/benchmarks/performance/ouius09l",
   "tg_tokens": null,
   "tps": 129.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ouius09l-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 1dwleykd, ydx1avkw; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2480.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 2,480 232.8 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ouius09l",
   "source_url": "https://omlx.ai/benchmarks/performance/ouius09l",
   "tg_tokens": null,
   "tps": 232.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-lfz4lw2l-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: k5m22npn; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 50.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 4bit; 1k 50.4 5.4 15.8 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lfz4lw2l",
   "source_url": "https://omlx.ai/benchmarks/performance/lfz4lw2l",
   "tg_tokens": null,
   "tps": 5.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-lfz4lw2l-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: k5m22npn; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 51.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 4bit; 4k 51.9 5.2 17.2 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lfz4lw2l",
   "source_url": "https://omlx.ai/benchmarks/performance/lfz4lw2l",
   "tg_tokens": null,
   "tps": 5.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-rcm9yjr5-ctx1024",
   "model": "Qwen3.8-27B-heretic-ara",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 154.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-heretic-ara-oQ4e-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 1k 154.5 19.5 18.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rcm9yjr5",
   "source_url": "https://omlx.ai/benchmarks/performance/rcm9yjr5",
   "tg_tokens": null,
   "tps": 19.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-rcm9yjr5-ctx32768",
   "model": "Qwen3.8-27B-heretic-ara",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.6 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 121.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-heretic-ara-oQ4e-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 32k 121.6 19.7 25.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rcm9yjr5",
   "source_url": "https://omlx.ai/benchmarks/performance/rcm9yjr5",
   "tg_tokens": null,
   "tps": 19.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-xhhcbt68-ctx1024",
   "model": "gemma-4-26B-A4B-it-OptiQ-4bit-REAP-14B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 10.1 GB; TTFT 2985 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 343.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-OptiQ-4bit-REAP-14B on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 4bit; 1k tokens 343.1 PP tok/s 22.2 TG tok/s 2985 TTFT (ms) 10.1 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xhhcbt68",
   "source_url": "https://omlx.ai/benchmarks/performance/xhhcbt68",
   "tg_tokens": null,
   "tps": 22.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-b18eqjtp-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 14.5 GB at this context; duplicate traces of the same session: whrw1u3r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4163.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,163 117.7 14.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b18eqjtp",
   "source_url": "https://omlx.ai/benchmarks/performance/b18eqjtp",
   "tg_tokens": null,
   "tps": 117.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-b18eqjtp-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: whrw1u3r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4495.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 4,495 113.9 14.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b18eqjtp",
   "source_url": "https://omlx.ai/benchmarks/performance/b18eqjtp",
   "tg_tokens": null,
   "tps": 113.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qbgi2jho-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4199.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,199 101.4 15.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qbgi2jho",
   "source_url": "https://omlx.ai/benchmarks/performance/qbgi2jho",
   "tg_tokens": null,
   "tps": 101.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qbgi2jho-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4457.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 4,457 124.5 15.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qbgi2jho",
   "source_url": "https://omlx.ai/benchmarks/performance/qbgi2jho",
   "tg_tokens": null,
   "tps": 124.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-l6nk4mof-ctx1024",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: b52mxgd9, r04vj95h; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 896.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 896.7 102.9 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l6nk4mof",
   "source_url": "https://omlx.ai/benchmarks/performance/l6nk4mof",
   "tg_tokens": null,
   "tps": 102.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-l6nk4mof-ctx4096",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: b52mxgd9, r04vj95h; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1041.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,041 90.6 23.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l6nk4mof",
   "source_url": "https://omlx.ai/benchmarks/performance/l6nk4mof",
   "tg_tokens": null,
   "tps": 90.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-l6nk4mof-ctx8192",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: b52mxgd9, r04vj95h; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 983.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 983.5 91.5 23.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l6nk4mof",
   "source_url": "https://omlx.ai/benchmarks/performance/l6nk4mof",
   "tg_tokens": null,
   "tps": 91.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-l6nk4mof-ctx16384",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: b52mxgd9, r04vj95h; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 862.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 862.4 97.0 24.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l6nk4mof",
   "source_url": "https://omlx.ai/benchmarks/performance/l6nk4mof",
   "tg_tokens": null,
   "tps": 97.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-l6nk4mof-ctx32768",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: b52mxgd9, r04vj95h; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 660.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 660.9 78.8 25.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l6nk4mof",
   "source_url": "https://omlx.ai/benchmarks/performance/l6nk4mof",
   "tg_tokens": null,
   "tps": 78.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-l6nk4mof-ctx65536",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: b52mxgd9, r04vj95h; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 486.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 486.5 57.9 26.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l6nk4mof",
   "source_url": "https://omlx.ai/benchmarks/performance/l6nk4mof",
   "tg_tokens": null,
   "tps": 57.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-jm918vvw-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: pncziryf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2473.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,473 101.1 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jm918vvw",
   "source_url": "https://omlx.ai/benchmarks/performance/jm918vvw",
   "tg_tokens": null,
   "tps": 101.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-jm918vvw-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: pncziryf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3336.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,336 96.5 22.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jm918vvw",
   "source_url": "https://omlx.ai/benchmarks/performance/jm918vvw",
   "tg_tokens": null,
   "tps": 96.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-jm918vvw-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: pncziryf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3359.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 3,359 114.4 22.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jm918vvw",
   "source_url": "https://omlx.ai/benchmarks/performance/jm918vvw",
   "tg_tokens": null,
   "tps": 114.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-jm918vvw-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: pncziryf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3030.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 3,030 129.7 22.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jm918vvw",
   "source_url": "https://omlx.ai/benchmarks/performance/jm918vvw",
   "tg_tokens": null,
   "tps": 129.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-jm918vvw-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: pncziryf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2480.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 2,480 232.8 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jm918vvw",
   "source_url": "https://omlx.ai/benchmarks/performance/jm918vvw",
   "tg_tokens": null,
   "tps": 232.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-10kkb6rt-ctx1024",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 9ytzd2og, ccm0qdyp, ten0kvvd; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1397.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 1,397 98.5 16.6 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/10kkb6rt",
   "source_url": "https://omlx.ai/benchmarks/performance/10kkb6rt",
   "tg_tokens": null,
   "tps": 98.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-10kkb6rt-ctx4096",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 9ytzd2og, ccm0qdyp, ten0kvvd; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1326.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 1,326 86.3 17.0 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/10kkb6rt",
   "source_url": "https://omlx.ai/benchmarks/performance/10kkb6rt",
   "tg_tokens": null,
   "tps": 86.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-10kkb6rt-ctx8192",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: 9ytzd2og, ccm0qdyp, ten0kvvd; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1040.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 1,040 75.8 17.4 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/10kkb6rt",
   "source_url": "https://omlx.ai/benchmarks/performance/10kkb6rt",
   "tg_tokens": null,
   "tps": 75.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-10kkb6rt-ctx16384",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: 9ytzd2og, ccm0qdyp, ten0kvvd; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 721.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 721.9 64.7 18.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/10kkb6rt",
   "source_url": "https://omlx.ai/benchmarks/performance/10kkb6rt",
   "tg_tokens": null,
   "tps": 64.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-qyr6vtim-ctx1024",
   "model": "Qwen3.8-27B-heretic-ara",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 154.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-heretic-ara-oQ4e-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 1k 154.5 19.5 18.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qyr6vtim",
   "source_url": "https://omlx.ai/benchmarks/performance/qyr6vtim",
   "tg_tokens": null,
   "tps": 19.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-qyr6vtim-ctx32768",
   "model": "Qwen3.8-27B-heretic-ara",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.6 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 121.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-heretic-ara-oQ4e-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 32k 121.6 19.7 25.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qyr6vtim",
   "source_url": "https://omlx.ai/benchmarks/performance/qyr6vtim",
   "tg_tokens": null,
   "tps": 19.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-8t9oqm34-ctx1024",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace, single run at 1k context; peak memory 11.5 GB; TTFT 21894 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 46.8,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 2bit; 1k tokens 46.8 PP tok/s 7.4 TG tok/s 21894 TTFT (ms) 11.5 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8t9oqm34",
   "source_url": "https://omlx.ai/benchmarks/performance/8t9oqm34",
   "tg_tokens": null,
   "tps": 7.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5m2c8tzp-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4199.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,199 101.4 15.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5m2c8tzp",
   "source_url": "https://omlx.ai/benchmarks/performance/5m2c8tzp",
   "tg_tokens": null,
   "tps": 101.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5m2c8tzp-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4457.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 4,457 124.5 15.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5m2c8tzp",
   "source_url": "https://omlx.ai/benchmarks/performance/5m2c8tzp",
   "tg_tokens": null,
   "tps": 124.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rrokvjrp-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 102.4 GB; TTFT 23829 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1375.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "qwen3.8-flash-next-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k tokens 1,375 PP tok/s 70.4 TG tok/s 23829 TTFT (ms) 102.4 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rrokvjrp",
   "source_url": "https://omlx.ai/benchmarks/performance/rrokvjrp",
   "tg_tokens": null,
   "tps": 70.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-1budesmc-ctx1024",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: ndnpqp9b, xv4ewugc; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 896.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 896.7 102.9 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1budesmc",
   "source_url": "https://omlx.ai/benchmarks/performance/1budesmc",
   "tg_tokens": null,
   "tps": 102.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-1budesmc-ctx4096",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: ndnpqp9b, xv4ewugc; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1041.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,041 90.6 23.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1budesmc",
   "source_url": "https://omlx.ai/benchmarks/performance/1budesmc",
   "tg_tokens": null,
   "tps": 90.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-1budesmc-ctx8192",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: ndnpqp9b, xv4ewugc; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 983.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 983.5 91.5 23.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1budesmc",
   "source_url": "https://omlx.ai/benchmarks/performance/1budesmc",
   "tg_tokens": null,
   "tps": 91.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-1budesmc-ctx16384",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: ndnpqp9b, xv4ewugc; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 862.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 862.4 97.0 24.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1budesmc",
   "source_url": "https://omlx.ai/benchmarks/performance/1budesmc",
   "tg_tokens": null,
   "tps": 97.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-1budesmc-ctx32768",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: ndnpqp9b, xv4ewugc; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 660.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 660.9 78.8 25.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1budesmc",
   "source_url": "https://omlx.ai/benchmarks/performance/1budesmc",
   "tg_tokens": null,
   "tps": 78.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-1budesmc-ctx65536",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: ndnpqp9b, xv4ewugc; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 486.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 486.5 57.9 26.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1budesmc",
   "source_url": "https://omlx.ai/benchmarks/performance/1budesmc",
   "tg_tokens": null,
   "tps": 57.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-390cfq4u-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.3 GB at this context; duplicate traces of the same session: jngq0j22, lvot8lk4, oi51d4rp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1320.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 1k 1,320 48.3 10.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/390cfq4u",
   "source_url": "https://omlx.ai/benchmarks/performance/390cfq4u",
   "tg_tokens": null,
   "tps": 48.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-390cfq4u-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.2 GB at this context; duplicate traces of the same session: jngq0j22, lvot8lk4, oi51d4rp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1481.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 4k 1,481 83.6 11.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/390cfq4u",
   "source_url": "https://omlx.ai/benchmarks/performance/390cfq4u",
   "tg_tokens": null,
   "tps": 83.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-390cfq4u-ctx8192",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.6 GB at this context; duplicate traces of the same session: jngq0j22, lvot8lk4, oi51d4rp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1606.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 8k 1,606 38.1 11.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/390cfq4u",
   "source_url": "https://omlx.ai/benchmarks/performance/390cfq4u",
   "tg_tokens": null,
   "tps": 38.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-390cfq4u-ctx16384",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 12.3 GB at this context; duplicate traces of the same session: jngq0j22, lvot8lk4, oi51d4rp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1467.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 16k 1,467 42.0 12.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/390cfq4u",
   "source_url": "https://omlx.ai/benchmarks/performance/390cfq4u",
   "tg_tokens": null,
   "tps": 42.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-390cfq4u-ctx32768",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 13.6 GB at this context; duplicate traces of the same session: jngq0j22, lvot8lk4, oi51d4rp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1254.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 32k 1,254 45.5 13.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/390cfq4u",
   "source_url": "https://omlx.ai/benchmarks/performance/390cfq4u",
   "tg_tokens": null,
   "tps": 45.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-390cfq4u-ctx65536",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: jngq0j22, lvot8lk4, oi51d4rp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1110.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 64k 1,110 55.1 16.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/390cfq4u",
   "source_url": "https://omlx.ai/benchmarks/performance/390cfq4u",
   "tg_tokens": null,
   "tps": 55.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-390cfq4u-ctx131072",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: jngq0j22, lvot8lk4, oi51d4rp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 930.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 128k 930.8 36.5 21.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/390cfq4u",
   "source_url": "https://omlx.ai/benchmarks/performance/390cfq4u",
   "tg_tokens": null,
   "tps": 36.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-vq8hi5g1-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "UD MLX build; oMLX community benchmark trace; peak memory 16.5 GB at this context; duplicate traces of the same session: 8g2sw5yo; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 411.8,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-3bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 3bit; 1k 411.8 60.2 16.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vq8hi5g1",
   "source_url": "https://omlx.ai/benchmarks/performance/vq8hi5g1",
   "tg_tokens": null,
   "tps": 60.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-vq8hi5g1-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "UD MLX build; oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: 8g2sw5yo; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 427.7,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-3bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 3bit; 4k 427.7 57.7 17.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vq8hi5g1",
   "source_url": "https://omlx.ai/benchmarks/performance/vq8hi5g1",
   "tg_tokens": null,
   "tps": 57.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-vq8hi5g1-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "UD MLX build; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: 8g2sw5yo; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 443.9,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-3bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 3bit; 8k 443.9 54.8 17.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vq8hi5g1",
   "source_url": "https://omlx.ai/benchmarks/performance/vq8hi5g1",
   "tg_tokens": null,
   "tps": 54.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-vq8hi5g1-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "UD MLX build; oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: 8g2sw5yo; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 421.4,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-3bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 3bit; 16k 421.4 50.6 18.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vq8hi5g1",
   "source_url": "https://omlx.ai/benchmarks/performance/vq8hi5g1",
   "tg_tokens": null,
   "tps": 50.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2ov3xyou-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 6c90ikjx, ij51rfk1, rgwpp6zi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 927.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 927.2 33.6 16.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ov3xyou",
   "source_url": "https://omlx.ai/benchmarks/performance/2ov3xyou",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2ov3xyou-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 6c90ikjx, ij51rfk1, rgwpp6zi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 985.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 985.2 32.6 17.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ov3xyou",
   "source_url": "https://omlx.ai/benchmarks/performance/2ov3xyou",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2ov3xyou-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 6c90ikjx, ij51rfk1, rgwpp6zi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 769.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 769.1 29.1 18.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ov3xyou",
   "source_url": "https://omlx.ai/benchmarks/performance/2ov3xyou",
   "tg_tokens": null,
   "tps": 29.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2ov3xyou-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 6c90ikjx, ij51rfk1, rgwpp6zi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 659.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 659.1 25.1 20.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ov3xyou",
   "source_url": "https://omlx.ai/benchmarks/performance/2ov3xyou",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-0b3bkze5-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 52l5cknl, f2grg6zx, sn2cnn70, xievqsqk, xn02pdsg; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2734.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 2,734 104.6 20.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0b3bkze5",
   "source_url": "https://omlx.ai/benchmarks/performance/0b3bkze5",
   "tg_tokens": null,
   "tps": 104.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0b3bkze5-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 52l5cknl, f2grg6zx, sn2cnn70, xievqsqk, xn02pdsg; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 570.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 570.3 89.2 20.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0b3bkze5",
   "source_url": "https://omlx.ai/benchmarks/performance/0b3bkze5",
   "tg_tokens": null,
   "tps": 89.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-0b3bkze5-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 52l5cknl, f2grg6zx, sn2cnn70, xievqsqk, xn02pdsg; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3061.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 3,061 80.6 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0b3bkze5",
   "source_url": "https://omlx.ai/benchmarks/performance/0b3bkze5",
   "tg_tokens": null,
   "tps": 80.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0b3bkze5-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: 52l5cknl, f2grg6zx, sn2cnn70, xievqsqk, xn02pdsg; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2837.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 2,837 69.8 20.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0b3bkze5",
   "source_url": "https://omlx.ai/benchmarks/performance/0b3bkze5",
   "tg_tokens": null,
   "tps": 69.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0b3bkze5-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 52l5cknl, f2grg6zx, sn2cnn70, xievqsqk, xn02pdsg; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2438.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 2,438 53.7 20.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0b3bkze5",
   "source_url": "https://omlx.ai/benchmarks/performance/0b3bkze5",
   "tg_tokens": null,
   "tps": 53.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0b3bkze5-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 52l5cknl, f2grg6zx, sn2cnn70, xievqsqk, xn02pdsg; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1806.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 1,806 35.3 21.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0b3bkze5",
   "source_url": "https://omlx.ai/benchmarks/performance/0b3bkze5",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-zp54fm19-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.3 GB at this context; duplicate traces of the same session: 7597ar2x, hytur4h3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1320.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 1k 1,320 48.3 10.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zp54fm19",
   "source_url": "https://omlx.ai/benchmarks/performance/zp54fm19",
   "tg_tokens": null,
   "tps": 48.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-zp54fm19-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.2 GB at this context; duplicate traces of the same session: 7597ar2x, hytur4h3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1481.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 4k 1,481 83.6 11.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zp54fm19",
   "source_url": "https://omlx.ai/benchmarks/performance/zp54fm19",
   "tg_tokens": null,
   "tps": 83.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-zp54fm19-ctx8192",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.6 GB at this context; duplicate traces of the same session: 7597ar2x, hytur4h3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1606.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 8k 1,606 38.1 11.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zp54fm19",
   "source_url": "https://omlx.ai/benchmarks/performance/zp54fm19",
   "tg_tokens": null,
   "tps": 38.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-zp54fm19-ctx16384",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 12.3 GB at this context; duplicate traces of the same session: 7597ar2x, hytur4h3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1467.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 16k 1,467 42.0 12.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zp54fm19",
   "source_url": "https://omlx.ai/benchmarks/performance/zp54fm19",
   "tg_tokens": null,
   "tps": 42.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-zp54fm19-ctx32768",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 13.6 GB at this context; duplicate traces of the same session: 7597ar2x, hytur4h3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1254.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 32k 1,254 45.5 13.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zp54fm19",
   "source_url": "https://omlx.ai/benchmarks/performance/zp54fm19",
   "tg_tokens": null,
   "tps": 45.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-zp54fm19-ctx65536",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: 7597ar2x, hytur4h3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1110.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 64k 1,110 55.1 16.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zp54fm19",
   "source_url": "https://omlx.ai/benchmarks/performance/zp54fm19",
   "tg_tokens": null,
   "tps": 55.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-zp54fm19-ctx131072",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 7597ar2x, hytur4h3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 930.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 128k 930.8 36.5 21.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zp54fm19",
   "source_url": "https://omlx.ai/benchmarks/performance/zp54fm19",
   "tg_tokens": null,
   "tps": 36.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-abc5ibzg-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "UD MLX build; oMLX community benchmark trace; peak memory 16.5 GB at this context; duplicate traces of the same session: xjx51ibz; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 411.8,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-3bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 3bit; 1k 411.8 60.2 16.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/abc5ibzg",
   "source_url": "https://omlx.ai/benchmarks/performance/abc5ibzg",
   "tg_tokens": null,
   "tps": 60.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-abc5ibzg-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "UD MLX build; oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: xjx51ibz; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 427.7,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-3bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 3bit; 4k 427.7 57.7 17.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/abc5ibzg",
   "source_url": "https://omlx.ai/benchmarks/performance/abc5ibzg",
   "tg_tokens": null,
   "tps": 57.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-abc5ibzg-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "UD MLX build; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: xjx51ibz; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 443.9,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-3bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 3bit; 8k 443.9 54.8 17.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/abc5ibzg",
   "source_url": "https://omlx.ai/benchmarks/performance/abc5ibzg",
   "tg_tokens": null,
   "tps": 54.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-abc5ibzg-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "UD MLX build; oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: xjx51ibz; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 421.4,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-3bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 3bit; 16k 421.4 50.6 18.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/abc5ibzg",
   "source_url": "https://omlx.ai/benchmarks/performance/abc5ibzg",
   "tg_tokens": null,
   "tps": 50.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-u2re6r23-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: mb0qc3hp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1076.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,076 67.9 19.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u2re6r23",
   "source_url": "https://omlx.ai/benchmarks/performance/u2re6r23",
   "tg_tokens": null,
   "tps": 67.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-u2re6r23-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: mb0qc3hp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1328.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,328 59.5 20.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u2re6r23",
   "source_url": "https://omlx.ai/benchmarks/performance/u2re6r23",
   "tg_tokens": null,
   "tps": 59.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-u2re6r23-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: mb0qc3hp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1354.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,354 68.2 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u2re6r23",
   "source_url": "https://omlx.ai/benchmarks/performance/u2re6r23",
   "tg_tokens": null,
   "tps": 68.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bkpjjolm-ctx4096",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "MTPLX speed-optimized build; oMLX community benchmark trace; peak memory 40.7 GB at this context; duplicate traces of the same session: 8paal4fd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 720.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Abliterated-MTPLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 720.6 32.9 40.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bkpjjolm",
   "source_url": "https://omlx.ai/benchmarks/performance/bkpjjolm",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bkpjjolm-ctx16384",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "MTPLX speed-optimized build; oMLX community benchmark trace; peak memory 41.4 GB at this context; duplicate traces of the same session: 8paal4fd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 596.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Abliterated-MTPLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 596.1 47.2 41.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bkpjjolm",
   "source_url": "https://omlx.ai/benchmarks/performance/bkpjjolm",
   "tg_tokens": null,
   "tps": 47.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-1dc8ysmn-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 19.6 GB; TTFT 31587 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 129.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 24 GB \u00b7 4bit; 4k tokens 129.7 PP tok/s 23.5 TG tok/s 31587 TTFT (ms) 19.6 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1dc8ysmn",
   "source_url": "https://omlx.ai/benchmarks/performance/1dc8ysmn",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-z75p1dho-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 538.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 538.4 66.7 70.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z75p1dho",
   "source_url": "https://omlx.ai/benchmarks/performance/z75p1dho",
   "tg_tokens": null,
   "tps": 66.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-z75p1dho-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 756.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 756.3 62.5 71.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z75p1dho",
   "source_url": "https://omlx.ai/benchmarks/performance/z75p1dho",
   "tg_tokens": null,
   "tps": 62.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-z75p1dho-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 822.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 822.6 57.8 72.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z75p1dho",
   "source_url": "https://omlx.ai/benchmarks/performance/z75p1dho",
   "tg_tokens": null,
   "tps": 57.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-eqefy9y1-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1076.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,076 67.9 19.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eqefy9y1",
   "source_url": "https://omlx.ai/benchmarks/performance/eqefy9y1",
   "tg_tokens": null,
   "tps": 67.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-eqefy9y1-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1328.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,328 59.5 20.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eqefy9y1",
   "source_url": "https://omlx.ai/benchmarks/performance/eqefy9y1",
   "tg_tokens": null,
   "tps": 59.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-eqefy9y1-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1354.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,354 68.2 21.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eqefy9y1",
   "source_url": "https://omlx.ai/benchmarks/performance/eqefy9y1",
   "tg_tokens": null,
   "tps": 68.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-89e2a82g-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: grr8n1fk, grwrro7s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 255.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 255.3 53.6 21.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/89e2a82g",
   "source_url": "https://omlx.ai/benchmarks/performance/89e2a82g",
   "tg_tokens": null,
   "tps": 53.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-89e2a82g-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: grr8n1fk, grwrro7s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 340.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 340.2 57.2 25.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/89e2a82g",
   "source_url": "https://omlx.ai/benchmarks/performance/89e2a82g",
   "tg_tokens": null,
   "tps": 57.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-89e2a82g-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.6 GB at this context; duplicate traces of the same session: grr8n1fk, grwrro7s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 342.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 342.5 60.9 25.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/89e2a82g",
   "source_url": "https://omlx.ai/benchmarks/performance/89e2a82g",
   "tg_tokens": null,
   "tps": 60.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-89e2a82g-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: grr8n1fk, grwrro7s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 325.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 325.7 55.2 26.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/89e2a82g",
   "source_url": "https://omlx.ai/benchmarks/performance/89e2a82g",
   "tg_tokens": null,
   "tps": 55.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-89e2a82g-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.4 GB at this context; duplicate traces of the same session: grr8n1fk, grwrro7s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 295.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 295.6 117.2 28.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/89e2a82g",
   "source_url": "https://omlx.ai/benchmarks/performance/89e2a82g",
   "tg_tokens": null,
   "tps": 117.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-89e2a82g-ctx65536",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.1 GB at this context; duplicate traces of the same session: grr8n1fk, grwrro7s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 249.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 249.5 47.4 31.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/89e2a82g",
   "source_url": "https://omlx.ai/benchmarks/performance/89e2a82g",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ytwbgphd-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 40.1 GB at this context; duplicate traces of the same session: 90eyyt37, oip7aq73, w42tc73d, xw0okdg8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 252.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 252.2 44.8 40.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ytwbgphd",
   "source_url": "https://omlx.ai/benchmarks/performance/ytwbgphd",
   "tg_tokens": null,
   "tps": 44.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ytwbgphd-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.4 GB at this context; duplicate traces of the same session: 90eyyt37, oip7aq73, w42tc73d, xw0okdg8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 332.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 332.7 33.4 43.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ytwbgphd",
   "source_url": "https://omlx.ai/benchmarks/performance/ytwbgphd",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ytwbgphd-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.8 GB at this context; duplicate traces of the same session: 90eyyt37, oip7aq73, w42tc73d, xw0okdg8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 334.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 334.2 41.4 43.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ytwbgphd",
   "source_url": "https://omlx.ai/benchmarks/performance/ytwbgphd",
   "tg_tokens": null,
   "tps": 41.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ytwbgphd-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 44.9 GB at this context; duplicate traces of the same session: 90eyyt37, oip7aq73, w42tc73d, xw0okdg8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 318.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 318.5 48.4 44.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ytwbgphd",
   "source_url": "https://omlx.ai/benchmarks/performance/ytwbgphd",
   "tg_tokens": null,
   "tps": 48.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ytwbgphd-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 46.7 GB at this context; duplicate traces of the same session: 90eyyt37, oip7aq73, w42tc73d, xw0okdg8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 289.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 289.7 64.4 46.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ytwbgphd",
   "source_url": "https://omlx.ai/benchmarks/performance/ytwbgphd",
   "tg_tokens": null,
   "tps": 64.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-s8cw790z-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 21.4 GB; TTFT 1306 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 784.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k tokens 784.0 PP tok/s 92.6 TG tok/s 1306 TTFT (ms) 21.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s8cw790z",
   "source_url": "https://omlx.ai/benchmarks/performance/s8cw790z",
   "tg_tokens": null,
   "tps": 92.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-1wjgzmuq-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 21.4 GB; TTFT 1307 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 783.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k tokens 783.7 PP tok/s 105.2 TG tok/s 1307 TTFT (ms) 21.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1wjgzmuq",
   "source_url": "https://omlx.ai/benchmarks/performance/1wjgzmuq",
   "tg_tokens": null,
   "tps": 105.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-fp75qykl-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 21.4 GB; TTFT 1322 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 774.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k tokens 774.6 PP tok/s 95.0 TG tok/s 1322 TTFT (ms) 21.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fp75qykl",
   "source_url": "https://omlx.ai/benchmarks/performance/fp75qykl",
   "tg_tokens": null,
   "tps": 95.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-w2yt01xn-ctx131072",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace, single run at 128k context; peak memory 28.9 GB; TTFT 419949 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 312.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 128k tokens 312.1 PP tok/s 9.8 TG tok/s 419949 TTFT (ms) 28.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w2yt01xn",
   "source_url": "https://omlx.ai/benchmarks/performance/w2yt01xn",
   "tg_tokens": null,
   "tps": 9.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-v5gkl119-ctx4096",
   "model": "Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-uncher",
   "notes": "bf16 last4 8bit vision MTP build; oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: t5ttk660; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 145.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-uncher-oQ5e-6target-bf16-last4_8bit-vision-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 5bit; 4k 145.0 14.4 26.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v5gkl119",
   "source_url": "https://omlx.ai/benchmarks/performance/v5gkl119",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-v5gkl119-ctx16384",
   "model": "Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-uncher",
   "notes": "bf16 last4 8bit vision MTP build; oMLX community benchmark trace; peak memory 29.1 GB at this context; duplicate traces of the same session: t5ttk660; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 134.6,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-uncher-oQ5e-6target-bf16-last4_8bit-vision-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 5bit; 16k 134.6 15.4 29.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v5gkl119",
   "source_url": "https://omlx.ai/benchmarks/performance/v5gkl119",
   "tg_tokens": null,
   "tps": 15.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-3o53va4f-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: svu0d74o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 538.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 538.4 66.7 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3o53va4f",
   "source_url": "https://omlx.ai/benchmarks/performance/3o53va4f",
   "tg_tokens": null,
   "tps": 66.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-3o53va4f-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: svu0d74o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 756.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 756.3 62.5 71.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3o53va4f",
   "source_url": "https://omlx.ai/benchmarks/performance/3o53va4f",
   "tg_tokens": null,
   "tps": 62.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-3o53va4f-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.3 GB at this context; duplicate traces of the same session: svu0d74o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 822.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 822.6 57.8 72.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3o53va4f",
   "source_url": "https://omlx.ai/benchmarks/performance/3o53va4f",
   "tg_tokens": null,
   "tps": 57.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-k1boj2ew-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: izm03noe, sfj3t5ah; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 255.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 255.3 53.6 21.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k1boj2ew",
   "source_url": "https://omlx.ai/benchmarks/performance/k1boj2ew",
   "tg_tokens": null,
   "tps": 53.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-k1boj2ew-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: izm03noe, sfj3t5ah; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 340.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 340.2 57.2 25.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k1boj2ew",
   "source_url": "https://omlx.ai/benchmarks/performance/k1boj2ew",
   "tg_tokens": null,
   "tps": 57.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-k1boj2ew-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.6 GB at this context; duplicate traces of the same session: izm03noe, sfj3t5ah; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 342.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 342.5 60.9 25.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k1boj2ew",
   "source_url": "https://omlx.ai/benchmarks/performance/k1boj2ew",
   "tg_tokens": null,
   "tps": 60.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-k1boj2ew-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: izm03noe, sfj3t5ah; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 325.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 325.7 55.2 26.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k1boj2ew",
   "source_url": "https://omlx.ai/benchmarks/performance/k1boj2ew",
   "tg_tokens": null,
   "tps": 55.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-k1boj2ew-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.4 GB at this context; duplicate traces of the same session: izm03noe, sfj3t5ah; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 295.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 295.6 117.2 28.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k1boj2ew",
   "source_url": "https://omlx.ai/benchmarks/performance/k1boj2ew",
   "tg_tokens": null,
   "tps": 117.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-k1boj2ew-ctx65536",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.1 GB at this context; duplicate traces of the same session: izm03noe, sfj3t5ah; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 249.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 249.5 47.4 31.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k1boj2ew",
   "source_url": "https://omlx.ai/benchmarks/performance/k1boj2ew",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-n05xgnwj-ctx4096",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 197.4,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 5bit; 4k 197.4 23.3 21.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n05xgnwj",
   "source_url": "https://omlx.ai/benchmarks/performance/n05xgnwj",
   "tg_tokens": null,
   "tps": 23.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-n05xgnwj-ctx16384",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 74.8,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 5bit; 16k 74.8 12.8 24.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n05xgnwj",
   "source_url": "https://omlx.ai/benchmarks/performance/n05xgnwj",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-fdg9oluc-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 21.4 GB; TTFT 1316 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 778.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k tokens 778.3 PP tok/s 95.1 TG tok/s 1316 TTFT (ms) 21.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fdg9oluc",
   "source_url": "https://omlx.ai/benchmarks/performance/fdg9oluc",
   "tg_tokens": null,
   "tps": 95.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-2jghf7hb-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 0u29fmbj, 61ocz24k; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 778.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 778.0 105.4 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2jghf7hb",
   "source_url": "https://omlx.ai/benchmarks/performance/2jghf7hb",
   "tg_tokens": null,
   "tps": 105.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-2jghf7hb-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: 0u29fmbj, 61ocz24k; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 857.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 857.6 82.7 23.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2jghf7hb",
   "source_url": "https://omlx.ai/benchmarks/performance/2jghf7hb",
   "tg_tokens": null,
   "tps": 82.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-2jghf7hb-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: 0u29fmbj, 61ocz24k; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 800.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 8k 800.6 77.9 23.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2jghf7hb",
   "source_url": "https://omlx.ai/benchmarks/performance/2jghf7hb",
   "tg_tokens": null,
   "tps": 77.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-2jghf7hb-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: 0u29fmbj, 61ocz24k; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 710.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 16k 710.1 74.8 24.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2jghf7hb",
   "source_url": "https://omlx.ai/benchmarks/performance/2jghf7hb",
   "tg_tokens": null,
   "tps": 74.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-wzv23f4m-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 67.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 1k 67.2 14.9 15.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wzv23f4m",
   "source_url": "https://omlx.ai/benchmarks/performance/wzv23f4m",
   "tg_tokens": null,
   "tps": 14.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-wzv23f4m-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 67.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k 67.5 14.6 17.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wzv23f4m",
   "source_url": "https://omlx.ai/benchmarks/performance/wzv23f4m",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-wzv23f4m-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.9 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 67.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 8k 67.7 14.2 17.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wzv23f4m",
   "source_url": "https://omlx.ai/benchmarks/performance/wzv23f4m",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-wzv23f4m-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 66.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 16k 66.5 13.1 19.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wzv23f4m",
   "source_url": "https://omlx.ai/benchmarks/performance/wzv23f4m",
   "tg_tokens": null,
   "tps": 13.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-wsnut08y-ctx4096",
   "model": "Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-uncher",
   "notes": "bf16 last4 8bit vision MTP build; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: 11wqkqo4; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 149.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-uncher-oQ5e-6target-bf16-last4_8bit-vision-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 5bit; 4k 149.0 18.3 25.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wsnut08y",
   "source_url": "https://omlx.ai/benchmarks/performance/wsnut08y",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-wsnut08y-ctx16384",
   "model": "Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-uncher",
   "notes": "bf16 last4 8bit vision MTP build; oMLX community benchmark trace; peak memory 27.2 GB at this context; duplicate traces of the same session: 11wqkqo4; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 138.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-uncher-oQ5e-6target-bf16-last4_8bit-vision-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 5bit; 16k 138.1 16.5 27.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wsnut08y",
   "source_url": "https://omlx.ai/benchmarks/performance/wsnut08y",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-4knapccv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: lahmfa4c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 230.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 230.1 35.6 22.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4knapccv",
   "source_url": "https://omlx.ai/benchmarks/performance/4knapccv",
   "tg_tokens": null,
   "tps": 35.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-4knapccv-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: lahmfa4c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 203.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 203.7 43.6 20.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4knapccv",
   "source_url": "https://omlx.ai/benchmarks/performance/4knapccv",
   "tg_tokens": null,
   "tps": 43.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-kbm2qihz-ctx4096",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 197.4,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 5bit; 4k 197.4 23.3 21.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kbm2qihz",
   "source_url": "https://omlx.ai/benchmarks/performance/kbm2qihz",
   "tg_tokens": null,
   "tps": 23.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-kbm2qihz-ctx16384",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 74.8,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 5bit; 16k 74.8 12.8 24.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kbm2qihz",
   "source_url": "https://omlx.ai/benchmarks/performance/kbm2qihz",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-22lxrlj8-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 778.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 778.0 105.4 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/22lxrlj8",
   "source_url": "https://omlx.ai/benchmarks/performance/22lxrlj8",
   "tg_tokens": null,
   "tps": 105.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-22lxrlj8-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 857.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 857.6 82.7 23.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/22lxrlj8",
   "source_url": "https://omlx.ai/benchmarks/performance/22lxrlj8",
   "tg_tokens": null,
   "tps": 82.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-22lxrlj8-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 800.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 8k 800.6 77.9 23.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/22lxrlj8",
   "source_url": "https://omlx.ai/benchmarks/performance/22lxrlj8",
   "tg_tokens": null,
   "tps": 77.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-22lxrlj8-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 710.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 16k 710.1 74.8 24.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/22lxrlj8",
   "source_url": "https://omlx.ai/benchmarks/performance/22lxrlj8",
   "tg_tokens": null,
   "tps": 74.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-8cf4cn1m-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: tynbojge, y8k46lmr; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 745.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 745.5 101.9 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8cf4cn1m",
   "source_url": "https://omlx.ai/benchmarks/performance/8cf4cn1m",
   "tg_tokens": null,
   "tps": 101.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-8cf4cn1m-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: tynbojge, y8k46lmr; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 836.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 8k 836.9 73.3 23.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8cf4cn1m",
   "source_url": "https://omlx.ai/benchmarks/performance/8cf4cn1m",
   "tg_tokens": null,
   "tps": 73.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-8cf4cn1m-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: tynbojge, y8k46lmr; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 740.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 16k 740.7 69.7 24.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8cf4cn1m",
   "source_url": "https://omlx.ai/benchmarks/performance/8cf4cn1m",
   "tg_tokens": null,
   "tps": 69.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-tf8lwzpm-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: cwbzwrgw, osgt3ih2; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 67.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 1k 67.2 14.9 15.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tf8lwzpm",
   "source_url": "https://omlx.ai/benchmarks/performance/tf8lwzpm",
   "tg_tokens": null,
   "tps": 14.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-tf8lwzpm-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: cwbzwrgw, osgt3ih2; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 67.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k 67.5 14.6 17.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tf8lwzpm",
   "source_url": "https://omlx.ai/benchmarks/performance/tf8lwzpm",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-tf8lwzpm-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.9 GB at this context; duplicate traces of the same session: cwbzwrgw, osgt3ih2; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 67.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 8k 67.7 14.2 17.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tf8lwzpm",
   "source_url": "https://omlx.ai/benchmarks/performance/tf8lwzpm",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-tf8lwzpm-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: cwbzwrgw, osgt3ih2; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 66.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 16k 66.5 13.1 19.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tf8lwzpm",
   "source_url": "https://omlx.ai/benchmarks/performance/tf8lwzpm",
   "tg_tokens": null,
   "tps": 13.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-s66tpn1v-ctx4096",
   "model": "Qwen3.8-27B-TWIN-TURBO-f-c-f-709-l-unc",
   "notes": "abbreviated TWIN-TURBO-Fable-Cold-Fusion-709-L build; aura6 last4 bf16 vision MTP; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 144.6,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-TWIN-TURBO-f-c-f-709-l-unc-oQ5e-aura6-last4-bf16-vision-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 5bit; 4k 144.6 19.4 24.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s66tpn1v",
   "source_url": "https://omlx.ai/benchmarks/performance/s66tpn1v",
   "tg_tokens": null,
   "tps": 19.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-s66tpn1v-ctx16384",
   "model": "Qwen3.8-27B-TWIN-TURBO-f-c-f-709-l-unc",
   "notes": "abbreviated TWIN-TURBO-Fable-Cold-Fusion-709-L build; aura6 last4 bf16 vision MTP; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 133.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-TWIN-TURBO-f-c-f-709-l-unc-oQ5e-aura6-last4-bf16-vision-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 5bit; 16k 133.1 18.0 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s66tpn1v",
   "source_url": "https://omlx.ai/benchmarks/performance/s66tpn1v",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gh4aiwx6-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 954.7,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 954.7 57.0 20.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gh4aiwx6",
   "source_url": "https://omlx.ai/benchmarks/performance/gh4aiwx6",
   "tg_tokens": null,
   "tps": 57.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-gh4aiwx6-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1358.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,358 87.7 21.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gh4aiwx6",
   "source_url": "https://omlx.ai/benchmarks/performance/gh4aiwx6",
   "tg_tokens": null,
   "tps": 87.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gh4aiwx6-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1327.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,327 42.1 22.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gh4aiwx6",
   "source_url": "https://omlx.ai/benchmarks/performance/gh4aiwx6",
   "tg_tokens": null,
   "tps": 42.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gh4aiwx6-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1269.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,269 58.0 22.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gh4aiwx6",
   "source_url": "https://omlx.ai/benchmarks/performance/gh4aiwx6",
   "tg_tokens": null,
   "tps": 58.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qfb6bv19-ctx1024",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 9.8 GB at this context; duplicate traces of the same session: 0wgb7r0h, 2e96vnx9, 61qnb8mj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2337.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2-12B-A2.5B-Instruct-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 1k 2,337 103.5 9.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qfb6bv19",
   "source_url": "https://omlx.ai/benchmarks/performance/qfb6bv19",
   "tg_tokens": null,
   "tps": 103.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qfb6bv19-ctx4096",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 9.9 GB at this context; duplicate traces of the same session: 0wgb7r0h, 2e96vnx9, 61qnb8mj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3138.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2-12B-A2.5B-Instruct-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 4k 3,138 97.9 9.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qfb6bv19",
   "source_url": "https://omlx.ai/benchmarks/performance/qfb6bv19",
   "tg_tokens": null,
   "tps": 97.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qfb6bv19-ctx8192",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 10.0 GB at this context; duplicate traces of the same session: 0wgb7r0h, 2e96vnx9, 61qnb8mj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3299.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2-12B-A2.5B-Instruct-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 8k 3,299 91.4 10.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qfb6bv19",
   "source_url": "https://omlx.ai/benchmarks/performance/qfb6bv19",
   "tg_tokens": null,
   "tps": 91.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qfb6bv19-ctx16384",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 10.1 GB at this context; duplicate traces of the same session: 0wgb7r0h, 2e96vnx9, 61qnb8mj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3148.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2-12B-A2.5B-Instruct-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 16k 3,148 81.3 10.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qfb6bv19",
   "source_url": "https://omlx.ai/benchmarks/performance/qfb6bv19",
   "tg_tokens": null,
   "tps": 81.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qfb6bv19-ctx32768",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 10.3 GB at this context; duplicate traces of the same session: 0wgb7r0h, 2e96vnx9, 61qnb8mj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2870.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2-12B-A2.5B-Instruct-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 32k 2,870 66.5 10.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qfb6bv19",
   "source_url": "https://omlx.ai/benchmarks/performance/qfb6bv19",
   "tg_tokens": null,
   "tps": 66.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qfb6bv19-ctx65536",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 11.0 GB at this context; duplicate traces of the same session: 0wgb7r0h, 2e96vnx9, 61qnb8mj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2300.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2-12B-A2.5B-Instruct-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 64k 2,300 48.3 11.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qfb6bv19",
   "source_url": "https://omlx.ai/benchmarks/performance/qfb6bv19",
   "tg_tokens": null,
   "tps": 48.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qfb6bv19-ctx131072",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: 0wgb7r0h, 2e96vnx9, 61qnb8mj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1749.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2-12B-A2.5B-Instruct-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 128k 1,749 31.5 12.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qfb6bv19",
   "source_url": "https://omlx.ai/benchmarks/performance/qfb6bv19",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-ow27dhav-ctx4096",
   "model": "Qwen3.8-27B-TWIN-TURBO-f-c-f-709-l-unc",
   "notes": "abbreviated TWIN-TURBO-Fable-Cold-Fusion-709-L build; aura6 last4 bf16 vision MTP; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 144.6,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-TWIN-TURBO-f-c-f-709-l-unc-oQ5e-aura6-last4-bf16-vision-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 5bit; 4k 144.6 19.4 24.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ow27dhav",
   "source_url": "https://omlx.ai/benchmarks/performance/ow27dhav",
   "tg_tokens": null,
   "tps": 19.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-ow27dhav-ctx16384",
   "model": "Qwen3.8-27B-TWIN-TURBO-f-c-f-709-l-unc",
   "notes": "abbreviated TWIN-TURBO-Fable-Cold-Fusion-709-L build; aura6 last4 bf16 vision MTP; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 133.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-TWIN-TURBO-f-c-f-709-l-unc-oQ5e-aura6-last4-bf16-vision-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 5bit; 16k 133.1 18.0 26.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ow27dhav",
   "source_url": "https://omlx.ai/benchmarks/performance/ow27dhav",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8bkg7oi0-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: 1nur8y5c, xgcbqhuc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 954.7,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 954.7 57.0 20.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8bkg7oi0",
   "source_url": "https://omlx.ai/benchmarks/performance/8bkg7oi0",
   "tg_tokens": null,
   "tps": 57.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8bkg7oi0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 1nur8y5c, xgcbqhuc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1358.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,358 87.7 21.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8bkg7oi0",
   "source_url": "https://omlx.ai/benchmarks/performance/8bkg7oi0",
   "tg_tokens": null,
   "tps": 87.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8bkg7oi0-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 1nur8y5c, xgcbqhuc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1327.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,327 42.1 22.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8bkg7oi0",
   "source_url": "https://omlx.ai/benchmarks/performance/8bkg7oi0",
   "tg_tokens": null,
   "tps": 42.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8bkg7oi0-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 1nur8y5c, xgcbqhuc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1269.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,269 58.0 22.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8bkg7oi0",
   "source_url": "https://omlx.ai/benchmarks/performance/8bkg7oi0",
   "tg_tokens": null,
   "tps": 58.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-0kz7nf51-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace, single run at 32k context; peak memory 26.0 GB; TTFT 23754 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1379.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k tokens 1,379 PP tok/s 112.9 TG tok/s 23754 TTFT (ms) 26.0 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0kz7nf51",
   "source_url": "https://omlx.ai/benchmarks/performance/0kz7nf51",
   "tg_tokens": null,
   "tps": 112.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-mf9vwg1u-ctx1024",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 9.8 GB at this context; duplicate traces of the same session: auo0pvtl, qb4tgd8m; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2337.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2-12B-A2.5B-Instruct-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 1k 2,337 103.5 9.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mf9vwg1u",
   "source_url": "https://omlx.ai/benchmarks/performance/mf9vwg1u",
   "tg_tokens": null,
   "tps": 103.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-mf9vwg1u-ctx4096",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 9.9 GB at this context; duplicate traces of the same session: auo0pvtl, qb4tgd8m; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3138.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2-12B-A2.5B-Instruct-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 4k 3,138 97.9 9.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mf9vwg1u",
   "source_url": "https://omlx.ai/benchmarks/performance/mf9vwg1u",
   "tg_tokens": null,
   "tps": 97.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-mf9vwg1u-ctx8192",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 10.0 GB at this context; duplicate traces of the same session: auo0pvtl, qb4tgd8m; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3299.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2-12B-A2.5B-Instruct-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 8k 3,299 91.4 10.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mf9vwg1u",
   "source_url": "https://omlx.ai/benchmarks/performance/mf9vwg1u",
   "tg_tokens": null,
   "tps": 91.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-mf9vwg1u-ctx16384",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 10.1 GB at this context; duplicate traces of the same session: auo0pvtl, qb4tgd8m; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3148.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2-12B-A2.5B-Instruct-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 16k 3,148 81.3 10.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mf9vwg1u",
   "source_url": "https://omlx.ai/benchmarks/performance/mf9vwg1u",
   "tg_tokens": null,
   "tps": 81.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-mf9vwg1u-ctx32768",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 10.3 GB at this context; duplicate traces of the same session: auo0pvtl, qb4tgd8m; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2870.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2-12B-A2.5B-Instruct-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 32k 2,870 66.5 10.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mf9vwg1u",
   "source_url": "https://omlx.ai/benchmarks/performance/mf9vwg1u",
   "tg_tokens": null,
   "tps": 66.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-mf9vwg1u-ctx65536",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 11.0 GB at this context; duplicate traces of the same session: auo0pvtl, qb4tgd8m; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2300.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2-12B-A2.5B-Instruct-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 64k 2,300 48.3 11.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mf9vwg1u",
   "source_url": "https://omlx.ai/benchmarks/performance/mf9vwg1u",
   "tg_tokens": null,
   "tps": 48.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-mf9vwg1u-ctx131072",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: auo0pvtl, qb4tgd8m; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1749.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2-12B-A2.5B-Instruct-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 128k 1,749 31.5 12.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mf9vwg1u",
   "source_url": "https://omlx.ai/benchmarks/performance/mf9vwg1u",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-46lrrh9y-ctx4096",
   "model": "Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-uncher",
   "notes": "bf16 last4 8bit vision MTP build; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: 5fq1bjzr; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 147.6,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-uncher-oQ5e-6target-bf16-last4_8bit-vision-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 5bit; 4k 147.6 19.7 25.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/46lrrh9y",
   "source_url": "https://omlx.ai/benchmarks/performance/46lrrh9y",
   "tg_tokens": null,
   "tps": 19.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-46lrrh9y-ctx16384",
   "model": "Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-uncher",
   "notes": "bf16 last4 8bit vision MTP build; oMLX community benchmark trace; peak memory 27.2 GB at this context; duplicate traces of the same session: 5fq1bjzr; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 135.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-uncher-oQ5e-6target-bf16-last4_8bit-vision-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 5bit; 16k 135.9 18.2 27.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/46lrrh9y",
   "source_url": "https://omlx.ai/benchmarks/performance/46lrrh9y",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fteoybsr-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 91.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2573.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 4k 2,573 78.3 91.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fteoybsr",
   "source_url": "https://omlx.ai/benchmarks/performance/fteoybsr",
   "tg_tokens": null,
   "tps": 78.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fteoybsr-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2849.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 8k 2,849 93.9 92.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fteoybsr",
   "source_url": "https://omlx.ai/benchmarks/performance/fteoybsr",
   "tg_tokens": null,
   "tps": 93.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fteoybsr-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 93.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2668.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 16k 2,668 97.3 93.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fteoybsr",
   "source_url": "https://omlx.ai/benchmarks/performance/fteoybsr",
   "tg_tokens": null,
   "tps": 97.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fteoybsr-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 93.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2592.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 32k 2,592 65.7 93.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fteoybsr",
   "source_url": "https://omlx.ai/benchmarks/performance/fteoybsr",
   "tg_tokens": null,
   "tps": 65.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fteoybsr-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2472.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 64k 2,472 76.8 96.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fteoybsr",
   "source_url": "https://omlx.ai/benchmarks/performance/fteoybsr",
   "tg_tokens": null,
   "tps": 76.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fteoybsr-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2393.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 128k 2,393 69.4 100.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fteoybsr",
   "source_url": "https://omlx.ai/benchmarks/performance/fteoybsr",
   "tg_tokens": null,
   "tps": 69.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fteoybsr-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2307.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 195k 2,307 56.5 102.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fteoybsr",
   "source_url": "https://omlx.ai/benchmarks/performance/fteoybsr",
   "tg_tokens": null,
   "tps": 56.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-6m62infm-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 91.7 GB at this context; duplicate traces of the same session: 1xy6k40l, 7mpb7z8r, ki6ngn53; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2573.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 4k 2,573 78.3 91.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6m62infm",
   "source_url": "https://omlx.ai/benchmarks/performance/6m62infm",
   "tg_tokens": null,
   "tps": 78.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-6m62infm-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.9 GB at this context; duplicate traces of the same session: 1xy6k40l, 7mpb7z8r, ki6ngn53; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2849.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 8k 2,849 93.9 92.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6m62infm",
   "source_url": "https://omlx.ai/benchmarks/performance/6m62infm",
   "tg_tokens": null,
   "tps": 93.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6m62infm-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 93.3 GB at this context; duplicate traces of the same session: 1xy6k40l, 7mpb7z8r, ki6ngn53; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2668.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 16k 2,668 97.3 93.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6m62infm",
   "source_url": "https://omlx.ai/benchmarks/performance/6m62infm",
   "tg_tokens": null,
   "tps": 97.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6m62infm-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 93.9 GB at this context; duplicate traces of the same session: 1xy6k40l, 7mpb7z8r, ki6ngn53; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2592.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 32k 2,592 65.7 93.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6m62infm",
   "source_url": "https://omlx.ai/benchmarks/performance/6m62infm",
   "tg_tokens": null,
   "tps": 65.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-6m62infm-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.3 GB at this context; duplicate traces of the same session: 1xy6k40l, 7mpb7z8r, ki6ngn53; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2472.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 64k 2,472 76.8 96.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6m62infm",
   "source_url": "https://omlx.ai/benchmarks/performance/6m62infm",
   "tg_tokens": null,
   "tps": 76.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-6m62infm-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: 1xy6k40l, 7mpb7z8r, ki6ngn53; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2393.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 128k 2,393 69.4 100.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6m62infm",
   "source_url": "https://omlx.ai/benchmarks/performance/6m62infm",
   "tg_tokens": null,
   "tps": 69.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-6m62infm-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.4 GB at this context; duplicate traces of the same session: 1xy6k40l, 7mpb7z8r, ki6ngn53; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2307.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 195k 2,307 56.5 102.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6m62infm",
   "source_url": "https://omlx.ai/benchmarks/performance/6m62infm",
   "tg_tokens": null,
   "tps": 56.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-tvgu2e9c-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 91.7 GB at this context; duplicate traces of the same session: hlrcd4mx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2573.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 4k 2,573 78.3 91.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tvgu2e9c",
   "source_url": "https://omlx.ai/benchmarks/performance/tvgu2e9c",
   "tg_tokens": null,
   "tps": 78.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-tvgu2e9c-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.9 GB at this context; duplicate traces of the same session: hlrcd4mx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2849.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 8k 2,849 93.9 92.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tvgu2e9c",
   "source_url": "https://omlx.ai/benchmarks/performance/tvgu2e9c",
   "tg_tokens": null,
   "tps": 93.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tvgu2e9c-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 93.3 GB at this context; duplicate traces of the same session: hlrcd4mx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2668.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 16k 2,668 97.3 93.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tvgu2e9c",
   "source_url": "https://omlx.ai/benchmarks/performance/tvgu2e9c",
   "tg_tokens": null,
   "tps": 97.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tvgu2e9c-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 93.9 GB at this context; duplicate traces of the same session: hlrcd4mx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2592.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 32k 2,592 65.7 93.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tvgu2e9c",
   "source_url": "https://omlx.ai/benchmarks/performance/tvgu2e9c",
   "tg_tokens": null,
   "tps": 65.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-tvgu2e9c-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.3 GB at this context; duplicate traces of the same session: hlrcd4mx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2472.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 64k 2,472 76.8 96.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tvgu2e9c",
   "source_url": "https://omlx.ai/benchmarks/performance/tvgu2e9c",
   "tg_tokens": null,
   "tps": 76.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-tvgu2e9c-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: hlrcd4mx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2393.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 128k 2,393 69.4 100.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tvgu2e9c",
   "source_url": "https://omlx.ai/benchmarks/performance/tvgu2e9c",
   "tg_tokens": null,
   "tps": 69.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-tvgu2e9c-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.4 GB at this context; duplicate traces of the same session: hlrcd4mx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2307.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 195k 2,307 56.5 102.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tvgu2e9c",
   "source_url": "https://omlx.ai/benchmarks/performance/tvgu2e9c",
   "tg_tokens": null,
   "tps": 56.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0xkmnmdr-ctx1024",
   "model": "Coder390",
   "notes": "abbreviated Coder390 title; fp16 MTP build; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: 2t7fi379, tdx3yiow, zx2n9bnr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 939.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Coder390-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 939.2 43.2 20.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0xkmnmdr",
   "source_url": "https://omlx.ai/benchmarks/performance/0xkmnmdr",
   "tg_tokens": null,
   "tps": 43.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0xkmnmdr-ctx4096",
   "model": "Coder390",
   "notes": "abbreviated Coder390 title; fp16 MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 2t7fi379, tdx3yiow, zx2n9bnr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 816.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Coder390-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 816.3 44.6 22.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0xkmnmdr",
   "source_url": "https://omlx.ai/benchmarks/performance/0xkmnmdr",
   "tg_tokens": null,
   "tps": 44.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0xkmnmdr-ctx8192",
   "model": "Coder390",
   "notes": "abbreviated Coder390 title; fp16 MTP build; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: 2t7fi379, tdx3yiow, zx2n9bnr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 753.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Coder390-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 753.5 40.2 23.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0xkmnmdr",
   "source_url": "https://omlx.ai/benchmarks/performance/0xkmnmdr",
   "tg_tokens": null,
   "tps": 40.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0xkmnmdr-ctx16384",
   "model": "Coder390",
   "notes": "abbreviated Coder390 title; fp16 MTP build; oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: 2t7fi379, tdx3yiow, zx2n9bnr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 718.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Coder390-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 718.8 49.9 24.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0xkmnmdr",
   "source_url": "https://omlx.ai/benchmarks/performance/0xkmnmdr",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0xkmnmdr-ctx32768",
   "model": "Coder390",
   "notes": "abbreviated Coder390 title; fp16 MTP build; oMLX community benchmark trace; peak memory 27.6 GB at this context; duplicate traces of the same session: 2t7fi379, tdx3yiow, zx2n9bnr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 668.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Coder390-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 668.1 25.1 27.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0xkmnmdr",
   "source_url": "https://omlx.ai/benchmarks/performance/0xkmnmdr",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0xkmnmdr-ctx65536",
   "model": "Coder390",
   "notes": "abbreviated Coder390 title; fp16 MTP build; oMLX community benchmark trace; peak memory 32.9 GB at this context; duplicate traces of the same session: 2t7fi379, tdx3yiow, zx2n9bnr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 529.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Coder390-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 529.5 14.1 32.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0xkmnmdr",
   "source_url": "https://omlx.ai/benchmarks/performance/0xkmnmdr",
   "tg_tokens": null,
   "tps": 14.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ak1yxkq6-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 151.1 GB at this context; duplicate traces of the same session: 0n0yc138, 3wplfsjj, 7tbuna4x; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 390.7,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "GLM-5.3-Flash-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 1k 390.7 55.4 151.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ak1yxkq6",
   "source_url": "https://omlx.ai/benchmarks/performance/ak1yxkq6",
   "tg_tokens": null,
   "tps": 55.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ak1yxkq6-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 152.1 GB at this context; duplicate traces of the same session: 0n0yc138, 3wplfsjj, 7tbuna4x; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 452.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "GLM-5.3-Flash-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 4k 452.0 51.2 152.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ak1yxkq6",
   "source_url": "https://omlx.ai/benchmarks/performance/ak1yxkq6",
   "tg_tokens": null,
   "tps": 51.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ak1yxkq6-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 152.4 GB at this context; duplicate traces of the same session: 0n0yc138, 3wplfsjj, 7tbuna4x; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 416.5,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "GLM-5.3-Flash-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 8k 416.5 58.1 152.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ak1yxkq6",
   "source_url": "https://omlx.ai/benchmarks/performance/ak1yxkq6",
   "tg_tokens": null,
   "tps": 58.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ak1yxkq6-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 152.6 GB at this context; duplicate traces of the same session: 0n0yc138, 3wplfsjj, 7tbuna4x; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 409.6,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "GLM-5.3-Flash-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 16k 409.6 83.3 152.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ak1yxkq6",
   "source_url": "https://omlx.ai/benchmarks/performance/ak1yxkq6",
   "tg_tokens": null,
   "tps": 83.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rdq4waxk-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.7 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 406.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 406.7 41.9 173.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rdq4waxk",
   "source_url": "https://omlx.ai/benchmarks/performance/rdq4waxk",
   "tg_tokens": null,
   "tps": 41.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rdq4waxk-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.8 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 460.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 460.4 40.0 174.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rdq4waxk",
   "source_url": "https://omlx.ai/benchmarks/performance/rdq4waxk",
   "tg_tokens": null,
   "tps": 40.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rdq4waxk-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.1 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 423.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 423.4 40.2 175.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rdq4waxk",
   "source_url": "https://omlx.ai/benchmarks/performance/rdq4waxk",
   "tg_tokens": null,
   "tps": 40.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rdq4waxk-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.2 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 416.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 416.4 39.6 175.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rdq4waxk",
   "source_url": "https://omlx.ai/benchmarks/performance/rdq4waxk",
   "tg_tokens": null,
   "tps": 39.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-ghn5uoc3-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: psiiy4sv; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 176.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 176.4 22.1 5.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ghn5uoc3",
   "source_url": "https://omlx.ai/benchmarks/performance/ghn5uoc3",
   "tg_tokens": null,
   "tps": 22.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-ghn5uoc3-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.5 GB at this context; duplicate traces of the same session: psiiy4sv; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 180.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 180.9 21.0 7.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ghn5uoc3",
   "source_url": "https://omlx.ai/benchmarks/performance/ghn5uoc3",
   "tg_tokens": null,
   "tps": 21.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-rk7oiip0-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: exh1auen; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 147.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 147.0 19.1 5.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rk7oiip0",
   "source_url": "https://omlx.ai/benchmarks/performance/rk7oiip0",
   "tg_tokens": null,
   "tps": 19.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-rk7oiip0-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.5 GB at this context; duplicate traces of the same session: exh1auen; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 191.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 191.4 19.7 7.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rk7oiip0",
   "source_url": "https://omlx.ai/benchmarks/performance/rk7oiip0",
   "tg_tokens": null,
   "tps": 19.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-hu7ymlva-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: vnejvjfe; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 180.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 180.4 24.5 5.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hu7ymlva",
   "source_url": "https://omlx.ai/benchmarks/performance/hu7ymlva",
   "tg_tokens": null,
   "tps": 24.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-hu7ymlva-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.5 GB at this context; duplicate traces of the same session: vnejvjfe; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 200.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 200.8 23.2 7.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hu7ymlva",
   "source_url": "https://omlx.ai/benchmarks/performance/hu7ymlva",
   "tg_tokens": null,
   "tps": 23.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-s03x66ch-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: pgb9mrcm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1896.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,896 146.8 22.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s03x66ch",
   "source_url": "https://omlx.ai/benchmarks/performance/s03x66ch",
   "tg_tokens": null,
   "tps": 146.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-s03x66ch-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: pgb9mrcm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1367.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 1,367 165.3 22.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s03x66ch",
   "source_url": "https://omlx.ai/benchmarks/performance/s03x66ch",
   "tg_tokens": null,
   "tps": 165.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-800frysn-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: ntdmryys; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 985.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 985.3 94.1 22.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/800frysn",
   "source_url": "https://omlx.ai/benchmarks/performance/800frysn",
   "tg_tokens": null,
   "tps": 94.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-800frysn-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: ntdmryys; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 923.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 923.8 94.4 22.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/800frysn",
   "source_url": "https://omlx.ai/benchmarks/performance/800frysn",
   "tg_tokens": null,
   "tps": 94.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-800frysn-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: ntdmryys; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 829.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 829.9 98.9 22.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/800frysn",
   "source_url": "https://omlx.ai/benchmarks/performance/800frysn",
   "tg_tokens": null,
   "tps": 98.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-800frysn-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: ntdmryys; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 679.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 679.6 84.6 23.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/800frysn",
   "source_url": "https://omlx.ai/benchmarks/performance/800frysn",
   "tg_tokens": null,
   "tps": 84.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-800frysn-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: ntdmryys; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 491.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 491.2 65.3 25.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/800frysn",
   "source_url": "https://omlx.ai/benchmarks/performance/800frysn",
   "tg_tokens": null,
   "tps": 65.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-dbg4kmfg-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: ui5jp7sx; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 209.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 209.3 22.4 16.0 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dbg4kmfg",
   "source_url": "https://omlx.ai/benchmarks/performance/dbg4kmfg",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-dbg4kmfg-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: ui5jp7sx; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 215.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 215.9 21.8 17.0 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dbg4kmfg",
   "source_url": "https://omlx.ai/benchmarks/performance/dbg4kmfg",
   "tg_tokens": null,
   "tps": 21.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-v32jxcbg-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: br08ympg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 239.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 239.8 25.9 17.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v32jxcbg",
   "source_url": "https://omlx.ai/benchmarks/performance/v32jxcbg",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-v32jxcbg-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.9 GB at this context; duplicate traces of the same session: br08ympg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 211.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 211.0 24.8 17.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v32jxcbg",
   "source_url": "https://omlx.ai/benchmarks/performance/v32jxcbg",
   "tg_tokens": null,
   "tps": 24.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-dprzuno3-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 8k context; peak memory 17.4 GB; TTFT 76710 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 8k tokens 106.8 PP tok/s 14.9 TG tok/s 76710 TTFT (ms) 17.4 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dprzuno3",
   "source_url": "https://omlx.ai/benchmarks/performance/dprzuno3",
   "tg_tokens": null,
   "tps": 14.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-qkp9oock-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 16.0 GB; TTFT 9653 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k tokens 106.1 PP tok/s 15.8 TG tok/s 9653 TTFT (ms) 16.0 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qkp9oock",
   "source_url": "https://omlx.ai/benchmarks/performance/qkp9oock",
   "tg_tokens": null,
   "tps": 15.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-mewvrkb3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 17.0 GB; TTFT 37762 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 108.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k tokens 108.5 PP tok/s 15.4 TG tok/s 37762 TTFT (ms) 17.0 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mewvrkb3",
   "source_url": "https://omlx.ai/benchmarks/performance/mewvrkb3",
   "tg_tokens": null,
   "tps": 15.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-gjt16mge-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: f04ai98f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 264.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 264.1 28.2 22.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gjt16mge",
   "source_url": "https://omlx.ai/benchmarks/performance/gjt16mge",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-gjt16mge-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: f04ai98f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 234.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 234.6 33.9 25.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gjt16mge",
   "source_url": "https://omlx.ai/benchmarks/performance/gjt16mge",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ow250fl0-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 29.9 GB; TTFT 1630414 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 80.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 128k tokens 80.4 PP tok/s 20.9 TG tok/s 1630414 TTFT (ms) 29.9 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ow250fl0",
   "source_url": "https://omlx.ai/benchmarks/performance/ow250fl0",
   "tg_tokens": null,
   "tps": 20.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-pvsjm3zg-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 84.7 GB at this context; duplicate traces of the same session: 4j29t2y0, f7lusiot, j99nuhm1, mud4eay0, u328rr4b, wxd4fvky; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 581.5,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 581.5 60.2 84.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pvsjm3zg",
   "source_url": "https://omlx.ai/benchmarks/performance/pvsjm3zg",
   "tg_tokens": null,
   "tps": 60.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-pvsjm3zg-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 85.2 GB at this context; duplicate traces of the same session: 4j29t2y0, f7lusiot, j99nuhm1, mud4eay0, u328rr4b, wxd4fvky; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 624.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 624.9 54.1 85.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pvsjm3zg",
   "source_url": "https://omlx.ai/benchmarks/performance/pvsjm3zg",
   "tg_tokens": null,
   "tps": 54.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-pvsjm3zg-ctx8192",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 85.3 GB at this context; duplicate traces of the same session: 4j29t2y0, f7lusiot, j99nuhm1, mud4eay0, u328rr4b, wxd4fvky; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 620.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 620.0 53.5 85.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pvsjm3zg",
   "source_url": "https://omlx.ai/benchmarks/performance/pvsjm3zg",
   "tg_tokens": null,
   "tps": 53.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-pvsjm3zg-ctx32768",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 86.0 GB at this context; duplicate traces of the same session: 4j29t2y0, f7lusiot, j99nuhm1, mud4eay0, u328rr4b, wxd4fvky; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 572.8,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 572.8 48.1 86.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pvsjm3zg",
   "source_url": "https://omlx.ai/benchmarks/performance/pvsjm3zg",
   "tg_tokens": null,
   "tps": 48.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-pvsjm3zg-ctx65536",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 87.1 GB at this context; duplicate traces of the same session: 4j29t2y0, f7lusiot, j99nuhm1, mud4eay0, u328rr4b, wxd4fvky; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 381.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 381.4 42.2 87.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pvsjm3zg",
   "source_url": "https://omlx.ai/benchmarks/performance/pvsjm3zg",
   "tg_tokens": null,
   "tps": 42.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-pvsjm3zg-ctx131072",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 90.3 GB at this context; duplicate traces of the same session: 4j29t2y0, f7lusiot, j99nuhm1, mud4eay0, u328rr4b, wxd4fvky; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 429.8,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 429.8 32.1 90.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pvsjm3zg",
   "source_url": "https://omlx.ai/benchmarks/performance/pvsjm3zg",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-pvsjm3zg-ctx199680",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 94.3 GB at this context; duplicate traces of the same session: 4j29t2y0, f7lusiot, j99nuhm1, mud4eay0, u328rr4b, wxd4fvky; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 413.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 5bit; 195k 413.4 30.4 94.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pvsjm3zg",
   "source_url": "https://omlx.ai/benchmarks/performance/pvsjm3zg",
   "tg_tokens": null,
   "tps": 30.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8v6x0xsq-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 77.6 GB; TTFT 896 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1143.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k tokens 1,143 PP tok/s 83.9 TG tok/s 896 TTFT (ms) 77.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8v6x0xsq",
   "source_url": "https://omlx.ai/benchmarks/performance/8v6x0xsq",
   "tg_tokens": null,
   "tps": 83.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-1ey92nt2-ctx4096",
   "model": "Schattenblume-31B",
   "notes": "oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: jstz1j86; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 126.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Schattenblume-31B-oQ5e on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 5bit; 4k 126.7 14.0 22.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ey92nt2",
   "source_url": "https://omlx.ai/benchmarks/performance/1ey92nt2",
   "tg_tokens": null,
   "tps": 14.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-1ey92nt2-ctx16384",
   "model": "Schattenblume-31B",
   "notes": "oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: jstz1j86; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 123.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Schattenblume-31B-oQ5e on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 5bit; 16k 123.4 13.0 24.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ey92nt2",
   "source_url": "https://omlx.ai/benchmarks/performance/1ey92nt2",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-cnx051k6-ctx4096",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: ze435o00; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 120.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 120.6 15.8 20.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cnx051k6",
   "source_url": "https://omlx.ai/benchmarks/performance/cnx051k6",
   "tg_tokens": null,
   "tps": 15.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-cnx051k6-ctx16384",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: ze435o00; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 113.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k 113.1 15.2 22.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cnx051k6",
   "source_url": "https://omlx.ai/benchmarks/performance/cnx051k6",
   "tg_tokens": null,
   "tps": 15.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-pggelkr4-ctx4096",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 2t4h7tg2; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 117.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 117.9 12.6 23.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pggelkr4",
   "source_url": "https://omlx.ai/benchmarks/performance/pggelkr4",
   "tg_tokens": null,
   "tps": 12.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-pggelkr4-ctx16384",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: 2t4h7tg2; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 112.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k 112.5 22.2 25.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pggelkr4",
   "source_url": "https://omlx.ai/benchmarks/performance/pggelkr4",
   "tg_tokens": null,
   "tps": 22.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-u82xxlwz-ctx4096",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: wmz5stx1; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 112.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 112.2 13.1 20.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u82xxlwz",
   "source_url": "https://omlx.ai/benchmarks/performance/u82xxlwz",
   "tg_tokens": null,
   "tps": 13.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-u82xxlwz-ctx16384",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: wmz5stx1; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k 106.9 12.3 22.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u82xxlwz",
   "source_url": "https://omlx.ai/benchmarks/performance/u82xxlwz",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-8r2hmh1s-ctx4096",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: qpn784rf; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 125.2 17.5 23.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8r2hmh1s",
   "source_url": "https://omlx.ai/benchmarks/performance/8r2hmh1s",
   "tg_tokens": null,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-8r2hmh1s-ctx16384",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: qpn784rf; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 116.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k 116.7 14.5 25.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8r2hmh1s",
   "source_url": "https://omlx.ai/benchmarks/performance/8r2hmh1s",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-2lfuu5r1-ctx4096",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: s9l3c24z; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 135.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 135.9 16.7 25.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2lfuu5r1",
   "source_url": "https://omlx.ai/benchmarks/performance/2lfuu5r1",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-2lfuu5r1-ctx16384",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: s9l3c24z; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 120.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k 120.2 18.2 25.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2lfuu5r1",
   "source_url": "https://omlx.ai/benchmarks/performance/2lfuu5r1",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-4mj91ugu-ctx4096",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: s3whvof1; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 138.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 138.9 17.5 23.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4mj91ugu",
   "source_url": "https://omlx.ai/benchmarks/performance/4mj91ugu",
   "tg_tokens": null,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-4mj91ugu-ctx16384",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: s3whvof1; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k 130.2 16.1 25.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4mj91ugu",
   "source_url": "https://omlx.ai/benchmarks/performance/4mj91ugu",
   "tg_tokens": null,
   "tps": 16.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-5vzfakeo-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift build; oQ4e/oQ6e mixed quant; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 77.5 GB at this context; duplicate traces of the same session: 4kmiovn7, hf2zprqx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 594.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 594.8 71.3 77.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5vzfakeo",
   "source_url": "https://omlx.ai/benchmarks/performance/5vzfakeo",
   "tg_tokens": null,
   "tps": 71.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-5vzfakeo-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift build; oQ4e/oQ6e mixed quant; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.1 GB at this context; duplicate traces of the same session: 4kmiovn7, hf2zprqx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 721.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 721.4 63.8 78.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5vzfakeo",
   "source_url": "https://omlx.ai/benchmarks/performance/5vzfakeo",
   "tg_tokens": null,
   "tps": 63.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-5vzfakeo-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift build; oQ4e/oQ6e mixed quant; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.3 GB at this context; duplicate traces of the same session: 4kmiovn7, hf2zprqx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 828.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 828.6 66.1 78.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5vzfakeo",
   "source_url": "https://omlx.ai/benchmarks/performance/5vzfakeo",
   "tg_tokens": null,
   "tps": 66.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-5tik17ks-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift build; oQ4e/oQ6e mixed quant; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 77.5 GB at this context; duplicate traces of the same session: 2a1w2a3k, cvdyt51w, x0ua7u8d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 542.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 542.6 46.2 77.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5tik17ks",
   "source_url": "https://omlx.ai/benchmarks/performance/5tik17ks",
   "tg_tokens": null,
   "tps": 46.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-5tik17ks-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift build; oQ4e/oQ6e mixed quant; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.1 GB at this context; duplicate traces of the same session: 2a1w2a3k, cvdyt51w, x0ua7u8d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 675.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 675.6 47.4 78.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5tik17ks",
   "source_url": "https://omlx.ai/benchmarks/performance/5tik17ks",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-5tik17ks-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift build; oQ4e/oQ6e mixed quant; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.6 GB at this context; duplicate traces of the same session: 2a1w2a3k, cvdyt51w, x0ua7u8d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 822.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 822.3 59.6 78.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5tik17ks",
   "source_url": "https://omlx.ai/benchmarks/performance/5tik17ks",
   "tg_tokens": null,
   "tps": 59.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-5tik17ks-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift build; oQ4e/oQ6e mixed quant; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.2 GB at this context; duplicate traces of the same session: 2a1w2a3k, cvdyt51w, x0ua7u8d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 815.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 815.8 60.3 79.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5tik17ks",
   "source_url": "https://omlx.ai/benchmarks/performance/5tik17ks",
   "tg_tokens": null,
   "tps": 60.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-8s907lqw-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift build; oQ4e/oQ6e mixed quant; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 77.5 GB at this context; duplicate traces of the same session: xmah3c3s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 596.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 596.6 75.0 77.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8s907lqw",
   "source_url": "https://omlx.ai/benchmarks/performance/8s907lqw",
   "tg_tokens": null,
   "tps": 75.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-8s907lqw-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift build; oQ4e/oQ6e mixed quant; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.3 GB at this context; duplicate traces of the same session: xmah3c3s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 827.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 827.2 63.6 78.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8s907lqw",
   "source_url": "https://omlx.ai/benchmarks/performance/8s907lqw",
   "tg_tokens": null,
   "tps": 63.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-d105qeu1-ctx1024",
   "model": "Coder390",
   "notes": "abbreviated Coder390 title; fp16 MTP build; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: os817z5b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 939.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Coder390-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 939.2 43.2 20.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d105qeu1",
   "source_url": "https://omlx.ai/benchmarks/performance/d105qeu1",
   "tg_tokens": null,
   "tps": 43.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-d105qeu1-ctx4096",
   "model": "Coder390",
   "notes": "abbreviated Coder390 title; fp16 MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: os817z5b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 816.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Coder390-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 816.3 44.6 22.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d105qeu1",
   "source_url": "https://omlx.ai/benchmarks/performance/d105qeu1",
   "tg_tokens": null,
   "tps": 44.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-d105qeu1-ctx8192",
   "model": "Coder390",
   "notes": "abbreviated Coder390 title; fp16 MTP build; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: os817z5b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 753.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Coder390-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 753.5 40.2 23.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d105qeu1",
   "source_url": "https://omlx.ai/benchmarks/performance/d105qeu1",
   "tg_tokens": null,
   "tps": 40.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-d105qeu1-ctx16384",
   "model": "Coder390",
   "notes": "abbreviated Coder390 title; fp16 MTP build; oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: os817z5b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 718.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Coder390-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 718.8 49.9 24.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d105qeu1",
   "source_url": "https://omlx.ai/benchmarks/performance/d105qeu1",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-d105qeu1-ctx32768",
   "model": "Coder390",
   "notes": "abbreviated Coder390 title; fp16 MTP build; oMLX community benchmark trace; peak memory 27.6 GB at this context; duplicate traces of the same session: os817z5b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 668.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Coder390-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 668.1 25.1 27.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d105qeu1",
   "source_url": "https://omlx.ai/benchmarks/performance/d105qeu1",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-d105qeu1-ctx65536",
   "model": "Coder390",
   "notes": "abbreviated Coder390 title; fp16 MTP build; oMLX community benchmark trace; peak memory 32.9 GB at this context; duplicate traces of the same session: os817z5b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 529.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Coder390-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 529.5 14.1 32.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d105qeu1",
   "source_url": "https://omlx.ai/benchmarks/performance/d105qeu1",
   "tg_tokens": null,
   "tps": 14.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-km43np2r-ctx1024",
   "model": "Coder390",
   "notes": "abbreviated Coder390 title; fp16 MTP build; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 9xl7tco4, l3zg1xx6, oky6l49v, oui9n0tj, x9ew3wnu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 717.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Coder390-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 717.6 40.2 19.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/km43np2r",
   "source_url": "https://omlx.ai/benchmarks/performance/km43np2r",
   "tg_tokens": null,
   "tps": 40.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-km43np2r-ctx4096",
   "model": "Coder390",
   "notes": "abbreviated Coder390 title; fp16 MTP build; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 9xl7tco4, l3zg1xx6, oky6l49v, oui9n0tj, x9ew3wnu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 872.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Coder390-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 872.8 51.9 21.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/km43np2r",
   "source_url": "https://omlx.ai/benchmarks/performance/km43np2r",
   "tg_tokens": null,
   "tps": 51.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-km43np2r-ctx8192",
   "model": "Coder390",
   "notes": "abbreviated Coder390 title; fp16 MTP build; oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: 9xl7tco4, l3zg1xx6, oky6l49v, oui9n0tj, x9ew3wnu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 780.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Coder390-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 780.1 57.7 21.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/km43np2r",
   "source_url": "https://omlx.ai/benchmarks/performance/km43np2r",
   "tg_tokens": null,
   "tps": 57.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-km43np2r-ctx16384",
   "model": "Coder390",
   "notes": "abbreviated Coder390 title; fp16 MTP build; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: 9xl7tco4, l3zg1xx6, oky6l49v, oui9n0tj, x9ew3wnu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 739.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Coder390-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 739.5 57.3 22.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/km43np2r",
   "source_url": "https://omlx.ai/benchmarks/performance/km43np2r",
   "tg_tokens": null,
   "tps": 57.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-km43np2r-ctx32768",
   "model": "Coder390",
   "notes": "abbreviated Coder390 title; fp16 MTP build; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: 9xl7tco4, l3zg1xx6, oky6l49v, oui9n0tj, x9ew3wnu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 687.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Coder390-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 687.5 114.2 25.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/km43np2r",
   "source_url": "https://omlx.ai/benchmarks/performance/km43np2r",
   "tg_tokens": null,
   "tps": 114.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-km43np2r-ctx65536",
   "model": "Coder390",
   "notes": "abbreviated Coder390 title; fp16 MTP build; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: 9xl7tco4, l3zg1xx6, oky6l49v, oui9n0tj, x9ew3wnu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 605.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Coder390-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 605.0 51.0 29.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/km43np2r",
   "source_url": "https://omlx.ai/benchmarks/performance/km43np2r",
   "tg_tokens": null,
   "tps": 51.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-idfqoluf-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 172.6 GB at this context; duplicate traces of the same session: iyqdvh89; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 397.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 397.9 48.6 172.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/idfqoluf",
   "source_url": "https://omlx.ai/benchmarks/performance/idfqoluf",
   "tg_tokens": null,
   "tps": 48.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-idfqoluf-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 173.7 GB at this context; duplicate traces of the same session: iyqdvh89; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 445.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 445.6 44.7 173.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/idfqoluf",
   "source_url": "https://omlx.ai/benchmarks/performance/idfqoluf",
   "tg_tokens": null,
   "tps": 44.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-a1he7ypn-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 172.4 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1281.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,281 78.5 172.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a1he7ypn",
   "source_url": "https://omlx.ai/benchmarks/performance/a1he7ypn",
   "tg_tokens": null,
   "tps": 78.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-a1he7ypn-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 173.2 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1923.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,923 68.5 173.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a1he7ypn",
   "source_url": "https://omlx.ai/benchmarks/performance/a1he7ypn",
   "tg_tokens": null,
   "tps": 68.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-a1he7ypn-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 173.8 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1924.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 1,924 75.8 173.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a1he7ypn",
   "source_url": "https://omlx.ai/benchmarks/performance/a1he7ypn",
   "tg_tokens": null,
   "tps": 75.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-a1he7ypn-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 174.0 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1913.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,913 150.3 174.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a1he7ypn",
   "source_url": "https://omlx.ai/benchmarks/performance/a1he7ypn",
   "tg_tokens": null,
   "tps": 150.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-a1he7ypn-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 174.4 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1901.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,901 95.6 174.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a1he7ypn",
   "source_url": "https://omlx.ai/benchmarks/performance/a1he7ypn",
   "tg_tokens": null,
   "tps": 95.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-a1he7ypn-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 175.0 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1878.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,878 82.5 175.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a1he7ypn",
   "source_url": "https://omlx.ai/benchmarks/performance/a1he7ypn",
   "tg_tokens": null,
   "tps": 82.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-a1he7ypn-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 176.3 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1839.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 1,839 144.0 176.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a1he7ypn",
   "source_url": "https://omlx.ai/benchmarks/performance/a1he7ypn",
   "tg_tokens": null,
   "tps": 144.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-a1he7ypn-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.5 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1812.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 1,812 73.6 177.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a1he7ypn",
   "source_url": "https://omlx.ai/benchmarks/performance/a1he7ypn",
   "tg_tokens": null,
   "tps": 73.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ist34xq6-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.7 GB at this context; duplicate traces of the same session: 9k6nr2ih, lfyt6jmu; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 406.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 406.7 41.9 173.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ist34xq6",
   "source_url": "https://omlx.ai/benchmarks/performance/ist34xq6",
   "tg_tokens": null,
   "tps": 41.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ist34xq6-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.8 GB at this context; duplicate traces of the same session: 9k6nr2ih, lfyt6jmu; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 460.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 460.4 40.0 174.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ist34xq6",
   "source_url": "https://omlx.ai/benchmarks/performance/ist34xq6",
   "tg_tokens": null,
   "tps": 40.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ist34xq6-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.1 GB at this context; duplicate traces of the same session: 9k6nr2ih, lfyt6jmu; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 423.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 423.4 40.2 175.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ist34xq6",
   "source_url": "https://omlx.ai/benchmarks/performance/ist34xq6",
   "tg_tokens": null,
   "tps": 40.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ist34xq6-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.2 GB at this context; duplicate traces of the same session: 9k6nr2ih, lfyt6jmu; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 416.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 416.4 39.6 175.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ist34xq6",
   "source_url": "https://omlx.ai/benchmarks/performance/ist34xq6",
   "tg_tokens": null,
   "tps": 39.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-zq2y82p6-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.4 GB at this context; duplicate traces of the same session: v2gncbm8; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 821.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 821.7 56.2 178.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zq2y82p6",
   "source_url": "https://omlx.ai/benchmarks/performance/zq2y82p6",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-zq2y82p6-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.3 GB at this context; duplicate traces of the same session: v2gncbm8; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1911.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,911 56.5 179.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zq2y82p6",
   "source_url": "https://omlx.ai/benchmarks/performance/zq2y82p6",
   "tg_tokens": null,
   "tps": 56.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-6ptg5kfu-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.4 GB at this context; duplicate traces of the same session: u1mnbnwx; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1847.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,847 48.4 178.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ptg5kfu",
   "source_url": "https://omlx.ai/benchmarks/performance/6ptg5kfu",
   "tg_tokens": null,
   "tps": 48.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-6ptg5kfu-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.1 GB at this context; duplicate traces of the same session: u1mnbnwx; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1899.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,899 75.4 179.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ptg5kfu",
   "source_url": "https://omlx.ai/benchmarks/performance/6ptg5kfu",
   "tg_tokens": null,
   "tps": 75.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ui5nb1mf-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.4 GB at this context; duplicate traces of the same session: jcuyw4f7; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1885.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,885 61.6 178.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ui5nb1mf",
   "source_url": "https://omlx.ai/benchmarks/performance/ui5nb1mf",
   "tg_tokens": null,
   "tps": 61.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ui5nb1mf-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.1 GB at this context; duplicate traces of the same session: jcuyw4f7; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1901.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,901 82.6 179.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ui5nb1mf",
   "source_url": "https://omlx.ai/benchmarks/performance/ui5nb1mf",
   "tg_tokens": null,
   "tps": 82.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-28jfbygq-ctx4096",
   "model": "Qwen3-4B-Instruct-2507",
   "notes": "oMLX community benchmark trace; peak memory 3.2 GB at this context; duplicate traces of the same session: iwzwj4mm; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 390.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-4B-Instruct-2507-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 390.7 31.7 3.2 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/28jfbygq",
   "source_url": "https://omlx.ai/benchmarks/performance/28jfbygq",
   "tg_tokens": null,
   "tps": 31.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-28jfbygq-ctx16384",
   "model": "Qwen3-4B-Instruct-2507",
   "notes": "oMLX community benchmark trace; peak memory 4.8 GB at this context; duplicate traces of the same session: iwzwj4mm; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 274.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-4B-Instruct-2507-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 16k 274.5 16.9 4.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/28jfbygq",
   "source_url": "https://omlx.ai/benchmarks/performance/28jfbygq",
   "tg_tokens": null,
   "tps": 16.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-m3hqljbv-ctx1024",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "MLX; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: suwnlfk1; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 504.6,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-MLX-5bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 5bit; 1k 504.6 45.5 20.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m3hqljbv",
   "source_url": "https://omlx.ai/benchmarks/performance/m3hqljbv",
   "tg_tokens": null,
   "tps": 45.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-m3hqljbv-ctx4096",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "MLX; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: suwnlfk1; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 504.4,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-MLX-5bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 5bit; 4k 504.4 40.0 20.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m3hqljbv",
   "source_url": "https://omlx.ai/benchmarks/performance/m3hqljbv",
   "tg_tokens": null,
   "tps": 40.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fejgxg0u-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: b76vrez2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2306.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 2,306 107.8 19.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fejgxg0u",
   "source_url": "https://omlx.ai/benchmarks/performance/fejgxg0u",
   "tg_tokens": null,
   "tps": 107.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fejgxg0u-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: b76vrez2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2945.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,945 103.9 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fejgxg0u",
   "source_url": "https://omlx.ai/benchmarks/performance/fejgxg0u",
   "tg_tokens": null,
   "tps": 103.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-1g93fljm-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: imajhw8l, qc60n0ay, t970v16q; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 507.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 507.3 49.5 21.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1g93fljm",
   "source_url": "https://omlx.ai/benchmarks/performance/1g93fljm",
   "tg_tokens": null,
   "tps": 49.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-1g93fljm-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: imajhw8l, qc60n0ay, t970v16q; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 820.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 820.7 66.5 23.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1g93fljm",
   "source_url": "https://omlx.ai/benchmarks/performance/1g93fljm",
   "tg_tokens": null,
   "tps": 66.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-1g93fljm-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: imajhw8l, qc60n0ay, t970v16q; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 696.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 696.9 19.7 24.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1g93fljm",
   "source_url": "https://omlx.ai/benchmarks/performance/1g93fljm",
   "tg_tokens": null,
   "tps": 19.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-1g93fljm-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: imajhw8l, qc60n0ay, t970v16q; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 646.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 646.0 63.3 23.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1g93fljm",
   "source_url": "https://omlx.ai/benchmarks/performance/1g93fljm",
   "tg_tokens": null,
   "tps": 63.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-1g93fljm-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: imajhw8l, qc60n0ay, t970v16q; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 538.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 32k 538.0 65.7 24.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1g93fljm",
   "source_url": "https://omlx.ai/benchmarks/performance/1g93fljm",
   "tg_tokens": null,
   "tps": 65.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-n3yjha03-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: 7t61pmjw, zmvwfh9l; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 985.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 985.3 94.1 22.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n3yjha03",
   "source_url": "https://omlx.ai/benchmarks/performance/n3yjha03",
   "tg_tokens": null,
   "tps": 94.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-n3yjha03-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: 7t61pmjw, zmvwfh9l; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 923.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 923.8 94.4 22.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n3yjha03",
   "source_url": "https://omlx.ai/benchmarks/performance/n3yjha03",
   "tg_tokens": null,
   "tps": 94.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-n3yjha03-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: 7t61pmjw, zmvwfh9l; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 829.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 829.9 98.9 22.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n3yjha03",
   "source_url": "https://omlx.ai/benchmarks/performance/n3yjha03",
   "tg_tokens": null,
   "tps": 98.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-n3yjha03-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 7t61pmjw, zmvwfh9l; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 679.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 679.6 84.6 23.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n3yjha03",
   "source_url": "https://omlx.ai/benchmarks/performance/n3yjha03",
   "tg_tokens": null,
   "tps": 84.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-n3yjha03-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: 7t61pmjw, zmvwfh9l; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 491.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 491.2 65.3 25.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n3yjha03",
   "source_url": "https://omlx.ai/benchmarks/performance/n3yjha03",
   "tg_tokens": null,
   "tps": 65.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-69x0pobb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 330.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 330.2 45.4 29.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/69x0pobb",
   "source_url": "https://omlx.ai/benchmarks/performance/69x0pobb",
   "tg_tokens": null,
   "tps": 45.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-69x0pobb-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 31.1 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 336.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 336.3 39.5 31.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/69x0pobb",
   "source_url": "https://omlx.ai/benchmarks/performance/69x0pobb",
   "tg_tokens": null,
   "tps": 39.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ytj2fd2g-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: hd49vn2v; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 407.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 407.1 18.0 16.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ytj2fd2g",
   "source_url": "https://omlx.ai/benchmarks/performance/ytj2fd2g",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ytj2fd2g-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: hd49vn2v; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 426.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 426.6 17.8 17.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ytj2fd2g",
   "source_url": "https://omlx.ai/benchmarks/performance/ytj2fd2g",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-wour3tq1-ctx1024",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "MLX; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 4jy2kpb9; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 276.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 276.9 20.3 21.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wour3tq1",
   "source_url": "https://omlx.ai/benchmarks/performance/wour3tq1",
   "tg_tokens": null,
   "tps": 20.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-wour3tq1-ctx4096",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "MLX; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: 4jy2kpb9; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 280.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 280.0 21.4 23.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wour3tq1",
   "source_url": "https://omlx.ai/benchmarks/performance/wour3tq1",
   "tg_tokens": null,
   "tps": 21.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-bmai21fm-ctx1024",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: zpb23rou; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1336.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 1,336 75.8 6.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bmai21fm",
   "source_url": "https://omlx.ai/benchmarks/performance/bmai21fm",
   "tg_tokens": null,
   "tps": 75.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-bmai21fm-ctx4096",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: zpb23rou; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1443.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 1,443 79.7 6.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bmai21fm",
   "source_url": "https://omlx.ai/benchmarks/performance/bmai21fm",
   "tg_tokens": null,
   "tps": 79.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-iihoasrn-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 100.7 GB; TTFT 893 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1147.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k tokens 1,147 PP tok/s 102.6 TG tok/s 893 TTFT (ms) 100.7 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iihoasrn",
   "source_url": "https://omlx.ai/benchmarks/performance/iihoasrn",
   "tg_tokens": null,
   "tps": 102.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5d93gbau-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 100.5 GB; TTFT 1294 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 791.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k tokens 791.1 PP tok/s 58.7 TG tok/s 1294 TTFT (ms) 100.5 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5d93gbau",
   "source_url": "https://omlx.ai/benchmarks/performance/5d93gbau",
   "tg_tokens": null,
   "tps": 58.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-zm78mxua-ctx16384",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace, single run at 16k context; peak memory 25.9 GB; TTFT 136531 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 120.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k tokens 120.0 PP tok/s 20.3 TG tok/s 136531 TTFT (ms) 25.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zm78mxua",
   "source_url": "https://omlx.ai/benchmarks/performance/zm78mxua",
   "tg_tokens": null,
   "tps": 20.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-lw6pi0v4-ctx16384",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace, single run at 16k context; peak memory 23.3 GB; TTFT 132457 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 123.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k tokens 123.7 PP tok/s 15.1 TG tok/s 132457 TTFT (ms) 23.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lw6pi0v4",
   "source_url": "https://omlx.ai/benchmarks/performance/lw6pi0v4",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-w4i6mbrx-ctx16384",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace, single run at 16k context; peak memory 22.5 GB; TTFT 152492 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 107.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k tokens 107.4 PP tok/s 13.1 TG tok/s 152492 TTFT (ms) 22.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w4i6mbrx",
   "source_url": "https://omlx.ai/benchmarks/performance/w4i6mbrx",
   "tg_tokens": null,
   "tps": 13.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-gx42x32f-ctx16384",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace, single run at 16k context; peak memory 25.1 GB; TTFT 144136 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 113.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k tokens 113.7 PP tok/s 20.0 TG tok/s 144136 TTFT (ms) 25.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gx42x32f",
   "source_url": "https://omlx.ai/benchmarks/performance/gx42x32f",
   "tg_tokens": null,
   "tps": 20.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-tc71w5cv-ctx4096",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 9vr9l2h5; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 122.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 122.3 17.3 23.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tc71w5cv",
   "source_url": "https://omlx.ai/benchmarks/performance/tc71w5cv",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-tc71w5cv-ctx16384",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: 9vr9l2h5; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 113.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k 113.7 19.6 25.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tc71w5cv",
   "source_url": "https://omlx.ai/benchmarks/performance/tc71w5cv",
   "tg_tokens": null,
   "tps": 19.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-2xplg02m-ctx1024",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: dz241qmx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 205.7 30.5 34.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2xplg02m",
   "source_url": "https://omlx.ai/benchmarks/performance/2xplg02m",
   "tg_tokens": null,
   "tps": 30.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-2xplg02m-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 35.0 GB at this context; duplicate traces of the same session: dz241qmx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 186.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 186.2 28.2 35.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2xplg02m",
   "source_url": "https://omlx.ai/benchmarks/performance/2xplg02m",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-qc5w2ugb-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 4.6 GB at this context; duplicate traces of the same session: 6rzgq9zz; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 568.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 568.8 32.2 4.6 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qc5w2ugb",
   "source_url": "https://omlx.ai/benchmarks/performance/qc5w2ugb",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-qc5w2ugb-ctx16384",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 4.8 GB at this context; duplicate traces of the same session: 6rzgq9zz; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 548.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 16k 548.9 28.8 4.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qc5w2ugb",
   "source_url": "https://omlx.ai/benchmarks/performance/qc5w2ugb",
   "tg_tokens": null,
   "tps": 28.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-namchm7s-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Unsloth UD build; oMLX community benchmark trace, single run at 16k context; peak memory 24.1 GB; TTFT 34677 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 472.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "unsloth--Qwen3.6-35B-A3B-UD-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k tokens 472.5 PP tok/s 48.1 TG tok/s 34677 TTFT (ms) 24.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/namchm7s",
   "source_url": "https://omlx.ai/benchmarks/performance/namchm7s",
   "tg_tokens": null,
   "tps": 48.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-ejwv0idp-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "Unsloth UD build; oMLX community benchmark trace, single run at 1k context; peak memory 20.6 GB; TTFT 1943 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 527.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "unsloth--Qwen3.6-35B-A3B-UD-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k tokens 527.0 PP tok/s 57.7 TG tok/s 1943 TTFT (ms) 20.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ejwv0idp",
   "source_url": "https://omlx.ai/benchmarks/performance/ejwv0idp",
   "tg_tokens": null,
   "tps": 57.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-mvypkk5o-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 172.4 GB at this context; duplicate traces of the same session: 0wh741n8, 7kcvil3z, 8xllkk30, geka9uq0, ot2wa2k0, r4wv1jnr; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1281.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,281 78.5 172.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mvypkk5o",
   "source_url": "https://omlx.ai/benchmarks/performance/mvypkk5o",
   "tg_tokens": null,
   "tps": 78.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-mvypkk5o-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 173.2 GB at this context; duplicate traces of the same session: 0wh741n8, 7kcvil3z, 8xllkk30, geka9uq0, ot2wa2k0, r4wv1jnr; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1923.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,923 68.5 173.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mvypkk5o",
   "source_url": "https://omlx.ai/benchmarks/performance/mvypkk5o",
   "tg_tokens": null,
   "tps": 68.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-mvypkk5o-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 173.8 GB at this context; duplicate traces of the same session: 0wh741n8, 7kcvil3z, 8xllkk30, geka9uq0, ot2wa2k0, r4wv1jnr; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1924.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 1,924 75.8 173.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mvypkk5o",
   "source_url": "https://omlx.ai/benchmarks/performance/mvypkk5o",
   "tg_tokens": null,
   "tps": 75.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-mvypkk5o-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 174.0 GB at this context; duplicate traces of the same session: 0wh741n8, 7kcvil3z, 8xllkk30, geka9uq0, ot2wa2k0, r4wv1jnr; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1913.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,913 150.3 174.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mvypkk5o",
   "source_url": "https://omlx.ai/benchmarks/performance/mvypkk5o",
   "tg_tokens": null,
   "tps": 150.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-mvypkk5o-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 174.4 GB at this context; duplicate traces of the same session: 0wh741n8, 7kcvil3z, 8xllkk30, geka9uq0, ot2wa2k0, r4wv1jnr; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1901.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,901 95.6 174.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mvypkk5o",
   "source_url": "https://omlx.ai/benchmarks/performance/mvypkk5o",
   "tg_tokens": null,
   "tps": 95.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-mvypkk5o-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 175.0 GB at this context; duplicate traces of the same session: 0wh741n8, 7kcvil3z, 8xllkk30, geka9uq0, ot2wa2k0, r4wv1jnr; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1878.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,878 82.5 175.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mvypkk5o",
   "source_url": "https://omlx.ai/benchmarks/performance/mvypkk5o",
   "tg_tokens": null,
   "tps": 82.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-mvypkk5o-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 176.3 GB at this context; duplicate traces of the same session: 0wh741n8, 7kcvil3z, 8xllkk30, geka9uq0, ot2wa2k0, r4wv1jnr; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1839.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 1,839 144.0 176.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mvypkk5o",
   "source_url": "https://omlx.ai/benchmarks/performance/mvypkk5o",
   "tg_tokens": null,
   "tps": 144.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-mvypkk5o-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.5 GB at this context; duplicate traces of the same session: 0wh741n8, 7kcvil3z, 8xllkk30, geka9uq0, ot2wa2k0, r4wv1jnr; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1812.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 1,812 73.6 177.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mvypkk5o",
   "source_url": "https://omlx.ai/benchmarks/performance/mvypkk5o",
   "tg_tokens": null,
   "tps": 73.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-p2idqvbe-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 172.4 GB at this context; duplicate traces of the same session: 2chlmfdz, 4i9317g7, gotyyc3r, iw95bqtt, kqlk9pat, uxey0o2m, yu3ve97a; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1301.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,301 69.2 172.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p2idqvbe",
   "source_url": "https://omlx.ai/benchmarks/performance/p2idqvbe",
   "tg_tokens": null,
   "tps": 69.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-p2idqvbe-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 173.2 GB at this context; duplicate traces of the same session: 2chlmfdz, 4i9317g7, gotyyc3r, iw95bqtt, kqlk9pat, uxey0o2m, yu3ve97a; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2009.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 2,009 70.2 173.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p2idqvbe",
   "source_url": "https://omlx.ai/benchmarks/performance/p2idqvbe",
   "tg_tokens": null,
   "tps": 70.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-p2idqvbe-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 174.2 GB at this context; duplicate traces of the same session: 2chlmfdz, 4i9317g7, gotyyc3r, iw95bqtt, kqlk9pat, uxey0o2m, yu3ve97a; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2008.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 2,008 70.6 174.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p2idqvbe",
   "source_url": "https://omlx.ai/benchmarks/performance/p2idqvbe",
   "tg_tokens": null,
   "tps": 70.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-p2idqvbe-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 174.3 GB at this context; duplicate traces of the same session: 2chlmfdz, 4i9317g7, gotyyc3r, iw95bqtt, kqlk9pat, uxey0o2m, yu3ve97a; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2003.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 2,003 70.4 174.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p2idqvbe",
   "source_url": "https://omlx.ai/benchmarks/performance/p2idqvbe",
   "tg_tokens": null,
   "tps": 70.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-p2idqvbe-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 174.5 GB at this context; duplicate traces of the same session: 2chlmfdz, 4i9317g7, gotyyc3r, iw95bqtt, kqlk9pat, uxey0o2m, yu3ve97a; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1981.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,981 69.2 174.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p2idqvbe",
   "source_url": "https://omlx.ai/benchmarks/performance/p2idqvbe",
   "tg_tokens": null,
   "tps": 69.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-p2idqvbe-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 174.8 GB at this context; duplicate traces of the same session: 2chlmfdz, 4i9317g7, gotyyc3r, iw95bqtt, kqlk9pat, uxey0o2m, yu3ve97a; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1963.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,963 68.5 174.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p2idqvbe",
   "source_url": "https://omlx.ai/benchmarks/performance/p2idqvbe",
   "tg_tokens": null,
   "tps": 68.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-p2idqvbe-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 175.5 GB at this context; duplicate traces of the same session: 2chlmfdz, 4i9317g7, gotyyc3r, iw95bqtt, kqlk9pat, uxey0o2m, yu3ve97a; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1919.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 1,919 65.0 175.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p2idqvbe",
   "source_url": "https://omlx.ai/benchmarks/performance/p2idqvbe",
   "tg_tokens": null,
   "tps": 65.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-p2idqvbe-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 176.3 GB at this context; duplicate traces of the same session: 2chlmfdz, 4i9317g7, gotyyc3r, iw95bqtt, kqlk9pat, uxey0o2m, yu3ve97a; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1879.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 1,879 60.8 176.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p2idqvbe",
   "source_url": "https://omlx.ai/benchmarks/performance/p2idqvbe",
   "tg_tokens": null,
   "tps": 60.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-djbbq7jy-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 172.4 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1290.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,290 70.0 172.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/djbbq7jy",
   "source_url": "https://omlx.ai/benchmarks/performance/djbbq7jy",
   "tg_tokens": null,
   "tps": 70.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-djbbq7jy-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 173.2 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1980.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,980 78.0 173.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/djbbq7jy",
   "source_url": "https://omlx.ai/benchmarks/performance/djbbq7jy",
   "tg_tokens": null,
   "tps": 78.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-djbbq7jy-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 173.8 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1978.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 1,978 76.4 173.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/djbbq7jy",
   "source_url": "https://omlx.ai/benchmarks/performance/djbbq7jy",
   "tg_tokens": null,
   "tps": 76.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-djbbq7jy-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 174.0 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1967.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,967 154.5 174.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/djbbq7jy",
   "source_url": "https://omlx.ai/benchmarks/performance/djbbq7jy",
   "tg_tokens": null,
   "tps": 154.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-djbbq7jy-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 174.4 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1954.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,954 97.9 174.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/djbbq7jy",
   "source_url": "https://omlx.ai/benchmarks/performance/djbbq7jy",
   "tg_tokens": null,
   "tps": 97.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-djbbq7jy-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 175.0 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1933.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,933 84.7 175.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/djbbq7jy",
   "source_url": "https://omlx.ai/benchmarks/performance/djbbq7jy",
   "tg_tokens": null,
   "tps": 84.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-djbbq7jy-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 176.3 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1887.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 1,887 128.9 176.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/djbbq7jy",
   "source_url": "https://omlx.ai/benchmarks/performance/djbbq7jy",
   "tg_tokens": null,
   "tps": 128.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-djbbq7jy-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.5 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1842.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 1,842 64.4 177.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/djbbq7jy",
   "source_url": "https://omlx.ai/benchmarks/performance/djbbq7jy",
   "tg_tokens": null,
   "tps": 64.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4upnz802-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: k4edphp1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 779.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 779.7 111.1 35.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4upnz802",
   "source_url": "https://omlx.ai/benchmarks/performance/4upnz802",
   "tg_tokens": null,
   "tps": 111.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4upnz802-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: k4edphp1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2509.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 2,509 109.2 36.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4upnz802",
   "source_url": "https://omlx.ai/benchmarks/performance/4upnz802",
   "tg_tokens": null,
   "tps": 109.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-bhyvi6jc-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 507.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 507.3 49.5 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhyvi6jc",
   "source_url": "https://omlx.ai/benchmarks/performance/bhyvi6jc",
   "tg_tokens": null,
   "tps": 49.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-bhyvi6jc-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 820.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 820.7 66.5 23.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhyvi6jc",
   "source_url": "https://omlx.ai/benchmarks/performance/bhyvi6jc",
   "tg_tokens": null,
   "tps": 66.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-bhyvi6jc-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 696.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 696.9 19.7 24.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhyvi6jc",
   "source_url": "https://omlx.ai/benchmarks/performance/bhyvi6jc",
   "tg_tokens": null,
   "tps": 19.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-bhyvi6jc-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 646.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 646.0 63.3 23.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhyvi6jc",
   "source_url": "https://omlx.ai/benchmarks/performance/bhyvi6jc",
   "tg_tokens": null,
   "tps": 63.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-bhyvi6jc-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 538.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 32k 538.0 65.7 24.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhyvi6jc",
   "source_url": "https://omlx.ai/benchmarks/performance/bhyvi6jc",
   "tg_tokens": null,
   "tps": 65.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-z05a9wby-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1918.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,918 146.9 22.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z05a9wby",
   "source_url": "https://omlx.ai/benchmarks/performance/z05a9wby",
   "tg_tokens": null,
   "tps": 146.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-z05a9wby-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1571.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 1,571 167.7 22.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z05a9wby",
   "source_url": "https://omlx.ai/benchmarks/performance/z05a9wby",
   "tg_tokens": null,
   "tps": 167.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-n4nenbye-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 12myfgwo; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 430.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 430.0 17.4 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n4nenbye",
   "source_url": "https://omlx.ai/benchmarks/performance/n4nenbye",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-n4nenbye-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 12myfgwo; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 466.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 466.4 17.3 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n4nenbye",
   "source_url": "https://omlx.ai/benchmarks/performance/n4nenbye",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lglzcxfv-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 330.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 330.2 45.4 29.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lglzcxfv",
   "source_url": "https://omlx.ai/benchmarks/performance/lglzcxfv",
   "tg_tokens": null,
   "tps": 45.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lglzcxfv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 31.1 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 336.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 336.3 39.5 31.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lglzcxfv",
   "source_url": "https://omlx.ai/benchmarks/performance/lglzcxfv",
   "tg_tokens": null,
   "tps": 39.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-uviph7wc-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: jr1ofimq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 325.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 325.8 23.0 28.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uviph7wc",
   "source_url": "https://omlx.ai/benchmarks/performance/uviph7wc",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-uviph7wc-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 31.7 GB at this context; duplicate traces of the same session: jr1ofimq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 342.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 342.7 22.7 31.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uviph7wc",
   "source_url": "https://omlx.ai/benchmarks/performance/uviph7wc",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-608pzkmv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace, single run at 4k context; peak memory 37.7 GB; TTFT 12210 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 335.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k tokens 335.5 PP tok/s 40.6 TG tok/s 12210 TTFT (ms) 37.7 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/608pzkmv",
   "source_url": "https://omlx.ai/benchmarks/performance/608pzkmv",
   "tg_tokens": null,
   "tps": 40.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-drnppw4o-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: 84bw4wia; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 417.3,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 417.3 62.9 36.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/drnppw4o",
   "source_url": "https://omlx.ai/benchmarks/performance/drnppw4o",
   "tg_tokens": null,
   "tps": 62.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-drnppw4o-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.9 GB at this context; duplicate traces of the same session: 84bw4wia; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 343.5,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k 343.5 50.9 39.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/drnppw4o",
   "source_url": "https://omlx.ai/benchmarks/performance/drnppw4o",
   "tg_tokens": null,
   "tps": 50.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-wu606z4z-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 4104l636, ktzev3dn, rf2xxzbb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 121.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 121.8 28.5 19.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wu606z4z",
   "source_url": "https://omlx.ai/benchmarks/performance/wu606z4z",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-wu606z4z-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: 4104l636, ktzev3dn, rf2xxzbb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 122.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 122.7 31.9 20.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wu606z4z",
   "source_url": "https://omlx.ai/benchmarks/performance/wu606z4z",
   "tg_tokens": null,
   "tps": 31.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-wu606z4z-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 4104l636, ktzev3dn, rf2xxzbb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 116.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 116.2 31.5 21.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wu606z4z",
   "source_url": "https://omlx.ai/benchmarks/performance/wu606z4z",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-wu606z4z-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: 4104l636, ktzev3dn, rf2xxzbb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 108.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 108.2 30.7 24.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wu606z4z",
   "source_url": "https://omlx.ai/benchmarks/performance/wu606z4z",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-53k5kwf6-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: jgow0aup; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 317.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 317.2 44.9 33.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/53k5kwf6",
   "source_url": "https://omlx.ai/benchmarks/performance/53k5kwf6",
   "tg_tokens": null,
   "tps": 44.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-53k5kwf6-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: jgow0aup; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 333.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 333.0 45.4 37.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/53k5kwf6",
   "source_url": "https://omlx.ai/benchmarks/performance/53k5kwf6",
   "tg_tokens": null,
   "tps": 45.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tymvd446-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: 1tiyelc6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 931.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 931.1 18.5 29.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tymvd446",
   "source_url": "https://omlx.ai/benchmarks/performance/tymvd446",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tymvd446-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: 1tiyelc6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 925.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 925.3 17.8 30.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tymvd446",
   "source_url": "https://omlx.ai/benchmarks/performance/tymvd446",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-h922y5xt-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: 1rs19l6x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 325.2,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 3bit; 4k 325.2 16.3 29.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h922y5xt",
   "source_url": "https://omlx.ai/benchmarks/performance/h922y5xt",
   "tg_tokens": null,
   "tps": 16.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-h922y5xt-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: 1rs19l6x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 405.3,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 3bit; 8k 405.3 21.8 29.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h922y5xt",
   "source_url": "https://omlx.ai/benchmarks/performance/h922y5xt",
   "tg_tokens": null,
   "tps": 21.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-un7eke03-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 29.9 GB; TTFT 22644 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 361.8,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 3bit; 8k tokens 361.8 PP tok/s 23.8 TG tok/s 22644 TTFT (ms) 29.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/un7eke03",
   "source_url": "https://omlx.ai/benchmarks/performance/un7eke03",
   "tg_tokens": null,
   "tps": 23.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2te97dwz-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 44.1 GB at this context; duplicate traces of the same session: 2t3hw543; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 380.2,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 3bit; 8k 380.2 23.9 44.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2te97dwz",
   "source_url": "https://omlx.ai/benchmarks/performance/2te97dwz",
   "tg_tokens": null,
   "tps": 23.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2te97dwz-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 45.6 GB at this context; duplicate traces of the same session: 2t3hw543; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 330.1,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 3bit; 32k 330.1 9.0 45.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2te97dwz",
   "source_url": "https://omlx.ai/benchmarks/performance/2te97dwz",
   "tg_tokens": null,
   "tps": 9.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4toxohgh-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: dinpyqia, wld5btsg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1481.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,481 62.5 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4toxohgh",
   "source_url": "https://omlx.ai/benchmarks/performance/4toxohgh",
   "tg_tokens": null,
   "tps": 62.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4toxohgh-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.2 GB at this context; duplicate traces of the same session: dinpyqia, wld5btsg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2195.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,195 49.3 71.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4toxohgh",
   "source_url": "https://omlx.ai/benchmarks/performance/4toxohgh",
   "tg_tokens": null,
   "tps": 49.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4toxohgh-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.1 GB at this context; duplicate traces of the same session: dinpyqia, wld5btsg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1933.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 1,933 60.0 74.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4toxohgh",
   "source_url": "https://omlx.ai/benchmarks/performance/4toxohgh",
   "tg_tokens": null,
   "tps": 60.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-tia5pldc-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.7 GB at this context; duplicate traces of the same session: jheu4w5k; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 250.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 250.7 10.1 28.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tia5pldc",
   "source_url": "https://omlx.ai/benchmarks/performance/tia5pldc",
   "tg_tokens": null,
   "tps": 10.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-tia5pldc-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.1 GB at this context; duplicate traces of the same session: jheu4w5k; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 367.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 367.7 11.8 30.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tia5pldc",
   "source_url": "https://omlx.ai/benchmarks/performance/tia5pldc",
   "tg_tokens": null,
   "tps": 11.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-drud38xe-ctx16384",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace, single run at 16k context; peak memory 22.5 GB; TTFT 148390 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 110.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k tokens 110.4 PP tok/s 13.7 TG tok/s 148390 TTFT (ms) 22.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/drud38xe",
   "source_url": "https://omlx.ai/benchmarks/performance/drud38xe",
   "tg_tokens": null,
   "tps": 13.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-4bpc28ua-ctx16384",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace, single run at 16k context; peak memory 22.5 GB; TTFT 156363 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k tokens 104.8 PP tok/s 12.2 TG tok/s 156363 TTFT (ms) 22.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4bpc28ua",
   "source_url": "https://omlx.ai/benchmarks/performance/4bpc28ua",
   "tg_tokens": null,
   "tps": 12.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-hlarklj7-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 8.2 GB at this context; duplicate traces of the same session: eirv48pa, gcn2649f, gwq4wnfn, jaex9reg, w7gl8y86, za3mezr3; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 254.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 1k 254.7 21.8 8.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hlarklj7",
   "source_url": "https://omlx.ai/benchmarks/performance/hlarklj7",
   "tg_tokens": null,
   "tps": 21.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-hlarklj7-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 8.6 GB at this context; duplicate traces of the same session: eirv48pa, gcn2649f, gwq4wnfn, jaex9reg, w7gl8y86, za3mezr3; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 300.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 4k 300.8 20.7 8.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hlarklj7",
   "source_url": "https://omlx.ai/benchmarks/performance/hlarklj7",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-hlarklj7-ctx8192",
   "model": "gemma-4-26B-A4B-it",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 8.8 GB at this context; duplicate traces of the same session: eirv48pa, gcn2649f, gwq4wnfn, jaex9reg, w7gl8y86, za3mezr3; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 297.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 8k 297.6 22.0 8.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hlarklj7",
   "source_url": "https://omlx.ai/benchmarks/performance/hlarklj7",
   "tg_tokens": null,
   "tps": 22.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-hlarklj7-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 9.2 GB at this context; duplicate traces of the same session: eirv48pa, gcn2649f, gwq4wnfn, jaex9reg, w7gl8y86, za3mezr3; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 286.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 16k 286.1 20.2 9.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hlarklj7",
   "source_url": "https://omlx.ai/benchmarks/performance/hlarklj7",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-hlarklj7-ctx32768",
   "model": "gemma-4-26B-A4B-it",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 10.0 GB at this context; duplicate traces of the same session: eirv48pa, gcn2649f, gwq4wnfn, jaex9reg, w7gl8y86, za3mezr3; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 260.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 32k 260.9 19.4 10.0 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hlarklj7",
   "source_url": "https://omlx.ai/benchmarks/performance/hlarklj7",
   "tg_tokens": null,
   "tps": 19.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-hlarklj7-ctx65536",
   "model": "gemma-4-26B-A4B-it",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 11.7 GB at this context; duplicate traces of the same session: eirv48pa, gcn2649f, gwq4wnfn, jaex9reg, w7gl8y86, za3mezr3; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 214.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 64k 214.7 14.9 11.7 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hlarklj7",
   "source_url": "https://omlx.ai/benchmarks/performance/hlarklj7",
   "tg_tokens": null,
   "tps": 14.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-hlarklj7-ctx131072",
   "model": "gemma-4-26B-A4B-it",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 10.9 GB at this context; duplicate traces of the same session: eirv48pa, gcn2649f, gwq4wnfn, jaex9reg, w7gl8y86, za3mezr3; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 128.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 128k 128.7 7.8 10.9 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hlarklj7",
   "source_url": "https://omlx.ai/benchmarks/performance/hlarklj7",
   "tg_tokens": null,
   "tps": 7.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-8zefvc34-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 172.4 GB at this context; duplicate traces of the same session: 2e0nxcuj, 6tqdf8to, b3cufvtf, g3spptdh, gbtytsf1, n64numet; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1290.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,290 70.0 172.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8zefvc34",
   "source_url": "https://omlx.ai/benchmarks/performance/8zefvc34",
   "tg_tokens": null,
   "tps": 70.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-8zefvc34-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 173.2 GB at this context; duplicate traces of the same session: 2e0nxcuj, 6tqdf8to, b3cufvtf, g3spptdh, gbtytsf1, n64numet; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1980.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,980 78.0 173.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8zefvc34",
   "source_url": "https://omlx.ai/benchmarks/performance/8zefvc34",
   "tg_tokens": null,
   "tps": 78.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-8zefvc34-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 173.8 GB at this context; duplicate traces of the same session: 2e0nxcuj, 6tqdf8to, b3cufvtf, g3spptdh, gbtytsf1, n64numet; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1978.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 1,978 76.4 173.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8zefvc34",
   "source_url": "https://omlx.ai/benchmarks/performance/8zefvc34",
   "tg_tokens": null,
   "tps": 76.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-8zefvc34-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 174.0 GB at this context; duplicate traces of the same session: 2e0nxcuj, 6tqdf8to, b3cufvtf, g3spptdh, gbtytsf1, n64numet; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1967.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,967 154.5 174.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8zefvc34",
   "source_url": "https://omlx.ai/benchmarks/performance/8zefvc34",
   "tg_tokens": null,
   "tps": 154.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-8zefvc34-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 174.4 GB at this context; duplicate traces of the same session: 2e0nxcuj, 6tqdf8to, b3cufvtf, g3spptdh, gbtytsf1, n64numet; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1954.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,954 97.9 174.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8zefvc34",
   "source_url": "https://omlx.ai/benchmarks/performance/8zefvc34",
   "tg_tokens": null,
   "tps": 97.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-8zefvc34-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 175.0 GB at this context; duplicate traces of the same session: 2e0nxcuj, 6tqdf8to, b3cufvtf, g3spptdh, gbtytsf1, n64numet; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1933.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,933 84.7 175.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8zefvc34",
   "source_url": "https://omlx.ai/benchmarks/performance/8zefvc34",
   "tg_tokens": null,
   "tps": 84.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-8zefvc34-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 176.3 GB at this context; duplicate traces of the same session: 2e0nxcuj, 6tqdf8to, b3cufvtf, g3spptdh, gbtytsf1, n64numet; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1887.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 1,887 128.9 176.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8zefvc34",
   "source_url": "https://omlx.ai/benchmarks/performance/8zefvc34",
   "tg_tokens": null,
   "tps": 128.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-8zefvc34-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.5 GB at this context; duplicate traces of the same session: 2e0nxcuj, 6tqdf8to, b3cufvtf, g3spptdh, gbtytsf1, n64numet; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1842.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 1,842 64.4 177.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8zefvc34",
   "source_url": "https://omlx.ai/benchmarks/performance/8zefvc34",
   "tg_tokens": null,
   "tps": 64.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-0ekoff1i-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 172.4 GB at this context; duplicate traces of the same session: 2vax6une, 384irv38, epi3ycm4, jcrbwxty, pn7g37no, s2rd67uy, tg1wqxss; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1306.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,306 84.7 172.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ekoff1i",
   "source_url": "https://omlx.ai/benchmarks/performance/0ekoff1i",
   "tg_tokens": null,
   "tps": 84.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-0ekoff1i-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 173.2 GB at this context; duplicate traces of the same session: 2vax6une, 384irv38, epi3ycm4, jcrbwxty, pn7g37no, s2rd67uy, tg1wqxss; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1983.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,983 74.0 173.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ekoff1i",
   "source_url": "https://omlx.ai/benchmarks/performance/0ekoff1i",
   "tg_tokens": null,
   "tps": 74.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-0ekoff1i-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 173.8 GB at this context; duplicate traces of the same session: 2vax6une, 384irv38, epi3ycm4, jcrbwxty, pn7g37no, s2rd67uy, tg1wqxss; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1970.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 1,970 75.3 173.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ekoff1i",
   "source_url": "https://omlx.ai/benchmarks/performance/0ekoff1i",
   "tg_tokens": null,
   "tps": 75.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-0ekoff1i-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 174.0 GB at this context; duplicate traces of the same session: 2vax6une, 384irv38, epi3ycm4, jcrbwxty, pn7g37no, s2rd67uy, tg1wqxss; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1964.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,964 153.7 174.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ekoff1i",
   "source_url": "https://omlx.ai/benchmarks/performance/0ekoff1i",
   "tg_tokens": null,
   "tps": 153.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-0ekoff1i-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 174.4 GB at this context; duplicate traces of the same session: 2vax6une, 384irv38, epi3ycm4, jcrbwxty, pn7g37no, s2rd67uy, tg1wqxss; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1955.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,955 98.6 174.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ekoff1i",
   "source_url": "https://omlx.ai/benchmarks/performance/0ekoff1i",
   "tg_tokens": null,
   "tps": 98.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-0ekoff1i-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 175.0 GB at this context; duplicate traces of the same session: 2vax6une, 384irv38, epi3ycm4, jcrbwxty, pn7g37no, s2rd67uy, tg1wqxss; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1929.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,929 85.3 175.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ekoff1i",
   "source_url": "https://omlx.ai/benchmarks/performance/0ekoff1i",
   "tg_tokens": null,
   "tps": 85.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-0ekoff1i-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 176.3 GB at this context; duplicate traces of the same session: 2vax6une, 384irv38, epi3ycm4, jcrbwxty, pn7g37no, s2rd67uy, tg1wqxss; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1882.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 1,882 129.3 176.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ekoff1i",
   "source_url": "https://omlx.ai/benchmarks/performance/0ekoff1i",
   "tg_tokens": null,
   "tps": 129.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-0ekoff1i-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.5 GB at this context; duplicate traces of the same session: 2vax6une, 384irv38, epi3ycm4, jcrbwxty, pn7g37no, s2rd67uy, tg1wqxss; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1839.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 1,839 76.1 177.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ekoff1i",
   "source_url": "https://omlx.ai/benchmarks/performance/0ekoff1i",
   "tg_tokens": null,
   "tps": 76.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-gba5jq1c-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 172.4 GB at this context; duplicate traces of the same session: 1ghh8w1f, 1moehyft, 5p98an1j, 6oq9cq5r, emqf05pb, ero28zhv, lfsr2703; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1308.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,308 77.6 172.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gba5jq1c",
   "source_url": "https://omlx.ai/benchmarks/performance/gba5jq1c",
   "tg_tokens": null,
   "tps": 77.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-gba5jq1c-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 173.2 GB at this context; duplicate traces of the same session: 1ghh8w1f, 1moehyft, 5p98an1j, 6oq9cq5r, emqf05pb, ero28zhv, lfsr2703; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2029.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 2,029 70.2 173.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gba5jq1c",
   "source_url": "https://omlx.ai/benchmarks/performance/gba5jq1c",
   "tg_tokens": null,
   "tps": 70.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-gba5jq1c-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 174.2 GB at this context; duplicate traces of the same session: 1ghh8w1f, 1moehyft, 5p98an1j, 6oq9cq5r, emqf05pb, ero28zhv, lfsr2703; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2007.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 2,007 70.7 174.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gba5jq1c",
   "source_url": "https://omlx.ai/benchmarks/performance/gba5jq1c",
   "tg_tokens": null,
   "tps": 70.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-gba5jq1c-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 174.3 GB at this context; duplicate traces of the same session: 1ghh8w1f, 1moehyft, 5p98an1j, 6oq9cq5r, emqf05pb, ero28zhv, lfsr2703; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2002.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 2,002 70.3 174.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gba5jq1c",
   "source_url": "https://omlx.ai/benchmarks/performance/gba5jq1c",
   "tg_tokens": null,
   "tps": 70.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-gba5jq1c-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 174.5 GB at this context; duplicate traces of the same session: 1ghh8w1f, 1moehyft, 5p98an1j, 6oq9cq5r, emqf05pb, ero28zhv, lfsr2703; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1993.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,993 70.2 174.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gba5jq1c",
   "source_url": "https://omlx.ai/benchmarks/performance/gba5jq1c",
   "tg_tokens": null,
   "tps": 70.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-gba5jq1c-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 174.8 GB at this context; duplicate traces of the same session: 1ghh8w1f, 1moehyft, 5p98an1j, 6oq9cq5r, emqf05pb, ero28zhv, lfsr2703; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1965.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,965 68.6 174.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gba5jq1c",
   "source_url": "https://omlx.ai/benchmarks/performance/gba5jq1c",
   "tg_tokens": null,
   "tps": 68.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-gba5jq1c-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 175.5 GB at this context; duplicate traces of the same session: 1ghh8w1f, 1moehyft, 5p98an1j, 6oq9cq5r, emqf05pb, ero28zhv, lfsr2703; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1920.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 1,920 65.1 175.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gba5jq1c",
   "source_url": "https://omlx.ai/benchmarks/performance/gba5jq1c",
   "tg_tokens": null,
   "tps": 65.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-gba5jq1c-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 176.3 GB at this context; duplicate traces of the same session: 1ghh8w1f, 1moehyft, 5p98an1j, 6oq9cq5r, emqf05pb, ero28zhv, lfsr2703; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1880.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 1,880 60.8 176.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gba5jq1c",
   "source_url": "https://omlx.ai/benchmarks/performance/gba5jq1c",
   "tg_tokens": null,
   "tps": 60.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-m9zceysw-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 182.7 GB at this context; duplicate traces of the same session: 2jjp6scg; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2354.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,354 107.9 182.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m9zceysw",
   "source_url": "https://omlx.ai/benchmarks/performance/m9zceysw",
   "tg_tokens": null,
   "tps": 107.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-m9zceysw-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 184.0 GB at this context; duplicate traces of the same session: 2jjp6scg; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2325.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,325 96.1 184.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m9zceysw",
   "source_url": "https://omlx.ai/benchmarks/performance/m9zceysw",
   "tg_tokens": null,
   "tps": 96.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-m9zceysw-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 186.9 GB at this context; duplicate traces of the same session: 2jjp6scg; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2281.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,281 145.1 186.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m9zceysw",
   "source_url": "https://omlx.ai/benchmarks/performance/m9zceysw",
   "tg_tokens": null,
   "tps": 145.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-4hn1c4k4-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 011kej4h, u8cljp7w; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1265.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 1,265 91.1 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4hn1c4k4",
   "source_url": "https://omlx.ai/benchmarks/performance/4hn1c4k4",
   "tg_tokens": null,
   "tps": 91.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-4hn1c4k4-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 011kej4h, u8cljp7w; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1518.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 1,518 89.4 21.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4hn1c4k4",
   "source_url": "https://omlx.ai/benchmarks/performance/4hn1c4k4",
   "tg_tokens": null,
   "tps": 89.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-4hn1c4k4-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 011kej4h, u8cljp7w; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1237.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 16k 1,237 80.6 22.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4hn1c4k4",
   "source_url": "https://omlx.ai/benchmarks/performance/4hn1c4k4",
   "tg_tokens": null,
   "tps": 80.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-d3h2c9ay-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1918.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,918 146.9 22.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d3h2c9ay",
   "source_url": "https://omlx.ai/benchmarks/performance/d3h2c9ay",
   "tg_tokens": null,
   "tps": 146.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-d3h2c9ay-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1571.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 1,571 167.7 22.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d3h2c9ay",
   "source_url": "https://omlx.ai/benchmarks/performance/d3h2c9ay",
   "tg_tokens": null,
   "tps": 167.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ktuj8ft5-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: m8vwtfq3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 211.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 211.1 9.9 15.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ktuj8ft5",
   "source_url": "https://omlx.ai/benchmarks/performance/ktuj8ft5",
   "tg_tokens": null,
   "tps": 9.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ktuj8ft5-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: m8vwtfq3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 205.4 13.9 17.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ktuj8ft5",
   "source_url": "https://omlx.ai/benchmarks/performance/ktuj8ft5",
   "tg_tokens": null,
   "tps": 13.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-wwk7lmvd-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 34.6 GB at this context; duplicate traces of the same session: rwfzfh5l; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 330.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 330.6 44.6 34.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wwk7lmvd",
   "source_url": "https://omlx.ai/benchmarks/performance/wwk7lmvd",
   "tg_tokens": null,
   "tps": 44.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-wwk7lmvd-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 37.9 GB at this context; duplicate traces of the same session: rwfzfh5l; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 335.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 335.4 39.3 37.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wwk7lmvd",
   "source_url": "https://omlx.ai/benchmarks/performance/wwk7lmvd",
   "tg_tokens": null,
   "tps": 39.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-bt8ahzdi-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: vy9x86c0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 330.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 330.5 42.3 29.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bt8ahzdi",
   "source_url": "https://omlx.ai/benchmarks/performance/bt8ahzdi",
   "tg_tokens": null,
   "tps": 42.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-bt8ahzdi-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 31.1 GB at this context; duplicate traces of the same session: vy9x86c0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 336.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 336.3 38.0 31.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bt8ahzdi",
   "source_url": "https://omlx.ai/benchmarks/performance/bt8ahzdi",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gibkkbh5-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: sd0scvcf; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 325.6,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 325.6 72.0 19.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gibkkbh5",
   "source_url": "https://omlx.ai/benchmarks/performance/gibkkbh5",
   "tg_tokens": null,
   "tps": 72.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gibkkbh5-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: sd0scvcf; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 358.3,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 358.3 73.5 22.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gibkkbh5",
   "source_url": "https://omlx.ai/benchmarks/performance/gibkkbh5",
   "tg_tokens": null,
   "tps": 73.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nesd5v0w-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: 1iqj5q98, 2m7kvu1y, 7dj51g3w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 910.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 910.9 57.6 20.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nesd5v0w",
   "source_url": "https://omlx.ai/benchmarks/performance/nesd5v0w",
   "tg_tokens": null,
   "tps": 57.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nesd5v0w-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: 1iqj5q98, 2m7kvu1y, 7dj51g3w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 758.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 758.6 102.2 22.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nesd5v0w",
   "source_url": "https://omlx.ai/benchmarks/performance/nesd5v0w",
   "tg_tokens": null,
   "tps": 102.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nesd5v0w-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: 1iqj5q98, 2m7kvu1y, 7dj51g3w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 653.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 653.3 47.2 27.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nesd5v0w",
   "source_url": "https://omlx.ai/benchmarks/performance/nesd5v0w",
   "tg_tokens": null,
   "tps": 47.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nesd5v0w-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.1 GB at this context; duplicate traces of the same session: 1iqj5q98, 2m7kvu1y, 7dj51g3w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 498.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 498.9 42.0 37.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nesd5v0w",
   "source_url": "https://omlx.ai/benchmarks/performance/nesd5v0w",
   "tg_tokens": null,
   "tps": 42.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-6b10b064-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 17.7 GB; TTFT 2677 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 382.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k tokens 382.5 PP tok/s 33.9 TG tok/s 2677 TTFT (ms) 17.7 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6b10b064",
   "source_url": "https://omlx.ai/benchmarks/performance/6b10b064",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-crw03gu9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: 8cfe03s4, vf4n67ef; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 783.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 783.1 38.0 24.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/crw03gu9",
   "source_url": "https://omlx.ai/benchmarks/performance/crw03gu9",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-crw03gu9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: 8cfe03s4, vf4n67ef; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1054.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 1,054 51.6 25.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/crw03gu9",
   "source_url": "https://omlx.ai/benchmarks/performance/crw03gu9",
   "tg_tokens": null,
   "tps": 51.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-crw03gu9-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: 8cfe03s4, vf4n67ef; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1103.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 1,103 58.5 26.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/crw03gu9",
   "source_url": "https://omlx.ai/benchmarks/performance/crw03gu9",
   "tg_tokens": null,
   "tps": 58.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-crw03gu9-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: 8cfe03s4, vf4n67ef; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1056.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 1,056 52.6 27.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/crw03gu9",
   "source_url": "https://omlx.ai/benchmarks/performance/crw03gu9",
   "tg_tokens": null,
   "tps": 52.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-crw03gu9-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.8 GB at this context; duplicate traces of the same session: 8cfe03s4, vf4n67ef; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 970.6,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 970.6 128.6 29.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/crw03gu9",
   "source_url": "https://omlx.ai/benchmarks/performance/crw03gu9",
   "tg_tokens": null,
   "tps": 128.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-muq6k65p-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 36.4 GB; TTFT 313952 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 417.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 128k tokens 417.5 PP tok/s 30.5 TG tok/s 313952 TTFT (ms) 36.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/muq6k65p",
   "source_url": "https://omlx.ai/benchmarks/performance/muq6k65p",
   "tg_tokens": null,
   "tps": 30.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-d52bacj1-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: fnd1j5su; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 318.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 318.1 46.7 33.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d52bacj1",
   "source_url": "https://omlx.ai/benchmarks/performance/d52bacj1",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-d52bacj1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: fnd1j5su; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 333.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 333.7 50.2 37.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d52bacj1",
   "source_url": "https://omlx.ai/benchmarks/performance/d52bacj1",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-aegecp0a-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.5 GB at this context; duplicate traces of the same session: 6llayxao; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2232.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,232 57.8 69.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aegecp0a",
   "source_url": "https://omlx.ai/benchmarks/performance/aegecp0a",
   "tg_tokens": null,
   "tps": 57.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-aegecp0a-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.0 GB at this context; duplicate traces of the same session: 6llayxao; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2003.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,003 49.3 72.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aegecp0a",
   "source_url": "https://omlx.ai/benchmarks/performance/aegecp0a",
   "tg_tokens": null,
   "tps": 49.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vtk66jfz-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: h16olla9, o95zosy3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1724.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,724 49.6 71.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vtk66jfz",
   "source_url": "https://omlx.ai/benchmarks/performance/vtk66jfz",
   "tg_tokens": null,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vtk66jfz-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.7 GB at this context; duplicate traces of the same session: h16olla9, o95zosy3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2340.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,340 72.9 72.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vtk66jfz",
   "source_url": "https://omlx.ai/benchmarks/performance/vtk66jfz",
   "tg_tokens": null,
   "tps": 72.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vtk66jfz-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.5 GB at this context; duplicate traces of the same session: h16olla9, o95zosy3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1971.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 1,971 60.1 78.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vtk66jfz",
   "source_url": "https://omlx.ai/benchmarks/performance/vtk66jfz",
   "tg_tokens": null,
   "tps": 60.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-b3es7jcc-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: il9ul21w; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 436.3,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-MLX-oQ4-MTP on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 436.3 62.9 177.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b3es7jcc",
   "source_url": "https://omlx.ai/benchmarks/performance/b3es7jcc",
   "tg_tokens": null,
   "tps": 62.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-b3es7jcc-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: il9ul21w; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1137.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-MLX-oQ4-MTP on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 1,137 57.0 178.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b3es7jcc",
   "source_url": "https://omlx.ai/benchmarks/performance/b3es7jcc",
   "tg_tokens": null,
   "tps": 57.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-jgg90rpp-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: 9bx63oyv, fe33b3wi, hc7jagai, ofra494j; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1261.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,261 68.7 177.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jgg90rpp",
   "source_url": "https://omlx.ai/benchmarks/performance/jgg90rpp",
   "tg_tokens": null,
   "tps": 68.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-jgg90rpp-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: 9bx63oyv, fe33b3wi, hc7jagai, ofra494j; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1939.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,939 64.4 178.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jgg90rpp",
   "source_url": "https://omlx.ai/benchmarks/performance/jgg90rpp",
   "tg_tokens": null,
   "tps": 64.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-jgg90rpp-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.9 GB at this context; duplicate traces of the same session: 9bx63oyv, fe33b3wi, hc7jagai, ofra494j; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1931.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 1,931 75.9 178.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jgg90rpp",
   "source_url": "https://omlx.ai/benchmarks/performance/jgg90rpp",
   "tg_tokens": null,
   "tps": 75.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-jgg90rpp-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.1 GB at this context; duplicate traces of the same session: 9bx63oyv, fe33b3wi, hc7jagai, ofra494j; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1917.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,917 138.8 179.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jgg90rpp",
   "source_url": "https://omlx.ai/benchmarks/performance/jgg90rpp",
   "tg_tokens": null,
   "tps": 138.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-jgg90rpp-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.5 GB at this context; duplicate traces of the same session: 9bx63oyv, fe33b3wi, hc7jagai, ofra494j; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1914.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,914 94.3 179.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jgg90rpp",
   "source_url": "https://omlx.ai/benchmarks/performance/jgg90rpp",
   "tg_tokens": null,
   "tps": 94.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-jgg90rpp-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.1 GB at this context; duplicate traces of the same session: 9bx63oyv, fe33b3wi, hc7jagai, ofra494j; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1885.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,885 71.1 180.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jgg90rpp",
   "source_url": "https://omlx.ai/benchmarks/performance/jgg90rpp",
   "tg_tokens": null,
   "tps": 71.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-jgg90rpp-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 182.6 GB at this context; duplicate traces of the same session: 9bx63oyv, fe33b3wi, hc7jagai, ofra494j; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1807.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 1,807 65.4 182.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jgg90rpp",
   "source_url": "https://omlx.ai/benchmarks/performance/jgg90rpp",
   "tg_tokens": null,
   "tps": 65.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-cw6qs5bx-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: 2przex7f, dx3s8xue, j3ld24b5, jabh2ff1, t21hbdlk, yg9vw3g6, zof3ux22; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 961.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 961.8 59.0 177.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cw6qs5bx",
   "source_url": "https://omlx.ai/benchmarks/performance/cw6qs5bx",
   "tg_tokens": null,
   "tps": 59.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-cw6qs5bx-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: 2przex7f, dx3s8xue, j3ld24b5, jabh2ff1, t21hbdlk, yg9vw3g6, zof3ux22; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1783.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,783 66.6 178.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cw6qs5bx",
   "source_url": "https://omlx.ai/benchmarks/performance/cw6qs5bx",
   "tg_tokens": null,
   "tps": 66.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-cw6qs5bx-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.9 GB at this context; duplicate traces of the same session: 2przex7f, dx3s8xue, j3ld24b5, jabh2ff1, t21hbdlk, yg9vw3g6, zof3ux22; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1884.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 1,884 73.1 178.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cw6qs5bx",
   "source_url": "https://omlx.ai/benchmarks/performance/cw6qs5bx",
   "tg_tokens": null,
   "tps": 73.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-cw6qs5bx-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.1 GB at this context; duplicate traces of the same session: 2przex7f, dx3s8xue, j3ld24b5, jabh2ff1, t21hbdlk, yg9vw3g6, zof3ux22; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1881.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,881 136.9 179.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cw6qs5bx",
   "source_url": "https://omlx.ai/benchmarks/performance/cw6qs5bx",
   "tg_tokens": null,
   "tps": 136.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-cw6qs5bx-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.5 GB at this context; duplicate traces of the same session: 2przex7f, dx3s8xue, j3ld24b5, jabh2ff1, t21hbdlk, yg9vw3g6, zof3ux22; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1869.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,869 83.3 179.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cw6qs5bx",
   "source_url": "https://omlx.ai/benchmarks/performance/cw6qs5bx",
   "tg_tokens": null,
   "tps": 83.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-cw6qs5bx-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.1 GB at this context; duplicate traces of the same session: 2przex7f, dx3s8xue, j3ld24b5, jabh2ff1, t21hbdlk, yg9vw3g6, zof3ux22; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1839.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,839 64.3 180.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cw6qs5bx",
   "source_url": "https://omlx.ai/benchmarks/performance/cw6qs5bx",
   "tg_tokens": null,
   "tps": 64.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-cw6qs5bx-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.3 GB at this context; duplicate traces of the same session: 2przex7f, dx3s8xue, j3ld24b5, jabh2ff1, t21hbdlk, yg9vw3g6, zof3ux22; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1805.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 1,805 116.3 181.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cw6qs5bx",
   "source_url": "https://omlx.ai/benchmarks/performance/cw6qs5bx",
   "tg_tokens": null,
   "tps": 116.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-cw6qs5bx-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 182.6 GB at this context; duplicate traces of the same session: 2przex7f, dx3s8xue, j3ld24b5, jabh2ff1, t21hbdlk, yg9vw3g6, zof3ux22; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1765.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 1,765 65.8 182.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cw6qs5bx",
   "source_url": "https://omlx.ai/benchmarks/performance/cw6qs5bx",
   "tg_tokens": null,
   "tps": 65.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-auzzfyqi-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 181.5 GB; TTFT 71460 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1834.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k tokens 1,834 PP tok/s 75.3 TG tok/s 71460 TTFT (ms) 181.5 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/auzzfyqi",
   "source_url": "https://omlx.ai/benchmarks/performance/auzzfyqi",
   "tg_tokens": null,
   "tps": 75.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-qfisqt8f-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 182.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2354.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,354 107.9 182.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qfisqt8f",
   "source_url": "https://omlx.ai/benchmarks/performance/qfisqt8f",
   "tg_tokens": null,
   "tps": 107.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-qfisqt8f-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 184.0 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2325.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,325 96.1 184.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qfisqt8f",
   "source_url": "https://omlx.ai/benchmarks/performance/qfisqt8f",
   "tg_tokens": null,
   "tps": 96.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-qfisqt8f-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 186.9 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2281.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,281 145.1 186.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qfisqt8f",
   "source_url": "https://omlx.ai/benchmarks/performance/qfisqt8f",
   "tg_tokens": null,
   "tps": 145.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-f1olsqbs-ctx1024",
   "model": "Ornith-1.5-35B-A3B-REAP192",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 14.4 GB; TTFT 1074 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 953.4,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Ornith-1.5-35B-A3B-REAP192-mxfp4-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k tokens 953.4 PP tok/s 98.8 TG tok/s 1074 TTFT (ms) 14.4 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f1olsqbs",
   "source_url": "https://omlx.ai/benchmarks/performance/f1olsqbs",
   "tg_tokens": null,
   "tps": 98.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-y86va00l-ctx32768",
   "model": "Ornith-1.5-35B-A3B-REAP192",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 15.8 GB; TTFT 45147 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 725.8,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Ornith-1.5-35B-A3B-REAP192-mxfp4-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k tokens 725.8 PP tok/s 63.7 TG tok/s 45147 TTFT (ms) 15.8 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y86va00l",
   "source_url": "https://omlx.ai/benchmarks/performance/y86va00l",
   "tg_tokens": null,
   "tps": 63.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-zp1yqoxs-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: 0tmtxouj; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 386.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 1k 386.1 58.5 5.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zp1yqoxs",
   "source_url": "https://omlx.ai/benchmarks/performance/zp1yqoxs",
   "tg_tokens": null,
   "tps": 58.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-zp1yqoxs-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: 0tmtxouj; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 380.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 4k 380.4 56.5 6.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zp1yqoxs",
   "source_url": "https://omlx.ai/benchmarks/performance/zp1yqoxs",
   "tg_tokens": null,
   "tps": 56.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-hono48qc-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: tgh59mqj; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 235.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 235.0 21.9 5.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hono48qc",
   "source_url": "https://omlx.ai/benchmarks/performance/hono48qc",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-hono48qc-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: tgh59mqj; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 242.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 242.3 21.5 6.2 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hono48qc",
   "source_url": "https://omlx.ai/benchmarks/performance/hono48qc",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-m084w0dm-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: jqmc1rrr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1667.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 1,667 152.6 29.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m084w0dm",
   "source_url": "https://omlx.ai/benchmarks/performance/m084w0dm",
   "tg_tokens": null,
   "tps": 152.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-m084w0dm-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: jqmc1rrr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1833.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 1,833 143.7 32.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m084w0dm",
   "source_url": "https://omlx.ai/benchmarks/performance/m084w0dm",
   "tg_tokens": null,
   "tps": 143.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-xginw6qu-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: atumskkb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1681.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 1,681 167.7 37.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xginw6qu",
   "source_url": "https://omlx.ai/benchmarks/performance/xginw6qu",
   "tg_tokens": null,
   "tps": 167.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-xginw6qu-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.5 GB at this context; duplicate traces of the same session: atumskkb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1860.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 1,860 123.0 39.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xginw6qu",
   "source_url": "https://omlx.ai/benchmarks/performance/xginw6qu",
   "tg_tokens": null,
   "tps": 123.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ospt4hty-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: dvqm7ui9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1759.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 1,759 96.6 35.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ospt4hty",
   "source_url": "https://omlx.ai/benchmarks/performance/ospt4hty",
   "tg_tokens": null,
   "tps": 96.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ospt4hty-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.5 GB at this context; duplicate traces of the same session: dvqm7ui9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1961.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 1,961 94.6 37.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ospt4hty",
   "source_url": "https://omlx.ai/benchmarks/performance/ospt4hty",
   "tg_tokens": null,
   "tps": 94.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-fjlncy9z-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1680.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 1,680 158.3 37.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fjlncy9z",
   "source_url": "https://omlx.ai/benchmarks/performance/fjlncy9z",
   "tg_tokens": null,
   "tps": 158.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-fjlncy9z-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1859.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 1,859 136.3 39.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fjlncy9z",
   "source_url": "https://omlx.ai/benchmarks/performance/fjlncy9z",
   "tg_tokens": null,
   "tps": 136.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-ahh3wcse-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: ekn53v2l; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 215.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 215.7 46.7 18.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ahh3wcse",
   "source_url": "https://omlx.ai/benchmarks/performance/ahh3wcse",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-ahh3wcse-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: ekn53v2l; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 199.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 199.7 41.5 22.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ahh3wcse",
   "source_url": "https://omlx.ai/benchmarks/performance/ahh3wcse",
   "tg_tokens": null,
   "tps": 41.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bqc9yghv-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: 582va6ph; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 783.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 783.1 38.0 24.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bqc9yghv",
   "source_url": "https://omlx.ai/benchmarks/performance/bqc9yghv",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bqc9yghv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: 582va6ph; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1054.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 1,054 51.6 25.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bqc9yghv",
   "source_url": "https://omlx.ai/benchmarks/performance/bqc9yghv",
   "tg_tokens": null,
   "tps": 51.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bqc9yghv-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: 582va6ph; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1103.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 1,103 58.5 26.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bqc9yghv",
   "source_url": "https://omlx.ai/benchmarks/performance/bqc9yghv",
   "tg_tokens": null,
   "tps": 58.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bqc9yghv-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: 582va6ph; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1056.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 1,056 52.6 27.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bqc9yghv",
   "source_url": "https://omlx.ai/benchmarks/performance/bqc9yghv",
   "tg_tokens": null,
   "tps": 52.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bqc9yghv-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.8 GB at this context; duplicate traces of the same session: 582va6ph; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 970.6,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 970.6 128.6 29.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bqc9yghv",
   "source_url": "https://omlx.ai/benchmarks/performance/bqc9yghv",
   "tg_tokens": null,
   "tps": 128.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bexllozv-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 945.5,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 945.5 55.7 24.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bexllozv",
   "source_url": "https://omlx.ai/benchmarks/performance/bexllozv",
   "tg_tokens": null,
   "tps": 55.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bexllozv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1055.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 1,055 54.1 25.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bexllozv",
   "source_url": "https://omlx.ai/benchmarks/performance/bexllozv",
   "tg_tokens": null,
   "tps": 54.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bexllozv-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1075.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 1,075 59.0 26.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bexllozv",
   "source_url": "https://omlx.ai/benchmarks/performance/bexllozv",
   "tg_tokens": null,
   "tps": 59.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bexllozv-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1034.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 1,034 78.4 27.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bexllozv",
   "source_url": "https://omlx.ai/benchmarks/performance/bexllozv",
   "tg_tokens": null,
   "tps": 78.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bexllozv-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 948.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 948.9 115.2 29.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bexllozv",
   "source_url": "https://omlx.ai/benchmarks/performance/bexllozv",
   "tg_tokens": null,
   "tps": 115.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-umht6s2r-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: 20ce46ys; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 923.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 923.2 18.2 29.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/umht6s2r",
   "source_url": "https://omlx.ai/benchmarks/performance/umht6s2r",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-umht6s2r-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: 20ce46ys; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 904.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 904.2 17.6 30.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/umht6s2r",
   "source_url": "https://omlx.ai/benchmarks/performance/umht6s2r",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jp6sg5zm-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.1 GB at this context; duplicate traces of the same session: y7qhpge1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 902.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 902.2 30.3 35.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jp6sg5zm",
   "source_url": "https://omlx.ai/benchmarks/performance/jp6sg5zm",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jp6sg5zm-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.7 GB at this context; duplicate traces of the same session: y7qhpge1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 881.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 881.3 54.6 35.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jp6sg5zm",
   "source_url": "https://omlx.ai/benchmarks/performance/jp6sg5zm",
   "tg_tokens": null,
   "tps": 54.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gxiv5rc1-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 195k context; peak memory 48.3 GB; TTFT 555196 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 360.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 195k tokens 360.2 PP tok/s 18.9 TG tok/s 555196 TTFT (ms) 48.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gxiv5rc1",
   "source_url": "https://omlx.ai/benchmarks/performance/gxiv5rc1",
   "tg_tokens": null,
   "tps": 18.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-oqbootf3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: 4cv3ctzp, f20q0bqo, jwkujet4, ovt12hgb, qetjizck, w6ejkypr, zagd33fm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 802.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 802.3 50.2 33.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oqbootf3",
   "source_url": "https://omlx.ai/benchmarks/performance/oqbootf3",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-oqbootf3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: 4cv3ctzp, f20q0bqo, jwkujet4, ovt12hgb, qetjizck, w6ejkypr, zagd33fm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 920.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 920.1 38.3 34.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oqbootf3",
   "source_url": "https://omlx.ai/benchmarks/performance/oqbootf3",
   "tg_tokens": null,
   "tps": 38.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-oqbootf3-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: 4cv3ctzp, f20q0bqo, jwkujet4, ovt12hgb, qetjizck, w6ejkypr, zagd33fm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 921.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 921.7 50.4 35.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oqbootf3",
   "source_url": "https://omlx.ai/benchmarks/performance/oqbootf3",
   "tg_tokens": null,
   "tps": 50.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-oqbootf3-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: 4cv3ctzp, f20q0bqo, jwkujet4, ovt12hgb, qetjizck, w6ejkypr, zagd33fm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 898.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 898.3 56.7 35.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oqbootf3",
   "source_url": "https://omlx.ai/benchmarks/performance/oqbootf3",
   "tg_tokens": null,
   "tps": 56.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-oqbootf3-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: 4cv3ctzp, f20q0bqo, jwkujet4, ovt12hgb, qetjizck, w6ejkypr, zagd33fm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 842.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 842.4 58.8 36.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oqbootf3",
   "source_url": "https://omlx.ai/benchmarks/performance/oqbootf3",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-oqbootf3-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: 4cv3ctzp, f20q0bqo, jwkujet4, ovt12hgb, qetjizck, w6ejkypr, zagd33fm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 734.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 734.8 41.9 39.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oqbootf3",
   "source_url": "https://omlx.ai/benchmarks/performance/oqbootf3",
   "tg_tokens": null,
   "tps": 41.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-oqbootf3-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 44.0 GB at this context; duplicate traces of the same session: 4cv3ctzp, f20q0bqo, jwkujet4, ovt12hgb, qetjizck, w6ejkypr, zagd33fm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 582.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 582.2 40.8 44.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oqbootf3",
   "source_url": "https://omlx.ai/benchmarks/performance/oqbootf3",
   "tg_tokens": null,
   "tps": 40.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-oqbootf3-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 55.5 GB at this context; duplicate traces of the same session: 4cv3ctzp, f20q0bqo, jwkujet4, ovt12hgb, qetjizck, w6ejkypr, zagd33fm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 463.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 195k 463.0 12.2 55.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oqbootf3",
   "source_url": "https://omlx.ai/benchmarks/performance/oqbootf3",
   "tg_tokens": null,
   "tps": 12.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ul5o5358-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP320",
   "notes": "MTP Vision build; oMLX community benchmark trace; peak memory 36.6 GB at this context; duplicate traces of the same session: 3xx27xwg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 962.3,
   "provenance": "community",
   "quant": "oQ3e-DWQ",
   "quote": "Qwen3.8-Flash-Next-REAP320-oQ3e-DWQ-MTP-Vision-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 3bit; 1k 962.3 65.8 36.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ul5o5358",
   "source_url": "https://omlx.ai/benchmarks/performance/ul5o5358",
   "tg_tokens": null,
   "tps": 65.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ul5o5358-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP320",
   "notes": "MTP Vision build; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: 3xx27xwg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1832.0,
   "provenance": "community",
   "quant": "oQ3e-DWQ",
   "quote": "Qwen3.8-Flash-Next-REAP320-oQ3e-DWQ-MTP-Vision-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 3bit; 4k 1,832 57.3 36.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ul5o5358",
   "source_url": "https://omlx.ai/benchmarks/performance/ul5o5358",
   "tg_tokens": null,
   "tps": 57.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xc1c209o-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.3 GB at this context; duplicate traces of the same session: 5wszns10; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 390.9,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 3bit; 8k 390.9 22.9 35.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xc1c209o",
   "source_url": "https://omlx.ai/benchmarks/performance/xc1c209o",
   "tg_tokens": null,
   "tps": 22.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xc1c209o-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.9 GB at this context; duplicate traces of the same session: 5wszns10; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 446.9,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 3bit; 16k 446.9 21.6 35.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xc1c209o",
   "source_url": "https://omlx.ai/benchmarks/performance/xc1c209o",
   "tg_tokens": null,
   "tps": 21.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ee7l4x48-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: cuzes0mb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1055.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,055 60.2 71.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ee7l4x48",
   "source_url": "https://omlx.ai/benchmarks/performance/ee7l4x48",
   "tg_tokens": null,
   "tps": 60.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ee7l4x48-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: cuzes0mb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1051.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 1,051 97.2 71.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ee7l4x48",
   "source_url": "https://omlx.ai/benchmarks/performance/ee7l4x48",
   "tg_tokens": null,
   "tps": 97.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ikqm08dz-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: 8hzmgnsy; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1049.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,049 75.0 71.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ikqm08dz",
   "source_url": "https://omlx.ai/benchmarks/performance/ikqm08dz",
   "tg_tokens": null,
   "tps": 75.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ikqm08dz-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.9 GB at this context; duplicate traces of the same session: 8hzmgnsy; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1052.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 1,052 85.2 71.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ikqm08dz",
   "source_url": "https://omlx.ai/benchmarks/performance/ikqm08dz",
   "tg_tokens": null,
   "tps": 85.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ut5mvdmu-ctx1024",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: zadi7rnq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 180.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 180.0 19.1 34.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ut5mvdmu",
   "source_url": "https://omlx.ai/benchmarks/performance/ut5mvdmu",
   "tg_tokens": null,
   "tps": 19.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ut5mvdmu-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 35.0 GB at this context; duplicate traces of the same session: zadi7rnq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 120.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 120.7 18.5 35.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ut5mvdmu",
   "source_url": "https://omlx.ai/benchmarks/performance/ut5mvdmu",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-x1wt7yre-ctx4096",
   "model": "Qwen3.6-27B",
   "notes": "mlx-community OptiQ 4bit build; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: mdq2q2qf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-27B-OptiQ-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 131.3 12.9 20.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x1wt7yre",
   "source_url": "https://omlx.ai/benchmarks/performance/x1wt7yre",
   "tg_tokens": null,
   "tps": 12.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-x1wt7yre-ctx16384",
   "model": "Qwen3.6-27B",
   "notes": "mlx-community OptiQ 4bit build; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: mdq2q2qf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 126.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-27B-OptiQ-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 126.8 12.1 21.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x1wt7yre",
   "source_url": "https://omlx.ai/benchmarks/performance/x1wt7yre",
   "tg_tokens": null,
   "tps": 12.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1pirzdfi-ctx1024",
   "model": "GLM-5.3-Flash-Abliterated",
   "notes": "MLX; oMLX community benchmark trace; peak memory 177.8 GB at this context; duplicate traces of the same session: 30isal8m, ctdugq6s, dt92wa32; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 512.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-Abliterated-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 512.2 46.8 177.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1pirzdfi",
   "source_url": "https://omlx.ai/benchmarks/performance/1pirzdfi",
   "tg_tokens": null,
   "tps": 46.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1pirzdfi-ctx4096",
   "model": "GLM-5.3-Flash-Abliterated",
   "notes": "MLX; oMLX community benchmark trace; peak memory 178.9 GB at this context; duplicate traces of the same session: 30isal8m, ctdugq6s, dt92wa32; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 593.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-Abliterated-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 593.8 43.9 178.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1pirzdfi",
   "source_url": "https://omlx.ai/benchmarks/performance/1pirzdfi",
   "tg_tokens": null,
   "tps": 43.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1pirzdfi-ctx8192",
   "model": "GLM-5.3-Flash-Abliterated",
   "notes": "MLX; oMLX community benchmark trace; peak memory 179.3 GB at this context; duplicate traces of the same session: 30isal8m, ctdugq6s, dt92wa32; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 550.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-Abliterated-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 550.8 48.8 179.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1pirzdfi",
   "source_url": "https://omlx.ai/benchmarks/performance/1pirzdfi",
   "tg_tokens": null,
   "tps": 48.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1pirzdfi-ctx16384",
   "model": "GLM-5.3-Flash-Abliterated",
   "notes": "MLX; oMLX community benchmark trace; peak memory 179.4 GB at this context; duplicate traces of the same session: 30isal8m, ctdugq6s, dt92wa32; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 540.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-Abliterated-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 540.6 85.6 179.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1pirzdfi",
   "source_url": "https://omlx.ai/benchmarks/performance/1pirzdfi",
   "tg_tokens": null,
   "tps": 85.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bbh7r8fy-ctx1024",
   "model": "GLM-5.3-Flash-Abliterated",
   "notes": "MLX; oMLX community benchmark trace; peak memory 177.8 GB at this context; duplicate traces of the same session: oknxysom, puefqnh0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 513.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-Abliterated-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 513.7 48.4 177.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bbh7r8fy",
   "source_url": "https://omlx.ai/benchmarks/performance/bbh7r8fy",
   "tg_tokens": null,
   "tps": 48.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bbh7r8fy-ctx4096",
   "model": "GLM-5.3-Flash-Abliterated",
   "notes": "MLX; oMLX community benchmark trace; peak memory 178.9 GB at this context; duplicate traces of the same session: oknxysom, puefqnh0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 594.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-Abliterated-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 594.9 43.8 178.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bbh7r8fy",
   "source_url": "https://omlx.ai/benchmarks/performance/bbh7r8fy",
   "tg_tokens": null,
   "tps": 43.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bbh7r8fy-ctx8192",
   "model": "GLM-5.3-Flash-Abliterated",
   "notes": "MLX; oMLX community benchmark trace; peak memory 179.3 GB at this context; duplicate traces of the same session: oknxysom, puefqnh0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 551.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-Abliterated-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 551.3 50.1 179.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bbh7r8fy",
   "source_url": "https://omlx.ai/benchmarks/performance/bbh7r8fy",
   "tg_tokens": null,
   "tps": 50.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bbh7r8fy-ctx16384",
   "model": "GLM-5.3-Flash-Abliterated",
   "notes": "MLX; oMLX community benchmark trace; peak memory 179.4 GB at this context; duplicate traces of the same session: oknxysom, puefqnh0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 541.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-Abliterated-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 541.6 85.5 179.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bbh7r8fy",
   "source_url": "https://omlx.ai/benchmarks/performance/bbh7r8fy",
   "tg_tokens": null,
   "tps": 85.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-i3p220ef-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 172.4 GB at this context; duplicate traces of the same session: 135hb18k, 6c8bw9ny, a74tnelm, c3q60m8n, epkier0o, tg3qatmt, uqa9sqo2; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1306.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,306 77.1 172.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i3p220ef",
   "source_url": "https://omlx.ai/benchmarks/performance/i3p220ef",
   "tg_tokens": null,
   "tps": 77.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-i3p220ef-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 173.2 GB at this context; duplicate traces of the same session: 135hb18k, 6c8bw9ny, a74tnelm, c3q60m8n, epkier0o, tg3qatmt, uqa9sqo2; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2029.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 2,029 70.2 173.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i3p220ef",
   "source_url": "https://omlx.ai/benchmarks/performance/i3p220ef",
   "tg_tokens": null,
   "tps": 70.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-i3p220ef-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 174.2 GB at this context; duplicate traces of the same session: 135hb18k, 6c8bw9ny, a74tnelm, c3q60m8n, epkier0o, tg3qatmt, uqa9sqo2; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2013.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 2,013 70.9 174.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i3p220ef",
   "source_url": "https://omlx.ai/benchmarks/performance/i3p220ef",
   "tg_tokens": null,
   "tps": 70.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-i3p220ef-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 174.3 GB at this context; duplicate traces of the same session: 135hb18k, 6c8bw9ny, a74tnelm, c3q60m8n, epkier0o, tg3qatmt, uqa9sqo2; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1996.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,996 70.2 174.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i3p220ef",
   "source_url": "https://omlx.ai/benchmarks/performance/i3p220ef",
   "tg_tokens": null,
   "tps": 70.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-i3p220ef-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 174.5 GB at this context; duplicate traces of the same session: 135hb18k, 6c8bw9ny, a74tnelm, c3q60m8n, epkier0o, tg3qatmt, uqa9sqo2; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1986.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,986 69.7 174.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i3p220ef",
   "source_url": "https://omlx.ai/benchmarks/performance/i3p220ef",
   "tg_tokens": null,
   "tps": 69.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-i3p220ef-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 174.8 GB at this context; duplicate traces of the same session: 135hb18k, 6c8bw9ny, a74tnelm, c3q60m8n, epkier0o, tg3qatmt, uqa9sqo2; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1969.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,969 68.5 174.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i3p220ef",
   "source_url": "https://omlx.ai/benchmarks/performance/i3p220ef",
   "tg_tokens": null,
   "tps": 68.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-i3p220ef-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 175.5 GB at this context; duplicate traces of the same session: 135hb18k, 6c8bw9ny, a74tnelm, c3q60m8n, epkier0o, tg3qatmt, uqa9sqo2; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1921.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 1,921 64.4 175.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i3p220ef",
   "source_url": "https://omlx.ai/benchmarks/performance/i3p220ef",
   "tg_tokens": null,
   "tps": 64.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-i3p220ef-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 176.3 GB at this context; duplicate traces of the same session: 135hb18k, 6c8bw9ny, a74tnelm, c3q60m8n, epkier0o, tg3qatmt, uqa9sqo2; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1880.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 1,880 61.0 176.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i3p220ef",
   "source_url": "https://omlx.ai/benchmarks/performance/i3p220ef",
   "tg_tokens": null,
   "tps": 61.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-dc7qee61-ctx1024",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 177.6 GB at this context; duplicate traces of the same session: 37gecy2v, o5c79ch4, u10dqe8o; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 525.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 525.3 40.0 177.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dc7qee61",
   "source_url": "https://omlx.ai/benchmarks/performance/dc7qee61",
   "tg_tokens": null,
   "tps": 40.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-dc7qee61-ctx4096",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 178.8 GB at this context; duplicate traces of the same session: 37gecy2v, o5c79ch4, u10dqe8o; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 606.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 606.2 38.3 178.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dc7qee61",
   "source_url": "https://omlx.ai/benchmarks/performance/dc7qee61",
   "tg_tokens": null,
   "tps": 38.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-dc7qee61-ctx8192",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 179.4 GB at this context; duplicate traces of the same session: 37gecy2v, o5c79ch4, u10dqe8o; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 561.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 561.2 38.4 179.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dc7qee61",
   "source_url": "https://omlx.ai/benchmarks/performance/dc7qee61",
   "tg_tokens": null,
   "tps": 38.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-dc7qee61-ctx16384",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 179.5 GB at this context; duplicate traces of the same session: 37gecy2v, o5c79ch4, u10dqe8o; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 550.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 550.1 38.0 179.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dc7qee61",
   "source_url": "https://omlx.ai/benchmarks/performance/dc7qee61",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ez0530ep-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 180.2 GB; TTFT 35151 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1864.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k tokens 1,864 PP tok/s 66.7 TG tok/s 35151 TTFT (ms) 180.2 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ez0530ep",
   "source_url": "https://omlx.ai/benchmarks/performance/ez0530ep",
   "tg_tokens": null,
   "tps": 66.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-658pv1br-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: 9tuagt6u, f9wjdco8, g2nqok03, myoo14ud, scp7wj7f, y2qade2j, y30chns4; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1262.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,262 62.1 177.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/658pv1br",
   "source_url": "https://omlx.ai/benchmarks/performance/658pv1br",
   "tg_tokens": null,
   "tps": 62.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-658pv1br-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: 9tuagt6u, f9wjdco8, g2nqok03, myoo14ud, scp7wj7f, y2qade2j, y30chns4; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1973.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,973 57.2 178.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/658pv1br",
   "source_url": "https://omlx.ai/benchmarks/performance/658pv1br",
   "tg_tokens": null,
   "tps": 57.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-658pv1br-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.3 GB at this context; duplicate traces of the same session: 9tuagt6u, f9wjdco8, g2nqok03, myoo14ud, scp7wj7f, y2qade2j, y30chns4; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1969.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 1,969 57.4 179.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/658pv1br",
   "source_url": "https://omlx.ai/benchmarks/performance/658pv1br",
   "tg_tokens": null,
   "tps": 57.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-658pv1br-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.3 GB at this context; duplicate traces of the same session: 9tuagt6u, f9wjdco8, g2nqok03, myoo14ud, scp7wj7f, y2qade2j, y30chns4; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1964.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,964 56.9 179.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/658pv1br",
   "source_url": "https://omlx.ai/benchmarks/performance/658pv1br",
   "tg_tokens": null,
   "tps": 56.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-658pv1br-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.6 GB at this context; duplicate traces of the same session: 9tuagt6u, f9wjdco8, g2nqok03, myoo14ud, scp7wj7f, y2qade2j, y30chns4; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1944.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,944 56.9 179.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/658pv1br",
   "source_url": "https://omlx.ai/benchmarks/performance/658pv1br",
   "tg_tokens": null,
   "tps": 56.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-658pv1br-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.9 GB at this context; duplicate traces of the same session: 9tuagt6u, f9wjdco8, g2nqok03, myoo14ud, scp7wj7f, y2qade2j, y30chns4; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1923.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,923 55.7 179.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/658pv1br",
   "source_url": "https://omlx.ai/benchmarks/performance/658pv1br",
   "tg_tokens": null,
   "tps": 55.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-658pv1br-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.6 GB at this context; duplicate traces of the same session: 9tuagt6u, f9wjdco8, g2nqok03, myoo14ud, scp7wj7f, y2qade2j, y30chns4; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1883.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 1,883 53.5 180.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/658pv1br",
   "source_url": "https://omlx.ai/benchmarks/performance/658pv1br",
   "tg_tokens": null,
   "tps": 53.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-658pv1br-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.4 GB at this context; duplicate traces of the same session: 9tuagt6u, f9wjdco8, g2nqok03, myoo14ud, scp7wj7f, y2qade2j, y30chns4; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1839.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 1,839 50.7 181.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/658pv1br",
   "source_url": "https://omlx.ai/benchmarks/performance/658pv1br",
   "tg_tokens": null,
   "tps": 50.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-lyujvegu-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: 3xml1eb6, cye0z4k5, deo6kmgk, qvoq9tcl, urnzj0zn; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1274.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,274 72.2 177.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lyujvegu",
   "source_url": "https://omlx.ai/benchmarks/performance/lyujvegu",
   "tg_tokens": null,
   "tps": 72.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-lyujvegu-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: 3xml1eb6, cye0z4k5, deo6kmgk, qvoq9tcl, urnzj0zn; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1926.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,926 70.9 178.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lyujvegu",
   "source_url": "https://omlx.ai/benchmarks/performance/lyujvegu",
   "tg_tokens": null,
   "tps": 70.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-lyujvegu-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.9 GB at this context; duplicate traces of the same session: 3xml1eb6, cye0z4k5, deo6kmgk, qvoq9tcl, urnzj0zn; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1929.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 1,929 85.6 178.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lyujvegu",
   "source_url": "https://omlx.ai/benchmarks/performance/lyujvegu",
   "tg_tokens": null,
   "tps": 85.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-lyujvegu-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.1 GB at this context; duplicate traces of the same session: 3xml1eb6, cye0z4k5, deo6kmgk, qvoq9tcl, urnzj0zn; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1922.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,922 139.0 179.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lyujvegu",
   "source_url": "https://omlx.ai/benchmarks/performance/lyujvegu",
   "tg_tokens": null,
   "tps": 139.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-lyujvegu-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.5 GB at this context; duplicate traces of the same session: 3xml1eb6, cye0z4k5, deo6kmgk, qvoq9tcl, urnzj0zn; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1914.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,914 101.5 179.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lyujvegu",
   "source_url": "https://omlx.ai/benchmarks/performance/lyujvegu",
   "tg_tokens": null,
   "tps": 101.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-lyujvegu-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.1 GB at this context; duplicate traces of the same session: 3xml1eb6, cye0z4k5, deo6kmgk, qvoq9tcl, urnzj0zn; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1867.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,867 72.3 180.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lyujvegu",
   "source_url": "https://omlx.ai/benchmarks/performance/lyujvegu",
   "tg_tokens": null,
   "tps": 72.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-lyujvegu-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.3 GB at this context; duplicate traces of the same session: 3xml1eb6, cye0z4k5, deo6kmgk, qvoq9tcl, urnzj0zn; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1844.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 1,844 118.4 181.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lyujvegu",
   "source_url": "https://omlx.ai/benchmarks/performance/lyujvegu",
   "tg_tokens": null,
   "tps": 118.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-lyujvegu-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 182.6 GB at this context; duplicate traces of the same session: 3xml1eb6, cye0z4k5, deo6kmgk, qvoq9tcl, urnzj0zn; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1806.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 1,806 64.4 182.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lyujvegu",
   "source_url": "https://omlx.ai/benchmarks/performance/lyujvegu",
   "tg_tokens": null,
   "tps": 64.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-x4a9t9t6-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1680.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 1,680 158.3 37.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x4a9t9t6",
   "source_url": "https://omlx.ai/benchmarks/performance/x4a9t9t6",
   "tg_tokens": null,
   "tps": 158.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-x4a9t9t6-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1859.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 1,859 136.3 39.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x4a9t9t6",
   "source_url": "https://omlx.ai/benchmarks/performance/x4a9t9t6",
   "tg_tokens": null,
   "tps": 136.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-5vykdw5h-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: rdt82tn9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1439.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 1,439 78.0 37.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5vykdw5h",
   "source_url": "https://omlx.ai/benchmarks/performance/5vykdw5h",
   "tg_tokens": null,
   "tps": 78.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-5vykdw5h-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.5 GB at this context; duplicate traces of the same session: rdt82tn9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1859.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 1,859 116.1 39.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5vykdw5h",
   "source_url": "https://omlx.ai/benchmarks/performance/5vykdw5h",
   "tg_tokens": null,
   "tps": 116.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-y8dtwyf9-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: s06hyfls; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1579.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 1,579 96.5 35.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y8dtwyf9",
   "source_url": "https://omlx.ai/benchmarks/performance/y8dtwyf9",
   "tg_tokens": null,
   "tps": 96.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-y8dtwyf9-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.5 GB at this context; duplicate traces of the same session: s06hyfls; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1923.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 1,923 94.7 37.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y8dtwyf9",
   "source_url": "https://omlx.ai/benchmarks/performance/y8dtwyf9",
   "tg_tokens": null,
   "tps": 94.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-9szd01fg-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: d8ze0r91; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 486.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 486.8 17.9 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9szd01fg",
   "source_url": "https://omlx.ai/benchmarks/performance/9szd01fg",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-9szd01fg-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: d8ze0r91; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 514.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 514.9 17.4 17.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9szd01fg",
   "source_url": "https://omlx.ai/benchmarks/performance/9szd01fg",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-txo1fdo2-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: j0ahm9wd, kmxubcw6, unucvxa7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 945.5,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 945.5 55.7 24.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/txo1fdo2",
   "source_url": "https://omlx.ai/benchmarks/performance/txo1fdo2",
   "tg_tokens": null,
   "tps": 55.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-txo1fdo2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: j0ahm9wd, kmxubcw6, unucvxa7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1055.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 1,055 54.1 25.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/txo1fdo2",
   "source_url": "https://omlx.ai/benchmarks/performance/txo1fdo2",
   "tg_tokens": null,
   "tps": 54.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-txo1fdo2-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: j0ahm9wd, kmxubcw6, unucvxa7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1075.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 1,075 59.0 26.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/txo1fdo2",
   "source_url": "https://omlx.ai/benchmarks/performance/txo1fdo2",
   "tg_tokens": null,
   "tps": 59.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-txo1fdo2-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: j0ahm9wd, kmxubcw6, unucvxa7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1034.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 1,034 78.4 27.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/txo1fdo2",
   "source_url": "https://omlx.ai/benchmarks/performance/txo1fdo2",
   "tg_tokens": null,
   "tps": 78.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-txo1fdo2-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.8 GB at this context; duplicate traces of the same session: j0ahm9wd, kmxubcw6, unucvxa7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 948.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 948.9 115.2 29.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/txo1fdo2",
   "source_url": "https://omlx.ai/benchmarks/performance/txo1fdo2",
   "tg_tokens": null,
   "tps": 115.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hbq0t7lf-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 75.3 GB; TTFT 1539 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 665.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 665.5 PP tok/s 39.6 TG tok/s 1539 TTFT (ms) 75.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hbq0t7lf",
   "source_url": "https://omlx.ai/benchmarks/performance/hbq0t7lf",
   "tg_tokens": null,
   "tps": 39.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-43xdjaz2-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 78.5 GB; TTFT 56390 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2324.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k tokens 2,324 PP tok/s 56.7 TG tok/s 56390 TTFT (ms) 78.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/43xdjaz2",
   "source_url": "https://omlx.ai/benchmarks/performance/43xdjaz2",
   "tg_tokens": null,
   "tps": 56.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-kxhxz2ls-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 99.0 GB; TTFT 2064 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 496.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k tokens 496.2 PP tok/s 129.3 TG tok/s 2064 TTFT (ms) 99.0 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kxhxz2ls",
   "source_url": "https://omlx.ai/benchmarks/performance/kxhxz2ls",
   "tg_tokens": null,
   "tps": 129.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-nvsdn4ti-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace, single run at 1k context; peak memory 14.3 GB; TTFT 453 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2260.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k tokens 2,260 PP tok/s 79.4 TG tok/s 453 TTFT (ms) 14.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nvsdn4ti",
   "source_url": "https://omlx.ai/benchmarks/performance/nvsdn4ti",
   "tg_tokens": null,
   "tps": 79.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-zkf1z6wy-ctx1024",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: y77mpca4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 207.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31B-it-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 207.4 23.4 17.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zkf1z6wy",
   "source_url": "https://omlx.ai/benchmarks/performance/zkf1z6wy",
   "tg_tokens": null,
   "tps": 23.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-zkf1z6wy-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: y77mpca4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 188.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31B-it-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 188.3 23.1 18.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zkf1z6wy",
   "source_url": "https://omlx.ai/benchmarks/performance/zkf1z6wy",
   "tg_tokens": null,
   "tps": 23.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hgr53vji-ctx1024",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 34.8 GB; TTFT 1781 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 575.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k tokens 575.1 PP tok/s 30.1 TG tok/s 1781 TTFT (ms) 34.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hgr53vji",
   "source_url": "https://omlx.ai/benchmarks/performance/hgr53vji",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-8rhkc8jz-ctx4096",
   "model": "Qwen3.6-27B",
   "notes": "mlx-community OptiQ 4bit build; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: 6g209kq6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 132.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-27B-OptiQ-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 132.3 12.8 20.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8rhkc8jz",
   "source_url": "https://omlx.ai/benchmarks/performance/8rhkc8jz",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-8rhkc8jz-ctx16384",
   "model": "Qwen3.6-27B",
   "notes": "mlx-community OptiQ 4bit build; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 6g209kq6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-27B-OptiQ-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 127.1 12.1 21.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8rhkc8jz",
   "source_url": "https://omlx.ai/benchmarks/performance/8rhkc8jz",
   "tg_tokens": null,
   "tps": 12.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-3fxwe57n-ctx1024",
   "model": "Occamy-1.0",
   "notes": "oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: 2f9oj5ek, 7s2b50qh, g2kwkyo2, os6ub61o, r20b436j; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 640.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "occamy-1.0-MLX-6bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 1k 640.7 61.5 27.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3fxwe57n",
   "source_url": "https://omlx.ai/benchmarks/performance/3fxwe57n",
   "tg_tokens": null,
   "tps": 61.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-3fxwe57n-ctx4096",
   "model": "Occamy-1.0",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: 2f9oj5ek, 7s2b50qh, g2kwkyo2, os6ub61o, r20b436j; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 725.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "occamy-1.0-MLX-6bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 4k 725.3 59.3 29.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3fxwe57n",
   "source_url": "https://omlx.ai/benchmarks/performance/3fxwe57n",
   "tg_tokens": null,
   "tps": 59.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-3fxwe57n-ctx8192",
   "model": "Occamy-1.0",
   "notes": "oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: 2f9oj5ek, 7s2b50qh, g2kwkyo2, os6ub61o, r20b436j; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 677.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "occamy-1.0-MLX-6bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 8k 677.3 57.0 29.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3fxwe57n",
   "source_url": "https://omlx.ai/benchmarks/performance/3fxwe57n",
   "tg_tokens": null,
   "tps": 57.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-3fxwe57n-ctx16384",
   "model": "Occamy-1.0",
   "notes": "oMLX community benchmark trace; peak memory 30.8 GB at this context; duplicate traces of the same session: 2f9oj5ek, 7s2b50qh, g2kwkyo2, os6ub61o, r20b436j; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 583.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "occamy-1.0-MLX-6bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 16k 583.9 53.5 30.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3fxwe57n",
   "source_url": "https://omlx.ai/benchmarks/performance/3fxwe57n",
   "tg_tokens": null,
   "tps": 53.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-3fxwe57n-ctx32768",
   "model": "Occamy-1.0",
   "notes": "oMLX community benchmark trace; peak memory 31.7 GB at this context; duplicate traces of the same session: 2f9oj5ek, 7s2b50qh, g2kwkyo2, os6ub61o, r20b436j; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 456.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "occamy-1.0-MLX-6bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 32k 456.5 48.5 31.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3fxwe57n",
   "source_url": "https://omlx.ai/benchmarks/performance/3fxwe57n",
   "tg_tokens": null,
   "tps": 48.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-3fxwe57n-ctx65536",
   "model": "Occamy-1.0",
   "notes": "oMLX community benchmark trace; peak memory 32.3 GB at this context; duplicate traces of the same session: 2f9oj5ek, 7s2b50qh, g2kwkyo2, os6ub61o, r20b436j; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 319.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "occamy-1.0-MLX-6bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 64k 319.3 38.8 32.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3fxwe57n",
   "source_url": "https://omlx.ai/benchmarks/performance/3fxwe57n",
   "tg_tokens": null,
   "tps": 38.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-njxhym3z-ctx1024",
   "model": "GLM-5.3-Flash-Abliterated",
   "notes": "MLX; oMLX community benchmark trace; peak memory 177.8 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 513.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-Abliterated-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 513.7 48.4 177.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/njxhym3z",
   "source_url": "https://omlx.ai/benchmarks/performance/njxhym3z",
   "tg_tokens": null,
   "tps": 48.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-njxhym3z-ctx4096",
   "model": "GLM-5.3-Flash-Abliterated",
   "notes": "MLX; oMLX community benchmark trace; peak memory 178.9 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 594.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-Abliterated-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 594.9 43.8 178.9 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/njxhym3z",
   "source_url": "https://omlx.ai/benchmarks/performance/njxhym3z",
   "tg_tokens": null,
   "tps": 43.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-njxhym3z-ctx8192",
   "model": "GLM-5.3-Flash-Abliterated",
   "notes": "MLX; oMLX community benchmark trace; peak memory 179.3 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 551.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-Abliterated-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 551.3 50.1 179.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/njxhym3z",
   "source_url": "https://omlx.ai/benchmarks/performance/njxhym3z",
   "tg_tokens": null,
   "tps": 50.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-njxhym3z-ctx16384",
   "model": "GLM-5.3-Flash-Abliterated",
   "notes": "MLX; oMLX community benchmark trace; peak memory 179.4 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 541.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-Abliterated-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 541.6 85.5 179.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/njxhym3z",
   "source_url": "https://omlx.ai/benchmarks/performance/njxhym3z",
   "tg_tokens": null,
   "tps": 85.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3qjjypkn-ctx1024",
   "model": "GLM-5.3-Flash-Abliterated",
   "notes": "MLX; oMLX community benchmark trace; peak memory 177.8 GB at this context; duplicate traces of the same session: 2fbbmbk0, c0jboud6, zoo65rof; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 514.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-Abliterated-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 514.7 52.6 177.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3qjjypkn",
   "source_url": "https://omlx.ai/benchmarks/performance/3qjjypkn",
   "tg_tokens": null,
   "tps": 52.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3qjjypkn-ctx4096",
   "model": "GLM-5.3-Flash-Abliterated",
   "notes": "MLX; oMLX community benchmark trace; peak memory 178.9 GB at this context; duplicate traces of the same session: 2fbbmbk0, c0jboud6, zoo65rof; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 587.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-Abliterated-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 587.2 45.1 178.9 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3qjjypkn",
   "source_url": "https://omlx.ai/benchmarks/performance/3qjjypkn",
   "tg_tokens": null,
   "tps": 45.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3qjjypkn-ctx8192",
   "model": "GLM-5.3-Flash-Abliterated",
   "notes": "MLX; oMLX community benchmark trace; peak memory 179.3 GB at this context; duplicate traces of the same session: 2fbbmbk0, c0jboud6, zoo65rof; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 552.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-Abliterated-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 552.3 49.0 179.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3qjjypkn",
   "source_url": "https://omlx.ai/benchmarks/performance/3qjjypkn",
   "tg_tokens": null,
   "tps": 49.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3qjjypkn-ctx16384",
   "model": "GLM-5.3-Flash-Abliterated",
   "notes": "MLX; oMLX community benchmark trace; peak memory 179.4 GB at this context; duplicate traces of the same session: 2fbbmbk0, c0jboud6, zoo65rof; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 541.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-Abliterated-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 541.3 88.0 179.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3qjjypkn",
   "source_url": "https://omlx.ai/benchmarks/performance/3qjjypkn",
   "tg_tokens": null,
   "tps": 88.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-8dgxsivx-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.2 GB at this context; duplicate traces of the same session: 6k10wc07, a87ihu23, amgnfa57, bk79o46v, gnrq0yd6, mt7jqove, vn97ut5s; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1233.0,
   "provenance": "community",
   "quant": "mixed 4/8bit",
   "quote": "GLM-5.3-Flash-MLX-mixed-4_8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,233 60.2 174.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8dgxsivx",
   "source_url": "https://omlx.ai/benchmarks/performance/8dgxsivx",
   "tg_tokens": null,
   "tps": 60.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-8dgxsivx-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.0 GB at this context; duplicate traces of the same session: 6k10wc07, a87ihu23, amgnfa57, bk79o46v, gnrq0yd6, mt7jqove, vn97ut5s; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1978.0,
   "provenance": "community",
   "quant": "mixed 4/8bit",
   "quote": "GLM-5.3-Flash-MLX-mixed-4_8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,978 55.0 175.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8dgxsivx",
   "source_url": "https://omlx.ai/benchmarks/performance/8dgxsivx",
   "tg_tokens": null,
   "tps": 55.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-8dgxsivx-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 176.0 GB at this context; duplicate traces of the same session: 6k10wc07, a87ihu23, amgnfa57, bk79o46v, gnrq0yd6, mt7jqove, vn97ut5s; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1970.0,
   "provenance": "community",
   "quant": "mixed 4/8bit",
   "quote": "GLM-5.3-Flash-MLX-mixed-4_8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 1,970 55.3 176.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8dgxsivx",
   "source_url": "https://omlx.ai/benchmarks/performance/8dgxsivx",
   "tg_tokens": null,
   "tps": 55.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-8dgxsivx-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 176.1 GB at this context; duplicate traces of the same session: 6k10wc07, a87ihu23, amgnfa57, bk79o46v, gnrq0yd6, mt7jqove, vn97ut5s; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1954.0,
   "provenance": "community",
   "quant": "mixed 4/8bit",
   "quote": "GLM-5.3-Flash-MLX-mixed-4_8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,954 55.5 176.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8dgxsivx",
   "source_url": "https://omlx.ai/benchmarks/performance/8dgxsivx",
   "tg_tokens": null,
   "tps": 55.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-8dgxsivx-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 176.3 GB at this context; duplicate traces of the same session: 6k10wc07, a87ihu23, amgnfa57, bk79o46v, gnrq0yd6, mt7jqove, vn97ut5s; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1950.0,
   "provenance": "community",
   "quant": "mixed 4/8bit",
   "quote": "GLM-5.3-Flash-MLX-mixed-4_8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,950 54.2 176.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8dgxsivx",
   "source_url": "https://omlx.ai/benchmarks/performance/8dgxsivx",
   "tg_tokens": null,
   "tps": 54.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-8dgxsivx-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 176.7 GB at this context; duplicate traces of the same session: 6k10wc07, a87ihu23, amgnfa57, bk79o46v, gnrq0yd6, mt7jqove, vn97ut5s; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1924.0,
   "provenance": "community",
   "quant": "mixed 4/8bit",
   "quote": "GLM-5.3-Flash-MLX-mixed-4_8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,924 53.7 176.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8dgxsivx",
   "source_url": "https://omlx.ai/benchmarks/performance/8dgxsivx",
   "tg_tokens": null,
   "tps": 53.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-8dgxsivx-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.3 GB at this context; duplicate traces of the same session: 6k10wc07, a87ihu23, amgnfa57, bk79o46v, gnrq0yd6, mt7jqove, vn97ut5s; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1890.0,
   "provenance": "community",
   "quant": "mixed 4/8bit",
   "quote": "GLM-5.3-Flash-MLX-mixed-4_8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 1,890 52.1 177.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8dgxsivx",
   "source_url": "https://omlx.ai/benchmarks/performance/8dgxsivx",
   "tg_tokens": null,
   "tps": 52.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-8dgxsivx-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.1 GB at this context; duplicate traces of the same session: 6k10wc07, a87ihu23, amgnfa57, bk79o46v, gnrq0yd6, mt7jqove, vn97ut5s; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1846.0,
   "provenance": "community",
   "quant": "mixed 4/8bit",
   "quote": "GLM-5.3-Flash-MLX-mixed-4_8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 1,846 49.3 178.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8dgxsivx",
   "source_url": "https://omlx.ai/benchmarks/performance/8dgxsivx",
   "tg_tokens": null,
   "tps": 49.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-vggts813-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: urh0gz16; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1274.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,274 72.2 177.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vggts813",
   "source_url": "https://omlx.ai/benchmarks/performance/vggts813",
   "tg_tokens": null,
   "tps": 72.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-vggts813-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: urh0gz16; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1926.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,926 70.9 178.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vggts813",
   "source_url": "https://omlx.ai/benchmarks/performance/vggts813",
   "tg_tokens": null,
   "tps": 70.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-vggts813-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.9 GB at this context; duplicate traces of the same session: urh0gz16; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1929.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 1,929 85.6 178.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vggts813",
   "source_url": "https://omlx.ai/benchmarks/performance/vggts813",
   "tg_tokens": null,
   "tps": 85.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-vggts813-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.1 GB at this context; duplicate traces of the same session: urh0gz16; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1922.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,922 139.0 179.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vggts813",
   "source_url": "https://omlx.ai/benchmarks/performance/vggts813",
   "tg_tokens": null,
   "tps": 139.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-vggts813-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.5 GB at this context; duplicate traces of the same session: urh0gz16; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1914.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,914 101.5 179.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vggts813",
   "source_url": "https://omlx.ai/benchmarks/performance/vggts813",
   "tg_tokens": null,
   "tps": 101.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-vggts813-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.1 GB at this context; duplicate traces of the same session: urh0gz16; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1867.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,867 72.3 180.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vggts813",
   "source_url": "https://omlx.ai/benchmarks/performance/vggts813",
   "tg_tokens": null,
   "tps": 72.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-vggts813-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.3 GB at this context; duplicate traces of the same session: urh0gz16; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1844.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 1,844 118.4 181.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vggts813",
   "source_url": "https://omlx.ai/benchmarks/performance/vggts813",
   "tg_tokens": null,
   "tps": 118.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-vggts813-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 182.6 GB at this context; duplicate traces of the same session: urh0gz16; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1806.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 1,806 64.4 182.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vggts813",
   "source_url": "https://omlx.ai/benchmarks/performance/vggts813",
   "tg_tokens": null,
   "tps": 64.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-a210j6nr-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: a8uvdxrg, dpn2hqyn, n18t64ou, t2j66c1s, tgrf9atn, wljyuway; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1250.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,250 73.6 177.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a210j6nr",
   "source_url": "https://omlx.ai/benchmarks/performance/a210j6nr",
   "tg_tokens": null,
   "tps": 73.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-a210j6nr-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: a8uvdxrg, dpn2hqyn, n18t64ou, t2j66c1s, tgrf9atn, wljyuway; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1919.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,919 58.9 178.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a210j6nr",
   "source_url": "https://omlx.ai/benchmarks/performance/a210j6nr",
   "tg_tokens": null,
   "tps": 58.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-a210j6nr-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.9 GB at this context; duplicate traces of the same session: a8uvdxrg, dpn2hqyn, n18t64ou, t2j66c1s, tgrf9atn, wljyuway; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1924.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 1,924 68.4 178.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a210j6nr",
   "source_url": "https://omlx.ai/benchmarks/performance/a210j6nr",
   "tg_tokens": null,
   "tps": 68.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-a210j6nr-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.1 GB at this context; duplicate traces of the same session: a8uvdxrg, dpn2hqyn, n18t64ou, t2j66c1s, tgrf9atn, wljyuway; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1923.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,923 138.6 179.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a210j6nr",
   "source_url": "https://omlx.ai/benchmarks/performance/a210j6nr",
   "tg_tokens": null,
   "tps": 138.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-a210j6nr-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.5 GB at this context; duplicate traces of the same session: a8uvdxrg, dpn2hqyn, n18t64ou, t2j66c1s, tgrf9atn, wljyuway; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1912.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,912 98.1 179.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a210j6nr",
   "source_url": "https://omlx.ai/benchmarks/performance/a210j6nr",
   "tg_tokens": null,
   "tps": 98.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-a210j6nr-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.3 GB at this context; duplicate traces of the same session: a8uvdxrg, dpn2hqyn, n18t64ou, t2j66c1s, tgrf9atn, wljyuway; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1845.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 1,845 131.2 181.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a210j6nr",
   "source_url": "https://omlx.ai/benchmarks/performance/a210j6nr",
   "tg_tokens": null,
   "tps": 131.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-a210j6nr-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 182.6 GB at this context; duplicate traces of the same session: a8uvdxrg, dpn2hqyn, n18t64ou, t2j66c1s, tgrf9atn, wljyuway; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1803.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 1,803 64.9 182.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a210j6nr",
   "source_url": "https://omlx.ai/benchmarks/performance/a210j6nr",
   "tg_tokens": null,
   "tps": 64.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-2gfk7aox-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: 14l1g8tm; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1261.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,261 68.7 177.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2gfk7aox",
   "source_url": "https://omlx.ai/benchmarks/performance/2gfk7aox",
   "tg_tokens": null,
   "tps": 68.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-2gfk7aox-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: 14l1g8tm; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1939.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,939 64.4 178.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2gfk7aox",
   "source_url": "https://omlx.ai/benchmarks/performance/2gfk7aox",
   "tg_tokens": null,
   "tps": 64.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-2gfk7aox-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.9 GB at this context; duplicate traces of the same session: 14l1g8tm; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1931.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 1,931 75.9 178.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2gfk7aox",
   "source_url": "https://omlx.ai/benchmarks/performance/2gfk7aox",
   "tg_tokens": null,
   "tps": 75.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-2gfk7aox-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.1 GB at this context; duplicate traces of the same session: 14l1g8tm; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1917.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,917 138.8 179.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2gfk7aox",
   "source_url": "https://omlx.ai/benchmarks/performance/2gfk7aox",
   "tg_tokens": null,
   "tps": 138.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-2gfk7aox-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.5 GB at this context; duplicate traces of the same session: 14l1g8tm; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1914.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,914 94.3 179.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2gfk7aox",
   "source_url": "https://omlx.ai/benchmarks/performance/2gfk7aox",
   "tg_tokens": null,
   "tps": 94.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-2gfk7aox-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.1 GB at this context; duplicate traces of the same session: 14l1g8tm; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1885.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,885 71.1 180.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2gfk7aox",
   "source_url": "https://omlx.ai/benchmarks/performance/2gfk7aox",
   "tg_tokens": null,
   "tps": 71.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-2gfk7aox-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 182.6 GB at this context; duplicate traces of the same session: 14l1g8tm; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1807.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 1,807 65.4 182.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2gfk7aox",
   "source_url": "https://omlx.ai/benchmarks/performance/2gfk7aox",
   "tg_tokens": null,
   "tps": 65.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ik6cxeei-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: 3dnwvvf2, 3hlqlio5, 3pmfrvpv, romjeots; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1260.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,260 62.0 177.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ik6cxeei",
   "source_url": "https://omlx.ai/benchmarks/performance/ik6cxeei",
   "tg_tokens": null,
   "tps": 62.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ik6cxeei-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: 3dnwvvf2, 3hlqlio5, 3pmfrvpv, romjeots; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1982.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,982 57.1 178.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ik6cxeei",
   "source_url": "https://omlx.ai/benchmarks/performance/ik6cxeei",
   "tg_tokens": null,
   "tps": 57.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ik6cxeei-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.3 GB at this context; duplicate traces of the same session: 3dnwvvf2, 3hlqlio5, 3pmfrvpv, romjeots; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1969.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 1,969 57.5 179.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ik6cxeei",
   "source_url": "https://omlx.ai/benchmarks/performance/ik6cxeei",
   "tg_tokens": null,
   "tps": 57.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ik6cxeei-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.3 GB at this context; duplicate traces of the same session: 3dnwvvf2, 3hlqlio5, 3pmfrvpv, romjeots; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1964.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,964 57.1 179.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ik6cxeei",
   "source_url": "https://omlx.ai/benchmarks/performance/ik6cxeei",
   "tg_tokens": null,
   "tps": 57.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ik6cxeei-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.6 GB at this context; duplicate traces of the same session: 3dnwvvf2, 3hlqlio5, 3pmfrvpv, romjeots; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1952.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,952 56.9 179.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ik6cxeei",
   "source_url": "https://omlx.ai/benchmarks/performance/ik6cxeei",
   "tg_tokens": null,
   "tps": 56.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ik6cxeei-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.9 GB at this context; duplicate traces of the same session: 3dnwvvf2, 3hlqlio5, 3pmfrvpv, romjeots; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1924.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,924 55.8 179.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ik6cxeei",
   "source_url": "https://omlx.ai/benchmarks/performance/ik6cxeei",
   "tg_tokens": null,
   "tps": 55.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ik6cxeei-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.6 GB at this context; duplicate traces of the same session: 3dnwvvf2, 3hlqlio5, 3pmfrvpv, romjeots; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1887.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 1,887 53.7 180.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ik6cxeei",
   "source_url": "https://omlx.ai/benchmarks/performance/ik6cxeei",
   "tg_tokens": null,
   "tps": 53.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ik6cxeei-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.4 GB at this context; duplicate traces of the same session: 3dnwvvf2, 3hlqlio5, 3pmfrvpv, romjeots; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1843.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 1,843 50.5 181.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ik6cxeei",
   "source_url": "https://omlx.ai/benchmarks/performance/ik6cxeei",
   "tg_tokens": null,
   "tps": 50.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-x0v8etzs-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 182.7 GB at this context; duplicate traces of the same session: 4ea2qh58, ptk439dv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2352.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,352 114.2 182.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x0v8etzs",
   "source_url": "https://omlx.ai/benchmarks/performance/x0v8etzs",
   "tg_tokens": null,
   "tps": 114.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-x0v8etzs-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 184.0 GB at this context; duplicate traces of the same session: 4ea2qh58, ptk439dv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2326.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,326 82.6 184.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x0v8etzs",
   "source_url": "https://omlx.ai/benchmarks/performance/x0v8etzs",
   "tg_tokens": null,
   "tps": 82.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-x0v8etzs-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 186.9 GB at this context; duplicate traces of the same session: 4ea2qh58, ptk439dv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2283.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,283 135.3 186.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x0v8etzs",
   "source_url": "https://omlx.ai/benchmarks/performance/x0v8etzs",
   "tg_tokens": null,
   "tps": 135.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-l77e8fnt-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: 2fav4lw9, z3qmh8yb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2299.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,299 96.1 178.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l77e8fnt",
   "source_url": "https://omlx.ai/benchmarks/performance/l77e8fnt",
   "tg_tokens": null,
   "tps": 96.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-l77e8fnt-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.7 GB at this context; duplicate traces of the same session: 2fav4lw9, z3qmh8yb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2269.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,269 83.8 178.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l77e8fnt",
   "source_url": "https://omlx.ai/benchmarks/performance/l77e8fnt",
   "tg_tokens": null,
   "tps": 83.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-l77e8fnt-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.0 GB at this context; duplicate traces of the same session: 2fav4lw9, z3qmh8yb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2223.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,223 112.3 180.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l77e8fnt",
   "source_url": "https://omlx.ai/benchmarks/performance/l77e8fnt",
   "tg_tokens": null,
   "tps": 112.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-nh0m6l4r-ctx1024",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 6.5 GB; TTFT 1846 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 554.7,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-mlx-mxfp4 on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 4bit; 1k tokens 554.7 PP tok/s 57.3 TG tok/s 1846 TTFT (ms) 6.5 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nh0m6l4r",
   "source_url": "https://omlx.ai/benchmarks/performance/nh0m6l4r",
   "tg_tokens": null,
   "tps": 57.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-ffzfmnig-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: vv7id2u3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1317.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 1,317 125.6 21.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ffzfmnig",
   "source_url": "https://omlx.ai/benchmarks/performance/ffzfmnig",
   "tg_tokens": null,
   "tps": 125.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-ffzfmnig-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: vv7id2u3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1485.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 1,485 96.2 23.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ffzfmnig",
   "source_url": "https://omlx.ai/benchmarks/performance/ffzfmnig",
   "tg_tokens": null,
   "tps": 96.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-m6gt4pjz-ctx1024",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4099.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,099 198.5 21.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m6gt4pjz",
   "source_url": "https://omlx.ai/benchmarks/performance/m6gt4pjz",
   "tg_tokens": null,
   "tps": 198.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-m6gt4pjz-ctx4096",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5603.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 5,603 161.0 22.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m6gt4pjz",
   "source_url": "https://omlx.ai/benchmarks/performance/m6gt4pjz",
   "tg_tokens": null,
   "tps": 161.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-66f20jyy-ctx1024",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4113.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,113 194.8 21.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/66f20jyy",
   "source_url": "https://omlx.ai/benchmarks/performance/66f20jyy",
   "tg_tokens": null,
   "tps": 194.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-66f20jyy-ctx4096",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5506.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 5,506 164.4 22.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/66f20jyy",
   "source_url": "https://omlx.ai/benchmarks/performance/66f20jyy",
   "tg_tokens": null,
   "tps": 164.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-92lkajrs-ctx16384",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 22.4 GB; TTFT 7767 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2109.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k tokens 2,109 PP tok/s 107.2 TG tok/s 7767 TTFT (ms) 22.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/92lkajrs",
   "source_url": "https://omlx.ai/benchmarks/performance/92lkajrs",
   "tg_tokens": null,
   "tps": 107.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-8357t7gm-ctx4096",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 22.1 GB; TTFT 1664 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2461.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k tokens 2,461 PP tok/s 83.1 TG tok/s 1664 TTFT (ms) 22.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8357t7gm",
   "source_url": "https://omlx.ai/benchmarks/performance/8357t7gm",
   "tg_tokens": null,
   "tps": 83.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-yovuy0x5-ctx4096",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 22.1 GB; TTFT 1682 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2435.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k tokens 2,435 PP tok/s 91.6 TG tok/s 1682 TTFT (ms) 22.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yovuy0x5",
   "source_url": "https://omlx.ai/benchmarks/performance/yovuy0x5",
   "tg_tokens": null,
   "tps": 91.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-za3hjebm-ctx4096",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 20.8 GB; TTFT 2192 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1869.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k tokens 1,869 PP tok/s 75.7 TG tok/s 2192 TTFT (ms) 20.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/za3hjebm",
   "source_url": "https://omlx.ai/benchmarks/performance/za3hjebm",
   "tg_tokens": null,
   "tps": 75.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-tij6fexm-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: edk314dr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 215.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 215.1 46.2 18.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tij6fexm",
   "source_url": "https://omlx.ai/benchmarks/performance/tij6fexm",
   "tg_tokens": null,
   "tps": 46.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-tij6fexm-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: edk314dr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 199.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 199.8 44.0 22.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tij6fexm",
   "source_url": "https://omlx.ai/benchmarks/performance/tij6fexm",
   "tg_tokens": null,
   "tps": 44.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-r9vmruzi-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: en20zklt, heregxxi, ogokptif, s7mhhhi9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 895.8,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 895.8 54.6 24.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r9vmruzi",
   "source_url": "https://omlx.ai/benchmarks/performance/r9vmruzi",
   "tg_tokens": null,
   "tps": 54.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-r9vmruzi-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: en20zklt, heregxxi, ogokptif, s7mhhhi9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1137.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 1,137 58.4 25.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r9vmruzi",
   "source_url": "https://omlx.ai/benchmarks/performance/r9vmruzi",
   "tg_tokens": null,
   "tps": 58.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-r9vmruzi-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: en20zklt, heregxxi, ogokptif, s7mhhhi9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1084.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 1,084 50.3 26.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r9vmruzi",
   "source_url": "https://omlx.ai/benchmarks/performance/r9vmruzi",
   "tg_tokens": null,
   "tps": 50.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-r9vmruzi-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: en20zklt, heregxxi, ogokptif, s7mhhhi9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 882.8,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 882.8 75.5 27.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r9vmruzi",
   "source_url": "https://omlx.ai/benchmarks/performance/r9vmruzi",
   "tg_tokens": null,
   "tps": 75.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-r9vmruzi-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.8 GB at this context; duplicate traces of the same session: en20zklt, heregxxi, ogokptif, s7mhhhi9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 926.3,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 926.3 123.6 29.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r9vmruzi",
   "source_url": "https://omlx.ai/benchmarks/performance/r9vmruzi",
   "tg_tokens": null,
   "tps": 123.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-tf45yt2i-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.0 GB at this context; duplicate traces of the same session: cdqtmnbo; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3148.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 3,148 129.9 99.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tf45yt2i",
   "source_url": "https://omlx.ai/benchmarks/performance/tf45yt2i",
   "tg_tokens": null,
   "tps": 129.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-tf45yt2i-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.4 GB at this context; duplicate traces of the same session: cdqtmnbo; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3325.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 3,325 117.6 100.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tf45yt2i",
   "source_url": "https://omlx.ai/benchmarks/performance/tf45yt2i",
   "tg_tokens": null,
   "tps": 117.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-i2fga9wo-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 74.2 GB; TTFT 161285 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 406.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k tokens 406.3 PP tok/s 10.4 TG tok/s 161285 TTFT (ms) 74.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i2fga9wo",
   "source_url": "https://omlx.ai/benchmarks/performance/i2fga9wo",
   "tg_tokens": null,
   "tps": 10.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-bt4ninv0-ctx1024",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: ctc0emd0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 208.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31B-it-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 208.5 31.9 20.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bt4ninv0",
   "source_url": "https://omlx.ai/benchmarks/performance/bt4ninv0",
   "tg_tokens": null,
   "tps": 31.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-bt4ninv0-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: ctc0emd0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 186.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31B-it-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 186.2 12.5 20.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bt4ninv0",
   "source_url": "https://omlx.ai/benchmarks/performance/bt4ninv0",
   "tg_tokens": null,
   "tps": 12.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-egdl0ihg-ctx1024",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: aujch9og; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 167.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 167.8 33.1 34.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/egdl0ihg",
   "source_url": "https://omlx.ai/benchmarks/performance/egdl0ihg",
   "tg_tokens": null,
   "tps": 33.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-egdl0ihg-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 35.0 GB at this context; duplicate traces of the same session: aujch9og; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 172.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 172.7 28.4 35.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/egdl0ihg",
   "source_url": "https://omlx.ai/benchmarks/performance/egdl0ihg",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-l3p21pnz-ctx1024",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: oosnezbz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 140.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 140.8 12.5 32.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l3p21pnz",
   "source_url": "https://omlx.ai/benchmarks/performance/l3p21pnz",
   "tg_tokens": null,
   "tps": 12.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-l3p21pnz-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 33.1 GB at this context; duplicate traces of the same session: oosnezbz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 160.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 160.6 13.0 33.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l3p21pnz",
   "source_url": "https://omlx.ai/benchmarks/performance/l3p21pnz",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-cbcvhm7e-ctx4096",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: 3oug1lqq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 213.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "AREX-2-6bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 6bit; 4k 213.9 16.3 25.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cbcvhm7e",
   "source_url": "https://omlx.ai/benchmarks/performance/cbcvhm7e",
   "tg_tokens": null,
   "tps": 16.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-cbcvhm7e-ctx16384",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 28.4 GB at this context; duplicate traces of the same session: 3oug1lqq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 186.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "AREX-2-6bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 6bit; 16k 186.8 14.2 28.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cbcvhm7e",
   "source_url": "https://omlx.ai/benchmarks/performance/cbcvhm7e",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-0wjy28k7-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: 9pwz5v73, f09t7hky; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1260.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,260 62.0 177.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0wjy28k7",
   "source_url": "https://omlx.ai/benchmarks/performance/0wjy28k7",
   "tg_tokens": null,
   "tps": 62.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-0wjy28k7-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: 9pwz5v73, f09t7hky; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1982.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,982 57.1 178.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0wjy28k7",
   "source_url": "https://omlx.ai/benchmarks/performance/0wjy28k7",
   "tg_tokens": null,
   "tps": 57.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-0wjy28k7-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.3 GB at this context; duplicate traces of the same session: 9pwz5v73, f09t7hky; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1969.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 1,969 57.5 179.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0wjy28k7",
   "source_url": "https://omlx.ai/benchmarks/performance/0wjy28k7",
   "tg_tokens": null,
   "tps": 57.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-0wjy28k7-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.3 GB at this context; duplicate traces of the same session: 9pwz5v73, f09t7hky; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1964.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,964 57.1 179.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0wjy28k7",
   "source_url": "https://omlx.ai/benchmarks/performance/0wjy28k7",
   "tg_tokens": null,
   "tps": 57.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-0wjy28k7-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.6 GB at this context; duplicate traces of the same session: 9pwz5v73, f09t7hky; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1952.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,952 56.9 179.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0wjy28k7",
   "source_url": "https://omlx.ai/benchmarks/performance/0wjy28k7",
   "tg_tokens": null,
   "tps": 56.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-0wjy28k7-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.9 GB at this context; duplicate traces of the same session: 9pwz5v73, f09t7hky; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1924.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,924 55.8 179.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0wjy28k7",
   "source_url": "https://omlx.ai/benchmarks/performance/0wjy28k7",
   "tg_tokens": null,
   "tps": 55.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-0wjy28k7-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.6 GB at this context; duplicate traces of the same session: 9pwz5v73, f09t7hky; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1887.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 1,887 53.7 180.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0wjy28k7",
   "source_url": "https://omlx.ai/benchmarks/performance/0wjy28k7",
   "tg_tokens": null,
   "tps": 53.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-0wjy28k7-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.4 GB at this context; duplicate traces of the same session: 9pwz5v73, f09t7hky; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1843.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 1,843 50.5 181.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0wjy28k7",
   "source_url": "https://omlx.ai/benchmarks/performance/0wjy28k7",
   "tg_tokens": null,
   "tps": 50.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-4kgaqd0a-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.5 GB at this context; duplicate traces of the same session: 96f5pzqv, ftessehm, j47olpe9, m491k777; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1593.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,593 74.0 177.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4kgaqd0a",
   "source_url": "https://omlx.ai/benchmarks/performance/4kgaqd0a",
   "tg_tokens": null,
   "tps": 74.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-4kgaqd0a-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.3 GB at this context; duplicate traces of the same session: 96f5pzqv, ftessehm, j47olpe9, m491k777; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2328.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 2,328 63.1 178.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4kgaqd0a",
   "source_url": "https://omlx.ai/benchmarks/performance/4kgaqd0a",
   "tg_tokens": null,
   "tps": 63.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-4kgaqd0a-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.1 GB at this context; duplicate traces of the same session: 96f5pzqv, ftessehm, j47olpe9, m491k777; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2341.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 2,341 54.5 179.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4kgaqd0a",
   "source_url": "https://omlx.ai/benchmarks/performance/4kgaqd0a",
   "tg_tokens": null,
   "tps": 54.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-4kgaqd0a-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.5 GB at this context; duplicate traces of the same session: 96f5pzqv, ftessehm, j47olpe9, m491k777; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2336.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 2,336 142.1 179.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4kgaqd0a",
   "source_url": "https://omlx.ai/benchmarks/performance/4kgaqd0a",
   "tg_tokens": null,
   "tps": 142.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-4kgaqd0a-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.4 GB at this context; duplicate traces of the same session: 96f5pzqv, ftessehm, j47olpe9, m491k777; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2325.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,325 106.1 180.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4kgaqd0a",
   "source_url": "https://omlx.ai/benchmarks/performance/4kgaqd0a",
   "tg_tokens": null,
   "tps": 106.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-4kgaqd0a-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 182.2 GB at this context; duplicate traces of the same session: 96f5pzqv, ftessehm, j47olpe9, m491k777; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2293.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,293 72.1 182.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4kgaqd0a",
   "source_url": "https://omlx.ai/benchmarks/performance/4kgaqd0a",
   "tg_tokens": null,
   "tps": 72.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-4kgaqd0a-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 185.7 GB at this context; duplicate traces of the same session: 96f5pzqv, ftessehm, j47olpe9, m491k777; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2248.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,248 145.9 185.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4kgaqd0a",
   "source_url": "https://omlx.ai/benchmarks/performance/4kgaqd0a",
   "tg_tokens": null,
   "tps": 145.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-4kgaqd0a-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 189.2 GB at this context; duplicate traces of the same session: 96f5pzqv, ftessehm, j47olpe9, m491k777; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2199.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 2,199 68.6 189.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4kgaqd0a",
   "source_url": "https://omlx.ai/benchmarks/performance/4kgaqd0a",
   "tg_tokens": null,
   "tps": 68.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8q1nbr3d-ctx1024",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: bs1srp43, j15oj6k0, qgh1q076; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2877.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,877 143.6 20.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8q1nbr3d",
   "source_url": "https://omlx.ai/benchmarks/performance/8q1nbr3d",
   "tg_tokens": null,
   "tps": 143.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8q1nbr3d-ctx4096",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: bs1srp43, j15oj6k0, qgh1q076; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4192.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 4,192 130.1 21.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8q1nbr3d",
   "source_url": "https://omlx.ai/benchmarks/performance/8q1nbr3d",
   "tg_tokens": null,
   "tps": 130.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8q1nbr3d-ctx8192",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: bs1srp43, j15oj6k0, qgh1q076; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4221.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 4,221 142.3 21.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8q1nbr3d",
   "source_url": "https://omlx.ai/benchmarks/performance/8q1nbr3d",
   "tg_tokens": null,
   "tps": 142.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8q1nbr3d-ctx16384",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: bs1srp43, j15oj6k0, qgh1q076; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3947.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 3,947 130.2 21.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8q1nbr3d",
   "source_url": "https://omlx.ai/benchmarks/performance/8q1nbr3d",
   "tg_tokens": null,
   "tps": 130.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8q1nbr3d-ctx32768",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: bs1srp43, j15oj6k0, qgh1q076; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3687.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 3,687 113.1 21.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8q1nbr3d",
   "source_url": "https://omlx.ai/benchmarks/performance/8q1nbr3d",
   "tg_tokens": null,
   "tps": 113.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8q1nbr3d-ctx65536",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: bs1srp43, j15oj6k0, qgh1q076; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2830.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,830 91.2 21.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8q1nbr3d",
   "source_url": "https://omlx.ai/benchmarks/performance/8q1nbr3d",
   "tg_tokens": null,
   "tps": 91.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-1t29h9bx-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 21.5 GB; TTFT 6927 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 591.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k tokens 591.3 PP tok/s 68.5 TG tok/s 6927 TTFT (ms) 21.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1t29h9bx",
   "source_url": "https://omlx.ai/benchmarks/performance/1t29h9bx",
   "tg_tokens": null,
   "tps": 68.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-0ozp7mbf-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 21.5 GB; TTFT 6922 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 591.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k tokens 591.7 PP tok/s 68.4 TG tok/s 6922 TTFT (ms) 21.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ozp7mbf",
   "source_url": "https://omlx.ai/benchmarks/performance/0ozp7mbf",
   "tg_tokens": null,
   "tps": 68.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hi7yjk5j-ctx1024",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4113.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,113 194.8 21.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hi7yjk5j",
   "source_url": "https://omlx.ai/benchmarks/performance/hi7yjk5j",
   "tg_tokens": null,
   "tps": 194.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hi7yjk5j-ctx4096",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5506.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 5,506 164.4 22.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hi7yjk5j",
   "source_url": "https://omlx.ai/benchmarks/performance/hi7yjk5j",
   "tg_tokens": null,
   "tps": 164.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-k8oyrhbc-ctx1024",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4099.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,099 198.5 21.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k8oyrhbc",
   "source_url": "https://omlx.ai/benchmarks/performance/k8oyrhbc",
   "tg_tokens": null,
   "tps": 198.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-k8oyrhbc-ctx4096",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5603.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 5,603 161.0 22.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k8oyrhbc",
   "source_url": "https://omlx.ai/benchmarks/performance/k8oyrhbc",
   "tg_tokens": null,
   "tps": 161.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-g5pqjjnl-ctx1024",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: jiwxjxqp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4117.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,117 142.5 21.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g5pqjjnl",
   "source_url": "https://omlx.ai/benchmarks/performance/g5pqjjnl",
   "tg_tokens": null,
   "tps": 142.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-g5pqjjnl-ctx4096",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: jiwxjxqp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5663.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 5,663 175.3 22.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g5pqjjnl",
   "source_url": "https://omlx.ai/benchmarks/performance/g5pqjjnl",
   "tg_tokens": null,
   "tps": 175.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-396annn5-ctx1024",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: ul39bbyc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4133.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,133 179.0 21.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/396annn5",
   "source_url": "https://omlx.ai/benchmarks/performance/396annn5",
   "tg_tokens": null,
   "tps": 179.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-396annn5-ctx4096",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: ul39bbyc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5474.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 5,474 154.9 22.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/396annn5",
   "source_url": "https://omlx.ai/benchmarks/performance/396annn5",
   "tg_tokens": null,
   "tps": 154.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hxa4wk9g-ctx1024",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: d9jcmtvw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4194.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,194 184.0 21.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hxa4wk9g",
   "source_url": "https://omlx.ai/benchmarks/performance/hxa4wk9g",
   "tg_tokens": null,
   "tps": 184.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hxa4wk9g-ctx4096",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: d9jcmtvw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5630.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 5,630 172.7 22.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hxa4wk9g",
   "source_url": "https://omlx.ai/benchmarks/performance/hxa4wk9g",
   "tg_tokens": null,
   "tps": 172.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lprbbaxj-ctx1024",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 0luvm2p0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4139.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,139 168.1 21.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lprbbaxj",
   "source_url": "https://omlx.ai/benchmarks/performance/lprbbaxj",
   "tg_tokens": null,
   "tps": 168.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lprbbaxj-ctx4096",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 0luvm2p0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5541.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 5,541 180.9 22.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lprbbaxj",
   "source_url": "https://omlx.ai/benchmarks/performance/lprbbaxj",
   "tg_tokens": null,
   "tps": 180.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-skvyj60c-ctx1024",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 0adhkm3o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4190.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,190 186.7 21.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/skvyj60c",
   "source_url": "https://omlx.ai/benchmarks/performance/skvyj60c",
   "tg_tokens": null,
   "tps": 186.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-skvyj60c-ctx4096",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 0adhkm3o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5754.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 5,754 173.2 22.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/skvyj60c",
   "source_url": "https://omlx.ai/benchmarks/performance/skvyj60c",
   "tg_tokens": null,
   "tps": 173.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-r104ahpk-ctx1024",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: b8rth7ag; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4175.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,175 172.8 21.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r104ahpk",
   "source_url": "https://omlx.ai/benchmarks/performance/r104ahpk",
   "tg_tokens": null,
   "tps": 172.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-r104ahpk-ctx4096",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: b8rth7ag; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5139.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 5,139 182.4 22.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r104ahpk",
   "source_url": "https://omlx.ai/benchmarks/performance/r104ahpk",
   "tg_tokens": null,
   "tps": 182.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4npg9x0a-ctx1024",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: xj87b9go; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4168.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,168 186.0 21.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4npg9x0a",
   "source_url": "https://omlx.ai/benchmarks/performance/4npg9x0a",
   "tg_tokens": null,
   "tps": 186.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4npg9x0a-ctx4096",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: xj87b9go; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5332.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 5,332 178.1 22.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4npg9x0a",
   "source_url": "https://omlx.ai/benchmarks/performance/4npg9x0a",
   "tg_tokens": null,
   "tps": 178.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-srmmiih4-ctx1024",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: f8ltazi3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4390.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,390 184.9 21.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/srmmiih4",
   "source_url": "https://omlx.ai/benchmarks/performance/srmmiih4",
   "tg_tokens": null,
   "tps": 184.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-srmmiih4-ctx4096",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: f8ltazi3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5892.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 5,892 177.2 22.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/srmmiih4",
   "source_url": "https://omlx.ai/benchmarks/performance/srmmiih4",
   "tg_tokens": null,
   "tps": 177.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-sa0kw9dr-ctx16384",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 21.0 GB; TTFT 6924 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2366.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k tokens 2,366 PP tok/s 70.1 TG tok/s 6924 TTFT (ms) 21.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sa0kw9dr",
   "source_url": "https://omlx.ai/benchmarks/performance/sa0kw9dr",
   "tg_tokens": null,
   "tps": 70.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5j2dtlmp-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: z3q2jeig; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1249.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,249 49.3 17.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5j2dtlmp",
   "source_url": "https://omlx.ai/benchmarks/performance/5j2dtlmp",
   "tg_tokens": null,
   "tps": 49.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5j2dtlmp-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: z3q2jeig; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1474.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,474 42.9 18.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5j2dtlmp",
   "source_url": "https://omlx.ai/benchmarks/performance/5j2dtlmp",
   "tg_tokens": null,
   "tps": 42.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jgnknm91-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: w31k1lpa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 956.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 956.1 45.3 21.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jgnknm91",
   "source_url": "https://omlx.ai/benchmarks/performance/jgnknm91",
   "tg_tokens": null,
   "tps": 45.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jgnknm91-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: w31k1lpa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 889.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 889.1 27.0 24.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jgnknm91",
   "source_url": "https://omlx.ai/benchmarks/performance/jgnknm91",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-su9gq20t-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: avgfs0ak; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1321.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k 1,321 55.4 16.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/su9gq20t",
   "source_url": "https://omlx.ai/benchmarks/performance/su9gq20t",
   "tg_tokens": null,
   "tps": 55.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-su9gq20t-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: avgfs0ak; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1484.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 4k 1,484 53.8 17.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/su9gq20t",
   "source_url": "https://omlx.ai/benchmarks/performance/su9gq20t",
   "tg_tokens": null,
   "tps": 53.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1k24p6b7-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: 0gildp5k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 868.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k 868.8 40.1 27.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1k24p6b7",
   "source_url": "https://omlx.ai/benchmarks/performance/1k24p6b7",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1k24p6b7-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.1 GB at this context; duplicate traces of the same session: 0gildp5k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 824.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 824.4 25.7 29.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1k24p6b7",
   "source_url": "https://omlx.ai/benchmarks/performance/1k24p6b7",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-pxv3gwlz-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: md0gk58j; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1169.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 1k 1,169 33.1 28.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pxv3gwlz",
   "source_url": "https://omlx.ai/benchmarks/performance/pxv3gwlz",
   "tg_tokens": null,
   "tps": 33.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-pxv3gwlz-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: md0gk58j; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1420.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 4k 1,420 32.9 29.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pxv3gwlz",
   "source_url": "https://omlx.ai/benchmarks/performance/pxv3gwlz",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0zmwtej2-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 43.2 GB at this context; duplicate traces of the same session: 4siyugx0, lacg0yl2, mk5wo84c, qt2o138d, qzfsrvfp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 776.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 776.9 47.0 43.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0zmwtej2",
   "source_url": "https://omlx.ai/benchmarks/performance/0zmwtej2",
   "tg_tokens": null,
   "tps": 47.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0zmwtej2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 44.4 GB at this context; duplicate traces of the same session: 4siyugx0, lacg0yl2, mk5wo84c, qt2o138d, qzfsrvfp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 706.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 706.4 46.5 44.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0zmwtej2",
   "source_url": "https://omlx.ai/benchmarks/performance/0zmwtej2",
   "tg_tokens": null,
   "tps": 46.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0zmwtej2-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 44.9 GB at this context; duplicate traces of the same session: 4siyugx0, lacg0yl2, mk5wo84c, qt2o138d, qzfsrvfp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 722.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 722.9 44.1 44.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0zmwtej2",
   "source_url": "https://omlx.ai/benchmarks/performance/0zmwtej2",
   "tg_tokens": null,
   "tps": 44.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0zmwtej2-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 46.1 GB at this context; duplicate traces of the same session: 4siyugx0, lacg0yl2, mk5wo84c, qt2o138d, qzfsrvfp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 693.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 693.8 53.4 46.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0zmwtej2",
   "source_url": "https://omlx.ai/benchmarks/performance/0zmwtej2",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0zmwtej2-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 48.4 GB at this context; duplicate traces of the same session: 4siyugx0, lacg0yl2, mk5wo84c, qt2o138d, qzfsrvfp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 632.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 632.8 63.6 48.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0zmwtej2",
   "source_url": "https://omlx.ai/benchmarks/performance/0zmwtej2",
   "tg_tokens": null,
   "tps": 63.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0zmwtej2-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 52.9 GB at this context; duplicate traces of the same session: 4siyugx0, lacg0yl2, mk5wo84c, qt2o138d, qzfsrvfp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 570.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 570.4 33.8 52.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0zmwtej2",
   "source_url": "https://omlx.ai/benchmarks/performance/0zmwtej2",
   "tg_tokens": null,
   "tps": 33.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-y4bkh13j-ctx1024",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: u84id3qe; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 233.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 233.8 21.5 5.8 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y4bkh13j",
   "source_url": "https://omlx.ai/benchmarks/performance/y4bkh13j",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-y4bkh13j-ctx4096",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: u84id3qe; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 240.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 240.8 21.0 6.2 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y4bkh13j",
   "source_url": "https://omlx.ai/benchmarks/performance/y4bkh13j",
   "tg_tokens": null,
   "tps": 21.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-fs8cz85t-ctx1024",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 27.3 GB at this context; duplicate traces of the same session: f41ptdt4; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 191.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31B-it-MLX-4bit on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 1k 191.2 19.7 27.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fs8cz85t",
   "source_url": "https://omlx.ai/benchmarks/performance/fs8cz85t",
   "tg_tokens": null,
   "tps": 19.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-fs8cz85t-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 28.4 GB at this context; duplicate traces of the same session: f41ptdt4; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 185.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31B-it-MLX-4bit on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 4k 185.7 19.0 28.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fs8cz85t",
   "source_url": "https://omlx.ai/benchmarks/performance/fs8cz85t",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-1dqipig7-ctx1024",
   "model": "gemma-4-E4B-it-ultra-uncensored-heretic",
   "notes": "oMLX community benchmark trace; peak memory 5.4 GB at this context; duplicate traces of the same session: jkqunu1f; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 650.3,
   "provenance": "community",
   "quant": "mixed 4+6bit",
   "quote": "gemma-4-E4B-it-ultra-uncensored-heretic-MLX-4bit-mixed_4_6 on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 650.3 31.6 5.4 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1dqipig7",
   "source_url": "https://omlx.ai/benchmarks/performance/1dqipig7",
   "tg_tokens": null,
   "tps": 31.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-1dqipig7-ctx4096",
   "model": "gemma-4-E4B-it-ultra-uncensored-heretic",
   "notes": "oMLX community benchmark trace; peak memory 5.5 GB at this context; duplicate traces of the same session: jkqunu1f; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 662.1,
   "provenance": "community",
   "quant": "mixed 4+6bit",
   "quote": "gemma-4-E4B-it-ultra-uncensored-heretic-MLX-4bit-mixed_4_6 on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 662.1 31.0 5.5 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1dqipig7",
   "source_url": "https://omlx.ai/benchmarks/performance/1dqipig7",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ean5kuky-ctx1024",
   "model": "Occamy-1.0",
   "notes": "oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: damreyhz, xedsqw5z; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 657.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "occamy-1.0-MLX-6bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 1k 657.1 60.9 27.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ean5kuky",
   "source_url": "https://omlx.ai/benchmarks/performance/ean5kuky",
   "tg_tokens": null,
   "tps": 60.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ean5kuky-ctx4096",
   "model": "Occamy-1.0",
   "notes": "oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: damreyhz, xedsqw5z; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 715.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "occamy-1.0-MLX-6bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 4k 715.4 59.0 29.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ean5kuky",
   "source_url": "https://omlx.ai/benchmarks/performance/ean5kuky",
   "tg_tokens": null,
   "tps": 59.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ean5kuky-ctx32768",
   "model": "Occamy-1.0",
   "notes": "oMLX community benchmark trace; peak memory 31.8 GB at this context; duplicate traces of the same session: damreyhz, xedsqw5z; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 461.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "occamy-1.0-MLX-6bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 32k 461.4 47.8 31.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ean5kuky",
   "source_url": "https://omlx.ai/benchmarks/performance/ean5kuky",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ey3027ql-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.2 GB at this context; duplicate traces of the same session: 1lwp0ce0, 92dvhrr0, gv3e74ru, ooymokxl, umumwzws, w3nwpkpb, zfyfmrje; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1269.0,
   "provenance": "community",
   "quant": "mixed 4/8bit",
   "quote": "GLM-5.3-Flash-MLX-mixed-4_8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,269 59.9 174.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ey3027ql",
   "source_url": "https://omlx.ai/benchmarks/performance/ey3027ql",
   "tg_tokens": null,
   "tps": 59.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ey3027ql-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.0 GB at this context; duplicate traces of the same session: 1lwp0ce0, 92dvhrr0, gv3e74ru, ooymokxl, umumwzws, w3nwpkpb, zfyfmrje; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1971.0,
   "provenance": "community",
   "quant": "mixed 4/8bit",
   "quote": "GLM-5.3-Flash-MLX-mixed-4_8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,971 55.3 175.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ey3027ql",
   "source_url": "https://omlx.ai/benchmarks/performance/ey3027ql",
   "tg_tokens": null,
   "tps": 55.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ey3027ql-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 176.0 GB at this context; duplicate traces of the same session: 1lwp0ce0, 92dvhrr0, gv3e74ru, ooymokxl, umumwzws, w3nwpkpb, zfyfmrje; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1971.0,
   "provenance": "community",
   "quant": "mixed 4/8bit",
   "quote": "GLM-5.3-Flash-MLX-mixed-4_8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 1,971 55.4 176.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ey3027ql",
   "source_url": "https://omlx.ai/benchmarks/performance/ey3027ql",
   "tg_tokens": null,
   "tps": 55.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ey3027ql-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 176.1 GB at this context; duplicate traces of the same session: 1lwp0ce0, 92dvhrr0, gv3e74ru, ooymokxl, umumwzws, w3nwpkpb, zfyfmrje; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1969.0,
   "provenance": "community",
   "quant": "mixed 4/8bit",
   "quote": "GLM-5.3-Flash-MLX-mixed-4_8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,969 55.5 176.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ey3027ql",
   "source_url": "https://omlx.ai/benchmarks/performance/ey3027ql",
   "tg_tokens": null,
   "tps": 55.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ey3027ql-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 176.3 GB at this context; duplicate traces of the same session: 1lwp0ce0, 92dvhrr0, gv3e74ru, ooymokxl, umumwzws, w3nwpkpb, zfyfmrje; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1958.0,
   "provenance": "community",
   "quant": "mixed 4/8bit",
   "quote": "GLM-5.3-Flash-MLX-mixed-4_8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,958 55.2 176.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ey3027ql",
   "source_url": "https://omlx.ai/benchmarks/performance/ey3027ql",
   "tg_tokens": null,
   "tps": 55.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ey3027ql-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 176.7 GB at this context; duplicate traces of the same session: 1lwp0ce0, 92dvhrr0, gv3e74ru, ooymokxl, umumwzws, w3nwpkpb, zfyfmrje; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1921.0,
   "provenance": "community",
   "quant": "mixed 4/8bit",
   "quote": "GLM-5.3-Flash-MLX-mixed-4_8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,921 53.2 176.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ey3027ql",
   "source_url": "https://omlx.ai/benchmarks/performance/ey3027ql",
   "tg_tokens": null,
   "tps": 53.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ey3027ql-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.3 GB at this context; duplicate traces of the same session: 1lwp0ce0, 92dvhrr0, gv3e74ru, ooymokxl, umumwzws, w3nwpkpb, zfyfmrje; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1883.0,
   "provenance": "community",
   "quant": "mixed 4/8bit",
   "quote": "GLM-5.3-Flash-MLX-mixed-4_8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 1,883 51.5 177.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ey3027ql",
   "source_url": "https://omlx.ai/benchmarks/performance/ey3027ql",
   "tg_tokens": null,
   "tps": 51.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ey3027ql-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.1 GB at this context; duplicate traces of the same session: 1lwp0ce0, 92dvhrr0, gv3e74ru, ooymokxl, umumwzws, w3nwpkpb, zfyfmrje; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1841.0,
   "provenance": "community",
   "quant": "mixed 4/8bit",
   "quote": "GLM-5.3-Flash-MLX-mixed-4_8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 1,841 48.9 178.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ey3027ql",
   "source_url": "https://omlx.ai/benchmarks/performance/ey3027ql",
   "tg_tokens": null,
   "tps": 48.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-szyqfr82-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.5 GB at this context; duplicate traces of the same session: n1e3jeaf, rqqdihq0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1593.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,593 74.0 177.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/szyqfr82",
   "source_url": "https://omlx.ai/benchmarks/performance/szyqfr82",
   "tg_tokens": null,
   "tps": 74.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-szyqfr82-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.3 GB at this context; duplicate traces of the same session: n1e3jeaf, rqqdihq0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2328.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 2,328 63.1 178.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/szyqfr82",
   "source_url": "https://omlx.ai/benchmarks/performance/szyqfr82",
   "tg_tokens": null,
   "tps": 63.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-szyqfr82-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.1 GB at this context; duplicate traces of the same session: n1e3jeaf, rqqdihq0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2341.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 2,341 54.5 179.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/szyqfr82",
   "source_url": "https://omlx.ai/benchmarks/performance/szyqfr82",
   "tg_tokens": null,
   "tps": 54.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-szyqfr82-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.5 GB at this context; duplicate traces of the same session: n1e3jeaf, rqqdihq0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2336.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 2,336 142.1 179.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/szyqfr82",
   "source_url": "https://omlx.ai/benchmarks/performance/szyqfr82",
   "tg_tokens": null,
   "tps": 142.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-szyqfr82-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.4 GB at this context; duplicate traces of the same session: n1e3jeaf, rqqdihq0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2325.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,325 106.1 180.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/szyqfr82",
   "source_url": "https://omlx.ai/benchmarks/performance/szyqfr82",
   "tg_tokens": null,
   "tps": 106.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-szyqfr82-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 182.2 GB at this context; duplicate traces of the same session: n1e3jeaf, rqqdihq0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2293.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,293 72.1 182.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/szyqfr82",
   "source_url": "https://omlx.ai/benchmarks/performance/szyqfr82",
   "tg_tokens": null,
   "tps": 72.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-szyqfr82-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 185.7 GB at this context; duplicate traces of the same session: n1e3jeaf, rqqdihq0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2248.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,248 145.9 185.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/szyqfr82",
   "source_url": "https://omlx.ai/benchmarks/performance/szyqfr82",
   "tg_tokens": null,
   "tps": 145.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-szyqfr82-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 189.2 GB at this context; duplicate traces of the same session: n1e3jeaf, rqqdihq0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2199.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 2,199 68.6 189.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/szyqfr82",
   "source_url": "https://omlx.ai/benchmarks/performance/szyqfr82",
   "tg_tokens": null,
   "tps": 68.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-aw00h4rl-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.7 GB at this context; duplicate traces of the same session: cj8tkaqo, hxu41f8q, ixvkpw02, ocucgk48, p4m7qa5k, uqaqtfgz, xccaracj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1556.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,556 73.3 179.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aw00h4rl",
   "source_url": "https://omlx.ai/benchmarks/performance/aw00h4rl",
   "tg_tokens": null,
   "tps": 73.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-aw00h4rl-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.6 GB at this context; duplicate traces of the same session: cj8tkaqo, hxu41f8q, ixvkpw02, ocucgk48, p4m7qa5k, uqaqtfgz, xccaracj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2320.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 2,320 64.7 180.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aw00h4rl",
   "source_url": "https://omlx.ai/benchmarks/performance/aw00h4rl",
   "tg_tokens": null,
   "tps": 64.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-aw00h4rl-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.4 GB at this context; duplicate traces of the same session: cj8tkaqo, hxu41f8q, ixvkpw02, ocucgk48, p4m7qa5k, uqaqtfgz, xccaracj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2361.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 2,361 104.3 181.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aw00h4rl",
   "source_url": "https://omlx.ai/benchmarks/performance/aw00h4rl",
   "tg_tokens": null,
   "tps": 104.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-aw00h4rl-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.8 GB at this context; duplicate traces of the same session: cj8tkaqo, hxu41f8q, ixvkpw02, ocucgk48, p4m7qa5k, uqaqtfgz, xccaracj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2370.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 2,370 133.7 181.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aw00h4rl",
   "source_url": "https://omlx.ai/benchmarks/performance/aw00h4rl",
   "tg_tokens": null,
   "tps": 133.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-aw00h4rl-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 182.5 GB at this context; duplicate traces of the same session: cj8tkaqo, hxu41f8q, ixvkpw02, ocucgk48, p4m7qa5k, uqaqtfgz, xccaracj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2364.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,364 109.0 182.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aw00h4rl",
   "source_url": "https://omlx.ai/benchmarks/performance/aw00h4rl",
   "tg_tokens": null,
   "tps": 109.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-aw00h4rl-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 184.0 GB at this context; duplicate traces of the same session: cj8tkaqo, hxu41f8q, ixvkpw02, ocucgk48, p4m7qa5k, uqaqtfgz, xccaracj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2330.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,330 100.0 184.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aw00h4rl",
   "source_url": "https://omlx.ai/benchmarks/performance/aw00h4rl",
   "tg_tokens": null,
   "tps": 100.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-aw00h4rl-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 186.9 GB at this context; duplicate traces of the same session: cj8tkaqo, hxu41f8q, ixvkpw02, ocucgk48, p4m7qa5k, uqaqtfgz, xccaracj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2287.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,287 135.9 186.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aw00h4rl",
   "source_url": "https://omlx.ai/benchmarks/performance/aw00h4rl",
   "tg_tokens": null,
   "tps": 135.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-aw00h4rl-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 189.8 GB at this context; duplicate traces of the same session: cj8tkaqo, hxu41f8q, ixvkpw02, ocucgk48, p4m7qa5k, uqaqtfgz, xccaracj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2234.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 2,234 74.8 189.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aw00h4rl",
   "source_url": "https://omlx.ai/benchmarks/performance/aw00h4rl",
   "tg_tokens": null,
   "tps": 74.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ntkx62zm-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 249.1 GB at this context; duplicate traces of the same session: enxqewo7, jp42h7hc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 488.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 1k 488.8 40.7 249.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ntkx62zm",
   "source_url": "https://omlx.ai/benchmarks/performance/ntkx62zm",
   "tg_tokens": null,
   "tps": 40.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ntkx62zm-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 250.2 GB at this context; duplicate traces of the same session: enxqewo7, jp42h7hc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 570.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 4k 570.7 40.9 250.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ntkx62zm",
   "source_url": "https://omlx.ai/benchmarks/performance/ntkx62zm",
   "tg_tokens": null,
   "tps": 40.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ntkx62zm-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 250.6 GB at this context; duplicate traces of the same session: enxqewo7, jp42h7hc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 530.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 8k 530.6 45.1 250.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ntkx62zm",
   "source_url": "https://omlx.ai/benchmarks/performance/ntkx62zm",
   "tg_tokens": null,
   "tps": 45.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ntkx62zm-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 250.7 GB at this context; duplicate traces of the same session: enxqewo7, jp42h7hc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 520.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 16k 520.6 71.0 250.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ntkx62zm",
   "source_url": "https://omlx.ai/benchmarks/performance/ntkx62zm",
   "tg_tokens": null,
   "tps": 71.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bybk9q7m-ctx1024",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: 8c9x8g6c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2877.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,877 143.6 20.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bybk9q7m",
   "source_url": "https://omlx.ai/benchmarks/performance/bybk9q7m",
   "tg_tokens": null,
   "tps": 143.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bybk9q7m-ctx4096",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 8c9x8g6c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4192.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 4,192 130.1 21.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bybk9q7m",
   "source_url": "https://omlx.ai/benchmarks/performance/bybk9q7m",
   "tg_tokens": null,
   "tps": 130.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bybk9q7m-ctx8192",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 8c9x8g6c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4221.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 4,221 142.3 21.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bybk9q7m",
   "source_url": "https://omlx.ai/benchmarks/performance/bybk9q7m",
   "tg_tokens": null,
   "tps": 142.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bybk9q7m-ctx16384",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 8c9x8g6c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3947.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 3,947 130.2 21.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bybk9q7m",
   "source_url": "https://omlx.ai/benchmarks/performance/bybk9q7m",
   "tg_tokens": null,
   "tps": 130.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bybk9q7m-ctx32768",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 8c9x8g6c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3687.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 3,687 113.1 21.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bybk9q7m",
   "source_url": "https://omlx.ai/benchmarks/performance/bybk9q7m",
   "tg_tokens": null,
   "tps": 113.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bybk9q7m-ctx65536",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: 8c9x8g6c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2830.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,830 91.2 21.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bybk9q7m",
   "source_url": "https://omlx.ai/benchmarks/performance/bybk9q7m",
   "tg_tokens": null,
   "tps": 91.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-s6htogf9-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 30x9uo44, v7qp0574; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4021.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,021 127.6 20.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s6htogf9",
   "source_url": "https://omlx.ai/benchmarks/performance/s6htogf9",
   "tg_tokens": null,
   "tps": 127.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-s6htogf9-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 30x9uo44, v7qp0574; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4636.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 4,636 104.3 21.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s6htogf9",
   "source_url": "https://omlx.ai/benchmarks/performance/s6htogf9",
   "tg_tokens": null,
   "tps": 104.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-s6htogf9-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 30x9uo44, v7qp0574; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3238.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 3,238 85.4 21.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s6htogf9",
   "source_url": "https://omlx.ai/benchmarks/performance/s6htogf9",
   "tg_tokens": null,
   "tps": 85.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-5gmdiflq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: f9opnvdv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 454.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 454.9 17.6 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5gmdiflq",
   "source_url": "https://omlx.ai/benchmarks/performance/5gmdiflq",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-5gmdiflq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: f9opnvdv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 497.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 497.8 17.5 17.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5gmdiflq",
   "source_url": "https://omlx.ai/benchmarks/performance/5gmdiflq",
   "tg_tokens": null,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mgk0ijtq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 16.8 GB at this context; duplicate traces of the same session: pwflv0ry; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 778.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 778.8 55.3 16.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mgk0ijtq",
   "source_url": "https://omlx.ai/benchmarks/performance/mgk0ijtq",
   "tg_tokens": null,
   "tps": 55.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mgk0ijtq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: pwflv0ry; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 967.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 967.6 70.6 18.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mgk0ijtq",
   "source_url": "https://omlx.ai/benchmarks/performance/mgk0ijtq",
   "tg_tokens": null,
   "tps": 70.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-g6bpji8s-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 2xxi6p9i, 63ejdap8, a29b0y1m, d3sadp99, v2rkerjv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 900.3,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 900.3 70.5 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g6bpji8s",
   "source_url": "https://omlx.ai/benchmarks/performance/g6bpji8s",
   "tg_tokens": null,
   "tps": 70.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-g6bpji8s-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: 2xxi6p9i, 63ejdap8, a29b0y1m, d3sadp99, v2rkerjv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 932.8,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 932.8 60.0 20.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g6bpji8s",
   "source_url": "https://omlx.ai/benchmarks/performance/g6bpji8s",
   "tg_tokens": null,
   "tps": 60.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-g6bpji8s-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 2xxi6p9i, 63ejdap8, a29b0y1m, d3sadp99, v2rkerjv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 749.7,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 749.7 65.6 21.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g6bpji8s",
   "source_url": "https://omlx.ai/benchmarks/performance/g6bpji8s",
   "tg_tokens": null,
   "tps": 65.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-g6bpji8s-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: 2xxi6p9i, 63ejdap8, a29b0y1m, d3sadp99, v2rkerjv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 705.3,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 705.3 75.8 22.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g6bpji8s",
   "source_url": "https://omlx.ai/benchmarks/performance/g6bpji8s",
   "tg_tokens": null,
   "tps": 75.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-g6bpji8s-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: 2xxi6p9i, 63ejdap8, a29b0y1m, d3sadp99, v2rkerjv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 654.3,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 654.3 123.3 24.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g6bpji8s",
   "source_url": "https://omlx.ai/benchmarks/performance/g6bpji8s",
   "tg_tokens": null,
   "tps": 123.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-g6bpji8s-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.1 GB at this context; duplicate traces of the same session: 2xxi6p9i, 63ejdap8, a29b0y1m, d3sadp99, v2rkerjv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 572.1,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 572.1 49.6 29.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g6bpji8s",
   "source_url": "https://omlx.ai/benchmarks/performance/g6bpji8s",
   "tg_tokens": null,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ii2na09r-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: 693dc4yd; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 357.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 357.6 26.6 17.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ii2na09r",
   "source_url": "https://omlx.ai/benchmarks/performance/ii2na09r",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ii2na09r-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 693dc4yd; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 358.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 358.2 24.6 18.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ii2na09r",
   "source_url": "https://omlx.ai/benchmarks/performance/ii2na09r",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-mqj80hki-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: 0lqen2kr, 7iyqgov1, n7w2iqwn, pmpwksl2, q88sg73v; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1171.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 1k 1,171 79.1 33.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mqj80hki",
   "source_url": "https://omlx.ai/benchmarks/performance/mqj80hki",
   "tg_tokens": null,
   "tps": 79.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-mqj80hki-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: 0lqen2kr, 7iyqgov1, n7w2iqwn, pmpwksl2, q88sg73v; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1407.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 4k 1,407 65.6 34.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mqj80hki",
   "source_url": "https://omlx.ai/benchmarks/performance/mqj80hki",
   "tg_tokens": null,
   "tps": 65.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-mqj80hki-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: 0lqen2kr, 7iyqgov1, n7w2iqwn, pmpwksl2, q88sg73v; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1410.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 8k 1,410 80.5 35.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mqj80hki",
   "source_url": "https://omlx.ai/benchmarks/performance/mqj80hki",
   "tg_tokens": null,
   "tps": 80.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-mqj80hki-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: 0lqen2kr, 7iyqgov1, n7w2iqwn, pmpwksl2, q88sg73v; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1372.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 16k 1,372 89.6 35.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mqj80hki",
   "source_url": "https://omlx.ai/benchmarks/performance/mqj80hki",
   "tg_tokens": null,
   "tps": 89.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-mqj80hki-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: 0lqen2kr, 7iyqgov1, n7w2iqwn, pmpwksl2, q88sg73v; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1283.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 32k 1,283 92.9 36.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mqj80hki",
   "source_url": "https://omlx.ai/benchmarks/performance/mqj80hki",
   "tg_tokens": null,
   "tps": 92.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-mqj80hki-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: 0lqen2kr, 7iyqgov1, n7w2iqwn, pmpwksl2, q88sg73v; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1124.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 64k 1,124 72.2 39.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mqj80hki",
   "source_url": "https://omlx.ai/benchmarks/performance/mqj80hki",
   "tg_tokens": null,
   "tps": 72.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-60m2go32-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 77.9 GB at this context; duplicate traces of the same session: 82h1vhd0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 950.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 950.6 57.2 77.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/60m2go32",
   "source_url": "https://omlx.ai/benchmarks/performance/60m2go32",
   "tg_tokens": null,
   "tps": 57.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-60m2go32-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.0 GB at this context; duplicate traces of the same session: 82h1vhd0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1300.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,300 56.1 78.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/60m2go32",
   "source_url": "https://omlx.ai/benchmarks/performance/60m2go32",
   "tg_tokens": null,
   "tps": 56.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-uauat9pm-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MLX; oMLX community benchmark trace, single run at 4k context; peak memory 75.5 GB; TTFT 3335 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1228.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 1,228 PP tok/s 28.7 TG tok/s 3335 TTFT (ms) 75.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uauat9pm",
   "source_url": "https://omlx.ai/benchmarks/performance/uauat9pm",
   "tg_tokens": null,
   "tps": 28.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nqxlxtb8-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: 1yxkh739, qb72yznm; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 600.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 600.6 64.2 70.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nqxlxtb8",
   "source_url": "https://omlx.ai/benchmarks/performance/nqxlxtb8",
   "tg_tokens": null,
   "tps": 64.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nqxlxtb8-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: 1yxkh739, qb72yznm; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 769.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 769.9 63.1 71.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nqxlxtb8",
   "source_url": "https://omlx.ai/benchmarks/performance/nqxlxtb8",
   "tg_tokens": null,
   "tps": 63.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nqxlxtb8-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: 1yxkh739, qb72yznm; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1060.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,060 60.9 71.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nqxlxtb8",
   "source_url": "https://omlx.ai/benchmarks/performance/nqxlxtb8",
   "tg_tokens": null,
   "tps": 60.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nqxlxtb8-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 1yxkh739, qb72yznm; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1060.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 1,060 62.7 71.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nqxlxtb8",
   "source_url": "https://omlx.ai/benchmarks/performance/nqxlxtb8",
   "tg_tokens": null,
   "tps": 62.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-q87gmcpb-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: fixjcmca, orec8b8w, xldae0qs; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 661.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 661.2 66.0 70.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q87gmcpb",
   "source_url": "https://omlx.ai/benchmarks/performance/q87gmcpb",
   "tg_tokens": null,
   "tps": 66.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-q87gmcpb-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: fixjcmca, orec8b8w, xldae0qs; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 809.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 809.5 62.1 71.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q87gmcpb",
   "source_url": "https://omlx.ai/benchmarks/performance/q87gmcpb",
   "tg_tokens": null,
   "tps": 62.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-q87gmcpb-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: fixjcmca, orec8b8w, xldae0qs; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1065.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,065 63.5 71.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q87gmcpb",
   "source_url": "https://omlx.ai/benchmarks/performance/q87gmcpb",
   "tg_tokens": null,
   "tps": 63.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-q87gmcpb-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: fixjcmca, orec8b8w, xldae0qs; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1059.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 1,059 60.8 71.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q87gmcpb",
   "source_url": "https://omlx.ai/benchmarks/performance/q87gmcpb",
   "tg_tokens": null,
   "tps": 60.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-uvf7dldt-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: dqd5jhf2, hdoq5euh, lzgt1w9u; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 700.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 700.3 94.5 70.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uvf7dldt",
   "source_url": "https://omlx.ai/benchmarks/performance/uvf7dldt",
   "tg_tokens": null,
   "tps": 94.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-uvf7dldt-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: dqd5jhf2, hdoq5euh, lzgt1w9u; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 858.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 858.0 58.8 71.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uvf7dldt",
   "source_url": "https://omlx.ai/benchmarks/performance/uvf7dldt",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-uvf7dldt-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: dqd5jhf2, hdoq5euh, lzgt1w9u; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1062.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,062 64.1 71.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uvf7dldt",
   "source_url": "https://omlx.ai/benchmarks/performance/uvf7dldt",
   "tg_tokens": null,
   "tps": 64.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-uvf7dldt-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: dqd5jhf2, hdoq5euh, lzgt1w9u; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1059.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 1,059 76.9 71.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uvf7dldt",
   "source_url": "https://omlx.ai/benchmarks/performance/uvf7dldt",
   "tg_tokens": null,
   "tps": 76.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-kimc2691-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 121.1 GB; TTFT 3936 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1041.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 4k tokens 1,041 PP tok/s 70.6 TG tok/s 3936 TTFT (ms) 121.1 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kimc2691",
   "source_url": "https://omlx.ai/benchmarks/performance/kimc2691",
   "tg_tokens": null,
   "tps": 70.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ycz7885v-ctx1024",
   "model": "Coder390",
   "notes": "abbreviated Coder390 title; fp16 MTP build; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: anqa7hjz, bgm5g822, f97s1obd, oaryr1q3, qijfz8f0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 866.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Coder390-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 866.1 47.9 20.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ycz7885v",
   "source_url": "https://omlx.ai/benchmarks/performance/ycz7885v",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ycz7885v-ctx4096",
   "model": "Coder390",
   "notes": "abbreviated Coder390 title; fp16 MTP build; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: anqa7hjz, bgm5g822, f97s1obd, oaryr1q3, qijfz8f0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1002.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Coder390-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,002 48.1 22.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ycz7885v",
   "source_url": "https://omlx.ai/benchmarks/performance/ycz7885v",
   "tg_tokens": null,
   "tps": 48.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ycz7885v-ctx8192",
   "model": "Coder390",
   "notes": "abbreviated Coder390 title; fp16 MTP build; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: anqa7hjz, bgm5g822, f97s1obd, oaryr1q3, qijfz8f0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1074.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Coder390-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,074 52.9 23.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ycz7885v",
   "source_url": "https://omlx.ai/benchmarks/performance/ycz7885v",
   "tg_tokens": null,
   "tps": 52.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ycz7885v-ctx16384",
   "model": "Coder390",
   "notes": "abbreviated Coder390 title; fp16 MTP build; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: anqa7hjz, bgm5g822, f97s1obd, oaryr1q3, qijfz8f0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1066.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Coder390-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,066 85.5 24.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ycz7885v",
   "source_url": "https://omlx.ai/benchmarks/performance/ycz7885v",
   "tg_tokens": null,
   "tps": 85.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ycz7885v-ctx32768",
   "model": "Coder390",
   "notes": "abbreviated Coder390 title; fp16 MTP build; oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: anqa7hjz, bgm5g822, f97s1obd, oaryr1q3, qijfz8f0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 984.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Coder390-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 984.0 116.6 26.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ycz7885v",
   "source_url": "https://omlx.ai/benchmarks/performance/ycz7885v",
   "tg_tokens": null,
   "tps": 116.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ycz7885v-ctx65536",
   "model": "Coder390",
   "notes": "abbreviated Coder390 title; fp16 MTP build; oMLX community benchmark trace; peak memory 31.1 GB at this context; duplicate traces of the same session: anqa7hjz, bgm5g822, f97s1obd, oaryr1q3, qijfz8f0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 784.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Coder390-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 784.6 41.2 31.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ycz7885v",
   "source_url": "https://omlx.ai/benchmarks/performance/ycz7885v",
   "tg_tokens": null,
   "tps": 41.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-mkyhmn1j-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 151.1 GB at this context; duplicate traces of the same session: 5h9kic82, 9ipq2fhc, m3wj0tpk; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 385.7,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "GLM-5.3-Flash-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 1k 385.7 57.3 151.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mkyhmn1j",
   "source_url": "https://omlx.ai/benchmarks/performance/mkyhmn1j",
   "tg_tokens": null,
   "tps": 57.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-mkyhmn1j-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 152.1 GB at this context; duplicate traces of the same session: 5h9kic82, 9ipq2fhc, m3wj0tpk; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 452.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "GLM-5.3-Flash-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 4k 452.0 51.2 152.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mkyhmn1j",
   "source_url": "https://omlx.ai/benchmarks/performance/mkyhmn1j",
   "tg_tokens": null,
   "tps": 51.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-mkyhmn1j-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 152.4 GB at this context; duplicate traces of the same session: 5h9kic82, 9ipq2fhc, m3wj0tpk; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 417.9,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "GLM-5.3-Flash-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 8k 417.9 56.2 152.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mkyhmn1j",
   "source_url": "https://omlx.ai/benchmarks/performance/mkyhmn1j",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-mkyhmn1j-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 152.6 GB at this context; duplicate traces of the same session: 5h9kic82, 9ipq2fhc, m3wj0tpk; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 409.5,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "GLM-5.3-Flash-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 16k 409.5 46.4 152.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mkyhmn1j",
   "source_url": "https://omlx.ai/benchmarks/performance/mkyhmn1j",
   "tg_tokens": null,
   "tps": 46.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-o6g4v79h-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 151.1 GB at this context; duplicate traces of the same session: 2hicq9r0, qwz1fx9c, wnrkhyt4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 388.8,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "GLM-5.3-Flash-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 1k 388.8 63.1 151.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o6g4v79h",
   "source_url": "https://omlx.ai/benchmarks/performance/o6g4v79h",
   "tg_tokens": null,
   "tps": 63.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-o6g4v79h-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 152.1 GB at this context; duplicate traces of the same session: 2hicq9r0, qwz1fx9c, wnrkhyt4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 451.6,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "GLM-5.3-Flash-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 4k 451.6 47.8 152.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o6g4v79h",
   "source_url": "https://omlx.ai/benchmarks/performance/o6g4v79h",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-o6g4v79h-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 152.4 GB at this context; duplicate traces of the same session: 2hicq9r0, qwz1fx9c, wnrkhyt4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 418.5,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "GLM-5.3-Flash-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 8k 418.5 41.5 152.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o6g4v79h",
   "source_url": "https://omlx.ai/benchmarks/performance/o6g4v79h",
   "tg_tokens": null,
   "tps": 41.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-o6g4v79h-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 152.6 GB at this context; duplicate traces of the same session: 2hicq9r0, qwz1fx9c, wnrkhyt4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 410.1,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "GLM-5.3-Flash-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 16k 410.1 41.3 152.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o6g4v79h",
   "source_url": "https://omlx.ai/benchmarks/performance/o6g4v79h",
   "tg_tokens": null,
   "tps": 41.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-7vywf9dh-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 249.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 488.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 1k 488.8 40.7 249.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7vywf9dh",
   "source_url": "https://omlx.ai/benchmarks/performance/7vywf9dh",
   "tg_tokens": null,
   "tps": 40.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-7vywf9dh-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 250.2 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 570.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 4k 570.7 40.9 250.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7vywf9dh",
   "source_url": "https://omlx.ai/benchmarks/performance/7vywf9dh",
   "tg_tokens": null,
   "tps": 40.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-7vywf9dh-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 250.6 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 530.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 8k 530.6 45.1 250.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7vywf9dh",
   "source_url": "https://omlx.ai/benchmarks/performance/7vywf9dh",
   "tg_tokens": null,
   "tps": 45.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-7vywf9dh-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 250.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 520.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 16k 520.6 71.0 250.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7vywf9dh",
   "source_url": "https://omlx.ai/benchmarks/performance/7vywf9dh",
   "tg_tokens": null,
   "tps": 71.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2",
   "id": "omlx-1las0h29-ctx4096",
   "model": "Qwen3 8B",
   "notes": "oMLX community benchmark trace; peak memory 5.4 GB at this context; duplicate traces of the same session: sizr5f4j; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 128.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-8B-4bit on M2 (10c); M2 (10c) \u00b7 16 GB \u00b7 4bit; 4k 128.9 16.7 5.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1las0h29",
   "source_url": "https://omlx.ai/benchmarks/performance/1las0h29",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2",
   "id": "omlx-1las0h29-ctx16384",
   "model": "Qwen3 8B",
   "notes": "oMLX community benchmark trace; peak memory 7.1 GB at this context; duplicate traces of the same session: sizr5f4j; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-8B-4bit on M2 (10c); M2 (10c) \u00b7 16 GB \u00b7 4bit; 16k 105.0 11.9 7.1 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1las0h29",
   "source_url": "https://omlx.ai/benchmarks/performance/1las0h29",
   "tg_tokens": null,
   "tps": 11.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-smtveig5-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 9hvzpgf3, f63rsosw, g04cq1fb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2161.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 2,161 144.1 21.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/smtveig5",
   "source_url": "https://omlx.ai/benchmarks/performance/smtveig5",
   "tg_tokens": null,
   "tps": 144.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-smtveig5-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 9hvzpgf3, f63rsosw, g04cq1fb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2630.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 2,630 156.3 23.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/smtveig5",
   "source_url": "https://omlx.ai/benchmarks/performance/smtveig5",
   "tg_tokens": null,
   "tps": 156.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-smtveig5-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: 9hvzpgf3, f63rsosw, g04cq1fb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2493.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 2,493 135.0 23.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/smtveig5",
   "source_url": "https://omlx.ai/benchmarks/performance/smtveig5",
   "tg_tokens": null,
   "tps": 135.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-smtveig5-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 9hvzpgf3, f63rsosw, g04cq1fb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2222.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 2,222 108.3 23.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/smtveig5",
   "source_url": "https://omlx.ai/benchmarks/performance/smtveig5",
   "tg_tokens": null,
   "tps": 108.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xs4fceie-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: k306jgat; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 860.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 860.9 33.8 16.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xs4fceie",
   "source_url": "https://omlx.ai/benchmarks/performance/xs4fceie",
   "tg_tokens": null,
   "tps": 33.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xs4fceie-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: k306jgat; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 922.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 922.9 33.2 17.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xs4fceie",
   "source_url": "https://omlx.ai/benchmarks/performance/xs4fceie",
   "tg_tokens": null,
   "tps": 33.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7h5ntet6-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: fns5of5j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 766.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 766.5 13.9 25.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7h5ntet6",
   "source_url": "https://omlx.ai/benchmarks/performance/7h5ntet6",
   "tg_tokens": null,
   "tps": 13.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7h5ntet6-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 35.3 GB at this context; duplicate traces of the same session: fns5of5j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 583.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 583.2 10.9 35.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7h5ntet6",
   "source_url": "https://omlx.ai/benchmarks/performance/7h5ntet6",
   "tg_tokens": null,
   "tps": 10.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-np41ehem-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 5wuj95tp, f9yafa0k, v5auyhoy; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 325.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 325.3 76.5 18.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/np41ehem",
   "source_url": "https://omlx.ai/benchmarks/performance/np41ehem",
   "tg_tokens": null,
   "tps": 76.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-np41ehem-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 5wuj95tp, f9yafa0k, v5auyhoy; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 346.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 346.5 58.0 22.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/np41ehem",
   "source_url": "https://omlx.ai/benchmarks/performance/np41ehem",
   "tg_tokens": null,
   "tps": 58.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-np41ehem-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: 5wuj95tp, f9yafa0k, v5auyhoy; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 340.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 340.8 46.0 22.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/np41ehem",
   "source_url": "https://omlx.ai/benchmarks/performance/np41ehem",
   "tg_tokens": null,
   "tps": 46.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-np41ehem-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 5wuj95tp, f9yafa0k, v5auyhoy; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 328.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 328.0 37.1 23.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/np41ehem",
   "source_url": "https://omlx.ai/benchmarks/performance/np41ehem",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-f1nhx6j0-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: agculhpw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 860.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 860.9 42.6 17.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f1nhx6j0",
   "source_url": "https://omlx.ai/benchmarks/performance/f1nhx6j0",
   "tg_tokens": null,
   "tps": 42.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-f1nhx6j0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: agculhpw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 979.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 979.0 46.3 18.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f1nhx6j0",
   "source_url": "https://omlx.ai/benchmarks/performance/f1nhx6j0",
   "tg_tokens": null,
   "tps": 46.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-lu1dcfn4-ctx1024",
   "model": "Qwen3.8-3.6-27B-blend",
   "notes": "blend of Qwen3.8 and Qwen3.6 27B; oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 6ncf51mq, fneyfl80; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 481.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-3.6-27B-blend-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 481.4 17.6 16.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lu1dcfn4",
   "source_url": "https://omlx.ai/benchmarks/performance/lu1dcfn4",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-lu1dcfn4-ctx4096",
   "model": "Qwen3.8-3.6-27B-blend",
   "notes": "blend of Qwen3.8 and Qwen3.6 27B; oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 6ncf51mq, fneyfl80; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 473.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-3.6-27B-blend-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 473.9 17.3 17.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lu1dcfn4",
   "source_url": "https://omlx.ai/benchmarks/performance/lu1dcfn4",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-lu1dcfn4-ctx8192",
   "model": "Qwen3.8-3.6-27B-blend",
   "notes": "blend of Qwen3.8 and Qwen3.6 27B; oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 6ncf51mq, fneyfl80; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 420.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-3.6-27B-blend-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 420.0 16.6 17.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lu1dcfn4",
   "source_url": "https://omlx.ai/benchmarks/performance/lu1dcfn4",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-lu1dcfn4-ctx16384",
   "model": "Qwen3.8-3.6-27B-blend",
   "notes": "blend of Qwen3.8 and Qwen3.6 27B; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 6ncf51mq, fneyfl80; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 415.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-3.6-27B-blend-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 415.6 16.2 17.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lu1dcfn4",
   "source_url": "https://omlx.ai/benchmarks/performance/lu1dcfn4",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-lu1dcfn4-ctx32768",
   "model": "Qwen3.8-3.6-27B-blend",
   "notes": "blend of Qwen3.8 and Qwen3.6 27B; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 6ncf51mq, fneyfl80; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 395.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-3.6-27B-blend-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 395.8 15.2 18.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lu1dcfn4",
   "source_url": "https://omlx.ai/benchmarks/performance/lu1dcfn4",
   "tg_tokens": null,
   "tps": 15.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-lu1dcfn4-ctx65536",
   "model": "Qwen3.8-3.6-27B-blend",
   "notes": "blend of Qwen3.8 and Qwen3.6 27B; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 6ncf51mq, fneyfl80; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 345.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-3.6-27B-blend-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 345.6 13.3 20.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lu1dcfn4",
   "source_url": "https://omlx.ai/benchmarks/performance/lu1dcfn4",
   "tg_tokens": null,
   "tps": 13.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mes8d6r6-ctx1024",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "MiniPlus build; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 4sfjs1bs, opwmnjxt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2046.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-35B-A3B-Distill-MiniPlus on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,046 51.6 21.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mes8d6r6",
   "source_url": "https://omlx.ai/benchmarks/performance/mes8d6r6",
   "tg_tokens": null,
   "tps": 51.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mes8d6r6-ctx4096",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "MiniPlus build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 4sfjs1bs, opwmnjxt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2877.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-35B-A3B-Distill-MiniPlus on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,877 50.9 22.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mes8d6r6",
   "source_url": "https://omlx.ai/benchmarks/performance/mes8d6r6",
   "tg_tokens": null,
   "tps": 50.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mes8d6r6-ctx8192",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "MiniPlus build; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: 4sfjs1bs, opwmnjxt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3000.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-35B-A3B-Distill-MiniPlus on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 3,000 49.8 22.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mes8d6r6",
   "source_url": "https://omlx.ai/benchmarks/performance/mes8d6r6",
   "tg_tokens": null,
   "tps": 49.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-h3jcj8xd-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 39.8 GB at this context; duplicate traces of the same session: 5ome7h5c; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 744.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k 744.4 64.9 39.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h3jcj8xd",
   "source_url": "https://omlx.ai/benchmarks/performance/h3jcj8xd",
   "tg_tokens": null,
   "tps": 64.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-h3jcj8xd-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 40.0 GB at this context; duplicate traces of the same session: 5ome7h5c; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1554.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 4k 1,554 62.7 40.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h3jcj8xd",
   "source_url": "https://omlx.ai/benchmarks/performance/h3jcj8xd",
   "tg_tokens": null,
   "tps": 62.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-15gdd5d5-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 50.2 GB at this context; duplicate traces of the same session: t2zdwb3o; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 233.7,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 2bit; 1k 233.7 19.7 50.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/15gdd5d5",
   "source_url": "https://omlx.ai/benchmarks/performance/15gdd5d5",
   "tg_tokens": null,
   "tps": 19.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-15gdd5d5-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 51.3 GB at this context; duplicate traces of the same session: t2zdwb3o; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 617.4,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 2bit; 16k 617.4 25.7 51.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/15gdd5d5",
   "source_url": "https://omlx.ai/benchmarks/performance/15gdd5d5",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9s286mnq-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 61.4 GB at this context; duplicate traces of the same session: 5hrvylfj, r5xf8nuv; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 697.1,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 1k 697.1 62.5 61.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9s286mnq",
   "source_url": "https://omlx.ai/benchmarks/performance/9s286mnq",
   "tg_tokens": null,
   "tps": 62.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9s286mnq-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 61.9 GB at this context; duplicate traces of the same session: 5hrvylfj, r5xf8nuv; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 859.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 4k 859.0 59.5 61.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9s286mnq",
   "source_url": "https://omlx.ai/benchmarks/performance/9s286mnq",
   "tg_tokens": null,
   "tps": 59.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9s286mnq-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 62.1 GB at this context; duplicate traces of the same session: 5hrvylfj, r5xf8nuv; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1057.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 8k 1,057 88.8 62.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9s286mnq",
   "source_url": "https://omlx.ai/benchmarks/performance/9s286mnq",
   "tg_tokens": null,
   "tps": 88.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ea979jrz-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 61.3 GB at this context; duplicate traces of the same session: a31eduxd, h0g8vdcy, myesswxc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 591.4,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 1k 591.4 48.5 61.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ea979jrz",
   "source_url": "https://omlx.ai/benchmarks/performance/ea979jrz",
   "tg_tokens": null,
   "tps": 48.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ea979jrz-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 61.9 GB at this context; duplicate traces of the same session: a31eduxd, h0g8vdcy, myesswxc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 789.5,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 4k 789.5 48.9 61.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ea979jrz",
   "source_url": "https://omlx.ai/benchmarks/performance/ea979jrz",
   "tg_tokens": null,
   "tps": 48.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ea979jrz-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 62.1 GB at this context; duplicate traces of the same session: a31eduxd, h0g8vdcy, myesswxc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1052.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 8k 1,052 63.9 62.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ea979jrz",
   "source_url": "https://omlx.ai/benchmarks/performance/ea979jrz",
   "tg_tokens": null,
   "tps": 63.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ea979jrz-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 62.4 GB at this context; duplicate traces of the same session: a31eduxd, h0g8vdcy, myesswxc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1054.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 16k 1,054 62.1 62.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ea979jrz",
   "source_url": "https://omlx.ai/benchmarks/performance/ea979jrz",
   "tg_tokens": null,
   "tps": 62.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-i9fomhip-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 600.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 600.6 64.2 70.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i9fomhip",
   "source_url": "https://omlx.ai/benchmarks/performance/i9fomhip",
   "tg_tokens": null,
   "tps": 64.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-i9fomhip-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 769.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 769.9 63.1 71.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i9fomhip",
   "source_url": "https://omlx.ai/benchmarks/performance/i9fomhip",
   "tg_tokens": null,
   "tps": 63.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-i9fomhip-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1060.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,060 60.9 71.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i9fomhip",
   "source_url": "https://omlx.ai/benchmarks/performance/i9fomhip",
   "tg_tokens": null,
   "tps": 60.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-i9fomhip-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1060.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 1,060 62.7 71.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i9fomhip",
   "source_url": "https://omlx.ai/benchmarks/performance/i9fomhip",
   "tg_tokens": null,
   "tps": 62.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2bw12yvv-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: e1a6h6l2, folcrfs1, u3s0p23b; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 594.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 594.5 62.4 70.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2bw12yvv",
   "source_url": "https://omlx.ai/benchmarks/performance/2bw12yvv",
   "tg_tokens": null,
   "tps": 62.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2bw12yvv-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: e1a6h6l2, folcrfs1, u3s0p23b; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 788.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 788.5 60.5 71.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2bw12yvv",
   "source_url": "https://omlx.ai/benchmarks/performance/2bw12yvv",
   "tg_tokens": null,
   "tps": 60.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2bw12yvv-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: e1a6h6l2, folcrfs1, u3s0p23b; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1059.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,059 74.4 71.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2bw12yvv",
   "source_url": "https://omlx.ai/benchmarks/performance/2bw12yvv",
   "tg_tokens": null,
   "tps": 74.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2bw12yvv-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: e1a6h6l2, folcrfs1, u3s0p23b; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1058.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 1,058 105.2 71.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2bw12yvv",
   "source_url": "https://omlx.ai/benchmarks/performance/2bw12yvv",
   "tg_tokens": null,
   "tps": 105.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-amgnxm03-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: 2e844ef9, 8qzprl3h, eevr2wr1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1872.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 1,872 106.3 16.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/amgnxm03",
   "source_url": "https://omlx.ai/benchmarks/performance/amgnxm03",
   "tg_tokens": null,
   "tps": 106.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-amgnxm03-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 2e844ef9, 8qzprl3h, eevr2wr1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1815.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 1,815 78.7 16.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/amgnxm03",
   "source_url": "https://omlx.ai/benchmarks/performance/amgnxm03",
   "tg_tokens": null,
   "tps": 78.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-amgnxm03-ctx8192",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: 2e844ef9, 8qzprl3h, eevr2wr1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1836.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,836 85.5 16.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/amgnxm03",
   "source_url": "https://omlx.ai/benchmarks/performance/amgnxm03",
   "tg_tokens": null,
   "tps": 85.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-amgnxm03-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 2e844ef9, 8qzprl3h, eevr2wr1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1732.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 1,732 76.3 17.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/amgnxm03",
   "source_url": "https://omlx.ai/benchmarks/performance/amgnxm03",
   "tg_tokens": null,
   "tps": 76.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-y8qxft5v-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 15.1 GB at this context; duplicate traces of the same session: jzvrvo8h; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 296.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 16k 296.1 29.1 15.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y8qxft5v",
   "source_url": "https://omlx.ai/benchmarks/performance/y8qxft5v",
   "tg_tokens": null,
   "tps": 29.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-y8qxft5v-ctx32768",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: jzvrvo8h; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 240.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 32k 240.9 22.3 15.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y8qxft5v",
   "source_url": "https://omlx.ai/benchmarks/performance/y8qxft5v",
   "tg_tokens": null,
   "tps": 22.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-txzaz90l-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace, single run at 4k context; peak memory 18.6 GB; TTFT 2898 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1413.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 1,413 PP tok/s 32.7 TG tok/s 2898 TTFT (ms) 18.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/txzaz90l",
   "source_url": "https://omlx.ai/benchmarks/performance/txzaz90l",
   "tg_tokens": null,
   "tps": 32.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5ct0lfa9-ctx1024",
   "model": "GLM-5.3-Flash-Abliterated",
   "notes": "MLX; oMLX community benchmark trace; peak memory 177.8 GB at this context; duplicate traces of the same session: ai582ztv, e4q769bx, x145ajq5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 512.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-Abliterated-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 512.8 49.4 177.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ct0lfa9",
   "source_url": "https://omlx.ai/benchmarks/performance/5ct0lfa9",
   "tg_tokens": null,
   "tps": 49.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5ct0lfa9-ctx4096",
   "model": "GLM-5.3-Flash-Abliterated",
   "notes": "MLX; oMLX community benchmark trace; peak memory 178.9 GB at this context; duplicate traces of the same session: ai582ztv, e4q769bx, x145ajq5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 594.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-Abliterated-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 594.1 44.3 178.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ct0lfa9",
   "source_url": "https://omlx.ai/benchmarks/performance/5ct0lfa9",
   "tg_tokens": null,
   "tps": 44.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5ct0lfa9-ctx8192",
   "model": "GLM-5.3-Flash-Abliterated",
   "notes": "MLX; oMLX community benchmark trace; peak memory 179.3 GB at this context; duplicate traces of the same session: ai582ztv, e4q769bx, x145ajq5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 551.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-Abliterated-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 551.1 53.0 179.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ct0lfa9",
   "source_url": "https://omlx.ai/benchmarks/performance/5ct0lfa9",
   "tg_tokens": null,
   "tps": 53.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5ct0lfa9-ctx16384",
   "model": "GLM-5.3-Flash-Abliterated",
   "notes": "MLX; oMLX community benchmark trace; peak memory 179.4 GB at this context; duplicate traces of the same session: ai582ztv, e4q769bx, x145ajq5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 540.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-Abliterated-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 540.9 91.9 179.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ct0lfa9",
   "source_url": "https://omlx.ai/benchmarks/performance/5ct0lfa9",
   "tg_tokens": null,
   "tps": 91.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-zmc749y5-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: g5yu1qpe, o6fgcjeo; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 606.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 8k 606.3 16.9 6.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zmc749y5",
   "source_url": "https://omlx.ai/benchmarks/performance/zmc749y5",
   "tg_tokens": null,
   "tps": 16.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-zmc749y5-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.6 GB at this context; duplicate traces of the same session: g5yu1qpe, o6fgcjeo; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 588.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 16k 588.3 16.0 6.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zmc749y5",
   "source_url": "https://omlx.ai/benchmarks/performance/zmc749y5",
   "tg_tokens": null,
   "tps": 16.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-zmc749y5-ctx32768",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.1 GB at this context; duplicate traces of the same session: g5yu1qpe, o6fgcjeo; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 471.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 32k 471.4 12.6 7.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zmc749y5",
   "source_url": "https://omlx.ai/benchmarks/performance/zmc749y5",
   "tg_tokens": null,
   "tps": 12.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lvxuk5h6-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: ffravmy7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 939.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 939.0 109.3 35.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lvxuk5h6",
   "source_url": "https://omlx.ai/benchmarks/performance/lvxuk5h6",
   "tg_tokens": null,
   "tps": 109.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lvxuk5h6-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: ffravmy7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2539.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 2,539 108.2 36.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lvxuk5h6",
   "source_url": "https://omlx.ai/benchmarks/performance/lvxuk5h6",
   "tg_tokens": null,
   "tps": 108.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hk8xbn6e-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 68a5pt3g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5497.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 5,497 98.1 21.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hk8xbn6e",
   "source_url": "https://omlx.ai/benchmarks/performance/hk8xbn6e",
   "tg_tokens": null,
   "tps": 98.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hk8xbn6e-ctx16384",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: 68a5pt3g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5680.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 5,680 93.4 21.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hk8xbn6e",
   "source_url": "https://omlx.ai/benchmarks/performance/hk8xbn6e",
   "tg_tokens": null,
   "tps": 93.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-svm7fbcm-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 1vx5kbmy, os7mb7jk, tj0sovg1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2195.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 2,195 97.1 21.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/svm7fbcm",
   "source_url": "https://omlx.ai/benchmarks/performance/svm7fbcm",
   "tg_tokens": null,
   "tps": 97.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-svm7fbcm-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 1vx5kbmy, os7mb7jk, tj0sovg1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2637.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 2,637 146.2 23.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/svm7fbcm",
   "source_url": "https://omlx.ai/benchmarks/performance/svm7fbcm",
   "tg_tokens": null,
   "tps": 146.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-svm7fbcm-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: 1vx5kbmy, os7mb7jk, tj0sovg1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2491.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 2,491 103.0 23.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/svm7fbcm",
   "source_url": "https://omlx.ai/benchmarks/performance/svm7fbcm",
   "tg_tokens": null,
   "tps": 103.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-svm7fbcm-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 1vx5kbmy, os7mb7jk, tj0sovg1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2220.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 2,220 114.4 23.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/svm7fbcm",
   "source_url": "https://omlx.ai/benchmarks/performance/svm7fbcm",
   "tg_tokens": null,
   "tps": 114.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-460dpzw0-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 70owrapf, 9jspzejd, uf77r7rn; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2189.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 2,189 112.9 21.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/460dpzw0",
   "source_url": "https://omlx.ai/benchmarks/performance/460dpzw0",
   "tg_tokens": null,
   "tps": 112.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-460dpzw0-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 70owrapf, 9jspzejd, uf77r7rn; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2634.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 2,634 143.1 23.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/460dpzw0",
   "source_url": "https://omlx.ai/benchmarks/performance/460dpzw0",
   "tg_tokens": null,
   "tps": 143.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-460dpzw0-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: 70owrapf, 9jspzejd, uf77r7rn; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2491.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 2,491 156.7 23.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/460dpzw0",
   "source_url": "https://omlx.ai/benchmarks/performance/460dpzw0",
   "tg_tokens": null,
   "tps": 156.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-460dpzw0-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 70owrapf, 9jspzejd, uf77r7rn; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2222.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 2,222 107.2 23.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/460dpzw0",
   "source_url": "https://omlx.ai/benchmarks/performance/460dpzw0",
   "tg_tokens": null,
   "tps": 107.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-p3gt9de8-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 31.2 GB; TTFT 8554 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3831.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k tokens 3,831 PP tok/s 79.0 TG tok/s 8554 TTFT (ms) 31.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p3gt9de8",
   "source_url": "https://omlx.ai/benchmarks/performance/p3gt9de8",
   "tg_tokens": null,
   "tps": 79.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ulztg1g2-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 32k context; peak memory 38.5 GB; TTFT 7060 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4641.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k tokens 4,641 PP tok/s 171.3 TG tok/s 7060 TTFT (ms) 38.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ulztg1g2",
   "source_url": "https://omlx.ai/benchmarks/performance/ulztg1g2",
   "tg_tokens": null,
   "tps": 171.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-80oor4yd-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 32k context; peak memory 38.5 GB; TTFT 8119 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4036.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k tokens 4,036 PP tok/s 80.7 TG tok/s 8119 TTFT (ms) 38.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/80oor4yd",
   "source_url": "https://omlx.ai/benchmarks/performance/80oor4yd",
   "tg_tokens": null,
   "tps": 80.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-74wkc5q8-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: n9i881d4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 224.0 22.7 16.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/74wkc5q8",
   "source_url": "https://omlx.ai/benchmarks/performance/74wkc5q8",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-74wkc5q8-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: n9i881d4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 237.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 237.4 22.2 19.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/74wkc5q8",
   "source_url": "https://omlx.ai/benchmarks/performance/74wkc5q8",
   "tg_tokens": null,
   "tps": 22.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-783zwtxm-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: xl7ky0ap; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 908.6,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-27B-nvfp4 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 908.6 26.6 17.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/783zwtxm",
   "source_url": "https://omlx.ai/benchmarks/performance/783zwtxm",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-783zwtxm-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: xl7ky0ap; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 973.6,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-27B-nvfp4 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 973.6 27.1 18.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/783zwtxm",
   "source_url": "https://omlx.ai/benchmarks/performance/783zwtxm",
   "tg_tokens": null,
   "tps": 27.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-seumxpmh-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 68.6 GB at this context; duplicate traces of the same session: mxpg2fc3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 140.8,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 140.8 14.7 68.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/seumxpmh",
   "source_url": "https://omlx.ai/benchmarks/performance/seumxpmh",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-seumxpmh-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.8 GB at this context; duplicate traces of the same session: mxpg2fc3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 70.5,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 70.5 11.8 73.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/seumxpmh",
   "source_url": "https://omlx.ai/benchmarks/performance/seumxpmh",
   "tg_tokens": null,
   "tps": 11.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xlpfd483-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 18.9 GB; TTFT 4135 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 990.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 990.7 PP tok/s 63.7 TG tok/s 4135 TTFT (ms) 18.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xlpfd483",
   "source_url": "https://omlx.ai/benchmarks/performance/xlpfd483",
   "tg_tokens": null,
   "tps": 63.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qi0cmjl3-ctx1024",
   "model": "Qwen3.8-3.6-27B-blend",
   "notes": "blend of Qwen3.8 and Qwen3.6 27B; oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 8kgbltnx, mc7xkwio; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 481.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-3.6-27B-blend-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 481.4 17.6 16.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qi0cmjl3",
   "source_url": "https://omlx.ai/benchmarks/performance/qi0cmjl3",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qi0cmjl3-ctx4096",
   "model": "Qwen3.8-3.6-27B-blend",
   "notes": "blend of Qwen3.8 and Qwen3.6 27B; oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 8kgbltnx, mc7xkwio; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 473.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-3.6-27B-blend-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 473.9 17.3 17.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qi0cmjl3",
   "source_url": "https://omlx.ai/benchmarks/performance/qi0cmjl3",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qi0cmjl3-ctx8192",
   "model": "Qwen3.8-3.6-27B-blend",
   "notes": "blend of Qwen3.8 and Qwen3.6 27B; oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 8kgbltnx, mc7xkwio; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 420.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-3.6-27B-blend-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 420.0 16.6 17.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qi0cmjl3",
   "source_url": "https://omlx.ai/benchmarks/performance/qi0cmjl3",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qi0cmjl3-ctx16384",
   "model": "Qwen3.8-3.6-27B-blend",
   "notes": "blend of Qwen3.8 and Qwen3.6 27B; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 8kgbltnx, mc7xkwio; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 415.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-3.6-27B-blend-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 415.6 16.2 17.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qi0cmjl3",
   "source_url": "https://omlx.ai/benchmarks/performance/qi0cmjl3",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qi0cmjl3-ctx32768",
   "model": "Qwen3.8-3.6-27B-blend",
   "notes": "blend of Qwen3.8 and Qwen3.6 27B; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 8kgbltnx, mc7xkwio; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 395.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-3.6-27B-blend-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 395.8 15.2 18.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qi0cmjl3",
   "source_url": "https://omlx.ai/benchmarks/performance/qi0cmjl3",
   "tg_tokens": null,
   "tps": 15.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qi0cmjl3-ctx65536",
   "model": "Qwen3.8-3.6-27B-blend",
   "notes": "blend of Qwen3.8 and Qwen3.6 27B; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 8kgbltnx, mc7xkwio; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 345.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-3.6-27B-blend-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 345.6 13.3 20.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qi0cmjl3",
   "source_url": "https://omlx.ai/benchmarks/performance/qi0cmjl3",
   "tg_tokens": null,
   "tps": 13.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-m7x0x4vo-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 61.4 GB at this context; duplicate traces of the same session: 0a31m2n6, 6iwu0uhf, lorv9fif; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 656.8,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 1k 656.8 56.8 61.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m7x0x4vo",
   "source_url": "https://omlx.ai/benchmarks/performance/m7x0x4vo",
   "tg_tokens": null,
   "tps": 56.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-m7x0x4vo-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 61.9 GB at this context; duplicate traces of the same session: 0a31m2n6, 6iwu0uhf, lorv9fif; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 815.5,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 4k 815.5 53.6 61.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m7x0x4vo",
   "source_url": "https://omlx.ai/benchmarks/performance/m7x0x4vo",
   "tg_tokens": null,
   "tps": 53.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-m7x0x4vo-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 62.1 GB at this context; duplicate traces of the same session: 0a31m2n6, 6iwu0uhf, lorv9fif; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1055.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 8k 1,055 53.0 62.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m7x0x4vo",
   "source_url": "https://omlx.ai/benchmarks/performance/m7x0x4vo",
   "tg_tokens": null,
   "tps": 53.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-m7x0x4vo-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 62.4 GB at this context; duplicate traces of the same session: 0a31m2n6, 6iwu0uhf, lorv9fif; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1056.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 16k 1,056 55.0 62.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m7x0x4vo",
   "source_url": "https://omlx.ai/benchmarks/performance/m7x0x4vo",
   "tg_tokens": null,
   "tps": 55.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-q5jymttf-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 61.3 GB at this context; duplicate traces of the same session: 7njrqc4c, jokwitl0, rzaqddqv; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 590.1,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 1k 590.1 53.7 61.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q5jymttf",
   "source_url": "https://omlx.ai/benchmarks/performance/q5jymttf",
   "tg_tokens": null,
   "tps": 53.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-q5jymttf-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 61.9 GB at this context; duplicate traces of the same session: 7njrqc4c, jokwitl0, rzaqddqv; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 758.8,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 4k 758.8 52.9 61.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q5jymttf",
   "source_url": "https://omlx.ai/benchmarks/performance/q5jymttf",
   "tg_tokens": null,
   "tps": 52.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-q5jymttf-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 62.1 GB at this context; duplicate traces of the same session: 7njrqc4c, jokwitl0, rzaqddqv; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1057.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 8k 1,057 53.9 62.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q5jymttf",
   "source_url": "https://omlx.ai/benchmarks/performance/q5jymttf",
   "tg_tokens": null,
   "tps": 53.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-q5jymttf-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 62.4 GB at this context; duplicate traces of the same session: 7njrqc4c, jokwitl0, rzaqddqv; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1054.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 16k 1,054 52.7 62.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q5jymttf",
   "source_url": "https://omlx.ai/benchmarks/performance/q5jymttf",
   "tg_tokens": null,
   "tps": 52.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-mj460et2-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: 3h8uubpg, 3qcjgg9m, 8tykqsb0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 660.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 660.0 67.4 70.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mj460et2",
   "source_url": "https://omlx.ai/benchmarks/performance/mj460et2",
   "tg_tokens": null,
   "tps": 67.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-mj460et2-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: 3h8uubpg, 3qcjgg9m, 8tykqsb0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 802.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 802.1 63.0 71.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mj460et2",
   "source_url": "https://omlx.ai/benchmarks/performance/mj460et2",
   "tg_tokens": null,
   "tps": 63.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-mj460et2-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: 3h8uubpg, 3qcjgg9m, 8tykqsb0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1068.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,068 63.0 71.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mj460et2",
   "source_url": "https://omlx.ai/benchmarks/performance/mj460et2",
   "tg_tokens": null,
   "tps": 63.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-mj460et2-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 3h8uubpg, 3qcjgg9m, 8tykqsb0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1061.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 1,061 63.7 71.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mj460et2",
   "source_url": "https://omlx.ai/benchmarks/performance/mj460et2",
   "tg_tokens": null,
   "tps": 63.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rym3rmki-ctx65536",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: ekswf7xr, iu2519ox; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 791.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 791.7 25.3 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rym3rmki",
   "source_url": "https://omlx.ai/benchmarks/performance/rym3rmki",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rym3rmki-ctx131072",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: ekswf7xr, iu2519ox; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 641.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 641.8 20.9 25.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rym3rmki",
   "source_url": "https://omlx.ai/benchmarks/performance/rym3rmki",
   "tg_tokens": null,
   "tps": 20.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-rym3rmki-ctx199680",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: ekswf7xr, iu2519ox; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 524.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 195k 524.7 17.4 29.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rym3rmki",
   "source_url": "https://omlx.ai/benchmarks/performance/rym3rmki",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1g8er3x9-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; Swift 1.5 build; oMLX community benchmark trace, single run at 1k context; peak memory 88.0 GB; TTFT 6868 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 149.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 149.1 PP tok/s 21.7 TG tok/s 6868 TTFT (ms) 88.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1g8er3x9",
   "source_url": "https://omlx.ai/benchmarks/performance/1g8er3x9",
   "tg_tokens": null,
   "tps": 21.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hu3a0l98-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1871.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 1,871 113.9 16.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hu3a0l98",
   "source_url": "https://omlx.ai/benchmarks/performance/hu3a0l98",
   "tg_tokens": null,
   "tps": 113.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hu3a0l98-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1889.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 1,889 103.9 16.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hu3a0l98",
   "source_url": "https://omlx.ai/benchmarks/performance/hu3a0l98",
   "tg_tokens": null,
   "tps": 103.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hu3a0l98-ctx8192",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1837.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,837 77.0 16.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hu3a0l98",
   "source_url": "https://omlx.ai/benchmarks/performance/hu3a0l98",
   "tg_tokens": null,
   "tps": 77.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hu3a0l98-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1731.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 1,731 75.9 17.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hu3a0l98",
   "source_url": "https://omlx.ai/benchmarks/performance/hu3a0l98",
   "tg_tokens": null,
   "tps": 75.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-dsap1e24-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 15.1 GB at this context; duplicate traces of the same session: tuqgos3q; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 296.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 16k 296.1 29.1 15.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dsap1e24",
   "source_url": "https://omlx.ai/benchmarks/performance/dsap1e24",
   "tg_tokens": null,
   "tps": 29.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-dsap1e24-ctx32768",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: tuqgos3q; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 239.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 32k 239.7 21.2 15.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dsap1e24",
   "source_url": "https://omlx.ai/benchmarks/performance/dsap1e24",
   "tg_tokens": null,
   "tps": 21.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-gfpf5gsg-ctx16384",
   "model": "gemma-4-26B-A4B-it-qat",
   "notes": "q4_0 mlx-aligned build; oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: jdixv56g; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 272.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-q4_0-mlx-aligned on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 16k 272.5 27.2 16.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gfpf5gsg",
   "source_url": "https://omlx.ai/benchmarks/performance/gfpf5gsg",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-gfpf5gsg-ctx32768",
   "model": "gemma-4-26B-A4B-it-qat",
   "notes": "q4_0 mlx-aligned build; oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: jdixv56g; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 214.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-q4_0-mlx-aligned on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 32k 214.2 20.8 17.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gfpf5gsg",
   "source_url": "https://omlx.ai/benchmarks/performance/gfpf5gsg",
   "tg_tokens": null,
   "tps": 20.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-dncqstsv-ctx1024",
   "model": "Gemma-4-31B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: ucidznrq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 179.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31b-it-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 179.9 18.0 17.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dncqstsv",
   "source_url": "https://omlx.ai/benchmarks/performance/dncqstsv",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-dncqstsv-ctx4096",
   "model": "Gemma-4-31B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: ucidznrq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 175.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31b-it-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 175.9 17.9 18.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dncqstsv",
   "source_url": "https://omlx.ai/benchmarks/performance/dncqstsv",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0ky5ltmi-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace, single run at 4k context; peak memory 18.6 GB; TTFT 3349 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1223.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 1,223 PP tok/s 31.0 TG tok/s 3349 TTFT (ms) 18.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ky5ltmi",
   "source_url": "https://omlx.ai/benchmarks/performance/0ky5ltmi",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-oh5t5kji-ctx16384",
   "model": "gemma-4-26B-A4B-it-qat",
   "notes": "unsloth; oMLX community benchmark trace; peak memory 15.3 GB at this context; duplicate traces of the same session: 5pts3c2i; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 295.1,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "unsloth-gemma-4-26B-A4B-it-qat-oQ4 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 16k 295.1 27.7 15.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oh5t5kji",
   "source_url": "https://omlx.ai/benchmarks/performance/oh5t5kji",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-oh5t5kji-ctx32768",
   "model": "gemma-4-26B-A4B-it-qat",
   "notes": "unsloth; oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: 5pts3c2i; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 233.5,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "unsloth-gemma-4-26B-A4B-it-qat-oQ4 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 32k 233.5 20.9 16.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oh5t5kji",
   "source_url": "https://omlx.ai/benchmarks/performance/oh5t5kji",
   "tg_tokens": null,
   "tps": 20.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-dfre5x4w-ctx4096",
   "model": "llama-3-2-1b-instruct",
   "notes": "oMLX community benchmark trace; peak memory 2.9 GB at this context; duplicate traces of the same session: emheaw27; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4157.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Llama-3.2-1B-Instruct on M5 (10c); M5 (10c) \u00b7 16 GB \u00b7 bf16; 4k 4,157 50.2 2.9 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dfre5x4w",
   "source_url": "https://omlx.ai/benchmarks/performance/dfre5x4w",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-dfre5x4w-ctx16384",
   "model": "llama-3-2-1b-instruct",
   "notes": "oMLX community benchmark trace; peak memory 3.3 GB at this context; duplicate traces of the same session: emheaw27; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3651.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Llama-3.2-1B-Instruct on M5 (10c); M5 (10c) \u00b7 16 GB \u00b7 bf16; 16k 3,651 42.7 3.3 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dfre5x4w",
   "source_url": "https://omlx.ai/benchmarks/performance/dfre5x4w",
   "tg_tokens": null,
   "tps": 42.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-u14v7r1e-ctx1024",
   "model": "GLM-4.7-Flash",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 23.3 GB; TTFT 1037 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 987.2,
   "provenance": "community",
   "quant": "6bit",
   "quote": "GLM-4.7-Flash-MLX-6bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k tokens 987.2 PP tok/s 61.5 TG tok/s 1037 TTFT (ms) 23.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u14v7r1e",
   "source_url": "https://omlx.ai/benchmarks/performance/u14v7r1e",
   "tg_tokens": null,
   "tps": 61.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5eklraam-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 151.1 GB at this context; duplicate traces of the same session: fryyhpaf, x5a0702x, xglmhivp; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 386.5,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "GLM-5.3-Flash-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 1k 386.5 50.7 151.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5eklraam",
   "source_url": "https://omlx.ai/benchmarks/performance/5eklraam",
   "tg_tokens": null,
   "tps": 50.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5eklraam-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 152.1 GB at this context; duplicate traces of the same session: fryyhpaf, x5a0702x, xglmhivp; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 451.8,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "GLM-5.3-Flash-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 4k 451.8 43.2 152.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5eklraam",
   "source_url": "https://omlx.ai/benchmarks/performance/5eklraam",
   "tg_tokens": null,
   "tps": 43.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5eklraam-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 152.4 GB at this context; duplicate traces of the same session: fryyhpaf, x5a0702x, xglmhivp; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 418.6,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "GLM-5.3-Flash-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 8k 418.6 48.7 152.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5eklraam",
   "source_url": "https://omlx.ai/benchmarks/performance/5eklraam",
   "tg_tokens": null,
   "tps": 48.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5eklraam-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 152.6 GB at this context; duplicate traces of the same session: fryyhpaf, x5a0702x, xglmhivp; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 410.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "GLM-5.3-Flash-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 3bit; 16k 410.0 43.3 152.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5eklraam",
   "source_url": "https://omlx.ai/benchmarks/performance/5eklraam",
   "tg_tokens": null,
   "tps": 43.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-w48z5vcg-ctx1024",
   "model": "Qwen2.5-Coder-14B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 8.5 GB at this context; duplicate traces of the same session: 8d1rh77x, d9zkqb2p; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 789.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-14B-Instruct-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 789.5 32.5 8.5 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w48z5vcg",
   "source_url": "https://omlx.ai/benchmarks/performance/w48z5vcg",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-w48z5vcg-ctx4096",
   "model": "Qwen2.5-Coder-14B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 9.1 GB at this context; duplicate traces of the same session: 8d1rh77x, d9zkqb2p; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 785.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-14B-Instruct-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 785.1 30.4 9.1 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w48z5vcg",
   "source_url": "https://omlx.ai/benchmarks/performance/w48z5vcg",
   "tg_tokens": null,
   "tps": 30.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-w48z5vcg-ctx8192",
   "model": "Qwen2.5-Coder-14B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 9.9 GB at this context; duplicate traces of the same session: 8d1rh77x, d9zkqb2p; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 739.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-14B-Instruct-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 8k 739.3 27.2 9.9 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w48z5vcg",
   "source_url": "https://omlx.ai/benchmarks/performance/w48z5vcg",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nd1pgt4g-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 3vmbicbf, b91w65k7, z2wezf3g; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2193.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 2,193 97.3 21.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nd1pgt4g",
   "source_url": "https://omlx.ai/benchmarks/performance/nd1pgt4g",
   "tg_tokens": null,
   "tps": 97.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-nd1pgt4g-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 3vmbicbf, b91w65k7, z2wezf3g; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2630.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 2,630 153.0 23.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nd1pgt4g",
   "source_url": "https://omlx.ai/benchmarks/performance/nd1pgt4g",
   "tg_tokens": null,
   "tps": 153.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nd1pgt4g-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: 3vmbicbf, b91w65k7, z2wezf3g; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2489.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 2,489 90.0 23.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nd1pgt4g",
   "source_url": "https://omlx.ai/benchmarks/performance/nd1pgt4g",
   "tg_tokens": null,
   "tps": 90.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nd1pgt4g-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 3vmbicbf, b91w65k7, z2wezf3g; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2222.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 2,222 95.7 23.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nd1pgt4g",
   "source_url": "https://omlx.ai/benchmarks/performance/nd1pgt4g",
   "tg_tokens": null,
   "tps": 95.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-te26y1d7-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 5hp0ucay, rsbjudp2, yajx79rr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2190.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 2,190 90.6 21.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/te26y1d7",
   "source_url": "https://omlx.ai/benchmarks/performance/te26y1d7",
   "tg_tokens": null,
   "tps": 90.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-te26y1d7-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 5hp0ucay, rsbjudp2, yajx79rr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2626.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 2,626 162.3 23.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/te26y1d7",
   "source_url": "https://omlx.ai/benchmarks/performance/te26y1d7",
   "tg_tokens": null,
   "tps": 162.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-te26y1d7-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: 5hp0ucay, rsbjudp2, yajx79rr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2491.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 2,491 88.6 23.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/te26y1d7",
   "source_url": "https://omlx.ai/benchmarks/performance/te26y1d7",
   "tg_tokens": null,
   "tps": 88.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-te26y1d7-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 5hp0ucay, rsbjudp2, yajx79rr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2222.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 2,222 95.5 23.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/te26y1d7",
   "source_url": "https://omlx.ai/benchmarks/performance/te26y1d7",
   "tg_tokens": null,
   "tps": 95.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-dd7j2vk2-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: mlopwwz7, nmd7w1vk, y5buoei1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 325.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 325.0 40.7 18.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dd7j2vk2",
   "source_url": "https://omlx.ai/benchmarks/performance/dd7j2vk2",
   "tg_tokens": null,
   "tps": 40.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-dd7j2vk2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: mlopwwz7, nmd7w1vk, y5buoei1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 346.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 346.5 38.0 22.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dd7j2vk2",
   "source_url": "https://omlx.ai/benchmarks/performance/dd7j2vk2",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-dd7j2vk2-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: mlopwwz7, nmd7w1vk, y5buoei1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 340.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 340.9 40.8 22.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dd7j2vk2",
   "source_url": "https://omlx.ai/benchmarks/performance/dd7j2vk2",
   "tg_tokens": null,
   "tps": 40.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-dd7j2vk2-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: mlopwwz7, nmd7w1vk, y5buoei1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 327.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 327.9 35.3 23.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dd7j2vk2",
   "source_url": "https://omlx.ai/benchmarks/performance/dd7j2vk2",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3dbgj5k9-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: jby9ds3v, ktj48ine; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 799.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 799.0 25.3 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3dbgj5k9",
   "source_url": "https://omlx.ai/benchmarks/performance/3dbgj5k9",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3dbgj5k9-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: jby9ds3v, ktj48ine; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 633.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 633.8 20.5 25.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3dbgj5k9",
   "source_url": "https://omlx.ai/benchmarks/performance/3dbgj5k9",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3dbgj5k9-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: jby9ds3v, ktj48ine; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 523.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 195k 523.3 17.6 29.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3dbgj5k9",
   "source_url": "https://omlx.ai/benchmarks/performance/3dbgj5k9",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-4xcqwqo7-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 50.2 GB at this context; duplicate traces of the same session: ukeohrah; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 236.8,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 2bit; 1k 236.8 19.0 50.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4xcqwqo7",
   "source_url": "https://omlx.ai/benchmarks/performance/4xcqwqo7",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-4xcqwqo7-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 51.3 GB at this context; duplicate traces of the same session: ukeohrah; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 638.4,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 2bit; 16k 638.4 15.8 51.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4xcqwqo7",
   "source_url": "https://omlx.ai/benchmarks/performance/4xcqwqo7",
   "tg_tokens": null,
   "tps": 15.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-gf9jtfng-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 47.8 GB at this context; duplicate traces of the same session: 1yw30ifb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 521.5,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 2bit; 4k 521.5 37.9 47.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gf9jtfng",
   "source_url": "https://omlx.ai/benchmarks/performance/gf9jtfng",
   "tg_tokens": null,
   "tps": 37.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-gf9jtfng-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 48.1 GB at this context; duplicate traces of the same session: 1yw30ifb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 666.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 2bit; 16k 666.0 11.4 48.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gf9jtfng",
   "source_url": "https://omlx.ai/benchmarks/performance/gf9jtfng",
   "tg_tokens": null,
   "tps": 11.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-kdjspuig-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 47.3 GB at this context; duplicate traces of the same session: lrxqowqc, q85w96pi; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 343.4,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 2bit; 1k 343.4 32.7 47.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kdjspuig",
   "source_url": "https://omlx.ai/benchmarks/performance/kdjspuig",
   "tg_tokens": null,
   "tps": 32.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-kdjspuig-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 47.8 GB at this context; duplicate traces of the same session: lrxqowqc, q85w96pi; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 361.8,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 2bit; 4k 361.8 31.3 47.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kdjspuig",
   "source_url": "https://omlx.ai/benchmarks/performance/kdjspuig",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-kdjspuig-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 48.1 GB at this context; duplicate traces of the same session: lrxqowqc, q85w96pi; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 664.4,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 2bit; 16k 664.4 11.0 48.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kdjspuig",
   "source_url": "https://omlx.ai/benchmarks/performance/kdjspuig",
   "tg_tokens": null,
   "tps": 11.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-52gtfail-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 50.2 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 431.4,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 2bit; 1k 431.4 25.6 50.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/52gtfail",
   "source_url": "https://omlx.ai/benchmarks/performance/52gtfail",
   "tg_tokens": null,
   "tps": 25.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-52gtfail-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 50.8 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 342.4,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 2bit; 4k 342.4 40.3 50.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/52gtfail",
   "source_url": "https://omlx.ai/benchmarks/performance/52gtfail",
   "tg_tokens": null,
   "tps": 40.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-52gtfail-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 51.3 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 637.7,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 2bit; 16k 637.7 17.5 51.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/52gtfail",
   "source_url": "https://omlx.ai/benchmarks/performance/52gtfail",
   "tg_tokens": null,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-tl8wtaoj-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 79.9 GB; TTFT 322303 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 406.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 128k tokens 406.7 PP tok/s 32.6 TG tok/s 322303 TTFT (ms) 79.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tl8wtaoj",
   "source_url": "https://omlx.ai/benchmarks/performance/tl8wtaoj",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-sunuyc3i-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 1upqij6u, llq56stc, vvddoips; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 552.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 552.8 54.2 70.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sunuyc3i",
   "source_url": "https://omlx.ai/benchmarks/performance/sunuyc3i",
   "tg_tokens": null,
   "tps": 54.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-sunuyc3i-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: 1upqij6u, llq56stc, vvddoips; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 726.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 726.1 52.3 71.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sunuyc3i",
   "source_url": "https://omlx.ai/benchmarks/performance/sunuyc3i",
   "tg_tokens": null,
   "tps": 52.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-sunuyc3i-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: 1upqij6u, llq56stc, vvddoips; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1046.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,046 53.4 71.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sunuyc3i",
   "source_url": "https://omlx.ai/benchmarks/performance/sunuyc3i",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-sunuyc3i-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 1upqij6u, llq56stc, vvddoips; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1056.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 1,056 57.2 71.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sunuyc3i",
   "source_url": "https://omlx.ai/benchmarks/performance/sunuyc3i",
   "tg_tokens": null,
   "tps": 57.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-ogud2fpn-ctx1024",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: 0mj5fh9w, a3daplgg, cder0nze; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 59.5,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 2bit; 1k 59.5 9.0 11.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ogud2fpn",
   "source_url": "https://omlx.ai/benchmarks/performance/ogud2fpn",
   "tg_tokens": null,
   "tps": 9.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-ogud2fpn-ctx4096",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 14.5 GB at this context; duplicate traces of the same session: 0mj5fh9w, a3daplgg, cder0nze; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 58.0,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 2bit; 4k 58.0 8.8 14.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ogud2fpn",
   "source_url": "https://omlx.ai/benchmarks/performance/ogud2fpn",
   "tg_tokens": null,
   "tps": 8.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-ogud2fpn-ctx8192",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: 0mj5fh9w, a3daplgg, cder0nze; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 56.6,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 2bit; 8k 56.6 8.2 15.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ogud2fpn",
   "source_url": "https://omlx.ai/benchmarks/performance/ogud2fpn",
   "tg_tokens": null,
   "tps": 8.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-ogud2fpn-ctx16384",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: 0mj5fh9w, a3daplgg, cder0nze; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 53.1,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 2bit; 16k 53.1 7.7 16.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ogud2fpn",
   "source_url": "https://omlx.ai/benchmarks/performance/ogud2fpn",
   "tg_tokens": null,
   "tps": 7.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-tn71ngjp-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: 028wr3vk, p6tdafsf; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1871.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 1,871 113.9 16.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tn71ngjp",
   "source_url": "https://omlx.ai/benchmarks/performance/tn71ngjp",
   "tg_tokens": null,
   "tps": 113.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-tn71ngjp-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 028wr3vk, p6tdafsf; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1889.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 1,889 103.9 16.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tn71ngjp",
   "source_url": "https://omlx.ai/benchmarks/performance/tn71ngjp",
   "tg_tokens": null,
   "tps": 103.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-tn71ngjp-ctx8192",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: 028wr3vk, p6tdafsf; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1837.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,837 77.0 16.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tn71ngjp",
   "source_url": "https://omlx.ai/benchmarks/performance/tn71ngjp",
   "tg_tokens": null,
   "tps": 77.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-tn71ngjp-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 028wr3vk, p6tdafsf; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1731.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 1,731 75.9 17.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tn71ngjp",
   "source_url": "https://omlx.ai/benchmarks/performance/tn71ngjp",
   "tg_tokens": null,
   "tps": 75.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-6mrt8jfy-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: gae6h9yq, k0z1tiim, yyoqe6pq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1873.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 1,873 105.7 16.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6mrt8jfy",
   "source_url": "https://omlx.ai/benchmarks/performance/6mrt8jfy",
   "tg_tokens": null,
   "tps": 105.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-6mrt8jfy-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: gae6h9yq, k0z1tiim, yyoqe6pq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1844.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 1,844 88.8 16.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6mrt8jfy",
   "source_url": "https://omlx.ai/benchmarks/performance/6mrt8jfy",
   "tg_tokens": null,
   "tps": 88.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-6mrt8jfy-ctx8192",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: gae6h9yq, k0z1tiim, yyoqe6pq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1839.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,839 85.2 16.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6mrt8jfy",
   "source_url": "https://omlx.ai/benchmarks/performance/6mrt8jfy",
   "tg_tokens": null,
   "tps": 85.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-6mrt8jfy-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: gae6h9yq, k0z1tiim, yyoqe6pq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1730.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 1,730 75.4 17.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6mrt8jfy",
   "source_url": "https://omlx.ai/benchmarks/performance/6mrt8jfy",
   "tg_tokens": null,
   "tps": 75.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ma4igz4l-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: c8d1mn0n, zlyfebxd; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1867.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 1,867 169.0 16.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ma4igz4l",
   "source_url": "https://omlx.ai/benchmarks/performance/ma4igz4l",
   "tg_tokens": null,
   "tps": 169.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ma4igz4l-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: c8d1mn0n, zlyfebxd; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1826.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 1,826 115.6 16.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ma4igz4l",
   "source_url": "https://omlx.ai/benchmarks/performance/ma4igz4l",
   "tg_tokens": null,
   "tps": 115.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ma4igz4l-ctx8192",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: c8d1mn0n, zlyfebxd; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1821.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,821 85.2 16.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ma4igz4l",
   "source_url": "https://omlx.ai/benchmarks/performance/ma4igz4l",
   "tg_tokens": null,
   "tps": 85.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-43vkcnjo-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: 0whrq9x0, e2qvzwgz, upx73cu3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1861.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 1,861 98.3 16.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/43vkcnjo",
   "source_url": "https://omlx.ai/benchmarks/performance/43vkcnjo",
   "tg_tokens": null,
   "tps": 98.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-43vkcnjo-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 0whrq9x0, e2qvzwgz, upx73cu3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1886.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 1,886 113.9 16.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/43vkcnjo",
   "source_url": "https://omlx.ai/benchmarks/performance/43vkcnjo",
   "tg_tokens": null,
   "tps": 113.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-43vkcnjo-ctx8192",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: 0whrq9x0, e2qvzwgz, upx73cu3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1825.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,825 106.5 16.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/43vkcnjo",
   "source_url": "https://omlx.ai/benchmarks/performance/43vkcnjo",
   "tg_tokens": null,
   "tps": 106.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-43vkcnjo-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 0whrq9x0, e2qvzwgz, upx73cu3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1714.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 1,714 64.7 17.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/43vkcnjo",
   "source_url": "https://omlx.ai/benchmarks/performance/43vkcnjo",
   "tg_tokens": null,
   "tps": 64.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-cb5d24mp-ctx16384",
   "model": "gemma-4-26B-A4B-it-qat",
   "notes": "q4_0 mlx-aligned build; oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: oynziblz; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 290.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-q4_0-mlx-aligned on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 16k 290.3 28.9 16.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cb5d24mp",
   "source_url": "https://omlx.ai/benchmarks/performance/cb5d24mp",
   "tg_tokens": null,
   "tps": 28.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-cb5d24mp-ctx32768",
   "model": "gemma-4-26B-A4B-it-qat",
   "notes": "q4_0 mlx-aligned build; oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: oynziblz; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 237.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-q4_0-mlx-aligned on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 32k 237.8 21.3 17.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cb5d24mp",
   "source_url": "https://omlx.ai/benchmarks/performance/cb5d24mp",
   "tg_tokens": null,
   "tps": 21.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-us09wwy1-ctx4096",
   "model": "glm53",
   "notes": "abbreviated GLM-5.3 trace title; oMLX community benchmark trace; peak memory 178.9 GB at this context; duplicate traces of the same session: yucpf9cm; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 361.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "glm53 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 361.5 18.0 178.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/us09wwy1",
   "source_url": "https://omlx.ai/benchmarks/performance/us09wwy1",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-us09wwy1-ctx8192",
   "model": "glm53",
   "notes": "abbreviated GLM-5.3 trace title; oMLX community benchmark trace; peak memory 179.0 GB at this context; duplicate traces of the same session: yucpf9cm; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 254.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "glm53 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 254.5 19.6 179.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/us09wwy1",
   "source_url": "https://omlx.ai/benchmarks/performance/us09wwy1",
   "tg_tokens": null,
   "tps": 19.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-ktwspagd-ctx16384",
   "model": "gemma-4-26B-A4B-it-qat",
   "notes": "unsloth; oMLX community benchmark trace; peak memory 15.3 GB at this context; duplicate traces of the same session: v2saerfs; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 296.2,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "unsloth-gemma-4-26B-A4B-it-qat-oQ4 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 16k 296.2 27.7 15.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ktwspagd",
   "source_url": "https://omlx.ai/benchmarks/performance/ktwspagd",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-ktwspagd-ctx32768",
   "model": "gemma-4-26B-A4B-it-qat",
   "notes": "unsloth; oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: v2saerfs; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 246.1,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "unsloth-gemma-4-26B-A4B-it-qat-oQ4 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 32k 246.1 21.9 16.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ktwspagd",
   "source_url": "https://omlx.ai/benchmarks/performance/ktwspagd",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gv0ytczr-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 354.8 GB at this context; duplicate traces of the same session: 4gyjky6p, f13lqu1z, ltf72iw1, o33jk5op, yrxouup8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 252.3,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-oQ3e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 252.3 19.4 354.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gv0ytczr",
   "source_url": "https://omlx.ai/benchmarks/performance/gv0ytczr",
   "tg_tokens": null,
   "tps": 19.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gv0ytczr-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 355.3 GB at this context; duplicate traces of the same session: 4gyjky6p, f13lqu1z, ltf72iw1, o33jk5op, yrxouup8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 332.4,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-oQ3e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 332.4 19.0 355.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gv0ytczr",
   "source_url": "https://omlx.ai/benchmarks/performance/gv0ytczr",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gv0ytczr-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 355.6 GB at this context; duplicate traces of the same session: 4gyjky6p, f13lqu1z, ltf72iw1, o33jk5op, yrxouup8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 359.3,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-oQ3e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 8k 359.3 18.6 355.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gv0ytczr",
   "source_url": "https://omlx.ai/benchmarks/performance/gv0ytczr",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gv0ytczr-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 356.1 GB at this context; duplicate traces of the same session: 4gyjky6p, f13lqu1z, ltf72iw1, o33jk5op, yrxouup8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 330.8,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-oQ3e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 16k 330.8 21.1 356.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gv0ytczr",
   "source_url": "https://omlx.ai/benchmarks/performance/gv0ytczr",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gv0ytczr-ctx32768",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 356.1 GB at this context; duplicate traces of the same session: 4gyjky6p, f13lqu1z, ltf72iw1, o33jk5op, yrxouup8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 374.7,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-oQ3e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 32k 374.7 20.8 356.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gv0ytczr",
   "source_url": "https://omlx.ai/benchmarks/performance/gv0ytczr",
   "tg_tokens": null,
   "tps": 20.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gv0ytczr-ctx65536",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 356.2 GB at this context; duplicate traces of the same session: 4gyjky6p, f13lqu1z, ltf72iw1, o33jk5op, yrxouup8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 366.2,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-oQ3e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 64k 366.2 20.8 356.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gv0ytczr",
   "source_url": "https://omlx.ai/benchmarks/performance/gv0ytczr",
   "tg_tokens": null,
   "tps": 20.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-v0r5n4f2-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "TextOnly build; oMLX community benchmark trace, single run at 64k context; peak memory 15.7 GB; TTFT 615022 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.6,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Qwen3.8-27B-2bit-MLX-TextOnly on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 2bit; 64k tokens 106.6 PP tok/s 12.8 TG tok/s 615022 TTFT (ms) 15.7 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v0r5n4f2",
   "source_url": "https://omlx.ai/benchmarks/performance/v0r5n4f2",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qcmw2f7z-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 855.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 855.1 36.4 33.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qcmw2f7z",
   "source_url": "https://omlx.ai/benchmarks/performance/qcmw2f7z",
   "tg_tokens": null,
   "tps": 36.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qcmw2f7z-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 954.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 954.0 32.7 34.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qcmw2f7z",
   "source_url": "https://omlx.ai/benchmarks/performance/qcmw2f7z",
   "tg_tokens": null,
   "tps": 32.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-60uskvku-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: xineeloy; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 456.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 456.4 18.2 16.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/60uskvku",
   "source_url": "https://omlx.ai/benchmarks/performance/60uskvku",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-60uskvku-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: xineeloy; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 506.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 506.4 17.8 17.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/60uskvku",
   "source_url": "https://omlx.ai/benchmarks/performance/60uskvku",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hftww8yl-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: ncjw1gzt, twyeheyw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 925.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-mlx-4Bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 925.7 32.8 16.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hftww8yl",
   "source_url": "https://omlx.ai/benchmarks/performance/hftww8yl",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hftww8yl-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: ncjw1gzt, twyeheyw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 912.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-mlx-4Bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 912.0 32.1 17.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hftww8yl",
   "source_url": "https://omlx.ai/benchmarks/performance/hftww8yl",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hftww8yl-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: ncjw1gzt, twyeheyw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 787.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-mlx-4Bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 787.6 30.9 17.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hftww8yl",
   "source_url": "https://omlx.ai/benchmarks/performance/hftww8yl",
   "tg_tokens": null,
   "tps": 30.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-x6sqg4k0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 19.6 GB; TTFT 39157 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k tokens 104.6 PP tok/s 24.5 TG tok/s 39157 TTFT (ms) 19.6 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x6sqg4k0",
   "source_url": "https://omlx.ai/benchmarks/performance/x6sqg4k0",
   "tg_tokens": null,
   "tps": 24.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-3qksioab-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 51.0 GB at this context; duplicate traces of the same session: 9zckoq94, cf5ukxbs, in5w2634, ni0s5z7j, tn8cw5gu, yfm439dr, yrgwb1jm; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 299.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 8bit; 1k 299.3 42.1 51.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3qksioab",
   "source_url": "https://omlx.ai/benchmarks/performance/3qksioab",
   "tg_tokens": null,
   "tps": 42.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-3qksioab-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 54.2 GB at this context; duplicate traces of the same session: 9zckoq94, cf5ukxbs, in5w2634, ni0s5z7j, tn8cw5gu, yfm439dr, yrgwb1jm; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 447.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 8bit; 4k 447.2 40.1 54.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3qksioab",
   "source_url": "https://omlx.ai/benchmarks/performance/3qksioab",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-3qksioab-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 55.0 GB at this context; duplicate traces of the same session: 9zckoq94, cf5ukxbs, in5w2634, ni0s5z7j, tn8cw5gu, yfm439dr, yrgwb1jm; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 354.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 8bit; 8k 354.1 30.6 55.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3qksioab",
   "source_url": "https://omlx.ai/benchmarks/performance/3qksioab",
   "tg_tokens": null,
   "tps": 30.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-3qksioab-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 56.4 GB at this context; duplicate traces of the same session: 9zckoq94, cf5ukxbs, in5w2634, ni0s5z7j, tn8cw5gu, yfm439dr, yrgwb1jm; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 320.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 8bit; 16k 320.4 36.6 56.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3qksioab",
   "source_url": "https://omlx.ai/benchmarks/performance/3qksioab",
   "tg_tokens": null,
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-3qksioab-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 57.9 GB at this context; duplicate traces of the same session: 9zckoq94, cf5ukxbs, in5w2634, ni0s5z7j, tn8cw5gu, yfm439dr, yrgwb1jm; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 327.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 8bit; 32k 327.9 36.4 57.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3qksioab",
   "source_url": "https://omlx.ai/benchmarks/performance/3qksioab",
   "tg_tokens": null,
   "tps": 36.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-3qksioab-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 60.1 GB at this context; duplicate traces of the same session: 9zckoq94, cf5ukxbs, in5w2634, ni0s5z7j, tn8cw5gu, yfm439dr, yrgwb1jm; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 269.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 8bit; 64k 269.7 33.7 60.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3qksioab",
   "source_url": "https://omlx.ai/benchmarks/performance/3qksioab",
   "tg_tokens": null,
   "tps": 33.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-3qksioab-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 64.5 GB at this context; duplicate traces of the same session: 9zckoq94, cf5ukxbs, in5w2634, ni0s5z7j, tn8cw5gu, yfm439dr, yrgwb1jm; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 218.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 8bit; 128k 218.1 27.5 64.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3qksioab",
   "source_url": "https://omlx.ai/benchmarks/performance/3qksioab",
   "tg_tokens": null,
   "tps": 27.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-3qksioab-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 69.5 GB at this context; duplicate traces of the same session: 9zckoq94, cf5ukxbs, in5w2634, ni0s5z7j, tn8cw5gu, yfm439dr, yrgwb1jm; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 169.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 8bit; 195k 169.3 21.4 69.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3qksioab",
   "source_url": "https://omlx.ai/benchmarks/performance/3qksioab",
   "tg_tokens": null,
   "tps": 21.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-u4tn4uju-ctx1024",
   "model": "Qwen3.8-3.6-27B-blend",
   "notes": "blend of Qwen3.8 and Qwen3.6 27B; oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: 6zq0tdcc, 8mmmaiac, iy7l466b, lwj70chp, z9u0ni4o; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 496.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-3.6-27B-blend-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 496.5 30.4 16.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u4tn4uju",
   "source_url": "https://omlx.ai/benchmarks/performance/u4tn4uju",
   "tg_tokens": null,
   "tps": 30.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-u4tn4uju-ctx4096",
   "model": "Qwen3.8-3.6-27B-blend",
   "notes": "blend of Qwen3.8 and Qwen3.6 27B; oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: 6zq0tdcc, 8mmmaiac, iy7l466b, lwj70chp, z9u0ni4o; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 469.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-3.6-27B-blend-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 469.1 22.9 18.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u4tn4uju",
   "source_url": "https://omlx.ai/benchmarks/performance/u4tn4uju",
   "tg_tokens": null,
   "tps": 22.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-u4tn4uju-ctx8192",
   "model": "Qwen3.8-3.6-27B-blend",
   "notes": "blend of Qwen3.8 and Qwen3.6 27B; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 6zq0tdcc, 8mmmaiac, iy7l466b, lwj70chp, z9u0ni4o; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 427.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-3.6-27B-blend-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 427.0 22.2 18.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u4tn4uju",
   "source_url": "https://omlx.ai/benchmarks/performance/u4tn4uju",
   "tg_tokens": null,
   "tps": 22.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-u4tn4uju-ctx16384",
   "model": "Qwen3.8-3.6-27B-blend",
   "notes": "blend of Qwen3.8 and Qwen3.6 27B; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: 6zq0tdcc, 8mmmaiac, iy7l466b, lwj70chp, z9u0ni4o; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 415.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-3.6-27B-blend-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 415.1 27.4 18.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u4tn4uju",
   "source_url": "https://omlx.ai/benchmarks/performance/u4tn4uju",
   "tg_tokens": null,
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-u4tn4uju-ctx32768",
   "model": "Qwen3.8-3.6-27B-blend",
   "notes": "blend of Qwen3.8 and Qwen3.6 27B; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 6zq0tdcc, 8mmmaiac, iy7l466b, lwj70chp, z9u0ni4o; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 410.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-3.6-27B-blend-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 410.8 24.5 19.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u4tn4uju",
   "source_url": "https://omlx.ai/benchmarks/performance/u4tn4uju",
   "tg_tokens": null,
   "tps": 24.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-u4tn4uju-ctx65536",
   "model": "Qwen3.8-3.6-27B-blend",
   "notes": "blend of Qwen3.8 and Qwen3.6 27B; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 6zq0tdcc, 8mmmaiac, iy7l466b, lwj70chp, z9u0ni4o; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 368.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-3.6-27B-blend-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 368.4 22.8 21.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u4tn4uju",
   "source_url": "https://omlx.ai/benchmarks/performance/u4tn4uju",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-nrpt4uf1-ctx1024",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "MiniPlus build; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2628.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-35B-A3B-Distill-MiniPlus on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,628 51.4 21.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nrpt4uf1",
   "source_url": "https://omlx.ai/benchmarks/performance/nrpt4uf1",
   "tg_tokens": null,
   "tps": 51.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-nrpt4uf1-ctx4096",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "MiniPlus build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3261.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-35B-A3B-Distill-MiniPlus on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 3,261 50.7 22.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nrpt4uf1",
   "source_url": "https://omlx.ai/benchmarks/performance/nrpt4uf1",
   "tg_tokens": null,
   "tps": 50.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-nrpt4uf1-ctx8192",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "MiniPlus build; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3017.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-35B-A3B-Distill-MiniPlus on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 3,017 49.6 22.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nrpt4uf1",
   "source_url": "https://omlx.ai/benchmarks/performance/nrpt4uf1",
   "tg_tokens": null,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-3c62f661-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 50.2 GB at this context; duplicate traces of the same session: u0kfp1a1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 431.4,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 2bit; 1k 431.4 25.6 50.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3c62f661",
   "source_url": "https://omlx.ai/benchmarks/performance/3c62f661",
   "tg_tokens": null,
   "tps": 25.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-3c62f661-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 50.8 GB at this context; duplicate traces of the same session: u0kfp1a1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 342.4,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 2bit; 4k 342.4 40.3 50.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3c62f661",
   "source_url": "https://omlx.ai/benchmarks/performance/3c62f661",
   "tg_tokens": null,
   "tps": 40.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-3c62f661-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 51.3 GB at this context; duplicate traces of the same session: u0kfp1a1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 637.7,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 2bit; 16k 637.7 17.5 51.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3c62f661",
   "source_url": "https://omlx.ai/benchmarks/performance/3c62f661",
   "tg_tokens": null,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-b8nt5vb3-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 3q0d6qat, 3qekuuev, 5sfu3ozz, s46rvk7f, zbpuv98q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1071.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,071 66.3 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b8nt5vb3",
   "source_url": "https://omlx.ai/benchmarks/performance/b8nt5vb3",
   "tg_tokens": null,
   "tps": 66.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-b8nt5vb3-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: 3q0d6qat, 3qekuuev, 5sfu3ozz, s46rvk7f, zbpuv98q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1898.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,898 59.6 70.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b8nt5vb3",
   "source_url": "https://omlx.ai/benchmarks/performance/b8nt5vb3",
   "tg_tokens": null,
   "tps": 59.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-b8nt5vb3-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 3q0d6qat, 3qekuuev, 5sfu3ozz, s46rvk7f, zbpuv98q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2379.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,379 81.3 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b8nt5vb3",
   "source_url": "https://omlx.ai/benchmarks/performance/b8nt5vb3",
   "tg_tokens": null,
   "tps": 81.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-b8nt5vb3-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.7 GB at this context; duplicate traces of the same session: 3q0d6qat, 3qekuuev, 5sfu3ozz, s46rvk7f, zbpuv98q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2279.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,279 71.9 74.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b8nt5vb3",
   "source_url": "https://omlx.ai/benchmarks/performance/b8nt5vb3",
   "tg_tokens": null,
   "tps": 71.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-b8nt5vb3-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.1 GB at this context; duplicate traces of the same session: 3q0d6qat, 3qekuuev, 5sfu3ozz, s46rvk7f, zbpuv98q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2216.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,216 67.8 79.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b8nt5vb3",
   "source_url": "https://omlx.ai/benchmarks/performance/b8nt5vb3",
   "tg_tokens": null,
   "tps": 67.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-b8nt5vb3-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.9 GB at this context; duplicate traces of the same session: 3q0d6qat, 3qekuuev, 5sfu3ozz, s46rvk7f, zbpuv98q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2166.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,166 58.6 80.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b8nt5vb3",
   "source_url": "https://omlx.ai/benchmarks/performance/b8nt5vb3",
   "tg_tokens": null,
   "tps": 58.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0pca06ow-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: ad96o06l, e44g4jtt, jfbwa6j0, oi97epc3, uj24cbz9, vhklz541; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1182.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,182 85.1 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0pca06ow",
   "source_url": "https://omlx.ai/benchmarks/performance/0pca06ow",
   "tg_tokens": null,
   "tps": 85.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0pca06ow-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: ad96o06l, e44g4jtt, jfbwa6j0, oi97epc3, uj24cbz9, vhklz541; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2030.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,030 77.4 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0pca06ow",
   "source_url": "https://omlx.ai/benchmarks/performance/0pca06ow",
   "tg_tokens": null,
   "tps": 77.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0pca06ow-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: ad96o06l, e44g4jtt, jfbwa6j0, oi97epc3, uj24cbz9, vhklz541; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2433.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,433 72.1 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0pca06ow",
   "source_url": "https://omlx.ai/benchmarks/performance/0pca06ow",
   "tg_tokens": null,
   "tps": 72.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0pca06ow-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: ad96o06l, e44g4jtt, jfbwa6j0, oi97epc3, uj24cbz9, vhklz541; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2542.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,542 74.7 73.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0pca06ow",
   "source_url": "https://omlx.ai/benchmarks/performance/0pca06ow",
   "tg_tokens": null,
   "tps": 74.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0pca06ow-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.6 GB at this context; duplicate traces of the same session: ad96o06l, e44g4jtt, jfbwa6j0, oi97epc3, uj24cbz9, vhklz541; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2397.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,397 81.0 74.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0pca06ow",
   "source_url": "https://omlx.ai/benchmarks/performance/0pca06ow",
   "tg_tokens": null,
   "tps": 81.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0pca06ow-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.0 GB at this context; duplicate traces of the same session: ad96o06l, e44g4jtt, jfbwa6j0, oi97epc3, uj24cbz9, vhklz541; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2312.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,312 71.6 79.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0pca06ow",
   "source_url": "https://omlx.ai/benchmarks/performance/0pca06ow",
   "tg_tokens": null,
   "tps": 71.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0pca06ow-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.9 GB at this context; duplicate traces of the same session: ad96o06l, e44g4jtt, jfbwa6j0, oi97epc3, uj24cbz9, vhklz541; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2412.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,412 61.2 80.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0pca06ow",
   "source_url": "https://omlx.ai/benchmarks/performance/0pca06ow",
   "tg_tokens": null,
   "tps": 61.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3nzz130z-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: 50tinjty, 7j3yv2kb, e3kuno0r, fxw1mfwz, ue4c6xgb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 335.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 335.4 64.6 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3nzz130z",
   "source_url": "https://omlx.ai/benchmarks/performance/3nzz130z",
   "tg_tokens": null,
   "tps": 64.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3nzz130z-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: 50tinjty, 7j3yv2kb, e3kuno0r, fxw1mfwz, ue4c6xgb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 929.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 929.4 28.4 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3nzz130z",
   "source_url": "https://omlx.ai/benchmarks/performance/3nzz130z",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3nzz130z-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: 50tinjty, 7j3yv2kb, e3kuno0r, fxw1mfwz, ue4c6xgb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2601.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,601 53.7 72.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3nzz130z",
   "source_url": "https://omlx.ai/benchmarks/performance/3nzz130z",
   "tg_tokens": null,
   "tps": 53.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3nzz130z-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.5 GB at this context; duplicate traces of the same session: 50tinjty, 7j3yv2kb, e3kuno0r, fxw1mfwz, ue4c6xgb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2410.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,410 66.5 74.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3nzz130z",
   "source_url": "https://omlx.ai/benchmarks/performance/3nzz130z",
   "tg_tokens": null,
   "tps": 66.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3nzz130z-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.7 GB at this context; duplicate traces of the same session: 50tinjty, 7j3yv2kb, e3kuno0r, fxw1mfwz, ue4c6xgb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2349.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,349 64.9 78.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3nzz130z",
   "source_url": "https://omlx.ai/benchmarks/performance/3nzz130z",
   "tg_tokens": null,
   "tps": 64.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3nzz130z-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.9 GB at this context; duplicate traces of the same session: 50tinjty, 7j3yv2kb, e3kuno0r, fxw1mfwz, ue4c6xgb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2397.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,397 64.3 80.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3nzz130z",
   "source_url": "https://omlx.ai/benchmarks/performance/3nzz130z",
   "tg_tokens": null,
   "tps": 64.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5y7eu2xs-ctx32768",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.2 GB at this context; duplicate traces of the same session: 4h3f5v12, kttfjg1h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3178.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 3,178 99.9 16.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5y7eu2xs",
   "source_url": "https://omlx.ai/benchmarks/performance/5y7eu2xs",
   "tg_tokens": null,
   "tps": 99.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5y7eu2xs-ctx65536",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 4h3f5v12, kttfjg1h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2752.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 2,752 84.3 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5y7eu2xs",
   "source_url": "https://omlx.ai/benchmarks/performance/5y7eu2xs",
   "tg_tokens": null,
   "tps": 84.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5y7eu2xs-ctx131072",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 4h3f5v12, kttfjg1h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1855.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 1,855 65.0 21.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5y7eu2xs",
   "source_url": "https://omlx.ai/benchmarks/performance/5y7eu2xs",
   "tg_tokens": null,
   "tps": 65.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8uysmy3d-ctx32768",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: f2caredu, gnqk92mw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3167.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "gemma-4-26B-A4B-it-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 3,167 86.6 22.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8uysmy3d",
   "source_url": "https://omlx.ai/benchmarks/performance/8uysmy3d",
   "tg_tokens": null,
   "tps": 86.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8uysmy3d-ctx65536",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: f2caredu, gnqk92mw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2621.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "gemma-4-26B-A4B-it-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 64k 2,621 73.8 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8uysmy3d",
   "source_url": "https://omlx.ai/benchmarks/performance/8uysmy3d",
   "tg_tokens": null,
   "tps": 73.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8uysmy3d-ctx131072",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.9 GB at this context; duplicate traces of the same session: f2caredu, gnqk92mw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1787.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "gemma-4-26B-A4B-it-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 128k 1,787 58.7 26.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8uysmy3d",
   "source_url": "https://omlx.ai/benchmarks/performance/8uysmy3d",
   "tg_tokens": null,
   "tps": 58.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-wy7xks4x-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 15.1 GB at this context; duplicate traces of the same session: 67rzbns1; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 299.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 16k 299.8 29.1 15.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wy7xks4x",
   "source_url": "https://omlx.ai/benchmarks/performance/wy7xks4x",
   "tg_tokens": null,
   "tps": 29.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-wy7xks4x-ctx32768",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: 67rzbns1; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 258.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 32k 258.7 24.3 15.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wy7xks4x",
   "source_url": "https://omlx.ai/benchmarks/performance/wy7xks4x",
   "tg_tokens": null,
   "tps": 24.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-uvguz0ub-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "mlx-community repack; oMLX community benchmark trace; peak memory 43.1 GB at this context; duplicate traces of the same session: mdhsd67m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1108.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3-Coder-Next-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,108 54.4 43.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uvguz0ub",
   "source_url": "https://omlx.ai/benchmarks/performance/uvguz0ub",
   "tg_tokens": null,
   "tps": 54.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-uvguz0ub-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "mlx-community repack; oMLX community benchmark trace; peak memory 43.9 GB at this context; duplicate traces of the same session: mdhsd67m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1363.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3-Coder-Next-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,363 54.5 43.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uvguz0ub",
   "source_url": "https://omlx.ai/benchmarks/performance/uvguz0ub",
   "tg_tokens": null,
   "tps": 54.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-i0zgtemw-ctx32768",
   "model": "Laguna XS 2.1",
   "notes": "oMLX community benchmark trace; peak memory 27.1 GB at this context; duplicate traces of the same session: 9erfohhv, vj7b0gri; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2417.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Laguna-XS-2.1-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 2,417 90.0 27.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i0zgtemw",
   "source_url": "https://omlx.ai/benchmarks/performance/i0zgtemw",
   "tg_tokens": null,
   "tps": 90.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-i0zgtemw-ctx65536",
   "model": "Laguna XS 2.1",
   "notes": "oMLX community benchmark trace; peak memory 28.3 GB at this context; duplicate traces of the same session: 9erfohhv, vj7b0gri; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2014.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Laguna-XS-2.1-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 64k 2,014 72.8 28.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i0zgtemw",
   "source_url": "https://omlx.ai/benchmarks/performance/i0zgtemw",
   "tg_tokens": null,
   "tps": 72.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-i0zgtemw-ctx131072",
   "model": "Laguna XS 2.1",
   "notes": "oMLX community benchmark trace; peak memory 30.9 GB at this context; duplicate traces of the same session: 9erfohhv, vj7b0gri; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1430.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Laguna-XS-2.1-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 128k 1,430 54.5 30.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i0zgtemw",
   "source_url": "https://omlx.ai/benchmarks/performance/i0zgtemw",
   "tg_tokens": null,
   "tps": 54.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jo39oha6-ctx32768",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: 3nmsmh42, gdjqbxpk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3438.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ6 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 3,438 115.1 27.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jo39oha6",
   "source_url": "https://omlx.ai/benchmarks/performance/jo39oha6",
   "tg_tokens": null,
   "tps": 115.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jo39oha6-ctx65536",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: 3nmsmh42, gdjqbxpk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3253.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ6 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 64k 3,253 103.3 27.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jo39oha6",
   "source_url": "https://omlx.ai/benchmarks/performance/jo39oha6",
   "tg_tokens": null,
   "tps": 103.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jo39oha6-ctx131072",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.2 GB at this context; duplicate traces of the same session: 3nmsmh42, gdjqbxpk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2685.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ6 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 128k 2,685 90.8 28.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jo39oha6",
   "source_url": "https://omlx.ai/benchmarks/performance/jo39oha6",
   "tg_tokens": null,
   "tps": 90.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qkg9p6wt-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: 2pcb0k5n, x1s8jkl2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3548.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "Ornith-1.5-35B-A3B-oQ5-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 3,548 134.1 26.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qkg9p6wt",
   "source_url": "https://omlx.ai/benchmarks/performance/qkg9p6wt",
   "tg_tokens": null,
   "tps": 134.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qkg9p6wt-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: 2pcb0k5n, x1s8jkl2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4924.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "Ornith-1.5-35B-A3B-oQ5-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 4,924 145.5 26.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qkg9p6wt",
   "source_url": "https://omlx.ai/benchmarks/performance/qkg9p6wt",
   "tg_tokens": null,
   "tps": 145.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qkg9p6wt-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: 2pcb0k5n, x1s8jkl2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4874.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "Ornith-1.5-35B-A3B-oQ5-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 4,874 168.7 26.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qkg9p6wt",
   "source_url": "https://omlx.ai/benchmarks/performance/qkg9p6wt",
   "tg_tokens": null,
   "tps": 168.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4h5oncy0-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: c6zumuf0, rsuoo9gm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3578.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "Qwen3.6-35B-A3B-oQ5-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 3,578 147.5 26.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4h5oncy0",
   "source_url": "https://omlx.ai/benchmarks/performance/4h5oncy0",
   "tg_tokens": null,
   "tps": 147.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4h5oncy0-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: c6zumuf0, rsuoo9gm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4914.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "Qwen3.6-35B-A3B-oQ5-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 4,914 152.1 26.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4h5oncy0",
   "source_url": "https://omlx.ai/benchmarks/performance/4h5oncy0",
   "tg_tokens": null,
   "tps": 152.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4h5oncy0-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: c6zumuf0, rsuoo9gm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4874.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "Qwen3.6-35B-A3B-oQ5-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 4,874 163.7 26.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4h5oncy0",
   "source_url": "https://omlx.ai/benchmarks/performance/4h5oncy0",
   "tg_tokens": null,
   "tps": 163.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-l07as8xc-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: fwcg5bn3, xnep1y4z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3679.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "Qwen3.6-35B-A3B-oQ5-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 3,679 103.7 26.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l07as8xc",
   "source_url": "https://omlx.ai/benchmarks/performance/l07as8xc",
   "tg_tokens": null,
   "tps": 103.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-l07as8xc-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: fwcg5bn3, xnep1y4z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5043.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "Qwen3.6-35B-A3B-oQ5-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 5,043 124.8 26.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l07as8xc",
   "source_url": "https://omlx.ai/benchmarks/performance/l07as8xc",
   "tg_tokens": null,
   "tps": 124.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-l07as8xc-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: fwcg5bn3, xnep1y4z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4985.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "Qwen3.6-35B-A3B-oQ5-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 4,985 121.6 26.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l07as8xc",
   "source_url": "https://omlx.ai/benchmarks/performance/l07as8xc",
   "tg_tokens": null,
   "tps": 121.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ab1sgnr6-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: jpjng5xx; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 398.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 398.8 23.5 16.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ab1sgnr6",
   "source_url": "https://omlx.ai/benchmarks/performance/ab1sgnr6",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ab1sgnr6-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: jpjng5xx; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 382.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 382.1 24.8 18.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ab1sgnr6",
   "source_url": "https://omlx.ai/benchmarks/performance/ab1sgnr6",
   "tg_tokens": null,
   "tps": 24.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vwfqj3xo-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 855.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 855.1 36.4 33.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vwfqj3xo",
   "source_url": "https://omlx.ai/benchmarks/performance/vwfqj3xo",
   "tg_tokens": null,
   "tps": 36.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vwfqj3xo-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 954.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 954.0 32.7 34.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vwfqj3xo",
   "source_url": "https://omlx.ai/benchmarks/performance/vwfqj3xo",
   "tg_tokens": null,
   "tps": 32.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-dm10kah4-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 8tre1gb8; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 74.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 74.7 12.5 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dm10kah4",
   "source_url": "https://omlx.ai/benchmarks/performance/dm10kah4",
   "tg_tokens": null,
   "tps": 12.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-dm10kah4-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 8tre1gb8; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 82.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 82.4 11.2 19.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dm10kah4",
   "source_url": "https://omlx.ai/benchmarks/performance/dm10kah4",
   "tg_tokens": null,
   "tps": 11.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-erdvwe32-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: v2f3031c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 220.9 29.1 16.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/erdvwe32",
   "source_url": "https://omlx.ai/benchmarks/performance/erdvwe32",
   "tg_tokens": null,
   "tps": 29.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-erdvwe32-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: v2f3031c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 255.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 255.7 28.2 19.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/erdvwe32",
   "source_url": "https://omlx.ai/benchmarks/performance/erdvwe32",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-6cm1rk1f-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 1cdyr2ke, 9ou9tqwn, b91baq4z; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 215.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 215.9 41.0 18.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6cm1rk1f",
   "source_url": "https://omlx.ai/benchmarks/performance/6cm1rk1f",
   "tg_tokens": null,
   "tps": 41.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-6cm1rk1f-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: 1cdyr2ke, 9ou9tqwn, b91baq4z; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 219.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 219.9 38.9 22.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6cm1rk1f",
   "source_url": "https://omlx.ai/benchmarks/performance/6cm1rk1f",
   "tg_tokens": null,
   "tps": 38.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-6cm1rk1f-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 1cdyr2ke, 9ou9tqwn, b91baq4z; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 218.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 8k 218.0 36.3 23.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6cm1rk1f",
   "source_url": "https://omlx.ai/benchmarks/performance/6cm1rk1f",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-6cm1rk1f-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 1cdyr2ke, 9ou9tqwn, b91baq4z; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 16k 205.9 37.7 22.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6cm1rk1f",
   "source_url": "https://omlx.ai/benchmarks/performance/6cm1rk1f",
   "tg_tokens": null,
   "tps": 37.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-28z9ys28-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: k3qsq7gi; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 127.2 21.9 18.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/28z9ys28",
   "source_url": "https://omlx.ai/benchmarks/performance/28z9ys28",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-28z9ys28-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: k3qsq7gi; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 131.4 27.5 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/28z9ys28",
   "source_url": "https://omlx.ai/benchmarks/performance/28z9ys28",
   "tg_tokens": null,
   "tps": 27.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-v1s3l2rm-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: 1ckg98m8, 1xvzf62e, fmy3v8pj, p6o2bgdo, sov4rjn6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 900.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 900.9 59.4 16.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v1s3l2rm",
   "source_url": "https://omlx.ai/benchmarks/performance/v1s3l2rm",
   "tg_tokens": null,
   "tps": 59.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-v1s3l2rm-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.9 GB at this context; duplicate traces of the same session: 1ckg98m8, 1xvzf62e, fmy3v8pj, p6o2bgdo, sov4rjn6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 935.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 935.2 51.7 17.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v1s3l2rm",
   "source_url": "https://omlx.ai/benchmarks/performance/v1s3l2rm",
   "tg_tokens": null,
   "tps": 51.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-v1s3l2rm-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.2 GB at this context; duplicate traces of the same session: 1ckg98m8, 1xvzf62e, fmy3v8pj, p6o2bgdo, sov4rjn6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 781.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 781.5 59.2 18.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v1s3l2rm",
   "source_url": "https://omlx.ai/benchmarks/performance/v1s3l2rm",
   "tg_tokens": null,
   "tps": 59.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-v1s3l2rm-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 1ckg98m8, 1xvzf62e, fmy3v8pj, p6o2bgdo, sov4rjn6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 733.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 733.6 59.6 18.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v1s3l2rm",
   "source_url": "https://omlx.ai/benchmarks/performance/v1s3l2rm",
   "tg_tokens": null,
   "tps": 59.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-v1s3l2rm-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 1ckg98m8, 1xvzf62e, fmy3v8pj, p6o2bgdo, sov4rjn6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 657.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 657.6 56.5 19.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v1s3l2rm",
   "source_url": "https://omlx.ai/benchmarks/performance/v1s3l2rm",
   "tg_tokens": null,
   "tps": 56.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-v1s3l2rm-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: 1ckg98m8, 1xvzf62e, fmy3v8pj, p6o2bgdo, sov4rjn6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 576.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 576.0 31.8 22.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v1s3l2rm",
   "source_url": "https://omlx.ai/benchmarks/performance/v1s3l2rm",
   "tg_tokens": null,
   "tps": 31.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ixjqol8x-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.7 GB at this context; duplicate traces of the same session: efy5l4t2, pi40qcxq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 840.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 840.7 21.5 25.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ixjqol8x",
   "source_url": "https://omlx.ai/benchmarks/performance/ixjqol8x",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ixjqol8x-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: efy5l4t2, pi40qcxq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 740.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 64k 740.9 19.6 27.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ixjqol8x",
   "source_url": "https://omlx.ai/benchmarks/performance/ixjqol8x",
   "tg_tokens": null,
   "tps": 19.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ixjqol8x-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.5 GB at this context; duplicate traces of the same session: efy5l4t2, pi40qcxq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 585.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 128k 585.4 16.8 31.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ixjqol8x",
   "source_url": "https://omlx.ai/benchmarks/performance/ixjqol8x",
   "tg_tokens": null,
   "tps": 16.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mvztx8nm-ctx1024",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "MiniPlus build; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: hgqqnq9k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2628.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-35B-A3B-Distill-MiniPlus on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,628 51.4 21.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mvztx8nm",
   "source_url": "https://omlx.ai/benchmarks/performance/mvztx8nm",
   "tg_tokens": null,
   "tps": 51.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mvztx8nm-ctx4096",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "MiniPlus build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: hgqqnq9k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3261.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-35B-A3B-Distill-MiniPlus on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 3,261 50.7 22.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mvztx8nm",
   "source_url": "https://omlx.ai/benchmarks/performance/mvztx8nm",
   "tg_tokens": null,
   "tps": 50.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mvztx8nm-ctx8192",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "MiniPlus build; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: hgqqnq9k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3017.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-35B-A3B-Distill-MiniPlus on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 3,017 49.6 22.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mvztx8nm",
   "source_url": "https://omlx.ai/benchmarks/performance/mvztx8nm",
   "tg_tokens": null,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-sk5ygc6n-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 70.3 GB; TTFT 530 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1931.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,931 PP tok/s 89.1 TG tok/s 530 TTFT (ms) 70.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sk5ygc6n",
   "source_url": "https://omlx.ai/benchmarks/performance/sk5ygc6n",
   "tg_tokens": null,
   "tps": 89.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ah0ssd8e-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: fd9c15ty, u0lykeeb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1201.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,201 92.4 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ah0ssd8e",
   "source_url": "https://omlx.ai/benchmarks/performance/ah0ssd8e",
   "tg_tokens": null,
   "tps": 92.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ah0ssd8e-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: fd9c15ty, u0lykeeb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2038.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,038 85.0 70.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ah0ssd8e",
   "source_url": "https://omlx.ai/benchmarks/performance/ah0ssd8e",
   "tg_tokens": null,
   "tps": 85.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ah0ssd8e-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: fd9c15ty, u0lykeeb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2519.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,519 87.8 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ah0ssd8e",
   "source_url": "https://omlx.ai/benchmarks/performance/ah0ssd8e",
   "tg_tokens": null,
   "tps": 87.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-63i2gy9v-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 26ptuvr0, 3gmwhhyx, cjjn844j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 406.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 406.9 27.9 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/63i2gy9v",
   "source_url": "https://omlx.ai/benchmarks/performance/63i2gy9v",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-63i2gy9v-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: 26ptuvr0, 3gmwhhyx, cjjn844j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1067.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,067 67.0 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/63i2gy9v",
   "source_url": "https://omlx.ai/benchmarks/performance/63i2gy9v",
   "tg_tokens": null,
   "tps": 67.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-63i2gy9v-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 26ptuvr0, 3gmwhhyx, cjjn844j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1701.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,701 55.0 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/63i2gy9v",
   "source_url": "https://omlx.ai/benchmarks/performance/63i2gy9v",
   "tg_tokens": null,
   "tps": 55.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-63i2gy9v-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: 26ptuvr0, 3gmwhhyx, cjjn844j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2596.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,596 22.1 73.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/63i2gy9v",
   "source_url": "https://omlx.ai/benchmarks/performance/63i2gy9v",
   "tg_tokens": null,
   "tps": 22.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-63i2gy9v-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.8 GB at this context; duplicate traces of the same session: 26ptuvr0, 3gmwhhyx, cjjn844j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2448.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,448 71.8 74.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/63i2gy9v",
   "source_url": "https://omlx.ai/benchmarks/performance/63i2gy9v",
   "tg_tokens": null,
   "tps": 71.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-63i2gy9v-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.4 GB at this context; duplicate traces of the same session: 26ptuvr0, 3gmwhhyx, cjjn844j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2402.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,402 62.8 79.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/63i2gy9v",
   "source_url": "https://omlx.ai/benchmarks/performance/63i2gy9v",
   "tg_tokens": null,
   "tps": 62.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-63i2gy9v-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.9 GB at this context; duplicate traces of the same session: 26ptuvr0, 3gmwhhyx, cjjn844j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2413.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,413 63.9 80.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/63i2gy9v",
   "source_url": "https://omlx.ai/benchmarks/performance/63i2gy9v",
   "tg_tokens": null,
   "tps": 63.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-1o16gv0a-ctx1024",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 7.4 GB at this context; duplicate traces of the same session: ba6m88li; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 230.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 230.5 28.6 7.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1o16gv0a",
   "source_url": "https://omlx.ai/benchmarks/performance/1o16gv0a",
   "tg_tokens": null,
   "tps": 28.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-1o16gv0a-ctx4096",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 7.7 GB at this context; duplicate traces of the same session: ba6m88li; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 189.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 189.3 24.5 7.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1o16gv0a",
   "source_url": "https://omlx.ai/benchmarks/performance/1o16gv0a",
   "tg_tokens": null,
   "tps": 24.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-2d601e5l-ctx16384",
   "model": "gemma-4-26B-A4B-it-qat",
   "notes": "q4_0 mlx-aligned build; oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: n7ocx0od; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 269.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-q4_0-mlx-aligned on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 16k 269.4 26.7 16.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2d601e5l",
   "source_url": "https://omlx.ai/benchmarks/performance/2d601e5l",
   "tg_tokens": null,
   "tps": 26.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-2d601e5l-ctx32768",
   "model": "gemma-4-26B-A4B-it-qat",
   "notes": "q4_0 mlx-aligned build; oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: n7ocx0od; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 207.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-q4_0-mlx-aligned on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 32k 207.1 20.5 17.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2d601e5l",
   "source_url": "https://omlx.ai/benchmarks/performance/2d601e5l",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gz8cdz9k-ctx32768",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: pnd3w653, z08y3x55; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 627.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 627.1 21.9 23.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gz8cdz9k",
   "source_url": "https://omlx.ai/benchmarks/performance/gz8cdz9k",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-gz8cdz9k-ctx65536",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: pnd3w653, z08y3x55; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 533.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 533.7 19.4 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gz8cdz9k",
   "source_url": "https://omlx.ai/benchmarks/performance/gz8cdz9k",
   "tg_tokens": null,
   "tps": 19.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-gz8cdz9k-ctx131072",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.8 GB at this context; duplicate traces of the same session: pnd3w653, z08y3x55; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 411.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 411.8 15.1 36.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gz8cdz9k",
   "source_url": "https://omlx.ai/benchmarks/performance/gz8cdz9k",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-sxaqzcg5-ctx16384",
   "model": "gemma-4-26B-A4B-it-qat",
   "notes": "unsloth; oMLX community benchmark trace; peak memory 15.3 GB at this context; duplicate traces of the same session: atovogd9; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 292.5,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "unsloth-gemma-4-26B-A4B-it-qat-oQ4 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 16k 292.5 27.8 15.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sxaqzcg5",
   "source_url": "https://omlx.ai/benchmarks/performance/sxaqzcg5",
   "tg_tokens": null,
   "tps": 27.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-sxaqzcg5-ctx32768",
   "model": "gemma-4-26B-A4B-it-qat",
   "notes": "unsloth; oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: atovogd9; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 227.1,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "unsloth-gemma-4-26B-A4B-it-qat-oQ4 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 32k 227.1 20.4 16.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sxaqzcg5",
   "source_url": "https://omlx.ai/benchmarks/performance/sxaqzcg5",
   "tg_tokens": null,
   "tps": 20.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-hufcy77i-ctx16384",
   "model": "gemma-4-26B-A4B-it-qat",
   "notes": "unsloth; oMLX community benchmark trace; peak memory 15.3 GB at this context; duplicate traces of the same session: k2ir1xh0; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 260.2,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "unsloth-gemma-4-26B-A4B-it-qat-oQ4 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 16k 260.2 26.0 15.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hufcy77i",
   "source_url": "https://omlx.ai/benchmarks/performance/hufcy77i",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-hufcy77i-ctx32768",
   "model": "gemma-4-26B-A4B-it-qat",
   "notes": "unsloth; oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: k2ir1xh0; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 216.6,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "unsloth-gemma-4-26B-A4B-it-qat-oQ4 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 32k 216.6 20.7 16.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hufcy77i",
   "source_url": "https://omlx.ai/benchmarks/performance/hufcy77i",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-rxc43w39-ctx16384",
   "model": "gemma-4-26B-A4B-it-qat",
   "notes": "unsloth; oMLX community benchmark trace; peak memory 15.3 GB at this context; duplicate traces of the same session: zlikqlzj; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 294.2,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "unsloth-gemma-4-26B-A4B-it-qat-oQ4 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 16k 294.2 27.7 15.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rxc43w39",
   "source_url": "https://omlx.ai/benchmarks/performance/rxc43w39",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-rxc43w39-ctx32768",
   "model": "gemma-4-26B-A4B-it-qat",
   "notes": "unsloth; oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: zlikqlzj; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 240.7,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "unsloth-gemma-4-26B-A4B-it-qat-oQ4 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 32k 240.7 21.7 16.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rxc43w39",
   "source_url": "https://omlx.ai/benchmarks/performance/rxc43w39",
   "tg_tokens": null,
   "tps": 21.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-q1ykhvxf-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 240.2 GB at this context; duplicate traces of the same session: 3u8xporm, 6p2u31bk, 8gotsepw, gonah5ay, n322vdoi; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 422.2,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-oQ3e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 422.2 21.3 240.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q1ykhvxf",
   "source_url": "https://omlx.ai/benchmarks/performance/q1ykhvxf",
   "tg_tokens": null,
   "tps": 21.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-q1ykhvxf-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 240.7 GB at this context; duplicate traces of the same session: 3u8xporm, 6p2u31bk, 8gotsepw, gonah5ay, n322vdoi; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 608.0,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-oQ3e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 608.0 21.2 240.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q1ykhvxf",
   "source_url": "https://omlx.ai/benchmarks/performance/q1ykhvxf",
   "tg_tokens": null,
   "tps": 21.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-q1ykhvxf-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 240.8 GB at this context; duplicate traces of the same session: 3u8xporm, 6p2u31bk, 8gotsepw, gonah5ay, n322vdoi; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 640.7,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-oQ3e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 8k 640.7 21.1 240.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q1ykhvxf",
   "source_url": "https://omlx.ai/benchmarks/performance/q1ykhvxf",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-q1ykhvxf-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 240.8 GB at this context; duplicate traces of the same session: 3u8xporm, 6p2u31bk, 8gotsepw, gonah5ay, n322vdoi; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 645.7,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-oQ3e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 16k 645.7 20.9 240.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q1ykhvxf",
   "source_url": "https://omlx.ai/benchmarks/performance/q1ykhvxf",
   "tg_tokens": null,
   "tps": 20.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-q1ykhvxf-ctx32768",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 240.8 GB at this context; duplicate traces of the same session: 3u8xporm, 6p2u31bk, 8gotsepw, gonah5ay, n322vdoi; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 655.2,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-oQ3e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 32k 655.2 20.8 240.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q1ykhvxf",
   "source_url": "https://omlx.ai/benchmarks/performance/q1ykhvxf",
   "tg_tokens": null,
   "tps": 20.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-q1ykhvxf-ctx65536",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 240.8 GB at this context; duplicate traces of the same session: 3u8xporm, 6p2u31bk, 8gotsepw, gonah5ay, n322vdoi; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 646.3,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-oQ3e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 64k 646.3 20.6 240.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q1ykhvxf",
   "source_url": "https://omlx.ai/benchmarks/performance/q1ykhvxf",
   "tg_tokens": null,
   "tps": 20.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-xbo17glt-ctx1024",
   "model": "Huihui-Ornith-1.5-35B-A3B-abliterated-VL",
   "notes": "VL; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: 2bl4v010, 2opsq8ss; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1625.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Huihui-Ornith-1.5-35B-A3B-abliterated-VL-oQ5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 5bit; 1k 1,625 64.7 26.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xbo17glt",
   "source_url": "https://omlx.ai/benchmarks/performance/xbo17glt",
   "tg_tokens": null,
   "tps": 64.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-xbo17glt-ctx4096",
   "model": "Huihui-Ornith-1.5-35B-A3B-abliterated-VL",
   "notes": "VL; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: 2bl4v010, 2opsq8ss; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1817.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Huihui-Ornith-1.5-35B-A3B-abliterated-VL-oQ5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 5bit; 4k 1,817 110.4 28.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xbo17glt",
   "source_url": "https://omlx.ai/benchmarks/performance/xbo17glt",
   "tg_tokens": null,
   "tps": 110.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-xbo17glt-ctx8192",
   "model": "Huihui-Ornith-1.5-35B-A3B-abliterated-VL",
   "notes": "VL; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.2 GB at this context; duplicate traces of the same session: 2bl4v010, 2opsq8ss; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1838.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Huihui-Ornith-1.5-35B-A3B-abliterated-VL-oQ5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 5bit; 8k 1,838 118.7 28.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xbo17glt",
   "source_url": "https://omlx.ai/benchmarks/performance/xbo17glt",
   "tg_tokens": null,
   "tps": 118.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-xbo17glt-ctx16384",
   "model": "Huihui-Ornith-1.5-35B-A3B-abliterated-VL",
   "notes": "VL; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: 2bl4v010, 2opsq8ss; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1685.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Huihui-Ornith-1.5-35B-A3B-abliterated-VL-oQ5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 5bit; 16k 1,685 99.3 28.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xbo17glt",
   "source_url": "https://omlx.ai/benchmarks/performance/xbo17glt",
   "tg_tokens": null,
   "tps": 99.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-xbo17glt-ctx32768",
   "model": "Huihui-Ornith-1.5-35B-A3B-abliterated-VL",
   "notes": "VL; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.3 GB at this context; duplicate traces of the same session: 2bl4v010, 2opsq8ss; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1378.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Huihui-Ornith-1.5-35B-A3B-abliterated-VL-oQ5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 5bit; 32k 1,378 72.6 29.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xbo17glt",
   "source_url": "https://omlx.ai/benchmarks/performance/xbo17glt",
   "tg_tokens": null,
   "tps": 72.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-xg6ys8qr-ctx1024",
   "model": "Nail-Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: 28bqx9iv, heme5i01, j24a4a09, jnr0xdir, ln2gwxyy, ptswmazp, zfdov9oz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4520.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,520 110.0 20.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xg6ys8qr",
   "source_url": "https://omlx.ai/benchmarks/performance/xg6ys8qr",
   "tg_tokens": null,
   "tps": 110.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-xg6ys8qr-ctx4096",
   "model": "Nail-Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 28bqx9iv, heme5i01, j24a4a09, jnr0xdir, ln2gwxyy, ptswmazp, zfdov9oz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5618.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 5,618 103.5 21.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xg6ys8qr",
   "source_url": "https://omlx.ai/benchmarks/performance/xg6ys8qr",
   "tg_tokens": null,
   "tps": 103.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-xg6ys8qr-ctx8192",
   "model": "Nail-Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 28bqx9iv, heme5i01, j24a4a09, jnr0xdir, ln2gwxyy, ptswmazp, zfdov9oz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5790.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 5,790 99.2 21.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xg6ys8qr",
   "source_url": "https://omlx.ai/benchmarks/performance/xg6ys8qr",
   "tg_tokens": null,
   "tps": 99.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-xg6ys8qr-ctx16384",
   "model": "Nail-Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 28bqx9iv, heme5i01, j24a4a09, jnr0xdir, ln2gwxyy, ptswmazp, zfdov9oz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5645.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 5,645 91.1 21.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xg6ys8qr",
   "source_url": "https://omlx.ai/benchmarks/performance/xg6ys8qr",
   "tg_tokens": null,
   "tps": 91.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-xg6ys8qr-ctx32768",
   "model": "Nail-Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: 28bqx9iv, heme5i01, j24a4a09, jnr0xdir, ln2gwxyy, ptswmazp, zfdov9oz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4678.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 4,678 79.4 21.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xg6ys8qr",
   "source_url": "https://omlx.ai/benchmarks/performance/xg6ys8qr",
   "tg_tokens": null,
   "tps": 79.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-xg6ys8qr-ctx65536",
   "model": "Nail-Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 28bqx9iv, heme5i01, j24a4a09, jnr0xdir, ln2gwxyy, ptswmazp, zfdov9oz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3447.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 3,447 58.7 22.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xg6ys8qr",
   "source_url": "https://omlx.ai/benchmarks/performance/xg6ys8qr",
   "tg_tokens": null,
   "tps": 58.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-xg6ys8qr-ctx131072",
   "model": "Nail-Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 28bqx9iv, heme5i01, j24a4a09, jnr0xdir, ln2gwxyy, ptswmazp, zfdov9oz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2287.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,287 39.2 23.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xg6ys8qr",
   "source_url": "https://omlx.ai/benchmarks/performance/xg6ys8qr",
   "tg_tokens": null,
   "tps": 39.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-xg6ys8qr-ctx199680",
   "model": "Nail-Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 25.7 GB at this context; duplicate traces of the same session: 28bqx9iv, heme5i01, j24a4a09, jnr0xdir, ln2gwxyy, ptswmazp, zfdov9oz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1747.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 1,747 29.7 25.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xg6ys8qr",
   "source_url": "https://omlx.ai/benchmarks/performance/xg6ys8qr",
   "tg_tokens": null,
   "tps": 29.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-0gts9wzg-ctx16384",
   "model": "OmniCoder-9B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 15.2 GB; TTFT 105543 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 155.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "OmniCoder-9B on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 8bit; 16k tokens 155.2 PP tok/s 17.1 TG tok/s 105543 TTFT (ms) 15.2 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0gts9wzg",
   "source_url": "https://omlx.ai/benchmarks/performance/0gts9wzg",
   "tg_tokens": null,
   "tps": 17.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5pdfvz20-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.8 GB at this context; duplicate traces of the same session: 11tpv4e7, 17xssj75; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4361.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 4,361 101.4 28.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5pdfvz20",
   "source_url": "https://omlx.ai/benchmarks/performance/5pdfvz20",
   "tg_tokens": null,
   "tps": 101.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5pdfvz20-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 29.3 GB at this context; duplicate traces of the same session: 11tpv4e7, 17xssj75; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3431.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 64k 3,431 81.2 29.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5pdfvz20",
   "source_url": "https://omlx.ai/benchmarks/performance/5pdfvz20",
   "tg_tokens": null,
   "tps": 81.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5pdfvz20-ctx131072",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 30.6 GB at this context; duplicate traces of the same session: 11tpv4e7, 17xssj75; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2362.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 128k 2,362 68.7 30.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5pdfvz20",
   "source_url": "https://omlx.ai/benchmarks/performance/5pdfvz20",
   "tg_tokens": null,
   "tps": 68.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0gi5lttr-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: d3r3ecnn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 735.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 735.7 102.0 20.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0gi5lttr",
   "source_url": "https://omlx.ai/benchmarks/performance/0gi5lttr",
   "tg_tokens": null,
   "tps": 102.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0gi5lttr-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: d3r3ecnn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2162.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,162 104.0 21.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0gi5lttr",
   "source_url": "https://omlx.ai/benchmarks/performance/0gi5lttr",
   "tg_tokens": null,
   "tps": 104.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9jzdvphz-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "UD build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 4dx76vvu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2100.0,
   "provenance": "community",
   "quant": "Q4_K_XL",
   "quote": "Qwen3.6-35B-A3B-UD-Q4_K_XL on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 2,100 91.8 22.3 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9jzdvphz",
   "source_url": "https://omlx.ai/benchmarks/performance/9jzdvphz",
   "tg_tokens": null,
   "tps": 91.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9jzdvphz-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "UD build; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: 4dx76vvu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4402.0,
   "provenance": "community",
   "quant": "Q4_K_XL",
   "quote": "Qwen3.6-35B-A3B-UD-Q4_K_XL on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 4,402 84.7 22.5 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9jzdvphz",
   "source_url": "https://omlx.ai/benchmarks/performance/9jzdvphz",
   "tg_tokens": null,
   "tps": 84.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-deggenu9-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "UD build; oMLX community benchmark trace, single run at 32k context; peak memory 22.8 GB; TTFT 8765 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3739.0,
   "provenance": "community",
   "quant": "Q4_K_XL",
   "quote": "Qwen3.6-35B-A3B-UD-Q4_K_XL on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 32k tokens 3,739 PP tok/s 74.6 TG tok/s 8765 TTFT (ms) 22.8 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/deggenu9",
   "source_url": "https://omlx.ai/benchmarks/performance/deggenu9",
   "tg_tokens": null,
   "tps": 74.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-6gkehi7p-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 37y6m2ld, fu3s1ifq, h80rl3sz, k81no2gn; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1291.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 1,291 61.4 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6gkehi7p",
   "source_url": "https://omlx.ai/benchmarks/performance/6gkehi7p",
   "tg_tokens": null,
   "tps": 61.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-6gkehi7p-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: 37y6m2ld, fu3s1ifq, h80rl3sz, k81no2gn; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1470.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 1,470 95.8 23.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6gkehi7p",
   "source_url": "https://omlx.ai/benchmarks/performance/6gkehi7p",
   "tg_tokens": null,
   "tps": 95.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-6gkehi7p-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: 37y6m2ld, fu3s1ifq, h80rl3sz, k81no2gn; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1494.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 8k 1,494 93.1 24.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6gkehi7p",
   "source_url": "https://omlx.ai/benchmarks/performance/6gkehi7p",
   "tg_tokens": null,
   "tps": 93.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-6gkehi7p-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 37y6m2ld, fu3s1ifq, h80rl3sz, k81no2gn; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1271.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 16k 1,271 98.7 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6gkehi7p",
   "source_url": "https://omlx.ai/benchmarks/performance/6gkehi7p",
   "tg_tokens": null,
   "tps": 98.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-6gkehi7p-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 37y6m2ld, fu3s1ifq, h80rl3sz, k81no2gn; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1013.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 32k 1,013 67.2 23.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6gkehi7p",
   "source_url": "https://omlx.ai/benchmarks/performance/6gkehi7p",
   "tg_tokens": null,
   "tps": 67.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ageehp0b-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: eie87f8y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 929.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 929.0 33.8 16.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ageehp0b",
   "source_url": "https://omlx.ai/benchmarks/performance/ageehp0b",
   "tg_tokens": null,
   "tps": 33.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ageehp0b-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: eie87f8y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 841.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 841.5 33.0 17.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ageehp0b",
   "source_url": "https://omlx.ai/benchmarks/performance/ageehp0b",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-scdxm616-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: oj77rvoe; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 460.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 460.1 17.7 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/scdxm616",
   "source_url": "https://omlx.ai/benchmarks/performance/scdxm616",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-scdxm616-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: oj77rvoe; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 489.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 489.1 17.7 17.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/scdxm616",
   "source_url": "https://omlx.ai/benchmarks/performance/scdxm616",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-j6pcnsgb-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: wknxk9us; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 459.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 459.2 9.9 29.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j6pcnsgb",
   "source_url": "https://omlx.ai/benchmarks/performance/j6pcnsgb",
   "tg_tokens": null,
   "tps": 9.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-j6pcnsgb-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: wknxk9us; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 376.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 16k 376.1 9.4 30.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j6pcnsgb",
   "source_url": "https://omlx.ai/benchmarks/performance/j6pcnsgb",
   "tg_tokens": null,
   "tps": 9.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-025w35ul-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: 9qcx19x3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 140.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 140.6 5.2 29.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/025w35ul",
   "source_url": "https://omlx.ai/benchmarks/performance/025w35ul",
   "tg_tokens": null,
   "tps": 5.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-025w35ul-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: 9qcx19x3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 136.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 16k 136.5 4.9 30.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/025w35ul",
   "source_url": "https://omlx.ai/benchmarks/performance/025w35ul",
   "tg_tokens": null,
   "tps": 4.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pp6m89zj-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: yexl7afc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 861.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 861.9 32.0 16.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pp6m89zj",
   "source_url": "https://omlx.ai/benchmarks/performance/pp6m89zj",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pp6m89zj-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: yexl7afc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 889.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 889.3 31.6 17.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pp6m89zj",
   "source_url": "https://omlx.ai/benchmarks/performance/pp6m89zj",
   "tg_tokens": null,
   "tps": 31.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-52zytcb2-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 6yat50xf, 7anvz8dh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 406.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 406.9 27.9 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/52zytcb2",
   "source_url": "https://omlx.ai/benchmarks/performance/52zytcb2",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-52zytcb2-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: 6yat50xf, 7anvz8dh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1067.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,067 67.0 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/52zytcb2",
   "source_url": "https://omlx.ai/benchmarks/performance/52zytcb2",
   "tg_tokens": null,
   "tps": 67.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-52zytcb2-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 6yat50xf, 7anvz8dh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1701.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,701 55.0 71.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/52zytcb2",
   "source_url": "https://omlx.ai/benchmarks/performance/52zytcb2",
   "tg_tokens": null,
   "tps": 55.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-52zytcb2-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: 6yat50xf, 7anvz8dh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2596.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,596 22.1 73.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/52zytcb2",
   "source_url": "https://omlx.ai/benchmarks/performance/52zytcb2",
   "tg_tokens": null,
   "tps": 22.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-52zytcb2-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.8 GB at this context; duplicate traces of the same session: 6yat50xf, 7anvz8dh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2448.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,448 71.8 74.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/52zytcb2",
   "source_url": "https://omlx.ai/benchmarks/performance/52zytcb2",
   "tg_tokens": null,
   "tps": 71.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-52zytcb2-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.4 GB at this context; duplicate traces of the same session: 6yat50xf, 7anvz8dh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2402.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,402 62.8 79.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/52zytcb2",
   "source_url": "https://omlx.ai/benchmarks/performance/52zytcb2",
   "tg_tokens": null,
   "tps": 62.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-52zytcb2-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.9 GB at this context; duplicate traces of the same session: 6yat50xf, 7anvz8dh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2413.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,413 63.9 80.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/52zytcb2",
   "source_url": "https://omlx.ai/benchmarks/performance/52zytcb2",
   "tg_tokens": null,
   "tps": 63.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6dqav5n6-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 39tpnoc7, 4d39swa1, lelyylcm, x5ffulyx, y0e5p170; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1083.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,083 67.7 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6dqav5n6",
   "source_url": "https://omlx.ai/benchmarks/performance/6dqav5n6",
   "tg_tokens": null,
   "tps": 67.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6dqav5n6-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: 39tpnoc7, 4d39swa1, lelyylcm, x5ffulyx, y0e5p170; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1914.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,914 57.2 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6dqav5n6",
   "source_url": "https://omlx.ai/benchmarks/performance/6dqav5n6",
   "tg_tokens": null,
   "tps": 57.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6dqav5n6-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 39tpnoc7, 4d39swa1, lelyylcm, x5ffulyx, y0e5p170; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2420.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,420 59.5 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6dqav5n6",
   "source_url": "https://omlx.ai/benchmarks/performance/6dqav5n6",
   "tg_tokens": null,
   "tps": 59.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6dqav5n6-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: 39tpnoc7, 4d39swa1, lelyylcm, x5ffulyx, y0e5p170; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2632.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,632 80.8 72.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6dqav5n6",
   "source_url": "https://omlx.ai/benchmarks/performance/6dqav5n6",
   "tg_tokens": null,
   "tps": 80.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6dqav5n6-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 75.3 GB at this context; duplicate traces of the same session: 39tpnoc7, 4d39swa1, lelyylcm, x5ffulyx, y0e5p170; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2404.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,404 78.0 75.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6dqav5n6",
   "source_url": "https://omlx.ai/benchmarks/performance/6dqav5n6",
   "tg_tokens": null,
   "tps": 78.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6dqav5n6-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.1 GB at this context; duplicate traces of the same session: 39tpnoc7, 4d39swa1, lelyylcm, x5ffulyx, y0e5p170; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2336.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,336 67.7 79.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6dqav5n6",
   "source_url": "https://omlx.ai/benchmarks/performance/6dqav5n6",
   "tg_tokens": null,
   "tps": 67.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6dqav5n6-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.9 GB at this context; duplicate traces of the same session: 39tpnoc7, 4d39swa1, lelyylcm, x5ffulyx, y0e5p170; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2377.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,377 60.3 80.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6dqav5n6",
   "source_url": "https://omlx.ai/benchmarks/performance/6dqav5n6",
   "tg_tokens": null,
   "tps": 60.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-4p9zclri-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.2 GB at this context; duplicate traces of the same session: 3qrlfkz5, o335stz0, wq6ep8c6; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 655.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 1k 655.3 34.8 183.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4p9zclri",
   "source_url": "https://omlx.ai/benchmarks/performance/4p9zclri",
   "tg_tokens": null,
   "tps": 34.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-4p9zclri-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.7 GB at this context; duplicate traces of the same session: 3qrlfkz5, o335stz0, wq6ep8c6; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1523.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 4k 1,523 96.4 183.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4p9zclri",
   "source_url": "https://omlx.ai/benchmarks/performance/4p9zclri",
   "tg_tokens": null,
   "tps": 96.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-4p9zclri-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 184.7 GB at this context; duplicate traces of the same session: 3qrlfkz5, o335stz0, wq6ep8c6; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2737.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 8k 2,737 113.4 184.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4p9zclri",
   "source_url": "https://omlx.ai/benchmarks/performance/4p9zclri",
   "tg_tokens": null,
   "tps": 113.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-4p9zclri-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 187.0 GB at this context; duplicate traces of the same session: 3qrlfkz5, o335stz0, wq6ep8c6; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4418.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 64k 4,418 69.0 187.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4p9zclri",
   "source_url": "https://omlx.ai/benchmarks/performance/4p9zclri",
   "tg_tokens": null,
   "tps": 69.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-qyueo7rr-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 15.1 GB at this context; duplicate traces of the same session: nkh2i1yv; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 293.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 16k 293.7 28.5 15.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qyueo7rr",
   "source_url": "https://omlx.ai/benchmarks/performance/qyueo7rr",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-qyueo7rr-ctx32768",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: nkh2i1yv; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 206.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 32k 206.9 19.4 15.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qyueo7rr",
   "source_url": "https://omlx.ai/benchmarks/performance/qyueo7rr",
   "tg_tokens": null,
   "tps": 19.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-hapwjeny-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 15.1 GB at this context; duplicate traces of the same session: qu34toqp; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 238.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 16k 238.1 24.3 15.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hapwjeny",
   "source_url": "https://omlx.ai/benchmarks/performance/hapwjeny",
   "tg_tokens": null,
   "tps": 24.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-hapwjeny-ctx32768",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: qu34toqp; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 194.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 32k 194.7 19.2 15.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hapwjeny",
   "source_url": "https://omlx.ai/benchmarks/performance/hapwjeny",
   "tg_tokens": null,
   "tps": 19.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-frjkf8rq-ctx16384",
   "model": "gemma-4-26B-A4B-it-qat",
   "notes": "q4_0 mlx-aligned build; oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: 8ccn4p7y; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 248.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-q4_0-mlx-aligned on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 16k 248.8 25.1 16.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/frjkf8rq",
   "source_url": "https://omlx.ai/benchmarks/performance/frjkf8rq",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-frjkf8rq-ctx32768",
   "model": "gemma-4-26B-A4B-it-qat",
   "notes": "q4_0 mlx-aligned build; oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: 8ccn4p7y; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-q4_0-mlx-aligned on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 32k 205.9 18.5 17.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/frjkf8rq",
   "source_url": "https://omlx.ai/benchmarks/performance/frjkf8rq",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2417kfvp-ctx1024",
   "model": "Qwen3.6-27B",
   "notes": "Dagger build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 53.7 GB at this context; duplicate traces of the same session: 216i9o02, b6chnrx3, ibrg5x5w, lo2a84ne, n6qraxjh, oi5w1axd, rv70p064; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 616.1,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Dagger-Qwen3.6-27B-MLX-MTP-BF16 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 bf16; 1k 616.1 17.9 53.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2417kfvp",
   "source_url": "https://omlx.ai/benchmarks/performance/2417kfvp",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2417kfvp-ctx4096",
   "model": "Qwen3.6-27B",
   "notes": "Dagger build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 55.0 GB at this context; duplicate traces of the same session: 216i9o02, b6chnrx3, ibrg5x5w, lo2a84ne, n6qraxjh, oi5w1axd, rv70p064; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 923.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Dagger-Qwen3.6-27B-MLX-MTP-BF16 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 bf16; 4k 923.0 20.0 55.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2417kfvp",
   "source_url": "https://omlx.ai/benchmarks/performance/2417kfvp",
   "tg_tokens": null,
   "tps": 20.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2417kfvp-ctx8192",
   "model": "Qwen3.6-27B",
   "notes": "Dagger build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 55.3 GB at this context; duplicate traces of the same session: 216i9o02, b6chnrx3, ibrg5x5w, lo2a84ne, n6qraxjh, oi5w1axd, rv70p064; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 988.2,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Dagger-Qwen3.6-27B-MLX-MTP-BF16 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 bf16; 8k 988.2 19.9 55.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2417kfvp",
   "source_url": "https://omlx.ai/benchmarks/performance/2417kfvp",
   "tg_tokens": null,
   "tps": 19.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2417kfvp-ctx16384",
   "model": "Qwen3.6-27B",
   "notes": "Dagger build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 55.9 GB at this context; duplicate traces of the same session: 216i9o02, b6chnrx3, ibrg5x5w, lo2a84ne, n6qraxjh, oi5w1axd, rv70p064; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 825.4,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Dagger-Qwen3.6-27B-MLX-MTP-BF16 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 bf16; 16k 825.4 18.6 55.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2417kfvp",
   "source_url": "https://omlx.ai/benchmarks/performance/2417kfvp",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2417kfvp-ctx32768",
   "model": "Qwen3.6-27B",
   "notes": "Dagger build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 57.0 GB at this context; duplicate traces of the same session: 216i9o02, b6chnrx3, ibrg5x5w, lo2a84ne, n6qraxjh, oi5w1axd, rv70p064; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 790.1,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Dagger-Qwen3.6-27B-MLX-MTP-BF16 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 bf16; 32k 790.1 21.7 57.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2417kfvp",
   "source_url": "https://omlx.ai/benchmarks/performance/2417kfvp",
   "tg_tokens": null,
   "tps": 21.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2417kfvp-ctx65536",
   "model": "Qwen3.6-27B",
   "notes": "Dagger build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 59.3 GB at this context; duplicate traces of the same session: 216i9o02, b6chnrx3, ibrg5x5w, lo2a84ne, n6qraxjh, oi5w1axd, rv70p064; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 700.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Dagger-Qwen3.6-27B-MLX-MTP-BF16 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 bf16; 64k 700.0 15.2 59.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2417kfvp",
   "source_url": "https://omlx.ai/benchmarks/performance/2417kfvp",
   "tg_tokens": null,
   "tps": 15.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2417kfvp-ctx131072",
   "model": "Qwen3.6-27B",
   "notes": "Dagger build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 64.1 GB at this context; duplicate traces of the same session: 216i9o02, b6chnrx3, ibrg5x5w, lo2a84ne, n6qraxjh, oi5w1axd, rv70p064; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 576.9,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Dagger-Qwen3.6-27B-MLX-MTP-BF16 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 bf16; 128k 576.9 12.3 64.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2417kfvp",
   "source_url": "https://omlx.ai/benchmarks/performance/2417kfvp",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2417kfvp-ctx199680",
   "model": "Qwen3.6-27B",
   "notes": "Dagger build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 216i9o02, b6chnrx3, ibrg5x5w, lo2a84ne, n6qraxjh, oi5w1axd, rv70p064; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 479.6,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Dagger-Qwen3.6-27B-MLX-MTP-BF16 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 bf16; 195k 479.6 10.3 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2417kfvp",
   "source_url": "https://omlx.ai/benchmarks/performance/2417kfvp",
   "tg_tokens": null,
   "tps": 10.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-pdy6kvpr-ctx1024",
   "model": "Huihui-Ornith-1.5-35B-A3B-abliterated-VL",
   "notes": "VL; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: ws6liq4v; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1625.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Huihui-Ornith-1.5-35B-A3B-abliterated-VL-oQ5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 5bit; 1k 1,625 64.7 26.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pdy6kvpr",
   "source_url": "https://omlx.ai/benchmarks/performance/pdy6kvpr",
   "tg_tokens": null,
   "tps": 64.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-pdy6kvpr-ctx4096",
   "model": "Huihui-Ornith-1.5-35B-A3B-abliterated-VL",
   "notes": "VL; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: ws6liq4v; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1817.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Huihui-Ornith-1.5-35B-A3B-abliterated-VL-oQ5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 5bit; 4k 1,817 110.4 28.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pdy6kvpr",
   "source_url": "https://omlx.ai/benchmarks/performance/pdy6kvpr",
   "tg_tokens": null,
   "tps": 110.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-pdy6kvpr-ctx8192",
   "model": "Huihui-Ornith-1.5-35B-A3B-abliterated-VL",
   "notes": "VL; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.2 GB at this context; duplicate traces of the same session: ws6liq4v; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1838.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Huihui-Ornith-1.5-35B-A3B-abliterated-VL-oQ5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 5bit; 8k 1,838 118.7 28.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pdy6kvpr",
   "source_url": "https://omlx.ai/benchmarks/performance/pdy6kvpr",
   "tg_tokens": null,
   "tps": 118.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-pdy6kvpr-ctx16384",
   "model": "Huihui-Ornith-1.5-35B-A3B-abliterated-VL",
   "notes": "VL; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: ws6liq4v; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1685.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Huihui-Ornith-1.5-35B-A3B-abliterated-VL-oQ5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 5bit; 16k 1,685 99.3 28.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pdy6kvpr",
   "source_url": "https://omlx.ai/benchmarks/performance/pdy6kvpr",
   "tg_tokens": null,
   "tps": 99.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-pdy6kvpr-ctx32768",
   "model": "Huihui-Ornith-1.5-35B-A3B-abliterated-VL",
   "notes": "VL; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.3 GB at this context; duplicate traces of the same session: ws6liq4v; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1378.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Huihui-Ornith-1.5-35B-A3B-abliterated-VL-oQ5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 5bit; 32k 1,378 72.6 29.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pdy6kvpr",
   "source_url": "https://omlx.ai/benchmarks/performance/pdy6kvpr",
   "tg_tokens": null,
   "tps": 72.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-9jnrh9v5-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: 98mxe8ej; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 478.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 478.1 15.3 6.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9jnrh9v5",
   "source_url": "https://omlx.ai/benchmarks/performance/9jnrh9v5",
   "tg_tokens": null,
   "tps": 15.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-9jnrh9v5-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.1 GB at this context; duplicate traces of the same session: 98mxe8ej; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 462.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 16k 462.2 13.2 7.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9jnrh9v5",
   "source_url": "https://omlx.ai/benchmarks/performance/9jnrh9v5",
   "tg_tokens": null,
   "tps": 13.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bz3odzck-ctx1024",
   "model": "Qwen3.5-9B-Uncensored-HauhauCS-Aggressive",
   "notes": "oMLX community benchmark trace; peak memory 10.4 GB at this context; duplicate traces of the same session: 1d2y4jkt, 5bxhi580, 5opwnohu, ga5cxhsb, qo7t0cyb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1090.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B-Uncensored-HauhauCS-Aggressive-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 1,090 67.8 10.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bz3odzck",
   "source_url": "https://omlx.ai/benchmarks/performance/bz3odzck",
   "tg_tokens": null,
   "tps": 67.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bz3odzck-ctx4096",
   "model": "Qwen3.5-9B-Uncensored-HauhauCS-Aggressive",
   "notes": "oMLX community benchmark trace; peak memory 11.9 GB at this context; duplicate traces of the same session: 1d2y4jkt, 5bxhi580, 5opwnohu, ga5cxhsb, qo7t0cyb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1204.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B-Uncensored-HauhauCS-Aggressive-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 1,204 66.9 11.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bz3odzck",
   "source_url": "https://omlx.ai/benchmarks/performance/bz3odzck",
   "tg_tokens": null,
   "tps": 66.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bz3odzck-ctx8192",
   "model": "Qwen3.5-9B-Uncensored-HauhauCS-Aggressive",
   "notes": "oMLX community benchmark trace; peak memory 12.0 GB at this context; duplicate traces of the same session: 1d2y4jkt, 5bxhi580, 5opwnohu, ga5cxhsb, qo7t0cyb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1186.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B-Uncensored-HauhauCS-Aggressive-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 8k 1,186 66.0 12.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bz3odzck",
   "source_url": "https://omlx.ai/benchmarks/performance/bz3odzck",
   "tg_tokens": null,
   "tps": 66.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bz3odzck-ctx16384",
   "model": "Qwen3.5-9B-Uncensored-HauhauCS-Aggressive",
   "notes": "oMLX community benchmark trace; peak memory 12.4 GB at this context; duplicate traces of the same session: 1d2y4jkt, 5bxhi580, 5opwnohu, ga5cxhsb, qo7t0cyb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1143.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B-Uncensored-HauhauCS-Aggressive-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 16k 1,143 63.7 12.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bz3odzck",
   "source_url": "https://omlx.ai/benchmarks/performance/bz3odzck",
   "tg_tokens": null,
   "tps": 63.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bz3odzck-ctx32768",
   "model": "Qwen3.5-9B-Uncensored-HauhauCS-Aggressive",
   "notes": "oMLX community benchmark trace; peak memory 13.2 GB at this context; duplicate traces of the same session: 1d2y4jkt, 5bxhi580, 5opwnohu, ga5cxhsb, qo7t0cyb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1057.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B-Uncensored-HauhauCS-Aggressive-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 32k 1,057 59.9 13.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bz3odzck",
   "source_url": "https://omlx.ai/benchmarks/performance/bz3odzck",
   "tg_tokens": null,
   "tps": 59.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bz3odzck-ctx65536",
   "model": "Qwen3.5-9B-Uncensored-HauhauCS-Aggressive",
   "notes": "oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: 1d2y4jkt, 5bxhi580, 5opwnohu, ga5cxhsb, qo7t0cyb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 913.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B-Uncensored-HauhauCS-Aggressive-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 64k 913.5 52.9 14.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bz3odzck",
   "source_url": "https://omlx.ai/benchmarks/performance/bz3odzck",
   "tg_tokens": null,
   "tps": 52.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dktij2wf-ctx199680",
   "model": "Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive",
   "notes": "OptiQ MLX build; oMLX community benchmark trace, single run at 195k context; peak memory 24.3 GB; TTFT 120243 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1663.0,
   "provenance": "community",
   "quant": "6bpw",
   "quote": "Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-OptiQ-6bpw-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k tokens 1,663 PP tok/s 58.5 TG tok/s 120243 TTFT (ms) 24.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dktij2wf",
   "source_url": "https://omlx.ai/benchmarks/performance/dktij2wf",
   "tg_tokens": null,
   "tps": 58.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0un8ed3e-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive",
   "notes": "OptiQ MLX build; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 5kvdmy00; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 960.2,
   "provenance": "community",
   "quant": "6bpw",
   "quote": "Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-OptiQ-6bpw-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 960.2 97.0 19.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0un8ed3e",
   "source_url": "https://omlx.ai/benchmarks/performance/0un8ed3e",
   "tg_tokens": null,
   "tps": 97.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0un8ed3e-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive",
   "notes": "OptiQ MLX build; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: 5kvdmy00; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2238.0,
   "provenance": "community",
   "quant": "6bpw",
   "quote": "Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-OptiQ-6bpw-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,238 95.4 20.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0un8ed3e",
   "source_url": "https://omlx.ai/benchmarks/performance/0un8ed3e",
   "tg_tokens": null,
   "tps": 95.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1m46kqtr-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: f90029u8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4471.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,471 113.0 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1m46kqtr",
   "source_url": "https://omlx.ai/benchmarks/performance/1m46kqtr",
   "tg_tokens": null,
   "tps": 113.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1m46kqtr-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: f90029u8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6116.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,116 109.7 20.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1m46kqtr",
   "source_url": "https://omlx.ai/benchmarks/performance/1m46kqtr",
   "tg_tokens": null,
   "tps": 109.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yn77y6rn-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: hm9qac15; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4063.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,063 110.9 19.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yn77y6rn",
   "source_url": "https://omlx.ai/benchmarks/performance/yn77y6rn",
   "tg_tokens": null,
   "tps": 110.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yn77y6rn-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: hm9qac15; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5775.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 5,775 108.6 20.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yn77y6rn",
   "source_url": "https://omlx.ai/benchmarks/performance/yn77y6rn",
   "tg_tokens": null,
   "tps": 108.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ovg4netg-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: 9gy0epvn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 489.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 489.5 17.2 17.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ovg4netg",
   "source_url": "https://omlx.ai/benchmarks/performance/ovg4netg",
   "tg_tokens": null,
   "tps": 17.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ovg4netg-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 9gy0epvn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 446.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 446.2 16.7 17.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ovg4netg",
   "source_url": "https://omlx.ai/benchmarks/performance/ovg4netg",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-imnx23w5-ctx1024",
   "model": "Qwen3.8-27B-VL",
   "notes": "VL; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.3 GB at this context; duplicate traces of the same session: 22idmi5t, kkmiapdv, s6afoy47, zdrz0cfw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 246.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-VL-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 246.4 39.7 33.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/imnx23w5",
   "source_url": "https://omlx.ai/benchmarks/performance/imnx23w5",
   "tg_tokens": null,
   "tps": 39.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-imnx23w5-ctx4096",
   "model": "Qwen3.8-27B-VL",
   "notes": "VL; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.6 GB at this context; duplicate traces of the same session: 22idmi5t, kkmiapdv, s6afoy47, zdrz0cfw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-VL-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 323.2 40.4 36.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/imnx23w5",
   "source_url": "https://omlx.ai/benchmarks/performance/imnx23w5",
   "tg_tokens": null,
   "tps": 40.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-imnx23w5-ctx8192",
   "model": "Qwen3.8-27B-VL",
   "notes": "VL; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.1 GB at this context; duplicate traces of the same session: 22idmi5t, kkmiapdv, s6afoy47, zdrz0cfw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 325.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-VL-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 8k 325.3 40.0 37.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/imnx23w5",
   "source_url": "https://omlx.ai/benchmarks/performance/imnx23w5",
   "tg_tokens": null,
   "tps": 40.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-imnx23w5-ctx16384",
   "model": "Qwen3.8-27B-VL",
   "notes": "VL; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.7 GB at this context; duplicate traces of the same session: 22idmi5t, kkmiapdv, s6afoy47, zdrz0cfw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 245.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-VL-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k 245.7 32.0 31.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/imnx23w5",
   "source_url": "https://omlx.ai/benchmarks/performance/imnx23w5",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-imnx23w5-ctx32768",
   "model": "Qwen3.8-27B-VL",
   "notes": "VL; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.5 GB at this context; duplicate traces of the same session: 22idmi5t, kkmiapdv, s6afoy47, zdrz0cfw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 232.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-VL-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 232.0 31.2 33.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/imnx23w5",
   "source_url": "https://omlx.ai/benchmarks/performance/imnx23w5",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-g7qdpotl-ctx1024",
   "model": "Qwen3.8-27B-VL",
   "notes": "VL; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.3 GB at this context; duplicate traces of the same session: 1l0d943i, e1y2l54l, ff5inmg9, tw345726; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 244.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-VL-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 244.4 42.5 33.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g7qdpotl",
   "source_url": "https://omlx.ai/benchmarks/performance/g7qdpotl",
   "tg_tokens": null,
   "tps": 42.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-g7qdpotl-ctx4096",
   "model": "Qwen3.8-27B-VL",
   "notes": "VL; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.5 GB at this context; duplicate traces of the same session: 1l0d943i, e1y2l54l, ff5inmg9, tw345726; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 319.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-VL-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 319.9 33.6 36.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g7qdpotl",
   "source_url": "https://omlx.ai/benchmarks/performance/g7qdpotl",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-g7qdpotl-ctx8192",
   "model": "Qwen3.8-27B-VL",
   "notes": "VL; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.1 GB at this context; duplicate traces of the same session: 1l0d943i, e1y2l54l, ff5inmg9, tw345726; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-VL-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 8k 323.8 31.7 37.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g7qdpotl",
   "source_url": "https://omlx.ai/benchmarks/performance/g7qdpotl",
   "tg_tokens": null,
   "tps": 31.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-g7qdpotl-ctx16384",
   "model": "Qwen3.8-27B-VL",
   "notes": "VL; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.7 GB at this context; duplicate traces of the same session: 1l0d943i, e1y2l54l, ff5inmg9, tw345726; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 235.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-VL-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k 235.1 35.4 31.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g7qdpotl",
   "source_url": "https://omlx.ai/benchmarks/performance/g7qdpotl",
   "tg_tokens": null,
   "tps": 35.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-g7qdpotl-ctx32768",
   "model": "Qwen3.8-27B-VL",
   "notes": "VL; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.5 GB at this context; duplicate traces of the same session: 1l0d943i, e1y2l54l, ff5inmg9, tw345726; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 231.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-VL-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 231.3 30.4 33.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g7qdpotl",
   "source_url": "https://omlx.ai/benchmarks/performance/g7qdpotl",
   "tg_tokens": null,
   "tps": 30.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-z7ued8vj-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 115.7 GB; TTFT 98441 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 665.7,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ6-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 64k tokens 665.7 PP tok/s 51.2 TG tok/s 98441 TTFT (ms) 115.7 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z7ued8vj",
   "source_url": "https://omlx.ai/benchmarks/performance/z7ued8vj",
   "tg_tokens": null,
   "tps": 51.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-f15jxjm7-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 39.8 GB at this context; duplicate traces of the same session: nx6aufi6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 479.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 479.3 63.5 39.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f15jxjm7",
   "source_url": "https://omlx.ai/benchmarks/performance/f15jxjm7",
   "tg_tokens": null,
   "tps": 63.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-f15jxjm7-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 40.0 GB at this context; duplicate traces of the same session: nx6aufi6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1115.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,115 46.5 40.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f15jxjm7",
   "source_url": "https://omlx.ai/benchmarks/performance/f15jxjm7",
   "tg_tokens": null,
   "tps": 46.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-cdsiw1rc-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: 0rvxm64r, 96ty0en2, qbfc1u6u, vlbn8q15, xjxnt2ek, zb12ufic; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 330.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 330.5 39.0 71.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cdsiw1rc",
   "source_url": "https://omlx.ai/benchmarks/performance/cdsiw1rc",
   "tg_tokens": null,
   "tps": 39.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-cdsiw1rc-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: 0rvxm64r, 96ty0en2, qbfc1u6u, vlbn8q15, xjxnt2ek, zb12ufic; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 528.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 528.8 33.0 71.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cdsiw1rc",
   "source_url": "https://omlx.ai/benchmarks/performance/cdsiw1rc",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Max",
   "id": "omlx-cdsiw1rc-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.0 GB at this context; duplicate traces of the same session: 0rvxm64r, 96ty0en2, qbfc1u6u, vlbn8q15, xjxnt2ek, zb12ufic; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 429.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 429.5 11.8 72.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cdsiw1rc",
   "source_url": "https://omlx.ai/benchmarks/performance/cdsiw1rc",
   "tg_tokens": null,
   "tps": 11.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-cdsiw1rc-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.5 GB at this context; duplicate traces of the same session: 0rvxm64r, 96ty0en2, qbfc1u6u, vlbn8q15, xjxnt2ek, zb12ufic; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 427.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 427.2 31.1 72.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cdsiw1rc",
   "source_url": "https://omlx.ai/benchmarks/performance/cdsiw1rc",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-cdsiw1rc-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.7 GB at this context; duplicate traces of the same session: 0rvxm64r, 96ty0en2, qbfc1u6u, vlbn8q15, xjxnt2ek, zb12ufic; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 303.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 303.4 24.9 73.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cdsiw1rc",
   "source_url": "https://omlx.ai/benchmarks/performance/cdsiw1rc",
   "tg_tokens": null,
   "tps": 24.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-cdsiw1rc-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 76.0 GB at this context; duplicate traces of the same session: 0rvxm64r, 96ty0en2, qbfc1u6u, vlbn8q15, xjxnt2ek, zb12ufic; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 255.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 255.4 23.2 76.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cdsiw1rc",
   "source_url": "https://omlx.ai/benchmarks/performance/cdsiw1rc",
   "tg_tokens": null,
   "tps": 23.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-cdsiw1rc-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.9 GB at this context; duplicate traces of the same session: 0rvxm64r, 96ty0en2, qbfc1u6u, vlbn8q15, xjxnt2ek, zb12ufic; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 273.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 273.8 18.6 78.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cdsiw1rc",
   "source_url": "https://omlx.ai/benchmarks/performance/cdsiw1rc",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ldmylo35-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1083.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,083 67.7 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ldmylo35",
   "source_url": "https://omlx.ai/benchmarks/performance/ldmylo35",
   "tg_tokens": null,
   "tps": 67.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ldmylo35-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1914.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,914 57.2 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ldmylo35",
   "source_url": "https://omlx.ai/benchmarks/performance/ldmylo35",
   "tg_tokens": null,
   "tps": 57.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ldmylo35-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2420.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,420 59.5 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ldmylo35",
   "source_url": "https://omlx.ai/benchmarks/performance/ldmylo35",
   "tg_tokens": null,
   "tps": 59.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ldmylo35-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2632.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,632 80.8 72.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ldmylo35",
   "source_url": "https://omlx.ai/benchmarks/performance/ldmylo35",
   "tg_tokens": null,
   "tps": 80.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ldmylo35-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 75.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2404.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,404 78.0 75.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ldmylo35",
   "source_url": "https://omlx.ai/benchmarks/performance/ldmylo35",
   "tg_tokens": null,
   "tps": 78.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ldmylo35-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2336.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,336 67.7 79.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ldmylo35",
   "source_url": "https://omlx.ai/benchmarks/performance/ldmylo35",
   "tg_tokens": null,
   "tps": 67.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ldmylo35-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2377.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,377 60.3 80.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ldmylo35",
   "source_url": "https://omlx.ai/benchmarks/performance/ldmylo35",
   "tg_tokens": null,
   "tps": 60.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i8ilv0ki-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 86.5 GB at this context; duplicate traces of the same session: is53ewb0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 931.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 931.2 60.5 86.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i8ilv0ki",
   "source_url": "https://omlx.ai/benchmarks/performance/i8ilv0ki",
   "tg_tokens": null,
   "tps": 60.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-i8ilv0ki-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 86.7 GB at this context; duplicate traces of the same session: is53ewb0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1691.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 1,691 96.4 86.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i8ilv0ki",
   "source_url": "https://omlx.ai/benchmarks/performance/i8ilv0ki",
   "tg_tokens": null,
   "tps": 96.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i8ilv0ki-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 87.8 GB at this context; duplicate traces of the same session: is53ewb0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2172.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 2,172 57.6 87.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i8ilv0ki",
   "source_url": "https://omlx.ai/benchmarks/performance/i8ilv0ki",
   "tg_tokens": null,
   "tps": 57.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-i8ilv0ki-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 88.9 GB at this context; duplicate traces of the same session: is53ewb0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2456.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 2,456 74.1 88.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i8ilv0ki",
   "source_url": "https://omlx.ai/benchmarks/performance/i8ilv0ki",
   "tg_tokens": null,
   "tps": 74.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-i8ilv0ki-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 89.5 GB at this context; duplicate traces of the same session: is53ewb0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2184.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 2,184 69.8 89.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i8ilv0ki",
   "source_url": "https://omlx.ai/benchmarks/performance/i8ilv0ki",
   "tg_tokens": null,
   "tps": 69.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-v3pvckku-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 15.1 GB at this context; duplicate traces of the same session: krjfse9q; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 294.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 16k 294.7 28.9 15.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v3pvckku",
   "source_url": "https://omlx.ai/benchmarks/performance/v3pvckku",
   "tg_tokens": null,
   "tps": 28.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-v3pvckku-ctx32768",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: krjfse9q; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 212.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 32k 212.6 19.4 15.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v3pvckku",
   "source_url": "https://omlx.ai/benchmarks/performance/v3pvckku",
   "tg_tokens": null,
   "tps": 19.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-7hnkwhzy-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 15.1 GB at this context; duplicate traces of the same session: lzkqk6zf; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 292.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 16k 292.7 28.5 15.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7hnkwhzy",
   "source_url": "https://omlx.ai/benchmarks/performance/7hnkwhzy",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-7hnkwhzy-ctx32768",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: lzkqk6zf; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 199.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 32k 199.5 18.9 15.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7hnkwhzy",
   "source_url": "https://omlx.ai/benchmarks/performance/7hnkwhzy",
   "tg_tokens": null,
   "tps": 18.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-zo4xnu7s-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 15.1 GB at this context; duplicate traces of the same session: 3hniuzmd; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 294.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 16k 294.7 29.0 15.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zo4xnu7s",
   "source_url": "https://omlx.ai/benchmarks/performance/zo4xnu7s",
   "tg_tokens": null,
   "tps": 29.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-zo4xnu7s-ctx32768",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: 3hniuzmd; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 32k 220.9 19.6 15.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zo4xnu7s",
   "source_url": "https://omlx.ai/benchmarks/performance/zo4xnu7s",
   "tg_tokens": null,
   "tps": 19.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-x8zj8yk1-ctx1024",
   "model": "Gemma-4-31B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: jphcsfwz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31b-it-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 103.8 11.9 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x8zj8yk1",
   "source_url": "https://omlx.ai/benchmarks/performance/x8zj8yk1",
   "tg_tokens": null,
   "tps": 11.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-x8zj8yk1-ctx4096",
   "model": "Gemma-4-31B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: jphcsfwz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 92.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31b-it-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 92.1 10.8 18.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x8zj8yk1",
   "source_url": "https://omlx.ai/benchmarks/performance/x8zj8yk1",
   "tg_tokens": null,
   "tps": 10.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-cqxj7011-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "zp build; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: g1i28id1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 248.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "qwen3.8-27b-zp-oQ6e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 248.5 39.7 23.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cqxj7011",
   "source_url": "https://omlx.ai/benchmarks/performance/cqxj7011",
   "tg_tokens": null,
   "tps": 39.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-cqxj7011-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "zp build; oMLX community benchmark trace; peak memory 28.4 GB at this context; duplicate traces of the same session: g1i28id1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "qwen3.8-27b-zp-oQ6e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 220.6 43.6 28.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cqxj7011",
   "source_url": "https://omlx.ai/benchmarks/performance/cqxj7011",
   "tg_tokens": null,
   "tps": 43.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lwcc79cg-ctx1024",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 06zi33vd, 9yijs1wg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 838.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "AREX-2-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 838.6 38.9 23.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lwcc79cg",
   "source_url": "https://omlx.ai/benchmarks/performance/lwcc79cg",
   "tg_tokens": null,
   "tps": 38.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lwcc79cg-ctx4096",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: 06zi33vd, 9yijs1wg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 860.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "AREX-2-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 860.8 40.0 24.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lwcc79cg",
   "source_url": "https://omlx.ai/benchmarks/performance/lwcc79cg",
   "tg_tokens": null,
   "tps": 40.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lwcc79cg-ctx65536",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: 06zi33vd, 9yijs1wg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 424.2,
   "provenance": "community",
   "quant": "6bit",
   "quote": "AREX-2-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 64k 424.2 12.3 29.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lwcc79cg",
   "source_url": "https://omlx.ai/benchmarks/performance/lwcc79cg",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5esahotf-ctx1024",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: cddbes7l, i4r48ndi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 788.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "AREX-2-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 788.1 23.8 22.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5esahotf",
   "source_url": "https://omlx.ai/benchmarks/performance/5esahotf",
   "tg_tokens": null,
   "tps": 23.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5esahotf-ctx4096",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: cddbes7l, i4r48ndi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 843.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "AREX-2-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 843.3 23.9 23.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5esahotf",
   "source_url": "https://omlx.ai/benchmarks/performance/5esahotf",
   "tg_tokens": null,
   "tps": 23.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5esahotf-ctx65536",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 27.0 GB at this context; duplicate traces of the same session: cddbes7l, i4r48ndi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 385.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "AREX-2-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 64k 385.3 11.9 27.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5esahotf",
   "source_url": "https://omlx.ai/benchmarks/performance/5esahotf",
   "tg_tokens": null,
   "tps": 11.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2",
   "id": "omlx-xi03j708-ctx1024",
   "model": "Qwen3.5-4B-Claude-4.6-Opus-Reasoning-Distilled",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: nrz70n0g; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 251.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MLX-Qwen3.5-4B-Claude-4.6-Opus-Reasoning-Distilled-4bit on M2 (10c); M2 (10c) \u00b7 16 GB \u00b7 4bit; 1k 251.8 32.5 3.4 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xi03j708",
   "source_url": "https://omlx.ai/benchmarks/performance/xi03j708",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2",
   "id": "omlx-xi03j708-ctx4096",
   "model": "Qwen3.5-4B-Claude-4.6-Opus-Reasoning-Distilled",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 3.8 GB at this context; duplicate traces of the same session: nrz70n0g; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 235.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MLX-Qwen3.5-4B-Claude-4.6-Opus-Reasoning-Distilled-4bit on M2 (10c); M2 (10c) \u00b7 16 GB \u00b7 4bit; 4k 235.5 30.1 3.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xi03j708",
   "source_url": "https://omlx.ai/benchmarks/performance/xi03j708",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-w40qv74b-ctx1024",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "text-only build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 6.0 GB at this context; duplicate traces of the same session: sqdychl1; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 232.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-Text-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 232.9 19.9 6.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w40qv74b",
   "source_url": "https://omlx.ai/benchmarks/performance/w40qv74b",
   "tg_tokens": null,
   "tps": 19.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-w40qv74b-ctx4096",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "text-only build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 6.6 GB at this context; duplicate traces of the same session: sqdychl1; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 233.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-Text-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 233.7 19.5 6.6 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w40qv74b",
   "source_url": "https://omlx.ai/benchmarks/performance/w40qv74b",
   "tg_tokens": null,
   "tps": 19.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-11m6vz20-ctx1024",
   "model": "Qwen3.5-35B-A3B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 11.4 GB at this context; duplicate traces of the same session: 9vcek97t; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 467.8,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Qwen3.5-35B-A3B-MLX-2bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 2bit; 1k 467.8 72.3 11.4 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/11m6vz20",
   "source_url": "https://omlx.ai/benchmarks/performance/11m6vz20",
   "tg_tokens": null,
   "tps": 72.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-11m6vz20-ctx4096",
   "model": "Qwen3.5-35B-A3B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 11.9 GB at this context; duplicate traces of the same session: 9vcek97t; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 515.5,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Qwen3.5-35B-A3B-MLX-2bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 2bit; 4k 515.5 68.4 11.9 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/11m6vz20",
   "source_url": "https://omlx.ai/benchmarks/performance/11m6vz20",
   "tg_tokens": null,
   "tps": 68.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-0yf5hwsq-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: gvggcrvf; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1396.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 1,396 104.5 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0yf5hwsq",
   "source_url": "https://omlx.ai/benchmarks/performance/0yf5hwsq",
   "tg_tokens": null,
   "tps": 104.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-0yf5hwsq-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: gvggcrvf; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1506.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 1,506 101.8 21.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0yf5hwsq",
   "source_url": "https://omlx.ai/benchmarks/performance/0yf5hwsq",
   "tg_tokens": null,
   "tps": 101.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-63pxaiwl-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 81qfrup1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 924.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 924.7 93.6 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/63pxaiwl",
   "source_url": "https://omlx.ai/benchmarks/performance/63pxaiwl",
   "tg_tokens": null,
   "tps": 93.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-63pxaiwl-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 81qfrup1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1074.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,074 90.2 21.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/63pxaiwl",
   "source_url": "https://omlx.ai/benchmarks/performance/63pxaiwl",
   "tg_tokens": null,
   "tps": 90.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-okkn6u41-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: m9eaeho1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4476.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,476 115.7 19.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/okkn6u41",
   "source_url": "https://omlx.ai/benchmarks/performance/okkn6u41",
   "tg_tokens": null,
   "tps": 115.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-okkn6u41-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: m9eaeho1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6038.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,038 109.0 20.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/okkn6u41",
   "source_url": "https://omlx.ai/benchmarks/performance/okkn6u41",
   "tg_tokens": null,
   "tps": 109.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fstp1axc-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: d2w5qdxm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4241.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,241 111.4 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fstp1axc",
   "source_url": "https://omlx.ai/benchmarks/performance/fstp1axc",
   "tg_tokens": null,
   "tps": 111.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fstp1axc-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: d2w5qdxm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5893.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 5,893 109.3 20.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fstp1axc",
   "source_url": "https://omlx.ai/benchmarks/performance/fstp1axc",
   "tg_tokens": null,
   "tps": 109.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-oglo5312-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: llt1vb2l; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 221.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 221.9 33.7 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oglo5312",
   "source_url": "https://omlx.ai/benchmarks/performance/oglo5312",
   "tg_tokens": null,
   "tps": 33.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-oglo5312-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: llt1vb2l; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 205.9 32.0 19.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oglo5312",
   "source_url": "https://omlx.ai/benchmarks/performance/oglo5312",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-q7vwlhcb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: jap825y4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 224.1 23.2 19.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q7vwlhcb",
   "source_url": "https://omlx.ai/benchmarks/performance/q7vwlhcb",
   "tg_tokens": null,
   "tps": 23.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-q7vwlhcb-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: jap825y4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 255.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 255.7 22.3 22.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q7vwlhcb",
   "source_url": "https://omlx.ai/benchmarks/performance/q7vwlhcb",
   "tg_tokens": null,
   "tps": 22.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-iacmhjmg-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 5rcdhovm, 6l3tpbnm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 259.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 259.9 44.5 20.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iacmhjmg",
   "source_url": "https://omlx.ai/benchmarks/performance/iacmhjmg",
   "tg_tokens": null,
   "tps": 44.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-iacmhjmg-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 5rcdhovm, 6l3tpbnm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 264.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 264.3 41.1 21.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iacmhjmg",
   "source_url": "https://omlx.ai/benchmarks/performance/iacmhjmg",
   "tg_tokens": null,
   "tps": 41.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-iacmhjmg-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: 5rcdhovm, 6l3tpbnm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 233.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 233.9 7.3 24.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iacmhjmg",
   "source_url": "https://omlx.ai/benchmarks/performance/iacmhjmg",
   "tg_tokens": null,
   "tps": 7.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Pro",
   "id": "omlx-f8h7ztgj-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 15.4 GB at this context; duplicate traces of the same session: cd3wpnnq; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 96.4,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-27B-oQ3e-mtp on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 3bit; 1k 96.4 12.4 15.4 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f8h7ztgj",
   "source_url": "https://omlx.ai/benchmarks/performance/f8h7ztgj",
   "tg_tokens": null,
   "tps": 12.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-f8h7ztgj-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: cd3wpnnq; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 97.2,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-27B-oQ3e-mtp on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 3bit; 4k 97.2 13.7 16.3 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f8h7ztgj",
   "source_url": "https://omlx.ai/benchmarks/performance/f8h7ztgj",
   "tg_tokens": null,
   "tps": 13.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-7900wg15-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: a5390vqe, kksq3hkn, sy0urn06; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 589.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 589.5 35.3 17.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7900wg15",
   "source_url": "https://omlx.ai/benchmarks/performance/7900wg15",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-7900wg15-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: a5390vqe, kksq3hkn, sy0urn06; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 688.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 688.7 34.8 19.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7900wg15",
   "source_url": "https://omlx.ai/benchmarks/performance/7900wg15",
   "tg_tokens": null,
   "tps": 34.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-7900wg15-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: a5390vqe, kksq3hkn, sy0urn06; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 680.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 680.7 26.1 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7900wg15",
   "source_url": "https://omlx.ai/benchmarks/performance/7900wg15",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-7900wg15-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: a5390vqe, kksq3hkn, sy0urn06; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 633.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 633.8 38.4 19.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7900wg15",
   "source_url": "https://omlx.ai/benchmarks/performance/7900wg15",
   "tg_tokens": null,
   "tps": 38.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-zo1kjdk9-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 20.9 GB; TTFT 53975 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 607.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k tokens 607.1 PP tok/s 28.4 TG tok/s 53975 TTFT (ms) 20.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zo1kjdk9",
   "source_url": "https://omlx.ai/benchmarks/performance/zo1kjdk9",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-d1hkwaj5-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 86.5 GB at this context; duplicate traces of the same session: o0zxv1kf, w7eovqut; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 931.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 931.2 60.5 86.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d1hkwaj5",
   "source_url": "https://omlx.ai/benchmarks/performance/d1hkwaj5",
   "tg_tokens": null,
   "tps": 60.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-d1hkwaj5-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 86.7 GB at this context; duplicate traces of the same session: o0zxv1kf, w7eovqut; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1691.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 1,691 96.4 86.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d1hkwaj5",
   "source_url": "https://omlx.ai/benchmarks/performance/d1hkwaj5",
   "tg_tokens": null,
   "tps": 96.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-d1hkwaj5-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 87.8 GB at this context; duplicate traces of the same session: o0zxv1kf, w7eovqut; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2172.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 2,172 57.6 87.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d1hkwaj5",
   "source_url": "https://omlx.ai/benchmarks/performance/d1hkwaj5",
   "tg_tokens": null,
   "tps": 57.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-d1hkwaj5-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 88.9 GB at this context; duplicate traces of the same session: o0zxv1kf, w7eovqut; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2456.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 2,456 74.1 88.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d1hkwaj5",
   "source_url": "https://omlx.ai/benchmarks/performance/d1hkwaj5",
   "tg_tokens": null,
   "tps": 74.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-d1hkwaj5-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 89.5 GB at this context; duplicate traces of the same session: o0zxv1kf, w7eovqut; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2184.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 2,184 69.8 89.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d1hkwaj5",
   "source_url": "https://omlx.ai/benchmarks/performance/d1hkwaj5",
   "tg_tokens": null,
   "tps": 69.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-kh9ar29h-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace, single run at 64k context; peak memory 198.3 GB; TTFT 125921 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 520.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 64k tokens 520.5 PP tok/s 38.3 TG tok/s 125921 TTFT (ms) 198.3 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kh9ar29h",
   "source_url": "https://omlx.ai/benchmarks/performance/kh9ar29h",
   "tg_tokens": null,
   "tps": 38.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-tnalurk0-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: 01avt1j8, 9w8zk0tp, ebzznjtw, nsqannxr; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 997.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 1k 997.6 72.9 16.3 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tnalurk0",
   "source_url": "https://omlx.ai/benchmarks/performance/tnalurk0",
   "tg_tokens": null,
   "tps": 72.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-tnalurk0-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 01avt1j8, 9w8zk0tp, ebzznjtw, nsqannxr; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 980.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 4k 980.6 62.1 16.6 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tnalurk0",
   "source_url": "https://omlx.ai/benchmarks/performance/tnalurk0",
   "tg_tokens": null,
   "tps": 62.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-tnalurk0-ctx8192",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: 01avt1j8, 9w8zk0tp, ebzznjtw, nsqannxr; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 950.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 8k 950.0 63.1 16.9 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tnalurk0",
   "source_url": "https://omlx.ai/benchmarks/performance/tnalurk0",
   "tg_tokens": null,
   "tps": 63.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-tnalurk0-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 01avt1j8, 9w8zk0tp, ebzznjtw, nsqannxr; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 887.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 16k 887.0 61.3 17.3 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tnalurk0",
   "source_url": "https://omlx.ai/benchmarks/performance/tnalurk0",
   "tg_tokens": null,
   "tps": 61.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-tnalurk0-ctx32768",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: 01avt1j8, 9w8zk0tp, ebzznjtw, nsqannxr; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 762.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 32k 762.9 48.0 18.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tnalurk0",
   "source_url": "https://omlx.ai/benchmarks/performance/tnalurk0",
   "tg_tokens": null,
   "tps": 48.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-tclj6j10-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: io7ypvuq, vkxapu87, w68y6ynr, wsxebpj3; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1018.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 1k 1,018 74.1 14.3 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tclj6j10",
   "source_url": "https://omlx.ai/benchmarks/performance/tclj6j10",
   "tg_tokens": null,
   "tps": 74.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-tclj6j10-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: io7ypvuq, vkxapu87, w68y6ynr, wsxebpj3; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 984.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 4k 984.4 71.8 14.6 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tclj6j10",
   "source_url": "https://omlx.ai/benchmarks/performance/tclj6j10",
   "tg_tokens": null,
   "tps": 71.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-tclj6j10-ctx8192",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: io7ypvuq, vkxapu87, w68y6ynr, wsxebpj3; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 947.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 8k 947.9 68.9 14.8 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tclj6j10",
   "source_url": "https://omlx.ai/benchmarks/performance/tclj6j10",
   "tg_tokens": null,
   "tps": 68.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-tclj6j10-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 15.2 GB at this context; duplicate traces of the same session: io7ypvuq, vkxapu87, w68y6ynr, wsxebpj3; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 887.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 16k 887.3 63.7 15.2 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tclj6j10",
   "source_url": "https://omlx.ai/benchmarks/performance/tclj6j10",
   "tg_tokens": null,
   "tps": 63.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-tclj6j10-ctx32768",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: io7ypvuq, vkxapu87, w68y6ynr, wsxebpj3; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 735.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 32k 735.5 55.6 16.0 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tclj6j10",
   "source_url": "https://omlx.ai/benchmarks/performance/tclj6j10",
   "tg_tokens": null,
   "tps": 55.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Pro",
   "id": "omlx-u3g0uupw-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 15.2 GB at this context; duplicate traces of the same session: tknscend; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 418.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 4k 418.7 27.0 15.2 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u3g0uupw",
   "source_url": "https://omlx.ai/benchmarks/performance/u3g0uupw",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-u3g0uupw-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: tknscend; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 453.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 16k 453.7 47.2 16.6 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u3g0uupw",
   "source_url": "https://omlx.ai/benchmarks/performance/u3g0uupw",
   "tg_tokens": null,
   "tps": 47.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-alvnasm8-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: lqr7fifu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 676.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "gemma-4-31B-it-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 676.9 35.2 28.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/alvnasm8",
   "source_url": "https://omlx.ai/benchmarks/performance/alvnasm8",
   "tg_tokens": null,
   "tps": 35.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-alvnasm8-ctx16384",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.1 GB at this context; duplicate traces of the same session: lqr7fifu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 651.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "gemma-4-31B-it-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 651.8 28.9 30.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/alvnasm8",
   "source_url": "https://omlx.ai/benchmarks/performance/alvnasm8",
   "tg_tokens": null,
   "tps": 28.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-alvnasm8-ctx65536",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.0 GB at this context; duplicate traces of the same session: lqr7fifu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 503.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "gemma-4-31B-it-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k 503.1 15.8 37.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/alvnasm8",
   "source_url": "https://omlx.ai/benchmarks/performance/alvnasm8",
   "tg_tokens": null,
   "tps": 15.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-li7iixq5-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 6.0 GB at this context; duplicate traces of the same session: 7pzz76ni; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 568.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E4B-it-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 568.1 22.6 6.0 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/li7iixq5",
   "source_url": "https://omlx.ai/benchmarks/performance/li7iixq5",
   "tg_tokens": null,
   "tps": 22.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-li7iixq5-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 6.1 GB at this context; duplicate traces of the same session: 7pzz76ni; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 577.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E4B-it-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 577.0 22.2 6.1 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/li7iixq5",
   "source_url": "https://omlx.ai/benchmarks/performance/li7iixq5",
   "tg_tokens": null,
   "tps": 22.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-wtmjpc9q-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.3 GB at this context; duplicate traces of the same session: 48echwkl; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 642.4,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "gpt-oss-20b-OptiQ-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 1k 642.4 52.8 11.3 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wtmjpc9q",
   "source_url": "https://omlx.ai/benchmarks/performance/wtmjpc9q",
   "tg_tokens": null,
   "tps": 52.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-wtmjpc9q-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.4 GB at this context; duplicate traces of the same session: 48echwkl; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 635.6,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "gpt-oss-20b-OptiQ-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 4k 635.6 50.8 11.4 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wtmjpc9q",
   "source_url": "https://omlx.ai/benchmarks/performance/wtmjpc9q",
   "tg_tokens": null,
   "tps": 50.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-en1b219f-ctx1024",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "TensorFold MLX build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 156.7 GB at this context; duplicate traces of the same session: 5v7k9b1a, 8dcz0u6d, vwkhutly; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 531.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "models--TensorFold--MiMo-V2.6-Flash-RL-MLX-4bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 531.9 51.2 156.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/en1b219f",
   "source_url": "https://omlx.ai/benchmarks/performance/en1b219f",
   "tg_tokens": null,
   "tps": 51.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-en1b219f-ctx4096",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "TensorFold MLX build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 157.8 GB at this context; duplicate traces of the same session: 5v7k9b1a, 8dcz0u6d, vwkhutly; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 636.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "models--TensorFold--MiMo-V2.6-Flash-RL-MLX-4bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 636.6 54.9 157.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/en1b219f",
   "source_url": "https://omlx.ai/benchmarks/performance/en1b219f",
   "tg_tokens": null,
   "tps": 54.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-en1b219f-ctx8192",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "TensorFold MLX build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 158.9 GB at this context; duplicate traces of the same session: 5v7k9b1a, 8dcz0u6d, vwkhutly; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 621.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "models--TensorFold--MiMo-V2.6-Flash-RL-MLX-4bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 621.0 54.3 158.9 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/en1b219f",
   "source_url": "https://omlx.ai/benchmarks/performance/en1b219f",
   "tg_tokens": null,
   "tps": 54.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-en1b219f-ctx16384",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "TensorFold MLX build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 161.1 GB at this context; duplicate traces of the same session: 5v7k9b1a, 8dcz0u6d, vwkhutly; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 587.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "models--TensorFold--MiMo-V2.6-Flash-RL-MLX-4bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 587.4 47.1 161.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/en1b219f",
   "source_url": "https://omlx.ai/benchmarks/performance/en1b219f",
   "tg_tokens": null,
   "tps": 47.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ckljih7d-ctx4096",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5396.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 5,396 165.2 23.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ckljih7d",
   "source_url": "https://omlx.ai/benchmarks/performance/ckljih7d",
   "tg_tokens": null,
   "tps": 165.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ckljih7d-ctx16384",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5155.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 5,155 187.7 23.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ckljih7d",
   "source_url": "https://omlx.ai/benchmarks/performance/ckljih7d",
   "tg_tokens": null,
   "tps": 187.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-2x34g6wf-ctx1024",
   "model": "KAT-Coder-V2.5-Dev",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: i1hnnvji; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 890.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "KAT-Coder-V2.5-Dev-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 890.5 69.7 21.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2x34g6wf",
   "source_url": "https://omlx.ai/benchmarks/performance/2x34g6wf",
   "tg_tokens": null,
   "tps": 69.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-2x34g6wf-ctx4096",
   "model": "KAT-Coder-V2.5-Dev",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: i1hnnvji; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 960.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "KAT-Coder-V2.5-Dev-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 960.1 66.7 22.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2x34g6wf",
   "source_url": "https://omlx.ai/benchmarks/performance/2x34g6wf",
   "tg_tokens": null,
   "tps": 66.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ybgc47jj-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: qn59gscc; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 869.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 869.8 67.2 21.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ybgc47jj",
   "source_url": "https://omlx.ai/benchmarks/performance/ybgc47jj",
   "tg_tokens": null,
   "tps": 67.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ybgc47jj-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: qn59gscc; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 943.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 943.0 74.7 22.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ybgc47jj",
   "source_url": "https://omlx.ai/benchmarks/performance/ybgc47jj",
   "tg_tokens": null,
   "tps": 74.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-i465hrag-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: t196lfz3, u3vpidc0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 357.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 357.1 16.0 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i465hrag",
   "source_url": "https://omlx.ai/benchmarks/performance/i465hrag",
   "tg_tokens": null,
   "tps": 16.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-i465hrag-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: t196lfz3, u3vpidc0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 357.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 357.1 13.6 17.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i465hrag",
   "source_url": "https://omlx.ai/benchmarks/performance/i465hrag",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-i465hrag-ctx16384",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: t196lfz3, u3vpidc0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 307.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 16k 307.2 12.5 18.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i465hrag",
   "source_url": "https://omlx.ai/benchmarks/performance/i465hrag",
   "tg_tokens": null,
   "tps": 12.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-paxmlunl-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: m4rpbq8l; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 393.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 393.1 33.1 21.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/paxmlunl",
   "source_url": "https://omlx.ai/benchmarks/performance/paxmlunl",
   "tg_tokens": null,
   "tps": 33.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-paxmlunl-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: m4rpbq8l; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 363.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 363.8 41.0 22.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/paxmlunl",
   "source_url": "https://omlx.ai/benchmarks/performance/paxmlunl",
   "tg_tokens": null,
   "tps": 41.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-iyvgah5y-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: njeiy9eb; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 493.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 493.0 17.3 17.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iyvgah5y",
   "source_url": "https://omlx.ai/benchmarks/performance/iyvgah5y",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-iyvgah5y-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: njeiy9eb; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 556.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 556.3 16.9 19.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iyvgah5y",
   "source_url": "https://omlx.ai/benchmarks/performance/iyvgah5y",
   "tg_tokens": null,
   "tps": 16.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0m504aqe-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: g7og2ui6; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 501.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 501.4 33.2 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0m504aqe",
   "source_url": "https://omlx.ai/benchmarks/performance/0m504aqe",
   "tg_tokens": null,
   "tps": 33.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0m504aqe-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: g7og2ui6; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 547.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 547.5 33.5 19.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0m504aqe",
   "source_url": "https://omlx.ai/benchmarks/performance/0m504aqe",
   "tg_tokens": null,
   "tps": 33.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-91yrv57k-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: shuc0qwb; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 457.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 457.4 17.2 17.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/91yrv57k",
   "source_url": "https://omlx.ai/benchmarks/performance/91yrv57k",
   "tg_tokens": null,
   "tps": 17.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-91yrv57k-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: shuc0qwb; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 526.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 526.5 17.0 19.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/91yrv57k",
   "source_url": "https://omlx.ai/benchmarks/performance/91yrv57k",
   "tg_tokens": null,
   "tps": 17.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-89tsg296-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: r3ucdntc; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 499.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 499.3 34.8 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/89tsg296",
   "source_url": "https://omlx.ai/benchmarks/performance/89tsg296",
   "tg_tokens": null,
   "tps": 34.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-89tsg296-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: r3ucdntc; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 529.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 529.6 41.3 19.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/89tsg296",
   "source_url": "https://omlx.ai/benchmarks/performance/89tsg296",
   "tg_tokens": null,
   "tps": 41.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-zus56oqs-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 367.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 367.7 31.0 17.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zus56oqs",
   "source_url": "https://omlx.ai/benchmarks/performance/zus56oqs",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-zus56oqs-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 363.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 363.1 38.9 18.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zus56oqs",
   "source_url": "https://omlx.ai/benchmarks/performance/zus56oqs",
   "tg_tokens": null,
   "tps": 38.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-8gcjxdk2-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 4hympgvr; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 503.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 503.1 32.8 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8gcjxdk2",
   "source_url": "https://omlx.ai/benchmarks/performance/8gcjxdk2",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-8gcjxdk2-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 4hympgvr; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 546.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 546.8 33.5 19.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8gcjxdk2",
   "source_url": "https://omlx.ai/benchmarks/performance/8gcjxdk2",
   "tg_tokens": null,
   "tps": 33.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-2ded4oan-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 02v1rt3n; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 500.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 500.7 31.5 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ded4oan",
   "source_url": "https://omlx.ai/benchmarks/performance/2ded4oan",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-2ded4oan-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 02v1rt3n; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 546.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 546.3 38.2 19.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ded4oan",
   "source_url": "https://omlx.ai/benchmarks/performance/2ded4oan",
   "tg_tokens": null,
   "tps": 38.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-td5t69ac-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 7hlslsvg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 493.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 493.7 34.7 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/td5t69ac",
   "source_url": "https://omlx.ai/benchmarks/performance/td5t69ac",
   "tg_tokens": null,
   "tps": 34.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-td5t69ac-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 7hlslsvg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 541.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 541.3 36.4 19.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/td5t69ac",
   "source_url": "https://omlx.ai/benchmarks/performance/td5t69ac",
   "tg_tokens": null,
   "tps": 36.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-a1w9k39d-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 499.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 499.5 32.7 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a1w9k39d",
   "source_url": "https://omlx.ai/benchmarks/performance/a1w9k39d",
   "tg_tokens": null,
   "tps": 32.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-a1w9k39d-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 543.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 543.7 31.0 19.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a1w9k39d",
   "source_url": "https://omlx.ai/benchmarks/performance/a1w9k39d",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-uitxp3mn-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 7out0cvk; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 136.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 136.4 26.4 20.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uitxp3mn",
   "source_url": "https://omlx.ai/benchmarks/performance/uitxp3mn",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-uitxp3mn-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: 7out0cvk; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 177.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 177.8 24.5 24.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uitxp3mn",
   "source_url": "https://omlx.ai/benchmarks/performance/uitxp3mn",
   "tg_tokens": null,
   "tps": 24.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Pro",
   "id": "omlx-nfxhugij-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: nahk2dd9; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 97.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 1k 97.6 7.9 16.6 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nfxhugij",
   "source_url": "https://omlx.ai/benchmarks/performance/nfxhugij",
   "tg_tokens": null,
   "tps": 7.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Pro",
   "id": "omlx-nfxhugij-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: nahk2dd9; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 100.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 4k 100.4 7.7 17.6 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nfxhugij",
   "source_url": "https://omlx.ai/benchmarks/performance/nfxhugij",
   "tg_tokens": null,
   "tps": 7.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-zjh9bs54-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: k5y7ms7p, ph4gv380; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 414.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 414.3 45.1 21.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zjh9bs54",
   "source_url": "https://omlx.ai/benchmarks/performance/zjh9bs54",
   "tg_tokens": null,
   "tps": 45.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-zjh9bs54-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: k5y7ms7p, ph4gv380; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 539.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 539.4 69.5 25.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zjh9bs54",
   "source_url": "https://omlx.ai/benchmarks/performance/zjh9bs54",
   "tg_tokens": null,
   "tps": 69.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-zjh9bs54-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: k5y7ms7p, ph4gv380; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 572.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 572.7 76.8 25.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zjh9bs54",
   "source_url": "https://omlx.ai/benchmarks/performance/zjh9bs54",
   "tg_tokens": null,
   "tps": 76.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-bgq8l7k3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 73yxca2z; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 126.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 126.6 26.3 18.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bgq8l7k3",
   "source_url": "https://omlx.ai/benchmarks/performance/bgq8l7k3",
   "tg_tokens": null,
   "tps": 26.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-bgq8l7k3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 73yxca2z; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 126.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 126.4 27.2 19.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bgq8l7k3",
   "source_url": "https://omlx.ai/benchmarks/performance/bgq8l7k3",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-upb9maa7-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 20.9 GB; TTFT 53888 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 608.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k tokens 608.1 PP tok/s 77.0 TG tok/s 53888 TTFT (ms) 20.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/upb9maa7",
   "source_url": "https://omlx.ai/benchmarks/performance/upb9maa7",
   "tg_tokens": null,
   "tps": 77.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-5as71wh2-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 20.9 GB; TTFT 55555 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 589.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k tokens 589.8 PP tok/s 55.0 TG tok/s 55555 TTFT (ms) 20.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5as71wh2",
   "source_url": "https://omlx.ai/benchmarks/performance/5as71wh2",
   "tg_tokens": null,
   "tps": 55.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ir21sqds-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 20.9 GB; TTFT 53964 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 607.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k tokens 607.2 PP tok/s 72.8 TG tok/s 53964 TTFT (ms) 20.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ir21sqds",
   "source_url": "https://omlx.ai/benchmarks/performance/ir21sqds",
   "tg_tokens": null,
   "tps": 72.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-m2y124tx-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 19.9 GB; TTFT 24260 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 675.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k tokens 675.4 PP tok/s 44.3 TG tok/s 24260 TTFT (ms) 19.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m2y124tx",
   "source_url": "https://omlx.ai/benchmarks/performance/m2y124tx",
   "tg_tokens": null,
   "tps": 44.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ud3i0f0g-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 23.5 GB; TTFT 81911 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 400.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k tokens 400.0 PP tok/s 17.2 TG tok/s 81911 TTFT (ms) 23.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ud3i0f0g",
   "source_url": "https://omlx.ai/benchmarks/performance/ud3i0f0g",
   "tg_tokens": null,
   "tps": 17.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-dfcc46v1-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 22.4 GB; TTFT 35106 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 466.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k tokens 466.7 PP tok/s 26.6 TG tok/s 35106 TTFT (ms) 22.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dfcc46v1",
   "source_url": "https://omlx.ai/benchmarks/performance/dfcc46v1",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-08pppc45-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 115.7 GB; TTFT 97134 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 674.7,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ6-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 64k tokens 674.7 PP tok/s 66.9 TG tok/s 97134 TTFT (ms) 115.7 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/08pppc45",
   "source_url": "https://omlx.ai/benchmarks/performance/08pppc45",
   "tg_tokens": null,
   "tps": 66.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7syp5mcg-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 69.6 GB at this context; duplicate traces of the same session: oh0705ka; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1934.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,934 67.1 69.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7syp5mcg",
   "source_url": "https://omlx.ai/benchmarks/performance/7syp5mcg",
   "tg_tokens": null,
   "tps": 67.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7syp5mcg-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 70.2 GB at this context; duplicate traces of the same session: oh0705ka; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1201.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,201 63.8 70.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7syp5mcg",
   "source_url": "https://omlx.ai/benchmarks/performance/7syp5mcg",
   "tg_tokens": null,
   "tps": 63.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-d9wnm1h0-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.5 GB at this context; duplicate traces of the same session: hrzphyno, k2h1pm33; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 72.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 72.8 9.7 37.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d9wnm1h0",
   "source_url": "https://omlx.ai/benchmarks/performance/d9wnm1h0",
   "tg_tokens": null,
   "tps": 9.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-d9wnm1h0-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.2 GB at this context; duplicate traces of the same session: hrzphyno, k2h1pm33; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 146.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 146.8 9.2 38.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d9wnm1h0",
   "source_url": "https://omlx.ai/benchmarks/performance/d9wnm1h0",
   "tg_tokens": null,
   "tps": 9.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-d9wnm1h0-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.5 GB at this context; duplicate traces of the same session: hrzphyno, k2h1pm33; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 181.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 181.1 9.2 39.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d9wnm1h0",
   "source_url": "https://omlx.ai/benchmarks/performance/d9wnm1h0",
   "tg_tokens": null,
   "tps": 9.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0zggd2a2-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 100.6 GB; TTFT 7009 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2338.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k tokens 2,338 PP tok/s 33.7 TG tok/s 7009 TTFT (ms) 100.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0zggd2a2",
   "source_url": "https://omlx.ai/benchmarks/performance/0zggd2a2",
   "tg_tokens": null,
   "tps": 33.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-7naa2xm6-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 94.1 GB; TTFT 96450 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 679.5,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 64k tokens 679.5 PP tok/s 56.4 TG tok/s 96450 TTFT (ms) 94.1 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7naa2xm6",
   "source_url": "https://omlx.ai/benchmarks/performance/7naa2xm6",
   "tg_tokens": null,
   "tps": 56.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-32epyqbt-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 102.8 GB at this context; duplicate traces of the same session: 7ah9tyry; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 659.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 4k 659.7 88.2 102.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/32epyqbt",
   "source_url": "https://omlx.ai/benchmarks/performance/32epyqbt",
   "tg_tokens": null,
   "tps": 88.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-32epyqbt-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 108.5 GB at this context; duplicate traces of the same session: 7ah9tyry; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 676.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 64k 676.1 59.9 108.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/32epyqbt",
   "source_url": "https://omlx.ai/benchmarks/performance/32epyqbt",
   "tg_tokens": null,
   "tps": 59.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Pro",
   "id": "omlx-li09p4xf-ctx1024",
   "model": "gemma-4-12B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 12.8 GB at this context; duplicate traces of the same session: 10pgo7j6; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 211.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-12B-it-qat-8bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 8bit; 1k 211.4 10.2 12.8 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/li09p4xf",
   "source_url": "https://omlx.ai/benchmarks/performance/li09p4xf",
   "tg_tokens": null,
   "tps": 10.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Pro",
   "id": "omlx-li09p4xf-ctx4096",
   "model": "gemma-4-12B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 13.2 GB at this context; duplicate traces of the same session: 10pgo7j6; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 208.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-12B-it-qat-8bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 8bit; 4k 208.5 9.9 13.2 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/li09p4xf",
   "source_url": "https://omlx.ai/benchmarks/performance/li09p4xf",
   "tg_tokens": null,
   "tps": 9.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-cuul22ud-ctx1024",
   "model": "gemma-4-12B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 14.2 GB at this context; duplicate traces of the same session: fdbgosff; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 213.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-12B-it-qat-8bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 8bit; 1k 213.1 17.2 14.2 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cuul22ud",
   "source_url": "https://omlx.ai/benchmarks/performance/cuul22ud",
   "tg_tokens": null,
   "tps": 17.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-cuul22ud-ctx4096",
   "model": "gemma-4-12B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 15.5 GB at this context; duplicate traces of the same session: fdbgosff; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 202.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-12B-it-qat-8bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 8bit; 4k 202.1 14.2 15.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cuul22ud",
   "source_url": "https://omlx.ai/benchmarks/performance/cuul22ud",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-8gy00flo-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: r7blhxxx; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 730.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 730.5 58.8 16.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8gy00flo",
   "source_url": "https://omlx.ai/benchmarks/performance/8gy00flo",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-8gy00flo-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: r7blhxxx; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 770.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 770.8 50.8 16.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8gy00flo",
   "source_url": "https://omlx.ai/benchmarks/performance/8gy00flo",
   "tg_tokens": null,
   "tps": 50.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xhj041t1-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 676.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "gemma-4-31B-it-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 676.9 35.2 28.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xhj041t1",
   "source_url": "https://omlx.ai/benchmarks/performance/xhj041t1",
   "tg_tokens": null,
   "tps": 35.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xhj041t1-ctx16384",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 651.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "gemma-4-31B-it-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 651.8 28.9 30.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xhj041t1",
   "source_url": "https://omlx.ai/benchmarks/performance/xhj041t1",
   "tg_tokens": null,
   "tps": 28.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xhj041t1-ctx65536",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 503.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "gemma-4-31B-it-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k 503.1 15.8 37.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xhj041t1",
   "source_url": "https://omlx.ai/benchmarks/performance/xhj041t1",
   "tg_tokens": null,
   "tps": 15.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-das4vdz8-ctx65536",
   "model": "Ornith-1.5-35B-A3B-REAP192",
   "notes": "mchen04 MLX build; oMLX community benchmark trace, single run at 64k context; peak memory 17.9 GB; TTFT 402042 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 163.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "mchen04--Ornith-1.5-35B-A3B-REAP192-mxfp4-MLX on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 64k tokens 163.0 PP tok/s 29.1 TG tok/s 402042 TTFT (ms) 17.9 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/das4vdz8",
   "source_url": "https://omlx.ai/benchmarks/performance/das4vdz8",
   "tg_tokens": null,
   "tps": 29.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-vuq3ci85-ctx16384",
   "model": "Ornith-1.5-35B-A3B-REAP192",
   "notes": "mchen04 MLX build; oMLX community benchmark trace, single run at 16k context; peak memory 15.7 GB; TTFT 53274 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 307.5,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "mchen04--Ornith-1.5-35B-A3B-REAP192-mxfp4-MLX on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k tokens 307.5 PP tok/s 44.3 TG tok/s 53274 TTFT (ms) 15.7 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vuq3ci85",
   "source_url": "https://omlx.ai/benchmarks/performance/vuq3ci85",
   "tg_tokens": null,
   "tps": 44.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-7zzdvxky-ctx4096",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5396.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 5,396 165.2 23.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7zzdvxky",
   "source_url": "https://omlx.ai/benchmarks/performance/7zzdvxky",
   "tg_tokens": null,
   "tps": 165.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-7zzdvxky-ctx16384",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5155.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 5,155 187.7 23.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7zzdvxky",
   "source_url": "https://omlx.ai/benchmarks/performance/7zzdvxky",
   "tg_tokens": null,
   "tps": 187.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qtvzwyb0-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.7 GB at this context; duplicate traces of the same session: 4rhrgx53, 812to39u, 86lw8gil, bzqsn9aj, dcs6e6gt, xuh97do5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 528.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 528.7 40.7 173.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qtvzwyb0",
   "source_url": "https://omlx.ai/benchmarks/performance/qtvzwyb0",
   "tg_tokens": null,
   "tps": 40.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qtvzwyb0-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.8 GB at this context; duplicate traces of the same session: 4rhrgx53, 812to39u, 86lw8gil, bzqsn9aj, dcs6e6gt, xuh97do5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 603.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 603.9 38.9 174.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qtvzwyb0",
   "source_url": "https://omlx.ai/benchmarks/performance/qtvzwyb0",
   "tg_tokens": null,
   "tps": 38.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qtvzwyb0-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.1 GB at this context; duplicate traces of the same session: 4rhrgx53, 812to39u, 86lw8gil, bzqsn9aj, dcs6e6gt, xuh97do5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 557.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 557.8 39.0 175.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qtvzwyb0",
   "source_url": "https://omlx.ai/benchmarks/performance/qtvzwyb0",
   "tg_tokens": null,
   "tps": 39.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qtvzwyb0-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.2 GB at this context; duplicate traces of the same session: 4rhrgx53, 812to39u, 86lw8gil, bzqsn9aj, dcs6e6gt, xuh97do5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 547.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 547.2 38.6 175.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qtvzwyb0",
   "source_url": "https://omlx.ai/benchmarks/performance/qtvzwyb0",
   "tg_tokens": null,
   "tps": 38.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qtvzwyb0-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.3 GB at this context; duplicate traces of the same session: 4rhrgx53, 812to39u, 86lw8gil, bzqsn9aj, dcs6e6gt, xuh97do5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 539.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 539.3 38.1 175.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qtvzwyb0",
   "source_url": "https://omlx.ai/benchmarks/performance/qtvzwyb0",
   "tg_tokens": null,
   "tps": 38.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qtvzwyb0-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.6 GB at this context; duplicate traces of the same session: 4rhrgx53, 812to39u, 86lw8gil, bzqsn9aj, dcs6e6gt, xuh97do5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 530.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 530.9 37.2 175.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qtvzwyb0",
   "source_url": "https://omlx.ai/benchmarks/performance/qtvzwyb0",
   "tg_tokens": null,
   "tps": 37.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qtvzwyb0-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 176.4 GB at this context; duplicate traces of the same session: 4rhrgx53, 812to39u, 86lw8gil, bzqsn9aj, dcs6e6gt, xuh97do5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 519.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 519.2 35.8 176.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qtvzwyb0",
   "source_url": "https://omlx.ai/benchmarks/performance/qtvzwyb0",
   "tg_tokens": null,
   "tps": 35.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-mrinjcfm-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 181.5 GB; TTFT 71311 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1838.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k tokens 1,838 PP tok/s 74.6 TG tok/s 71311 TTFT (ms) 181.5 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mrinjcfm",
   "source_url": "https://omlx.ai/benchmarks/performance/mrinjcfm",
   "tg_tokens": null,
   "tps": 74.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-x8ihh9lq-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 181.5 GB; TTFT 71068 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1844.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k tokens 1,844 PP tok/s 71.9 TG tok/s 71068 TTFT (ms) 181.5 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x8ihh9lq",
   "source_url": "https://omlx.ai/benchmarks/performance/x8ihh9lq",
   "tg_tokens": null,
   "tps": 71.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-2jkpog81-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 499.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 499.5 32.7 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2jkpog81",
   "source_url": "https://omlx.ai/benchmarks/performance/2jkpog81",
   "tg_tokens": null,
   "tps": 32.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-2jkpog81-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 543.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 543.7 31.0 19.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2jkpog81",
   "source_url": "https://omlx.ai/benchmarks/performance/2jkpog81",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-7zlh66jf-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: wpr4cini; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 513.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 513.9 33.9 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7zlh66jf",
   "source_url": "https://omlx.ai/benchmarks/performance/7zlh66jf",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-7zlh66jf-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: wpr4cini; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 577.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 577.5 37.0 19.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7zlh66jf",
   "source_url": "https://omlx.ai/benchmarks/performance/7zlh66jf",
   "tg_tokens": null,
   "tps": 37.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3vzp8to5-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: u1yocsay; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 510.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 510.4 33.3 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3vzp8to5",
   "source_url": "https://omlx.ai/benchmarks/performance/3vzp8to5",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3vzp8to5-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: u1yocsay; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 553.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 553.7 31.7 19.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3vzp8to5",
   "source_url": "https://omlx.ai/benchmarks/performance/3vzp8to5",
   "tg_tokens": null,
   "tps": 31.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-rcefn0q9-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: jefnq1i6; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 498.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 498.9 35.6 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rcefn0q9",
   "source_url": "https://omlx.ai/benchmarks/performance/rcefn0q9",
   "tg_tokens": null,
   "tps": 35.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-rcefn0q9-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: jefnq1i6; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 543.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 543.8 39.6 19.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rcefn0q9",
   "source_url": "https://omlx.ai/benchmarks/performance/rcefn0q9",
   "tg_tokens": null,
   "tps": 39.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-pg91azrv-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: 7oi0zpli; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 393.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 393.8 33.9 17.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pg91azrv",
   "source_url": "https://omlx.ai/benchmarks/performance/pg91azrv",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-pg91azrv-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 7oi0zpli; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 398.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 398.0 40.5 18.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pg91azrv",
   "source_url": "https://omlx.ai/benchmarks/performance/pg91azrv",
   "tg_tokens": null,
   "tps": 40.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3pw801qi-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 451.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 451.2 25.6 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3pw801qi",
   "source_url": "https://omlx.ai/benchmarks/performance/3pw801qi",
   "tg_tokens": null,
   "tps": 25.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3pw801qi-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 506.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 506.9 28.4 19.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3pw801qi",
   "source_url": "https://omlx.ai/benchmarks/performance/3pw801qi",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-vthxwhk1-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 452.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 452.5 24.7 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vthxwhk1",
   "source_url": "https://omlx.ai/benchmarks/performance/vthxwhk1",
   "tg_tokens": null,
   "tps": 24.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-vthxwhk1-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 506.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 506.2 22.8 19.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vthxwhk1",
   "source_url": "https://omlx.ai/benchmarks/performance/vthxwhk1",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-jnj6tlrl-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: dii1fame; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 347.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 347.4 22.7 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jnj6tlrl",
   "source_url": "https://omlx.ai/benchmarks/performance/jnj6tlrl",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-jnj6tlrl-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: dii1fame; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 351.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 351.5 20.3 27.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jnj6tlrl",
   "source_url": "https://omlx.ai/benchmarks/performance/jnj6tlrl",
   "tg_tokens": null,
   "tps": 20.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-fnw8c81t-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: c7pprfkk; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 301.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 301.6 20.2 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fnw8c81t",
   "source_url": "https://omlx.ai/benchmarks/performance/fnw8c81t",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-fnw8c81t-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: c7pprfkk; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 343.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 343.5 20.9 27.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fnw8c81t",
   "source_url": "https://omlx.ai/benchmarks/performance/fnw8c81t",
   "tg_tokens": null,
   "tps": 20.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-pbhqeqg1-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: osm4ckfg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 332.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 332.9 19.8 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pbhqeqg1",
   "source_url": "https://omlx.ai/benchmarks/performance/pbhqeqg1",
   "tg_tokens": null,
   "tps": 19.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-pbhqeqg1-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: osm4ckfg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 321.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 321.8 21.7 27.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pbhqeqg1",
   "source_url": "https://omlx.ai/benchmarks/performance/pbhqeqg1",
   "tg_tokens": null,
   "tps": 21.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-lq8dcbcv-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: qj4b4ow0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 273.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 273.8 18.1 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lq8dcbcv",
   "source_url": "https://omlx.ai/benchmarks/performance/lq8dcbcv",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-lq8dcbcv-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: qj4b4ow0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 307.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 307.7 19.9 27.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lq8dcbcv",
   "source_url": "https://omlx.ai/benchmarks/performance/lq8dcbcv",
   "tg_tokens": null,
   "tps": 19.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3qh6pcxn-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: ekfg4ygn; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 310.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 310.7 17.9 26.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3qh6pcxn",
   "source_url": "https://omlx.ai/benchmarks/performance/3qh6pcxn",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3qh6pcxn-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: ekfg4ygn; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 309.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 309.4 21.1 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3qh6pcxn",
   "source_url": "https://omlx.ai/benchmarks/performance/3qh6pcxn",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-xbj1v10k-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: kuipi6gq; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 311.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 311.5 19.3 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xbj1v10k",
   "source_url": "https://omlx.ai/benchmarks/performance/xbj1v10k",
   "tg_tokens": null,
   "tps": 19.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-xbj1v10k-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: kuipi6gq; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 310.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 310.1 17.4 27.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xbj1v10k",
   "source_url": "https://omlx.ai/benchmarks/performance/xbj1v10k",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-g7fk3tmk-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: lav3gez7; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 306.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 306.9 19.7 26.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g7fk3tmk",
   "source_url": "https://omlx.ai/benchmarks/performance/g7fk3tmk",
   "tg_tokens": null,
   "tps": 19.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-g7fk3tmk-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: lav3gez7; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 303.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 303.9 23.4 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g7fk3tmk",
   "source_url": "https://omlx.ai/benchmarks/performance/g7fk3tmk",
   "tg_tokens": null,
   "tps": 23.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-c9laprhn-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: vle6dn6d; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 306.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 306.8 15.3 26.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c9laprhn",
   "source_url": "https://omlx.ai/benchmarks/performance/c9laprhn",
   "tg_tokens": null,
   "tps": 15.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-c9laprhn-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: vle6dn6d; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 286.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 286.4 18.8 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c9laprhn",
   "source_url": "https://omlx.ai/benchmarks/performance/c9laprhn",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-i2b6uxig-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: vssgceo2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 302.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 302.1 19.3 26.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i2b6uxig",
   "source_url": "https://omlx.ai/benchmarks/performance/i2b6uxig",
   "tg_tokens": null,
   "tps": 19.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-i2b6uxig-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: vssgceo2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 286.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 286.0 22.7 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i2b6uxig",
   "source_url": "https://omlx.ai/benchmarks/performance/i2b6uxig",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-4iti179d-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 32.4 GB at this context; duplicate traces of the same session: reozvfbp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 320.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 320.0 19.7 32.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4iti179d",
   "source_url": "https://omlx.ai/benchmarks/performance/4iti179d",
   "tg_tokens": null,
   "tps": 19.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-4iti179d-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 33.4 GB at this context; duplicate traces of the same session: reozvfbp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 344.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 344.0 19.0 33.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4iti179d",
   "source_url": "https://omlx.ai/benchmarks/performance/4iti179d",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1mx7gvo7-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: vqd97r2s; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 273.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 273.7 14.4 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1mx7gvo7",
   "source_url": "https://omlx.ai/benchmarks/performance/1mx7gvo7",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1mx7gvo7-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: vqd97r2s; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 277.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 277.2 18.3 27.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1mx7gvo7",
   "source_url": "https://omlx.ai/benchmarks/performance/1mx7gvo7",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-145dh3md-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: quv118j9; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 286.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 286.5 17.7 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/145dh3md",
   "source_url": "https://omlx.ai/benchmarks/performance/145dh3md",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-145dh3md-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: quv118j9; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 294.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 294.5 17.0 27.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/145dh3md",
   "source_url": "https://omlx.ai/benchmarks/performance/145dh3md",
   "tg_tokens": null,
   "tps": 17.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-um1so507-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: c3pvdldy; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 321.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 321.9 20.7 26.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/um1so507",
   "source_url": "https://omlx.ai/benchmarks/performance/um1so507",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-um1so507-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: c3pvdldy; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 311.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 311.8 20.4 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/um1so507",
   "source_url": "https://omlx.ai/benchmarks/performance/um1so507",
   "tg_tokens": null,
   "tps": 20.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ijpr7i1l-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: pu92gg9u; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 316.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 316.3 19.1 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ijpr7i1l",
   "source_url": "https://omlx.ai/benchmarks/performance/ijpr7i1l",
   "tg_tokens": null,
   "tps": 19.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ijpr7i1l-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: pu92gg9u; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 319.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 319.9 24.7 27.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ijpr7i1l",
   "source_url": "https://omlx.ai/benchmarks/performance/ijpr7i1l",
   "tg_tokens": null,
   "tps": 24.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1lfjxegh-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 70.3 GB; TTFT 1122 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3650.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 3,650 PP tok/s 73.0 TG tok/s 1122 TTFT (ms) 70.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1lfjxegh",
   "source_url": "https://omlx.ai/benchmarks/performance/1lfjxegh",
   "tg_tokens": null,
   "tps": 73.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-a7hbyjqv-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 69.6 GB at this context; duplicate traces of the same session: 3p5k5glz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2141.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,141 77.0 69.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a7hbyjqv",
   "source_url": "https://omlx.ai/benchmarks/performance/a7hbyjqv",
   "tg_tokens": null,
   "tps": 77.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-a7hbyjqv-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 70.2 GB at this context; duplicate traces of the same session: 3p5k5glz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3060.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 3,060 73.6 70.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a7hbyjqv",
   "source_url": "https://omlx.ai/benchmarks/performance/a7hbyjqv",
   "tg_tokens": null,
   "tps": 73.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-l7fvxi16-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 69.6 GB; TTFT 517 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1982.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,982 PP tok/s 67.9 TG tok/s 517 TTFT (ms) 69.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l7fvxi16",
   "source_url": "https://omlx.ai/benchmarks/performance/l7fvxi16",
   "tg_tokens": null,
   "tps": 67.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-opan4s20-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.5 GB at this context; duplicate traces of the same session: ct4esdur; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2355.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,355 80.3 74.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/opan4s20",
   "source_url": "https://omlx.ai/benchmarks/performance/opan4s20",
   "tg_tokens": null,
   "tps": 80.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-opan4s20-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.8 GB at this context; duplicate traces of the same session: ct4esdur; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2255.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,255 56.6 78.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/opan4s20",
   "source_url": "https://omlx.ai/benchmarks/performance/opan4s20",
   "tg_tokens": null,
   "tps": 56.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-f4g4tox4-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: f8re5qm4, yzhhpyio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 688.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "gemma-4-31B-it-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 688.5 29.7 29.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f4g4tox4",
   "source_url": "https://omlx.ai/benchmarks/performance/f4g4tox4",
   "tg_tokens": null,
   "tps": 29.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-f4g4tox4-ctx16384",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.5 GB at this context; duplicate traces of the same session: f8re5qm4, yzhhpyio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 644.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "gemma-4-31B-it-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 644.9 37.6 33.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f4g4tox4",
   "source_url": "https://omlx.ai/benchmarks/performance/f4g4tox4",
   "tg_tokens": null,
   "tps": 37.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-f4g4tox4-ctx65536",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 47.0 GB at this context; duplicate traces of the same session: f8re5qm4, yzhhpyio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 491.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "gemma-4-31B-it-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k 491.4 32.0 47.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f4g4tox4",
   "source_url": "https://omlx.ai/benchmarks/performance/f4g4tox4",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-gtj3zkpl-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 4.5 GB at this context; duplicate traces of the same session: 50yorf7b; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 564.6,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "gemma-4-e4b-it-mxfp4 on M2 Pro (16c); M2 Pro (16c) \u00b7 16 GB \u00b7 4bit; 1k 564.6 50.0 4.5 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gtj3zkpl",
   "source_url": "https://omlx.ai/benchmarks/performance/gtj3zkpl",
   "tg_tokens": null,
   "tps": 50.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-gtj3zkpl-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 4.5 GB at this context; duplicate traces of the same session: 50yorf7b; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 589.9,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "gemma-4-e4b-it-mxfp4 on M2 Pro (16c); M2 Pro (16c) \u00b7 16 GB \u00b7 4bit; 4k 589.9 47.8 4.5 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gtj3zkpl",
   "source_url": "https://omlx.ai/benchmarks/performance/gtj3zkpl",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-xk1qdypi-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 451.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 451.2 25.6 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xk1qdypi",
   "source_url": "https://omlx.ai/benchmarks/performance/xk1qdypi",
   "tg_tokens": null,
   "tps": 25.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-xk1qdypi-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 506.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 506.9 28.4 19.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xk1qdypi",
   "source_url": "https://omlx.ai/benchmarks/performance/xk1qdypi",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-wse4yx88-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 452.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 452.5 24.7 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wse4yx88",
   "source_url": "https://omlx.ai/benchmarks/performance/wse4yx88",
   "tg_tokens": null,
   "tps": 24.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-wse4yx88-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 506.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 506.2 22.8 19.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wse4yx88",
   "source_url": "https://omlx.ai/benchmarks/performance/wse4yx88",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-d4518713-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: e2hvnthp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 464.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 464.3 29.0 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d4518713",
   "source_url": "https://omlx.ai/benchmarks/performance/d4518713",
   "tg_tokens": null,
   "tps": 29.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-d4518713-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: e2hvnthp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 512.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 512.2 35.7 19.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d4518713",
   "source_url": "https://omlx.ai/benchmarks/performance/d4518713",
   "tg_tokens": null,
   "tps": 35.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-cs9d0hy3-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 4eehjh6o; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 464.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 464.4 29.9 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cs9d0hy3",
   "source_url": "https://omlx.ai/benchmarks/performance/cs9d0hy3",
   "tg_tokens": null,
   "tps": 29.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-cs9d0hy3-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 4eehjh6o; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 485.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 485.3 36.6 19.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cs9d0hy3",
   "source_url": "https://omlx.ai/benchmarks/performance/cs9d0hy3",
   "tg_tokens": null,
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-x3knu411-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: vuz2lolw; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 438.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 438.3 26.4 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x3knu411",
   "source_url": "https://omlx.ai/benchmarks/performance/x3knu411",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-x3knu411-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: vuz2lolw; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 474.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 474.2 28.3 19.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x3knu411",
   "source_url": "https://omlx.ai/benchmarks/performance/x3knu411",
   "tg_tokens": null,
   "tps": 28.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-hf8tiq4u-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 31l4b6d8; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 437.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 437.2 24.8 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hf8tiq4u",
   "source_url": "https://omlx.ai/benchmarks/performance/hf8tiq4u",
   "tg_tokens": null,
   "tps": 24.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-hf8tiq4u-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 31l4b6d8; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 472.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 472.7 25.3 19.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hf8tiq4u",
   "source_url": "https://omlx.ai/benchmarks/performance/hf8tiq4u",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ogxyphlo-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: rlwqc74p; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 479.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 479.2 24.4 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ogxyphlo",
   "source_url": "https://omlx.ai/benchmarks/performance/ogxyphlo",
   "tg_tokens": null,
   "tps": 24.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ogxyphlo-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: rlwqc74p; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 488.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 488.2 26.2 19.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ogxyphlo",
   "source_url": "https://omlx.ai/benchmarks/performance/ogxyphlo",
   "tg_tokens": null,
   "tps": 26.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ysne7psb-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 2vo9g0wd; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 450.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 450.9 26.1 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ysne7psb",
   "source_url": "https://omlx.ai/benchmarks/performance/ysne7psb",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ysne7psb-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 2vo9g0wd; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 476.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 476.5 33.7 19.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ysne7psb",
   "source_url": "https://omlx.ai/benchmarks/performance/ysne7psb",
   "tg_tokens": null,
   "tps": 33.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-mcqkfkcv-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: m5whj0m2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 351.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 351.6 25.0 17.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mcqkfkcv",
   "source_url": "https://omlx.ai/benchmarks/performance/mcqkfkcv",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-mcqkfkcv-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: m5whj0m2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 315.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 315.5 25.8 18.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mcqkfkcv",
   "source_url": "https://omlx.ai/benchmarks/performance/mcqkfkcv",
   "tg_tokens": null,
   "tps": 25.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-71dbv8aw-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: 0kuoa9xp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 344.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 344.7 29.8 17.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/71dbv8aw",
   "source_url": "https://omlx.ai/benchmarks/performance/71dbv8aw",
   "tg_tokens": null,
   "tps": 29.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-71dbv8aw-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 0kuoa9xp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 323.2 35.0 18.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/71dbv8aw",
   "source_url": "https://omlx.ai/benchmarks/performance/71dbv8aw",
   "tg_tokens": null,
   "tps": 35.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-a3a4o7q9-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 0mxvwxs4; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 382.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 382.4 30.0 21.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a3a4o7q9",
   "source_url": "https://omlx.ai/benchmarks/performance/a3a4o7q9",
   "tg_tokens": null,
   "tps": 30.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-a3a4o7q9-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: 0mxvwxs4; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 359.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 359.5 33.6 22.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a3a4o7q9",
   "source_url": "https://omlx.ai/benchmarks/performance/a3a4o7q9",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-bhaa1fda-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 94x17q1k; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 448.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 448.9 28.5 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhaa1fda",
   "source_url": "https://omlx.ai/benchmarks/performance/bhaa1fda",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-bhaa1fda-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 94x17q1k; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 475.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 475.4 28.5 19.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhaa1fda",
   "source_url": "https://omlx.ai/benchmarks/performance/bhaa1fda",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-vvj90egl-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: 053j3rxm; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 342.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 342.4 28.2 17.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vvj90egl",
   "source_url": "https://omlx.ai/benchmarks/performance/vvj90egl",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-vvj90egl-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 053j3rxm; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 336.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 336.8 35.1 18.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vvj90egl",
   "source_url": "https://omlx.ai/benchmarks/performance/vvj90egl",
   "tg_tokens": null,
   "tps": 35.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qbqiq0u3-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: v2rsxjbt; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 316.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 316.2 23.6 17.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qbqiq0u3",
   "source_url": "https://omlx.ai/benchmarks/performance/qbqiq0u3",
   "tg_tokens": null,
   "tps": 23.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qbqiq0u3-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: v2rsxjbt; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 337.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 337.6 24.1 18.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qbqiq0u3",
   "source_url": "https://omlx.ai/benchmarks/performance/qbqiq0u3",
   "tg_tokens": null,
   "tps": 24.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-mxfkmtzd-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: liy9wppd; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 345.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 345.5 27.4 17.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mxfkmtzd",
   "source_url": "https://omlx.ai/benchmarks/performance/mxfkmtzd",
   "tg_tokens": null,
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-mxfkmtzd-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: liy9wppd; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 326.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 326.0 36.1 18.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mxfkmtzd",
   "source_url": "https://omlx.ai/benchmarks/performance/mxfkmtzd",
   "tg_tokens": null,
   "tps": 36.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-zkxrr0yd-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: 828v9xqv; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 380.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 380.6 29.9 17.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zkxrr0yd",
   "source_url": "https://omlx.ai/benchmarks/performance/zkxrr0yd",
   "tg_tokens": null,
   "tps": 29.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-zkxrr0yd-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 828v9xqv; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 348.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 348.9 34.5 18.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zkxrr0yd",
   "source_url": "https://omlx.ai/benchmarks/performance/zkxrr0yd",
   "tg_tokens": null,
   "tps": 34.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-jahxi63o-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace, single run at 1k context; peak memory 17.8 GB; TTFT 2180 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 469.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k tokens 469.7 PP tok/s 29.4 TG tok/s 2180 TTFT (ms) 17.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jahxi63o",
   "source_url": "https://omlx.ai/benchmarks/performance/jahxi63o",
   "tg_tokens": null,
   "tps": 29.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-hierwyje-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace, single run at 1k context; peak memory 17.7 GB; TTFT 2756 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 371.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k tokens 371.5 PP tok/s 27.7 TG tok/s 2756 TTFT (ms) 17.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hierwyje",
   "source_url": "https://omlx.ai/benchmarks/performance/hierwyje",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-mub8sjhm-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace, single run at 1k context; peak memory 17.7 GB; TTFT 2637 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 388.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k tokens 388.4 PP tok/s 28.7 TG tok/s 2637 TTFT (ms) 17.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mub8sjhm",
   "source_url": "https://omlx.ai/benchmarks/performance/mub8sjhm",
   "tg_tokens": null,
   "tps": 28.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0k9tav9c-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: pyrpbsxk; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 279.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 279.4 18.1 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0k9tav9c",
   "source_url": "https://omlx.ai/benchmarks/performance/0k9tav9c",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0k9tav9c-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: pyrpbsxk; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 291.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 291.2 20.1 27.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0k9tav9c",
   "source_url": "https://omlx.ai/benchmarks/performance/0k9tav9c",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-70mvvde3-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: 3stlcjaq; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 322.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 322.3 20.2 26.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/70mvvde3",
   "source_url": "https://omlx.ai/benchmarks/performance/70mvvde3",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-70mvvde3-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: 3stlcjaq; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 313.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 313.1 22.5 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/70mvvde3",
   "source_url": "https://omlx.ai/benchmarks/performance/70mvvde3",
   "tg_tokens": null,
   "tps": 22.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ozcwp0hw-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: frs175hu; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 348.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 348.6 12.4 22.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ozcwp0hw",
   "source_url": "https://omlx.ai/benchmarks/performance/ozcwp0hw",
   "tg_tokens": null,
   "tps": 12.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ozcwp0hw-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: frs175hu; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 351.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 351.8 12.3 23.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ozcwp0hw",
   "source_url": "https://omlx.ai/benchmarks/performance/ozcwp0hw",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-i5p75a7q-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: i92bfvhb; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 354.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 354.7 22.6 26.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i5p75a7q",
   "source_url": "https://omlx.ai/benchmarks/performance/i5p75a7q",
   "tg_tokens": null,
   "tps": 22.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-i5p75a7q-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: i92bfvhb; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 349.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 349.2 25.2 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i5p75a7q",
   "source_url": "https://omlx.ai/benchmarks/performance/i5p75a7q",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-rijt18h0-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: nwlnl3pf; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 343.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 343.8 22.4 26.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rijt18h0",
   "source_url": "https://omlx.ai/benchmarks/performance/rijt18h0",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-rijt18h0-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: nwlnl3pf; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 347.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 347.3 22.9 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rijt18h0",
   "source_url": "https://omlx.ai/benchmarks/performance/rijt18h0",
   "tg_tokens": null,
   "tps": 22.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-yg01vy1k-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: o4l0gzjs; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 351.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 351.6 12.4 22.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yg01vy1k",
   "source_url": "https://omlx.ai/benchmarks/performance/yg01vy1k",
   "tg_tokens": null,
   "tps": 12.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-yg01vy1k-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: o4l0gzjs; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 353.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 353.7 12.3 23.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yg01vy1k",
   "source_url": "https://omlx.ai/benchmarks/performance/yg01vy1k",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-j51os670-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: s0m8jin5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 344.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 344.5 19.8 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j51os670",
   "source_url": "https://omlx.ai/benchmarks/performance/j51os670",
   "tg_tokens": null,
   "tps": 19.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-j51os670-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: s0m8jin5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 350.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 350.4 25.3 27.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j51os670",
   "source_url": "https://omlx.ai/benchmarks/performance/j51os670",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-15yyemrg-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 348.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 348.2 23.8 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/15yyemrg",
   "source_url": "https://omlx.ai/benchmarks/performance/15yyemrg",
   "tg_tokens": null,
   "tps": 23.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-15yyemrg-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 353.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 353.7 25.3 27.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/15yyemrg",
   "source_url": "https://omlx.ai/benchmarks/performance/15yyemrg",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-m0vidrpx-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 15.4 GB at this context; duplicate traces of the same session: a2hjm2v9, dald638o; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 96.1,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-27B-oQ3e-mtp on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 3bit; 1k 96.1 13.8 15.4 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m0vidrpx",
   "source_url": "https://omlx.ai/benchmarks/performance/m0vidrpx",
   "tg_tokens": null,
   "tps": 13.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-m0vidrpx-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: a2hjm2v9, dald638o; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 97.4,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-27B-oQ3e-mtp on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 3bit; 4k 97.4 12.6 16.3 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m0vidrpx",
   "source_url": "https://omlx.ai/benchmarks/performance/m0vidrpx",
   "tg_tokens": null,
   "tps": 12.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-m0vidrpx-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.5 GB at this context; duplicate traces of the same session: a2hjm2v9, dald638o; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 92.4,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-27B-oQ3e-mtp on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 3bit; 16k 92.4 12.9 17.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m0vidrpx",
   "source_url": "https://omlx.ai/benchmarks/performance/m0vidrpx",
   "tg_tokens": null,
   "tps": 12.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-n92wb3se-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: bwdxc6qp, r06ugic0, w6rjn57r; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 589.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 589.5 35.5 17.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n92wb3se",
   "source_url": "https://omlx.ai/benchmarks/performance/n92wb3se",
   "tg_tokens": null,
   "tps": 35.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-n92wb3se-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: bwdxc6qp, r06ugic0, w6rjn57r; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 689.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 689.0 45.3 19.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n92wb3se",
   "source_url": "https://omlx.ai/benchmarks/performance/n92wb3se",
   "tg_tokens": null,
   "tps": 45.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-n92wb3se-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: bwdxc6qp, r06ugic0, w6rjn57r; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 706.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 706.2 22.8 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n92wb3se",
   "source_url": "https://omlx.ai/benchmarks/performance/n92wb3se",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-n92wb3se-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: bwdxc6qp, r06ugic0, w6rjn57r; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 668.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 668.6 39.6 19.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n92wb3se",
   "source_url": "https://omlx.ai/benchmarks/performance/n92wb3se",
   "tg_tokens": null,
   "tps": 39.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1i2axhy8-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 69.6 GB; TTFT 517 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1983.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,983 PP tok/s 67.5 TG tok/s 517 TTFT (ms) 69.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1i2axhy8",
   "source_url": "https://omlx.ai/benchmarks/performance/1i2axhy8",
   "tg_tokens": null,
   "tps": 67.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1fyav1sh-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 40.0 GB; TTFT 2061 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1988.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 1,988 PP tok/s 54.8 TG tok/s 2061 TTFT (ms) 40.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1fyav1sh",
   "source_url": "https://omlx.ai/benchmarks/performance/1fyav1sh",
   "tg_tokens": null,
   "tps": 54.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tc7ye24a-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 74.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2149.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,149 52.0 74.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tc7ye24a",
   "source_url": "https://omlx.ai/benchmarks/performance/tc7ye24a",
   "tg_tokens": null,
   "tps": 52.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tc7ye24a-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 75.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3101.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 3,101 54.9 75.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tc7ye24a",
   "source_url": "https://omlx.ai/benchmarks/performance/tc7ye24a",
   "tg_tokens": null,
   "tps": 54.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-wxki2fhy-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.7 GB at this context; duplicate traces of the same session: sgnrc57u; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1050.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,050 101.3 100.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wxki2fhy",
   "source_url": "https://omlx.ai/benchmarks/performance/wxki2fhy",
   "tg_tokens": null,
   "tps": 101.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-wxki2fhy-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: sgnrc57u; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1416.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,416 103.9 101.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wxki2fhy",
   "source_url": "https://omlx.ai/benchmarks/performance/wxki2fhy",
   "tg_tokens": null,
   "tps": 103.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-wxki2fhy-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.9 GB at this context; duplicate traces of the same session: sgnrc57u; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1394.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 1,394 85.0 101.9 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wxki2fhy",
   "source_url": "https://omlx.ai/benchmarks/performance/wxki2fhy",
   "tg_tokens": null,
   "tps": 85.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-wxki2fhy-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 106.1 GB at this context; duplicate traces of the same session: sgnrc57u; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1367.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 1,367 72.7 106.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wxki2fhy",
   "source_url": "https://omlx.ai/benchmarks/performance/wxki2fhy",
   "tg_tokens": null,
   "tps": 72.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-wxki2fhy-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 112.4 GB at this context; duplicate traces of the same session: sgnrc57u; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1339.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 1,339 56.0 112.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wxki2fhy",
   "source_url": "https://omlx.ai/benchmarks/performance/wxki2fhy",
   "tg_tokens": null,
   "tps": 56.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-jv7gu4bm-ctx1024",
   "model": "Qwen3-VL-8B-Instruct",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: ewwlh4kd; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 158.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-VL-8B-Instruct-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 158.6 20.5 4.9 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jv7gu4bm",
   "source_url": "https://omlx.ai/benchmarks/performance/jv7gu4bm",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-jv7gu4bm-ctx4096",
   "model": "Qwen3-VL-8B-Instruct",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 5.4 GB at this context; duplicate traces of the same session: ewwlh4kd; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 152.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-VL-8B-Instruct-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 152.8 18.4 5.4 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jv7gu4bm",
   "source_url": "https://omlx.ai/benchmarks/performance/jv7gu4bm",
   "tg_tokens": null,
   "tps": 18.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-hvhk62kq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS bare speed FP16 build; oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: 8qoqtsp4; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 115.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Bare-Speed-FP16 on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 1k 115.1 17.8 15.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hvhk62kq",
   "source_url": "https://omlx.ai/benchmarks/performance/hvhk62kq",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-hvhk62kq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS bare speed FP16 build; oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: 8qoqtsp4; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 119.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Bare-Speed-FP16 on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k 119.1 16.8 17.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hvhk62kq",
   "source_url": "https://omlx.ai/benchmarks/performance/hvhk62kq",
   "tg_tokens": null,
   "tps": 16.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-mql3cass-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: fnf78vd2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 487.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 487.8 17.3 17.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mql3cass",
   "source_url": "https://omlx.ai/benchmarks/performance/mql3cass",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-mql3cass-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: fnf78vd2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 558.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 558.9 17.0 19.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mql3cass",
   "source_url": "https://omlx.ai/benchmarks/performance/mql3cass",
   "tg_tokens": null,
   "tps": 17.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-p65f54ra-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: rabe4t3a; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 487.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 487.5 32.6 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p65f54ra",
   "source_url": "https://omlx.ai/benchmarks/performance/p65f54ra",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-p65f54ra-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: rabe4t3a; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 524.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 524.3 31.4 19.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p65f54ra",
   "source_url": "https://omlx.ai/benchmarks/performance/p65f54ra",
   "tg_tokens": null,
   "tps": 31.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-5oaaf50g-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: l101zxyx; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 482.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 482.5 17.4 17.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5oaaf50g",
   "source_url": "https://omlx.ai/benchmarks/performance/5oaaf50g",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-5oaaf50g-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: l101zxyx; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 537.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 537.0 17.0 19.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5oaaf50g",
   "source_url": "https://omlx.ai/benchmarks/performance/5oaaf50g",
   "tg_tokens": null,
   "tps": 17.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qk5glrms-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 32.4 GB at this context; duplicate traces of the same session: 3y2asnsa; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 361.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 361.6 23.0 32.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qk5glrms",
   "source_url": "https://omlx.ai/benchmarks/performance/qk5glrms",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qk5glrms-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 33.4 GB at this context; duplicate traces of the same session: 3y2asnsa; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 353.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 353.7 24.3 33.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qk5glrms",
   "source_url": "https://omlx.ai/benchmarks/performance/qk5glrms",
   "tg_tokens": null,
   "tps": 24.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-g0thx7rv-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 348.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 348.2 23.8 26.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g0thx7rv",
   "source_url": "https://omlx.ai/benchmarks/performance/g0thx7rv",
   "tg_tokens": null,
   "tps": 23.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-g0thx7rv-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 353.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 353.7 25.3 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g0thx7rv",
   "source_url": "https://omlx.ai/benchmarks/performance/g0thx7rv",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-t4p6611e-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: by4nx6p4; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 349.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 349.9 23.4 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t4p6611e",
   "source_url": "https://omlx.ai/benchmarks/performance/t4p6611e",
   "tg_tokens": null,
   "tps": 23.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-t4p6611e-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: by4nx6p4; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 353.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 353.1 23.5 27.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t4p6611e",
   "source_url": "https://omlx.ai/benchmarks/performance/t4p6611e",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-66bqjvdu-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: dxioncad; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 349.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 349.4 23.5 26.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/66bqjvdu",
   "source_url": "https://omlx.ai/benchmarks/performance/66bqjvdu",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-66bqjvdu-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: dxioncad; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 353.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 353.8 25.2 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/66bqjvdu",
   "source_url": "https://omlx.ai/benchmarks/performance/66bqjvdu",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-i7iscyye-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: t9zs74q4; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 346.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 346.4 23.3 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i7iscyye",
   "source_url": "https://omlx.ai/benchmarks/performance/i7iscyye",
   "tg_tokens": null,
   "tps": 23.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-i7iscyye-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: t9zs74q4; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 352.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 352.6 25.1 27.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i7iscyye",
   "source_url": "https://omlx.ai/benchmarks/performance/i7iscyye",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-p8884v51-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: jhp1gjwn; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 348.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 348.8 20.9 26.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p8884v51",
   "source_url": "https://omlx.ai/benchmarks/performance/p8884v51",
   "tg_tokens": null,
   "tps": 20.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-p8884v51-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: jhp1gjwn; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 353.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 353.0 23.1 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p8884v51",
   "source_url": "https://omlx.ai/benchmarks/performance/p8884v51",
   "tg_tokens": null,
   "tps": 23.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-66p3wdul-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: y96r27cg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 348.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 348.9 23.6 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/66p3wdul",
   "source_url": "https://omlx.ai/benchmarks/performance/66p3wdul",
   "tg_tokens": null,
   "tps": 23.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-66p3wdul-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: y96r27cg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 353.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 353.6 24.1 27.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/66p3wdul",
   "source_url": "https://omlx.ai/benchmarks/performance/66p3wdul",
   "tg_tokens": null,
   "tps": 24.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-upx43v2k-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: 10az1ukl; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 345.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 345.7 21.7 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/upx43v2k",
   "source_url": "https://omlx.ai/benchmarks/performance/upx43v2k",
   "tg_tokens": null,
   "tps": 21.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-upx43v2k-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: 10az1ukl; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 351.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 351.2 21.6 27.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/upx43v2k",
   "source_url": "https://omlx.ai/benchmarks/performance/upx43v2k",
   "tg_tokens": null,
   "tps": 21.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-y2kld75q-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: n2gppdus; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 345.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 345.6 22.3 26.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y2kld75q",
   "source_url": "https://omlx.ai/benchmarks/performance/y2kld75q",
   "tg_tokens": null,
   "tps": 22.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-y2kld75q-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: n2gppdus; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 348.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 348.7 25.1 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y2kld75q",
   "source_url": "https://omlx.ai/benchmarks/performance/y2kld75q",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-unii8hm0-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: v1zmg490; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 345.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 345.3 12.4 22.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/unii8hm0",
   "source_url": "https://omlx.ai/benchmarks/performance/unii8hm0",
   "tg_tokens": null,
   "tps": 12.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-unii8hm0-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: v1zmg490; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 354.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 354.4 12.3 23.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/unii8hm0",
   "source_url": "https://omlx.ai/benchmarks/performance/unii8hm0",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-dt371nx5-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 69s7mkya; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 335.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 335.1 12.4 22.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dt371nx5",
   "source_url": "https://omlx.ai/benchmarks/performance/dt371nx5",
   "tg_tokens": null,
   "tps": 12.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-dt371nx5-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: 69s7mkya; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 336.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 336.5 12.3 23.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dt371nx5",
   "source_url": "https://omlx.ai/benchmarks/performance/dt371nx5",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-o8k9azio-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: hklfqpj0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 353.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 353.2 24.6 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o8k9azio",
   "source_url": "https://omlx.ai/benchmarks/performance/o8k9azio",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-o8k9azio-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: hklfqpj0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 355.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 355.0 25.3 27.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o8k9azio",
   "source_url": "https://omlx.ai/benchmarks/performance/o8k9azio",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-w8t4uk2i-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: qfad4t1p; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 344.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 344.3 20.9 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w8t4uk2i",
   "source_url": "https://omlx.ai/benchmarks/performance/w8t4uk2i",
   "tg_tokens": null,
   "tps": 20.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-w8t4uk2i-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: qfad4t1p; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 330.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 330.0 23.1 27.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w8t4uk2i",
   "source_url": "https://omlx.ai/benchmarks/performance/w8t4uk2i",
   "tg_tokens": null,
   "tps": 23.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0rkb5vl8-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 32.2 GB at this context; duplicate traces of the same session: sjl0geay; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 357.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 357.5 22.9 32.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0rkb5vl8",
   "source_url": "https://omlx.ai/benchmarks/performance/0rkb5vl8",
   "tg_tokens": null,
   "tps": 22.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0rkb5vl8-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 33.4 GB at this context; duplicate traces of the same session: sjl0geay; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 354.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 354.7 23.2 33.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0rkb5vl8",
   "source_url": "https://omlx.ai/benchmarks/performance/0rkb5vl8",
   "tg_tokens": null,
   "tps": 23.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-538wrth0-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: h2zfk0d2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 345.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 345.6 21.5 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/538wrth0",
   "source_url": "https://omlx.ai/benchmarks/performance/538wrth0",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-538wrth0-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: h2zfk0d2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 351.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 351.1 21.9 27.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/538wrth0",
   "source_url": "https://omlx.ai/benchmarks/performance/538wrth0",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-klh924c5-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: rvj3zqvf; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 344.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 344.7 23.0 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/klh924c5",
   "source_url": "https://omlx.ai/benchmarks/performance/klh924c5",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-klh924c5-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: rvj3zqvf; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 350.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 350.9 23.1 27.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/klh924c5",
   "source_url": "https://omlx.ai/benchmarks/performance/klh924c5",
   "tg_tokens": null,
   "tps": 23.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-lt8iqi87-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: gemuy911; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 345.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 345.0 22.2 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lt8iqi87",
   "source_url": "https://omlx.ai/benchmarks/performance/lt8iqi87",
   "tg_tokens": null,
   "tps": 22.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-lt8iqi87-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: gemuy911; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 351.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 351.7 25.1 27.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lt8iqi87",
   "source_url": "https://omlx.ai/benchmarks/performance/lt8iqi87",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-efxztid2-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: 49e1izzu; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 341.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 341.5 22.9 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/efxztid2",
   "source_url": "https://omlx.ai/benchmarks/performance/efxztid2",
   "tg_tokens": null,
   "tps": 22.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-efxztid2-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: 49e1izzu; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 347.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 347.7 25.0 27.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/efxztid2",
   "source_url": "https://omlx.ai/benchmarks/performance/efxztid2",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-tvgowxw9-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: uxvro8y6; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 338.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 338.1 23.7 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tvgowxw9",
   "source_url": "https://omlx.ai/benchmarks/performance/tvgowxw9",
   "tg_tokens": null,
   "tps": 23.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-tvgowxw9-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: uxvro8y6; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 337.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 337.9 21.1 27.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tvgowxw9",
   "source_url": "https://omlx.ai/benchmarks/performance/tvgowxw9",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-mssp2zri-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: a4zeuq75; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 340.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 340.4 23.5 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mssp2zri",
   "source_url": "https://omlx.ai/benchmarks/performance/mssp2zri",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-mssp2zri-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: a4zeuq75; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 344.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 344.8 23.9 27.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mssp2zri",
   "source_url": "https://omlx.ai/benchmarks/performance/mssp2zri",
   "tg_tokens": null,
   "tps": 23.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-sgd70l2d-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: l0pc4oq3; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 340.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 340.3 23.3 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sgd70l2d",
   "source_url": "https://omlx.ai/benchmarks/performance/sgd70l2d",
   "tg_tokens": null,
   "tps": 23.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-sgd70l2d-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: l0pc4oq3; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 345.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 345.6 24.8 27.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sgd70l2d",
   "source_url": "https://omlx.ai/benchmarks/performance/sgd70l2d",
   "tg_tokens": null,
   "tps": 24.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-y1aeunap-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 74.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2149.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,149 52.0 74.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y1aeunap",
   "source_url": "https://omlx.ai/benchmarks/performance/y1aeunap",
   "tg_tokens": null,
   "tps": 52.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-y1aeunap-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 75.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3101.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 3,101 54.9 75.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y1aeunap",
   "source_url": "https://omlx.ai/benchmarks/performance/y1aeunap",
   "tg_tokens": null,
   "tps": 54.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9sumthy1-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 74.7 GB at this context; duplicate traces of the same session: 2n0ken3m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2178.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,178 43.0 74.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9sumthy1",
   "source_url": "https://omlx.ai/benchmarks/performance/9sumthy1",
   "tg_tokens": null,
   "tps": 43.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9sumthy1-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 75.1 GB at this context; duplicate traces of the same session: 2n0ken3m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3104.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 3,104 57.2 75.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9sumthy1",
   "source_url": "https://omlx.ai/benchmarks/performance/9sumthy1",
   "tg_tokens": null,
   "tps": 57.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0lqfsjuy-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 74.7 GB; TTFT 466 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2199.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 2,199 PP tok/s 60.7 TG tok/s 466 TTFT (ms) 74.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0lqfsjuy",
   "source_url": "https://omlx.ai/benchmarks/performance/0lqfsjuy",
   "tg_tokens": null,
   "tps": 60.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nw8ourfz-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.7 GB at this context; duplicate traces of the same session: gfe345jv, zrvjw7z1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1050.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,050 101.3 100.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nw8ourfz",
   "source_url": "https://omlx.ai/benchmarks/performance/nw8ourfz",
   "tg_tokens": null,
   "tps": 101.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-nw8ourfz-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: gfe345jv, zrvjw7z1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1416.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,416 103.9 101.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nw8ourfz",
   "source_url": "https://omlx.ai/benchmarks/performance/nw8ourfz",
   "tg_tokens": null,
   "tps": 103.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nw8ourfz-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.9 GB at this context; duplicate traces of the same session: gfe345jv, zrvjw7z1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1394.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 1,394 85.0 101.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nw8ourfz",
   "source_url": "https://omlx.ai/benchmarks/performance/nw8ourfz",
   "tg_tokens": null,
   "tps": 85.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nw8ourfz-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 106.1 GB at this context; duplicate traces of the same session: gfe345jv, zrvjw7z1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1367.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 1,367 72.7 106.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nw8ourfz",
   "source_url": "https://omlx.ai/benchmarks/performance/nw8ourfz",
   "tg_tokens": null,
   "tps": 72.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nw8ourfz-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 112.4 GB at this context; duplicate traces of the same session: gfe345jv, zrvjw7z1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1339.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 1,339 56.0 112.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nw8ourfz",
   "source_url": "https://omlx.ai/benchmarks/performance/nw8ourfz",
   "tg_tokens": null,
   "tps": 56.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-21ys5uij-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 100.0 GB; TTFT 11584 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1414.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k tokens 1,414 PP tok/s 70.3 TG tok/s 11584 TTFT (ms) 100.0 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/21ys5uij",
   "source_url": "https://omlx.ai/benchmarks/performance/21ys5uij",
   "tg_tokens": null,
   "tps": 70.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-36oy84rm-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.7 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 320.3,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 320.3 34.0 11.7 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/36oy84rm",
   "source_url": "https://omlx.ai/benchmarks/performance/36oy84rm",
   "tg_tokens": null,
   "tps": 34.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-36oy84rm-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.7 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 336.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 336.0 32.1 11.7 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/36oy84rm",
   "source_url": "https://omlx.ai/benchmarks/performance/36oy84rm",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-3lm7x8si-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: w2czabdw; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 727.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 1k 727.1 90.7 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3lm7x8si",
   "source_url": "https://omlx.ai/benchmarks/performance/3lm7x8si",
   "tg_tokens": null,
   "tps": 90.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-3lm7x8si-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: w2czabdw; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 809.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k 809.4 88.0 21.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3lm7x8si",
   "source_url": "https://omlx.ai/benchmarks/performance/3lm7x8si",
   "tg_tokens": null,
   "tps": 88.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-aumdazpf-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 20.5 GB; TTFT 1390 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 736.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 1k tokens 736.9 PP tok/s 74.0 TG tok/s 1390 TTFT (ms) 20.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aumdazpf",
   "source_url": "https://omlx.ai/benchmarks/performance/aumdazpf",
   "tg_tokens": null,
   "tps": 74.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-jrni4608-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 1kalyp0b, 2lx10bbo, erowsioh, sb30edot; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1134.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,134 96.2 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jrni4608",
   "source_url": "https://omlx.ai/benchmarks/performance/jrni4608",
   "tg_tokens": null,
   "tps": 96.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-jrni4608-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 1kalyp0b, 2lx10bbo, erowsioh, sb30edot; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1919.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,919 119.8 23.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jrni4608",
   "source_url": "https://omlx.ai/benchmarks/performance/jrni4608",
   "tg_tokens": null,
   "tps": 119.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-jrni4608-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: 1kalyp0b, 2lx10bbo, erowsioh, sb30edot; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1807.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,807 28.3 24.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jrni4608",
   "source_url": "https://omlx.ai/benchmarks/performance/jrni4608",
   "tg_tokens": null,
   "tps": 28.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-jrni4608-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 1kalyp0b, 2lx10bbo, erowsioh, sb30edot; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1636.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,636 150.6 23.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jrni4608",
   "source_url": "https://omlx.ai/benchmarks/performance/jrni4608",
   "tg_tokens": null,
   "tps": 150.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-jrni4608-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 1kalyp0b, 2lx10bbo, erowsioh, sb30edot; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1237.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 1,237 190.1 23.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jrni4608",
   "source_url": "https://omlx.ai/benchmarks/performance/jrni4608",
   "tg_tokens": null,
   "tps": 190.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-swne4fn3-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 2tk952jy; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 392.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 392.2 33.1 21.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/swne4fn3",
   "source_url": "https://omlx.ai/benchmarks/performance/swne4fn3",
   "tg_tokens": null,
   "tps": 33.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-swne4fn3-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: 2tk952jy; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 374.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 374.2 39.2 22.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/swne4fn3",
   "source_url": "https://omlx.ai/benchmarks/performance/swne4fn3",
   "tg_tokens": null,
   "tps": 39.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-7xbojcvu-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: ybkj2g9g; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 498.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 498.1 35.5 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7xbojcvu",
   "source_url": "https://omlx.ai/benchmarks/performance/7xbojcvu",
   "tg_tokens": null,
   "tps": 35.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-7xbojcvu-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: ybkj2g9g; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 544.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 544.7 29.5 19.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7xbojcvu",
   "source_url": "https://omlx.ai/benchmarks/performance/7xbojcvu",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0lk46xn2-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: p33chp1b; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 502.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 502.2 32.2 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0lk46xn2",
   "source_url": "https://omlx.ai/benchmarks/performance/0lk46xn2",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0lk46xn2-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: p33chp1b; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 542.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 542.0 31.9 19.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0lk46xn2",
   "source_url": "https://omlx.ai/benchmarks/performance/0lk46xn2",
   "tg_tokens": null,
   "tps": 31.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3geuwo4w-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: vkzmcv2d; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 492.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 492.3 32.2 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3geuwo4w",
   "source_url": "https://omlx.ai/benchmarks/performance/3geuwo4w",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-3geuwo4w-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: vkzmcv2d; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 538.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 538.0 38.7 19.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3geuwo4w",
   "source_url": "https://omlx.ai/benchmarks/performance/3geuwo4w",
   "tg_tokens": null,
   "tps": 38.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-lji75wdb-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: d58tx9zh; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 509.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 509.9 32.6 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lji75wdb",
   "source_url": "https://omlx.ai/benchmarks/performance/lji75wdb",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-lji75wdb-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: d58tx9zh; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 552.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 552.0 35.8 19.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lji75wdb",
   "source_url": "https://omlx.ai/benchmarks/performance/lji75wdb",
   "tg_tokens": null,
   "tps": 35.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-lyioujjm-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: xz3dr1ax; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 511.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 511.2 34.2 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lyioujjm",
   "source_url": "https://omlx.ai/benchmarks/performance/lyioujjm",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-lyioujjm-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: xz3dr1ax; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 559.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 559.9 34.3 19.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lyioujjm",
   "source_url": "https://omlx.ai/benchmarks/performance/lyioujjm",
   "tg_tokens": null,
   "tps": 34.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-pgyzv6ft-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: nx5krqss; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 507.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 507.3 31.8 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pgyzv6ft",
   "source_url": "https://omlx.ai/benchmarks/performance/pgyzv6ft",
   "tg_tokens": null,
   "tps": 31.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-pgyzv6ft-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: nx5krqss; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 558.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 558.2 31.5 19.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pgyzv6ft",
   "source_url": "https://omlx.ai/benchmarks/performance/pgyzv6ft",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-dqh5obvv-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 2vhusy5i; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 507.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 507.0 31.0 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dqh5obvv",
   "source_url": "https://omlx.ai/benchmarks/performance/dqh5obvv",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-dqh5obvv-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 2vhusy5i; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 558.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 558.2 39.3 19.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dqh5obvv",
   "source_url": "https://omlx.ai/benchmarks/performance/dqh5obvv",
   "tg_tokens": null,
   "tps": 39.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-t8mn76gz-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: mswuhkn1; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 511.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 511.4 33.7 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t8mn76gz",
   "source_url": "https://omlx.ai/benchmarks/performance/t8mn76gz",
   "tg_tokens": null,
   "tps": 33.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-t8mn76gz-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: mswuhkn1; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 562.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 562.6 32.2 19.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t8mn76gz",
   "source_url": "https://omlx.ai/benchmarks/performance/t8mn76gz",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ni0d7wad-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace, single run at 4k context; peak memory 18.8 GB; TTFT 11099 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 369.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k tokens 369.1 PP tok/s 35.0 TG tok/s 11099 TTFT (ms) 18.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ni0d7wad",
   "source_url": "https://omlx.ai/benchmarks/performance/ni0d7wad",
   "tg_tokens": null,
   "tps": 35.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-gd4gjn0d-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: y252va4h; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 499.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 499.7 17.4 17.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gd4gjn0d",
   "source_url": "https://omlx.ai/benchmarks/performance/gd4gjn0d",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-gd4gjn0d-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: y252va4h; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 552.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 552.2 17.1 19.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gd4gjn0d",
   "source_url": "https://omlx.ai/benchmarks/performance/gd4gjn0d",
   "tg_tokens": null,
   "tps": 17.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-emc4a3kn-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: t8gxfqi0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 500.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 500.4 32.8 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/emc4a3kn",
   "source_url": "https://omlx.ai/benchmarks/performance/emc4a3kn",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-emc4a3kn-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: t8gxfqi0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 541.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 541.0 34.5 19.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/emc4a3kn",
   "source_url": "https://omlx.ai/benchmarks/performance/emc4a3kn",
   "tg_tokens": null,
   "tps": 34.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-jhufdp07-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: zxefzzk9; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 392.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 392.5 33.3 21.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jhufdp07",
   "source_url": "https://omlx.ai/benchmarks/performance/jhufdp07",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-jhufdp07-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: zxefzzk9; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 378.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 378.4 40.0 22.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jhufdp07",
   "source_url": "https://omlx.ai/benchmarks/performance/jhufdp07",
   "tg_tokens": null,
   "tps": 40.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0965jiqk-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace, single run at 1k context; peak memory 17.6 GB; TTFT 2062 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 496.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k tokens 496.7 PP tok/s 17.4 TG tok/s 2062 TTFT (ms) 17.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0965jiqk",
   "source_url": "https://omlx.ai/benchmarks/performance/0965jiqk",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1u0f24vb-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: dtwovbb9; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 498.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 498.0 32.2 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1u0f24vb",
   "source_url": "https://omlx.ai/benchmarks/performance/1u0f24vb",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-1u0f24vb-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: dtwovbb9; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 545.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 545.6 38.2 19.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1u0f24vb",
   "source_url": "https://omlx.ai/benchmarks/performance/1u0f24vb",
   "tg_tokens": null,
   "tps": 38.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1d9hg9x5-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: cvf7021o; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 507.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 507.7 30.2 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1d9hg9x5",
   "source_url": "https://omlx.ai/benchmarks/performance/1d9hg9x5",
   "tg_tokens": null,
   "tps": 30.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1d9hg9x5-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: cvf7021o; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 549.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 549.9 30.2 19.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1d9hg9x5",
   "source_url": "https://omlx.ai/benchmarks/performance/1d9hg9x5",
   "tg_tokens": null,
   "tps": 30.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-e7ds2l7t-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 49knccy2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 509.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 509.7 33.3 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e7ds2l7t",
   "source_url": "https://omlx.ai/benchmarks/performance/e7ds2l7t",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-e7ds2l7t-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 49knccy2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 558.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 558.0 36.7 19.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e7ds2l7t",
   "source_url": "https://omlx.ai/benchmarks/performance/e7ds2l7t",
   "tg_tokens": null,
   "tps": 36.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qon3x19s-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: hukikthy; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 504.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 504.4 33.2 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qon3x19s",
   "source_url": "https://omlx.ai/benchmarks/performance/qon3x19s",
   "tg_tokens": null,
   "tps": 33.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qon3x19s-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: hukikthy; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 554.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 554.7 33.9 19.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qon3x19s",
   "source_url": "https://omlx.ai/benchmarks/performance/qon3x19s",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-nz32extl-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: qmytuk46; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 512.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 512.0 28.9 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nz32extl",
   "source_url": "https://omlx.ai/benchmarks/performance/nz32extl",
   "tg_tokens": null,
   "tps": 28.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-nz32extl-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: qmytuk46; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 556.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 556.9 31.1 19.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nz32extl",
   "source_url": "https://omlx.ai/benchmarks/performance/nz32extl",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-b720xjv6-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 367.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 367.7 31.0 17.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b720xjv6",
   "source_url": "https://omlx.ai/benchmarks/performance/b720xjv6",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-b720xjv6-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 363.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 363.1 38.9 18.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b720xjv6",
   "source_url": "https://omlx.ai/benchmarks/performance/b720xjv6",
   "tg_tokens": null,
   "tps": 38.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-c7mx7rt4-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: x1lajsxf; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 507.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 507.2 31.6 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c7mx7rt4",
   "source_url": "https://omlx.ai/benchmarks/performance/c7mx7rt4",
   "tg_tokens": null,
   "tps": 31.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-c7mx7rt4-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: x1lajsxf; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 556.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 556.4 29.8 19.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c7mx7rt4",
   "source_url": "https://omlx.ai/benchmarks/performance/c7mx7rt4",
   "tg_tokens": null,
   "tps": 29.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-esmwbdh8-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 499.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 499.8 32.8 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/esmwbdh8",
   "source_url": "https://omlx.ai/benchmarks/performance/esmwbdh8",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-esmwbdh8-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 548.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 548.6 28.9 19.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/esmwbdh8",
   "source_url": "https://omlx.ai/benchmarks/performance/esmwbdh8",
   "tg_tokens": null,
   "tps": 28.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-f2gv3fvs-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: qi8jr483; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 127.8 27.9 16.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f2gv3fvs",
   "source_url": "https://omlx.ai/benchmarks/performance/f2gv3fvs",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-f2gv3fvs-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.0 GB at this context; duplicate traces of the same session: qi8jr483; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 131.5 27.0 18.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f2gv3fvs",
   "source_url": "https://omlx.ai/benchmarks/performance/f2gv3fvs",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-tq528r6c-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: sml8p2tb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 125.8 22.0 16.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tq528r6c",
   "source_url": "https://omlx.ai/benchmarks/performance/tq528r6c",
   "tg_tokens": null,
   "tps": 22.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-tq528r6c-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.0 GB at this context; duplicate traces of the same session: sml8p2tb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 131.3 26.9 18.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tq528r6c",
   "source_url": "https://omlx.ai/benchmarks/performance/tq528r6c",
   "tg_tokens": null,
   "tps": 26.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-oz7jdnu6-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: 6wn2a4zz, 9e6bjrez, vlldje1e; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1152.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,152 27.0 101.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oz7jdnu6",
   "source_url": "https://omlx.ai/benchmarks/performance/oz7jdnu6",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-oz7jdnu6-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.9 GB at this context; duplicate traces of the same session: 6wn2a4zz, 9e6bjrez, vlldje1e; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1386.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 1,386 83.1 101.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oz7jdnu6",
   "source_url": "https://omlx.ai/benchmarks/performance/oz7jdnu6",
   "tg_tokens": null,
   "tps": 83.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-oz7jdnu6-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 106.1 GB at this context; duplicate traces of the same session: 6wn2a4zz, 9e6bjrez, vlldje1e; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1356.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 1,356 47.3 106.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oz7jdnu6",
   "source_url": "https://omlx.ai/benchmarks/performance/oz7jdnu6",
   "tg_tokens": null,
   "tps": 47.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-oz7jdnu6-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 114.8 GB at this context; duplicate traces of the same session: 6wn2a4zz, 9e6bjrez, vlldje1e; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1320.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 195k 1,320 47.9 114.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oz7jdnu6",
   "source_url": "https://omlx.ai/benchmarks/performance/oz7jdnu6",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1yvzoyo7-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.3 GB at this context; duplicate traces of the same session: c4e2900a, oyvd70j5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2307.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,307 75.4 73.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1yvzoyo7",
   "source_url": "https://omlx.ai/benchmarks/performance/1yvzoyo7",
   "tg_tokens": null,
   "tps": 75.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1yvzoyo7-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.5 GB at this context; duplicate traces of the same session: c4e2900a, oyvd70j5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2330.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,330 83.2 74.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1yvzoyo7",
   "source_url": "https://omlx.ai/benchmarks/performance/1yvzoyo7",
   "tg_tokens": null,
   "tps": 83.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1yvzoyo7-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.9 GB at this context; duplicate traces of the same session: c4e2900a, oyvd70j5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2231.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,231 71.7 78.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1yvzoyo7",
   "source_url": "https://omlx.ai/benchmarks/performance/1yvzoyo7",
   "tg_tokens": null,
   "tps": 71.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-boiqbzpu-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.7 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 320.3,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 320.3 34.0 11.7 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/boiqbzpu",
   "source_url": "https://omlx.ai/benchmarks/performance/boiqbzpu",
   "tg_tokens": null,
   "tps": 34.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-boiqbzpu-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.7 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 336.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 336.0 32.1 11.7 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/boiqbzpu",
   "source_url": "https://omlx.ai/benchmarks/performance/boiqbzpu",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-e6n14r59-ctx1024",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 9.8 GB at this context; duplicate traces of the same session: 0dp6h0eq, 0fd5238n, 2wdiu1hg, aaagktws, f7e4izqy, fbwuzy2z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1859.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2-12B-A2.5B-Instruct-6bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 1,859 118.5 9.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e6n14r59",
   "source_url": "https://omlx.ai/benchmarks/performance/e6n14r59",
   "tg_tokens": null,
   "tps": 118.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-e6n14r59-ctx4096",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 9.9 GB at this context; duplicate traces of the same session: 0dp6h0eq, 0fd5238n, 2wdiu1hg, aaagktws, f7e4izqy, fbwuzy2z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1925.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2-12B-A2.5B-Instruct-6bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 1,925 113.1 9.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e6n14r59",
   "source_url": "https://omlx.ai/benchmarks/performance/e6n14r59",
   "tg_tokens": null,
   "tps": 113.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-e6n14r59-ctx8192",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 10.0 GB at this context; duplicate traces of the same session: 0dp6h0eq, 0fd5238n, 2wdiu1hg, aaagktws, f7e4izqy, fbwuzy2z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1874.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2-12B-A2.5B-Instruct-6bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 1,874 106.0 10.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e6n14r59",
   "source_url": "https://omlx.ai/benchmarks/performance/e6n14r59",
   "tg_tokens": null,
   "tps": 106.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-e6n14r59-ctx16384",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 10.1 GB at this context; duplicate traces of the same session: 0dp6h0eq, 0fd5238n, 2wdiu1hg, aaagktws, f7e4izqy, fbwuzy2z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1697.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2-12B-A2.5B-Instruct-6bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 1,697 95.1 10.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e6n14r59",
   "source_url": "https://omlx.ai/benchmarks/performance/e6n14r59",
   "tg_tokens": null,
   "tps": 95.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-e6n14r59-ctx32768",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 10.3 GB at this context; duplicate traces of the same session: 0dp6h0eq, 0fd5238n, 2wdiu1hg, aaagktws, f7e4izqy, fbwuzy2z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1304.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2-12B-A2.5B-Instruct-6bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 1,304 76.5 10.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e6n14r59",
   "source_url": "https://omlx.ai/benchmarks/performance/e6n14r59",
   "tg_tokens": null,
   "tps": 76.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-e6n14r59-ctx65536",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 11.0 GB at this context; duplicate traces of the same session: 0dp6h0eq, 0fd5238n, 2wdiu1hg, aaagktws, f7e4izqy, fbwuzy2z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1099.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2-12B-A2.5B-Instruct-6bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k 1,099 53.9 11.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e6n14r59",
   "source_url": "https://omlx.ai/benchmarks/performance/e6n14r59",
   "tg_tokens": null,
   "tps": 53.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-e6n14r59-ctx131072",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: 0dp6h0eq, 0fd5238n, 2wdiu1hg, aaagktws, f7e4izqy, fbwuzy2z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 881.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2-12B-A2.5B-Instruct-6bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 128k 881.4 35.9 12.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e6n14r59",
   "source_url": "https://omlx.ai/benchmarks/performance/e6n14r59",
   "tg_tokens": null,
   "tps": 35.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-wzzllknh-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 8ssqjjgs; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 501.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 501.9 33.7 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wzzllknh",
   "source_url": "https://omlx.ai/benchmarks/performance/wzzllknh",
   "tg_tokens": null,
   "tps": 33.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-wzzllknh-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 8ssqjjgs; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 552.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 552.1 32.1 19.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wzzllknh",
   "source_url": "https://omlx.ai/benchmarks/performance/wzzllknh",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-wysl9u1d-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 499.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 499.8 32.8 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wysl9u1d",
   "source_url": "https://omlx.ai/benchmarks/performance/wysl9u1d",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-wysl9u1d-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 548.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 548.6 28.9 19.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wysl9u1d",
   "source_url": "https://omlx.ai/benchmarks/performance/wysl9u1d",
   "tg_tokens": null,
   "tps": 28.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ta8mocun-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: yii7a74b; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 509.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 509.6 35.9 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ta8mocun",
   "source_url": "https://omlx.ai/benchmarks/performance/ta8mocun",
   "tg_tokens": null,
   "tps": 35.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ta8mocun-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: yii7a74b; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 561.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 561.1 38.3 19.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ta8mocun",
   "source_url": "https://omlx.ai/benchmarks/performance/ta8mocun",
   "tg_tokens": null,
   "tps": 38.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-2t74k7r5-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: 7rs69g0r; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 393.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 393.4 33.1 17.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2t74k7r5",
   "source_url": "https://omlx.ai/benchmarks/performance/2t74k7r5",
   "tg_tokens": null,
   "tps": 33.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-2t74k7r5-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 7rs69g0r; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 387.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 387.4 33.5 18.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2t74k7r5",
   "source_url": "https://omlx.ai/benchmarks/performance/2t74k7r5",
   "tg_tokens": null,
   "tps": 33.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-h3kyleh0-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: ragxbpbt; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 509.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 509.0 35.1 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h3kyleh0",
   "source_url": "https://omlx.ai/benchmarks/performance/h3kyleh0",
   "tg_tokens": null,
   "tps": 35.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-h3kyleh0-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: ragxbpbt; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 562.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 562.3 35.0 19.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h3kyleh0",
   "source_url": "https://omlx.ai/benchmarks/performance/h3kyleh0",
   "tg_tokens": null,
   "tps": 35.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ai4npvzu-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: q1td5uwr; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 514.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 514.5 36.2 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ai4npvzu",
   "source_url": "https://omlx.ai/benchmarks/performance/ai4npvzu",
   "tg_tokens": null,
   "tps": 36.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ai4npvzu-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: q1td5uwr; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 562.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 562.8 35.9 19.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ai4npvzu",
   "source_url": "https://omlx.ai/benchmarks/performance/ai4npvzu",
   "tg_tokens": null,
   "tps": 35.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-6u13mosq-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: zf2so9of; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 505.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 505.8 33.8 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6u13mosq",
   "source_url": "https://omlx.ai/benchmarks/performance/6u13mosq",
   "tg_tokens": null,
   "tps": 33.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-6u13mosq-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: zf2so9of; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 562.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 562.1 34.0 19.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6u13mosq",
   "source_url": "https://omlx.ai/benchmarks/performance/6u13mosq",
   "tg_tokens": null,
   "tps": 34.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-brbw31oz-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 20a3o47z; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 510.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 510.0 33.0 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/brbw31oz",
   "source_url": "https://omlx.ai/benchmarks/performance/brbw31oz",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-brbw31oz-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 20a3o47z; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 562.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 562.5 34.3 19.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/brbw31oz",
   "source_url": "https://omlx.ai/benchmarks/performance/brbw31oz",
   "tg_tokens": null,
   "tps": 34.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-wzws9t2u-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: ae1d5rbx; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 501.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 501.0 34.8 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wzws9t2u",
   "source_url": "https://omlx.ai/benchmarks/performance/wzws9t2u",
   "tg_tokens": null,
   "tps": 34.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-wzws9t2u-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: ae1d5rbx; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 557.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 557.5 33.9 19.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wzws9t2u",
   "source_url": "https://omlx.ai/benchmarks/performance/wzws9t2u",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-h2jumlif-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: jyxjbj8k; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 386.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 386.0 35.8 17.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h2jumlif",
   "source_url": "https://omlx.ai/benchmarks/performance/h2jumlif",
   "tg_tokens": null,
   "tps": 35.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-h2jumlif-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: jyxjbj8k; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 375.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 375.9 36.0 18.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h2jumlif",
   "source_url": "https://omlx.ai/benchmarks/performance/h2jumlif",
   "tg_tokens": null,
   "tps": 36.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-5hinvni1-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 2o1kba9z; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 507.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 507.1 35.4 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5hinvni1",
   "source_url": "https://omlx.ai/benchmarks/performance/5hinvni1",
   "tg_tokens": null,
   "tps": 35.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-5hinvni1-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 2o1kba9z; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 547.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 547.0 34.9 19.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5hinvni1",
   "source_url": "https://omlx.ai/benchmarks/performance/5hinvni1",
   "tg_tokens": null,
   "tps": 34.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-1f6ylal6-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: a3zaguvb; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 392.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 392.5 37.6 17.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1f6ylal6",
   "source_url": "https://omlx.ai/benchmarks/performance/1f6ylal6",
   "tg_tokens": null,
   "tps": 37.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1f6ylal6-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: a3zaguvb; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 381.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 381.3 35.3 18.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1f6ylal6",
   "source_url": "https://omlx.ai/benchmarks/performance/1f6ylal6",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-j496st70-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: cqv9vcue; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 349.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 349.3 22.8 26.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j496st70",
   "source_url": "https://omlx.ai/benchmarks/performance/j496st70",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-j496st70-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: cqv9vcue; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 352.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 352.2 23.5 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j496st70",
   "source_url": "https://omlx.ai/benchmarks/performance/j496st70",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-63q7ro73-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: uwmmw8vi; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 348.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 348.7 29.7 26.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/63q7ro73",
   "source_url": "https://omlx.ai/benchmarks/performance/63q7ro73",
   "tg_tokens": null,
   "tps": 29.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-63q7ro73-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: uwmmw8vi; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 352.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 352.1 24.0 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/63q7ro73",
   "source_url": "https://omlx.ai/benchmarks/performance/63q7ro73",
   "tg_tokens": null,
   "tps": 24.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-00hdu9lz-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: 25c63or2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 348.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 348.6 29.3 26.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/00hdu9lz",
   "source_url": "https://omlx.ai/benchmarks/performance/00hdu9lz",
   "tg_tokens": null,
   "tps": 29.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-00hdu9lz-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: 25c63or2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 351.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 351.9 23.1 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/00hdu9lz",
   "source_url": "https://omlx.ai/benchmarks/performance/00hdu9lz",
   "tg_tokens": null,
   "tps": 23.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-90zn0eml-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: le5oq2oh; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 345.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 345.1 25.1 26.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/90zn0eml",
   "source_url": "https://omlx.ai/benchmarks/performance/90zn0eml",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-90zn0eml-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: le5oq2oh; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 351.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 351.9 25.5 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/90zn0eml",
   "source_url": "https://omlx.ai/benchmarks/performance/90zn0eml",
   "tg_tokens": null,
   "tps": 25.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-dzhlbajk-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: fq5fnz2k; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 349.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 349.0 25.3 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dzhlbajk",
   "source_url": "https://omlx.ai/benchmarks/performance/dzhlbajk",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-dzhlbajk-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: fq5fnz2k; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 354.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 354.6 25.5 27.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dzhlbajk",
   "source_url": "https://omlx.ai/benchmarks/performance/dzhlbajk",
   "tg_tokens": null,
   "tps": 25.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-lgl3pm71-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: zestx864; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 346.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 346.6 26.5 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lgl3pm71",
   "source_url": "https://omlx.ai/benchmarks/performance/lgl3pm71",
   "tg_tokens": null,
   "tps": 26.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-lgl3pm71-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: zestx864; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 352.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 352.2 24.0 27.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lgl3pm71",
   "source_url": "https://omlx.ai/benchmarks/performance/lgl3pm71",
   "tg_tokens": null,
   "tps": 24.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-9ebrbbwl-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: mu988gp0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 347.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 347.1 24.5 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ebrbbwl",
   "source_url": "https://omlx.ai/benchmarks/performance/9ebrbbwl",
   "tg_tokens": null,
   "tps": 24.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-9ebrbbwl-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: mu988gp0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 353.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 353.1 25.1 27.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ebrbbwl",
   "source_url": "https://omlx.ai/benchmarks/performance/9ebrbbwl",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-quqak575-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: tp0phldu; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 348.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 348.3 29.1 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/quqak575",
   "source_url": "https://omlx.ai/benchmarks/performance/quqak575",
   "tg_tokens": null,
   "tps": 29.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-quqak575-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: tp0phldu; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 354.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 354.6 23.6 27.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/quqak575",
   "source_url": "https://omlx.ai/benchmarks/performance/quqak575",
   "tg_tokens": null,
   "tps": 23.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-cmpvqt62-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: p7f3xuqu; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 353.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 353.1 12.4 22.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cmpvqt62",
   "source_url": "https://omlx.ai/benchmarks/performance/cmpvqt62",
   "tg_tokens": null,
   "tps": 12.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-cmpvqt62-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: p7f3xuqu; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 355.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 355.2 12.3 23.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cmpvqt62",
   "source_url": "https://omlx.ai/benchmarks/performance/cmpvqt62",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-33fjgkuf-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 355.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 355.3 23.7 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/33fjgkuf",
   "source_url": "https://omlx.ai/benchmarks/performance/33fjgkuf",
   "tg_tokens": null,
   "tps": 23.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-33fjgkuf-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 353.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 353.1 25.3 27.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/33fjgkuf",
   "source_url": "https://omlx.ai/benchmarks/performance/33fjgkuf",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-juc1eidz-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 81.5 GB at this context; duplicate traces of the same session: 3pjblq32; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2087.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,087 79.8 81.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/juc1eidz",
   "source_url": "https://omlx.ai/benchmarks/performance/juc1eidz",
   "tg_tokens": null,
   "tps": 79.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-juc1eidz-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 82.1 GB at this context; duplicate traces of the same session: 3pjblq32; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3011.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 3,011 74.2 82.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/juc1eidz",
   "source_url": "https://omlx.ai/benchmarks/performance/juc1eidz",
   "tg_tokens": null,
   "tps": 74.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-lq78qdtt-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 56.5 GB; TTFT 1986 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 515.6,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-Flash-Next-oQ3e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 3bit; 1k tokens 515.6 PP tok/s 28.8 TG tok/s 1986 TTFT (ms) 56.5 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lq78qdtt",
   "source_url": "https://omlx.ai/benchmarks/performance/lq78qdtt",
   "tg_tokens": null,
   "tps": 28.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-oq2ka6rt-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: 0shgaw3p, 5kgvbrrd, bejqv3p9, lo6ubwi7, t6g3uob3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 570.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 570.6 61.5 70.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oq2ka6rt",
   "source_url": "https://omlx.ai/benchmarks/performance/oq2ka6rt",
   "tg_tokens": null,
   "tps": 61.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-oq2ka6rt-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: 0shgaw3p, 5kgvbrrd, bejqv3p9, lo6ubwi7, t6g3uob3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 772.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 772.5 55.4 71.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oq2ka6rt",
   "source_url": "https://omlx.ai/benchmarks/performance/oq2ka6rt",
   "tg_tokens": null,
   "tps": 55.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-oq2ka6rt-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: 0shgaw3p, 5kgvbrrd, bejqv3p9, lo6ubwi7, t6g3uob3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1059.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 1,059 64.6 71.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oq2ka6rt",
   "source_url": "https://omlx.ai/benchmarks/performance/oq2ka6rt",
   "tg_tokens": null,
   "tps": 64.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-oq2ka6rt-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.1 GB at this context; duplicate traces of the same session: 0shgaw3p, 5kgvbrrd, bejqv3p9, lo6ubwi7, t6g3uob3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1050.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 16k 1,050 37.7 72.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oq2ka6rt",
   "source_url": "https://omlx.ai/benchmarks/performance/oq2ka6rt",
   "tg_tokens": null,
   "tps": 37.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-oq2ka6rt-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 76.5 GB at this context; duplicate traces of the same session: 0shgaw3p, 5kgvbrrd, bejqv3p9, lo6ubwi7, t6g3uob3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1023.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k 1,023 34.6 76.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oq2ka6rt",
   "source_url": "https://omlx.ai/benchmarks/performance/oq2ka6rt",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-oq2ka6rt-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 82.6 GB at this context; duplicate traces of the same session: 0shgaw3p, 5kgvbrrd, bejqv3p9, lo6ubwi7, t6g3uob3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 944.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 128k 944.1 41.1 82.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oq2ka6rt",
   "source_url": "https://omlx.ai/benchmarks/performance/oq2ka6rt",
   "tg_tokens": null,
   "tps": 41.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5rzwtunj-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.5 GB at this context; duplicate traces of the same session: p5wjzewn; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 311.5,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 311.5 41.3 39.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5rzwtunj",
   "source_url": "https://omlx.ai/benchmarks/performance/5rzwtunj",
   "tg_tokens": null,
   "tps": 41.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5rzwtunj-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 42.8 GB at this context; duplicate traces of the same session: p5wjzewn; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 416.3,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 416.3 39.2 42.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5rzwtunj",
   "source_url": "https://omlx.ai/benchmarks/performance/5rzwtunj",
   "tg_tokens": null,
   "tps": 39.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5rzwtunj-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 43.5 GB at this context; duplicate traces of the same session: p5wjzewn; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 429.7,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 8k 429.7 54.9 43.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5rzwtunj",
   "source_url": "https://omlx.ai/benchmarks/performance/5rzwtunj",
   "tg_tokens": null,
   "tps": 54.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-q77t2mot-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 174.9 GB at this context; duplicate traces of the same session: 2osegkrq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 410.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 410.8 40.2 174.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q77t2mot",
   "source_url": "https://omlx.ai/benchmarks/performance/q77t2mot",
   "tg_tokens": null,
   "tps": 40.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-q77t2mot-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 175.5 GB at this context; duplicate traces of the same session: 2osegkrq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 404.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 404.2 39.7 175.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q77t2mot",
   "source_url": "https://omlx.ai/benchmarks/performance/q77t2mot",
   "tg_tokens": null,
   "tps": 39.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-khaefnio-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.2 GB at this context; duplicate traces of the same session: bjzgcpi6, mx8kvofh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2423.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 2,423 76.9 28.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/khaefnio",
   "source_url": "https://omlx.ai/benchmarks/performance/khaefnio",
   "tg_tokens": null,
   "tps": 76.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-khaefnio-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.4 GB at this context; duplicate traces of the same session: bjzgcpi6, mx8kvofh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2471.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 16k 2,471 72.8 28.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/khaefnio",
   "source_url": "https://omlx.ai/benchmarks/performance/khaefnio",
   "tg_tokens": null,
   "tps": 72.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-khaefnio-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 29.3 GB at this context; duplicate traces of the same session: bjzgcpi6, mx8kvofh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1700.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 64k 1,700 54.0 29.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/khaefnio",
   "source_url": "https://omlx.ai/benchmarks/performance/khaefnio",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-k7fj6d8d-ctx4096",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 67.5 GB; TTFT 2228 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1839.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.5-35B-A3B-bf16 on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 bf16; 4k tokens 1,839 PP tok/s 48.7 TG tok/s 2228 TTFT (ms) 67.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k7fj6d8d",
   "source_url": "https://omlx.ai/benchmarks/performance/k7fj6d8d",
   "tg_tokens": null,
   "tps": 48.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-w9uux452-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: vugofocx; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 553.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 553.4 61.0 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w9uux452",
   "source_url": "https://omlx.ai/benchmarks/performance/w9uux452",
   "tg_tokens": null,
   "tps": 61.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-w9uux452-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: vugofocx; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 550.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 550.3 63.8 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w9uux452",
   "source_url": "https://omlx.ai/benchmarks/performance/w9uux452",
   "tg_tokens": null,
   "tps": 63.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qvf1bkxi-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 224 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4582.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 4,582 PP tok/s 152.9 TG tok/s 224 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qvf1bkxi",
   "source_url": "https://omlx.ai/benchmarks/performance/qvf1bkxi",
   "tg_tokens": null,
   "tps": 152.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-tgez2s7i-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: twsblcah; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 586.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 586.2 68.0 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tgez2s7i",
   "source_url": "https://omlx.ai/benchmarks/performance/tgez2s7i",
   "tg_tokens": null,
   "tps": 68.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-tgez2s7i-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: twsblcah; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 562.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 562.0 63.9 21.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tgez2s7i",
   "source_url": "https://omlx.ai/benchmarks/performance/tgez2s7i",
   "tg_tokens": null,
   "tps": 63.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mz15noon-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: 64enmuwy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 324.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 324.8 6.5 29.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mz15noon",
   "source_url": "https://omlx.ai/benchmarks/performance/mz15noon",
   "tg_tokens": null,
   "tps": 6.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mz15noon-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: 64enmuwy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 316.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 316.3 6.1 30.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mz15noon",
   "source_url": "https://omlx.ai/benchmarks/performance/mz15noon",
   "tg_tokens": null,
   "tps": 6.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rqrm5hrt-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 823.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 823.0 19.0 28.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rqrm5hrt",
   "source_url": "https://omlx.ai/benchmarks/performance/rqrm5hrt",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rqrm5hrt-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 950.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 950.7 18.8 29.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rqrm5hrt",
   "source_url": "https://omlx.ai/benchmarks/performance/rqrm5hrt",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-iwybogzd-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: 80cez85y; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1257.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 4k 1,257 31.8 29.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iwybogzd",
   "source_url": "https://omlx.ai/benchmarks/performance/iwybogzd",
   "tg_tokens": null,
   "tps": 31.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-iwybogzd-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: 80cez85y; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1320.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 16k 1,320 30.6 30.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iwybogzd",
   "source_url": "https://omlx.ai/benchmarks/performance/iwybogzd",
   "tg_tokens": null,
   "tps": 30.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qwwim9pr-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 26vagpib; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 352.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 352.0 12.4 22.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qwwim9pr",
   "source_url": "https://omlx.ai/benchmarks/performance/qwwim9pr",
   "tg_tokens": null,
   "tps": 12.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qwwim9pr-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: 26vagpib; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 352.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 352.7 12.3 23.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qwwim9pr",
   "source_url": "https://omlx.ai/benchmarks/performance/qwwim9pr",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-upqcitj6-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: p04dtgff; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 346.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 346.4 20.3 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/upqcitj6",
   "source_url": "https://omlx.ai/benchmarks/performance/upqcitj6",
   "tg_tokens": null,
   "tps": 20.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-upqcitj6-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: p04dtgff; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 352.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 352.8 25.1 27.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/upqcitj6",
   "source_url": "https://omlx.ai/benchmarks/performance/upqcitj6",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ybn18ged-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: 3ifx6p6y; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 348.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 348.6 24.2 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ybn18ged",
   "source_url": "https://omlx.ai/benchmarks/performance/ybn18ged",
   "tg_tokens": null,
   "tps": 24.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ybn18ged-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: 3ifx6p6y; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 354.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 354.3 25.0 27.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ybn18ged",
   "source_url": "https://omlx.ai/benchmarks/performance/ybn18ged",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ajxafb5k-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 32.4 GB at this context; duplicate traces of the same session: zr1wnli9; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 362.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 362.0 21.9 32.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ajxafb5k",
   "source_url": "https://omlx.ai/benchmarks/performance/ajxafb5k",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ajxafb5k-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 33.4 GB at this context; duplicate traces of the same session: zr1wnli9; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 355.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 355.0 24.5 33.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ajxafb5k",
   "source_url": "https://omlx.ai/benchmarks/performance/ajxafb5k",
   "tg_tokens": null,
   "tps": 24.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-e19374jq-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: 8zweayr5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 348.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 348.8 24.2 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e19374jq",
   "source_url": "https://omlx.ai/benchmarks/performance/e19374jq",
   "tg_tokens": null,
   "tps": 24.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-e19374jq-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: 8zweayr5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 354.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 354.5 25.3 27.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e19374jq",
   "source_url": "https://omlx.ai/benchmarks/performance/e19374jq",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-21mbldc1-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 355.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 355.3 23.7 26.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/21mbldc1",
   "source_url": "https://omlx.ai/benchmarks/performance/21mbldc1",
   "tg_tokens": null,
   "tps": 23.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-21mbldc1-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 353.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 353.1 25.3 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/21mbldc1",
   "source_url": "https://omlx.ai/benchmarks/performance/21mbldc1",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-bllb1a57-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: gpwb44jr; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 349.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 349.1 24.5 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bllb1a57",
   "source_url": "https://omlx.ai/benchmarks/performance/bllb1a57",
   "tg_tokens": null,
   "tps": 24.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-bllb1a57-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: gpwb44jr; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 353.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 353.9 24.1 27.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bllb1a57",
   "source_url": "https://omlx.ai/benchmarks/performance/bllb1a57",
   "tg_tokens": null,
   "tps": 24.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-xhib7b1z-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: ogs26cqd; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 346.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 346.7 24.5 26.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xhib7b1z",
   "source_url": "https://omlx.ai/benchmarks/performance/xhib7b1z",
   "tg_tokens": null,
   "tps": 24.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-xhib7b1z-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: ogs26cqd; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 352.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 352.9 24.6 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xhib7b1z",
   "source_url": "https://omlx.ai/benchmarks/performance/xhib7b1z",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1i09z4ud-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: 0ymjfefw; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 344.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 344.0 22.1 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1i09z4ud",
   "source_url": "https://omlx.ai/benchmarks/performance/1i09z4ud",
   "tg_tokens": null,
   "tps": 22.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-1i09z4ud-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: 0ymjfefw; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 350.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 350.9 25.6 27.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1i09z4ud",
   "source_url": "https://omlx.ai/benchmarks/performance/1i09z4ud",
   "tg_tokens": null,
   "tps": 25.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-83j2b6kz-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: elgz6o3l, y5thovfb; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 353.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 353.9 24.2 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/83j2b6kz",
   "source_url": "https://omlx.ai/benchmarks/performance/83j2b6kz",
   "tg_tokens": null,
   "tps": 24.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-83j2b6kz-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: elgz6o3l, y5thovfb; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 352.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 352.3 24.2 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/83j2b6kz",
   "source_url": "https://omlx.ai/benchmarks/performance/83j2b6kz",
   "tg_tokens": null,
   "tps": 24.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-83j2b6kz-ctx8192",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: elgz6o3l, y5thovfb; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 352.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 8k 352.6 26.7 27.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/83j2b6kz",
   "source_url": "https://omlx.ai/benchmarks/performance/83j2b6kz",
   "tg_tokens": null,
   "tps": 26.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0a7utmn4-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: ru4b0vvu, xxd9oi93; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 353.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 353.3 25.9 26.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0a7utmn4",
   "source_url": "https://omlx.ai/benchmarks/performance/0a7utmn4",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0a7utmn4-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: ru4b0vvu, xxd9oi93; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 353.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 353.4 24.3 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0a7utmn4",
   "source_url": "https://omlx.ai/benchmarks/performance/0a7utmn4",
   "tg_tokens": null,
   "tps": 24.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-0a7utmn4-ctx8192",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: ru4b0vvu, xxd9oi93; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 354.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 8k 354.2 28.9 27.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0a7utmn4",
   "source_url": "https://omlx.ai/benchmarks/performance/0a7utmn4",
   "tg_tokens": null,
   "tps": 28.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-so726y57-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: 45lng5gb, cs35p820; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 354.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 354.7 23.7 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/so726y57",
   "source_url": "https://omlx.ai/benchmarks/performance/so726y57",
   "tg_tokens": null,
   "tps": 23.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-so726y57-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: 45lng5gb, cs35p820; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 353.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 353.3 23.3 27.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/so726y57",
   "source_url": "https://omlx.ai/benchmarks/performance/so726y57",
   "tg_tokens": null,
   "tps": 23.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-so726y57-ctx8192",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: 45lng5gb, cs35p820; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 353.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 8k 353.5 27.1 27.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/so726y57",
   "source_url": "https://omlx.ai/benchmarks/performance/so726y57",
   "tg_tokens": null,
   "tps": 27.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-gi38nwf7-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: 0rba7lvb, 8dblrcwq; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 345.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 345.1 25.0 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gi38nwf7",
   "source_url": "https://omlx.ai/benchmarks/performance/gi38nwf7",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-gi38nwf7-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: 0rba7lvb, 8dblrcwq; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 348.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 348.2 23.0 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gi38nwf7",
   "source_url": "https://omlx.ai/benchmarks/performance/gi38nwf7",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-gi38nwf7-ctx8192",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: 0rba7lvb, 8dblrcwq; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 355.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 8k 355.3 25.9 27.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gi38nwf7",
   "source_url": "https://omlx.ai/benchmarks/performance/gi38nwf7",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-9aiciyh2-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: ddgnjggy, omf1zg9p; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 361.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 361.9 25.8 26.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9aiciyh2",
   "source_url": "https://omlx.ai/benchmarks/performance/9aiciyh2",
   "tg_tokens": null,
   "tps": 25.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-9aiciyh2-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: ddgnjggy, omf1zg9p; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 384.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 384.6 23.5 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9aiciyh2",
   "source_url": "https://omlx.ai/benchmarks/performance/9aiciyh2",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-9aiciyh2-ctx8192",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: ddgnjggy, omf1zg9p; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 343.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 8k 343.9 28.5 27.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9aiciyh2",
   "source_url": "https://omlx.ai/benchmarks/performance/9aiciyh2",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-7q54ws67-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: qa634p4q; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 79.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 79.6 14.6 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7q54ws67",
   "source_url": "https://omlx.ai/benchmarks/performance/7q54ws67",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-7q54ws67-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: qa634p4q; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 77.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 77.6 15.1 20.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7q54ws67",
   "source_url": "https://omlx.ai/benchmarks/performance/7q54ws67",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fjyrzm2t-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 75.1 GB; TTFT 1316 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3113.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 3,113 PP tok/s 54.3 TG tok/s 1316 TTFT (ms) 75.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fjyrzm2t",
   "source_url": "https://omlx.ai/benchmarks/performance/fjyrzm2t",
   "tg_tokens": null,
   "tps": 54.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-h8zr2m6r-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 82.2 GB; TTFT 1140 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3592.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 3,592 PP tok/s 74.8 TG tok/s 1140 TTFT (ms) 82.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h8zr2m6r",
   "source_url": "https://omlx.ai/benchmarks/performance/h8zr2m6r",
   "tg_tokens": null,
   "tps": 74.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-vgpx4kek-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 81.5 GB at this context; duplicate traces of the same session: 4e40f6mf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2091.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,091 77.9 81.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vgpx4kek",
   "source_url": "https://omlx.ai/benchmarks/performance/vgpx4kek",
   "tg_tokens": null,
   "tps": 77.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-vgpx4kek-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 82.1 GB at this context; duplicate traces of the same session: 4e40f6mf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3017.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 3,017 74.3 82.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vgpx4kek",
   "source_url": "https://omlx.ai/benchmarks/performance/vgpx4kek",
   "tg_tokens": null,
   "tps": 74.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-k8o0pxcn-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 81.5 GB at this context; duplicate traces of the same session: 37ri8qco; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2089.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,089 81.6 81.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k8o0pxcn",
   "source_url": "https://omlx.ai/benchmarks/performance/k8o0pxcn",
   "tg_tokens": null,
   "tps": 81.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-k8o0pxcn-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 82.1 GB at this context; duplicate traces of the same session: 37ri8qco; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3008.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 3,008 76.4 82.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k8o0pxcn",
   "source_url": "https://omlx.ai/benchmarks/performance/k8o0pxcn",
   "tg_tokens": null,
   "tps": 76.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-4sxeszcx-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 81.8 GB; TTFT 132261 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 991.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 128k tokens 991.0 PP tok/s 42.0 TG tok/s 132261 TTFT (ms) 81.8 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4sxeszcx",
   "source_url": "https://omlx.ai/benchmarks/performance/4sxeszcx",
   "tg_tokens": null,
   "tps": 42.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dqckgjlc-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.7 GB at this context; duplicate traces of the same session: tj4we4xt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2669.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,669 67.5 74.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dqckgjlc",
   "source_url": "https://omlx.ai/benchmarks/performance/dqckgjlc",
   "tg_tokens": null,
   "tps": 67.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dqckgjlc-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.2 GB at this context; duplicate traces of the same session: tj4we4xt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2524.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,524 65.2 79.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dqckgjlc",
   "source_url": "https://omlx.ai/benchmarks/performance/dqckgjlc",
   "tg_tokens": null,
   "tps": 65.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-no7clqsj-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.5 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 311.5,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 311.5 41.3 39.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/no7clqsj",
   "source_url": "https://omlx.ai/benchmarks/performance/no7clqsj",
   "tg_tokens": null,
   "tps": 41.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-no7clqsj-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 42.8 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 416.3,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 416.3 39.2 42.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/no7clqsj",
   "source_url": "https://omlx.ai/benchmarks/performance/no7clqsj",
   "tg_tokens": null,
   "tps": 39.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-no7clqsj-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 43.5 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 429.7,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 8k 429.7 54.9 43.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/no7clqsj",
   "source_url": "https://omlx.ai/benchmarks/performance/no7clqsj",
   "tg_tokens": null,
   "tps": 54.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-y3hgbh2f-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: sj2w5qbu; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1524.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 1,524 58.0 19.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y3hgbh2f",
   "source_url": "https://omlx.ai/benchmarks/performance/y3hgbh2f",
   "tg_tokens": null,
   "tps": 58.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-y3hgbh2f-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: sj2w5qbu; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1334.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 16k 1,334 44.5 20.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y3hgbh2f",
   "source_url": "https://omlx.ai/benchmarks/performance/y3hgbh2f",
   "tg_tokens": null,
   "tps": 44.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-i0i2z9n3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: w51ihbql; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 252.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 252.4 28.8 16.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i0i2z9n3",
   "source_url": "https://omlx.ai/benchmarks/performance/i0i2z9n3",
   "tg_tokens": null,
   "tps": 28.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-i0i2z9n3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: w51ihbql; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 267.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 267.6 27.8 19.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i0i2z9n3",
   "source_url": "https://omlx.ai/benchmarks/performance/i0i2z9n3",
   "tg_tokens": null,
   "tps": 27.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hcvt67jj-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 823.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 823.0 19.0 28.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hcvt67jj",
   "source_url": "https://omlx.ai/benchmarks/performance/hcvt67jj",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hcvt67jj-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 950.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 950.7 18.8 29.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hcvt67jj",
   "source_url": "https://omlx.ai/benchmarks/performance/hcvt67jj",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-b752iim4-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: vd4enouj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 509.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 509.2 34.9 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b752iim4",
   "source_url": "https://omlx.ai/benchmarks/performance/b752iim4",
   "tg_tokens": null,
   "tps": 34.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-b752iim4-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: vd4enouj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 562.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 562.4 32.2 19.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b752iim4",
   "source_url": "https://omlx.ai/benchmarks/performance/b752iim4",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-41wvj3wk-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: ftr63xlk; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 504.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 504.7 34.1 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/41wvj3wk",
   "source_url": "https://omlx.ai/benchmarks/performance/41wvj3wk",
   "tg_tokens": null,
   "tps": 34.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-41wvj3wk-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: ftr63xlk; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 561.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 561.9 33.4 19.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/41wvj3wk",
   "source_url": "https://omlx.ai/benchmarks/performance/41wvj3wk",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-k5nr58nn-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: vlrpc38o; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 387.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 387.4 37.9 17.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k5nr58nn",
   "source_url": "https://omlx.ai/benchmarks/performance/k5nr58nn",
   "tg_tokens": null,
   "tps": 37.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-k5nr58nn-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: vlrpc38o; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 376.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 376.8 33.7 18.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k5nr58nn",
   "source_url": "https://omlx.ai/benchmarks/performance/k5nr58nn",
   "tg_tokens": null,
   "tps": 33.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ryaisjys-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: bm88k5ub; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 505.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 505.9 36.0 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ryaisjys",
   "source_url": "https://omlx.ai/benchmarks/performance/ryaisjys",
   "tg_tokens": null,
   "tps": 36.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ryaisjys-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: bm88k5ub; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 563.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 563.3 37.6 19.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ryaisjys",
   "source_url": "https://omlx.ai/benchmarks/performance/ryaisjys",
   "tg_tokens": null,
   "tps": 37.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-tps34j7w-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace, single run at 1k context; peak memory 17.8 GB; TTFT 1989 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 514.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k tokens 514.7 PP tok/s 30.1 TG tok/s 1989 TTFT (ms) 17.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tps34j7w",
   "source_url": "https://omlx.ai/benchmarks/performance/tps34j7w",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-rpr6t0f5-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace, single run at 4k context; peak memory 19.0 GB; TTFT 7021 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 583.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k tokens 583.4 PP tok/s 32.9 TG tok/s 7021 TTFT (ms) 19.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rpr6t0f5",
   "source_url": "https://omlx.ai/benchmarks/performance/rpr6t0f5",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-b5z2u9xq-ctx16384",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace, single run at 16k context; peak memory 19.8 GB; TTFT 28586 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 573.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 16k tokens 573.2 PP tok/s 40.7 TG tok/s 28586 TTFT (ms) 19.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b5z2u9xq",
   "source_url": "https://omlx.ai/benchmarks/performance/b5z2u9xq",
   "tg_tokens": null,
   "tps": 40.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-5l3w5i49-ctx65536",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace, single run at 64k context; peak memory 23.1 GB; TTFT 151349 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 433.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 64k tokens 433.0 PP tok/s 22.3 TG tok/s 151349 TTFT (ms) 23.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5l3w5i49",
   "source_url": "https://omlx.ai/benchmarks/performance/5l3w5i49",
   "tg_tokens": null,
   "tps": 22.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-584qmqar-ctx32768",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace, single run at 32k context; peak memory 20.9 GB; TTFT 63192 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 518.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 32k tokens 518.5 PP tok/s 32.1 TG tok/s 63192 TTFT (ms) 20.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/584qmqar",
   "source_url": "https://omlx.ai/benchmarks/performance/584qmqar",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-9h1u3309-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 32.4 GB at this context; duplicate traces of the same session: vewqxqj3; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 361.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 361.1 22.7 32.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9h1u3309",
   "source_url": "https://omlx.ai/benchmarks/performance/9h1u3309",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-9h1u3309-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 33.4 GB at this context; duplicate traces of the same session: vewqxqj3; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 345.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 345.4 25.3 33.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9h1u3309",
   "source_url": "https://omlx.ai/benchmarks/performance/9h1u3309",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-a2xojkt4-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: 2p6osxkp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 348.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 348.3 26.5 26.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a2xojkt4",
   "source_url": "https://omlx.ai/benchmarks/performance/a2xojkt4",
   "tg_tokens": null,
   "tps": 26.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-a2xojkt4-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: 2p6osxkp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 351.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 351.4 26.1 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a2xojkt4",
   "source_url": "https://omlx.ai/benchmarks/performance/a2xojkt4",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-wnimv6vj-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: nk84w168; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 345.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 345.1 24.6 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wnimv6vj",
   "source_url": "https://omlx.ai/benchmarks/performance/wnimv6vj",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-wnimv6vj-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: nk84w168; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 351.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 351.8 23.6 27.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wnimv6vj",
   "source_url": "https://omlx.ai/benchmarks/performance/wnimv6vj",
   "tg_tokens": null,
   "tps": 23.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-16jpwg0y-ctx65536",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace, single run at 64k context; peak memory 31.8 GB; TTFT 226639 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 289.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 64k tokens 289.2 PP tok/s 21.2 TG tok/s 226639 TTFT (ms) 31.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/16jpwg0y",
   "source_url": "https://omlx.ai/benchmarks/performance/16jpwg0y",
   "tg_tokens": null,
   "tps": 21.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-kgwe9nls-ctx16384",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace, single run at 16k context; peak memory 28.5 GB; TTFT 48241 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 339.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 16k tokens 339.6 PP tok/s 25.1 TG tok/s 48241 TTFT (ms) 28.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kgwe9nls",
   "source_url": "https://omlx.ai/benchmarks/performance/kgwe9nls",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-1sitm6ut-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace, single run at 4k context; peak memory 27.7 GB; TTFT 11012 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 371.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k tokens 371.9 PP tok/s 27.5 TG tok/s 11012 TTFT (ms) 27.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1sitm6ut",
   "source_url": "https://omlx.ai/benchmarks/performance/1sitm6ut",
   "tg_tokens": null,
   "tps": 27.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-e4uw7goa-ctx32768",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace, single run at 32k context; peak memory 29.5 GB; TTFT 100951 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 324.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 32k tokens 324.6 PP tok/s 24.9 TG tok/s 100951 TTFT (ms) 29.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e4uw7goa",
   "source_url": "https://omlx.ai/benchmarks/performance/e4uw7goa",
   "tg_tokens": null,
   "tps": 24.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-r1pn0dyp-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: 9ovn1v71, nw7t08o5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 349.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 349.9 22.7 26.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r1pn0dyp",
   "source_url": "https://omlx.ai/benchmarks/performance/r1pn0dyp",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-r1pn0dyp-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: 9ovn1v71, nw7t08o5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 355.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 355.2 24.2 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r1pn0dyp",
   "source_url": "https://omlx.ai/benchmarks/performance/r1pn0dyp",
   "tg_tokens": null,
   "tps": 24.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-r1pn0dyp-ctx8192",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: 9ovn1v71, nw7t08o5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 354.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 8k 354.7 26.4 27.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r1pn0dyp",
   "source_url": "https://omlx.ai/benchmarks/performance/r1pn0dyp",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-izj28r4r-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace, single run at 1k context; peak memory 26.6 GB; TTFT 2846 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 359.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k tokens 359.8 PP tok/s 24.6 TG tok/s 2846 TTFT (ms) 26.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/izj28r4r",
   "source_url": "https://omlx.ai/benchmarks/performance/izj28r4r",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0y9cworg-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 349.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 349.2 24.8 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0y9cworg",
   "source_url": "https://omlx.ai/benchmarks/performance/0y9cworg",
   "tg_tokens": null,
   "tps": 24.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0y9cworg-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 351.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 351.1 23.6 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0y9cworg",
   "source_url": "https://omlx.ai/benchmarks/performance/0y9cworg",
   "tg_tokens": null,
   "tps": 23.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-0y9cworg-ctx16384",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 339.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 16k 339.6 26.6 28.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0y9cworg",
   "source_url": "https://omlx.ai/benchmarks/performance/0y9cworg",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-heyrhrkv-ctx1024",
   "model": "Qwen3.8-3.6-27B-blend",
   "notes": "blend of Qwen3.8 and Qwen3.6 27B; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 09lw1y7m, 695mb14b, b63fe55u, ebyvwdq0, jpfzi0zn, uqrhi2pw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 228.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-3.6-27B-blend-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 228.3 36.2 19.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/heyrhrkv",
   "source_url": "https://omlx.ai/benchmarks/performance/heyrhrkv",
   "tg_tokens": null,
   "tps": 36.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-heyrhrkv-ctx4096",
   "model": "Qwen3.8-3.6-27B-blend",
   "notes": "blend of Qwen3.8 and Qwen3.6 27B; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: 09lw1y7m, 695mb14b, b63fe55u, ebyvwdq0, jpfzi0zn, uqrhi2pw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-3.6-27B-blend-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 220.0 31.6 22.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/heyrhrkv",
   "source_url": "https://omlx.ai/benchmarks/performance/heyrhrkv",
   "tg_tokens": null,
   "tps": 31.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-heyrhrkv-ctx8192",
   "model": "Qwen3.8-3.6-27B-blend",
   "notes": "blend of Qwen3.8 and Qwen3.6 27B; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 09lw1y7m, 695mb14b, b63fe55u, ebyvwdq0, jpfzi0zn, uqrhi2pw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 188.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-3.6-27B-blend-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 188.0 21.2 23.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/heyrhrkv",
   "source_url": "https://omlx.ai/benchmarks/performance/heyrhrkv",
   "tg_tokens": null,
   "tps": 21.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-heyrhrkv-ctx16384",
   "model": "Qwen3.8-3.6-27B-blend",
   "notes": "blend of Qwen3.8 and Qwen3.6 27B; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: 09lw1y7m, 695mb14b, b63fe55u, ebyvwdq0, jpfzi0zn, uqrhi2pw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 197.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-3.6-27B-blend-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 197.6 28.2 23.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/heyrhrkv",
   "source_url": "https://omlx.ai/benchmarks/performance/heyrhrkv",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-heyrhrkv-ctx32768",
   "model": "Qwen3.8-3.6-27B-blend",
   "notes": "blend of Qwen3.8 and Qwen3.6 27B; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: 09lw1y7m, 695mb14b, b63fe55u, ebyvwdq0, jpfzi0zn, uqrhi2pw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 191.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-3.6-27B-blend-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 191.4 26.7 24.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/heyrhrkv",
   "source_url": "https://omlx.ai/benchmarks/performance/heyrhrkv",
   "tg_tokens": null,
   "tps": 26.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-heyrhrkv-ctx65536",
   "model": "Qwen3.8-3.6-27B-blend",
   "notes": "blend of Qwen3.8 and Qwen3.6 27B; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: 09lw1y7m, 695mb14b, b63fe55u, ebyvwdq0, jpfzi0zn, uqrhi2pw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 170.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-3.6-27B-blend-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 170.0 26.5 26.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/heyrhrkv",
   "source_url": "https://omlx.ai/benchmarks/performance/heyrhrkv",
   "tg_tokens": null,
   "tps": 26.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-heyrhrkv-ctx131072",
   "model": "Qwen3.8-3.6-27B-blend",
   "notes": "blend of Qwen3.8 and Qwen3.6 27B; oMLX community benchmark trace; peak memory 30.1 GB at this context; duplicate traces of the same session: 09lw1y7m, 695mb14b, b63fe55u, ebyvwdq0, jpfzi0zn, uqrhi2pw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 134.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-3.6-27B-blend-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 134.2 22.5 30.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/heyrhrkv",
   "source_url": "https://omlx.ai/benchmarks/performance/heyrhrkv",
   "tg_tokens": null,
   "tps": 22.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-os1itg4j-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 40.5 GB at this context; duplicate traces of the same session: f7pm1qhq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1294.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,294 76.1 40.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/os1itg4j",
   "source_url": "https://omlx.ai/benchmarks/performance/os1itg4j",
   "tg_tokens": null,
   "tps": 76.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-os1itg4j-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 40.8 GB at this context; duplicate traces of the same session: f7pm1qhq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2258.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,258 74.7 40.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/os1itg4j",
   "source_url": "https://omlx.ai/benchmarks/performance/os1itg4j",
   "tg_tokens": null,
   "tps": 74.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9vizmxsj-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 40.5 GB at this context; duplicate traces of the same session: 01dfx6vr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1003.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,003 67.8 40.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9vizmxsj",
   "source_url": "https://omlx.ai/benchmarks/performance/9vizmxsj",
   "tg_tokens": null,
   "tps": 67.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9vizmxsj-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 40.8 GB at this context; duplicate traces of the same session: 01dfx6vr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2090.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,090 68.0 40.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9vizmxsj",
   "source_url": "https://omlx.ai/benchmarks/performance/9vizmxsj",
   "tg_tokens": null,
   "tps": 68.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-eic8omce-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 81.5 GB at this context; duplicate traces of the same session: cd4y1fr9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2083.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,083 77.8 81.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eic8omce",
   "source_url": "https://omlx.ai/benchmarks/performance/eic8omce",
   "tg_tokens": null,
   "tps": 77.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-eic8omce-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 82.1 GB at this context; duplicate traces of the same session: cd4y1fr9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3009.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 3,009 74.3 82.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eic8omce",
   "source_url": "https://omlx.ai/benchmarks/performance/eic8omce",
   "tg_tokens": null,
   "tps": 74.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-xjaom1ia-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 81.5 GB at this context; duplicate traces of the same session: d33zzkep; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2080.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,080 78.0 81.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xjaom1ia",
   "source_url": "https://omlx.ai/benchmarks/performance/xjaom1ia",
   "tg_tokens": null,
   "tps": 78.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-xjaom1ia-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 82.1 GB at this context; duplicate traces of the same session: d33zzkep; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3013.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 3,013 74.6 82.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xjaom1ia",
   "source_url": "https://omlx.ai/benchmarks/performance/xjaom1ia",
   "tg_tokens": null,
   "tps": 74.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8sw5h0td-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 81.5 GB; TTFT 492 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2083.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 2,083 PP tok/s 78.5 TG tok/s 492 TTFT (ms) 81.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8sw5h0td",
   "source_url": "https://omlx.ai/benchmarks/performance/8sw5h0td",
   "tg_tokens": null,
   "tps": 78.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xxpf6wap-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 86.6 GB at this context; duplicate traces of the same session: 15vo7d91; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2138.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,138 60.6 86.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xxpf6wap",
   "source_url": "https://omlx.ai/benchmarks/performance/xxpf6wap",
   "tg_tokens": null,
   "tps": 60.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xxpf6wap-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 87.0 GB at this context; duplicate traces of the same session: 15vo7d91; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3043.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 3,043 53.0 87.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xxpf6wap",
   "source_url": "https://omlx.ai/benchmarks/performance/xxpf6wap",
   "tg_tokens": null,
   "tps": 53.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-v4cyajt4-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 86.6 GB; TTFT 478 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2144.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 2,144 PP tok/s 61.1 TG tok/s 478 TTFT (ms) 86.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v4cyajt4",
   "source_url": "https://omlx.ai/benchmarks/performance/v4cyajt4",
   "tg_tokens": null,
   "tps": 61.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-44z8iwla-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 86.7 GB; TTFT 481 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2129.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 2,129 PP tok/s 66.9 TG tok/s 481 TTFT (ms) 86.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/44z8iwla",
   "source_url": "https://omlx.ai/benchmarks/performance/44z8iwla",
   "tg_tokens": null,
   "tps": 66.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-e3uyxapu-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 86.6 GB at this context; duplicate traces of the same session: x5eeuc6m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2143.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,143 66.3 86.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e3uyxapu",
   "source_url": "https://omlx.ai/benchmarks/performance/e3uyxapu",
   "tg_tokens": null,
   "tps": 66.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-e3uyxapu-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 87.0 GB at this context; duplicate traces of the same session: x5eeuc6m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3042.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 3,042 59.3 87.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e3uyxapu",
   "source_url": "https://omlx.ai/benchmarks/performance/e3uyxapu",
   "tg_tokens": null,
   "tps": 59.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-yc98niy9-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 50.3 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 510.3,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 2bit; 1k 510.3 45.0 50.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yc98niy9",
   "source_url": "https://omlx.ai/benchmarks/performance/yc98niy9",
   "tg_tokens": null,
   "tps": 45.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-yc98niy9-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 51.4 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1240.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 2bit; 8k 1,240 49.1 51.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yc98niy9",
   "source_url": "https://omlx.ai/benchmarks/performance/yc98niy9",
   "tg_tokens": null,
   "tps": 49.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-gq06gaqw-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 4.5 GB at this context; duplicate traces of the same session: ac8kd2sp; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 629.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 629.1 30.9 4.5 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gq06gaqw",
   "source_url": "https://omlx.ai/benchmarks/performance/gq06gaqw",
   "tg_tokens": null,
   "tps": 30.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-gq06gaqw-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 4.6 GB at this context; duplicate traces of the same session: ac8kd2sp; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 654.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 654.5 30.5 4.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gq06gaqw",
   "source_url": "https://omlx.ai/benchmarks/performance/gq06gaqw",
   "tg_tokens": null,
   "tps": 30.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-n2zs24zj-ctx1024",
   "model": "Humanizer-12B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 12.8 GB at this context; duplicate traces of the same session: zhommt33; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1138.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "humanizer-12b-mlx-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 1,138 39.1 12.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n2zs24zj",
   "source_url": "https://omlx.ai/benchmarks/performance/n2zs24zj",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-n2zs24zj-ctx4096",
   "model": "Humanizer-12B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 13.2 GB at this context; duplicate traces of the same session: zhommt33; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1537.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "humanizer-12b-mlx-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 1,537 37.8 13.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n2zs24zj",
   "source_url": "https://omlx.ai/benchmarks/performance/n2zs24zj",
   "tg_tokens": null,
   "tps": 37.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-4uhv23hz-ctx65536",
   "model": "DeepSeek-Coder-V2-Lite-Instruct",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 34.1 GB; TTFT 99843 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 656.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-Coder-V2-Lite-Instruct-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k tokens 656.4 PP tok/s 23.3 TG tok/s 99843 TTFT (ms) 34.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4uhv23hz",
   "source_url": "https://omlx.ai/benchmarks/performance/4uhv23hz",
   "tg_tokens": null,
   "tps": 23.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-h90wc61u-ctx1024",
   "model": "DeepSeek-Coder-V2-Lite-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 9.6 GB at this context; duplicate traces of the same session: vm7jnri3, yeuasenf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2297.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-Coder-V2-Lite-Instruct-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 2,297 148.7 9.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h90wc61u",
   "source_url": "https://omlx.ai/benchmarks/performance/h90wc61u",
   "tg_tokens": null,
   "tps": 148.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-h90wc61u-ctx4096",
   "model": "DeepSeek-Coder-V2-Lite-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 10.6 GB at this context; duplicate traces of the same session: vm7jnri3, yeuasenf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2146.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-Coder-V2-Lite-Instruct-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 2,146 118.7 10.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h90wc61u",
   "source_url": "https://omlx.ai/benchmarks/performance/h90wc61u",
   "tg_tokens": null,
   "tps": 118.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-h90wc61u-ctx8192",
   "model": "DeepSeek-Coder-V2-Lite-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 12.2 GB at this context; duplicate traces of the same session: vm7jnri3, yeuasenf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1866.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-Coder-V2-Lite-Instruct-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,866 91.5 12.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h90wc61u",
   "source_url": "https://omlx.ai/benchmarks/performance/h90wc61u",
   "tg_tokens": null,
   "tps": 91.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-m36jqeoq-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: l3a3nhe1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4878.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,878 133.8 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m36jqeoq",
   "source_url": "https://omlx.ai/benchmarks/performance/m36jqeoq",
   "tg_tokens": null,
   "tps": 133.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-m36jqeoq-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: l3a3nhe1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6531.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,531 125.3 20.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m36jqeoq",
   "source_url": "https://omlx.ai/benchmarks/performance/m36jqeoq",
   "tg_tokens": null,
   "tps": 125.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hbij07uv-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: dd3jzsnf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4809.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,809 132.3 19.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hbij07uv",
   "source_url": "https://omlx.ai/benchmarks/performance/hbij07uv",
   "tg_tokens": null,
   "tps": 132.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hbij07uv-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: dd3jzsnf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6517.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,517 125.2 20.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hbij07uv",
   "source_url": "https://omlx.ai/benchmarks/performance/hbij07uv",
   "tg_tokens": null,
   "tps": 125.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-shwty58x-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 5x9wg58u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4845.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,845 132.6 19.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/shwty58x",
   "source_url": "https://omlx.ai/benchmarks/performance/shwty58x",
   "tg_tokens": null,
   "tps": 132.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-shwty58x-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 5x9wg58u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6525.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,525 125.7 20.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/shwty58x",
   "source_url": "https://omlx.ai/benchmarks/performance/shwty58x",
   "tg_tokens": null,
   "tps": 125.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i9vhj2at-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: v1j6jg9c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4779.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,779 130.5 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i9vhj2at",
   "source_url": "https://omlx.ai/benchmarks/performance/i9vhj2at",
   "tg_tokens": null,
   "tps": 130.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i9vhj2at-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: v1j6jg9c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6521.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,521 123.8 20.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i9vhj2at",
   "source_url": "https://omlx.ai/benchmarks/performance/i9vhj2at",
   "tg_tokens": null,
   "tps": 123.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-0l5qpfr2-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 11vp1obp; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 1k 205.5 29.7 16.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0l5qpfr2",
   "source_url": "https://omlx.ai/benchmarks/performance/0l5qpfr2",
   "tg_tokens": null,
   "tps": 29.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-0l5qpfr2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: 11vp1obp; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 194.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 4k 194.5 28.7 19.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0l5qpfr2",
   "source_url": "https://omlx.ai/benchmarks/performance/0l5qpfr2",
   "tg_tokens": null,
   "tps": 28.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-044mr070-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: i12fjzke; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 863.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 863.5 33.0 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/044mr070",
   "source_url": "https://omlx.ai/benchmarks/performance/044mr070",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-044mr070-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: i12fjzke; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 949.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 949.9 33.0 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/044mr070",
   "source_url": "https://omlx.ai/benchmarks/performance/044mr070",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ntohftsx-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: bdi6lqm2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 349.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 349.2 24.8 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ntohftsx",
   "source_url": "https://omlx.ai/benchmarks/performance/ntohftsx",
   "tg_tokens": null,
   "tps": 24.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ntohftsx-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: bdi6lqm2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 351.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 351.1 23.6 27.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ntohftsx",
   "source_url": "https://omlx.ai/benchmarks/performance/ntohftsx",
   "tg_tokens": null,
   "tps": 23.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ntohftsx-ctx16384",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: bdi6lqm2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 339.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 16k 339.6 26.6 28.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ntohftsx",
   "source_url": "https://omlx.ai/benchmarks/performance/ntohftsx",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-k3dkmii0-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 31.0 GB at this context; duplicate traces of the same session: 1acf6cvu, 2idlqdhw; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 362.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 362.9 22.2 31.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k3dkmii0",
   "source_url": "https://omlx.ai/benchmarks/performance/k3dkmii0",
   "tg_tokens": null,
   "tps": 22.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-k3dkmii0-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 31.9 GB at this context; duplicate traces of the same session: 1acf6cvu, 2idlqdhw; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 413.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 413.2 23.8 31.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k3dkmii0",
   "source_url": "https://omlx.ai/benchmarks/performance/k3dkmii0",
   "tg_tokens": null,
   "tps": 23.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-k3dkmii0-ctx16384",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 32.7 GB at this context; duplicate traces of the same session: 1acf6cvu, 2idlqdhw; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 377.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 16k 377.1 27.7 32.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k3dkmii0",
   "source_url": "https://omlx.ai/benchmarks/performance/k3dkmii0",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-h5cbn20c-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 31.0 GB at this context; duplicate traces of the same session: 33oqoo7a, z82z2qrs; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 361.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 361.9 21.9 31.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h5cbn20c",
   "source_url": "https://omlx.ai/benchmarks/performance/h5cbn20c",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-h5cbn20c-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 31.9 GB at this context; duplicate traces of the same session: 33oqoo7a, z82z2qrs; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 411.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 411.9 24.3 31.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h5cbn20c",
   "source_url": "https://omlx.ai/benchmarks/performance/h5cbn20c",
   "tg_tokens": null,
   "tps": 24.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-h5cbn20c-ctx16384",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 32.8 GB at this context; duplicate traces of the same session: 33oqoo7a, z82z2qrs; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 389.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 16k 389.3 28.7 32.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h5cbn20c",
   "source_url": "https://omlx.ai/benchmarks/performance/h5cbn20c",
   "tg_tokens": null,
   "tps": 28.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-51uzf2wj-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: 5ckkzdye, 9k3fa7ym; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 362.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 362.8 22.6 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/51uzf2wj",
   "source_url": "https://omlx.ai/benchmarks/performance/51uzf2wj",
   "tg_tokens": null,
   "tps": 22.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-51uzf2wj-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: 5ckkzdye, 9k3fa7ym; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 385.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 385.4 24.8 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/51uzf2wj",
   "source_url": "https://omlx.ai/benchmarks/performance/51uzf2wj",
   "tg_tokens": null,
   "tps": 24.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-51uzf2wj-ctx16384",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 5ckkzdye, 9k3fa7ym; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 341.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 16k 341.0 27.5 28.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/51uzf2wj",
   "source_url": "https://omlx.ai/benchmarks/performance/51uzf2wj",
   "tg_tokens": null,
   "tps": 27.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-sne7umvj-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: lmxk23b3; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 350.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 350.2 25.1 26.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sne7umvj",
   "source_url": "https://omlx.ai/benchmarks/performance/sne7umvj",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-sne7umvj-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: lmxk23b3; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 351.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 351.6 26.6 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sne7umvj",
   "source_url": "https://omlx.ai/benchmarks/performance/sne7umvj",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-7ejn2q1f-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: och0e1ao; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 348.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 348.3 28.4 26.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ejn2q1f",
   "source_url": "https://omlx.ai/benchmarks/performance/7ejn2q1f",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-7ejn2q1f-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: och0e1ao; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 348.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 348.9 23.5 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ejn2q1f",
   "source_url": "https://omlx.ai/benchmarks/performance/7ejn2q1f",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qq0vhdjb-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 32.4 GB at this context; duplicate traces of the same session: grfweza6; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 362.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 362.5 24.1 32.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qq0vhdjb",
   "source_url": "https://omlx.ai/benchmarks/performance/qq0vhdjb",
   "tg_tokens": null,
   "tps": 24.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qq0vhdjb-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 33.4 GB at this context; duplicate traces of the same session: grfweza6; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 344.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 344.4 24.6 33.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qq0vhdjb",
   "source_url": "https://omlx.ai/benchmarks/performance/qq0vhdjb",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-iw3wjo2q-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 348.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 348.3 24.6 26.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iw3wjo2q",
   "source_url": "https://omlx.ai/benchmarks/performance/iw3wjo2q",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-iw3wjo2q-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 350.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 350.8 23.9 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iw3wjo2q",
   "source_url": "https://omlx.ai/benchmarks/performance/iw3wjo2q",
   "tg_tokens": null,
   "tps": 23.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-vznjhdzv-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 347.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 347.8 26.9 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vznjhdzv",
   "source_url": "https://omlx.ai/benchmarks/performance/vznjhdzv",
   "tg_tokens": null,
   "tps": 26.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-vznjhdzv-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 346.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 346.2 24.8 27.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vznjhdzv",
   "source_url": "https://omlx.ai/benchmarks/performance/vznjhdzv",
   "tg_tokens": null,
   "tps": 24.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-vamls70o-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: 0o1cmc6f, 35vb2k28, 6t7rv4g6, sws7el66; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 426.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 426.7 71.7 21.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vamls70o",
   "source_url": "https://omlx.ai/benchmarks/performance/vamls70o",
   "tg_tokens": null,
   "tps": 71.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-vamls70o-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: 0o1cmc6f, 35vb2k28, 6t7rv4g6, sws7el66; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 512.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 512.5 75.6 25.0 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vamls70o",
   "source_url": "https://omlx.ai/benchmarks/performance/vamls70o",
   "tg_tokens": null,
   "tps": 75.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-vamls70o-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: 0o1cmc6f, 35vb2k28, 6t7rv4g6, sws7el66; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 565.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 565.7 97.7 25.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vamls70o",
   "source_url": "https://omlx.ai/benchmarks/performance/vamls70o",
   "tg_tokens": null,
   "tps": 97.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-vamls70o-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.2 GB at this context; duplicate traces of the same session: 0o1cmc6f, 35vb2k28, 6t7rv4g6, sws7el66; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 539.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 539.4 82.7 26.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vamls70o",
   "source_url": "https://omlx.ai/benchmarks/performance/vamls70o",
   "tg_tokens": null,
   "tps": 82.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vamls70o-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: 0o1cmc6f, 35vb2k28, 6t7rv4g6, sws7el66; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 491.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 491.9 164.5 27.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vamls70o",
   "source_url": "https://omlx.ai/benchmarks/performance/vamls70o",
   "tg_tokens": null,
   "tps": 164.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-u2naw9ud-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 54.4 GB at this context; duplicate traces of the same session: oqgsen4p, wf93zofj; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 487.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 487.7 46.1 54.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u2naw9ud",
   "source_url": "https://omlx.ai/benchmarks/performance/u2naw9ud",
   "tg_tokens": null,
   "tps": 46.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-u2naw9ud-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 54.8 GB at this context; duplicate traces of the same session: oqgsen4p, wf93zofj; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 474.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 8k 474.9 52.3 54.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u2naw9ud",
   "source_url": "https://omlx.ai/benchmarks/performance/u2naw9ud",
   "tg_tokens": null,
   "tps": 52.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-u2naw9ud-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 56.1 GB at this context; duplicate traces of the same session: oqgsen4p, wf93zofj; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 449.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 16k 449.7 63.1 56.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u2naw9ud",
   "source_url": "https://omlx.ai/benchmarks/performance/u2naw9ud",
   "tg_tokens": null,
   "tps": 63.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-f6jdtgee-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 43.4 GB at this context; duplicate traces of the same session: 52pu7t29, wu33jqj6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 371.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 371.1 30.9 43.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f6jdtgee",
   "source_url": "https://omlx.ai/benchmarks/performance/f6jdtgee",
   "tg_tokens": null,
   "tps": 30.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-f6jdtgee-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 43.9 GB at this context; duplicate traces of the same session: 52pu7t29, wu33jqj6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 393.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 8k 393.3 39.9 43.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f6jdtgee",
   "source_url": "https://omlx.ai/benchmarks/performance/f6jdtgee",
   "tg_tokens": null,
   "tps": 39.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-f6jdtgee-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 45.1 GB at this context; duplicate traces of the same session: 52pu7t29, wu33jqj6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 393.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 16k 393.3 36.9 45.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f6jdtgee",
   "source_url": "https://omlx.ai/benchmarks/performance/f6jdtgee",
   "tg_tokens": null,
   "tps": 36.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gxy5kkoo-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 54.4 GB at this context; duplicate traces of the same session: 2bb7ejsx, muuj9oji; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 490.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 490.1 49.1 54.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gxy5kkoo",
   "source_url": "https://omlx.ai/benchmarks/performance/gxy5kkoo",
   "tg_tokens": null,
   "tps": 49.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gxy5kkoo-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 54.8 GB at this context; duplicate traces of the same session: 2bb7ejsx, muuj9oji; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 485.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 8k 485.5 51.6 54.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gxy5kkoo",
   "source_url": "https://omlx.ai/benchmarks/performance/gxy5kkoo",
   "tg_tokens": null,
   "tps": 51.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gxy5kkoo-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 56.1 GB at this context; duplicate traces of the same session: 2bb7ejsx, muuj9oji; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 457.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 16k 457.5 62.3 56.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gxy5kkoo",
   "source_url": "https://omlx.ai/benchmarks/performance/gxy5kkoo",
   "tg_tokens": null,
   "tps": 62.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zu1jdij3-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 86.6 GB at this context; duplicate traces of the same session: 6hc69ljc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2145.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,145 71.9 86.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zu1jdij3",
   "source_url": "https://omlx.ai/benchmarks/performance/zu1jdij3",
   "tg_tokens": null,
   "tps": 71.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zu1jdij3-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 87.0 GB at this context; duplicate traces of the same session: 6hc69ljc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3039.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 3,039 57.6 87.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zu1jdij3",
   "source_url": "https://omlx.ai/benchmarks/performance/zu1jdij3",
   "tg_tokens": null,
   "tps": 57.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-2k004v27-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 50.3 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 510.3,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 2bit; 1k 510.3 45.0 50.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2k004v27",
   "source_url": "https://omlx.ai/benchmarks/performance/2k004v27",
   "tg_tokens": null,
   "tps": 45.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-2k004v27-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 51.4 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1240.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 2bit; 8k 1,240 49.1 51.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2k004v27",
   "source_url": "https://omlx.ai/benchmarks/performance/2k004v27",
   "tg_tokens": null,
   "tps": 49.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gr5j8sfz-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.7 GB at this context; duplicate traces of the same session: 1gmggulo, rdv42voh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2585.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,585 97.1 72.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gr5j8sfz",
   "source_url": "https://omlx.ai/benchmarks/performance/gr5j8sfz",
   "tg_tokens": null,
   "tps": 97.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-gr5j8sfz-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.0 GB at this context; duplicate traces of the same session: 1gmggulo, rdv42voh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2423.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,423 72.4 79.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gr5j8sfz",
   "source_url": "https://omlx.ai/benchmarks/performance/gr5j8sfz",
   "tg_tokens": null,
   "tps": 72.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gr5j8sfz-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.8 GB at this context; duplicate traces of the same session: 1gmggulo, rdv42voh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2325.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,325 60.4 80.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gr5j8sfz",
   "source_url": "https://omlx.ai/benchmarks/performance/gr5j8sfz",
   "tg_tokens": null,
   "tps": 60.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qqj2lxjy-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.7 GB at this context; duplicate traces of the same session: zfccpyes; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2491.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,491 112.7 72.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qqj2lxjy",
   "source_url": "https://omlx.ai/benchmarks/performance/qqj2lxjy",
   "tg_tokens": null,
   "tps": 112.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qqj2lxjy-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.5 GB at this context; duplicate traces of the same session: zfccpyes; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2429.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,429 55.1 78.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qqj2lxjy",
   "source_url": "https://omlx.ai/benchmarks/performance/qqj2lxjy",
   "tg_tokens": null,
   "tps": 55.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-r0agocsu-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: dya5tidc, f8lnwu80; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2324.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,324 59.4 71.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r0agocsu",
   "source_url": "https://omlx.ai/benchmarks/performance/r0agocsu",
   "tg_tokens": null,
   "tps": 59.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-r0agocsu-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.7 GB at this context; duplicate traces of the same session: dya5tidc, f8lnwu80; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2509.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,509 95.5 72.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r0agocsu",
   "source_url": "https://omlx.ai/benchmarks/performance/r0agocsu",
   "tg_tokens": null,
   "tps": 95.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-r0agocsu-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.4 GB at this context; duplicate traces of the same session: dya5tidc, f8lnwu80; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2409.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,409 66.4 78.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r0agocsu",
   "source_url": "https://omlx.ai/benchmarks/performance/r0agocsu",
   "tg_tokens": null,
   "tps": 66.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-g84rgkgd-ctx1024",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 177.7 GB at this context; duplicate traces of the same session: 3b99zco8, cpr3zi2f, gei5lzk7, p8f8ck7k, pv1omm7q; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 516.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 516.5 47.9 177.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g84rgkgd",
   "source_url": "https://omlx.ai/benchmarks/performance/g84rgkgd",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-g84rgkgd-ctx4096",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 178.8 GB at this context; duplicate traces of the same session: 3b99zco8, cpr3zi2f, gei5lzk7, p8f8ck7k, pv1omm7q; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 591.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 591.8 63.9 178.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g84rgkgd",
   "source_url": "https://omlx.ai/benchmarks/performance/g84rgkgd",
   "tg_tokens": null,
   "tps": 63.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-g84rgkgd-ctx8192",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 179.1 GB at this context; duplicate traces of the same session: 3b99zco8, cpr3zi2f, gei5lzk7, p8f8ck7k, pv1omm7q; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 547.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 547.3 49.6 179.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g84rgkgd",
   "source_url": "https://omlx.ai/benchmarks/performance/g84rgkgd",
   "tg_tokens": null,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-g84rgkgd-ctx16384",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 179.3 GB at this context; duplicate traces of the same session: 3b99zco8, cpr3zi2f, gei5lzk7, p8f8ck7k, pv1omm7q; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 537.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 537.3 43.8 179.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g84rgkgd",
   "source_url": "https://omlx.ai/benchmarks/performance/g84rgkgd",
   "tg_tokens": null,
   "tps": 43.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-g84rgkgd-ctx32768",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 179.6 GB at this context; duplicate traces of the same session: 3b99zco8, cpr3zi2f, gei5lzk7, p8f8ck7k, pv1omm7q; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 531.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 531.7 43.2 179.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g84rgkgd",
   "source_url": "https://omlx.ai/benchmarks/performance/g84rgkgd",
   "tg_tokens": null,
   "tps": 43.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-g84rgkgd-ctx65536",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 180.1 GB at this context; duplicate traces of the same session: 3b99zco8, cpr3zi2f, gei5lzk7, p8f8ck7k, pv1omm7q; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 522.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 522.0 51.5 180.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g84rgkgd",
   "source_url": "https://omlx.ai/benchmarks/performance/g84rgkgd",
   "tg_tokens": null,
   "tps": 51.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ja4hketr-ctx1024",
   "model": "Qwen3 14B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: abkb3q01, ljdh061k; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 383.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen--Qwen3-14B-MLX-8bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 8bit; 1k 383.3 23.5 14.8 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ja4hketr",
   "source_url": "https://omlx.ai/benchmarks/performance/ja4hketr",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ja4hketr-ctx4096",
   "model": "Qwen3 14B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 15.5 GB at this context; duplicate traces of the same session: abkb3q01, ljdh061k; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 371.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen--Qwen3-14B-MLX-8bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 8bit; 4k 371.1 22.9 15.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ja4hketr",
   "source_url": "https://omlx.ai/benchmarks/performance/ja4hketr",
   "tg_tokens": null,
   "tps": 22.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ja4hketr-ctx8192",
   "model": "Qwen3 14B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: abkb3q01, ljdh061k; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 357.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen--Qwen3-14B-MLX-8bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 8bit; 8k 357.2 22.0 16.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ja4hketr",
   "source_url": "https://omlx.ai/benchmarks/performance/ja4hketr",
   "tg_tokens": null,
   "tps": 22.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ia2qk32t-ctx1024",
   "model": "Qwen3 8B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 8.6 GB at this context; duplicate traces of the same session: pu8mn385, v6raa02w; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 738.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen--Qwen3-8B-MLX-8bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 8bit; 1k 738.0 44.1 8.6 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ia2qk32t",
   "source_url": "https://omlx.ai/benchmarks/performance/ia2qk32t",
   "tg_tokens": null,
   "tps": 44.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ia2qk32t-ctx4096",
   "model": "Qwen3 8B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 9.0 GB at this context; duplicate traces of the same session: pu8mn385, v6raa02w; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 719.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen--Qwen3-8B-MLX-8bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 8bit; 4k 719.1 41.1 9.0 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ia2qk32t",
   "source_url": "https://omlx.ai/benchmarks/performance/ia2qk32t",
   "tg_tokens": null,
   "tps": 41.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ia2qk32t-ctx8192",
   "model": "Qwen3 8B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 9.6 GB at this context; duplicate traces of the same session: pu8mn385, v6raa02w; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 663.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen--Qwen3-8B-MLX-8bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 8bit; 8k 663.2 36.7 9.6 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ia2qk32t",
   "source_url": "https://omlx.ai/benchmarks/performance/ia2qk32t",
   "tg_tokens": null,
   "tps": 36.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-a22so0kk-ctx1024",
   "model": "Qwen3-8B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 7.2 GB at this context; duplicate traces of the same session: 27gty8bb, gmhgfmw7; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 721.6,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3-8B-Uncensored-oQ6 on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 6bit; 1k 721.6 51.7 7.2 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a22so0kk",
   "source_url": "https://omlx.ai/benchmarks/performance/a22so0kk",
   "tg_tokens": null,
   "tps": 51.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-a22so0kk-ctx4096",
   "model": "Qwen3-8B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 7.6 GB at this context; duplicate traces of the same session: 27gty8bb, gmhgfmw7; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 704.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3-8B-Uncensored-oQ6 on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 6bit; 4k 704.0 47.4 7.6 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a22so0kk",
   "source_url": "https://omlx.ai/benchmarks/performance/a22so0kk",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-a22so0kk-ctx8192",
   "model": "Qwen3-8B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 8.1 GB at this context; duplicate traces of the same session: 27gty8bb, gmhgfmw7; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 641.5,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3-8B-Uncensored-oQ6 on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 6bit; 8k 641.5 40.8 8.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a22so0kk",
   "source_url": "https://omlx.ai/benchmarks/performance/a22so0kk",
   "tg_tokens": null,
   "tps": 40.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-qt2ml8aq-ctx1024",
   "model": "Qwen3-8B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 8.9 GB at this context; duplicate traces of the same session: 7q5lucgr, b2dp6l3t; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 738.6,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3-8B-Uncensored-oQ8 on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 8bit; 1k 738.6 43.0 8.9 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qt2ml8aq",
   "source_url": "https://omlx.ai/benchmarks/performance/qt2ml8aq",
   "tg_tokens": null,
   "tps": 43.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-qt2ml8aq-ctx4096",
   "model": "Qwen3-8B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 9.3 GB at this context; duplicate traces of the same session: 7q5lucgr, b2dp6l3t; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 717.2,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3-8B-Uncensored-oQ8 on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 8bit; 4k 717.2 39.6 9.3 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qt2ml8aq",
   "source_url": "https://omlx.ai/benchmarks/performance/qt2ml8aq",
   "tg_tokens": null,
   "tps": 39.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-qt2ml8aq-ctx8192",
   "model": "Qwen3-8B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 9.8 GB at this context; duplicate traces of the same session: 7q5lucgr, b2dp6l3t; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 642.9,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3-8B-Uncensored-oQ8 on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 8bit; 8k 642.9 36.8 9.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qt2ml8aq",
   "source_url": "https://omlx.ai/benchmarks/performance/qt2ml8aq",
   "tg_tokens": null,
   "tps": 36.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2",
   "id": "omlx-jnw2s22m-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: sex80r0l, wo2kolea; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 86.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 (8c); M2 (8c) \u00b7 16 GB \u00b7 4bit; 4k 86.4 18.5 6.3 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jnw2s22m",
   "source_url": "https://omlx.ai/benchmarks/performance/jnw2s22m",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2",
   "id": "omlx-jnw2s22m-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: sex80r0l, wo2kolea; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 85.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 (8c); M2 (8c) \u00b7 16 GB \u00b7 4bit; 8k 85.4 17.4 6.9 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jnw2s22m",
   "source_url": "https://omlx.ai/benchmarks/performance/jnw2s22m",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2",
   "id": "omlx-jnw2s22m-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.6 GB at this context; duplicate traces of the same session: sex80r0l, wo2kolea; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 43.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M2 (8c); M2 (8c) \u00b7 16 GB \u00b7 4bit; 16k 43.8 14.6 7.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jnw2s22m",
   "source_url": "https://omlx.ai/benchmarks/performance/jnw2s22m",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-a2s3qirq-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: 6tyiu00w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5480.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 5,480 193.5 20.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a2s3qirq",
   "source_url": "https://omlx.ai/benchmarks/performance/a2s3qirq",
   "tg_tokens": null,
   "tps": 193.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-a2s3qirq-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 6tyiu00w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6564.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,564 157.6 21.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a2s3qirq",
   "source_url": "https://omlx.ai/benchmarks/performance/a2s3qirq",
   "tg_tokens": null,
   "tps": 157.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kke3lin2-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 8uhpiem5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5493.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 5,493 161.4 20.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kke3lin2",
   "source_url": "https://omlx.ai/benchmarks/performance/kke3lin2",
   "tg_tokens": null,
   "tps": 161.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-kke3lin2-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 8uhpiem5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6561.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,561 146.4 21.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kke3lin2",
   "source_url": "https://omlx.ai/benchmarks/performance/kke3lin2",
   "tg_tokens": null,
   "tps": 146.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9ytngq5o-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 0wovpfit; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4141.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,141 131.5 19.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ytngq5o",
   "source_url": "https://omlx.ai/benchmarks/performance/9ytngq5o",
   "tg_tokens": null,
   "tps": 131.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9ytngq5o-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: 0wovpfit; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6388.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,388 133.1 20.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ytngq5o",
   "source_url": "https://omlx.ai/benchmarks/performance/9ytngq5o",
   "tg_tokens": null,
   "tps": 133.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vkti3gbl-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: nomlnhh2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4207.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,207 123.0 19.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vkti3gbl",
   "source_url": "https://omlx.ai/benchmarks/performance/vkti3gbl",
   "tg_tokens": null,
   "tps": 123.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vkti3gbl-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: nomlnhh2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6434.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,434 141.6 20.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vkti3gbl",
   "source_url": "https://omlx.ai/benchmarks/performance/vkti3gbl",
   "tg_tokens": null,
   "tps": 141.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jcf3kj7c-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 8qr67aq6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4217.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,217 123.8 19.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jcf3kj7c",
   "source_url": "https://omlx.ai/benchmarks/performance/jcf3kj7c",
   "tg_tokens": null,
   "tps": 123.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jcf3kj7c-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: 8qr67aq6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6442.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,442 144.1 20.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jcf3kj7c",
   "source_url": "https://omlx.ai/benchmarks/performance/jcf3kj7c",
   "tg_tokens": null,
   "tps": 144.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fs55mrbb-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: q2c9vd1v; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4178.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,178 131.9 19.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fs55mrbb",
   "source_url": "https://omlx.ai/benchmarks/performance/fs55mrbb",
   "tg_tokens": null,
   "tps": 131.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fs55mrbb-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: q2c9vd1v; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6409.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,409 129.7 20.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fs55mrbb",
   "source_url": "https://omlx.ai/benchmarks/performance/fs55mrbb",
   "tg_tokens": null,
   "tps": 129.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-p7lze6ur-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4451.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,451 132.9 19.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p7lze6ur",
   "source_url": "https://omlx.ai/benchmarks/performance/p7lze6ur",
   "tg_tokens": null,
   "tps": 132.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-p7lze6ur-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6759.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,759 126.3 20.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p7lze6ur",
   "source_url": "https://omlx.ai/benchmarks/performance/p7lze6ur",
   "tg_tokens": null,
   "tps": 126.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-y1w7xfbl-ctx131072",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace, single run at 128k context; peak memory 27.6 GB; TTFT 411539 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 318.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 128k tokens 318.5 PP tok/s 22.9 TG tok/s 411539 TTFT (ms) 27.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y1w7xfbl",
   "source_url": "https://omlx.ai/benchmarks/performance/y1w7xfbl",
   "tg_tokens": null,
   "tps": 22.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-kiggswh7-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 348.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 348.3 24.6 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kiggswh7",
   "source_url": "https://omlx.ai/benchmarks/performance/kiggswh7",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-kiggswh7-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 350.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 350.8 23.9 27.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kiggswh7",
   "source_url": "https://omlx.ai/benchmarks/performance/kiggswh7",
   "tg_tokens": null,
   "tps": 23.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-52po5uyv-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 347.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 347.8 26.9 26.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/52po5uyv",
   "source_url": "https://omlx.ai/benchmarks/performance/52po5uyv",
   "tg_tokens": null,
   "tps": 26.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-52po5uyv-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 346.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 346.2 24.8 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/52po5uyv",
   "source_url": "https://omlx.ai/benchmarks/performance/52po5uyv",
   "tg_tokens": null,
   "tps": 24.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1sykst2f-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 32.4 GB at this context; duplicate traces of the same session: agcod3pg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 362.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 362.2 25.0 32.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1sykst2f",
   "source_url": "https://omlx.ai/benchmarks/performance/1sykst2f",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1sykst2f-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 33.4 GB at this context; duplicate traces of the same session: agcod3pg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 350.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 350.2 23.3 33.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1sykst2f",
   "source_url": "https://omlx.ai/benchmarks/performance/1sykst2f",
   "tg_tokens": null,
   "tps": 23.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-wbk8g13v-ctx32768",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace, single run at 32k context; peak memory 33.8 GB; TTFT 88824 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 368.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 32k tokens 368.9 PP tok/s 24.4 TG tok/s 88824 TTFT (ms) 33.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wbk8g13v",
   "source_url": "https://omlx.ai/benchmarks/performance/wbk8g13v",
   "tg_tokens": null,
   "tps": 24.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-9c1z6505-ctx65536",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace, single run at 64k context; peak memory 36.0 GB; TTFT 206988 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 316.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 64k tokens 316.6 PP tok/s 18.7 TG tok/s 206988 TTFT (ms) 36.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9c1z6505",
   "source_url": "https://omlx.ai/benchmarks/performance/9c1z6505",
   "tg_tokens": null,
   "tps": 18.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-v65cnjco-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: 4wzw39r9, ge5metor, otx8n8rl, wfctbhf5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 425.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 425.5 73.5 21.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v65cnjco",
   "source_url": "https://omlx.ai/benchmarks/performance/v65cnjco",
   "tg_tokens": null,
   "tps": 73.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-v65cnjco-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: 4wzw39r9, ge5metor, otx8n8rl, wfctbhf5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 507.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 507.4 72.1 25.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v65cnjco",
   "source_url": "https://omlx.ai/benchmarks/performance/v65cnjco",
   "tg_tokens": null,
   "tps": 72.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-v65cnjco-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: 4wzw39r9, ge5metor, otx8n8rl, wfctbhf5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 565.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 565.8 75.3 25.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v65cnjco",
   "source_url": "https://omlx.ai/benchmarks/performance/v65cnjco",
   "tg_tokens": null,
   "tps": 75.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-v65cnjco-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.2 GB at this context; duplicate traces of the same session: 4wzw39r9, ge5metor, otx8n8rl, wfctbhf5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 539.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 539.6 100.8 26.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v65cnjco",
   "source_url": "https://omlx.ai/benchmarks/performance/v65cnjco",
   "tg_tokens": null,
   "tps": 100.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-v65cnjco-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: 4wzw39r9, ge5metor, otx8n8rl, wfctbhf5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 491.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 491.7 154.7 27.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v65cnjco",
   "source_url": "https://omlx.ai/benchmarks/performance/v65cnjco",
   "tg_tokens": null,
   "tps": 154.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-v65cnjco-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.0 GB at this context; duplicate traces of the same session: 4wzw39r9, ge5metor, otx8n8rl, wfctbhf5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 415.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 415.8 67.7 31.0 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v65cnjco",
   "source_url": "https://omlx.ai/benchmarks/performance/v65cnjco",
   "tg_tokens": null,
   "tps": 67.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-l1bod8j0-ctx1024",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 7.1 GB at this context; duplicate traces of the same session: kbj0jsjp; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 141.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 141.9 12.5 7.1 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l1bod8j0",
   "source_url": "https://omlx.ai/benchmarks/performance/l1bod8j0",
   "tg_tokens": null,
   "tps": 12.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-l1bod8j0-ctx4096",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 7.5 GB at this context; duplicate traces of the same session: kbj0jsjp; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 138.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 138.8 12.4 7.5 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l1bod8j0",
   "source_url": "https://omlx.ai/benchmarks/performance/l1bod8j0",
   "tg_tokens": null,
   "tps": 12.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-846giq6v-ctx1024",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 5.1 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1950.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-E2B-it-MLX-8bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 8bit; 1k 1,950 35.1 5.1 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/846giq6v",
   "source_url": "https://omlx.ai/benchmarks/performance/846giq6v",
   "tg_tokens": null,
   "tps": 35.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-846giq6v-ctx4096",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 5.1 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2219.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-E2B-it-MLX-8bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 8bit; 4k 2,219 34.7 5.1 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/846giq6v",
   "source_url": "https://omlx.ai/benchmarks/performance/846giq6v",
   "tg_tokens": null,
   "tps": 34.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-lpdribla-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 5.7 GB at this context; duplicate traces of the same session: nsykiel0; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2207.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 2,207 95.6 5.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lpdribla",
   "source_url": "https://omlx.ai/benchmarks/performance/lpdribla",
   "tg_tokens": null,
   "tps": 95.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-lpdribla-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: nsykiel0; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2589.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 2,589 58.0 5.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lpdribla",
   "source_url": "https://omlx.ai/benchmarks/performance/lpdribla",
   "tg_tokens": null,
   "tps": 58.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-zy3rgtxy-ctx1024",
   "model": "Devstral-Small-2505",
   "notes": "oMLX community benchmark trace; peak memory 13.2 GB at this context; duplicate traces of the same session: j32nlq1s, jn5ymkn9; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 236.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Devstral-Small-2505-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 236.0 26.3 13.2 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zy3rgtxy",
   "source_url": "https://omlx.ai/benchmarks/performance/zy3rgtxy",
   "tg_tokens": null,
   "tps": 26.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-zy3rgtxy-ctx4096",
   "model": "Devstral-Small-2505",
   "notes": "oMLX community benchmark trace; peak memory 13.9 GB at this context; duplicate traces of the same session: j32nlq1s, jn5ymkn9; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Devstral-Small-2505-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 220.5 24.3 13.9 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zy3rgtxy",
   "source_url": "https://omlx.ai/benchmarks/performance/zy3rgtxy",
   "tg_tokens": null,
   "tps": 24.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-zy3rgtxy-ctx8192",
   "model": "Devstral-Small-2505",
   "notes": "oMLX community benchmark trace; peak memory 14.5 GB at this context; duplicate traces of the same session: j32nlq1s, jn5ymkn9; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 213.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Devstral-Small-2505-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 213.8 22.7 14.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zy3rgtxy",
   "source_url": "https://omlx.ai/benchmarks/performance/zy3rgtxy",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-eosugbzw-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: 6z47m5uz, ie3cxsvt; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1315.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.5-4B-MLX-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 1,315 117.3 3.4 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eosugbzw",
   "source_url": "https://omlx.ai/benchmarks/performance/eosugbzw",
   "tg_tokens": null,
   "tps": 117.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-eosugbzw-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.1 GB at this context; duplicate traces of the same session: 6z47m5uz, ie3cxsvt; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1066.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.5-4B-MLX-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 1,066 112.8 4.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eosugbzw",
   "source_url": "https://omlx.ai/benchmarks/performance/eosugbzw",
   "tg_tokens": null,
   "tps": 112.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-eosugbzw-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.5 GB at this context; duplicate traces of the same session: 6z47m5uz, ie3cxsvt; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1082.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.5-4B-MLX-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 1,082 108.5 4.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eosugbzw",
   "source_url": "https://omlx.ai/benchmarks/performance/eosugbzw",
   "tg_tokens": null,
   "tps": 108.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-u37ot2x1-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "MLX; mlx-community; oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: 4gqf0c17, qg605ja1; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 742.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.5-9B-MLX-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 742.9 69.8 5.9 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u37ot2x1",
   "source_url": "https://omlx.ai/benchmarks/performance/u37ot2x1",
   "tg_tokens": null,
   "tps": 69.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-u37ot2x1-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "MLX; mlx-community; oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: 4gqf0c17, qg605ja1; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 751.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.5-9B-MLX-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 751.4 67.9 6.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u37ot2x1",
   "source_url": "https://omlx.ai/benchmarks/performance/u37ot2x1",
   "tg_tokens": null,
   "tps": 67.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-u37ot2x1-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "MLX; mlx-community; oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: 4gqf0c17, qg605ja1; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 627.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.5-9B-MLX-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 627.9 66.0 6.9 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u37ot2x1",
   "source_url": "https://omlx.ai/benchmarks/performance/u37ot2x1",
   "tg_tokens": null,
   "tps": 66.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-bylnchll-ctx1024",
   "model": "Bonsai-27B",
   "notes": "oMLX community benchmark trace; peak memory 7.3 GB at this context; duplicate traces of the same session: ke5krtme; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 43.9,
   "provenance": "community",
   "quant": "1bit",
   "quote": "Bonsai-27B-mlx-1bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 unknown; 1k 43.9 14.9 7.3 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bylnchll",
   "source_url": "https://omlx.ai/benchmarks/performance/bylnchll",
   "tg_tokens": null,
   "tps": 14.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-bylnchll-ctx4096",
   "model": "Bonsai-27B",
   "notes": "oMLX community benchmark trace; peak memory 8.3 GB at this context; duplicate traces of the same session: ke5krtme; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 54.1,
   "provenance": "community",
   "quant": "1bit",
   "quote": "Bonsai-27B-mlx-1bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 unknown; 4k 54.1 14.4 8.3 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bylnchll",
   "source_url": "https://omlx.ai/benchmarks/performance/bylnchll",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-4bdekwsd-ctx1024",
   "model": "Bonsai-8B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: tnbhxyge; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 235.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Bonsai-8B-unpacked-mlx-4Bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 235.1 20.1 4.9 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4bdekwsd",
   "source_url": "https://omlx.ai/benchmarks/performance/4bdekwsd",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-4bdekwsd-ctx4096",
   "model": "Bonsai-8B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 5.4 GB at this context; duplicate traces of the same session: tnbhxyge; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Bonsai-8B-unpacked-mlx-4Bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 220.3 18.2 5.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4bdekwsd",
   "source_url": "https://omlx.ai/benchmarks/performance/4bdekwsd",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-vepmjqif-ctx1024",
   "model": "Gemma-4-26B-A4B-Hermes-VLM",
   "notes": "MLX; oMLX community benchmark trace; peak memory 14.0 GB at this context; duplicate traces of the same session: 5g12h9se; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 630.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Gemma-4-26B-A4B-Hermes-VLM-MLX on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 630.4 45.5 14.0 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vepmjqif",
   "source_url": "https://omlx.ai/benchmarks/performance/vepmjqif",
   "tg_tokens": null,
   "tps": 45.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-vepmjqif-ctx4096",
   "model": "Gemma-4-26B-A4B-Hermes-VLM",
   "notes": "MLX; oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: 5g12h9se; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 962.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Gemma-4-26B-A4B-Hermes-VLM-MLX on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 962.9 43.4 14.3 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vepmjqif",
   "source_url": "https://omlx.ai/benchmarks/performance/vepmjqif",
   "tg_tokens": null,
   "tps": 43.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mcqqe18c-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.5 GB at this context; duplicate traces of the same session: nythpu1p; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2872.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-9B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 2,872 64.4 8.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mcqqe18c",
   "source_url": "https://omlx.ai/benchmarks/performance/mcqqe18c",
   "tg_tokens": null,
   "tps": 64.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mcqqe18c-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.8 GB at this context; duplicate traces of the same session: nythpu1p; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1843.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-9B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 1,843 54.7 8.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mcqqe18c",
   "source_url": "https://omlx.ai/benchmarks/performance/mcqqe18c",
   "tg_tokens": null,
   "tps": 54.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-woqcrqw1-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 214 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4790.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 4,790 PP tok/s 152.4 TG tok/s 214 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/woqcrqw1",
   "source_url": "https://omlx.ai/benchmarks/performance/woqcrqw1",
   "tg_tokens": null,
   "tps": 152.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hfojjacj-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4451.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,451 132.9 19.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hfojjacj",
   "source_url": "https://omlx.ai/benchmarks/performance/hfojjacj",
   "tg_tokens": null,
   "tps": 132.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hfojjacj-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6759.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,759 126.3 20.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hfojjacj",
   "source_url": "https://omlx.ai/benchmarks/performance/hfojjacj",
   "tg_tokens": null,
   "tps": 126.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-d3qjdpzl-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: mh8s11ry; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4808.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,808 133.3 19.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d3qjdpzl",
   "source_url": "https://omlx.ai/benchmarks/performance/d3qjdpzl",
   "tg_tokens": null,
   "tps": 133.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-d3qjdpzl-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: mh8s11ry; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6530.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,530 126.2 20.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d3qjdpzl",
   "source_url": "https://omlx.ai/benchmarks/performance/d3qjdpzl",
   "tg_tokens": null,
   "tps": 126.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-46ftqiwg-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: mf33q8y0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4815.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,815 133.0 19.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/46ftqiwg",
   "source_url": "https://omlx.ai/benchmarks/performance/46ftqiwg",
   "tg_tokens": null,
   "tps": 133.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-46ftqiwg-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: mf33q8y0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6518.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,518 126.6 20.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/46ftqiwg",
   "source_url": "https://omlx.ai/benchmarks/performance/46ftqiwg",
   "tg_tokens": null,
   "tps": 126.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-zf2936mx-ctx4096",
   "model": "Qwen3.6-35B-A3B-oQ8-mtp",
   "notes": "oMLX community benchmark trace; peak memory 38.0 GB at this context; duplicate traces of the same session: 0ij3ni9q, gegkdcm3, guzd7yfq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5709.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.6-35B-A3B-oQ8-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 5,709 149.4 38.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zf2936mx",
   "source_url": "https://omlx.ai/benchmarks/performance/zf2936mx",
   "tg_tokens": null,
   "tps": 149.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zf2936mx-ctx8192",
   "model": "Qwen3.6-35B-A3B-oQ8-mtp",
   "notes": "oMLX community benchmark trace; peak memory 38.2 GB at this context; duplicate traces of the same session: 0ij3ni9q, gegkdcm3, guzd7yfq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5735.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.6-35B-A3B-oQ8-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 5,735 154.2 38.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zf2936mx",
   "source_url": "https://omlx.ai/benchmarks/performance/zf2936mx",
   "tg_tokens": null,
   "tps": 154.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-zf2936mx-ctx16384",
   "model": "Qwen3.6-35B-A3B-oQ8-mtp",
   "notes": "oMLX community benchmark trace; peak memory 38.6 GB at this context; duplicate traces of the same session: 0ij3ni9q, gegkdcm3, guzd7yfq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5268.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.6-35B-A3B-oQ8-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 5,268 191.8 38.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zf2936mx",
   "source_url": "https://omlx.ai/benchmarks/performance/zf2936mx",
   "tg_tokens": null,
   "tps": 191.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-zf2936mx-ctx32768",
   "model": "Qwen3.6-35B-A3B-oQ8-mtp",
   "notes": "oMLX community benchmark trace; peak memory 39.5 GB at this context; duplicate traces of the same session: 0ij3ni9q, gegkdcm3, guzd7yfq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4344.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.6-35B-A3B-oQ8-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 4,344 201.9 39.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zf2936mx",
   "source_url": "https://omlx.ai/benchmarks/performance/zf2936mx",
   "tg_tokens": null,
   "tps": 201.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wulh2ph0-ctx4096",
   "model": "Qwen3.6-35B-A3B-oQ8-mtp",
   "notes": "oMLX community benchmark trace; peak memory 38.0 GB at this context; duplicate traces of the same session: 0maj0umy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1663.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.6-35B-A3B-oQ8-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 1,663 112.7 38.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wulh2ph0",
   "source_url": "https://omlx.ai/benchmarks/performance/wulh2ph0",
   "tg_tokens": null,
   "tps": 112.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wulh2ph0-ctx8192",
   "model": "Qwen3.6-35B-A3B-oQ8-mtp",
   "notes": "oMLX community benchmark trace; peak memory 38.2 GB at this context; duplicate traces of the same session: 0maj0umy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5698.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.6-35B-A3B-oQ8-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 5,698 42.1 38.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wulh2ph0",
   "source_url": "https://omlx.ai/benchmarks/performance/wulh2ph0",
   "tg_tokens": null,
   "tps": 42.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wulh2ph0-ctx16384",
   "model": "Qwen3.6-35B-A3B-oQ8-mtp",
   "notes": "oMLX community benchmark trace; peak memory 38.6 GB at this context; duplicate traces of the same session: 0maj0umy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5296.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.6-35B-A3B-oQ8-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 5,296 155.3 38.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wulh2ph0",
   "source_url": "https://omlx.ai/benchmarks/performance/wulh2ph0",
   "tg_tokens": null,
   "tps": 155.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wulh2ph0-ctx32768",
   "model": "Qwen3.6-35B-A3B-oQ8-mtp",
   "notes": "oMLX community benchmark trace; peak memory 39.5 GB at this context; duplicate traces of the same session: 0maj0umy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4233.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.6-35B-A3B-oQ8-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 4,233 191.0 39.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wulh2ph0",
   "source_url": "https://omlx.ai/benchmarks/performance/wulh2ph0",
   "tg_tokens": null,
   "tps": 191.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-6z3oawn2-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 33.5 GB; TTFT 58189 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1126.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 64k tokens 1,126 PP tok/s 27.5 TG tok/s 58189 TTFT (ms) 33.5 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6z3oawn2",
   "source_url": "https://omlx.ai/benchmarks/performance/6z3oawn2",
   "tg_tokens": null,
   "tps": 27.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-snx3x819-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 1pu01czf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 923.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 923.1 33.8 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/snx3x819",
   "source_url": "https://omlx.ai/benchmarks/performance/snx3x819",
   "tg_tokens": null,
   "tps": 33.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-snx3x819-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 1pu01czf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1014.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,014 33.3 17.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/snx3x819",
   "source_url": "https://omlx.ai/benchmarks/performance/snx3x819",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-rzjuojkt-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 1trzawqt; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 212.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 8bit; 1k 212.5 19.6 28.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rzjuojkt",
   "source_url": "https://omlx.ai/benchmarks/performance/rzjuojkt",
   "tg_tokens": null,
   "tps": 19.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-rzjuojkt-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.9 GB at this context; duplicate traces of the same session: 1trzawqt; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 219.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 8bit; 4k 219.6 19.2 31.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rzjuojkt",
   "source_url": "https://omlx.ai/benchmarks/performance/rzjuojkt",
   "tg_tokens": null,
   "tps": 19.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qzbnzo8e-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: gy3isy8q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 372.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 372.4 13.7 28.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qzbnzo8e",
   "source_url": "https://omlx.ai/benchmarks/performance/qzbnzo8e",
   "tg_tokens": null,
   "tps": 13.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qzbnzo8e-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: gy3isy8q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 508.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 508.5 14.5 29.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qzbnzo8e",
   "source_url": "https://omlx.ai/benchmarks/performance/qzbnzo8e",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-1tfpfbtk-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 23to0v7v, 7315rsfp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 253.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 253.0 45.4 21.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1tfpfbtk",
   "source_url": "https://omlx.ai/benchmarks/performance/1tfpfbtk",
   "tg_tokens": null,
   "tps": 45.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-1tfpfbtk-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: 23to0v7v, 7315rsfp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 336.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 336.9 52.5 25.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1tfpfbtk",
   "source_url": "https://omlx.ai/benchmarks/performance/1tfpfbtk",
   "tg_tokens": null,
   "tps": 52.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-1tfpfbtk-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.3 GB at this context; duplicate traces of the same session: 23to0v7v, 7315rsfp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 293.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 293.4 34.6 28.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1tfpfbtk",
   "source_url": "https://omlx.ai/benchmarks/performance/1tfpfbtk",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-fez9twgg-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: vto015rs; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 252.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 252.9 47.0 21.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fez9twgg",
   "source_url": "https://omlx.ai/benchmarks/performance/fez9twgg",
   "tg_tokens": null,
   "tps": 47.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-fez9twgg-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: vto015rs; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 339.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 339.2 46.3 25.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fez9twgg",
   "source_url": "https://omlx.ai/benchmarks/performance/fez9twgg",
   "tg_tokens": null,
   "tps": 46.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-fez9twgg-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.3 GB at this context; duplicate traces of the same session: vto015rs; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 293.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 293.8 35.3 28.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fez9twgg",
   "source_url": "https://omlx.ai/benchmarks/performance/fez9twgg",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-5ql7o28t-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 253.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 253.3 47.8 21.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ql7o28t",
   "source_url": "https://omlx.ai/benchmarks/performance/5ql7o28t",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-5ql7o28t-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 340.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 340.0 46.4 25.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ql7o28t",
   "source_url": "https://omlx.ai/benchmarks/performance/5ql7o28t",
   "tg_tokens": null,
   "tps": 46.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-5ql7o28t-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 293.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 293.7 34.6 28.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ql7o28t",
   "source_url": "https://omlx.ai/benchmarks/performance/5ql7o28t",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-jm39e7vq-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: ub9ecclo, znw1kbgn; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 338.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 338.3 21.6 26.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jm39e7vq",
   "source_url": "https://omlx.ai/benchmarks/performance/jm39e7vq",
   "tg_tokens": null,
   "tps": 21.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-jm39e7vq-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: ub9ecclo, znw1kbgn; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 338.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 338.9 22.5 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jm39e7vq",
   "source_url": "https://omlx.ai/benchmarks/performance/jm39e7vq",
   "tg_tokens": null,
   "tps": 22.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-jm39e7vq-ctx16384",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: ub9ecclo, znw1kbgn; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 336.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 16k 336.9 26.3 28.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jm39e7vq",
   "source_url": "https://omlx.ai/benchmarks/performance/jm39e7vq",
   "tg_tokens": null,
   "tps": 26.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-4b1632so-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 31.0 GB at this context; duplicate traces of the same session: fs2m5b3r, o2swms03; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 362.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 362.4 22.4 31.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4b1632so",
   "source_url": "https://omlx.ai/benchmarks/performance/4b1632so",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-4b1632so-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 31.9 GB at this context; duplicate traces of the same session: fs2m5b3r, o2swms03; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 411.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 411.8 23.8 31.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4b1632so",
   "source_url": "https://omlx.ai/benchmarks/performance/4b1632so",
   "tg_tokens": null,
   "tps": 23.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-4b1632so-ctx16384",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 32.8 GB at this context; duplicate traces of the same session: fs2m5b3r, o2swms03; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 371.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 16k 371.7 27.6 32.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4b1632so",
   "source_url": "https://omlx.ai/benchmarks/performance/4b1632so",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-h8tpr42z-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 4610 ms; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 222.1,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 1k tokens 222.1 PP tok/s 32.2 TG tok/s 4610 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h8tpr42z",
   "source_url": "https://omlx.ai/benchmarks/performance/h8tpr42z",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-z6vt89dx-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 26u31d36; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 220.5 29.9 18.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z6vt89dx",
   "source_url": "https://omlx.ai/benchmarks/performance/z6vt89dx",
   "tg_tokens": null,
   "tps": 29.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-z6vt89dx-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 26u31d36; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 213.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 213.9 27.6 22.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z6vt89dx",
   "source_url": "https://omlx.ai/benchmarks/performance/z6vt89dx",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5ys2bbry-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 425.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 425.5 73.5 21.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ys2bbry",
   "source_url": "https://omlx.ai/benchmarks/performance/5ys2bbry",
   "tg_tokens": null,
   "tps": 73.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5ys2bbry-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 507.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 507.4 72.1 25.0 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ys2bbry",
   "source_url": "https://omlx.ai/benchmarks/performance/5ys2bbry",
   "tg_tokens": null,
   "tps": 72.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5ys2bbry-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 565.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 565.8 75.3 25.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ys2bbry",
   "source_url": "https://omlx.ai/benchmarks/performance/5ys2bbry",
   "tg_tokens": null,
   "tps": 75.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5ys2bbry-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.2 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 539.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 539.6 100.8 26.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ys2bbry",
   "source_url": "https://omlx.ai/benchmarks/performance/5ys2bbry",
   "tg_tokens": null,
   "tps": 100.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5ys2bbry-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 491.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 491.7 154.7 27.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ys2bbry",
   "source_url": "https://omlx.ai/benchmarks/performance/5ys2bbry",
   "tg_tokens": null,
   "tps": 154.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5ys2bbry-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.0 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 415.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 415.8 67.7 31.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ys2bbry",
   "source_url": "https://omlx.ai/benchmarks/performance/5ys2bbry",
   "tg_tokens": null,
   "tps": 67.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-jmtuzfhy-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: lv9ro36h; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 201.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 8bit; 1k 201.3 21.2 33.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jmtuzfhy",
   "source_url": "https://omlx.ai/benchmarks/performance/jmtuzfhy",
   "tg_tokens": null,
   "tps": 21.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-jmtuzfhy-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.5 GB at this context; duplicate traces of the same session: lv9ro36h; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 213.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 8bit; 4k 213.4 42.2 37.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jmtuzfhy",
   "source_url": "https://omlx.ai/benchmarks/performance/jmtuzfhy",
   "tg_tokens": null,
   "tps": 42.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-y9h4psqi-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 50.6 GB at this context; duplicate traces of the same session: 6hwh16vb, a57guuje; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2165.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 2bit; 4k 2,165 79.8 50.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y9h4psqi",
   "source_url": "https://omlx.ai/benchmarks/performance/y9h4psqi",
   "tg_tokens": null,
   "tps": 79.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-y9h4psqi-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 52.3 GB at this context; duplicate traces of the same session: 6hwh16vb, a57guuje; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3533.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 2bit; 16k 3,533 82.6 52.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y9h4psqi",
   "source_url": "https://omlx.ai/benchmarks/performance/y9h4psqi",
   "tg_tokens": null,
   "tps": 82.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-y9h4psqi-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 54.1 GB at this context; duplicate traces of the same session: 6hwh16vb, a57guuje; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4067.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 2bit; 64k 4,067 58.5 54.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y9h4psqi",
   "source_url": "https://omlx.ai/benchmarks/performance/y9h4psqi",
   "tg_tokens": null,
   "tps": 58.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-u25w407x-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 130.5 GB at this context; duplicate traces of the same session: 2diu39hc, i68iftf3, un33wutw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 821.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 16k 821.4 58.7 130.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u25w407x",
   "source_url": "https://omlx.ai/benchmarks/performance/u25w407x",
   "tg_tokens": null,
   "tps": 58.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-u25w407x-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 131.1 GB at this context; duplicate traces of the same session: 2diu39hc, i68iftf3, un33wutw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1015.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 32k 1,015 52.7 131.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u25w407x",
   "source_url": "https://omlx.ai/benchmarks/performance/u25w407x",
   "tg_tokens": null,
   "tps": 52.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-u25w407x-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 133.6 GB at this context; duplicate traces of the same session: 2diu39hc, i68iftf3, un33wutw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1003.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 64k 1,003 63.5 133.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u25w407x",
   "source_url": "https://omlx.ai/benchmarks/performance/u25w407x",
   "tg_tokens": null,
   "tps": 63.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-u25w407x-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 137.9 GB at this context; duplicate traces of the same session: 2diu39hc, i68iftf3, un33wutw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 987.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 128k 987.6 54.3 137.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u25w407x",
   "source_url": "https://omlx.ai/benchmarks/performance/u25w407x",
   "tg_tokens": null,
   "tps": 54.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-e5re6ryh-ctx1024",
   "model": "Swift 1.5 Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 142.7 GB at this context; duplicate traces of the same session: mapflo9x; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 884.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 1k 884.9 89.9 142.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e5re6ryh",
   "source_url": "https://omlx.ai/benchmarks/performance/e5re6ryh",
   "tg_tokens": null,
   "tps": 89.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-e5re6ryh-ctx4096",
   "model": "Swift 1.5 Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 144.6 GB at this context; duplicate traces of the same session: mapflo9x; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 809.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 4k 809.7 79.9 144.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e5re6ryh",
   "source_url": "https://omlx.ai/benchmarks/performance/e5re6ryh",
   "tg_tokens": null,
   "tps": 79.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nghj9wcu-ctx1024",
   "model": "Swift 1.5 Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 139.5 GB at this context; duplicate traces of the same session: 9ialq5ug; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 914.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 1k 914.5 66.2 139.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nghj9wcu",
   "source_url": "https://omlx.ai/benchmarks/performance/nghj9wcu",
   "tg_tokens": null,
   "tps": 66.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nghj9wcu-ctx4096",
   "model": "Swift 1.5 Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 141.3 GB at this context; duplicate traces of the same session: 9ialq5ug; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 841.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 4k 841.4 60.4 141.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nghj9wcu",
   "source_url": "https://omlx.ai/benchmarks/performance/nghj9wcu",
   "tg_tokens": null,
   "tps": 60.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-n3x80fwh-ctx1024",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 5.1 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1950.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-E2B-it-MLX-8bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 8bit; 1k 1,950 35.1 5.1 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n3x80fwh",
   "source_url": "https://omlx.ai/benchmarks/performance/n3x80fwh",
   "tg_tokens": null,
   "tps": 35.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-n3x80fwh-ctx4096",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 5.1 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2219.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-E2B-it-MLX-8bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 8bit; 4k 2,219 34.7 5.1 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n3x80fwh",
   "source_url": "https://omlx.ai/benchmarks/performance/n3x80fwh",
   "tg_tokens": null,
   "tps": 34.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-czggy6it-ctx1024",
   "model": "gemma-4-E2B-it",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 3.8 GB; TTFT 451 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2270.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e2b-it on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k tokens 2,270 PP tok/s 64.4 TG tok/s 451 TTFT (ms) 3.8 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/czggy6it",
   "source_url": "https://omlx.ai/benchmarks/performance/czggy6it",
   "tg_tokens": null,
   "tps": 64.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-bag0p0dk-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 4.5 GB at this context; duplicate traces of the same session: n2fydwl3; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 652.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 652.3 31.3 4.5 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bag0p0dk",
   "source_url": "https://omlx.ai/benchmarks/performance/bag0p0dk",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-bag0p0dk-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 4.6 GB at this context; duplicate traces of the same session: n2fydwl3; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 664.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 664.0 31.1 4.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bag0p0dk",
   "source_url": "https://omlx.ai/benchmarks/performance/bag0p0dk",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9k5f2t9p-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.3 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 386.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 386.0 56.1 179.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9k5f2t9p",
   "source_url": "https://omlx.ai/benchmarks/performance/9k5f2t9p",
   "tg_tokens": null,
   "tps": 56.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9k5f2t9p-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.5 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 399.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 399.2 50.2 179.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9k5f2t9p",
   "source_url": "https://omlx.ai/benchmarks/performance/9k5f2t9p",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9k5f2t9p-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.2 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 396.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 396.2 47.4 180.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9k5f2t9p",
   "source_url": "https://omlx.ai/benchmarks/performance/9k5f2t9p",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9gezimog-ctx4096",
   "model": "GPT-OSS-Coder",
   "notes": "MLX; oMLX community benchmark trace; peak memory 61.3 GB at this context; duplicate traces of the same session: ioysr08l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1463.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GPT-OSS-Coder-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,463 78.2 61.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9gezimog",
   "source_url": "https://omlx.ai/benchmarks/performance/9gezimog",
   "tg_tokens": null,
   "tps": 78.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9gezimog-ctx16384",
   "model": "GPT-OSS-Coder",
   "notes": "MLX; oMLX community benchmark trace; peak memory 61.7 GB at this context; duplicate traces of the same session: ioysr08l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2066.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GPT-OSS-Coder-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,066 68.6 61.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9gezimog",
   "source_url": "https://omlx.ai/benchmarks/performance/9gezimog",
   "tg_tokens": null,
   "tps": 68.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0de8glmc-ctx1024",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: 6udk31d1, gmb3p8zx, gr5vlsnm; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2312.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 2,312 142.9 19.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0de8glmc",
   "source_url": "https://omlx.ai/benchmarks/performance/0de8glmc",
   "tg_tokens": null,
   "tps": 142.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0de8glmc-ctx4096",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 6udk31d1, gmb3p8zx, gr5vlsnm; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2525.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 2,525 140.3 20.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0de8glmc",
   "source_url": "https://omlx.ai/benchmarks/performance/0de8glmc",
   "tg_tokens": null,
   "tps": 140.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0de8glmc-ctx8192",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 6udk31d1, gmb3p8zx, gr5vlsnm; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2507.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 2,507 136.4 20.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0de8glmc",
   "source_url": "https://omlx.ai/benchmarks/performance/0de8glmc",
   "tg_tokens": null,
   "tps": 136.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0de8glmc-ctx131072",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 6udk31d1, gmb3p8zx, gr5vlsnm; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1501.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 1,501 96.5 21.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0de8glmc",
   "source_url": "https://omlx.ai/benchmarks/performance/0de8glmc",
   "tg_tokens": null,
   "tps": 96.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-48jug9w2-ctx1024",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 237 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4323.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 4,323 PP tok/s 136.3 TG tok/s 237 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/48jug9w2",
   "source_url": "https://omlx.ai/benchmarks/performance/48jug9w2",
   "tg_tokens": null,
   "tps": 136.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-so1fnq4c-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: rillgmla; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 752.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 752.9 108.1 21.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/so1fnq4c",
   "source_url": "https://omlx.ai/benchmarks/performance/so1fnq4c",
   "tg_tokens": null,
   "tps": 108.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-so1fnq4c-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: rillgmla; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 804.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 8k 804.9 103.7 23.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/so1fnq4c",
   "source_url": "https://omlx.ai/benchmarks/performance/so1fnq4c",
   "tg_tokens": null,
   "tps": 103.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3sbuetqh-ctx4096",
   "model": "Qwen3.6-35B-A3B-oQ8-mtp",
   "notes": "oMLX community benchmark trace; peak memory 38.0 GB at this context; duplicate traces of the same session: tle9ejc8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1663.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.6-35B-A3B-oQ8-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 1,663 112.7 38.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3sbuetqh",
   "source_url": "https://omlx.ai/benchmarks/performance/3sbuetqh",
   "tg_tokens": null,
   "tps": 112.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3sbuetqh-ctx8192",
   "model": "Qwen3.6-35B-A3B-oQ8-mtp",
   "notes": "oMLX community benchmark trace; peak memory 38.2 GB at this context; duplicate traces of the same session: tle9ejc8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5698.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.6-35B-A3B-oQ8-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 5,698 42.1 38.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3sbuetqh",
   "source_url": "https://omlx.ai/benchmarks/performance/3sbuetqh",
   "tg_tokens": null,
   "tps": 42.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3sbuetqh-ctx16384",
   "model": "Qwen3.6-35B-A3B-oQ8-mtp",
   "notes": "oMLX community benchmark trace; peak memory 38.6 GB at this context; duplicate traces of the same session: tle9ejc8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5296.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.6-35B-A3B-oQ8-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 5,296 155.3 38.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3sbuetqh",
   "source_url": "https://omlx.ai/benchmarks/performance/3sbuetqh",
   "tg_tokens": null,
   "tps": 155.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3sbuetqh-ctx32768",
   "model": "Qwen3.6-35B-A3B-oQ8-mtp",
   "notes": "oMLX community benchmark trace; peak memory 39.5 GB at this context; duplicate traces of the same session: tle9ejc8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4233.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.6-35B-A3B-oQ8-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 4,233 191.0 39.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3sbuetqh",
   "source_url": "https://omlx.ai/benchmarks/performance/3sbuetqh",
   "tg_tokens": null,
   "tps": 191.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-6uf9ph1k-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 6rj8yq34; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 129.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k 129.0 19.2 19.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6uf9ph1k",
   "source_url": "https://omlx.ai/benchmarks/performance/6uf9ph1k",
   "tg_tokens": null,
   "tps": 19.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-6uf9ph1k-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 6rj8yq34; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 118.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 16k 118.0 14.0 21.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6uf9ph1k",
   "source_url": "https://omlx.ai/benchmarks/performance/6uf9ph1k",
   "tg_tokens": null,
   "tps": 14.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-wztg483x-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: bpr34tea; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 139.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 8bit; 4k 139.9 11.7 32.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wztg483x",
   "source_url": "https://omlx.ai/benchmarks/performance/wztg483x",
   "tg_tokens": null,
   "tps": 11.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-wztg483x-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.2 GB at this context; duplicate traces of the same session: bpr34tea; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 8bit; 16k 131.2 11.2 34.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wztg483x",
   "source_url": "https://omlx.ai/benchmarks/performance/wztg483x",
   "tg_tokens": null,
   "tps": 11.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-1u03su11-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 37.5 GB at this context; duplicate traces of the same session: 0b0nkgwk; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 141.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 8bit; 4k 141.6 12.7 37.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1u03su11",
   "source_url": "https://omlx.ai/benchmarks/performance/1u03su11",
   "tg_tokens": null,
   "tps": 12.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-1u03su11-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 39.5 GB at this context; duplicate traces of the same session: 0b0nkgwk; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 8bit; 16k 131.2 17.7 39.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1u03su11",
   "source_url": "https://omlx.ai/benchmarks/performance/1u03su11",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-jjtjcghk-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.3 GB at this context; duplicate traces of the same session: 1mf57jts; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 126.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 1k 126.4 8.4 28.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jjtjcghk",
   "source_url": "https://omlx.ai/benchmarks/performance/jjtjcghk",
   "tg_tokens": null,
   "tps": 8.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-jjtjcghk-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: 1mf57jts; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 4k 127.7 8.4 29.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jjtjcghk",
   "source_url": "https://omlx.ai/benchmarks/performance/jjtjcghk",
   "tg_tokens": null,
   "tps": 8.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-dlohmlwr-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: heyq944v; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 111.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 1k 111.8 19.4 16.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dlohmlwr",
   "source_url": "https://omlx.ai/benchmarks/performance/dlohmlwr",
   "tg_tokens": null,
   "tps": 19.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-dlohmlwr-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: heyq944v; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 107.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 4k 107.6 18.8 17.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dlohmlwr",
   "source_url": "https://omlx.ai/benchmarks/performance/dlohmlwr",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-hybrm42n-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: epr0y2bm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 253.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 253.3 47.8 21.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hybrm42n",
   "source_url": "https://omlx.ai/benchmarks/performance/hybrm42n",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-hybrm42n-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: epr0y2bm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 340.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 340.0 46.4 25.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hybrm42n",
   "source_url": "https://omlx.ai/benchmarks/performance/hybrm42n",
   "tg_tokens": null,
   "tps": 46.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-hybrm42n-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.3 GB at this context; duplicate traces of the same session: epr0y2bm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 293.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 293.7 34.6 28.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hybrm42n",
   "source_url": "https://omlx.ai/benchmarks/performance/hybrm42n",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-97h4723x-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 252.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 252.9 47.0 21.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/97h4723x",
   "source_url": "https://omlx.ai/benchmarks/performance/97h4723x",
   "tg_tokens": null,
   "tps": 47.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-97h4723x-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 339.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 339.2 46.3 25.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/97h4723x",
   "source_url": "https://omlx.ai/benchmarks/performance/97h4723x",
   "tg_tokens": null,
   "tps": 46.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-97h4723x-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 293.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 293.8 35.3 28.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/97h4723x",
   "source_url": "https://omlx.ai/benchmarks/performance/97h4723x",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-t8azsgn9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: mf1g2w05; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 218.1,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 1k 218.1 36.6 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t8azsgn9",
   "source_url": "https://omlx.ai/benchmarks/performance/t8azsgn9",
   "tg_tokens": null,
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-t8azsgn9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: mf1g2w05; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 238.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 4k 238.0 41.9 23.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t8azsgn9",
   "source_url": "https://omlx.ai/benchmarks/performance/t8azsgn9",
   "tg_tokens": null,
   "tps": 41.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-9w9hu8k1-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: w5vetroc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 220.5 45.7 18.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9w9hu8k1",
   "source_url": "https://omlx.ai/benchmarks/performance/9w9hu8k1",
   "tg_tokens": null,
   "tps": 45.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-9w9hu8k1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: w5vetroc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 214.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 214.8 26.1 22.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9w9hu8k1",
   "source_url": "https://omlx.ai/benchmarks/performance/9w9hu8k1",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9a6ljq3i-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: wghvnul0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 692.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 692.4 57.0 19.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9a6ljq3i",
   "source_url": "https://omlx.ai/benchmarks/performance/9a6ljq3i",
   "tg_tokens": null,
   "tps": 57.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9a6ljq3i-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: wghvnul0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 688.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 688.9 81.2 20.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9a6ljq3i",
   "source_url": "https://omlx.ai/benchmarks/performance/9a6ljq3i",
   "tg_tokens": null,
   "tps": 81.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-2mohp7r0-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: xtgkrao2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 471.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 471.0 33.1 17.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2mohp7r0",
   "source_url": "https://omlx.ai/benchmarks/performance/2mohp7r0",
   "tg_tokens": null,
   "tps": 33.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-2mohp7r0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: xtgkrao2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 563.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 563.6 35.0 19.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2mohp7r0",
   "source_url": "https://omlx.ai/benchmarks/performance/2mohp7r0",
   "tg_tokens": null,
   "tps": 35.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-iy5folnn-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: 0rjo3hob; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 354.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 354.9 27.2 17.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iy5folnn",
   "source_url": "https://omlx.ai/benchmarks/performance/iy5folnn",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-iy5folnn-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 0rjo3hob; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 371.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 371.1 31.6 18.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iy5folnn",
   "source_url": "https://omlx.ai/benchmarks/performance/iy5folnn",
   "tg_tokens": null,
   "tps": 31.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-z22jrpat-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.6 GB at this context; duplicate traces of the same session: 8gnvh1kx, 8i5nwvo2, zz58ayg7; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 314.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 16k 314.2 41.8 38.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z22jrpat",
   "source_url": "https://omlx.ai/benchmarks/performance/z22jrpat",
   "tg_tokens": null,
   "tps": 41.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-z22jrpat-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 40.3 GB at this context; duplicate traces of the same session: 8gnvh1kx, 8i5nwvo2, zz58ayg7; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 292.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 32k 292.4 75.9 40.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z22jrpat",
   "source_url": "https://omlx.ai/benchmarks/performance/z22jrpat",
   "tg_tokens": null,
   "tps": 75.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-z22jrpat-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.1 GB at this context; duplicate traces of the same session: 8gnvh1kx, 8i5nwvo2, zz58ayg7; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 256.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 64k 256.0 45.3 43.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z22jrpat",
   "source_url": "https://omlx.ai/benchmarks/performance/z22jrpat",
   "tg_tokens": null,
   "tps": 45.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-z22jrpat-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 47.6 GB at this context; duplicate traces of the same session: 8gnvh1kx, 8i5nwvo2, zz58ayg7; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 128k 205.5 40.1 47.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z22jrpat",
   "source_url": "https://omlx.ai/benchmarks/performance/z22jrpat",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-66e6xatw-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.7 GB at this context; duplicate traces of the same session: hjboq62p, ntezc67x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2673.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,673 102.3 72.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/66e6xatw",
   "source_url": "https://omlx.ai/benchmarks/performance/66e6xatw",
   "tg_tokens": null,
   "tps": 102.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-66e6xatw-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.5 GB at this context; duplicate traces of the same session: hjboq62p, ntezc67x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2409.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,409 68.4 78.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/66e6xatw",
   "source_url": "https://omlx.ai/benchmarks/performance/66e6xatw",
   "tg_tokens": null,
   "tps": 68.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-66e6xatw-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.8 GB at this context; duplicate traces of the same session: hjboq62p, ntezc67x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2362.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,362 62.9 80.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/66e6xatw",
   "source_url": "https://omlx.ai/benchmarks/performance/66e6xatw",
   "tg_tokens": null,
   "tps": 62.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5xgyifn0-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: 1xrq0lcj, 7kfydu3h, 7m93vj1i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2310.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,310 58.7 71.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5xgyifn0",
   "source_url": "https://omlx.ai/benchmarks/performance/5xgyifn0",
   "tg_tokens": null,
   "tps": 58.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5xgyifn0-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.7 GB at this context; duplicate traces of the same session: 1xrq0lcj, 7kfydu3h, 7m93vj1i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2498.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,498 110.6 72.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5xgyifn0",
   "source_url": "https://omlx.ai/benchmarks/performance/5xgyifn0",
   "tg_tokens": null,
   "tps": 110.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5xgyifn0-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.5 GB at this context; duplicate traces of the same session: 1xrq0lcj, 7kfydu3h, 7m93vj1i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2427.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,427 67.0 78.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5xgyifn0",
   "source_url": "https://omlx.ai/benchmarks/performance/5xgyifn0",
   "tg_tokens": null,
   "tps": 67.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5xgyifn0-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.8 GB at this context; duplicate traces of the same session: 1xrq0lcj, 7kfydu3h, 7m93vj1i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2348.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,348 63.1 80.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5xgyifn0",
   "source_url": "https://omlx.ai/benchmarks/performance/5xgyifn0",
   "tg_tokens": null,
   "tps": 63.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3oxw6atw-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.7 GB at this context; duplicate traces of the same session: q8hcqtx6, uvta5t9w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2599.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,599 108.7 72.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3oxw6atw",
   "source_url": "https://omlx.ai/benchmarks/performance/3oxw6atw",
   "tg_tokens": null,
   "tps": 108.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3oxw6atw-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.6 GB at this context; duplicate traces of the same session: q8hcqtx6, uvta5t9w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2441.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,441 67.4 78.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3oxw6atw",
   "source_url": "https://omlx.ai/benchmarks/performance/3oxw6atw",
   "tg_tokens": null,
   "tps": 67.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3oxw6atw-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.8 GB at this context; duplicate traces of the same session: q8hcqtx6, uvta5t9w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2353.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,353 50.2 80.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3oxw6atw",
   "source_url": "https://omlx.ai/benchmarks/performance/3oxw6atw",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-cx3ko7k3-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: 95v5waqr; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 612.1,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 612.1 73.4 11.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cx3ko7k3",
   "source_url": "https://omlx.ai/benchmarks/performance/cx3ko7k3",
   "tg_tokens": null,
   "tps": 73.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-cx3ko7k3-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: 95v5waqr; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 662.5,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 662.5 69.5 11.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cx3ko7k3",
   "source_url": "https://omlx.ai/benchmarks/performance/cx3ko7k3",
   "tg_tokens": null,
   "tps": 69.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-cx3ko7k3-ctx8192",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.9 GB at this context; duplicate traces of the same session: 95v5waqr; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 640.2,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 8k 640.2 63.1 11.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cx3ko7k3",
   "source_url": "https://omlx.ai/benchmarks/performance/cx3ko7k3",
   "tg_tokens": null,
   "tps": 63.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-cx3ko7k3-ctx16384",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 12.1 GB at this context; duplicate traces of the same session: 95v5waqr; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 589.3,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 16k 589.3 55.9 12.1 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cx3ko7k3",
   "source_url": "https://omlx.ai/benchmarks/performance/cx3ko7k3",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-cx3ko7k3-ctx32768",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 12.4 GB at this context; duplicate traces of the same session: 95v5waqr; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 504.9,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 32k 504.9 45.8 12.4 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cx3ko7k3",
   "source_url": "https://omlx.ai/benchmarks/performance/cx3ko7k3",
   "tg_tokens": null,
   "tps": 45.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-cx3ko7k3-ctx65536",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 13.2 GB at this context; duplicate traces of the same session: 95v5waqr; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 397.5,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 64k 397.5 32.5 13.2 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cx3ko7k3",
   "source_url": "https://omlx.ai/benchmarks/performance/cx3ko7k3",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-cx3ko7k3-ctx131072",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: 95v5waqr; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 276.5,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 128k 276.5 21.1 14.6 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cx3ko7k3",
   "source_url": "https://omlx.ai/benchmarks/performance/cx3ko7k3",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-cx3ko7k3-ctx199680",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 16.2 GB at this context; duplicate traces of the same session: 95v5waqr; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 208.7,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 195k 208.7 15.4 16.2 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cx3ko7k3",
   "source_url": "https://omlx.ai/benchmarks/performance/cx3ko7k3",
   "tg_tokens": null,
   "tps": 15.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-rx5yn1j9-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: nzw504up; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 660.8,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 660.8 63.4 11.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rx5yn1j9",
   "source_url": "https://omlx.ai/benchmarks/performance/rx5yn1j9",
   "tg_tokens": null,
   "tps": 63.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-rx5yn1j9-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: nzw504up; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 739.9,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 739.9 62.5 11.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rx5yn1j9",
   "source_url": "https://omlx.ai/benchmarks/performance/rx5yn1j9",
   "tg_tokens": null,
   "tps": 62.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-kbkgm4a5-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: fuamkfez; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 785.6,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 785.6 72.5 11.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kbkgm4a5",
   "source_url": "https://omlx.ai/benchmarks/performance/kbkgm4a5",
   "tg_tokens": null,
   "tps": 72.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-kbkgm4a5-ctx32768",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 12.4 GB at this context; duplicate traces of the same session: fuamkfez; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 621.6,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 621.6 48.6 12.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kbkgm4a5",
   "source_url": "https://omlx.ai/benchmarks/performance/kbkgm4a5",
   "tg_tokens": null,
   "tps": 48.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-9me03s95-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 660.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "gpt-oss-20b-oQ6e-bf16-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 6bit; 1k 660.9 62.4 16.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9me03s95",
   "source_url": "https://omlx.ai/benchmarks/performance/9me03s95",
   "tg_tokens": null,
   "tps": 62.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-9me03s95-ctx32768",
   "model": "gpt-oss 20B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 601.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "gpt-oss-20b-oQ6e-bf16-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 6bit; 32k 601.7 44.2 17.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9me03s95",
   "source_url": "https://omlx.ai/benchmarks/performance/9me03s95",
   "tg_tokens": null,
   "tps": 44.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-x0uisaic-ctx1024",
   "model": "granite-4.2-3b",
   "notes": "MLX; oMLX community benchmark trace; peak memory 2.8 GB at this context; duplicate traces of the same session: 0pwzhrpz, 674azzpo, 6imjl0bj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3759.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "granite-4.2-3b-q4-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 3,759 110.6 2.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x0uisaic",
   "source_url": "https://omlx.ai/benchmarks/performance/x0uisaic",
   "tg_tokens": null,
   "tps": 110.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-x0uisaic-ctx4096",
   "model": "granite-4.2-3b",
   "notes": "MLX; oMLX community benchmark trace; peak memory 3.0 GB at this context; duplicate traces of the same session: 0pwzhrpz, 674azzpo, 6imjl0bj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3706.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "granite-4.2-3b-q4-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,706 99.0 3.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x0uisaic",
   "source_url": "https://omlx.ai/benchmarks/performance/x0uisaic",
   "tg_tokens": null,
   "tps": 99.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-x0uisaic-ctx8192",
   "model": "granite-4.2-3b",
   "notes": "MLX; oMLX community benchmark trace; peak memory 3.3 GB at this context; duplicate traces of the same session: 0pwzhrpz, 674azzpo, 6imjl0bj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3447.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "granite-4.2-3b-q4-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 3,447 78.1 3.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x0uisaic",
   "source_url": "https://omlx.ai/benchmarks/performance/x0uisaic",
   "tg_tokens": null,
   "tps": 78.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-x0uisaic-ctx16384",
   "model": "granite-4.2-3b",
   "notes": "MLX; oMLX community benchmark trace; peak memory 3.8 GB at this context; duplicate traces of the same session: 0pwzhrpz, 674azzpo, 6imjl0bj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2943.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "granite-4.2-3b-q4-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,943 64.2 3.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x0uisaic",
   "source_url": "https://omlx.ai/benchmarks/performance/x0uisaic",
   "tg_tokens": null,
   "tps": 64.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ag78mdo1-ctx1024",
   "model": "Artemis-31B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: 487p9t00; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 88.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Artemis-31B-v1.1-mlx-6Bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 1k 88.5 11.7 24.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ag78mdo1",
   "source_url": "https://omlx.ai/benchmarks/performance/ag78mdo1",
   "tg_tokens": null,
   "tps": 11.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ag78mdo1-ctx4096",
   "model": "Artemis-31B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: 487p9t00; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 88.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Artemis-31B-v1.1-mlx-6Bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 4k 88.0 11.3 25.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ag78mdo1",
   "source_url": "https://omlx.ai/benchmarks/performance/ag78mdo1",
   "tg_tokens": null,
   "tps": 11.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-kifre1ok-ctx16384",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.4 GB at this context; duplicate traces of the same session: 3gthjdv5, 86uw0tyd, yj93tjs0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 433.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 433.9 35.0 156.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kifre1ok",
   "source_url": "https://omlx.ai/benchmarks/performance/kifre1ok",
   "tg_tokens": null,
   "tps": 35.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-kifre1ok-ctx32768",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.5 GB at this context; duplicate traces of the same session: 3gthjdv5, 86uw0tyd, yj93tjs0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 433.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 433.8 47.2 156.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kifre1ok",
   "source_url": "https://omlx.ai/benchmarks/performance/kifre1ok",
   "tg_tokens": null,
   "tps": 47.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-kifre1ok-ctx65536",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.8 GB at this context; duplicate traces of the same session: 3gthjdv5, 86uw0tyd, yj93tjs0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 411.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 411.3 39.3 156.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kifre1ok",
   "source_url": "https://omlx.ai/benchmarks/performance/kifre1ok",
   "tg_tokens": null,
   "tps": 39.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-kifre1ok-ctx131072",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 157.1 GB at this context; duplicate traces of the same session: 3gthjdv5, 86uw0tyd, yj93tjs0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 372.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 128k 372.5 32.8 157.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kifre1ok",
   "source_url": "https://omlx.ai/benchmarks/performance/kifre1ok",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-fwaz8bl8-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.3 GB at this context; duplicate traces of the same session: 1qbedlpl; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 386.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 386.0 56.1 179.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fwaz8bl8",
   "source_url": "https://omlx.ai/benchmarks/performance/fwaz8bl8",
   "tg_tokens": null,
   "tps": 56.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-fwaz8bl8-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.5 GB at this context; duplicate traces of the same session: 1qbedlpl; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 399.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 399.2 50.2 179.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fwaz8bl8",
   "source_url": "https://omlx.ai/benchmarks/performance/fwaz8bl8",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-fwaz8bl8-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.2 GB at this context; duplicate traces of the same session: 1qbedlpl; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 396.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 396.2 47.4 180.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fwaz8bl8",
   "source_url": "https://omlx.ai/benchmarks/performance/fwaz8bl8",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3t88ce0y-ctx1024",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 97kdnfkb, t48r789q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4666.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,666 154.2 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3t88ce0y",
   "source_url": "https://omlx.ai/benchmarks/performance/3t88ce0y",
   "tg_tokens": null,
   "tps": 154.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3t88ce0y-ctx32768",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 97kdnfkb, t48r789q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4914.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 4,914 124.9 20.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3t88ce0y",
   "source_url": "https://omlx.ai/benchmarks/performance/3t88ce0y",
   "tg_tokens": null,
   "tps": 124.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3t88ce0y-ctx65536",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 97kdnfkb, t48r789q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3674.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 3,674 103.5 21.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3t88ce0y",
   "source_url": "https://omlx.ai/benchmarks/performance/3t88ce0y",
   "tg_tokens": null,
   "tps": 103.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3t88ce0y-ctx131072",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: 97kdnfkb, t48r789q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2414.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,414 76.4 22.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3t88ce0y",
   "source_url": "https://omlx.ai/benchmarks/performance/3t88ce0y",
   "tg_tokens": null,
   "tps": 76.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3t88ce0y-ctx199680",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: 97kdnfkb, t48r789q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1827.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 1,827 63.5 23.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3t88ce0y",
   "source_url": "https://omlx.ai/benchmarks/performance/3t88ce0y",
   "tg_tokens": null,
   "tps": 63.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-9kmplhc4-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: b6tmvwos; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 278.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 278.7 18.3 17.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9kmplhc4",
   "source_url": "https://omlx.ai/benchmarks/performance/9kmplhc4",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-9kmplhc4-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: b6tmvwos; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 206.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 206.3 17.3 18.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9kmplhc4",
   "source_url": "https://omlx.ai/benchmarks/performance/9kmplhc4",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i9uvw501-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: bhnizs4e; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 788.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 788.8 25.9 17.2 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i9uvw501",
   "source_url": "https://omlx.ai/benchmarks/performance/i9uvw501",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-i9uvw501-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: bhnizs4e; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 797.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 797.2 24.6 17.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i9uvw501",
   "source_url": "https://omlx.ai/benchmarks/performance/i9uvw501",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-u46hmqst-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: nq3ynule; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 397.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 397.8 27.0 23.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u46hmqst",
   "source_url": "https://omlx.ai/benchmarks/performance/u46hmqst",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-u46hmqst-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: nq3ynule; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 421.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 421.9 25.0 24.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u46hmqst",
   "source_url": "https://omlx.ai/benchmarks/performance/u46hmqst",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9f8cca8h-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 6320zuxf, bu5to86l, cd3photo, ggoi9flq, q3s4bo85; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 428.5,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 428.5 116.6 19.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9f8cca8h",
   "source_url": "https://omlx.ai/benchmarks/performance/9f8cca8h",
   "tg_tokens": null,
   "tps": 116.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9f8cca8h-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: 6320zuxf, bu5to86l, cd3photo, ggoi9flq, q3s4bo85; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 471.4,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 471.4 104.9 22.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9f8cca8h",
   "source_url": "https://omlx.ai/benchmarks/performance/9f8cca8h",
   "tg_tokens": null,
   "tps": 104.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9f8cca8h-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 6320zuxf, bu5to86l, cd3photo, ggoi9flq, q3s4bo85; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 462.8,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 462.8 61.1 23.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9f8cca8h",
   "source_url": "https://omlx.ai/benchmarks/performance/9f8cca8h",
   "tg_tokens": null,
   "tps": 61.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9f8cca8h-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: 6320zuxf, bu5to86l, cd3photo, ggoi9flq, q3s4bo85; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 444.1,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 444.1 79.3 24.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9f8cca8h",
   "source_url": "https://omlx.ai/benchmarks/performance/9f8cca8h",
   "tg_tokens": null,
   "tps": 79.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9f8cca8h-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: 6320zuxf, bu5to86l, cd3photo, ggoi9flq, q3s4bo85; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 411.5,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 411.5 67.0 25.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9f8cca8h",
   "source_url": "https://omlx.ai/benchmarks/performance/9f8cca8h",
   "tg_tokens": null,
   "tps": 67.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9f8cca8h-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.7 GB at this context; duplicate traces of the same session: 6320zuxf, bu5to86l, cd3photo, ggoi9flq, q3s4bo85; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 354.8,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 354.8 61.0 28.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9f8cca8h",
   "source_url": "https://omlx.ai/benchmarks/performance/9f8cca8h",
   "tg_tokens": null,
   "tps": 61.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-36i7yy1n-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: ollwo6rn, s8hgscyp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 942.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 942.3 48.8 18.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/36i7yy1n",
   "source_url": "https://omlx.ai/benchmarks/performance/36i7yy1n",
   "tg_tokens": null,
   "tps": 48.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-36i7yy1n-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: ollwo6rn, s8hgscyp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 798.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 798.8 84.5 19.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/36i7yy1n",
   "source_url": "https://omlx.ai/benchmarks/performance/36i7yy1n",
   "tg_tokens": null,
   "tps": 84.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-36i7yy1n-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: ollwo6rn, s8hgscyp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 645.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 645.4 75.3 20.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/36i7yy1n",
   "source_url": "https://omlx.ai/benchmarks/performance/36i7yy1n",
   "tg_tokens": null,
   "tps": 75.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-gomhptmc-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: k7z7cdzl; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 501.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 501.4 33.7 17.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gomhptmc",
   "source_url": "https://omlx.ai/benchmarks/performance/gomhptmc",
   "tg_tokens": null,
   "tps": 33.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-gomhptmc-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: k7z7cdzl; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 569.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 569.6 31.2 19.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gomhptmc",
   "source_url": "https://omlx.ai/benchmarks/performance/gomhptmc",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ma22qwh7-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: wa9y26du; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 360.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 360.9 22.9 20.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ma22qwh7",
   "source_url": "https://omlx.ai/benchmarks/performance/ma22qwh7",
   "tg_tokens": null,
   "tps": 22.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ma22qwh7-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: wa9y26du; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 377.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 377.9 25.7 21.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ma22qwh7",
   "source_url": "https://omlx.ai/benchmarks/performance/ma22qwh7",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lyl8n2ul-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 36.4 GB; TTFT 259409 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 505.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 128k tokens 505.3 PP tok/s 44.5 TG tok/s 259409 TTFT (ms) 36.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lyl8n2ul",
   "source_url": "https://omlx.ai/benchmarks/performance/lyl8n2ul",
   "tg_tokens": null,
   "tps": 44.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-cm9gx3d7-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 36.4 GB; TTFT 246946 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 530.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 128k tokens 530.8 PP tok/s 46.0 TG tok/s 246946 TTFT (ms) 36.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cm9gx3d7",
   "source_url": "https://omlx.ai/benchmarks/performance/cm9gx3d7",
   "tg_tokens": null,
   "tps": 46.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xok3rf3v-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: ay1kxd37, z75oypzj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 926.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 926.4 58.9 34.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xok3rf3v",
   "source_url": "https://omlx.ai/benchmarks/performance/xok3rf3v",
   "tg_tokens": null,
   "tps": 58.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xok3rf3v-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: ay1kxd37, z75oypzj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 934.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 934.1 33.0 35.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xok3rf3v",
   "source_url": "https://omlx.ai/benchmarks/performance/xok3rf3v",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xok3rf3v-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: ay1kxd37, z75oypzj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 911.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 911.0 46.1 35.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xok3rf3v",
   "source_url": "https://omlx.ai/benchmarks/performance/xok3rf3v",
   "tg_tokens": null,
   "tps": 46.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2lo6emaf-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 76.3 GB at this context; duplicate traces of the same session: yrk23pq5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 102.5 18.0 76.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2lo6emaf",
   "source_url": "https://omlx.ai/benchmarks/performance/2lo6emaf",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2lo6emaf-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 76.5 GB at this context; duplicate traces of the same session: yrk23pq5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 129.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 129.4 15.6 76.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2lo6emaf",
   "source_url": "https://omlx.ai/benchmarks/performance/2lo6emaf",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-2ph6dgnp-ctx1024",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: u0l5ji9y; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 117.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 117.8 18.6 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ph6dgnp",
   "source_url": "https://omlx.ai/benchmarks/performance/2ph6dgnp",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-2ph6dgnp-ctx4096",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: u0l5ji9y; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 125.7 18.0 20.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ph6dgnp",
   "source_url": "https://omlx.ai/benchmarks/performance/2ph6dgnp",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-epe30ao2-ctx1024",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 116.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 116.1 18.6 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/epe30ao2",
   "source_url": "https://omlx.ai/benchmarks/performance/epe30ao2",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-epe30ao2-ctx4096",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 125.6 18.1 20.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/epe30ao2",
   "source_url": "https://omlx.ai/benchmarks/performance/epe30ao2",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-8fuvt8sq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 2vb5crao; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 377.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 377.8 25.1 19.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8fuvt8sq",
   "source_url": "https://omlx.ai/benchmarks/performance/8fuvt8sq",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-8fuvt8sq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 2vb5crao; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 386.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 386.1 25.2 21.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8fuvt8sq",
   "source_url": "https://omlx.ai/benchmarks/performance/8fuvt8sq",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-mtphca8t-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: 3jrlfpgk, a66kvw3r, eydsrmdh, i0jj1eit, tfxs7kw1; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 612.1,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 612.1 73.4 11.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mtphca8t",
   "source_url": "https://omlx.ai/benchmarks/performance/mtphca8t",
   "tg_tokens": null,
   "tps": 73.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-mtphca8t-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: 3jrlfpgk, a66kvw3r, eydsrmdh, i0jj1eit, tfxs7kw1; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 662.5,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 662.5 69.5 11.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mtphca8t",
   "source_url": "https://omlx.ai/benchmarks/performance/mtphca8t",
   "tg_tokens": null,
   "tps": 69.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-mtphca8t-ctx8192",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.9 GB at this context; duplicate traces of the same session: 3jrlfpgk, a66kvw3r, eydsrmdh, i0jj1eit, tfxs7kw1; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 640.2,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 8k 640.2 63.1 11.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mtphca8t",
   "source_url": "https://omlx.ai/benchmarks/performance/mtphca8t",
   "tg_tokens": null,
   "tps": 63.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-mtphca8t-ctx16384",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 12.1 GB at this context; duplicate traces of the same session: 3jrlfpgk, a66kvw3r, eydsrmdh, i0jj1eit, tfxs7kw1; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 589.3,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 16k 589.3 55.9 12.1 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mtphca8t",
   "source_url": "https://omlx.ai/benchmarks/performance/mtphca8t",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-mtphca8t-ctx32768",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 12.4 GB at this context; duplicate traces of the same session: 3jrlfpgk, a66kvw3r, eydsrmdh, i0jj1eit, tfxs7kw1; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 504.9,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 32k 504.9 45.8 12.4 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mtphca8t",
   "source_url": "https://omlx.ai/benchmarks/performance/mtphca8t",
   "tg_tokens": null,
   "tps": 45.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-mtphca8t-ctx65536",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 13.2 GB at this context; duplicate traces of the same session: 3jrlfpgk, a66kvw3r, eydsrmdh, i0jj1eit, tfxs7kw1; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 397.5,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 64k 397.5 32.5 13.2 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mtphca8t",
   "source_url": "https://omlx.ai/benchmarks/performance/mtphca8t",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-mtphca8t-ctx131072",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: 3jrlfpgk, a66kvw3r, eydsrmdh, i0jj1eit, tfxs7kw1; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 276.5,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 128k 276.5 21.1 14.6 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mtphca8t",
   "source_url": "https://omlx.ai/benchmarks/performance/mtphca8t",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-mtphca8t-ctx199680",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 16.2 GB at this context; duplicate traces of the same session: 3jrlfpgk, a66kvw3r, eydsrmdh, i0jj1eit, tfxs7kw1; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 208.7,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 195k 208.7 15.4 16.2 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mtphca8t",
   "source_url": "https://omlx.ai/benchmarks/performance/mtphca8t",
   "tg_tokens": null,
   "tps": 15.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-b9a3a78l-ctx16384",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 12.1 GB at this context; duplicate traces of the same session: oaaaq1j3, qttwbvp4; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 712.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 712.0 56.8 12.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b9a3a78l",
   "source_url": "https://omlx.ai/benchmarks/performance/b9a3a78l",
   "tg_tokens": null,
   "tps": 56.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-b9a3a78l-ctx32768",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 12.6 GB at this context; duplicate traces of the same session: oaaaq1j3, qttwbvp4; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 614.9,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 614.9 48.1 12.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b9a3a78l",
   "source_url": "https://omlx.ai/benchmarks/performance/b9a3a78l",
   "tg_tokens": null,
   "tps": 48.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-b9a3a78l-ctx65536",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 13.5 GB at this context; duplicate traces of the same session: oaaaq1j3, qttwbvp4; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 459.3,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 459.3 34.6 13.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b9a3a78l",
   "source_url": "https://omlx.ai/benchmarks/performance/b9a3a78l",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-30h60jj6-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 660.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "gpt-oss-20b-oQ6e-bf16-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 6bit; 1k 660.9 62.4 16.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/30h60jj6",
   "source_url": "https://omlx.ai/benchmarks/performance/30h60jj6",
   "tg_tokens": null,
   "tps": 62.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-30h60jj6-ctx32768",
   "model": "gpt-oss 20B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 601.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "gpt-oss-20b-oQ6e-bf16-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 6bit; 32k 601.7 44.2 17.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/30h60jj6",
   "source_url": "https://omlx.ai/benchmarks/performance/30h60jj6",
   "tg_tokens": null,
   "tps": 44.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-tqk34k6s-ctx4096",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 7ks5r4ja; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 208.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "AREX-2-6bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 6bit; 4k 208.9 14.9 23.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tqk34k6s",
   "source_url": "https://omlx.ai/benchmarks/performance/tqk34k6s",
   "tg_tokens": null,
   "tps": 14.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-tqk34k6s-ctx8192",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 7ks5r4ja; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 202.2,
   "provenance": "community",
   "quant": "6bit",
   "quote": "AREX-2-6bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 6bit; 8k 202.2 12.5 23.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tqk34k6s",
   "source_url": "https://omlx.ai/benchmarks/performance/tqk34k6s",
   "tg_tokens": null,
   "tps": 12.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-tmjykgml-ctx4096",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: infioj96; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 226.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "AREX-2-6bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 6bit; 4k 226.0 15.4 23.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tmjykgml",
   "source_url": "https://omlx.ai/benchmarks/performance/tmjykgml",
   "tg_tokens": null,
   "tps": 15.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-tmjykgml-ctx8192",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: infioj96; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 203.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "AREX-2-6bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 6bit; 8k 203.7 14.2 23.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tmjykgml",
   "source_url": "https://omlx.ai/benchmarks/performance/tmjykgml",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ushfpucl-ctx1024",
   "model": "JOSIE-2-4B",
   "notes": "oMLX community benchmark trace; peak memory 9.7 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 808.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "JOSIE-2-4B-Preview on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 unknown; 1k 808.0 34.7 9.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ushfpucl",
   "source_url": "https://omlx.ai/benchmarks/performance/ushfpucl",
   "tg_tokens": null,
   "tps": 34.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ushfpucl-ctx4096",
   "model": "JOSIE-2-4B",
   "notes": "oMLX community benchmark trace; peak memory 11.4 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 864.3,
   "provenance": "community",
   "quant": "unknown",
   "quote": "JOSIE-2-4B-Preview on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 unknown; 4k 864.3 53.7 11.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ushfpucl",
   "source_url": "https://omlx.ai/benchmarks/performance/ushfpucl",
   "tg_tokens": null,
   "tps": 53.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-hqsds8p7-ctx1024",
   "model": "JOSIE-2-9B-OSS",
   "notes": "oMLX community benchmark trace; peak memory 17.5 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 500.6,
   "provenance": "community",
   "quant": "unknown",
   "quote": "JOSIE-2-9B-OSS on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 unknown; 1k 500.6 20.4 17.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hqsds8p7",
   "source_url": "https://omlx.ai/benchmarks/performance/hqsds8p7",
   "tg_tokens": null,
   "tps": 20.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-hqsds8p7-ctx4096",
   "model": "JOSIE-2-9B-OSS",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 508.6,
   "provenance": "community",
   "quant": "unknown",
   "quote": "JOSIE-2-9B-OSS on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 unknown; 4k 508.6 19.9 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hqsds8p7",
   "source_url": "https://omlx.ai/benchmarks/performance/hqsds8p7",
   "tg_tokens": null,
   "tps": 19.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-m46vo6hc-ctx1024",
   "model": "Laguna XS 2.1",
   "notes": "MLX; oMLX community benchmark trace; peak memory 33.7 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 589.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Laguna-XS-2.1-MLX-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 1k 589.7 46.1 33.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m46vo6hc",
   "source_url": "https://omlx.ai/benchmarks/performance/m46vo6hc",
   "tg_tokens": null,
   "tps": 46.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-m46vo6hc-ctx4096",
   "model": "Laguna XS 2.1",
   "notes": "MLX; oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 593.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Laguna-XS-2.1-MLX-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k 593.7 42.8 33.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m46vo6hc",
   "source_url": "https://omlx.ai/benchmarks/performance/m46vo6hc",
   "tg_tokens": null,
   "tps": 42.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-d5rs00os-ctx1024",
   "model": "Llama-3.2-3B-Instruct-uncensored",
   "notes": "oMLX community benchmark trace; peak memory 3.1 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1246.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Llama-3.2-3B-Instruct-uncensored-6bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 1k 1,246 103.9 3.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d5rs00os",
   "source_url": "https://omlx.ai/benchmarks/performance/d5rs00os",
   "tg_tokens": null,
   "tps": 103.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-d5rs00os-ctx4096",
   "model": "Llama-3.2-3B-Instruct-uncensored",
   "notes": "oMLX community benchmark trace; peak memory 3.5 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1171.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Llama-3.2-3B-Instruct-uncensored-6bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 4k 1,171 92.0 3.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d5rs00os",
   "source_url": "https://omlx.ai/benchmarks/performance/d5rs00os",
   "tg_tokens": null,
   "tps": 92.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-y541vzvs-ctx1024",
   "model": "MagenticBrain",
   "notes": "oMLX community benchmark trace; peak memory 15.3 GB at this context; duplicate traces of the same session: yusou20x; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 212.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "MagenticBrain-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 1k 212.4 20.8 15.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y541vzvs",
   "source_url": "https://omlx.ai/benchmarks/performance/y541vzvs",
   "tg_tokens": null,
   "tps": 20.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-y541vzvs-ctx4096",
   "model": "MagenticBrain",
   "notes": "oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: yusou20x; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 207.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "MagenticBrain-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k 207.3 19.7 15.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y541vzvs",
   "source_url": "https://omlx.ai/benchmarks/performance/y541vzvs",
   "tg_tokens": null,
   "tps": 19.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-pwq3mhgx-ctx1024",
   "model": "Nex-N2-mini-MLX-VLM-4bit-MTP",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: q825rc4n; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 678.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Nex-N2-mini-MLX-VLM-4bit-MTP on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 678.2 43.6 21.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pwq3mhgx",
   "source_url": "https://omlx.ai/benchmarks/performance/pwq3mhgx",
   "tg_tokens": null,
   "tps": 43.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-pwq3mhgx-ctx4096",
   "model": "Nex-N2-mini-MLX-VLM-4bit-MTP",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: q825rc4n; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 756.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Nex-N2-mini-MLX-VLM-4bit-MTP on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 756.4 46.3 23.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pwq3mhgx",
   "source_url": "https://omlx.ai/benchmarks/performance/pwq3mhgx",
   "tg_tokens": null,
   "tps": 46.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-9mx97zaj-ctx1024",
   "model": "Nex-N2.5-mini",
   "notes": "MLX; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: o9puo9ej; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 692.6,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Nex-N2.5-mini-MLX-oQ4 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 692.6 63.6 20.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9mx97zaj",
   "source_url": "https://omlx.ai/benchmarks/performance/9mx97zaj",
   "tg_tokens": null,
   "tps": 63.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-9mx97zaj-ctx4096",
   "model": "Nex-N2.5-mini",
   "notes": "MLX; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: o9puo9ej; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 767.8,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Nex-N2.5-mini-MLX-oQ4 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 767.8 57.1 22.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9mx97zaj",
   "source_url": "https://omlx.ai/benchmarks/performance/9mx97zaj",
   "tg_tokens": null,
   "tps": 57.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-nk8ur97f-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: 0xly5434; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 730.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 730.4 58.9 19.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nk8ur97f",
   "source_url": "https://omlx.ai/benchmarks/performance/nk8ur97f",
   "tg_tokens": null,
   "tps": 58.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-nk8ur97f-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: 0xly5434; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 720.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 720.7 50.3 20.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nk8ur97f",
   "source_url": "https://omlx.ai/benchmarks/performance/nk8ur97f",
   "tg_tokens": null,
   "tps": 50.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-src4w323-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 36.3 GB at this context; duplicate traces of the same session: 011yt0nm; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 752.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 1k 752.8 51.4 36.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/src4w323",
   "source_url": "https://omlx.ai/benchmarks/performance/src4w323",
   "tg_tokens": null,
   "tps": 51.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-src4w323-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: 011yt0nm; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 742.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k 742.0 43.9 37.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/src4w323",
   "source_url": "https://omlx.ai/benchmarks/performance/src4w323",
   "tg_tokens": null,
   "tps": 43.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-v3tq2ara-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: 96brug5h; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1006.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 1,006 31.1 5.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v3tq2ara",
   "source_url": "https://omlx.ai/benchmarks/performance/v3tq2ara",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-v3tq2ara-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: 96brug5h; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1077.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,077 31.0 6.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v3tq2ara",
   "source_url": "https://omlx.ai/benchmarks/performance/v3tq2ara",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-upn7xv5e-ctx1024",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: nw60zvod; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4666.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,666 154.2 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/upn7xv5e",
   "source_url": "https://omlx.ai/benchmarks/performance/upn7xv5e",
   "tg_tokens": null,
   "tps": 154.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-upn7xv5e-ctx32768",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: nw60zvod; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4914.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 4,914 124.9 20.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/upn7xv5e",
   "source_url": "https://omlx.ai/benchmarks/performance/upn7xv5e",
   "tg_tokens": null,
   "tps": 124.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-upn7xv5e-ctx65536",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: nw60zvod; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3674.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 3,674 103.5 21.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/upn7xv5e",
   "source_url": "https://omlx.ai/benchmarks/performance/upn7xv5e",
   "tg_tokens": null,
   "tps": 103.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-upn7xv5e-ctx131072",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: nw60zvod; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2414.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,414 76.4 22.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/upn7xv5e",
   "source_url": "https://omlx.ai/benchmarks/performance/upn7xv5e",
   "tg_tokens": null,
   "tps": 76.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-upn7xv5e-ctx199680",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: nw60zvod; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1827.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 1,827 63.5 23.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/upn7xv5e",
   "source_url": "https://omlx.ai/benchmarks/performance/upn7xv5e",
   "tg_tokens": null,
   "tps": 63.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-c5ro24yy-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: 31389lug, 6eydt27q, c7shxii8, jrdw1aw9, l9ol1e8g, ot421q0f, tv8aob1b; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 379.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 379.9 50.9 5.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c5ro24yy",
   "source_url": "https://omlx.ai/benchmarks/performance/c5ro24yy",
   "tg_tokens": null,
   "tps": 50.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-c5ro24yy-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: 31389lug, 6eydt27q, c7shxii8, jrdw1aw9, l9ol1e8g, ot421q0f, tv8aob1b; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 379.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 379.1 49.8 6.5 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c5ro24yy",
   "source_url": "https://omlx.ai/benchmarks/performance/c5ro24yy",
   "tg_tokens": null,
   "tps": 49.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-c5ro24yy-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: 31389lug, 6eydt27q, c7shxii8, jrdw1aw9, l9ol1e8g, ot421q0f, tv8aob1b; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 373.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 8k 373.9 48.1 6.9 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c5ro24yy",
   "source_url": "https://omlx.ai/benchmarks/performance/c5ro24yy",
   "tg_tokens": null,
   "tps": 48.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-c5ro24yy-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: 31389lug, 6eydt27q, c7shxii8, jrdw1aw9, l9ol1e8g, ot421q0f, tv8aob1b; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 355.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 16k 355.8 45.1 7.0 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c5ro24yy",
   "source_url": "https://omlx.ai/benchmarks/performance/c5ro24yy",
   "tg_tokens": null,
   "tps": 45.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-c5ro24yy-ctx32768",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.5 GB at this context; duplicate traces of the same session: 31389lug, 6eydt27q, c7shxii8, jrdw1aw9, l9ol1e8g, ot421q0f, tv8aob1b; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 326.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 32k 326.1 39.9 7.5 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c5ro24yy",
   "source_url": "https://omlx.ai/benchmarks/performance/c5ro24yy",
   "tg_tokens": null,
   "tps": 39.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-c5ro24yy-ctx65536",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.5 GB at this context; duplicate traces of the same session: 31389lug, 6eydt27q, c7shxii8, jrdw1aw9, l9ol1e8g, ot421q0f, tv8aob1b; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 280.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 64k 280.1 31.9 8.5 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c5ro24yy",
   "source_url": "https://omlx.ai/benchmarks/performance/c5ro24yy",
   "tg_tokens": null,
   "tps": 31.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-c5ro24yy-ctx131072",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.5 GB at this context; duplicate traces of the same session: 31389lug, 6eydt27q, c7shxii8, jrdw1aw9, l9ol1e8g, ot421q0f, tv8aob1b; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 217.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 128k 217.5 22.2 10.5 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c5ro24yy",
   "source_url": "https://omlx.ai/benchmarks/performance/c5ro24yy",
   "tg_tokens": null,
   "tps": 22.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-c5ro24yy-ctx199680",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 12.6 GB at this context; duplicate traces of the same session: 31389lug, 6eydt27q, c7shxii8, jrdw1aw9, l9ol1e8g, ot421q0f, tv8aob1b; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 176.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 195k 176.2 16.7 12.6 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c5ro24yy",
   "source_url": "https://omlx.ai/benchmarks/performance/c5ro24yy",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-6fkb584r-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: cd3nwnb5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 331.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 331.2 27.1 17.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6fkb584r",
   "source_url": "https://omlx.ai/benchmarks/performance/6fkb584r",
   "tg_tokens": null,
   "tps": 27.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-6fkb584r-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: cd3nwnb5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 331.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 331.3 31.2 18.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6fkb584r",
   "source_url": "https://omlx.ai/benchmarks/performance/6fkb584r",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-f1k29l6s-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: r9dxcenn; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 451.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 451.7 26.0 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f1k29l6s",
   "source_url": "https://omlx.ai/benchmarks/performance/f1k29l6s",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-f1k29l6s-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: r9dxcenn; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 461.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 461.8 29.0 19.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f1k29l6s",
   "source_url": "https://omlx.ai/benchmarks/performance/f1k29l6s",
   "tg_tokens": null,
   "tps": 29.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-lrmswmli-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: 524z3irh; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 320.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 320.9 28.2 17.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lrmswmli",
   "source_url": "https://omlx.ai/benchmarks/performance/lrmswmli",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-lrmswmli-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 524z3irh; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 317.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 317.0 29.6 18.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lrmswmli",
   "source_url": "https://omlx.ai/benchmarks/performance/lrmswmli",
   "tg_tokens": null,
   "tps": 29.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-l4lqirza-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: fkm4yvv5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 451.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 451.8 22.7 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l4lqirza",
   "source_url": "https://omlx.ai/benchmarks/performance/l4lqirza",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-l4lqirza-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: fkm4yvv5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 500.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 500.7 23.1 19.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l4lqirza",
   "source_url": "https://omlx.ai/benchmarks/performance/l4lqirza",
   "tg_tokens": null,
   "tps": 23.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-h2f3sm17-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 31.0 GB at this context; duplicate traces of the same session: b2qjnpxt; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 348.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 348.8 20.7 31.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h2f3sm17",
   "source_url": "https://omlx.ai/benchmarks/performance/h2f3sm17",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-h2f3sm17-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 31.9 GB at this context; duplicate traces of the same session: b2qjnpxt; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 390.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 390.4 23.4 31.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h2f3sm17",
   "source_url": "https://omlx.ai/benchmarks/performance/h2f3sm17",
   "tg_tokens": null,
   "tps": 23.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-d4jakxpj-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: iq6mg1fm; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 326.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 326.5 20.4 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d4jakxpj",
   "source_url": "https://omlx.ai/benchmarks/performance/d4jakxpj",
   "tg_tokens": null,
   "tps": 20.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-d4jakxpj-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: iq6mg1fm; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 282.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 282.1 20.7 27.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d4jakxpj",
   "source_url": "https://omlx.ai/benchmarks/performance/d4jakxpj",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-mpqudsu0-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: fmooh05x; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 287.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 287.2 18.5 26.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mpqudsu0",
   "source_url": "https://omlx.ai/benchmarks/performance/mpqudsu0",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-mpqudsu0-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: fmooh05x; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 295.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 295.5 20.2 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mpqudsu0",
   "source_url": "https://omlx.ai/benchmarks/performance/mpqudsu0",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-91x0czmk-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 31.0 GB at this context; duplicate traces of the same session: f4o4m23l; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 289.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 289.5 17.3 31.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/91x0czmk",
   "source_url": "https://omlx.ai/benchmarks/performance/91x0czmk",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-91x0czmk-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 31.9 GB at this context; duplicate traces of the same session: f4o4m23l; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 317.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 317.9 17.9 31.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/91x0czmk",
   "source_url": "https://omlx.ai/benchmarks/performance/91x0czmk",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-h8494skp-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 31.0 GB at this context; duplicate traces of the same session: 5a540s30; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 312.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 312.8 18.7 31.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h8494skp",
   "source_url": "https://omlx.ai/benchmarks/performance/h8494skp",
   "tg_tokens": null,
   "tps": 18.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-h8494skp-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 31.9 GB at this context; duplicate traces of the same session: 5a540s30; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 354.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 354.0 19.8 31.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h8494skp",
   "source_url": "https://omlx.ai/benchmarks/performance/h8494skp",
   "tg_tokens": null,
   "tps": 19.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9rse1okh-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace, single run at 8k context; peak memory 21.3 GB; TTFT 22464 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 364.7,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k tokens 364.7 PP tok/s 20.2 TG tok/s 22464 TTFT (ms) 21.3 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9rse1okh",
   "source_url": "https://omlx.ai/benchmarks/performance/9rse1okh",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-zwpi70zw-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 27.7 GB; TTFT 11554 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 354.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-Uncensored-oQ6e-mtp on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 6bit; 4k tokens 354.5 PP tok/s 37.9 TG tok/s 11554 TTFT (ms) 27.7 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zwpi70zw",
   "source_url": "https://omlx.ai/benchmarks/performance/zwpi70zw",
   "tg_tokens": null,
   "tps": 37.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-r3sro0pw-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 9vsyc4f2, bqlkpcfi, cr6sawma, gpwa8kg3, qph6aqpz, vf36ve5b, xtk6sbkw; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 223.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 1k 223.0 36.6 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r3sro0pw",
   "source_url": "https://omlx.ai/benchmarks/performance/r3sro0pw",
   "tg_tokens": null,
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-r3sro0pw-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: 9vsyc4f2, bqlkpcfi, cr6sawma, gpwa8kg3, qph6aqpz, vf36ve5b, xtk6sbkw; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 233.1,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 4k 233.1 38.1 23.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r3sro0pw",
   "source_url": "https://omlx.ai/benchmarks/performance/r3sro0pw",
   "tg_tokens": null,
   "tps": 38.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-r3sro0pw-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: 9vsyc4f2, bqlkpcfi, cr6sawma, gpwa8kg3, qph6aqpz, vf36ve5b, xtk6sbkw; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 230.9,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 8k 230.9 36.7 24.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r3sro0pw",
   "source_url": "https://omlx.ai/benchmarks/performance/r3sro0pw",
   "tg_tokens": null,
   "tps": 36.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-r3sro0pw-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: 9vsyc4f2, bqlkpcfi, cr6sawma, gpwa8kg3, qph6aqpz, vf36ve5b, xtk6sbkw; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 221.6,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 16k 221.6 44.4 25.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r3sro0pw",
   "source_url": "https://omlx.ai/benchmarks/performance/r3sro0pw",
   "tg_tokens": null,
   "tps": 44.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-r3sro0pw-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: 9vsyc4f2, bqlkpcfi, cr6sawma, gpwa8kg3, qph6aqpz, vf36ve5b, xtk6sbkw; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 203.4,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 32k 203.4 64.1 26.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r3sro0pw",
   "source_url": "https://omlx.ai/benchmarks/performance/r3sro0pw",
   "tg_tokens": null,
   "tps": 64.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-r3sro0pw-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.7 GB at this context; duplicate traces of the same session: 9vsyc4f2, bqlkpcfi, cr6sawma, gpwa8kg3, qph6aqpz, vf36ve5b, xtk6sbkw; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 173.5,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 64k 173.5 34.0 28.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r3sro0pw",
   "source_url": "https://omlx.ai/benchmarks/performance/r3sro0pw",
   "tg_tokens": null,
   "tps": 34.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-r3sro0pw-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.2 GB at this context; duplicate traces of the same session: 9vsyc4f2, bqlkpcfi, cr6sawma, gpwa8kg3, qph6aqpz, vf36ve5b, xtk6sbkw; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 133.2,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 128k 133.2 29.1 33.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r3sro0pw",
   "source_url": "https://omlx.ai/benchmarks/performance/r3sro0pw",
   "tg_tokens": null,
   "tps": 29.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-r3sro0pw-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.1 GB at this context; duplicate traces of the same session: 9vsyc4f2, bqlkpcfi, cr6sawma, gpwa8kg3, qph6aqpz, vf36ve5b, xtk6sbkw; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 108.1,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 195k 108.1 19.3 38.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r3sro0pw",
   "source_url": "https://omlx.ai/benchmarks/performance/r3sro0pw",
   "tg_tokens": null,
   "tps": 19.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fx8bh7lr-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 36.4 GB; TTFT 246431 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 531.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 128k tokens 531.9 PP tok/s 46.8 TG tok/s 246431 TTFT (ms) 36.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fx8bh7lr",
   "source_url": "https://omlx.ai/benchmarks/performance/fx8bh7lr",
   "tg_tokens": null,
   "tps": 46.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-yr7rg68y-ctx1024",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 116.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 116.1 18.6 16.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yr7rg68y",
   "source_url": "https://omlx.ai/benchmarks/performance/yr7rg68y",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-yr7rg68y-ctx4096",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 125.6 18.1 20.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yr7rg68y",
   "source_url": "https://omlx.ai/benchmarks/performance/yr7rg68y",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-odb8f0iw-ctx4096",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 0nyr2zf2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 191.5,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 5bit; 4k 191.5 14.8 20.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/odb8f0iw",
   "source_url": "https://omlx.ai/benchmarks/performance/odb8f0iw",
   "tg_tokens": null,
   "tps": 14.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-odb8f0iw-ctx8192",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: 0nyr2zf2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 176.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 5bit; 8k 176.0 13.7 20.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/odb8f0iw",
   "source_url": "https://omlx.ai/benchmarks/performance/odb8f0iw",
   "tg_tokens": null,
   "tps": 13.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-qticja8i-ctx4096",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: qm9rzr31; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 217.2,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 5bit; 4k 217.2 17.8 20.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qticja8i",
   "source_url": "https://omlx.ai/benchmarks/performance/qticja8i",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-qticja8i-ctx8192",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: qm9rzr31; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 214.4,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 5bit; 8k 214.4 15.9 20.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qticja8i",
   "source_url": "https://omlx.ai/benchmarks/performance/qticja8i",
   "tg_tokens": null,
   "tps": 15.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-kcacxkyk-ctx4096",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: c6sv33tv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 166.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "AREX-2-6bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 6bit; 4k 166.3 9.9 24.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kcacxkyk",
   "source_url": "https://omlx.ai/benchmarks/performance/kcacxkyk",
   "tg_tokens": null,
   "tps": 9.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-kcacxkyk-ctx8192",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 25.6 GB at this context; duplicate traces of the same session: c6sv33tv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 162.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "AREX-2-6bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 6bit; 8k 162.6 10.2 25.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kcacxkyk",
   "source_url": "https://omlx.ai/benchmarks/performance/kcacxkyk",
   "tg_tokens": null,
   "tps": 10.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-i1z1i0hh-ctx1024",
   "model": "Artemis-31B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 31.8 GB at this context; duplicate traces of the same session: it22qx64; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 90.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Artemis-31B-v1.1-mlx-8Bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 1k 90.5 9.5 31.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i1z1i0hh",
   "source_url": "https://omlx.ai/benchmarks/performance/i1z1i0hh",
   "tg_tokens": null,
   "tps": 9.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-i1z1i0hh-ctx4096",
   "model": "Artemis-31B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 33.1 GB at this context; duplicate traces of the same session: it22qx64; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 90.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Artemis-31B-v1.1-mlx-8Bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k 90.8 9.3 33.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i1z1i0hh",
   "source_url": "https://omlx.ai/benchmarks/performance/i1z1i0hh",
   "tg_tokens": null,
   "tps": 9.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ozlik4c2-ctx1024",
   "model": "Artemis-31B",
   "notes": "oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: g2w4xh83; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 90.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Artemis-31B-v1.1-oQ4e on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 90.7 14.0 18.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ozlik4c2",
   "source_url": "https://omlx.ai/benchmarks/performance/ozlik4c2",
   "tg_tokens": null,
   "tps": 14.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ozlik4c2-ctx4096",
   "model": "Artemis-31B",
   "notes": "oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: g2w4xh83; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 90.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Artemis-31B-v1.1-oQ4e on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 90.5 13.8 19.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ozlik4c2",
   "source_url": "https://omlx.ai/benchmarks/performance/ozlik4c2",
   "tg_tokens": null,
   "tps": 13.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-2zfjsi88-ctx1024",
   "model": "Fara1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.7 GB at this context; duplicate traces of the same session: tf20ax02; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 390.4,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Fara1.5-9B-OptiQ-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 390.4 47.6 7.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2zfjsi88",
   "source_url": "https://omlx.ai/benchmarks/performance/2zfjsi88",
   "tg_tokens": null,
   "tps": 47.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-2zfjsi88-ctx4096",
   "model": "Fara1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 9.3 GB at this context; duplicate traces of the same session: tf20ax02; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 376.2,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Fara1.5-9B-OptiQ-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 376.2 46.1 9.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2zfjsi88",
   "source_url": "https://omlx.ai/benchmarks/performance/2zfjsi88",
   "tg_tokens": null,
   "tps": 46.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-gzzsrs8p-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 176.5 GB; TTFT 7648 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1071.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M1 (80c); M1 (80c) \u00b7 256 GB \u00b7 4bit; 8k tokens 1,071 PP tok/s 53.5 TG tok/s 7648 TTFT (ms) 176.5 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gzzsrs8p",
   "source_url": "https://omlx.ai/benchmarks/performance/gzzsrs8p",
   "tg_tokens": null,
   "tps": 53.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1",
   "id": "omlx-u99s41j6-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 176.3 GB at this context; duplicate traces of the same session: nqbt76x6; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1057.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M1 (80c); M1 (80c) \u00b7 256 GB \u00b7 4bit; 8k 1,057 36.1 176.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u99s41j6",
   "source_url": "https://omlx.ai/benchmarks/performance/u99s41j6",
   "tg_tokens": null,
   "tps": 36.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-u99s41j6-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 176.4 GB at this context; duplicate traces of the same session: nqbt76x6; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1050.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M1 (80c); M1 (80c) \u00b7 256 GB \u00b7 4bit; 16k 1,050 36.4 176.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u99s41j6",
   "source_url": "https://omlx.ai/benchmarks/performance/u99s41j6",
   "tg_tokens": null,
   "tps": 36.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ghe776qo-ctx1024",
   "model": "JOSIE-2-4B",
   "notes": "oMLX community benchmark trace; peak memory 9.7 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 808.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "JOSIE-2-4B-Preview on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 unknown; 1k 808.0 34.7 9.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ghe776qo",
   "source_url": "https://omlx.ai/benchmarks/performance/ghe776qo",
   "tg_tokens": null,
   "tps": 34.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ghe776qo-ctx4096",
   "model": "JOSIE-2-4B",
   "notes": "oMLX community benchmark trace; peak memory 11.4 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 864.3,
   "provenance": "community",
   "quant": "unknown",
   "quote": "JOSIE-2-4B-Preview on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 unknown; 4k 864.3 53.7 11.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ghe776qo",
   "source_url": "https://omlx.ai/benchmarks/performance/ghe776qo",
   "tg_tokens": null,
   "tps": 53.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-v4iap2nk-ctx1024",
   "model": "JOSIE-2-9B-OSS",
   "notes": "oMLX community benchmark trace; peak memory 17.5 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 500.6,
   "provenance": "community",
   "quant": "unknown",
   "quote": "JOSIE-2-9B-OSS on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 unknown; 1k 500.6 20.4 17.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v4iap2nk",
   "source_url": "https://omlx.ai/benchmarks/performance/v4iap2nk",
   "tg_tokens": null,
   "tps": 20.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-v4iap2nk-ctx4096",
   "model": "JOSIE-2-9B-OSS",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 508.6,
   "provenance": "community",
   "quant": "unknown",
   "quote": "JOSIE-2-9B-OSS on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 unknown; 4k 508.6 19.9 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v4iap2nk",
   "source_url": "https://omlx.ai/benchmarks/performance/v4iap2nk",
   "tg_tokens": null,
   "tps": 19.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-utn87o95-ctx1024",
   "model": "Laguna XS 2.1",
   "notes": "MLX; oMLX community benchmark trace; peak memory 33.7 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 589.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Laguna-XS-2.1-MLX-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 1k 589.7 46.1 33.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/utn87o95",
   "source_url": "https://omlx.ai/benchmarks/performance/utn87o95",
   "tg_tokens": null,
   "tps": 46.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-utn87o95-ctx4096",
   "model": "Laguna XS 2.1",
   "notes": "MLX; oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 593.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Laguna-XS-2.1-MLX-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k 593.7 42.8 33.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/utn87o95",
   "source_url": "https://omlx.ai/benchmarks/performance/utn87o95",
   "tg_tokens": null,
   "tps": 42.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-apeut5ak-ctx1024",
   "model": "Llama-3.2-3B-Instruct-uncensored",
   "notes": "oMLX community benchmark trace; peak memory 3.1 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1246.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Llama-3.2-3B-Instruct-uncensored-6bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 1k 1,246 103.9 3.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/apeut5ak",
   "source_url": "https://omlx.ai/benchmarks/performance/apeut5ak",
   "tg_tokens": null,
   "tps": 103.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-apeut5ak-ctx4096",
   "model": "Llama-3.2-3B-Instruct-uncensored",
   "notes": "oMLX community benchmark trace; peak memory 3.5 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1171.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Llama-3.2-3B-Instruct-uncensored-6bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 4k 1,171 92.0 3.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/apeut5ak",
   "source_url": "https://omlx.ai/benchmarks/performance/apeut5ak",
   "tg_tokens": null,
   "tps": 92.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-535br4v9-ctx199680",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace, single run at 195k context; peak memory 32.4 GB; TTFT 945099 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 211.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ4e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 195k tokens 211.6 PP tok/s 13.8 TG tok/s 945099 TTFT (ms) 32.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/535br4v9",
   "source_url": "https://omlx.ai/benchmarks/performance/535br4v9",
   "tg_tokens": null,
   "tps": 13.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-46p42zbh-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 22.8 GB; TTFT 44325 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 739.3,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 32k tokens 739.3 PP tok/s 18.1 TG tok/s 44325 TTFT (ms) 22.8 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/46p42zbh",
   "source_url": "https://omlx.ai/benchmarks/performance/46p42zbh",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-89hih3v1-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: 4qco4z4w, 8dk3eodj, lcl3101z, m0wcvw91, nuu208vx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 428.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 428.0 60.1 22.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/89hih3v1",
   "source_url": "https://omlx.ai/benchmarks/performance/89hih3v1",
   "tg_tokens": null,
   "tps": 60.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-89hih3v1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: 4qco4z4w, 8dk3eodj, lcl3101z, m0wcvw91, nuu208vx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 447.5,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 447.5 84.8 25.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/89hih3v1",
   "source_url": "https://omlx.ai/benchmarks/performance/89hih3v1",
   "tg_tokens": null,
   "tps": 84.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-89hih3v1-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: 4qco4z4w, 8dk3eodj, lcl3101z, m0wcvw91, nuu208vx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 524.3,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 524.3 84.2 25.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/89hih3v1",
   "source_url": "https://omlx.ai/benchmarks/performance/89hih3v1",
   "tg_tokens": null,
   "tps": 84.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-89hih3v1-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: 4qco4z4w, 8dk3eodj, lcl3101z, m0wcvw91, nuu208vx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 518.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 518.0 96.9 26.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/89hih3v1",
   "source_url": "https://omlx.ai/benchmarks/performance/89hih3v1",
   "tg_tokens": null,
   "tps": 96.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-89hih3v1-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.2 GB at this context; duplicate traces of the same session: 4qco4z4w, 8dk3eodj, lcl3101z, m0wcvw91, nuu208vx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 479.6,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 479.6 121.1 28.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/89hih3v1",
   "source_url": "https://omlx.ai/benchmarks/performance/89hih3v1",
   "tg_tokens": null,
   "tps": 121.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-89hih3v1-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.4 GB at this context; duplicate traces of the same session: 4qco4z4w, 8dk3eodj, lcl3101z, m0wcvw91, nuu208vx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 405.4,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 405.4 71.0 31.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/89hih3v1",
   "source_url": "https://omlx.ai/benchmarks/performance/89hih3v1",
   "tg_tokens": null,
   "tps": 71.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-9c9muif2-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 153.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 1k 153.2 20.1 18.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9c9muif2",
   "source_url": "https://omlx.ai/benchmarks/performance/9c9muif2",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-9c9muif2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 148.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 4k 148.8 30.3 22.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9c9muif2",
   "source_url": "https://omlx.ai/benchmarks/performance/9c9muif2",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-9c9muif2-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 145.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 8k 145.3 36.3 23.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9c9muif2",
   "source_url": "https://omlx.ai/benchmarks/performance/9c9muif2",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-9c9muif2-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 136.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 16k 136.4 27.3 24.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9c9muif2",
   "source_url": "https://omlx.ai/benchmarks/performance/9c9muif2",
   "tg_tokens": null,
   "tps": 27.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-9c9muif2-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.6 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 121.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 32k 121.3 48.6 25.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9c9muif2",
   "source_url": "https://omlx.ai/benchmarks/performance/9c9muif2",
   "tg_tokens": null,
   "tps": 48.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-rzz0f236-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: 53rrki50, 7niqkv2m, 8f2lkamm, da0ypsc9, jy032tpx, lfbxmyyu, smmmhnhc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 807.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 807.2 59.1 26.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rzz0f236",
   "source_url": "https://omlx.ai/benchmarks/performance/rzz0f236",
   "tg_tokens": null,
   "tps": 59.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rzz0f236-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: 53rrki50, 7niqkv2m, 8f2lkamm, da0ypsc9, jy032tpx, lfbxmyyu, smmmhnhc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 907.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 907.4 54.6 27.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rzz0f236",
   "source_url": "https://omlx.ai/benchmarks/performance/rzz0f236",
   "tg_tokens": null,
   "tps": 54.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rzz0f236-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: 53rrki50, 7niqkv2m, 8f2lkamm, da0ypsc9, jy032tpx, lfbxmyyu, smmmhnhc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 736.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 736.6 51.7 28.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rzz0f236",
   "source_url": "https://omlx.ai/benchmarks/performance/rzz0f236",
   "tg_tokens": null,
   "tps": 51.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rzz0f236-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: 53rrki50, 7niqkv2m, 8f2lkamm, da0ypsc9, jy032tpx, lfbxmyyu, smmmhnhc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 712.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 712.9 59.8 28.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rzz0f236",
   "source_url": "https://omlx.ai/benchmarks/performance/rzz0f236",
   "tg_tokens": null,
   "tps": 59.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rzz0f236-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: 53rrki50, 7niqkv2m, 8f2lkamm, da0ypsc9, jy032tpx, lfbxmyyu, smmmhnhc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 709.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 709.2 63.7 29.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rzz0f236",
   "source_url": "https://omlx.ai/benchmarks/performance/rzz0f236",
   "tg_tokens": null,
   "tps": 63.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rzz0f236-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: 53rrki50, 7niqkv2m, 8f2lkamm, da0ypsc9, jy032tpx, lfbxmyyu, smmmhnhc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 636.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k 636.8 44.7 32.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rzz0f236",
   "source_url": "https://omlx.ai/benchmarks/performance/rzz0f236",
   "tg_tokens": null,
   "tps": 44.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-rzz0f236-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: 53rrki50, 7niqkv2m, 8f2lkamm, da0ypsc9, jy032tpx, lfbxmyyu, smmmhnhc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 526.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 128k 526.1 45.7 36.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rzz0f236",
   "source_url": "https://omlx.ai/benchmarks/performance/rzz0f236",
   "tg_tokens": null,
   "tps": 45.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rzz0f236-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 42.1 GB at this context; duplicate traces of the same session: 53rrki50, 7niqkv2m, 8f2lkamm, da0ypsc9, jy032tpx, lfbxmyyu, smmmhnhc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 435.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 195k 435.4 31.6 42.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rzz0f236",
   "source_url": "https://omlx.ai/benchmarks/performance/rzz0f236",
   "tg_tokens": null,
   "tps": 31.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fuaykzg3-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: f3osamrt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1660.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,660 52.6 71.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fuaykzg3",
   "source_url": "https://omlx.ai/benchmarks/performance/fuaykzg3",
   "tg_tokens": null,
   "tps": 52.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fuaykzg3-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.9 GB at this context; duplicate traces of the same session: f3osamrt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2540.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,540 51.5 72.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fuaykzg3",
   "source_url": "https://omlx.ai/benchmarks/performance/fuaykzg3",
   "tg_tokens": null,
   "tps": 51.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-oyeg7w2c-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1750.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,750 62.5 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oyeg7w2c",
   "source_url": "https://omlx.ai/benchmarks/performance/oyeg7w2c",
   "tg_tokens": null,
   "tps": 62.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-oyeg7w2c-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2930.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,930 104.6 72.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oyeg7w2c",
   "source_url": "https://omlx.ai/benchmarks/performance/oyeg7w2c",
   "tg_tokens": null,
   "tps": 104.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-judmkmjl-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 421.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 421.4 19.8 39.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/judmkmjl",
   "source_url": "https://omlx.ai/benchmarks/performance/judmkmjl",
   "tg_tokens": null,
   "tps": 19.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-judmkmjl-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 41.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 948.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 948.3 16.1 41.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/judmkmjl",
   "source_url": "https://omlx.ai/benchmarks/performance/judmkmjl",
   "tg_tokens": null,
   "tps": 16.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4i252osi-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1097.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,097 79.3 70.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4i252osi",
   "source_url": "https://omlx.ai/benchmarks/performance/4i252osi",
   "tg_tokens": null,
   "tps": 79.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4i252osi-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2143.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,143 75.0 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4i252osi",
   "source_url": "https://omlx.ai/benchmarks/performance/4i252osi",
   "tg_tokens": null,
   "tps": 75.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4i252osi-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2920.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,920 110.1 72.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4i252osi",
   "source_url": "https://omlx.ai/benchmarks/performance/4i252osi",
   "tg_tokens": null,
   "tps": 110.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4i252osi-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2855.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,855 69.6 73.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4i252osi",
   "source_url": "https://omlx.ai/benchmarks/performance/4i252osi",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-p10ixo47-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 25.8 GB; TTFT 19641 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 834.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k tokens 834.2 PP tok/s 13.5 TG tok/s 19641 TTFT (ms) 25.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p10ixo47",
   "source_url": "https://omlx.ai/benchmarks/performance/p10ixo47",
   "tg_tokens": null,
   "tps": 13.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-i90gequ0-ctx1024",
   "model": "Rocinante-X-12B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 12.8 GB at this context; duplicate traces of the same session: 5xmpfkk1; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 257.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Rocinante-X-12B-v1-absolute-heresy-mlx-8Bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 1k 257.7 25.5 12.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i90gequ0",
   "source_url": "https://omlx.ai/benchmarks/performance/i90gequ0",
   "tg_tokens": null,
   "tps": 25.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-i90gequ0-ctx4096",
   "model": "Rocinante-X-12B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 13.3 GB at this context; duplicate traces of the same session: 5xmpfkk1; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 251.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Rocinante-X-12B-v1-absolute-heresy-mlx-8Bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k 251.2 23.8 13.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i90gequ0",
   "source_url": "https://omlx.ai/benchmarks/performance/i90gequ0",
   "tg_tokens": null,
   "tps": 23.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-wazyqym9-ctx1024",
   "model": "Skyfall-31B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: y323x2zh; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 83.0,
   "provenance": "community",
   "quant": "Mixed-4-8",
   "quote": "Skyfall-31B-v4.2-MLX-Mixed-4-8 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 83.0 12.9 23.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wazyqym9",
   "source_url": "https://omlx.ai/benchmarks/performance/wazyqym9",
   "tg_tokens": null,
   "tps": 12.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-wazyqym9-ctx4096",
   "model": "Skyfall-31B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: y323x2zh; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 82.6,
   "provenance": "community",
   "quant": "Mixed-4-8",
   "quote": "Skyfall-31B-v4.2-MLX-Mixed-4-8 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 82.6 12.6 24.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wazyqym9",
   "source_url": "https://omlx.ai/benchmarks/performance/wazyqym9",
   "tg_tokens": null,
   "tps": 12.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-kk7rs0ij-ctx1024",
   "model": "Skyfall-31B-v4.2-heretic",
   "notes": "MLX; oMLX community benchmark trace; peak memory 31.8 GB at this context; duplicate traces of the same session: nfy5i72v; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 83.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Skyfall-31B-v4.2-heretic-mlx-8Bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 1k 83.9 9.7 31.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kk7rs0ij",
   "source_url": "https://omlx.ai/benchmarks/performance/kk7rs0ij",
   "tg_tokens": null,
   "tps": 9.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-kk7rs0ij-ctx4096",
   "model": "Skyfall-31B-v4.2-heretic",
   "notes": "MLX; oMLX community benchmark trace; peak memory 32.8 GB at this context; duplicate traces of the same session: nfy5i72v; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 83.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Skyfall-31B-v4.2-heretic-mlx-8Bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k 83.8 9.4 32.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kk7rs0ij",
   "source_url": "https://omlx.ai/benchmarks/performance/kk7rs0ij",
   "tg_tokens": null,
   "tps": 9.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-t0bbblxg-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 69.2 GB at this context; duplicate traces of the same session: 61f2pku5; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 587.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k 587.0 62.7 69.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t0bbblxg",
   "source_url": "https://omlx.ai/benchmarks/performance/t0bbblxg",
   "tg_tokens": null,
   "tps": 62.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-t0bbblxg-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 69.3 GB at this context; duplicate traces of the same session: 61f2pku5; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1650.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 4k 1,650 48.3 69.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t0bbblxg",
   "source_url": "https://omlx.ai/benchmarks/performance/t0bbblxg",
   "tg_tokens": null,
   "tps": 48.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-5kc2oenc-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.6 GB at this context; duplicate traces of the same session: niisqmfc, r5p937v9; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 811.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-oQ8e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 8bit; 1k 811.5 82.1 27.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5kc2oenc",
   "source_url": "https://omlx.ai/benchmarks/performance/5kc2oenc",
   "tg_tokens": null,
   "tps": 82.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-5kc2oenc-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: niisqmfc, r5p937v9; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 828.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-oQ8e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 8bit; 4k 828.4 56.3 28.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5kc2oenc",
   "source_url": "https://omlx.ai/benchmarks/performance/5kc2oenc",
   "tg_tokens": null,
   "tps": 56.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-5kc2oenc-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: niisqmfc, r5p937v9; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 776.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-oQ8e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 8bit; 16k 776.1 56.9 28.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5kc2oenc",
   "source_url": "https://omlx.ai/benchmarks/performance/5kc2oenc",
   "tg_tokens": null,
   "tps": 56.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-enm4pqmo-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: 9h15s1f6, rlm19bfv; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 735.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-oQ8e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 8bit; 1k 735.2 52.6 25.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/enm4pqmo",
   "source_url": "https://omlx.ai/benchmarks/performance/enm4pqmo",
   "tg_tokens": null,
   "tps": 52.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-enm4pqmo-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: 9h15s1f6, rlm19bfv; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 793.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-oQ8e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 8bit; 4k 793.9 51.8 26.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/enm4pqmo",
   "source_url": "https://omlx.ai/benchmarks/performance/enm4pqmo",
   "tg_tokens": null,
   "tps": 51.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-enm4pqmo-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: 9h15s1f6, rlm19bfv; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 763.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-oQ8e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 8bit; 16k 763.1 46.9 26.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/enm4pqmo",
   "source_url": "https://omlx.ai/benchmarks/performance/enm4pqmo",
   "tg_tokens": null,
   "tps": 46.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-73vmzasy-ctx32768",
   "model": "gemma-4-26B-A4B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 15.3 GB at this context; duplicate traces of the same session: 9d1e78gd; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 516.5,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "gemma-4-26B-A4B-it-qat-mxfp4 on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 32k 516.5 50.4 15.3 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/73vmzasy",
   "source_url": "https://omlx.ai/benchmarks/performance/73vmzasy",
   "tg_tokens": null,
   "tps": 50.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-73vmzasy-ctx65536",
   "model": "gemma-4-26B-A4B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: 9d1e78gd; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 428.4,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "gemma-4-26B-A4B-it-qat-mxfp4 on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 64k 428.4 39.6 16.9 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/73vmzasy",
   "source_url": "https://omlx.ai/benchmarks/performance/73vmzasy",
   "tg_tokens": null,
   "tps": 39.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-2zmkxv3m-ctx32768",
   "model": "gemma-4-26B-A4B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 15.3 GB at this context; duplicate traces of the same session: 0ezel56g; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 516.2,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "gemma-4-26B-A4B-it-qat-mxfp4 on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 32k 516.2 50.4 15.3 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2zmkxv3m",
   "source_url": "https://omlx.ai/benchmarks/performance/2zmkxv3m",
   "tg_tokens": null,
   "tps": 50.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-2zmkxv3m-ctx65536",
   "model": "gemma-4-26B-A4B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: 0ezel56g; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 426.4,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "gemma-4-26B-A4B-it-qat-mxfp4 on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 64k 426.4 39.7 16.9 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2zmkxv3m",
   "source_url": "https://omlx.ai/benchmarks/performance/2zmkxv3m",
   "tg_tokens": null,
   "tps": 39.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-obvd5zad-ctx4096",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: c7335a4o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 213.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "AREX-2-6bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 6bit; 4k 213.8 15.4 23.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/obvd5zad",
   "source_url": "https://omlx.ai/benchmarks/performance/obvd5zad",
   "tg_tokens": null,
   "tps": 15.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-obvd5zad-ctx8192",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: c7335a4o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 214.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "AREX-2-6bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 6bit; 8k 214.0 13.1 23.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/obvd5zad",
   "source_url": "https://omlx.ai/benchmarks/performance/obvd5zad",
   "tg_tokens": null,
   "tps": 13.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-zkvhvvqw-ctx4096",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: lt3gtbdm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 214.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "AREX-2-6bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 6bit; 4k 214.6 15.0 23.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zkvhvvqw",
   "source_url": "https://omlx.ai/benchmarks/performance/zkvhvvqw",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-zkvhvvqw-ctx8192",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: lt3gtbdm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "AREX-2-6bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 6bit; 8k 205.3 13.7 23.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zkvhvvqw",
   "source_url": "https://omlx.ai/benchmarks/performance/zkvhvvqw",
   "tg_tokens": null,
   "tps": 13.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-rxtgz4re-ctx4096",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "MLX; oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: cz9or6f5; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 504.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k 504.7 49.6 17.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rxtgz4re",
   "source_url": "https://omlx.ai/benchmarks/performance/rxtgz4re",
   "tg_tokens": null,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-rxtgz4re-ctx16384",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "MLX; oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: cz9or6f5; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 332.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 16k 332.6 33.0 18.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rxtgz4re",
   "source_url": "https://omlx.ai/benchmarks/performance/rxtgz4re",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-50ugoym4-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: 17jqz7c2, at0ct0mn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 772.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 772.6 88.8 22.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/50ugoym4",
   "source_url": "https://omlx.ai/benchmarks/performance/50ugoym4",
   "tg_tokens": null,
   "tps": 88.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-50ugoym4-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 17jqz7c2, at0ct0mn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 643.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 643.6 77.9 23.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/50ugoym4",
   "source_url": "https://omlx.ai/benchmarks/performance/50ugoym4",
   "tg_tokens": null,
   "tps": 77.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-50ugoym4-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: 17jqz7c2, at0ct0mn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 472.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 472.3 56.6 25.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/50ugoym4",
   "source_url": "https://omlx.ai/benchmarks/performance/50ugoym4",
   "tg_tokens": null,
   "tps": 56.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-sc76iluh-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: v9t3oyxa; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 842.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 842.1 68.9 21.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sc76iluh",
   "source_url": "https://omlx.ai/benchmarks/performance/sc76iluh",
   "tg_tokens": null,
   "tps": 68.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-sc76iluh-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: v9t3oyxa; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 915.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 915.7 84.9 22.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sc76iluh",
   "source_url": "https://omlx.ai/benchmarks/performance/sc76iluh",
   "tg_tokens": null,
   "tps": 84.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-m8b11o0s-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 32h5sxap, 85mowpgw, eirdsq9w, g9gfppyi, v66wj1q4; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2408.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 6bit; 1k 2,408 54.2 28.5 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m8b11o0s",
   "source_url": "https://omlx.ai/benchmarks/performance/m8b11o0s",
   "tg_tokens": null,
   "tps": 54.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-m8b11o0s-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.4 GB at this context; duplicate traces of the same session: 32h5sxap, 85mowpgw, eirdsq9w, g9gfppyi, v66wj1q4; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2481.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 6bit; 4k 2,481 89.5 29.4 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m8b11o0s",
   "source_url": "https://omlx.ai/benchmarks/performance/m8b11o0s",
   "tg_tokens": null,
   "tps": 89.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-m8b11o0s-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: 32h5sxap, 85mowpgw, eirdsq9w, g9gfppyi, v66wj1q4; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2346.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 6bit; 8k 2,346 25.3 29.5 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m8b11o0s",
   "source_url": "https://omlx.ai/benchmarks/performance/m8b11o0s",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-m8b11o0s-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.8 GB at this context; duplicate traces of the same session: 32h5sxap, 85mowpgw, eirdsq9w, g9gfppyi, v66wj1q4; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2107.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 6bit; 16k 2,107 19.8 29.8 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m8b11o0s",
   "source_url": "https://omlx.ai/benchmarks/performance/m8b11o0s",
   "tg_tokens": null,
   "tps": 19.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-m8b11o0s-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: 32h5sxap, 85mowpgw, eirdsq9w, g9gfppyi, v66wj1q4; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1704.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 6bit; 32k 1,704 11.6 30.3 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m8b11o0s",
   "source_url": "https://omlx.ai/benchmarks/performance/m8b11o0s",
   "tg_tokens": null,
   "tps": 11.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-m8b11o0s-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.5 GB at this context; duplicate traces of the same session: 32h5sxap, 85mowpgw, eirdsq9w, g9gfppyi, v66wj1q4; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1218.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 6bit; 64k 1,218 6.9 31.5 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m8b11o0s",
   "source_url": "https://omlx.ai/benchmarks/performance/m8b11o0s",
   "tg_tokens": null,
   "tps": 6.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-vl82ludy-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 5cabd8zh, rtmwlqeb, u4c8z7du, v3m7ihfd, ylqs5uso; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2261.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 6bit; 1k 2,261 76.0 28.5 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vl82ludy",
   "source_url": "https://omlx.ai/benchmarks/performance/vl82ludy",
   "tg_tokens": null,
   "tps": 76.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-vl82ludy-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.4 GB at this context; duplicate traces of the same session: 5cabd8zh, rtmwlqeb, u4c8z7du, v3m7ihfd, ylqs5uso; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2484.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 6bit; 4k 2,484 79.1 29.4 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vl82ludy",
   "source_url": "https://omlx.ai/benchmarks/performance/vl82ludy",
   "tg_tokens": null,
   "tps": 79.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-vl82ludy-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: 5cabd8zh, rtmwlqeb, u4c8z7du, v3m7ihfd, ylqs5uso; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2336.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 6bit; 8k 2,336 31.6 29.5 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vl82ludy",
   "source_url": "https://omlx.ai/benchmarks/performance/vl82ludy",
   "tg_tokens": null,
   "tps": 31.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-vl82ludy-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.8 GB at this context; duplicate traces of the same session: 5cabd8zh, rtmwlqeb, u4c8z7du, v3m7ihfd, ylqs5uso; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2067.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 6bit; 16k 2,067 59.1 29.8 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vl82ludy",
   "source_url": "https://omlx.ai/benchmarks/performance/vl82ludy",
   "tg_tokens": null,
   "tps": 59.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-vl82ludy-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: 5cabd8zh, rtmwlqeb, u4c8z7du, v3m7ihfd, ylqs5uso; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1666.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 6bit; 32k 1,666 22.3 30.3 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vl82ludy",
   "source_url": "https://omlx.ai/benchmarks/performance/vl82ludy",
   "tg_tokens": null,
   "tps": 22.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-vl82ludy-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.5 GB at this context; duplicate traces of the same session: 5cabd8zh, rtmwlqeb, u4c8z7du, v3m7ihfd, ylqs5uso; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1207.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 6bit; 64k 1,207 15.6 31.5 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vl82ludy",
   "source_url": "https://omlx.ai/benchmarks/performance/vl82ludy",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-05d0n34v-ctx199680",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace, single run at 195k context; peak memory 41.0 GB; TTFT 1107430 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 180.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 195k tokens 180.6 PP tok/s 11.9 TG tok/s 1107430 TTFT (ms) 41.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/05d0n34v",
   "source_url": "https://omlx.ai/benchmarks/performance/05d0n34v",
   "tg_tokens": null,
   "tps": 11.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-jyxaoj3u-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: dgbpqipq, ihm81mvl, u3zaw200; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 153.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 1k 153.2 20.1 18.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jyxaoj3u",
   "source_url": "https://omlx.ai/benchmarks/performance/jyxaoj3u",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-jyxaoj3u-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: dgbpqipq, ihm81mvl, u3zaw200; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 148.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 4k 148.8 30.3 22.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jyxaoj3u",
   "source_url": "https://omlx.ai/benchmarks/performance/jyxaoj3u",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-jyxaoj3u-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: dgbpqipq, ihm81mvl, u3zaw200; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 145.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 8k 145.3 36.3 23.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jyxaoj3u",
   "source_url": "https://omlx.ai/benchmarks/performance/jyxaoj3u",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-jyxaoj3u-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: dgbpqipq, ihm81mvl, u3zaw200; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 136.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 16k 136.4 27.3 24.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jyxaoj3u",
   "source_url": "https://omlx.ai/benchmarks/performance/jyxaoj3u",
   "tg_tokens": null,
   "tps": 27.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-jyxaoj3u-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.6 GB at this context; duplicate traces of the same session: dgbpqipq, ihm81mvl, u3zaw200; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 121.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 32k 121.3 48.6 25.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jyxaoj3u",
   "source_url": "https://omlx.ai/benchmarks/performance/jyxaoj3u",
   "tg_tokens": null,
   "tps": 48.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ufbobq0f-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: rtnm2lq6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 324.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 324.9 22.9 34.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ufbobq0f",
   "source_url": "https://omlx.ai/benchmarks/performance/ufbobq0f",
   "tg_tokens": null,
   "tps": 22.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ufbobq0f-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.8 GB at this context; duplicate traces of the same session: rtnm2lq6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 370.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 370.1 22.6 37.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ufbobq0f",
   "source_url": "https://omlx.ai/benchmarks/performance/ufbobq0f",
   "tg_tokens": null,
   "tps": 22.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-4tmmgcux-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP320",
   "notes": "DWQ MTP Vision build; oMLX community benchmark trace, single run at 4k context; peak memory 25.1 GB; TTFT 15057 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 272.0,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-Flash-Next-REAP320-oQ3e-fp16-DWQ-MTP-Vision-MLX on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 3bit; 4k tokens 272.0 PP tok/s 8.5 TG tok/s 15057 TTFT (ms) 25.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4tmmgcux",
   "source_url": "https://omlx.ai/benchmarks/performance/4tmmgcux",
   "tg_tokens": null,
   "tps": 8.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ld4syqa2-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 61.5 GB at this context; duplicate traces of the same session: b6jcpq0x, ht4johy2, rg80otgb; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 289.1,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 3bit; 1k 289.1 33.7 61.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ld4syqa2",
   "source_url": "https://omlx.ai/benchmarks/performance/ld4syqa2",
   "tg_tokens": null,
   "tps": 33.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ld4syqa2-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 61.9 GB at this context; duplicate traces of the same session: b6jcpq0x, ht4johy2, rg80otgb; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 322.8,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 3bit; 4k 322.8 36.9 61.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ld4syqa2",
   "source_url": "https://omlx.ai/benchmarks/performance/ld4syqa2",
   "tg_tokens": null,
   "tps": 36.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-ld4syqa2-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 62.6 GB at this context; duplicate traces of the same session: b6jcpq0x, ht4johy2, rg80otgb; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 495.1,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 3bit; 16k 495.1 49.1 62.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ld4syqa2",
   "source_url": "https://omlx.ai/benchmarks/performance/ld4syqa2",
   "tg_tokens": null,
   "tps": 49.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ld4syqa2-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 66.8 GB at this context; duplicate traces of the same session: b6jcpq0x, ht4johy2, rg80otgb; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 489.4,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 3bit; 64k 489.4 37.1 66.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ld4syqa2",
   "source_url": "https://omlx.ai/benchmarks/performance/ld4syqa2",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dxhh3ir3-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1750.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,750 62.5 70.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dxhh3ir3",
   "source_url": "https://omlx.ai/benchmarks/performance/dxhh3ir3",
   "tg_tokens": null,
   "tps": 62.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dxhh3ir3-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2930.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,930 104.6 72.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dxhh3ir3",
   "source_url": "https://omlx.ai/benchmarks/performance/dxhh3ir3",
   "tg_tokens": null,
   "tps": 104.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-sc5z6c7x-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.5 GB at this context; duplicate traces of the same session: 0bz8rshi, dli0l3em; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1097.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,097 79.3 70.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sc5z6c7x",
   "source_url": "https://omlx.ai/benchmarks/performance/sc5z6c7x",
   "tg_tokens": null,
   "tps": 79.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-sc5z6c7x-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: 0bz8rshi, dli0l3em; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2143.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,143 75.0 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sc5z6c7x",
   "source_url": "https://omlx.ai/benchmarks/performance/sc5z6c7x",
   "tg_tokens": null,
   "tps": 75.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-sc5z6c7x-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: 0bz8rshi, dli0l3em; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2920.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,920 110.1 72.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sc5z6c7x",
   "source_url": "https://omlx.ai/benchmarks/performance/sc5z6c7x",
   "tg_tokens": null,
   "tps": 110.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-sc5z6c7x-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: 0bz8rshi, dli0l3em; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2855.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,855 69.6 73.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sc5z6c7x",
   "source_url": "https://omlx.ai/benchmarks/performance/sc5z6c7x",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4ql70clx-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 421.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 421.4 19.8 39.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ql70clx",
   "source_url": "https://omlx.ai/benchmarks/performance/4ql70clx",
   "tg_tokens": null,
   "tps": 19.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4ql70clx-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 41.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 948.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 948.3 16.1 41.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ql70clx",
   "source_url": "https://omlx.ai/benchmarks/performance/4ql70clx",
   "tg_tokens": null,
   "tps": 16.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2m3rpk2d-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 3qdut765, eq52ky7n, hyhj8mgo, u7cch6q6, z94buxsd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 965.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 965.4 72.3 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2m3rpk2d",
   "source_url": "https://omlx.ai/benchmarks/performance/2m3rpk2d",
   "tg_tokens": null,
   "tps": 72.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2m3rpk2d-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: 3qdut765, eq52ky7n, hyhj8mgo, u7cch6q6, z94buxsd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1990.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,990 69.3 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2m3rpk2d",
   "source_url": "https://omlx.ai/benchmarks/performance/2m3rpk2d",
   "tg_tokens": null,
   "tps": 69.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2m3rpk2d-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 3qdut765, eq52ky7n, hyhj8mgo, u7cch6q6, z94buxsd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2218.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,218 81.9 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2m3rpk2d",
   "source_url": "https://omlx.ai/benchmarks/performance/2m3rpk2d",
   "tg_tokens": null,
   "tps": 81.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2m3rpk2d-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.6 GB at this context; duplicate traces of the same session: 3qdut765, eq52ky7n, hyhj8mgo, u7cch6q6, z94buxsd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2484.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,484 83.4 74.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2m3rpk2d",
   "source_url": "https://omlx.ai/benchmarks/performance/2m3rpk2d",
   "tg_tokens": null,
   "tps": 83.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2m3rpk2d-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.6 GB at this context; duplicate traces of the same session: 3qdut765, eq52ky7n, hyhj8mgo, u7cch6q6, z94buxsd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2400.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,400 65.3 78.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2m3rpk2d",
   "source_url": "https://omlx.ai/benchmarks/performance/2m3rpk2d",
   "tg_tokens": null,
   "tps": 65.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2m3rpk2d-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.9 GB at this context; duplicate traces of the same session: 3qdut765, eq52ky7n, hyhj8mgo, u7cch6q6, z94buxsd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2330.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,330 59.6 80.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2m3rpk2d",
   "source_url": "https://omlx.ai/benchmarks/performance/2m3rpk2d",
   "tg_tokens": null,
   "tps": 59.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-qjwpwnd4-ctx1024",
   "model": "gemma-4-12B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: 2o29bc2d; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 297.4,
   "provenance": "community",
   "quant": "bf16",
   "quote": "gemma-4-12B-it-qat-bf16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 bf16; 1k 297.4 12.9 23.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qjwpwnd4",
   "source_url": "https://omlx.ai/benchmarks/performance/qjwpwnd4",
   "tg_tokens": null,
   "tps": 12.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-qjwpwnd4-ctx4096",
   "model": "gemma-4-12B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 2o29bc2d; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 307.9,
   "provenance": "community",
   "quant": "bf16",
   "quote": "gemma-4-12B-it-qat-bf16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 bf16; 4k 307.9 12.8 23.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qjwpwnd4",
   "source_url": "https://omlx.ai/benchmarks/performance/qjwpwnd4",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-gy1j6gqr-ctx1024",
   "model": "gemma-4-26B-A4B-it-oQ8e-fp16-mtp",
   "notes": "oMLX community benchmark trace; peak memory 28.7 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 938.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 8bit; 1k 938.1 52.3 28.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gy1j6gqr",
   "source_url": "https://omlx.ai/benchmarks/performance/gy1j6gqr",
   "tg_tokens": null,
   "tps": 52.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-gy1j6gqr-ctx4096",
   "model": "gemma-4-26B-A4B-it-oQ8e-fp16-mtp",
   "notes": "oMLX community benchmark trace; peak memory 29.0 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1072.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 8bit; 4k 1,072 58.2 29.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gy1j6gqr",
   "source_url": "https://omlx.ai/benchmarks/performance/gy1j6gqr",
   "tg_tokens": null,
   "tps": 58.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-gy1j6gqr-ctx16384",
   "model": "gemma-4-26B-A4B-it-oQ8e-fp16-mtp",
   "notes": "oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1027.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 8bit; 16k 1,027 77.7 29.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gy1j6gqr",
   "source_url": "https://omlx.ai/benchmarks/performance/gy1j6gqr",
   "tg_tokens": null,
   "tps": 77.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-oe25bday-ctx1024",
   "model": "gemma-4-26B-A4B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 30.9 GB at this context; duplicate traces of the same session: 8q8vw9d7; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 781.9,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "gemma-4-26B-A4B-it-qat-oQ8-fp16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 1k 781.9 12.2 30.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oe25bday",
   "source_url": "https://omlx.ai/benchmarks/performance/oe25bday",
   "tg_tokens": null,
   "tps": 12.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-oe25bday-ctx4096",
   "model": "gemma-4-26B-A4B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 31.1 GB at this context; duplicate traces of the same session: 8q8vw9d7; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 807.6,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "gemma-4-26B-A4B-it-qat-oQ8-fp16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k 807.6 9.8 31.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oe25bday",
   "source_url": "https://omlx.ai/benchmarks/performance/oe25bday",
   "tg_tokens": null,
   "tps": 9.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-4mmeew4e-ctx1024",
   "model": "gemma-4-26B-A4B-it-uncensored-heretic-ara",
   "notes": "MLX; oMLX community benchmark trace; peak memory 29.1 GB at this context; duplicate traces of the same session: u1peh00l; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 635.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-uncensored-heretic-ara-MLX-8bit-int8-affine-betterCache on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 1k 635.4 44.5 29.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4mmeew4e",
   "source_url": "https://omlx.ai/benchmarks/performance/4mmeew4e",
   "tg_tokens": null,
   "tps": 44.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-4mmeew4e-ctx4096",
   "model": "gemma-4-26B-A4B-it-uncensored-heretic-ara",
   "notes": "MLX; oMLX community benchmark trace; peak memory 29.4 GB at this context; duplicate traces of the same session: u1peh00l; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 653.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-uncensored-heretic-ara-MLX-8bit-int8-affine-betterCache on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k 653.4 43.0 29.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4mmeew4e",
   "source_url": "https://omlx.ai/benchmarks/performance/4mmeew4e",
   "tg_tokens": null,
   "tps": 43.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-l8oujcj3-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 27.6 GB at this context; duplicate traces of the same session: 1ifanvuh; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 653.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26b-a4b-it-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 1k 653.2 48.7 27.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l8oujcj3",
   "source_url": "https://omlx.ai/benchmarks/performance/l8oujcj3",
   "tg_tokens": null,
   "tps": 48.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-l8oujcj3-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: 1ifanvuh; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 662.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26b-a4b-it-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k 662.0 42.7 28.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l8oujcj3",
   "source_url": "https://omlx.ai/benchmarks/performance/l8oujcj3",
   "tg_tokens": null,
   "tps": 42.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-yjfky5u1-ctx1024",
   "model": "gemma-4-E2B-Heretic-Uncensored",
   "notes": "MLX; oMLX community benchmark trace; peak memory 3.1 GB at this context; duplicate traces of the same session: z8en10fr; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3079.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E2B-Heretic-Uncensored-mlx-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 3,079 107.3 3.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yjfky5u1",
   "source_url": "https://omlx.ai/benchmarks/performance/yjfky5u1",
   "tg_tokens": null,
   "tps": 107.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-yjfky5u1-ctx4096",
   "model": "gemma-4-E2B-Heretic-Uncensored",
   "notes": "MLX; oMLX community benchmark trace; peak memory 3.1 GB at this context; duplicate traces of the same session: z8en10fr; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3665.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E2B-Heretic-Uncensored-mlx-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 3,665 102.3 3.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yjfky5u1",
   "source_url": "https://omlx.ai/benchmarks/performance/yjfky5u1",
   "tg_tokens": null,
   "tps": 102.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-wgndzovx-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 10.9 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2300.0,
   "provenance": "community",
   "quant": "MXFP4-Q4",
   "quote": "gpt-oss-20b-MXFP4-Q4 on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 2,300 99.8 10.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wgndzovx",
   "source_url": "https://omlx.ai/benchmarks/performance/wgndzovx",
   "tg_tokens": null,
   "tps": 99.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-wgndzovx-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2449.0,
   "provenance": "community",
   "quant": "MXFP4-Q4",
   "quote": "gpt-oss-20b-MXFP4-Q4 on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 2,449 94.0 11.0 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wgndzovx",
   "source_url": "https://omlx.ai/benchmarks/performance/wgndzovx",
   "tg_tokens": null,
   "tps": 94.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-wgndzovx-ctx8192",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2432.0,
   "provenance": "community",
   "quant": "MXFP4-Q4",
   "quote": "gpt-oss-20b-MXFP4-Q4 on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 8k 2,432 85.4 11.0 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wgndzovx",
   "source_url": "https://omlx.ai/benchmarks/performance/wgndzovx",
   "tg_tokens": null,
   "tps": 85.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-zfli714a-ctx1024",
   "model": "Devstral-Small-2-24B-Instruct-2512",
   "notes": "MLX; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: ng0fj7s7; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 113.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mistralai_Devstral-Small-2-24B-Instruct-2512-MLX-8Bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 1k 113.1 13.2 24.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zfli714a",
   "source_url": "https://omlx.ai/benchmarks/performance/zfli714a",
   "tg_tokens": null,
   "tps": 13.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-zfli714a-ctx4096",
   "model": "Devstral-Small-2-24B-Instruct-2512",
   "notes": "MLX; oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: ng0fj7s7; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 112.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mistralai_Devstral-Small-2-24B-Instruct-2512-MLX-8Bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k 112.7 12.9 25.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zfli714a",
   "source_url": "https://omlx.ai/benchmarks/performance/zfli714a",
   "tg_tokens": null,
   "tps": 12.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ejw2cz9r-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: vvn29anr; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1438.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,438 101.6 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ejw2cz9r",
   "source_url": "https://omlx.ai/benchmarks/performance/ejw2cz9r",
   "tg_tokens": null,
   "tps": 101.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ejw2cz9r-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: vvn29anr; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2970.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,970 89.8 20.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ejw2cz9r",
   "source_url": "https://omlx.ai/benchmarks/performance/ejw2cz9r",
   "tg_tokens": null,
   "tps": 89.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-oyzps2tr-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 567dvqqs; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 548.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 548.4 43.0 22.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oyzps2tr",
   "source_url": "https://omlx.ai/benchmarks/performance/oyzps2tr",
   "tg_tokens": null,
   "tps": 43.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-oyzps2tr-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: 567dvqqs; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 402.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 402.7 39.6 23.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oyzps2tr",
   "source_url": "https://omlx.ai/benchmarks/performance/oyzps2tr",
   "tg_tokens": null,
   "tps": 39.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-xjf914qs-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: 4yzydqa3; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 372.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 5bit; 4k 372.9 20.7 26.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xjf914qs",
   "source_url": "https://omlx.ai/benchmarks/performance/xjf914qs",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-xjf914qs-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: 4yzydqa3; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 404.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 5bit; 16k 404.7 20.7 27.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xjf914qs",
   "source_url": "https://omlx.ai/benchmarks/performance/xjf914qs",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ode3614q-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 92j6bsd0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 941.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 941.5 31.7 17.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ode3614q",
   "source_url": "https://omlx.ai/benchmarks/performance/ode3614q",
   "tg_tokens": null,
   "tps": 31.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ode3614q-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 92j6bsd0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 833.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 833.6 30.3 17.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ode3614q",
   "source_url": "https://omlx.ai/benchmarks/performance/ode3614q",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-83jahkn4-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: bcn5yxv6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 591.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 591.2 26.4 17.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/83jahkn4",
   "source_url": "https://omlx.ai/benchmarks/performance/83jahkn4",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-83jahkn4-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: bcn5yxv6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 609.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 609.6 26.0 17.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/83jahkn4",
   "source_url": "https://omlx.ai/benchmarks/performance/83jahkn4",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-s309hwie-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: 8bnonamd, jhe69jws, nsm43qbm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 476.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 476.6 18.1 15.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s309hwie",
   "source_url": "https://omlx.ai/benchmarks/performance/s309hwie",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-s309hwie-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 8bnonamd, jhe69jws, nsm43qbm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 508.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 508.1 17.8 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s309hwie",
   "source_url": "https://omlx.ai/benchmarks/performance/s309hwie",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-s309hwie-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 8bnonamd, jhe69jws, nsm43qbm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 503.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 503.1 17.1 17.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s309hwie",
   "source_url": "https://omlx.ai/benchmarks/performance/s309hwie",
   "tg_tokens": null,
   "tps": 17.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-s309hwie-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 8bnonamd, jhe69jws, nsm43qbm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 462.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 462.5 16.7 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s309hwie",
   "source_url": "https://omlx.ai/benchmarks/performance/s309hwie",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-3pycbbzy-ctx131072",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace, single run at 128k context; peak memory 36.3 GB; TTFT 624152 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 210.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ6e-1M on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 128k tokens 210.0 PP tok/s 16.4 TG tok/s 624152 TTFT (ms) 36.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3pycbbzy",
   "source_url": "https://omlx.ai/benchmarks/performance/3pycbbzy",
   "tg_tokens": null,
   "tps": 16.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-3blw5lpk-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 26.3 GB; TTFT 101090 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 324.1,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 32k tokens 324.1 PP tok/s 28.1 TG tok/s 101090 TTFT (ms) 26.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3blw5lpk",
   "source_url": "https://omlx.ai/benchmarks/performance/3blw5lpk",
   "tg_tokens": null,
   "tps": 28.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-vu91te09-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 353.9,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 353.9 22.8 26.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vu91te09",
   "source_url": "https://omlx.ai/benchmarks/performance/vu91te09",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-vu91te09-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 27.6 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 347.3,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 347.3 27.5 27.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vu91te09",
   "source_url": "https://omlx.ai/benchmarks/performance/vu91te09",
   "tg_tokens": null,
   "tps": 27.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-am5kbvk7-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: wto7ac44; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 940.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 940.2 56.0 18.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/am5kbvk7",
   "source_url": "https://omlx.ai/benchmarks/performance/am5kbvk7",
   "tg_tokens": null,
   "tps": 56.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-am5kbvk7-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: wto7ac44; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 931.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 931.3 50.5 20.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/am5kbvk7",
   "source_url": "https://omlx.ai/benchmarks/performance/am5kbvk7",
   "tg_tokens": null,
   "tps": 50.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-soljqfwu-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 20.9 GB; TTFT 53385 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 613.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k tokens 613.8 PP tok/s 73.9 TG tok/s 53385 TTFT (ms) 20.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/soljqfwu",
   "source_url": "https://omlx.ai/benchmarks/performance/soljqfwu",
   "tg_tokens": null,
   "tps": 73.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-evgpltpw-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 40.0 GB at this context; duplicate traces of the same session: 32jn4eux, cmx2qnhz, mz7wz8d6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 303.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 8bit; 1k 303.0 29.3 40.0 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/evgpltpw",
   "source_url": "https://omlx.ai/benchmarks/performance/evgpltpw",
   "tg_tokens": null,
   "tps": 29.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-evgpltpw-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 43.5 GB at this context; duplicate traces of the same session: 32jn4eux, cmx2qnhz, mz7wz8d6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 347.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 8bit; 4k 347.7 32.4 43.5 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/evgpltpw",
   "source_url": "https://omlx.ai/benchmarks/performance/evgpltpw",
   "tg_tokens": null,
   "tps": 32.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-evgpltpw-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 44.3 GB at this context; duplicate traces of the same session: 32jn4eux, cmx2qnhz, mz7wz8d6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 372.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 8bit; 8k 372.0 33.6 44.3 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/evgpltpw",
   "source_url": "https://omlx.ai/benchmarks/performance/evgpltpw",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-evgpltpw-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 45.9 GB at this context; duplicate traces of the same session: 32jn4eux, cmx2qnhz, mz7wz8d6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 368.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 8bit; 16k 368.1 41.3 45.9 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/evgpltpw",
   "source_url": "https://omlx.ai/benchmarks/performance/evgpltpw",
   "tg_tokens": null,
   "tps": 41.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-lbgt4ta7-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: jhwc6rvy, ty77gubh, v6ozvh46; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 412.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 412.4 21.1 32.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lbgt4ta7",
   "source_url": "https://omlx.ai/benchmarks/performance/lbgt4ta7",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-lbgt4ta7-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.7 GB at this context; duplicate traces of the same session: jhwc6rvy, ty77gubh, v6ozvh46; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 417.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 417.6 15.5 33.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lbgt4ta7",
   "source_url": "https://omlx.ai/benchmarks/performance/lbgt4ta7",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-lbgt4ta7-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.6 GB at this context; duplicate traces of the same session: jhwc6rvy, ty77gubh, v6ozvh46; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 374.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 374.5 20.4 34.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lbgt4ta7",
   "source_url": "https://omlx.ai/benchmarks/performance/lbgt4ta7",
   "tg_tokens": null,
   "tps": 20.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-lbgt4ta7-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.0 GB at this context; duplicate traces of the same session: jhwc6rvy, ty77gubh, v6ozvh46; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 402.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 16k 402.0 27.1 36.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lbgt4ta7",
   "source_url": "https://omlx.ai/benchmarks/performance/lbgt4ta7",
   "tg_tokens": null,
   "tps": 27.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-g39zze5t-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "fp16 MTP build; oMLX community benchmark trace, single run at 16k context; peak memory 75.3 GB; TTFT 35858 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 456.9,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-fp16-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 3bit; 16k tokens 456.9 PP tok/s 28.2 TG tok/s 35858 TTFT (ms) 75.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g39zze5t",
   "source_url": "https://omlx.ai/benchmarks/performance/g39zze5t",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-a7w2adlc-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "fp16 MTP build; oMLX community benchmark trace, single run at 4k context; peak memory 72.0 GB; TTFT 9425 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 434.6,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-fp16-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 3bit; 4k tokens 434.6 PP tok/s 20.0 TG tok/s 9425 TTFT (ms) 72.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a7w2adlc",
   "source_url": "https://omlx.ai/benchmarks/performance/a7w2adlc",
   "tg_tokens": null,
   "tps": 20.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-dw5mk8cu-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace, single run at 1k context; peak memory 74.3 GB; TTFT 4725 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 216.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k tokens 216.7 PP tok/s 12.8 TG tok/s 4725 TTFT (ms) 74.3 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dw5mk8cu",
   "source_url": "https://omlx.ai/benchmarks/performance/dw5mk8cu",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-jsj5tkxq-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 42.1 GB at this context; duplicate traces of the same session: 2ipctlbs, xnxuafk7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 97.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 97.6 14.4 42.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jsj5tkxq",
   "source_url": "https://omlx.ai/benchmarks/performance/jsj5tkxq",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-jsj5tkxq-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 42.9 GB at this context; duplicate traces of the same session: 2ipctlbs, xnxuafk7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 272.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 272.2 11.7 42.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jsj5tkxq",
   "source_url": "https://omlx.ai/benchmarks/performance/jsj5tkxq",
   "tg_tokens": null,
   "tps": 11.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-jsj5tkxq-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 44.1 GB at this context; duplicate traces of the same session: 2ipctlbs, xnxuafk7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 425.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 425.8 13.7 44.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jsj5tkxq",
   "source_url": "https://omlx.ai/benchmarks/performance/jsj5tkxq",
   "tg_tokens": null,
   "tps": 13.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1okvy2m2-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: 0nqtsrqx, 87qr0omn, kl3c21cl, lrer7n75, m2tkzsu3, tslofglr, u1bkaytr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1045.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,045 75.9 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1okvy2m2",
   "source_url": "https://omlx.ai/benchmarks/performance/1okvy2m2",
   "tg_tokens": null,
   "tps": 75.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1okvy2m2-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 0nqtsrqx, 87qr0omn, kl3c21cl, lrer7n75, m2tkzsu3, tslofglr, u1bkaytr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2062.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,062 79.7 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1okvy2m2",
   "source_url": "https://omlx.ai/benchmarks/performance/1okvy2m2",
   "tg_tokens": null,
   "tps": 79.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1okvy2m2-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 0nqtsrqx, 87qr0omn, kl3c21cl, lrer7n75, m2tkzsu3, tslofglr, u1bkaytr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2379.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,379 86.0 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1okvy2m2",
   "source_url": "https://omlx.ai/benchmarks/performance/1okvy2m2",
   "tg_tokens": null,
   "tps": 86.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1okvy2m2-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: 0nqtsrqx, 87qr0omn, kl3c21cl, lrer7n75, m2tkzsu3, tslofglr, u1bkaytr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2582.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,582 76.2 72.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1okvy2m2",
   "source_url": "https://omlx.ai/benchmarks/performance/1okvy2m2",
   "tg_tokens": null,
   "tps": 76.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1okvy2m2-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: 0nqtsrqx, 87qr0omn, kl3c21cl, lrer7n75, m2tkzsu3, tslofglr, u1bkaytr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2576.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,576 87.2 73.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1okvy2m2",
   "source_url": "https://omlx.ai/benchmarks/performance/1okvy2m2",
   "tg_tokens": null,
   "tps": 87.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1okvy2m2-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.5 GB at this context; duplicate traces of the same session: 0nqtsrqx, 87qr0omn, kl3c21cl, lrer7n75, m2tkzsu3, tslofglr, u1bkaytr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2482.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,482 79.2 74.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1okvy2m2",
   "source_url": "https://omlx.ai/benchmarks/performance/1okvy2m2",
   "tg_tokens": null,
   "tps": 79.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1okvy2m2-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.7 GB at this context; duplicate traces of the same session: 0nqtsrqx, 87qr0omn, kl3c21cl, lrer7n75, m2tkzsu3, tslofglr, u1bkaytr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2397.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,397 67.4 78.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1okvy2m2",
   "source_url": "https://omlx.ai/benchmarks/performance/1okvy2m2",
   "tg_tokens": null,
   "tps": 67.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1okvy2m2-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.9 GB at this context; duplicate traces of the same session: 0nqtsrqx, 87qr0omn, kl3c21cl, lrer7n75, m2tkzsu3, tslofglr, u1bkaytr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2289.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,289 58.8 80.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1okvy2m2",
   "source_url": "https://omlx.ai/benchmarks/performance/1okvy2m2",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-evewzou7-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 150.8 GB at this context; duplicate traces of the same session: p9naa2jt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4339.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 128k 4,339 99.8 150.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/evewzou7",
   "source_url": "https://omlx.ai/benchmarks/performance/evewzou7",
   "tg_tokens": null,
   "tps": 99.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-evewzou7-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 153.0 GB at this context; duplicate traces of the same session: p9naa2jt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4210.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 195k 4,210 87.8 153.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/evewzou7",
   "source_url": "https://omlx.ai/benchmarks/performance/evewzou7",
   "tg_tokens": null,
   "tps": 87.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-13gzrvu5-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.5 GB at this context; duplicate traces of the same session: 3onk20mn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2057.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 2,057 149.4 142.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/13gzrvu5",
   "source_url": "https://omlx.ai/benchmarks/performance/13gzrvu5",
   "tg_tokens": null,
   "tps": 149.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-13gzrvu5-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: 3onk20mn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3328.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 3,328 121.6 143.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/13gzrvu5",
   "source_url": "https://omlx.ai/benchmarks/performance/13gzrvu5",
   "tg_tokens": null,
   "tps": 121.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ylscj605-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: o6qgicn9; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 967.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k 967.8 26.0 70.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ylscj605",
   "source_url": "https://omlx.ai/benchmarks/performance/ylscj605",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ylscj605-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: o6qgicn9; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1669.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 4k 1,669 43.4 70.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ylscj605",
   "source_url": "https://omlx.ai/benchmarks/performance/ylscj605",
   "tg_tokens": null,
   "tps": 43.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-nt3pg0r7-ctx1024",
   "model": "Anubis-70B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 37.8 GB at this context; duplicate traces of the same session: nagwidr6; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 41.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "anubis-70b-v1.2-mlx-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 41.1 7.4 37.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nt3pg0r7",
   "source_url": "https://omlx.ai/benchmarks/performance/nt3pg0r7",
   "tg_tokens": null,
   "tps": 7.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-nt3pg0r7-ctx4096",
   "model": "Anubis-70B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 39.0 GB at this context; duplicate traces of the same session: nagwidr6; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 40.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "anubis-70b-v1.2-mlx-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 40.4 7.0 39.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nt3pg0r7",
   "source_url": "https://omlx.ai/benchmarks/performance/nt3pg0r7",
   "tg_tokens": null,
   "tps": 7.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-svinemd7-ctx1024",
   "model": "gemma-4-26B-A4B-it-oQ8e-fp16-mtp",
   "notes": "oMLX community benchmark trace; peak memory 28.7 GB at this context; duplicate traces of the same session: 0og0zshg; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 938.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 8bit; 1k 938.1 52.3 28.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/svinemd7",
   "source_url": "https://omlx.ai/benchmarks/performance/svinemd7",
   "tg_tokens": null,
   "tps": 52.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-svinemd7-ctx4096",
   "model": "gemma-4-26B-A4B-it-oQ8e-fp16-mtp",
   "notes": "oMLX community benchmark trace; peak memory 29.0 GB at this context; duplicate traces of the same session: 0og0zshg; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1072.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 8bit; 4k 1,072 58.2 29.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/svinemd7",
   "source_url": "https://omlx.ai/benchmarks/performance/svinemd7",
   "tg_tokens": null,
   "tps": 58.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-svinemd7-ctx16384",
   "model": "gemma-4-26B-A4B-it-oQ8e-fp16-mtp",
   "notes": "oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: 0og0zshg; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1027.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 8bit; 16k 1,027 77.7 29.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/svinemd7",
   "source_url": "https://omlx.ai/benchmarks/performance/svinemd7",
   "tg_tokens": null,
   "tps": 77.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-72w5a4oj-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: o0xejwxg; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2204.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 2,204 136.9 28.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/72w5a4oj",
   "source_url": "https://omlx.ai/benchmarks/performance/72w5a4oj",
   "tg_tokens": null,
   "tps": 136.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-72w5a4oj-ctx8192",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.2 GB at this context; duplicate traces of the same session: o0xejwxg; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2253.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 2,253 78.4 28.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/72w5a4oj",
   "source_url": "https://omlx.ai/benchmarks/performance/72w5a4oj",
   "tg_tokens": null,
   "tps": 78.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-72w5a4oj-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: o0xejwxg; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2132.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k 2,132 84.1 28.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/72w5a4oj",
   "source_url": "https://omlx.ai/benchmarks/performance/72w5a4oj",
   "tg_tokens": null,
   "tps": 84.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-72w5a4oj-ctx32768",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.4 GB at this context; duplicate traces of the same session: o0xejwxg; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1918.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 32k 1,918 65.4 29.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/72w5a4oj",
   "source_url": "https://omlx.ai/benchmarks/performance/72w5a4oj",
   "tg_tokens": null,
   "tps": 65.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-zl72g491-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 10.9 GB at this context; duplicate traces of the same session: pvik5ug1; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2300.0,
   "provenance": "community",
   "quant": "MXFP4-Q4",
   "quote": "gpt-oss-20b-MXFP4-Q4 on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 2,300 99.8 10.9 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zl72g491",
   "source_url": "https://omlx.ai/benchmarks/performance/zl72g491",
   "tg_tokens": null,
   "tps": 99.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-zl72g491-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.0 GB at this context; duplicate traces of the same session: pvik5ug1; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2449.0,
   "provenance": "community",
   "quant": "MXFP4-Q4",
   "quote": "gpt-oss-20b-MXFP4-Q4 on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 2,449 94.0 11.0 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zl72g491",
   "source_url": "https://omlx.ai/benchmarks/performance/zl72g491",
   "tg_tokens": null,
   "tps": 94.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-zl72g491-ctx8192",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.0 GB at this context; duplicate traces of the same session: pvik5ug1; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2432.0,
   "provenance": "community",
   "quant": "MXFP4-Q4",
   "quote": "gpt-oss-20b-MXFP4-Q4 on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 8k 2,432 85.4 11.0 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zl72g491",
   "source_url": "https://omlx.ai/benchmarks/performance/zl72g491",
   "tg_tokens": null,
   "tps": 85.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-84owe4uq-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: vuj89806; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 622.7,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 4k 622.7 69.9 11.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/84owe4uq",
   "source_url": "https://omlx.ai/benchmarks/performance/84owe4uq",
   "tg_tokens": null,
   "tps": 69.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-84owe4uq-ctx16384",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 12.1 GB at this context; duplicate traces of the same session: vuj89806; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 564.7,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 16k 564.7 55.4 12.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/84owe4uq",
   "source_url": "https://omlx.ai/benchmarks/performance/84owe4uq",
   "tg_tokens": null,
   "tps": 55.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-i2d9c8sa-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: nhqqgq5j; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4274.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 4,274 105.1 11.8 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i2d9c8sa",
   "source_url": "https://omlx.ai/benchmarks/performance/i2d9c8sa",
   "tg_tokens": null,
   "tps": 105.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-i2d9c8sa-ctx16384",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 12.1 GB at this context; duplicate traces of the same session: nhqqgq5j; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4018.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 4,018 90.1 12.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i2d9c8sa",
   "source_url": "https://omlx.ai/benchmarks/performance/i2d9c8sa",
   "tg_tokens": null,
   "tps": 90.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-c8pjl5e7-ctx4096",
   "model": "Phi-3.5-mini-instruct",
   "notes": "oMLX community benchmark trace; peak memory 4.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3484.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Phi-3.5-mini-instruct-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 3,484 102.1 4.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c8pjl5e7",
   "source_url": "https://omlx.ai/benchmarks/performance/c8pjl5e7",
   "tg_tokens": null,
   "tps": 102.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-c8pjl5e7-ctx16384",
   "model": "Phi-3.5-mini-instruct",
   "notes": "oMLX community benchmark trace; peak memory 8.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2392.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Phi-3.5-mini-instruct-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,392 47.4 8.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c8pjl5e7",
   "source_url": "https://omlx.ai/benchmarks/performance/c8pjl5e7",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-k1al2h31-ctx4096",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 2.3 GB at this context; duplicate traces of the same session: jlox8kbm; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 990.5,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.5-0.8B-oQ6 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 4k 990.5 43.4 2.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k1al2h31",
   "source_url": "https://omlx.ai/benchmarks/performance/k1al2h31",
   "tg_tokens": null,
   "tps": 43.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-k1al2h31-ctx16384",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 2.8 GB at this context; duplicate traces of the same session: jlox8kbm; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1141.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.5-0.8B-oQ6 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 16k 1,141 34.6 2.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k1al2h31",
   "source_url": "https://omlx.ai/benchmarks/performance/k1al2h31",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-xt7f66nl-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: ly4gqzzn; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 663.9,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 663.9 36.7 22.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xt7f66nl",
   "source_url": "https://omlx.ai/benchmarks/performance/xt7f66nl",
   "tg_tokens": null,
   "tps": 36.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-xt7f66nl-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: ly4gqzzn; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 501.4,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 501.4 36.5 23.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xt7f66nl",
   "source_url": "https://omlx.ai/benchmarks/performance/xt7f66nl",
   "tg_tokens": null,
   "tps": 36.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-3oe9sa0u-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace, single run at 32k context; peak memory 42.6 GB; TTFT 23414 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1400.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 8bit; 32k tokens 1,400 PP tok/s 60.3 TG tok/s 23414 TTFT (ms) 42.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3oe9sa0u",
   "source_url": "https://omlx.ai/benchmarks/performance/3oe9sa0u",
   "tg_tokens": null,
   "tps": 60.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-u7v8abn5-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace, single run at 32k context; peak memory 42.6 GB; TTFT 23536 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1392.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 8bit; 32k tokens 1,392 PP tok/s 121.7 TG tok/s 23536 TTFT (ms) 42.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u7v8abn5",
   "source_url": "https://omlx.ai/benchmarks/performance/u7v8abn5",
   "tg_tokens": null,
   "tps": 121.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-vsepc69t-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 353.9,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 353.9 22.8 26.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vsepc69t",
   "source_url": "https://omlx.ai/benchmarks/performance/vsepc69t",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-vsepc69t-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 27.6 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 347.3,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 347.3 27.5 27.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vsepc69t",
   "source_url": "https://omlx.ai/benchmarks/performance/vsepc69t",
   "tg_tokens": null,
   "tps": 27.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-7ixovgtn-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: yqog7vmc; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 351.5,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 351.5 24.1 26.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ixovgtn",
   "source_url": "https://omlx.ai/benchmarks/performance/7ixovgtn",
   "tg_tokens": null,
   "tps": 24.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-7ixovgtn-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 27.6 GB at this context; duplicate traces of the same session: yqog7vmc; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 356.1,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 356.1 28.1 27.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ixovgtn",
   "source_url": "https://omlx.ai/benchmarks/performance/7ixovgtn",
   "tg_tokens": null,
   "tps": 28.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-d5r3a1yn-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: 5ltm88mb; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 345.1,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 345.1 22.0 23.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d5r3a1yn",
   "source_url": "https://omlx.ai/benchmarks/performance/d5r3a1yn",
   "tg_tokens": null,
   "tps": 22.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-d5r3a1yn-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: 5ltm88mb; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 322.5,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 322.5 23.8 24.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d5r3a1yn",
   "source_url": "https://omlx.ai/benchmarks/performance/d5r3a1yn",
   "tg_tokens": null,
   "tps": 23.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-t7s0bo0s-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: jafsmovp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 405.7,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 405.7 25.4 23.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t7s0bo0s",
   "source_url": "https://omlx.ai/benchmarks/performance/t7s0bo0s",
   "tg_tokens": null,
   "tps": 25.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-t7s0bo0s-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: jafsmovp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 436.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 436.0 25.7 24.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t7s0bo0s",
   "source_url": "https://omlx.ai/benchmarks/performance/t7s0bo0s",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-vrqxtadk-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: lsdy8fog; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 442.8,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 442.8 27.2 23.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vrqxtadk",
   "source_url": "https://omlx.ai/benchmarks/performance/vrqxtadk",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-vrqxtadk-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: lsdy8fog; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 466.4,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 466.4 25.7 24.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vrqxtadk",
   "source_url": "https://omlx.ai/benchmarks/performance/vrqxtadk",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-r0txciuh-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: u6elsw6j; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 375.5,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 375.5 26.1 23.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r0txciuh",
   "source_url": "https://omlx.ai/benchmarks/performance/r0txciuh",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-r0txciuh-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: u6elsw6j; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 373.8,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 373.8 22.8 24.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r0txciuh",
   "source_url": "https://omlx.ai/benchmarks/performance/r0txciuh",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-9s0fsjq9-ctx65536",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 28.5 GB; TTFT 232955 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 281.3,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 64k tokens 281.3 PP tok/s 24.9 TG tok/s 232955 TTFT (ms) 28.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9s0fsjq9",
   "source_url": "https://omlx.ai/benchmarks/performance/9s0fsjq9",
   "tg_tokens": null,
   "tps": 24.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-d7qbzsly-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: bm1ge2iz; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 355.2,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 355.2 22.4 23.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d7qbzsly",
   "source_url": "https://omlx.ai/benchmarks/performance/d7qbzsly",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-d7qbzsly-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: bm1ge2iz; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 371.3,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 371.3 24.7 24.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d7qbzsly",
   "source_url": "https://omlx.ai/benchmarks/performance/d7qbzsly",
   "tg_tokens": null,
   "tps": 24.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-mklbuld0-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: ykv7zrpm; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 430.6,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 430.6 24.9 23.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mklbuld0",
   "source_url": "https://omlx.ai/benchmarks/performance/mklbuld0",
   "tg_tokens": null,
   "tps": 24.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-mklbuld0-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: ykv7zrpm; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 440.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 440.0 25.7 24.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mklbuld0",
   "source_url": "https://omlx.ai/benchmarks/performance/mklbuld0",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-xjf1mgah-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: v96r1f19; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 374.7,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 374.7 26.1 23.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xjf1mgah",
   "source_url": "https://omlx.ai/benchmarks/performance/xjf1mgah",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-xjf1mgah-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: v96r1f19; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 357.1,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 357.1 23.6 24.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xjf1mgah",
   "source_url": "https://omlx.ai/benchmarks/performance/xjf1mgah",
   "tg_tokens": null,
   "tps": 23.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-624fl2lf-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: e5914v2c; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 322.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 322.0 20.3 23.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/624fl2lf",
   "source_url": "https://omlx.ai/benchmarks/performance/624fl2lf",
   "tg_tokens": null,
   "tps": 20.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-624fl2lf-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: e5914v2c; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 319.3,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 319.3 24.6 24.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/624fl2lf",
   "source_url": "https://omlx.ai/benchmarks/performance/624fl2lf",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-l6ahin3e-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: yky2wkib; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 437.3,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 437.3 29.9 23.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l6ahin3e",
   "source_url": "https://omlx.ai/benchmarks/performance/l6ahin3e",
   "tg_tokens": null,
   "tps": 29.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-l6ahin3e-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: yky2wkib; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 444.9,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 444.9 23.3 24.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l6ahin3e",
   "source_url": "https://omlx.ai/benchmarks/performance/l6ahin3e",
   "tg_tokens": null,
   "tps": 23.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-yrx4vd88-ctx131072",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 33.0 GB; TTFT 485974 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 269.7,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 128k tokens 269.7 PP tok/s 16.6 TG tok/s 485974 TTFT (ms) 33.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yrx4vd88",
   "source_url": "https://omlx.ai/benchmarks/performance/yrx4vd88",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-9y3jbhxi-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: brzizhyo; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 425.2,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 425.2 28.6 23.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9y3jbhxi",
   "source_url": "https://omlx.ai/benchmarks/performance/9y3jbhxi",
   "tg_tokens": null,
   "tps": 28.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-9y3jbhxi-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: brzizhyo; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 432.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 432.0 24.8 24.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9y3jbhxi",
   "source_url": "https://omlx.ai/benchmarks/performance/9y3jbhxi",
   "tg_tokens": null,
   "tps": 24.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-s0de9301-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 2t8r2gyz; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 428.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 428.0 23.4 23.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s0de9301",
   "source_url": "https://omlx.ai/benchmarks/performance/s0de9301",
   "tg_tokens": null,
   "tps": 23.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-s0de9301-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: 2t8r2gyz; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 472.2,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 472.2 25.0 24.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s0de9301",
   "source_url": "https://omlx.ai/benchmarks/performance/s0de9301",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-srpr0o8k-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 2f1ucbi8; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 414.3,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 414.3 23.2 23.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/srpr0o8k",
   "source_url": "https://omlx.ai/benchmarks/performance/srpr0o8k",
   "tg_tokens": null,
   "tps": 23.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-srpr0o8k-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: 2f1ucbi8; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 437.7,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 437.7 21.5 24.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/srpr0o8k",
   "source_url": "https://omlx.ai/benchmarks/performance/srpr0o8k",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-9f5hh6ri-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 4bpn8hu4; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 440.6,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 440.6 23.8 23.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9f5hh6ri",
   "source_url": "https://omlx.ai/benchmarks/performance/9f5hh6ri",
   "tg_tokens": null,
   "tps": 23.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-9f5hh6ri-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: 4bpn8hu4; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 491.7,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 491.7 26.0 24.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9f5hh6ri",
   "source_url": "https://omlx.ai/benchmarks/performance/9f5hh6ri",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-yymjw4oo-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 75.3 GB at this context; duplicate traces of the same session: sxqqutkg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 471.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 471.3 34.6 75.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yymjw4oo",
   "source_url": "https://omlx.ai/benchmarks/performance/yymjw4oo",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-yymjw4oo-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 75.8 GB at this context; duplicate traces of the same session: sxqqutkg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 593.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 593.1 30.7 75.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yymjw4oo",
   "source_url": "https://omlx.ai/benchmarks/performance/yymjw4oo",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ua2dn0z6-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 57.6 GB; TTFT 20244 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 809.3,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-Flash-Next-oQ3e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 3bit; 16k tokens 809.3 PP tok/s 78.7 TG tok/s 20244 TTFT (ms) 57.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ua2dn0z6",
   "source_url": "https://omlx.ai/benchmarks/performance/ua2dn0z6",
   "tg_tokens": null,
   "tps": 78.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-zybgjov0-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.9 GB; TTFT 4285 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 239.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k tokens 239.0 PP tok/s 35.1 TG tok/s 4285 TTFT (ms) 70.9 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zybgjov0",
   "source_url": "https://omlx.ai/benchmarks/performance/zybgjov0",
   "tg_tokens": null,
   "tps": 35.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-9p73rcj5-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.4 GB at this context; duplicate traces of the same session: vkbhw74d; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 500.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 4k 500.1 39.8 71.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9p73rcj5",
   "source_url": "https://omlx.ai/benchmarks/performance/9p73rcj5",
   "tg_tokens": null,
   "tps": 39.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-9p73rcj5-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: vkbhw74d; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 622.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 16k 622.2 52.8 71.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9p73rcj5",
   "source_url": "https://omlx.ai/benchmarks/performance/9p73rcj5",
   "tg_tokens": null,
   "tps": 52.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-9p73rcj5-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.4 GB at this context; duplicate traces of the same session: vkbhw74d; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 613.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 64k 613.5 46.1 74.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9p73rcj5",
   "source_url": "https://omlx.ai/benchmarks/performance/9p73rcj5",
   "tg_tokens": null,
   "tps": 46.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-i0lmv1da-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 622.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 16k 622.0 18.2 71.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i0lmv1da",
   "source_url": "https://omlx.ai/benchmarks/performance/i0lmv1da",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-i0lmv1da-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.4 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 610.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 64k 610.7 44.4 74.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i0lmv1da",
   "source_url": "https://omlx.ai/benchmarks/performance/i0lmv1da",
   "tg_tokens": null,
   "tps": 44.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2wtcga3j-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.7 GB at this context; duplicate traces of the same session: 2axkqajo, 963zaumm, ac3jt2gu; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1331.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 1,331 56.6 183.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2wtcga3j",
   "source_url": "https://omlx.ai/benchmarks/performance/2wtcga3j",
   "tg_tokens": null,
   "tps": 56.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2wtcga3j-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.9 GB at this context; duplicate traces of the same session: 2axkqajo, 963zaumm, ac3jt2gu; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1347.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 1,347 66.4 183.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2wtcga3j",
   "source_url": "https://omlx.ai/benchmarks/performance/2wtcga3j",
   "tg_tokens": null,
   "tps": 66.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2wtcga3j-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 184.2 GB at this context; duplicate traces of the same session: 2axkqajo, 963zaumm, ac3jt2gu; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1338.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k 1,338 75.6 184.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2wtcga3j",
   "source_url": "https://omlx.ai/benchmarks/performance/2wtcga3j",
   "tg_tokens": null,
   "tps": 75.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2wtcga3j-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 184.8 GB at this context; duplicate traces of the same session: 2axkqajo, 963zaumm, ac3jt2gu; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1334.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 32k 1,334 75.1 184.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2wtcga3j",
   "source_url": "https://omlx.ai/benchmarks/performance/2wtcga3j",
   "tg_tokens": null,
   "tps": 75.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-a92beaq2-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.7 GB at this context; duplicate traces of the same session: 1fm35io3, nsvi7ha6, te28qons; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1368.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 1,368 56.5 180.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a92beaq2",
   "source_url": "https://omlx.ai/benchmarks/performance/a92beaq2",
   "tg_tokens": null,
   "tps": 56.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-a92beaq2-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.8 GB at this context; duplicate traces of the same session: 1fm35io3, nsvi7ha6, te28qons; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1396.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 1,396 56.1 180.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a92beaq2",
   "source_url": "https://omlx.ai/benchmarks/performance/a92beaq2",
   "tg_tokens": null,
   "tps": 56.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-a92beaq2-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.0 GB at this context; duplicate traces of the same session: 1fm35io3, nsvi7ha6, te28qons; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1386.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k 1,386 51.9 181.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a92beaq2",
   "source_url": "https://omlx.ai/benchmarks/performance/a92beaq2",
   "tg_tokens": null,
   "tps": 51.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-a92beaq2-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.6 GB at this context; duplicate traces of the same session: 1fm35io3, nsvi7ha6, te28qons; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1375.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 32k 1,375 46.3 181.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a92beaq2",
   "source_url": "https://omlx.ai/benchmarks/performance/a92beaq2",
   "tg_tokens": null,
   "tps": 46.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0chzn2pk-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: wrwnw5ne; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2204.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 2,204 136.9 28.0 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0chzn2pk",
   "source_url": "https://omlx.ai/benchmarks/performance/0chzn2pk",
   "tg_tokens": null,
   "tps": 136.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0chzn2pk-ctx8192",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.2 GB at this context; duplicate traces of the same session: wrwnw5ne; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2253.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 2,253 78.4 28.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0chzn2pk",
   "source_url": "https://omlx.ai/benchmarks/performance/0chzn2pk",
   "tg_tokens": null,
   "tps": 78.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0chzn2pk-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: wrwnw5ne; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2132.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k 2,132 84.1 28.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0chzn2pk",
   "source_url": "https://omlx.ai/benchmarks/performance/0chzn2pk",
   "tg_tokens": null,
   "tps": 84.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-12",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0chzn2pk-ctx32768",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.4 GB at this context; duplicate traces of the same session: wrwnw5ne; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1918.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 32k 1,918 65.4 29.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0chzn2pk",
   "source_url": "https://omlx.ai/benchmarks/performance/0chzn2pk",
   "tg_tokens": null,
   "tps": 65.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-uhk41xpa-ctx4096",
   "model": "gemma-4-E2B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 3.3 GB at this context; duplicate traces of the same session: fspeginr; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 869.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E2B-it-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 869.4 39.1 3.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uhk41xpa",
   "source_url": "https://omlx.ai/benchmarks/performance/uhk41xpa",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-uhk41xpa-ctx16384",
   "model": "gemma-4-E2B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 3.5 GB at this context; duplicate traces of the same session: fspeginr; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1198.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E2B-it-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 1,198 31.0 3.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uhk41xpa",
   "source_url": "https://omlx.ai/benchmarks/performance/uhk41xpa",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-3gx9zziu-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 5.0 GB at this context; duplicate traces of the same session: w8yl0c4s; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 457.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 457.3 22.2 5.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3gx9zziu",
   "source_url": "https://omlx.ai/benchmarks/performance/3gx9zziu",
   "tg_tokens": null,
   "tps": 22.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-3gx9zziu-ctx16384",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 5.7 GB at this context; duplicate traces of the same session: w8yl0c4s; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 287.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 287.9 34.0 5.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3gx9zziu",
   "source_url": "https://omlx.ai/benchmarks/performance/3gx9zziu",
   "tg_tokens": null,
   "tps": 34.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-qptnt0bj-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 76.0 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 192.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-Flash-Next-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 4k 192.9 43.2 76.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qptnt0bj",
   "source_url": "https://omlx.ai/benchmarks/performance/qptnt0bj",
   "tg_tokens": null,
   "tps": 43.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-qptnt0bj-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 76.2 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 653.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-Flash-Next-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 16k 653.5 49.6 76.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qptnt0bj",
   "source_url": "https://omlx.ai/benchmarks/performance/qptnt0bj",
   "tg_tokens": null,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-qptnt0bj-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 77.7 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 643.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-Flash-Next-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 64k 643.8 44.9 77.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qptnt0bj",
   "source_url": "https://omlx.ai/benchmarks/performance/qptnt0bj",
   "tg_tokens": null,
   "tps": 44.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-6voxeklf-ctx32768",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 179.5 GB; TTFT 18142 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1806.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k tokens 1,806 PP tok/s 94.9 TG tok/s 18142 TTFT (ms) 179.5 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6voxeklf",
   "source_url": "https://omlx.ai/benchmarks/performance/6voxeklf",
   "tg_tokens": null,
   "tps": 94.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-26xo87ty-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 179.5 GB; TTFT 18027 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1818.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k tokens 1,818 PP tok/s 79.0 TG tok/s 18027 TTFT (ms) 179.5 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/26xo87ty",
   "source_url": "https://omlx.ai/benchmarks/performance/26xo87ty",
   "tg_tokens": null,
   "tps": 79.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-q6um3txr-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 179.5 GB; TTFT 18154 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1805.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k tokens 1,805 PP tok/s 82.8 TG tok/s 18154 TTFT (ms) 179.5 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q6um3txr",
   "source_url": "https://omlx.ai/benchmarks/performance/q6um3txr",
   "tg_tokens": null,
   "tps": 82.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-jwpcmif5-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 179.4 GB; TTFT 17981 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1822.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k tokens 1,822 PP tok/s 84.2 TG tok/s 17981 TTFT (ms) 179.4 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jwpcmif5",
   "source_url": "https://omlx.ai/benchmarks/performance/jwpcmif5",
   "tg_tokens": null,
   "tps": 84.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-kx1885j0-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 179.5 GB; TTFT 18057 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1815.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k tokens 1,815 PP tok/s 95.4 TG tok/s 18057 TTFT (ms) 179.5 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kx1885j0",
   "source_url": "https://omlx.ai/benchmarks/performance/kx1885j0",
   "tg_tokens": null,
   "tps": 95.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-afwg6e2e-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 179.5 GB; TTFT 18004 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1820.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k tokens 1,820 PP tok/s 96.9 TG tok/s 18004 TTFT (ms) 179.5 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/afwg6e2e",
   "source_url": "https://omlx.ai/benchmarks/performance/afwg6e2e",
   "tg_tokens": null,
   "tps": 96.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-dog4iz8b-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 179.5 GB; TTFT 18131 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1807.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k tokens 1,807 PP tok/s 90.1 TG tok/s 18131 TTFT (ms) 179.5 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dog4iz8b",
   "source_url": "https://omlx.ai/benchmarks/performance/dog4iz8b",
   "tg_tokens": null,
   "tps": 90.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5v7alt17-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 250.2 GB at this context; duplicate traces of the same session: 1flcbd04, m2xfq66j, ydozrxn3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 573.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 4k 573.4 35.5 250.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5v7alt17",
   "source_url": "https://omlx.ai/benchmarks/performance/5v7alt17",
   "tg_tokens": null,
   "tps": 35.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5v7alt17-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 250.9 GB at this context; duplicate traces of the same session: 1flcbd04, m2xfq66j, ydozrxn3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 534.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 8k 534.8 35.4 250.9 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5v7alt17",
   "source_url": "https://omlx.ai/benchmarks/performance/5v7alt17",
   "tg_tokens": null,
   "tps": 35.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5v7alt17-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 251.0 GB at this context; duplicate traces of the same session: 1flcbd04, m2xfq66j, ydozrxn3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 526.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 16k 526.0 33.1 251.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5v7alt17",
   "source_url": "https://omlx.ai/benchmarks/performance/5v7alt17",
   "tg_tokens": null,
   "tps": 33.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5v7alt17-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 251.1 GB at this context; duplicate traces of the same session: 1flcbd04, m2xfq66j, ydozrxn3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 516.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 32k 516.7 34.7 251.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5v7alt17",
   "source_url": "https://omlx.ai/benchmarks/performance/5v7alt17",
   "tg_tokens": null,
   "tps": 34.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-6yd6yibn-ctx1024",
   "model": "Occamy-1.0",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: k6t0at5q; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2082.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Occamy-1.0-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 1k 2,082 82.7 28.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6yd6yibn",
   "source_url": "https://omlx.ai/benchmarks/performance/6yd6yibn",
   "tg_tokens": null,
   "tps": 82.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-6yd6yibn-ctx16384",
   "model": "Occamy-1.0",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.1 GB at this context; duplicate traces of the same session: k6t0at5q; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2237.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Occamy-1.0-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 16k 2,237 68.7 30.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6yd6yibn",
   "source_url": "https://omlx.ai/benchmarks/performance/6yd6yibn",
   "tg_tokens": null,
   "tps": 68.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-7t6nhn29-ctx1024",
   "model": "Occamy-1.0",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.2 GB at this context; duplicate traces of the same session: aa936wjy; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2015.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Occamy-1.0-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 1k 2,015 151.2 30.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7t6nhn29",
   "source_url": "https://omlx.ai/benchmarks/performance/7t6nhn29",
   "tg_tokens": null,
   "tps": 151.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-7t6nhn29-ctx16384",
   "model": "Occamy-1.0",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.6 GB at this context; duplicate traces of the same session: aa936wjy; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2142.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Occamy-1.0-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 16k 2,142 166.5 32.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7t6nhn29",
   "source_url": "https://omlx.ai/benchmarks/performance/7t6nhn29",
   "tg_tokens": null,
   "tps": 166.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-z3wvikbx-ctx1024",
   "model": "Occamy-1.0",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: psi7w7n3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2104.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Occamy-1.0-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 1k 2,104 92.6 28.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z3wvikbx",
   "source_url": "https://omlx.ai/benchmarks/performance/z3wvikbx",
   "tg_tokens": null,
   "tps": 92.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-z3wvikbx-ctx16384",
   "model": "Occamy-1.0",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.1 GB at this context; duplicate traces of the same session: psi7w7n3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2249.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Occamy-1.0-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 16k 2,249 83.9 30.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z3wvikbx",
   "source_url": "https://omlx.ai/benchmarks/performance/z3wvikbx",
   "tg_tokens": null,
   "tps": 83.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-e4l04ilf-ctx1024",
   "model": "Occamy-1.0",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.7 GB at this context; duplicate traces of the same session: c3l5zpj3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2187.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Occamy-1.0-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 1k 2,187 113.3 28.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e4l04ilf",
   "source_url": "https://omlx.ai/benchmarks/performance/e4l04ilf",
   "tg_tokens": null,
   "tps": 113.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-e4l04ilf-ctx16384",
   "model": "Occamy-1.0",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: c3l5zpj3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2026.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Occamy-1.0-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 16k 2,026 114.6 29.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e4l04ilf",
   "source_url": "https://omlx.ai/benchmarks/performance/e4l04ilf",
   "tg_tokens": null,
   "tps": 114.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-iyb648rn-ctx1024",
   "model": "Occamy-1.0",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: 3jm2cl59; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2099.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Occamy-1.0-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 1k 2,099 92.9 28.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iyb648rn",
   "source_url": "https://omlx.ai/benchmarks/performance/iyb648rn",
   "tg_tokens": null,
   "tps": 92.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-iyb648rn-ctx16384",
   "model": "Occamy-1.0",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.1 GB at this context; duplicate traces of the same session: 3jm2cl59; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2248.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Occamy-1.0-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 16k 2,248 84.0 30.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iyb648rn",
   "source_url": "https://omlx.ai/benchmarks/performance/iyb648rn",
   "tg_tokens": null,
   "tps": 84.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9t63cjll-ctx32768",
   "model": "Occamy-1.0",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 32.8 GB; TTFT 20334 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1612.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Occamy-1.0-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 32k tokens 1,612 PP tok/s 26.5 TG tok/s 20334 TTFT (ms) 32.8 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9t63cjll",
   "source_url": "https://omlx.ai/benchmarks/performance/9t63cjll",
   "tg_tokens": null,
   "tps": 26.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-jeihd6ob-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 19.9 GB; TTFT 1931 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2121.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k tokens 2,121 PP tok/s 67.0 TG tok/s 1931 TTFT (ms) 19.9 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jeihd6ob",
   "source_url": "https://omlx.ai/benchmarks/performance/jeihd6ob",
   "tg_tokens": null,
   "tps": 67.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-ia3bskez-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 21.1 GB; TTFT 57224 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1145.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 64k tokens 1,145 PP tok/s 37.9 TG tok/s 57224 TTFT (ms) 21.1 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ia3bskez",
   "source_url": "https://omlx.ai/benchmarks/performance/ia3bskez",
   "tg_tokens": null,
   "tps": 37.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-qgk3b6rg-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 19.9 GB; TTFT 1964 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2085.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k tokens 2,085 PP tok/s 69.8 TG tok/s 1964 TTFT (ms) 19.9 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qgk3b6rg",
   "source_url": "https://omlx.ai/benchmarks/performance/qgk3b6rg",
   "tg_tokens": null,
   "tps": 69.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-mto2egxc-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.3 GB; TTFT 647 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1582.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k tokens 1,582 PP tok/s 69.6 TG tok/s 647 TTFT (ms) 19.3 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mto2egxc",
   "source_url": "https://omlx.ai/benchmarks/performance/mto2egxc",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-shg67qpa-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: g5kqbda5; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2093.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 2,093 69.5 19.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/shg67qpa",
   "source_url": "https://omlx.ai/benchmarks/performance/shg67qpa",
   "tg_tokens": null,
   "tps": 69.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-shg67qpa-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: g5kqbda5; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2077.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 8k 2,077 67.2 19.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/shg67qpa",
   "source_url": "https://omlx.ai/benchmarks/performance/shg67qpa",
   "tg_tokens": null,
   "tps": 67.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-p7izffdd-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 20.1 GB; TTFT 8639 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1897.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k tokens 1,897 PP tok/s 61.4 TG tok/s 8639 TTFT (ms) 20.1 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p7izffdd",
   "source_url": "https://omlx.ai/benchmarks/performance/p7izffdd",
   "tg_tokens": null,
   "tps": 61.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7hook3s6-ctx4096",
   "model": "Phi-3.5-mini-instruct",
   "notes": "oMLX community benchmark trace; peak memory 4.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3484.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Phi-3.5-mini-instruct-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 3,484 102.1 4.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7hook3s6",
   "source_url": "https://omlx.ai/benchmarks/performance/7hook3s6",
   "tg_tokens": null,
   "tps": 102.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-7hook3s6-ctx16384",
   "model": "Phi-3.5-mini-instruct",
   "notes": "oMLX community benchmark trace; peak memory 8.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2392.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Phi-3.5-mini-instruct-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,392 47.4 8.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7hook3s6",
   "source_url": "https://omlx.ai/benchmarks/performance/7hook3s6",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-2xomvudh-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: idynx2ad; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 584.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 584.1 53.5 19.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2xomvudh",
   "source_url": "https://omlx.ai/benchmarks/performance/2xomvudh",
   "tg_tokens": null,
   "tps": 53.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-2xomvudh-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: idynx2ad; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1090.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 1,090 56.0 20.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2xomvudh",
   "source_url": "https://omlx.ai/benchmarks/performance/2xomvudh",
   "tg_tokens": null,
   "tps": 56.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-yszhd8zc-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: nx631ver; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 245.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 245.2 36.5 19.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yszhd8zc",
   "source_url": "https://omlx.ai/benchmarks/performance/yszhd8zc",
   "tg_tokens": null,
   "tps": 36.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-yszhd8zc-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: nx631ver; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 247.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 247.5 64.1 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yszhd8zc",
   "source_url": "https://omlx.ai/benchmarks/performance/yszhd8zc",
   "tg_tokens": null,
   "tps": 64.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-t48scpmq-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: zq96e59o; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 424.7,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 424.7 25.1 23.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t48scpmq",
   "source_url": "https://omlx.ai/benchmarks/performance/t48scpmq",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-t48scpmq-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: zq96e59o; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 454.6,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 454.6 22.9 24.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t48scpmq",
   "source_url": "https://omlx.ai/benchmarks/performance/t48scpmq",
   "tg_tokens": null,
   "tps": 22.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dof35y9k-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 863.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 863.7 53.7 21.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dof35y9k",
   "source_url": "https://omlx.ai/benchmarks/performance/dof35y9k",
   "tg_tokens": null,
   "tps": 53.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dof35y9k-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 708.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 708.4 47.4 22.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dof35y9k",
   "source_url": "https://omlx.ai/benchmarks/performance/dof35y9k",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dof35y9k-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 637.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 637.2 56.4 23.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dof35y9k",
   "source_url": "https://omlx.ai/benchmarks/performance/dof35y9k",
   "tg_tokens": null,
   "tps": 56.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-8yodyo0s-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: 5ipramtb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 253.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 253.8 38.1 23.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8yodyo0s",
   "source_url": "https://omlx.ai/benchmarks/performance/8yodyo0s",
   "tg_tokens": null,
   "tps": 38.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-8yodyo0s-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: 5ipramtb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 262.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 262.4 34.4 24.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8yodyo0s",
   "source_url": "https://omlx.ai/benchmarks/performance/8yodyo0s",
   "tg_tokens": null,
   "tps": 34.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-b2jx3hss-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 98n82ioo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 233.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 233.6 42.6 18.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b2jx3hss",
   "source_url": "https://omlx.ai/benchmarks/performance/b2jx3hss",
   "tg_tokens": null,
   "tps": 42.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-b2jx3hss-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: 98n82ioo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 244.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 244.4 75.5 22.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b2jx3hss",
   "source_url": "https://omlx.ai/benchmarks/performance/b2jx3hss",
   "tg_tokens": null,
   "tps": 75.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-agbbygpt-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: efpan28h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 836.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 836.9 59.5 26.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/agbbygpt",
   "source_url": "https://omlx.ai/benchmarks/performance/agbbygpt",
   "tg_tokens": null,
   "tps": 59.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-agbbygpt-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: efpan28h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 441.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 441.4 45.8 27.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/agbbygpt",
   "source_url": "https://omlx.ai/benchmarks/performance/agbbygpt",
   "tg_tokens": null,
   "tps": 45.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-o1fout6c-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.9 GB at this context; duplicate traces of the same session: b3ldh0mv, pktoqzm1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 761.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 761.6 48.2 35.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o1fout6c",
   "source_url": "https://omlx.ai/benchmarks/performance/o1fout6c",
   "tg_tokens": null,
   "tps": 48.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-o1fout6c-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: b3ldh0mv, pktoqzm1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 732.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 732.8 37.4 36.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o1fout6c",
   "source_url": "https://omlx.ai/benchmarks/performance/o1fout6c",
   "tg_tokens": null,
   "tps": 37.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-o1fout6c-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.2 GB at this context; duplicate traces of the same session: b3ldh0mv, pktoqzm1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 631.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 631.8 38.6 39.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o1fout6c",
   "source_url": "https://omlx.ai/benchmarks/performance/o1fout6c",
   "tg_tokens": null,
   "tps": 38.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-o1fout6c-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 44.0 GB at this context; duplicate traces of the same session: b3ldh0mv, pktoqzm1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 488.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 488.4 38.4 44.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o1fout6c",
   "source_url": "https://omlx.ai/benchmarks/performance/o1fout6c",
   "tg_tokens": null,
   "tps": 38.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-o1fout6c-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 49.2 GB at this context; duplicate traces of the same session: b3ldh0mv, pktoqzm1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 407.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 195k 407.4 33.6 49.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o1fout6c",
   "source_url": "https://omlx.ai/benchmarks/performance/o1fout6c",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-edvah6a9-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 195.5 GB; TTFT 14682 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1116.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-Flash-Next-MLX-8bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k tokens 1,116 PP tok/s 46.4 TG tok/s 14682 TTFT (ms) 195.5 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/edvah6a9",
   "source_url": "https://omlx.ai/benchmarks/performance/edvah6a9",
   "tg_tokens": null,
   "tps": 46.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-5i0369zd-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 622.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 16k 622.0 18.2 71.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5i0369zd",
   "source_url": "https://omlx.ai/benchmarks/performance/5i0369zd",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-5i0369zd-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.4 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 610.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 64k 610.7 44.4 74.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5i0369zd",
   "source_url": "https://omlx.ai/benchmarks/performance/5i0369zd",
   "tg_tokens": null,
   "tps": 44.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-ml3pl54m-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.4 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 500.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 4k 500.1 39.8 71.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ml3pl54m",
   "source_url": "https://omlx.ai/benchmarks/performance/ml3pl54m",
   "tg_tokens": null,
   "tps": 39.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-ml3pl54m-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 622.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 16k 622.2 52.8 71.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ml3pl54m",
   "source_url": "https://omlx.ai/benchmarks/performance/ml3pl54m",
   "tg_tokens": null,
   "tps": 52.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-ml3pl54m-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.4 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 613.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 64k 613.5 46.1 74.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ml3pl54m",
   "source_url": "https://omlx.ai/benchmarks/performance/ml3pl54m",
   "tg_tokens": null,
   "tps": 46.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-w9n4ppt5-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 88.9 GB; TTFT 8214 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1995.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 5bit; 16k tokens 1,995 PP tok/s 83.2 TG tok/s 8214 TTFT (ms) 88.9 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w9n4ppt5",
   "source_url": "https://omlx.ai/benchmarks/performance/w9n4ppt5",
   "tg_tokens": null,
   "tps": 83.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-x4e5fyc6-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 89.5 GB; TTFT 11900 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2754.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 5bit; 32k tokens 2,754 PP tok/s 83.1 TG tok/s 11900 TTFT (ms) 89.5 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x4e5fyc6",
   "source_url": "https://omlx.ai/benchmarks/performance/x4e5fyc6",
   "tg_tokens": null,
   "tps": 83.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-vxwyh9c7-ctx16384",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MTPLX Optimized Speed build; oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 1rpds11t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 552.5,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Ternary-Bonsai-2-27B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 unknown; 16k 552.5 34.8 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vxwyh9c7",
   "source_url": "https://omlx.ai/benchmarks/performance/vxwyh9c7",
   "tg_tokens": null,
   "tps": 34.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-vxwyh9c7-ctx131072",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MTPLX Optimized Speed build; oMLX community benchmark trace; peak memory 30.2 GB at this context; duplicate traces of the same session: 1rpds11t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 236.9,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Ternary-Bonsai-2-27B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 unknown; 128k 236.9 16.4 30.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vxwyh9c7",
   "source_url": "https://omlx.ai/benchmarks/performance/vxwyh9c7",
   "tg_tokens": null,
   "tps": 16.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-h1xocrq2-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 28.2 GB at this context; duplicate traces of the same session: g26iv8re, vgn39m49; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1921.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 1,921 79.8 28.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h1xocrq2",
   "source_url": "https://omlx.ai/benchmarks/performance/h1xocrq2",
   "tg_tokens": null,
   "tps": 79.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-h1xocrq2-ctx16384",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: g26iv8re, vgn39m49; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3269.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 3,269 84.6 28.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h1xocrq2",
   "source_url": "https://omlx.ai/benchmarks/performance/h1xocrq2",
   "tg_tokens": null,
   "tps": 84.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-h1xocrq2-ctx65536",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 31.0 GB at this context; duplicate traces of the same session: g26iv8re, vgn39m49; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2098.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 2,098 54.3 31.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h1xocrq2",
   "source_url": "https://omlx.ai/benchmarks/performance/h1xocrq2",
   "tg_tokens": null,
   "tps": 54.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-kvaobp2s-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 76.0 GB at this context; duplicate traces of the same session: 1ngr006u; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 192.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-Flash-Next-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 4k 192.9 43.2 76.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kvaobp2s",
   "source_url": "https://omlx.ai/benchmarks/performance/kvaobp2s",
   "tg_tokens": null,
   "tps": 43.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-kvaobp2s-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 76.2 GB at this context; duplicate traces of the same session: 1ngr006u; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 653.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-Flash-Next-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 16k 653.5 49.6 76.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kvaobp2s",
   "source_url": "https://omlx.ai/benchmarks/performance/kvaobp2s",
   "tg_tokens": null,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-kvaobp2s-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 77.7 GB at this context; duplicate traces of the same session: 1ngr006u; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 643.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-Flash-Next-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 64k 643.8 44.9 77.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kvaobp2s",
   "source_url": "https://omlx.ai/benchmarks/performance/kvaobp2s",
   "tg_tokens": null,
   "tps": 44.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-xajr0mft-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 76.0 GB at this context; duplicate traces of the same session: uo3rzv89; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 530.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-Flash-Next-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 4k 530.4 47.8 76.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xajr0mft",
   "source_url": "https://omlx.ai/benchmarks/performance/xajr0mft",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-xajr0mft-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 76.2 GB at this context; duplicate traces of the same session: uo3rzv89; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 652.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-Flash-Next-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 16k 652.4 48.2 76.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xajr0mft",
   "source_url": "https://omlx.ai/benchmarks/performance/xajr0mft",
   "tg_tokens": null,
   "tps": 48.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-pan8ifm4-ctx131072",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "REAP-19B pruned variant",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 112.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-oQ4e-REAP-19B on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 128k tokens 112.8 PP tok/s 15.0 TG tok/s 1162485 TTFT (ms) 19.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pan8ifm4",
   "source_url": "https://omlx.ai/benchmarks/performance/pan8ifm4",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-1emp1ffy-ctx16384",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 179.2 GB; TTFT 8981 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1824.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k tokens 1,824 PP tok/s 59.0 TG tok/s 8981 TTFT (ms) 179.2 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1emp1ffy",
   "source_url": "https://omlx.ai/benchmarks/performance/1emp1ffy",
   "tg_tokens": null,
   "tps": 59.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-vnv0rffg-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 179.1 GB; TTFT 8952 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1830.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k tokens 1,830 PP tok/s 59.2 TG tok/s 8952 TTFT (ms) 179.1 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vnv0rffg",
   "source_url": "https://omlx.ai/benchmarks/performance/vnv0rffg",
   "tg_tokens": null,
   "tps": 59.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-2zgnr05t-ctx1024",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: njwa9odf; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 481.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Gemma-4-26B-A4B-IT-6Bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 6bit; 1k 481.6 43.9 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2zgnr05t",
   "source_url": "https://omlx.ai/benchmarks/performance/2zgnr05t",
   "tg_tokens": null,
   "tps": 43.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-2zgnr05t-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: njwa9odf; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 489.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Gemma-4-26B-A4B-IT-6Bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 6bit; 4k 489.5 42.1 20.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2zgnr05t",
   "source_url": "https://omlx.ai/benchmarks/performance/2zgnr05t",
   "tg_tokens": null,
   "tps": 42.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-2zgnr05t-ctx8192",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: njwa9odf; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 480.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Gemma-4-26B-A4B-IT-6Bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 6bit; 8k 480.1 40.9 20.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2zgnr05t",
   "source_url": "https://omlx.ai/benchmarks/performance/2zgnr05t",
   "tg_tokens": null,
   "tps": 40.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9vzqlj7s-ctx32768",
   "model": "Occamy-1.0",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 33.3 GB; TTFT 18671 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1755.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Occamy-1.0-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 32k tokens 1,755 PP tok/s 158.8 TG tok/s 18671 TTFT (ms) 33.3 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9vzqlj7s",
   "source_url": "https://omlx.ai/benchmarks/performance/9vzqlj7s",
   "tg_tokens": null,
   "tps": 158.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8lc4jxsc-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.1 GB at this context; duplicate traces of the same session: noi4qysx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1084.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,084 80.3 43.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8lc4jxsc",
   "source_url": "https://omlx.ai/benchmarks/performance/8lc4jxsc",
   "tg_tokens": null,
   "tps": 80.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8lc4jxsc-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.9 GB at this context; duplicate traces of the same session: noi4qysx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3372.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 3,372 79.7 43.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8lc4jxsc",
   "source_url": "https://omlx.ai/benchmarks/performance/8lc4jxsc",
   "tg_tokens": null,
   "tps": 79.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-b21h390g-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 52.0 GB; TTFT 1774 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 577.2,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.8-27B on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k tokens 577.2 PP tok/s 10.2 TG tok/s 1774 TTFT (ms) 52.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b21h390g",
   "source_url": "https://omlx.ai/benchmarks/performance/b21h390g",
   "tg_tokens": null,
   "tps": 10.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-84qsrnzf-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: gj5zhz0z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 911.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 911.0 32.9 16.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/84qsrnzf",
   "source_url": "https://omlx.ai/benchmarks/performance/84qsrnzf",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-84qsrnzf-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: gj5zhz0z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 817.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 817.5 32.4 17.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/84qsrnzf",
   "source_url": "https://omlx.ai/benchmarks/performance/84qsrnzf",
   "tg_tokens": null,
   "tps": 32.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dke50j8x-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 8mraqdbu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 925.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 925.6 33.1 16.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dke50j8x",
   "source_url": "https://omlx.ai/benchmarks/performance/dke50j8x",
   "tg_tokens": null,
   "tps": 33.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dke50j8x-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 8mraqdbu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 860.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 860.7 32.5 17.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dke50j8x",
   "source_url": "https://omlx.ai/benchmarks/performance/dke50j8x",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-37wltthy-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: tadl4aa9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 865.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 865.5 32.9 16.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/37wltthy",
   "source_url": "https://omlx.ai/benchmarks/performance/37wltthy",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-37wltthy-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: tadl4aa9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 999.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 999.2 32.3 17.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/37wltthy",
   "source_url": "https://omlx.ai/benchmarks/performance/37wltthy",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-2i9v74g1-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 22.8 GB; TTFT 82849 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 395.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k tokens 395.5 PP tok/s 72.3 TG tok/s 82849 TTFT (ms) 22.8 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2i9v74g1",
   "source_url": "https://omlx.ai/benchmarks/performance/2i9v74g1",
   "tg_tokens": null,
   "tps": 72.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1s935tsp-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 600hnugb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 863.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 863.7 53.7 21.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1s935tsp",
   "source_url": "https://omlx.ai/benchmarks/performance/1s935tsp",
   "tg_tokens": null,
   "tps": 53.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1s935tsp-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: 600hnugb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 708.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 708.4 47.4 22.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1s935tsp",
   "source_url": "https://omlx.ai/benchmarks/performance/1s935tsp",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1s935tsp-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: 600hnugb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 637.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 637.2 56.4 23.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1s935tsp",
   "source_url": "https://omlx.ai/benchmarks/performance/1s935tsp",
   "tg_tokens": null,
   "tps": 56.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-w1q0iyca-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: qqoaf29i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1338.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,338 44.6 19.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w1q0iyca",
   "source_url": "https://omlx.ai/benchmarks/performance/w1q0iyca",
   "tg_tokens": null,
   "tps": 44.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-w1q0iyca-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: qqoaf29i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1396.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,396 50.1 20.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w1q0iyca",
   "source_url": "https://omlx.ai/benchmarks/performance/w1q0iyca",
   "tg_tokens": null,
   "tps": 50.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-d19nln6e-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 0i0hm9bn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1312.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,312 40.4 19.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d19nln6e",
   "source_url": "https://omlx.ai/benchmarks/performance/d19nln6e",
   "tg_tokens": null,
   "tps": 40.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-d19nln6e-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 0i0hm9bn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1368.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,368 42.9 19.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d19nln6e",
   "source_url": "https://omlx.ai/benchmarks/performance/d19nln6e",
   "tg_tokens": null,
   "tps": 42.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-v4hyzavk-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: qi7wsv5d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 738.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 738.7 54.7 22.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v4hyzavk",
   "source_url": "https://omlx.ai/benchmarks/performance/v4hyzavk",
   "tg_tokens": null,
   "tps": 54.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-v4hyzavk-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: qi7wsv5d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 630.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 630.4 57.4 23.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v4hyzavk",
   "source_url": "https://omlx.ai/benchmarks/performance/v4hyzavk",
   "tg_tokens": null,
   "tps": 57.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5c3g50qj-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 20szsg6e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1314.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,314 40.6 19.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5c3g50qj",
   "source_url": "https://omlx.ai/benchmarks/performance/5c3g50qj",
   "tg_tokens": null,
   "tps": 40.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5c3g50qj-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 20szsg6e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1361.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,361 44.2 19.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5c3g50qj",
   "source_url": "https://omlx.ai/benchmarks/performance/5c3g50qj",
   "tg_tokens": null,
   "tps": 44.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-esnopptc-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 0izdgjgo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1313.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,313 41.0 19.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/esnopptc",
   "source_url": "https://omlx.ai/benchmarks/performance/esnopptc",
   "tg_tokens": null,
   "tps": 41.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-esnopptc-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 0izdgjgo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1361.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,361 36.9 19.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/esnopptc",
   "source_url": "https://omlx.ai/benchmarks/performance/esnopptc",
   "tg_tokens": null,
   "tps": 36.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5dr3zpec-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: 2ata6i97; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1208.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp-fp16 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,208 58.2 21.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5dr3zpec",
   "source_url": "https://omlx.ai/benchmarks/performance/5dr3zpec",
   "tg_tokens": null,
   "tps": 58.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5dr3zpec-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 2ata6i97; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1381.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp-fp16 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,381 74.1 21.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5dr3zpec",
   "source_url": "https://omlx.ai/benchmarks/performance/5dr3zpec",
   "tg_tokens": null,
   "tps": 74.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pl1x3opc-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 5v59m8u7, ea0cgque, yygn7tn1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1375.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,375 64.2 20.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pl1x3opc",
   "source_url": "https://omlx.ai/benchmarks/performance/pl1x3opc",
   "tg_tokens": null,
   "tps": 64.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-pl1x3opc-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: 5v59m8u7, ea0cgque, yygn7tn1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1256.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 1,256 140.8 21.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pl1x3opc",
   "source_url": "https://omlx.ai/benchmarks/performance/pl1x3opc",
   "tg_tokens": null,
   "tps": 140.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-pl1x3opc-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 5v59m8u7, ea0cgque, yygn7tn1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1046.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 1,046 58.8 23.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pl1x3opc",
   "source_url": "https://omlx.ai/benchmarks/performance/pl1x3opc",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pl1x3opc-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: 5v59m8u7, ea0cgque, yygn7tn1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 766.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 766.7 45.2 28.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pl1x3opc",
   "source_url": "https://omlx.ai/benchmarks/performance/pl1x3opc",
   "tg_tokens": null,
   "tps": 45.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-e5oux6yn-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.9 GB at this context; duplicate traces of the same session: 3pvnlm5l, f034kizq, xtlvr00a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 900.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k 900.6 62.7 28.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e5oux6yn",
   "source_url": "https://omlx.ai/benchmarks/performance/e5oux6yn",
   "tg_tokens": null,
   "tps": 62.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-e5oux6yn-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: 3pvnlm5l, f034kizq, xtlvr00a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 844.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 844.6 83.4 29.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e5oux6yn",
   "source_url": "https://omlx.ai/benchmarks/performance/e5oux6yn",
   "tg_tokens": null,
   "tps": 83.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-e5oux6yn-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: 3pvnlm5l, f034kizq, xtlvr00a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 741.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 64k 741.8 48.1 32.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e5oux6yn",
   "source_url": "https://omlx.ai/benchmarks/performance/e5oux6yn",
   "tg_tokens": null,
   "tps": 48.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-e5oux6yn-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.8 GB at this context; duplicate traces of the same session: 3pvnlm5l, f034kizq, xtlvr00a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 592.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 128k 592.5 40.9 36.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e5oux6yn",
   "source_url": "https://omlx.ai/benchmarks/performance/e5oux6yn",
   "tg_tokens": null,
   "tps": 40.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-54n8wpeb-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.9 GB at this context; duplicate traces of the same session: x6opciq3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 761.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 761.6 48.2 35.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/54n8wpeb",
   "source_url": "https://omlx.ai/benchmarks/performance/54n8wpeb",
   "tg_tokens": null,
   "tps": 48.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-54n8wpeb-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: x6opciq3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 732.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 732.8 37.4 36.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/54n8wpeb",
   "source_url": "https://omlx.ai/benchmarks/performance/54n8wpeb",
   "tg_tokens": null,
   "tps": 37.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-54n8wpeb-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.2 GB at this context; duplicate traces of the same session: x6opciq3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 631.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 631.8 38.6 39.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/54n8wpeb",
   "source_url": "https://omlx.ai/benchmarks/performance/54n8wpeb",
   "tg_tokens": null,
   "tps": 38.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-54n8wpeb-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 44.0 GB at this context; duplicate traces of the same session: x6opciq3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 488.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 488.4 38.4 44.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/54n8wpeb",
   "source_url": "https://omlx.ai/benchmarks/performance/54n8wpeb",
   "tg_tokens": null,
   "tps": 38.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-54n8wpeb-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 49.2 GB at this context; duplicate traces of the same session: x6opciq3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 407.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 195k 407.4 33.6 49.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/54n8wpeb",
   "source_url": "https://omlx.ai/benchmarks/performance/54n8wpeb",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-g673n1qi-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 142.6 GB at this context; duplicate traces of the same session: gqm69ue5, k1b92yn6, q8tmgji8, vch27jhf, vkso28re; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1096.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 1,096 90.2 142.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g673n1qi",
   "source_url": "https://omlx.ai/benchmarks/performance/g673n1qi",
   "tg_tokens": null,
   "tps": 90.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-g673n1qi-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: gqm69ue5, k1b92yn6, q8tmgji8, vch27jhf, vkso28re; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1358.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 1,358 101.0 143.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g673n1qi",
   "source_url": "https://omlx.ai/benchmarks/performance/g673n1qi",
   "tg_tokens": null,
   "tps": 101.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-g673n1qi-ctx8192",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 143.4 GB at this context; duplicate traces of the same session: gqm69ue5, k1b92yn6, q8tmgji8, vch27jhf, vkso28re; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1334.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 8k 1,334 84.6 143.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g673n1qi",
   "source_url": "https://omlx.ai/benchmarks/performance/g673n1qi",
   "tg_tokens": null,
   "tps": 84.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-g673n1qi-ctx16384",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 143.7 GB at this context; duplicate traces of the same session: gqm69ue5, k1b92yn6, q8tmgji8, vch27jhf, vkso28re; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1346.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 16k 1,346 97.9 143.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g673n1qi",
   "source_url": "https://omlx.ai/benchmarks/performance/g673n1qi",
   "tg_tokens": null,
   "tps": 97.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-g673n1qi-ctx65536",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 146.4 GB at this context; duplicate traces of the same session: gqm69ue5, k1b92yn6, q8tmgji8, vch27jhf, vkso28re; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1321.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 64k 1,321 79.6 146.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g673n1qi",
   "source_url": "https://omlx.ai/benchmarks/performance/g673n1qi",
   "tg_tokens": null,
   "tps": 79.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-g673n1qi-ctx131072",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 150.6 GB at this context; duplicate traces of the same session: gqm69ue5, k1b92yn6, q8tmgji8, vch27jhf, vkso28re; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1297.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 128k 1,297 61.6 150.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g673n1qi",
   "source_url": "https://omlx.ai/benchmarks/performance/g673n1qi",
   "tg_tokens": null,
   "tps": 61.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-g673n1qi-ctx199680",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 151.9 GB at this context; duplicate traces of the same session: gqm69ue5, k1b92yn6, q8tmgji8, vch27jhf, vkso28re; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1271.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 195k 1,271 54.0 151.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g673n1qi",
   "source_url": "https://omlx.ai/benchmarks/performance/g673n1qi",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-i17m1p2r-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "fp16 MTP build; oMLX community benchmark trace, single run at 1k context; peak memory 21.2 GB; TTFT 11786 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 86.9,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 2bit; 1k tokens 86.9 PP tok/s 6.4 TG tok/s 11786 TTFT (ms) 21.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i17m1p2r",
   "source_url": "https://omlx.ai/benchmarks/performance/i17m1p2r",
   "tg_tokens": null,
   "tps": 6.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-p019nqlw-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.4 GB at this context; duplicate traces of the same session: 1amt4fh9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 783.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 783.6 54.5 78.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p019nqlw",
   "source_url": "https://omlx.ai/benchmarks/performance/p019nqlw",
   "tg_tokens": null,
   "tps": 54.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-p019nqlw-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.8 GB at this context; duplicate traces of the same session: 1amt4fh9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 753.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 753.2 50.7 79.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p019nqlw",
   "source_url": "https://omlx.ai/benchmarks/performance/p019nqlw",
   "tg_tokens": null,
   "tps": 50.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2fvhkxu1-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: 2j610llu, dvb3dbd6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 987.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 987.8 63.3 70.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2fvhkxu1",
   "source_url": "https://omlx.ai/benchmarks/performance/2fvhkxu1",
   "tg_tokens": null,
   "tps": 63.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2fvhkxu1-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.9 GB at this context; duplicate traces of the same session: 2j610llu, dvb3dbd6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2458.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,458 89.8 72.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2fvhkxu1",
   "source_url": "https://omlx.ai/benchmarks/performance/2fvhkxu1",
   "tg_tokens": null,
   "tps": 89.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2fvhkxu1-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.7 GB at this context; duplicate traces of the same session: 2j610llu, dvb3dbd6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2174.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,174 26.7 73.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2fvhkxu1",
   "source_url": "https://omlx.ai/benchmarks/performance/2fvhkxu1",
   "tg_tokens": null,
   "tps": 26.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-7x9p3qs7-ctx1024",
   "model": "Qwopus3.8-27B-Flash-V2",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: 0ad0ey92; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 819.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwopus3.8-27B-Flash-V2-oQ8e-fp16-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 1k 819.2 44.7 34.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7x9p3qs7",
   "source_url": "https://omlx.ai/benchmarks/performance/7x9p3qs7",
   "tg_tokens": null,
   "tps": 44.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-7x9p3qs7-ctx4096",
   "model": "Qwopus3.8-27B-Flash-V2",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: 0ad0ey92; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1278.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwopus3.8-27B-Flash-V2-oQ8e-fp16-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 4k 1,278 67.5 35.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7x9p3qs7",
   "source_url": "https://omlx.ai/benchmarks/performance/7x9p3qs7",
   "tg_tokens": null,
   "tps": 67.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-3iwvo70t-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 22.8 GB; TTFT 81526 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 401.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k tokens 401.9 PP tok/s 73.4 TG tok/s 81526 TTFT (ms) 22.8 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3iwvo70t",
   "source_url": "https://omlx.ai/benchmarks/performance/3iwvo70t",
   "tg_tokens": null,
   "tps": 73.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-cum4ccd8-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 22.8 GB; TTFT 81217 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 403.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k tokens 403.5 PP tok/s 73.7 TG tok/s 81217 TTFT (ms) 22.8 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cum4ccd8",
   "source_url": "https://omlx.ai/benchmarks/performance/cum4ccd8",
   "tg_tokens": null,
   "tps": 73.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-95i32duq-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 22.8 GB; TTFT 85095 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 385.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k tokens 385.1 PP tok/s 72.4 TG tok/s 85095 TTFT (ms) 22.8 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/95i32duq",
   "source_url": "https://omlx.ai/benchmarks/performance/95i32duq",
   "tg_tokens": null,
   "tps": 72.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-u8e9v9ul-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: bddj2wms; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 793.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k 793.5 36.5 70.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u8e9v9ul",
   "source_url": "https://omlx.ai/benchmarks/performance/u8e9v9ul",
   "tg_tokens": null,
   "tps": 36.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-u8e9v9ul-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: bddj2wms; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1529.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 4k 1,529 39.5 70.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u8e9v9ul",
   "source_url": "https://omlx.ai/benchmarks/performance/u8e9v9ul",
   "tg_tokens": null,
   "tps": 39.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-cmfwmny5-ctx4096",
   "model": "Phi-3.5-mini-instruct",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 4.2 GB; TTFT 1234 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3319.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Phi-3.5-mini-instruct-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 4k tokens 3,319 PP tok/s 117.3 TG tok/s 1234 TTFT (ms) 4.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cmfwmny5",
   "source_url": "https://omlx.ai/benchmarks/performance/cmfwmny5",
   "tg_tokens": null,
   "tps": 117.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-cnb6xm3l-ctx1024",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 481.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Gemma-4-26B-A4B-IT-6Bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 6bit; 1k 481.6 43.9 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cnb6xm3l",
   "source_url": "https://omlx.ai/benchmarks/performance/cnb6xm3l",
   "tg_tokens": null,
   "tps": 43.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-cnb6xm3l-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 489.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Gemma-4-26B-A4B-IT-6Bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 6bit; 4k 489.5 42.1 20.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cnb6xm3l",
   "source_url": "https://omlx.ai/benchmarks/performance/cnb6xm3l",
   "tg_tokens": null,
   "tps": 42.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-cnb6xm3l-ctx8192",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 480.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Gemma-4-26B-A4B-IT-6Bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 6bit; 8k 480.1 40.9 20.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cnb6xm3l",
   "source_url": "https://omlx.ai/benchmarks/performance/cnb6xm3l",
   "tg_tokens": null,
   "tps": 40.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-bml8zgpp-ctx1024",
   "model": "LFM2.5-8B-A1B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 8.9 GB at this context; duplicate traces of the same session: ln265khl; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1545.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-8B-A1B-MLX-8bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 8bit; 1k 1,545 86.8 8.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bml8zgpp",
   "source_url": "https://omlx.ai/benchmarks/performance/bml8zgpp",
   "tg_tokens": null,
   "tps": 86.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-bml8zgpp-ctx4096",
   "model": "LFM2.5-8B-A1B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 9.3 GB at this context; duplicate traces of the same session: ln265khl; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1643.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-8B-A1B-MLX-8bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 8bit; 4k 1,643 85.7 9.3 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bml8zgpp",
   "source_url": "https://omlx.ai/benchmarks/performance/bml8zgpp",
   "tg_tokens": null,
   "tps": 85.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-doubaemd-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 0khr8vay, a6fmc1u0, dy9i3jkx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2111.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,111 128.7 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/doubaemd",
   "source_url": "https://omlx.ai/benchmarks/performance/doubaemd",
   "tg_tokens": null,
   "tps": 128.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-doubaemd-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 0khr8vay, a6fmc1u0, dy9i3jkx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3971.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 3,971 124.1 20.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/doubaemd",
   "source_url": "https://omlx.ai/benchmarks/performance/doubaemd",
   "tg_tokens": null,
   "tps": 124.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-doubaemd-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 0khr8vay, a6fmc1u0, dy9i3jkx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4329.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 4,329 115.5 20.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/doubaemd",
   "source_url": "https://omlx.ai/benchmarks/performance/doubaemd",
   "tg_tokens": null,
   "tps": 115.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-doubaemd-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 0khr8vay, a6fmc1u0, dy9i3jkx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4020.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 4,020 113.8 21.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/doubaemd",
   "source_url": "https://omlx.ai/benchmarks/performance/doubaemd",
   "tg_tokens": null,
   "tps": 113.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-pmtuas2l-ctx1024",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: j2qmp8e8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 865.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 865.0 31.2 16.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pmtuas2l",
   "source_url": "https://omlx.ai/benchmarks/performance/pmtuas2l",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-pmtuas2l-ctx4096",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: j2qmp8e8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 889.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 889.9 31.1 17.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pmtuas2l",
   "source_url": "https://omlx.ai/benchmarks/performance/pmtuas2l",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-g1i6wg23-ctx1024",
   "model": "Qwen3.5-9B-Claude-4.6-HighIQ-INSTRUCT-HERETIC-UNCENSORED",
   "notes": "oMLX community benchmark trace; peak memory 9.7 GB at this context; duplicate traces of the same session: cgn70x71; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2203.0,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Qwen3.5-9B-Claude-4.6-HighIQ-INSTRUCT-HERETIC-UNCENSORED-MLX-mxfp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 2,203 56.4 9.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g1i6wg23",
   "source_url": "https://omlx.ai/benchmarks/performance/g1i6wg23",
   "tg_tokens": null,
   "tps": 56.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-g1i6wg23-ctx4096",
   "model": "Qwen3.5-9B-Claude-4.6-HighIQ-INSTRUCT-HERETIC-UNCENSORED",
   "notes": "oMLX community benchmark trace; peak memory 10.3 GB at this context; duplicate traces of the same session: cgn70x71; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1781.0,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Qwen3.5-9B-Claude-4.6-HighIQ-INSTRUCT-HERETIC-UNCENSORED-MLX-mxfp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 1,781 56.6 10.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g1i6wg23",
   "source_url": "https://omlx.ai/benchmarks/performance/g1i6wg23",
   "tg_tokens": null,
   "tps": 56.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ec2v34ya-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: jsp806f0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 373.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 373.5 50.5 5.9 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ec2v34ya",
   "source_url": "https://omlx.ai/benchmarks/performance/ec2v34ya",
   "tg_tokens": null,
   "tps": 50.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ec2v34ya-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: jsp806f0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 383.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 383.0 49.0 6.3 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ec2v34ya",
   "source_url": "https://omlx.ai/benchmarks/performance/ec2v34ya",
   "tg_tokens": null,
   "tps": 49.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-z1ot0hpj-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: 4z8nn283; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 372.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 372.0 49.4 5.9 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z1ot0hpj",
   "source_url": "https://omlx.ai/benchmarks/performance/z1ot0hpj",
   "tg_tokens": null,
   "tps": 49.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-z1ot0hpj-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: 4z8nn283; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 362.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 362.2 48.9 6.3 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z1ot0hpj",
   "source_url": "https://omlx.ai/benchmarks/performance/z1ot0hpj",
   "tg_tokens": null,
   "tps": 48.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-jqzrmjrd-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: huod2cwn, rzhzjf5v; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1673.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 1,673 96.7 35.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jqzrmjrd",
   "source_url": "https://omlx.ai/benchmarks/performance/jqzrmjrd",
   "tg_tokens": null,
   "tps": 96.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-jqzrmjrd-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: huod2cwn, rzhzjf5v; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2102.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 2,102 94.6 37.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jqzrmjrd",
   "source_url": "https://omlx.ai/benchmarks/performance/jqzrmjrd",
   "tg_tokens": null,
   "tps": 94.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-jqzrmjrd-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.5 GB at this context; duplicate traces of the same session: huod2cwn, rzhzjf5v; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1989.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 8k 1,989 93.6 37.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jqzrmjrd",
   "source_url": "https://omlx.ai/benchmarks/performance/jqzrmjrd",
   "tg_tokens": null,
   "tps": 93.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-z40w7lbf-ctx1024",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: 8oet8kw9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1016.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,016 113.0 12.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z40w7lbf",
   "source_url": "https://omlx.ai/benchmarks/performance/z40w7lbf",
   "tg_tokens": null,
   "tps": 113.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-z40w7lbf-ctx4096",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: 8oet8kw9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1696.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,696 108.8 13.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z40w7lbf",
   "source_url": "https://omlx.ai/benchmarks/performance/z40w7lbf",
   "tg_tokens": null,
   "tps": 108.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vm5s6dkh-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: bjfrid6u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 440.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 440.4 24.8 16.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vm5s6dkh",
   "source_url": "https://omlx.ai/benchmarks/performance/vm5s6dkh",
   "tg_tokens": null,
   "tps": 24.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vm5s6dkh-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: bjfrid6u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 521.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 521.4 24.3 17.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vm5s6dkh",
   "source_url": "https://omlx.ai/benchmarks/performance/vm5s6dkh",
   "tg_tokens": null,
   "tps": 24.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bat5ljjs-ctx16384",
   "model": "Qwen3.8-27B-Coder390-LynnStyle-Mixed",
   "notes": "A4toA8 build; MTPFix; oMLX community benchmark trace, single run at 16k context; peak memory 22.1 GB; TTFT 19981 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 820.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Coder390-LynnStyle-Mixed-MLX-A4toA8-MTPFix on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k tokens 820.0 PP tok/s 79.3 TG tok/s 19981 TTFT (ms) 22.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bat5ljjs",
   "source_url": "https://omlx.ai/benchmarks/performance/bat5ljjs",
   "tg_tokens": null,
   "tps": 79.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lwmye68v-ctx199680",
   "model": "Qwen3.8-27B-Coder390-LynnStyle-Mixed",
   "notes": "A4toA8 build; MTPFix; oMLX community benchmark trace, single run at 195k context; peak memory 34.6 GB; TTFT 480011 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 416.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Coder390-LynnStyle-Mixed-MLX-A4toA8-MTPFix on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 195k tokens 416.7 PP tok/s 25.1 TG tok/s 480011 TTFT (ms) 34.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lwmye68v",
   "source_url": "https://omlx.ai/benchmarks/performance/lwmye68v",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-07obvhc2-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: oisgoh78; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1275.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,275 46.3 19.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/07obvhc2",
   "source_url": "https://omlx.ai/benchmarks/performance/07obvhc2",
   "tg_tokens": null,
   "tps": 46.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-07obvhc2-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: oisgoh78; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1337.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,337 54.8 19.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/07obvhc2",
   "source_url": "https://omlx.ai/benchmarks/performance/07obvhc2",
   "tg_tokens": null,
   "tps": 54.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wq8rbuea-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 25a572gt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 955.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 955.6 46.4 21.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wq8rbuea",
   "source_url": "https://omlx.ai/benchmarks/performance/wq8rbuea",
   "tg_tokens": null,
   "tps": 46.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wq8rbuea-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: 25a572gt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 936.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 936.7 38.3 23.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wq8rbuea",
   "source_url": "https://omlx.ai/benchmarks/performance/wq8rbuea",
   "tg_tokens": null,
   "tps": 38.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yp8jq6n3-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: fd5ylta2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1302.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,302 29.0 19.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yp8jq6n3",
   "source_url": "https://omlx.ai/benchmarks/performance/yp8jq6n3",
   "tg_tokens": null,
   "tps": 29.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yp8jq6n3-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: fd5ylta2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1380.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,380 26.8 19.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yp8jq6n3",
   "source_url": "https://omlx.ai/benchmarks/performance/yp8jq6n3",
   "tg_tokens": null,
   "tps": 26.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kh33utx1-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: f38gd67k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1299.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,299 75.2 19.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kh33utx1",
   "source_url": "https://omlx.ai/benchmarks/performance/kh33utx1",
   "tg_tokens": null,
   "tps": 75.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-kh33utx1-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: f38gd67k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1362.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,362 95.6 19.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kh33utx1",
   "source_url": "https://omlx.ai/benchmarks/performance/kh33utx1",
   "tg_tokens": null,
   "tps": 95.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-86rwt5kf-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 50dm1w1l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1299.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,299 55.4 19.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/86rwt5kf",
   "source_url": "https://omlx.ai/benchmarks/performance/86rwt5kf",
   "tg_tokens": null,
   "tps": 55.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-86rwt5kf-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 50dm1w1l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1345.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,345 58.9 19.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/86rwt5kf",
   "source_url": "https://omlx.ai/benchmarks/performance/86rwt5kf",
   "tg_tokens": null,
   "tps": 58.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2wtiobmt-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: a5tqw09d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 950.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 950.5 59.5 18.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2wtiobmt",
   "source_url": "https://omlx.ai/benchmarks/performance/2wtiobmt",
   "tg_tokens": null,
   "tps": 59.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2wtiobmt-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: a5tqw09d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 943.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 943.0 57.7 19.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2wtiobmt",
   "source_url": "https://omlx.ai/benchmarks/performance/2wtiobmt",
   "tg_tokens": null,
   "tps": 57.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-c2lk4n6e-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: sg4sonjl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1291.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,291 58.7 19.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c2lk4n6e",
   "source_url": "https://omlx.ai/benchmarks/performance/c2lk4n6e",
   "tg_tokens": null,
   "tps": 58.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-c2lk4n6e-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: sg4sonjl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1347.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,347 54.8 19.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c2lk4n6e",
   "source_url": "https://omlx.ai/benchmarks/performance/c2lk4n6e",
   "tg_tokens": null,
   "tps": 54.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gsgwvwnu-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: d81euf0g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1238.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,238 56.5 19.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gsgwvwnu",
   "source_url": "https://omlx.ai/benchmarks/performance/gsgwvwnu",
   "tg_tokens": null,
   "tps": 56.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gsgwvwnu-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: d81euf0g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1342.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,342 56.5 19.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gsgwvwnu",
   "source_url": "https://omlx.ai/benchmarks/performance/gsgwvwnu",
   "tg_tokens": null,
   "tps": 56.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-y73z6jdv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.3 GB at this context; duplicate traces of the same session: fc5yiim0, opet76s9, uwa1y85o, ztolr3gk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 848.4,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Qwen3.8-27B-mxfp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 848.4 18.8 29.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y73z6jdv",
   "source_url": "https://omlx.ai/benchmarks/performance/y73z6jdv",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-y73z6jdv-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 30.9 GB at this context; duplicate traces of the same session: fc5yiim0, opet76s9, uwa1y85o, ztolr3gk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 772.8,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Qwen3.8-27B-mxfp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 772.8 18.2 30.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y73z6jdv",
   "source_url": "https://omlx.ai/benchmarks/performance/y73z6jdv",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-y73z6jdv-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.4 GB at this context; duplicate traces of the same session: fc5yiim0, opet76s9, uwa1y85o, ztolr3gk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 678.1,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Qwen3.8-27B-mxfp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 678.1 17.6 33.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y73z6jdv",
   "source_url": "https://omlx.ai/benchmarks/performance/y73z6jdv",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-y73z6jdv-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 48.8 GB at this context; duplicate traces of the same session: fc5yiim0, opet76s9, uwa1y85o, ztolr3gk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 416.1,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Qwen3.8-27B-mxfp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 416.1 14.2 48.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y73z6jdv",
   "source_url": "https://omlx.ai/benchmarks/performance/y73z6jdv",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-y73z6jdv-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 59.5 GB at this context; duplicate traces of the same session: fc5yiim0, opet76s9, uwa1y85o, ztolr3gk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 307.9,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Qwen3.8-27B-mxfp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 195k 307.9 12.1 59.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y73z6jdv",
   "source_url": "https://omlx.ai/benchmarks/performance/y73z6jdv",
   "tg_tokens": null,
   "tps": 12.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-d103425d-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: zsf8nhcn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 890.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 890.2 71.9 22.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d103425d",
   "source_url": "https://omlx.ai/benchmarks/performance/d103425d",
   "tg_tokens": null,
   "tps": 71.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-d103425d-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: zsf8nhcn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 978.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 978.4 96.1 24.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d103425d",
   "source_url": "https://omlx.ai/benchmarks/performance/d103425d",
   "tg_tokens": null,
   "tps": 96.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-d103425d-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: zsf8nhcn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 924.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 924.3 50.2 24.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d103425d",
   "source_url": "https://omlx.ai/benchmarks/performance/d103425d",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-d103425d-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: zsf8nhcn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 877.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 877.2 64.1 25.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d103425d",
   "source_url": "https://omlx.ai/benchmarks/performance/d103425d",
   "tg_tokens": null,
   "tps": 64.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-d103425d-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: zsf8nhcn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 803.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 803.5 59.6 25.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d103425d",
   "source_url": "https://omlx.ai/benchmarks/performance/d103425d",
   "tg_tokens": null,
   "tps": 59.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-d103425d-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.2 GB at this context; duplicate traces of the same session: zsf8nhcn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 720.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 720.4 53.3 28.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d103425d",
   "source_url": "https://omlx.ai/benchmarks/performance/d103425d",
   "tg_tokens": null,
   "tps": 53.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-d103425d-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: zsf8nhcn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 545.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 545.1 47.4 32.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d103425d",
   "source_url": "https://omlx.ai/benchmarks/performance/d103425d",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ejr98t7f-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 142.6 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1096.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 1,096 90.2 142.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ejr98t7f",
   "source_url": "https://omlx.ai/benchmarks/performance/ejr98t7f",
   "tg_tokens": null,
   "tps": 90.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ejr98t7f-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1358.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 1,358 101.0 143.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ejr98t7f",
   "source_url": "https://omlx.ai/benchmarks/performance/ejr98t7f",
   "tg_tokens": null,
   "tps": 101.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ejr98t7f-ctx8192",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 143.4 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1334.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 8k 1,334 84.6 143.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ejr98t7f",
   "source_url": "https://omlx.ai/benchmarks/performance/ejr98t7f",
   "tg_tokens": null,
   "tps": 84.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ejr98t7f-ctx16384",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 143.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1346.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 16k 1,346 97.9 143.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ejr98t7f",
   "source_url": "https://omlx.ai/benchmarks/performance/ejr98t7f",
   "tg_tokens": null,
   "tps": 97.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ejr98t7f-ctx65536",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 146.4 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1321.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 64k 1,321 79.6 146.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ejr98t7f",
   "source_url": "https://omlx.ai/benchmarks/performance/ejr98t7f",
   "tg_tokens": null,
   "tps": 79.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ejr98t7f-ctx131072",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 150.6 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1297.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 128k 1,297 61.6 150.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ejr98t7f",
   "source_url": "https://omlx.ai/benchmarks/performance/ejr98t7f",
   "tg_tokens": null,
   "tps": 61.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ejr98t7f-ctx199680",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 151.9 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1271.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 195k 1,271 54.0 151.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ejr98t7f",
   "source_url": "https://omlx.ai/benchmarks/performance/ejr98t7f",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-6thcgvek-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.5 GB at this context; duplicate traces of the same session: xgzxls8t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 783.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 783.3 56.7 78.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6thcgvek",
   "source_url": "https://omlx.ai/benchmarks/performance/6thcgvek",
   "tg_tokens": null,
   "tps": 56.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-6thcgvek-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.8 GB at this context; duplicate traces of the same session: xgzxls8t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 777.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 777.2 52.5 79.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6thcgvek",
   "source_url": "https://omlx.ai/benchmarks/performance/6thcgvek",
   "tg_tokens": null,
   "tps": 52.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-m8bp5o6e-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.5 GB at this context; duplicate traces of the same session: ic82f142; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 783.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 783.7 56.3 78.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m8bp5o6e",
   "source_url": "https://omlx.ai/benchmarks/performance/m8bp5o6e",
   "tg_tokens": null,
   "tps": 56.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-m8bp5o6e-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.8 GB at this context; duplicate traces of the same session: ic82f142; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 775.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 775.4 52.6 79.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m8bp5o6e",
   "source_url": "https://omlx.ai/benchmarks/performance/m8bp5o6e",
   "tg_tokens": null,
   "tps": 52.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-xrslxx5e-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: peneydht; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1415.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k 1,415 53.0 70.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xrslxx5e",
   "source_url": "https://omlx.ai/benchmarks/performance/xrslxx5e",
   "tg_tokens": null,
   "tps": 53.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-xrslxx5e-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: peneydht; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2440.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 4k 2,440 58.0 70.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xrslxx5e",
   "source_url": "https://omlx.ai/benchmarks/performance/xrslxx5e",
   "tg_tokens": null,
   "tps": 58.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-9a0n4hcz-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: syn81lsu; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 953.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k 953.0 39.8 70.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9a0n4hcz",
   "source_url": "https://omlx.ai/benchmarks/performance/9a0n4hcz",
   "tg_tokens": null,
   "tps": 39.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-9a0n4hcz-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: syn81lsu; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1631.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 4k 1,631 39.4 70.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9a0n4hcz",
   "source_url": "https://omlx.ai/benchmarks/performance/9a0n4hcz",
   "tg_tokens": null,
   "tps": 39.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-m0ad5jv1-ctx1024",
   "model": "gemma-4-12B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 12.8 GB at this context; duplicate traces of the same session: n6qenki5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1096.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-12B-it-qat-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 1,096 36.3 12.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m0ad5jv1",
   "source_url": "https://omlx.ai/benchmarks/performance/m0ad5jv1",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-m0ad5jv1-ctx4096",
   "model": "gemma-4-12B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 14.2 GB at this context; duplicate traces of the same session: n6qenki5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1517.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-12B-it-qat-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 1,517 35.5 14.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m0ad5jv1",
   "source_url": "https://omlx.ai/benchmarks/performance/m0ad5jv1",
   "tg_tokens": null,
   "tps": 35.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-b6iz41ub-ctx1024",
   "model": "gemma-4-12b-coder-fable5-composer2.5",
   "notes": "community coder fine-tune; oMLX community benchmark trace; peak memory 7.4 GB at this context; duplicate traces of the same session: je4jtldr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1587.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12b-coder-fable5-composer2.5-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,587 60.7 7.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b6iz41ub",
   "source_url": "https://omlx.ai/benchmarks/performance/b6iz41ub",
   "tg_tokens": null,
   "tps": 60.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-b6iz41ub-ctx4096",
   "model": "gemma-4-12b-coder-fable5-composer2.5",
   "notes": "community coder fine-tune; oMLX community benchmark trace; peak memory 8.6 GB at this context; duplicate traces of the same session: je4jtldr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1618.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12b-coder-fable5-composer2.5-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,618 57.7 8.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b6iz41ub",
   "source_url": "https://omlx.ai/benchmarks/performance/b6iz41ub",
   "tg_tokens": null,
   "tps": 57.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-kn7y6fhx-ctx1024",
   "model": "Agents-A1",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 229 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4470.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Agents-A1-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 4,470 PP tok/s 127.8 TG tok/s 229 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kn7y6fhx",
   "source_url": "https://omlx.ai/benchmarks/performance/kn7y6fhx",
   "tg_tokens": null,
   "tps": 127.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wvy8d38l-ctx1024",
   "model": "Bonsai-27B",
   "notes": "oMLX community benchmark trace; peak memory 7.3 GB at this context; duplicate traces of the same session: xtn84480; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 289.2,
   "provenance": "community",
   "quant": "1bit",
   "quote": "Bonsai-27B-mlx-1bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 289.2 51.9 7.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wvy8d38l",
   "source_url": "https://omlx.ai/benchmarks/performance/wvy8d38l",
   "tg_tokens": null,
   "tps": 51.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wvy8d38l-ctx4096",
   "model": "Bonsai-27B",
   "notes": "oMLX community benchmark trace; peak memory 8.4 GB at this context; duplicate traces of the same session: xtn84480; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 459.9,
   "provenance": "community",
   "quant": "1bit",
   "quote": "Bonsai-27B-mlx-1bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 4k 459.9 52.5 8.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wvy8d38l",
   "source_url": "https://omlx.ai/benchmarks/performance/wvy8d38l",
   "tg_tokens": null,
   "tps": 52.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-hcp3pzh2-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 151.1 GB at this context; duplicate traces of the same session: 4e5yufqe, 6znxx43z, m6vjmbca, mbnwleod, o6vcek7j, pr283dzk; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 508.2,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "GLM-5.3-Flash-oQ3.5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 3bit; 1k 508.2 46.5 151.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hcp3pzh2",
   "source_url": "https://omlx.ai/benchmarks/performance/hcp3pzh2",
   "tg_tokens": null,
   "tps": 46.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-hcp3pzh2-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 152.1 GB at this context; duplicate traces of the same session: 4e5yufqe, 6znxx43z, m6vjmbca, mbnwleod, o6vcek7j, pr283dzk; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 588.4,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "GLM-5.3-Flash-oQ3.5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 3bit; 4k 588.4 53.1 152.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hcp3pzh2",
   "source_url": "https://omlx.ai/benchmarks/performance/hcp3pzh2",
   "tg_tokens": null,
   "tps": 53.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-hcp3pzh2-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 152.4 GB at this context; duplicate traces of the same session: 4e5yufqe, 6znxx43z, m6vjmbca, mbnwleod, o6vcek7j, pr283dzk; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 548.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "GLM-5.3-Flash-oQ3.5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 3bit; 8k 548.0 50.7 152.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hcp3pzh2",
   "source_url": "https://omlx.ai/benchmarks/performance/hcp3pzh2",
   "tg_tokens": null,
   "tps": 50.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-hcp3pzh2-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 152.6 GB at this context; duplicate traces of the same session: 4e5yufqe, 6znxx43z, m6vjmbca, mbnwleod, o6vcek7j, pr283dzk; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 533.9,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "GLM-5.3-Flash-oQ3.5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 3bit; 16k 533.9 89.2 152.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hcp3pzh2",
   "source_url": "https://omlx.ai/benchmarks/performance/hcp3pzh2",
   "tg_tokens": null,
   "tps": 89.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-hcp3pzh2-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 152.7 GB at this context; duplicate traces of the same session: 4e5yufqe, 6znxx43z, m6vjmbca, mbnwleod, o6vcek7j, pr283dzk; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 528.7,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "GLM-5.3-Flash-oQ3.5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 3bit; 32k 528.7 66.8 152.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hcp3pzh2",
   "source_url": "https://omlx.ai/benchmarks/performance/hcp3pzh2",
   "tg_tokens": null,
   "tps": 66.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-hcp3pzh2-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 153.3 GB at this context; duplicate traces of the same session: 4e5yufqe, 6znxx43z, m6vjmbca, mbnwleod, o6vcek7j, pr283dzk; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 521.5,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "GLM-5.3-Flash-oQ3.5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 3bit; 64k 521.5 58.6 153.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hcp3pzh2",
   "source_url": "https://omlx.ai/benchmarks/performance/hcp3pzh2",
   "tg_tokens": null,
   "tps": 58.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-hcp3pzh2-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 154.6 GB at this context; duplicate traces of the same session: 4e5yufqe, 6znxx43z, m6vjmbca, mbnwleod, o6vcek7j, pr283dzk; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 509.9,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "GLM-5.3-Flash-oQ3.5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 3bit; 128k 509.9 97.5 154.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hcp3pzh2",
   "source_url": "https://omlx.ai/benchmarks/performance/hcp3pzh2",
   "tg_tokens": null,
   "tps": 97.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-i0jre1j3-ctx4096",
   "model": "KAT-Coder-V2.5-Dev-VL",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.9 GB at this context; duplicate traces of the same session: 74preoia, s8izc2kq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4339.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "KAT-Coder-V2.5-Dev-VL-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 4,339 101.7 30.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i0jre1j3",
   "source_url": "https://omlx.ai/benchmarks/performance/i0jre1j3",
   "tg_tokens": null,
   "tps": 101.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-i0jre1j3-ctx16384",
   "model": "KAT-Coder-V2.5-Dev-VL",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.1 GB at this context; duplicate traces of the same session: 74preoia, s8izc2kq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4382.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "KAT-Coder-V2.5-Dev-VL-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 4,382 123.7 31.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i0jre1j3",
   "source_url": "https://omlx.ai/benchmarks/performance/i0jre1j3",
   "tg_tokens": null,
   "tps": 123.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-i0jre1j3-ctx65536",
   "model": "KAT-Coder-V2.5-Dev-VL",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.2 GB at this context; duplicate traces of the same session: 74preoia, s8izc2kq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2515.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "KAT-Coder-V2.5-Dev-VL-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k 2,515 111.1 32.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i0jre1j3",
   "source_url": "https://omlx.ai/benchmarks/performance/i0jre1j3",
   "tg_tokens": null,
   "tps": 111.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-134t2zo4-ctx1024",
   "model": "Qwen3.5-9B-OmniCoder-Claude-Polaris",
   "notes": "oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: ead26gq7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2813.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.5-9B-OmniCoder-Claude-Polaris on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 2,813 31.2 17.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/134t2zo4",
   "source_url": "https://omlx.ai/benchmarks/performance/134t2zo4",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-134t2zo4-ctx4096",
   "model": "Qwen3.5-9B-OmniCoder-Claude-Polaris",
   "notes": "oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: ead26gq7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3101.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.5-9B-OmniCoder-Claude-Polaris on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 4k 3,101 31.4 18.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/134t2zo4",
   "source_url": "https://omlx.ai/benchmarks/performance/134t2zo4",
   "tg_tokens": null,
   "tps": 31.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-t10b0wo4-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 232 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4424.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 4,424 PP tok/s 115.4 TG tok/s 232 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t10b0wo4",
   "source_url": "https://omlx.ai/benchmarks/performance/t10b0wo4",
   "tg_tokens": null,
   "tps": 115.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-gf1k91p2-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: nnv0tz68, xhxhzfkp, zckp6sad; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1770.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 1,770 96.3 35.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gf1k91p2",
   "source_url": "https://omlx.ai/benchmarks/performance/gf1k91p2",
   "tg_tokens": null,
   "tps": 96.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-gf1k91p2-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: nnv0tz68, xhxhzfkp, zckp6sad; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2102.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 2,102 94.6 37.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gf1k91p2",
   "source_url": "https://omlx.ai/benchmarks/performance/gf1k91p2",
   "tg_tokens": null,
   "tps": 94.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-gf1k91p2-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.5 GB at this context; duplicate traces of the same session: nnv0tz68, xhxhzfkp, zckp6sad; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1991.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 8k 1,991 93.5 37.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gf1k91p2",
   "source_url": "https://omlx.ai/benchmarks/performance/gf1k91p2",
   "tg_tokens": null,
   "tps": 93.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-gf1k91p2-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.8 GB at this context; duplicate traces of the same session: nnv0tz68, xhxhzfkp, zckp6sad; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1788.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 16k 1,788 86.8 37.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gf1k91p2",
   "source_url": "https://omlx.ai/benchmarks/performance/gf1k91p2",
   "tg_tokens": null,
   "tps": 86.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-c5zjz54r-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 7m5r8sa4; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 565.3,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 565.3 58.4 21.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c5zjz54r",
   "source_url": "https://omlx.ai/benchmarks/performance/c5zjz54r",
   "tg_tokens": null,
   "tps": 58.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-c5zjz54r-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 7m5r8sa4; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 603.1,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 603.1 55.3 23.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c5zjz54r",
   "source_url": "https://omlx.ai/benchmarks/performance/c5zjz54r",
   "tg_tokens": null,
   "tps": 55.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-81uh7e4q-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V13-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: c3np05q3, h1b70red, t42zffhq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1867.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V13-dequantized-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,867 116.8 21.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/81uh7e4q",
   "source_url": "https://omlx.ai/benchmarks/performance/81uh7e4q",
   "tg_tokens": null,
   "tps": 116.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-81uh7e4q-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V13-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: c3np05q3, h1b70red, t42zffhq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1945.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V13-dequantized-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,945 95.8 22.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/81uh7e4q",
   "source_url": "https://omlx.ai/benchmarks/performance/81uh7e4q",
   "tg_tokens": null,
   "tps": 95.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-81uh7e4q-ctx8192",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V13-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: c3np05q3, h1b70red, t42zffhq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1833.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V13-dequantized-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,833 65.7 22.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/81uh7e4q",
   "source_url": "https://omlx.ai/benchmarks/performance/81uh7e4q",
   "tg_tokens": null,
   "tps": 65.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-81uh7e4q-ctx16384",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V13-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: c3np05q3, h1b70red, t42zffhq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1610.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V13-dequantized-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,610 53.6 22.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/81uh7e4q",
   "source_url": "https://omlx.ai/benchmarks/performance/81uh7e4q",
   "tg_tokens": null,
   "tps": 53.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-dzqj6eqs-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V13-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 1pj5yatc, eldf0oy8, n88bbmfw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1452.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V13-dequantized-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,452 58.5 21.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dzqj6eqs",
   "source_url": "https://omlx.ai/benchmarks/performance/dzqj6eqs",
   "tg_tokens": null,
   "tps": 58.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-dzqj6eqs-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V13-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 1pj5yatc, eldf0oy8, n88bbmfw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2031.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V13-dequantized-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,031 103.9 23.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dzqj6eqs",
   "source_url": "https://omlx.ai/benchmarks/performance/dzqj6eqs",
   "tg_tokens": null,
   "tps": 103.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-dzqj6eqs-ctx8192",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V13-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 1pj5yatc, eldf0oy8, n88bbmfw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1921.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V13-dequantized-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,921 58.0 23.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dzqj6eqs",
   "source_url": "https://omlx.ai/benchmarks/performance/dzqj6eqs",
   "tg_tokens": null,
   "tps": 58.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-dzqj6eqs-ctx16384",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V13-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 1pj5yatc, eldf0oy8, n88bbmfw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1702.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V13-dequantized-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,702 101.3 23.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dzqj6eqs",
   "source_url": "https://omlx.ai/benchmarks/performance/dzqj6eqs",
   "tg_tokens": null,
   "tps": 101.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-7ho68b3b-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: ljkttgph; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2328.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 2,328 107.7 19.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ho68b3b",
   "source_url": "https://omlx.ai/benchmarks/performance/7ho68b3b",
   "tg_tokens": null,
   "tps": 107.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-7ho68b3b-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: ljkttgph; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2954.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,954 103.8 20.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ho68b3b",
   "source_url": "https://omlx.ai/benchmarks/performance/7ho68b3b",
   "tg_tokens": null,
   "tps": 103.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-qkpt1qqv-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 51.8 GB at this context; duplicate traces of the same session: fnk7yirk; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 245.4,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.8-27B on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 unknown; 1k 245.4 10.7 51.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qkpt1qqv",
   "source_url": "https://omlx.ai/benchmarks/performance/qkpt1qqv",
   "tg_tokens": null,
   "tps": 10.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-qkpt1qqv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 55.4 GB at this context; duplicate traces of the same session: fnk7yirk; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 258.2,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.8-27B on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 unknown; 4k 258.2 10.6 55.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qkpt1qqv",
   "source_url": "https://omlx.ai/benchmarks/performance/qkpt1qqv",
   "tg_tokens": null,
   "tps": 10.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-efs2qtbp-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 51.8 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 225.2,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.8-27B on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 unknown; 1k 225.2 10.7 51.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/efs2qtbp",
   "source_url": "https://omlx.ai/benchmarks/performance/efs2qtbp",
   "tg_tokens": null,
   "tps": 10.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-efs2qtbp-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 55.4 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 255.4,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.8-27B on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 unknown; 4k 255.4 10.6 55.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/efs2qtbp",
   "source_url": "https://omlx.ai/benchmarks/performance/efs2qtbp",
   "tg_tokens": null,
   "tps": 10.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-d8cxgvcr-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: bd3c7n5h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 927.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 927.8 33.4 18.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d8cxgvcr",
   "source_url": "https://omlx.ai/benchmarks/performance/d8cxgvcr",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-d8cxgvcr-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: bd3c7n5h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 739.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 739.5 32.8 19.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d8cxgvcr",
   "source_url": "https://omlx.ai/benchmarks/performance/d8cxgvcr",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-lo20y0nf-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: qi09f11p; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 216.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 8bit; 1k 216.6 19.5 28.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lo20y0nf",
   "source_url": "https://omlx.ai/benchmarks/performance/lo20y0nf",
   "tg_tokens": null,
   "tps": 19.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-lo20y0nf-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.9 GB at this context; duplicate traces of the same session: qi09f11p; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 219.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 8bit; 4k 219.6 19.1 31.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lo20y0nf",
   "source_url": "https://omlx.ai/benchmarks/performance/lo20y0nf",
   "tg_tokens": null,
   "tps": 19.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-w6iffh48-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: ckrc94et; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 822.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 822.0 18.6 28.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w6iffh48",
   "source_url": "https://omlx.ai/benchmarks/performance/w6iffh48",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-w6iffh48-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: ckrc94et; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 804.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 804.3 18.4 29.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w6iffh48",
   "source_url": "https://omlx.ai/benchmarks/performance/w6iffh48",
   "tg_tokens": null,
   "tps": 18.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ki4lfo4r-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 1sbad37e; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 406.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 406.8 18.1 16.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ki4lfo4r",
   "source_url": "https://omlx.ai/benchmarks/performance/ki4lfo4r",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ki4lfo4r-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: 1sbad37e; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 426.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 426.5 17.7 17.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ki4lfo4r",
   "source_url": "https://omlx.ai/benchmarks/performance/ki4lfo4r",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wjg5w79f-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: fcd7w232, mxr9xhdy, ok566g3m, uxgjekt6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 890.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 890.2 71.9 22.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wjg5w79f",
   "source_url": "https://omlx.ai/benchmarks/performance/wjg5w79f",
   "tg_tokens": null,
   "tps": 71.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wjg5w79f-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: fcd7w232, mxr9xhdy, ok566g3m, uxgjekt6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 978.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 978.4 96.1 24.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wjg5w79f",
   "source_url": "https://omlx.ai/benchmarks/performance/wjg5w79f",
   "tg_tokens": null,
   "tps": 96.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wjg5w79f-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: fcd7w232, mxr9xhdy, ok566g3m, uxgjekt6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 924.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 924.3 50.2 24.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wjg5w79f",
   "source_url": "https://omlx.ai/benchmarks/performance/wjg5w79f",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wjg5w79f-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: fcd7w232, mxr9xhdy, ok566g3m, uxgjekt6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 877.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 877.2 64.1 25.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wjg5w79f",
   "source_url": "https://omlx.ai/benchmarks/performance/wjg5w79f",
   "tg_tokens": null,
   "tps": 64.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wjg5w79f-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: fcd7w232, mxr9xhdy, ok566g3m, uxgjekt6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 803.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 803.5 59.6 25.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wjg5w79f",
   "source_url": "https://omlx.ai/benchmarks/performance/wjg5w79f",
   "tg_tokens": null,
   "tps": 59.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wjg5w79f-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.2 GB at this context; duplicate traces of the same session: fcd7w232, mxr9xhdy, ok566g3m, uxgjekt6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 720.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 720.4 53.3 28.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wjg5w79f",
   "source_url": "https://omlx.ai/benchmarks/performance/wjg5w79f",
   "tg_tokens": null,
   "tps": 53.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wjg5w79f-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: fcd7w232, mxr9xhdy, ok566g3m, uxgjekt6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 545.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 545.1 47.4 32.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wjg5w79f",
   "source_url": "https://omlx.ai/benchmarks/performance/wjg5w79f",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-clixqex4-ctx1024",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: 2ak0n09b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 355.6,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 1k 355.6 37.5 11.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/clixqex4",
   "source_url": "https://omlx.ai/benchmarks/performance/clixqex4",
   "tg_tokens": null,
   "tps": 37.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-clixqex4-ctx4096",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: 2ak0n09b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 445.8,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 4k 445.8 35.9 14.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/clixqex4",
   "source_url": "https://omlx.ai/benchmarks/performance/clixqex4",
   "tg_tokens": null,
   "tps": 35.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-j5191bus-ctx1024",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: lbllhkas; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 558.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 558.3 19.0 7.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j5191bus",
   "source_url": "https://omlx.ai/benchmarks/performance/j5191bus",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-j5191bus-ctx4096",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 7.5 GB at this context; duplicate traces of the same session: lbllhkas; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 609.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 609.7 18.8 7.5 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j5191bus",
   "source_url": "https://omlx.ai/benchmarks/performance/j5191bus",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-abc9vh7c-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 35.4 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 966.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26b-a4b-it-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 1k 966.3 12.1 35.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/abc9vh7c",
   "source_url": "https://omlx.ai/benchmarks/performance/abc9vh7c",
   "tg_tokens": null,
   "tps": 12.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-abc9vh7c-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 35.9 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1097.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26b-a4b-it-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 4k 1,097 18.0 35.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/abc9vh7c",
   "source_url": "https://omlx.ai/benchmarks/performance/abc9vh7c",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-gn4br2oe-ctx1024",
   "model": "Gemma-4-31B",
   "notes": "oMLX community benchmark trace; peak memory 18.0 GB at this context; duplicate traces of the same session: uhfle3pz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 705.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31b-it-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 705.8 24.8 18.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gn4br2oe",
   "source_url": "https://omlx.ai/benchmarks/performance/gn4br2oe",
   "tg_tokens": null,
   "tps": 24.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-gn4br2oe-ctx4096",
   "model": "Gemma-4-31B",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: uhfle3pz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 644.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31b-it-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 644.0 25.0 21.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gn4br2oe",
   "source_url": "https://omlx.ai/benchmarks/performance/gn4br2oe",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-ljtg2cae-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: l0x6wvjg; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 629.7,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "mlx-community--Ornith-1.5-35B-A3B-OptiQ-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 629.7 14.0 21.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ljtg2cae",
   "source_url": "https://omlx.ai/benchmarks/performance/ljtg2cae",
   "tg_tokens": null,
   "tps": 14.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-ljtg2cae-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: l0x6wvjg; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 74.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "mlx-community--Ornith-1.5-35B-A3B-OptiQ-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 74.0 5.6 23.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ljtg2cae",
   "source_url": "https://omlx.ai/benchmarks/performance/ljtg2cae",
   "tg_tokens": null,
   "tps": 5.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-1trryssd-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 140yqjwj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1714.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "mlx-community--Ornith-1.5-35B-A3B-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,714 110.5 21.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1trryssd",
   "source_url": "https://omlx.ai/benchmarks/performance/1trryssd",
   "tg_tokens": null,
   "tps": 110.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-1trryssd-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 140yqjwj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1938.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "mlx-community--Ornith-1.5-35B-A3B-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,938 106.8 23.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1trryssd",
   "source_url": "https://omlx.ai/benchmarks/performance/1trryssd",
   "tg_tokens": null,
   "tps": 106.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2baluegj-ctx65536",
   "model": "pplx-computer-qwen-3-6-35b-a3b",
   "notes": "MLX build 20260709; oMLX community benchmark trace, single run at 64k context; peak memory 21.2 GB; TTFT 23974 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2734.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "perplexity-ai--pplx-computer-qwen-3-6-35b-a3b-mlx-20260709 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k tokens 2,734 PP tok/s 86.3 TG tok/s 23974 TTFT (ms) 21.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2baluegj",
   "source_url": "https://omlx.ai/benchmarks/performance/2baluegj",
   "tg_tokens": null,
   "tps": 86.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3ec3f82a-ctx4096",
   "model": "DeepSeek-Coder-V2-Lite-Instruct",
   "notes": "MLX; oMLX community benchmark trace, single run at 4k context; peak memory 10.2 GB; TTFT 1360 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3011.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-Coder-V2-Lite-Instruct-4bit-mlx on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 3,011 PP tok/s 105.0 TG tok/s 1360 TTFT (ms) 10.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ec3f82a",
   "source_url": "https://omlx.ai/benchmarks/performance/3ec3f82a",
   "tg_tokens": null,
   "tps": 105.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-hj0k69im-ctx1024",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 155.8 GB at this context; duplicate traces of the same session: 52inipck, 8i67yzl1, tcwin3e2, vwy1it1f; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 418.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 418.8 41.6 155.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hj0k69im",
   "source_url": "https://omlx.ai/benchmarks/performance/hj0k69im",
   "tg_tokens": null,
   "tps": 41.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-hj0k69im-ctx4096",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.3 GB at this context; duplicate traces of the same session: 52inipck, 8i67yzl1, tcwin3e2, vwy1it1f; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 554.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 554.2 42.0 156.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hj0k69im",
   "source_url": "https://omlx.ai/benchmarks/performance/hj0k69im",
   "tg_tokens": null,
   "tps": 42.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hj0k69im-ctx8192",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.3 GB at this context; duplicate traces of the same session: 52inipck, 8i67yzl1, tcwin3e2, vwy1it1f; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 602.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 602.9 35.5 156.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hj0k69im",
   "source_url": "https://omlx.ai/benchmarks/performance/hj0k69im",
   "tg_tokens": null,
   "tps": 35.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hj0k69im-ctx16384",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.4 GB at this context; duplicate traces of the same session: 52inipck, 8i67yzl1, tcwin3e2, vwy1it1f; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 591.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 591.6 42.3 156.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hj0k69im",
   "source_url": "https://omlx.ai/benchmarks/performance/hj0k69im",
   "tg_tokens": null,
   "tps": 42.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hj0k69im-ctx32768",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.5 GB at this context; duplicate traces of the same session: 52inipck, 8i67yzl1, tcwin3e2, vwy1it1f; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 573.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 573.9 46.1 156.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hj0k69im",
   "source_url": "https://omlx.ai/benchmarks/performance/hj0k69im",
   "tg_tokens": null,
   "tps": 46.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-hj0k69im-ctx65536",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.8 GB at this context; duplicate traces of the same session: 52inipck, 8i67yzl1, tcwin3e2, vwy1it1f; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 536.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 536.8 40.8 156.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hj0k69im",
   "source_url": "https://omlx.ai/benchmarks/performance/hj0k69im",
   "tg_tokens": null,
   "tps": 40.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-hj0k69im-ctx131072",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 157.1 GB at this context; duplicate traces of the same session: 52inipck, 8i67yzl1, tcwin3e2, vwy1it1f; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 463.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 463.8 38.2 157.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hj0k69im",
   "source_url": "https://omlx.ai/benchmarks/performance/hj0k69im",
   "tg_tokens": null,
   "tps": 38.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-lic467vq-ctx1024",
   "model": "Devstral-4bit-mlx",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 13.0 GB; TTFT 3396 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 301.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Devstral-4bit-mlx on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 1k tokens 301.6 PP tok/s 10.9 TG tok/s 3396 TTFT (ms) 13.0 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lic467vq",
   "source_url": "https://omlx.ai/benchmarks/performance/lic467vq",
   "tg_tokens": null,
   "tps": 10.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5wd11k3s-ctx1024",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "JK build; oMLX community benchmark trace; peak memory 162.2 GB at this context; duplicate traces of the same session: 3nk201hw, e7nj1ayy, k0kjicp1, rhp2m0mr, uh1bhwbz, wdpt8g1g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 483.5,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "GLM-5.3-Flash-UNCENSORED-JK-oQ3.5e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 3bit; 1k 483.5 53.9 162.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5wd11k3s",
   "source_url": "https://omlx.ai/benchmarks/performance/5wd11k3s",
   "tg_tokens": null,
   "tps": 53.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5wd11k3s-ctx4096",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "JK build; oMLX community benchmark trace; peak memory 163.3 GB at this context; duplicate traces of the same session: 3nk201hw, e7nj1ayy, k0kjicp1, rhp2m0mr, uh1bhwbz, wdpt8g1g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 574.3,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "GLM-5.3-Flash-UNCENSORED-JK-oQ3.5e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 3bit; 4k 574.3 48.3 163.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5wd11k3s",
   "source_url": "https://omlx.ai/benchmarks/performance/5wd11k3s",
   "tg_tokens": null,
   "tps": 48.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5wd11k3s-ctx8192",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "JK build; oMLX community benchmark trace; peak memory 163.7 GB at this context; duplicate traces of the same session: 3nk201hw, e7nj1ayy, k0kjicp1, rhp2m0mr, uh1bhwbz, wdpt8g1g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 540.1,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "GLM-5.3-Flash-UNCENSORED-JK-oQ3.5e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 3bit; 8k 540.1 59.5 163.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5wd11k3s",
   "source_url": "https://omlx.ai/benchmarks/performance/5wd11k3s",
   "tg_tokens": null,
   "tps": 59.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5wd11k3s-ctx16384",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "JK build; oMLX community benchmark trace; peak memory 163.8 GB at this context; duplicate traces of the same session: 3nk201hw, e7nj1ayy, k0kjicp1, rhp2m0mr, uh1bhwbz, wdpt8g1g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 529.4,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "GLM-5.3-Flash-UNCENSORED-JK-oQ3.5e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 3bit; 16k 529.4 99.0 163.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5wd11k3s",
   "source_url": "https://omlx.ai/benchmarks/performance/5wd11k3s",
   "tg_tokens": null,
   "tps": 99.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5wd11k3s-ctx32768",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "JK build; oMLX community benchmark trace; peak memory 164.1 GB at this context; duplicate traces of the same session: 3nk201hw, e7nj1ayy, k0kjicp1, rhp2m0mr, uh1bhwbz, wdpt8g1g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 523.2,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "GLM-5.3-Flash-UNCENSORED-JK-oQ3.5e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 3bit; 32k 523.2 68.0 164.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5wd11k3s",
   "source_url": "https://omlx.ai/benchmarks/performance/5wd11k3s",
   "tg_tokens": null,
   "tps": 68.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5wd11k3s-ctx65536",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "JK build; oMLX community benchmark trace; peak memory 164.5 GB at this context; duplicate traces of the same session: 3nk201hw, e7nj1ayy, k0kjicp1, rhp2m0mr, uh1bhwbz, wdpt8g1g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 428.6,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "GLM-5.3-Flash-UNCENSORED-JK-oQ3.5e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 3bit; 64k 428.6 59.9 164.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5wd11k3s",
   "source_url": "https://omlx.ai/benchmarks/performance/5wd11k3s",
   "tg_tokens": null,
   "tps": 59.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5wd11k3s-ctx131072",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "JK build; oMLX community benchmark trace; peak memory 166.0 GB at this context; duplicate traces of the same session: 3nk201hw, e7nj1ayy, k0kjicp1, rhp2m0mr, uh1bhwbz, wdpt8g1g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 509.1,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "GLM-5.3-Flash-UNCENSORED-JK-oQ3.5e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 3bit; 128k 509.1 98.0 166.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5wd11k3s",
   "source_url": "https://omlx.ai/benchmarks/performance/5wd11k3s",
   "tg_tokens": null,
   "tps": 98.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gzqlndor-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.8 GB at this context; duplicate traces of the same session: 1h3hipco, 3oznr8k1, 6rx783v5, 82mudzhr, dfb9vhpf, sl8hayor; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 443.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 443.4 37.4 175.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gzqlndor",
   "source_url": "https://omlx.ai/benchmarks/performance/gzqlndor",
   "tg_tokens": null,
   "tps": 37.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gzqlndor-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: 1h3hipco, 3oznr8k1, 6rx783v5, 82mudzhr, dfb9vhpf, sl8hayor; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 462.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 462.6 57.0 177.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gzqlndor",
   "source_url": "https://omlx.ai/benchmarks/performance/gzqlndor",
   "tg_tokens": null,
   "tps": 57.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gzqlndor-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.0 GB at this context; duplicate traces of the same session: 1h3hipco, 3oznr8k1, 6rx783v5, 82mudzhr, dfb9vhpf, sl8hayor; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 525.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 525.6 37.3 178.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gzqlndor",
   "source_url": "https://omlx.ai/benchmarks/performance/gzqlndor",
   "tg_tokens": null,
   "tps": 37.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gzqlndor-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.7 GB at this context; duplicate traces of the same session: 1h3hipco, 3oznr8k1, 6rx783v5, 82mudzhr, dfb9vhpf, sl8hayor; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 516.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 516.3 29.4 178.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gzqlndor",
   "source_url": "https://omlx.ai/benchmarks/performance/gzqlndor",
   "tg_tokens": null,
   "tps": 29.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gzqlndor-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 180.1 GB at this context; duplicate traces of the same session: 1h3hipco, 3oznr8k1, 6rx783v5, 82mudzhr, dfb9vhpf, sl8hayor; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 508.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 508.8 32.1 180.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gzqlndor",
   "source_url": "https://omlx.ai/benchmarks/performance/gzqlndor",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gzqlndor-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 182.8 GB at this context; duplicate traces of the same session: 1h3hipco, 3oznr8k1, 6rx783v5, 82mudzhr, dfb9vhpf, sl8hayor; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 495.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 495.1 43.3 182.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gzqlndor",
   "source_url": "https://omlx.ai/benchmarks/performance/gzqlndor",
   "tg_tokens": null,
   "tps": 43.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gzqlndor-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 188.5 GB at this context; duplicate traces of the same session: 1h3hipco, 3oznr8k1, 6rx783v5, 82mudzhr, dfb9vhpf, sl8hayor; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 473.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 473.8 34.3 188.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gzqlndor",
   "source_url": "https://omlx.ai/benchmarks/performance/gzqlndor",
   "tg_tokens": null,
   "tps": 34.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-7a19498j-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: 06rtt4nf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 958.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 958.8 106.5 35.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7a19498j",
   "source_url": "https://omlx.ai/benchmarks/performance/7a19498j",
   "tg_tokens": null,
   "tps": 106.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-7a19498j-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: 06rtt4nf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2839.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 2,839 104.2 36.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7a19498j",
   "source_url": "https://omlx.ai/benchmarks/performance/7a19498j",
   "tg_tokens": null,
   "tps": 104.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-sq2wwr1r-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: rd0esain; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 560.9,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-OptiQ-4bit-MLX on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 560.9 66.9 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sq2wwr1r",
   "source_url": "https://omlx.ai/benchmarks/performance/sq2wwr1r",
   "tg_tokens": null,
   "tps": 66.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-sq2wwr1r-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: rd0esain; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 589.7,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-OptiQ-4bit-MLX on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 589.7 65.6 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sq2wwr1r",
   "source_url": "https://omlx.ai/benchmarks/performance/sq2wwr1r",
   "tg_tokens": null,
   "tps": 65.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-uivk0t3p-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 9fjjal7k; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2926.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,926 104.2 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uivk0t3p",
   "source_url": "https://omlx.ai/benchmarks/performance/uivk0t3p",
   "tg_tokens": null,
   "tps": 104.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-uivk0t3p-ctx8192",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: 9fjjal7k; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2846.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 2,846 100.5 20.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uivk0t3p",
   "source_url": "https://omlx.ai/benchmarks/performance/uivk0t3p",
   "tg_tokens": null,
   "tps": 100.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-g8t0oogb-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 21.4 GB; TTFT 669 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1531.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 1,531 PP tok/s 144.3 TG tok/s 669 TTFT (ms) 21.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g8t0oogb",
   "source_url": "https://omlx.ai/benchmarks/performance/g8t0oogb",
   "tg_tokens": null,
   "tps": 144.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-prcns8ob-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 21.4 GB; TTFT 669 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1530.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 1,530 PP tok/s 147.2 TG tok/s 669 TTFT (ms) 21.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/prcns8ob",
   "source_url": "https://omlx.ai/benchmarks/performance/prcns8ob",
   "tg_tokens": null,
   "tps": 147.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-ag6xnur5-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 51.8 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 225.2,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.8-27B on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 unknown; 1k 225.2 10.7 51.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ag6xnur5",
   "source_url": "https://omlx.ai/benchmarks/performance/ag6xnur5",
   "tg_tokens": null,
   "tps": 10.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-ag6xnur5-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 55.4 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 255.4,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.8-27B on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 unknown; 4k 255.4 10.6 55.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ag6xnur5",
   "source_url": "https://omlx.ai/benchmarks/performance/ag6xnur5",
   "tg_tokens": null,
   "tps": 10.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-fmhu8iqs-ctx1024",
   "model": "Qwen3.8-27B-heretic-ara",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: 1dacvjwt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 367.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-heretic-ara-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 1k 367.3 14.5 33.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fmhu8iqs",
   "source_url": "https://omlx.ai/benchmarks/performance/fmhu8iqs",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-fmhu8iqs-ctx4096",
   "model": "Qwen3.8-27B-heretic-ara",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.7 GB at this context; duplicate traces of the same session: 1dacvjwt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 424.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-heretic-ara-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 4k 424.5 30.1 37.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fmhu8iqs",
   "source_url": "https://omlx.ai/benchmarks/performance/fmhu8iqs",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-m7lxyxzb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 18.5 GB; TTFT 4683 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 218.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 218.7 PP tok/s 52.6 TG tok/s 4683 TTFT (ms) 18.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m7lxyxzb",
   "source_url": "https://omlx.ai/benchmarks/performance/m7lxyxzb",
   "tg_tokens": null,
   "tps": 52.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-4ysmdxz3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 38.9 GB; TTFT 4860 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 210.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 210.7 PP tok/s 31.5 TG tok/s 4860 TTFT (ms) 38.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ysmdxz3",
   "source_url": "https://omlx.ai/benchmarks/performance/4ysmdxz3",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-f9ai3jsg-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 40.0 GB at this context; duplicate traces of the same session: 0ft6hfof, gblvgn85, hovra6yi; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 302.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 8bit; 1k 302.3 32.0 40.0 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f9ai3jsg",
   "source_url": "https://omlx.ai/benchmarks/performance/f9ai3jsg",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-f9ai3jsg-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 43.5 GB at this context; duplicate traces of the same session: 0ft6hfof, gblvgn85, hovra6yi; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 349.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 8bit; 4k 349.1 30.8 43.5 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f9ai3jsg",
   "source_url": "https://omlx.ai/benchmarks/performance/f9ai3jsg",
   "tg_tokens": null,
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-f9ai3jsg-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 44.3 GB at this context; duplicate traces of the same session: 0ft6hfof, gblvgn85, hovra6yi; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 377.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 8bit; 8k 377.4 37.1 44.3 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f9ai3jsg",
   "source_url": "https://omlx.ai/benchmarks/performance/f9ai3jsg",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-f9ai3jsg-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 45.8 GB at this context; duplicate traces of the same session: 0ft6hfof, gblvgn85, hovra6yi; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 369.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 8bit; 16k 369.2 41.7 45.8 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f9ai3jsg",
   "source_url": "https://omlx.ai/benchmarks/performance/f9ai3jsg",
   "tg_tokens": null,
   "tps": 41.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-l39a73at-ctx1024",
   "model": "Qwen3.8-Flash-Next-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 78.0 GB at this context; duplicate traces of the same session: 53k3v6mv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 893.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-Abliterated-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 893.6 45.7 78.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l39a73at",
   "source_url": "https://omlx.ai/benchmarks/performance/l39a73at",
   "tg_tokens": null,
   "tps": 45.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-l39a73at-ctx4096",
   "model": "Qwen3.8-Flash-Next-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 78.1 GB at this context; duplicate traces of the same session: 53k3v6mv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1703.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-Abliterated-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,703 57.9 78.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l39a73at",
   "source_url": "https://omlx.ai/benchmarks/performance/l39a73at",
   "tg_tokens": null,
   "tps": 57.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-roa237co-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 102.1 GB at this context; duplicate traces of the same session: 0si4nbl7, g32hvps3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 694.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 4bit; 1k 694.1 44.9 102.1 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/roa237co",
   "source_url": "https://omlx.ai/benchmarks/performance/roa237co",
   "tg_tokens": null,
   "tps": 44.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-roa237co-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 104.6 GB at this context; duplicate traces of the same session: 0si4nbl7, g32hvps3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 855.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 4bit; 4k 855.6 41.5 104.6 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/roa237co",
   "source_url": "https://omlx.ai/benchmarks/performance/roa237co",
   "tg_tokens": null,
   "tps": 41.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-roa237co-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 104.7 GB at this context; duplicate traces of the same session: 0si4nbl7, g32hvps3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 854.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 4bit; 8k 854.7 39.5 104.7 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/roa237co",
   "source_url": "https://omlx.ai/benchmarks/performance/roa237co",
   "tg_tokens": null,
   "tps": 39.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-fk8axm5s-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 100.7 GB at this context; duplicate traces of the same session: fcahe3aj, s94771bh; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 720.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 4bit; 1k 720.9 26.1 100.7 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fk8axm5s",
   "source_url": "https://omlx.ai/benchmarks/performance/fk8axm5s",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-fk8axm5s-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 103.2 GB at this context; duplicate traces of the same session: fcahe3aj, s94771bh; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 886.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 4bit; 4k 886.1 25.4 103.2 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fk8axm5s",
   "source_url": "https://omlx.ai/benchmarks/performance/fk8axm5s",
   "tg_tokens": null,
   "tps": 25.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-fk8axm5s-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 103.2 GB at this context; duplicate traces of the same session: fcahe3aj, s94771bh; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 882.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 4bit; 8k 882.0 25.1 103.2 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fk8axm5s",
   "source_url": "https://omlx.ai/benchmarks/performance/fk8axm5s",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-418iyy0f-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 18.5 GB; TTFT 4671 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 219.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 219.2 PP tok/s 53.8 TG tok/s 4671 TTFT (ms) 18.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/418iyy0f",
   "source_url": "https://omlx.ai/benchmarks/performance/418iyy0f",
   "tg_tokens": null,
   "tps": 53.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jetlm0vz-ctx1024",
   "model": "Unlimited-OCR",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 6.4 GB; TTFT 375 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2730.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Unlimited-OCR on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k tokens 2,730 PP tok/s 336.5 TG tok/s 375 TTFT (ms) 6.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jetlm0vz",
   "source_url": "https://omlx.ai/benchmarks/performance/jetlm0vz",
   "tg_tokens": null,
   "tps": 336.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qrsmts91-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: r85wj3fd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3945.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 3,945 122.8 16.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qrsmts91",
   "source_url": "https://omlx.ai/benchmarks/performance/qrsmts91",
   "tg_tokens": null,
   "tps": 122.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qrsmts91-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: r85wj3fd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4652.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 4,652 104.9 17.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qrsmts91",
   "source_url": "https://omlx.ai/benchmarks/performance/qrsmts91",
   "tg_tokens": null,
   "tps": 104.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-vcnmqtxa-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 14.5 GB at this context; duplicate traces of the same session: uq6pw9dc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 995.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 995.5 118.3 14.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vcnmqtxa",
   "source_url": "https://omlx.ai/benchmarks/performance/vcnmqtxa",
   "tg_tokens": null,
   "tps": 118.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vcnmqtxa-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 15.4 GB at this context; duplicate traces of the same session: uq6pw9dc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3573.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 3,573 113.7 15.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vcnmqtxa",
   "source_url": "https://omlx.ai/benchmarks/performance/vcnmqtxa",
   "tg_tokens": null,
   "tps": 113.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ngpnux3v-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 35.4 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 966.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26b-a4b-it-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 1k 966.3 12.1 35.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ngpnux3v",
   "source_url": "https://omlx.ai/benchmarks/performance/ngpnux3v",
   "tg_tokens": null,
   "tps": 12.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ngpnux3v-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 35.9 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1097.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26b-a4b-it-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 4k 1,097 18.0 35.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ngpnux3v",
   "source_url": "https://omlx.ai/benchmarks/performance/ngpnux3v",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-pka5xp3j-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 4.7 GB; TTFT 491 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2087.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-4bit on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 2,087 PP tok/s 92.2 TG tok/s 491 TTFT (ms) 4.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pka5xp3j",
   "source_url": "https://omlx.ai/benchmarks/performance/pka5xp3j",
   "tg_tokens": null,
   "tps": 92.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M1 Max",
   "id": "omlx-d8bd9q4n-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: xbiqm41i; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 11.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 11.2 1.5 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d8bd9q4n",
   "source_url": "https://omlx.ai/benchmarks/performance/d8bd9q4n",
   "tg_tokens": null,
   "tps": 1.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M1 Max",
   "id": "omlx-d8bd9q4n-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: xbiqm41i; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 10.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 10.8 1.5 19.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d8bd9q4n",
   "source_url": "https://omlx.ai/benchmarks/performance/d8bd9q4n",
   "tg_tokens": null,
   "tps": 1.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M1 Max",
   "id": "omlx-nz9q39qt-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: eahdwkvt; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 10.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 10.2 1.4 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nz9q39qt",
   "source_url": "https://omlx.ai/benchmarks/performance/nz9q39qt",
   "tg_tokens": null,
   "tps": 1.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M1 Max",
   "id": "omlx-nz9q39qt-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: eahdwkvt; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 22.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 22.3 1.5 19.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nz9q39qt",
   "source_url": "https://omlx.ai/benchmarks/performance/nz9q39qt",
   "tg_tokens": null,
   "tps": 1.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-kdb90pp8-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: m6tns5zg; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1430.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "qwen35-9b-4bit on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 1k 1,430 39.5 6.7 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kdb90pp8",
   "source_url": "https://omlx.ai/benchmarks/performance/kdb90pp8",
   "tg_tokens": null,
   "tps": 39.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-kdb90pp8-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.1 GB at this context; duplicate traces of the same session: m6tns5zg; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1485.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "qwen35-9b-4bit on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 4k 1,485 40.4 7.1 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kdb90pp8",
   "source_url": "https://omlx.ai/benchmarks/performance/kdb90pp8",
   "tg_tokens": null,
   "tps": 40.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-a3n59ftr-ctx1024",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 155.8 GB at this context; duplicate traces of the same session: be3hxams; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 418.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 418.8 41.6 155.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a3n59ftr",
   "source_url": "https://omlx.ai/benchmarks/performance/a3n59ftr",
   "tg_tokens": null,
   "tps": 41.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-a3n59ftr-ctx4096",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.3 GB at this context; duplicate traces of the same session: be3hxams; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 554.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 554.2 42.0 156.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a3n59ftr",
   "source_url": "https://omlx.ai/benchmarks/performance/a3n59ftr",
   "tg_tokens": null,
   "tps": 42.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-a3n59ftr-ctx8192",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.3 GB at this context; duplicate traces of the same session: be3hxams; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 602.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 602.9 35.5 156.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a3n59ftr",
   "source_url": "https://omlx.ai/benchmarks/performance/a3n59ftr",
   "tg_tokens": null,
   "tps": 35.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-a3n59ftr-ctx16384",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.4 GB at this context; duplicate traces of the same session: be3hxams; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 591.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 591.6 42.3 156.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a3n59ftr",
   "source_url": "https://omlx.ai/benchmarks/performance/a3n59ftr",
   "tg_tokens": null,
   "tps": 42.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-a3n59ftr-ctx32768",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.5 GB at this context; duplicate traces of the same session: be3hxams; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 573.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 573.9 46.1 156.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a3n59ftr",
   "source_url": "https://omlx.ai/benchmarks/performance/a3n59ftr",
   "tg_tokens": null,
   "tps": 46.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-a3n59ftr-ctx65536",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.8 GB at this context; duplicate traces of the same session: be3hxams; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 536.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 536.8 40.8 156.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a3n59ftr",
   "source_url": "https://omlx.ai/benchmarks/performance/a3n59ftr",
   "tg_tokens": null,
   "tps": 40.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-a3n59ftr-ctx131072",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 157.1 GB at this context; duplicate traces of the same session: be3hxams; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 463.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 463.8 38.2 157.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a3n59ftr",
   "source_url": "https://omlx.ai/benchmarks/performance/a3n59ftr",
   "tg_tokens": null,
   "tps": 38.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vgrasnho-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 248.9 GB at this context; duplicate traces of the same session: 1349bc0c, aggnc1fr, bmcl63vt, gwjorhrr, klf0wxz7, ox89ud14, t8h9sapm; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 473.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 1k 473.6 42.2 248.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vgrasnho",
   "source_url": "https://omlx.ai/benchmarks/performance/vgrasnho",
   "tg_tokens": null,
   "tps": 42.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vgrasnho-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 250.2 GB at this context; duplicate traces of the same session: 1349bc0c, aggnc1fr, bmcl63vt, gwjorhrr, klf0wxz7, ox89ud14, t8h9sapm; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 558.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 4k 558.3 37.4 250.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vgrasnho",
   "source_url": "https://omlx.ai/benchmarks/performance/vgrasnho",
   "tg_tokens": null,
   "tps": 37.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vgrasnho-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 250.6 GB at this context; duplicate traces of the same session: 1349bc0c, aggnc1fr, bmcl63vt, gwjorhrr, klf0wxz7, ox89ud14, t8h9sapm; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 521.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 8k 521.8 41.1 250.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vgrasnho",
   "source_url": "https://omlx.ai/benchmarks/performance/vgrasnho",
   "tg_tokens": null,
   "tps": 41.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vgrasnho-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 250.7 GB at this context; duplicate traces of the same session: 1349bc0c, aggnc1fr, bmcl63vt, gwjorhrr, klf0wxz7, ox89ud14, t8h9sapm; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 512.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 16k 512.4 75.3 250.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vgrasnho",
   "source_url": "https://omlx.ai/benchmarks/performance/vgrasnho",
   "tg_tokens": null,
   "tps": 75.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vgrasnho-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 250.9 GB at this context; duplicate traces of the same session: 1349bc0c, aggnc1fr, bmcl63vt, gwjorhrr, klf0wxz7, ox89ud14, t8h9sapm; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 493.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 32k 493.0 53.9 250.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vgrasnho",
   "source_url": "https://omlx.ai/benchmarks/performance/vgrasnho",
   "tg_tokens": null,
   "tps": 53.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-vgrasnho-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 251.5 GB at this context; duplicate traces of the same session: 1349bc0c, aggnc1fr, bmcl63vt, gwjorhrr, klf0wxz7, ox89ud14, t8h9sapm; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 494.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 64k 494.3 41.4 251.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vgrasnho",
   "source_url": "https://omlx.ai/benchmarks/performance/vgrasnho",
   "tg_tokens": null,
   "tps": 41.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vgrasnho-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 252.9 GB at this context; duplicate traces of the same session: 1349bc0c, aggnc1fr, bmcl63vt, gwjorhrr, klf0wxz7, ox89ud14, t8h9sapm; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 490.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 128k 490.4 72.9 252.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vgrasnho",
   "source_url": "https://omlx.ai/benchmarks/performance/vgrasnho",
   "tg_tokens": null,
   "tps": 72.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-vgrasnho-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 254.3 GB at this context; duplicate traces of the same session: 1349bc0c, aggnc1fr, bmcl63vt, gwjorhrr, klf0wxz7, ox89ud14, t8h9sapm; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 483.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 195k 483.5 37.3 254.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vgrasnho",
   "source_url": "https://omlx.ai/benchmarks/performance/vgrasnho",
   "tg_tokens": null,
   "tps": 37.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-s35ijojf-ctx4096",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "MLX; oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: f10hi0j9; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1045.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 1,045 47.2 17.0 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s35ijojf",
   "source_url": "https://omlx.ai/benchmarks/performance/s35ijojf",
   "tg_tokens": null,
   "tps": 47.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-s35ijojf-ctx16384",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "MLX; oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: f10hi0j9; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 858.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 16k 858.4 32.0 18.1 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s35ijojf",
   "source_url": "https://omlx.ai/benchmarks/performance/s35ijojf",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-5y0fnybt-ctx1024",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "MLX; oMLX community benchmark trace; peak memory 16.8 GB at this context; duplicate traces of the same session: 201u0yef, jhxal77v, r0w875xj; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1077.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 1,077 53.4 16.8 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5y0fnybt",
   "source_url": "https://omlx.ai/benchmarks/performance/5y0fnybt",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-5y0fnybt-ctx4096",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "MLX; oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 201u0yef, jhxal77v, r0w875xj; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1377.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 1,377 47.7 17.0 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5y0fnybt",
   "source_url": "https://omlx.ai/benchmarks/performance/5y0fnybt",
   "tg_tokens": null,
   "tps": 47.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-5y0fnybt-ctx8192",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "MLX; oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 201u0yef, jhxal77v, r0w875xj; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1150.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 8k 1,150 41.9 17.3 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5y0fnybt",
   "source_url": "https://omlx.ai/benchmarks/performance/5y0fnybt",
   "tg_tokens": null,
   "tps": 41.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-5y0fnybt-ctx16384",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "MLX; oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: 201u0yef, jhxal77v, r0w875xj; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 830.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 16k 830.3 32.2 18.1 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5y0fnybt",
   "source_url": "https://omlx.ai/benchmarks/performance/5y0fnybt",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-615hqq8d-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: qq6j2tro; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-6bit-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 127.4 11.0 22.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/615hqq8d",
   "source_url": "https://omlx.ai/benchmarks/performance/615hqq8d",
   "tg_tokens": null,
   "tps": 11.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-615hqq8d-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 25.6 GB at this context; duplicate traces of the same session: qq6j2tro; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 135.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-6bit-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 135.6 10.9 25.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/615hqq8d",
   "source_url": "https://omlx.ai/benchmarks/performance/615hqq8d",
   "tg_tokens": null,
   "tps": 10.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-nb30are7-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 44.4 GB at this context; duplicate traces of the same session: 30reo320; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 155.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 155.2 25.6 44.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nb30are7",
   "source_url": "https://omlx.ai/benchmarks/performance/nb30are7",
   "tg_tokens": null,
   "tps": 25.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-nb30are7-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 49.6 GB at this context; duplicate traces of the same session: 30reo320; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 117.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 195k 117.7 15.6 49.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nb30are7",
   "source_url": "https://omlx.ai/benchmarks/performance/nb30are7",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-6jnr1fud-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 29.1 GB at this context; duplicate traces of the same session: xrktottw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 241.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 241.5 15.2 29.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6jnr1fud",
   "source_url": "https://omlx.ai/benchmarks/performance/6jnr1fud",
   "tg_tokens": null,
   "tps": 15.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-6jnr1fud-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: xrktottw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 238.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 16k 238.2 14.8 32.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6jnr1fud",
   "source_url": "https://omlx.ai/benchmarks/performance/6jnr1fud",
   "tg_tokens": null,
   "tps": 14.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-g5eutps3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 3mzcy3f5, byemxcoe; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 501.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 501.4 28.6 17.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g5eutps3",
   "source_url": "https://omlx.ai/benchmarks/performance/g5eutps3",
   "tg_tokens": null,
   "tps": 28.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-g5eutps3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 3mzcy3f5, byemxcoe; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 572.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 572.9 27.7 19.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g5eutps3",
   "source_url": "https://omlx.ai/benchmarks/performance/g5eutps3",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-g5eutps3-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: 3mzcy3f5, byemxcoe; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 315.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 128k 315.4 9.4 27.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g5eutps3",
   "source_url": "https://omlx.ai/benchmarks/performance/g5eutps3",
   "tg_tokens": null,
   "tps": 9.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-6kpcoec9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 5rulxf0u, te9o349b; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 486.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 486.8 31.3 17.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6kpcoec9",
   "source_url": "https://omlx.ai/benchmarks/performance/6kpcoec9",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-6kpcoec9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 5rulxf0u, te9o349b; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 558.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 558.6 35.5 19.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6kpcoec9",
   "source_url": "https://omlx.ai/benchmarks/performance/6kpcoec9",
   "tg_tokens": null,
   "tps": 35.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-6kpcoec9-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.5 GB at this context; duplicate traces of the same session: 5rulxf0u, te9o349b; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 242.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 195k 242.7 15.6 32.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6kpcoec9",
   "source_url": "https://omlx.ai/benchmarks/performance/6kpcoec9",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-key7zybc-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: lnanj2bf; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 1k 102.5 19.7 33.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/key7zybc",
   "source_url": "https://omlx.ai/benchmarks/performance/key7zybc",
   "tg_tokens": null,
   "tps": 19.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-key7zybc-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.5 GB at this context; duplicate traces of the same session: lnanj2bf; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 4k 103.9 19.1 37.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/key7zybc",
   "source_url": "https://omlx.ai/benchmarks/performance/key7zybc",
   "tg_tokens": null,
   "tps": 19.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8y81ecyt-ctx1024",
   "model": "Qwen3.8-Flash-Next-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 78.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1100.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-Abliterated-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,100 68.9 78.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8y81ecyt",
   "source_url": "https://omlx.ai/benchmarks/performance/8y81ecyt",
   "tg_tokens": null,
   "tps": 68.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8y81ecyt-ctx4096",
   "model": "Qwen3.8-Flash-Next-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 78.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1785.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-Abliterated-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,785 65.5 78.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8y81ecyt",
   "source_url": "https://omlx.ai/benchmarks/performance/8y81ecyt",
   "tg_tokens": null,
   "tps": 65.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-65bj95hl-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 102.1 GB at this context; duplicate traces of the same session: 7azob84v, gvs2wid8, susjrwve; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 704.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 4bit; 1k 704.4 46.7 102.1 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/65bj95hl",
   "source_url": "https://omlx.ai/benchmarks/performance/65bj95hl",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-65bj95hl-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 104.6 GB at this context; duplicate traces of the same session: 7azob84v, gvs2wid8, susjrwve; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 862.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 4bit; 4k 862.5 47.5 104.6 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/65bj95hl",
   "source_url": "https://omlx.ai/benchmarks/performance/65bj95hl",
   "tg_tokens": null,
   "tps": 47.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-65bj95hl-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 104.7 GB at this context; duplicate traces of the same session: 7azob84v, gvs2wid8, susjrwve; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 858.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 4bit; 8k 858.0 55.6 104.7 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/65bj95hl",
   "source_url": "https://omlx.ai/benchmarks/performance/65bj95hl",
   "tg_tokens": null,
   "tps": 55.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-65bj95hl-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 105.0 GB at this context; duplicate traces of the same session: 7azob84v, gvs2wid8, susjrwve; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 851.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 4bit; 16k 851.0 45.3 105.0 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/65bj95hl",
   "source_url": "https://omlx.ai/benchmarks/performance/65bj95hl",
   "tg_tokens": null,
   "tps": 45.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-5ti5kled-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 102.3 GB at this context; duplicate traces of the same session: 8retayrs, f3pwqit1, umn9ovbv; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 708.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 4bit; 1k 708.1 45.7 102.3 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ti5kled",
   "source_url": "https://omlx.ai/benchmarks/performance/5ti5kled",
   "tg_tokens": null,
   "tps": 45.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-5ti5kled-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 104.6 GB at this context; duplicate traces of the same session: 8retayrs, f3pwqit1, umn9ovbv; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 863.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 4bit; 4k 863.4 41.2 104.6 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ti5kled",
   "source_url": "https://omlx.ai/benchmarks/performance/5ti5kled",
   "tg_tokens": null,
   "tps": 41.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-5ti5kled-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 104.7 GB at this context; duplicate traces of the same session: 8retayrs, f3pwqit1, umn9ovbv; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 859.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 4bit; 8k 859.9 30.7 104.7 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ti5kled",
   "source_url": "https://omlx.ai/benchmarks/performance/5ti5kled",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-5ti5kled-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 105.0 GB at this context; duplicate traces of the same session: 8retayrs, f3pwqit1, umn9ovbv; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 853.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 4bit; 16k 853.5 35.3 105.0 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ti5kled",
   "source_url": "https://omlx.ai/benchmarks/performance/5ti5kled",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-k6d30mqf-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 102.2 GB at this context; duplicate traces of the same session: 9x2jydo5, eqldcjlg, q1p2tnyr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 697.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 4bit; 1k 697.7 46.4 102.2 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k6d30mqf",
   "source_url": "https://omlx.ai/benchmarks/performance/k6d30mqf",
   "tg_tokens": null,
   "tps": 46.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-k6d30mqf-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 104.6 GB at this context; duplicate traces of the same session: 9x2jydo5, eqldcjlg, q1p2tnyr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 860.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 4bit; 4k 860.1 44.0 104.6 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k6d30mqf",
   "source_url": "https://omlx.ai/benchmarks/performance/k6d30mqf",
   "tg_tokens": null,
   "tps": 44.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-k6d30mqf-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 104.7 GB at this context; duplicate traces of the same session: 9x2jydo5, eqldcjlg, q1p2tnyr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 856.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 4bit; 8k 856.1 43.3 104.7 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k6d30mqf",
   "source_url": "https://omlx.ai/benchmarks/performance/k6d30mqf",
   "tg_tokens": null,
   "tps": 43.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-k6d30mqf-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 105.0 GB at this context; duplicate traces of the same session: 9x2jydo5, eqldcjlg, q1p2tnyr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 849.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 4bit; 16k 849.2 49.1 105.0 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k6d30mqf",
   "source_url": "https://omlx.ai/benchmarks/performance/k6d30mqf",
   "tg_tokens": null,
   "tps": 49.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-xaveaf39-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6 mixed-quant MTP v2 build; oMLX community benchmark trace; peak memory 64.4 GB at this context; duplicate traces of the same session: drcep822; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 96.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp-v2 on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 96.0 26.2 64.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xaveaf39",
   "source_url": "https://omlx.ai/benchmarks/performance/xaveaf39",
   "tg_tokens": null,
   "tps": 26.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-xaveaf39-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6 mixed-quant MTP v2 build; oMLX community benchmark trace; peak memory 64.9 GB at this context; duplicate traces of the same session: drcep822; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 284.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp-v2 on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 284.2 34.8 64.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xaveaf39",
   "source_url": "https://omlx.ai/benchmarks/performance/xaveaf39",
   "tg_tokens": null,
   "tps": 34.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jxtd1aiy-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.6 GB; TTFT 1031 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 993.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 993.3 PP tok/s 72.3 TG tok/s 1031 TTFT (ms) 70.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jxtd1aiy",
   "source_url": "https://omlx.ai/benchmarks/performance/jxtd1aiy",
   "tg_tokens": null,
   "tps": 72.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1n8uds3v-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: ayujzdgo, mm7nd2cm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 766.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 766.3 41.2 71.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1n8uds3v",
   "source_url": "https://omlx.ai/benchmarks/performance/1n8uds3v",
   "tg_tokens": null,
   "tps": 41.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1n8uds3v-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.3 GB at this context; duplicate traces of the same session: ayujzdgo, mm7nd2cm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1285.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,285 38.5 72.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1n8uds3v",
   "source_url": "https://omlx.ai/benchmarks/performance/1n8uds3v",
   "tg_tokens": null,
   "tps": 38.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1n8uds3v-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.3 GB at this context; duplicate traces of the same session: ayujzdgo, mm7nd2cm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1456.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,456 42.6 72.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1n8uds3v",
   "source_url": "https://omlx.ai/benchmarks/performance/1n8uds3v",
   "tg_tokens": null,
   "tps": 42.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-fx6br0pb-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.2 GB at this context; duplicate traces of the same session: besztcxd, blkgi4ez, c931wrxu; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1005.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 1,005 84.8 183.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fx6br0pb",
   "source_url": "https://omlx.ai/benchmarks/performance/fx6br0pb",
   "tg_tokens": null,
   "tps": 84.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-fx6br0pb-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.9 GB at this context; duplicate traces of the same session: besztcxd, blkgi4ez, c931wrxu; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1323.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 1,323 90.4 183.9 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fx6br0pb",
   "source_url": "https://omlx.ai/benchmarks/performance/fx6br0pb",
   "tg_tokens": null,
   "tps": 90.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-fx6br0pb-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 184.2 GB at this context; duplicate traces of the same session: besztcxd, blkgi4ez, c931wrxu; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1318.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k 1,318 65.9 184.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fx6br0pb",
   "source_url": "https://omlx.ai/benchmarks/performance/fx6br0pb",
   "tg_tokens": null,
   "tps": 65.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-fx6br0pb-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 184.8 GB at this context; duplicate traces of the same session: besztcxd, blkgi4ez, c931wrxu; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1308.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 32k 1,308 72.1 184.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fx6br0pb",
   "source_url": "https://omlx.ai/benchmarks/performance/fx6br0pb",
   "tg_tokens": null,
   "tps": 72.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-32pwh8xz-ctx1024",
   "model": "gemma-4-E2B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.1 GB at this context; duplicate traces of the same session: 50o6y553, 6g23zb72, amtanwny, ufbptlji; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4022.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E2B-it-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k 4,022 185.7 4.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/32pwh8xz",
   "source_url": "https://omlx.ai/benchmarks/performance/32pwh8xz",
   "tg_tokens": null,
   "tps": 185.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-32pwh8xz-ctx4096",
   "model": "gemma-4-E2B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: 50o6y553, 6g23zb72, amtanwny, ufbptlji; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5897.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E2B-it-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 5,897 156.2 4.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/32pwh8xz",
   "source_url": "https://omlx.ai/benchmarks/performance/32pwh8xz",
   "tg_tokens": null,
   "tps": 156.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-32pwh8xz-ctx8192",
   "model": "gemma-4-E2B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: 50o6y553, 6g23zb72, amtanwny, ufbptlji; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5519.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E2B-it-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 8k 5,519 149.3 4.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/32pwh8xz",
   "source_url": "https://omlx.ai/benchmarks/performance/32pwh8xz",
   "tg_tokens": null,
   "tps": 149.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-32pwh8xz-ctx16384",
   "model": "gemma-4-E2B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.3 GB at this context; duplicate traces of the same session: 50o6y553, 6g23zb72, amtanwny, ufbptlji; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4915.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E2B-it-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 16k 4,915 108.1 4.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/32pwh8xz",
   "source_url": "https://omlx.ai/benchmarks/performance/32pwh8xz",
   "tg_tokens": null,
   "tps": 108.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-32pwh8xz-ctx32768",
   "model": "gemma-4-E2B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.7 GB at this context; duplicate traces of the same session: 50o6y553, 6g23zb72, amtanwny, ufbptlji; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4619.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E2B-it-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 32k 4,619 53.2 4.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/32pwh8xz",
   "source_url": "https://omlx.ai/benchmarks/performance/32pwh8xz",
   "tg_tokens": null,
   "tps": 53.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-n7oquyv8-ctx1024",
   "model": "gemma-4-E2B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.1 GB at this context; duplicate traces of the same session: d80yfbyx; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2642.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E2B-it-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k 2,642 255.2 4.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n7oquyv8",
   "source_url": "https://omlx.ai/benchmarks/performance/n7oquyv8",
   "tg_tokens": null,
   "tps": 255.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-n7oquyv8-ctx4096",
   "model": "gemma-4-E2B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: d80yfbyx; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4034.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E2B-it-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 4,034 145.8 4.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n7oquyv8",
   "source_url": "https://omlx.ai/benchmarks/performance/n7oquyv8",
   "tg_tokens": null,
   "tps": 145.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nd61g6dx-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 86.0 GB at this context; duplicate traces of the same session: ajwguqvl, c73lkb6k; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 88.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 88.0 12.5 86.0 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nd61g6dx",
   "source_url": "https://omlx.ai/benchmarks/performance/nd61g6dx",
   "tg_tokens": null,
   "tps": 12.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nd61g6dx-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 86.3 GB at this context; duplicate traces of the same session: ajwguqvl, c73lkb6k; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 156.1,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 156.1 12.7 86.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nd61g6dx",
   "source_url": "https://omlx.ai/benchmarks/performance/nd61g6dx",
   "tg_tokens": null,
   "tps": 12.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-nd61g6dx-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 86.3 GB at this context; duplicate traces of the same session: ajwguqvl, c73lkb6k; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 158.6,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 8k 158.6 13.6 86.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nd61g6dx",
   "source_url": "https://omlx.ai/benchmarks/performance/nd61g6dx",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-8xa4rbry-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 150.8 GB at this context; duplicate traces of the same session: 1eaczdcq, 9r2gojql; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 175.2,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "GLM-5.3-Flash-oQ3.5e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 3bit; 1k 175.2 27.3 150.8 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8xa4rbry",
   "source_url": "https://omlx.ai/benchmarks/performance/8xa4rbry",
   "tg_tokens": null,
   "tps": 27.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-8xa4rbry-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 151.8 GB at this context; duplicate traces of the same session: 1eaczdcq, 9r2gojql; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 238.8,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "GLM-5.3-Flash-oQ3.5e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 3bit; 4k 238.8 24.3 151.8 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8xa4rbry",
   "source_url": "https://omlx.ai/benchmarks/performance/8xa4rbry",
   "tg_tokens": null,
   "tps": 24.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-8xa4rbry-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 152.1 GB at this context; duplicate traces of the same session: 1eaczdcq, 9r2gojql; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.7,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "GLM-5.3-Flash-oQ3.5e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 3bit; 8k 204.7 28.6 152.1 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8xa4rbry",
   "source_url": "https://omlx.ai/benchmarks/performance/8xa4rbry",
   "tg_tokens": null,
   "tps": 28.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-jgj05csl-ctx1024",
   "model": "Jev-Style-Qwen3.5-2B-Decision-v2",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: 0f1y1fti; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 601.8,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Jev-Style-Qwen3.5-2B-Decision-v2-MLX-bf16 on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 bf16; 1k 601.8 4.9 6.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jgj05csl",
   "source_url": "https://omlx.ai/benchmarks/performance/jgj05csl",
   "tg_tokens": null,
   "tps": 4.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-jgj05csl-ctx4096",
   "model": "Jev-Style-Qwen3.5-2B-Decision-v2",
   "notes": "oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: 0f1y1fti; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 715.6,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Jev-Style-Qwen3.5-2B-Decision-v2-MLX-bf16 on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 bf16; 4k 715.6 4.8 6.3 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jgj05csl",
   "source_url": "https://omlx.ai/benchmarks/performance/jgj05csl",
   "tg_tokens": null,
   "tps": 4.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-dyqjvrsa-ctx4096",
   "model": "Qwen3 8B",
   "notes": "oMLX community benchmark trace; peak memory 5.5 GB at this context; duplicate traces of the same session: qun4eqbt; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 355.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-8B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 355.4 45.7 5.5 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dyqjvrsa",
   "source_url": "https://omlx.ai/benchmarks/performance/dyqjvrsa",
   "tg_tokens": null,
   "tps": 45.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-dyqjvrsa-ctx16384",
   "model": "Qwen3 8B",
   "notes": "oMLX community benchmark trace; peak memory 7.1 GB at this context; duplicate traces of the same session: qun4eqbt; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 291.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-8B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 16k 291.2 28.5 7.1 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dyqjvrsa",
   "source_url": "https://omlx.ai/benchmarks/performance/dyqjvrsa",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-loqbv6lc-ctx1024",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: oscjdft9; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 131.2 15.5 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/loqbv6lc",
   "source_url": "https://omlx.ai/benchmarks/performance/loqbv6lc",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-loqbv6lc-ctx4096",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: oscjdft9; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 137.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 137.3 15.1 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/loqbv6lc",
   "source_url": "https://omlx.ai/benchmarks/performance/loqbv6lc",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-dpg8nst7-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: g4y6w7yp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 922.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 922.6 93.3 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dpg8nst7",
   "source_url": "https://omlx.ai/benchmarks/performance/dpg8nst7",
   "tg_tokens": null,
   "tps": 93.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-dpg8nst7-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: g4y6w7yp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1058.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,058 90.3 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dpg8nst7",
   "source_url": "https://omlx.ai/benchmarks/performance/dpg8nst7",
   "tg_tokens": null,
   "tps": 90.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4euj6724-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 218 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4699.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 4,699 PP tok/s 148.7 TG tok/s 218 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4euj6724",
   "source_url": "https://omlx.ai/benchmarks/performance/4euj6724",
   "tg_tokens": null,
   "tps": 148.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-zhu4xjpg-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 0qx7zmwn; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 791.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 791.0 92.8 21.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zhu4xjpg",
   "source_url": "https://omlx.ai/benchmarks/performance/zhu4xjpg",
   "tg_tokens": null,
   "tps": 92.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-zhu4xjpg-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: 0qx7zmwn; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 869.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 869.7 93.1 23.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zhu4xjpg",
   "source_url": "https://omlx.ai/benchmarks/performance/zhu4xjpg",
   "tg_tokens": null,
   "tps": 93.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-udjpi6f8-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: mipxvig8; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 790.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 790.3 107.2 21.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/udjpi6f8",
   "source_url": "https://omlx.ai/benchmarks/performance/udjpi6f8",
   "tg_tokens": null,
   "tps": 107.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-udjpi6f8-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: mipxvig8; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 900.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 900.0 88.3 23.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/udjpi6f8",
   "source_url": "https://omlx.ai/benchmarks/performance/udjpi6f8",
   "tg_tokens": null,
   "tps": 88.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-jxmtook5-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 791.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 791.5 96.4 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jxmtook5",
   "source_url": "https://omlx.ai/benchmarks/performance/jxmtook5",
   "tg_tokens": null,
   "tps": 96.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-jxmtook5-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 869.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 869.2 87.5 23.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jxmtook5",
   "source_url": "https://omlx.ai/benchmarks/performance/jxmtook5",
   "tg_tokens": null,
   "tps": 87.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-jxmtook5-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 813.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 8k 813.7 112.4 23.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jxmtook5",
   "source_url": "https://omlx.ai/benchmarks/performance/jxmtook5",
   "tg_tokens": null,
   "tps": 112.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-kohernsv-ctx4096",
   "model": "Qwen3.6-40B-Claude-4.6-Opus-Deckard-Heretic",
   "notes": "hi MLX build; oMLX community benchmark trace; peak memory 29.3 GB at this context; duplicate traces of the same session: fuv733bg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 263.1,
   "provenance": "community",
   "quant": "qx64",
   "quote": "Qwen3.6-40B-Claude-4.6-Opus-Deckard-Heretic-qx64-hi-mlx on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 263.1 10.0 29.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kohernsv",
   "source_url": "https://omlx.ai/benchmarks/performance/kohernsv",
   "tg_tokens": null,
   "tps": 10.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-kohernsv-ctx16384",
   "model": "Qwen3.6-40B-Claude-4.6-Opus-Deckard-Heretic",
   "notes": "hi MLX build; oMLX community benchmark trace; peak memory 30.2 GB at this context; duplicate traces of the same session: fuv733bg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.0,
   "provenance": "community",
   "quant": "qx64",
   "quote": "Qwen3.6-40B-Claude-4.6-Opus-Deckard-Heretic-qx64-hi-mlx on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 16k 224.0 8.3 30.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kohernsv",
   "source_url": "https://omlx.ai/benchmarks/performance/kohernsv",
   "tg_tokens": null,
   "tps": 8.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-0nxe71cg-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: qseg3uc7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 225.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 225.7 23.3 16.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0nxe71cg",
   "source_url": "https://omlx.ai/benchmarks/performance/0nxe71cg",
   "tg_tokens": null,
   "tps": 23.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-0nxe71cg-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: qseg3uc7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 231.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 231.9 22.8 19.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0nxe71cg",
   "source_url": "https://omlx.ai/benchmarks/performance/0nxe71cg",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-7rz0ex5i-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 1cj8adju; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 131.0 15.5 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7rz0ex5i",
   "source_url": "https://omlx.ai/benchmarks/performance/7rz0ex5i",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-7rz0ex5i-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 1cj8adju; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 135.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 135.6 15.0 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7rz0ex5i",
   "source_url": "https://omlx.ai/benchmarks/performance/7rz0ex5i",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-bkn8la0r-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 9mfid33s, gt02li1v, zqhbae01; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1185.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 1k 1,185 33.0 28.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bkn8la0r",
   "source_url": "https://omlx.ai/benchmarks/performance/bkn8la0r",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-bkn8la0r-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: 9mfid33s, gt02li1v, zqhbae01; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1426.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 4k 1,426 32.8 29.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bkn8la0r",
   "source_url": "https://omlx.ai/benchmarks/performance/bkn8la0r",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-bkn8la0r-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.8 GB at this context; duplicate traces of the same session: 9mfid33s, gt02li1v, zqhbae01; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1430.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 8k 1,430 32.2 29.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bkn8la0r",
   "source_url": "https://omlx.ai/benchmarks/performance/bkn8la0r",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-bkn8la0r-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: 9mfid33s, gt02li1v, zqhbae01; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1392.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 16k 1,392 31.5 30.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bkn8la0r",
   "source_url": "https://omlx.ai/benchmarks/performance/bkn8la0r",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-a41ns6ci-ctx1024",
   "model": "Qwen3.8-27B-CRACK",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 13060djj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 906.0,
   "provenance": "community",
   "quant": "DWQ-4bit",
   "quote": "Qwen3.8-27B-CRACK-DWQ-4bit-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 906.0 44.1 19.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a41ns6ci",
   "source_url": "https://omlx.ai/benchmarks/performance/a41ns6ci",
   "tg_tokens": null,
   "tps": 44.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-a41ns6ci-ctx4096",
   "model": "Qwen3.8-27B-CRACK",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 13060djj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 942.9,
   "provenance": "community",
   "quant": "DWQ-4bit",
   "quote": "Qwen3.8-27B-CRACK-DWQ-4bit-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 942.9 55.8 21.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a41ns6ci",
   "source_url": "https://omlx.ai/benchmarks/performance/a41ns6ci",
   "tg_tokens": null,
   "tps": 55.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-p7mmgdm0-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 2iu16v6h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 925.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 925.0 33.8 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p7mmgdm0",
   "source_url": "https://omlx.ai/benchmarks/performance/p7mmgdm0",
   "tg_tokens": null,
   "tps": 33.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-p7mmgdm0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 2iu16v6h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1012.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,012 33.3 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p7mmgdm0",
   "source_url": "https://omlx.ai/benchmarks/performance/p7mmgdm0",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-34khwhlt-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 83h0ahbs; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 457.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 457.8 18.2 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/34khwhlt",
   "source_url": "https://omlx.ai/benchmarks/performance/34khwhlt",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-34khwhlt-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 83h0ahbs; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 429.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 429.6 17.6 17.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/34khwhlt",
   "source_url": "https://omlx.ai/benchmarks/performance/34khwhlt",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-tq0q4bts-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: sa7m125z; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 305.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 305.9 9.6 28.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tq0q4bts",
   "source_url": "https://omlx.ai/benchmarks/performance/tq0q4bts",
   "tg_tokens": null,
   "tps": 9.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-tq0q4bts-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: sa7m125z; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 315.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 315.7 9.6 29.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tq0q4bts",
   "source_url": "https://omlx.ai/benchmarks/performance/tq0q4bts",
   "tg_tokens": null,
   "tps": 9.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-rw043tuv-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: jx18w2x5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 276.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 276.1 9.8 28.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rw043tuv",
   "source_url": "https://omlx.ai/benchmarks/performance/rw043tuv",
   "tg_tokens": null,
   "tps": 9.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-rw043tuv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: jx18w2x5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 296.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 296.3 9.7 29.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rw043tuv",
   "source_url": "https://omlx.ai/benchmarks/performance/rw043tuv",
   "tg_tokens": null,
   "tps": 9.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-iaccu756-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: it8g8qzr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 806.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-6Bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 806.7 24.3 22.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iaccu756",
   "source_url": "https://omlx.ai/benchmarks/performance/iaccu756",
   "tg_tokens": null,
   "tps": 24.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-iaccu756-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: it8g8qzr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 909.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-6Bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 909.3 23.7 23.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iaccu756",
   "source_url": "https://omlx.ai/benchmarks/performance/iaccu756",
   "tg_tokens": null,
   "tps": 23.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n4r373qg-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: kiuec6hq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 812.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ8e-1M on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 812.9 18.3 28.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n4r373qg",
   "source_url": "https://omlx.ai/benchmarks/performance/n4r373qg",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n4r373qg-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: kiuec6hq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 922.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ8e-1M on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 922.7 18.6 29.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n4r373qg",
   "source_url": "https://omlx.ai/benchmarks/performance/n4r373qg",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-4s8rznx8-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: i8m0m3q1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1824.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 1,824 51.0 17.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4s8rznx8",
   "source_url": "https://omlx.ai/benchmarks/performance/4s8rznx8",
   "tg_tokens": null,
   "tps": 51.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-4s8rznx8-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: i8m0m3q1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1803.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 1,803 48.2 18.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4s8rznx8",
   "source_url": "https://omlx.ai/benchmarks/performance/4s8rznx8",
   "tg_tokens": null,
   "tps": 48.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-8o12k41b-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.6 GB at this context; duplicate traces of the same session: 1qd8xntq; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 100.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 1k 100.0 21.6 37.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8o12k41b",
   "source_url": "https://omlx.ai/benchmarks/performance/8o12k41b",
   "tg_tokens": null,
   "tps": 21.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-8o12k41b-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.5 GB at this context; duplicate traces of the same session: 1qd8xntq; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 4k 102.1 17.6 37.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8o12k41b",
   "source_url": "https://omlx.ai/benchmarks/performance/8o12k41b",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wgstibpc-ctx1024",
   "model": "Qwen3.8-Flash-Next-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 78.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1100.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-Abliterated-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,100 68.9 78.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wgstibpc",
   "source_url": "https://omlx.ai/benchmarks/performance/wgstibpc",
   "tg_tokens": null,
   "tps": 68.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wgstibpc-ctx4096",
   "model": "Qwen3.8-Flash-Next-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 78.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1785.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-Abliterated-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,785 65.5 78.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wgstibpc",
   "source_url": "https://omlx.ai/benchmarks/performance/wgstibpc",
   "tg_tokens": null,
   "tps": 65.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-uc2dr22w-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 99.7 GB; TTFT 853 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4803.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k tokens 4,803 PP tok/s 116.9 TG tok/s 853 TTFT (ms) 99.7 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uc2dr22w",
   "source_url": "https://omlx.ai/benchmarks/performance/uc2dr22w",
   "tg_tokens": null,
   "tps": 116.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-c8j1fzlb-ctx1024",
   "model": "Swift-1.5-3bit",
   "notes": "TextOnly; oMLX community benchmark trace, single run at 1k context; peak memory 13.0 GB; TTFT 9306 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 110.0,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Swift-1.5-3bit-MLX-TextOnly on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 3bit; 1k tokens 110.0 PP tok/s 17.1 TG tok/s 9306 TTFT (ms) 13.0 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c8j1fzlb",
   "source_url": "https://omlx.ai/benchmarks/performance/c8j1fzlb",
   "tg_tokens": null,
   "tps": 17.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3elzvi59-ctx1024",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: 81f39rtb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1340.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 1,340 93.1 25.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3elzvi59",
   "source_url": "https://omlx.ai/benchmarks/performance/3elzvi59",
   "tg_tokens": null,
   "tps": 93.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3elzvi59-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: 81f39rtb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3690.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 3,690 90.6 26.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3elzvi59",
   "source_url": "https://omlx.ai/benchmarks/performance/3elzvi59",
   "tg_tokens": null,
   "tps": 90.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-svfog910-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: 6obz5tti, a5vd2cpl; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1407.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "qwen35-9b-4bit on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 1k 1,407 37.2 6.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/svfog910",
   "source_url": "https://omlx.ai/benchmarks/performance/svfog910",
   "tg_tokens": null,
   "tps": 37.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-svfog910-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.1 GB at this context; duplicate traces of the same session: 6obz5tti, a5vd2cpl; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1542.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "qwen35-9b-4bit on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 4k 1,542 35.3 7.1 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/svfog910",
   "source_url": "https://omlx.ai/benchmarks/performance/svfog910",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-svfog910-ctx199680",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 13.1 GB at this context; duplicate traces of the same session: 6obz5tti, a5vd2cpl; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 596.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "qwen35-9b-4bit on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 195k 596.7 22.2 13.1 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/svfog910",
   "source_url": "https://omlx.ai/benchmarks/performance/svfog910",
   "tg_tokens": null,
   "tps": 22.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-jwcgrqv2-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: tw43b3vc; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1465.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "qwen35-9b-4bit on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 1k 1,465 38.3 6.7 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jwcgrqv2",
   "source_url": "https://omlx.ai/benchmarks/performance/jwcgrqv2",
   "tg_tokens": null,
   "tps": 38.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-jwcgrqv2-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.1 GB at this context; duplicate traces of the same session: tw43b3vc; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1545.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "qwen35-9b-4bit on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 4k 1,545 34.6 7.1 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jwcgrqv2",
   "source_url": "https://omlx.ai/benchmarks/performance/jwcgrqv2",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-59emro09-ctx1024",
   "model": "Qwen3.8-27B-aeon-ultimate-uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: e7yi34ni; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 890.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "swdjojo--Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 890.6 54.7 17.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/59emro09",
   "source_url": "https://omlx.ai/benchmarks/performance/59emro09",
   "tg_tokens": null,
   "tps": 54.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-59emro09-ctx4096",
   "model": "Qwen3.8-27B-aeon-ultimate-uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: e7yi34ni; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 969.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "swdjojo--Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 969.1 62.8 19.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/59emro09",
   "source_url": "https://omlx.ai/benchmarks/performance/59emro09",
   "tg_tokens": null,
   "tps": 62.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-8p35um0e-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "oMLX community benchmark trace; peak memory 288.9 GB at this context; duplicate traces of the same session: dvp1bn0s, hesi5rbn, w1yxbr1v; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 541.7,
   "provenance": "community",
   "quant": "unknown",
   "quote": "DeepSeek-V4.1-Flash on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 541.7 30.8 288.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8p35um0e",
   "source_url": "https://omlx.ai/benchmarks/performance/8p35um0e",
   "tg_tokens": null,
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-8p35um0e-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "oMLX community benchmark trace; peak memory 289.3 GB at this context; duplicate traces of the same session: dvp1bn0s, hesi5rbn, w1yxbr1v; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 711.4,
   "provenance": "community",
   "quant": "unknown",
   "quote": "DeepSeek-V4.1-Flash on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 711.4 32.2 289.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8p35um0e",
   "source_url": "https://omlx.ai/benchmarks/performance/8p35um0e",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-8p35um0e-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "oMLX community benchmark trace; peak memory 289.3 GB at this context; duplicate traces of the same session: dvp1bn0s, hesi5rbn, w1yxbr1v; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 736.2,
   "provenance": "community",
   "quant": "unknown",
   "quote": "DeepSeek-V4.1-Flash on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 8k 736.2 36.0 289.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8p35um0e",
   "source_url": "https://omlx.ai/benchmarks/performance/8p35um0e",
   "tg_tokens": null,
   "tps": 36.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-8p35um0e-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "oMLX community benchmark trace; peak memory 289.3 GB at this context; duplicate traces of the same session: dvp1bn0s, hesi5rbn, w1yxbr1v; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 748.8,
   "provenance": "community",
   "quant": "unknown",
   "quote": "DeepSeek-V4.1-Flash on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 16k 748.8 35.2 289.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8p35um0e",
   "source_url": "https://omlx.ai/benchmarks/performance/8p35um0e",
   "tg_tokens": null,
   "tps": 35.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-e1gcehr5-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "oMLX community benchmark trace; peak memory 288.9 GB at this context; duplicate traces of the same session: 22b6sa83, 28b0217t, vd4ful4d; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 361.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "DeepSeek-V4.1-Flash on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 361.0 29.0 288.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e1gcehr5",
   "source_url": "https://omlx.ai/benchmarks/performance/e1gcehr5",
   "tg_tokens": null,
   "tps": 29.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-e1gcehr5-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "oMLX community benchmark trace; peak memory 289.4 GB at this context; duplicate traces of the same session: 22b6sa83, 28b0217t, vd4ful4d; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 418.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "DeepSeek-V4.1-Flash on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 418.0 33.6 289.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e1gcehr5",
   "source_url": "https://omlx.ai/benchmarks/performance/e1gcehr5",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-e1gcehr5-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "oMLX community benchmark trace; peak memory 289.7 GB at this context; duplicate traces of the same session: 22b6sa83, 28b0217t, vd4ful4d; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 422.8,
   "provenance": "community",
   "quant": "unknown",
   "quote": "DeepSeek-V4.1-Flash on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 8k 422.8 37.0 289.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e1gcehr5",
   "source_url": "https://omlx.ai/benchmarks/performance/e1gcehr5",
   "tg_tokens": null,
   "tps": 37.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-e1gcehr5-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "oMLX community benchmark trace; peak memory 290.1 GB at this context; duplicate traces of the same session: 22b6sa83, 28b0217t, vd4ful4d; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 421.1,
   "provenance": "community",
   "quant": "unknown",
   "quote": "DeepSeek-V4.1-Flash on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 16k 421.1 38.1 290.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e1gcehr5",
   "source_url": "https://omlx.ai/benchmarks/performance/e1gcehr5",
   "tg_tokens": null,
   "tps": 38.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5dyjd7k4-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.0 GB at this context; duplicate traces of the same session: 34a5kd76, 4x3wt4jq, lthdqiy6; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 88.6,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4.1-Flash-oQ2e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 88.6 14.6 100.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5dyjd7k4",
   "source_url": "https://omlx.ai/benchmarks/performance/5dyjd7k4",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5dyjd7k4-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.4 GB at this context; duplicate traces of the same session: 34a5kd76, 4x3wt4jq, lthdqiy6; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 192.9,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4.1-Flash-oQ2e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 192.9 15.1 100.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5dyjd7k4",
   "source_url": "https://omlx.ai/benchmarks/performance/5dyjd7k4",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5dyjd7k4-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.4 GB at this context; duplicate traces of the same session: 34a5kd76, 4x3wt4jq, lthdqiy6; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 195.4,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4.1-Flash-oQ2e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 8k 195.4 17.0 100.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5dyjd7k4",
   "source_url": "https://omlx.ai/benchmarks/performance/5dyjd7k4",
   "tg_tokens": null,
   "tps": 17.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5dyjd7k4-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: 34a5kd76, 4x3wt4jq, lthdqiy6; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 192.7,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4.1-Flash-oQ2e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 16k 192.7 19.2 100.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5dyjd7k4",
   "source_url": "https://omlx.ai/benchmarks/performance/5dyjd7k4",
   "tg_tokens": null,
   "tps": 19.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-vme2oy04-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.4 GB at this context; duplicate traces of the same session: 274hs8ub; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 837.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 837.5 57.3 178.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vme2oy04",
   "source_url": "https://omlx.ai/benchmarks/performance/vme2oy04",
   "tg_tokens": null,
   "tps": 57.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-vme2oy04-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.3 GB at this context; duplicate traces of the same session: 274hs8ub; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2317.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 2,317 57.3 179.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vme2oy04",
   "source_url": "https://omlx.ai/benchmarks/performance/vme2oy04",
   "tg_tokens": null,
   "tps": 57.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-4fj2ssas-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.4 GB at this context; duplicate traces of the same session: ej1si183; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1990.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 1,990 57.4 178.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4fj2ssas",
   "source_url": "https://omlx.ai/benchmarks/performance/4fj2ssas",
   "tg_tokens": null,
   "tps": 57.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-4fj2ssas-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.3 GB at this context; duplicate traces of the same session: ej1si183; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2354.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 2,354 57.5 179.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4fj2ssas",
   "source_url": "https://omlx.ai/benchmarks/performance/4fj2ssas",
   "tg_tokens": null,
   "tps": 57.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-p772qk22-ctx1024",
   "model": "Qwen2.5-VL-3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 2.2 GB at this context; duplicate traces of the same session: 3gvdzfb0; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1795.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2___5-VL-3B-Instruct-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k 1,795 63.4 2.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p772qk22",
   "source_url": "https://omlx.ai/benchmarks/performance/p772qk22",
   "tg_tokens": null,
   "tps": 63.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-p772qk22-ctx4096",
   "model": "Qwen2.5-VL-3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 2.4 GB at this context; duplicate traces of the same session: 3gvdzfb0; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1701.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2___5-VL-3B-Instruct-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 1,701 57.4 2.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p772qk22",
   "source_url": "https://omlx.ai/benchmarks/performance/p772qk22",
   "tg_tokens": null,
   "tps": 57.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-61cl1s03-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 214 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4788.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 4,788 PP tok/s 153.9 TG tok/s 214 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/61cl1s03",
   "source_url": "https://omlx.ai/benchmarks/performance/61cl1s03",
   "tg_tokens": null,
   "tps": 153.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yruiwxe4-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 213 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4814.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 4,814 PP tok/s 149.5 TG tok/s 213 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yruiwxe4",
   "source_url": "https://omlx.ai/benchmarks/performance/yruiwxe4",
   "tg_tokens": null,
   "tps": 149.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-84wv03d7-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: ops2tl6b; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 791.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 791.5 96.4 21.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/84wv03d7",
   "source_url": "https://omlx.ai/benchmarks/performance/84wv03d7",
   "tg_tokens": null,
   "tps": 96.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-84wv03d7-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: ops2tl6b; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 869.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 869.2 87.5 23.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/84wv03d7",
   "source_url": "https://omlx.ai/benchmarks/performance/84wv03d7",
   "tg_tokens": null,
   "tps": 87.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-84wv03d7-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: ops2tl6b; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 813.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 8k 813.7 112.4 23.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/84wv03d7",
   "source_url": "https://omlx.ai/benchmarks/performance/84wv03d7",
   "tg_tokens": null,
   "tps": 112.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-jqv4xwtj-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: gwuob2yp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 784.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 784.2 81.9 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jqv4xwtj",
   "source_url": "https://omlx.ai/benchmarks/performance/jqv4xwtj",
   "tg_tokens": null,
   "tps": 81.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-jqv4xwtj-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: gwuob2yp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 848.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 848.4 63.6 23.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jqv4xwtj",
   "source_url": "https://omlx.ai/benchmarks/performance/jqv4xwtj",
   "tg_tokens": null,
   "tps": 63.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-uo73mme3-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: rcp49j7w; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 791.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 791.1 104.3 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uo73mme3",
   "source_url": "https://omlx.ai/benchmarks/performance/uo73mme3",
   "tg_tokens": null,
   "tps": 104.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-uo73mme3-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: rcp49j7w; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 869.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 869.0 88.0 23.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uo73mme3",
   "source_url": "https://omlx.ai/benchmarks/performance/uo73mme3",
   "tg_tokens": null,
   "tps": 88.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-nz89blnw-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 8htg0mvs; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 791.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 791.8 98.4 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nz89blnw",
   "source_url": "https://omlx.ai/benchmarks/performance/nz89blnw",
   "tg_tokens": null,
   "tps": 98.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-nz89blnw-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: 8htg0mvs; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 911.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 911.5 87.6 23.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nz89blnw",
   "source_url": "https://omlx.ai/benchmarks/performance/nz89blnw",
   "tg_tokens": null,
   "tps": 87.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-y5k3y7kl-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: d8u5qmqr; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 791.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 791.6 102.3 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y5k3y7kl",
   "source_url": "https://omlx.ai/benchmarks/performance/y5k3y7kl",
   "tg_tokens": null,
   "tps": 102.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-y5k3y7kl-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: d8u5qmqr; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 869.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 869.2 88.7 23.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y5k3y7kl",
   "source_url": "https://omlx.ai/benchmarks/performance/y5k3y7kl",
   "tg_tokens": null,
   "tps": 88.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-jppll7d0-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: wzofxe6m; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 106.1 15.5 16.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jppll7d0",
   "source_url": "https://omlx.ai/benchmarks/performance/jppll7d0",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-jppll7d0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: wzofxe6m; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 106.8 14.6 17.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jppll7d0",
   "source_url": "https://omlx.ai/benchmarks/performance/jppll7d0",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-3ddlvbbu-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: fc0tqa5u; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 105.9 15.5 16.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ddlvbbu",
   "source_url": "https://omlx.ai/benchmarks/performance/3ddlvbbu",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-3ddlvbbu-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: fc0tqa5u; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 108.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 108.6 15.1 17.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ddlvbbu",
   "source_url": "https://omlx.ai/benchmarks/performance/3ddlvbbu",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-n3fyzw1a-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 220.6 48.8 18.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n3fyzw1a",
   "source_url": "https://omlx.ai/benchmarks/performance/n3fyzw1a",
   "tg_tokens": null,
   "tps": 48.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-n3fyzw1a-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 199.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 199.1 41.0 22.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n3fyzw1a",
   "source_url": "https://omlx.ai/benchmarks/performance/n3fyzw1a",
   "tg_tokens": null,
   "tps": 41.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-n3fyzw1a-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 67.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 67.3 15.7 25.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n3fyzw1a",
   "source_url": "https://omlx.ai/benchmarks/performance/n3fyzw1a",
   "tg_tokens": null,
   "tps": 15.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-fxt66mun-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: v03k2f1j; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 104.2 25.3 18.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fxt66mun",
   "source_url": "https://omlx.ai/benchmarks/performance/fxt66mun",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-fxt66mun-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: v03k2f1j; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 109.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 109.0 29.3 22.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fxt66mun",
   "source_url": "https://omlx.ai/benchmarks/performance/fxt66mun",
   "tg_tokens": null,
   "tps": 29.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-fxt66mun-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: v03k2f1j; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 107.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 8k 107.0 27.0 23.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fxt66mun",
   "source_url": "https://omlx.ai/benchmarks/performance/fxt66mun",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6euj1t3x-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 0hw208p0, n4yxyrpp, t9agz5j1, x6v55iw7, y765t7mg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 725.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 725.2 49.0 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6euj1t3x",
   "source_url": "https://omlx.ai/benchmarks/performance/6euj1t3x",
   "tg_tokens": null,
   "tps": 49.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6euj1t3x-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 0hw208p0, n4yxyrpp, t9agz5j1, x6v55iw7, y765t7mg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 523.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 523.5 62.5 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6euj1t3x",
   "source_url": "https://omlx.ai/benchmarks/performance/6euj1t3x",
   "tg_tokens": null,
   "tps": 62.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6euj1t3x-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 0hw208p0, n4yxyrpp, t9agz5j1, x6v55iw7, y765t7mg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 576.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 576.0 71.0 22.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6euj1t3x",
   "source_url": "https://omlx.ai/benchmarks/performance/6euj1t3x",
   "tg_tokens": null,
   "tps": 71.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6euj1t3x-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 0hw208p0, n4yxyrpp, t9agz5j1, x6v55iw7, y765t7mg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 568.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 568.2 82.7 23.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6euj1t3x",
   "source_url": "https://omlx.ai/benchmarks/performance/6euj1t3x",
   "tg_tokens": null,
   "tps": 82.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6euj1t3x-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: 0hw208p0, n4yxyrpp, t9agz5j1, x6v55iw7, y765t7mg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 544.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 544.9 80.7 24.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6euj1t3x",
   "source_url": "https://omlx.ai/benchmarks/performance/6euj1t3x",
   "tg_tokens": null,
   "tps": 80.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-6euj1t3x-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: 0hw208p0, n4yxyrpp, t9agz5j1, x6v55iw7, y765t7mg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 498.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 498.9 55.6 26.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6euj1t3x",
   "source_url": "https://omlx.ai/benchmarks/performance/6euj1t3x",
   "tg_tokens": null,
   "tps": 55.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-9aelip2d-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 31.8 GB; TTFT 91119 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 359.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 32k tokens 359.6 PP tok/s 42.1 TG tok/s 91119 TTFT (ms) 31.8 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9aelip2d",
   "source_url": "https://omlx.ai/benchmarks/performance/9aelip2d",
   "tg_tokens": null,
   "tps": 42.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-kpys5nzk-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 31.8 GB; TTFT 87981 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 372.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 32k tokens 372.4 PP tok/s 43.8 TG tok/s 87981 TTFT (ms) 31.8 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kpys5nzk",
   "source_url": "https://omlx.ai/benchmarks/performance/kpys5nzk",
   "tg_tokens": null,
   "tps": 43.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-pz2shtpp-ctx4096",
   "model": "Qwen3.8-9B-Distill-Uncensored-Heretic",
   "notes": "MTP build (mlx); oMLX community benchmark trace; peak memory 11.2 GB at this context; duplicate traces of the same session: un7urkiy; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 403.5,
   "provenance": "community",
   "quant": "Q6G64",
   "quote": "Qwen3.8-9B-Distill-Uncensored-Heretic-MTPLX-Q6G64 on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 6bit; 4k 403.5 29.9 11.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pz2shtpp",
   "source_url": "https://omlx.ai/benchmarks/performance/pz2shtpp",
   "tg_tokens": null,
   "tps": 29.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-pz2shtpp-ctx16384",
   "model": "Qwen3.8-9B-Distill-Uncensored-Heretic",
   "notes": "MTP build (mlx); oMLX community benchmark trace; peak memory 12.4 GB at this context; duplicate traces of the same session: un7urkiy; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 382.1,
   "provenance": "community",
   "quant": "Q6G64",
   "quote": "Qwen3.8-9B-Distill-Uncensored-Heretic-MTPLX-Q6G64 on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 6bit; 16k 382.1 26.5 12.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pz2shtpp",
   "source_url": "https://omlx.ai/benchmarks/performance/pz2shtpp",
   "tg_tokens": null,
   "tps": 26.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-2kiq2frh-ctx1024",
   "model": "Qwythos-9B-Claude-Mythos-5",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: v9w3q4tt; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 225.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwythos-9B-Claude-Mythos-5-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 225.8 18.6 5.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2kiq2frh",
   "source_url": "https://omlx.ai/benchmarks/performance/2kiq2frh",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-2kiq2frh-ctx4096",
   "model": "Qwythos-9B-Claude-Mythos-5",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: v9w3q4tt; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 226.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwythos-9B-Claude-Mythos-5-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 226.7 19.4 6.4 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2kiq2frh",
   "source_url": "https://omlx.ai/benchmarks/performance/2kiq2frh",
   "tg_tokens": null,
   "tps": 19.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3etfmouq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: ega75jnj, g9bmyzb7, lzi75tgo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1044.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,044 22.5 22.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3etfmouq",
   "source_url": "https://omlx.ai/benchmarks/performance/3etfmouq",
   "tg_tokens": null,
   "tps": 22.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3etfmouq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: ega75jnj, g9bmyzb7, lzi75tgo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 831.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 831.4 21.2 24.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3etfmouq",
   "source_url": "https://omlx.ai/benchmarks/performance/3etfmouq",
   "tg_tokens": null,
   "tps": 21.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3etfmouq-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: ega75jnj, g9bmyzb7, lzi75tgo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 858.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 858.9 18.9 24.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3etfmouq",
   "source_url": "https://omlx.ai/benchmarks/performance/3etfmouq",
   "tg_tokens": null,
   "tps": 18.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3etfmouq-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace; peak memory 28.2 GB at this context; duplicate traces of the same session: ega75jnj, g9bmyzb7, lzi75tgo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 625.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 625.7 14.6 28.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3etfmouq",
   "source_url": "https://omlx.ai/benchmarks/performance/3etfmouq",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q62iyt1q-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: 6x2u8poy, eos66rnd, l7cm8xc0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 665.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 665.0 19.5 22.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q62iyt1q",
   "source_url": "https://omlx.ai/benchmarks/performance/q62iyt1q",
   "tg_tokens": null,
   "tps": 19.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q62iyt1q-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: 6x2u8poy, eos66rnd, l7cm8xc0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 765.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 765.0 20.6 24.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q62iyt1q",
   "source_url": "https://omlx.ai/benchmarks/performance/q62iyt1q",
   "tg_tokens": null,
   "tps": 20.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q62iyt1q-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: 6x2u8poy, eos66rnd, l7cm8xc0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 828.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 828.3 17.2 24.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q62iyt1q",
   "source_url": "https://omlx.ai/benchmarks/performance/q62iyt1q",
   "tg_tokens": null,
   "tps": 17.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q62iyt1q-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace; peak memory 28.2 GB at this context; duplicate traces of the same session: 6x2u8poy, eos66rnd, l7cm8xc0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 630.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 630.3 14.4 28.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q62iyt1q",
   "source_url": "https://omlx.ai/benchmarks/performance/q62iyt1q",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-013w1rnv-ctx32768",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 11.5 GB; TTFT 105143 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 311.7,
   "provenance": "community",
   "quant": "MXFP4-Q4",
   "quote": "gpt-oss-20b-MXFP4-Q4 on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 32k tokens 311.7 PP tok/s 27.4 TG tok/s 105143 TTFT (ms) 11.5 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/013w1rnv",
   "source_url": "https://omlx.ai/benchmarks/performance/013w1rnv",
   "tg_tokens": null,
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-v3v72fuh-ctx1024",
   "model": "Hy-MT2-1.8B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 1.8 GB at this context; duplicate traces of the same session: s3m1aajt; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 847.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Hy-MT2-1.8B-MLX-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 847.4 105.8 1.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v3v72fuh",
   "source_url": "https://omlx.ai/benchmarks/performance/v3v72fuh",
   "tg_tokens": null,
   "tps": 105.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-v3v72fuh-ctx4096",
   "model": "Hy-MT2-1.8B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 1.9 GB at this context; duplicate traces of the same session: s3m1aajt; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 774.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Hy-MT2-1.8B-MLX-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 774.7 91.7 1.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v3v72fuh",
   "source_url": "https://omlx.ai/benchmarks/performance/v3v72fuh",
   "tg_tokens": null,
   "tps": 91.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7m1qpvxw-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 22.9 GB; TTFT 245 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4180.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 4,180 PP tok/s 145.8 TG tok/s 245 TTFT (ms) 22.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7m1qpvxw",
   "source_url": "https://omlx.ai/benchmarks/performance/7m1qpvxw",
   "tg_tokens": null,
   "tps": 145.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-qbpd4r4t-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 9.8 GB at this context; duplicate traces of the same session: 43v5kqvm; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 892.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B-MLX-8bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 8bit; 1k 892.6 17.4 9.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qbpd4r4t",
   "source_url": "https://omlx.ai/benchmarks/performance/qbpd4r4t",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-qbpd4r4t-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 10.3 GB at this context; duplicate traces of the same session: 43v5kqvm; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 973.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B-MLX-8bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 8bit; 4k 973.6 17.1 10.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qbpd4r4t",
   "source_url": "https://omlx.ai/benchmarks/performance/qbpd4r4t",
   "tg_tokens": null,
   "tps": 17.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-pk5kznl1-ctx1024",
   "model": "Qwen3.5-VL-2B",
   "notes": "MLX CRACK build; oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: gf95zazs; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 11785.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-VL-2B-8bit-MLX-CRACK on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 11,785 125.1 3.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pk5kznl1",
   "source_url": "https://omlx.ai/benchmarks/performance/pk5kznl1",
   "tg_tokens": null,
   "tps": 125.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-pk5kznl1-ctx4096",
   "model": "Qwen3.5-VL-2B",
   "notes": "MLX CRACK build; oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: gf95zazs; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 13444.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-VL-2B-8bit-MLX-CRACK on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 13,444 201.7 3.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pk5kznl1",
   "source_url": "https://omlx.ai/benchmarks/performance/pk5kznl1",
   "tg_tokens": null,
   "tps": 201.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9owozprk-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 22.4 GB; TTFT 61946 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2116.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k tokens 2,116 PP tok/s 71.8 TG tok/s 61946 TTFT (ms) 22.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9owozprk",
   "source_url": "https://omlx.ai/benchmarks/performance/9owozprk",
   "tg_tokens": null,
   "tps": 71.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-x0p9d84f-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: fv5ecn21; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1687.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 1,687 72.4 19.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x0p9d84f",
   "source_url": "https://omlx.ai/benchmarks/performance/x0p9d84f",
   "tg_tokens": null,
   "tps": 72.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-x0p9d84f-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: fv5ecn21; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2149.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 2,149 70.4 19.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x0p9d84f",
   "source_url": "https://omlx.ai/benchmarks/performance/x0p9d84f",
   "tg_tokens": null,
   "tps": 70.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-scs2jtdc-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: big84i75; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3920.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 3,920 225.0 25.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/scs2jtdc",
   "source_url": "https://omlx.ai/benchmarks/performance/scs2jtdc",
   "tg_tokens": null,
   "tps": 225.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-scs2jtdc-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 27.0 GB at this context; duplicate traces of the same session: big84i75; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5915.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 5,915 197.3 27.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/scs2jtdc",
   "source_url": "https://omlx.ai/benchmarks/performance/scs2jtdc",
   "tg_tokens": null,
   "tps": 197.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-pafhp99d-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: nuepxri4, yjn12khp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 770.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 1k 770.7 80.4 37.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pafhp99d",
   "source_url": "https://omlx.ai/benchmarks/performance/pafhp99d",
   "tg_tokens": null,
   "tps": 80.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-pafhp99d-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.4 GB at this context; duplicate traces of the same session: nuepxri4, yjn12khp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 849.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 4k 849.1 73.7 39.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pafhp99d",
   "source_url": "https://omlx.ai/benchmarks/performance/pafhp99d",
   "tg_tokens": null,
   "tps": 73.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-pafhp99d-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.7 GB at this context; duplicate traces of the same session: nuepxri4, yjn12khp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 796.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 8k 796.3 58.5 39.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pafhp99d",
   "source_url": "https://omlx.ai/benchmarks/performance/pafhp99d",
   "tg_tokens": null,
   "tps": 58.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-df9g56ps-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: ltjy72zh; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 164.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 1k 164.7 25.7 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/df9g56ps",
   "source_url": "https://omlx.ai/benchmarks/performance/df9g56ps",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-df9g56ps-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: ltjy72zh; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 153.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 4k 153.7 24.9 19.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/df9g56ps",
   "source_url": "https://omlx.ai/benchmarks/performance/df9g56ps",
   "tg_tokens": null,
   "tps": 24.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lks6sqqp-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 0yt7mgyg, q5z6vow8, u2jsrjn2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 351.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 351.8 20.6 16.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lks6sqqp",
   "source_url": "https://omlx.ai/benchmarks/performance/lks6sqqp",
   "tg_tokens": null,
   "tps": 20.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lks6sqqp-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 0yt7mgyg, q5z6vow8, u2jsrjn2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 423.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 423.0 17.5 17.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lks6sqqp",
   "source_url": "https://omlx.ai/benchmarks/performance/lks6sqqp",
   "tg_tokens": null,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lks6sqqp-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 0yt7mgyg, q5z6vow8, u2jsrjn2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 447.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 447.7 19.6 17.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lks6sqqp",
   "source_url": "https://omlx.ai/benchmarks/performance/lks6sqqp",
   "tg_tokens": null,
   "tps": 19.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lks6sqqp-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 0yt7mgyg, q5z6vow8, u2jsrjn2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 462.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 462.6 20.1 17.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lks6sqqp",
   "source_url": "https://omlx.ai/benchmarks/performance/lks6sqqp",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-dgowl10l-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: y3qdu294; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 162.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 8bit; 1k 162.7 18.5 28.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dgowl10l",
   "source_url": "https://omlx.ai/benchmarks/performance/dgowl10l",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-dgowl10l-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: y3qdu294; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 162.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 8bit; 4k 162.6 18.1 32.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dgowl10l",
   "source_url": "https://omlx.ai/benchmarks/performance/dgowl10l",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-r063dagp-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: h6441n0p; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 220.6 48.8 18.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r063dagp",
   "source_url": "https://omlx.ai/benchmarks/performance/r063dagp",
   "tg_tokens": null,
   "tps": 48.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-r063dagp-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: h6441n0p; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 199.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 199.1 41.0 22.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r063dagp",
   "source_url": "https://omlx.ai/benchmarks/performance/r063dagp",
   "tg_tokens": null,
   "tps": 41.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-r063dagp-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: h6441n0p; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 67.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 67.3 15.7 25.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r063dagp",
   "source_url": "https://omlx.ai/benchmarks/performance/r063dagp",
   "tg_tokens": null,
   "tps": 15.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-92s1fpee-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: iabaoyev, tc8hs016; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 173.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 173.8 25.1 18.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/92s1fpee",
   "source_url": "https://omlx.ai/benchmarks/performance/92s1fpee",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-92s1fpee-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: iabaoyev, tc8hs016; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 198.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 198.5 28.7 22.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/92s1fpee",
   "source_url": "https://omlx.ai/benchmarks/performance/92s1fpee",
   "tg_tokens": null,
   "tps": 28.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-92s1fpee-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: iabaoyev, tc8hs016; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 62.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 62.3 9.1 25.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/92s1fpee",
   "source_url": "https://omlx.ai/benchmarks/performance/92s1fpee",
   "tg_tokens": null,
   "tps": 9.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-eb249tzr-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 104.2 25.3 18.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eb249tzr",
   "source_url": "https://omlx.ai/benchmarks/performance/eb249tzr",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-eb249tzr-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 109.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 109.0 29.3 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eb249tzr",
   "source_url": "https://omlx.ai/benchmarks/performance/eb249tzr",
   "tg_tokens": null,
   "tps": 29.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-eb249tzr-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 107.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 8k 107.0 27.0 23.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eb249tzr",
   "source_url": "https://omlx.ai/benchmarks/performance/eb249tzr",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-btj0zmxu-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: 12c5yv6f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1096.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,096 72.7 23.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/btj0zmxu",
   "source_url": "https://omlx.ai/benchmarks/performance/btj0zmxu",
   "tg_tokens": null,
   "tps": 72.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-btj0zmxu-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.6 GB at this context; duplicate traces of the same session: 12c5yv6f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1333.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,333 56.4 25.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/btj0zmxu",
   "source_url": "https://omlx.ai/benchmarks/performance/btj0zmxu",
   "tg_tokens": null,
   "tps": 56.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-65lwwvv0-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: agvrjlvu, dak802em, dk07lslu, rc43x6r2, zgexz9do; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 814.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 814.7 55.1 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/65lwwvv0",
   "source_url": "https://omlx.ai/benchmarks/performance/65lwwvv0",
   "tg_tokens": null,
   "tps": 55.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-65lwwvv0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: agvrjlvu, dak802em, dk07lslu, rc43x6r2, zgexz9do; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 931.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 931.5 55.2 27.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/65lwwvv0",
   "source_url": "https://omlx.ai/benchmarks/performance/65lwwvv0",
   "tg_tokens": null,
   "tps": 55.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-65lwwvv0-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: agvrjlvu, dak802em, dk07lslu, rc43x6r2, zgexz9do; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 930.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 8k 930.5 57.3 28.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/65lwwvv0",
   "source_url": "https://omlx.ai/benchmarks/performance/65lwwvv0",
   "tg_tokens": null,
   "tps": 57.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-65lwwvv0-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: agvrjlvu, dak802em, dk07lslu, rc43x6r2, zgexz9do; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 905.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k 905.9 72.0 28.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/65lwwvv0",
   "source_url": "https://omlx.ai/benchmarks/performance/65lwwvv0",
   "tg_tokens": null,
   "tps": 72.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-65lwwvv0-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: agvrjlvu, dak802em, dk07lslu, rc43x6r2, zgexz9do; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 849.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 849.2 90.3 29.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/65lwwvv0",
   "source_url": "https://omlx.ai/benchmarks/performance/65lwwvv0",
   "tg_tokens": null,
   "tps": 90.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-65lwwvv0-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: agvrjlvu, dak802em, dk07lslu, rc43x6r2, zgexz9do; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 748.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 64k 748.0 48.0 32.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/65lwwvv0",
   "source_url": "https://omlx.ai/benchmarks/performance/65lwwvv0",
   "tg_tokens": null,
   "tps": 48.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-b9s6mxw7-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 36.3 GB; TTFT 204466 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 320.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 64k tokens 320.5 PP tok/s 22.8 TG tok/s 204466 TTFT (ms) 36.3 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b9s6mxw7",
   "source_url": "https://omlx.ai/benchmarks/performance/b9s6mxw7",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-wpg4iinp-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: 4e2wuhmw, i78vo2kp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 1k 102.9 18.6 33.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wpg4iinp",
   "source_url": "https://omlx.ai/benchmarks/performance/wpg4iinp",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-wpg4iinp-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.5 GB at this context; duplicate traces of the same session: 4e2wuhmw, i78vo2kp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 4k 103.9 18.3 37.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wpg4iinp",
   "source_url": "https://omlx.ai/benchmarks/performance/wpg4iinp",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-wpg4iinp-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.3 GB at this context; duplicate traces of the same session: 4e2wuhmw, i78vo2kp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 100.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 8k 100.4 21.0 38.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wpg4iinp",
   "source_url": "https://omlx.ai/benchmarks/performance/wpg4iinp",
   "tg_tokens": null,
   "tps": 21.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-iidys3dl-ctx1024",
   "model": "Qwythos-9B-v2",
   "notes": "oMLX community benchmark trace; peak memory 7.8 GB at this context; duplicate traces of the same session: 4qsblwxn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 895.5,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwythos-9B-v2-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 895.5 68.1 7.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iidys3dl",
   "source_url": "https://omlx.ai/benchmarks/performance/iidys3dl",
   "tg_tokens": null,
   "tps": 68.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-iidys3dl-ctx4096",
   "model": "Qwythos-9B-v2",
   "notes": "oMLX community benchmark trace; peak memory 9.1 GB at this context; duplicate traces of the same session: 4qsblwxn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 932.4,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwythos-9B-v2-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 932.4 66.4 9.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iidys3dl",
   "source_url": "https://omlx.ai/benchmarks/performance/iidys3dl",
   "tg_tokens": null,
   "tps": 66.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-y7xrpo4l-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: ayw1i0z3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 455.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 455.5 31.0 17.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y7xrpo4l",
   "source_url": "https://omlx.ai/benchmarks/performance/y7xrpo4l",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-y7xrpo4l-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: ayw1i0z3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 463.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 463.0 31.6 18.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y7xrpo4l",
   "source_url": "https://omlx.ai/benchmarks/performance/y7xrpo4l",
   "tg_tokens": null,
   "tps": 31.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-n76gbayn-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift1.5 build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 86.5 GB at this context; duplicate traces of the same session: r6yxo5c0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1062.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 1,062 87.2 86.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n76gbayn",
   "source_url": "https://omlx.ai/benchmarks/performance/n76gbayn",
   "tg_tokens": null,
   "tps": 87.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-n76gbayn-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift1.5 build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 87.8 GB at this context; duplicate traces of the same session: r6yxo5c0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2347.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 2,347 78.3 87.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n76gbayn",
   "source_url": "https://omlx.ai/benchmarks/performance/n76gbayn",
   "tg_tokens": null,
   "tps": 78.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yr36qe8i-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift1.5 build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 86.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1077.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 1,077 72.3 86.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yr36qe8i",
   "source_url": "https://omlx.ai/benchmarks/performance/yr36qe8i",
   "tg_tokens": null,
   "tps": 72.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-yr36qe8i-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift1.5 build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 87.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2342.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 2,342 79.2 87.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yr36qe8i",
   "source_url": "https://omlx.ai/benchmarks/performance/yr36qe8i",
   "tg_tokens": null,
   "tps": 79.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bap2pt3n-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift1.5 build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.1 GB at this context; duplicate traces of the same session: f1wzdtb6, jej2bhoa, m0rwcdrc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 143.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 143.9 31.3 79.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bap2pt3n",
   "source_url": "https://omlx.ai/benchmarks/performance/bap2pt3n",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bap2pt3n-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift1.5 build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.5 GB at this context; duplicate traces of the same session: f1wzdtb6, jej2bhoa, m0rwcdrc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 823.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 823.7 42.0 79.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bap2pt3n",
   "source_url": "https://omlx.ai/benchmarks/performance/bap2pt3n",
   "tg_tokens": null,
   "tps": 42.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bap2pt3n-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift1.5 build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 81.5 GB at this context; duplicate traces of the same session: f1wzdtb6, jej2bhoa, m0rwcdrc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1210.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 1,210 34.8 81.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bap2pt3n",
   "source_url": "https://omlx.ai/benchmarks/performance/bap2pt3n",
   "tg_tokens": null,
   "tps": 34.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bap2pt3n-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift1.5 build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 81.9 GB at this context; duplicate traces of the same session: f1wzdtb6, jej2bhoa, m0rwcdrc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1444.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 1,444 42.5 81.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bap2pt3n",
   "source_url": "https://omlx.ai/benchmarks/performance/bap2pt3n",
   "tg_tokens": null,
   "tps": 42.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mesj828c-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift1.5 build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 82.9 GB at this context; duplicate traces of the same session: 6gn6n255; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 150.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 150.2 44.7 82.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mesj828c",
   "source_url": "https://omlx.ai/benchmarks/performance/mesj828c",
   "tg_tokens": null,
   "tps": 44.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mesj828c-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift1.5 build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 83.2 GB at this context; duplicate traces of the same session: 6gn6n255; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 962.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 962.0 47.5 83.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mesj828c",
   "source_url": "https://omlx.ai/benchmarks/performance/mesj828c",
   "tg_tokens": null,
   "tps": 47.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mesj828c-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift1.5 build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 85.3 GB at this context; duplicate traces of the same session: 6gn6n255; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1384.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 1,384 35.0 85.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mesj828c",
   "source_url": "https://omlx.ai/benchmarks/performance/mesj828c",
   "tg_tokens": null,
   "tps": 35.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-p0kd5ecu-ctx1024",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: erjjz5y9; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2072.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 2,072 54.0 25.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p0kd5ecu",
   "source_url": "https://omlx.ai/benchmarks/performance/p0kd5ecu",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-p0kd5ecu-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: erjjz5y9; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1533.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 1,533 53.2 26.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p0kd5ecu",
   "source_url": "https://omlx.ai/benchmarks/performance/p0kd5ecu",
   "tg_tokens": null,
   "tps": 53.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-vnro5hmx-ctx1024",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: eywynlj6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 807.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 807.8 51.5 25.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vnro5hmx",
   "source_url": "https://omlx.ai/benchmarks/performance/vnro5hmx",
   "tg_tokens": null,
   "tps": 51.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-vnro5hmx-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: eywynlj6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2097.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 2,097 50.9 26.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vnro5hmx",
   "source_url": "https://omlx.ai/benchmarks/performance/vnro5hmx",
   "tg_tokens": null,
   "tps": 50.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-02zosvio-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: 21lwu356; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 537.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 537.3 61.2 14.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/02zosvio",
   "source_url": "https://omlx.ai/benchmarks/performance/02zosvio",
   "tg_tokens": null,
   "tps": 61.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-02zosvio-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: 21lwu356; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1569.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,569 65.7 14.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/02zosvio",
   "source_url": "https://omlx.ai/benchmarks/performance/02zosvio",
   "tg_tokens": null,
   "tps": 65.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-cbuletgv-ctx1024",
   "model": "Apodex-1.1-mini",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.1 GB at this context; duplicate traces of the same session: 4uxvvkdy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1479.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Apodex-1.1-mini-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 1,479 79.7 30.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cbuletgv",
   "source_url": "https://omlx.ai/benchmarks/performance/cbuletgv",
   "tg_tokens": null,
   "tps": 79.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-cbuletgv-ctx4096",
   "model": "Apodex-1.1-mini",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: 4uxvvkdy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1955.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Apodex-1.1-mini-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 1,955 98.6 32.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cbuletgv",
   "source_url": "https://omlx.ai/benchmarks/performance/cbuletgv",
   "tg_tokens": null,
   "tps": 98.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-9sbz66tl-ctx1024",
   "model": "Huihui-gemma-4-26B-A4B-it-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.6 GB at this context; duplicate traces of the same session: e5s3z3l7, nrcaz6im; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 666.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Huihui-gemma-4-26B-A4B-it-abliterated-oQ8e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 1k 666.8 45.4 27.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9sbz66tl",
   "source_url": "https://omlx.ai/benchmarks/performance/9sbz66tl",
   "tg_tokens": null,
   "tps": 45.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-9sbz66tl-ctx4096",
   "model": "Huihui-gemma-4-26B-A4B-it-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: e5s3z3l7, nrcaz6im; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 645.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Huihui-gemma-4-26B-A4B-it-abliterated-oQ8e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 4k 645.6 45.2 28.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9sbz66tl",
   "source_url": "https://omlx.ai/benchmarks/performance/9sbz66tl",
   "tg_tokens": null,
   "tps": 45.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-9sbz66tl-ctx8192",
   "model": "Huihui-gemma-4-26B-A4B-it-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.2 GB at this context; duplicate traces of the same session: e5s3z3l7, nrcaz6im; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 623.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Huihui-gemma-4-26B-A4B-it-abliterated-oQ8e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 8k 623.0 49.1 28.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9sbz66tl",
   "source_url": "https://omlx.ai/benchmarks/performance/9sbz66tl",
   "tg_tokens": null,
   "tps": 49.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-mre7jzrn-ctx1024",
   "model": "Huihui-gemma-4-31B-it-abliterated-v2",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: h5tffb2c, hu8iqhqw; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 85.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Huihui-gemma-4-31B-it-abliterated-v2-oQ8e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 1k 85.6 20.3 33.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mre7jzrn",
   "source_url": "https://omlx.ai/benchmarks/performance/mre7jzrn",
   "tg_tokens": null,
   "tps": 20.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-mre7jzrn-ctx4096",
   "model": "Huihui-gemma-4-31B-it-abliterated-v2",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.0 GB at this context; duplicate traces of the same session: h5tffb2c, hu8iqhqw; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 84.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Huihui-gemma-4-31B-it-abliterated-v2-oQ8e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 4k 84.4 17.1 35.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mre7jzrn",
   "source_url": "https://omlx.ai/benchmarks/performance/mre7jzrn",
   "tg_tokens": null,
   "tps": 17.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-mre7jzrn-ctx8192",
   "model": "Huihui-gemma-4-31B-it-abliterated-v2",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: h5tffb2c, hu8iqhqw; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 82.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Huihui-gemma-4-31B-it-abliterated-v2-oQ8e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 8k 82.0 17.5 35.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mre7jzrn",
   "source_url": "https://omlx.ai/benchmarks/performance/mre7jzrn",
   "tg_tokens": null,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-hg3i02wp-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 3.3 GB; TTFT 2468 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 415.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k tokens 415.0 PP tok/s 36.3 TG tok/s 2468 TTFT (ms) 3.3 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hg3i02wp",
   "source_url": "https://omlx.ai/benchmarks/performance/hg3i02wp",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-vcly0wxf-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 18qo6l7l; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2502.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,502 93.3 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vcly0wxf",
   "source_url": "https://omlx.ai/benchmarks/performance/vcly0wxf",
   "tg_tokens": null,
   "tps": 93.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-vcly0wxf-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 18qo6l7l; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1837.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,837 89.2 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vcly0wxf",
   "source_url": "https://omlx.ai/benchmarks/performance/vcly0wxf",
   "tg_tokens": null,
   "tps": 89.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-54j65zpq-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: q2spo7h4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4019.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,019 181.5 25.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/54j65zpq",
   "source_url": "https://omlx.ai/benchmarks/performance/54j65zpq",
   "tg_tokens": null,
   "tps": 181.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-54j65zpq-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 27.0 GB at this context; duplicate traces of the same session: q2spo7h4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5964.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 5,964 176.1 27.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/54j65zpq",
   "source_url": "https://omlx.ai/benchmarks/performance/54j65zpq",
   "tg_tokens": null,
   "tps": 176.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-v3bp7uaq-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 25.2 GB at this context; duplicate traces of the same session: 35ache4j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4039.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,039 221.7 25.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v3bp7uaq",
   "source_url": "https://omlx.ai/benchmarks/performance/v3bp7uaq",
   "tg_tokens": null,
   "tps": 221.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-v3bp7uaq-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 27.0 GB at this context; duplicate traces of the same session: 35ache4j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5953.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 5,953 172.7 27.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v3bp7uaq",
   "source_url": "https://omlx.ai/benchmarks/performance/v3bp7uaq",
   "tg_tokens": null,
   "tps": 172.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i34maxhu-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: dwiu5w01; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 923.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 923.3 55.0 19.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i34maxhu",
   "source_url": "https://omlx.ai/benchmarks/performance/i34maxhu",
   "tg_tokens": null,
   "tps": 55.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i34maxhu-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: dwiu5w01; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 956.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 956.8 51.2 21.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i34maxhu",
   "source_url": "https://omlx.ai/benchmarks/performance/i34maxhu",
   "tg_tokens": null,
   "tps": 51.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-d8em5ol9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 1lnuoius, 7rad5fbw, 8785cdny, ch8774t9, epwm2njz, jwlkg270, si7i1l66; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 959.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 959.3 44.2 19.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d8em5ol9",
   "source_url": "https://omlx.ai/benchmarks/performance/d8em5ol9",
   "tg_tokens": null,
   "tps": 44.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-d8em5ol9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 1lnuoius, 7rad5fbw, 8785cdny, ch8774t9, epwm2njz, jwlkg270, si7i1l66; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 712.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 712.6 32.8 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d8em5ol9",
   "source_url": "https://omlx.ai/benchmarks/performance/d8em5ol9",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-d8em5ol9-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 1lnuoius, 7rad5fbw, 8785cdny, ch8774t9, epwm2njz, jwlkg270, si7i1l66; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1006.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,006 32.0 17.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d8em5ol9",
   "source_url": "https://omlx.ai/benchmarks/performance/d8em5ol9",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-d8em5ol9-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 1lnuoius, 7rad5fbw, 8785cdny, ch8774t9, epwm2njz, jwlkg270, si7i1l66; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 976.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 976.0 30.9 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d8em5ol9",
   "source_url": "https://omlx.ai/benchmarks/performance/d8em5ol9",
   "tg_tokens": null,
   "tps": 30.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-d8em5ol9-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 1lnuoius, 7rad5fbw, 8785cdny, ch8774t9, epwm2njz, jwlkg270, si7i1l66; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 911.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 911.4 29.2 18.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d8em5ol9",
   "source_url": "https://omlx.ai/benchmarks/performance/d8em5ol9",
   "tg_tokens": null,
   "tps": 29.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-d8em5ol9-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 1lnuoius, 7rad5fbw, 8785cdny, ch8774t9, epwm2njz, jwlkg270, si7i1l66; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 799.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 799.4 26.0 20.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d8em5ol9",
   "source_url": "https://omlx.ai/benchmarks/performance/d8em5ol9",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-d8em5ol9-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: 1lnuoius, 7rad5fbw, 8785cdny, ch8774t9, epwm2njz, jwlkg270, si7i1l66; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 636.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 636.3 21.3 24.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d8em5ol9",
   "source_url": "https://omlx.ai/benchmarks/performance/d8em5ol9",
   "tg_tokens": null,
   "tps": 21.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-d8em5ol9-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.0 GB at this context; duplicate traces of the same session: 1lnuoius, 7rad5fbw, 8785cdny, ch8774t9, epwm2njz, jwlkg270, si7i1l66; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 505.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 195k 505.3 16.3 29.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d8em5ol9",
   "source_url": "https://omlx.ai/benchmarks/performance/d8em5ol9",
   "tg_tokens": null,
   "tps": 16.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-ghoh3g8r-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 28.5 GB; TTFT 6361 ms; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 161.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 8bit; 1k tokens 161.0 PP tok/s 18.5 TG tok/s 6361 TTFT (ms) 28.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ghoh3g8r",
   "source_url": "https://omlx.ai/benchmarks/performance/ghoh3g8r",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6z19ruq9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 3k68rr8l, 678uia74, a2ez38kw, rafs0ih0, vzinyoud; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 791.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 791.0 18.3 28.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6z19ruq9",
   "source_url": "https://omlx.ai/benchmarks/performance/6z19ruq9",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6z19ruq9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: 3k68rr8l, 678uia74, a2ez38kw, rafs0ih0, vzinyoud; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 935.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 935.0 18.6 29.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6z19ruq9",
   "source_url": "https://omlx.ai/benchmarks/performance/6z19ruq9",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6z19ruq9-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.8 GB at this context; duplicate traces of the same session: 3k68rr8l, 678uia74, a2ez38kw, rafs0ih0, vzinyoud; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 955.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 8k 955.0 18.0 29.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6z19ruq9",
   "source_url": "https://omlx.ai/benchmarks/performance/6z19ruq9",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6z19ruq9-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: 3k68rr8l, 678uia74, a2ez38kw, rafs0ih0, vzinyoud; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 931.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 16k 931.1 18.0 30.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6z19ruq9",
   "source_url": "https://omlx.ai/benchmarks/performance/6z19ruq9",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6z19ruq9-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.3 GB at this context; duplicate traces of the same session: 3k68rr8l, 678uia74, a2ez38kw, rafs0ih0, vzinyoud; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 869.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 32k 869.7 17.3 31.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6z19ruq9",
   "source_url": "https://omlx.ai/benchmarks/performance/6z19ruq9",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6z19ruq9-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.3 GB at this context; duplicate traces of the same session: 3k68rr8l, 678uia74, a2ez38kw, rafs0ih0, vzinyoud; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 768.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 64k 768.1 16.2 33.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6z19ruq9",
   "source_url": "https://omlx.ai/benchmarks/performance/6z19ruq9",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wk8vbg9c-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: akvumnp7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 820.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 820.0 40.9 24.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wk8vbg9c",
   "source_url": "https://omlx.ai/benchmarks/performance/wk8vbg9c",
   "tg_tokens": null,
   "tps": 40.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wk8vbg9c-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: akvumnp7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 946.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 946.0 41.5 25.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wk8vbg9c",
   "source_url": "https://omlx.ai/benchmarks/performance/wk8vbg9c",
   "tg_tokens": null,
   "tps": 41.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-oqxe94kd-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: da37o2yy; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 407.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 407.0 18.1 16.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oqxe94kd",
   "source_url": "https://omlx.ai/benchmarks/performance/oqxe94kd",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-oqxe94kd-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: da37o2yy; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 426.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 426.5 17.7 17.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oqxe94kd",
   "source_url": "https://omlx.ai/benchmarks/performance/oqxe94kd",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-sbat6bns-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 14.1 GB at this context; duplicate traces of the same session: w0pobphc, y704fh4h; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 99.7,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-27B-oQ2.5e-mtp on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 2bit; 1k 99.7 11.8 14.1 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sbat6bns",
   "source_url": "https://omlx.ai/benchmarks/performance/sbat6bns",
   "tg_tokens": null,
   "tps": 11.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-sbat6bns-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 15.1 GB at this context; duplicate traces of the same session: w0pobphc, y704fh4h; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.5,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-27B-oQ2.5e-mtp on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 2bit; 4k 103.5 11.8 15.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sbat6bns",
   "source_url": "https://omlx.ai/benchmarks/performance/sbat6bns",
   "tg_tokens": null,
   "tps": 11.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-sbat6bns-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: w0pobphc, y704fh4h; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 99.6,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-27B-oQ2.5e-mtp on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 2bit; 16k 99.6 13.7 16.3 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sbat6bns",
   "source_url": "https://omlx.ai/benchmarks/performance/sbat6bns",
   "tg_tokens": null,
   "tps": 13.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-fj9x10nx-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 221.8,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 1k 221.8 28.3 16.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fj9x10nx",
   "source_url": "https://omlx.ai/benchmarks/performance/fj9x10nx",
   "tg_tokens": null,
   "tps": 28.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-fj9x10nx-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 233.2,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 4k 233.2 27.5 20.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fj9x10nx",
   "source_url": "https://omlx.ai/benchmarks/performance/fj9x10nx",
   "tg_tokens": null,
   "tps": 27.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-6subh1vf-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: ehazxojx; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 323.0 45.5 23.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6subh1vf",
   "source_url": "https://omlx.ai/benchmarks/performance/6subh1vf",
   "tg_tokens": null,
   "tps": 45.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-6subh1vf-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: ehazxojx; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 303.5,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 16k 303.5 44.0 25.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6subh1vf",
   "source_url": "https://omlx.ai/benchmarks/performance/6subh1vf",
   "tg_tokens": null,
   "tps": 44.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-u97i9vqs-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 31.8 GB; TTFT 88488 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 370.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 32k tokens 370.3 PP tok/s 42.1 TG tok/s 88488 TTFT (ms) 31.8 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u97i9vqs",
   "source_url": "https://omlx.ai/benchmarks/performance/u97i9vqs",
   "tg_tokens": null,
   "tps": 42.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-5fs6o1pe-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 8bit; 1k 224.9 18.4 28.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5fs6o1pe",
   "source_url": "https://omlx.ai/benchmarks/performance/5fs6o1pe",
   "tg_tokens": null,
   "tps": 18.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-5fs6o1pe-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 227.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 8bit; 4k 227.9 18.0 32.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5fs6o1pe",
   "source_url": "https://omlx.ai/benchmarks/performance/5fs6o1pe",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-08im9aid-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 74.1 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 287.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 1k 287.9 18.6 74.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/08im9aid",
   "source_url": "https://omlx.ai/benchmarks/performance/08im9aid",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-08im9aid-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 81.3 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 284.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 4k 284.3 14.9 81.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/08im9aid",
   "source_url": "https://omlx.ai/benchmarks/performance/08im9aid",
   "tg_tokens": null,
   "tps": 14.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-5kxgdgfh-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.3 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 283.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 1k 283.4 39.5 69.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5kxgdgfh",
   "source_url": "https://omlx.ai/benchmarks/performance/5kxgdgfh",
   "tg_tokens": null,
   "tps": 39.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-5kxgdgfh-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.9 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 312.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 4k 312.7 37.7 69.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5kxgdgfh",
   "source_url": "https://omlx.ai/benchmarks/performance/5kxgdgfh",
   "tg_tokens": null,
   "tps": 37.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-276l2300-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 22.8 GB; TTFT 86534 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 378.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k tokens 378.7 PP tok/s 74.4 TG tok/s 86534 TTFT (ms) 22.8 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/276l2300",
   "source_url": "https://omlx.ai/benchmarks/performance/276l2300",
   "tg_tokens": null,
   "tps": 74.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-4mlcfd2o-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 28.3 GB; TTFT 91332 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 358.8,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-bf16-mtp-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 5bit; 32k tokens 358.8 PP tok/s 54.7 TG tok/s 91332 TTFT (ms) 28.3 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4mlcfd2o",
   "source_url": "https://omlx.ai/benchmarks/performance/4mlcfd2o",
   "tg_tokens": null,
   "tps": 54.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rovx903o-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift1.5 build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 82.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 150.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 150.2 44.7 82.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rovx903o",
   "source_url": "https://omlx.ai/benchmarks/performance/rovx903o",
   "tg_tokens": null,
   "tps": 44.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rovx903o-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift1.5 build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 83.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 962.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 962.0 47.5 83.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rovx903o",
   "source_url": "https://omlx.ai/benchmarks/performance/rovx903o",
   "tg_tokens": null,
   "tps": 47.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rovx903o-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift1.5 build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 85.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1384.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 1,384 35.0 85.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rovx903o",
   "source_url": "https://omlx.ai/benchmarks/performance/rovx903o",
   "tg_tokens": null,
   "tps": 35.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-k8h71s8l-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift1.5 build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 86.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1077.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 1,077 72.3 86.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k8h71s8l",
   "source_url": "https://omlx.ai/benchmarks/performance/k8h71s8l",
   "tg_tokens": null,
   "tps": 72.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-k8h71s8l-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift1.5 build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 87.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2342.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 2,342 79.2 87.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k8h71s8l",
   "source_url": "https://omlx.ai/benchmarks/performance/k8h71s8l",
   "tg_tokens": null,
   "tps": 79.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-32he9nvq-ctx1024",
   "model": "gemma-3-4b-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 3.2 GB at this context; duplicate traces of the same session: qpp7er5x; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 358.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-3-4b-it-qat-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 1k 358.9 51.7 3.2 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/32he9nvq",
   "source_url": "https://omlx.ai/benchmarks/performance/32he9nvq",
   "tg_tokens": null,
   "tps": 51.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-32he9nvq-ctx4096",
   "model": "gemma-3-4b-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 3.8 GB at this context; duplicate traces of the same session: qpp7er5x; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 349.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-3-4b-it-qat-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 4k 349.3 50.8 3.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/32he9nvq",
   "source_url": "https://omlx.ai/benchmarks/performance/32he9nvq",
   "tg_tokens": null,
   "tps": 50.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-4brt7cr1-ctx1024",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: sigu2fke, yhsm2vnc; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 85.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-31B-it-oQ8e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 1k 85.2 15.5 33.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4brt7cr1",
   "source_url": "https://omlx.ai/benchmarks/performance/4brt7cr1",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-4brt7cr1-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.0 GB at this context; duplicate traces of the same session: sigu2fke, yhsm2vnc; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 83.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-31B-it-oQ8e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 4k 83.8 17.9 35.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4brt7cr1",
   "source_url": "https://omlx.ai/benchmarks/performance/4brt7cr1",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-4brt7cr1-ctx8192",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: sigu2fke, yhsm2vnc; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 81.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-31B-it-oQ8e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 8k 81.2 14.6 35.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4brt7cr1",
   "source_url": "https://omlx.ai/benchmarks/performance/4brt7cr1",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-ycln1evr-ctx32768",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 11.9 GB; TTFT 105946 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 309.3,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "gpt-oss-20b-OptiQ-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 32k tokens 309.3 PP tok/s 25.6 TG tok/s 105946 TTFT (ms) 11.9 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ycln1evr",
   "source_url": "https://omlx.ai/benchmarks/performance/ycln1evr",
   "tg_tokens": null,
   "tps": 25.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ti4yxm17-ctx1024",
   "model": "granite-4.2-8b",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 9.5 GB at this context; duplicate traces of the same session: kjgrxnjb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 796.7,
   "provenance": "community",
   "quant": "q8",
   "quote": "granite-4.2-8b-q8-mlx on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 796.7 49.4 9.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ti4yxm17",
   "source_url": "https://omlx.ai/benchmarks/performance/ti4yxm17",
   "tg_tokens": null,
   "tps": 49.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ti4yxm17-ctx4096",
   "model": "granite-4.2-8b",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 9.9 GB at this context; duplicate traces of the same session: kjgrxnjb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 771.7,
   "provenance": "community",
   "quant": "q8",
   "quote": "granite-4.2-8b-q8-mlx on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 771.7 46.9 9.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ti4yxm17",
   "source_url": "https://omlx.ai/benchmarks/performance/ti4yxm17",
   "tg_tokens": null,
   "tps": 46.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-1289nl5r-ctx1024",
   "model": "Qwen3.8-27B-aeon-ultimate-uncensored",
   "notes": "vision MTP build; oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: mybo64p9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.5,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "qwen3.8-27b-aeon-ultimate-uncensored-mxfp4-vision-mtplx on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 204.5 28.6 16.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1289nl5r",
   "source_url": "https://omlx.ai/benchmarks/performance/1289nl5r",
   "tg_tokens": null,
   "tps": 28.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-1289nl5r-ctx4096",
   "model": "Qwen3.8-27B-aeon-ultimate-uncensored",
   "notes": "vision MTP build; oMLX community benchmark trace; peak memory 18.0 GB at this context; duplicate traces of the same session: mybo64p9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 194.3,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "qwen3.8-27b-aeon-ultimate-uncensored-mxfp4-vision-mtplx on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 194.3 30.3 18.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1289nl5r",
   "source_url": "https://omlx.ai/benchmarks/performance/1289nl5r",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-8f03nn1a-ctx1024",
   "model": "GLM-4.7-Flash",
   "notes": "REAP-23B-A3B pruned variant; oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: 5x786duc; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 833.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "GLM-4.7-Flash-REAP-23B-A3B-oQ5 on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 5bit; 1k 833.0 36.7 15.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8f03nn1a",
   "source_url": "https://omlx.ai/benchmarks/performance/8f03nn1a",
   "tg_tokens": null,
   "tps": 36.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-8f03nn1a-ctx4096",
   "model": "GLM-4.7-Flash",
   "notes": "REAP-23B-A3B pruned variant; oMLX community benchmark trace; peak memory 16.8 GB at this context; duplicate traces of the same session: 5x786duc; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 613.5,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "GLM-4.7-Flash-REAP-23B-A3B-oQ5 on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 5bit; 4k 613.5 31.4 16.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8f03nn1a",
   "source_url": "https://omlx.ai/benchmarks/performance/8f03nn1a",
   "tg_tokens": null,
   "tps": 31.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-a659c02y-ctx1024",
   "model": "Huihui-Qwen3.8-27B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: be1mn1ua, djaucjaz; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Huihui-Qwen3.8-27B-abliterated-oQ8e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 1k 102.9 20.9 33.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a659c02y",
   "source_url": "https://omlx.ai/benchmarks/performance/a659c02y",
   "tg_tokens": null,
   "tps": 20.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-a659c02y-ctx4096",
   "model": "Huihui-Qwen3.8-27B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.5 GB at this context; duplicate traces of the same session: be1mn1ua, djaucjaz; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 100.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Huihui-Qwen3.8-27B-abliterated-oQ8e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 4k 100.9 16.2 37.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a659c02y",
   "source_url": "https://omlx.ai/benchmarks/performance/a659c02y",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-a659c02y-ctx8192",
   "model": "Huihui-Qwen3.8-27B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.3 GB at this context; duplicate traces of the same session: be1mn1ua, djaucjaz; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 99.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Huihui-Qwen3.8-27B-abliterated-oQ8e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 8k 99.0 23.5 38.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a659c02y",
   "source_url": "https://omlx.ai/benchmarks/performance/a659c02y",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qhd8ud2u-ctx199680",
   "model": "Ornith-1.5-35B-A3B-uncensored",
   "notes": "fixed MTP vision MLX build; oMLX community benchmark trace, single run at 195k context; peak memory 31.9 GB; TTFT 118278 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1691.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-35B-A3B-uncensored-MLX-6bit-fixed-mtp-vision on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 195k tokens 1,691 PP tok/s 52.8 TG tok/s 118278 TTFT (ms) 31.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qhd8ud2u",
   "source_url": "https://omlx.ai/benchmarks/performance/qhd8ud2u",
   "tg_tokens": null,
   "tps": 52.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-mvv1k2iv-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 7.6 GB; TTFT 4567 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.2,
   "provenance": "community",
   "quant": "OptiQ",
   "quote": "Qwen3.5-9B-OptiQ on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k tokens 224.2 PP tok/s 14.2 TG tok/s 4567 TTFT (ms) 7.6 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mvv1k2iv",
   "source_url": "https://omlx.ai/benchmarks/performance/mvv1k2iv",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-e99byxsi-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 3fxqqnut; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 510.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 510.2 69.2 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e99byxsi",
   "source_url": "https://omlx.ai/benchmarks/performance/e99byxsi",
   "tg_tokens": null,
   "tps": 69.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-e99byxsi-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 3fxqqnut; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 565.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 565.5 67.3 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e99byxsi",
   "source_url": "https://omlx.ai/benchmarks/performance/e99byxsi",
   "tg_tokens": null,
   "tps": 67.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-y2mhoxjm-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 89aumerx, 9sokk9n6, dx4iv3fw; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2696.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 2,696 161.7 21.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y2mhoxjm",
   "source_url": "https://omlx.ai/benchmarks/performance/y2mhoxjm",
   "tg_tokens": null,
   "tps": 161.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-y2mhoxjm-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: 89aumerx, 9sokk9n6, dx4iv3fw; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3108.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 3,108 97.6 23.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y2mhoxjm",
   "source_url": "https://omlx.ai/benchmarks/performance/y2mhoxjm",
   "tg_tokens": null,
   "tps": 97.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-y2mhoxjm-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: 89aumerx, 9sokk9n6, dx4iv3fw; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2795.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 2,795 111.1 23.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y2mhoxjm",
   "source_url": "https://omlx.ai/benchmarks/performance/y2mhoxjm",
   "tg_tokens": null,
   "tps": 111.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-y2mhoxjm-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: 89aumerx, 9sokk9n6, dx4iv3fw; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2307.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 2,307 118.5 24.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y2mhoxjm",
   "source_url": "https://omlx.ai/benchmarks/performance/y2mhoxjm",
   "tg_tokens": null,
   "tps": 118.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-flqq1ro4-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 28.5 GB; TTFT 6367 ms; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 160.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 8bit; 1k tokens 160.8 PP tok/s 18.5 TG tok/s 6367 TTFT (ms) 28.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/flqq1ro4",
   "source_url": "https://omlx.ai/benchmarks/performance/flqq1ro4",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-qirnhiea-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: tve6zgrv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 129.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 4k 129.0 8.4 29.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qirnhiea",
   "source_url": "https://omlx.ai/benchmarks/performance/qirnhiea",
   "tg_tokens": null,
   "tps": 8.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-qirnhiea-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 30.6 GB at this context; duplicate traces of the same session: tve6zgrv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 118.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 16k 118.4 8.2 30.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qirnhiea",
   "source_url": "https://omlx.ai/benchmarks/performance/qirnhiea",
   "tg_tokens": null,
   "tps": 8.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-r5fr3dn1-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 221.8,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 1k 221.8 28.3 16.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r5fr3dn1",
   "source_url": "https://omlx.ai/benchmarks/performance/r5fr3dn1",
   "tg_tokens": null,
   "tps": 28.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-r5fr3dn1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 233.2,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 4k 233.2 27.5 20.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r5fr3dn1",
   "source_url": "https://omlx.ai/benchmarks/performance/r5fr3dn1",
   "tg_tokens": null,
   "tps": 27.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-w1lfcu5j-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: 1sfua4bp, 3obqa6bk, 52e10r86, mbyzcd96, p0u3dxbn, r2lieenj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 429.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 429.5 30.2 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w1lfcu5j",
   "source_url": "https://omlx.ai/benchmarks/performance/w1lfcu5j",
   "tg_tokens": null,
   "tps": 30.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-w1lfcu5j-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: 1sfua4bp, 3obqa6bk, 52e10r86, mbyzcd96, p0u3dxbn, r2lieenj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 461.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 461.4 29.9 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w1lfcu5j",
   "source_url": "https://omlx.ai/benchmarks/performance/w1lfcu5j",
   "tg_tokens": null,
   "tps": 29.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-w1lfcu5j-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: 1sfua4bp, 3obqa6bk, 52e10r86, mbyzcd96, p0u3dxbn, r2lieenj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 431.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 8k 431.9 17.1 28.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w1lfcu5j",
   "source_url": "https://omlx.ai/benchmarks/performance/w1lfcu5j",
   "tg_tokens": null,
   "tps": 17.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-w1lfcu5j-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 1sfua4bp, 3obqa6bk, 52e10r86, mbyzcd96, p0u3dxbn, r2lieenj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 396.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 16k 396.6 18.2 28.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w1lfcu5j",
   "source_url": "https://omlx.ai/benchmarks/performance/w1lfcu5j",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-w1lfcu5j-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: 1sfua4bp, 3obqa6bk, 52e10r86, mbyzcd96, p0u3dxbn, r2lieenj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 355.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 32k 355.4 16.8 29.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w1lfcu5j",
   "source_url": "https://omlx.ai/benchmarks/performance/w1lfcu5j",
   "tg_tokens": null,
   "tps": 16.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-w1lfcu5j-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: 1sfua4bp, 3obqa6bk, 52e10r86, mbyzcd96, p0u3dxbn, r2lieenj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 280.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 64k 280.4 12.5 32.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w1lfcu5j",
   "source_url": "https://omlx.ai/benchmarks/performance/w1lfcu5j",
   "tg_tokens": null,
   "tps": 12.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-w1lfcu5j-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.0 GB at this context; duplicate traces of the same session: 1sfua4bp, 3obqa6bk, 52e10r86, mbyzcd96, p0u3dxbn, r2lieenj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 213.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 128k 213.2 7.3 37.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w1lfcu5j",
   "source_url": "https://omlx.ai/benchmarks/performance/w1lfcu5j",
   "tg_tokens": null,
   "tps": 7.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-52cwjau2-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 8bit; 1k 224.9 18.4 28.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/52cwjau2",
   "source_url": "https://omlx.ai/benchmarks/performance/52cwjau2",
   "tg_tokens": null,
   "tps": 18.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-52cwjau2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 227.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 8bit; 4k 227.9 18.0 32.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/52cwjau2",
   "source_url": "https://omlx.ai/benchmarks/performance/52cwjau2",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2f22p6iv-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 16yeml7v, 5p9bc72f, 85srq3sb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 256.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 256.9 16.6 28.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2f22p6iv",
   "source_url": "https://omlx.ai/benchmarks/performance/2f22p6iv",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2f22p6iv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.2 GB at this context; duplicate traces of the same session: 16yeml7v, 5p9bc72f, 85srq3sb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 259.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 259.6 16.4 32.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2f22p6iv",
   "source_url": "https://omlx.ai/benchmarks/performance/2f22p6iv",
   "tg_tokens": null,
   "tps": 16.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2f22p6iv-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.2 GB at this context; duplicate traces of the same session: 16yeml7v, 5p9bc72f, 85srq3sb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 256.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 256.3 16.1 33.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2f22p6iv",
   "source_url": "https://omlx.ai/benchmarks/performance/2f22p6iv",
   "tg_tokens": null,
   "tps": 16.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2f22p6iv-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.4 GB at this context; duplicate traces of the same session: 16yeml7v, 5p9bc72f, 85srq3sb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 247.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 247.8 15.7 32.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2f22p6iv",
   "source_url": "https://omlx.ai/benchmarks/performance/2f22p6iv",
   "tg_tokens": null,
   "tps": 15.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2f22p6iv-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.4 GB at this context; duplicate traces of the same session: 16yeml7v, 5p9bc72f, 85srq3sb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 230.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 230.3 15.1 33.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2f22p6iv",
   "source_url": "https://omlx.ai/benchmarks/performance/2f22p6iv",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2f22p6iv-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.4 GB at this context; duplicate traces of the same session: 16yeml7v, 5p9bc72f, 85srq3sb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 202.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 202.1 13.9 35.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2f22p6iv",
   "source_url": "https://omlx.ai/benchmarks/performance/2f22p6iv",
   "tg_tokens": null,
   "tps": 13.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-2f22p6iv-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.4 GB at this context; duplicate traces of the same session: 16yeml7v, 5p9bc72f, 85srq3sb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 162.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 162.0 12.1 39.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2f22p6iv",
   "source_url": "https://omlx.ai/benchmarks/performance/2f22p6iv",
   "tg_tokens": null,
   "tps": 12.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-yr6rpquv-ctx65536",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 76.4 GB; TTFT 28718 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2282.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k tokens 2,282 PP tok/s 66.8 TG tok/s 28718 TTFT (ms) 76.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yr6rpquv",
   "source_url": "https://omlx.ai/benchmarks/performance/yr6rpquv",
   "tg_tokens": null,
   "tps": 66.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-zohob11w-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 74.1 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 287.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 1k 287.9 18.6 74.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zohob11w",
   "source_url": "https://omlx.ai/benchmarks/performance/zohob11w",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-zohob11w-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 81.3 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 284.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 4k 284.3 14.9 81.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zohob11w",
   "source_url": "https://omlx.ai/benchmarks/performance/zohob11w",
   "tg_tokens": null,
   "tps": 14.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-avtks0uk-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 72.6 GB at this context; duplicate traces of the same session: indsfb3u; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 339.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 1k 339.4 11.9 72.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/avtks0uk",
   "source_url": "https://omlx.ai/benchmarks/performance/avtks0uk",
   "tg_tokens": null,
   "tps": 11.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-avtks0uk-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 79.3 GB at this context; duplicate traces of the same session: indsfb3u; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 349.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 4k 349.6 11.4 79.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/avtks0uk",
   "source_url": "https://omlx.ai/benchmarks/performance/avtks0uk",
   "tg_tokens": null,
   "tps": 11.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-d4iashxa-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.3 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 283.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 1k 283.4 39.5 69.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d4iashxa",
   "source_url": "https://omlx.ai/benchmarks/performance/d4iashxa",
   "tg_tokens": null,
   "tps": 39.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-d4iashxa-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.9 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 312.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 4k 312.7 37.7 69.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d4iashxa",
   "source_url": "https://omlx.ai/benchmarks/performance/d4iashxa",
   "tg_tokens": null,
   "tps": 37.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6pvt3pcq-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.6 GB; TTFT 991 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1033.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,033 PP tok/s 71.0 TG tok/s 991 TTFT (ms) 70.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6pvt3pcq",
   "source_url": "https://omlx.ai/benchmarks/performance/6pvt3pcq",
   "tg_tokens": null,
   "tps": 71.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-lvm5t57e-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 29.7 GB; TTFT 65211 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 502.5,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-bf16-mtp-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 5bit; 32k tokens 502.5 PP tok/s 57.9 TG tok/s 65211 TTFT (ms) 29.7 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lvm5t57e",
   "source_url": "https://omlx.ai/benchmarks/performance/lvm5t57e",
   "tg_tokens": null,
   "tps": 57.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-tdt1hn24-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "bf16; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 31.7 GB; TTFT 81045 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 404.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-bf16-mtp-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 32k tokens 404.3 PP tok/s 48.7 TG tok/s 81045 TTFT (ms) 31.7 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tdt1hn24",
   "source_url": "https://omlx.ai/benchmarks/performance/tdt1hn24",
   "tg_tokens": null,
   "tps": 48.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-41n2doah-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "bf16; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 31.7 GB; TTFT 81874 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 400.2,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-bf16-mtp-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 32k tokens 400.2 PP tok/s 50.2 TG tok/s 81874 TTFT (ms) 31.7 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/41n2doah",
   "source_url": "https://omlx.ai/benchmarks/performance/41n2doah",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-374fole2-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "bf16; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 31.7 GB; TTFT 88470 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 370.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-bf16-mtp-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 32k tokens 370.4 PP tok/s 44.7 TG tok/s 88470 TTFT (ms) 31.7 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/374fole2",
   "source_url": "https://omlx.ai/benchmarks/performance/374fole2",
   "tg_tokens": null,
   "tps": 44.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-h5iz8pmw-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "bf16; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 31.7 GB; TTFT 87812 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 373.2,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-bf16-mtp-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 32k tokens 373.2 PP tok/s 43.7 TG tok/s 87812 TTFT (ms) 31.7 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h5iz8pmw",
   "source_url": "https://omlx.ai/benchmarks/performance/h5iz8pmw",
   "tg_tokens": null,
   "tps": 43.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-mlx3zuqw-ctx1024",
   "model": "Gemma 4 26B-A4B-it-uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: bbiv0zx8, epyoirre, y8iaoza3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1889.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 1,889 89.6 25.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mlx3zuqw",
   "source_url": "https://omlx.ai/benchmarks/performance/mlx3zuqw",
   "tg_tokens": null,
   "tps": 89.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-mlx3zuqw-ctx4096",
   "model": "Gemma 4 26B-A4B-it-uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: bbiv0zx8, epyoirre, y8iaoza3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1877.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 1,877 89.1 26.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mlx3zuqw",
   "source_url": "https://omlx.ai/benchmarks/performance/mlx3zuqw",
   "tg_tokens": null,
   "tps": 89.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-mlx3zuqw-ctx8192",
   "model": "Gemma 4 26B-A4B-it-uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: bbiv0zx8, epyoirre, y8iaoza3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1812.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 8k 1,812 87.2 26.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mlx3zuqw",
   "source_url": "https://omlx.ai/benchmarks/performance/mlx3zuqw",
   "tg_tokens": null,
   "tps": 87.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-mlx3zuqw-ctx16384",
   "model": "Gemma 4 26B-A4B-it-uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: bbiv0zx8, epyoirre, y8iaoza3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1708.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 16k 1,708 83.3 26.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mlx3zuqw",
   "source_url": "https://omlx.ai/benchmarks/performance/mlx3zuqw",
   "tg_tokens": null,
   "tps": 83.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-mutyujhb-ctx1024",
   "model": "gemma-4-31B-it-uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: cgw250lt, gxgamdot, ussttkhw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 259.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "gemma-4-31B-it-uncensored-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 1k 259.0 44.2 24.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mutyujhb",
   "source_url": "https://omlx.ai/benchmarks/performance/mutyujhb",
   "tg_tokens": null,
   "tps": 44.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-mutyujhb-ctx4096",
   "model": "gemma-4-31B-it-uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.2 GB at this context; duplicate traces of the same session: cgw250lt, gxgamdot, ussttkhw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 257.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "gemma-4-31B-it-uncensored-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 4k 257.2 38.5 25.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mutyujhb",
   "source_url": "https://omlx.ai/benchmarks/performance/mutyujhb",
   "tg_tokens": null,
   "tps": 38.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-mutyujhb-ctx8192",
   "model": "gemma-4-31B-it-uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: cgw250lt, gxgamdot, ussttkhw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 251.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "gemma-4-31B-it-uncensored-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 8k 251.9 61.3 25.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mutyujhb",
   "source_url": "https://omlx.ai/benchmarks/performance/mutyujhb",
   "tg_tokens": null,
   "tps": 61.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-mutyujhb-ctx16384",
   "model": "gemma-4-31B-it-uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: cgw250lt, gxgamdot, ussttkhw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 242.6,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "gemma-4-31B-it-uncensored-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 16k 242.6 37.5 26.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mutyujhb",
   "source_url": "https://omlx.ai/benchmarks/performance/mutyujhb",
   "tg_tokens": null,
   "tps": 37.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vnla8u8q-ctx1024",
   "model": "gemma-4-31B-it-uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.6 GB at this context; duplicate traces of the same session: 7smimtkh, oohy49x1, zuklrgo0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 260.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-31B-it-uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 260.5 37.6 34.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vnla8u8q",
   "source_url": "https://omlx.ai/benchmarks/performance/vnla8u8q",
   "tg_tokens": null,
   "tps": 37.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-vnla8u8q-ctx4096",
   "model": "gemma-4-31B-it-uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.1 GB at this context; duplicate traces of the same session: 7smimtkh, oohy49x1, zuklrgo0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 262.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-31B-it-uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 262.0 38.8 35.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vnla8u8q",
   "source_url": "https://omlx.ai/benchmarks/performance/vnla8u8q",
   "tg_tokens": null,
   "tps": 38.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-vnla8u8q-ctx8192",
   "model": "gemma-4-31B-it-uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: 7smimtkh, oohy49x1, zuklrgo0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 256.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-31B-it-uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 8k 256.6 32.6 35.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vnla8u8q",
   "source_url": "https://omlx.ai/benchmarks/performance/vnla8u8q",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vnla8u8q-ctx16384",
   "model": "gemma-4-31B-it-uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.7 GB at this context; duplicate traces of the same session: 7smimtkh, oohy49x1, zuklrgo0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 246.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-31B-it-uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 16k 246.8 36.3 36.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vnla8u8q",
   "source_url": "https://omlx.ai/benchmarks/performance/vnla8u8q",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-dyswpfnz-ctx1024",
   "model": "gemma-4-31B-it-uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: 39wnqg6b, 4wncogwu, wbimiuc5; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 263.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-31B-it-uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 263.7 18.8 32.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dyswpfnz",
   "source_url": "https://omlx.ai/benchmarks/performance/dyswpfnz",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-dyswpfnz-ctx4096",
   "model": "gemma-4-31B-it-uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.0 GB at this context; duplicate traces of the same session: 39wnqg6b, 4wncogwu, wbimiuc5; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 262.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-31B-it-uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 262.4 18.9 33.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dyswpfnz",
   "source_url": "https://omlx.ai/benchmarks/performance/dyswpfnz",
   "tg_tokens": null,
   "tps": 18.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-dyswpfnz-ctx8192",
   "model": "gemma-4-31B-it-uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.6 GB at this context; duplicate traces of the same session: 39wnqg6b, 4wncogwu, wbimiuc5; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 256.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-31B-it-uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 8k 256.8 18.6 33.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dyswpfnz",
   "source_url": "https://omlx.ai/benchmarks/performance/dyswpfnz",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-dyswpfnz-ctx16384",
   "model": "gemma-4-31B-it-uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: 39wnqg6b, 4wncogwu, wbimiuc5; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 246.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-31B-it-uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 16k 246.9 18.0 34.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dyswpfnz",
   "source_url": "https://omlx.ai/benchmarks/performance/dyswpfnz",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fc14y3l2-ctx131072",
   "model": "Alice-27B",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 37.0 GB; TTFT 583837 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Alice-27B on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 128k tokens 224.5 PP tok/s 8.5 TG tok/s 583837 TTFT (ms) 37.0 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fc14y3l2",
   "source_url": "https://omlx.ai/benchmarks/performance/fc14y3l2",
   "tg_tokens": null,
   "tps": 8.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q4e709h0-ctx16384",
   "model": "Ornith-1.5-35B-A3B-uncensored",
   "notes": "fixed MTP vision MLX build; oMLX community benchmark trace, single run at 16k context; peak memory 28.4 GB; TTFT 3293 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4976.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-35B-A3B-uncensored-MLX-6bit-fixed-mtp-vision on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k tokens 4,976 PP tok/s 100.7 TG tok/s 3293 TTFT (ms) 28.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q4e709h0",
   "source_url": "https://omlx.ai/benchmarks/performance/q4e709h0",
   "tg_tokens": null,
   "tps": 100.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ryc00i5y-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.7 GB at this context; duplicate traces of the same session: 1kemnasi; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 358.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-9B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 358.1 44.3 7.7 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ryc00i5y",
   "source_url": "https://omlx.ai/benchmarks/performance/ryc00i5y",
   "tg_tokens": null,
   "tps": 44.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ryc00i5y-ctx16384",
   "model": "Ornith-1.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 8.2 GB at this context; duplicate traces of the same session: 1kemnasi; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 343.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-9B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 16k 343.7 42.8 8.2 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ryc00i5y",
   "source_url": "https://omlx.ai/benchmarks/performance/ryc00i5y",
   "tg_tokens": null,
   "tps": 42.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-3b1yvkqc-ctx4096",
   "model": "Ornith-1.5-9B-uncensored",
   "notes": "oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: 0pyrmia6; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 382.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-9B-uncensored-oQ4e on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 382.4 45.9 6.5 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3b1yvkqc",
   "source_url": "https://omlx.ai/benchmarks/performance/3b1yvkqc",
   "tg_tokens": null,
   "tps": 45.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-3b1yvkqc-ctx16384",
   "model": "Ornith-1.5-9B-uncensored",
   "notes": "oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: 0pyrmia6; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 365.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-9B-uncensored-oQ4e on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 16k 365.3 41.7 6.9 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3b1yvkqc",
   "source_url": "https://omlx.ai/benchmarks/performance/3b1yvkqc",
   "tg_tokens": null,
   "tps": 41.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-k218oqbl-ctx1024",
   "model": "Qwen3-1.7B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 1.6 GB at this context; duplicate traces of the same session: 73fsv188, segmqmhg; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1039.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-1.7B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 1,039 71.3 1.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k218oqbl",
   "source_url": "https://omlx.ai/benchmarks/performance/k218oqbl",
   "tg_tokens": null,
   "tps": 71.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-k218oqbl-ctx4096",
   "model": "Qwen3-1.7B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 2.0 GB at this context; duplicate traces of the same session: 73fsv188, segmqmhg; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 909.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-1.7B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 909.6 55.3 2.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k218oqbl",
   "source_url": "https://omlx.ai/benchmarks/performance/k218oqbl",
   "tg_tokens": null,
   "tps": 55.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-k218oqbl-ctx8192",
   "model": "Qwen3-1.7B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 2.3 GB at this context; duplicate traces of the same session: 73fsv188, segmqmhg; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 776.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-1.7B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 8k 776.5 41.3 2.3 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k218oqbl",
   "source_url": "https://omlx.ai/benchmarks/performance/k218oqbl",
   "tg_tokens": null,
   "tps": 41.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-8ys83t3e-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.7 GB at this context; duplicate traces of the same session: fzw3s2oq; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 226.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 226.4 20.4 5.7 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8ys83t3e",
   "source_url": "https://omlx.ai/benchmarks/performance/8ys83t3e",
   "tg_tokens": null,
   "tps": 20.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-8ys83t3e-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: fzw3s2oq; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 234.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 234.6 20.0 6.2 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8ys83t3e",
   "source_url": "https://omlx.ai/benchmarks/performance/8ys83t3e",
   "tg_tokens": null,
   "tps": 20.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-fy4mkxr5-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: 1chrffpi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1583.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 1,583 92.9 35.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fy4mkxr5",
   "source_url": "https://omlx.ai/benchmarks/performance/fy4mkxr5",
   "tg_tokens": null,
   "tps": 92.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-fy4mkxr5-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 37.5 GB at this context; duplicate traces of the same session: 1chrffpi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1825.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 1,825 91.0 37.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fy4mkxr5",
   "source_url": "https://omlx.ai/benchmarks/performance/fy4mkxr5",
   "tg_tokens": null,
   "tps": 91.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-tewt29uo-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: n8bm6vvg; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 351.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 351.3 37.9 19.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tewt29uo",
   "source_url": "https://omlx.ai/benchmarks/performance/tewt29uo",
   "tg_tokens": null,
   "tps": 37.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-tewt29uo-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: n8bm6vvg; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 335.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 335.3 35.3 20.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tewt29uo",
   "source_url": "https://omlx.ai/benchmarks/performance/tewt29uo",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-10hxyhxe-ctx16384",
   "model": "Qwen3.8-27B-Coder390-LynnStyle-Mixed",
   "notes": "A4toA8 build; MTPFix; oMLX community benchmark trace, single run at 16k context; peak memory 22.1 GB; TTFT 18430 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 889.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Coder390-LynnStyle-Mixed-MLX-A4toA8-MTPFix on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k tokens 889.0 PP tok/s 98.4 TG tok/s 18430 TTFT (ms) 22.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/10hxyhxe",
   "source_url": "https://omlx.ai/benchmarks/performance/10hxyhxe",
   "tg_tokens": null,
   "tps": 98.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-zp32ud7y-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: qhfge1mm; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 211.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k 211.1 8.4 15.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zp32ud7y",
   "source_url": "https://omlx.ai/benchmarks/performance/zp32ud7y",
   "tg_tokens": null,
   "tps": 8.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-zp32ud7y-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: qhfge1mm; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 212.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 212.5 8.3 17.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zp32ud7y",
   "source_url": "https://omlx.ai/benchmarks/performance/zp32ud7y",
   "tg_tokens": null,
   "tps": 8.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-eq4xngqb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: v19h08n9; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 129.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 1k 129.1 8.6 28.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eq4xngqb",
   "source_url": "https://omlx.ai/benchmarks/performance/eq4xngqb",
   "tg_tokens": null,
   "tps": 8.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-eq4xngqb-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: v19h08n9; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 129.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 4k 129.2 8.4 29.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eq4xngqb",
   "source_url": "https://omlx.ai/benchmarks/performance/eq4xngqb",
   "tg_tokens": null,
   "tps": 8.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-3o08idz7-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-NEO-CODER",
   "notes": "1M context build; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: dd47dwfj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 254.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-NEO-CODER-mlx-oQ8e-1M on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 254.4 16.4 28.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3o08idz7",
   "source_url": "https://omlx.ai/benchmarks/performance/3o08idz7",
   "tg_tokens": null,
   "tps": 16.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-3o08idz7-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-NEO-CODER",
   "notes": "1M context build; oMLX community benchmark trace; peak memory 31.7 GB at this context; duplicate traces of the same session: dd47dwfj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 250.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-NEO-CODER-mlx-oQ8e-1M on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 250.9 16.3 31.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3o08idz7",
   "source_url": "https://omlx.ai/benchmarks/performance/3o08idz7",
   "tg_tokens": null,
   "tps": 16.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-fdfxo0bf-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 131.3 13.8 16.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fdfxo0bf",
   "source_url": "https://omlx.ai/benchmarks/performance/fdfxo0bf",
   "tg_tokens": null,
   "tps": 13.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-fdfxo0bf-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 119.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 119.8 13.3 20.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fdfxo0bf",
   "source_url": "https://omlx.ai/benchmarks/performance/fdfxo0bf",
   "tg_tokens": null,
   "tps": 13.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-2v865d9c-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 126.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 126.2 21.8 18.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2v865d9c",
   "source_url": "https://omlx.ai/benchmarks/performance/2v865d9c",
   "tg_tokens": null,
   "tps": 21.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-2v865d9c-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 115.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 115.8 19.1 22.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2v865d9c",
   "source_url": "https://omlx.ai/benchmarks/performance/2v865d9c",
   "tg_tokens": null,
   "tps": 19.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-7vo4r8e2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: 3kn3siaz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 346.2,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-27B-oQ6 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 4k 346.2 27.3 26.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7vo4r8e2",
   "source_url": "https://omlx.ai/benchmarks/performance/7vo4r8e2",
   "tg_tokens": null,
   "tps": 27.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-7vo4r8e2-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 27.3 GB at this context; duplicate traces of the same session: 3kn3siaz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 330.6,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-27B-oQ6 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 16k 330.6 26.2 27.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7vo4r8e2",
   "source_url": "https://omlx.ai/benchmarks/performance/7vo4r8e2",
   "tg_tokens": null,
   "tps": 26.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-9cs5bnsc-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: r62hn41z, veo0fp3o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 256.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 256.9 16.6 28.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9cs5bnsc",
   "source_url": "https://omlx.ai/benchmarks/performance/9cs5bnsc",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-9cs5bnsc-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.2 GB at this context; duplicate traces of the same session: r62hn41z, veo0fp3o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 259.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 259.6 16.4 32.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9cs5bnsc",
   "source_url": "https://omlx.ai/benchmarks/performance/9cs5bnsc",
   "tg_tokens": null,
   "tps": 16.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-9cs5bnsc-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.2 GB at this context; duplicate traces of the same session: r62hn41z, veo0fp3o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 256.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 256.3 16.1 33.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9cs5bnsc",
   "source_url": "https://omlx.ai/benchmarks/performance/9cs5bnsc",
   "tg_tokens": null,
   "tps": 16.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-9cs5bnsc-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.4 GB at this context; duplicate traces of the same session: r62hn41z, veo0fp3o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 247.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 247.8 15.7 32.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9cs5bnsc",
   "source_url": "https://omlx.ai/benchmarks/performance/9cs5bnsc",
   "tg_tokens": null,
   "tps": 15.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-9cs5bnsc-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.4 GB at this context; duplicate traces of the same session: r62hn41z, veo0fp3o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 230.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 230.3 15.1 33.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9cs5bnsc",
   "source_url": "https://omlx.ai/benchmarks/performance/9cs5bnsc",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-9cs5bnsc-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.4 GB at this context; duplicate traces of the same session: r62hn41z, veo0fp3o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 202.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 202.1 13.9 35.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9cs5bnsc",
   "source_url": "https://omlx.ai/benchmarks/performance/9cs5bnsc",
   "tg_tokens": null,
   "tps": 13.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-9cs5bnsc-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.4 GB at this context; duplicate traces of the same session: r62hn41z, veo0fp3o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 162.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 162.0 12.1 39.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9cs5bnsc",
   "source_url": "https://omlx.ai/benchmarks/performance/9cs5bnsc",
   "tg_tokens": null,
   "tps": 12.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-qh5zvl53-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 37vs2wxa, 80syuiqx, dvr8l5x0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 319.8,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-uncensored-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 1k 319.8 57.1 23.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qh5zvl53",
   "source_url": "https://omlx.ai/benchmarks/performance/qh5zvl53",
   "tg_tokens": null,
   "tps": 57.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-qh5zvl53-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: 37vs2wxa, 80syuiqx, dvr8l5x0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 344.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-uncensored-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 4k 344.1 53.1 26.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qh5zvl53",
   "source_url": "https://omlx.ai/benchmarks/performance/qh5zvl53",
   "tg_tokens": null,
   "tps": 53.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-qh5zvl53-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.3 GB at this context; duplicate traces of the same session: 37vs2wxa, 80syuiqx, dvr8l5x0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 338.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-uncensored-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 8k 338.4 64.3 27.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qh5zvl53",
   "source_url": "https://omlx.ai/benchmarks/performance/qh5zvl53",
   "tg_tokens": null,
   "tps": 64.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-qh5zvl53-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.2 GB at this context; duplicate traces of the same session: 37vs2wxa, 80syuiqx, dvr8l5x0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 325.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-uncensored-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 16k 325.7 62.6 28.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qh5zvl53",
   "source_url": "https://omlx.ai/benchmarks/performance/qh5zvl53",
   "tg_tokens": null,
   "tps": 62.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1yfhl15t-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: f72xv266, gsnv4tcs, jjya9oj7; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 320.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 320.5 55.2 33.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1yfhl15t",
   "source_url": "https://omlx.ai/benchmarks/performance/1yfhl15t",
   "tg_tokens": null,
   "tps": 55.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1yfhl15t-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: f72xv266, gsnv4tcs, jjya9oj7; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 334.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 334.5 48.4 37.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1yfhl15t",
   "source_url": "https://omlx.ai/benchmarks/performance/1yfhl15t",
   "tg_tokens": null,
   "tps": 48.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1yfhl15t-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.8 GB at this context; duplicate traces of the same session: f72xv266, gsnv4tcs, jjya9oj7; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 329.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 8k 329.3 63.9 37.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1yfhl15t",
   "source_url": "https://omlx.ai/benchmarks/performance/1yfhl15t",
   "tg_tokens": null,
   "tps": 63.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1yfhl15t-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.7 GB at this context; duplicate traces of the same session: f72xv266, gsnv4tcs, jjya9oj7; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 317.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 16k 317.5 63.1 38.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1yfhl15t",
   "source_url": "https://omlx.ai/benchmarks/performance/1yfhl15t",
   "tg_tokens": null,
   "tps": 63.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-o57tm400-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift build; oQ4e/oQ8e mixed quant; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 82.8 GB at this context; duplicate traces of the same session: 173h34yg, 1p05bgj9, 3zd1phgq, losw5qsb, q05lz31z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 608.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 608.6 72.9 82.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o57tm400",
   "source_url": "https://omlx.ai/benchmarks/performance/o57tm400",
   "tg_tokens": null,
   "tps": 72.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-o57tm400-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift build; oQ4e/oQ8e mixed quant; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 83.2 GB at this context; duplicate traces of the same session: 173h34yg, 1p05bgj9, 3zd1phgq, losw5qsb, q05lz31z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 710.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 710.8 60.0 83.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o57tm400",
   "source_url": "https://omlx.ai/benchmarks/performance/o57tm400",
   "tg_tokens": null,
   "tps": 60.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-o57tm400-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift build; oQ4e/oQ8e mixed quant; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 83.4 GB at this context; duplicate traces of the same session: 173h34yg, 1p05bgj9, 3zd1phgq, losw5qsb, q05lz31z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 825.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 825.0 69.8 83.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o57tm400",
   "source_url": "https://omlx.ai/benchmarks/performance/o57tm400",
   "tg_tokens": null,
   "tps": 69.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-o57tm400-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift build; oQ4e/oQ8e mixed quant; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 83.7 GB at this context; duplicate traces of the same session: 173h34yg, 1p05bgj9, 3zd1phgq, losw5qsb, q05lz31z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 818.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 818.8 64.6 83.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o57tm400",
   "source_url": "https://omlx.ai/benchmarks/performance/o57tm400",
   "tg_tokens": null,
   "tps": 64.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-o57tm400-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift build; oQ4e/oQ8e mixed quant; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 84.5 GB at this context; duplicate traces of the same session: 173h34yg, 1p05bgj9, 3zd1phgq, losw5qsb, q05lz31z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 812.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 812.8 63.7 84.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o57tm400",
   "source_url": "https://omlx.ai/benchmarks/performance/o57tm400",
   "tg_tokens": null,
   "tps": 63.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-o57tm400-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift build; oQ4e/oQ8e mixed quant; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 86.6 GB at this context; duplicate traces of the same session: 173h34yg, 1p05bgj9, 3zd1phgq, losw5qsb, q05lz31z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 802.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 802.3 57.8 86.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o57tm400",
   "source_url": "https://omlx.ai/benchmarks/performance/o57tm400",
   "tg_tokens": null,
   "tps": 57.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-dw20igs6-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift build; oQ4e/oQ8e mixed quant; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 82.6 GB at this context; duplicate traces of the same session: niwhg4n5, rx61ypq1, v4awrk4q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 548.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 548.2 56.1 82.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dw20igs6",
   "source_url": "https://omlx.ai/benchmarks/performance/dw20igs6",
   "tg_tokens": null,
   "tps": 56.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-dw20igs6-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift build; oQ4e/oQ8e mixed quant; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 83.2 GB at this context; duplicate traces of the same session: niwhg4n5, rx61ypq1, v4awrk4q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 607.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 607.6 52.9 83.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dw20igs6",
   "source_url": "https://omlx.ai/benchmarks/performance/dw20igs6",
   "tg_tokens": null,
   "tps": 52.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-dw20igs6-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift build; oQ4e/oQ8e mixed quant; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 83.4 GB at this context; duplicate traces of the same session: niwhg4n5, rx61ypq1, v4awrk4q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 822.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 822.3 59.7 83.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dw20igs6",
   "source_url": "https://omlx.ai/benchmarks/performance/dw20igs6",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-dw20igs6-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift build; oQ4e/oQ8e mixed quant; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 86.3 GB at this context; duplicate traces of the same session: niwhg4n5, rx61ypq1, v4awrk4q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 802.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 802.2 55.9 86.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dw20igs6",
   "source_url": "https://omlx.ai/benchmarks/performance/dw20igs6",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-quica021-ctx1024",
   "model": "gemma-3-4b-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 3.2 GB at this context; duplicate traces of the same session: agwcmae0, alht03zm, m2x8a8er; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 358.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-3-4b-it-qat-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 1k 358.3 52.0 3.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/quica021",
   "source_url": "https://omlx.ai/benchmarks/performance/quica021",
   "tg_tokens": null,
   "tps": 52.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-quica021-ctx4096",
   "model": "gemma-3-4b-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 3.8 GB at this context; duplicate traces of the same session: agwcmae0, alht03zm, m2x8a8er; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 352.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-3-4b-it-qat-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 4k 352.4 51.2 3.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/quica021",
   "source_url": "https://omlx.ai/benchmarks/performance/quica021",
   "tg_tokens": null,
   "tps": 51.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-quica021-ctx8192",
   "model": "gemma-3-4b-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 4.3 GB at this context; duplicate traces of the same session: agwcmae0, alht03zm, m2x8a8er; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 343.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-3-4b-it-qat-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 8k 343.2 49.7 4.3 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/quica021",
   "source_url": "https://omlx.ai/benchmarks/performance/quica021",
   "tg_tokens": null,
   "tps": 49.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-quica021-ctx16384",
   "model": "gemma-3-4b-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 5.4 GB at this context; duplicate traces of the same session: agwcmae0, alht03zm, m2x8a8er; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 326.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-3-4b-it-qat-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 16k 326.4 47.4 5.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/quica021",
   "source_url": "https://omlx.ai/benchmarks/performance/quica021",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-ayow3bhw-ctx1024",
   "model": "gemma-3-4b-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 3.2 GB at this context; duplicate traces of the same session: tre83hdw; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 357.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-3-4b-it-qat-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 1k 357.8 52.0 3.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ayow3bhw",
   "source_url": "https://omlx.ai/benchmarks/performance/ayow3bhw",
   "tg_tokens": null,
   "tps": 52.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-ayow3bhw-ctx4096",
   "model": "gemma-3-4b-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 3.8 GB at this context; duplicate traces of the same session: tre83hdw; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 346.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-3-4b-it-qat-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 4k 346.5 51.0 3.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ayow3bhw",
   "source_url": "https://omlx.ai/benchmarks/performance/ayow3bhw",
   "tg_tokens": null,
   "tps": 51.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-h5myk1bb-ctx4096",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 7.8 GB at this context; duplicate traces of the same session: r3l27k97; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 245.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12b-it-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 4k 245.5 35.4 7.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h5myk1bb",
   "source_url": "https://omlx.ai/benchmarks/performance/h5myk1bb",
   "tg_tokens": null,
   "tps": 35.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-h5myk1bb-ctx16384",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 8.3 GB at this context; duplicate traces of the same session: r3l27k97; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 234.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12b-it-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 16k 234.7 34.0 8.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h5myk1bb",
   "source_url": "https://omlx.ai/benchmarks/performance/h5myk1bb",
   "tg_tokens": null,
   "tps": 34.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Pro",
   "id": "omlx-vn8e2ale-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.3 GB at this context; duplicate traces of the same session: hr6aspk0; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 619.3,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "gpt-oss-20b-OptiQ-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 1k 619.3 13.8 11.3 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vn8e2ale",
   "source_url": "https://omlx.ai/benchmarks/performance/vn8e2ale",
   "tg_tokens": null,
   "tps": 13.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Pro",
   "id": "omlx-vn8e2ale-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.4 GB at this context; duplicate traces of the same session: hr6aspk0; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 526.4,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "gpt-oss-20b-OptiQ-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 4k 526.4 26.1 11.4 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vn8e2ale",
   "source_url": "https://omlx.ai/benchmarks/performance/vn8e2ale",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-02do2444-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 22.4 GB; TTFT 41178 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 397.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "swdjojo--Qwen3.8-27B-MLX-oQ4e on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k tokens 397.9 PP tok/s 28.6 TG tok/s 41178 TTFT (ms) 22.4 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/02do2444",
   "source_url": "https://omlx.ai/benchmarks/performance/02do2444",
   "tg_tokens": null,
   "tps": 28.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-bcqkb4a0-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 23.5 GB; TTFT 90960 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 360.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "swdjojo--Qwen3.8-27B-MLX-oQ4e on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k tokens 360.2 PP tok/s 13.9 TG tok/s 90960 TTFT (ms) 23.5 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bcqkb4a0",
   "source_url": "https://omlx.ai/benchmarks/performance/bcqkb4a0",
   "tg_tokens": null,
   "tps": 13.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-spf41c6l-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 8k context; peak memory 21.8 GB; TTFT 23259 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 352.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "swdjojo--Qwen3.8-27B-MLX-oQ4e on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k tokens 352.2 PP tok/s 28.6 TG tok/s 23259 TTFT (ms) 21.8 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/spf41c6l",
   "source_url": "https://omlx.ai/benchmarks/performance/spf41c6l",
   "tg_tokens": null,
   "tps": 28.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-k9vdaek2-ctx1024",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: kqd81yg0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 334.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 334.5 39.2 16.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k9vdaek2",
   "source_url": "https://omlx.ai/benchmarks/performance/k9vdaek2",
   "tg_tokens": null,
   "tps": 39.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-k9vdaek2-ctx4096",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: kqd81yg0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 356.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 356.6 37.9 19.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k9vdaek2",
   "source_url": "https://omlx.ai/benchmarks/performance/k9vdaek2",
   "tg_tokens": null,
   "tps": 37.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-w8hukpne-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 9.8 GB at this context; duplicate traces of the same session: 9gd4gkmp; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 217.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B-MLX-8bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 8bit; 1k 217.4 10.7 9.8 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w8hukpne",
   "source_url": "https://omlx.ai/benchmarks/performance/w8hukpne",
   "tg_tokens": null,
   "tps": 10.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-w8hukpne-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 10.3 GB at this context; duplicate traces of the same session: 9gd4gkmp; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B-MLX-8bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 8bit; 4k 220.4 10.7 10.3 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w8hukpne",
   "source_url": "https://omlx.ai/benchmarks/performance/w8hukpne",
   "tg_tokens": null,
   "tps": 10.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ph4oepq9-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: ae2eoj20; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4508.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,508 210.7 21.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ph4oepq9",
   "source_url": "https://omlx.ai/benchmarks/performance/ph4oepq9",
   "tg_tokens": null,
   "tps": 210.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ph4oepq9-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: ae2eoj20; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6252.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,252 183.8 22.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ph4oepq9",
   "source_url": "https://omlx.ai/benchmarks/performance/ph4oepq9",
   "tg_tokens": null,
   "tps": 183.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-r3mjgfxd-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: fc6yszza; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 589.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 589.1 83.8 21.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r3mjgfxd",
   "source_url": "https://omlx.ai/benchmarks/performance/r3mjgfxd",
   "tg_tokens": null,
   "tps": 83.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-r3mjgfxd-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: fc6yszza; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2252.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,252 193.0 22.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r3mjgfxd",
   "source_url": "https://omlx.ai/benchmarks/performance/r3mjgfxd",
   "tg_tokens": null,
   "tps": 193.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hi605sum-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: tjhua5jr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 352.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 352.7 37.8 19.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hi605sum",
   "source_url": "https://omlx.ai/benchmarks/performance/hi605sum",
   "tg_tokens": null,
   "tps": 37.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-hi605sum-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: tjhua5jr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 333.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 333.9 35.1 20.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hi605sum",
   "source_url": "https://omlx.ai/benchmarks/performance/hi605sum",
   "tg_tokens": null,
   "tps": 35.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-sud8hpp7-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: cy91p9l4, ep822nlz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 345.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 345.5 37.2 19.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sud8hpp7",
   "source_url": "https://omlx.ai/benchmarks/performance/sud8hpp7",
   "tg_tokens": null,
   "tps": 37.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-sud8hpp7-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: cy91p9l4, ep822nlz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 325.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 325.3 34.6 20.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sud8hpp7",
   "source_url": "https://omlx.ai/benchmarks/performance/sud8hpp7",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-sud8hpp7-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 26.9 GB at this context; duplicate traces of the same session: cy91p9l4, ep822nlz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 264.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 264.5 27.4 26.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sud8hpp7",
   "source_url": "https://omlx.ai/benchmarks/performance/sud8hpp7",
   "tg_tokens": null,
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-64jnrb3m-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: 6cmustcm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 861.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 861.7 32.2 33.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/64jnrb3m",
   "source_url": "https://omlx.ai/benchmarks/performance/64jnrb3m",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-64jnrb3m-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 35.0 GB at this context; duplicate traces of the same session: 6cmustcm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 933.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 933.9 34.5 35.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/64jnrb3m",
   "source_url": "https://omlx.ai/benchmarks/performance/64jnrb3m",
   "tg_tokens": null,
   "tps": 34.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vf0u94j3-ctx1024",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "ET-O5.5-G6A-G4.7-DSV4-K3-RLOO MTP DFlash2 build; oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: 82iptiam; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 834.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-ET-O5.5-G6A-G4.7-DSV4-K3-RLOO-MTP-DFlash2-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 834.3 48.5 18.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vf0u94j3",
   "source_url": "https://omlx.ai/benchmarks/performance/vf0u94j3",
   "tg_tokens": null,
   "tps": 48.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vf0u94j3-ctx4096",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "ET-O5.5-G6A-G4.7-DSV4-K3-RLOO MTP DFlash2 build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 82iptiam; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 951.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-ET-O5.5-G6A-G4.7-DSV4-K3-RLOO-MTP-DFlash2-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 951.5 41.2 19.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vf0u94j3",
   "source_url": "https://omlx.ai/benchmarks/performance/vf0u94j3",
   "tg_tokens": null,
   "tps": 41.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vf0u94j3-ctx8192",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "ET-O5.5-G6A-G4.7-DSV4-K3-RLOO MTP DFlash2 build; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 82iptiam; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1027.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-ET-O5.5-G6A-G4.7-DSV4-K3-RLOO-MTP-DFlash2-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,027 44.3 19.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vf0u94j3",
   "source_url": "https://omlx.ai/benchmarks/performance/vf0u94j3",
   "tg_tokens": null,
   "tps": 44.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9zekwijo-ctx16384",
   "model": "Qwen3.8-27B-Coder390-LynnStyle-Mixed",
   "notes": "A4toA8 build; MTPFix; oMLX community benchmark trace, single run at 16k context; peak memory 22.1 GB; TTFT 19285 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 849.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Coder390-LynnStyle-Mixed-MLX-A4toA8-MTPFix on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k tokens 849.6 PP tok/s 89.2 TG tok/s 19285 TTFT (ms) 22.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9zekwijo",
   "source_url": "https://omlx.ai/benchmarks/performance/9zekwijo",
   "tg_tokens": null,
   "tps": 89.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7xjhei6r-ctx16384",
   "model": "Qwen3.8-27B-Coder390-LynnStyle-Mixed",
   "notes": "A4toA8 build; MTPFix; oMLX community benchmark trace, single run at 16k context; peak memory 22.3 GB; TTFT 18744 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 874.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Coder390-LynnStyle-Mixed-MLX-A4toA8-MTPFix on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k tokens 874.1 PP tok/s 85.9 TG tok/s 18744 TTFT (ms) 22.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7xjhei6r",
   "source_url": "https://omlx.ai/benchmarks/performance/7xjhei6r",
   "tg_tokens": null,
   "tps": 85.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vhecmjcq-ctx1024",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1105.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,105 31.7 17.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vhecmjcq",
   "source_url": "https://omlx.ai/benchmarks/performance/vhecmjcq",
   "tg_tokens": null,
   "tps": 31.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vhecmjcq-ctx4096",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1382.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,382 31.5 19.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vhecmjcq",
   "source_url": "https://omlx.ai/benchmarks/performance/vhecmjcq",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vhecmjcq-ctx8192",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1400.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,400 31.0 19.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vhecmjcq",
   "source_url": "https://omlx.ai/benchmarks/performance/vhecmjcq",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-h29d0uqn-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: j7xu4r0z; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 130.5 12.7 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h29d0uqn",
   "source_url": "https://omlx.ai/benchmarks/performance/h29d0uqn",
   "tg_tokens": null,
   "tps": 12.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-h29d0uqn-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: j7xu4r0z; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 129.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 129.3 12.3 21.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h29d0uqn",
   "source_url": "https://omlx.ai/benchmarks/performance/h29d0uqn",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-134o62s4-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: 8hlkikvy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 938.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 938.4 52.0 18.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/134o62s4",
   "source_url": "https://omlx.ai/benchmarks/performance/134o62s4",
   "tg_tokens": null,
   "tps": 52.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-134o62s4-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: 8hlkikvy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 813.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 813.7 52.3 20.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/134o62s4",
   "source_url": "https://omlx.ai/benchmarks/performance/134o62s4",
   "tg_tokens": null,
   "tps": 52.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-7x614rzw-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 131.3 13.8 16.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7x614rzw",
   "source_url": "https://omlx.ai/benchmarks/performance/7x614rzw",
   "tg_tokens": null,
   "tps": 13.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-7x614rzw-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 119.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 119.8 13.3 20.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7x614rzw",
   "source_url": "https://omlx.ai/benchmarks/performance/7x614rzw",
   "tg_tokens": null,
   "tps": 13.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-kiqbbhyy-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 126.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 126.2 21.8 18.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kiqbbhyy",
   "source_url": "https://omlx.ai/benchmarks/performance/kiqbbhyy",
   "tg_tokens": null,
   "tps": 21.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-kiqbbhyy-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 115.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 115.8 19.1 22.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kiqbbhyy",
   "source_url": "https://omlx.ai/benchmarks/performance/kiqbbhyy",
   "tg_tokens": null,
   "tps": 19.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-j0e74rhv-ctx4096",
   "model": "Qwen3.8-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: pzbcmr51; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5890.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-35B-A3B-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 5,890 98.6 21.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j0e74rhv",
   "source_url": "https://omlx.ai/benchmarks/performance/j0e74rhv",
   "tg_tokens": null,
   "tps": 98.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-j0e74rhv-ctx16384",
   "model": "Qwen3.8-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: pzbcmr51; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5484.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-35B-A3B-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 5,484 68.5 21.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j0e74rhv",
   "source_url": "https://omlx.ai/benchmarks/performance/j0e74rhv",
   "tg_tokens": null,
   "tps": 68.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-e8822hdl-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 73.6 GB at this context; duplicate traces of the same session: 1u7i2ld5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 373.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 373.2 69.9 73.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e8822hdl",
   "source_url": "https://omlx.ai/benchmarks/performance/e8822hdl",
   "tg_tokens": null,
   "tps": 69.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-e8822hdl-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 73.8 GB at this context; duplicate traces of the same session: 1u7i2ld5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 925.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 925.9 28.6 73.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e8822hdl",
   "source_url": "https://omlx.ai/benchmarks/performance/e8822hdl",
   "tg_tokens": null,
   "tps": 28.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-i7ckb98e-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 100.7 GB at this context; duplicate traces of the same session: djfu78xl, pee5i15x, y1ryidwj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 595.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 595.9 35.8 100.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i7ckb98e",
   "source_url": "https://omlx.ai/benchmarks/performance/i7ckb98e",
   "tg_tokens": null,
   "tps": 35.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-i7ckb98e-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 103.0 GB at this context; duplicate traces of the same session: djfu78xl, pee5i15x, y1ryidwj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 603.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 603.0 34.5 103.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i7ckb98e",
   "source_url": "https://omlx.ai/benchmarks/performance/i7ckb98e",
   "tg_tokens": null,
   "tps": 34.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-i7ckb98e-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 104.0 GB at this context; duplicate traces of the same session: djfu78xl, pee5i15x, y1ryidwj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 576.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 576.1 34.1 104.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i7ckb98e",
   "source_url": "https://omlx.ai/benchmarks/performance/i7ckb98e",
   "tg_tokens": null,
   "tps": 34.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-i7ckb98e-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 108.8 GB at this context; duplicate traces of the same session: djfu78xl, pee5i15x, y1ryidwj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 514.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 514.7 32.1 108.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i7ckb98e",
   "source_url": "https://omlx.ai/benchmarks/performance/i7ckb98e",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3wgs20ar-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.5 GB at this context; duplicate traces of the same session: a37bwme4, bj05i0i8, jqj1v07j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 479.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 479.7 39.3 70.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3wgs20ar",
   "source_url": "https://omlx.ai/benchmarks/performance/3wgs20ar",
   "tg_tokens": null,
   "tps": 39.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3wgs20ar-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: a37bwme4, bj05i0i8, jqj1v07j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1083.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,083 32.3 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3wgs20ar",
   "source_url": "https://omlx.ai/benchmarks/performance/3wgs20ar",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3wgs20ar-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.9 GB at this context; duplicate traces of the same session: a37bwme4, bj05i0i8, jqj1v07j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2272.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,272 42.3 74.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3wgs20ar",
   "source_url": "https://omlx.ai/benchmarks/performance/3wgs20ar",
   "tg_tokens": null,
   "tps": 42.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3wgs20ar-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.2 GB at this context; duplicate traces of the same session: a37bwme4, bj05i0i8, jqj1v07j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1788.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 1,788 37.9 79.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3wgs20ar",
   "source_url": "https://omlx.ai/benchmarks/performance/3wgs20ar",
   "tg_tokens": null,
   "tps": 37.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-02z4v061-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: 3xorctw5, 6ubd1fr6, aewnosrq, fz3e6bpj, od79a439, u6vh11vr; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2480.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 2,480 148.7 100.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/02z4v061",
   "source_url": "https://omlx.ai/benchmarks/performance/02z4v061",
   "tg_tokens": null,
   "tps": 148.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-02z4v061-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.0 GB at this context; duplicate traces of the same session: 3xorctw5, 6ubd1fr6, aewnosrq, fz3e6bpj, od79a439, u6vh11vr; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4166.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 4,166 129.4 101.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/02z4v061",
   "source_url": "https://omlx.ai/benchmarks/performance/02z4v061",
   "tg_tokens": null,
   "tps": 129.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-02z4v061-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.1 GB at this context; duplicate traces of the same session: 3xorctw5, 6ubd1fr6, aewnosrq, fz3e6bpj, od79a439, u6vh11vr; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4695.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 4,695 136.3 102.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/02z4v061",
   "source_url": "https://omlx.ai/benchmarks/performance/02z4v061",
   "tg_tokens": null,
   "tps": 136.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-02z4v061-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.5 GB at this context; duplicate traces of the same session: 3xorctw5, 6ubd1fr6, aewnosrq, fz3e6bpj, od79a439, u6vh11vr; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4723.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 4,723 165.9 102.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/02z4v061",
   "source_url": "https://omlx.ai/benchmarks/performance/02z4v061",
   "tg_tokens": null,
   "tps": 165.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-02z4v061-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.3 GB at this context; duplicate traces of the same session: 3xorctw5, 6ubd1fr6, aewnosrq, fz3e6bpj, od79a439, u6vh11vr; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4621.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 4,621 128.4 104.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/02z4v061",
   "source_url": "https://omlx.ai/benchmarks/performance/02z4v061",
   "tg_tokens": null,
   "tps": 128.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-02z4v061-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.0 GB at this context; duplicate traces of the same session: 3xorctw5, 6ubd1fr6, aewnosrq, fz3e6bpj, od79a439, u6vh11vr; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4481.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 4,481 104.6 109.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/02z4v061",
   "source_url": "https://omlx.ai/benchmarks/performance/02z4v061",
   "tg_tokens": null,
   "tps": 104.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-02z4v061-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.7 GB at this context; duplicate traces of the same session: 3xorctw5, 6ubd1fr6, aewnosrq, fz3e6bpj, od79a439, u6vh11vr; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4354.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 4,354 84.9 110.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/02z4v061",
   "source_url": "https://omlx.ai/benchmarks/performance/02z4v061",
   "tg_tokens": null,
   "tps": 84.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-zuxgq1tx-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.4 GB at this context; duplicate traces of the same session: 9du80pzt, gx18qctk, q0apwuxl, xirbigkx; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1360.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,360 140.2 100.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zuxgq1tx",
   "source_url": "https://omlx.ai/benchmarks/performance/zuxgq1tx",
   "tg_tokens": null,
   "tps": 140.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-zuxgq1tx-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.2 GB at this context; duplicate traces of the same session: 9du80pzt, gx18qctk, q0apwuxl, xirbigkx; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5304.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 5,304 129.6 102.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zuxgq1tx",
   "source_url": "https://omlx.ai/benchmarks/performance/zuxgq1tx",
   "tg_tokens": null,
   "tps": 129.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-zuxgq1tx-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.3 GB at this context; duplicate traces of the same session: 9du80pzt, gx18qctk, q0apwuxl, xirbigkx; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5180.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 5,180 126.8 104.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zuxgq1tx",
   "source_url": "https://omlx.ai/benchmarks/performance/zuxgq1tx",
   "tg_tokens": null,
   "tps": 126.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-zuxgq1tx-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.0 GB at this context; duplicate traces of the same session: 9du80pzt, gx18qctk, q0apwuxl, xirbigkx; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5015.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 5,015 101.5 109.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zuxgq1tx",
   "source_url": "https://omlx.ai/benchmarks/performance/zuxgq1tx",
   "tg_tokens": null,
   "tps": 101.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-zuxgq1tx-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.7 GB at this context; duplicate traces of the same session: 9du80pzt, gx18qctk, q0apwuxl, xirbigkx; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4855.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 4,855 89.2 110.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zuxgq1tx",
   "source_url": "https://omlx.ai/benchmarks/performance/zuxgq1tx",
   "tg_tokens": null,
   "tps": 89.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-h8r5l6hv-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 25.3 GB; TTFT 36652 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 223.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k tokens 223.5 PP tok/s 29.8 TG tok/s 36652 TTFT (ms) 25.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h8r5l6hv",
   "source_url": "https://omlx.ai/benchmarks/performance/h8r5l6hv",
   "tg_tokens": null,
   "tps": 29.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-pi2u022b-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace, single run at 4k context; peak memory 71.5 GB; TTFT 7251 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 564.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 564.9 PP tok/s 60.7 TG tok/s 7251 TTFT (ms) 71.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pi2u022b",
   "source_url": "https://omlx.ai/benchmarks/performance/pi2u022b",
   "tg_tokens": null,
   "tps": 60.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-c0ecdd2s-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: 0j626cr2, 6gpa0ged, 8oq25x5r, ceqz9ql9, monit7a9, ncj0jjdh, ziksaoxn; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1323.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,323 55.0 177.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c0ecdd2s",
   "source_url": "https://omlx.ai/benchmarks/performance/c0ecdd2s",
   "tg_tokens": null,
   "tps": 55.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-c0ecdd2s-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: 0j626cr2, 6gpa0ged, 8oq25x5r, ceqz9ql9, monit7a9, ncj0jjdh, ziksaoxn; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1894.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,894 61.2 178.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c0ecdd2s",
   "source_url": "https://omlx.ai/benchmarks/performance/c0ecdd2s",
   "tg_tokens": null,
   "tps": 61.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-c0ecdd2s-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.9 GB at this context; duplicate traces of the same session: 0j626cr2, 6gpa0ged, 8oq25x5r, ceqz9ql9, monit7a9, ncj0jjdh, ziksaoxn; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1922.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 1,922 55.9 178.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c0ecdd2s",
   "source_url": "https://omlx.ai/benchmarks/performance/c0ecdd2s",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-c0ecdd2s-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.1 GB at this context; duplicate traces of the same session: 0j626cr2, 6gpa0ged, 8oq25x5r, ceqz9ql9, monit7a9, ncj0jjdh, ziksaoxn; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1922.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,922 134.5 179.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c0ecdd2s",
   "source_url": "https://omlx.ai/benchmarks/performance/c0ecdd2s",
   "tg_tokens": null,
   "tps": 134.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-c0ecdd2s-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.4 GB at this context; duplicate traces of the same session: 0j626cr2, 6gpa0ged, 8oq25x5r, ceqz9ql9, monit7a9, ncj0jjdh, ziksaoxn; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1912.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,912 88.0 179.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c0ecdd2s",
   "source_url": "https://omlx.ai/benchmarks/performance/c0ecdd2s",
   "tg_tokens": null,
   "tps": 88.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-c0ecdd2s-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.1 GB at this context; duplicate traces of the same session: 0j626cr2, 6gpa0ged, 8oq25x5r, ceqz9ql9, monit7a9, ncj0jjdh, ziksaoxn; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1880.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,880 79.0 180.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c0ecdd2s",
   "source_url": "https://omlx.ai/benchmarks/performance/c0ecdd2s",
   "tg_tokens": null,
   "tps": 79.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-c0ecdd2s-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.3 GB at this context; duplicate traces of the same session: 0j626cr2, 6gpa0ged, 8oq25x5r, ceqz9ql9, monit7a9, ncj0jjdh, ziksaoxn; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1847.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 1,847 132.4 181.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c0ecdd2s",
   "source_url": "https://omlx.ai/benchmarks/performance/c0ecdd2s",
   "tg_tokens": null,
   "tps": 132.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-c0ecdd2s-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 182.6 GB at this context; duplicate traces of the same session: 0j626cr2, 6gpa0ged, 8oq25x5r, ceqz9ql9, monit7a9, ncj0jjdh, ziksaoxn; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1792.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 1,792 63.3 182.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c0ecdd2s",
   "source_url": "https://omlx.ai/benchmarks/performance/c0ecdd2s",
   "tg_tokens": null,
   "tps": 63.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gpqth3xj-ctx16384",
   "model": "Ornith-1.5-35B-A3B-uncensored",
   "notes": "fixed MTP vision MLX build; oMLX community benchmark trace, single run at 16k context; peak memory 27.7 GB; TTFT 4024 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4072.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-35B-A3B-uncensored-MLX-6bit-fixed-mtp-vision on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k tokens 4,072 PP tok/s 102.3 TG tok/s 4024 TTFT (ms) 27.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gpqth3xj",
   "source_url": "https://omlx.ai/benchmarks/performance/gpqth3xj",
   "tg_tokens": null,
   "tps": 102.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0qjofmpk-ctx1024",
   "model": "Qwen3 8B",
   "notes": "oMLX community benchmark trace; peak memory 5.0 GB at this context; duplicate traces of the same session: 31wt0a8g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3386.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-8B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 3,386 104.7 5.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0qjofmpk",
   "source_url": "https://omlx.ai/benchmarks/performance/0qjofmpk",
   "tg_tokens": null,
   "tps": 104.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0qjofmpk-ctx4096",
   "model": "Qwen3 8B",
   "notes": "oMLX community benchmark trace; peak memory 5.5 GB at this context; duplicate traces of the same session: 31wt0a8g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3468.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-8B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 3,468 98.3 5.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0qjofmpk",
   "source_url": "https://omlx.ai/benchmarks/performance/0qjofmpk",
   "tg_tokens": null,
   "tps": 98.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-l7g5386a-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 6.9 GB; TTFT 45741 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 358.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 16k tokens 358.2 PP tok/s 43.6 TG tok/s 45741 TTFT (ms) 6.9 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l7g5386a",
   "source_url": "https://omlx.ai/benchmarks/performance/l7g5386a",
   "tg_tokens": null,
   "tps": 43.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-y2tq60zu-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 7.2 GB; TTFT 9204 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 111.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k tokens 111.3 PP tok/s 14.5 TG tok/s 9204 TTFT (ms) 7.2 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y2tq60zu",
   "source_url": "https://omlx.ai/benchmarks/performance/y2tq60zu",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-eflg5jgy-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: pv8vktp8; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 8bit; 1k 224.5 21.3 28.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eflg5jgy",
   "source_url": "https://omlx.ai/benchmarks/performance/eflg5jgy",
   "tg_tokens": null,
   "tps": 21.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-eflg5jgy-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.9 GB at this context; duplicate traces of the same session: pv8vktp8; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 229.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 8bit; 4k 229.1 20.8 31.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eflg5jgy",
   "source_url": "https://omlx.ai/benchmarks/performance/eflg5jgy",
   "tg_tokens": null,
   "tps": 20.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-674nrtfs-ctx1024",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "ET-O5.5-G6A-G4.7-DSV4-K3-RLOO MTP DFlash2 build; oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 834.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-ET-O5.5-G6A-G4.7-DSV4-K3-RLOO-MTP-DFlash2-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 834.3 48.5 18.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/674nrtfs",
   "source_url": "https://omlx.ai/benchmarks/performance/674nrtfs",
   "tg_tokens": null,
   "tps": 48.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-674nrtfs-ctx4096",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "ET-O5.5-G6A-G4.7-DSV4-K3-RLOO MTP DFlash2 build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 951.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-ET-O5.5-G6A-G4.7-DSV4-K3-RLOO-MTP-DFlash2-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 951.5 41.2 19.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/674nrtfs",
   "source_url": "https://omlx.ai/benchmarks/performance/674nrtfs",
   "tg_tokens": null,
   "tps": 41.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-674nrtfs-ctx8192",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "ET-O5.5-G6A-G4.7-DSV4-K3-RLOO MTP DFlash2 build; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1027.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-ET-O5.5-G6A-G4.7-DSV4-K3-RLOO-MTP-DFlash2-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,027 44.3 19.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/674nrtfs",
   "source_url": "https://omlx.ai/benchmarks/performance/674nrtfs",
   "tg_tokens": null,
   "tps": 44.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ce6qj5au-ctx1024",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "ET-O5.5-G6A-G4.7-DSV4-K3-RLOO MTP DFlash2 build; oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: 097c3zq3, ejp1hr4g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1102.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-ET-O5.5-G6A-G4.7-DSV4-K3-RLOO-MTP-DFlash2-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,102 44.1 18.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ce6qj5au",
   "source_url": "https://omlx.ai/benchmarks/performance/ce6qj5au",
   "tg_tokens": null,
   "tps": 44.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ce6qj5au-ctx4096",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "ET-O5.5-G6A-G4.7-DSV4-K3-RLOO MTP DFlash2 build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 097c3zq3, ejp1hr4g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1361.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-ET-O5.5-G6A-G4.7-DSV4-K3-RLOO-MTP-DFlash2-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,361 51.8 19.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ce6qj5au",
   "source_url": "https://omlx.ai/benchmarks/performance/ce6qj5au",
   "tg_tokens": null,
   "tps": 51.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ce6qj5au-ctx8192",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "ET-O5.5-G6A-G4.7-DSV4-K3-RLOO MTP DFlash2 build; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 097c3zq3, ejp1hr4g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1370.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-ET-O5.5-G6A-G4.7-DSV4-K3-RLOO-MTP-DFlash2-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,370 53.2 19.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ce6qj5au",
   "source_url": "https://omlx.ai/benchmarks/performance/ce6qj5au",
   "tg_tokens": null,
   "tps": 53.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hb8n92j4-ctx16384",
   "model": "Qwen3.8-27B-Coder390-LynnStyle-Mixed",
   "notes": "A4toA8 build; MTPFix; oMLX community benchmark trace, single run at 16k context; peak memory 21.1 GB; TTFT 18950 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 864.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Coder390-LynnStyle-Mixed-MLX-A4toA8-MTPFix on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k tokens 864.6 PP tok/s 84.9 TG tok/s 18950 TTFT (ms) 21.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hb8n92j4",
   "source_url": "https://omlx.ai/benchmarks/performance/hb8n92j4",
   "tg_tokens": null,
   "tps": 84.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-966poiu5-ctx16384",
   "model": "Qwen3.8-27B-Coder390-LynnStyle-Mixed",
   "notes": "A4toA8 build; MTPFix; oMLX community benchmark trace, single run at 16k context; peak memory 21.1 GB; TTFT 18767 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 873.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Coder390-LynnStyle-Mixed-MLX-A4toA8-MTPFix on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k tokens 873.0 PP tok/s 86.0 TG tok/s 18767 TTFT (ms) 21.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/966poiu5",
   "source_url": "https://omlx.ai/benchmarks/performance/966poiu5",
   "tg_tokens": null,
   "tps": 86.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-c503qlgp-ctx1024",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: sb4s52wh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1105.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,105 31.7 17.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c503qlgp",
   "source_url": "https://omlx.ai/benchmarks/performance/c503qlgp",
   "tg_tokens": null,
   "tps": 31.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-c503qlgp-ctx4096",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: sb4s52wh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1382.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,382 31.5 19.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c503qlgp",
   "source_url": "https://omlx.ai/benchmarks/performance/c503qlgp",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-c503qlgp-ctx8192",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: sb4s52wh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1400.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,400 31.0 19.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c503qlgp",
   "source_url": "https://omlx.ai/benchmarks/performance/c503qlgp",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-7nwoloou-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: 610ciqng; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 247.7,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "Qwen3.8-27B-oQ2e-fp16-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 2bit; 1k 247.7 52.5 14.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7nwoloou",
   "source_url": "https://omlx.ai/benchmarks/performance/7nwoloou",
   "tg_tokens": null,
   "tps": 52.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-7nwoloou-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: 610ciqng; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 246.6,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "Qwen3.8-27B-oQ2e-fp16-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 2bit; 4k 246.6 47.4 18.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7nwoloou",
   "source_url": "https://omlx.ai/benchmarks/performance/7nwoloou",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-rjomo1ma-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: umnvqj6q; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 171.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 171.4 18.6 24.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rjomo1ma",
   "source_url": "https://omlx.ai/benchmarks/performance/rjomo1ma",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-rjomo1ma-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: umnvqj6q; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 163.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 163.3 16.2 25.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rjomo1ma",
   "source_url": "https://omlx.ai/benchmarks/performance/rjomo1ma",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-10a5un1r-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 20.9 GB; TTFT 265923 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 61.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 16k tokens 61.6 PP tok/s 9.3 TG tok/s 265923 TTFT (ms) 20.9 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/10a5un1r",
   "source_url": "https://omlx.ai/benchmarks/performance/10a5un1r",
   "tg_tokens": null,
   "tps": 9.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-q53kk5at-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 2dzh50x2, tdvs1neu, tyczirid, u1ogndpk; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 160.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 1k 160.7 11.8 28.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q53kk5at",
   "source_url": "https://omlx.ai/benchmarks/performance/q53kk5at",
   "tg_tokens": null,
   "tps": 11.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-q53kk5at-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: 2dzh50x2, tdvs1neu, tyczirid, u1ogndpk; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 162.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 4k 162.6 11.3 32.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q53kk5at",
   "source_url": "https://omlx.ai/benchmarks/performance/q53kk5at",
   "tg_tokens": null,
   "tps": 11.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-q53kk5at-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 33.1 GB at this context; duplicate traces of the same session: 2dzh50x2, tdvs1neu, tyczirid, u1ogndpk; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 150.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 8k 150.8 11.5 33.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q53kk5at",
   "source_url": "https://omlx.ai/benchmarks/performance/q53kk5at",
   "tg_tokens": null,
   "tps": 11.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-q53kk5at-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 34.2 GB at this context; duplicate traces of the same session: 2dzh50x2, tdvs1neu, tyczirid, u1ogndpk; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 152.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 16k 152.6 11.2 34.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q53kk5at",
   "source_url": "https://omlx.ai/benchmarks/performance/q53kk5at",
   "tg_tokens": null,
   "tps": 11.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-q53kk5at-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: 2dzh50x2, tdvs1neu, tyczirid, u1ogndpk; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 137.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 32k 137.8 10.2 35.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q53kk5at",
   "source_url": "https://omlx.ai/benchmarks/performance/q53kk5at",
   "tg_tokens": null,
   "tps": 10.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-efk14pum-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: 3u3nz57k; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 298.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 8bit; 1k 298.1 35.6 35.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/efk14pum",
   "source_url": "https://omlx.ai/benchmarks/performance/efk14pum",
   "tg_tokens": null,
   "tps": 35.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-efk14pum-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 38.4 GB at this context; duplicate traces of the same session: 3u3nz57k; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 313.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 8bit; 4k 313.9 34.3 38.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/efk14pum",
   "source_url": "https://omlx.ai/benchmarks/performance/efk14pum",
   "tg_tokens": null,
   "tps": 34.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-ig0vu341-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 72.6 GB at this context; duplicate traces of the same session: aw3vqzma, r4rbfqq9; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 285.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 1k 285.8 11.2 72.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ig0vu341",
   "source_url": "https://omlx.ai/benchmarks/performance/ig0vu341",
   "tg_tokens": null,
   "tps": 11.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-ig0vu341-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 79.3 GB at this context; duplicate traces of the same session: aw3vqzma, r4rbfqq9; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 295.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 4k 295.1 11.4 79.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ig0vu341",
   "source_url": "https://omlx.ai/benchmarks/performance/ig0vu341",
   "tg_tokens": null,
   "tps": 11.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-ig0vu341-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 82.2 GB at this context; duplicate traces of the same session: aw3vqzma, r4rbfqq9; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 412.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 8k 412.0 11.4 82.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ig0vu341",
   "source_url": "https://omlx.ai/benchmarks/performance/ig0vu341",
   "tg_tokens": null,
   "tps": 11.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-2t5v7yj9-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.3 GB at this context; duplicate traces of the same session: tahd7aa2; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 279.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 1k 279.1 38.8 69.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2t5v7yj9",
   "source_url": "https://omlx.ai/benchmarks/performance/2t5v7yj9",
   "tg_tokens": null,
   "tps": 38.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-2t5v7yj9-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.9 GB at this context; duplicate traces of the same session: tahd7aa2; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 292.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 4k 292.2 34.7 69.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2t5v7yj9",
   "source_url": "https://omlx.ai/benchmarks/performance/2t5v7yj9",
   "tg_tokens": null,
   "tps": 34.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-zw2dzefr-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 0e8bnfvh, fnsxqw13, lr3zzf32, mr241py7, x8xndh47; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 520.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 520.0 48.3 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zw2dzefr",
   "source_url": "https://omlx.ai/benchmarks/performance/zw2dzefr",
   "tg_tokens": null,
   "tps": 48.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-zw2dzefr-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.5 GB at this context; duplicate traces of the same session: 0e8bnfvh, fnsxqw13, lr3zzf32, mr241py7, x8xndh47; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 580.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 580.7 50.0 72.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zw2dzefr",
   "source_url": "https://omlx.ai/benchmarks/performance/zw2dzefr",
   "tg_tokens": null,
   "tps": 50.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-zw2dzefr-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.1 GB at this context; duplicate traces of the same session: 0e8bnfvh, fnsxqw13, lr3zzf32, mr241py7, x8xndh47; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 588.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 588.1 52.1 74.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zw2dzefr",
   "source_url": "https://omlx.ai/benchmarks/performance/zw2dzefr",
   "tg_tokens": null,
   "tps": 52.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-zw2dzefr-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.4 GB at this context; duplicate traces of the same session: 0e8bnfvh, fnsxqw13, lr3zzf32, mr241py7, x8xndh47; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 564.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 564.2 88.0 74.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zw2dzefr",
   "source_url": "https://omlx.ai/benchmarks/performance/zw2dzefr",
   "tg_tokens": null,
   "tps": 88.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-zw2dzefr-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 77.4 GB at this context; duplicate traces of the same session: 0e8bnfvh, fnsxqw13, lr3zzf32, mr241py7, x8xndh47; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 546.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 546.9 65.1 77.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zw2dzefr",
   "source_url": "https://omlx.ai/benchmarks/performance/zw2dzefr",
   "tg_tokens": null,
   "tps": 65.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-zw2dzefr-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.5 GB at this context; duplicate traces of the same session: 0e8bnfvh, fnsxqw13, lr3zzf32, mr241py7, x8xndh47; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 531.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 531.5 61.1 78.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zw2dzefr",
   "source_url": "https://omlx.ai/benchmarks/performance/zw2dzefr",
   "tg_tokens": null,
   "tps": 61.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-37eft70a-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.1 GB at this context; duplicate traces of the same session: 5xhe3rj5, auj0et3l, svfpk2g8; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2445.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 1k 2,445 124.9 183.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/37eft70a",
   "source_url": "https://omlx.ai/benchmarks/performance/37eft70a",
   "tg_tokens": null,
   "tps": 124.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-37eft70a-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 187.2 GB at this context; duplicate traces of the same session: 5xhe3rj5, auj0et3l, svfpk2g8; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4615.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 64k 4,615 111.4 187.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/37eft70a",
   "source_url": "https://omlx.ai/benchmarks/performance/37eft70a",
   "tg_tokens": null,
   "tps": 111.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-37eft70a-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 191.6 GB at this context; duplicate traces of the same session: 5xhe3rj5, auj0et3l, svfpk2g8; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4480.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 128k 4,480 96.4 191.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/37eft70a",
   "source_url": "https://omlx.ai/benchmarks/performance/37eft70a",
   "tg_tokens": null,
   "tps": 96.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-37eft70a-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 193.3 GB at this context; duplicate traces of the same session: 5xhe3rj5, auj0et3l, svfpk2g8; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4342.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 195k 4,342 84.7 193.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/37eft70a",
   "source_url": "https://omlx.ai/benchmarks/performance/37eft70a",
   "tg_tokens": null,
   "tps": 84.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-1wbwokzp-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.0 GB at this context; duplicate traces of the same session: 0sk1nwgn, 738g1f3b, i46pkwlr, sb11fl3q; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2343.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 1k 2,343 109.9 183.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1wbwokzp",
   "source_url": "https://omlx.ai/benchmarks/performance/1wbwokzp",
   "tg_tokens": null,
   "tps": 109.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-1wbwokzp-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.7 GB at this context; duplicate traces of the same session: 0sk1nwgn, 738g1f3b, i46pkwlr, sb11fl3q; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4147.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 4k 4,147 110.3 183.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1wbwokzp",
   "source_url": "https://omlx.ai/benchmarks/performance/1wbwokzp",
   "tg_tokens": null,
   "tps": 110.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-1wbwokzp-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 187.2 GB at this context; duplicate traces of the same session: 0sk1nwgn, 738g1f3b, i46pkwlr, sb11fl3q; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4446.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 64k 4,446 113.2 187.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1wbwokzp",
   "source_url": "https://omlx.ai/benchmarks/performance/1wbwokzp",
   "tg_tokens": null,
   "tps": 113.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-1wbwokzp-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 191.6 GB at this context; duplicate traces of the same session: 0sk1nwgn, 738g1f3b, i46pkwlr, sb11fl3q; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4314.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 128k 4,314 97.8 191.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1wbwokzp",
   "source_url": "https://omlx.ai/benchmarks/performance/1wbwokzp",
   "tg_tokens": null,
   "tps": 97.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-1wbwokzp-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 193.3 GB at this context; duplicate traces of the same session: 0sk1nwgn, 738g1f3b, i46pkwlr, sb11fl3q; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4192.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 195k 4,192 84.1 193.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1wbwokzp",
   "source_url": "https://omlx.ai/benchmarks/performance/1wbwokzp",
   "tg_tokens": null,
   "tps": 84.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-mliinmst-ctx4096",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 7.6 GB at this context; duplicate traces of the same session: 3gf3ht6c; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 567.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 567.3 18.1 7.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mliinmst",
   "source_url": "https://omlx.ai/benchmarks/performance/mliinmst",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-mliinmst-ctx16384",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 8.0 GB at this context; duplicate traces of the same session: 3gf3ht6c; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 534.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 534.7 14.9 8.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mliinmst",
   "source_url": "https://omlx.ai/benchmarks/performance/mliinmst",
   "tg_tokens": null,
   "tps": 14.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-thfomb5x-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: dopiitx9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 644.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "gemma-4-31b-it-6bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 6bit; 4k 644.7 18.8 26.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/thfomb5x",
   "source_url": "https://omlx.ai/benchmarks/performance/thfomb5x",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-thfomb5x-ctx16384",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: dopiitx9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 510.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "gemma-4-31b-it-6bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 6bit; 16k 510.4 17.4 27.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/thfomb5x",
   "source_url": "https://omlx.ai/benchmarks/performance/thfomb5x",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-cw0w7o5s-ctx1024",
   "model": "K2-Horizon-7B",
   "notes": "oMLX community benchmark trace; peak memory 7.7 GB at this context; duplicate traces of the same session: a1mvjy3i; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "K2-Horizon-7B-oQ6e on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 6bit; 1k 224.1 13.2 7.7 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cw0w7o5s",
   "source_url": "https://omlx.ai/benchmarks/performance/cw0w7o5s",
   "tg_tokens": null,
   "tps": 13.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-cw0w7o5s-ctx4096",
   "model": "K2-Horizon-7B",
   "notes": "oMLX community benchmark trace; peak memory 8.2 GB at this context; duplicate traces of the same session: a1mvjy3i; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 216.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "K2-Horizon-7B-oQ6e on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 6bit; 4k 216.7 12.1 8.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cw0w7o5s",
   "source_url": "https://omlx.ai/benchmarks/performance/cw0w7o5s",
   "tg_tokens": null,
   "tps": 12.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-bclsffv2-ctx1024",
   "model": "LFM2.5-8B-A1B",
   "notes": "oMLX community benchmark trace; peak memory 5.1 GB at this context; duplicate traces of the same session: 41l6tdue; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 965.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-8B-A1B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 965.1 76.0 5.1 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bclsffv2",
   "source_url": "https://omlx.ai/benchmarks/performance/bclsffv2",
   "tg_tokens": null,
   "tps": 76.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-bclsffv2-ctx4096",
   "model": "LFM2.5-8B-A1B",
   "notes": "oMLX community benchmark trace; peak memory 5.3 GB at this context; duplicate traces of the same session: 41l6tdue; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1006.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-8B-A1B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 1,006 66.9 5.3 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bclsffv2",
   "source_url": "https://omlx.ai/benchmarks/performance/bclsffv2",
   "tg_tokens": null,
   "tps": 66.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-3uj3y13o-ctx1024",
   "model": "Magidonia-24B-v4.3",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: oloajwex; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 474.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Magidonia-24B-v4.3-mlx-6Bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 474.7 14.8 18.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3uj3y13o",
   "source_url": "https://omlx.ai/benchmarks/performance/3uj3y13o",
   "tg_tokens": null,
   "tps": 14.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-3uj3y13o-ctx4096",
   "model": "Magidonia-24B-v4.3",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: oloajwex; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 509.2,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Magidonia-24B-v4.3-mlx-6Bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 509.2 14.3 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3uj3y13o",
   "source_url": "https://omlx.ai/benchmarks/performance/3uj3y13o",
   "tg_tokens": null,
   "tps": 14.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-h23mwdx7-ctx1024",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 22.9 GB; TTFT 283 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3619.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 3,619 PP tok/s 130.9 TG tok/s 283 TTFT (ms) 22.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h23mwdx7",
   "source_url": "https://omlx.ai/benchmarks/performance/h23mwdx7",
   "tg_tokens": null,
   "tps": 130.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-kpp199mb-ctx1024",
   "model": "NeoHorse-1-9B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 1k context; peak memory 10.0 GB; TTFT 2777 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 368.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "NeoHorse-1-9B-MLX-8bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 8bit; 1k tokens 368.8 PP tok/s 27.5 TG tok/s 2777 TTFT (ms) 10.0 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kpp199mb",
   "source_url": "https://omlx.ai/benchmarks/performance/kpp199mb",
   "tg_tokens": null,
   "tps": 27.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-34dvyw7b-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: vc152lo5, zat51hhw; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 787.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 787.7 97.7 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/34dvyw7b",
   "source_url": "https://omlx.ai/benchmarks/performance/34dvyw7b",
   "tg_tokens": null,
   "tps": 97.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-34dvyw7b-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: vc152lo5, zat51hhw; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 868.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 868.9 86.3 23.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/34dvyw7b",
   "source_url": "https://omlx.ai/benchmarks/performance/34dvyw7b",
   "tg_tokens": null,
   "tps": 86.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-34dvyw7b-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: vc152lo5, zat51hhw; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 813.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 8k 813.1 89.2 23.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/34dvyw7b",
   "source_url": "https://omlx.ai/benchmarks/performance/34dvyw7b",
   "tg_tokens": null,
   "tps": 89.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-xw1bwnns-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 775.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 1k 775.8 72.6 37.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xw1bwnns",
   "source_url": "https://omlx.ai/benchmarks/performance/xw1bwnns",
   "tg_tokens": null,
   "tps": 72.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-xw1bwnns-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 848.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 4k 848.6 74.7 39.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xw1bwnns",
   "source_url": "https://omlx.ai/benchmarks/performance/xw1bwnns",
   "tg_tokens": null,
   "tps": 74.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-xw1bwnns-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.7 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 796.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 8k 796.4 70.2 39.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xw1bwnns",
   "source_url": "https://omlx.ai/benchmarks/performance/xw1bwnns",
   "tg_tokens": null,
   "tps": 70.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-tk2k23kq-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.8 GB at this context; duplicate traces of the same session: hvnu9axe; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 225.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 6bit; 1k 225.0 13.7 7.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tk2k23kq",
   "source_url": "https://omlx.ai/benchmarks/performance/tk2k23kq",
   "tg_tokens": null,
   "tps": 13.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-tk2k23kq-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.2 GB at this context; duplicate traces of the same session: hvnu9axe; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 226.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 6bit; 4k 226.8 13.6 8.2 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tk2k23kq",
   "source_url": "https://omlx.ai/benchmarks/performance/tk2k23kq",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ku2mq522-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 11.1 GB at this context; duplicate traces of the same session: 4uc87n10, f42r2g9g, gmaopd3l, ywnbtqfl; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 347.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-9B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 1k 347.8 38.9 11.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ku2mq522",
   "source_url": "https://omlx.ai/benchmarks/performance/ku2mq522",
   "tg_tokens": null,
   "tps": 38.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ku2mq522-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 11.7 GB at this context; duplicate traces of the same session: 4uc87n10, f42r2g9g, gmaopd3l, ywnbtqfl; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 412.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-9B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 4k 412.8 46.1 11.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ku2mq522",
   "source_url": "https://omlx.ai/benchmarks/performance/ku2mq522",
   "tg_tokens": null,
   "tps": 46.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ku2mq522-ctx8192",
   "model": "Ornith-1.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 11.9 GB at this context; duplicate traces of the same session: 4uc87n10, f42r2g9g, gmaopd3l, ywnbtqfl; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 440.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-9B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 8k 440.7 29.6 11.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ku2mq522",
   "source_url": "https://omlx.ai/benchmarks/performance/ku2mq522",
   "tg_tokens": null,
   "tps": 29.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ku2mq522-ctx16384",
   "model": "Ornith-1.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 12.3 GB at this context; duplicate traces of the same session: 4uc87n10, f42r2g9g, gmaopd3l, ywnbtqfl; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 434.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-9B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 16k 434.9 38.1 12.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ku2mq522",
   "source_url": "https://omlx.ai/benchmarks/performance/ku2mq522",
   "tg_tokens": null,
   "tps": 38.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ku2mq522-ctx32768",
   "model": "Ornith-1.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 13.1 GB at this context; duplicate traces of the same session: 4uc87n10, f42r2g9g, gmaopd3l, ywnbtqfl; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 388.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-9B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 32k 388.7 26.9 13.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ku2mq522",
   "source_url": "https://omlx.ai/benchmarks/performance/ku2mq522",
   "tg_tokens": null,
   "tps": 26.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-9veuhh6x-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 11.1 GB at this context; duplicate traces of the same session: 5uwomqqk, bguj5vb8, f768peko, m9sa1ul0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 279.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-9B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 1k 279.5 54.2 11.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9veuhh6x",
   "source_url": "https://omlx.ai/benchmarks/performance/9veuhh6x",
   "tg_tokens": null,
   "tps": 54.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-9veuhh6x-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 11.7 GB at this context; duplicate traces of the same session: 5uwomqqk, bguj5vb8, f768peko, m9sa1ul0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 345.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-9B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 4k 345.4 47.1 11.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9veuhh6x",
   "source_url": "https://omlx.ai/benchmarks/performance/9veuhh6x",
   "tg_tokens": null,
   "tps": 47.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-9veuhh6x-ctx8192",
   "model": "Ornith-1.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 11.9 GB at this context; duplicate traces of the same session: 5uwomqqk, bguj5vb8, f768peko, m9sa1ul0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 396.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-9B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 8k 396.5 35.8 11.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9veuhh6x",
   "source_url": "https://omlx.ai/benchmarks/performance/9veuhh6x",
   "tg_tokens": null,
   "tps": 35.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-9veuhh6x-ctx16384",
   "model": "Ornith-1.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 12.3 GB at this context; duplicate traces of the same session: 5uwomqqk, bguj5vb8, f768peko, m9sa1ul0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 419.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-9B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 16k 419.2 35.4 12.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9veuhh6x",
   "source_url": "https://omlx.ai/benchmarks/performance/9veuhh6x",
   "tg_tokens": null,
   "tps": 35.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-9veuhh6x-ctx32768",
   "model": "Ornith-1.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 13.1 GB at this context; duplicate traces of the same session: 5uwomqqk, bguj5vb8, f768peko, m9sa1ul0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 383.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-9B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 32k 383.1 36.6 13.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9veuhh6x",
   "source_url": "https://omlx.ai/benchmarks/performance/9veuhh6x",
   "tg_tokens": null,
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-s7pux90s-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: to5rxq2s; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 681.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 1k 681.4 58.9 19.4 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s7pux90s",
   "source_url": "https://omlx.ai/benchmarks/performance/s7pux90s",
   "tg_tokens": null,
   "tps": 58.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-s7pux90s-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: to5rxq2s; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 788.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 4k 788.8 57.0 20.1 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s7pux90s",
   "source_url": "https://omlx.ai/benchmarks/performance/s7pux90s",
   "tg_tokens": null,
   "tps": 57.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9v2rual7-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 65.8 GB at this context; duplicate traces of the same session: b1igtwgm; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1461.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.6-35B-A3B-bf16 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 bf16; 1k 1,461 68.6 65.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9v2rual7",
   "source_url": "https://omlx.ai/benchmarks/performance/9v2rual7",
   "tg_tokens": null,
   "tps": 68.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9v2rual7-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 67.8 GB at this context; duplicate traces of the same session: b1igtwgm; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3135.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.6-35B-A3B-bf16 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 bf16; 4k 3,135 67.6 67.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9v2rual7",
   "source_url": "https://omlx.ai/benchmarks/performance/9v2rual7",
   "tg_tokens": null,
   "tps": 67.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rt4dbx9u-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 29.7 GB; TTFT 4461 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 918.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k tokens 918.3 PP tok/s 18.8 TG tok/s 4461 TTFT (ms) 29.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rt4dbx9u",
   "source_url": "https://omlx.ai/benchmarks/performance/rt4dbx9u",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-aewcprky-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 30.4 GB; TTFT 17491 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 936.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 16k tokens 936.7 PP tok/s 18.1 TG tok/s 17491 TTFT (ms) 30.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aewcprky",
   "source_url": "https://omlx.ai/benchmarks/performance/aewcprky",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0asw91vl-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 28.5 GB; TTFT 1317 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 777.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k tokens 777.4 PP tok/s 19.0 TG tok/s 1317 TTFT (ms) 28.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0asw91vl",
   "source_url": "https://omlx.ai/benchmarks/performance/0asw91vl",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9qjg4blz-ctx16384",
   "model": "Qwen3.8-27B-Coder390-LynnStyle-Mixed",
   "notes": "A4toA8 build; MTPFix; oMLX community benchmark trace, single run at 16k context; peak memory 22.1 GB; TTFT 18275 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 896.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Coder390-LynnStyle-Mixed-MLX-A4toA8-MTPFix on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k tokens 896.6 PP tok/s 116.2 TG tok/s 18275 TTFT (ms) 22.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9qjg4blz",
   "source_url": "https://omlx.ai/benchmarks/performance/9qjg4blz",
   "tg_tokens": null,
   "tps": 116.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-5ff33zfs-ctx1024",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "MTP text build; oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: fgaosj6w; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 218.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-oQ4e-MTP-Text on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 218.5 30.7 17.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ff33zfs",
   "source_url": "https://omlx.ai/benchmarks/performance/5ff33zfs",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-5ff33zfs-ctx4096",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "MTP text build; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: fgaosj6w; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 206.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-oQ4e-MTP-Text on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 206.3 33.6 20.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ff33zfs",
   "source_url": "https://omlx.ai/benchmarks/performance/5ff33zfs",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-nan8jud8-ctx1024",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: 3o0t3r13; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 161.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-6bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 1k 161.6 12.7 25.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nan8jud8",
   "source_url": "https://omlx.ai/benchmarks/performance/nan8jud8",
   "tg_tokens": null,
   "tps": 12.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-nan8jud8-ctx4096",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 32.8 GB at this context; duplicate traces of the same session: 3o0t3r13; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 160.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-6bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 4k 160.4 12.6 32.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nan8jud8",
   "source_url": "https://omlx.ai/benchmarks/performance/nan8jud8",
   "tg_tokens": null,
   "tps": 12.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-iqc2d88v-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: 93dshrfc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 302.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 8bit; 1k 302.8 37.7 34.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iqc2d88v",
   "source_url": "https://omlx.ai/benchmarks/performance/iqc2d88v",
   "tg_tokens": null,
   "tps": 37.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-iqc2d88v-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 38.4 GB at this context; duplicate traces of the same session: 93dshrfc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 316.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 8bit; 4k 316.1 30.9 38.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iqc2d88v",
   "source_url": "https://omlx.ai/benchmarks/performance/iqc2d88v",
   "tg_tokens": null,
   "tps": 30.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-fkrbtxde-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 101.5 GB; TTFT 11842 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1384.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k tokens 1,384 PP tok/s 93.5 TG tok/s 11842 TTFT (ms) 101.5 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fkrbtxde",
   "source_url": "https://omlx.ai/benchmarks/performance/fkrbtxde",
   "tg_tokens": null,
   "tps": 93.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-td86eyxr-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 71.1 GB; TTFT 2663 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1538.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 1,538 PP tok/s 63.2 TG tok/s 2663 TTFT (ms) 71.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/td86eyxr",
   "source_url": "https://omlx.ai/benchmarks/performance/td86eyxr",
   "tg_tokens": null,
   "tps": 63.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-74izwpgq-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace, single run at 1k context; peak memory 38.9 GB; TTFT 6751 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 151.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 151.7 PP tok/s 20.0 TG tok/s 6751 TTFT (ms) 38.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/74izwpgq",
   "source_url": "https://omlx.ai/benchmarks/performance/74izwpgq",
   "tg_tokens": null,
   "tps": 20.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-f5zlic8q-ctx1024",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 7.7 GB at this context; duplicate traces of the same session: gzh7ilsc; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 141.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12b-it-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 141.5 21.0 7.7 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f5zlic8q",
   "source_url": "https://omlx.ai/benchmarks/performance/f5zlic8q",
   "tg_tokens": null,
   "tps": 21.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-f5zlic8q-ctx4096",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 7.7 GB at this context; duplicate traces of the same session: gzh7ilsc; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 138.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12b-it-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 138.0 20.5 7.7 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f5zlic8q",
   "source_url": "https://omlx.ai/benchmarks/performance/f5zlic8q",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-yzzdwhfh-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.7 GB at this context; duplicate traces of the same session: cc3w73lx, gfaiz138, j2vqcfe1, qxs48al8; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 653.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 653.2 50.7 16.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yzzdwhfh",
   "source_url": "https://omlx.ai/benchmarks/performance/yzzdwhfh",
   "tg_tokens": null,
   "tps": 50.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-yzzdwhfh-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: cc3w73lx, gfaiz138, j2vqcfe1, qxs48al8; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 652.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 652.7 57.3 17.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yzzdwhfh",
   "source_url": "https://omlx.ai/benchmarks/performance/yzzdwhfh",
   "tg_tokens": null,
   "tps": 57.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-yzzdwhfh-ctx8192",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: cc3w73lx, gfaiz138, j2vqcfe1, qxs48al8; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 628.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 628.9 47.6 17.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yzzdwhfh",
   "source_url": "https://omlx.ai/benchmarks/performance/yzzdwhfh",
   "tg_tokens": null,
   "tps": 47.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-yzzdwhfh-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: cc3w73lx, gfaiz138, j2vqcfe1, qxs48al8; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 590.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 590.2 46.7 17.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yzzdwhfh",
   "source_url": "https://omlx.ai/benchmarks/performance/yzzdwhfh",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-yzzdwhfh-ctx32768",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: cc3w73lx, gfaiz138, j2vqcfe1, qxs48al8; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 525.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 525.5 33.4 18.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yzzdwhfh",
   "source_url": "https://omlx.ai/benchmarks/performance/yzzdwhfh",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-4pmt5zk9-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: rykmq9nn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 784.9,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 784.9 71.2 11.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4pmt5zk9",
   "source_url": "https://omlx.ai/benchmarks/performance/4pmt5zk9",
   "tg_tokens": null,
   "tps": 71.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-4pmt5zk9-ctx16384",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 12.1 GB at this context; duplicate traces of the same session: rykmq9nn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 714.7,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 714.7 57.4 12.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4pmt5zk9",
   "source_url": "https://omlx.ai/benchmarks/performance/4pmt5zk9",
   "tg_tokens": null,
   "tps": 57.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-uvey13y0-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: 9chsv39o; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 775.4,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 775.4 70.6 11.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uvey13y0",
   "source_url": "https://omlx.ai/benchmarks/performance/uvey13y0",
   "tg_tokens": null,
   "tps": 70.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-uvey13y0-ctx16384",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 12.1 GB at this context; duplicate traces of the same session: 9chsv39o; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 713.9,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 713.9 57.7 12.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uvey13y0",
   "source_url": "https://omlx.ai/benchmarks/performance/uvey13y0",
   "tg_tokens": null,
   "tps": 57.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-8dx65da8-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai build; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: seclo8q6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 909.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 909.1 68.9 20.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8dx65da8",
   "source_url": "https://omlx.ai/benchmarks/performance/8dx65da8",
   "tg_tokens": null,
   "tps": 68.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-8dx65da8-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai build; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: seclo8q6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 639.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 639.3 48.5 22.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8dx65da8",
   "source_url": "https://omlx.ai/benchmarks/performance/8dx65da8",
   "tg_tokens": null,
   "tps": 48.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-iqq0b7hl-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai build; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: eii6g42g; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 905.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 905.4 101.3 20.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iqq0b7hl",
   "source_url": "https://omlx.ai/benchmarks/performance/iqq0b7hl",
   "tg_tokens": null,
   "tps": 101.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-iqq0b7hl-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai build; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: eii6g42g; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 763.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 763.5 68.7 22.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iqq0b7hl",
   "source_url": "https://omlx.ai/benchmarks/performance/iqq0b7hl",
   "tg_tokens": null,
   "tps": 68.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-b62kucjv-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai build; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: 0lxa2nid; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 807.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 807.6 86.1 20.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b62kucjv",
   "source_url": "https://omlx.ai/benchmarks/performance/b62kucjv",
   "tg_tokens": null,
   "tps": 86.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-b62kucjv-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai build; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: 0lxa2nid; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 659.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 659.2 52.4 22.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b62kucjv",
   "source_url": "https://omlx.ai/benchmarks/performance/b62kucjv",
   "tg_tokens": null,
   "tps": 52.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-iet0j2ja-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai build; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: 0v50h8rn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 909.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 909.4 72.8 20.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iet0j2ja",
   "source_url": "https://omlx.ai/benchmarks/performance/iet0j2ja",
   "tg_tokens": null,
   "tps": 72.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-iet0j2ja-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai build; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: 0v50h8rn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 794.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 794.2 53.8 22.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iet0j2ja",
   "source_url": "https://omlx.ai/benchmarks/performance/iet0j2ja",
   "tg_tokens": null,
   "tps": 53.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-tc4xkxcl-ctx1024",
   "model": "DeepSeek-R1-Distill-Qwen-32B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 1k context; peak memory 33.1 GB; TTFT 5610 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 182.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "DeepSeek-R1-Distill-Qwen-32B-MLX-8Bit on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 8bit; 1k tokens 182.5 PP tok/s 16.5 TG tok/s 5610 TTFT (ms) 33.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tc4xkxcl",
   "source_url": "https://omlx.ai/benchmarks/performance/tc4xkxcl",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-6f13j3p9-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.7 GB at this context; duplicate traces of the same session: 0458hy7i, 1bxfzt64, 8u453v3j, gp2h0urg, oac5o5p7, r2r3hvwu, uahwz3jj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1560.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,560 71.9 179.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6f13j3p9",
   "source_url": "https://omlx.ai/benchmarks/performance/6f13j3p9",
   "tg_tokens": null,
   "tps": 71.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-6f13j3p9-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.6 GB at this context; duplicate traces of the same session: 0458hy7i, 1bxfzt64, 8u453v3j, gp2h0urg, oac5o5p7, r2r3hvwu, uahwz3jj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2325.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 2,325 68.1 180.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6f13j3p9",
   "source_url": "https://omlx.ai/benchmarks/performance/6f13j3p9",
   "tg_tokens": null,
   "tps": 68.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-6f13j3p9-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.4 GB at this context; duplicate traces of the same session: 0458hy7i, 1bxfzt64, 8u453v3j, gp2h0urg, oac5o5p7, r2r3hvwu, uahwz3jj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2357.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 2,357 113.9 181.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6f13j3p9",
   "source_url": "https://omlx.ai/benchmarks/performance/6f13j3p9",
   "tg_tokens": null,
   "tps": 113.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-6f13j3p9-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.8 GB at this context; duplicate traces of the same session: 0458hy7i, 1bxfzt64, 8u453v3j, gp2h0urg, oac5o5p7, r2r3hvwu, uahwz3jj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2365.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 2,365 140.9 181.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6f13j3p9",
   "source_url": "https://omlx.ai/benchmarks/performance/6f13j3p9",
   "tg_tokens": null,
   "tps": 140.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-6f13j3p9-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 182.5 GB at this context; duplicate traces of the same session: 0458hy7i, 1bxfzt64, 8u453v3j, gp2h0urg, oac5o5p7, r2r3hvwu, uahwz3jj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2358.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,358 114.9 182.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6f13j3p9",
   "source_url": "https://omlx.ai/benchmarks/performance/6f13j3p9",
   "tg_tokens": null,
   "tps": 114.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-6f13j3p9-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 184.0 GB at this context; duplicate traces of the same session: 0458hy7i, 1bxfzt64, 8u453v3j, gp2h0urg, oac5o5p7, r2r3hvwu, uahwz3jj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2328.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,328 89.9 184.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6f13j3p9",
   "source_url": "https://omlx.ai/benchmarks/performance/6f13j3p9",
   "tg_tokens": null,
   "tps": 89.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-6f13j3p9-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 186.9 GB at this context; duplicate traces of the same session: 0458hy7i, 1bxfzt64, 8u453v3j, gp2h0urg, oac5o5p7, r2r3hvwu, uahwz3jj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2280.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,280 142.7 186.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6f13j3p9",
   "source_url": "https://omlx.ai/benchmarks/performance/6f13j3p9",
   "tg_tokens": null,
   "tps": 142.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-6f13j3p9-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 189.8 GB at this context; duplicate traces of the same session: 0458hy7i, 1bxfzt64, 8u453v3j, gp2h0urg, oac5o5p7, r2r3hvwu, uahwz3jj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2225.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 2,225 73.5 189.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6f13j3p9",
   "source_url": "https://omlx.ai/benchmarks/performance/6f13j3p9",
   "tg_tokens": null,
   "tps": 73.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-nnv0d1bi-ctx4096",
   "model": "MiniCPM5-2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 2.3 GB at this context; duplicate traces of the same session: vi34qibs; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 866.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM5-2B-mlx-4Bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 866.8 93.4 2.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nnv0d1bi",
   "source_url": "https://omlx.ai/benchmarks/performance/nnv0d1bi",
   "tg_tokens": null,
   "tps": 93.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-nnv0d1bi-ctx16384",
   "model": "MiniCPM5-2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 2.6 GB at this context; duplicate traces of the same session: vi34qibs; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 616.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM5-2B-mlx-4Bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k 616.0 66.2 2.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nnv0d1bi",
   "source_url": "https://omlx.ai/benchmarks/performance/nnv0d1bi",
   "tg_tokens": null,
   "tps": 66.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-o2rayc1b-ctx4096",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 2.4 GB at this context; duplicate traces of the same session: 5nmfm41k; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 876.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiniCPM5-2B-oQ4e on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 876.9 88.4 2.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o2rayc1b",
   "source_url": "https://omlx.ai/benchmarks/performance/o2rayc1b",
   "tg_tokens": null,
   "tps": 88.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-o2rayc1b-ctx16384",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 2.6 GB at this context; duplicate traces of the same session: 5nmfm41k; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 615.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiniCPM5-2B-oQ4e on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k 615.3 63.6 2.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o2rayc1b",
   "source_url": "https://omlx.ai/benchmarks/performance/o2rayc1b",
   "tg_tokens": null,
   "tps": 63.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-52aljn1m-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 5bs7g2t7, eu5l9z8a, gxf1a5oy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3563.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 3,563 127.5 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/52aljn1m",
   "source_url": "https://omlx.ai/benchmarks/performance/52aljn1m",
   "tg_tokens": null,
   "tps": 127.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-52aljn1m-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 5bs7g2t7, eu5l9z8a, gxf1a5oy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4500.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 4,500 122.0 20.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/52aljn1m",
   "source_url": "https://omlx.ai/benchmarks/performance/52aljn1m",
   "tg_tokens": null,
   "tps": 122.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-52aljn1m-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 5bs7g2t7, eu5l9z8a, gxf1a5oy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4402.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 4,402 119.3 20.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/52aljn1m",
   "source_url": "https://omlx.ai/benchmarks/performance/52aljn1m",
   "tg_tokens": null,
   "tps": 119.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-52aljn1m-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 5bs7g2t7, eu5l9z8a, gxf1a5oy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4064.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 4,064 113.9 21.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/52aljn1m",
   "source_url": "https://omlx.ai/benchmarks/performance/52aljn1m",
   "tg_tokens": null,
   "tps": 113.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-4yjpzkav-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 39.5 GB; TTFT 2788 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1469.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 8bit; 4k tokens 1,469 PP tok/s 73.7 TG tok/s 2788 TTFT (ms) 39.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4yjpzkav",
   "source_url": "https://omlx.ai/benchmarks/performance/4yjpzkav",
   "tg_tokens": null,
   "tps": 73.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-mv9phe9s-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 38.5 GB; TTFT 2809 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1458.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 8bit; 4k tokens 1,458 PP tok/s 70.1 TG tok/s 2809 TTFT (ms) 38.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mv9phe9s",
   "source_url": "https://omlx.ai/benchmarks/performance/mv9phe9s",
   "tg_tokens": null,
   "tps": 70.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-2mivbmyk-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 37.4 GB; TTFT 795 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1289.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 8bit; 1k tokens 1,289 PP tok/s 73.4 TG tok/s 795 TTFT (ms) 37.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2mivbmyk",
   "source_url": "https://omlx.ai/benchmarks/performance/2mivbmyk",
   "tg_tokens": null,
   "tps": 73.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-gzaha84d-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 36.4 GB; TTFT 769 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1331.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 8bit; 1k tokens 1,331 PP tok/s 69.9 TG tok/s 769 TTFT (ms) 36.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gzaha84d",
   "source_url": "https://omlx.ai/benchmarks/performance/gzaha84d",
   "tg_tokens": null,
   "tps": 69.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-7621hssw-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: ctpfl0t9; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 775.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 1k 775.8 72.6 37.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7621hssw",
   "source_url": "https://omlx.ai/benchmarks/performance/7621hssw",
   "tg_tokens": null,
   "tps": 72.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-7621hssw-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.4 GB at this context; duplicate traces of the same session: ctpfl0t9; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 848.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 4k 848.6 74.7 39.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7621hssw",
   "source_url": "https://omlx.ai/benchmarks/performance/7621hssw",
   "tg_tokens": null,
   "tps": 74.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-7621hssw-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.7 GB at this context; duplicate traces of the same session: ctpfl0t9; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 796.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 8k 796.4 70.2 39.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7621hssw",
   "source_url": "https://omlx.ai/benchmarks/performance/7621hssw",
   "tg_tokens": null,
   "tps": 70.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-wc78etio-ctx1024",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 16.0 GB; TTFT 4713 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 217.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled-MLX-4bit on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 1k tokens 217.3 PP tok/s 30.3 TG tok/s 4713 TTFT (ms) 16.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wc78etio",
   "source_url": "https://omlx.ai/benchmarks/performance/wc78etio",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-m382a6zn-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: fs468kje, xi1jme64; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1323.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 8k 1,323 44.7 3.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m382a6zn",
   "source_url": "https://omlx.ai/benchmarks/performance/m382a6zn",
   "tg_tokens": null,
   "tps": 44.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-m382a6zn-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.1 GB at this context; duplicate traces of the same session: fs468kje, xi1jme64; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1204.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 16k 1,204 40.2 4.1 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m382a6zn",
   "source_url": "https://omlx.ai/benchmarks/performance/m382a6zn",
   "tg_tokens": null,
   "tps": 40.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-m382a6zn-ctx32768",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.6 GB at this context; duplicate traces of the same session: fs468kje, xi1jme64; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 912.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 32k 912.8 23.6 4.6 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m382a6zn",
   "source_url": "https://omlx.ai/benchmarks/performance/m382a6zn",
   "tg_tokens": null,
   "tps": 23.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-7u6ulqsw-ctx8192",
   "model": "Qwen3.5-4B-OptiQ",
   "notes": "oMLX community benchmark trace; peak memory 4.7 GB at this context; duplicate traces of the same session: 5edq7563, ma6o3flw; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 727.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-OptiQ-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 8k 727.4 21.3 4.7 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7u6ulqsw",
   "source_url": "https://omlx.ai/benchmarks/performance/7u6ulqsw",
   "tg_tokens": null,
   "tps": 21.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-7u6ulqsw-ctx16384",
   "model": "Qwen3.5-4B-OptiQ",
   "notes": "oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: 5edq7563, ma6o3flw; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 724.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-OptiQ-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 16k 724.8 20.4 4.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7u6ulqsw",
   "source_url": "https://omlx.ai/benchmarks/performance/7u6ulqsw",
   "tg_tokens": null,
   "tps": 20.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-7u6ulqsw-ctx32768",
   "model": "Qwen3.5-4B-OptiQ",
   "notes": "oMLX community benchmark trace; peak memory 5.4 GB at this context; duplicate traces of the same session: 5edq7563, ma6o3flw; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 664.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-OptiQ-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 32k 664.9 16.1 5.4 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7u6ulqsw",
   "source_url": "https://omlx.ai/benchmarks/performance/7u6ulqsw",
   "tg_tokens": null,
   "tps": 16.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1",
   "id": "omlx-uwc9f05g-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: cvdzzvyz; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 39.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 4bit; 1k 39.4 6.1 5.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uwc9f05g",
   "source_url": "https://omlx.ai/benchmarks/performance/uwc9f05g",
   "tg_tokens": null,
   "tps": 6.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1",
   "id": "omlx-uwc9f05g-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: cvdzzvyz; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 39.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 4bit; 4k 39.0 5.7 6.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uwc9f05g",
   "source_url": "https://omlx.ai/benchmarks/performance/uwc9f05g",
   "tg_tokens": null,
   "tps": 5.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-ae5oo8kg-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 28.5 GB; TTFT 4722 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 216.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 8bit; 1k tokens 216.8 PP tok/s 19.4 TG tok/s 4722 TTFT (ms) 28.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ae5oo8kg",
   "source_url": "https://omlx.ai/benchmarks/performance/ae5oo8kg",
   "tg_tokens": null,
   "tps": 19.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-r911tns0-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 73mtho2q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 830.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 830.2 18.3 28.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r911tns0",
   "source_url": "https://omlx.ai/benchmarks/performance/r911tns0",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-r911tns0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: 73mtho2q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 898.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 898.5 18.2 29.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r911tns0",
   "source_url": "https://omlx.ai/benchmarks/performance/r911tns0",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-z7cb2p8k-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 1ehcxvzh; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 228.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 8bit; 1k 228.1 18.6 28.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z7cb2p8k",
   "source_url": "https://omlx.ai/benchmarks/performance/z7cb2p8k",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-z7cb2p8k-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: 1ehcxvzh; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 232.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 8bit; 4k 232.2 18.3 32.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z7cb2p8k",
   "source_url": "https://omlx.ai/benchmarks/performance/z7cb2p8k",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-57aw3pfh-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 85.2 GB; TTFT 6642 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 616.7,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.8-Flash-Next-MLX-3bit-MTP on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 3bit; 4k tokens 616.7 PP tok/s 20.4 TG tok/s 6642 TTFT (ms) 85.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/57aw3pfh",
   "source_url": "https://omlx.ai/benchmarks/performance/57aw3pfh",
   "tg_tokens": null,
   "tps": 20.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-0ecd9tt1-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 87.4 GB; TTFT 6802 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 602.1,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.8-Flash-Next-MLX-3bit-MTP on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 3bit; 4k tokens 602.1 PP tok/s 25.9 TG tok/s 6802 TTFT (ms) 87.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ecd9tt1",
   "source_url": "https://omlx.ai/benchmarks/performance/0ecd9tt1",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-lgxaujq4-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 87.5 GB; TTFT 6804 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 602.0,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.8-Flash-Next-MLX-3bit-MTP on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 3bit; 4k tokens 602.0 PP tok/s 23.2 TG tok/s 6804 TTFT (ms) 87.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lgxaujq4",
   "source_url": "https://omlx.ai/benchmarks/performance/lgxaujq4",
   "tg_tokens": null,
   "tps": 23.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-y9ot23k8-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 87.2 GB; TTFT 6807 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 601.8,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.8-Flash-Next-MLX-3bit-MTP on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 3bit; 4k tokens 601.8 PP tok/s 25.2 TG tok/s 6807 TTFT (ms) 87.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y9ot23k8",
   "source_url": "https://omlx.ai/benchmarks/performance/y9ot23k8",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-imksyzq5-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 87.2 GB; TTFT 6804 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 602.0,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.8-Flash-Next-MLX-3bit-MTP on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 3bit; 4k tokens 602.0 PP tok/s 24.7 TG tok/s 6804 TTFT (ms) 87.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/imksyzq5",
   "source_url": "https://omlx.ai/benchmarks/performance/imksyzq5",
   "tg_tokens": null,
   "tps": 24.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-h86d5i5p-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 87.2 GB; TTFT 6833 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 599.4,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.8-Flash-Next-MLX-3bit-MTP on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 3bit; 4k tokens 599.4 PP tok/s 27.3 TG tok/s 6833 TTFT (ms) 87.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h86d5i5p",
   "source_url": "https://omlx.ai/benchmarks/performance/h86d5i5p",
   "tg_tokens": null,
   "tps": 27.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-l3p48vux-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 87.2 GB; TTFT 6806 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 601.9,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.8-Flash-Next-MLX-3bit-MTP on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 3bit; 4k tokens 601.9 PP tok/s 24.0 TG tok/s 6806 TTFT (ms) 87.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l3p48vux",
   "source_url": "https://omlx.ai/benchmarks/performance/l3p48vux",
   "tg_tokens": null,
   "tps": 24.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-zek55ui1-ctx131072",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 104.7 GB; TTFT 94804 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1383.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k tokens 1,383 PP tok/s 61.4 TG tok/s 94804 TTFT (ms) 104.7 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zek55ui1",
   "source_url": "https://omlx.ai/benchmarks/performance/zek55ui1",
   "tg_tokens": null,
   "tps": 61.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-gbavqg45-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 72.6 GB at this context; duplicate traces of the same session: 75ng2ce9; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 252.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 1k 252.6 11.2 72.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gbavqg45",
   "source_url": "https://omlx.ai/benchmarks/performance/gbavqg45",
   "tg_tokens": null,
   "tps": 11.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-gbavqg45-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 79.3 GB at this context; duplicate traces of the same session: 75ng2ce9; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 289.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 4k 289.6 11.0 79.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gbavqg45",
   "source_url": "https://omlx.ai/benchmarks/performance/gbavqg45",
   "tg_tokens": null,
   "tps": 11.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nweevduc-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 77.6 GB; TTFT 1078 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 949.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k tokens 949.9 PP tok/s 68.3 TG tok/s 1078 TTFT (ms) 77.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nweevduc",
   "source_url": "https://omlx.ai/benchmarks/performance/nweevduc",
   "tg_tokens": null,
   "tps": 68.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-d0r71xbt-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.3 GB at this context; duplicate traces of the same session: p1f2c2b5; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 236.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 1k 236.2 34.5 69.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d0r71xbt",
   "source_url": "https://omlx.ai/benchmarks/performance/d0r71xbt",
   "tg_tokens": null,
   "tps": 34.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-d0r71xbt-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.8 GB at this context; duplicate traces of the same session: p1f2c2b5; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 303.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 4k 303.1 33.9 69.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d0r71xbt",
   "source_url": "https://omlx.ai/benchmarks/performance/d0r71xbt",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dorv1g3v-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 69.2 GB; TTFT 1052 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 973.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 973.5 PP tok/s 64.2 TG tok/s 1052 TTFT (ms) 69.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dorv1g3v",
   "source_url": "https://omlx.ai/benchmarks/performance/dorv1g3v",
   "tg_tokens": null,
   "tps": 64.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-u565xgi7-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: h33k7hzq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 946.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 946.6 63.1 70.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u565xgi7",
   "source_url": "https://omlx.ai/benchmarks/performance/u565xgi7",
   "tg_tokens": null,
   "tps": 63.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-u565xgi7-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: h33k7hzq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1927.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,927 74.8 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u565xgi7",
   "source_url": "https://omlx.ai/benchmarks/performance/u565xgi7",
   "tg_tokens": null,
   "tps": 74.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-zg1s607g-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 102.0 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 908.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 16k 908.6 73.6 102.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zg1s607g",
   "source_url": "https://omlx.ai/benchmarks/performance/zg1s607g",
   "tg_tokens": null,
   "tps": 73.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-zg1s607g-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 104.4 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1006.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 64k 1,006 69.6 104.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zg1s607g",
   "source_url": "https://omlx.ai/benchmarks/performance/zg1s607g",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-50u4blxd-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: rlkgaxuj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1201.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,201 96.3 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/50u4blxd",
   "source_url": "https://omlx.ai/benchmarks/performance/50u4blxd",
   "tg_tokens": null,
   "tps": 96.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-50u4blxd-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: rlkgaxuj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2073.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,073 58.3 70.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/50u4blxd",
   "source_url": "https://omlx.ai/benchmarks/performance/50u4blxd",
   "tg_tokens": null,
   "tps": 58.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-hk98axr2-ctx1024",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: td81j6ev; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 182.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3-Coder-30B-A3B-Instruct-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 182.0 31.3 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hk98axr2",
   "source_url": "https://omlx.ai/benchmarks/performance/hk98axr2",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-hk98axr2-ctx4096",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: td81j6ev; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 173.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3-Coder-30B-A3B-Instruct-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 173.8 25.4 17.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hk98axr2",
   "source_url": "https://omlx.ai/benchmarks/performance/hk98axr2",
   "tg_tokens": null,
   "tps": 25.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-7vhfeeow-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 122.2 GB at this context; duplicate traces of the same session: 8bqq78ac; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1063.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "mlx-community--Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 5bit; 1k 1,063 98.6 122.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7vhfeeow",
   "source_url": "https://omlx.ai/benchmarks/performance/7vhfeeow",
   "tg_tokens": null,
   "tps": 98.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-7vhfeeow-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 126.1 GB at this context; duplicate traces of the same session: 8bqq78ac; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 960.3,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "mlx-community--Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 5bit; 4k 960.3 29.6 126.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7vhfeeow",
   "source_url": "https://omlx.ai/benchmarks/performance/7vhfeeow",
   "tg_tokens": null,
   "tps": 29.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-d7uxhuly-ctx1024",
   "model": "GLM-4.7-Flash",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 16.4 GB; TTFT 1475 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 694.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-4.7-Flash-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 694.2 PP tok/s 93.2 TG tok/s 1475 TTFT (ms) 16.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d7uxhuly",
   "source_url": "https://omlx.ai/benchmarks/performance/d7uxhuly",
   "tg_tokens": null,
   "tps": 93.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-1jx3dtht-ctx8192",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 2.7 GB at this context; duplicate traces of the same session: j8cj7eer, jut8v35w; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1220.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM5-2B-OptiQ-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 8k 1,220 33.8 2.7 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1jx3dtht",
   "source_url": "https://omlx.ai/benchmarks/performance/1jx3dtht",
   "tg_tokens": null,
   "tps": 33.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-1jx3dtht-ctx16384",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 3.1 GB at this context; duplicate traces of the same session: j8cj7eer, jut8v35w; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 951.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM5-2B-OptiQ-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 16k 951.2 31.5 3.1 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1jx3dtht",
   "source_url": "https://omlx.ai/benchmarks/performance/1jx3dtht",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-1jx3dtht-ctx32768",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 3.6 GB at this context; duplicate traces of the same session: j8cj7eer, jut8v35w; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 707.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM5-2B-OptiQ-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 32k 707.6 22.0 3.6 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1jx3dtht",
   "source_url": "https://omlx.ai/benchmarks/performance/1jx3dtht",
   "tg_tokens": null,
   "tps": 22.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-puywsxrb-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 9.5 GB at this context; duplicate traces of the same session: 970gf9nw, a88aa1do, h49k0hy4, pk88yt8q; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 369.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-9B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 1k 369.4 33.3 9.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/puywsxrb",
   "source_url": "https://omlx.ai/benchmarks/performance/puywsxrb",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-puywsxrb-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 10.0 GB at this context; duplicate traces of the same session: 970gf9nw, a88aa1do, h49k0hy4, pk88yt8q; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 413.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-9B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 4k 413.0 33.2 10.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/puywsxrb",
   "source_url": "https://omlx.ai/benchmarks/performance/puywsxrb",
   "tg_tokens": null,
   "tps": 33.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-puywsxrb-ctx8192",
   "model": "Ornith-1.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 10.2 GB at this context; duplicate traces of the same session: 970gf9nw, a88aa1do, h49k0hy4, pk88yt8q; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 462.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-9B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 8k 462.9 32.0 10.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/puywsxrb",
   "source_url": "https://omlx.ai/benchmarks/performance/puywsxrb",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-puywsxrb-ctx16384",
   "model": "Ornith-1.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 10.5 GB at this context; duplicate traces of the same session: 970gf9nw, a88aa1do, h49k0hy4, pk88yt8q; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 460.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-9B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 16k 460.2 30.3 10.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/puywsxrb",
   "source_url": "https://omlx.ai/benchmarks/performance/puywsxrb",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-puywsxrb-ctx32768",
   "model": "Ornith-1.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 11.2 GB at this context; duplicate traces of the same session: 970gf9nw, a88aa1do, h49k0hy4, pk88yt8q; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 425.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-9B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 32k 425.5 27.2 11.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/puywsxrb",
   "source_url": "https://omlx.ai/benchmarks/performance/puywsxrb",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-eqjjvpcp-ctx1024",
   "model": "Qwen3.5-2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 1.9 GB at this context; duplicate traces of the same session: jy6ksp0c; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3548.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-2B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 3,548 100.8 1.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eqjjvpcp",
   "source_url": "https://omlx.ai/benchmarks/performance/eqjjvpcp",
   "tg_tokens": null,
   "tps": 100.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-eqjjvpcp-ctx4096",
   "model": "Qwen3.5-2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 2.2 GB at this context; duplicate traces of the same session: jy6ksp0c; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3934.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-2B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 3,934 97.8 2.2 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eqjjvpcp",
   "source_url": "https://omlx.ai/benchmarks/performance/eqjjvpcp",
   "tg_tokens": null,
   "tps": 97.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-h9ffyar6-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.7 GB at this context; duplicate traces of the same session: 653501ky; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 149.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 4bit; 4k 149.5 20.6 3.7 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h9ffyar6",
   "source_url": "https://omlx.ai/benchmarks/performance/h9ffyar6",
   "tg_tokens": null,
   "tps": 20.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-h9ffyar6-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.3 GB at this context; duplicate traces of the same session: 653501ky; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 136.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 4bit; 16k 136.0 18.1 4.3 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h9ffyar6",
   "source_url": "https://omlx.ai/benchmarks/performance/h9ffyar6",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-vhxp9oxy-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: hjkt18ff; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1474.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 1,474 46.7 3.4 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vhxp9oxy",
   "source_url": "https://omlx.ai/benchmarks/performance/vhxp9oxy",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-vhxp9oxy-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.7 GB at this context; duplicate traces of the same session: hjkt18ff; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1575.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 1,575 44.9 3.7 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vhxp9oxy",
   "source_url": "https://omlx.ai/benchmarks/performance/vhxp9oxy",
   "tg_tokens": null,
   "tps": 44.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-7cwj353c-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: 5mndpk81; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1463.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 1,463 46.7 3.4 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7cwj353c",
   "source_url": "https://omlx.ai/benchmarks/performance/7cwj353c",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-7cwj353c-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.7 GB at this context; duplicate traces of the same session: 5mndpk81; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1577.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 1,577 44.8 3.7 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7cwj353c",
   "source_url": "https://omlx.ai/benchmarks/performance/7cwj353c",
   "tg_tokens": null,
   "tps": 44.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-swf9gogl-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: 5japrk59; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 235.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 235.4 20.9 6.2 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/swf9gogl",
   "source_url": "https://omlx.ai/benchmarks/performance/swf9gogl",
   "tg_tokens": null,
   "tps": 20.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-swf9gogl-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: 5japrk59; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 234.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 8k 234.6 20.4 6.4 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/swf9gogl",
   "source_url": "https://omlx.ai/benchmarks/performance/swf9gogl",
   "tg_tokens": null,
   "tps": 20.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-wyl7mcdk-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "bf16-pt checkpoint; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: 8uuq90fp; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1766.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-bf16-pt-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 1k 1,766 132.5 29.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wyl7mcdk",
   "source_url": "https://omlx.ai/benchmarks/performance/wyl7mcdk",
   "tg_tokens": null,
   "tps": 132.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-wyl7mcdk-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "bf16-pt checkpoint; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.9 GB at this context; duplicate traces of the same session: 8uuq90fp; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2352.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-bf16-pt-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 4k 2,352 135.5 31.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wyl7mcdk",
   "source_url": "https://omlx.ai/benchmarks/performance/wyl7mcdk",
   "tg_tokens": null,
   "tps": 135.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1i55x7tk-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.0 GB at this context; duplicate traces of the same session: 7lgdtisd, wiggzwb5, wj1jpggq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2665.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 1k 2,665 180.0 30.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1i55x7tk",
   "source_url": "https://omlx.ai/benchmarks/performance/1i55x7tk",
   "tg_tokens": null,
   "tps": 180.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1i55x7tk-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.8 GB at this context; duplicate traces of the same session: 7lgdtisd, wiggzwb5, wj1jpggq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3307.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 4k 3,307 165.9 31.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1i55x7tk",
   "source_url": "https://omlx.ai/benchmarks/performance/1i55x7tk",
   "tg_tokens": null,
   "tps": 165.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1i55x7tk-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: 7lgdtisd, wiggzwb5, wj1jpggq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3135.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 8k 3,135 167.4 32.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1i55x7tk",
   "source_url": "https://omlx.ai/benchmarks/performance/1i55x7tk",
   "tg_tokens": null,
   "tps": 167.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1i55x7tk-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.5 GB at this context; duplicate traces of the same session: 7lgdtisd, wiggzwb5, wj1jpggq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2821.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 16k 2,821 187.9 32.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1i55x7tk",
   "source_url": "https://omlx.ai/benchmarks/performance/1i55x7tk",
   "tg_tokens": null,
   "tps": 187.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qx3oyp4n-ctx32768",
   "model": "Qwen3.8-27B-Uncensored-HauhauCS-Aggressive",
   "notes": "MTPLX build; oMLX community benchmark trace, single run at 32k context; peak memory 19.7 GB; TTFT 45669 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 717.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTPLX-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 32k tokens 717.5 PP tok/s 34.5 TG tok/s 45669 TTFT (ms) 19.7 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qx3oyp4n",
   "source_url": "https://omlx.ai/benchmarks/performance/qx3oyp4n",
   "tg_tokens": null,
   "tps": 34.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ec0d53je-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 0shq5913, 6uoi4jao, mqg39fmg, omacwf2d; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 161.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 1k 161.3 11.6 28.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ec0d53je",
   "source_url": "https://omlx.ai/benchmarks/performance/ec0d53je",
   "tg_tokens": null,
   "tps": 11.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-ec0d53je-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: 0shq5913, 6uoi4jao, mqg39fmg, omacwf2d; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 162.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 4k 162.9 11.5 32.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ec0d53je",
   "source_url": "https://omlx.ai/benchmarks/performance/ec0d53je",
   "tg_tokens": null,
   "tps": 11.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ec0d53je-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 33.1 GB at this context; duplicate traces of the same session: 0shq5913, 6uoi4jao, mqg39fmg, omacwf2d; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 157.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 8k 157.7 11.4 33.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ec0d53je",
   "source_url": "https://omlx.ai/benchmarks/performance/ec0d53je",
   "tg_tokens": null,
   "tps": 11.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-ec0d53je-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 34.2 GB at this context; duplicate traces of the same session: 0shq5913, 6uoi4jao, mqg39fmg, omacwf2d; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 152.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 16k 152.6 11.1 34.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ec0d53je",
   "source_url": "https://omlx.ai/benchmarks/performance/ec0d53je",
   "tg_tokens": null,
   "tps": 11.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ec0d53je-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: 0shq5913, 6uoi4jao, mqg39fmg, omacwf2d; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 140.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 32k 140.5 10.6 35.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ec0d53je",
   "source_url": "https://omlx.ai/benchmarks/performance/ec0d53je",
   "tg_tokens": null,
   "tps": 10.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-6t1gkzvh-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 87.3 GB; TTFT 6806 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 601.9,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.8-Flash-Next-MLX-3bit-MTP on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 3bit; 4k tokens 601.9 PP tok/s 23.1 TG tok/s 6806 TTFT (ms) 87.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6t1gkzvh",
   "source_url": "https://omlx.ai/benchmarks/performance/6t1gkzvh",
   "tg_tokens": null,
   "tps": 23.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-ann9zz0z-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 59.3 GB; TTFT 21810 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 187.8,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.8-Flash-Next-MLX-3bit-MTP on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 3bit; 4k tokens 187.8 PP tok/s 8.5 TG tok/s 21810 TTFT (ms) 59.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ann9zz0z",
   "source_url": "https://omlx.ai/benchmarks/performance/ann9zz0z",
   "tg_tokens": null,
   "tps": 8.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-cjw7bc0t-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 57.1 GB; TTFT 21756 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 188.3,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.8-Flash-Next-MLX-3bit-MTP on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 3bit; 4k tokens 188.3 PP tok/s 8.6 TG tok/s 21756 TTFT (ms) 57.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cjw7bc0t",
   "source_url": "https://omlx.ai/benchmarks/performance/cjw7bc0t",
   "tg_tokens": null,
   "tps": 8.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lfp96yms-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: gnhvpukt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 970.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 970.9 44.1 100.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lfp96yms",
   "source_url": "https://omlx.ai/benchmarks/performance/lfp96yms",
   "tg_tokens": null,
   "tps": 44.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-lfp96yms-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: gnhvpukt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1414.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 1,414 92.2 101.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lfp96yms",
   "source_url": "https://omlx.ai/benchmarks/performance/lfp96yms",
   "tg_tokens": null,
   "tps": 92.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-c767nnrj-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.5 GB at this context; duplicate traces of the same session: u9vj7e57; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1130.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,130 75.3 70.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c767nnrj",
   "source_url": "https://omlx.ai/benchmarks/performance/c767nnrj",
   "tg_tokens": null,
   "tps": 75.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-c767nnrj-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.0 GB at this context; duplicate traces of the same session: u9vj7e57; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2927.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,927 85.1 72.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c767nnrj",
   "source_url": "https://omlx.ai/benchmarks/performance/c767nnrj",
   "tg_tokens": null,
   "tps": 85.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4zp4fw61-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 75.3 GB; TTFT 24022 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2728.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k tokens 2,728 PP tok/s 81.3 TG tok/s 24022 TTFT (ms) 75.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4zp4fw61",
   "source_url": "https://omlx.ai/benchmarks/performance/4zp4fw61",
   "tg_tokens": null,
   "tps": 81.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-e8x9s4rd-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 69.2 GB; TTFT 1038 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 986.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 986.8 PP tok/s 62.0 TG tok/s 1038 TTFT (ms) 69.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e8x9s4rd",
   "source_url": "https://omlx.ai/benchmarks/performance/e8x9s4rd",
   "tg_tokens": null,
   "tps": 62.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yayr3x1p-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 69.2 GB; TTFT 859 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1192.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,192 PP tok/s 74.8 TG tok/s 859 TTFT (ms) 69.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yayr3x1p",
   "source_url": "https://omlx.ai/benchmarks/performance/yayr3x1p",
   "tg_tokens": null,
   "tps": 74.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-iszgwjrf-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: 2w6tt6sk; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3934.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 3,934 107.4 101.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iszgwjrf",
   "source_url": "https://omlx.ai/benchmarks/performance/iszgwjrf",
   "tg_tokens": null,
   "tps": 107.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-iszgwjrf-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.0 GB at this context; duplicate traces of the same session: 2w6tt6sk; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3918.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 3,918 143.7 102.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iszgwjrf",
   "source_url": "https://omlx.ai/benchmarks/performance/iszgwjrf",
   "tg_tokens": null,
   "tps": 143.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-jxcmymsn-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.8 GB at this context; duplicate traces of the same session: dau89ca3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3672.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 3,672 160.3 100.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jxcmymsn",
   "source_url": "https://omlx.ai/benchmarks/performance/jxcmymsn",
   "tg_tokens": null,
   "tps": 160.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-jxcmymsn-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.4 GB at this context; duplicate traces of the same session: dau89ca3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3909.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 3,909 126.2 102.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jxcmymsn",
   "source_url": "https://omlx.ai/benchmarks/performance/jxcmymsn",
   "tg_tokens": null,
   "tps": 126.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-lb19hamv-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.9 GB at this context; duplicate traces of the same session: 6sd95ltd, lk774bkv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3778.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 3,778 123.4 100.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lb19hamv",
   "source_url": "https://omlx.ai/benchmarks/performance/lb19hamv",
   "tg_tokens": null,
   "tps": 123.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-lb19hamv-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.2 GB at this context; duplicate traces of the same session: 6sd95ltd, lk774bkv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3868.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 3,868 146.1 101.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lb19hamv",
   "source_url": "https://omlx.ai/benchmarks/performance/lb19hamv",
   "tg_tokens": null,
   "tps": 146.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-lb19hamv-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.1 GB at this context; duplicate traces of the same session: 6sd95ltd, lk774bkv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3880.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 3,880 158.3 102.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lb19hamv",
   "source_url": "https://omlx.ai/benchmarks/performance/lb19hamv",
   "tg_tokens": null,
   "tps": 158.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ak3v9lx1-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.9 GB at this context; duplicate traces of the same session: 5i1327if, ty04gw13; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3531.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 3,531 83.0 100.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ak3v9lx1",
   "source_url": "https://omlx.ai/benchmarks/performance/ak3v9lx1",
   "tg_tokens": null,
   "tps": 83.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ak3v9lx1-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.0 GB at this context; duplicate traces of the same session: 5i1327if, ty04gw13; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3920.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 3,920 138.3 101.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ak3v9lx1",
   "source_url": "https://omlx.ai/benchmarks/performance/ak3v9lx1",
   "tg_tokens": null,
   "tps": 138.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ak3v9lx1-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.3 GB at this context; duplicate traces of the same session: 5i1327if, ty04gw13; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3913.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 3,913 99.1 102.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ak3v9lx1",
   "source_url": "https://omlx.ai/benchmarks/performance/ak3v9lx1",
   "tg_tokens": null,
   "tps": 99.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9o9etfa4-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 102.0 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 908.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 16k 908.6 73.6 102.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9o9etfa4",
   "source_url": "https://omlx.ai/benchmarks/performance/9o9etfa4",
   "tg_tokens": null,
   "tps": 73.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9o9etfa4-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 104.4 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1006.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 64k 1,006 69.6 104.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9o9etfa4",
   "source_url": "https://omlx.ai/benchmarks/performance/9o9etfa4",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-4tldbhz1-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.9 GB at this context; duplicate traces of the same session: 4wzbw4zt, k36deov9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2846.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 2,846 77.2 142.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4tldbhz1",
   "source_url": "https://omlx.ai/benchmarks/performance/4tldbhz1",
   "tg_tokens": null,
   "tps": 77.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-4tldbhz1-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.0 GB at this context; duplicate traces of the same session: 4wzbw4zt, k36deov9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3290.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 16k 3,290 125.6 143.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4tldbhz1",
   "source_url": "https://omlx.ai/benchmarks/performance/4tldbhz1",
   "tg_tokens": null,
   "tps": 125.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-4tldbhz1-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.7 GB at this context; duplicate traces of the same session: 4wzbw4zt, k36deov9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3307.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 32k 3,307 138.4 143.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4tldbhz1",
   "source_url": "https://omlx.ai/benchmarks/performance/4tldbhz1",
   "tg_tokens": null,
   "tps": 138.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-eygze1dp-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 14.5 GB at this context; duplicate traces of the same session: gq9mxtou, n27z6ngu, o89pjxfv, od39nx59; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 683.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 683.3 63.2 14.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eygze1dp",
   "source_url": "https://omlx.ai/benchmarks/performance/eygze1dp",
   "tg_tokens": null,
   "tps": 63.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-eygze1dp-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: gq9mxtou, n27z6ngu, o89pjxfv, od39nx59; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 643.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 643.2 60.9 14.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eygze1dp",
   "source_url": "https://omlx.ai/benchmarks/performance/eygze1dp",
   "tg_tokens": null,
   "tps": 60.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-eygze1dp-ctx8192",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 15.0 GB at this context; duplicate traces of the same session: gq9mxtou, n27z6ngu, o89pjxfv, od39nx59; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 624.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 624.0 58.0 15.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eygze1dp",
   "source_url": "https://omlx.ai/benchmarks/performance/eygze1dp",
   "tg_tokens": null,
   "tps": 58.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-eygze1dp-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 15.4 GB at this context; duplicate traces of the same session: gq9mxtou, n27z6ngu, o89pjxfv, od39nx59; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 590.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 590.8 52.9 15.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eygze1dp",
   "source_url": "https://omlx.ai/benchmarks/performance/eygze1dp",
   "tg_tokens": null,
   "tps": 52.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-eygze1dp-ctx32768",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: gq9mxtou, n27z6ngu, o89pjxfv, od39nx59; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 525.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 525.5 45.3 16.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eygze1dp",
   "source_url": "https://omlx.ai/benchmarks/performance/eygze1dp",
   "tg_tokens": null,
   "tps": 45.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-u0ezna44-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "Jundot build; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 39.9 GB; TTFT 121917 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 268.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Jundot--Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k tokens 268.8 PP tok/s 48.6 TG tok/s 121917 TTFT (ms) 39.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u0ezna44",
   "source_url": "https://omlx.ai/benchmarks/performance/u0ezna44",
   "tg_tokens": null,
   "tps": 48.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ca9uolt8-ctx1024",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: 3tdnzvkd, dbijq25d; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Muse-Glimmer-30B-6bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 1k 106.0 14.2 21.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ca9uolt8",
   "source_url": "https://omlx.ai/benchmarks/performance/ca9uolt8",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ca9uolt8-ctx4096",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 3tdnzvkd, dbijq25d; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Muse-Glimmer-30B-6bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 4k 104.8 13.6 21.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ca9uolt8",
   "source_url": "https://omlx.ai/benchmarks/performance/ca9uolt8",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ca9uolt8-ctx8192",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 3tdnzvkd, dbijq25d; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Muse-Glimmer-30B-6bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 8k 102.9 13.6 22.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ca9uolt8",
   "source_url": "https://omlx.ai/benchmarks/performance/ca9uolt8",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-wziihzjm-ctx4096",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 16.2 GB at this context; duplicate traces of the same session: 846c978n; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 481.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 481.3 17.5 16.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wziihzjm",
   "source_url": "https://omlx.ai/benchmarks/performance/wziihzjm",
   "tg_tokens": null,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-wziihzjm-ctx16384",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: 846c978n; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 487.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 487.5 17.1 16.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wziihzjm",
   "source_url": "https://omlx.ai/benchmarks/performance/wziihzjm",
   "tg_tokens": null,
   "tps": 17.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-x512jusk-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: 78wnxwze, pmxu27ht, td1qgvcc, twge8i84; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 755.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 1k 755.8 65.1 28.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x512jusk",
   "source_url": "https://omlx.ai/benchmarks/performance/x512jusk",
   "tg_tokens": null,
   "tps": 65.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-x512jusk-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.8 GB at this context; duplicate traces of the same session: 78wnxwze, pmxu27ht, td1qgvcc, twge8i84; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 839.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 4k 839.0 63.2 28.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x512jusk",
   "source_url": "https://omlx.ai/benchmarks/performance/x512jusk",
   "tg_tokens": null,
   "tps": 63.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-x512jusk-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.9 GB at this context; duplicate traces of the same session: 78wnxwze, pmxu27ht, td1qgvcc, twge8i84; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 822.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 8k 822.7 59.8 28.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x512jusk",
   "source_url": "https://omlx.ai/benchmarks/performance/x512jusk",
   "tg_tokens": null,
   "tps": 59.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-x512jusk-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.2 GB at this context; duplicate traces of the same session: 78wnxwze, pmxu27ht, td1qgvcc, twge8i84; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 735.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 16k 735.3 56.3 29.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x512jusk",
   "source_url": "https://omlx.ai/benchmarks/performance/x512jusk",
   "tg_tokens": null,
   "tps": 56.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-x512jusk-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: 78wnxwze, pmxu27ht, td1qgvcc, twge8i84; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 607.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 32k 607.0 49.9 29.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x512jusk",
   "source_url": "https://omlx.ai/benchmarks/performance/x512jusk",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-nimk5md6-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 9.4 GB at this context; duplicate traces of the same session: crm8d64h, hhk1w6bw, n1rtnikt, w98sa062; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 364.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-9B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 1k 364.9 34.3 9.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nimk5md6",
   "source_url": "https://omlx.ai/benchmarks/performance/nimk5md6",
   "tg_tokens": null,
   "tps": 34.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-nimk5md6-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 10.0 GB at this context; duplicate traces of the same session: crm8d64h, hhk1w6bw, n1rtnikt, w98sa062; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 434.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-9B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 4k 434.4 33.9 10.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nimk5md6",
   "source_url": "https://omlx.ai/benchmarks/performance/nimk5md6",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-nimk5md6-ctx8192",
   "model": "Ornith-1.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 10.2 GB at this context; duplicate traces of the same session: crm8d64h, hhk1w6bw, n1rtnikt, w98sa062; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 464.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-9B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 8k 464.3 33.3 10.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nimk5md6",
   "source_url": "https://omlx.ai/benchmarks/performance/nimk5md6",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-nimk5md6-ctx16384",
   "model": "Ornith-1.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 10.5 GB at this context; duplicate traces of the same session: crm8d64h, hhk1w6bw, n1rtnikt, w98sa062; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 461.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-9B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 16k 461.2 31.6 10.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nimk5md6",
   "source_url": "https://omlx.ai/benchmarks/performance/nimk5md6",
   "tg_tokens": null,
   "tps": 31.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-nimk5md6-ctx32768",
   "model": "Ornith-1.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 11.2 GB at this context; duplicate traces of the same session: crm8d64h, hhk1w6bw, n1rtnikt, w98sa062; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 426.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-9B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 32k 426.4 28.6 11.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nimk5md6",
   "source_url": "https://omlx.ai/benchmarks/performance/nimk5md6",
   "tg_tokens": null,
   "tps": 28.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-z53tmvmw-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: hw3yrtq7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 226.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 226.0 22.8 16.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z53tmvmw",
   "source_url": "https://omlx.ai/benchmarks/performance/z53tmvmw",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-z53tmvmw-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: hw3yrtq7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 239.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 239.1 22.3 19.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z53tmvmw",
   "source_url": "https://omlx.ai/benchmarks/performance/z53tmvmw",
   "tg_tokens": null,
   "tps": 22.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-sk8ng67r-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: nx7tak4h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 391.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 391.1 16.0 28.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sk8ng67r",
   "source_url": "https://omlx.ai/benchmarks/performance/sk8ng67r",
   "tg_tokens": null,
   "tps": 16.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-sk8ng67r-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: nx7tak4h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 537.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 537.6 16.6 29.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sk8ng67r",
   "source_url": "https://omlx.ai/benchmarks/performance/sk8ng67r",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-fsnatflr-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.7 GB at this context; duplicate traces of the same session: cn4cu4bv, q5db2g8e; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1178.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,178 103.9 100.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fsnatflr",
   "source_url": "https://omlx.ai/benchmarks/performance/fsnatflr",
   "tg_tokens": null,
   "tps": 103.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-fsnatflr-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: cn4cu4bv, q5db2g8e; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1428.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,428 93.1 101.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fsnatflr",
   "source_url": "https://omlx.ai/benchmarks/performance/fsnatflr",
   "tg_tokens": null,
   "tps": 93.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-fsnatflr-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: cn4cu4bv, q5db2g8e; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1411.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 1,411 88.4 101.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fsnatflr",
   "source_url": "https://omlx.ai/benchmarks/performance/fsnatflr",
   "tg_tokens": null,
   "tps": 88.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-x45edoj6-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.9 GB at this context; duplicate traces of the same session: 8so6spvh, b1v2rbnq, sywm9wy3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3737.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 3,737 110.1 100.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x45edoj6",
   "source_url": "https://omlx.ai/benchmarks/performance/x45edoj6",
   "tg_tokens": null,
   "tps": 110.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-x45edoj6-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.0 GB at this context; duplicate traces of the same session: 8so6spvh, b1v2rbnq, sywm9wy3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3864.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 3,864 150.8 101.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x45edoj6",
   "source_url": "https://omlx.ai/benchmarks/performance/x45edoj6",
   "tg_tokens": null,
   "tps": 150.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-x45edoj6-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.8 GB at this context; duplicate traces of the same session: 8so6spvh, b1v2rbnq, sywm9wy3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3909.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 3,909 174.4 102.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x45edoj6",
   "source_url": "https://omlx.ai/benchmarks/performance/x45edoj6",
   "tg_tokens": null,
   "tps": 174.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-x45edoj6-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 108.4 GB at this context; duplicate traces of the same session: 8so6spvh, b1v2rbnq, sywm9wy3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3557.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 3,557 110.3 108.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x45edoj6",
   "source_url": "https://omlx.ai/benchmarks/performance/x45edoj6",
   "tg_tokens": null,
   "tps": 110.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-zg3luluf-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.8 GB at this context; duplicate traces of the same session: 479w7j0n, guilqq36, wm2yuiez; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3732.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 3,732 163.2 100.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zg3luluf",
   "source_url": "https://omlx.ai/benchmarks/performance/zg3luluf",
   "tg_tokens": null,
   "tps": 163.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-zg3luluf-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.2 GB at this context; duplicate traces of the same session: 479w7j0n, guilqq36, wm2yuiez; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3907.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 3,907 145.5 101.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zg3luluf",
   "source_url": "https://omlx.ai/benchmarks/performance/zg3luluf",
   "tg_tokens": null,
   "tps": 145.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-zg3luluf-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.1 GB at this context; duplicate traces of the same session: 479w7j0n, guilqq36, wm2yuiez; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3950.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 3,950 133.3 102.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zg3luluf",
   "source_url": "https://omlx.ai/benchmarks/performance/zg3luluf",
   "tg_tokens": null,
   "tps": 133.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-zg3luluf-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 108.4 GB at this context; duplicate traces of the same session: 479w7j0n, guilqq36, wm2yuiez; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3593.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 3,593 102.5 108.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zg3luluf",
   "source_url": "https://omlx.ai/benchmarks/performance/zg3luluf",
   "tg_tokens": null,
   "tps": 102.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-o9cdo2s3-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.0 GB at this context; duplicate traces of the same session: 8f0b58ot, e3nk0dyn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3144.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 16k 3,144 122.7 143.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9cdo2s3",
   "source_url": "https://omlx.ai/benchmarks/performance/o9cdo2s3",
   "tg_tokens": null,
   "tps": 122.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-o9cdo2s3-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.7 GB at this context; duplicate traces of the same session: 8f0b58ot, e3nk0dyn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3212.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 32k 3,212 132.3 143.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9cdo2s3",
   "source_url": "https://omlx.ai/benchmarks/performance/o9cdo2s3",
   "tg_tokens": null,
   "tps": 132.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-o9cdo2s3-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 149.8 GB at this context; duplicate traces of the same session: 8f0b58ot, e3nk0dyn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3014.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 128k 3,014 87.6 149.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9cdo2s3",
   "source_url": "https://omlx.ai/benchmarks/performance/o9cdo2s3",
   "tg_tokens": null,
   "tps": 87.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-mnw37ty4-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.9 GB at this context; duplicate traces of the same session: 5o44mhgo, nticst1x; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3159.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 3,159 154.3 142.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mnw37ty4",
   "source_url": "https://omlx.ai/benchmarks/performance/mnw37ty4",
   "tg_tokens": null,
   "tps": 154.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-mnw37ty4-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.0 GB at this context; duplicate traces of the same session: 5o44mhgo, nticst1x; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3273.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 16k 3,273 140.2 143.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mnw37ty4",
   "source_url": "https://omlx.ai/benchmarks/performance/mnw37ty4",
   "tg_tokens": null,
   "tps": 140.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-mnw37ty4-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.7 GB at this context; duplicate traces of the same session: 5o44mhgo, nticst1x; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3272.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 32k 3,272 108.5 143.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mnw37ty4",
   "source_url": "https://omlx.ai/benchmarks/performance/mnw37ty4",
   "tg_tokens": null,
   "tps": 108.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-us6o4dlt-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.0 GB at this context; duplicate traces of the same session: 4uqe8ii7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3249.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 16k 3,249 128.9 143.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/us6o4dlt",
   "source_url": "https://omlx.ai/benchmarks/performance/us6o4dlt",
   "tg_tokens": null,
   "tps": 128.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-us6o4dlt-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.8 GB at this context; duplicate traces of the same session: 4uqe8ii7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3269.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 32k 3,269 182.3 143.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/us6o4dlt",
   "source_url": "https://omlx.ai/benchmarks/performance/us6o4dlt",
   "tg_tokens": null,
   "tps": 182.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-mtf23jo1-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.0 GB at this context; duplicate traces of the same session: h171f5l2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3039.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 16k 3,039 134.3 143.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mtf23jo1",
   "source_url": "https://omlx.ai/benchmarks/performance/mtf23jo1",
   "tg_tokens": null,
   "tps": 134.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-mtf23jo1-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.7 GB at this context; duplicate traces of the same session: h171f5l2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3286.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 32k 3,286 149.5 143.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mtf23jo1",
   "source_url": "https://omlx.ai/benchmarks/performance/mtf23jo1",
   "tg_tokens": null,
   "tps": 149.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-am1uzzt3-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.0 GB at this context; duplicate traces of the same session: jyc9xgc9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3267.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 16k 3,267 130.7 143.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/am1uzzt3",
   "source_url": "https://omlx.ai/benchmarks/performance/am1uzzt3",
   "tg_tokens": null,
   "tps": 130.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-am1uzzt3-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.7 GB at this context; duplicate traces of the same session: jyc9xgc9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3270.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 32k 3,270 135.7 143.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/am1uzzt3",
   "source_url": "https://omlx.ai/benchmarks/performance/am1uzzt3",
   "tg_tokens": null,
   "tps": 135.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-2fraht4o-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.0 GB at this context; duplicate traces of the same session: 523ilmgj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3264.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 16k 3,264 125.6 143.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2fraht4o",
   "source_url": "https://omlx.ai/benchmarks/performance/2fraht4o",
   "tg_tokens": null,
   "tps": 125.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-2fraht4o-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.7 GB at this context; duplicate traces of the same session: 523ilmgj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3281.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 32k 3,281 106.3 143.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2fraht4o",
   "source_url": "https://omlx.ai/benchmarks/performance/2fraht4o",
   "tg_tokens": null,
   "tps": 106.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-32y2jpah-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.0 GB at this context; duplicate traces of the same session: u4uvoi87; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3295.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 16k 3,295 112.3 143.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/32y2jpah",
   "source_url": "https://omlx.ai/benchmarks/performance/32y2jpah",
   "tg_tokens": null,
   "tps": 112.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-32y2jpah-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.7 GB at this context; duplicate traces of the same session: u4uvoi87; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3312.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 32k 3,312 142.0 143.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/32y2jpah",
   "source_url": "https://omlx.ai/benchmarks/performance/32y2jpah",
   "tg_tokens": null,
   "tps": 142.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-wa5vi06r-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.0 GB at this context; duplicate traces of the same session: f3t0a6sl; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3268.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 16k 3,268 151.3 143.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wa5vi06r",
   "source_url": "https://omlx.ai/benchmarks/performance/wa5vi06r",
   "tg_tokens": null,
   "tps": 151.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-wa5vi06r-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.7 GB at this context; duplicate traces of the same session: f3t0a6sl; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3274.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 32k 3,274 138.3 143.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wa5vi06r",
   "source_url": "https://omlx.ai/benchmarks/performance/wa5vi06r",
   "tg_tokens": null,
   "tps": 138.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-hvgj3qqv-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace, single run at 32k context; peak memory 72.7 GB; TTFT 41210 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 795.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k tokens 795.1 PP tok/s 62.5 TG tok/s 41210 TTFT (ms) 72.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hvgj3qqv",
   "source_url": "https://omlx.ai/benchmarks/performance/hvgj3qqv",
   "tg_tokens": null,
   "tps": 62.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-497917xo-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace, single run at 32k context; peak memory 102.3 GB; TTFT 40494 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 809.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k tokens 809.2 PP tok/s 71.6 TG tok/s 40494 TTFT (ms) 102.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/497917xo",
   "source_url": "https://omlx.ai/benchmarks/performance/497917xo",
   "tg_tokens": null,
   "tps": 71.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-wsxhbr0t-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: 3mzb79tw, a209x4ft, e8f61xsv, ufva6efg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 663.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 663.3 64.1 14.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wsxhbr0t",
   "source_url": "https://omlx.ai/benchmarks/performance/wsxhbr0t",
   "tg_tokens": null,
   "tps": 64.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-wsxhbr0t-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: 3mzb79tw, a209x4ft, e8f61xsv, ufva6efg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 590.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 590.3 62.4 14.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wsxhbr0t",
   "source_url": "https://omlx.ai/benchmarks/performance/wsxhbr0t",
   "tg_tokens": null,
   "tps": 62.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-wsxhbr0t-ctx8192",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 15.0 GB at this context; duplicate traces of the same session: 3mzb79tw, a209x4ft, e8f61xsv, ufva6efg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 519.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 519.6 60.3 15.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wsxhbr0t",
   "source_url": "https://omlx.ai/benchmarks/performance/wsxhbr0t",
   "tg_tokens": null,
   "tps": 60.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-wsxhbr0t-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 15.4 GB at this context; duplicate traces of the same session: 3mzb79tw, a209x4ft, e8f61xsv, ufva6efg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 498.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 498.9 57.0 15.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wsxhbr0t",
   "source_url": "https://omlx.ai/benchmarks/performance/wsxhbr0t",
   "tg_tokens": null,
   "tps": 57.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-wsxhbr0t-ctx32768",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: 3mzb79tw, a209x4ft, e8f61xsv, ufva6efg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 475.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 475.7 50.1 16.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wsxhbr0t",
   "source_url": "https://omlx.ai/benchmarks/performance/wsxhbr0t",
   "tg_tokens": null,
   "tps": 50.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ylcut9yy-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: s7ld3ppy; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 793.7,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 793.7 72.6 11.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ylcut9yy",
   "source_url": "https://omlx.ai/benchmarks/performance/ylcut9yy",
   "tg_tokens": null,
   "tps": 72.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ylcut9yy-ctx131072",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: s7ld3ppy; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 339.7,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 128k 339.7 24.3 14.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ylcut9yy",
   "source_url": "https://omlx.ai/benchmarks/performance/ylcut9yy",
   "tg_tokens": null,
   "tps": 24.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-9ke5m92r-ctx1024",
   "model": "supergemma4-26b-uncensored",
   "notes": "oMLX community benchmark trace; peak memory 14.1 GB at this context; duplicate traces of the same session: 0r6s5n2w; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 818.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "supergemma4-26b-uncensored-mlx-4bit-v2 on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 818.5 74.7 14.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ke5m92r",
   "source_url": "https://omlx.ai/benchmarks/performance/9ke5m92r",
   "tg_tokens": null,
   "tps": 74.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-9ke5m92r-ctx4096",
   "model": "supergemma4-26b-uncensored",
   "notes": "oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: 0r6s5n2w; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 801.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "supergemma4-26b-uncensored-mlx-4bit-v2 on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 801.0 73.6 14.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ke5m92r",
   "source_url": "https://omlx.ai/benchmarks/performance/9ke5m92r",
   "tg_tokens": null,
   "tps": 73.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-9emhdx82-ctx1024",
   "model": "Laguna XS 2.1",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 14.2 GB; TTFT 1630 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 628.3,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Laguna-XS-2.1-3bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 3bit; 1k tokens 628.3 PP tok/s 73.1 TG tok/s 1630 TTFT (ms) 14.2 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9emhdx82",
   "source_url": "https://omlx.ai/benchmarks/performance/9emhdx82",
   "tg_tokens": null,
   "tps": 73.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-4l6oxp6i-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: fta4e4q9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1491.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 1,491 137.4 22.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4l6oxp6i",
   "source_url": "https://omlx.ai/benchmarks/performance/4l6oxp6i",
   "tg_tokens": null,
   "tps": 137.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-4l6oxp6i-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: fta4e4q9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2141.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 8k 2,141 128.9 24.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4l6oxp6i",
   "source_url": "https://omlx.ai/benchmarks/performance/4l6oxp6i",
   "tg_tokens": null,
   "tps": 128.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-lcwde2pi-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: b8q211lb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1767.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 1,767 105.2 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lcwde2pi",
   "source_url": "https://omlx.ai/benchmarks/performance/lcwde2pi",
   "tg_tokens": null,
   "tps": 105.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-lcwde2pi-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: b8q211lb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2184.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 8k 2,184 96.8 22.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lcwde2pi",
   "source_url": "https://omlx.ai/benchmarks/performance/lcwde2pi",
   "tg_tokens": null,
   "tps": 96.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-zryy5dud-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.2 GB at this context; duplicate traces of the same session: 0fb3drgw, 24j7xm7p, k8b2oap0, xdds41g6; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 731.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 1k 731.5 78.1 30.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zryy5dud",
   "source_url": "https://omlx.ai/benchmarks/performance/zryy5dud",
   "tg_tokens": null,
   "tps": 78.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-zryy5dud-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.0 GB at this context; duplicate traces of the same session: 0fb3drgw, 24j7xm7p, k8b2oap0, xdds41g6; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 814.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 4k 814.5 70.6 31.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zryy5dud",
   "source_url": "https://omlx.ai/benchmarks/performance/zryy5dud",
   "tg_tokens": null,
   "tps": 70.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-zryy5dud-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.1 GB at this context; duplicate traces of the same session: 0fb3drgw, 24j7xm7p, k8b2oap0, xdds41g6; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 789.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 8k 789.5 76.2 31.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zryy5dud",
   "source_url": "https://omlx.ai/benchmarks/performance/zryy5dud",
   "tg_tokens": null,
   "tps": 76.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-zryy5dud-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.4 GB at this context; duplicate traces of the same session: 0fb3drgw, 24j7xm7p, k8b2oap0, xdds41g6; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 704.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 16k 704.2 66.3 31.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zryy5dud",
   "source_url": "https://omlx.ai/benchmarks/performance/zryy5dud",
   "tg_tokens": null,
   "tps": 66.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-zryy5dud-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fixed build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.9 GB at this context; duplicate traces of the same session: 0fb3drgw, 24j7xm7p, k8b2oap0, xdds41g6; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 572.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fixed-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 32k 572.8 95.2 31.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zryy5dud",
   "source_url": "https://omlx.ai/benchmarks/performance/zryy5dud",
   "tg_tokens": null,
   "tps": 95.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ior6sddm-ctx16384",
   "model": "Ornith-1.5-35B-A3B-uncensored",
   "notes": "fixed MTP vision MLX build; oMLX community benchmark trace, single run at 16k context; peak memory 30.5 GB; TTFT 3425 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4784.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-35B-A3B-uncensored-MLX-6bit-fixed-mtp-vision on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k tokens 4,784 PP tok/s 202.3 TG tok/s 3425 TTFT (ms) 30.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ior6sddm",
   "source_url": "https://omlx.ai/benchmarks/performance/ior6sddm",
   "tg_tokens": null,
   "tps": 202.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-z1684dkp-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: rwsgkggj; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 219.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-9B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 219.9 30.2 7.0 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z1684dkp",
   "source_url": "https://omlx.ai/benchmarks/performance/z1684dkp",
   "tg_tokens": null,
   "tps": 30.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-z1684dkp-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.7 GB at this context; duplicate traces of the same session: rwsgkggj; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 222.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-9B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 222.8 28.1 7.7 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z1684dkp",
   "source_url": "https://omlx.ai/benchmarks/performance/z1684dkp",
   "tg_tokens": null,
   "tps": 28.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-gr9dzmxb-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: c1kdew2b; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 215.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-9B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 215.6 26.1 7.0 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gr9dzmxb",
   "source_url": "https://omlx.ai/benchmarks/performance/gr9dzmxb",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-gr9dzmxb-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.7 GB at this context; duplicate traces of the same session: c1kdew2b; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 221.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-9B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 221.9 26.6 7.7 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gr9dzmxb",
   "source_url": "https://omlx.ai/benchmarks/performance/gr9dzmxb",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-i4els5c0-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.7 GB at this context; duplicate traces of the same session: giis663r; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 225.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k 225.6 20.4 5.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i4els5c0",
   "source_url": "https://omlx.ai/benchmarks/performance/i4els5c0",
   "tg_tokens": null,
   "tps": 20.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-i4els5c0-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: giis663r; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 234.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 234.5 19.4 6.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i4els5c0",
   "source_url": "https://omlx.ai/benchmarks/performance/i4els5c0",
   "tg_tokens": null,
   "tps": 19.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jp03bady-ctx1024",
   "model": "Qwen3.6-27B-Abliterated-Heretic-Uncensored",
   "notes": "MLX build; oMLX community benchmark trace, single run at 1k context; peak memory 22.7 GB; TTFT 1218 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 841.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-27B-Abliterated-Heretic-Uncensored-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k tokens 841.0 PP tok/s 22.7 TG tok/s 1218 TTFT (ms) 22.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jp03bady",
   "source_url": "https://omlx.ai/benchmarks/performance/jp03bady",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-hs9rfc3r-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MLX TextOnly build; oMLX community benchmark trace, single run at 1k context; peak memory 12.7 GB; TTFT 4913 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 208.4,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.8-27B-3bit-MLX-TextOnly on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 3bit; 1k tokens 208.4 PP tok/s 10.5 TG tok/s 4913 TTFT (ms) 12.7 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hs9rfc3r",
   "source_url": "https://omlx.ai/benchmarks/performance/hs9rfc3r",
   "tg_tokens": null,
   "tps": 10.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-wdbivh6i-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: s82ujxgl; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 333.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 333.9 39.0 16.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wdbivh6i",
   "source_url": "https://omlx.ai/benchmarks/performance/wdbivh6i",
   "tg_tokens": null,
   "tps": 39.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-wdbivh6i-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: s82ujxgl; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 356.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 356.1 37.7 19.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wdbivh6i",
   "source_url": "https://omlx.ai/benchmarks/performance/wdbivh6i",
   "tg_tokens": null,
   "tps": 37.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xqjpuz9u-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 4t820fv6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 937.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 937.0 33.9 16.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xqjpuz9u",
   "source_url": "https://omlx.ai/benchmarks/performance/xqjpuz9u",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xqjpuz9u-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 4t820fv6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 992.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 992.8 33.3 17.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xqjpuz9u",
   "source_url": "https://omlx.ai/benchmarks/performance/xqjpuz9u",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rznrcnik-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 0s0cxrmn, 417c2fvq, 5xiuvgap, alg3se6g, la8yie4x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 866.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 866.3 32.8 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rznrcnik",
   "source_url": "https://omlx.ai/benchmarks/performance/rznrcnik",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rznrcnik-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 0s0cxrmn, 417c2fvq, 5xiuvgap, alg3se6g, la8yie4x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 986.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 986.0 32.7 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rznrcnik",
   "source_url": "https://omlx.ai/benchmarks/performance/rznrcnik",
   "tg_tokens": null,
   "tps": 32.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rznrcnik-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 0s0cxrmn, 417c2fvq, 5xiuvgap, alg3se6g, la8yie4x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 882.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 882.7 31.8 17.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rznrcnik",
   "source_url": "https://omlx.ai/benchmarks/performance/rznrcnik",
   "tg_tokens": null,
   "tps": 31.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rznrcnik-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 0s0cxrmn, 417c2fvq, 5xiuvgap, alg3se6g, la8yie4x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 795.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 795.9 30.8 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rznrcnik",
   "source_url": "https://omlx.ai/benchmarks/performance/rznrcnik",
   "tg_tokens": null,
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rznrcnik-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 0s0cxrmn, 417c2fvq, 5xiuvgap, alg3se6g, la8yie4x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 775.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 775.7 29.2 18.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rznrcnik",
   "source_url": "https://omlx.ai/benchmarks/performance/rznrcnik",
   "tg_tokens": null,
   "tps": 29.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rznrcnik-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 0s0cxrmn, 417c2fvq, 5xiuvgap, alg3se6g, la8yie4x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 639.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 639.5 24.5 20.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rznrcnik",
   "source_url": "https://omlx.ai/benchmarks/performance/rznrcnik",
   "tg_tokens": null,
   "tps": 24.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-ky4karaw-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; MLX; oMLX community benchmark trace, single run at 4k context; peak memory 37.6 GB; TTFT 19167 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 213.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit-MTP-MLX on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 8bit; 4k tokens 213.7 PP tok/s 21.7 TG tok/s 19167 TTFT (ms) 37.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ky4karaw",
   "source_url": "https://omlx.ai/benchmarks/performance/ky4karaw",
   "tg_tokens": null,
   "tps": 21.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-t0h6xev6-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; MLX; oMLX community benchmark trace, single run at 4k context; peak memory 32.1 GB; TTFT 18660 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 219.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit-MTP-MLX on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 8bit; 4k tokens 219.5 PP tok/s 19.0 TG tok/s 18660 TTFT (ms) 32.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t0h6xev6",
   "source_url": "https://omlx.ai/benchmarks/performance/t0h6xev6",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-ql33xzxb-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; MLX; oMLX community benchmark trace, single run at 4k context; peak memory 37.6 GB; TTFT 19164 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 213.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit-MTP-MLX on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 8bit; 4k tokens 213.7 PP tok/s 25.4 TG tok/s 19164 TTFT (ms) 37.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ql33xzxb",
   "source_url": "https://omlx.ai/benchmarks/performance/ql33xzxb",
   "tg_tokens": null,
   "tps": 25.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-x5sgy90n-ctx16384",
   "model": "Qwen3.8-27B-Coder390-LynnStyle-Mixed",
   "notes": "A4toA8 build; MTPFix; oMLX community benchmark trace, single run at 16k context; peak memory 22.1 GB; TTFT 18160 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 902.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Coder390-LynnStyle-Mixed-MLX-A4toA8-MTPFix on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k tokens 902.2 PP tok/s 94.5 TG tok/s 18160 TTFT (ms) 22.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x5sgy90n",
   "source_url": "https://omlx.ai/benchmarks/performance/x5sgy90n",
   "tg_tokens": null,
   "tps": 94.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-j0oqknh7-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: s5v78xfh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 811.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 811.9 18.3 28.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j0oqknh7",
   "source_url": "https://omlx.ai/benchmarks/performance/j0oqknh7",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-j0oqknh7-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: s5v78xfh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 897.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 897.5 18.1 29.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j0oqknh7",
   "source_url": "https://omlx.ai/benchmarks/performance/j0oqknh7",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-xp5for4m-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: pcn3p050; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 422.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 422.6 26.6 23.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xp5for4m",
   "source_url": "https://omlx.ai/benchmarks/performance/xp5for4m",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-xp5for4m-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: pcn3p050; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 438.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 438.8 23.8 24.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xp5for4m",
   "source_url": "https://omlx.ai/benchmarks/performance/xp5for4m",
   "tg_tokens": null,
   "tps": 23.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qg5d6718-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 856.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 856.7 31.6 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qg5d6718",
   "source_url": "https://omlx.ai/benchmarks/performance/qg5d6718",
   "tg_tokens": null,
   "tps": 31.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qg5d6718-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 952.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 952.6 31.1 17.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qg5d6718",
   "source_url": "https://omlx.ai/benchmarks/performance/qg5d6718",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qg5d6718-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 912.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 912.3 30.6 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qg5d6718",
   "source_url": "https://omlx.ai/benchmarks/performance/qg5d6718",
   "tg_tokens": null,
   "tps": 30.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qg5d6718-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 773.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 773.8 29.4 18.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qg5d6718",
   "source_url": "https://omlx.ai/benchmarks/performance/qg5d6718",
   "tg_tokens": null,
   "tps": 29.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qg5d6718-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 745.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 745.8 27.9 19.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qg5d6718",
   "source_url": "https://omlx.ai/benchmarks/performance/qg5d6718",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qg5d6718-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 594.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 594.8 22.6 21.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qg5d6718",
   "source_url": "https://omlx.ai/benchmarks/performance/qg5d6718",
   "tg_tokens": null,
   "tps": 22.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-9l19h3e3-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 192.3 GB at this context; duplicate traces of the same session: 9olj7trf, mshdbz00; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4303.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ8-MTP on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 16k 4,303 67.5 192.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9l19h3e3",
   "source_url": "https://omlx.ai/benchmarks/performance/9l19h3e3",
   "tg_tokens": null,
   "tps": 67.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-9l19h3e3-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 194.2 GB at this context; duplicate traces of the same session: 9olj7trf, mshdbz00; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4319.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ8-MTP on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 64k 4,319 91.8 194.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9l19h3e3",
   "source_url": "https://omlx.ai/benchmarks/performance/9l19h3e3",
   "tg_tokens": null,
   "tps": 91.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-9l19h3e3-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 198.6 GB at this context; duplicate traces of the same session: 9olj7trf, mshdbz00; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4204.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ8-MTP on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 128k 4,204 81.6 198.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9l19h3e3",
   "source_url": "https://omlx.ai/benchmarks/performance/9l19h3e3",
   "tg_tokens": null,
   "tps": 81.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hmxjrepk-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.7 GB at this context; duplicate traces of the same session: 9ynnwbni, zoz911vu, zp43l8wx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1169.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,169 97.6 100.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hmxjrepk",
   "source_url": "https://omlx.ai/benchmarks/performance/hmxjrepk",
   "tg_tokens": null,
   "tps": 97.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hmxjrepk-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: 9ynnwbni, zoz911vu, zp43l8wx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1428.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,428 100.6 101.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hmxjrepk",
   "source_url": "https://omlx.ai/benchmarks/performance/hmxjrepk",
   "tg_tokens": null,
   "tps": 100.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-hmxjrepk-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: 9ynnwbni, zoz911vu, zp43l8wx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1412.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 1,412 102.4 101.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hmxjrepk",
   "source_url": "https://omlx.ai/benchmarks/performance/hmxjrepk",
   "tg_tokens": null,
   "tps": 102.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hmxjrepk-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.1 GB at this context; duplicate traces of the same session: 9ynnwbni, zoz911vu, zp43l8wx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1401.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 1,401 37.9 102.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hmxjrepk",
   "source_url": "https://omlx.ai/benchmarks/performance/hmxjrepk",
   "tg_tokens": null,
   "tps": 37.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-fa8xl8jv-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: 833168in; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1165.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,165 91.9 100.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fa8xl8jv",
   "source_url": "https://omlx.ai/benchmarks/performance/fa8xl8jv",
   "tg_tokens": null,
   "tps": 91.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-fa8xl8jv-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: 833168in; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1411.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 1,411 88.1 101.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fa8xl8jv",
   "source_url": "https://omlx.ai/benchmarks/performance/fa8xl8jv",
   "tg_tokens": null,
   "tps": 88.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1rss2xsz-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.4 GB at this context; duplicate traces of the same session: 29v56p32, klcwdbv2, s1b90mj0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 250.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 250.3 38.6 71.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1rss2xsz",
   "source_url": "https://omlx.ai/benchmarks/performance/1rss2xsz",
   "tg_tokens": null,
   "tps": 38.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1rss2xsz-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.5 GB at this context; duplicate traces of the same session: 29v56p32, klcwdbv2, s1b90mj0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 333.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 333.9 47.0 72.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1rss2xsz",
   "source_url": "https://omlx.ai/benchmarks/performance/1rss2xsz",
   "tg_tokens": null,
   "tps": 47.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1rss2xsz-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.5 GB at this context; duplicate traces of the same session: 29v56p32, klcwdbv2, s1b90mj0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 529.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 529.7 41.5 72.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1rss2xsz",
   "source_url": "https://omlx.ai/benchmarks/performance/1rss2xsz",
   "tg_tokens": null,
   "tps": 41.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1rss2xsz-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.7 GB at this context; duplicate traces of the same session: 29v56p32, klcwdbv2, s1b90mj0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 673.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 673.4 41.5 72.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1rss2xsz",
   "source_url": "https://omlx.ai/benchmarks/performance/1rss2xsz",
   "tg_tokens": null,
   "tps": 41.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-9idnybov-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.9 GB at this context; duplicate traces of the same session: 55o87532, f41oiz46, h09ngvhs; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3629.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 3,629 154.6 100.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9idnybov",
   "source_url": "https://omlx.ai/benchmarks/performance/9idnybov",
   "tg_tokens": null,
   "tps": 154.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-9idnybov-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.4 GB at this context; duplicate traces of the same session: 55o87532, f41oiz46, h09ngvhs; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3888.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 3,888 164.9 101.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9idnybov",
   "source_url": "https://omlx.ai/benchmarks/performance/9idnybov",
   "tg_tokens": null,
   "tps": 164.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-9idnybov-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.1 GB at this context; duplicate traces of the same session: 55o87532, f41oiz46, h09ngvhs; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3887.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 3,887 142.8 102.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9idnybov",
   "source_url": "https://omlx.ai/benchmarks/performance/9idnybov",
   "tg_tokens": null,
   "tps": 142.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-9idnybov-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 108.8 GB at this context; duplicate traces of the same session: 55o87532, f41oiz46, h09ngvhs; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3418.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 3,418 81.4 108.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9idnybov",
   "source_url": "https://omlx.ai/benchmarks/performance/9idnybov",
   "tg_tokens": null,
   "tps": 81.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-vxodhtow-ctx1024",
   "model": "Ternary-Bonsai-27B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 1k context; peak memory 8.8 GB; TTFT 4597 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 222.8,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-27B-mlx-2bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 2bit; 1k tokens 222.8 PP tok/s 17.4 TG tok/s 4597 TTFT (ms) 8.8 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vxodhtow",
   "source_url": "https://omlx.ai/benchmarks/performance/vxodhtow",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-r4vdkhjc-ctx1024",
   "model": "DeepSeek-V4-Flash",
   "notes": "oMLX community benchmark trace; peak memory 284.4 GB at this context; duplicate traces of the same session: kolg2btk; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 338.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "deepseek-ai-DeepSeek-V4-Flash-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 338.3 8.6 284.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r4vdkhjc",
   "source_url": "https://omlx.ai/benchmarks/performance/r4vdkhjc",
   "tg_tokens": null,
   "tps": 8.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-r4vdkhjc-ctx4096",
   "model": "DeepSeek-V4-Flash",
   "notes": "oMLX community benchmark trace; peak memory 288.6 GB at this context; duplicate traces of the same session: kolg2btk; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 339.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "deepseek-ai-DeepSeek-V4-Flash-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 339.4 8.2 288.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r4vdkhjc",
   "source_url": "https://omlx.ai/benchmarks/performance/r4vdkhjc",
   "tg_tokens": null,
   "tps": 8.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-96a33twq-ctx1024",
   "model": "gemma-4-12B-it-qat",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 8.2 GB at this context; duplicate traces of the same session: yis2uaij; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 136.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-12B-it-qat-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 136.0 22.1 8.2 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/96a33twq",
   "source_url": "https://omlx.ai/benchmarks/performance/96a33twq",
   "tg_tokens": null,
   "tps": 22.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-96a33twq-ctx4096",
   "model": "gemma-4-12B-it-qat",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 8.6 GB at this context; duplicate traces of the same session: yis2uaij; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 137.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-12B-it-qat-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 137.4 13.3 8.6 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/96a33twq",
   "source_url": "https://omlx.ai/benchmarks/performance/96a33twq",
   "tg_tokens": null,
   "tps": 13.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-aul6qghe-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 323.6 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 433.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 433.8 31.4 323.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aul6qghe",
   "source_url": "https://omlx.ai/benchmarks/performance/aul6qghe",
   "tg_tokens": null,
   "tps": 31.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-aul6qghe-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 324.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 552.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 552.7 28.2 324.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aul6qghe",
   "source_url": "https://omlx.ai/benchmarks/performance/aul6qghe",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-aul6qghe-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 325.0 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 526.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 526.9 30.0 325.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aul6qghe",
   "source_url": "https://omlx.ai/benchmarks/performance/aul6qghe",
   "tg_tokens": null,
   "tps": 30.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-aul6qghe-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 325.2 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 518.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k 518.8 35.4 325.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aul6qghe",
   "source_url": "https://omlx.ai/benchmarks/performance/aul6qghe",
   "tg_tokens": null,
   "tps": 35.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-aul6qghe-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 325.4 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 512.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 32k 512.2 35.9 325.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aul6qghe",
   "source_url": "https://omlx.ai/benchmarks/performance/aul6qghe",
   "tg_tokens": null,
   "tps": 35.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-aul6qghe-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 326.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 505.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 64k 505.2 31.0 326.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aul6qghe",
   "source_url": "https://omlx.ai/benchmarks/performance/aul6qghe",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-aul6qghe-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 327.4 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 495.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 128k 495.0 32.4 327.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aul6qghe",
   "source_url": "https://omlx.ai/benchmarks/performance/aul6qghe",
   "tg_tokens": null,
   "tps": 32.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-aul6qghe-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 328.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 486.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 195k 486.0 22.6 328.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aul6qghe",
   "source_url": "https://omlx.ai/benchmarks/performance/aul6qghe",
   "tg_tokens": null,
   "tps": 22.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i8xuevnd-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: 6psa1i9n, 74q8na5r, 8ypagf0y, bzb137me, dyw47v3y, izsoafoc, k51g2naq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4651.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 4,651 113.4 20.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i8xuevnd",
   "source_url": "https://omlx.ai/benchmarks/performance/i8xuevnd",
   "tg_tokens": null,
   "tps": 113.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i8xuevnd-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 6psa1i9n, 74q8na5r, 8ypagf0y, bzb137me, dyw47v3y, izsoafoc, k51g2naq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4498.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 4,498 96.0 21.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i8xuevnd",
   "source_url": "https://omlx.ai/benchmarks/performance/i8xuevnd",
   "tg_tokens": null,
   "tps": 96.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i8xuevnd-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 6psa1i9n, 74q8na5r, 8ypagf0y, bzb137me, dyw47v3y, izsoafoc, k51g2naq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4146.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 4,146 105.9 21.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i8xuevnd",
   "source_url": "https://omlx.ai/benchmarks/performance/i8xuevnd",
   "tg_tokens": null,
   "tps": 105.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i8xuevnd-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 6psa1i9n, 74q8na5r, 8ypagf0y, bzb137me, dyw47v3y, izsoafoc, k51g2naq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4420.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 4,420 99.9 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i8xuevnd",
   "source_url": "https://omlx.ai/benchmarks/performance/i8xuevnd",
   "tg_tokens": null,
   "tps": 99.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i8xuevnd-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: 6psa1i9n, 74q8na5r, 8ypagf0y, bzb137me, dyw47v3y, izsoafoc, k51g2naq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3868.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 3,868 100.3 21.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i8xuevnd",
   "source_url": "https://omlx.ai/benchmarks/performance/i8xuevnd",
   "tg_tokens": null,
   "tps": 100.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i8xuevnd-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 6psa1i9n, 74q8na5r, 8ypagf0y, bzb137me, dyw47v3y, izsoafoc, k51g2naq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3140.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 3,140 84.4 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i8xuevnd",
   "source_url": "https://omlx.ai/benchmarks/performance/i8xuevnd",
   "tg_tokens": null,
   "tps": 84.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i8xuevnd-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 6psa1i9n, 74q8na5r, 8ypagf0y, bzb137me, dyw47v3y, izsoafoc, k51g2naq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2160.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 2,160 65.7 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i8xuevnd",
   "source_url": "https://omlx.ai/benchmarks/performance/i8xuevnd",
   "tg_tokens": null,
   "tps": 65.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i8xuevnd-ctx199680",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: 6psa1i9n, 74q8na5r, 8ypagf0y, bzb137me, dyw47v3y, izsoafoc, k51g2naq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1570.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 195k 1,570 53.1 24.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i8xuevnd",
   "source_url": "https://omlx.ai/benchmarks/performance/i8xuevnd",
   "tg_tokens": null,
   "tps": 53.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5oyzl9bn-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP mlx build; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: coie7sdr, vz225vm3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 830.6,
   "provenance": "community",
   "quant": "UD-Q4_K_XL-AWQ",
   "quote": "Qwen3.8-27B-UD-Q4_K_XL-AWQ-MTP-mlx on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 830.6 39.9 21.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5oyzl9bn",
   "source_url": "https://omlx.ai/benchmarks/performance/5oyzl9bn",
   "tg_tokens": null,
   "tps": 39.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5oyzl9bn-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP mlx build; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: coie7sdr, vz225vm3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 953.7,
   "provenance": "community",
   "quant": "UD-Q4_K_XL-AWQ",
   "quote": "Qwen3.8-27B-UD-Q4_K_XL-AWQ-MTP-mlx on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 953.7 41.2 22.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5oyzl9bn",
   "source_url": "https://omlx.ai/benchmarks/performance/5oyzl9bn",
   "tg_tokens": null,
   "tps": 41.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5oyzl9bn-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP mlx build; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: coie7sdr, vz225vm3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 975.7,
   "provenance": "community",
   "quant": "UD-Q4_K_XL-AWQ",
   "quote": "Qwen3.8-27B-UD-Q4_K_XL-AWQ-MTP-mlx on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 8k 975.7 39.2 22.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5oyzl9bn",
   "source_url": "https://omlx.ai/benchmarks/performance/5oyzl9bn",
   "tg_tokens": null,
   "tps": 39.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-tg5eyd67-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 13.8 GB at this context; duplicate traces of the same session: f7i55kyl, r4cpjod3, ygd7dlpk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 463.4,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-27B-oQ3e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 3bit; 1k 463.4 21.1 13.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tg5eyd67",
   "source_url": "https://omlx.ai/benchmarks/performance/tg5eyd67",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-tg5eyd67-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: f7i55kyl, r4cpjod3, ygd7dlpk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 510.8,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-27B-oQ3e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 3bit; 4k 510.8 20.6 14.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tg5eyd67",
   "source_url": "https://omlx.ai/benchmarks/performance/tg5eyd67",
   "tg_tokens": null,
   "tps": 20.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-tg5eyd67-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 15.4 GB at this context; duplicate traces of the same session: f7i55kyl, r4cpjod3, ygd7dlpk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 491.2,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-27B-oQ3e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 3bit; 16k 491.2 19.3 15.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tg5eyd67",
   "source_url": "https://omlx.ai/benchmarks/performance/tg5eyd67",
   "tg_tokens": null,
   "tps": 19.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-tg5eyd67-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: f7i55kyl, r4cpjod3, ygd7dlpk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 402.8,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-27B-oQ3e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 3bit; 64k 402.8 15.3 18.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tg5eyd67",
   "source_url": "https://omlx.ai/benchmarks/performance/tg5eyd67",
   "tg_tokens": null,
   "tps": 15.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-etdedyam-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 138s0o55, 4vyhpqp0, oo8w945m, pb2auvll; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 856.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 856.7 31.6 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/etdedyam",
   "source_url": "https://omlx.ai/benchmarks/performance/etdedyam",
   "tg_tokens": null,
   "tps": 31.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-etdedyam-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 138s0o55, 4vyhpqp0, oo8w945m, pb2auvll; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 952.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 952.6 31.1 17.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/etdedyam",
   "source_url": "https://omlx.ai/benchmarks/performance/etdedyam",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-etdedyam-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 138s0o55, 4vyhpqp0, oo8w945m, pb2auvll; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 912.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 912.3 30.6 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/etdedyam",
   "source_url": "https://omlx.ai/benchmarks/performance/etdedyam",
   "tg_tokens": null,
   "tps": 30.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-etdedyam-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: 138s0o55, 4vyhpqp0, oo8w945m, pb2auvll; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 773.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 773.8 29.4 18.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/etdedyam",
   "source_url": "https://omlx.ai/benchmarks/performance/etdedyam",
   "tg_tokens": null,
   "tps": 29.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-etdedyam-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 138s0o55, 4vyhpqp0, oo8w945m, pb2auvll; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 745.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 745.8 27.9 19.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/etdedyam",
   "source_url": "https://omlx.ai/benchmarks/performance/etdedyam",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-etdedyam-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 138s0o55, 4vyhpqp0, oo8w945m, pb2auvll; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 594.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 594.8 22.6 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/etdedyam",
   "source_url": "https://omlx.ai/benchmarks/performance/etdedyam",
   "tg_tokens": null,
   "tps": 22.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bm0n74d6-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: 85kghx4v, 9ex1csak, ann60jkj, o6odc8jl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 911.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 911.0 61.6 24.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bm0n74d6",
   "source_url": "https://omlx.ai/benchmarks/performance/bm0n74d6",
   "tg_tokens": null,
   "tps": 61.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bm0n74d6-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: 85kghx4v, 9ex1csak, ann60jkj, o6odc8jl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1161.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 1,161 46.8 25.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bm0n74d6",
   "source_url": "https://omlx.ai/benchmarks/performance/bm0n74d6",
   "tg_tokens": null,
   "tps": 46.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bm0n74d6-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: 85kghx4v, 9ex1csak, ann60jkj, o6odc8jl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1096.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 1,096 52.9 26.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bm0n74d6",
   "source_url": "https://omlx.ai/benchmarks/performance/bm0n74d6",
   "tg_tokens": null,
   "tps": 52.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bm0n74d6-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: 85kghx4v, 9ex1csak, ann60jkj, o6odc8jl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1050.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 1,050 74.3 27.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bm0n74d6",
   "source_url": "https://omlx.ai/benchmarks/performance/bm0n74d6",
   "tg_tokens": null,
   "tps": 74.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bm0n74d6-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.8 GB at this context; duplicate traces of the same session: 85kghx4v, 9ex1csak, ann60jkj, o6odc8jl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 972.3,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 972.3 117.4 29.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bm0n74d6",
   "source_url": "https://omlx.ai/benchmarks/performance/bm0n74d6",
   "tg_tokens": null,
   "tps": 117.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0moe2jf6-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 8ap28gfa, bzi9oyke, etvr8aiu, y9ddn34w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 797.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 797.0 62.8 23.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0moe2jf6",
   "source_url": "https://omlx.ai/benchmarks/performance/0moe2jf6",
   "tg_tokens": null,
   "tps": 62.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0moe2jf6-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: 8ap28gfa, bzi9oyke, etvr8aiu, y9ddn34w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 847.6,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 847.6 52.8 24.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0moe2jf6",
   "source_url": "https://omlx.ai/benchmarks/performance/0moe2jf6",
   "tg_tokens": null,
   "tps": 52.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0moe2jf6-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: 8ap28gfa, bzi9oyke, etvr8aiu, y9ddn34w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 796.5,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 796.5 52.1 25.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0moe2jf6",
   "source_url": "https://omlx.ai/benchmarks/performance/0moe2jf6",
   "tg_tokens": null,
   "tps": 52.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0moe2jf6-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: 8ap28gfa, bzi9oyke, etvr8aiu, y9ddn34w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 759.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 759.2 71.2 26.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0moe2jf6",
   "source_url": "https://omlx.ai/benchmarks/performance/0moe2jf6",
   "tg_tokens": null,
   "tps": 71.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0moe2jf6-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.4 GB at this context; duplicate traces of the same session: 8ap28gfa, bzi9oyke, etvr8aiu, y9ddn34w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 760.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 760.1 125.8 28.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0moe2jf6",
   "source_url": "https://omlx.ai/benchmarks/performance/0moe2jf6",
   "tg_tokens": null,
   "tps": 125.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-qct7poei-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: ntbzc0dh; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1184.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 1k 1,184 33.2 28.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qct7poei",
   "source_url": "https://omlx.ai/benchmarks/performance/qct7poei",
   "tg_tokens": null,
   "tps": 33.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-qct7poei-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: ntbzc0dh; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1441.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 4k 1,441 33.0 29.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qct7poei",
   "source_url": "https://omlx.ai/benchmarks/performance/qct7poei",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5s3xuim6-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.4 GB at this context; duplicate traces of the same session: e3v0hn4s, haum4fee; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1033.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,033 43.0 71.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5s3xuim6",
   "source_url": "https://omlx.ai/benchmarks/performance/5s3xuim6",
   "tg_tokens": null,
   "tps": 43.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5s3xuim6-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.5 GB at this context; duplicate traces of the same session: e3v0hn4s, haum4fee; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 752.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 752.8 27.8 72.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5s3xuim6",
   "source_url": "https://omlx.ai/benchmarks/performance/5s3xuim6",
   "tg_tokens": null,
   "tps": 27.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5s3xuim6-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.7 GB at this context; duplicate traces of the same session: e3v0hn4s, haum4fee; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 834.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 834.2 44.3 72.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5s3xuim6",
   "source_url": "https://omlx.ai/benchmarks/performance/5s3xuim6",
   "tg_tokens": null,
   "tps": 44.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vw3kgaj0-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.4 GB at this context; duplicate traces of the same session: 749vqo2b, saksmo2t, x4gzcqty; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 843.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 843.7 50.4 71.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vw3kgaj0",
   "source_url": "https://omlx.ai/benchmarks/performance/vw3kgaj0",
   "tg_tokens": null,
   "tps": 50.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vw3kgaj0-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.5 GB at this context; duplicate traces of the same session: 749vqo2b, saksmo2t, x4gzcqty; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1059.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,059 44.7 72.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vw3kgaj0",
   "source_url": "https://omlx.ai/benchmarks/performance/vw3kgaj0",
   "tg_tokens": null,
   "tps": 44.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vw3kgaj0-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.5 GB at this context; duplicate traces of the same session: 749vqo2b, saksmo2t, x4gzcqty; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1159.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,159 43.8 72.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vw3kgaj0",
   "source_url": "https://omlx.ai/benchmarks/performance/vw3kgaj0",
   "tg_tokens": null,
   "tps": 43.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vw3kgaj0-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.7 GB at this context; duplicate traces of the same session: 749vqo2b, saksmo2t, x4gzcqty; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1252.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,252 47.4 72.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vw3kgaj0",
   "source_url": "https://omlx.ai/benchmarks/performance/vw3kgaj0",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5mhv50xp-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: sh1ruuce; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1085.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,085 90.4 100.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5mhv50xp",
   "source_url": "https://omlx.ai/benchmarks/performance/5mhv50xp",
   "tg_tokens": null,
   "tps": 90.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5mhv50xp-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.2 GB at this context; duplicate traces of the same session: sh1ruuce; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3126.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 3,126 78.1 102.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5mhv50xp",
   "source_url": "https://omlx.ai/benchmarks/performance/5mhv50xp",
   "tg_tokens": null,
   "tps": 78.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5mhv50xp-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.9 GB at this context; duplicate traces of the same session: sh1ruuce; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2963.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,963 110.5 102.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5mhv50xp",
   "source_url": "https://omlx.ai/benchmarks/performance/5mhv50xp",
   "tg_tokens": null,
   "tps": 110.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-k9izu59e-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.5 GB at this context; duplicate traces of the same session: 227nqbr3, 3z5atq1n; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5691.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 5,691 146.4 102.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k9izu59e",
   "source_url": "https://omlx.ai/benchmarks/performance/k9izu59e",
   "tg_tokens": null,
   "tps": 146.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-k9izu59e-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.1 GB at this context; duplicate traces of the same session: 227nqbr3, 3z5atq1n; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5663.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 5,663 153.5 103.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k9izu59e",
   "source_url": "https://omlx.ai/benchmarks/performance/k9izu59e",
   "tg_tokens": null,
   "tps": 153.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-k9izu59e-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 108.9 GB at this context; duplicate traces of the same session: 227nqbr3, 3z5atq1n; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5385.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,385 109.2 108.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k9izu59e",
   "source_url": "https://omlx.ai/benchmarks/performance/k9izu59e",
   "tg_tokens": null,
   "tps": 109.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-gy0v2py5-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.5 GB at this context; duplicate traces of the same session: 6v2yl358, cfo68egh; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3698.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 3,698 139.5 102.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gy0v2py5",
   "source_url": "https://omlx.ai/benchmarks/performance/gy0v2py5",
   "tg_tokens": null,
   "tps": 139.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-gy0v2py5-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.1 GB at this context; duplicate traces of the same session: 6v2yl358, cfo68egh; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5579.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 5,579 185.0 103.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gy0v2py5",
   "source_url": "https://omlx.ai/benchmarks/performance/gy0v2py5",
   "tg_tokens": null,
   "tps": 185.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-gy0v2py5-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.0 GB at this context; duplicate traces of the same session: 6v2yl358, cfo68egh; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5297.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,297 99.2 109.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gy0v2py5",
   "source_url": "https://omlx.ai/benchmarks/performance/gy0v2py5",
   "tg_tokens": null,
   "tps": 99.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-oiie3d4c-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.3 GB at this context; duplicate traces of the same session: iszgkgfa, or176qbp, xcffj116; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4300.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 4,300 131.0 143.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oiie3d4c",
   "source_url": "https://omlx.ai/benchmarks/performance/oiie3d4c",
   "tg_tokens": null,
   "tps": 131.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-oiie3d4c-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.1 GB at this context; duplicate traces of the same session: iszgkgfa, or176qbp, xcffj116; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4955.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 16k 4,955 124.3 145.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oiie3d4c",
   "source_url": "https://omlx.ai/benchmarks/performance/oiie3d4c",
   "tg_tokens": null,
   "tps": 124.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-oiie3d4c-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.7 GB at this context; duplicate traces of the same session: iszgkgfa, or176qbp, xcffj116; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4937.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 32k 4,937 135.8 145.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oiie3d4c",
   "source_url": "https://omlx.ai/benchmarks/performance/oiie3d4c",
   "tg_tokens": null,
   "tps": 135.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-oiie3d4c-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 150.5 GB at this context; duplicate traces of the same session: iszgkgfa, or176qbp, xcffj116; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4710.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 128k 4,710 102.9 150.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oiie3d4c",
   "source_url": "https://omlx.ai/benchmarks/performance/oiie3d4c",
   "tg_tokens": null,
   "tps": 102.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-iq0zzboe-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.1 GB at this context; duplicate traces of the same session: 0y0htune, n2i0ewzy; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4782.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 16k 4,782 118.4 145.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iq0zzboe",
   "source_url": "https://omlx.ai/benchmarks/performance/iq0zzboe",
   "tg_tokens": null,
   "tps": 118.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-iq0zzboe-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.7 GB at this context; duplicate traces of the same session: 0y0htune, n2i0ewzy; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4921.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 32k 4,921 146.5 145.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iq0zzboe",
   "source_url": "https://omlx.ai/benchmarks/performance/iq0zzboe",
   "tg_tokens": null,
   "tps": 146.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-iq0zzboe-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 150.5 GB at this context; duplicate traces of the same session: 0y0htune, n2i0ewzy; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4700.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 128k 4,700 101.3 150.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iq0zzboe",
   "source_url": "https://omlx.ai/benchmarks/performance/iq0zzboe",
   "tg_tokens": null,
   "tps": 101.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-9aw6o2hk-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: z3daw3bn, zlfs6m5e; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 476.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 476.8 16.7 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9aw6o2hk",
   "source_url": "https://omlx.ai/benchmarks/performance/9aw6o2hk",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-9aw6o2hk-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: z3daw3bn, zlfs6m5e; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 509.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 509.9 16.2 17.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9aw6o2hk",
   "source_url": "https://omlx.ai/benchmarks/performance/9aw6o2hk",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-9aw6o2hk-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: z3daw3bn, zlfs6m5e; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 489.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 489.9 14.0 18.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9aw6o2hk",
   "source_url": "https://omlx.ai/benchmarks/performance/9aw6o2hk",
   "tg_tokens": null,
   "tps": 14.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gu0mesxv-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: jw6yws58; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 741.2,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "mlx-community--Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 741.2 24.4 21.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gu0mesxv",
   "source_url": "https://omlx.ai/benchmarks/performance/gu0mesxv",
   "tg_tokens": null,
   "tps": 24.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-gu0mesxv-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: jw6yws58; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 764.1,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "mlx-community--Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 764.1 23.6 21.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gu0mesxv",
   "source_url": "https://omlx.ai/benchmarks/performance/gu0mesxv",
   "tg_tokens": null,
   "tps": 23.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-7ywpbkqk-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 323.6 GB at this context; duplicate traces of the same session: 1cw3stvt, o8acl1fh, rto9y2xt, tzzqna26, v6lw5ikh, zi70u2kn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 433.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 433.8 31.4 323.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ywpbkqk",
   "source_url": "https://omlx.ai/benchmarks/performance/7ywpbkqk",
   "tg_tokens": null,
   "tps": 31.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-7ywpbkqk-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 324.7 GB at this context; duplicate traces of the same session: 1cw3stvt, o8acl1fh, rto9y2xt, tzzqna26, v6lw5ikh, zi70u2kn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 552.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 552.7 28.2 324.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ywpbkqk",
   "source_url": "https://omlx.ai/benchmarks/performance/7ywpbkqk",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-7ywpbkqk-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 325.0 GB at this context; duplicate traces of the same session: 1cw3stvt, o8acl1fh, rto9y2xt, tzzqna26, v6lw5ikh, zi70u2kn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 526.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 526.9 30.0 325.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ywpbkqk",
   "source_url": "https://omlx.ai/benchmarks/performance/7ywpbkqk",
   "tg_tokens": null,
   "tps": 30.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-7ywpbkqk-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 325.2 GB at this context; duplicate traces of the same session: 1cw3stvt, o8acl1fh, rto9y2xt, tzzqna26, v6lw5ikh, zi70u2kn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 518.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k 518.8 35.4 325.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ywpbkqk",
   "source_url": "https://omlx.ai/benchmarks/performance/7ywpbkqk",
   "tg_tokens": null,
   "tps": 35.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-7ywpbkqk-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 325.4 GB at this context; duplicate traces of the same session: 1cw3stvt, o8acl1fh, rto9y2xt, tzzqna26, v6lw5ikh, zi70u2kn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 512.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 32k 512.2 35.9 325.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ywpbkqk",
   "source_url": "https://omlx.ai/benchmarks/performance/7ywpbkqk",
   "tg_tokens": null,
   "tps": 35.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-7ywpbkqk-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 326.1 GB at this context; duplicate traces of the same session: 1cw3stvt, o8acl1fh, rto9y2xt, tzzqna26, v6lw5ikh, zi70u2kn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 505.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 64k 505.2 31.0 326.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ywpbkqk",
   "source_url": "https://omlx.ai/benchmarks/performance/7ywpbkqk",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-7ywpbkqk-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 327.4 GB at this context; duplicate traces of the same session: 1cw3stvt, o8acl1fh, rto9y2xt, tzzqna26, v6lw5ikh, zi70u2kn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 495.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 128k 495.0 32.4 327.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ywpbkqk",
   "source_url": "https://omlx.ai/benchmarks/performance/7ywpbkqk",
   "tg_tokens": null,
   "tps": 32.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-7ywpbkqk-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 328.7 GB at this context; duplicate traces of the same session: 1cw3stvt, o8acl1fh, rto9y2xt, tzzqna26, v6lw5ikh, zi70u2kn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 486.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 195k 486.0 22.6 328.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ywpbkqk",
   "source_url": "https://omlx.ai/benchmarks/performance/7ywpbkqk",
   "tg_tokens": null,
   "tps": 22.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-ceba8p1d-ctx8192",
   "model": "Jev-Style-Qwen3.5-2B-Decision",
   "notes": "oMLX community benchmark trace, single run at 8k context; peak memory 4.8 GB; TTFT 3322 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2466.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Jev-Style-Qwen3.5-2B-Decision on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 bf16; 8k tokens 2,466 PP tok/s 31.0 TG tok/s 3322 TTFT (ms) 4.8 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ceba8p1d",
   "source_url": "https://omlx.ai/benchmarks/performance/ceba8p1d",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-0pb4ld3e-ctx1024",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 6.5 GB; TTFT 4217 ms; duplicate traces of the same session: none; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 242.8,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-mlx-mxfp4 on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 4bit; 1k tokens 242.8 PP tok/s 34.1 TG tok/s 4217 TTFT (ms) 6.5 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0pb4ld3e",
   "source_url": "https://omlx.ai/benchmarks/performance/0pb4ld3e",
   "tg_tokens": null,
   "tps": 34.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-len1coji-ctx4096",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 15.1 GB; TTFT 20216 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 202.6,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Muse-Glimmer-30B-mxfp4 on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 4k tokens 202.6 PP tok/s 26.6 TG tok/s 20216 TTFT (ms) 15.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/len1coji",
   "source_url": "https://omlx.ai/benchmarks/performance/len1coji",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-recajhn2-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: u6l6wwdd; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 931.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 931.9 57.9 19.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/recajhn2",
   "source_url": "https://omlx.ai/benchmarks/performance/recajhn2",
   "tg_tokens": null,
   "tps": 57.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-recajhn2-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: u6l6wwdd; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 960.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 960.1 75.4 20.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/recajhn2",
   "source_url": "https://omlx.ai/benchmarks/performance/recajhn2",
   "tg_tokens": null,
   "tps": 75.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4mmk7706-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 145mjmf7, 7cfo24yz, cadehynk, hnyew9ms, w65b5848, ytf1od8a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4602.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 4,602 140.9 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4mmk7706",
   "source_url": "https://omlx.ai/benchmarks/performance/4mmk7706",
   "tg_tokens": null,
   "tps": 140.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4mmk7706-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 145mjmf7, 7cfo24yz, cadehynk, hnyew9ms, w65b5848, ytf1od8a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4451.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 4,451 130.0 20.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4mmk7706",
   "source_url": "https://omlx.ai/benchmarks/performance/4mmk7706",
   "tg_tokens": null,
   "tps": 130.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4mmk7706-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: 145mjmf7, 7cfo24yz, cadehynk, hnyew9ms, w65b5848, ytf1od8a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4084.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 4,084 118.4 20.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4mmk7706",
   "source_url": "https://omlx.ai/benchmarks/performance/4mmk7706",
   "tg_tokens": null,
   "tps": 118.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4mmk7706-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 145mjmf7, 7cfo24yz, cadehynk, hnyew9ms, w65b5848, ytf1od8a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3586.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 3,586 113.9 20.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4mmk7706",
   "source_url": "https://omlx.ai/benchmarks/performance/4mmk7706",
   "tg_tokens": null,
   "tps": 113.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4mmk7706-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 145mjmf7, 7cfo24yz, cadehynk, hnyew9ms, w65b5848, ytf1od8a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2897.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 2,897 98.9 21.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4mmk7706",
   "source_url": "https://omlx.ai/benchmarks/performance/4mmk7706",
   "tg_tokens": null,
   "tps": 98.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4mmk7706-ctx131072",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: 145mjmf7, 7cfo24yz, cadehynk, hnyew9ms, w65b5848, ytf1od8a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1987.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 1,987 76.1 22.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4mmk7706",
   "source_url": "https://omlx.ai/benchmarks/performance/4mmk7706",
   "tg_tokens": null,
   "tps": 76.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4mmk7706-ctx199680",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: 145mjmf7, 7cfo24yz, cadehynk, hnyew9ms, w65b5848, ytf1od8a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1578.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 195k 1,578 59.5 23.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4mmk7706",
   "source_url": "https://omlx.ai/benchmarks/performance/4mmk7706",
   "tg_tokens": null,
   "tps": 59.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8nw1jwvu-ctx1024",
   "model": "Qwen2.5-32B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 17.9 GB at this context; duplicate traces of the same session: opkkyvti; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 787.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-32B-Instruct-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 787.3 28.0 17.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8nw1jwvu",
   "source_url": "https://omlx.ai/benchmarks/performance/8nw1jwvu",
   "tg_tokens": null,
   "tps": 28.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8nw1jwvu-ctx4096",
   "model": "Qwen2.5-32B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: opkkyvti; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 807.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-32B-Instruct-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 807.3 26.7 19.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8nw1jwvu",
   "source_url": "https://omlx.ai/benchmarks/performance/8nw1jwvu",
   "tg_tokens": null,
   "tps": 26.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9fpm88qc-ctx1024",
   "model": "Qwen2.5-32B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 17.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 734.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-32B-Instruct-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 734.9 28.1 17.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9fpm88qc",
   "source_url": "https://omlx.ai/benchmarks/performance/9fpm88qc",
   "tg_tokens": null,
   "tps": 28.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9fpm88qc-ctx4096",
   "model": "Qwen2.5-32B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 794.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-32B-Instruct-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 794.1 26.9 19.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9fpm88qc",
   "source_url": "https://omlx.ai/benchmarks/performance/9fpm88qc",
   "tg_tokens": null,
   "tps": 26.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-clxpsg4r-ctx1024",
   "model": "Qwen3.5-0.8B",
   "notes": "MLX MTP build; oMLX community benchmark trace; peak memory 2.2 GB at this context; duplicate traces of the same session: y9xqtvz5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5203.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.5-0.8B-MLX-MTP-5bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 5,203 141.7 2.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/clxpsg4r",
   "source_url": "https://omlx.ai/benchmarks/performance/clxpsg4r",
   "tg_tokens": null,
   "tps": 141.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-clxpsg4r-ctx4096",
   "model": "Qwen3.5-0.8B",
   "notes": "MLX MTP build; oMLX community benchmark trace; peak memory 2.5 GB at this context; duplicate traces of the same session: y9xqtvz5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 17974.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.5-0.8B-MLX-MTP-5bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 17,974 331.5 2.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/clxpsg4r",
   "source_url": "https://omlx.ai/benchmarks/performance/clxpsg4r",
   "tg_tokens": null,
   "tps": 331.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-clxpsg4r-ctx8192",
   "model": "Qwen3.5-0.8B",
   "notes": "MLX MTP build; oMLX community benchmark trace; peak memory 2.6 GB at this context; duplicate traces of the same session: y9xqtvz5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 19711.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.5-0.8B-MLX-MTP-5bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 19,711 389.0 2.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/clxpsg4r",
   "source_url": "https://omlx.ai/benchmarks/performance/clxpsg4r",
   "tg_tokens": null,
   "tps": 389.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-clxpsg4r-ctx32768",
   "model": "Qwen3.5-0.8B",
   "notes": "MLX MTP build; oMLX community benchmark trace; peak memory 3.1 GB at this context; duplicate traces of the same session: y9xqtvz5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 17226.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.5-0.8B-MLX-MTP-5bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 17,226 622.0 3.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/clxpsg4r",
   "source_url": "https://omlx.ai/benchmarks/performance/clxpsg4r",
   "tg_tokens": null,
   "tps": 622.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-clxpsg4r-ctx65536",
   "model": "Qwen3.5-0.8B",
   "notes": "MLX MTP build; oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: y9xqtvz5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 12634.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.5-0.8B-MLX-MTP-5bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 12,634 275.6 3.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/clxpsg4r",
   "source_url": "https://omlx.ai/benchmarks/performance/clxpsg4r",
   "tg_tokens": null,
   "tps": 275.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-clxpsg4r-ctx131072",
   "model": "Qwen3.5-0.8B",
   "notes": "MLX MTP build; oMLX community benchmark trace; peak memory 5.5 GB at this context; duplicate traces of the same session: y9xqtvz5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 8002.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.5-0.8B-MLX-MTP-5bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 8,002 227.0 5.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/clxpsg4r",
   "source_url": "https://omlx.ai/benchmarks/performance/clxpsg4r",
   "tg_tokens": null,
   "tps": 227.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-clxpsg4r-ctx199680",
   "model": "Qwen3.5-0.8B",
   "notes": "MLX MTP build; oMLX community benchmark trace; peak memory 7.2 GB at this context; duplicate traces of the same session: y9xqtvz5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5736.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.5-0.8B-MLX-MTP-5bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 195k 5,736 158.7 7.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/clxpsg4r",
   "source_url": "https://omlx.ai/benchmarks/performance/clxpsg4r",
   "tg_tokens": null,
   "tps": 158.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8tg0xmqs-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: 4s7x26wu, 8otgfvim; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4699.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 4,699 114.6 20.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8tg0xmqs",
   "source_url": "https://omlx.ai/benchmarks/performance/8tg0xmqs",
   "tg_tokens": null,
   "tps": 114.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8tg0xmqs-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: 4s7x26wu, 8otgfvim; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4689.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 4,689 99.6 21.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8tg0xmqs",
   "source_url": "https://omlx.ai/benchmarks/performance/8tg0xmqs",
   "tg_tokens": null,
   "tps": 99.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8tg0xmqs-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 4s7x26wu, 8otgfvim; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3556.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 3,556 85.1 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8tg0xmqs",
   "source_url": "https://omlx.ai/benchmarks/performance/8tg0xmqs",
   "tg_tokens": null,
   "tps": 85.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wx6emssh-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 35.7 GB at this context; duplicate traces of the same session: 37mppr1z, iudbsqr7, yrm2ewns; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3580.0,
   "provenance": "community",
   "quant": "UD-Q8_K_XL",
   "quote": "Qwen3.6-35B-A3B-UD-Q8_K_XL-mlx on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 3,580 94.5 35.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wx6emssh",
   "source_url": "https://omlx.ai/benchmarks/performance/wx6emssh",
   "tg_tokens": null,
   "tps": 94.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wx6emssh-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 36.4 GB at this context; duplicate traces of the same session: 37mppr1z, iudbsqr7, yrm2ewns; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3182.0,
   "provenance": "community",
   "quant": "UD-Q8_K_XL",
   "quote": "Qwen3.6-35B-A3B-UD-Q8_K_XL-mlx on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 3,182 96.2 36.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wx6emssh",
   "source_url": "https://omlx.ai/benchmarks/performance/wx6emssh",
   "tg_tokens": null,
   "tps": 96.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wx6emssh-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 37.0 GB at this context; duplicate traces of the same session: 37mppr1z, iudbsqr7, yrm2ewns; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4511.0,
   "provenance": "community",
   "quant": "UD-Q8_K_XL",
   "quote": "Qwen3.6-35B-A3B-UD-Q8_K_XL-mlx on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 32k 4,511 86.1 37.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wx6emssh",
   "source_url": "https://omlx.ai/benchmarks/performance/wx6emssh",
   "tg_tokens": null,
   "tps": 86.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wx6emssh-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 37.6 GB at this context; duplicate traces of the same session: 37mppr1z, iudbsqr7, yrm2ewns; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3322.0,
   "provenance": "community",
   "quant": "UD-Q8_K_XL",
   "quote": "Qwen3.6-35B-A3B-UD-Q8_K_XL-mlx on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 64k 3,322 75.1 37.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wx6emssh",
   "source_url": "https://omlx.ai/benchmarks/performance/wx6emssh",
   "tg_tokens": null,
   "tps": 75.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-p1rvr4gg-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: zrlwkn9e; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1216.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k 1,216 54.3 16.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p1rvr4gg",
   "source_url": "https://omlx.ai/benchmarks/performance/p1rvr4gg",
   "tg_tokens": null,
   "tps": 54.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-p1rvr4gg-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: zrlwkn9e; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1480.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 4k 1,480 53.2 17.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p1rvr4gg",
   "source_url": "https://omlx.ai/benchmarks/performance/p1rvr4gg",
   "tg_tokens": null,
   "tps": 53.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-7fzr31k9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: xl5jcfks; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 468.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 468.1 18.0 16.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7fzr31k9",
   "source_url": "https://omlx.ai/benchmarks/performance/7fzr31k9",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-7fzr31k9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: xl5jcfks; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 505.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 505.4 17.7 17.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7fzr31k9",
   "source_url": "https://omlx.ai/benchmarks/performance/7fzr31k9",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-4rqyr19t-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 0oqism36; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 476.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 476.8 17.7 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4rqyr19t",
   "source_url": "https://omlx.ai/benchmarks/performance/4rqyr19t",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-4rqyr19t-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 0oqism36; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 441.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 441.7 17.0 17.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4rqyr19t",
   "source_url": "https://omlx.ai/benchmarks/performance/4rqyr19t",
   "tg_tokens": null,
   "tps": 17.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-g5plbd59-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 929.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 929.2 32.2 16.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g5plbd59",
   "source_url": "https://omlx.ai/benchmarks/performance/g5plbd59",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-g5plbd59-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1011.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,011 31.4 17.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g5plbd59",
   "source_url": "https://omlx.ai/benchmarks/performance/g5plbd59",
   "tg_tokens": null,
   "tps": 31.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-w6pvtwjw-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 29.9 GB; TTFT 68465 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 478.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k tokens 478.6 PP tok/s 13.9 TG tok/s 68465 TTFT (ms) 29.9 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w6pvtwjw",
   "source_url": "https://omlx.ai/benchmarks/performance/w6pvtwjw",
   "tg_tokens": null,
   "tps": 13.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-m0ha3zoq-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1085.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,085 90.4 100.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m0ha3zoq",
   "source_url": "https://omlx.ai/benchmarks/performance/m0ha3zoq",
   "tg_tokens": null,
   "tps": 90.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-m0ha3zoq-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3126.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 3,126 78.1 102.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m0ha3zoq",
   "source_url": "https://omlx.ai/benchmarks/performance/m0ha3zoq",
   "tg_tokens": null,
   "tps": 78.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-m0ha3zoq-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2963.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,963 110.5 102.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m0ha3zoq",
   "source_url": "https://omlx.ai/benchmarks/performance/m0ha3zoq",
   "tg_tokens": null,
   "tps": 110.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4t2i9vk3-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: gf5c9opr, nbqwqk1x, oyelh8p5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1745.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,745 68.6 100.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4t2i9vk3",
   "source_url": "https://omlx.ai/benchmarks/performance/4t2i9vk3",
   "tg_tokens": null,
   "tps": 68.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4t2i9vk3-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.1 GB at this context; duplicate traces of the same session: gf5c9opr, nbqwqk1x, oyelh8p5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2699.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,699 59.4 102.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4t2i9vk3",
   "source_url": "https://omlx.ai/benchmarks/performance/4t2i9vk3",
   "tg_tokens": null,
   "tps": 59.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4t2i9vk3-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.9 GB at this context; duplicate traces of the same session: gf5c9opr, nbqwqk1x, oyelh8p5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2558.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,558 71.6 102.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4t2i9vk3",
   "source_url": "https://omlx.ai/benchmarks/performance/4t2i9vk3",
   "tg_tokens": null,
   "tps": 71.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4t2i9vk3-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.7 GB at this context; duplicate traces of the same session: gf5c9opr, nbqwqk1x, oyelh8p5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2461.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,461 81.4 103.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4t2i9vk3",
   "source_url": "https://omlx.ai/benchmarks/performance/4t2i9vk3",
   "tg_tokens": null,
   "tps": 81.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-j4sr65e7-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: s91vj2ub, vmszike2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1755.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,755 81.9 100.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j4sr65e7",
   "source_url": "https://omlx.ai/benchmarks/performance/j4sr65e7",
   "tg_tokens": null,
   "tps": 81.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-j4sr65e7-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.1 GB at this context; duplicate traces of the same session: s91vj2ub, vmszike2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2023.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,023 85.3 102.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j4sr65e7",
   "source_url": "https://omlx.ai/benchmarks/performance/j4sr65e7",
   "tg_tokens": null,
   "tps": 85.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-j4sr65e7-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.9 GB at this context; duplicate traces of the same session: s91vj2ub, vmszike2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2611.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,611 62.1 102.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j4sr65e7",
   "source_url": "https://omlx.ai/benchmarks/performance/j4sr65e7",
   "tg_tokens": null,
   "tps": 62.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-paa0dwkn-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 37ul4ayv, j71vgqbm, o8lb3pvl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 545.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 545.3 50.8 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/paa0dwkn",
   "source_url": "https://omlx.ai/benchmarks/performance/paa0dwkn",
   "tg_tokens": null,
   "tps": 50.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-paa0dwkn-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: 37ul4ayv, j71vgqbm, o8lb3pvl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1321.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,321 63.7 71.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/paa0dwkn",
   "source_url": "https://omlx.ai/benchmarks/performance/paa0dwkn",
   "tg_tokens": null,
   "tps": 63.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-paa0dwkn-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.0 GB at this context; duplicate traces of the same session: 37ul4ayv, j71vgqbm, o8lb3pvl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2062.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,062 74.7 72.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/paa0dwkn",
   "source_url": "https://omlx.ai/benchmarks/performance/paa0dwkn",
   "tg_tokens": null,
   "tps": 74.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-paa0dwkn-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.0 GB at this context; duplicate traces of the same session: 37ul4ayv, j71vgqbm, o8lb3pvl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2523.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,523 46.2 73.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/paa0dwkn",
   "source_url": "https://omlx.ai/benchmarks/performance/paa0dwkn",
   "tg_tokens": null,
   "tps": 46.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-kijwadqj-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: fn9mwlsg, lgic20pw, m4j9d6ps; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4917.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 4,917 149.1 101.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kijwadqj",
   "source_url": "https://omlx.ai/benchmarks/performance/kijwadqj",
   "tg_tokens": null,
   "tps": 149.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-kijwadqj-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.5 GB at this context; duplicate traces of the same session: fn9mwlsg, lgic20pw, m4j9d6ps; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5678.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 5,678 133.1 102.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kijwadqj",
   "source_url": "https://omlx.ai/benchmarks/performance/kijwadqj",
   "tg_tokens": null,
   "tps": 133.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-kijwadqj-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.1 GB at this context; duplicate traces of the same session: fn9mwlsg, lgic20pw, m4j9d6ps; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5659.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 5,659 160.1 103.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kijwadqj",
   "source_url": "https://omlx.ai/benchmarks/performance/kijwadqj",
   "tg_tokens": null,
   "tps": 160.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-kijwadqj-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 108.9 GB at this context; duplicate traces of the same session: fn9mwlsg, lgic20pw, m4j9d6ps; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5383.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,383 106.1 108.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kijwadqj",
   "source_url": "https://omlx.ai/benchmarks/performance/kijwadqj",
   "tg_tokens": null,
   "tps": 106.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-p1fmx2h7-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.5 GB at this context; duplicate traces of the same session: mcir2fyj, wfs2tv4r; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5581.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 5,581 142.6 102.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p1fmx2h7",
   "source_url": "https://omlx.ai/benchmarks/performance/p1fmx2h7",
   "tg_tokens": null,
   "tps": 142.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-p1fmx2h7-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.1 GB at this context; duplicate traces of the same session: mcir2fyj, wfs2tv4r; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5598.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 5,598 193.6 103.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p1fmx2h7",
   "source_url": "https://omlx.ai/benchmarks/performance/p1fmx2h7",
   "tg_tokens": null,
   "tps": 193.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-p1fmx2h7-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 108.5 GB at this context; duplicate traces of the same session: mcir2fyj, wfs2tv4r; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5330.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,330 113.8 108.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p1fmx2h7",
   "source_url": "https://omlx.ai/benchmarks/performance/p1fmx2h7",
   "tg_tokens": null,
   "tps": 113.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-tl2ok0i2-ctx1024",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: drf7gdky; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2028.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 2,028 54.6 25.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tl2ok0i2",
   "source_url": "https://omlx.ai/benchmarks/performance/tl2ok0i2",
   "tg_tokens": null,
   "tps": 54.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-tl2ok0i2-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: drf7gdky; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2150.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 2,150 53.2 26.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tl2ok0i2",
   "source_url": "https://omlx.ai/benchmarks/performance/tl2ok0i2",
   "tg_tokens": null,
   "tps": 53.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-78das9mf-ctx1024",
   "model": "Llama-3-Groq-8B-Tool-Use",
   "notes": "oMLX community benchmark trace; peak memory 5.0 GB at this context; duplicate traces of the same session: xqvbppt4; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 492.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Llama-3-Groq-8B-Tool-Use-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 492.8 61.1 5.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/78das9mf",
   "source_url": "https://omlx.ai/benchmarks/performance/78das9mf",
   "tg_tokens": null,
   "tps": 61.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-78das9mf-ctx4096",
   "model": "Llama-3-Groq-8B-Tool-Use",
   "notes": "oMLX community benchmark trace; peak memory 5.4 GB at this context; duplicate traces of the same session: xqvbppt4; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 479.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Llama-3-Groq-8B-Tool-Use-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 479.3 55.7 5.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/78das9mf",
   "source_url": "https://omlx.ai/benchmarks/performance/78das9mf",
   "tg_tokens": null,
   "tps": 55.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-oixg0s3f-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: p0hquqma, riqfx64i; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 773.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 773.0 90.6 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oixg0s3f",
   "source_url": "https://omlx.ai/benchmarks/performance/oixg0s3f",
   "tg_tokens": null,
   "tps": 90.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-oixg0s3f-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: p0hquqma, riqfx64i; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 923.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 923.6 86.5 21.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oixg0s3f",
   "source_url": "https://omlx.ai/benchmarks/performance/oixg0s3f",
   "tg_tokens": null,
   "tps": 86.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-oixg0s3f-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: p0hquqma, riqfx64i; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 616.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 32k 616.2 62.8 22.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oixg0s3f",
   "source_url": "https://omlx.ai/benchmarks/performance/oixg0s3f",
   "tg_tokens": null,
   "tps": 62.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-v579d2ub-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 14l8pfxw; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 743.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 743.6 89.3 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v579d2ub",
   "source_url": "https://omlx.ai/benchmarks/performance/v579d2ub",
   "tg_tokens": null,
   "tps": 89.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-v579d2ub-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 14l8pfxw; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 885.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 885.5 86.7 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v579d2ub",
   "source_url": "https://omlx.ai/benchmarks/performance/v579d2ub",
   "tg_tokens": null,
   "tps": 86.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-v579d2ub-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: 14l8pfxw; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 589.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 32k 589.1 62.2 23.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v579d2ub",
   "source_url": "https://omlx.ai/benchmarks/performance/v579d2ub",
   "tg_tokens": null,
   "tps": 62.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-371id0k3-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: kovo4900; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 944.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 944.9 89.3 19.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/371id0k3",
   "source_url": "https://omlx.ai/benchmarks/performance/371id0k3",
   "tg_tokens": null,
   "tps": 89.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-371id0k3-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: kovo4900; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 989.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 989.4 86.5 20.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/371id0k3",
   "source_url": "https://omlx.ai/benchmarks/performance/371id0k3",
   "tg_tokens": null,
   "tps": 86.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-oitqlswq-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: nrf3joho; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 942.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 942.3 91.8 19.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oitqlswq",
   "source_url": "https://omlx.ai/benchmarks/performance/oitqlswq",
   "tg_tokens": null,
   "tps": 91.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-oitqlswq-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: nrf3joho; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 988.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 988.3 86.5 20.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oitqlswq",
   "source_url": "https://omlx.ai/benchmarks/performance/oitqlswq",
   "tg_tokens": null,
   "tps": 86.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-vhyq3f47-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 4sfditjx; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 937.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 937.0 91.8 19.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vhyq3f47",
   "source_url": "https://omlx.ai/benchmarks/performance/vhyq3f47",
   "tg_tokens": null,
   "tps": 91.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-vhyq3f47-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 4sfditjx; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 990.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 990.5 86.6 20.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vhyq3f47",
   "source_url": "https://omlx.ai/benchmarks/performance/vhyq3f47",
   "tg_tokens": null,
   "tps": 86.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-1v35y06y-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 61h0kswm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 939.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 939.6 90.5 19.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1v35y06y",
   "source_url": "https://omlx.ai/benchmarks/performance/1v35y06y",
   "tg_tokens": null,
   "tps": 90.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-1v35y06y-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 61h0kswm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 988.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 988.6 86.2 20.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1v35y06y",
   "source_url": "https://omlx.ai/benchmarks/performance/1v35y06y",
   "tg_tokens": null,
   "tps": 86.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-fsri2w2u-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: nynl2v3k; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 938.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 938.3 90.3 19.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fsri2w2u",
   "source_url": "https://omlx.ai/benchmarks/performance/fsri2w2u",
   "tg_tokens": null,
   "tps": 90.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-fsri2w2u-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: nynl2v3k; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 988.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 988.3 86.3 20.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fsri2w2u",
   "source_url": "https://omlx.ai/benchmarks/performance/fsri2w2u",
   "tg_tokens": null,
   "tps": 86.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-pziuj7oo-ctx1024",
   "model": "Qwen2.5-32B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 17.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 734.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-32B-Instruct-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 734.9 28.1 17.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pziuj7oo",
   "source_url": "https://omlx.ai/benchmarks/performance/pziuj7oo",
   "tg_tokens": null,
   "tps": 28.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-pziuj7oo-ctx4096",
   "model": "Qwen2.5-32B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 794.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-32B-Instruct-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 794.1 26.9 19.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pziuj7oo",
   "source_url": "https://omlx.ai/benchmarks/performance/pziuj7oo",
   "tg_tokens": null,
   "tps": 26.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9hrgltx3-ctx1024",
   "model": "Qwen3.5-0.8B",
   "notes": "MLX MTP build; oMLX community benchmark trace; peak memory 2.2 GB at this context; duplicate traces of the same session: 9lq996wf, ayg77rfc, nwdjag3d, qpl99ura; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5203.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.5-0.8B-MLX-MTP-5bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 5,203 141.7 2.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9hrgltx3",
   "source_url": "https://omlx.ai/benchmarks/performance/9hrgltx3",
   "tg_tokens": null,
   "tps": 141.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9hrgltx3-ctx4096",
   "model": "Qwen3.5-0.8B",
   "notes": "MLX MTP build; oMLX community benchmark trace; peak memory 2.5 GB at this context; duplicate traces of the same session: 9lq996wf, ayg77rfc, nwdjag3d, qpl99ura; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 17974.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.5-0.8B-MLX-MTP-5bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 17,974 331.5 2.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9hrgltx3",
   "source_url": "https://omlx.ai/benchmarks/performance/9hrgltx3",
   "tg_tokens": null,
   "tps": 331.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9hrgltx3-ctx8192",
   "model": "Qwen3.5-0.8B",
   "notes": "MLX MTP build; oMLX community benchmark trace; peak memory 2.6 GB at this context; duplicate traces of the same session: 9lq996wf, ayg77rfc, nwdjag3d, qpl99ura; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 19711.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.5-0.8B-MLX-MTP-5bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 19,711 389.0 2.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9hrgltx3",
   "source_url": "https://omlx.ai/benchmarks/performance/9hrgltx3",
   "tg_tokens": null,
   "tps": 389.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9hrgltx3-ctx32768",
   "model": "Qwen3.5-0.8B",
   "notes": "MLX MTP build; oMLX community benchmark trace; peak memory 3.1 GB at this context; duplicate traces of the same session: 9lq996wf, ayg77rfc, nwdjag3d, qpl99ura; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 17226.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.5-0.8B-MLX-MTP-5bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 17,226 622.0 3.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9hrgltx3",
   "source_url": "https://omlx.ai/benchmarks/performance/9hrgltx3",
   "tg_tokens": null,
   "tps": 622.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9hrgltx3-ctx65536",
   "model": "Qwen3.5-0.8B",
   "notes": "MLX MTP build; oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: 9lq996wf, ayg77rfc, nwdjag3d, qpl99ura; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 12634.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.5-0.8B-MLX-MTP-5bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 12,634 275.6 3.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9hrgltx3",
   "source_url": "https://omlx.ai/benchmarks/performance/9hrgltx3",
   "tg_tokens": null,
   "tps": 275.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9hrgltx3-ctx131072",
   "model": "Qwen3.5-0.8B",
   "notes": "MLX MTP build; oMLX community benchmark trace; peak memory 5.5 GB at this context; duplicate traces of the same session: 9lq996wf, ayg77rfc, nwdjag3d, qpl99ura; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 8002.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.5-0.8B-MLX-MTP-5bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 8,002 227.0 5.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9hrgltx3",
   "source_url": "https://omlx.ai/benchmarks/performance/9hrgltx3",
   "tg_tokens": null,
   "tps": 227.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9hrgltx3-ctx199680",
   "model": "Qwen3.5-0.8B",
   "notes": "MLX MTP build; oMLX community benchmark trace; peak memory 7.2 GB at this context; duplicate traces of the same session: 9lq996wf, ayg77rfc, nwdjag3d, qpl99ura; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5736.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.5-0.8B-MLX-MTP-5bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 195k 5,736 158.7 7.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9hrgltx3",
   "source_url": "https://omlx.ai/benchmarks/performance/9hrgltx3",
   "tg_tokens": null,
   "tps": 158.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-254gmu2i-ctx1024",
   "model": "Qwen3.5-2B",
   "notes": "oMLX community benchmark trace; peak memory 2.3 GB at this context; duplicate traces of the same session: u7x4ypks; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1027.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.5-2B-OptiQ-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 1,027 48.2 2.3 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/254gmu2i",
   "source_url": "https://omlx.ai/benchmarks/performance/254gmu2i",
   "tg_tokens": null,
   "tps": 48.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-254gmu2i-ctx4096",
   "model": "Qwen3.5-2B",
   "notes": "oMLX community benchmark trace; peak memory 2.6 GB at this context; duplicate traces of the same session: u7x4ypks; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 900.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.5-2B-OptiQ-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 900.0 43.4 2.6 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/254gmu2i",
   "source_url": "https://omlx.ai/benchmarks/performance/254gmu2i",
   "tg_tokens": null,
   "tps": 43.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jep9zgst-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 35.7 GB at this context; duplicate traces of the same session: 0795oda5, ecaosczk, qbadlq09, xoyadwp6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4044.0,
   "provenance": "community",
   "quant": "UD-Q8_K_XL",
   "quote": "Qwen3.6-35B-A3B-UD-Q8_K_XL-mlx on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 4,044 99.1 35.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jep9zgst",
   "source_url": "https://omlx.ai/benchmarks/performance/jep9zgst",
   "tg_tokens": null,
   "tps": 99.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jep9zgst-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 37.0 GB at this context; duplicate traces of the same session: 0795oda5, ecaosczk, qbadlq09, xoyadwp6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4418.0,
   "provenance": "community",
   "quant": "UD-Q8_K_XL",
   "quote": "Qwen3.6-35B-A3B-UD-Q8_K_XL-mlx on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 32k 4,418 83.4 37.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jep9zgst",
   "source_url": "https://omlx.ai/benchmarks/performance/jep9zgst",
   "tg_tokens": null,
   "tps": 83.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jep9zgst-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 37.6 GB at this context; duplicate traces of the same session: 0795oda5, ecaosczk, qbadlq09, xoyadwp6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3372.0,
   "provenance": "community",
   "quant": "UD-Q8_K_XL",
   "quote": "Qwen3.6-35B-A3B-UD-Q8_K_XL-mlx on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 64k 3,372 76.1 37.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jep9zgst",
   "source_url": "https://omlx.ai/benchmarks/performance/jep9zgst",
   "tg_tokens": null,
   "tps": 76.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jep9zgst-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 38.9 GB at this context; duplicate traces of the same session: 0795oda5, ecaosczk, qbadlq09, xoyadwp6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2277.0,
   "provenance": "community",
   "quant": "UD-Q8_K_XL",
   "quote": "Qwen3.6-35B-A3B-UD-Q8_K_XL-mlx on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 128k 2,277 61.1 38.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jep9zgst",
   "source_url": "https://omlx.ai/benchmarks/performance/jep9zgst",
   "tg_tokens": null,
   "tps": 61.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jep9zgst-ctx199680",
   "model": "Qwen3.6-35B-A3B",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 40.2 GB at this context; duplicate traces of the same session: 0795oda5, ecaosczk, qbadlq09, xoyadwp6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1568.0,
   "provenance": "community",
   "quant": "UD-Q8_K_XL",
   "quote": "Qwen3.6-35B-A3B-UD-Q8_K_XL-mlx on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 195k 1,568 50.1 40.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jep9zgst",
   "source_url": "https://omlx.ai/benchmarks/performance/jep9zgst",
   "tg_tokens": null,
   "tps": 50.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0kl8vh1a-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 2gv00yfr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4451.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,451 186.8 21.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0kl8vh1a",
   "source_url": "https://omlx.ai/benchmarks/performance/0kl8vh1a",
   "tg_tokens": null,
   "tps": 186.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0kl8vh1a-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 2gv00yfr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6255.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,255 209.7 22.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0kl8vh1a",
   "source_url": "https://omlx.ai/benchmarks/performance/0kl8vh1a",
   "tg_tokens": null,
   "tps": 209.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-gm1wt2t3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: a1dajg7r; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 4k 105.5 17.7 17.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gm1wt2t3",
   "source_url": "https://omlx.ai/benchmarks/performance/gm1wt2t3",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-gm1wt2t3-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: a1dajg7r; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 91.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 16k 91.6 14.2 19.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gm1wt2t3",
   "source_url": "https://omlx.ai/benchmarks/performance/gm1wt2t3",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tfqrohi5-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 929.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 929.2 32.2 16.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tfqrohi5",
   "source_url": "https://omlx.ai/benchmarks/performance/tfqrohi5",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tfqrohi5-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1011.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,011 31.4 17.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tfqrohi5",
   "source_url": "https://omlx.ai/benchmarks/performance/tfqrohi5",
   "tg_tokens": null,
   "tps": 31.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dre0asf3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 2pgdbx3w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 930.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 930.6 32.6 16.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dre0asf3",
   "source_url": "https://omlx.ai/benchmarks/performance/dre0asf3",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dre0asf3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 2pgdbx3w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1015.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,015 32.1 17.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dre0asf3",
   "source_url": "https://omlx.ai/benchmarks/performance/dre0asf3",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wt4qmg77-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: b5l7t5l4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 923.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 923.4 32.0 16.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wt4qmg77",
   "source_url": "https://omlx.ai/benchmarks/performance/wt4qmg77",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wt4qmg77-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: b5l7t5l4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1009.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,009 31.5 17.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wt4qmg77",
   "source_url": "https://omlx.ai/benchmarks/performance/wt4qmg77",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-emwcx8j4-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 19.3 GB; TTFT 71522 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 458.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k tokens 458.2 PP tok/s 15.0 TG tok/s 71522 TTFT (ms) 19.3 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/emwcx8j4",
   "source_url": "https://omlx.ai/benchmarks/performance/emwcx8j4",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jqc3gi8g-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 69.6 GB; TTFT 477 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2147.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 2,147 PP tok/s 79.0 TG tok/s 477 TTFT (ms) 69.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jqc3gi8g",
   "source_url": "https://omlx.ai/benchmarks/performance/jqc3gi8g",
   "tg_tokens": null,
   "tps": 79.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-09my8afp-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 61.4 GB at this context; duplicate traces of the same session: 106ish9b, 41472c3w, i7nkm9bu, qiglbnp0, w3s1m245; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 974.4,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 1k 974.4 61.2 61.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/09my8afp",
   "source_url": "https://omlx.ai/benchmarks/performance/09my8afp",
   "tg_tokens": null,
   "tps": 61.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-09my8afp-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 61.6 GB at this context; duplicate traces of the same session: 106ish9b, 41472c3w, i7nkm9bu, qiglbnp0, w3s1m245; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1865.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 4k 1,865 60.2 61.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/09my8afp",
   "source_url": "https://omlx.ai/benchmarks/performance/09my8afp",
   "tg_tokens": null,
   "tps": 60.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-09my8afp-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 62.5 GB at this context; duplicate traces of the same session: 106ish9b, 41472c3w, i7nkm9bu, qiglbnp0, w3s1m245; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2293.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 8k 2,293 77.4 62.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/09my8afp",
   "source_url": "https://omlx.ai/benchmarks/performance/09my8afp",
   "tg_tokens": null,
   "tps": 77.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-09my8afp-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 63.5 GB at this context; duplicate traces of the same session: 106ish9b, 41472c3w, i7nkm9bu, qiglbnp0, w3s1m245; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2558.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 16k 2,558 108.2 63.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/09my8afp",
   "source_url": "https://omlx.ai/benchmarks/performance/09my8afp",
   "tg_tokens": null,
   "tps": 108.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-09my8afp-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 64.1 GB at this context; duplicate traces of the same session: 106ish9b, 41472c3w, i7nkm9bu, qiglbnp0, w3s1m245; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2517.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 32k 2,517 75.1 64.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/09my8afp",
   "source_url": "https://omlx.ai/benchmarks/performance/09my8afp",
   "tg_tokens": null,
   "tps": 75.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-09my8afp-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 65.3 GB at this context; duplicate traces of the same session: 106ish9b, 41472c3w, i7nkm9bu, qiglbnp0, w3s1m245; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2393.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 64k 2,393 58.2 65.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/09my8afp",
   "source_url": "https://omlx.ai/benchmarks/performance/09my8afp",
   "tg_tokens": null,
   "tps": 58.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vo6xfbms-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: 3n4kv4vn, 3yzz50xz, 9e5wum5h, ucwfp940, xri2fw7v, yh18vo8m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 958.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 958.6 59.7 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vo6xfbms",
   "source_url": "https://omlx.ai/benchmarks/performance/vo6xfbms",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vo6xfbms-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: 3n4kv4vn, 3yzz50xz, 9e5wum5h, ucwfp940, xri2fw7v, yh18vo8m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1932.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,932 70.5 71.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vo6xfbms",
   "source_url": "https://omlx.ai/benchmarks/performance/vo6xfbms",
   "tg_tokens": null,
   "tps": 70.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vo6xfbms-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.0 GB at this context; duplicate traces of the same session: 3n4kv4vn, 3yzz50xz, 9e5wum5h, ucwfp940, xri2fw7v, yh18vo8m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2909.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,909 85.6 73.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vo6xfbms",
   "source_url": "https://omlx.ai/benchmarks/performance/vo6xfbms",
   "tg_tokens": null,
   "tps": 85.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-vo6xfbms-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.9 GB at this context; duplicate traces of the same session: 3n4kv4vn, 3yzz50xz, 9e5wum5h, ucwfp940, xri2fw7v, yh18vo8m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2814.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,814 85.5 73.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vo6xfbms",
   "source_url": "https://omlx.ai/benchmarks/performance/vo6xfbms",
   "tg_tokens": null,
   "tps": 85.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vo6xfbms-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 75.5 GB at this context; duplicate traces of the same session: 3n4kv4vn, 3yzz50xz, 9e5wum5h, ucwfp940, xri2fw7v, yh18vo8m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2724.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,724 76.6 75.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vo6xfbms",
   "source_url": "https://omlx.ai/benchmarks/performance/vo6xfbms",
   "tg_tokens": null,
   "tps": 76.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vo6xfbms-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.6 GB at this context; duplicate traces of the same session: 3n4kv4vn, 3yzz50xz, 9e5wum5h, ucwfp940, xri2fw7v, yh18vo8m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2597.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,597 64.7 78.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vo6xfbms",
   "source_url": "https://omlx.ai/benchmarks/performance/vo6xfbms",
   "tg_tokens": null,
   "tps": 64.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vo6xfbms-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 82.3 GB at this context; duplicate traces of the same session: 3n4kv4vn, 3yzz50xz, 9e5wum5h, ucwfp940, xri2fw7v, yh18vo8m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2515.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,515 59.4 82.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vo6xfbms",
   "source_url": "https://omlx.ai/benchmarks/performance/vo6xfbms",
   "tg_tokens": null,
   "tps": 59.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dvicpjpn-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 71.1 GB; TTFT 1779 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 575.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 575.5 PP tok/s 36.6 TG tok/s 1779 TTFT (ms) 71.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dvicpjpn",
   "source_url": "https://omlx.ai/benchmarks/performance/dvicpjpn",
   "tg_tokens": null,
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8zuty4eh-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1648.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,648 87.2 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8zuty4eh",
   "source_url": "https://omlx.ai/benchmarks/performance/8zuty4eh",
   "tg_tokens": null,
   "tps": 87.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8zuty4eh-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3098.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 3,098 101.9 72.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8zuty4eh",
   "source_url": "https://omlx.ai/benchmarks/performance/8zuty4eh",
   "tg_tokens": null,
   "tps": 101.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8zuty4eh-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2755.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,755 68.1 73.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8zuty4eh",
   "source_url": "https://omlx.ai/benchmarks/performance/8zuty4eh",
   "tg_tokens": null,
   "tps": 68.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8zuty4eh-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 75.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2575.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,575 81.3 75.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8zuty4eh",
   "source_url": "https://omlx.ai/benchmarks/performance/8zuty4eh",
   "tg_tokens": null,
   "tps": 81.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-x2xmmo1w-ctx1024",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: yl0j54kq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 377.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31B-it-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 377.4 14.7 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x2xmmo1w",
   "source_url": "https://omlx.ai/benchmarks/performance/x2xmmo1w",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-x2xmmo1w-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: yl0j54kq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 342.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31B-it-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 342.3 14.1 18.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x2xmmo1w",
   "source_url": "https://omlx.ai/benchmarks/performance/x2xmmo1w",
   "tg_tokens": null,
   "tps": 14.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-vdtpisow-ctx1024",
   "model": "supergemma4-26b-abliterated-multimodal",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: tdi119qy; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 809.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "supergemma4-26b-abliterated-multimodal-mlx-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 809.9 48.0 25.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vdtpisow",
   "source_url": "https://omlx.ai/benchmarks/performance/vdtpisow",
   "tg_tokens": null,
   "tps": 48.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-vdtpisow-ctx4096",
   "model": "supergemma4-26b-abliterated-multimodal",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: tdi119qy; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 781.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "supergemma4-26b-abliterated-multimodal-mlx-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 781.2 47.0 26.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vdtpisow",
   "source_url": "https://omlx.ai/benchmarks/performance/vdtpisow",
   "tg_tokens": null,
   "tps": 47.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-nn828r34-ctx1024",
   "model": "DeepSeek-R1-Distill-Llama-70B",
   "notes": "oMLX community benchmark trace; peak memory 37.8 GB at this context; duplicate traces of the same session: yihhh9er; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 360.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-R1-Distill-Llama-70B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 360.7 13.4 37.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nn828r34",
   "source_url": "https://omlx.ai/benchmarks/performance/nn828r34",
   "tg_tokens": null,
   "tps": 13.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-nn828r34-ctx4096",
   "model": "DeepSeek-R1-Distill-Llama-70B",
   "notes": "oMLX community benchmark trace; peak memory 39.0 GB at this context; duplicate traces of the same session: yihhh9er; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 370.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-R1-Distill-Llama-70B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 370.1 13.0 39.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nn828r34",
   "source_url": "https://omlx.ai/benchmarks/performance/nn828r34",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-9bhatffw-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.6 GB at this context; duplicate traces of the same session: 0lwoilfh, 77kcodqj, 7zhafh4r, bee78qd0, c6i7w4hd, nu86ys13, xmy05uk4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1647.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,647 86.2 174.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9bhatffw",
   "source_url": "https://omlx.ai/benchmarks/performance/9bhatffw",
   "tg_tokens": null,
   "tps": 86.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-9bhatffw-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.5 GB at this context; duplicate traces of the same session: 0lwoilfh, 77kcodqj, 7zhafh4r, bee78qd0, c6i7w4hd, nu86ys13, xmy05uk4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2400.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 2,400 78.8 175.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9bhatffw",
   "source_url": "https://omlx.ai/benchmarks/performance/9bhatffw",
   "tg_tokens": null,
   "tps": 78.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-9bhatffw-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 176.3 GB at this context; duplicate traces of the same session: 0lwoilfh, 77kcodqj, 7zhafh4r, bee78qd0, c6i7w4hd, nu86ys13, xmy05uk4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2418.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 2,418 81.4 176.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9bhatffw",
   "source_url": "https://omlx.ai/benchmarks/performance/9bhatffw",
   "tg_tokens": null,
   "tps": 81.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-9bhatffw-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 176.6 GB at this context; duplicate traces of the same session: 0lwoilfh, 77kcodqj, 7zhafh4r, bee78qd0, c6i7w4hd, nu86ys13, xmy05uk4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2417.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 2,417 161.4 176.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9bhatffw",
   "source_url": "https://omlx.ai/benchmarks/performance/9bhatffw",
   "tg_tokens": null,
   "tps": 161.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-9bhatffw-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: 0lwoilfh, 77kcodqj, 7zhafh4r, bee78qd0, c6i7w4hd, nu86ys13, xmy05uk4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2408.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,408 129.9 177.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9bhatffw",
   "source_url": "https://omlx.ai/benchmarks/performance/9bhatffw",
   "tg_tokens": null,
   "tps": 129.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-9bhatffw-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.9 GB at this context; duplicate traces of the same session: 0lwoilfh, 77kcodqj, 7zhafh4r, bee78qd0, c6i7w4hd, nu86ys13, xmy05uk4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2381.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,381 92.0 178.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9bhatffw",
   "source_url": "https://omlx.ai/benchmarks/performance/9bhatffw",
   "tg_tokens": null,
   "tps": 92.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-9bhatffw-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 181.8 GB at this context; duplicate traces of the same session: 0lwoilfh, 77kcodqj, 7zhafh4r, bee78qd0, c6i7w4hd, nu86ys13, xmy05uk4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2330.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,330 162.8 181.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9bhatffw",
   "source_url": "https://omlx.ai/benchmarks/performance/9bhatffw",
   "tg_tokens": null,
   "tps": 162.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-9bhatffw-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 184.6 GB at this context; duplicate traces of the same session: 0lwoilfh, 77kcodqj, 7zhafh4r, bee78qd0, c6i7w4hd, nu86ys13, xmy05uk4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2276.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 2,276 76.3 184.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9bhatffw",
   "source_url": "https://omlx.ai/benchmarks/performance/9bhatffw",
   "tg_tokens": null,
   "tps": 76.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-kx5dtpjl-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.7 GB at this context; duplicate traces of the same session: 2n42kzzr, 7wn5kqbv, bi5klnnb, ghyui8mh, i62f6mgj, u5l70axh, v0iiw0l0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1551.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,551 73.2 179.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kx5dtpjl",
   "source_url": "https://omlx.ai/benchmarks/performance/kx5dtpjl",
   "tg_tokens": null,
   "tps": 73.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-kx5dtpjl-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.6 GB at this context; duplicate traces of the same session: 2n42kzzr, 7wn5kqbv, bi5klnnb, ghyui8mh, i62f6mgj, u5l70axh, v0iiw0l0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2330.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 2,330 67.7 180.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kx5dtpjl",
   "source_url": "https://omlx.ai/benchmarks/performance/kx5dtpjl",
   "tg_tokens": null,
   "tps": 67.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-kx5dtpjl-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.4 GB at this context; duplicate traces of the same session: 2n42kzzr, 7wn5kqbv, bi5klnnb, ghyui8mh, i62f6mgj, u5l70axh, v0iiw0l0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2360.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 2,360 82.6 181.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kx5dtpjl",
   "source_url": "https://omlx.ai/benchmarks/performance/kx5dtpjl",
   "tg_tokens": null,
   "tps": 82.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-kx5dtpjl-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.8 GB at this context; duplicate traces of the same session: 2n42kzzr, 7wn5kqbv, bi5klnnb, ghyui8mh, i62f6mgj, u5l70axh, v0iiw0l0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2365.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 2,365 141.4 181.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kx5dtpjl",
   "source_url": "https://omlx.ai/benchmarks/performance/kx5dtpjl",
   "tg_tokens": null,
   "tps": 141.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-kx5dtpjl-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 182.5 GB at this context; duplicate traces of the same session: 2n42kzzr, 7wn5kqbv, bi5klnnb, ghyui8mh, i62f6mgj, u5l70axh, v0iiw0l0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2356.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,356 115.5 182.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kx5dtpjl",
   "source_url": "https://omlx.ai/benchmarks/performance/kx5dtpjl",
   "tg_tokens": null,
   "tps": 115.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-kx5dtpjl-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 184.0 GB at this context; duplicate traces of the same session: 2n42kzzr, 7wn5kqbv, bi5klnnb, ghyui8mh, i62f6mgj, u5l70axh, v0iiw0l0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2323.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,323 103.2 184.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kx5dtpjl",
   "source_url": "https://omlx.ai/benchmarks/performance/kx5dtpjl",
   "tg_tokens": null,
   "tps": 103.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-kx5dtpjl-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 186.9 GB at this context; duplicate traces of the same session: 2n42kzzr, 7wn5kqbv, bi5klnnb, ghyui8mh, i62f6mgj, u5l70axh, v0iiw0l0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2278.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,278 153.3 186.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kx5dtpjl",
   "source_url": "https://omlx.ai/benchmarks/performance/kx5dtpjl",
   "tg_tokens": null,
   "tps": 153.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-kx5dtpjl-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 189.8 GB at this context; duplicate traces of the same session: 2n42kzzr, 7wn5kqbv, bi5klnnb, ghyui8mh, i62f6mgj, u5l70axh, v0iiw0l0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2229.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 2,229 75.4 189.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kx5dtpjl",
   "source_url": "https://omlx.ai/benchmarks/performance/kx5dtpjl",
   "tg_tokens": null,
   "tps": 75.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-fhumjjpe-ctx4096",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 2.4 GB at this context; duplicate traces of the same session: bj4nwsr2; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 890.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiniCPM5-2B-oQ4e on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 890.0 88.3 2.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fhumjjpe",
   "source_url": "https://omlx.ai/benchmarks/performance/fhumjjpe",
   "tg_tokens": null,
   "tps": 88.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-fhumjjpe-ctx16384",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 2.6 GB at this context; duplicate traces of the same session: bj4nwsr2; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 615.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiniCPM5-2B-oQ4e on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k 615.3 63.7 2.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fhumjjpe",
   "source_url": "https://omlx.ai/benchmarks/performance/fhumjjpe",
   "tg_tokens": null,
   "tps": 63.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-zxefgkc5-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 743.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 743.6 89.3 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zxefgkc5",
   "source_url": "https://omlx.ai/benchmarks/performance/zxefgkc5",
   "tg_tokens": null,
   "tps": 89.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-zxefgkc5-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 885.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 885.5 86.7 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zxefgkc5",
   "source_url": "https://omlx.ai/benchmarks/performance/zxefgkc5",
   "tg_tokens": null,
   "tps": 86.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-zxefgkc5-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 589.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 32k 589.1 62.2 23.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zxefgkc5",
   "source_url": "https://omlx.ai/benchmarks/performance/zxefgkc5",
   "tg_tokens": null,
   "tps": 62.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-9aojw72k-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 0ke5t26u, 6unjvjo5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 775.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 775.5 91.0 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9aojw72k",
   "source_url": "https://omlx.ai/benchmarks/performance/9aojw72k",
   "tg_tokens": null,
   "tps": 91.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-9aojw72k-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 0ke5t26u, 6unjvjo5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 886.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 886.2 86.7 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9aojw72k",
   "source_url": "https://omlx.ai/benchmarks/performance/9aojw72k",
   "tg_tokens": null,
   "tps": 86.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-9aojw72k-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: 0ke5t26u, 6unjvjo5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 608.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 32k 608.8 63.1 23.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9aojw72k",
   "source_url": "https://omlx.ai/benchmarks/performance/9aojw72k",
   "tg_tokens": null,
   "tps": 63.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-nczbzv97-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: g4qcta2c; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 939.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 939.6 91.4 19.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nczbzv97",
   "source_url": "https://omlx.ai/benchmarks/performance/nczbzv97",
   "tg_tokens": null,
   "tps": 91.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-nczbzv97-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: g4qcta2c; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 988.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 988.5 86.3 20.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nczbzv97",
   "source_url": "https://omlx.ai/benchmarks/performance/nczbzv97",
   "tg_tokens": null,
   "tps": 86.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ddpu5h4l-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: zrj9pcwf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 939.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 939.8 89.2 19.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ddpu5h4l",
   "source_url": "https://omlx.ai/benchmarks/performance/ddpu5h4l",
   "tg_tokens": null,
   "tps": 89.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ddpu5h4l-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: zrj9pcwf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 987.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 987.3 85.4 20.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ddpu5h4l",
   "source_url": "https://omlx.ai/benchmarks/performance/ddpu5h4l",
   "tg_tokens": null,
   "tps": 85.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-lboe55xr-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: z5bnwqjo; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 921.5,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 921.5 76.1 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lboe55xr",
   "source_url": "https://omlx.ai/benchmarks/performance/lboe55xr",
   "tg_tokens": null,
   "tps": 76.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-lboe55xr-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: z5bnwqjo; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1154.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 1,154 71.9 22.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lboe55xr",
   "source_url": "https://omlx.ai/benchmarks/performance/lboe55xr",
   "tg_tokens": null,
   "tps": 71.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-jexedf25-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: m8w0rbnh; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 953.2,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 953.2 76.5 20.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jexedf25",
   "source_url": "https://omlx.ai/benchmarks/performance/jexedf25",
   "tg_tokens": null,
   "tps": 76.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-jexedf25-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: m8w0rbnh; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1154.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 1,154 73.1 22.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jexedf25",
   "source_url": "https://omlx.ai/benchmarks/performance/jexedf25",
   "tg_tokens": null,
   "tps": 73.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-l2a2z27t-ctx1024",
   "model": "Qwen3-4B-Instruct-2507",
   "notes": "oMLX community benchmark trace; peak memory 2.9 GB at this context; duplicate traces of the same session: m4p5vtk3; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 734.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-4B-Instruct-2507-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 734.1 94.8 2.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l2a2z27t",
   "source_url": "https://omlx.ai/benchmarks/performance/l2a2z27t",
   "tg_tokens": null,
   "tps": 94.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-l2a2z27t-ctx4096",
   "model": "Qwen3-4B-Instruct-2507",
   "notes": "oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: m4p5vtk3; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 670.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-4B-Instruct-2507-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 670.5 81.2 3.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l2a2z27t",
   "source_url": "https://omlx.ai/benchmarks/performance/l2a2z27t",
   "tg_tokens": null,
   "tps": 81.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fxoylkr9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: 3s8gy3s0; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 719.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 719.7 25.2 15.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fxoylkr9",
   "source_url": "https://omlx.ai/benchmarks/performance/fxoylkr9",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fxoylkr9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 3s8gy3s0; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 740.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 740.0 24.7 17.3 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fxoylkr9",
   "source_url": "https://omlx.ai/benchmarks/performance/fxoylkr9",
   "tg_tokens": null,
   "tps": 24.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-n02z0iml-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 139.4 GB at this context; duplicate traces of the same session: 2qf20tvw, 3an2fknz, cg9nqv6x, hchmtj5d, oonnaz15, rr83it8s; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 800.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 1k 800.2 65.8 139.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n02z0iml",
   "source_url": "https://omlx.ai/benchmarks/performance/n02z0iml",
   "tg_tokens": null,
   "tps": 65.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-n02z0iml-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 140.0 GB at this context; duplicate traces of the same session: 2qf20tvw, 3an2fknz, cg9nqv6x, hchmtj5d, oonnaz15, rr83it8s; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1210.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 4k 1,210 57.4 140.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n02z0iml",
   "source_url": "https://omlx.ai/benchmarks/performance/n02z0iml",
   "tg_tokens": null,
   "tps": 57.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-n02z0iml-ctx8192",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 140.1 GB at this context; duplicate traces of the same session: 2qf20tvw, 3an2fknz, cg9nqv6x, hchmtj5d, oonnaz15, rr83it8s; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1384.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 8k 1,384 61.8 140.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n02z0iml",
   "source_url": "https://omlx.ai/benchmarks/performance/n02z0iml",
   "tg_tokens": null,
   "tps": 61.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-n02z0iml-ctx16384",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 140.3 GB at this context; duplicate traces of the same session: 2qf20tvw, 3an2fknz, cg9nqv6x, hchmtj5d, oonnaz15, rr83it8s; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1371.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 16k 1,371 60.9 140.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n02z0iml",
   "source_url": "https://omlx.ai/benchmarks/performance/n02z0iml",
   "tg_tokens": null,
   "tps": 60.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-n02z0iml-ctx65536",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 141.8 GB at this context; duplicate traces of the same session: 2qf20tvw, 3an2fknz, cg9nqv6x, hchmtj5d, oonnaz15, rr83it8s; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1359.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 64k 1,359 39.5 141.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n02z0iml",
   "source_url": "https://omlx.ai/benchmarks/performance/n02z0iml",
   "tg_tokens": null,
   "tps": 39.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-n02z0iml-ctx131072",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 145.0 GB at this context; duplicate traces of the same session: 2qf20tvw, 3an2fknz, cg9nqv6x, hchmtj5d, oonnaz15, rr83it8s; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1336.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 128k 1,336 53.6 145.0 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n02z0iml",
   "source_url": "https://omlx.ai/benchmarks/performance/n02z0iml",
   "tg_tokens": null,
   "tps": 53.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-n02z0iml-ctx199680",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 149.0 GB at this context; duplicate traces of the same session: 2qf20tvw, 3an2fknz, cg9nqv6x, hchmtj5d, oonnaz15, rr83it8s; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1318.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 195k 1,318 53.1 149.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n02z0iml",
   "source_url": "https://omlx.ai/benchmarks/performance/n02z0iml",
   "tg_tokens": null,
   "tps": 53.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-t3lls4ib-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: 1mqr7tqk, ehgxwafj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1648.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,648 87.2 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t3lls4ib",
   "source_url": "https://omlx.ai/benchmarks/performance/t3lls4ib",
   "tg_tokens": null,
   "tps": 87.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-t3lls4ib-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: 1mqr7tqk, ehgxwafj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3098.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 3,098 101.9 72.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t3lls4ib",
   "source_url": "https://omlx.ai/benchmarks/performance/t3lls4ib",
   "tg_tokens": null,
   "tps": 101.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-t3lls4ib-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: 1mqr7tqk, ehgxwafj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2755.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,755 68.1 73.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t3lls4ib",
   "source_url": "https://omlx.ai/benchmarks/performance/t3lls4ib",
   "tg_tokens": null,
   "tps": 68.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-t3lls4ib-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 75.2 GB at this context; duplicate traces of the same session: 1mqr7tqk, ehgxwafj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2575.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,575 81.3 75.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t3lls4ib",
   "source_url": "https://omlx.ai/benchmarks/performance/t3lls4ib",
   "tg_tokens": null,
   "tps": 81.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8byfb6sn-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.5 GB at this context; duplicate traces of the same session: pd7xar1k, u3srxav7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 905.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 905.8 39.7 72.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8byfb6sn",
   "source_url": "https://omlx.ai/benchmarks/performance/8byfb6sn",
   "tg_tokens": null,
   "tps": 39.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8byfb6sn-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.7 GB at this context; duplicate traces of the same session: pd7xar1k, u3srxav7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 825.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 825.1 35.0 72.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8byfb6sn",
   "source_url": "https://omlx.ai/benchmarks/performance/8byfb6sn",
   "tg_tokens": null,
   "tps": 35.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8byfb6sn-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.2 GB at this context; duplicate traces of the same session: pd7xar1k, u3srxav7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 800.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 800.2 40.1 73.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8byfb6sn",
   "source_url": "https://omlx.ai/benchmarks/performance/8byfb6sn",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8byfb6sn-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.8 GB at this context; duplicate traces of the same session: pd7xar1k, u3srxav7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 789.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 789.4 30.4 74.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8byfb6sn",
   "source_url": "https://omlx.ai/benchmarks/performance/8byfb6sn",
   "tg_tokens": null,
   "tps": 30.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-y0qvjnnf-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.4 GB at this context; duplicate traces of the same session: b213lzk6, dcqwtn0b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 886.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 886.4 50.5 71.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y0qvjnnf",
   "source_url": "https://omlx.ai/benchmarks/performance/y0qvjnnf",
   "tg_tokens": null,
   "tps": 50.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-y0qvjnnf-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.5 GB at this context; duplicate traces of the same session: b213lzk6, dcqwtn0b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1192.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,192 32.0 72.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y0qvjnnf",
   "source_url": "https://omlx.ai/benchmarks/performance/y0qvjnnf",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-y0qvjnnf-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.5 GB at this context; duplicate traces of the same session: b213lzk6, dcqwtn0b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1188.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,188 43.5 72.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y0qvjnnf",
   "source_url": "https://omlx.ai/benchmarks/performance/y0qvjnnf",
   "tg_tokens": null,
   "tps": 43.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ivm9h0n2-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.4 GB at this context; duplicate traces of the same session: ft7wy4t9, ybqasjn7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1013.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,013 30.0 71.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ivm9h0n2",
   "source_url": "https://omlx.ai/benchmarks/performance/ivm9h0n2",
   "tg_tokens": null,
   "tps": 30.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ivm9h0n2-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.5 GB at this context; duplicate traces of the same session: ft7wy4t9, ybqasjn7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 744.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 744.2 46.4 72.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ivm9h0n2",
   "source_url": "https://omlx.ai/benchmarks/performance/ivm9h0n2",
   "tg_tokens": null,
   "tps": 46.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ivm9h0n2-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.5 GB at this context; duplicate traces of the same session: ft7wy4t9, ybqasjn7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 968.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 968.6 32.3 72.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ivm9h0n2",
   "source_url": "https://omlx.ai/benchmarks/performance/ivm9h0n2",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i6cobu2p-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.4 GB at this context; duplicate traces of the same session: d2p0i8bb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 955.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 955.9 56.3 71.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i6cobu2p",
   "source_url": "https://omlx.ai/benchmarks/performance/i6cobu2p",
   "tg_tokens": null,
   "tps": 56.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i6cobu2p-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.5 GB at this context; duplicate traces of the same session: d2p0i8bb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1149.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,149 41.0 72.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i6cobu2p",
   "source_url": "https://omlx.ai/benchmarks/performance/i6cobu2p",
   "tg_tokens": null,
   "tps": 41.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-7c86nd3t-ctx1024",
   "model": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed",
   "notes": "FP16 variant; Youssofal build; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: hkyx1ohv; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 875.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Youssofal--Qwen3.6-35B-A3B-MTPLX-Optimized-Speed-FP16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 875.1 80.9 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7c86nd3t",
   "source_url": "https://omlx.ai/benchmarks/performance/7c86nd3t",
   "tg_tokens": null,
   "tps": 80.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-7c86nd3t-ctx4096",
   "model": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed",
   "notes": "FP16 variant; Youssofal build; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: hkyx1ohv; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1165.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Youssofal--Qwen3.6-35B-A3B-MTPLX-Optimized-Speed-FP16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 1,165 79.7 21.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7c86nd3t",
   "source_url": "https://omlx.ai/benchmarks/performance/7c86nd3t",
   "tg_tokens": null,
   "tps": 79.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-3g2lbt6o-ctx1024",
   "model": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed",
   "notes": "FP16 variant; Youssofal build; oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 1049 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 976.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Youssofal--Qwen3.6-35B-A3B-MTPLX-Optimized-Speed-FP16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k tokens 976.0 PP tok/s 81.2 TG tok/s 1049 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3g2lbt6o",
   "source_url": "https://omlx.ai/benchmarks/performance/3g2lbt6o",
   "tg_tokens": null,
   "tps": 81.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-xaa8qq3i-ctx1024",
   "model": "Gemma4-26B-MoE",
   "notes": "oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: 7u108mjl; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3405.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Gemma4-26B-MoE-Q4 on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 3,405 106.6 14.3 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xaa8qq3i",
   "source_url": "https://omlx.ai/benchmarks/performance/xaa8qq3i",
   "tg_tokens": null,
   "tps": 106.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-xaa8qq3i-ctx4096",
   "model": "Gemma4-26B-MoE",
   "notes": "oMLX community benchmark trace; peak memory 14.5 GB at this context; duplicate traces of the same session: 7u108mjl; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2806.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Gemma4-26B-MoE-Q4 on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 2,806 105.5 14.5 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xaa8qq3i",
   "source_url": "https://omlx.ai/benchmarks/performance/xaa8qq3i",
   "tg_tokens": null,
   "tps": 105.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-au2wktui-ctx1024",
   "model": "Huihui-ThinkingCap-Qwen3.6-27B-abliterated",
   "notes": "oMLX MTP build; oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: g9tb1prq, jrplfhqz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 258.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Huihui-ThinkingCap-Qwen3.6-27B-abliterated-MLX-4bit-oMLX-MTP on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 258.1 29.2 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/au2wktui",
   "source_url": "https://omlx.ai/benchmarks/performance/au2wktui",
   "tg_tokens": null,
   "tps": 29.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-au2wktui-ctx4096",
   "model": "Huihui-ThinkingCap-Qwen3.6-27B-abliterated",
   "notes": "oMLX MTP build; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: g9tb1prq, jrplfhqz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 239.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Huihui-ThinkingCap-Qwen3.6-27B-abliterated-MLX-4bit-oMLX-MTP on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 239.8 27.4 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/au2wktui",
   "source_url": "https://omlx.ai/benchmarks/performance/au2wktui",
   "tg_tokens": null,
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-au2wktui-ctx8192",
   "model": "Huihui-ThinkingCap-Qwen3.6-27B-abliterated",
   "notes": "oMLX MTP build; oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: g9tb1prq, jrplfhqz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 233.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Huihui-ThinkingCap-Qwen3.6-27B-abliterated-MLX-4bit-oMLX-MTP on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 233.3 26.4 19.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/au2wktui",
   "source_url": "https://omlx.ai/benchmarks/performance/au2wktui",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-17gcw83p-ctx32768",
   "model": "LFM2.5-8B-A1B",
   "notes": "oMLX community benchmark trace; peak memory 39.6 GB at this context; duplicate traces of the same session: 6zcflbrw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2866.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "LFM2.5-8B-A1B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 unknown; 32k 2,866 37.5 39.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/17gcw83p",
   "source_url": "https://omlx.ai/benchmarks/performance/17gcw83p",
   "tg_tokens": null,
   "tps": 37.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-17gcw83p-ctx65536",
   "model": "LFM2.5-8B-A1B",
   "notes": "oMLX community benchmark trace; peak memory 44.9 GB at this context; duplicate traces of the same session: 6zcflbrw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1262.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "LFM2.5-8B-A1B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 unknown; 64k 1,262 74.1 44.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/17gcw83p",
   "source_url": "https://omlx.ai/benchmarks/performance/17gcw83p",
   "tg_tokens": null,
   "tps": 74.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-jew0hojj-ctx1024",
   "model": "LFM2.5-8B-A1B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 8.9 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1369.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-8B-A1B-MLX-8bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 8bit; 1k 1,369 107.0 8.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jew0hojj",
   "source_url": "https://omlx.ai/benchmarks/performance/jew0hojj",
   "tg_tokens": null,
   "tps": 107.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-jew0hojj-ctx4096",
   "model": "LFM2.5-8B-A1B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 9.3 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1619.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-8B-A1B-MLX-8bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 8bit; 4k 1,619 106.1 9.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jew0hojj",
   "source_url": "https://omlx.ai/benchmarks/performance/jew0hojj",
   "tg_tokens": null,
   "tps": 106.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-jew0hojj-ctx8192",
   "model": "LFM2.5-8B-A1B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 9.3 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1582.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-8B-A1B-MLX-8bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 8bit; 8k 1,582 103.1 9.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jew0hojj",
   "source_url": "https://omlx.ai/benchmarks/performance/jew0hojj",
   "tg_tokens": null,
   "tps": 103.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-jew0hojj-ctx16384",
   "model": "LFM2.5-8B-A1B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 9.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1495.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-8B-A1B-MLX-8bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 8bit; 16k 1,495 94.5 9.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jew0hojj",
   "source_url": "https://omlx.ai/benchmarks/performance/jew0hojj",
   "tg_tokens": null,
   "tps": 94.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-jew0hojj-ctx32768",
   "model": "LFM2.5-8B-A1B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 9.6 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1338.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-8B-A1B-MLX-8bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 8bit; 32k 1,338 81.1 9.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jew0hojj",
   "source_url": "https://omlx.ai/benchmarks/performance/jew0hojj",
   "tg_tokens": null,
   "tps": 81.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jq9xgc5q-ctx131072",
   "model": "LFM2.5-VL-1.6B",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 9.9 GB; TTFT 42109 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3113.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "LFM2.5-VL-1.6B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 unknown; 128k tokens 3,113 PP tok/s 134.2 TG tok/s 42109 TTFT (ms) 9.9 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jq9xgc5q",
   "source_url": "https://omlx.ai/benchmarks/performance/jq9xgc5q",
   "tg_tokens": null,
   "tps": 134.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kzsxjslm-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 29eq1h5d, kct74bu8, o003zx38; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3438.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 3,438 105.6 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kzsxjslm",
   "source_url": "https://omlx.ai/benchmarks/performance/kzsxjslm",
   "tg_tokens": null,
   "tps": 105.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kzsxjslm-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 29eq1h5d, kct74bu8, o003zx38; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4382.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 4,382 105.4 20.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kzsxjslm",
   "source_url": "https://omlx.ai/benchmarks/performance/kzsxjslm",
   "tg_tokens": null,
   "tps": 105.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kzsxjslm-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 29eq1h5d, kct74bu8, o003zx38; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4131.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 4,131 99.6 20.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kzsxjslm",
   "source_url": "https://omlx.ai/benchmarks/performance/kzsxjslm",
   "tg_tokens": null,
   "tps": 99.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kzsxjslm-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 29eq1h5d, kct74bu8, o003zx38; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3634.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 3,634 105.6 21.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kzsxjslm",
   "source_url": "https://omlx.ai/benchmarks/performance/kzsxjslm",
   "tg_tokens": null,
   "tps": 105.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-y17d0kfj-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 2x9940sk, 3frzk2nn, mdxvxq8u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3463.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 3,463 126.3 19.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y17d0kfj",
   "source_url": "https://omlx.ai/benchmarks/performance/y17d0kfj",
   "tg_tokens": null,
   "tps": 126.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-y17d0kfj-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 2x9940sk, 3frzk2nn, mdxvxq8u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4479.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 4,479 124.4 20.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y17d0kfj",
   "source_url": "https://omlx.ai/benchmarks/performance/y17d0kfj",
   "tg_tokens": null,
   "tps": 124.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-y17d0kfj-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 2x9940sk, 3frzk2nn, mdxvxq8u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4304.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 4,304 118.7 20.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y17d0kfj",
   "source_url": "https://omlx.ai/benchmarks/performance/y17d0kfj",
   "tg_tokens": null,
   "tps": 118.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-y17d0kfj-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 2x9940sk, 3frzk2nn, mdxvxq8u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3752.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 3,752 102.5 21.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y17d0kfj",
   "source_url": "https://omlx.ai/benchmarks/performance/y17d0kfj",
   "tg_tokens": null,
   "tps": 102.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8i4ry3id-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: a4a28f08, lm84t2yp, xnm490p7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3378.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 3,378 97.3 19.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8i4ry3id",
   "source_url": "https://omlx.ai/benchmarks/performance/8i4ry3id",
   "tg_tokens": null,
   "tps": 97.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8i4ry3id-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: a4a28f08, lm84t2yp, xnm490p7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2640.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,640 99.8 20.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8i4ry3id",
   "source_url": "https://omlx.ai/benchmarks/performance/8i4ry3id",
   "tg_tokens": null,
   "tps": 99.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8i4ry3id-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: a4a28f08, lm84t2yp, xnm490p7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2807.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,807 101.5 20.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8i4ry3id",
   "source_url": "https://omlx.ai/benchmarks/performance/8i4ry3id",
   "tg_tokens": null,
   "tps": 101.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8i4ry3id-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: a4a28f08, lm84t2yp, xnm490p7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2787.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,787 94.4 21.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8i4ry3id",
   "source_url": "https://omlx.ai/benchmarks/performance/8i4ry3id",
   "tg_tokens": null,
   "tps": 94.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-ko4wws8u-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: qzytnjlx; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 312.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 312.4 19.1 6.5 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ko4wws8u",
   "source_url": "https://omlx.ai/benchmarks/performance/ko4wws8u",
   "tg_tokens": null,
   "tps": 19.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-ko4wws8u-ctx16384",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.1 GB at this context; duplicate traces of the same session: qzytnjlx; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 314.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 16k 314.0 18.4 7.1 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ko4wws8u",
   "source_url": "https://omlx.ai/benchmarks/performance/ko4wws8u",
   "tg_tokens": null,
   "tps": 18.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-lwziiiwf-ctx1024",
   "model": "Qwen2.5-0.5B-Instruct",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 0.9 GB at this context; duplicate traces of the same session: nomf7gp5; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3657.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-0.5B-Instruct-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 3,657 198.3 0.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lwziiiwf",
   "source_url": "https://omlx.ai/benchmarks/performance/lwziiiwf",
   "tg_tokens": null,
   "tps": 198.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-lwziiiwf-ctx4096",
   "model": "Qwen2.5-0.5B-Instruct",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 1.1 GB at this context; duplicate traces of the same session: nomf7gp5; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3530.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-0.5B-Instruct-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 3,530 201.1 1.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lwziiiwf",
   "source_url": "https://omlx.ai/benchmarks/performance/lwziiiwf",
   "tg_tokens": null,
   "tps": 201.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2g4bspu2-ctx1024",
   "model": "Qwen2.5-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 4.8 GB at this context; duplicate traces of the same session: s2vuu3kh; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2871.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-7B-Instruct-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 2,871 87.7 4.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2g4bspu2",
   "source_url": "https://omlx.ai/benchmarks/performance/2g4bspu2",
   "tg_tokens": null,
   "tps": 87.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2g4bspu2-ctx4096",
   "model": "Qwen2.5-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: s2vuu3kh; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3091.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-7B-Instruct-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 3,091 86.8 4.9 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2g4bspu2",
   "source_url": "https://omlx.ai/benchmarks/performance/2g4bspu2",
   "tg_tokens": null,
   "tps": 86.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-zu9ellgr-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 21.3 GB; TTFT 1722 ms; duplicate traces of the same session: none; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 594.5,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 1k tokens 594.5 PP tok/s 40.6 TG tok/s 1722 TTFT (ms) 21.3 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zu9ellgr",
   "source_url": "https://omlx.ai/benchmarks/performance/zu9ellgr",
   "tg_tokens": null,
   "tps": 40.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-f06wv81j-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 20.0 GB; TTFT 11866 ms; duplicate traces of the same session: none; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 86.3,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 1k tokens 86.3 PP tok/s 6.7 TG tok/s 11866 TTFT (ms) 20.0 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f06wv81j",
   "source_url": "https://omlx.ai/benchmarks/performance/f06wv81j",
   "tg_tokens": null,
   "tps": 6.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-a3klq3gr-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 28.6 GB; TTFT 85126 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 384.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k tokens 384.9 PP tok/s 63.0 TG tok/s 85126 TTFT (ms) 28.6 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a3klq3gr",
   "source_url": "https://omlx.ai/benchmarks/performance/a3klq3gr",
   "tg_tokens": null,
   "tps": 63.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-jmero0dq-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 19.3 GB; TTFT 72042 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 454.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k tokens 454.8 PP tok/s 15.1 TG tok/s 72042 TTFT (ms) 19.3 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jmero0dq",
   "source_url": "https://omlx.ai/benchmarks/performance/jmero0dq",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-bc1phoai-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 20.6 GB; TTFT 75787 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 432.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k tokens 432.4 PP tok/s 41.3 TG tok/s 75787 TTFT (ms) 20.6 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bc1phoai",
   "source_url": "https://omlx.ai/benchmarks/performance/bc1phoai",
   "tg_tokens": null,
   "tps": 41.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-96htbmz3-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 142.6 GB at this context; duplicate traces of the same session: 4xily9vw, cb92joya, cv4cil1y, hnoh85q0, tp69c4nu; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1055.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 1k 1,055 83.6 142.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/96htbmz3",
   "source_url": "https://omlx.ai/benchmarks/performance/96htbmz3",
   "tg_tokens": null,
   "tps": 83.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-96htbmz3-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: 4xily9vw, cb92joya, cv4cil1y, hnoh85q0, tp69c4nu; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1301.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 4k 1,301 84.3 143.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/96htbmz3",
   "source_url": "https://omlx.ai/benchmarks/performance/96htbmz3",
   "tg_tokens": null,
   "tps": 84.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-96htbmz3-ctx16384",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 143.7 GB at this context; duplicate traces of the same session: 4xily9vw, cb92joya, cv4cil1y, hnoh85q0, tp69c4nu; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1338.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 16k 1,338 85.6 143.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/96htbmz3",
   "source_url": "https://omlx.ai/benchmarks/performance/96htbmz3",
   "tg_tokens": null,
   "tps": 85.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-96htbmz3-ctx65536",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 146.6 GB at this context; duplicate traces of the same session: 4xily9vw, cb92joya, cv4cil1y, hnoh85q0, tp69c4nu; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1315.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 64k 1,315 86.6 146.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/96htbmz3",
   "source_url": "https://omlx.ai/benchmarks/performance/96htbmz3",
   "tg_tokens": null,
   "tps": 86.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-96htbmz3-ctx131072",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 151.2 GB at this context; duplicate traces of the same session: 4xily9vw, cb92joya, cv4cil1y, hnoh85q0, tp69c4nu; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1292.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 128k 1,292 63.5 151.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/96htbmz3",
   "source_url": "https://omlx.ai/benchmarks/performance/96htbmz3",
   "tg_tokens": null,
   "tps": 63.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-96htbmz3-ctx199680",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 152.2 GB at this context; duplicate traces of the same session: 4xily9vw, cb92joya, cv4cil1y, hnoh85q0, tp69c4nu; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1275.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 195k 1,275 52.0 152.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/96htbmz3",
   "source_url": "https://omlx.ai/benchmarks/performance/96htbmz3",
   "tg_tokens": null,
   "tps": 52.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-yd3avf3k-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 70.8 GB; TTFT 82499 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 397.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 32k tokens 397.2 PP tok/s 32.8 TG tok/s 82499 TTFT (ms) 70.8 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yd3avf3k",
   "source_url": "https://omlx.ai/benchmarks/performance/yd3avf3k",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-hxc9ucw0-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.7 GB; TTFT 2254 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 454.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 454.3 PP tok/s 61.7 TG tok/s 2254 TTFT (ms) 70.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hxc9ucw0",
   "source_url": "https://omlx.ai/benchmarks/performance/hxc9ucw0",
   "tg_tokens": null,
   "tps": 61.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-onmimkqa-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.7 GB; TTFT 3885 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 263.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 263.6 PP tok/s 57.5 TG tok/s 3885 TTFT (ms) 70.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/onmimkqa",
   "source_url": "https://omlx.ai/benchmarks/performance/onmimkqa",
   "tg_tokens": null,
   "tps": 57.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5ojksjrb-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.4 GB at this context; duplicate traces of the same session: 2ypkbphy, mcu34qlg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 386.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 386.5 43.7 71.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ojksjrb",
   "source_url": "https://omlx.ai/benchmarks/performance/5ojksjrb",
   "tg_tokens": null,
   "tps": 43.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5ojksjrb-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.5 GB at this context; duplicate traces of the same session: 2ypkbphy, mcu34qlg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 606.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 606.0 44.5 72.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ojksjrb",
   "source_url": "https://omlx.ai/benchmarks/performance/5ojksjrb",
   "tg_tokens": null,
   "tps": 44.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5ojksjrb-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.7 GB at this context; duplicate traces of the same session: 2ypkbphy, mcu34qlg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 838.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 838.8 53.7 72.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ojksjrb",
   "source_url": "https://omlx.ai/benchmarks/performance/5ojksjrb",
   "tg_tokens": null,
   "tps": 53.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ztknk1qj-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.4 GB at this context; duplicate traces of the same session: 60yoghd4, 67xewd7x, wqzltixp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 986.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 986.8 45.7 71.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ztknk1qj",
   "source_url": "https://omlx.ai/benchmarks/performance/ztknk1qj",
   "tg_tokens": null,
   "tps": 45.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ztknk1qj-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.5 GB at this context; duplicate traces of the same session: 60yoghd4, 67xewd7x, wqzltixp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1009.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,009 48.7 72.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ztknk1qj",
   "source_url": "https://omlx.ai/benchmarks/performance/ztknk1qj",
   "tg_tokens": null,
   "tps": 48.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ztknk1qj-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.7 GB at this context; duplicate traces of the same session: 60yoghd4, 67xewd7x, wqzltixp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 916.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 916.6 36.8 72.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ztknk1qj",
   "source_url": "https://omlx.ai/benchmarks/performance/ztknk1qj",
   "tg_tokens": null,
   "tps": 36.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ztknk1qj-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.5 GB at this context; duplicate traces of the same session: 60yoghd4, 67xewd7x, wqzltixp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 828.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 828.9 43.1 73.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ztknk1qj",
   "source_url": "https://omlx.ai/benchmarks/performance/ztknk1qj",
   "tg_tokens": null,
   "tps": 43.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2ntlkzt6-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.4 GB at this context; duplicate traces of the same session: 2pm4t74m, i7yg4w8c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 980.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 980.8 45.9 71.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ntlkzt6",
   "source_url": "https://omlx.ai/benchmarks/performance/2ntlkzt6",
   "tg_tokens": null,
   "tps": 45.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2ntlkzt6-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.5 GB at this context; duplicate traces of the same session: 2pm4t74m, i7yg4w8c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1177.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,177 15.1 72.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ntlkzt6",
   "source_url": "https://omlx.ai/benchmarks/performance/2ntlkzt6",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2ntlkzt6-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.5 GB at this context; duplicate traces of the same session: 2pm4t74m, i7yg4w8c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1103.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,103 50.9 73.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ntlkzt6",
   "source_url": "https://omlx.ai/benchmarks/performance/2ntlkzt6",
   "tg_tokens": null,
   "tps": 50.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ervppd8d-ctx1024",
   "model": "ThinkingCap-Qwen3.6-27B",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 13.3 GB at this context; duplicate traces of the same session: j7mr3ejb, ko4d7o8g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 260.3,
   "provenance": "community",
   "quant": "3bit",
   "quote": "ThinkingCap-Qwen3.6-27B-3bit-mlx on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 260.3 34.8 13.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ervppd8d",
   "source_url": "https://omlx.ai/benchmarks/performance/ervppd8d",
   "tg_tokens": null,
   "tps": 34.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ervppd8d-ctx4096",
   "model": "ThinkingCap-Qwen3.6-27B",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: j7mr3ejb, ko4d7o8g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 215.5,
   "provenance": "community",
   "quant": "3bit",
   "quote": "ThinkingCap-Qwen3.6-27B-3bit-mlx on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 215.5 26.0 16.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ervppd8d",
   "source_url": "https://omlx.ai/benchmarks/performance/ervppd8d",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ervppd8d-ctx8192",
   "model": "ThinkingCap-Qwen3.6-27B",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 16.8 GB at this context; duplicate traces of the same session: j7mr3ejb, ko4d7o8g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 197.4,
   "provenance": "community",
   "quant": "3bit",
   "quote": "ThinkingCap-Qwen3.6-27B-3bit-mlx on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 197.4 23.8 16.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ervppd8d",
   "source_url": "https://omlx.ai/benchmarks/performance/ervppd8d",
   "tg_tokens": null,
   "tps": 23.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ahz7r3pe-ctx1024",
   "model": "ThinkingCap-Qwen3.6-27B",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 13.3 GB at this context; duplicate traces of the same session: 47tni6ed, zuxjndpv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 259.4,
   "provenance": "community",
   "quant": "3bit",
   "quote": "ThinkingCap-Qwen3.6-27B-3bit-mlx on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 259.4 34.7 13.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ahz7r3pe",
   "source_url": "https://omlx.ai/benchmarks/performance/ahz7r3pe",
   "tg_tokens": null,
   "tps": 34.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ahz7r3pe-ctx4096",
   "model": "ThinkingCap-Qwen3.6-27B",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: 47tni6ed, zuxjndpv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 217.2,
   "provenance": "community",
   "quant": "3bit",
   "quote": "ThinkingCap-Qwen3.6-27B-3bit-mlx on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 217.2 29.6 16.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ahz7r3pe",
   "source_url": "https://omlx.ai/benchmarks/performance/ahz7r3pe",
   "tg_tokens": null,
   "tps": 29.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ahz7r3pe-ctx8192",
   "model": "ThinkingCap-Qwen3.6-27B",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 16.8 GB at this context; duplicate traces of the same session: 47tni6ed, zuxjndpv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 207.5,
   "provenance": "community",
   "quant": "3bit",
   "quote": "ThinkingCap-Qwen3.6-27B-3bit-mlx on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 207.5 28.4 16.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ahz7r3pe",
   "source_url": "https://omlx.ai/benchmarks/performance/ahz7r3pe",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-9ldzzrce-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace, single run at 16k context; peak memory 18.2 GB; TTFT 50674 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k tokens 323.3 PP tok/s 32.2 TG tok/s 50674 TTFT (ms) 18.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ldzzrce",
   "source_url": "https://omlx.ai/benchmarks/performance/9ldzzrce",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-eqwb83mb-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 15.3 GB at this context; duplicate traces of the same session: l6t97frp; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 340.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 340.6 36.0 15.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eqwb83mb",
   "source_url": "https://omlx.ai/benchmarks/performance/eqwb83mb",
   "tg_tokens": null,
   "tps": 36.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-eqwb83mb-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: l6t97frp; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 293.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k 293.8 20.5 23.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eqwb83mb",
   "source_url": "https://omlx.ai/benchmarks/performance/eqwb83mb",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-4gn3a4ru-ctx1024",
   "model": "LFM2.5-8B-A1B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 8.9 GB at this context; duplicate traces of the same session: c0ve41sx, fxxyceh0, s00x8riy; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1369.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-8B-A1B-MLX-8bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 8bit; 1k 1,369 107.0 8.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4gn3a4ru",
   "source_url": "https://omlx.ai/benchmarks/performance/4gn3a4ru",
   "tg_tokens": null,
   "tps": 107.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-4gn3a4ru-ctx4096",
   "model": "LFM2.5-8B-A1B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 9.3 GB at this context; duplicate traces of the same session: c0ve41sx, fxxyceh0, s00x8riy; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1619.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-8B-A1B-MLX-8bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 8bit; 4k 1,619 106.1 9.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4gn3a4ru",
   "source_url": "https://omlx.ai/benchmarks/performance/4gn3a4ru",
   "tg_tokens": null,
   "tps": 106.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-4gn3a4ru-ctx8192",
   "model": "LFM2.5-8B-A1B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 9.3 GB at this context; duplicate traces of the same session: c0ve41sx, fxxyceh0, s00x8riy; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1582.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-8B-A1B-MLX-8bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 8bit; 8k 1,582 103.1 9.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4gn3a4ru",
   "source_url": "https://omlx.ai/benchmarks/performance/4gn3a4ru",
   "tg_tokens": null,
   "tps": 103.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-4gn3a4ru-ctx16384",
   "model": "LFM2.5-8B-A1B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 9.4 GB at this context; duplicate traces of the same session: c0ve41sx, fxxyceh0, s00x8riy; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1495.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-8B-A1B-MLX-8bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 8bit; 16k 1,495 94.5 9.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4gn3a4ru",
   "source_url": "https://omlx.ai/benchmarks/performance/4gn3a4ru",
   "tg_tokens": null,
   "tps": 94.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-4gn3a4ru-ctx32768",
   "model": "LFM2.5-8B-A1B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 9.6 GB at this context; duplicate traces of the same session: c0ve41sx, fxxyceh0, s00x8riy; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1338.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-8B-A1B-MLX-8bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 8bit; 32k 1,338 81.1 9.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4gn3a4ru",
   "source_url": "https://omlx.ai/benchmarks/performance/4gn3a4ru",
   "tg_tokens": null,
   "tps": 81.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yh5n85f2-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 0szo93vl, 3j9j1cmm, 41fdxx1k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3483.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 3,483 119.9 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yh5n85f2",
   "source_url": "https://omlx.ai/benchmarks/performance/yh5n85f2",
   "tg_tokens": null,
   "tps": 119.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yh5n85f2-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: 0szo93vl, 3j9j1cmm, 41fdxx1k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4568.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 4,568 117.7 20.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yh5n85f2",
   "source_url": "https://omlx.ai/benchmarks/performance/yh5n85f2",
   "tg_tokens": null,
   "tps": 117.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yh5n85f2-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: 0szo93vl, 3j9j1cmm, 41fdxx1k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4352.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 4,352 110.3 20.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yh5n85f2",
   "source_url": "https://omlx.ai/benchmarks/performance/yh5n85f2",
   "tg_tokens": null,
   "tps": 110.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yh5n85f2-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: 0szo93vl, 3j9j1cmm, 41fdxx1k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4083.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 4,083 112.9 21.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yh5n85f2",
   "source_url": "https://omlx.ai/benchmarks/performance/yh5n85f2",
   "tg_tokens": null,
   "tps": 112.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2feu2xg5-ctx199680",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 195k context; peak memory 30.8 GB; TTFT 128876 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1552.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-Safetensors-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 5bit; 195k tokens 1,552 PP tok/s 128.2 TG tok/s 128876 TTFT (ms) 30.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2feu2xg5",
   "source_url": "https://omlx.ai/benchmarks/performance/2feu2xg5",
   "tg_tokens": null,
   "tps": 128.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4u9wlink-ctx16384",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 27.0 GB; TTFT 3733 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4389.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-Safetensors-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 5bit; 16k tokens 4,389 PP tok/s 178.1 TG tok/s 3733 TTFT (ms) 27.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4u9wlink",
   "source_url": "https://omlx.ai/benchmarks/performance/4u9wlink",
   "tg_tokens": null,
   "tps": 178.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qgauuwun-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 7dcx28od, 9yinbuhg, hv30k3oh, jsenlf7f, l2h5m4oa, od3heikk, xdzpms9d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 929.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 929.2 32.9 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qgauuwun",
   "source_url": "https://omlx.ai/benchmarks/performance/qgauuwun",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qgauuwun-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 7dcx28od, 9yinbuhg, hv30k3oh, jsenlf7f, l2h5m4oa, od3heikk, xdzpms9d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1024.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,024 32.2 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qgauuwun",
   "source_url": "https://omlx.ai/benchmarks/performance/qgauuwun",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qgauuwun-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 7dcx28od, 9yinbuhg, hv30k3oh, jsenlf7f, l2h5m4oa, od3heikk, xdzpms9d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 935.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 935.7 31.5 17.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qgauuwun",
   "source_url": "https://omlx.ai/benchmarks/performance/qgauuwun",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qgauuwun-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 7dcx28od, 9yinbuhg, hv30k3oh, jsenlf7f, l2h5m4oa, od3heikk, xdzpms9d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 830.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 830.9 30.7 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qgauuwun",
   "source_url": "https://omlx.ai/benchmarks/performance/qgauuwun",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qgauuwun-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 7dcx28od, 9yinbuhg, hv30k3oh, jsenlf7f, l2h5m4oa, od3heikk, xdzpms9d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 770.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 770.7 28.5 18.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qgauuwun",
   "source_url": "https://omlx.ai/benchmarks/performance/qgauuwun",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qgauuwun-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 7dcx28od, 9yinbuhg, hv30k3oh, jsenlf7f, l2h5m4oa, od3heikk, xdzpms9d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 655.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 655.7 24.6 20.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qgauuwun",
   "source_url": "https://omlx.ai/benchmarks/performance/qgauuwun",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qgauuwun-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: 7dcx28od, 9yinbuhg, hv30k3oh, jsenlf7f, l2h5m4oa, od3heikk, xdzpms9d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 475.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 475.4 18.6 24.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qgauuwun",
   "source_url": "https://omlx.ai/benchmarks/performance/qgauuwun",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qgauuwun-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.0 GB at this context; duplicate traces of the same session: 7dcx28od, 9yinbuhg, hv30k3oh, jsenlf7f, l2h5m4oa, od3heikk, xdzpms9d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 369.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 195k 369.7 14.4 29.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qgauuwun",
   "source_url": "https://omlx.ai/benchmarks/performance/qgauuwun",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-y0bv35qu-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: gxditjoc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1334.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,334 58.9 19.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y0bv35qu",
   "source_url": "https://omlx.ai/benchmarks/performance/y0bv35qu",
   "tg_tokens": null,
   "tps": 58.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-y0bv35qu-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: gxditjoc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 948.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 948.0 57.1 20.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y0bv35qu",
   "source_url": "https://omlx.ai/benchmarks/performance/y0bv35qu",
   "tg_tokens": null,
   "tps": 57.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mtuo91in-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 26f80qfi, g9bpnyer; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 910.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 910.8 31.2 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mtuo91in",
   "source_url": "https://omlx.ai/benchmarks/performance/mtuo91in",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mtuo91in-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 26f80qfi, g9bpnyer; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1004.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,004 30.6 17.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mtuo91in",
   "source_url": "https://omlx.ai/benchmarks/performance/mtuo91in",
   "tg_tokens": null,
   "tps": 30.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mtuo91in-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 26f80qfi, g9bpnyer; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 926.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 926.8 29.6 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mtuo91in",
   "source_url": "https://omlx.ai/benchmarks/performance/mtuo91in",
   "tg_tokens": null,
   "tps": 29.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mtuo91in-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: 26f80qfi, g9bpnyer; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 795.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 795.6 29.0 18.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mtuo91in",
   "source_url": "https://omlx.ai/benchmarks/performance/mtuo91in",
   "tg_tokens": null,
   "tps": 29.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mtuo91in-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 26f80qfi, g9bpnyer; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 732.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 732.5 27.5 19.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mtuo91in",
   "source_url": "https://omlx.ai/benchmarks/performance/mtuo91in",
   "tg_tokens": null,
   "tps": 27.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mtuo91in-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 26f80qfi, g9bpnyer; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 555.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 555.8 21.1 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mtuo91in",
   "source_url": "https://omlx.ai/benchmarks/performance/mtuo91in",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mtuo91in-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: 26f80qfi, g9bpnyer; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 437.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 437.1 16.5 25.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mtuo91in",
   "source_url": "https://omlx.ai/benchmarks/performance/mtuo91in",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mtuo91in-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: 26f80qfi, g9bpnyer; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 358.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 195k 358.1 13.9 29.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mtuo91in",
   "source_url": "https://omlx.ai/benchmarks/performance/mtuo91in",
   "tg_tokens": null,
   "tps": 13.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rl6urpvp-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 69.6 GB at this context; duplicate traces of the same session: otiswftz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2037.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,037 72.0 69.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rl6urpvp",
   "source_url": "https://omlx.ai/benchmarks/performance/rl6urpvp",
   "tg_tokens": null,
   "tps": 72.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rl6urpvp-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 70.2 GB at this context; duplicate traces of the same session: otiswftz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2886.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,886 69.1 70.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rl6urpvp",
   "source_url": "https://omlx.ai/benchmarks/performance/rl6urpvp",
   "tg_tokens": null,
   "tps": 69.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-snx83emo-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 69.6 GB at this context; duplicate traces of the same session: br77tw4z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2043.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,043 72.6 69.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/snx83emo",
   "source_url": "https://omlx.ai/benchmarks/performance/snx83emo",
   "tg_tokens": null,
   "tps": 72.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-snx83emo-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 70.2 GB at this context; duplicate traces of the same session: br77tw4z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2902.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,902 69.6 70.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/snx83emo",
   "source_url": "https://omlx.ai/benchmarks/performance/snx83emo",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-g5fe0buv-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 69.6 GB at this context; duplicate traces of the same session: a9mi0iyf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2048.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,048 69.2 69.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g5fe0buv",
   "source_url": "https://omlx.ai/benchmarks/performance/g5fe0buv",
   "tg_tokens": null,
   "tps": 69.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-g5fe0buv-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 70.2 GB at this context; duplicate traces of the same session: a9mi0iyf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2829.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,829 64.9 70.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g5fe0buv",
   "source_url": "https://omlx.ai/benchmarks/performance/g5fe0buv",
   "tg_tokens": null,
   "tps": 64.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hwtsslc6-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 69.6 GB; TTFT 512 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2001.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 2,001 PP tok/s 69.8 TG tok/s 512 TTFT (ms) 69.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hwtsslc6",
   "source_url": "https://omlx.ai/benchmarks/performance/hwtsslc6",
   "tg_tokens": null,
   "tps": 69.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-w2t0qwky-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 69.6 GB at this context; duplicate traces of the same session: q2a0cz4l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1974.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,974 69.7 69.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w2t0qwky",
   "source_url": "https://omlx.ai/benchmarks/performance/w2t0qwky",
   "tg_tokens": null,
   "tps": 69.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-w2t0qwky-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 70.2 GB at this context; duplicate traces of the same session: q2a0cz4l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2839.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,839 65.0 70.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w2t0qwky",
   "source_url": "https://omlx.ai/benchmarks/performance/w2t0qwky",
   "tg_tokens": null,
   "tps": 65.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q81xyetn-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 69.6 GB at this context; duplicate traces of the same session: y4pkoot2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2040.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,040 68.8 69.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q81xyetn",
   "source_url": "https://omlx.ai/benchmarks/performance/q81xyetn",
   "tg_tokens": null,
   "tps": 68.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q81xyetn-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 70.2 GB at this context; duplicate traces of the same session: y4pkoot2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2869.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,869 66.4 70.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q81xyetn",
   "source_url": "https://omlx.ai/benchmarks/performance/q81xyetn",
   "tg_tokens": null,
   "tps": 66.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qid0r9qz-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 70.2 GB; TTFT 1459 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2808.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 2,808 PP tok/s 65.5 TG tok/s 1459 TTFT (ms) 70.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qid0r9qz",
   "source_url": "https://omlx.ai/benchmarks/performance/qid0r9qz",
   "tg_tokens": null,
   "tps": 65.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-x39eigyh-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 69.6 GB at this context; duplicate traces of the same session: 28grdk0z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1917.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,917 67.6 69.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x39eigyh",
   "source_url": "https://omlx.ai/benchmarks/performance/x39eigyh",
   "tg_tokens": null,
   "tps": 67.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-x39eigyh-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 70.2 GB at this context; duplicate traces of the same session: 28grdk0z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2746.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,746 64.9 70.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x39eigyh",
   "source_url": "https://omlx.ai/benchmarks/performance/x39eigyh",
   "tg_tokens": null,
   "tps": 64.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-kqyzkn6y-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: npjlqn5z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 220.9 25.8 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kqyzkn6y",
   "source_url": "https://omlx.ai/benchmarks/performance/kqyzkn6y",
   "tg_tokens": null,
   "tps": 25.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-kqyzkn6y-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: npjlqn5z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 282.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 282.5 27.8 71.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kqyzkn6y",
   "source_url": "https://omlx.ai/benchmarks/performance/kqyzkn6y",
   "tg_tokens": null,
   "tps": 27.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-ri5k5day-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: cg0wxky5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 298.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 298.7 36.0 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ri5k5day",
   "source_url": "https://omlx.ai/benchmarks/performance/ri5k5day",
   "tg_tokens": null,
   "tps": 36.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-ri5k5day-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: cg0wxky5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 323.7 35.1 71.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ri5k5day",
   "source_url": "https://omlx.ai/benchmarks/performance/ri5k5day",
   "tg_tokens": null,
   "tps": 35.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-tm0kxf68-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.7 GB; TTFT 4077 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 251.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 251.2 PP tok/s 31.5 TG tok/s 4077 TTFT (ms) 70.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tm0kxf68",
   "source_url": "https://omlx.ai/benchmarks/performance/tm0kxf68",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-csl24qot-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 430.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 430.8 45.1 70.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/csl24qot",
   "source_url": "https://omlx.ai/benchmarks/performance/csl24qot",
   "tg_tokens": null,
   "tps": 45.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-csl24qot-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 418.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 418.9 37.5 71.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/csl24qot",
   "source_url": "https://omlx.ai/benchmarks/performance/csl24qot",
   "tg_tokens": null,
   "tps": 37.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-dsnksg11-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 71.7 GB; TTFT 21050 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 778.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k tokens 778.3 PP tok/s 32.1 TG tok/s 21050 TTFT (ms) 71.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dsnksg11",
   "source_url": "https://omlx.ai/benchmarks/performance/dsnksg11",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ojr61t1a-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 905.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 905.8 39.7 72.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ojr61t1a",
   "source_url": "https://omlx.ai/benchmarks/performance/ojr61t1a",
   "tg_tokens": null,
   "tps": 39.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ojr61t1a-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 825.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 825.1 35.0 72.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ojr61t1a",
   "source_url": "https://omlx.ai/benchmarks/performance/ojr61t1a",
   "tg_tokens": null,
   "tps": 35.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ojr61t1a-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 800.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 800.2 40.1 73.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ojr61t1a",
   "source_url": "https://omlx.ai/benchmarks/performance/ojr61t1a",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ojr61t1a-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 789.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 789.4 30.4 74.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ojr61t1a",
   "source_url": "https://omlx.ai/benchmarks/performance/ojr61t1a",
   "tg_tokens": null,
   "tps": 30.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-a4ajktpy-ctx1024",
   "model": "ThinkingCap-Qwen3.6-27B",
   "notes": "DWQ MTP Vision MLX build; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 1xjifx29, a9rj7afo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 246.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "ThinkingCap-Qwen3.6-27B-oQ4e-DWQ-MTP-Vision-MLX on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 246.8 47.2 19.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a4ajktpy",
   "source_url": "https://omlx.ai/benchmarks/performance/a4ajktpy",
   "tg_tokens": null,
   "tps": 47.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-a4ajktpy-ctx4096",
   "model": "ThinkingCap-Qwen3.6-27B",
   "notes": "DWQ MTP Vision MLX build; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: 1xjifx29, a9rj7afo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 231.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "ThinkingCap-Qwen3.6-27B-oQ4e-DWQ-MTP-Vision-MLX on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 231.7 35.9 22.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a4ajktpy",
   "source_url": "https://omlx.ai/benchmarks/performance/a4ajktpy",
   "tg_tokens": null,
   "tps": 35.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-a4ajktpy-ctx8192",
   "model": "ThinkingCap-Qwen3.6-27B",
   "notes": "DWQ MTP Vision MLX build; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: 1xjifx29, a9rj7afo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 221.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "ThinkingCap-Qwen3.6-27B-oQ4e-DWQ-MTP-Vision-MLX on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 221.6 32.1 23.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a4ajktpy",
   "source_url": "https://omlx.ai/benchmarks/performance/a4ajktpy",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hi473410-ctx1024",
   "model": "Qwen3.6-27B-MTPLX-Optimized-Speed-V2",
   "notes": "Youssofal build; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: q7axil2n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 722.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Youssofal--Qwen3.6-27B-MTPLX-Optimized-Speed-V2 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 722.1 22.9 20.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hi473410",
   "source_url": "https://omlx.ai/benchmarks/performance/hi473410",
   "tg_tokens": null,
   "tps": 22.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hi473410-ctx4096",
   "model": "Qwen3.6-27B-MTPLX-Optimized-Speed-V2",
   "notes": "Youssofal build; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: q7axil2n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 746.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Youssofal--Qwen3.6-27B-MTPLX-Optimized-Speed-V2 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 746.0 22.2 21.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hi473410",
   "source_url": "https://omlx.ai/benchmarks/performance/hi473410",
   "tg_tokens": null,
   "tps": 22.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-oh0328yi-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: j2zaw0me; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 394.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "qwen3.8-27b-4bit-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 394.9 34.2 17.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oh0328yi",
   "source_url": "https://omlx.ai/benchmarks/performance/oh0328yi",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-oh0328yi-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.2 GB at this context; duplicate traces of the same session: j2zaw0me; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 396.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "qwen3.8-27b-4bit-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 396.3 35.5 18.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oh0328yi",
   "source_url": "https://omlx.ai/benchmarks/performance/oh0328yi",
   "tg_tokens": null,
   "tps": 35.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-07oygrvk-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.1 GB at this context; duplicate traces of the same session: 0r4vjc2w, ackwx9vg, qos49bic; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 66.2,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4.1-Flash-oQ2e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 66.2 13.1 100.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/07oygrvk",
   "source_url": "https://omlx.ai/benchmarks/performance/07oygrvk",
   "tg_tokens": null,
   "tps": 13.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-07oygrvk-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.4 GB at this context; duplicate traces of the same session: 0r4vjc2w, ackwx9vg, qos49bic; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 194.2,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4.1-Flash-oQ2e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 194.2 15.4 100.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/07oygrvk",
   "source_url": "https://omlx.ai/benchmarks/performance/07oygrvk",
   "tg_tokens": null,
   "tps": 15.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-07oygrvk-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: 0r4vjc2w, ackwx9vg, qos49bic; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 206.5,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4.1-Flash-oQ2e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 8k 206.5 17.1 100.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/07oygrvk",
   "source_url": "https://omlx.ai/benchmarks/performance/07oygrvk",
   "tg_tokens": null,
   "tps": 17.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-07oygrvk-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.4 GB at this context; duplicate traces of the same session: 0r4vjc2w, ackwx9vg, qos49bic; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.1,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4.1-Flash-oQ2e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 16k 204.1 18.8 100.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/07oygrvk",
   "source_url": "https://omlx.ai/benchmarks/performance/07oygrvk",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-scd14kcq-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.8 GB at this context; duplicate traces of the same session: fybobtd8, k4l5ius0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 507.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 507.0 37.6 175.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/scd14kcq",
   "source_url": "https://omlx.ai/benchmarks/performance/scd14kcq",
   "tg_tokens": null,
   "tps": 37.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-scd14kcq-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: fybobtd8, k4l5ius0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 554.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 554.1 67.5 177.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/scd14kcq",
   "source_url": "https://omlx.ai/benchmarks/performance/scd14kcq",
   "tg_tokens": null,
   "tps": 67.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-scd14kcq-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.0 GB at this context; duplicate traces of the same session: fybobtd8, k4l5ius0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 536.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 536.6 32.6 178.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/scd14kcq",
   "source_url": "https://omlx.ai/benchmarks/performance/scd14kcq",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-scd14kcq-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.7 GB at this context; duplicate traces of the same session: fybobtd8, k4l5ius0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 526.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 526.3 23.7 178.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/scd14kcq",
   "source_url": "https://omlx.ai/benchmarks/performance/scd14kcq",
   "tg_tokens": null,
   "tps": 23.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-j3ur090r-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 01ls98k0, 7e2j1nna, 80rso0km, sg55nxww, sogqm9oz, sqx8ytop; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2769.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,769 101.1 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j3ur090r",
   "source_url": "https://omlx.ai/benchmarks/performance/j3ur090r",
   "tg_tokens": null,
   "tps": 101.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-j3ur090r-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 01ls98k0, 7e2j1nna, 80rso0km, sg55nxww, sogqm9oz, sqx8ytop; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3302.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,302 106.1 20.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j3ur090r",
   "source_url": "https://omlx.ai/benchmarks/performance/j3ur090r",
   "tg_tokens": null,
   "tps": 106.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-j3ur090r-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 01ls98k0, 7e2j1nna, 80rso0km, sg55nxww, sogqm9oz, sqx8ytop; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3188.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 3,188 101.2 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j3ur090r",
   "source_url": "https://omlx.ai/benchmarks/performance/j3ur090r",
   "tg_tokens": null,
   "tps": 101.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-j3ur090r-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: 01ls98k0, 7e2j1nna, 80rso0km, sg55nxww, sogqm9oz, sqx8ytop; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2934.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,934 94.0 20.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j3ur090r",
   "source_url": "https://omlx.ai/benchmarks/performance/j3ur090r",
   "tg_tokens": null,
   "tps": 94.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-j3ur090r-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 01ls98k0, 7e2j1nna, 80rso0km, sg55nxww, sogqm9oz, sqx8ytop; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2493.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 2,493 81.8 20.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j3ur090r",
   "source_url": "https://omlx.ai/benchmarks/performance/j3ur090r",
   "tg_tokens": null,
   "tps": 81.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-j3ur090r-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 01ls98k0, 7e2j1nna, 80rso0km, sg55nxww, sogqm9oz, sqx8ytop; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1821.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 1,821 65.0 21.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j3ur090r",
   "source_url": "https://omlx.ai/benchmarks/performance/j3ur090r",
   "tg_tokens": null,
   "tps": 65.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-j3ur090r-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: 01ls98k0, 7e2j1nna, 80rso0km, sg55nxww, sogqm9oz, sqx8ytop; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1264.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 128k 1,264 46.7 22.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j3ur090r",
   "source_url": "https://omlx.ai/benchmarks/performance/j3ur090r",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ooyrmf8x-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: pqy1mcqs; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4576.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,576 200.9 21.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ooyrmf8x",
   "source_url": "https://omlx.ai/benchmarks/performance/ooyrmf8x",
   "tg_tokens": null,
   "tps": 200.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ooyrmf8x-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: pqy1mcqs; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6261.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,261 206.1 22.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ooyrmf8x",
   "source_url": "https://omlx.ai/benchmarks/performance/ooyrmf8x",
   "tg_tokens": null,
   "tps": 206.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-naw107e4-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: vlst2m3m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4513.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,513 219.3 21.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/naw107e4",
   "source_url": "https://omlx.ai/benchmarks/performance/naw107e4",
   "tg_tokens": null,
   "tps": 219.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-naw107e4-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: vlst2m3m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5725.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 5,725 190.5 22.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/naw107e4",
   "source_url": "https://omlx.ai/benchmarks/performance/naw107e4",
   "tg_tokens": null,
   "tps": 190.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2b8mge9c-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 21.4 GB; TTFT 227 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4515.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 4,515 PP tok/s 211.7 TG tok/s 227 TTFT (ms) 21.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2b8mge9c",
   "source_url": "https://omlx.ai/benchmarks/performance/2b8mge9c",
   "tg_tokens": null,
   "tps": 211.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fr25n6xl-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 1k context; peak memory 37.4 GB; TTFT 241 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4255.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k tokens 4,255 PP tok/s 136.9 TG tok/s 241 TTFT (ms) 37.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fr25n6xl",
   "source_url": "https://omlx.ai/benchmarks/performance/fr25n6xl",
   "tg_tokens": null,
   "tps": 136.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-p35oxotm-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 1k context; peak memory 37.4 GB; TTFT 242 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4229.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k tokens 4,229 PP tok/s 148.0 TG tok/s 242 TTFT (ms) 37.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p35oxotm",
   "source_url": "https://omlx.ai/benchmarks/performance/p35oxotm",
   "tg_tokens": null,
   "tps": 148.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-raaiq936-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 1k context; peak memory 37.4 GB; TTFT 240 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4274.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k tokens 4,274 PP tok/s 131.2 TG tok/s 240 TTFT (ms) 37.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/raaiq936",
   "source_url": "https://omlx.ai/benchmarks/performance/raaiq936",
   "tg_tokens": null,
   "tps": 131.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0puh2xvu-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: 50s2z2xh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4171.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 4,171 131.7 37.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0puh2xvu",
   "source_url": "https://omlx.ai/benchmarks/performance/0puh2xvu",
   "tg_tokens": null,
   "tps": 131.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0puh2xvu-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 38.0 GB at this context; duplicate traces of the same session: 50s2z2xh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5375.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 5,375 129.4 38.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0puh2xvu",
   "source_url": "https://omlx.ai/benchmarks/performance/0puh2xvu",
   "tg_tokens": null,
   "tps": 129.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-k8a6emf3-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 1k context; peak memory 35.5 GB; TTFT 227 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4509.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k tokens 4,509 PP tok/s 108.9 TG tok/s 227 TTFT (ms) 35.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k8a6emf3",
   "source_url": "https://omlx.ai/benchmarks/performance/k8a6emf3",
   "tg_tokens": null,
   "tps": 108.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-z27ydpz6-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 1k context; peak memory 35.5 GB; TTFT 223 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4587.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k tokens 4,587 PP tok/s 109.1 TG tok/s 223 TTFT (ms) 35.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z27ydpz6",
   "source_url": "https://omlx.ai/benchmarks/performance/z27ydpz6",
   "tg_tokens": null,
   "tps": 109.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9dl8rc6j-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: 2h48hkdk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 985.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 985.8 109.7 35.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9dl8rc6j",
   "source_url": "https://omlx.ai/benchmarks/performance/9dl8rc6j",
   "tg_tokens": null,
   "tps": 109.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9dl8rc6j-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: 2h48hkdk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3023.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 3,023 108.5 36.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9dl8rc6j",
   "source_url": "https://omlx.ai/benchmarks/performance/9dl8rc6j",
   "tg_tokens": null,
   "tps": 108.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-dr602o2l-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 20.0 GB; TTFT 18243 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 224.5 PP tok/s 20.1 TG tok/s 18243 TTFT (ms) 20.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dr602o2l",
   "source_url": "https://omlx.ai/benchmarks/performance/dr602o2l",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-f57ni21d-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 20.0 GB; TTFT 17287 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 236.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 236.9 PP tok/s 20.8 TG tok/s 17287 TTFT (ms) 20.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f57ni21d",
   "source_url": "https://omlx.ai/benchmarks/performance/f57ni21d",
   "tg_tokens": null,
   "tps": 20.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dvrsrgdo-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 6pympt0s, p2h94mjw, pvx6o9se, rm6xigfw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 910.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 910.8 31.2 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dvrsrgdo",
   "source_url": "https://omlx.ai/benchmarks/performance/dvrsrgdo",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dvrsrgdo-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 6pympt0s, p2h94mjw, pvx6o9se, rm6xigfw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1004.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,004 30.6 17.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dvrsrgdo",
   "source_url": "https://omlx.ai/benchmarks/performance/dvrsrgdo",
   "tg_tokens": null,
   "tps": 30.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dvrsrgdo-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 6pympt0s, p2h94mjw, pvx6o9se, rm6xigfw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 926.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 926.8 29.6 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dvrsrgdo",
   "source_url": "https://omlx.ai/benchmarks/performance/dvrsrgdo",
   "tg_tokens": null,
   "tps": 29.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dvrsrgdo-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: 6pympt0s, p2h94mjw, pvx6o9se, rm6xigfw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 795.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 795.6 29.0 18.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dvrsrgdo",
   "source_url": "https://omlx.ai/benchmarks/performance/dvrsrgdo",
   "tg_tokens": null,
   "tps": 29.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dvrsrgdo-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 6pympt0s, p2h94mjw, pvx6o9se, rm6xigfw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 732.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 732.5 27.5 19.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dvrsrgdo",
   "source_url": "https://omlx.ai/benchmarks/performance/dvrsrgdo",
   "tg_tokens": null,
   "tps": 27.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dvrsrgdo-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 6pympt0s, p2h94mjw, pvx6o9se, rm6xigfw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 555.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 555.8 21.1 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dvrsrgdo",
   "source_url": "https://omlx.ai/benchmarks/performance/dvrsrgdo",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dvrsrgdo-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: 6pympt0s, p2h94mjw, pvx6o9se, rm6xigfw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 437.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 437.1 16.5 25.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dvrsrgdo",
   "source_url": "https://omlx.ai/benchmarks/performance/dvrsrgdo",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dvrsrgdo-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: 6pympt0s, p2h94mjw, pvx6o9se, rm6xigfw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 358.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 195k 358.1 13.9 29.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dvrsrgdo",
   "source_url": "https://omlx.ai/benchmarks/performance/dvrsrgdo",
   "tg_tokens": null,
   "tps": 13.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cschump6-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 69.6 GB at this context; duplicate traces of the same session: mx1u7htn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1861.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,861 59.1 69.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cschump6",
   "source_url": "https://omlx.ai/benchmarks/performance/cschump6",
   "tg_tokens": null,
   "tps": 59.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cschump6-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 70.2 GB at this context; duplicate traces of the same session: mx1u7htn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2792.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,792 61.2 70.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cschump6",
   "source_url": "https://omlx.ai/benchmarks/performance/cschump6",
   "tg_tokens": null,
   "tps": 61.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lglwyj9g-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 69.6 GB at this context; duplicate traces of the same session: hwmx7xkm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2157.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,157 74.7 69.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lglwyj9g",
   "source_url": "https://omlx.ai/benchmarks/performance/lglwyj9g",
   "tg_tokens": null,
   "tps": 74.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lglwyj9g-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 70.2 GB at this context; duplicate traces of the same session: hwmx7xkm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3059.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 3,059 71.1 70.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lglwyj9g",
   "source_url": "https://omlx.ai/benchmarks/performance/lglwyj9g",
   "tg_tokens": null,
   "tps": 71.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pbez6sex-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 70.2 GB; TTFT 1480 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2768.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 2,768 PP tok/s 60.0 TG tok/s 1480 TTFT (ms) 70.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pbez6sex",
   "source_url": "https://omlx.ai/benchmarks/performance/pbez6sex",
   "tg_tokens": null,
   "tps": 60.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8o3xjs3y-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 69.6 GB at this context; duplicate traces of the same session: 4yk7sfpp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2148.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,148 75.6 69.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8o3xjs3y",
   "source_url": "https://omlx.ai/benchmarks/performance/8o3xjs3y",
   "tg_tokens": null,
   "tps": 75.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8o3xjs3y-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 70.2 GB at this context; duplicate traces of the same session: 4yk7sfpp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3041.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 3,041 71.2 70.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8o3xjs3y",
   "source_url": "https://omlx.ai/benchmarks/performance/8o3xjs3y",
   "tg_tokens": null,
   "tps": 71.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pi07xpmv-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 70.2 GB; TTFT 1358 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3016.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 3,016 PP tok/s 71.0 TG tok/s 1358 TTFT (ms) 70.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pi07xpmv",
   "source_url": "https://omlx.ai/benchmarks/performance/pi07xpmv",
   "tg_tokens": null,
   "tps": 71.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-pzk1ii7e-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 430.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 430.8 45.1 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pzk1ii7e",
   "source_url": "https://omlx.ai/benchmarks/performance/pzk1ii7e",
   "tg_tokens": null,
   "tps": 45.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-pzk1ii7e-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 418.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 418.9 37.5 71.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pzk1ii7e",
   "source_url": "https://omlx.ai/benchmarks/performance/pzk1ii7e",
   "tg_tokens": null,
   "tps": 37.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8flm2nw6-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 4mpebx3e, pk0nj6s3, wk1tt7f7, yd0bvp65; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 863.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 863.4 60.8 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8flm2nw6",
   "source_url": "https://omlx.ai/benchmarks/performance/8flm2nw6",
   "tg_tokens": null,
   "tps": 60.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8flm2nw6-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: 4mpebx3e, pk0nj6s3, wk1tt7f7, yd0bvp65; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 152.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 152.6 63.0 73.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8flm2nw6",
   "source_url": "https://omlx.ai/benchmarks/performance/8flm2nw6",
   "tg_tokens": null,
   "tps": 63.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8flm2nw6-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.0 GB at this context; duplicate traces of the same session: 4mpebx3e, pk0nj6s3, wk1tt7f7, yd0bvp65; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1525.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,525 63.6 72.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8flm2nw6",
   "source_url": "https://omlx.ai/benchmarks/performance/8flm2nw6",
   "tg_tokens": null,
   "tps": 63.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8flm2nw6-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.0 GB at this context; duplicate traces of the same session: 4mpebx3e, pk0nj6s3, wk1tt7f7, yd0bvp65; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1642.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,642 63.1 73.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8flm2nw6",
   "source_url": "https://omlx.ai/benchmarks/performance/8flm2nw6",
   "tg_tokens": null,
   "tps": 63.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8flm2nw6-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 76.6 GB at this context; duplicate traces of the same session: 4mpebx3e, pk0nj6s3, wk1tt7f7, yd0bvp65; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1358.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,358 45.7 76.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8flm2nw6",
   "source_url": "https://omlx.ai/benchmarks/performance/8flm2nw6",
   "tg_tokens": null,
   "tps": 45.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-s1rq31gx-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.5 GB at this context; duplicate traces of the same session: 29qu9p0a, 5n4grj2f, miawty6r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 973.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 973.4 62.2 70.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s1rq31gx",
   "source_url": "https://omlx.ai/benchmarks/performance/s1rq31gx",
   "tg_tokens": null,
   "tps": 62.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-s1rq31gx-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 29qu9p0a, 5n4grj2f, miawty6r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1546.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,546 62.3 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s1rq31gx",
   "source_url": "https://omlx.ai/benchmarks/performance/s1rq31gx",
   "tg_tokens": null,
   "tps": 62.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-s1rq31gx-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: 29qu9p0a, 5n4grj2f, miawty6r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1575.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,575 71.9 72.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s1rq31gx",
   "source_url": "https://omlx.ai/benchmarks/performance/s1rq31gx",
   "tg_tokens": null,
   "tps": 71.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-s1rq31gx-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.5 GB at this context; duplicate traces of the same session: 29qu9p0a, 5n4grj2f, miawty6r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1425.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,425 63.9 73.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s1rq31gx",
   "source_url": "https://omlx.ai/benchmarks/performance/s1rq31gx",
   "tg_tokens": null,
   "tps": 63.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-z16g5zra-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.5 GB at this context; duplicate traces of the same session: 39tpc40f, 576h4y58, l37nwdjt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1003.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,003 57.7 70.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z16g5zra",
   "source_url": "https://omlx.ai/benchmarks/performance/z16g5zra",
   "tg_tokens": null,
   "tps": 57.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-z16g5zra-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: 39tpc40f, 576h4y58, l37nwdjt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1205.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,205 66.3 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z16g5zra",
   "source_url": "https://omlx.ai/benchmarks/performance/z16g5zra",
   "tg_tokens": null,
   "tps": 66.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-z16g5zra-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 39tpc40f, 576h4y58, l37nwdjt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1353.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,353 60.5 71.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z16g5zra",
   "source_url": "https://omlx.ai/benchmarks/performance/z16g5zra",
   "tg_tokens": null,
   "tps": 60.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-z16g5zra-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.5 GB at this context; duplicate traces of the same session: 39tpc40f, 576h4y58, l37nwdjt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1707.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,707 78.5 73.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z16g5zra",
   "source_url": "https://omlx.ai/benchmarks/performance/z16g5zra",
   "tg_tokens": null,
   "tps": 78.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-y4b219si-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 20.5 GB; TTFT 2359 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1736.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k tokens 1,736 PP tok/s 67.4 TG tok/s 2359 TTFT (ms) 20.5 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y4b219si",
   "source_url": "https://omlx.ai/benchmarks/performance/y4b219si",
   "tg_tokens": null,
   "tps": 67.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-2zy7wha6-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 20.5 GB; TTFT 2354 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1740.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k tokens 1,740 PP tok/s 66.1 TG tok/s 2354 TTFT (ms) 20.5 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2zy7wha6",
   "source_url": "https://omlx.ai/benchmarks/performance/2zy7wha6",
   "tg_tokens": null,
   "tps": 66.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-fitkphun-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 20.5 GB; TTFT 2364 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1733.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k tokens 1,733 PP tok/s 60.9 TG tok/s 2364 TTFT (ms) 20.5 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fitkphun",
   "source_url": "https://omlx.ai/benchmarks/performance/fitkphun",
   "tg_tokens": null,
   "tps": 60.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ftcd3hav-ctx1024",
   "model": "DeepSeek-R1-Distill-Llama-70B",
   "notes": "oMLX community benchmark trace; peak memory 58.6 GB at this context; duplicate traces of the same session: deqjfwww; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 186.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-R1-Distill-Llama-70B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 186.4 10.9 58.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ftcd3hav",
   "source_url": "https://omlx.ai/benchmarks/performance/ftcd3hav",
   "tg_tokens": null,
   "tps": 10.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ftcd3hav-ctx4096",
   "model": "DeepSeek-R1-Distill-Llama-70B",
   "notes": "oMLX community benchmark trace; peak memory 65.1 GB at this context; duplicate traces of the same session: deqjfwww; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 279.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-R1-Distill-Llama-70B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 279.4 12.3 65.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ftcd3hav",
   "source_url": "https://omlx.ai/benchmarks/performance/ftcd3hav",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-abezwahv-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.8 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 507.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 507.0 37.6 175.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/abezwahv",
   "source_url": "https://omlx.ai/benchmarks/performance/abezwahv",
   "tg_tokens": null,
   "tps": 37.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-abezwahv-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 554.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 554.1 67.5 177.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/abezwahv",
   "source_url": "https://omlx.ai/benchmarks/performance/abezwahv",
   "tg_tokens": null,
   "tps": 67.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-abezwahv-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.0 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 536.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 536.6 32.6 178.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/abezwahv",
   "source_url": "https://omlx.ai/benchmarks/performance/abezwahv",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-abezwahv-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 526.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 526.3 23.7 178.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/abezwahv",
   "source_url": "https://omlx.ai/benchmarks/performance/abezwahv",
   "tg_tokens": null,
   "tps": 23.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jgemmt5e-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.2 GB at this context; duplicate traces of the same session: edndddd7, tc2fy0in, zkt9shcg; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 507.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 507.0 49.8 174.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jgemmt5e",
   "source_url": "https://omlx.ai/benchmarks/performance/jgemmt5e",
   "tg_tokens": null,
   "tps": 49.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-jgemmt5e-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.8 GB at this context; duplicate traces of the same session: edndddd7, tc2fy0in, zkt9shcg; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 552.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 552.9 60.8 175.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jgemmt5e",
   "source_url": "https://omlx.ai/benchmarks/performance/jgemmt5e",
   "tg_tokens": null,
   "tps": 60.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-jgemmt5e-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 176.4 GB at this context; duplicate traces of the same session: edndddd7, tc2fy0in, zkt9shcg; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 536.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 536.8 35.2 176.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jgemmt5e",
   "source_url": "https://omlx.ai/benchmarks/performance/jgemmt5e",
   "tg_tokens": null,
   "tps": 35.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-jgemmt5e-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.2 GB at this context; duplicate traces of the same session: edndddd7, tc2fy0in, zkt9shcg; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 527.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 527.0 26.2 177.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jgemmt5e",
   "source_url": "https://omlx.ai/benchmarks/performance/jgemmt5e",
   "tg_tokens": null,
   "tps": 26.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-sezrb3ss-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.2 GB at this context; duplicate traces of the same session: 32s6czll, 35gw8vjn, f1zczwqn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 507.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 507.4 41.3 174.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sezrb3ss",
   "source_url": "https://omlx.ai/benchmarks/performance/sezrb3ss",
   "tg_tokens": null,
   "tps": 41.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-sezrb3ss-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.8 GB at this context; duplicate traces of the same session: 32s6czll, 35gw8vjn, f1zczwqn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 552.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 552.8 61.6 175.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sezrb3ss",
   "source_url": "https://omlx.ai/benchmarks/performance/sezrb3ss",
   "tg_tokens": null,
   "tps": 61.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-sezrb3ss-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 176.4 GB at this context; duplicate traces of the same session: 32s6czll, 35gw8vjn, f1zczwqn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 537.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 537.2 28.4 176.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sezrb3ss",
   "source_url": "https://omlx.ai/benchmarks/performance/sezrb3ss",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-sezrb3ss-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.2 GB at this context; duplicate traces of the same session: 32s6czll, 35gw8vjn, f1zczwqn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 527.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 527.1 20.2 177.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sezrb3ss",
   "source_url": "https://omlx.ai/benchmarks/performance/sezrb3ss",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-94ixbfec-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 10.4 GB at this context; duplicate traces of the same session: arb32ite, y95x79lo; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 708.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B-MLX-8bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 8bit; 4k 708.5 14.3 10.4 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/94ixbfec",
   "source_url": "https://omlx.ai/benchmarks/performance/94ixbfec",
   "tg_tokens": null,
   "tps": 14.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-94ixbfec-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 10.4 GB at this context; duplicate traces of the same session: arb32ite, y95x79lo; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 707.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B-MLX-8bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 8bit; 8k 707.6 14.1 10.4 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/94ixbfec",
   "source_url": "https://omlx.ai/benchmarks/performance/94ixbfec",
   "tg_tokens": null,
   "tps": 14.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-94ixbfec-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 10.7 GB at this context; duplicate traces of the same session: arb32ite, y95x79lo; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 664.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B-MLX-8bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 8bit; 16k 664.1 13.6 10.7 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/94ixbfec",
   "source_url": "https://omlx.ai/benchmarks/performance/94ixbfec",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-e8jtjvo8-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 19.4 GB; TTFT 17562 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 233.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 233.2 PP tok/s 21.9 TG tok/s 17562 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e8jtjvo8",
   "source_url": "https://omlx.ai/benchmarks/performance/e8jtjvo8",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-f9m9x7ty-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 19.4 GB; TTFT 19245 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 212.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 212.8 PP tok/s 19.3 TG tok/s 19245 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f9m9x7ty",
   "source_url": "https://omlx.ai/benchmarks/performance/f9m9x7ty",
   "tg_tokens": null,
   "tps": 19.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Pro",
   "id": "omlx-2kxnpjy2-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 16.0 GB; TTFT 11268 ms; duplicate traces of the same session: none; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 90.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 1k tokens 90.9 PP tok/s 8.2 TG tok/s 11268 TTFT (ms) 16.0 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2kxnpjy2",
   "source_url": "https://omlx.ai/benchmarks/performance/2kxnpjy2",
   "tg_tokens": null,
   "tps": 8.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-3oa449oh-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: 2sxe9ipi; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 209.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k 209.8 8.5 15.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3oa449oh",
   "source_url": "https://omlx.ai/benchmarks/performance/3oa449oh",
   "tg_tokens": null,
   "tps": 8.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-3oa449oh-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 2sxe9ipi; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 209.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 209.9 8.4 17.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3oa449oh",
   "source_url": "https://omlx.ai/benchmarks/performance/3oa449oh",
   "tg_tokens": null,
   "tps": 8.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ubxcd7j3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 647mbzq8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 453.7,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.8-27B-MLX-5bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 5bit; 1k 453.7 14.9 19.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ubxcd7j3",
   "source_url": "https://omlx.ai/benchmarks/performance/ubxcd7j3",
   "tg_tokens": null,
   "tps": 14.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ubxcd7j3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 647mbzq8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 485.1,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.8-27B-MLX-5bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 5bit; 4k 485.1 14.7 20.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ubxcd7j3",
   "source_url": "https://omlx.ai/benchmarks/performance/ubxcd7j3",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lmbh3uua-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 24.2 GB; TTFT 17887 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 916.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Uncensored-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k tokens 916.0 PP tok/s 23.4 TG tok/s 17887 TTFT (ms) 24.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lmbh3uua",
   "source_url": "https://omlx.ai/benchmarks/performance/lmbh3uua",
   "tg_tokens": null,
   "tps": 23.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hpl4os11-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 23.4 GB; TTFT 4345 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 942.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Uncensored-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k tokens 942.7 PP tok/s 24.3 TG tok/s 4345 TTFT (ms) 23.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hpl4os11",
   "source_url": "https://omlx.ai/benchmarks/performance/hpl4os11",
   "tg_tokens": null,
   "tps": 24.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ezs947jq-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 23.4 GB; TTFT 4445 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 921.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Uncensored-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k tokens 921.4 PP tok/s 24.2 TG tok/s 4445 TTFT (ms) 23.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ezs947jq",
   "source_url": "https://omlx.ai/benchmarks/performance/ezs947jq",
   "tg_tokens": null,
   "tps": 24.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-w6hc9k74-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 20.0 GB; TTFT 22988 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 178.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 178.2 PP tok/s 12.7 TG tok/s 22988 TTFT (ms) 20.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w6hc9k74",
   "source_url": "https://omlx.ai/benchmarks/performance/w6hc9k74",
   "tg_tokens": null,
   "tps": 12.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-4ubkdnbh-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 6qcyndx8, gb3f8mx9, w44iecv2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 425.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 425.3 61.4 18.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ubkdnbh",
   "source_url": "https://omlx.ai/benchmarks/performance/4ubkdnbh",
   "tg_tokens": null,
   "tps": 61.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-4ubkdnbh-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 6qcyndx8, gb3f8mx9, w44iecv2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 454.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 454.0 87.8 21.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ubkdnbh",
   "source_url": "https://omlx.ai/benchmarks/performance/4ubkdnbh",
   "tg_tokens": null,
   "tps": 87.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-4ubkdnbh-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: 6qcyndx8, gb3f8mx9, w44iecv2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 447.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 447.1 50.2 22.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ubkdnbh",
   "source_url": "https://omlx.ai/benchmarks/performance/4ubkdnbh",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-4ubkdnbh-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: 6qcyndx8, gb3f8mx9, w44iecv2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 430.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 430.8 50.8 23.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ubkdnbh",
   "source_url": "https://omlx.ai/benchmarks/performance/4ubkdnbh",
   "tg_tokens": null,
   "tps": 50.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6bhy08k6-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 70.2 GB; TTFT 1514 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2705.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 2,705 PP tok/s 62.7 TG tok/s 1514 TTFT (ms) 70.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6bhy08k6",
   "source_url": "https://omlx.ai/benchmarks/performance/6bhy08k6",
   "tg_tokens": null,
   "tps": 62.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tgidigny-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 70.2 GB; TTFT 1531 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2675.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 2,675 PP tok/s 59.7 TG tok/s 1531 TTFT (ms) 70.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tgidigny",
   "source_url": "https://omlx.ai/benchmarks/performance/tgidigny",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bzi6zh4y-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 70.2 GB; TTFT 1511 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2710.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 2,710 PP tok/s 60.3 TG tok/s 1511 TTFT (ms) 70.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bzi6zh4y",
   "source_url": "https://omlx.ai/benchmarks/performance/bzi6zh4y",
   "tg_tokens": null,
   "tps": 60.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-03afgtdj-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 69.6 GB; TTFT 536 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1909.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,909 PP tok/s 65.6 TG tok/s 536 TTFT (ms) 69.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/03afgtdj",
   "source_url": "https://omlx.ai/benchmarks/performance/03afgtdj",
   "tg_tokens": null,
   "tps": 65.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2bxp18qx-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 69.6 GB at this context; duplicate traces of the same session: 4x24vlkf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1914.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,914 64.5 69.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2bxp18qx",
   "source_url": "https://omlx.ai/benchmarks/performance/2bxp18qx",
   "tg_tokens": null,
   "tps": 64.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2bxp18qx-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 70.2 GB at this context; duplicate traces of the same session: 4x24vlkf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2703.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,703 62.9 70.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2bxp18qx",
   "source_url": "https://omlx.ai/benchmarks/performance/2bxp18qx",
   "tg_tokens": null,
   "tps": 62.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-m1keola5-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 69.6 GB at this context; duplicate traces of the same session: qzjglxcm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1916.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,916 64.2 69.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m1keola5",
   "source_url": "https://omlx.ai/benchmarks/performance/m1keola5",
   "tg_tokens": null,
   "tps": 64.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-m1keola5-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 70.2 GB at this context; duplicate traces of the same session: qzjglxcm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2697.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,697 61.6 70.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m1keola5",
   "source_url": "https://omlx.ai/benchmarks/performance/m1keola5",
   "tg_tokens": null,
   "tps": 61.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pr7wv0ke-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 69.6 GB at this context; duplicate traces of the same session: ze4e12iw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1918.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,918 65.8 69.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pr7wv0ke",
   "source_url": "https://omlx.ai/benchmarks/performance/pr7wv0ke",
   "tg_tokens": null,
   "tps": 65.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pr7wv0ke-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 70.2 GB at this context; duplicate traces of the same session: ze4e12iw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2692.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,692 61.2 70.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pr7wv0ke",
   "source_url": "https://omlx.ai/benchmarks/performance/pr7wv0ke",
   "tg_tokens": null,
   "tps": 61.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0cpe5pvw-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 69.6 GB at this context; duplicate traces of the same session: 8s910mtg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1917.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,917 66.3 69.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0cpe5pvw",
   "source_url": "https://omlx.ai/benchmarks/performance/0cpe5pvw",
   "tg_tokens": null,
   "tps": 66.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0cpe5pvw-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 70.2 GB at this context; duplicate traces of the same session: 8s910mtg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2709.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,709 63.2 70.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0cpe5pvw",
   "source_url": "https://omlx.ai/benchmarks/performance/0cpe5pvw",
   "tg_tokens": null,
   "tps": 63.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xbc138h2-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 69.6 GB; TTFT 537 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1905.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,905 PP tok/s 65.4 TG tok/s 537 TTFT (ms) 69.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xbc138h2",
   "source_url": "https://omlx.ai/benchmarks/performance/xbc138h2",
   "tg_tokens": null,
   "tps": 65.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gy0h0joc-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 69.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1903.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,903 65.2 69.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gy0h0joc",
   "source_url": "https://omlx.ai/benchmarks/performance/gy0h0joc",
   "tg_tokens": null,
   "tps": 65.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gy0h0joc-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 70.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2715.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,715 61.7 70.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gy0h0joc",
   "source_url": "https://omlx.ai/benchmarks/performance/gy0h0joc",
   "tg_tokens": null,
   "tps": 61.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-119gg1x0-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: 6qief3sb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1109.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,109 88.8 100.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/119gg1x0",
   "source_url": "https://omlx.ai/benchmarks/performance/119gg1x0",
   "tg_tokens": null,
   "tps": 88.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-119gg1x0-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: 6qief3sb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1385.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,385 99.3 101.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/119gg1x0",
   "source_url": "https://omlx.ai/benchmarks/performance/119gg1x0",
   "tg_tokens": null,
   "tps": 99.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-119gg1x0-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: 6qief3sb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1378.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 1,378 67.5 101.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/119gg1x0",
   "source_url": "https://omlx.ai/benchmarks/performance/119gg1x0",
   "tg_tokens": null,
   "tps": 67.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-119gg1x0-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.6 GB at this context; duplicate traces of the same session: 6qief3sb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1371.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 1,371 92.1 101.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/119gg1x0",
   "source_url": "https://omlx.ai/benchmarks/performance/119gg1x0",
   "tg_tokens": null,
   "tps": 92.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2ler8jjd-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.6 GB at this context; duplicate traces of the same session: c66p7at0, ea6zcytg, kdzbaqkt, zombyfxm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1152.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,152 88.6 73.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ler8jjd",
   "source_url": "https://omlx.ai/benchmarks/performance/2ler8jjd",
   "tg_tokens": null,
   "tps": 88.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2ler8jjd-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.7 GB at this context; duplicate traces of the same session: c66p7at0, ea6zcytg, kdzbaqkt, zombyfxm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1899.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,899 77.6 73.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ler8jjd",
   "source_url": "https://omlx.ai/benchmarks/performance/2ler8jjd",
   "tg_tokens": null,
   "tps": 77.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2ler8jjd-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 75.7 GB at this context; duplicate traces of the same session: c66p7at0, ea6zcytg, kdzbaqkt, zombyfxm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2167.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,167 77.7 75.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ler8jjd",
   "source_url": "https://omlx.ai/benchmarks/performance/2ler8jjd",
   "tg_tokens": null,
   "tps": 77.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2ler8jjd-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 76.3 GB at this context; duplicate traces of the same session: c66p7at0, ea6zcytg, kdzbaqkt, zombyfxm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2019.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,019 66.1 76.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ler8jjd",
   "source_url": "https://omlx.ai/benchmarks/performance/2ler8jjd",
   "tg_tokens": null,
   "tps": 66.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2ler8jjd-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 77.5 GB at this context; duplicate traces of the same session: c66p7at0, ea6zcytg, kdzbaqkt, zombyfxm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1789.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 1,789 81.3 77.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ler8jjd",
   "source_url": "https://omlx.ai/benchmarks/performance/2ler8jjd",
   "tg_tokens": null,
   "tps": 81.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5grhfjkp-ctx1024",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 3al17p1s; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 612.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31B-it-MLX-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 612.4 21.9 17.6 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5grhfjkp",
   "source_url": "https://omlx.ai/benchmarks/performance/5grhfjkp",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5grhfjkp-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 3al17p1s; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 589.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31B-it-MLX-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 589.5 21.3 18.8 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5grhfjkp",
   "source_url": "https://omlx.ai/benchmarks/performance/5grhfjkp",
   "tg_tokens": null,
   "tps": 21.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-p2gtwo5m-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 5.8 GB; TTFT 1560 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2626.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 4k tokens 2,626 PP tok/s 72.4 TG tok/s 1560 TTFT (ms) 5.8 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p2gtwo5m",
   "source_url": "https://omlx.ai/benchmarks/performance/p2gtwo5m",
   "tg_tokens": null,
   "tps": 72.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-3ywz01rm-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.7 GB at this context; duplicate traces of the same session: gqdd2gu3; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1436.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 1k 1,436 45.9 4.7 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ywz01rm",
   "source_url": "https://omlx.ai/benchmarks/performance/3ywz01rm",
   "tg_tokens": null,
   "tps": 45.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-3ywz01rm-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.7 GB at this context; duplicate traces of the same session: gqdd2gu3; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2327.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 4k 2,327 44.7 4.7 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ywz01rm",
   "source_url": "https://omlx.ai/benchmarks/performance/3ywz01rm",
   "tg_tokens": null,
   "tps": 44.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-xqxlss1w-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 20.5 GB; TTFT 2352 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1742.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k tokens 1,742 PP tok/s 62.9 TG tok/s 2352 TTFT (ms) 20.5 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xqxlss1w",
   "source_url": "https://omlx.ai/benchmarks/performance/xqxlss1w",
   "tg_tokens": null,
   "tps": 62.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-b6wiorfo-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 20.5 GB; TTFT 2350 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1743.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k tokens 1,743 PP tok/s 68.7 TG tok/s 2350 TTFT (ms) 20.5 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b6wiorfo",
   "source_url": "https://omlx.ai/benchmarks/performance/b6wiorfo",
   "tg_tokens": null,
   "tps": 68.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-rx4nqyww-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 20.5 GB; TTFT 2374 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1726.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k tokens 1,726 PP tok/s 64.3 TG tok/s 2374 TTFT (ms) 20.5 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rx4nqyww",
   "source_url": "https://omlx.ai/benchmarks/performance/rx4nqyww",
   "tg_tokens": null,
   "tps": 64.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-z70tiewn-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 19.9 GB; TTFT 3956 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1036.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k tokens 1,036 PP tok/s 51.6 TG tok/s 3956 TTFT (ms) 19.9 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z70tiewn",
   "source_url": "https://omlx.ai/benchmarks/performance/z70tiewn",
   "tg_tokens": null,
   "tps": 51.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-y3sqfcl3-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "spicyneuron MLX build; oMLX community benchmark trace, single run at 8k context; peak memory 25.9 GB; TTFT 10599 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 772.9,
   "provenance": "community",
   "quant": "5.4bit",
   "quote": "spicyneuron--Qwen3.6-35B-A3B-MLX-5.4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 8k tokens 772.9 PP tok/s 61.8 TG tok/s 10599 TTFT (ms) 25.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y3sqfcl3",
   "source_url": "https://omlx.ai/benchmarks/performance/y3sqfcl3",
   "tg_tokens": null,
   "tps": 61.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ae57szhv-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2580.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 2,580 111.1 22.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ae57szhv",
   "source_url": "https://omlx.ai/benchmarks/performance/ae57szhv",
   "tg_tokens": null,
   "tps": 111.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ae57szhv-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3408.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 3,408 117.6 24.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ae57szhv",
   "source_url": "https://omlx.ai/benchmarks/performance/ae57szhv",
   "tg_tokens": null,
   "tps": 117.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ae57szhv-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3233.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 3,233 120.3 24.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ae57szhv",
   "source_url": "https://omlx.ai/benchmarks/performance/ae57szhv",
   "tg_tokens": null,
   "tps": 120.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ae57szhv-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2877.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 2,877 127.6 24.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ae57szhv",
   "source_url": "https://omlx.ai/benchmarks/performance/ae57szhv",
   "tg_tokens": null,
   "tps": 127.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Pro",
   "id": "omlx-nfaf1327-ctx1024",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 16.0 GB; TTFT 11390 ms; duplicate traces of the same session: none; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 89.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-27B-MLX-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 1k tokens 89.9 PP tok/s 6.5 TG tok/s 11390 TTFT (ms) 16.0 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nfaf1327",
   "source_url": "https://omlx.ai/benchmarks/performance/nfaf1327",
   "tg_tokens": null,
   "tps": 6.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-syrljh75-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 2lfjbosy, 4y7dipko, 6s1aqond, br7qvdlr, h5rqwrjo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3872.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 3,872 148.3 20.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/syrljh75",
   "source_url": "https://omlx.ai/benchmarks/performance/syrljh75",
   "tg_tokens": null,
   "tps": 148.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-syrljh75-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: 2lfjbosy, 4y7dipko, 6s1aqond, br7qvdlr, h5rqwrjo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6535.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,535 141.9 21.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/syrljh75",
   "source_url": "https://omlx.ai/benchmarks/performance/syrljh75",
   "tg_tokens": null,
   "tps": 141.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-syrljh75-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: 2lfjbosy, 4y7dipko, 6s1aqond, br7qvdlr, h5rqwrjo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6897.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 6,897 137.0 21.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/syrljh75",
   "source_url": "https://omlx.ai/benchmarks/performance/syrljh75",
   "tg_tokens": null,
   "tps": 137.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-syrljh75-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 2lfjbosy, 4y7dipko, 6s1aqond, br7qvdlr, h5rqwrjo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6225.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 6,225 133.7 21.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/syrljh75",
   "source_url": "https://omlx.ai/benchmarks/performance/syrljh75",
   "tg_tokens": null,
   "tps": 133.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-syrljh75-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 2lfjbosy, 4y7dipko, 6s1aqond, br7qvdlr, h5rqwrjo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4768.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 4,768 122.5 21.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/syrljh75",
   "source_url": "https://omlx.ai/benchmarks/performance/syrljh75",
   "tg_tokens": null,
   "tps": 122.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-syrljh75-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 2lfjbosy, 4y7dipko, 6s1aqond, br7qvdlr, h5rqwrjo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3226.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 3,226 91.8 22.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/syrljh75",
   "source_url": "https://omlx.ai/benchmarks/performance/syrljh75",
   "tg_tokens": null,
   "tps": 91.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0wuqfoue-ctx16384",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 27.0 GB; TTFT 3354 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4885.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-Safetensors-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 5bit; 16k tokens 4,885 PP tok/s 213.9 TG tok/s 3354 TTFT (ms) 27.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0wuqfoue",
   "source_url": "https://omlx.ai/benchmarks/performance/0wuqfoue",
   "tg_tokens": null,
   "tps": 213.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-is1wyzyb-ctx199680",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "MLX build; oMLX community benchmark trace, single run at 195k context; peak memory 38.7 GB; TTFT 135802 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1473.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 195k tokens 1,473 PP tok/s 24.5 TG tok/s 135802 TTFT (ms) 38.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/is1wyzyb",
   "source_url": "https://omlx.ai/benchmarks/performance/is1wyzyb",
   "tg_tokens": null,
   "tps": 24.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hc8n88mq-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 20.4 GB; TTFT 36926 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 887.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k tokens 887.4 PP tok/s 29.3 TG tok/s 36926 TTFT (ms) 20.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hc8n88mq",
   "source_url": "https://omlx.ai/benchmarks/performance/hc8n88mq",
   "tg_tokens": null,
   "tps": 29.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zwtv6uch-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: 64nrwvtc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 711.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 711.2 39.2 33.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zwtv6uch",
   "source_url": "https://omlx.ai/benchmarks/performance/zwtv6uch",
   "tg_tokens": null,
   "tps": 39.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zwtv6uch-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: 64nrwvtc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 873.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 873.9 38.1 34.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zwtv6uch",
   "source_url": "https://omlx.ai/benchmarks/performance/zwtv6uch",
   "tg_tokens": null,
   "tps": 38.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kgdyru7k-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 69.6 GB at this context; duplicate traces of the same session: l728f013; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1892.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,892 64.7 69.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kgdyru7k",
   "source_url": "https://omlx.ai/benchmarks/performance/kgdyru7k",
   "tg_tokens": null,
   "tps": 64.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kgdyru7k-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 70.2 GB at this context; duplicate traces of the same session: l728f013; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2709.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,709 61.4 70.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kgdyru7k",
   "source_url": "https://omlx.ai/benchmarks/performance/kgdyru7k",
   "tg_tokens": null,
   "tps": 61.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mhx6udwp-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 69.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1903.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,903 65.2 69.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mhx6udwp",
   "source_url": "https://omlx.ai/benchmarks/performance/mhx6udwp",
   "tg_tokens": null,
   "tps": 65.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mhx6udwp-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 70.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2715.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,715 61.7 70.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mhx6udwp",
   "source_url": "https://omlx.ai/benchmarks/performance/mhx6udwp",
   "tg_tokens": null,
   "tps": 61.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-m42mw9id-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 69.6 GB; TTFT 540 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1897.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,897 PP tok/s 65.6 TG tok/s 540 TTFT (ms) 69.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m42mw9id",
   "source_url": "https://omlx.ai/benchmarks/performance/m42mw9id",
   "tg_tokens": null,
   "tps": 65.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-r0h1ve91-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 69.6 GB at this context; duplicate traces of the same session: x2x25yf9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1910.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,910 64.8 69.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r0h1ve91",
   "source_url": "https://omlx.ai/benchmarks/performance/r0h1ve91",
   "tg_tokens": null,
   "tps": 64.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-r0h1ve91-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 70.2 GB at this context; duplicate traces of the same session: x2x25yf9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2688.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,688 60.1 70.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r0h1ve91",
   "source_url": "https://omlx.ai/benchmarks/performance/r0h1ve91",
   "tg_tokens": null,
   "tps": 60.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8cqmvht4-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 69.6 GB; TTFT 538 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1904.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,904 PP tok/s 66.0 TG tok/s 538 TTFT (ms) 69.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8cqmvht4",
   "source_url": "https://omlx.ai/benchmarks/performance/8cqmvht4",
   "tg_tokens": null,
   "tps": 66.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qoytbmz0-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 69.6 GB; TTFT 539 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1901.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,901 PP tok/s 65.6 TG tok/s 539 TTFT (ms) 69.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qoytbmz0",
   "source_url": "https://omlx.ai/benchmarks/performance/qoytbmz0",
   "tg_tokens": null,
   "tps": 65.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-h50ufs82-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 69.6 GB; TTFT 549 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1865.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,865 PP tok/s 65.2 TG tok/s 549 TTFT (ms) 69.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h50ufs82",
   "source_url": "https://omlx.ai/benchmarks/performance/h50ufs82",
   "tg_tokens": null,
   "tps": 65.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-43pbgzgy-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 69.6 GB at this context; duplicate traces of the same session: vgohld47; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2055.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,055 61.0 69.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/43pbgzgy",
   "source_url": "https://omlx.ai/benchmarks/performance/43pbgzgy",
   "tg_tokens": null,
   "tps": 61.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-43pbgzgy-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 70.2 GB at this context; duplicate traces of the same session: vgohld47; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2890.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,890 65.7 70.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/43pbgzgy",
   "source_url": "https://omlx.ai/benchmarks/performance/43pbgzgy",
   "tg_tokens": null,
   "tps": 65.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-b8324mcz-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 69.6 GB at this context; duplicate traces of the same session: ymjip94g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2048.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,048 70.4 69.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b8324mcz",
   "source_url": "https://omlx.ai/benchmarks/performance/b8324mcz",
   "tg_tokens": null,
   "tps": 70.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-b8324mcz-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 70.2 GB at this context; duplicate traces of the same session: ymjip94g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2904.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,904 62.2 70.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b8324mcz",
   "source_url": "https://omlx.ai/benchmarks/performance/b8324mcz",
   "tg_tokens": null,
   "tps": 62.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-d262btrq-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 69.6 GB at this context; duplicate traces of the same session: l6tneef2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2100.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,100 73.4 69.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d262btrq",
   "source_url": "https://omlx.ai/benchmarks/performance/d262btrq",
   "tg_tokens": null,
   "tps": 73.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-d262btrq-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 70.2 GB at this context; duplicate traces of the same session: l6tneef2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2928.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,928 62.4 70.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d262btrq",
   "source_url": "https://omlx.ai/benchmarks/performance/d262btrq",
   "tg_tokens": null,
   "tps": 62.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zjwgvab2-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 69.6 GB; TTFT 507 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2019.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 2,019 PP tok/s 71.2 TG tok/s 507 TTFT (ms) 69.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zjwgvab2",
   "source_url": "https://omlx.ai/benchmarks/performance/zjwgvab2",
   "tg_tokens": null,
   "tps": 71.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6vqqdxd1-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 69.6 GB; TTFT 524 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1954.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,954 PP tok/s 66.5 TG tok/s 524 TTFT (ms) 69.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6vqqdxd1",
   "source_url": "https://omlx.ai/benchmarks/performance/6vqqdxd1",
   "tg_tokens": null,
   "tps": 66.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qh9tp22q-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 70.2 GB; TTFT 1410 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2906.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 2,906 PP tok/s 65.4 TG tok/s 1410 TTFT (ms) 70.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qh9tp22q",
   "source_url": "https://omlx.ai/benchmarks/performance/qh9tp22q",
   "tg_tokens": null,
   "tps": 65.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-96tqtu53-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 69.6 GB at this context; duplicate traces of the same session: 3zo24h3n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2046.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,046 72.5 69.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/96tqtu53",
   "source_url": "https://omlx.ai/benchmarks/performance/96tqtu53",
   "tg_tokens": null,
   "tps": 72.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-96tqtu53-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 70.2 GB at this context; duplicate traces of the same session: 3zo24h3n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2882.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,882 68.6 70.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/96tqtu53",
   "source_url": "https://omlx.ai/benchmarks/performance/96tqtu53",
   "tg_tokens": null,
   "tps": 68.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-d5o9yyn5-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 69.6 GB at this context; duplicate traces of the same session: 8eyc941h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2063.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,063 71.0 69.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d5o9yyn5",
   "source_url": "https://omlx.ai/benchmarks/performance/d5o9yyn5",
   "tg_tokens": null,
   "tps": 71.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-d5o9yyn5-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 70.2 GB at this context; duplicate traces of the same session: 8eyc941h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2889.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,889 68.6 70.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d5o9yyn5",
   "source_url": "https://omlx.ai/benchmarks/performance/d5o9yyn5",
   "tg_tokens": null,
   "tps": 68.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-09qig58s-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 69.6 GB at this context; duplicate traces of the same session: yv59pzk2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2047.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,047 72.3 69.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/09qig58s",
   "source_url": "https://omlx.ai/benchmarks/performance/09qig58s",
   "tg_tokens": null,
   "tps": 72.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-09qig58s-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 70.2 GB at this context; duplicate traces of the same session: yv59pzk2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2888.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,888 68.3 70.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/09qig58s",
   "source_url": "https://omlx.ai/benchmarks/performance/09qig58s",
   "tg_tokens": null,
   "tps": 68.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fgkzxv0o-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 74.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1983.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,983 51.9 74.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fgkzxv0o",
   "source_url": "https://omlx.ai/benchmarks/performance/fgkzxv0o",
   "tg_tokens": null,
   "tps": 51.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fgkzxv0o-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 75.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2791.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,791 42.2 75.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fgkzxv0o",
   "source_url": "https://omlx.ai/benchmarks/performance/fgkzxv0o",
   "tg_tokens": null,
   "tps": 42.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-lr1tlc1y-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: rk88zr8l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 467.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 467.7 59.3 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lr1tlc1y",
   "source_url": "https://omlx.ai/benchmarks/performance/lr1tlc1y",
   "tg_tokens": null,
   "tps": 59.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-lr1tlc1y-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: rk88zr8l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 505.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 505.6 41.9 71.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lr1tlc1y",
   "source_url": "https://omlx.ai/benchmarks/performance/lr1tlc1y",
   "tg_tokens": null,
   "tps": 41.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-86nbk2cp-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: 1odb5ies; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1109.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,109 88.8 100.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/86nbk2cp",
   "source_url": "https://omlx.ai/benchmarks/performance/86nbk2cp",
   "tg_tokens": null,
   "tps": 88.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-86nbk2cp-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: 1odb5ies; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1385.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,385 99.3 101.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/86nbk2cp",
   "source_url": "https://omlx.ai/benchmarks/performance/86nbk2cp",
   "tg_tokens": null,
   "tps": 99.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-86nbk2cp-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: 1odb5ies; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1378.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 1,378 67.5 101.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/86nbk2cp",
   "source_url": "https://omlx.ai/benchmarks/performance/86nbk2cp",
   "tg_tokens": null,
   "tps": 67.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-86nbk2cp-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.6 GB at this context; duplicate traces of the same session: 1odb5ies; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1371.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 1,371 92.1 101.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/86nbk2cp",
   "source_url": "https://omlx.ai/benchmarks/performance/86nbk2cp",
   "tg_tokens": null,
   "tps": 92.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-o3eiaicr-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: 0zhoravo, p6w5pv44; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1110.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,110 98.0 100.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o3eiaicr",
   "source_url": "https://omlx.ai/benchmarks/performance/o3eiaicr",
   "tg_tokens": null,
   "tps": 98.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-o3eiaicr-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: 0zhoravo, p6w5pv44; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1386.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,386 83.4 101.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o3eiaicr",
   "source_url": "https://omlx.ai/benchmarks/performance/o3eiaicr",
   "tg_tokens": null,
   "tps": 83.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-o3eiaicr-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: 0zhoravo, p6w5pv44; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1379.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 1,379 80.7 101.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o3eiaicr",
   "source_url": "https://omlx.ai/benchmarks/performance/o3eiaicr",
   "tg_tokens": null,
   "tps": 80.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nu5bs19k-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: 1a6xmdmi, vitfzp6x; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1104.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,104 82.6 100.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nu5bs19k",
   "source_url": "https://omlx.ai/benchmarks/performance/nu5bs19k",
   "tg_tokens": null,
   "tps": 82.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nu5bs19k-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: 1a6xmdmi, vitfzp6x; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1381.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 1,381 76.6 101.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nu5bs19k",
   "source_url": "https://omlx.ai/benchmarks/performance/nu5bs19k",
   "tg_tokens": null,
   "tps": 76.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nu5bs19k-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.6 GB at this context; duplicate traces of the same session: 1a6xmdmi, vitfzp6x; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1373.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 1,373 86.4 101.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nu5bs19k",
   "source_url": "https://omlx.ai/benchmarks/performance/nu5bs19k",
   "tg_tokens": null,
   "tps": 86.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-g5clp72k-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: 63gsy1ci, qpg2gm2k; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1112.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,112 89.0 100.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g5clp72k",
   "source_url": "https://omlx.ai/benchmarks/performance/g5clp72k",
   "tg_tokens": null,
   "tps": 89.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-g5clp72k-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: 63gsy1ci, qpg2gm2k; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1390.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,390 74.7 101.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g5clp72k",
   "source_url": "https://omlx.ai/benchmarks/performance/g5clp72k",
   "tg_tokens": null,
   "tps": 74.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-g5clp72k-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: 63gsy1ci, qpg2gm2k; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1381.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 1,381 92.9 101.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g5clp72k",
   "source_url": "https://omlx.ai/benchmarks/performance/g5clp72k",
   "tg_tokens": null,
   "tps": 92.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2gkb5lnq-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 71.8 GB; TTFT 4507 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1817.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k tokens 1,817 PP tok/s 81.5 TG tok/s 4507 TTFT (ms) 71.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2gkb5lnq",
   "source_url": "https://omlx.ai/benchmarks/performance/2gkb5lnq",
   "tg_tokens": null,
   "tps": 81.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-e86jf7l6-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 77.6 GB; TTFT 4997 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1639.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k tokens 1,639 PP tok/s 84.0 TG tok/s 4997 TTFT (ms) 77.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e86jf7l6",
   "source_url": "https://omlx.ai/benchmarks/performance/e86jf7l6",
   "tg_tokens": null,
   "tps": 84.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rbkn1nox-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 77.6 GB; TTFT 4603 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1780.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k tokens 1,780 PP tok/s 75.2 TG tok/s 4603 TTFT (ms) 77.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rbkn1nox",
   "source_url": "https://omlx.ai/benchmarks/performance/rbkn1nox",
   "tg_tokens": null,
   "tps": 75.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zsm3coha-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 77.6 GB; TTFT 5574 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1470.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k tokens 1,470 PP tok/s 46.8 TG tok/s 5574 TTFT (ms) 77.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zsm3coha",
   "source_url": "https://omlx.ai/benchmarks/performance/zsm3coha",
   "tg_tokens": null,
   "tps": 46.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9ct2y5lx-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 71.1 GB; TTFT 2209 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1854.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 1,854 PP tok/s 71.0 TG tok/s 2209 TTFT (ms) 71.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ct2y5lx",
   "source_url": "https://omlx.ai/benchmarks/performance/9ct2y5lx",
   "tg_tokens": null,
   "tps": 71.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q1n1x3cs-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 77.0 GB; TTFT 2205 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1858.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 1,858 PP tok/s 60.8 TG tok/s 2205 TTFT (ms) 77.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q1n1x3cs",
   "source_url": "https://omlx.ai/benchmarks/performance/q1n1x3cs",
   "tg_tokens": null,
   "tps": 60.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-z9n76072-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 70.9 GB; TTFT 2129 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1924.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 1,924 PP tok/s 58.0 TG tok/s 2129 TTFT (ms) 70.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z9n76072",
   "source_url": "https://omlx.ai/benchmarks/performance/z9n76072",
   "tg_tokens": null,
   "tps": 58.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bcvc87br-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 71.1 GB; TTFT 2282 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1795.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 1,795 PP tok/s 62.3 TG tok/s 2282 TTFT (ms) 71.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bcvc87br",
   "source_url": "https://omlx.ai/benchmarks/performance/bcvc87br",
   "tg_tokens": null,
   "tps": 62.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8xwikxht-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 70.9 GB; TTFT 3687 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1111.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 1,111 PP tok/s 36.8 TG tok/s 3687 TTFT (ms) 70.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8xwikxht",
   "source_url": "https://omlx.ai/benchmarks/performance/8xwikxht",
   "tg_tokens": null,
   "tps": 36.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-247cizdw-ctx1024",
   "model": "DeepSeek-R1-Distill-Qwen-7B",
   "notes": "oMLX community benchmark trace; peak memory 4.8 GB at this context; duplicate traces of the same session: mme887lr; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1661.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-R1-Distill-Qwen-7B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 1,661 64.2 4.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/247cizdw",
   "source_url": "https://omlx.ai/benchmarks/performance/247cizdw",
   "tg_tokens": null,
   "tps": 64.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-247cizdw-ctx4096",
   "model": "DeepSeek-R1-Distill-Qwen-7B",
   "notes": "oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: mme887lr; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1898.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-R1-Distill-Qwen-7B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,898 61.7 4.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/247cizdw",
   "source_url": "https://omlx.ai/benchmarks/performance/247cizdw",
   "tg_tokens": null,
   "tps": 61.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-n2t3ppfl-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 107.0 GB at this context; duplicate traces of the same session: 6picjwft, k9jv9hjx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 256.5,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4.1-Flash-oQ2e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 256.5 15.7 107.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n2t3ppfl",
   "source_url": "https://omlx.ai/benchmarks/performance/n2t3ppfl",
   "tg_tokens": null,
   "tps": 15.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-n2t3ppfl-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 108.8 GB at this context; duplicate traces of the same session: 6picjwft, k9jv9hjx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 438.0,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4.1-Flash-oQ2e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 8k 438.0 18.9 108.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n2t3ppfl",
   "source_url": "https://omlx.ai/benchmarks/performance/n2t3ppfl",
   "tg_tokens": null,
   "tps": 18.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-n2t3ppfl-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 108.9 GB at this context; duplicate traces of the same session: 6picjwft, k9jv9hjx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 452.6,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4.1-Flash-oQ2e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 16k 452.6 21.1 108.9 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n2t3ppfl",
   "source_url": "https://omlx.ai/benchmarks/performance/n2t3ppfl",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-s6gn3bck-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 106.3 GB at this context; duplicate traces of the same session: dyzgjm8j, lusr71pi, u7wx2gnm; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 92.7,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4.1-Flash-oQ2e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 92.7 16.3 106.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s6gn3bck",
   "source_url": "https://omlx.ai/benchmarks/performance/s6gn3bck",
   "tg_tokens": null,
   "tps": 16.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-s6gn3bck-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 107.0 GB at this context; duplicate traces of the same session: dyzgjm8j, lusr71pi, u7wx2gnm; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 324.9,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4.1-Flash-oQ2e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 324.9 17.6 107.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s6gn3bck",
   "source_url": "https://omlx.ai/benchmarks/performance/s6gn3bck",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-s6gn3bck-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 108.8 GB at this context; duplicate traces of the same session: dyzgjm8j, lusr71pi, u7wx2gnm; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 440.2,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4.1-Flash-oQ2e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 8k 440.2 18.5 108.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s6gn3bck",
   "source_url": "https://omlx.ai/benchmarks/performance/s6gn3bck",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-s6gn3bck-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 108.9 GB at this context; duplicate traces of the same session: dyzgjm8j, lusr71pi, u7wx2gnm; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 454.3,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4.1-Flash-oQ2e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 16k 454.3 20.4 108.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s6gn3bck",
   "source_url": "https://omlx.ai/benchmarks/performance/s6gn3bck",
   "tg_tokens": null,
   "tps": 20.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-y306bb5v-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 3gxz3h3o, 86kvtjdg, dw3zjdr1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2757.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 2,757 109.2 19.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y306bb5v",
   "source_url": "https://omlx.ai/benchmarks/performance/y306bb5v",
   "tg_tokens": null,
   "tps": 109.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-y306bb5v-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 3gxz3h3o, 86kvtjdg, dw3zjdr1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3460.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 3,460 107.2 21.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y306bb5v",
   "source_url": "https://omlx.ai/benchmarks/performance/y306bb5v",
   "tg_tokens": null,
   "tps": 107.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-y306bb5v-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 3gxz3h3o, 86kvtjdg, dw3zjdr1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3312.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 3,312 100.7 21.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y306bb5v",
   "source_url": "https://omlx.ai/benchmarks/performance/y306bb5v",
   "tg_tokens": null,
   "tps": 100.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-y306bb5v-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 3gxz3h3o, 86kvtjdg, dw3zjdr1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2985.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 2,985 95.1 21.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y306bb5v",
   "source_url": "https://omlx.ai/benchmarks/performance/y306bb5v",
   "tg_tokens": null,
   "tps": 95.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ik5967cy-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: skzxhjnw, t5mbq2re; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2580.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 2,580 111.1 22.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ik5967cy",
   "source_url": "https://omlx.ai/benchmarks/performance/ik5967cy",
   "tg_tokens": null,
   "tps": 111.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ik5967cy-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: skzxhjnw, t5mbq2re; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3408.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 3,408 117.6 24.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ik5967cy",
   "source_url": "https://omlx.ai/benchmarks/performance/ik5967cy",
   "tg_tokens": null,
   "tps": 117.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ik5967cy-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: skzxhjnw, t5mbq2re; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3233.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 3,233 120.3 24.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ik5967cy",
   "source_url": "https://omlx.ai/benchmarks/performance/ik5967cy",
   "tg_tokens": null,
   "tps": 120.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ik5967cy-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: skzxhjnw, t5mbq2re; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2877.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 2,877 127.6 24.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ik5967cy",
   "source_url": "https://omlx.ai/benchmarks/performance/ik5967cy",
   "tg_tokens": null,
   "tps": 127.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vp1co0gc-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: d5ysglai; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2594.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 2,594 119.8 22.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vp1co0gc",
   "source_url": "https://omlx.ai/benchmarks/performance/vp1co0gc",
   "tg_tokens": null,
   "tps": 119.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vp1co0gc-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: d5ysglai; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3401.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 3,401 116.6 24.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vp1co0gc",
   "source_url": "https://omlx.ai/benchmarks/performance/vp1co0gc",
   "tg_tokens": null,
   "tps": 116.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vp1co0gc-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: d5ysglai; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3228.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 3,228 128.4 24.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vp1co0gc",
   "source_url": "https://omlx.ai/benchmarks/performance/vp1co0gc",
   "tg_tokens": null,
   "tps": 128.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vp1co0gc-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: d5ysglai; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2879.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 2,879 127.8 24.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vp1co0gc",
   "source_url": "https://omlx.ai/benchmarks/performance/vp1co0gc",
   "tg_tokens": null,
   "tps": 127.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ndjcfg37-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: el1bq64o, fz2s8z9s, u07v65k4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2375.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.6-35B-A3B-mxfp4 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 2,375 110.7 18.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ndjcfg37",
   "source_url": "https://omlx.ai/benchmarks/performance/ndjcfg37",
   "tg_tokens": null,
   "tps": 110.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ndjcfg37-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: el1bq64o, fz2s8z9s, u07v65k4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3403.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.6-35B-A3B-mxfp4 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 3,403 108.9 20.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ndjcfg37",
   "source_url": "https://omlx.ai/benchmarks/performance/ndjcfg37",
   "tg_tokens": null,
   "tps": 108.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ndjcfg37-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: el1bq64o, fz2s8z9s, u07v65k4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3235.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.6-35B-A3B-mxfp4 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 3,235 101.4 20.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ndjcfg37",
   "source_url": "https://omlx.ai/benchmarks/performance/ndjcfg37",
   "tg_tokens": null,
   "tps": 101.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ndjcfg37-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: el1bq64o, fz2s8z9s, u07v65k4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2946.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.6-35B-A3B-mxfp4 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 2,946 97.6 20.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ndjcfg37",
   "source_url": "https://omlx.ai/benchmarks/performance/ndjcfg37",
   "tg_tokens": null,
   "tps": 97.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-twqmti57-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: tuz7dhpx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4120.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 4,120 192.2 37.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/twqmti57",
   "source_url": "https://omlx.ai/benchmarks/performance/twqmti57",
   "tg_tokens": null,
   "tps": 192.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-twqmti57-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 38.0 GB at this context; duplicate traces of the same session: tuz7dhpx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5502.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 5,502 154.7 38.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/twqmti57",
   "source_url": "https://omlx.ai/benchmarks/performance/twqmti57",
   "tg_tokens": null,
   "tps": 154.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ho0geb4h-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: vlgrp5ah; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 967.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 967.7 56.2 17.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ho0geb4h",
   "source_url": "https://omlx.ai/benchmarks/performance/ho0geb4h",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ho0geb4h-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: vlgrp5ah; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1014.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,014 58.2 18.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ho0geb4h",
   "source_url": "https://omlx.ai/benchmarks/performance/ho0geb4h",
   "tg_tokens": null,
   "tps": 58.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qqg9kvhg-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: qpm4kg82; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 969.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 969.3 46.4 17.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qqg9kvhg",
   "source_url": "https://omlx.ai/benchmarks/performance/qqg9kvhg",
   "tg_tokens": null,
   "tps": 46.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qqg9kvhg-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: qpm4kg82; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1012.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,012 55.8 18.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qqg9kvhg",
   "source_url": "https://omlx.ai/benchmarks/performance/qqg9kvhg",
   "tg_tokens": null,
   "tps": 55.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4d7hrwlz-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: gbt7vrk9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 970.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 970.2 43.9 17.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4d7hrwlz",
   "source_url": "https://omlx.ai/benchmarks/performance/4d7hrwlz",
   "tg_tokens": null,
   "tps": 43.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4d7hrwlz-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: gbt7vrk9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1015.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,015 64.0 18.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4d7hrwlz",
   "source_url": "https://omlx.ai/benchmarks/performance/4d7hrwlz",
   "tg_tokens": null,
   "tps": 64.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-h4yft8wi-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: dxi0ssnu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 902.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 902.7 35.8 23.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h4yft8wi",
   "source_url": "https://omlx.ai/benchmarks/performance/h4yft8wi",
   "tg_tokens": null,
   "tps": 35.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-h4yft8wi-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: dxi0ssnu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 933.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 933.7 38.8 24.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h4yft8wi",
   "source_url": "https://omlx.ai/benchmarks/performance/h4yft8wi",
   "tg_tokens": null,
   "tps": 38.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-teh7edht-ctx1024",
   "model": "Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored",
   "notes": "MTP MLX build; oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: 3qfs3vvf, e3l0t7ag, vupevv7h; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 427.3,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored-oQ4-mtp-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 427.3 49.9 18.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/teh7edht",
   "source_url": "https://omlx.ai/benchmarks/performance/teh7edht",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-teh7edht-ctx4096",
   "model": "Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored",
   "notes": "MTP MLX build; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 3qfs3vvf, e3l0t7ag, vupevv7h; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 455.8,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored-oQ4-mtp-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 455.8 49.7 22.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/teh7edht",
   "source_url": "https://omlx.ai/benchmarks/performance/teh7edht",
   "tg_tokens": null,
   "tps": 49.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-teh7edht-ctx8192",
   "model": "Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored",
   "notes": "MTP MLX build; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: 3qfs3vvf, e3l0t7ag, vupevv7h; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 448.5,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored-oQ4-mtp-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 448.5 66.2 22.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/teh7edht",
   "source_url": "https://omlx.ai/benchmarks/performance/teh7edht",
   "tg_tokens": null,
   "tps": 66.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-teh7edht-ctx16384",
   "model": "Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored",
   "notes": "MTP MLX build; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: 3qfs3vvf, e3l0t7ag, vupevv7h; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 432.2,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored-oQ4-mtp-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 432.2 52.2 23.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/teh7edht",
   "source_url": "https://omlx.ai/benchmarks/performance/teh7edht",
   "tg_tokens": null,
   "tps": 52.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-mo6gjsx2-ctx1024",
   "model": "Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored",
   "notes": "MTP MLX build; oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: hii6a8dw, v1xjk34x; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 428.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored-oQ4-mtp-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 428.0 74.8 18.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mo6gjsx2",
   "source_url": "https://omlx.ai/benchmarks/performance/mo6gjsx2",
   "tg_tokens": null,
   "tps": 74.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-mo6gjsx2-ctx4096",
   "model": "Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored",
   "notes": "MTP MLX build; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: hii6a8dw, v1xjk34x; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 455.5,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored-oQ4-mtp-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 455.5 108.2 22.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mo6gjsx2",
   "source_url": "https://omlx.ai/benchmarks/performance/mo6gjsx2",
   "tg_tokens": null,
   "tps": 108.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-mo6gjsx2-ctx16384",
   "model": "Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored",
   "notes": "MTP MLX build; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: hii6a8dw, v1xjk34x; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 432.4,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored-oQ4-mtp-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 432.4 67.3 23.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mo6gjsx2",
   "source_url": "https://omlx.ai/benchmarks/performance/mo6gjsx2",
   "tg_tokens": null,
   "tps": 67.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5ahxc6vp-ctx1024",
   "model": "Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored",
   "notes": "MTP MLX build; oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: 8b2koier, p29u4eex, wpv4muk8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 427.4,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored-oQ4-mtp-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 427.4 69.7 18.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ahxc6vp",
   "source_url": "https://omlx.ai/benchmarks/performance/5ahxc6vp",
   "tg_tokens": null,
   "tps": 69.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5ahxc6vp-ctx4096",
   "model": "Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored",
   "notes": "MTP MLX build; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 8b2koier, p29u4eex, wpv4muk8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 456.2,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored-oQ4-mtp-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 456.2 70.5 22.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ahxc6vp",
   "source_url": "https://omlx.ai/benchmarks/performance/5ahxc6vp",
   "tg_tokens": null,
   "tps": 70.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5ahxc6vp-ctx8192",
   "model": "Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored",
   "notes": "MTP MLX build; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: 8b2koier, p29u4eex, wpv4muk8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 448.7,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored-oQ4-mtp-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 448.7 79.3 22.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ahxc6vp",
   "source_url": "https://omlx.ai/benchmarks/performance/5ahxc6vp",
   "tg_tokens": null,
   "tps": 79.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5ahxc6vp-ctx16384",
   "model": "Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored",
   "notes": "MTP MLX build; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: 8b2koier, p29u4eex, wpv4muk8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 432.3,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored-oQ4-mtp-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 432.3 90.9 23.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ahxc6vp",
   "source_url": "https://omlx.ai/benchmarks/performance/5ahxc6vp",
   "tg_tokens": null,
   "tps": 90.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2v0jmnz3-ctx1024",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: uj5y5dc5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 601.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 601.9 29.3 26.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2v0jmnz3",
   "source_url": "https://omlx.ai/benchmarks/performance/2v0jmnz3",
   "tg_tokens": null,
   "tps": 29.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2v0jmnz3-ctx4096",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 29.4 GB at this context; duplicate traces of the same session: uj5y5dc5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 648.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 648.6 23.5 29.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2v0jmnz3",
   "source_url": "https://omlx.ai/benchmarks/performance/2v0jmnz3",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4ql5m55v-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: dkqald74; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1137.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,137 72.3 17.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ql5m55v",
   "source_url": "https://omlx.ai/benchmarks/performance/4ql5m55v",
   "tg_tokens": null,
   "tps": 72.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4ql5m55v-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: dkqald74; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1348.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,348 98.2 19.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ql5m55v",
   "source_url": "https://omlx.ai/benchmarks/performance/4ql5m55v",
   "tg_tokens": null,
   "tps": 98.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-xnmp1p8t-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 20.6 GB; TTFT 36681 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 893.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k tokens 893.3 PP tok/s 136.5 TG tok/s 36681 TTFT (ms) 20.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xnmp1p8t",
   "source_url": "https://omlx.ai/benchmarks/performance/xnmp1p8t",
   "tg_tokens": null,
   "tps": 136.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n4zo6lhk-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: yq9bukxh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 908.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 908.6 70.5 17.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n4zo6lhk",
   "source_url": "https://omlx.ai/benchmarks/performance/n4zo6lhk",
   "tg_tokens": null,
   "tps": 70.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n4zo6lhk-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: yq9bukxh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 985.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 985.4 65.0 18.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n4zo6lhk",
   "source_url": "https://omlx.ai/benchmarks/performance/n4zo6lhk",
   "tg_tokens": null,
   "tps": 65.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2qjdjph6-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 74.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1983.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,983 51.9 74.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2qjdjph6",
   "source_url": "https://omlx.ai/benchmarks/performance/2qjdjph6",
   "tg_tokens": null,
   "tps": 51.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2qjdjph6-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 75.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2791.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,791 42.2 75.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2qjdjph6",
   "source_url": "https://omlx.ai/benchmarks/performance/2qjdjph6",
   "tg_tokens": null,
   "tps": 42.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1e9oqvl9-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 74.7 GB; TTFT 525 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1952.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,952 PP tok/s 49.5 TG tok/s 525 TTFT (ms) 74.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1e9oqvl9",
   "source_url": "https://omlx.ai/benchmarks/performance/1e9oqvl9",
   "tg_tokens": null,
   "tps": 49.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ro7gfa7y-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 74.7 GB at this context; duplicate traces of the same session: n896km2u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1884.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,884 50.9 74.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ro7gfa7y",
   "source_url": "https://omlx.ai/benchmarks/performance/ro7gfa7y",
   "tg_tokens": null,
   "tps": 50.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ro7gfa7y-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 75.1 GB at this context; duplicate traces of the same session: n896km2u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2736.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,736 44.5 75.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ro7gfa7y",
   "source_url": "https://omlx.ai/benchmarks/performance/ro7gfa7y",
   "tg_tokens": null,
   "tps": 44.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pc3tgq1g-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 74.7 GB; TTFT 518 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1978.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,978 PP tok/s 44.2 TG tok/s 518 TTFT (ms) 74.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pc3tgq1g",
   "source_url": "https://omlx.ai/benchmarks/performance/pc3tgq1g",
   "tg_tokens": null,
   "tps": 44.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i23tn7u7-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 75.1 GB; TTFT 1452 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2821.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 2,821 PP tok/s 42.4 TG tok/s 1452 TTFT (ms) 75.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i23tn7u7",
   "source_url": "https://omlx.ai/benchmarks/performance/i23tn7u7",
   "tg_tokens": null,
   "tps": 42.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3wfw6jln-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 74.7 GB at this context; duplicate traces of the same session: vjxrqxv7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1900.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,900 46.2 74.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3wfw6jln",
   "source_url": "https://omlx.ai/benchmarks/performance/3wfw6jln",
   "tg_tokens": null,
   "tps": 46.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3wfw6jln-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 75.1 GB at this context; duplicate traces of the same session: vjxrqxv7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2791.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,791 39.4 75.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3wfw6jln",
   "source_url": "https://omlx.ai/benchmarks/performance/3wfw6jln",
   "tg_tokens": null,
   "tps": 39.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dhyod5ia-ctx1024",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "mlx-community mixed build; oMLX community benchmark trace, single run at 1k context; peak memory 80.0 GB; TTFT 1243 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 823.6,
   "provenance": "community",
   "quant": "2.4bit",
   "quote": "mlx-community-DeepSeek-V4-Flash-0731-2.4bit-mixed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 1k tokens 823.6 PP tok/s 41.0 TG tok/s 1243 TTFT (ms) 80.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dhyod5ia",
   "source_url": "https://omlx.ai/benchmarks/performance/dhyod5ia",
   "tg_tokens": null,
   "tps": 41.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ylotgmvn-ctx1024",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "mlx-community mixed build; oMLX community benchmark trace, single run at 1k context; peak memory 87.8 GB; TTFT 1280 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 799.8,
   "provenance": "community",
   "quant": "2.4bit",
   "quote": "mlx-community-DeepSeek-V4-Flash-0731-2.4bit-mixed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 1k tokens 799.8 PP tok/s 44.0 TG tok/s 1280 TTFT (ms) 87.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ylotgmvn",
   "source_url": "https://omlx.ai/benchmarks/performance/ylotgmvn",
   "tg_tokens": null,
   "tps": 44.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-l5ajh8ug-ctx1024",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 155.8 GB at this context; duplicate traces of the same session: q2i4q90v, wlcdvrxp, xl8o2vzp; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 565.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 565.3 36.1 155.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l5ajh8ug",
   "source_url": "https://omlx.ai/benchmarks/performance/l5ajh8ug",
   "tg_tokens": null,
   "tps": 36.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-l5ajh8ug-ctx4096",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.3 GB at this context; duplicate traces of the same session: q2i4q90v, wlcdvrxp, xl8o2vzp; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 618.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 618.8 37.8 156.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l5ajh8ug",
   "source_url": "https://omlx.ai/benchmarks/performance/l5ajh8ug",
   "tg_tokens": null,
   "tps": 37.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-l5ajh8ug-ctx8192",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.3 GB at this context; duplicate traces of the same session: q2i4q90v, wlcdvrxp, xl8o2vzp; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 608.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 608.5 48.5 156.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l5ajh8ug",
   "source_url": "https://omlx.ai/benchmarks/performance/l5ajh8ug",
   "tg_tokens": null,
   "tps": 48.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-l5ajh8ug-ctx16384",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.4 GB at this context; duplicate traces of the same session: q2i4q90v, wlcdvrxp, xl8o2vzp; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 598.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 598.5 43.7 156.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l5ajh8ug",
   "source_url": "https://omlx.ai/benchmarks/performance/l5ajh8ug",
   "tg_tokens": null,
   "tps": 43.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-wt5vakym-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.0 GB at this context; duplicate traces of the same session: 4z3b77u7, 9cpnx3qx, y7lidqck; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 94.2,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4.1-Flash-oQ2e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 94.2 16.0 100.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wt5vakym",
   "source_url": "https://omlx.ai/benchmarks/performance/wt5vakym",
   "tg_tokens": null,
   "tps": 16.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-wt5vakym-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.0 GB at this context; duplicate traces of the same session: 4z3b77u7, 9cpnx3qx, y7lidqck; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 315.5,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4.1-Flash-oQ2e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 315.5 16.2 101.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wt5vakym",
   "source_url": "https://omlx.ai/benchmarks/performance/wt5vakym",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-wt5vakym-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.7 GB at this context; duplicate traces of the same session: 4z3b77u7, 9cpnx3qx, y7lidqck; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 429.6,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4.1-Flash-oQ2e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 8k 429.6 15.7 102.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wt5vakym",
   "source_url": "https://omlx.ai/benchmarks/performance/wt5vakym",
   "tg_tokens": null,
   "tps": 15.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-wt5vakym-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.7 GB at this context; duplicate traces of the same session: 4z3b77u7, 9cpnx3qx, y7lidqck; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 442.1,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4.1-Flash-oQ2e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 16k 442.1 19.2 102.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wt5vakym",
   "source_url": "https://omlx.ai/benchmarks/performance/wt5vakym",
   "tg_tokens": null,
   "tps": 19.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9091gkes-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.2 GB at this context; duplicate traces of the same session: 5s3abg33, 8not62pr, i79owm6f; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 505.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 505.9 35.5 174.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9091gkes",
   "source_url": "https://omlx.ai/benchmarks/performance/9091gkes",
   "tg_tokens": null,
   "tps": 35.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9091gkes-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.8 GB at this context; duplicate traces of the same session: 5s3abg33, 8not62pr, i79owm6f; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 554.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 554.2 25.3 175.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9091gkes",
   "source_url": "https://omlx.ai/benchmarks/performance/9091gkes",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9091gkes-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 176.4 GB at this context; duplicate traces of the same session: 5s3abg33, 8not62pr, i79owm6f; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 536.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 536.7 38.7 176.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9091gkes",
   "source_url": "https://omlx.ai/benchmarks/performance/9091gkes",
   "tg_tokens": null,
   "tps": 38.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9091gkes-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.2 GB at this context; duplicate traces of the same session: 5s3abg33, 8not62pr, i79owm6f; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 526.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 526.3 60.8 177.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9091gkes",
   "source_url": "https://omlx.ai/benchmarks/performance/9091gkes",
   "tg_tokens": null,
   "tps": 60.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-jpo4d0bd-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.2 GB at this context; duplicate traces of the same session: fiwier88, h85akc3r, i68s9zln; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 507.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 507.0 38.9 174.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jpo4d0bd",
   "source_url": "https://omlx.ai/benchmarks/performance/jpo4d0bd",
   "tg_tokens": null,
   "tps": 38.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jpo4d0bd-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.8 GB at this context; duplicate traces of the same session: fiwier88, h85akc3r, i68s9zln; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 553.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 553.1 69.1 175.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jpo4d0bd",
   "source_url": "https://omlx.ai/benchmarks/performance/jpo4d0bd",
   "tg_tokens": null,
   "tps": 69.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-jpo4d0bd-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 176.4 GB at this context; duplicate traces of the same session: fiwier88, h85akc3r, i68s9zln; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 537.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 537.0 35.9 176.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jpo4d0bd",
   "source_url": "https://omlx.ai/benchmarks/performance/jpo4d0bd",
   "tg_tokens": null,
   "tps": 35.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-jpo4d0bd-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.2 GB at this context; duplicate traces of the same session: fiwier88, h85akc3r, i68s9zln; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 526.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 526.6 50.0 177.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jpo4d0bd",
   "source_url": "https://omlx.ai/benchmarks/performance/jpo4d0bd",
   "tg_tokens": null,
   "tps": 50.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-fd4xucbr-ctx1024",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: bqifrrdf, e1nbnm8m, vo9nddgc, wn8tzs8p; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2288.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 2,288 117.6 19.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fd4xucbr",
   "source_url": "https://omlx.ai/benchmarks/performance/fd4xucbr",
   "tg_tokens": null,
   "tps": 117.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-fd4xucbr-ctx4096",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: bqifrrdf, e1nbnm8m, vo9nddgc, wn8tzs8p; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2357.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 2,357 113.8 20.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fd4xucbr",
   "source_url": "https://omlx.ai/benchmarks/performance/fd4xucbr",
   "tg_tokens": null,
   "tps": 113.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-fd4xucbr-ctx8192",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: bqifrrdf, e1nbnm8m, vo9nddgc, wn8tzs8p; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2439.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 2,439 124.1 20.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fd4xucbr",
   "source_url": "https://omlx.ai/benchmarks/performance/fd4xucbr",
   "tg_tokens": null,
   "tps": 124.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-fd4xucbr-ctx16384",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: bqifrrdf, e1nbnm8m, vo9nddgc, wn8tzs8p; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2401.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 2,401 117.0 20.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fd4xucbr",
   "source_url": "https://omlx.ai/benchmarks/performance/fd4xucbr",
   "tg_tokens": null,
   "tps": 117.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-fd4xucbr-ctx32768",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: bqifrrdf, e1nbnm8m, vo9nddgc, wn8tzs8p; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2223.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,223 109.4 20.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fd4xucbr",
   "source_url": "https://omlx.ai/benchmarks/performance/fd4xucbr",
   "tg_tokens": null,
   "tps": 109.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-krea2v4q-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: skc85fif; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2594.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 2,594 119.8 22.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/krea2v4q",
   "source_url": "https://omlx.ai/benchmarks/performance/krea2v4q",
   "tg_tokens": null,
   "tps": 119.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-krea2v4q-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: skc85fif; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3401.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 3,401 116.6 24.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/krea2v4q",
   "source_url": "https://omlx.ai/benchmarks/performance/krea2v4q",
   "tg_tokens": null,
   "tps": 116.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-krea2v4q-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: skc85fif; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3228.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 3,228 128.4 24.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/krea2v4q",
   "source_url": "https://omlx.ai/benchmarks/performance/krea2v4q",
   "tg_tokens": null,
   "tps": 128.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-krea2v4q-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: skc85fif; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2879.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 2,879 127.8 24.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/krea2v4q",
   "source_url": "https://omlx.ai/benchmarks/performance/krea2v4q",
   "tg_tokens": null,
   "tps": 127.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-zgmr0s5g-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: aot8fec7; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 81.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 81.5 11.7 19.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zgmr0s5g",
   "source_url": "https://omlx.ai/benchmarks/performance/zgmr0s5g",
   "tg_tokens": null,
   "tps": 11.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-zgmr0s5g-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: aot8fec7; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 78.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 78.7 11.8 23.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zgmr0s5g",
   "source_url": "https://omlx.ai/benchmarks/performance/zgmr0s5g",
   "tg_tokens": null,
   "tps": 11.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-l6hzpsrm-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 0fuu0m0m, 1e86c6ro, alnuhj7f, bys3pm70, kmn06v7v, p7yy20me; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 433.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 433.0 10.0 28.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l6hzpsrm",
   "source_url": "https://omlx.ai/benchmarks/performance/l6hzpsrm",
   "tg_tokens": null,
   "tps": 10.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-l6hzpsrm-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: 0fuu0m0m, 1e86c6ro, alnuhj7f, bys3pm70, kmn06v7v, p7yy20me; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 473.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 473.2 9.9 29.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l6hzpsrm",
   "source_url": "https://omlx.ai/benchmarks/performance/l6hzpsrm",
   "tg_tokens": null,
   "tps": 9.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-l6hzpsrm-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 29.8 GB at this context; duplicate traces of the same session: 0fuu0m0m, 1e86c6ro, alnuhj7f, bys3pm70, kmn06v7v, p7yy20me; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 416.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 416.5 9.8 29.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l6hzpsrm",
   "source_url": "https://omlx.ai/benchmarks/performance/l6hzpsrm",
   "tg_tokens": null,
   "tps": 9.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-l6hzpsrm-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: 0fuu0m0m, 1e86c6ro, alnuhj7f, bys3pm70, kmn06v7v, p7yy20me; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 401.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 16k 401.7 9.6 30.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l6hzpsrm",
   "source_url": "https://omlx.ai/benchmarks/performance/l6hzpsrm",
   "tg_tokens": null,
   "tps": 9.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-l6hzpsrm-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 31.3 GB at this context; duplicate traces of the same session: 0fuu0m0m, 1e86c6ro, alnuhj7f, bys3pm70, kmn06v7v, p7yy20me; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 394.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 32k 394.6 9.2 31.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l6hzpsrm",
   "source_url": "https://omlx.ai/benchmarks/performance/l6hzpsrm",
   "tg_tokens": null,
   "tps": 9.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-l6hzpsrm-ctx65536",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 33.3 GB at this context; duplicate traces of the same session: 0fuu0m0m, 1e86c6ro, alnuhj7f, bys3pm70, kmn06v7v, p7yy20me; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 351.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 64k 351.2 8.5 33.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l6hzpsrm",
   "source_url": "https://omlx.ai/benchmarks/performance/l6hzpsrm",
   "tg_tokens": null,
   "tps": 8.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-l6hzpsrm-ctx131072",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: 0fuu0m0m, 1e86c6ro, alnuhj7f, bys3pm70, kmn06v7v, p7yy20me; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 286.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 128k 286.7 7.5 37.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l6hzpsrm",
   "source_url": "https://omlx.ai/benchmarks/performance/l6hzpsrm",
   "tg_tokens": null,
   "tps": 7.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-s9l0nqob-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 0z2uuqeo; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 465.1,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 465.1 16.9 20.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s9l0nqob",
   "source_url": "https://omlx.ai/benchmarks/performance/s9l0nqob",
   "tg_tokens": null,
   "tps": 16.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-s9l0nqob-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 0z2uuqeo; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 460.2,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 460.2 15.3 19.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s9l0nqob",
   "source_url": "https://omlx.ai/benchmarks/performance/s9l0nqob",
   "tg_tokens": null,
   "tps": 15.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-h83dq6rl-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 2yi8l7nq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 595.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 595.1 27.4 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h83dq6rl",
   "source_url": "https://omlx.ai/benchmarks/performance/h83dq6rl",
   "tg_tokens": null,
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-h83dq6rl-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 2yi8l7nq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 633.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 633.3 30.6 19.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h83dq6rl",
   "source_url": "https://omlx.ai/benchmarks/performance/h83dq6rl",
   "tg_tokens": null,
   "tps": 30.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yvk7yc6b-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 74.7 GB; TTFT 524 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1954.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,954 PP tok/s 46.7 TG tok/s 524 TTFT (ms) 74.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yvk7yc6b",
   "source_url": "https://omlx.ai/benchmarks/performance/yvk7yc6b",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0jwiaiyq-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 74.7 GB at this context; duplicate traces of the same session: x26njoh3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1946.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,946 51.9 74.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0jwiaiyq",
   "source_url": "https://omlx.ai/benchmarks/performance/0jwiaiyq",
   "tg_tokens": null,
   "tps": 51.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0jwiaiyq-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 75.1 GB at this context; duplicate traces of the same session: x26njoh3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2764.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,764 48.7 75.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0jwiaiyq",
   "source_url": "https://omlx.ai/benchmarks/performance/0jwiaiyq",
   "tg_tokens": null,
   "tps": 48.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nu4i4982-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 74.7 GB at this context; duplicate traces of the same session: 3tab7ek2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2023.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,023 52.9 74.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nu4i4982",
   "source_url": "https://omlx.ai/benchmarks/performance/nu4i4982",
   "tg_tokens": null,
   "tps": 52.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nu4i4982-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 75.1 GB at this context; duplicate traces of the same session: 3tab7ek2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2829.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,829 46.1 75.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nu4i4982",
   "source_url": "https://omlx.ai/benchmarks/performance/nu4i4982",
   "tg_tokens": null,
   "tps": 46.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zw6exuv2-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 75.1 GB; TTFT 1508 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2716.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 2,716 PP tok/s 47.8 TG tok/s 1508 TTFT (ms) 75.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zw6exuv2",
   "source_url": "https://omlx.ai/benchmarks/performance/zw6exuv2",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-95prua07-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 74.7 GB at this context; duplicate traces of the same session: j27c1roy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1970.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,970 54.7 74.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/95prua07",
   "source_url": "https://omlx.ai/benchmarks/performance/95prua07",
   "tg_tokens": null,
   "tps": 54.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-95prua07-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 75.1 GB at this context; duplicate traces of the same session: j27c1roy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2750.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,750 46.4 75.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/95prua07",
   "source_url": "https://omlx.ai/benchmarks/performance/95prua07",
   "tg_tokens": null,
   "tps": 46.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-64xwhbjy-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 69.6 GB at this context; duplicate traces of the same session: 31xhmtr5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2010.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,010 65.9 69.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/64xwhbjy",
   "source_url": "https://omlx.ai/benchmarks/performance/64xwhbjy",
   "tg_tokens": null,
   "tps": 65.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-64xwhbjy-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 70.2 GB at this context; duplicate traces of the same session: 31xhmtr5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2834.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,834 63.0 70.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/64xwhbjy",
   "source_url": "https://omlx.ai/benchmarks/performance/64xwhbjy",
   "tg_tokens": null,
   "tps": 63.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-sjjna1tk-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 81.5 GB at this context; duplicate traces of the same session: oeg1rnni; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1913.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,913 65.9 81.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sjjna1tk",
   "source_url": "https://omlx.ai/benchmarks/performance/sjjna1tk",
   "tg_tokens": null,
   "tps": 65.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-sjjna1tk-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 82.1 GB at this context; duplicate traces of the same session: oeg1rnni; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2804.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,804 65.1 82.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sjjna1tk",
   "source_url": "https://omlx.ai/benchmarks/performance/sjjna1tk",
   "tg_tokens": null,
   "tps": 65.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-28q1oqf4-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 81.5 GB at this context; duplicate traces of the same session: 34vsk3sv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1964.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,964 70.1 81.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/28q1oqf4",
   "source_url": "https://omlx.ai/benchmarks/performance/28q1oqf4",
   "tg_tokens": null,
   "tps": 70.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-28q1oqf4-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 82.1 GB at this context; duplicate traces of the same session: 34vsk3sv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2824.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,824 65.3 82.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/28q1oqf4",
   "source_url": "https://omlx.ai/benchmarks/performance/28q1oqf4",
   "tg_tokens": null,
   "tps": 65.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jgryewca-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 81.5 GB at this context; duplicate traces of the same session: a0py7yr3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1964.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,964 65.4 81.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jgryewca",
   "source_url": "https://omlx.ai/benchmarks/performance/jgryewca",
   "tg_tokens": null,
   "tps": 65.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jgryewca-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 82.1 GB at this context; duplicate traces of the same session: a0py7yr3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2840.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,840 68.1 82.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jgryewca",
   "source_url": "https://omlx.ai/benchmarks/performance/jgryewca",
   "tg_tokens": null,
   "tps": 68.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qpabcowv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: zc6jrd1r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 985.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 985.4 20.9 24.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qpabcowv",
   "source_url": "https://omlx.ai/benchmarks/performance/qpabcowv",
   "tg_tokens": null,
   "tps": 20.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qpabcowv-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: zc6jrd1r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 959.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 959.9 23.6 24.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qpabcowv",
   "source_url": "https://omlx.ai/benchmarks/performance/qpabcowv",
   "tg_tokens": null,
   "tps": 23.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-db36vl0v-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: s8awpkrx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1005.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,005 25.1 22.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/db36vl0v",
   "source_url": "https://omlx.ai/benchmarks/performance/db36vl0v",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-db36vl0v-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: s8awpkrx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 967.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 967.2 24.2 23.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/db36vl0v",
   "source_url": "https://omlx.ai/benchmarks/performance/db36vl0v",
   "tg_tokens": null,
   "tps": 24.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wd7k2xv6-ctx1024",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "mlx-community mixed build; oMLX community benchmark trace; peak memory 80.0 GB at this context; duplicate traces of the same session: 890rujdf, uoywbes1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 827.1,
   "provenance": "community",
   "quant": "2.4bit",
   "quote": "mlx-community-DeepSeek-V4-Flash-0731-2.4bit-mixed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 1k 827.1 40.7 80.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wd7k2xv6",
   "source_url": "https://omlx.ai/benchmarks/performance/wd7k2xv6",
   "tg_tokens": null,
   "tps": 40.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wd7k2xv6-ctx4096",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "mlx-community mixed build; oMLX community benchmark trace; peak memory 80.8 GB at this context; duplicate traces of the same session: 890rujdf, uoywbes1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 751.4,
   "provenance": "community",
   "quant": "2.4bit",
   "quote": "mlx-community-DeepSeek-V4-Flash-0731-2.4bit-mixed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 4k 751.4 37.7 80.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wd7k2xv6",
   "source_url": "https://omlx.ai/benchmarks/performance/wd7k2xv6",
   "tg_tokens": null,
   "tps": 37.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wd7k2xv6-ctx8192",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "mlx-community mixed build; oMLX community benchmark trace; peak memory 81.4 GB at this context; duplicate traces of the same session: 890rujdf, uoywbes1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 830.6,
   "provenance": "community",
   "quant": "2.4bit",
   "quote": "mlx-community-DeepSeek-V4-Flash-0731-2.4bit-mixed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 8k 830.6 36.8 81.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wd7k2xv6",
   "source_url": "https://omlx.ai/benchmarks/performance/wd7k2xv6",
   "tg_tokens": null,
   "tps": 36.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-73ux8lfw-ctx1024",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "mlx-community mixed build; oMLX community benchmark trace, single run at 1k context; peak memory 80.0 GB; TTFT 2037 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 502.7,
   "provenance": "community",
   "quant": "2.4bit",
   "quote": "mlx-community-DeepSeek-V4-Flash-0731-2.4bit-mixed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 1k tokens 502.7 PP tok/s 40.9 TG tok/s 2037 TTFT (ms) 80.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/73ux8lfw",
   "source_url": "https://omlx.ai/benchmarks/performance/73ux8lfw",
   "tg_tokens": null,
   "tps": 40.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-a2r3u4gx-ctx1024",
   "model": "GLM-4.5-Air",
   "notes": "oMLX community benchmark trace; peak memory 57.0 GB at this context; duplicate traces of the same session: l0ldi033; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2131.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-4.5-Air-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k 2,131 72.3 57.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a2r3u4gx",
   "source_url": "https://omlx.ai/benchmarks/performance/a2r3u4gx",
   "tg_tokens": null,
   "tps": 72.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-a2r3u4gx-ctx4096",
   "model": "GLM-4.5-Air",
   "notes": "oMLX community benchmark trace; peak memory 58.0 GB at this context; duplicate traces of the same session: l0ldi033; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2239.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-4.5-Air-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 4k 2,239 63.3 58.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a2r3u4gx",
   "source_url": "https://omlx.ai/benchmarks/performance/a2r3u4gx",
   "tg_tokens": null,
   "tps": 63.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-4lxzv69g-ctx1024",
   "model": "Llama-3-Groq-8B-Tool-Use",
   "notes": "oMLX community benchmark trace; peak memory 5.0 GB at this context; duplicate traces of the same session: ieie3rm1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 475.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Llama-3-Groq-8B-Tool-Use-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 24 GB \u00b7 4bit; 1k 475.1 53.1 5.0 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4lxzv69g",
   "source_url": "https://omlx.ai/benchmarks/performance/4lxzv69g",
   "tg_tokens": null,
   "tps": 53.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-4lxzv69g-ctx4096",
   "model": "Llama-3-Groq-8B-Tool-Use",
   "notes": "oMLX community benchmark trace; peak memory 5.4 GB at this context; duplicate traces of the same session: ieie3rm1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 457.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Llama-3-Groq-8B-Tool-Use-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 24 GB \u00b7 4bit; 4k 457.4 47.9 5.4 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4lxzv69g",
   "source_url": "https://omlx.ai/benchmarks/performance/4lxzv69g",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-lbqpb44f-ctx1024",
   "model": "Llama-3.2-3B-Instruct-uncensored",
   "notes": "oMLX community benchmark trace; peak memory 3.1 GB at this context; duplicate traces of the same session: a2kimyu0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1049.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Llama-3.2-3B-Instruct-uncensored-6bit on M4 Pro (20c); M4 Pro (20c) \u00b7 24 GB \u00b7 6bit; 1k 1,049 76.6 3.1 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lbqpb44f",
   "source_url": "https://omlx.ai/benchmarks/performance/lbqpb44f",
   "tg_tokens": null,
   "tps": 76.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-lbqpb44f-ctx4096",
   "model": "Llama-3.2-3B-Instruct-uncensored",
   "notes": "oMLX community benchmark trace; peak memory 3.5 GB at this context; duplicate traces of the same session: a2kimyu0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1049.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Llama-3.2-3B-Instruct-uncensored-6bit on M4 Pro (20c); M4 Pro (20c) \u00b7 24 GB \u00b7 6bit; 4k 1,049 72.5 3.5 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lbqpb44f",
   "source_url": "https://omlx.ai/benchmarks/performance/lbqpb44f",
   "tg_tokens": null,
   "tps": 72.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-tfyedmi4-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: bjh8r4g6, h826rv1f; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 754.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 754.6 65.7 20.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tfyedmi4",
   "source_url": "https://omlx.ai/benchmarks/performance/tfyedmi4",
   "tg_tokens": null,
   "tps": 65.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-tfyedmi4-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: bjh8r4g6, h826rv1f; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 794.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 794.6 56.5 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tfyedmi4",
   "source_url": "https://omlx.ai/benchmarks/performance/tfyedmi4",
   "tg_tokens": null,
   "tps": 56.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-tfyedmi4-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: bjh8r4g6, h826rv1f; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 549.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 32k 549.8 22.9 22.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tfyedmi4",
   "source_url": "https://omlx.ai/benchmarks/performance/tfyedmi4",
   "tg_tokens": null,
   "tps": 22.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-247wlclz-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: ksdffutk; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 756.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 756.9 80.7 20.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/247wlclz",
   "source_url": "https://omlx.ai/benchmarks/performance/247wlclz",
   "tg_tokens": null,
   "tps": 80.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-247wlclz-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: ksdffutk; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 794.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 794.3 57.5 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/247wlclz",
   "source_url": "https://omlx.ai/benchmarks/performance/247wlclz",
   "tg_tokens": null,
   "tps": 57.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-017tqoh2-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 30.6 GB; TTFT 11910 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1376.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 6bit; 16k tokens 1,376 PP tok/s 101.3 TG tok/s 11910 TTFT (ms) 30.6 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/017tqoh2",
   "source_url": "https://omlx.ai/benchmarks/performance/017tqoh2",
   "tg_tokens": null,
   "tps": 101.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-su92qg5v-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: 8izp46kn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 470.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 470.1 9.9 29.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/su92qg5v",
   "source_url": "https://omlx.ai/benchmarks/performance/su92qg5v",
   "tg_tokens": null,
   "tps": 9.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-su92qg5v-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: 8izp46kn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 448.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 16k 448.0 9.6 30.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/su92qg5v",
   "source_url": "https://omlx.ai/benchmarks/performance/su92qg5v",
   "tg_tokens": null,
   "tps": 9.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-vahvzpgr-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: j0p0fm1t, nv47kvtv; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 170.1,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-27B-nvfp4 on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 1k 170.1 14.5 17.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vahvzpgr",
   "source_url": "https://omlx.ai/benchmarks/performance/vahvzpgr",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-vahvzpgr-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: j0p0fm1t, nv47kvtv; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 157.1,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-27B-nvfp4 on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 4k 157.1 13.0 18.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vahvzpgr",
   "source_url": "https://omlx.ai/benchmarks/performance/vahvzpgr",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-vahvzpgr-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: j0p0fm1t, nv47kvtv; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 140.1,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-27B-nvfp4 on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 8k 140.1 8.9 18.4 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vahvzpgr",
   "source_url": "https://omlx.ai/benchmarks/performance/vahvzpgr",
   "tg_tokens": null,
   "tps": 8.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-o65qzjf6-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 17.8 GB; TTFT 1046 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 979.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k tokens 979.2 PP tok/s 55.8 TG tok/s 1046 TTFT (ms) 17.8 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o65qzjf6",
   "source_url": "https://omlx.ai/benchmarks/performance/o65qzjf6",
   "tg_tokens": null,
   "tps": 55.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5aco8p94-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 75.1 GB; TTFT 1446 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2832.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 2,832 PP tok/s 51.3 TG tok/s 1446 TTFT (ms) 75.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5aco8p94",
   "source_url": "https://omlx.ai/benchmarks/performance/5aco8p94",
   "tg_tokens": null,
   "tps": 51.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-w4que3bv-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 75.1 GB; TTFT 1405 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2915.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 2,915 PP tok/s 51.2 TG tok/s 1405 TTFT (ms) 75.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w4que3bv",
   "source_url": "https://omlx.ai/benchmarks/performance/w4que3bv",
   "tg_tokens": null,
   "tps": 51.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9wqtvvac-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 74.7 GB at this context; duplicate traces of the same session: jh7kiecc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1997.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,997 56.8 74.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9wqtvvac",
   "source_url": "https://omlx.ai/benchmarks/performance/9wqtvvac",
   "tg_tokens": null,
   "tps": 56.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9wqtvvac-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 75.1 GB at this context; duplicate traces of the same session: jh7kiecc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2805.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,805 52.4 75.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9wqtvvac",
   "source_url": "https://omlx.ai/benchmarks/performance/9wqtvvac",
   "tg_tokens": null,
   "tps": 52.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vjun18qm-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 74.9 GB at this context; duplicate traces of the same session: cprfztm3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1945.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,945 49.1 74.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vjun18qm",
   "source_url": "https://omlx.ai/benchmarks/performance/vjun18qm",
   "tg_tokens": null,
   "tps": 49.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vjun18qm-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 75.1 GB at this context; duplicate traces of the same session: cprfztm3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2784.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,784 45.6 75.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vjun18qm",
   "source_url": "https://omlx.ai/benchmarks/performance/vjun18qm",
   "tg_tokens": null,
   "tps": 45.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zjzu0e32-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 74.8 GB at this context; duplicate traces of the same session: d75a4a22; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1948.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,948 47.3 74.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zjzu0e32",
   "source_url": "https://omlx.ai/benchmarks/performance/zjzu0e32",
   "tg_tokens": null,
   "tps": 47.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zjzu0e32-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 75.1 GB at this context; duplicate traces of the same session: d75a4a22; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2773.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,773 43.6 75.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zjzu0e32",
   "source_url": "https://omlx.ai/benchmarks/performance/zjzu0e32",
   "tg_tokens": null,
   "tps": 43.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rv46gn00-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 74.7 GB at this context; duplicate traces of the same session: nwtxoix1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1985.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,985 46.1 74.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rv46gn00",
   "source_url": "https://omlx.ai/benchmarks/performance/rv46gn00",
   "tg_tokens": null,
   "tps": 46.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rv46gn00-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 75.1 GB at this context; duplicate traces of the same session: nwtxoix1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2775.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,775 49.6 75.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rv46gn00",
   "source_url": "https://omlx.ai/benchmarks/performance/rv46gn00",
   "tg_tokens": null,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-62avi6hw-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 74.9 GB at this context; duplicate traces of the same session: 85edcz7n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2184.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,184 59.8 74.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/62avi6hw",
   "source_url": "https://omlx.ai/benchmarks/performance/62avi6hw",
   "tg_tokens": null,
   "tps": 59.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-62avi6hw-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 75.1 GB at this context; duplicate traces of the same session: 85edcz7n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3101.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 3,101 59.0 75.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/62avi6hw",
   "source_url": "https://omlx.ai/benchmarks/performance/62avi6hw",
   "tg_tokens": null,
   "tps": 59.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bsk8e6w4-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 75.1 GB; TTFT 1327 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3087.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 3,087 PP tok/s 62.2 TG tok/s 1327 TTFT (ms) 75.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bsk8e6w4",
   "source_url": "https://omlx.ai/benchmarks/performance/bsk8e6w4",
   "tg_tokens": null,
   "tps": 62.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xoclrmgq-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 74.7 GB; TTFT 463 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2210.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 2,210 PP tok/s 59.0 TG tok/s 463 TTFT (ms) 74.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xoclrmgq",
   "source_url": "https://omlx.ai/benchmarks/performance/xoclrmgq",
   "tg_tokens": null,
   "tps": 59.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5kbewvm4-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 74.7 GB at this context; duplicate traces of the same session: e5oz86di; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2184.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,184 56.9 74.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5kbewvm4",
   "source_url": "https://omlx.ai/benchmarks/performance/5kbewvm4",
   "tg_tokens": null,
   "tps": 56.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5kbewvm4-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 75.1 GB at this context; duplicate traces of the same session: e5oz86di; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3093.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 3,093 54.2 75.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5kbewvm4",
   "source_url": "https://omlx.ai/benchmarks/performance/5kbewvm4",
   "tg_tokens": null,
   "tps": 54.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-35f3xynn-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 74.7 GB at this context; duplicate traces of the same session: h1y1ov1r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2038.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,038 51.0 74.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/35f3xynn",
   "source_url": "https://omlx.ai/benchmarks/performance/35f3xynn",
   "tg_tokens": null,
   "tps": 51.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-35f3xynn-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 75.1 GB at this context; duplicate traces of the same session: h1y1ov1r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2864.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,864 44.3 75.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/35f3xynn",
   "source_url": "https://omlx.ai/benchmarks/performance/35f3xynn",
   "tg_tokens": null,
   "tps": 44.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ewo3s7xy-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 74.7 GB at this context; duplicate traces of the same session: tcafo7d4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2005.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,005 50.1 74.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ewo3s7xy",
   "source_url": "https://omlx.ai/benchmarks/performance/ewo3s7xy",
   "tg_tokens": null,
   "tps": 50.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ewo3s7xy-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 75.1 GB at this context; duplicate traces of the same session: tcafo7d4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2834.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,834 53.1 75.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ewo3s7xy",
   "source_url": "https://omlx.ai/benchmarks/performance/ewo3s7xy",
   "tg_tokens": null,
   "tps": 53.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ah63xqmr-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 74.7 GB at this context; duplicate traces of the same session: z0hlypo9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2032.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,032 51.4 74.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ah63xqmr",
   "source_url": "https://omlx.ai/benchmarks/performance/ah63xqmr",
   "tg_tokens": null,
   "tps": 51.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ah63xqmr-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 75.1 GB at this context; duplicate traces of the same session: z0hlypo9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2788.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,788 42.8 75.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ah63xqmr",
   "source_url": "https://omlx.ai/benchmarks/performance/ah63xqmr",
   "tg_tokens": null,
   "tps": 42.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-b2swu1xd-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 74.9 GB at this context; duplicate traces of the same session: 9156rgi2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2055.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,055 52.7 74.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b2swu1xd",
   "source_url": "https://omlx.ai/benchmarks/performance/b2swu1xd",
   "tg_tokens": null,
   "tps": 52.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-b2swu1xd-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 75.1 GB at this context; duplicate traces of the same session: 9156rgi2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2895.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,895 50.4 75.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b2swu1xd",
   "source_url": "https://omlx.ai/benchmarks/performance/b2swu1xd",
   "tg_tokens": null,
   "tps": 50.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-t3lbc0xr-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 74.7 GB at this context; duplicate traces of the same session: 3xb0pnus; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2049.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,049 51.9 74.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t3lbc0xr",
   "source_url": "https://omlx.ai/benchmarks/performance/t3lbc0xr",
   "tg_tokens": null,
   "tps": 51.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-t3lbc0xr-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 75.1 GB at this context; duplicate traces of the same session: 3xb0pnus; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2867.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,867 37.7 75.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t3lbc0xr",
   "source_url": "https://omlx.ai/benchmarks/performance/t3lbc0xr",
   "tg_tokens": null,
   "tps": 37.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-o13p9aro-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 74.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2026.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,026 49.5 74.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o13p9aro",
   "source_url": "https://omlx.ai/benchmarks/performance/o13p9aro",
   "tg_tokens": null,
   "tps": 49.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-o13p9aro-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 75.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2829.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,829 51.7 75.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o13p9aro",
   "source_url": "https://omlx.ai/benchmarks/performance/o13p9aro",
   "tg_tokens": null,
   "tps": 51.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rmd1p2d7-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 74.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2027.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,027 52.6 74.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rmd1p2d7",
   "source_url": "https://omlx.ai/benchmarks/performance/rmd1p2d7",
   "tg_tokens": null,
   "tps": 52.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rmd1p2d7-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 75.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2790.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,790 53.4 75.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rmd1p2d7",
   "source_url": "https://omlx.ai/benchmarks/performance/rmd1p2d7",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0czy0xgl-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 82.1 GB; TTFT 1432 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2860.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 2,860 PP tok/s 67.7 TG tok/s 1432 TTFT (ms) 82.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0czy0xgl",
   "source_url": "https://omlx.ai/benchmarks/performance/0czy0xgl",
   "tg_tokens": null,
   "tps": 67.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-x6qvytlw-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 73.3 GB; TTFT 118320 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 553.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k tokens 553.9 PP tok/s 37.1 TG tok/s 118320 TTFT (ms) 73.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x6qvytlw",
   "source_url": "https://omlx.ai/benchmarks/performance/x6qvytlw",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-1l8m0orx-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: zaxdilhe; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 413.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 413.1 19.9 70.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1l8m0orx",
   "source_url": "https://omlx.ai/benchmarks/performance/1l8m0orx",
   "tg_tokens": null,
   "tps": 19.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-1l8m0orx-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: zaxdilhe; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 495.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 495.5 31.3 71.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1l8m0orx",
   "source_url": "https://omlx.ai/benchmarks/performance/1l8m0orx",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-ybxn26ut-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 69.3 GB; TTFT 1987 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 515.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 515.3 PP tok/s 53.9 TG tok/s 1987 TTFT (ms) 69.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ybxn26ut",
   "source_url": "https://omlx.ai/benchmarks/performance/ybxn26ut",
   "tg_tokens": null,
   "tps": 53.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-5sy20zo7-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.6 GB at this context; duplicate traces of the same session: sr44v2bq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 819.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 819.3 138.1 142.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5sy20zo7",
   "source_url": "https://omlx.ai/benchmarks/performance/5sy20zo7",
   "tg_tokens": null,
   "tps": 138.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-5sy20zo7-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 144.3 GB at this context; duplicate traces of the same session: sr44v2bq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4164.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 8k 4,164 55.8 144.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5sy20zo7",
   "source_url": "https://omlx.ai/benchmarks/performance/5sy20zo7",
   "tg_tokens": null,
   "tps": 55.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7p8xgd6u-ctx1024",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: lreb0sdq, rlytp62a, vt64k6te; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 444.8,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 2bit; 1k 444.8 38.2 11.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7p8xgd6u",
   "source_url": "https://omlx.ai/benchmarks/performance/7p8xgd6u",
   "tg_tokens": null,
   "tps": 38.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7p8xgd6u-ctx4096",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 14.2 GB at this context; duplicate traces of the same session: lreb0sdq, rlytp62a, vt64k6te; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 477.0,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 2bit; 4k 477.0 37.4 14.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7p8xgd6u",
   "source_url": "https://omlx.ai/benchmarks/performance/7p8xgd6u",
   "tg_tokens": null,
   "tps": 37.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-7p8xgd6u-ctx8192",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 14.7 GB at this context; duplicate traces of the same session: lreb0sdq, rlytp62a, vt64k6te; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 562.5,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 2bit; 8k 562.5 39.7 14.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7p8xgd6u",
   "source_url": "https://omlx.ai/benchmarks/performance/7p8xgd6u",
   "tg_tokens": null,
   "tps": 39.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7p8xgd6u-ctx16384",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 15.7 GB at this context; duplicate traces of the same session: lreb0sdq, rlytp62a, vt64k6te; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 479.6,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 2bit; 16k 479.6 32.5 15.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7p8xgd6u",
   "source_url": "https://omlx.ai/benchmarks/performance/7p8xgd6u",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bc0athvb-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace, single run at 4k context; peak memory 24.2 GB; TTFT 4223 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 970.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 970.0 PP tok/s 24.5 TG tok/s 4223 TTFT (ms) 24.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bc0athvb",
   "source_url": "https://omlx.ai/benchmarks/performance/bc0athvb",
   "tg_tokens": null,
   "tps": 24.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-la1s7nja-ctx1024",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 3.8 GB at this context; duplicate traces of the same session: tavheakg; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1394.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E2B-it-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 1,394 89.2 3.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/la1s7nja",
   "source_url": "https://omlx.ai/benchmarks/performance/la1s7nja",
   "tg_tokens": null,
   "tps": 89.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-la1s7nja-ctx4096",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 3.8 GB at this context; duplicate traces of the same session: tavheakg; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3152.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E2B-it-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 3,152 88.8 3.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/la1s7nja",
   "source_url": "https://omlx.ai/benchmarks/performance/la1s7nja",
   "tg_tokens": null,
   "tps": 88.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-xdujjl9w-ctx1024",
   "model": "MiniCPM5-2B-Abliterated-Uncensored",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 2.2 GB at this context; duplicate traces of the same session: 70i2yu02; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1451.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM5-2B-Abliterated-Uncensored-MLX-4Bit on M4 Pro (20c); M4 Pro (20c) \u00b7 24 GB \u00b7 4bit; 1k 1,451 129.5 2.2 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xdujjl9w",
   "source_url": "https://omlx.ai/benchmarks/performance/xdujjl9w",
   "tg_tokens": null,
   "tps": 129.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-xdujjl9w-ctx4096",
   "model": "MiniCPM5-2B-Abliterated-Uncensored",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 2.3 GB at this context; duplicate traces of the same session: 70i2yu02; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1354.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM5-2B-Abliterated-Uncensored-MLX-4Bit on M4 Pro (20c); M4 Pro (20c) \u00b7 24 GB \u00b7 4bit; 4k 1,354 112.2 2.3 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xdujjl9w",
   "source_url": "https://omlx.ai/benchmarks/performance/xdujjl9w",
   "tg_tokens": null,
   "tps": 112.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-gk1igksa-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 2yydf23p, flhyi6lv; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 775.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 775.2 90.9 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gk1igksa",
   "source_url": "https://omlx.ai/benchmarks/performance/gk1igksa",
   "tg_tokens": null,
   "tps": 90.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-gk1igksa-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 2yydf23p, flhyi6lv; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 884.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 884.8 86.5 21.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gk1igksa",
   "source_url": "https://omlx.ai/benchmarks/performance/gk1igksa",
   "tg_tokens": null,
   "tps": 86.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-gk1igksa-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: 2yydf23p, flhyi6lv; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 592.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 32k 592.1 61.3 23.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gk1igksa",
   "source_url": "https://omlx.ai/benchmarks/performance/gk1igksa",
   "tg_tokens": null,
   "tps": 61.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-cppov1r8-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: c2k0959z, xruojt38; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 770.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 770.6 90.6 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cppov1r8",
   "source_url": "https://omlx.ai/benchmarks/performance/cppov1r8",
   "tg_tokens": null,
   "tps": 90.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-cppov1r8-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: c2k0959z, xruojt38; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 912.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 912.8 86.6 21.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cppov1r8",
   "source_url": "https://omlx.ai/benchmarks/performance/cppov1r8",
   "tg_tokens": null,
   "tps": 86.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-cppov1r8-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: c2k0959z, xruojt38; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 604.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 32k 604.4 60.7 22.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cppov1r8",
   "source_url": "https://omlx.ai/benchmarks/performance/cppov1r8",
   "tg_tokens": null,
   "tps": 60.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-u6lluvrj-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: 31tip6ne, y4r1shpr; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 755.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 755.4 67.8 20.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u6lluvrj",
   "source_url": "https://omlx.ai/benchmarks/performance/u6lluvrj",
   "tg_tokens": null,
   "tps": 67.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-u6lluvrj-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 31tip6ne, y4r1shpr; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 793.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 793.7 59.7 21.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u6lluvrj",
   "source_url": "https://omlx.ai/benchmarks/performance/u6lluvrj",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-u6lluvrj-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 31tip6ne, y4r1shpr; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 536.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 32k 536.9 14.9 22.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u6lluvrj",
   "source_url": "https://omlx.ai/benchmarks/performance/u6lluvrj",
   "tg_tokens": null,
   "tps": 14.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-krru1433-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: xnw6u7pe; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1165.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 4bit; 1k 1,165 55.0 16.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/krru1433",
   "source_url": "https://omlx.ai/benchmarks/performance/krru1433",
   "tg_tokens": null,
   "tps": 55.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-krru1433-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: xnw6u7pe; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1664.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 4bit; 4k 1,664 53.6 17.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/krru1433",
   "source_url": "https://omlx.ai/benchmarks/performance/krru1433",
   "tg_tokens": null,
   "tps": 53.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5n8qndpq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: li0152ot; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 902.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 902.9 40.6 23.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5n8qndpq",
   "source_url": "https://omlx.ai/benchmarks/performance/5n8qndpq",
   "tg_tokens": null,
   "tps": 40.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5n8qndpq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: li0152ot; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 941.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 941.0 42.1 24.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5n8qndpq",
   "source_url": "https://omlx.ai/benchmarks/performance/5n8qndpq",
   "tg_tokens": null,
   "tps": 42.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ps4sls57-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 1eg2iclb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 904.2,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 904.2 36.4 23.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ps4sls57",
   "source_url": "https://omlx.ai/benchmarks/performance/ps4sls57",
   "tg_tokens": null,
   "tps": 36.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ps4sls57-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: 1eg2iclb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 942.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 942.1 39.3 24.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ps4sls57",
   "source_url": "https://omlx.ai/benchmarks/performance/ps4sls57",
   "tg_tokens": null,
   "tps": 39.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-6w48v0kv-ctx1024",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: y7b0jxtz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 636.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 636.9 28.4 26.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6w48v0kv",
   "source_url": "https://omlx.ai/benchmarks/performance/6w48v0kv",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-6w48v0kv-ctx4096",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 29.4 GB at this context; duplicate traces of the same session: y7b0jxtz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 661.2,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 661.2 24.9 29.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6w48v0kv",
   "source_url": "https://omlx.ai/benchmarks/performance/6w48v0kv",
   "tg_tokens": null,
   "tps": 24.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-b2d3f211-ctx1024",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: vlhs2mru; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 593.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 593.5 30.4 26.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b2d3f211",
   "source_url": "https://omlx.ai/benchmarks/performance/b2d3f211",
   "tg_tokens": null,
   "tps": 30.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-b2d3f211-ctx4096",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 29.4 GB at this context; duplicate traces of the same session: vlhs2mru; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 650.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 650.0 26.6 29.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b2d3f211",
   "source_url": "https://omlx.ai/benchmarks/performance/b2d3f211",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q1148h2d-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 11s6q7qg, 3pvn9x0h, 6quiyg27, mzxi8pbx, tllexnlf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 834.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 834.2 48.5 17.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q1148h2d",
   "source_url": "https://omlx.ai/benchmarks/performance/q1148h2d",
   "tg_tokens": null,
   "tps": 48.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q1148h2d-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 11s6q7qg, 3pvn9x0h, 6quiyg27, mzxi8pbx, tllexnlf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1120.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,120 60.7 19.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q1148h2d",
   "source_url": "https://omlx.ai/benchmarks/performance/q1148h2d",
   "tg_tokens": null,
   "tps": 60.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q1148h2d-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 11s6q7qg, 3pvn9x0h, 6quiyg27, mzxi8pbx, tllexnlf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1101.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,101 36.3 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q1148h2d",
   "source_url": "https://omlx.ai/benchmarks/performance/q1148h2d",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q1148h2d-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 11s6q7qg, 3pvn9x0h, 6quiyg27, mzxi8pbx, tllexnlf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1006.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,006 58.6 19.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q1148h2d",
   "source_url": "https://omlx.ai/benchmarks/performance/q1148h2d",
   "tg_tokens": null,
   "tps": 58.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q1148h2d-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 11s6q7qg, 3pvn9x0h, 6quiyg27, mzxi8pbx, tllexnlf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 849.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 849.4 50.5 21.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q1148h2d",
   "source_url": "https://omlx.ai/benchmarks/performance/q1148h2d",
   "tg_tokens": null,
   "tps": 50.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q1148h2d-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: 11s6q7qg, 3pvn9x0h, 6quiyg27, mzxi8pbx, tllexnlf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 624.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 624.2 22.8 26.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q1148h2d",
   "source_url": "https://omlx.ai/benchmarks/performance/q1148h2d",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-jes5mq6p-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 8yyhf9py, a7jrmv4x, en4b50ex; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 275.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 275.7 20.4 17.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jes5mq6p",
   "source_url": "https://omlx.ai/benchmarks/performance/jes5mq6p",
   "tg_tokens": null,
   "tps": 20.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-jes5mq6p-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: 8yyhf9py, a7jrmv4x, en4b50ex; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 226.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 226.1 18.4 18.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jes5mq6p",
   "source_url": "https://omlx.ai/benchmarks/performance/jes5mq6p",
   "tg_tokens": null,
   "tps": 18.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-jes5mq6p-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: 8yyhf9py, a7jrmv4x, en4b50ex; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 228.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 8k 228.0 27.0 18.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jes5mq6p",
   "source_url": "https://omlx.ai/benchmarks/performance/jes5mq6p",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-jes5mq6p-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: 8yyhf9py, a7jrmv4x, en4b50ex; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 222.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 222.6 22.4 19.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jes5mq6p",
   "source_url": "https://omlx.ai/benchmarks/performance/jes5mq6p",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wti6l4jz-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: lrax9ull; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 788.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 788.9 45.7 33.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wti6l4jz",
   "source_url": "https://omlx.ai/benchmarks/performance/wti6l4jz",
   "tg_tokens": null,
   "tps": 45.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wti6l4jz-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: lrax9ull; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 873.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 873.3 45.1 34.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wti6l4jz",
   "source_url": "https://omlx.ai/benchmarks/performance/wti6l4jz",
   "tg_tokens": null,
   "tps": 45.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hbrz7wuk-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 74.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2026.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,026 49.5 74.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hbrz7wuk",
   "source_url": "https://omlx.ai/benchmarks/performance/hbrz7wuk",
   "tg_tokens": null,
   "tps": 49.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hbrz7wuk-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 75.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2829.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,829 51.7 75.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hbrz7wuk",
   "source_url": "https://omlx.ai/benchmarks/performance/hbrz7wuk",
   "tg_tokens": null,
   "tps": 51.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ekbkua7w-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 74.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2027.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,027 52.6 74.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ekbkua7w",
   "source_url": "https://omlx.ai/benchmarks/performance/ekbkua7w",
   "tg_tokens": null,
   "tps": 52.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ekbkua7w-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 75.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2790.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,790 53.4 75.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ekbkua7w",
   "source_url": "https://omlx.ai/benchmarks/performance/ekbkua7w",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0nhzbw92-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 82.1 GB; TTFT 1517 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2700.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 2,700 PP tok/s 60.4 TG tok/s 1517 TTFT (ms) 82.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0nhzbw92",
   "source_url": "https://omlx.ai/benchmarks/performance/0nhzbw92",
   "tg_tokens": null,
   "tps": 60.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bbclfarl-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 81.5 GB at this context; duplicate traces of the same session: nqmy17tt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1877.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,877 63.0 81.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bbclfarl",
   "source_url": "https://omlx.ai/benchmarks/performance/bbclfarl",
   "tg_tokens": null,
   "tps": 63.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bbclfarl-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 82.1 GB at this context; duplicate traces of the same session: nqmy17tt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2704.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,704 61.4 82.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bbclfarl",
   "source_url": "https://omlx.ai/benchmarks/performance/bbclfarl",
   "tg_tokens": null,
   "tps": 61.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ll5910jp-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 51.7 GB at this context; duplicate traces of the same session: zy1krhj1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 980.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 980.8 54.6 51.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ll5910jp",
   "source_url": "https://omlx.ai/benchmarks/performance/ll5910jp",
   "tg_tokens": null,
   "tps": 54.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ll5910jp-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 51.9 GB at this context; duplicate traces of the same session: zy1krhj1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1881.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,881 52.2 51.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ll5910jp",
   "source_url": "https://omlx.ai/benchmarks/performance/ll5910jp",
   "tg_tokens": null,
   "tps": 52.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-z2s4iu5e-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 81.5 GB at this context; duplicate traces of the same session: o2iklfp5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1874.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,874 65.1 81.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z2s4iu5e",
   "source_url": "https://omlx.ai/benchmarks/performance/z2s4iu5e",
   "tg_tokens": null,
   "tps": 65.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-z2s4iu5e-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 82.1 GB at this context; duplicate traces of the same session: o2iklfp5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2711.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,711 60.2 82.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z2s4iu5e",
   "source_url": "https://omlx.ai/benchmarks/performance/z2s4iu5e",
   "tg_tokens": null,
   "tps": 60.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tovcauwi-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 82.1 GB; TTFT 1525 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2686.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 2,686 PP tok/s 60.7 TG tok/s 1525 TTFT (ms) 82.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tovcauwi",
   "source_url": "https://omlx.ai/benchmarks/performance/tovcauwi",
   "tg_tokens": null,
   "tps": 60.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q0ejaqrb-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 82.1 GB; TTFT 1518 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2699.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 2,699 PP tok/s 59.8 TG tok/s 1518 TTFT (ms) 82.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q0ejaqrb",
   "source_url": "https://omlx.ai/benchmarks/performance/q0ejaqrb",
   "tg_tokens": null,
   "tps": 59.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i33e4ggh-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 81.5 GB at this context; duplicate traces of the same session: pxlrh8i5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1854.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,854 60.9 81.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i33e4ggh",
   "source_url": "https://omlx.ai/benchmarks/performance/i33e4ggh",
   "tg_tokens": null,
   "tps": 60.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i33e4ggh-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 82.1 GB at this context; duplicate traces of the same session: pxlrh8i5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2692.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,692 55.5 82.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i33e4ggh",
   "source_url": "https://omlx.ai/benchmarks/performance/i33e4ggh",
   "tg_tokens": null,
   "tps": 55.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2pouqcn9-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 81.5 GB at this context; duplicate traces of the same session: p4fvw2gy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1851.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,851 60.3 81.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2pouqcn9",
   "source_url": "https://omlx.ai/benchmarks/performance/2pouqcn9",
   "tg_tokens": null,
   "tps": 60.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2pouqcn9-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 82.1 GB at this context; duplicate traces of the same session: p4fvw2gy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2686.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,686 58.3 82.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2pouqcn9",
   "source_url": "https://omlx.ai/benchmarks/performance/2pouqcn9",
   "tg_tokens": null,
   "tps": 58.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-idvw5t7z-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 81.5 GB; TTFT 552 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1855.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,855 PP tok/s 59.3 TG tok/s 552 TTFT (ms) 81.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/idvw5t7z",
   "source_url": "https://omlx.ai/benchmarks/performance/idvw5t7z",
   "tg_tokens": null,
   "tps": 59.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6n4mxhjg-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 81.5 GB at this context; duplicate traces of the same session: pi352jbl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1843.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,843 59.4 81.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6n4mxhjg",
   "source_url": "https://omlx.ai/benchmarks/performance/6n4mxhjg",
   "tg_tokens": null,
   "tps": 59.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6n4mxhjg-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 82.1 GB at this context; duplicate traces of the same session: pi352jbl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2686.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,686 57.5 82.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6n4mxhjg",
   "source_url": "https://omlx.ai/benchmarks/performance/6n4mxhjg",
   "tg_tokens": null,
   "tps": 57.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7rv5sa0p-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 81.5 GB at this context; duplicate traces of the same session: 7fiebgsy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1850.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,850 60.9 81.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7rv5sa0p",
   "source_url": "https://omlx.ai/benchmarks/performance/7rv5sa0p",
   "tg_tokens": null,
   "tps": 60.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7rv5sa0p-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 82.1 GB at this context; duplicate traces of the same session: 7fiebgsy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2688.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,688 57.8 82.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7rv5sa0p",
   "source_url": "https://omlx.ai/benchmarks/performance/7rv5sa0p",
   "tg_tokens": null,
   "tps": 57.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1sde0rtw-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 82.1 GB; TTFT 1524 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2688.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 2,688 PP tok/s 58.7 TG tok/s 1524 TTFT (ms) 82.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1sde0rtw",
   "source_url": "https://omlx.ai/benchmarks/performance/1sde0rtw",
   "tg_tokens": null,
   "tps": 58.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yqdnvjz4-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 81.5 GB at this context; duplicate traces of the same session: f8v5t4e1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1852.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,852 61.8 81.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yqdnvjz4",
   "source_url": "https://omlx.ai/benchmarks/performance/yqdnvjz4",
   "tg_tokens": null,
   "tps": 61.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yqdnvjz4-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 82.1 GB at this context; duplicate traces of the same session: f8v5t4e1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2689.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,689 61.7 82.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yqdnvjz4",
   "source_url": "https://omlx.ai/benchmarks/performance/yqdnvjz4",
   "tg_tokens": null,
   "tps": 61.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yja2jc83-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 81.5 GB; TTFT 550 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1863.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,863 PP tok/s 63.5 TG tok/s 550 TTFT (ms) 81.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yja2jc83",
   "source_url": "https://omlx.ai/benchmarks/performance/yja2jc83",
   "tg_tokens": null,
   "tps": 63.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-l6yunwsn-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: h36j3f52; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1007.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,007 25.5 22.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l6yunwsn",
   "source_url": "https://omlx.ai/benchmarks/performance/l6yunwsn",
   "tg_tokens": null,
   "tps": 25.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-l6yunwsn-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: h36j3f52; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 978.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 978.6 24.4 23.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l6yunwsn",
   "source_url": "https://omlx.ai/benchmarks/performance/l6yunwsn",
   "tg_tokens": null,
   "tps": 24.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ab6aray1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1004.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,004 25.3 22.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ab6aray1",
   "source_url": "https://omlx.ai/benchmarks/performance/ab6aray1",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ab6aray1-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 979.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 979.9 24.3 23.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ab6aray1",
   "source_url": "https://omlx.ai/benchmarks/performance/ab6aray1",
   "tg_tokens": null,
   "tps": 24.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-dllt8l6t-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 187.5 GB at this context; duplicate traces of the same session: 1eqvawk5, f73ym0ih, jg5xdehf; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 398.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 398.9 48.8 187.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dllt8l6t",
   "source_url": "https://omlx.ai/benchmarks/performance/dllt8l6t",
   "tg_tokens": null,
   "tps": 48.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-dllt8l6t-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 188.6 GB at this context; duplicate traces of the same session: 1eqvawk5, f73ym0ih, jg5xdehf; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 450.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 450.7 46.1 188.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dllt8l6t",
   "source_url": "https://omlx.ai/benchmarks/performance/dllt8l6t",
   "tg_tokens": null,
   "tps": 46.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-dllt8l6t-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 189.0 GB at this context; duplicate traces of the same session: 1eqvawk5, f73ym0ih, jg5xdehf; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 417.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 417.5 59.0 189.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dllt8l6t",
   "source_url": "https://omlx.ai/benchmarks/performance/dllt8l6t",
   "tg_tokens": null,
   "tps": 59.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-dllt8l6t-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 189.1 GB at this context; duplicate traces of the same session: 1eqvawk5, f73ym0ih, jg5xdehf; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 408.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 408.7 85.1 189.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dllt8l6t",
   "source_url": "https://omlx.ai/benchmarks/performance/dllt8l6t",
   "tg_tokens": null,
   "tps": 85.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ct1f4v97-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 187.5 GB at this context; duplicate traces of the same session: 994nafqk, br0x5nim, n5rjfdx9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 343.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 343.5 39.0 187.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ct1f4v97",
   "source_url": "https://omlx.ai/benchmarks/performance/ct1f4v97",
   "tg_tokens": null,
   "tps": 39.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ct1f4v97-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 188.6 GB at this context; duplicate traces of the same session: 994nafqk, br0x5nim, n5rjfdx9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 420.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 420.7 41.4 188.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ct1f4v97",
   "source_url": "https://omlx.ai/benchmarks/performance/ct1f4v97",
   "tg_tokens": null,
   "tps": 41.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ct1f4v97-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 189.0 GB at this context; duplicate traces of the same session: 994nafqk, br0x5nim, n5rjfdx9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 386.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 386.1 45.8 189.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ct1f4v97",
   "source_url": "https://omlx.ai/benchmarks/performance/ct1f4v97",
   "tg_tokens": null,
   "tps": 45.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ct1f4v97-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 189.1 GB at this context; duplicate traces of the same session: 994nafqk, br0x5nim, n5rjfdx9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 401.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 401.1 83.7 189.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ct1f4v97",
   "source_url": "https://omlx.ai/benchmarks/performance/ct1f4v97",
   "tg_tokens": null,
   "tps": 83.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-97o09yt7-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 1152 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 888.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k tokens 888.8 PP tok/s 94.0 TG tok/s 1152 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/97o09yt7",
   "source_url": "https://omlx.ai/benchmarks/performance/97o09yt7",
   "tg_tokens": null,
   "tps": 94.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-w5q5uvpj-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: 19d3q1yo, 7onkd6fu, 8nm99o0t, bhfe86l8, nd3q0zy3, r7sh8ii1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2572.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,572 79.6 20.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w5q5uvpj",
   "source_url": "https://omlx.ai/benchmarks/performance/w5q5uvpj",
   "tg_tokens": null,
   "tps": 79.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-w5q5uvpj-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 19d3q1yo, 7onkd6fu, 8nm99o0t, bhfe86l8, nd3q0zy3, r7sh8ii1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3309.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,309 77.3 21.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w5q5uvpj",
   "source_url": "https://omlx.ai/benchmarks/performance/w5q5uvpj",
   "tg_tokens": null,
   "tps": 77.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-w5q5uvpj-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 19d3q1yo, 7onkd6fu, 8nm99o0t, bhfe86l8, nd3q0zy3, r7sh8ii1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3186.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 3,186 76.3 21.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w5q5uvpj",
   "source_url": "https://omlx.ai/benchmarks/performance/w5q5uvpj",
   "tg_tokens": null,
   "tps": 76.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-w5q5uvpj-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 19d3q1yo, 7onkd6fu, 8nm99o0t, bhfe86l8, nd3q0zy3, r7sh8ii1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2932.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,932 70.9 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w5q5uvpj",
   "source_url": "https://omlx.ai/benchmarks/performance/w5q5uvpj",
   "tg_tokens": null,
   "tps": 70.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-w5q5uvpj-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: 19d3q1yo, 7onkd6fu, 8nm99o0t, bhfe86l8, nd3q0zy3, r7sh8ii1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2530.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 2,530 63.9 21.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w5q5uvpj",
   "source_url": "https://omlx.ai/benchmarks/performance/w5q5uvpj",
   "tg_tokens": null,
   "tps": 63.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-w5q5uvpj-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 19d3q1yo, 7onkd6fu, 8nm99o0t, bhfe86l8, nd3q0zy3, r7sh8ii1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1865.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 1,865 53.1 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w5q5uvpj",
   "source_url": "https://omlx.ai/benchmarks/performance/w5q5uvpj",
   "tg_tokens": null,
   "tps": 53.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-w5q5uvpj-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 19d3q1yo, 7onkd6fu, 8nm99o0t, bhfe86l8, nd3q0zy3, r7sh8ii1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1291.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 128k 1,291 39.9 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w5q5uvpj",
   "source_url": "https://omlx.ai/benchmarks/performance/w5q5uvpj",
   "tg_tokens": null,
   "tps": 39.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-p6tqw7rv-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: 53kdlg0c, k03u2zt2, p7nv4usb, wloialze; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 82.0,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "Qwen3.8-27B-oQ2e-fp16-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 2bit; 1k 82.0 11.6 14.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p6tqw7rv",
   "source_url": "https://omlx.ai/benchmarks/performance/p6tqw7rv",
   "tg_tokens": null,
   "tps": 11.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-p6tqw7rv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 15.6 GB at this context; duplicate traces of the same session: 53kdlg0c, k03u2zt2, p7nv4usb, wloialze; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 85.7,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "Qwen3.8-27B-oQ2e-fp16-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 2bit; 4k 85.7 12.6 15.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p6tqw7rv",
   "source_url": "https://omlx.ai/benchmarks/performance/p6tqw7rv",
   "tg_tokens": null,
   "tps": 12.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-p6tqw7rv-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: 53kdlg0c, k03u2zt2, p7nv4usb, wloialze; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 84.9,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "Qwen3.8-27B-oQ2e-fp16-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 2bit; 8k 84.9 12.2 16.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p6tqw7rv",
   "source_url": "https://omlx.ai/benchmarks/performance/p6tqw7rv",
   "tg_tokens": null,
   "tps": 12.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-p6tqw7rv-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: 53kdlg0c, k03u2zt2, p7nv4usb, wloialze; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 81.2,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "Qwen3.8-27B-oQ2e-fp16-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 2bit; 16k 81.2 12.2 17.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p6tqw7rv",
   "source_url": "https://omlx.ai/benchmarks/performance/p6tqw7rv",
   "tg_tokens": null,
   "tps": 12.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-p6tqw7rv-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 53kdlg0c, k03u2zt2, p7nv4usb, wloialze; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 72.1,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "Qwen3.8-27B-oQ2e-fp16-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 2bit; 32k 72.1 15.7 19.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p6tqw7rv",
   "source_url": "https://omlx.ai/benchmarks/performance/p6tqw7rv",
   "tg_tokens": null,
   "tps": 15.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-57hq4cff-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: c8foq9sp, jp4qneju; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 341.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 341.9 55.5 25.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/57hq4cff",
   "source_url": "https://omlx.ai/benchmarks/performance/57hq4cff",
   "tg_tokens": null,
   "tps": 55.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-57hq4cff-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: c8foq9sp, jp4qneju; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 319.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 319.4 68.3 26.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/57hq4cff",
   "source_url": "https://omlx.ai/benchmarks/performance/57hq4cff",
   "tg_tokens": null,
   "tps": 68.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-57hq4cff-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: c8foq9sp, jp4qneju; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 203.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 203.0 47.2 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/57hq4cff",
   "source_url": "https://omlx.ai/benchmarks/performance/57hq4cff",
   "tg_tokens": null,
   "tps": 47.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-zd73u950-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 3ucfl9tb, y6r5nwfl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 270.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 270.9 27.3 19.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zd73u950",
   "source_url": "https://omlx.ai/benchmarks/performance/zd73u950",
   "tg_tokens": null,
   "tps": 27.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-zd73u950-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 3ucfl9tb, y6r5nwfl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 257.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 257.1 25.2 21.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zd73u950",
   "source_url": "https://omlx.ai/benchmarks/performance/zd73u950",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-zd73u950-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: 3ucfl9tb, y6r5nwfl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 209.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 209.6 21.0 25.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zd73u950",
   "source_url": "https://omlx.ai/benchmarks/performance/zd73u950",
   "tg_tokens": null,
   "tps": 21.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-1h9eudc3-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 40.3 GB; TTFT 181399 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 180.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k tokens 180.6 PP tok/s 30.9 TG tok/s 181399 TTFT (ms) 40.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1h9eudc3",
   "source_url": "https://omlx.ai/benchmarks/performance/1h9eudc3",
   "tg_tokens": null,
   "tps": 30.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n1istcgz-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 37.3 GB; TTFT 287555 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 455.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k tokens 455.8 PP tok/s 11.4 TG tok/s 287555 TTFT (ms) 37.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n1istcgz",
   "source_url": "https://omlx.ai/benchmarks/performance/n1istcgz",
   "tg_tokens": null,
   "tps": 11.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-u1moymd0-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 81.5 GB at this context; duplicate traces of the same session: ho0irfi5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1897.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,897 65.1 81.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u1moymd0",
   "source_url": "https://omlx.ai/benchmarks/performance/u1moymd0",
   "tg_tokens": null,
   "tps": 65.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-u1moymd0-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 82.1 GB at this context; duplicate traces of the same session: ho0irfi5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2717.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,717 63.0 82.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u1moymd0",
   "source_url": "https://omlx.ai/benchmarks/performance/u1moymd0",
   "tg_tokens": null,
   "tps": 63.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-u8p3ln7f-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 81.5 GB at this context; duplicate traces of the same session: x20eftz4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1886.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,886 64.8 81.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u8p3ln7f",
   "source_url": "https://omlx.ai/benchmarks/performance/u8p3ln7f",
   "tg_tokens": null,
   "tps": 64.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-u8p3ln7f-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 82.1 GB at this context; duplicate traces of the same session: x20eftz4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2715.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,715 62.6 82.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u8p3ln7f",
   "source_url": "https://omlx.ai/benchmarks/performance/u8p3ln7f",
   "tg_tokens": null,
   "tps": 62.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7y3zkt1m-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 81.5 GB at this context; duplicate traces of the same session: 3b7y8xv4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1883.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,883 64.2 81.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7y3zkt1m",
   "source_url": "https://omlx.ai/benchmarks/performance/7y3zkt1m",
   "tg_tokens": null,
   "tps": 64.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7y3zkt1m-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 82.1 GB at this context; duplicate traces of the same session: 3b7y8xv4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2713.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,713 63.2 82.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7y3zkt1m",
   "source_url": "https://omlx.ai/benchmarks/performance/7y3zkt1m",
   "tg_tokens": null,
   "tps": 63.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ijmjqf1n-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 81.5 GB at this context; duplicate traces of the same session: puxmpirq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1881.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,881 64.5 81.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ijmjqf1n",
   "source_url": "https://omlx.ai/benchmarks/performance/ijmjqf1n",
   "tg_tokens": null,
   "tps": 64.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ijmjqf1n-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 82.1 GB at this context; duplicate traces of the same session: puxmpirq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2720.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,720 62.4 82.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ijmjqf1n",
   "source_url": "https://omlx.ai/benchmarks/performance/ijmjqf1n",
   "tg_tokens": null,
   "tps": 62.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xxrv3s0i-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 81.5 GB at this context; duplicate traces of the same session: fjrmkfta; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1888.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,888 63.9 81.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xxrv3s0i",
   "source_url": "https://omlx.ai/benchmarks/performance/xxrv3s0i",
   "tg_tokens": null,
   "tps": 63.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xxrv3s0i-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 82.1 GB at this context; duplicate traces of the same session: fjrmkfta; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2718.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,718 61.7 82.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xxrv3s0i",
   "source_url": "https://omlx.ai/benchmarks/performance/xxrv3s0i",
   "tg_tokens": null,
   "tps": 61.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9t8a1hh4-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 82.1 GB; TTFT 1509 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2715.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 2,715 PP tok/s 61.3 TG tok/s 1509 TTFT (ms) 82.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9t8a1hh4",
   "source_url": "https://omlx.ai/benchmarks/performance/9t8a1hh4",
   "tg_tokens": null,
   "tps": 61.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nke0osv9-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 81.5 GB at this context; duplicate traces of the same session: z03okbxh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1875.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,875 65.2 81.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nke0osv9",
   "source_url": "https://omlx.ai/benchmarks/performance/nke0osv9",
   "tg_tokens": null,
   "tps": 65.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nke0osv9-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 82.1 GB at this context; duplicate traces of the same session: z03okbxh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2721.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,721 62.5 82.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nke0osv9",
   "source_url": "https://omlx.ai/benchmarks/performance/nke0osv9",
   "tg_tokens": null,
   "tps": 62.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qr2b2ba9-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 81.5 GB; TTFT 544 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1881.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,881 PP tok/s 63.6 TG tok/s 544 TTFT (ms) 81.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qr2b2ba9",
   "source_url": "https://omlx.ai/benchmarks/performance/qr2b2ba9",
   "tg_tokens": null,
   "tps": 63.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3c4h19si-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 81.5 GB at this context; duplicate traces of the same session: qpbnflpf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1881.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,881 64.2 81.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3c4h19si",
   "source_url": "https://omlx.ai/benchmarks/performance/3c4h19si",
   "tg_tokens": null,
   "tps": 64.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3c4h19si-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 82.1 GB at this context; duplicate traces of the same session: qpbnflpf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2701.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,701 61.7 82.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3c4h19si",
   "source_url": "https://omlx.ai/benchmarks/performance/3c4h19si",
   "tg_tokens": null,
   "tps": 61.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yewy7x7f-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 81.5 GB at this context; duplicate traces of the same session: 4x4bjm6t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1896.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,896 63.9 81.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yewy7x7f",
   "source_url": "https://omlx.ai/benchmarks/performance/yewy7x7f",
   "tg_tokens": null,
   "tps": 63.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yewy7x7f-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 82.1 GB at this context; duplicate traces of the same session: 4x4bjm6t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2706.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,706 61.1 82.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yewy7x7f",
   "source_url": "https://omlx.ai/benchmarks/performance/yewy7x7f",
   "tg_tokens": null,
   "tps": 61.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0br5qxaw-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 81.5 GB at this context; duplicate traces of the same session: makzsklo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1880.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,880 64.4 81.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0br5qxaw",
   "source_url": "https://omlx.ai/benchmarks/performance/0br5qxaw",
   "tg_tokens": null,
   "tps": 64.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0br5qxaw-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 82.1 GB at this context; duplicate traces of the same session: makzsklo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2693.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,693 61.7 82.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0br5qxaw",
   "source_url": "https://omlx.ai/benchmarks/performance/0br5qxaw",
   "tg_tokens": null,
   "tps": 61.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ypsfy19w-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 81.5 GB at this context; duplicate traces of the same session: nz9ex7uz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1878.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,878 64.3 81.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ypsfy19w",
   "source_url": "https://omlx.ai/benchmarks/performance/ypsfy19w",
   "tg_tokens": null,
   "tps": 64.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ypsfy19w-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 82.1 GB at this context; duplicate traces of the same session: nz9ex7uz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2698.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,698 61.5 82.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ypsfy19w",
   "source_url": "https://omlx.ai/benchmarks/performance/ypsfy19w",
   "tg_tokens": null,
   "tps": 61.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jw9153n0-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 81.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1892.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,892 63.2 81.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jw9153n0",
   "source_url": "https://omlx.ai/benchmarks/performance/jw9153n0",
   "tg_tokens": null,
   "tps": 63.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jw9153n0-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 82.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2710.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,710 61.4 82.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jw9153n0",
   "source_url": "https://omlx.ai/benchmarks/performance/jw9153n0",
   "tg_tokens": null,
   "tps": 61.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5yxv0cr6-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 82.1 GB; TTFT 1508 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2716.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 2,716 PP tok/s 62.2 TG tok/s 1508 TTFT (ms) 82.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5yxv0cr6",
   "source_url": "https://omlx.ai/benchmarks/performance/5yxv0cr6",
   "tg_tokens": null,
   "tps": 62.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-70hswy05-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 81.5 GB; TTFT 554 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1847.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,847 PP tok/s 63.6 TG tok/s 554 TTFT (ms) 81.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/70hswy05",
   "source_url": "https://omlx.ai/benchmarks/performance/70hswy05",
   "tg_tokens": null,
   "tps": 63.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mn9nvnz1-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 81.5 GB at this context; duplicate traces of the same session: 87aizvyu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1881.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,881 62.4 81.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mn9nvnz1",
   "source_url": "https://omlx.ai/benchmarks/performance/mn9nvnz1",
   "tg_tokens": null,
   "tps": 62.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mn9nvnz1-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 82.1 GB at this context; duplicate traces of the same session: 87aizvyu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2699.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,699 61.1 82.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mn9nvnz1",
   "source_url": "https://omlx.ai/benchmarks/performance/mn9nvnz1",
   "tg_tokens": null,
   "tps": 61.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qciykgpo-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 81.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1869.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,869 63.8 81.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qciykgpo",
   "source_url": "https://omlx.ai/benchmarks/performance/qciykgpo",
   "tg_tokens": null,
   "tps": 63.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qciykgpo-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 82.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2708.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,708 61.1 82.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qciykgpo",
   "source_url": "https://omlx.ai/benchmarks/performance/qciykgpo",
   "tg_tokens": null,
   "tps": 61.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-hoqiekyf-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.3 GB at this context; duplicate traces of the same session: g1v64b5s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 322.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 322.1 44.5 69.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hoqiekyf",
   "source_url": "https://omlx.ai/benchmarks/performance/hoqiekyf",
   "tg_tokens": null,
   "tps": 44.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-hoqiekyf-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.8 GB at this context; duplicate traces of the same session: g1v64b5s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 370.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 370.6 47.4 69.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hoqiekyf",
   "source_url": "https://omlx.ai/benchmarks/performance/hoqiekyf",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vpgcm3wj-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1004.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,004 25.3 22.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vpgcm3wj",
   "source_url": "https://omlx.ai/benchmarks/performance/vpgcm3wj",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vpgcm3wj-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 979.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 979.9 24.3 23.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vpgcm3wj",
   "source_url": "https://omlx.ai/benchmarks/performance/vpgcm3wj",
   "tg_tokens": null,
   "tps": 24.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-1re83s43-ctx1024",
   "model": "Agents-A1",
   "notes": "mlx build; 4B variant; oMLX community benchmark trace, single run at 1k context; peak memory 6.0 GB; TTFT 1517 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 674.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Agents-A1-4B-mlx-8bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 8bit; 1k tokens 674.9 PP tok/s 47.4 TG tok/s 1517 TTFT (ms) 6.0 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1re83s43",
   "source_url": "https://omlx.ai/benchmarks/performance/1re83s43",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-l00u7djj-ctx1024",
   "model": "Qwen2.5-1.5B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 2.2 GB at this context; duplicate traces of the same session: 41h1dw1y; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 660.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen2.5-1.5B-oQ8e-fp16 on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 8bit; 1k 660.4 30.1 2.2 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l00u7djj",
   "source_url": "https://omlx.ai/benchmarks/performance/l00u7djj",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-l00u7djj-ctx4096",
   "model": "Qwen2.5-1.5B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 2.2 GB at this context; duplicate traces of the same session: 41h1dw1y; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 598.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen2.5-1.5B-oQ8e-fp16 on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 8bit; 4k 598.0 26.9 2.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l00u7djj",
   "source_url": "https://omlx.ai/benchmarks/performance/l00u7djj",
   "tg_tokens": null,
   "tps": 26.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-5c3xhnpt-ctx1024",
   "model": "Qwen2.5-Coder-7B-Instruct",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 4.8 GB; TTFT 244 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4203.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-7B-Instruct-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 4bit; 1k tokens 4,203 PP tok/s 185.9 TG tok/s 244 TTFT (ms) 4.8 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5c3xhnpt",
   "source_url": "https://omlx.ai/benchmarks/performance/5c3xhnpt",
   "tg_tokens": null,
   "tps": 185.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-nq67y7h5-ctx1024",
   "model": "Qwen3.5-35B-A3B-Heretic",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: k63ao0wi, rxrr2mrs, yqt20395, zfthgi54; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4309.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-35B-A3B-Heretic-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 4,309 110.3 35.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nq67y7h5",
   "source_url": "https://omlx.ai/benchmarks/performance/nq67y7h5",
   "tg_tokens": null,
   "tps": 110.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-nq67y7h5-ctx32768",
   "model": "Qwen3.5-35B-A3B-Heretic",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 36.7 GB at this context; duplicate traces of the same session: k63ao0wi, rxrr2mrs, yqt20395, zfthgi54; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4772.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-35B-A3B-Heretic-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 4,772 94.7 36.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nq67y7h5",
   "source_url": "https://omlx.ai/benchmarks/performance/nq67y7h5",
   "tg_tokens": null,
   "tps": 94.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-nq67y7h5-ctx65536",
   "model": "Qwen3.5-35B-A3B-Heretic",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: k63ao0wi, rxrr2mrs, yqt20395, zfthgi54; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3721.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-35B-A3B-Heretic-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 3,721 82.5 37.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nq67y7h5",
   "source_url": "https://omlx.ai/benchmarks/performance/nq67y7h5",
   "tg_tokens": null,
   "tps": 82.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-nq67y7h5-ctx131072",
   "model": "Qwen3.5-35B-A3B-Heretic",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 38.6 GB at this context; duplicate traces of the same session: k63ao0wi, rxrr2mrs, yqt20395, zfthgi54; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2568.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-35B-A3B-Heretic-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 2,568 66.2 38.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nq67y7h5",
   "source_url": "https://omlx.ai/benchmarks/performance/nq67y7h5",
   "tg_tokens": null,
   "tps": 66.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-nq67y7h5-ctx199680",
   "model": "Qwen3.5-35B-A3B-Heretic",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 39.9 GB at this context; duplicate traces of the same session: k63ao0wi, rxrr2mrs, yqt20395, zfthgi54; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1891.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-35B-A3B-Heretic-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 195k 1,891 54.0 39.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nq67y7h5",
   "source_url": "https://omlx.ai/benchmarks/performance/nq67y7h5",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-f5285kwy-ctx1024",
   "model": "Qwen3.5-35B-A3B-Heretic",
   "notes": "MLX build; oMLX community benchmark trace, single run at 1k context; peak memory 35.5 GB; TTFT 250 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4092.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-35B-A3B-Heretic-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k tokens 4,092 PP tok/s 107.0 TG tok/s 250 TTFT (ms) 35.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f5285kwy",
   "source_url": "https://omlx.ai/benchmarks/performance/f5285kwy",
   "tg_tokens": null,
   "tps": 107.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1z0kmj5r-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 1p5wk5gr, 5a104wkt; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1855.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 1,855 111.1 21.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1z0kmj5r",
   "source_url": "https://omlx.ai/benchmarks/performance/1z0kmj5r",
   "tg_tokens": null,
   "tps": 111.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1z0kmj5r-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 1p5wk5gr, 5a104wkt; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2659.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 2,659 159.5 23.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1z0kmj5r",
   "source_url": "https://omlx.ai/benchmarks/performance/1z0kmj5r",
   "tg_tokens": null,
   "tps": 159.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1z0kmj5r-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: 1p5wk5gr, 5a104wkt; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2508.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 2,508 117.8 23.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1z0kmj5r",
   "source_url": "https://omlx.ai/benchmarks/performance/1z0kmj5r",
   "tg_tokens": null,
   "tps": 117.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-w323zpgo-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 30.3 GB; TTFT 19465 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 841.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k tokens 841.7 PP tok/s 17.2 TG tok/s 19465 TTFT (ms) 30.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w323zpgo",
   "source_url": "https://omlx.ai/benchmarks/performance/w323zpgo",
   "tg_tokens": null,
   "tps": 17.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-m1g8h98e-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 35.7 GB; TTFT 19983 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 819.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k tokens 819.9 PP tok/s 58.4 TG tok/s 19983 TTFT (ms) 35.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m1g8h98e",
   "source_url": "https://omlx.ai/benchmarks/performance/m1g8h98e",
   "tg_tokens": null,
   "tps": 58.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2cal7ssg-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 43.5 GB; TTFT 317487 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 412.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k tokens 412.8 PP tok/s 29.3 TG tok/s 317487 TTFT (ms) 43.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2cal7ssg",
   "source_url": "https://omlx.ai/benchmarks/performance/2cal7ssg",
   "tg_tokens": null,
   "tps": 29.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-k41qtom5-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 180.3 GB at this context; duplicate traces of the same session: 5ezw45t3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 759.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-Flash-Next-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 759.0 42.9 180.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k41qtom5",
   "source_url": "https://omlx.ai/benchmarks/performance/k41qtom5",
   "tg_tokens": null,
   "tps": 42.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-k41qtom5-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 182.1 GB at this context; duplicate traces of the same session: 5ezw45t3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 800.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-Flash-Next-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 800.4 38.6 182.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k41qtom5",
   "source_url": "https://omlx.ai/benchmarks/performance/k41qtom5",
   "tg_tokens": null,
   "tps": 38.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kdzwg6k2-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 81.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1869.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,869 63.8 81.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kdzwg6k2",
   "source_url": "https://omlx.ai/benchmarks/performance/kdzwg6k2",
   "tg_tokens": null,
   "tps": 63.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kdzwg6k2-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 82.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2708.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,708 61.1 82.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kdzwg6k2",
   "source_url": "https://omlx.ai/benchmarks/performance/kdzwg6k2",
   "tg_tokens": null,
   "tps": 61.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1lav53d5-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 81.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1892.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,892 63.2 81.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1lav53d5",
   "source_url": "https://omlx.ai/benchmarks/performance/1lav53d5",
   "tg_tokens": null,
   "tps": 63.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1lav53d5-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 82.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2710.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,710 61.4 82.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1lav53d5",
   "source_url": "https://omlx.ai/benchmarks/performance/1lav53d5",
   "tg_tokens": null,
   "tps": 61.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vrh8isst-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 86.6 GB; TTFT 527 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1942.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,942 PP tok/s 40.6 TG tok/s 527 TTFT (ms) 86.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vrh8isst",
   "source_url": "https://omlx.ai/benchmarks/performance/vrh8isst",
   "tg_tokens": null,
   "tps": 40.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1hrcdhy2-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 86.7 GB at this context; duplicate traces of the same session: xixzp5mj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1933.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,933 52.8 86.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1hrcdhy2",
   "source_url": "https://omlx.ai/benchmarks/performance/1hrcdhy2",
   "tg_tokens": null,
   "tps": 52.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1hrcdhy2-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 87.0 GB at this context; duplicate traces of the same session: xixzp5mj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2743.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,743 50.2 87.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1hrcdhy2",
   "source_url": "https://omlx.ai/benchmarks/performance/1hrcdhy2",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-y79o5fz5-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 86.6 GB at this context; duplicate traces of the same session: wfwofg4g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1894.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,894 56.0 86.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y79o5fz5",
   "source_url": "https://omlx.ai/benchmarks/performance/y79o5fz5",
   "tg_tokens": null,
   "tps": 56.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-y79o5fz5-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 87.0 GB at this context; duplicate traces of the same session: wfwofg4g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2748.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,748 46.6 87.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y79o5fz5",
   "source_url": "https://omlx.ai/benchmarks/performance/y79o5fz5",
   "tg_tokens": null,
   "tps": 46.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3jjwoh4n-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 86.8 GB at this context; duplicate traces of the same session: 2u3g48gc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1932.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,932 61.6 86.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3jjwoh4n",
   "source_url": "https://omlx.ai/benchmarks/performance/3jjwoh4n",
   "tg_tokens": null,
   "tps": 61.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3jjwoh4n-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 87.0 GB at this context; duplicate traces of the same session: 2u3g48gc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2751.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,751 35.7 87.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3jjwoh4n",
   "source_url": "https://omlx.ai/benchmarks/performance/3jjwoh4n",
   "tg_tokens": null,
   "tps": 35.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-b6tceoxq-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 87.0 GB; TTFT 1492 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2746.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 2,746 PP tok/s 48.9 TG tok/s 1492 TTFT (ms) 87.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b6tceoxq",
   "source_url": "https://omlx.ai/benchmarks/performance/b6tceoxq",
   "tg_tokens": null,
   "tps": 48.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2s4tabja-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 86.6 GB at this context; duplicate traces of the same session: fm2s424c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1947.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,947 56.4 86.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2s4tabja",
   "source_url": "https://omlx.ai/benchmarks/performance/2s4tabja",
   "tg_tokens": null,
   "tps": 56.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2s4tabja-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 87.0 GB at this context; duplicate traces of the same session: fm2s424c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2744.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,744 51.4 87.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2s4tabja",
   "source_url": "https://omlx.ai/benchmarks/performance/2s4tabja",
   "tg_tokens": null,
   "tps": 51.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jjrkh35w-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 86.8 GB at this context; duplicate traces of the same session: cafwjnv1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1911.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,911 56.7 86.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jjrkh35w",
   "source_url": "https://omlx.ai/benchmarks/performance/jjrkh35w",
   "tg_tokens": null,
   "tps": 56.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jjrkh35w-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 87.0 GB at this context; duplicate traces of the same session: cafwjnv1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2742.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,742 48.6 87.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jjrkh35w",
   "source_url": "https://omlx.ai/benchmarks/performance/jjrkh35w",
   "tg_tokens": null,
   "tps": 48.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-afhssso0-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 86.6 GB at this context; duplicate traces of the same session: 704qrv5y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1910.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,910 46.1 86.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/afhssso0",
   "source_url": "https://omlx.ai/benchmarks/performance/afhssso0",
   "tg_tokens": null,
   "tps": 46.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-afhssso0-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 87.0 GB at this context; duplicate traces of the same session: 704qrv5y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2728.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,728 45.0 87.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/afhssso0",
   "source_url": "https://omlx.ai/benchmarks/performance/afhssso0",
   "tg_tokens": null,
   "tps": 45.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ch0ykkcu-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 86.6 GB; TTFT 533 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1920.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,920 PP tok/s 61.8 TG tok/s 533 TTFT (ms) 86.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ch0ykkcu",
   "source_url": "https://omlx.ai/benchmarks/performance/ch0ykkcu",
   "tg_tokens": null,
   "tps": 61.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-b4zkn2w1-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 86.6 GB at this context; duplicate traces of the same session: ljicsf10; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1921.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,921 40.7 86.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b4zkn2w1",
   "source_url": "https://omlx.ai/benchmarks/performance/b4zkn2w1",
   "tg_tokens": null,
   "tps": 40.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-b4zkn2w1-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 87.0 GB at this context; duplicate traces of the same session: ljicsf10; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2742.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,742 43.9 87.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b4zkn2w1",
   "source_url": "https://omlx.ai/benchmarks/performance/b4zkn2w1",
   "tg_tokens": null,
   "tps": 43.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cvbppy01-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 86.7 GB at this context; duplicate traces of the same session: vqvsijyi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1930.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,930 64.6 86.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cvbppy01",
   "source_url": "https://omlx.ai/benchmarks/performance/cvbppy01",
   "tg_tokens": null,
   "tps": 64.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cvbppy01-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 87.0 GB at this context; duplicate traces of the same session: vqvsijyi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2762.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,762 49.0 87.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cvbppy01",
   "source_url": "https://omlx.ai/benchmarks/performance/cvbppy01",
   "tg_tokens": null,
   "tps": 49.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-uk76cf8e-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 87.0 GB; TTFT 1489 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2752.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 2,752 PP tok/s 45.8 TG tok/s 1489 TTFT (ms) 87.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uk76cf8e",
   "source_url": "https://omlx.ai/benchmarks/performance/uk76cf8e",
   "tg_tokens": null,
   "tps": 45.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-12pz39bz-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 86.7 GB at this context; duplicate traces of the same session: 0fg0iaf5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1924.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,924 57.8 86.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/12pz39bz",
   "source_url": "https://omlx.ai/benchmarks/performance/12pz39bz",
   "tg_tokens": null,
   "tps": 57.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-12pz39bz-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 87.0 GB at this context; duplicate traces of the same session: 0fg0iaf5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2742.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,742 48.8 87.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/12pz39bz",
   "source_url": "https://omlx.ai/benchmarks/performance/12pz39bz",
   "tg_tokens": null,
   "tps": 48.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-h3r2d82f-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 56.8 GB at this context; duplicate traces of the same session: fj0yk7up; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 990.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 990.0 47.9 56.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h3r2d82f",
   "source_url": "https://omlx.ai/benchmarks/performance/h3r2d82f",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-h3r2d82f-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 56.8 GB at this context; duplicate traces of the same session: fj0yk7up; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1921.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,921 45.1 56.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h3r2d82f",
   "source_url": "https://omlx.ai/benchmarks/performance/h3r2d82f",
   "tg_tokens": null,
   "tps": 45.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-u08hm9d6-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 87.0 GB; TTFT 1490 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2748.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 2,748 PP tok/s 49.4 TG tok/s 1490 TTFT (ms) 87.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u08hm9d6",
   "source_url": "https://omlx.ai/benchmarks/performance/u08hm9d6",
   "tg_tokens": null,
   "tps": 49.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-x7rd7wyt-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 86.6 GB at this context; duplicate traces of the same session: 9xdsi5p5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1947.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,947 59.1 86.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x7rd7wyt",
   "source_url": "https://omlx.ai/benchmarks/performance/x7rd7wyt",
   "tg_tokens": null,
   "tps": 59.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-x7rd7wyt-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 87.0 GB at this context; duplicate traces of the same session: 9xdsi5p5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2750.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,750 48.2 87.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x7rd7wyt",
   "source_url": "https://omlx.ai/benchmarks/performance/x7rd7wyt",
   "tg_tokens": null,
   "tps": 48.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lh78pz2m-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 86.7 GB; TTFT 527 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1942.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,942 PP tok/s 50.4 TG tok/s 527 TTFT (ms) 86.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lh78pz2m",
   "source_url": "https://omlx.ai/benchmarks/performance/lh78pz2m",
   "tg_tokens": null,
   "tps": 50.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-02h5baru-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 86.6 GB at this context; duplicate traces of the same session: voq2axpq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1938.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,938 57.0 86.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/02h5baru",
   "source_url": "https://omlx.ai/benchmarks/performance/02h5baru",
   "tg_tokens": null,
   "tps": 57.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-02h5baru-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 87.0 GB at this context; duplicate traces of the same session: voq2axpq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2744.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,744 49.8 87.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/02h5baru",
   "source_url": "https://omlx.ai/benchmarks/performance/02h5baru",
   "tg_tokens": null,
   "tps": 49.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-f6nz8akw-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 87.0 GB; TTFT 1489 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2751.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 2,751 PP tok/s 48.5 TG tok/s 1489 TTFT (ms) 87.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f6nz8akw",
   "source_url": "https://omlx.ai/benchmarks/performance/f6nz8akw",
   "tg_tokens": null,
   "tps": 48.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ec1032tu-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 86.6 GB at this context; duplicate traces of the same session: 8aqv2fg2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1940.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,940 56.0 86.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ec1032tu",
   "source_url": "https://omlx.ai/benchmarks/performance/ec1032tu",
   "tg_tokens": null,
   "tps": 56.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ec1032tu-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 87.0 GB at this context; duplicate traces of the same session: 8aqv2fg2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2748.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,748 50.3 87.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ec1032tu",
   "source_url": "https://omlx.ai/benchmarks/performance/ec1032tu",
   "tg_tokens": null,
   "tps": 50.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-w20sozaz-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 86.6 GB at this context; duplicate traces of the same session: jr2o8e9s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1937.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,937 56.4 86.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w20sozaz",
   "source_url": "https://omlx.ai/benchmarks/performance/w20sozaz",
   "tg_tokens": null,
   "tps": 56.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-w20sozaz-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 87.0 GB at this context; duplicate traces of the same session: jr2o8e9s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2747.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,747 50.2 87.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w20sozaz",
   "source_url": "https://omlx.ai/benchmarks/performance/w20sozaz",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2mmhlpsi-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace, single run at 4k context; peak memory 87.0 GB; TTFT 3071 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1334.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k tokens 1,334 PP tok/s 55.5 TG tok/s 3071 TTFT (ms) 87.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2mmhlpsi",
   "source_url": "https://omlx.ai/benchmarks/performance/2mmhlpsi",
   "tg_tokens": null,
   "tps": 55.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nk4w10d0-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: 2euijsex; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 675.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 675.4 20.2 100.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nk4w10d0",
   "source_url": "https://omlx.ai/benchmarks/performance/nk4w10d0",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nk4w10d0-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: 2euijsex; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1076.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,076 62.5 101.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nk4w10d0",
   "source_url": "https://omlx.ai/benchmarks/performance/nk4w10d0",
   "tg_tokens": null,
   "tps": 62.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-5pravonn-ctx4096",
   "model": "gemma-4-12b-coder-fable5-composer2.5",
   "notes": "community coder fine-tune; oMLX community benchmark trace; peak memory 7.7 GB at this context; duplicate traces of the same session: dd49g8pt; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 170.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12b-coder-fable5-composer2.5-4bit on M3 Pro (14c); M3 Pro (14c) \u00b7 18 GB \u00b7 4bit; 4k 170.1 18.0 7.7 GB view",
   "ram_gb": 18,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5pravonn",
   "source_url": "https://omlx.ai/benchmarks/performance/5pravonn",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-5pravonn-ctx16384",
   "model": "gemma-4-12b-coder-fable5-composer2.5",
   "notes": "community coder fine-tune; oMLX community benchmark trace; peak memory 8.1 GB at this context; duplicate traces of the same session: dd49g8pt; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 159.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12b-coder-fable5-composer2.5-4bit on M3 Pro (14c); M3 Pro (14c) \u00b7 18 GB \u00b7 4bit; 16k 159.3 17.3 8.1 GB current",
   "ram_gb": 18,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5pravonn",
   "source_url": "https://omlx.ai/benchmarks/performance/5pravonn",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-zh5ddh37-ctx4096",
   "model": "Coder390",
   "notes": "27B build; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: j6upneqk; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 138.7,
   "provenance": "community",
   "quant": "oq4e",
   "quote": "Coder390-27B-oq4e on M2 Max (38c); M2 Max (38c) \u00b7 32 GB \u00b7 4bit; 4k 138.7 19.7 17.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zh5ddh37",
   "source_url": "https://omlx.ai/benchmarks/performance/zh5ddh37",
   "tg_tokens": null,
   "tps": 19.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-zh5ddh37-ctx16384",
   "model": "Coder390",
   "notes": "27B build; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: j6upneqk; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 124.5,
   "provenance": "community",
   "quant": "oq4e",
   "quote": "Coder390-27B-oq4e on M2 Max (38c); M2 Max (38c) \u00b7 32 GB \u00b7 4bit; 16k 124.5 18.2 18.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zh5ddh37",
   "source_url": "https://omlx.ai/benchmarks/performance/zh5ddh37",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-2oarvh4y-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: 3zqyp8qu; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 664.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 664.0 79.3 13.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2oarvh4y",
   "source_url": "https://omlx.ai/benchmarks/performance/2oarvh4y",
   "tg_tokens": null,
   "tps": 79.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-2oarvh4y-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: 3zqyp8qu; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 770.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 770.3 75.1 13.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2oarvh4y",
   "source_url": "https://omlx.ai/benchmarks/performance/2oarvh4y",
   "tg_tokens": null,
   "tps": 75.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-9fl20v30-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.1 GB at this context; duplicate traces of the same session: icmloq8l; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1764.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 1,764 75.9 43.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9fl20v30",
   "source_url": "https://omlx.ai/benchmarks/performance/9fl20v30",
   "tg_tokens": null,
   "tps": 75.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-9fl20v30-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.9 GB at this context; duplicate traces of the same session: icmloq8l; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2298.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,298 74.4 43.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9fl20v30",
   "source_url": "https://omlx.ai/benchmarks/performance/9fl20v30",
   "tg_tokens": null,
   "tps": 74.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-incoo4q0-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: iz903wkz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4775.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,775 128.5 19.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/incoo4q0",
   "source_url": "https://omlx.ai/benchmarks/performance/incoo4q0",
   "tg_tokens": null,
   "tps": 128.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-incoo4q0-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: iz903wkz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6448.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,448 122.5 20.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/incoo4q0",
   "source_url": "https://omlx.ai/benchmarks/performance/incoo4q0",
   "tg_tokens": null,
   "tps": 122.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ddtqd5vo-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: ohxpn1wl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4739.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,739 128.4 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ddtqd5vo",
   "source_url": "https://omlx.ai/benchmarks/performance/ddtqd5vo",
   "tg_tokens": null,
   "tps": 128.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ddtqd5vo-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: ohxpn1wl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6462.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,462 118.7 20.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ddtqd5vo",
   "source_url": "https://omlx.ai/benchmarks/performance/ddtqd5vo",
   "tg_tokens": null,
   "tps": 118.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ap20owgh-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 6sykc1hl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4785.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,785 124.4 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ap20owgh",
   "source_url": "https://omlx.ai/benchmarks/performance/ap20owgh",
   "tg_tokens": null,
   "tps": 124.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ap20owgh-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 6sykc1hl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6472.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,472 119.8 20.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ap20owgh",
   "source_url": "https://omlx.ai/benchmarks/performance/ap20owgh",
   "tg_tokens": null,
   "tps": 119.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4qdvniek-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: sf11phjg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4832.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,832 125.2 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4qdvniek",
   "source_url": "https://omlx.ai/benchmarks/performance/4qdvniek",
   "tg_tokens": null,
   "tps": 125.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4qdvniek-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: sf11phjg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6460.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,460 120.9 20.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4qdvniek",
   "source_url": "https://omlx.ai/benchmarks/performance/4qdvniek",
   "tg_tokens": null,
   "tps": 120.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-iu0ct8xe-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: xlhvd3hq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4740.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,740 126.7 19.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iu0ct8xe",
   "source_url": "https://omlx.ai/benchmarks/performance/iu0ct8xe",
   "tg_tokens": null,
   "tps": 126.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-iu0ct8xe-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: xlhvd3hq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6452.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,452 120.9 20.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iu0ct8xe",
   "source_url": "https://omlx.ai/benchmarks/performance/iu0ct8xe",
   "tg_tokens": null,
   "tps": 120.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-otouwp6w-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: eenxmbtm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4819.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,819 128.3 19.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/otouwp6w",
   "source_url": "https://omlx.ai/benchmarks/performance/otouwp6w",
   "tg_tokens": null,
   "tps": 128.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-otouwp6w-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: eenxmbtm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6450.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,450 120.7 20.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/otouwp6w",
   "source_url": "https://omlx.ai/benchmarks/performance/otouwp6w",
   "tg_tokens": null,
   "tps": 120.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-73jfsfjm-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: qoxm54x0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4735.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,735 127.8 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/73jfsfjm",
   "source_url": "https://omlx.ai/benchmarks/performance/73jfsfjm",
   "tg_tokens": null,
   "tps": 127.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-73jfsfjm-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: qoxm54x0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6448.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,448 122.5 20.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/73jfsfjm",
   "source_url": "https://omlx.ai/benchmarks/performance/73jfsfjm",
   "tg_tokens": null,
   "tps": 122.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bi43e779-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: q0ev4k6g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4812.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,812 127.3 19.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bi43e779",
   "source_url": "https://omlx.ai/benchmarks/performance/bi43e779",
   "tg_tokens": null,
   "tps": 127.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bi43e779-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: q0ev4k6g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6449.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,449 121.7 20.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bi43e779",
   "source_url": "https://omlx.ai/benchmarks/performance/bi43e779",
   "tg_tokens": null,
   "tps": 121.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-eak7q0rp-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: hsvg4hjk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4797.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,797 126.1 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eak7q0rp",
   "source_url": "https://omlx.ai/benchmarks/performance/eak7q0rp",
   "tg_tokens": null,
   "tps": 126.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-eak7q0rp-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: hsvg4hjk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6464.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,464 120.9 20.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eak7q0rp",
   "source_url": "https://omlx.ai/benchmarks/performance/eak7q0rp",
   "tg_tokens": null,
   "tps": 120.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2l5ypmva-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: lgqthh6v; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4780.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,780 126.8 19.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2l5ypmva",
   "source_url": "https://omlx.ai/benchmarks/performance/2l5ypmva",
   "tg_tokens": null,
   "tps": 126.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2l5ypmva-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: lgqthh6v; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6462.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,462 118.0 20.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2l5ypmva",
   "source_url": "https://omlx.ai/benchmarks/performance/2l5ypmva",
   "tg_tokens": null,
   "tps": 118.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mce0d5kg-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: lrg8y8vo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4816.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,816 127.9 19.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mce0d5kg",
   "source_url": "https://omlx.ai/benchmarks/performance/mce0d5kg",
   "tg_tokens": null,
   "tps": 127.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mce0d5kg-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: lrg8y8vo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6448.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,448 119.6 20.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mce0d5kg",
   "source_url": "https://omlx.ai/benchmarks/performance/mce0d5kg",
   "tg_tokens": null,
   "tps": 119.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lys0f1vx-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: o7jb5zbs; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4704.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,704 129.8 19.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lys0f1vx",
   "source_url": "https://omlx.ai/benchmarks/performance/lys0f1vx",
   "tg_tokens": null,
   "tps": 129.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lys0f1vx-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: o7jb5zbs; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6419.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,419 121.6 20.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lys0f1vx",
   "source_url": "https://omlx.ai/benchmarks/performance/lys0f1vx",
   "tg_tokens": null,
   "tps": 121.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hs3cg7md-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 4o25wir0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4426.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,426 114.8 19.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hs3cg7md",
   "source_url": "https://omlx.ai/benchmarks/performance/hs3cg7md",
   "tg_tokens": null,
   "tps": 114.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hs3cg7md-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 4o25wir0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5928.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 5,928 110.6 20.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hs3cg7md",
   "source_url": "https://omlx.ai/benchmarks/performance/hs3cg7md",
   "tg_tokens": null,
   "tps": 110.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7m30iv3i-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: whv9trx6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4412.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,412 118.6 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7m30iv3i",
   "source_url": "https://omlx.ai/benchmarks/performance/7m30iv3i",
   "tg_tokens": null,
   "tps": 118.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7m30iv3i-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: whv9trx6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5969.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 5,969 106.7 20.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7m30iv3i",
   "source_url": "https://omlx.ai/benchmarks/performance/7m30iv3i",
   "tg_tokens": null,
   "tps": 106.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-sovggnph-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 39blh0il; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4410.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,410 116.0 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sovggnph",
   "source_url": "https://omlx.ai/benchmarks/performance/sovggnph",
   "tg_tokens": null,
   "tps": 116.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-sovggnph-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 39blh0il; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5908.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 5,908 112.8 20.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sovggnph",
   "source_url": "https://omlx.ai/benchmarks/performance/sovggnph",
   "tg_tokens": null,
   "tps": 112.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-aqvu3qbp-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 1wko7xan; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4440.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,440 114.6 19.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aqvu3qbp",
   "source_url": "https://omlx.ai/benchmarks/performance/aqvu3qbp",
   "tg_tokens": null,
   "tps": 114.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-aqvu3qbp-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 1wko7xan; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5910.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 5,910 108.9 20.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aqvu3qbp",
   "source_url": "https://omlx.ai/benchmarks/performance/aqvu3qbp",
   "tg_tokens": null,
   "tps": 108.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yg3x77e7-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: ro7m89xb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4436.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,436 115.6 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yg3x77e7",
   "source_url": "https://omlx.ai/benchmarks/performance/yg3x77e7",
   "tg_tokens": null,
   "tps": 115.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yg3x77e7-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: ro7m89xb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5940.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 5,940 110.4 20.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yg3x77e7",
   "source_url": "https://omlx.ai/benchmarks/performance/yg3x77e7",
   "tg_tokens": null,
   "tps": 110.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-h03mjods-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 1ufuoigz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4384.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,384 110.3 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h03mjods",
   "source_url": "https://omlx.ai/benchmarks/performance/h03mjods",
   "tg_tokens": null,
   "tps": 110.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-h03mjods-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 1ufuoigz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5937.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 5,937 103.9 20.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h03mjods",
   "source_url": "https://omlx.ai/benchmarks/performance/h03mjods",
   "tg_tokens": null,
   "tps": 103.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5estfvxp-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 4hxi20u1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4297.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,297 108.5 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5estfvxp",
   "source_url": "https://omlx.ai/benchmarks/performance/5estfvxp",
   "tg_tokens": null,
   "tps": 108.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5estfvxp-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 4hxi20u1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5893.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 5,893 104.1 20.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5estfvxp",
   "source_url": "https://omlx.ai/benchmarks/performance/5estfvxp",
   "tg_tokens": null,
   "tps": 104.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-72p1gfw2-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 3bbls42h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4366.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,366 108.8 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/72p1gfw2",
   "source_url": "https://omlx.ai/benchmarks/performance/72p1gfw2",
   "tg_tokens": null,
   "tps": 108.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-72p1gfw2-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 3bbls42h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5895.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 5,895 104.7 20.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/72p1gfw2",
   "source_url": "https://omlx.ai/benchmarks/performance/72p1gfw2",
   "tg_tokens": null,
   "tps": 104.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-sx2hvle7-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: bneiwh0m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4400.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,400 108.6 19.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sx2hvle7",
   "source_url": "https://omlx.ai/benchmarks/performance/sx2hvle7",
   "tg_tokens": null,
   "tps": 108.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-sx2hvle7-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: bneiwh0m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5950.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 5,950 103.4 20.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sx2hvle7",
   "source_url": "https://omlx.ai/benchmarks/performance/sx2hvle7",
   "tg_tokens": null,
   "tps": 103.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bqab1ywl-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: b66faasa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4354.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,354 108.8 19.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bqab1ywl",
   "source_url": "https://omlx.ai/benchmarks/performance/bqab1ywl",
   "tg_tokens": null,
   "tps": 108.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bqab1ywl-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: b66faasa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5896.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 5,896 103.3 20.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bqab1ywl",
   "source_url": "https://omlx.ai/benchmarks/performance/bqab1ywl",
   "tg_tokens": null,
   "tps": 103.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qoye7nz4-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: hkeg1o1b; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 487.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 487.8 18.2 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qoye7nz4",
   "source_url": "https://omlx.ai/benchmarks/performance/qoye7nz4",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qoye7nz4-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: hkeg1o1b; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 517.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 517.6 17.9 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qoye7nz4",
   "source_url": "https://omlx.ai/benchmarks/performance/qoye7nz4",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-kex2jqni-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 19.6 GB; TTFT 18234 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 898.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k tokens 898.5 PP tok/s 85.7 TG tok/s 18234 TTFT (ms) 19.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kex2jqni",
   "source_url": "https://omlx.ai/benchmarks/performance/kex2jqni",
   "tg_tokens": null,
   "tps": 85.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-71772w2w-ctx4096",
   "model": "ThinkingCap-Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 26.9 GB; TTFT 12148 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 337.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "ThinkingCap-Qwen3.8-27B-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 4k tokens 337.2 PP tok/s 38.3 TG tok/s 12148 TTFT (ms) 26.9 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/71772w2w",
   "source_url": "https://omlx.ai/benchmarks/performance/71772w2w",
   "tg_tokens": null,
   "tps": 38.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-1uqldnlb-ctx1024",
   "model": "gemma-4-26B-A4B-heretic-v2-qat-aligned",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 12.0 GB at this context; duplicate traces of the same session: 0hpr3s1g, 17jd60d7; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 468.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-heretic-v2-qat-aligned-4bit-mlx on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 468.4 30.7 12.0 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1uqldnlb",
   "source_url": "https://omlx.ai/benchmarks/performance/1uqldnlb",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-1uqldnlb-ctx4096",
   "model": "gemma-4-26B-A4B-heretic-v2-qat-aligned",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 12.4 GB at this context; duplicate traces of the same session: 0hpr3s1g, 17jd60d7; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 741.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-heretic-v2-qat-aligned-4bit-mlx on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 741.5 29.6 12.4 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1uqldnlb",
   "source_url": "https://omlx.ai/benchmarks/performance/1uqldnlb",
   "tg_tokens": null,
   "tps": 29.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-1uqldnlb-ctx8192",
   "model": "gemma-4-26B-A4B-heretic-v2-qat-aligned",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 12.6 GB at this context; duplicate traces of the same session: 0hpr3s1g, 17jd60d7; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 726.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-heretic-v2-qat-aligned-4bit-mlx on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 8k 726.7 28.4 12.6 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1uqldnlb",
   "source_url": "https://omlx.ai/benchmarks/performance/1uqldnlb",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-4mt9cumh-ctx1024",
   "model": "gemma-4-26B-A4B-heretic-v2-qat-aligned",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 12.6 GB at this context; duplicate traces of the same session: 0eci39sc, 50jvx219; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 456.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-heretic-v2-qat-aligned-4bit-mlx on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 456.8 30.4 12.6 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4mt9cumh",
   "source_url": "https://omlx.ai/benchmarks/performance/4mt9cumh",
   "tg_tokens": null,
   "tps": 30.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-4mt9cumh-ctx4096",
   "model": "gemma-4-26B-A4B-heretic-v2-qat-aligned",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 13.0 GB at this context; duplicate traces of the same session: 0eci39sc, 50jvx219; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 783.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-heretic-v2-qat-aligned-4bit-mlx on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 783.4 30.0 13.0 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4mt9cumh",
   "source_url": "https://omlx.ai/benchmarks/performance/4mt9cumh",
   "tg_tokens": null,
   "tps": 30.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-4mt9cumh-ctx8192",
   "model": "gemma-4-26B-A4B-heretic-v2-qat-aligned",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 13.2 GB at this context; duplicate traces of the same session: 0eci39sc, 50jvx219; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 771.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-heretic-v2-qat-aligned-4bit-mlx on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 8k 771.6 29.4 13.2 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4mt9cumh",
   "source_url": "https://omlx.ai/benchmarks/performance/4mt9cumh",
   "tg_tokens": null,
   "tps": 29.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-gnm2tf4p-ctx1024",
   "model": "LFM2-8B-A1B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 9.8 GB at this context; duplicate traces of the same session: 4llcdg0z, 79z8vvvb, krfz2b6t, qbn65poy, qs3a96mi; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1399.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2-8B-A1B-8bit-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 1k 1,399 111.3 9.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gnm2tf4p",
   "source_url": "https://omlx.ai/benchmarks/performance/gnm2tf4p",
   "tg_tokens": null,
   "tps": 111.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-gnm2tf4p-ctx4096",
   "model": "LFM2-8B-A1B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 10.6 GB at this context; duplicate traces of the same session: 4llcdg0z, 79z8vvvb, krfz2b6t, qbn65poy, qs3a96mi; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1035.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2-8B-A1B-8bit-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 4k 1,035 108.9 10.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gnm2tf4p",
   "source_url": "https://omlx.ai/benchmarks/performance/gnm2tf4p",
   "tg_tokens": null,
   "tps": 108.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-gnm2tf4p-ctx8192",
   "model": "LFM2-8B-A1B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 10.7 GB at this context; duplicate traces of the same session: 4llcdg0z, 79z8vvvb, krfz2b6t, qbn65poy, qs3a96mi; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1173.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2-8B-A1B-8bit-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 8k 1,173 103.2 10.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gnm2tf4p",
   "source_url": "https://omlx.ai/benchmarks/performance/gnm2tf4p",
   "tg_tokens": null,
   "tps": 103.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-gnm2tf4p-ctx16384",
   "model": "LFM2-8B-A1B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 10.8 GB at this context; duplicate traces of the same session: 4llcdg0z, 79z8vvvb, krfz2b6t, qbn65poy, qs3a96mi; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1125.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2-8B-A1B-8bit-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 16k 1,125 93.2 10.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gnm2tf4p",
   "source_url": "https://omlx.ai/benchmarks/performance/gnm2tf4p",
   "tg_tokens": null,
   "tps": 93.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-gnm2tf4p-ctx32768",
   "model": "LFM2-8B-A1B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 11.2 GB at this context; duplicate traces of the same session: 4llcdg0z, 79z8vvvb, krfz2b6t, qbn65poy, qs3a96mi; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 962.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2-8B-A1B-8bit-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 32k 962.1 80.1 11.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gnm2tf4p",
   "source_url": "https://omlx.ai/benchmarks/performance/gnm2tf4p",
   "tg_tokens": null,
   "tps": 80.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-gnm2tf4p-ctx65536",
   "model": "LFM2-8B-A1B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 12.0 GB at this context; duplicate traces of the same session: 4llcdg0z, 79z8vvvb, krfz2b6t, qbn65poy, qs3a96mi; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 860.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2-8B-A1B-8bit-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 64k 860.5 56.3 12.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gnm2tf4p",
   "source_url": "https://omlx.ai/benchmarks/performance/gnm2tf4p",
   "tg_tokens": null,
   "tps": 56.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-y7xywtvk-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX FP16 build; oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: atnbhk6x; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 922.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit-FP16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 922.1 90.5 13.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y7xywtvk",
   "source_url": "https://omlx.ai/benchmarks/performance/y7xywtvk",
   "tg_tokens": null,
   "tps": 90.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-y7xywtvk-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX FP16 build; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: atnbhk6x; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1037.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit-FP16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 1,037 87.7 13.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y7xywtvk",
   "source_url": "https://omlx.ai/benchmarks/performance/y7xywtvk",
   "tg_tokens": null,
   "tps": 87.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-69gapd4i-ctx1024",
   "model": "Ling-3.0-tiny",
   "notes": "fp16; oMLX community benchmark trace; peak memory 5.2 GB at this context; duplicate traces of the same session: ywl1ifej; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1794.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-tiny-oQ4e-fp16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 1,794 113.0 5.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/69gapd4i",
   "source_url": "https://omlx.ai/benchmarks/performance/69gapd4i",
   "tg_tokens": null,
   "tps": 113.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-69gapd4i-ctx4096",
   "model": "Ling-3.0-tiny",
   "notes": "fp16; oMLX community benchmark trace; peak memory 5.7 GB at this context; duplicate traces of the same session: ywl1ifej; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2046.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-tiny-oQ4e-fp16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 2,046 110.5 5.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/69gapd4i",
   "source_url": "https://omlx.ai/benchmarks/performance/69gapd4i",
   "tg_tokens": null,
   "tps": 110.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-60q68s35-ctx1024",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 7.3 GB at this context; duplicate traces of the same session: fe155ln8; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1019.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-oQ4e-fp16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 1,019 89.2 7.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/60q68s35",
   "source_url": "https://omlx.ai/benchmarks/performance/60q68s35",
   "tg_tokens": null,
   "tps": 89.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-60q68s35-ctx4096",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 7.4 GB at this context; duplicate traces of the same session: fe155ln8; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 920.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-oQ4e-fp16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 920.6 86.7 7.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/60q68s35",
   "source_url": "https://omlx.ai/benchmarks/performance/60q68s35",
   "tg_tokens": null,
   "tps": 86.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8hbcmhxt-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: gmq03x6c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4132.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,132 124.4 19.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8hbcmhxt",
   "source_url": "https://omlx.ai/benchmarks/performance/8hbcmhxt",
   "tg_tokens": null,
   "tps": 124.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8hbcmhxt-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: gmq03x6c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6357.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,357 134.8 20.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8hbcmhxt",
   "source_url": "https://omlx.ai/benchmarks/performance/8hbcmhxt",
   "tg_tokens": null,
   "tps": 134.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q48pggt4-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 54dnfvew; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4087.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,087 115.4 19.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q48pggt4",
   "source_url": "https://omlx.ai/benchmarks/performance/q48pggt4",
   "tg_tokens": null,
   "tps": 115.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q48pggt4-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: 54dnfvew; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6354.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,354 110.7 20.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q48pggt4",
   "source_url": "https://omlx.ai/benchmarks/performance/q48pggt4",
   "tg_tokens": null,
   "tps": 110.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-62043xll-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: k2yq9m9a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4122.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,122 122.3 19.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/62043xll",
   "source_url": "https://omlx.ai/benchmarks/performance/62043xll",
   "tg_tokens": null,
   "tps": 122.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-62043xll-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: k2yq9m9a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6322.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,322 140.8 20.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/62043xll",
   "source_url": "https://omlx.ai/benchmarks/performance/62043xll",
   "tg_tokens": null,
   "tps": 140.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9i1lci74-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: y6s6xefm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4140.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,140 129.8 19.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9i1lci74",
   "source_url": "https://omlx.ai/benchmarks/performance/9i1lci74",
   "tg_tokens": null,
   "tps": 129.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9i1lci74-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: y6s6xefm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6341.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,341 135.5 20.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9i1lci74",
   "source_url": "https://omlx.ai/benchmarks/performance/9i1lci74",
   "tg_tokens": null,
   "tps": 135.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nat563og-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 7h7yn4ai; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4139.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,139 117.1 19.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nat563og",
   "source_url": "https://omlx.ai/benchmarks/performance/nat563og",
   "tg_tokens": null,
   "tps": 117.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nat563og-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: 7h7yn4ai; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6399.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,399 132.4 20.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nat563og",
   "source_url": "https://omlx.ai/benchmarks/performance/nat563og",
   "tg_tokens": null,
   "tps": 132.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fhnjo4gl-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: ybg8zlwu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4079.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,079 116.3 19.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fhnjo4gl",
   "source_url": "https://omlx.ai/benchmarks/performance/fhnjo4gl",
   "tg_tokens": null,
   "tps": 116.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fhnjo4gl-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: ybg8zlwu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6344.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,344 134.7 20.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fhnjo4gl",
   "source_url": "https://omlx.ai/benchmarks/performance/fhnjo4gl",
   "tg_tokens": null,
   "tps": 134.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6o036kpr-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: p6m0g9u5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4105.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,105 127.7 19.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6o036kpr",
   "source_url": "https://omlx.ai/benchmarks/performance/6o036kpr",
   "tg_tokens": null,
   "tps": 127.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6o036kpr-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: p6m0g9u5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6310.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,310 130.4 20.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6o036kpr",
   "source_url": "https://omlx.ai/benchmarks/performance/6o036kpr",
   "tg_tokens": null,
   "tps": 130.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-79gu2i1g-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: cd59ovzc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4089.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,089 121.5 19.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/79gu2i1g",
   "source_url": "https://omlx.ai/benchmarks/performance/79gu2i1g",
   "tg_tokens": null,
   "tps": 121.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-79gu2i1g-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: cd59ovzc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6342.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,342 139.3 20.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/79gu2i1g",
   "source_url": "https://omlx.ai/benchmarks/performance/79gu2i1g",
   "tg_tokens": null,
   "tps": 139.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6e985inx-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 1kddgena; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4132.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,132 121.6 19.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6e985inx",
   "source_url": "https://omlx.ai/benchmarks/performance/6e985inx",
   "tg_tokens": null,
   "tps": 121.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6e985inx-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: 1kddgena; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6331.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,331 134.5 20.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6e985inx",
   "source_url": "https://omlx.ai/benchmarks/performance/6e985inx",
   "tg_tokens": null,
   "tps": 134.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-d3ni50yj-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: lkhzjq1y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3978.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 3,978 128.8 19.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d3ni50yj",
   "source_url": "https://omlx.ai/benchmarks/performance/d3ni50yj",
   "tg_tokens": null,
   "tps": 128.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-d3ni50yj-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: lkhzjq1y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6303.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,303 137.8 20.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d3ni50yj",
   "source_url": "https://omlx.ai/benchmarks/performance/d3ni50yj",
   "tg_tokens": null,
   "tps": 137.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kv3satbc-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: nhiqdn5t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4083.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,083 129.9 19.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kv3satbc",
   "source_url": "https://omlx.ai/benchmarks/performance/kv3satbc",
   "tg_tokens": null,
   "tps": 129.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kv3satbc-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: nhiqdn5t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6354.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,354 137.2 20.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kv3satbc",
   "source_url": "https://omlx.ai/benchmarks/performance/kv3satbc",
   "tg_tokens": null,
   "tps": 137.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6d5b1xxb-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: z68jdf6n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4069.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,069 125.7 19.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6d5b1xxb",
   "source_url": "https://omlx.ai/benchmarks/performance/6d5b1xxb",
   "tg_tokens": null,
   "tps": 125.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6d5b1xxb-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: z68jdf6n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6374.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,374 133.2 20.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6d5b1xxb",
   "source_url": "https://omlx.ai/benchmarks/performance/6d5b1xxb",
   "tg_tokens": null,
   "tps": 133.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-p3xdureh-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: zzatgoio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4125.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,125 122.5 19.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p3xdureh",
   "source_url": "https://omlx.ai/benchmarks/performance/p3xdureh",
   "tg_tokens": null,
   "tps": 122.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-p3xdureh-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: zzatgoio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6345.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,345 139.4 20.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p3xdureh",
   "source_url": "https://omlx.ai/benchmarks/performance/p3xdureh",
   "tg_tokens": null,
   "tps": 139.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-13tye0ds-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: g2mtazqn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4134.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,134 121.4 19.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/13tye0ds",
   "source_url": "https://omlx.ai/benchmarks/performance/13tye0ds",
   "tg_tokens": null,
   "tps": 121.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-13tye0ds-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: g2mtazqn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6320.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,320 135.1 20.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/13tye0ds",
   "source_url": "https://omlx.ai/benchmarks/performance/13tye0ds",
   "tg_tokens": null,
   "tps": 135.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-h04mqxmm-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: 5s0qqg0v; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 205.0 21.3 17.2 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h04mqxmm",
   "source_url": "https://omlx.ai/benchmarks/performance/h04mqxmm",
   "tg_tokens": null,
   "tps": 21.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-h04mqxmm-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.9 GB at this context; duplicate traces of the same session: 5s0qqg0v; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 185.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 185.4 19.5 17.9 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h04mqxmm",
   "source_url": "https://omlx.ai/benchmarks/performance/h04mqxmm",
   "tg_tokens": null,
   "tps": 19.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-bqfeqybl-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: z93vkoog; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1384.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 4bit; 1k 1,384 45.2 20.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bqfeqybl",
   "source_url": "https://omlx.ai/benchmarks/performance/bqfeqybl",
   "tg_tokens": null,
   "tps": 45.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-bqfeqybl-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: z93vkoog; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1840.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 4bit; 4k 1,840 44.7 21.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bqfeqybl",
   "source_url": "https://omlx.ai/benchmarks/performance/bqfeqybl",
   "tg_tokens": null,
   "tps": 44.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-583qpjbx-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 28.6 GB; TTFT 20004 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 819.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k tokens 819.0 PP tok/s 62.4 TG tok/s 20004 TTFT (ms) 28.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/583qpjbx",
   "source_url": "https://omlx.ai/benchmarks/performance/583qpjbx",
   "tg_tokens": null,
   "tps": 62.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-043rrnet-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 87.0 GB; TTFT 1491 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2748.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 2,748 PP tok/s 34.1 TG tok/s 1491 TTFT (ms) 87.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/043rrnet",
   "source_url": "https://omlx.ai/benchmarks/performance/043rrnet",
   "tg_tokens": null,
   "tps": 34.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2yz9djga-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 86.8 GB at this context; duplicate traces of the same session: oo7j2c0p; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1922.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,922 45.3 86.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2yz9djga",
   "source_url": "https://omlx.ai/benchmarks/performance/2yz9djga",
   "tg_tokens": null,
   "tps": 45.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2yz9djga-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 87.0 GB at this context; duplicate traces of the same session: oo7j2c0p; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2745.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,745 47.7 87.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2yz9djga",
   "source_url": "https://omlx.ai/benchmarks/performance/2yz9djga",
   "tg_tokens": null,
   "tps": 47.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8yjuydj9-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 86.6 GB at this context; duplicate traces of the same session: 1hqah67k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1942.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,942 39.8 86.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8yjuydj9",
   "source_url": "https://omlx.ai/benchmarks/performance/8yjuydj9",
   "tg_tokens": null,
   "tps": 39.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8yjuydj9-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace; peak memory 87.0 GB at this context; duplicate traces of the same session: 1hqah67k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2747.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,747 36.2 87.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8yjuydj9",
   "source_url": "https://omlx.ai/benchmarks/performance/8yjuydj9",
   "tg_tokens": null,
   "tps": 36.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0y4yhm26-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 87.0 GB; TTFT 1492 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2745.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 2,745 PP tok/s 48.8 TG tok/s 1492 TTFT (ms) 87.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0y4yhm26",
   "source_url": "https://omlx.ai/benchmarks/performance/0y4yhm26",
   "tg_tokens": null,
   "tps": 48.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-p3sep0s9-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 87.0 GB; TTFT 1487 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2756.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 2,756 PP tok/s 41.9 TG tok/s 1487 TTFT (ms) 87.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p3sep0s9",
   "source_url": "https://omlx.ai/benchmarks/performance/p3sep0s9",
   "tg_tokens": null,
   "tps": 41.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9j0v8r2s-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-384",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 87.0 GB; TTFT 1496 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2738.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-384-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 2,738 PP tok/s 45.8 TG tok/s 1496 TTFT (ms) 87.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9j0v8r2s",
   "source_url": "https://omlx.ai/benchmarks/performance/9j0v8r2s",
   "tg_tokens": null,
   "tps": 45.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2hzimd0g-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: ud4lk00n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 166.4,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 1k 166.4 17.0 20.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2hzimd0g",
   "source_url": "https://omlx.ai/benchmarks/performance/2hzimd0g",
   "tg_tokens": null,
   "tps": 17.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2hzimd0g-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: ud4lk00n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 526.5,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 4k 526.5 15.9 21.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2hzimd0g",
   "source_url": "https://omlx.ai/benchmarks/performance/2hzimd0g",
   "tg_tokens": null,
   "tps": 15.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-uf8nsab9-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 34.2 GB at this context; duplicate traces of the same session: jckcrqla, ta4k9g7l, zy2ukoi9; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 109.3,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 1k 109.3 16.5 34.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uf8nsab9",
   "source_url": "https://omlx.ai/benchmarks/performance/uf8nsab9",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-uf8nsab9-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.7 GB at this context; duplicate traces of the same session: jckcrqla, ta4k9g7l, zy2ukoi9; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 110.3,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k 110.3 18.2 37.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uf8nsab9",
   "source_url": "https://omlx.ai/benchmarks/performance/uf8nsab9",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-uf8nsab9-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 38.8 GB at this context; duplicate traces of the same session: jckcrqla, ta4k9g7l, zy2ukoi9; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 109.4,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 8k 109.4 16.6 38.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uf8nsab9",
   "source_url": "https://omlx.ai/benchmarks/performance/uf8nsab9",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-uf8nsab9-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.9 GB at this context; duplicate traces of the same session: jckcrqla, ta4k9g7l, zy2ukoi9; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.4,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 16k 102.4 19.4 37.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uf8nsab9",
   "source_url": "https://omlx.ai/benchmarks/performance/uf8nsab9",
   "tg_tokens": null,
   "tps": 19.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-v1scqdc8-ctx16384",
   "model": "ThinkingCap-Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 28.1 GB; TTFT 50894 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 321.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "ThinkingCap-Qwen3.8-27B-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 16k tokens 321.9 PP tok/s 21.8 TG tok/s 50894 TTFT (ms) 28.1 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v1scqdc8",
   "source_url": "https://omlx.ai/benchmarks/performance/v1scqdc8",
   "tg_tokens": null,
   "tps": 21.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-u0hqb4wt-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace, single run at 16k context; peak memory 21.5 GB; TTFT 98938 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 165.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ4e on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 16k tokens 165.6 PP tok/s 15.0 TG tok/s 98938 TTFT (ms) 21.5 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u0hqb4wt",
   "source_url": "https://omlx.ai/benchmarks/performance/u0hqb4wt",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-ezpdlroe-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: yvov1b97; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 637.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 637.9 77.4 13.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ezpdlroe",
   "source_url": "https://omlx.ai/benchmarks/performance/ezpdlroe",
   "tg_tokens": null,
   "tps": 77.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-ezpdlroe-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: yvov1b97; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 755.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 755.4 71.3 13.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ezpdlroe",
   "source_url": "https://omlx.ai/benchmarks/performance/ezpdlroe",
   "tg_tokens": null,
   "tps": 71.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-o7iguoyv-ctx199680",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "MLX; oMLX community benchmark trace, single run at 195k context; peak memory 77.1 GB; TTFT 329296 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 607.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k tokens 607.4 PP tok/s 50.6 TG tok/s 329296 TTFT (ms) 77.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o7iguoyv",
   "source_url": "https://omlx.ai/benchmarks/performance/o7iguoyv",
   "tg_tokens": null,
   "tps": 50.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qscq1h39-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: pzd5mcff; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4196.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,196 125.0 19.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qscq1h39",
   "source_url": "https://omlx.ai/benchmarks/performance/qscq1h39",
   "tg_tokens": null,
   "tps": 125.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qscq1h39-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: pzd5mcff; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6380.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,380 143.0 20.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qscq1h39",
   "source_url": "https://omlx.ai/benchmarks/performance/qscq1h39",
   "tg_tokens": null,
   "tps": 143.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ltosfsag-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: g2cthpe9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4159.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,159 124.4 19.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ltosfsag",
   "source_url": "https://omlx.ai/benchmarks/performance/ltosfsag",
   "tg_tokens": null,
   "tps": 124.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ltosfsag-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: g2cthpe9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6361.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,361 140.3 20.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ltosfsag",
   "source_url": "https://omlx.ai/benchmarks/performance/ltosfsag",
   "tg_tokens": null,
   "tps": 140.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0ud8kzim-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 8vw58pgp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4129.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,129 116.4 19.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ud8kzim",
   "source_url": "https://omlx.ai/benchmarks/performance/0ud8kzim",
   "tg_tokens": null,
   "tps": 116.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0ud8kzim-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: 8vw58pgp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6392.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,392 133.8 20.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ud8kzim",
   "source_url": "https://omlx.ai/benchmarks/performance/0ud8kzim",
   "tg_tokens": null,
   "tps": 133.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rt8crgq1-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: f916j2kt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4147.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,147 123.7 19.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rt8crgq1",
   "source_url": "https://omlx.ai/benchmarks/performance/rt8crgq1",
   "tg_tokens": null,
   "tps": 123.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-rt8crgq1-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: f916j2kt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6378.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,378 142.6 20.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rt8crgq1",
   "source_url": "https://omlx.ai/benchmarks/performance/rt8crgq1",
   "tg_tokens": null,
   "tps": 142.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5flbrggm-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: pqx2qh9a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4172.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,172 120.5 19.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5flbrggm",
   "source_url": "https://omlx.ai/benchmarks/performance/5flbrggm",
   "tg_tokens": null,
   "tps": 120.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5flbrggm-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: pqx2qh9a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6357.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,357 135.0 20.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5flbrggm",
   "source_url": "https://omlx.ai/benchmarks/performance/5flbrggm",
   "tg_tokens": null,
   "tps": 135.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wo49ac3x-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: xe5ahgiu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4444.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,444 128.0 19.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wo49ac3x",
   "source_url": "https://omlx.ai/benchmarks/performance/wo49ac3x",
   "tg_tokens": null,
   "tps": 128.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wo49ac3x-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: xe5ahgiu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6692.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,692 120.9 20.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wo49ac3x",
   "source_url": "https://omlx.ai/benchmarks/performance/wo49ac3x",
   "tg_tokens": null,
   "tps": 120.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3dhmvac0-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: eeq2ve4y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4370.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,370 126.9 19.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3dhmvac0",
   "source_url": "https://omlx.ai/benchmarks/performance/3dhmvac0",
   "tg_tokens": null,
   "tps": 126.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3dhmvac0-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: eeq2ve4y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6700.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,700 122.9 20.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3dhmvac0",
   "source_url": "https://omlx.ai/benchmarks/performance/3dhmvac0",
   "tg_tokens": null,
   "tps": 122.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wbebo1x7-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: hf6j8d5s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4182.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,182 122.3 19.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wbebo1x7",
   "source_url": "https://omlx.ai/benchmarks/performance/wbebo1x7",
   "tg_tokens": null,
   "tps": 122.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wbebo1x7-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: hf6j8d5s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6371.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,371 139.6 20.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wbebo1x7",
   "source_url": "https://omlx.ai/benchmarks/performance/wbebo1x7",
   "tg_tokens": null,
   "tps": 139.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-260p3sas-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: evs0xx9c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4438.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,438 123.7 19.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/260p3sas",
   "source_url": "https://omlx.ai/benchmarks/performance/260p3sas",
   "tg_tokens": null,
   "tps": 123.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-260p3sas-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: evs0xx9c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6705.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,705 121.5 20.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/260p3sas",
   "source_url": "https://omlx.ai/benchmarks/performance/260p3sas",
   "tg_tokens": null,
   "tps": 121.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3i8mxi2l-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: boh3osnx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4442.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,442 125.1 19.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3i8mxi2l",
   "source_url": "https://omlx.ai/benchmarks/performance/3i8mxi2l",
   "tg_tokens": null,
   "tps": 125.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3i8mxi2l-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: boh3osnx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6667.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,667 121.4 20.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3i8mxi2l",
   "source_url": "https://omlx.ai/benchmarks/performance/3i8mxi2l",
   "tg_tokens": null,
   "tps": 121.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-52fz2ltt-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: pxwqhusj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4798.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,798 128.1 19.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/52fz2ltt",
   "source_url": "https://omlx.ai/benchmarks/performance/52fz2ltt",
   "tg_tokens": null,
   "tps": 128.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-52fz2ltt-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: pxwqhusj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6466.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,466 121.2 20.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/52fz2ltt",
   "source_url": "https://omlx.ai/benchmarks/performance/52fz2ltt",
   "tg_tokens": null,
   "tps": 121.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kq3nkltn-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: dwzda1ft; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4777.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,777 128.0 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kq3nkltn",
   "source_url": "https://omlx.ai/benchmarks/performance/kq3nkltn",
   "tg_tokens": null,
   "tps": 128.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kq3nkltn-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: dwzda1ft; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6465.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,465 121.1 20.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kq3nkltn",
   "source_url": "https://omlx.ai/benchmarks/performance/kq3nkltn",
   "tg_tokens": null,
   "tps": 121.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-brr39gqp-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: bwz0jlvl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4303.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,303 126.3 19.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/brr39gqp",
   "source_url": "https://omlx.ai/benchmarks/performance/brr39gqp",
   "tg_tokens": null,
   "tps": 126.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-brr39gqp-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: bwz0jlvl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6692.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,692 122.4 20.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/brr39gqp",
   "source_url": "https://omlx.ai/benchmarks/performance/brr39gqp",
   "tg_tokens": null,
   "tps": 122.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5tygd5pi-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: zs0sqjhc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4752.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,752 124.9 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5tygd5pi",
   "source_url": "https://omlx.ai/benchmarks/performance/5tygd5pi",
   "tg_tokens": null,
   "tps": 124.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5tygd5pi-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: zs0sqjhc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6478.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,478 121.7 20.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5tygd5pi",
   "source_url": "https://omlx.ai/benchmarks/performance/5tygd5pi",
   "tg_tokens": null,
   "tps": 121.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hsufsp69-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: d3e0f2hz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4764.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,764 129.0 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hsufsp69",
   "source_url": "https://omlx.ai/benchmarks/performance/hsufsp69",
   "tg_tokens": null,
   "tps": 129.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hsufsp69-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: d3e0f2hz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6467.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,467 119.3 20.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hsufsp69",
   "source_url": "https://omlx.ai/benchmarks/performance/hsufsp69",
   "tg_tokens": null,
   "tps": 119.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7ua2cynb-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: ba598pd1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4786.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,786 129.0 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ua2cynb",
   "source_url": "https://omlx.ai/benchmarks/performance/7ua2cynb",
   "tg_tokens": null,
   "tps": 129.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7ua2cynb-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: ba598pd1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6454.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,454 122.1 20.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ua2cynb",
   "source_url": "https://omlx.ai/benchmarks/performance/7ua2cynb",
   "tg_tokens": null,
   "tps": 122.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-w2bf971f-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: bsf6psh1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4787.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,787 130.0 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w2bf971f",
   "source_url": "https://omlx.ai/benchmarks/performance/w2bf971f",
   "tg_tokens": null,
   "tps": 130.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-w2bf971f-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: bsf6psh1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6452.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,452 120.4 20.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w2bf971f",
   "source_url": "https://omlx.ai/benchmarks/performance/w2bf971f",
   "tg_tokens": null,
   "tps": 120.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-4tejqqu7-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 391.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 8bit; 4k 391.1 9.9 29.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4tejqqu7",
   "source_url": "https://omlx.ai/benchmarks/performance/4tejqqu7",
   "tg_tokens": null,
   "tps": 9.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-4tejqqu7-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 370.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 8bit; 16k 370.9 9.6 30.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4tejqqu7",
   "source_url": "https://omlx.ai/benchmarks/performance/4tejqqu7",
   "tg_tokens": null,
   "tps": 9.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-s5yje908-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 270.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 270.2 15.0 17.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s5yje908",
   "source_url": "https://omlx.ai/benchmarks/performance/s5yje908",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-s5yje908-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 267.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 267.6 14.4 17.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s5yje908",
   "source_url": "https://omlx.ai/benchmarks/performance/s5yje908",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-f1m6r830-ctx4096",
   "model": "Qwen3.8-27B-TTURBO-Cold-Fusion-709-L-Uncensored",
   "notes": "mlx 1M context build; oMLX community benchmark trace; peak memory 31.8 GB at this context; duplicate traces of the same session: 941wjx86; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 227.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TTURBO-Cold-Fusion-709-L-Uncensored-mlx-oQ6e-1M on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 6bit; 4k 227.0 24.2 31.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f1m6r830",
   "source_url": "https://omlx.ai/benchmarks/performance/f1m6r830",
   "tg_tokens": null,
   "tps": 24.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-f1m6r830-ctx16384",
   "model": "Qwen3.8-27B-TTURBO-Cold-Fusion-709-L-Uncensored",
   "notes": "mlx 1M context build; oMLX community benchmark trace; peak memory 33.3 GB at this context; duplicate traces of the same session: 941wjx86; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 210.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TTURBO-Cold-Fusion-709-L-Uncensored-mlx-oQ6e-1M on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 6bit; 16k 210.7 26.5 33.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f1m6r830",
   "source_url": "https://omlx.ai/benchmarks/performance/f1m6r830",
   "tg_tokens": null,
   "tps": 26.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-c8yi2t5j-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: 6qx1roto; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 76.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 76.1 18.0 22.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c8yi2t5j",
   "source_url": "https://omlx.ai/benchmarks/performance/c8yi2t5j",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-c8yi2t5j-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: 6qx1roto; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 69.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k 69.4 21.1 24.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c8yi2t5j",
   "source_url": "https://omlx.ai/benchmarks/performance/c8yi2t5j",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4ovaszty-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 9lscskfw, aqfmw0lo, gvos2gfb, jyvf60zu, snqiavol, vaogh18w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1142.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,142 89.7 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ovaszty",
   "source_url": "https://omlx.ai/benchmarks/performance/4ovaszty",
   "tg_tokens": null,
   "tps": 89.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4ovaszty-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 9lscskfw, aqfmw0lo, gvos2gfb, jyvf60zu, snqiavol, vaogh18w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2027.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,027 69.0 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ovaszty",
   "source_url": "https://omlx.ai/benchmarks/performance/4ovaszty",
   "tg_tokens": null,
   "tps": 69.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4ovaszty-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 9lscskfw, aqfmw0lo, gvos2gfb, jyvf60zu, snqiavol, vaogh18w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2444.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,444 76.0 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ovaszty",
   "source_url": "https://omlx.ai/benchmarks/performance/4ovaszty",
   "tg_tokens": null,
   "tps": 76.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4ovaszty-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: 9lscskfw, aqfmw0lo, gvos2gfb, jyvf60zu, snqiavol, vaogh18w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2459.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,459 67.1 73.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ovaszty",
   "source_url": "https://omlx.ai/benchmarks/performance/4ovaszty",
   "tg_tokens": null,
   "tps": 67.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4ovaszty-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.8 GB at this context; duplicate traces of the same session: 9lscskfw, aqfmw0lo, gvos2gfb, jyvf60zu, snqiavol, vaogh18w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2336.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,336 83.4 74.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ovaszty",
   "source_url": "https://omlx.ai/benchmarks/performance/4ovaszty",
   "tg_tokens": null,
   "tps": 83.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4ovaszty-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.9 GB at this context; duplicate traces of the same session: 9lscskfw, aqfmw0lo, gvos2gfb, jyvf60zu, snqiavol, vaogh18w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2352.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,352 66.3 78.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ovaszty",
   "source_url": "https://omlx.ai/benchmarks/performance/4ovaszty",
   "tg_tokens": null,
   "tps": 66.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4ovaszty-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.9 GB at this context; duplicate traces of the same session: 9lscskfw, aqfmw0lo, gvos2gfb, jyvf60zu, snqiavol, vaogh18w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2310.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,310 57.8 80.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ovaszty",
   "source_url": "https://omlx.ai/benchmarks/performance/4ovaszty",
   "tg_tokens": null,
   "tps": 57.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pfvkh6ss-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: 9h4k75ut; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1802.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,802 79.8 71.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pfvkh6ss",
   "source_url": "https://omlx.ai/benchmarks/performance/pfvkh6ss",
   "tg_tokens": null,
   "tps": 79.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-pfvkh6ss-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.8 GB at this context; duplicate traces of the same session: 9h4k75ut; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2423.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,423 82.6 74.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pfvkh6ss",
   "source_url": "https://omlx.ai/benchmarks/performance/pfvkh6ss",
   "tg_tokens": null,
   "tps": 82.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-kouecorn-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.2 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 92.9,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 32k 92.9 17.9 39.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kouecorn",
   "source_url": "https://omlx.ai/benchmarks/performance/kouecorn",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-kouecorn-ctx65536",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 41.3 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 78.4,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 64k 78.4 14.0 41.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kouecorn",
   "source_url": "https://omlx.ai/benchmarks/performance/kouecorn",
   "tg_tokens": null,
   "tps": 14.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-04z30ut8-ctx1024",
   "model": "gemma-4-26B-A4B-heretic-v2-qat-aligned",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 12.6 GB at this context; duplicate traces of the same session: wvd6qjy5, xxkcietg; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 460.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-heretic-v2-qat-aligned-4bit-mlx on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 460.6 30.7 12.6 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/04z30ut8",
   "source_url": "https://omlx.ai/benchmarks/performance/04z30ut8",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-04z30ut8-ctx4096",
   "model": "gemma-4-26B-A4B-heretic-v2-qat-aligned",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 13.0 GB at this context; duplicate traces of the same session: wvd6qjy5, xxkcietg; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 787.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-heretic-v2-qat-aligned-4bit-mlx on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 787.3 29.5 13.0 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/04z30ut8",
   "source_url": "https://omlx.ai/benchmarks/performance/04z30ut8",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-04z30ut8-ctx8192",
   "model": "gemma-4-26B-A4B-heretic-v2-qat-aligned",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 13.2 GB at this context; duplicate traces of the same session: wvd6qjy5, xxkcietg; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 778.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-heretic-v2-qat-aligned-4bit-mlx on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 8k 778.0 28.9 13.2 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/04z30ut8",
   "source_url": "https://omlx.ai/benchmarks/performance/04z30ut8",
   "tg_tokens": null,
   "tps": 28.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-vqf6c8mf-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 6.1 GB at this context; duplicate traces of the same session: djf050b6; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1979.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E4B-it-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 1,979 24.8 6.1 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vqf6c8mf",
   "source_url": "https://omlx.ai/benchmarks/performance/vqf6c8mf",
   "tg_tokens": null,
   "tps": 24.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-vqf6c8mf-ctx16384",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: djf050b6; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1881.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E4B-it-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 16k 1,881 23.4 6.4 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vqf6c8mf",
   "source_url": "https://omlx.ai/benchmarks/performance/vqf6c8mf",
   "tg_tokens": null,
   "tps": 23.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-538tkxno-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 6.0 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1800.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E4B-it-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 1,800 27.0 6.0 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/538tkxno",
   "source_url": "https://omlx.ai/benchmarks/performance/538tkxno",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-538tkxno-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 6.1 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1988.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E4B-it-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 1,988 25.0 6.1 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/538tkxno",
   "source_url": "https://omlx.ai/benchmarks/performance/538tkxno",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ws0hp07g-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: i7zr0hxt, o9n337y6, qh8fo4zz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1688.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 1,688 86.7 24.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ws0hp07g",
   "source_url": "https://omlx.ai/benchmarks/performance/ws0hp07g",
   "tg_tokens": null,
   "tps": 86.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-ws0hp07g-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: i7zr0hxt, o9n337y6, qh8fo4zz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2100.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 2,100 77.4 25.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ws0hp07g",
   "source_url": "https://omlx.ai/benchmarks/performance/ws0hp07g",
   "tg_tokens": null,
   "tps": 77.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ws0hp07g-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: i7zr0hxt, o9n337y6, qh8fo4zz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2074.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 2,074 80.2 25.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ws0hp07g",
   "source_url": "https://omlx.ai/benchmarks/performance/ws0hp07g",
   "tg_tokens": null,
   "tps": 80.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-ws0hp07g-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: i7zr0hxt, o9n337y6, qh8fo4zz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1955.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 1,955 71.2 26.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ws0hp07g",
   "source_url": "https://omlx.ai/benchmarks/performance/ws0hp07g",
   "tg_tokens": null,
   "tps": 71.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ppp1b7jq-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: vv7c91xp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1665.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 1,665 91.5 24.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ppp1b7jq",
   "source_url": "https://omlx.ai/benchmarks/performance/ppp1b7jq",
   "tg_tokens": null,
   "tps": 91.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ppp1b7jq-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: vv7c91xp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2087.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 2,087 91.3 25.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ppp1b7jq",
   "source_url": "https://omlx.ai/benchmarks/performance/ppp1b7jq",
   "tg_tokens": null,
   "tps": 91.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ppp1b7jq-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: vv7c91xp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2044.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 8bit; 8k 2,044 88.3 25.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ppp1b7jq",
   "source_url": "https://omlx.ai/benchmarks/performance/ppp1b7jq",
   "tg_tokens": null,
   "tps": 88.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ppp1b7jq-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: vv7c91xp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1769.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 8bit; 16k 1,769 83.3 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ppp1b7jq",
   "source_url": "https://omlx.ai/benchmarks/performance/ppp1b7jq",
   "tg_tokens": null,
   "tps": 83.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ppp1b7jq-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 28.8 GB at this context; duplicate traces of the same session: vv7c91xp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1381.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 8bit; 32k 1,381 73.7 28.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ppp1b7jq",
   "source_url": "https://omlx.ai/benchmarks/performance/ppp1b7jq",
   "tg_tokens": null,
   "tps": 73.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ppp1b7jq-ctx65536",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 33.2 GB at this context; duplicate traces of the same session: vv7c91xp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1103.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 8bit; 64k 1,103 55.8 33.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ppp1b7jq",
   "source_url": "https://omlx.ai/benchmarks/performance/ppp1b7jq",
   "tg_tokens": null,
   "tps": 55.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ppp1b7jq-ctx131072",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 42.0 GB at this context; duplicate traces of the same session: vv7c91xp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 816.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 8bit; 128k 816.2 29.1 42.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ppp1b7jq",
   "source_url": "https://omlx.ai/benchmarks/performance/ppp1b7jq",
   "tg_tokens": null,
   "tps": 29.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-gruy9j97-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "Preview checkpoint; FP16 build; oMLX community benchmark trace; peak memory 13.3 GB at this context; duplicate traces of the same session: qx2lco9l; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 921.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-Preview-4bit-FP16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 921.9 90.0 13.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gruy9j97",
   "source_url": "https://omlx.ai/benchmarks/performance/gruy9j97",
   "tg_tokens": null,
   "tps": 90.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-gruy9j97-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "Preview checkpoint; FP16 build; oMLX community benchmark trace; peak memory 13.6 GB at this context; duplicate traces of the same session: qx2lco9l; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1034.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-Preview-4bit-FP16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 1,034 89.8 13.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gruy9j97",
   "source_url": "https://omlx.ai/benchmarks/performance/gruy9j97",
   "tg_tokens": null,
   "tps": 89.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-m9mhmsgr-ctx4096",
   "model": "MiniCPM5-1B",
   "notes": "oMLX community benchmark trace; peak memory 1.3 GB at this context; duplicate traces of the same session: 4bq46p60; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 766.6,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "MiniCPM5-1B-oQ4 on M1 (8c); M1 (8c) \u00b7 8 GB \u00b7 4bit; 4k 766.6 69.6 1.3 GB view",
   "ram_gb": 8,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m9mhmsgr",
   "source_url": "https://omlx.ai/benchmarks/performance/m9mhmsgr",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-m9mhmsgr-ctx16384",
   "model": "MiniCPM5-1B",
   "notes": "oMLX community benchmark trace; peak memory 1.5 GB at this context; duplicate traces of the same session: 4bq46p60; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 489.2,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "MiniCPM5-1B-oQ4 on M1 (8c); M1 (8c) \u00b7 8 GB \u00b7 4bit; 16k 489.2 49.1 1.5 GB current",
   "ram_gb": 8,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m9mhmsgr",
   "source_url": "https://omlx.ai/benchmarks/performance/m9mhmsgr",
   "tg_tokens": null,
   "tps": 49.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-mi6luc16-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: k9mn2huk; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 194.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 4bit; 1k 194.0 19.1 5.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mi6luc16",
   "source_url": "https://omlx.ai/benchmarks/performance/mi6luc16",
   "tg_tokens": null,
   "tps": 19.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-mi6luc16-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: k9mn2huk; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 193.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 4bit; 4k 193.7 18.5 6.3 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mi6luc16",
   "source_url": "https://omlx.ai/benchmarks/performance/mi6luc16",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-vwyleoi3-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: xnwqhmnc; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 836.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 836.6 25.9 5.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vwyleoi3",
   "source_url": "https://omlx.ai/benchmarks/performance/vwyleoi3",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-vwyleoi3-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: xnwqhmnc; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 838.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 838.4 25.8 6.2 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vwyleoi3",
   "source_url": "https://omlx.ai/benchmarks/performance/vwyleoi3",
   "tg_tokens": null,
   "tps": 25.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-gp11k8sq-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: ln77ye91; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 948.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 8bit; 1k 948.8 123.8 35.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gp11k8sq",
   "source_url": "https://omlx.ai/benchmarks/performance/gp11k8sq",
   "tg_tokens": null,
   "tps": 123.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-gp11k8sq-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: ln77ye91; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3508.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 8bit; 4k 3,508 121.4 36.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gp11k8sq",
   "source_url": "https://omlx.ai/benchmarks/performance/gp11k8sq",
   "tg_tokens": null,
   "tps": 121.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-iu8yb71o-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: xu4gig0p; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5400.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 5,400 149.9 20.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iu8yb71o",
   "source_url": "https://omlx.ai/benchmarks/performance/iu8yb71o",
   "tg_tokens": null,
   "tps": 149.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-iu8yb71o-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: xu4gig0p; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6479.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,479 134.0 21.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iu8yb71o",
   "source_url": "https://omlx.ai/benchmarks/performance/iu8yb71o",
   "tg_tokens": null,
   "tps": 134.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qy2iw05e-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 5bra9pg2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5300.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 5,300 162.9 20.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qy2iw05e",
   "source_url": "https://omlx.ai/benchmarks/performance/qy2iw05e",
   "tg_tokens": null,
   "tps": 162.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qy2iw05e-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 5bra9pg2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6486.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,486 132.3 21.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qy2iw05e",
   "source_url": "https://omlx.ai/benchmarks/performance/qy2iw05e",
   "tg_tokens": null,
   "tps": 132.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nku5vxew-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: pltd8tjc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5372.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 5,372 148.0 20.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nku5vxew",
   "source_url": "https://omlx.ai/benchmarks/performance/nku5vxew",
   "tg_tokens": null,
   "tps": 148.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-nku5vxew-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: pltd8tjc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6479.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,479 145.6 21.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nku5vxew",
   "source_url": "https://omlx.ai/benchmarks/performance/nku5vxew",
   "tg_tokens": null,
   "tps": 145.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ytijgb84-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: yf6ti1n0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5408.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 5,408 155.2 20.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ytijgb84",
   "source_url": "https://omlx.ai/benchmarks/performance/ytijgb84",
   "tg_tokens": null,
   "tps": 155.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ytijgb84-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: yf6ti1n0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6483.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,483 146.1 21.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ytijgb84",
   "source_url": "https://omlx.ai/benchmarks/performance/ytijgb84",
   "tg_tokens": null,
   "tps": 146.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pzfutqyk-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: rt1mgiep; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5377.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 5,377 152.4 20.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pzfutqyk",
   "source_url": "https://omlx.ai/benchmarks/performance/pzfutqyk",
   "tg_tokens": null,
   "tps": 152.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pzfutqyk-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: rt1mgiep; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6480.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,480 137.0 21.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pzfutqyk",
   "source_url": "https://omlx.ai/benchmarks/performance/pzfutqyk",
   "tg_tokens": null,
   "tps": 137.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3owxm60k-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: 1kr89py0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5373.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 5,373 148.7 20.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3owxm60k",
   "source_url": "https://omlx.ai/benchmarks/performance/3owxm60k",
   "tg_tokens": null,
   "tps": 148.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3owxm60k-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 1kr89py0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6486.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,486 146.2 21.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3owxm60k",
   "source_url": "https://omlx.ai/benchmarks/performance/3owxm60k",
   "tg_tokens": null,
   "tps": 146.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mz0dho44-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: 1669owhx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5387.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 5,387 149.7 20.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mz0dho44",
   "source_url": "https://omlx.ai/benchmarks/performance/mz0dho44",
   "tg_tokens": null,
   "tps": 149.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mz0dho44-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 1669owhx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6494.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,494 157.5 21.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mz0dho44",
   "source_url": "https://omlx.ai/benchmarks/performance/mz0dho44",
   "tg_tokens": null,
   "tps": 157.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-swvn2t01-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: rsgxui2q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5379.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 5,379 154.9 20.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/swvn2t01",
   "source_url": "https://omlx.ai/benchmarks/performance/swvn2t01",
   "tg_tokens": null,
   "tps": 154.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-swvn2t01-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: rsgxui2q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6478.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,478 139.6 21.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/swvn2t01",
   "source_url": "https://omlx.ai/benchmarks/performance/swvn2t01",
   "tg_tokens": null,
   "tps": 139.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1blws921-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: ew8j8h3f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5362.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 5,362 158.8 20.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1blws921",
   "source_url": "https://omlx.ai/benchmarks/performance/1blws921",
   "tg_tokens": null,
   "tps": 158.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1blws921-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: ew8j8h3f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6496.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,496 148.3 21.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1blws921",
   "source_url": "https://omlx.ai/benchmarks/performance/1blws921",
   "tg_tokens": null,
   "tps": 148.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gha9defr-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: 4klpo3wh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5403.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 5,403 165.4 20.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gha9defr",
   "source_url": "https://omlx.ai/benchmarks/performance/gha9defr",
   "tg_tokens": null,
   "tps": 165.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-gha9defr-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 4klpo3wh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6470.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,470 145.3 21.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gha9defr",
   "source_url": "https://omlx.ai/benchmarks/performance/gha9defr",
   "tg_tokens": null,
   "tps": 145.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-u46a2wn0-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: sbkz5ze9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5406.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 5,406 159.1 20.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u46a2wn0",
   "source_url": "https://omlx.ai/benchmarks/performance/u46a2wn0",
   "tg_tokens": null,
   "tps": 159.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-u46a2wn0-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: sbkz5ze9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6466.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,466 133.2 21.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u46a2wn0",
   "source_url": "https://omlx.ai/benchmarks/performance/u46a2wn0",
   "tg_tokens": null,
   "tps": 133.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-iwj8nran-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: clglov6b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5411.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 5,411 154.2 20.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iwj8nran",
   "source_url": "https://omlx.ai/benchmarks/performance/iwj8nran",
   "tg_tokens": null,
   "tps": 154.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-iwj8nran-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: clglov6b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6467.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,467 139.6 21.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iwj8nran",
   "source_url": "https://omlx.ai/benchmarks/performance/iwj8nran",
   "tg_tokens": null,
   "tps": 139.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bd83512i-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: ejdivy5s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5394.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 5,394 165.4 20.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bd83512i",
   "source_url": "https://omlx.ai/benchmarks/performance/bd83512i",
   "tg_tokens": null,
   "tps": 165.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bd83512i-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: ejdivy5s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6479.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,479 146.4 21.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bd83512i",
   "source_url": "https://omlx.ai/benchmarks/performance/bd83512i",
   "tg_tokens": null,
   "tps": 146.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-21ygkoof-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: wu2z9not; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5335.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 5,335 163.2 20.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/21ygkoof",
   "source_url": "https://omlx.ai/benchmarks/performance/21ygkoof",
   "tg_tokens": null,
   "tps": 163.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-21ygkoof-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: wu2z9not; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6462.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,462 130.0 21.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/21ygkoof",
   "source_url": "https://omlx.ai/benchmarks/performance/21ygkoof",
   "tg_tokens": null,
   "tps": 130.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cfly66q5-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 2cp97280; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5361.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 5,361 163.3 20.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cfly66q5",
   "source_url": "https://omlx.ai/benchmarks/performance/cfly66q5",
   "tg_tokens": null,
   "tps": 163.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cfly66q5-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 2cp97280; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6469.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,469 140.4 21.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cfly66q5",
   "source_url": "https://omlx.ai/benchmarks/performance/cfly66q5",
   "tg_tokens": null,
   "tps": 140.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-74e2l8gl-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: l2q4tr8g; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 4k 105.6 17.8 17.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/74e2l8gl",
   "source_url": "https://omlx.ai/benchmarks/performance/74e2l8gl",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-74e2l8gl-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: l2q4tr8g; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 95.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 16k 95.6 16.1 18.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/74e2l8gl",
   "source_url": "https://omlx.ai/benchmarks/performance/74e2l8gl",
   "tg_tokens": null,
   "tps": 16.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-kaj8puu6-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 391.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 8bit; 4k 391.1 9.9 29.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kaj8puu6",
   "source_url": "https://omlx.ai/benchmarks/performance/kaj8puu6",
   "tg_tokens": null,
   "tps": 9.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-kaj8puu6-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 370.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 8bit; 16k 370.9 9.6 30.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kaj8puu6",
   "source_url": "https://omlx.ai/benchmarks/performance/kaj8puu6",
   "tg_tokens": null,
   "tps": 9.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1fl5ltxq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 270.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 270.2 15.0 17.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1fl5ltxq",
   "source_url": "https://omlx.ai/benchmarks/performance/1fl5ltxq",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1fl5ltxq-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 267.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 267.6 14.4 17.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1fl5ltxq",
   "source_url": "https://omlx.ai/benchmarks/performance/1fl5ltxq",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ogxtmaje-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: 7nrn8ym7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 531.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 531.3 34.8 71.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ogxtmaje",
   "source_url": "https://omlx.ai/benchmarks/performance/ogxtmaje",
   "tg_tokens": null,
   "tps": 34.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ogxtmaje-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.5 GB at this context; duplicate traces of the same session: 7nrn8ym7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 952.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 952.5 25.4 74.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ogxtmaje",
   "source_url": "https://omlx.ai/benchmarks/performance/ogxtmaje",
   "tg_tokens": null,
   "tps": 25.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-qwz10ywz-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.2 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 92.9,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 32k 92.9 17.9 39.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qwz10ywz",
   "source_url": "https://omlx.ai/benchmarks/performance/qwz10ywz",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-qwz10ywz-ctx65536",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 41.3 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 78.4,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 64k 78.4 14.0 41.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qwz10ywz",
   "source_url": "https://omlx.ai/benchmarks/performance/qwz10ywz",
   "tg_tokens": null,
   "tps": 14.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-9j61nvom-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 6.0 GB at this context; duplicate traces of the same session: gkxwmao5; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 527.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E4B-it-MLX-4bit on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 4bit; 1k 527.3 20.2 6.0 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9j61nvom",
   "source_url": "https://omlx.ai/benchmarks/performance/9j61nvom",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-9j61nvom-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 6.1 GB at this context; duplicate traces of the same session: gkxwmao5; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 546.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E4B-it-MLX-4bit on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 4bit; 4k 546.2 19.5 6.1 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9j61nvom",
   "source_url": "https://omlx.ai/benchmarks/performance/9j61nvom",
   "tg_tokens": null,
   "tps": 19.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-0dswml39-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 6.0 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1800.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E4B-it-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 1,800 27.0 6.0 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0dswml39",
   "source_url": "https://omlx.ai/benchmarks/performance/0dswml39",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-0dswml39-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 6.1 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1988.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E4B-it-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 1,988 25.0 6.1 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0dswml39",
   "source_url": "https://omlx.ai/benchmarks/performance/0dswml39",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-ib62rxk3-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 6.0 GB at this context; duplicate traces of the same session: 72fl6wko; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1161.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E4B-it-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 1,161 25.8 6.0 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ib62rxk3",
   "source_url": "https://omlx.ai/benchmarks/performance/ib62rxk3",
   "tg_tokens": null,
   "tps": 25.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-ib62rxk3-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 6.1 GB at this context; duplicate traces of the same session: 72fl6wko; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1736.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E4B-it-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 1,736 25.3 6.1 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ib62rxk3",
   "source_url": "https://omlx.ai/benchmarks/performance/ib62rxk3",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-7z4927cv-ctx1024",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 3.0 GB at this context; duplicate traces of the same session: qaeqy7ds; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1436.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e2b-it-4bit on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 4bit; 1k 1,436 61.5 3.0 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7z4927cv",
   "source_url": "https://omlx.ai/benchmarks/performance/7z4927cv",
   "tg_tokens": null,
   "tps": 61.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-7z4927cv-ctx4096",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 3.0 GB at this context; duplicate traces of the same session: qaeqy7ds; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1900.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e2b-it-4bit on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 4bit; 4k 1,900 58.1 3.0 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7z4927cv",
   "source_url": "https://omlx.ai/benchmarks/performance/7z4927cv",
   "tg_tokens": null,
   "tps": 58.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-il5fx1jc-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: 3ds0xanw, 3tn1pq8u, g8wya4zu, slg9xm0z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1665.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 1,665 91.5 24.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/il5fx1jc",
   "source_url": "https://omlx.ai/benchmarks/performance/il5fx1jc",
   "tg_tokens": null,
   "tps": 91.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-il5fx1jc-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: 3ds0xanw, 3tn1pq8u, g8wya4zu, slg9xm0z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2087.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 2,087 91.3 25.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/il5fx1jc",
   "source_url": "https://omlx.ai/benchmarks/performance/il5fx1jc",
   "tg_tokens": null,
   "tps": 91.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-il5fx1jc-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: 3ds0xanw, 3tn1pq8u, g8wya4zu, slg9xm0z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2044.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 8bit; 8k 2,044 88.3 25.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/il5fx1jc",
   "source_url": "https://omlx.ai/benchmarks/performance/il5fx1jc",
   "tg_tokens": null,
   "tps": 88.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-il5fx1jc-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: 3ds0xanw, 3tn1pq8u, g8wya4zu, slg9xm0z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1769.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 8bit; 16k 1,769 83.3 26.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/il5fx1jc",
   "source_url": "https://omlx.ai/benchmarks/performance/il5fx1jc",
   "tg_tokens": null,
   "tps": 83.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-il5fx1jc-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 28.8 GB at this context; duplicate traces of the same session: 3ds0xanw, 3tn1pq8u, g8wya4zu, slg9xm0z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1381.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 8bit; 32k 1,381 73.7 28.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/il5fx1jc",
   "source_url": "https://omlx.ai/benchmarks/performance/il5fx1jc",
   "tg_tokens": null,
   "tps": 73.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-il5fx1jc-ctx65536",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 33.2 GB at this context; duplicate traces of the same session: 3ds0xanw, 3tn1pq8u, g8wya4zu, slg9xm0z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1103.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 8bit; 64k 1,103 55.8 33.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/il5fx1jc",
   "source_url": "https://omlx.ai/benchmarks/performance/il5fx1jc",
   "tg_tokens": null,
   "tps": 55.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-il5fx1jc-ctx131072",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 42.0 GB at this context; duplicate traces of the same session: 3ds0xanw, 3tn1pq8u, g8wya4zu, slg9xm0z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 816.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 8bit; 128k 816.2 29.1 42.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/il5fx1jc",
   "source_url": "https://omlx.ai/benchmarks/performance/il5fx1jc",
   "tg_tokens": null,
   "tps": 29.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-x929d10y-ctx1024",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 6.0 GB at this context; duplicate traces of the same session: lf2lk0x4, mqv3100t, pngpj6sd, qw1zcihm; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 230.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 1k 230.3 34.0 6.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x929d10y",
   "source_url": "https://omlx.ai/benchmarks/performance/x929d10y",
   "tg_tokens": null,
   "tps": 34.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-x929d10y-ctx4096",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: lf2lk0x4, mqv3100t, pngpj6sd, qw1zcihm; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 243.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 4k 243.1 33.3 6.5 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x929d10y",
   "source_url": "https://omlx.ai/benchmarks/performance/x929d10y",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-x929d10y-ctx8192",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: lf2lk0x4, mqv3100t, pngpj6sd, qw1zcihm; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 249.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 8k 249.5 32.3 6.7 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x929d10y",
   "source_url": "https://omlx.ai/benchmarks/performance/x929d10y",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-x929d10y-ctx16384",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 7.2 GB at this context; duplicate traces of the same session: lf2lk0x4, mqv3100t, pngpj6sd, qw1zcihm; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 236.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 16k 236.6 26.5 7.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x929d10y",
   "source_url": "https://omlx.ai/benchmarks/performance/x929d10y",
   "tg_tokens": null,
   "tps": 26.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-x929d10y-ctx32768",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 8.1 GB at this context; duplicate traces of the same session: lf2lk0x4, mqv3100t, pngpj6sd, qw1zcihm; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 214.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 32k 214.6 26.6 8.1 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x929d10y",
   "source_url": "https://omlx.ai/benchmarks/performance/x929d10y",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qany1g2m-ctx4096",
   "model": "Muse-Glimmer-30B-Abliterated",
   "notes": "MLX build; oMLX community benchmark trace, single run at 4k context; peak memory 14.6 GB; TTFT 4060 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1009.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Muse-Glimmer-30B-Abliterated-MLX-mxfp4 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 1,009 PP tok/s 33.5 TG tok/s 4060 TTFT (ms) 14.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qany1g2m",
   "source_url": "https://omlx.ai/benchmarks/performance/qany1g2m",
   "tg_tokens": null,
   "tps": 33.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vd7ysizz-ctx4096",
   "model": "Muse-Glimmer-30B-Abliterated",
   "notes": "MLX build; oMLX community benchmark trace, single run at 4k context; peak memory 14.6 GB; TTFT 4223 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 969.9,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Muse-Glimmer-30B-Abliterated-MLX-mxfp4 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 969.9 PP tok/s 34.3 TG tok/s 4223 TTFT (ms) 14.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vd7ysizz",
   "source_url": "https://omlx.ai/benchmarks/performance/vd7ysizz",
   "tg_tokens": null,
   "tps": 34.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hjm921vt-ctx4096",
   "model": "muse-glimmer 30B",
   "notes": "OptiQ; oMLX community benchmark trace, single run at 4k context; peak memory 19.5 GB; TTFT 4226 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 969.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-30B-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 969.2 PP tok/s 26.5 TG tok/s 4226 TTFT (ms) 19.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hjm921vt",
   "source_url": "https://omlx.ai/benchmarks/performance/hjm921vt",
   "tg_tokens": null,
   "tps": 26.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-punbufk2-ctx1024",
   "model": "NeoHorse-1-9B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 1k context; peak memory 6.6 GB; TTFT 2710 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 377.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NeoHorse-1-9B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k tokens 377.9 PP tok/s 49.1 TG tok/s 2710 TTFT (ms) 6.6 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/punbufk2",
   "source_url": "https://omlx.ai/benchmarks/performance/punbufk2",
   "tg_tokens": null,
   "tps": 49.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-yznzy9g8-ctx1024",
   "model": "NeoHorse-1-9B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 1k context; peak memory 10.0 GB; TTFT 2764 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 370.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "NeoHorse-1-9B-MLX-8bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 8bit; 1k tokens 370.5 PP tok/s 27.7 TG tok/s 2764 TTFT (ms) 10.0 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yznzy9g8",
   "source_url": "https://omlx.ai/benchmarks/performance/yznzy9g8",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-485zk3p1-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 12.3 GB; TTFT 2732 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 374.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B-8bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 8bit; 1k tokens 374.8 PP tok/s 37.8 TG tok/s 2732 TTFT (ms) 12.3 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/485zk3p1",
   "source_url": "https://omlx.ai/benchmarks/performance/485zk3p1",
   "tg_tokens": null,
   "tps": 37.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-8i2z7uu0-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 12.3 GB; TTFT 2748 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 372.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B-8bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 8bit; 1k tokens 372.6 PP tok/s 38.2 TG tok/s 2748 TTFT (ms) 12.3 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8i2z7uu0",
   "source_url": "https://omlx.ai/benchmarks/performance/8i2z7uu0",
   "tg_tokens": null,
   "tps": 38.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-gf90nwlt-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "mlx-lm build; oMLX community benchmark trace, single run at 1k context; peak memory 5.7 GB; TTFT 2732 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 374.8,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.5-9B-mlx-lm-mxfp4 on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k tokens 374.8 PP tok/s 50.1 TG tok/s 2732 TTFT (ms) 5.7 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gf90nwlt",
   "source_url": "https://omlx.ai/benchmarks/performance/gf90nwlt",
   "tg_tokens": null,
   "tps": 50.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-w9v4zbev-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: jd2zslzy; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 800.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 800.0 76.3 22.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w9v4zbev",
   "source_url": "https://omlx.ai/benchmarks/performance/w9v4zbev",
   "tg_tokens": null,
   "tps": 76.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-w9v4zbev-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: jd2zslzy; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 640.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 16k 640.1 66.9 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w9v4zbev",
   "source_url": "https://omlx.ai/benchmarks/performance/w9v4zbev",
   "tg_tokens": null,
   "tps": 66.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-ci41lxh1-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 9s7di8r4; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 765.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 765.7 73.7 23.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ci41lxh1",
   "source_url": "https://omlx.ai/benchmarks/performance/ci41lxh1",
   "tg_tokens": null,
   "tps": 73.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-ci41lxh1-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: 9s7di8r4; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 602.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 16k 602.8 68.1 24.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ci41lxh1",
   "source_url": "https://omlx.ai/benchmarks/performance/ci41lxh1",
   "tg_tokens": null,
   "tps": 68.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-76wunl3m-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.2 GB at this context; duplicate traces of the same session: 8bh6e3mz; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 740.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 6bit; 4k 740.8 65.5 32.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/76wunl3m",
   "source_url": "https://omlx.ai/benchmarks/performance/76wunl3m",
   "tg_tokens": null,
   "tps": 65.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-76wunl3m-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.5 GB at this context; duplicate traces of the same session: 8bh6e3mz; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 595.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 6bit; 16k 595.3 61.0 33.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/76wunl3m",
   "source_url": "https://omlx.ai/benchmarks/performance/76wunl3m",
   "tg_tokens": null,
   "tps": 61.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-u2yxegzc-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: f0dz5qjg, ucdfecs4; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 737.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 6bit; 4k 737.1 61.9 32.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u2yxegzc",
   "source_url": "https://omlx.ai/benchmarks/performance/u2yxegzc",
   "tg_tokens": null,
   "tps": 61.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-u2yxegzc-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.5 GB at this context; duplicate traces of the same session: f0dz5qjg, ucdfecs4; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 593.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 6bit; 16k 593.9 60.6 33.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u2yxegzc",
   "source_url": "https://omlx.ai/benchmarks/performance/u2yxegzc",
   "tg_tokens": null,
   "tps": 60.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-u2yxegzc-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: f0dz5qjg, ucdfecs4; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 460.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 6bit; 32k 460.4 53.5 34.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u2yxegzc",
   "source_url": "https://omlx.ai/benchmarks/performance/u2yxegzc",
   "tg_tokens": null,
   "tps": 53.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-nuanejbt-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.2 GB at this context; duplicate traces of the same session: yzjmd5j7; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 740.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 6bit; 4k 740.1 64.1 32.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nuanejbt",
   "source_url": "https://omlx.ai/benchmarks/performance/nuanejbt",
   "tg_tokens": null,
   "tps": 64.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-nuanejbt-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.5 GB at this context; duplicate traces of the same session: yzjmd5j7; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 595.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 6bit; 16k 595.1 60.8 33.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nuanejbt",
   "source_url": "https://omlx.ai/benchmarks/performance/nuanejbt",
   "tg_tokens": null,
   "tps": 60.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ycuyjda8-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.8 GB at this context; duplicate traces of the same session: pf4ecis6, x9g45os2; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 754.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 6bit; 4k 754.9 68.1 29.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ycuyjda8",
   "source_url": "https://omlx.ai/benchmarks/performance/ycuyjda8",
   "tg_tokens": null,
   "tps": 68.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-ycuyjda8-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.1 GB at this context; duplicate traces of the same session: pf4ecis6, x9g45os2; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 619.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 6bit; 16k 619.8 59.9 31.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ycuyjda8",
   "source_url": "https://omlx.ai/benchmarks/performance/ycuyjda8",
   "tg_tokens": null,
   "tps": 59.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-ycuyjda8-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.4 GB at this context; duplicate traces of the same session: pf4ecis6, x9g45os2; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 487.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 6bit; 32k 487.4 53.1 32.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ycuyjda8",
   "source_url": "https://omlx.ai/benchmarks/performance/ycuyjda8",
   "tg_tokens": null,
   "tps": 53.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6343ht9d-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 09uqyl9b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 921.8,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 921.8 32.1 16.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6343ht9d",
   "source_url": "https://omlx.ai/benchmarks/performance/6343ht9d",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6343ht9d-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 09uqyl9b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 976.8,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 976.8 31.5 17.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6343ht9d",
   "source_url": "https://omlx.ai/benchmarks/performance/6343ht9d",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-xv37qxhp-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: aoca2k0u, ma92686x; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 105.2 20.1 18.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xv37qxhp",
   "source_url": "https://omlx.ai/benchmarks/performance/xv37qxhp",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-xv37qxhp-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: aoca2k0u, ma92686x; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 106.0 18.9 22.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xv37qxhp",
   "source_url": "https://omlx.ai/benchmarks/performance/xv37qxhp",
   "tg_tokens": null,
   "tps": 18.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-xv37qxhp-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: aoca2k0u, ma92686x; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 104.5 16.9 23.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xv37qxhp",
   "source_url": "https://omlx.ai/benchmarks/performance/xv37qxhp",
   "tg_tokens": null,
   "tps": 16.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-79qwzkus-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 6db2xxbk, pkba3g2q; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 106.9 17.4 18.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/79qwzkus",
   "source_url": "https://omlx.ai/benchmarks/performance/79qwzkus",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-79qwzkus-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: 6db2xxbk, pkba3g2q; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 107.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 107.6 17.5 22.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/79qwzkus",
   "source_url": "https://omlx.ai/benchmarks/performance/79qwzkus",
   "tg_tokens": null,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-79qwzkus-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 6db2xxbk, pkba3g2q; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 106.6 16.0 23.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/79qwzkus",
   "source_url": "https://omlx.ai/benchmarks/performance/79qwzkus",
   "tg_tokens": null,
   "tps": 16.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-svwxf0p8-ctx1024",
   "model": "Qwen3.8-9B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 5.6 GB at this context; duplicate traces of the same session: lw1it0n1; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 377.1,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Qwen3.8-9B-MLX-MXFP4 on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 377.1 51.7 5.6 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/svwxf0p8",
   "source_url": "https://omlx.ai/benchmarks/performance/svwxf0p8",
   "tg_tokens": null,
   "tps": 51.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-svwxf0p8-ctx16384",
   "model": "Qwen3.8-9B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 6.6 GB at this context; duplicate traces of the same session: lw1it0n1; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 347.9,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Qwen3.8-9B-MLX-MXFP4 on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 16k 347.9 45.4 6.6 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/svwxf0p8",
   "source_url": "https://omlx.ai/benchmarks/performance/svwxf0p8",
   "tg_tokens": null,
   "tps": 45.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-5qwwi7sa-ctx1024",
   "model": "Qwen3.8-9B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 1k context; peak memory 5.6 GB; TTFT 2717 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 376.9,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Qwen3.8-9B-MLX-MXFP4 on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k tokens 376.9 PP tok/s 51.2 TG tok/s 2717 TTFT (ms) 5.6 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5qwwi7sa",
   "source_url": "https://omlx.ai/benchmarks/performance/5qwwi7sa",
   "tg_tokens": null,
   "tps": 51.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-vas1uhqf-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 86.6 GB; TTFT 67733 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1935.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k tokens 1,935 PP tok/s 59.8 TG tok/s 67733 TTFT (ms) 86.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vas1uhqf",
   "source_url": "https://omlx.ai/benchmarks/performance/vas1uhqf",
   "tg_tokens": null,
   "tps": 59.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hhnzmqh7-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.3 GB at this context; duplicate traces of the same session: rbg0w1sz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1083.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,083 59.2 78.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hhnzmqh7",
   "source_url": "https://omlx.ai/benchmarks/performance/hhnzmqh7",
   "tg_tokens": null,
   "tps": 59.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hhnzmqh7-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.3 GB at this context; duplicate traces of the same session: rbg0w1sz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1767.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,767 66.4 78.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hhnzmqh7",
   "source_url": "https://omlx.ai/benchmarks/performance/hhnzmqh7",
   "tg_tokens": null,
   "tps": 66.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rcuks4so-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.0 GB at this context; duplicate traces of the same session: b0kxcx6a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1095.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,095 63.6 78.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rcuks4so",
   "source_url": "https://omlx.ai/benchmarks/performance/rcuks4so",
   "tg_tokens": null,
   "tps": 63.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rcuks4so-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.0 GB at this context; duplicate traces of the same session: b0kxcx6a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1796.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,796 62.4 78.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rcuks4so",
   "source_url": "https://omlx.ai/benchmarks/performance/rcuks4so",
   "tg_tokens": null,
   "tps": 62.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-55i22uox-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 78.0 GB; TTFT 1034 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 990.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 990.2 PP tok/s 67.5 TG tok/s 1034 TTFT (ms) 78.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/55i22uox",
   "source_url": "https://omlx.ai/benchmarks/performance/55i22uox",
   "tg_tokens": null,
   "tps": 67.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-udk9vs4y-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6 mixed-quant MTP v2 build; oMLX community benchmark trace; peak memory 77.6 GB at this context; duplicate traces of the same session: c003vnc2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 980.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp-v2 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 980.1 60.3 77.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/udk9vs4y",
   "source_url": "https://omlx.ai/benchmarks/performance/udk9vs4y",
   "tg_tokens": null,
   "tps": 60.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-udk9vs4y-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6 mixed-quant MTP v2 build; oMLX community benchmark trace; peak memory 77.8 GB at this context; duplicate traces of the same session: c003vnc2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1846.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp-v2 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 1,846 69.3 77.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/udk9vs4y",
   "source_url": "https://omlx.ai/benchmarks/performance/udk9vs4y",
   "tg_tokens": null,
   "tps": 69.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tqsx5qd4-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 86.5 GB; TTFT 1223 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 837.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k tokens 837.0 PP tok/s 41.7 TG tok/s 1223 TTFT (ms) 86.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tqsx5qd4",
   "source_url": "https://omlx.ai/benchmarks/performance/tqsx5qd4",
   "tg_tokens": null,
   "tps": 41.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-p6ed66jl-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 40.0 GB at this context; duplicate traces of the same session: 2i6nlwcv; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 128.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 32k 128.6 17.4 40.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p6ed66jl",
   "source_url": "https://omlx.ai/benchmarks/performance/p6ed66jl",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-p6ed66jl-ctx65536",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 42.1 GB at this context; duplicate traces of the same session: 2i6nlwcv; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 111.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 64k 111.8 13.8 42.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p6ed66jl",
   "source_url": "https://omlx.ai/benchmarks/performance/p6ed66jl",
   "tg_tokens": null,
   "tps": 13.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fd2zpnd6-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 69.2 GB at this context; duplicate traces of the same session: b5x673z4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 954.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 954.5 64.0 69.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fd2zpnd6",
   "source_url": "https://omlx.ai/benchmarks/performance/fd2zpnd6",
   "tg_tokens": null,
   "tps": 64.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fd2zpnd6-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 69.3 GB at this context; duplicate traces of the same session: b5x673z4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1928.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,928 62.9 69.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fd2zpnd6",
   "source_url": "https://omlx.ai/benchmarks/performance/fd2zpnd6",
   "tg_tokens": null,
   "tps": 62.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-1w006puk-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: ilw1syj0; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 396.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k 396.4 33.8 14.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1w006puk",
   "source_url": "https://omlx.ai/benchmarks/performance/1w006puk",
   "tg_tokens": null,
   "tps": 33.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-1w006puk-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.5 GB at this context; duplicate traces of the same session: ilw1syj0; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 379.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 379.7 32.9 14.5 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1w006puk",
   "source_url": "https://omlx.ai/benchmarks/performance/1w006puk",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-m7aytrkz-ctx16384",
   "model": "Qwen3 14B",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 11.1 GB at this context; duplicate traces of the same session: 31trzrjk; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 678.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3-14B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 16k 678.3 24.7 11.1 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m7aytrkz",
   "source_url": "https://omlx.ai/benchmarks/performance/m7aytrkz",
   "tg_tokens": null,
   "tps": 24.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-m7aytrkz-ctx32768",
   "model": "Qwen3 14B",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 13.3 GB at this context; duplicate traces of the same session: 31trzrjk; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 523.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3-14B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 32k 523.7 19.6 13.3 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m7aytrkz",
   "source_url": "https://omlx.ai/benchmarks/performance/m7aytrkz",
   "tg_tokens": null,
   "tps": 19.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-4ti8ogm0-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.7 GB at this context; duplicate traces of the same session: 33dmogj5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 421.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 421.9 40.9 173.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ti8ogm0",
   "source_url": "https://omlx.ai/benchmarks/performance/4ti8ogm0",
   "tg_tokens": null,
   "tps": 40.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-4ti8ogm0-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.2 GB at this context; duplicate traces of the same session: 33dmogj5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 523.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 523.4 38.8 175.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ti8ogm0",
   "source_url": "https://omlx.ai/benchmarks/performance/4ti8ogm0",
   "tg_tokens": null,
   "tps": 38.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-34s03nvm-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: mgp9vd0w; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 472.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 472.8 46.3 13.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/34s03nvm",
   "source_url": "https://omlx.ai/benchmarks/performance/34s03nvm",
   "tg_tokens": null,
   "tps": 46.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-34s03nvm-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: mgp9vd0w; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 535.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 535.0 53.4 13.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/34s03nvm",
   "source_url": "https://omlx.ai/benchmarks/performance/34s03nvm",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-2usyvv4f-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 1k context; peak memory 24.3 GB; TTFT 1483 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 690.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 1k tokens 690.4 PP tok/s 61.6 TG tok/s 1483 TTFT (ms) 24.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2usyvv4f",
   "source_url": "https://omlx.ai/benchmarks/performance/2usyvv4f",
   "tg_tokens": null,
   "tps": 61.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-s0s0vd6r-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: 1hj7nyw0, 48swx15j, 5phxp24o, e2b0fa27, m2ehy526; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1681.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-24B-A2B-oQ8e-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,681 71.6 24.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s0s0vd6r",
   "source_url": "https://omlx.ai/benchmarks/performance/s0s0vd6r",
   "tg_tokens": null,
   "tps": 71.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-s0s0vd6r-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: 1hj7nyw0, 48swx15j, 5phxp24o, e2b0fa27, m2ehy526; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1916.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-24B-A2B-oQ8e-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 1,916 70.9 24.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s0s0vd6r",
   "source_url": "https://omlx.ai/benchmarks/performance/s0s0vd6r",
   "tg_tokens": null,
   "tps": 70.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-s0s0vd6r-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: 1hj7nyw0, 48swx15j, 5phxp24o, e2b0fa27, m2ehy526; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1839.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-24B-A2B-oQ8e-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 8k 1,839 71.0 24.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s0s0vd6r",
   "source_url": "https://omlx.ai/benchmarks/performance/s0s0vd6r",
   "tg_tokens": null,
   "tps": 71.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-s0s0vd6r-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: 1hj7nyw0, 48swx15j, 5phxp24o, e2b0fa27, m2ehy526; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1723.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-24B-A2B-oQ8e-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 16k 1,723 71.8 25.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s0s0vd6r",
   "source_url": "https://omlx.ai/benchmarks/performance/s0s0vd6r",
   "tg_tokens": null,
   "tps": 71.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-s0s0vd6r-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: 1hj7nyw0, 48swx15j, 5phxp24o, e2b0fa27, m2ehy526; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1516.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-24B-A2B-oQ8e-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 32k 1,516 60.9 25.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s0s0vd6r",
   "source_url": "https://omlx.ai/benchmarks/performance/s0s0vd6r",
   "tg_tokens": null,
   "tps": 60.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-s0s0vd6r-ctx65536",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: 1hj7nyw0, 48swx15j, 5phxp24o, e2b0fa27, m2ehy526; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1201.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-24B-A2B-oQ8e-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 64k 1,201 54.5 25.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s0s0vd6r",
   "source_url": "https://omlx.ai/benchmarks/performance/s0s0vd6r",
   "tg_tokens": null,
   "tps": 54.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qyhyj6cw-ctx4096",
   "model": "Muse-Glimmer-30B-Abliterated",
   "notes": "MLX build; oMLX community benchmark trace, single run at 4k context; peak memory 14.6 GB; TTFT 4065 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1008.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Muse-Glimmer-30B-Abliterated-MLX-mxfp4 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 1,008 PP tok/s 33.5 TG tok/s 4065 TTFT (ms) 14.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qyhyj6cw",
   "source_url": "https://omlx.ai/benchmarks/performance/qyhyj6cw",
   "tg_tokens": null,
   "tps": 33.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-nfxn4yzf-ctx4096",
   "model": "Qwen2.5-Coder-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: hdphzzk6; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 428.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-7B-Instruct-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k 428.5 59.3 4.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nfxn4yzf",
   "source_url": "https://omlx.ai/benchmarks/performance/nfxn4yzf",
   "tg_tokens": null,
   "tps": 59.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-nfxn4yzf-ctx8192",
   "model": "Qwen2.5-Coder-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 5.1 GB at this context; duplicate traces of the same session: hdphzzk6; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 405.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-7B-Instruct-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 8k 405.6 51.1 5.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nfxn4yzf",
   "source_url": "https://omlx.ai/benchmarks/performance/nfxn4yzf",
   "tg_tokens": null,
   "tps": 51.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-fv9kzu45-ctx4096",
   "model": "Qwen2.5-Coder-7B-Instruct",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 4.9 GB; TTFT 9580 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 427.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-7B-Instruct-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k tokens 427.6 PP tok/s 59.4 TG tok/s 9580 TTFT (ms) 4.9 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fv9kzu45",
   "source_url": "https://omlx.ai/benchmarks/performance/fv9kzu45",
   "tg_tokens": null,
   "tps": 59.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-58v9tlmk-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: 8bfbgra2, cpfm6sd1, cz7725vv; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 495.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-9B-oQ4e-fp16-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 495.4 40.4 11.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/58v9tlmk",
   "source_url": "https://omlx.ai/benchmarks/performance/58v9tlmk",
   "tg_tokens": null,
   "tps": 40.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-58v9tlmk-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 12.0 GB at this context; duplicate traces of the same session: 8bfbgra2, cpfm6sd1, cz7725vv; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 476.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-9B-oQ4e-fp16-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 8k 476.9 44.9 12.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/58v9tlmk",
   "source_url": "https://omlx.ai/benchmarks/performance/58v9tlmk",
   "tg_tokens": null,
   "tps": 44.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-58v9tlmk-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 12.5 GB at this context; duplicate traces of the same session: 8bfbgra2, cpfm6sd1, cz7725vv; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 441.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-9B-oQ4e-fp16-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k 441.2 63.1 12.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/58v9tlmk",
   "source_url": "https://omlx.ai/benchmarks/performance/58v9tlmk",
   "tg_tokens": null,
   "tps": 63.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-58v9tlmk-ctx32768",
   "model": "Qwen3.5-9B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 13.5 GB at this context; duplicate traces of the same session: 8bfbgra2, cpfm6sd1, cz7725vv; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 361.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-9B-oQ4e-fp16-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 32k 361.1 91.9 13.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/58v9tlmk",
   "source_url": "https://omlx.ai/benchmarks/performance/58v9tlmk",
   "tg_tokens": null,
   "tps": 91.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-d8vuq2pe-ctx4096",
   "model": "Qwen3.6-35B-A3B-Brainwaves-Nex-Uncensored",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 31.1 GB at this context; duplicate traces of the same session: 5dikr4rx; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1336.0,
   "provenance": "community",
   "quant": "qx64-hi",
   "quote": "Qwen3.6-35B-A3B-Brainwaves-Nex-Uncensored-qx64-hi-mlx on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 6bit; 4k 1,336 68.3 31.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d8vuq2pe",
   "source_url": "https://omlx.ai/benchmarks/performance/d8vuq2pe",
   "tg_tokens": null,
   "tps": 68.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-d8vuq2pe-ctx16384",
   "model": "Qwen3.6-35B-A3B-Brainwaves-Nex-Uncensored",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 31.6 GB at this context; duplicate traces of the same session: 5dikr4rx; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1142.0,
   "provenance": "community",
   "quant": "qx64-hi",
   "quote": "Qwen3.6-35B-A3B-Brainwaves-Nex-Uncensored-qx64-hi-mlx on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 6bit; 16k 1,142 62.8 31.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d8vuq2pe",
   "source_url": "https://omlx.ai/benchmarks/performance/d8vuq2pe",
   "tg_tokens": null,
   "tps": 62.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-wq8q97nu-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 17.3 GB; TTFT 51466 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 79.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k tokens 79.6 PP tok/s 15.2 TG tok/s 51466 TTFT (ms) 17.3 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wq8q97nu",
   "source_url": "https://omlx.ai/benchmarks/performance/wq8q97nu",
   "tg_tokens": null,
   "tps": 15.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-y0ton3yk-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 16.0 GB; TTFT 12361 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 82.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 1k tokens 82.8 PP tok/s 15.6 TG tok/s 12361 TTFT (ms) 16.0 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y0ton3yk",
   "source_url": "https://omlx.ai/benchmarks/performance/y0ton3yk",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-0uejmvv8-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 17.3 GB; TTFT 52128 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 78.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k tokens 78.6 PP tok/s 15.1 TG tok/s 52128 TTFT (ms) 17.3 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0uejmvv8",
   "source_url": "https://omlx.ai/benchmarks/performance/0uejmvv8",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-cqx2fwc6-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: bsau72vi; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 76.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 76.9 10.1 16.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cqx2fwc6",
   "source_url": "https://omlx.ai/benchmarks/performance/cqx2fwc6",
   "tg_tokens": null,
   "tps": 10.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-cqx2fwc6-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: bsau72vi; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 77.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 77.7 9.6 17.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cqx2fwc6",
   "source_url": "https://omlx.ai/benchmarks/performance/cqx2fwc6",
   "tg_tokens": null,
   "tps": 9.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-175brvyp-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: b0gc1jvu; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 75.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 75.5 9.8 17.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/175brvyp",
   "source_url": "https://omlx.ai/benchmarks/performance/175brvyp",
   "tg_tokens": null,
   "tps": 9.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-175brvyp-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: b0gc1jvu; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 73.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 73.6 8.4 18.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/175brvyp",
   "source_url": "https://omlx.ai/benchmarks/performance/175brvyp",
   "tg_tokens": null,
   "tps": 8.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-fp2wei9c-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: g6jpuvjz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 225.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 225.4 22.5 16.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fp2wei9c",
   "source_url": "https://omlx.ai/benchmarks/performance/fp2wei9c",
   "tg_tokens": null,
   "tps": 22.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-fp2wei9c-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: g6jpuvjz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 236.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 236.9 21.9 19.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fp2wei9c",
   "source_url": "https://omlx.ai/benchmarks/performance/fp2wei9c",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-t53uibil-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: tg8pj2k5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 378.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 378.3 17.6 15.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t53uibil",
   "source_url": "https://omlx.ai/benchmarks/performance/t53uibil",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-t53uibil-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: tg8pj2k5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 400.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 400.2 17.7 17.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t53uibil",
   "source_url": "https://omlx.ai/benchmarks/performance/t53uibil",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-71wr10qv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: e379f12a; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 338.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 6bit; 4k 338.9 12.2 23.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/71wr10qv",
   "source_url": "https://omlx.ai/benchmarks/performance/71wr10qv",
   "tg_tokens": null,
   "tps": 12.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-71wr10qv-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: e379f12a; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 348.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 6bit; 8k 348.6 11.5 23.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/71wr10qv",
   "source_url": "https://omlx.ai/benchmarks/performance/71wr10qv",
   "tg_tokens": null,
   "tps": 11.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-rfy73ito-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "oMLX community benchmark trace; peak memory 29.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 118.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 1k 118.6 19.7 29.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rfy73ito",
   "source_url": "https://omlx.ai/benchmarks/performance/rfy73ito",
   "tg_tokens": null,
   "tps": 19.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-rfy73ito-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "oMLX community benchmark trace; peak memory 30.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 121.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 4k 121.2 14.7 30.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rfy73ito",
   "source_url": "https://omlx.ai/benchmarks/performance/rfy73ito",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-en63q0xb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: i5grwvj7, msrz3qoy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 838.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 838.6 24.1 22.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/en63q0xb",
   "source_url": "https://omlx.ai/benchmarks/performance/en63q0xb",
   "tg_tokens": null,
   "tps": 24.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-en63q0xb-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: i5grwvj7, msrz3qoy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 900.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 900.9 23.4 23.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/en63q0xb",
   "source_url": "https://omlx.ai/benchmarks/performance/en63q0xb",
   "tg_tokens": null,
   "tps": 23.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-en63q0xb-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: i5grwvj7, msrz3qoy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 868.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 868.7 21.7 25.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/en63q0xb",
   "source_url": "https://omlx.ai/benchmarks/performance/en63q0xb",
   "tg_tokens": null,
   "tps": 21.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pmfj2djb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: 9se1f28b, anrzig42; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 816.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 816.3 40.2 33.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pmfj2djb",
   "source_url": "https://omlx.ai/benchmarks/performance/pmfj2djb",
   "tg_tokens": null,
   "tps": 40.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pmfj2djb-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: 9se1f28b, anrzig42; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 892.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 8k 892.7 35.7 35.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pmfj2djb",
   "source_url": "https://omlx.ai/benchmarks/performance/pmfj2djb",
   "tg_tokens": null,
   "tps": 35.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pmfj2djb-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.8 GB at this context; duplicate traces of the same session: 9se1f28b, anrzig42; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 718.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 32k 718.2 40.9 36.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pmfj2djb",
   "source_url": "https://omlx.ai/benchmarks/performance/pmfj2djb",
   "tg_tokens": null,
   "tps": 40.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-z4dl9i6h-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: 30aafupw, 8xkxppnv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 811.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 811.7 48.7 33.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z4dl9i6h",
   "source_url": "https://omlx.ai/benchmarks/performance/z4dl9i6h",
   "tg_tokens": null,
   "tps": 48.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-z4dl9i6h-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: 30aafupw, 8xkxppnv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 887.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 8k 887.7 28.2 35.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z4dl9i6h",
   "source_url": "https://omlx.ai/benchmarks/performance/z4dl9i6h",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-z4dl9i6h-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.8 GB at this context; duplicate traces of the same session: 30aafupw, 8xkxppnv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 762.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 32k 762.1 59.7 36.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z4dl9i6h",
   "source_url": "https://omlx.ai/benchmarks/performance/z4dl9i6h",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-m7eqam3c-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: q02e08pz, t1ac0f8w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 810.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 810.2 74.2 33.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m7eqam3c",
   "source_url": "https://omlx.ai/benchmarks/performance/m7eqam3c",
   "tg_tokens": null,
   "tps": 74.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-m7eqam3c-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: q02e08pz, t1ac0f8w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 894.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 8k 894.8 29.5 35.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m7eqam3c",
   "source_url": "https://omlx.ai/benchmarks/performance/m7eqam3c",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-m7eqam3c-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.8 GB at this context; duplicate traces of the same session: q02e08pz, t1ac0f8w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 717.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 32k 717.6 39.2 36.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m7eqam3c",
   "source_url": "https://omlx.ai/benchmarks/performance/m7eqam3c",
   "tg_tokens": null,
   "tps": 39.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-eu6w0wnq-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP320",
   "notes": "MTP Vision build; oMLX community benchmark trace, single run at 1k context; peak memory 22.1 GB; TTFT 10703 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 95.7,
   "provenance": "community",
   "quant": "oQ3e-DWQ",
   "quote": "Qwen3.8-Flash-Next-REAP320-oQ3e-DWQ-MTP-Vision-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 3bit; 1k tokens 95.7 PP tok/s 12.0 TG tok/s 10703 TTFT (ms) 22.1 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eu6w0wnq",
   "source_url": "https://omlx.ai/benchmarks/performance/eu6w0wnq",
   "tg_tokens": null,
   "tps": 12.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-zqhkxcvg-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP320",
   "notes": "MTP Vision build; oMLX community benchmark trace, single run at 4k context; peak memory 21.6 GB; TTFT 31659 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 129.4,
   "provenance": "community",
   "quant": "oQ3e-DWQ",
   "quote": "Qwen3.8-Flash-Next-REAP320-oQ3e-DWQ-MTP-Vision-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 3bit; 4k tokens 129.4 PP tok/s 11.0 TG tok/s 31659 TTFT (ms) 21.6 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zqhkxcvg",
   "source_url": "https://omlx.ai/benchmarks/performance/zqhkxcvg",
   "tg_tokens": null,
   "tps": 11.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-0sit0ii7-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP320",
   "notes": "MTP Vision build; oMLX community benchmark trace, single run at 1k context; peak memory 21.1 GB; TTFT 10721 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 95.5,
   "provenance": "community",
   "quant": "oQ3e-DWQ",
   "quote": "Qwen3.8-Flash-Next-REAP320-oQ3e-DWQ-MTP-Vision-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 3bit; 1k tokens 95.5 PP tok/s 11.2 TG tok/s 10721 TTFT (ms) 21.1 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0sit0ii7",
   "source_url": "https://omlx.ai/benchmarks/performance/0sit0ii7",
   "tg_tokens": null,
   "tps": 11.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-xdeue6fx-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP320",
   "notes": "MTP Vision build; oMLX community benchmark trace, single run at 1k context; peak memory 9.6 GB; TTFT 11325 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 90.4,
   "provenance": "community",
   "quant": "oQ3e-DWQ",
   "quote": "Qwen3.8-Flash-Next-REAP320-oQ3e-DWQ-MTP-Vision-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 3bit; 1k tokens 90.4 PP tok/s 8.2 TG tok/s 11325 TTFT (ms) 9.6 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xdeue6fx",
   "source_url": "https://omlx.ai/benchmarks/performance/xdeue6fx",
   "tg_tokens": null,
   "tps": 8.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-h039j1lt-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP320",
   "notes": "MTP Vision build; oMLX community benchmark trace, single run at 4k context; peak memory 10.1 GB; TTFT 31455 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.2,
   "provenance": "community",
   "quant": "oQ3e-DWQ",
   "quote": "Qwen3.8-Flash-Next-REAP320-oQ3e-DWQ-MTP-Vision-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 3bit; 4k tokens 130.2 PP tok/s 7.4 TG tok/s 31455 TTFT (ms) 10.1 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h039j1lt",
   "source_url": "https://omlx.ai/benchmarks/performance/h039j1lt",
   "tg_tokens": null,
   "tps": 7.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-wesrz6pd-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.4 GB at this context; duplicate traces of the same session: aqep2tye; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2735.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 (80c); M1 (80c) \u00b7 256 GB \u00b7 4bit; 16k 2,735 36.2 102.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wesrz6pd",
   "source_url": "https://omlx.ai/benchmarks/performance/wesrz6pd",
   "tg_tokens": null,
   "tps": 36.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-wesrz6pd-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.3 GB at this context; duplicate traces of the same session: aqep2tye; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2655.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 (80c); M1 (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,655 65.4 104.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wesrz6pd",
   "source_url": "https://omlx.ai/benchmarks/performance/wesrz6pd",
   "tg_tokens": null,
   "tps": 65.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1",
   "id": "omlx-ab8qtzc4-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.7 GB at this context; duplicate traces of the same session: gqos8yty; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2372.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 (80c); M1 (80c) \u00b7 256 GB \u00b7 4bit; 16k 2,372 31.6 100.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ab8qtzc4",
   "source_url": "https://omlx.ai/benchmarks/performance/ab8qtzc4",
   "tg_tokens": null,
   "tps": 31.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1",
   "id": "omlx-ab8qtzc4-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.4 GB at this context; duplicate traces of the same session: gqos8yty; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2725.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 (80c); M1 (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,725 57.0 102.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ab8qtzc4",
   "source_url": "https://omlx.ai/benchmarks/performance/ab8qtzc4",
   "tg_tokens": null,
   "tps": 57.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-8575lc1h-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: jtrn9jke; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 699.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 1k 699.1 74.4 21.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8575lc1h",
   "source_url": "https://omlx.ai/benchmarks/performance/8575lc1h",
   "tg_tokens": null,
   "tps": 74.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-8575lc1h-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: jtrn9jke; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 778.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 4k 778.8 69.9 22.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8575lc1h",
   "source_url": "https://omlx.ai/benchmarks/performance/8575lc1h",
   "tg_tokens": null,
   "tps": 69.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-3r32bc3r-ctx1024",
   "model": "gemma-4-26b-a4b-it-UD",
   "notes": "oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: djxijy0m; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 477.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-UD-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 477.9 48.1 14.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3r32bc3r",
   "source_url": "https://omlx.ai/benchmarks/performance/3r32bc3r",
   "tg_tokens": null,
   "tps": 48.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-3r32bc3r-ctx4096",
   "model": "gemma-4-26b-a4b-it-UD",
   "notes": "oMLX community benchmark trace; peak memory 15.1 GB at this context; duplicate traces of the same session: djxijy0m; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 487.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-UD-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 487.8 47.0 15.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3r32bc3r",
   "source_url": "https://omlx.ai/benchmarks/performance/3r32bc3r",
   "tg_tokens": null,
   "tps": 47.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-bbp24yhx-ctx1024",
   "model": "NeoHorse-1-9B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 6.6 GB at this context; duplicate traces of the same session: 7o1wjbrp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 377.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NeoHorse-1-9B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 377.5 49.1 6.6 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bbp24yhx",
   "source_url": "https://omlx.ai/benchmarks/performance/bbp24yhx",
   "tg_tokens": null,
   "tps": 49.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-bbp24yhx-ctx16384",
   "model": "NeoHorse-1-9B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 7.1 GB at this context; duplicate traces of the same session: 7o1wjbrp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 477.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NeoHorse-1-9B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 16k 477.6 43.4 7.1 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bbp24yhx",
   "source_url": "https://omlx.ai/benchmarks/performance/bbp24yhx",
   "tg_tokens": null,
   "tps": 43.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-v88nrydy-ctx1024",
   "model": "NeoHorse-1-9B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 10.0 GB at this context; duplicate traces of the same session: 6z2k7e3j; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 371.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "NeoHorse-1-9B-MLX-8bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 8bit; 1k 371.7 27.7 10.0 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v88nrydy",
   "source_url": "https://omlx.ai/benchmarks/performance/v88nrydy",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-v88nrydy-ctx16384",
   "model": "NeoHorse-1-9B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 11.0 GB at this context; duplicate traces of the same session: 6z2k7e3j; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 345.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "NeoHorse-1-9B-MLX-8bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 8bit; 16k 345.9 25.6 11.0 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v88nrydy",
   "source_url": "https://omlx.ai/benchmarks/performance/v88nrydy",
   "tg_tokens": null,
   "tps": 25.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-8it1tm6t-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 12.3 GB at this context; duplicate traces of the same session: iywinff7; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 376.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B-8bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 8bit; 1k 376.4 38.3 12.3 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8it1tm6t",
   "source_url": "https://omlx.ai/benchmarks/performance/8it1tm6t",
   "tg_tokens": null,
   "tps": 38.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-8it1tm6t-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 13.3 GB at this context; duplicate traces of the same session: iywinff7; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 344.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B-8bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 8bit; 16k 344.4 44.8 13.3 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8it1tm6t",
   "source_url": "https://omlx.ai/benchmarks/performance/8it1tm6t",
   "tg_tokens": null,
   "tps": 44.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-uq9kour3-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 8.0 GB at this context; duplicate traces of the same session: bpw6pclt, hlk8bmgc, upj6joi6, w8ybeyev; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 276.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-9B-oQ4e-fp16-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 276.7 35.8 8.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uq9kour3",
   "source_url": "https://omlx.ai/benchmarks/performance/uq9kour3",
   "tg_tokens": null,
   "tps": 35.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-uq9kour3-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 8.5 GB at this context; duplicate traces of the same session: bpw6pclt, hlk8bmgc, upj6joi6, w8ybeyev; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 298.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-9B-oQ4e-fp16-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 298.2 35.7 8.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uq9kour3",
   "source_url": "https://omlx.ai/benchmarks/performance/uq9kour3",
   "tg_tokens": null,
   "tps": 35.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-uq9kour3-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 8.8 GB at this context; duplicate traces of the same session: bpw6pclt, hlk8bmgc, upj6joi6, w8ybeyev; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 295.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-9B-oQ4e-fp16-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 8k 295.2 36.2 8.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uq9kour3",
   "source_url": "https://omlx.ai/benchmarks/performance/uq9kour3",
   "tg_tokens": null,
   "tps": 36.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-uq9kour3-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 9.3 GB at this context; duplicate traces of the same session: bpw6pclt, hlk8bmgc, upj6joi6, w8ybeyev; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 281.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-9B-oQ4e-fp16-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k 281.4 55.2 9.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uq9kour3",
   "source_url": "https://omlx.ai/benchmarks/performance/uq9kour3",
   "tg_tokens": null,
   "tps": 55.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-uq9kour3-ctx32768",
   "model": "Qwen3.5-9B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 10.3 GB at this context; duplicate traces of the same session: bpw6pclt, hlk8bmgc, upj6joi6, w8ybeyev; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 253.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-9B-oQ4e-fp16-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 32k 253.9 79.0 10.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uq9kour3",
   "source_url": "https://omlx.ai/benchmarks/performance/uq9kour3",
   "tg_tokens": null,
   "tps": 79.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-4a4kkxo4-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 10.4 GB at this context; duplicate traces of the same session: hbk05v5r, q632gm9p, r0mtlhlm, x13mmdqz; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 273.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-9B-oQ6e-fp16-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 6bit; 1k 273.3 28.4 10.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4a4kkxo4",
   "source_url": "https://omlx.ai/benchmarks/performance/4a4kkxo4",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-4a4kkxo4-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 11.1 GB at this context; duplicate traces of the same session: hbk05v5r, q632gm9p, r0mtlhlm, x13mmdqz; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 285.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-9B-oQ6e-fp16-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 6bit; 4k 285.5 28.9 11.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4a4kkxo4",
   "source_url": "https://omlx.ai/benchmarks/performance/4a4kkxo4",
   "tg_tokens": null,
   "tps": 28.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-4a4kkxo4-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 11.3 GB at this context; duplicate traces of the same session: hbk05v5r, q632gm9p, r0mtlhlm, x13mmdqz; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 288.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-9B-oQ6e-fp16-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 6bit; 8k 288.0 27.8 11.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4a4kkxo4",
   "source_url": "https://omlx.ai/benchmarks/performance/4a4kkxo4",
   "tg_tokens": null,
   "tps": 27.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-4a4kkxo4-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: hbk05v5r, q632gm9p, r0mtlhlm, x13mmdqz; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 276.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-9B-oQ6e-fp16-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 6bit; 16k 276.1 30.7 11.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4a4kkxo4",
   "source_url": "https://omlx.ai/benchmarks/performance/4a4kkxo4",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-4a4kkxo4-ctx32768",
   "model": "Qwen3.5-9B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 12.8 GB at this context; duplicate traces of the same session: hbk05v5r, q632gm9p, r0mtlhlm, x13mmdqz; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 250.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-9B-oQ6e-fp16-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 6bit; 32k 250.2 37.7 12.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4a4kkxo4",
   "source_url": "https://omlx.ai/benchmarks/performance/4a4kkxo4",
   "tg_tokens": null,
   "tps": 37.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-cyev8die-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: tcyx2mhj; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1626.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 1,626 58.8 19.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cyev8die",
   "source_url": "https://omlx.ai/benchmarks/performance/cyev8die",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-cyev8die-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: tcyx2mhj; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1468.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,468 57.9 20.5 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cyev8die",
   "source_url": "https://omlx.ai/benchmarks/performance/cyev8die",
   "tg_tokens": null,
   "tps": 57.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ukvugblw-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.3 GB at this context; duplicate traces of the same session: 603xujzj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 348.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 8bit; 1k 348.3 9.3 28.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ukvugblw",
   "source_url": "https://omlx.ai/benchmarks/performance/ukvugblw",
   "tg_tokens": null,
   "tps": 9.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ukvugblw-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: 603xujzj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 340.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 8bit; 4k 340.5 9.7 29.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ukvugblw",
   "source_url": "https://omlx.ai/benchmarks/performance/ukvugblw",
   "tg_tokens": null,
   "tps": 9.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-1mxk8but-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: af8dmme7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1110.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 1k 1,110 32.5 28.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1mxk8but",
   "source_url": "https://omlx.ai/benchmarks/performance/1mxk8but",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-1mxk8but-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: af8dmme7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1603.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 4k 1,603 32.3 29.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1mxk8but",
   "source_url": "https://omlx.ai/benchmarks/performance/1mxk8but",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ug7gqope-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "oMLX community benchmark trace; peak memory 29.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 118.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 1k 118.6 19.7 29.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ug7gqope",
   "source_url": "https://omlx.ai/benchmarks/performance/ug7gqope",
   "tg_tokens": null,
   "tps": 19.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ug7gqope-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "oMLX community benchmark trace; peak memory 30.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 121.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 4k 121.2 14.7 30.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ug7gqope",
   "source_url": "https://omlx.ai/benchmarks/performance/ug7gqope",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-b18zb679-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 83dhmzwl, xgzpxz6t; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 126.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 126.5 22.7 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b18zb679",
   "source_url": "https://omlx.ai/benchmarks/performance/b18zb679",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-b18zb679-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 27.2 GB at this context; duplicate traces of the same session: 83dhmzwl, xgzpxz6t; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 129.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 129.3 22.6 27.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b18zb679",
   "source_url": "https://omlx.ai/benchmarks/performance/b18zb679",
   "tg_tokens": null,
   "tps": 22.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-b18zb679-ctx8192",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: 83dhmzwl, xgzpxz6t; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 124.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 124.8 23.0 28.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b18zb679",
   "source_url": "https://omlx.ai/benchmarks/performance/b18zb679",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-69px1xku-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 6pxx44rj, jtz99eoa; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 584.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 584.6 37.3 17.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/69px1xku",
   "source_url": "https://omlx.ai/benchmarks/performance/69px1xku",
   "tg_tokens": null,
   "tps": 37.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-69px1xku-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 6pxx44rj, jtz99eoa; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 687.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 687.2 39.1 19.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/69px1xku",
   "source_url": "https://omlx.ai/benchmarks/performance/69px1xku",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-69px1xku-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 6pxx44rj, jtz99eoa; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 654.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 654.3 39.2 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/69px1xku",
   "source_url": "https://omlx.ai/benchmarks/performance/69px1xku",
   "tg_tokens": null,
   "tps": 39.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-m0o7oer2-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 78p2gl41, n1l8g22c; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 586.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 586.6 34.2 17.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m0o7oer2",
   "source_url": "https://omlx.ai/benchmarks/performance/m0o7oer2",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-m0o7oer2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 78p2gl41, n1l8g22c; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 670.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 670.8 36.1 19.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m0o7oer2",
   "source_url": "https://omlx.ai/benchmarks/performance/m0o7oer2",
   "tg_tokens": null,
   "tps": 36.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-m0o7oer2-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 78p2gl41, n1l8g22c; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 699.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 699.5 51.1 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m0o7oer2",
   "source_url": "https://omlx.ai/benchmarks/performance/m0o7oer2",
   "tg_tokens": null,
   "tps": 51.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-u4j52b7m-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: 473hmpc2, m98mxalk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 815.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 815.4 55.5 33.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u4j52b7m",
   "source_url": "https://omlx.ai/benchmarks/performance/u4j52b7m",
   "tg_tokens": null,
   "tps": 55.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-u4j52b7m-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: 473hmpc2, m98mxalk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 899.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 8k 899.3 31.8 35.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u4j52b7m",
   "source_url": "https://omlx.ai/benchmarks/performance/u4j52b7m",
   "tg_tokens": null,
   "tps": 31.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-u4j52b7m-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.8 GB at this context; duplicate traces of the same session: 473hmpc2, m98mxalk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 748.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 32k 748.8 61.4 36.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u4j52b7m",
   "source_url": "https://omlx.ai/benchmarks/performance/u4j52b7m",
   "tg_tokens": null,
   "tps": 61.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qr1ehxbc-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 77.8 GB; TTFT 913 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1122.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,122 PP tok/s 68.3 TG tok/s 913 TTFT (ms) 77.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qr1ehxbc",
   "source_url": "https://omlx.ai/benchmarks/performance/qr1ehxbc",
   "tg_tokens": null,
   "tps": 68.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hrvrt54a-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 77.8 GB; TTFT 914 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1121.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,121 PP tok/s 67.7 TG tok/s 914 TTFT (ms) 77.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hrvrt54a",
   "source_url": "https://omlx.ai/benchmarks/performance/hrvrt54a",
   "tg_tokens": null,
   "tps": 67.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fsty0g5w-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 78.1 GB; TTFT 2613 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1568.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 1,568 PP tok/s 69.9 TG tok/s 2613 TTFT (ms) 78.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fsty0g5w",
   "source_url": "https://omlx.ai/benchmarks/performance/fsty0g5w",
   "tg_tokens": null,
   "tps": 69.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-n661apeo-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.1 GB at this context; duplicate traces of the same session: hvhgvt8d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1123.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,123 70.7 78.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n661apeo",
   "source_url": "https://omlx.ai/benchmarks/performance/n661apeo",
   "tg_tokens": null,
   "tps": 70.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-n661apeo-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.0 GB at this context; duplicate traces of the same session: hvhgvt8d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1816.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,816 79.6 78.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n661apeo",
   "source_url": "https://omlx.ai/benchmarks/performance/n661apeo",
   "tg_tokens": null,
   "tps": 79.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-skl3im14-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 69.6 GB at this context; duplicate traces of the same session: 4f9m3wvj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 559.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 559.3 77.9 69.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/skl3im14",
   "source_url": "https://omlx.ai/benchmarks/performance/skl3im14",
   "tg_tokens": null,
   "tps": 77.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-skl3im14-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 70.2 GB at this context; duplicate traces of the same session: 4f9m3wvj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1167.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,167 62.2 70.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/skl3im14",
   "source_url": "https://omlx.ai/benchmarks/performance/skl3im14",
   "tg_tokens": null,
   "tps": 62.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rdwkas5d-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: g2ty0nt1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1190.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,190 67.6 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rdwkas5d",
   "source_url": "https://omlx.ai/benchmarks/performance/rdwkas5d",
   "tg_tokens": null,
   "tps": 67.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rdwkas5d-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: g2ty0nt1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1665.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,665 68.7 70.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rdwkas5d",
   "source_url": "https://omlx.ai/benchmarks/performance/rdwkas5d",
   "tg_tokens": null,
   "tps": 68.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rdwkas5d-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.5 GB at this context; duplicate traces of the same session: g2ty0nt1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2084.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,084 79.8 74.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rdwkas5d",
   "source_url": "https://omlx.ai/benchmarks/performance/rdwkas5d",
   "tg_tokens": null,
   "tps": 79.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-xhjuifdd-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 93.9 GB; TTFT 37016 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3541.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 5bit; 128k tokens 3,541 PP tok/s 93.3 TG tok/s 37016 TTFT (ms) 93.9 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xhjuifdd",
   "source_url": "https://omlx.ai/benchmarks/performance/xhjuifdd",
   "tg_tokens": null,
   "tps": 93.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-u9egcg74-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 90.6 GB; TTFT 19644 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3336.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 5bit; 64k tokens 3,336 PP tok/s 105.4 TG tok/s 19644 TTFT (ms) 90.6 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u9egcg74",
   "source_url": "https://omlx.ai/benchmarks/performance/u9egcg74",
   "tg_tokens": null,
   "tps": 105.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-k26q5ybd-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 90.6 GB; TTFT 19483 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3364.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 5bit; 64k tokens 3,364 PP tok/s 102.5 TG tok/s 19483 TTFT (ms) 90.6 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k26q5ybd",
   "source_url": "https://omlx.ai/benchmarks/performance/k26q5ybd",
   "tg_tokens": null,
   "tps": 102.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-knbc44qh-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 133.2 GB; TTFT 16713 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3921.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 64k tokens 3,921 PP tok/s 78.8 TG tok/s 16713 TTFT (ms) 133.2 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/knbc44qh",
   "source_url": "https://omlx.ai/benchmarks/performance/knbc44qh",
   "tg_tokens": null,
   "tps": 78.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-cs09f7ph-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 137.6 GB; TTFT 32300 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4058.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 128k tokens 4,058 PP tok/s 82.9 TG tok/s 32300 TTFT (ms) 137.6 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cs09f7ph",
   "source_url": "https://omlx.ai/benchmarks/performance/cs09f7ph",
   "tg_tokens": null,
   "tps": 82.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-y8scz20l-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 133.4 GB; TTFT 16829 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3894.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 64k tokens 3,894 PP tok/s 95.2 TG tok/s 16829 TTFT (ms) 133.4 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y8scz20l",
   "source_url": "https://omlx.ai/benchmarks/performance/y8scz20l",
   "tg_tokens": null,
   "tps": 95.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-5cxy06l3-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 130.5 GB; TTFT 4908 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1669.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 8k tokens 1,669 PP tok/s 34.5 TG tok/s 4908 TTFT (ms) 130.5 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5cxy06l3",
   "source_url": "https://omlx.ai/benchmarks/performance/5cxy06l3",
   "tg_tokens": null,
   "tps": 34.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-o73hcidu-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: vn0316vl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 835.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 835.9 64.0 27.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o73hcidu",
   "source_url": "https://omlx.ai/benchmarks/performance/o73hcidu",
   "tg_tokens": null,
   "tps": 64.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-o73hcidu-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: vn0316vl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 919.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 919.6 54.8 28.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o73hcidu",
   "source_url": "https://omlx.ai/benchmarks/performance/o73hcidu",
   "tg_tokens": null,
   "tps": 54.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-o73hcidu-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 30.4 GB at this context; duplicate traces of the same session: vn0316vl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 702.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 702.6 84.8 30.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o73hcidu",
   "source_url": "https://omlx.ai/benchmarks/performance/o73hcidu",
   "tg_tokens": null,
   "tps": 84.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-o73hcidu-ctx65536",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 32.8 GB at this context; duplicate traces of the same session: vn0316vl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 549.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 64k 549.8 35.4 32.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o73hcidu",
   "source_url": "https://omlx.ai/benchmarks/performance/o73hcidu",
   "tg_tokens": null,
   "tps": 35.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-lh30wsmz-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: 7f81ccbs, h69oyxpo, mcm84gmc, ue3xjub7; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 637.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 637.1 64.7 14.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lh30wsmz",
   "source_url": "https://omlx.ai/benchmarks/performance/lh30wsmz",
   "tg_tokens": null,
   "tps": 64.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-lh30wsmz-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: 7f81ccbs, h69oyxpo, mcm84gmc, ue3xjub7; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 639.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 639.7 59.5 14.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lh30wsmz",
   "source_url": "https://omlx.ai/benchmarks/performance/lh30wsmz",
   "tg_tokens": null,
   "tps": 59.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-lh30wsmz-ctx8192",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 15.0 GB at this context; duplicate traces of the same session: 7f81ccbs, h69oyxpo, mcm84gmc, ue3xjub7; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 616.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 616.4 59.0 15.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lh30wsmz",
   "source_url": "https://omlx.ai/benchmarks/performance/lh30wsmz",
   "tg_tokens": null,
   "tps": 59.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-lh30wsmz-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 15.4 GB at this context; duplicate traces of the same session: 7f81ccbs, h69oyxpo, mcm84gmc, ue3xjub7; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 552.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 552.1 49.7 15.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lh30wsmz",
   "source_url": "https://omlx.ai/benchmarks/performance/lh30wsmz",
   "tg_tokens": null,
   "tps": 49.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-lh30wsmz-ctx32768",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: 7f81ccbs, h69oyxpo, mcm84gmc, ue3xjub7; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 504.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 504.0 44.3 16.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lh30wsmz",
   "source_url": "https://omlx.ai/benchmarks/performance/lh30wsmz",
   "tg_tokens": null,
   "tps": 44.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-3wfgadhf-ctx16384",
   "model": "Qwen3 14B",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 11.2 GB at this context; duplicate traces of the same session: ewqtdto4; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 680.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3-14B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 16k 680.2 24.7 11.2 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3wfgadhf",
   "source_url": "https://omlx.ai/benchmarks/performance/3wfgadhf",
   "tg_tokens": null,
   "tps": 24.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-3wfgadhf-ctx32768",
   "model": "Qwen3 14B",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 13.3 GB at this context; duplicate traces of the same session: ewqtdto4; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 546.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3-14B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 32k 546.9 19.6 13.3 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3wfgadhf",
   "source_url": "https://omlx.ai/benchmarks/performance/3wfgadhf",
   "tg_tokens": null,
   "tps": 19.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-i7tuq13v-ctx1024",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 10.0 GB at this context; duplicate traces of the same session: 0l92sbqm, afleyve4, ks1yclxz, lqwzazgr; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-oQ8e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 8bit; 1k 220.4 20.1 10.0 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i7tuq13v",
   "source_url": "https://omlx.ai/benchmarks/performance/i7tuq13v",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-i7tuq13v-ctx4096",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 10.6 GB at this context; duplicate traces of the same session: 0l92sbqm, afleyve4, ks1yclxz, lqwzazgr; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 247.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-oQ8e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 8bit; 4k 247.8 19.7 10.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i7tuq13v",
   "source_url": "https://omlx.ai/benchmarks/performance/i7tuq13v",
   "tg_tokens": null,
   "tps": 19.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-i7tuq13v-ctx8192",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 10.8 GB at this context; duplicate traces of the same session: 0l92sbqm, afleyve4, ks1yclxz, lqwzazgr; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 242.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-oQ8e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 8bit; 8k 242.7 19.4 10.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i7tuq13v",
   "source_url": "https://omlx.ai/benchmarks/performance/i7tuq13v",
   "tg_tokens": null,
   "tps": 19.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-i7tuq13v-ctx16384",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 11.2 GB at this context; duplicate traces of the same session: 0l92sbqm, afleyve4, ks1yclxz, lqwzazgr; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 230.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-oQ8e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 8bit; 16k 230.7 18.5 11.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i7tuq13v",
   "source_url": "https://omlx.ai/benchmarks/performance/i7tuq13v",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-i7tuq13v-ctx32768",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 12.0 GB at this context; duplicate traces of the same session: 0l92sbqm, afleyve4, ks1yclxz, lqwzazgr; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 209.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-oQ8e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 8bit; 32k 209.3 17.2 12.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i7tuq13v",
   "source_url": "https://omlx.ai/benchmarks/performance/i7tuq13v",
   "tg_tokens": null,
   "tps": 17.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-gu0lg35i-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 64k context; peak memory 23.6 GB; TTFT 143673 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 456.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k tokens 456.1 PP tok/s 58.7 TG tok/s 143673 TTFT (ms) 23.6 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gu0lg35i",
   "source_url": "https://omlx.ai/benchmarks/performance/gu0lg35i",
   "tg_tokens": null,
   "tps": 58.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-vmlwvbsj-ctx65536",
   "model": "Ornith-1.5-35B-A3B-uncensored",
   "notes": "MLX build; oMLX community benchmark trace, single run at 64k context; peak memory 20.3 GB; TTFT 34247 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1914.0,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Ornith-1.5-35B-A3B-uncensored-MLX-MXFP4 on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k tokens 1,914 PP tok/s 63.4 TG tok/s 34247 TTFT (ms) 20.3 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vmlwvbsj",
   "source_url": "https://omlx.ai/benchmarks/performance/vmlwvbsj",
   "tg_tokens": null,
   "tps": 63.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-4eq5h6wv-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 10.0 GB at this context; duplicate traces of the same session: 0sgb3u4q, 3w2mvozq, 7qystnpf, ed4epzpx; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 213.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-9B-oQ8e-fp16-mtp on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 8bit; 1k 213.3 19.4 10.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4eq5h6wv",
   "source_url": "https://omlx.ai/benchmarks/performance/4eq5h6wv",
   "tg_tokens": null,
   "tps": 19.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-4eq5h6wv-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 10.6 GB at this context; duplicate traces of the same session: 0sgb3u4q, 3w2mvozq, 7qystnpf, ed4epzpx; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 246.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-9B-oQ8e-fp16-mtp on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 8bit; 4k 246.9 18.8 10.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4eq5h6wv",
   "source_url": "https://omlx.ai/benchmarks/performance/4eq5h6wv",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-4eq5h6wv-ctx8192",
   "model": "Ornith-1.5-9B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 10.8 GB at this context; duplicate traces of the same session: 0sgb3u4q, 3w2mvozq, 7qystnpf, ed4epzpx; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 242.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-9B-oQ8e-fp16-mtp on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 8bit; 8k 242.6 19.3 10.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4eq5h6wv",
   "source_url": "https://omlx.ai/benchmarks/performance/4eq5h6wv",
   "tg_tokens": null,
   "tps": 19.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-4eq5h6wv-ctx16384",
   "model": "Ornith-1.5-9B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 11.2 GB at this context; duplicate traces of the same session: 0sgb3u4q, 3w2mvozq, 7qystnpf, ed4epzpx; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 230.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-9B-oQ8e-fp16-mtp on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 8bit; 16k 230.6 18.5 11.2 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4eq5h6wv",
   "source_url": "https://omlx.ai/benchmarks/performance/4eq5h6wv",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-4eq5h6wv-ctx32768",
   "model": "Ornith-1.5-9B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 12.1 GB at this context; duplicate traces of the same session: 0sgb3u4q, 3w2mvozq, 7qystnpf, ed4epzpx; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 209.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-9B-oQ8e-fp16-mtp on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 8bit; 32k 209.1 17.1 12.1 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4eq5h6wv",
   "source_url": "https://omlx.ai/benchmarks/performance/4eq5h6wv",
   "tg_tokens": null,
   "tps": 17.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-rg4488w2-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 10.0 GB at this context; duplicate traces of the same session: 2aywumtk, dkm0k558, paqoaaf1, waam36qd; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-9B-oQ8e-fp16-mtp on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 8bit; 1k 220.8 20.1 10.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rg4488w2",
   "source_url": "https://omlx.ai/benchmarks/performance/rg4488w2",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-rg4488w2-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 10.6 GB at this context; duplicate traces of the same session: 2aywumtk, dkm0k558, paqoaaf1, waam36qd; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 248.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-9B-oQ8e-fp16-mtp on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 8bit; 4k 248.3 19.7 10.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rg4488w2",
   "source_url": "https://omlx.ai/benchmarks/performance/rg4488w2",
   "tg_tokens": null,
   "tps": 19.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-rg4488w2-ctx8192",
   "model": "Ornith-1.5-9B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 10.8 GB at this context; duplicate traces of the same session: 2aywumtk, dkm0k558, paqoaaf1, waam36qd; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 242.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-9B-oQ8e-fp16-mtp on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 8bit; 8k 242.7 19.4 10.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rg4488w2",
   "source_url": "https://omlx.ai/benchmarks/performance/rg4488w2",
   "tg_tokens": null,
   "tps": 19.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-rg4488w2-ctx16384",
   "model": "Ornith-1.5-9B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 11.2 GB at this context; duplicate traces of the same session: 2aywumtk, dkm0k558, paqoaaf1, waam36qd; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 230.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-9B-oQ8e-fp16-mtp on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 8bit; 16k 230.8 18.6 11.2 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rg4488w2",
   "source_url": "https://omlx.ai/benchmarks/performance/rg4488w2",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-rg4488w2-ctx32768",
   "model": "Ornith-1.5-9B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 12.1 GB at this context; duplicate traces of the same session: 2aywumtk, dkm0k558, paqoaaf1, waam36qd; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 208.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-9B-oQ8e-fp16-mtp on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 8bit; 32k 208.4 17.2 12.1 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rg4488w2",
   "source_url": "https://omlx.ai/benchmarks/performance/rg4488w2",
   "tg_tokens": null,
   "tps": 17.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-gpye7kgv-ctx1024",
   "model": "Qwen2.5-Coder-0.5B-Instruct",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 1.2 GB at this context; duplicate traces of the same session: 6ymg8aiv, vu22r5go, zqfzidyh; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3840.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen2.5-Coder-0.5B-Instruct-oQ8e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 8bit; 1k 3,840 176.4 1.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gpye7kgv",
   "source_url": "https://omlx.ai/benchmarks/performance/gpye7kgv",
   "tg_tokens": null,
   "tps": 176.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-gpye7kgv-ctx4096",
   "model": "Qwen2.5-Coder-0.5B-Instruct",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 1.3 GB at this context; duplicate traces of the same session: 6ymg8aiv, vu22r5go, zqfzidyh; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3677.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen2.5-Coder-0.5B-Instruct-oQ8e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 8bit; 4k 3,677 154.7 1.3 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gpye7kgv",
   "source_url": "https://omlx.ai/benchmarks/performance/gpye7kgv",
   "tg_tokens": null,
   "tps": 154.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-gpye7kgv-ctx8192",
   "model": "Qwen2.5-Coder-0.5B-Instruct",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 1.3 GB at this context; duplicate traces of the same session: 6ymg8aiv, vu22r5go, zqfzidyh; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3078.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen2.5-Coder-0.5B-Instruct-oQ8e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 8bit; 8k 3,078 125.5 1.3 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gpye7kgv",
   "source_url": "https://omlx.ai/benchmarks/performance/gpye7kgv",
   "tg_tokens": null,
   "tps": 125.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-gpye7kgv-ctx16384",
   "model": "Qwen2.5-Coder-0.5B-Instruct",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 1.4 GB at this context; duplicate traces of the same session: 6ymg8aiv, vu22r5go, zqfzidyh; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2350.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen2.5-Coder-0.5B-Instruct-oQ8e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 8bit; 16k 2,350 98.1 1.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gpye7kgv",
   "source_url": "https://omlx.ai/benchmarks/performance/gpye7kgv",
   "tg_tokens": null,
   "tps": 98.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-gpye7kgv-ctx32768",
   "model": "Qwen2.5-Coder-0.5B-Instruct",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 1.6 GB at this context; duplicate traces of the same session: 6ymg8aiv, vu22r5go, zqfzidyh; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1501.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen2.5-Coder-0.5B-Instruct-oQ8e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 8bit; 32k 1,501 61.9 1.6 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gpye7kgv",
   "source_url": "https://omlx.ai/benchmarks/performance/gpye7kgv",
   "tg_tokens": null,
   "tps": 61.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ljqpekt1-ctx1024",
   "model": "Qwen3-Coder-Next-REAM",
   "notes": "oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: fmwl6p5l, zkb0t3t4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1301.0,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3-Coder-Next-REAM-mlx-3Bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 3bit; 1k 1,301 91.0 26.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ljqpekt1",
   "source_url": "https://omlx.ai/benchmarks/performance/ljqpekt1",
   "tg_tokens": null,
   "tps": 91.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ljqpekt1-ctx4096",
   "model": "Qwen3-Coder-Next-REAM",
   "notes": "oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: fmwl6p5l, zkb0t3t4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1472.0,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3-Coder-Next-REAM-mlx-3Bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 3bit; 4k 1,472 86.6 26.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ljqpekt1",
   "source_url": "https://omlx.ai/benchmarks/performance/ljqpekt1",
   "tg_tokens": null,
   "tps": 86.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ljqpekt1-ctx8192",
   "model": "Qwen3-Coder-Next-REAM",
   "notes": "oMLX community benchmark trace; peak memory 26.9 GB at this context; duplicate traces of the same session: fmwl6p5l, zkb0t3t4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1257.0,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3-Coder-Next-REAM-mlx-3Bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 3bit; 8k 1,257 84.5 26.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ljqpekt1",
   "source_url": "https://omlx.ai/benchmarks/performance/ljqpekt1",
   "tg_tokens": null,
   "tps": 84.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-pijm6hym-ctx1024",
   "model": "Qwen3-Coder-Next-REAP-40B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: nld4a6p8, rqha1tda; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1377.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3-Coder-Next-REAP-40B-A3B-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,377 85.2 23.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pijm6hym",
   "source_url": "https://omlx.ai/benchmarks/performance/pijm6hym",
   "tg_tokens": null,
   "tps": 85.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-pijm6hym-ctx4096",
   "model": "Qwen3-Coder-Next-REAP-40B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: nld4a6p8, rqha1tda; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1530.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3-Coder-Next-REAP-40B-A3B-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,530 82.7 24.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pijm6hym",
   "source_url": "https://omlx.ai/benchmarks/performance/pijm6hym",
   "tg_tokens": null,
   "tps": 82.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-pijm6hym-ctx8192",
   "model": "Qwen3-Coder-Next-REAP-40B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: nld4a6p8, rqha1tda; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1327.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3-Coder-Next-REAP-40B-A3B-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,327 77.3 24.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pijm6hym",
   "source_url": "https://omlx.ai/benchmarks/performance/pijm6hym",
   "tg_tokens": null,
   "tps": 77.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-z7ovbd3x-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 2z773j76; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 76.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 76.8 13.8 18.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z7ovbd3x",
   "source_url": "https://omlx.ai/benchmarks/performance/z7ovbd3x",
   "tg_tokens": null,
   "tps": 13.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-z7ovbd3x-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 2z773j76; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 71.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 71.5 14.4 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z7ovbd3x",
   "source_url": "https://omlx.ai/benchmarks/performance/z7ovbd3x",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i1jthr1u-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1190.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,190 67.6 70.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i1jthr1u",
   "source_url": "https://omlx.ai/benchmarks/performance/i1jthr1u",
   "tg_tokens": null,
   "tps": 67.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i1jthr1u-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1665.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,665 68.7 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i1jthr1u",
   "source_url": "https://omlx.ai/benchmarks/performance/i1jthr1u",
   "tg_tokens": null,
   "tps": 68.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i1jthr1u-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2084.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,084 79.8 74.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i1jthr1u",
   "source_url": "https://omlx.ai/benchmarks/performance/i1jthr1u",
   "tg_tokens": null,
   "tps": 79.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5w44iewp-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: h8i9he9y, nq8gh94p; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1156.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,156 66.4 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5w44iewp",
   "source_url": "https://omlx.ai/benchmarks/performance/5w44iewp",
   "tg_tokens": null,
   "tps": 66.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5w44iewp-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: h8i9he9y, nq8gh94p; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1645.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,645 73.9 70.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5w44iewp",
   "source_url": "https://omlx.ai/benchmarks/performance/5w44iewp",
   "tg_tokens": null,
   "tps": 73.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5w44iewp-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.7 GB at this context; duplicate traces of the same session: h8i9he9y, nq8gh94p; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2127.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,127 79.1 74.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5w44iewp",
   "source_url": "https://omlx.ai/benchmarks/performance/5w44iewp",
   "tg_tokens": null,
   "tps": 79.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q5i8f8qt-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: hzqyzq9r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1207.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,207 84.7 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q5i8f8qt",
   "source_url": "https://omlx.ai/benchmarks/performance/q5i8f8qt",
   "tg_tokens": null,
   "tps": 84.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q5i8f8qt-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.5 GB at this context; duplicate traces of the same session: hzqyzq9r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2150.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,150 69.2 74.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q5i8f8qt",
   "source_url": "https://omlx.ai/benchmarks/performance/q5i8f8qt",
   "tg_tokens": null,
   "tps": 69.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xiv0ldh1-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.2 GB at this context; duplicate traces of the same session: 0w90k84u, 38sj60vb, 6e07dbm4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1162.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,162 66.0 69.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xiv0ldh1",
   "source_url": "https://omlx.ai/benchmarks/performance/xiv0ldh1",
   "tg_tokens": null,
   "tps": 66.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xiv0ldh1-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.4 GB at this context; duplicate traces of the same session: 0w90k84u, 38sj60vb, 6e07dbm4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1488.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,488 63.7 69.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xiv0ldh1",
   "source_url": "https://omlx.ai/benchmarks/performance/xiv0ldh1",
   "tg_tokens": null,
   "tps": 63.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xiv0ldh1-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.4 GB at this context; duplicate traces of the same session: 0w90k84u, 38sj60vb, 6e07dbm4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2201.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,201 68.7 71.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xiv0ldh1",
   "source_url": "https://omlx.ai/benchmarks/performance/xiv0ldh1",
   "tg_tokens": null,
   "tps": 68.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xiv0ldh1-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.4 GB at this context; duplicate traces of the same session: 0w90k84u, 38sj60vb, 6e07dbm4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2338.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,338 66.3 72.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xiv0ldh1",
   "source_url": "https://omlx.ai/benchmarks/performance/xiv0ldh1",
   "tg_tokens": null,
   "tps": 66.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rp440vgs-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: lg4xfaph; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1024.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,024 61.3 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rp440vgs",
   "source_url": "https://omlx.ai/benchmarks/performance/rp440vgs",
   "tg_tokens": null,
   "tps": 61.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rp440vgs-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: lg4xfaph; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1933.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,933 61.9 70.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rp440vgs",
   "source_url": "https://omlx.ai/benchmarks/performance/rp440vgs",
   "tg_tokens": null,
   "tps": 61.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rp440vgs-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: lg4xfaph; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2438.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,438 68.8 73.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rp440vgs",
   "source_url": "https://omlx.ai/benchmarks/performance/rp440vgs",
   "tg_tokens": null,
   "tps": 68.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rp440vgs-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.6 GB at this context; duplicate traces of the same session: lg4xfaph; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2403.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,403 77.3 74.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rp440vgs",
   "source_url": "https://omlx.ai/benchmarks/performance/rp440vgs",
   "tg_tokens": null,
   "tps": 77.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-a739p9tr-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.6 GB at this context; duplicate traces of the same session: 14u2ysxh, ey6go4ay, fu2bpbzz, pmcprcqj, wseyu9uz, z1d3aybo; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2464.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 2,464 133.7 142.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a739p9tr",
   "source_url": "https://omlx.ai/benchmarks/performance/a739p9tr",
   "tg_tokens": null,
   "tps": 133.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-a739p9tr-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 144.5 GB at this context; duplicate traces of the same session: 14u2ysxh, ey6go4ay, fu2bpbzz, pmcprcqj, wseyu9uz, z1d3aybo; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4876.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 8k 4,876 109.9 144.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a739p9tr",
   "source_url": "https://omlx.ai/benchmarks/performance/a739p9tr",
   "tg_tokens": null,
   "tps": 109.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-a739p9tr-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.1 GB at this context; duplicate traces of the same session: 14u2ysxh, ey6go4ay, fu2bpbzz, pmcprcqj, wseyu9uz, z1d3aybo; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4477.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 16k 4,477 120.7 145.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a739p9tr",
   "source_url": "https://omlx.ai/benchmarks/performance/a739p9tr",
   "tg_tokens": null,
   "tps": 120.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-a739p9tr-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.7 GB at this context; duplicate traces of the same session: 14u2ysxh, ey6go4ay, fu2bpbzz, pmcprcqj, wseyu9uz, z1d3aybo; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4896.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 32k 4,896 67.3 145.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a739p9tr",
   "source_url": "https://omlx.ai/benchmarks/performance/a739p9tr",
   "tg_tokens": null,
   "tps": 67.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-a739p9tr-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 146.8 GB at this context; duplicate traces of the same session: 14u2ysxh, ey6go4ay, fu2bpbzz, pmcprcqj, wseyu9uz, z1d3aybo; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4773.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 64k 4,773 59.5 146.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a739p9tr",
   "source_url": "https://omlx.ai/benchmarks/performance/a739p9tr",
   "tg_tokens": null,
   "tps": 59.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-a739p9tr-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 150.0 GB at this context; duplicate traces of the same session: 14u2ysxh, ey6go4ay, fu2bpbzz, pmcprcqj, wseyu9uz, z1d3aybo; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4650.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 128k 4,650 94.1 150.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a739p9tr",
   "source_url": "https://omlx.ai/benchmarks/performance/a739p9tr",
   "tg_tokens": null,
   "tps": 94.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-a739p9tr-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 153.3 GB at this context; duplicate traces of the same session: 14u2ysxh, ey6go4ay, fu2bpbzz, pmcprcqj, wseyu9uz, z1d3aybo; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4505.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 195k 4,505 92.2 153.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a739p9tr",
   "source_url": "https://omlx.ai/benchmarks/performance/a739p9tr",
   "tg_tokens": null,
   "tps": 92.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-8qct34d4-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 133.2 GB; TTFT 16556 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3958.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 64k tokens 3,958 PP tok/s 69.5 TG tok/s 16556 TTFT (ms) 133.2 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8qct34d4",
   "source_url": "https://omlx.ai/benchmarks/performance/8qct34d4",
   "tg_tokens": null,
   "tps": 69.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-s47p7m6y-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: zej3trsf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 835.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 835.9 64.0 27.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s47p7m6y",
   "source_url": "https://omlx.ai/benchmarks/performance/s47p7m6y",
   "tg_tokens": null,
   "tps": 64.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-s47p7m6y-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: zej3trsf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 919.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 919.6 54.8 28.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s47p7m6y",
   "source_url": "https://omlx.ai/benchmarks/performance/s47p7m6y",
   "tg_tokens": null,
   "tps": 54.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-s47p7m6y-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 30.4 GB at this context; duplicate traces of the same session: zej3trsf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 702.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 702.6 84.8 30.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s47p7m6y",
   "source_url": "https://omlx.ai/benchmarks/performance/s47p7m6y",
   "tg_tokens": null,
   "tps": 84.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-s47p7m6y-ctx65536",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 32.8 GB at this context; duplicate traces of the same session: zej3trsf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 549.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 64k 549.8 35.4 32.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s47p7m6y",
   "source_url": "https://omlx.ai/benchmarks/performance/s47p7m6y",
   "tg_tokens": null,
   "tps": 35.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zlhhpc2r-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: 5av75q6a, gulw2m49, p7l0lnqk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 833.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 833.6 57.2 27.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zlhhpc2r",
   "source_url": "https://omlx.ai/benchmarks/performance/zlhhpc2r",
   "tg_tokens": null,
   "tps": 57.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-zlhhpc2r-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 5av75q6a, gulw2m49, p7l0lnqk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 918.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 918.4 57.7 28.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zlhhpc2r",
   "source_url": "https://omlx.ai/benchmarks/performance/zlhhpc2r",
   "tg_tokens": null,
   "tps": 57.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-zlhhpc2r-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 30.4 GB at this context; duplicate traces of the same session: 5av75q6a, gulw2m49, p7l0lnqk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 722.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 722.5 86.3 30.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zlhhpc2r",
   "source_url": "https://omlx.ai/benchmarks/performance/zlhhpc2r",
   "tg_tokens": null,
   "tps": 86.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-zlhhpc2r-ctx65536",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "MTP speculative decoding enabled; fp16 weights; oMLX community benchmark trace; peak memory 32.8 GB at this context; duplicate traces of the same session: 5av75q6a, gulw2m49, p7l0lnqk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 601.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 64k 601.6 37.3 32.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zlhhpc2r",
   "source_url": "https://omlx.ai/benchmarks/performance/zlhhpc2r",
   "tg_tokens": null,
   "tps": 37.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1bh4kt9t-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace, single run at 64k context; peak memory 26.5 GB; TTFT 81324 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 805.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k tokens 805.9 PP tok/s 21.1 TG tok/s 81324 TTFT (ms) 26.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1bh4kt9t",
   "source_url": "https://omlx.ai/benchmarks/performance/1bh4kt9t",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-h1j3hxn5-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: p6hagcz5; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 598.7,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 1k 598.7 69.4 11.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h1j3hxn5",
   "source_url": "https://omlx.ai/benchmarks/performance/h1j3hxn5",
   "tg_tokens": null,
   "tps": 69.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-h1j3hxn5-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: p6hagcz5; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 615.7,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 4k 615.7 64.7 11.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h1j3hxn5",
   "source_url": "https://omlx.ai/benchmarks/performance/h1j3hxn5",
   "tg_tokens": null,
   "tps": 64.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-yvmxjsjb-ctx1024",
   "model": "Kolibri-1",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 33.2 GB at this context; duplicate traces of the same session: dqfb1bq2; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 308.8,
   "provenance": "community",
   "quant": "3bit",
   "quote": "here-be-dragons-ai--Kolibri-1-MLX-3bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 3bit; 1k 308.8 38.7 33.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yvmxjsjb",
   "source_url": "https://omlx.ai/benchmarks/performance/yvmxjsjb",
   "tg_tokens": null,
   "tps": 38.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-yvmxjsjb-ctx4096",
   "model": "Kolibri-1",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 33.3 GB at this context; duplicate traces of the same session: dqfb1bq2; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 327.6,
   "provenance": "community",
   "quant": "3bit",
   "quote": "here-be-dragons-ai--Kolibri-1-MLX-3bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 3bit; 4k 327.6 37.2 33.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yvmxjsjb",
   "source_url": "https://omlx.ai/benchmarks/performance/yvmxjsjb",
   "tg_tokens": null,
   "tps": 37.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-klerzwqf-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 16.7 GB at this context; duplicate traces of the same session: b0cafhej; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 461.1,
   "provenance": "community",
   "quant": "5bit",
   "quote": "LFM2.5-24B-A2B-MLX-5bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 5bit; 16k 461.1 43.7 16.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/klerzwqf",
   "source_url": "https://omlx.ai/benchmarks/performance/klerzwqf",
   "tg_tokens": null,
   "tps": 43.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-klerzwqf-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: b0cafhej; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 422.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "LFM2.5-24B-A2B-MLX-5bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 5bit; 32k 422.0 36.8 17.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/klerzwqf",
   "source_url": "https://omlx.ai/benchmarks/performance/klerzwqf",
   "tg_tokens": null,
   "tps": 36.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-jxbr0qr3-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 8k context; peak memory 16.5 GB; TTFT 17118 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 478.6,
   "provenance": "community",
   "quant": "5bit",
   "quote": "LFM2.5-24B-A2B-MLX-5bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 5bit; 8k tokens 478.6 PP tok/s 48.7 TG tok/s 17118 TTFT (ms) 16.5 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jxbr0qr3",
   "source_url": "https://omlx.ai/benchmarks/performance/jxbr0qr3",
   "tg_tokens": null,
   "tps": 48.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-zagfw8ha-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: tiyrhxjv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 907.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 907.3 89.9 21.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zagfw8ha",
   "source_url": "https://omlx.ai/benchmarks/performance/zagfw8ha",
   "tg_tokens": null,
   "tps": 89.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-zagfw8ha-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: tiyrhxjv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 926.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 926.6 100.1 22.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zagfw8ha",
   "source_url": "https://omlx.ai/benchmarks/performance/zagfw8ha",
   "tg_tokens": null,
   "tps": 100.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-8a3mi2kj-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: olo2x4zp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 906.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 906.3 97.1 21.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8a3mi2kj",
   "source_url": "https://omlx.ai/benchmarks/performance/8a3mi2kj",
   "tg_tokens": null,
   "tps": 97.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-8a3mi2kj-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: olo2x4zp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 953.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 953.9 97.5 22.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8a3mi2kj",
   "source_url": "https://omlx.ai/benchmarks/performance/8a3mi2kj",
   "tg_tokens": null,
   "tps": 97.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-5dc171gq-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: rs2ovrl3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 906.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 906.3 99.1 21.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5dc171gq",
   "source_url": "https://omlx.ai/benchmarks/performance/5dc171gq",
   "tg_tokens": null,
   "tps": 99.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-5dc171gq-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: rs2ovrl3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 951.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 951.9 101.1 22.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5dc171gq",
   "source_url": "https://omlx.ai/benchmarks/performance/5dc171gq",
   "tg_tokens": null,
   "tps": 101.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-f1pfx82g-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 56nzma9h; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 904.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 904.1 68.3 21.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f1pfx82g",
   "source_url": "https://omlx.ai/benchmarks/performance/f1pfx82g",
   "tg_tokens": null,
   "tps": 68.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-f1pfx82g-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: 56nzma9h; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 951.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 951.1 70.2 22.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f1pfx82g",
   "source_url": "https://omlx.ai/benchmarks/performance/f1pfx82g",
   "tg_tokens": null,
   "tps": 70.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-a6d0ndm5-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: i73uomzv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 938.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 938.9 80.5 20.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a6d0ndm5",
   "source_url": "https://omlx.ai/benchmarks/performance/a6d0ndm5",
   "tg_tokens": null,
   "tps": 80.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-a6d0ndm5-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: i73uomzv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 988.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 988.0 77.8 20.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a6d0ndm5",
   "source_url": "https://omlx.ai/benchmarks/performance/a6d0ndm5",
   "tg_tokens": null,
   "tps": 77.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-tdl0wrv9-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: p6a68cpb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 907.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 907.7 91.0 21.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tdl0wrv9",
   "source_url": "https://omlx.ai/benchmarks/performance/tdl0wrv9",
   "tg_tokens": null,
   "tps": 91.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-tdl0wrv9-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: p6a68cpb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 954.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 954.5 85.7 22.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tdl0wrv9",
   "source_url": "https://omlx.ai/benchmarks/performance/tdl0wrv9",
   "tg_tokens": null,
   "tps": 85.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-xd15saj3-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: 2wwzrmsh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 453.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 24 GB \u00b7 4bit; 1k 453.1 49.9 5.9 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xd15saj3",
   "source_url": "https://omlx.ai/benchmarks/performance/xd15saj3",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-xd15saj3-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: 2wwzrmsh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 475.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 24 GB \u00b7 4bit; 4k 475.5 48.7 6.3 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xd15saj3",
   "source_url": "https://omlx.ai/benchmarks/performance/xd15saj3",
   "tg_tokens": null,
   "tps": 48.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-i526y8ll-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 6.0 GB at this context; duplicate traces of the same session: bkfci0g5, g17gvenw, owqs508l, rx1csoud; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 228.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-9B-oQ4e-fp16-mtp on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 1k 228.9 34.5 6.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i526y8ll",
   "source_url": "https://omlx.ai/benchmarks/performance/i526y8ll",
   "tg_tokens": null,
   "tps": 34.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-i526y8ll-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: bkfci0g5, g17gvenw, owqs508l, rx1csoud; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 251.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-9B-oQ4e-fp16-mtp on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 4k 251.5 33.5 6.5 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i526y8ll",
   "source_url": "https://omlx.ai/benchmarks/performance/i526y8ll",
   "tg_tokens": null,
   "tps": 33.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-i526y8ll-ctx8192",
   "model": "Ornith-1.5-9B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: bkfci0g5, g17gvenw, owqs508l, rx1csoud; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 249.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-9B-oQ4e-fp16-mtp on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 8k 249.5 32.2 6.7 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i526y8ll",
   "source_url": "https://omlx.ai/benchmarks/performance/i526y8ll",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-i526y8ll-ctx16384",
   "model": "Ornith-1.5-9B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 7.2 GB at this context; duplicate traces of the same session: bkfci0g5, g17gvenw, owqs508l, rx1csoud; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 238.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-9B-oQ4e-fp16-mtp on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 16k 238.6 30.1 7.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i526y8ll",
   "source_url": "https://omlx.ai/benchmarks/performance/i526y8ll",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-i526y8ll-ctx32768",
   "model": "Ornith-1.5-9B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 8.1 GB at this context; duplicate traces of the same session: bkfci0g5, g17gvenw, owqs508l, rx1csoud; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 216.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-9B-oQ4e-fp16-mtp on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 32k 216.6 26.6 8.1 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i526y8ll",
   "source_url": "https://omlx.ai/benchmarks/performance/i526y8ll",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-a6xk9sk7-ctx1024",
   "model": "Qwen2.5-Coder-0.5B-Instruct",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 1.2 GB at this context; duplicate traces of the same session: none; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3840.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen2.5-Coder-0.5B-Instruct-oQ8e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 8bit; 1k 3,840 176.4 1.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a6xk9sk7",
   "source_url": "https://omlx.ai/benchmarks/performance/a6xk9sk7",
   "tg_tokens": null,
   "tps": 176.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-a6xk9sk7-ctx4096",
   "model": "Qwen2.5-Coder-0.5B-Instruct",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 1.3 GB at this context; duplicate traces of the same session: none; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3677.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen2.5-Coder-0.5B-Instruct-oQ8e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 8bit; 4k 3,677 154.7 1.3 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a6xk9sk7",
   "source_url": "https://omlx.ai/benchmarks/performance/a6xk9sk7",
   "tg_tokens": null,
   "tps": 154.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-a6xk9sk7-ctx8192",
   "model": "Qwen2.5-Coder-0.5B-Instruct",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 1.3 GB at this context; duplicate traces of the same session: none; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3078.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen2.5-Coder-0.5B-Instruct-oQ8e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 8bit; 8k 3,078 125.5 1.3 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a6xk9sk7",
   "source_url": "https://omlx.ai/benchmarks/performance/a6xk9sk7",
   "tg_tokens": null,
   "tps": 125.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-a6xk9sk7-ctx16384",
   "model": "Qwen2.5-Coder-0.5B-Instruct",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 1.4 GB at this context; duplicate traces of the same session: none; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2350.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen2.5-Coder-0.5B-Instruct-oQ8e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 8bit; 16k 2,350 98.1 1.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a6xk9sk7",
   "source_url": "https://omlx.ai/benchmarks/performance/a6xk9sk7",
   "tg_tokens": null,
   "tps": 98.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-a6xk9sk7-ctx32768",
   "model": "Qwen2.5-Coder-0.5B-Instruct",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 1.6 GB at this context; duplicate traces of the same session: none; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1501.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen2.5-Coder-0.5B-Instruct-oQ8e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 8bit; 32k 1,501 61.9 1.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a6xk9sk7",
   "source_url": "https://omlx.ai/benchmarks/performance/a6xk9sk7",
   "tg_tokens": null,
   "tps": 61.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-zec3ndgb-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 14.9 GB at this context; duplicate traces of the same session: b0vi3hik, o51j6z6i, waouahtg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 475.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-9B-oQ6e-fp16-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 6bit; 4k 475.4 28.9 14.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zec3ndgb",
   "source_url": "https://omlx.ai/benchmarks/performance/zec3ndgb",
   "tg_tokens": null,
   "tps": 28.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-zec3ndgb-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 15.1 GB at this context; duplicate traces of the same session: b0vi3hik, o51j6z6i, waouahtg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 459.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-9B-oQ6e-fp16-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 6bit; 8k 459.7 28.0 15.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zec3ndgb",
   "source_url": "https://omlx.ai/benchmarks/performance/zec3ndgb",
   "tg_tokens": null,
   "tps": 28.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-zec3ndgb-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 15.6 GB at this context; duplicate traces of the same session: b0vi3hik, o51j6z6i, waouahtg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 417.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-9B-oQ6e-fp16-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 6bit; 16k 417.3 32.2 15.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zec3ndgb",
   "source_url": "https://omlx.ai/benchmarks/performance/zec3ndgb",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-zec3ndgb-ctx32768",
   "model": "Qwen3.5-9B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: b0vi3hik, o51j6z6i, waouahtg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 329.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.5-9B-oQ6e-fp16-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 6bit; 32k 329.5 38.7 16.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zec3ndgb",
   "source_url": "https://omlx.ai/benchmarks/performance/zec3ndgb",
   "tg_tokens": null,
   "tps": 38.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-2j5i0a3j-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: qkbfw9f4; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 532.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 532.3 65.9 19.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2j5i0a3j",
   "source_url": "https://omlx.ai/benchmarks/performance/2j5i0a3j",
   "tg_tokens": null,
   "tps": 65.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-2j5i0a3j-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: qkbfw9f4; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 614.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 614.0 63.7 20.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2j5i0a3j",
   "source_url": "https://omlx.ai/benchmarks/performance/2j5i0a3j",
   "tg_tokens": null,
   "tps": 63.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-pdxj5l2h-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: m03yurl8, oywiakvp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 861.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 861.1 108.5 35.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pdxj5l2h",
   "source_url": "https://omlx.ai/benchmarks/performance/pdxj5l2h",
   "tg_tokens": null,
   "tps": 108.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-pdxj5l2h-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: m03yurl8, oywiakvp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2809.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 2,809 106.0 36.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pdxj5l2h",
   "source_url": "https://omlx.ai/benchmarks/performance/pdxj5l2h",
   "tg_tokens": null,
   "tps": 106.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-pdxj5l2h-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 36.7 GB at this context; duplicate traces of the same session: m03yurl8, oywiakvp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4500.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 32k 4,500 93.3 36.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pdxj5l2h",
   "source_url": "https://omlx.ai/benchmarks/performance/pdxj5l2h",
   "tg_tokens": null,
   "tps": 93.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-c6usdau5-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 1jedywwk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 437.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 437.5 15.9 18.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c6usdau5",
   "source_url": "https://omlx.ai/benchmarks/performance/c6usdau5",
   "tg_tokens": null,
   "tps": 15.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-c6usdau5-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 1jedywwk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 461.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 461.2 13.9 19.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c6usdau5",
   "source_url": "https://omlx.ai/benchmarks/performance/c6usdau5",
   "tg_tokens": null,
   "tps": 13.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-kc3c1pcx-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: rmv79uxb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 461.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 461.5 18.1 18.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kc3c1pcx",
   "source_url": "https://omlx.ai/benchmarks/performance/kc3c1pcx",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-kc3c1pcx-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: rmv79uxb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 474.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 474.5 17.8 19.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kc3c1pcx",
   "source_url": "https://omlx.ai/benchmarks/performance/kc3c1pcx",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qmt0peyh-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: 999z692s; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 463.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 463.5 43.9 22.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qmt0peyh",
   "source_url": "https://omlx.ai/benchmarks/performance/qmt0peyh",
   "tg_tokens": null,
   "tps": 43.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qmt0peyh-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 999z692s; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 486.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 486.9 33.6 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qmt0peyh",
   "source_url": "https://omlx.ai/benchmarks/performance/qmt0peyh",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pemye1xf-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: blc67jh1, rivyxsgq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 768.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 768.4 41.6 26.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pemye1xf",
   "source_url": "https://omlx.ai/benchmarks/performance/pemye1xf",
   "tg_tokens": null,
   "tps": 41.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pemye1xf-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: blc67jh1, rivyxsgq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 922.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 922.6 46.3 27.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pemye1xf",
   "source_url": "https://omlx.ai/benchmarks/performance/pemye1xf",
   "tg_tokens": null,
   "tps": 46.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pemye1xf-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: blc67jh1, rivyxsgq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 844.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 844.5 66.5 29.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pemye1xf",
   "source_url": "https://omlx.ai/benchmarks/performance/pemye1xf",
   "tg_tokens": null,
   "tps": 66.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jeno8ni9-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: dti3a4w4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1024.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,024 61.3 70.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jeno8ni9",
   "source_url": "https://omlx.ai/benchmarks/performance/jeno8ni9",
   "tg_tokens": null,
   "tps": 61.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jeno8ni9-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: dti3a4w4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1933.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,933 61.9 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jeno8ni9",
   "source_url": "https://omlx.ai/benchmarks/performance/jeno8ni9",
   "tg_tokens": null,
   "tps": 61.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jeno8ni9-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: dti3a4w4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2438.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,438 68.8 73.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jeno8ni9",
   "source_url": "https://omlx.ai/benchmarks/performance/jeno8ni9",
   "tg_tokens": null,
   "tps": 68.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jeno8ni9-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.6 GB at this context; duplicate traces of the same session: dti3a4w4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2403.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,403 77.3 74.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jeno8ni9",
   "source_url": "https://omlx.ai/benchmarks/performance/jeno8ni9",
   "tg_tokens": null,
   "tps": 77.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nhyemde4-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: anfoz88z, fblvxjtc, hhs9dps7, jm3t3ac4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1190.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,190 64.2 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nhyemde4",
   "source_url": "https://omlx.ai/benchmarks/performance/nhyemde4",
   "tg_tokens": null,
   "tps": 64.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nhyemde4-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: anfoz88z, fblvxjtc, hhs9dps7, jm3t3ac4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1599.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,599 71.6 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nhyemde4",
   "source_url": "https://omlx.ai/benchmarks/performance/nhyemde4",
   "tg_tokens": null,
   "tps": 71.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nhyemde4-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: anfoz88z, fblvxjtc, hhs9dps7, jm3t3ac4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1993.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,993 73.8 72.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nhyemde4",
   "source_url": "https://omlx.ai/benchmarks/performance/nhyemde4",
   "tg_tokens": null,
   "tps": 73.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nhyemde4-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: anfoz88z, fblvxjtc, hhs9dps7, jm3t3ac4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2111.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,111 63.6 73.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nhyemde4",
   "source_url": "https://omlx.ai/benchmarks/performance/nhyemde4",
   "tg_tokens": null,
   "tps": 63.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nhyemde4-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.5 GB at this context; duplicate traces of the same session: anfoz88z, fblvxjtc, hhs9dps7, jm3t3ac4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2091.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,091 77.3 74.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nhyemde4",
   "source_url": "https://omlx.ai/benchmarks/performance/nhyemde4",
   "tg_tokens": null,
   "tps": 77.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-u5owmcp5-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.2 GB at this context; duplicate traces of the same session: dkicyvla, sez2vuye; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1044.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,044 72.2 69.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u5owmcp5",
   "source_url": "https://omlx.ai/benchmarks/performance/u5owmcp5",
   "tg_tokens": null,
   "tps": 72.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-u5owmcp5-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.4 GB at this context; duplicate traces of the same session: dkicyvla, sez2vuye; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1551.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,551 63.7 69.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u5owmcp5",
   "source_url": "https://omlx.ai/benchmarks/performance/u5owmcp5",
   "tg_tokens": null,
   "tps": 63.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-u5owmcp5-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.4 GB at this context; duplicate traces of the same session: dkicyvla, sez2vuye; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2337.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,337 57.7 72.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u5owmcp5",
   "source_url": "https://omlx.ai/benchmarks/performance/u5owmcp5",
   "tg_tokens": null,
   "tps": 57.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-udina6si-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 98.9 GB at this context; duplicate traces of the same session: culu6rh7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1833.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,833 79.5 98.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/udina6si",
   "source_url": "https://omlx.ai/benchmarks/performance/udina6si",
   "tg_tokens": null,
   "tps": 79.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-udina6si-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.5 GB at this context; duplicate traces of the same session: culu6rh7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2852.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,852 75.6 99.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/udina6si",
   "source_url": "https://omlx.ai/benchmarks/performance/udina6si",
   "tg_tokens": null,
   "tps": 75.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gqq36v0r-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 86.5 GB at this context; duplicate traces of the same session: rxx6ij88; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1066.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 1,066 56.0 86.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gqq36v0r",
   "source_url": "https://omlx.ai/benchmarks/performance/gqq36v0r",
   "tg_tokens": null,
   "tps": 56.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gqq36v0r-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 86.7 GB at this context; duplicate traces of the same session: rxx6ij88; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1776.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 1,776 62.9 86.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gqq36v0r",
   "source_url": "https://omlx.ai/benchmarks/performance/gqq36v0r",
   "tg_tokens": null,
   "tps": 62.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gpgkh55b-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: mfi2oudg, mrz2w3s4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1029.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,029 25.5 28.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gpgkh55b",
   "source_url": "https://omlx.ai/benchmarks/performance/gpgkh55b",
   "tg_tokens": null,
   "tps": 25.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gpgkh55b-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace; peak memory 29.0 GB at this context; duplicate traces of the same session: mfi2oudg, mrz2w3s4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1014.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,014 24.5 29.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gpgkh55b",
   "source_url": "https://omlx.ai/benchmarks/performance/gpgkh55b",
   "tg_tokens": null,
   "tps": 24.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gpgkh55b-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace; peak memory 36.0 GB at this context; duplicate traces of the same session: mfi2oudg, mrz2w3s4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 661.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 661.1 18.0 36.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gpgkh55b",
   "source_url": "https://omlx.ai/benchmarks/performance/gpgkh55b",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9g6kx890-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: w68xp47k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1002.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,002 25.5 22.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9g6kx890",
   "source_url": "https://omlx.ai/benchmarks/performance/9g6kx890",
   "tg_tokens": null,
   "tps": 25.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9g6kx890-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: w68xp47k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 979.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 979.3 24.4 23.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9g6kx890",
   "source_url": "https://omlx.ai/benchmarks/performance/9g6kx890",
   "tg_tokens": null,
   "tps": 24.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9g6kx890-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: w68xp47k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 646.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 646.1 18.1 30.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9g6kx890",
   "source_url": "https://omlx.ai/benchmarks/performance/9g6kx890",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-lf092ao6-ctx1024",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 9330swso; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 211.7,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 5bit; 1k 211.7 18.6 19.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lf092ao6",
   "source_url": "https://omlx.ai/benchmarks/performance/lf092ao6",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-lf092ao6-ctx4096",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 9330swso; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 214.9,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 5bit; 4k 214.9 18.2 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lf092ao6",
   "source_url": "https://omlx.ai/benchmarks/performance/lf092ao6",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-08eik32y-ctx1024",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 52x0d6d7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 213.6,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 5bit; 1k 213.6 29.3 20.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/08eik32y",
   "source_url": "https://omlx.ai/benchmarks/performance/08eik32y",
   "tg_tokens": null,
   "tps": 29.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-08eik32y-ctx4096",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 52x0d6d7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 213.6,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 5bit; 4k 213.6 45.7 21.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/08eik32y",
   "source_url": "https://omlx.ai/benchmarks/performance/08eik32y",
   "tg_tokens": null,
   "tps": 45.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-jllj6d5b-ctx1024",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: cvxtxzr7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 206.4,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 5bit; 1k 206.4 27.5 20.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jllj6d5b",
   "source_url": "https://omlx.ai/benchmarks/performance/jllj6d5b",
   "tg_tokens": null,
   "tps": 27.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-jllj6d5b-ctx4096",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: cvxtxzr7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 212.4,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 5bit; 4k 212.4 47.9 21.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jllj6d5b",
   "source_url": "https://omlx.ai/benchmarks/performance/jllj6d5b",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-4a9jkktt-ctx1024",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: c3smii2r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 209.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "AREX-2-6bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 6bit; 1k 209.9 31.6 25.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4a9jkktt",
   "source_url": "https://omlx.ai/benchmarks/performance/4a9jkktt",
   "tg_tokens": null,
   "tps": 31.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-4a9jkktt-ctx4096",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: c3smii2r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 203.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "AREX-2-6bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 6bit; 4k 203.3 41.5 27.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4a9jkktt",
   "source_url": "https://omlx.ai/benchmarks/performance/4a9jkktt",
   "tg_tokens": null,
   "tps": 41.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3wkngdlo-ctx1024",
   "model": "Huihui-Ornith-1.5-35B-A3B-abliterated",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.9 GB at this context; duplicate traces of the same session: da5f0ajm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3915.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Huihui-Ornith-1.5-35B-A3B-abliterated-oQ5e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 3,915 132.8 26.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3wkngdlo",
   "source_url": "https://omlx.ai/benchmarks/performance/3wkngdlo",
   "tg_tokens": null,
   "tps": 132.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3wkngdlo-ctx4096",
   "model": "Huihui-Ornith-1.5-35B-A3B-abliterated",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: da5f0ajm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3904.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Huihui-Ornith-1.5-35B-A3B-abliterated-oQ5e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 3,904 139.6 27.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3wkngdlo",
   "source_url": "https://omlx.ai/benchmarks/performance/3wkngdlo",
   "tg_tokens": null,
   "tps": 139.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bhu7n4bj-ctx1024",
   "model": "Huihui-Ornith-1.5-35B-A3B-abliterated",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: dzkkypwb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3936.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Huihui-Ornith-1.5-35B-A3B-abliterated-oQ5e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 3,936 132.3 26.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhu7n4bj",
   "source_url": "https://omlx.ai/benchmarks/performance/bhu7n4bj",
   "tg_tokens": null,
   "tps": 132.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bhu7n4bj-ctx4096",
   "model": "Huihui-Ornith-1.5-35B-A3B-abliterated",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: dzkkypwb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5332.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Huihui-Ornith-1.5-35B-A3B-abliterated-oQ5e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 5,332 141.2 27.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhu7n4bj",
   "source_url": "https://omlx.ai/benchmarks/performance/bhu7n4bj",
   "tg_tokens": null,
   "tps": 141.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wyhn7dfd-ctx1024",
   "model": "Huihui-Ornith-1.5-35B-A3B-abliterated",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: 90ofugmz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3906.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Huihui-Ornith-1.5-35B-A3B-abliterated-oQ5e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 3,906 134.8 26.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wyhn7dfd",
   "source_url": "https://omlx.ai/benchmarks/performance/wyhn7dfd",
   "tg_tokens": null,
   "tps": 134.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wyhn7dfd-ctx4096",
   "model": "Huihui-Ornith-1.5-35B-A3B-abliterated",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: 90ofugmz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5315.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Huihui-Ornith-1.5-35B-A3B-abliterated-oQ5e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 5,315 145.6 27.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wyhn7dfd",
   "source_url": "https://omlx.ai/benchmarks/performance/wyhn7dfd",
   "tg_tokens": null,
   "tps": 145.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-8o4zpxi6-ctx1024",
   "model": "Hy-MT2-30B-A3B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 10.6 GB at this context; duplicate traces of the same session: gzmu1kp8, i9o6r55s; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1250.0,
   "provenance": "community",
   "quant": "oQ2",
   "quote": "Hy-MT2-30B-A3B-oQ2-MLX on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 2bit; 1k 1,250 80.0 10.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8o4zpxi6",
   "source_url": "https://omlx.ai/benchmarks/performance/8o4zpxi6",
   "tg_tokens": null,
   "tps": 80.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-8o4zpxi6-ctx4096",
   "model": "Hy-MT2-30B-A3B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 11.1 GB at this context; duplicate traces of the same session: gzmu1kp8, i9o6r55s; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1235.0,
   "provenance": "community",
   "quant": "oQ2",
   "quote": "Hy-MT2-30B-A3B-oQ2-MLX on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 2bit; 4k 1,235 69.6 11.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8o4zpxi6",
   "source_url": "https://omlx.ai/benchmarks/performance/8o4zpxi6",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-8o4zpxi6-ctx8192",
   "model": "Hy-MT2-30B-A3B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: gzmu1kp8, i9o6r55s; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1086.0,
   "provenance": "community",
   "quant": "oQ2",
   "quote": "Hy-MT2-30B-A3B-oQ2-MLX on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 2bit; 8k 1,086 68.3 11.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8o4zpxi6",
   "source_url": "https://omlx.ai/benchmarks/performance/8o4zpxi6",
   "tg_tokens": null,
   "tps": 68.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-hw7u2aqx-ctx4096",
   "model": "Index-Translate-2B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 4k context; peak memory 2.2 GB; TTFT 1642 ms; duplicate traces of the same session: none; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2495.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Index-Translate-2B-MLX-4bit on M5 (8c); M5 (8c) \u00b7 16 GB \u00b7 4bit; 4k tokens 2,495 PP tok/s 98.2 TG tok/s 1642 TTFT (ms) 2.2 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hw7u2aqx",
   "source_url": "https://omlx.ai/benchmarks/performance/hw7u2aqx",
   "tg_tokens": null,
   "tps": 98.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-amtqn7yb-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: yqmczlty, zkw9e865; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3126.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 3,126 85.6 24.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/amtqn7yb",
   "source_url": "https://omlx.ai/benchmarks/performance/amtqn7yb",
   "tg_tokens": null,
   "tps": 85.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-amtqn7yb-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: yqmczlty, zkw9e865; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3457.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 3,457 84.4 24.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/amtqn7yb",
   "source_url": "https://omlx.ai/benchmarks/performance/amtqn7yb",
   "tg_tokens": null,
   "tps": 84.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-amtqn7yb-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: yqmczlty, zkw9e865; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4347.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 4,347 81.7 24.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/amtqn7yb",
   "source_url": "https://omlx.ai/benchmarks/performance/amtqn7yb",
   "tg_tokens": null,
   "tps": 81.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-amtqn7yb-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: yqmczlty, zkw9e865; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4178.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 16k 4,178 76.5 25.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/amtqn7yb",
   "source_url": "https://omlx.ai/benchmarks/performance/amtqn7yb",
   "tg_tokens": null,
   "tps": 76.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-amtqn7yb-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: yqmczlty, zkw9e865; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3825.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 32k 3,825 67.5 25.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/amtqn7yb",
   "source_url": "https://omlx.ai/benchmarks/performance/amtqn7yb",
   "tg_tokens": null,
   "tps": 67.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-amtqn7yb-ctx65536",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: yqmczlty, zkw9e865; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3013.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 64k 3,013 54.3 26.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/amtqn7yb",
   "source_url": "https://omlx.ai/benchmarks/performance/amtqn7yb",
   "tg_tokens": null,
   "tps": 54.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-uswizad0-ctx1024",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 10.0 GB at this context; duplicate traces of the same session: 36x2c5c6, 3jc64oqj, 9rtuds44, 9shmxc1z; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 242.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-oQ8e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 8bit; 1k 242.0 19.1 10.0 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uswizad0",
   "source_url": "https://omlx.ai/benchmarks/performance/uswizad0",
   "tg_tokens": null,
   "tps": 19.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-uswizad0-ctx4096",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 10.6 GB at this context; duplicate traces of the same session: 36x2c5c6, 3jc64oqj, 9rtuds44, 9shmxc1z; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 228.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-oQ8e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 8bit; 4k 228.0 17.3 10.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uswizad0",
   "source_url": "https://omlx.ai/benchmarks/performance/uswizad0",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-uswizad0-ctx8192",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 10.8 GB at this context; duplicate traces of the same session: 36x2c5c6, 3jc64oqj, 9rtuds44, 9shmxc1z; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 235.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-oQ8e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 8bit; 8k 235.8 19.2 10.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uswizad0",
   "source_url": "https://omlx.ai/benchmarks/performance/uswizad0",
   "tg_tokens": null,
   "tps": 19.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-uswizad0-ctx16384",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 11.2 GB at this context; duplicate traces of the same session: 36x2c5c6, 3jc64oqj, 9rtuds44, 9shmxc1z; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 227.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-oQ8e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 8bit; 16k 227.2 18.4 11.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uswizad0",
   "source_url": "https://omlx.ai/benchmarks/performance/uswizad0",
   "tg_tokens": null,
   "tps": 18.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-uswizad0-ctx32768",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 12.1 GB at this context; duplicate traces of the same session: 36x2c5c6, 3jc64oqj, 9rtuds44, 9shmxc1z; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 206.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-oQ8e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 8bit; 32k 206.1 17.0 12.1 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uswizad0",
   "source_url": "https://omlx.ai/benchmarks/performance/uswizad0",
   "tg_tokens": null,
   "tps": 17.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-weconks7-ctx1024",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "MLX REAP50 MTP build; oMLX community benchmark trace, single run at 1k context; peak memory 82.3 GB; TTFT 792 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1293.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "MiMo-V2.6-Flash-RL-MLX-REAP50-mxfp4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k tokens 1,293 PP tok/s 47.4 TG tok/s 792 TTFT (ms) 82.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/weconks7",
   "source_url": "https://omlx.ai/benchmarks/performance/weconks7",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-tqxjthbb-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: tzh2rh4d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4386.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,386 103.2 21.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tqxjthbb",
   "source_url": "https://omlx.ai/benchmarks/performance/tqxjthbb",
   "tg_tokens": null,
   "tps": 103.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-tqxjthbb-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: tzh2rh4d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6072.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,072 100.4 22.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tqxjthbb",
   "source_url": "https://omlx.ai/benchmarks/performance/tqxjthbb",
   "tg_tokens": null,
   "tps": 100.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-b23dkjx2-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: i6xhffh7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4201.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,201 101.5 21.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b23dkjx2",
   "source_url": "https://omlx.ai/benchmarks/performance/b23dkjx2",
   "tg_tokens": null,
   "tps": 101.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-b23dkjx2-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: i6xhffh7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5115.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 5,115 98.8 22.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b23dkjx2",
   "source_url": "https://omlx.ai/benchmarks/performance/b23dkjx2",
   "tg_tokens": null,
   "tps": 98.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-c5ypfl4x-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: m2eeugpf, rocn6spi; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1842.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 1,842 112.5 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c5ypfl4x",
   "source_url": "https://omlx.ai/benchmarks/performance/c5ypfl4x",
   "tg_tokens": null,
   "tps": 112.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-c5ypfl4x-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: m2eeugpf, rocn6spi; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2289.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 2,289 129.9 24.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c5ypfl4x",
   "source_url": "https://omlx.ai/benchmarks/performance/c5ypfl4x",
   "tg_tokens": null,
   "tps": 129.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-c5ypfl4x-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: m2eeugpf, rocn6spi; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2148.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 8k 2,148 105.0 24.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c5ypfl4x",
   "source_url": "https://omlx.ai/benchmarks/performance/c5ypfl4x",
   "tg_tokens": null,
   "tps": 105.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-kjgg8v4s-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: vmqsk1gf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 786.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 786.0 61.2 21.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kjgg8v4s",
   "source_url": "https://omlx.ai/benchmarks/performance/kjgg8v4s",
   "tg_tokens": null,
   "tps": 61.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-kjgg8v4s-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: vmqsk1gf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 954.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 954.5 90.5 22.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kjgg8v4s",
   "source_url": "https://omlx.ai/benchmarks/performance/kjgg8v4s",
   "tg_tokens": null,
   "tps": 90.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1cbsci68-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.1 GB at this context; duplicate traces of the same session: gp2jjzbe; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2961.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,961 82.5 43.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1cbsci68",
   "source_url": "https://omlx.ai/benchmarks/performance/1cbsci68",
   "tg_tokens": null,
   "tps": 82.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1cbsci68-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.9 GB at this context; duplicate traces of the same session: gp2jjzbe; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4107.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 4,107 79.7 43.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1cbsci68",
   "source_url": "https://omlx.ai/benchmarks/performance/1cbsci68",
   "tg_tokens": null,
   "tps": 79.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-j91tzubo-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 6.7 GB; TTFT 370 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2766.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 2,766 PP tok/s 142.5 TG tok/s 370 TTFT (ms) 6.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j91tzubo",
   "source_url": "https://omlx.ai/benchmarks/performance/j91tzubo",
   "tg_tokens": null,
   "tps": 142.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-w34lgc6a-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: wjn3j10c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2300.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,300 166.7 22.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w34lgc6a",
   "source_url": "https://omlx.ai/benchmarks/performance/w34lgc6a",
   "tg_tokens": null,
   "tps": 166.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-w34lgc6a-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: wjn3j10c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4440.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 4,440 198.6 22.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w34lgc6a",
   "source_url": "https://omlx.ai/benchmarks/performance/w34lgc6a",
   "tg_tokens": null,
   "tps": 198.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ul1lsqo7-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 0f4rwxrd; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 132.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 32 GB \u00b7 4bit; 1k 132.6 20.0 16.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ul1lsqo7",
   "source_url": "https://omlx.ai/benchmarks/performance/ul1lsqo7",
   "tg_tokens": null,
   "tps": 20.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-ul1lsqo7-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: 0f4rwxrd; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 32 GB \u00b7 4bit; 4k 130.2 18.6 17.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ul1lsqo7",
   "source_url": "https://omlx.ai/benchmarks/performance/ul1lsqo7",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9opwpemi-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.1 GB at this context; duplicate traces of the same session: p9wxyoy0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 888.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 888.6 40.8 35.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9opwpemi",
   "source_url": "https://omlx.ai/benchmarks/performance/9opwpemi",
   "tg_tokens": null,
   "tps": 40.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9opwpemi-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.8 GB at this context; duplicate traces of the same session: p9wxyoy0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 842.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 842.6 56.5 36.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9opwpemi",
   "source_url": "https://omlx.ai/benchmarks/performance/9opwpemi",
   "tg_tokens": null,
   "tps": 56.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-0corharg-ctx16384",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 41.5 GB at this context; duplicate traces of the same session: 6742mq9e; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1221.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 16k 1,221 45.2 41.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0corharg",
   "source_url": "https://omlx.ai/benchmarks/performance/0corharg",
   "tg_tokens": null,
   "tps": 45.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-0corharg-ctx32768",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 42.0 GB at this context; duplicate traces of the same session: 6742mq9e; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1120.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 32k 1,120 41.9 42.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0corharg",
   "source_url": "https://omlx.ai/benchmarks/performance/0corharg",
   "tg_tokens": null,
   "tps": 41.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-l4gl492s-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 69.3 GB; TTFT 4416 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 927.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 927.5 PP tok/s 39.3 TG tok/s 4416 TTFT (ms) 69.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l4gl492s",
   "source_url": "https://omlx.ai/benchmarks/performance/l4gl492s",
   "tg_tokens": null,
   "tps": 39.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-6wucuxj5-ctx1024",
   "model": "Ternary-Bonsai-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 9.1 GB at this context; duplicate traces of the same session: wtrdh7lt; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 184.2,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-27B-mlx-2bit on M2 Max (38c); M2 Max (38c) \u00b7 32 GB \u00b7 2bit; 1k 184.2 23.0 9.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6wucuxj5",
   "source_url": "https://omlx.ai/benchmarks/performance/6wucuxj5",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-6wucuxj5-ctx4096",
   "model": "Ternary-Bonsai-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 10.2 GB at this context; duplicate traces of the same session: wtrdh7lt; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 197.6,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-27B-mlx-2bit on M2 Max (38c); M2 Max (38c) \u00b7 32 GB \u00b7 2bit; 4k 197.6 30.5 10.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6wucuxj5",
   "source_url": "https://omlx.ai/benchmarks/performance/6wucuxj5",
   "tg_tokens": null,
   "tps": 30.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-87up0bvc-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1002.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,002 25.5 22.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/87up0bvc",
   "source_url": "https://omlx.ai/benchmarks/performance/87up0bvc",
   "tg_tokens": null,
   "tps": 25.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-87up0bvc-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 979.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 979.3 24.4 23.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/87up0bvc",
   "source_url": "https://omlx.ai/benchmarks/performance/87up0bvc",
   "tg_tokens": null,
   "tps": 24.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-87up0bvc-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 646.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 646.1 18.1 30.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/87up0bvc",
   "source_url": "https://omlx.ai/benchmarks/performance/87up0bvc",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-uovah7q5-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: 33naxdjj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2130.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,130 77.4 16.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uovah7q5",
   "source_url": "https://omlx.ai/benchmarks/performance/uovah7q5",
   "tg_tokens": null,
   "tps": 77.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-uovah7q5-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 33naxdjj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2315.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,315 74.3 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uovah7q5",
   "source_url": "https://omlx.ai/benchmarks/performance/uovah7q5",
   "tg_tokens": null,
   "tps": 74.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-6wz9rim5-ctx1024",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 2tsjbnrv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 210.7,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 5bit; 1k 210.7 26.4 20.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6wz9rim5",
   "source_url": "https://omlx.ai/benchmarks/performance/6wz9rim5",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-6wz9rim5-ctx4096",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 2tsjbnrv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 208.9,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 5bit; 4k 208.9 45.1 22.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6wz9rim5",
   "source_url": "https://omlx.ai/benchmarks/performance/6wz9rim5",
   "tg_tokens": null,
   "tps": 45.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-14mteq3r-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 150.8 GB at this context; duplicate traces of the same session: fsg8gpza, orec5crp, r1x6pn8o; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 198.5,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "GLM-5.3-Flash-oQ3.5e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 3bit; 1k 198.5 27.6 150.8 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/14mteq3r",
   "source_url": "https://omlx.ai/benchmarks/performance/14mteq3r",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-14mteq3r-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 151.8 GB at this context; duplicate traces of the same session: fsg8gpza, orec5crp, r1x6pn8o; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 246.7,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "GLM-5.3-Flash-oQ3.5e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 3bit; 4k 246.7 26.4 151.8 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/14mteq3r",
   "source_url": "https://omlx.ai/benchmarks/performance/14mteq3r",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-14mteq3r-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 152.1 GB at this context; duplicate traces of the same session: fsg8gpza, orec5crp, r1x6pn8o; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 215.8,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "GLM-5.3-Flash-oQ3.5e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 3bit; 8k 215.8 29.4 152.1 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/14mteq3r",
   "source_url": "https://omlx.ai/benchmarks/performance/14mteq3r",
   "tg_tokens": null,
   "tps": 29.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-14mteq3r-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 152.2 GB at this context; duplicate traces of the same session: fsg8gpza, orec5crp, r1x6pn8o; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 210.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "GLM-5.3-Flash-oQ3.5e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 3bit; 16k 210.0 33.3 152.2 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/14mteq3r",
   "source_url": "https://omlx.ai/benchmarks/performance/14mteq3r",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yndq4tdb-ctx1024",
   "model": "Huihui-Ornith-1.5-35B-A3B-abliterated",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.9 GB at this context; duplicate traces of the same session: t5ckhd4t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1273.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Huihui-Ornith-1.5-35B-A3B-abliterated-oQ5e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 1,273 56.4 26.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yndq4tdb",
   "source_url": "https://omlx.ai/benchmarks/performance/yndq4tdb",
   "tg_tokens": null,
   "tps": 56.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yndq4tdb-ctx4096",
   "model": "Huihui-Ornith-1.5-35B-A3B-abliterated",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: t5ckhd4t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3312.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Huihui-Ornith-1.5-35B-A3B-abliterated-oQ5e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 3,312 126.9 27.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yndq4tdb",
   "source_url": "https://omlx.ai/benchmarks/performance/yndq4tdb",
   "tg_tokens": null,
   "tps": 126.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-exk4zftg-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 8k context; peak memory 13.8 GB; TTFT 15994 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 512.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 8k tokens 512.2 PP tok/s 57.2 TG tok/s 15994 TTFT (ms) 13.8 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/exk4zftg",
   "source_url": "https://omlx.ai/benchmarks/performance/exk4zftg",
   "tg_tokens": null,
   "tps": 57.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-1xjr72ho-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: 3dpim9yp, ka81iidd; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3126.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 3,126 85.6 24.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1xjr72ho",
   "source_url": "https://omlx.ai/benchmarks/performance/1xjr72ho",
   "tg_tokens": null,
   "tps": 85.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-1xjr72ho-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: 3dpim9yp, ka81iidd; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3457.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 3,457 84.4 24.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1xjr72ho",
   "source_url": "https://omlx.ai/benchmarks/performance/1xjr72ho",
   "tg_tokens": null,
   "tps": 84.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-1xjr72ho-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: 3dpim9yp, ka81iidd; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4347.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 4,347 81.7 24.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1xjr72ho",
   "source_url": "https://omlx.ai/benchmarks/performance/1xjr72ho",
   "tg_tokens": null,
   "tps": 81.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-1xjr72ho-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: 3dpim9yp, ka81iidd; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4178.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 16k 4,178 76.5 25.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1xjr72ho",
   "source_url": "https://omlx.ai/benchmarks/performance/1xjr72ho",
   "tg_tokens": null,
   "tps": 76.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-1xjr72ho-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: 3dpim9yp, ka81iidd; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3825.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 32k 3,825 67.5 25.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1xjr72ho",
   "source_url": "https://omlx.ai/benchmarks/performance/1xjr72ho",
   "tg_tokens": null,
   "tps": 67.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1xjr72ho-ctx65536",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: 3dpim9yp, ka81iidd; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3013.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 64k 3,013 54.3 26.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1xjr72ho",
   "source_url": "https://omlx.ai/benchmarks/performance/1xjr72ho",
   "tg_tokens": null,
   "tps": 54.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-dziarxhu-ctx1024",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 6.0 GB at this context; duplicate traces of the same session: 04i1ugy4, 2idaimg3, 2tem2o52, 92v2lzxg, k0290vz9; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 228.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 1k 228.5 33.9 6.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dziarxhu",
   "source_url": "https://omlx.ai/benchmarks/performance/dziarxhu",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-dziarxhu-ctx4096",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: 04i1ugy4, 2idaimg3, 2tem2o52, 92v2lzxg, k0290vz9; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 247.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 4k 247.3 32.7 6.5 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dziarxhu",
   "source_url": "https://omlx.ai/benchmarks/performance/dziarxhu",
   "tg_tokens": null,
   "tps": 32.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-dziarxhu-ctx8192",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: 04i1ugy4, 2idaimg3, 2tem2o52, 92v2lzxg, k0290vz9; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 248.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 8k 248.8 32.3 6.7 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dziarxhu",
   "source_url": "https://omlx.ai/benchmarks/performance/dziarxhu",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-dziarxhu-ctx16384",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 7.2 GB at this context; duplicate traces of the same session: 04i1ugy4, 2idaimg3, 2tem2o52, 92v2lzxg, k0290vz9; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 237.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 16k 237.9 30.1 7.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dziarxhu",
   "source_url": "https://omlx.ai/benchmarks/performance/dziarxhu",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-dziarxhu-ctx32768",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 8.1 GB at this context; duplicate traces of the same session: 04i1ugy4, 2idaimg3, 2tem2o52, 92v2lzxg, k0290vz9; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 32k 205.5 24.3 8.1 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dziarxhu",
   "source_url": "https://omlx.ai/benchmarks/performance/dziarxhu",
   "tg_tokens": null,
   "tps": 24.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-dziarxhu-ctx65536",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 9.9 GB at this context; duplicate traces of the same session: 04i1ugy4, 2idaimg3, 2tem2o52, 92v2lzxg, k0290vz9; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 177.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 64k 177.2 21.7 9.9 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dziarxhu",
   "source_url": "https://omlx.ai/benchmarks/performance/dziarxhu",
   "tg_tokens": null,
   "tps": 21.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-7pwa9req-ctx1024",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "MLX REAP50 MTP build; oMLX community benchmark trace; peak memory 82.3 GB at this context; duplicate traces of the same session: owatt704, vyugrp71, w91qks28; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 579.6,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "MiMo-V2.6-Flash-RL-MLX-REAP50-mxfp4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 579.6 47.4 82.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7pwa9req",
   "source_url": "https://omlx.ai/benchmarks/performance/7pwa9req",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-7pwa9req-ctx4096",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "MLX REAP50 MTP build; oMLX community benchmark trace; peak memory 83.6 GB at this context; duplicate traces of the same session: owatt704, vyugrp71, w91qks28; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1300.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "MiMo-V2.6-Flash-RL-MLX-REAP50-mxfp4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 1,300 46.4 83.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7pwa9req",
   "source_url": "https://omlx.ai/benchmarks/performance/7pwa9req",
   "tg_tokens": null,
   "tps": 46.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-7pwa9req-ctx8192",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "MLX REAP50 MTP build; oMLX community benchmark trace; peak memory 84.9 GB at this context; duplicate traces of the same session: owatt704, vyugrp71, w91qks28; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1753.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "MiMo-V2.6-Flash-RL-MLX-REAP50-mxfp4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 1,753 46.0 84.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7pwa9req",
   "source_url": "https://omlx.ai/benchmarks/performance/7pwa9req",
   "tg_tokens": null,
   "tps": 46.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-7pwa9req-ctx16384",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "MLX REAP50 MTP build; oMLX community benchmark trace; peak memory 85.3 GB at this context; duplicate traces of the same session: owatt704, vyugrp71, w91qks28; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1461.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "MiMo-V2.6-Flash-RL-MLX-REAP50-mxfp4-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 1,461 44.9 85.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7pwa9req",
   "source_url": "https://omlx.ai/benchmarks/performance/7pwa9req",
   "tg_tokens": null,
   "tps": 44.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-1cphzl35-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: 14rcc6hk, 86o989be; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 234.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-9B-oQ4e on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 234.2 21.2 5.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1cphzl35",
   "source_url": "https://omlx.ai/benchmarks/performance/1cphzl35",
   "tg_tokens": null,
   "tps": 21.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-1cphzl35-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: 14rcc6hk, 86o989be; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 240.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-9B-oQ4e on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 240.9 20.6 6.3 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1cphzl35",
   "source_url": "https://omlx.ai/benchmarks/performance/1cphzl35",
   "tg_tokens": null,
   "tps": 20.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-1cphzl35-ctx8192",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: 14rcc6hk, 86o989be; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 239.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-9B-oQ4e on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 8k 239.2 19.9 6.4 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1cphzl35",
   "source_url": "https://omlx.ai/benchmarks/performance/1cphzl35",
   "tg_tokens": null,
   "tps": 19.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-p6wr31ed-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: vnp4hdqw, wu3o8eij; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 225.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-9B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 225.3 29.5 7.0 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p6wr31ed",
   "source_url": "https://omlx.ai/benchmarks/performance/p6wr31ed",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-p6wr31ed-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.5 GB at this context; duplicate traces of the same session: vnp4hdqw, wu3o8eij; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 231.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-9B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 231.3 27.0 7.5 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p6wr31ed",
   "source_url": "https://omlx.ai/benchmarks/performance/p6wr31ed",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-p6wr31ed-ctx8192",
   "model": "Ornith-1.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.7 GB at this context; duplicate traces of the same session: vnp4hdqw, wu3o8eij; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 229.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-9B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 8k 229.1 27.2 7.7 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p6wr31ed",
   "source_url": "https://omlx.ai/benchmarks/performance/p6wr31ed",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-wrexw711-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: rpf90uwb, x6ja98zd; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 414.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-4B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 414.8 50.4 4.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wrexw711",
   "source_url": "https://omlx.ai/benchmarks/performance/wrexw711",
   "tg_tokens": null,
   "tps": 50.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-wrexw711-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.7 GB at this context; duplicate traces of the same session: rpf90uwb, x6ja98zd; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 419.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-4B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 419.9 45.4 4.7 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wrexw711",
   "source_url": "https://omlx.ai/benchmarks/performance/wrexw711",
   "tg_tokens": null,
   "tps": 45.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-wrexw711-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 5.0 GB at this context; duplicate traces of the same session: rpf90uwb, x6ja98zd; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 408.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-4B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 8k 408.1 52.9 5.0 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wrexw711",
   "source_url": "https://omlx.ai/benchmarks/performance/wrexw711",
   "tg_tokens": null,
   "tps": 52.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-dem97abd-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 17.0 GB; TTFT 1874 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 546.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k tokens 546.4 PP tok/s 27.7 TG tok/s 1874 TTFT (ms) 17.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dem97abd",
   "source_url": "https://omlx.ai/benchmarks/performance/dem97abd",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-0km5bh3m-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: n0mjjym1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 212.3,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 212.3 18.1 21.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0km5bh3m",
   "source_url": "https://omlx.ai/benchmarks/performance/0km5bh3m",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-0km5bh3m-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: n0mjjym1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 165.7,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 165.7 8.3 21.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0km5bh3m",
   "source_url": "https://omlx.ai/benchmarks/performance/0km5bh3m",
   "tg_tokens": null,
   "tps": 8.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3p9zkrn8-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "DWQ MTP MLX build; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: 7y479qdj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 487.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-DWQ-MTP-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 487.3 26.1 22.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3p9zkrn8",
   "source_url": "https://omlx.ai/benchmarks/performance/3p9zkrn8",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3p9zkrn8-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "DWQ MTP MLX build; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: 7y479qdj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 466.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-DWQ-MTP-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 466.7 30.8 23.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3p9zkrn8",
   "source_url": "https://omlx.ai/benchmarks/performance/3p9zkrn8",
   "tg_tokens": null,
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-d9xw2mzc-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 62.8 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 128k 205.1 43.9 62.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d9xw2mzc",
   "source_url": "https://omlx.ai/benchmarks/performance/d9xw2mzc",
   "tg_tokens": null,
   "tps": 43.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-d9xw2mzc-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 53.0 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 167.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 195k 167.3 30.8 53.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d9xw2mzc",
   "source_url": "https://omlx.ai/benchmarks/performance/d9xw2mzc",
   "tg_tokens": null,
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dpin5tj7-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 75.3 GB at this context; duplicate traces of the same session: g210q6f7, h6lq12nk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1091.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,091 46.6 75.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dpin5tj7",
   "source_url": "https://omlx.ai/benchmarks/performance/dpin5tj7",
   "tg_tokens": null,
   "tps": 46.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dpin5tj7-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 77.2 GB at this context; duplicate traces of the same session: g210q6f7, h6lq12nk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2407.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,407 45.1 77.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dpin5tj7",
   "source_url": "https://omlx.ai/benchmarks/performance/dpin5tj7",
   "tg_tokens": null,
   "tps": 45.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dpin5tj7-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 80.9 GB at this context; duplicate traces of the same session: g210q6f7, h6lq12nk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2329.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,329 38.5 80.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dpin5tj7",
   "source_url": "https://omlx.ai/benchmarks/performance/dpin5tj7",
   "tg_tokens": null,
   "tps": 38.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-00odjsgy-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 75.3 GB at this context; duplicate traces of the same session: sujivsxe; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1114.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,114 45.7 75.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/00odjsgy",
   "source_url": "https://omlx.ai/benchmarks/performance/00odjsgy",
   "tg_tokens": null,
   "tps": 45.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-00odjsgy-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 77.2 GB at this context; duplicate traces of the same session: sujivsxe; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2389.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,389 33.1 77.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/00odjsgy",
   "source_url": "https://omlx.ai/benchmarks/performance/00odjsgy",
   "tg_tokens": null,
   "tps": 33.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-g4qmt1pm-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 75.3 GB; TTFT 1810 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 565.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 565.9 PP tok/s 40.7 TG tok/s 1810 TTFT (ms) 75.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g4qmt1pm",
   "source_url": "https://omlx.ai/benchmarks/performance/g4qmt1pm",
   "tg_tokens": null,
   "tps": 40.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wdf22lwz-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 75.3 GB at this context; duplicate traces of the same session: f74rv2gc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1135.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,135 41.4 75.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wdf22lwz",
   "source_url": "https://omlx.ai/benchmarks/performance/wdf22lwz",
   "tg_tokens": null,
   "tps": 41.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wdf22lwz-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 75.5 GB at this context; duplicate traces of the same session: f74rv2gc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1156.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,156 35.1 75.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wdf22lwz",
   "source_url": "https://omlx.ai/benchmarks/performance/wdf22lwz",
   "tg_tokens": null,
   "tps": 35.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-k30ca0jp-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 75.3 GB; TTFT 1652 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 620.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 620.1 PP tok/s 40.3 TG tok/s 1652 TTFT (ms) 75.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k30ca0jp",
   "source_url": "https://omlx.ai/benchmarks/performance/k30ca0jp",
   "tg_tokens": null,
   "tps": 40.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-5highzro-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 187.3 GB at this context; duplicate traces of the same session: uvzewen5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 579.4,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ8-MTP on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 1k 579.4 85.2 187.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5highzro",
   "source_url": "https://omlx.ai/benchmarks/performance/5highzro",
   "tg_tokens": null,
   "tps": 85.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-5highzro-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 187.8 GB at this context; duplicate traces of the same session: uvzewen5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1612.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ8-MTP on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 4k 1,612 62.4 187.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5highzro",
   "source_url": "https://omlx.ai/benchmarks/performance/5highzro",
   "tg_tokens": null,
   "tps": 62.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i0ib4jbw-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: klk1rhfe; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2326.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,326 62.4 73.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i0ib4jbw",
   "source_url": "https://omlx.ai/benchmarks/performance/i0ib4jbw",
   "tg_tokens": null,
   "tps": 62.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i0ib4jbw-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 75.0 GB at this context; duplicate traces of the same session: klk1rhfe; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2065.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,065 65.4 75.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i0ib4jbw",
   "source_url": "https://omlx.ai/benchmarks/performance/i0ib4jbw",
   "tg_tokens": null,
   "tps": 65.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-u4zczabr-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: 15u834na, ckixds77, x9qnhmpv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 375.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 375.9 71.8 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u4zczabr",
   "source_url": "https://omlx.ai/benchmarks/performance/u4zczabr",
   "tg_tokens": null,
   "tps": 71.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-u4zczabr-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 15u834na, ckixds77, x9qnhmpv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1198.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,198 64.6 71.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u4zczabr",
   "source_url": "https://omlx.ai/benchmarks/performance/u4zczabr",
   "tg_tokens": null,
   "tps": 64.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-u4zczabr-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: 15u834na, ckixds77, x9qnhmpv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2429.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,429 69.8 72.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u4zczabr",
   "source_url": "https://omlx.ai/benchmarks/performance/u4zczabr",
   "tg_tokens": null,
   "tps": 69.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-u4zczabr-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.7 GB at this context; duplicate traces of the same session: 15u834na, ckixds77, x9qnhmpv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2392.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,392 66.6 73.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u4zczabr",
   "source_url": "https://omlx.ai/benchmarks/performance/u4zczabr",
   "tg_tokens": null,
   "tps": 66.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-uc645j7m-ctx4096",
   "model": "Spark-X2.5-4B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: 19rhctld; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1217.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Spark-X2.5-4B-MLX-8bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 8bit; 4k 1,217 69.4 4.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uc645j7m",
   "source_url": "https://omlx.ai/benchmarks/performance/uc645j7m",
   "tg_tokens": null,
   "tps": 69.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-uc645j7m-ctx16384",
   "model": "Spark-X2.5-4B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 5.3 GB at this context; duplicate traces of the same session: 19rhctld; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1082.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Spark-X2.5-4B-MLX-8bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 8bit; 16k 1,082 63.7 5.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uc645j7m",
   "source_url": "https://omlx.ai/benchmarks/performance/uc645j7m",
   "tg_tokens": null,
   "tps": 63.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-w4g16imn-ctx1024",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: c2p0j2j2; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 54.1,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 2bit; 1k 54.1 8.2 11.5 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w4g16imn",
   "source_url": "https://omlx.ai/benchmarks/performance/w4g16imn",
   "tg_tokens": null,
   "tps": 8.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-w4g16imn-ctx4096",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 14.0 GB at this context; duplicate traces of the same session: c2p0j2j2; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 53.9,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 2bit; 4k 53.9 8.1 14.0 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w4g16imn",
   "source_url": "https://omlx.ai/benchmarks/performance/w4g16imn",
   "tg_tokens": null,
   "tps": 8.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-f0c3bczk-ctx8192",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 15.0 GB at this context; duplicate traces of the same session: aju5esh6; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1152.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 1,152 79.8 15.0 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f0c3bczk",
   "source_url": "https://omlx.ai/benchmarks/performance/f0c3bczk",
   "tg_tokens": null,
   "tps": 79.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-f0c3bczk-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 15.4 GB at this context; duplicate traces of the same session: aju5esh6; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1116.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 1,116 75.2 15.4 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f0c3bczk",
   "source_url": "https://omlx.ai/benchmarks/performance/f0c3bczk",
   "tg_tokens": null,
   "tps": 75.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dlbrflqd-ctx4096",
   "model": "Huihui-gemma-4-26B-A4B-it",
   "notes": "XL mlx build; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: 2ncrlxk1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3999.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Huihui-gemma-4-26B-A4B-it-8bit-XL-mlx on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 3,999 85.2 26.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dlbrflqd",
   "source_url": "https://omlx.ai/benchmarks/performance/dlbrflqd",
   "tg_tokens": null,
   "tps": 85.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dlbrflqd-ctx16384",
   "model": "Huihui-gemma-4-26B-A4B-it",
   "notes": "XL mlx build; oMLX community benchmark trace; peak memory 27.2 GB at this context; duplicate traces of the same session: 2ncrlxk1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3860.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Huihui-gemma-4-26B-A4B-it-8bit-XL-mlx on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 16k 3,860 76.9 27.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dlbrflqd",
   "source_url": "https://omlx.ai/benchmarks/performance/dlbrflqd",
   "tg_tokens": null,
   "tps": 76.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-28fb9nv1-ctx4096",
   "model": "Huihui-gemma-4-26B-A4B-it",
   "notes": "XL mlx build; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: 2m2afqz4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2089.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Huihui-gemma-4-26B-A4B-it-8bit-XL-mlx on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 2,089 85.2 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/28fb9nv1",
   "source_url": "https://omlx.ai/benchmarks/performance/28fb9nv1",
   "tg_tokens": null,
   "tps": 85.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-28fb9nv1-ctx16384",
   "model": "Huihui-gemma-4-26B-A4B-it",
   "notes": "XL mlx build; oMLX community benchmark trace; peak memory 27.2 GB at this context; duplicate traces of the same session: 2m2afqz4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3804.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Huihui-gemma-4-26B-A4B-it-8bit-XL-mlx on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 16k 3,804 76.2 27.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/28fb9nv1",
   "source_url": "https://omlx.ai/benchmarks/performance/28fb9nv1",
   "tg_tokens": null,
   "tps": 76.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-31uaymyl-ctx16384",
   "model": "LFM2.5-2.6B",
   "notes": "oMLX community benchmark trace; peak memory 7.2 GB at this context; duplicate traces of the same session: dia6n8br, mdt6j316, nyovpsl7; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3186.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "LFM2.5-2.6B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 unknown; 16k 3,186 103.5 7.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/31uaymyl",
   "source_url": "https://omlx.ai/benchmarks/performance/31uaymyl",
   "tg_tokens": null,
   "tps": 103.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-31uaymyl-ctx32768",
   "model": "LFM2.5-2.6B",
   "notes": "oMLX community benchmark trace; peak memory 8.8 GB at this context; duplicate traces of the same session: dia6n8br, mdt6j316, nyovpsl7; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2899.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "LFM2.5-2.6B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 unknown; 32k 2,899 98.9 8.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/31uaymyl",
   "source_url": "https://omlx.ai/benchmarks/performance/31uaymyl",
   "tg_tokens": null,
   "tps": 98.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-31uaymyl-ctx65536",
   "model": "LFM2.5-2.6B",
   "notes": "oMLX community benchmark trace; peak memory 12.0 GB at this context; duplicate traces of the same session: dia6n8br, mdt6j316, nyovpsl7; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2434.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "LFM2.5-2.6B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 unknown; 64k 2,434 89.7 12.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/31uaymyl",
   "source_url": "https://omlx.ai/benchmarks/performance/31uaymyl",
   "tg_tokens": null,
   "tps": 89.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-31uaymyl-ctx131072",
   "model": "LFM2.5-2.6B",
   "notes": "oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: dia6n8br, mdt6j316, nyovpsl7; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1831.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "LFM2.5-2.6B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 unknown; 128k 1,831 73.7 18.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/31uaymyl",
   "source_url": "https://omlx.ai/benchmarks/performance/31uaymyl",
   "tg_tokens": null,
   "tps": 73.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-7kduebt5-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 16k context; peak memory 13.9 GB; TTFT 8794 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1863.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k tokens 1,863 PP tok/s 127.1 TG tok/s 8794 TTFT (ms) 13.9 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7kduebt5",
   "source_url": "https://omlx.ai/benchmarks/performance/7kduebt5",
   "tg_tokens": null,
   "tps": 127.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5p8rlzxw-ctx4096",
   "model": "Muse-Glimmer-30B-Abliterated",
   "notes": "MLX build; oMLX community benchmark trace, single run at 4k context; peak memory 20.3 GB; TTFT 5492 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 745.8,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Muse-Glimmer-30B-Abliterated-MLX-mxfp4 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 745.8 PP tok/s 43.9 TG tok/s 5492 TTFT (ms) 20.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5p8rlzxw",
   "source_url": "https://omlx.ai/benchmarks/performance/5p8rlzxw",
   "tg_tokens": null,
   "tps": 43.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dilph9g8-ctx4096",
   "model": "Muse-Glimmer-30B-Abliterated",
   "notes": "MLX build; oMLX community benchmark trace, single run at 4k context; peak memory 20.3 GB; TTFT 5270 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 777.2,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Muse-Glimmer-30B-Abliterated-MLX-mxfp4 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 777.2 PP tok/s 42.9 TG tok/s 5270 TTFT (ms) 20.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dilph9g8",
   "source_url": "https://omlx.ai/benchmarks/performance/dilph9g8",
   "tg_tokens": null,
   "tps": 42.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4",
   "id": "omlx-wiuqvg12-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 8.7 GB at this context; duplicate traces of the same session: u710jn9d; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 104.7 10.0 8.7 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wiuqvg12",
   "source_url": "https://omlx.ai/benchmarks/performance/wiuqvg12",
   "tg_tokens": null,
   "tps": 10.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4",
   "id": "omlx-wiuqvg12-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 9.6 GB at this context; duplicate traces of the same session: u710jn9d; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 149.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 149.9 10.5 9.6 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wiuqvg12",
   "source_url": "https://omlx.ai/benchmarks/performance/wiuqvg12",
   "tg_tokens": null,
   "tps": 10.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-w78z8qq6-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: pe8abjw5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 848.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 848.2 33.7 16.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w78z8qq6",
   "source_url": "https://omlx.ai/benchmarks/performance/w78z8qq6",
   "tg_tokens": null,
   "tps": 33.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-w78z8qq6-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: pe8abjw5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 934.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 934.9 33.0 17.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w78z8qq6",
   "source_url": "https://omlx.ai/benchmarks/performance/w78z8qq6",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rb7bqer3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: y8wfp7ry; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 929.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 929.9 33.3 16.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rb7bqer3",
   "source_url": "https://omlx.ai/benchmarks/performance/rb7bqer3",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rb7bqer3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: y8wfp7ry; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 929.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 929.0 32.6 17.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rb7bqer3",
   "source_url": "https://omlx.ai/benchmarks/performance/rb7bqer3",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rb7bqer3-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: y8wfp7ry; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 796.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 796.5 32.0 17.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rb7bqer3",
   "source_url": "https://omlx.ai/benchmarks/performance/rb7bqer3",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rb7bqer3-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: y8wfp7ry; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 715.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 715.9 29.2 17.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rb7bqer3",
   "source_url": "https://omlx.ai/benchmarks/performance/rb7bqer3",
   "tg_tokens": null,
   "tps": 29.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-zta4sh2o-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 17.0 GB; TTFT 1876 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 545.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k tokens 545.7 PP tok/s 26.1 TG tok/s 1876 TTFT (ms) 17.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zta4sh2o",
   "source_url": "https://omlx.ai/benchmarks/performance/zta4sh2o",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-fhuc3ywv-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 72riqrs6; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 545.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 545.4 26.0 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fhuc3ywv",
   "source_url": "https://omlx.ai/benchmarks/performance/fhuc3ywv",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fhuc3ywv-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 72riqrs6; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 449.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 64k 449.7 26.2 21.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fhuc3ywv",
   "source_url": "https://omlx.ai/benchmarks/performance/fhuc3ywv",
   "tg_tokens": null,
   "tps": 26.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-hcnxmwrp-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 17.0 GB; TTFT 1871 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 547.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k tokens 547.2 PP tok/s 26.4 TG tok/s 1871 TTFT (ms) 17.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hcnxmwrp",
   "source_url": "https://omlx.ai/benchmarks/performance/hcnxmwrp",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-vjoxcde0-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 17.0 GB; TTFT 1869 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 547.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k tokens 547.8 PP tok/s 27.4 TG tok/s 1869 TTFT (ms) 17.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vjoxcde0",
   "source_url": "https://omlx.ai/benchmarks/performance/vjoxcde0",
   "tg_tokens": null,
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-4x0r8mbd-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 17.0 GB; TTFT 1871 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 547.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k tokens 547.3 PP tok/s 24.5 TG tok/s 1871 TTFT (ms) 17.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4x0r8mbd",
   "source_url": "https://omlx.ai/benchmarks/performance/4x0r8mbd",
   "tg_tokens": null,
   "tps": 24.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1otx6yfg-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 17.0 GB; TTFT 2021 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 506.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k tokens 506.6 PP tok/s 23.0 TG tok/s 2021 TTFT (ms) 17.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1otx6yfg",
   "source_url": "https://omlx.ai/benchmarks/performance/1otx6yfg",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-wj40qc3k-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 17.0 GB; TTFT 1875 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 546.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k tokens 546.1 PP tok/s 30.7 TG tok/s 1875 TTFT (ms) 17.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wj40qc3k",
   "source_url": "https://omlx.ai/benchmarks/performance/wj40qc3k",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-2wmdogda-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: hz3jk54l; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 521.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 521.0 17.9 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2wmdogda",
   "source_url": "https://omlx.ai/benchmarks/performance/2wmdogda",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-2wmdogda-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: hz3jk54l; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 623.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 8k 623.7 16.8 18.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2wmdogda",
   "source_url": "https://omlx.ai/benchmarks/performance/2wmdogda",
   "tg_tokens": null,
   "tps": 16.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-h8pa2wru-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 17.0 GB; TTFT 1874 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 546.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k tokens 546.4 PP tok/s 22.1 TG tok/s 1874 TTFT (ms) 17.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h8pa2wru",
   "source_url": "https://omlx.ai/benchmarks/performance/h8pa2wru",
   "tg_tokens": null,
   "tps": 22.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-lk8ydcbt-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 17.0 GB; TTFT 1976 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 518.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k tokens 518.2 PP tok/s 17.6 TG tok/s 1976 TTFT (ms) 17.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lk8ydcbt",
   "source_url": "https://omlx.ai/benchmarks/performance/lk8ydcbt",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-utu73foc-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 16.0 GB; TTFT 2637 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 388.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k tokens 388.4 PP tok/s 17.9 TG tok/s 2637 TTFT (ms) 16.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/utu73foc",
   "source_url": "https://omlx.ai/benchmarks/performance/utu73foc",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-j56m76ph-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 17.0 GB; TTFT 2008 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 509.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k tokens 509.9 PP tok/s 17.7 TG tok/s 2008 TTFT (ms) 17.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j56m76ph",
   "source_url": "https://omlx.ai/benchmarks/performance/j56m76ph",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-webyefkv-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 17.0 GB; TTFT 1971 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 519.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k tokens 519.4 PP tok/s 17.7 TG tok/s 1971 TTFT (ms) 17.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/webyefkv",
   "source_url": "https://omlx.ai/benchmarks/performance/webyefkv",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-5nczj1zy-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 52e22ycj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 124.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 124.1 22.1 16.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5nczj1zy",
   "source_url": "https://omlx.ai/benchmarks/performance/5nczj1zy",
   "tg_tokens": null,
   "tps": 22.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-5nczj1zy-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: 52e22ycj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 123.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 123.2 26.0 18.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5nczj1zy",
   "source_url": "https://omlx.ai/benchmarks/performance/5nczj1zy",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-219ejg5x-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: cw64muxh; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 144.1,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 144.1 23.7 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/219ejg5x",
   "source_url": "https://omlx.ai/benchmarks/performance/219ejg5x",
   "tg_tokens": null,
   "tps": 23.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-219ejg5x-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: cw64muxh; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 160.6,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 160.6 28.8 23.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/219ejg5x",
   "source_url": "https://omlx.ai/benchmarks/performance/219ejg5x",
   "tg_tokens": null,
   "tps": 28.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-r1m7iyhj-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: fm2ycj5o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 989.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 989.1 74.3 18.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r1m7iyhj",
   "source_url": "https://omlx.ai/benchmarks/performance/r1m7iyhj",
   "tg_tokens": null,
   "tps": 74.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-r1m7iyhj-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: fm2ycj5o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 968.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 968.0 77.8 19.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r1m7iyhj",
   "source_url": "https://omlx.ai/benchmarks/performance/r1m7iyhj",
   "tg_tokens": null,
   "tps": 77.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vofa9dud-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: rxoqixfe; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 988.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 988.9 46.7 18.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vofa9dud",
   "source_url": "https://omlx.ai/benchmarks/performance/vofa9dud",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vofa9dud-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: rxoqixfe; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 968.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 968.3 50.2 19.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vofa9dud",
   "source_url": "https://omlx.ai/benchmarks/performance/vofa9dud",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-7m9xqs8e-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 62.8 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 128k 205.1 43.9 62.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7m9xqs8e",
   "source_url": "https://omlx.ai/benchmarks/performance/7m9xqs8e",
   "tg_tokens": null,
   "tps": 43.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-7m9xqs8e-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 53.0 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 167.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 195k 167.3 30.8 53.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7m9xqs8e",
   "source_url": "https://omlx.ai/benchmarks/performance/7m9xqs8e",
   "tg_tokens": null,
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-7h6xjt0h-ctx131072",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 75.3 GB at this context; duplicate traces of the same session: avzwv4om; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1088.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 128k 1,088 46.6 75.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7h6xjt0h",
   "source_url": "https://omlx.ai/benchmarks/performance/7h6xjt0h",
   "tg_tokens": null,
   "tps": 46.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-7h6xjt0h-ctx199680",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 79.3 GB at this context; duplicate traces of the same session: avzwv4om; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1075.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 195k 1,075 46.6 79.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7h6xjt0h",
   "source_url": "https://omlx.ai/benchmarks/performance/7h6xjt0h",
   "tg_tokens": null,
   "tps": 46.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0vcvc2px-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: rbgfq9bt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1703.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,703 70.0 72.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0vcvc2px",
   "source_url": "https://omlx.ai/benchmarks/performance/0vcvc2px",
   "tg_tokens": null,
   "tps": 70.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0vcvc2px-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.4 GB at this context; duplicate traces of the same session: rbgfq9bt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1989.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 1,989 67.0 74.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0vcvc2px",
   "source_url": "https://omlx.ai/benchmarks/performance/0vcvc2px",
   "tg_tokens": null,
   "tps": 67.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-exe2agtc-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: c95jarim, lxkkzdgl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1101.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,101 71.0 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/exe2agtc",
   "source_url": "https://omlx.ai/benchmarks/performance/exe2agtc",
   "tg_tokens": null,
   "tps": 71.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-exe2agtc-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: c95jarim, lxkkzdgl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1540.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,540 71.7 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/exe2agtc",
   "source_url": "https://omlx.ai/benchmarks/performance/exe2agtc",
   "tg_tokens": null,
   "tps": 71.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-exe2agtc-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.5 GB at this context; duplicate traces of the same session: c95jarim, lxkkzdgl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2080.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,080 66.0 74.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/exe2agtc",
   "source_url": "https://omlx.ai/benchmarks/performance/exe2agtc",
   "tg_tokens": null,
   "tps": 66.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xzjt3rsk-ctx16384",
   "model": "Spark-X2.5-1.7B",
   "notes": "oMLX community benchmark trace; peak memory 4.4 GB at this context; duplicate traces of the same session: h2dkpv87, n3fm7pd1, syh2plzj; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3731.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Spark-X2.5-1.7B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 unknown; 16k 3,731 131.2 4.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xzjt3rsk",
   "source_url": "https://omlx.ai/benchmarks/performance/xzjt3rsk",
   "tg_tokens": null,
   "tps": 131.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xzjt3rsk-ctx32768",
   "model": "Spark-X2.5-1.7B",
   "notes": "oMLX community benchmark trace; peak memory 5.3 GB at this context; duplicate traces of the same session: h2dkpv87, n3fm7pd1, syh2plzj; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3134.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Spark-X2.5-1.7B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 unknown; 32k 3,134 117.1 5.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xzjt3rsk",
   "source_url": "https://omlx.ai/benchmarks/performance/xzjt3rsk",
   "tg_tokens": null,
   "tps": 117.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xzjt3rsk-ctx65536",
   "model": "Spark-X2.5-1.7B",
   "notes": "oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: h2dkpv87, n3fm7pd1, syh2plzj; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2452.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Spark-X2.5-1.7B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 unknown; 64k 2,452 106.0 7.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xzjt3rsk",
   "source_url": "https://omlx.ai/benchmarks/performance/xzjt3rsk",
   "tg_tokens": null,
   "tps": 106.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xzjt3rsk-ctx131072",
   "model": "Spark-X2.5-1.7B",
   "notes": "oMLX community benchmark trace; peak memory 10.5 GB at this context; duplicate traces of the same session: h2dkpv87, n3fm7pd1, syh2plzj; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1672.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Spark-X2.5-1.7B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 unknown; 128k 1,672 85.8 10.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xzjt3rsk",
   "source_url": "https://omlx.ai/benchmarks/performance/xzjt3rsk",
   "tg_tokens": null,
   "tps": 85.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-u50mpdle-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: ag84wz7s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 587.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 587.0 25.7 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u50mpdle",
   "source_url": "https://omlx.ai/benchmarks/performance/u50mpdle",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-u50mpdle-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: ag84wz7s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 955.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 955.7 25.3 22.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u50mpdle",
   "source_url": "https://omlx.ai/benchmarks/performance/u50mpdle",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-29rt7t9u-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: 0gha22cb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3527.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "gemma-4-26B-A4B-it-oQ5 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 5bit; 4k 3,527 102.1 17.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/29rt7t9u",
   "source_url": "https://omlx.ai/benchmarks/performance/29rt7t9u",
   "tg_tokens": null,
   "tps": 102.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-29rt7t9u-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: 0gha22cb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3453.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "gemma-4-26B-A4B-it-oQ5 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 5bit; 16k 3,453 91.0 18.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/29rt7t9u",
   "source_url": "https://omlx.ai/benchmarks/performance/29rt7t9u",
   "tg_tokens": null,
   "tps": 91.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0tky5nto-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: 1p3yn035; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3498.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "gemma-4-26B-A4B-it-oQ5 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 5bit; 4k 3,498 103.2 17.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0tky5nto",
   "source_url": "https://omlx.ai/benchmarks/performance/0tky5nto",
   "tg_tokens": null,
   "tps": 103.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-13",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0tky5nto-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: 1p3yn035; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3400.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "gemma-4-26B-A4B-it-oQ5 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 5bit; 16k 3,400 91.1 18.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0tky5nto",
   "source_url": "https://omlx.ai/benchmarks/performance/0tky5nto",
   "tg_tokens": null,
   "tps": 91.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-nvr9e8e0-ctx4096",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 21.5 GB; TTFT 3072 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1333.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiniCPM-V-4.7-35B-A3B-oQ4e on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 4k tokens 1,333 PP tok/s 82.4 TG tok/s 3072 TTFT (ms) 21.5 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nvr9e8e0",
   "source_url": "https://omlx.ai/benchmarks/performance/nvr9e8e0",
   "tg_tokens": null,
   "tps": 82.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-j2ox3frg-ctx4096",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: oqa2tbb2, skojri3h; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1294.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiniCPM-V-4.7-35B-A3B-oQ4e on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 4k 1,294 79.6 21.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j2ox3frg",
   "source_url": "https://omlx.ai/benchmarks/performance/j2ox3frg",
   "tg_tokens": null,
   "tps": 79.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-j2ox3frg-ctx8192",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: oqa2tbb2, skojri3h; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1257.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiniCPM-V-4.7-35B-A3B-oQ4e on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 8k 1,257 74.8 21.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j2ox3frg",
   "source_url": "https://omlx.ai/benchmarks/performance/j2ox3frg",
   "tg_tokens": null,
   "tps": 74.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-j2ox3frg-ctx16384",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: oqa2tbb2, skojri3h; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1144.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiniCPM-V-4.7-35B-A3B-oQ4e on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 16k 1,144 66.9 21.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j2ox3frg",
   "source_url": "https://omlx.ai/benchmarks/performance/j2ox3frg",
   "tg_tokens": null,
   "tps": 66.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-g4u9tltg-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: lkrkbuu3, m2y4goix; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2223.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 2,223 102.8 23.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g4u9tltg",
   "source_url": "https://omlx.ai/benchmarks/performance/g4u9tltg",
   "tg_tokens": null,
   "tps": 102.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-g4u9tltg-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: lkrkbuu3, m2y4goix; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1815.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 1,815 86.4 24.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g4u9tltg",
   "source_url": "https://omlx.ai/benchmarks/performance/g4u9tltg",
   "tg_tokens": null,
   "tps": 86.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-g4u9tltg-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: lkrkbuu3, m2y4goix; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1321.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 1,321 58.8 25.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g4u9tltg",
   "source_url": "https://omlx.ai/benchmarks/performance/g4u9tltg",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-fn3gl4nb-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 0e6v00hf; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 792.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 792.1 106.2 21.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fn3gl4nb",
   "source_url": "https://omlx.ai/benchmarks/performance/fn3gl4nb",
   "tg_tokens": null,
   "tps": 106.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-fn3gl4nb-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: 0e6v00hf; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 868.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 868.9 93.2 23.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fn3gl4nb",
   "source_url": "https://omlx.ai/benchmarks/performance/fn3gl4nb",
   "tg_tokens": null,
   "tps": 93.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-huirqnko-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace, single run at 64k context; peak memory 40.4 GB; TTFT 118426 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 553.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 64k tokens 553.4 PP tok/s 41.0 TG tok/s 118426 TTFT (ms) 40.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/huirqnko",
   "source_url": "https://omlx.ai/benchmarks/performance/huirqnko",
   "tg_tokens": null,
   "tps": 41.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-iswoz7ih-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 37.6 GB at this context; duplicate traces of the same session: 23qq4xo5, 3cgtznf1, bl4ara77; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1098.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k 1,098 56.5 37.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iswoz7ih",
   "source_url": "https://omlx.ai/benchmarks/performance/iswoz7ih",
   "tg_tokens": null,
   "tps": 56.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-iswoz7ih-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 37.9 GB at this context; duplicate traces of the same session: 23qq4xo5, 3cgtznf1, bl4ara77; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1049.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 8k 1,049 56.7 37.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iswoz7ih",
   "source_url": "https://omlx.ai/benchmarks/performance/iswoz7ih",
   "tg_tokens": null,
   "tps": 56.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-iswoz7ih-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 38.8 GB at this context; duplicate traces of the same session: 23qq4xo5, 3cgtznf1, bl4ara77; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 932.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 16k 932.7 52.1 38.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iswoz7ih",
   "source_url": "https://omlx.ai/benchmarks/performance/iswoz7ih",
   "tg_tokens": null,
   "tps": 52.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-iswoz7ih-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 39.7 GB at this context; duplicate traces of the same session: 23qq4xo5, 3cgtznf1, bl4ara77; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 755.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 32k 755.2 48.6 39.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iswoz7ih",
   "source_url": "https://omlx.ai/benchmarks/performance/iswoz7ih",
   "tg_tokens": null,
   "tps": 48.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-k6v2qlfq-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.4 GB at this context; duplicate traces of the same session: uekiusll, xx4t21ew; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2591.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 2,591 138.4 39.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k6v2qlfq",
   "source_url": "https://omlx.ai/benchmarks/performance/k6v2qlfq",
   "tg_tokens": null,
   "tps": 138.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-k6v2qlfq-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.5 GB at this context; duplicate traces of the same session: uekiusll, xx4t21ew; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2463.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 2,463 140.3 39.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k6v2qlfq",
   "source_url": "https://omlx.ai/benchmarks/performance/k6v2qlfq",
   "tg_tokens": null,
   "tps": 140.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-k6v2qlfq-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.8 GB at this context; duplicate traces of the same session: uekiusll, xx4t21ew; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2203.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 16k 2,203 85.2 39.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k6v2qlfq",
   "source_url": "https://omlx.ai/benchmarks/performance/k6v2qlfq",
   "tg_tokens": null,
   "tps": 85.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-wrx6kbrj-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.8 GB at this context; duplicate traces of the same session: 9x0f4xin, jwz11yi0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2202.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 16k 2,202 102.6 39.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wrx6kbrj",
   "source_url": "https://omlx.ai/benchmarks/performance/wrx6kbrj",
   "tg_tokens": null,
   "tps": 102.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-wrx6kbrj-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 40.5 GB at this context; duplicate traces of the same session: 9x0f4xin, jwz11yi0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1797.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 32k 1,797 86.3 40.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wrx6kbrj",
   "source_url": "https://omlx.ai/benchmarks/performance/wrx6kbrj",
   "tg_tokens": null,
   "tps": 86.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-wrx6kbrj-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 41.8 GB at this context; duplicate traces of the same session: 9x0f4xin, jwz11yi0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1313.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 64k 1,313 65.1 41.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wrx6kbrj",
   "source_url": "https://omlx.ai/benchmarks/performance/wrx6kbrj",
   "tg_tokens": null,
   "tps": 65.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-y00a02v7-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.4 GB at this context; duplicate traces of the same session: 0ad7ef1i, dsleregt; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2580.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 2,580 129.3 39.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y00a02v7",
   "source_url": "https://omlx.ai/benchmarks/performance/y00a02v7",
   "tg_tokens": null,
   "tps": 129.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-y00a02v7-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.5 GB at this context; duplicate traces of the same session: 0ad7ef1i, dsleregt; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2462.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 2,462 111.7 39.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y00a02v7",
   "source_url": "https://omlx.ai/benchmarks/performance/y00a02v7",
   "tg_tokens": null,
   "tps": 111.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-y00a02v7-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.8 GB at this context; duplicate traces of the same session: 0ad7ef1i, dsleregt; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2197.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 16k 2,197 112.5 39.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y00a02v7",
   "source_url": "https://omlx.ai/benchmarks/performance/y00a02v7",
   "tg_tokens": null,
   "tps": 112.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ubnmu7t1-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 8tq5buyr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 929.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 929.9 33.3 16.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ubnmu7t1",
   "source_url": "https://omlx.ai/benchmarks/performance/ubnmu7t1",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ubnmu7t1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 8tq5buyr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 929.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 929.0 32.6 17.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ubnmu7t1",
   "source_url": "https://omlx.ai/benchmarks/performance/ubnmu7t1",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ubnmu7t1-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 8tq5buyr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 796.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 796.5 32.0 17.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ubnmu7t1",
   "source_url": "https://omlx.ai/benchmarks/performance/ubnmu7t1",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ubnmu7t1-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 8tq5buyr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 715.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 715.9 29.2 17.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ubnmu7t1",
   "source_url": "https://omlx.ai/benchmarks/performance/ubnmu7t1",
   "tg_tokens": null,
   "tps": 29.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-h6y3vs57-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized speed FP16 build; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: 394sn657, 4m4mw3hj; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 154.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed-FP16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 154.9 16.3 24.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h6y3vs57",
   "source_url": "https://omlx.ai/benchmarks/performance/h6y3vs57",
   "tg_tokens": null,
   "tps": 16.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-h6y3vs57-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized speed FP16 build; oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: 394sn657, 4m4mw3hj; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 145.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed-FP16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 145.4 16.0 24.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h6y3vs57",
   "source_url": "https://omlx.ai/benchmarks/performance/h6y3vs57",
   "tg_tokens": null,
   "tps": 16.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-h6y3vs57-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized speed FP16 build; oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: 394sn657, 4m4mw3hj; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 145.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed-FP16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 145.5 15.5 25.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h6y3vs57",
   "source_url": "https://omlx.ai/benchmarks/performance/h6y3vs57",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-h6y3vs57-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized speed FP16 build; oMLX community benchmark trace; peak memory 26.9 GB at this context; duplicate traces of the same session: 394sn657, 4m4mw3hj; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 134.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed-FP16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 32k 134.0 13.6 26.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h6y3vs57",
   "source_url": "https://omlx.ai/benchmarks/performance/h6y3vs57",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-sqzze32e-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 3eb1vswt; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 103.8 23.8 20.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sqzze32e",
   "source_url": "https://omlx.ai/benchmarks/performance/sqzze32e",
   "tg_tokens": null,
   "tps": 23.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-sqzze32e-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: 3eb1vswt; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 154.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 154.8 34.2 24.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sqzze32e",
   "source_url": "https://omlx.ai/benchmarks/performance/sqzze32e",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-704groyv-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 17.7 GB; TTFT 1135 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 902.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 902.4 PP tok/s 73.7 TG tok/s 1135 TTFT (ms) 17.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/704groyv",
   "source_url": "https://omlx.ai/benchmarks/performance/704groyv",
   "tg_tokens": null,
   "tps": 73.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-infh8823-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 21.3 GB; TTFT 94733 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 691.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k tokens 691.8 PP tok/s 23.6 TG tok/s 94733 TTFT (ms) 21.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/infh8823",
   "source_url": "https://omlx.ai/benchmarks/performance/infh8823",
   "tg_tokens": null,
   "tps": 23.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-00v05bqv-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 31.9 GB; TTFT 102465 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 639.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k tokens 639.6 PP tok/s 42.1 TG tok/s 102465 TTFT (ms) 31.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/00v05bqv",
   "source_url": "https://omlx.ai/benchmarks/performance/00v05bqv",
   "tg_tokens": null,
   "tps": 42.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-53v9iccu-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 101.2 GB; TTFT 2101 ms; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 487.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 1k tokens 487.4 PP tok/s 35.0 TG tok/s 2101 TTFT (ms) 101.2 Peak mem",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/53v9iccu",
   "source_url": "https://omlx.ai/benchmarks/performance/53v9iccu",
   "tg_tokens": null,
   "tps": 35.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-23ol3j8m-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 39.0 GB; TTFT 6681 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 153.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 153.3 PP tok/s 21.0 TG tok/s 6681 TTFT (ms) 39.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/23ol3j8m",
   "source_url": "https://omlx.ai/benchmarks/performance/23ol3j8m",
   "tg_tokens": null,
   "tps": 21.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mo5rd6nj-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 23.3 GB; TTFT 6063 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 168.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 168.9 PP tok/s 14.7 TG tok/s 6063 TTFT (ms) 23.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mo5rd6nj",
   "source_url": "https://omlx.ai/benchmarks/performance/mo5rd6nj",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pbdbvx1b-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 22.9 GB; TTFT 6878 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 148.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 148.9 PP tok/s 12.7 TG tok/s 6878 TTFT (ms) 22.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pbdbvx1b",
   "source_url": "https://omlx.ai/benchmarks/performance/pbdbvx1b",
   "tg_tokens": null,
   "tps": 12.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mg4zu3op-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: 24s8v2nb, evpt7fon, mb702noz, p9113gcf, t36atkb5, w3y6mi3c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1154.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,154 78.1 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mg4zu3op",
   "source_url": "https://omlx.ai/benchmarks/performance/mg4zu3op",
   "tg_tokens": null,
   "tps": 78.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mg4zu3op-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 24s8v2nb, evpt7fon, mb702noz, p9113gcf, t36atkb5, w3y6mi3c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1841.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,841 75.2 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mg4zu3op",
   "source_url": "https://omlx.ai/benchmarks/performance/mg4zu3op",
   "tg_tokens": null,
   "tps": 75.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mg4zu3op-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 24s8v2nb, evpt7fon, mb702noz, p9113gcf, t36atkb5, w3y6mi3c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2249.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,249 77.3 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mg4zu3op",
   "source_url": "https://omlx.ai/benchmarks/performance/mg4zu3op",
   "tg_tokens": null,
   "tps": 77.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mg4zu3op-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: 24s8v2nb, evpt7fon, mb702noz, p9113gcf, t36atkb5, w3y6mi3c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2278.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,278 72.8 72.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mg4zu3op",
   "source_url": "https://omlx.ai/benchmarks/performance/mg4zu3op",
   "tg_tokens": null,
   "tps": 72.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mg4zu3op-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: 24s8v2nb, evpt7fon, mb702noz, p9113gcf, t36atkb5, w3y6mi3c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2149.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,149 52.6 73.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mg4zu3op",
   "source_url": "https://omlx.ai/benchmarks/performance/mg4zu3op",
   "tg_tokens": null,
   "tps": 52.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mg4zu3op-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.8 GB at this context; duplicate traces of the same session: 24s8v2nb, evpt7fon, mb702noz, p9113gcf, t36atkb5, w3y6mi3c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2143.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,143 76.6 74.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mg4zu3op",
   "source_url": "https://omlx.ai/benchmarks/performance/mg4zu3op",
   "tg_tokens": null,
   "tps": 76.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mg4zu3op-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.2 GB at this context; duplicate traces of the same session: 24s8v2nb, evpt7fon, mb702noz, p9113gcf, t36atkb5, w3y6mi3c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2157.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,157 65.4 79.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mg4zu3op",
   "source_url": "https://omlx.ai/benchmarks/performance/mg4zu3op",
   "tg_tokens": null,
   "tps": 65.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4uhsilni-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.7 GB; TTFT 2807 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 364.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 364.8 PP tok/s 29.3 TG tok/s 2807 TTFT (ms) 70.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4uhsilni",
   "source_url": "https://omlx.ai/benchmarks/performance/4uhsilni",
   "tg_tokens": null,
   "tps": 29.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-veocwsr2-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 4xaz3i4t, s0balggm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1110.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,110 80.9 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/veocwsr2",
   "source_url": "https://omlx.ai/benchmarks/performance/veocwsr2",
   "tg_tokens": null,
   "tps": 80.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-veocwsr2-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 4xaz3i4t, s0balggm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1825.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,825 60.4 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/veocwsr2",
   "source_url": "https://omlx.ai/benchmarks/performance/veocwsr2",
   "tg_tokens": null,
   "tps": 60.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-veocwsr2-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 4xaz3i4t, s0balggm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2216.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,216 76.4 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/veocwsr2",
   "source_url": "https://omlx.ai/benchmarks/performance/veocwsr2",
   "tg_tokens": null,
   "tps": 76.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-veocwsr2-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: 4xaz3i4t, s0balggm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2319.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,319 74.9 73.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/veocwsr2",
   "source_url": "https://omlx.ai/benchmarks/performance/veocwsr2",
   "tg_tokens": null,
   "tps": 74.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-veocwsr2-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.5 GB at this context; duplicate traces of the same session: 4xaz3i4t, s0balggm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2227.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,227 77.0 74.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/veocwsr2",
   "source_url": "https://omlx.ai/benchmarks/performance/veocwsr2",
   "tg_tokens": null,
   "tps": 77.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-veocwsr2-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.2 GB at this context; duplicate traces of the same session: 4xaz3i4t, s0balggm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2170.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,170 63.9 79.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/veocwsr2",
   "source_url": "https://omlx.ai/benchmarks/performance/veocwsr2",
   "tg_tokens": null,
   "tps": 63.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-09ol2x3s-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 86.5 GB at this context; duplicate traces of the same session: 99l824w2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1044.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 1,044 66.2 86.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/09ol2x3s",
   "source_url": "https://omlx.ai/benchmarks/performance/09ol2x3s",
   "tg_tokens": null,
   "tps": 66.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-09ol2x3s-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 86.7 GB at this context; duplicate traces of the same session: 99l824w2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1811.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 1,811 73.6 86.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/09ol2x3s",
   "source_url": "https://omlx.ai/benchmarks/performance/09ol2x3s",
   "tg_tokens": null,
   "tps": 73.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5ghbb4pv-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 28.2 GB; TTFT 6844 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 149.6,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k tokens 149.6 PP tok/s 11.7 TG tok/s 6844 TTFT (ms) 28.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ghbb4pv",
   "source_url": "https://omlx.ai/benchmarks/performance/5ghbb4pv",
   "tg_tokens": null,
   "tps": 11.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ucr9hwr9-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 86.5 GB; TTFT 1148 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 892.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k tokens 892.1 PP tok/s 69.6 TG tok/s 1148 TTFT (ms) 86.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ucr9hwr9",
   "source_url": "https://omlx.ai/benchmarks/performance/ucr9hwr9",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-v3yg36a9-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.6 GB at this context; duplicate traces of the same session: 549g7lwc, cmehvhhk, elyyjxh0, kci3h0gy, oemblapa, w9k221lf; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2530.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 2,530 154.6 142.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v3yg36a9",
   "source_url": "https://omlx.ai/benchmarks/performance/v3yg36a9",
   "tg_tokens": null,
   "tps": 154.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-v3yg36a9-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: 549g7lwc, cmehvhhk, elyyjxh0, kci3h0gy, oemblapa, w9k221lf; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4463.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 4,463 107.4 143.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v3yg36a9",
   "source_url": "https://omlx.ai/benchmarks/performance/v3yg36a9",
   "tg_tokens": null,
   "tps": 107.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-v3yg36a9-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 144.4 GB at this context; duplicate traces of the same session: 549g7lwc, cmehvhhk, elyyjxh0, kci3h0gy, oemblapa, w9k221lf; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4950.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 8k 4,950 114.5 144.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v3yg36a9",
   "source_url": "https://omlx.ai/benchmarks/performance/v3yg36a9",
   "tg_tokens": null,
   "tps": 114.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-v3yg36a9-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.7 GB at this context; duplicate traces of the same session: 549g7lwc, cmehvhhk, elyyjxh0, kci3h0gy, oemblapa, w9k221lf; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4964.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 32k 4,964 104.8 145.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v3yg36a9",
   "source_url": "https://omlx.ai/benchmarks/performance/v3yg36a9",
   "tg_tokens": null,
   "tps": 104.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-v3yg36a9-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 146.8 GB at this context; duplicate traces of the same session: 549g7lwc, cmehvhhk, elyyjxh0, kci3h0gy, oemblapa, w9k221lf; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4865.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 64k 4,865 108.1 146.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v3yg36a9",
   "source_url": "https://omlx.ai/benchmarks/performance/v3yg36a9",
   "tg_tokens": null,
   "tps": 108.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-v3yg36a9-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 150.3 GB at this context; duplicate traces of the same session: 549g7lwc, cmehvhhk, elyyjxh0, kci3h0gy, oemblapa, w9k221lf; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4725.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 128k 4,725 86.7 150.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v3yg36a9",
   "source_url": "https://omlx.ai/benchmarks/performance/v3yg36a9",
   "tg_tokens": null,
   "tps": 86.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-v3yg36a9-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 153.3 GB at this context; duplicate traces of the same session: 549g7lwc, cmehvhhk, elyyjxh0, kci3h0gy, oemblapa, w9k221lf; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4588.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 195k 4,588 84.0 153.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v3yg36a9",
   "source_url": "https://omlx.ai/benchmarks/performance/v3yg36a9",
   "tg_tokens": null,
   "tps": 84.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ml4qre7o-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 186.2 GB; TTFT 24523 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1336.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 32k tokens 1,336 PP tok/s 86.0 TG tok/s 24523 TTFT (ms) 186.2 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ml4qre7o",
   "source_url": "https://omlx.ai/benchmarks/performance/ml4qre7o",
   "tg_tokens": null,
   "tps": 86.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-nqv6u26t-ctx1024",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: r87o5jmb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 215.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "AREX-2-6bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 6bit; 1k 215.5 28.6 25.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nqv6u26t",
   "source_url": "https://omlx.ai/benchmarks/performance/nqv6u26t",
   "tg_tokens": null,
   "tps": 28.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-nqv6u26t-ctx4096",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: r87o5jmb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 214.2,
   "provenance": "community",
   "quant": "6bit",
   "quote": "AREX-2-6bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 6bit; 4k 214.2 41.9 27.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nqv6u26t",
   "source_url": "https://omlx.ai/benchmarks/performance/nqv6u26t",
   "tg_tokens": null,
   "tps": 41.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-61gykwlm-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 105.1 GB at this context; duplicate traces of the same session: 3gokkf2s, 5i1cfuzh, u275vapd; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 61.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 61.0 11.6 105.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/61gykwlm",
   "source_url": "https://omlx.ai/benchmarks/performance/61gykwlm",
   "tg_tokens": null,
   "tps": 11.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-61gykwlm-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 106.1 GB at this context; duplicate traces of the same session: 3gokkf2s, 5i1cfuzh, u275vapd; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 249.3,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 249.3 12.6 106.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/61gykwlm",
   "source_url": "https://omlx.ai/benchmarks/performance/61gykwlm",
   "tg_tokens": null,
   "tps": 12.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-61gykwlm-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 107.2 GB at this context; duplicate traces of the same session: 3gokkf2s, 5i1cfuzh, u275vapd; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 356.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 8k 356.0 12.9 107.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/61gykwlm",
   "source_url": "https://omlx.ai/benchmarks/performance/61gykwlm",
   "tg_tokens": null,
   "tps": 12.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-61gykwlm-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 107.2 GB at this context; duplicate traces of the same session: 3gokkf2s, 5i1cfuzh, u275vapd; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 367.8,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 16k 367.8 14.3 107.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/61gykwlm",
   "source_url": "https://omlx.ai/benchmarks/performance/61gykwlm",
   "tg_tokens": null,
   "tps": 14.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jo4geo8k-ctx1024",
   "model": "Huihui-Ornith-1.5-35B-A3B-abliterated",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: foypulcw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3982.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Huihui-Ornith-1.5-35B-A3B-abliterated-oQ5e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 3,982 170.9 26.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jo4geo8k",
   "source_url": "https://omlx.ai/benchmarks/performance/jo4geo8k",
   "tg_tokens": null,
   "tps": 170.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jo4geo8k-ctx4096",
   "model": "Huihui-Ornith-1.5-35B-A3B-abliterated",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: foypulcw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5412.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Huihui-Ornith-1.5-35B-A3B-abliterated-oQ5e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 5,412 159.0 27.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jo4geo8k",
   "source_url": "https://omlx.ai/benchmarks/performance/jo4geo8k",
   "tg_tokens": null,
   "tps": 159.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-xgt4pmgi-ctx1024",
   "model": "Huihui-Qwen3.8-27B-abliterated",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: agdx15eg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 850.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Huihui-Qwen3.8-27B-abliterated-mlx-8Bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 850.5 47.8 32.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xgt4pmgi",
   "source_url": "https://omlx.ai/benchmarks/performance/xgt4pmgi",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-xgt4pmgi-ctx4096",
   "model": "Huihui-Qwen3.8-27B-abliterated",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 33.7 GB at this context; duplicate traces of the same session: agdx15eg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 606.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Huihui-Qwen3.8-27B-abliterated-mlx-8Bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 606.0 36.9 33.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xgt4pmgi",
   "source_url": "https://omlx.ai/benchmarks/performance/xgt4pmgi",
   "tg_tokens": null,
   "tps": 36.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dd2z2e10-ctx1024",
   "model": "Huihui-Qwen3.8-27B-abliterated",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: bq3vz56a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 814.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Huihui-Qwen3.8-27B-abliterated-mlx-8Bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 814.6 18.8 28.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dd2z2e10",
   "source_url": "https://omlx.ai/benchmarks/performance/dd2z2e10",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dd2z2e10-ctx4096",
   "model": "Huihui-Qwen3.8-27B-abliterated",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: bq3vz56a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 925.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Huihui-Qwen3.8-27B-abliterated-mlx-8Bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 925.5 18.6 29.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dd2z2e10",
   "source_url": "https://omlx.ai/benchmarks/performance/dd2z2e10",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-wctaarl3-ctx4096",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 4wd4vlcb, uekbgu2w; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1335.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiniCPM-V-4.7-35B-A3B-oQ4e on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 4k 1,335 82.4 21.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wctaarl3",
   "source_url": "https://omlx.ai/benchmarks/performance/wctaarl3",
   "tg_tokens": null,
   "tps": 82.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-wctaarl3-ctx8192",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: 4wd4vlcb, uekbgu2w; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1278.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiniCPM-V-4.7-35B-A3B-oQ4e on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 8k 1,278 79.4 21.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wctaarl3",
   "source_url": "https://omlx.ai/benchmarks/performance/wctaarl3",
   "tg_tokens": null,
   "tps": 79.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-wctaarl3-ctx16384",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: 4wd4vlcb, uekbgu2w; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1137.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiniCPM-V-4.7-35B-A3B-oQ4e on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 16k 1,137 73.2 21.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wctaarl3",
   "source_url": "https://omlx.ai/benchmarks/performance/wctaarl3",
   "tg_tokens": null,
   "tps": 73.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ddy6ls23-ctx4096",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 21.5 GB; TTFT 3227 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1269.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiniCPM-V-4.7-35B-A3B-oQ4e on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 4k tokens 1,269 PP tok/s 82.4 TG tok/s 3227 TTFT (ms) 21.5 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ddy6ls23",
   "source_url": "https://omlx.ai/benchmarks/performance/ddy6ls23",
   "tg_tokens": null,
   "tps": 82.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-l4tj190h-ctx4096",
   "model": "Muse-Glimmer-30B-Abliterated",
   "notes": "MM build; oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: ixjcq0y7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 218.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-Abliterated-MM-oQ4e on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 218.7 20.8 16.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l4tj190h",
   "source_url": "https://omlx.ai/benchmarks/performance/l4tj190h",
   "tg_tokens": null,
   "tps": 20.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-l4tj190h-ctx16384",
   "model": "Muse-Glimmer-30B-Abliterated",
   "notes": "MM build; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: ixjcq0y7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 194.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-Abliterated-MM-oQ4e on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 194.4 19.5 16.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l4tj190h",
   "source_url": "https://omlx.ai/benchmarks/performance/l4tj190h",
   "tg_tokens": null,
   "tps": 19.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-5zl1cwt0-ctx4096",
   "model": "Muse-Glimmer-30B-Abliterated",
   "notes": "MM build; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: ufv3gtt5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 213.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Muse-Glimmer-30B-Abliterated-MM-oQ6e on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 213.7 15.1 22.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5zl1cwt0",
   "source_url": "https://omlx.ai/benchmarks/performance/5zl1cwt0",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-5zl1cwt0-ctx16384",
   "model": "Muse-Glimmer-30B-Abliterated",
   "notes": "MM build; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: ufv3gtt5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 183.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Muse-Glimmer-30B-Abliterated-MM-oQ6e on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 183.6 14.3 22.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5zl1cwt0",
   "source_url": "https://omlx.ai/benchmarks/performance/5zl1cwt0",
   "tg_tokens": null,
   "tps": 14.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8vkl8bxz-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: qomx3d11; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2574.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 2,574 79.5 22.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8vkl8bxz",
   "source_url": "https://omlx.ai/benchmarks/performance/8vkl8bxz",
   "tg_tokens": null,
   "tps": 79.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8vkl8bxz-ctx199680",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: qomx3d11; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1920.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 195k 1,920 63.3 23.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8vkl8bxz",
   "source_url": "https://omlx.ai/benchmarks/performance/8vkl8bxz",
   "tg_tokens": null,
   "tps": 63.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-300jbn39-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 25.2 GB at this context; duplicate traces of the same session: f93u9i0u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4164.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,164 208.8 25.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/300jbn39",
   "source_url": "https://omlx.ai/benchmarks/performance/300jbn39",
   "tg_tokens": null,
   "tps": 208.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-300jbn39-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 27.0 GB at this context; duplicate traces of the same session: f93u9i0u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5880.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 5,880 172.2 27.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/300jbn39",
   "source_url": "https://omlx.ai/benchmarks/performance/300jbn39",
   "tg_tokens": null,
   "tps": 172.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4vvw9y67-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: d6g5th92, h6p4pdhn, hzp05c3p; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2590.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 2,590 210.0 21.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4vvw9y67",
   "source_url": "https://omlx.ai/benchmarks/performance/4vvw9y67",
   "tg_tokens": null,
   "tps": 210.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4vvw9y67-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: d6g5th92, h6p4pdhn, hzp05c3p; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3182.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 3,182 134.5 23.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4vvw9y67",
   "source_url": "https://omlx.ai/benchmarks/performance/4vvw9y67",
   "tg_tokens": null,
   "tps": 134.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-4vvw9y67-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: d6g5th92, h6p4pdhn, hzp05c3p; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2850.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 2,850 123.2 23.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4vvw9y67",
   "source_url": "https://omlx.ai/benchmarks/performance/4vvw9y67",
   "tg_tokens": null,
   "tps": 123.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4vvw9y67-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: d6g5th92, h6p4pdhn, hzp05c3p; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2352.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 2,352 142.0 24.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4vvw9y67",
   "source_url": "https://omlx.ai/benchmarks/performance/4vvw9y67",
   "tg_tokens": null,
   "tps": 142.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-uk1ki276-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 8k context; peak memory 12.1 GB; TTFT 42833 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 191.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 8bit; 8k tokens 191.3 PP tok/s 6.8 TG tok/s 42833 TTFT (ms) 12.1 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uk1ki276",
   "source_url": "https://omlx.ai/benchmarks/performance/uk1ki276",
   "tg_tokens": null,
   "tps": 6.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nax874cc-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.5 GB at this context; duplicate traces of the same session: k22ayclb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2465.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 2,465 106.7 39.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nax874cc",
   "source_url": "https://omlx.ai/benchmarks/performance/nax874cc",
   "tg_tokens": null,
   "tps": 106.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nax874cc-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.8 GB at this context; duplicate traces of the same session: k22ayclb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2200.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 16k 2,200 93.6 39.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nax874cc",
   "source_url": "https://omlx.ai/benchmarks/performance/nax874cc",
   "tg_tokens": null,
   "tps": 93.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-t18es8p1-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP fp16 build; oMLX community benchmark trace, single run at 8k context; peak memory 15.3 GB; TTFT 34195 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 239.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp-fp16 on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 8bit; 8k tokens 239.6 PP tok/s 6.9 TG tok/s 34195 TTFT (ms) 15.3 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t18es8p1",
   "source_url": "https://omlx.ai/benchmarks/performance/t18es8p1",
   "tg_tokens": null,
   "tps": 6.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-j92z2vej-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP fp16 build; oMLX community benchmark trace, single run at 4k context; peak memory 18.4 GB; TTFT 17487 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 234.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp-fp16 on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 8bit; 4k tokens 234.2 PP tok/s 9.6 TG tok/s 17487 TTFT (ms) 18.4 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j92z2vej",
   "source_url": "https://omlx.ai/benchmarks/performance/j92z2vej",
   "tg_tokens": null,
   "tps": 9.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9k65p6xi-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: pqcpywvj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 991.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 991.7 32.7 17.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9k65p6xi",
   "source_url": "https://omlx.ai/benchmarks/performance/9k65p6xi",
   "tg_tokens": null,
   "tps": 32.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9k65p6xi-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: pqcpywvj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 805.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 805.1 30.1 17.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9k65p6xi",
   "source_url": "https://omlx.ai/benchmarks/performance/9k65p6xi",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-od6a06vm-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: o258dffr; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 466.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 466.6 18.2 16.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/od6a06vm",
   "source_url": "https://omlx.ai/benchmarks/performance/od6a06vm",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-od6a06vm-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: o258dffr; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 510.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 510.9 17.9 17.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/od6a06vm",
   "source_url": "https://omlx.ai/benchmarks/performance/od6a06vm",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-4bte7739-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "fresh rebuild; oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: zyuneotr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 260.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-fresh on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 260.2 29.1 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4bte7739",
   "source_url": "https://omlx.ai/benchmarks/performance/4bte7739",
   "tg_tokens": null,
   "tps": 29.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-4bte7739-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "fresh rebuild; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: zyuneotr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 274.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-fresh on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 274.3 28.3 19.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4bte7739",
   "source_url": "https://omlx.ai/benchmarks/performance/4bte7739",
   "tg_tokens": null,
   "tps": 28.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tdz3fhme-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: bnyxexk5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 864.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 864.5 34.9 33.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tdz3fhme",
   "source_url": "https://omlx.ai/benchmarks/performance/tdz3fhme",
   "tg_tokens": null,
   "tps": 34.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tdz3fhme-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: bnyxexk5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 812.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 812.6 32.4 34.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tdz3fhme",
   "source_url": "https://omlx.ai/benchmarks/performance/tdz3fhme",
   "tg_tokens": null,
   "tps": 32.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-stihv69l-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: k7holpye; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 825.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 825.6 21.5 25.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/stihv69l",
   "source_url": "https://omlx.ai/benchmarks/performance/stihv69l",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-stihv69l-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: k7holpye; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 893.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 893.6 27.2 27.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/stihv69l",
   "source_url": "https://omlx.ai/benchmarks/performance/stihv69l",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7s1ut0e3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: uy7l2a1d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 828.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 828.9 48.4 25.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7s1ut0e3",
   "source_url": "https://omlx.ai/benchmarks/performance/7s1ut0e3",
   "tg_tokens": null,
   "tps": 48.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7s1ut0e3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: uy7l2a1d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 900.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 900.7 42.8 27.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7s1ut0e3",
   "source_url": "https://omlx.ai/benchmarks/performance/7s1ut0e3",
   "tg_tokens": null,
   "tps": 42.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-yya6uja3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized speed FP16 build; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 154.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed-FP16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 154.9 16.3 24.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yya6uja3",
   "source_url": "https://omlx.ai/benchmarks/performance/yya6uja3",
   "tg_tokens": null,
   "tps": 16.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-yya6uja3-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized speed FP16 build; oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 145.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed-FP16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 145.4 16.0 24.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yya6uja3",
   "source_url": "https://omlx.ai/benchmarks/performance/yya6uja3",
   "tg_tokens": null,
   "tps": 16.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-yya6uja3-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized speed FP16 build; oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 145.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed-FP16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 145.5 15.5 25.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yya6uja3",
   "source_url": "https://omlx.ai/benchmarks/performance/yya6uja3",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-yya6uja3-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized speed FP16 build; oMLX community benchmark trace; peak memory 26.9 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 134.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed-FP16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 32k 134.0 13.6 26.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yya6uja3",
   "source_url": "https://omlx.ai/benchmarks/performance/yya6uja3",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-8xw9rr9d-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: o5zl1xee; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 325.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 325.3 36.9 16.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8xw9rr9d",
   "source_url": "https://omlx.ai/benchmarks/performance/8xw9rr9d",
   "tg_tokens": null,
   "tps": 36.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-8xw9rr9d-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: o5zl1xee; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 351.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 351.6 36.0 19.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8xw9rr9d",
   "source_url": "https://omlx.ai/benchmarks/performance/8xw9rr9d",
   "tg_tokens": null,
   "tps": 36.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rt1nxi48-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 36.8 GB; TTFT 247207 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 530.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 128k tokens 530.2 PP tok/s 36.0 TG tok/s 247207 TTFT (ms) 36.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rt1nxi48",
   "source_url": "https://omlx.ai/benchmarks/performance/rt1nxi48",
   "tg_tokens": null,
   "tps": 36.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lpqdndx8-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 105.0 GB at this context; duplicate traces of the same session: ytv0al16; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1138.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,138 36.7 105.0 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lpqdndx8",
   "source_url": "https://omlx.ai/benchmarks/performance/lpqdndx8",
   "tg_tokens": null,
   "tps": 36.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lpqdndx8-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 105.6 GB at this context; duplicate traces of the same session: ytv0al16; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1325.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,325 34.0 105.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lpqdndx8",
   "source_url": "https://omlx.ai/benchmarks/performance/lpqdndx8",
   "tg_tokens": null,
   "tps": 34.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-618v6tzz-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 0o9n2009, swz5ickg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1110.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,110 80.9 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/618v6tzz",
   "source_url": "https://omlx.ai/benchmarks/performance/618v6tzz",
   "tg_tokens": null,
   "tps": 80.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-618v6tzz-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 0o9n2009, swz5ickg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1825.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,825 60.4 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/618v6tzz",
   "source_url": "https://omlx.ai/benchmarks/performance/618v6tzz",
   "tg_tokens": null,
   "tps": 60.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-618v6tzz-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 0o9n2009, swz5ickg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2216.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,216 76.4 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/618v6tzz",
   "source_url": "https://omlx.ai/benchmarks/performance/618v6tzz",
   "tg_tokens": null,
   "tps": 76.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-618v6tzz-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: 0o9n2009, swz5ickg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2319.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,319 74.9 73.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/618v6tzz",
   "source_url": "https://omlx.ai/benchmarks/performance/618v6tzz",
   "tg_tokens": null,
   "tps": 74.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-618v6tzz-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.5 GB at this context; duplicate traces of the same session: 0o9n2009, swz5ickg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2227.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,227 77.0 74.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/618v6tzz",
   "source_url": "https://omlx.ai/benchmarks/performance/618v6tzz",
   "tg_tokens": null,
   "tps": 77.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-618v6tzz-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.2 GB at this context; duplicate traces of the same session: 0o9n2009, swz5ickg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2170.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,170 63.9 79.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/618v6tzz",
   "source_url": "https://omlx.ai/benchmarks/performance/618v6tzz",
   "tg_tokens": null,
   "tps": 63.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-t0fb18g2-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.5 GB at this context; duplicate traces of the same session: js1w5azt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1059.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,059 71.9 70.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t0fb18g2",
   "source_url": "https://omlx.ai/benchmarks/performance/t0fb18g2",
   "tg_tokens": null,
   "tps": 71.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-t0fb18g2-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 82.9 GB at this context; duplicate traces of the same session: js1w5azt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2400.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,400 61.9 82.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t0fb18g2",
   "source_url": "https://omlx.ai/benchmarks/performance/t0fb18g2",
   "tg_tokens": null,
   "tps": 61.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-t3yniyhe-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 101.0 GB; TTFT 3988 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1027.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k tokens 1,027 PP tok/s 26.7 TG tok/s 3988 TTFT (ms) 101.0 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t3yniyhe",
   "source_url": "https://omlx.ai/benchmarks/performance/t3yniyhe",
   "tg_tokens": null,
   "tps": 26.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-3qoyqfpm-ctx1024",
   "model": "medgemma-4b-it",
   "notes": "MLX build; oMLX community benchmark trace, single run at 1k context; peak memory 3.2 GB; TTFT 757 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1352.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "medgemma-4b-it-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k tokens 1,352 PP tok/s 47.9 TG tok/s 757 TTFT (ms) 3.2 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3qoyqfpm",
   "source_url": "https://omlx.ai/benchmarks/performance/3qoyqfpm",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-iz5jj11k-ctx1024",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: rrd3es3o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 216.2,
   "provenance": "community",
   "quant": "6bit",
   "quote": "AREX-2-6bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 6bit; 1k 216.2 28.5 25.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iz5jj11k",
   "source_url": "https://omlx.ai/benchmarks/performance/iz5jj11k",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-iz5jj11k-ctx4096",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: rrd3es3o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 221.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "AREX-2-6bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 6bit; 4k 221.7 42.1 27.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iz5jj11k",
   "source_url": "https://omlx.ai/benchmarks/performance/iz5jj11k",
   "tg_tokens": null,
   "tps": 42.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kun4v8bj-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.0 GB at this context; duplicate traces of the same session: dypiuiwm, hef80473; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 87.4,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 87.4 9.6 92.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kun4v8bj",
   "source_url": "https://omlx.ai/benchmarks/performance/kun4v8bj",
   "tg_tokens": null,
   "tps": 9.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kun4v8bj-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 93.9 GB at this context; duplicate traces of the same session: dypiuiwm, hef80473; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 257.3,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 257.3 12.0 93.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kun4v8bj",
   "source_url": "https://omlx.ai/benchmarks/performance/kun4v8bj",
   "tg_tokens": null,
   "tps": 12.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kun4v8bj-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 94.1 GB at this context; duplicate traces of the same session: dypiuiwm, hef80473; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 122.6,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 122.6 11.8 94.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kun4v8bj",
   "source_url": "https://omlx.ai/benchmarks/performance/kun4v8bj",
   "tg_tokens": null,
   "tps": 11.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-exms02m0-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.0 GB at this context; duplicate traces of the same session: dzdldh22, iubfxtty; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 90.5,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 90.5 10.1 92.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/exms02m0",
   "source_url": "https://omlx.ai/benchmarks/performance/exms02m0",
   "tg_tokens": null,
   "tps": 10.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-exms02m0-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 94.1 GB at this context; duplicate traces of the same session: dzdldh22, iubfxtty; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 283.4,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 283.4 10.3 94.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/exms02m0",
   "source_url": "https://omlx.ai/benchmarks/performance/exms02m0",
   "tg_tokens": null,
   "tps": 10.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-exms02m0-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 94.1 GB at this context; duplicate traces of the same session: dzdldh22, iubfxtty; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 126.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 126.0 11.5 94.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/exms02m0",
   "source_url": "https://omlx.ai/benchmarks/performance/exms02m0",
   "tg_tokens": null,
   "tps": 11.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-cwfzri2c-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 218.9 GB at this context; duplicate traces of the same session: 00rmzxmx, 6h0okd15, d8foyuhj; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 397.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "GLM-5.3-Flash-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 32k 397.0 63.6 218.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cwfzri2c",
   "source_url": "https://omlx.ai/benchmarks/performance/cwfzri2c",
   "tg_tokens": null,
   "tps": 63.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-cwfzri2c-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 221.6 GB at this context; duplicate traces of the same session: 00rmzxmx, 6h0okd15, d8foyuhj; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 391.5,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "GLM-5.3-Flash-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 64k 391.5 41.2 221.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cwfzri2c",
   "source_url": "https://omlx.ai/benchmarks/performance/cwfzri2c",
   "tg_tokens": null,
   "tps": 41.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-cwfzri2c-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 227.4 GB at this context; duplicate traces of the same session: 00rmzxmx, 6h0okd15, d8foyuhj; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 384.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "GLM-5.3-Flash-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 128k 384.2 53.2 227.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cwfzri2c",
   "source_url": "https://omlx.ai/benchmarks/performance/cwfzri2c",
   "tg_tokens": null,
   "tps": 53.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-cwfzri2c-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 233.2 GB at this context; duplicate traces of the same session: 00rmzxmx, 6h0okd15, d8foyuhj; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 377.3,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "GLM-5.3-Flash-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 195k 377.3 32.6 233.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cwfzri2c",
   "source_url": "https://omlx.ai/benchmarks/performance/cwfzri2c",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hoxa0efo-ctx1024",
   "model": "Huihui-Ornith-1.5-35B-A3B-abliterated",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: j6f00znw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3951.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Huihui-Ornith-1.5-35B-A3B-abliterated-oQ5e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 3,951 170.2 26.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hoxa0efo",
   "source_url": "https://omlx.ai/benchmarks/performance/hoxa0efo",
   "tg_tokens": null,
   "tps": 170.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hoxa0efo-ctx4096",
   "model": "Huihui-Ornith-1.5-35B-A3B-abliterated",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: j6f00znw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5380.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Huihui-Ornith-1.5-35B-A3B-abliterated-oQ5e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 5,380 167.5 27.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hoxa0efo",
   "source_url": "https://omlx.ai/benchmarks/performance/hoxa0efo",
   "tg_tokens": null,
   "tps": 167.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-qpoee3i7-ctx1024",
   "model": "Qwen2.5-Coder-1.5B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 2.3 GB at this context; duplicate traces of the same session: 22duqvxk, awgqpkt4, b20ln70c, bdw22q6s; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1064.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen2.5-Coder-1.5B-oQ8e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 8bit; 1k 1,064 83.0 2.3 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qpoee3i7",
   "source_url": "https://omlx.ai/benchmarks/performance/qpoee3i7",
   "tg_tokens": null,
   "tps": 83.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-qpoee3i7-ctx4096",
   "model": "Qwen2.5-Coder-1.5B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 2.3 GB at this context; duplicate traces of the same session: 22duqvxk, awgqpkt4, b20ln70c, bdw22q6s; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1162.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen2.5-Coder-1.5B-oQ8e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 8bit; 4k 1,162 76.4 2.3 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qpoee3i7",
   "source_url": "https://omlx.ai/benchmarks/performance/qpoee3i7",
   "tg_tokens": null,
   "tps": 76.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-qpoee3i7-ctx8192",
   "model": "Qwen2.5-Coder-1.5B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 2.4 GB at this context; duplicate traces of the same session: 22duqvxk, awgqpkt4, b20ln70c, bdw22q6s; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1037.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen2.5-Coder-1.5B-oQ8e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 8bit; 8k 1,037 69.3 2.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qpoee3i7",
   "source_url": "https://omlx.ai/benchmarks/performance/qpoee3i7",
   "tg_tokens": null,
   "tps": 69.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-qpoee3i7-ctx16384",
   "model": "Qwen2.5-Coder-1.5B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 2.6 GB at this context; duplicate traces of the same session: 22duqvxk, awgqpkt4, b20ln70c, bdw22q6s; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 851.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen2.5-Coder-1.5B-oQ8e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 8bit; 16k 851.0 59.1 2.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qpoee3i7",
   "source_url": "https://omlx.ai/benchmarks/performance/qpoee3i7",
   "tg_tokens": null,
   "tps": 59.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-qpoee3i7-ctx32768",
   "model": "Qwen2.5-Coder-1.5B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 3.0 GB at this context; duplicate traces of the same session: 22duqvxk, awgqpkt4, b20ln70c, bdw22q6s; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 621.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen2.5-Coder-1.5B-oQ8e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 8bit; 32k 621.7 44.6 3.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qpoee3i7",
   "source_url": "https://omlx.ai/benchmarks/performance/qpoee3i7",
   "tg_tokens": null,
   "tps": 44.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-knfe4awb-ctx65536",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: luokfosq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1396.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 64k 1,396 45.4 22.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/knfe4awb",
   "source_url": "https://omlx.ai/benchmarks/performance/knfe4awb",
   "tg_tokens": null,
   "tps": 45.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-knfe4awb-ctx131072",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: luokfosq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 464.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 128k 464.5 15.6 28.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/knfe4awb",
   "source_url": "https://omlx.ai/benchmarks/performance/knfe4awb",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-954uxyjz-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 24.4 GB; TTFT 781795 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 167.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 128k tokens 167.7 PP tok/s 12.1 TG tok/s 781795 TTFT (ms) 24.4 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/954uxyjz",
   "source_url": "https://omlx.ai/benchmarks/performance/954uxyjz",
   "tg_tokens": null,
   "tps": 12.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-a8zb99ru-ctx65536",
   "model": "Qwen3.6-35B-A3B-TurboQuant",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: adqa20ku; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2193.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-TurboQuant-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 64k 2,193 92.0 21.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a8zb99ru",
   "source_url": "https://omlx.ai/benchmarks/performance/a8zb99ru",
   "tg_tokens": null,
   "tps": 92.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-a8zb99ru-ctx131072",
   "model": "Qwen3.6-35B-A3B-TurboQuant",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: adqa20ku; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1293.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-TurboQuant-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 128k 1,293 57.0 22.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a8zb99ru",
   "source_url": "https://omlx.ai/benchmarks/performance/a8zb99ru",
   "tg_tokens": null,
   "tps": 57.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-os7k2pxd-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: xvzt46m3; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 131.0 14.3 16.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/os7k2pxd",
   "source_url": "https://omlx.ai/benchmarks/performance/os7k2pxd",
   "tg_tokens": null,
   "tps": 14.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-os7k2pxd-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: xvzt46m3; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 118.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 118.3 13.0 19.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/os7k2pxd",
   "source_url": "https://omlx.ai/benchmarks/performance/os7k2pxd",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4s5agtoi-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: lsi06w75; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 843.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 843.1 33.1 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4s5agtoi",
   "source_url": "https://omlx.ai/benchmarks/performance/4s5agtoi",
   "tg_tokens": null,
   "tps": 33.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4s5agtoi-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: lsi06w75; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 946.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 946.5 32.3 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4s5agtoi",
   "source_url": "https://omlx.ai/benchmarks/performance/4s5agtoi",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-6pzt4b33-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 43.8 GB at this context; duplicate traces of the same session: 5e49ouu6, xg7b34pa; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 419.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 16k 419.0 43.9 43.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6pzt4b33",
   "source_url": "https://omlx.ai/benchmarks/performance/6pzt4b33",
   "tg_tokens": null,
   "tps": 43.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-6pzt4b33-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 48.0 GB at this context; duplicate traces of the same session: 5e49ouu6, xg7b34pa; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 319.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 64k 319.3 22.7 48.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6pzt4b33",
   "source_url": "https://omlx.ai/benchmarks/performance/6pzt4b33",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-6pzt4b33-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 52.5 GB at this context; duplicate traces of the same session: 5e49ouu6, xg7b34pa; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 244.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 128k 244.7 17.7 52.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6pzt4b33",
   "source_url": "https://omlx.ai/benchmarks/performance/6pzt4b33",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-05tysss6-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 19.6 GB; TTFT 62296 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 65.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k tokens 65.8 PP tok/s 13.2 TG tok/s 62296 TTFT (ms) 19.6 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/05tysss6",
   "source_url": "https://omlx.ai/benchmarks/performance/05tysss6",
   "tg_tokens": null,
   "tps": 13.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-5clp31i0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 19.6 GB; TTFT 62425 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 65.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k tokens 65.6 PP tok/s 12.2 TG tok/s 62425 TTFT (ms) 19.6 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5clp31i0",
   "source_url": "https://omlx.ai/benchmarks/performance/5clp31i0",
   "tg_tokens": null,
   "tps": 12.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-x6472b8q-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 17.8 GB; TTFT 61239 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 66.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k tokens 66.9 PP tok/s 6.2 TG tok/s 61239 TTFT (ms) 17.8 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x6472b8q",
   "source_url": "https://omlx.ai/benchmarks/performance/x6472b8q",
   "tg_tokens": null,
   "tps": 6.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-89xyu6eb-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.8 GB at this context; duplicate traces of the same session: 1mqnkg83, f4krw3cv; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 418.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 16k 418.3 42.4 43.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/89xyu6eb",
   "source_url": "https://omlx.ai/benchmarks/performance/89xyu6eb",
   "tg_tokens": null,
   "tps": 42.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-89xyu6eb-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 48.1 GB at this context; duplicate traces of the same session: 1mqnkg83, f4krw3cv; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 320.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 64k 320.2 22.9 48.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/89xyu6eb",
   "source_url": "https://omlx.ai/benchmarks/performance/89xyu6eb",
   "tg_tokens": null,
   "tps": 22.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-89xyu6eb-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 52.6 GB at this context; duplicate traces of the same session: 1mqnkg83, f4krw3cv; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 244.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 128k 244.8 16.9 52.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/89xyu6eb",
   "source_url": "https://omlx.ai/benchmarks/performance/89xyu6eb",
   "tg_tokens": null,
   "tps": 16.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-1dzyl3rw-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "GGUF exact f32 build; oMLX community benchmark trace; peak memory 88.2 GB at this context; duplicate traces of the same session: 65il0x6j; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 514.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-GGUF-exact-f32 on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 1k 514.7 25.3 88.2 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1dzyl3rw",
   "source_url": "https://omlx.ai/benchmarks/performance/1dzyl3rw",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-1dzyl3rw-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "GGUF exact f32 build; oMLX community benchmark trace; peak memory 96.0 GB at this context; duplicate traces of the same session: 65il0x6j; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 563.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-GGUF-exact-f32 on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 4k 563.9 15.4 96.0 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1dzyl3rw",
   "source_url": "https://omlx.ai/benchmarks/performance/1dzyl3rw",
   "tg_tokens": null,
   "tps": 15.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-1cefunse-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: j1v8mlqi, wmmwg0x3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 580.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 580.6 73.6 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1cefunse",
   "source_url": "https://omlx.ai/benchmarks/performance/1cefunse",
   "tg_tokens": null,
   "tps": 73.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-1cefunse-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: j1v8mlqi, wmmwg0x3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 689.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 689.5 57.9 71.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1cefunse",
   "source_url": "https://omlx.ai/benchmarks/performance/1cefunse",
   "tg_tokens": null,
   "tps": 57.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-1cefunse-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: j1v8mlqi, wmmwg0x3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 774.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 774.8 69.9 71.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1cefunse",
   "source_url": "https://omlx.ai/benchmarks/performance/1cefunse",
   "tg_tokens": null,
   "tps": 69.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-h1xyv8xl-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: d8akdihr, o5fwehcx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 515.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 515.9 49.7 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h1xyv8xl",
   "source_url": "https://omlx.ai/benchmarks/performance/h1xyv8xl",
   "tg_tokens": null,
   "tps": 49.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-h1xyv8xl-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: d8akdihr, o5fwehcx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 654.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 654.8 60.7 71.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h1xyv8xl",
   "source_url": "https://omlx.ai/benchmarks/performance/h1xyv8xl",
   "tg_tokens": null,
   "tps": 60.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-h1xyv8xl-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: d8akdihr, o5fwehcx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 772.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 772.5 60.9 71.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h1xyv8xl",
   "source_url": "https://omlx.ai/benchmarks/performance/h1xyv8xl",
   "tg_tokens": null,
   "tps": 60.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bdib5nuu-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: 2vv8smnq, 9t8z57jz, arcdweh3, dmhz1v56, krx01dpe, lsm5likz, qjvuqx1i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1087.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,087 71.1 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bdib5nuu",
   "source_url": "https://omlx.ai/benchmarks/performance/bdib5nuu",
   "tg_tokens": null,
   "tps": 71.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bdib5nuu-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: 2vv8smnq, 9t8z57jz, arcdweh3, dmhz1v56, krx01dpe, lsm5likz, qjvuqx1i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1841.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,841 72.2 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bdib5nuu",
   "source_url": "https://omlx.ai/benchmarks/performance/bdib5nuu",
   "tg_tokens": null,
   "tps": 72.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bdib5nuu-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 2vv8smnq, 9t8z57jz, arcdweh3, dmhz1v56, krx01dpe, lsm5likz, qjvuqx1i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2174.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,174 83.4 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bdib5nuu",
   "source_url": "https://omlx.ai/benchmarks/performance/bdib5nuu",
   "tg_tokens": null,
   "tps": 83.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bdib5nuu-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: 2vv8smnq, 9t8z57jz, arcdweh3, dmhz1v56, krx01dpe, lsm5likz, qjvuqx1i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2276.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,276 65.9 72.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bdib5nuu",
   "source_url": "https://omlx.ai/benchmarks/performance/bdib5nuu",
   "tg_tokens": null,
   "tps": 65.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bdib5nuu-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: 2vv8smnq, 9t8z57jz, arcdweh3, dmhz1v56, krx01dpe, lsm5likz, qjvuqx1i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2028.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,028 64.9 73.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bdib5nuu",
   "source_url": "https://omlx.ai/benchmarks/performance/bdib5nuu",
   "tg_tokens": null,
   "tps": 64.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bdib5nuu-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 75.5 GB at this context; duplicate traces of the same session: 2vv8smnq, 9t8z57jz, arcdweh3, dmhz1v56, krx01dpe, lsm5likz, qjvuqx1i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2073.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,073 69.7 75.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bdib5nuu",
   "source_url": "https://omlx.ai/benchmarks/performance/bdib5nuu",
   "tg_tokens": null,
   "tps": 69.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bdib5nuu-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.3 GB at this context; duplicate traces of the same session: 2vv8smnq, 9t8z57jz, arcdweh3, dmhz1v56, krx01dpe, lsm5likz, qjvuqx1i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2030.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,030 63.6 79.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bdib5nuu",
   "source_url": "https://omlx.ai/benchmarks/performance/bdib5nuu",
   "tg_tokens": null,
   "tps": 63.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bdib5nuu-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.9 GB at this context; duplicate traces of the same session: 2vv8smnq, 9t8z57jz, arcdweh3, dmhz1v56, krx01dpe, lsm5likz, qjvuqx1i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1996.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 1,996 52.2 80.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bdib5nuu",
   "source_url": "https://omlx.ai/benchmarks/performance/bdib5nuu",
   "tg_tokens": null,
   "tps": 52.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-40iuxvlr-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 11ia9z6t, aeip93ov, er03lxji, jtasb3hh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 933.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 933.3 55.2 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/40iuxvlr",
   "source_url": "https://omlx.ai/benchmarks/performance/40iuxvlr",
   "tg_tokens": null,
   "tps": 55.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-40iuxvlr-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 11ia9z6t, aeip93ov, er03lxji, jtasb3hh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2011.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,011 65.8 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/40iuxvlr",
   "source_url": "https://omlx.ai/benchmarks/performance/40iuxvlr",
   "tg_tokens": null,
   "tps": 65.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-40iuxvlr-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: 11ia9z6t, aeip93ov, er03lxji, jtasb3hh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2093.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,093 51.5 73.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/40iuxvlr",
   "source_url": "https://omlx.ai/benchmarks/performance/40iuxvlr",
   "tg_tokens": null,
   "tps": 51.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-40iuxvlr-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.8 GB at this context; duplicate traces of the same session: 11ia9z6t, aeip93ov, er03lxji, jtasb3hh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2056.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,056 65.0 74.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/40iuxvlr",
   "source_url": "https://omlx.ai/benchmarks/performance/40iuxvlr",
   "tg_tokens": null,
   "tps": 65.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-40iuxvlr-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.1 GB at this context; duplicate traces of the same session: 11ia9z6t, aeip93ov, er03lxji, jtasb3hh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2007.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,007 54.5 79.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/40iuxvlr",
   "source_url": "https://omlx.ai/benchmarks/performance/40iuxvlr",
   "tg_tokens": null,
   "tps": 54.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-40iuxvlr-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.9 GB at this context; duplicate traces of the same session: 11ia9z6t, aeip93ov, er03lxji, jtasb3hh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1993.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 1,993 52.4 80.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/40iuxvlr",
   "source_url": "https://omlx.ai/benchmarks/performance/40iuxvlr",
   "tg_tokens": null,
   "tps": 52.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-o2vdzk8k-ctx65536",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 13.2 GB at this context; duplicate traces of the same session: salwb70g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1657.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 64k 1,657 45.2 13.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o2vdzk8k",
   "source_url": "https://omlx.ai/benchmarks/performance/o2vdzk8k",
   "tg_tokens": null,
   "tps": 45.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-o2vdzk8k-ctx131072",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: salwb70g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1131.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 128k 1,131 30.0 14.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o2vdzk8k",
   "source_url": "https://omlx.ai/benchmarks/performance/o2vdzk8k",
   "tg_tokens": null,
   "tps": 30.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-s5qwpk8e-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 214.7 GB at this context; duplicate traces of the same session: 27u2mljm, f0j1yhwq, f9ilml6s, k66oystc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 403.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "GLM-5.3-Flash-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 1k 403.9 35.4 214.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s5qwpk8e",
   "source_url": "https://omlx.ai/benchmarks/performance/s5qwpk8e",
   "tg_tokens": null,
   "tps": 35.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-s5qwpk8e-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 216.1 GB at this context; duplicate traces of the same session: 27u2mljm, f0j1yhwq, f9ilml6s, k66oystc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 422.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "GLM-5.3-Flash-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 4k 422.1 31.2 216.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s5qwpk8e",
   "source_url": "https://omlx.ai/benchmarks/performance/s5qwpk8e",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-s5qwpk8e-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 216.9 GB at this context; duplicate traces of the same session: 27u2mljm, f0j1yhwq, f9ilml6s, k66oystc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 409.8,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "GLM-5.3-Flash-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 8k 409.8 47.4 216.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s5qwpk8e",
   "source_url": "https://omlx.ai/benchmarks/performance/s5qwpk8e",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-s5qwpk8e-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 217.6 GB at this context; duplicate traces of the same session: 27u2mljm, f0j1yhwq, f9ilml6s, k66oystc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 402.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "GLM-5.3-Flash-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 16k 402.1 62.6 217.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s5qwpk8e",
   "source_url": "https://omlx.ai/benchmarks/performance/s5qwpk8e",
   "tg_tokens": null,
   "tps": 62.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-s5qwpk8e-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 218.9 GB at this context; duplicate traces of the same session: 27u2mljm, f0j1yhwq, f9ilml6s, k66oystc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 396.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "GLM-5.3-Flash-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 32k 396.9 65.4 218.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s5qwpk8e",
   "source_url": "https://omlx.ai/benchmarks/performance/s5qwpk8e",
   "tg_tokens": null,
   "tps": 65.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2v3kdjvu-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 212.5 GB at this context; duplicate traces of the same session: gkw142g8, k941t2v8, rsi1slm7, ys6ojye1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 388.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "GLM-5.3-Flash-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 1k 388.0 46.2 212.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2v3kdjvu",
   "source_url": "https://omlx.ai/benchmarks/performance/2v3kdjvu",
   "tg_tokens": null,
   "tps": 46.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2v3kdjvu-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 213.6 GB at this context; duplicate traces of the same session: gkw142g8, k941t2v8, rsi1slm7, ys6ojye1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 438.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "GLM-5.3-Flash-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 4k 438.0 45.1 213.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2v3kdjvu",
   "source_url": "https://omlx.ai/benchmarks/performance/2v3kdjvu",
   "tg_tokens": null,
   "tps": 45.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2v3kdjvu-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 214.1 GB at this context; duplicate traces of the same session: gkw142g8, k941t2v8, rsi1slm7, ys6ojye1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 405.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "GLM-5.3-Flash-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 8k 405.9 52.4 214.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2v3kdjvu",
   "source_url": "https://omlx.ai/benchmarks/performance/2v3kdjvu",
   "tg_tokens": null,
   "tps": 52.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2v3kdjvu-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 214.5 GB at this context; duplicate traces of the same session: gkw142g8, k941t2v8, rsi1slm7, ys6ojye1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 398.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "GLM-5.3-Flash-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 16k 398.1 74.2 214.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2v3kdjvu",
   "source_url": "https://omlx.ai/benchmarks/performance/2v3kdjvu",
   "tg_tokens": null,
   "tps": 74.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2v3kdjvu-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 215.3 GB at this context; duplicate traces of the same session: gkw142g8, k941t2v8, rsi1slm7, ys6ojye1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 392.8,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "GLM-5.3-Flash-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 32k 392.8 58.7 215.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2v3kdjvu",
   "source_url": "https://omlx.ai/benchmarks/performance/2v3kdjvu",
   "tg_tokens": null,
   "tps": 58.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-z17r4237-ctx1024",
   "model": "Laguna XS 2.1",
   "notes": "oMLX community benchmark trace; peak memory 18.2 GB at this context; duplicate traces of the same session: ga9i02jw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 973.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Laguna-XS-2.1-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 973.2 97.1 18.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z17r4237",
   "source_url": "https://omlx.ai/benchmarks/performance/z17r4237",
   "tg_tokens": null,
   "tps": 97.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-z17r4237-ctx4096",
   "model": "Laguna XS 2.1",
   "notes": "oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: ga9i02jw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1101.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Laguna-XS-2.1-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 1,101 91.3 18.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z17r4237",
   "source_url": "https://omlx.ai/benchmarks/performance/z17r4237",
   "tg_tokens": null,
   "tps": 91.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-suh0jwpr-ctx8192",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "oMLX community benchmark trace, single run at 8k context; peak memory 8.3 GB; TTFT 35018 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 233.9,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-OptiQ-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 8k tokens 233.9 PP tok/s 14.6 TG tok/s 35018 TTFT (ms) 8.3 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/suh0jwpr",
   "source_url": "https://omlx.ai/benchmarks/performance/suh0jwpr",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-k6aoa6h5-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: 5kacp3pp, 8wfqw3fi, dogcdrxm, slduo5i2, tz020ekv; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3459.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 1k 3,459 89.1 37.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k6aoa6h5",
   "source_url": "https://omlx.ai/benchmarks/performance/k6aoa6h5",
   "tg_tokens": null,
   "tps": 89.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-k6aoa6h5-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.0 GB at this context; duplicate traces of the same session: 5kacp3pp, 8wfqw3fi, dogcdrxm, slduo5i2, tz020ekv; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 8452.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 4k 8,452 169.8 38.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k6aoa6h5",
   "source_url": "https://omlx.ai/benchmarks/performance/k6aoa6h5",
   "tg_tokens": null,
   "tps": 169.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-k6aoa6h5-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.1 GB at this context; duplicate traces of the same session: 5kacp3pp, 8wfqw3fi, dogcdrxm, slduo5i2, tz020ekv; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 8514.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 8k 8,514 218.4 38.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k6aoa6h5",
   "source_url": "https://omlx.ai/benchmarks/performance/k6aoa6h5",
   "tg_tokens": null,
   "tps": 218.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-k6aoa6h5-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.2 GB at this context; duplicate traces of the same session: 5kacp3pp, 8wfqw3fi, dogcdrxm, slduo5i2, tz020ekv; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 8016.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 16k 8,016 267.9 38.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k6aoa6h5",
   "source_url": "https://omlx.ai/benchmarks/performance/k6aoa6h5",
   "tg_tokens": null,
   "tps": 267.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-k6aoa6h5-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.6 GB at this context; duplicate traces of the same session: 5kacp3pp, 8wfqw3fi, dogcdrxm, slduo5i2, tz020ekv; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6957.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 32k 6,957 213.3 38.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k6aoa6h5",
   "source_url": "https://omlx.ai/benchmarks/performance/k6aoa6h5",
   "tg_tokens": null,
   "tps": 213.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-k6aoa6h5-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.4 GB at this context; duplicate traces of the same session: 5kacp3pp, 8wfqw3fi, dogcdrxm, slduo5i2, tz020ekv; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5394.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 64k 5,394 182.0 39.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k6aoa6h5",
   "source_url": "https://omlx.ai/benchmarks/performance/k6aoa6h5",
   "tg_tokens": null,
   "tps": 182.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-th9me5dg-ctx8192",
   "model": "Qwen3 8B",
   "notes": "oMLX community benchmark trace, single run at 8k context; peak memory 5.7 GB; TTFT 38886 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 210.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-8B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 8k tokens 210.7 PP tok/s 18.3 TG tok/s 38886 TTFT (ms) 5.7 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/th9me5dg",
   "source_url": "https://omlx.ai/benchmarks/performance/th9me5dg",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-np3t3ury-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 3.9 GB; TTFT 9379 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 436.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-4B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k tokens 436.7 PP tok/s 34.9 TG tok/s 9379 TTFT (ms) 3.9 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/np3t3ury",
   "source_url": "https://omlx.ai/benchmarks/performance/np3t3ury",
   "tg_tokens": null,
   "tps": 34.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-19ze3r97-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 11.3 GB at this context; duplicate traces of the same session: t6cl33h4; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 185.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B-MLX-8bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 8bit; 1k 185.1 18.2 11.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/19ze3r97",
   "source_url": "https://omlx.ai/benchmarks/performance/19ze3r97",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-19ze3r97-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 11.4 GB at this context; duplicate traces of the same session: t6cl33h4; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 247.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B-MLX-8bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 8bit; 4k 247.9 17.5 11.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/19ze3r97",
   "source_url": "https://omlx.ai/benchmarks/performance/19ze3r97",
   "tg_tokens": null,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-1oth1x22-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 22.3 GB; TTFT 187330 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 699.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 128k tokens 699.7 PP tok/s 23.1 TG tok/s 187330 TTFT (ms) 22.3 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1oth1x22",
   "source_url": "https://omlx.ai/benchmarks/performance/1oth1x22",
   "tg_tokens": null,
   "tps": 23.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-n4h8c7q9-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 21.0 GB; TTFT 58107 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1128.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 64k tokens 1,128 PP tok/s 36.7 TG tok/s 58107 TTFT (ms) 21.0 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n4h8c7q9",
   "source_url": "https://omlx.ai/benchmarks/performance/n4h8c7q9",
   "tg_tokens": null,
   "tps": 36.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-azr2e7n7-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.2 GB; TTFT 1526 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 670.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k tokens 670.8 PP tok/s 67.9 TG tok/s 1526 TTFT (ms) 19.2 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/azr2e7n7",
   "source_url": "https://omlx.ai/benchmarks/performance/azr2e7n7",
   "tg_tokens": null,
   "tps": 67.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-erj0rpsj-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 20.4 GB; TTFT 22027 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1488.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 32k tokens 1,488 PP tok/s 49.9 TG tok/s 22027 TTFT (ms) 20.4 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/erj0rpsj",
   "source_url": "https://omlx.ai/benchmarks/performance/erj0rpsj",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-fxhg1n7s-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 226.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 226.2 23.0 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fxhg1n7s",
   "source_url": "https://omlx.ai/benchmarks/performance/fxhg1n7s",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-fxhg1n7s-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 239.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 239.3 22.7 19.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fxhg1n7s",
   "source_url": "https://omlx.ai/benchmarks/performance/fxhg1n7s",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5mhpzecj-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: hgw7kll2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 935.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 935.5 33.7 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5mhpzecj",
   "source_url": "https://omlx.ai/benchmarks/performance/5mhpzecj",
   "tg_tokens": null,
   "tps": 33.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5mhpzecj-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: hgw7kll2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1024.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,024 32.9 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5mhpzecj",
   "source_url": "https://omlx.ai/benchmarks/performance/5mhpzecj",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-x4pdj7x1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 20.9 GB; TTFT 14252 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 287.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k tokens 287.4 PP tok/s 18.3 TG tok/s 14252 TTFT (ms) 20.9 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x4pdj7x1",
   "source_url": "https://omlx.ai/benchmarks/performance/x4pdj7x1",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-glbcidaa-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.6 GB; TTFT 3763 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 272.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k tokens 272.1 PP tok/s 13.5 TG tok/s 3763 TTFT (ms) 19.6 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/glbcidaa",
   "source_url": "https://omlx.ai/benchmarks/performance/glbcidaa",
   "tg_tokens": null,
   "tps": 13.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-pglc3gc9-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 22.9 GB; TTFT 134710 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 243.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 32k tokens 243.2 PP tok/s 5.8 TG tok/s 134710 TTFT (ms) 22.9 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pglc3gc9",
   "source_url": "https://omlx.ai/benchmarks/performance/pglc3gc9",
   "tg_tokens": null,
   "tps": 5.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-j5h3vc1g-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 08n1176s; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 427.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 427.3 22.1 28.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j5h3vc1g",
   "source_url": "https://omlx.ai/benchmarks/performance/j5h3vc1g",
   "tg_tokens": null,
   "tps": 22.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-j5h3vc1g-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.7 GB at this context; duplicate traces of the same session: 08n1176s; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 444.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 444.7 21.9 31.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j5h3vc1g",
   "source_url": "https://omlx.ai/benchmarks/performance/j5h3vc1g",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-94ormxfe-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: qd9zi3os; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 429.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 429.5 21.7 28.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/94ormxfe",
   "source_url": "https://omlx.ai/benchmarks/performance/94ormxfe",
   "tg_tokens": null,
   "tps": 21.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-94ormxfe-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.7 GB at this context; duplicate traces of the same session: qd9zi3os; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 450.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 450.8 21.3 31.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/94ormxfe",
   "source_url": "https://omlx.ai/benchmarks/performance/94ormxfe",
   "tg_tokens": null,
   "tps": 21.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-iq4dbin8-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: xo2nuk29; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 833.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ8e-1M on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 833.8 18.8 28.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iq4dbin8",
   "source_url": "https://omlx.ai/benchmarks/performance/iq4dbin8",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-iq4dbin8-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: xo2nuk29; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 906.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ8e-1M on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 906.9 18.5 29.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iq4dbin8",
   "source_url": "https://omlx.ai/benchmarks/performance/iq4dbin8",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-86xqc8ln-ctx1024",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 33.0 GB; TTFT 4806 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 213.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k tokens 213.1 PP tok/s 22.8 TG tok/s 4806 TTFT (ms) 33.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/86xqc8ln",
   "source_url": "https://omlx.ai/benchmarks/performance/86xqc8ln",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ex14iytk-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 18.5 GB; TTFT 4054 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 252.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 252.6 PP tok/s 65.0 TG tok/s 4054 TTFT (ms) 18.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ex14iytk",
   "source_url": "https://omlx.ai/benchmarks/performance/ex14iytk",
   "tg_tokens": null,
   "tps": 65.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-216rrg3z-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 17.8 GB; TTFT 61055 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 67.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k tokens 67.1 PP tok/s 6.0 TG tok/s 61055 TTFT (ms) 17.8 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/216rrg3z",
   "source_url": "https://omlx.ai/benchmarks/performance/216rrg3z",
   "tg_tokens": null,
   "tps": 6.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-gd6qyjq0-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: 3beygko2, 46putd0h, f1hl71sj, vz3ju1te; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 249.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 249.3 41.5 33.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gd6qyjq0",
   "source_url": "https://omlx.ai/benchmarks/performance/gd6qyjq0",
   "tg_tokens": null,
   "tps": 41.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-gd6qyjq0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: 3beygko2, 46putd0h, f1hl71sj, vz3ju1te; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 243.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 243.3 30.3 37.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gd6qyjq0",
   "source_url": "https://omlx.ai/benchmarks/performance/gd6qyjq0",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-gd6qyjq0-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.8 GB at this context; duplicate traces of the same session: 3beygko2, 46putd0h, f1hl71sj, vz3ju1te; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 216.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 216.5 38.7 37.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gd6qyjq0",
   "source_url": "https://omlx.ai/benchmarks/performance/gd6qyjq0",
   "tg_tokens": null,
   "tps": 38.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-gd6qyjq0-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.7 GB at this context; duplicate traces of the same session: 3beygko2, 46putd0h, f1hl71sj, vz3ju1te; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 215.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 215.9 46.2 38.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gd6qyjq0",
   "source_url": "https://omlx.ai/benchmarks/performance/gd6qyjq0",
   "tg_tokens": null,
   "tps": 46.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-gd6qyjq0-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 40.4 GB at this context; duplicate traces of the same session: 3beygko2, 46putd0h, f1hl71sj, vz3ju1te; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 168.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 168.2 33.6 40.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gd6qyjq0",
   "source_url": "https://omlx.ai/benchmarks/performance/gd6qyjq0",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-esm95a7z-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 105.2 GB at this context; duplicate traces of the same session: cmd7877u; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 810.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 810.8 37.8 105.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/esm95a7z",
   "source_url": "https://omlx.ai/benchmarks/performance/esm95a7z",
   "tg_tokens": null,
   "tps": 37.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-esm95a7z-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 105.6 GB at this context; duplicate traces of the same session: cmd7877u; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1082.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,082 33.3 105.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/esm95a7z",
   "source_url": "https://omlx.ai/benchmarks/performance/esm95a7z",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-339235d3-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 106.0 GB at this context; duplicate traces of the same session: 2yw71gd8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1299.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 1,299 34.6 106.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/339235d3",
   "source_url": "https://omlx.ai/benchmarks/performance/339235d3",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-339235d3-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 106.3 GB at this context; duplicate traces of the same session: 2yw71gd8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1293.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 1,293 22.1 106.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/339235d3",
   "source_url": "https://omlx.ai/benchmarks/performance/339235d3",
   "tg_tokens": null,
   "tps": 22.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0mbsn4bk-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 175.0 GB at this context; duplicate traces of the same session: m87d3fpk; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 795.2,
   "provenance": "community",
   "quant": "FP8",
   "quote": "Qwen3.8-Flash-Next-FP8 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 795.2 32.6 175.0 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0mbsn4bk",
   "source_url": "https://omlx.ai/benchmarks/performance/0mbsn4bk",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0mbsn4bk-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 175.5 GB at this context; duplicate traces of the same session: m87d3fpk; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1017.0,
   "provenance": "community",
   "quant": "FP8",
   "quote": "Qwen3.8-Flash-Next-FP8 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 1,017 30.0 175.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0mbsn4bk",
   "source_url": "https://omlx.ai/benchmarks/performance/0mbsn4bk",
   "tg_tokens": null,
   "tps": 30.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-hd2wg6b4-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace, single run at 1k context; peak memory 86.4 GB; TTFT 2136 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 479.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k tokens 479.4 PP tok/s 49.3 TG tok/s 2136 TTFT (ms) 86.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hd2wg6b4",
   "source_url": "https://omlx.ai/benchmarks/performance/hd2wg6b4",
   "tg_tokens": null,
   "tps": 49.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-94w1qy1i-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.6 GB; TTFT 3130 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 327.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 327.1 PP tok/s 41.9 TG tok/s 3130 TTFT (ms) 70.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/94w1qy1i",
   "source_url": "https://omlx.ai/benchmarks/performance/94w1qy1i",
   "tg_tokens": null,
   "tps": 41.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-o9klrgo5-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 933.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 933.3 55.2 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9klrgo5",
   "source_url": "https://omlx.ai/benchmarks/performance/o9klrgo5",
   "tg_tokens": null,
   "tps": 55.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-o9klrgo5-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2011.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,011 65.8 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9klrgo5",
   "source_url": "https://omlx.ai/benchmarks/performance/o9klrgo5",
   "tg_tokens": null,
   "tps": 65.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-o9klrgo5-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2093.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,093 51.5 73.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9klrgo5",
   "source_url": "https://omlx.ai/benchmarks/performance/o9klrgo5",
   "tg_tokens": null,
   "tps": 51.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-o9klrgo5-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2056.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,056 65.0 74.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9klrgo5",
   "source_url": "https://omlx.ai/benchmarks/performance/o9klrgo5",
   "tg_tokens": null,
   "tps": 65.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-o9klrgo5-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2007.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,007 54.5 79.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9klrgo5",
   "source_url": "https://omlx.ai/benchmarks/performance/o9klrgo5",
   "tg_tokens": null,
   "tps": 54.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-o9klrgo5-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1993.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 1,993 52.4 80.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9klrgo5",
   "source_url": "https://omlx.ai/benchmarks/performance/o9klrgo5",
   "tg_tokens": null,
   "tps": 52.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-by05ofu8-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: if2k8nyt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 259.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 259.6 28.2 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/by05ofu8",
   "source_url": "https://omlx.ai/benchmarks/performance/by05ofu8",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-by05ofu8-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: if2k8nyt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 274.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 274.1 27.2 19.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/by05ofu8",
   "source_url": "https://omlx.ai/benchmarks/performance/by05ofu8",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-mb7v3kk5-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 5.8 GB; TTFT 6387 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 641.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k tokens 641.3 PP tok/s 46.5 TG tok/s 6387 TTFT (ms) 5.8 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mb7v3kk5",
   "source_url": "https://omlx.ai/benchmarks/performance/mb7v3kk5",
   "tg_tokens": null,
   "tps": 46.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Pro",
   "id": "omlx-51i697ga-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 4.7 GB; TTFT 1323 ms; duplicate traces of the same session: none; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 774.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 1k tokens 774.2 PP tok/s 35.8 TG tok/s 1323 TTFT (ms) 4.7 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/51i697ga",
   "source_url": "https://omlx.ai/benchmarks/performance/51i697ga",
   "tg_tokens": null,
   "tps": 35.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-b3cuygh7-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.9 GB at this context; duplicate traces of the same session: ec8xr0y5, w6zs5izv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1260.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 1,260 104.0 13.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b3cuygh7",
   "source_url": "https://omlx.ai/benchmarks/performance/b3cuygh7",
   "tg_tokens": null,
   "tps": 104.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-b3cuygh7-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 15.5 GB at this context; duplicate traces of the same session: ec8xr0y5, w6zs5izv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1177.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 1,177 88.2 15.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b3cuygh7",
   "source_url": "https://omlx.ai/benchmarks/performance/b3cuygh7",
   "tg_tokens": null,
   "tps": 88.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-b3cuygh7-ctx65536",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: ec8xr0y5, w6zs5izv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 753.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 753.9 54.2 22.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b3cuygh7",
   "source_url": "https://omlx.ai/benchmarks/performance/b3cuygh7",
   "tg_tokens": null,
   "tps": 54.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ico9ek2y-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 1k context; peak memory 24.3 GB; TTFT 474 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2162.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k tokens 2,162 PP tok/s 128.6 TG tok/s 474 TTFT (ms) 24.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ico9ek2y",
   "source_url": "https://omlx.ai/benchmarks/performance/ico9ek2y",
   "tg_tokens": null,
   "tps": 128.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-fb5a1rsx-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: c8dqg3nh, wjv6s8nw, xjk53pnk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 973.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 973.4 92.2 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fb5a1rsx",
   "source_url": "https://omlx.ai/benchmarks/performance/fb5a1rsx",
   "tg_tokens": null,
   "tps": 92.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-fb5a1rsx-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: c8dqg3nh, wjv6s8nw, xjk53pnk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1083.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,083 88.9 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fb5a1rsx",
   "source_url": "https://omlx.ai/benchmarks/performance/fb5a1rsx",
   "tg_tokens": null,
   "tps": 88.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-fb5a1rsx-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: c8dqg3nh, wjv6s8nw, xjk53pnk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1021.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 1,021 83.0 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fb5a1rsx",
   "source_url": "https://omlx.ai/benchmarks/performance/fb5a1rsx",
   "tg_tokens": null,
   "tps": 83.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-fb5a1rsx-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: c8dqg3nh, wjv6s8nw, xjk53pnk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 892.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 892.9 75.9 21.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fb5a1rsx",
   "source_url": "https://omlx.ai/benchmarks/performance/fb5a1rsx",
   "tg_tokens": null,
   "tps": 75.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-4fwdnvw4-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 12nefqs0, 6kxx2n99, ikx9hqz2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 870.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 870.1 92.0 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4fwdnvw4",
   "source_url": "https://omlx.ai/benchmarks/performance/4fwdnvw4",
   "tg_tokens": null,
   "tps": 92.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-4fwdnvw4-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 12nefqs0, 6kxx2n99, ikx9hqz2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1045.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,045 88.7 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4fwdnvw4",
   "source_url": "https://omlx.ai/benchmarks/performance/4fwdnvw4",
   "tg_tokens": null,
   "tps": 88.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-4fwdnvw4-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 12nefqs0, 6kxx2n99, ikx9hqz2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1018.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 1,018 83.6 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4fwdnvw4",
   "source_url": "https://omlx.ai/benchmarks/performance/4fwdnvw4",
   "tg_tokens": null,
   "tps": 83.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-4fwdnvw4-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 12nefqs0, 6kxx2n99, ikx9hqz2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 910.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 910.3 76.2 21.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4fwdnvw4",
   "source_url": "https://omlx.ai/benchmarks/performance/4fwdnvw4",
   "tg_tokens": null,
   "tps": 76.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-vs9ia5cc-ctx65536",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: 6moagz9j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1338.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 64k 1,338 44.7 22.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vs9ia5cc",
   "source_url": "https://omlx.ai/benchmarks/performance/vs9ia5cc",
   "tg_tokens": null,
   "tps": 44.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-vs9ia5cc-ctx131072",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 6moagz9j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 486.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 128k 486.1 17.2 28.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vs9ia5cc",
   "source_url": "https://omlx.ai/benchmarks/performance/vs9ia5cc",
   "tg_tokens": null,
   "tps": 17.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-zg6m3vzo-ctx199680",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace, single run at 195k context; peak memory 12.2 GB; TTFT 422147 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 473.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 195k tokens 473.8 PP tok/s 10.1 TG tok/s 422147 TTFT (ms) 12.2 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zg6m3vzo",
   "source_url": "https://omlx.ai/benchmarks/performance/zg6m3vzo",
   "tg_tokens": null,
   "tps": 10.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-tzb7yg8u-ctx131072",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 10.1 GB; TTFT 221809 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 590.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 128k tokens 590.9 PP tok/s 14.0 TG tok/s 221809 TTFT (ms) 10.1 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tzb7yg8u",
   "source_url": "https://omlx.ai/benchmarks/performance/tzb7yg8u",
   "tg_tokens": null,
   "tps": 14.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-bb8n0j08-ctx32768",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 7.1 GB; TTFT 34265 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 956.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 32k tokens 956.3 PP tok/s 24.5 TG tok/s 34265 TTFT (ms) 7.1 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bb8n0j08",
   "source_url": "https://omlx.ai/benchmarks/performance/bb8n0j08",
   "tg_tokens": null,
   "tps": 24.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-dflok438-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 5.7 GB; TTFT 1051 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 974.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k tokens 974.6 PP tok/s 30.8 TG tok/s 1051 TTFT (ms) 5.7 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dflok438",
   "source_url": "https://omlx.ai/benchmarks/performance/dflok438",
   "tg_tokens": null,
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-v2w12rfj-ctx65536",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 8.1 GB; TTFT 82407 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 795.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 64k tokens 795.3 PP tok/s 20.1 TG tok/s 82407 TTFT (ms) 8.1 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v2w12rfj",
   "source_url": "https://omlx.ai/benchmarks/performance/v2w12rfj",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-ydajbgst-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: k63vwzh7, uovr2p2i; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 960.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 1k 960.4 59.6 21.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ydajbgst",
   "source_url": "https://omlx.ai/benchmarks/performance/ydajbgst",
   "tg_tokens": null,
   "tps": 59.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-ydajbgst-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: k63vwzh7, uovr2p2i; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1290.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 4k 1,290 72.4 21.8 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ydajbgst",
   "source_url": "https://omlx.ai/benchmarks/performance/ydajbgst",
   "tg_tokens": null,
   "tps": 72.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-ydajbgst-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: k63vwzh7, uovr2p2i; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1255.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 8k 1,255 71.1 21.9 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ydajbgst",
   "source_url": "https://omlx.ai/benchmarks/performance/ydajbgst",
   "tg_tokens": null,
   "tps": 71.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-0ju9y6go-ctx1024",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 3euqwb28, 5d5zuli4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1765.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,765 114.4 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ju9y6go",
   "source_url": "https://omlx.ai/benchmarks/performance/0ju9y6go",
   "tg_tokens": null,
   "tps": 114.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-0ju9y6go-ctx4096",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 3euqwb28, 5d5zuli4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2032.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 2,032 106.8 21.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ju9y6go",
   "source_url": "https://omlx.ai/benchmarks/performance/0ju9y6go",
   "tg_tokens": null,
   "tps": 106.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-0ju9y6go-ctx8192",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 3euqwb28, 5d5zuli4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1775.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,775 99.7 22.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ju9y6go",
   "source_url": "https://omlx.ai/benchmarks/performance/0ju9y6go",
   "tg_tokens": null,
   "tps": 99.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-u8ltlg2v-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.3 GB; TTFT 224 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4563.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 4,563 PP tok/s 149.4 TG tok/s 224 TTFT (ms) 19.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u8ltlg2v",
   "source_url": "https://omlx.ai/benchmarks/performance/u8ltlg2v",
   "tg_tokens": null,
   "tps": 149.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-4gp26ahi-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: lxmffrpg, o5dq3vj9; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1188.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 6bit; 4k 1,188 74.6 29.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4gp26ahi",
   "source_url": "https://omlx.ai/benchmarks/performance/4gp26ahi",
   "tg_tokens": null,
   "tps": 74.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-4gp26ahi-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 31.1 GB at this context; duplicate traces of the same session: lxmffrpg, o5dq3vj9; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 988.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 6bit; 16k 988.3 67.2 31.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4gp26ahi",
   "source_url": "https://omlx.ai/benchmarks/performance/4gp26ahi",
   "tg_tokens": null,
   "tps": 67.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-4gp26ahi-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 32.4 GB at this context; duplicate traces of the same session: lxmffrpg, o5dq3vj9; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 795.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 6bit; 32k 795.6 60.5 32.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4gp26ahi",
   "source_url": "https://omlx.ai/benchmarks/performance/4gp26ahi",
   "tg_tokens": null,
   "tps": 60.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-flp4bmvp-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 226.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 226.2 23.0 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/flp4bmvp",
   "source_url": "https://omlx.ai/benchmarks/performance/flp4bmvp",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-flp4bmvp-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 239.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 239.3 22.7 19.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/flp4bmvp",
   "source_url": "https://omlx.ai/benchmarks/performance/flp4bmvp",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ga59forh-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 17.2 GB; TTFT 4008 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1022.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 1,022 PP tok/s 33.0 TG tok/s 4008 TTFT (ms) 17.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ga59forh",
   "source_url": "https://omlx.ai/benchmarks/performance/ga59forh",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-3u07x1um-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized speed FP16 build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 9jx1785u, hagyyifb, pt9a59r3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed-FP16 on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 130.0 12.7 20.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3u07x1um",
   "source_url": "https://omlx.ai/benchmarks/performance/3u07x1um",
   "tg_tokens": null,
   "tps": 12.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-3u07x1um-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized speed FP16 build; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: 9jx1785u, hagyyifb, pt9a59r3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed-FP16 on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 131.4 12.5 23.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3u07x1um",
   "source_url": "https://omlx.ai/benchmarks/performance/3u07x1um",
   "tg_tokens": null,
   "tps": 12.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-3u07x1um-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized speed FP16 build; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: 9jx1785u, hagyyifb, pt9a59r3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 120.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed-FP16 on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 120.3 12.3 24.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3u07x1um",
   "source_url": "https://omlx.ai/benchmarks/performance/3u07x1um",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-3u07x1um-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized speed FP16 build; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: 9jx1785u, hagyyifb, pt9a59r3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 117.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed-FP16 on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 117.8 11.8 25.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3u07x1um",
   "source_url": "https://omlx.ai/benchmarks/performance/3u07x1um",
   "tg_tokens": null,
   "tps": 11.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-oljhuicy-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: ga7gloxj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1130.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,130 69.6 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oljhuicy",
   "source_url": "https://omlx.ai/benchmarks/performance/oljhuicy",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-oljhuicy-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: ga7gloxj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1377.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,377 63.9 19.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oljhuicy",
   "source_url": "https://omlx.ai/benchmarks/performance/oljhuicy",
   "tg_tokens": null,
   "tps": 63.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-h12wy78u-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 19.2 GB; TTFT 3002 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1365.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 1,365 PP tok/s 68.4 TG tok/s 3002 TTFT (ms) 19.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h12wy78u",
   "source_url": "https://omlx.ai/benchmarks/performance/h12wy78u",
   "tg_tokens": null,
   "tps": 68.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-h25i51n4-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: acqi0dnv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 425.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 425.4 24.6 17.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h25i51n4",
   "source_url": "https://omlx.ai/benchmarks/performance/h25i51n4",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-h25i51n4-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: acqi0dnv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 456.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 456.1 26.7 18.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h25i51n4",
   "source_url": "https://omlx.ai/benchmarks/performance/h25i51n4",
   "tg_tokens": null,
   "tps": 26.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-3x7pnaj7-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 12.1 GB at this context; duplicate traces of the same session: 1cnp3mkb; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 14.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 14.6 1.4 12.1 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3x7pnaj7",
   "source_url": "https://omlx.ai/benchmarks/performance/3x7pnaj7",
   "tg_tokens": null,
   "tps": 1.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-3x7pnaj7-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 12.3 GB at this context; duplicate traces of the same session: 1cnp3mkb; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 24.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 24.4 1.3 12.3 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3x7pnaj7",
   "source_url": "https://omlx.ai/benchmarks/performance/3x7pnaj7",
   "tg_tokens": null,
   "tps": 1.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-13q89p81-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: 0mz1oww1, 2bug3i2i, i4rh0ep1, jvwq5dre, lbph4feh, yk5svwnr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 689.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 689.5 20.8 100.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/13q89p81",
   "source_url": "https://omlx.ai/benchmarks/performance/13q89p81",
   "tg_tokens": null,
   "tps": 20.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-13q89p81-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: 0mz1oww1, 2bug3i2i, i4rh0ep1, jvwq5dre, lbph4feh, yk5svwnr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 923.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 923.5 52.3 101.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/13q89p81",
   "source_url": "https://omlx.ai/benchmarks/performance/13q89p81",
   "tg_tokens": null,
   "tps": 52.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-13q89p81-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: 0mz1oww1, 2bug3i2i, i4rh0ep1, jvwq5dre, lbph4feh, yk5svwnr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1072.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,072 89.2 101.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/13q89p81",
   "source_url": "https://omlx.ai/benchmarks/performance/13q89p81",
   "tg_tokens": null,
   "tps": 89.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-13q89p81-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 102.5 GB at this context; duplicate traces of the same session: 0mz1oww1, 2bug3i2i, i4rh0ep1, jvwq5dre, lbph4feh, yk5svwnr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1055.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 1,055 44.3 102.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/13q89p81",
   "source_url": "https://omlx.ai/benchmarks/performance/13q89p81",
   "tg_tokens": null,
   "tps": 44.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-13q89p81-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 104.1 GB at this context; duplicate traces of the same session: 0mz1oww1, 2bug3i2i, i4rh0ep1, jvwq5dre, lbph4feh, yk5svwnr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1041.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 1,041 49.8 104.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/13q89p81",
   "source_url": "https://omlx.ai/benchmarks/performance/13q89p81",
   "tg_tokens": null,
   "tps": 49.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-13q89p81-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 108.6 GB at this context; duplicate traces of the same session: 0mz1oww1, 2bug3i2i, i4rh0ep1, jvwq5dre, lbph4feh, yk5svwnr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1023.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 128k 1,023 64.0 108.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/13q89p81",
   "source_url": "https://omlx.ai/benchmarks/performance/13q89p81",
   "tg_tokens": null,
   "tps": 64.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-13q89p81-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 109.8 GB at this context; duplicate traces of the same session: 0mz1oww1, 2bug3i2i, i4rh0ep1, jvwq5dre, lbph4feh, yk5svwnr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1010.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 195k 1,010 57.6 109.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/13q89p81",
   "source_url": "https://omlx.ai/benchmarks/performance/13q89p81",
   "tg_tokens": null,
   "tps": 57.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-46fbtzjp-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.7 GB at this context; duplicate traces of the same session: j1sy6zgu; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 912.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 4k 912.3 71.3 14.7 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/46fbtzjp",
   "source_url": "https://omlx.ai/benchmarks/performance/46fbtzjp",
   "tg_tokens": null,
   "tps": 71.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-46fbtzjp-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 15.2 GB at this context; duplicate traces of the same session: j1sy6zgu; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 893.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 16k 893.4 62.8 15.2 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/46fbtzjp",
   "source_url": "https://omlx.ai/benchmarks/performance/46fbtzjp",
   "tg_tokens": null,
   "tps": 62.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yry63rhf-ctx8192",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "mlx-community mixed build; oMLX community benchmark trace; peak memory 87.7 GB at this context; duplicate traces of the same session: 0suk0d7d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 642.8,
   "provenance": "community",
   "quant": "2.4bit",
   "quote": "mlx-community-DeepSeek-V4-Flash-0731-2.4bit-mixed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 8k 642.8 42.9 87.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yry63rhf",
   "source_url": "https://omlx.ai/benchmarks/performance/yry63rhf",
   "tg_tokens": null,
   "tps": 42.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yry63rhf-ctx32768",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "mlx-community mixed build; oMLX community benchmark trace; peak memory 88.6 GB at this context; duplicate traces of the same session: 0suk0d7d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 563.0,
   "provenance": "community",
   "quant": "2.4bit",
   "quote": "mlx-community-DeepSeek-V4-Flash-0731-2.4bit-mixed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 32k 563.0 48.5 88.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yry63rhf",
   "source_url": "https://omlx.ai/benchmarks/performance/yry63rhf",
   "tg_tokens": null,
   "tps": 48.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-efrzj0yc-ctx8192",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "mlx-community mixed build; oMLX community benchmark trace; peak memory 87.7 GB at this context; duplicate traces of the same session: cihzed7y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 642.6,
   "provenance": "community",
   "quant": "2.4bit",
   "quote": "mlx-community-DeepSeek-V4-Flash-0731-2.4bit-mixed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 8k 642.6 43.5 87.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/efrzj0yc",
   "source_url": "https://omlx.ai/benchmarks/performance/efrzj0yc",
   "tg_tokens": null,
   "tps": 43.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-efrzj0yc-ctx32768",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "mlx-community mixed build; oMLX community benchmark trace; peak memory 88.6 GB at this context; duplicate traces of the same session: cihzed7y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 563.4,
   "provenance": "community",
   "quant": "2.4bit",
   "quote": "mlx-community-DeepSeek-V4-Flash-0731-2.4bit-mixed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 32k 563.4 51.5 88.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/efrzj0yc",
   "source_url": "https://omlx.ai/benchmarks/performance/efrzj0yc",
   "tg_tokens": null,
   "tps": 51.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-gddop7tm-ctx8192",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "mlx-community mixed build; oMLX community benchmark trace; peak memory 89.3 GB at this context; duplicate traces of the same session: 7x9k5m1i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 800.2,
   "provenance": "community",
   "quant": "2.4bit",
   "quote": "mlx-community-DeepSeek-V4-Flash-0731-2.4bit-mixed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 8k 800.2 45.1 89.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gddop7tm",
   "source_url": "https://omlx.ai/benchmarks/performance/gddop7tm",
   "tg_tokens": null,
   "tps": 45.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gddop7tm-ctx32768",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "mlx-community mixed build; oMLX community benchmark trace; peak memory 92.8 GB at this context; duplicate traces of the same session: 7x9k5m1i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 668.2,
   "provenance": "community",
   "quant": "2.4bit",
   "quote": "mlx-community-DeepSeek-V4-Flash-0731-2.4bit-mixed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 32k 668.2 52.1 92.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gddop7tm",
   "source_url": "https://omlx.ai/benchmarks/performance/gddop7tm",
   "tg_tokens": null,
   "tps": 52.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-sb1bfj29-ctx8192",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "mlx-community mixed build; oMLX community benchmark trace; peak memory 88.1 GB at this context; duplicate traces of the same session: i10a6wnr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 736.8,
   "provenance": "community",
   "quant": "2.4bit",
   "quote": "mlx-community-DeepSeek-V4-Flash-0731-2.4bit-mixed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 8k 736.8 46.4 88.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sb1bfj29",
   "source_url": "https://omlx.ai/benchmarks/performance/sb1bfj29",
   "tg_tokens": null,
   "tps": 46.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-sb1bfj29-ctx32768",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "mlx-community mixed build; oMLX community benchmark trace; peak memory 90.0 GB at this context; duplicate traces of the same session: i10a6wnr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 629.2,
   "provenance": "community",
   "quant": "2.4bit",
   "quote": "mlx-community-DeepSeek-V4-Flash-0731-2.4bit-mixed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 32k 629.2 51.9 90.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sb1bfj29",
   "source_url": "https://omlx.ai/benchmarks/performance/sb1bfj29",
   "tg_tokens": null,
   "tps": 51.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-v6k1087a-ctx8192",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "mlx-community mixed build; oMLX community benchmark trace; peak memory 89.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 798.9,
   "provenance": "community",
   "quant": "2.4bit",
   "quote": "mlx-community-DeepSeek-V4-Flash-0731-2.4bit-mixed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 8k 798.9 44.7 89.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v6k1087a",
   "source_url": "https://omlx.ai/benchmarks/performance/v6k1087a",
   "tg_tokens": null,
   "tps": 44.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-v6k1087a-ctx32768",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "mlx-community mixed build; oMLX community benchmark trace; peak memory 92.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 668.8,
   "provenance": "community",
   "quant": "2.4bit",
   "quote": "mlx-community-DeepSeek-V4-Flash-0731-2.4bit-mixed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 32k 668.8 58.2 92.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v6k1087a",
   "source_url": "https://omlx.ai/benchmarks/performance/v6k1087a",
   "tg_tokens": null,
   "tps": 58.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8lf3aatg-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 94.0 GB at this context; duplicate traces of the same session: 5lzddgyt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 305.3,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 305.3 11.7 94.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8lf3aatg",
   "source_url": "https://omlx.ai/benchmarks/performance/8lf3aatg",
   "tg_tokens": null,
   "tps": 11.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8lf3aatg-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 94.1 GB at this context; duplicate traces of the same session: 5lzddgyt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 136.8,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 136.8 11.8 94.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8lf3aatg",
   "source_url": "https://omlx.ai/benchmarks/performance/8lf3aatg",
   "tg_tokens": null,
   "tps": 11.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kskhqw17-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 88.0 GB at this context; duplicate traces of the same session: 43am089z, 5y07n46e, kx6zq5q0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 89.6,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 89.6 8.8 88.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kskhqw17",
   "source_url": "https://omlx.ai/benchmarks/performance/kskhqw17",
   "tg_tokens": null,
   "tps": 8.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-kskhqw17-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 88.7 GB at this context; duplicate traces of the same session: 43am089z, 5y07n46e, kx6zq5q0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 279.2,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 4k 279.2 10.4 88.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kskhqw17",
   "source_url": "https://omlx.ai/benchmarks/performance/kskhqw17",
   "tg_tokens": null,
   "tps": 10.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kskhqw17-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 90.1 GB at this context; duplicate traces of the same session: 43am089z, 5y07n46e, kx6zq5q0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 199.2,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 199.2 10.8 90.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kskhqw17",
   "source_url": "https://omlx.ai/benchmarks/performance/kskhqw17",
   "tg_tokens": null,
   "tps": 10.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kskhqw17-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 90.3 GB at this context; duplicate traces of the same session: 43am089z, 5y07n46e, kx6zq5q0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 147.3,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 147.3 12.2 90.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kskhqw17",
   "source_url": "https://omlx.ai/benchmarks/performance/kskhqw17",
   "tg_tokens": null,
   "tps": 12.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vhlu2col-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 86.0 GB at this context; duplicate traces of the same session: 68xez47v, j7v7xqc6, yfqwdooa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 92.4,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 92.4 10.0 86.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vhlu2col",
   "source_url": "https://omlx.ai/benchmarks/performance/vhlu2col",
   "tg_tokens": null,
   "tps": 10.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-vhlu2col-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 86.7 GB at this context; duplicate traces of the same session: 68xez47v, j7v7xqc6, yfqwdooa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 288.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 4k 288.0 10.2 86.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vhlu2col",
   "source_url": "https://omlx.ai/benchmarks/performance/vhlu2col",
   "tg_tokens": null,
   "tps": 10.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vhlu2col-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 88.3 GB at this context; duplicate traces of the same session: 68xez47v, j7v7xqc6, yfqwdooa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 346.1,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 346.1 10.5 88.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vhlu2col",
   "source_url": "https://omlx.ai/benchmarks/performance/vhlu2col",
   "tg_tokens": null,
   "tps": 10.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vhlu2col-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 88.4 GB at this context; duplicate traces of the same session: 68xez47v, j7v7xqc6, yfqwdooa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 190.5,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 190.5 11.2 88.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vhlu2col",
   "source_url": "https://omlx.ai/benchmarks/performance/vhlu2col",
   "tg_tokens": null,
   "tps": 11.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-83miv4x1-ctx32768",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace, single run at 32k context; peak memory 404.7 GB; TTFT 81207 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 403.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 32k tokens 403.5 PP tok/s 36.1 TG tok/s 81207 TTFT (ms) 404.7 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/83miv4x1",
   "source_url": "https://omlx.ai/benchmarks/performance/83miv4x1",
   "tg_tokens": null,
   "tps": 36.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-xhrvep68-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: 4oem1pz6, 846cixy9, btur7b0v, hwm0bjil; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 935.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 935.0 110.3 13.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xhrvep68",
   "source_url": "https://omlx.ai/benchmarks/performance/xhrvep68",
   "tg_tokens": null,
   "tps": 110.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-xhrvep68-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: 4oem1pz6, 846cixy9, btur7b0v, hwm0bjil; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1036.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 1,036 106.8 13.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xhrvep68",
   "source_url": "https://omlx.ai/benchmarks/performance/xhrvep68",
   "tg_tokens": null,
   "tps": 106.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-xhrvep68-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.8 GB at this context; duplicate traces of the same session: 4oem1pz6, 846cixy9, btur7b0v, hwm0bjil; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1005.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 1,005 98.9 13.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xhrvep68",
   "source_url": "https://omlx.ai/benchmarks/performance/xhrvep68",
   "tg_tokens": null,
   "tps": 98.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-xhrvep68-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.9 GB at this context; duplicate traces of the same session: 4oem1pz6, 846cixy9, btur7b0v, hwm0bjil; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 945.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 945.2 87.2 13.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xhrvep68",
   "source_url": "https://omlx.ai/benchmarks/performance/xhrvep68",
   "tg_tokens": null,
   "tps": 87.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-xhrvep68-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 14.2 GB at this context; duplicate traces of the same session: 4oem1pz6, 846cixy9, btur7b0v, hwm0bjil; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 850.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 850.0 69.2 14.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xhrvep68",
   "source_url": "https://omlx.ai/benchmarks/performance/xhrvep68",
   "tg_tokens": null,
   "tps": 69.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-7ongtfk0-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: 9w6t0x8f; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3934.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 3,934 122.5 13.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ongtfk0",
   "source_url": "https://omlx.ai/benchmarks/performance/7ongtfk0",
   "tg_tokens": null,
   "tps": 122.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-7ongtfk0-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: 9w6t0x8f; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3474.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,474 123.0 13.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ongtfk0",
   "source_url": "https://omlx.ai/benchmarks/performance/7ongtfk0",
   "tg_tokens": null,
   "tps": 123.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-7ongtfk0-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.9 GB at this context; duplicate traces of the same session: 9w6t0x8f; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4488.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 4,488 104.5 13.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ongtfk0",
   "source_url": "https://omlx.ai/benchmarks/performance/7ongtfk0",
   "tg_tokens": null,
   "tps": 104.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-7ongtfk0-ctx65536",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 14.9 GB at this context; duplicate traces of the same session: 9w6t0x8f; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3448.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 3,448 68.2 14.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ongtfk0",
   "source_url": "https://omlx.ai/benchmarks/performance/7ongtfk0",
   "tg_tokens": null,
   "tps": 68.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-swjck84w-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 13.6 GB at this context; duplicate traces of the same session: epifc2j8, jk9ysmyt, uroib52m, wdirpne5, xezgx83z, ytwukh6k; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3172.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 3,172 119.2 13.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/swjck84w",
   "source_url": "https://omlx.ai/benchmarks/performance/swjck84w",
   "tg_tokens": null,
   "tps": 119.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-swjck84w-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 13.9 GB at this context; duplicate traces of the same session: epifc2j8, jk9ysmyt, uroib52m, wdirpne5, xezgx83z, ytwukh6k; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3400.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,400 91.6 13.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/swjck84w",
   "source_url": "https://omlx.ai/benchmarks/performance/swjck84w",
   "tg_tokens": null,
   "tps": 91.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-swjck84w-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 13.9 GB at this context; duplicate traces of the same session: epifc2j8, jk9ysmyt, uroib52m, wdirpne5, xezgx83z, ytwukh6k; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4066.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 4,066 84.5 13.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/swjck84w",
   "source_url": "https://omlx.ai/benchmarks/performance/swjck84w",
   "tg_tokens": null,
   "tps": 84.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-swjck84w-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 14.1 GB at this context; duplicate traces of the same session: epifc2j8, jk9ysmyt, uroib52m, wdirpne5, xezgx83z, ytwukh6k; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4412.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 4,412 100.4 14.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/swjck84w",
   "source_url": "https://omlx.ai/benchmarks/performance/swjck84w",
   "tg_tokens": null,
   "tps": 100.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-swjck84w-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: epifc2j8, jk9ysmyt, uroib52m, wdirpne5, xezgx83z, ytwukh6k; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4016.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 4,016 84.3 14.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/swjck84w",
   "source_url": "https://omlx.ai/benchmarks/performance/swjck84w",
   "tg_tokens": null,
   "tps": 84.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-swjck84w-ctx65536",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 15.0 GB at this context; duplicate traces of the same session: epifc2j8, jk9ysmyt, uroib52m, wdirpne5, xezgx83z, ytwukh6k; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3163.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 3,163 64.4 15.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/swjck84w",
   "source_url": "https://omlx.ai/benchmarks/performance/swjck84w",
   "tg_tokens": null,
   "tps": 64.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-swjck84w-ctx131072",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: epifc2j8, jk9ysmyt, uroib52m, wdirpne5, xezgx83z, ytwukh6k; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2301.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 128k 2,301 44.5 16.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/swjck84w",
   "source_url": "https://omlx.ai/benchmarks/performance/swjck84w",
   "tg_tokens": null,
   "tps": 44.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-mnc0nvd3-ctx1024",
   "model": "Laguna XS 2.1",
   "notes": "oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: 4oqlx9a7, 703lm700, d1nie6b0, non5xuc2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 555.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Laguna-XS-2.1-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 555.1 92.1 18.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mnc0nvd3",
   "source_url": "https://omlx.ai/benchmarks/performance/mnc0nvd3",
   "tg_tokens": null,
   "tps": 92.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-mnc0nvd3-ctx4096",
   "model": "Laguna XS 2.1",
   "notes": "oMLX community benchmark trace; peak memory 18.2 GB at this context; duplicate traces of the same session: 4oqlx9a7, 703lm700, d1nie6b0, non5xuc2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 616.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Laguna-XS-2.1-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 616.1 86.7 18.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mnc0nvd3",
   "source_url": "https://omlx.ai/benchmarks/performance/mnc0nvd3",
   "tg_tokens": null,
   "tps": 86.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-mnc0nvd3-ctx8192",
   "model": "Laguna XS 2.1",
   "notes": "oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 4oqlx9a7, 703lm700, d1nie6b0, non5xuc2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 582.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Laguna-XS-2.1-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 582.0 78.7 18.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mnc0nvd3",
   "source_url": "https://omlx.ai/benchmarks/performance/mnc0nvd3",
   "tg_tokens": null,
   "tps": 78.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-mnc0nvd3-ctx16384",
   "model": "Laguna XS 2.1",
   "notes": "oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: 4oqlx9a7, 703lm700, d1nie6b0, non5xuc2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 523.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Laguna-XS-2.1-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 523.9 68.4 18.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mnc0nvd3",
   "source_url": "https://omlx.ai/benchmarks/performance/mnc0nvd3",
   "tg_tokens": null,
   "tps": 68.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-mnc0nvd3-ctx32768",
   "model": "Laguna XS 2.1",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 4oqlx9a7, 703lm700, d1nie6b0, non5xuc2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 436.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Laguna-XS-2.1-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 436.6 54.5 19.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mnc0nvd3",
   "source_url": "https://omlx.ai/benchmarks/performance/mnc0nvd3",
   "tg_tokens": null,
   "tps": 54.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-imrjqhwm-ctx1024",
   "model": "Qwen2.5-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: o12m9h93, qje7p11b; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 256.3,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "Qwen2.5-7B-Instruct-AWQ on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 256.3 21.8 5.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/imrjqhwm",
   "source_url": "https://omlx.ai/benchmarks/performance/imrjqhwm",
   "tg_tokens": null,
   "tps": 21.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-imrjqhwm-ctx4096",
   "model": "Qwen2.5-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 6.1 GB at this context; duplicate traces of the same session: o12m9h93, qje7p11b; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 247.5,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "Qwen2.5-7B-Instruct-AWQ on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 247.5 20.9 6.1 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/imrjqhwm",
   "source_url": "https://omlx.ai/benchmarks/performance/imrjqhwm",
   "tg_tokens": null,
   "tps": 20.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-imrjqhwm-ctx8192",
   "model": "Qwen2.5-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: o12m9h93, qje7p11b; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 235.5,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "Qwen2.5-7B-Instruct-AWQ on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 8k 235.5 19.8 6.3 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/imrjqhwm",
   "source_url": "https://omlx.ai/benchmarks/performance/imrjqhwm",
   "tg_tokens": null,
   "tps": 19.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-trl2v9nz-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: 7kolqi18; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5493.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 5,493 166.9 20.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/trl2v9nz",
   "source_url": "https://omlx.ai/benchmarks/performance/trl2v9nz",
   "tg_tokens": null,
   "tps": 166.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-trl2v9nz-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 7kolqi18; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6573.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,573 146.7 21.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/trl2v9nz",
   "source_url": "https://omlx.ai/benchmarks/performance/trl2v9nz",
   "tg_tokens": null,
   "tps": 146.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-w3wlbww3-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 206 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4973.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 4,973 PP tok/s 150.1 TG tok/s 206 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w3wlbww3",
   "source_url": "https://omlx.ai/benchmarks/performance/w3wlbww3",
   "tg_tokens": null,
   "tps": 150.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-7xfm7f3o-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.3 GB; TTFT 228 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4486.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 4,486 PP tok/s 150.9 TG tok/s 228 TTFT (ms) 19.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7xfm7f3o",
   "source_url": "https://omlx.ai/benchmarks/performance/7xfm7f3o",
   "tg_tokens": null,
   "tps": 150.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8dlngr4h-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: q7jh9oer; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5485.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 5,485 157.3 20.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8dlngr4h",
   "source_url": "https://omlx.ai/benchmarks/performance/8dlngr4h",
   "tg_tokens": null,
   "tps": 157.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8dlngr4h-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: q7jh9oer; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6583.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,583 135.6 21.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8dlngr4h",
   "source_url": "https://omlx.ai/benchmarks/performance/8dlngr4h",
   "tg_tokens": null,
   "tps": 135.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5o5w328w-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 206 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4976.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 4,976 PP tok/s 150.8 TG tok/s 206 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5o5w328w",
   "source_url": "https://omlx.ai/benchmarks/performance/5o5w328w",
   "tg_tokens": null,
   "tps": 150.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-16ep4pol-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 222 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4613.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 4,613 PP tok/s 142.3 TG tok/s 222 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/16ep4pol",
   "source_url": "https://omlx.ai/benchmarks/performance/16ep4pol",
   "tg_tokens": null,
   "tps": 142.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-6jy1rjo1-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: ns35rgrx; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 489.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 489.0 18.3 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6jy1rjo1",
   "source_url": "https://omlx.ai/benchmarks/performance/6jy1rjo1",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-6jy1rjo1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: ns35rgrx; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 517.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 517.6 17.7 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6jy1rjo1",
   "source_url": "https://omlx.ai/benchmarks/performance/6jy1rjo1",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-rwlk019a-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: lxz3r8zq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 222.4,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 222.4 28.2 20.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rwlk019a",
   "source_url": "https://omlx.ai/benchmarks/performance/rwlk019a",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-rwlk019a-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: lxz3r8zq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 214.3,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 214.3 27.4 23.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rwlk019a",
   "source_url": "https://omlx.ai/benchmarks/performance/rwlk019a",
   "tg_tokens": null,
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hzzt7ppp-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: cjcmumug, ioc2y1r7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 773.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 773.8 31.4 16.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hzzt7ppp",
   "source_url": "https://omlx.ai/benchmarks/performance/hzzt7ppp",
   "tg_tokens": null,
   "tps": 31.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hzzt7ppp-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: cjcmumug, ioc2y1r7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 841.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 841.8 31.5 17.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hzzt7ppp",
   "source_url": "https://omlx.ai/benchmarks/performance/hzzt7ppp",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hzzt7ppp-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: cjcmumug, ioc2y1r7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 756.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 756.4 30.8 17.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hzzt7ppp",
   "source_url": "https://omlx.ai/benchmarks/performance/hzzt7ppp",
   "tg_tokens": null,
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4qk4gm3b-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: jn0lyjhp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 886.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 886.7 32.1 16.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4qk4gm3b",
   "source_url": "https://omlx.ai/benchmarks/performance/4qk4gm3b",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4qk4gm3b-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: jn0lyjhp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 910.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 910.1 31.8 17.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4qk4gm3b",
   "source_url": "https://omlx.ai/benchmarks/performance/4qk4gm3b",
   "tg_tokens": null,
   "tps": 31.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-6qwlr37h-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 185.1 GB; TTFT 24678 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1328.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 32k tokens 1,328 PP tok/s 51.8 TG tok/s 24678 TTFT (ms) 185.1 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6qwlr37h",
   "source_url": "https://omlx.ai/benchmarks/performance/6qwlr37h",
   "tg_tokens": null,
   "tps": 51.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-ggayjz8e-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: dslvzgzj; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1377.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 1,377 89.3 20.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ggayjz8e",
   "source_url": "https://omlx.ai/benchmarks/performance/ggayjz8e",
   "tg_tokens": null,
   "tps": 89.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-ggayjz8e-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: dslvzgzj; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1543.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 1,543 84.5 21.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ggayjz8e",
   "source_url": "https://omlx.ai/benchmarks/performance/ggayjz8e",
   "tg_tokens": null,
   "tps": 84.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-10vq5e9u-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 7nvqn3m2, a1xglicb, kl1utq0s; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2622.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,622 86.5 20.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/10vq5e9u",
   "source_url": "https://omlx.ai/benchmarks/performance/10vq5e9u",
   "tg_tokens": null,
   "tps": 86.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-10vq5e9u-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: 7nvqn3m2, a1xglicb, kl1utq0s; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3119.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,119 81.7 20.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/10vq5e9u",
   "source_url": "https://omlx.ai/benchmarks/performance/10vq5e9u",
   "tg_tokens": null,
   "tps": 81.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-10vq5e9u-ctx16384",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 7nvqn3m2, a1xglicb, kl1utq0s; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2825.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,825 63.5 20.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/10vq5e9u",
   "source_url": "https://omlx.ai/benchmarks/performance/10vq5e9u",
   "tg_tokens": null,
   "tps": 63.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-10vq5e9u-ctx65536",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 7nvqn3m2, a1xglicb, kl1utq0s; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1880.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 1,880 33.6 21.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/10vq5e9u",
   "source_url": "https://omlx.ai/benchmarks/performance/10vq5e9u",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-t6y32byy-ctx8192",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "mlx-community mixed build; oMLX community benchmark trace; peak memory 89.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 798.9,
   "provenance": "community",
   "quant": "2.4bit",
   "quote": "mlx-community-DeepSeek-V4-Flash-0731-2.4bit-mixed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 8k 798.9 44.7 89.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t6y32byy",
   "source_url": "https://omlx.ai/benchmarks/performance/t6y32byy",
   "tg_tokens": null,
   "tps": 44.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-t6y32byy-ctx32768",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "mlx-community mixed build; oMLX community benchmark trace; peak memory 92.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 668.8,
   "provenance": "community",
   "quant": "2.4bit",
   "quote": "mlx-community-DeepSeek-V4-Flash-0731-2.4bit-mixed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 32k 668.8 58.2 92.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t6y32byy",
   "source_url": "https://omlx.ai/benchmarks/performance/t6y32byy",
   "tg_tokens": null,
   "tps": 58.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dnnb7et5-ctx8192",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "mlx-community mixed build; oMLX community benchmark trace; peak memory 89.3 GB at this context; duplicate traces of the same session: bp6fujbc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 799.0,
   "provenance": "community",
   "quant": "2.4bit",
   "quote": "mlx-community-DeepSeek-V4-Flash-0731-2.4bit-mixed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 8k 799.0 46.0 89.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dnnb7et5",
   "source_url": "https://omlx.ai/benchmarks/performance/dnnb7et5",
   "tg_tokens": null,
   "tps": 46.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dnnb7et5-ctx32768",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "mlx-community mixed build; oMLX community benchmark trace; peak memory 92.8 GB at this context; duplicate traces of the same session: bp6fujbc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 668.8,
   "provenance": "community",
   "quant": "2.4bit",
   "quote": "mlx-community-DeepSeek-V4-Flash-0731-2.4bit-mixed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 32k 668.8 52.9 92.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dnnb7et5",
   "source_url": "https://omlx.ai/benchmarks/performance/dnnb7et5",
   "tg_tokens": null,
   "tps": 52.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hz8dobhw-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 86.0 GB at this context; duplicate traces of the same session: 3q07dnha, 5ltf28uj, nw0messk, zcowlr71; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 90.6,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 90.6 9.7 86.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hz8dobhw",
   "source_url": "https://omlx.ai/benchmarks/performance/hz8dobhw",
   "tg_tokens": null,
   "tps": 9.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hz8dobhw-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 86.7 GB at this context; duplicate traces of the same session: 3q07dnha, 5ltf28uj, nw0messk, zcowlr71; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 278.8,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 4k 278.8 10.4 86.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hz8dobhw",
   "source_url": "https://omlx.ai/benchmarks/performance/hz8dobhw",
   "tg_tokens": null,
   "tps": 10.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hz8dobhw-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 88.3 GB at this context; duplicate traces of the same session: 3q07dnha, 5ltf28uj, nw0messk, zcowlr71; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 247.4,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 247.4 9.6 88.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hz8dobhw",
   "source_url": "https://omlx.ai/benchmarks/performance/hz8dobhw",
   "tg_tokens": null,
   "tps": 9.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hz8dobhw-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 88.4 GB at this context; duplicate traces of the same session: 3q07dnha, 5ltf28uj, nw0messk, zcowlr71; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 210.5,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 210.5 12.0 88.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hz8dobhw",
   "source_url": "https://omlx.ai/benchmarks/performance/hz8dobhw",
   "tg_tokens": null,
   "tps": 12.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hz8dobhw-ctx32768",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 88.5 GB at this context; duplicate traces of the same session: 3q07dnha, 5ltf28uj, nw0messk, zcowlr71; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 160.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 32k 160.0 11.2 88.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hz8dobhw",
   "source_url": "https://omlx.ai/benchmarks/performance/hz8dobhw",
   "tg_tokens": null,
   "tps": 11.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ttl2jj75-ctx1024",
   "model": "Jev-Style-0.8B-Decision-v3",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 1.9 GB at this context; duplicate traces of the same session: ks5uvsky, p1l00922; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1187.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Jev-Style-0.8B-Decision-v3-MLX-oQ8e on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 8bit; 1k 1,187 34.0 1.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ttl2jj75",
   "source_url": "https://omlx.ai/benchmarks/performance/ttl2jj75",
   "tg_tokens": null,
   "tps": 34.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ttl2jj75-ctx4096",
   "model": "Jev-Style-0.8B-Decision-v3",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 2.1 GB at this context; duplicate traces of the same session: ks5uvsky, p1l00922; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2085.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Jev-Style-0.8B-Decision-v3-MLX-oQ8e on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 8bit; 4k 2,085 29.9 2.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ttl2jj75",
   "source_url": "https://omlx.ai/benchmarks/performance/ttl2jj75",
   "tg_tokens": null,
   "tps": 29.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ttl2jj75-ctx16384",
   "model": "Jev-Style-0.8B-Decision-v3",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 2.3 GB at this context; duplicate traces of the same session: ks5uvsky, p1l00922; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2438.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Jev-Style-0.8B-Decision-v3-MLX-oQ8e on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 8bit; 16k 2,438 39.8 2.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ttl2jj75",
   "source_url": "https://omlx.ai/benchmarks/performance/ttl2jj75",
   "tg_tokens": null,
   "tps": 39.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-80b96xmr-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: htruhrsh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3934.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 3,934 122.5 13.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/80b96xmr",
   "source_url": "https://omlx.ai/benchmarks/performance/80b96xmr",
   "tg_tokens": null,
   "tps": 122.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-80b96xmr-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: htruhrsh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3474.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,474 123.0 13.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/80b96xmr",
   "source_url": "https://omlx.ai/benchmarks/performance/80b96xmr",
   "tg_tokens": null,
   "tps": 123.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-80b96xmr-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.9 GB at this context; duplicate traces of the same session: htruhrsh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4488.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 4,488 104.5 13.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/80b96xmr",
   "source_url": "https://omlx.ai/benchmarks/performance/80b96xmr",
   "tg_tokens": null,
   "tps": 104.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-80b96xmr-ctx65536",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 14.9 GB at this context; duplicate traces of the same session: htruhrsh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3448.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 3,448 68.2 14.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/80b96xmr",
   "source_url": "https://omlx.ai/benchmarks/performance/80b96xmr",
   "tg_tokens": null,
   "tps": 68.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-tffhf3bn-ctx131072",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 25.0 GB; TTFT 292435 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 448.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiniCPM-V-4.7-35B-A3B-oQ4e on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 128k tokens 448.2 PP tok/s 27.6 TG tok/s 292435 TTFT (ms) 25.0 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tffhf3bn",
   "source_url": "https://omlx.ai/benchmarks/performance/tffhf3bn",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-awbryjlu-ctx32768",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 22.3 GB; TTFT 34525 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 949.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiniCPM-V-4.7-35B-A3B-oQ4e on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 32k tokens 949.1 PP tok/s 73.8 TG tok/s 34525 TTFT (ms) 22.3 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/awbryjlu",
   "source_url": "https://omlx.ai/benchmarks/performance/awbryjlu",
   "tg_tokens": null,
   "tps": 73.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-ud9rptyr-ctx1024",
   "model": "Qwen2.5-Coder-1.5B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 2.3 GB at this context; duplicate traces of the same session: b4kw1nzk, btzo6emw, ck4j6ind, svirnc9j, xx6rpyzg, ybn1lmyd; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 899.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen2.5-Coder-1.5B-oQ8e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 8bit; 1k 899.9 64.4 2.3 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ud9rptyr",
   "source_url": "https://omlx.ai/benchmarks/performance/ud9rptyr",
   "tg_tokens": null,
   "tps": 64.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-ud9rptyr-ctx4096",
   "model": "Qwen2.5-Coder-1.5B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 2.3 GB at this context; duplicate traces of the same session: b4kw1nzk, btzo6emw, ck4j6ind, svirnc9j, xx6rpyzg, ybn1lmyd; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 709.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen2.5-Coder-1.5B-oQ8e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 8bit; 4k 709.0 77.0 2.3 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ud9rptyr",
   "source_url": "https://omlx.ai/benchmarks/performance/ud9rptyr",
   "tg_tokens": null,
   "tps": 77.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-ud9rptyr-ctx8192",
   "model": "Qwen2.5-Coder-1.5B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 2.4 GB at this context; duplicate traces of the same session: b4kw1nzk, btzo6emw, ck4j6ind, svirnc9j, xx6rpyzg, ybn1lmyd; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 812.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen2.5-Coder-1.5B-oQ8e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 8bit; 8k 812.9 68.4 2.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ud9rptyr",
   "source_url": "https://omlx.ai/benchmarks/performance/ud9rptyr",
   "tg_tokens": null,
   "tps": 68.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-ud9rptyr-ctx16384",
   "model": "Qwen2.5-Coder-1.5B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 2.6 GB at this context; duplicate traces of the same session: b4kw1nzk, btzo6emw, ck4j6ind, svirnc9j, xx6rpyzg, ybn1lmyd; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 683.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen2.5-Coder-1.5B-oQ8e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 8bit; 16k 683.6 56.3 2.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ud9rptyr",
   "source_url": "https://omlx.ai/benchmarks/performance/ud9rptyr",
   "tg_tokens": null,
   "tps": 56.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-ud9rptyr-ctx32768",
   "model": "Qwen2.5-Coder-1.5B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 3.0 GB at this context; duplicate traces of the same session: b4kw1nzk, btzo6emw, ck4j6ind, svirnc9j, xx6rpyzg, ybn1lmyd; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 511.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen2.5-Coder-1.5B-oQ8e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 8bit; 32k 511.9 41.8 3.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ud9rptyr",
   "source_url": "https://omlx.ai/benchmarks/performance/ud9rptyr",
   "tg_tokens": null,
   "tps": 41.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-ud9rptyr-ctx65536",
   "model": "Qwen2.5-Coder-1.5B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 3.8 GB at this context; duplicate traces of the same session: b4kw1nzk, btzo6emw, ck4j6ind, svirnc9j, xx6rpyzg, ybn1lmyd; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 338.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen2.5-Coder-1.5B-oQ8e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 8bit; 64k 338.8 26.0 3.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ud9rptyr",
   "source_url": "https://omlx.ai/benchmarks/performance/ud9rptyr",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-ud9rptyr-ctx131072",
   "model": "Qwen2.5-Coder-1.5B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 5.5 GB at this context; duplicate traces of the same session: b4kw1nzk, btzo6emw, ck4j6ind, svirnc9j, xx6rpyzg, ybn1lmyd; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 206.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen2.5-Coder-1.5B-oQ8e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 8bit; 128k 206.8 14.6 5.5 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ud9rptyr",
   "source_url": "https://omlx.ai/benchmarks/performance/ud9rptyr",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-75f8h49i-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.1 GB at this context; duplicate traces of the same session: 22dolonh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1715.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 1,715 77.4 43.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/75f8h49i",
   "source_url": "https://omlx.ai/benchmarks/performance/75f8h49i",
   "tg_tokens": null,
   "tps": 77.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-75f8h49i-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.9 GB at this context; duplicate traces of the same session: 22dolonh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2283.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,283 72.2 43.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/75f8h49i",
   "source_url": "https://omlx.ai/benchmarks/performance/75f8h49i",
   "tg_tokens": null,
   "tps": 72.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3x2j02ej-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 638ueq0x, hi0zd77i, rc4kfpn8, z26yawc0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2812.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,812 89.8 20.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3x2j02ej",
   "source_url": "https://omlx.ai/benchmarks/performance/3x2j02ej",
   "tg_tokens": null,
   "tps": 89.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3x2j02ej-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 638ueq0x, hi0zd77i, rc4kfpn8, z26yawc0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2666.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 2,666 80.1 20.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3x2j02ej",
   "source_url": "https://omlx.ai/benchmarks/performance/3x2j02ej",
   "tg_tokens": null,
   "tps": 80.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3x2j02ej-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: 638ueq0x, hi0zd77i, rc4kfpn8, z26yawc0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1580.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 64k 1,580 54.4 23.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3x2j02ej",
   "source_url": "https://omlx.ai/benchmarks/performance/3x2j02ej",
   "tg_tokens": null,
   "tps": 54.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3x2j02ej-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: 638ueq0x, hi0zd77i, rc4kfpn8, z26yawc0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 952.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 128k 952.1 37.4 26.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3x2j02ej",
   "source_url": "https://omlx.ai/benchmarks/performance/3x2j02ej",
   "tg_tokens": null,
   "tps": 37.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-3x2j02ej-ctx199680",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 29.4 GB at this context; duplicate traces of the same session: 638ueq0x, hi0zd77i, rc4kfpn8, z26yawc0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 729.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 195k 729.5 30.9 29.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3x2j02ej",
   "source_url": "https://omlx.ai/benchmarks/performance/3x2j02ej",
   "tg_tokens": null,
   "tps": 30.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-migvrrk8-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: zts12do0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 734.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 734.5 109.7 35.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/migvrrk8",
   "source_url": "https://omlx.ai/benchmarks/performance/migvrrk8",
   "tg_tokens": null,
   "tps": 109.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-migvrrk8-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: zts12do0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2434.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 2,434 108.3 36.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/migvrrk8",
   "source_url": "https://omlx.ai/benchmarks/performance/migvrrk8",
   "tg_tokens": null,
   "tps": 108.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-1fmjosgb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 128.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-6bit-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 128.0 11.0 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1fmjosgb",
   "source_url": "https://omlx.ai/benchmarks/performance/1fmjosgb",
   "tg_tokens": null,
   "tps": 11.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-1fmjosgb-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 25.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 133.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-6bit-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 133.5 10.7 25.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1fmjosgb",
   "source_url": "https://omlx.ai/benchmarks/performance/1fmjosgb",
   "tg_tokens": null,
   "tps": 10.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-fsmdatlu-ctx1024",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "ET-O5.5-G6A-G4.7-DSV4-K3-RLOO MTP DFlash2 build; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: e5hy6dng; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 271.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-ET-O5.5-G6A-G4.7-DSV4-K3-RLOO-MTP-DFlash2-oQ4e on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 1k 271.4 56.5 19.3 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fsmdatlu",
   "source_url": "https://omlx.ai/benchmarks/performance/fsmdatlu",
   "tg_tokens": null,
   "tps": 56.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-fsmdatlu-ctx4096",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "ET-O5.5-G6A-G4.7-DSV4-K3-RLOO MTP DFlash2 build; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: e5hy6dng; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 258.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-ET-O5.5-G6A-G4.7-DSV4-K3-RLOO-MTP-DFlash2-oQ4e on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 4k 258.8 47.5 22.9 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fsmdatlu",
   "source_url": "https://omlx.ai/benchmarks/performance/fsmdatlu",
   "tg_tokens": null,
   "tps": 47.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-df88oh97-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 5ao0qabt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 253.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 253.8 16.6 28.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/df88oh97",
   "source_url": "https://omlx.ai/benchmarks/performance/df88oh97",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-df88oh97-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.7 GB at this context; duplicate traces of the same session: 5ao0qabt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 259.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 259.8 16.4 31.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/df88oh97",
   "source_url": "https://omlx.ai/benchmarks/performance/df88oh97",
   "tg_tokens": null,
   "tps": 16.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-6ev98obx-ctx1024",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 7ad2kk9m; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 111.4,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-5bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 5bit; 1k 111.4 9.7 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ev98obx",
   "source_url": "https://omlx.ai/benchmarks/performance/6ev98obx",
   "tg_tokens": null,
   "tps": 9.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-6ev98obx-ctx4096",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 29.8 GB at this context; duplicate traces of the same session: 7ad2kk9m; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 98.6,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-5bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 5bit; 4k 98.6 9.5 29.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ev98obx",
   "source_url": "https://omlx.ai/benchmarks/performance/6ev98obx",
   "tg_tokens": null,
   "tps": 9.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2zrowrio-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 34.8 GB; TTFT 802296 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 163.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k tokens 163.4 PP tok/s 29.3 TG tok/s 802296 TTFT (ms) 34.8 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2zrowrio",
   "source_url": "https://omlx.ai/benchmarks/performance/2zrowrio",
   "tg_tokens": null,
   "tps": 29.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-2gx34l81-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: emxd3cd4; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1637.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k 1,637 95.7 17.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2gx34l81",
   "source_url": "https://omlx.ai/benchmarks/performance/2gx34l81",
   "tg_tokens": null,
   "tps": 95.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-2gx34l81-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: emxd3cd4; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1165.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 128k 1,165 83.9 27.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2gx34l81",
   "source_url": "https://omlx.ai/benchmarks/performance/2gx34l81",
   "tg_tokens": null,
   "tps": 83.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-i6kxnepd-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: v9ekz2yz; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1358.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 1,358 40.1 22.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i6kxnepd",
   "source_url": "https://omlx.ai/benchmarks/performance/i6kxnepd",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-i6kxnepd-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: v9ekz2yz; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1745.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 8k 1,745 39.2 24.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i6kxnepd",
   "source_url": "https://omlx.ai/benchmarks/performance/i6kxnepd",
   "tg_tokens": null,
   "tps": 39.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-i6kxnepd-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: v9ekz2yz; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1613.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 32k 1,613 36.3 25.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i6kxnepd",
   "source_url": "https://omlx.ai/benchmarks/performance/i6kxnepd",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5ebjzc4r-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace, single run at 16k context; peak memory 103.4 GB; TTFT 16231 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1009.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k tokens 1,009 PP tok/s 32.4 TG tok/s 16231 TTFT (ms) 103.4 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ebjzc4r",
   "source_url": "https://omlx.ai/benchmarks/performance/5ebjzc4r",
   "tg_tokens": null,
   "tps": 32.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-aq2pwegi-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 100.0 GB; TTFT 15026 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1090.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k tokens 1,090 PP tok/s 69.3 TG tok/s 15026 TTFT (ms) 100.0 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aq2pwegi",
   "source_url": "https://omlx.ai/benchmarks/performance/aq2pwegi",
   "tg_tokens": null,
   "tps": 69.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-0cnpsmos-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 100.0 GB; TTFT 15046 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1089.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k tokens 1,089 PP tok/s 71.3 TG tok/s 15046 TTFT (ms) 100.0 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0cnpsmos",
   "source_url": "https://omlx.ai/benchmarks/performance/0cnpsmos",
   "tg_tokens": null,
   "tps": 71.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qjlvswr4-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 101.5 GB; TTFT 15503 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1057.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k tokens 1,057 PP tok/s 31.7 TG tok/s 15503 TTFT (ms) 101.5 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qjlvswr4",
   "source_url": "https://omlx.ai/benchmarks/performance/qjlvswr4",
   "tg_tokens": null,
   "tps": 31.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-69z9s57k-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.8 GB at this context; duplicate traces of the same session: 5k0xjghv, 62pszm6g, bl6ek2kz, nucpdl4q, z82e76ir; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1093.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,093 82.8 104.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/69z9s57k",
   "source_url": "https://omlx.ai/benchmarks/performance/69z9s57k",
   "tg_tokens": null,
   "tps": 82.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-69z9s57k-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 107.7 GB at this context; duplicate traces of the same session: 5k0xjghv, 62pszm6g, bl6ek2kz, nucpdl4q, z82e76ir; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1369.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 1,369 73.4 107.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/69z9s57k",
   "source_url": "https://omlx.ai/benchmarks/performance/69z9s57k",
   "tg_tokens": null,
   "tps": 73.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-69z9s57k-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 108.0 GB at this context; duplicate traces of the same session: 5k0xjghv, 62pszm6g, bl6ek2kz, nucpdl4q, z82e76ir; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1365.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 1,365 82.3 108.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/69z9s57k",
   "source_url": "https://omlx.ai/benchmarks/performance/69z9s57k",
   "tg_tokens": null,
   "tps": 82.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-69z9s57k-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 108.7 GB at this context; duplicate traces of the same session: 5k0xjghv, 62pszm6g, bl6ek2kz, nucpdl4q, z82e76ir; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1354.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 1,354 96.4 108.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/69z9s57k",
   "source_url": "https://omlx.ai/benchmarks/performance/69z9s57k",
   "tg_tokens": null,
   "tps": 96.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-69z9s57k-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.9 GB at this context; duplicate traces of the same session: 5k0xjghv, 62pszm6g, bl6ek2kz, nucpdl4q, z82e76ir; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1342.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 1,342 73.2 109.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/69z9s57k",
   "source_url": "https://omlx.ai/benchmarks/performance/69z9s57k",
   "tg_tokens": null,
   "tps": 73.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-69z9s57k-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 112.6 GB at this context; duplicate traces of the same session: 5k0xjghv, 62pszm6g, bl6ek2kz, nucpdl4q, z82e76ir; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1320.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 1,320 58.2 112.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/69z9s57k",
   "source_url": "https://omlx.ai/benchmarks/performance/69z9s57k",
   "tg_tokens": null,
   "tps": 58.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1uf3guug-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 121.3 GB; TTFT 15421 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1062.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 16k tokens 1,062 PP tok/s 66.6 TG tok/s 15421 TTFT (ms) 121.3 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1uf3guug",
   "source_url": "https://omlx.ai/benchmarks/performance/1uf3guug",
   "tg_tokens": null,
   "tps": 66.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-bxtorijv-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 183.1 GB; TTFT 1055 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 971.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k tokens 971.1 PP tok/s 77.1 TG tok/s 1055 TTFT (ms) 183.1 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bxtorijv",
   "source_url": "https://omlx.ai/benchmarks/performance/bxtorijv",
   "tg_tokens": null,
   "tps": 77.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-4xdi8dmk-ctx4096",
   "model": "Qwopus3.6-35B-A3B-Coder",
   "notes": "oMLX community benchmark trace; peak memory 28.2 GB at this context; duplicate traces of the same session: ywnl8i7c; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2299.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwopus3.6-35B-A3B-Coder-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 2,299 70.9 28.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4xdi8dmk",
   "source_url": "https://omlx.ai/benchmarks/performance/4xdi8dmk",
   "tg_tokens": null,
   "tps": 70.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-4xdi8dmk-ctx32768",
   "model": "Qwopus3.6-35B-A3B-Coder",
   "notes": "oMLX community benchmark trace; peak memory 28.7 GB at this context; duplicate traces of the same session: ywnl8i7c; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2112.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwopus3.6-35B-A3B-Coder-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 32k 2,112 45.3 28.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4xdi8dmk",
   "source_url": "https://omlx.ai/benchmarks/performance/4xdi8dmk",
   "tg_tokens": null,
   "tps": 45.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-82whsqkv-ctx4096",
   "model": "Qwopus3.6-35B-A3B-Coder",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.6 GB at this context; duplicate traces of the same session: sebt6kmp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2417.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwopus3.6-35B-A3B-Coder-oQ6-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 2,417 75.9 30.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/82whsqkv",
   "source_url": "https://omlx.ai/benchmarks/performance/82whsqkv",
   "tg_tokens": null,
   "tps": 75.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-82whsqkv-ctx32768",
   "model": "Qwopus3.6-35B-A3B-Coder",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.2 GB at this context; duplicate traces of the same session: sebt6kmp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2057.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwopus3.6-35B-A3B-Coder-oQ6-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 32k 2,057 62.4 31.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/82whsqkv",
   "source_url": "https://omlx.ai/benchmarks/performance/82whsqkv",
   "tg_tokens": null,
   "tps": 62.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-2dxo82wm-ctx1024",
   "model": "granite-4.2-8b",
   "notes": "MLX build; oMLX community benchmark trace, single run at 1k context; peak memory 5.3 GB; TTFT 1025 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 998.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "granite-4.2-8b-MLX-4bit on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 1k tokens 998.7 PP tok/s 28.7 TG tok/s 1025 TTFT (ms) 5.3 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2dxo82wm",
   "source_url": "https://omlx.ai/benchmarks/performance/2dxo82wm",
   "tg_tokens": null,
   "tps": 28.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-ai17yp03-ctx1024",
   "model": "granite-4.2-8b",
   "notes": "MLX build; oMLX community benchmark trace, single run at 1k context; peak memory 5.3 GB; TTFT 1078 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 949.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "granite-4.2-8b-MLX-4bit on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 1k tokens 949.7 PP tok/s 28.6 TG tok/s 1078 TTFT (ms) 5.3 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ai17yp03",
   "source_url": "https://omlx.ai/benchmarks/performance/ai17yp03",
   "tg_tokens": null,
   "tps": 28.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-hrxn3pei-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "spacecomx fp16 MTP build; oMLX community benchmark trace; peak memory 36.3 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 512.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "spacecomx--Qwen3.6-35B-A3B-oQ6e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 64k 512.0 40.9 36.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hrxn3pei",
   "source_url": "https://omlx.ai/benchmarks/performance/hrxn3pei",
   "tg_tokens": null,
   "tps": 40.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-hrxn3pei-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "spacecomx fp16 MTP build; oMLX community benchmark trace; peak memory 37.7 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 333.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "spacecomx--Qwen3.6-35B-A3B-oQ6e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 128k 333.0 35.3 37.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hrxn3pei",
   "source_url": "https://omlx.ai/benchmarks/performance/hrxn3pei",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-t8yz7v0x-ctx4096",
   "model": "Qwen3.6-27B",
   "notes": "unsloth MLX build; oMLX community benchmark trace; peak memory 34.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 852.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "unsloth--Qwen3.6-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 852.9 15.6 34.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t8yz7v0x",
   "source_url": "https://omlx.ai/benchmarks/performance/t8yz7v0x",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-t8yz7v0x-ctx16384",
   "model": "Qwen3.6-27B",
   "notes": "unsloth MLX build; oMLX community benchmark trace; peak memory 35.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 888.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "unsloth--Qwen3.6-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 16k 888.3 15.2 35.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t8yz7v0x",
   "source_url": "https://omlx.ai/benchmarks/performance/t8yz7v0x",
   "tg_tokens": null,
   "tps": 15.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tb4rf48o-ctx1024",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 22.3 GB; TTFT 1671 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 612.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "AREX-2-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k tokens 612.8 PP tok/s 23.9 TG tok/s 1671 TTFT (ms) 22.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tb4rf48o",
   "source_url": "https://omlx.ai/benchmarks/performance/tb4rf48o",
   "tg_tokens": null,
   "tps": 23.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-tzwle271-ctx1024",
   "model": "AREX-Turbo",
   "notes": "oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: yihxuhcp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1354.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "AREX-Turbo-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 1,354 116.1 3.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tzwle271",
   "source_url": "https://omlx.ai/benchmarks/performance/tzwle271",
   "tg_tokens": null,
   "tps": 116.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-tzwle271-ctx4096",
   "model": "AREX-Turbo",
   "notes": "oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: yihxuhcp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1460.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "AREX-Turbo-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,460 112.0 3.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tzwle271",
   "source_url": "https://omlx.ai/benchmarks/performance/tzwle271",
   "tg_tokens": null,
   "tps": 112.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-zdn96eyf-ctx1024",
   "model": "AREX-Turbo",
   "notes": "oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: jsvg1srv, v67iv27o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1394.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "AREX-Turbo-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 1,394 110.7 3.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zdn96eyf",
   "source_url": "https://omlx.ai/benchmarks/performance/zdn96eyf",
   "tg_tokens": null,
   "tps": 110.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-zdn96eyf-ctx4096",
   "model": "AREX-Turbo",
   "notes": "oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: jsvg1srv, v67iv27o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1461.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "AREX-Turbo-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,461 110.4 3.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zdn96eyf",
   "source_url": "https://omlx.ai/benchmarks/performance/zdn96eyf",
   "tg_tokens": null,
   "tps": 110.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-zdn96eyf-ctx16384",
   "model": "AREX-Turbo",
   "notes": "oMLX community benchmark trace; peak memory 4.4 GB at this context; duplicate traces of the same session: jsvg1srv, v67iv27o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1307.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "AREX-Turbo-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 1,307 97.0 4.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zdn96eyf",
   "source_url": "https://omlx.ai/benchmarks/performance/zdn96eyf",
   "tg_tokens": null,
   "tps": 97.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9g8j63c2-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: maiz8hfd, o9jgzllh, q191qbup, u7ta0o6w; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 83.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 83.0 13.0 101.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9g8j63c2",
   "source_url": "https://omlx.ai/benchmarks/performance/9g8j63c2",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9g8j63c2-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.2 GB at this context; duplicate traces of the same session: maiz8hfd, o9jgzllh, q191qbup, u7ta0o6w; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 264.7,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 264.7 13.0 102.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9g8j63c2",
   "source_url": "https://omlx.ai/benchmarks/performance/9g8j63c2",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9g8j63c2-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.3 GB at this context; duplicate traces of the same session: maiz8hfd, o9jgzllh, q191qbup, u7ta0o6w; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 373.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 8k 373.0 13.2 103.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9g8j63c2",
   "source_url": "https://omlx.ai/benchmarks/performance/9g8j63c2",
   "tg_tokens": null,
   "tps": 13.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9g8j63c2-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.3 GB at this context; duplicate traces of the same session: maiz8hfd, o9jgzllh, q191qbup, u7ta0o6w; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 379.8,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 16k 379.8 15.0 103.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9g8j63c2",
   "source_url": "https://omlx.ai/benchmarks/performance/9g8j63c2",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9g8j63c2-ctx32768",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.4 GB at this context; duplicate traces of the same session: maiz8hfd, o9jgzllh, q191qbup, u7ta0o6w; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 369.8,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 32k 369.8 17.4 103.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9g8j63c2",
   "source_url": "https://omlx.ai/benchmarks/performance/9g8j63c2",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-tdao4jvg-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.4 GB at this context; duplicate traces of the same session: 4glhatd5, huihd0rh, kriiqyuu, stvmuosk, zjmcusaw; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 83.9,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 83.9 15.0 101.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tdao4jvg",
   "source_url": "https://omlx.ai/benchmarks/performance/tdao4jvg",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-tdao4jvg-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.2 GB at this context; duplicate traces of the same session: 4glhatd5, huihd0rh, kriiqyuu, stvmuosk, zjmcusaw; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 270.6,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 270.6 14.7 102.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tdao4jvg",
   "source_url": "https://omlx.ai/benchmarks/performance/tdao4jvg",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-tdao4jvg-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.3 GB at this context; duplicate traces of the same session: 4glhatd5, huihd0rh, kriiqyuu, stvmuosk, zjmcusaw; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 383.8,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 8k 383.8 14.7 103.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tdao4jvg",
   "source_url": "https://omlx.ai/benchmarks/performance/tdao4jvg",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-tdao4jvg-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.3 GB at this context; duplicate traces of the same session: 4glhatd5, huihd0rh, kriiqyuu, stvmuosk, zjmcusaw; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 388.1,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 16k 388.1 17.6 103.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tdao4jvg",
   "source_url": "https://omlx.ai/benchmarks/performance/tdao4jvg",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-tdao4jvg-ctx32768",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.4 GB at this context; duplicate traces of the same session: 4glhatd5, huihd0rh, kriiqyuu, stvmuosk, zjmcusaw; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 382.3,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 32k 382.3 16.6 103.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tdao4jvg",
   "source_url": "https://omlx.ai/benchmarks/performance/tdao4jvg",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-tdao4jvg-ctx65536",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.5 GB at this context; duplicate traces of the same session: 4glhatd5, huihd0rh, kriiqyuu, stvmuosk, zjmcusaw; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 376.4,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 64k 376.4 18.4 103.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tdao4jvg",
   "source_url": "https://omlx.ai/benchmarks/performance/tdao4jvg",
   "tg_tokens": null,
   "tps": 18.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-q01wus5n-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: 6pgijzt4, 7pusy8kt, 906rtuma, pwpo9sr2, stxawc8r, sy8aor6w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2236.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 2,236 161.1 13.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q01wus5n",
   "source_url": "https://omlx.ai/benchmarks/performance/q01wus5n",
   "tg_tokens": null,
   "tps": 161.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-q01wus5n-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: 6pgijzt4, 7pusy8kt, 906rtuma, pwpo9sr2, stxawc8r, sy8aor6w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2504.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 2,504 161.4 13.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q01wus5n",
   "source_url": "https://omlx.ai/benchmarks/performance/q01wus5n",
   "tg_tokens": null,
   "tps": 161.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-q01wus5n-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.8 GB at this context; duplicate traces of the same session: 6pgijzt4, 7pusy8kt, 906rtuma, pwpo9sr2, stxawc8r, sy8aor6w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2361.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 2,361 147.9 13.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q01wus5n",
   "source_url": "https://omlx.ai/benchmarks/performance/q01wus5n",
   "tg_tokens": null,
   "tps": 147.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-q01wus5n-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.9 GB at this context; duplicate traces of the same session: 6pgijzt4, 7pusy8kt, 906rtuma, pwpo9sr2, stxawc8r, sy8aor6w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1988.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,988 146.5 13.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q01wus5n",
   "source_url": "https://omlx.ai/benchmarks/performance/q01wus5n",
   "tg_tokens": null,
   "tps": 146.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-q01wus5n-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 14.2 GB at this context; duplicate traces of the same session: 6pgijzt4, 7pusy8kt, 906rtuma, pwpo9sr2, stxawc8r, sy8aor6w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1700.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 1,700 125.2 14.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q01wus5n",
   "source_url": "https://omlx.ai/benchmarks/performance/q01wus5n",
   "tg_tokens": null,
   "tps": 125.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-q01wus5n-ctx65536",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 14.9 GB at this context; duplicate traces of the same session: 6pgijzt4, 7pusy8kt, 906rtuma, pwpo9sr2, stxawc8r, sy8aor6w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1314.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 1,314 96.1 14.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q01wus5n",
   "source_url": "https://omlx.ai/benchmarks/performance/q01wus5n",
   "tg_tokens": null,
   "tps": 96.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-q01wus5n-ctx131072",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 6pgijzt4, 7pusy8kt, 906rtuma, pwpo9sr2, stxawc8r, sy8aor6w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 896.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 896.4 68.9 17.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q01wus5n",
   "source_url": "https://omlx.ai/benchmarks/performance/q01wus5n",
   "tg_tokens": null,
   "tps": 68.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-49uo0sp2-ctx131072",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 25.0 GB; TTFT 336787 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 389.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiniCPM-V-4.7-35B-A3B-oQ4e on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 128k tokens 389.2 PP tok/s 22.1 TG tok/s 336787 TTFT (ms) 25.0 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/49uo0sp2",
   "source_url": "https://omlx.ai/benchmarks/performance/49uo0sp2",
   "tg_tokens": null,
   "tps": 22.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-t00y1wn3-ctx65536",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 23.2 GB; TTFT 96166 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 681.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiniCPM-V-4.7-35B-A3B-oQ4e on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 64k tokens 681.5 PP tok/s 58.5 TG tok/s 96166 TTFT (ms) 23.2 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t00y1wn3",
   "source_url": "https://omlx.ai/benchmarks/performance/t00y1wn3",
   "tg_tokens": null,
   "tps": 58.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-m29by6lx-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: g3frzjxw, k6mypp10; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4697.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,697 146.4 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m29by6lx",
   "source_url": "https://omlx.ai/benchmarks/performance/m29by6lx",
   "tg_tokens": null,
   "tps": 146.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-m29by6lx-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: g3frzjxw, k6mypp10; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3130.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 3,130 144.7 20.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m29by6lx",
   "source_url": "https://omlx.ai/benchmarks/performance/m29by6lx",
   "tg_tokens": null,
   "tps": 144.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-m29by6lx-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: g3frzjxw, k6mypp10; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4898.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 4,898 119.9 21.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m29by6lx",
   "source_url": "https://omlx.ai/benchmarks/performance/m29by6lx",
   "tg_tokens": null,
   "tps": 119.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0gcfpit1-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: pd1c036z; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2374.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,374 83.1 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0gcfpit1",
   "source_url": "https://omlx.ai/benchmarks/performance/0gcfpit1",
   "tg_tokens": null,
   "tps": 83.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0gcfpit1-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: pd1c036z; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2207.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 2,207 71.2 20.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0gcfpit1",
   "source_url": "https://omlx.ai/benchmarks/performance/0gcfpit1",
   "tg_tokens": null,
   "tps": 71.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-rk4kn42p-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 20.1 GB; TTFT 1626 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2519.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k tokens 2,519 PP tok/s 96.9 TG tok/s 1626 TTFT (ms) 20.1 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rk4kn42p",
   "source_url": "https://omlx.ai/benchmarks/performance/rk4kn42p",
   "tg_tokens": null,
   "tps": 96.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-np3drbly-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 5778mf2y; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2352.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,352 82.7 20.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/np3drbly",
   "source_url": "https://omlx.ai/benchmarks/performance/np3drbly",
   "tg_tokens": null,
   "tps": 82.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-np3drbly-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 5778mf2y; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2567.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 2,567 84.5 20.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/np3drbly",
   "source_url": "https://omlx.ai/benchmarks/performance/np3drbly",
   "tg_tokens": null,
   "tps": 84.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-jykinpui-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 22.4 GB; TTFT 6738 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2432.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k tokens 2,432 PP tok/s 78.5 TG tok/s 6738 TTFT (ms) 22.4 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jykinpui",
   "source_url": "https://omlx.ai/benchmarks/performance/jykinpui",
   "tg_tokens": null,
   "tps": 78.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-eqx9dtom-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 5ras4ut3; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2339.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,339 70.8 21.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eqx9dtom",
   "source_url": "https://omlx.ai/benchmarks/performance/eqx9dtom",
   "tg_tokens": null,
   "tps": 70.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-eqx9dtom-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: 5ras4ut3; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2440.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 2,440 75.6 22.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eqx9dtom",
   "source_url": "https://omlx.ai/benchmarks/performance/eqx9dtom",
   "tg_tokens": null,
   "tps": 75.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ri3v9d8b-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: bqv7yock; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2469.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,469 81.1 20.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ri3v9d8b",
   "source_url": "https://omlx.ai/benchmarks/performance/ri3v9d8b",
   "tg_tokens": null,
   "tps": 81.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ri3v9d8b-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: bqv7yock; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2438.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 2,438 72.0 20.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ri3v9d8b",
   "source_url": "https://omlx.ai/benchmarks/performance/ri3v9d8b",
   "tg_tokens": null,
   "tps": 72.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-s4nk0emd-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 20.1 GB; TTFT 1860 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2202.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k tokens 2,202 PP tok/s 89.6 TG tok/s 1860 TTFT (ms) 20.1 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s4nk0emd",
   "source_url": "https://omlx.ai/benchmarks/performance/s4nk0emd",
   "tg_tokens": null,
   "tps": 89.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-rj6nysua-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 22.4 GB; TTFT 7021 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2334.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k tokens 2,334 PP tok/s 91.1 TG tok/s 7021 TTFT (ms) 22.4 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rj6nysua",
   "source_url": "https://omlx.ai/benchmarks/performance/rj6nysua",
   "tg_tokens": null,
   "tps": 91.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-owhsx59j-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 89gljuh3; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2446.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,446 78.1 21.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/owhsx59j",
   "source_url": "https://omlx.ai/benchmarks/performance/owhsx59j",
   "tg_tokens": null,
   "tps": 78.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-owhsx59j-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: 89gljuh3; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2451.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 2,451 66.7 22.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/owhsx59j",
   "source_url": "https://omlx.ai/benchmarks/performance/owhsx59j",
   "tg_tokens": null,
   "tps": 66.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1bhwqf8i-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: i1zx3r3x; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2099.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,099 81.0 20.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1bhwqf8i",
   "source_url": "https://omlx.ai/benchmarks/performance/1bhwqf8i",
   "tg_tokens": null,
   "tps": 81.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1bhwqf8i-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: i1zx3r3x; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2108.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 2,108 71.9 20.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1bhwqf8i",
   "source_url": "https://omlx.ai/benchmarks/performance/1bhwqf8i",
   "tg_tokens": null,
   "tps": 71.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-dj1qgkbx-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2343.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,343 73.5 21.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dj1qgkbx",
   "source_url": "https://omlx.ai/benchmarks/performance/dj1qgkbx",
   "tg_tokens": null,
   "tps": 73.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-dj1qgkbx-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2520.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 2,520 82.1 22.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dj1qgkbx",
   "source_url": "https://omlx.ai/benchmarks/performance/dj1qgkbx",
   "tg_tokens": null,
   "tps": 82.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-c5mjy1nx-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2477.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,477 71.0 21.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c5mjy1nx",
   "source_url": "https://omlx.ai/benchmarks/performance/c5mjy1nx",
   "tg_tokens": null,
   "tps": 71.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-c5mjy1nx-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2445.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 2,445 69.1 22.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c5mjy1nx",
   "source_url": "https://omlx.ai/benchmarks/performance/c5mjy1nx",
   "tg_tokens": null,
   "tps": 69.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-3364dh96-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 128.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-6bit-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 128.0 11.0 22.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3364dh96",
   "source_url": "https://omlx.ai/benchmarks/performance/3364dh96",
   "tg_tokens": null,
   "tps": 11.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-3364dh96-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 25.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 133.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-6bit-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 133.5 10.7 25.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3364dh96",
   "source_url": "https://omlx.ai/benchmarks/performance/3364dh96",
   "tg_tokens": null,
   "tps": 10.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-q2uj653f-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 56vtj23a; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 130.4 12.6 20.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q2uj653f",
   "source_url": "https://omlx.ai/benchmarks/performance/q2uj653f",
   "tg_tokens": null,
   "tps": 12.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-q2uj653f-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: 56vtj23a; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 131.4 12.4 23.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q2uj653f",
   "source_url": "https://omlx.ai/benchmarks/performance/q2uj653f",
   "tg_tokens": null,
   "tps": 12.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-xu1pumpo-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 32.4 GB at this context; duplicate traces of the same session: 80095zx6; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 110.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k 110.3 11.4 32.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xu1pumpo",
   "source_url": "https://omlx.ai/benchmarks/performance/xu1pumpo",
   "tg_tokens": null,
   "tps": 11.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-xu1pumpo-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 34.2 GB at this context; duplicate traces of the same session: 80095zx6; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 100.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 16k 100.6 9.9 34.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xu1pumpo",
   "source_url": "https://omlx.ai/benchmarks/performance/xu1pumpo",
   "tg_tokens": null,
   "tps": 9.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-k6mnam54-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1358.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 1,358 40.1 22.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k6mnam54",
   "source_url": "https://omlx.ai/benchmarks/performance/k6mnam54",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-k6mnam54-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1745.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 8k 1,745 39.2 24.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k6mnam54",
   "source_url": "https://omlx.ai/benchmarks/performance/k6mnam54",
   "tg_tokens": null,
   "tps": 39.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-k6mnam54-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1613.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 32k 1,613 36.3 25.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k6mnam54",
   "source_url": "https://omlx.ai/benchmarks/performance/k6mnam54",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bkdzv97t-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace, single run at 16k context; peak memory 108.8 GB; TTFT 15629 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1048.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ4-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k tokens 1,048 PP tok/s 23.8 TG tok/s 15629 TTFT (ms) 108.8 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bkdzv97t",
   "source_url": "https://omlx.ai/benchmarks/performance/bkdzv97t",
   "tg_tokens": null,
   "tps": 23.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-k696lsld-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 180.2 GB; TTFT 1002 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1022.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k tokens 1,022 PP tok/s 59.1 TG tok/s 1002 TTFT (ms) 180.2 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k696lsld",
   "source_url": "https://omlx.ai/benchmarks/performance/k696lsld",
   "tg_tokens": null,
   "tps": 59.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-825g681q-ctx1024",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "mlx-community mixed build; oMLX community benchmark trace; peak memory 87.7 GB at this context; duplicate traces of the same session: b0naqzru, z21s7750; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 683.9,
   "provenance": "community",
   "quant": "2.4bit",
   "quote": "mlx-community-DeepSeek-V4-Flash-0731-2.4bit-mixed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 1k 683.9 27.5 87.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/825g681q",
   "source_url": "https://omlx.ai/benchmarks/performance/825g681q",
   "tg_tokens": null,
   "tps": 27.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-825g681q-ctx8192",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "mlx-community mixed build; oMLX community benchmark trace; peak memory 87.7 GB at this context; duplicate traces of the same session: b0naqzru, z21s7750; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 645.5,
   "provenance": "community",
   "quant": "2.4bit",
   "quote": "mlx-community-DeepSeek-V4-Flash-0731-2.4bit-mixed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 8k 645.5 42.0 87.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/825g681q",
   "source_url": "https://omlx.ai/benchmarks/performance/825g681q",
   "tg_tokens": null,
   "tps": 42.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-825g681q-ctx32768",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "mlx-community mixed build; oMLX community benchmark trace; peak memory 88.6 GB at this context; duplicate traces of the same session: b0naqzru, z21s7750; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 563.4,
   "provenance": "community",
   "quant": "2.4bit",
   "quote": "mlx-community-DeepSeek-V4-Flash-0731-2.4bit-mixed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 32k 563.4 46.8 88.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/825g681q",
   "source_url": "https://omlx.ai/benchmarks/performance/825g681q",
   "tg_tokens": null,
   "tps": 46.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-sut1kfbh-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "spacecomx fp16 MTP build; oMLX community benchmark trace; peak memory 36.3 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 512.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "spacecomx--Qwen3.6-35B-A3B-oQ6e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 64k 512.0 40.9 36.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sut1kfbh",
   "source_url": "https://omlx.ai/benchmarks/performance/sut1kfbh",
   "tg_tokens": null,
   "tps": 40.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-sut1kfbh-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "spacecomx fp16 MTP build; oMLX community benchmark trace; peak memory 37.7 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 333.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "spacecomx--Qwen3.6-35B-A3B-oQ6e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 128k 333.0 35.3 37.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sut1kfbh",
   "source_url": "https://omlx.ai/benchmarks/performance/sut1kfbh",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-7br4atew-ctx4096",
   "model": "Qwen3.6-27B",
   "notes": "unsloth MLX build; oMLX community benchmark trace; peak memory 34.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 852.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "unsloth--Qwen3.6-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 852.9 15.6 34.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7br4atew",
   "source_url": "https://omlx.ai/benchmarks/performance/7br4atew",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-7br4atew-ctx16384",
   "model": "Qwen3.6-27B",
   "notes": "unsloth MLX build; oMLX community benchmark trace; peak memory 35.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 888.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "unsloth--Qwen3.6-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 16k 888.3 15.2 35.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7br4atew",
   "source_url": "https://omlx.ai/benchmarks/performance/7br4atew",
   "tg_tokens": null,
   "tps": 15.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-4irlxr04-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 105.1 GB at this context; duplicate traces of the same session: 0y5f4bl2, 1yromg4c, 7brlzg4n, msq7pnip, x1hmj7ba; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 59.4,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 59.4 11.8 105.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4irlxr04",
   "source_url": "https://omlx.ai/benchmarks/performance/4irlxr04",
   "tg_tokens": null,
   "tps": 11.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4irlxr04-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 106.1 GB at this context; duplicate traces of the same session: 0y5f4bl2, 1yromg4c, 7brlzg4n, msq7pnip, x1hmj7ba; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 267.6,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 267.6 16.5 106.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4irlxr04",
   "source_url": "https://omlx.ai/benchmarks/performance/4irlxr04",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4irlxr04-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 107.2 GB at this context; duplicate traces of the same session: 0y5f4bl2, 1yromg4c, 7brlzg4n, msq7pnip, x1hmj7ba; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 368.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 8k 368.0 14.2 107.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4irlxr04",
   "source_url": "https://omlx.ai/benchmarks/performance/4irlxr04",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4irlxr04-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 107.2 GB at this context; duplicate traces of the same session: 0y5f4bl2, 1yromg4c, 7brlzg4n, msq7pnip, x1hmj7ba; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 381.6,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 16k 381.6 17.6 107.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4irlxr04",
   "source_url": "https://omlx.ai/benchmarks/performance/4irlxr04",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4irlxr04-ctx32768",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 107.3 GB at this context; duplicate traces of the same session: 0y5f4bl2, 1yromg4c, 7brlzg4n, msq7pnip, x1hmj7ba; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 379.2,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 32k 379.2 16.0 107.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4irlxr04",
   "source_url": "https://omlx.ai/benchmarks/performance/4irlxr04",
   "tg_tokens": null,
   "tps": 16.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4irlxr04-ctx65536",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 107.3 GB at this context; duplicate traces of the same session: 0y5f4bl2, 1yromg4c, 7brlzg4n, msq7pnip, x1hmj7ba; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 376.3,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 64k 376.3 19.3 107.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4irlxr04",
   "source_url": "https://omlx.ai/benchmarks/performance/4irlxr04",
   "tg_tokens": null,
   "tps": 19.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-6pm728hi-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 1k context; peak memory 13.4 GB; TTFT 1285 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 797.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 1k tokens 797.0 PP tok/s 78.0 TG tok/s 1285 TTFT (ms) 13.4 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6pm728hi",
   "source_url": "https://omlx.ai/benchmarks/performance/6pm728hi",
   "tg_tokens": null,
   "tps": 78.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-424mi3cl-ctx65536",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 23.2 GB; TTFT 110032 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 595.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiniCPM-V-4.7-35B-A3B-oQ4e on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 64k tokens 595.6 PP tok/s 39.5 TG tok/s 110032 TTFT (ms) 23.2 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/424mi3cl",
   "source_url": "https://omlx.ai/benchmarks/performance/424mi3cl",
   "tg_tokens": null,
   "tps": 39.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-kzrhz4i7-ctx4096",
   "model": "Qwen3 30B A3B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 4k context; peak memory 16.1 GB; TTFT 9606 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 426.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-30B-A3B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k tokens 426.4 PP tok/s 31.3 TG tok/s 9606 TTFT (ms) 16.1 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kzrhz4i7",
   "source_url": "https://omlx.ai/benchmarks/performance/kzrhz4i7",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-8xl2lycg-ctx1024",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 1.4 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2114.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-0.8B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 2,114 181.3 1.4 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8xl2lycg",
   "source_url": "https://omlx.ai/benchmarks/performance/8xl2lycg",
   "tg_tokens": null,
   "tps": 181.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-8xl2lycg-ctx4096",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 1.6 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2161.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-0.8B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 2,161 166.3 1.6 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8xl2lycg",
   "source_url": "https://omlx.ai/benchmarks/performance/8xl2lycg",
   "tg_tokens": null,
   "tps": 166.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-tx14b26v-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 5.9 GB; TTFT 2818 ms; duplicate traces of the same session: none; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 363.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 1k tokens 363.3 PP tok/s 28.2 TG tok/s 2818 TTFT (ms) 5.9 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tx14b26v",
   "source_url": "https://omlx.ai/benchmarks/performance/tx14b26v",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-rjrbtw6f-ctx1024",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 16.0 GB; TTFT 10012 ms; duplicate traces of the same session: none; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-27B-MLX-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 1k tokens 102.3 PP tok/s 8.8 TG tok/s 10012 TTFT (ms) 16.0 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rjrbtw6f",
   "source_url": "https://omlx.ai/benchmarks/performance/rjrbtw6f",
   "tg_tokens": null,
   "tps": 8.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-z7em5wpd-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: hful4jis; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 899.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 899.0 94.0 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z7em5wpd",
   "source_url": "https://omlx.ai/benchmarks/performance/z7em5wpd",
   "tg_tokens": null,
   "tps": 94.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-z7em5wpd-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: hful4jis; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1077.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,077 89.5 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z7em5wpd",
   "source_url": "https://omlx.ai/benchmarks/performance/z7em5wpd",
   "tg_tokens": null,
   "tps": 89.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-vb95apk9-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2477.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,477 71.0 21.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vb95apk9",
   "source_url": "https://omlx.ai/benchmarks/performance/vb95apk9",
   "tg_tokens": null,
   "tps": 71.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-vb95apk9-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2445.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 2,445 69.1 22.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vb95apk9",
   "source_url": "https://omlx.ai/benchmarks/performance/vb95apk9",
   "tg_tokens": null,
   "tps": 69.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-19tjz54x-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2343.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,343 73.5 21.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/19tjz54x",
   "source_url": "https://omlx.ai/benchmarks/performance/19tjz54x",
   "tg_tokens": null,
   "tps": 73.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-19tjz54x-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2520.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 2,520 82.1 22.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/19tjz54x",
   "source_url": "https://omlx.ai/benchmarks/performance/19tjz54x",
   "tg_tokens": null,
   "tps": 82.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-h9vm1ca1-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: k0dk8lri; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2399.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,399 83.3 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h9vm1ca1",
   "source_url": "https://omlx.ai/benchmarks/performance/h9vm1ca1",
   "tg_tokens": null,
   "tps": 83.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-h9vm1ca1-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: k0dk8lri; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2223.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 2,223 68.6 20.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h9vm1ca1",
   "source_url": "https://omlx.ai/benchmarks/performance/h9vm1ca1",
   "tg_tokens": null,
   "tps": 68.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-hqd5d4g5-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: mm6du6oj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2376.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,376 84.5 20.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hqd5d4g5",
   "source_url": "https://omlx.ai/benchmarks/performance/hqd5d4g5",
   "tg_tokens": null,
   "tps": 84.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-hqd5d4g5-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: mm6du6oj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2228.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 2,228 72.9 20.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hqd5d4g5",
   "source_url": "https://omlx.ai/benchmarks/performance/hqd5d4g5",
   "tg_tokens": null,
   "tps": 72.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0g72vfm0-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 20.7 GB; TTFT 7338 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2233.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k tokens 2,233 PP tok/s 73.1 TG tok/s 7338 TTFT (ms) 20.7 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0g72vfm0",
   "source_url": "https://omlx.ai/benchmarks/performance/0g72vfm0",
   "tg_tokens": null,
   "tps": 73.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-oi3grngr-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: ssogep2h; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2651.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,651 97.1 20.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oi3grngr",
   "source_url": "https://omlx.ai/benchmarks/performance/oi3grngr",
   "tg_tokens": null,
   "tps": 97.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-oi3grngr-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: ssogep2h; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2712.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 2,712 83.7 20.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oi3grngr",
   "source_url": "https://omlx.ai/benchmarks/performance/oi3grngr",
   "tg_tokens": null,
   "tps": 83.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-pgkqnw5f-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: c29ivpf5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2364.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,364 84.6 20.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pgkqnw5f",
   "source_url": "https://omlx.ai/benchmarks/performance/pgkqnw5f",
   "tg_tokens": null,
   "tps": 84.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-pgkqnw5f-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: c29ivpf5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2449.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 2,449 68.9 20.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pgkqnw5f",
   "source_url": "https://omlx.ai/benchmarks/performance/pgkqnw5f",
   "tg_tokens": null,
   "tps": 68.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-apbuxvny-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 20.8 GB; TTFT 6666 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2458.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k tokens 2,458 PP tok/s 69.0 TG tok/s 6666 TTFT (ms) 20.8 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/apbuxvny",
   "source_url": "https://omlx.ai/benchmarks/performance/apbuxvny",
   "tg_tokens": null,
   "tps": 69.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-cp2y03y1-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: t9s8qpbd; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2625.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,625 93.6 20.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cp2y03y1",
   "source_url": "https://omlx.ai/benchmarks/performance/cp2y03y1",
   "tg_tokens": null,
   "tps": 93.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-cp2y03y1-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: t9s8qpbd; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2691.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 2,691 82.7 20.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cp2y03y1",
   "source_url": "https://omlx.ai/benchmarks/performance/cp2y03y1",
   "tg_tokens": null,
   "tps": 82.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-82bpwi0d-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: af1cyv7x; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2668.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,668 94.6 20.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/82bpwi0d",
   "source_url": "https://omlx.ai/benchmarks/performance/82bpwi0d",
   "tg_tokens": null,
   "tps": 94.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-82bpwi0d-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: af1cyv7x; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2516.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 2,516 72.7 20.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/82bpwi0d",
   "source_url": "https://omlx.ai/benchmarks/performance/82bpwi0d",
   "tg_tokens": null,
   "tps": 72.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-huxamho0-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: exqdjstg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2655.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,655 96.7 20.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/huxamho0",
   "source_url": "https://omlx.ai/benchmarks/performance/huxamho0",
   "tg_tokens": null,
   "tps": 96.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-huxamho0-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: exqdjstg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2702.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 2,702 83.0 20.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/huxamho0",
   "source_url": "https://omlx.ai/benchmarks/performance/huxamho0",
   "tg_tokens": null,
   "tps": 83.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1b9vt58z-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: v6ul3zpa; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2662.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,662 81.5 21.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1b9vt58z",
   "source_url": "https://omlx.ai/benchmarks/performance/1b9vt58z",
   "tg_tokens": null,
   "tps": 81.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1b9vt58z-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: v6ul3zpa; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2564.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 2,564 82.3 22.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1b9vt58z",
   "source_url": "https://omlx.ai/benchmarks/performance/1b9vt58z",
   "tg_tokens": null,
   "tps": 82.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1kj40wzw-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: nqicdvfr; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2681.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,681 82.8 21.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1kj40wzw",
   "source_url": "https://omlx.ai/benchmarks/performance/1kj40wzw",
   "tg_tokens": null,
   "tps": 82.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-1kj40wzw-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: nqicdvfr; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2710.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 2,710 94.5 22.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1kj40wzw",
   "source_url": "https://omlx.ai/benchmarks/performance/1kj40wzw",
   "tg_tokens": null,
   "tps": 94.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-rb32b1q2-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: f2v8dyaf; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2567.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,567 84.0 21.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rb32b1q2",
   "source_url": "https://omlx.ai/benchmarks/performance/rb32b1q2",
   "tg_tokens": null,
   "tps": 84.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-rb32b1q2-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: f2v8dyaf; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2722.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 2,722 91.2 22.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rb32b1q2",
   "source_url": "https://omlx.ai/benchmarks/performance/rb32b1q2",
   "tg_tokens": null,
   "tps": 91.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-vvz40n3x-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2282.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,282 85.1 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vvz40n3x",
   "source_url": "https://omlx.ai/benchmarks/performance/vvz40n3x",
   "tg_tokens": null,
   "tps": 85.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-vvz40n3x-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2269.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 2,269 84.7 20.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vvz40n3x",
   "source_url": "https://omlx.ai/benchmarks/performance/vvz40n3x",
   "tg_tokens": null,
   "tps": 84.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-tjq5440o-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: ym8oagw0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 688.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 688.4 108.2 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tjq5440o",
   "source_url": "https://omlx.ai/benchmarks/performance/tjq5440o",
   "tg_tokens": null,
   "tps": 108.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-tjq5440o-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: ym8oagw0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1709.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 1,709 104.4 20.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tjq5440o",
   "source_url": "https://omlx.ai/benchmarks/performance/tjq5440o",
   "tg_tokens": null,
   "tps": 104.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-tjq5440o-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: ym8oagw0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1634.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 64k 1,634 61.1 21.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tjq5440o",
   "source_url": "https://omlx.ai/benchmarks/performance/tjq5440o",
   "tg_tokens": null,
   "tps": 61.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Pro",
   "id": "omlx-zhukjfsb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 16.0 GB; TTFT 10020 ms; duplicate traces of the same session: none; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 1k tokens 102.2 PP tok/s 8.8 TG tok/s 10020 TTFT (ms) 16.0 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zhukjfsb",
   "source_url": "https://omlx.ai/benchmarks/performance/zhukjfsb",
   "tg_tokens": null,
   "tps": 8.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qgee7qda-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.4 GB at this context; duplicate traces of the same session: rre1tz6w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 857.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 857.4 34.9 33.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qgee7qda",
   "source_url": "https://omlx.ai/benchmarks/performance/qgee7qda",
   "tg_tokens": null,
   "tps": 34.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qgee7qda-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: rre1tz6w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 714.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 714.8 36.0 34.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qgee7qda",
   "source_url": "https://omlx.ai/benchmarks/performance/qgee7qda",
   "tg_tokens": null,
   "tps": 36.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mluy2s84-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: zlwsjggc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 775.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 775.6 18.3 28.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mluy2s84",
   "source_url": "https://omlx.ai/benchmarks/performance/mluy2s84",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mluy2s84-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: zlwsjggc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 931.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 931.5 18.2 29.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mluy2s84",
   "source_url": "https://omlx.ai/benchmarks/performance/mluy2s84",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-y2ds4euk-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: zo9el4wj; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 169.0,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-27B-nvfp4 on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 1k 169.0 17.5 16.0 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y2ds4euk",
   "source_url": "https://omlx.ai/benchmarks/performance/y2ds4euk",
   "tg_tokens": null,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-y2ds4euk-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: zo9el4wj; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 167.5,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-27B-nvfp4 on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 4k 167.5 17.2 17.0 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y2ds4euk",
   "source_url": "https://omlx.ai/benchmarks/performance/y2ds4euk",
   "tg_tokens": null,
   "tps": 17.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q9nocx8v-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1305.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 1,305 34.6 35.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q9nocx8v",
   "source_url": "https://omlx.ai/benchmarks/performance/q9nocx8v",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q9nocx8v-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1008.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 1,008 63.1 36.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q9nocx8v",
   "source_url": "https://omlx.ai/benchmarks/performance/q9nocx8v",
   "tg_tokens": null,
   "tps": 63.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9cr48y2v-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace, single run at 16k context; peak memory 108.8 GB; TTFT 15725 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1042.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ4-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k tokens 1,042 PP tok/s 24.3 TG tok/s 15725 TTFT (ms) 108.8 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9cr48y2v",
   "source_url": "https://omlx.ai/benchmarks/performance/9cr48y2v",
   "tg_tokens": null,
   "tps": 24.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1370w0eb-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 70.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1110.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,110 52.5 70.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1370w0eb",
   "source_url": "https://omlx.ai/benchmarks/performance/1370w0eb",
   "tg_tokens": null,
   "tps": 52.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1370w0eb-ctx32768",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3290.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 3,290 24.9 71.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1370w0eb",
   "source_url": "https://omlx.ai/benchmarks/performance/1370w0eb",
   "tg_tokens": null,
   "tps": 24.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-kwx2ymo4-ctx1024",
   "model": "Gemma 4 12B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 12.8 GB; TTFT 4790 ms; duplicate traces of the same session: none; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 213.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-12B-it-8bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 8bit; 1k tokens 213.8 PP tok/s 10.2 TG tok/s 4790 TTFT (ms) 12.8 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kwx2ymo4",
   "source_url": "https://omlx.ai/benchmarks/performance/kwx2ymo4",
   "tg_tokens": null,
   "tps": 10.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-g7rt5sv4-ctx1024",
   "model": "Gemma 4 12B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 12.8 GB; TTFT 4815 ms; duplicate traces of the same session: none; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 212.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-12B-it-8bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 8bit; 1k tokens 212.7 PP tok/s 10.2 TG tok/s 4815 TTFT (ms) 12.8 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g7rt5sv4",
   "source_url": "https://omlx.ai/benchmarks/performance/g7rt5sv4",
   "tg_tokens": null,
   "tps": 10.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-k0szooqf-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 4.7 GB; TTFT 1028 ms; duplicate traces of the same session: none; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 996.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 1k tokens 996.1 PP tok/s 43.2 TG tok/s 1028 TTFT (ms) 4.7 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k0szooqf",
   "source_url": "https://omlx.ai/benchmarks/performance/k0szooqf",
   "tg_tokens": null,
   "tps": 43.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-sid73efk-ctx1024",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "mlx-community mixed build; oMLX community benchmark trace; peak memory 87.8 GB at this context; duplicate traces of the same session: 2fyd549l, rroj0hed; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 806.7,
   "provenance": "community",
   "quant": "2.4bit",
   "quote": "mlx-community-DeepSeek-V4-Flash-0731-2.4bit-mixed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 1k 806.7 28.7 87.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sid73efk",
   "source_url": "https://omlx.ai/benchmarks/performance/sid73efk",
   "tg_tokens": null,
   "tps": 28.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-sid73efk-ctx8192",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "mlx-community mixed build; oMLX community benchmark trace; peak memory 89.3 GB at this context; duplicate traces of the same session: 2fyd549l, rroj0hed; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 803.9,
   "provenance": "community",
   "quant": "2.4bit",
   "quote": "mlx-community-DeepSeek-V4-Flash-0731-2.4bit-mixed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 8k 803.9 48.1 89.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sid73efk",
   "source_url": "https://omlx.ai/benchmarks/performance/sid73efk",
   "tg_tokens": null,
   "tps": 48.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-sid73efk-ctx32768",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "mlx-community mixed build; oMLX community benchmark trace; peak memory 92.8 GB at this context; duplicate traces of the same session: 2fyd549l, rroj0hed; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 668.1,
   "provenance": "community",
   "quant": "2.4bit",
   "quote": "mlx-community-DeepSeek-V4-Flash-0731-2.4bit-mixed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 32k 668.1 52.6 92.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sid73efk",
   "source_url": "https://omlx.ai/benchmarks/performance/sid73efk",
   "tg_tokens": null,
   "tps": 52.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kax4f2fr-ctx1024",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "mlx-community mixed build; oMLX community benchmark trace; peak memory 87.8 GB at this context; duplicate traces of the same session: 0a34m69n, na0p1o2g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 799.2,
   "provenance": "community",
   "quant": "2.4bit",
   "quote": "mlx-community-DeepSeek-V4-Flash-0731-2.4bit-mixed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 1k 799.2 28.7 87.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kax4f2fr",
   "source_url": "https://omlx.ai/benchmarks/performance/kax4f2fr",
   "tg_tokens": null,
   "tps": 28.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-kax4f2fr-ctx8192",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "mlx-community mixed build; oMLX community benchmark trace; peak memory 89.3 GB at this context; duplicate traces of the same session: 0a34m69n, na0p1o2g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 803.8,
   "provenance": "community",
   "quant": "2.4bit",
   "quote": "mlx-community-DeepSeek-V4-Flash-0731-2.4bit-mixed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 8k 803.8 49.5 89.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kax4f2fr",
   "source_url": "https://omlx.ai/benchmarks/performance/kax4f2fr",
   "tg_tokens": null,
   "tps": 49.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kax4f2fr-ctx32768",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "mlx-community mixed build; oMLX community benchmark trace; peak memory 92.8 GB at this context; duplicate traces of the same session: 0a34m69n, na0p1o2g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 668.2,
   "provenance": "community",
   "quant": "2.4bit",
   "quote": "mlx-community-DeepSeek-V4-Flash-0731-2.4bit-mixed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 32k 668.2 49.2 92.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kax4f2fr",
   "source_url": "https://omlx.ai/benchmarks/performance/kax4f2fr",
   "tg_tokens": null,
   "tps": 49.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-polz607b-ctx4096",
   "model": "GLM-4.7-Flash",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 17.4 GB; TTFT 14288 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 286.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-4.7-Flash-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k tokens 286.7 PP tok/s 29.9 TG tok/s 14288 TTFT (ms) 17.4 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/polz607b",
   "source_url": "https://omlx.ai/benchmarks/performance/polz607b",
   "tg_tokens": null,
   "tps": 29.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-v70xfzth-ctx1024",
   "model": "GLM-4.7-Flash",
   "notes": "oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: mtibgmn0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1002.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "GLM-4.7-Flash-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 8bit; 1k 1,002 52.2 30.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v70xfzth",
   "source_url": "https://omlx.ai/benchmarks/performance/v70xfzth",
   "tg_tokens": null,
   "tps": 52.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-v70xfzth-ctx4096",
   "model": "GLM-4.7-Flash",
   "notes": "oMLX community benchmark trace; peak memory 31.3 GB at this context; duplicate traces of the same session: mtibgmn0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1301.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "GLM-4.7-Flash-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 8bit; 4k 1,301 43.3 31.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v70xfzth",
   "source_url": "https://omlx.ai/benchmarks/performance/v70xfzth",
   "tg_tokens": null,
   "tps": 43.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-d45fngc0-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.2 GB at this context; duplicate traces of the same session: kqkxxv4n, mxr9vecb, wnhmmz4g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2400.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 2,400 81.1 177.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d45fngc0",
   "source_url": "https://omlx.ai/benchmarks/performance/d45fngc0",
   "tg_tokens": null,
   "tps": 81.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-d45fngc0-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.8 GB at this context; duplicate traces of the same session: kqkxxv4n, mxr9vecb, wnhmmz4g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2417.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,417 120.3 177.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d45fngc0",
   "source_url": "https://omlx.ai/benchmarks/performance/d45fngc0",
   "tg_tokens": null,
   "tps": 120.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-d45fngc0-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 179.3 GB at this context; duplicate traces of the same session: kqkxxv4n, mxr9vecb, wnhmmz4g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2393.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,393 100.2 179.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d45fngc0",
   "source_url": "https://omlx.ai/benchmarks/performance/d45fngc0",
   "tg_tokens": null,
   "tps": 100.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-d45fngc0-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 182.3 GB at this context; duplicate traces of the same session: kqkxxv4n, mxr9vecb, wnhmmz4g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2338.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,338 114.5 182.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d45fngc0",
   "source_url": "https://omlx.ai/benchmarks/performance/d45fngc0",
   "tg_tokens": null,
   "tps": 114.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-qlskpd0b-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.6 GB at this context; duplicate traces of the same session: blamuh4x, xvnih38o; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2274.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 2,274 136.5 174.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qlskpd0b",
   "source_url": "https://omlx.ai/benchmarks/performance/qlskpd0b",
   "tg_tokens": null,
   "tps": 136.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-qlskpd0b-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.3 GB at this context; duplicate traces of the same session: blamuh4x, xvnih38o; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2369.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,369 121.6 175.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qlskpd0b",
   "source_url": "https://omlx.ai/benchmarks/performance/qlskpd0b",
   "tg_tokens": null,
   "tps": 121.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-qlskpd0b-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.1 GB at this context; duplicate traces of the same session: blamuh4x, xvnih38o; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2344.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,344 99.9 177.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qlskpd0b",
   "source_url": "https://omlx.ai/benchmarks/performance/qlskpd0b",
   "tg_tokens": null,
   "tps": 99.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-qlskpd0b-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 180.5 GB at this context; duplicate traces of the same session: blamuh4x, xvnih38o; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2292.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,292 150.5 180.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qlskpd0b",
   "source_url": "https://omlx.ai/benchmarks/performance/qlskpd0b",
   "tg_tokens": null,
   "tps": 150.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lmnqt76u-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 214.7 GB at this context; duplicate traces of the same session: 026bct6y, 5qbzo75m, aqfiz9cz, ay77atva, giay0wzg, qgzde7nn, x35tx97w; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 403.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "GLM-5.3-Flash-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 1k 403.7 36.5 214.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lmnqt76u",
   "source_url": "https://omlx.ai/benchmarks/performance/lmnqt76u",
   "tg_tokens": null,
   "tps": 36.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lmnqt76u-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 216.1 GB at this context; duplicate traces of the same session: 026bct6y, 5qbzo75m, aqfiz9cz, ay77atva, giay0wzg, qgzde7nn, x35tx97w; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 422.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "GLM-5.3-Flash-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 4k 422.4 35.5 216.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lmnqt76u",
   "source_url": "https://omlx.ai/benchmarks/performance/lmnqt76u",
   "tg_tokens": null,
   "tps": 35.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lmnqt76u-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 216.9 GB at this context; duplicate traces of the same session: 026bct6y, 5qbzo75m, aqfiz9cz, ay77atva, giay0wzg, qgzde7nn, x35tx97w; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 410.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "GLM-5.3-Flash-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 8k 410.2 45.0 216.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lmnqt76u",
   "source_url": "https://omlx.ai/benchmarks/performance/lmnqt76u",
   "tg_tokens": null,
   "tps": 45.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lmnqt76u-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 217.6 GB at this context; duplicate traces of the same session: 026bct6y, 5qbzo75m, aqfiz9cz, ay77atva, giay0wzg, qgzde7nn, x35tx97w; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 402.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "GLM-5.3-Flash-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 16k 402.4 65.0 217.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lmnqt76u",
   "source_url": "https://omlx.ai/benchmarks/performance/lmnqt76u",
   "tg_tokens": null,
   "tps": 65.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-lmnqt76u-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 218.9 GB at this context; duplicate traces of the same session: 026bct6y, 5qbzo75m, aqfiz9cz, ay77atva, giay0wzg, qgzde7nn, x35tx97w; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 397.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "GLM-5.3-Flash-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 32k 397.2 65.5 218.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lmnqt76u",
   "source_url": "https://omlx.ai/benchmarks/performance/lmnqt76u",
   "tg_tokens": null,
   "tps": 65.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-lmnqt76u-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 221.6 GB at this context; duplicate traces of the same session: 026bct6y, 5qbzo75m, aqfiz9cz, ay77atva, giay0wzg, qgzde7nn, x35tx97w; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 391.8,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "GLM-5.3-Flash-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 64k 391.8 44.1 221.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lmnqt76u",
   "source_url": "https://omlx.ai/benchmarks/performance/lmnqt76u",
   "tg_tokens": null,
   "tps": 44.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-lmnqt76u-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 227.4 GB at this context; duplicate traces of the same session: 026bct6y, 5qbzo75m, aqfiz9cz, ay77atva, giay0wzg, qgzde7nn, x35tx97w; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 384.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "GLM-5.3-Flash-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 128k 384.4 70.1 227.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lmnqt76u",
   "source_url": "https://omlx.ai/benchmarks/performance/lmnqt76u",
   "tg_tokens": null,
   "tps": 70.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lmnqt76u-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 233.2 GB at this context; duplicate traces of the same session: 026bct6y, 5qbzo75m, aqfiz9cz, ay77atva, giay0wzg, qgzde7nn, x35tx97w; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 377.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "GLM-5.3-Flash-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 195k 377.4 35.3 233.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lmnqt76u",
   "source_url": "https://omlx.ai/benchmarks/performance/lmnqt76u",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-c4ag469x-ctx1024",
   "model": "Qwen2.5-Coder-7B-Instruct",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 5.1 GB at this context; duplicate traces of the same session: 2aru483u, ag44kqh5, j3fqmd7x, k0vsd8i9, lihppzd6; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 250.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen2.5-Coder-7B-Instruct-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 1k 250.6 36.6 5.1 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c4ag469x",
   "source_url": "https://omlx.ai/benchmarks/performance/c4ag469x",
   "tg_tokens": null,
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-c4ag469x-ctx4096",
   "model": "Qwen2.5-Coder-7B-Instruct",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 5.0 GB at this context; duplicate traces of the same session: 2aru483u, ag44kqh5, j3fqmd7x, k0vsd8i9, lihppzd6; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 250.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen2.5-Coder-7B-Instruct-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 4k 250.1 33.6 5.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c4ag469x",
   "source_url": "https://omlx.ai/benchmarks/performance/c4ag469x",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-c4ag469x-ctx8192",
   "model": "Qwen2.5-Coder-7B-Instruct",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 5.3 GB at this context; duplicate traces of the same session: 2aru483u, ag44kqh5, j3fqmd7x, k0vsd8i9, lihppzd6; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 236.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen2.5-Coder-7B-Instruct-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 8k 236.0 29.9 5.3 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c4ag469x",
   "source_url": "https://omlx.ai/benchmarks/performance/c4ag469x",
   "tg_tokens": null,
   "tps": 29.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-c4ag469x-ctx16384",
   "model": "Qwen2.5-Coder-7B-Instruct",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 5.7 GB at this context; duplicate traces of the same session: 2aru483u, ag44kqh5, j3fqmd7x, k0vsd8i9, lihppzd6; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 211.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen2.5-Coder-7B-Instruct-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 16k 211.5 25.7 5.7 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c4ag469x",
   "source_url": "https://omlx.ai/benchmarks/performance/c4ag469x",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-c4ag469x-ctx32768",
   "model": "Qwen2.5-Coder-7B-Instruct",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: 2aru483u, ag44kqh5, j3fqmd7x, k0vsd8i9, lihppzd6; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 174.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen2.5-Coder-7B-Instruct-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 32k 174.8 19.0 6.5 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c4ag469x",
   "source_url": "https://omlx.ai/benchmarks/performance/c4ag469x",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-c4ag469x-ctx65536",
   "model": "Qwen2.5-Coder-7B-Instruct",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 8.3 GB at this context; duplicate traces of the same session: 2aru483u, ag44kqh5, j3fqmd7x, k0vsd8i9, lihppzd6; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 129.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen2.5-Coder-7B-Instruct-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 64k 129.7 12.9 8.3 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c4ag469x",
   "source_url": "https://omlx.ai/benchmarks/performance/c4ag469x",
   "tg_tokens": null,
   "tps": 12.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-rg2zj2uk-ctx1024",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 1.4 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2114.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-0.8B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 2,114 181.3 1.4 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rg2zj2uk",
   "source_url": "https://omlx.ai/benchmarks/performance/rg2zj2uk",
   "tg_tokens": null,
   "tps": 181.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-rg2zj2uk-ctx4096",
   "model": "Qwen3.5-0.8B",
   "notes": "oMLX community benchmark trace; peak memory 1.6 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2161.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-0.8B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 2,161 166.3 1.6 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rg2zj2uk",
   "source_url": "https://omlx.ai/benchmarks/performance/rg2zj2uk",
   "tg_tokens": null,
   "tps": 166.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-jlkmr6ps-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace, single run at 8k context; peak memory 6.6 GB; TTFT 21949 ms; duplicate traces of the same session: none; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 373.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 8k tokens 373.2 PP tok/s 26.6 TG tok/s 21949 TTFT (ms) 6.6 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jlkmr6ps",
   "source_url": "https://omlx.ai/benchmarks/performance/jlkmr6ps",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-f4s7hihi-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 5.9 GB; TTFT 2821 ms; duplicate traces of the same session: none; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 363.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 1k tokens 363.0 PP tok/s 28.3 TG tok/s 2821 TTFT (ms) 5.9 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f4s7hihi",
   "source_url": "https://omlx.ai/benchmarks/performance/f4s7hihi",
   "tg_tokens": null,
   "tps": 28.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-bpmrju1u-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.3 GB; TTFT 1448 ms; duplicate traces of the same session: none; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 707.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 1k tokens 707.4 PP tok/s 57.6 TG tok/s 1448 TTFT (ms) 19.3 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bpmrju1u",
   "source_url": "https://omlx.ai/benchmarks/performance/bpmrju1u",
   "tg_tokens": null,
   "tps": 57.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-uf48fwel-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.3 GB; TTFT 1441 ms; duplicate traces of the same session: none; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 710.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 1k tokens 710.7 PP tok/s 57.8 TG tok/s 1441 TTFT (ms) 19.3 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uf48fwel",
   "source_url": "https://omlx.ai/benchmarks/performance/uf48fwel",
   "tg_tokens": null,
   "tps": 57.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-dyof9m82-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 688.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 688.4 108.2 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dyof9m82",
   "source_url": "https://omlx.ai/benchmarks/performance/dyof9m82",
   "tg_tokens": null,
   "tps": 108.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-dyof9m82-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1709.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 1,709 104.4 20.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dyof9m82",
   "source_url": "https://omlx.ai/benchmarks/performance/dyof9m82",
   "tg_tokens": null,
   "tps": 104.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-dyof9m82-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1634.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 64k 1,634 61.1 21.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dyof9m82",
   "source_url": "https://omlx.ai/benchmarks/performance/dyof9m82",
   "tg_tokens": null,
   "tps": 61.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ehqa9egw-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: 1g4b58zx; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1173.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 1,173 102.4 24.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ehqa9egw",
   "source_url": "https://omlx.ai/benchmarks/performance/ehqa9egw",
   "tg_tokens": null,
   "tps": 102.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-ehqa9egw-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: 1g4b58zx; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 971.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 16k 971.8 69.7 25.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ehqa9egw",
   "source_url": "https://omlx.ai/benchmarks/performance/ehqa9egw",
   "tg_tokens": null,
   "tps": 69.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-bfakt5fx-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 0qc3oflp; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1233.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 1,233 82.1 22.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bfakt5fx",
   "source_url": "https://omlx.ai/benchmarks/performance/bfakt5fx",
   "tg_tokens": null,
   "tps": 82.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-bfakt5fx-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 0qc3oflp; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1009.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 16k 1,009 78.1 23.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bfakt5fx",
   "source_url": "https://omlx.ai/benchmarks/performance/bfakt5fx",
   "tg_tokens": null,
   "tps": 78.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-wkktycqi-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: gw63nena; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 84.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 1k 84.6 15.6 16.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wkktycqi",
   "source_url": "https://omlx.ai/benchmarks/performance/wkktycqi",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-wkktycqi-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: gw63nena; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 81.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k 81.4 15.1 17.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wkktycqi",
   "source_url": "https://omlx.ai/benchmarks/performance/wkktycqi",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Pro",
   "id": "omlx-oee1kte5-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 16.0 GB; TTFT 9981 ms; duplicate traces of the same session: none; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 1k tokens 102.6 PP tok/s 8.8 TG tok/s 9981 TTFT (ms) 16.0 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oee1kte5",
   "source_url": "https://omlx.ai/benchmarks/performance/oee1kte5",
   "tg_tokens": null,
   "tps": 8.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-3mq7e7il-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: avsur6tg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 259.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 259.3 29.3 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3mq7e7il",
   "source_url": "https://omlx.ai/benchmarks/performance/3mq7e7il",
   "tg_tokens": null,
   "tps": 29.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-3mq7e7il-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: avsur6tg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 237.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 237.1 25.4 19.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3mq7e7il",
   "source_url": "https://omlx.ai/benchmarks/performance/3mq7e7il",
   "tg_tokens": null,
   "tps": 25.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-tw0oveqq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: uasqdedp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 467.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 467.2 18.4 16.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tw0oveqq",
   "source_url": "https://omlx.ai/benchmarks/performance/tw0oveqq",
   "tg_tokens": null,
   "tps": 18.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-tw0oveqq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: uasqdedp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 510.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 510.8 18.1 17.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tw0oveqq",
   "source_url": "https://omlx.ai/benchmarks/performance/tw0oveqq",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-o4rnh1s3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: ur37zjry; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 255.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 255.8 16.6 28.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o4rnh1s3",
   "source_url": "https://omlx.ai/benchmarks/performance/o4rnh1s3",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-o4rnh1s3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.7 GB at this context; duplicate traces of the same session: ur37zjry; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 212.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 212.8 15.6 31.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o4rnh1s3",
   "source_url": "https://omlx.ai/benchmarks/performance/o4rnh1s3",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Pro",
   "id": "omlx-mbqn5yo2-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 88ac1kuo; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 99.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 1k 99.7 18.0 18.4 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mbqn5yo2",
   "source_url": "https://omlx.ai/benchmarks/performance/mbqn5yo2",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-mbqn5yo2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: 88ac1kuo; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 4k 102.4 20.0 19.7 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mbqn5yo2",
   "source_url": "https://omlx.ai/benchmarks/performance/mbqn5yo2",
   "tg_tokens": null,
   "tps": 20.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-h2192i8m-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 20.6 GB; TTFT 41352 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 792.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k tokens 792.4 PP tok/s 62.0 TG tok/s 41352 TTFT (ms) 20.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h2192i8m",
   "source_url": "https://omlx.ai/benchmarks/performance/h2192i8m",
   "tg_tokens": null,
   "tps": 62.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4z9gi12a-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1305.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 1,305 34.6 35.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4z9gi12a",
   "source_url": "https://omlx.ai/benchmarks/performance/4z9gi12a",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4z9gi12a-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1008.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 1,008 63.1 36.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4z9gi12a",
   "source_url": "https://omlx.ai/benchmarks/performance/4z9gi12a",
   "tg_tokens": null,
   "tps": 63.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9f75a9gl-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.4 GB at this context; duplicate traces of the same session: 0i37xdgt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 867.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 867.1 30.7 33.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9f75a9gl",
   "source_url": "https://omlx.ai/benchmarks/performance/9f75a9gl",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9f75a9gl-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: 0i37xdgt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 706.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 706.4 38.2 36.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9f75a9gl",
   "source_url": "https://omlx.ai/benchmarks/performance/9f75a9gl",
   "tg_tokens": null,
   "tps": 38.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-7pbrcxw5-ctx1024",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: knt8bqp0; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 698.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 698.8 25.1 5.8 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7pbrcxw5",
   "source_url": "https://omlx.ai/benchmarks/performance/7pbrcxw5",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-7pbrcxw5-ctx4096",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: knt8bqp0; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 746.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 746.9 25.5 6.2 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7pbrcxw5",
   "source_url": "https://omlx.ai/benchmarks/performance/7pbrcxw5",
   "tg_tokens": null,
   "tps": 25.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-r6k8m3w7-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 70.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1110.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,110 52.5 70.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r6k8m3w7",
   "source_url": "https://omlx.ai/benchmarks/performance/r6k8m3w7",
   "tg_tokens": null,
   "tps": 52.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-r6k8m3w7-ctx32768",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3290.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 3,290 24.9 71.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r6k8m3w7",
   "source_url": "https://omlx.ai/benchmarks/performance/r6k8m3w7",
   "tg_tokens": null,
   "tps": 24.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-csjjz31w-ctx1024",
   "model": "Gemma 4 12B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 12.8 GB; TTFT 4730 ms; duplicate traces of the same session: none; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 216.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-12B-it-8bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 8bit; 1k tokens 216.5 PP tok/s 10.1 TG tok/s 4730 TTFT (ms) 12.8 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/csjjz31w",
   "source_url": "https://omlx.ai/benchmarks/performance/csjjz31w",
   "tg_tokens": null,
   "tps": 10.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2",
   "id": "omlx-6jl2zdih-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 4.5 GB; TTFT 2600 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 393.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-4bit on M2 (10c); M2 (10c) \u00b7 16 GB \u00b7 4bit; 1k tokens 393.9 PP tok/s 29.1 TG tok/s 2600 TTFT (ms) 4.5 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6jl2zdih",
   "source_url": "https://omlx.ai/benchmarks/performance/6jl2zdih",
   "tg_tokens": null,
   "tps": 29.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-cmsaegyt-ctx1024",
   "model": "Qwen2-57B-A14B-Instruct",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 30.8 GB at this context; duplicate traces of the same session: l526z6tl; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 230.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen2-57B-A14B-Instruct-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 unknown; 1k 230.6 29.2 30.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cmsaegyt",
   "source_url": "https://omlx.ai/benchmarks/performance/cmsaegyt",
   "tg_tokens": null,
   "tps": 29.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-cmsaegyt-ctx4096",
   "model": "Qwen2-57B-A14B-Instruct",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 31.3 GB at this context; duplicate traces of the same session: l526z6tl; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 236.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen2-57B-A14B-Instruct-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 unknown; 4k 236.7 28.4 31.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cmsaegyt",
   "source_url": "https://omlx.ai/benchmarks/performance/cmsaegyt",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fzk0273p-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 90.0 GB at this context; duplicate traces of the same session: 6j3k2bgo, mb8mi3i9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 87.2,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 87.2 9.0 90.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fzk0273p",
   "source_url": "https://omlx.ai/benchmarks/performance/fzk0273p",
   "tg_tokens": null,
   "tps": 9.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fzk0273p-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.2 GB at this context; duplicate traces of the same session: 6j3k2bgo, mb8mi3i9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 283.4,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 283.4 10.5 92.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fzk0273p",
   "source_url": "https://omlx.ai/benchmarks/performance/fzk0273p",
   "tg_tokens": null,
   "tps": 10.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fzk0273p-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.3 GB at this context; duplicate traces of the same session: 6j3k2bgo, mb8mi3i9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 315.2,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 315.2 14.4 92.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fzk0273p",
   "source_url": "https://omlx.ai/benchmarks/performance/fzk0273p",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-31qxq398-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.6 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2274.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 2,274 136.5 174.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/31qxq398",
   "source_url": "https://omlx.ai/benchmarks/performance/31qxq398",
   "tg_tokens": null,
   "tps": 136.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-31qxq398-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.3 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2369.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,369 121.6 175.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/31qxq398",
   "source_url": "https://omlx.ai/benchmarks/performance/31qxq398",
   "tg_tokens": null,
   "tps": 121.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-31qxq398-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2344.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,344 99.9 177.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/31qxq398",
   "source_url": "https://omlx.ai/benchmarks/performance/31qxq398",
   "tg_tokens": null,
   "tps": 99.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-31qxq398-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 180.5 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2292.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,292 150.5 180.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/31qxq398",
   "source_url": "https://omlx.ai/benchmarks/performance/31qxq398",
   "tg_tokens": null,
   "tps": 150.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-eq63wf8z-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.5 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 449.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 449.5 30.8 173.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eq63wf8z",
   "source_url": "https://omlx.ai/benchmarks/performance/eq63wf8z",
   "tg_tokens": null,
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-eq63wf8z-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 545.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 545.5 24.7 174.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eq63wf8z",
   "source_url": "https://omlx.ai/benchmarks/performance/eq63wf8z",
   "tg_tokens": null,
   "tps": 24.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-p4eg1nty-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 316.2 GB at this context; duplicate traces of the same session: 69ujcmmp; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 441.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 441.6 25.7 316.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p4eg1nty",
   "source_url": "https://omlx.ai/benchmarks/performance/p4eg1nty",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-p4eg1nty-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 317.3 GB at this context; duplicate traces of the same session: 69ujcmmp; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 573.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 573.5 22.8 317.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p4eg1nty",
   "source_url": "https://omlx.ai/benchmarks/performance/p4eg1nty",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xspqh35z-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Jundot MTP build; oMLX community benchmark trace; peak memory 129.6 GB at this context; duplicate traces of the same session: 27nulsk1, iwu30ah1, jbk77vkd, t0eyilod; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 645.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Jundot:Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 645.7 73.5 129.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xspqh35z",
   "source_url": "https://omlx.ai/benchmarks/performance/xspqh35z",
   "tg_tokens": null,
   "tps": 73.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xspqh35z-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Jundot MTP build; oMLX community benchmark trace; peak memory 130.0 GB at this context; duplicate traces of the same session: 27nulsk1, iwu30ah1, jbk77vkd, t0eyilod; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 831.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Jundot:Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 831.7 69.1 130.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xspqh35z",
   "source_url": "https://omlx.ai/benchmarks/performance/xspqh35z",
   "tg_tokens": null,
   "tps": 69.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xspqh35z-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Jundot MTP build; oMLX community benchmark trace; peak memory 130.2 GB at this context; duplicate traces of the same session: 27nulsk1, iwu30ah1, jbk77vkd, t0eyilod; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1027.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Jundot:Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 1,027 74.8 130.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xspqh35z",
   "source_url": "https://omlx.ai/benchmarks/performance/xspqh35z",
   "tg_tokens": null,
   "tps": 74.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xspqh35z-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Jundot MTP build; oMLX community benchmark trace; peak memory 133.6 GB at this context; duplicate traces of the same session: 27nulsk1, iwu30ah1, jbk77vkd, t0eyilod; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1006.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Jundot:Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 64k 1,006 64.3 133.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xspqh35z",
   "source_url": "https://omlx.ai/benchmarks/performance/xspqh35z",
   "tg_tokens": null,
   "tps": 64.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xspqh35z-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Jundot MTP build; oMLX community benchmark trace; peak memory 137.9 GB at this context; duplicate traces of the same session: 27nulsk1, iwu30ah1, jbk77vkd, t0eyilod; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 990.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Jundot:Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 128k 990.5 55.5 137.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xspqh35z",
   "source_url": "https://omlx.ai/benchmarks/performance/xspqh35z",
   "tg_tokens": null,
   "tps": 55.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4e4ctbuj-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Jundot MTP build; oMLX community benchmark trace; peak memory 129.5 GB at this context; duplicate traces of the same session: 03uo3q2d; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 598.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Jundot:Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 598.4 75.4 129.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4e4ctbuj",
   "source_url": "https://omlx.ai/benchmarks/performance/4e4ctbuj",
   "tg_tokens": null,
   "tps": 75.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-4e4ctbuj-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Jundot MTP build; oMLX community benchmark trace; peak memory 130.0 GB at this context; duplicate traces of the same session: 03uo3q2d; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 812.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Jundot:Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 812.1 66.8 130.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4e4ctbuj",
   "source_url": "https://omlx.ai/benchmarks/performance/4e4ctbuj",
   "tg_tokens": null,
   "tps": 66.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-ue9hra3l-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 13.3 GB at this context; duplicate traces of the same session: hp1o7pz2; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1815.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k 1,815 66.5 13.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ue9hra3l",
   "source_url": "https://omlx.ai/benchmarks/performance/ue9hra3l",
   "tg_tokens": null,
   "tps": 66.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-ue9hra3l-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: hp1o7pz2; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1784.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 1,784 64.5 13.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ue9hra3l",
   "source_url": "https://omlx.ai/benchmarks/performance/ue9hra3l",
   "tg_tokens": null,
   "tps": 64.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-x15ff660-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 9.7 GB at this context; duplicate traces of the same session: iy5af1ob, t2kok7ux, tz9sbi12, wykcapv4; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 348.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 1k 348.3 15.6 9.7 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x15ff660",
   "source_url": "https://omlx.ai/benchmarks/performance/x15ff660",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-x15ff660-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 10.2 GB at this context; duplicate traces of the same session: iy5af1ob, t2kok7ux, tz9sbi12, wykcapv4; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 517.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 4k 517.9 19.7 10.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x15ff660",
   "source_url": "https://omlx.ai/benchmarks/performance/x15ff660",
   "tg_tokens": null,
   "tps": 19.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-x15ff660-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 10.3 GB at this context; duplicate traces of the same session: iy5af1ob, t2kok7ux, tz9sbi12, wykcapv4; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 515.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 8k 515.9 16.5 10.3 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x15ff660",
   "source_url": "https://omlx.ai/benchmarks/performance/x15ff660",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-x15ff660-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 10.5 GB at this context; duplicate traces of the same session: iy5af1ob, t2kok7ux, tz9sbi12, wykcapv4; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 494.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 16k 494.0 19.3 10.5 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x15ff660",
   "source_url": "https://omlx.ai/benchmarks/performance/x15ff660",
   "tg_tokens": null,
   "tps": 19.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-x15ff660-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 10.8 GB at this context; duplicate traces of the same session: iy5af1ob, t2kok7ux, tz9sbi12, wykcapv4; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 448.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 32k 448.7 16.5 10.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x15ff660",
   "source_url": "https://omlx.ai/benchmarks/performance/x15ff660",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-2h5aszb7-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 9.7 GB at this context; duplicate traces of the same session: 3t9c3eg4, cs3pu9sb, kluxjwgd, zgdfk7yi; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 343.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 1k 343.1 16.1 9.7 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2h5aszb7",
   "source_url": "https://omlx.ai/benchmarks/performance/2h5aszb7",
   "tg_tokens": null,
   "tps": 16.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-2h5aszb7-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 10.2 GB at this context; duplicate traces of the same session: 3t9c3eg4, cs3pu9sb, kluxjwgd, zgdfk7yi; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 514.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 4k 514.4 20.5 10.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2h5aszb7",
   "source_url": "https://omlx.ai/benchmarks/performance/2h5aszb7",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-2h5aszb7-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 10.3 GB at this context; duplicate traces of the same session: 3t9c3eg4, cs3pu9sb, kluxjwgd, zgdfk7yi; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 519.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 8k 519.6 18.0 10.3 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2h5aszb7",
   "source_url": "https://omlx.ai/benchmarks/performance/2h5aszb7",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-2h5aszb7-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 10.5 GB at this context; duplicate traces of the same session: 3t9c3eg4, cs3pu9sb, kluxjwgd, zgdfk7yi; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 496.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 16k 496.9 19.7 10.5 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2h5aszb7",
   "source_url": "https://omlx.ai/benchmarks/performance/2h5aszb7",
   "tg_tokens": null,
   "tps": 19.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-2h5aszb7-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 10.8 GB at this context; duplicate traces of the same session: 3t9c3eg4, cs3pu9sb, kluxjwgd, zgdfk7yi; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 449.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 32k 449.9 16.9 10.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2h5aszb7",
   "source_url": "https://omlx.ai/benchmarks/performance/2h5aszb7",
   "tg_tokens": null,
   "tps": 16.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-jyjg48f5-ctx1024",
   "model": "Mamba-Codestral-7B-v0.1",
   "notes": "oMLX community benchmark trace; peak memory 15.2 GB at this context; duplicate traces of the same session: iq3ptvhy, yv00uh8d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 598.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Mamba-Codestral-7B-v0.1-8bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 598.4 52.0 15.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jyjg48f5",
   "source_url": "https://omlx.ai/benchmarks/performance/jyjg48f5",
   "tg_tokens": null,
   "tps": 52.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-jyjg48f5-ctx4096",
   "model": "Mamba-Codestral-7B-v0.1",
   "notes": "oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: iq3ptvhy, yv00uh8d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 579.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Mamba-Codestral-7B-v0.1-8bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 579.9 52.8 19.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jyjg48f5",
   "source_url": "https://omlx.ai/benchmarks/performance/jyjg48f5",
   "tg_tokens": null,
   "tps": 52.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-jyjg48f5-ctx8192",
   "model": "Mamba-Codestral-7B-v0.1",
   "notes": "oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: iq3ptvhy, yv00uh8d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 517.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Mamba-Codestral-7B-v0.1-8bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 8k 517.7 52.6 19.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jyjg48f5",
   "source_url": "https://omlx.ai/benchmarks/performance/jyjg48f5",
   "tg_tokens": null,
   "tps": 52.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-lvjlxh93-ctx1024",
   "model": "MiniCPM5-1B",
   "notes": "OptiQ build; oMLX community benchmark trace; peak memory 1.5 GB at this context; duplicate traces of the same session: egwax4q5; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2122.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM5-1B-OptiQ-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k 2,122 117.6 1.5 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lvjlxh93",
   "source_url": "https://omlx.ai/benchmarks/performance/lvjlxh93",
   "tg_tokens": null,
   "tps": 117.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-lvjlxh93-ctx4096",
   "model": "MiniCPM5-1B",
   "notes": "OptiQ build; oMLX community benchmark trace; peak memory 1.6 GB at this context; duplicate traces of the same session: egwax4q5; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1784.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM5-1B-OptiQ-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 1,784 102.0 1.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lvjlxh93",
   "source_url": "https://omlx.ai/benchmarks/performance/lvjlxh93",
   "tg_tokens": null,
   "tps": 102.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-iqa4hgju-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: gitlrxuv; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2123.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 2,123 79.7 21.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iqa4hgju",
   "source_url": "https://omlx.ai/benchmarks/performance/iqa4hgju",
   "tg_tokens": null,
   "tps": 79.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-iqa4hgju-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: gitlrxuv; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2712.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 2,712 77.6 22.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iqa4hgju",
   "source_url": "https://omlx.ai/benchmarks/performance/iqa4hgju",
   "tg_tokens": null,
   "tps": 77.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-nf0zti83-ctx1024",
   "model": "Qwen3.5-2B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 1k context; peak memory 1.9 GB; TTFT 247 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4154.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-2B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k tokens 4,154 PP tok/s 126.3 TG tok/s 247 TTFT (ms) 1.9 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nf0zti83",
   "source_url": "https://omlx.ai/benchmarks/performance/nf0zti83",
   "tg_tokens": null,
   "tps": 126.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-xtmk31z6-ctx32768",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 7.4 GB; TTFT 92515 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 354.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-4B-MLX-8bit on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 8bit; 32k tokens 354.2 PP tok/s 36.2 TG tok/s 92515 TTFT (ms) 7.4 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xtmk31z6",
   "source_url": "https://omlx.ai/benchmarks/performance/xtmk31z6",
   "tg_tokens": null,
   "tps": 36.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-7t1ygpav-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 5.3 GB at this context; duplicate traces of the same session: gz6xlds3; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 501.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-4B-MLX-8bit on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 8bit; 1k 501.9 47.3 5.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7t1ygpav",
   "source_url": "https://omlx.ai/benchmarks/performance/7t1ygpav",
   "tg_tokens": null,
   "tps": 47.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-7t1ygpav-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: gz6xlds3; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 490.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-4B-MLX-8bit on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 8bit; 4k 490.5 46.1 5.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7t1ygpav",
   "source_url": "https://omlx.ai/benchmarks/performance/7t1ygpav",
   "tg_tokens": null,
   "tps": 46.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-03hqh5dy-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: ok1r051l; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 667.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 667.8 70.1 19.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/03hqh5dy",
   "source_url": "https://omlx.ai/benchmarks/performance/03hqh5dy",
   "tg_tokens": null,
   "tps": 70.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-03hqh5dy-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: ok1r051l; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1454.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,454 70.5 19.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/03hqh5dy",
   "source_url": "https://omlx.ai/benchmarks/performance/03hqh5dy",
   "tg_tokens": null,
   "tps": 70.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-27607k88-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: j7kigtw8, ux6z5r0s, vthlhs4v; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 536.0,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Qwen3.8-27B-mxfp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 536.0 18.5 27.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/27607k88",
   "source_url": "https://omlx.ai/benchmarks/performance/27607k88",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-27607k88-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.9 GB at this context; duplicate traces of the same session: j7kigtw8, ux6z5r0s, vthlhs4v; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 620.8,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Qwen3.8-27B-mxfp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 620.8 18.6 28.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/27607k88",
   "source_url": "https://omlx.ai/benchmarks/performance/27607k88",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-27607k88-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 30.5 GB at this context; duplicate traces of the same session: j7kigtw8, ux6z5r0s, vthlhs4v; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 617.7,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Qwen3.8-27B-mxfp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 617.7 18.2 30.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/27607k88",
   "source_url": "https://omlx.ai/benchmarks/performance/27607k88",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-27607k88-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.6 GB at this context; duplicate traces of the same session: j7kigtw8, ux6z5r0s, vthlhs4v; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 631.8,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Qwen3.8-27B-mxfp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 631.8 16.6 32.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/27607k88",
   "source_url": "https://omlx.ai/benchmarks/performance/27607k88",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-s0dapan3-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAM-288",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 42.1 GB at this context; duplicate traces of the same session: aa52ca4n, uzh8f0k1, y6vyb7ag; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 494.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-REAM-288-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 494.6 57.0 42.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s0dapan3",
   "source_url": "https://omlx.ai/benchmarks/performance/s0dapan3",
   "tg_tokens": null,
   "tps": 57.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-s0dapan3-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAM-288",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 42.7 GB at this context; duplicate traces of the same session: aa52ca4n, uzh8f0k1, y6vyb7ag; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 612.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-REAM-288-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 612.3 61.7 42.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s0dapan3",
   "source_url": "https://omlx.ai/benchmarks/performance/s0dapan3",
   "tg_tokens": null,
   "tps": 61.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-s0dapan3-ctx8192",
   "model": "Qwen3.8-Flash-Next-REAM-288",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 42.9 GB at this context; duplicate traces of the same session: aa52ca4n, uzh8f0k1, y6vyb7ag; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 733.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-REAM-288-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 733.6 54.9 42.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s0dapan3",
   "source_url": "https://omlx.ai/benchmarks/performance/s0dapan3",
   "tg_tokens": null,
   "tps": 54.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-s0dapan3-ctx16384",
   "model": "Qwen3.8-Flash-Next-REAM-288",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.2 GB at this context; duplicate traces of the same session: aa52ca4n, uzh8f0k1, y6vyb7ag; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 702.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-REAM-288-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 702.1 59.0 43.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s0dapan3",
   "source_url": "https://omlx.ai/benchmarks/performance/s0dapan3",
   "tg_tokens": null,
   "tps": 59.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-5rgat57p-ctx16384",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 40.8 GB; TTFT 58035 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 282.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k tokens 282.3 PP tok/s 34.4 TG tok/s 58035 TTFT (ms) 40.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5rgat57p",
   "source_url": "https://omlx.ai/benchmarks/performance/5rgat57p",
   "tg_tokens": null,
   "tps": 34.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-bgmoame1-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6 mixed-quant MTP v2 build; oMLX community benchmark trace; peak memory 77.6 GB at this context; duplicate traces of the same session: svrl5b7k; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 429.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp-v2 on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 6bit; 1k 429.2 46.9 77.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bgmoame1",
   "source_url": "https://omlx.ai/benchmarks/performance/bgmoame1",
   "tg_tokens": null,
   "tps": 46.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-bgmoame1-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6 mixed-quant MTP v2 build; oMLX community benchmark trace; peak memory 78.3 GB at this context; duplicate traces of the same session: svrl5b7k; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 611.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp-v2 on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 6bit; 8k 611.3 50.3 78.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bgmoame1",
   "source_url": "https://omlx.ai/benchmarks/performance/bgmoame1",
   "tg_tokens": null,
   "tps": 50.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-5q0xktj3-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6 mixed-quant MTP v2 build; oMLX community benchmark trace; peak memory 77.5 GB at this context; duplicate traces of the same session: btnsmy8d, ypdg9v1q; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 398.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp-v2 on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 6bit; 1k 398.5 44.9 77.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5q0xktj3",
   "source_url": "https://omlx.ai/benchmarks/performance/5q0xktj3",
   "tg_tokens": null,
   "tps": 44.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-5q0xktj3-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6 mixed-quant MTP v2 build; oMLX community benchmark trace; peak memory 78.3 GB at this context; duplicate traces of the same session: btnsmy8d, ypdg9v1q; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 611.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp-v2 on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 6bit; 8k 611.9 45.8 78.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5q0xktj3",
   "source_url": "https://omlx.ai/benchmarks/performance/5q0xktj3",
   "tg_tokens": null,
   "tps": 45.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-5q0xktj3-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6 mixed-quant MTP v2 build; oMLX community benchmark trace; peak memory 78.6 GB at this context; duplicate traces of the same session: btnsmy8d, ypdg9v1q; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 611.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp-v2 on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 6bit; 16k 611.0 50.5 78.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5q0xktj3",
   "source_url": "https://omlx.ai/benchmarks/performance/5q0xktj3",
   "tg_tokens": null,
   "tps": 50.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xfe7xx5l-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: 852bzvtl, cr08s4xe; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1025.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,025 64.3 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xfe7xx5l",
   "source_url": "https://omlx.ai/benchmarks/performance/xfe7xx5l",
   "tg_tokens": null,
   "tps": 64.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xfe7xx5l-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: 852bzvtl, cr08s4xe; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2040.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,040 70.9 72.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xfe7xx5l",
   "source_url": "https://omlx.ai/benchmarks/performance/xfe7xx5l",
   "tg_tokens": null,
   "tps": 70.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-xfe7xx5l-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.6 GB at this context; duplicate traces of the same session: 852bzvtl, cr08s4xe; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2195.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,195 84.5 73.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xfe7xx5l",
   "source_url": "https://omlx.ai/benchmarks/performance/xfe7xx5l",
   "tg_tokens": null,
   "tps": 84.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-t4j7pthw-ctx1024",
   "model": "Ternary-Bonsai-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 9.1 GB at this context; duplicate traces of the same session: 1j7mng1m; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 312.5,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-27B-mlx-2bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 2bit; 1k 312.5 43.7 9.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t4j7pthw",
   "source_url": "https://omlx.ai/benchmarks/performance/t4j7pthw",
   "tg_tokens": null,
   "tps": 43.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-t4j7pthw-ctx4096",
   "model": "Ternary-Bonsai-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 12.4 GB at this context; duplicate traces of the same session: 1j7mng1m; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 337.8,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-27B-mlx-2bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 2bit; 4k 337.8 42.7 12.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t4j7pthw",
   "source_url": "https://omlx.ai/benchmarks/performance/t4j7pthw",
   "tg_tokens": null,
   "tps": 42.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-fu3g1y3d-ctx1024",
   "model": "4-bit",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: 92pxzu9o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 247.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "4-bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 247.2 28.0 15.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fu3g1y3d",
   "source_url": "https://omlx.ai/benchmarks/performance/fu3g1y3d",
   "tg_tokens": null,
   "tps": 28.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-fu3g1y3d-ctx4096",
   "model": "4-bit",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 92pxzu9o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 240.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "4-bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 240.2 25.8 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fu3g1y3d",
   "source_url": "https://omlx.ai/benchmarks/performance/fu3g1y3d",
   "tg_tokens": null,
   "tps": 25.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-l0992o5o-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.0 GB at this context; duplicate traces of the same session: 1nb2ipb3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 322.6,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 322.6 11.4 92.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l0992o5o",
   "source_url": "https://omlx.ai/benchmarks/performance/l0992o5o",
   "tg_tokens": null,
   "tps": 11.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-l0992o5o-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.2 GB at this context; duplicate traces of the same session: 1nb2ipb3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 312.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 312.0 12.3 92.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l0992o5o",
   "source_url": "https://omlx.ai/benchmarks/performance/l0992o5o",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qahwwwy3-ctx32768",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 92.5 GB; TTFT 110679 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 296.1,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 32k tokens 296.1 PP tok/s 11.9 TG tok/s 110679 TTFT (ms) 92.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qahwwwy3",
   "source_url": "https://omlx.ai/benchmarks/performance/qahwwwy3",
   "tg_tokens": null,
   "tps": 11.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-j0m5e0lt-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 90.0 GB at this context; duplicate traces of the same session: 1gm9rig6, jn8jrcot; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 85.6,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 85.6 10.5 90.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j0m5e0lt",
   "source_url": "https://omlx.ai/benchmarks/performance/j0m5e0lt",
   "tg_tokens": null,
   "tps": 10.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-j0m5e0lt-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.0 GB at this context; duplicate traces of the same session: 1gm9rig6, jn8jrcot; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 311.5,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 311.5 11.1 92.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j0m5e0lt",
   "source_url": "https://omlx.ai/benchmarks/performance/j0m5e0lt",
   "tg_tokens": null,
   "tps": 11.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-j0m5e0lt-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.2 GB at this context; duplicate traces of the same session: 1gm9rig6, jn8jrcot; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 326.3,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 326.3 12.1 92.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j0m5e0lt",
   "source_url": "https://omlx.ai/benchmarks/performance/j0m5e0lt",
   "tg_tokens": null,
   "tps": 12.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-466tum7a-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 90.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 88.5,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 88.5 10.1 90.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/466tum7a",
   "source_url": "https://omlx.ai/benchmarks/performance/466tum7a",
   "tg_tokens": null,
   "tps": 10.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-466tum7a-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 278.5,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 278.5 11.2 92.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/466tum7a",
   "source_url": "https://omlx.ai/benchmarks/performance/466tum7a",
   "tg_tokens": null,
   "tps": 11.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-466tum7a-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 332.1,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 332.1 11.8 92.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/466tum7a",
   "source_url": "https://omlx.ai/benchmarks/performance/466tum7a",
   "tg_tokens": null,
   "tps": 11.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-3ri6ao7l-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.5 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 449.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 449.5 30.8 173.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ri6ao7l",
   "source_url": "https://omlx.ai/benchmarks/performance/3ri6ao7l",
   "tg_tokens": null,
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-3ri6ao7l-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 545.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 545.5 24.7 174.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ri6ao7l",
   "source_url": "https://omlx.ai/benchmarks/performance/3ri6ao7l",
   "tg_tokens": null,
   "tps": 24.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-kwtufsvf-ctx8192",
   "model": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: 97il9l4l, k5m5kmnq, tjqvjucn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 993.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated-MLX-6bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 8k 993.4 66.2 29.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kwtufsvf",
   "source_url": "https://omlx.ai/benchmarks/performance/kwtufsvf",
   "tg_tokens": null,
   "tps": 66.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-kwtufsvf-ctx16384",
   "model": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 30.0 GB at this context; duplicate traces of the same session: 97il9l4l, k5m5kmnq, tjqvjucn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 893.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated-MLX-6bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 16k 893.6 60.7 30.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kwtufsvf",
   "source_url": "https://omlx.ai/benchmarks/performance/kwtufsvf",
   "tg_tokens": null,
   "tps": 60.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-kwtufsvf-ctx32768",
   "model": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 31.0 GB at this context; duplicate traces of the same session: 97il9l4l, k5m5kmnq, tjqvjucn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 719.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated-MLX-6bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 32k 719.1 53.4 31.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kwtufsvf",
   "source_url": "https://omlx.ai/benchmarks/performance/kwtufsvf",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-kwtufsvf-ctx65536",
   "model": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 32.3 GB at this context; duplicate traces of the same session: 97il9l4l, k5m5kmnq, tjqvjucn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 540.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated-MLX-6bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 64k 540.0 42.4 32.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kwtufsvf",
   "source_url": "https://omlx.ai/benchmarks/performance/kwtufsvf",
   "tg_tokens": null,
   "tps": 42.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-vwkizztt-ctx1024",
   "model": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: 3zo71hfj, zqjr5nrq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 866.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated-MLX-6bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 866.1 71.1 27.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vwkizztt",
   "source_url": "https://omlx.ai/benchmarks/performance/vwkizztt",
   "tg_tokens": null,
   "tps": 71.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-vwkizztt-ctx4096",
   "model": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 29.3 GB at this context; duplicate traces of the same session: 3zo71hfj, zqjr5nrq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1050.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated-MLX-6bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 1,050 69.1 29.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vwkizztt",
   "source_url": "https://omlx.ai/benchmarks/performance/vwkizztt",
   "tg_tokens": null,
   "tps": 69.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-vwkizztt-ctx8192",
   "model": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: 3zo71hfj, zqjr5nrq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 995.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated-MLX-6bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 8k 995.7 66.1 29.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vwkizztt",
   "source_url": "https://omlx.ai/benchmarks/performance/vwkizztt",
   "tg_tokens": null,
   "tps": 66.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-1um6y4a5-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 10.1 GB at this context; duplicate traces of the same session: 4q78jk7m, enowxgi8, mof5grct; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 340.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 1k 340.6 14.5 10.1 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1um6y4a5",
   "source_url": "https://omlx.ai/benchmarks/performance/1um6y4a5",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-1um6y4a5-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 10.6 GB at this context; duplicate traces of the same session: 4q78jk7m, enowxgi8, mof5grct; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 515.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 4k 515.5 18.5 10.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1um6y4a5",
   "source_url": "https://omlx.ai/benchmarks/performance/1um6y4a5",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-1um6y4a5-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 10.7 GB at this context; duplicate traces of the same session: 4q78jk7m, enowxgi8, mof5grct; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 520.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 8k 520.6 16.6 10.7 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1um6y4a5",
   "source_url": "https://omlx.ai/benchmarks/performance/1um6y4a5",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-1um6y4a5-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 10.8 GB at this context; duplicate traces of the same session: 4q78jk7m, enowxgi8, mof5grct; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 497.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 16k 497.9 19.2 10.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1um6y4a5",
   "source_url": "https://omlx.ai/benchmarks/performance/1um6y4a5",
   "tg_tokens": null,
   "tps": 19.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-mkq6iefd-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 7.2 GB at this context; duplicate traces of the same session: lkzwhdh4, rvqxzeap, z9jtgosf; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 337.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 1k 337.3 7.7 7.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mkq6iefd",
   "source_url": "https://omlx.ai/benchmarks/performance/mkq6iefd",
   "tg_tokens": null,
   "tps": 7.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-mkq6iefd-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 7.6 GB at this context; duplicate traces of the same session: lkzwhdh4, rvqxzeap, z9jtgosf; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 463.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 4k 463.9 14.2 7.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mkq6iefd",
   "source_url": "https://omlx.ai/benchmarks/performance/mkq6iefd",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-mkq6iefd-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 7.7 GB at this context; duplicate traces of the same session: lkzwhdh4, rvqxzeap, z9jtgosf; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 463.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 8k 463.9 11.0 7.7 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mkq6iefd",
   "source_url": "https://omlx.ai/benchmarks/performance/mkq6iefd",
   "tg_tokens": null,
   "tps": 11.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-mkq6iefd-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 7.8 GB at this context; duplicate traces of the same session: lkzwhdh4, rvqxzeap, z9jtgosf; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 444.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 16k 444.8 15.5 7.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mkq6iefd",
   "source_url": "https://omlx.ai/benchmarks/performance/mkq6iefd",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-krmmzjr8-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 4.1 GB at this context; duplicate traces of the same session: 2mz0icxb, nugvxw6l, shvla74l; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 265.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 1k 265.6 8.9 4.1 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/krmmzjr8",
   "source_url": "https://omlx.ai/benchmarks/performance/krmmzjr8",
   "tg_tokens": null,
   "tps": 8.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-krmmzjr8-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 4.6 GB at this context; duplicate traces of the same session: 2mz0icxb, nugvxw6l, shvla74l; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 425.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 4k 425.1 10.9 4.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/krmmzjr8",
   "source_url": "https://omlx.ai/benchmarks/performance/krmmzjr8",
   "tg_tokens": null,
   "tps": 10.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-krmmzjr8-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 4.7 GB at this context; duplicate traces of the same session: 2mz0icxb, nugvxw6l, shvla74l; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 426.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 8k 426.0 10.6 4.7 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/krmmzjr8",
   "source_url": "https://omlx.ai/benchmarks/performance/krmmzjr8",
   "tg_tokens": null,
   "tps": 10.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-krmmzjr8-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 4.8 GB at this context; duplicate traces of the same session: 2mz0icxb, nugvxw6l, shvla74l; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 411.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 16k 411.6 9.0 4.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/krmmzjr8",
   "source_url": "https://omlx.ai/benchmarks/performance/krmmzjr8",
   "tg_tokens": null,
   "tps": 9.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9i1n3d7b-ctx1024",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 156.9 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 573.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "MiMo-V2.6-Flash-RL-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 573.2 44.7 156.9 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9i1n3d7b",
   "source_url": "https://omlx.ai/benchmarks/performance/9i1n3d7b",
   "tg_tokens": null,
   "tps": 44.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9i1n3d7b-ctx4096",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 158.2 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 644.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "MiMo-V2.6-Flash-RL-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 644.6 44.1 158.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9i1n3d7b",
   "source_url": "https://omlx.ai/benchmarks/performance/9i1n3d7b",
   "tg_tokens": null,
   "tps": 44.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-859zqkgm-ctx65536",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 16.9 GB; TTFT 601822 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 108.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k tokens 108.9 PP tok/s 13.4 TG tok/s 601822 TTFT (ms) 16.9 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/859zqkgm",
   "source_url": "https://omlx.ai/benchmarks/performance/859zqkgm",
   "tg_tokens": null,
   "tps": 13.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-smgqt2r2-ctx1024",
   "model": "Qwen3 32B",
   "notes": "oMLX community benchmark trace; peak memory 18.0 GB at this context; duplicate traces of the same session: 67sciqin, qz9fwo3x; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-32B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 323.0 14.6 18.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/smgqt2r2",
   "source_url": "https://omlx.ai/benchmarks/performance/smgqt2r2",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-smgqt2r2-ctx4096",
   "model": "Qwen3 32B",
   "notes": "oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: 67sciqin, qz9fwo3x; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 325.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-32B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 325.9 13.5 18.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/smgqt2r2",
   "source_url": "https://omlx.ai/benchmarks/performance/smgqt2r2",
   "tg_tokens": null,
   "tps": 13.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-smgqt2r2-ctx8192",
   "model": "Qwen3 32B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 67sciqin, qz9fwo3x; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 301.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-32B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 301.8 12.7 19.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/smgqt2r2",
   "source_url": "https://omlx.ai/benchmarks/performance/smgqt2r2",
   "tg_tokens": null,
   "tps": 12.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-xqj1aj28-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 65.6 GB; TTFT 623 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1645.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,645 PP tok/s 70.4 TG tok/s 623 TTFT (ms) 65.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xqj1aj28",
   "source_url": "https://omlx.ai/benchmarks/performance/xqj1aj28",
   "tg_tokens": null,
   "tps": 70.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-0sfttzag-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: iso11u9x; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 506.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 506.7 24.5 5.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0sfttzag",
   "source_url": "https://omlx.ai/benchmarks/performance/0sfttzag",
   "tg_tokens": null,
   "tps": 24.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-0sfttzag-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: iso11u9x; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 560.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 560.4 23.3 6.3 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0sfttzag",
   "source_url": "https://omlx.ai/benchmarks/performance/0sfttzag",
   "tg_tokens": null,
   "tps": 23.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-hiq1ojy2-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: kgseahh4; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 728.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 728.2 107.5 19.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hiq1ojy2",
   "source_url": "https://omlx.ai/benchmarks/performance/hiq1ojy2",
   "tg_tokens": null,
   "tps": 107.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-hiq1ojy2-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: kgseahh4; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2742.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 2,742 98.0 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hiq1ojy2",
   "source_url": "https://omlx.ai/benchmarks/performance/hiq1ojy2",
   "tg_tokens": null,
   "tps": 98.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-c0g25dru-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 16.0 GB; TTFT 7783 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k tokens 131.6 PP tok/s 15.4 TG tok/s 7783 TTFT (ms) 16.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c0g25dru",
   "source_url": "https://omlx.ai/benchmarks/performance/c0g25dru",
   "tg_tokens": null,
   "tps": 15.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6x4loi5o-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.8 GB at this context; duplicate traces of the same session: dckqzpr7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 925.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 925.0 33.1 28.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6x4loi5o",
   "source_url": "https://omlx.ai/benchmarks/performance/6x4loi5o",
   "tg_tokens": null,
   "tps": 33.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6x4loi5o-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.8 GB at this context; duplicate traces of the same session: dckqzpr7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1011.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,011 32.6 29.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6x4loi5o",
   "source_url": "https://omlx.ai/benchmarks/performance/6x4loi5o",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Pro",
   "id": "omlx-i7kcyn68-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: ybls7xg8; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 95.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 1k 95.6 15.1 17.6 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i7kcyn68",
   "source_url": "https://omlx.ai/benchmarks/performance/i7kcyn68",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Pro",
   "id": "omlx-i7kcyn68-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: ybls7xg8; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 96.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 4k 96.7 15.9 19.3 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i7kcyn68",
   "source_url": "https://omlx.ai/benchmarks/performance/i7kcyn68",
   "tg_tokens": null,
   "tps": 15.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Pro",
   "id": "omlx-7afmptsp-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: 9ehf2nib; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 101.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 1k 101.5 8.1 16.4 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7afmptsp",
   "source_url": "https://omlx.ai/benchmarks/performance/7afmptsp",
   "tg_tokens": null,
   "tps": 8.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Pro",
   "id": "omlx-7afmptsp-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 9ehf2nib; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 4k 104.2 7.8 17.6 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7afmptsp",
   "source_url": "https://omlx.ai/benchmarks/performance/7afmptsp",
   "tg_tokens": null,
   "tps": 7.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-b7zc7ssx-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.9 GB at this context; duplicate traces of the same session: 4s5x3zcq, 92dw2nzt, iuadjz5t; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2327.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 2,327 138.3 142.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b7zc7ssx",
   "source_url": "https://omlx.ai/benchmarks/performance/b7zc7ssx",
   "tg_tokens": null,
   "tps": 138.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-b7zc7ssx-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.3 GB at this context; duplicate traces of the same session: 4s5x3zcq, 92dw2nzt, iuadjz5t; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4115.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 4,115 121.7 143.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b7zc7ssx",
   "source_url": "https://omlx.ai/benchmarks/performance/b7zc7ssx",
   "tg_tokens": null,
   "tps": 121.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-b7zc7ssx-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 146.9 GB at this context; duplicate traces of the same session: 4s5x3zcq, 92dw2nzt, iuadjz5t; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4588.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 64k 4,588 136.3 146.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b7zc7ssx",
   "source_url": "https://omlx.ai/benchmarks/performance/b7zc7ssx",
   "tg_tokens": null,
   "tps": 136.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-b7zc7ssx-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 150.9 GB at this context; duplicate traces of the same session: 4s5x3zcq, 92dw2nzt, iuadjz5t; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4469.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 128k 4,469 108.0 150.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b7zc7ssx",
   "source_url": "https://omlx.ai/benchmarks/performance/b7zc7ssx",
   "tg_tokens": null,
   "tps": 108.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-eiybpwmi-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.0 GB at this context; duplicate traces of the same session: dz3l035d, hortzthq, ujq6kml6; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 873.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 1k 873.7 111.8 183.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eiybpwmi",
   "source_url": "https://omlx.ai/benchmarks/performance/eiybpwmi",
   "tg_tokens": null,
   "tps": 111.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-eiybpwmi-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.7 GB at this context; duplicate traces of the same session: dz3l035d, hortzthq, ujq6kml6; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2674.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 4k 2,674 97.3 183.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eiybpwmi",
   "source_url": "https://omlx.ai/benchmarks/performance/eiybpwmi",
   "tg_tokens": null,
   "tps": 97.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-eiybpwmi-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 186.9 GB at this context; duplicate traces of the same session: dz3l035d, hortzthq, ujq6kml6; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5243.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 64k 5,243 106.2 186.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eiybpwmi",
   "source_url": "https://omlx.ai/benchmarks/performance/eiybpwmi",
   "tg_tokens": null,
   "tps": 106.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-eiybpwmi-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 191.3 GB at this context; duplicate traces of the same session: dz3l035d, hortzthq, ujq6kml6; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5084.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 128k 5,084 100.9 191.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eiybpwmi",
   "source_url": "https://omlx.ai/benchmarks/performance/eiybpwmi",
   "tg_tokens": null,
   "tps": 100.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-q4s0mg02-ctx1024",
   "model": "Gemma 4 12B",
   "notes": "oMLX community benchmark trace; peak memory 12.8 GB at this context; duplicate traces of the same session: ypafvimg; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 210.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-12B-it-8bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 8bit; 1k 210.9 10.1 12.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q4s0mg02",
   "source_url": "https://omlx.ai/benchmarks/performance/q4s0mg02",
   "tg_tokens": null,
   "tps": 10.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-q4s0mg02-ctx4096",
   "model": "Gemma 4 12B",
   "notes": "oMLX community benchmark trace; peak memory 13.2 GB at this context; duplicate traces of the same session: ypafvimg; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 215.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-12B-it-8bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 8bit; 4k 215.2 10.1 13.2 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q4s0mg02",
   "source_url": "https://omlx.ai/benchmarks/performance/q4s0mg02",
   "tg_tokens": null,
   "tps": 10.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-hieoyu3e-ctx1024",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 7.4 GB at this context; duplicate traces of the same session: 44xmflob; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 221.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-MLX-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 1k 221.1 17.3 7.4 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hieoyu3e",
   "source_url": "https://omlx.ai/benchmarks/performance/hieoyu3e",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-hieoyu3e-ctx4096",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 7.7 GB at this context; duplicate traces of the same session: 44xmflob; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 217.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-MLX-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 4k 217.0 17.9 7.7 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hieoyu3e",
   "source_url": "https://omlx.ai/benchmarks/performance/hieoyu3e",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-osejewm0-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 90.0 GB at this context; duplicate traces of the same session: nts1vll8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 88.5,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 88.5 10.1 90.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/osejewm0",
   "source_url": "https://omlx.ai/benchmarks/performance/osejewm0",
   "tg_tokens": null,
   "tps": 10.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-osejewm0-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.2 GB at this context; duplicate traces of the same session: nts1vll8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 278.5,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 278.5 11.2 92.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/osejewm0",
   "source_url": "https://omlx.ai/benchmarks/performance/osejewm0",
   "tg_tokens": null,
   "tps": 11.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-osejewm0-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.3 GB at this context; duplicate traces of the same session: nts1vll8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 332.1,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 332.1 11.8 92.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/osejewm0",
   "source_url": "https://omlx.ai/benchmarks/performance/osejewm0",
   "tg_tokens": null,
   "tps": 11.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kzmrmcla-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.2 GB at this context; duplicate traces of the same session: nnwi7tog; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 261.6,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 261.6 10.8 92.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kzmrmcla",
   "source_url": "https://omlx.ai/benchmarks/performance/kzmrmcla",
   "tg_tokens": null,
   "tps": 10.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kzmrmcla-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.3 GB at this context; duplicate traces of the same session: nnwi7tog; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 292.6,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 292.6 12.9 92.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kzmrmcla",
   "source_url": "https://omlx.ai/benchmarks/performance/kzmrmcla",
   "tg_tokens": null,
   "tps": 12.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-r61k8tpc-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 90.0 GB at this context; duplicate traces of the same session: j4sw8bun, olza0j1m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 84.1,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 84.1 8.9 90.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r61k8tpc",
   "source_url": "https://omlx.ai/benchmarks/performance/r61k8tpc",
   "tg_tokens": null,
   "tps": 8.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-r61k8tpc-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.0 GB at this context; duplicate traces of the same session: j4sw8bun, olza0j1m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 197.3,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 197.3 9.1 92.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r61k8tpc",
   "source_url": "https://omlx.ai/benchmarks/performance/r61k8tpc",
   "tg_tokens": null,
   "tps": 9.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-r61k8tpc-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.3 GB at this context; duplicate traces of the same session: j4sw8bun, olza0j1m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 275.9,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 275.9 10.7 92.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r61k8tpc",
   "source_url": "https://omlx.ai/benchmarks/performance/r61k8tpc",
   "tg_tokens": null,
   "tps": 10.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5qx5eu8i-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 90.0 GB at this context; duplicate traces of the same session: cqt8uu47, cr58q0w1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 85.8,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 85.8 10.6 90.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5qx5eu8i",
   "source_url": "https://omlx.ai/benchmarks/performance/5qx5eu8i",
   "tg_tokens": null,
   "tps": 10.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5qx5eu8i-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.0 GB at this context; duplicate traces of the same session: cqt8uu47, cr58q0w1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 191.2,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 191.2 10.6 92.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5qx5eu8i",
   "source_url": "https://omlx.ai/benchmarks/performance/5qx5eu8i",
   "tg_tokens": null,
   "tps": 10.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5qx5eu8i-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.3 GB at this context; duplicate traces of the same session: cqt8uu47, cr58q0w1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 161.1,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 161.1 10.6 92.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5qx5eu8i",
   "source_url": "https://omlx.ai/benchmarks/performance/5qx5eu8i",
   "tg_tokens": null,
   "tps": 10.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1s8qjer2-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 90.0 GB at this context; duplicate traces of the same session: 7i7q2mya, jagl7nsk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 87.4,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 87.4 11.4 90.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1s8qjer2",
   "source_url": "https://omlx.ai/benchmarks/performance/1s8qjer2",
   "tg_tokens": null,
   "tps": 11.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1s8qjer2-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.2 GB at this context; duplicate traces of the same session: 7i7q2mya, jagl7nsk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 257.3,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 257.3 10.4 92.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1s8qjer2",
   "source_url": "https://omlx.ai/benchmarks/performance/1s8qjer2",
   "tg_tokens": null,
   "tps": 10.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1s8qjer2-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.3 GB at this context; duplicate traces of the same session: 7i7q2mya, jagl7nsk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 369.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 369.0 11.7 92.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1s8qjer2",
   "source_url": "https://omlx.ai/benchmarks/performance/1s8qjer2",
   "tg_tokens": null,
   "tps": 11.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-zbz91yld-ctx1024",
   "model": "GLM-5.3-Flash-Abliterated",
   "notes": "MLX; oMLX community benchmark trace; peak memory 177.6 GB at this context; duplicate traces of the same session: kkp0y6tv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 520.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-Abliterated-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 520.3 49.9 177.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zbz91yld",
   "source_url": "https://omlx.ai/benchmarks/performance/zbz91yld",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-zbz91yld-ctx32768",
   "model": "GLM-5.3-Flash-Abliterated",
   "notes": "MLX; oMLX community benchmark trace; peak memory 179.7 GB at this context; duplicate traces of the same session: kkp0y6tv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 524.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-Abliterated-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 524.4 60.4 179.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zbz91yld",
   "source_url": "https://omlx.ai/benchmarks/performance/zbz91yld",
   "tg_tokens": null,
   "tps": 60.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-hw3md8d1-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.6 GB at this context; duplicate traces of the same session: 2g2bt2ax, 6twkhkb5, 7ovzs349, syv3mejv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2365.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 2,365 166.3 174.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hw3md8d1",
   "source_url": "https://omlx.ai/benchmarks/performance/hw3md8d1",
   "tg_tokens": null,
   "tps": 166.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-hw3md8d1-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.3 GB at this context; duplicate traces of the same session: 2g2bt2ax, 6twkhkb5, 7ovzs349, syv3mejv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2371.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,371 123.1 175.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hw3md8d1",
   "source_url": "https://omlx.ai/benchmarks/performance/hw3md8d1",
   "tg_tokens": null,
   "tps": 123.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-hw3md8d1-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.1 GB at this context; duplicate traces of the same session: 2g2bt2ax, 6twkhkb5, 7ovzs349, syv3mejv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2340.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,340 99.3 177.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hw3md8d1",
   "source_url": "https://omlx.ai/benchmarks/performance/hw3md8d1",
   "tg_tokens": null,
   "tps": 99.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-hw3md8d1-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 180.5 GB at this context; duplicate traces of the same session: 2g2bt2ax, 6twkhkb5, 7ovzs349, syv3mejv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2293.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,293 143.5 180.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hw3md8d1",
   "source_url": "https://omlx.ai/benchmarks/performance/hw3md8d1",
   "tg_tokens": null,
   "tps": 143.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-hw3md8d1-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 184.1 GB at this context; duplicate traces of the same session: 2g2bt2ax, 6twkhkb5, 7ovzs349, syv3mejv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2235.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 2,235 81.5 184.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hw3md8d1",
   "source_url": "https://omlx.ai/benchmarks/performance/hw3md8d1",
   "tg_tokens": null,
   "tps": 81.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-axtr1mc8-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.1 GB at this context; duplicate traces of the same session: uhh24ct8; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 503.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 4bit; 1k 503.1 52.3 13.1 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/axtr1mc8",
   "source_url": "https://omlx.ai/benchmarks/performance/axtr1mc8",
   "tg_tokens": null,
   "tps": 52.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-axtr1mc8-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.5 GB at this context; duplicate traces of the same session: uhh24ct8; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 536.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 4bit; 4k 536.6 50.8 13.5 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/axtr1mc8",
   "source_url": "https://omlx.ai/benchmarks/performance/axtr1mc8",
   "tg_tokens": null,
   "tps": 50.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-axtr1mc8-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.6 GB at this context; duplicate traces of the same session: uhh24ct8; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 510.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 4bit; 8k 510.6 48.1 13.6 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/axtr1mc8",
   "source_url": "https://omlx.ai/benchmarks/performance/axtr1mc8",
   "tg_tokens": null,
   "tps": 48.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-axtr1mc8-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.8 GB at this context; duplicate traces of the same session: uhh24ct8; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 447.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 4bit; 16k 447.1 41.9 13.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/axtr1mc8",
   "source_url": "https://omlx.ai/benchmarks/performance/axtr1mc8",
   "tg_tokens": null,
   "tps": 41.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-axtr1mc8-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 14.1 GB at this context; duplicate traces of the same session: uhh24ct8; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 255.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 4bit; 32k 255.1 24.8 14.1 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/axtr1mc8",
   "source_url": "https://omlx.ai/benchmarks/performance/axtr1mc8",
   "tg_tokens": null,
   "tps": 24.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qf7v16et-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: jr0ddw9t, n5rymox6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3536.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 3,536 104.2 13.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qf7v16et",
   "source_url": "https://omlx.ai/benchmarks/performance/qf7v16et",
   "tg_tokens": null,
   "tps": 104.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qf7v16et-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: jr0ddw9t, n5rymox6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3793.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,793 104.2 13.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qf7v16et",
   "source_url": "https://omlx.ai/benchmarks/performance/qf7v16et",
   "tg_tokens": null,
   "tps": 104.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qf7v16et-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.8 GB at this context; duplicate traces of the same session: jr0ddw9t, n5rymox6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4068.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 4,068 99.5 13.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qf7v16et",
   "source_url": "https://omlx.ai/benchmarks/performance/qf7v16et",
   "tg_tokens": null,
   "tps": 99.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-zmjmzuh8-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: 8eg8omf7, gfy0gsak; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3850.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 3,850 118.5 13.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zmjmzuh8",
   "source_url": "https://omlx.ai/benchmarks/performance/zmjmzuh8",
   "tg_tokens": null,
   "tps": 118.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-zmjmzuh8-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: 8eg8omf7, gfy0gsak; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4273.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 4,273 114.1 13.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zmjmzuh8",
   "source_url": "https://omlx.ai/benchmarks/performance/zmjmzuh8",
   "tg_tokens": null,
   "tps": 114.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-zmjmzuh8-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.8 GB at this context; duplicate traces of the same session: 8eg8omf7, gfy0gsak; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4271.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 4,271 120.5 13.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zmjmzuh8",
   "source_url": "https://omlx.ai/benchmarks/performance/zmjmzuh8",
   "tg_tokens": null,
   "tps": 120.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-25lmamfo-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: w30ul6ck; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3551.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 3,551 128.3 13.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/25lmamfo",
   "source_url": "https://omlx.ai/benchmarks/performance/25lmamfo",
   "tg_tokens": null,
   "tps": 128.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-25lmamfo-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: w30ul6ck; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3057.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,057 111.2 13.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/25lmamfo",
   "source_url": "https://omlx.ai/benchmarks/performance/25lmamfo",
   "tg_tokens": null,
   "tps": 111.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-dul11a5j-ctx1024",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 156.9 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 573.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "MiMo-V2.6-Flash-RL-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 573.2 44.7 156.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dul11a5j",
   "source_url": "https://omlx.ai/benchmarks/performance/dul11a5j",
   "tg_tokens": null,
   "tps": 44.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-dul11a5j-ctx4096",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 158.2 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 644.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "MiMo-V2.6-Flash-RL-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 644.6 44.1 158.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dul11a5j",
   "source_url": "https://omlx.ai/benchmarks/performance/dul11a5j",
   "tg_tokens": null,
   "tps": 44.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-dh2t8j8v-ctx1024",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 5.0 GB at this context; duplicate traces of the same session: qyk2mk5k; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 797.1,
   "provenance": "community",
   "quant": "bf16",
   "quote": "MiniCPM5-2B-bf16 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 bf16; 1k 797.1 21.1 5.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dh2t8j8v",
   "source_url": "https://omlx.ai/benchmarks/performance/dh2t8j8v",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-dh2t8j8v-ctx4096",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 5.3 GB at this context; duplicate traces of the same session: qyk2mk5k; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 760.9,
   "provenance": "community",
   "quant": "bf16",
   "quote": "MiniCPM5-2B-bf16 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 bf16; 4k 760.9 20.1 5.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dh2t8j8v",
   "source_url": "https://omlx.ai/benchmarks/performance/dh2t8j8v",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-tl00z1d2-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.0 GB at this context; duplicate traces of the same session: fopljfee; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 396.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-9B-MLX-8bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 8bit; 1k 396.2 37.8 10.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tl00z1d2",
   "source_url": "https://omlx.ai/benchmarks/performance/tl00z1d2",
   "tg_tokens": null,
   "tps": 37.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-tl00z1d2-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.5 GB at this context; duplicate traces of the same session: fopljfee; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 349.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-9B-MLX-8bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 8bit; 4k 349.3 31.6 10.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tl00z1d2",
   "source_url": "https://omlx.ai/benchmarks/performance/tl00z1d2",
   "tg_tokens": null,
   "tps": 31.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-wusej4or-ctx65536",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 14.5 GB at this context; duplicate traces of the same session: ca0d4q4n; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1657.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 64k 1,657 51.0 14.5 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wusej4or",
   "source_url": "https://omlx.ai/benchmarks/performance/wusej4or",
   "tg_tokens": null,
   "tps": 51.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-wusej4or-ctx131072",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: ca0d4q4n; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1103.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 128k 1,103 37.4 15.8 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wusej4or",
   "source_url": "https://omlx.ai/benchmarks/performance/wusej4or",
   "tg_tokens": null,
   "tps": 37.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-sv2lyle4-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: vgg7mdla; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2375.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 2,375 106.9 19.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sv2lyle4",
   "source_url": "https://omlx.ai/benchmarks/performance/sv2lyle4",
   "tg_tokens": null,
   "tps": 106.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-sv2lyle4-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: vgg7mdla; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2941.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,941 103.0 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sv2lyle4",
   "source_url": "https://omlx.ai/benchmarks/performance/sv2lyle4",
   "tg_tokens": null,
   "tps": 103.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5swf3c4m-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: dgo329bf, es5xgapp, gijhycy5, n9bk06xa, u70f9zcm; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2648.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 1k 2,648 169.2 29.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5swf3c4m",
   "source_url": "https://omlx.ai/benchmarks/performance/5swf3c4m",
   "tg_tokens": null,
   "tps": 169.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5swf3c4m-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.8 GB at this context; duplicate traces of the same session: dgo329bf, es5xgapp, gijhycy5, n9bk06xa, u70f9zcm; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3307.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 4k 3,307 155.7 31.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5swf3c4m",
   "source_url": "https://omlx.ai/benchmarks/performance/5swf3c4m",
   "tg_tokens": null,
   "tps": 155.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5swf3c4m-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: dgo329bf, es5xgapp, gijhycy5, n9bk06xa, u70f9zcm; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3145.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 8k 3,145 156.8 32.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5swf3c4m",
   "source_url": "https://omlx.ai/benchmarks/performance/5swf3c4m",
   "tg_tokens": null,
   "tps": 156.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5swf3c4m-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.5 GB at this context; duplicate traces of the same session: dgo329bf, es5xgapp, gijhycy5, n9bk06xa, u70f9zcm; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2820.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 16k 2,820 200.6 32.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5swf3c4m",
   "source_url": "https://omlx.ai/benchmarks/performance/5swf3c4m",
   "tg_tokens": null,
   "tps": 200.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5swf3c4m-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.3 GB at this context; duplicate traces of the same session: dgo329bf, es5xgapp, gijhycy5, n9bk06xa, u70f9zcm; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2324.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 32k 2,324 236.1 33.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5swf3c4m",
   "source_url": "https://omlx.ai/benchmarks/performance/5swf3c4m",
   "tg_tokens": null,
   "tps": 236.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5swf3c4m-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.0 GB at this context; duplicate traces of the same session: dgo329bf, es5xgapp, gijhycy5, n9bk06xa, u70f9zcm; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1715.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 64k 1,715 131.4 35.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5swf3c4m",
   "source_url": "https://omlx.ai/benchmarks/performance/5swf3c4m",
   "tg_tokens": null,
   "tps": 131.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5swf3c4m-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.3 GB at this context; duplicate traces of the same session: dgo329bf, es5xgapp, gijhycy5, n9bk06xa, u70f9zcm; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1124.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 128k 1,124 111.5 38.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5swf3c4m",
   "source_url": "https://omlx.ai/benchmarks/performance/5swf3c4m",
   "tg_tokens": null,
   "tps": 111.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-l3cmrimv-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: hsqchnxa; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 214.9,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 214.9 6.6 21.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l3cmrimv",
   "source_url": "https://omlx.ai/benchmarks/performance/l3cmrimv",
   "tg_tokens": null,
   "tps": 6.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-l3cmrimv-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: hsqchnxa; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 207.3,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 16k 207.3 6.2 21.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l3cmrimv",
   "source_url": "https://omlx.ai/benchmarks/performance/l3cmrimv",
   "tg_tokens": null,
   "tps": 6.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-ey0xzyxl-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 66.3 GB; TTFT 144384 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 453.9,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 3bit; 64k tokens 453.9 PP tok/s 32.7 TG tok/s 144384 TTFT (ms) 66.3 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ey0xzyxl",
   "source_url": "https://omlx.ai/benchmarks/performance/ey0xzyxl",
   "tg_tokens": null,
   "tps": 32.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-nk2ltpdp-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 68.0 GB; TTFT 137055 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 478.2,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 3bit; 64k tokens 478.2 PP tok/s 30.9 TG tok/s 137055 TTFT (ms) 68.0 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nk2ltpdp",
   "source_url": "https://omlx.ai/benchmarks/performance/nk2ltpdp",
   "tg_tokens": null,
   "tps": 30.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-vhjpybxg-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 61.5 GB at this context; duplicate traces of the same session: 9phke2qj; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 308.8,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 3bit; 1k 308.8 31.2 61.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vhjpybxg",
   "source_url": "https://omlx.ai/benchmarks/performance/vhjpybxg",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-vhjpybxg-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 61.9 GB at this context; duplicate traces of the same session: 9phke2qj; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 334.9,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 3bit; 4k 334.9 40.1 61.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vhjpybxg",
   "source_url": "https://omlx.ai/benchmarks/performance/vhjpybxg",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-vhjpybxg-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 63.1 GB at this context; duplicate traces of the same session: 9phke2qj; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 482.1,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 3bit; 16k 482.1 57.9 63.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vhjpybxg",
   "source_url": "https://omlx.ai/benchmarks/performance/vhjpybxg",
   "tg_tokens": null,
   "tps": 57.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-vhjpybxg-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 66.8 GB at this context; duplicate traces of the same session: 9phke2qj; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 477.2,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 3bit; 64k 477.2 34.2 66.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vhjpybxg",
   "source_url": "https://omlx.ai/benchmarks/performance/vhjpybxg",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-vhjpybxg-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: 9phke2qj; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 418.4,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 3bit; 128k 418.4 29.1 72.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vhjpybxg",
   "source_url": "https://omlx.ai/benchmarks/performance/vhjpybxg",
   "tg_tokens": null,
   "tps": 29.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-vhjpybxg-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 75.7 GB at this context; duplicate traces of the same session: 9phke2qj; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 435.3,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 3bit; 195k 435.3 26.1 75.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vhjpybxg",
   "source_url": "https://omlx.ai/benchmarks/performance/vhjpybxg",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lexx0tsp-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 100.7 GB; TTFT 1253 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 817.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k tokens 817.2 PP tok/s 93.5 TG tok/s 1253 TTFT (ms) 100.7 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lexx0tsp",
   "source_url": "https://omlx.ai/benchmarks/performance/lexx0tsp",
   "tg_tokens": null,
   "tps": 93.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hd45n3w5-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.4 GB at this context; duplicate traces of the same session: fe779tah; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1637.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,637 67.1 69.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hd45n3w5",
   "source_url": "https://omlx.ai/benchmarks/performance/hd45n3w5",
   "tg_tokens": null,
   "tps": 67.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hd45n3w5-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: fe779tah; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2444.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,444 69.4 70.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hd45n3w5",
   "source_url": "https://omlx.ai/benchmarks/performance/hd45n3w5",
   "tg_tokens": null,
   "tps": 69.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-2826u9au-ctx1024",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 7.4 GB; TTFT 4720 ms; duplicate traces of the same session: none; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 217.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 1k tokens 217.0 PP tok/s 18.1 TG tok/s 4720 TTFT (ms) 7.4 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2826u9au",
   "source_url": "https://omlx.ai/benchmarks/performance/2826u9au",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-bcndk90g-ctx1024",
   "model": "gemma-4-26B-A4B-it-qat",
   "notes": "QAT 6bit; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: mieq6fee, prm2b6ij; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1762.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "gemma-4-26B-A4B-it-qat-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 1,762 63.8 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bcndk90g",
   "source_url": "https://omlx.ai/benchmarks/performance/bcndk90g",
   "tg_tokens": null,
   "tps": 63.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-bcndk90g-ctx4096",
   "model": "gemma-4-26B-A4B-it-qat",
   "notes": "QAT 6bit; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: mieq6fee, prm2b6ij; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1818.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "gemma-4-26B-A4B-it-qat-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 1,818 60.3 20.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bcndk90g",
   "source_url": "https://omlx.ai/benchmarks/performance/bcndk90g",
   "tg_tokens": null,
   "tps": 60.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-bcndk90g-ctx16384",
   "model": "gemma-4-26B-A4B-it-qat",
   "notes": "QAT 6bit; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: mieq6fee, prm2b6ij; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1735.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "gemma-4-26B-A4B-it-qat-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 16k 1,735 53.0 20.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bcndk90g",
   "source_url": "https://omlx.ai/benchmarks/performance/bcndk90g",
   "tg_tokens": null,
   "tps": 53.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-y55lvf1t-ctx4096",
   "model": "gemma-4-E2B-Heretic-Uncensored",
   "notes": "MLX; oMLX community benchmark trace; peak memory 3.0 GB at this context; duplicate traces of the same session: 321spd3j; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 974.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E2B-Heretic-Uncensored-mlx-4bit on M1 (8c); M1 (8c) \u00b7 8 GB \u00b7 4bit; 4k 974.4 36.8 3.0 GB view",
   "ram_gb": 8,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y55lvf1t",
   "source_url": "https://omlx.ai/benchmarks/performance/y55lvf1t",
   "tg_tokens": null,
   "tps": 36.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-y55lvf1t-ctx16384",
   "model": "gemma-4-E2B-Heretic-Uncensored",
   "notes": "MLX; oMLX community benchmark trace; peak memory 3.1 GB at this context; duplicate traces of the same session: 321spd3j; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 919.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E2B-Heretic-Uncensored-mlx-4bit on M1 (8c); M1 (8c) \u00b7 8 GB \u00b7 4bit; 16k 919.2 32.1 3.1 GB current",
   "ram_gb": 8,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y55lvf1t",
   "source_url": "https://omlx.ai/benchmarks/performance/y55lvf1t",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-otoesawq-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 316.2 GB at this context; duplicate traces of the same session: cqza7dmb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 413.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "GLM-5.3-Flash-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 413.9 31.5 316.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/otoesawq",
   "source_url": "https://omlx.ai/benchmarks/performance/otoesawq",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-otoesawq-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 317.3 GB at this context; duplicate traces of the same session: cqza7dmb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 525.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "GLM-5.3-Flash-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 525.4 30.1 317.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/otoesawq",
   "source_url": "https://omlx.ai/benchmarks/performance/otoesawq",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-yve1y0g8-ctx1024",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.6 GB at this context; duplicate traces of the same session: lukznv76; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 509.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 509.9 39.4 177.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yve1y0g8",
   "source_url": "https://omlx.ai/benchmarks/performance/yve1y0g8",
   "tg_tokens": null,
   "tps": 39.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-yve1y0g8-ctx32768",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.4 GB at this context; duplicate traces of the same session: lukznv76; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 537.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 537.1 36.1 179.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yve1y0g8",
   "source_url": "https://omlx.ai/benchmarks/performance/yve1y0g8",
   "tg_tokens": null,
   "tps": 36.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-v4kxsdwc-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.1 GB at this context; duplicate traces of the same session: ewmf73sw, rku24w73; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 503.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 4bit; 1k 503.1 52.3 13.1 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v4kxsdwc",
   "source_url": "https://omlx.ai/benchmarks/performance/v4kxsdwc",
   "tg_tokens": null,
   "tps": 52.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-v4kxsdwc-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.5 GB at this context; duplicate traces of the same session: ewmf73sw, rku24w73; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 536.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 4bit; 4k 536.6 50.8 13.5 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v4kxsdwc",
   "source_url": "https://omlx.ai/benchmarks/performance/v4kxsdwc",
   "tg_tokens": null,
   "tps": 50.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-v4kxsdwc-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.6 GB at this context; duplicate traces of the same session: ewmf73sw, rku24w73; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 510.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 4bit; 8k 510.6 48.1 13.6 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v4kxsdwc",
   "source_url": "https://omlx.ai/benchmarks/performance/v4kxsdwc",
   "tg_tokens": null,
   "tps": 48.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-v4kxsdwc-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.8 GB at this context; duplicate traces of the same session: ewmf73sw, rku24w73; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 447.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 4bit; 16k 447.1 41.9 13.8 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v4kxsdwc",
   "source_url": "https://omlx.ai/benchmarks/performance/v4kxsdwc",
   "tg_tokens": null,
   "tps": 41.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-v4kxsdwc-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 14.1 GB at this context; duplicate traces of the same session: ewmf73sw, rku24w73; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 255.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 4bit; 32k 255.1 24.8 14.1 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v4kxsdwc",
   "source_url": "https://omlx.ai/benchmarks/performance/v4kxsdwc",
   "tg_tokens": null,
   "tps": 24.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-n6xu7wyj-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.1 GB at this context; duplicate traces of the same session: fh48huvq; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 502.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 4bit; 1k 502.2 51.7 13.1 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n6xu7wyj",
   "source_url": "https://omlx.ai/benchmarks/performance/n6xu7wyj",
   "tg_tokens": null,
   "tps": 51.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-n6xu7wyj-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.5 GB at this context; duplicate traces of the same session: fh48huvq; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 535.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 4bit; 4k 535.4 50.0 13.5 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n6xu7wyj",
   "source_url": "https://omlx.ai/benchmarks/performance/n6xu7wyj",
   "tg_tokens": null,
   "tps": 50.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-szz76tz6-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: 7dsvndib; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1937.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 1,937 139.5 13.7 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/szz76tz6",
   "source_url": "https://omlx.ai/benchmarks/performance/szz76tz6",
   "tg_tokens": null,
   "tps": 139.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-szz76tz6-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.9 GB at this context; duplicate traces of the same session: 7dsvndib; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1854.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 1,854 124.8 13.9 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/szz76tz6",
   "source_url": "https://omlx.ai/benchmarks/performance/szz76tz6",
   "tg_tokens": null,
   "tps": 124.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-srca68r9-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 9.4 GB at this context; duplicate traces of the same session: c6z2ovw9, mzw1wcs8, on3x72o8, va0j6l2m; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 335.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 1k 335.2 16.5 9.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/srca68r9",
   "source_url": "https://omlx.ai/benchmarks/performance/srca68r9",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-srca68r9-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 9.8 GB at this context; duplicate traces of the same session: c6z2ovw9, mzw1wcs8, on3x72o8, va0j6l2m; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 505.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 4k 505.0 19.5 9.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/srca68r9",
   "source_url": "https://omlx.ai/benchmarks/performance/srca68r9",
   "tg_tokens": null,
   "tps": 19.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-srca68r9-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 9.9 GB at this context; duplicate traces of the same session: c6z2ovw9, mzw1wcs8, on3x72o8, va0j6l2m; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 507.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 8k 507.7 16.3 9.9 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/srca68r9",
   "source_url": "https://omlx.ai/benchmarks/performance/srca68r9",
   "tg_tokens": null,
   "tps": 16.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-srca68r9-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 10.1 GB at this context; duplicate traces of the same session: c6z2ovw9, mzw1wcs8, on3x72o8, va0j6l2m; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 487.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 16k 487.8 13.9 10.1 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/srca68r9",
   "source_url": "https://omlx.ai/benchmarks/performance/srca68r9",
   "tg_tokens": null,
   "tps": 13.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-srca68r9-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 10.4 GB at this context; duplicate traces of the same session: c6z2ovw9, mzw1wcs8, on3x72o8, va0j6l2m; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 444.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 32k 444.0 14.7 10.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/srca68r9",
   "source_url": "https://omlx.ai/benchmarks/performance/srca68r9",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-kgpbyf6a-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 9.7 GB at this context; duplicate traces of the same session: 9r3y6vtu, e1g2tn8i, jsqqyx2s, o53ycqso; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 347.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 1k 347.6 27.7 9.7 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kgpbyf6a",
   "source_url": "https://omlx.ai/benchmarks/performance/kgpbyf6a",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-kgpbyf6a-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 10.2 GB at this context; duplicate traces of the same session: 9r3y6vtu, e1g2tn8i, jsqqyx2s, o53ycqso; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 517.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 4k 517.2 26.0 10.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kgpbyf6a",
   "source_url": "https://omlx.ai/benchmarks/performance/kgpbyf6a",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-kgpbyf6a-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 10.3 GB at this context; duplicate traces of the same session: 9r3y6vtu, e1g2tn8i, jsqqyx2s, o53ycqso; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 516.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 8k 516.1 16.1 10.3 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kgpbyf6a",
   "source_url": "https://omlx.ai/benchmarks/performance/kgpbyf6a",
   "tg_tokens": null,
   "tps": 16.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-kgpbyf6a-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 10.5 GB at this context; duplicate traces of the same session: 9r3y6vtu, e1g2tn8i, jsqqyx2s, o53ycqso; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 495.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 16k 495.6 19.0 10.5 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kgpbyf6a",
   "source_url": "https://omlx.ai/benchmarks/performance/kgpbyf6a",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-kgpbyf6a-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 10.8 GB at this context; duplicate traces of the same session: 9r3y6vtu, e1g2tn8i, jsqqyx2s, o53ycqso; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 449.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 32k 449.6 15.9 10.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kgpbyf6a",
   "source_url": "https://omlx.ai/benchmarks/performance/kgpbyf6a",
   "tg_tokens": null,
   "tps": 15.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-kgpbyf6a-ctx65536",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 11.4 GB at this context; duplicate traces of the same session: 9r3y6vtu, e1g2tn8i, jsqqyx2s, o53ycqso; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 375.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 64k 375.9 10.2 11.4 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kgpbyf6a",
   "source_url": "https://omlx.ai/benchmarks/performance/kgpbyf6a",
   "tg_tokens": null,
   "tps": 10.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-kgpbyf6a-ctx131072",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 12.2 GB at this context; duplicate traces of the same session: 9r3y6vtu, e1g2tn8i, jsqqyx2s, o53ycqso; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 236.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 128k 236.7 5.1 12.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kgpbyf6a",
   "source_url": "https://omlx.ai/benchmarks/performance/kgpbyf6a",
   "tg_tokens": null,
   "tps": 5.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-q1bpz4ul-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 5qd3d4m7; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1766.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 1,766 104.8 22.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q1bpz4ul",
   "source_url": "https://omlx.ai/benchmarks/performance/q1bpz4ul",
   "tg_tokens": null,
   "tps": 104.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-q1bpz4ul-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: 5qd3d4m7; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2212.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 2,212 105.5 24.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q1bpz4ul",
   "source_url": "https://omlx.ai/benchmarks/performance/q1bpz4ul",
   "tg_tokens": null,
   "tps": 105.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-h3981loj-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: nf2sfaot, x9gpby21, zftduvpu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 735.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 735.0 114.9 22.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h3981loj",
   "source_url": "https://omlx.ai/benchmarks/performance/h3981loj",
   "tg_tokens": null,
   "tps": 114.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-h3981loj-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: nf2sfaot, x9gpby21, zftduvpu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2561.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,561 114.8 22.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h3981loj",
   "source_url": "https://omlx.ai/benchmarks/performance/h3981loj",
   "tg_tokens": null,
   "tps": 114.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-h3981loj-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: nf2sfaot, x9gpby21, zftduvpu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5907.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 5,907 148.9 22.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h3981loj",
   "source_url": "https://omlx.ai/benchmarks/performance/h3981loj",
   "tg_tokens": null,
   "tps": 148.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-h3981loj-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: nf2sfaot, x9gpby21, zftduvpu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5503.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 5,503 116.4 23.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h3981loj",
   "source_url": "https://omlx.ai/benchmarks/performance/h3981loj",
   "tg_tokens": null,
   "tps": 116.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-216v3nlt-ctx1024",
   "model": "Qwable-v2",
   "notes": "DWQ; MTP speculative decoding enabled; Vision; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: zxafsslj; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1334.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwable-v2-oQ4e-DWQ-MTP-Vision-MLX on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 1,334 83.2 21.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/216v3nlt",
   "source_url": "https://omlx.ai/benchmarks/performance/216v3nlt",
   "tg_tokens": null,
   "tps": 83.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-216v3nlt-ctx4096",
   "model": "Qwable-v2",
   "notes": "DWQ; MTP speculative decoding enabled; Vision; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: zxafsslj; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1484.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwable-v2-oQ4e-DWQ-MTP-Vision-MLX on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 1,484 81.0 23.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/216v3nlt",
   "source_url": "https://omlx.ai/benchmarks/performance/216v3nlt",
   "tg_tokens": null,
   "tps": 81.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-5r8irrfz-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 5.9 GB; TTFT 2912 ms; duplicate traces of the same session: none; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 351.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 1k tokens 351.7 PP tok/s 27.9 TG tok/s 2912 TTFT (ms) 5.9 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5r8irrfz",
   "source_url": "https://omlx.ai/benchmarks/performance/5r8irrfz",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-781gdme0-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2648.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 1k 2,648 169.2 29.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/781gdme0",
   "source_url": "https://omlx.ai/benchmarks/performance/781gdme0",
   "tg_tokens": null,
   "tps": 169.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-781gdme0-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.8 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3307.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 4k 3,307 155.7 31.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/781gdme0",
   "source_url": "https://omlx.ai/benchmarks/performance/781gdme0",
   "tg_tokens": null,
   "tps": 155.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-781gdme0-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3145.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 8k 3,145 156.8 32.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/781gdme0",
   "source_url": "https://omlx.ai/benchmarks/performance/781gdme0",
   "tg_tokens": null,
   "tps": 156.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-781gdme0-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.5 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2820.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 16k 2,820 200.6 32.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/781gdme0",
   "source_url": "https://omlx.ai/benchmarks/performance/781gdme0",
   "tg_tokens": null,
   "tps": 200.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-781gdme0-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.3 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2324.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 32k 2,324 236.1 33.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/781gdme0",
   "source_url": "https://omlx.ai/benchmarks/performance/781gdme0",
   "tg_tokens": null,
   "tps": 236.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-781gdme0-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.0 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1715.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 64k 1,715 131.4 35.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/781gdme0",
   "source_url": "https://omlx.ai/benchmarks/performance/781gdme0",
   "tg_tokens": null,
   "tps": 131.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-781gdme0-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.3 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1124.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 128k 1,124 111.5 38.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/781gdme0",
   "source_url": "https://omlx.ai/benchmarks/performance/781gdme0",
   "tg_tokens": null,
   "tps": 111.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-hmazyvh2-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: 4253p5qk, c8kn35tk, xndxupyf; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2694.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 1k 2,694 184.6 29.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hmazyvh2",
   "source_url": "https://omlx.ai/benchmarks/performance/hmazyvh2",
   "tg_tokens": null,
   "tps": 184.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-hmazyvh2-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.8 GB at this context; duplicate traces of the same session: 4253p5qk, c8kn35tk, xndxupyf; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3313.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 4k 3,313 166.2 31.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hmazyvh2",
   "source_url": "https://omlx.ai/benchmarks/performance/hmazyvh2",
   "tg_tokens": null,
   "tps": 166.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-hmazyvh2-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: 4253p5qk, c8kn35tk, xndxupyf; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3143.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 8k 3,143 155.2 32.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hmazyvh2",
   "source_url": "https://omlx.ai/benchmarks/performance/hmazyvh2",
   "tg_tokens": null,
   "tps": 155.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-hmazyvh2-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.5 GB at this context; duplicate traces of the same session: 4253p5qk, c8kn35tk, xndxupyf; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2818.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 16k 2,818 190.2 32.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hmazyvh2",
   "source_url": "https://omlx.ai/benchmarks/performance/hmazyvh2",
   "tg_tokens": null,
   "tps": 190.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-lq5taw45-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: jljaph3g; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 167.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 167.8 23.4 24.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lq5taw45",
   "source_url": "https://omlx.ai/benchmarks/performance/lq5taw45",
   "tg_tokens": null,
   "tps": 23.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-lq5taw45-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.2 GB at this context; duplicate traces of the same session: jljaph3g; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 152.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 152.9 19.8 26.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lq5taw45",
   "source_url": "https://omlx.ai/benchmarks/performance/lq5taw45",
   "tg_tokens": null,
   "tps": 19.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-byr8vz7i-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 28.6 GB; TTFT 86609 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 378.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k tokens 378.3 PP tok/s 100.3 TG tok/s 86609 TTFT (ms) 28.6 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/byr8vz7i",
   "source_url": "https://omlx.ai/benchmarks/performance/byr8vz7i",
   "tg_tokens": null,
   "tps": 100.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-wzgb7pv7-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 65.2 GB; TTFT 122375 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 267.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k tokens 267.8 PP tok/s 34.9 TG tok/s 122375 TTFT (ms) 65.2 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wzgb7pv7",
   "source_url": "https://omlx.ai/benchmarks/performance/wzgb7pv7",
   "tg_tokens": null,
   "tps": 34.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gbyu1jw7-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 103.4 GB at this context; duplicate traces of the same session: r36hdp2y; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 862.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 862.9 86.0 103.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gbyu1jw7",
   "source_url": "https://omlx.ai/benchmarks/performance/gbyu1jw7",
   "tg_tokens": null,
   "tps": 86.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gbyu1jw7-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 104.1 GB at this context; duplicate traces of the same session: r36hdp2y; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1044.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 1,044 70.2 104.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gbyu1jw7",
   "source_url": "https://omlx.ai/benchmarks/performance/gbyu1jw7",
   "tg_tokens": null,
   "tps": 70.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-m60a34d1-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 61.5 GB at this context; duplicate traces of the same session: 5epaszbb, rg2o76fe, ykwdwe3y; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 308.8,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 3bit; 1k 308.8 31.2 61.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m60a34d1",
   "source_url": "https://omlx.ai/benchmarks/performance/m60a34d1",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-m60a34d1-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 61.9 GB at this context; duplicate traces of the same session: 5epaszbb, rg2o76fe, ykwdwe3y; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 334.9,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 3bit; 4k 334.9 40.1 61.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m60a34d1",
   "source_url": "https://omlx.ai/benchmarks/performance/m60a34d1",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-m60a34d1-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 63.1 GB at this context; duplicate traces of the same session: 5epaszbb, rg2o76fe, ykwdwe3y; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 482.1,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 3bit; 16k 482.1 57.9 63.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m60a34d1",
   "source_url": "https://omlx.ai/benchmarks/performance/m60a34d1",
   "tg_tokens": null,
   "tps": 57.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-m60a34d1-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 66.8 GB at this context; duplicate traces of the same session: 5epaszbb, rg2o76fe, ykwdwe3y; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 477.2,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 3bit; 64k 477.2 34.2 66.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m60a34d1",
   "source_url": "https://omlx.ai/benchmarks/performance/m60a34d1",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-m60a34d1-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: 5epaszbb, rg2o76fe, ykwdwe3y; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 418.4,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 3bit; 128k 418.4 29.1 72.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m60a34d1",
   "source_url": "https://omlx.ai/benchmarks/performance/m60a34d1",
   "tg_tokens": null,
   "tps": 29.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-m60a34d1-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 75.7 GB at this context; duplicate traces of the same session: 5epaszbb, rg2o76fe, ykwdwe3y; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 435.3,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 3bit; 195k 435.3 26.1 75.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m60a34d1",
   "source_url": "https://omlx.ai/benchmarks/performance/m60a34d1",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-95w2qcwe-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: acrurbh6; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1172.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,172 103.7 100.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/95w2qcwe",
   "source_url": "https://omlx.ai/benchmarks/performance/95w2qcwe",
   "tg_tokens": null,
   "tps": 103.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-95w2qcwe-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: acrurbh6; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1397.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,397 74.2 101.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/95w2qcwe",
   "source_url": "https://omlx.ai/benchmarks/performance/95w2qcwe",
   "tg_tokens": null,
   "tps": 74.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-95w2qcwe-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: acrurbh6; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1414.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 1,414 93.2 101.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/95w2qcwe",
   "source_url": "https://omlx.ai/benchmarks/performance/95w2qcwe",
   "tg_tokens": null,
   "tps": 93.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-95w2qcwe-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.1 GB at this context; duplicate traces of the same session: acrurbh6; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1402.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 1,402 70.0 102.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/95w2qcwe",
   "source_url": "https://omlx.ai/benchmarks/performance/95w2qcwe",
   "tg_tokens": null,
   "tps": 70.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-c85o8bmk-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1176.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,176 104.9 100.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c85o8bmk",
   "source_url": "https://omlx.ai/benchmarks/performance/c85o8bmk",
   "tg_tokens": null,
   "tps": 104.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-c85o8bmk-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1406.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 1,406 86.1 101.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c85o8bmk",
   "source_url": "https://omlx.ai/benchmarks/performance/c85o8bmk",
   "tg_tokens": null,
   "tps": 86.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-c85o8bmk-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1390.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 1,390 121.9 102.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c85o8bmk",
   "source_url": "https://omlx.ai/benchmarks/performance/c85o8bmk",
   "tg_tokens": null,
   "tps": 121.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-gedq7ti5-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: devbel0x, mq9dfvnk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1975.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,975 100.1 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gedq7ti5",
   "source_url": "https://omlx.ai/benchmarks/performance/gedq7ti5",
   "tg_tokens": null,
   "tps": 100.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gedq7ti5-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: devbel0x, mq9dfvnk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2338.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,338 64.8 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gedq7ti5",
   "source_url": "https://omlx.ai/benchmarks/performance/gedq7ti5",
   "tg_tokens": null,
   "tps": 64.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gedq7ti5-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: devbel0x, mq9dfvnk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2582.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,582 82.6 72.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gedq7ti5",
   "source_url": "https://omlx.ai/benchmarks/performance/gedq7ti5",
   "tg_tokens": null,
   "tps": 82.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zokqxiuo-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.5 GB; TTFT 1085 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 943.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 943.5 PP tok/s 70.5 TG tok/s 1085 TTFT (ms) 70.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zokqxiuo",
   "source_url": "https://omlx.ai/benchmarks/performance/zokqxiuo",
   "tg_tokens": null,
   "tps": 70.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mlae0g2r-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: fd09tcuy, hb7kgd53; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 375.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 375.3 57.9 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mlae0g2r",
   "source_url": "https://omlx.ai/benchmarks/performance/mlae0g2r",
   "tg_tokens": null,
   "tps": 57.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mlae0g2r-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: fd09tcuy, hb7kgd53; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1002.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,002 53.5 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mlae0g2r",
   "source_url": "https://omlx.ai/benchmarks/performance/mlae0g2r",
   "tg_tokens": null,
   "tps": 53.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mlae0g2r-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: fd09tcuy, hb7kgd53; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1788.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,788 64.9 71.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mlae0g2r",
   "source_url": "https://omlx.ai/benchmarks/performance/mlae0g2r",
   "tg_tokens": null,
   "tps": 64.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-u4aiv3sy-ctx1024",
   "model": "StartLux-Decision-4B",
   "notes": "oMLX community benchmark trace; peak memory 9.0 GB at this context; duplicate traces of the same session: 393tfbto; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4788.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "StartLux-Decision-4B on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 4,788 52.9 9.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u4aiv3sy",
   "source_url": "https://omlx.ai/benchmarks/performance/u4aiv3sy",
   "tg_tokens": null,
   "tps": 52.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-u4aiv3sy-ctx4096",
   "model": "StartLux-Decision-4B",
   "notes": "oMLX community benchmark trace; peak memory 9.4 GB at this context; duplicate traces of the same session: 393tfbto; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4001.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "StartLux-Decision-4B on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 4k 4,001 52.7 9.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u4aiv3sy",
   "source_url": "https://omlx.ai/benchmarks/performance/u4aiv3sy",
   "tg_tokens": null,
   "tps": 52.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-xuzjrn0g-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 27.9 GB; TTFT 12790 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 320.3,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 5bit; 4k tokens 320.3 PP tok/s 52.9 TG tok/s 12790 TTFT (ms) 27.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xuzjrn0g",
   "source_url": "https://omlx.ai/benchmarks/performance/xuzjrn0g",
   "tg_tokens": null,
   "tps": 52.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-iabytrl0-ctx1024",
   "model": "supergemma4-26b-uncensored",
   "notes": "oMLX community benchmark trace; peak memory 14.1 GB at this context; duplicate traces of the same session: nivx0g9d; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 859.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "supergemma4-26b-uncensored-mlx-4bit-v2 on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 859.5 84.0 14.1 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iabytrl0",
   "source_url": "https://omlx.ai/benchmarks/performance/iabytrl0",
   "tg_tokens": null,
   "tps": 84.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-iabytrl0-ctx4096",
   "model": "supergemma4-26b-uncensored",
   "notes": "oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: nivx0g9d; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1712.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "supergemma4-26b-uncensored-mlx-4bit-v2 on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 1,712 81.1 14.4 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iabytrl0",
   "source_url": "https://omlx.ai/benchmarks/performance/iabytrl0",
   "tg_tokens": null,
   "tps": 81.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-j3vgjvpc-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 90.0 GB at this context; duplicate traces of the same session: 5zb5kqak, nthqafvu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 86.9,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 86.9 10.1 90.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j3vgjvpc",
   "source_url": "https://omlx.ai/benchmarks/performance/j3vgjvpc",
   "tg_tokens": null,
   "tps": 10.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-j3vgjvpc-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 90.3 GB at this context; duplicate traces of the same session: 5zb5kqak, nthqafvu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 164.2,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 164.2 12.3 90.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j3vgjvpc",
   "source_url": "https://omlx.ai/benchmarks/performance/j3vgjvpc",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-j3vgjvpc-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 90.5 GB at this context; duplicate traces of the same session: 5zb5kqak, nthqafvu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 165.5,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 165.5 13.1 90.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j3vgjvpc",
   "source_url": "https://omlx.ai/benchmarks/performance/j3vgjvpc",
   "tg_tokens": null,
   "tps": 13.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-avso66yf-ctx16384",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 179.6 GB; TTFT 7035 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2329.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k tokens 2,329 PP tok/s 57.3 TG tok/s 7035 TTFT (ms) 179.6 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/avso66yf",
   "source_url": "https://omlx.ai/benchmarks/performance/avso66yf",
   "tg_tokens": null,
   "tps": 57.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ow9ag4yi-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 323.6 GB at this context; duplicate traces of the same session: fa31xdm6, vs1iwnon, vvdpfwo2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 493.3,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "GLM-5.3-Flash-oQ8-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 493.3 38.6 323.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ow9ag4yi",
   "source_url": "https://omlx.ai/benchmarks/performance/ow9ag4yi",
   "tg_tokens": null,
   "tps": 38.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ow9ag4yi-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 324.7 GB at this context; duplicate traces of the same session: fa31xdm6, vs1iwnon, vvdpfwo2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 569.2,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "GLM-5.3-Flash-oQ8-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 569.2 38.4 324.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ow9ag4yi",
   "source_url": "https://omlx.ai/benchmarks/performance/ow9ag4yi",
   "tg_tokens": null,
   "tps": 38.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ow9ag4yi-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 325.0 GB at this context; duplicate traces of the same session: fa31xdm6, vs1iwnon, vvdpfwo2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 527.9,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "GLM-5.3-Flash-oQ8-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 527.9 41.9 325.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ow9ag4yi",
   "source_url": "https://omlx.ai/benchmarks/performance/ow9ag4yi",
   "tg_tokens": null,
   "tps": 41.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ow9ag4yi-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 325.4 GB at this context; duplicate traces of the same session: fa31xdm6, vs1iwnon, vvdpfwo2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 510.3,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "GLM-5.3-Flash-oQ8-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 32k 510.3 49.1 325.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ow9ag4yi",
   "source_url": "https://omlx.ai/benchmarks/performance/ow9ag4yi",
   "tg_tokens": null,
   "tps": 49.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-tosdomdj-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: izqq2yk0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3861.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 3,861 123.3 13.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tosdomdj",
   "source_url": "https://omlx.ai/benchmarks/performance/tosdomdj",
   "tg_tokens": null,
   "tps": 123.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-tosdomdj-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: izqq2yk0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4705.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 4,705 126.8 13.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tosdomdj",
   "source_url": "https://omlx.ai/benchmarks/performance/tosdomdj",
   "tg_tokens": null,
   "tps": 126.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-8va02c9g-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: f1qnjn9f; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1700.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "LFM2.5-24B-A2B-MLX-6bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 6bit; 1k 1,700 120.0 18.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8va02c9g",
   "source_url": "https://omlx.ai/benchmarks/performance/8va02c9g",
   "tg_tokens": null,
   "tps": 120.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-8va02c9g-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: f1qnjn9f; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1946.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "LFM2.5-24B-A2B-MLX-6bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 6bit; 4k 1,946 117.0 19.2 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8va02c9g",
   "source_url": "https://omlx.ai/benchmarks/performance/8va02c9g",
   "tg_tokens": null,
   "tps": 117.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-78x5vb6f-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 9.7 GB at this context; duplicate traces of the same session: 9avztykz; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 347.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 1k 347.6 27.7 9.7 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/78x5vb6f",
   "source_url": "https://omlx.ai/benchmarks/performance/78x5vb6f",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-78x5vb6f-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 10.2 GB at this context; duplicate traces of the same session: 9avztykz; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 517.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 4k 517.2 26.0 10.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/78x5vb6f",
   "source_url": "https://omlx.ai/benchmarks/performance/78x5vb6f",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-78x5vb6f-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 10.3 GB at this context; duplicate traces of the same session: 9avztykz; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 516.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 8k 516.1 16.1 10.3 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/78x5vb6f",
   "source_url": "https://omlx.ai/benchmarks/performance/78x5vb6f",
   "tg_tokens": null,
   "tps": 16.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-78x5vb6f-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 10.5 GB at this context; duplicate traces of the same session: 9avztykz; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 495.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 16k 495.6 19.0 10.5 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/78x5vb6f",
   "source_url": "https://omlx.ai/benchmarks/performance/78x5vb6f",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-78x5vb6f-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 10.8 GB at this context; duplicate traces of the same session: 9avztykz; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 449.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 32k 449.6 15.9 10.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/78x5vb6f",
   "source_url": "https://omlx.ai/benchmarks/performance/78x5vb6f",
   "tg_tokens": null,
   "tps": 15.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-78x5vb6f-ctx65536",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 11.4 GB at this context; duplicate traces of the same session: 9avztykz; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 375.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 64k 375.9 10.2 11.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/78x5vb6f",
   "source_url": "https://omlx.ai/benchmarks/performance/78x5vb6f",
   "tg_tokens": null,
   "tps": 10.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-78x5vb6f-ctx131072",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 12.2 GB at this context; duplicate traces of the same session: 9avztykz; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 236.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e-fp16 on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 128k 236.7 5.1 12.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/78x5vb6f",
   "source_url": "https://omlx.ai/benchmarks/performance/78x5vb6f",
   "tg_tokens": null,
   "tps": 5.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-4ghtfrqt-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "UD MLX build; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 6nqxvsr8; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 475.7,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-3bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 3bit; 1k 475.7 36.7 16.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ghtfrqt",
   "source_url": "https://omlx.ai/benchmarks/performance/4ghtfrqt",
   "tg_tokens": null,
   "tps": 36.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-4ghtfrqt-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "UD MLX build; oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: 6nqxvsr8; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 490.6,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-3bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 3bit; 4k 490.6 35.7 17.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ghtfrqt",
   "source_url": "https://omlx.ai/benchmarks/performance/4ghtfrqt",
   "tg_tokens": null,
   "tps": 35.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-mkpc7kel-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: wxsgpayp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 133.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 133.1 13.8 17.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mkpc7kel",
   "source_url": "https://omlx.ai/benchmarks/performance/mkpc7kel",
   "tg_tokens": null,
   "tps": 13.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-mkpc7kel-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: wxsgpayp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 113.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 113.4 12.8 18.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mkpc7kel",
   "source_url": "https://omlx.ai/benchmarks/performance/mkpc7kel",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2xx1bq3v-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 8037qgbt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 949.2,
   "provenance": "community",
   "quant": "OptiQ 4bit",
   "quote": "Qwen3.8-27B-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 949.2 49.4 21.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2xx1bq3v",
   "source_url": "https://omlx.ai/benchmarks/performance/2xx1bq3v",
   "tg_tokens": null,
   "tps": 49.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2xx1bq3v-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: 8037qgbt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 971.6,
   "provenance": "community",
   "quant": "OptiQ 4bit",
   "quote": "Qwen3.8-27B-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 971.6 47.7 22.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2xx1bq3v",
   "source_url": "https://omlx.ai/benchmarks/performance/2xx1bq3v",
   "tg_tokens": null,
   "tps": 47.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-r1vm4gwt-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: o0pkqrt5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 157.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 157.5 21.5 24.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r1vm4gwt",
   "source_url": "https://omlx.ai/benchmarks/performance/r1vm4gwt",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-r1vm4gwt-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.2 GB at this context; duplicate traces of the same session: o0pkqrt5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 143.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 143.0 18.1 26.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r1vm4gwt",
   "source_url": "https://omlx.ai/benchmarks/performance/r1vm4gwt",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-27l7tikd-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: r14b1sv9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 797.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 797.5 53.2 33.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/27l7tikd",
   "source_url": "https://omlx.ai/benchmarks/performance/27l7tikd",
   "tg_tokens": null,
   "tps": 53.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-27l7tikd-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: r14b1sv9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 910.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 910.8 41.1 34.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/27l7tikd",
   "source_url": "https://omlx.ai/benchmarks/performance/27l7tikd",
   "tg_tokens": null,
   "tps": 41.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ixjy69j6-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 103.4 GB at this context; duplicate traces of the same session: yzafkp1a; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 802.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 802.0 81.6 103.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ixjy69j6",
   "source_url": "https://omlx.ai/benchmarks/performance/ixjy69j6",
   "tg_tokens": null,
   "tps": 81.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ixjy69j6-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 104.1 GB at this context; duplicate traces of the same session: yzafkp1a; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 986.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 986.6 74.8 104.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ixjy69j6",
   "source_url": "https://omlx.ai/benchmarks/performance/ixjy69j6",
   "tg_tokens": null,
   "tps": 74.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-f9d9vjxj-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 101.2 GB at this context; duplicate traces of the same session: xezvbkir; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 874.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 874.1 65.1 101.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f9d9vjxj",
   "source_url": "https://omlx.ai/benchmarks/performance/f9d9vjxj",
   "tg_tokens": null,
   "tps": 65.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-f9d9vjxj-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 101.8 GB at this context; duplicate traces of the same session: xezvbkir; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1073.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 1,073 62.8 101.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f9d9vjxj",
   "source_url": "https://omlx.ai/benchmarks/performance/f9d9vjxj",
   "tg_tokens": null,
   "tps": 62.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-risclzur-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace, single run at 1k context; peak memory 101.2 GB; TTFT 1192 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 859.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k tokens 859.4 PP tok/s 63.5 TG tok/s 1192 TTFT (ms) 101.2 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/risclzur",
   "source_url": "https://omlx.ai/benchmarks/performance/risclzur",
   "tg_tokens": null,
   "tps": 63.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ecc5k9z3-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "g32 MTP build; oMLX community benchmark trace; peak memory 107.7 GB at this context; duplicate traces of the same session: esk66wiu, ijeavnd8, vbwjxh81, vzx4axro; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1115.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-g32-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,115 96.4 107.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ecc5k9z3",
   "source_url": "https://omlx.ai/benchmarks/performance/ecc5k9z3",
   "tg_tokens": null,
   "tps": 96.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ecc5k9z3-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "g32 MTP build; oMLX community benchmark trace; peak memory 108.2 GB at this context; duplicate traces of the same session: esk66wiu, ijeavnd8, vbwjxh81, vzx4axro; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1357.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-g32-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 1,357 96.0 108.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ecc5k9z3",
   "source_url": "https://omlx.ai/benchmarks/performance/ecc5k9z3",
   "tg_tokens": null,
   "tps": 96.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ecc5k9z3-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "g32 MTP build; oMLX community benchmark trace; peak memory 108.4 GB at this context; duplicate traces of the same session: esk66wiu, ijeavnd8, vbwjxh81, vzx4axro; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1339.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-g32-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 1,339 88.2 108.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ecc5k9z3",
   "source_url": "https://omlx.ai/benchmarks/performance/ecc5k9z3",
   "tg_tokens": null,
   "tps": 88.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ecc5k9z3-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "g32 MTP build; oMLX community benchmark trace; peak memory 108.7 GB at this context; duplicate traces of the same session: esk66wiu, ijeavnd8, vbwjxh81, vzx4axro; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1327.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-g32-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 1,327 97.6 108.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ecc5k9z3",
   "source_url": "https://omlx.ai/benchmarks/performance/ecc5k9z3",
   "tg_tokens": null,
   "tps": 97.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ecc5k9z3-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "g32 MTP build; oMLX community benchmark trace; peak memory 109.3 GB at this context; duplicate traces of the same session: esk66wiu, ijeavnd8, vbwjxh81, vzx4axro; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1313.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-g32-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 1,313 81.3 109.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ecc5k9z3",
   "source_url": "https://omlx.ai/benchmarks/performance/ecc5k9z3",
   "tg_tokens": null,
   "tps": 81.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ck4k2oz5-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 87.8 GB; TTFT 79505 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1649.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 128k tokens 1,649 PP tok/s 46.9 TG tok/s 79505 TTFT (ms) 87.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ck4k2oz5",
   "source_url": "https://omlx.ai/benchmarks/performance/ck4k2oz5",
   "tg_tokens": null,
   "tps": 46.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bllfb5j5-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 84.0 GB; TTFT 30470 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2151.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 64k tokens 2,151 PP tok/s 52.8 TG tok/s 30470 TTFT (ms) 84.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bllfb5j5",
   "source_url": "https://omlx.ai/benchmarks/performance/bllfb5j5",
   "tg_tokens": null,
   "tps": 52.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-sa2ywe4g-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: hmj5ye0u; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1172.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,172 103.7 100.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sa2ywe4g",
   "source_url": "https://omlx.ai/benchmarks/performance/sa2ywe4g",
   "tg_tokens": null,
   "tps": 103.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-sa2ywe4g-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: hmj5ye0u; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1397.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,397 74.2 101.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sa2ywe4g",
   "source_url": "https://omlx.ai/benchmarks/performance/sa2ywe4g",
   "tg_tokens": null,
   "tps": 74.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-sa2ywe4g-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: hmj5ye0u; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1414.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 1,414 93.2 101.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sa2ywe4g",
   "source_url": "https://omlx.ai/benchmarks/performance/sa2ywe4g",
   "tg_tokens": null,
   "tps": 93.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-sa2ywe4g-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.1 GB at this context; duplicate traces of the same session: hmj5ye0u; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1402.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 1,402 70.0 102.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sa2ywe4g",
   "source_url": "https://omlx.ai/benchmarks/performance/sa2ywe4g",
   "tg_tokens": null,
   "tps": 70.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1m3xqzw5-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.7 GB at this context; duplicate traces of the same session: b6m4dt7b; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1176.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,176 104.9 100.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1m3xqzw5",
   "source_url": "https://omlx.ai/benchmarks/performance/1m3xqzw5",
   "tg_tokens": null,
   "tps": 104.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1m3xqzw5-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: b6m4dt7b; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1406.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 1,406 86.1 101.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1m3xqzw5",
   "source_url": "https://omlx.ai/benchmarks/performance/1m3xqzw5",
   "tg_tokens": null,
   "tps": 86.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1m3xqzw5-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.1 GB at this context; duplicate traces of the same session: b6m4dt7b; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1390.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 1,390 121.9 102.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1m3xqzw5",
   "source_url": "https://omlx.ai/benchmarks/performance/1m3xqzw5",
   "tg_tokens": null,
   "tps": 121.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qd78prbs-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.1 GB at this context; duplicate traces of the same session: 4ydlngza, c2gfiw81, df2rfzyx, ffa0ix1a, fu8gh62g, k0phtx4t; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 765.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 765.6 84.3 99.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qd78prbs",
   "source_url": "https://omlx.ai/benchmarks/performance/qd78prbs",
   "tg_tokens": null,
   "tps": 84.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qd78prbs-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.6 GB at this context; duplicate traces of the same session: 4ydlngza, c2gfiw81, df2rfzyx, ffa0ix1a, fu8gh62g, k0phtx4t; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1103.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,103 67.6 99.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qd78prbs",
   "source_url": "https://omlx.ai/benchmarks/performance/qd78prbs",
   "tg_tokens": null,
   "tps": 67.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qd78prbs-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.7 GB at this context; duplicate traces of the same session: 4ydlngza, c2gfiw81, df2rfzyx, ffa0ix1a, fu8gh62g, k0phtx4t; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1443.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 1,443 76.9 99.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qd78prbs",
   "source_url": "https://omlx.ai/benchmarks/performance/qd78prbs",
   "tg_tokens": null,
   "tps": 76.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qd78prbs-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.0 GB at this context; duplicate traces of the same session: 4ydlngza, c2gfiw81, df2rfzyx, ffa0ix1a, fu8gh62g, k0phtx4t; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1426.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 1,426 75.8 100.0 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qd78prbs",
   "source_url": "https://omlx.ai/benchmarks/performance/qd78prbs",
   "tg_tokens": null,
   "tps": 75.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qd78prbs-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: 4ydlngza, c2gfiw81, df2rfzyx, ffa0ix1a, fu8gh62g, k0phtx4t; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1427.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 1,427 35.6 100.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qd78prbs",
   "source_url": "https://omlx.ai/benchmarks/performance/qd78prbs",
   "tg_tokens": null,
   "tps": 35.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qd78prbs-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: 4ydlngza, c2gfiw81, df2rfzyx, ffa0ix1a, fu8gh62g, k0phtx4t; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1403.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 1,403 50.3 101.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qd78prbs",
   "source_url": "https://omlx.ai/benchmarks/performance/qd78prbs",
   "tg_tokens": null,
   "tps": 50.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qd78prbs-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.7 GB at this context; duplicate traces of the same session: 4ydlngza, c2gfiw81, df2rfzyx, ffa0ix1a, fu8gh62g, k0phtx4t; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1383.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 1,383 58.4 104.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qd78prbs",
   "source_url": "https://omlx.ai/benchmarks/performance/qd78prbs",
   "tg_tokens": null,
   "tps": 58.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8jn1r6x7-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: 4xfg9ite, hyiu77hz, nitmlfet, o1sy99fa, oly0allz, xjf6ic3z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1001.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,001 47.4 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8jn1r6x7",
   "source_url": "https://omlx.ai/benchmarks/performance/8jn1r6x7",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8jn1r6x7-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: 4xfg9ite, hyiu77hz, nitmlfet, o1sy99fa, oly0allz, xjf6ic3z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1783.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,783 75.4 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8jn1r6x7",
   "source_url": "https://omlx.ai/benchmarks/performance/8jn1r6x7",
   "tg_tokens": null,
   "tps": 75.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8jn1r6x7-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 4xfg9ite, hyiu77hz, nitmlfet, o1sy99fa, oly0allz, xjf6ic3z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2177.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,177 88.5 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8jn1r6x7",
   "source_url": "https://omlx.ai/benchmarks/performance/8jn1r6x7",
   "tg_tokens": null,
   "tps": 88.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8jn1r6x7-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.9 GB at this context; duplicate traces of the same session: 4xfg9ite, hyiu77hz, nitmlfet, o1sy99fa, oly0allz, xjf6ic3z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2542.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,542 82.7 72.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8jn1r6x7",
   "source_url": "https://omlx.ai/benchmarks/performance/8jn1r6x7",
   "tg_tokens": null,
   "tps": 82.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8jn1r6x7-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 75.3 GB at this context; duplicate traces of the same session: 4xfg9ite, hyiu77hz, nitmlfet, o1sy99fa, oly0allz, xjf6ic3z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2438.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,438 82.7 75.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8jn1r6x7",
   "source_url": "https://omlx.ai/benchmarks/performance/8jn1r6x7",
   "tg_tokens": null,
   "tps": 82.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8jn1r6x7-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 81.0 GB at this context; duplicate traces of the same session: 4xfg9ite, hyiu77hz, nitmlfet, o1sy99fa, oly0allz, xjf6ic3z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2254.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,254 71.6 81.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8jn1r6x7",
   "source_url": "https://omlx.ai/benchmarks/performance/8jn1r6x7",
   "tg_tokens": null,
   "tps": 71.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8jn1r6x7-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 83.4 GB at this context; duplicate traces of the same session: 4xfg9ite, hyiu77hz, nitmlfet, o1sy99fa, oly0allz, xjf6ic3z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2017.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,017 62.1 83.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8jn1r6x7",
   "source_url": "https://omlx.ai/benchmarks/performance/8jn1r6x7",
   "tg_tokens": null,
   "tps": 62.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-zuryjklq-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace, single run at 1k context; peak memory 70.8 GB; TTFT 2383 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 429.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 429.8 PP tok/s 44.9 TG tok/s 2383 TTFT (ms) 70.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zuryjklq",
   "source_url": "https://omlx.ai/benchmarks/performance/zuryjklq",
   "tg_tokens": null,
   "tps": 44.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-k6bg2110-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 48.2 GB at this context; duplicate traces of the same session: 0pjgvq1s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1438.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "gemma-4-26B-A4B-it on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 4k 1,438 58.6 48.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k6bg2110",
   "source_url": "https://omlx.ai/benchmarks/performance/k6bg2110",
   "tg_tokens": null,
   "tps": 58.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-k6bg2110-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 48.7 GB at this context; duplicate traces of the same session: 0pjgvq1s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2849.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "gemma-4-26B-A4B-it on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 2,849 54.8 48.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k6bg2110",
   "source_url": "https://omlx.ai/benchmarks/performance/k6bg2110",
   "tg_tokens": null,
   "tps": 54.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-yshuqghe-ctx1024",
   "model": "gemma-4-31B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 365.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31B-it-qat-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 365.0 11.6 22.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yshuqghe",
   "source_url": "https://omlx.ai/benchmarks/performance/yshuqghe",
   "tg_tokens": null,
   "tps": 11.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-yshuqghe-ctx4096",
   "model": "gemma-4-31B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 359.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31B-it-qat-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 359.8 11.3 23.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yshuqghe",
   "source_url": "https://omlx.ai/benchmarks/performance/yshuqghe",
   "tg_tokens": null,
   "tps": 11.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-yshuqghe-ctx16384",
   "model": "gemma-4-31B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 316.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31B-it-qat-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 316.5 10.2 25.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yshuqghe",
   "source_url": "https://omlx.ai/benchmarks/performance/yshuqghe",
   "tg_tokens": null,
   "tps": 10.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2",
   "id": "omlx-d444rknc-ctx1024",
   "model": "Bonsai-4B",
   "notes": "mlx build; oMLX community benchmark trace, single run at 1k context; peak memory 1.5 GB; TTFT 3940 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 259.9,
   "provenance": "community",
   "quant": "1bit",
   "quote": "Bonsai-4B-mlx-1bit on M2 (10c); M2 (10c) \u00b7 16 GB \u00b7 unknown; 1k tokens 259.9 PP tok/s 43.7 TG tok/s 3940 TTFT (ms) 1.5 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d444rknc",
   "source_url": "https://omlx.ai/benchmarks/performance/d444rknc",
   "tg_tokens": null,
   "tps": 43.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-kj1rtvi0-ctx1024",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: dvgx6i8z, o6m31amq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 986.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Cyber-Tiel-Coder-35B-A3BoQ6e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 986.1 76.9 27.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kj1rtvi0",
   "source_url": "https://omlx.ai/benchmarks/performance/kj1rtvi0",
   "tg_tokens": null,
   "tps": 76.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-kj1rtvi0-ctx4096",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: dvgx6i8z, o6m31amq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1751.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Cyber-Tiel-Coder-35B-A3BoQ6e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 1,751 74.3 28.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kj1rtvi0",
   "source_url": "https://omlx.ai/benchmarks/performance/kj1rtvi0",
   "tg_tokens": null,
   "tps": 74.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-kj1rtvi0-ctx8192",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 29.0 GB at this context; duplicate traces of the same session: dvgx6i8z, o6m31amq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2169.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Cyber-Tiel-Coder-35B-A3BoQ6e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 8k 2,169 71.8 29.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kj1rtvi0",
   "source_url": "https://omlx.ai/benchmarks/performance/kj1rtvi0",
   "tg_tokens": null,
   "tps": 71.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-a14oetsr-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 90.0 GB at this context; duplicate traces of the same session: gtc3fu4s, nvnddnyp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 87.4,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 87.4 10.6 90.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a14oetsr",
   "source_url": "https://omlx.ai/benchmarks/performance/a14oetsr",
   "tg_tokens": null,
   "tps": 10.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-a14oetsr-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.0 GB at this context; duplicate traces of the same session: gtc3fu4s, nvnddnyp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 286.9,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 286.9 11.0 92.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a14oetsr",
   "source_url": "https://omlx.ai/benchmarks/performance/a14oetsr",
   "tg_tokens": null,
   "tps": 11.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-a14oetsr-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.2 GB at this context; duplicate traces of the same session: gtc3fu4s, nvnddnyp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 391.7,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 391.7 13.7 92.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a14oetsr",
   "source_url": "https://omlx.ai/benchmarks/performance/a14oetsr",
   "tg_tokens": null,
   "tps": 13.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-oy6svrj8-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 90.0 GB at this context; duplicate traces of the same session: gqresa9j, ppkj5k2m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 89.9,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 89.9 9.6 90.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oy6svrj8",
   "source_url": "https://omlx.ai/benchmarks/performance/oy6svrj8",
   "tg_tokens": null,
   "tps": 9.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-oy6svrj8-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.2 GB at this context; duplicate traces of the same session: gqresa9j, ppkj5k2m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 306.4,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 306.4 9.9 92.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oy6svrj8",
   "source_url": "https://omlx.ai/benchmarks/performance/oy6svrj8",
   "tg_tokens": null,
   "tps": 9.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-oy6svrj8-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.3 GB at this context; duplicate traces of the same session: gqresa9j, ppkj5k2m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 389.9,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 389.9 13.5 92.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oy6svrj8",
   "source_url": "https://omlx.ai/benchmarks/performance/oy6svrj8",
   "tg_tokens": null,
   "tps": 13.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3fyijvc8-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 90.0 GB at this context; duplicate traces of the same session: 8le2unsq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 88.6,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 88.6 10.3 90.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3fyijvc8",
   "source_url": "https://omlx.ai/benchmarks/performance/3fyijvc8",
   "tg_tokens": null,
   "tps": 10.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3fyijvc8-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.2 GB at this context; duplicate traces of the same session: 8le2unsq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 327.4,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 327.4 10.2 92.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3fyijvc8",
   "source_url": "https://omlx.ai/benchmarks/performance/3fyijvc8",
   "tg_tokens": null,
   "tps": 10.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3fyijvc8-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.3 GB at this context; duplicate traces of the same session: 8le2unsq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 391.7,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 391.7 12.9 92.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3fyijvc8",
   "source_url": "https://omlx.ai/benchmarks/performance/3fyijvc8",
   "tg_tokens": null,
   "tps": 12.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-unuo4ftj-ctx65536",
   "model": "MiniCPM-V-4.6",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 6.3 GB; TTFT 16951 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3866.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "MiniCPM-V-4.6 on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 bf16; 64k tokens 3,866 PP tok/s 105.0 TG tok/s 16951 TTFT (ms) 6.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/unuo4ftj",
   "source_url": "https://omlx.ai/benchmarks/performance/unuo4ftj",
   "tg_tokens": null,
   "tps": 105.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Max",
   "id": "omlx-0tumyrvs-ctx1024",
   "model": "Ornith-1.5-35B-A3B-uncensored",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: pa8xwf17; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 203.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-uncensored-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 203.9 17.1 6.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0tumyrvs",
   "source_url": "https://omlx.ai/benchmarks/performance/0tumyrvs",
   "tg_tokens": null,
   "tps": 17.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-0tumyrvs-ctx4096",
   "model": "Ornith-1.5-35B-A3B-uncensored",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: pa8xwf17; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 363.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-uncensored-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 363.2 14.4 7.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0tumyrvs",
   "source_url": "https://omlx.ai/benchmarks/performance/0tumyrvs",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-8uas43um-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.1 GB at this context; duplicate traces of the same session: iq4axw2v; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 591.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 1k 591.1 60.6 43.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8uas43um",
   "source_url": "https://omlx.ai/benchmarks/performance/8uas43um",
   "tg_tokens": null,
   "tps": 60.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-8uas43um-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.9 GB at this context; duplicate traces of the same session: iq4axw2v; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 672.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k 672.7 60.2 43.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8uas43um",
   "source_url": "https://omlx.ai/benchmarks/performance/8uas43um",
   "tg_tokens": null,
   "tps": 60.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-lgdl1jxx-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.1 GB at this context; duplicate traces of the same session: 1moyq3rd, 6x26ny86, 9qyio5fb, kg8ij5wq, q6ahhz1x; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1678.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 1,678 73.7 43.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lgdl1jxx",
   "source_url": "https://omlx.ai/benchmarks/performance/lgdl1jxx",
   "tg_tokens": null,
   "tps": 73.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-lgdl1jxx-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.9 GB at this context; duplicate traces of the same session: 1moyq3rd, 6x26ny86, 9qyio5fb, kg8ij5wq, q6ahhz1x; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2347.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,347 75.7 43.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lgdl1jxx",
   "source_url": "https://omlx.ai/benchmarks/performance/lgdl1jxx",
   "tg_tokens": null,
   "tps": 75.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-lgdl1jxx-ctx8192",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 44.0 GB at this context; duplicate traces of the same session: 1moyq3rd, 6x26ny86, 9qyio5fb, kg8ij5wq, q6ahhz1x; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2278.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 2,278 66.1 44.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lgdl1jxx",
   "source_url": "https://omlx.ai/benchmarks/performance/lgdl1jxx",
   "tg_tokens": null,
   "tps": 66.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-lgdl1jxx-ctx32768",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 44.6 GB at this context; duplicate traces of the same session: 1moyq3rd, 6x26ny86, 9qyio5fb, kg8ij5wq, q6ahhz1x; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1815.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 1,815 61.1 44.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lgdl1jxx",
   "source_url": "https://omlx.ai/benchmarks/performance/lgdl1jxx",
   "tg_tokens": null,
   "tps": 61.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-lgdl1jxx-ctx65536",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 45.3 GB at this context; duplicate traces of the same session: 1moyq3rd, 6x26ny86, 9qyio5fb, kg8ij5wq, q6ahhz1x; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1382.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 1,382 49.7 45.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lgdl1jxx",
   "source_url": "https://omlx.ai/benchmarks/performance/lgdl1jxx",
   "tg_tokens": null,
   "tps": 49.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-lgdl1jxx-ctx131072",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 46.8 GB at this context; duplicate traces of the same session: 1moyq3rd, 6x26ny86, 9qyio5fb, kg8ij5wq, q6ahhz1x; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1014.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 128k 1,014 37.4 46.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lgdl1jxx",
   "source_url": "https://omlx.ai/benchmarks/performance/lgdl1jxx",
   "tg_tokens": null,
   "tps": 37.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-tlxuryg5-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: rz4l22iv; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 777.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 1k 777.1 89.5 37.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tlxuryg5",
   "source_url": "https://omlx.ai/benchmarks/performance/tlxuryg5",
   "tg_tokens": null,
   "tps": 89.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-tlxuryg5-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.4 GB at this context; duplicate traces of the same session: rz4l22iv; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 848.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 4k 848.3 78.6 39.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tlxuryg5",
   "source_url": "https://omlx.ai/benchmarks/performance/tlxuryg5",
   "tg_tokens": null,
   "tps": 78.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-vjz8br0v-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: ymjpqd2o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 217.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 217.5 22.1 15.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vjz8br0v",
   "source_url": "https://omlx.ai/benchmarks/performance/vjz8br0v",
   "tg_tokens": null,
   "tps": 22.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-vjz8br0v-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: ymjpqd2o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 108.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 108.0 9.0 21.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vjz8br0v",
   "source_url": "https://omlx.ai/benchmarks/performance/vjz8br0v",
   "tg_tokens": null,
   "tps": 9.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-0kcuzeek-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: d2i2aups; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 84.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 84.8 15.7 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0kcuzeek",
   "source_url": "https://omlx.ai/benchmarks/performance/0kcuzeek",
   "tg_tokens": null,
   "tps": 15.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-0kcuzeek-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: d2i2aups; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 79.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 79.1 15.2 20.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0kcuzeek",
   "source_url": "https://omlx.ai/benchmarks/performance/0kcuzeek",
   "tg_tokens": null,
   "tps": 15.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-m2hv8jw9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 18.8 GB; TTFT 1095 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 935.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 935.2 PP tok/s 33.5 TG tok/s 1095 TTFT (ms) 18.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m2hv8jw9",
   "source_url": "https://omlx.ai/benchmarks/performance/m2hv8jw9",
   "tg_tokens": null,
   "tps": 33.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-xlq9wrw0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 18.5 GB; TTFT 13253 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 309.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 4k tokens 309.1 PP tok/s 20.8 TG tok/s 13253 TTFT (ms) 18.5 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xlq9wrw0",
   "source_url": "https://omlx.ai/benchmarks/performance/xlq9wrw0",
   "tg_tokens": null,
   "tps": 20.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-zcbdsu0m-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: 063anf5j, perg5b49; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 335.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 335.9 22.9 32.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zcbdsu0m",
   "source_url": "https://omlx.ai/benchmarks/performance/zcbdsu0m",
   "tg_tokens": null,
   "tps": 22.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-zcbdsu0m-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.3 GB at this context; duplicate traces of the same session: 063anf5j, perg5b49; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 324.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 16k 324.1 21.8 33.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zcbdsu0m",
   "source_url": "https://omlx.ai/benchmarks/performance/zcbdsu0m",
   "tg_tokens": null,
   "tps": 21.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-zcbdsu0m-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 39.4 GB at this context; duplicate traces of the same session: 063anf5j, perg5b49; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 266.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 64k 266.7 18.9 39.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zcbdsu0m",
   "source_url": "https://omlx.ai/benchmarks/performance/zcbdsu0m",
   "tg_tokens": null,
   "tps": 18.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-sapw2kp0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 322.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 322.0 57.1 23.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sapw2kp0",
   "source_url": "https://omlx.ai/benchmarks/performance/sapw2kp0",
   "tg_tokens": null,
   "tps": 57.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-sapw2kp0-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 303.2,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 16k 303.2 66.9 25.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sapw2kp0",
   "source_url": "https://omlx.ai/benchmarks/performance/sapw2kp0",
   "tg_tokens": null,
   "tps": 66.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-e540uz2c-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: zjdyfehh; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 1k 102.9 20.9 33.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e540uz2c",
   "source_url": "https://omlx.ai/benchmarks/performance/e540uz2c",
   "tg_tokens": null,
   "tps": 20.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-e540uz2c-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.5 GB at this context; duplicate traces of the same session: zjdyfehh; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 8bit; 4k 102.8 16.6 37.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e540uz2c",
   "source_url": "https://omlx.ai/benchmarks/performance/e540uz2c",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8xvtlw52-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 44.3 GB; TTFT 125149 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 523.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k tokens 523.7 PP tok/s 21.3 TG tok/s 125149 TTFT (ms) 44.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8xvtlw52",
   "source_url": "https://omlx.ai/benchmarks/performance/8xvtlw52",
   "tg_tokens": null,
   "tps": 21.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3vzch3ma-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 39.0 GB; TTFT 98712 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 663.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k tokens 663.9 PP tok/s 19.5 TG tok/s 98712 TTFT (ms) 39.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3vzch3ma",
   "source_url": "https://omlx.ai/benchmarks/performance/3vzch3ma",
   "tg_tokens": null,
   "tps": 19.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ec8kfhfq-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.2 GB at this context; duplicate traces of the same session: 311rs2so; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 472.7,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 1k 472.7 68.0 80.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ec8kfhfq",
   "source_url": "https://omlx.ai/benchmarks/performance/ec8kfhfq",
   "tg_tokens": null,
   "tps": 68.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ec8kfhfq-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.6 GB at this context; duplicate traces of the same session: 311rs2so; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 920.7,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 4k 920.7 73.2 80.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ec8kfhfq",
   "source_url": "https://omlx.ai/benchmarks/performance/ec8kfhfq",
   "tg_tokens": null,
   "tps": 73.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2rfd5taq-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 62.7 GB at this context; duplicate traces of the same session: m4hsotjq, pn83ahgw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1911.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 8k 1,911 59.9 62.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2rfd5taq",
   "source_url": "https://omlx.ai/benchmarks/performance/2rfd5taq",
   "tg_tokens": null,
   "tps": 59.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2rfd5taq-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 63.7 GB at this context; duplicate traces of the same session: m4hsotjq, pn83ahgw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2680.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 16k 2,680 74.3 63.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2rfd5taq",
   "source_url": "https://omlx.ai/benchmarks/performance/2rfd5taq",
   "tg_tokens": null,
   "tps": 74.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2rfd5taq-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 64.3 GB at this context; duplicate traces of the same session: m4hsotjq, pn83ahgw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2551.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 32k 2,551 78.1 64.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2rfd5taq",
   "source_url": "https://omlx.ai/benchmarks/performance/2rfd5taq",
   "tg_tokens": null,
   "tps": 78.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-yjeld8mv-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: iqs6f9ci, x2dggvjg; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1407.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 1,407 35.5 101.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yjeld8mv",
   "source_url": "https://omlx.ai/benchmarks/performance/yjeld8mv",
   "tg_tokens": null,
   "tps": 35.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-yjeld8mv-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.6 GB at this context; duplicate traces of the same session: iqs6f9ci, x2dggvjg; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1395.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 1,395 49.1 101.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yjeld8mv",
   "source_url": "https://omlx.ai/benchmarks/performance/yjeld8mv",
   "tg_tokens": null,
   "tps": 49.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-yjeld8mv-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.3 GB at this context; duplicate traces of the same session: iqs6f9ci, x2dggvjg; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1358.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 1,358 63.5 104.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yjeld8mv",
   "source_url": "https://omlx.ai/benchmarks/performance/yjeld8mv",
   "tg_tokens": null,
   "tps": 63.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-hes5cgr1-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 3649f3lf, nbetm7kx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 253.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 253.4 61.7 21.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hes5cgr1",
   "source_url": "https://omlx.ai/benchmarks/performance/hes5cgr1",
   "tg_tokens": null,
   "tps": 61.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-hes5cgr1-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: 3649f3lf, nbetm7kx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 300.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 300.5 75.2 25.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hes5cgr1",
   "source_url": "https://omlx.ai/benchmarks/performance/hes5cgr1",
   "tg_tokens": null,
   "tps": 75.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-hes5cgr1-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: 3649f3lf, nbetm7kx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 267.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 267.5 37.4 25.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hes5cgr1",
   "source_url": "https://omlx.ai/benchmarks/performance/hes5cgr1",
   "tg_tokens": null,
   "tps": 37.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-xcpgkah0-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 40.1 GB at this context; duplicate traces of the same session: mxqg7p22; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 250.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 250.5 34.4 40.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xcpgkah0",
   "source_url": "https://omlx.ai/benchmarks/performance/xcpgkah0",
   "tg_tokens": null,
   "tps": 34.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-xcpgkah0-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.3 GB at this context; duplicate traces of the same session: mxqg7p22; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 328.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 328.8 46.1 43.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xcpgkah0",
   "source_url": "https://omlx.ai/benchmarks/performance/xcpgkah0",
   "tg_tokens": null,
   "tps": 46.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-6epbz9rz-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "OptiQ build; oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 804.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-OptiQ-4bit on M2 Max (38c); M2 Max (38c) \u00b7 32 GB \u00b7 4bit; 1k 804.3 63.5 17.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6epbz9rz",
   "source_url": "https://omlx.ai/benchmarks/performance/6epbz9rz",
   "tg_tokens": null,
   "tps": 63.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-6epbz9rz-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "OptiQ build; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 809.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-OptiQ-4bit on M2 Max (38c); M2 Max (38c) \u00b7 32 GB \u00b7 4bit; 4k 809.9 61.6 17.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6epbz9rz",
   "source_url": "https://omlx.ai/benchmarks/performance/6epbz9rz",
   "tg_tokens": null,
   "tps": 61.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-8dk1zgiq-ctx1024",
   "model": "gemma-4-26B-A4B-it-uncensored-heretic-ara",
   "notes": "oMLX community benchmark trace; peak memory 15.1 GB at this context; duplicate traces of the same session: 5ub5lg8k, 7ptkjclh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1539.0,
   "provenance": "community",
   "quant": "mixed 4+6bit",
   "quote": "gemma-4-26B-A4B-it-uncensored-heretic-ara-MLX-4bit-mixed_4_6 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,539 108.7 15.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8dk1zgiq",
   "source_url": "https://omlx.ai/benchmarks/performance/8dk1zgiq",
   "tg_tokens": null,
   "tps": 108.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-8dk1zgiq-ctx4096",
   "model": "gemma-4-26B-A4B-it-uncensored-heretic-ara",
   "notes": "oMLX community benchmark trace; peak memory 15.4 GB at this context; duplicate traces of the same session: 5ub5lg8k, 7ptkjclh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1502.0,
   "provenance": "community",
   "quant": "mixed 4+6bit",
   "quote": "gemma-4-26B-A4B-it-uncensored-heretic-ara-MLX-4bit-mixed_4_6 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,502 106.6 15.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8dk1zgiq",
   "source_url": "https://omlx.ai/benchmarks/performance/8dk1zgiq",
   "tg_tokens": null,
   "tps": 106.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-8dk1zgiq-ctx8192",
   "model": "gemma-4-26B-A4B-it-uncensored-heretic-ara",
   "notes": "oMLX community benchmark trace; peak memory 15.6 GB at this context; duplicate traces of the same session: 5ub5lg8k, 7ptkjclh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1350.0,
   "provenance": "community",
   "quant": "mixed 4+6bit",
   "quote": "gemma-4-26B-A4B-it-uncensored-heretic-ara-MLX-4bit-mixed_4_6 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,350 101.8 15.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8dk1zgiq",
   "source_url": "https://omlx.ai/benchmarks/performance/8dk1zgiq",
   "tg_tokens": null,
   "tps": 101.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-1d716rzm-ctx1024",
   "model": "gemma-4-31B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: 5mk6uzic; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 365.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31B-it-qat-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 365.0 11.6 22.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1d716rzm",
   "source_url": "https://omlx.ai/benchmarks/performance/1d716rzm",
   "tg_tokens": null,
   "tps": 11.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-1d716rzm-ctx4096",
   "model": "gemma-4-31B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: 5mk6uzic; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 359.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31B-it-qat-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 359.8 11.3 23.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1d716rzm",
   "source_url": "https://omlx.ai/benchmarks/performance/1d716rzm",
   "tg_tokens": null,
   "tps": 11.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-1d716rzm-ctx16384",
   "model": "gemma-4-31B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: 5mk6uzic; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 316.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31B-it-qat-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 316.5 10.2 25.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1d716rzm",
   "source_url": "https://omlx.ai/benchmarks/performance/1d716rzm",
   "tg_tokens": null,
   "tps": 10.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ua82v53q-ctx16384",
   "model": "Occamy-1.0",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 21.1 GB; TTFT 6331 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2588.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "occamy-1.0-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k tokens 2,588 PP tok/s 52.3 TG tok/s 6331 TTFT (ms) 21.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ua82v53q",
   "source_url": "https://omlx.ai/benchmarks/performance/ua82v53q",
   "tg_tokens": null,
   "tps": 52.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-59db33ei-ctx1024",
   "model": "Occamy-1.0",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 20.2 GB; TTFT 269 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3802.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "occamy-1.0-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 3,802 PP tok/s 88.1 TG tok/s 269 TTFT (ms) 20.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/59db33ei",
   "source_url": "https://omlx.ai/benchmarks/performance/59db33ei",
   "tg_tokens": null,
   "tps": 88.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-by6ikmu8-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 90.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 88.6,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 88.6 10.3 90.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/by6ikmu8",
   "source_url": "https://omlx.ai/benchmarks/performance/by6ikmu8",
   "tg_tokens": null,
   "tps": 10.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-by6ikmu8-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 327.4,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 327.4 10.2 92.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/by6ikmu8",
   "source_url": "https://omlx.ai/benchmarks/performance/by6ikmu8",
   "tg_tokens": null,
   "tps": 10.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-by6ikmu8-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 391.7,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 391.7 12.9 92.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/by6ikmu8",
   "source_url": "https://omlx.ai/benchmarks/performance/by6ikmu8",
   "tg_tokens": null,
   "tps": 12.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xj4h3w42-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 90.0 GB at this context; duplicate traces of the same session: 406jf3ku, v15azui3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 90.1,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 90.1 9.5 90.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xj4h3w42",
   "source_url": "https://omlx.ai/benchmarks/performance/xj4h3w42",
   "tg_tokens": null,
   "tps": 9.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xj4h3w42-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.0 GB at this context; duplicate traces of the same session: 406jf3ku, v15azui3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 330.4,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 330.4 11.9 92.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xj4h3w42",
   "source_url": "https://omlx.ai/benchmarks/performance/xj4h3w42",
   "tg_tokens": null,
   "tps": 11.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xj4h3w42-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.2 GB at this context; duplicate traces of the same session: 406jf3ku, v15azui3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 386.7,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 386.7 13.1 92.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xj4h3w42",
   "source_url": "https://omlx.ai/benchmarks/performance/xj4h3w42",
   "tg_tokens": null,
   "tps": 13.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-unw5tk4z-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 90.0 GB at this context; duplicate traces of the same session: 0sd8y107, u830piby; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 88.4,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 88.4 9.7 90.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/unw5tk4z",
   "source_url": "https://omlx.ai/benchmarks/performance/unw5tk4z",
   "tg_tokens": null,
   "tps": 9.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-unw5tk4z-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.2 GB at this context; duplicate traces of the same session: 0sd8y107, u830piby; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 334.1,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 334.1 11.9 92.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/unw5tk4z",
   "source_url": "https://omlx.ai/benchmarks/performance/unw5tk4z",
   "tg_tokens": null,
   "tps": 11.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-unw5tk4z-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.0 GB at this context; duplicate traces of the same session: 0sd8y107, u830piby; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 390.3,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 390.3 12.1 92.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/unw5tk4z",
   "source_url": "https://omlx.ai/benchmarks/performance/unw5tk4z",
   "tg_tokens": null,
   "tps": 12.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nqvpzi9j-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 90.0 GB at this context; duplicate traces of the same session: mzmuwded, vme6s549; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 86.4,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 86.4 10.5 90.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nqvpzi9j",
   "source_url": "https://omlx.ai/benchmarks/performance/nqvpzi9j",
   "tg_tokens": null,
   "tps": 10.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nqvpzi9j-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 91.8 GB at this context; duplicate traces of the same session: mzmuwded, vme6s549; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 372.5,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 372.5 10.4 91.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nqvpzi9j",
   "source_url": "https://omlx.ai/benchmarks/performance/nqvpzi9j",
   "tg_tokens": null,
   "tps": 10.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nqvpzi9j-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.2 GB at this context; duplicate traces of the same session: mzmuwded, vme6s549; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 408.9,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 408.9 12.2 92.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nqvpzi9j",
   "source_url": "https://omlx.ai/benchmarks/performance/nqvpzi9j",
   "tg_tokens": null,
   "tps": 12.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qee37u2s-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 90.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 87.9,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 87.9 9.7 90.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qee37u2s",
   "source_url": "https://omlx.ai/benchmarks/performance/qee37u2s",
   "tg_tokens": null,
   "tps": 9.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qee37u2s-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 333.5,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 333.5 11.4 92.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qee37u2s",
   "source_url": "https://omlx.ai/benchmarks/performance/qee37u2s",
   "tg_tokens": null,
   "tps": 11.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qee37u2s-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 390.4,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 390.4 11.8 92.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qee37u2s",
   "source_url": "https://omlx.ai/benchmarks/performance/qee37u2s",
   "tg_tokens": null,
   "tps": 11.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-e6jf2k11-ctx1024",
   "model": "GLM-4.7-Flash",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: eu2pye0l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 621.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "GLM-4.7-Flash on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 621.8 43.5 23.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e6jf2k11",
   "source_url": "https://omlx.ai/benchmarks/performance/e6jf2k11",
   "tg_tokens": null,
   "tps": 43.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-e6jf2k11-ctx65536",
   "model": "GLM-4.7-Flash",
   "notes": "oMLX community benchmark trace; peak memory 47.4 GB at this context; duplicate traces of the same session: eu2pye0l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 120.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "GLM-4.7-Flash on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k 120.9 11.9 47.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e6jf2k11",
   "source_url": "https://omlx.ai/benchmarks/performance/e6jf2k11",
   "tg_tokens": null,
   "tps": 11.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-0lm01k1o-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.6 GB at this context; duplicate traces of the same session: k33336ej, q7lcbhgi, tzkwgd2l; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2368.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 2,368 167.2 174.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0lm01k1o",
   "source_url": "https://omlx.ai/benchmarks/performance/0lm01k1o",
   "tg_tokens": null,
   "tps": 167.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-0lm01k1o-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.3 GB at this context; duplicate traces of the same session: k33336ej, q7lcbhgi, tzkwgd2l; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2369.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,369 122.9 175.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0lm01k1o",
   "source_url": "https://omlx.ai/benchmarks/performance/0lm01k1o",
   "tg_tokens": null,
   "tps": 122.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-0lm01k1o-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.1 GB at this context; duplicate traces of the same session: k33336ej, q7lcbhgi, tzkwgd2l; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2345.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,345 91.2 177.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0lm01k1o",
   "source_url": "https://omlx.ai/benchmarks/performance/0lm01k1o",
   "tg_tokens": null,
   "tps": 91.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-0lm01k1o-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 180.5 GB at this context; duplicate traces of the same session: k33336ej, q7lcbhgi, tzkwgd2l; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2294.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,294 153.6 180.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0lm01k1o",
   "source_url": "https://omlx.ai/benchmarks/performance/0lm01k1o",
   "tg_tokens": null,
   "tps": 153.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-j2feexih-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.7 GB at this context; duplicate traces of the same session: okf42to1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 401.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 401.9 43.3 177.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j2feexih",
   "source_url": "https://omlx.ai/benchmarks/performance/j2feexih",
   "tg_tokens": null,
   "tps": 43.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-j2feexih-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.8 GB at this context; duplicate traces of the same session: okf42to1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 442.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 442.0 47.4 178.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j2feexih",
   "source_url": "https://omlx.ai/benchmarks/performance/j2feexih",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-9c86350r-ctx4096",
   "model": "Josiefied-Qwen3-4B-abliterated-v1",
   "notes": "oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: rps4bu38; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 651.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Josiefied-Qwen3-4B-abliterated-v1-4bit on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 4k 651.1 75.5 3.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9c86350r",
   "source_url": "https://omlx.ai/benchmarks/performance/9c86350r",
   "tg_tokens": null,
   "tps": 75.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-9c86350r-ctx16384",
   "model": "Josiefied-Qwen3-4B-abliterated-v1",
   "notes": "oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: rps4bu38; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 450.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Josiefied-Qwen3-4B-abliterated-v1-4bit on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 16k 450.8 47.4 4.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9c86350r",
   "source_url": "https://omlx.ai/benchmarks/performance/9c86350r",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-mgc6uu7n-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: y7n02ybk; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 917.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 1k 917.4 71.5 13.4 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mgc6uu7n",
   "source_url": "https://omlx.ai/benchmarks/performance/mgc6uu7n",
   "tg_tokens": null,
   "tps": 71.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-mgc6uu7n-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: y7n02ybk; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 989.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 4k 989.9 68.2 13.7 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mgc6uu7n",
   "source_url": "https://omlx.ai/benchmarks/performance/mgc6uu7n",
   "tg_tokens": null,
   "tps": 68.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-toqjo8wy-ctx32768",
   "model": "Qwen-AgentWorld-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 23.4 GB; TTFT 17545 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1868.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen-AgentWorld-35B-A3B-MLX-oQ4 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k tokens 1,868 PP tok/s 86.4 TG tok/s 17545 TTFT (ms) 23.4 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/toqjo8wy",
   "source_url": "https://omlx.ai/benchmarks/performance/toqjo8wy",
   "tg_tokens": null,
   "tps": 86.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-umfsmtsg-ctx1024",
   "model": "Qwen3 14B",
   "notes": "oMLX community benchmark trace; peak memory 8.5 GB at this context; duplicate traces of the same session: 5gmmntgg, engxcs8y, sr2v7ewl, txblu2oz, z4d86xok; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 941.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-14B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 941.1 33.0 8.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/umfsmtsg",
   "source_url": "https://omlx.ai/benchmarks/performance/umfsmtsg",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-umfsmtsg-ctx4096",
   "model": "Qwen3 14B",
   "notes": "oMLX community benchmark trace; peak memory 9.0 GB at this context; duplicate traces of the same session: 5gmmntgg, engxcs8y, sr2v7ewl, txblu2oz, z4d86xok; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 956.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-14B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 956.1 31.5 9.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/umfsmtsg",
   "source_url": "https://omlx.ai/benchmarks/performance/umfsmtsg",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-umfsmtsg-ctx8192",
   "model": "Qwen3 14B",
   "notes": "oMLX community benchmark trace; peak memory 9.6 GB at this context; duplicate traces of the same session: 5gmmntgg, engxcs8y, sr2v7ewl, txblu2oz, z4d86xok; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 934.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-14B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 934.1 28.6 9.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/umfsmtsg",
   "source_url": "https://omlx.ai/benchmarks/performance/umfsmtsg",
   "tg_tokens": null,
   "tps": 28.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-umfsmtsg-ctx16384",
   "model": "Qwen3 14B",
   "notes": "oMLX community benchmark trace; peak memory 10.8 GB at this context; duplicate traces of the same session: 5gmmntgg, engxcs8y, sr2v7ewl, txblu2oz, z4d86xok; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 831.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-14B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 831.7 25.1 10.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/umfsmtsg",
   "source_url": "https://omlx.ai/benchmarks/performance/umfsmtsg",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-umfsmtsg-ctx32768",
   "model": "Qwen3 14B",
   "notes": "oMLX community benchmark trace; peak memory 13.3 GB at this context; duplicate traces of the same session: 5gmmntgg, engxcs8y, sr2v7ewl, txblu2oz, z4d86xok; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 645.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-14B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 645.5 20.1 13.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/umfsmtsg",
   "source_url": "https://omlx.ai/benchmarks/performance/umfsmtsg",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-umfsmtsg-ctx65536",
   "model": "Qwen3 14B",
   "notes": "oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: 5gmmntgg, engxcs8y, sr2v7ewl, txblu2oz, z4d86xok; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 502.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-14B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 502.4 14.1 18.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/umfsmtsg",
   "source_url": "https://omlx.ai/benchmarks/performance/umfsmtsg",
   "tg_tokens": null,
   "tps": 14.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-umfsmtsg-ctx131072",
   "model": "Qwen3 14B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 5gmmntgg, engxcs8y, sr2v7ewl, txblu2oz, z4d86xok; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 332.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-14B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 128k 332.1 9.2 28.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/umfsmtsg",
   "source_url": "https://omlx.ai/benchmarks/performance/umfsmtsg",
   "tg_tokens": null,
   "tps": 9.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qrc966rt-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: k2tuasdj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2484.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,484 85.4 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qrc966rt",
   "source_url": "https://omlx.ai/benchmarks/performance/qrc966rt",
   "tg_tokens": null,
   "tps": 85.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qrc966rt-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: k2tuasdj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2308.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,308 76.7 20.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qrc966rt",
   "source_url": "https://omlx.ai/benchmarks/performance/qrc966rt",
   "tg_tokens": null,
   "tps": 76.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-urotv31h-ctx1024",
   "model": "Qwen3.6-35B-A3B-MTPLX-Optimized-Balance",
   "notes": "balance build; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 19y4bwti; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 501.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-MTPLX-Optimized-Balance on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 6bit; 1k 501.4 49.6 28.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/urotv31h",
   "source_url": "https://omlx.ai/benchmarks/performance/urotv31h",
   "tg_tokens": null,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-urotv31h-ctx4096",
   "model": "Qwen3.6-35B-A3B-MTPLX-Optimized-Balance",
   "notes": "balance build; oMLX community benchmark trace; peak memory 30.4 GB at this context; duplicate traces of the same session: 19y4bwti; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 526.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-MTPLX-Optimized-Balance on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 6bit; 4k 526.0 48.1 30.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/urotv31h",
   "source_url": "https://omlx.ai/benchmarks/performance/urotv31h",
   "tg_tokens": null,
   "tps": 48.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-vixskc8x-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: rd8zv6d4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1110.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,110 94.4 19.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vixskc8x",
   "source_url": "https://omlx.ai/benchmarks/performance/vixskc8x",
   "tg_tokens": null,
   "tps": 94.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-vixskc8x-ctx65536",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: rd8zv6d4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 791.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 791.9 58.6 21.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vixskc8x",
   "source_url": "https://omlx.ai/benchmarks/performance/vixskc8x",
   "tg_tokens": null,
   "tps": 58.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mpzyn672-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 19.2 GB; TTFT 20281 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 807.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k tokens 807.8 PP tok/s 41.4 TG tok/s 20281 TTFT (ms) 19.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mpzyn672",
   "source_url": "https://omlx.ai/benchmarks/performance/mpzyn672",
   "tg_tokens": null,
   "tps": 41.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-e7ic1354-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: 8rvs2xow, jai7w1ul, m6xnvl66; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 971.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 971.0 46.7 17.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e7ic1354",
   "source_url": "https://omlx.ai/benchmarks/performance/e7ic1354",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-e7ic1354-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: 8rvs2xow, jai7w1ul, m6xnvl66; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 894.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 894.2 50.5 18.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e7ic1354",
   "source_url": "https://omlx.ai/benchmarks/performance/e7ic1354",
   "tg_tokens": null,
   "tps": 50.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-e7ic1354-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: 8rvs2xow, jai7w1ul, m6xnvl66; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 783.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 783.5 51.1 19.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e7ic1354",
   "source_url": "https://omlx.ai/benchmarks/performance/e7ic1354",
   "tg_tokens": null,
   "tps": 51.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-e7ic1354-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 8rvs2xow, jai7w1ul, m6xnvl66; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 745.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 745.4 21.0 20.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e7ic1354",
   "source_url": "https://omlx.ai/benchmarks/performance/e7ic1354",
   "tg_tokens": null,
   "tps": 21.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-uj49p7nf-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 17.0 GB; TTFT 3373 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 303.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 1k tokens 303.6 PP tok/s 18.5 TG tok/s 3373 TTFT (ms) 17.0 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uj49p7nf",
   "source_url": "https://omlx.ai/benchmarks/performance/uj49p7nf",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ycjuw389-ctx1024",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: ddasu4in, vhxbopg5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1103.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,103 62.9 20.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ycjuw389",
   "source_url": "https://omlx.ai/benchmarks/performance/ycjuw389",
   "tg_tokens": null,
   "tps": 62.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ycjuw389-ctx4096",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: ddasu4in, vhxbopg5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1361.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,361 57.5 21.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ycjuw389",
   "source_url": "https://omlx.ai/benchmarks/performance/ycjuw389",
   "tg_tokens": null,
   "tps": 57.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ycjuw389-ctx8192",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: ddasu4in, vhxbopg5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1365.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,365 67.3 22.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ycjuw389",
   "source_url": "https://omlx.ai/benchmarks/performance/ycjuw389",
   "tg_tokens": null,
   "tps": 67.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-e2n75xjk-ctx1024",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: fhsng4m0, vttn0e1r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1094.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,094 62.6 20.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e2n75xjk",
   "source_url": "https://omlx.ai/benchmarks/performance/e2n75xjk",
   "tg_tokens": null,
   "tps": 62.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-e2n75xjk-ctx4096",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: fhsng4m0, vttn0e1r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1352.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,352 58.1 21.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e2n75xjk",
   "source_url": "https://omlx.ai/benchmarks/performance/e2n75xjk",
   "tg_tokens": null,
   "tps": 58.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-e2n75xjk-ctx8192",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: fhsng4m0, vttn0e1r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1372.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,372 59.2 22.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e2n75xjk",
   "source_url": "https://omlx.ai/benchmarks/performance/e2n75xjk",
   "tg_tokens": null,
   "tps": 59.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-ldkyoqam-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 322.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 322.0 57.1 23.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ldkyoqam",
   "source_url": "https://omlx.ai/benchmarks/performance/ldkyoqam",
   "tg_tokens": null,
   "tps": 57.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-ldkyoqam-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 303.2,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 16k 303.2 66.9 25.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ldkyoqam",
   "source_url": "https://omlx.ai/benchmarks/performance/ldkyoqam",
   "tg_tokens": null,
   "tps": 66.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-z5bh7amb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 18.5 GB; TTFT 9930 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k tokens 103.1 PP tok/s 21.8 TG tok/s 9930 TTFT (ms) 18.5 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z5bh7amb",
   "source_url": "https://omlx.ai/benchmarks/performance/z5bh7amb",
   "tg_tokens": null,
   "tps": 21.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-zo70a1vz-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 99v12z9f; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 100.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 100.6 21.5 17.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zo70a1vz",
   "source_url": "https://omlx.ai/benchmarks/performance/zo70a1vz",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-zo70a1vz-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: 99v12z9f; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 103.6 21.3 19.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zo70a1vz",
   "source_url": "https://omlx.ai/benchmarks/performance/zo70a1vz",
   "tg_tokens": null,
   "tps": 21.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-8rzcbgic-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 17.6 GB; TTFT 10201 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 100.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k tokens 100.4 PP tok/s 24.3 TG tok/s 10201 TTFT (ms) 17.6 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8rzcbgic",
   "source_url": "https://omlx.ai/benchmarks/performance/8rzcbgic",
   "tg_tokens": null,
   "tps": 24.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-svupthh1-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.8 GB; TTFT 1600 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 639.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k tokens 639.9 PP tok/s 72.9 TG tok/s 1600 TTFT (ms) 70.8 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/svupthh1",
   "source_url": "https://omlx.ai/benchmarks/performance/svupthh1",
   "tg_tokens": null,
   "tps": 72.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hlzag2gv-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.8 GB; TTFT 1814 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 564.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k tokens 564.6 PP tok/s 50.1 TG tok/s 1814 TTFT (ms) 70.8 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hlzag2gv",
   "source_url": "https://omlx.ai/benchmarks/performance/hlzag2gv",
   "tg_tokens": null,
   "tps": 50.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jwnolmt8-ctx4096",
   "model": "gemma-4-12b-it",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.3 GB at this context; duplicate traces of the same session: 8bfw4q3l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1462.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-12B-it-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 1,462 38.6 13.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jwnolmt8",
   "source_url": "https://omlx.ai/benchmarks/performance/jwnolmt8",
   "tg_tokens": null,
   "tps": 38.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jwnolmt8-ctx16384",
   "model": "gemma-4-12b-it",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.8 GB at this context; duplicate traces of the same session: 8bfw4q3l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1742.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-12B-it-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 16k 1,742 38.2 13.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jwnolmt8",
   "source_url": "https://omlx.ai/benchmarks/performance/jwnolmt8",
   "tg_tokens": null,
   "tps": 38.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-3zm5py5w-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: arrddmf2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1058.0,
   "provenance": "community",
   "quant": "OptiQ",
   "quote": "gemma-4-26B-A4B-it-OptiQ on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,058 60.7 17.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3zm5py5w",
   "source_url": "https://omlx.ai/benchmarks/performance/3zm5py5w",
   "tg_tokens": null,
   "tps": 60.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-3zm5py5w-ctx65536",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 30.2 GB at this context; duplicate traces of the same session: arrddmf2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 733.7,
   "provenance": "community",
   "quant": "OptiQ",
   "quote": "gemma-4-26B-A4B-it-OptiQ on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 733.7 36.0 30.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3zm5py5w",
   "source_url": "https://omlx.ai/benchmarks/performance/3zm5py5w",
   "tg_tokens": null,
   "tps": 36.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-bbrqc61l-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "OptiQ build; oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 804.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-OptiQ-4bit on M2 Max (38c); M2 Max (38c) \u00b7 32 GB \u00b7 4bit; 1k 804.3 63.5 17.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bbrqc61l",
   "source_url": "https://omlx.ai/benchmarks/performance/bbrqc61l",
   "tg_tokens": null,
   "tps": 63.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-bbrqc61l-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "OptiQ build; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 809.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-OptiQ-4bit on M2 Max (38c); M2 Max (38c) \u00b7 32 GB \u00b7 4bit; 4k 809.9 61.6 17.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bbrqc61l",
   "source_url": "https://omlx.ai/benchmarks/performance/bbrqc61l",
   "tg_tokens": null,
   "tps": 61.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2123aozu-ctx1024",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 80.0 GB at this context; duplicate traces of the same session: oii2i0l9, wqdgwffu; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 416.1,
   "provenance": "community",
   "quant": "2.4bit mixed",
   "quote": "DeepSeek-V4-Flash-0731-2.4bit-mixed on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 1k 416.1 30.9 80.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2123aozu",
   "source_url": "https://omlx.ai/benchmarks/performance/2123aozu",
   "tg_tokens": null,
   "tps": 30.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2123aozu-ctx16384",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 80.5 GB at this context; duplicate traces of the same session: oii2i0l9, wqdgwffu; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 449.3,
   "provenance": "community",
   "quant": "2.4bit mixed",
   "quote": "DeepSeek-V4-Flash-0731-2.4bit-mixed on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 16k 449.3 26.7 80.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2123aozu",
   "source_url": "https://omlx.ai/benchmarks/performance/2123aozu",
   "tg_tokens": null,
   "tps": 26.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2123aozu-ctx131072",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 81.3 GB at this context; duplicate traces of the same session: oii2i0l9, wqdgwffu; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 377.6,
   "provenance": "community",
   "quant": "2.4bit mixed",
   "quote": "DeepSeek-V4-Flash-0731-2.4bit-mixed on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 128k 377.6 23.8 81.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2123aozu",
   "source_url": "https://omlx.ai/benchmarks/performance/2123aozu",
   "tg_tokens": null,
   "tps": 23.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ow2cp9j1-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 90.0 GB at this context; duplicate traces of the same session: 13g7iozc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 87.9,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 87.9 9.7 90.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ow2cp9j1",
   "source_url": "https://omlx.ai/benchmarks/performance/ow2cp9j1",
   "tg_tokens": null,
   "tps": 9.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ow2cp9j1-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.2 GB at this context; duplicate traces of the same session: 13g7iozc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 333.5,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 333.5 11.4 92.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ow2cp9j1",
   "source_url": "https://omlx.ai/benchmarks/performance/ow2cp9j1",
   "tg_tokens": null,
   "tps": 11.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ow2cp9j1-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.0 GB at this context; duplicate traces of the same session: 13g7iozc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 390.4,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 390.4 11.8 92.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ow2cp9j1",
   "source_url": "https://omlx.ai/benchmarks/performance/ow2cp9j1",
   "tg_tokens": null,
   "tps": 11.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-x53w3u3n-ctx1024",
   "model": "LFM2.5-2.6B",
   "notes": "oMLX community benchmark trace; peak memory 2.4 GB at this context; duplicate traces of the same session: 83fakks0, ei9ap5j4, tsieqg90, v10oxaut; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3950.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-2.6B-oQ4e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 3,950 151.9 2.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x53w3u3n",
   "source_url": "https://omlx.ai/benchmarks/performance/x53w3u3n",
   "tg_tokens": null,
   "tps": 151.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-x53w3u3n-ctx4096",
   "model": "LFM2.5-2.6B",
   "notes": "oMLX community benchmark trace; peak memory 2.6 GB at this context; duplicate traces of the same session: 83fakks0, ei9ap5j4, tsieqg90, v10oxaut; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4518.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-2.6B-oQ4e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 4,518 137.8 2.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x53w3u3n",
   "source_url": "https://omlx.ai/benchmarks/performance/x53w3u3n",
   "tg_tokens": null,
   "tps": 137.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-x53w3u3n-ctx8192",
   "model": "LFM2.5-2.6B",
   "notes": "oMLX community benchmark trace; peak memory 2.7 GB at this context; duplicate traces of the same session: 83fakks0, ei9ap5j4, tsieqg90, v10oxaut; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4461.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-2.6B-oQ4e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 4,461 135.3 2.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x53w3u3n",
   "source_url": "https://omlx.ai/benchmarks/performance/x53w3u3n",
   "tg_tokens": null,
   "tps": 135.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-x53w3u3n-ctx16384",
   "model": "LFM2.5-2.6B",
   "notes": "oMLX community benchmark trace; peak memory 2.7 GB at this context; duplicate traces of the same session: 83fakks0, ei9ap5j4, tsieqg90, v10oxaut; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4451.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-2.6B-oQ4e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 4,451 126.6 2.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x53w3u3n",
   "source_url": "https://omlx.ai/benchmarks/performance/x53w3u3n",
   "tg_tokens": null,
   "tps": 126.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-x53w3u3n-ctx32768",
   "model": "LFM2.5-2.6B",
   "notes": "oMLX community benchmark trace; peak memory 2.9 GB at this context; duplicate traces of the same session: 83fakks0, ei9ap5j4, tsieqg90, v10oxaut; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4155.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-2.6B-oQ4e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 4,155 104.5 2.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x53w3u3n",
   "source_url": "https://omlx.ai/benchmarks/performance/x53w3u3n",
   "tg_tokens": null,
   "tps": 104.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-x53w3u3n-ctx65536",
   "model": "LFM2.5-2.6B",
   "notes": "oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: 83fakks0, ei9ap5j4, tsieqg90, v10oxaut; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3513.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-2.6B-oQ4e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 3,513 82.9 3.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x53w3u3n",
   "source_url": "https://omlx.ai/benchmarks/performance/x53w3u3n",
   "tg_tokens": null,
   "tps": 82.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-x53w3u3n-ctx131072",
   "model": "LFM2.5-2.6B",
   "notes": "oMLX community benchmark trace; peak memory 4.4 GB at this context; duplicate traces of the same session: 83fakks0, ei9ap5j4, tsieqg90, v10oxaut; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2672.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-2.6B-oQ4e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 128k 2,672 54.0 4.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x53w3u3n",
   "source_url": "https://omlx.ai/benchmarks/performance/x53w3u3n",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-owfuy0p4-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: bxfq1md9, whv52dlg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 965.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 965.6 111.2 13.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/owfuy0p4",
   "source_url": "https://omlx.ai/benchmarks/performance/owfuy0p4",
   "tg_tokens": null,
   "tps": 111.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-owfuy0p4-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 14.2 GB at this context; duplicate traces of the same session: bxfq1md9, whv52dlg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 850.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 850.2 70.0 14.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/owfuy0p4",
   "source_url": "https://omlx.ai/benchmarks/performance/owfuy0p4",
   "tg_tokens": null,
   "tps": 70.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-owfuy0p4-ctx65536",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 14.9 GB at this context; duplicate traces of the same session: bxfq1md9, whv52dlg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 703.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 64k 703.0 49.6 14.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/owfuy0p4",
   "source_url": "https://omlx.ai/benchmarks/performance/owfuy0p4",
   "tg_tokens": null,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-967w6ro1-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: b6qvypl5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 950.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 950.3 109.6 13.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/967w6ro1",
   "source_url": "https://omlx.ai/benchmarks/performance/967w6ro1",
   "tg_tokens": null,
   "tps": 109.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-967w6ro1-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: b6qvypl5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1031.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 1,031 106.6 13.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/967w6ro1",
   "source_url": "https://omlx.ai/benchmarks/performance/967w6ro1",
   "tg_tokens": null,
   "tps": 106.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-967w6ro1-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.8 GB at this context; duplicate traces of the same session: b6qvypl5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1006.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 1,006 99.9 13.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/967w6ro1",
   "source_url": "https://omlx.ai/benchmarks/performance/967w6ro1",
   "tg_tokens": null,
   "tps": 99.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-c2wsfpvg-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: x3s9po0a; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 953.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 953.6 109.7 13.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c2wsfpvg",
   "source_url": "https://omlx.ai/benchmarks/performance/c2wsfpvg",
   "tg_tokens": null,
   "tps": 109.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-c2wsfpvg-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: x3s9po0a; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1028.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 1,028 106.5 13.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c2wsfpvg",
   "source_url": "https://omlx.ai/benchmarks/performance/c2wsfpvg",
   "tg_tokens": null,
   "tps": 106.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-c2wsfpvg-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.8 GB at this context; duplicate traces of the same session: x3s9po0a; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1006.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 1,006 99.8 13.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c2wsfpvg",
   "source_url": "https://omlx.ai/benchmarks/performance/c2wsfpvg",
   "tg_tokens": null,
   "tps": 99.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-13fxr9zf-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX FP16 build; oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 922.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit-FP16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 922.4 90.3 13.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/13fxr9zf",
   "source_url": "https://omlx.ai/benchmarks/performance/13fxr9zf",
   "tg_tokens": null,
   "tps": 90.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-13fxr9zf-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX FP16 build; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1035.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit-FP16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 1,035 87.7 13.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/13fxr9zf",
   "source_url": "https://omlx.ai/benchmarks/performance/13fxr9zf",
   "tg_tokens": null,
   "tps": 87.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-axcq985b-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: fwaoa1oa; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1745.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "LFM2.5-24B-A2B-MLX-6bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 6bit; 1k 1,745 120.1 18.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/axcq985b",
   "source_url": "https://omlx.ai/benchmarks/performance/axcq985b",
   "tg_tokens": null,
   "tps": 120.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-axcq985b-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: fwaoa1oa; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1952.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "LFM2.5-24B-A2B-MLX-6bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 6bit; 4k 1,952 116.8 19.2 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/axcq985b",
   "source_url": "https://omlx.ai/benchmarks/performance/axcq985b",
   "tg_tokens": null,
   "tps": 116.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-vfavr5rs-ctx1024",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 2.6 GB at this context; duplicate traces of the same session: go6wyqa8; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 667.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM5-2B-OptiQ-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 16 GB \u00b7 4bit; 1k 667.6 74.2 2.6 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vfavr5rs",
   "source_url": "https://omlx.ai/benchmarks/performance/vfavr5rs",
   "tg_tokens": null,
   "tps": 74.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-vfavr5rs-ctx4096",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 2.7 GB at this context; duplicate traces of the same session: go6wyqa8; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 610.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM5-2B-OptiQ-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 16 GB \u00b7 4bit; 4k 610.0 64.0 2.7 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vfavr5rs",
   "source_url": "https://omlx.ai/benchmarks/performance/vfavr5rs",
   "tg_tokens": null,
   "tps": 64.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-c94f08sn-ctx1024",
   "model": "Qwen3 14B",
   "notes": "oMLX community benchmark trace; peak memory 8.5 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 941.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-14B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 941.1 33.0 8.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c94f08sn",
   "source_url": "https://omlx.ai/benchmarks/performance/c94f08sn",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-c94f08sn-ctx4096",
   "model": "Qwen3 14B",
   "notes": "oMLX community benchmark trace; peak memory 9.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 956.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-14B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 956.1 31.5 9.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c94f08sn",
   "source_url": "https://omlx.ai/benchmarks/performance/c94f08sn",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-c94f08sn-ctx8192",
   "model": "Qwen3 14B",
   "notes": "oMLX community benchmark trace; peak memory 9.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 934.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-14B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 934.1 28.6 9.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c94f08sn",
   "source_url": "https://omlx.ai/benchmarks/performance/c94f08sn",
   "tg_tokens": null,
   "tps": 28.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-c94f08sn-ctx16384",
   "model": "Qwen3 14B",
   "notes": "oMLX community benchmark trace; peak memory 10.8 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 831.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-14B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 831.7 25.1 10.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c94f08sn",
   "source_url": "https://omlx.ai/benchmarks/performance/c94f08sn",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-c94f08sn-ctx32768",
   "model": "Qwen3 14B",
   "notes": "oMLX community benchmark trace; peak memory 13.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 645.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-14B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 645.5 20.1 13.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c94f08sn",
   "source_url": "https://omlx.ai/benchmarks/performance/c94f08sn",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-c94f08sn-ctx65536",
   "model": "Qwen3 14B",
   "notes": "oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 502.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-14B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 502.4 14.1 18.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c94f08sn",
   "source_url": "https://omlx.ai/benchmarks/performance/c94f08sn",
   "tg_tokens": null,
   "tps": 14.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-c94f08sn-ctx131072",
   "model": "Qwen3 14B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 332.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-14B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 128k 332.1 9.2 28.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c94f08sn",
   "source_url": "https://omlx.ai/benchmarks/performance/c94f08sn",
   "tg_tokens": null,
   "tps": 9.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-yzzdq3y3-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 34.6 GB at this context; duplicate traces of the same session: ynd8euap; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 876.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 16k 876.4 48.3 34.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yzzdq3y3",
   "source_url": "https://omlx.ai/benchmarks/performance/yzzdq3y3",
   "tg_tokens": null,
   "tps": 48.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-yzzdq3y3-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: ynd8euap; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 708.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 32k 708.2 56.6 35.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yzzdq3y3",
   "source_url": "https://omlx.ai/benchmarks/performance/yzzdq3y3",
   "tg_tokens": null,
   "tps": 56.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5stu9v2m-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 2s7xzmfa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1003.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,003 63.5 18.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5stu9v2m",
   "source_url": "https://omlx.ai/benchmarks/performance/5stu9v2m",
   "tg_tokens": null,
   "tps": 63.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5stu9v2m-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 2s7xzmfa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 763.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 763.7 41.6 19.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5stu9v2m",
   "source_url": "https://omlx.ai/benchmarks/performance/5stu9v2m",
   "tg_tokens": null,
   "tps": 41.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-z7c01r6l-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 2patf64o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 925.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 925.2 54.0 18.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z7c01r6l",
   "source_url": "https://omlx.ai/benchmarks/performance/z7c01r6l",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-z7c01r6l-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 2patf64o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 705.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 705.2 8.2 21.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z7c01r6l",
   "source_url": "https://omlx.ai/benchmarks/performance/z7c01r6l",
   "tg_tokens": null,
   "tps": 8.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0sh1kmd7-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: 7764spq6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 731.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 731.7 11.8 20.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0sh1kmd7",
   "source_url": "https://omlx.ai/benchmarks/performance/0sh1kmd7",
   "tg_tokens": null,
   "tps": 11.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0sh1kmd7-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: 7764spq6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 618.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 618.1 3.4 25.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0sh1kmd7",
   "source_url": "https://omlx.ai/benchmarks/performance/0sh1kmd7",
   "tg_tokens": null,
   "tps": 3.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-55prg9w9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: t50pcjus; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 302.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 1k 302.4 21.9 17.1 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/55prg9w9",
   "source_url": "https://omlx.ai/benchmarks/performance/55prg9w9",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-55prg9w9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: t50pcjus; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 309.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 4k 309.6 23.7 18.7 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/55prg9w9",
   "source_url": "https://omlx.ai/benchmarks/performance/55prg9w9",
   "tg_tokens": null,
   "tps": 23.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-w0t2jfzy-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.0 GB at this context; duplicate traces of the same session: p10t15rk, ue9vnl1b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 795.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 795.8 27.2 18.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w0t2jfzy",
   "source_url": "https://omlx.ai/benchmarks/performance/w0t2jfzy",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-w0t2jfzy-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: p10t15rk, ue9vnl1b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 679.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 679.1 24.2 19.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w0t2jfzy",
   "source_url": "https://omlx.ai/benchmarks/performance/w0t2jfzy",
   "tg_tokens": null,
   "tps": 24.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-w0t2jfzy-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: p10t15rk, ue9vnl1b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 380.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 380.4 14.5 21.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w0t2jfzy",
   "source_url": "https://omlx.ai/benchmarks/performance/w0t2jfzy",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lzxhk7lw-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 73.5 GB at this context; duplicate traces of the same session: 0fzeu3tg, bdd4kkt5, szgzneno; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1069.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,069 57.9 73.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lzxhk7lw",
   "source_url": "https://omlx.ai/benchmarks/performance/lzxhk7lw",
   "tg_tokens": null,
   "tps": 57.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lzxhk7lw-ctx8192",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 74.9 GB at this context; duplicate traces of the same session: 0fzeu3tg, bdd4kkt5, szgzneno; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2695.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,695 58.6 74.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lzxhk7lw",
   "source_url": "https://omlx.ai/benchmarks/performance/lzxhk7lw",
   "tg_tokens": null,
   "tps": 58.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lzxhk7lw-ctx65536",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 77.6 GB at this context; duplicate traces of the same session: 0fzeu3tg, bdd4kkt5, szgzneno; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2915.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,915 61.6 77.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lzxhk7lw",
   "source_url": "https://omlx.ai/benchmarks/performance/lzxhk7lw",
   "tg_tokens": null,
   "tps": 61.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lzxhk7lw-ctx131072",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 79.9 GB at this context; duplicate traces of the same session: 0fzeu3tg, bdd4kkt5, szgzneno; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2798.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,798 63.4 79.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lzxhk7lw",
   "source_url": "https://omlx.ai/benchmarks/performance/lzxhk7lw",
   "tg_tokens": null,
   "tps": 63.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-v3phwf2c-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "100K MTP plefp8 build; oMLX community benchmark trace; peak memory 73.5 GB at this context; duplicate traces of the same session: 8qtda37m, ahtvb93h, qbiiiy5o, xtqzy2u7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1464.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP-plefp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,464 63.2 73.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v3phwf2c",
   "source_url": "https://omlx.ai/benchmarks/performance/v3phwf2c",
   "tg_tokens": null,
   "tps": 63.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-v3phwf2c-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "100K MTP plefp8 build; oMLX community benchmark trace; peak memory 74.2 GB at this context; duplicate traces of the same session: 8qtda37m, ahtvb93h, qbiiiy5o, xtqzy2u7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2173.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP-plefp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,173 73.3 74.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v3phwf2c",
   "source_url": "https://omlx.ai/benchmarks/performance/v3phwf2c",
   "tg_tokens": null,
   "tps": 73.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-v3phwf2c-ctx8192",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "100K MTP plefp8 build; oMLX community benchmark trace; peak memory 74.9 GB at this context; duplicate traces of the same session: 8qtda37m, ahtvb93h, qbiiiy5o, xtqzy2u7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2820.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP-plefp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,820 56.2 74.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v3phwf2c",
   "source_url": "https://omlx.ai/benchmarks/performance/v3phwf2c",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-v3phwf2c-ctx65536",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "100K MTP plefp8 build; oMLX community benchmark trace; peak memory 77.6 GB at this context; duplicate traces of the same session: 8qtda37m, ahtvb93h, qbiiiy5o, xtqzy2u7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2871.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP-plefp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,871 69.0 77.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v3phwf2c",
   "source_url": "https://omlx.ai/benchmarks/performance/v3phwf2c",
   "tg_tokens": null,
   "tps": 69.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-v3phwf2c-ctx131072",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "100K MTP plefp8 build; oMLX community benchmark trace; peak memory 79.9 GB at this context; duplicate traces of the same session: 8qtda37m, ahtvb93h, qbiiiy5o, xtqzy2u7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2629.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP-plefp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,629 61.9 79.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v3phwf2c",
   "source_url": "https://omlx.ai/benchmarks/performance/v3phwf2c",
   "tg_tokens": null,
   "tps": 61.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-o22ehg7i-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "100K MTP plefp8 build; oMLX community benchmark trace; peak memory 73.5 GB at this context; duplicate traces of the same session: 0jz18ydx, jg8oq0ft, r81cgagu, sxex27x4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1247.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP-plefp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,247 82.5 73.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o22ehg7i",
   "source_url": "https://omlx.ai/benchmarks/performance/o22ehg7i",
   "tg_tokens": null,
   "tps": 82.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-o22ehg7i-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "100K MTP plefp8 build; oMLX community benchmark trace; peak memory 74.1 GB at this context; duplicate traces of the same session: 0jz18ydx, jg8oq0ft, r81cgagu, sxex27x4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2449.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP-plefp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,449 91.8 74.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o22ehg7i",
   "source_url": "https://omlx.ai/benchmarks/performance/o22ehg7i",
   "tg_tokens": null,
   "tps": 91.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-o22ehg7i-ctx8192",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "100K MTP plefp8 build; oMLX community benchmark trace; peak memory 74.9 GB at this context; duplicate traces of the same session: 0jz18ydx, jg8oq0ft, r81cgagu, sxex27x4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3186.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP-plefp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 3,186 72.4 74.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o22ehg7i",
   "source_url": "https://omlx.ai/benchmarks/performance/o22ehg7i",
   "tg_tokens": null,
   "tps": 72.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-o22ehg7i-ctx65536",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "100K MTP plefp8 build; oMLX community benchmark trace; peak memory 77.6 GB at this context; duplicate traces of the same session: 0jz18ydx, jg8oq0ft, r81cgagu, sxex27x4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2964.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP-plefp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,964 53.7 77.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o22ehg7i",
   "source_url": "https://omlx.ai/benchmarks/performance/o22ehg7i",
   "tg_tokens": null,
   "tps": 53.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-o22ehg7i-ctx131072",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "100K MTP plefp8 build; oMLX community benchmark trace; peak memory 79.9 GB at this context; duplicate traces of the same session: 0jz18ydx, jg8oq0ft, r81cgagu, sxex27x4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2824.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP-plefp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,824 68.4 79.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o22ehg7i",
   "source_url": "https://omlx.ai/benchmarks/performance/o22ehg7i",
   "tg_tokens": null,
   "tps": 68.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-0u1ybsvw-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: or3xzwz6, u4rb4pfx, wt1h0qwn; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 687.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 687.3 59.7 70.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0u1ybsvw",
   "source_url": "https://omlx.ai/benchmarks/performance/0u1ybsvw",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-0u1ybsvw-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.1 GB at this context; duplicate traces of the same session: or3xzwz6, u4rb4pfx, wt1h0qwn; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1057.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 16k 1,057 90.5 72.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0u1ybsvw",
   "source_url": "https://omlx.ai/benchmarks/performance/0u1ybsvw",
   "tg_tokens": null,
   "tps": 90.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-0u1ybsvw-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 76.4 GB at this context; duplicate traces of the same session: or3xzwz6, u4rb4pfx, wt1h0qwn; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1031.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k 1,031 45.8 76.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0u1ybsvw",
   "source_url": "https://omlx.ai/benchmarks/performance/0u1ybsvw",
   "tg_tokens": null,
   "tps": 45.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-0u1ybsvw-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 82.7 GB at this context; duplicate traces of the same session: or3xzwz6, u4rb4pfx, wt1h0qwn; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 965.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 128k 965.9 47.1 82.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0u1ybsvw",
   "source_url": "https://omlx.ai/benchmarks/performance/0u1ybsvw",
   "tg_tokens": null,
   "tps": 47.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-q1bhisqr-ctx1024",
   "model": "Qwopus3.8-27B-Flash-V2",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: hxb2okyc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 780.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwopus3.8-27B-Flash-V2-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 780.8 17.8 28.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q1bhisqr",
   "source_url": "https://omlx.ai/benchmarks/performance/q1bhisqr",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-q1bhisqr-ctx4096",
   "model": "Qwopus3.8-27B-Flash-V2",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: hxb2okyc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 882.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwopus3.8-27B-Flash-V2-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 882.3 17.6 29.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q1bhisqr",
   "source_url": "https://omlx.ai/benchmarks/performance/q1bhisqr",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-0k8cih50-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 950.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 950.3 109.6 13.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0k8cih50",
   "source_url": "https://omlx.ai/benchmarks/performance/0k8cih50",
   "tg_tokens": null,
   "tps": 109.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-0k8cih50-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1031.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 1,031 106.6 13.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0k8cih50",
   "source_url": "https://omlx.ai/benchmarks/performance/0k8cih50",
   "tg_tokens": null,
   "tps": 106.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-0k8cih50-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.8 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1006.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 1,006 99.9 13.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0k8cih50",
   "source_url": "https://omlx.ai/benchmarks/performance/0k8cih50",
   "tg_tokens": null,
   "tps": 99.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-tvols3ex-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 953.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 953.6 109.7 13.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tvols3ex",
   "source_url": "https://omlx.ai/benchmarks/performance/tvols3ex",
   "tg_tokens": null,
   "tps": 109.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-tvols3ex-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1028.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 1,028 106.5 13.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tvols3ex",
   "source_url": "https://omlx.ai/benchmarks/performance/tvols3ex",
   "tg_tokens": null,
   "tps": 106.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-tvols3ex-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.8 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1006.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 1,006 99.8 13.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tvols3ex",
   "source_url": "https://omlx.ai/benchmarks/performance/tvols3ex",
   "tg_tokens": null,
   "tps": 99.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-etbah9d7-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: c40wwsdu, zy9us3nx; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 966.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 966.0 110.3 13.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/etbah9d7",
   "source_url": "https://omlx.ai/benchmarks/performance/etbah9d7",
   "tg_tokens": null,
   "tps": 110.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-etbah9d7-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: c40wwsdu, zy9us3nx; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1030.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 1,030 107.5 13.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/etbah9d7",
   "source_url": "https://omlx.ai/benchmarks/performance/etbah9d7",
   "tg_tokens": null,
   "tps": 107.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-etbah9d7-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.8 GB at this context; duplicate traces of the same session: c40wwsdu, zy9us3nx; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1007.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 1,007 100.5 13.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/etbah9d7",
   "source_url": "https://omlx.ai/benchmarks/performance/etbah9d7",
   "tg_tokens": null,
   "tps": 100.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-773690tw-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 1k context; peak memory 13.4 GB; TTFT 1084 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 944.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k tokens 944.8 PP tok/s 110.9 TG tok/s 1084 TTFT (ms) 13.4 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/773690tw",
   "source_url": "https://omlx.ai/benchmarks/performance/773690tw",
   "tg_tokens": null,
   "tps": 110.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-qdfddnj5-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX FP16 build; oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 922.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit-FP16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 922.4 90.3 13.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qdfddnj5",
   "source_url": "https://omlx.ai/benchmarks/performance/qdfddnj5",
   "tg_tokens": null,
   "tps": 90.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-qdfddnj5-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX FP16 build; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1035.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit-FP16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 1,035 87.7 13.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qdfddnj5",
   "source_url": "https://omlx.ai/benchmarks/performance/qdfddnj5",
   "tg_tokens": null,
   "tps": 87.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-bbsmaiwg-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 13.6 GB at this context; duplicate traces of the same session: jv0gppf9; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 922.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e-fp16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 922.8 82.0 13.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bbsmaiwg",
   "source_url": "https://omlx.ai/benchmarks/performance/bbsmaiwg",
   "tg_tokens": null,
   "tps": 82.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-bbsmaiwg-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 13.9 GB at this context; duplicate traces of the same session: jv0gppf9; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1029.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e-fp16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 1,029 79.3 13.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bbsmaiwg",
   "source_url": "https://omlx.ai/benchmarks/performance/bbsmaiwg",
   "tg_tokens": null,
   "tps": 79.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-hrcz9qvc-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: 0h0dkeoo, es6swwk3, ezun2akn, gcfsl718, hg06cg3a, rgva6rir; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2254.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "LFM2.5-24B-A2B-oQ5e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 5bit; 1k 2,254 104.5 16.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hrcz9qvc",
   "source_url": "https://omlx.ai/benchmarks/performance/hrcz9qvc",
   "tg_tokens": null,
   "tps": 104.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-hrcz9qvc-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 0h0dkeoo, es6swwk3, ezun2akn, gcfsl718, hg06cg3a, rgva6rir; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3688.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "LFM2.5-24B-A2B-oQ5e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 5bit; 4k 3,688 103.8 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hrcz9qvc",
   "source_url": "https://omlx.ai/benchmarks/performance/hrcz9qvc",
   "tg_tokens": null,
   "tps": 103.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-hrcz9qvc-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 0h0dkeoo, es6swwk3, ezun2akn, gcfsl718, hg06cg3a, rgva6rir; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3691.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "LFM2.5-24B-A2B-oQ5e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 5bit; 8k 3,691 99.5 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hrcz9qvc",
   "source_url": "https://omlx.ai/benchmarks/performance/hrcz9qvc",
   "tg_tokens": null,
   "tps": 99.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-hrcz9qvc-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 16.8 GB at this context; duplicate traces of the same session: 0h0dkeoo, es6swwk3, ezun2akn, gcfsl718, hg06cg3a, rgva6rir; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3749.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "LFM2.5-24B-A2B-oQ5e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 5bit; 16k 3,749 91.6 16.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hrcz9qvc",
   "source_url": "https://omlx.ai/benchmarks/performance/hrcz9qvc",
   "tg_tokens": null,
   "tps": 91.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-hrcz9qvc-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: 0h0dkeoo, es6swwk3, ezun2akn, gcfsl718, hg06cg3a, rgva6rir; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3465.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "LFM2.5-24B-A2B-oQ5e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 5bit; 32k 3,465 79.5 17.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hrcz9qvc",
   "source_url": "https://omlx.ai/benchmarks/performance/hrcz9qvc",
   "tg_tokens": null,
   "tps": 79.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-hrcz9qvc-ctx65536",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 0h0dkeoo, es6swwk3, ezun2akn, gcfsl718, hg06cg3a, rgva6rir; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2962.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "LFM2.5-24B-A2B-oQ5e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 5bit; 64k 2,962 61.8 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hrcz9qvc",
   "source_url": "https://omlx.ai/benchmarks/performance/hrcz9qvc",
   "tg_tokens": null,
   "tps": 61.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-hrcz9qvc-ctx131072",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 0h0dkeoo, es6swwk3, ezun2akn, gcfsl718, hg06cg3a, rgva6rir; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2148.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "LFM2.5-24B-A2B-oQ5e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 5bit; 128k 2,148 42.7 19.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hrcz9qvc",
   "source_url": "https://omlx.ai/benchmarks/performance/hrcz9qvc",
   "tg_tokens": null,
   "tps": 42.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-dygagvzn-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 1ye9to0o, 2x50u2tq, 41qfwtvf, 7t2kdbx3, 8ie5ni3f, ie5id3sf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2210.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "LFM2.5-24B-A2B-oQ6e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 2,210 100.3 18.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dygagvzn",
   "source_url": "https://omlx.ai/benchmarks/performance/dygagvzn",
   "tg_tokens": null,
   "tps": 100.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-dygagvzn-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 1ye9to0o, 2x50u2tq, 41qfwtvf, 7t2kdbx3, 8ie5ni3f, ie5id3sf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3511.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "LFM2.5-24B-A2B-oQ6e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 3,511 96.2 19.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dygagvzn",
   "source_url": "https://omlx.ai/benchmarks/performance/dygagvzn",
   "tg_tokens": null,
   "tps": 96.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-dygagvzn-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 1ye9to0o, 2x50u2tq, 41qfwtvf, 7t2kdbx3, 8ie5ni3f, ie5id3sf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3857.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "LFM2.5-24B-A2B-oQ6e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 8k 3,857 93.0 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dygagvzn",
   "source_url": "https://omlx.ai/benchmarks/performance/dygagvzn",
   "tg_tokens": null,
   "tps": 93.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-dygagvzn-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: 1ye9to0o, 2x50u2tq, 41qfwtvf, 7t2kdbx3, 8ie5ni3f, ie5id3sf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3713.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "LFM2.5-24B-A2B-oQ6e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 16k 3,713 84.5 19.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dygagvzn",
   "source_url": "https://omlx.ai/benchmarks/performance/dygagvzn",
   "tg_tokens": null,
   "tps": 84.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-dygagvzn-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 1ye9to0o, 2x50u2tq, 41qfwtvf, 7t2kdbx3, 8ie5ni3f, ie5id3sf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3258.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "LFM2.5-24B-A2B-oQ6e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 32k 3,258 75.2 19.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dygagvzn",
   "source_url": "https://omlx.ai/benchmarks/performance/dygagvzn",
   "tg_tokens": null,
   "tps": 75.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-dygagvzn-ctx65536",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: 1ye9to0o, 2x50u2tq, 41qfwtvf, 7t2kdbx3, 8ie5ni3f, ie5id3sf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2954.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "LFM2.5-24B-A2B-oQ6e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 64k 2,954 59.4 20.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dygagvzn",
   "source_url": "https://omlx.ai/benchmarks/performance/dygagvzn",
   "tg_tokens": null,
   "tps": 59.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-dygagvzn-ctx131072",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: 1ye9to0o, 2x50u2tq, 41qfwtvf, 7t2kdbx3, 8ie5ni3f, ie5id3sf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2164.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "LFM2.5-24B-A2B-oQ6e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 128k 2,164 41.0 21.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dygagvzn",
   "source_url": "https://omlx.ai/benchmarks/performance/dygagvzn",
   "tg_tokens": null,
   "tps": 41.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-bge8263c-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: f52s2syr, k7xev5xm, mhmb9rl2, oabth2ji, qpx8nn2x, yclxtgsd; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2969.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-24B-A2B-oQ8e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 2,969 77.6 24.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bge8263c",
   "source_url": "https://omlx.ai/benchmarks/performance/bge8263c",
   "tg_tokens": null,
   "tps": 77.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-bge8263c-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: f52s2syr, k7xev5xm, mhmb9rl2, oabth2ji, qpx8nn2x, yclxtgsd; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4295.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-24B-A2B-oQ8e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 4,295 78.7 24.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bge8263c",
   "source_url": "https://omlx.ai/benchmarks/performance/bge8263c",
   "tg_tokens": null,
   "tps": 78.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-bge8263c-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: f52s2syr, k7xev5xm, mhmb9rl2, oabth2ji, qpx8nn2x, yclxtgsd; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4179.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-24B-A2B-oQ8e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 4,179 80.2 24.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bge8263c",
   "source_url": "https://omlx.ai/benchmarks/performance/bge8263c",
   "tg_tokens": null,
   "tps": 80.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-bge8263c-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: f52s2syr, k7xev5xm, mhmb9rl2, oabth2ji, qpx8nn2x, yclxtgsd; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4099.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-24B-A2B-oQ8e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 16k 4,099 73.5 25.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bge8263c",
   "source_url": "https://omlx.ai/benchmarks/performance/bge8263c",
   "tg_tokens": null,
   "tps": 73.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-bge8263c-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: f52s2syr, k7xev5xm, mhmb9rl2, oabth2ji, qpx8nn2x, yclxtgsd; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3678.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-24B-A2B-oQ8e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 32k 3,678 64.2 25.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bge8263c",
   "source_url": "https://omlx.ai/benchmarks/performance/bge8263c",
   "tg_tokens": null,
   "tps": 64.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-bge8263c-ctx65536",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: f52s2syr, k7xev5xm, mhmb9rl2, oabth2ji, qpx8nn2x, yclxtgsd; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3096.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-24B-A2B-oQ8e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 64k 3,096 53.4 25.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bge8263c",
   "source_url": "https://omlx.ai/benchmarks/performance/bge8263c",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-bge8263c-ctx131072",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 27.2 GB at this context; duplicate traces of the same session: f52s2syr, k7xev5xm, mhmb9rl2, oabth2ji, qpx8nn2x, yclxtgsd; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2316.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-24B-A2B-oQ8e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 128k 2,316 30.5 27.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bge8263c",
   "source_url": "https://omlx.ai/benchmarks/performance/bge8263c",
   "tg_tokens": null,
   "tps": 30.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-x98mk19y-ctx1024",
   "model": "Qwen3 8B",
   "notes": "oMLX community benchmark trace; peak memory 5.0 GB at this context; duplicate traces of the same session: bcziody6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 467.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-8B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 467.5 51.0 5.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x98mk19y",
   "source_url": "https://omlx.ai/benchmarks/performance/x98mk19y",
   "tg_tokens": null,
   "tps": 51.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-x98mk19y-ctx4096",
   "model": "Qwen3 8B",
   "notes": "oMLX community benchmark trace; peak memory 5.5 GB at this context; duplicate traces of the same session: bcziody6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 448.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-8B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 448.2 45.6 5.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x98mk19y",
   "source_url": "https://omlx.ai/benchmarks/performance/x98mk19y",
   "tg_tokens": null,
   "tps": 45.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-one3tdh6-ctx16384",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 18.1 GB; TTFT 4109 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3988.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k tokens 3,988 PP tok/s 94.6 TG tok/s 4109 TTFT (ms) 18.1 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/one3tdh6",
   "source_url": "https://omlx.ai/benchmarks/performance/one3tdh6",
   "tg_tokens": null,
   "tps": 94.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-8yuvora4-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace, single run at 8k context; peak memory 33.7 GB; TTFT 8351 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 981.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 8k tokens 981.0 PP tok/s 52.1 TG tok/s 8351 TTFT (ms) 33.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8yuvora4",
   "source_url": "https://omlx.ai/benchmarks/performance/8yuvora4",
   "tg_tokens": null,
   "tps": 52.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-c41q3m0r-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace, single run at 1k context; peak memory 28.0 GB; TTFT 1281 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 799.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 1k tokens 799.7 PP tok/s 58.8 TG tok/s 1281 TTFT (ms) 28.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c41q3m0r",
   "source_url": "https://omlx.ai/benchmarks/performance/c41q3m0r",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-3rn4nz7y-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace, single run at 1k context; peak memory 28.0 GB; TTFT 1274 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 804.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 1k tokens 804.1 PP tok/s 60.5 TG tok/s 1274 TTFT (ms) 28.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3rn4nz7y",
   "source_url": "https://omlx.ai/benchmarks/performance/3rn4nz7y",
   "tg_tokens": null,
   "tps": 60.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-2vzi5ufa-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: 1db2fofx; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 463.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 463.3 35.9 18.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2vzi5ufa",
   "source_url": "https://omlx.ai/benchmarks/performance/2vzi5ufa",
   "tg_tokens": null,
   "tps": 35.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-2vzi5ufa-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: 1db2fofx; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 459.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 459.2 34.3 18.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2vzi5ufa",
   "source_url": "https://omlx.ai/benchmarks/performance/2vzi5ufa",
   "tg_tokens": null,
   "tps": 34.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-qlvgn46d-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 19.5 GB; TTFT 73375 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 223.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k tokens 223.3 PP tok/s 28.1 TG tok/s 73375 TTFT (ms) 19.5 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qlvgn46d",
   "source_url": "https://omlx.ai/benchmarks/performance/qlvgn46d",
   "tg_tokens": null,
   "tps": 28.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-so7hbtxm-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.3 GB at this context; duplicate traces of the same session: 08n9ctgg, 2mck25xd, uwbl9lyu, xgykp1oi, ztiptrp5; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 626.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 626.2 60.4 69.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/so7hbtxm",
   "source_url": "https://omlx.ai/benchmarks/performance/so7hbtxm",
   "tg_tokens": null,
   "tps": 60.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-so7hbtxm-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.0 GB at this context; duplicate traces of the same session: 08n9ctgg, 2mck25xd, uwbl9lyu, xgykp1oi, ztiptrp5; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1021.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 1,021 56.5 70.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/so7hbtxm",
   "source_url": "https://omlx.ai/benchmarks/performance/so7hbtxm",
   "tg_tokens": null,
   "tps": 56.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-so7hbtxm-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.5 GB at this context; duplicate traces of the same session: 08n9ctgg, 2mck25xd, uwbl9lyu, xgykp1oi, ztiptrp5; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1007.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 16k 1,007 59.5 70.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/so7hbtxm",
   "source_url": "https://omlx.ai/benchmarks/performance/so7hbtxm",
   "tg_tokens": null,
   "tps": 59.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-so7hbtxm-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.4 GB at this context; duplicate traces of the same session: 08n9ctgg, 2mck25xd, uwbl9lyu, xgykp1oi, ztiptrp5; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1022.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 1,022 61.1 71.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/so7hbtxm",
   "source_url": "https://omlx.ai/benchmarks/performance/so7hbtxm",
   "tg_tokens": null,
   "tps": 61.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-so7hbtxm-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.3 GB at this context; duplicate traces of the same session: 08n9ctgg, 2mck25xd, uwbl9lyu, xgykp1oi, ztiptrp5; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 990.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k 990.2 52.3 73.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/so7hbtxm",
   "source_url": "https://omlx.ai/benchmarks/performance/so7hbtxm",
   "tg_tokens": null,
   "tps": 52.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-so7hbtxm-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.3 GB at this context; duplicate traces of the same session: 08n9ctgg, 2mck25xd, uwbl9lyu, xgykp1oi, ztiptrp5; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 973.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 128k 973.0 44.1 78.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/so7hbtxm",
   "source_url": "https://omlx.ai/benchmarks/performance/so7hbtxm",
   "tg_tokens": null,
   "tps": 44.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-d747x7e6-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: r3ehkgfl, swzkjskl, uownw1b1, z0w1wjb5; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 677.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 677.8 62.9 70.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d747x7e6",
   "source_url": "https://omlx.ai/benchmarks/performance/d747x7e6",
   "tg_tokens": null,
   "tps": 62.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-d747x7e6-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: r3ehkgfl, swzkjskl, uownw1b1, z0w1wjb5; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1061.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 1,061 74.0 71.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d747x7e6",
   "source_url": "https://omlx.ai/benchmarks/performance/d747x7e6",
   "tg_tokens": null,
   "tps": 74.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-d747x7e6-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.3 GB at this context; duplicate traces of the same session: r3ehkgfl, swzkjskl, uownw1b1, z0w1wjb5; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1057.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 16k 1,057 73.8 72.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d747x7e6",
   "source_url": "https://omlx.ai/benchmarks/performance/d747x7e6",
   "tg_tokens": null,
   "tps": 73.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-d747x7e6-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 76.0 GB at this context; duplicate traces of the same session: r3ehkgfl, swzkjskl, uownw1b1, z0w1wjb5; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1037.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k 1,037 60.8 76.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d747x7e6",
   "source_url": "https://omlx.ai/benchmarks/performance/d747x7e6",
   "tg_tokens": null,
   "tps": 60.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-d747x7e6-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 82.6 GB at this context; duplicate traces of the same session: r3ehkgfl, swzkjskl, uownw1b1, z0w1wjb5; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 945.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 128k 945.2 41.3 82.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d747x7e6",
   "source_url": "https://omlx.ai/benchmarks/performance/d747x7e6",
   "tg_tokens": null,
   "tps": 41.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-00v6f97u-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 55.0 GB at this context; duplicate traces of the same session: 0co83ezy, 6oakh293, khohsqoc, tmnzluij, ya99frtn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 154.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 154.6 26.1 55.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/00v6f97u",
   "source_url": "https://omlx.ai/benchmarks/performance/00v6f97u",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-00v6f97u-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 55.8 GB at this context; duplicate traces of the same session: 0co83ezy, 6oakh293, khohsqoc, tmnzluij, ya99frtn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 791.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 791.5 34.4 55.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/00v6f97u",
   "source_url": "https://omlx.ai/benchmarks/performance/00v6f97u",
   "tg_tokens": null,
   "tps": 34.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-00v6f97u-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 57.7 GB at this context; duplicate traces of the same session: 0co83ezy, 6oakh293, khohsqoc, tmnzluij, ya99frtn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1114.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,114 31.5 57.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/00v6f97u",
   "source_url": "https://omlx.ai/benchmarks/performance/00v6f97u",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-00v6f97u-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 57.8 GB at this context; duplicate traces of the same session: 0co83ezy, 6oakh293, khohsqoc, tmnzluij, ya99frtn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1316.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,316 37.6 57.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/00v6f97u",
   "source_url": "https://omlx.ai/benchmarks/performance/00v6f97u",
   "tg_tokens": null,
   "tps": 37.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-00v6f97u-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 59.2 GB at this context; duplicate traces of the same session: 0co83ezy, 6oakh293, khohsqoc, tmnzluij, ya99frtn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1294.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 1,294 28.2 59.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/00v6f97u",
   "source_url": "https://omlx.ai/benchmarks/performance/00v6f97u",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-00v6f97u-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 65.6 GB at this context; duplicate traces of the same session: 0co83ezy, 6oakh293, khohsqoc, tmnzluij, ya99frtn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1227.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 1,227 24.8 65.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/00v6f97u",
   "source_url": "https://omlx.ai/benchmarks/performance/00v6f97u",
   "tg_tokens": null,
   "tps": 24.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-6rs0gnkp-ctx1024",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 94.4 GB at this context; duplicate traces of the same session: 4ohv1dpl, 5zujn3kr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 406.1,
   "provenance": "community",
   "quant": "2.4bit mixed",
   "quote": "DeepSeek-V4-Flash-0731-2.4bit-mixed on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 1k 406.1 17.4 94.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6rs0gnkp",
   "source_url": "https://omlx.ai/benchmarks/performance/6rs0gnkp",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-6rs0gnkp-ctx16384",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 88.4 GB at this context; duplicate traces of the same session: 4ohv1dpl, 5zujn3kr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 434.9,
   "provenance": "community",
   "quant": "2.4bit mixed",
   "quote": "DeepSeek-V4-Flash-0731-2.4bit-mixed on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 16k 434.9 37.4 88.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6rs0gnkp",
   "source_url": "https://omlx.ai/benchmarks/performance/6rs0gnkp",
   "tg_tokens": null,
   "tps": 37.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-6rs0gnkp-ctx131072",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 89.1 GB at this context; duplicate traces of the same session: 4ohv1dpl, 5zujn3kr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 367.3,
   "provenance": "community",
   "quant": "2.4bit mixed",
   "quote": "DeepSeek-V4-Flash-0731-2.4bit-mixed on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 128k 367.3 35.2 89.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6rs0gnkp",
   "source_url": "https://omlx.ai/benchmarks/performance/6rs0gnkp",
   "tg_tokens": null,
   "tps": 35.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-svt9zh5r-ctx4096",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 173.7 GB at this context; duplicate traces of the same session: drvpl34c, i45djpwt, mrkxj447; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 618.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 618.7 27.1 173.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/svt9zh5r",
   "source_url": "https://omlx.ai/benchmarks/performance/svt9zh5r",
   "tg_tokens": null,
   "tps": 27.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-svt9zh5r-ctx16384",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 173.9 GB at this context; duplicate traces of the same session: drvpl34c, i45djpwt, mrkxj447; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 609.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 609.6 25.8 173.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/svt9zh5r",
   "source_url": "https://omlx.ai/benchmarks/performance/svt9zh5r",
   "tg_tokens": null,
   "tps": 25.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-svt9zh5r-ctx65536",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 194.8 GB at this context; duplicate traces of the same session: drvpl34c, i45djpwt, mrkxj447; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 523.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 523.3 24.2 194.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/svt9zh5r",
   "source_url": "https://omlx.ai/benchmarks/performance/svt9zh5r",
   "tg_tokens": null,
   "tps": 24.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-svt9zh5r-ctx199680",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 198.1 GB at this context; duplicate traces of the same session: drvpl34c, i45djpwt, mrkxj447; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 393.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 393.7 22.5 198.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/svt9zh5r",
   "source_url": "https://omlx.ai/benchmarks/performance/svt9zh5r",
   "tg_tokens": null,
   "tps": 22.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lv3d457u-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 88.9 GB at this context; duplicate traces of the same session: 0uk6kp4f, 1jc69w16; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 96.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 96.0 9.8 88.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lv3d457u",
   "source_url": "https://omlx.ai/benchmarks/performance/lv3d457u",
   "tg_tokens": null,
   "tps": 9.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lv3d457u-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.2 GB at this context; duplicate traces of the same session: 0uk6kp4f, 1jc69w16; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 328.7,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 328.7 11.6 92.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lv3d457u",
   "source_url": "https://omlx.ai/benchmarks/performance/lv3d457u",
   "tg_tokens": null,
   "tps": 11.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lv3d457u-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 91.9 GB at this context; duplicate traces of the same session: 0uk6kp4f, 1jc69w16; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 437.4,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 437.4 12.9 91.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lv3d457u",
   "source_url": "https://omlx.ai/benchmarks/performance/lv3d457u",
   "tg_tokens": null,
   "tps": 12.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-qyc1fsd3-ctx4096",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 557.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-mlx-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 557.6 67.0 21.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qyc1fsd3",
   "source_url": "https://omlx.ai/benchmarks/performance/qyc1fsd3",
   "tg_tokens": null,
   "tps": 67.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-qyc1fsd3-ctx16384",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 456.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-mlx-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k 456.6 56.1 23.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qyc1fsd3",
   "source_url": "https://omlx.ai/benchmarks/performance/qyc1fsd3",
   "tg_tokens": null,
   "tps": 56.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-6too02pl-ctx1024",
   "model": "LFM2.5-2.6B",
   "notes": "oMLX community benchmark trace; peak memory 2.4 GB at this context; duplicate traces of the same session: bbb6zfon; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3950.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-2.6B-oQ4e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 3,950 151.9 2.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6too02pl",
   "source_url": "https://omlx.ai/benchmarks/performance/6too02pl",
   "tg_tokens": null,
   "tps": 151.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-6too02pl-ctx4096",
   "model": "LFM2.5-2.6B",
   "notes": "oMLX community benchmark trace; peak memory 2.6 GB at this context; duplicate traces of the same session: bbb6zfon; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4518.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-2.6B-oQ4e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 4,518 137.8 2.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6too02pl",
   "source_url": "https://omlx.ai/benchmarks/performance/6too02pl",
   "tg_tokens": null,
   "tps": 137.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-6too02pl-ctx8192",
   "model": "LFM2.5-2.6B",
   "notes": "oMLX community benchmark trace; peak memory 2.7 GB at this context; duplicate traces of the same session: bbb6zfon; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4461.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-2.6B-oQ4e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 4,461 135.3 2.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6too02pl",
   "source_url": "https://omlx.ai/benchmarks/performance/6too02pl",
   "tg_tokens": null,
   "tps": 135.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-6too02pl-ctx16384",
   "model": "LFM2.5-2.6B",
   "notes": "oMLX community benchmark trace; peak memory 2.7 GB at this context; duplicate traces of the same session: bbb6zfon; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4451.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-2.6B-oQ4e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 4,451 126.6 2.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6too02pl",
   "source_url": "https://omlx.ai/benchmarks/performance/6too02pl",
   "tg_tokens": null,
   "tps": 126.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-6too02pl-ctx32768",
   "model": "LFM2.5-2.6B",
   "notes": "oMLX community benchmark trace; peak memory 2.9 GB at this context; duplicate traces of the same session: bbb6zfon; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4155.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-2.6B-oQ4e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 4,155 104.5 2.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6too02pl",
   "source_url": "https://omlx.ai/benchmarks/performance/6too02pl",
   "tg_tokens": null,
   "tps": 104.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-6too02pl-ctx65536",
   "model": "LFM2.5-2.6B",
   "notes": "oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: bbb6zfon; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3513.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-2.6B-oQ4e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 3,513 82.9 3.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6too02pl",
   "source_url": "https://omlx.ai/benchmarks/performance/6too02pl",
   "tg_tokens": null,
   "tps": 82.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-6too02pl-ctx131072",
   "model": "LFM2.5-2.6B",
   "notes": "oMLX community benchmark trace; peak memory 4.4 GB at this context; duplicate traces of the same session: bbb6zfon; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2672.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-2.6B-oQ4e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 128k 2,672 54.0 4.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6too02pl",
   "source_url": "https://omlx.ai/benchmarks/performance/6too02pl",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-lh93tl0y-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX fp16 build; oMLX community benchmark trace, single run at 8k context; peak memory 13.8 GB; TTFT 12624 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 648.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit-fp16 on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 8k tokens 648.9 PP tok/s 59.2 TG tok/s 12624 TTFT (ms) 13.8 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lh93tl0y",
   "source_url": "https://omlx.ai/benchmarks/performance/lh93tl0y",
   "tg_tokens": null,
   "tps": 59.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-9virdwei-ctx1024",
   "model": "MiniCPM5-2B-chat",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 2.0 GB at this context; duplicate traces of the same session: 19fxhax7, hsksz1na, ozgitejs; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 793.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM5-2B-chat-mlx-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 793.0 68.9 2.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9virdwei",
   "source_url": "https://omlx.ai/benchmarks/performance/9virdwei",
   "tg_tokens": null,
   "tps": 68.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-9virdwei-ctx4096",
   "model": "MiniCPM5-2B-chat",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 2.1 GB at this context; duplicate traces of the same session: 19fxhax7, hsksz1na, ozgitejs; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 714.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM5-2B-chat-mlx-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 714.6 59.9 2.1 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9virdwei",
   "source_url": "https://omlx.ai/benchmarks/performance/9virdwei",
   "tg_tokens": null,
   "tps": 59.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-9virdwei-ctx8192",
   "model": "MiniCPM5-2B-chat",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: 19fxhax7, hsksz1na, ozgitejs; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 581.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM5-2B-chat-mlx-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 8k 581.6 19.8 6.2 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9virdwei",
   "source_url": "https://omlx.ai/benchmarks/performance/9virdwei",
   "tg_tokens": null,
   "tps": 19.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-9virdwei-ctx16384",
   "model": "MiniCPM5-2B-chat",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: 19fxhax7, hsksz1na, ozgitejs; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 490.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM5-2B-chat-mlx-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 16k 490.7 24.0 6.7 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9virdwei",
   "source_url": "https://omlx.ai/benchmarks/performance/9virdwei",
   "tg_tokens": null,
   "tps": 24.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-fhvhd54d-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: b965i3g1, h6ofjtmw, y5x9d4pf; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1921.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 1,921 136.4 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fhvhd54d",
   "source_url": "https://omlx.ai/benchmarks/performance/fhvhd54d",
   "tg_tokens": null,
   "tps": 136.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-fhvhd54d-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: b965i3g1, h6ofjtmw, y5x9d4pf; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2317.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 2,317 130.7 24.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fhvhd54d",
   "source_url": "https://omlx.ai/benchmarks/performance/fhvhd54d",
   "tg_tokens": null,
   "tps": 130.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-fhvhd54d-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: b965i3g1, h6ofjtmw, y5x9d4pf; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2163.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 8k 2,163 97.2 24.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fhvhd54d",
   "source_url": "https://omlx.ai/benchmarks/performance/fhvhd54d",
   "tg_tokens": null,
   "tps": 97.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-fhvhd54d-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.2 GB at this context; duplicate traces of the same session: b965i3g1, h6ofjtmw, y5x9d4pf; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1892.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 16k 1,892 169.8 25.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fhvhd54d",
   "source_url": "https://omlx.ai/benchmarks/performance/fhvhd54d",
   "tg_tokens": null,
   "tps": 169.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-fhvhd54d-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: b965i3g1, h6ofjtmw, y5x9d4pf; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1508.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 32k 1,508 156.6 26.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fhvhd54d",
   "source_url": "https://omlx.ai/benchmarks/performance/fhvhd54d",
   "tg_tokens": null,
   "tps": 156.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-fhvhd54d-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: b965i3g1, h6ofjtmw, y5x9d4pf; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1073.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 64k 1,073 91.9 27.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fhvhd54d",
   "source_url": "https://omlx.ai/benchmarks/performance/fhvhd54d",
   "tg_tokens": null,
   "tps": 91.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-fhvhd54d-ctx131072",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.1 GB at this context; duplicate traces of the same session: b965i3g1, h6ofjtmw, y5x9d4pf; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 688.5,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 128k 688.5 73.4 29.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fhvhd54d",
   "source_url": "https://omlx.ai/benchmarks/performance/fhvhd54d",
   "tg_tokens": null,
   "tps": 73.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-fhvhd54d-ctx199680",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.9 GB at this context; duplicate traces of the same session: b965i3g1, h6ofjtmw, y5x9d4pf; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 499.9,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 195k 499.9 50.9 30.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fhvhd54d",
   "source_url": "https://omlx.ai/benchmarks/performance/fhvhd54d",
   "tg_tokens": null,
   "tps": 50.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-vuf6xc2c-ctx4096",
   "model": "Qwen2.5-VL-3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 2.6 GB at this context; duplicate traces of the same session: jagm5wzy; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 327.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-VL-3B-Instruct-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 4k 327.5 35.0 2.6 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vuf6xc2c",
   "source_url": "https://omlx.ai/benchmarks/performance/vuf6xc2c",
   "tg_tokens": null,
   "tps": 35.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-vuf6xc2c-ctx16384",
   "model": "Qwen2.5-VL-3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 3.0 GB at this context; duplicate traces of the same session: jagm5wzy; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 337.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-VL-3B-Instruct-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 16k 337.7 44.8 3.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vuf6xc2c",
   "source_url": "https://omlx.ai/benchmarks/performance/vuf6xc2c",
   "tg_tokens": null,
   "tps": 44.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-331xwaco-ctx1024",
   "model": "Qwen3-4B-Instruct-2507",
   "notes": "oMLX community benchmark trace; peak memory 2.9 GB at this context; duplicate traces of the same session: 8gclgr75; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1554.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-4B-Instruct-2507-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,554 136.2 2.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/331xwaco",
   "source_url": "https://omlx.ai/benchmarks/performance/331xwaco",
   "tg_tokens": null,
   "tps": 136.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-331xwaco-ctx4096",
   "model": "Qwen3-4B-Instruct-2507",
   "notes": "oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: 8gclgr75; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1396.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-4B-Instruct-2507-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,396 119.8 3.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/331xwaco",
   "source_url": "https://omlx.ai/benchmarks/performance/331xwaco",
   "tg_tokens": null,
   "tps": 119.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-aya7byus-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: jmhlf8s6; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 234.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 234.8 21.8 5.8 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aya7byus",
   "source_url": "https://omlx.ai/benchmarks/performance/aya7byus",
   "tg_tokens": null,
   "tps": 21.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-aya7byus-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: jmhlf8s6; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 241.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 241.9 21.3 6.2 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aya7byus",
   "source_url": "https://omlx.ai/benchmarks/performance/aya7byus",
   "tg_tokens": null,
   "tps": 21.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-njbjvko8-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: a6wbkyqa; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 234.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 234.3 21.6 5.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/njbjvko8",
   "source_url": "https://omlx.ai/benchmarks/performance/njbjvko8",
   "tg_tokens": null,
   "tps": 21.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-njbjvko8-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: a6wbkyqa; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 240.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 240.8 21.2 6.2 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/njbjvko8",
   "source_url": "https://omlx.ai/benchmarks/performance/njbjvko8",
   "tg_tokens": null,
   "tps": 21.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ahr7ngn6-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 86bhgw6k; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 941.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 941.5 91.6 19.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ahr7ngn6",
   "source_url": "https://omlx.ai/benchmarks/performance/ahr7ngn6",
   "tg_tokens": null,
   "tps": 91.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ahr7ngn6-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 86bhgw6k; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 986.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 986.1 88.3 20.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ahr7ngn6",
   "source_url": "https://omlx.ai/benchmarks/performance/ahr7ngn6",
   "tg_tokens": null,
   "tps": 88.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-i3voo7x5-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: 2bqaym4a; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1420.0,
   "provenance": "community",
   "quant": "OptiQ-5bpw",
   "quote": "Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-OptiQ-5bpw-MLX on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 1,420 57.8 19.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i3voo7x5",
   "source_url": "https://omlx.ai/benchmarks/performance/i3voo7x5",
   "tg_tokens": null,
   "tps": 57.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-i3voo7x5-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 2bqaym4a; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1823.0,
   "provenance": "community",
   "quant": "OptiQ-5bpw",
   "quote": "Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-OptiQ-5bpw-MLX on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,823 56.5 20.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i3voo7x5",
   "source_url": "https://omlx.ai/benchmarks/performance/i3voo7x5",
   "tg_tokens": null,
   "tps": 56.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-fh09p71u-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 2c5y2ll1; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1941.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,941 56.0 19.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fh09p71u",
   "source_url": "https://omlx.ai/benchmarks/performance/fh09p71u",
   "tg_tokens": null,
   "tps": 56.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-fh09p71u-ctx16384",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 2c5y2ll1; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1766.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 1,766 52.7 20.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fh09p71u",
   "source_url": "https://omlx.ai/benchmarks/performance/fh09p71u",
   "tg_tokens": null,
   "tps": 52.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ljxp2qzo-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 380u39p0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 900.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 900.9 97.3 21.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ljxp2qzo",
   "source_url": "https://omlx.ai/benchmarks/performance/ljxp2qzo",
   "tg_tokens": null,
   "tps": 97.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ljxp2qzo-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: 380u39p0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 946.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 946.9 106.4 22.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ljxp2qzo",
   "source_url": "https://omlx.ai/benchmarks/performance/ljxp2qzo",
   "tg_tokens": null,
   "tps": 106.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-8wzopgjz-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: ia2l6w1w; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 890.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 890.2 109.1 21.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8wzopgjz",
   "source_url": "https://omlx.ai/benchmarks/performance/8wzopgjz",
   "tg_tokens": null,
   "tps": 109.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-8wzopgjz-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: ia2l6w1w; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 941.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 941.2 98.1 22.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8wzopgjz",
   "source_url": "https://omlx.ai/benchmarks/performance/8wzopgjz",
   "tg_tokens": null,
   "tps": 98.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-w8o9k592-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: q1qbilcu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 129.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 129.9 14.9 16.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w8o9k592",
   "source_url": "https://omlx.ai/benchmarks/performance/w8o9k592",
   "tg_tokens": null,
   "tps": 14.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-w8o9k592-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: q1qbilcu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 127.7 14.5 17.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w8o9k592",
   "source_url": "https://omlx.ai/benchmarks/performance/w8o9k592",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ckusyece-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: ezq8qjai; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1176.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 1,176 18.9 29.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ckusyece",
   "source_url": "https://omlx.ai/benchmarks/performance/ckusyece",
   "tg_tokens": null,
   "tps": 18.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ckusyece-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: ezq8qjai; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1317.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 1,317 18.1 30.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ckusyece",
   "source_url": "https://omlx.ai/benchmarks/performance/ckusyece",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-h27beu0d-ctx1024",
   "model": "Qwen3.8-27B-OBLITERATED",
   "notes": "oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: fb779yfw, zo5wjko3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 332.4,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-OBLITERATED-OptiQ-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 332.4 32.5 22.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h27beu0d",
   "source_url": "https://omlx.ai/benchmarks/performance/h27beu0d",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-h27beu0d-ctx4096",
   "model": "Qwen3.8-27B-OBLITERATED",
   "notes": "oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: fb779yfw, zo5wjko3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 425.6,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-OBLITERATED-OptiQ-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 425.6 32.1 25.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h27beu0d",
   "source_url": "https://omlx.ai/benchmarks/performance/h27beu0d",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-h27beu0d-ctx32768",
   "model": "Qwen3.8-27B-OBLITERATED",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: fb779yfw, zo5wjko3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 386.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-OBLITERATED-OptiQ-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 386.0 27.7 28.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h27beu0d",
   "source_url": "https://omlx.ai/benchmarks/performance/h27beu0d",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-eq4hvwsx-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: sbhxkqzp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 392.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 392.6 15.9 18.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eq4hvwsx",
   "source_url": "https://omlx.ai/benchmarks/performance/eq4hvwsx",
   "tg_tokens": null,
   "tps": 15.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-eq4hvwsx-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: sbhxkqzp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 369.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 369.7 14.3 19.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eq4hvwsx",
   "source_url": "https://omlx.ai/benchmarks/performance/eq4hvwsx",
   "tg_tokens": null,
   "tps": 14.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-s9c71umg-ctx131072",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 27.7 GB; TTFT 701799 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 186.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 128k tokens 186.8 PP tok/s 10.7 TG tok/s 701799 TTFT (ms) 27.7 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s9c71umg",
   "source_url": "https://omlx.ai/benchmarks/performance/s9c71umg",
   "tg_tokens": null,
   "tps": 10.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-inppecpt-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.3 GB at this context; duplicate traces of the same session: 3jeua6c1, a3trrb7u, ctfevc3w, lmxjumk3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 677.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 677.7 62.8 69.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/inppecpt",
   "source_url": "https://omlx.ai/benchmarks/performance/inppecpt",
   "tg_tokens": null,
   "tps": 62.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-inppecpt-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.0 GB at this context; duplicate traces of the same session: 3jeua6c1, a3trrb7u, ctfevc3w, lmxjumk3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1037.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 1,037 62.2 70.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/inppecpt",
   "source_url": "https://omlx.ai/benchmarks/performance/inppecpt",
   "tg_tokens": null,
   "tps": 62.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-inppecpt-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.5 GB at this context; duplicate traces of the same session: 3jeua6c1, a3trrb7u, ctfevc3w, lmxjumk3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1090.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 16k 1,090 67.5 70.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/inppecpt",
   "source_url": "https://omlx.ai/benchmarks/performance/inppecpt",
   "tg_tokens": null,
   "tps": 67.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-inppecpt-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.3 GB at this context; duplicate traces of the same session: 3jeua6c1, a3trrb7u, ctfevc3w, lmxjumk3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1045.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k 1,045 54.7 73.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/inppecpt",
   "source_url": "https://omlx.ai/benchmarks/performance/inppecpt",
   "tg_tokens": null,
   "tps": 54.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-inppecpt-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.3 GB at this context; duplicate traces of the same session: 3jeua6c1, a3trrb7u, ctfevc3w, lmxjumk3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1007.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 128k 1,007 51.3 78.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/inppecpt",
   "source_url": "https://omlx.ai/benchmarks/performance/inppecpt",
   "tg_tokens": null,
   "tps": 51.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-teh8i4yv-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: bqhmlxg8; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1826.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 4k 1,826 89.8 70.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/teh8i4yv",
   "source_url": "https://omlx.ai/benchmarks/performance/teh8i4yv",
   "tg_tokens": null,
   "tps": 89.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-teh8i4yv-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: bqhmlxg8; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2962.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 8k 2,962 60.9 71.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/teh8i4yv",
   "source_url": "https://omlx.ai/benchmarks/performance/teh8i4yv",
   "tg_tokens": null,
   "tps": 60.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-92t8hxop-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "Youssofal MTPLX Optimized Quality; oMLX community benchmark trace, single run at 64k context; peak memory 39.0 GB; TTFT 98415 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 665.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Youssofal--Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k tokens 665.9 PP tok/s 21.9 TG tok/s 98415 TTFT (ms) 39.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/92t8hxop",
   "source_url": "https://omlx.ai/benchmarks/performance/92t8hxop",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-l9v2pkp3-ctx1024",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 155.8 GB at this context; duplicate traces of the same session: iv0wdh1y, tywmke75; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 448.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 448.2 37.8 155.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l9v2pkp3",
   "source_url": "https://omlx.ai/benchmarks/performance/l9v2pkp3",
   "tg_tokens": null,
   "tps": 37.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-l9v2pkp3-ctx16384",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.4 GB at this context; duplicate traces of the same session: iv0wdh1y, tywmke75; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 465.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 465.8 47.1 156.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l9v2pkp3",
   "source_url": "https://omlx.ai/benchmarks/performance/l9v2pkp3",
   "tg_tokens": null,
   "tps": 47.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-l9v2pkp3-ctx131072",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 157.1 GB at this context; duplicate traces of the same session: iv0wdh1y, tywmke75; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 385.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 128k 385.4 35.2 157.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l9v2pkp3",
   "source_url": "https://omlx.ai/benchmarks/performance/l9v2pkp3",
   "tg_tokens": null,
   "tps": 35.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0vjqch24-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 90.0 GB at this context; duplicate traces of the same session: afw4lrpw, xuwh8lts; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 88.6,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 88.6 10.5 90.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0vjqch24",
   "source_url": "https://omlx.ai/benchmarks/performance/0vjqch24",
   "tg_tokens": null,
   "tps": 10.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0vjqch24-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.2 GB at this context; duplicate traces of the same session: afw4lrpw, xuwh8lts; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 320.2,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 320.2 10.6 92.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0vjqch24",
   "source_url": "https://omlx.ai/benchmarks/performance/0vjqch24",
   "tg_tokens": null,
   "tps": 10.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0vjqch24-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 91.9 GB at this context; duplicate traces of the same session: afw4lrpw, xuwh8lts; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 424.9,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 424.9 12.3 91.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0vjqch24",
   "source_url": "https://omlx.ai/benchmarks/performance/0vjqch24",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-nki95cla-ctx4096",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 557.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-mlx-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 557.6 67.0 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nki95cla",
   "source_url": "https://omlx.ai/benchmarks/performance/nki95cla",
   "tg_tokens": null,
   "tps": 67.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-nki95cla-ctx16384",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 456.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-mlx-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k 456.6 56.1 23.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nki95cla",
   "source_url": "https://omlx.ai/benchmarks/performance/nki95cla",
   "tg_tokens": null,
   "tps": 56.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ybqwarua-ctx1024",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 17.9 GB at this context; duplicate traces of the same session: t8d6oxoh, u5inhbgz, zpucxspw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 196.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 196.1 54.9 17.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ybqwarua",
   "source_url": "https://omlx.ai/benchmarks/performance/ybqwarua",
   "tg_tokens": null,
   "tps": 54.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ybqwarua-ctx4096",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 18.2 GB at this context; duplicate traces of the same session: t8d6oxoh, u5inhbgz, zpucxspw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 233.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 233.9 77.5 18.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ybqwarua",
   "source_url": "https://omlx.ai/benchmarks/performance/ybqwarua",
   "tg_tokens": null,
   "tps": 77.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ybqwarua-ctx8192",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: t8d6oxoh, u5inhbgz, zpucxspw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 236.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 236.5 38.8 18.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ybqwarua",
   "source_url": "https://omlx.ai/benchmarks/performance/ybqwarua",
   "tg_tokens": null,
   "tps": 38.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ybqwarua-ctx16384",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: t8d6oxoh, u5inhbgz, zpucxspw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 231.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 231.9 30.1 19.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ybqwarua",
   "source_url": "https://omlx.ai/benchmarks/performance/ybqwarua",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-jkn29b8e-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: oh3ks7c3, s4tm69u2, wty6ok0s; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1921.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 1,921 136.4 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jkn29b8e",
   "source_url": "https://omlx.ai/benchmarks/performance/jkn29b8e",
   "tg_tokens": null,
   "tps": 136.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-jkn29b8e-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: oh3ks7c3, s4tm69u2, wty6ok0s; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2317.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 2,317 130.7 24.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jkn29b8e",
   "source_url": "https://omlx.ai/benchmarks/performance/jkn29b8e",
   "tg_tokens": null,
   "tps": 130.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-jkn29b8e-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: oh3ks7c3, s4tm69u2, wty6ok0s; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2163.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 8k 2,163 97.2 24.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jkn29b8e",
   "source_url": "https://omlx.ai/benchmarks/performance/jkn29b8e",
   "tg_tokens": null,
   "tps": 97.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-jkn29b8e-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.2 GB at this context; duplicate traces of the same session: oh3ks7c3, s4tm69u2, wty6ok0s; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1892.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 16k 1,892 169.8 25.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jkn29b8e",
   "source_url": "https://omlx.ai/benchmarks/performance/jkn29b8e",
   "tg_tokens": null,
   "tps": 169.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-jkn29b8e-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: oh3ks7c3, s4tm69u2, wty6ok0s; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1508.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 32k 1,508 156.6 26.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jkn29b8e",
   "source_url": "https://omlx.ai/benchmarks/performance/jkn29b8e",
   "tg_tokens": null,
   "tps": 156.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-jkn29b8e-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: oh3ks7c3, s4tm69u2, wty6ok0s; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1073.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 64k 1,073 91.9 27.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jkn29b8e",
   "source_url": "https://omlx.ai/benchmarks/performance/jkn29b8e",
   "tg_tokens": null,
   "tps": 91.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-jkn29b8e-ctx131072",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.1 GB at this context; duplicate traces of the same session: oh3ks7c3, s4tm69u2, wty6ok0s; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 688.5,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 128k 688.5 73.4 29.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jkn29b8e",
   "source_url": "https://omlx.ai/benchmarks/performance/jkn29b8e",
   "tg_tokens": null,
   "tps": 73.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-jkn29b8e-ctx199680",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.9 GB at this context; duplicate traces of the same session: oh3ks7c3, s4tm69u2, wty6ok0s; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 499.9,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 195k 499.9 50.9 30.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jkn29b8e",
   "source_url": "https://omlx.ai/benchmarks/performance/jkn29b8e",
   "tg_tokens": null,
   "tps": 50.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-76vx73g2-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 08ryfp3u, cclnt49u, ybiwsvtd; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1911.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 1,911 134.4 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/76vx73g2",
   "source_url": "https://omlx.ai/benchmarks/performance/76vx73g2",
   "tg_tokens": null,
   "tps": 134.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-76vx73g2-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: 08ryfp3u, cclnt49u, ybiwsvtd; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2315.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 2,315 129.6 24.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/76vx73g2",
   "source_url": "https://omlx.ai/benchmarks/performance/76vx73g2",
   "tg_tokens": null,
   "tps": 129.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-76vx73g2-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: 08ryfp3u, cclnt49u, ybiwsvtd; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2164.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 8k 2,164 155.1 24.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/76vx73g2",
   "source_url": "https://omlx.ai/benchmarks/performance/76vx73g2",
   "tg_tokens": null,
   "tps": 155.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-76vx73g2-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.2 GB at this context; duplicate traces of the same session: 08ryfp3u, cclnt49u, ybiwsvtd; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1896.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 16k 1,896 171.0 25.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/76vx73g2",
   "source_url": "https://omlx.ai/benchmarks/performance/76vx73g2",
   "tg_tokens": null,
   "tps": 171.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-wk65aoo6-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: d1t1rqjy; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1721.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 1,721 148.8 22.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wk65aoo6",
   "source_url": "https://omlx.ai/benchmarks/performance/wk65aoo6",
   "tg_tokens": null,
   "tps": 148.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-wk65aoo6-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: d1t1rqjy; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2310.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 2,310 126.0 24.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wk65aoo6",
   "source_url": "https://omlx.ai/benchmarks/performance/wk65aoo6",
   "tg_tokens": null,
   "tps": 126.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-1jyw214o-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: mncfmlpo; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1902.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 1,902 133.8 22.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1jyw214o",
   "source_url": "https://omlx.ai/benchmarks/performance/1jyw214o",
   "tg_tokens": null,
   "tps": 133.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-1jyw214o-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: mncfmlpo; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2313.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 2,313 143.9 24.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1jyw214o",
   "source_url": "https://omlx.ai/benchmarks/performance/1jyw214o",
   "tg_tokens": null,
   "tps": 143.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-i8ecnfnb-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: 9ad772fk; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 429.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 429.5 39.0 3.4 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i8ecnfnb",
   "source_url": "https://omlx.ai/benchmarks/performance/i8ecnfnb",
   "tg_tokens": null,
   "tps": 39.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-i8ecnfnb-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.7 GB at this context; duplicate traces of the same session: 9ad772fk; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 439.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 439.3 37.7 3.7 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i8ecnfnb",
   "source_url": "https://omlx.ai/benchmarks/performance/i8ecnfnb",
   "tg_tokens": null,
   "tps": 37.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2",
   "id": "omlx-3oed4wzh-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: ihm88mxr; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M2 (8c); M2 (8c) \u00b7 16 GB \u00b7 4bit; 4k 104.7 17.6 6.3 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3oed4wzh",
   "source_url": "https://omlx.ai/benchmarks/performance/3oed4wzh",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2",
   "id": "omlx-3oed4wzh-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: ihm88mxr; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 77.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M2 (8c); M2 (8c) \u00b7 16 GB \u00b7 4bit; 16k 77.9 10.2 6.9 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3oed4wzh",
   "source_url": "https://omlx.ai/benchmarks/performance/3oed4wzh",
   "tg_tokens": null,
   "tps": 10.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-d14xgz1g-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.8 GB at this context; duplicate traces of the same session: 1ni75cjv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 125.8 8.5 31.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d14xgz1g",
   "source_url": "https://omlx.ai/benchmarks/performance/d14xgz1g",
   "tg_tokens": null,
   "tps": 8.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-d14xgz1g-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.6 GB at this context; duplicate traces of the same session: 1ni75cjv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 113.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 16k 113.8 8.2 33.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d14xgz1g",
   "source_url": "https://omlx.ai/benchmarks/performance/d14xgz1g",
   "tg_tokens": null,
   "tps": 8.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2gzaj4pz-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 31.6 GB; TTFT 27129 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1208.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k tokens 1,208 PP tok/s 17.4 TG tok/s 27129 TTFT (ms) 31.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2gzaj4pz",
   "source_url": "https://omlx.ai/benchmarks/performance/2gzaj4pz",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-6paynwbh-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: lc2oqqri, obbb3cvt; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 383.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 383.2 15.1 18.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6paynwbh",
   "source_url": "https://omlx.ai/benchmarks/performance/6paynwbh",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-6paynwbh-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: lc2oqqri, obbb3cvt; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 355.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 355.4 14.7 19.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6paynwbh",
   "source_url": "https://omlx.ai/benchmarks/performance/6paynwbh",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-6paynwbh-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: lc2oqqri, obbb3cvt; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 319.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 64k 319.8 12.9 21.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6paynwbh",
   "source_url": "https://omlx.ai/benchmarks/performance/6paynwbh",
   "tg_tokens": null,
   "tps": 12.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-edt7msc2-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.2 GB at this context; duplicate traces of the same session: bmqzr2hu, buu0cnu1, ue8rf90u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 245.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 245.8 33.3 39.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/edt7msc2",
   "source_url": "https://omlx.ai/benchmarks/performance/edt7msc2",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-edt7msc2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 42.4 GB at this context; duplicate traces of the same session: bmqzr2hu, buu0cnu1, ue8rf90u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 279.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 279.4 37.0 42.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/edt7msc2",
   "source_url": "https://omlx.ai/benchmarks/performance/edt7msc2",
   "tg_tokens": null,
   "tps": 37.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-edt7msc2-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 42.9 GB at this context; duplicate traces of the same session: bmqzr2hu, buu0cnu1, ue8rf90u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 293.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 293.0 29.6 42.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/edt7msc2",
   "source_url": "https://omlx.ai/benchmarks/performance/edt7msc2",
   "tg_tokens": null,
   "tps": 29.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-edt7msc2-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.9 GB at this context; duplicate traces of the same session: bmqzr2hu, buu0cnu1, ue8rf90u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 289.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 289.3 45.4 43.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/edt7msc2",
   "source_url": "https://omlx.ai/benchmarks/performance/edt7msc2",
   "tg_tokens": null,
   "tps": 45.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ahol6n18-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 48jiqbe7, hehy7qyf, k9kyssk7, thn6xqhe, wp0aa50b; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 636.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 636.6 72.6 70.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ahol6n18",
   "source_url": "https://omlx.ai/benchmarks/performance/ahol6n18",
   "tg_tokens": null,
   "tps": 72.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ahol6n18-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: 48jiqbe7, hehy7qyf, k9kyssk7, thn6xqhe, wp0aa50b; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1059.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 1,059 59.4 71.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ahol6n18",
   "source_url": "https://omlx.ai/benchmarks/performance/ahol6n18",
   "tg_tokens": null,
   "tps": 59.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ahol6n18-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.5 GB at this context; duplicate traces of the same session: 48jiqbe7, hehy7qyf, k9kyssk7, thn6xqhe, wp0aa50b; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1057.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 16k 1,057 111.1 72.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ahol6n18",
   "source_url": "https://omlx.ai/benchmarks/performance/ahol6n18",
   "tg_tokens": null,
   "tps": 111.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ahol6n18-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.5 GB at this context; duplicate traces of the same session: 48jiqbe7, hehy7qyf, k9kyssk7, thn6xqhe, wp0aa50b; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1048.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 1,048 59.2 73.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ahol6n18",
   "source_url": "https://omlx.ai/benchmarks/performance/ahol6n18",
   "tg_tokens": null,
   "tps": 59.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ahol6n18-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 76.2 GB at this context; duplicate traces of the same session: 48jiqbe7, hehy7qyf, k9kyssk7, thn6xqhe, wp0aa50b; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1019.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k 1,019 49.3 76.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ahol6n18",
   "source_url": "https://omlx.ai/benchmarks/performance/ahol6n18",
   "tg_tokens": null,
   "tps": 49.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ahol6n18-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 82.6 GB at this context; duplicate traces of the same session: 48jiqbe7, hehy7qyf, k9kyssk7, thn6xqhe, wp0aa50b; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 962.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 128k 962.5 47.3 82.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ahol6n18",
   "source_url": "https://omlx.ai/benchmarks/performance/ahol6n18",
   "tg_tokens": null,
   "tps": 47.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qxitizo1-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: 10u0wgus, 915rjg99, m9ppzj45, vdtgc3w2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 554.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 554.2 54.1 70.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qxitizo1",
   "source_url": "https://omlx.ai/benchmarks/performance/qxitizo1",
   "tg_tokens": null,
   "tps": 54.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qxitizo1-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: 10u0wgus, 915rjg99, m9ppzj45, vdtgc3w2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 771.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 771.8 63.1 71.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qxitizo1",
   "source_url": "https://omlx.ai/benchmarks/performance/qxitizo1",
   "tg_tokens": null,
   "tps": 63.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qxitizo1-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: 10u0wgus, 915rjg99, m9ppzj45, vdtgc3w2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1045.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 1,045 58.2 71.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qxitizo1",
   "source_url": "https://omlx.ai/benchmarks/performance/qxitizo1",
   "tg_tokens": null,
   "tps": 58.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-qxitizo1-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.1 GB at this context; duplicate traces of the same session: 10u0wgus, 915rjg99, m9ppzj45, vdtgc3w2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1043.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 16k 1,043 100.8 72.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qxitizo1",
   "source_url": "https://omlx.ai/benchmarks/performance/qxitizo1",
   "tg_tokens": null,
   "tps": 100.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qxitizo1-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.3 GB at this context; duplicate traces of the same session: 10u0wgus, 915rjg99, m9ppzj45, vdtgc3w2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1036.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 1,036 51.4 73.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qxitizo1",
   "source_url": "https://omlx.ai/benchmarks/performance/qxitizo1",
   "tg_tokens": null,
   "tps": 51.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qxitizo1-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 76.0 GB at this context; duplicate traces of the same session: 10u0wgus, 915rjg99, m9ppzj45, vdtgc3w2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 993.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k 993.9 53.1 76.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qxitizo1",
   "source_url": "https://omlx.ai/benchmarks/performance/qxitizo1",
   "tg_tokens": null,
   "tps": 53.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qxitizo1-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 82.3 GB at this context; duplicate traces of the same session: 10u0wgus, 915rjg99, m9ppzj45, vdtgc3w2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 972.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 128k 972.4 44.3 82.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qxitizo1",
   "source_url": "https://omlx.ai/benchmarks/performance/qxitizo1",
   "tg_tokens": null,
   "tps": 44.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-po7onl42-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 0i1hm2mz, 46fn7end, 8g5hy6b5, b04j70yy, f3vgjgnd, qnjfpw6j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 902.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 902.4 91.0 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/po7onl42",
   "source_url": "https://omlx.ai/benchmarks/performance/po7onl42",
   "tg_tokens": null,
   "tps": 91.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-po7onl42-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: 0i1hm2mz, 46fn7end, 8g5hy6b5, b04j70yy, f3vgjgnd, qnjfpw6j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1863.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,863 64.6 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/po7onl42",
   "source_url": "https://omlx.ai/benchmarks/performance/po7onl42",
   "tg_tokens": null,
   "tps": 64.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-po7onl42-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 0i1hm2mz, 46fn7end, 8g5hy6b5, b04j70yy, f3vgjgnd, qnjfpw6j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2357.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,357 71.4 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/po7onl42",
   "source_url": "https://omlx.ai/benchmarks/performance/po7onl42",
   "tg_tokens": null,
   "tps": 71.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-po7onl42-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: 0i1hm2mz, 46fn7end, 8g5hy6b5, b04j70yy, f3vgjgnd, qnjfpw6j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2865.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,865 93.1 72.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/po7onl42",
   "source_url": "https://omlx.ai/benchmarks/performance/po7onl42",
   "tg_tokens": null,
   "tps": 93.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-po7onl42-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.9 GB at this context; duplicate traces of the same session: 0i1hm2mz, 46fn7end, 8g5hy6b5, b04j70yy, f3vgjgnd, qnjfpw6j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2650.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,650 70.4 74.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/po7onl42",
   "source_url": "https://omlx.ai/benchmarks/performance/po7onl42",
   "tg_tokens": null,
   "tps": 70.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-po7onl42-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.3 GB at this context; duplicate traces of the same session: 0i1hm2mz, 46fn7end, 8g5hy6b5, b04j70yy, f3vgjgnd, qnjfpw6j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2485.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,485 67.0 79.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/po7onl42",
   "source_url": "https://omlx.ai/benchmarks/performance/po7onl42",
   "tg_tokens": null,
   "tps": 67.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-po7onl42-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.9 GB at this context; duplicate traces of the same session: 0i1hm2mz, 46fn7end, 8g5hy6b5, b04j70yy, f3vgjgnd, qnjfpw6j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2459.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,459 57.4 80.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/po7onl42",
   "source_url": "https://omlx.ai/benchmarks/performance/po7onl42",
   "tg_tokens": null,
   "tps": 57.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-w60ei9vf-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 71.0 GB; TTFT 6199 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 660.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 4k tokens 660.7 PP tok/s 26.8 TG tok/s 6199 TTFT (ms) 71.0 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w60ei9vf",
   "source_url": "https://omlx.ai/benchmarks/performance/w60ei9vf",
   "tg_tokens": null,
   "tps": 26.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-mxulemql-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: 62aeihmp; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1141.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k 1,141 80.2 70.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mxulemql",
   "source_url": "https://omlx.ai/benchmarks/performance/mxulemql",
   "tg_tokens": null,
   "tps": 80.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-mxulemql-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: 62aeihmp; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1732.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 8k 1,732 86.1 71.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mxulemql",
   "source_url": "https://omlx.ai/benchmarks/performance/mxulemql",
   "tg_tokens": null,
   "tps": 86.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-j2fgd1ii-ctx1024",
   "model": "gemma-3-1b-it",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 1.1 GB at this context; duplicate traces of the same session: vlr1j57w; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 389.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--gemma-3-1b-it-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k 389.1 54.2 1.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j2fgd1ii",
   "source_url": "https://omlx.ai/benchmarks/performance/j2fgd1ii",
   "tg_tokens": null,
   "tps": 54.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-j2fgd1ii-ctx4096",
   "model": "gemma-3-1b-it",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 1.2 GB at this context; duplicate traces of the same session: vlr1j57w; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1293.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--gemma-3-1b-it-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 1,293 55.6 1.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j2fgd1ii",
   "source_url": "https://omlx.ai/benchmarks/performance/j2fgd1ii",
   "tg_tokens": null,
   "tps": 55.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-eglc2223-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.5 GB at this context; duplicate traces of the same session: 5v8ofnlj, ed31ahiz, k69hball, lgtjn0d5, mfcps9vo, p0rse8e3, xjgs6rrg; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 921.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 921.4 56.9 179.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eglc2223",
   "source_url": "https://omlx.ai/benchmarks/performance/eglc2223",
   "tg_tokens": null,
   "tps": 56.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-eglc2223-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.9 GB at this context; duplicate traces of the same session: 5v8ofnlj, ed31ahiz, k69hball, lgtjn0d5, mfcps9vo, p0rse8e3, xjgs6rrg; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1891.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,891 48.5 180.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eglc2223",
   "source_url": "https://omlx.ai/benchmarks/performance/eglc2223",
   "tg_tokens": null,
   "tps": 48.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-eglc2223-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.9 GB at this context; duplicate traces of the same session: 5v8ofnlj, ed31ahiz, k69hball, lgtjn0d5, mfcps9vo, p0rse8e3, xjgs6rrg; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1987.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 1,987 89.5 181.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eglc2223",
   "source_url": "https://omlx.ai/benchmarks/performance/eglc2223",
   "tg_tokens": null,
   "tps": 89.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-eglc2223-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 182.6 GB at this context; duplicate traces of the same session: 5v8ofnlj, ed31ahiz, k69hball, lgtjn0d5, mfcps9vo, p0rse8e3, xjgs6rrg; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1966.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,966 104.6 182.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eglc2223",
   "source_url": "https://omlx.ai/benchmarks/performance/eglc2223",
   "tg_tokens": null,
   "tps": 104.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-eglc2223-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 184.0 GB at this context; duplicate traces of the same session: 5v8ofnlj, ed31ahiz, k69hball, lgtjn0d5, mfcps9vo, p0rse8e3, xjgs6rrg; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1950.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,950 93.9 184.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eglc2223",
   "source_url": "https://omlx.ai/benchmarks/performance/eglc2223",
   "tg_tokens": null,
   "tps": 93.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-eglc2223-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 186.8 GB at this context; duplicate traces of the same session: 5v8ofnlj, ed31ahiz, k69hball, lgtjn0d5, mfcps9vo, p0rse8e3, xjgs6rrg; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1921.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,921 76.6 186.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eglc2223",
   "source_url": "https://omlx.ai/benchmarks/performance/eglc2223",
   "tg_tokens": null,
   "tps": 76.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-eglc2223-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 192.6 GB at this context; duplicate traces of the same session: 5v8ofnlj, ed31ahiz, k69hball, lgtjn0d5, mfcps9vo, p0rse8e3, xjgs6rrg; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1883.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 1,883 112.6 192.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eglc2223",
   "source_url": "https://omlx.ai/benchmarks/performance/eglc2223",
   "tg_tokens": null,
   "tps": 112.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-eglc2223-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 198.4 GB at this context; duplicate traces of the same session: 5v8ofnlj, ed31ahiz, k69hball, lgtjn0d5, mfcps9vo, p0rse8e3, xjgs6rrg; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1844.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 1,844 56.2 198.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eglc2223",
   "source_url": "https://omlx.ai/benchmarks/performance/eglc2223",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-a8c64ywj-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: 0nguh736, 7fjz4dkw, 7ib5xj0z, a5ooi2qr, dobcyn7z, gsgu8y70, i7m36vhz; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 374.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 374.8 65.6 177.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a8c64ywj",
   "source_url": "https://omlx.ai/benchmarks/performance/a8c64ywj",
   "tg_tokens": null,
   "tps": 65.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-a8c64ywj-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: 0nguh736, 7fjz4dkw, 7ib5xj0z, a5ooi2qr, dobcyn7z, gsgu8y70, i7m36vhz; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 897.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 897.3 63.6 178.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a8c64ywj",
   "source_url": "https://omlx.ai/benchmarks/performance/a8c64ywj",
   "tg_tokens": null,
   "tps": 63.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-a8c64ywj-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.9 GB at this context; duplicate traces of the same session: 0nguh736, 7fjz4dkw, 7ib5xj0z, a5ooi2qr, dobcyn7z, gsgu8y70, i7m36vhz; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1866.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 1,866 54.4 178.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a8c64ywj",
   "source_url": "https://omlx.ai/benchmarks/performance/a8c64ywj",
   "tg_tokens": null,
   "tps": 54.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-a8c64ywj-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.1 GB at this context; duplicate traces of the same session: 0nguh736, 7fjz4dkw, 7ib5xj0z, a5ooi2qr, dobcyn7z, gsgu8y70, i7m36vhz; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1919.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,919 141.0 179.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a8c64ywj",
   "source_url": "https://omlx.ai/benchmarks/performance/a8c64ywj",
   "tg_tokens": null,
   "tps": 141.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-a8c64ywj-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.4 GB at this context; duplicate traces of the same session: 0nguh736, 7fjz4dkw, 7ib5xj0z, a5ooi2qr, dobcyn7z, gsgu8y70, i7m36vhz; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1911.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,911 100.4 179.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a8c64ywj",
   "source_url": "https://omlx.ai/benchmarks/performance/a8c64ywj",
   "tg_tokens": null,
   "tps": 100.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-a8c64ywj-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.1 GB at this context; duplicate traces of the same session: 0nguh736, 7fjz4dkw, 7ib5xj0z, a5ooi2qr, dobcyn7z, gsgu8y70, i7m36vhz; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1882.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,882 80.6 180.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a8c64ywj",
   "source_url": "https://omlx.ai/benchmarks/performance/a8c64ywj",
   "tg_tokens": null,
   "tps": 80.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-a8c64ywj-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.3 GB at this context; duplicate traces of the same session: 0nguh736, 7fjz4dkw, 7ib5xj0z, a5ooi2qr, dobcyn7z, gsgu8y70, i7m36vhz; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1848.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 1,848 138.5 181.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a8c64ywj",
   "source_url": "https://omlx.ai/benchmarks/performance/a8c64ywj",
   "tg_tokens": null,
   "tps": 138.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-a8c64ywj-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 182.6 GB at this context; duplicate traces of the same session: 0nguh736, 7fjz4dkw, 7ib5xj0z, a5ooi2qr, dobcyn7z, gsgu8y70, i7m36vhz; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1794.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 1,794 66.7 182.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a8c64ywj",
   "source_url": "https://omlx.ai/benchmarks/performance/a8c64ywj",
   "tg_tokens": null,
   "tps": 66.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-mzlgwmcd-ctx1024",
   "model": "Qwen2.5-Coder-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 4.8 GB at this context; duplicate traces of the same session: u0mmooqd; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 444.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-7B-Instruct-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 1k 444.4 67.3 4.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mzlgwmcd",
   "source_url": "https://omlx.ai/benchmarks/performance/mzlgwmcd",
   "tg_tokens": null,
   "tps": 67.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-mzlgwmcd-ctx4096",
   "model": "Qwen2.5-Coder-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: u0mmooqd; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 434.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-7B-Instruct-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k 434.6 61.7 4.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mzlgwmcd",
   "source_url": "https://omlx.ai/benchmarks/performance/mzlgwmcd",
   "tg_tokens": null,
   "tps": 61.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-hlg8mgh1-ctx4096",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 070kuxm7, m6aiqoyx; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 511.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k 511.8 47.2 17.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hlg8mgh1",
   "source_url": "https://omlx.ai/benchmarks/performance/hlg8mgh1",
   "tg_tokens": null,
   "tps": 47.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-hlg8mgh1-ctx16384",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: 070kuxm7, m6aiqoyx; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 326.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 16k 326.6 31.8 18.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hlg8mgh1",
   "source_url": "https://omlx.ai/benchmarks/performance/hlg8mgh1",
   "tg_tokens": null,
   "tps": 31.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-hlg8mgh1-ctx65536",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: 070kuxm7, m6aiqoyx; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 132.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 64k 132.6 12.7 22.5 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hlg8mgh1",
   "source_url": "https://omlx.ai/benchmarks/performance/hlg8mgh1",
   "tg_tokens": null,
   "tps": 12.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Pro",
   "id": "omlx-pda7vzj6-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 7ernl7l0, p72nho6y; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 1k 102.9 9.0 16.0 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pda7vzj6",
   "source_url": "https://omlx.ai/benchmarks/performance/pda7vzj6",
   "tg_tokens": null,
   "tps": 9.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Pro",
   "id": "omlx-pda7vzj6-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 7ernl7l0, p72nho6y; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 4k 106.2 8.7 17.0 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pda7vzj6",
   "source_url": "https://omlx.ai/benchmarks/performance/pda7vzj6",
   "tg_tokens": null,
   "tps": 8.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-pda7vzj6-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: 7ernl7l0, p72nho6y; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 36.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 64k 36.2 1.7 22.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pda7vzj6",
   "source_url": "https://omlx.ai/benchmarks/performance/pda7vzj6",
   "tg_tokens": null,
   "tps": 1.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-bgxjjq1a-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: vwntpckj; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 205.8 22.5 16.0 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bgxjjq1a",
   "source_url": "https://omlx.ai/benchmarks/performance/bgxjjq1a",
   "tg_tokens": null,
   "tps": 22.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-bgxjjq1a-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: vwntpckj; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 213.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 213.9 21.9 17.0 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bgxjjq1a",
   "source_url": "https://omlx.ai/benchmarks/performance/bgxjjq1a",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-201u29hc-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 31.8 GB; TTFT 32202 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k tokens 127.2 PP tok/s 8.4 TG tok/s 32202 TTFT (ms) 31.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/201u29hc",
   "source_url": "https://omlx.ai/benchmarks/performance/201u29hc",
   "tg_tokens": null,
   "tps": 8.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ywj2jom1-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: yvznh0qr; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1073.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 1k 1,073 33.2 28.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ywj2jom1",
   "source_url": "https://omlx.ai/benchmarks/performance/ywj2jom1",
   "tg_tokens": null,
   "tps": 33.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ywj2jom1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: yvznh0qr; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1371.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 4k 1,371 32.9 29.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ywj2jom1",
   "source_url": "https://omlx.ai/benchmarks/performance/ywj2jom1",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-otz6w61s-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 707.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 707.3 46.9 21.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/otz6w61s",
   "source_url": "https://omlx.ai/benchmarks/performance/otz6w61s",
   "tg_tokens": null,
   "tps": 46.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-otz6w61s-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 670.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 670.9 62.1 23.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/otz6w61s",
   "source_url": "https://omlx.ai/benchmarks/performance/otz6w61s",
   "tg_tokens": null,
   "tps": 62.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-fpga391g-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: pnziv20s; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 565.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 565.5 19.0 20.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fpga391g",
   "source_url": "https://omlx.ai/benchmarks/performance/fpga391g",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-fpga391g-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: pnziv20s; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 506.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 506.3 26.2 21.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fpga391g",
   "source_url": "https://omlx.ai/benchmarks/performance/fpga391g",
   "tg_tokens": null,
   "tps": 26.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q4ucuil5-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: 1yjg5ecj, 6vt6kunu, mnaxhds1, yxzvq02u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1077.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,077 65.6 71.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q4ucuil5",
   "source_url": "https://omlx.ai/benchmarks/performance/q4ucuil5",
   "tg_tokens": null,
   "tps": 65.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q4ucuil5-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 71.6 GB at this context; duplicate traces of the same session: 1yjg5ecj, 6vt6kunu, mnaxhds1, yxzvq02u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1888.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,888 63.3 71.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q4ucuil5",
   "source_url": "https://omlx.ai/benchmarks/performance/q4ucuil5",
   "tg_tokens": null,
   "tps": 63.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q4ucuil5-ctx32768",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 73.5 GB at this context; duplicate traces of the same session: 1yjg5ecj, 6vt6kunu, mnaxhds1, yxzvq02u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2593.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,593 63.2 73.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q4ucuil5",
   "source_url": "https://omlx.ai/benchmarks/performance/q4ucuil5",
   "tg_tokens": null,
   "tps": 63.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q4ucuil5-ctx65536",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 74.6 GB at this context; duplicate traces of the same session: 1yjg5ecj, 6vt6kunu, mnaxhds1, yxzvq02u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2577.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,577 57.7 74.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q4ucuil5",
   "source_url": "https://omlx.ai/benchmarks/performance/q4ucuil5",
   "tg_tokens": null,
   "tps": 57.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q4ucuil5-ctx131072",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 77.0 GB at this context; duplicate traces of the same session: 1yjg5ecj, 6vt6kunu, mnaxhds1, yxzvq02u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2483.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,483 55.9 77.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q4ucuil5",
   "source_url": "https://omlx.ai/benchmarks/performance/q4ucuil5",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-t2etu35w-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace, single run at 1k context; peak memory 71.3 GB; TTFT 2901 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 352.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 352.9 PP tok/s 57.2 TG tok/s 2901 TTFT (ms) 71.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t2etu35w",
   "source_url": "https://omlx.ai/benchmarks/performance/t2etu35w",
   "tg_tokens": null,
   "tps": 57.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-y9p24688-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: 17nwoo8y, 1806g8tq, 3oof4vvz, e46q6ru4, o4txa7v0, tb1xwdai; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 465.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 465.0 60.4 71.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y9p24688",
   "source_url": "https://omlx.ai/benchmarks/performance/y9p24688",
   "tg_tokens": null,
   "tps": 60.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-y9p24688-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.1 GB at this context; duplicate traces of the same session: 17nwoo8y, 1806g8tq, 3oof4vvz, e46q6ru4, o4txa7v0, tb1xwdai; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 468.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 468.5 45.3 74.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y9p24688",
   "source_url": "https://omlx.ai/benchmarks/performance/y9p24688",
   "tg_tokens": null,
   "tps": 45.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-y9p24688-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.4 GB at this context; duplicate traces of the same session: 17nwoo8y, 1806g8tq, 3oof4vvz, e46q6ru4, o4txa7v0, tb1xwdai; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 404.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 404.3 52.6 74.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y9p24688",
   "source_url": "https://omlx.ai/benchmarks/performance/y9p24688",
   "tg_tokens": null,
   "tps": 52.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-y9p24688-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 77.4 GB at this context; duplicate traces of the same session: 17nwoo8y, 1806g8tq, 3oof4vvz, e46q6ru4, o4txa7v0, tb1xwdai; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 395.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 395.5 56.4 77.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y9p24688",
   "source_url": "https://omlx.ai/benchmarks/performance/y9p24688",
   "tg_tokens": null,
   "tps": 56.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-y9p24688-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.5 GB at this context; duplicate traces of the same session: 17nwoo8y, 1806g8tq, 3oof4vvz, e46q6ru4, o4txa7v0, tb1xwdai; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 375.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 375.5 41.1 78.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y9p24688",
   "source_url": "https://omlx.ai/benchmarks/performance/y9p24688",
   "tg_tokens": null,
   "tps": 41.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-y9p24688-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.4 GB at this context; duplicate traces of the same session: 17nwoo8y, 1806g8tq, 3oof4vvz, e46q6ru4, o4txa7v0, tb1xwdai; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 378.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 378.1 42.0 80.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y9p24688",
   "source_url": "https://omlx.ai/benchmarks/performance/y9p24688",
   "tg_tokens": null,
   "tps": 42.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-y9p24688-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 84.9 GB at this context; duplicate traces of the same session: 17nwoo8y, 1806g8tq, 3oof4vvz, e46q6ru4, o4txa7v0, tb1xwdai; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 352.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 352.5 37.6 84.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y9p24688",
   "source_url": "https://omlx.ai/benchmarks/performance/y9p24688",
   "tg_tokens": null,
   "tps": 37.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-k0mv92z7-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: r1tnhq0i; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 554.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 554.2 54.1 70.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k0mv92z7",
   "source_url": "https://omlx.ai/benchmarks/performance/k0mv92z7",
   "tg_tokens": null,
   "tps": 54.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-k0mv92z7-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: r1tnhq0i; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 771.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 771.8 63.1 71.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k0mv92z7",
   "source_url": "https://omlx.ai/benchmarks/performance/k0mv92z7",
   "tg_tokens": null,
   "tps": 63.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-k0mv92z7-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: r1tnhq0i; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1045.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 1,045 58.2 71.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k0mv92z7",
   "source_url": "https://omlx.ai/benchmarks/performance/k0mv92z7",
   "tg_tokens": null,
   "tps": 58.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-k0mv92z7-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.1 GB at this context; duplicate traces of the same session: r1tnhq0i; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1043.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 16k 1,043 100.8 72.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k0mv92z7",
   "source_url": "https://omlx.ai/benchmarks/performance/k0mv92z7",
   "tg_tokens": null,
   "tps": 100.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-k0mv92z7-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.3 GB at this context; duplicate traces of the same session: r1tnhq0i; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1036.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 1,036 51.4 73.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k0mv92z7",
   "source_url": "https://omlx.ai/benchmarks/performance/k0mv92z7",
   "tg_tokens": null,
   "tps": 51.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-k0mv92z7-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 76.0 GB at this context; duplicate traces of the same session: r1tnhq0i; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 993.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k 993.9 53.1 76.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k0mv92z7",
   "source_url": "https://omlx.ai/benchmarks/performance/k0mv92z7",
   "tg_tokens": null,
   "tps": 53.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-k0mv92z7-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 82.3 GB at this context; duplicate traces of the same session: r1tnhq0i; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 972.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 128k 972.4 44.3 82.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k0mv92z7",
   "source_url": "https://omlx.ai/benchmarks/performance/k0mv92z7",
   "tg_tokens": null,
   "tps": 44.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-camacnfk-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: bhx1oatw, d302ktmu, fahd6ajk, hc38g4k1, jz75vuoc, vrytasz1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 335.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 335.2 28.4 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/camacnfk",
   "source_url": "https://omlx.ai/benchmarks/performance/camacnfk",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-camacnfk-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: bhx1oatw, d302ktmu, fahd6ajk, hc38g4k1, jz75vuoc, vrytasz1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1448.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,448 31.2 71.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/camacnfk",
   "source_url": "https://omlx.ai/benchmarks/performance/camacnfk",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-camacnfk-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.7 GB at this context; duplicate traces of the same session: bhx1oatw, d302ktmu, fahd6ajk, hc38g4k1, jz75vuoc, vrytasz1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2237.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,237 61.2 72.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/camacnfk",
   "source_url": "https://omlx.ai/benchmarks/performance/camacnfk",
   "tg_tokens": null,
   "tps": 61.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-camacnfk-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: bhx1oatw, d302ktmu, fahd6ajk, hc38g4k1, jz75vuoc, vrytasz1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2267.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,267 57.1 73.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/camacnfk",
   "source_url": "https://omlx.ai/benchmarks/performance/camacnfk",
   "tg_tokens": null,
   "tps": 57.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-camacnfk-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.9 GB at this context; duplicate traces of the same session: bhx1oatw, d302ktmu, fahd6ajk, hc38g4k1, jz75vuoc, vrytasz1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2218.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,218 50.2 74.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/camacnfk",
   "source_url": "https://omlx.ai/benchmarks/performance/camacnfk",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-camacnfk-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.8 GB at this context; duplicate traces of the same session: bhx1oatw, d302ktmu, fahd6ajk, hc38g4k1, jz75vuoc, vrytasz1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2104.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,104 36.4 79.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/camacnfk",
   "source_url": "https://omlx.ai/benchmarks/performance/camacnfk",
   "tg_tokens": null,
   "tps": 36.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-camacnfk-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.9 GB at this context; duplicate traces of the same session: bhx1oatw, d302ktmu, fahd6ajk, hc38g4k1, jz75vuoc, vrytasz1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2043.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,043 66.2 80.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/camacnfk",
   "source_url": "https://omlx.ai/benchmarks/performance/camacnfk",
   "tg_tokens": null,
   "tps": 66.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-19yvfiql-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 71.1 GB; TTFT 2718 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1507.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 1,507 PP tok/s 72.6 TG tok/s 2718 TTFT (ms) 71.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/19yvfiql",
   "source_url": "https://omlx.ai/benchmarks/performance/19yvfiql",
   "tg_tokens": null,
   "tps": 72.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-835stej6-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 74.5 GB; TTFT 23377 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2804.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k tokens 2,804 PP tok/s 70.1 TG tok/s 23377 TTFT (ms) 74.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/835stej6",
   "source_url": "https://omlx.ai/benchmarks/performance/835stej6",
   "tg_tokens": null,
   "tps": 70.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cl31yd25-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 71.0 GB; TTFT 2247 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1823.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 1,823 PP tok/s 69.6 TG tok/s 2247 TTFT (ms) 71.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cl31yd25",
   "source_url": "https://omlx.ai/benchmarks/performance/cl31yd25",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-200bs7al-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.2 GB at this context; duplicate traces of the same session: mhwono6h, norqc5j4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1157.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,157 78.0 69.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/200bs7al",
   "source_url": "https://omlx.ai/benchmarks/performance/200bs7al",
   "tg_tokens": null,
   "tps": 78.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-200bs7al-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.4 GB at this context; duplicate traces of the same session: mhwono6h, norqc5j4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2306.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,306 77.2 69.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/200bs7al",
   "source_url": "https://omlx.ai/benchmarks/performance/200bs7al",
   "tg_tokens": null,
   "tps": 77.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-200bs7al-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.3 GB at this context; duplicate traces of the same session: mhwono6h, norqc5j4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2604.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,604 73.1 70.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/200bs7al",
   "source_url": "https://omlx.ai/benchmarks/performance/200bs7al",
   "tg_tokens": null,
   "tps": 73.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-200bs7al-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: mhwono6h, norqc5j4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2661.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,661 71.3 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/200bs7al",
   "source_url": "https://omlx.ai/benchmarks/performance/200bs7al",
   "tg_tokens": null,
   "tps": 71.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-200bs7al-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.4 GB at this context; duplicate traces of the same session: mhwono6h, norqc5j4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2515.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,515 70.2 71.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/200bs7al",
   "source_url": "https://omlx.ai/benchmarks/performance/200bs7al",
   "tg_tokens": null,
   "tps": 70.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-200bs7al-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.5 GB at this context; duplicate traces of the same session: mhwono6h, norqc5j4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2425.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,425 67.6 72.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/200bs7al",
   "source_url": "https://omlx.ai/benchmarks/performance/200bs7al",
   "tg_tokens": null,
   "tps": 67.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-0nhwjvvn-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: hq5ncfsm; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 589.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 1k 589.4 41.1 5.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0nhwjvvn",
   "source_url": "https://omlx.ai/benchmarks/performance/0nhwjvvn",
   "tg_tokens": null,
   "tps": 41.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-0nhwjvvn-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.2 GB at this context; duplicate traces of the same session: hq5ncfsm; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 758.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 4k 758.3 87.4 7.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0nhwjvvn",
   "source_url": "https://omlx.ai/benchmarks/performance/0nhwjvvn",
   "tg_tokens": null,
   "tps": 87.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ue7ny62b-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1814.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-MLX-6bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 6bit; 1k 1,814 82.4 27.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ue7ny62b",
   "source_url": "https://omlx.ai/benchmarks/performance/ue7ny62b",
   "tg_tokens": null,
   "tps": 82.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ue7ny62b-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.2 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2254.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-MLX-6bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 6bit; 8k 2,254 77.0 28.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ue7ny62b",
   "source_url": "https://omlx.ai/benchmarks/performance/ue7ny62b",
   "tg_tokens": null,
   "tps": 77.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ue7ny62b-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2099.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-MLX-6bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 6bit; 16k 2,099 72.6 28.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ue7ny62b",
   "source_url": "https://omlx.ai/benchmarks/performance/ue7ny62b",
   "tg_tokens": null,
   "tps": 72.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ue7ny62b-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.7 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1824.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-MLX-6bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 6bit; 32k 1,824 64.0 28.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ue7ny62b",
   "source_url": "https://omlx.ai/benchmarks/performance/ue7ny62b",
   "tg_tokens": null,
   "tps": 64.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ue7ny62b-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 29.3 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1431.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-MLX-6bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 6bit; 64k 1,431 51.9 29.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ue7ny62b",
   "source_url": "https://omlx.ai/benchmarks/performance/ue7ny62b",
   "tg_tokens": null,
   "tps": 51.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ue7ny62b-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 30.6 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 996.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-MLX-6bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 6bit; 128k 996.4 38.0 30.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ue7ny62b",
   "source_url": "https://omlx.ai/benchmarks/performance/ue7ny62b",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ue7ny62b-ctx199680",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 31.9 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 755.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-MLX-6bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 6bit; 195k 755.6 29.5 31.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ue7ny62b",
   "source_url": "https://omlx.ai/benchmarks/performance/ue7ny62b",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8msahh3p-ctx1024",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 0b9drwmh, 3slb2en3, 5i5g9in4, giw4rpro, ryhoxob3, vhe1y1iy, z0ctl9d1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3064.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 3,064 197.6 21.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8msahh3p",
   "source_url": "https://omlx.ai/benchmarks/performance/8msahh3p",
   "tg_tokens": null,
   "tps": 197.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8msahh3p-ctx4096",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 0b9drwmh, 3slb2en3, 5i5g9in4, giw4rpro, ryhoxob3, vhe1y1iy, z0ctl9d1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5511.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 5,511 177.4 22.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8msahh3p",
   "source_url": "https://omlx.ai/benchmarks/performance/8msahh3p",
   "tg_tokens": null,
   "tps": 177.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8msahh3p-ctx8192",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 0b9drwmh, 3slb2en3, 5i5g9in4, giw4rpro, ryhoxob3, vhe1y1iy, z0ctl9d1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5975.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 5,975 150.6 22.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8msahh3p",
   "source_url": "https://omlx.ai/benchmarks/performance/8msahh3p",
   "tg_tokens": null,
   "tps": 150.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8msahh3p-ctx16384",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 0b9drwmh, 3slb2en3, 5i5g9in4, giw4rpro, ryhoxob3, vhe1y1iy, z0ctl9d1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5800.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 5,800 183.8 22.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8msahh3p",
   "source_url": "https://omlx.ai/benchmarks/performance/8msahh3p",
   "tg_tokens": null,
   "tps": 183.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8msahh3p-ctx32768",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 0b9drwmh, 3slb2en3, 5i5g9in4, giw4rpro, ryhoxob3, vhe1y1iy, z0ctl9d1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4764.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 4,764 167.0 22.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8msahh3p",
   "source_url": "https://omlx.ai/benchmarks/performance/8msahh3p",
   "tg_tokens": null,
   "tps": 167.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8msahh3p-ctx65536",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: 0b9drwmh, 3slb2en3, 5i5g9in4, giw4rpro, ryhoxob3, vhe1y1iy, z0ctl9d1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2977.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,977 115.3 23.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8msahh3p",
   "source_url": "https://omlx.ai/benchmarks/performance/8msahh3p",
   "tg_tokens": null,
   "tps": 115.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8msahh3p-ctx131072",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: 0b9drwmh, 3slb2en3, 5i5g9in4, giw4rpro, ryhoxob3, vhe1y1iy, z0ctl9d1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2212.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,212 99.2 24.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8msahh3p",
   "source_url": "https://omlx.ai/benchmarks/performance/8msahh3p",
   "tg_tokens": null,
   "tps": 99.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8msahh3p-ctx199680",
   "model": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.3 GB at this context; duplicate traces of the same session: 0b9drwmh, 3slb2en3, 5i5g9in4, giw4rpro, ryhoxob3, vhe1y1iy, z0ctl9d1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1650.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 1,650 64.5 27.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8msahh3p",
   "source_url": "https://omlx.ai/benchmarks/performance/8msahh3p",
   "tg_tokens": null,
   "tps": 64.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-guicp9pn-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: nxtbwsgm; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 245.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 1k 245.3 29.6 16.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/guicp9pn",
   "source_url": "https://omlx.ai/benchmarks/performance/guicp9pn",
   "tg_tokens": null,
   "tps": 29.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-guicp9pn-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: nxtbwsgm; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 230.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 4k 230.0 30.1 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/guicp9pn",
   "source_url": "https://omlx.ai/benchmarks/performance/guicp9pn",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-63o0tcys-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: dz01oas9; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4-bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 130.0 13.9 17.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/63o0tcys",
   "source_url": "https://omlx.ai/benchmarks/performance/63o0tcys",
   "tg_tokens": null,
   "tps": 13.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-63o0tcys-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: dz01oas9; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4-bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 127.1 13.6 18.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/63o0tcys",
   "source_url": "https://omlx.ai/benchmarks/performance/63o0tcys",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wwf7zrht-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 707.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 707.3 46.9 21.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wwf7zrht",
   "source_url": "https://omlx.ai/benchmarks/performance/wwf7zrht",
   "tg_tokens": null,
   "tps": 46.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wwf7zrht-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 670.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 670.9 62.1 23.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wwf7zrht",
   "source_url": "https://omlx.ai/benchmarks/performance/wwf7zrht",
   "tg_tokens": null,
   "tps": 62.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-7d1uhye9-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: lon60v95; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 507.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 507.5 33.7 17.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7d1uhye9",
   "source_url": "https://omlx.ai/benchmarks/performance/7d1uhye9",
   "tg_tokens": null,
   "tps": 33.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-7d1uhye9-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: lon60v95; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 691.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 691.2 27.6 19.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7d1uhye9",
   "source_url": "https://omlx.ai/benchmarks/performance/7d1uhye9",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rcrz7hy6-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: ucn1lzkv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 753.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 753.1 45.1 17.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rcrz7hy6",
   "source_url": "https://omlx.ai/benchmarks/performance/rcrz7hy6",
   "tg_tokens": null,
   "tps": 45.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rcrz7hy6-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: ucn1lzkv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 907.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 907.5 59.0 18.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rcrz7hy6",
   "source_url": "https://omlx.ai/benchmarks/performance/rcrz7hy6",
   "tg_tokens": null,
   "tps": 59.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0ia6t4lx-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.5 GB at this context; duplicate traces of the same session: 0w5aqg6p, 1fbqnbzm, 7xmqm5dn, qsakvmg6, xfd7405y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1121.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,121 32.2 17.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ia6t4lx",
   "source_url": "https://omlx.ai/benchmarks/performance/0ia6t4lx",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0ia6t4lx-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 0w5aqg6p, 1fbqnbzm, 7xmqm5dn, qsakvmg6, xfd7405y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1319.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,319 31.6 19.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ia6t4lx",
   "source_url": "https://omlx.ai/benchmarks/performance/0ia6t4lx",
   "tg_tokens": null,
   "tps": 31.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0ia6t4lx-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 0w5aqg6p, 1fbqnbzm, 7xmqm5dn, qsakvmg6, xfd7405y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1169.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,169 31.0 19.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ia6t4lx",
   "source_url": "https://omlx.ai/benchmarks/performance/0ia6t4lx",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0ia6t4lx-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 0w5aqg6p, 1fbqnbzm, 7xmqm5dn, qsakvmg6, xfd7405y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1120.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,120 28.9 19.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ia6t4lx",
   "source_url": "https://omlx.ai/benchmarks/performance/0ia6t4lx",
   "tg_tokens": null,
   "tps": 28.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0ia6t4lx-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 0w5aqg6p, 1fbqnbzm, 7xmqm5dn, qsakvmg6, xfd7405y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1010.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,010 28.3 20.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ia6t4lx",
   "source_url": "https://omlx.ai/benchmarks/performance/0ia6t4lx",
   "tg_tokens": null,
   "tps": 28.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0ia6t4lx-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: 0w5aqg6p, 1fbqnbzm, 7xmqm5dn, qsakvmg6, xfd7405y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 828.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 828.2 22.0 22.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ia6t4lx",
   "source_url": "https://omlx.ai/benchmarks/performance/0ia6t4lx",
   "tg_tokens": null,
   "tps": 22.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-j0zny9u1-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 7ljas9c5, o3ql3um2, o8vmesq4, ph219mic, qhdizv0j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1135.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,135 70.7 17.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j0zny9u1",
   "source_url": "https://omlx.ai/benchmarks/performance/j0zny9u1",
   "tg_tokens": null,
   "tps": 70.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-j0zny9u1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 7ljas9c5, o3ql3um2, o8vmesq4, ph219mic, qhdizv0j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1304.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,304 64.7 19.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j0zny9u1",
   "source_url": "https://omlx.ai/benchmarks/performance/j0zny9u1",
   "tg_tokens": null,
   "tps": 64.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-j0zny9u1-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 7ljas9c5, o3ql3um2, o8vmesq4, ph219mic, qhdizv0j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1182.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,182 56.6 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j0zny9u1",
   "source_url": "https://omlx.ai/benchmarks/performance/j0zny9u1",
   "tg_tokens": null,
   "tps": 56.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-j0zny9u1-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 7ljas9c5, o3ql3um2, o8vmesq4, ph219mic, qhdizv0j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1089.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,089 86.8 19.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j0zny9u1",
   "source_url": "https://omlx.ai/benchmarks/performance/j0zny9u1",
   "tg_tokens": null,
   "tps": 86.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-j0zny9u1-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 7ljas9c5, o3ql3um2, o8vmesq4, ph219mic, qhdizv0j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 949.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 949.1 126.1 21.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j0zny9u1",
   "source_url": "https://omlx.ai/benchmarks/performance/j0zny9u1",
   "tg_tokens": null,
   "tps": 126.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-j0zny9u1-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 7ljas9c5, o3ql3um2, o8vmesq4, ph219mic, qhdizv0j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 766.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 766.2 45.0 23.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j0zny9u1",
   "source_url": "https://omlx.ai/benchmarks/performance/j0zny9u1",
   "tg_tokens": null,
   "tps": 45.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-i2ldqpx2-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: 2hkbossw, 9yneuslw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 509.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 509.1 24.4 71.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i2ldqpx2",
   "source_url": "https://omlx.ai/benchmarks/performance/i2ldqpx2",
   "tg_tokens": null,
   "tps": 24.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-i2ldqpx2-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: 2hkbossw, 9yneuslw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 586.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 586.8 42.7 71.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i2ldqpx2",
   "source_url": "https://omlx.ai/benchmarks/performance/i2ldqpx2",
   "tg_tokens": null,
   "tps": 42.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-i2ldqpx2-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: 2hkbossw, 9yneuslw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 819.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 819.9 55.3 71.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i2ldqpx2",
   "source_url": "https://omlx.ai/benchmarks/performance/i2ldqpx2",
   "tg_tokens": null,
   "tps": 55.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-i2ldqpx2-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 2hkbossw, 9yneuslw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 820.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 820.2 30.8 71.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i2ldqpx2",
   "source_url": "https://omlx.ai/benchmarks/performance/i2ldqpx2",
   "tg_tokens": null,
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-g2b7hq59-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.2 GB at this context; duplicate traces of the same session: qq51guei, vggc7qnp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1157.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,157 78.0 69.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g2b7hq59",
   "source_url": "https://omlx.ai/benchmarks/performance/g2b7hq59",
   "tg_tokens": null,
   "tps": 78.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-g2b7hq59-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.4 GB at this context; duplicate traces of the same session: qq51guei, vggc7qnp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2306.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,306 77.2 69.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g2b7hq59",
   "source_url": "https://omlx.ai/benchmarks/performance/g2b7hq59",
   "tg_tokens": null,
   "tps": 77.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-g2b7hq59-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.3 GB at this context; duplicate traces of the same session: qq51guei, vggc7qnp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2604.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,604 73.1 70.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g2b7hq59",
   "source_url": "https://omlx.ai/benchmarks/performance/g2b7hq59",
   "tg_tokens": null,
   "tps": 73.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-g2b7hq59-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: qq51guei, vggc7qnp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2661.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,661 71.3 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g2b7hq59",
   "source_url": "https://omlx.ai/benchmarks/performance/g2b7hq59",
   "tg_tokens": null,
   "tps": 71.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-g2b7hq59-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.4 GB at this context; duplicate traces of the same session: qq51guei, vggc7qnp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2515.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,515 70.2 71.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g2b7hq59",
   "source_url": "https://omlx.ai/benchmarks/performance/g2b7hq59",
   "tg_tokens": null,
   "tps": 70.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-g2b7hq59-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.5 GB at this context; duplicate traces of the same session: qq51guei, vggc7qnp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2425.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,425 67.6 72.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g2b7hq59",
   "source_url": "https://omlx.ai/benchmarks/performance/g2b7hq59",
   "tg_tokens": null,
   "tps": 67.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mfeap0fl-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.2 GB at this context; duplicate traces of the same session: dfevsik7, dnklbw4q, jk7hbmks, s2hn25ai, tivv4n1x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1158.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,158 72.8 69.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mfeap0fl",
   "source_url": "https://omlx.ai/benchmarks/performance/mfeap0fl",
   "tg_tokens": null,
   "tps": 72.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mfeap0fl-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.5 GB at this context; duplicate traces of the same session: dfevsik7, dnklbw4q, jk7hbmks, s2hn25ai, tivv4n1x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1668.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,668 68.9 69.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mfeap0fl",
   "source_url": "https://omlx.ai/benchmarks/performance/mfeap0fl",
   "tg_tokens": null,
   "tps": 68.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mfeap0fl-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.1 GB at this context; duplicate traces of the same session: dfevsik7, dnklbw4q, jk7hbmks, s2hn25ai, tivv4n1x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2028.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,028 68.3 70.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mfeap0fl",
   "source_url": "https://omlx.ai/benchmarks/performance/mfeap0fl",
   "tg_tokens": null,
   "tps": 68.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mfeap0fl-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: dfevsik7, dnklbw4q, jk7hbmks, s2hn25ai, tivv4n1x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2197.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,197 69.4 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mfeap0fl",
   "source_url": "https://omlx.ai/benchmarks/performance/mfeap0fl",
   "tg_tokens": null,
   "tps": 69.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mfeap0fl-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.4 GB at this context; duplicate traces of the same session: dfevsik7, dnklbw4q, jk7hbmks, s2hn25ai, tivv4n1x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2269.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,269 68.5 71.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mfeap0fl",
   "source_url": "https://omlx.ai/benchmarks/performance/mfeap0fl",
   "tg_tokens": null,
   "tps": 68.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mfeap0fl-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.5 GB at this context; duplicate traces of the same session: dfevsik7, dnklbw4q, jk7hbmks, s2hn25ai, tivv4n1x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2126.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,126 62.9 72.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mfeap0fl",
   "source_url": "https://omlx.ai/benchmarks/performance/mfeap0fl",
   "tg_tokens": null,
   "tps": 62.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vh28d3fa-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.2 GB at this context; duplicate traces of the same session: 7rjhx5oc, fh0z3dbh, kt3046z0, xrldn3go, zhhqpxrv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1014.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,014 70.2 69.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vh28d3fa",
   "source_url": "https://omlx.ai/benchmarks/performance/vh28d3fa",
   "tg_tokens": null,
   "tps": 70.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vh28d3fa-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.4 GB at this context; duplicate traces of the same session: 7rjhx5oc, fh0z3dbh, kt3046z0, xrldn3go, zhhqpxrv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1958.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,958 69.0 69.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vh28d3fa",
   "source_url": "https://omlx.ai/benchmarks/performance/vh28d3fa",
   "tg_tokens": null,
   "tps": 69.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vh28d3fa-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.1 GB at this context; duplicate traces of the same session: 7rjhx5oc, fh0z3dbh, kt3046z0, xrldn3go, zhhqpxrv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2308.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,308 69.0 70.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vh28d3fa",
   "source_url": "https://omlx.ai/benchmarks/performance/vh28d3fa",
   "tg_tokens": null,
   "tps": 69.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vh28d3fa-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: 7rjhx5oc, fh0z3dbh, kt3046z0, xrldn3go, zhhqpxrv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2555.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,555 67.9 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vh28d3fa",
   "source_url": "https://omlx.ai/benchmarks/performance/vh28d3fa",
   "tg_tokens": null,
   "tps": 67.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vh28d3fa-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.4 GB at this context; duplicate traces of the same session: 7rjhx5oc, fh0z3dbh, kt3046z0, xrldn3go, zhhqpxrv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2476.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,476 67.1 71.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vh28d3fa",
   "source_url": "https://omlx.ai/benchmarks/performance/vh28d3fa",
   "tg_tokens": null,
   "tps": 67.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vh28d3fa-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.5 GB at this context; duplicate traces of the same session: 7rjhx5oc, fh0z3dbh, kt3046z0, xrldn3go, zhhqpxrv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2379.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,379 62.8 72.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vh28d3fa",
   "source_url": "https://omlx.ai/benchmarks/performance/vh28d3fa",
   "tg_tokens": null,
   "tps": 62.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mrl5f378-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.2 GB at this context; duplicate traces of the same session: 9h22t197, cks8h82r, oyqerbew, s7t3o3gl, tgh3o3v3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1148.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,148 77.7 69.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mrl5f378",
   "source_url": "https://omlx.ai/benchmarks/performance/mrl5f378",
   "tg_tokens": null,
   "tps": 77.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mrl5f378-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.4 GB at this context; duplicate traces of the same session: 9h22t197, cks8h82r, oyqerbew, s7t3o3gl, tgh3o3v3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2222.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,222 72.9 69.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mrl5f378",
   "source_url": "https://omlx.ai/benchmarks/performance/mrl5f378",
   "tg_tokens": null,
   "tps": 72.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mrl5f378-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.3 GB at this context; duplicate traces of the same session: 9h22t197, cks8h82r, oyqerbew, s7t3o3gl, tgh3o3v3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2400.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,400 71.4 70.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mrl5f378",
   "source_url": "https://omlx.ai/benchmarks/performance/mrl5f378",
   "tg_tokens": null,
   "tps": 71.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mrl5f378-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: 9h22t197, cks8h82r, oyqerbew, s7t3o3gl, tgh3o3v3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2648.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,648 68.2 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mrl5f378",
   "source_url": "https://omlx.ai/benchmarks/performance/mrl5f378",
   "tg_tokens": null,
   "tps": 68.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mrl5f378-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.4 GB at this context; duplicate traces of the same session: 9h22t197, cks8h82r, oyqerbew, s7t3o3gl, tgh3o3v3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2548.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,548 70.4 71.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mrl5f378",
   "source_url": "https://omlx.ai/benchmarks/performance/mrl5f378",
   "tg_tokens": null,
   "tps": 70.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mrl5f378-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.5 GB at this context; duplicate traces of the same session: 9h22t197, cks8h82r, oyqerbew, s7t3o3gl, tgh3o3v3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2486.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,486 64.0 72.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mrl5f378",
   "source_url": "https://omlx.ai/benchmarks/performance/mrl5f378",
   "tg_tokens": null,
   "tps": 64.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dponqcvt-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: h6wk5zqo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2787.0,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "gemma-4-26B-A4B-it-qat-nvfp4 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 2,787 119.9 16.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dponqcvt",
   "source_url": "https://omlx.ai/benchmarks/performance/dponqcvt",
   "tg_tokens": null,
   "tps": 119.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dponqcvt-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 16.5 GB at this context; duplicate traces of the same session: h6wk5zqo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3648.0,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "gemma-4-26B-A4B-it-qat-nvfp4 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 3,648 103.2 16.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dponqcvt",
   "source_url": "https://omlx.ai/benchmarks/performance/dponqcvt",
   "tg_tokens": null,
   "tps": 103.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-k0j01zqm-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 6.1 GB at this context; duplicate traces of the same session: 21m1xhfd; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1943.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E4B-it-MLX-4bit on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 1k 1,943 67.5 6.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k0j01zqm",
   "source_url": "https://omlx.ai/benchmarks/performance/k0j01zqm",
   "tg_tokens": null,
   "tps": 67.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-k0j01zqm-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: 21m1xhfd; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1896.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E4B-it-MLX-4bit on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 4k 1,896 65.8 6.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k0j01zqm",
   "source_url": "https://omlx.ai/benchmarks/performance/k0j01zqm",
   "tg_tokens": null,
   "tps": 65.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-lrcjt5sk-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 16k context; peak memory 16.7 GB; TTFT 6406 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2558.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "LFM2.5-24B-A2B-MLX-5bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 16k tokens 2,558 PP tok/s 61.2 TG tok/s 6406 TTFT (ms) 16.7 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lrcjt5sk",
   "source_url": "https://omlx.ai/benchmarks/performance/lrcjt5sk",
   "tg_tokens": null,
   "tps": 61.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-wf765xo4-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 32k context; peak memory 17.0 GB; TTFT 13813 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2372.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "LFM2.5-24B-A2B-MLX-5bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 32k tokens 2,372 PP tok/s 59.8 TG tok/s 13813 TTFT (ms) 17.0 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wf765xo4",
   "source_url": "https://omlx.ai/benchmarks/performance/wf765xo4",
   "tg_tokens": null,
   "tps": 59.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-ozu1jfay-ctx16384",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 3.1 GB at this context; duplicate traces of the same session: ezb5oqbl; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1100.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM5-2B-OptiQ-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 16k 1,100 35.3 3.1 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ozu1jfay",
   "source_url": "https://omlx.ai/benchmarks/performance/ozu1jfay",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-ozu1jfay-ctx32768",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 3.6 GB at this context; duplicate traces of the same session: ezb5oqbl; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 827.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM5-2B-OptiQ-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 32k 827.8 26.3 3.6 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ozu1jfay",
   "source_url": "https://omlx.ai/benchmarks/performance/ozu1jfay",
   "tg_tokens": null,
   "tps": 26.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ra9vodv4-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.5 GB at this context; duplicate traces of the same session: 2dplzq1t, i3kovjg1; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 666.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 6bit; 4k 666.1 39.8 8.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ra9vodv4",
   "source_url": "https://omlx.ai/benchmarks/performance/ra9vodv4",
   "tg_tokens": null,
   "tps": 39.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ra9vodv4-ctx16384",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.8 GB at this context; duplicate traces of the same session: 2dplzq1t, i3kovjg1; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 695.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 6bit; 16k 695.4 43.8 8.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ra9vodv4",
   "source_url": "https://omlx.ai/benchmarks/performance/ra9vodv4",
   "tg_tokens": null,
   "tps": 43.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ra9vodv4-ctx131072",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 13.1 GB at this context; duplicate traces of the same session: 2dplzq1t, i3kovjg1; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 350.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 6bit; 128k 350.3 21.2 13.1 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ra9vodv4",
   "source_url": "https://omlx.ai/benchmarks/performance/ra9vodv4",
   "tg_tokens": null,
   "tps": 21.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-on5baehp-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: 4rhbbff1; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 861.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 1k 861.6 86.8 5.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/on5baehp",
   "source_url": "https://omlx.ai/benchmarks/performance/on5baehp",
   "tg_tokens": null,
   "tps": 86.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-on5baehp-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.2 GB at this context; duplicate traces of the same session: 4rhbbff1; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 533.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 4k 533.3 15.7 7.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/on5baehp",
   "source_url": "https://omlx.ai/benchmarks/performance/on5baehp",
   "tg_tokens": null,
   "tps": 15.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-kdjkk35f-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 1gvikj68; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1054.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 1k 1,054 69.7 19.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kdjkk35f",
   "source_url": "https://omlx.ai/benchmarks/performance/kdjkk35f",
   "tg_tokens": null,
   "tps": 69.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-kdjkk35f-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 1gvikj68; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1262.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 4k 1,262 37.3 21.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kdjkk35f",
   "source_url": "https://omlx.ai/benchmarks/performance/kdjkk35f",
   "tg_tokens": null,
   "tps": 37.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-3raeyase-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: 0hap0tvz, 9dsst9q7, rjq10iu2, w4j0nl1x, wlxagfy1; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1814.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-MLX-6bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 6bit; 1k 1,814 82.4 27.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3raeyase",
   "source_url": "https://omlx.ai/benchmarks/performance/3raeyase",
   "tg_tokens": null,
   "tps": 82.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-3raeyase-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.2 GB at this context; duplicate traces of the same session: 0hap0tvz, 9dsst9q7, rjq10iu2, w4j0nl1x, wlxagfy1; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2254.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-MLX-6bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 6bit; 8k 2,254 77.0 28.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3raeyase",
   "source_url": "https://omlx.ai/benchmarks/performance/3raeyase",
   "tg_tokens": null,
   "tps": 77.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-3raeyase-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.4 GB at this context; duplicate traces of the same session: 0hap0tvz, 9dsst9q7, rjq10iu2, w4j0nl1x, wlxagfy1; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2099.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-MLX-6bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 6bit; 16k 2,099 72.6 28.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3raeyase",
   "source_url": "https://omlx.ai/benchmarks/performance/3raeyase",
   "tg_tokens": null,
   "tps": 72.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-3raeyase-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.7 GB at this context; duplicate traces of the same session: 0hap0tvz, 9dsst9q7, rjq10iu2, w4j0nl1x, wlxagfy1; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1824.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-MLX-6bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 6bit; 32k 1,824 64.0 28.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3raeyase",
   "source_url": "https://omlx.ai/benchmarks/performance/3raeyase",
   "tg_tokens": null,
   "tps": 64.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-3raeyase-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 29.3 GB at this context; duplicate traces of the same session: 0hap0tvz, 9dsst9q7, rjq10iu2, w4j0nl1x, wlxagfy1; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1431.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-MLX-6bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 6bit; 64k 1,431 51.9 29.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3raeyase",
   "source_url": "https://omlx.ai/benchmarks/performance/3raeyase",
   "tg_tokens": null,
   "tps": 51.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-3raeyase-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 30.6 GB at this context; duplicate traces of the same session: 0hap0tvz, 9dsst9q7, rjq10iu2, w4j0nl1x, wlxagfy1; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 996.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-MLX-6bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 6bit; 128k 996.4 38.0 30.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3raeyase",
   "source_url": "https://omlx.ai/benchmarks/performance/3raeyase",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-3raeyase-ctx199680",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 31.9 GB at this context; duplicate traces of the same session: 0hap0tvz, 9dsst9q7, rjq10iu2, w4j0nl1x, wlxagfy1; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 755.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-MLX-6bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 6bit; 195k 755.6 29.5 31.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3raeyase",
   "source_url": "https://omlx.ai/benchmarks/performance/3raeyase",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-yymihoju-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: tngneikx; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1812.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-MLX-6bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 6bit; 1k 1,812 81.7 27.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yymihoju",
   "source_url": "https://omlx.ai/benchmarks/performance/yymihoju",
   "tg_tokens": null,
   "tps": 81.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-yymihoju-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.2 GB at this context; duplicate traces of the same session: tngneikx; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2248.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-MLX-6bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 6bit; 8k 2,248 76.5 28.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yymihoju",
   "source_url": "https://omlx.ai/benchmarks/performance/yymihoju",
   "tg_tokens": null,
   "tps": 76.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-yymihoju-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.4 GB at this context; duplicate traces of the same session: tngneikx; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2065.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-MLX-6bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 6bit; 16k 2,065 72.2 28.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yymihoju",
   "source_url": "https://omlx.ai/benchmarks/performance/yymihoju",
   "tg_tokens": null,
   "tps": 72.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-w9f05pac-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: gp44pmsx; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 752.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 752.9 25.5 16.0 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w9f05pac",
   "source_url": "https://omlx.ai/benchmarks/performance/w9f05pac",
   "tg_tokens": null,
   "tps": 25.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-w9f05pac-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: gp44pmsx; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 454.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 454.5 18.5 17.0 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w9f05pac",
   "source_url": "https://omlx.ai/benchmarks/performance/w9f05pac",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-2uaxzz76-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: m45kbtz4; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4-bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 130.8 17.9 21.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2uaxzz76",
   "source_url": "https://omlx.ai/benchmarks/performance/2uaxzz76",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-2uaxzz76-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: m45kbtz4; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4-bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 125.1 20.8 21.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2uaxzz76",
   "source_url": "https://omlx.ai/benchmarks/performance/2uaxzz76",
   "tg_tokens": null,
   "tps": 20.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-lzn83k7l-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: rx0n6ntb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 100.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 100.8 25.3 25.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lzn83k7l",
   "source_url": "https://omlx.ai/benchmarks/performance/lzn83k7l",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-lzn83k7l-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.2 GB at this context; duplicate traces of the same session: rx0n6ntb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 81.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 128k 81.2 20.5 30.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lzn83k7l",
   "source_url": "https://omlx.ai/benchmarks/performance/lzn83k7l",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xpytd3qt-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.3 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 543.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 543.1 20.5 72.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xpytd3qt",
   "source_url": "https://omlx.ai/benchmarks/performance/xpytd3qt",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xpytd3qt-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.5 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 921.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k 921.8 20.2 74.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xpytd3qt",
   "source_url": "https://omlx.ai/benchmarks/performance/xpytd3qt",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-hkt3b0pw-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 509.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 509.1 24.4 71.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hkt3b0pw",
   "source_url": "https://omlx.ai/benchmarks/performance/hkt3b0pw",
   "tg_tokens": null,
   "tps": 24.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-hkt3b0pw-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 586.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 586.8 42.7 71.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hkt3b0pw",
   "source_url": "https://omlx.ai/benchmarks/performance/hkt3b0pw",
   "tg_tokens": null,
   "tps": 42.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-hkt3b0pw-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 819.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 819.9 55.3 71.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hkt3b0pw",
   "source_url": "https://omlx.ai/benchmarks/performance/hkt3b0pw",
   "tg_tokens": null,
   "tps": 55.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-hkt3b0pw-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 820.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 820.2 30.8 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hkt3b0pw",
   "source_url": "https://omlx.ai/benchmarks/performance/hkt3b0pw",
   "tg_tokens": null,
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-otoswfjh-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.1 GB at this context; duplicate traces of the same session: b6qwt5e6, beys2rib, l35qa46e, p4bac9ca, x38dewhl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1119.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,119 71.6 69.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/otoswfjh",
   "source_url": "https://omlx.ai/benchmarks/performance/otoswfjh",
   "tg_tokens": null,
   "tps": 71.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-otoswfjh-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.4 GB at this context; duplicate traces of the same session: b6qwt5e6, beys2rib, l35qa46e, p4bac9ca, x38dewhl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2340.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,340 72.3 69.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/otoswfjh",
   "source_url": "https://omlx.ai/benchmarks/performance/otoswfjh",
   "tg_tokens": null,
   "tps": 72.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-otoswfjh-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.1 GB at this context; duplicate traces of the same session: b6qwt5e6, beys2rib, l35qa46e, p4bac9ca, x38dewhl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2587.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,587 72.2 70.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/otoswfjh",
   "source_url": "https://omlx.ai/benchmarks/performance/otoswfjh",
   "tg_tokens": null,
   "tps": 72.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-otoswfjh-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: b6qwt5e6, beys2rib, l35qa46e, p4bac9ca, x38dewhl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2676.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,676 73.2 70.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/otoswfjh",
   "source_url": "https://omlx.ai/benchmarks/performance/otoswfjh",
   "tg_tokens": null,
   "tps": 73.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-otoswfjh-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.4 GB at this context; duplicate traces of the same session: b6qwt5e6, beys2rib, l35qa46e, p4bac9ca, x38dewhl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2460.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,460 67.7 71.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/otoswfjh",
   "source_url": "https://omlx.ai/benchmarks/performance/otoswfjh",
   "tg_tokens": null,
   "tps": 67.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-otoswfjh-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.5 GB at this context; duplicate traces of the same session: b6qwt5e6, beys2rib, l35qa46e, p4bac9ca, x38dewhl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2370.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,370 66.2 72.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/otoswfjh",
   "source_url": "https://omlx.ai/benchmarks/performance/otoswfjh",
   "tg_tokens": null,
   "tps": 66.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-upranmv5-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.1 GB at this context; duplicate traces of the same session: bql9iows, n4pdl1iz, rvdjc546, xfw7nha4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1029.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,029 70.8 69.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/upranmv5",
   "source_url": "https://omlx.ai/benchmarks/performance/upranmv5",
   "tg_tokens": null,
   "tps": 70.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-upranmv5-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.4 GB at this context; duplicate traces of the same session: bql9iows, n4pdl1iz, rvdjc546, xfw7nha4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2183.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,183 67.5 69.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/upranmv5",
   "source_url": "https://omlx.ai/benchmarks/performance/upranmv5",
   "tg_tokens": null,
   "tps": 67.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-upranmv5-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.3 GB at this context; duplicate traces of the same session: bql9iows, n4pdl1iz, rvdjc546, xfw7nha4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2430.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,430 66.5 70.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/upranmv5",
   "source_url": "https://omlx.ai/benchmarks/performance/upranmv5",
   "tg_tokens": null,
   "tps": 66.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-upranmv5-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: bql9iows, n4pdl1iz, rvdjc546, xfw7nha4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2660.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,660 67.3 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/upranmv5",
   "source_url": "https://omlx.ai/benchmarks/performance/upranmv5",
   "tg_tokens": null,
   "tps": 67.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-upranmv5-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.5 GB at this context; duplicate traces of the same session: bql9iows, n4pdl1iz, rvdjc546, xfw7nha4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2414.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,414 62.8 72.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/upranmv5",
   "source_url": "https://omlx.ai/benchmarks/performance/upranmv5",
   "tg_tokens": null,
   "tps": 62.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-nd21gwl2-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 3bjgxm7b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 218.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 218.3 33.2 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nd21gwl2",
   "source_url": "https://omlx.ai/benchmarks/performance/nd21gwl2",
   "tg_tokens": null,
   "tps": 33.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-nd21gwl2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 3bjgxm7b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 221.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 221.4 31.8 21.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nd21gwl2",
   "source_url": "https://omlx.ai/benchmarks/performance/nd21gwl2",
   "tg_tokens": null,
   "tps": 31.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-au80hyq5-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: wmdy6r1z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3806.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 3,806 103.2 14.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/au80hyq5",
   "source_url": "https://omlx.ai/benchmarks/performance/au80hyq5",
   "tg_tokens": null,
   "tps": 103.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-au80hyq5-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: wmdy6r1z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3741.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 3,741 99.6 14.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/au80hyq5",
   "source_url": "https://omlx.ai/benchmarks/performance/au80hyq5",
   "tg_tokens": null,
   "tps": 99.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-78pzhabj-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace, single run at 1k context; peak memory 14.3 GB; TTFT 262 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3904.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 3,904 PP tok/s 109.4 TG tok/s 262 TTFT (ms) 14.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/78pzhabj",
   "source_url": "https://omlx.ai/benchmarks/performance/78pzhabj",
   "tg_tokens": null,
   "tps": 109.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2fxor85j-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: f6n01ghd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3139.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 3,139 100.4 14.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2fxor85j",
   "source_url": "https://omlx.ai/benchmarks/performance/2fxor85j",
   "tg_tokens": null,
   "tps": 100.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2fxor85j-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: f6n01ghd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2669.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 2,669 99.9 14.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2fxor85j",
   "source_url": "https://omlx.ai/benchmarks/performance/2fxor85j",
   "tg_tokens": null,
   "tps": 99.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-zu974p84-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: awm9zbpp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3670.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 3,670 108.2 14.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zu974p84",
   "source_url": "https://omlx.ai/benchmarks/performance/zu974p84",
   "tg_tokens": null,
   "tps": 108.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zu974p84-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: awm9zbpp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3837.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 3,837 106.8 14.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zu974p84",
   "source_url": "https://omlx.ai/benchmarks/performance/zu974p84",
   "tg_tokens": null,
   "tps": 106.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ui4afshp-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.3 GB at this context; duplicate traces of the same session: 8gk1hkd9, akt39hnj, g660dlkh, pt1z1xjy, qw6kij9r, tp5khghb, ultsg5yd; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2430.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "gpt-oss-20b-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 2,430 94.8 11.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ui4afshp",
   "source_url": "https://omlx.ai/benchmarks/performance/ui4afshp",
   "tg_tokens": null,
   "tps": 94.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ui4afshp-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.4 GB at this context; duplicate traces of the same session: 8gk1hkd9, akt39hnj, g660dlkh, pt1z1xjy, qw6kij9r, tp5khghb, ultsg5yd; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2526.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "gpt-oss-20b-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,526 89.0 11.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ui4afshp",
   "source_url": "https://omlx.ai/benchmarks/performance/ui4afshp",
   "tg_tokens": null,
   "tps": 89.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ui4afshp-ctx8192",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.4 GB at this context; duplicate traces of the same session: 8gk1hkd9, akt39hnj, g660dlkh, pt1z1xjy, qw6kij9r, tp5khghb, ultsg5yd; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2440.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "gpt-oss-20b-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 2,440 81.5 11.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ui4afshp",
   "source_url": "https://omlx.ai/benchmarks/performance/ui4afshp",
   "tg_tokens": null,
   "tps": 81.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ui4afshp-ctx16384",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.6 GB at this context; duplicate traces of the same session: 8gk1hkd9, akt39hnj, g660dlkh, pt1z1xjy, qw6kij9r, tp5khghb, ultsg5yd; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2272.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "gpt-oss-20b-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 2,272 71.5 11.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ui4afshp",
   "source_url": "https://omlx.ai/benchmarks/performance/ui4afshp",
   "tg_tokens": null,
   "tps": 71.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ui4afshp-ctx32768",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 12.0 GB at this context; duplicate traces of the same session: 8gk1hkd9, akt39hnj, g660dlkh, pt1z1xjy, qw6kij9r, tp5khghb, ultsg5yd; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1913.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "gpt-oss-20b-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 1,913 57.4 12.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ui4afshp",
   "source_url": "https://omlx.ai/benchmarks/performance/ui4afshp",
   "tg_tokens": null,
   "tps": 57.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ui4afshp-ctx65536",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: 8gk1hkd9, akt39hnj, g660dlkh, pt1z1xjy, qw6kij9r, tp5khghb, ultsg5yd; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1536.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "gpt-oss-20b-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 64k 1,536 41.1 12.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ui4afshp",
   "source_url": "https://omlx.ai/benchmarks/performance/ui4afshp",
   "tg_tokens": null,
   "tps": 41.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ui4afshp-ctx131072",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: 8gk1hkd9, akt39hnj, g660dlkh, pt1z1xjy, qw6kij9r, tp5khghb, ultsg5yd; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1145.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "gpt-oss-20b-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 128k 1,145 25.7 14.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ui4afshp",
   "source_url": "https://omlx.ai/benchmarks/performance/ui4afshp",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ui4afshp-ctx199680",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: 8gk1hkd9, akt39hnj, g660dlkh, pt1z1xjy, qw6kij9r, tp5khghb, ultsg5yd; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 868.6,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "gpt-oss-20b-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 195k 868.6 18.3 15.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ui4afshp",
   "source_url": "https://omlx.ai/benchmarks/performance/ui4afshp",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-06rq9lrc-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.3 GB at this context; duplicate traces of the same session: 7u46dvw5, tj62ktye; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2427.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "gpt-oss-20b-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 2,427 94.6 11.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/06rq9lrc",
   "source_url": "https://omlx.ai/benchmarks/performance/06rq9lrc",
   "tg_tokens": null,
   "tps": 94.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-06rq9lrc-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.4 GB at this context; duplicate traces of the same session: 7u46dvw5, tj62ktye; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2518.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "gpt-oss-20b-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,518 89.1 11.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/06rq9lrc",
   "source_url": "https://omlx.ai/benchmarks/performance/06rq9lrc",
   "tg_tokens": null,
   "tps": 89.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-06rq9lrc-ctx8192",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.4 GB at this context; duplicate traces of the same session: 7u46dvw5, tj62ktye; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2431.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "gpt-oss-20b-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 2,431 81.5 11.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/06rq9lrc",
   "source_url": "https://omlx.ai/benchmarks/performance/06rq9lrc",
   "tg_tokens": null,
   "tps": 81.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-06rq9lrc-ctx16384",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.6 GB at this context; duplicate traces of the same session: 7u46dvw5, tj62ktye; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2246.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "gpt-oss-20b-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 2,246 72.5 11.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/06rq9lrc",
   "source_url": "https://omlx.ai/benchmarks/performance/06rq9lrc",
   "tg_tokens": null,
   "tps": 72.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ptrcked4-ctx1024",
   "model": "gpt-oss-120b-MXFP4-Q8",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 59.5 GB at this context; duplicate traces of the same session: gloe41js; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2021.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community__gpt-oss-120b-MXFP4-Q8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,021 94.2 59.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ptrcked4",
   "source_url": "https://omlx.ai/benchmarks/performance/ptrcked4",
   "tg_tokens": null,
   "tps": 94.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ptrcked4-ctx4096",
   "model": "gpt-oss-120b-MXFP4-Q8",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 59.6 GB at this context; duplicate traces of the same session: gloe41js; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2193.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community__gpt-oss-120b-MXFP4-Q8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,193 90.2 59.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ptrcked4",
   "source_url": "https://omlx.ai/benchmarks/performance/ptrcked4",
   "tg_tokens": null,
   "tps": 90.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hf41kfe8-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 90.0 GB at this context; duplicate traces of the same session: lsg7fv44, qhp4j7mo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 86.8,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 86.8 9.7 90.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hf41kfe8",
   "source_url": "https://omlx.ai/benchmarks/performance/hf41kfe8",
   "tg_tokens": null,
   "tps": 9.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hf41kfe8-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.0 GB at this context; duplicate traces of the same session: lsg7fv44, qhp4j7mo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 373.8,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 373.8 11.0 92.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hf41kfe8",
   "source_url": "https://omlx.ai/benchmarks/performance/hf41kfe8",
   "tg_tokens": null,
   "tps": 11.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hf41kfe8-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 91.9 GB at this context; duplicate traces of the same session: lsg7fv44, qhp4j7mo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 417.9,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 417.9 12.3 91.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hf41kfe8",
   "source_url": "https://omlx.ai/benchmarks/performance/hf41kfe8",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2iwm7d22-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.7 GB at this context; duplicate traces of the same session: 1gzmrkt4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 440.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 440.1 40.6 173.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2iwm7d22",
   "source_url": "https://omlx.ai/benchmarks/performance/2iwm7d22",
   "tg_tokens": null,
   "tps": 40.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2iwm7d22-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.8 GB at this context; duplicate traces of the same session: 1gzmrkt4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 547.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 547.2 38.8 174.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2iwm7d22",
   "source_url": "https://omlx.ai/benchmarks/performance/2iwm7d22",
   "tg_tokens": null,
   "tps": 38.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2iwm7d22-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.3 GB at this context; duplicate traces of the same session: 1gzmrkt4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 528.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 528.1 38.0 175.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2iwm7d22",
   "source_url": "https://omlx.ai/benchmarks/performance/2iwm7d22",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-mpcxyr7u-ctx1024",
   "model": "Gemma-4-E4B-Claude-Opus-Reasoning",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 4.5 GB at this context; duplicate traces of the same session: s9bv8ql4; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 655.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Gemma-4-E4B-Claude-Opus-Reasoning-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 655.9 31.3 4.5 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mpcxyr7u",
   "source_url": "https://omlx.ai/benchmarks/performance/mpcxyr7u",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-mpcxyr7u-ctx4096",
   "model": "Gemma-4-E4B-Claude-Opus-Reasoning",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 4.5 GB at this context; duplicate traces of the same session: s9bv8ql4; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 662.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Gemma-4-E4B-Claude-Opus-Reasoning-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 662.5 30.8 4.5 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mpcxyr7u",
   "source_url": "https://omlx.ai/benchmarks/performance/mpcxyr7u",
   "tg_tokens": null,
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-n4h9xp2o-ctx1024",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "g64; oMLX community benchmark trace; peak memory 9.8 GB at this context; duplicate traces of the same session: 32lbkk0v, tkwe4k2e, zzc89hcj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2248.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-mlx-6bit-g64 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 2,248 109.5 9.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n4h9xp2o",
   "source_url": "https://omlx.ai/benchmarks/performance/n4h9xp2o",
   "tg_tokens": null,
   "tps": 109.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-n4h9xp2o-ctx4096",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "g64; oMLX community benchmark trace; peak memory 9.9 GB at this context; duplicate traces of the same session: 32lbkk0v, tkwe4k2e, zzc89hcj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2606.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-mlx-6bit-g64 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 2,606 108.7 9.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n4h9xp2o",
   "source_url": "https://omlx.ai/benchmarks/performance/n4h9xp2o",
   "tg_tokens": null,
   "tps": 108.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-n4h9xp2o-ctx16384",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "g64; oMLX community benchmark trace; peak memory 10.1 GB at this context; duplicate traces of the same session: 32lbkk0v, tkwe4k2e, zzc89hcj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3190.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-mlx-6bit-g64 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 16k 3,190 99.5 10.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n4h9xp2o",
   "source_url": "https://omlx.ai/benchmarks/performance/n4h9xp2o",
   "tg_tokens": null,
   "tps": 99.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-n4h9xp2o-ctx65536",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "g64; oMLX community benchmark trace; peak memory 10.7 GB at this context; duplicate traces of the same session: 32lbkk0v, tkwe4k2e, zzc89hcj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2464.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-mlx-6bit-g64 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 64k 2,464 77.0 10.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n4h9xp2o",
   "source_url": "https://omlx.ai/benchmarks/performance/n4h9xp2o",
   "tg_tokens": null,
   "tps": 77.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-1silqhsh-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 889eptxy, bwlxw0pe, fj3yp8f5, nbfha9w0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1664.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,664 141.5 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1silqhsh",
   "source_url": "https://omlx.ai/benchmarks/performance/1silqhsh",
   "tg_tokens": null,
   "tps": 141.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-1silqhsh-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 889eptxy, bwlxw0pe, fj3yp8f5, nbfha9w0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2092.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,092 137.7 21.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1silqhsh",
   "source_url": "https://omlx.ai/benchmarks/performance/1silqhsh",
   "tg_tokens": null,
   "tps": 137.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-1silqhsh-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 889eptxy, bwlxw0pe, fj3yp8f5, nbfha9w0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1849.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,849 126.7 22.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1silqhsh",
   "source_url": "https://omlx.ai/benchmarks/performance/1silqhsh",
   "tg_tokens": null,
   "tps": 126.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-1silqhsh-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 889eptxy, bwlxw0pe, fj3yp8f5, nbfha9w0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1517.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,517 118.0 21.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1silqhsh",
   "source_url": "https://omlx.ai/benchmarks/performance/1silqhsh",
   "tg_tokens": null,
   "tps": 118.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-1silqhsh-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: 889eptxy, bwlxw0pe, fj3yp8f5, nbfha9w0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1191.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,191 102.6 21.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1silqhsh",
   "source_url": "https://omlx.ai/benchmarks/performance/1silqhsh",
   "tg_tokens": null,
   "tps": 102.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3s03zzj7-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 8edp0jxw, j17svyke; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4814.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,814 151.9 19.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3s03zzj7",
   "source_url": "https://omlx.ai/benchmarks/performance/3s03zzj7",
   "tg_tokens": null,
   "tps": 151.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3s03zzj7-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 8edp0jxw, j17svyke; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2333.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,333 148.8 20.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3s03zzj7",
   "source_url": "https://omlx.ai/benchmarks/performance/3s03zzj7",
   "tg_tokens": null,
   "tps": 148.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3s03zzj7-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 8edp0jxw, j17svyke; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4498.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 4,498 122.6 20.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3s03zzj7",
   "source_url": "https://omlx.ai/benchmarks/performance/3s03zzj7",
   "tg_tokens": null,
   "tps": 122.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kikudftp-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.3 GB; TTFT 1245 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 822.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 822.8 PP tok/s 146.4 TG tok/s 1245 TTFT (ms) 19.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kikudftp",
   "source_url": "https://omlx.ai/benchmarks/performance/kikudftp",
   "tg_tokens": null,
   "tps": 146.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-sc7za9xy-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: ngd9yxcl, oy6gn7sp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1787.0,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Ornith-1.5-35B-A3B-mlx-nvfp4 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,787 118.3 20.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sc7za9xy",
   "source_url": "https://omlx.ai/benchmarks/performance/sc7za9xy",
   "tg_tokens": null,
   "tps": 118.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-sc7za9xy-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: ngd9yxcl, oy6gn7sp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2078.0,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Ornith-1.5-35B-A3B-mlx-nvfp4 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 2,078 114.1 22.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sc7za9xy",
   "source_url": "https://omlx.ai/benchmarks/performance/sc7za9xy",
   "tg_tokens": null,
   "tps": 114.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-sc7za9xy-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: ngd9yxcl, oy6gn7sp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1817.0,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Ornith-1.5-35B-A3B-mlx-nvfp4 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,817 107.2 22.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sc7za9xy",
   "source_url": "https://omlx.ai/benchmarks/performance/sc7za9xy",
   "tg_tokens": null,
   "tps": 107.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-hq2e8uji-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: apmlhdn2, ls6ac3ud; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1776.0,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Ornith-1.5-35B-A3B-mlx-nvfp4 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,776 119.0 20.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hq2e8uji",
   "source_url": "https://omlx.ai/benchmarks/performance/hq2e8uji",
   "tg_tokens": null,
   "tps": 119.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-hq2e8uji-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: apmlhdn2, ls6ac3ud; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2073.0,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Ornith-1.5-35B-A3B-mlx-nvfp4 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 2,073 115.4 22.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hq2e8uji",
   "source_url": "https://omlx.ai/benchmarks/performance/hq2e8uji",
   "tg_tokens": null,
   "tps": 115.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-hq2e8uji-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: apmlhdn2, ls6ac3ud; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1885.0,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Ornith-1.5-35B-A3B-mlx-nvfp4 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,885 111.2 22.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hq2e8uji",
   "source_url": "https://omlx.ai/benchmarks/performance/hq2e8uji",
   "tg_tokens": null,
   "tps": 111.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-wty323kk-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 26kxw99p, ciqqxxhs; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1665.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,665 116.5 21.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wty323kk",
   "source_url": "https://omlx.ai/benchmarks/performance/wty323kk",
   "tg_tokens": null,
   "tps": 116.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-wty323kk-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 26kxw99p, ciqqxxhs; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2031.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 2,031 118.2 23.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wty323kk",
   "source_url": "https://omlx.ai/benchmarks/performance/wty323kk",
   "tg_tokens": null,
   "tps": 118.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-wty323kk-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 26kxw99p, ciqqxxhs; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1761.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,761 97.7 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wty323kk",
   "source_url": "https://omlx.ai/benchmarks/performance/wty323kk",
   "tg_tokens": null,
   "tps": 97.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-2polhuw5-ctx1024",
   "model": "Qwen2.5-Coder-7B-Instruct",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 4.8 GB; TTFT 3275 ms; duplicate traces of the same session: none; config=19c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 312.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-7B-Instruct-4bit on M2 Pro (19c); M2 Pro (19c) \u00b7 16 GB \u00b7 4bit; 1k tokens 312.7 PP tok/s 38.4 TG tok/s 3275 TTFT (ms) 4.8 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2polhuw5",
   "source_url": "https://omlx.ai/benchmarks/performance/2polhuw5",
   "tg_tokens": null,
   "tps": 38.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-r3hdxrlc-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.8 GB at this context; duplicate traces of the same session: r41wod97; config=5c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 166.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M1 (5c); M1 (5c) \u00b7 8 GB \u00b7 4bit; 4k 166.3 18.5 3.8 GB view",
   "ram_gb": 8,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r3hdxrlc",
   "source_url": "https://omlx.ai/benchmarks/performance/r3hdxrlc",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-r3hdxrlc-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.5 GB at this context; duplicate traces of the same session: r41wod97; config=5c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M1 (5c); M1 (5c) \u00b7 8 GB \u00b7 4bit; 16k 130.2 15.6 4.5 GB current",
   "ram_gb": 8,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r3hdxrlc",
   "source_url": "https://omlx.ai/benchmarks/performance/r3hdxrlc",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-u8mg2kya-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: nr8p9rs7; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 195.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 4bit; 1k 195.7 18.9 5.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u8mg2kya",
   "source_url": "https://omlx.ai/benchmarks/performance/u8mg2kya",
   "tg_tokens": null,
   "tps": 18.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-u8mg2kya-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: nr8p9rs7; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 202.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 4bit; 4k 202.3 18.7 6.2 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u8mg2kya",
   "source_url": "https://omlx.ai/benchmarks/performance/u8mg2kya",
   "tg_tokens": null,
   "tps": 18.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-gldjtz98-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: bsurgvgk, dwtiui22; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1858.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-MLX-6bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 6bit; 1k 1,858 82.3 27.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gldjtz98",
   "source_url": "https://omlx.ai/benchmarks/performance/gldjtz98",
   "tg_tokens": null,
   "tps": 82.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-gldjtz98-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.2 GB at this context; duplicate traces of the same session: bsurgvgk, dwtiui22; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2295.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-MLX-6bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 6bit; 8k 2,295 76.8 28.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gldjtz98",
   "source_url": "https://omlx.ai/benchmarks/performance/gldjtz98",
   "tg_tokens": null,
   "tps": 76.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-gldjtz98-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.4 GB at this context; duplicate traces of the same session: bsurgvgk, dwtiui22; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2137.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-MLX-6bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 6bit; 16k 2,137 72.1 28.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gldjtz98",
   "source_url": "https://omlx.ai/benchmarks/performance/gldjtz98",
   "tg_tokens": null,
   "tps": 72.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-bnsqbfhk-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1812.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-MLX-6bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 6bit; 1k 1,812 81.7 27.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bnsqbfhk",
   "source_url": "https://omlx.ai/benchmarks/performance/bnsqbfhk",
   "tg_tokens": null,
   "tps": 81.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-bnsqbfhk-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.2 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2248.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-MLX-6bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 6bit; 8k 2,248 76.5 28.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bnsqbfhk",
   "source_url": "https://omlx.ai/benchmarks/performance/bnsqbfhk",
   "tg_tokens": null,
   "tps": 76.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-bnsqbfhk-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2065.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-MLX-6bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 6bit; 16k 2,065 72.2 28.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bnsqbfhk",
   "source_url": "https://omlx.ai/benchmarks/performance/bnsqbfhk",
   "tg_tokens": null,
   "tps": 72.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-bo2m2xgh-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: ej68rkph, ledld07k, o4i90yn1; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 403.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 403.4 17.9 16.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bo2m2xgh",
   "source_url": "https://omlx.ai/benchmarks/performance/bo2m2xgh",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-bo2m2xgh-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: ej68rkph, ledld07k, o4i90yn1; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 419.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 419.5 17.4 17.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bo2m2xgh",
   "source_url": "https://omlx.ai/benchmarks/performance/bo2m2xgh",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-bo2m2xgh-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: ej68rkph, ledld07k, o4i90yn1; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 385.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 385.8 16.8 17.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bo2m2xgh",
   "source_url": "https://omlx.ai/benchmarks/performance/bo2m2xgh",
   "tg_tokens": null,
   "tps": 16.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-bo2m2xgh-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: ej68rkph, ledld07k, o4i90yn1; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 377.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 377.9 15.8 17.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bo2m2xgh",
   "source_url": "https://omlx.ai/benchmarks/performance/bo2m2xgh",
   "tg_tokens": null,
   "tps": 15.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-rlpaid1z-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: ceortkf4, nzs30smb; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 402.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 402.3 17.8 16.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rlpaid1z",
   "source_url": "https://omlx.ai/benchmarks/performance/rlpaid1z",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-rlpaid1z-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: ceortkf4, nzs30smb; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 418.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 418.8 17.4 17.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rlpaid1z",
   "source_url": "https://omlx.ai/benchmarks/performance/rlpaid1z",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-rlpaid1z-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: ceortkf4, nzs30smb; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 418.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 418.0 16.7 17.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rlpaid1z",
   "source_url": "https://omlx.ai/benchmarks/performance/rlpaid1z",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-f7ggcgg2-ctx16384",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 40.9 GB at this context; duplicate traces of the same session: o91tobpl; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 309.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 309.7 31.3 40.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f7ggcgg2",
   "source_url": "https://omlx.ai/benchmarks/performance/f7ggcgg2",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-f7ggcgg2-ctx32768",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 41.3 GB at this context; duplicate traces of the same session: o91tobpl; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 332.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 32k 332.1 23.3 41.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f7ggcgg2",
   "source_url": "https://omlx.ai/benchmarks/performance/f7ggcgg2",
   "tg_tokens": null,
   "tps": 23.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-f7ggcgg2-ctx65536",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 42.3 GB at this context; duplicate traces of the same session: o91tobpl; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 328.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 64k 328.7 26.3 42.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f7ggcgg2",
   "source_url": "https://omlx.ai/benchmarks/performance/f7ggcgg2",
   "tg_tokens": null,
   "tps": 26.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-0g1gzajd-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.3 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 543.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 543.1 20.5 72.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0g1gzajd",
   "source_url": "https://omlx.ai/benchmarks/performance/0g1gzajd",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-0g1gzajd-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.5 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 921.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k 921.8 20.2 74.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0g1gzajd",
   "source_url": "https://omlx.ai/benchmarks/performance/0g1gzajd",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-67wpvtpm-ctx1024",
   "model": "gemma-4-31B-it-uncensored",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 0k8lq01c, 4pcxxser; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 189.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31B-it-uncensored-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 189.5 21.3 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/67wpvtpm",
   "source_url": "https://omlx.ai/benchmarks/performance/67wpvtpm",
   "tg_tokens": null,
   "tps": 21.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-67wpvtpm-ctx4096",
   "model": "gemma-4-31B-it-uncensored",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 0k8lq01c, 4pcxxser; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 160.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31B-it-uncensored-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 160.2 20.0 18.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/67wpvtpm",
   "source_url": "https://omlx.ai/benchmarks/performance/67wpvtpm",
   "tg_tokens": null,
   "tps": 20.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-67wpvtpm-ctx8192",
   "model": "gemma-4-31B-it-uncensored",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 0k8lq01c, 4pcxxser; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 152.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31B-it-uncensored-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 152.4 18.0 19.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/67wpvtpm",
   "source_url": "https://omlx.ai/benchmarks/performance/67wpvtpm",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-xr9rdo9u-ctx1024",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 5.1 GB at this context; duplicate traces of the same session: 3eqdyxjt; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2207.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-e2b-8bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 8bit; 1k 2,207 34.9 5.1 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xr9rdo9u",
   "source_url": "https://omlx.ai/benchmarks/performance/xr9rdo9u",
   "tg_tokens": null,
   "tps": 34.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-xr9rdo9u-ctx4096",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 5.2 GB at this context; duplicate traces of the same session: 3eqdyxjt; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2306.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-e2b-8bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 8bit; 4k 2,306 34.0 5.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xr9rdo9u",
   "source_url": "https://omlx.ai/benchmarks/performance/xr9rdo9u",
   "tg_tokens": null,
   "tps": 34.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-a5jf9e1e-ctx1024",
   "model": "gemma-4-E2B-it",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 4.5 GB at this context; duplicate traces of the same session: zbcwhdwx; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2272.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "gemma-4-e2b-it-qat-OptiQ-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 2,272 46.2 4.5 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a5jf9e1e",
   "source_url": "https://omlx.ai/benchmarks/performance/a5jf9e1e",
   "tg_tokens": null,
   "tps": 46.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-a5jf9e1e-ctx4096",
   "model": "gemma-4-E2B-it",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 4.5 GB at this context; duplicate traces of the same session: zbcwhdwx; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2315.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "gemma-4-e2b-it-qat-OptiQ-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 2,315 45.5 4.5 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a5jf9e1e",
   "source_url": "https://omlx.ai/benchmarks/performance/a5jf9e1e",
   "tg_tokens": null,
   "tps": 45.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-d6541z1v-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 4.5 GB at this context; duplicate traces of the same session: gdws6oa4; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 654.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 654.3 31.2 4.5 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d6541z1v",
   "source_url": "https://omlx.ai/benchmarks/performance/d6541z1v",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-d6541z1v-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 4.6 GB at this context; duplicate traces of the same session: gdws6oa4; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 662.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 662.6 31.2 4.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d6541z1v",
   "source_url": "https://omlx.ai/benchmarks/performance/d6541z1v",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-jbjag4dk-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 4.5 GB at this context; duplicate traces of the same session: qsy3z4n0; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 654.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 654.5 31.2 4.5 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jbjag4dk",
   "source_url": "https://omlx.ai/benchmarks/performance/jbjag4dk",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-jbjag4dk-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 4.6 GB at this context; duplicate traces of the same session: qsy3z4n0; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 662.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 662.0 30.7 4.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jbjag4dk",
   "source_url": "https://omlx.ai/benchmarks/performance/jbjag4dk",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-inqfkky7-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.3 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2427.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "gpt-oss-20b-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 2,427 94.6 11.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/inqfkky7",
   "source_url": "https://omlx.ai/benchmarks/performance/inqfkky7",
   "tg_tokens": null,
   "tps": 94.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-inqfkky7-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2518.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "gpt-oss-20b-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,518 89.1 11.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/inqfkky7",
   "source_url": "https://omlx.ai/benchmarks/performance/inqfkky7",
   "tg_tokens": null,
   "tps": 89.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-inqfkky7-ctx8192",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2431.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "gpt-oss-20b-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 2,431 81.5 11.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/inqfkky7",
   "source_url": "https://omlx.ai/benchmarks/performance/inqfkky7",
   "tg_tokens": null,
   "tps": 81.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-inqfkky7-ctx16384",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.6 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2246.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "gpt-oss-20b-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 2,246 72.5 11.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/inqfkky7",
   "source_url": "https://omlx.ai/benchmarks/performance/inqfkky7",
   "tg_tokens": null,
   "tps": 72.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-o75hmaxi-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 173.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 440.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 440.1 40.6 173.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o75hmaxi",
   "source_url": "https://omlx.ai/benchmarks/performance/o75hmaxi",
   "tg_tokens": null,
   "tps": 40.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-o75hmaxi-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.8 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 547.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 547.2 38.8 174.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o75hmaxi",
   "source_url": "https://omlx.ai/benchmarks/performance/o75hmaxi",
   "tg_tokens": null,
   "tps": 38.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-o75hmaxi-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.3 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 528.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 528.1 38.0 175.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o75hmaxi",
   "source_url": "https://omlx.ai/benchmarks/performance/o75hmaxi",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-mppx3zjx-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 192.6 GB at this context; duplicate traces of the same session: a5v0qsaw; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2277.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,277 117.8 192.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mppx3zjx",
   "source_url": "https://omlx.ai/benchmarks/performance/mppx3zjx",
   "tg_tokens": null,
   "tps": 117.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-mppx3zjx-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 198.4 GB at this context; duplicate traces of the same session: a5v0qsaw; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2228.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 2,228 62.0 198.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mppx3zjx",
   "source_url": "https://omlx.ai/benchmarks/performance/mppx3zjx",
   "tg_tokens": null,
   "tps": 62.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-29lf7hea-ctx1024",
   "model": "Gemma-4-E4B-Claude-Opus-Reasoning",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 4.5 GB at this context; duplicate traces of the same session: nqylavz8; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 655.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Gemma-4-E4B-Claude-Opus-Reasoning-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 655.2 31.3 4.5 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/29lf7hea",
   "source_url": "https://omlx.ai/benchmarks/performance/29lf7hea",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-29lf7hea-ctx4096",
   "model": "Gemma-4-E4B-Claude-Opus-Reasoning",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 4.5 GB at this context; duplicate traces of the same session: nqylavz8; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 661.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Gemma-4-E4B-Claude-Opus-Reasoning-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 661.7 30.7 4.5 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/29lf7hea",
   "source_url": "https://omlx.ai/benchmarks/performance/29lf7hea",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ja2ir31o-ctx1024",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 3b9lueiw, jaf89ej7, tgs4ghib; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2408.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,408 100.0 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ja2ir31o",
   "source_url": "https://omlx.ai/benchmarks/performance/ja2ir31o",
   "tg_tokens": null,
   "tps": 100.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ja2ir31o-ctx8192",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 3b9lueiw, jaf89ej7, tgs4ghib; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3137.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 3,137 100.1 22.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ja2ir31o",
   "source_url": "https://omlx.ai/benchmarks/performance/ja2ir31o",
   "tg_tokens": null,
   "tps": 100.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ja2ir31o-ctx16384",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 3b9lueiw, jaf89ej7, tgs4ghib; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2876.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,876 86.3 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ja2ir31o",
   "source_url": "https://omlx.ai/benchmarks/performance/ja2ir31o",
   "tg_tokens": null,
   "tps": 86.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ja2ir31o-ctx32768",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 3b9lueiw, jaf89ej7, tgs4ghib; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2378.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 2,378 66.7 22.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ja2ir31o",
   "source_url": "https://omlx.ai/benchmarks/performance/ja2ir31o",
   "tg_tokens": null,
   "tps": 66.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ja2ir31o-ctx65536",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 3b9lueiw, jaf89ej7, tgs4ghib; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1736.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 1,736 65.4 23.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ja2ir31o",
   "source_url": "https://omlx.ai/benchmarks/performance/ja2ir31o",
   "tg_tokens": null,
   "tps": 65.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-re5or61z-ctx1024",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2367.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,367 107.6 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/re5or61z",
   "source_url": "https://omlx.ai/benchmarks/performance/re5or61z",
   "tg_tokens": null,
   "tps": 107.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-re5or61z-ctx4096",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3266.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,266 91.7 22.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/re5or61z",
   "source_url": "https://omlx.ai/benchmarks/performance/re5or61z",
   "tg_tokens": null,
   "tps": 91.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-re5or61z-ctx8192",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3176.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 3,176 96.6 22.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/re5or61z",
   "source_url": "https://omlx.ai/benchmarks/performance/re5or61z",
   "tg_tokens": null,
   "tps": 96.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-re5or61z-ctx16384",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2891.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,891 92.5 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/re5or61z",
   "source_url": "https://omlx.ai/benchmarks/performance/re5or61z",
   "tg_tokens": null,
   "tps": 92.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-re5or61z-ctx32768",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2459.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 2,459 77.3 22.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/re5or61z",
   "source_url": "https://omlx.ai/benchmarks/performance/re5or61z",
   "tg_tokens": null,
   "tps": 77.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-1lrb5p77-ctx1024",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: bk8ht2sj; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 109.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 109.8 18.4 20.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1lrb5p77",
   "source_url": "https://omlx.ai/benchmarks/performance/1lrb5p77",
   "tg_tokens": null,
   "tps": 18.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-1lrb5p77-ctx4096",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: bk8ht2sj; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 105.0 18.0 21.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1lrb5p77",
   "source_url": "https://omlx.ai/benchmarks/performance/1lrb5p77",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-9n2cru66-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: 720tv5pz; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 232.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 232.9 18.5 5.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9n2cru66",
   "source_url": "https://omlx.ai/benchmarks/performance/9n2cru66",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-9n2cru66-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: 720tv5pz; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 239.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 239.4 18.5 6.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9n2cru66",
   "source_url": "https://omlx.ai/benchmarks/performance/9n2cru66",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-i1jfcu2w-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: n49osmb1; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 587.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 587.4 69.5 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i1jfcu2w",
   "source_url": "https://omlx.ai/benchmarks/performance/i1jfcu2w",
   "tg_tokens": null,
   "tps": 69.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-i1jfcu2w-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: n49osmb1; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 603.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 603.8 67.2 21.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i1jfcu2w",
   "source_url": "https://omlx.ai/benchmarks/performance/i1jfcu2w",
   "tg_tokens": null,
   "tps": 67.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-v3babl47-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: ejoa7t2i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1595.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,595 176.2 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v3babl47",
   "source_url": "https://omlx.ai/benchmarks/performance/v3babl47",
   "tg_tokens": null,
   "tps": 176.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-v3babl47-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: ejoa7t2i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1997.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,997 150.2 23.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v3babl47",
   "source_url": "https://omlx.ai/benchmarks/performance/v3babl47",
   "tg_tokens": null,
   "tps": 150.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-m7rbrnls-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: girop37r, qw2nez96; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1597.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,597 175.9 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m7rbrnls",
   "source_url": "https://omlx.ai/benchmarks/performance/m7rbrnls",
   "tg_tokens": null,
   "tps": 175.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-m7rbrnls-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: girop37r, qw2nez96; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1999.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,999 150.0 23.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m7rbrnls",
   "source_url": "https://omlx.ai/benchmarks/performance/m7rbrnls",
   "tg_tokens": null,
   "tps": 150.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-m7rbrnls-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: girop37r, qw2nez96; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1870.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,870 178.3 23.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m7rbrnls",
   "source_url": "https://omlx.ai/benchmarks/performance/m7rbrnls",
   "tg_tokens": null,
   "tps": 178.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-icz1xarx-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 2rid3w5u; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 111.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 111.1 19.4 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/icz1xarx",
   "source_url": "https://omlx.ai/benchmarks/performance/icz1xarx",
   "tg_tokens": null,
   "tps": 19.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-icz1xarx-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 2rid3w5u; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 104.1 18.8 19.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/icz1xarx",
   "source_url": "https://omlx.ai/benchmarks/performance/icz1xarx",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-nu838orp-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: oma8rgny, r4n1e8mu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 128.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 128.7 14.4 16.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nu838orp",
   "source_url": "https://omlx.ai/benchmarks/performance/nu838orp",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-nu838orp-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: oma8rgny, r4n1e8mu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 128.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 128.8 13.6 17.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nu838orp",
   "source_url": "https://omlx.ai/benchmarks/performance/nu838orp",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-nu838orp-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: oma8rgny, r4n1e8mu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 121.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 121.0 13.6 18.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nu838orp",
   "source_url": "https://omlx.ai/benchmarks/performance/nu838orp",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-nu838orp-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: oma8rgny, r4n1e8mu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 116.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 116.2 12.8 18.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nu838orp",
   "source_url": "https://omlx.ai/benchmarks/performance/nu838orp",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-iyrxbzvm-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 19.1 GB; TTFT 3252 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1260.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 1,260 PP tok/s 43.8 TG tok/s 3252 TTFT (ms) 19.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iyrxbzvm",
   "source_url": "https://omlx.ai/benchmarks/performance/iyrxbzvm",
   "tg_tokens": null,
   "tps": 43.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zhd04eic-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 21.5 GB; TTFT 4602 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 890.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 890.0 PP tok/s 32.6 TG tok/s 4602 TTFT (ms) 21.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zhd04eic",
   "source_url": "https://omlx.ai/benchmarks/performance/zhd04eic",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-651jid5o-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 19.1 GB; TTFT 3029 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1352.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 1,352 PP tok/s 61.1 TG tok/s 3029 TTFT (ms) 19.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/651jid5o",
   "source_url": "https://omlx.ai/benchmarks/performance/651jid5o",
   "tg_tokens": null,
   "tps": 61.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yzrbx3hn-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 19.1 GB; TTFT 3062 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1338.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 1,338 PP tok/s 52.6 TG tok/s 3062 TTFT (ms) 19.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yzrbx3hn",
   "source_url": "https://omlx.ai/benchmarks/performance/yzrbx3hn",
   "tg_tokens": null,
   "tps": 52.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7jjbtwcf-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 19.1 GB; TTFT 3146 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1302.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 1,302 PP tok/s 34.4 TG tok/s 3146 TTFT (ms) 19.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7jjbtwcf",
   "source_url": "https://omlx.ai/benchmarks/performance/7jjbtwcf",
   "tg_tokens": null,
   "tps": 34.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ulaun11f-ctx16384",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 40.9 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 309.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 309.7 31.3 40.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ulaun11f",
   "source_url": "https://omlx.ai/benchmarks/performance/ulaun11f",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ulaun11f-ctx32768",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 41.3 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 332.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 32k 332.1 23.3 41.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ulaun11f",
   "source_url": "https://omlx.ai/benchmarks/performance/ulaun11f",
   "tg_tokens": null,
   "tps": 23.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ulaun11f-ctx65536",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 42.3 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 328.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 64k 328.7 26.3 42.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ulaun11f",
   "source_url": "https://omlx.ai/benchmarks/performance/ulaun11f",
   "tg_tokens": null,
   "tps": 26.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cokj9fvi-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 45.2 GB at this context; duplicate traces of the same session: 4gmjfhtb, 4wxihrxt, fkx41x43, kaxn6n01, lr295wzu, oponssv0, ovodi6ju; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 125.8 25.5 45.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cokj9fvi",
   "source_url": "https://omlx.ai/benchmarks/performance/cokj9fvi",
   "tg_tokens": null,
   "tps": 25.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cokj9fvi-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 46.0 GB at this context; duplicate traces of the same session: 4gmjfhtb, 4wxihrxt, fkx41x43, kaxn6n01, lr295wzu, oponssv0, ovodi6ju; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 556.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 556.7 22.5 46.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cokj9fvi",
   "source_url": "https://omlx.ai/benchmarks/performance/cokj9fvi",
   "tg_tokens": null,
   "tps": 22.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cokj9fvi-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 47.1 GB at this context; duplicate traces of the same session: 4gmjfhtb, 4wxihrxt, fkx41x43, kaxn6n01, lr295wzu, oponssv0, ovodi6ju; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 873.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 873.1 13.0 47.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cokj9fvi",
   "source_url": "https://omlx.ai/benchmarks/performance/cokj9fvi",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cokj9fvi-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 48.1 GB at this context; duplicate traces of the same session: 4gmjfhtb, 4wxihrxt, fkx41x43, kaxn6n01, lr295wzu, oponssv0, ovodi6ju; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 909.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 909.0 17.7 48.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cokj9fvi",
   "source_url": "https://omlx.ai/benchmarks/performance/cokj9fvi",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cokj9fvi-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 48.6 GB at this context; duplicate traces of the same session: 4gmjfhtb, 4wxihrxt, fkx41x43, kaxn6n01, lr295wzu, oponssv0, ovodi6ju; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 937.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 937.5 21.1 48.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cokj9fvi",
   "source_url": "https://omlx.ai/benchmarks/performance/cokj9fvi",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cokj9fvi-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 49.8 GB at this context; duplicate traces of the same session: 4gmjfhtb, 4wxihrxt, fkx41x43, kaxn6n01, lr295wzu, oponssv0, ovodi6ju; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 943.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 943.9 23.1 49.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cokj9fvi",
   "source_url": "https://omlx.ai/benchmarks/performance/cokj9fvi",
   "tg_tokens": null,
   "tps": 23.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cokj9fvi-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 54.6 GB at this context; duplicate traces of the same session: 4gmjfhtb, 4wxihrxt, fkx41x43, kaxn6n01, lr295wzu, oponssv0, ovodi6ju; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 954.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 954.1 19.3 54.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cokj9fvi",
   "source_url": "https://omlx.ai/benchmarks/performance/cokj9fvi",
   "tg_tokens": null,
   "tps": 19.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-14",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cokj9fvi-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 56.2 GB at this context; duplicate traces of the same session: 4gmjfhtb, 4wxihrxt, fkx41x43, kaxn6n01, lr295wzu, oponssv0, ovodi6ju; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 944.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 944.0 19.7 56.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cokj9fvi",
   "source_url": "https://omlx.ai/benchmarks/performance/cokj9fvi",
   "tg_tokens": null,
   "tps": 19.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5molrqga-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 2ckjl4p6, 3s7mrovc, cqbmtpjn, djcjsg0f, g4k2x2hj, h4vuyeyz, i17qduem; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 902.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 902.4 65.9 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5molrqga",
   "source_url": "https://omlx.ai/benchmarks/performance/5molrqga",
   "tg_tokens": null,
   "tps": 65.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5molrqga-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: 2ckjl4p6, 3s7mrovc, cqbmtpjn, djcjsg0f, g4k2x2hj, h4vuyeyz, i17qduem; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1840.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,840 78.2 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5molrqga",
   "source_url": "https://omlx.ai/benchmarks/performance/5molrqga",
   "tg_tokens": null,
   "tps": 78.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5molrqga-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 2ckjl4p6, 3s7mrovc, cqbmtpjn, djcjsg0f, g4k2x2hj, h4vuyeyz, i17qduem; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2232.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,232 53.4 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5molrqga",
   "source_url": "https://omlx.ai/benchmarks/performance/5molrqga",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5molrqga-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: 2ckjl4p6, 3s7mrovc, cqbmtpjn, djcjsg0f, g4k2x2hj, h4vuyeyz, i17qduem; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2521.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,521 59.6 72.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5molrqga",
   "source_url": "https://omlx.ai/benchmarks/performance/5molrqga",
   "tg_tokens": null,
   "tps": 59.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5molrqga-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: 2ckjl4p6, 3s7mrovc, cqbmtpjn, djcjsg0f, g4k2x2hj, h4vuyeyz, i17qduem; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2539.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,539 68.6 73.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5molrqga",
   "source_url": "https://omlx.ai/benchmarks/performance/5molrqga",
   "tg_tokens": null,
   "tps": 68.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5molrqga-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 75.0 GB at this context; duplicate traces of the same session: 2ckjl4p6, 3s7mrovc, cqbmtpjn, djcjsg0f, g4k2x2hj, h4vuyeyz, i17qduem; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2457.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,457 64.2 75.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5molrqga",
   "source_url": "https://omlx.ai/benchmarks/performance/5molrqga",
   "tg_tokens": null,
   "tps": 64.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5molrqga-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.9 GB at this context; duplicate traces of the same session: 2ckjl4p6, 3s7mrovc, cqbmtpjn, djcjsg0f, g4k2x2hj, h4vuyeyz, i17qduem; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2358.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,358 45.0 79.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5molrqga",
   "source_url": "https://omlx.ai/benchmarks/performance/5molrqga",
   "tg_tokens": null,
   "tps": 45.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5molrqga-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.9 GB at this context; duplicate traces of the same session: 2ckjl4p6, 3s7mrovc, cqbmtpjn, djcjsg0f, g4k2x2hj, h4vuyeyz, i17qduem; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2273.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,273 47.6 80.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5molrqga",
   "source_url": "https://omlx.ai/benchmarks/performance/5molrqga",
   "tg_tokens": null,
   "tps": 47.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-gh82x1f1-ctx1024",
   "model": "gemma-4-12b-it",
   "notes": "OptiQ; oMLX community benchmark trace; peak memory 9.4 GB at this context; duplicate traces of the same session: zqiywnon; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 135.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-OptiQ-4bit on M2 Pro (16c); M2 Pro (16c) \u00b7 16 GB \u00b7 4bit; 1k 135.8 18.0 9.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gh82x1f1",
   "source_url": "https://omlx.ai/benchmarks/performance/gh82x1f1",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-gh82x1f1-ctx4096",
   "model": "gemma-4-12b-it",
   "notes": "OptiQ; oMLX community benchmark trace; peak memory 9.7 GB at this context; duplicate traces of the same session: zqiywnon; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 133.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-OptiQ-4bit on M2 Pro (16c); M2 Pro (16c) \u00b7 16 GB \u00b7 4bit; 4k 133.5 17.7 9.7 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gh82x1f1",
   "source_url": "https://omlx.ai/benchmarks/performance/gh82x1f1",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-r3i82r5j-ctx4096",
   "model": "gemma-4-26B-A4B-it-qat-uncensored-heretic",
   "notes": "MLX VLM build; oMLX community benchmark trace, single run at 4k context; peak memory 14.4 GB; TTFT 1006 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4071.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-uncensored-heretic-mlx-vlm-4Bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 4,071 PP tok/s 130.2 TG tok/s 1006 TTFT (ms) 14.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r3i82r5j",
   "source_url": "https://omlx.ai/benchmarks/performance/r3i82r5j",
   "tg_tokens": null,
   "tps": 130.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-96phhyv9-ctx4096",
   "model": "gemma-4-26B-A4B-it-qat-uncensored-heretic",
   "notes": "MLX VLM build; oMLX community benchmark trace, single run at 4k context; peak memory 14.4 GB; TTFT 1556 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2632.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-uncensored-heretic-mlx-vlm-4Bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 2,632 PP tok/s 130.2 TG tok/s 1556 TTFT (ms) 14.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/96phhyv9",
   "source_url": "https://omlx.ai/benchmarks/performance/96phhyv9",
   "tg_tokens": null,
   "tps": 130.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-pp8h2peg-ctx1024",
   "model": "gemma-4-E4B-it",
   "notes": "oMLX community benchmark trace; peak memory 7.9 GB at this context; duplicate traces of the same session: 0wkks5oe; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 631.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-E4B-it-8bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 8bit; 1k 631.8 17.4 7.9 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pp8h2peg",
   "source_url": "https://omlx.ai/benchmarks/performance/pp8h2peg",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-pp8h2peg-ctx4096",
   "model": "gemma-4-E4B-it",
   "notes": "oMLX community benchmark trace; peak memory 7.9 GB at this context; duplicate traces of the same session: 0wkks5oe; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 655.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-E4B-it-8bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 8bit; 4k 655.2 17.0 7.9 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pp8h2peg",
   "source_url": "https://omlx.ai/benchmarks/performance/pp8h2peg",
   "tg_tokens": null,
   "tps": 17.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-02rtirc3-ctx1024",
   "model": "gemma-4-E4B-it-ultra-uncensored-heretic",
   "notes": "oMLX community benchmark trace; peak memory 5.4 GB at this context; duplicate traces of the same session: m9803e2m; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 646.5,
   "provenance": "community",
   "quant": "mixed 4+6bit",
   "quote": "gemma-4-E4B-it-ultra-uncensored-heretic-MLX-4bit-mixed_4_6 on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 646.5 28.2 5.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/02rtirc3",
   "source_url": "https://omlx.ai/benchmarks/performance/02rtirc3",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-02rtirc3-ctx4096",
   "model": "gemma-4-E4B-it-ultra-uncensored-heretic",
   "notes": "oMLX community benchmark trace; peak memory 5.5 GB at this context; duplicate traces of the same session: m9803e2m; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 658.4,
   "provenance": "community",
   "quant": "mixed 4+6bit",
   "quote": "gemma-4-E4B-it-ultra-uncensored-heretic-MLX-4bit-mixed_4_6 on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 658.4 29.0 5.5 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/02rtirc3",
   "source_url": "https://omlx.ai/benchmarks/performance/02rtirc3",
   "tg_tokens": null,
   "tps": 29.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-rfaftm50-ctx1024",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 5.1 GB at this context; duplicate traces of the same session: rly3db6g; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2222.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-e2b-8bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 8bit; 1k 2,222 34.5 5.1 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rfaftm50",
   "source_url": "https://omlx.ai/benchmarks/performance/rfaftm50",
   "tg_tokens": null,
   "tps": 34.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-rfaftm50-ctx4096",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 5.2 GB at this context; duplicate traces of the same session: rly3db6g; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2291.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-e2b-8bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 8bit; 4k 2,291 34.2 5.2 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rfaftm50",
   "source_url": "https://omlx.ai/benchmarks/performance/rfaftm50",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-iur2ylqb-ctx1024",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 5.1 GB at this context; duplicate traces of the same session: mtl7g006; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2222.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-e2b-8bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 8bit; 1k 2,222 34.9 5.1 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iur2ylqb",
   "source_url": "https://omlx.ai/benchmarks/performance/iur2ylqb",
   "tg_tokens": null,
   "tps": 34.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-iur2ylqb-ctx4096",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 5.2 GB at this context; duplicate traces of the same session: mtl7g006; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2259.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-e2b-8bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 8bit; 4k 2,259 34.6 5.2 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iur2ylqb",
   "source_url": "https://omlx.ai/benchmarks/performance/iur2ylqb",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-185pz6do-ctx1024",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2408.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,408 100.0 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/185pz6do",
   "source_url": "https://omlx.ai/benchmarks/performance/185pz6do",
   "tg_tokens": null,
   "tps": 100.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-185pz6do-ctx8192",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3137.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 3,137 100.1 22.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/185pz6do",
   "source_url": "https://omlx.ai/benchmarks/performance/185pz6do",
   "tg_tokens": null,
   "tps": 100.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-185pz6do-ctx16384",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2876.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,876 86.3 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/185pz6do",
   "source_url": "https://omlx.ai/benchmarks/performance/185pz6do",
   "tg_tokens": null,
   "tps": 86.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-185pz6do-ctx32768",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2378.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 2,378 66.7 22.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/185pz6do",
   "source_url": "https://omlx.ai/benchmarks/performance/185pz6do",
   "tg_tokens": null,
   "tps": 66.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-185pz6do-ctx65536",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1736.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 1,736 65.4 23.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/185pz6do",
   "source_url": "https://omlx.ai/benchmarks/performance/185pz6do",
   "tg_tokens": null,
   "tps": 65.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-0gl2yp8s-ctx1024",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: lyiebx2x, qd544t6e, u1ma3hk9; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2367.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,367 107.6 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0gl2yp8s",
   "source_url": "https://omlx.ai/benchmarks/performance/0gl2yp8s",
   "tg_tokens": null,
   "tps": 107.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0gl2yp8s-ctx4096",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: lyiebx2x, qd544t6e, u1ma3hk9; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3266.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,266 91.7 22.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0gl2yp8s",
   "source_url": "https://omlx.ai/benchmarks/performance/0gl2yp8s",
   "tg_tokens": null,
   "tps": 91.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0gl2yp8s-ctx8192",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: lyiebx2x, qd544t6e, u1ma3hk9; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3176.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 3,176 96.6 22.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0gl2yp8s",
   "source_url": "https://omlx.ai/benchmarks/performance/0gl2yp8s",
   "tg_tokens": null,
   "tps": 96.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0gl2yp8s-ctx16384",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: lyiebx2x, qd544t6e, u1ma3hk9; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2891.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,891 92.5 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0gl2yp8s",
   "source_url": "https://omlx.ai/benchmarks/performance/0gl2yp8s",
   "tg_tokens": null,
   "tps": 92.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0gl2yp8s-ctx32768",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: lyiebx2x, qd544t6e, u1ma3hk9; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2459.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 2,459 77.3 22.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0gl2yp8s",
   "source_url": "https://omlx.ai/benchmarks/performance/0gl2yp8s",
   "tg_tokens": null,
   "tps": 77.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-fauz2klp-ctx1024",
   "model": "KAT-Dev",
   "notes": "oMLX community benchmark trace; peak memory 33.3 GB at this context; duplicate traces of the same session: 86l41bnv, kod8gxf3; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 86.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "KAT-Dev-MLX-8bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 1k 86.8 8.0 33.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fauz2klp",
   "source_url": "https://omlx.ai/benchmarks/performance/fauz2klp",
   "tg_tokens": null,
   "tps": 8.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-fauz2klp-ctx4096",
   "model": "KAT-Dev",
   "notes": "oMLX community benchmark trace; peak memory 34.2 GB at this context; duplicate traces of the same session: 86l41bnv, kod8gxf3; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 77.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "KAT-Dev-MLX-8bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 4k 77.2 7.2 34.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fauz2klp",
   "source_url": "https://omlx.ai/benchmarks/performance/fauz2klp",
   "tg_tokens": null,
   "tps": 7.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-fauz2klp-ctx8192",
   "model": "KAT-Dev",
   "notes": "oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: 86l41bnv, kod8gxf3; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 70.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "KAT-Dev-MLX-8bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 8k 70.5 6.3 35.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fauz2klp",
   "source_url": "https://omlx.ai/benchmarks/performance/fauz2klp",
   "tg_tokens": null,
   "tps": 6.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-rmzld28q-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX DWQ v2 build; oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: mqox4fde; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1018.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-DWQ-4bit-v2 on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 1,018 106.4 13.4 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rmzld28q",
   "source_url": "https://omlx.ai/benchmarks/performance/rmzld28q",
   "tg_tokens": null,
   "tps": 106.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-rmzld28q-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX DWQ v2 build; oMLX community benchmark trace; peak memory 13.9 GB at this context; duplicate traces of the same session: mqox4fde; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 958.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-DWQ-4bit-v2 on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 16k 958.7 86.4 13.9 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rmzld28q",
   "source_url": "https://omlx.ai/benchmarks/performance/rmzld28q",
   "tg_tokens": null,
   "tps": 86.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-da9kikzc-ctx131072",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace, single run at 128k context; peak memory 16.3 GB; TTFT 208907 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 627.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e-fp16 on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 128k tokens 627.4 PP tok/s 36.6 TG tok/s 208907 TTFT (ms) 16.3 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/da9kikzc",
   "source_url": "https://omlx.ai/benchmarks/performance/da9kikzc",
   "tg_tokens": null,
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-z6qpfsd8-ctx1024",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: 84413eoh, b56pbh5z; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 875.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-mlx-mxfp4 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 875.0 89.7 6.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z6qpfsd8",
   "source_url": "https://omlx.ai/benchmarks/performance/z6qpfsd8",
   "tg_tokens": null,
   "tps": 89.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-z6qpfsd8-ctx4096",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: 84413eoh, b56pbh5z; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 835.3,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-mlx-mxfp4 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 835.3 86.7 6.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z6qpfsd8",
   "source_url": "https://omlx.ai/benchmarks/performance/z6qpfsd8",
   "tg_tokens": null,
   "tps": 86.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-z6qpfsd8-ctx32768",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 7.2 GB at this context; duplicate traces of the same session: 84413eoh, b56pbh5z; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 655.9,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-mlx-mxfp4 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 32k 655.9 69.6 7.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z6qpfsd8",
   "source_url": "https://omlx.ai/benchmarks/performance/z6qpfsd8",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-kizo21up-ctx1024",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 12.8 GB at this context; duplicate traces of the same session: 8ymc9j1d, r4o0a9zp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 436.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 436.1 48.3 12.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kizo21up",
   "source_url": "https://omlx.ai/benchmarks/performance/kizo21up",
   "tg_tokens": null,
   "tps": 48.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-kizo21up-ctx4096",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: 8ymc9j1d, r4o0a9zp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 445.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 445.7 45.5 13.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kizo21up",
   "source_url": "https://omlx.ai/benchmarks/performance/kizo21up",
   "tg_tokens": null,
   "tps": 45.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-kizo21up-ctx8192",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.6 GB at this context; duplicate traces of the same session: 8ymc9j1d, r4o0a9zp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 408.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 8k 408.4 42.3 13.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kizo21up",
   "source_url": "https://omlx.ai/benchmarks/performance/kizo21up",
   "tg_tokens": null,
   "tps": 42.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-ai79pw9m-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 24.2 GB; TTFT 19745 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 829.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k tokens 829.8 PP tok/s 62.6 TG tok/s 19745 TTFT (ms) 24.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ai79pw9m",
   "source_url": "https://omlx.ai/benchmarks/performance/ai79pw9m",
   "tg_tokens": null,
   "tps": 62.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-o42jmc9h-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 24.3 GB; TTFT 19519 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 839.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k tokens 839.4 PP tok/s 76.0 TG tok/s 19519 TTFT (ms) 24.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o42jmc9h",
   "source_url": "https://omlx.ai/benchmarks/performance/o42jmc9h",
   "tg_tokens": null,
   "tps": 76.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-8dya0fjd-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 24.2 GB; TTFT 19730 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 830.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k tokens 830.4 PP tok/s 61.7 TG tok/s 19730 TTFT (ms) 24.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8dya0fjd",
   "source_url": "https://omlx.ai/benchmarks/performance/8dya0fjd",
   "tg_tokens": null,
   "tps": 61.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-x33f5msy-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 24.2 GB; TTFT 19576 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 837.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k tokens 837.0 PP tok/s 74.8 TG tok/s 19576 TTFT (ms) 24.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x33f5msy",
   "source_url": "https://omlx.ai/benchmarks/performance/x33f5msy",
   "tg_tokens": null,
   "tps": 74.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M1 Max",
   "id": "omlx-3xpw0yao-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 10.2 GB; TTFT 18471 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 443.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k tokens 443.5 PP tok/s 14.7 TG tok/s 18471 TTFT (ms) 10.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3xpw0yao",
   "source_url": "https://omlx.ai/benchmarks/performance/3xpw0yao",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ohfz115v-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 23.4 GB; TTFT 7901 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1037.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k tokens 1,037 PP tok/s 74.5 TG tok/s 7901 TTFT (ms) 23.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ohfz115v",
   "source_url": "https://omlx.ai/benchmarks/performance/ohfz115v",
   "tg_tokens": null,
   "tps": 74.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-e87xbfwn-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: dpqoysjr, nqixnl7q, nr79h7r8; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 424.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 424.7 17.9 17.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e87xbfwn",
   "source_url": "https://omlx.ai/benchmarks/performance/e87xbfwn",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-e87xbfwn-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: dpqoysjr, nqixnl7q, nr79h7r8; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 398.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 398.2 16.7 17.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e87xbfwn",
   "source_url": "https://omlx.ai/benchmarks/performance/e87xbfwn",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-e87xbfwn-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: dpqoysjr, nqixnl7q, nr79h7r8; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 326.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 64k 326.7 13.5 20.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e87xbfwn",
   "source_url": "https://omlx.ai/benchmarks/performance/e87xbfwn",
   "tg_tokens": null,
   "tps": 13.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-e87xbfwn-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: dpqoysjr, nqixnl7q, nr79h7r8; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 260.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 128k 260.4 10.8 24.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e87xbfwn",
   "source_url": "https://omlx.ai/benchmarks/performance/e87xbfwn",
   "tg_tokens": null,
   "tps": 10.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ya8djcv0-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 195k context; peak memory 29.1 GB; TTFT 869269 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 230.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 195k tokens 230.1 PP tok/s 8.9 TG tok/s 869269 TTFT (ms) 29.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ya8djcv0",
   "source_url": "https://omlx.ai/benchmarks/performance/ya8djcv0",
   "tg_tokens": null,
   "tps": 8.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-vqohhgnn-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: 4qxlxdjv, dm45vg4t; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 419.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 419.0 17.9 17.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vqohhgnn",
   "source_url": "https://omlx.ai/benchmarks/performance/vqohhgnn",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-vqohhgnn-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 4qxlxdjv, dm45vg4t; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 396.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 396.3 16.8 17.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vqohhgnn",
   "source_url": "https://omlx.ai/benchmarks/performance/vqohhgnn",
   "tg_tokens": null,
   "tps": 16.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-vqohhgnn-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: 4qxlxdjv, dm45vg4t; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 262.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 128k 262.7 10.5 24.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vqohhgnn",
   "source_url": "https://omlx.ai/benchmarks/performance/vqohhgnn",
   "tg_tokens": null,
   "tps": 10.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-g2ocrgxd-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: zasd8zgb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 464.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 464.2 18.1 16.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g2ocrgxd",
   "source_url": "https://omlx.ai/benchmarks/performance/g2ocrgxd",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-g2ocrgxd-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: zasd8zgb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 500.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 500.1 17.7 17.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g2ocrgxd",
   "source_url": "https://omlx.ai/benchmarks/performance/g2ocrgxd",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-vcnr388z-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: hy4421es; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 252.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 252.2 49.8 23.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vcnr388z",
   "source_url": "https://omlx.ai/benchmarks/performance/vcnr388z",
   "tg_tokens": null,
   "tps": 49.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-vcnr388z-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: hy4421es; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 243.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 243.4 47.1 27.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vcnr388z",
   "source_url": "https://omlx.ai/benchmarks/performance/vcnr388z",
   "tg_tokens": null,
   "tps": 47.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-35hndobc-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: yq4d9myk; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 212.4,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-27B-nvfp4 on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 212.4 5.4 17.0 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/35hndobc",
   "source_url": "https://omlx.ai/benchmarks/performance/35hndobc",
   "tg_tokens": null,
   "tps": 5.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-35hndobc-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: yq4d9myk; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 213.7,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-27B-nvfp4 on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 213.7 5.4 18.1 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/35hndobc",
   "source_url": "https://omlx.ai/benchmarks/performance/35hndobc",
   "tg_tokens": null,
   "tps": 5.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-u91mar7p-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 128.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 128.7 14.4 16.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u91mar7p",
   "source_url": "https://omlx.ai/benchmarks/performance/u91mar7p",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-u91mar7p-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 128.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 128.8 13.6 17.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u91mar7p",
   "source_url": "https://omlx.ai/benchmarks/performance/u91mar7p",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-u91mar7p-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 121.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 121.0 13.6 18.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u91mar7p",
   "source_url": "https://omlx.ai/benchmarks/performance/u91mar7p",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-u91mar7p-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 116.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 116.2 12.8 18.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u91mar7p",
   "source_url": "https://omlx.ai/benchmarks/performance/u91mar7p",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-tr62slaz-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "plefp8 build; oMLX community benchmark trace; peak memory 74.7 GB at this context; duplicate traces of the same session: a9ff0xhy, sjxi95i6, vsz7gbgo, xvfszazy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1258.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-Uncensored-plefp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,258 70.5 74.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tr62slaz",
   "source_url": "https://omlx.ai/benchmarks/performance/tr62slaz",
   "tg_tokens": null,
   "tps": 70.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-tr62slaz-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "plefp8 build; oMLX community benchmark trace; peak memory 75.1 GB at this context; duplicate traces of the same session: a9ff0xhy, sjxi95i6, vsz7gbgo, xvfszazy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1648.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-Uncensored-plefp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,648 69.5 75.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tr62slaz",
   "source_url": "https://omlx.ai/benchmarks/performance/tr62slaz",
   "tg_tokens": null,
   "tps": 69.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-tr62slaz-ctx8192",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "plefp8 build; oMLX community benchmark trace; peak memory 75.9 GB at this context; duplicate traces of the same session: a9ff0xhy, sjxi95i6, vsz7gbgo, xvfszazy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1807.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-Uncensored-plefp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,807 59.8 75.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tr62slaz",
   "source_url": "https://omlx.ai/benchmarks/performance/tr62slaz",
   "tg_tokens": null,
   "tps": 59.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-tr62slaz-ctx16384",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "plefp8 build; oMLX community benchmark trace; peak memory 76.9 GB at this context; duplicate traces of the same session: a9ff0xhy, sjxi95i6, vsz7gbgo, xvfszazy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1970.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-Uncensored-plefp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,970 60.5 76.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tr62slaz",
   "source_url": "https://omlx.ai/benchmarks/performance/tr62slaz",
   "tg_tokens": null,
   "tps": 60.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tr62slaz-ctx65536",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "plefp8 build; oMLX community benchmark trace; peak memory 78.6 GB at this context; duplicate traces of the same session: a9ff0xhy, sjxi95i6, vsz7gbgo, xvfszazy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1988.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-Uncensored-plefp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 1,988 64.8 78.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tr62slaz",
   "source_url": "https://omlx.ai/benchmarks/performance/tr62slaz",
   "tg_tokens": null,
   "tps": 64.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-tr62slaz-ctx131072",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "plefp8 build; oMLX community benchmark trace; peak memory 80.9 GB at this context; duplicate traces of the same session: a9ff0xhy, sjxi95i6, vsz7gbgo, xvfszazy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1906.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-Uncensored-plefp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 1,906 68.3 80.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tr62slaz",
   "source_url": "https://omlx.ai/benchmarks/performance/tr62slaz",
   "tg_tokens": null,
   "tps": 68.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-g6vx0ib6-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace, single run at 16k context; peak memory 75.3 GB; TTFT 31733 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 516.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k tokens 516.3 PP tok/s 33.4 TG tok/s 31733 TTFT (ms) 75.3 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g6vx0ib6",
   "source_url": "https://omlx.ai/benchmarks/performance/g6vx0ib6",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-oxbjdplw-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace, single run at 8k context; peak memory 75.1 GB; TTFT 16253 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 504.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k tokens 504.0 PP tok/s 31.7 TG tok/s 16253 TTFT (ms) 75.1 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oxbjdplw",
   "source_url": "https://omlx.ai/benchmarks/performance/oxbjdplw",
   "tg_tokens": null,
   "tps": 31.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-gfirxfds-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace, single run at 4k context; peak memory 74.8 GB; TTFT 12682 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k tokens 323.0 PP tok/s 27.4 TG tok/s 12682 TTFT (ms) 74.8 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gfirxfds",
   "source_url": "https://omlx.ai/benchmarks/performance/gfirxfds",
   "tg_tokens": null,
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-8mjblvt3-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 71.9 GB; TTFT 49010 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 334.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k tokens 334.3 PP tok/s 37.1 TG tok/s 49010 TTFT (ms) 71.9 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8mjblvt3",
   "source_url": "https://omlx.ai/benchmarks/performance/8mjblvt3",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9twpzzlu-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 101.0 GB; TTFT 1475 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2778.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 2,778 PP tok/s 94.2 TG tok/s 1475 TTFT (ms) 101.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9twpzzlu",
   "source_url": "https://omlx.ai/benchmarks/performance/9twpzzlu",
   "tg_tokens": null,
   "tps": 94.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ur6mqdnx-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 100.8 GB; TTFT 540 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1897.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,897 PP tok/s 91.7 TG tok/s 540 TTFT (ms) 100.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ur6mqdnx",
   "source_url": "https://omlx.ai/benchmarks/performance/ur6mqdnx",
   "tg_tokens": null,
   "tps": 91.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vhk53wg6-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 99.0 GB; TTFT 520 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1971.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,971 PP tok/s 82.0 TG tok/s 520 TTFT (ms) 99.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vhk53wg6",
   "source_url": "https://omlx.ai/benchmarks/performance/vhk53wg6",
   "tg_tokens": null,
   "tps": 82.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cuntp70y-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 101.0 GB; TTFT 1855 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2208.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 2,208 PP tok/s 53.8 TG tok/s 1855 TTFT (ms) 101.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cuntp70y",
   "source_url": "https://omlx.ai/benchmarks/performance/cuntp70y",
   "tg_tokens": null,
   "tps": 53.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dfm8cirf-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: xtnsnenp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1217.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,217 75.0 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dfm8cirf",
   "source_url": "https://omlx.ai/benchmarks/performance/dfm8cirf",
   "tg_tokens": null,
   "tps": 75.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dfm8cirf-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: xtnsnenp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1931.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,931 67.6 70.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dfm8cirf",
   "source_url": "https://omlx.ai/benchmarks/performance/dfm8cirf",
   "tg_tokens": null,
   "tps": 67.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-4b7p3tpd-ctx1024",
   "model": "gemma-4-26B-A4B-heretic-v2-qat-aligned",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: jhwsxgjn, qk83m2ty; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 478.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-heretic-v2-qat-aligned-4bit-mlx on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 478.2 31.5 12.7 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4b7p3tpd",
   "source_url": "https://omlx.ai/benchmarks/performance/4b7p3tpd",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-4b7p3tpd-ctx4096",
   "model": "gemma-4-26B-A4B-heretic-v2-qat-aligned",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 13.0 GB at this context; duplicate traces of the same session: jhwsxgjn, qk83m2ty; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 971.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-heretic-v2-qat-aligned-4bit-mlx on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 971.9 30.8 13.0 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4b7p3tpd",
   "source_url": "https://omlx.ai/benchmarks/performance/4b7p3tpd",
   "tg_tokens": null,
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-4b7p3tpd-ctx32768",
   "model": "gemma-4-26B-A4B-heretic-v2-qat-aligned",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 14.5 GB at this context; duplicate traces of the same session: jhwsxgjn, qk83m2ty; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 789.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-heretic-v2-qat-aligned-4bit-mlx on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 32k 789.2 23.6 14.5 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4b7p3tpd",
   "source_url": "https://omlx.ai/benchmarks/performance/4b7p3tpd",
   "tg_tokens": null,
   "tps": 23.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-e87qtcbs-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 49.6 GB at this context; duplicate traces of the same session: 7pno1r42; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 768.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "gemma-4-26b-a4b-it-bf16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 bf16; 1k 768.0 29.7 49.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e87qtcbs",
   "source_url": "https://omlx.ai/benchmarks/performance/e87qtcbs",
   "tg_tokens": null,
   "tps": 29.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-e87qtcbs-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 50.0 GB at this context; duplicate traces of the same session: 7pno1r42; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 811.2,
   "provenance": "community",
   "quant": "bf16",
   "quote": "gemma-4-26b-a4b-it-bf16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 bf16; 4k 811.2 18.3 50.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e87qtcbs",
   "source_url": "https://omlx.ai/benchmarks/performance/e87qtcbs",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-4p9hmx2f-ctx1024",
   "model": "gemma-4-E4B-Agentic-Sol-Fable-Reasoning-GeminiCLI",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: d1ps9lw1; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 803.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E4B-Agentic-Sol-Fable-Reasoning-GeminiCLI-mlx-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 803.2 18.0 17.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4p9hmx2f",
   "source_url": "https://omlx.ai/benchmarks/performance/4p9hmx2f",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-4p9hmx2f-ctx4096",
   "model": "gemma-4-E4B-Agentic-Sol-Fable-Reasoning-GeminiCLI",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 17.5 GB at this context; duplicate traces of the same session: d1ps9lw1; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 995.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E4B-Agentic-Sol-Fable-Reasoning-GeminiCLI-mlx-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 995.8 17.4 17.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4p9hmx2f",
   "source_url": "https://omlx.ai/benchmarks/performance/4p9hmx2f",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-7u7f8w1f-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.9 GB at this context; duplicate traces of the same session: nnk2w401; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1148.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,148 107.7 13.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7u7f8w1f",
   "source_url": "https://omlx.ai/benchmarks/performance/7u7f8w1f",
   "tg_tokens": null,
   "tps": 107.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-7u7f8w1f-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 14.2 GB at this context; duplicate traces of the same session: nnk2w401; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 719.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 719.1 75.7 14.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7u7f8w1f",
   "source_url": "https://omlx.ai/benchmarks/performance/7u7f8w1f",
   "tg_tokens": null,
   "tps": 75.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ujnkbs3n-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.6 GB at this context; duplicate traces of the same session: 99i1jn40, dxf216k8, itizrwp1, zqgpsglo; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 936.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-MLX-oQ4e on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 936.7 101.6 13.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ujnkbs3n",
   "source_url": "https://omlx.ai/benchmarks/performance/ujnkbs3n",
   "tg_tokens": null,
   "tps": 101.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ujnkbs3n-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.9 GB at this context; duplicate traces of the same session: 99i1jn40, dxf216k8, itizrwp1, zqgpsglo; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1039.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-MLX-oQ4e on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 1,039 97.3 13.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ujnkbs3n",
   "source_url": "https://omlx.ai/benchmarks/performance/ujnkbs3n",
   "tg_tokens": null,
   "tps": 97.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ujnkbs3n-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.9 GB at this context; duplicate traces of the same session: 99i1jn40, dxf216k8, itizrwp1, zqgpsglo; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1011.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-MLX-oQ4e on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 1,011 91.5 13.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ujnkbs3n",
   "source_url": "https://omlx.ai/benchmarks/performance/ujnkbs3n",
   "tg_tokens": null,
   "tps": 91.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ujnkbs3n-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 14.1 GB at this context; duplicate traces of the same session: 99i1jn40, dxf216k8, itizrwp1, zqgpsglo; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 952.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-MLX-oQ4e on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 952.7 81.5 14.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ujnkbs3n",
   "source_url": "https://omlx.ai/benchmarks/performance/ujnkbs3n",
   "tg_tokens": null,
   "tps": 81.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ujnkbs3n-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: 99i1jn40, dxf216k8, itizrwp1, zqgpsglo; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 849.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-MLX-oQ4e on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 849.4 66.1 14.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ujnkbs3n",
   "source_url": "https://omlx.ai/benchmarks/performance/ujnkbs3n",
   "tg_tokens": null,
   "tps": 66.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-icq62exq-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 65.8 GB at this context; duplicate traces of the same session: 2zomz31r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1666.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,666 66.4 65.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/icq62exq",
   "source_url": "https://omlx.ai/benchmarks/performance/icq62exq",
   "tg_tokens": null,
   "tps": 66.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-icq62exq-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.7 GB at this context; duplicate traces of the same session: 2zomz31r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2266.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,266 65.0 66.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/icq62exq",
   "source_url": "https://omlx.ai/benchmarks/performance/icq62exq",
   "tg_tokens": null,
   "tps": 65.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-paga8y79-ctx1024",
   "model": "Qwen3.5-9B-Uncensored-HauhauCS-Aggressive",
   "notes": "MLX build; oMLX community benchmark trace, single run at 1k context; peak memory 9.6 GB; TTFT 6777 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 151.1,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Qwen3.5-9B-Uncensored-HauhauCS-Aggressive-MLX-mxfp8 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 8bit; 1k tokens 151.1 PP tok/s 10.8 TG tok/s 6777 TTFT (ms) 9.6 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/paga8y79",
   "source_url": "https://omlx.ai/benchmarks/performance/paga8y79",
   "tg_tokens": null,
   "tps": 10.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-do40dy2t-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: indhv0ro, y6sjp1ai; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 769.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 769.7 74.5 20.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/do40dy2t",
   "source_url": "https://omlx.ai/benchmarks/performance/do40dy2t",
   "tg_tokens": null,
   "tps": 74.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-do40dy2t-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: indhv0ro, y6sjp1ai; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 644.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 32k 644.8 60.1 25.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/do40dy2t",
   "source_url": "https://omlx.ai/benchmarks/performance/do40dy2t",
   "tg_tokens": null,
   "tps": 60.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-do40dy2t-ctx199680",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.8 GB at this context; duplicate traces of the same session: indhv0ro, y6sjp1ai; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 240.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 195k 240.7 20.7 28.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/do40dy2t",
   "source_url": "https://omlx.ai/benchmarks/performance/do40dy2t",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xwd8811p-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 0mgvw0f9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4634.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,634 196.0 21.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xwd8811p",
   "source_url": "https://omlx.ai/benchmarks/performance/xwd8811p",
   "tg_tokens": null,
   "tps": 196.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xwd8811p-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 0mgvw0f9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6076.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,076 188.7 22.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xwd8811p",
   "source_url": "https://omlx.ai/benchmarks/performance/xwd8811p",
   "tg_tokens": null,
   "tps": 188.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qda3up38-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: 9seh0xdo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4228.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 4,228 182.7 37.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qda3up38",
   "source_url": "https://omlx.ai/benchmarks/performance/qda3up38",
   "tg_tokens": null,
   "tps": 182.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qda3up38-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 38.0 GB at this context; duplicate traces of the same session: 9seh0xdo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5755.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 5,755 153.1 38.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qda3up38",
   "source_url": "https://omlx.ai/benchmarks/performance/qda3up38",
   "tg_tokens": null,
   "tps": 153.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qzef9cxn-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: tvczndx1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4265.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 4,265 146.3 37.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qzef9cxn",
   "source_url": "https://omlx.ai/benchmarks/performance/qzef9cxn",
   "tg_tokens": null,
   "tps": 146.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qzef9cxn-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 38.0 GB at this context; duplicate traces of the same session: tvczndx1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5965.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 5,965 172.5 38.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qzef9cxn",
   "source_url": "https://omlx.ai/benchmarks/performance/qzef9cxn",
   "tg_tokens": null,
   "tps": 172.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-tscwz7ne-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: 903je29b, c483mo10, dw600zui, s9785ub9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4192.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 4,192 190.4 37.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tscwz7ne",
   "source_url": "https://omlx.ai/benchmarks/performance/tscwz7ne",
   "tg_tokens": null,
   "tps": 190.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tscwz7ne-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 38.0 GB at this context; duplicate traces of the same session: 903je29b, c483mo10, dw600zui, s9785ub9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5949.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 5,949 153.7 38.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tscwz7ne",
   "source_url": "https://omlx.ai/benchmarks/performance/tscwz7ne",
   "tg_tokens": null,
   "tps": 153.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-tscwz7ne-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 38.1 GB at this context; duplicate traces of the same session: 903je29b, c483mo10, dw600zui, s9785ub9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5835.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 5,835 155.1 38.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tscwz7ne",
   "source_url": "https://omlx.ai/benchmarks/performance/tscwz7ne",
   "tg_tokens": null,
   "tps": 155.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-tscwz7ne-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 38.2 GB at this context; duplicate traces of the same session: 903je29b, c483mo10, dw600zui, s9785ub9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5412.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 5,412 171.5 38.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tscwz7ne",
   "source_url": "https://omlx.ai/benchmarks/performance/tscwz7ne",
   "tg_tokens": null,
   "tps": 171.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-tscwz7ne-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 38.6 GB at this context; duplicate traces of the same session: 903je29b, c483mo10, dw600zui, s9785ub9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4608.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 4,608 168.1 38.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tscwz7ne",
   "source_url": "https://omlx.ai/benchmarks/performance/tscwz7ne",
   "tg_tokens": null,
   "tps": 168.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jnqahzul-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: ies65rwo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4205.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 4,205 189.2 37.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jnqahzul",
   "source_url": "https://omlx.ai/benchmarks/performance/jnqahzul",
   "tg_tokens": null,
   "tps": 189.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jnqahzul-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 38.0 GB at this context; duplicate traces of the same session: ies65rwo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5960.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 5,960 145.1 38.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jnqahzul",
   "source_url": "https://omlx.ai/benchmarks/performance/jnqahzul",
   "tg_tokens": null,
   "tps": 145.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-0ggtlt0t-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: rcd7z5vz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 129.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 129.5 15.2 16.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ggtlt0t",
   "source_url": "https://omlx.ai/benchmarks/performance/0ggtlt0t",
   "tg_tokens": null,
   "tps": 15.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-0ggtlt0t-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: rcd7z5vz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 134.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 134.6 15.0 17.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ggtlt0t",
   "source_url": "https://omlx.ai/benchmarks/performance/0ggtlt0t",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-9nkh8p27-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: 4e5jho1n, kshdo689, tvu9az25; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 424.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 424.6 17.9 17.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9nkh8p27",
   "source_url": "https://omlx.ai/benchmarks/performance/9nkh8p27",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-9nkh8p27-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 4e5jho1n, kshdo689, tvu9az25; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 397.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 397.3 16.7 17.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9nkh8p27",
   "source_url": "https://omlx.ai/benchmarks/performance/9nkh8p27",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-9nkh8p27-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 4e5jho1n, kshdo689, tvu9az25; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 330.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 64k 330.1 13.6 20.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9nkh8p27",
   "source_url": "https://omlx.ai/benchmarks/performance/9nkh8p27",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-9nkh8p27-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: 4e5jho1n, kshdo689, tvu9az25; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 265.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 128k 265.4 10.9 24.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9nkh8p27",
   "source_url": "https://omlx.ai/benchmarks/performance/9nkh8p27",
   "tg_tokens": null,
   "tps": 10.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ct35kt82-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 4ahw52gx; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 470.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 470.9 18.3 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ct35kt82",
   "source_url": "https://omlx.ai/benchmarks/performance/ct35kt82",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ct35kt82-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 4ahw52gx; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 504.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 504.5 17.7 17.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ct35kt82",
   "source_url": "https://omlx.ai/benchmarks/performance/ct35kt82",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-w720utbb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 15.9 GB; TTFT 4874 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 210.1,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-27B-nvfp4 on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k tokens 210.1 PP tok/s 7.9 TG tok/s 4874 TTFT (ms) 15.9 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w720utbb",
   "source_url": "https://omlx.ai/benchmarks/performance/w720utbb",
   "tg_tokens": null,
   "tps": 7.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-53dewg4e-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 22.7 GB; TTFT 110352 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 296.9,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-27B-oQ3e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 3bit; 32k tokens 296.9 PP tok/s 43.5 TG tok/s 110352 TTFT (ms) 22.7 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/53dewg4e",
   "source_url": "https://omlx.ai/benchmarks/performance/53dewg4e",
   "tg_tokens": null,
   "tps": 43.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-wyou7bj7-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: k4gvq5fr, yl8wzms8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 127.7 26.4 18.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wyou7bj7",
   "source_url": "https://omlx.ai/benchmarks/performance/wyou7bj7",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-wyou7bj7-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: k4gvq5fr, yl8wzms8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 125.9 28.3 19.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wyou7bj7",
   "source_url": "https://omlx.ai/benchmarks/performance/wyou7bj7",
   "tg_tokens": null,
   "tps": 28.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-wyou7bj7-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: k4gvq5fr, yl8wzms8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 118.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 118.9 33.4 20.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wyou7bj7",
   "source_url": "https://omlx.ai/benchmarks/performance/wyou7bj7",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-wyou7bj7-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: k4gvq5fr, yl8wzms8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 114.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 114.6 35.6 20.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wyou7bj7",
   "source_url": "https://omlx.ai/benchmarks/performance/wyou7bj7",
   "tg_tokens": null,
   "tps": 35.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-k3z8obu1-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 394.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 1k 394.7 25.2 26.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k3z8obu1",
   "source_url": "https://omlx.ai/benchmarks/performance/k3z8obu1",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-k3z8obu1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 457.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 4k 457.7 31.5 27.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k3z8obu1",
   "source_url": "https://omlx.ai/benchmarks/performance/k3z8obu1",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-51a28w60-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: z335hvxb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 814.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 814.4 47.7 33.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/51a28w60",
   "source_url": "https://omlx.ai/benchmarks/performance/51a28w60",
   "tg_tokens": null,
   "tps": 47.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-51a28w60-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: z335hvxb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 911.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 911.8 47.3 34.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/51a28w60",
   "source_url": "https://omlx.ai/benchmarks/performance/51a28w60",
   "tg_tokens": null,
   "tps": 47.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gj3ldfce-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: 90eq0hj0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 839.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 839.8 36.5 32.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gj3ldfce",
   "source_url": "https://omlx.ai/benchmarks/performance/gj3ldfce",
   "tg_tokens": null,
   "tps": 36.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gj3ldfce-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.7 GB at this context; duplicate traces of the same session: 90eq0hj0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 839.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 839.4 36.3 33.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gj3ldfce",
   "source_url": "https://omlx.ai/benchmarks/performance/gj3ldfce",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-uk2yzljo-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 39.8 GB at this context; duplicate traces of the same session: oivwt235; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 685.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 685.3 38.2 39.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uk2yzljo",
   "source_url": "https://omlx.ai/benchmarks/performance/uk2yzljo",
   "tg_tokens": null,
   "tps": 38.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-uk2yzljo-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 40.0 GB at this context; duplicate traces of the same session: oivwt235; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 779.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 779.5 35.0 40.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uk2yzljo",
   "source_url": "https://omlx.ai/benchmarks/performance/uk2yzljo",
   "tg_tokens": null,
   "tps": 35.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3n1y0aoe-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "plefp8 build; oMLX community benchmark trace; peak memory 74.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1258.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-Uncensored-plefp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,258 70.5 74.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3n1y0aoe",
   "source_url": "https://omlx.ai/benchmarks/performance/3n1y0aoe",
   "tg_tokens": null,
   "tps": 70.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3n1y0aoe-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "plefp8 build; oMLX community benchmark trace; peak memory 75.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1648.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-Uncensored-plefp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,648 69.5 75.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3n1y0aoe",
   "source_url": "https://omlx.ai/benchmarks/performance/3n1y0aoe",
   "tg_tokens": null,
   "tps": 69.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3n1y0aoe-ctx8192",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "plefp8 build; oMLX community benchmark trace; peak memory 75.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1807.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-Uncensored-plefp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,807 59.8 75.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3n1y0aoe",
   "source_url": "https://omlx.ai/benchmarks/performance/3n1y0aoe",
   "tg_tokens": null,
   "tps": 59.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3n1y0aoe-ctx16384",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "plefp8 build; oMLX community benchmark trace; peak memory 76.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1970.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-Uncensored-plefp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,970 60.5 76.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3n1y0aoe",
   "source_url": "https://omlx.ai/benchmarks/performance/3n1y0aoe",
   "tg_tokens": null,
   "tps": 60.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3n1y0aoe-ctx65536",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "plefp8 build; oMLX community benchmark trace; peak memory 78.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1988.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-Uncensored-plefp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 1,988 64.8 78.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3n1y0aoe",
   "source_url": "https://omlx.ai/benchmarks/performance/3n1y0aoe",
   "tg_tokens": null,
   "tps": 64.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3n1y0aoe-ctx131072",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "plefp8 build; oMLX community benchmark trace; peak memory 80.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1906.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-Uncensored-plefp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 1,906 68.3 80.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3n1y0aoe",
   "source_url": "https://omlx.ai/benchmarks/performance/3n1y0aoe",
   "tg_tokens": null,
   "tps": 68.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rk5f8s0m-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 99.0 GB; TTFT 2622 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 390.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 390.6 PP tok/s 81.9 TG tok/s 2622 TTFT (ms) 99.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rk5f8s0m",
   "source_url": "https://omlx.ai/benchmarks/performance/rk5f8s0m",
   "tg_tokens": null,
   "tps": 81.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-uu9s1t8q-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.7 GB; TTFT 3166 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 323.4 PP tok/s 43.6 TG tok/s 3166 TTFT (ms) 70.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uu9s1t8q",
   "source_url": "https://omlx.ai/benchmarks/performance/uu9s1t8q",
   "tg_tokens": null,
   "tps": 43.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ap5w04ol-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 69.2 GB; TTFT 992 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1032.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,032 PP tok/s 67.9 TG tok/s 992 TTFT (ms) 69.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ap5w04ol",
   "source_url": "https://omlx.ai/benchmarks/performance/ap5w04ol",
   "tg_tokens": null,
   "tps": 67.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-lyihwczy-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 49.6 GB at this context; duplicate traces of the same session: 21a2jr21; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 818.8,
   "provenance": "community",
   "quant": "bf16",
   "quote": "gemma-4-26b-a4b-it-bf16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 bf16; 1k 818.8 24.7 49.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lyihwczy",
   "source_url": "https://omlx.ai/benchmarks/performance/lyihwczy",
   "tg_tokens": null,
   "tps": 24.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-lyihwczy-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 50.0 GB at this context; duplicate traces of the same session: 21a2jr21; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 786.7,
   "provenance": "community",
   "quant": "bf16",
   "quote": "gemma-4-26b-a4b-it-bf16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 bf16; 4k 786.7 22.0 50.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lyihwczy",
   "source_url": "https://omlx.ai/benchmarks/performance/lyihwczy",
   "tg_tokens": null,
   "tps": 22.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-d5x2bl90-ctx32768",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 5.6 GB; TTFT 49336 ms; duplicate traces of the same session: none; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 664.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-E2B-it-MLX-8bit on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 8bit; 32k tokens 664.2 PP tok/s 15.7 TG tok/s 49336 TTFT (ms) 5.6 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d5x2bl90",
   "source_url": "https://omlx.ai/benchmarks/performance/d5x2bl90",
   "tg_tokens": null,
   "tps": 15.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-rfyuv133-ctx1024",
   "model": "gemma-4-E4B-Agentic-Sol-Fable-Reasoning-GeminiCLI",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: lzmba0j2; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 861.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E4B-Agentic-Sol-Fable-Reasoning-GeminiCLI-mlx-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 861.6 17.8 17.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rfyuv133",
   "source_url": "https://omlx.ai/benchmarks/performance/rfyuv133",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-rfyuv133-ctx4096",
   "model": "gemma-4-E4B-Agentic-Sol-Fable-Reasoning-GeminiCLI",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 17.5 GB at this context; duplicate traces of the same session: lzmba0j2; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 983.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E4B-Agentic-Sol-Fable-Reasoning-GeminiCLI-mlx-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 983.0 17.4 17.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rfyuv133",
   "source_url": "https://omlx.ai/benchmarks/performance/rfyuv133",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-72kjymgw-ctx1024",
   "model": "granite-4.1-3b",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 2.5 GB; TTFT 519 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1973.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "granite-4.1-3b-mxfp4 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 1,973 PP tok/s 168.5 TG tok/s 519 TTFT (ms) 2.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/72kjymgw",
   "source_url": "https://omlx.ai/benchmarks/performance/72kjymgw",
   "tg_tokens": null,
   "tps": 168.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-z7evg4u1-ctx4096",
   "model": "Huihui-Qwen3.5-4B-Claude-4.6-Opus-abliterated",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 3.8 GB at this context; duplicate traces of the same session: 7cobtq2h; config=5c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 166.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Huihui-Qwen3.5-4B-Claude-4.6-Opus-abliterated-4bit on M1 (5c); M1 (5c) \u00b7 8 GB \u00b7 4bit; 4k 166.8 18.2 3.8 GB view",
   "ram_gb": 8,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z7evg4u1",
   "source_url": "https://omlx.ai/benchmarks/performance/z7evg4u1",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-z7evg4u1-ctx16384",
   "model": "Huihui-Qwen3.5-4B-Claude-4.6-Opus-abliterated",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 4.5 GB at this context; duplicate traces of the same session: 7cobtq2h; config=5c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 138.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Huihui-Qwen3.5-4B-Claude-4.6-Opus-abliterated-4bit on M1 (5c); M1 (5c) \u00b7 8 GB \u00b7 4bit; 16k 138.4 15.6 4.5 GB current",
   "ram_gb": 8,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z7evg4u1",
   "source_url": "https://omlx.ai/benchmarks/performance/z7evg4u1",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-nfasw703-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: qq77gmqt; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 767.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 767.2 91.0 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nfasw703",
   "source_url": "https://omlx.ai/benchmarks/performance/nfasw703",
   "tg_tokens": null,
   "tps": 91.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-nfasw703-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: qq77gmqt; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 866.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 866.9 87.2 21.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nfasw703",
   "source_url": "https://omlx.ai/benchmarks/performance/nfasw703",
   "tg_tokens": null,
   "tps": 87.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ekul5jj7-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 89.0 GB at this context; duplicate traces of the same session: 17abyteu, k7licst7, vry0tz2c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 96.6,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 96.6 10.0 89.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ekul5jj7",
   "source_url": "https://omlx.ai/benchmarks/performance/ekul5jj7",
   "tg_tokens": null,
   "tps": 10.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ekul5jj7-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 91.8 GB at this context; duplicate traces of the same session: 17abyteu, k7licst7, vry0tz2c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 332.2,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 332.2 11.1 91.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ekul5jj7",
   "source_url": "https://omlx.ai/benchmarks/performance/ekul5jj7",
   "tg_tokens": null,
   "tps": 11.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ekul5jj7-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.2 GB at this context; duplicate traces of the same session: 17abyteu, k7licst7, vry0tz2c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 463.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 463.0 13.2 92.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ekul5jj7",
   "source_url": "https://omlx.ai/benchmarks/performance/ekul5jj7",
   "tg_tokens": null,
   "tps": 13.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ekul5jj7-ctx32768",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.1 GB at this context; duplicate traces of the same session: 17abyteu, k7licst7, vry0tz2c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 420.4,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 32k 420.4 13.4 92.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ekul5jj7",
   "source_url": "https://omlx.ai/benchmarks/performance/ekul5jj7",
   "tg_tokens": null,
   "tps": 13.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fopsok6o-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 92.2 GB; TTFT 31671 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 258.7,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k tokens 258.7 PP tok/s 12.6 TG tok/s 31671 TTFT (ms) 92.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fopsok6o",
   "source_url": "https://omlx.ai/benchmarks/performance/fopsok6o",
   "tg_tokens": null,
   "tps": 12.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9sfxi31s-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 92.0 GB; TTFT 31882 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 256.9,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k tokens 256.9 PP tok/s 10.6 TG tok/s 31882 TTFT (ms) 92.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9sfxi31s",
   "source_url": "https://omlx.ai/benchmarks/performance/9sfxi31s",
   "tg_tokens": null,
   "tps": 10.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tm7o4595-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 92.2 GB; TTFT 32460 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 252.4,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k tokens 252.4 PP tok/s 11.9 TG tok/s 32460 TTFT (ms) 92.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tm7o4595",
   "source_url": "https://omlx.ai/benchmarks/performance/tm7o4595",
   "tg_tokens": null,
   "tps": 11.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1m6y7ng3-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.7 GB at this context; duplicate traces of the same session: miskr5mi, uwe1jmg0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 521.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 521.6 31.9 177.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1m6y7ng3",
   "source_url": "https://omlx.ai/benchmarks/performance/1m6y7ng3",
   "tg_tokens": null,
   "tps": 31.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1m6y7ng3-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.8 GB at this context; duplicate traces of the same session: miskr5mi, uwe1jmg0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 589.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 589.6 38.6 178.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1m6y7ng3",
   "source_url": "https://omlx.ai/benchmarks/performance/1m6y7ng3",
   "tg_tokens": null,
   "tps": 38.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1m6y7ng3-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.6 GB at this context; duplicate traces of the same session: miskr5mi, uwe1jmg0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 527.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 527.9 46.4 179.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1m6y7ng3",
   "source_url": "https://omlx.ai/benchmarks/performance/1m6y7ng3",
   "tg_tokens": null,
   "tps": 46.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-vr14nm2z-ctx8192",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "fp16 build; oMLX community benchmark trace, single run at 8k context; peak memory 7.5 GB; TTFT 15410 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 531.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-oQ4e-fp16 on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 8k tokens 531.6 PP tok/s 58.3 TG tok/s 15410 TTFT (ms) 7.5 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vr14nm2z",
   "source_url": "https://omlx.ai/benchmarks/performance/vr14nm2z",
   "tg_tokens": null,
   "tps": 58.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-n9xlf8rq-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.0 GB at this context; duplicate traces of the same session: 3wqoj2yx; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 397.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-9B on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 8bit; 1k 397.7 38.9 10.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n9xlf8rq",
   "source_url": "https://omlx.ai/benchmarks/performance/n9xlf8rq",
   "tg_tokens": null,
   "tps": 38.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-n9xlf8rq-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.5 GB at this context; duplicate traces of the same session: 3wqoj2yx; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 371.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-9B on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 8bit; 4k 371.6 33.5 10.5 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n9xlf8rq",
   "source_url": "https://omlx.ai/benchmarks/performance/n9xlf8rq",
   "tg_tokens": null,
   "tps": 33.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-56x00s8f-ctx1024",
   "model": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed",
   "notes": "FP16; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 1icvlc9i, 48uqr4z4, bxawelqb, mzmmomin; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 781.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed-FP16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 781.3 77.0 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/56x00s8f",
   "source_url": "https://omlx.ai/benchmarks/performance/56x00s8f",
   "tg_tokens": null,
   "tps": 77.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-56x00s8f-ctx4096",
   "model": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed",
   "notes": "FP16; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 1icvlc9i, 48uqr4z4, bxawelqb, mzmmomin; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 886.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed-FP16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 886.7 75.8 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/56x00s8f",
   "source_url": "https://omlx.ai/benchmarks/performance/56x00s8f",
   "tg_tokens": null,
   "tps": 75.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-56x00s8f-ctx8192",
   "model": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed",
   "notes": "FP16; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 1icvlc9i, 48uqr4z4, bxawelqb, mzmmomin; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 834.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed-FP16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 8k 834.7 68.9 22.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/56x00s8f",
   "source_url": "https://omlx.ai/benchmarks/performance/56x00s8f",
   "tg_tokens": null,
   "tps": 68.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-56x00s8f-ctx16384",
   "model": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed",
   "notes": "FP16; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: 1icvlc9i, 48uqr4z4, bxawelqb, mzmmomin; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 737.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed-FP16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k 737.9 64.2 23.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/56x00s8f",
   "source_url": "https://omlx.ai/benchmarks/performance/56x00s8f",
   "tg_tokens": null,
   "tps": 64.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-56x00s8f-ctx32768",
   "model": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed",
   "notes": "FP16; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 1icvlc9i, 48uqr4z4, bxawelqb, mzmmomin; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 592.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed-FP16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 32k 592.6 55.3 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/56x00s8f",
   "source_url": "https://omlx.ai/benchmarks/performance/56x00s8f",
   "tg_tokens": null,
   "tps": 55.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-2w100iu0-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Official MLX build; oMLX community benchmark trace; peak memory not reported at this context; duplicate traces of the same session: kmkbq1fz, t2j9hvkz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 296.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Official-MLX on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 296.0 50.7 \u2014 current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2w100iu0",
   "source_url": "https://omlx.ai/benchmarks/performance/2w100iu0",
   "tg_tokens": null,
   "tps": 50.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-2w100iu0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "Official MLX build; oMLX community benchmark trace; peak memory not reported at this context; duplicate traces of the same session: kmkbq1fz, t2j9hvkz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 348.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Official-MLX on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 348.7 58.8 \u2014 view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2w100iu0",
   "source_url": "https://omlx.ai/benchmarks/performance/2w100iu0",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-2w100iu0-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "Official MLX build; oMLX community benchmark trace; peak memory not reported at this context; duplicate traces of the same session: kmkbq1fz, t2j9hvkz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 312.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Official-MLX on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 8k 312.0 46.8 \u2014 view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2w100iu0",
   "source_url": "https://omlx.ai/benchmarks/performance/2w100iu0",
   "tg_tokens": null,
   "tps": 46.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-b1xdoyte-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: iahrl0dd, ptbncmvj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 76.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-6bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 76.5 11.6 29.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b1xdoyte",
   "source_url": "https://omlx.ai/benchmarks/performance/b1xdoyte",
   "tg_tokens": null,
   "tps": 11.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-b1xdoyte-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 31.0 GB at this context; duplicate traces of the same session: iahrl0dd, ptbncmvj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 88.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-6bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 8k 88.7 16.2 31.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b1xdoyte",
   "source_url": "https://omlx.ai/benchmarks/performance/b1xdoyte",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-b1xdoyte-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 33.6 GB at this context; duplicate traces of the same session: iahrl0dd, ptbncmvj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 107.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-6bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 107.9 20.4 33.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b1xdoyte",
   "source_url": "https://omlx.ai/benchmarks/performance/b1xdoyte",
   "tg_tokens": null,
   "tps": 20.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-eaactx8p-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MLX build; oMLX community benchmark trace, single run at 1k context; peak memory 22.3 GB; TTFT 5186 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 197.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-6bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k tokens 197.4 PP tok/s 18.3 TG tok/s 5186 TTFT (ms) 22.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eaactx8p",
   "source_url": "https://omlx.ai/benchmarks/performance/eaactx8p",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-78fl5jlq-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 27.0 GB at this context; duplicate traces of the same session: 32qi09le; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 176.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-6bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 176.5 30.0 27.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/78fl5jlq",
   "source_url": "https://omlx.ai/benchmarks/performance/78fl5jlq",
   "tg_tokens": null,
   "tps": 30.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-78fl5jlq-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 31.0 GB at this context; duplicate traces of the same session: 32qi09le; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 203.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-6bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 8k 203.7 29.6 31.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/78fl5jlq",
   "source_url": "https://omlx.ai/benchmarks/performance/78fl5jlq",
   "tg_tokens": null,
   "tps": 29.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-b0kf60al-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MLX build; oMLX community benchmark trace, single run at 1k context; peak memory 27.0 GB; TTFT 5557 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 184.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-6bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k tokens 184.3 PP tok/s 18.0 TG tok/s 5557 TTFT (ms) 27.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b0kf60al",
   "source_url": "https://omlx.ai/benchmarks/performance/b0kf60al",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-pzm33ygo-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 453v8nqu, 66zijhv7, odjffeti, pp423uq1; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 111.4,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 111.4 13.8 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pzm33ygo",
   "source_url": "https://omlx.ai/benchmarks/performance/pzm33ygo",
   "tg_tokens": null,
   "tps": 13.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-pzm33ygo-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 453v8nqu, 66zijhv7, odjffeti, pp423uq1; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 118.4,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 118.4 15.5 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pzm33ygo",
   "source_url": "https://omlx.ai/benchmarks/performance/pzm33ygo",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-pzm33ygo-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: 453v8nqu, 66zijhv7, odjffeti, pp423uq1; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 117.2,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 8k 117.2 14.4 24.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pzm33ygo",
   "source_url": "https://omlx.ai/benchmarks/performance/pzm33ygo",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-pzm33ygo-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: 453v8nqu, 66zijhv7, odjffeti, pp423uq1; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 110.5,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k 110.5 13.6 25.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pzm33ygo",
   "source_url": "https://omlx.ai/benchmarks/performance/pzm33ygo",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-pzm33ygo-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: 453v8nqu, 66zijhv7, odjffeti, pp423uq1; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.3,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 32k 102.3 10.6 26.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pzm33ygo",
   "source_url": "https://omlx.ai/benchmarks/performance/pzm33ygo",
   "tg_tokens": null,
   "tps": 10.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-tjos262g-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 30.7 GB; TTFT 360863 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 181.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k tokens 181.6 PP tok/s 23.8 TG tok/s 360863 TTFT (ms) 30.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tjos262g",
   "source_url": "https://omlx.ai/benchmarks/performance/tjos262g",
   "tg_tokens": null,
   "tps": 23.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-3p5n82v3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: eysnp2p2, ojeobq26; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 103.4 26.5 20.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3p5n82v3",
   "source_url": "https://omlx.ai/benchmarks/performance/3p5n82v3",
   "tg_tokens": null,
   "tps": 26.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-3p5n82v3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: eysnp2p2, ojeobq26; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 149.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 149.1 24.6 21.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3p5n82v3",
   "source_url": "https://omlx.ai/benchmarks/performance/3p5n82v3",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-3p5n82v3-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: eysnp2p2, ojeobq26; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 99.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 99.5 26.9 20.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3p5n82v3",
   "source_url": "https://omlx.ai/benchmarks/performance/3p5n82v3",
   "tg_tokens": null,
   "tps": 26.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-h5pq0tin-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 8setgpnw; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 100.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 100.8 27.8 20.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h5pq0tin",
   "source_url": "https://omlx.ai/benchmarks/performance/h5pq0tin",
   "tg_tokens": null,
   "tps": 27.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-h5pq0tin-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: 8setgpnw; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 117.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 117.6 26.1 21.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h5pq0tin",
   "source_url": "https://omlx.ai/benchmarks/performance/h5pq0tin",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-srxvuyib-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 127.7 26.4 18.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/srxvuyib",
   "source_url": "https://omlx.ai/benchmarks/performance/srxvuyib",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-srxvuyib-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 125.9 28.3 19.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/srxvuyib",
   "source_url": "https://omlx.ai/benchmarks/performance/srxvuyib",
   "tg_tokens": null,
   "tps": 28.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-srxvuyib-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 118.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 118.9 33.4 20.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/srxvuyib",
   "source_url": "https://omlx.ai/benchmarks/performance/srxvuyib",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-srxvuyib-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 114.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 114.6 35.6 20.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/srxvuyib",
   "source_url": "https://omlx.ai/benchmarks/performance/srxvuyib",
   "tg_tokens": null,
   "tps": 35.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-hvj88c4a-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 699ip8wy; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 379.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 379.9 42.4 19.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hvj88c4a",
   "source_url": "https://omlx.ai/benchmarks/performance/hvj88c4a",
   "tg_tokens": null,
   "tps": 42.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-hvj88c4a-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: 699ip8wy; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 353.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 353.4 64.3 20.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hvj88c4a",
   "source_url": "https://omlx.ai/benchmarks/performance/hvj88c4a",
   "tg_tokens": null,
   "tps": 64.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-5d9fsdax-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: rp9ykckt, ww8dkvc7; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1543.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 4k 1,543 100.3 18.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5d9fsdax",
   "source_url": "https://omlx.ai/benchmarks/performance/5d9fsdax",
   "tg_tokens": null,
   "tps": 100.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-5d9fsdax-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: rp9ykckt, ww8dkvc7; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1504.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 16k 1,504 124.1 19.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5d9fsdax",
   "source_url": "https://omlx.ai/benchmarks/performance/5d9fsdax",
   "tg_tokens": null,
   "tps": 124.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-5d9fsdax-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: rp9ykckt, ww8dkvc7; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1189.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 64k 1,189 89.2 23.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5d9fsdax",
   "source_url": "https://omlx.ai/benchmarks/performance/5d9fsdax",
   "tg_tokens": null,
   "tps": 89.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ew7n4txt-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 22.4 GB; TTFT 46173 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 709.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 5bit; 32k tokens 709.7 PP tok/s 19.4 TG tok/s 46173 TTFT (ms) 22.4 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ew7n4txt",
   "source_url": "https://omlx.ai/benchmarks/performance/ew7n4txt",
   "tg_tokens": null,
   "tps": 19.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-2db6l86z-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 394.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 1k 394.7 25.2 26.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2db6l86z",
   "source_url": "https://omlx.ai/benchmarks/performance/2db6l86z",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-2db6l86z-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 457.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 4k 457.7 31.5 27.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2db6l86z",
   "source_url": "https://omlx.ai/benchmarks/performance/2db6l86z",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-126exmbo-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP320",
   "notes": "DWQ MTP Vision build; oMLX community benchmark trace; peak memory 67.8 GB at this context; duplicate traces of the same session: w5l1hvn7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 898.6,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-Flash-Next-REAP320-oQ3e-fp16-DWQ-MTP-Vision-MLX on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 3bit; 1k 898.6 39.7 67.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/126exmbo",
   "source_url": "https://omlx.ai/benchmarks/performance/126exmbo",
   "tg_tokens": null,
   "tps": 39.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-126exmbo-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP320",
   "notes": "DWQ MTP Vision build; oMLX community benchmark trace; peak memory 70.4 GB at this context; duplicate traces of the same session: w5l1hvn7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1767.0,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-Flash-Next-REAP320-oQ3e-fp16-DWQ-MTP-Vision-MLX on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 3bit; 4k 1,767 36.7 70.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/126exmbo",
   "source_url": "https://omlx.ai/benchmarks/performance/126exmbo",
   "tg_tokens": null,
   "tps": 36.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-90q208ia-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: l5uuledo, m3bpa6zu, mn1j6l6l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 375.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 375.0 49.1 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/90q208ia",
   "source_url": "https://omlx.ai/benchmarks/performance/90q208ia",
   "tg_tokens": null,
   "tps": 49.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-90q208ia-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: l5uuledo, m3bpa6zu, mn1j6l6l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 997.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 997.6 66.7 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/90q208ia",
   "source_url": "https://omlx.ai/benchmarks/performance/90q208ia",
   "tg_tokens": null,
   "tps": 66.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-90q208ia-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: l5uuledo, m3bpa6zu, mn1j6l6l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1738.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,738 39.9 71.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/90q208ia",
   "source_url": "https://omlx.ai/benchmarks/performance/90q208ia",
   "tg_tokens": null,
   "tps": 39.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-90q208ia-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: l5uuledo, m3bpa6zu, mn1j6l6l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2396.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,396 62.6 72.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/90q208ia",
   "source_url": "https://omlx.ai/benchmarks/performance/90q208ia",
   "tg_tokens": null,
   "tps": 62.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3adttvk1-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: b9cias41, md0v23f5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1227.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,227 90.9 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3adttvk1",
   "source_url": "https://omlx.ai/benchmarks/performance/3adttvk1",
   "tg_tokens": null,
   "tps": 90.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3adttvk1-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: b9cias41, md0v23f5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2020.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,020 63.2 70.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3adttvk1",
   "source_url": "https://omlx.ai/benchmarks/performance/3adttvk1",
   "tg_tokens": null,
   "tps": 63.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3adttvk1-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: b9cias41, md0v23f5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2112.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,112 65.0 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3adttvk1",
   "source_url": "https://omlx.ai/benchmarks/performance/3adttvk1",
   "tg_tokens": null,
   "tps": 65.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-p0wowgvq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: autgv1gc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 565.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 565.5 25.4 21.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p0wowgvq",
   "source_url": "https://omlx.ai/benchmarks/performance/p0wowgvq",
   "tg_tokens": null,
   "tps": 25.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-p0wowgvq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: autgv1gc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 888.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 888.1 24.7 22.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p0wowgvq",
   "source_url": "https://omlx.ai/benchmarks/performance/p0wowgvq",
   "tg_tokens": null,
   "tps": 24.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-9zdjfknw-ctx1024",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: bkjvkd8e; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1201.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "nerkyok-Qwen3.8-27B-Coder390-oQe6 on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 1k 1,201 63.7 23.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9zdjfknw",
   "source_url": "https://omlx.ai/benchmarks/performance/9zdjfknw",
   "tg_tokens": null,
   "tps": 63.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-9zdjfknw-ctx4096",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: bkjvkd8e; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1428.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "nerkyok-Qwen3.8-27B-Coder390-oQe6 on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 4k 1,428 71.1 25.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9zdjfknw",
   "source_url": "https://omlx.ai/benchmarks/performance/9zdjfknw",
   "tg_tokens": null,
   "tps": 71.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-53p6l7a9-ctx1024",
   "model": "DeepSeek-Coder-V2-Lite-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 9.1 GB at this context; duplicate traces of the same session: jhc3x1zo; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1411.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-Coder-V2-Lite-Instruct-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 1,411 121.3 9.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/53p6l7a9",
   "source_url": "https://omlx.ai/benchmarks/performance/53p6l7a9",
   "tg_tokens": null,
   "tps": 121.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-53p6l7a9-ctx4096",
   "model": "DeepSeek-Coder-V2-Lite-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 10.2 GB at this context; duplicate traces of the same session: jhc3x1zo; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2022.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-Coder-V2-Lite-Instruct-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 2,022 89.2 10.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/53p6l7a9",
   "source_url": "https://omlx.ai/benchmarks/performance/53p6l7a9",
   "tg_tokens": null,
   "tps": 89.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-gaed6n9j-ctx1024",
   "model": "DeepSeek-R1-Distill-Qwen-7B",
   "notes": "oMLX community benchmark trace; peak memory 4.8 GB at this context; duplicate traces of the same session: 7hn0stjz; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1556.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-R1-Distill-Qwen-7B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 1,556 64.7 4.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gaed6n9j",
   "source_url": "https://omlx.ai/benchmarks/performance/gaed6n9j",
   "tg_tokens": null,
   "tps": 64.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-gaed6n9j-ctx4096",
   "model": "DeepSeek-R1-Distill-Qwen-7B",
   "notes": "oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: 7hn0stjz; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1548.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-R1-Distill-Qwen-7B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 1,548 61.2 4.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gaed6n9j",
   "source_url": "https://omlx.ai/benchmarks/performance/gaed6n9j",
   "tg_tokens": null,
   "tps": 61.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cjflva5s-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 92.2 GB; TTFT 28720 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 285.2,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k tokens 285.2 PP tok/s 8.1 TG tok/s 28720 TTFT (ms) 92.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cjflva5s",
   "source_url": "https://omlx.ai/benchmarks/performance/cjflva5s",
   "tg_tokens": null,
   "tps": 8.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nsaec5vf-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 90.0 GB; TTFT 11703 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 87.5,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k tokens 87.5 PP tok/s 10.5 TG tok/s 11703 TTFT (ms) 90.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nsaec5vf",
   "source_url": "https://omlx.ai/benchmarks/performance/nsaec5vf",
   "tg_tokens": null,
   "tps": 10.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-n5k1gtao-ctx8192",
   "model": "Ling-3.0-tiny",
   "notes": "fp16 build; oMLX community benchmark trace, single run at 8k context; peak memory 9.4 GB; TTFT 6335 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1293.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ling-3.0-tiny-oQ8e-fp16 on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 8bit; 8k tokens 1,293 PP tok/s 67.7 TG tok/s 6335 TTFT (ms) 9.4 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n5k1gtao",
   "source_url": "https://omlx.ai/benchmarks/performance/n5k1gtao",
   "tg_tokens": null,
   "tps": 67.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-s06n204g-ctx1024",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 0fl979xk; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 874.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 1k 874.0 95.4 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s06n204g",
   "source_url": "https://omlx.ai/benchmarks/performance/s06n204g",
   "tg_tokens": null,
   "tps": 95.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-s06n204g-ctx4096",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 0fl979xk; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 987.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k 987.7 92.8 21.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s06n204g",
   "source_url": "https://omlx.ai/benchmarks/performance/s06n204g",
   "tg_tokens": null,
   "tps": 92.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-3pn1acq9-ctx1024",
   "model": "Ornith-1.5-35B-A3B-BigBang",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: xuw0efn9; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 648.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-BigBang-MTP-mlx-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 648.5 72.3 21.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3pn1acq9",
   "source_url": "https://omlx.ai/benchmarks/performance/3pn1acq9",
   "tg_tokens": null,
   "tps": 72.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-3pn1acq9-ctx4096",
   "model": "Ornith-1.5-35B-A3B-BigBang",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: xuw0efn9; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 767.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-BigBang-MTP-mlx-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 767.8 70.0 23.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3pn1acq9",
   "source_url": "https://omlx.ai/benchmarks/performance/3pn1acq9",
   "tg_tokens": null,
   "tps": 70.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-0vcb4txb-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: jxslhhko; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 946.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 1k 946.9 129.1 35.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0vcb4txb",
   "source_url": "https://omlx.ai/benchmarks/performance/0vcb4txb",
   "tg_tokens": null,
   "tps": 129.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-0vcb4txb-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: jxslhhko; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3475.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 4k 3,475 127.0 36.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0vcb4txb",
   "source_url": "https://omlx.ai/benchmarks/performance/0vcb4txb",
   "tg_tokens": null,
   "tps": 127.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9z7j94a9-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 6utmae58, jldl5lpc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2220.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 2,220 122.4 23.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9z7j94a9",
   "source_url": "https://omlx.ai/benchmarks/performance/9z7j94a9",
   "tg_tokens": null,
   "tps": 122.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9z7j94a9-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: 6utmae58, jldl5lpc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1812.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 1,812 85.1 24.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9z7j94a9",
   "source_url": "https://omlx.ai/benchmarks/performance/9z7j94a9",
   "tg_tokens": null,
   "tps": 85.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9z7j94a9-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: 6utmae58, jldl5lpc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1319.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 1,319 59.3 25.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9z7j94a9",
   "source_url": "https://omlx.ai/benchmarks/performance/9z7j94a9",
   "tg_tokens": null,
   "tps": 59.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-bx5vrd14-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 604bvc4s, a0x601z6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2210.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 2,210 119.5 23.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bx5vrd14",
   "source_url": "https://omlx.ai/benchmarks/performance/bx5vrd14",
   "tg_tokens": null,
   "tps": 119.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-bx5vrd14-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: 604bvc4s, a0x601z6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1805.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 1,805 79.3 24.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bx5vrd14",
   "source_url": "https://omlx.ai/benchmarks/performance/bx5vrd14",
   "tg_tokens": null,
   "tps": 79.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-bx5vrd14-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: 604bvc4s, a0x601z6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1319.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 1,319 58.8 25.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bx5vrd14",
   "source_url": "https://omlx.ai/benchmarks/performance/bx5vrd14",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-dhep4ej2-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.8 GB at this context; duplicate traces of the same session: phdz8kuc, qunn5huh; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2188.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 16k 2,188 101.0 39.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dhep4ej2",
   "source_url": "https://omlx.ai/benchmarks/performance/dhep4ej2",
   "tg_tokens": null,
   "tps": 101.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-dhep4ej2-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 40.5 GB at this context; duplicate traces of the same session: phdz8kuc, qunn5huh; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1786.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 32k 1,786 95.1 40.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dhep4ej2",
   "source_url": "https://omlx.ai/benchmarks/performance/dhep4ej2",
   "tg_tokens": null,
   "tps": 95.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-dhep4ej2-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 41.8 GB at this context; duplicate traces of the same session: phdz8kuc, qunn5huh; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1307.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 64k 1,307 56.8 41.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dhep4ej2",
   "source_url": "https://omlx.ai/benchmarks/performance/dhep4ej2",
   "tg_tokens": null,
   "tps": 56.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nmpaka8y-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.8 GB at this context; duplicate traces of the same session: e843xlyk, ncrboljd; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2196.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 16k 2,196 92.5 39.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nmpaka8y",
   "source_url": "https://omlx.ai/benchmarks/performance/nmpaka8y",
   "tg_tokens": null,
   "tps": 92.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nmpaka8y-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 40.5 GB at this context; duplicate traces of the same session: e843xlyk, ncrboljd; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1797.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 32k 1,797 85.4 40.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nmpaka8y",
   "source_url": "https://omlx.ai/benchmarks/performance/nmpaka8y",
   "tg_tokens": null,
   "tps": 85.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nmpaka8y-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 41.8 GB at this context; duplicate traces of the same session: e843xlyk, ncrboljd; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1310.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 64k 1,310 59.2 41.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nmpaka8y",
   "source_url": "https://omlx.ai/benchmarks/performance/nmpaka8y",
   "tg_tokens": null,
   "tps": 59.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2f0nn96o-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: lw2lxugs; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 772.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 772.4 18.4 28.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2f0nn96o",
   "source_url": "https://omlx.ai/benchmarks/performance/2f0nn96o",
   "tg_tokens": null,
   "tps": 18.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2f0nn96o-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: lw2lxugs; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 904.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 904.5 18.3 29.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2f0nn96o",
   "source_url": "https://omlx.ai/benchmarks/performance/2f0nn96o",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-6vehcubc-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MLX build; oMLX community benchmark trace, single run at 1k context; peak memory 27.0 GB; TTFT 5260 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 194.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-6bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k tokens 194.7 PP tok/s 19.7 TG tok/s 5260 TTFT (ms) 27.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6vehcubc",
   "source_url": "https://omlx.ai/benchmarks/performance/6vehcubc",
   "tg_tokens": null,
   "tps": 19.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-noky41tm-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 27.0 GB at this context; duplicate traces of the same session: 28tqrscy, r2hsqc8n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 157.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-6bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 157.3 24.6 27.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/noky41tm",
   "source_url": "https://omlx.ai/benchmarks/performance/noky41tm",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-noky41tm-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 31.0 GB at this context; duplicate traces of the same session: 28tqrscy, r2hsqc8n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 180.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-6bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 8k 180.9 24.6 31.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/noky41tm",
   "source_url": "https://omlx.ai/benchmarks/performance/noky41tm",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-noky41tm-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 33.6 GB at this context; duplicate traces of the same session: 28tqrscy, r2hsqc8n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 163.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-6bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 163.6 19.3 33.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/noky41tm",
   "source_url": "https://omlx.ai/benchmarks/performance/noky41tm",
   "tg_tokens": null,
   "tps": 19.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-29dhqpfy-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 4ib10y56, bid6470a, wwz3m3bf, zqhuks2j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 815.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 815.0 18.0 28.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/29dhqpfy",
   "source_url": "https://omlx.ai/benchmarks/performance/29dhqpfy",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-29dhqpfy-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: 4ib10y56, bid6470a, wwz3m3bf, zqhuks2j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 885.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 885.1 18.1 29.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/29dhqpfy",
   "source_url": "https://omlx.ai/benchmarks/performance/29dhqpfy",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-29dhqpfy-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: 4ib10y56, bid6470a, wwz3m3bf, zqhuks2j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 704.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 704.7 17.6 30.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/29dhqpfy",
   "source_url": "https://omlx.ai/benchmarks/performance/29dhqpfy",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-29dhqpfy-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 31.3 GB at this context; duplicate traces of the same session: 4ib10y56, bid6470a, wwz3m3bf, zqhuks2j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 652.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 652.3 16.5 31.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/29dhqpfy",
   "source_url": "https://omlx.ai/benchmarks/performance/29dhqpfy",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-29dhqpfy-ctx65536",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 33.3 GB at this context; duplicate traces of the same session: 4ib10y56, bid6470a, wwz3m3bf, zqhuks2j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 518.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 518.8 13.9 33.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/29dhqpfy",
   "source_url": "https://omlx.ai/benchmarks/performance/29dhqpfy",
   "tg_tokens": null,
   "tps": 13.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-g3rth751-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 33.7 GB; TTFT 1416 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 723.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k tokens 723.0 PP tok/s 33.6 TG tok/s 1416 TTFT (ms) 33.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g3rth751",
   "source_url": "https://omlx.ai/benchmarks/performance/g3rth751",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-p05sxpm9-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP320",
   "notes": "MTP Vision build; oMLX community benchmark trace; peak memory 36.7 GB at this context; duplicate traces of the same session: ctv4sulz; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 470.9,
   "provenance": "community",
   "quant": "oQ3e-DWQ",
   "quote": "Qwen3.8-Flash-Next-REAP320-oQ3e-DWQ-MTP-Vision-MLX on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 3bit; 1k 470.9 34.2 36.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p05sxpm9",
   "source_url": "https://omlx.ai/benchmarks/performance/p05sxpm9",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-p05sxpm9-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP320",
   "notes": "MTP Vision build; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: ctv4sulz; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 878.5,
   "provenance": "community",
   "quant": "oQ3e-DWQ",
   "quote": "Qwen3.8-Flash-Next-REAP320-oQ3e-DWQ-MTP-Vision-MLX on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 3bit; 4k 878.5 35.3 36.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p05sxpm9",
   "source_url": "https://omlx.ai/benchmarks/performance/p05sxpm9",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1h41vh1k-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 61.3 GB at this context; duplicate traces of the same session: 60x554ak, omq25cjd, rjxkp7oh; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 574.5,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 3bit; 1k 574.5 48.1 61.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1h41vh1k",
   "source_url": "https://omlx.ai/benchmarks/performance/1h41vh1k",
   "tg_tokens": null,
   "tps": 48.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1h41vh1k-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 61.9 GB at this context; duplicate traces of the same session: 60x554ak, omq25cjd, rjxkp7oh; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 796.3,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 3bit; 4k 796.3 65.1 61.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1h41vh1k",
   "source_url": "https://omlx.ai/benchmarks/performance/1h41vh1k",
   "tg_tokens": null,
   "tps": 65.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1h41vh1k-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 65.0 GB at this context; duplicate traces of the same session: 60x554ak, omq25cjd, rjxkp7oh; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1029.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 3bit; 64k 1,029 60.9 65.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1h41vh1k",
   "source_url": "https://omlx.ai/benchmarks/performance/1h41vh1k",
   "tg_tokens": null,
   "tps": 60.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1h41vh1k-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.7 GB at this context; duplicate traces of the same session: 60x554ak, omq25cjd, rjxkp7oh; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1011.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 3bit; 128k 1,011 50.6 69.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1h41vh1k",
   "source_url": "https://omlx.ai/benchmarks/performance/1h41vh1k",
   "tg_tokens": null,
   "tps": 50.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6mic5n02-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 86.5 GB; TTFT 1195 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 856.8,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k tokens 856.8 PP tok/s 45.5 TG tok/s 1195 TTFT (ms) 86.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6mic5n02",
   "source_url": "https://omlx.ai/benchmarks/performance/6mic5n02",
   "tg_tokens": null,
   "tps": 45.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-m08zox0w-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: nptpiez5, zjoyt3ne; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1424.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k 1,424 81.2 19.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m08zox0w",
   "source_url": "https://omlx.ai/benchmarks/performance/m08zox0w",
   "tg_tokens": null,
   "tps": 81.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-m08zox0w-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: nptpiez5, zjoyt3ne; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1495.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 8k 1,495 80.6 21.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m08zox0w",
   "source_url": "https://omlx.ai/benchmarks/performance/m08zox0w",
   "tg_tokens": null,
   "tps": 80.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-m08zox0w-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: nptpiez5, zjoyt3ne; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1337.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 32k 1,337 83.2 22.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m08zox0w",
   "source_url": "https://omlx.ai/benchmarks/performance/m08zox0w",
   "tg_tokens": null,
   "tps": 83.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-5vhyt94q-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1415.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k 1,415 77.5 19.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5vhyt94q",
   "source_url": "https://omlx.ai/benchmarks/performance/5vhyt94q",
   "tg_tokens": null,
   "tps": 77.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-5vhyt94q-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1490.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 8k 1,490 93.4 21.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5vhyt94q",
   "source_url": "https://omlx.ai/benchmarks/performance/5vhyt94q",
   "tg_tokens": null,
   "tps": 93.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-5vhyt94q-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1332.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 32k 1,332 86.1 22.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5vhyt94q",
   "source_url": "https://omlx.ai/benchmarks/performance/5vhyt94q",
   "tg_tokens": null,
   "tps": 86.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-yrd4evnk-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1421.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k 1,421 68.1 19.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yrd4evnk",
   "source_url": "https://omlx.ai/benchmarks/performance/yrd4evnk",
   "tg_tokens": null,
   "tps": 68.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-yrd4evnk-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1522.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 4k 1,522 79.0 20.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yrd4evnk",
   "source_url": "https://omlx.ai/benchmarks/performance/yrd4evnk",
   "tg_tokens": null,
   "tps": 79.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-adrjpkgw-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: bgpahobd, dckqlzwq, t6atjkkc, yd86b415; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 897.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 897.4 25.7 21.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/adrjpkgw",
   "source_url": "https://omlx.ai/benchmarks/performance/adrjpkgw",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-adrjpkgw-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: bgpahobd, dckqlzwq, t6atjkkc, yd86b415; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 914.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 914.0 25.4 22.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/adrjpkgw",
   "source_url": "https://omlx.ai/benchmarks/performance/adrjpkgw",
   "tg_tokens": null,
   "tps": 25.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-adrjpkgw-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: bgpahobd, dckqlzwq, t6atjkkc, yd86b415; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 903.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 8k 903.2 24.5 22.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/adrjpkgw",
   "source_url": "https://omlx.ai/benchmarks/performance/adrjpkgw",
   "tg_tokens": null,
   "tps": 24.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-adrjpkgw-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: bgpahobd, dckqlzwq, t6atjkkc, yd86b415; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 863.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 863.0 24.2 23.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/adrjpkgw",
   "source_url": "https://omlx.ai/benchmarks/performance/adrjpkgw",
   "tg_tokens": null,
   "tps": 24.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-14",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-adrjpkgw-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: bgpahobd, dckqlzwq, t6atjkkc, yd86b415; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 786.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 786.6 23.1 24.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/adrjpkgw",
   "source_url": "https://omlx.ai/benchmarks/performance/adrjpkgw",
   "tg_tokens": null,
   "tps": 23.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-k2uj9zed-ctx1024",
   "model": "deepseek-coder-6.7b-instruct-hf",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 4.8 GB at this context; duplicate traces of the same session: y1c6h8st; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1583.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "deepseek-coder-6.7b-instruct-hf-4bit-mlx on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 1,583 62.0 4.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k2uj9zed",
   "source_url": "https://omlx.ai/benchmarks/performance/k2uj9zed",
   "tg_tokens": null,
   "tps": 62.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-k2uj9zed-ctx4096",
   "model": "deepseek-coder-6.7b-instruct-hf",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: y1c6h8st; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1547.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "deepseek-coder-6.7b-instruct-hf-4bit-mlx on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 1,547 45.2 6.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k2uj9zed",
   "source_url": "https://omlx.ai/benchmarks/performance/k2uj9zed",
   "tg_tokens": null,
   "tps": 45.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-gmfw9pqd-ctx1024",
   "model": "deepseek-coder-6.7b-instruct-hf",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 4.8 GB at this context; duplicate traces of the same session: 32g3imuf; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1546.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "deepseek-coder-6.7b-instruct-hf-4bit-mlx on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 1,546 62.1 4.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gmfw9pqd",
   "source_url": "https://omlx.ai/benchmarks/performance/gmfw9pqd",
   "tg_tokens": null,
   "tps": 62.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-gmfw9pqd-ctx4096",
   "model": "deepseek-coder-6.7b-instruct-hf",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: 32g3imuf; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1502.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "deepseek-coder-6.7b-instruct-hf-4bit-mlx on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 1,502 45.2 6.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gmfw9pqd",
   "source_url": "https://omlx.ai/benchmarks/performance/gmfw9pqd",
   "tg_tokens": null,
   "tps": 45.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-gj2bqkkv-ctx1024",
   "model": "Occamy-1.0-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: g0upuc7f; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 648.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "occamy-1.0-abliterated-mtp-mlx-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 648.6 64.4 21.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gj2bqkkv",
   "source_url": "https://omlx.ai/benchmarks/performance/gj2bqkkv",
   "tg_tokens": null,
   "tps": 64.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-14",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-gj2bqkkv-ctx4096",
   "model": "Occamy-1.0-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: g0upuc7f; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 766.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "occamy-1.0-abliterated-mtp-mlx-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 766.9 70.5 23.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gj2bqkkv",
   "source_url": "https://omlx.ai/benchmarks/performance/gj2bqkkv",
   "tg_tokens": null,
   "tps": 70.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ckxycxxz-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 90.0 GB; TTFT 11728 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 87.3,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k tokens 87.3 PP tok/s 10.3 TG tok/s 11728 TTFT (ms) 90.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ckxycxxz",
   "source_url": "https://omlx.ai/benchmarks/performance/ckxycxxz",
   "tg_tokens": null,
   "tps": 10.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bk9mqavi-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 92.2 GB; TTFT 50969 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 321.5,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k tokens 321.5 PP tok/s 13.6 TG tok/s 50969 TTFT (ms) 92.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bk9mqavi",
   "source_url": "https://omlx.ai/benchmarks/performance/bk9mqavi",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-kdxiwals-ctx1024",
   "model": "GLM-4.7-Flash",
   "notes": "oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: 7omfelyy; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 295.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-4.7-Flash-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 295.8 33.6 16.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kdxiwals",
   "source_url": "https://omlx.ai/benchmarks/performance/kdxiwals",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-kdxiwals-ctx4096",
   "model": "GLM-4.7-Flash",
   "notes": "oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 7omfelyy; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 260.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-4.7-Flash-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 260.0 27.9 17.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kdxiwals",
   "source_url": "https://omlx.ai/benchmarks/performance/kdxiwals",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-lkh14pab-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX fp16 build; oMLX community benchmark trace, single run at 8k context; peak memory 13.8 GB; TTFT 12583 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 651.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit-fp16 on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 8k tokens 651.1 PP tok/s 61.6 TG tok/s 12583 TTFT (ms) 13.8 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lkh14pab",
   "source_url": "https://omlx.ai/benchmarks/performance/lkh14pab",
   "tg_tokens": null,
   "tps": 61.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-9a2ed4y1-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: 7ku74s0u, roe03til, ssoyccve, zjn00cws; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2034.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-24B-A2B-oQ8e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 2,034 109.2 24.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9a2ed4y1",
   "source_url": "https://omlx.ai/benchmarks/performance/9a2ed4y1",
   "tg_tokens": null,
   "tps": 109.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-9a2ed4y1-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: 7ku74s0u, roe03til, ssoyccve, zjn00cws; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2436.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-24B-A2B-oQ8e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 2,436 104.6 24.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9a2ed4y1",
   "source_url": "https://omlx.ai/benchmarks/performance/9a2ed4y1",
   "tg_tokens": null,
   "tps": 104.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-9a2ed4y1-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: 7ku74s0u, roe03til, ssoyccve, zjn00cws; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2421.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-24B-A2B-oQ8e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 8k 2,421 98.8 24.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9a2ed4y1",
   "source_url": "https://omlx.ai/benchmarks/performance/9a2ed4y1",
   "tg_tokens": null,
   "tps": 98.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-9a2ed4y1-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: 7ku74s0u, roe03til, ssoyccve, zjn00cws; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2331.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-24B-A2B-oQ8e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 16k 2,331 89.7 25.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9a2ed4y1",
   "source_url": "https://omlx.ai/benchmarks/performance/9a2ed4y1",
   "tg_tokens": null,
   "tps": 89.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-9a2ed4y1-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: 7ku74s0u, roe03til, ssoyccve, zjn00cws; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2084.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-24B-A2B-oQ8e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 32k 2,084 77.6 25.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9a2ed4y1",
   "source_url": "https://omlx.ai/benchmarks/performance/9a2ed4y1",
   "tg_tokens": null,
   "tps": 77.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-abibnok6-ctx32768",
   "model": "Qwen-AgentWorld-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 23.4 GB; TTFT 17178 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1908.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen-AgentWorld-35B-A3B-MLX-oQ4 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k tokens 1,908 PP tok/s 90.0 TG tok/s 17178 TTFT (ms) 23.4 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/abibnok6",
   "source_url": "https://omlx.ai/benchmarks/performance/abibnok6",
   "tg_tokens": null,
   "tps": 90.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-hs8t9dt3-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: mfcq9sq3; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 233.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-9B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 233.2 18.3 6.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hs8t9dt3",
   "source_url": "https://omlx.ai/benchmarks/performance/hs8t9dt3",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-hs8t9dt3-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: mfcq9sq3; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 212.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-9B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 16k 212.2 16.8 6.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hs8t9dt3",
   "source_url": "https://omlx.ai/benchmarks/performance/hs8t9dt3",
   "tg_tokens": null,
   "tps": 16.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-yan9i4ec-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: d9q934in; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 233.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-9B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k 233.0 19.5 5.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yan9i4ec",
   "source_url": "https://omlx.ai/benchmarks/performance/yan9i4ec",
   "tg_tokens": null,
   "tps": 19.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-yan9i4ec-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: d9q934in; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 239.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-9B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 239.6 19.0 6.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yan9i4ec",
   "source_url": "https://omlx.ai/benchmarks/performance/yan9i4ec",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-t0o4bkv6-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: fd9lfu2q; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2460.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,460 107.6 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t0o4bkv6",
   "source_url": "https://omlx.ai/benchmarks/performance/t0o4bkv6",
   "tg_tokens": null,
   "tps": 107.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-t0o4bkv6-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: fd9lfu2q; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3315.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,315 104.0 20.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t0o4bkv6",
   "source_url": "https://omlx.ai/benchmarks/performance/t0o4bkv6",
   "tg_tokens": null,
   "tps": 104.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-p6qnkfu4-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: 0vyqxqdo, 5ko401m7; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 769.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 769.0 52.5 20.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p6qnkfu4",
   "source_url": "https://omlx.ai/benchmarks/performance/p6qnkfu4",
   "tg_tokens": null,
   "tps": 52.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-p6qnkfu4-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: 0vyqxqdo, 5ko401m7; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 641.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 32k 641.1 40.3 25.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p6qnkfu4",
   "source_url": "https://omlx.ai/benchmarks/performance/p6qnkfu4",
   "tg_tokens": null,
   "tps": 40.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-p6qnkfu4-ctx199680",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.8 GB at this context; duplicate traces of the same session: 0vyqxqdo, 5ko401m7; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 237.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 195k 237.6 13.5 28.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p6qnkfu4",
   "source_url": "https://omlx.ai/benchmarks/performance/p6qnkfu4",
   "tg_tokens": null,
   "tps": 13.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-mo0rrvuv-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: nlo0pmvn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 500.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 500.6 47.7 24.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mo0rrvuv",
   "source_url": "https://omlx.ai/benchmarks/performance/mo0rrvuv",
   "tg_tokens": null,
   "tps": 47.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-mo0rrvuv-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: nlo0pmvn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 325.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 128k 325.6 48.2 26.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mo0rrvuv",
   "source_url": "https://omlx.ai/benchmarks/performance/mo0rrvuv",
   "tg_tokens": null,
   "tps": 48.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-l3263qy6-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: kujswqfg; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 325.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 325.5 23.1 28.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l3263qy6",
   "source_url": "https://omlx.ai/benchmarks/performance/l3263qy6",
   "tg_tokens": null,
   "tps": 23.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-l3263qy6-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.7 GB at this context; duplicate traces of the same session: kujswqfg; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 342.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 342.1 22.8 31.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l3263qy6",
   "source_url": "https://omlx.ai/benchmarks/performance/l3263qy6",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-6hrqaisq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 8vn5m1sq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 323.2 23.2 28.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6hrqaisq",
   "source_url": "https://omlx.ai/benchmarks/performance/6hrqaisq",
   "tg_tokens": null,
   "tps": 23.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-6hrqaisq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.7 GB at this context; duplicate traces of the same session: 8vn5m1sq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 341.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 341.3 22.9 31.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6hrqaisq",
   "source_url": "https://omlx.ai/benchmarks/performance/6hrqaisq",
   "tg_tokens": null,
   "tps": 22.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wkcctbat-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 17.0 GB; TTFT 4024 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1018.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 1,018 PP tok/s 33.1 TG tok/s 4024 TTFT (ms) 17.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wkcctbat",
   "source_url": "https://omlx.ai/benchmarks/performance/wkcctbat",
   "tg_tokens": null,
   "tps": 33.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3l115cea-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: ahcvy2hs, fxq000qy, iszcfp98, ludrsbcp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 548.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 548.5 18.3 28.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3l115cea",
   "source_url": "https://omlx.ai/benchmarks/performance/3l115cea",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3l115cea-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: ahcvy2hs, fxq000qy, iszcfp98, ludrsbcp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 634.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 634.0 18.3 29.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3l115cea",
   "source_url": "https://omlx.ai/benchmarks/performance/3l115cea",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3l115cea-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: ahcvy2hs, fxq000qy, iszcfp98, ludrsbcp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 657.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 657.9 17.4 30.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3l115cea",
   "source_url": "https://omlx.ai/benchmarks/performance/3l115cea",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3l115cea-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 31.3 GB at this context; duplicate traces of the same session: ahcvy2hs, fxq000qy, iszcfp98, ludrsbcp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 620.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 620.7 16.1 31.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3l115cea",
   "source_url": "https://omlx.ai/benchmarks/performance/3l115cea",
   "tg_tokens": null,
   "tps": 16.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3l115cea-ctx65536",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 33.3 GB at this context; duplicate traces of the same session: ahcvy2hs, fxq000qy, iszcfp98, ludrsbcp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 543.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 543.3 14.9 33.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3l115cea",
   "source_url": "https://omlx.ai/benchmarks/performance/3l115cea",
   "tg_tokens": null,
   "tps": 14.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-y8jta483-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: tlw9buo0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 308.1,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 308.1 32.9 16.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y8jta483",
   "source_url": "https://omlx.ai/benchmarks/performance/y8jta483",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-y8jta483-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: tlw9buo0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 335.4,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 335.4 32.2 20.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y8jta483",
   "source_url": "https://omlx.ai/benchmarks/performance/y8jta483",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-8loaq58k-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: jkmaayr1, xb1z8g7h; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 325.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 325.0 78.5 18.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8loaq58k",
   "source_url": "https://omlx.ai/benchmarks/performance/8loaq58k",
   "tg_tokens": null,
   "tps": 78.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-8loaq58k-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: jkmaayr1, xb1z8g7h; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 346.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 346.5 56.8 22.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8loaq58k",
   "source_url": "https://omlx.ai/benchmarks/performance/8loaq58k",
   "tg_tokens": null,
   "tps": 56.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-8loaq58k-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: jkmaayr1, xb1z8g7h; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 340.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 340.7 61.7 22.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8loaq58k",
   "source_url": "https://omlx.ai/benchmarks/performance/8loaq58k",
   "tg_tokens": null,
   "tps": 61.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4fury8x4-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: r25kkwbj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 853.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 853.5 55.4 18.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4fury8x4",
   "source_url": "https://omlx.ai/benchmarks/performance/4fury8x4",
   "tg_tokens": null,
   "tps": 55.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4fury8x4-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: r25kkwbj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 799.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 799.7 83.5 19.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4fury8x4",
   "source_url": "https://omlx.ai/benchmarks/performance/4fury8x4",
   "tg_tokens": null,
   "tps": 83.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-f05hriqy-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: hr5d7th8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 902.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 902.1 53.9 18.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f05hriqy",
   "source_url": "https://omlx.ai/benchmarks/performance/f05hriqy",
   "tg_tokens": null,
   "tps": 53.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-f05hriqy-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: hr5d7th8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 703.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 703.6 80.3 19.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f05hriqy",
   "source_url": "https://omlx.ai/benchmarks/performance/f05hriqy",
   "tg_tokens": null,
   "tps": 80.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-57bphuis-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 2w9z69p6, rnlggqxx, tqjobd37, z7ivtu5b; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 712.3,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 5bit; 1k 712.3 28.2 23.6 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/57bphuis",
   "source_url": "https://omlx.ai/benchmarks/performance/57bphuis",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-57bphuis-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: 2w9z69p6, rnlggqxx, tqjobd37, z7ivtu5b; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 774.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 5bit; 4k 774.9 27.6 24.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/57bphuis",
   "source_url": "https://omlx.ai/benchmarks/performance/57bphuis",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-57bphuis-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: 2w9z69p6, rnlggqxx, tqjobd37, z7ivtu5b; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 776.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 5bit; 8k 776.1 22.3 25.0 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/57bphuis",
   "source_url": "https://omlx.ai/benchmarks/performance/57bphuis",
   "tg_tokens": null,
   "tps": 22.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-57bphuis-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: 2w9z69p6, rnlggqxx, tqjobd37, z7ivtu5b; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 753.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 5bit; 16k 753.7 30.9 25.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/57bphuis",
   "source_url": "https://omlx.ai/benchmarks/performance/57bphuis",
   "tg_tokens": null,
   "tps": 30.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-57bphuis-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: 2w9z69p6, rnlggqxx, tqjobd37, z7ivtu5b; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 706.8,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 5bit; 32k 706.8 24.6 26.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/57bphuis",
   "source_url": "https://omlx.ai/benchmarks/performance/57bphuis",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-61g01l05-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 25.5 GB; TTFT 21755 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 753.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 5bit; 16k tokens 753.1 PP tok/s 30.1 TG tok/s 21755 TTFT (ms) 25.5 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/61g01l05",
   "source_url": "https://omlx.ai/benchmarks/performance/61g01l05",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1jjgxpaj-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.7 GB at this context; duplicate traces of the same session: mes9eb18, ujfcaepa; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 321.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 321.3 52.5 33.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1jjgxpaj",
   "source_url": "https://omlx.ai/benchmarks/performance/1jjgxpaj",
   "tg_tokens": null,
   "tps": 52.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1jjgxpaj-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: mes9eb18, ujfcaepa; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 333.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 333.2 42.2 37.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1jjgxpaj",
   "source_url": "https://omlx.ai/benchmarks/performance/1jjgxpaj",
   "tg_tokens": null,
   "tps": 42.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1jjgxpaj-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.8 GB at this context; duplicate traces of the same session: mes9eb18, ujfcaepa; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 328.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 328.4 46.9 37.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1jjgxpaj",
   "source_url": "https://omlx.ai/benchmarks/performance/1jjgxpaj",
   "tg_tokens": null,
   "tps": 46.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xg6lggck-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 195k context; peak memory 70.3 GB; TTFT 201745 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 991.3,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 3bit; 195k tokens 991.3 PP tok/s 44.2 TG tok/s 201745 TTFT (ms) 70.3 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xg6lggck",
   "source_url": "https://omlx.ai/benchmarks/performance/xg6lggck",
   "tg_tokens": null,
   "tps": 44.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-w4ifr70q-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.5 GB at this context; duplicate traces of the same session: h3hv8k30; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1043.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,043 83.6 70.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w4ifr70q",
   "source_url": "https://omlx.ai/benchmarks/performance/w4ifr70q",
   "tg_tokens": null,
   "tps": 83.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-w4ifr70q-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: h3hv8k30; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1968.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,968 71.0 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w4ifr70q",
   "source_url": "https://omlx.ai/benchmarks/performance/w4ifr70q",
   "tg_tokens": null,
   "tps": 71.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-pnjbiuaq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1421.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k 1,421 68.1 19.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pnjbiuaq",
   "source_url": "https://omlx.ai/benchmarks/performance/pnjbiuaq",
   "tg_tokens": null,
   "tps": 68.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-pnjbiuaq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1522.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 4k 1,522 79.0 20.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pnjbiuaq",
   "source_url": "https://omlx.ai/benchmarks/performance/pnjbiuaq",
   "tg_tokens": null,
   "tps": 79.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-f4uz5xtz-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: cfi2mw72; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1415.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k 1,415 77.5 19.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f4uz5xtz",
   "source_url": "https://omlx.ai/benchmarks/performance/f4uz5xtz",
   "tg_tokens": null,
   "tps": 77.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-f4uz5xtz-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: cfi2mw72; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1490.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 8k 1,490 93.4 21.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f4uz5xtz",
   "source_url": "https://omlx.ai/benchmarks/performance/f4uz5xtz",
   "tg_tokens": null,
   "tps": 93.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-f4uz5xtz-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: cfi2mw72; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1332.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 32k 1,332 86.1 22.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f4uz5xtz",
   "source_url": "https://omlx.ai/benchmarks/performance/f4uz5xtz",
   "tg_tokens": null,
   "tps": 86.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-6b1h929a-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.1 GB at this context; duplicate traces of the same session: yw9dd0g8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2369.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,369 83.4 14.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6b1h929a",
   "source_url": "https://omlx.ai/benchmarks/performance/6b1h929a",
   "tg_tokens": null,
   "tps": 83.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-6b1h929a-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: yw9dd0g8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2114.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,114 79.5 14.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6b1h929a",
   "source_url": "https://omlx.ai/benchmarks/performance/6b1h929a",
   "tg_tokens": null,
   "tps": 79.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-292n5kmx-ctx4096",
   "model": "Gemma-4-31B",
   "notes": "oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 5gysn927; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 174.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31b-it-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 174.5 24.8 19.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/292n5kmx",
   "source_url": "https://omlx.ai/benchmarks/performance/292n5kmx",
   "tg_tokens": null,
   "tps": 24.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-292n5kmx-ctx16384",
   "model": "Gemma-4-31B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: 5gysn927; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 171.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31b-it-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 16k 171.6 21.6 20.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/292n5kmx",
   "source_url": "https://omlx.ai/benchmarks/performance/292n5kmx",
   "tg_tokens": null,
   "tps": 21.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-8jv9ajn8-ctx1024",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: m1vcxg9p; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 65.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "AREX-2-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k 65.6 6.6 15.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8jv9ajn8",
   "source_url": "https://omlx.ai/benchmarks/performance/8jv9ajn8",
   "tg_tokens": null,
   "tps": 6.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-8jv9ajn8-ctx4096",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: m1vcxg9p; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 67.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "AREX-2-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 67.3 6.5 17.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8jv9ajn8",
   "source_url": "https://omlx.ai/benchmarks/performance/8jv9ajn8",
   "tg_tokens": null,
   "tps": 6.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1qp9vo7l-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 327.1 GB; TTFT 60719 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 539.7,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "GLM-5.3-Flash-oQ8-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 32k tokens 539.7 PP tok/s 50.6 TG tok/s 60719 TTFT (ms) 327.1 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1qp9vo7l",
   "source_url": "https://omlx.ai/benchmarks/performance/1qp9vo7l",
   "tg_tokens": null,
   "tps": 50.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-d7skb1fz-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: lz63juq5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 980.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 980.0 110.0 13.4 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d7skb1fz",
   "source_url": "https://omlx.ai/benchmarks/performance/d7skb1fz",
   "tg_tokens": null,
   "tps": 110.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-d7skb1fz-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.9 GB at this context; duplicate traces of the same session: lz63juq5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 960.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 16k 960.3 87.8 13.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d7skb1fz",
   "source_url": "https://omlx.ai/benchmarks/performance/d7skb1fz",
   "tg_tokens": null,
   "tps": 87.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-72qbcpih-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 4k context; peak memory 16.4 GB; TTFT 2680 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1529.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "LFM2.5-24B-A2B-MLX-5bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 5bit; 4k tokens 1,529 PP tok/s 112.3 TG tok/s 2680 TTFT (ms) 16.4 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/72qbcpih",
   "source_url": "https://omlx.ai/benchmarks/performance/72qbcpih",
   "tg_tokens": null,
   "tps": 112.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2at46036-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "bf16 MTP MLX build; oMLX community benchmark trace, single run at 16k context; peak memory 25.1 GB; TTFT 22461 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 729.5,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-Uncensored-oQ5e-bf16-mtp-MLX on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 5bit; 16k tokens 729.5 PP tok/s 18.4 TG tok/s 22461 TTFT (ms) 25.1 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2at46036",
   "source_url": "https://omlx.ai/benchmarks/performance/2at46036",
   "tg_tokens": null,
   "tps": 18.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-q0n6ix4w-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "bf16 MTP MLX build; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: 3hc1knii, bhsntmvz, y407u591; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 670.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-Uncensored-oQ5e-bf16-mtp-MLX on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 5bit; 1k 670.0 43.0 23.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q0n6ix4w",
   "source_url": "https://omlx.ai/benchmarks/performance/q0n6ix4w",
   "tg_tokens": null,
   "tps": 43.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-q0n6ix4w-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "bf16 MTP MLX build; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: 3hc1knii, bhsntmvz, y407u591; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 747.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-Uncensored-oQ5e-bf16-mtp-MLX on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 5bit; 4k 747.7 51.5 24.3 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q0n6ix4w",
   "source_url": "https://omlx.ai/benchmarks/performance/q0n6ix4w",
   "tg_tokens": null,
   "tps": 51.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-q0n6ix4w-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "bf16 MTP MLX build; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: 3hc1knii, bhsntmvz, y407u591; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 748.6,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-Uncensored-oQ5e-bf16-mtp-MLX on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 5bit; 8k 748.6 29.2 24.6 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q0n6ix4w",
   "source_url": "https://omlx.ai/benchmarks/performance/q0n6ix4w",
   "tg_tokens": null,
   "tps": 29.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-q0n6ix4w-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "bf16 MTP MLX build; oMLX community benchmark trace; peak memory 25.2 GB at this context; duplicate traces of the same session: 3hc1knii, bhsntmvz, y407u591; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 732.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-Uncensored-oQ5e-bf16-mtp-MLX on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 5bit; 16k 732.4 34.7 25.2 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q0n6ix4w",
   "source_url": "https://omlx.ai/benchmarks/performance/q0n6ix4w",
   "tg_tokens": null,
   "tps": 34.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-qi8of1d3-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: hn3cnq3v; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 221.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Uncensored-oQ8e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 8bit; 1k 221.5 21.1 28.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qi8of1d3",
   "source_url": "https://omlx.ai/benchmarks/performance/qi8of1d3",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-qi8of1d3-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 31.9 GB at this context; duplicate traces of the same session: hn3cnq3v; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 227.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Uncensored-oQ8e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 8bit; 4k 227.2 20.6 31.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qi8of1d3",
   "source_url": "https://omlx.ai/benchmarks/performance/qi8of1d3",
   "tg_tokens": null,
   "tps": 20.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bs2jvqe5-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.9 GB at this context; duplicate traces of the same session: 0gqio9jd, 1zb9geyg, h7urmwca, mh6wc114, qikjobgn, vzp21c7h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 825.3,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Qwen3.8-27B-mxfp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 825.3 12.5 28.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bs2jvqe5",
   "source_url": "https://omlx.ai/benchmarks/performance/bs2jvqe5",
   "tg_tokens": null,
   "tps": 12.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bs2jvqe5-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 30.0 GB at this context; duplicate traces of the same session: 0gqio9jd, 1zb9geyg, h7urmwca, mh6wc114, qikjobgn, vzp21c7h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 946.7,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Qwen3.8-27B-mxfp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 946.7 12.1 30.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bs2jvqe5",
   "source_url": "https://omlx.ai/benchmarks/performance/bs2jvqe5",
   "tg_tokens": null,
   "tps": 12.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bs2jvqe5-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: 0gqio9jd, 1zb9geyg, h7urmwca, mh6wc114, qikjobgn, vzp21c7h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 946.4,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Qwen3.8-27B-mxfp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 946.4 10.2 30.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bs2jvqe5",
   "source_url": "https://omlx.ai/benchmarks/performance/bs2jvqe5",
   "tg_tokens": null,
   "tps": 10.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bs2jvqe5-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 30.8 GB at this context; duplicate traces of the same session: 0gqio9jd, 1zb9geyg, h7urmwca, mh6wc114, qikjobgn, vzp21c7h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 922.9,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Qwen3.8-27B-mxfp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 922.9 12.4 30.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bs2jvqe5",
   "source_url": "https://omlx.ai/benchmarks/performance/bs2jvqe5",
   "tg_tokens": null,
   "tps": 12.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bs2jvqe5-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.8 GB at this context; duplicate traces of the same session: 0gqio9jd, 1zb9geyg, h7urmwca, mh6wc114, qikjobgn, vzp21c7h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 868.2,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Qwen3.8-27B-mxfp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 868.2 12.2 31.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bs2jvqe5",
   "source_url": "https://omlx.ai/benchmarks/performance/bs2jvqe5",
   "tg_tokens": null,
   "tps": 12.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bs2jvqe5-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: 0gqio9jd, 1zb9geyg, h7urmwca, mh6wc114, qikjobgn, vzp21c7h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 770.3,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Qwen3.8-27B-mxfp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 770.3 13.1 33.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bs2jvqe5",
   "source_url": "https://omlx.ai/benchmarks/performance/bs2jvqe5",
   "tg_tokens": null,
   "tps": 13.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bs2jvqe5-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 37.8 GB at this context; duplicate traces of the same session: 0gqio9jd, 1zb9geyg, h7urmwca, mh6wc114, qikjobgn, vzp21c7h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 624.3,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Qwen3.8-27B-mxfp8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 624.3 13.1 37.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bs2jvqe5",
   "source_url": "https://omlx.ai/benchmarks/performance/bs2jvqe5",
   "tg_tokens": null,
   "tps": 13.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-rykggym7-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: d26ynjw6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 307.9,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 307.9 33.1 16.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rykggym7",
   "source_url": "https://omlx.ai/benchmarks/performance/rykggym7",
   "tg_tokens": null,
   "tps": 33.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-rykggym7-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: d26ynjw6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 332.9,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 332.9 32.3 20.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rykggym7",
   "source_url": "https://omlx.ai/benchmarks/performance/rykggym7",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zy17h3zv-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 19.9 GB; TTFT 14969 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1095.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k tokens 1,095 PP tok/s 34.1 TG tok/s 14969 TTFT (ms) 19.9 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zy17h3zv",
   "source_url": "https://omlx.ai/benchmarks/performance/zy17h3zv",
   "tg_tokens": null,
   "tps": 34.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-64g4d2wv-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 7u8ue4pv, 8cc70p4k, cfkzdenz, yhx5qo26; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 888.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 888.9 48.4 17.6 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/64g4d2wv",
   "source_url": "https://omlx.ai/benchmarks/performance/64g4d2wv",
   "tg_tokens": null,
   "tps": 48.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-64g4d2wv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 7u8ue4pv, 8cc70p4k, cfkzdenz, yhx5qo26; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1082.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 1,082 57.4 19.1 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/64g4d2wv",
   "source_url": "https://omlx.ai/benchmarks/performance/64g4d2wv",
   "tg_tokens": null,
   "tps": 57.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-64g4d2wv-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 7u8ue4pv, 8cc70p4k, cfkzdenz, yhx5qo26; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1131.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 1,131 33.0 19.4 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/64g4d2wv",
   "source_url": "https://omlx.ai/benchmarks/performance/64g4d2wv",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-64g4d2wv-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 7u8ue4pv, 8cc70p4k, cfkzdenz, yhx5qo26; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1100.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 1,100 36.9 19.9 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/64g4d2wv",
   "source_url": "https://omlx.ai/benchmarks/performance/64g4d2wv",
   "tg_tokens": null,
   "tps": 36.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-64g4d2wv-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 7u8ue4pv, 8cc70p4k, cfkzdenz, yhx5qo26; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 996.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 32k 996.8 28.6 21.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/64g4d2wv",
   "source_url": "https://omlx.ai/benchmarks/performance/64g4d2wv",
   "tg_tokens": null,
   "tps": 28.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ree5hqvi-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 94u463ve; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 978.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 978.3 63.9 18.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ree5hqvi",
   "source_url": "https://omlx.ai/benchmarks/performance/ree5hqvi",
   "tg_tokens": null,
   "tps": 63.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ree5hqvi-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 94u463ve; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 832.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 832.5 83.3 19.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ree5hqvi",
   "source_url": "https://omlx.ai/benchmarks/performance/ree5hqvi",
   "tg_tokens": null,
   "tps": 83.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-vkyrvf3m-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 9chqdl1i, pg1ac682; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 675.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 675.9 31.1 19.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vkyrvf3m",
   "source_url": "https://omlx.ai/benchmarks/performance/vkyrvf3m",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-vkyrvf3m-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 9chqdl1i, pg1ac682; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 683.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 683.0 50.7 19.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vkyrvf3m",
   "source_url": "https://omlx.ai/benchmarks/performance/vkyrvf3m",
   "tg_tokens": null,
   "tps": 50.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-vkyrvf3m-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: 9chqdl1i, pg1ac682; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 583.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 583.4 73.7 21.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vkyrvf3m",
   "source_url": "https://omlx.ai/benchmarks/performance/vkyrvf3m",
   "tg_tokens": null,
   "tps": 73.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1se37h79-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: hh86eg2u; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 597.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 597.0 41.8 17.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1se37h79",
   "source_url": "https://omlx.ai/benchmarks/performance/1se37h79",
   "tg_tokens": null,
   "tps": 41.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1se37h79-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: hh86eg2u; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 689.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 689.0 28.4 19.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1se37h79",
   "source_url": "https://omlx.ai/benchmarks/performance/1se37h79",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-dk92zzna-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 39.2 GB; TTFT 58876 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1113.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 64k tokens 1,113 PP tok/s 65.2 TG tok/s 58876 TTFT (ms) 39.2 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dk92zzna",
   "source_url": "https://omlx.ai/benchmarks/performance/dk92zzna",
   "tg_tokens": null,
   "tps": 65.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-vc8pmdk3-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 39.2 GB; TTFT 58861 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1113.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 64k tokens 1,113 PP tok/s 69.2 TG tok/s 58861 TTFT (ms) 39.2 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vc8pmdk3",
   "source_url": "https://omlx.ai/benchmarks/performance/vc8pmdk3",
   "tg_tokens": null,
   "tps": 69.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-78tadjiv-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1159.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 1k 1,159 73.8 33.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/78tadjiv",
   "source_url": "https://omlx.ai/benchmarks/performance/78tadjiv",
   "tg_tokens": null,
   "tps": 73.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-78tadjiv-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1403.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 8k 1,403 85.8 35.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/78tadjiv",
   "source_url": "https://omlx.ai/benchmarks/performance/78tadjiv",
   "tg_tokens": null,
   "tps": 85.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-78tadjiv-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.8 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1274.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 32k 1,274 119.7 36.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/78tadjiv",
   "source_url": "https://omlx.ai/benchmarks/performance/78tadjiv",
   "tg_tokens": null,
   "tps": 119.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-bah1dvdn-ctx4096",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: nrsvaiwj; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1635.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ6e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 6bit; 4k 1,635 61.1 29.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bah1dvdn",
   "source_url": "https://omlx.ai/benchmarks/performance/bah1dvdn",
   "tg_tokens": null,
   "tps": 61.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-bah1dvdn-ctx16384",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 30.6 GB at this context; duplicate traces of the same session: nrsvaiwj; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1397.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ6e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 6bit; 16k 1,397 57.0 30.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bah1dvdn",
   "source_url": "https://omlx.ai/benchmarks/performance/bah1dvdn",
   "tg_tokens": null,
   "tps": 57.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-wrlrr7pp-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "REAP288 MTP MLX build; oMLX community benchmark trace, single run at 1k context; peak memory 44.5 GB; TTFT 4369 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 234.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-Uncensored-REAP288-MTP-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k tokens 234.4 PP tok/s 30.0 TG tok/s 4369 TTFT (ms) 44.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wrlrr7pp",
   "source_url": "https://omlx.ai/benchmarks/performance/wrlrr7pp",
   "tg_tokens": null,
   "tps": 30.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ydmlybh6-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 21.4 GB; TTFT 21721 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 754.3,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-bf16-mtp-MLX on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 5bit; 16k tokens 754.3 PP tok/s 20.6 TG tok/s 21721 TTFT (ms) 21.4 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ydmlybh6",
   "source_url": "https://omlx.ai/benchmarks/performance/ydmlybh6",
   "tg_tokens": null,
   "tps": 20.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-06l31k4o-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 22.4 GB; TTFT 46251 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 708.5,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-bf16-mtp-MLX on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 5bit; 32k tokens 708.5 PP tok/s 19.4 TG tok/s 46251 TTFT (ms) 22.4 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/06l31k4o",
   "source_url": "https://omlx.ai/benchmarks/performance/06l31k4o",
   "tg_tokens": null,
   "tps": 19.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lfkyjhtq-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: 46bjaryy, 6y5khufk, indh3l4v; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 666.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-bf16-mtp-MLX on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 5bit; 1k 666.2 42.0 23.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lfkyjhtq",
   "source_url": "https://omlx.ai/benchmarks/performance/lfkyjhtq",
   "tg_tokens": null,
   "tps": 42.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lfkyjhtq-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: 46bjaryy, 6y5khufk, indh3l4v; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 747.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-bf16-mtp-MLX on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 5bit; 4k 747.2 49.2 24.3 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lfkyjhtq",
   "source_url": "https://omlx.ai/benchmarks/performance/lfkyjhtq",
   "tg_tokens": null,
   "tps": 49.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lfkyjhtq-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: 46bjaryy, 6y5khufk, indh3l4v; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 744.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-bf16-mtp-MLX on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 5bit; 8k 744.2 31.8 24.6 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lfkyjhtq",
   "source_url": "https://omlx.ai/benchmarks/performance/lfkyjhtq",
   "tg_tokens": null,
   "tps": 31.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lfkyjhtq-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.2 GB at this context; duplicate traces of the same session: 46bjaryy, 6y5khufk, indh3l4v; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 731.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-bf16-mtp-MLX on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 5bit; 16k 731.4 36.3 25.2 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lfkyjhtq",
   "source_url": "https://omlx.ai/benchmarks/performance/lfkyjhtq",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-8qr47zsn-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 99.1 GB at this context; duplicate traces of the same session: 0rp0ktfv; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 908.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 908.6 70.3 99.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8qr47zsn",
   "source_url": "https://omlx.ai/benchmarks/performance/8qr47zsn",
   "tg_tokens": null,
   "tps": 70.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-8qr47zsn-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 108.7 GB at this context; duplicate traces of the same session: 0rp0ktfv; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1015.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 195k 1,015 55.8 108.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8qr47zsn",
   "source_url": "https://omlx.ai/benchmarks/performance/8qr47zsn",
   "tg_tokens": null,
   "tps": 55.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-enpjd933-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 99.1 GB at this context; duplicate traces of the same session: 9x5bg1fq, lpk3q2pk; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 837.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 837.2 80.1 99.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/enpjd933",
   "source_url": "https://omlx.ai/benchmarks/performance/enpjd933",
   "tg_tokens": null,
   "tps": 80.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-enpjd933-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 99.6 GB at this context; duplicate traces of the same session: 9x5bg1fq, lpk3q2pk; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1104.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 1,104 74.0 99.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/enpjd933",
   "source_url": "https://omlx.ai/benchmarks/performance/enpjd933",
   "tg_tokens": null,
   "tps": 74.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-enpjd933-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: 9x5bg1fq, lpk3q2pk; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1062.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 1,062 44.1 101.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/enpjd933",
   "source_url": "https://omlx.ai/benchmarks/performance/enpjd933",
   "tg_tokens": null,
   "tps": 44.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-z67upt9p-ctx4096",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 7.8 GB at this context; duplicate traces of the same session: dyvb0o76, gndf9j8n; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 858.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 858.2 34.9 7.8 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z67upt9p",
   "source_url": "https://omlx.ai/benchmarks/performance/z67upt9p",
   "tg_tokens": null,
   "tps": 34.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-z67upt9p-ctx16384",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 8.3 GB at this context; duplicate traces of the same session: dyvb0o76, gndf9j8n; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 804.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 16k 804.3 34.2 8.3 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z67upt9p",
   "source_url": "https://omlx.ai/benchmarks/performance/z67upt9p",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-z67upt9p-ctx32768",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 9.0 GB at this context; duplicate traces of the same session: dyvb0o76, gndf9j8n; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 729.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 32k 729.0 33.2 9.0 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z67upt9p",
   "source_url": "https://omlx.ai/benchmarks/performance/z67upt9p",
   "tg_tokens": null,
   "tps": 33.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-y4onyzsr-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: 4x7iexzb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2571.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,571 99.4 14.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y4onyzsr",
   "source_url": "https://omlx.ai/benchmarks/performance/y4onyzsr",
   "tg_tokens": null,
   "tps": 99.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-y4onyzsr-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 15.4 GB at this context; duplicate traces of the same session: 4x7iexzb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2461.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,461 77.4 15.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y4onyzsr",
   "source_url": "https://omlx.ai/benchmarks/performance/y4onyzsr",
   "tg_tokens": null,
   "tps": 77.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ubwr1s2a-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: lkfwvzpy; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2477.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,477 50.5 14.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ubwr1s2a",
   "source_url": "https://omlx.ai/benchmarks/performance/ubwr1s2a",
   "tg_tokens": null,
   "tps": 50.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ubwr1s2a-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 15.4 GB at this context; duplicate traces of the same session: lkfwvzpy; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1826.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,826 103.0 15.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ubwr1s2a",
   "source_url": "https://omlx.ai/benchmarks/performance/ubwr1s2a",
   "tg_tokens": null,
   "tps": 103.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-jyg7829b-ctx32768",
   "model": "gpt-oss 20B",
   "notes": "bf16 MLX build; oMLX community benchmark trace, single run at 32k context; peak memory 15.8 GB; TTFT 108569 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 301.8,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "gpt-oss-20b-oQ5e-bf16-MLX on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 5bit; 32k tokens 301.8 PP tok/s 22.4 TG tok/s 108569 TTFT (ms) 15.8 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jyg7829b",
   "source_url": "https://omlx.ai/benchmarks/performance/jyg7829b",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-lwn4i37s-ctx1024",
   "model": "GLM-4.7-Flash",
   "notes": "oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: t1h3fn7l; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 292.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-4.7-Flash-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 292.6 33.7 16.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lwn4i37s",
   "source_url": "https://omlx.ai/benchmarks/performance/lwn4i37s",
   "tg_tokens": null,
   "tps": 33.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-lwn4i37s-ctx4096",
   "model": "GLM-4.7-Flash",
   "notes": "oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: t1h3fn7l; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 260.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-4.7-Flash-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 260.2 28.6 17.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lwn4i37s",
   "source_url": "https://omlx.ai/benchmarks/performance/lwn4i37s",
   "tg_tokens": null,
   "tps": 28.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-s6i5qasz-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: r8r3ozgh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1060.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "LFM2.5-24B-A2B-MLX-5bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 5bit; 4k 1,060 89.7 16.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s6i5qasz",
   "source_url": "https://omlx.ai/benchmarks/performance/s6i5qasz",
   "tg_tokens": null,
   "tps": 89.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-s6i5qasz-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 16.7 GB at this context; duplicate traces of the same session: r8r3ozgh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1165.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "LFM2.5-24B-A2B-MLX-5bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 5bit; 16k 1,165 78.3 16.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s6i5qasz",
   "source_url": "https://omlx.ai/benchmarks/performance/s6i5qasz",
   "tg_tokens": null,
   "tps": 78.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-g8w9knwj-ctx4096",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 8.8 GB at this context; duplicate traces of the same session: wkyuo803, xo94l27u; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3638.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-oQ5e on M5 Pro (20c); M5 Pro (20c) \u00b7 24 GB \u00b7 5bit; 4k 3,638 112.8 8.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g8w9knwj",
   "source_url": "https://omlx.ai/benchmarks/performance/g8w9knwj",
   "tg_tokens": null,
   "tps": 112.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-g8w9knwj-ctx8192",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 8.8 GB at this context; duplicate traces of the same session: wkyuo803, xo94l27u; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3653.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-oQ5e on M5 Pro (20c); M5 Pro (20c) \u00b7 24 GB \u00b7 5bit; 8k 3,653 108.0 8.8 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g8w9knwj",
   "source_url": "https://omlx.ai/benchmarks/performance/g8w9knwj",
   "tg_tokens": null,
   "tps": 108.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-g8w9knwj-ctx16384",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 8.9 GB at this context; duplicate traces of the same session: wkyuo803, xo94l27u; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3463.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-oQ5e on M5 Pro (20c); M5 Pro (20c) \u00b7 24 GB \u00b7 5bit; 16k 3,463 105.0 8.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g8w9knwj",
   "source_url": "https://omlx.ai/benchmarks/performance/g8w9knwj",
   "tg_tokens": null,
   "tps": 105.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-d3kv8sif-ctx1024",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 2.6 GB at this context; duplicate traces of the same session: kj7vux8r; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 654.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM5-2B-OptiQ-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 16 GB \u00b7 4bit; 1k 654.9 74.2 2.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d3kv8sif",
   "source_url": "https://omlx.ai/benchmarks/performance/d3kv8sif",
   "tg_tokens": null,
   "tps": 74.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-d3kv8sif-ctx4096",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 2.7 GB at this context; duplicate traces of the same session: kj7vux8r; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 610.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM5-2B-OptiQ-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 16 GB \u00b7 4bit; 4k 610.0 64.3 2.7 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d3kv8sif",
   "source_url": "https://omlx.ai/benchmarks/performance/d3kv8sif",
   "tg_tokens": null,
   "tps": 64.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-e3lhje89-ctx131072",
   "model": "Ornith-1.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 16.5 GB; TTFT 1166521 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 112.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-9B-oQ6e-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 6bit; 128k tokens 112.4 PP tok/s 12.3 TG tok/s 1166521 TTFT (ms) 16.5 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e3lhje89",
   "source_url": "https://omlx.ai/benchmarks/performance/e3lhje89",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-d5e1sk5u-ctx1024",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 12.7 GB; TTFT 2060 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 497.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 497.2 PP tok/s 33.8 TG tok/s 2060 TTFT (ms) 12.7 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d5e1sk5u",
   "source_url": "https://omlx.ai/benchmarks/performance/d5e1sk5u",
   "tg_tokens": null,
   "tps": 33.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-xdq23nqp-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: 2cl3k3v8, 4iig60zv, llhtuahu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 496.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 496.5 18.1 15.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xdq23nqp",
   "source_url": "https://omlx.ai/benchmarks/performance/xdq23nqp",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-xdq23nqp-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 2cl3k3v8, 4iig60zv, llhtuahu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 516.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 516.9 17.8 17.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xdq23nqp",
   "source_url": "https://omlx.ai/benchmarks/performance/xdq23nqp",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-xdq23nqp-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 2cl3k3v8, 4iig60zv, llhtuahu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 510.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 510.9 17.4 17.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xdq23nqp",
   "source_url": "https://omlx.ai/benchmarks/performance/xdq23nqp",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-xdq23nqp-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 2cl3k3v8, 4iig60zv, llhtuahu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 461.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 461.1 16.8 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xdq23nqp",
   "source_url": "https://omlx.ai/benchmarks/performance/xdq23nqp",
   "tg_tokens": null,
   "tps": 16.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3t2j0bre-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: av8li558, o3thn1x5, yqfmkiv6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 496.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 496.2 18.0 15.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3t2j0bre",
   "source_url": "https://omlx.ai/benchmarks/performance/3t2j0bre",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3t2j0bre-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: av8li558, o3thn1x5, yqfmkiv6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 516.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 516.2 17.7 17.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3t2j0bre",
   "source_url": "https://omlx.ai/benchmarks/performance/3t2j0bre",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3t2j0bre-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: av8li558, o3thn1x5, yqfmkiv6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 500.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 500.7 17.4 17.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3t2j0bre",
   "source_url": "https://omlx.ai/benchmarks/performance/3t2j0bre",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3t2j0bre-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: av8li558, o3thn1x5, yqfmkiv6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 456.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 456.3 16.7 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3t2j0bre",
   "source_url": "https://omlx.ai/benchmarks/performance/3t2j0bre",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-01z87gj7-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.8 GB at this context; duplicate traces of the same session: tzrb2p52; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 272.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 272.9 18.3 31.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/01z87gj7",
   "source_url": "https://omlx.ai/benchmarks/performance/01z87gj7",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-01z87gj7-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.0 GB at this context; duplicate traces of the same session: tzrb2p52; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 260.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 260.0 17.5 33.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/01z87gj7",
   "source_url": "https://omlx.ai/benchmarks/performance/01z87gj7",
   "tg_tokens": null,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fnj4uw5u-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 1mqwrq0i, 3g46jqul, 8yhon1ac, pc8a777r; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 912.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 912.4 43.6 17.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fnj4uw5u",
   "source_url": "https://omlx.ai/benchmarks/performance/fnj4uw5u",
   "tg_tokens": null,
   "tps": 43.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fnj4uw5u-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 1mqwrq0i, 3g46jqul, 8yhon1ac, pc8a777r; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1082.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 1,082 59.6 19.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fnj4uw5u",
   "source_url": "https://omlx.ai/benchmarks/performance/fnj4uw5u",
   "tg_tokens": null,
   "tps": 59.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fnj4uw5u-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 1mqwrq0i, 3g46jqul, 8yhon1ac, pc8a777r; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1133.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 1,133 30.2 19.4 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fnj4uw5u",
   "source_url": "https://omlx.ai/benchmarks/performance/fnj4uw5u",
   "tg_tokens": null,
   "tps": 30.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fnj4uw5u-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 1mqwrq0i, 3g46jqul, 8yhon1ac, pc8a777r; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1101.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 1,101 36.5 19.9 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fnj4uw5u",
   "source_url": "https://omlx.ai/benchmarks/performance/fnj4uw5u",
   "tg_tokens": null,
   "tps": 36.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fnj4uw5u-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 1mqwrq0i, 3g46jqul, 8yhon1ac, pc8a777r; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1003.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 32k 1,003 25.8 21.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fnj4uw5u",
   "source_url": "https://omlx.ai/benchmarks/performance/fnj4uw5u",
   "tg_tokens": null,
   "tps": 25.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qju7tdmo-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: dg1oirgi; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 907.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 907.0 41.0 17.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qju7tdmo",
   "source_url": "https://omlx.ai/benchmarks/performance/qju7tdmo",
   "tg_tokens": null,
   "tps": 41.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qju7tdmo-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: dg1oirgi; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1083.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 1,083 47.9 19.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qju7tdmo",
   "source_url": "https://omlx.ai/benchmarks/performance/qju7tdmo",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qju7tdmo-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: dg1oirgi; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1133.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 1,133 31.7 19.4 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qju7tdmo",
   "source_url": "https://omlx.ai/benchmarks/performance/qju7tdmo",
   "tg_tokens": null,
   "tps": 31.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qju7tdmo-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: dg1oirgi; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1101.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 1,101 37.9 19.9 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qju7tdmo",
   "source_url": "https://omlx.ai/benchmarks/performance/qju7tdmo",
   "tg_tokens": null,
   "tps": 37.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qju7tdmo-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: dg1oirgi; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1003.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 32k 1,003 28.0 21.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qju7tdmo",
   "source_url": "https://omlx.ai/benchmarks/performance/qju7tdmo",
   "tg_tokens": null,
   "tps": 28.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-c7ll6gei-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: u8oz6o1i; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 466.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 466.7 27.4 18.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c7ll6gei",
   "source_url": "https://omlx.ai/benchmarks/performance/c7ll6gei",
   "tg_tokens": null,
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-c7ll6gei-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: u8oz6o1i; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 426.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 426.5 44.1 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c7ll6gei",
   "source_url": "https://omlx.ai/benchmarks/performance/c7ll6gei",
   "tg_tokens": null,
   "tps": 44.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-14uea0fw-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: rzn85fzj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 462.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 462.2 31.2 18.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/14uea0fw",
   "source_url": "https://omlx.ai/benchmarks/performance/14uea0fw",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-14uea0fw-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: rzn85fzj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 437.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 437.3 41.4 19.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/14uea0fw",
   "source_url": "https://omlx.ai/benchmarks/performance/14uea0fw",
   "tg_tokens": null,
   "tps": 41.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-t4in5fdi-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: g3gk8gr2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 534.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 534.9 37.1 17.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t4in5fdi",
   "source_url": "https://omlx.ai/benchmarks/performance/t4in5fdi",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-t4in5fdi-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: g3gk8gr2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 688.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 688.8 34.7 19.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t4in5fdi",
   "source_url": "https://omlx.ai/benchmarks/performance/t4in5fdi",
   "tg_tokens": null,
   "tps": 34.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-slzf7f50-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: rl11aale; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 533.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 533.4 28.2 17.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/slzf7f50",
   "source_url": "https://omlx.ai/benchmarks/performance/slzf7f50",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-slzf7f50-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: rl11aale; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 626.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 626.5 28.3 19.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/slzf7f50",
   "source_url": "https://omlx.ai/benchmarks/performance/slzf7f50",
   "tg_tokens": null,
   "tps": 28.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-a8scyyya-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 82wl4usn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 479.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 479.9 34.5 17.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a8scyyya",
   "source_url": "https://omlx.ai/benchmarks/performance/a8scyyya",
   "tg_tokens": null,
   "tps": 34.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-a8scyyya-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 82wl4usn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 500.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 500.0 26.4 19.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a8scyyya",
   "source_url": "https://omlx.ai/benchmarks/performance/a8scyyya",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3hqro6mn-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 3h4ji0bt; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 480.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 480.9 33.8 17.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3hqro6mn",
   "source_url": "https://omlx.ai/benchmarks/performance/3hqro6mn",
   "tg_tokens": null,
   "tps": 33.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3hqro6mn-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 3h4ji0bt; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 489.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 489.5 34.4 19.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3hqro6mn",
   "source_url": "https://omlx.ai/benchmarks/performance/3hqro6mn",
   "tg_tokens": null,
   "tps": 34.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-uqf8s3qb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: 92amw58d; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1159.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 1k 1,159 73.8 33.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uqf8s3qb",
   "source_url": "https://omlx.ai/benchmarks/performance/uqf8s3qb",
   "tg_tokens": null,
   "tps": 73.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-uqf8s3qb-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: 92amw58d; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1403.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 8k 1,403 85.8 35.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uqf8s3qb",
   "source_url": "https://omlx.ai/benchmarks/performance/uqf8s3qb",
   "tg_tokens": null,
   "tps": 85.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-uqf8s3qb-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.8 GB at this context; duplicate traces of the same session: 92amw58d; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1274.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 32k 1,274 119.7 36.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uqf8s3qb",
   "source_url": "https://omlx.ai/benchmarks/performance/uqf8s3qb",
   "tg_tokens": null,
   "tps": 119.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-4gvi7tji-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: ljv42m41, uhu9r8qm; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1149.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 1k 1,149 81.8 33.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4gvi7tji",
   "source_url": "https://omlx.ai/benchmarks/performance/4gvi7tji",
   "tg_tokens": null,
   "tps": 81.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-4gvi7tji-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: ljv42m41, uhu9r8qm; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1402.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 8k 1,402 74.1 35.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4gvi7tji",
   "source_url": "https://omlx.ai/benchmarks/performance/4gvi7tji",
   "tg_tokens": null,
   "tps": 74.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-4gvi7tji-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.8 GB at this context; duplicate traces of the same session: ljv42m41, uhu9r8qm; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1274.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 32k 1,274 117.8 36.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4gvi7tji",
   "source_url": "https://omlx.ai/benchmarks/performance/4gvi7tji",
   "tg_tokens": null,
   "tps": 117.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jwezrzxz-ctx1024",
   "model": "Qwen3.8-35B-A3B-Distill-Heretic",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 36.4 GB at this context; duplicate traces of the same session: gnyfab1d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3475.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-Heretic-oQ8-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 3,475 99.3 36.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jwezrzxz",
   "source_url": "https://omlx.ai/benchmarks/performance/jwezrzxz",
   "tg_tokens": null,
   "tps": 99.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jwezrzxz-ctx4096",
   "model": "Qwen3.8-35B-A3B-Distill-Heretic",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 37.0 GB at this context; duplicate traces of the same session: gnyfab1d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5319.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-Heretic-oQ8-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 5,319 124.7 37.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jwezrzxz",
   "source_url": "https://omlx.ai/benchmarks/performance/jwezrzxz",
   "tg_tokens": null,
   "tps": 124.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-m778zziy-ctx1024",
   "model": "Qwen3.8-35B-A3B-Distill-Heretic",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 35.4 GB at this context; duplicate traces of the same session: 3rlalrf2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 837.1,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-Heretic-oQ8-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 837.1 108.7 35.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m778zziy",
   "source_url": "https://omlx.ai/benchmarks/performance/m778zziy",
   "tg_tokens": null,
   "tps": 108.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-m778zziy-ctx4096",
   "model": "Qwen3.8-35B-A3B-Distill-Heretic",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: 3rlalrf2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2465.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-Heretic-oQ8-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 2,465 105.1 36.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m778zziy",
   "source_url": "https://omlx.ai/benchmarks/performance/m778zziy",
   "tg_tokens": null,
   "tps": 105.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-36148m61-ctx1024",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace, single run at 1k context; peak memory 30.7 GB; TTFT 1426 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 718.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ6e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 1k tokens 718.2 PP tok/s 79.6 TG tok/s 1426 TTFT (ms) 30.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/36148m61",
   "source_url": "https://omlx.ai/benchmarks/performance/36148m61",
   "tg_tokens": null,
   "tps": 79.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-k5tinhlw-ctx65536",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 11.1 GB; TTFT 105715 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 619.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 64k tokens 619.9 PP tok/s 30.6 TG tok/s 105715 TTFT (ms) 11.1 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k5tinhlw",
   "source_url": "https://omlx.ai/benchmarks/performance/k5tinhlw",
   "tg_tokens": null,
   "tps": 30.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-2wpcu7w2-ctx1024",
   "model": "gemma-4-26B-A4B-it-Claude-Opus-Distilled-v2",
   "notes": "mxfp8 MLX build; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: b3n7ey3e, mrdkl4nx, nfsfzahd, pl59q879, se8qw6gs, sj0oj6u8, vpgs9vs3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 804.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-Claude-Opus-Distilled-v2-MLX-8bit-mxfp8 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 804.1 53.8 25.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2wpcu7w2",
   "source_url": "https://omlx.ai/benchmarks/performance/2wpcu7w2",
   "tg_tokens": null,
   "tps": 53.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-2wpcu7w2-ctx4096",
   "model": "gemma-4-26B-A4B-it-Claude-Opus-Distilled-v2",
   "notes": "mxfp8 MLX build; oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: b3n7ey3e, mrdkl4nx, nfsfzahd, pl59q879, se8qw6gs, sj0oj6u8, vpgs9vs3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 748.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-Claude-Opus-Distilled-v2-MLX-8bit-mxfp8 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 748.9 49.1 25.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2wpcu7w2",
   "source_url": "https://omlx.ai/benchmarks/performance/2wpcu7w2",
   "tg_tokens": null,
   "tps": 49.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-2wpcu7w2-ctx8192",
   "model": "gemma-4-26B-A4B-it-Claude-Opus-Distilled-v2",
   "notes": "mxfp8 MLX build; oMLX community benchmark trace; peak memory 25.6 GB at this context; duplicate traces of the same session: b3n7ey3e, mrdkl4nx, nfsfzahd, pl59q879, se8qw6gs, sj0oj6u8, vpgs9vs3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 754.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-Claude-Opus-Distilled-v2-MLX-8bit-mxfp8 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 754.1 50.7 25.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2wpcu7w2",
   "source_url": "https://omlx.ai/benchmarks/performance/2wpcu7w2",
   "tg_tokens": null,
   "tps": 50.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-2wpcu7w2-ctx16384",
   "model": "gemma-4-26B-A4B-it-Claude-Opus-Distilled-v2",
   "notes": "mxfp8 MLX build; oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: b3n7ey3e, mrdkl4nx, nfsfzahd, pl59q879, se8qw6gs, sj0oj6u8, vpgs9vs3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 722.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-Claude-Opus-Distilled-v2-MLX-8bit-mxfp8 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 16k 722.7 47.9 26.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2wpcu7w2",
   "source_url": "https://omlx.ai/benchmarks/performance/2wpcu7w2",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-2wpcu7w2-ctx32768",
   "model": "gemma-4-26B-A4B-it-Claude-Opus-Distilled-v2",
   "notes": "mxfp8 MLX build; oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: b3n7ey3e, mrdkl4nx, nfsfzahd, pl59q879, se8qw6gs, sj0oj6u8, vpgs9vs3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 642.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-Claude-Opus-Distilled-v2-MLX-8bit-mxfp8 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 32k 642.2 41.6 26.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2wpcu7w2",
   "source_url": "https://omlx.ai/benchmarks/performance/2wpcu7w2",
   "tg_tokens": null,
   "tps": 41.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-2wpcu7w2-ctx65536",
   "model": "gemma-4-26B-A4B-it-Claude-Opus-Distilled-v2",
   "notes": "mxfp8 MLX build; oMLX community benchmark trace; peak memory 28.4 GB at this context; duplicate traces of the same session: b3n7ey3e, mrdkl4nx, nfsfzahd, pl59q879, se8qw6gs, sj0oj6u8, vpgs9vs3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 533.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-Claude-Opus-Distilled-v2-MLX-8bit-mxfp8 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 64k 533.9 34.6 28.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2wpcu7w2",
   "source_url": "https://omlx.ai/benchmarks/performance/2wpcu7w2",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-2wpcu7w2-ctx131072",
   "model": "gemma-4-26B-A4B-it-Claude-Opus-Distilled-v2",
   "notes": "mxfp8 MLX build; oMLX community benchmark trace; peak memory 31.6 GB at this context; duplicate traces of the same session: b3n7ey3e, mrdkl4nx, nfsfzahd, pl59q879, se8qw6gs, sj0oj6u8, vpgs9vs3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 396.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-Claude-Opus-Distilled-v2-MLX-8bit-mxfp8 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 128k 396.0 27.1 31.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2wpcu7w2",
   "source_url": "https://omlx.ai/benchmarks/performance/2wpcu7w2",
   "tg_tokens": null,
   "tps": 27.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-2wpcu7w2-ctx199680",
   "model": "gemma-4-26B-A4B-it-Claude-Opus-Distilled-v2",
   "notes": "mxfp8 MLX build; oMLX community benchmark trace; peak memory 35.1 GB at this context; duplicate traces of the same session: b3n7ey3e, mrdkl4nx, nfsfzahd, pl59q879, se8qw6gs, sj0oj6u8, vpgs9vs3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 305.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-Claude-Opus-Distilled-v2-MLX-8bit-mxfp8 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 195k 305.0 21.5 35.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2wpcu7w2",
   "source_url": "https://omlx.ai/benchmarks/performance/2wpcu7w2",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-9bb6f2h5-ctx1024",
   "model": "gemma-4-12b-it",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 369.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mlx-community--gemma-4-12B-it-8bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 8bit; 1k 369.8 9.3 12.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9bb6f2h5",
   "source_url": "https://omlx.ai/benchmarks/performance/9bb6f2h5",
   "tg_tokens": null,
   "tps": 9.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-9bb6f2h5-ctx4096",
   "model": "gemma-4-12b-it",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 13.9 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 428.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mlx-community--gemma-4-12B-it-8bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 8bit; 4k 428.9 8.8 13.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9bb6f2h5",
   "source_url": "https://omlx.ai/benchmarks/performance/9bb6f2h5",
   "tg_tokens": null,
   "tps": 8.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2gaehq5e-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 173.7 GB at this context; duplicate traces of the same session: 7kl47b5v; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 356.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 356.8 42.1 173.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2gaehq5e",
   "source_url": "https://omlx.ai/benchmarks/performance/2gaehq5e",
   "tg_tokens": null,
   "tps": 42.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2gaehq5e-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 174.8 GB at this context; duplicate traces of the same session: 7kl47b5v; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 432.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 432.2 40.1 174.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2gaehq5e",
   "source_url": "https://omlx.ai/benchmarks/performance/2gaehq5e",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-mr1xx62o-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.1 GB at this context; duplicate traces of the same session: 1wltn2kb, owf6pbhz, uoowonsd, v5kkacer, xv9ie2wn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 485.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 485.0 42.8 179.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mr1xx62o",
   "source_url": "https://omlx.ai/benchmarks/performance/mr1xx62o",
   "tg_tokens": null,
   "tps": 42.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-mr1xx62o-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.2 GB at this context; duplicate traces of the same session: 1wltn2kb, owf6pbhz, uoowonsd, v5kkacer, xv9ie2wn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 529.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 529.7 51.3 179.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mr1xx62o",
   "source_url": "https://omlx.ai/benchmarks/performance/mr1xx62o",
   "tg_tokens": null,
   "tps": 51.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-mr1xx62o-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.5 GB at this context; duplicate traces of the same session: 1wltn2kb, owf6pbhz, uoowonsd, v5kkacer, xv9ie2wn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 522.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 522.7 51.1 179.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mr1xx62o",
   "source_url": "https://omlx.ai/benchmarks/performance/mr1xx62o",
   "tg_tokens": null,
   "tps": 51.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-mr1xx62o-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.1 GB at this context; duplicate traces of the same session: 1wltn2kb, owf6pbhz, uoowonsd, v5kkacer, xv9ie2wn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 512.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 512.3 39.6 180.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mr1xx62o",
   "source_url": "https://omlx.ai/benchmarks/performance/mr1xx62o",
   "tg_tokens": null,
   "tps": 39.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-mr1xx62o-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.4 GB at this context; duplicate traces of the same session: 1wltn2kb, owf6pbhz, uoowonsd, v5kkacer, xv9ie2wn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 460.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 460.9 43.6 181.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mr1xx62o",
   "source_url": "https://omlx.ai/benchmarks/performance/mr1xx62o",
   "tg_tokens": null,
   "tps": 43.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-mr1xx62o-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 182.8 GB at this context; duplicate traces of the same session: 1wltn2kb, owf6pbhz, uoowonsd, v5kkacer, xv9ie2wn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 439.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 195k 439.6 30.2 182.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mr1xx62o",
   "source_url": "https://omlx.ai/benchmarks/performance/mr1xx62o",
   "tg_tokens": null,
   "tps": 30.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-m0lyua6c-ctx4096",
   "model": "LFM2.5-2.6B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 3.2 GB at this context; duplicate traces of the same session: obcvcjjl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1610.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "LFM2.5-2.6B-MLX-6bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 6bit; 4k 1,610 118.4 3.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m0lyua6c",
   "source_url": "https://omlx.ai/benchmarks/performance/m0lyua6c",
   "tg_tokens": null,
   "tps": 118.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-m0lyua6c-ctx16384",
   "model": "LFM2.5-2.6B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 3.3 GB at this context; duplicate traces of the same session: obcvcjjl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1447.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "LFM2.5-2.6B-MLX-6bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 6bit; 16k 1,447 119.1 3.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m0lyua6c",
   "source_url": "https://omlx.ai/benchmarks/performance/m0lyua6c",
   "tg_tokens": null,
   "tps": 119.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-hxyql6qf-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 1k context; peak memory 13.4 GB; TTFT 513 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1995.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 1,995 PP tok/s 136.3 TG tok/s 513 TTFT (ms) 13.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hxyql6qf",
   "source_url": "https://omlx.ai/benchmarks/performance/hxyql6qf",
   "tg_tokens": null,
   "tps": 136.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-3u1pgmer-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1250.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-24B-A2B-oQ8e-fp16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k 1,250 65.1 24.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3u1pgmer",
   "source_url": "https://omlx.ai/benchmarks/performance/3u1pgmer",
   "tg_tokens": null,
   "tps": 65.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-3u1pgmer-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1176.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-24B-A2B-oQ8e-fp16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 16k 1,176 57.0 25.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3u1pgmer",
   "source_url": "https://omlx.ai/benchmarks/performance/3u1pgmer",
   "tg_tokens": null,
   "tps": 57.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-3u1pgmer-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1056.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-24B-A2B-oQ8e-fp16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 32k 1,056 50.8 25.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3u1pgmer",
   "source_url": "https://omlx.ai/benchmarks/performance/3u1pgmer",
   "tg_tokens": null,
   "tps": 50.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-3u1pgmer-ctx65536",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 877.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-24B-A2B-oQ8e-fp16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 64k 877.4 41.7 25.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3u1pgmer",
   "source_url": "https://omlx.ai/benchmarks/performance/3u1pgmer",
   "tg_tokens": null,
   "tps": 41.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-9lt7q2t5-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 43.9 GB; TTFT 912 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4492.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 4bit; 4k tokens 4,492 PP tok/s 101.3 TG tok/s 912 TTFT (ms) 43.9 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9lt7q2t5",
   "source_url": "https://omlx.ai/benchmarks/performance/9lt7q2t5",
   "tg_tokens": null,
   "tps": 101.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-f3i3a5nk-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "MLX MTP build; oMLX community benchmark trace; peak memory 10.8 GB at this context; duplicate traces of the same session: 3soi69x2; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 160.1,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.5-9B-MLX-MTP-5bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 5bit; 1k 160.1 17.0 10.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f3i3a5nk",
   "source_url": "https://omlx.ai/benchmarks/performance/f3i3a5nk",
   "tg_tokens": null,
   "tps": 17.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-f3i3a5nk-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "MLX MTP build; oMLX community benchmark trace; peak memory 11.1 GB at this context; duplicate traces of the same session: 3soi69x2; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 242.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.5-9B-MLX-MTP-5bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 5bit; 4k 242.0 16.9 11.1 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f3i3a5nk",
   "source_url": "https://omlx.ai/benchmarks/performance/f3i3a5nk",
   "tg_tokens": null,
   "tps": 16.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-8pmltbbr-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.2 GB at this context; duplicate traces of the same session: k27f7sx7; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 152.2,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.5-9B-OptiQ-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 152.2 15.9 8.2 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8pmltbbr",
   "source_url": "https://omlx.ai/benchmarks/performance/8pmltbbr",
   "tg_tokens": null,
   "tps": 15.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-8pmltbbr-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.7 GB at this context; duplicate traces of the same session: k27f7sx7; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 198.1,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.5-9B-OptiQ-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 198.1 15.6 8.7 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8pmltbbr",
   "source_url": "https://omlx.ai/benchmarks/performance/8pmltbbr",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-eqqzhz8x-ctx4096",
   "model": "Qwen3.6-35B-A3B-Fable-Holo3.1-Qwopus-KAT-Coder-C-qx86-hi",
   "notes": "community merge; oMLX community benchmark trace; peak memory 37.9 GB at this context; duplicate traces of the same session: fpv7o2ry; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2301.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-Fable-Holo3.1-Qwopus-KAT-Coder-C-qx86-hi-mlx on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 2,301 103.8 37.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eqqzhz8x",
   "source_url": "https://omlx.ai/benchmarks/performance/eqqzhz8x",
   "tg_tokens": null,
   "tps": 103.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-eqqzhz8x-ctx32768",
   "model": "Qwen3.6-35B-A3B-Fable-Holo3.1-Qwopus-KAT-Coder-C-qx86-hi",
   "notes": "community merge; oMLX community benchmark trace; peak memory 38.3 GB at this context; duplicate traces of the same session: fpv7o2ry; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4386.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-Fable-Holo3.1-Qwopus-KAT-Coder-C-qx86-hi-mlx on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 4,386 89.0 38.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eqqzhz8x",
   "source_url": "https://omlx.ai/benchmarks/performance/eqqzhz8x",
   "tg_tokens": null,
   "tps": 89.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-c0iu429q-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: sp80nsoi; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2599.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 2,599 105.8 19.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c0iu429q",
   "source_url": "https://omlx.ai/benchmarks/performance/c0iu429q",
   "tg_tokens": null,
   "tps": 105.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-c0iu429q-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: sp80nsoi; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3373.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 3,373 101.4 20.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c0iu429q",
   "source_url": "https://omlx.ai/benchmarks/performance/c0iu429q",
   "tg_tokens": null,
   "tps": 101.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-8so4wu3q-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: t5ptd5or; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1775.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.6-35B-A3B-mxfp4 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,775 118.2 20.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8so4wu3q",
   "source_url": "https://omlx.ai/benchmarks/performance/8so4wu3q",
   "tg_tokens": null,
   "tps": 118.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-8so4wu3q-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: t5ptd5or; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1759.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.6-35B-A3B-mxfp4 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,759 101.1 20.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8so4wu3q",
   "source_url": "https://omlx.ai/benchmarks/performance/8so4wu3q",
   "tg_tokens": null,
   "tps": 101.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-ouivimm5-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: x1fcn1es; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 214.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 214.2 27.6 18.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ouivimm5",
   "source_url": "https://omlx.ai/benchmarks/performance/ouivimm5",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-ouivimm5-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: x1fcn1es; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 77.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 77.1 10.2 21.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ouivimm5",
   "source_url": "https://omlx.ai/benchmarks/performance/ouivimm5",
   "tg_tokens": null,
   "tps": 10.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-hy09g8ns-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: 6kaennej; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 133.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 133.7 14.9 17.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hy09g8ns",
   "source_url": "https://omlx.ai/benchmarks/performance/hy09g8ns",
   "tg_tokens": null,
   "tps": 14.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-hy09g8ns-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: 6kaennej; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 126.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 126.6 13.4 18.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hy09g8ns",
   "source_url": "https://omlx.ai/benchmarks/performance/hy09g8ns",
   "tg_tokens": null,
   "tps": 13.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lbheokw4-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 39.0 GB at this context; duplicate traces of the same session: ki3n12ub; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 435.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 435.0 18.8 39.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lbheokw4",
   "source_url": "https://omlx.ai/benchmarks/performance/lbheokw4",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lbheokw4-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 40.4 GB at this context; duplicate traces of the same session: ki3n12ub; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 399.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 399.8 17.6 40.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lbheokw4",
   "source_url": "https://omlx.ai/benchmarks/performance/lbheokw4",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ftsi05uo-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: bhcojjii; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 471.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 471.3 33.6 17.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ftsi05uo",
   "source_url": "https://omlx.ai/benchmarks/performance/ftsi05uo",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ftsi05uo-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: bhcojjii; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 468.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 468.9 37.7 18.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ftsi05uo",
   "source_url": "https://omlx.ai/benchmarks/performance/ftsi05uo",
   "tg_tokens": null,
   "tps": 37.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-zsg6y9ln-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: xyl6mg7i; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 469.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 469.2 37.8 17.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zsg6y9ln",
   "source_url": "https://omlx.ai/benchmarks/performance/zsg6y9ln",
   "tg_tokens": null,
   "tps": 37.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-zsg6y9ln-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: xyl6mg7i; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 474.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 474.6 33.9 18.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zsg6y9ln",
   "source_url": "https://omlx.ai/benchmarks/performance/zsg6y9ln",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-euhrb299-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: j285c5d1, tshwdlan; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 907.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 907.0 41.0 17.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/euhrb299",
   "source_url": "https://omlx.ai/benchmarks/performance/euhrb299",
   "tg_tokens": null,
   "tps": 41.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-euhrb299-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: j285c5d1, tshwdlan; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1083.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 1,083 47.9 19.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/euhrb299",
   "source_url": "https://omlx.ai/benchmarks/performance/euhrb299",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-euhrb299-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: j285c5d1, tshwdlan; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1133.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 1,133 31.7 19.4 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/euhrb299",
   "source_url": "https://omlx.ai/benchmarks/performance/euhrb299",
   "tg_tokens": null,
   "tps": 31.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-euhrb299-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: j285c5d1, tshwdlan; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1101.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 1,101 37.9 19.9 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/euhrb299",
   "source_url": "https://omlx.ai/benchmarks/performance/euhrb299",
   "tg_tokens": null,
   "tps": 37.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-euhrb299-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: j285c5d1, tshwdlan; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1003.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 32k 1,003 28.0 21.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/euhrb299",
   "source_url": "https://omlx.ai/benchmarks/performance/euhrb299",
   "tg_tokens": null,
   "tps": 28.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7qb64gfu-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: aggf0jo6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 901.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 901.6 46.9 18.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7qb64gfu",
   "source_url": "https://omlx.ai/benchmarks/performance/7qb64gfu",
   "tg_tokens": null,
   "tps": 46.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7qb64gfu-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: aggf0jo6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 835.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 835.8 56.4 19.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7qb64gfu",
   "source_url": "https://omlx.ai/benchmarks/performance/7qb64gfu",
   "tg_tokens": null,
   "tps": 56.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-m7dr29lb-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 39.0 GB; TTFT 78052 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 839.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k tokens 839.6 PP tok/s 39.7 TG tok/s 78052 TTFT (ms) 39.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m7dr29lb",
   "source_url": "https://omlx.ai/benchmarks/performance/m7dr29lb",
   "tg_tokens": null,
   "tps": 39.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-uik87jp3-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 39.0 GB; TTFT 98454 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 665.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k tokens 665.7 PP tok/s 39.9 TG tok/s 98454 TTFT (ms) 39.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uik87jp3",
   "source_url": "https://omlx.ai/benchmarks/performance/uik87jp3",
   "tg_tokens": null,
   "tps": 39.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-98hojc1u-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 33.3 GB; TTFT 93436 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 701.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k tokens 701.4 PP tok/s 16.1 TG tok/s 93436 TTFT (ms) 33.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/98hojc1u",
   "source_url": "https://omlx.ai/benchmarks/performance/98hojc1u",
   "tg_tokens": null,
   "tps": 16.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-nk8vq62c-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 39.0 GB; TTFT 96362 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 680.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k tokens 680.1 PP tok/s 41.8 TG tok/s 96362 TTFT (ms) 39.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nk8vq62c",
   "source_url": "https://omlx.ai/benchmarks/performance/nk8vq62c",
   "tg_tokens": null,
   "tps": 41.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i66uwibs-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 39.0 GB; TTFT 98032 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 668.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k tokens 668.5 PP tok/s 11.7 TG tok/s 98032 TTFT (ms) 39.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i66uwibs",
   "source_url": "https://omlx.ai/benchmarks/performance/i66uwibs",
   "tg_tokens": null,
   "tps": 11.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-kl9ownsb-ctx1024",
   "model": "Qwen3.8-35B-A3B-Distill-Heretic",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 38.2 GB at this context; duplicate traces of the same session: ne1qprfk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3457.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-Heretic-oQ8-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 3,457 134.3 38.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kl9ownsb",
   "source_url": "https://omlx.ai/benchmarks/performance/kl9ownsb",
   "tg_tokens": null,
   "tps": 134.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-kl9ownsb-ctx4096",
   "model": "Qwen3.8-35B-A3B-Distill-Heretic",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 38.8 GB at this context; duplicate traces of the same session: ne1qprfk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5696.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-Heretic-oQ8-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 5,696 136.5 38.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kl9ownsb",
   "source_url": "https://omlx.ai/benchmarks/performance/kl9ownsb",
   "tg_tokens": null,
   "tps": 136.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-kjup6kz9-ctx1024",
   "model": "Qwen3.8-35B-A3B-Distill-Heretic",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 38.2 GB at this context; duplicate traces of the same session: b5sql5y2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3434.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-Heretic-oQ8-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 3,434 136.0 38.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kjup6kz9",
   "source_url": "https://omlx.ai/benchmarks/performance/kjup6kz9",
   "tg_tokens": null,
   "tps": 136.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-kjup6kz9-ctx4096",
   "model": "Qwen3.8-35B-A3B-Distill-Heretic",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 38.8 GB at this context; duplicate traces of the same session: b5sql5y2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5368.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-Heretic-oQ8-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 5,368 145.6 38.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kjup6kz9",
   "source_url": "https://omlx.ai/benchmarks/performance/kjup6kz9",
   "tg_tokens": null,
   "tps": 145.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-unoy8gel-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP320",
   "notes": "DWQ MTP Vision build; oMLX community benchmark trace; peak memory 68.7 GB at this context; duplicate traces of the same session: 94tl3r8v; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2621.0,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-Flash-Next-REAP320-oQ3e-fp16-DWQ-MTP-Vision-MLX on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 3bit; 1k 2,621 81.5 68.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/unoy8gel",
   "source_url": "https://omlx.ai/benchmarks/performance/unoy8gel",
   "tg_tokens": null,
   "tps": 81.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-unoy8gel-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP320",
   "notes": "DWQ MTP Vision build; oMLX community benchmark trace; peak memory 71.2 GB at this context; duplicate traces of the same session: 94tl3r8v; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3480.0,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-Flash-Next-REAP320-oQ3e-fp16-DWQ-MTP-Vision-MLX on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 3bit; 4k 3,480 82.0 71.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/unoy8gel",
   "source_url": "https://omlx.ai/benchmarks/performance/unoy8gel",
   "tg_tokens": null,
   "tps": 82.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-39bw4kh3-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.8 GB; TTFT 1934 ms; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 529.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 1k tokens 529.4 PP tok/s 62.7 TG tok/s 1934 TTFT (ms) 70.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/39bw4kh3",
   "source_url": "https://omlx.ai/benchmarks/performance/39bw4kh3",
   "tg_tokens": null,
   "tps": 62.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-fua0xlyr-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.8 GB; TTFT 2740 ms; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 373.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 1k tokens 373.7 PP tok/s 21.9 TG tok/s 2740 TTFT (ms) 70.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fua0xlyr",
   "source_url": "https://omlx.ai/benchmarks/performance/fua0xlyr",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hv8w5awv-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.7 GB at this context; duplicate traces of the same session: llxwhk92; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 723.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 723.1 38.5 100.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hv8w5awv",
   "source_url": "https://omlx.ai/benchmarks/performance/hv8w5awv",
   "tg_tokens": null,
   "tps": 38.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hv8w5awv-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: llxwhk92; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1082.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,082 58.9 101.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hv8w5awv",
   "source_url": "https://omlx.ai/benchmarks/performance/hv8w5awv",
   "tg_tokens": null,
   "tps": 58.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-hmfvmlht-ctx4096",
   "model": "Swift 1.5 Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 143.3 GB at this context; duplicate traces of the same session: hdyj6hhk; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1045.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 4k 1,045 78.8 143.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hmfvmlht",
   "source_url": "https://omlx.ai/benchmarks/performance/hmfvmlht",
   "tg_tokens": null,
   "tps": 78.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-hmfvmlht-ctx65536",
   "model": "Swift 1.5 Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 156.6 GB at this context; duplicate traces of the same session: hdyj6hhk; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1008.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 64k 1,008 0.8 156.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hmfvmlht",
   "source_url": "https://omlx.ai/benchmarks/performance/hmfvmlht",
   "tg_tokens": null,
   "tps": 0.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-9xaezoqv-ctx65536",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 25.2 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1752.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 1,752 71.4 25.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9xaezoqv",
   "source_url": "https://omlx.ai/benchmarks/performance/9xaezoqv",
   "tg_tokens": null,
   "tps": 71.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-9xaezoqv-ctx131072",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1014.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 128k 1,014 64.2 28.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9xaezoqv",
   "source_url": "https://omlx.ai/benchmarks/performance/9xaezoqv",
   "tg_tokens": null,
   "tps": 64.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-0tt9fclv-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 10.9 GB; TTFT 2259 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 453.3,
   "provenance": "community",
   "quant": "MXFP4-Q4",
   "quote": "gpt-oss-20b-MXFP4-Q4 on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k tokens 453.3 PP tok/s 59.7 TG tok/s 2259 TTFT (ms) 10.9 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0tt9fclv",
   "source_url": "https://omlx.ai/benchmarks/performance/0tt9fclv",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-b94nkjur-ctx8192",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace, single run at 8k context; peak memory 11.0 GB; TTFT 18028 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 454.4,
   "provenance": "community",
   "quant": "MXFP4-Q4",
   "quote": "gpt-oss-20b-MXFP4-Q4 on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 8k tokens 454.4 PP tok/s 51.1 TG tok/s 18028 TTFT (ms) 11.0 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b94nkjur",
   "source_url": "https://omlx.ai/benchmarks/performance/b94nkjur",
   "tg_tokens": null,
   "tps": 51.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ishgmthl-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 10.9 GB; TTFT 2465 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 415.4,
   "provenance": "community",
   "quant": "MXFP4-Q4",
   "quote": "gpt-oss-20b-MXFP4-Q4 on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k tokens 415.4 PP tok/s 60.9 TG tok/s 2465 TTFT (ms) 10.9 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ishgmthl",
   "source_url": "https://omlx.ai/benchmarks/performance/ishgmthl",
   "tg_tokens": null,
   "tps": 60.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-3hcqachm-ctx1024",
   "model": "gemma-4-12b-it",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 369.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mlx-community--gemma-4-12B-it-8bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 8bit; 1k 369.8 9.3 12.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3hcqachm",
   "source_url": "https://omlx.ai/benchmarks/performance/3hcqachm",
   "tg_tokens": null,
   "tps": 9.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-3hcqachm-ctx4096",
   "model": "gemma-4-12b-it",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 13.9 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 428.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mlx-community--gemma-4-12B-it-8bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 8bit; 4k 428.9 8.8 13.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3hcqachm",
   "source_url": "https://omlx.ai/benchmarks/performance/3hcqachm",
   "tg_tokens": null,
   "tps": 8.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-hswl4kss-ctx1024",
   "model": "AREX-Turbo",
   "notes": "oMLX community benchmark trace; peak memory 5.2 GB at this context; duplicate traces of the same session: zi142jvt; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 407.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "AREX-Turbo on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 8bit; 1k 407.0 21.3 5.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hswl4kss",
   "source_url": "https://omlx.ai/benchmarks/performance/hswl4kss",
   "tg_tokens": null,
   "tps": 21.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-hswl4kss-ctx4096",
   "model": "AREX-Turbo",
   "notes": "oMLX community benchmark trace; peak memory 5.7 GB at this context; duplicate traces of the same session: zi142jvt; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 405.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "AREX-Turbo on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 8bit; 4k 405.1 20.8 5.7 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hswl4kss",
   "source_url": "https://omlx.ai/benchmarks/performance/hswl4kss",
   "tg_tokens": null,
   "tps": 20.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-z3tbdcri-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 176.1 GB at this context; duplicate traces of the same session: xfn2g4dh; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 519.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 519.5 41.5 176.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z3tbdcri",
   "source_url": "https://omlx.ai/benchmarks/performance/z3tbdcri",
   "tg_tokens": null,
   "tps": 41.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-z3tbdcri-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 176.6 GB at this context; duplicate traces of the same session: xfn2g4dh; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 542.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 542.6 39.9 176.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z3tbdcri",
   "source_url": "https://omlx.ai/benchmarks/performance/z3tbdcri",
   "tg_tokens": null,
   "tps": 39.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-2oeb0sg9-ctx32768",
   "model": "KAT-Coder-V2.5-Dev",
   "notes": "MLX build; oMLX community benchmark trace, single run at 32k context; peak memory 21.5 GB; TTFT 13367 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2452.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "KAT-Coder-V2.5-Dev-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k tokens 2,452 PP tok/s 72.1 TG tok/s 13367 TTFT (ms) 21.5 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2oeb0sg9",
   "source_url": "https://omlx.ai/benchmarks/performance/2oeb0sg9",
   "tg_tokens": null,
   "tps": 72.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-1omjb2ux-ctx32768",
   "model": "KAT-Coder-V2.5-Dev",
   "notes": "MLX build; oMLX community benchmark trace, single run at 32k context; peak memory 21.6 GB; TTFT 12124 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2703.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "KAT-Coder-V2.5-Dev-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k tokens 2,703 PP tok/s 76.1 TG tok/s 12124 TTFT (ms) 21.6 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1omjb2ux",
   "source_url": "https://omlx.ai/benchmarks/performance/1omjb2ux",
   "tg_tokens": null,
   "tps": 76.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-v8ddnimi-ctx32768",
   "model": "KAT-Coder-V2.5-Dev",
   "notes": "MLX build; oMLX community benchmark trace, single run at 32k context; peak memory 21.6 GB; TTFT 12069 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2715.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "KAT-Coder-V2.5-Dev-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k tokens 2,715 PP tok/s 76.1 TG tok/s 12069 TTFT (ms) 21.6 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v8ddnimi",
   "source_url": "https://omlx.ai/benchmarks/performance/v8ddnimi",
   "tg_tokens": null,
   "tps": 76.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-pq39ga1u-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: bhgltvdy; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1046.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 24 GB \u00b7 4bit; 1k 1,046 107.5 13.4 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pq39ga1u",
   "source_url": "https://omlx.ai/benchmarks/performance/pq39ga1u",
   "tg_tokens": null,
   "tps": 107.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-pq39ga1u-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: bhgltvdy; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1270.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 24 GB \u00b7 4bit; 4k 1,270 106.8 13.7 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pq39ga1u",
   "source_url": "https://omlx.ai/benchmarks/performance/pq39ga1u",
   "tg_tokens": null,
   "tps": 106.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-2zr3ym3b-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: a3ax2u9q, xqzqrhwf; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1250.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-24B-A2B-oQ8e-fp16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k 1,250 65.1 24.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2zr3ym3b",
   "source_url": "https://omlx.ai/benchmarks/performance/2zr3ym3b",
   "tg_tokens": null,
   "tps": 65.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-2zr3ym3b-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: a3ax2u9q, xqzqrhwf; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1176.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-24B-A2B-oQ8e-fp16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 16k 1,176 57.0 25.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2zr3ym3b",
   "source_url": "https://omlx.ai/benchmarks/performance/2zr3ym3b",
   "tg_tokens": null,
   "tps": 57.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-2zr3ym3b-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: a3ax2u9q, xqzqrhwf; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1056.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-24B-A2B-oQ8e-fp16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 32k 1,056 50.8 25.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2zr3ym3b",
   "source_url": "https://omlx.ai/benchmarks/performance/2zr3ym3b",
   "tg_tokens": null,
   "tps": 50.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-2zr3ym3b-ctx65536",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: a3ax2u9q, xqzqrhwf; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 877.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-24B-A2B-oQ8e-fp16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 64k 877.4 41.7 25.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2zr3ym3b",
   "source_url": "https://omlx.ai/benchmarks/performance/2zr3ym3b",
   "tg_tokens": null,
   "tps": 41.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-6da5w4zn-ctx4096",
   "model": "Mellum2-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 7.2 GB at this context; duplicate traces of the same session: cesk2bb1, cvk4ejko, iug18s12; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1155.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Mellum2-12B-A2.5B-Thinking-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 1,155 113.2 7.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6da5w4zn",
   "source_url": "https://omlx.ai/benchmarks/performance/6da5w4zn",
   "tg_tokens": null,
   "tps": 113.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-6da5w4zn-ctx16384",
   "model": "Mellum2-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 7.3 GB at this context; duplicate traces of the same session: cesk2bb1, cvk4ejko, iug18s12; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1074.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Mellum2-12B-A2.5B-Thinking-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 1,074 100.3 7.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6da5w4zn",
   "source_url": "https://omlx.ai/benchmarks/performance/6da5w4zn",
   "tg_tokens": null,
   "tps": 100.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-6da5w4zn-ctx32768",
   "model": "Mellum2-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 7.5 GB at this context; duplicate traces of the same session: cesk2bb1, cvk4ejko, iug18s12; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 929.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Mellum2-12B-A2.5B-Thinking-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 32k 929.2 90.7 7.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6da5w4zn",
   "source_url": "https://omlx.ai/benchmarks/performance/6da5w4zn",
   "tg_tokens": null,
   "tps": 90.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-6da5w4zn-ctx65536",
   "model": "Mellum2-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 7.9 GB at this context; duplicate traces of the same session: cesk2bb1, cvk4ejko, iug18s12; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 766.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Mellum2-12B-A2.5B-Thinking-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 64k 766.9 75.9 7.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6da5w4zn",
   "source_url": "https://omlx.ai/benchmarks/performance/6da5w4zn",
   "tg_tokens": null,
   "tps": 75.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-4ycvuduw-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: 49oax741, 63yuanl1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2717.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 2,717 76.2 21.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ycvuduw",
   "source_url": "https://omlx.ai/benchmarks/performance/4ycvuduw",
   "tg_tokens": null,
   "tps": 76.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-4ycvuduw-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: 49oax741, 63yuanl1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1967.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 1,967 59.4 23.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ycvuduw",
   "source_url": "https://omlx.ai/benchmarks/performance/4ycvuduw",
   "tg_tokens": null,
   "tps": 59.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-4ycvuduw-ctx131072",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: 49oax741, 63yuanl1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1201.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 128k 1,201 43.0 26.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ycvuduw",
   "source_url": "https://omlx.ai/benchmarks/performance/4ycvuduw",
   "tg_tokens": null,
   "tps": 43.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-o4sgwuxb-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 39.8 GB at this context; duplicate traces of the same session: j6a872tp; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 626.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 32k 626.8 48.8 39.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o4sgwuxb",
   "source_url": "https://omlx.ai/benchmarks/performance/o4sgwuxb",
   "tg_tokens": null,
   "tps": 48.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-o4sgwuxb-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 40.3 GB at this context; duplicate traces of the same session: j6a872tp; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 427.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 64k 427.5 41.6 40.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o4sgwuxb",
   "source_url": "https://omlx.ai/benchmarks/performance/o4sgwuxb",
   "tg_tokens": null,
   "tps": 41.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-a772z7pv-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 64k context; peak memory 21.9 GB; TTFT 34940 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1876.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k tokens 1,876 PP tok/s 57.5 TG tok/s 34940 TTFT (ms) 21.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a772z7pv",
   "source_url": "https://omlx.ai/benchmarks/performance/a772z7pv",
   "tg_tokens": null,
   "tps": 57.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-o67g216q-ctx8192",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.8 GB at this context; duplicate traces of the same session: omx49d15; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 436.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 436.5 57.8 7.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o67g216q",
   "source_url": "https://omlx.ai/benchmarks/performance/o67g216q",
   "tg_tokens": null,
   "tps": 57.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-o67g216q-ctx16384",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.4 GB at this context; duplicate traces of the same session: omx49d15; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 423.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 423.7 54.0 8.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o67g216q",
   "source_url": "https://omlx.ai/benchmarks/performance/o67g216q",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-r4iuwjm3-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 7.4 GB; TTFT 8851 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 462.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k tokens 462.8 PP tok/s 59.5 TG tok/s 8851 TTFT (ms) 7.4 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r4iuwjm3",
   "source_url": "https://omlx.ai/benchmarks/performance/r4iuwjm3",
   "tg_tokens": null,
   "tps": 59.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ej9jngxk-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.6 GB at this context; duplicate traces of the same session: cgzeuoso, dunbc4n0; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 493.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-9B-MLX-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 493.3 36.6 11.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ej9jngxk",
   "source_url": "https://omlx.ai/benchmarks/performance/ej9jngxk",
   "tg_tokens": null,
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ej9jngxk-ctx8192",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: cgzeuoso, dunbc4n0; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 481.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-9B-MLX-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 8k 481.2 35.9 11.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ej9jngxk",
   "source_url": "https://omlx.ai/benchmarks/performance/ej9jngxk",
   "tg_tokens": null,
   "tps": 35.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ej9jngxk-ctx16384",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 12.6 GB at this context; duplicate traces of the same session: cgzeuoso, dunbc4n0; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 449.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-9B-MLX-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 16k 449.8 34.3 12.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ej9jngxk",
   "source_url": "https://omlx.ai/benchmarks/performance/ej9jngxk",
   "tg_tokens": null,
   "tps": 34.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-gf3ges2a-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.4 GB at this context; duplicate traces of the same session: 3duhpj3l, lp3v3xs8, y0q0pf33; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1525.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 1,525 53.7 95.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gf3ges2a",
   "source_url": "https://omlx.ai/benchmarks/performance/gf3ges2a",
   "tg_tokens": null,
   "tps": 53.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-gf3ges2a-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.5 GB at this context; duplicate traces of the same session: 3duhpj3l, lp3v3xs8, y0q0pf33; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1784.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 1,784 50.4 95.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gf3ges2a",
   "source_url": "https://omlx.ai/benchmarks/performance/gf3ges2a",
   "tg_tokens": null,
   "tps": 50.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-gf3ges2a-ctx65536",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 96.6 GB at this context; duplicate traces of the same session: 3duhpj3l, lp3v3xs8, y0q0pf33; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1284.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k 1,284 40.1 96.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gf3ges2a",
   "source_url": "https://omlx.ai/benchmarks/performance/gf3ges2a",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-gf3ges2a-ctx131072",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 98.1 GB at this context; duplicate traces of the same session: 3duhpj3l, lp3v3xs8, y0q0pf33; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 929.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 128k 929.3 32.0 98.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gf3ges2a",
   "source_url": "https://omlx.ai/benchmarks/performance/gf3ges2a",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-gf3ges2a-ctx199680",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 99.7 GB at this context; duplicate traces of the same session: 3duhpj3l, lp3v3xs8, y0q0pf33; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 717.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 195k 717.1 26.4 99.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gf3ges2a",
   "source_url": "https://omlx.ai/benchmarks/performance/gf3ges2a",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-uou7vncw-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: 2yb4wh9t; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1969.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 1,969 73.5 35.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uou7vncw",
   "source_url": "https://omlx.ai/benchmarks/performance/uou7vncw",
   "tg_tokens": null,
   "tps": 73.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-uou7vncw-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 36.2 GB at this context; duplicate traces of the same session: 2yb4wh9t; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2147.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 2,147 75.2 36.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uou7vncw",
   "source_url": "https://omlx.ai/benchmarks/performance/uou7vncw",
   "tg_tokens": null,
   "tps": 75.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ransr1kg-ctx4096",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "fp16 MTP build; oMLX community benchmark trace, single run at 4k context; peak memory 24.8 GB; TTFT 5277 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 776.2,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-oQ5-fp16-mtp on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 5bit; 4k tokens 776.2 PP tok/s 57.1 TG tok/s 5277 TTFT (ms) 24.8 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ransr1kg",
   "source_url": "https://omlx.ai/benchmarks/performance/ransr1kg",
   "tg_tokens": null,
   "tps": 57.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-117ubc2w-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive",
   "notes": "OptiQ MLX build; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: lm4m23qd; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 938.9,
   "provenance": "community",
   "quant": "6bpw",
   "quote": "Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-OptiQ-6bpw-MLX on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 1k 938.9 81.8 19.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/117ubc2w",
   "source_url": "https://omlx.ai/benchmarks/performance/117ubc2w",
   "tg_tokens": null,
   "tps": 81.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-117ubc2w-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive",
   "notes": "OptiQ MLX build; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: lm4m23qd; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1016.0,
   "provenance": "community",
   "quant": "6bpw",
   "quote": "Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-OptiQ-6bpw-MLX on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k 1,016 81.6 21.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/117ubc2w",
   "source_url": "https://omlx.ai/benchmarks/performance/117ubc2w",
   "tg_tokens": null,
   "tps": 81.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-o9iamzl3-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive",
   "notes": "OptiQ MLX build; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: hpglrjyf; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 951.4,
   "provenance": "community",
   "quant": "6bpw",
   "quote": "Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-OptiQ-6bpw-MLX on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 1k 951.4 81.0 19.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9iamzl3",
   "source_url": "https://omlx.ai/benchmarks/performance/o9iamzl3",
   "tg_tokens": null,
   "tps": 81.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-o9iamzl3-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive",
   "notes": "OptiQ MLX build; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: hpglrjyf; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1035.0,
   "provenance": "community",
   "quant": "6bpw",
   "quote": "Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-OptiQ-6bpw-MLX on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k 1,035 78.4 21.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9iamzl3",
   "source_url": "https://omlx.ai/benchmarks/performance/o9iamzl3",
   "tg_tokens": null,
   "tps": 78.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-826yb7op-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: ah1n99mp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2251.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,251 108.6 20.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/826yb7op",
   "source_url": "https://omlx.ai/benchmarks/performance/826yb7op",
   "tg_tokens": null,
   "tps": 108.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-826yb7op-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: ah1n99mp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4400.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 4,400 101.9 21.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/826yb7op",
   "source_url": "https://omlx.ai/benchmarks/performance/826yb7op",
   "tg_tokens": null,
   "tps": 101.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-dljrjbxy-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 4j2isuh3; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2720.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 2,720 109.3 22.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dljrjbxy",
   "source_url": "https://omlx.ai/benchmarks/performance/dljrjbxy",
   "tg_tokens": null,
   "tps": 109.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-dljrjbxy-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 4j2isuh3; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2414.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 16k 2,414 102.4 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dljrjbxy",
   "source_url": "https://omlx.ai/benchmarks/performance/dljrjbxy",
   "tg_tokens": null,
   "tps": 102.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-v29i9lir-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: a97eqbky; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 208.3,
   "provenance": "community",
   "quant": "OptiQ 4bit",
   "quote": "Qwen3.8-27B-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 208.3 9.0 20.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v29i9lir",
   "source_url": "https://omlx.ai/benchmarks/performance/v29i9lir",
   "tg_tokens": null,
   "tps": 9.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-v29i9lir-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: a97eqbky; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 203.3,
   "provenance": "community",
   "quant": "OptiQ 4bit",
   "quote": "Qwen3.8-27B-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 203.3 8.0 21.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v29i9lir",
   "source_url": "https://omlx.ai/benchmarks/performance/v29i9lir",
   "tg_tokens": null,
   "tps": 8.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-8bkgvc2j-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable5.5-Cold-Fusion-L-Uncensored",
   "notes": "fp16 MTP build; oMLX community benchmark trace, single run at 4k context; peak memory 20.7 GB; TTFT 35345 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 115.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable5.5-Cold-Fusion-L-Uncensored-fp16-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 4k tokens 115.9 PP tok/s 16.5 TG tok/s 35345 TTFT (ms) 20.7 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8bkgvc2j",
   "source_url": "https://omlx.ai/benchmarks/performance/8bkgvc2j",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0d03hav5-ctx1024",
   "model": "Qwen3.8-35B-A3B-Distill-Heretic",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 36.4 GB at this context; duplicate traces of the same session: 68jvi1d3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3504.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-Heretic-oQ8-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 3,504 119.8 36.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0d03hav5",
   "source_url": "https://omlx.ai/benchmarks/performance/0d03hav5",
   "tg_tokens": null,
   "tps": 119.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0d03hav5-ctx4096",
   "model": "Qwen3.8-35B-A3B-Distill-Heretic",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 37.0 GB at this context; duplicate traces of the same session: 68jvi1d3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5439.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-Heretic-oQ8-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 5,439 133.4 37.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0d03hav5",
   "source_url": "https://omlx.ai/benchmarks/performance/0d03hav5",
   "tg_tokens": null,
   "tps": 133.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-f6ncuo2z-ctx1024",
   "model": "Qwen3.8-35B-A3B-Distill-Heretic",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 36.4 GB at this context; duplicate traces of the same session: 2piwp0pl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3470.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-Heretic-oQ8-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 3,470 117.8 36.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f6ncuo2z",
   "source_url": "https://omlx.ai/benchmarks/performance/f6ncuo2z",
   "tg_tokens": null,
   "tps": 117.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-f6ncuo2z-ctx4096",
   "model": "Qwen3.8-35B-A3B-Distill-Heretic",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 37.0 GB at this context; duplicate traces of the same session: 2piwp0pl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5335.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-Heretic-oQ8-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 5,335 128.6 37.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f6ncuo2z",
   "source_url": "https://omlx.ai/benchmarks/performance/f6ncuo2z",
   "tg_tokens": null,
   "tps": 128.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jrebjtyy-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 74.4 GB; TTFT 23804 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2753.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k tokens 2,753 PP tok/s 86.8 TG tok/s 23804 TTFT (ms) 74.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jrebjtyy",
   "source_url": "https://omlx.ai/benchmarks/performance/jrebjtyy",
   "tg_tokens": null,
   "tps": 86.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-eqjvl0pa-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 74.4 GB; TTFT 30757 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2131.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k tokens 2,131 PP tok/s 42.5 TG tok/s 30757 TTFT (ms) 74.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eqjvl0pa",
   "source_url": "https://omlx.ai/benchmarks/performance/eqjvl0pa",
   "tg_tokens": null,
   "tps": 42.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6bxiyd4n-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: thhca5aq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1671.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,671 71.7 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6bxiyd4n",
   "source_url": "https://omlx.ai/benchmarks/performance/6bxiyd4n",
   "tg_tokens": null,
   "tps": 71.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6bxiyd4n-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.5 GB at this context; duplicate traces of the same session: thhca5aq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2768.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,768 81.1 74.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6bxiyd4n",
   "source_url": "https://omlx.ai/benchmarks/performance/6bxiyd4n",
   "tg_tokens": null,
   "tps": 81.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-yfasmj71-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 13w3v7y5, 6yy16wgp; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 616.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 616.1 76.4 70.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yfasmj71",
   "source_url": "https://omlx.ai/benchmarks/performance/yfasmj71",
   "tg_tokens": null,
   "tps": 76.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-yfasmj71-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: 13w3v7y5, 6yy16wgp; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 807.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 807.6 69.9 71.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yfasmj71",
   "source_url": "https://omlx.ai/benchmarks/performance/yfasmj71",
   "tg_tokens": null,
   "tps": 69.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-yfasmj71-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: 13w3v7y5, 6yy16wgp; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1051.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,051 80.4 71.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yfasmj71",
   "source_url": "https://omlx.ai/benchmarks/performance/yfasmj71",
   "tg_tokens": null,
   "tps": 80.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-8ommtrj7-ctx65536",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 25.2 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1752.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 1,752 71.4 25.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8ommtrj7",
   "source_url": "https://omlx.ai/benchmarks/performance/8ommtrj7",
   "tg_tokens": null,
   "tps": 71.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-8ommtrj7-ctx131072",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1014.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 128k 1,014 64.2 28.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8ommtrj7",
   "source_url": "https://omlx.ai/benchmarks/performance/8ommtrj7",
   "tg_tokens": null,
   "tps": 64.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-hj6ou2z6-ctx32768",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 32k context; peak memory 23.5 GB; TTFT 13389 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2447.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k tokens 2,447 PP tok/s 120.2 TG tok/s 13389 TTFT (ms) 23.5 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hj6ou2z6",
   "source_url": "https://omlx.ai/benchmarks/performance/hj6ou2z6",
   "tg_tokens": null,
   "tps": 120.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-hffe9n2p-ctx16384",
   "model": "qwen4flashNext",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 70.9 GB; TTFT 6306 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2598.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "qwen4flashNext on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 16k tokens 2,598 PP tok/s 83.4 TG tok/s 6306 TTFT (ms) 70.9 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hffe9n2p",
   "source_url": "https://omlx.ai/benchmarks/performance/hffe9n2p",
   "tg_tokens": null,
   "tps": 83.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-zg5rxwsa-ctx16384",
   "model": "Kimi-VL-A3B-Instruct",
   "notes": "mlx build; oMLX community benchmark trace, single run at 16k context; peak memory 16.1 GB; TTFT 20677 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 792.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Kimi-VL-A3B-Instruct-mlx-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 16k tokens 792.4 PP tok/s 37.3 TG tok/s 20677 TTFT (ms) 16.1 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zg5rxwsa",
   "source_url": "https://omlx.ai/benchmarks/performance/zg5rxwsa",
   "tg_tokens": null,
   "tps": 37.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-bg6yyqcu-ctx16384",
   "model": "LFM2.5-1.2B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 16k context; peak memory 2.2 GB; TTFT 6154 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2662.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-1.2B-Instruct-MLX-8bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 8bit; 16k tokens 2,662 PP tok/s 105.1 TG tok/s 6154 TTFT (ms) 2.2 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bg6yyqcu",
   "source_url": "https://omlx.ai/benchmarks/performance/bg6yyqcu",
   "tg_tokens": null,
   "tps": 105.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-d8u74rt2-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 32k context; peak memory 14.4 GB; TTFT 38519 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 850.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-MLX-oQ4e on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k tokens 850.7 PP tok/s 66.1 TG tok/s 38519 TTFT (ms) 14.4 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d8u74rt2",
   "source_url": "https://omlx.ai/benchmarks/performance/d8u74rt2",
   "tg_tokens": null,
   "tps": 66.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-922m73d9-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.6 GB at this context; duplicate traces of the same session: 9qxfh5r4; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 963.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-MLX-oQ4e on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 963.6 98.3 13.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/922m73d9",
   "source_url": "https://omlx.ai/benchmarks/performance/922m73d9",
   "tg_tokens": null,
   "tps": 98.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-922m73d9-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: 9qxfh5r4; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 828.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-MLX-oQ4e on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 828.8 65.6 14.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/922m73d9",
   "source_url": "https://omlx.ai/benchmarks/performance/922m73d9",
   "tg_tokens": null,
   "tps": 65.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-fdgrp0b0-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 14.1 GB at this context; duplicate traces of the same session: mj42ajub; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 947.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-MLX-oQ4e on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 947.8 80.3 14.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fdgrp0b0",
   "source_url": "https://omlx.ai/benchmarks/performance/fdgrp0b0",
   "tg_tokens": null,
   "tps": 80.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-fdgrp0b0-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: mj42ajub; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 838.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-MLX-oQ4e on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 838.0 65.6 14.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fdgrp0b0",
   "source_url": "https://omlx.ai/benchmarks/performance/fdgrp0b0",
   "tg_tokens": null,
   "tps": 65.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-v9mjl8r8-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.6 GB at this context; duplicate traces of the same session: sr5ce1d8, vzmof2in, vzo3ex95, z5blaklx; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 931.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-MLX-oQ4e on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 931.9 98.6 13.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v9mjl8r8",
   "source_url": "https://omlx.ai/benchmarks/performance/v9mjl8r8",
   "tg_tokens": null,
   "tps": 98.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-v9mjl8r8-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.9 GB at this context; duplicate traces of the same session: sr5ce1d8, vzmof2in, vzo3ex95, z5blaklx; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1039.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-MLX-oQ4e on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 1,039 97.0 13.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v9mjl8r8",
   "source_url": "https://omlx.ai/benchmarks/performance/v9mjl8r8",
   "tg_tokens": null,
   "tps": 97.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-v9mjl8r8-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.9 GB at this context; duplicate traces of the same session: sr5ce1d8, vzmof2in, vzo3ex95, z5blaklx; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1009.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-MLX-oQ4e on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 1,009 91.2 13.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v9mjl8r8",
   "source_url": "https://omlx.ai/benchmarks/performance/v9mjl8r8",
   "tg_tokens": null,
   "tps": 91.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-v9mjl8r8-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 14.1 GB at this context; duplicate traces of the same session: sr5ce1d8, vzmof2in, vzo3ex95, z5blaklx; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 935.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-MLX-oQ4e on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 935.2 76.8 14.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v9mjl8r8",
   "source_url": "https://omlx.ai/benchmarks/performance/v9mjl8r8",
   "tg_tokens": null,
   "tps": 76.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-v9mjl8r8-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: sr5ce1d8, vzmof2in, vzo3ex95, z5blaklx; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 773.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-MLX-oQ4e on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 773.6 64.0 14.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v9mjl8r8",
   "source_url": "https://omlx.ai/benchmarks/performance/v9mjl8r8",
   "tg_tokens": null,
   "tps": 64.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-szkaez54-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.6 GB at this context; duplicate traces of the same session: mwitjwwr, nopx1zl2, wrbvjxfa, yukxh6z7; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 925.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-MLX-oQ4e on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 925.8 92.0 13.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/szkaez54",
   "source_url": "https://omlx.ai/benchmarks/performance/szkaez54",
   "tg_tokens": null,
   "tps": 92.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-szkaez54-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.9 GB at this context; duplicate traces of the same session: mwitjwwr, nopx1zl2, wrbvjxfa, yukxh6z7; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1011.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-MLX-oQ4e on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 1,011 82.8 13.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/szkaez54",
   "source_url": "https://omlx.ai/benchmarks/performance/szkaez54",
   "tg_tokens": null,
   "tps": 82.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-szkaez54-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.9 GB at this context; duplicate traces of the same session: mwitjwwr, nopx1zl2, wrbvjxfa, yukxh6z7; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 993.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-MLX-oQ4e on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 993.6 70.4 13.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/szkaez54",
   "source_url": "https://omlx.ai/benchmarks/performance/szkaez54",
   "tg_tokens": null,
   "tps": 70.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-szkaez54-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 14.1 GB at this context; duplicate traces of the same session: mwitjwwr, nopx1zl2, wrbvjxfa, yukxh6z7; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 947.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-MLX-oQ4e on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 947.7 53.3 14.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/szkaez54",
   "source_url": "https://omlx.ai/benchmarks/performance/szkaez54",
   "tg_tokens": null,
   "tps": 53.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-szkaez54-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: mwitjwwr, nopx1zl2, wrbvjxfa, yukxh6z7; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 833.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-MLX-oQ4e on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 833.1 35.7 14.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/szkaez54",
   "source_url": "https://omlx.ai/benchmarks/performance/szkaez54",
   "tg_tokens": null,
   "tps": 35.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-r7l6agb4-ctx16384",
   "model": "Ling-3.0-tiny",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 6.7 GB; TTFT 9602 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1706.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ling-3.0-tiny-oQ4e on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 16k tokens 1,706 PP tok/s 95.2 TG tok/s 9602 TTFT (ms) 6.7 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r7l6agb4",
   "source_url": "https://omlx.ai/benchmarks/performance/r7l6agb4",
   "tg_tokens": null,
   "tps": 95.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-lez31evr-ctx16384",
   "model": "Ling-3.0-tiny",
   "notes": "bf16; MLX; oMLX community benchmark trace, single run at 16k context; peak memory 10.0 GB; TTFT 10620 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1543.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ling-3.0-tiny-oQ8e-bf16-MLX on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 8bit; 16k tokens 1,543 PP tok/s 73.7 TG tok/s 10620 TTFT (ms) 10.0 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lez31evr",
   "source_url": "https://omlx.ai/benchmarks/performance/lez31evr",
   "tg_tokens": null,
   "tps": 73.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-89f0j5pt-ctx1024",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 9.8 GB at this context; duplicate traces of the same session: 848zln6p, 8pez9ups, hi07q3rd, jhduf423, yqos83pz, zkpzu51l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2002.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2-12B-A2.5B-Instruct-6bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 2,002 120.0 9.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/89f0j5pt",
   "source_url": "https://omlx.ai/benchmarks/performance/89f0j5pt",
   "tg_tokens": null,
   "tps": 120.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-89f0j5pt-ctx4096",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 9.9 GB at this context; duplicate traces of the same session: 848zln6p, 8pez9ups, hi07q3rd, jhduf423, yqos83pz, zkpzu51l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1926.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2-12B-A2.5B-Instruct-6bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 1,926 112.5 9.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/89f0j5pt",
   "source_url": "https://omlx.ai/benchmarks/performance/89f0j5pt",
   "tg_tokens": null,
   "tps": 112.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-89f0j5pt-ctx8192",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 10.0 GB at this context; duplicate traces of the same session: 848zln6p, 8pez9ups, hi07q3rd, jhduf423, yqos83pz, zkpzu51l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1846.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2-12B-A2.5B-Instruct-6bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 1,846 95.1 10.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/89f0j5pt",
   "source_url": "https://omlx.ai/benchmarks/performance/89f0j5pt",
   "tg_tokens": null,
   "tps": 95.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-89f0j5pt-ctx16384",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 10.1 GB at this context; duplicate traces of the same session: 848zln6p, 8pez9ups, hi07q3rd, jhduf423, yqos83pz, zkpzu51l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1731.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2-12B-A2.5B-Instruct-6bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 1,731 92.6 10.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/89f0j5pt",
   "source_url": "https://omlx.ai/benchmarks/performance/89f0j5pt",
   "tg_tokens": null,
   "tps": 92.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-89f0j5pt-ctx32768",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 10.3 GB at this context; duplicate traces of the same session: 848zln6p, 8pez9ups, hi07q3rd, jhduf423, yqos83pz, zkpzu51l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1530.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2-12B-A2.5B-Instruct-6bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 1,530 79.2 10.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/89f0j5pt",
   "source_url": "https://omlx.ai/benchmarks/performance/89f0j5pt",
   "tg_tokens": null,
   "tps": 79.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-89f0j5pt-ctx65536",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 11.0 GB at this context; duplicate traces of the same session: 848zln6p, 8pez9ups, hi07q3rd, jhduf423, yqos83pz, zkpzu51l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1200.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2-12B-A2.5B-Instruct-6bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k 1,200 56.5 11.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/89f0j5pt",
   "source_url": "https://omlx.ai/benchmarks/performance/89f0j5pt",
   "tg_tokens": null,
   "tps": 56.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-89f0j5pt-ctx131072",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: 848zln6p, 8pez9ups, hi07q3rd, jhduf423, yqos83pz, zkpzu51l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 764.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2-12B-A2.5B-Instruct-6bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 128k 764.4 31.7 12.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/89f0j5pt",
   "source_url": "https://omlx.ai/benchmarks/performance/89f0j5pt",
   "tg_tokens": null,
   "tps": 31.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-3n53r99o-ctx16384",
   "model": "Muse Glimmer",
   "notes": "Abliterated MM build; oMLX community benchmark trace, single run at 16k context; peak memory 23.6 GB; TTFT 159811 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.5,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Muse-Glimmer-30B-Abliterated-MM-oQ5e on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 5bit; 16k tokens 102.5 PP tok/s 6.9 TG tok/s 159811 TTFT (ms) 23.6 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3n53r99o",
   "source_url": "https://omlx.ai/benchmarks/performance/3n53r99o",
   "tg_tokens": null,
   "tps": 6.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-f22agwvn-ctx16384",
   "model": "Ornith-1.5-35B-A3B-BigBang",
   "notes": "MTP build; oMLX community benchmark trace, single run at 16k context; peak memory 22.9 GB; TTFT 16630 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 985.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-BigBang-MTP-mlx-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 16k tokens 985.2 PP tok/s 101.5 TG tok/s 16630 TTFT (ms) 22.9 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f22agwvn",
   "source_url": "https://omlx.ai/benchmarks/performance/f22agwvn",
   "tg_tokens": null,
   "tps": 101.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-xy2pmund-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: m6x0aovi, m7yw758m; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 759.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 759.2 53.9 20.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xy2pmund",
   "source_url": "https://omlx.ai/benchmarks/performance/xy2pmund",
   "tg_tokens": null,
   "tps": 53.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-xy2pmund-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: m6x0aovi, m7yw758m; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 793.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 793.1 55.3 21.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xy2pmund",
   "source_url": "https://omlx.ai/benchmarks/performance/xy2pmund",
   "tg_tokens": null,
   "tps": 55.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-xy2pmund-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: m6x0aovi, m7yw758m; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 533.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 32k 533.1 13.0 22.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xy2pmund",
   "source_url": "https://omlx.ai/benchmarks/performance/xy2pmund",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-nauci6t7-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.5 GB at this context; duplicate traces of the same session: vlpszfoq, wswfaxp7; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1028.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 1,028 59.5 37.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nauci6t7",
   "source_url": "https://omlx.ai/benchmarks/performance/nauci6t7",
   "tg_tokens": null,
   "tps": 59.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-nauci6t7-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.8 GB at this context; duplicate traces of the same session: vlpszfoq, wswfaxp7; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 954.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 8k 954.4 56.8 37.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nauci6t7",
   "source_url": "https://omlx.ai/benchmarks/performance/nauci6t7",
   "tg_tokens": null,
   "tps": 56.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-nauci6t7-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 38.4 GB at this context; duplicate traces of the same session: vlpszfoq, wswfaxp7; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 816.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 16k 816.0 54.2 38.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nauci6t7",
   "source_url": "https://omlx.ai/benchmarks/performance/nauci6t7",
   "tg_tokens": null,
   "tps": 54.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-uyxo631m-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1525.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 1,525 53.7 95.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uyxo631m",
   "source_url": "https://omlx.ai/benchmarks/performance/uyxo631m",
   "tg_tokens": null,
   "tps": 53.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-uyxo631m-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 95.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1784.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 1,784 50.4 95.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uyxo631m",
   "source_url": "https://omlx.ai/benchmarks/performance/uyxo631m",
   "tg_tokens": null,
   "tps": 50.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-uyxo631m-ctx65536",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 96.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1284.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k 1,284 40.1 96.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uyxo631m",
   "source_url": "https://omlx.ai/benchmarks/performance/uyxo631m",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-uyxo631m-ctx131072",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 98.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 929.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 128k 929.3 32.0 98.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uyxo631m",
   "source_url": "https://omlx.ai/benchmarks/performance/uyxo631m",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-uyxo631m-ctx199680",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 99.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 717.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-122B-A10B-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 195k 717.1 26.4 99.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uyxo631m",
   "source_url": "https://omlx.ai/benchmarks/performance/uyxo631m",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-vcexj3ix-ctx1024",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: fa13dh5q; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 135.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-27B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 1k 135.6 15.6 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vcexj3ix",
   "source_url": "https://omlx.ai/benchmarks/performance/vcexj3ix",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-vcexj3ix-ctx4096",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: fa13dh5q; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 124.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-27B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k 124.1 14.9 19.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vcexj3ix",
   "source_url": "https://omlx.ai/benchmarks/performance/vcexj3ix",
   "tg_tokens": null,
   "tps": 14.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-gnj9fiuc-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 18.8 GB; TTFT 10671 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 383.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k tokens 383.9 PP tok/s 22.5 TG tok/s 10671 TTFT (ms) 18.8 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gnj9fiuc",
   "source_url": "https://omlx.ai/benchmarks/performance/gnj9fiuc",
   "tg_tokens": null,
   "tps": 22.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5m78y723-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.2 GB at this context; duplicate traces of the same session: fnejrun4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 804.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 804.5 53.5 30.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5m78y723",
   "source_url": "https://omlx.ai/benchmarks/performance/5m78y723",
   "tg_tokens": null,
   "tps": 53.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5m78y723-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.2 GB at this context; duplicate traces of the same session: fnejrun4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 569.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 569.9 64.0 33.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5m78y723",
   "source_url": "https://omlx.ai/benchmarks/performance/5m78y723",
   "tg_tokens": null,
   "tps": 64.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ieljabxc-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.2 GB at this context; duplicate traces of the same session: e997kzh2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 802.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 802.2 39.2 30.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ieljabxc",
   "source_url": "https://omlx.ai/benchmarks/performance/ieljabxc",
   "tg_tokens": null,
   "tps": 39.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ieljabxc-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.2 GB at this context; duplicate traces of the same session: e997kzh2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 569.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 569.6 61.5 33.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ieljabxc",
   "source_url": "https://omlx.ai/benchmarks/performance/ieljabxc",
   "tg_tokens": null,
   "tps": 61.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0hvjfbaz-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1766.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,766 84.4 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0hvjfbaz",
   "source_url": "https://omlx.ai/benchmarks/performance/0hvjfbaz",
   "tg_tokens": null,
   "tps": 84.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0hvjfbaz-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2767.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,767 80.1 74.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0hvjfbaz",
   "source_url": "https://omlx.ai/benchmarks/performance/0hvjfbaz",
   "tg_tokens": null,
   "tps": 80.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-gmmt06c5-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.6 GB at this context; duplicate traces of the same session: dualiviu, jp74ke3l, u2e9a61d, x0qrv41d, zwup3607; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4540.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 1k 4,540 130.4 27.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gmmt06c5",
   "source_url": "https://omlx.ai/benchmarks/performance/gmmt06c5",
   "tg_tokens": null,
   "tps": 130.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-gmmt06c5-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: dualiviu, jp74ke3l, u2e9a61d, x0qrv41d, zwup3607; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4894.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 4k 4,894 119.2 28.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gmmt06c5",
   "source_url": "https://omlx.ai/benchmarks/performance/gmmt06c5",
   "tg_tokens": null,
   "tps": 119.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-gmmt06c5-ctx8192",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.2 GB at this context; duplicate traces of the same session: dualiviu, jp74ke3l, u2e9a61d, x0qrv41d, zwup3607; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5502.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 8k 5,502 154.5 28.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gmmt06c5",
   "source_url": "https://omlx.ai/benchmarks/performance/gmmt06c5",
   "tg_tokens": null,
   "tps": 154.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-gmmt06c5-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: dualiviu, jp74ke3l, u2e9a61d, x0qrv41d, zwup3607; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5038.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 16k 5,038 160.6 28.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gmmt06c5",
   "source_url": "https://omlx.ai/benchmarks/performance/gmmt06c5",
   "tg_tokens": null,
   "tps": 160.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-gmmt06c5-ctx32768",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.4 GB at this context; duplicate traces of the same session: dualiviu, jp74ke3l, u2e9a61d, x0qrv41d, zwup3607; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4693.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 32k 4,693 121.5 29.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gmmt06c5",
   "source_url": "https://omlx.ai/benchmarks/performance/gmmt06c5",
   "tg_tokens": null,
   "tps": 121.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-gmmt06c5-ctx65536",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.1 GB at this context; duplicate traces of the same session: dualiviu, jp74ke3l, u2e9a61d, x0qrv41d, zwup3607; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3708.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 64k 3,708 105.9 31.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gmmt06c5",
   "source_url": "https://omlx.ai/benchmarks/performance/gmmt06c5",
   "tg_tokens": null,
   "tps": 105.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-mjmype23-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT q4_0 MLX aligned build; oMLX community benchmark trace; peak memory 15.5 GB at this context; duplicate traces of the same session: wtv0b089; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 973.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--gemma-4-26B-A4B-it-qat-q4_0-mlx-aligned on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 973.8 42.0 15.5 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mjmype23",
   "source_url": "https://omlx.ai/benchmarks/performance/mjmype23",
   "tg_tokens": null,
   "tps": 42.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-mjmype23-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT q4_0 MLX aligned build; oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: wtv0b089; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1080.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--gemma-4-26B-A4B-it-qat-q4_0-mlx-aligned on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 1,080 39.1 15.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mjmype23",
   "source_url": "https://omlx.ai/benchmarks/performance/mjmype23",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-vany7vg6-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT q4_0 MLX aligned build; oMLX community benchmark trace; peak memory 11.9 GB at this context; duplicate traces of the same session: 1xsvdy3e; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 295.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--gemma-4-26B-A4B-it-qat-q4_0-mlx-aligned on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 295.2 29.9 11.9 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vany7vg6",
   "source_url": "https://omlx.ai/benchmarks/performance/vany7vg6",
   "tg_tokens": null,
   "tps": 29.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-vany7vg6-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT q4_0 MLX aligned build; oMLX community benchmark trace; peak memory 12.3 GB at this context; duplicate traces of the same session: 1xsvdy3e; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 169.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--gemma-4-26B-A4B-it-qat-q4_0-mlx-aligned on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 169.0 28.4 12.3 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vany7vg6",
   "source_url": "https://omlx.ai/benchmarks/performance/vany7vg6",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-rpd47i95-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT q4_0 MLX aligned build; oMLX community benchmark trace; peak memory 8.6 GB at this context; duplicate traces of the same session: 40s1v5cx; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 247.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--gemma-4-26B-A4B-it-qat-q4_0-mlx-aligned on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 247.9 23.8 8.6 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rpd47i95",
   "source_url": "https://omlx.ai/benchmarks/performance/rpd47i95",
   "tg_tokens": null,
   "tps": 23.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-rpd47i95-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT q4_0 MLX aligned build; oMLX community benchmark trace; peak memory 9.0 GB at this context; duplicate traces of the same session: 40s1v5cx; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 242.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--gemma-4-26B-A4B-it-qat-q4_0-mlx-aligned on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 242.3 23.3 9.0 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rpd47i95",
   "source_url": "https://omlx.ai/benchmarks/performance/rpd47i95",
   "tg_tokens": null,
   "tps": 23.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-2ki2o1ue-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT q4_0 MLX aligned build; oMLX community benchmark trace; peak memory 5.3 GB at this context; duplicate traces of the same session: 9txk0oy6; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 265.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--gemma-4-26B-A4B-it-qat-q4_0-mlx-aligned on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 265.5 17.3 5.3 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ki2o1ue",
   "source_url": "https://omlx.ai/benchmarks/performance/2ki2o1ue",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-2ki2o1ue-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT q4_0 MLX aligned build; oMLX community benchmark trace; peak memory 5.7 GB at this context; duplicate traces of the same session: 9txk0oy6; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 263.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--gemma-4-26B-A4B-it-qat-q4_0-mlx-aligned on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 263.4 17.5 5.7 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ki2o1ue",
   "source_url": "https://omlx.ai/benchmarks/performance/2ki2o1ue",
   "tg_tokens": null,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-uuy77vrw-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT q4_0 MLX aligned build; oMLX community benchmark trace; peak memory 15.5 GB at this context; duplicate traces of the same session: o1llviaa; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 596.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--gemma-4-26B-A4B-it-qat-q4_0-mlx-aligned on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 596.0 41.8 15.5 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uuy77vrw",
   "source_url": "https://omlx.ai/benchmarks/performance/uuy77vrw",
   "tg_tokens": null,
   "tps": 41.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-uuy77vrw-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT q4_0 MLX aligned build; oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: o1llviaa; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1072.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--gemma-4-26B-A4B-it-qat-q4_0-mlx-aligned on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 1,072 37.3 15.8 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uuy77vrw",
   "source_url": "https://omlx.ai/benchmarks/performance/uuy77vrw",
   "tg_tokens": null,
   "tps": 37.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-c0v1z23c-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT q4_0 MLX aligned build; oMLX community benchmark trace; peak memory 5.3 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 216.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--gemma-4-26B-A4B-it-qat-q4_0-mlx-aligned on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 216.9 14.2 5.3 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c0v1z23c",
   "source_url": "https://omlx.ai/benchmarks/performance/c0v1z23c",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-c0v1z23c-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT q4_0 MLX aligned build; oMLX community benchmark trace; peak memory 5.7 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 236.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--gemma-4-26B-A4B-it-qat-q4_0-mlx-aligned on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 236.4 14.8 5.7 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c0v1z23c",
   "source_url": "https://omlx.ai/benchmarks/performance/c0v1z23c",
   "tg_tokens": null,
   "tps": 14.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-v0m84wnj-ctx4096",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.9 GB at this context; duplicate traces of the same session: lc2szcal, qelj2x07; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 449.0,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4-Flash-0731-oQ2e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 4k 449.0 38.8 99.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v0m84wnj",
   "source_url": "https://omlx.ai/benchmarks/performance/v0m84wnj",
   "tg_tokens": null,
   "tps": 38.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-v0m84wnj-ctx16384",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.9 GB at this context; duplicate traces of the same session: lc2szcal, qelj2x07; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 454.3,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4-Flash-0731-oQ2e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 16k 454.3 47.8 99.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v0m84wnj",
   "source_url": "https://omlx.ai/benchmarks/performance/v0m84wnj",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-v0m84wnj-ctx65536",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.3 GB at this context; duplicate traces of the same session: lc2szcal, qelj2x07; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 419.7,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4-Flash-0731-oQ2e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 2bit; 64k 419.7 43.0 100.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v0m84wnj",
   "source_url": "https://omlx.ai/benchmarks/performance/v0m84wnj",
   "tg_tokens": null,
   "tps": 43.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gixq1ry8-ctx131072",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 128k context; peak memory 16.1 GB; TTFT 31715 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4133.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LiquidAI--LFM2.5-24B-A2B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k tokens 4,133 PP tok/s 82.9 TG tok/s 31715 TTFT (ms) 16.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gixq1ry8",
   "source_url": "https://omlx.ai/benchmarks/performance/gixq1ry8",
   "tg_tokens": null,
   "tps": 82.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-qrmtmlsk-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 9v57otm0, n7h1r0za; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 772.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 772.8 91.5 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qrmtmlsk",
   "source_url": "https://omlx.ai/benchmarks/performance/qrmtmlsk",
   "tg_tokens": null,
   "tps": 91.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-qrmtmlsk-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 9v57otm0, n7h1r0za; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 885.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 885.9 87.0 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qrmtmlsk",
   "source_url": "https://omlx.ai/benchmarks/performance/qrmtmlsk",
   "tg_tokens": null,
   "tps": 87.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-qrmtmlsk-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: 9v57otm0, n7h1r0za; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 579.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 32k 579.0 60.8 23.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qrmtmlsk",
   "source_url": "https://omlx.ai/benchmarks/performance/qrmtmlsk",
   "tg_tokens": null,
   "tps": 60.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-d6o31k33-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: 7gdyideq, yu0kohap, zhakoght; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1182.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 1,182 61.8 37.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d6o31k33",
   "source_url": "https://omlx.ai/benchmarks/performance/d6o31k33",
   "tg_tokens": null,
   "tps": 61.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-d6o31k33-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: 7gdyideq, yu0kohap, zhakoght; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1826.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 1,826 75.7 39.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d6o31k33",
   "source_url": "https://omlx.ai/benchmarks/performance/d6o31k33",
   "tg_tokens": null,
   "tps": 75.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-d6o31k33-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.5 GB at this context; duplicate traces of the same session: 7gdyideq, yu0kohap, zhakoght; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1719.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 1,719 89.8 39.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d6o31k33",
   "source_url": "https://omlx.ai/benchmarks/performance/d6o31k33",
   "tg_tokens": null,
   "tps": 89.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-d6o31k33-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.9 GB at this context; duplicate traces of the same session: 7gdyideq, yu0kohap, zhakoght; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1562.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 1,562 105.1 39.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d6o31k33",
   "source_url": "https://omlx.ai/benchmarks/performance/d6o31k33",
   "tg_tokens": null,
   "tps": 105.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-qofk45hn-ctx8192",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.8 GB at this context; duplicate traces of the same session: yht8o45a; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 450.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 450.3 57.4 7.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qofk45hn",
   "source_url": "https://omlx.ai/benchmarks/performance/qofk45hn",
   "tg_tokens": null,
   "tps": 57.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-qofk45hn-ctx16384",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.4 GB at this context; duplicate traces of the same session: yht8o45a; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 432.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 432.7 55.7 8.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qofk45hn",
   "source_url": "https://omlx.ai/benchmarks/performance/qofk45hn",
   "tg_tokens": null,
   "tps": 55.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-savob6kj-ctx16384",
   "model": "Qwen-AgentWorld-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 22.0 GB; TTFT 25049 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 654.1,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen-AgentWorld-35B-A3B-oQ4 on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 16k tokens 654.1 PP tok/s 47.6 TG tok/s 25049 TTFT (ms) 22.0 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/savob6kj",
   "source_url": "https://omlx.ai/benchmarks/performance/savob6kj",
   "tg_tokens": null,
   "tps": 47.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ulxs21x4-ctx1024",
   "model": "Qwen3-Coder-Next-REAP-48B-A3B",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 27.0 GB at this context; duplicate traces of the same session: exxsv66u; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 643.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next-REAP-48B-A3B-4bit-mlx on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 643.0 68.5 27.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ulxs21x4",
   "source_url": "https://omlx.ai/benchmarks/performance/ulxs21x4",
   "tg_tokens": null,
   "tps": 68.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ulxs21x4-ctx4096",
   "model": "Qwen3-Coder-Next-REAP-48B-A3B",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: exxsv66u; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 740.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next-REAP-48B-A3B-4bit-mlx on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 740.7 65.7 27.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ulxs21x4",
   "source_url": "https://omlx.ai/benchmarks/performance/ulxs21x4",
   "tg_tokens": null,
   "tps": 65.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-go85kdls-ctx1024",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 16.0 GB; TTFT 7491 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 136.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-27B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 1k tokens 136.7 PP tok/s 15.6 TG tok/s 7491 TTFT (ms) 16.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/go85kdls",
   "source_url": "https://omlx.ai/benchmarks/performance/go85kdls",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-u4g5r656-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 1085 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 943.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k tokens 943.9 PP tok/s 93.7 TG tok/s 1085 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u4g5r656",
   "source_url": "https://omlx.ai/benchmarks/performance/u4g5r656",
   "tg_tokens": null,
   "tps": 93.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-k082t0o4-ctx16384",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "fp16 MTP build; oMLX community benchmark trace, single run at 16k context; peak memory 25.8 GB; TTFT 16734 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 979.1,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-oQ5-fp16-mtp on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 5bit; 16k tokens 979.1 PP tok/s 60.9 TG tok/s 16734 TTFT (ms) 25.8 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k082t0o4",
   "source_url": "https://omlx.ai/benchmarks/performance/k082t0o4",
   "tg_tokens": null,
   "tps": 60.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-nxm4py6z-ctx16384",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "fp16 MTP build; oMLX community benchmark trace, single run at 16k context; peak memory 25.8 GB; TTFT 20411 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 802.7,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-oQ5-fp16-mtp on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 5bit; 16k tokens 802.7 PP tok/s 55.9 TG tok/s 20411 TTFT (ms) 25.8 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nxm4py6z",
   "source_url": "https://omlx.ai/benchmarks/performance/nxm4py6z",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-c867nxmb-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 4f3ymszm, jpje1enu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2243.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,243 149.7 23.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c867nxmb",
   "source_url": "https://omlx.ai/benchmarks/performance/c867nxmb",
   "tg_tokens": null,
   "tps": 149.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-c867nxmb-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 4f3ymszm, jpje1enu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2253.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,253 196.7 23.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c867nxmb",
   "source_url": "https://omlx.ai/benchmarks/performance/c867nxmb",
   "tg_tokens": null,
   "tps": 196.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-c867nxmb-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: 4f3ymszm, jpje1enu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1737.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,737 200.9 24.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c867nxmb",
   "source_url": "https://omlx.ai/benchmarks/performance/c867nxmb",
   "tg_tokens": null,
   "tps": 200.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-hs6sj02n-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 17.0 GB; TTFT 10344 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 396.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k tokens 396.0 PP tok/s 13.5 TG tok/s 10344 TTFT (ms) 17.0 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hs6sj02n",
   "source_url": "https://omlx.ai/benchmarks/performance/hs6sj02n",
   "tg_tokens": null,
   "tps": 13.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-l6ci30kr-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 17.0 GB; TTFT 10296 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 397.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k tokens 397.8 PP tok/s 13.3 TG tok/s 10296 TTFT (ms) 17.0 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l6ci30kr",
   "source_url": "https://omlx.ai/benchmarks/performance/l6ci30kr",
   "tg_tokens": null,
   "tps": 13.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-9s1m85r3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: 82io97z1, fh15ca5l, zsf22rxq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 497.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 497.4 18.1 15.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9s1m85r3",
   "source_url": "https://omlx.ai/benchmarks/performance/9s1m85r3",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-9s1m85r3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 82io97z1, fh15ca5l, zsf22rxq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 517.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 517.0 17.8 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9s1m85r3",
   "source_url": "https://omlx.ai/benchmarks/performance/9s1m85r3",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-9s1m85r3-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 82io97z1, fh15ca5l, zsf22rxq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 512.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 512.2 17.3 17.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9s1m85r3",
   "source_url": "https://omlx.ai/benchmarks/performance/9s1m85r3",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-9s1m85r3-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 82io97z1, fh15ca5l, zsf22rxq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 465.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 465.3 16.8 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9s1m85r3",
   "source_url": "https://omlx.ai/benchmarks/performance/9s1m85r3",
   "tg_tokens": null,
   "tps": 16.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-2ps2xldt-ctx16384",
   "model": "Qwen3.8-27B-TURBO-Fable5.5-Cold-Fusion-L-Uncensored",
   "notes": "fp16 MTP build; oMLX community benchmark trace, single run at 16k context; peak memory 22.7 GB; TTFT 120261 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 136.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable5.5-Cold-Fusion-L-Uncensored-fp16-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 16k tokens 136.2 PP tok/s 23.3 TG tok/s 120261 TTFT (ms) 22.7 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ps2xldt",
   "source_url": "https://omlx.ai/benchmarks/performance/2ps2xldt",
   "tg_tokens": null,
   "tps": 23.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-aicvycey-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 212.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 1k 212.1 45.1 20.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aicvycey",
   "source_url": "https://omlx.ai/benchmarks/performance/aicvycey",
   "tg_tokens": null,
   "tps": 45.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-aicvycey-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 272.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 4k 272.0 37.0 24.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aicvycey",
   "source_url": "https://omlx.ai/benchmarks/performance/aicvycey",
   "tg_tokens": null,
   "tps": 37.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-t36vven5-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 36.0 GB; TTFT 156219 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 419.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k tokens 419.5 PP tok/s 47.5 TG tok/s 156219 TTFT (ms) 36.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t36vven5",
   "source_url": "https://omlx.ai/benchmarks/performance/t36vven5",
   "tg_tokens": null,
   "tps": 47.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0qtvnqrl-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 36.0 GB; TTFT 155969 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 420.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k tokens 420.2 PP tok/s 50.1 TG tok/s 155969 TTFT (ms) 36.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0qtvnqrl",
   "source_url": "https://omlx.ai/benchmarks/performance/0qtvnqrl",
   "tg_tokens": null,
   "tps": 50.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-i4l111gc-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 761.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 761.5 31.2 100.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i4l111gc",
   "source_url": "https://omlx.ai/benchmarks/performance/i4l111gc",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-i4l111gc-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 936.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 936.8 52.5 101.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i4l111gc",
   "source_url": "https://omlx.ai/benchmarks/performance/i4l111gc",
   "tg_tokens": null,
   "tps": 52.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-i4l111gc-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.4 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1311.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 195k 1,311 58.6 109.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i4l111gc",
   "source_url": "https://omlx.ai/benchmarks/performance/i4l111gc",
   "tg_tokens": null,
   "tps": 58.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-o1noxkqf-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1766.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,766 84.4 70.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o1noxkqf",
   "source_url": "https://omlx.ai/benchmarks/performance/o1noxkqf",
   "tg_tokens": null,
   "tps": 84.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-o1noxkqf-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2767.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,767 80.1 74.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o1noxkqf",
   "source_url": "https://omlx.ai/benchmarks/performance/o1noxkqf",
   "tg_tokens": null,
   "tps": 80.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-168rb3j4-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 101.0 GB at this context; duplicate traces of the same session: kh78ctin, nctk98m7; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 550.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 1k 550.4 55.2 101.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/168rb3j4",
   "source_url": "https://omlx.ai/benchmarks/performance/168rb3j4",
   "tg_tokens": null,
   "tps": 55.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-168rb3j4-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 101.6 GB at this context; duplicate traces of the same session: kh78ctin, nctk98m7; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 772.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 4k 772.7 54.8 101.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/168rb3j4",
   "source_url": "https://omlx.ai/benchmarks/performance/168rb3j4",
   "tg_tokens": null,
   "tps": 54.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-168rb3j4-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 101.8 GB at this context; duplicate traces of the same session: kh78ctin, nctk98m7; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1024.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 8k 1,024 66.3 101.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/168rb3j4",
   "source_url": "https://omlx.ai/benchmarks/performance/168rb3j4",
   "tg_tokens": null,
   "tps": 66.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-2183zszi-ctx1024",
   "model": "gemma-4-12b-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 11.2 GB at this context; duplicate traces of the same session: 3nccer28, ao24u3ar, b8k3tmus, fiir9n3w, xkaj1bcr, z7f4jnya; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 446.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "gemma-4-12B-it-oQ6e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 446.9 54.7 11.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2183zszi",
   "source_url": "https://omlx.ai/benchmarks/performance/2183zszi",
   "tg_tokens": null,
   "tps": 54.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-2183zszi-ctx4096",
   "model": "gemma-4-12b-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 11.6 GB at this context; duplicate traces of the same session: 3nccer28, ao24u3ar, b8k3tmus, fiir9n3w, xkaj1bcr, z7f4jnya; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 446.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "gemma-4-12B-it-oQ6e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 446.3 46.8 11.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2183zszi",
   "source_url": "https://omlx.ai/benchmarks/performance/2183zszi",
   "tg_tokens": null,
   "tps": 46.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-2183zszi-ctx8192",
   "model": "gemma-4-12b-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: 3nccer28, ao24u3ar, b8k3tmus, fiir9n3w, xkaj1bcr, z7f4jnya; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 403.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "gemma-4-12B-it-oQ6e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 403.7 26.4 11.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2183zszi",
   "source_url": "https://omlx.ai/benchmarks/performance/2183zszi",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-2183zszi-ctx16384",
   "model": "gemma-4-12b-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 12.2 GB at this context; duplicate traces of the same session: 3nccer28, ao24u3ar, b8k3tmus, fiir9n3w, xkaj1bcr, z7f4jnya; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 383.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "gemma-4-12B-it-oQ6e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 383.4 25.3 12.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2183zszi",
   "source_url": "https://omlx.ai/benchmarks/performance/2183zszi",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-2183zszi-ctx32768",
   "model": "gemma-4-12b-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 12.9 GB at this context; duplicate traces of the same session: 3nccer28, ao24u3ar, b8k3tmus, fiir9n3w, xkaj1bcr, z7f4jnya; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 383.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "gemma-4-12B-it-oQ6e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 383.7 22.5 12.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2183zszi",
   "source_url": "https://omlx.ai/benchmarks/performance/2183zszi",
   "tg_tokens": null,
   "tps": 22.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-2183zszi-ctx65536",
   "model": "gemma-4-12b-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: 3nccer28, ao24u3ar, b8k3tmus, fiir9n3w, xkaj1bcr, z7f4jnya; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 352.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "gemma-4-12B-it-oQ6e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k 352.4 21.1 14.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2183zszi",
   "source_url": "https://omlx.ai/benchmarks/performance/2183zszi",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-2183zszi-ctx131072",
   "model": "gemma-4-12b-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.5 GB at this context; duplicate traces of the same session: 3nccer28, ao24u3ar, b8k3tmus, fiir9n3w, xkaj1bcr, z7f4jnya; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 275.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "gemma-4-12B-it-oQ6e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 128k 275.5 14.9 17.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2183zszi",
   "source_url": "https://omlx.ai/benchmarks/performance/2183zszi",
   "tg_tokens": null,
   "tps": 14.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-cnvfl32j-ctx1024",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 14.3 GB; TTFT 1271 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 805.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k tokens 805.4 PP tok/s 70.0 TG tok/s 1271 TTFT (ms) 14.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cnvfl32j",
   "source_url": "https://omlx.ai/benchmarks/performance/cnvfl32j",
   "tg_tokens": null,
   "tps": 70.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-nwru4nwi-ctx16384",
   "model": "gemma-4-31B-it-scotoma-2",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 21.0 GB; TTFT 147522 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 111.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-scotoma-2-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 16k tokens 111.1 PP tok/s 11.2 TG tok/s 147522 TTFT (ms) 21.0 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nwru4nwi",
   "source_url": "https://omlx.ai/benchmarks/performance/nwru4nwi",
   "tg_tokens": null,
   "tps": 11.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-mynmjzyc-ctx16384",
   "model": "gemma-4-31B-it-scotoma-2",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 23.0 GB; TTFT 146439 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 111.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-scotoma-2-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 16k tokens 111.9 PP tok/s 14.4 TG tok/s 146439 TTFT (ms) 23.0 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mynmjzyc",
   "source_url": "https://omlx.ai/benchmarks/performance/mynmjzyc",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-fx0ycevp-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: da2xx9ll, hdxc5p6e, niekdj1a, suxsuany, vahpn80u; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 624.1,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 1k 624.1 68.6 11.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fx0ycevp",
   "source_url": "https://omlx.ai/benchmarks/performance/fx0ycevp",
   "tg_tokens": null,
   "tps": 68.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-fx0ycevp-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: da2xx9ll, hdxc5p6e, niekdj1a, suxsuany, vahpn80u; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 609.6,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 4k 609.6 63.8 11.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fx0ycevp",
   "source_url": "https://omlx.ai/benchmarks/performance/fx0ycevp",
   "tg_tokens": null,
   "tps": 63.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-fx0ycevp-ctx8192",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.9 GB at this context; duplicate traces of the same session: da2xx9ll, hdxc5p6e, niekdj1a, suxsuany, vahpn80u; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 595.4,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 8k 595.4 57.8 11.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fx0ycevp",
   "source_url": "https://omlx.ai/benchmarks/performance/fx0ycevp",
   "tg_tokens": null,
   "tps": 57.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-fx0ycevp-ctx16384",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 12.1 GB at this context; duplicate traces of the same session: da2xx9ll, hdxc5p6e, niekdj1a, suxsuany, vahpn80u; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 553.1,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 16k 553.1 50.8 12.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fx0ycevp",
   "source_url": "https://omlx.ai/benchmarks/performance/fx0ycevp",
   "tg_tokens": null,
   "tps": 50.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-fx0ycevp-ctx32768",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 12.4 GB at this context; duplicate traces of the same session: da2xx9ll, hdxc5p6e, niekdj1a, suxsuany, vahpn80u; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 492.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 32k 492.0 40.4 12.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fx0ycevp",
   "source_url": "https://omlx.ai/benchmarks/performance/fx0ycevp",
   "tg_tokens": null,
   "tps": 40.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-fx0ycevp-ctx65536",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 13.2 GB at this context; duplicate traces of the same session: da2xx9ll, hdxc5p6e, niekdj1a, suxsuany, vahpn80u; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 400.5,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 64k 400.5 28.2 13.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fx0ycevp",
   "source_url": "https://omlx.ai/benchmarks/performance/fx0ycevp",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ylcdbm3g-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: ygq9ekep; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 623.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 1k 623.0 68.7 11.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ylcdbm3g",
   "source_url": "https://omlx.ai/benchmarks/performance/ylcdbm3g",
   "tg_tokens": null,
   "tps": 68.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ylcdbm3g-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: ygq9ekep; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 612.7,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 4k 612.7 63.7 11.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ylcdbm3g",
   "source_url": "https://omlx.ai/benchmarks/performance/ylcdbm3g",
   "tg_tokens": null,
   "tps": 63.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-nve63axe-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT q4_0 MLX aligned build; oMLX community benchmark trace; peak memory 5.3 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 216.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--gemma-4-26B-A4B-it-qat-q4_0-mlx-aligned on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 216.9 14.2 5.3 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nve63axe",
   "source_url": "https://omlx.ai/benchmarks/performance/nve63axe",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-nve63axe-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT q4_0 MLX aligned build; oMLX community benchmark trace; peak memory 5.7 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 236.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--gemma-4-26B-A4B-it-qat-q4_0-mlx-aligned on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 236.4 14.8 5.7 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nve63axe",
   "source_url": "https://omlx.ai/benchmarks/performance/nve63axe",
   "tg_tokens": null,
   "tps": 14.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-ds1qc3pd-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT q4_0 MLX aligned build; oMLX community benchmark trace, single run at 4k context; peak memory 17.0 GB; TTFT 6994 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 585.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--gemma-4-26B-A4B-it-qat-q4_0-mlx-aligned on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k tokens 585.7 PP tok/s 20.6 TG tok/s 6994 TTFT (ms) 17.0 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ds1qc3pd",
   "source_url": "https://omlx.ai/benchmarks/performance/ds1qc3pd",
   "tg_tokens": null,
   "tps": 20.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-7c82i2a6-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT q4_0 MLX aligned build; oMLX community benchmark trace, single run at 4k context; peak memory 17.0 GB; TTFT 5359 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 764.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--gemma-4-26B-A4B-it-qat-q4_0-mlx-aligned on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k tokens 764.4 PP tok/s 49.9 TG tok/s 5359 TTFT (ms) 17.0 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7c82i2a6",
   "source_url": "https://omlx.ai/benchmarks/performance/7c82i2a6",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-xlh3v5ru-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT q4_0 MLX aligned build; oMLX community benchmark trace, single run at 4k context; peak memory 9.0 GB; TTFT 17091 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 239.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--gemma-4-26B-A4B-it-qat-q4_0-mlx-aligned on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k tokens 239.7 PP tok/s 24.0 TG tok/s 17091 TTFT (ms) 9.0 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xlh3v5ru",
   "source_url": "https://omlx.ai/benchmarks/performance/xlh3v5ru",
   "tg_tokens": null,
   "tps": 24.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-sln6u2dv-ctx1024",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 146.9 GB at this context; duplicate traces of the same session: 4tq63tru, rivjcra1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 415.7,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "DeepSeek-V4-Flash-0731-MXFP4-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 415.7 28.8 146.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sln6u2dv",
   "source_url": "https://omlx.ai/benchmarks/performance/sln6u2dv",
   "tg_tokens": null,
   "tps": 28.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-sln6u2dv-ctx4096",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 147.9 GB at this context; duplicate traces of the same session: 4tq63tru, rivjcra1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 461.8,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "DeepSeek-V4-Flash-0731-MXFP4-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 461.8 26.5 147.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sln6u2dv",
   "source_url": "https://omlx.ai/benchmarks/performance/sln6u2dv",
   "tg_tokens": null,
   "tps": 26.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-sln6u2dv-ctx16384",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 150.9 GB at this context; duplicate traces of the same session: 4tq63tru, rivjcra1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 433.8,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "DeepSeek-V4-Flash-0731-MXFP4-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 433.8 25.1 150.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sln6u2dv",
   "source_url": "https://omlx.ai/benchmarks/performance/sln6u2dv",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6rr1bqpc-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.2 GB at this context; duplicate traces of the same session: 2tcbt3f6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 337.1,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 337.1 12.1 92.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6rr1bqpc",
   "source_url": "https://omlx.ai/benchmarks/performance/6rr1bqpc",
   "tg_tokens": null,
   "tps": 12.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6rr1bqpc-ctx32768",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.1 GB at this context; duplicate traces of the same session: 2tcbt3f6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 413.9,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 32k 413.9 12.0 92.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6rr1bqpc",
   "source_url": "https://omlx.ai/benchmarks/performance/6rr1bqpc",
   "tg_tokens": null,
   "tps": 12.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tb3ztuga-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 92.0 GB; TTFT 12369 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 82.8,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k tokens 82.8 PP tok/s 10.2 TG tok/s 12369 TTFT (ms) 92.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tb3ztuga",
   "source_url": "https://omlx.ai/benchmarks/performance/tb3ztuga",
   "tg_tokens": null,
   "tps": 10.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5wlyp0fl-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 90.3 GB at this context; duplicate traces of the same session: jo5u0ajm, trao9qe4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 91.4,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 91.4 12.0 90.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5wlyp0fl",
   "source_url": "https://omlx.ai/benchmarks/performance/5wlyp0fl",
   "tg_tokens": null,
   "tps": 12.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5wlyp0fl-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.2 GB at this context; duplicate traces of the same session: jo5u0ajm, trao9qe4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 366.9,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 366.9 12.2 92.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5wlyp0fl",
   "source_url": "https://omlx.ai/benchmarks/performance/5wlyp0fl",
   "tg_tokens": null,
   "tps": 12.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5wlyp0fl-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.2 GB at this context; duplicate traces of the same session: jo5u0ajm, trao9qe4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 431.8,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 431.8 12.6 92.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5wlyp0fl",
   "source_url": "https://omlx.ai/benchmarks/performance/5wlyp0fl",
   "tg_tokens": null,
   "tps": 12.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-y73at9r9-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 90.0 GB at this context; duplicate traces of the same session: g0zs12mv, l6tmylga; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 89.3,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 89.3 11.4 90.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y73at9r9",
   "source_url": "https://omlx.ai/benchmarks/performance/y73at9r9",
   "tg_tokens": null,
   "tps": 11.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-y73at9r9-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.2 GB at this context; duplicate traces of the same session: g0zs12mv, l6tmylga; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 325.4,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 325.4 12.4 92.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y73at9r9",
   "source_url": "https://omlx.ai/benchmarks/performance/y73at9r9",
   "tg_tokens": null,
   "tps": 12.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-y73at9r9-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.2 GB at this context; duplicate traces of the same session: g0zs12mv, l6tmylga; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 425.9,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 425.9 13.7 92.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y73at9r9",
   "source_url": "https://omlx.ai/benchmarks/performance/y73at9r9",
   "tg_tokens": null,
   "tps": 13.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-cyqe9kui-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: npr67vfm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2116.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 2,116 124.0 13.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cyqe9kui",
   "source_url": "https://omlx.ai/benchmarks/performance/cyqe9kui",
   "tg_tokens": null,
   "tps": 124.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-cyqe9kui-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.9 GB at this context; duplicate traces of the same session: npr67vfm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1931.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,931 112.8 13.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cyqe9kui",
   "source_url": "https://omlx.ai/benchmarks/performance/cyqe9kui",
   "tg_tokens": null,
   "tps": 112.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-1yyp8du0-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: y2ic7dm0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1796.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 1,796 86.0 24.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1yyp8du0",
   "source_url": "https://omlx.ai/benchmarks/performance/1yyp8du0",
   "tg_tokens": null,
   "tps": 86.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-1yyp8du0-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: y2ic7dm0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1703.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 8bit; 16k 1,703 80.5 25.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1yyp8du0",
   "source_url": "https://omlx.ai/benchmarks/performance/1yyp8du0",
   "tg_tokens": null,
   "tps": 80.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-e49q37yx-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 98xzf2w2; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 486.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k 486.5 49.0 19.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e49q37yx",
   "source_url": "https://omlx.ai/benchmarks/performance/e49q37yx",
   "tg_tokens": null,
   "tps": 49.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-e49q37yx-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 98xzf2w2; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 529.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 529.3 47.3 20.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e49q37yx",
   "source_url": "https://omlx.ai/benchmarks/performance/e49q37yx",
   "tg_tokens": null,
   "tps": 47.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-a9xocu2j-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "REAP-19B build; oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: 0e2n9d6t; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 485.3,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit-REAP-19B on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k 485.3 38.8 14.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a9xocu2j",
   "source_url": "https://omlx.ai/benchmarks/performance/a9xocu2j",
   "tg_tokens": null,
   "tps": 38.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-a9xocu2j-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "REAP-19B build; oMLX community benchmark trace; peak memory 15.3 GB at this context; duplicate traces of the same session: 0e2n9d6t; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 541.8,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit-REAP-19B on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 541.8 38.2 15.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a9xocu2j",
   "source_url": "https://omlx.ai/benchmarks/performance/a9xocu2j",
   "tg_tokens": null,
   "tps": 38.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-z07w655e-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "REAP-19B build; oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 507.4,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit-REAP-19B on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k 507.4 33.9 12.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z07w655e",
   "source_url": "https://omlx.ai/benchmarks/performance/z07w655e",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-z07w655e-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "REAP-19B build; oMLX community benchmark trace; peak memory 13.3 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 539.3,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit-REAP-19B on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 539.3 33.2 13.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z07w655e",
   "source_url": "https://omlx.ai/benchmarks/performance/z07w655e",
   "tg_tokens": null,
   "tps": 33.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-6gte7bdm-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.9 GB at this context; duplicate traces of the same session: kxrj9bgl, uk6pw4w2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1856.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 1k 1,856 95.6 30.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6gte7bdm",
   "source_url": "https://omlx.ai/benchmarks/performance/6gte7bdm",
   "tg_tokens": null,
   "tps": 95.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-6gte7bdm-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.7 GB at this context; duplicate traces of the same session: kxrj9bgl, uk6pw4w2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2589.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 4k 2,589 103.4 32.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6gte7bdm",
   "source_url": "https://omlx.ai/benchmarks/performance/6gte7bdm",
   "tg_tokens": null,
   "tps": 103.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-6gte7bdm-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.9 GB at this context; duplicate traces of the same session: kxrj9bgl, uk6pw4w2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2441.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 8k 2,441 107.7 32.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6gte7bdm",
   "source_url": "https://omlx.ai/benchmarks/performance/6gte7bdm",
   "tg_tokens": null,
   "tps": 107.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-q7dule5h-ctx4096",
   "model": "Qwen3.5-2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 2.2 GB at this context; duplicate traces of the same session: ue0o3sdd; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4510.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-2B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 16 GB \u00b7 4bit; 4k 4,510 113.3 2.2 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q7dule5h",
   "source_url": "https://omlx.ai/benchmarks/performance/q7dule5h",
   "tg_tokens": null,
   "tps": 113.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-q7dule5h-ctx16384",
   "model": "Qwen3.5-2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 2.2 GB at this context; duplicate traces of the same session: ue0o3sdd; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4625.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-2B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 16 GB \u00b7 4bit; 16k 4,625 101.4 2.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q7dule5h",
   "source_url": "https://omlx.ai/benchmarks/performance/q7dule5h",
   "tg_tokens": null,
   "tps": 101.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-3wo5nvvx-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 21.4 GB; TTFT 667 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1535.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 1,535 PP tok/s 156.9 TG tok/s 667 TTFT (ms) 21.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3wo5nvvx",
   "source_url": "https://omlx.ai/benchmarks/performance/3wo5nvvx",
   "tg_tokens": null,
   "tps": 156.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-7fme8ewq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 1ea8ydb3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 467.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 467.7 18.1 16.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7fme8ewq",
   "source_url": "https://omlx.ai/benchmarks/performance/7fme8ewq",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-7fme8ewq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 1ea8ydb3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 504.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 504.9 17.8 17.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7fme8ewq",
   "source_url": "https://omlx.ai/benchmarks/performance/7fme8ewq",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kxj0so9m-ctx1024",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 823.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 823.2 46.9 20.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kxj0so9m",
   "source_url": "https://omlx.ai/benchmarks/performance/kxj0so9m",
   "tg_tokens": null,
   "tps": 46.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-kxj0so9m-ctx4096",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1245.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,245 56.8 21.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kxj0so9m",
   "source_url": "https://omlx.ai/benchmarks/performance/kxj0so9m",
   "tg_tokens": null,
   "tps": 56.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kxj0so9m-ctx8192",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1364.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,364 46.3 22.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kxj0so9m",
   "source_url": "https://omlx.ai/benchmarks/performance/kxj0so9m",
   "tg_tokens": null,
   "tps": 46.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-innwd98u-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: qu60amx0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 258.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 258.6 23.5 21.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/innwd98u",
   "source_url": "https://omlx.ai/benchmarks/performance/innwd98u",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-innwd98u-ctx16384",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: qu60amx0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 233.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 233.9 22.2 22.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/innwd98u",
   "source_url": "https://omlx.ai/benchmarks/performance/innwd98u",
   "tg_tokens": null,
   "tps": 22.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3axkcvil-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: ububpriv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 907.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 907.5 26.7 20.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3axkcvil",
   "source_url": "https://omlx.ai/benchmarks/performance/3axkcvil",
   "tg_tokens": null,
   "tps": 26.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3axkcvil-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: ububpriv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 980.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 980.7 26.8 21.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3axkcvil",
   "source_url": "https://omlx.ai/benchmarks/performance/3axkcvil",
   "tg_tokens": null,
   "tps": 26.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-qlas8vcj-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 25.6 GB at this context; duplicate traces of the same session: kwdkfmh2, vwjahffh; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 276.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 276.4 55.1 25.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qlas8vcj",
   "source_url": "https://omlx.ai/benchmarks/performance/qlas8vcj",
   "tg_tokens": null,
   "tps": 55.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-qlas8vcj-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 29.1 GB at this context; duplicate traces of the same session: kwdkfmh2, vwjahffh; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 356.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 356.5 62.5 29.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qlas8vcj",
   "source_url": "https://omlx.ai/benchmarks/performance/qlas8vcj",
   "tg_tokens": null,
   "tps": 62.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-qlas8vcj-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: kwdkfmh2, vwjahffh; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 351.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 351.2 61.1 29.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qlas8vcj",
   "source_url": "https://omlx.ai/benchmarks/performance/qlas8vcj",
   "tg_tokens": null,
   "tps": 61.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-f95kn7y2-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 212.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 1k 212.1 45.1 20.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f95kn7y2",
   "source_url": "https://omlx.ai/benchmarks/performance/f95kn7y2",
   "tg_tokens": null,
   "tps": 45.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-f95kn7y2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 272.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 4k 272.0 37.0 24.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f95kn7y2",
   "source_url": "https://omlx.ai/benchmarks/performance/f95kn7y2",
   "tg_tokens": null,
   "tps": 37.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-54l4a7s8-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: r5y7caey; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 1k 220.8 30.9 19.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/54l4a7s8",
   "source_url": "https://omlx.ai/benchmarks/performance/54l4a7s8",
   "tg_tokens": null,
   "tps": 30.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-54l4a7s8-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: r5y7caey; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 283.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 4k 283.3 29.5 22.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/54l4a7s8",
   "source_url": "https://omlx.ai/benchmarks/performance/54l4a7s8",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-cpxtc42v-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: ograehth; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 217.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 1k 217.2 30.9 19.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cpxtc42v",
   "source_url": "https://omlx.ai/benchmarks/performance/cpxtc42v",
   "tg_tokens": null,
   "tps": 30.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-cpxtc42v-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: ograehth; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 283.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 4k 283.5 29.6 22.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cpxtc42v",
   "source_url": "https://omlx.ai/benchmarks/performance/cpxtc42v",
   "tg_tokens": null,
   "tps": 29.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-wr6r2pkb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: mzx9365a; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 207.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 1k 207.4 31.2 20.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wr6r2pkb",
   "source_url": "https://omlx.ai/benchmarks/performance/wr6r2pkb",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-wr6r2pkb-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: mzx9365a; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 271.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 4k 271.5 28.9 24.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wr6r2pkb",
   "source_url": "https://omlx.ai/benchmarks/performance/wr6r2pkb",
   "tg_tokens": null,
   "tps": 28.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-kozmbi7n-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 18.5 GB; TTFT 4673 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 219.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 219.1 PP tok/s 45.9 TG tok/s 4673 TTFT (ms) 18.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kozmbi7n",
   "source_url": "https://omlx.ai/benchmarks/performance/kozmbi7n",
   "tg_tokens": null,
   "tps": 45.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-uj47ccs7-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 2wdxjbo9, p734hem7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 258.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 258.8 28.1 16.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uj47ccs7",
   "source_url": "https://omlx.ai/benchmarks/performance/uj47ccs7",
   "tg_tokens": null,
   "tps": 28.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-uj47ccs7-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 2wdxjbo9, p734hem7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 262.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 262.4 27.3 17.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uj47ccs7",
   "source_url": "https://omlx.ai/benchmarks/performance/uj47ccs7",
   "tg_tokens": null,
   "tps": 27.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-uj47ccs7-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 2wdxjbo9, p734hem7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 245.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 8k 245.8 26.6 17.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uj47ccs7",
   "source_url": "https://omlx.ai/benchmarks/performance/uj47ccs7",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-lb3jqxz6-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "fp16 build; oMLX community benchmark trace, single run at 1k context; peak memory 43.1 GB; TTFT 17340 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 59.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k tokens 59.1 PP tok/s 3.7 TG tok/s 17340 TTFT (ms) 43.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lb3jqxz6",
   "source_url": "https://omlx.ai/benchmarks/performance/lb3jqxz6",
   "tg_tokens": null,
   "tps": 3.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-pf756x35-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "fp16 build; oMLX community benchmark trace, single run at 1k context; peak memory 43.1 GB; TTFT 17508 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 58.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k tokens 58.5 PP tok/s 3.7 TG tok/s 17508 TTFT (ms) 43.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pf756x35",
   "source_url": "https://omlx.ai/benchmarks/performance/pf756x35",
   "tg_tokens": null,
   "tps": 3.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-8whcib2c-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "fp16 build; oMLX community benchmark trace, single run at 1k context; peak memory 52.6 GB; TTFT 16375 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 62.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k tokens 62.5 PP tok/s 5.3 TG tok/s 16375 TTFT (ms) 52.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8whcib2c",
   "source_url": "https://omlx.ai/benchmarks/performance/8whcib2c",
   "tg_tokens": null,
   "tps": 5.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-eosz0glc-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: 9164jebq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 761.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 761.5 31.2 100.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eosz0glc",
   "source_url": "https://omlx.ai/benchmarks/performance/eosz0glc",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-eosz0glc-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: 9164jebq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 936.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 936.8 52.5 101.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eosz0glc",
   "source_url": "https://omlx.ai/benchmarks/performance/eosz0glc",
   "tg_tokens": null,
   "tps": 52.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-eosz0glc-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.4 GB at this context; duplicate traces of the same session: 9164jebq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1311.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 195k 1,311 58.6 109.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eosz0glc",
   "source_url": "https://omlx.ai/benchmarks/performance/eosz0glc",
   "tg_tokens": null,
   "tps": 58.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-g3iv7hxv-ctx8192",
   "model": "gemma-4-E2B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.3 GB at this context; duplicate traces of the same session: pqmuze7z; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 8370.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E2B-it-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 8,370 234.3 4.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g3iv7hxv",
   "source_url": "https://omlx.ai/benchmarks/performance/g3iv7hxv",
   "tg_tokens": null,
   "tps": 234.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-g3iv7hxv-ctx16384",
   "model": "gemma-4-E2B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.3 GB at this context; duplicate traces of the same session: pqmuze7z; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 8227.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E2B-it-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 16k 8,227 203.3 4.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g3iv7hxv",
   "source_url": "https://omlx.ai/benchmarks/performance/g3iv7hxv",
   "tg_tokens": null,
   "tps": 203.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-v0ckayop-ctx1024",
   "model": "gemma-4-E2B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.1 GB at this context; duplicate traces of the same session: pn407cki; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6115.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E2B-it-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 6,115 473.7 4.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v0ckayop",
   "source_url": "https://omlx.ai/benchmarks/performance/v0ckayop",
   "tg_tokens": null,
   "tps": 473.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-v0ckayop-ctx4096",
   "model": "gemma-4-E2B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: pn407cki; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 8830.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E2B-it-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 8,830 326.4 4.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v0ckayop",
   "source_url": "https://omlx.ai/benchmarks/performance/v0ckayop",
   "tg_tokens": null,
   "tps": 326.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-g4gcsba2-ctx1024",
   "model": "Occamy-1.0",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: 2q8yz4uw, xxkdty2g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4291.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "occamy-1.0-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 4,291 162.7 20.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g4gcsba2",
   "source_url": "https://omlx.ai/benchmarks/performance/g4gcsba2",
   "tg_tokens": null,
   "tps": 162.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-g4gcsba2-ctx8192",
   "model": "Occamy-1.0",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 2q8yz4uw, xxkdty2g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5795.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "occamy-1.0-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 5,795 123.5 21.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g4gcsba2",
   "source_url": "https://omlx.ai/benchmarks/performance/g4gcsba2",
   "tg_tokens": null,
   "tps": 123.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-g4gcsba2-ctx131072",
   "model": "Occamy-1.0",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: 2q8yz4uw, xxkdty2g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1924.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "occamy-1.0-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 1,924 69.6 24.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g4gcsba2",
   "source_url": "https://omlx.ai/benchmarks/performance/g4gcsba2",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-mpwtn41f-ctx1024",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: 4x61eb9t; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 65.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "AREX-2-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k 65.5 6.8 15.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mpwtn41f",
   "source_url": "https://omlx.ai/benchmarks/performance/mpwtn41f",
   "tg_tokens": null,
   "tps": 6.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-mpwtn41f-ctx4096",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: 4x61eb9t; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 67.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "AREX-2-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 67.5 6.7 17.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mpwtn41f",
   "source_url": "https://omlx.ai/benchmarks/performance/mpwtn41f",
   "tg_tokens": null,
   "tps": 6.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9i9zpv9b-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 94.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 82.9,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 82.9 11.4 94.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9i9zpv9b",
   "source_url": "https://omlx.ai/benchmarks/performance/9i9zpv9b",
   "tg_tokens": null,
   "tps": 11.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9i9zpv9b-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 95.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 364.7,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 364.7 12.0 95.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9i9zpv9b",
   "source_url": "https://omlx.ai/benchmarks/performance/9i9zpv9b",
   "tg_tokens": null,
   "tps": 12.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9i9zpv9b-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 394.7,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 394.7 12.9 96.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9i9zpv9b",
   "source_url": "https://omlx.ai/benchmarks/performance/9i9zpv9b",
   "tg_tokens": null,
   "tps": 12.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-pc4lfk1g-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.1 GB at this context; duplicate traces of the same session: u2hwrfkm; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 617.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k 617.6 60.4 13.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pc4lfk1g",
   "source_url": "https://omlx.ai/benchmarks/performance/pc4lfk1g",
   "tg_tokens": null,
   "tps": 60.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-pc4lfk1g-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.5 GB at this context; duplicate traces of the same session: u2hwrfkm; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 654.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 654.6 58.1 13.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pc4lfk1g",
   "source_url": "https://omlx.ai/benchmarks/performance/pc4lfk1g",
   "tg_tokens": null,
   "tps": 58.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-a3qmm2rm-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 4k context; peak memory 13.6 GB; TTFT 7384 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 554.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k tokens 554.7 PP tok/s 45.2 TG tok/s 7384 TTFT (ms) 13.6 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a3qmm2rm",
   "source_url": "https://omlx.ai/benchmarks/performance/a3qmm2rm",
   "tg_tokens": null,
   "tps": 45.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-gbz269o7-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "REAP-19B build; oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 507.4,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit-REAP-19B on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k 507.4 33.9 12.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gbz269o7",
   "source_url": "https://omlx.ai/benchmarks/performance/gbz269o7",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-gbz269o7-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "REAP-19B build; oMLX community benchmark trace; peak memory 13.3 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 539.3,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit-REAP-19B on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 539.3 33.2 13.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gbz269o7",
   "source_url": "https://omlx.ai/benchmarks/performance/gbz269o7",
   "tg_tokens": null,
   "tps": 33.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-mp2akgln-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: cdjhch37; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 468.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 24 GB \u00b7 4bit; 1k 468.5 50.5 5.9 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mp2akgln",
   "source_url": "https://omlx.ai/benchmarks/performance/mp2akgln",
   "tg_tokens": null,
   "tps": 50.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-mp2akgln-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: cdjhch37; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 480.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 24 GB \u00b7 4bit; 4k 480.6 49.5 6.3 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mp2akgln",
   "source_url": "https://omlx.ai/benchmarks/performance/mp2akgln",
   "tg_tokens": null,
   "tps": 49.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-3epkn1wy-ctx1024",
   "model": "Qwen2.5-Coder-3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 2.2 GB at this context; duplicate traces of the same session: o32918de; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2045.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-3B-Instruct-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 2,045 60.3 2.2 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3epkn1wy",
   "source_url": "https://omlx.ai/benchmarks/performance/3epkn1wy",
   "tg_tokens": null,
   "tps": 60.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-3epkn1wy-ctx4096",
   "model": "Qwen2.5-Coder-3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 2.3 GB at this context; duplicate traces of the same session: o32918de; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2066.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-3B-Instruct-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 2,066 56.4 2.3 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3epkn1wy",
   "source_url": "https://omlx.ai/benchmarks/performance/3epkn1wy",
   "tg_tokens": null,
   "tps": 56.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-4h6tmuiu-ctx1024",
   "model": "Qwen2.5-Coder-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 4.5 GB at this context; duplicate traces of the same session: 4kwatc7g; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 926.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-7B-Instruct-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 926.0 27.9 4.5 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4h6tmuiu",
   "source_url": "https://omlx.ai/benchmarks/performance/4h6tmuiu",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-4h6tmuiu-ctx4096",
   "model": "Qwen2.5-Coder-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: 4kwatc7g; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 893.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-7B-Instruct-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 893.8 26.6 4.9 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4h6tmuiu",
   "source_url": "https://omlx.ai/benchmarks/performance/4h6tmuiu",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ohmf43hr-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: 6swd0y38; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 380.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 380.2 50.5 5.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ohmf43hr",
   "source_url": "https://omlx.ai/benchmarks/performance/ohmf43hr",
   "tg_tokens": null,
   "tps": 50.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ohmf43hr-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: 6swd0y38; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 390.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 390.1 49.2 6.3 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ohmf43hr",
   "source_url": "https://omlx.ai/benchmarks/performance/ohmf43hr",
   "tg_tokens": null,
   "tps": 49.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-yk23y7tv-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 7hg9npc3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1284.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k 1,284 54.5 16.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yk23y7tv",
   "source_url": "https://omlx.ai/benchmarks/performance/yk23y7tv",
   "tg_tokens": null,
   "tps": 54.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-yk23y7tv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 7hg9npc3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1462.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 4k 1,462 53.1 17.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yk23y7tv",
   "source_url": "https://omlx.ai/benchmarks/performance/yk23y7tv",
   "tg_tokens": null,
   "tps": 53.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fazn1vwv-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 0h01enlp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 884.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 884.8 62.2 17.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fazn1vwv",
   "source_url": "https://omlx.ai/benchmarks/performance/fazn1vwv",
   "tg_tokens": null,
   "tps": 62.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fazn1vwv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: 0h01enlp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1000.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,000 84.3 18.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fazn1vwv",
   "source_url": "https://omlx.ai/benchmarks/performance/fazn1vwv",
   "tg_tokens": null,
   "tps": 84.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-439mcyqp-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 8sxmecct; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1137.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 1k 1,137 32.5 28.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/439mcyqp",
   "source_url": "https://omlx.ai/benchmarks/performance/439mcyqp",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-439mcyqp-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: 8sxmecct; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1386.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 4k 1,386 32.1 29.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/439mcyqp",
   "source_url": "https://omlx.ai/benchmarks/performance/439mcyqp",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9ndimw59-ctx1024",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: p0afh661; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 823.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 823.2 46.9 20.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ndimw59",
   "source_url": "https://omlx.ai/benchmarks/performance/9ndimw59",
   "tg_tokens": null,
   "tps": 46.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9ndimw59-ctx4096",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: p0afh661; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1245.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,245 56.8 21.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ndimw59",
   "source_url": "https://omlx.ai/benchmarks/performance/9ndimw59",
   "tg_tokens": null,
   "tps": 56.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9ndimw59-ctx8192",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: p0afh661; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1364.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,364 46.3 22.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ndimw59",
   "source_url": "https://omlx.ai/benchmarks/performance/9ndimw59",
   "tg_tokens": null,
   "tps": 46.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-bq6gofv8-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 28.5 GB; TTFT 2637 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 388.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k tokens 388.3 PP tok/s 8.8 TG tok/s 2637 TTFT (ms) 28.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bq6gofv8",
   "source_url": "https://omlx.ai/benchmarks/performance/bq6gofv8",
   "tg_tokens": null,
   "tps": 8.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-j4hw0lf2-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 73sy8g12; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 897.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 897.0 26.4 20.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j4hw0lf2",
   "source_url": "https://omlx.ai/benchmarks/performance/j4hw0lf2",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-j4hw0lf2-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 73sy8g12; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 988.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 988.4 25.9 21.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j4hw0lf2",
   "source_url": "https://omlx.ai/benchmarks/performance/j4hw0lf2",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-uvssrugn-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 59e6rx9r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 897.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 897.8 26.5 20.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uvssrugn",
   "source_url": "https://omlx.ai/benchmarks/performance/uvssrugn",
   "tg_tokens": null,
   "tps": 26.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-uvssrugn-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 59e6rx9r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 992.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 992.9 24.9 21.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uvssrugn",
   "source_url": "https://omlx.ai/benchmarks/performance/uvssrugn",
   "tg_tokens": null,
   "tps": 24.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-hk16ftug-ctx1024",
   "model": "Qwen3.8-27B-MindMeld-AREX",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 15.0 GB at this context; duplicate traces of the same session: gajjmm9i; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 65.8,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-MindMeld-AREX-mxfp4-mlx on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k 65.8 7.1 15.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hk16ftug",
   "source_url": "https://omlx.ai/benchmarks/performance/hk16ftug",
   "tg_tokens": null,
   "tps": 7.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-hk16ftug-ctx4096",
   "model": "Qwen3.8-27B-MindMeld-AREX",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: gajjmm9i; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 65.7,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-MindMeld-AREX-mxfp4-mlx on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 65.7 6.8 16.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hk16ftug",
   "source_url": "https://omlx.ai/benchmarks/performance/hk16ftug",
   "tg_tokens": null,
   "tps": 6.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-rgp38u28-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.1,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 323.1 60.0 23.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rgp38u28",
   "source_url": "https://omlx.ai/benchmarks/performance/rgp38u28",
   "tg_tokens": null,
   "tps": 60.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-rgp38u28-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 303.8,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 16k 303.8 67.1 25.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rgp38u28",
   "source_url": "https://omlx.ai/benchmarks/performance/rgp38u28",
   "tg_tokens": null,
   "tps": 67.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-0pron4j5-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 7b1lvuhi, n6sgjpjo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 252.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 252.5 67.1 18.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0pron4j5",
   "source_url": "https://omlx.ai/benchmarks/performance/0pron4j5",
   "tg_tokens": null,
   "tps": 67.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-0pron4j5-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 7b1lvuhi, n6sgjpjo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 255.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 255.4 51.3 19.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0pron4j5",
   "source_url": "https://omlx.ai/benchmarks/performance/0pron4j5",
   "tg_tokens": null,
   "tps": 51.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-0pron4j5-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 7b1lvuhi, n6sgjpjo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 245.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 8k 245.0 65.0 19.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0pron4j5",
   "source_url": "https://omlx.ai/benchmarks/performance/0pron4j5",
   "tg_tokens": null,
   "tps": 65.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ri9z8gea-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 0w9oxati, fm6z3nr4, qee34dkr; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 451.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 451.1 27.2 18.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ri9z8gea",
   "source_url": "https://omlx.ai/benchmarks/performance/ri9z8gea",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ri9z8gea-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 0w9oxati, fm6z3nr4, qee34dkr; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 449.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 449.3 20.8 19.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ri9z8gea",
   "source_url": "https://omlx.ai/benchmarks/performance/ri9z8gea",
   "tg_tokens": null,
   "tps": 20.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ri9z8gea-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: 0w9oxati, fm6z3nr4, qee34dkr; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 349.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 349.8 12.1 23.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ri9z8gea",
   "source_url": "https://omlx.ai/benchmarks/performance/ri9z8gea",
   "tg_tokens": null,
   "tps": 12.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ri9z8gea-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: 0w9oxati, fm6z3nr4, qee34dkr; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 251.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 128k 251.4 9.0 28.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ri9z8gea",
   "source_url": "https://omlx.ai/benchmarks/performance/ri9z8gea",
   "tg_tokens": null,
   "tps": 9.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-y1q8r5ex-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: jecn7a8s; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1117.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 1k 1,117 31.7 28.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y1q8r5ex",
   "source_url": "https://omlx.ai/benchmarks/performance/y1q8r5ex",
   "tg_tokens": null,
   "tps": 31.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-y1q8r5ex-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: jecn7a8s; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1365.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 4k 1,365 32.3 29.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y1q8r5ex",
   "source_url": "https://omlx.ai/benchmarks/performance/y1q8r5ex",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4a3paifg-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.5 GB at this context; duplicate traces of the same session: hmj8k31h, wvndeyg5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1694.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,694 66.5 69.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4a3paifg",
   "source_url": "https://omlx.ai/benchmarks/performance/4a3paifg",
   "tg_tokens": null,
   "tps": 66.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4a3paifg-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.1 GB at this context; duplicate traces of the same session: hmj8k31h, wvndeyg5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2486.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,486 69.6 70.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4a3paifg",
   "source_url": "https://omlx.ai/benchmarks/performance/4a3paifg",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4a3paifg-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.4 GB at this context; duplicate traces of the same session: hmj8k31h, wvndeyg5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2966.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,966 69.0 71.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4a3paifg",
   "source_url": "https://omlx.ai/benchmarks/performance/4a3paifg",
   "tg_tokens": null,
   "tps": 69.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rpqvmj05-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 63xec6mu, ueqvnm50; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 867.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 867.9 55.3 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rpqvmj05",
   "source_url": "https://omlx.ai/benchmarks/performance/rpqvmj05",
   "tg_tokens": null,
   "tps": 55.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-rpqvmj05-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 63xec6mu, ueqvnm50; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1243.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,243 61.7 19.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rpqvmj05",
   "source_url": "https://omlx.ai/benchmarks/performance/rpqvmj05",
   "tg_tokens": null,
   "tps": 61.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-rpqvmj05-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 63xec6mu, ueqvnm50; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1333.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,333 73.1 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rpqvmj05",
   "source_url": "https://omlx.ai/benchmarks/performance/rpqvmj05",
   "tg_tokens": null,
   "tps": 73.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-22zcsl6k-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: aa618ont, cpztrsr1, disarmua; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2772.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 2,772 120.2 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/22zcsl6k",
   "source_url": "https://omlx.ai/benchmarks/performance/22zcsl6k",
   "tg_tokens": null,
   "tps": 120.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-22zcsl6k-ctx8192",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: aa618ont, cpztrsr1, disarmua; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5723.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 5,723 312.1 22.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/22zcsl6k",
   "source_url": "https://omlx.ai/benchmarks/performance/22zcsl6k",
   "tg_tokens": null,
   "tps": 312.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-22zcsl6k-ctx32768",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: aa618ont, cpztrsr1, disarmua; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4276.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 4,276 192.3 22.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/22zcsl6k",
   "source_url": "https://omlx.ai/benchmarks/performance/22zcsl6k",
   "tg_tokens": null,
   "tps": 192.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-22zcsl6k-ctx131072",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: aa618ont, cpztrsr1, disarmua; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1916.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 1,916 168.8 24.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/22zcsl6k",
   "source_url": "https://omlx.ai/benchmarks/performance/22zcsl6k",
   "tg_tokens": null,
   "tps": 168.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-28d8kn7u-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4287.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 4,287 208.8 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/28d8kn7u",
   "source_url": "https://omlx.ai/benchmarks/performance/28d8kn7u",
   "tg_tokens": null,
   "tps": 208.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-28d8kn7u-ctx8192",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5896.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 5,896 303.1 22.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/28d8kn7u",
   "source_url": "https://omlx.ai/benchmarks/performance/28d8kn7u",
   "tg_tokens": null,
   "tps": 303.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-28d8kn7u-ctx32768",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4373.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 4,373 202.7 22.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/28d8kn7u",
   "source_url": "https://omlx.ai/benchmarks/performance/28d8kn7u",
   "tg_tokens": null,
   "tps": 202.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-28d8kn7u-ctx131072",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1958.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 1,958 166.1 24.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/28d8kn7u",
   "source_url": "https://omlx.ai/benchmarks/performance/28d8kn7u",
   "tg_tokens": null,
   "tps": 166.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-ql0d0ixq-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 15.5 GB; TTFT 19000 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 215.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k tokens 215.6 PP tok/s 28.2 TG tok/s 19000 TTFT (ms) 15.5 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ql0d0ixq",
   "source_url": "https://omlx.ai/benchmarks/performance/ql0d0ixq",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-f8r9f65p-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 11.3 GB; TTFT 14069 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 291.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k tokens 291.1 PP tok/s 19.3 TG tok/s 14069 TTFT (ms) 11.3 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f8r9f65p",
   "source_url": "https://omlx.ai/benchmarks/performance/f8r9f65p",
   "tg_tokens": null,
   "tps": 19.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-52piky80-ctx1024",
   "model": "Occamy-1.0",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: 46w1gchc, kieay8lg, qw76htl3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3586.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "occamy-1.0-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 3,586 171.2 20.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/52piky80",
   "source_url": "https://omlx.ai/benchmarks/performance/52piky80",
   "tg_tokens": null,
   "tps": 171.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-52piky80-ctx8192",
   "model": "Occamy-1.0",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 46w1gchc, kieay8lg, qw76htl3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5454.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "occamy-1.0-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 5,454 106.2 21.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/52piky80",
   "source_url": "https://omlx.ai/benchmarks/performance/52piky80",
   "tg_tokens": null,
   "tps": 106.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-52piky80-ctx32768",
   "model": "Occamy-1.0",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 46w1gchc, kieay8lg, qw76htl3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4109.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "occamy-1.0-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 4,109 117.2 21.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/52piky80",
   "source_url": "https://omlx.ai/benchmarks/performance/52piky80",
   "tg_tokens": null,
   "tps": 117.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-52piky80-ctx131072",
   "model": "Occamy-1.0",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: 46w1gchc, kieay8lg, qw76htl3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1918.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "occamy-1.0-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 1,918 67.9 24.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/52piky80",
   "source_url": "https://omlx.ai/benchmarks/performance/52piky80",
   "tg_tokens": null,
   "tps": 67.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-22pw7uzk-ctx1024",
   "model": "Occamy-1.0",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: 7178bmsa, e6d4g4ly; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 568.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "occamy-1.0-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 568.0 150.8 20.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/22pw7uzk",
   "source_url": "https://omlx.ai/benchmarks/performance/22pw7uzk",
   "tg_tokens": null,
   "tps": 150.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-22pw7uzk-ctx8192",
   "model": "Occamy-1.0",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 7178bmsa, e6d4g4ly; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3302.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "occamy-1.0-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 3,302 103.7 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/22pw7uzk",
   "source_url": "https://omlx.ai/benchmarks/performance/22pw7uzk",
   "tg_tokens": null,
   "tps": 103.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-22pw7uzk-ctx32768",
   "model": "Occamy-1.0",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 7178bmsa, e6d4g4ly; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4083.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "occamy-1.0-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 4,083 149.3 21.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/22pw7uzk",
   "source_url": "https://omlx.ai/benchmarks/performance/22pw7uzk",
   "tg_tokens": null,
   "tps": 149.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-22pw7uzk-ctx131072",
   "model": "Occamy-1.0",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: 7178bmsa, e6d4g4ly; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1882.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "occamy-1.0-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 1,882 69.8 24.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/22pw7uzk",
   "source_url": "https://omlx.ai/benchmarks/performance/22pw7uzk",
   "tg_tokens": null,
   "tps": 69.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ryklkx3w-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 94.0 GB at this context; duplicate traces of the same session: gd545pyr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 82.9,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 82.9 11.4 94.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ryklkx3w",
   "source_url": "https://omlx.ai/benchmarks/performance/ryklkx3w",
   "tg_tokens": null,
   "tps": 11.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ryklkx3w-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 95.7 GB at this context; duplicate traces of the same session: gd545pyr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 364.7,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 364.7 12.0 95.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ryklkx3w",
   "source_url": "https://omlx.ai/benchmarks/performance/ryklkx3w",
   "tg_tokens": null,
   "tps": 12.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ryklkx3w-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.0 GB at this context; duplicate traces of the same session: gd545pyr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 394.7,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 394.7 12.9 96.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ryklkx3w",
   "source_url": "https://omlx.ai/benchmarks/performance/ryklkx3w",
   "tg_tokens": null,
   "tps": 12.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-eubkdctp-ctx1024",
   "model": "Qwen3.5-2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 1.9 GB at this context; duplicate traces of the same session: 1dsvww5m; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1105.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-2B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 1,105 85.5 1.9 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eubkdctp",
   "source_url": "https://omlx.ai/benchmarks/performance/eubkdctp",
   "tg_tokens": null,
   "tps": 85.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-eubkdctp-ctx4096",
   "model": "Qwen3.5-2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 2.1 GB at this context; duplicate traces of the same session: 1dsvww5m; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1130.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-2B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 1,130 81.8 2.1 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eubkdctp",
   "source_url": "https://omlx.ai/benchmarks/performance/eubkdctp",
   "tg_tokens": null,
   "tps": 81.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-mxibzj8t-ctx1024",
   "model": "Qwen3.5-2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 1.9 GB at this context; duplicate traces of the same session: r4hevnhq; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1105.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-2B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 1,105 84.8 1.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mxibzj8t",
   "source_url": "https://omlx.ai/benchmarks/performance/mxibzj8t",
   "tg_tokens": null,
   "tps": 84.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-mxibzj8t-ctx4096",
   "model": "Qwen3.5-2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 2.1 GB at this context; duplicate traces of the same session: r4hevnhq; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1129.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-2B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 1,129 81.7 2.1 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mxibzj8t",
   "source_url": "https://omlx.ai/benchmarks/performance/mxibzj8t",
   "tg_tokens": null,
   "tps": 81.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-yr7eal4y-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.7 GB at this context; duplicate traces of the same session: faduh3tu; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 428.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 428.0 38.3 3.7 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yr7eal4y",
   "source_url": "https://omlx.ai/benchmarks/performance/yr7eal4y",
   "tg_tokens": null,
   "tps": 38.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-yr7eal4y-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.7 GB at this context; duplicate traces of the same session: faduh3tu; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 567.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 567.7 36.8 3.7 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yr7eal4y",
   "source_url": "https://omlx.ai/benchmarks/performance/yr7eal4y",
   "tg_tokens": null,
   "tps": 36.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-3lhwd4ye-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.7 GB at this context; duplicate traces of the same session: s1f1o8cz; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 427.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 427.8 37.7 3.7 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3lhwd4ye",
   "source_url": "https://omlx.ai/benchmarks/performance/3lhwd4ye",
   "tg_tokens": null,
   "tps": 37.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-3lhwd4ye-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.7 GB at this context; duplicate traces of the same session: s1f1o8cz; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 568.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 568.7 36.6 3.7 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3lhwd4ye",
   "source_url": "https://omlx.ai/benchmarks/performance/3lhwd4ye",
   "tg_tokens": null,
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-p5nej7ra-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "DWQ build; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: 4f22b2pg, gkojj1ce, gnj2hpgf, m9b2pdgi, pf3v4z6i; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 746.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit-DWQ on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 8bit; 1k 746.7 67.9 20.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p5nej7ra",
   "source_url": "https://omlx.ai/benchmarks/performance/p5nej7ra",
   "tg_tokens": null,
   "tps": 67.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-p5nej7ra-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "DWQ build; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 4f22b2pg, gkojj1ce, gnj2hpgf, m9b2pdgi, pf3v4z6i; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 754.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit-DWQ on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 8bit; 4k 754.6 64.6 21.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p5nej7ra",
   "source_url": "https://omlx.ai/benchmarks/performance/p5nej7ra",
   "tg_tokens": null,
   "tps": 64.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-p5nej7ra-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "DWQ build; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 4f22b2pg, gkojj1ce, gnj2hpgf, m9b2pdgi, pf3v4z6i; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 693.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit-DWQ on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 8bit; 8k 693.4 61.0 21.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p5nej7ra",
   "source_url": "https://omlx.ai/benchmarks/performance/p5nej7ra",
   "tg_tokens": null,
   "tps": 61.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-p5nej7ra-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "DWQ build; oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: 4f22b2pg, gkojj1ce, gnj2hpgf, m9b2pdgi, pf3v4z6i; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 593.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit-DWQ on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 8bit; 16k 593.3 57.8 21.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p5nej7ra",
   "source_url": "https://omlx.ai/benchmarks/performance/p5nej7ra",
   "tg_tokens": null,
   "tps": 57.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-p5nej7ra-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "DWQ build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 4f22b2pg, gkojj1ce, gnj2hpgf, m9b2pdgi, pf3v4z6i; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 459.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit-DWQ on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 8bit; 32k 459.8 52.0 22.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p5nej7ra",
   "source_url": "https://omlx.ai/benchmarks/performance/p5nej7ra",
   "tg_tokens": null,
   "tps": 52.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-p5nej7ra-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "DWQ build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 4f22b2pg, gkojj1ce, gnj2hpgf, m9b2pdgi, pf3v4z6i; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 316.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit-DWQ on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 8bit; 64k 316.7 42.6 23.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p5nej7ra",
   "source_url": "https://omlx.ai/benchmarks/performance/p5nej7ra",
   "tg_tokens": null,
   "tps": 42.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-091v1gp1-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 28.5 GB; TTFT 8381 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 122.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 8bit; 1k tokens 122.2 PP tok/s 9.7 TG tok/s 8381 TTFT (ms) 28.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/091v1gp1",
   "source_url": "https://omlx.ai/benchmarks/performance/091v1gp1",
   "tg_tokens": null,
   "tps": 9.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-puild5iw-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: ev4hi570; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 914.2,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 914.2 27.5 20.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/puild5iw",
   "source_url": "https://omlx.ai/benchmarks/performance/puild5iw",
   "tg_tokens": null,
   "tps": 27.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-puild5iw-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: ev4hi570; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 984.1,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 984.1 27.2 21.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/puild5iw",
   "source_url": "https://omlx.ai/benchmarks/performance/puild5iw",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-05l1rnwn-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.1,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 323.1 60.0 23.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/05l1rnwn",
   "source_url": "https://omlx.ai/benchmarks/performance/05l1rnwn",
   "tg_tokens": null,
   "tps": 60.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-05l1rnwn-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 303.8,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 16k 303.8 67.1 25.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/05l1rnwn",
   "source_url": "https://omlx.ai/benchmarks/performance/05l1rnwn",
   "tg_tokens": null,
   "tps": 67.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-q3628m3i-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: f9k81wxn; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.3,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 323.3 60.0 23.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q3628m3i",
   "source_url": "https://omlx.ai/benchmarks/performance/q3628m3i",
   "tg_tokens": null,
   "tps": 60.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-q3628m3i-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: f9k81wxn; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 303.2,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 16k 303.2 65.3 25.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q3628m3i",
   "source_url": "https://omlx.ai/benchmarks/performance/q3628m3i",
   "tg_tokens": null,
   "tps": 65.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-idw30d20-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 21.3 GB; TTFT 7907 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 129.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 1k tokens 129.5 PP tok/s 16.3 TG tok/s 7907 TTFT (ms) 21.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/idw30d20",
   "source_url": "https://omlx.ai/benchmarks/performance/idw30d20",
   "tg_tokens": null,
   "tps": 16.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-b2ea4ggk-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: kavsm0op; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 826.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 826.5 31.8 16.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b2ea4ggk",
   "source_url": "https://omlx.ai/benchmarks/performance/b2ea4ggk",
   "tg_tokens": null,
   "tps": 31.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-b2ea4ggk-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: kavsm0op; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1011.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,011 31.2 17.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b2ea4ggk",
   "source_url": "https://omlx.ai/benchmarks/performance/b2ea4ggk",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-7aajy209-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.5 GB at this context; duplicate traces of the same session: 52a7nv82, gh23wvkd, kdekpofe, mf5zylhx, ttodxa8e, v22fonkr, wrcuafvv; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1075.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 1k 1,075 54.8 34.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7aajy209",
   "source_url": "https://omlx.ai/benchmarks/performance/7aajy209",
   "tg_tokens": null,
   "tps": 54.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-7aajy209-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.6 GB at this context; duplicate traces of the same session: 52a7nv82, gh23wvkd, kdekpofe, mf5zylhx, ttodxa8e, v22fonkr, wrcuafvv; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1301.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 4k 1,301 67.5 35.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7aajy209",
   "source_url": "https://omlx.ai/benchmarks/performance/7aajy209",
   "tg_tokens": null,
   "tps": 67.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-7aajy209-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: 52a7nv82, gh23wvkd, kdekpofe, mf5zylhx, ttodxa8e, v22fonkr, wrcuafvv; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1387.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 8k 1,387 86.4 35.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7aajy209",
   "source_url": "https://omlx.ai/benchmarks/performance/7aajy209",
   "tg_tokens": null,
   "tps": 86.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-7aajy209-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: 52a7nv82, gh23wvkd, kdekpofe, mf5zylhx, ttodxa8e, v22fonkr, wrcuafvv; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1356.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 16k 1,356 77.5 35.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7aajy209",
   "source_url": "https://omlx.ai/benchmarks/performance/7aajy209",
   "tg_tokens": null,
   "tps": 77.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-7aajy209-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: 52a7nv82, gh23wvkd, kdekpofe, mf5zylhx, ttodxa8e, v22fonkr, wrcuafvv; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1270.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 32k 1,270 124.8 36.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7aajy209",
   "source_url": "https://omlx.ai/benchmarks/performance/7aajy209",
   "tg_tokens": null,
   "tps": 124.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-7aajy209-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: 52a7nv82, gh23wvkd, kdekpofe, mf5zylhx, ttodxa8e, v22fonkr, wrcuafvv; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1112.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 64k 1,112 73.3 39.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7aajy209",
   "source_url": "https://omlx.ai/benchmarks/performance/7aajy209",
   "tg_tokens": null,
   "tps": 73.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-7aajy209-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 44.0 GB at this context; duplicate traces of the same session: 52a7nv82, gh23wvkd, kdekpofe, mf5zylhx, ttodxa8e, v22fonkr, wrcuafvv; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 886.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 128k 886.2 71.5 44.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7aajy209",
   "source_url": "https://omlx.ai/benchmarks/performance/7aajy209",
   "tg_tokens": null,
   "tps": 71.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-7aajy209-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 49.2 GB at this context; duplicate traces of the same session: 52a7nv82, gh23wvkd, kdekpofe, mf5zylhx, ttodxa8e, v22fonkr, wrcuafvv; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 730.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 195k 730.9 54.4 49.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7aajy209",
   "source_url": "https://omlx.ai/benchmarks/performance/7aajy209",
   "tg_tokens": null,
   "tps": 54.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ksdcu0oc-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 45.7 GB; TTFT 32551 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2013.0,
   "provenance": "community",
   "quant": "GSQ-RCO-Q2",
   "quote": "Qwen3.8-Flash-Next-GSQ-RCO-Q2-mlx on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k tokens 2,013 PP tok/s 54.0 TG tok/s 32551 TTFT (ms) 45.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ksdcu0oc",
   "source_url": "https://omlx.ai/benchmarks/performance/ksdcu0oc",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wlqfpazq-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 49.9 GB; TTFT 71487 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1834.0,
   "provenance": "community",
   "quant": "GSQ-RCO-Q2",
   "quote": "Qwen3.8-Flash-Next-GSQ-RCO-Q2-mlx on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k tokens 1,834 PP tok/s 14.5 TG tok/s 71487 TTFT (ms) 49.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wlqfpazq",
   "source_url": "https://omlx.ai/benchmarks/performance/wlqfpazq",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-a34jdcsm-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: 1qnbriis, b88darsf, bm0qatwk, dx1g6pxy, ja0534gc, ue66f8nv; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 558.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 558.6 30.8 100.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a34jdcsm",
   "source_url": "https://omlx.ai/benchmarks/performance/a34jdcsm",
   "tg_tokens": null,
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-a34jdcsm-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.1 GB at this context; duplicate traces of the same session: 1qnbriis, b88darsf, bm0qatwk, dx1g6pxy, ja0534gc, ue66f8nv; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2934.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 2,934 86.0 102.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a34jdcsm",
   "source_url": "https://omlx.ai/benchmarks/performance/a34jdcsm",
   "tg_tokens": null,
   "tps": 86.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-a34jdcsm-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.7 GB at this context; duplicate traces of the same session: 1qnbriis, b88darsf, bm0qatwk, dx1g6pxy, ja0534gc, ue66f8nv; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4480.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 4,480 99.8 102.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a34jdcsm",
   "source_url": "https://omlx.ai/benchmarks/performance/a34jdcsm",
   "tg_tokens": null,
   "tps": 99.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-a34jdcsm-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.3 GB at this context; duplicate traces of the same session: 1qnbriis, b88darsf, bm0qatwk, dx1g6pxy, ja0534gc, ue66f8nv; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4678.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 4,678 90.3 103.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a34jdcsm",
   "source_url": "https://omlx.ai/benchmarks/performance/a34jdcsm",
   "tg_tokens": null,
   "tps": 90.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-a34jdcsm-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.5 GB at this context; duplicate traces of the same session: 1qnbriis, b88darsf, bm0qatwk, dx1g6pxy, ja0534gc, ue66f8nv; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4633.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 4,633 114.2 104.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a34jdcsm",
   "source_url": "https://omlx.ai/benchmarks/performance/a34jdcsm",
   "tg_tokens": null,
   "tps": 114.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-a34jdcsm-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 106.8 GB at this context; duplicate traces of the same session: 1qnbriis, b88darsf, bm0qatwk, dx1g6pxy, ja0534gc, ue66f8nv; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4526.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 4,526 99.1 106.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a34jdcsm",
   "source_url": "https://omlx.ai/benchmarks/performance/a34jdcsm",
   "tg_tokens": null,
   "tps": 99.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-a34jdcsm-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.5 GB at this context; duplicate traces of the same session: 1qnbriis, b88darsf, bm0qatwk, dx1g6pxy, ja0534gc, ue66f8nv; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4381.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 4,381 82.0 109.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a34jdcsm",
   "source_url": "https://omlx.ai/benchmarks/performance/a34jdcsm",
   "tg_tokens": null,
   "tps": 82.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-x3h2hpl2-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 0hnelctr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 259.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 259.8 57.2 22.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x3h2hpl2",
   "source_url": "https://omlx.ai/benchmarks/performance/x3h2hpl2",
   "tg_tokens": null,
   "tps": 57.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-x3h2hpl2-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 0hnelctr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 224.8 71.4 23.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x3h2hpl2",
   "source_url": "https://omlx.ai/benchmarks/performance/x3h2hpl2",
   "tg_tokens": null,
   "tps": 71.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-x3h2hpl2-ctx65536",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: 0hnelctr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 187.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 187.2 41.1 27.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x3h2hpl2",
   "source_url": "https://omlx.ai/benchmarks/performance/x3h2hpl2",
   "tg_tokens": null,
   "tps": 41.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-x3h2hpl2-ctx199680",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.7 GB at this context; duplicate traces of the same session: 0hnelctr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 129.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 129.0 26.3 36.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x3h2hpl2",
   "source_url": "https://omlx.ai/benchmarks/performance/x3h2hpl2",
   "tg_tokens": null,
   "tps": 26.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-vxamu9hg-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 4xw0ihae, fip4d8mb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4287.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 4,287 208.8 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vxamu9hg",
   "source_url": "https://omlx.ai/benchmarks/performance/vxamu9hg",
   "tg_tokens": null,
   "tps": 208.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-vxamu9hg-ctx8192",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 4xw0ihae, fip4d8mb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5896.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 5,896 303.1 22.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vxamu9hg",
   "source_url": "https://omlx.ai/benchmarks/performance/vxamu9hg",
   "tg_tokens": null,
   "tps": 303.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-vxamu9hg-ctx32768",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 4xw0ihae, fip4d8mb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4373.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 4,373 202.7 22.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vxamu9hg",
   "source_url": "https://omlx.ai/benchmarks/performance/vxamu9hg",
   "tg_tokens": null,
   "tps": 202.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-vxamu9hg-ctx131072",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: 4xw0ihae, fip4d8mb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1958.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 1,958 166.1 24.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vxamu9hg",
   "source_url": "https://omlx.ai/benchmarks/performance/vxamu9hg",
   "tg_tokens": null,
   "tps": 166.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fqih5u80-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: fppmnp3q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4220.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 4,220 194.2 21.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fqih5u80",
   "source_url": "https://omlx.ai/benchmarks/performance/fqih5u80",
   "tg_tokens": null,
   "tps": 194.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fqih5u80-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: fppmnp3q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5873.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 5,873 167.2 22.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fqih5u80",
   "source_url": "https://omlx.ai/benchmarks/performance/fqih5u80",
   "tg_tokens": null,
   "tps": 167.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-jy9yn9ju-ctx1024",
   "model": "gemma-4-12b-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 13.8 GB at this context; duplicate traces of the same session: ksvjh4p1, x8hqeoxi; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 214.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-12B-it-oQ8e-mtp on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 8bit; 1k 214.0 12.9 13.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jy9yn9ju",
   "source_url": "https://omlx.ai/benchmarks/performance/jy9yn9ju",
   "tg_tokens": null,
   "tps": 12.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-jy9yn9ju-ctx4096",
   "model": "gemma-4-12b-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 14.1 GB at this context; duplicate traces of the same session: ksvjh4p1, x8hqeoxi; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 212.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-12B-it-oQ8e-mtp on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 8bit; 4k 212.9 14.3 14.1 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jy9yn9ju",
   "source_url": "https://omlx.ai/benchmarks/performance/jy9yn9ju",
   "tg_tokens": null,
   "tps": 14.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-jy9yn9ju-ctx16384",
   "model": "gemma-4-12b-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 14.7 GB at this context; duplicate traces of the same session: ksvjh4p1, x8hqeoxi; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 196.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-12B-it-oQ8e-mtp on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 8bit; 16k 196.6 9.8 14.7 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jy9yn9ju",
   "source_url": "https://omlx.ai/benchmarks/performance/jy9yn9ju",
   "tg_tokens": null,
   "tps": 9.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-3n80u65i-ctx1024",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.6 GB at this context; duplicate traces of the same session: dum8ri0k, eviep4jw, lya53zvy, qx16vnj3, u1wheitc; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 911.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-31B-it-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 1k 911.1 44.5 34.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3n80u65i",
   "source_url": "https://omlx.ai/benchmarks/performance/3n80u65i",
   "tg_tokens": null,
   "tps": 44.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-3n80u65i-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.0 GB at this context; duplicate traces of the same session: dum8ri0k, eviep4jw, lya53zvy, qx16vnj3, u1wheitc; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 975.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-31B-it-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 4k 975.7 52.0 35.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3n80u65i",
   "source_url": "https://omlx.ai/benchmarks/performance/3n80u65i",
   "tg_tokens": null,
   "tps": 52.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-3n80u65i-ctx8192",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: dum8ri0k, eviep4jw, lya53zvy, qx16vnj3, u1wheitc; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 960.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-31B-it-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 8k 960.9 50.6 35.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3n80u65i",
   "source_url": "https://omlx.ai/benchmarks/performance/3n80u65i",
   "tg_tokens": null,
   "tps": 50.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-3n80u65i-ctx16384",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.7 GB at this context; duplicate traces of the same session: dum8ri0k, eviep4jw, lya53zvy, qx16vnj3, u1wheitc; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 923.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-31B-it-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 16k 923.8 59.6 36.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3n80u65i",
   "source_url": "https://omlx.ai/benchmarks/performance/3n80u65i",
   "tg_tokens": null,
   "tps": 59.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-3n80u65i-ctx32768",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.0 GB at this context; duplicate traces of the same session: dum8ri0k, eviep4jw, lya53zvy, qx16vnj3, u1wheitc; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 845.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-31B-it-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 32k 845.0 45.5 39.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3n80u65i",
   "source_url": "https://omlx.ai/benchmarks/performance/3n80u65i",
   "tg_tokens": null,
   "tps": 45.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-3n80u65i-ctx65536",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.7 GB at this context; duplicate traces of the same session: dum8ri0k, eviep4jw, lya53zvy, qx16vnj3, u1wheitc; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 723.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-31B-it-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 64k 723.8 34.9 43.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3n80u65i",
   "source_url": "https://omlx.ai/benchmarks/performance/3n80u65i",
   "tg_tokens": null,
   "tps": 34.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-k8hnsq2l-ctx1024",
   "model": "Occamy-1.0",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 568.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "occamy-1.0-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 568.0 150.8 20.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k8hnsq2l",
   "source_url": "https://omlx.ai/benchmarks/performance/k8hnsq2l",
   "tg_tokens": null,
   "tps": 150.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-k8hnsq2l-ctx8192",
   "model": "Occamy-1.0",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3302.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "occamy-1.0-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 3,302 103.7 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k8hnsq2l",
   "source_url": "https://omlx.ai/benchmarks/performance/k8hnsq2l",
   "tg_tokens": null,
   "tps": 103.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-k8hnsq2l-ctx32768",
   "model": "Occamy-1.0",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4083.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "occamy-1.0-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 4,083 149.3 21.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k8hnsq2l",
   "source_url": "https://omlx.ai/benchmarks/performance/k8hnsq2l",
   "tg_tokens": null,
   "tps": 149.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-k8hnsq2l-ctx131072",
   "model": "Occamy-1.0",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1882.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "occamy-1.0-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 1,882 69.8 24.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k8hnsq2l",
   "source_url": "https://omlx.ai/benchmarks/performance/k8hnsq2l",
   "tg_tokens": null,
   "tps": 69.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-y8tvyb63-ctx32768",
   "model": "Occamy-1.0",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 22.0 GB; TTFT 7493 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4373.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "occamy-1.0-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k tokens 4,373 PP tok/s 132.9 TG tok/s 7493 TTFT (ms) 22.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y8tvyb63",
   "source_url": "https://omlx.ai/benchmarks/performance/y8tvyb63",
   "tg_tokens": null,
   "tps": 132.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-lr3u9x2r-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 249.1 GB at this context; duplicate traces of the same session: 23upctoc, 6ksf0y33, ad4bphml, amu0houb, h6kj8fza; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 491.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 1k 491.6 41.6 249.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lr3u9x2r",
   "source_url": "https://omlx.ai/benchmarks/performance/lr3u9x2r",
   "tg_tokens": null,
   "tps": 41.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-lr3u9x2r-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 250.7 GB at this context; duplicate traces of the same session: 23upctoc, 6ksf0y33, ad4bphml, amu0houb, h6kj8fza; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 518.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 16k 518.8 69.4 250.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lr3u9x2r",
   "source_url": "https://omlx.ai/benchmarks/performance/lr3u9x2r",
   "tg_tokens": null,
   "tps": 69.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-lr3u9x2r-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 250.8 GB at this context; duplicate traces of the same session: 23upctoc, 6ksf0y33, ad4bphml, amu0houb, h6kj8fza; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 513.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 32k 513.6 53.4 250.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lr3u9x2r",
   "source_url": "https://omlx.ai/benchmarks/performance/lr3u9x2r",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-lr3u9x2r-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 251.3 GB at this context; duplicate traces of the same session: 23upctoc, 6ksf0y33, ad4bphml, amu0houb, h6kj8fza; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 506.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 64k 506.0 47.9 251.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lr3u9x2r",
   "source_url": "https://omlx.ai/benchmarks/performance/lr3u9x2r",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lr3u9x2r-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 252.2 GB at this context; duplicate traces of the same session: 23upctoc, 6ksf0y33, ad4bphml, amu0houb, h6kj8fza; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 496.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 128k 496.7 66.7 252.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lr3u9x2r",
   "source_url": "https://omlx.ai/benchmarks/performance/lr3u9x2r",
   "tg_tokens": null,
   "tps": 66.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lr3u9x2r-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 253.3 GB at this context; duplicate traces of the same session: 23upctoc, 6ksf0y33, ad4bphml, amu0houb, h6kj8fza; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 488.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 195k 488.4 37.3 253.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lr3u9x2r",
   "source_url": "https://omlx.ai/benchmarks/performance/lr3u9x2r",
   "tg_tokens": null,
   "tps": 37.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-cyadws5c-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 323.4 GB at this context; duplicate traces of the same session: sxrcf5y3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 496.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 496.5 38.0 323.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cyadws5c",
   "source_url": "https://omlx.ai/benchmarks/performance/cyadws5c",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-cyadws5c-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 324.7 GB at this context; duplicate traces of the same session: sxrcf5y3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 582.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 582.5 38.7 324.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cyadws5c",
   "source_url": "https://omlx.ai/benchmarks/performance/cyadws5c",
   "tg_tokens": null,
   "tps": 38.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-cyadws5c-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 325.0 GB at this context; duplicate traces of the same session: sxrcf5y3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 543.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 543.0 46.1 325.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cyadws5c",
   "source_url": "https://omlx.ai/benchmarks/performance/cyadws5c",
   "tg_tokens": null,
   "tps": 46.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-cyadws5c-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 325.2 GB at this context; duplicate traces of the same session: sxrcf5y3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 532.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k 532.5 55.9 325.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cyadws5c",
   "source_url": "https://omlx.ai/benchmarks/performance/cyadws5c",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-cyadws5c-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 325.3 GB at this context; duplicate traces of the same session: sxrcf5y3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 525.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 32k 525.1 47.8 325.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cyadws5c",
   "source_url": "https://omlx.ai/benchmarks/performance/cyadws5c",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-cyadws5c-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 326.0 GB at this context; duplicate traces of the same session: sxrcf5y3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 518.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 64k 518.1 38.9 326.0 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cyadws5c",
   "source_url": "https://omlx.ai/benchmarks/performance/cyadws5c",
   "tg_tokens": null,
   "tps": 38.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-tms49y79-ctx4096",
   "model": "LOCAL-SMART",
   "notes": "oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: j2rs5hi6; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 126.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LOCAL-SMART on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 126.9 27.1 22.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tms49y79",
   "source_url": "https://omlx.ai/benchmarks/performance/tms49y79",
   "tg_tokens": null,
   "tps": 27.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-tms49y79-ctx16384",
   "model": "LOCAL-SMART",
   "notes": "oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: j2rs5hi6; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 117.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LOCAL-SMART on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 117.1 34.0 24.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tms49y79",
   "source_url": "https://omlx.ai/benchmarks/performance/tms49y79",
   "tg_tokens": null,
   "tps": 34.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-uf59m3df-ctx1024",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: 09n0mw1u, msc33o4o, pvynn9f9, uzsbrgh9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 824.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 824.1 32.6 15.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uf59m3df",
   "source_url": "https://omlx.ai/benchmarks/performance/uf59m3df",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-uf59m3df-ctx4096",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: 09n0mw1u, msc33o4o, pvynn9f9, uzsbrgh9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 800.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 800.3 32.3 16.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uf59m3df",
   "source_url": "https://omlx.ai/benchmarks/performance/uf59m3df",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-uf59m3df-ctx8192",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 16.2 GB at this context; duplicate traces of the same session: 09n0mw1u, msc33o4o, pvynn9f9, uzsbrgh9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 717.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 717.7 31.9 16.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uf59m3df",
   "source_url": "https://omlx.ai/benchmarks/performance/uf59m3df",
   "tg_tokens": null,
   "tps": 31.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-uf59m3df-ctx16384",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: 09n0mw1u, msc33o4o, pvynn9f9, uzsbrgh9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 693.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 693.3 30.7 16.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uf59m3df",
   "source_url": "https://omlx.ai/benchmarks/performance/uf59m3df",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-uf59m3df-ctx32768",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 16.5 GB at this context; duplicate traces of the same session: 09n0mw1u, msc33o4o, pvynn9f9, uzsbrgh9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 661.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 661.9 29.1 16.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uf59m3df",
   "source_url": "https://omlx.ai/benchmarks/performance/uf59m3df",
   "tg_tokens": null,
   "tps": 29.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-62jal4gp-ctx1024",
   "model": "Qwen3 14B",
   "notes": "oMLX community benchmark trace; peak memory 8.4 GB at this context; duplicate traces of the same session: fx4f0c1e; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 33.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-14B-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 33.4 4.2 8.4 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/62jal4gp",
   "source_url": "https://omlx.ai/benchmarks/performance/62jal4gp",
   "tg_tokens": null,
   "tps": 4.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-62jal4gp-ctx4096",
   "model": "Qwen3 14B",
   "notes": "oMLX community benchmark trace; peak memory 8.9 GB at this context; duplicate traces of the same session: fx4f0c1e; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 41.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-14B-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 41.6 4.6 8.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/62jal4gp",
   "source_url": "https://omlx.ai/benchmarks/performance/62jal4gp",
   "tg_tokens": null,
   "tps": 4.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-mko3tacc-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 5.9 GB; TTFT 3648 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 280.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 1k tokens 280.7 PP tok/s 47.6 TG tok/s 3648 TTFT (ms) 5.9 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mko3tacc",
   "source_url": "https://omlx.ai/benchmarks/performance/mko3tacc",
   "tg_tokens": null,
   "tps": 47.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-jo8448hj-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.6 GB at this context; duplicate traces of the same session: 618p2pug; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 233.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 233.5 21.3 6.6 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jo8448hj",
   "source_url": "https://omlx.ai/benchmarks/performance/jo8448hj",
   "tg_tokens": null,
   "tps": 21.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-jo8448hj-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: 618p2pug; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 274.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 274.6 20.7 7.0 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jo8448hj",
   "source_url": "https://omlx.ai/benchmarks/performance/jo8448hj",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5fs2ye70-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: 1dcph2hx, 2b889d5c, 49bit2kv, 7gkv9d8f, d93a1x81, q0tq0vzz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3822.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 3,822 167.0 26.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5fs2ye70",
   "source_url": "https://omlx.ai/benchmarks/performance/5fs2ye70",
   "tg_tokens": null,
   "tps": 167.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5fs2ye70-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: 1dcph2hx, 2b889d5c, 49bit2kv, 7gkv9d8f, d93a1x81, q0tq0vzz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5303.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 5,303 158.7 26.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5fs2ye70",
   "source_url": "https://omlx.ai/benchmarks/performance/5fs2ye70",
   "tg_tokens": null,
   "tps": 158.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5fs2ye70-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: 1dcph2hx, 2b889d5c, 49bit2kv, 7gkv9d8f, d93a1x81, q0tq0vzz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5290.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 5,290 173.8 26.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5fs2ye70",
   "source_url": "https://omlx.ai/benchmarks/performance/5fs2ye70",
   "tg_tokens": null,
   "tps": 173.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5fs2ye70-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.9 GB at this context; duplicate traces of the same session: 1dcph2hx, 2b889d5c, 49bit2kv, 7gkv9d8f, d93a1x81, q0tq0vzz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5003.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 5,003 215.4 26.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5fs2ye70",
   "source_url": "https://omlx.ai/benchmarks/performance/5fs2ye70",
   "tg_tokens": null,
   "tps": 215.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5fs2ye70-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.3 GB at this context; duplicate traces of the same session: 1dcph2hx, 2b889d5c, 49bit2kv, 7gkv9d8f, d93a1x81, q0tq0vzz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4185.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 4,185 253.5 27.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5fs2ye70",
   "source_url": "https://omlx.ai/benchmarks/performance/5fs2ye70",
   "tg_tokens": null,
   "tps": 253.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5fs2ye70-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: 1dcph2hx, 2b889d5c, 49bit2kv, 7gkv9d8f, d93a1x81, q0tq0vzz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3134.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 3,134 139.4 28.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5fs2ye70",
   "source_url": "https://omlx.ai/benchmarks/performance/5fs2ye70",
   "tg_tokens": null,
   "tps": 139.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5fs2ye70-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: 1dcph2hx, 2b889d5c, 49bit2kv, 7gkv9d8f, d93a1x81, q0tq0vzz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2097.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 2,097 100.2 29.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5fs2ye70",
   "source_url": "https://omlx.ai/benchmarks/performance/5fs2ye70",
   "tg_tokens": null,
   "tps": 100.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5zp2tmdm-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: 1mfn85vz, fx60a3i8, q9s2pyny, t1jv2eg8, u0jnxj2i, w053ejq6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3658.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 3,658 198.6 26.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5zp2tmdm",
   "source_url": "https://omlx.ai/benchmarks/performance/5zp2tmdm",
   "tg_tokens": null,
   "tps": 198.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5zp2tmdm-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: 1mfn85vz, fx60a3i8, q9s2pyny, t1jv2eg8, u0jnxj2i, w053ejq6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5249.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 5,249 167.3 26.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5zp2tmdm",
   "source_url": "https://omlx.ai/benchmarks/performance/5zp2tmdm",
   "tg_tokens": null,
   "tps": 167.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5zp2tmdm-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: 1mfn85vz, fx60a3i8, q9s2pyny, t1jv2eg8, u0jnxj2i, w053ejq6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5251.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 5,251 117.2 26.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5zp2tmdm",
   "source_url": "https://omlx.ai/benchmarks/performance/5zp2tmdm",
   "tg_tokens": null,
   "tps": 117.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5zp2tmdm-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.0 GB at this context; duplicate traces of the same session: 1mfn85vz, fx60a3i8, q9s2pyny, t1jv2eg8, u0jnxj2i, w053ejq6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4837.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 4,837 186.7 27.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5zp2tmdm",
   "source_url": "https://omlx.ai/benchmarks/performance/5zp2tmdm",
   "tg_tokens": null,
   "tps": 186.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5zp2tmdm-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: 1mfn85vz, fx60a3i8, q9s2pyny, t1jv2eg8, u0jnxj2i, w053ejq6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4105.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 4,105 237.2 27.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5zp2tmdm",
   "source_url": "https://omlx.ai/benchmarks/performance/5zp2tmdm",
   "tg_tokens": null,
   "tps": 237.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5zp2tmdm-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: 1mfn85vz, fx60a3i8, q9s2pyny, t1jv2eg8, u0jnxj2i, w053ejq6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3058.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 3,058 139.2 28.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5zp2tmdm",
   "source_url": "https://omlx.ai/benchmarks/performance/5zp2tmdm",
   "tg_tokens": null,
   "tps": 139.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5zp2tmdm-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: 1mfn85vz, fx60a3i8, q9s2pyny, t1jv2eg8, u0jnxj2i, w053ejq6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2020.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 2,020 99.4 29.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5zp2tmdm",
   "source_url": "https://omlx.ai/benchmarks/performance/5zp2tmdm",
   "tg_tokens": null,
   "tps": 99.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-w3nu30ou-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 17.6 GB; TTFT 6450 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 158.8,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 1k tokens 158.8 PP tok/s 19.8 TG tok/s 6450 TTFT (ms) 17.6 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w3nu30ou",
   "source_url": "https://omlx.ai/benchmarks/performance/w3nu30ou",
   "tg_tokens": null,
   "tps": 19.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-9kxk50ga-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 19.1 GB; TTFT 25409 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 161.2,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 4k tokens 161.2 PP tok/s 19.8 TG tok/s 25409 TTFT (ms) 19.1 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9kxk50ga",
   "source_url": "https://omlx.ai/benchmarks/performance/9kxk50ga",
   "tg_tokens": null,
   "tps": 19.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-tgat98o3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 4n6szc9m, c1mwitfp, fanntar5, lwq14cf6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 252.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 252.0 56.0 18.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tgat98o3",
   "source_url": "https://omlx.ai/benchmarks/performance/tgat98o3",
   "tg_tokens": null,
   "tps": 56.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-tgat98o3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 4n6szc9m, c1mwitfp, fanntar5, lwq14cf6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 253.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 253.7 49.9 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tgat98o3",
   "source_url": "https://omlx.ai/benchmarks/performance/tgat98o3",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-tgat98o3-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 4n6szc9m, c1mwitfp, fanntar5, lwq14cf6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 234.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 8k 234.5 64.9 19.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tgat98o3",
   "source_url": "https://omlx.ai/benchmarks/performance/tgat98o3",
   "tg_tokens": null,
   "tps": 64.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-tgat98o3-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 4n6szc9m, c1mwitfp, fanntar5, lwq14cf6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 236.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 236.4 63.7 20.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tgat98o3",
   "source_url": "https://omlx.ai/benchmarks/performance/tgat98o3",
   "tg_tokens": null,
   "tps": 63.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-tgat98o3-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 4n6szc9m, c1mwitfp, fanntar5, lwq14cf6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 225.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 225.7 59.3 21.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tgat98o3",
   "source_url": "https://omlx.ai/benchmarks/performance/tgat98o3",
   "tg_tokens": null,
   "tps": 59.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-a9q3fqnr-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: j4shcnhw, wwfs877v; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 960.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 960.4 63.7 18.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a9q3fqnr",
   "source_url": "https://omlx.ai/benchmarks/performance/a9q3fqnr",
   "tg_tokens": null,
   "tps": 63.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-a9q3fqnr-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: j4shcnhw, wwfs877v; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 860.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 860.5 80.9 19.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a9q3fqnr",
   "source_url": "https://omlx.ai/benchmarks/performance/a9q3fqnr",
   "tg_tokens": null,
   "tps": 80.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-a9q3fqnr-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: j4shcnhw, wwfs877v; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 731.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 731.5 124.4 20.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a9q3fqnr",
   "source_url": "https://omlx.ai/benchmarks/performance/a9q3fqnr",
   "tg_tokens": null,
   "tps": 124.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pp60a3nr-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 50.9 GB; TTFT 76285 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1718.0,
   "provenance": "community",
   "quant": "GSQ-RCO-Q2",
   "quote": "Qwen3.8-Flash-Next-GSQ-RCO-Q2-mlx on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k tokens 1,718 PP tok/s 43.5 TG tok/s 76285 TTFT (ms) 50.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pp60a3nr",
   "source_url": "https://omlx.ai/benchmarks/performance/pp60a3nr",
   "tg_tokens": null,
   "tps": 43.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-2xzwmxw3-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 72.2 GB at this context; duplicate traces of the same session: jnmqtwsp; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 329.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 329.2 29.5 72.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2xzwmxw3",
   "source_url": "https://omlx.ai/benchmarks/performance/2xzwmxw3",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-2xzwmxw3-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 74.5 GB at this context; duplicate traces of the same session: jnmqtwsp; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 424.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 424.6 25.0 74.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2xzwmxw3",
   "source_url": "https://omlx.ai/benchmarks/performance/2xzwmxw3",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-2xzwmxw3-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 74.9 GB at this context; duplicate traces of the same session: jnmqtwsp; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 502.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 502.2 34.8 74.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2xzwmxw3",
   "source_url": "https://omlx.ai/benchmarks/performance/2xzwmxw3",
   "tg_tokens": null,
   "tps": 34.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-2xzwmxw3-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 75.2 GB at this context; duplicate traces of the same session: jnmqtwsp; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 507.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 507.9 31.9 75.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2xzwmxw3",
   "source_url": "https://omlx.ai/benchmarks/performance/2xzwmxw3",
   "tg_tokens": null,
   "tps": 31.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-2xzwmxw3-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 75.7 GB at this context; duplicate traces of the same session: jnmqtwsp; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 489.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 32k 489.4 27.6 75.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2xzwmxw3",
   "source_url": "https://omlx.ai/benchmarks/performance/2xzwmxw3",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-09kvw8n8-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace, single run at 1k context; peak memory 72.2 GB; TTFT 2172 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 471.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 471.4 PP tok/s 36.2 TG tok/s 2172 TTFT (ms) 72.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/09kvw8n8",
   "source_url": "https://omlx.ai/benchmarks/performance/09kvw8n8",
   "tg_tokens": null,
   "tps": 36.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dm8a3ifi-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace, single run at 1k context; peak memory 72.2 GB; TTFT 959 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1067.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,067 PP tok/s 64.6 TG tok/s 959 TTFT (ms) 72.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dm8a3ifi",
   "source_url": "https://omlx.ai/benchmarks/performance/dm8a3ifi",
   "tg_tokens": null,
   "tps": 64.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vh1mgqkc-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace, single run at 1k context; peak memory 76.2 GB; TTFT 1633 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 627.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 627.1 PP tok/s 18.5 TG tok/s 1633 TTFT (ms) 76.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vh1mgqkc",
   "source_url": "https://omlx.ai/benchmarks/performance/vh1mgqkc",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hgsbrrt1-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace, single run at 1k context; peak memory 76.0 GB; TTFT 1475 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 694.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 694.1 PP tok/s 21.4 TG tok/s 1475 TTFT (ms) 76.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hgsbrrt1",
   "source_url": "https://omlx.ai/benchmarks/performance/hgsbrrt1",
   "tg_tokens": null,
   "tps": 21.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7p703q4k-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.5 GB; TTFT 1051 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 974.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 974.6 PP tok/s 67.1 TG tok/s 1051 TTFT (ms) 70.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7p703q4k",
   "source_url": "https://omlx.ai/benchmarks/performance/7p703q4k",
   "tg_tokens": null,
   "tps": 67.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9vevwq9d-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 71.0 GB; TTFT 6030 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 679.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 679.3 PP tok/s 31.1 TG tok/s 6030 TTFT (ms) 71.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9vevwq9d",
   "source_url": "https://omlx.ai/benchmarks/performance/9vevwq9d",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-5hbtira5-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: oz2hpi9u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 259.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 259.8 57.2 22.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5hbtira5",
   "source_url": "https://omlx.ai/benchmarks/performance/5hbtira5",
   "tg_tokens": null,
   "tps": 57.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-5hbtira5-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: oz2hpi9u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 224.8 71.4 23.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5hbtira5",
   "source_url": "https://omlx.ai/benchmarks/performance/5hbtira5",
   "tg_tokens": null,
   "tps": 71.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-5hbtira5-ctx65536",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: oz2hpi9u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 187.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 187.2 41.1 27.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5hbtira5",
   "source_url": "https://omlx.ai/benchmarks/performance/5hbtira5",
   "tg_tokens": null,
   "tps": 41.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-5hbtira5-ctx199680",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.7 GB at this context; duplicate traces of the same session: oz2hpi9u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 129.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 129.0 26.3 36.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5hbtira5",
   "source_url": "https://omlx.ai/benchmarks/performance/5hbtira5",
   "tg_tokens": null,
   "tps": 26.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-xb7ofplp-ctx1024",
   "model": "gemma-4-12b-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 8.6 GB at this context; duplicate traces of the same session: e7crbc1p, xnccqct4; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 210.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-12B-it-oQ4e-mtp on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 1k 210.8 17.2 8.6 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xb7ofplp",
   "source_url": "https://omlx.ai/benchmarks/performance/xb7ofplp",
   "tg_tokens": null,
   "tps": 17.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-xb7ofplp-ctx4096",
   "model": "gemma-4-12b-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 8.8 GB at this context; duplicate traces of the same session: e7crbc1p, xnccqct4; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 207.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-12B-it-oQ4e-mtp on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 4k 207.4 26.6 8.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xb7ofplp",
   "source_url": "https://omlx.ai/benchmarks/performance/xb7ofplp",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-xb7ofplp-ctx16384",
   "model": "gemma-4-12b-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 9.4 GB at this context; duplicate traces of the same session: e7crbc1p, xnccqct4; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 196.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-12B-it-oQ4e-mtp on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 16k 196.5 11.6 9.4 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xb7ofplp",
   "source_url": "https://omlx.ai/benchmarks/performance/xb7ofplp",
   "tg_tokens": null,
   "tps": 11.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-opz8wttu-ctx1024",
   "model": "Occamy-1.0",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: 5w1pwc51; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4427.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "occamy-1.0-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 4,427 168.9 20.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/opz8wttu",
   "source_url": "https://omlx.ai/benchmarks/performance/opz8wttu",
   "tg_tokens": null,
   "tps": 168.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-opz8wttu-ctx4096",
   "model": "Occamy-1.0",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 5w1pwc51; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6002.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "occamy-1.0-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 6,002 150.6 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/opz8wttu",
   "source_url": "https://omlx.ai/benchmarks/performance/opz8wttu",
   "tg_tokens": null,
   "tps": 150.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-6po51h3d-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 323.4 GB at this context; duplicate traces of the same session: 4ahns392, 57fqkon5, vcalhiml; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 496.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 496.5 38.0 323.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6po51h3d",
   "source_url": "https://omlx.ai/benchmarks/performance/6po51h3d",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-6po51h3d-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 324.7 GB at this context; duplicate traces of the same session: 4ahns392, 57fqkon5, vcalhiml; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 582.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 582.5 38.7 324.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6po51h3d",
   "source_url": "https://omlx.ai/benchmarks/performance/6po51h3d",
   "tg_tokens": null,
   "tps": 38.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-6po51h3d-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 325.0 GB at this context; duplicate traces of the same session: 4ahns392, 57fqkon5, vcalhiml; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 543.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 543.0 46.1 325.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6po51h3d",
   "source_url": "https://omlx.ai/benchmarks/performance/6po51h3d",
   "tg_tokens": null,
   "tps": 46.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-6po51h3d-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 325.2 GB at this context; duplicate traces of the same session: 4ahns392, 57fqkon5, vcalhiml; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 532.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k 532.5 55.9 325.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6po51h3d",
   "source_url": "https://omlx.ai/benchmarks/performance/6po51h3d",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-6po51h3d-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 325.3 GB at this context; duplicate traces of the same session: 4ahns392, 57fqkon5, vcalhiml; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 525.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 32k 525.1 47.8 325.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6po51h3d",
   "source_url": "https://omlx.ai/benchmarks/performance/6po51h3d",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-6po51h3d-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 326.0 GB at this context; duplicate traces of the same session: 4ahns392, 57fqkon5, vcalhiml; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 518.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 64k 518.1 38.9 326.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6po51h3d",
   "source_url": "https://omlx.ai/benchmarks/performance/6po51h3d",
   "tg_tokens": null,
   "tps": 38.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-i9zlx5m2-ctx4096",
   "model": "LOCAL-SMART",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: htwroclr; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 129.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LOCAL-SMART on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 129.8 19.1 20.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i9zlx5m2",
   "source_url": "https://omlx.ai/benchmarks/performance/i9zlx5m2",
   "tg_tokens": null,
   "tps": 19.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-i9zlx5m2-ctx16384",
   "model": "LOCAL-SMART",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: htwroclr; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 121.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LOCAL-SMART on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 121.2 17.3 22.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i9zlx5m2",
   "source_url": "https://omlx.ai/benchmarks/performance/i9zlx5m2",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-1z8suhfz-ctx4096",
   "model": "LOCAL-SMART",
   "notes": "oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: svszbrz2; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 126.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LOCAL-SMART on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 126.6 31.6 22.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1z8suhfz",
   "source_url": "https://omlx.ai/benchmarks/performance/1z8suhfz",
   "tg_tokens": null,
   "tps": 31.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-1z8suhfz-ctx16384",
   "model": "LOCAL-SMART",
   "notes": "oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: svszbrz2; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 118.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LOCAL-SMART on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 118.0 33.5 24.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1z8suhfz",
   "source_url": "https://omlx.ai/benchmarks/performance/1z8suhfz",
   "tg_tokens": null,
   "tps": 33.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-37fu468y-ctx1024",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: epmmzxq9; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 725.7,
   "provenance": "community",
   "quant": "5bit",
   "quote": "MiniCPM-V-4.7-35B-A3B-5bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 5bit; 1k 725.7 48.9 26.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/37fu468y",
   "source_url": "https://omlx.ai/benchmarks/performance/37fu468y",
   "tg_tokens": null,
   "tps": 48.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-37fu468y-ctx4096",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: epmmzxq9; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 837.5,
   "provenance": "community",
   "quant": "5bit",
   "quote": "MiniCPM-V-4.7-35B-A3B-5bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 5bit; 4k 837.5 48.0 27.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/37fu468y",
   "source_url": "https://omlx.ai/benchmarks/performance/37fu468y",
   "tg_tokens": null,
   "tps": 48.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-oo4onq4s-ctx1024",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 17.9 GB at this context; duplicate traces of the same session: soonujxn, yc23abgv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 198.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 198.9 53.9 17.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oo4onq4s",
   "source_url": "https://omlx.ai/benchmarks/performance/oo4onq4s",
   "tg_tokens": null,
   "tps": 53.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-oo4onq4s-ctx4096",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 18.2 GB at this context; duplicate traces of the same session: soonujxn, yc23abgv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 233.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 233.9 78.9 18.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oo4onq4s",
   "source_url": "https://omlx.ai/benchmarks/performance/oo4onq4s",
   "tg_tokens": null,
   "tps": 78.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-oo4onq4s-ctx8192",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: soonujxn, yc23abgv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 236.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 236.7 37.9 18.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oo4onq4s",
   "source_url": "https://omlx.ai/benchmarks/performance/oo4onq4s",
   "tg_tokens": null,
   "tps": 37.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-oo4onq4s-ctx16384",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: soonujxn, yc23abgv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 229.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 229.7 29.4 19.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oo4onq4s",
   "source_url": "https://omlx.ai/benchmarks/performance/oo4onq4s",
   "tg_tokens": null,
   "tps": 29.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-mot774o5-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP MLX build; oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: j79kzt47, qg96x47w, ukz418fa; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 765.7,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.6-35B-A3B-oQ3e-MTP-MLX on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 3bit; 1k 765.7 70.8 16.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mot774o5",
   "source_url": "https://omlx.ai/benchmarks/performance/mot774o5",
   "tg_tokens": null,
   "tps": 70.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-mot774o5-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP MLX build; oMLX community benchmark trace; peak memory 16.8 GB at this context; duplicate traces of the same session: j79kzt47, qg96x47w, ukz418fa; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 744.9,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.6-35B-A3B-oQ3e-MTP-MLX on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 3bit; 4k 744.9 68.0 16.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mot774o5",
   "source_url": "https://omlx.ai/benchmarks/performance/mot774o5",
   "tg_tokens": null,
   "tps": 68.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-mot774o5-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP MLX build; oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: j79kzt47, qg96x47w, ukz418fa; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 685.9,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.6-35B-A3B-oQ3e-MTP-MLX on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 3bit; 8k 685.9 65.6 16.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mot774o5",
   "source_url": "https://omlx.ai/benchmarks/performance/mot774o5",
   "tg_tokens": null,
   "tps": 65.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-mot774o5-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP MLX build; oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: j79kzt47, qg96x47w, ukz418fa; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 588.4,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.6-35B-A3B-oQ3e-MTP-MLX on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 3bit; 16k 588.4 59.7 17.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mot774o5",
   "source_url": "https://omlx.ai/benchmarks/performance/mot774o5",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q7apdb8e-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: 5nq0mg2r, 6hb3voru, 6n3vvom0, 8ajhif1k, d0aas3xx, yo0ez3kv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 938.3,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 938.3 88.4 26.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q7apdb8e",
   "source_url": "https://omlx.ai/benchmarks/performance/q7apdb8e",
   "tg_tokens": null,
   "tps": 88.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-q7apdb8e-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: 5nq0mg2r, 6hb3voru, 6n3vvom0, 8ajhif1k, d0aas3xx, yo0ez3kv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3486.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 3,486 180.3 26.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q7apdb8e",
   "source_url": "https://omlx.ai/benchmarks/performance/q7apdb8e",
   "tg_tokens": null,
   "tps": 180.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q7apdb8e-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: 5nq0mg2r, 6hb3voru, 6n3vvom0, 8ajhif1k, d0aas3xx, yo0ez3kv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5209.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 5,209 37.1 26.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q7apdb8e",
   "source_url": "https://omlx.ai/benchmarks/performance/q7apdb8e",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q7apdb8e-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.0 GB at this context; duplicate traces of the same session: 5nq0mg2r, 6hb3voru, 6n3vvom0, 8ajhif1k, d0aas3xx, yo0ez3kv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4927.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 4,927 142.2 27.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q7apdb8e",
   "source_url": "https://omlx.ai/benchmarks/performance/q7apdb8e",
   "tg_tokens": null,
   "tps": 142.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q7apdb8e-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: 5nq0mg2r, 6hb3voru, 6n3vvom0, 8ajhif1k, d0aas3xx, yo0ez3kv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4115.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 4,115 222.0 27.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q7apdb8e",
   "source_url": "https://omlx.ai/benchmarks/performance/q7apdb8e",
   "tg_tokens": null,
   "tps": 222.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-q7apdb8e-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: 5nq0mg2r, 6hb3voru, 6n3vvom0, 8ajhif1k, d0aas3xx, yo0ez3kv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3048.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 3,048 138.4 28.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q7apdb8e",
   "source_url": "https://omlx.ai/benchmarks/performance/q7apdb8e",
   "tg_tokens": null,
   "tps": 138.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q7apdb8e-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: 5nq0mg2r, 6hb3voru, 6n3vvom0, 8ajhif1k, d0aas3xx, yo0ez3kv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2049.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.6-35B-A3B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 2,049 92.9 29.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q7apdb8e",
   "source_url": "https://omlx.ai/benchmarks/performance/q7apdb8e",
   "tg_tokens": null,
   "tps": 92.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-114deaeo-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: k6o53w43; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 32.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 32.3 6.1 15.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/114deaeo",
   "source_url": "https://omlx.ai/benchmarks/performance/114deaeo",
   "tg_tokens": null,
   "tps": 6.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-114deaeo-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: k6o53w43; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 36.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 36.3 5.6 17.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/114deaeo",
   "source_url": "https://omlx.ai/benchmarks/performance/114deaeo",
   "tg_tokens": null,
   "tps": 5.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1ec49biz-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: 2qqvrjzl, 53mhrjzo, 6m1cjbcg, 8ijczinn, hfoxk9u4, xelgizsr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 951.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 951.1 64.8 24.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ec49biz",
   "source_url": "https://omlx.ai/benchmarks/performance/1ec49biz",
   "tg_tokens": null,
   "tps": 64.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1ec49biz-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.6 GB at this context; duplicate traces of the same session: 2qqvrjzl, 53mhrjzo, 6m1cjbcg, 8ijczinn, hfoxk9u4, xelgizsr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1147.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 1,147 57.3 25.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ec49biz",
   "source_url": "https://omlx.ai/benchmarks/performance/1ec49biz",
   "tg_tokens": null,
   "tps": 57.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1ec49biz-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: 2qqvrjzl, 53mhrjzo, 6m1cjbcg, 8ijczinn, hfoxk9u4, xelgizsr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1082.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 1,082 69.0 25.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ec49biz",
   "source_url": "https://omlx.ai/benchmarks/performance/1ec49biz",
   "tg_tokens": null,
   "tps": 69.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1ec49biz-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: 2qqvrjzl, 53mhrjzo, 6m1cjbcg, 8ijczinn, hfoxk9u4, xelgizsr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 971.3,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 971.3 71.3 26.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ec49biz",
   "source_url": "https://omlx.ai/benchmarks/performance/1ec49biz",
   "tg_tokens": null,
   "tps": 71.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1ec49biz-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.6 GB at this context; duplicate traces of the same session: 2qqvrjzl, 53mhrjzo, 6m1cjbcg, 8ijczinn, hfoxk9u4, xelgizsr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 870.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 870.9 120.7 27.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ec49biz",
   "source_url": "https://omlx.ai/benchmarks/performance/1ec49biz",
   "tg_tokens": null,
   "tps": 120.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1ec49biz-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.0 GB at this context; duplicate traces of the same session: 2qqvrjzl, 53mhrjzo, 6m1cjbcg, 8ijczinn, hfoxk9u4, xelgizsr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 690.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 690.2 40.3 30.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ec49biz",
   "source_url": "https://omlx.ai/benchmarks/performance/1ec49biz",
   "tg_tokens": null,
   "tps": 40.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1ec49biz-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: 2qqvrjzl, 53mhrjzo, 6m1cjbcg, 8ijczinn, hfoxk9u4, xelgizsr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 530.6,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 530.6 38.0 34.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ec49biz",
   "source_url": "https://omlx.ai/benchmarks/performance/1ec49biz",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-3qf436dt-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 72.2 GB at this context; duplicate traces of the same session: 45pzb3r1, astkasnt; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 329.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 329.2 29.5 72.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3qf436dt",
   "source_url": "https://omlx.ai/benchmarks/performance/3qf436dt",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-3qf436dt-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 74.5 GB at this context; duplicate traces of the same session: 45pzb3r1, astkasnt; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 424.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 424.6 25.0 74.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3qf436dt",
   "source_url": "https://omlx.ai/benchmarks/performance/3qf436dt",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-3qf436dt-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 74.9 GB at this context; duplicate traces of the same session: 45pzb3r1, astkasnt; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 502.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 502.2 34.8 74.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3qf436dt",
   "source_url": "https://omlx.ai/benchmarks/performance/3qf436dt",
   "tg_tokens": null,
   "tps": 34.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-3qf436dt-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 75.2 GB at this context; duplicate traces of the same session: 45pzb3r1, astkasnt; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 507.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 507.9 31.9 75.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3qf436dt",
   "source_url": "https://omlx.ai/benchmarks/performance/3qf436dt",
   "tg_tokens": null,
   "tps": 31.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-3qf436dt-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 75.7 GB at this context; duplicate traces of the same session: 45pzb3r1, astkasnt; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 489.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 32k 489.4 27.6 75.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3qf436dt",
   "source_url": "https://omlx.ai/benchmarks/performance/3qf436dt",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-oj4roqa2-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 72.2 GB at this context; duplicate traces of the same session: e1rrkect, j9806jut, v17tvbdc, xq5cnxbw; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 307.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 307.1 25.5 72.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oj4roqa2",
   "source_url": "https://omlx.ai/benchmarks/performance/oj4roqa2",
   "tg_tokens": null,
   "tps": 25.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-oj4roqa2-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 74.5 GB at this context; duplicate traces of the same session: e1rrkect, j9806jut, v17tvbdc, xq5cnxbw; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 371.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 371.1 36.7 74.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oj4roqa2",
   "source_url": "https://omlx.ai/benchmarks/performance/oj4roqa2",
   "tg_tokens": null,
   "tps": 36.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-oj4roqa2-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 74.9 GB at this context; duplicate traces of the same session: e1rrkect, j9806jut, v17tvbdc, xq5cnxbw; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 529.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 529.6 37.5 74.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oj4roqa2",
   "source_url": "https://omlx.ai/benchmarks/performance/oj4roqa2",
   "tg_tokens": null,
   "tps": 37.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-oj4roqa2-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 75.2 GB at this context; duplicate traces of the same session: e1rrkect, j9806jut, v17tvbdc, xq5cnxbw; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 495.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 495.9 32.7 75.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oj4roqa2",
   "source_url": "https://omlx.ai/benchmarks/performance/oj4roqa2",
   "tg_tokens": null,
   "tps": 32.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-oj4roqa2-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 75.7 GB at this context; duplicate traces of the same session: e1rrkect, j9806jut, v17tvbdc, xq5cnxbw; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 500.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 32k 500.3 28.6 75.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oj4roqa2",
   "source_url": "https://omlx.ai/benchmarks/performance/oj4roqa2",
   "tg_tokens": null,
   "tps": 28.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-owagf081-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace, single run at 1k context; peak memory 76.0 GB; TTFT 1648 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 621.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 621.4 PP tok/s 18.3 TG tok/s 1648 TTFT (ms) 76.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/owagf081",
   "source_url": "https://omlx.ai/benchmarks/performance/owagf081",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ehzr9jzm-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6 mixed-quant MTP v2 build; oMLX community benchmark trace, single run at 16k context; peak memory 116.9 GB; TTFT 11547 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1419.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp-v2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 16k tokens 1,419 PP tok/s 61.4 TG tok/s 11547 TTFT (ms) 116.9 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ehzr9jzm",
   "source_url": "https://omlx.ai/benchmarks/performance/ehzr9jzm",
   "tg_tokens": null,
   "tps": 61.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-qrjz0izw-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: 0yor5b1i, 523wdn77, b5g94ly1, gwhmne8i, q14z8c36, uv2z8iq3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2471.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 2,471 134.0 100.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qrjz0izw",
   "source_url": "https://omlx.ai/benchmarks/performance/qrjz0izw",
   "tg_tokens": null,
   "tps": 134.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-qrjz0izw-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: 0yor5b1i, 523wdn77, b5g94ly1, gwhmne8i, q14z8c36, uv2z8iq3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4829.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 4,829 139.4 101.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qrjz0izw",
   "source_url": "https://omlx.ai/benchmarks/performance/qrjz0izw",
   "tg_tokens": null,
   "tps": 139.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-qrjz0izw-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.8 GB at this context; duplicate traces of the same session: 0yor5b1i, 523wdn77, b5g94ly1, gwhmne8i, q14z8c36, uv2z8iq3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5335.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 5,335 157.2 102.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qrjz0izw",
   "source_url": "https://omlx.ai/benchmarks/performance/qrjz0izw",
   "tg_tokens": null,
   "tps": 157.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-qrjz0izw-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.4 GB at this context; duplicate traces of the same session: 0yor5b1i, 523wdn77, b5g94ly1, gwhmne8i, q14z8c36, uv2z8iq3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5243.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 5,243 131.5 103.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qrjz0izw",
   "source_url": "https://omlx.ai/benchmarks/performance/qrjz0izw",
   "tg_tokens": null,
   "tps": 131.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-qrjz0izw-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.5 GB at this context; duplicate traces of the same session: 0yor5b1i, 523wdn77, b5g94ly1, gwhmne8i, q14z8c36, uv2z8iq3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5175.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 5,175 124.4 104.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qrjz0izw",
   "source_url": "https://omlx.ai/benchmarks/performance/qrjz0izw",
   "tg_tokens": null,
   "tps": 124.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-qrjz0izw-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 106.8 GB at this context; duplicate traces of the same session: 0yor5b1i, 523wdn77, b5g94ly1, gwhmne8i, q14z8c36, uv2z8iq3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5022.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 5,022 101.4 106.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qrjz0izw",
   "source_url": "https://omlx.ai/benchmarks/performance/qrjz0izw",
   "tg_tokens": null,
   "tps": 101.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-qrjz0izw-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.6 GB at this context; duplicate traces of the same session: 0yor5b1i, 523wdn77, b5g94ly1, gwhmne8i, q14z8c36, uv2z8iq3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4861.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 4,861 81.2 109.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qrjz0izw",
   "source_url": "https://omlx.ai/benchmarks/performance/qrjz0izw",
   "tg_tokens": null,
   "tps": 81.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-g36qjpkl-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: 0efhxhn3, 4f3z4292, 5o17y4qn, aitk9xoj, dpkw9g4u, ojvsl5el, ywy0qztu; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1742.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,742 140.0 100.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g36qjpkl",
   "source_url": "https://omlx.ai/benchmarks/performance/g36qjpkl",
   "tg_tokens": null,
   "tps": 140.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-g36qjpkl-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: 0efhxhn3, 4f3z4292, 5o17y4qn, aitk9xoj, dpkw9g4u, ojvsl5el, ywy0qztu; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4837.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 4,837 132.6 101.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g36qjpkl",
   "source_url": "https://omlx.ai/benchmarks/performance/g36qjpkl",
   "tg_tokens": null,
   "tps": 132.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-g36qjpkl-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.2 GB at this context; duplicate traces of the same session: 0efhxhn3, 4f3z4292, 5o17y4qn, aitk9xoj, dpkw9g4u, ojvsl5el, ywy0qztu; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5286.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 5,286 135.8 102.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g36qjpkl",
   "source_url": "https://omlx.ai/benchmarks/performance/g36qjpkl",
   "tg_tokens": null,
   "tps": 135.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-g36qjpkl-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.8 GB at this context; duplicate traces of the same session: 0efhxhn3, 4f3z4292, 5o17y4qn, aitk9xoj, dpkw9g4u, ojvsl5el, ywy0qztu; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5341.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 5,341 138.4 102.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g36qjpkl",
   "source_url": "https://omlx.ai/benchmarks/performance/g36qjpkl",
   "tg_tokens": null,
   "tps": 138.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-g36qjpkl-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.4 GB at this context; duplicate traces of the same session: 0efhxhn3, 4f3z4292, 5o17y4qn, aitk9xoj, dpkw9g4u, ojvsl5el, ywy0qztu; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5231.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 5,231 150.5 103.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g36qjpkl",
   "source_url": "https://omlx.ai/benchmarks/performance/g36qjpkl",
   "tg_tokens": null,
   "tps": 150.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-g36qjpkl-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.5 GB at this context; duplicate traces of the same session: 0efhxhn3, 4f3z4292, 5o17y4qn, aitk9xoj, dpkw9g4u, ojvsl5el, ywy0qztu; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5175.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 5,175 126.8 104.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g36qjpkl",
   "source_url": "https://omlx.ai/benchmarks/performance/g36qjpkl",
   "tg_tokens": null,
   "tps": 126.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-g36qjpkl-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 106.9 GB at this context; duplicate traces of the same session: 0efhxhn3, 4f3z4292, 5o17y4qn, aitk9xoj, dpkw9g4u, ojvsl5el, ywy0qztu; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5023.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 5,023 108.2 106.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g36qjpkl",
   "source_url": "https://omlx.ai/benchmarks/performance/g36qjpkl",
   "tg_tokens": null,
   "tps": 108.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-g36qjpkl-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.6 GB at this context; duplicate traces of the same session: 0efhxhn3, 4f3z4292, 5o17y4qn, aitk9xoj, dpkw9g4u, ojvsl5el, ywy0qztu; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4854.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 4,854 80.9 109.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g36qjpkl",
   "source_url": "https://omlx.ai/benchmarks/performance/g36qjpkl",
   "tg_tokens": null,
   "tps": 80.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-kui0nabd-ctx1024",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: gti0nbdc; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 626.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 626.4 94.2 18.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kui0nabd",
   "source_url": "https://omlx.ai/benchmarks/performance/kui0nabd",
   "tg_tokens": null,
   "tps": 94.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-kui0nabd-ctx4096",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: gti0nbdc; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 682.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 682.3 93.2 19.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kui0nabd",
   "source_url": "https://omlx.ai/benchmarks/performance/kui0nabd",
   "tg_tokens": null,
   "tps": 93.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-kui0nabd-ctx8192",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: gti0nbdc; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 677.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 8k 677.1 86.5 19.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kui0nabd",
   "source_url": "https://omlx.ai/benchmarks/performance/kui0nabd",
   "tg_tokens": null,
   "tps": 86.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jmw236ya-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.0 GB at this context; duplicate traces of the same session: d19pbkle, gxzxkupk; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 449.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 449.5 45.5 179.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jmw236ya",
   "source_url": "https://omlx.ai/benchmarks/performance/jmw236ya",
   "tg_tokens": null,
   "tps": 45.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-jmw236ya-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.2 GB at this context; duplicate traces of the same session: d19pbkle, gxzxkupk; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 416.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 416.8 52.9 179.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jmw236ya",
   "source_url": "https://omlx.ai/benchmarks/performance/jmw236ya",
   "tg_tokens": null,
   "tps": 52.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jmw236ya-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.3 GB at this context; duplicate traces of the same session: d19pbkle, gxzxkupk; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 408.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 408.7 85.2 179.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jmw236ya",
   "source_url": "https://omlx.ai/benchmarks/performance/jmw236ya",
   "tg_tokens": null,
   "tps": 85.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-wsrtau47-ctx1024",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 17.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 198.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 198.9 53.9 17.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wsrtau47",
   "source_url": "https://omlx.ai/benchmarks/performance/wsrtau47",
   "tg_tokens": null,
   "tps": 53.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-wsrtau47-ctx4096",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 18.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 233.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 233.9 78.9 18.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wsrtau47",
   "source_url": "https://omlx.ai/benchmarks/performance/wsrtau47",
   "tg_tokens": null,
   "tps": 78.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-wsrtau47-ctx8192",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 236.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 236.7 37.9 18.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wsrtau47",
   "source_url": "https://omlx.ai/benchmarks/performance/wsrtau47",
   "tg_tokens": null,
   "tps": 37.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-wsrtau47-ctx16384",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 229.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 229.7 29.4 19.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wsrtau47",
   "source_url": "https://omlx.ai/benchmarks/performance/wsrtau47",
   "tg_tokens": null,
   "tps": 29.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-jxq7pu8t-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 3.5 GB at this context; duplicate traces of the same session: 62gef2x9, nvs9mwc1, oo1tj90h; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 655.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-4B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 655.8 75.6 3.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jxq7pu8t",
   "source_url": "https://omlx.ai/benchmarks/performance/jxq7pu8t",
   "tg_tokens": null,
   "tps": 75.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-jxq7pu8t-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.1 GB at this context; duplicate traces of the same session: 62gef2x9, nvs9mwc1, oo1tj90h; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 687.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-4B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 687.7 75.0 4.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jxq7pu8t",
   "source_url": "https://omlx.ai/benchmarks/performance/jxq7pu8t",
   "tg_tokens": null,
   "tps": 75.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-jxq7pu8t-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.3 GB at this context; duplicate traces of the same session: 62gef2x9, nvs9mwc1, oo1tj90h; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 672.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-4B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 672.5 71.4 4.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jxq7pu8t",
   "source_url": "https://omlx.ai/benchmarks/performance/jxq7pu8t",
   "tg_tokens": null,
   "tps": 71.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-jxq7pu8t-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.6 GB at this context; duplicate traces of the same session: 62gef2x9, nvs9mwc1, oo1tj90h; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 629.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-4B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 629.6 64.5 4.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jxq7pu8t",
   "source_url": "https://omlx.ai/benchmarks/performance/jxq7pu8t",
   "tg_tokens": null,
   "tps": 64.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-6w5i7t9e-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 6.6 GB; TTFT 995 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1029.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-9B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 24 GB \u00b7 4bit; 1k tokens 1,029 PP tok/s 25.7 TG tok/s 995 TTFT (ms) 6.6 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6w5i7t9e",
   "source_url": "https://omlx.ai/benchmarks/performance/6w5i7t9e",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qxqwb6gf-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 1u0nr6id; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2380.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 2,380 107.9 19.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qxqwb6gf",
   "source_url": "https://omlx.ai/benchmarks/performance/qxqwb6gf",
   "tg_tokens": null,
   "tps": 107.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qxqwb6gf-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 1u0nr6id; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2928.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,928 104.0 20.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qxqwb6gf",
   "source_url": "https://omlx.ai/benchmarks/performance/qxqwb6gf",
   "tg_tokens": null,
   "tps": 104.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qxqwb6gf-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 1u0nr6id; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2155.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 2,155 75.7 20.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qxqwb6gf",
   "source_url": "https://omlx.ai/benchmarks/performance/qxqwb6gf",
   "tg_tokens": null,
   "tps": 75.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qxqwb6gf-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 1u0nr6id; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1584.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 64k 1,584 60.7 21.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qxqwb6gf",
   "source_url": "https://omlx.ai/benchmarks/performance/qxqwb6gf",
   "tg_tokens": null,
   "tps": 60.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qxqwb6gf-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: 1u0nr6id; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1083.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 128k 1,083 42.6 22.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qxqwb6gf",
   "source_url": "https://omlx.ai/benchmarks/performance/qxqwb6gf",
   "tg_tokens": null,
   "tps": 42.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-p5wra5aq-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "FP16 build; oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: 3horev01, fwoikbx6; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1068.0,
   "provenance": "community",
   "quant": "oQ3.5",
   "quote": "Qwen3.6-35B-A3B-oQ3.5-FP16 on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 3bit; 1k 1,068 86.2 17.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p5wra5aq",
   "source_url": "https://omlx.ai/benchmarks/performance/p5wra5aq",
   "tg_tokens": null,
   "tps": 86.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-p5wra5aq-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "FP16 build; oMLX community benchmark trace; peak memory 17.9 GB at this context; duplicate traces of the same session: 3horev01, fwoikbx6; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1045.0,
   "provenance": "community",
   "quant": "oQ3.5",
   "quote": "Qwen3.6-35B-A3B-oQ3.5-FP16 on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 3bit; 4k 1,045 82.6 17.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p5wra5aq",
   "source_url": "https://omlx.ai/benchmarks/performance/p5wra5aq",
   "tg_tokens": null,
   "tps": 82.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-p5wra5aq-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "FP16 build; oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: 3horev01, fwoikbx6; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1073.0,
   "provenance": "community",
   "quant": "oQ3.5",
   "quote": "Qwen3.6-35B-A3B-oQ3.5-FP16 on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 3bit; 8k 1,073 75.6 18.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p5wra5aq",
   "source_url": "https://omlx.ai/benchmarks/performance/p5wra5aq",
   "tg_tokens": null,
   "tps": 75.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-v3fbqnre-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "FP16 build; oMLX community benchmark trace, single run at 1k context; peak memory 17.2 GB; TTFT 1349 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 759.3,
   "provenance": "community",
   "quant": "oQ3.5",
   "quote": "Qwen3.6-35B-A3B-oQ3.5-FP16 on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 3bit; 1k tokens 759.3 PP tok/s 89.0 TG tok/s 1349 TTFT (ms) 17.2 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v3fbqnre",
   "source_url": "https://omlx.ai/benchmarks/performance/v3fbqnre",
   "tg_tokens": null,
   "tps": 89.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-id0ii1sv-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 457.4,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.6-35B-A3B-oQ3.5e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 3bit; 32k 457.4 54.0 19.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/id0ii1sv",
   "source_url": "https://omlx.ai/benchmarks/performance/id0ii1sv",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-id0ii1sv-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 315.8,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.6-35B-A3B-oQ3.5e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 3bit; 64k 315.8 44.1 20.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/id0ii1sv",
   "source_url": "https://omlx.ai/benchmarks/performance/id0ii1sv",
   "tg_tokens": null,
   "tps": 44.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-7agey0kg-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP MLX build; oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 2h9jiqxr, 8nk4j596, 8vywxs71, c5t0ybeh, yopecq5p; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 764.0,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.6-35B-A3B-oQ3e-MTP-MLX on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 3bit; 1k 764.0 71.1 16.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7agey0kg",
   "source_url": "https://omlx.ai/benchmarks/performance/7agey0kg",
   "tg_tokens": null,
   "tps": 71.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-7agey0kg-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP MLX build; oMLX community benchmark trace; peak memory 16.8 GB at this context; duplicate traces of the same session: 2h9jiqxr, 8nk4j596, 8vywxs71, c5t0ybeh, yopecq5p; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 742.1,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.6-35B-A3B-oQ3e-MTP-MLX on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 3bit; 4k 742.1 68.5 16.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7agey0kg",
   "source_url": "https://omlx.ai/benchmarks/performance/7agey0kg",
   "tg_tokens": null,
   "tps": 68.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-7agey0kg-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP MLX build; oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: 2h9jiqxr, 8nk4j596, 8vywxs71, c5t0ybeh, yopecq5p; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 683.5,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.6-35B-A3B-oQ3e-MTP-MLX on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 3bit; 8k 683.5 65.1 16.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7agey0kg",
   "source_url": "https://omlx.ai/benchmarks/performance/7agey0kg",
   "tg_tokens": null,
   "tps": 65.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-7agey0kg-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP MLX build; oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 2h9jiqxr, 8nk4j596, 8vywxs71, c5t0ybeh, yopecq5p; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 584.3,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.6-35B-A3B-oQ3e-MTP-MLX on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 3bit; 16k 584.3 60.2 17.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7agey0kg",
   "source_url": "https://omlx.ai/benchmarks/performance/7agey0kg",
   "tg_tokens": null,
   "tps": 60.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-7agey0kg-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP MLX build; oMLX community benchmark trace; peak memory 17.9 GB at this context; duplicate traces of the same session: 2h9jiqxr, 8nk4j596, 8vywxs71, c5t0ybeh, yopecq5p; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 454.3,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.6-35B-A3B-oQ3e-MTP-MLX on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 3bit; 32k 454.3 54.0 17.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7agey0kg",
   "source_url": "https://omlx.ai/benchmarks/performance/7agey0kg",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-7agey0kg-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP MLX build; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 2h9jiqxr, 8nk4j596, 8vywxs71, c5t0ybeh, yopecq5p; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 314.2,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.6-35B-A3B-oQ3e-MTP-MLX on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 3bit; 64k 314.2 44.0 19.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7agey0kg",
   "source_url": "https://omlx.ai/benchmarks/performance/7agey0kg",
   "tg_tokens": null,
   "tps": 44.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-guffud5p-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 19.6 GB; TTFT 51737 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 158.3,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 8k tokens 158.3 PP tok/s 19.4 TG tok/s 51737 TTFT (ms) 19.6 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/guffud5p",
   "source_url": "https://omlx.ai/benchmarks/performance/guffud5p",
   "tg_tokens": null,
   "tps": 19.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-jh09nmkq-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 18.5 GB; TTFT 51069 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 160.4,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 8k tokens 160.4 PP tok/s 20.8 TG tok/s 51069 TTFT (ms) 18.5 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jh09nmkq",
   "source_url": "https://omlx.ai/benchmarks/performance/jh09nmkq",
   "tg_tokens": null,
   "tps": 20.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-upvha7iv-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 0w6jrjyo, 3q1qihfi, a5wzjpea, qe77ui00, qscmxexa, v2b99we5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 730.8,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 730.8 63.5 23.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/upvha7iv",
   "source_url": "https://omlx.ai/benchmarks/performance/upvha7iv",
   "tg_tokens": null,
   "tps": 63.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-upvha7iv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: 0w6jrjyo, 3q1qihfi, a5wzjpea, qe77ui00, qscmxexa, v2b99we5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 779.6,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 779.6 47.7 24.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/upvha7iv",
   "source_url": "https://omlx.ai/benchmarks/performance/upvha7iv",
   "tg_tokens": null,
   "tps": 47.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-upvha7iv-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: 0w6jrjyo, 3q1qihfi, a5wzjpea, qe77ui00, qscmxexa, v2b99we5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 865.5,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 865.5 65.8 24.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/upvha7iv",
   "source_url": "https://omlx.ai/benchmarks/performance/upvha7iv",
   "tg_tokens": null,
   "tps": 65.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-upvha7iv-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: 0w6jrjyo, 3q1qihfi, a5wzjpea, qe77ui00, qscmxexa, v2b99we5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 784.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 784.9 79.5 25.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/upvha7iv",
   "source_url": "https://omlx.ai/benchmarks/performance/upvha7iv",
   "tg_tokens": null,
   "tps": 79.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-upvha7iv-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: 0w6jrjyo, 3q1qihfi, a5wzjpea, qe77ui00, qscmxexa, v2b99we5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 708.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 708.2 119.1 26.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/upvha7iv",
   "source_url": "https://omlx.ai/benchmarks/performance/upvha7iv",
   "tg_tokens": null,
   "tps": 119.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-upvha7iv-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.7 GB at this context; duplicate traces of the same session: 0w6jrjyo, 3q1qihfi, a5wzjpea, qe77ui00, qscmxexa, v2b99we5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 593.8,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 593.8 46.7 28.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/upvha7iv",
   "source_url": "https://omlx.ai/benchmarks/performance/upvha7iv",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-upvha7iv-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.4 GB at this context; duplicate traces of the same session: 0w6jrjyo, 3q1qihfi, a5wzjpea, qe77ui00, qscmxexa, v2b99we5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 456.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 456.7 43.8 33.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/upvha7iv",
   "source_url": "https://omlx.ai/benchmarks/performance/upvha7iv",
   "tg_tokens": null,
   "tps": 43.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-tv4bpycg-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: b9ufnh8p, fufofrd6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 446.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 446.7 54.1 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tv4bpycg",
   "source_url": "https://omlx.ai/benchmarks/performance/tv4bpycg",
   "tg_tokens": null,
   "tps": 54.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-tv4bpycg-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: b9ufnh8p, fufofrd6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 582.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 582.5 50.6 71.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tv4bpycg",
   "source_url": "https://omlx.ai/benchmarks/performance/tv4bpycg",
   "tg_tokens": null,
   "tps": 50.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-tv4bpycg-ctx16384",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.0 GB at this context; duplicate traces of the same session: b9ufnh8p, fufofrd6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 611.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 611.1 44.0 72.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tv4bpycg",
   "source_url": "https://omlx.ai/benchmarks/performance/tv4bpycg",
   "tg_tokens": null,
   "tps": 44.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-b6g80spl-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: sdvbha0x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 456.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 456.8 39.7 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b6g80spl",
   "source_url": "https://omlx.ai/benchmarks/performance/b6g80spl",
   "tg_tokens": null,
   "tps": 39.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-b6g80spl-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: sdvbha0x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 580.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 580.8 49.1 71.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b6g80spl",
   "source_url": "https://omlx.ai/benchmarks/performance/b6g80spl",
   "tg_tokens": null,
   "tps": 49.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-vr1b0sto-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: xqssb245; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 376.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 376.0 43.2 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vr1b0sto",
   "source_url": "https://omlx.ai/benchmarks/performance/vr1b0sto",
   "tg_tokens": null,
   "tps": 43.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-vr1b0sto-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: xqssb245; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 476.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 476.2 43.9 71.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vr1b0sto",
   "source_url": "https://omlx.ai/benchmarks/performance/vr1b0sto",
   "tg_tokens": null,
   "tps": 43.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-1n0h4ryd-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 71.5 GB; TTFT 6391 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 640.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 640.9 PP tok/s 55.2 TG tok/s 6391 TTFT (ms) 71.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1n0h4ryd",
   "source_url": "https://omlx.ai/benchmarks/performance/1n0h4ryd",
   "tg_tokens": null,
   "tps": 55.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-7nk3adh0-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.8 GB; TTFT 2144 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 477.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 477.7 PP tok/s 24.1 TG tok/s 2144 TTFT (ms) 70.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7nk3adh0",
   "source_url": "https://omlx.ai/benchmarks/performance/7nk3adh0",
   "tg_tokens": null,
   "tps": 24.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-075lryio-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 71.1 GB; TTFT 856 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1197.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,197 PP tok/s 68.3 TG tok/s 856 TTFT (ms) 71.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/075lryio",
   "source_url": "https://omlx.ai/benchmarks/performance/075lryio",
   "tg_tokens": null,
   "tps": 68.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4kn1nm72-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: 50uy335g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1260.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,260 83.9 71.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4kn1nm72",
   "source_url": "https://omlx.ai/benchmarks/performance/4kn1nm72",
   "tg_tokens": null,
   "tps": 83.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4kn1nm72-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.2 GB at this context; duplicate traces of the same session: 50uy335g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1984.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,984 69.7 71.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4kn1nm72",
   "source_url": "https://omlx.ai/benchmarks/performance/4kn1nm72",
   "tg_tokens": null,
   "tps": 69.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nxez8b4j-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: 14fmf7cr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1196.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,196 85.0 71.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nxez8b4j",
   "source_url": "https://omlx.ai/benchmarks/performance/nxez8b4j",
   "tg_tokens": null,
   "tps": 85.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nxez8b4j-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.2 GB at this context; duplicate traces of the same session: 14fmf7cr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1974.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,974 65.1 71.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nxez8b4j",
   "source_url": "https://omlx.ai/benchmarks/performance/nxez8b4j",
   "tg_tokens": null,
   "tps": 65.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-eux1pn14-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 71.1 GB; TTFT 796 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1287.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,287 PP tok/s 71.1 TG tok/s 796 TTFT (ms) 71.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eux1pn14",
   "source_url": "https://omlx.ai/benchmarks/performance/eux1pn14",
   "tg_tokens": null,
   "tps": 71.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-xzkqnq1g-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift build; oQ4e/oQ6e mixed quant; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 77.5 GB at this context; duplicate traces of the same session: dpsz00s7, t0thy8hg, y2tc9b6j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 625.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 625.3 65.6 77.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xzkqnq1g",
   "source_url": "https://omlx.ai/benchmarks/performance/xzkqnq1g",
   "tg_tokens": null,
   "tps": 65.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-xzkqnq1g-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift build; oQ4e/oQ6e mixed quant; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.1 GB at this context; duplicate traces of the same session: dpsz00s7, t0thy8hg, y2tc9b6j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 759.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 759.5 71.3 78.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xzkqnq1g",
   "source_url": "https://omlx.ai/benchmarks/performance/xzkqnq1g",
   "tg_tokens": null,
   "tps": 71.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-xzkqnq1g-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift build; oQ4e/oQ6e mixed quant; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.5 GB at this context; duplicate traces of the same session: dpsz00s7, t0thy8hg, y2tc9b6j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 826.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 826.5 55.1 78.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xzkqnq1g",
   "source_url": "https://omlx.ai/benchmarks/performance/xzkqnq1g",
   "tg_tokens": null,
   "tps": 55.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-xzkqnq1g-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift build; oQ4e/oQ6e mixed quant; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.4 GB at this context; duplicate traces of the same session: dpsz00s7, t0thy8hg, y2tc9b6j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 817.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 817.7 52.9 79.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xzkqnq1g",
   "source_url": "https://omlx.ai/benchmarks/performance/xzkqnq1g",
   "tg_tokens": null,
   "tps": 52.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-kjvvlt3v-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift build; oQ4e/oQ6e mixed quant; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 77.5 GB at this context; duplicate traces of the same session: 0859rcio, 6d1yaucc, mhodrjie; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 509.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 509.9 29.8 77.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kjvvlt3v",
   "source_url": "https://omlx.ai/benchmarks/performance/kjvvlt3v",
   "tg_tokens": null,
   "tps": 29.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-kjvvlt3v-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift build; oQ4e/oQ6e mixed quant; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.1 GB at this context; duplicate traces of the same session: 0859rcio, 6d1yaucc, mhodrjie; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 670.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 670.1 62.2 78.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kjvvlt3v",
   "source_url": "https://omlx.ai/benchmarks/performance/kjvvlt3v",
   "tg_tokens": null,
   "tps": 62.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-kjvvlt3v-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift build; oQ4e/oQ6e mixed quant; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.5 GB at this context; duplicate traces of the same session: 0859rcio, 6d1yaucc, mhodrjie; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 825.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 825.1 70.3 78.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kjvvlt3v",
   "source_url": "https://omlx.ai/benchmarks/performance/kjvvlt3v",
   "tg_tokens": null,
   "tps": 70.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-kjvvlt3v-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift build; oQ4e/oQ6e mixed quant; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.8 GB at this context; duplicate traces of the same session: 0859rcio, 6d1yaucc, mhodrjie; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 822.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 822.8 31.3 78.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kjvvlt3v",
   "source_url": "https://omlx.ai/benchmarks/performance/kjvvlt3v",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-mchhubyi-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift build; oQ4e/oQ6e mixed quant; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 77.5 GB at this context; duplicate traces of the same session: eo47aulx, gs14l12g, k6hu03pq, nzaxkygw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 619.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 619.9 73.5 77.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mchhubyi",
   "source_url": "https://omlx.ai/benchmarks/performance/mchhubyi",
   "tg_tokens": null,
   "tps": 73.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-mchhubyi-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift build; oQ4e/oQ6e mixed quant; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.1 GB at this context; duplicate traces of the same session: eo47aulx, gs14l12g, k6hu03pq, nzaxkygw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 757.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 757.4 66.8 78.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mchhubyi",
   "source_url": "https://omlx.ai/benchmarks/performance/mchhubyi",
   "tg_tokens": null,
   "tps": 66.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-mchhubyi-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift build; oQ4e/oQ6e mixed quant; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.5 GB at this context; duplicate traces of the same session: eo47aulx, gs14l12g, k6hu03pq, nzaxkygw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 828.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 828.7 72.4 78.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mchhubyi",
   "source_url": "https://omlx.ai/benchmarks/performance/mchhubyi",
   "tg_tokens": null,
   "tps": 72.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-mchhubyi-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift build; oQ4e/oQ6e mixed quant; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.8 GB at this context; duplicate traces of the same session: eo47aulx, gs14l12g, k6hu03pq, nzaxkygw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 823.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 823.4 81.9 78.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mchhubyi",
   "source_url": "https://omlx.ai/benchmarks/performance/mchhubyi",
   "tg_tokens": null,
   "tps": 81.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-mchhubyi-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift build; oQ4e/oQ6e mixed quant; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.4 GB at this context; duplicate traces of the same session: eo47aulx, gs14l12g, k6hu03pq, nzaxkygw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 817.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 817.4 51.6 79.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mchhubyi",
   "source_url": "https://omlx.ai/benchmarks/performance/mchhubyi",
   "tg_tokens": null,
   "tps": 51.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-himj1fj4-ctx1024",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 626.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 626.4 94.2 18.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/himj1fj4",
   "source_url": "https://omlx.ai/benchmarks/performance/himj1fj4",
   "tg_tokens": null,
   "tps": 94.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-himj1fj4-ctx4096",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 682.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 682.3 93.2 19.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/himj1fj4",
   "source_url": "https://omlx.ai/benchmarks/performance/himj1fj4",
   "tg_tokens": null,
   "tps": 93.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-himj1fj4-ctx8192",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 677.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 8k 677.1 86.5 19.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/himj1fj4",
   "source_url": "https://omlx.ai/benchmarks/performance/himj1fj4",
   "tg_tokens": null,
   "tps": 86.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-14qw86xi-ctx1024",
   "model": "Kolibri-1",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 42.7 GB at this context; duplicate traces of the same session: 5y5uqzsk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1333.0,
   "provenance": "community",
   "quant": "mixed-4-8-bit",
   "quote": "Kolibri-1-mlx-mixed-4-8-bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,333 84.3 42.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/14qw86xi",
   "source_url": "https://omlx.ai/benchmarks/performance/14qw86xi",
   "tg_tokens": null,
   "tps": 84.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-14qw86xi-ctx4096",
   "model": "Kolibri-1",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 42.8 GB at this context; duplicate traces of the same session: 5y5uqzsk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2525.0,
   "provenance": "community",
   "quant": "mixed-4-8-bit",
   "quote": "Kolibri-1-mlx-mixed-4-8-bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,525 83.3 42.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/14qw86xi",
   "source_url": "https://omlx.ai/benchmarks/performance/14qw86xi",
   "tg_tokens": null,
   "tps": 83.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-14qw86xi-ctx8192",
   "model": "Kolibri-1",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 42.8 GB at this context; duplicate traces of the same session: 5y5uqzsk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2325.0,
   "provenance": "community",
   "quant": "mixed-4-8-bit",
   "quote": "Kolibri-1-mlx-mixed-4-8-bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,325 80.3 42.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/14qw86xi",
   "source_url": "https://omlx.ai/benchmarks/performance/14qw86xi",
   "tg_tokens": null,
   "tps": 80.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-14qw86xi-ctx16384",
   "model": "Kolibri-1",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 43.0 GB at this context; duplicate traces of the same session: 5y5uqzsk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2068.0,
   "provenance": "community",
   "quant": "mixed-4-8-bit",
   "quote": "Kolibri-1-mlx-mixed-4-8-bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,068 77.7 43.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/14qw86xi",
   "source_url": "https://omlx.ai/benchmarks/performance/14qw86xi",
   "tg_tokens": null,
   "tps": 77.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-14qw86xi-ctx32768",
   "model": "Kolibri-1",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 43.3 GB at this context; duplicate traces of the same session: 5y5uqzsk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1811.0,
   "provenance": "community",
   "quant": "mixed-4-8-bit",
   "quote": "Kolibri-1-mlx-mixed-4-8-bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,811 72.1 43.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/14qw86xi",
   "source_url": "https://omlx.ai/benchmarks/performance/14qw86xi",
   "tg_tokens": null,
   "tps": 72.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-q5kt44d5-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.8 GB at this context; duplicate traces of the same session: 9tahl0hz, ptfeqkij; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 470.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 16k 470.9 48.0 13.8 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q5kt44d5",
   "source_url": "https://omlx.ai/benchmarks/performance/q5kt44d5",
   "tg_tokens": null,
   "tps": 48.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-q5kt44d5-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 14.1 GB at this context; duplicate traces of the same session: 9tahl0hz, ptfeqkij; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 418.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 32k 418.0 39.8 14.1 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q5kt44d5",
   "source_url": "https://omlx.ai/benchmarks/performance/q5kt44d5",
   "tg_tokens": null,
   "tps": 39.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-q5kt44d5-ctx65536",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 14.7 GB at this context; duplicate traces of the same session: 9tahl0hz, ptfeqkij; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 308.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 64k 308.1 26.4 14.7 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q5kt44d5",
   "source_url": "https://omlx.ai/benchmarks/performance/q5kt44d5",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-h9ow117v-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: s7kjc3b2, th04ds3k; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2380.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 2,380 107.9 19.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h9ow117v",
   "source_url": "https://omlx.ai/benchmarks/performance/h9ow117v",
   "tg_tokens": null,
   "tps": 107.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-h9ow117v-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: s7kjc3b2, th04ds3k; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2928.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,928 104.0 20.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h9ow117v",
   "source_url": "https://omlx.ai/benchmarks/performance/h9ow117v",
   "tg_tokens": null,
   "tps": 104.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-h9ow117v-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: s7kjc3b2, th04ds3k; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2155.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 2,155 75.7 20.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h9ow117v",
   "source_url": "https://omlx.ai/benchmarks/performance/h9ow117v",
   "tg_tokens": null,
   "tps": 75.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-h9ow117v-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: s7kjc3b2, th04ds3k; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1584.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 64k 1,584 60.7 21.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h9ow117v",
   "source_url": "https://omlx.ai/benchmarks/performance/h9ow117v",
   "tg_tokens": null,
   "tps": 60.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-h9ow117v-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: s7kjc3b2, th04ds3k; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1083.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 128k 1,083 42.6 22.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h9ow117v",
   "source_url": "https://omlx.ai/benchmarks/performance/h9ow117v",
   "tg_tokens": null,
   "tps": 42.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-tmzedgbp-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "DWQ build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: hb5j1e3n; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 459.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit-DWQ on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 8bit; 32k 459.9 52.2 22.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tmzedgbp",
   "source_url": "https://omlx.ai/benchmarks/performance/tmzedgbp",
   "tg_tokens": null,
   "tps": 52.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-tmzedgbp-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "DWQ build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: hb5j1e3n; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 317.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit-DWQ on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 8bit; 64k 317.1 42.9 23.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tmzedgbp",
   "source_url": "https://omlx.ai/benchmarks/performance/tmzedgbp",
   "tg_tokens": null,
   "tps": 42.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ghz6y5jb-ctx1024",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: 0dfz16l1, 1sd02650, st4k38b0, y8u7ge2b; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2499.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 2,499 81.8 12.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ghz6y5jb",
   "source_url": "https://omlx.ai/benchmarks/performance/ghz6y5jb",
   "tg_tokens": null,
   "tps": 81.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ghz6y5jb-ctx4096",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.3 GB at this context; duplicate traces of the same session: 0dfz16l1, 1sd02650, st4k38b0, y8u7ge2b; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3012.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 3,012 79.2 13.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ghz6y5jb",
   "source_url": "https://omlx.ai/benchmarks/performance/ghz6y5jb",
   "tg_tokens": null,
   "tps": 79.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ghz6y5jb-ctx32768",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.9 GB at this context; duplicate traces of the same session: 0dfz16l1, 1sd02650, st4k38b0, y8u7ge2b; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2202.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 2,202 61.7 13.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ghz6y5jb",
   "source_url": "https://omlx.ai/benchmarks/performance/ghz6y5jb",
   "tg_tokens": null,
   "tps": 61.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ghz6y5jb-ctx65536",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 14.5 GB at this context; duplicate traces of the same session: 0dfz16l1, 1sd02650, st4k38b0, y8u7ge2b; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1610.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 64k 1,610 51.4 14.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ghz6y5jb",
   "source_url": "https://omlx.ai/benchmarks/performance/ghz6y5jb",
   "tg_tokens": null,
   "tps": 51.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ghz6y5jb-ctx131072",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: 0dfz16l1, 1sd02650, st4k38b0, y8u7ge2b; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1097.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 128k 1,097 37.7 15.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ghz6y5jb",
   "source_url": "https://omlx.ai/benchmarks/performance/ghz6y5jb",
   "tg_tokens": null,
   "tps": 37.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-3vx05n43-ctx1024",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: 4rq1xz1g, 81sj42p6, 8ksofo00; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2506.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 2,506 81.3 12.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3vx05n43",
   "source_url": "https://omlx.ai/benchmarks/performance/3vx05n43",
   "tg_tokens": null,
   "tps": 81.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-3vx05n43-ctx32768",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.9 GB at this context; duplicate traces of the same session: 4rq1xz1g, 81sj42p6, 8ksofo00; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2201.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 2,201 61.9 13.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3vx05n43",
   "source_url": "https://omlx.ai/benchmarks/performance/3vx05n43",
   "tg_tokens": null,
   "tps": 61.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-3vx05n43-ctx65536",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 14.5 GB at this context; duplicate traces of the same session: 4rq1xz1g, 81sj42p6, 8ksofo00; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1608.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 64k 1,608 51.4 14.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3vx05n43",
   "source_url": "https://omlx.ai/benchmarks/performance/3vx05n43",
   "tg_tokens": null,
   "tps": 51.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-3vx05n43-ctx131072",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: 4rq1xz1g, 81sj42p6, 8ksofo00; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1098.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 128k 1,098 37.7 15.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3vx05n43",
   "source_url": "https://omlx.ai/benchmarks/performance/3vx05n43",
   "tg_tokens": null,
   "tps": 37.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ycb34oa5-ctx1024",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: rl6fop32, so1dlzsx, szhjcae7, yanens8p; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 382.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 382.7 81.3 12.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ycb34oa5",
   "source_url": "https://omlx.ai/benchmarks/performance/ycb34oa5",
   "tg_tokens": null,
   "tps": 81.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ycb34oa5-ctx4096",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.3 GB at this context; duplicate traces of the same session: rl6fop32, so1dlzsx, szhjcae7, yanens8p; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1157.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 1,157 79.0 13.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ycb34oa5",
   "source_url": "https://omlx.ai/benchmarks/performance/ycb34oa5",
   "tg_tokens": null,
   "tps": 79.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ycb34oa5-ctx32768",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.9 GB at this context; duplicate traces of the same session: rl6fop32, so1dlzsx, szhjcae7, yanens8p; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2205.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 2,205 61.9 13.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ycb34oa5",
   "source_url": "https://omlx.ai/benchmarks/performance/ycb34oa5",
   "tg_tokens": null,
   "tps": 61.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ycb34oa5-ctx65536",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 14.5 GB at this context; duplicate traces of the same session: rl6fop32, so1dlzsx, szhjcae7, yanens8p; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1649.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 64k 1,649 51.3 14.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ycb34oa5",
   "source_url": "https://omlx.ai/benchmarks/performance/ycb34oa5",
   "tg_tokens": null,
   "tps": 51.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ycb34oa5-ctx131072",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: rl6fop32, so1dlzsx, szhjcae7, yanens8p; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1097.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 128k 1,097 37.7 15.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ycb34oa5",
   "source_url": "https://omlx.ai/benchmarks/performance/ycb34oa5",
   "tg_tokens": null,
   "tps": 37.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-swjgxlb1-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 457.4,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.6-35B-A3B-oQ3.5e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 3bit; 32k 457.4 54.0 19.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/swjgxlb1",
   "source_url": "https://omlx.ai/benchmarks/performance/swjgxlb1",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-swjgxlb1-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 315.8,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.6-35B-A3B-oQ3.5e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 3bit; 64k 315.8 44.1 20.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/swjgxlb1",
   "source_url": "https://omlx.ai/benchmarks/performance/swjgxlb1",
   "tg_tokens": null,
   "tps": 44.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-56xcg5p2-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: z4q6uy5u; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 53.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 53.6 9.8 16.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/56xcg5p2",
   "source_url": "https://omlx.ai/benchmarks/performance/56xcg5p2",
   "tg_tokens": null,
   "tps": 9.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-56xcg5p2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: z4q6uy5u; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 53.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 53.0 9.5 17.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/56xcg5p2",
   "source_url": "https://omlx.ai/benchmarks/performance/56xcg5p2",
   "tg_tokens": null,
   "tps": 9.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-f0uagspb-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: rllnmwfe; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 464.1,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 464.1 14.5 20.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f0uagspb",
   "source_url": "https://omlx.ai/benchmarks/performance/f0uagspb",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-f0uagspb-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: rllnmwfe; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 498.6,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 498.6 14.4 21.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f0uagspb",
   "source_url": "https://omlx.ai/benchmarks/performance/f0uagspb",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-xm2g0tka-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 8d8wernk; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 146.8,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 146.8 26.9 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xm2g0tka",
   "source_url": "https://omlx.ai/benchmarks/performance/xm2g0tka",
   "tg_tokens": null,
   "tps": 26.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-xm2g0tka-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 8d8wernk; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 156.4,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 156.4 24.5 23.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xm2g0tka",
   "source_url": "https://omlx.ai/benchmarks/performance/xm2g0tka",
   "tg_tokens": null,
   "tps": 24.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-qciogaf0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.2 GB at this context; duplicate traces of the same session: a1lmxwnt, nzkgow8f, sawwjwpu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 333.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 333.8 57.0 25.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qciogaf0",
   "source_url": "https://omlx.ai/benchmarks/performance/qciogaf0",
   "tg_tokens": null,
   "tps": 57.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-qciogaf0-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: a1lmxwnt, nzkgow8f, sawwjwpu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 279.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 279.5 61.0 26.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qciogaf0",
   "source_url": "https://omlx.ai/benchmarks/performance/qciogaf0",
   "tg_tokens": null,
   "tps": 61.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-qciogaf0-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.9 GB at this context; duplicate traces of the same session: a1lmxwnt, nzkgow8f, sawwjwpu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 217.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 217.7 34.4 30.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qciogaf0",
   "source_url": "https://omlx.ai/benchmarks/performance/qciogaf0",
   "tg_tokens": null,
   "tps": 34.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-qciogaf0-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.8 GB at this context; duplicate traces of the same session: a1lmxwnt, nzkgow8f, sawwjwpu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 133.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 133.7 21.0 39.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qciogaf0",
   "source_url": "https://omlx.ai/benchmarks/performance/qciogaf0",
   "tg_tokens": null,
   "tps": 21.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jvkf6j8i-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 1pkqzc2u, csb8xyey, iiabtuzz, ov9ereak, r0d394q9, t321pm9n, x3p4j5nh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 770.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 770.0 46.5 23.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jvkf6j8i",
   "source_url": "https://omlx.ai/benchmarks/performance/jvkf6j8i",
   "tg_tokens": null,
   "tps": 46.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jvkf6j8i-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: 1pkqzc2u, csb8xyey, iiabtuzz, ov9ereak, r0d394q9, t321pm9n, x3p4j5nh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 874.3,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 874.3 56.1 24.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jvkf6j8i",
   "source_url": "https://omlx.ai/benchmarks/performance/jvkf6j8i",
   "tg_tokens": null,
   "tps": 56.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jvkf6j8i-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: 1pkqzc2u, csb8xyey, iiabtuzz, ov9ereak, r0d394q9, t321pm9n, x3p4j5nh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 827.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 827.2 51.4 24.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jvkf6j8i",
   "source_url": "https://omlx.ai/benchmarks/performance/jvkf6j8i",
   "tg_tokens": null,
   "tps": 51.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jvkf6j8i-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: 1pkqzc2u, csb8xyey, iiabtuzz, ov9ereak, r0d394q9, t321pm9n, x3p4j5nh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 798.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 798.4 66.8 25.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jvkf6j8i",
   "source_url": "https://omlx.ai/benchmarks/performance/jvkf6j8i",
   "tg_tokens": null,
   "tps": 66.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jvkf6j8i-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: 1pkqzc2u, csb8xyey, iiabtuzz, ov9ereak, r0d394q9, t321pm9n, x3p4j5nh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 739.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 739.0 132.2 26.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jvkf6j8i",
   "source_url": "https://omlx.ai/benchmarks/performance/jvkf6j8i",
   "tg_tokens": null,
   "tps": 132.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jvkf6j8i-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.7 GB at this context; duplicate traces of the same session: 1pkqzc2u, csb8xyey, iiabtuzz, ov9ereak, r0d394q9, t321pm9n, x3p4j5nh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 586.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 586.7 45.8 28.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jvkf6j8i",
   "source_url": "https://omlx.ai/benchmarks/performance/jvkf6j8i",
   "tg_tokens": null,
   "tps": 45.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jvkf6j8i-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.4 GB at this context; duplicate traces of the same session: 1pkqzc2u, csb8xyey, iiabtuzz, ov9ereak, r0d394q9, t321pm9n, x3p4j5nh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 453.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 453.7 43.3 33.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jvkf6j8i",
   "source_url": "https://omlx.ai/benchmarks/performance/jvkf6j8i",
   "tg_tokens": null,
   "tps": 43.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jvkf6j8i-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.6 GB at this context; duplicate traces of the same session: 1pkqzc2u, csb8xyey, iiabtuzz, ov9ereak, r0d394q9, t321pm9n, x3p4j5nh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 370.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 195k 370.1 27.1 38.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jvkf6j8i",
   "source_url": "https://omlx.ai/benchmarks/performance/jvkf6j8i",
   "tg_tokens": null,
   "tps": 27.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0fuqcg70-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: tv9kwox3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 847.5,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 847.5 67.0 23.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0fuqcg70",
   "source_url": "https://omlx.ai/benchmarks/performance/0fuqcg70",
   "tg_tokens": null,
   "tps": 67.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0fuqcg70-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: tv9kwox3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 907.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 907.4 55.2 24.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0fuqcg70",
   "source_url": "https://omlx.ai/benchmarks/performance/0fuqcg70",
   "tg_tokens": null,
   "tps": 55.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0fuqcg70-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: tv9kwox3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 844.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 844.4 70.9 24.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0fuqcg70",
   "source_url": "https://omlx.ai/benchmarks/performance/0fuqcg70",
   "tg_tokens": null,
   "tps": 70.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0fuqcg70-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: tv9kwox3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 751.6,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 751.6 76.5 25.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0fuqcg70",
   "source_url": "https://omlx.ai/benchmarks/performance/0fuqcg70",
   "tg_tokens": null,
   "tps": 76.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0fuqcg70-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: tv9kwox3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 667.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 667.2 130.4 26.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0fuqcg70",
   "source_url": "https://omlx.ai/benchmarks/performance/0fuqcg70",
   "tg_tokens": null,
   "tps": 130.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0fuqcg70-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.7 GB at this context; duplicate traces of the same session: tv9kwox3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 545.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 545.7 42.4 28.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0fuqcg70",
   "source_url": "https://omlx.ai/benchmarks/performance/0fuqcg70",
   "tg_tokens": null,
   "tps": 42.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0fuqcg70-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.4 GB at this context; duplicate traces of the same session: tv9kwox3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 444.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 444.7 44.5 33.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0fuqcg70",
   "source_url": "https://omlx.ai/benchmarks/performance/0fuqcg70",
   "tg_tokens": null,
   "tps": 44.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-bmj0klze-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.0 GB at this context; duplicate traces of the same session: jxmcpb4n; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 317.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 317.4 43.3 39.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bmj0klze",
   "source_url": "https://omlx.ai/benchmarks/performance/bmj0klze",
   "tg_tokens": null,
   "tps": 43.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-bmj0klze-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 42.3 GB at this context; duplicate traces of the same session: jxmcpb4n; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 411.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 411.4 47.0 42.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bmj0klze",
   "source_url": "https://omlx.ai/benchmarks/performance/bmj0klze",
   "tg_tokens": null,
   "tps": 47.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-fkjjkv15-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.0 GB at this context; duplicate traces of the same session: zdzpi8tw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 317.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 317.9 43.9 39.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fkjjkv15",
   "source_url": "https://omlx.ai/benchmarks/performance/fkjjkv15",
   "tg_tokens": null,
   "tps": 43.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-fkjjkv15-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 42.3 GB at this context; duplicate traces of the same session: zdzpi8tw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 410.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 410.9 41.8 42.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fkjjkv15",
   "source_url": "https://omlx.ai/benchmarks/performance/fkjjkv15",
   "tg_tokens": null,
   "tps": 41.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-i5bvfzlp-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: m40fxsfb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 316.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 316.5 47.3 33.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i5bvfzlp",
   "source_url": "https://omlx.ai/benchmarks/performance/i5bvfzlp",
   "tg_tokens": null,
   "tps": 47.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-i5bvfzlp-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: m40fxsfb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 331.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 331.7 43.9 37.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i5bvfzlp",
   "source_url": "https://omlx.ai/benchmarks/performance/i5bvfzlp",
   "tg_tokens": null,
   "tps": 43.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-z50sl5ae-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 71.1 GB; TTFT 838 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1222.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,222 PP tok/s 66.5 TG tok/s 838 TTFT (ms) 71.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z50sl5ae",
   "source_url": "https://omlx.ai/benchmarks/performance/z50sl5ae",
   "tg_tokens": null,
   "tps": 66.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3bp0d2uk-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: jw0nabve; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1022.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,022 61.5 71.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3bp0d2uk",
   "source_url": "https://omlx.ai/benchmarks/performance/3bp0d2uk",
   "tg_tokens": null,
   "tps": 61.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3bp0d2uk-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.2 GB at this context; duplicate traces of the same session: jw0nabve; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1324.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,324 40.7 71.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3bp0d2uk",
   "source_url": "https://omlx.ai/benchmarks/performance/3bp0d2uk",
   "tg_tokens": null,
   "tps": 40.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-mru8zyqt-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 86.5 GB at this context; duplicate traces of the same session: 7jyyyr46; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 491.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 491.0 43.4 86.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mru8zyqt",
   "source_url": "https://omlx.ai/benchmarks/performance/mru8zyqt",
   "tg_tokens": null,
   "tps": 43.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-mru8zyqt-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 88.3 GB at this context; duplicate traces of the same session: 7jyyyr46; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 521.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 521.4 58.4 88.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mru8zyqt",
   "source_url": "https://omlx.ai/benchmarks/performance/mru8zyqt",
   "tg_tokens": null,
   "tps": 58.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-96y7drqj-ctx32768",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 12.4 GB at this context; duplicate traces of the same session: mzhy66vj; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 489.8,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 32k 489.8 40.5 12.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/96y7drqj",
   "source_url": "https://omlx.ai/benchmarks/performance/96y7drqj",
   "tg_tokens": null,
   "tps": 40.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-96y7drqj-ctx65536",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 13.2 GB at this context; duplicate traces of the same session: mzhy66vj; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 399.2,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 64k 399.2 28.2 13.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/96y7drqj",
   "source_url": "https://omlx.ai/benchmarks/performance/96y7drqj",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-b7r6u0kz-ctx1024",
   "model": "Occamy-1.0",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: bv51qqyd; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 846.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "occamy-1.0-oQ4e-fp16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 846.4 72.0 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b7r6u0kz",
   "source_url": "https://omlx.ai/benchmarks/performance/b7r6u0kz",
   "tg_tokens": null,
   "tps": 72.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-b7r6u0kz-ctx4096",
   "model": "Occamy-1.0",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: bv51qqyd; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1146.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "occamy-1.0-oQ4e-fp16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 1,146 67.8 21.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b7r6u0kz",
   "source_url": "https://omlx.ai/benchmarks/performance/b7r6u0kz",
   "tg_tokens": null,
   "tps": 67.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-u7art0ae-ctx1024",
   "model": "Kolibri-1",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 42.7 GB at this context; duplicate traces of the same session: 7tvwx7np, ayh22f13; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1333.0,
   "provenance": "community",
   "quant": "mixed-4-8-bit",
   "quote": "Kolibri-1-mlx-mixed-4-8-bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,333 84.3 42.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u7art0ae",
   "source_url": "https://omlx.ai/benchmarks/performance/u7art0ae",
   "tg_tokens": null,
   "tps": 84.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-u7art0ae-ctx4096",
   "model": "Kolibri-1",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 42.8 GB at this context; duplicate traces of the same session: 7tvwx7np, ayh22f13; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2525.0,
   "provenance": "community",
   "quant": "mixed-4-8-bit",
   "quote": "Kolibri-1-mlx-mixed-4-8-bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,525 83.3 42.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u7art0ae",
   "source_url": "https://omlx.ai/benchmarks/performance/u7art0ae",
   "tg_tokens": null,
   "tps": 83.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-u7art0ae-ctx8192",
   "model": "Kolibri-1",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 42.8 GB at this context; duplicate traces of the same session: 7tvwx7np, ayh22f13; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2325.0,
   "provenance": "community",
   "quant": "mixed-4-8-bit",
   "quote": "Kolibri-1-mlx-mixed-4-8-bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,325 80.3 42.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u7art0ae",
   "source_url": "https://omlx.ai/benchmarks/performance/u7art0ae",
   "tg_tokens": null,
   "tps": 80.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-u7art0ae-ctx16384",
   "model": "Kolibri-1",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 43.0 GB at this context; duplicate traces of the same session: 7tvwx7np, ayh22f13; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2068.0,
   "provenance": "community",
   "quant": "mixed-4-8-bit",
   "quote": "Kolibri-1-mlx-mixed-4-8-bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,068 77.7 43.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u7art0ae",
   "source_url": "https://omlx.ai/benchmarks/performance/u7art0ae",
   "tg_tokens": null,
   "tps": 77.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-u7art0ae-ctx32768",
   "model": "Kolibri-1",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 43.3 GB at this context; duplicate traces of the same session: 7tvwx7np, ayh22f13; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1811.0,
   "provenance": "community",
   "quant": "mixed-4-8-bit",
   "quote": "Kolibri-1-mlx-mixed-4-8-bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,811 72.1 43.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u7art0ae",
   "source_url": "https://omlx.ai/benchmarks/performance/u7art0ae",
   "tg_tokens": null,
   "tps": 72.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-cnnw7by9-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: i7al7kux, ibgit0nv, kbxfcrmp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 960.6,
   "provenance": "community",
   "quant": "5bit",
   "quote": "LFM2.5-24B-A2B-MLX-5bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 5bit; 1k 960.6 98.7 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cnnw7by9",
   "source_url": "https://omlx.ai/benchmarks/performance/cnnw7by9",
   "tg_tokens": null,
   "tps": 98.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-cnnw7by9-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: i7al7kux, ibgit0nv, kbxfcrmp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1023.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "LFM2.5-24B-A2B-MLX-5bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 5bit; 4k 1,023 95.1 16.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cnnw7by9",
   "source_url": "https://omlx.ai/benchmarks/performance/cnnw7by9",
   "tg_tokens": null,
   "tps": 95.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-cnnw7by9-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: i7al7kux, ibgit0nv, kbxfcrmp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 995.7,
   "provenance": "community",
   "quant": "5bit",
   "quote": "LFM2.5-24B-A2B-MLX-5bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 5bit; 8k 995.7 91.5 16.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cnnw7by9",
   "source_url": "https://omlx.ai/benchmarks/performance/cnnw7by9",
   "tg_tokens": null,
   "tps": 91.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-cnnw7by9-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: i7al7kux, ibgit0nv, kbxfcrmp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 939.4,
   "provenance": "community",
   "quant": "5bit",
   "quote": "LFM2.5-24B-A2B-MLX-5bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 5bit; 16k 939.4 85.1 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cnnw7by9",
   "source_url": "https://omlx.ai/benchmarks/performance/cnnw7by9",
   "tg_tokens": null,
   "tps": 85.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-a9akquu6-ctx1024",
   "model": "Muse Glimmer",
   "notes": "oMLX community benchmark trace; peak memory 28.2 GB at this context; duplicate traces of the same session: 91fyi98b; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 355.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Muse-Glimmer-30B-8bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 8bit; 1k 355.5 10.0 28.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a9akquu6",
   "source_url": "https://omlx.ai/benchmarks/performance/a9akquu6",
   "tg_tokens": null,
   "tps": 10.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-a9akquu6-ctx4096",
   "model": "Muse Glimmer",
   "notes": "oMLX community benchmark trace; peak memory 28.4 GB at this context; duplicate traces of the same session: 91fyi98b; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 370.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Muse-Glimmer-30B-8bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 8bit; 4k 370.8 9.8 28.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a9akquu6",
   "source_url": "https://omlx.ai/benchmarks/performance/a9akquu6",
   "tg_tokens": null,
   "tps": 9.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-mvey8sw5-ctx1024",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: e6qqimo1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 192.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 192.6 49.3 21.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mvey8sw5",
   "source_url": "https://omlx.ai/benchmarks/performance/mvey8sw5",
   "tg_tokens": null,
   "tps": 49.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-mvey8sw5-ctx8192",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: e6qqimo1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 236.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 236.5 35.3 22.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mvey8sw5",
   "source_url": "https://omlx.ai/benchmarks/performance/mvey8sw5",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-mvey8sw5-ctx65536",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: e6qqimo1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 215.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 215.9 22.0 23.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mvey8sw5",
   "source_url": "https://omlx.ai/benchmarks/performance/mvey8sw5",
   "tg_tokens": null,
   "tps": 22.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-pujhyc7y-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.8 GB at this context; duplicate traces of the same session: r9hu5j65; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 222.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 6bit; 1k 222.4 14.9 7.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pujhyc7y",
   "source_url": "https://omlx.ai/benchmarks/performance/pujhyc7y",
   "tg_tokens": null,
   "tps": 14.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-pujhyc7y-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.2 GB at this context; duplicate traces of the same session: r9hu5j65; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 232.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 6bit; 4k 232.7 14.5 8.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pujhyc7y",
   "source_url": "https://omlx.ai/benchmarks/performance/pujhyc7y",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-uj7dm230-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP MLX build; oMLX community benchmark trace; peak memory 17.9 GB at this context; duplicate traces of the same session: 3xd8vm89; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 454.4,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.6-35B-A3B-oQ3e-MTP-MLX on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 3bit; 32k 454.4 53.9 17.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uj7dm230",
   "source_url": "https://omlx.ai/benchmarks/performance/uj7dm230",
   "tg_tokens": null,
   "tps": 53.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-uj7dm230-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP MLX build; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 3xd8vm89; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 314.2,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.6-35B-A3B-oQ3e-MTP-MLX on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 3bit; 64k 314.2 43.9 19.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uj7dm230",
   "source_url": "https://omlx.ai/benchmarks/performance/uj7dm230",
   "tg_tokens": null,
   "tps": 43.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-uzkmjzv7-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 43oevhwk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 157.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 157.2 18.4 23.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uzkmjzv7",
   "source_url": "https://omlx.ai/benchmarks/performance/uzkmjzv7",
   "tg_tokens": null,
   "tps": 18.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-uzkmjzv7-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 27.0 GB at this context; duplicate traces of the same session: 43oevhwk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 169.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 169.8 22.7 27.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uzkmjzv7",
   "source_url": "https://omlx.ai/benchmarks/performance/uzkmjzv7",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-cjf7thez-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: zyer39gt; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 128.8,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 128.8 35.1 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cjf7thez",
   "source_url": "https://omlx.ai/benchmarks/performance/cjf7thez",
   "tg_tokens": null,
   "tps": 35.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-cjf7thez-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: zyer39gt; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.9,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 127.9 29.1 23.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cjf7thez",
   "source_url": "https://omlx.ai/benchmarks/performance/cjf7thez",
   "tg_tokens": null,
   "tps": 29.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dk01b10s-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: t1viz09c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 883.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 883.3 59.9 17.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dk01b10s",
   "source_url": "https://omlx.ai/benchmarks/performance/dk01b10s",
   "tg_tokens": null,
   "tps": 59.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dk01b10s-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: t1viz09c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 850.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 850.8 62.9 18.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dk01b10s",
   "source_url": "https://omlx.ai/benchmarks/performance/dk01b10s",
   "tg_tokens": null,
   "tps": 62.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yrb3lhxa-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: sc7t9mdf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 916.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 916.8 59.2 17.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yrb3lhxa",
   "source_url": "https://omlx.ai/benchmarks/performance/yrb3lhxa",
   "tg_tokens": null,
   "tps": 59.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yrb3lhxa-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: sc7t9mdf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 939.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 939.1 64.6 18.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yrb3lhxa",
   "source_url": "https://omlx.ai/benchmarks/performance/yrb3lhxa",
   "tg_tokens": null,
   "tps": 64.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dyrjjlx7-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: iqi0gm1s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 909.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 909.1 66.8 17.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dyrjjlx7",
   "source_url": "https://omlx.ai/benchmarks/performance/dyrjjlx7",
   "tg_tokens": null,
   "tps": 66.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dyrjjlx7-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: iqi0gm1s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 897.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 897.8 68.9 18.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dyrjjlx7",
   "source_url": "https://omlx.ai/benchmarks/performance/dyrjjlx7",
   "tg_tokens": null,
   "tps": 68.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-zbfv8q9k-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 24zgpzvq, 6lu3hto3, ozvxq7z5; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1290.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k 1,290 64.5 20.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zbfv8q9k",
   "source_url": "https://omlx.ai/benchmarks/performance/zbfv8q9k",
   "tg_tokens": null,
   "tps": 64.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-zbfv8q9k-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 24zgpzvq, 6lu3hto3, ozvxq7z5; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1489.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 4k 1,489 71.3 21.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zbfv8q9k",
   "source_url": "https://omlx.ai/benchmarks/performance/zbfv8q9k",
   "tg_tokens": null,
   "tps": 71.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-zbfv8q9k-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 24zgpzvq, 6lu3hto3, ozvxq7z5; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1494.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 8k 1,494 63.5 21.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zbfv8q9k",
   "source_url": "https://omlx.ai/benchmarks/performance/zbfv8q9k",
   "tg_tokens": null,
   "tps": 63.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-zbfv8q9k-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 24zgpzvq, 6lu3hto3, ozvxq7z5; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1451.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 16k 1,451 75.4 22.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zbfv8q9k",
   "source_url": "https://omlx.ai/benchmarks/performance/zbfv8q9k",
   "tg_tokens": null,
   "tps": 75.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-y8ol84t3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: hj61ysd2, tt3njqe1, zwbtuhmu; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1128.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k 1,128 70.9 17.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y8ol84t3",
   "source_url": "https://omlx.ai/benchmarks/performance/y8ol84t3",
   "tg_tokens": null,
   "tps": 70.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-y8ol84t3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: hj61ysd2, tt3njqe1, zwbtuhmu; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1524.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 4k 1,524 98.8 18.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y8ol84t3",
   "source_url": "https://omlx.ai/benchmarks/performance/y8ol84t3",
   "tg_tokens": null,
   "tps": 98.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-y8ol84t3-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: hj61ysd2, tt3njqe1, zwbtuhmu; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1533.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 8k 1,533 71.7 19.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y8ol84t3",
   "source_url": "https://omlx.ai/benchmarks/performance/y8ol84t3",
   "tg_tokens": null,
   "tps": 71.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-y8ol84t3-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: hj61ysd2, tt3njqe1, zwbtuhmu; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1492.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 16k 1,492 89.8 19.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y8ol84t3",
   "source_url": "https://omlx.ai/benchmarks/performance/y8ol84t3",
   "tg_tokens": null,
   "tps": 89.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-909zunu6-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 4hvc3wuq; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1196.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 4k 1,196 42.1 21.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/909zunu6",
   "source_url": "https://omlx.ai/benchmarks/performance/909zunu6",
   "tg_tokens": null,
   "tps": 42.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-909zunu6-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 4hvc3wuq; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1490.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 8k 1,490 93.8 21.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/909zunu6",
   "source_url": "https://omlx.ai/benchmarks/performance/909zunu6",
   "tg_tokens": null,
   "tps": 93.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-909zunu6-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 4hvc3wuq; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1451.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 16k 1,451 97.3 22.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/909zunu6",
   "source_url": "https://omlx.ai/benchmarks/performance/909zunu6",
   "tg_tokens": null,
   "tps": 97.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2g7l0y56-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 240kj8io, qbub9zgt, r4yc1qlc, wztvbpol; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 847.5,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 847.5 67.0 23.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2g7l0y56",
   "source_url": "https://omlx.ai/benchmarks/performance/2g7l0y56",
   "tg_tokens": null,
   "tps": 67.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2g7l0y56-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: 240kj8io, qbub9zgt, r4yc1qlc, wztvbpol; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 907.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 907.4 55.2 24.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2g7l0y56",
   "source_url": "https://omlx.ai/benchmarks/performance/2g7l0y56",
   "tg_tokens": null,
   "tps": 55.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2g7l0y56-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: 240kj8io, qbub9zgt, r4yc1qlc, wztvbpol; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 844.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 844.4 70.9 24.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2g7l0y56",
   "source_url": "https://omlx.ai/benchmarks/performance/2g7l0y56",
   "tg_tokens": null,
   "tps": 70.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2g7l0y56-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: 240kj8io, qbub9zgt, r4yc1qlc, wztvbpol; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 751.6,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 751.6 76.5 25.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2g7l0y56",
   "source_url": "https://omlx.ai/benchmarks/performance/2g7l0y56",
   "tg_tokens": null,
   "tps": 76.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2g7l0y56-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: 240kj8io, qbub9zgt, r4yc1qlc, wztvbpol; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 667.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 667.2 130.4 26.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2g7l0y56",
   "source_url": "https://omlx.ai/benchmarks/performance/2g7l0y56",
   "tg_tokens": null,
   "tps": 130.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2g7l0y56-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.7 GB at this context; duplicate traces of the same session: 240kj8io, qbub9zgt, r4yc1qlc, wztvbpol; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 545.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 545.7 42.4 28.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2g7l0y56",
   "source_url": "https://omlx.ai/benchmarks/performance/2g7l0y56",
   "tg_tokens": null,
   "tps": 42.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2g7l0y56-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.4 GB at this context; duplicate traces of the same session: 240kj8io, qbub9zgt, r4yc1qlc, wztvbpol; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 444.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 444.7 44.5 33.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2g7l0y56",
   "source_url": "https://omlx.ai/benchmarks/performance/2g7l0y56",
   "tg_tokens": null,
   "tps": 44.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-t0ujsy8w-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: 2dewehd3, 8q5ela2r, bccvrvwh, bgyq0w6n, p4rkh20r, ref9flgu, wlcmtbln; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 682.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 682.8 37.2 33.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t0ujsy8w",
   "source_url": "https://omlx.ai/benchmarks/performance/t0ujsy8w",
   "tg_tokens": null,
   "tps": 37.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-t0ujsy8w-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: 2dewehd3, 8q5ela2r, bccvrvwh, bgyq0w6n, p4rkh20r, ref9flgu, wlcmtbln; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 876.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 876.2 41.4 34.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t0ujsy8w",
   "source_url": "https://omlx.ai/benchmarks/performance/t0ujsy8w",
   "tg_tokens": null,
   "tps": 41.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-t0ujsy8w-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: 2dewehd3, 8q5ela2r, bccvrvwh, bgyq0w6n, p4rkh20r, ref9flgu, wlcmtbln; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 814.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 814.8 51.4 35.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t0ujsy8w",
   "source_url": "https://omlx.ai/benchmarks/performance/t0ujsy8w",
   "tg_tokens": null,
   "tps": 51.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-t0ujsy8w-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: 2dewehd3, 8q5ela2r, bccvrvwh, bgyq0w6n, p4rkh20r, ref9flgu, wlcmtbln; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 766.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 766.1 57.8 35.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t0ujsy8w",
   "source_url": "https://omlx.ai/benchmarks/performance/t0ujsy8w",
   "tg_tokens": null,
   "tps": 57.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-t0ujsy8w-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: 2dewehd3, 8q5ela2r, bccvrvwh, bgyq0w6n, p4rkh20r, ref9flgu, wlcmtbln; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 720.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 720.1 83.6 36.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t0ujsy8w",
   "source_url": "https://omlx.ai/benchmarks/performance/t0ujsy8w",
   "tg_tokens": null,
   "tps": 83.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-t0ujsy8w-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: 2dewehd3, 8q5ela2r, bccvrvwh, bgyq0w6n, p4rkh20r, ref9flgu, wlcmtbln; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 613.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 613.9 41.8 39.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t0ujsy8w",
   "source_url": "https://omlx.ai/benchmarks/performance/t0ujsy8w",
   "tg_tokens": null,
   "tps": 41.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-t0ujsy8w-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 44.0 GB at this context; duplicate traces of the same session: 2dewehd3, 8q5ela2r, bccvrvwh, bgyq0w6n, p4rkh20r, ref9flgu, wlcmtbln; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 466.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 466.2 35.6 44.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t0ujsy8w",
   "source_url": "https://omlx.ai/benchmarks/performance/t0ujsy8w",
   "tg_tokens": null,
   "tps": 35.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-t0ujsy8w-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 49.2 GB at this context; duplicate traces of the same session: 2dewehd3, 8q5ela2r, bccvrvwh, bgyq0w6n, p4rkh20r, ref9flgu, wlcmtbln; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 364.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 195k 364.7 24.7 49.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t0ujsy8w",
   "source_url": "https://omlx.ai/benchmarks/performance/t0ujsy8w",
   "tg_tokens": null,
   "tps": 24.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-16zix35m-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: id98z2dl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 240.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 240.5 28.3 70.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/16zix35m",
   "source_url": "https://omlx.ai/benchmarks/performance/16zix35m",
   "tg_tokens": null,
   "tps": 28.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-16zix35m-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.4 GB at this context; duplicate traces of the same session: id98z2dl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 290.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 290.9 36.9 71.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/16zix35m",
   "source_url": "https://omlx.ai/benchmarks/performance/16zix35m",
   "tg_tokens": null,
   "tps": 36.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-03nnpceu-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 71.0 GB; TTFT 842 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1217.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,217 PP tok/s 70.1 TG tok/s 842 TTFT (ms) 71.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/03nnpceu",
   "source_url": "https://omlx.ai/benchmarks/performance/03nnpceu",
   "tg_tokens": null,
   "tps": 70.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-m8zsdjmg-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: 3cfre0sj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1249.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,249 82.5 71.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m8zsdjmg",
   "source_url": "https://omlx.ai/benchmarks/performance/m8zsdjmg",
   "tg_tokens": null,
   "tps": 82.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-m8zsdjmg-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.2 GB at this context; duplicate traces of the same session: 3cfre0sj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2002.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,002 46.4 71.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m8zsdjmg",
   "source_url": "https://omlx.ai/benchmarks/performance/m8zsdjmg",
   "tg_tokens": null,
   "tps": 46.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8zfklpjk-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: 1vikcqwd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1213.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,213 76.7 71.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8zfklpjk",
   "source_url": "https://omlx.ai/benchmarks/performance/8zfklpjk",
   "tg_tokens": null,
   "tps": 76.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8zfklpjk-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.2 GB at this context; duplicate traces of the same session: 1vikcqwd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1953.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,953 35.5 71.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8zfklpjk",
   "source_url": "https://omlx.ai/benchmarks/performance/8zfklpjk",
   "tg_tokens": null,
   "tps": 35.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-imulaki8-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 20.0 GB; TTFT 24707 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 663.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k tokens 663.1 PP tok/s 27.2 TG tok/s 24707 TTFT (ms) 20.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/imulaki8",
   "source_url": "https://omlx.ai/benchmarks/performance/imulaki8",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-qf46jq8v-ctx1024",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 7.7 GB at this context; duplicate traces of the same session: l2z78ime; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 143.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12b-it-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 143.5 24.3 7.7 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qf46jq8v",
   "source_url": "https://omlx.ai/benchmarks/performance/qf46jq8v",
   "tg_tokens": null,
   "tps": 24.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-qf46jq8v-ctx4096",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 7.7 GB at this context; duplicate traces of the same session: l2z78ime; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 139.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12b-it-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 139.0 24.6 7.7 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qf46jq8v",
   "source_url": "https://omlx.ai/benchmarks/performance/qf46jq8v",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-l6gorrk5-ctx1024",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: bw60n6ae; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 107.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31B-it-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 107.0 12.8 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l6gorrk5",
   "source_url": "https://omlx.ai/benchmarks/performance/l6gorrk5",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-l6gorrk5-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: bw60n6ae; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31B-it-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 103.2 12.6 18.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l6gorrk5",
   "source_url": "https://omlx.ai/benchmarks/performance/l6gorrk5",
   "tg_tokens": null,
   "tps": 12.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-989kghd2-ctx1024",
   "model": "muse-glimmer 30B",
   "notes": "OptiQ; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 395.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-30B-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 395.1 14.8 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/989kghd2",
   "source_url": "https://omlx.ai/benchmarks/performance/989kghd2",
   "tg_tokens": null,
   "tps": 14.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-989kghd2-ctx4096",
   "model": "muse-glimmer 30B",
   "notes": "OptiQ; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 409.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-30B-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 409.3 14.6 19.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/989kghd2",
   "source_url": "https://omlx.ai/benchmarks/performance/989kghd2",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-dxwtcvo8-ctx1024",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 192.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 192.6 49.3 21.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dxwtcvo8",
   "source_url": "https://omlx.ai/benchmarks/performance/dxwtcvo8",
   "tg_tokens": null,
   "tps": 49.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-dxwtcvo8-ctx8192",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 236.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 236.5 35.3 22.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dxwtcvo8",
   "source_url": "https://omlx.ai/benchmarks/performance/dxwtcvo8",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-dxwtcvo8-ctx65536",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 215.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 215.9 22.0 23.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dxwtcvo8",
   "source_url": "https://omlx.ai/benchmarks/performance/dxwtcvo8",
   "tg_tokens": null,
   "tps": 22.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-qyi5nech-ctx1024",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 17.9 GB at this context; duplicate traces of the same session: 0q07ae2q, 1hdwlus9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 196.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 196.8 55.3 17.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qyi5nech",
   "source_url": "https://omlx.ai/benchmarks/performance/qyi5nech",
   "tg_tokens": null,
   "tps": 55.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-qyi5nech-ctx8192",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: 0q07ae2q, 1hdwlus9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 237.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 237.1 37.6 18.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qyi5nech",
   "source_url": "https://omlx.ai/benchmarks/performance/qyi5nech",
   "tg_tokens": null,
   "tps": 37.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-qyi5nech-ctx65536",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 0q07ae2q, 1hdwlus9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 216.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 216.3 20.0 20.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qyi5nech",
   "source_url": "https://omlx.ai/benchmarks/performance/qyi5nech",
   "tg_tokens": null,
   "tps": 20.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-6l5e4zkr-ctx1024",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: brnrnma7, qzeihszi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 259.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 259.1 28.4 15.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6l5e4zkr",
   "source_url": "https://omlx.ai/benchmarks/performance/6l5e4zkr",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-6l5e4zkr-ctx8192",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 16.2 GB at this context; duplicate traces of the same session: brnrnma7, qzeihszi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 252.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 252.8 27.6 16.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6l5e4zkr",
   "source_url": "https://omlx.ai/benchmarks/performance/6l5e4zkr",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-6l5e4zkr-ctx65536",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: brnrnma7, qzeihszi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 227.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 227.3 24.5 16.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6l5e4zkr",
   "source_url": "https://omlx.ai/benchmarks/performance/6l5e4zkr",
   "tg_tokens": null,
   "tps": 24.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ibqo2k05-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: atcv4a9a, c86fpf23, kt0ib653; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2370.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 2,370 107.7 19.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ibqo2k05",
   "source_url": "https://omlx.ai/benchmarks/performance/ibqo2k05",
   "tg_tokens": null,
   "tps": 107.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ibqo2k05-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: atcv4a9a, c86fpf23, kt0ib653; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2149.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 2,149 75.8 20.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ibqo2k05",
   "source_url": "https://omlx.ai/benchmarks/performance/ibqo2k05",
   "tg_tokens": null,
   "tps": 75.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ibqo2k05-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: atcv4a9a, c86fpf23, kt0ib653; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1585.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 64k 1,585 60.9 21.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ibqo2k05",
   "source_url": "https://omlx.ai/benchmarks/performance/ibqo2k05",
   "tg_tokens": null,
   "tps": 60.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ibqo2k05-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: atcv4a9a, c86fpf23, kt0ib653; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1081.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 128k 1,081 42.3 22.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ibqo2k05",
   "source_url": "https://omlx.ai/benchmarks/performance/ibqo2k05",
   "tg_tokens": null,
   "tps": 42.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-4lgfylrs-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 25.6 GB at this context; duplicate traces of the same session: q8njz1ii; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 270.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 270.5 20.7 25.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4lgfylrs",
   "source_url": "https://omlx.ai/benchmarks/performance/4lgfylrs",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-4lgfylrs-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: q8njz1ii; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 256.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k 256.8 19.8 26.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4lgfylrs",
   "source_url": "https://omlx.ai/benchmarks/performance/4lgfylrs",
   "tg_tokens": null,
   "tps": 19.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-i3ho0dj0-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: iy1gq3cs; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 350.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 350.2 45.5 23.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i3ho0dj0",
   "source_url": "https://omlx.ai/benchmarks/performance/i3ho0dj0",
   "tg_tokens": null,
   "tps": 45.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-i3ho0dj0-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: iy1gq3cs; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 336.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 16k 336.5 39.9 24.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i3ho0dj0",
   "source_url": "https://omlx.ai/benchmarks/performance/i3ho0dj0",
   "tg_tokens": null,
   "tps": 39.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-v72n2282-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: hkye2bo0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.4,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 323.4 63.1 23.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v72n2282",
   "source_url": "https://omlx.ai/benchmarks/performance/v72n2282",
   "tg_tokens": null,
   "tps": 63.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-v72n2282-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: hkye2bo0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 303.1,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 16k 303.1 45.3 25.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v72n2282",
   "source_url": "https://omlx.ai/benchmarks/performance/v72n2282",
   "tg_tokens": null,
   "tps": 45.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-h5dpshpl-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: n8yxyg32; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 321.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 321.0 35.6 22.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h5dpshpl",
   "source_url": "https://omlx.ai/benchmarks/performance/h5dpshpl",
   "tg_tokens": null,
   "tps": 35.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-h5dpshpl-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: n8yxyg32; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 303.3,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 16k 303.3 43.8 25.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h5dpshpl",
   "source_url": "https://omlx.ai/benchmarks/performance/h5dpshpl",
   "tg_tokens": null,
   "tps": 43.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ppgqjvc5-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: iy57ljm1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 862.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 862.7 67.4 17.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ppgqjvc5",
   "source_url": "https://omlx.ai/benchmarks/performance/ppgqjvc5",
   "tg_tokens": null,
   "tps": 67.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ppgqjvc5-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: iy57ljm1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 846.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 846.2 66.8 18.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ppgqjvc5",
   "source_url": "https://omlx.ai/benchmarks/performance/ppgqjvc5",
   "tg_tokens": null,
   "tps": 66.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-6hylt2d9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1196.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 4k 1,196 42.1 21.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6hylt2d9",
   "source_url": "https://omlx.ai/benchmarks/performance/6hylt2d9",
   "tg_tokens": null,
   "tps": 42.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-6hylt2d9-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1490.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 8k 1,490 93.8 21.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6hylt2d9",
   "source_url": "https://omlx.ai/benchmarks/performance/6hylt2d9",
   "tg_tokens": null,
   "tps": 93.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-6hylt2d9-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1451.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 16k 1,451 97.3 22.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6hylt2d9",
   "source_url": "https://omlx.ai/benchmarks/performance/6hylt2d9",
   "tg_tokens": null,
   "tps": 97.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ytoe0uxc-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 195k context; peak memory 51.7 GB; TTFT 1164396 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 171.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 195k tokens 171.8 PP tok/s 11.9 TG tok/s 1164396 TTFT (ms) 51.7 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ytoe0uxc",
   "source_url": "https://omlx.ai/benchmarks/performance/ytoe0uxc",
   "tg_tokens": null,
   "tps": 11.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-fp3ts1br-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 517.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 517.2 60.3 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fp3ts1br",
   "source_url": "https://omlx.ai/benchmarks/performance/fp3ts1br",
   "tg_tokens": null,
   "tps": 60.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-fp3ts1br-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 654.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 654.8 55.9 71.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fp3ts1br",
   "source_url": "https://omlx.ai/benchmarks/performance/fp3ts1br",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-fp3ts1br-ctx16384",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 705.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 705.6 57.3 72.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fp3ts1br",
   "source_url": "https://omlx.ai/benchmarks/performance/fp3ts1br",
   "tg_tokens": null,
   "tps": 57.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-algade1z-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace, single run at 4k context; peak memory 34.0 GB; TTFT 25413 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 161.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 161.2 PP tok/s 7.6 TG tok/s 25413 TTFT (ms) 34.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/algade1z",
   "source_url": "https://omlx.ai/benchmarks/performance/algade1z",
   "tg_tokens": null,
   "tps": 7.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-o332kj6b-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: 93efi5j1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1265.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,265 90.0 71.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o332kj6b",
   "source_url": "https://omlx.ai/benchmarks/performance/o332kj6b",
   "tg_tokens": null,
   "tps": 90.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-o332kj6b-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.2 GB at this context; duplicate traces of the same session: 93efi5j1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1978.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,978 57.3 71.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o332kj6b",
   "source_url": "https://omlx.ai/benchmarks/performance/o332kj6b",
   "tg_tokens": null,
   "tps": 57.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n29kyypn-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: sz9ts3ri; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1251.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,251 85.7 71.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n29kyypn",
   "source_url": "https://omlx.ai/benchmarks/performance/n29kyypn",
   "tg_tokens": null,
   "tps": 85.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n29kyypn-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.2 GB at this context; duplicate traces of the same session: sz9ts3ri; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1789.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,789 57.9 71.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n29kyypn",
   "source_url": "https://omlx.ai/benchmarks/performance/n29kyypn",
   "tg_tokens": null,
   "tps": 57.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jpb121ta-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 71.1 GB; TTFT 833 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1230.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,230 PP tok/s 78.7 TG tok/s 833 TTFT (ms) 71.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jpb121ta",
   "source_url": "https://omlx.ai/benchmarks/performance/jpb121ta",
   "tg_tokens": null,
   "tps": 78.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tnv4oq69-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 71.0 GB; TTFT 862 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1188.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,188 PP tok/s 67.9 TG tok/s 862 TTFT (ms) 71.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tnv4oq69",
   "source_url": "https://omlx.ai/benchmarks/performance/tnv4oq69",
   "tg_tokens": null,
   "tps": 67.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-soj6n5l9-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: kjnz2twr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1143.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,143 66.3 71.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/soj6n5l9",
   "source_url": "https://omlx.ai/benchmarks/performance/soj6n5l9",
   "tg_tokens": null,
   "tps": 66.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-soj6n5l9-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.2 GB at this context; duplicate traces of the same session: kjnz2twr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1369.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,369 61.8 71.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/soj6n5l9",
   "source_url": "https://omlx.ai/benchmarks/performance/soj6n5l9",
   "tg_tokens": null,
   "tps": 61.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hj5y85uv-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 71.0 GB; TTFT 803 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1275.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,275 PP tok/s 90.9 TG tok/s 803 TTFT (ms) 71.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hj5y85uv",
   "source_url": "https://omlx.ai/benchmarks/performance/hj5y85uv",
   "tg_tokens": null,
   "tps": 90.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9oawpo8e-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 71.2 GB; TTFT 3030 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1352.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 1,352 PP tok/s 67.2 TG tok/s 3030 TTFT (ms) 71.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9oawpo8e",
   "source_url": "https://omlx.ai/benchmarks/performance/9oawpo8e",
   "tg_tokens": null,
   "tps": 67.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yckptsz9-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 71.0 GB; TTFT 857 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1196.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,196 PP tok/s 84.3 TG tok/s 857 TTFT (ms) 71.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yckptsz9",
   "source_url": "https://omlx.ai/benchmarks/performance/yckptsz9",
   "tg_tokens": null,
   "tps": 84.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2yk2luba-ctx1024",
   "model": "Qwopus3.8-27B-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: imbw0xj1, pjm47ro3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 395.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwopus3.8-27B-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 395.3 37.2 21.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2yk2luba",
   "source_url": "https://omlx.ai/benchmarks/performance/2yk2luba",
   "tg_tokens": null,
   "tps": 37.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2yk2luba-ctx4096",
   "model": "Qwopus3.8-27B-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: imbw0xj1, pjm47ro3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 473.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwopus3.8-27B-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 473.7 59.7 24.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2yk2luba",
   "source_url": "https://omlx.ai/benchmarks/performance/2yk2luba",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2yk2luba-ctx65536",
   "model": "Qwopus3.8-27B-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: imbw0xj1, pjm47ro3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 408.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwopus3.8-27B-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 408.8 63.0 30.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2yk2luba",
   "source_url": "https://omlx.ai/benchmarks/performance/2yk2luba",
   "tg_tokens": null,
   "tps": 63.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-xi2f9ogz-ctx1024",
   "model": "gemma-4-12b-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 8.5 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-12B-it-oQ4e-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 16 GB \u00b7 4bit; 1k 130.8 15.5 8.5 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xi2f9ogz",
   "source_url": "https://omlx.ai/benchmarks/performance/xi2f9ogz",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-xi2f9ogz-ctx4096",
   "model": "gemma-4-12b-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 8.8 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-12B-it-oQ4e-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 16 GB \u00b7 4bit; 4k 130.6 16.7 8.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xi2f9ogz",
   "source_url": "https://omlx.ai/benchmarks/performance/xi2f9ogz",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-dd8hws31-ctx8192",
   "model": "gemma-4-E2B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.3 GB at this context; duplicate traces of the same session: p9m10cg6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 8376.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E2B-it-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 8,376 210.1 4.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dd8hws31",
   "source_url": "https://omlx.ai/benchmarks/performance/dd8hws31",
   "tg_tokens": null,
   "tps": 210.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-dd8hws31-ctx16384",
   "model": "gemma-4-E2B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.3 GB at this context; duplicate traces of the same session: p9m10cg6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 8220.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E2B-it-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 16k 8,220 188.3 4.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dd8hws31",
   "source_url": "https://omlx.ai/benchmarks/performance/dd8hws31",
   "tg_tokens": null,
   "tps": 188.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Pro",
   "id": "omlx-b211db7w-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: 6tx8s0a3; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 603.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 16 GB \u00b7 4bit; 1k 603.2 20.2 5.9 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b211db7w",
   "source_url": "https://omlx.ai/benchmarks/performance/b211db7w",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-b211db7w-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: 6tx8s0a3; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 629.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 16 GB \u00b7 4bit; 4k 629.8 45.6 5.9 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b211db7w",
   "source_url": "https://omlx.ai/benchmarks/performance/b211db7w",
   "tg_tokens": null,
   "tps": 45.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-wpqnn17g-ctx8192",
   "model": "gemma-4-E4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 6.6 GB; TTFT 2913 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2813.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "gemma-4-E4B-it-oQ6-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 8k tokens 2,813 PP tok/s 80.3 TG tok/s 2913 TTFT (ms) 6.6 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wpqnn17g",
   "source_url": "https://omlx.ai/benchmarks/performance/wpqnn17g",
   "tg_tokens": null,
   "tps": 80.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bd48qj8x-ctx8192",
   "model": "gemma-4-E4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 6.6 GB at this context; duplicate traces of the same session: mfx4dvjb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2783.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "gemma-4-E4B-it-oQ6-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 8k 2,783 80.9 6.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bd48qj8x",
   "source_url": "https://omlx.ai/benchmarks/performance/bd48qj8x",
   "tg_tokens": null,
   "tps": 80.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bd48qj8x-ctx16384",
   "model": "gemma-4-E4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: mfx4dvjb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2792.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "gemma-4-E4B-it-oQ6-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 16k 2,792 72.7 6.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bd48qj8x",
   "source_url": "https://omlx.ai/benchmarks/performance/bd48qj8x",
   "tg_tokens": null,
   "tps": 72.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-6lesx4l4-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: 5k9e5co0, jotqny3x, v6ykoy1c, xc0ps5wy; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2423.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 2,423 84.3 11.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6lesx4l4",
   "source_url": "https://omlx.ai/benchmarks/performance/6lesx4l4",
   "tg_tokens": null,
   "tps": 84.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-6lesx4l4-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: 5k9e5co0, jotqny3x, v6ykoy1c, xc0ps5wy; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2530.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,530 79.7 11.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6lesx4l4",
   "source_url": "https://omlx.ai/benchmarks/performance/6lesx4l4",
   "tg_tokens": null,
   "tps": 79.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-6lesx4l4-ctx32768",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 12.4 GB at this context; duplicate traces of the same session: 5k9e5co0, jotqny3x, v6ykoy1c, xc0ps5wy; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2027.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 2,027 53.4 12.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6lesx4l4",
   "source_url": "https://omlx.ai/benchmarks/performance/6lesx4l4",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-6lesx4l4-ctx65536",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 13.2 GB at this context; duplicate traces of the same session: 5k9e5co0, jotqny3x, v6ykoy1c, xc0ps5wy; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1556.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 64k 1,556 38.7 13.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6lesx4l4",
   "source_url": "https://omlx.ai/benchmarks/performance/6lesx4l4",
   "tg_tokens": null,
   "tps": 38.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-6lesx4l4-ctx131072",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: 5k9e5co0, jotqny3x, v6ykoy1c, xc0ps5wy; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1143.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 128k 1,143 24.8 14.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6lesx4l4",
   "source_url": "https://omlx.ai/benchmarks/performance/6lesx4l4",
   "tg_tokens": null,
   "tps": 24.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-xmln8851-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: kg8wqt3p, s2rkppo9, v07ymrxx, xbzl6llz; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 983.5,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 983.5 84.2 11.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xmln8851",
   "source_url": "https://omlx.ai/benchmarks/performance/xmln8851",
   "tg_tokens": null,
   "tps": 84.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-xmln8851-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: kg8wqt3p, s2rkppo9, v07ymrxx, xbzl6llz; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2354.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,354 80.3 11.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xmln8851",
   "source_url": "https://omlx.ai/benchmarks/performance/xmln8851",
   "tg_tokens": null,
   "tps": 80.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-xmln8851-ctx32768",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 12.4 GB at this context; duplicate traces of the same session: kg8wqt3p, s2rkppo9, v07ymrxx, xbzl6llz; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2001.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 2,001 53.3 12.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xmln8851",
   "source_url": "https://omlx.ai/benchmarks/performance/xmln8851",
   "tg_tokens": null,
   "tps": 53.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-xmln8851-ctx65536",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 13.2 GB at this context; duplicate traces of the same session: kg8wqt3p, s2rkppo9, v07ymrxx, xbzl6llz; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1502.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 64k 1,502 37.2 13.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xmln8851",
   "source_url": "https://omlx.ai/benchmarks/performance/xmln8851",
   "tg_tokens": null,
   "tps": 37.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-xmln8851-ctx131072",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: kg8wqt3p, s2rkppo9, v07ymrxx, xbzl6llz; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1126.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 128k 1,126 24.7 14.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xmln8851",
   "source_url": "https://omlx.ai/benchmarks/performance/xmln8851",
   "tg_tokens": null,
   "tps": 24.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ceq50mqy-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: b68wlwuq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2059.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "LFM2.5-24B-A2B-oQ5e-fp16 on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 2,059 146.0 16.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ceq50mqy",
   "source_url": "https://omlx.ai/benchmarks/performance/ceq50mqy",
   "tg_tokens": null,
   "tps": 146.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ceq50mqy-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: b68wlwuq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2326.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "LFM2.5-24B-A2B-oQ5e-fp16 on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 2,326 142.8 16.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ceq50mqy",
   "source_url": "https://omlx.ai/benchmarks/performance/ceq50mqy",
   "tg_tokens": null,
   "tps": 142.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-arrc495o-ctx1024",
   "model": "muse-glimmer 30B",
   "notes": "OptiQ; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 395.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-30B-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 395.1 14.8 19.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/arrc495o",
   "source_url": "https://omlx.ai/benchmarks/performance/arrc495o",
   "tg_tokens": null,
   "tps": 14.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-arrc495o-ctx4096",
   "model": "muse-glimmer 30B",
   "notes": "OptiQ; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 409.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-30B-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 409.3 14.6 19.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/arrc495o",
   "source_url": "https://omlx.ai/benchmarks/performance/arrc495o",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-7dxdnn4g-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 882.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 882.8 71.7 20.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7dxdnn4g",
   "source_url": "https://omlx.ai/benchmarks/performance/7dxdnn4g",
   "tg_tokens": null,
   "tps": 71.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-7dxdnn4g-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 971.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 971.6 69.7 20.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7dxdnn4g",
   "source_url": "https://omlx.ai/benchmarks/performance/7dxdnn4g",
   "tg_tokens": null,
   "tps": 69.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-7dxdnn4g-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 929.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 929.4 66.6 21.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7dxdnn4g",
   "source_url": "https://omlx.ai/benchmarks/performance/7dxdnn4g",
   "tg_tokens": null,
   "tps": 66.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-7dxdnn4g-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 840.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 840.6 62.3 21.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7dxdnn4g",
   "source_url": "https://omlx.ai/benchmarks/performance/7dxdnn4g",
   "tg_tokens": null,
   "tps": 62.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-7dxdnn4g-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 677.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 677.5 54.9 21.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7dxdnn4g",
   "source_url": "https://omlx.ai/benchmarks/performance/7dxdnn4g",
   "tg_tokens": null,
   "tps": 54.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-7dxdnn4g-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 501.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 501.3 43.4 22.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7dxdnn4g",
   "source_url": "https://omlx.ai/benchmarks/performance/7dxdnn4g",
   "tg_tokens": null,
   "tps": 43.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-8o90i1eq-ctx8192",
   "model": "Qwen3-VL-8B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 6.1 GB at this context; duplicate traces of the same session: d2z10urn; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1044.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-VL-8B-Instruct-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 1,044 90.5 6.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8o90i1eq",
   "source_url": "https://omlx.ai/benchmarks/performance/8o90i1eq",
   "tg_tokens": null,
   "tps": 90.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-8o90i1eq-ctx16384",
   "model": "Qwen3-VL-8B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 7.1 GB at this context; duplicate traces of the same session: d2z10urn; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 930.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-VL-8B-Instruct-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 16k 930.5 72.7 7.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8o90i1eq",
   "source_url": "https://omlx.ai/benchmarks/performance/8o90i1eq",
   "tg_tokens": null,
   "tps": 72.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-z5gkbw7p-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 3.5 GB at this context; duplicate traces of the same session: 3cdvhi5s, if2xq11q, s4v5qbac; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 718.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-4B-oQ4e-mtp on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 1k 718.1 93.4 3.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z5gkbw7p",
   "source_url": "https://omlx.ai/benchmarks/performance/z5gkbw7p",
   "tg_tokens": null,
   "tps": 93.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-z5gkbw7p-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.1 GB at this context; duplicate traces of the same session: 3cdvhi5s, if2xq11q, s4v5qbac; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 691.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-4B-oQ4e-mtp on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 4k 691.2 89.5 4.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z5gkbw7p",
   "source_url": "https://omlx.ai/benchmarks/performance/z5gkbw7p",
   "tg_tokens": null,
   "tps": 89.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-z5gkbw7p-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.3 GB at this context; duplicate traces of the same session: 3cdvhi5s, if2xq11q, s4v5qbac; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 643.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-4B-oQ4e-mtp on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 8k 643.0 84.8 4.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z5gkbw7p",
   "source_url": "https://omlx.ai/benchmarks/performance/z5gkbw7p",
   "tg_tokens": null,
   "tps": 84.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-z5gkbw7p-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.7 GB at this context; duplicate traces of the same session: 3cdvhi5s, if2xq11q, s4v5qbac; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 575.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-4B-oQ4e-mtp on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 16k 575.3 76.7 4.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z5gkbw7p",
   "source_url": "https://omlx.ai/benchmarks/performance/z5gkbw7p",
   "tg_tokens": null,
   "tps": 76.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-o7ifnvxa-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 04751hn7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 780.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 780.4 99.4 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o7ifnvxa",
   "source_url": "https://omlx.ai/benchmarks/performance/o7ifnvxa",
   "tg_tokens": null,
   "tps": 99.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-o7ifnvxa-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 04751hn7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2251.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,251 100.9 20.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o7ifnvxa",
   "source_url": "https://omlx.ai/benchmarks/performance/o7ifnvxa",
   "tg_tokens": null,
   "tps": 100.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-ju5dsmvt-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 5bkv0y20; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1189.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 1,189 56.9 21.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ju5dsmvt",
   "source_url": "https://omlx.ai/benchmarks/performance/ju5dsmvt",
   "tg_tokens": null,
   "tps": 56.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-ju5dsmvt-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 5bkv0y20; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2045.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 2,045 62.0 21.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ju5dsmvt",
   "source_url": "https://omlx.ai/benchmarks/performance/ju5dsmvt",
   "tg_tokens": null,
   "tps": 62.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-7k2rba7h-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 62k4hpck, 6gui0b6c, kkfqr1qt; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 53.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 53.7 10.3 16.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7k2rba7h",
   "source_url": "https://omlx.ai/benchmarks/performance/7k2rba7h",
   "tg_tokens": null,
   "tps": 10.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-7k2rba7h-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: 62k4hpck, 6gui0b6c, kkfqr1qt; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 53.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 53.0 9.5 17.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7k2rba7h",
   "source_url": "https://omlx.ai/benchmarks/performance/7k2rba7h",
   "tg_tokens": null,
   "tps": 9.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-7k2rba7h-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: 62k4hpck, 6gui0b6c, kkfqr1qt; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 51.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 8k 51.6 9.8 17.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7k2rba7h",
   "source_url": "https://omlx.ai/benchmarks/performance/7k2rba7h",
   "tg_tokens": null,
   "tps": 9.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-7k2rba7h-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: 62k4hpck, 6gui0b6c, kkfqr1qt; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 47.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k 47.5 8.5 18.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7k2rba7h",
   "source_url": "https://omlx.ai/benchmarks/performance/7k2rba7h",
   "tg_tokens": null,
   "tps": 8.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-otypz16c-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 0r9vhu5d; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 106.6 16.0 18.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/otypz16c",
   "source_url": "https://omlx.ai/benchmarks/performance/otypz16c",
   "tg_tokens": null,
   "tps": 16.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-otypz16c-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: 0r9vhu5d; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 101.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 101.9 21.0 22.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/otypz16c",
   "source_url": "https://omlx.ai/benchmarks/performance/otypz16c",
   "tg_tokens": null,
   "tps": 21.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ycgxig4h-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: botlt9i9, k3l0b5pc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 260.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 260.6 28.5 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ycgxig4h",
   "source_url": "https://omlx.ai/benchmarks/performance/ycgxig4h",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ycgxig4h-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: botlt9i9, k3l0b5pc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 271.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 271.0 27.3 20.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ycgxig4h",
   "source_url": "https://omlx.ai/benchmarks/performance/ycgxig4h",
   "tg_tokens": null,
   "tps": 27.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ycgxig4h-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: botlt9i9, k3l0b5pc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 212.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 212.2 22.1 25.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ycgxig4h",
   "source_url": "https://omlx.ai/benchmarks/performance/ycgxig4h",
   "tg_tokens": null,
   "tps": 22.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q09jbb2b-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: z7erqlvf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 917.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 917.3 71.5 17.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q09jbb2b",
   "source_url": "https://omlx.ai/benchmarks/performance/q09jbb2b",
   "tg_tokens": null,
   "tps": 71.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q09jbb2b-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: z7erqlvf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 956.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 956.5 67.4 18.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q09jbb2b",
   "source_url": "https://omlx.ai/benchmarks/performance/q09jbb2b",
   "tg_tokens": null,
   "tps": 67.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kkmicuv0-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: lg13rzda; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 799.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 799.0 62.2 17.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kkmicuv0",
   "source_url": "https://omlx.ai/benchmarks/performance/kkmicuv0",
   "tg_tokens": null,
   "tps": 62.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kkmicuv0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: lg13rzda; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 797.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 797.0 62.8 18.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kkmicuv0",
   "source_url": "https://omlx.ai/benchmarks/performance/kkmicuv0",
   "tg_tokens": null,
   "tps": 62.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-m8pk37oh-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: ztg0yrkj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 898.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 898.3 61.3 17.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m8pk37oh",
   "source_url": "https://omlx.ai/benchmarks/performance/m8pk37oh",
   "tg_tokens": null,
   "tps": 61.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-m8pk37oh-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: ztg0yrkj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 887.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 887.6 60.1 18.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m8pk37oh",
   "source_url": "https://omlx.ai/benchmarks/performance/m8pk37oh",
   "tg_tokens": null,
   "tps": 60.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rqyq7aw5-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: wux6qhwz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 902.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 902.6 57.5 17.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rqyq7aw5",
   "source_url": "https://omlx.ai/benchmarks/performance/rqyq7aw5",
   "tg_tokens": null,
   "tps": 57.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rqyq7aw5-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: wux6qhwz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 935.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 935.0 60.7 18.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rqyq7aw5",
   "source_url": "https://omlx.ai/benchmarks/performance/rqyq7aw5",
   "tg_tokens": null,
   "tps": 60.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kigd66mz-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: ieaq542y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 859.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 859.0 62.0 17.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kigd66mz",
   "source_url": "https://omlx.ai/benchmarks/performance/kigd66mz",
   "tg_tokens": null,
   "tps": 62.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kigd66mz-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: ieaq542y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 820.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 820.2 55.0 18.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kigd66mz",
   "source_url": "https://omlx.ai/benchmarks/performance/kigd66mz",
   "tg_tokens": null,
   "tps": 55.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4cb4dfbm-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 8i7k56z1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 917.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 917.8 72.3 17.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4cb4dfbm",
   "source_url": "https://omlx.ai/benchmarks/performance/4cb4dfbm",
   "tg_tokens": null,
   "tps": 72.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4cb4dfbm-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: 8i7k56z1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 960.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 960.6 64.3 18.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4cb4dfbm",
   "source_url": "https://omlx.ai/benchmarks/performance/4cb4dfbm",
   "tg_tokens": null,
   "tps": 64.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-au6eb0p0-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 73.7 GB at this context; duplicate traces of the same session: 0eodperv, 20z9u8e8, 2s30rs2y, 7ea207c3, uu3t9rln, vuu2r86b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 322.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 322.2 40.2 73.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/au6eb0p0",
   "source_url": "https://omlx.ai/benchmarks/performance/au6eb0p0",
   "tg_tokens": null,
   "tps": 40.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-au6eb0p0-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 74.3 GB at this context; duplicate traces of the same session: 0eodperv, 20z9u8e8, 2s30rs2y, 7ea207c3, uu3t9rln, vuu2r86b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 349.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 349.2 40.5 74.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/au6eb0p0",
   "source_url": "https://omlx.ai/benchmarks/performance/au6eb0p0",
   "tg_tokens": null,
   "tps": 40.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-au6eb0p0-ctx8192",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 74.4 GB at this context; duplicate traces of the same session: 0eodperv, 20z9u8e8, 2s30rs2y, 7ea207c3, uu3t9rln, vuu2r86b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 554.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 554.6 35.0 74.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/au6eb0p0",
   "source_url": "https://omlx.ai/benchmarks/performance/au6eb0p0",
   "tg_tokens": null,
   "tps": 35.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-au6eb0p0-ctx16384",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 74.7 GB at this context; duplicate traces of the same session: 0eodperv, 20z9u8e8, 2s30rs2y, 7ea207c3, uu3t9rln, vuu2r86b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 511.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 511.5 39.2 74.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/au6eb0p0",
   "source_url": "https://omlx.ai/benchmarks/performance/au6eb0p0",
   "tg_tokens": null,
   "tps": 39.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-au6eb0p0-ctx65536",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 77.6 GB at this context; duplicate traces of the same session: 0eodperv, 20z9u8e8, 2s30rs2y, 7ea207c3, uu3t9rln, vuu2r86b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 492.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 492.5 26.4 77.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/au6eb0p0",
   "source_url": "https://omlx.ai/benchmarks/performance/au6eb0p0",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-au6eb0p0-ctx131072",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 81.6 GB at this context; duplicate traces of the same session: 0eodperv, 20z9u8e8, 2s30rs2y, 7ea207c3, uu3t9rln, vuu2r86b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 403.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 403.2 29.3 81.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/au6eb0p0",
   "source_url": "https://omlx.ai/benchmarks/performance/au6eb0p0",
   "tg_tokens": null,
   "tps": 29.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-au6eb0p0-ctx199680",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 83.0 GB at this context; duplicate traces of the same session: 0eodperv, 20z9u8e8, 2s30rs2y, 7ea207c3, uu3t9rln, vuu2r86b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 432.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 432.9 27.8 83.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/au6eb0p0",
   "source_url": "https://omlx.ai/benchmarks/performance/au6eb0p0",
   "tg_tokens": null,
   "tps": 27.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-1ctfye1g-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: 5j9fey27; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 517.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 517.2 60.3 70.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ctfye1g",
   "source_url": "https://omlx.ai/benchmarks/performance/1ctfye1g",
   "tg_tokens": null,
   "tps": 60.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-1ctfye1g-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: 5j9fey27; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 654.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 654.8 55.9 71.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ctfye1g",
   "source_url": "https://omlx.ai/benchmarks/performance/1ctfye1g",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-1ctfye1g-ctx16384",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.0 GB at this context; duplicate traces of the same session: 5j9fey27; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 705.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 705.6 57.3 72.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ctfye1g",
   "source_url": "https://omlx.ai/benchmarks/performance/1ctfye1g",
   "tg_tokens": null,
   "tps": 57.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-pcy40rtf-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: 2ij60qrm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 529.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 529.5 59.0 70.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pcy40rtf",
   "source_url": "https://omlx.ai/benchmarks/performance/pcy40rtf",
   "tg_tokens": null,
   "tps": 59.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-pcy40rtf-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: 2ij60qrm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 653.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 653.2 56.9 71.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pcy40rtf",
   "source_url": "https://omlx.ai/benchmarks/performance/pcy40rtf",
   "tg_tokens": null,
   "tps": 56.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-xwm29sn5-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.9 GB; TTFT 2035 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 503.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 503.2 PP tok/s 51.5 TG tok/s 2035 TTFT (ms) 70.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xwm29sn5",
   "source_url": "https://omlx.ai/benchmarks/performance/xwm29sn5",
   "tg_tokens": null,
   "tps": 51.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9w1zkekf-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: pfj1psms; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1259.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,259 78.5 71.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9w1zkekf",
   "source_url": "https://omlx.ai/benchmarks/performance/9w1zkekf",
   "tg_tokens": null,
   "tps": 78.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9w1zkekf-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.2 GB at this context; duplicate traces of the same session: pfj1psms; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1968.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,968 44.2 71.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9w1zkekf",
   "source_url": "https://omlx.ai/benchmarks/performance/9w1zkekf",
   "tg_tokens": null,
   "tps": 44.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-799v93ft-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: juks5620; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1173.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,173 76.5 71.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/799v93ft",
   "source_url": "https://omlx.ai/benchmarks/performance/799v93ft",
   "tg_tokens": null,
   "tps": 76.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-799v93ft-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.2 GB at this context; duplicate traces of the same session: juks5620; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1972.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,972 66.7 71.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/799v93ft",
   "source_url": "https://omlx.ai/benchmarks/performance/799v93ft",
   "tg_tokens": null,
   "tps": 66.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1vpe3buc-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 71.0 GB; TTFT 810 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1265.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,265 PP tok/s 82.1 TG tok/s 810 TTFT (ms) 71.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1vpe3buc",
   "source_url": "https://omlx.ai/benchmarks/performance/1vpe3buc",
   "tg_tokens": null,
   "tps": 82.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-43y00q5e-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 71.0 GB; TTFT 807 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1269.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,269 PP tok/s 83.0 TG tok/s 807 TTFT (ms) 71.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/43y00q5e",
   "source_url": "https://omlx.ai/benchmarks/performance/43y00q5e",
   "tg_tokens": null,
   "tps": 83.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mopb7b01-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 71.1 GB; TTFT 806 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1270.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,270 PP tok/s 90.4 TG tok/s 806 TTFT (ms) 71.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mopb7b01",
   "source_url": "https://omlx.ai/benchmarks/performance/mopb7b01",
   "tg_tokens": null,
   "tps": 90.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nkkxp1ag-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 71.1 GB; TTFT 809 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1265.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,265 PP tok/s 85.2 TG tok/s 809 TTFT (ms) 71.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nkkxp1ag",
   "source_url": "https://omlx.ai/benchmarks/performance/nkkxp1ag",
   "tg_tokens": null,
   "tps": 85.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-wu0gznx8-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 184.1 GB at this context; duplicate traces of the same session: ifkul7qb, xjjzvbfc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1047.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 1,047 30.8 184.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wu0gznx8",
   "source_url": "https://omlx.ai/benchmarks/performance/wu0gznx8",
   "tg_tokens": null,
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-wu0gznx8-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 184.7 GB at this context; duplicate traces of the same session: ifkul7qb, xjjzvbfc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1037.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 16k 1,037 66.8 184.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wu0gznx8",
   "source_url": "https://omlx.ai/benchmarks/performance/wu0gznx8",
   "tg_tokens": null,
   "tps": 66.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-wu0gznx8-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 185.7 GB at this context; duplicate traces of the same session: ifkul7qb, xjjzvbfc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1028.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 32k 1,028 74.7 185.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wu0gznx8",
   "source_url": "https://omlx.ai/benchmarks/performance/wu0gznx8",
   "tg_tokens": null,
   "tps": 74.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ahr8z2bd-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 904.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 904.3 101.8 21.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ahr8z2bd",
   "source_url": "https://omlx.ai/benchmarks/performance/ahr8z2bd",
   "tg_tokens": null,
   "tps": 101.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ahr8z2bd-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 924.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 924.0 98.9 22.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ahr8z2bd",
   "source_url": "https://omlx.ai/benchmarks/performance/ahr8z2bd",
   "tg_tokens": null,
   "tps": 98.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-u9nssrbh-ctx1024",
   "model": "gemma-4-12b-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 8.5 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-12B-it-oQ4e-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 16 GB \u00b7 4bit; 1k 130.8 15.5 8.5 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u9nssrbh",
   "source_url": "https://omlx.ai/benchmarks/performance/u9nssrbh",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-u9nssrbh-ctx4096",
   "model": "gemma-4-12b-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 8.8 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-12B-it-oQ4e-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 16 GB \u00b7 4bit; 4k 130.6 16.7 8.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u9nssrbh",
   "source_url": "https://omlx.ai/benchmarks/performance/u9nssrbh",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-315wtff5-ctx1024",
   "model": "MiniCPM-V-4.7-1B",
   "notes": "oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: mr9i7ota; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1982.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM-V-4.7-1B-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 1,982 79.3 3.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/315wtff5",
   "source_url": "https://omlx.ai/benchmarks/performance/315wtff5",
   "tg_tokens": null,
   "tps": 79.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-315wtff5-ctx4096",
   "model": "MiniCPM-V-4.7-1B",
   "notes": "oMLX community benchmark trace; peak memory 3.7 GB at this context; duplicate traces of the same session: mr9i7ota; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2437.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM-V-4.7-1B-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 2,437 78.9 3.7 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/315wtff5",
   "source_url": "https://omlx.ai/benchmarks/performance/315wtff5",
   "tg_tokens": null,
   "tps": 78.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-85lh8g0x-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 0k6jepdp, 8cqugvl0, pg69np2d, tg3ldjb2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 882.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 882.8 71.7 20.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/85lh8g0x",
   "source_url": "https://omlx.ai/benchmarks/performance/85lh8g0x",
   "tg_tokens": null,
   "tps": 71.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-85lh8g0x-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 0k6jepdp, 8cqugvl0, pg69np2d, tg3ldjb2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 971.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 971.6 69.7 20.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/85lh8g0x",
   "source_url": "https://omlx.ai/benchmarks/performance/85lh8g0x",
   "tg_tokens": null,
   "tps": 69.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-85lh8g0x-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: 0k6jepdp, 8cqugvl0, pg69np2d, tg3ldjb2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 929.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 929.4 66.6 21.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/85lh8g0x",
   "source_url": "https://omlx.ai/benchmarks/performance/85lh8g0x",
   "tg_tokens": null,
   "tps": 66.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-85lh8g0x-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 0k6jepdp, 8cqugvl0, pg69np2d, tg3ldjb2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 840.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 840.6 62.3 21.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/85lh8g0x",
   "source_url": "https://omlx.ai/benchmarks/performance/85lh8g0x",
   "tg_tokens": null,
   "tps": 62.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-85lh8g0x-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: 0k6jepdp, 8cqugvl0, pg69np2d, tg3ldjb2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 677.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 677.5 54.9 21.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/85lh8g0x",
   "source_url": "https://omlx.ai/benchmarks/performance/85lh8g0x",
   "tg_tokens": null,
   "tps": 54.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-85lh8g0x-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: 0k6jepdp, 8cqugvl0, pg69np2d, tg3ldjb2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 501.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 501.3 43.4 22.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/85lh8g0x",
   "source_url": "https://omlx.ai/benchmarks/performance/85lh8g0x",
   "tg_tokens": null,
   "tps": 43.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-16bv2otr-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: uyzuvfxv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 906.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 906.1 92.3 21.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/16bv2otr",
   "source_url": "https://omlx.ai/benchmarks/performance/16bv2otr",
   "tg_tokens": null,
   "tps": 92.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-16bv2otr-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: uyzuvfxv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 952.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 952.9 86.3 22.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/16bv2otr",
   "source_url": "https://omlx.ai/benchmarks/performance/16bv2otr",
   "tg_tokens": null,
   "tps": 86.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-gzjrcyvm-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 903.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 903.4 93.3 21.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gzjrcyvm",
   "source_url": "https://omlx.ai/benchmarks/performance/gzjrcyvm",
   "tg_tokens": null,
   "tps": 93.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-gzjrcyvm-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 952.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 952.6 97.3 22.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gzjrcyvm",
   "source_url": "https://omlx.ai/benchmarks/performance/gzjrcyvm",
   "tg_tokens": null,
   "tps": 97.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-16ufz0js-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: l720qfg2; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1581.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized-oQ4e-fp16-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 1,581 71.7 22.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/16ufz0js",
   "source_url": "https://omlx.ai/benchmarks/performance/16ufz0js",
   "tg_tokens": null,
   "tps": 71.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-16ufz0js-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: l720qfg2; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2036.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized-oQ4e-fp16-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 2,036 64.4 22.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/16ufz0js",
   "source_url": "https://omlx.ai/benchmarks/performance/16ufz0js",
   "tg_tokens": null,
   "tps": 64.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-lyrlt36z-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: olziifan; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 59.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 59.9 6.4 17.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lyrlt36z",
   "source_url": "https://omlx.ai/benchmarks/performance/lyrlt36z",
   "tg_tokens": null,
   "tps": 6.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-lyrlt36z-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: olziifan; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 51.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 16k 51.9 5.9 18.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lyrlt36z",
   "source_url": "https://omlx.ai/benchmarks/performance/lyrlt36z",
   "tg_tokens": null,
   "tps": 5.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-p3ktjfl5-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: nk0nvgeg; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1308.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k 1,308 54.7 16.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p3ktjfl5",
   "source_url": "https://omlx.ai/benchmarks/performance/p3ktjfl5",
   "tg_tokens": null,
   "tps": 54.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-p3ktjfl5-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: nk0nvgeg; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1495.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 4k 1,495 52.4 17.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p3ktjfl5",
   "source_url": "https://omlx.ai/benchmarks/performance/p3ktjfl5",
   "tg_tokens": null,
   "tps": 52.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-i6fz5lf7-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 8xfhqsj5; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.5,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 323.5 61.9 19.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i6fz5lf7",
   "source_url": "https://omlx.ai/benchmarks/performance/i6fz5lf7",
   "tg_tokens": null,
   "tps": 61.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-i6fz5lf7-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: 8xfhqsj5; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 356.8,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 356.8 65.1 22.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i6fz5lf7",
   "source_url": "https://omlx.ai/benchmarks/performance/i6fz5lf7",
   "tg_tokens": null,
   "tps": 65.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-83blac1q-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: oid531mt; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 321.4,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 321.4 43.9 19.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/83blac1q",
   "source_url": "https://omlx.ai/benchmarks/performance/83blac1q",
   "tg_tokens": null,
   "tps": 43.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-83blac1q-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: oid531mt; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 353.8,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 353.8 39.3 22.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/83blac1q",
   "source_url": "https://omlx.ai/benchmarks/performance/83blac1q",
   "tg_tokens": null,
   "tps": 39.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-f0jmqfc0-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: tu5smlgi; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 324.4,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 324.4 77.3 19.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f0jmqfc0",
   "source_url": "https://omlx.ai/benchmarks/performance/f0jmqfc0",
   "tg_tokens": null,
   "tps": 77.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-f0jmqfc0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: tu5smlgi; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 356.7,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 356.7 71.6 22.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f0jmqfc0",
   "source_url": "https://omlx.ai/benchmarks/performance/f0jmqfc0",
   "tg_tokens": null,
   "tps": 71.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-us7a7dx9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 19.8 GB; TTFT 135556 ms; duplicate traces of the same session: none; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 30.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 4bit; 4k tokens 30.2 PP tok/s 5.2 TG tok/s 135556 TTFT (ms) 19.8 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/us7a7dx9",
   "source_url": "https://omlx.ai/benchmarks/performance/us7a7dx9",
   "tg_tokens": null,
   "tps": 5.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-gvfnhckc-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 17.7 GB; TTFT 118459 ms; duplicate traces of the same session: none; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 34.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 4bit; 4k tokens 34.6 PP tok/s 3.4 TG tok/s 118459 TTFT (ms) 17.7 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gvfnhckc",
   "source_url": "https://omlx.ai/benchmarks/performance/gvfnhckc",
   "tg_tokens": null,
   "tps": 3.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-z68l0ll7-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 17.7 GB; TTFT 132624 ms; duplicate traces of the same session: none; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 30.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 4bit; 4k tokens 30.9 PP tok/s 3.1 TG tok/s 132624 TTFT (ms) 17.7 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z68l0ll7",
   "source_url": "https://omlx.ai/benchmarks/performance/z68l0ll7",
   "tg_tokens": null,
   "tps": 3.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-qm6hkpyt-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 5f0fd3cx, s4i4rg8p; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 248.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 248.7 52.6 21.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qm6hkpyt",
   "source_url": "https://omlx.ai/benchmarks/performance/qm6hkpyt",
   "tg_tokens": null,
   "tps": 52.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-qm6hkpyt-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.6 GB at this context; duplicate traces of the same session: 5f0fd3cx, s4i4rg8p; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 336.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 336.1 37.3 25.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qm6hkpyt",
   "source_url": "https://omlx.ai/benchmarks/performance/qm6hkpyt",
   "tg_tokens": null,
   "tps": 37.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-qm6hkpyt-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.8 GB at this context; duplicate traces of the same session: 5f0fd3cx, s4i4rg8p; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 247.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 247.4 35.0 30.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qm6hkpyt",
   "source_url": "https://omlx.ai/benchmarks/performance/qm6hkpyt",
   "tg_tokens": null,
   "tps": 35.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ynzfvtkn-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: ukrrnaeu; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1278.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k 1,278 51.5 16.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ynzfvtkn",
   "source_url": "https://omlx.ai/benchmarks/performance/ynzfvtkn",
   "tg_tokens": null,
   "tps": 51.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ynzfvtkn-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: ukrrnaeu; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1491.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 4k 1,491 50.0 17.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ynzfvtkn",
   "source_url": "https://omlx.ai/benchmarks/performance/ynzfvtkn",
   "tg_tokens": null,
   "tps": 50.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-pfyfb3ye-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1253.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k 1,253 52.2 16.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pfyfb3ye",
   "source_url": "https://omlx.ai/benchmarks/performance/pfyfb3ye",
   "tg_tokens": null,
   "tps": 52.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-pfyfb3ye-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1452.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 4k 1,452 50.7 17.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pfyfb3ye",
   "source_url": "https://omlx.ai/benchmarks/performance/pfyfb3ye",
   "tg_tokens": null,
   "tps": 50.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3x8a81b3-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 177.8 GB at this context; duplicate traces of the same session: 6vdtv047, al23eai9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 805.1,
   "provenance": "community",
   "quant": "FP8",
   "quote": "Qwen3.8-Flash-Next-FP8 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 1k 805.1 61.2 177.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3x8a81b3",
   "source_url": "https://omlx.ai/benchmarks/performance/3x8a81b3",
   "tg_tokens": null,
   "tps": 61.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3x8a81b3-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 178.1 GB at this context; duplicate traces of the same session: 6vdtv047, al23eai9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 953.2,
   "provenance": "community",
   "quant": "FP8",
   "quote": "Qwen3.8-Flash-Next-FP8 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 4k 953.2 51.7 178.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3x8a81b3",
   "source_url": "https://omlx.ai/benchmarks/performance/3x8a81b3",
   "tg_tokens": null,
   "tps": 51.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3x8a81b3-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 178.5 GB at this context; duplicate traces of the same session: 6vdtv047, al23eai9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 935.8,
   "provenance": "community",
   "quant": "FP8",
   "quote": "Qwen3.8-Flash-Next-FP8 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 8k 935.8 50.1 178.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3x8a81b3",
   "source_url": "https://omlx.ai/benchmarks/performance/3x8a81b3",
   "tg_tokens": null,
   "tps": 50.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1y6jri6h-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 177.8 GB at this context; duplicate traces of the same session: 9g3p16zq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 614.4,
   "provenance": "community",
   "quant": "FP8",
   "quote": "Qwen3.8-Flash-Next-FP8 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 1k 614.4 39.5 177.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1y6jri6h",
   "source_url": "https://omlx.ai/benchmarks/performance/1y6jri6h",
   "tg_tokens": null,
   "tps": 39.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1y6jri6h-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 178.1 GB at this context; duplicate traces of the same session: 9g3p16zq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 818.7,
   "provenance": "community",
   "quant": "FP8",
   "quote": "Qwen3.8-Flash-Next-FP8 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 4k 818.7 48.5 178.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1y6jri6h",
   "source_url": "https://omlx.ai/benchmarks/performance/1y6jri6h",
   "tg_tokens": null,
   "tps": 48.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-t7qfkbn9-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 45.7 GB; TTFT 36643 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1789.0,
   "provenance": "community",
   "quant": "GSQ-RCO-Q2",
   "quote": "Qwen3.8-Flash-Next-GSQ-RCO-Q2-mlx on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k tokens 1,789 PP tok/s 36.7 TG tok/s 36643 TTFT (ms) 45.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t7qfkbn9",
   "source_url": "https://omlx.ai/benchmarks/performance/t7qfkbn9",
   "tg_tokens": null,
   "tps": 36.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-xai6pnrb-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 39.8 GB at this context; duplicate traces of the same session: ahntcl69; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1051.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k 1,051 69.1 39.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xai6pnrb",
   "source_url": "https://omlx.ai/benchmarks/performance/xai6pnrb",
   "tg_tokens": null,
   "tps": 69.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-xai6pnrb-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 40.0 GB at this context; duplicate traces of the same session: ahntcl69; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1926.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 4k 1,926 63.1 40.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xai6pnrb",
   "source_url": "https://omlx.ai/benchmarks/performance/xai6pnrb",
   "tg_tokens": null,
   "tps": 63.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-59o529wx-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 904.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 904.3 101.8 21.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/59o529wx",
   "source_url": "https://omlx.ai/benchmarks/performance/59o529wx",
   "tg_tokens": null,
   "tps": 101.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-59o529wx-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 924.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 924.0 98.9 22.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/59o529wx",
   "source_url": "https://omlx.ai/benchmarks/performance/59o529wx",
   "tg_tokens": null,
   "tps": 98.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-0euhyfgm-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 65rb2l1x; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 895.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 895.9 88.5 21.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0euhyfgm",
   "source_url": "https://omlx.ai/benchmarks/performance/0euhyfgm",
   "tg_tokens": null,
   "tps": 88.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-0euhyfgm-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 65rb2l1x; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 908.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 908.5 89.6 22.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0euhyfgm",
   "source_url": "https://omlx.ai/benchmarks/performance/0euhyfgm",
   "tg_tokens": null,
   "tps": 89.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-tz6hsa1h-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: b61le2u9; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 892.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 892.5 86.1 21.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tz6hsa1h",
   "source_url": "https://omlx.ai/benchmarks/performance/tz6hsa1h",
   "tg_tokens": null,
   "tps": 86.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-tz6hsa1h-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: b61le2u9; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 896.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 896.3 88.7 22.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tz6hsa1h",
   "source_url": "https://omlx.ai/benchmarks/performance/tz6hsa1h",
   "tg_tokens": null,
   "tps": 88.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-iu8untun-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: m9snfzjf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 893.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 893.3 102.3 21.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iu8untun",
   "source_url": "https://omlx.ai/benchmarks/performance/iu8untun",
   "tg_tokens": null,
   "tps": 102.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-iu8untun-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: m9snfzjf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 906.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 906.7 90.2 22.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iu8untun",
   "source_url": "https://omlx.ai/benchmarks/performance/iu8untun",
   "tg_tokens": null,
   "tps": 90.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-7kjt0qw4-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: youtakki; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 906.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 906.2 95.1 21.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7kjt0qw4",
   "source_url": "https://omlx.ai/benchmarks/performance/7kjt0qw4",
   "tg_tokens": null,
   "tps": 95.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-7kjt0qw4-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: youtakki; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 935.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 935.1 97.1 22.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7kjt0qw4",
   "source_url": "https://omlx.ai/benchmarks/performance/7kjt0qw4",
   "tg_tokens": null,
   "tps": 97.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-xwn88wzy-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: gh9mj6q7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 890.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 890.7 95.6 21.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xwn88wzy",
   "source_url": "https://omlx.ai/benchmarks/performance/xwn88wzy",
   "tg_tokens": null,
   "tps": 95.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-xwn88wzy-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: gh9mj6q7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 900.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 900.8 91.8 22.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xwn88wzy",
   "source_url": "https://omlx.ai/benchmarks/performance/xwn88wzy",
   "tg_tokens": null,
   "tps": 91.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ryj3kxwl-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: txamdb1k; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 903.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 903.7 87.7 21.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ryj3kxwl",
   "source_url": "https://omlx.ai/benchmarks/performance/ryj3kxwl",
   "tg_tokens": null,
   "tps": 87.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ryj3kxwl-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: txamdb1k; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 950.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 950.5 84.8 22.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ryj3kxwl",
   "source_url": "https://omlx.ai/benchmarks/performance/ryj3kxwl",
   "tg_tokens": null,
   "tps": 84.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-z0b1ctu3-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: iiccw4ji; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 906.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 906.0 97.7 21.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z0b1ctu3",
   "source_url": "https://omlx.ai/benchmarks/performance/z0b1ctu3",
   "tg_tokens": null,
   "tps": 97.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-z0b1ctu3-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: iiccw4ji; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 928.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 928.1 86.5 22.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z0b1ctu3",
   "source_url": "https://omlx.ai/benchmarks/performance/z0b1ctu3",
   "tg_tokens": null,
   "tps": 86.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-a5hixrx2-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: onyrnzc0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 902.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 902.5 89.0 21.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a5hixrx2",
   "source_url": "https://omlx.ai/benchmarks/performance/a5hixrx2",
   "tg_tokens": null,
   "tps": 89.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-a5hixrx2-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: onyrnzc0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 917.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 917.9 86.2 22.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a5hixrx2",
   "source_url": "https://omlx.ai/benchmarks/performance/a5hixrx2",
   "tg_tokens": null,
   "tps": 86.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-dn6mi8vu-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 6yncmkw1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 906.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 906.6 105.4 21.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dn6mi8vu",
   "source_url": "https://omlx.ai/benchmarks/performance/dn6mi8vu",
   "tg_tokens": null,
   "tps": 105.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-dn6mi8vu-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 6yncmkw1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 954.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 954.0 95.8 22.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dn6mi8vu",
   "source_url": "https://omlx.ai/benchmarks/performance/dn6mi8vu",
   "tg_tokens": null,
   "tps": 95.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-jetdaavv-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: b73b0xuh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 905.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 905.2 96.5 21.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jetdaavv",
   "source_url": "https://omlx.ai/benchmarks/performance/jetdaavv",
   "tg_tokens": null,
   "tps": 96.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-jetdaavv-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: b73b0xuh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 951.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 951.6 98.7 22.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jetdaavv",
   "source_url": "https://omlx.ai/benchmarks/performance/jetdaavv",
   "tg_tokens": null,
   "tps": 98.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-bdksib2t-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 288.6 GB at this context; duplicate traces of the same session: n0i61vdq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 409.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 409.1 28.9 288.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bdksib2t",
   "source_url": "https://omlx.ai/benchmarks/performance/bdksib2t",
   "tg_tokens": null,
   "tps": 28.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-bdksib2t-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 289.3 GB at this context; duplicate traces of the same session: n0i61vdq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 640.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 16k 640.2 34.2 289.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bdksib2t",
   "source_url": "https://omlx.ai/benchmarks/performance/bdksib2t",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bdksib2t-ctx32768",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 289.3 GB at this context; duplicate traces of the same session: n0i61vdq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 676.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 32k 676.4 41.4 289.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bdksib2t",
   "source_url": "https://omlx.ai/benchmarks/performance/bdksib2t",
   "tg_tokens": null,
   "tps": 41.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-bdksib2t-ctx65536",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 289.3 GB at this context; duplicate traces of the same session: n0i61vdq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 692.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 64k 692.2 33.7 289.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bdksib2t",
   "source_url": "https://omlx.ai/benchmarks/performance/bdksib2t",
   "tg_tokens": null,
   "tps": 33.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-bdksib2t-ctx131072",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 289.4 GB at this context; duplicate traces of the same session: n0i61vdq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 667.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 128k 667.1 29.8 289.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bdksib2t",
   "source_url": "https://omlx.ai/benchmarks/performance/bdksib2t",
   "tg_tokens": null,
   "tps": 29.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bdksib2t-ctx199680",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 289.5 GB at this context; duplicate traces of the same session: n0i61vdq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 650.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 195k 650.8 36.6 289.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bdksib2t",
   "source_url": "https://omlx.ai/benchmarks/performance/bdksib2t",
   "tg_tokens": null,
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fhs0z4x1-ctx1024",
   "model": "Muse Glimmer",
   "notes": "oMLX community benchmark trace; peak memory 15.2 GB at this context; duplicate traces of the same session: gp6d7d1d; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 406.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-30B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 406.5 18.2 15.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fhs0z4x1",
   "source_url": "https://omlx.ai/benchmarks/performance/fhs0z4x1",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fhs0z4x1-ctx4096",
   "model": "Muse Glimmer",
   "notes": "oMLX community benchmark trace; peak memory 15.5 GB at this context; duplicate traces of the same session: gp6d7d1d; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 418.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-30B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 418.4 17.8 15.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fhs0z4x1",
   "source_url": "https://omlx.ai/benchmarks/performance/fhs0z4x1",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-1mmow0mi-ctx1024",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: 9bm9opoe, x5d258s7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 259.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 259.2 28.4 15.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1mmow0mi",
   "source_url": "https://omlx.ai/benchmarks/performance/1mmow0mi",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-1mmow0mi-ctx8192",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 16.2 GB at this context; duplicate traces of the same session: 9bm9opoe, x5d258s7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 252.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 252.8 27.7 16.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1mmow0mi",
   "source_url": "https://omlx.ai/benchmarks/performance/1mmow0mi",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-1mmow0mi-ctx65536",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: 9bm9opoe, x5d258s7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 227.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 227.3 24.5 16.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1mmow0mi",
   "source_url": "https://omlx.ai/benchmarks/performance/1mmow0mi",
   "tg_tokens": null,
   "tps": 24.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-1shagcup-ctx1024",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 33.5 GB; TTFT 2157 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 474.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-8bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 1k tokens 474.7 PP tok/s 60.8 TG tok/s 2157 TTFT (ms) 33.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1shagcup",
   "source_url": "https://omlx.ai/benchmarks/performance/1shagcup",
   "tg_tokens": null,
   "tps": 60.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-a14vqm5o-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: y1qvxr0z; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 744.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 744.4 68.8 35.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a14vqm5o",
   "source_url": "https://omlx.ai/benchmarks/performance/a14vqm5o",
   "tg_tokens": null,
   "tps": 68.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-a14vqm5o-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 36.2 GB at this context; duplicate traces of the same session: y1qvxr0z; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1927.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 1,927 67.9 36.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a14vqm5o",
   "source_url": "https://omlx.ai/benchmarks/performance/a14vqm5o",
   "tg_tokens": null,
   "tps": 67.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-s90dm24u-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 903.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 903.4 93.3 21.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s90dm24u",
   "source_url": "https://omlx.ai/benchmarks/performance/s90dm24u",
   "tg_tokens": null,
   "tps": 93.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-s90dm24u-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 952.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 952.6 97.3 22.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s90dm24u",
   "source_url": "https://omlx.ai/benchmarks/performance/s90dm24u",
   "tg_tokens": null,
   "tps": 97.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-wsb96pc3-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 9.8 GB at this context; duplicate traces of the same session: 0fepqnsn; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 643.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-9B-MLX-8bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 8bit; 1k 643.3 15.0 9.8 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wsb96pc3",
   "source_url": "https://omlx.ai/benchmarks/performance/wsb96pc3",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-wsb96pc3-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.3 GB at this context; duplicate traces of the same session: 0fepqnsn; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 718.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-9B-MLX-8bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 8bit; 4k 718.6 14.6 10.3 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wsb96pc3",
   "source_url": "https://omlx.ai/benchmarks/performance/wsb96pc3",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-vj9vk414-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.6 GB at this context; duplicate traces of the same session: q75vxj2m; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 685.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 685.5 27.0 6.6 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vj9vk414",
   "source_url": "https://omlx.ai/benchmarks/performance/vj9vk414",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-vj9vk414-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: q75vxj2m; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 829.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 829.4 26.7 7.0 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vj9vk414",
   "source_url": "https://omlx.ai/benchmarks/performance/vj9vk414",
   "tg_tokens": null,
   "tps": 26.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-egzf0olv-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: gdkgvx8m; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 521.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized-oQ4e-fp16-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 521.6 54.6 22.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/egzf0olv",
   "source_url": "https://omlx.ai/benchmarks/performance/egzf0olv",
   "tg_tokens": null,
   "tps": 54.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-egzf0olv-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: gdkgvx8m; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1418.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized-oQ4e-fp16-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,418 61.0 22.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/egzf0olv",
   "source_url": "https://omlx.ai/benchmarks/performance/egzf0olv",
   "tg_tokens": null,
   "tps": 61.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-6k6j8kod-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "FP16 build; oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: ljse02yx; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1067.0,
   "provenance": "community",
   "quant": "oQ3.5",
   "quote": "Qwen3.6-35B-A3B-oQ3.5-FP16 on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 3bit; 1k 1,067 89.8 17.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6k6j8kod",
   "source_url": "https://omlx.ai/benchmarks/performance/6k6j8kod",
   "tg_tokens": null,
   "tps": 89.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-6k6j8kod-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "FP16 build; oMLX community benchmark trace; peak memory 17.9 GB at this context; duplicate traces of the same session: ljse02yx; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1165.0,
   "provenance": "community",
   "quant": "oQ3.5",
   "quote": "Qwen3.6-35B-A3B-oQ3.5-FP16 on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 3bit; 4k 1,165 83.3 17.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6k6j8kod",
   "source_url": "https://omlx.ai/benchmarks/performance/6k6j8kod",
   "tg_tokens": null,
   "tps": 83.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-2y0yvnmp-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "FP16 build; oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: cofgzml4; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1065.0,
   "provenance": "community",
   "quant": "oQ3.5",
   "quote": "Qwen3.6-35B-A3B-oQ3.5-FP16 on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 3bit; 1k 1,065 89.1 17.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2y0yvnmp",
   "source_url": "https://omlx.ai/benchmarks/performance/2y0yvnmp",
   "tg_tokens": null,
   "tps": 89.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-2y0yvnmp-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "FP16 build; oMLX community benchmark trace; peak memory 17.9 GB at this context; duplicate traces of the same session: cofgzml4; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1165.0,
   "provenance": "community",
   "quant": "oQ3.5",
   "quote": "Qwen3.6-35B-A3B-oQ3.5-FP16 on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 3bit; 4k 1,165 82.9 17.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2y0yvnmp",
   "source_url": "https://omlx.ai/benchmarks/performance/2y0yvnmp",
   "tg_tokens": null,
   "tps": 82.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ldyz00kp-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 0ylwev5t, 4ira31re, c9owb8yi, qiu3n4u1, txhe9izn, ukwuzgrm; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2118.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 2,118 141.4 21.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ldyz00kp",
   "source_url": "https://omlx.ai/benchmarks/performance/ldyz00kp",
   "tg_tokens": null,
   "tps": 141.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ldyz00kp-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: 0ylwev5t, 4ira31re, c9owb8yi, qiu3n4u1, txhe9izn, ukwuzgrm; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2469.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 2,469 121.4 22.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ldyz00kp",
   "source_url": "https://omlx.ai/benchmarks/performance/ldyz00kp",
   "tg_tokens": null,
   "tps": 121.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ldyz00kp-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: 0ylwev5t, 4ira31re, c9owb8yi, qiu3n4u1, txhe9izn, ukwuzgrm; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2564.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 2,564 106.3 23.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ldyz00kp",
   "source_url": "https://omlx.ai/benchmarks/performance/ldyz00kp",
   "tg_tokens": null,
   "tps": 106.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ldyz00kp-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 0ylwev5t, 4ira31re, c9owb8yi, qiu3n4u1, txhe9izn, ukwuzgrm; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2337.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 2,337 139.2 23.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ldyz00kp",
   "source_url": "https://omlx.ai/benchmarks/performance/ldyz00kp",
   "tg_tokens": null,
   "tps": 139.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ldyz00kp-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: 0ylwev5t, 4ira31re, c9owb8yi, qiu3n4u1, txhe9izn, ukwuzgrm; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1929.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 1,929 99.2 24.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ldyz00kp",
   "source_url": "https://omlx.ai/benchmarks/performance/ldyz00kp",
   "tg_tokens": null,
   "tps": 99.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ldyz00kp-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: 0ylwev5t, 4ira31re, c9owb8yi, qiu3n4u1, txhe9izn, ukwuzgrm; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1415.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 1,415 89.3 26.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ldyz00kp",
   "source_url": "https://omlx.ai/benchmarks/performance/ldyz00kp",
   "tg_tokens": null,
   "tps": 89.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ldyz00kp-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.0 GB at this context; duplicate traces of the same session: 0ylwev5t, 4ira31re, c9owb8yi, qiu3n4u1, txhe9izn, ukwuzgrm; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 918.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 128k 918.6 63.3 29.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ldyz00kp",
   "source_url": "https://omlx.ai/benchmarks/performance/ldyz00kp",
   "tg_tokens": null,
   "tps": 63.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-jjqsmzoe-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: ihlrrktw, pspgwmij; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 716.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 716.3 39.7 21.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jjqsmzoe",
   "source_url": "https://omlx.ai/benchmarks/performance/jjqsmzoe",
   "tg_tokens": null,
   "tps": 39.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-jjqsmzoe-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: ihlrrktw, pspgwmij; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 973.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 973.5 57.8 22.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jjqsmzoe",
   "source_url": "https://omlx.ai/benchmarks/performance/jjqsmzoe",
   "tg_tokens": null,
   "tps": 57.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-jjqsmzoe-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: ihlrrktw, pspgwmij; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 962.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 962.5 89.5 22.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jjqsmzoe",
   "source_url": "https://omlx.ai/benchmarks/performance/jjqsmzoe",
   "tg_tokens": null,
   "tps": 89.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-d24hqfh5-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: 754aizhi; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 168.3,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 168.3 32.7 23.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d24hqfh5",
   "source_url": "https://omlx.ai/benchmarks/performance/d24hqfh5",
   "tg_tokens": null,
   "tps": 32.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-d24hqfh5-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: 754aizhi; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 163.8,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 163.8 38.4 24.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d24hqfh5",
   "source_url": "https://omlx.ai/benchmarks/performance/d24hqfh5",
   "tg_tokens": null,
   "tps": 38.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-zsqb2t0e-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: 9p53ut96, gl9a9sv2, sio8hyz4, woo1zdjs; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.4,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 127.4 25.8 22.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zsqb2t0e",
   "source_url": "https://omlx.ai/benchmarks/performance/zsqb2t0e",
   "tg_tokens": null,
   "tps": 25.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-zsqb2t0e-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: 9p53ut96, gl9a9sv2, sio8hyz4, woo1zdjs; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 167.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 167.0 26.9 23.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zsqb2t0e",
   "source_url": "https://omlx.ai/benchmarks/performance/zsqb2t0e",
   "tg_tokens": null,
   "tps": 26.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-zsqb2t0e-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: 9p53ut96, gl9a9sv2, sio8hyz4, woo1zdjs; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 173.6,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 173.6 33.9 23.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zsqb2t0e",
   "source_url": "https://omlx.ai/benchmarks/performance/zsqb2t0e",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-zsqb2t0e-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 9p53ut96, gl9a9sv2, sio8hyz4, woo1zdjs; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 137.3,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 137.3 32.1 23.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zsqb2t0e",
   "source_url": "https://omlx.ai/benchmarks/performance/zsqb2t0e",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-zsqb2t0e-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: 9p53ut96, gl9a9sv2, sio8hyz4, woo1zdjs; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 119.8,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 119.8 61.5 23.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zsqb2t0e",
   "source_url": "https://omlx.ai/benchmarks/performance/zsqb2t0e",
   "tg_tokens": null,
   "tps": 61.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-rifawg2z-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 19.8 GB; TTFT 88845 ms; duplicate traces of the same session: none; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 46.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 4bit; 4k tokens 46.1 PP tok/s 8.8 TG tok/s 88845 TTFT (ms) 19.8 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rifawg2z",
   "source_url": "https://omlx.ai/benchmarks/performance/rifawg2z",
   "tg_tokens": null,
   "tps": 8.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-0d6hfz9a-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: ung13qka; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 53.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 53.7 13.5 18.4 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0d6hfz9a",
   "source_url": "https://omlx.ai/benchmarks/performance/0d6hfz9a",
   "tg_tokens": null,
   "tps": 13.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-0d6hfz9a-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: ung13qka; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 54.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 54.0 11.0 19.6 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0d6hfz9a",
   "source_url": "https://omlx.ai/benchmarks/performance/0d6hfz9a",
   "tg_tokens": null,
   "tps": 11.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-a7p1hb5z-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: h0qxf2lv; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 53.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 53.0 13.3 18.4 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a7p1hb5z",
   "source_url": "https://omlx.ai/benchmarks/performance/a7p1hb5z",
   "tg_tokens": null,
   "tps": 13.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-a7p1hb5z-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: h0qxf2lv; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 52.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 52.9 12.1 19.6 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a7p1hb5z",
   "source_url": "https://omlx.ai/benchmarks/performance/a7p1hb5z",
   "tg_tokens": null,
   "tps": 12.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-qeycgyd6-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 1ghzpi2m; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 53.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 53.2 13.4 18.4 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qeycgyd6",
   "source_url": "https://omlx.ai/benchmarks/performance/qeycgyd6",
   "tg_tokens": null,
   "tps": 13.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-qeycgyd6-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 1ghzpi2m; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 53.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 53.8 11.5 19.6 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qeycgyd6",
   "source_url": "https://omlx.ai/benchmarks/performance/qeycgyd6",
   "tg_tokens": null,
   "tps": 11.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-pj3cgvs1-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 1v09e7wi; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 54.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 54.2 13.4 18.4 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pj3cgvs1",
   "source_url": "https://omlx.ai/benchmarks/performance/pj3cgvs1",
   "tg_tokens": null,
   "tps": 13.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-pj3cgvs1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 1v09e7wi; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 55.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 55.5 13.0 19.6 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pj3cgvs1",
   "source_url": "https://omlx.ai/benchmarks/performance/pj3cgvs1",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ccyr50tg-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1253.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k 1,253 52.2 16.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ccyr50tg",
   "source_url": "https://omlx.ai/benchmarks/performance/ccyr50tg",
   "tg_tokens": null,
   "tps": 52.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ccyr50tg-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1452.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 4k 1,452 50.7 17.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ccyr50tg",
   "source_url": "https://omlx.ai/benchmarks/performance/ccyr50tg",
   "tg_tokens": null,
   "tps": 50.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xd8ldcyw-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: lua036qg; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 494.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 494.8 51.9 70.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xd8ldcyw",
   "source_url": "https://omlx.ai/benchmarks/performance/xd8ldcyw",
   "tg_tokens": null,
   "tps": 51.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xd8ldcyw-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: lua036qg; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 743.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 743.9 52.7 71.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xd8ldcyw",
   "source_url": "https://omlx.ai/benchmarks/performance/xd8ldcyw",
   "tg_tokens": null,
   "tps": 52.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-de7gf8w7-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.7 GB at this context; duplicate traces of the same session: 4kkodh54, ky6bhr0o; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1155.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,155 102.1 100.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/de7gf8w7",
   "source_url": "https://omlx.ai/benchmarks/performance/de7gf8w7",
   "tg_tokens": null,
   "tps": 102.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-de7gf8w7-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: 4kkodh54, ky6bhr0o; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1422.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,422 100.4 101.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/de7gf8w7",
   "source_url": "https://omlx.ai/benchmarks/performance/de7gf8w7",
   "tg_tokens": null,
   "tps": 100.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-de7gf8w7-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.4 GB at this context; duplicate traces of the same session: 4kkodh54, ky6bhr0o; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1371.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 1,371 89.5 103.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/de7gf8w7",
   "source_url": "https://omlx.ai/benchmarks/performance/de7gf8w7",
   "tg_tokens": null,
   "tps": 89.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-0bisr5vl-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: n9hchhm3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 977.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 977.6 18.6 101.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0bisr5vl",
   "source_url": "https://omlx.ai/benchmarks/performance/0bisr5vl",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-0bisr5vl-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.4 GB at this context; duplicate traces of the same session: n9hchhm3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1347.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 1,347 42.0 103.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0bisr5vl",
   "source_url": "https://omlx.ai/benchmarks/performance/0bisr5vl",
   "tg_tokens": null,
   "tps": 42.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-l6gu4o5l-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: 25sqi80t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 873.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 873.8 62.0 70.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l6gu4o5l",
   "source_url": "https://omlx.ai/benchmarks/performance/l6gu4o5l",
   "tg_tokens": null,
   "tps": 62.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-l6gu4o5l-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: 25sqi80t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1625.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,625 71.6 71.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l6gu4o5l",
   "source_url": "https://omlx.ai/benchmarks/performance/l6gu4o5l",
   "tg_tokens": null,
   "tps": 71.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-f0gh4pax-ctx1024",
   "model": "RavenX-CyberAgent-Qwen3.6-35B-A3B-Opus-4.7-OpenMythos-Pentester-BugHunter-RATH",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: 7tax0zky; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2092.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "RavenX-CyberAgent-Qwen3.6-35B-A3B-Opus-4.7-OpenMythos-Pentester-BugHunter-RATH-mlx-mlx-8Bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 2,092 66.5 35.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f0gh4pax",
   "source_url": "https://omlx.ai/benchmarks/performance/f0gh4pax",
   "tg_tokens": null,
   "tps": 66.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-f0gh4pax-ctx4096",
   "model": "RavenX-CyberAgent-Qwen3.6-35B-A3B-Opus-4.7-OpenMythos-Pentester-BugHunter-RATH",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 36.2 GB at this context; duplicate traces of the same session: 7tax0zky; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3104.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "RavenX-CyberAgent-Qwen3.6-35B-A3B-Opus-4.7-OpenMythos-Pentester-BugHunter-RATH-mlx-mlx-8Bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 3,104 65.8 36.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f0gh4pax",
   "source_url": "https://omlx.ai/benchmarks/performance/f0gh4pax",
   "tg_tokens": null,
   "tps": 65.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-11y6g2uz-ctx1024",
   "model": "granite-4.1-3b",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 2.4 GB; TTFT 524 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1955.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "granite-4.1-3b-mxfp4 on M5 (10c); M5 (10c) \u00b7 16 GB \u00b7 4bit; 1k tokens 1,955 PP tok/s 62.8 TG tok/s 524 TTFT (ms) 2.4 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/11y6g2uz",
   "source_url": "https://omlx.ai/benchmarks/performance/11y6g2uz",
   "tg_tokens": null,
   "tps": 62.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lmftetup-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 94.1 GB at this context; duplicate traces of the same session: y8st07c9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 274.7,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 274.7 11.8 94.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lmftetup",
   "source_url": "https://omlx.ai/benchmarks/performance/lmftetup",
   "tg_tokens": null,
   "tps": 11.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lmftetup-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 94.1 GB at this context; duplicate traces of the same session: y8st07c9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 405.7,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 405.7 14.5 94.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lmftetup",
   "source_url": "https://omlx.ai/benchmarks/performance/lmftetup",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kwzbs6re-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 94.1 GB at this context; duplicate traces of the same session: qmzbck1y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 241.9,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 241.9 11.5 94.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kwzbs6re",
   "source_url": "https://omlx.ai/benchmarks/performance/kwzbs6re",
   "tg_tokens": null,
   "tps": 11.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kwzbs6re-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 94.1 GB at this context; duplicate traces of the same session: qmzbck1y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 402.7,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 402.7 15.1 94.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kwzbs6re",
   "source_url": "https://omlx.ai/benchmarks/performance/kwzbs6re",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-w0ajzdqa-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 288.6 GB at this context; duplicate traces of the same session: 2xc29js4, tuyhbd6m, x0uearo7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 409.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 409.1 28.9 288.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w0ajzdqa",
   "source_url": "https://omlx.ai/benchmarks/performance/w0ajzdqa",
   "tg_tokens": null,
   "tps": 28.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-w0ajzdqa-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 289.3 GB at this context; duplicate traces of the same session: 2xc29js4, tuyhbd6m, x0uearo7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 640.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 16k 640.2 34.2 289.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w0ajzdqa",
   "source_url": "https://omlx.ai/benchmarks/performance/w0ajzdqa",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-w0ajzdqa-ctx32768",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 289.3 GB at this context; duplicate traces of the same session: 2xc29js4, tuyhbd6m, x0uearo7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 676.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 32k 676.4 41.4 289.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w0ajzdqa",
   "source_url": "https://omlx.ai/benchmarks/performance/w0ajzdqa",
   "tg_tokens": null,
   "tps": 41.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-w0ajzdqa-ctx65536",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 289.3 GB at this context; duplicate traces of the same session: 2xc29js4, tuyhbd6m, x0uearo7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 692.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 64k 692.2 33.7 289.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w0ajzdqa",
   "source_url": "https://omlx.ai/benchmarks/performance/w0ajzdqa",
   "tg_tokens": null,
   "tps": 33.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-w0ajzdqa-ctx131072",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 289.4 GB at this context; duplicate traces of the same session: 2xc29js4, tuyhbd6m, x0uearo7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 667.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 128k 667.1 29.8 289.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w0ajzdqa",
   "source_url": "https://omlx.ai/benchmarks/performance/w0ajzdqa",
   "tg_tokens": null,
   "tps": 29.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-w0ajzdqa-ctx199680",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 289.5 GB at this context; duplicate traces of the same session: 2xc29js4, tuyhbd6m, x0uearo7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 650.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 195k 650.8 36.6 289.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w0ajzdqa",
   "source_url": "https://omlx.ai/benchmarks/performance/w0ajzdqa",
   "tg_tokens": null,
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-y8zhif0j-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 1k context; peak memory 24.3 GB; TTFT 1361 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 752.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 1k tokens 752.6 PP tok/s 64.4 TG tok/s 1361 TTFT (ms) 24.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y8zhif0j",
   "source_url": "https://omlx.ai/benchmarks/performance/y8zhif0j",
   "tg_tokens": null,
   "tps": 64.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-rl9lrk34-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: u1wl2m20; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 484.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "LFM2.5-24B-A2B-oQ5e-fp16 on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 5bit; 1k 484.7 57.0 16.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rl9lrk34",
   "source_url": "https://omlx.ai/benchmarks/performance/rl9lrk34",
   "tg_tokens": null,
   "tps": 57.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-rl9lrk34-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: u1wl2m20; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 608.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "LFM2.5-24B-A2B-oQ5e-fp16 on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 5bit; 4k 608.9 54.6 16.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rl9lrk34",
   "source_url": "https://omlx.ai/benchmarks/performance/rl9lrk34",
   "tg_tokens": null,
   "tps": 54.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-majgsk95-ctx1024",
   "model": "MiniCPM5-1B",
   "notes": "OptiQ build; oMLX community benchmark trace; peak memory 1.5 GB at this context; duplicate traces of the same session: va9712x6; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2145.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM5-1B-OptiQ-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k 2,145 123.9 1.5 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/majgsk95",
   "source_url": "https://omlx.ai/benchmarks/performance/majgsk95",
   "tg_tokens": null,
   "tps": 123.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-majgsk95-ctx4096",
   "model": "MiniCPM5-1B",
   "notes": "OptiQ build; oMLX community benchmark trace; peak memory 1.6 GB at this context; duplicate traces of the same session: va9712x6; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1879.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM5-1B-OptiQ-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 1,879 106.3 1.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/majgsk95",
   "source_url": "https://omlx.ai/benchmarks/performance/majgsk95",
   "tg_tokens": null,
   "tps": 106.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-7x2xurob-ctx1024",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 5.0 GB at this context; duplicate traces of the same session: 3sv8tspi; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 807.4,
   "provenance": "community",
   "quant": "bf16",
   "quote": "MiniCPM5-2B-bf16 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 bf16; 1k 807.4 21.2 5.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7x2xurob",
   "source_url": "https://omlx.ai/benchmarks/performance/7x2xurob",
   "tg_tokens": null,
   "tps": 21.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-7x2xurob-ctx4096",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 5.3 GB at this context; duplicate traces of the same session: 3sv8tspi; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 763.7,
   "provenance": "community",
   "quant": "bf16",
   "quote": "MiniCPM5-2B-bf16 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 bf16; 4k 763.7 20.0 5.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7x2xurob",
   "source_url": "https://omlx.ai/benchmarks/performance/7x2xurob",
   "tg_tokens": null,
   "tps": 20.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ve2xc83y-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: axopdtzf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2052.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 2,052 67.7 35.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ve2xc83y",
   "source_url": "https://omlx.ai/benchmarks/performance/ve2xc83y",
   "tg_tokens": null,
   "tps": 67.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ve2xc83y-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 36.2 GB at this context; duplicate traces of the same session: axopdtzf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3069.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 3,069 66.3 36.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ve2xc83y",
   "source_url": "https://omlx.ai/benchmarks/performance/ve2xc83y",
   "tg_tokens": null,
   "tps": 66.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-den74txi-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: o9hz1yzv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2604.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,604 104.0 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/den74txi",
   "source_url": "https://omlx.ai/benchmarks/performance/den74txi",
   "tg_tokens": null,
   "tps": 104.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-den74txi-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: o9hz1yzv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3338.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,338 96.7 20.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/den74txi",
   "source_url": "https://omlx.ai/benchmarks/performance/den74txi",
   "tg_tokens": null,
   "tps": 96.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-gtd4s9th-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: yp062gq8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2602.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,602 106.5 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gtd4s9th",
   "source_url": "https://omlx.ai/benchmarks/performance/gtd4s9th",
   "tg_tokens": null,
   "tps": 106.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-gtd4s9th-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: yp062gq8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3345.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,345 103.2 20.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gtd4s9th",
   "source_url": "https://omlx.ai/benchmarks/performance/gtd4s9th",
   "tg_tokens": null,
   "tps": 103.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-v3seqvog-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 4o3i4wzk, 64e7e5sz, k09r0bms, mm4mj37j, q2g5xkjf, qhvhlaow; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2232.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 2,232 63.2 21.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v3seqvog",
   "source_url": "https://omlx.ai/benchmarks/performance/v3seqvog",
   "tg_tokens": null,
   "tps": 63.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-v3seqvog-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: 4o3i4wzk, 64e7e5sz, k09r0bms, mm4mj37j, q2g5xkjf, qhvhlaow; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2684.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 2,684 126.9 23.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v3seqvog",
   "source_url": "https://omlx.ai/benchmarks/performance/v3seqvog",
   "tg_tokens": null,
   "tps": 126.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-v3seqvog-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 4o3i4wzk, 64e7e5sz, k09r0bms, mm4mj37j, q2g5xkjf, qhvhlaow; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2527.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 2,527 45.8 23.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v3seqvog",
   "source_url": "https://omlx.ai/benchmarks/performance/v3seqvog",
   "tg_tokens": null,
   "tps": 45.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-v3seqvog-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 4o3i4wzk, 64e7e5sz, k09r0bms, mm4mj37j, q2g5xkjf, qhvhlaow; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2243.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 2,243 136.7 23.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v3seqvog",
   "source_url": "https://omlx.ai/benchmarks/performance/v3seqvog",
   "tg_tokens": null,
   "tps": 136.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-v3seqvog-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: 4o3i4wzk, 64e7e5sz, k09r0bms, mm4mj37j, q2g5xkjf, qhvhlaow; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1824.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 1,824 190.1 24.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v3seqvog",
   "source_url": "https://omlx.ai/benchmarks/performance/v3seqvog",
   "tg_tokens": null,
   "tps": 190.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-v3seqvog-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: 4o3i4wzk, 64e7e5sz, k09r0bms, mm4mj37j, q2g5xkjf, qhvhlaow; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1325.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 1,325 137.6 26.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v3seqvog",
   "source_url": "https://omlx.ai/benchmarks/performance/v3seqvog",
   "tg_tokens": null,
   "tps": 137.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-v3seqvog-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.8 GB at this context; duplicate traces of the same session: 4o3i4wzk, 64e7e5sz, k09r0bms, mm4mj37j, q2g5xkjf, qhvhlaow; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 854.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 128k 854.7 110.6 29.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v3seqvog",
   "source_url": "https://omlx.ai/benchmarks/performance/v3seqvog",
   "tg_tokens": null,
   "tps": 110.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Pro",
   "id": "omlx-gn0b0u7w-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: x1fjx474; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 1k 102.9 8.9 16.0 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gn0b0u7w",
   "source_url": "https://omlx.ai/benchmarks/performance/gn0b0u7w",
   "tg_tokens": null,
   "tps": 8.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Pro",
   "id": "omlx-gn0b0u7w-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: x1fjx474; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 99.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 4k 99.3 8.7 17.0 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gn0b0u7w",
   "source_url": "https://omlx.ai/benchmarks/performance/gn0b0u7w",
   "tg_tokens": null,
   "tps": 8.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-0vwheqtv-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: 8nzvncx5, nv4q41t1, sxx38ti9; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 121.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k 121.5 3.4 15.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0vwheqtv",
   "source_url": "https://omlx.ai/benchmarks/performance/0vwheqtv",
   "tg_tokens": null,
   "tps": 3.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5",
   "id": "omlx-0vwheqtv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 8nzvncx5, nv4q41t1, sxx38ti9; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 148.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 148.3 3.3 17.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0vwheqtv",
   "source_url": "https://omlx.ai/benchmarks/performance/0vwheqtv",
   "tg_tokens": null,
   "tps": 3.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-0vwheqtv-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 8nzvncx5, nv4q41t1, sxx38ti9; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 152.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 8k 152.4 3.7 17.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0vwheqtv",
   "source_url": "https://omlx.ai/benchmarks/performance/0vwheqtv",
   "tg_tokens": null,
   "tps": 3.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-0vwheqtv-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 8nzvncx5, nv4q41t1, sxx38ti9; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 156.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 16k 156.6 3.9 17.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0vwheqtv",
   "source_url": "https://omlx.ai/benchmarks/performance/0vwheqtv",
   "tg_tokens": null,
   "tps": 3.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-991y7d1y-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.2 GB at this context; duplicate traces of the same session: wmo07s5g; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 141.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 8bit; 4k 141.9 12.1 32.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/991y7d1y",
   "source_url": "https://omlx.ai/benchmarks/performance/991y7d1y",
   "tg_tokens": null,
   "tps": 12.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-991y7d1y-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.2 GB at this context; duplicate traces of the same session: wmo07s5g; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 8bit; 16k 127.0 11.7 34.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/991y7d1y",
   "source_url": "https://omlx.ai/benchmarks/performance/991y7d1y",
   "tg_tokens": null,
   "tps": 11.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-h682plm8-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: c4lc39ic; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 421.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 421.5 16.0 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h682plm8",
   "source_url": "https://omlx.ai/benchmarks/performance/h682plm8",
   "tg_tokens": null,
   "tps": 16.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-h682plm8-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: c4lc39ic; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 471.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 471.5 16.3 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h682plm8",
   "source_url": "https://omlx.ai/benchmarks/performance/h682plm8",
   "tg_tokens": null,
   "tps": 16.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-op7t13ij-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 28.6 GB; TTFT 85514 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 383.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k tokens 383.2 PP tok/s 39.1 TG tok/s 85514 TTFT (ms) 28.6 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/op7t13ij",
   "source_url": "https://omlx.ai/benchmarks/performance/op7t13ij",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-vk1r0e9p-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 28.6 GB; TTFT 85588 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 382.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k tokens 382.9 PP tok/s 59.9 TG tok/s 85588 TTFT (ms) 28.6 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vk1r0e9p",
   "source_url": "https://omlx.ai/benchmarks/performance/vk1r0e9p",
   "tg_tokens": null,
   "tps": 59.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-43igyyfd-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 325.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 325.6 64.6 21.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/43igyyfd",
   "source_url": "https://omlx.ai/benchmarks/performance/43igyyfd",
   "tg_tokens": null,
   "tps": 64.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-43igyyfd-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 378.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 378.4 59.7 28.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/43igyyfd",
   "source_url": "https://omlx.ai/benchmarks/performance/43igyyfd",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-lduepp27-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: 147o4wkp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 275.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 275.3 31.1 24.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lduepp27",
   "source_url": "https://omlx.ai/benchmarks/performance/lduepp27",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-lduepp27-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: 147o4wkp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 247.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 247.3 47.1 24.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lduepp27",
   "source_url": "https://omlx.ai/benchmarks/performance/lduepp27",
   "tg_tokens": null,
   "tps": 47.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-0ps5c8c0-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 3dg7ecyg, yq6ce4cr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 254.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 254.5 63.5 18.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ps5c8c0",
   "source_url": "https://omlx.ai/benchmarks/performance/0ps5c8c0",
   "tg_tokens": null,
   "tps": 63.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-0ps5c8c0-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 3dg7ecyg, yq6ce4cr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 264.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 264.9 45.4 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ps5c8c0",
   "source_url": "https://omlx.ai/benchmarks/performance/0ps5c8c0",
   "tg_tokens": null,
   "tps": 45.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-0ps5c8c0-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.6 GB at this context; duplicate traces of the same session: 3dg7ecyg, yq6ce4cr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 205.6 40.9 27.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ps5c8c0",
   "source_url": "https://omlx.ai/benchmarks/performance/0ps5c8c0",
   "tg_tokens": null,
   "tps": 40.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-rvgt1lk3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: 0gyc18c4; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 364.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 1k 364.6 28.2 26.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rvgt1lk3",
   "source_url": "https://omlx.ai/benchmarks/performance/rvgt1lk3",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-rvgt1lk3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: 0gyc18c4; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 386.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 386.6 28.2 27.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rvgt1lk3",
   "source_url": "https://omlx.ai/benchmarks/performance/rvgt1lk3",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-269boskq-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: 5gmgo3j2, 5uan1qq6, edzu1zuc, iq72ka8g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 480.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 480.2 51.3 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/269boskq",
   "source_url": "https://omlx.ai/benchmarks/performance/269boskq",
   "tg_tokens": null,
   "tps": 51.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-269boskq-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: 5gmgo3j2, 5uan1qq6, edzu1zuc, iq72ka8g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 652.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 652.0 62.0 71.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/269boskq",
   "source_url": "https://omlx.ai/benchmarks/performance/269boskq",
   "tg_tokens": null,
   "tps": 62.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-269boskq-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: 5gmgo3j2, 5uan1qq6, edzu1zuc, iq72ka8g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 817.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 817.6 74.4 71.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/269boskq",
   "source_url": "https://omlx.ai/benchmarks/performance/269boskq",
   "tg_tokens": null,
   "tps": 74.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-269boskq-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 5gmgo3j2, 5uan1qq6, edzu1zuc, iq72ka8g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 813.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 813.8 57.8 71.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/269boskq",
   "source_url": "https://omlx.ai/benchmarks/performance/269boskq",
   "tg_tokens": null,
   "tps": 57.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-269boskq-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.4 GB at this context; duplicate traces of the same session: 5gmgo3j2, 5uan1qq6, edzu1zuc, iq72ka8g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 810.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 810.9 52.1 72.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/269boskq",
   "source_url": "https://omlx.ai/benchmarks/performance/269boskq",
   "tg_tokens": null,
   "tps": 52.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-6i394kf8-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: f7lot4ju; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 267.1,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "gemma-4-26B-A4B-it-oQ2.5e-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 16 GB \u00b7 2bit; 1k 267.1 22.3 6.2 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6i394kf8",
   "source_url": "https://omlx.ai/benchmarks/performance/6i394kf8",
   "tg_tokens": null,
   "tps": 22.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-6i394kf8-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: f7lot4ju; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 304.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "gemma-4-26B-A4B-it-oQ2.5e-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 16 GB \u00b7 2bit; 4k 304.0 23.6 6.5 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6i394kf8",
   "source_url": "https://omlx.ai/benchmarks/performance/6i394kf8",
   "tg_tokens": null,
   "tps": 23.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-9nz1i8ov-ctx1024",
   "model": "granite-4.1-3b",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 2.4 GB; TTFT 523 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1958.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "granite-4.1-3b-mxfp4 on M5 (10c); M5 (10c) \u00b7 16 GB \u00b7 4bit; 1k tokens 1,958 PP tok/s 61.2 TG tok/s 523 TTFT (ms) 2.4 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9nz1i8ov",
   "source_url": "https://omlx.ai/benchmarks/performance/9nz1i8ov",
   "tg_tokens": null,
   "tps": 61.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-04oy6jaj-ctx1024",
   "model": "granite-4.1-3b",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 2.4 GB; TTFT 535 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1913.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "granite-4.1-3b-mxfp4 on M5 (10c); M5 (10c) \u00b7 16 GB \u00b7 4bit; 1k tokens 1,913 PP tok/s 60.9 TG tok/s 535 TTFT (ms) 2.4 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/04oy6jaj",
   "source_url": "https://omlx.ai/benchmarks/performance/04oy6jaj",
   "tg_tokens": null,
   "tps": 60.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zvnzmmb7-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.0 GB at this context; duplicate traces of the same session: shh524pu, y1xhigfw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 77.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 77.0 11.5 92.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zvnzmmb7",
   "source_url": "https://omlx.ai/benchmarks/performance/zvnzmmb7",
   "tg_tokens": null,
   "tps": 11.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zvnzmmb7-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 93.9 GB at this context; duplicate traces of the same session: shh524pu, y1xhigfw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 311.8,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 311.8 11.5 93.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zvnzmmb7",
   "source_url": "https://omlx.ai/benchmarks/performance/zvnzmmb7",
   "tg_tokens": null,
   "tps": 11.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zvnzmmb7-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 93.8 GB at this context; duplicate traces of the same session: shh524pu, y1xhigfw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 372.1,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 372.1 12.4 93.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zvnzmmb7",
   "source_url": "https://omlx.ai/benchmarks/performance/zvnzmmb7",
   "tg_tokens": null,
   "tps": 12.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-1r2szcqv-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 671.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 1k 671.5 63.4 24.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1r2szcqv",
   "source_url": "https://omlx.ai/benchmarks/performance/1r2szcqv",
   "tg_tokens": null,
   "tps": 63.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-1r2szcqv-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 676.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 32k 676.2 43.7 25.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1r2szcqv",
   "source_url": "https://omlx.ai/benchmarks/performance/1r2szcqv",
   "tg_tokens": null,
   "tps": 43.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-u9093agz-ctx1024",
   "model": "Ling-3.0-tiny",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 8.7 GB at this context; duplicate traces of the same session: agoihirs; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1854.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ling-3.0-tiny-oQ8-fp16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 1k 1,854 84.3 8.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u9093agz",
   "source_url": "https://omlx.ai/benchmarks/performance/u9093agz",
   "tg_tokens": null,
   "tps": 84.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-u9093agz-ctx65536",
   "model": "Ling-3.0-tiny",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 13.8 GB at this context; duplicate traces of the same session: agoihirs; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1154.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ling-3.0-tiny-oQ8-fp16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 64k 1,154 56.9 13.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u9093agz",
   "source_url": "https://omlx.ai/benchmarks/performance/u9093agz",
   "tg_tokens": null,
   "tps": 56.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-7059l6yy-ctx1024",
   "model": "Ling-3.0-tiny",
   "notes": "fp16 build; oMLX community benchmark trace, single run at 1k context; peak memory 8.7 GB; TTFT 669 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1531.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ling-3.0-tiny-oQ8-fp16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 1k tokens 1,531 PP tok/s 89.0 TG tok/s 669 TTFT (ms) 8.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7059l6yy",
   "source_url": "https://omlx.ai/benchmarks/performance/7059l6yy",
   "tg_tokens": null,
   "tps": 89.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-50iu82r4-ctx1024",
   "model": "Muse-Glimmer-30B-heretic",
   "notes": "oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: m34kc0i9; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-heretic-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 130.0 14.9 15.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/50iu82r4",
   "source_url": "https://omlx.ai/benchmarks/performance/50iu82r4",
   "tg_tokens": null,
   "tps": 14.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-50iu82r4-ctx4096",
   "model": "Muse-Glimmer-30B-heretic",
   "notes": "oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: m34kc0i9; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 128.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-heretic-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 128.2 14.5 16.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/50iu82r4",
   "source_url": "https://omlx.ai/benchmarks/performance/50iu82r4",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-lqkl047n-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 722.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 722.0 86.5 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lqkl047n",
   "source_url": "https://omlx.ai/benchmarks/performance/lqkl047n",
   "tg_tokens": null,
   "tps": 86.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-lqkl047n-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 918.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 918.0 83.6 21.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lqkl047n",
   "source_url": "https://omlx.ai/benchmarks/performance/lqkl047n",
   "tg_tokens": null,
   "tps": 83.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-lqkl047n-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 609.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 32k 609.4 62.9 22.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lqkl047n",
   "source_url": "https://omlx.ai/benchmarks/performance/lqkl047n",
   "tg_tokens": null,
   "tps": 62.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-4ncq5tip-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: 9qesbh8w; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2351.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 2,351 94.4 3.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ncq5tip",
   "source_url": "https://omlx.ai/benchmarks/performance/4ncq5tip",
   "tg_tokens": null,
   "tps": 94.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-4ncq5tip-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: 9qesbh8w; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2313.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,313 90.6 3.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ncq5tip",
   "source_url": "https://omlx.ai/benchmarks/performance/4ncq5tip",
   "tg_tokens": null,
   "tps": 90.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-mick8rc9-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: 3y91knqg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1409.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 1,409 57.0 5.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mick8rc9",
   "source_url": "https://omlx.ai/benchmarks/performance/mick8rc9",
   "tg_tokens": null,
   "tps": 57.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-mick8rc9-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: 3y91knqg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1491.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 1,491 55.4 6.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mick8rc9",
   "source_url": "https://omlx.ai/benchmarks/performance/mick8rc9",
   "tg_tokens": null,
   "tps": 55.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-hwb16tpg-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "bf16 pt MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: 18jo0cg3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1984.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-bf16-pt-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 1,984 168.8 37.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hwb16tpg",
   "source_url": "https://omlx.ai/benchmarks/performance/hwb16tpg",
   "tg_tokens": null,
   "tps": 168.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-hwb16tpg-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "bf16 pt MTP build; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: 18jo0cg3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2467.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-bf16-pt-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 2,467 138.5 39.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hwb16tpg",
   "source_url": "https://omlx.ai/benchmarks/performance/hwb16tpg",
   "tg_tokens": null,
   "tps": 138.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-5wsxsn62-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1467.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 1,467 68.8 28.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5wsxsn62",
   "source_url": "https://omlx.ai/benchmarks/performance/5wsxsn62",
   "tg_tokens": null,
   "tps": 68.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-5wsxsn62-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 534.8,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k 534.8 72.2 33.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5wsxsn62",
   "source_url": "https://omlx.ai/benchmarks/performance/5wsxsn62",
   "tg_tokens": null,
   "tps": 72.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xg5xy1h5-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 33.3 GB; TTFT 18466 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1775.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 32k tokens 1,775 PP tok/s 186.4 TG tok/s 18466 TTFT (ms) 33.3 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xg5xy1h5",
   "source_url": "https://omlx.ai/benchmarks/performance/xg5xy1h5",
   "tg_tokens": null,
   "tps": 186.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-cxa0ylcj-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 33.3 GB; TTFT 18463 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1775.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 32k tokens 1,775 PP tok/s 121.2 TG tok/s 18463 TTFT (ms) 33.3 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cxa0ylcj",
   "source_url": "https://omlx.ai/benchmarks/performance/cxa0ylcj",
   "tg_tokens": null,
   "tps": 121.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2lwdlkon-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: 63sekw0q; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2161.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 1k 2,161 162.2 29.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2lwdlkon",
   "source_url": "https://omlx.ai/benchmarks/performance/2lwdlkon",
   "tg_tokens": null,
   "tps": 162.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2lwdlkon-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.3 GB at this context; duplicate traces of the same session: 63sekw0q; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1777.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 32k 1,777 141.2 33.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2lwdlkon",
   "source_url": "https://omlx.ai/benchmarks/performance/2lwdlkon",
   "tg_tokens": null,
   "tps": 141.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-97hbtmf9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: 5q41g1m2; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 1k 102.3 16.3 17.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/97hbtmf9",
   "source_url": "https://omlx.ai/benchmarks/performance/97hbtmf9",
   "tg_tokens": null,
   "tps": 16.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Pro",
   "id": "omlx-97hbtmf9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: 5q41g1m2; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 4k 103.9 16.4 18.7 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/97hbtmf9",
   "source_url": "https://omlx.ai/benchmarks/performance/97hbtmf9",
   "tg_tokens": null,
   "tps": 16.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ux6452cq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: opqsqncw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1069.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,069 33.8 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ux6452cq",
   "source_url": "https://omlx.ai/benchmarks/performance/ux6452cq",
   "tg_tokens": null,
   "tps": 33.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ux6452cq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: opqsqncw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1444.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,444 33.1 18.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ux6452cq",
   "source_url": "https://omlx.ai/benchmarks/performance/ux6452cq",
   "tg_tokens": null,
   "tps": 33.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-p77d929g-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: eucky1yk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1162.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,162 33.7 17.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p77d929g",
   "source_url": "https://omlx.ai/benchmarks/performance/p77d929g",
   "tg_tokens": null,
   "tps": 33.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-p77d929g-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: eucky1yk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1448.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,448 33.1 18.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p77d929g",
   "source_url": "https://omlx.ai/benchmarks/performance/p77d929g",
   "tg_tokens": null,
   "tps": 33.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lr876f4r-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 2s56vq9h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 843.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 843.7 33.4 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lr876f4r",
   "source_url": "https://omlx.ai/benchmarks/performance/lr876f4r",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lr876f4r-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 2s56vq9h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 982.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 982.6 33.1 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lr876f4r",
   "source_url": "https://omlx.ai/benchmarks/performance/lr876f4r",
   "tg_tokens": null,
   "tps": 33.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-8l67srzd-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: vvv1yqp0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 352.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 8bit; 1k 352.9 10.0 28.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8l67srzd",
   "source_url": "https://omlx.ai/benchmarks/performance/8l67srzd",
   "tg_tokens": null,
   "tps": 10.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-8l67srzd-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: vvv1yqp0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 388.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 8bit; 4k 388.1 9.9 29.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8l67srzd",
   "source_url": "https://omlx.ai/benchmarks/performance/8l67srzd",
   "tg_tokens": null,
   "tps": 9.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-7he2lu7e-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 325.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 325.6 64.6 21.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7he2lu7e",
   "source_url": "https://omlx.ai/benchmarks/performance/7he2lu7e",
   "tg_tokens": null,
   "tps": 64.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-7he2lu7e-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 378.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 378.4 59.7 28.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7he2lu7e",
   "source_url": "https://omlx.ai/benchmarks/performance/7he2lu7e",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-e083kav4-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: wl83krlr; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 197.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k 197.3 15.5 17.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e083kav4",
   "source_url": "https://omlx.ai/benchmarks/performance/e083kav4",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-e083kav4-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: wl83krlr; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 200.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 200.7 16.2 18.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e083kav4",
   "source_url": "https://omlx.ai/benchmarks/performance/e083kav4",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-12eqoubr-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: p5z5wtnn, x5mv1s3s; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 503.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 503.8 27.9 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/12eqoubr",
   "source_url": "https://omlx.ai/benchmarks/performance/12eqoubr",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-12eqoubr-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: p5z5wtnn, x5mv1s3s; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 610.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 610.5 46.8 19.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/12eqoubr",
   "source_url": "https://omlx.ai/benchmarks/performance/12eqoubr",
   "tg_tokens": null,
   "tps": 46.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-12eqoubr-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: p5z5wtnn, x5mv1s3s; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 617.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 617.0 22.6 19.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/12eqoubr",
   "source_url": "https://omlx.ai/benchmarks/performance/12eqoubr",
   "tg_tokens": null,
   "tps": 22.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-w1ph5pzq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 3ovzbovg, zfob67el; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 466.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 466.7 39.5 17.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w1ph5pzq",
   "source_url": "https://omlx.ai/benchmarks/performance/w1ph5pzq",
   "tg_tokens": null,
   "tps": 39.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-w1ph5pzq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: 3ovzbovg, zfob67el; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 470.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 470.9 39.2 18.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w1ph5pzq",
   "source_url": "https://omlx.ai/benchmarks/performance/w1ph5pzq",
   "tg_tokens": null,
   "tps": 39.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-w1ph5pzq-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 3ovzbovg, zfob67el; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 438.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 438.5 19.8 19.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w1ph5pzq",
   "source_url": "https://omlx.ai/benchmarks/performance/w1ph5pzq",
   "tg_tokens": null,
   "tps": 19.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-b4x8ma6j-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 58n66fjg, bmyp5lpg; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 407.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 407.5 36.5 17.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b4x8ma6j",
   "source_url": "https://omlx.ai/benchmarks/performance/b4x8ma6j",
   "tg_tokens": null,
   "tps": 36.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-b4x8ma6j-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: 58n66fjg, bmyp5lpg; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 458.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 458.3 45.0 18.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b4x8ma6j",
   "source_url": "https://omlx.ai/benchmarks/performance/b4x8ma6j",
   "tg_tokens": null,
   "tps": 45.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-b4x8ma6j-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 58n66fjg, bmyp5lpg; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 461.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 461.0 27.0 19.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b4x8ma6j",
   "source_url": "https://omlx.ai/benchmarks/performance/b4x8ma6j",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-40qg9rdp-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 50.0 GB at this context; duplicate traces of the same session: b6phfxht; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 76.8,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 2bit; 1k 76.8 16.7 50.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/40qg9rdp",
   "source_url": "https://omlx.ai/benchmarks/performance/40qg9rdp",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-40qg9rdp-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 49.5 GB at this context; duplicate traces of the same session: b6phfxht; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 182.4,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 2bit; 4k 182.4 7.7 49.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/40qg9rdp",
   "source_url": "https://omlx.ai/benchmarks/performance/40qg9rdp",
   "tg_tokens": null,
   "tps": 7.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-uivxcupj-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: ku7fsi82; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.7,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 2bit; 1k 102.7 14.4 20.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uivxcupj",
   "source_url": "https://omlx.ai/benchmarks/performance/uivxcupj",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-uivxcupj-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: ku7fsi82; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 242.0,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 2bit; 4k 242.0 14.4 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uivxcupj",
   "source_url": "https://omlx.ai/benchmarks/performance/uivxcupj",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-rvcqr7e7-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 41.5 GB at this context; duplicate traces of the same session: 5xyov0zw; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 91.6,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 2bit; 1k 91.6 18.7 41.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rvcqr7e7",
   "source_url": "https://omlx.ai/benchmarks/performance/rvcqr7e7",
   "tg_tokens": null,
   "tps": 18.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-rvcqr7e7-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 42.3 GB at this context; duplicate traces of the same session: 5xyov0zw; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 296.8,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 2bit; 4k 296.8 21.3 42.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rvcqr7e7",
   "source_url": "https://omlx.ai/benchmarks/performance/rvcqr7e7",
   "tg_tokens": null,
   "tps": 21.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-9adjgj55-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 41.5 GB at this context; duplicate traces of the same session: qs50mgwl; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 90.7,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 2bit; 1k 90.7 20.2 41.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9adjgj55",
   "source_url": "https://omlx.ai/benchmarks/performance/9adjgj55",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-9adjgj55-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 42.3 GB at this context; duplicate traces of the same session: qs50mgwl; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 265.6,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 2bit; 4k 265.6 21.7 42.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9adjgj55",
   "source_url": "https://omlx.ai/benchmarks/performance/9adjgj55",
   "tg_tokens": null,
   "tps": 21.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-a5kxpj5l-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 28.7 GB; TTFT 15634 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 262.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k tokens 262.0 PP tok/s 12.0 TG tok/s 15634 TTFT (ms) 28.7 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a5kxpj5l",
   "source_url": "https://omlx.ai/benchmarks/performance/a5kxpj5l",
   "tg_tokens": null,
   "tps": 12.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-nkordxu9-ctx4096",
   "model": "Qwopus3.8-27B-Flash-V2",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.3 GB at this context; duplicate traces of the same session: 0j5hme9x; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 226.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwopus3.8-27B-Flash-V2-oQ6e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 6bit; 4k 226.7 22.1 35.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nkordxu9",
   "source_url": "https://omlx.ai/benchmarks/performance/nkordxu9",
   "tg_tokens": null,
   "tps": 22.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-nkordxu9-ctx16384",
   "model": "Qwopus3.8-27B-Flash-V2",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: 0j5hme9x; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 211.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwopus3.8-27B-Flash-V2-oQ6e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 6bit; 16k 211.6 29.9 36.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nkordxu9",
   "source_url": "https://omlx.ai/benchmarks/performance/nkordxu9",
   "tg_tokens": null,
   "tps": 29.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-3pwe81g8-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.7,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 127.7 14.6 33.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3pwe81g8",
   "source_url": "https://omlx.ai/benchmarks/performance/3pwe81g8",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-3pwe81g8-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 126.3,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 126.3 12.6 37.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3pwe81g8",
   "source_url": "https://omlx.ai/benchmarks/performance/3pwe81g8",
   "tg_tokens": null,
   "tps": 12.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-sdod1c7b-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift build; oQ4e/oQ6e mixed quant; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 77.5 GB at this context; duplicate traces of the same session: 5jpc133k, eepulavx, n56rzuq6, u7c7uo0l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 584.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 584.2 53.2 77.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sdod1c7b",
   "source_url": "https://omlx.ai/benchmarks/performance/sdod1c7b",
   "tg_tokens": null,
   "tps": 53.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-sdod1c7b-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift build; oQ4e/oQ6e mixed quant; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.5 GB at this context; duplicate traces of the same session: 5jpc133k, eepulavx, n56rzuq6, u7c7uo0l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 826.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 826.2 53.4 78.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sdod1c7b",
   "source_url": "https://omlx.ai/benchmarks/performance/sdod1c7b",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-sdod1c7b-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift build; oQ4e/oQ6e mixed quant; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.8 GB at this context; duplicate traces of the same session: 5jpc133k, eepulavx, n56rzuq6, u7c7uo0l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 824.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 824.3 81.1 78.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sdod1c7b",
   "source_url": "https://omlx.ai/benchmarks/performance/sdod1c7b",
   "tg_tokens": null,
   "tps": 81.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-sdod1c7b-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift build; oQ4e/oQ6e mixed quant; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.4 GB at this context; duplicate traces of the same session: 5jpc133k, eepulavx, n56rzuq6, u7c7uo0l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 817.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 817.6 58.3 79.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sdod1c7b",
   "source_url": "https://omlx.ai/benchmarks/performance/sdod1c7b",
   "tg_tokens": null,
   "tps": 58.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-sdod1c7b-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Swift build; oQ4e/oQ6e mixed quant; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.5 GB at this context; duplicate traces of the same session: 5jpc133k, eepulavx, n56rzuq6, u7c7uo0l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 807.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k 807.2 57.1 80.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sdod1c7b",
   "source_url": "https://omlx.ai/benchmarks/performance/sdod1c7b",
   "tg_tokens": null,
   "tps": 57.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-15940qq5-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 94.0 GB; TTFT 22035 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 371.8,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k tokens 371.8 PP tok/s 13.6 TG tok/s 22035 TTFT (ms) 94.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/15940qq5",
   "source_url": "https://omlx.ai/benchmarks/performance/15940qq5",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-rlwxdozo-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 671.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 1k 671.5 63.4 24.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rlwxdozo",
   "source_url": "https://omlx.ai/benchmarks/performance/rlwxdozo",
   "tg_tokens": null,
   "tps": 63.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-rlwxdozo-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 676.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 32k 676.2 43.7 25.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rlwxdozo",
   "source_url": "https://omlx.ai/benchmarks/performance/rlwxdozo",
   "tg_tokens": null,
   "tps": 43.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-xtkqcorz-ctx1024",
   "model": "muse-glimmer 30B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 194.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 194.4 20.8 15.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xtkqcorz",
   "source_url": "https://omlx.ai/benchmarks/performance/xtkqcorz",
   "tg_tokens": null,
   "tps": 20.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-xtkqcorz-ctx4096",
   "model": "muse-glimmer 30B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 189.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 189.2 20.4 16.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xtkqcorz",
   "source_url": "https://omlx.ai/benchmarks/performance/xtkqcorz",
   "tg_tokens": null,
   "tps": 20.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-xtkqcorz-ctx8192",
   "model": "muse-glimmer 30B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 16.2 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 171.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 171.0 20.2 16.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xtkqcorz",
   "source_url": "https://omlx.ai/benchmarks/performance/xtkqcorz",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-xtkqcorz-ctx16384",
   "model": "muse-glimmer 30B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 184.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 184.5 20.3 16.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xtkqcorz",
   "source_url": "https://omlx.ai/benchmarks/performance/xtkqcorz",
   "tg_tokens": null,
   "tps": 20.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-6l1xfu1j-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: tsju0tcj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 722.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 722.0 86.5 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6l1xfu1j",
   "source_url": "https://omlx.ai/benchmarks/performance/6l1xfu1j",
   "tg_tokens": null,
   "tps": 86.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-6l1xfu1j-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: tsju0tcj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 918.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 918.0 83.6 21.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6l1xfu1j",
   "source_url": "https://omlx.ai/benchmarks/performance/6l1xfu1j",
   "tg_tokens": null,
   "tps": 83.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-6l1xfu1j-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: tsju0tcj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 609.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 32k 609.4 62.9 22.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6l1xfu1j",
   "source_url": "https://omlx.ai/benchmarks/performance/6l1xfu1j",
   "tg_tokens": null,
   "tps": 62.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-tje3b1mb-ctx4096",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: pz3t2y02; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2219.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 2,219 75.1 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tje3b1mb",
   "source_url": "https://omlx.ai/benchmarks/performance/tje3b1mb",
   "tg_tokens": null,
   "tps": 75.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-tje3b1mb-ctx16384",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: pz3t2y02; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2129.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 16k 2,129 65.9 20.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tje3b1mb",
   "source_url": "https://omlx.ai/benchmarks/performance/tje3b1mb",
   "tg_tokens": null,
   "tps": 65.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-kam999nk-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "bf16 pt MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: 9d3jy3a8; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1966.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-bf16-pt-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 1,966 168.5 37.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kam999nk",
   "source_url": "https://omlx.ai/benchmarks/performance/kam999nk",
   "tg_tokens": null,
   "tps": 168.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-kam999nk-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "bf16 pt MTP build; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: 9d3jy3a8; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2465.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-bf16-pt-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 2,465 128.8 39.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kam999nk",
   "source_url": "https://omlx.ai/benchmarks/performance/kam999nk",
   "tg_tokens": null,
   "tps": 128.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-mwicv2rt-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "bf16 pt MTP build; oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: 1bxw3b6j; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2071.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-bf16-pt-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 2,071 90.2 35.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mwicv2rt",
   "source_url": "https://omlx.ai/benchmarks/performance/mwicv2rt",
   "tg_tokens": null,
   "tps": 90.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-mwicv2rt-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "bf16 pt MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: 1bxw3b6j; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2669.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-bf16-pt-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 2,669 85.2 37.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mwicv2rt",
   "source_url": "https://omlx.ai/benchmarks/performance/mwicv2rt",
   "tg_tokens": null,
   "tps": 85.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-51pz5wv5-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1467.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 1,467 68.8 28.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/51pz5wv5",
   "source_url": "https://omlx.ai/benchmarks/performance/51pz5wv5",
   "tg_tokens": null,
   "tps": 68.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-51pz5wv5-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 534.8,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k 534.8 72.2 33.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/51pz5wv5",
   "source_url": "https://omlx.ai/benchmarks/performance/51pz5wv5",
   "tg_tokens": null,
   "tps": 72.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-cfhmvvjz-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: byfnb29f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1737.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 1,737 125.4 28.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cfhmvvjz",
   "source_url": "https://omlx.ai/benchmarks/performance/cfhmvvjz",
   "tg_tokens": null,
   "tps": 125.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-cfhmvvjz-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.6 GB at this context; duplicate traces of the same session: byfnb29f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 799.3,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k 799.3 62.0 33.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cfhmvvjz",
   "source_url": "https://omlx.ai/benchmarks/performance/cfhmvvjz",
   "tg_tokens": null,
   "tps": 62.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-3ke1tpds-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.9 GB at this context; duplicate traces of the same session: w3d3z9o9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1616.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 1,616 109.3 28.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ke1tpds",
   "source_url": "https://omlx.ai/benchmarks/performance/3ke1tpds",
   "tg_tokens": null,
   "tps": 109.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-3ke1tpds-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: w3d3z9o9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 777.3,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k 777.3 63.4 33.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ke1tpds",
   "source_url": "https://omlx.ai/benchmarks/performance/3ke1tpds",
   "tg_tokens": null,
   "tps": 63.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-6qfiavf4-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: r63hnei2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1706.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 1,706 92.4 27.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6qfiavf4",
   "source_url": "https://omlx.ai/benchmarks/performance/6qfiavf4",
   "tg_tokens": null,
   "tps": 92.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-6qfiavf4-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.7 GB at this context; duplicate traces of the same session: r63hnei2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 893.5,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k 893.5 70.1 32.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6qfiavf4",
   "source_url": "https://omlx.ai/benchmarks/performance/6qfiavf4",
   "tg_tokens": null,
   "tps": 70.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-wh4kfj9a-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: z3ael79e; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 1k 102.2 15.8 17.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wh4kfj9a",
   "source_url": "https://omlx.ai/benchmarks/performance/wh4kfj9a",
   "tg_tokens": null,
   "tps": 15.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Pro",
   "id": "omlx-wh4kfj9a-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: z3ael79e; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 4k 103.9 16.6 18.7 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wh4kfj9a",
   "source_url": "https://omlx.ai/benchmarks/performance/wh4kfj9a",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-z94rje02-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 47feb84b, 4l03qjbk, b27uit5b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 532.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 532.1 33.3 19.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z94rje02",
   "source_url": "https://omlx.ai/benchmarks/performance/z94rje02",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-z94rje02-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: 47feb84b, 4l03qjbk, b27uit5b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 628.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 628.4 41.5 20.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z94rje02",
   "source_url": "https://omlx.ai/benchmarks/performance/z94rje02",
   "tg_tokens": null,
   "tps": 41.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-z94rje02-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: 47feb84b, 4l03qjbk, b27uit5b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 679.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 679.3 53.8 21.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z94rje02",
   "source_url": "https://omlx.ai/benchmarks/performance/z94rje02",
   "tg_tokens": null,
   "tps": 53.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-z94rje02-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 47feb84b, 4l03qjbk, b27uit5b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 680.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 680.6 64.2 21.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z94rje02",
   "source_url": "https://omlx.ai/benchmarks/performance/z94rje02",
   "tg_tokens": null,
   "tps": 64.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-839py3xk-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: f0js1vqb, qu8hkk61, zrctpzb1, zxld9toy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 892.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 892.5 31.4 16.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/839py3xk",
   "source_url": "https://omlx.ai/benchmarks/performance/839py3xk",
   "tg_tokens": null,
   "tps": 31.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-839py3xk-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: f0js1vqb, qu8hkk61, zrctpzb1, zxld9toy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 946.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 946.7 31.0 17.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/839py3xk",
   "source_url": "https://omlx.ai/benchmarks/performance/839py3xk",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-839py3xk-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: f0js1vqb, qu8hkk61, zrctpzb1, zxld9toy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 913.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 913.1 30.8 17.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/839py3xk",
   "source_url": "https://omlx.ai/benchmarks/performance/839py3xk",
   "tg_tokens": null,
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-839py3xk-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: f0js1vqb, qu8hkk61, zrctpzb1, zxld9toy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 815.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 815.0 30.3 17.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/839py3xk",
   "source_url": "https://omlx.ai/benchmarks/performance/839py3xk",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-839py3xk-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: f0js1vqb, qu8hkk61, zrctpzb1, zxld9toy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 695.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 695.7 27.6 18.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/839py3xk",
   "source_url": "https://omlx.ai/benchmarks/performance/839py3xk",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2i890c8n-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 0oyj6pv8, cz17wxdk, t7fmjgc7, z2ieaupj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 892.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 892.3 46.8 19.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2i890c8n",
   "source_url": "https://omlx.ai/benchmarks/performance/2i890c8n",
   "tg_tokens": null,
   "tps": 46.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2i890c8n-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: 0oyj6pv8, cz17wxdk, t7fmjgc7, z2ieaupj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 929.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 929.9 54.9 20.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2i890c8n",
   "source_url": "https://omlx.ai/benchmarks/performance/2i890c8n",
   "tg_tokens": null,
   "tps": 54.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2i890c8n-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: 0oyj6pv8, cz17wxdk, t7fmjgc7, z2ieaupj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 864.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 864.3 48.7 21.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2i890c8n",
   "source_url": "https://omlx.ai/benchmarks/performance/2i890c8n",
   "tg_tokens": null,
   "tps": 48.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2i890c8n-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 0oyj6pv8, cz17wxdk, t7fmjgc7, z2ieaupj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 798.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 798.7 65.1 21.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2i890c8n",
   "source_url": "https://omlx.ai/benchmarks/performance/2i890c8n",
   "tg_tokens": null,
   "tps": 65.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2i890c8n-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: 0oyj6pv8, cz17wxdk, t7fmjgc7, z2ieaupj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 683.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 683.3 39.7 22.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2i890c8n",
   "source_url": "https://omlx.ai/benchmarks/performance/2i890c8n",
   "tg_tokens": null,
   "tps": 39.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-syd49ihv-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: a9gvrvmi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 841.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 841.6 32.1 19.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/syd49ihv",
   "source_url": "https://omlx.ai/benchmarks/performance/syd49ihv",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-syd49ihv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: a9gvrvmi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 885.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 885.5 53.4 20.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/syd49ihv",
   "source_url": "https://omlx.ai/benchmarks/performance/syd49ihv",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-0etxziok-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: 28xrrxq7, n6i4th59; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 748.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 748.5 27.6 23.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0etxziok",
   "source_url": "https://omlx.ai/benchmarks/performance/0etxziok",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-0etxziok-ctx32768",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: 28xrrxq7, n6i4th59; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1342.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,342 41.7 26.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0etxziok",
   "source_url": "https://omlx.ai/benchmarks/performance/0etxziok",
   "tg_tokens": null,
   "tps": 41.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-0etxziok-ctx65536",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: 28xrrxq7, n6i4th59; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1185.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,185 34.8 28.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0etxziok",
   "source_url": "https://omlx.ai/benchmarks/performance/0etxziok",
   "tg_tokens": null,
   "tps": 34.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2xuei3z4-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 327.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 327.9 39.4 21.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2xuei3z4",
   "source_url": "https://omlx.ai/benchmarks/performance/2xuei3z4",
   "tg_tokens": null,
   "tps": 39.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2xuei3z4-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 416.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 416.9 69.8 24.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2xuei3z4",
   "source_url": "https://omlx.ai/benchmarks/performance/2xuei3z4",
   "tg_tokens": null,
   "tps": 69.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2xuei3z4-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.2 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 452.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 452.8 72.5 25.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2xuei3z4",
   "source_url": "https://omlx.ai/benchmarks/performance/2xuei3z4",
   "tg_tokens": null,
   "tps": 72.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2xuei3z4-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 430.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 430.6 53.9 26.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2xuei3z4",
   "source_url": "https://omlx.ai/benchmarks/performance/2xuei3z4",
   "tg_tokens": null,
   "tps": 53.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2xuei3z4-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 388.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 388.8 128.6 28.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2xuei3z4",
   "source_url": "https://omlx.ai/benchmarks/performance/2xuei3z4",
   "tg_tokens": null,
   "tps": 128.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1wogfqrm-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 3dy52jxl, 8t7il93y, c0rfsppd, m8przet5, ns4wwwyn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 862.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 862.7 33.6 21.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1wogfqrm",
   "source_url": "https://omlx.ai/benchmarks/performance/1wogfqrm",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1wogfqrm-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: 3dy52jxl, 8t7il93y, c0rfsppd, m8przet5, ns4wwwyn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1168.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,168 51.0 22.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1wogfqrm",
   "source_url": "https://omlx.ai/benchmarks/performance/1wogfqrm",
   "tg_tokens": null,
   "tps": 51.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1wogfqrm-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: 3dy52jxl, 8t7il93y, c0rfsppd, m8przet5, ns4wwwyn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1167.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,167 45.2 22.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1wogfqrm",
   "source_url": "https://omlx.ai/benchmarks/performance/1wogfqrm",
   "tg_tokens": null,
   "tps": 45.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1wogfqrm-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 3dy52jxl, 8t7il93y, c0rfsppd, m8przet5, ns4wwwyn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1112.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,112 71.6 23.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1wogfqrm",
   "source_url": "https://omlx.ai/benchmarks/performance/1wogfqrm",
   "tg_tokens": null,
   "tps": 71.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1wogfqrm-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: 3dy52jxl, 8t7il93y, c0rfsppd, m8przet5, ns4wwwyn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 981.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 981.7 45.3 24.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1wogfqrm",
   "source_url": "https://omlx.ai/benchmarks/performance/1wogfqrm",
   "tg_tokens": null,
   "tps": 45.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1wogfqrm-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: 3dy52jxl, 8t7il93y, c0rfsppd, m8przet5, ns4wwwyn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 832.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 832.8 36.6 26.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1wogfqrm",
   "source_url": "https://omlx.ai/benchmarks/performance/1wogfqrm",
   "tg_tokens": null,
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-9ktfm3dr-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 29.6 GB; TTFT 238944 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 274.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 64k tokens 274.3 PP tok/s 20.2 TG tok/s 238944 TTFT (ms) 29.6 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ktfm3dr",
   "source_url": "https://omlx.ai/benchmarks/performance/9ktfm3dr",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-111tesph-ctx4096",
   "model": "Qwopus3.8-27B-Flash-V2",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: y6wbcbic; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 163.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwopus3.8-27B-Flash-V2-oQ6e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 6bit; 4k 163.9 20.4 27.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/111tesph",
   "source_url": "https://omlx.ai/benchmarks/performance/111tesph",
   "tg_tokens": null,
   "tps": 20.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-111tesph-ctx16384",
   "model": "Qwopus3.8-27B-Flash-V2",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.4 GB at this context; duplicate traces of the same session: y6wbcbic; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 157.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwopus3.8-27B-Flash-V2-oQ6e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 6bit; 16k 157.5 19.6 29.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/111tesph",
   "source_url": "https://omlx.ai/benchmarks/performance/111tesph",
   "tg_tokens": null,
   "tps": 19.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-j81aztin-ctx4096",
   "model": "Qwopus3.8-27B-Flash-V2",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.3 GB at this context; duplicate traces of the same session: xwywjae0; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 227.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwopus3.8-27B-Flash-V2-oQ6e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 6bit; 4k 227.9 25.1 35.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j81aztin",
   "source_url": "https://omlx.ai/benchmarks/performance/j81aztin",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-j81aztin-ctx16384",
   "model": "Qwopus3.8-27B-Flash-V2",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: xwywjae0; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 211.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwopus3.8-27B-Flash-V2-oQ6e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 6bit; 16k 211.7 31.6 36.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j81aztin",
   "source_url": "https://omlx.ai/benchmarks/performance/j81aztin",
   "tg_tokens": null,
   "tps": 31.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-wjbcq65t-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.7,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 127.7 14.6 33.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wjbcq65t",
   "source_url": "https://omlx.ai/benchmarks/performance/wjbcq65t",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-wjbcq65t-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 126.3,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 126.3 12.6 37.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wjbcq65t",
   "source_url": "https://omlx.ai/benchmarks/performance/wjbcq65t",
   "tg_tokens": null,
   "tps": 12.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-zc8hm5r0-ctx4096",
   "model": "Swift-Qwen3.8-27B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 4k context; peak memory 23.4 GB; TTFT 88951 ms; duplicate traces of the same session: none; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 46.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-Qwen3.8-27B-MLX-6bit on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 6bit; 4k tokens 46.0 PP tok/s 6.5 TG tok/s 88951 TTFT (ms) 23.4 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zc8hm5r0",
   "source_url": "https://omlx.ai/benchmarks/performance/zc8hm5r0",
   "tg_tokens": null,
   "tps": 6.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-am2ccd1b-ctx8192",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: a9ezr3od, xwmxnd1r; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2138.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 8k 2,138 35.7 5.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/am2ccd1b",
   "source_url": "https://omlx.ai/benchmarks/performance/am2ccd1b",
   "tg_tokens": null,
   "tps": 35.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-am2ccd1b-ctx16384",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 6.1 GB at this context; duplicate traces of the same session: a9ezr3od, xwmxnd1r; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2444.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 2,444 76.8 6.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/am2ccd1b",
   "source_url": "https://omlx.ai/benchmarks/performance/am2ccd1b",
   "tg_tokens": null,
   "tps": 76.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-am2ccd1b-ctx32768",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: a9ezr3od, xwmxnd1r; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2151.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 32k 2,151 40.5 6.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/am2ccd1b",
   "source_url": "https://omlx.ai/benchmarks/performance/am2ccd1b",
   "tg_tokens": null,
   "tps": 40.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-8lm55ccy-ctx1024",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 5jw6mrdc, vkfj20zi, w3fmubc2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 359.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 359.4 45.8 20.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8lm55ccy",
   "source_url": "https://omlx.ai/benchmarks/performance/8lm55ccy",
   "tg_tokens": null,
   "tps": 45.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-8lm55ccy-ctx4096",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 5jw6mrdc, vkfj20zi, w3fmubc2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 398.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 398.1 43.7 20.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8lm55ccy",
   "source_url": "https://omlx.ai/benchmarks/performance/8lm55ccy",
   "tg_tokens": null,
   "tps": 43.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-8lm55ccy-ctx8192",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 5jw6mrdc, vkfj20zi, w3fmubc2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 386.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 8k 386.4 41.0 21.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8lm55ccy",
   "source_url": "https://omlx.ai/benchmarks/performance/8lm55ccy",
   "tg_tokens": null,
   "tps": 41.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-8lm55ccy-ctx16384",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 5jw6mrdc, vkfj20zi, w3fmubc2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 324.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k 324.9 38.2 21.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8lm55ccy",
   "source_url": "https://omlx.ai/benchmarks/performance/8lm55ccy",
   "tg_tokens": null,
   "tps": 38.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-51tl4x31-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 15.2 GB at this context; duplicate traces of the same session: adaas5o9, rkk6bzwu; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 939.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 939.8 108.3 15.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/51tl4x31",
   "source_url": "https://omlx.ai/benchmarks/performance/51tl4x31",
   "tg_tokens": null,
   "tps": 108.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-51tl4x31-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 15.4 GB at this context; duplicate traces of the same session: adaas5o9, rkk6bzwu; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1033.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 1,033 105.2 15.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/51tl4x31",
   "source_url": "https://omlx.ai/benchmarks/performance/51tl4x31",
   "tg_tokens": null,
   "tps": 105.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-51tl4x31-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 15.5 GB at this context; duplicate traces of the same session: adaas5o9, rkk6bzwu; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1006.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 1,006 99.0 15.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/51tl4x31",
   "source_url": "https://omlx.ai/benchmarks/performance/51tl4x31",
   "tg_tokens": null,
   "tps": 99.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-y7ljvhz1-ctx1024",
   "model": "muse-glimmer 30B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: cx15q4lq, qc84copk; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 194.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 194.4 20.8 15.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y7ljvhz1",
   "source_url": "https://omlx.ai/benchmarks/performance/y7ljvhz1",
   "tg_tokens": null,
   "tps": 20.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-y7ljvhz1-ctx4096",
   "model": "muse-glimmer 30B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: cx15q4lq, qc84copk; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 189.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 189.2 20.4 16.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y7ljvhz1",
   "source_url": "https://omlx.ai/benchmarks/performance/y7ljvhz1",
   "tg_tokens": null,
   "tps": 20.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-y7ljvhz1-ctx8192",
   "model": "muse-glimmer 30B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 16.2 GB at this context; duplicate traces of the same session: cx15q4lq, qc84copk; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 171.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 171.0 20.2 16.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y7ljvhz1",
   "source_url": "https://omlx.ai/benchmarks/performance/y7ljvhz1",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-y7ljvhz1-ctx16384",
   "model": "muse-glimmer 30B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: cx15q4lq, qc84copk; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 184.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 184.5 20.3 16.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y7ljvhz1",
   "source_url": "https://omlx.ai/benchmarks/performance/y7ljvhz1",
   "tg_tokens": null,
   "tps": 20.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-gg52tqc1-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: n15fgqtf, qkdf3d50; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 770.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 770.2 85.0 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gg52tqc1",
   "source_url": "https://omlx.ai/benchmarks/performance/gg52tqc1",
   "tg_tokens": null,
   "tps": 85.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-gg52tqc1-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: n15fgqtf, qkdf3d50; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 916.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 916.2 81.5 21.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gg52tqc1",
   "source_url": "https://omlx.ai/benchmarks/performance/gg52tqc1",
   "tg_tokens": null,
   "tps": 81.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-gg52tqc1-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: n15fgqtf, qkdf3d50; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 607.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 32k 607.7 52.8 22.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gg52tqc1",
   "source_url": "https://omlx.ai/benchmarks/performance/gg52tqc1",
   "tg_tokens": null,
   "tps": 52.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-hv4wv0sn-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 38.2 GB at this context; duplicate traces of the same session: 1uoaa627, 3dnkxlv7, 8eoc5o48; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1233.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,233 63.2 38.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hv4wv0sn",
   "source_url": "https://omlx.ai/benchmarks/performance/hv4wv0sn",
   "tg_tokens": null,
   "tps": 63.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-hv4wv0sn-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 40.1 GB at this context; duplicate traces of the same session: 1uoaa627, 3dnkxlv7, 8eoc5o48; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1426.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 1,426 71.7 40.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hv4wv0sn",
   "source_url": "https://omlx.ai/benchmarks/performance/hv4wv0sn",
   "tg_tokens": null,
   "tps": 71.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-hv4wv0sn-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 40.4 GB at this context; duplicate traces of the same session: 1uoaa627, 3dnkxlv7, 8eoc5o48; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1338.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 8k 1,338 61.0 40.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hv4wv0sn",
   "source_url": "https://omlx.ai/benchmarks/performance/hv4wv0sn",
   "tg_tokens": null,
   "tps": 61.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-hv4wv0sn-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 41.1 GB at this context; duplicate traces of the same session: 1uoaa627, 3dnkxlv7, 8eoc5o48; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1176.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 16k 1,176 89.4 41.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hv4wv0sn",
   "source_url": "https://omlx.ai/benchmarks/performance/hv4wv0sn",
   "tg_tokens": null,
   "tps": 89.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3plm9tl3-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 81.1 GB at this context; duplicate traces of the same session: ecuyzhfa, mj5slhde; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1559.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 1,559 69.6 81.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3plm9tl3",
   "source_url": "https://omlx.ai/benchmarks/performance/3plm9tl3",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3plm9tl3-ctx16384",
   "model": "Qwen3-Coder-Next",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 81.3 GB at this context; duplicate traces of the same session: ecuyzhfa, mj5slhde; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1535.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 16k 1,535 79.7 81.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3plm9tl3",
   "source_url": "https://omlx.ai/benchmarks/performance/3plm9tl3",
   "tg_tokens": null,
   "tps": 79.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3plm9tl3-ctx65536",
   "model": "Qwen3-Coder-Next",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 82.5 GB at this context; duplicate traces of the same session: ecuyzhfa, mj5slhde; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 985.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 64k 985.9 52.6 82.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3plm9tl3",
   "source_url": "https://omlx.ai/benchmarks/performance/3plm9tl3",
   "tg_tokens": null,
   "tps": 52.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-4ppbc8nz-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: rjv6m1pv; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 193.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 193.9 30.7 5.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ppbc8nz",
   "source_url": "https://omlx.ai/benchmarks/performance/4ppbc8nz",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-4ppbc8nz-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: rjv6m1pv; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 185.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 185.5 8.1 6.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ppbc8nz",
   "source_url": "https://omlx.ai/benchmarks/performance/4ppbc8nz",
   "tg_tokens": null,
   "tps": 8.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-rhjgviqm-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 07gfsa82, l3jit6sp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1499.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,499 88.7 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rhjgviqm",
   "source_url": "https://omlx.ai/benchmarks/performance/rhjgviqm",
   "tg_tokens": null,
   "tps": 88.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-rhjgviqm-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 07gfsa82, l3jit6sp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1645.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,645 79.8 21.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rhjgviqm",
   "source_url": "https://omlx.ai/benchmarks/performance/rhjgviqm",
   "tg_tokens": null,
   "tps": 79.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-rhjgviqm-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 07gfsa82, l3jit6sp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1482.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,482 71.4 21.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rhjgviqm",
   "source_url": "https://omlx.ai/benchmarks/performance/rhjgviqm",
   "tg_tokens": null,
   "tps": 71.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6t85t7ip-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 0t5jbe5x, 1het4ect, 41zaizpy, fgw6b5f0, h6bt18ub; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4580.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,580 129.6 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6t85t7ip",
   "source_url": "https://omlx.ai/benchmarks/performance/6t85t7ip",
   "tg_tokens": null,
   "tps": 129.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6t85t7ip-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 0t5jbe5x, 1het4ect, 41zaizpy, fgw6b5f0, h6bt18ub; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5924.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 5,924 125.2 20.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6t85t7ip",
   "source_url": "https://omlx.ai/benchmarks/performance/6t85t7ip",
   "tg_tokens": null,
   "tps": 125.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6t85t7ip-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 0t5jbe5x, 1het4ect, 41zaizpy, fgw6b5f0, h6bt18ub; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6041.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 6,041 117.2 20.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6t85t7ip",
   "source_url": "https://omlx.ai/benchmarks/performance/6t85t7ip",
   "tg_tokens": null,
   "tps": 117.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6t85t7ip-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: 0t5jbe5x, 1het4ect, 41zaizpy, fgw6b5f0, h6bt18ub; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5689.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 5,689 103.4 20.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6t85t7ip",
   "source_url": "https://omlx.ai/benchmarks/performance/6t85t7ip",
   "tg_tokens": null,
   "tps": 103.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6t85t7ip-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 0t5jbe5x, 1het4ect, 41zaizpy, fgw6b5f0, h6bt18ub; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4956.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 4,956 83.8 20.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6t85t7ip",
   "source_url": "https://omlx.ai/benchmarks/performance/6t85t7ip",
   "tg_tokens": null,
   "tps": 83.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6t85t7ip-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 0t5jbe5x, 1het4ect, 41zaizpy, fgw6b5f0, h6bt18ub; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3824.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 3,824 62.1 21.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6t85t7ip",
   "source_url": "https://omlx.ai/benchmarks/performance/6t85t7ip",
   "tg_tokens": null,
   "tps": 62.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-of0bgzkt-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: b9pz9lzd, ds099zrd, iis0stce, rrhhbftq, rt70z2u3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1779.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 1,779 53.8 29.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/of0bgzkt",
   "source_url": "https://omlx.ai/benchmarks/performance/of0bgzkt",
   "tg_tokens": null,
   "tps": 53.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-of0bgzkt-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.6 GB at this context; duplicate traces of the same session: b9pz9lzd, ds099zrd, iis0stce, rrhhbftq, rt70z2u3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3874.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 3,874 116.8 30.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/of0bgzkt",
   "source_url": "https://omlx.ai/benchmarks/performance/of0bgzkt",
   "tg_tokens": null,
   "tps": 116.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-of0bgzkt-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.7 GB at this context; duplicate traces of the same session: b9pz9lzd, ds099zrd, iis0stce, rrhhbftq, rt70z2u3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4728.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 4,728 67.0 30.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/of0bgzkt",
   "source_url": "https://omlx.ai/benchmarks/performance/of0bgzkt",
   "tg_tokens": null,
   "tps": 67.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-of0bgzkt-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.9 GB at this context; duplicate traces of the same session: b9pz9lzd, ds099zrd, iis0stce, rrhhbftq, rt70z2u3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4508.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 4,508 111.9 30.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/of0bgzkt",
   "source_url": "https://omlx.ai/benchmarks/performance/of0bgzkt",
   "tg_tokens": null,
   "tps": 111.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-of0bgzkt-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.2 GB at this context; duplicate traces of the same session: b9pz9lzd, ds099zrd, iis0stce, rrhhbftq, rt70z2u3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3989.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 3,989 102.9 31.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/of0bgzkt",
   "source_url": "https://omlx.ai/benchmarks/performance/of0bgzkt",
   "tg_tokens": null,
   "tps": 102.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-of0bgzkt-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: b9pz9lzd, ds099zrd, iis0stce, rrhhbftq, rt70z2u3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3156.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k 3,156 82.0 32.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/of0bgzkt",
   "source_url": "https://omlx.ai/benchmarks/performance/of0bgzkt",
   "tg_tokens": null,
   "tps": 82.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-d9y4dc9q-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 38.2 GB at this context; duplicate traces of the same session: 7k41l3s8, lgh72wcv, q8ltd5af; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1255.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,255 96.4 38.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d9y4dc9q",
   "source_url": "https://omlx.ai/benchmarks/performance/d9y4dc9q",
   "tg_tokens": null,
   "tps": 96.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-d9y4dc9q-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 40.1 GB at this context; duplicate traces of the same session: 7k41l3s8, lgh72wcv, q8ltd5af; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1421.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 1,421 87.0 40.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d9y4dc9q",
   "source_url": "https://omlx.ai/benchmarks/performance/d9y4dc9q",
   "tg_tokens": null,
   "tps": 87.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-d9y4dc9q-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 40.4 GB at this context; duplicate traces of the same session: 7k41l3s8, lgh72wcv, q8ltd5af; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1332.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 8k 1,332 41.6 40.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d9y4dc9q",
   "source_url": "https://omlx.ai/benchmarks/performance/d9y4dc9q",
   "tg_tokens": null,
   "tps": 41.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-d9y4dc9q-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 41.1 GB at this context; duplicate traces of the same session: 7k41l3s8, lgh72wcv, q8ltd5af; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1176.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 16k 1,176 96.3 41.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d9y4dc9q",
   "source_url": "https://omlx.ai/benchmarks/performance/d9y4dc9q",
   "tg_tokens": null,
   "tps": 96.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-tw54tnv0-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: jm72mf4f, krwf708z, p90ozk8f; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1337.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,337 56.0 35.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tw54tnv0",
   "source_url": "https://omlx.ai/benchmarks/performance/tw54tnv0",
   "tg_tokens": null,
   "tps": 56.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-tw54tnv0-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: jm72mf4f, krwf708z, p90ozk8f; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1529.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 1,529 57.5 37.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tw54tnv0",
   "source_url": "https://omlx.ai/benchmarks/performance/tw54tnv0",
   "tg_tokens": null,
   "tps": 57.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-tw54tnv0-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 37.6 GB at this context; duplicate traces of the same session: jm72mf4f, krwf708z, p90ozk8f; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1403.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 8k 1,403 60.2 37.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tw54tnv0",
   "source_url": "https://omlx.ai/benchmarks/performance/tw54tnv0",
   "tg_tokens": null,
   "tps": 60.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-tw54tnv0-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 38.2 GB at this context; duplicate traces of the same session: jm72mf4f, krwf708z, p90ozk8f; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1264.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 16k 1,264 57.8 38.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tw54tnv0",
   "source_url": "https://omlx.ai/benchmarks/performance/tw54tnv0",
   "tg_tokens": null,
   "tps": 57.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-7wgh59lv-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 38.2 GB at this context; duplicate traces of the same session: lhc68a8f; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1271.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,271 90.6 38.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7wgh59lv",
   "source_url": "https://omlx.ai/benchmarks/performance/7wgh59lv",
   "tg_tokens": null,
   "tps": 90.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-7wgh59lv-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 40.1 GB at this context; duplicate traces of the same session: lhc68a8f; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1431.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 1,431 87.1 40.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7wgh59lv",
   "source_url": "https://omlx.ai/benchmarks/performance/7wgh59lv",
   "tg_tokens": null,
   "tps": 87.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-7wgh59lv-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 40.4 GB at this context; duplicate traces of the same session: lhc68a8f; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1344.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 8k 1,344 72.2 40.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7wgh59lv",
   "source_url": "https://omlx.ai/benchmarks/performance/7wgh59lv",
   "tg_tokens": null,
   "tps": 72.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-7wgh59lv-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 41.1 GB at this context; duplicate traces of the same session: lhc68a8f; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1169.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 16k 1,169 66.6 41.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7wgh59lv",
   "source_url": "https://omlx.ai/benchmarks/performance/7wgh59lv",
   "tg_tokens": null,
   "tps": 66.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-wlluax6o-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 38.2 GB at this context; duplicate traces of the same session: 56n6lupt, 763jbeq7, zfnqzauz; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1261.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,261 97.0 38.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wlluax6o",
   "source_url": "https://omlx.ai/benchmarks/performance/wlluax6o",
   "tg_tokens": null,
   "tps": 97.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-wlluax6o-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 40.1 GB at this context; duplicate traces of the same session: 56n6lupt, 763jbeq7, zfnqzauz; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1433.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 1,433 85.3 40.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wlluax6o",
   "source_url": "https://omlx.ai/benchmarks/performance/wlluax6o",
   "tg_tokens": null,
   "tps": 85.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-wlluax6o-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 40.4 GB at this context; duplicate traces of the same session: 56n6lupt, 763jbeq7, zfnqzauz; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1344.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 8k 1,344 65.1 40.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wlluax6o",
   "source_url": "https://omlx.ai/benchmarks/performance/wlluax6o",
   "tg_tokens": null,
   "tps": 65.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-wlluax6o-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 41.1 GB at this context; duplicate traces of the same session: 56n6lupt, 763jbeq7, zfnqzauz; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1180.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 16k 1,180 87.2 41.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wlluax6o",
   "source_url": "https://omlx.ai/benchmarks/performance/wlluax6o",
   "tg_tokens": null,
   "tps": 87.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-kfo9kjso-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1358.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,358 66.3 35.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kfo9kjso",
   "source_url": "https://omlx.ai/benchmarks/performance/kfo9kjso",
   "tg_tokens": null,
   "tps": 66.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-kfo9kjso-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1536.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 1,536 64.3 37.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kfo9kjso",
   "source_url": "https://omlx.ai/benchmarks/performance/kfo9kjso",
   "tg_tokens": null,
   "tps": 64.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-kfo9kjso-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 37.6 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1432.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 8k 1,432 63.5 37.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kfo9kjso",
   "source_url": "https://omlx.ai/benchmarks/performance/kfo9kjso",
   "tg_tokens": null,
   "tps": 63.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-kfo9kjso-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 38.2 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1273.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 16k 1,273 61.1 38.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kfo9kjso",
   "source_url": "https://omlx.ai/benchmarks/performance/kfo9kjso",
   "tg_tokens": null,
   "tps": 61.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-s1khwo41-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 16.0 GB; TTFT 2599 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 394.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k tokens 394.0 PP tok/s 18.1 TG tok/s 2599 TTFT (ms) 16.0 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s1khwo41",
   "source_url": "https://omlx.ai/benchmarks/performance/s1khwo41",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-7bl3n12k-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 2gms6avf, 2r357az5, y38d1s8j; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 327.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 327.9 39.4 21.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7bl3n12k",
   "source_url": "https://omlx.ai/benchmarks/performance/7bl3n12k",
   "tg_tokens": null,
   "tps": 39.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-7bl3n12k-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: 2gms6avf, 2r357az5, y38d1s8j; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 416.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 416.9 69.8 24.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7bl3n12k",
   "source_url": "https://omlx.ai/benchmarks/performance/7bl3n12k",
   "tg_tokens": null,
   "tps": 69.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-7bl3n12k-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.2 GB at this context; duplicate traces of the same session: 2gms6avf, 2r357az5, y38d1s8j; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 452.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 452.8 72.5 25.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7bl3n12k",
   "source_url": "https://omlx.ai/benchmarks/performance/7bl3n12k",
   "tg_tokens": null,
   "tps": 72.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-7bl3n12k-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: 2gms6avf, 2r357az5, y38d1s8j; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 430.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 430.6 53.9 26.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7bl3n12k",
   "source_url": "https://omlx.ai/benchmarks/performance/7bl3n12k",
   "tg_tokens": null,
   "tps": 53.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-7bl3n12k-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: 2gms6avf, 2r357az5, y38d1s8j; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 388.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 388.8 128.6 28.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7bl3n12k",
   "source_url": "https://omlx.ai/benchmarks/performance/7bl3n12k",
   "tg_tokens": null,
   "tps": 128.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-91ggitin-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 46.0 GB; TTFT 36410 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1800.0,
   "provenance": "community",
   "quant": "GSQ-RCO-Q2",
   "quote": "Qwen3.8-Flash-Next-GSQ-RCO-Q2-mlx on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k tokens 1,800 PP tok/s 36.9 TG tok/s 36410 TTFT (ms) 46.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/91ggitin",
   "source_url": "https://omlx.ai/benchmarks/performance/91ggitin",
   "tg_tokens": null,
   "tps": 36.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-wzwjq1fd-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 14.1 GB; TTFT 1236 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 828.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k tokens 828.5 PP tok/s 80.6 TG tok/s 1236 TTFT (ms) 14.1 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wzwjq1fd",
   "source_url": "https://omlx.ai/benchmarks/performance/wzwjq1fd",
   "tg_tokens": null,
   "tps": 80.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-gj006eu2-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai build; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 448.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 448.2 35.4 21.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gj006eu2",
   "source_url": "https://omlx.ai/benchmarks/performance/gj006eu2",
   "tg_tokens": null,
   "tps": 35.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-gj006eu2-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 390.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 16k 390.6 39.3 21.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gj006eu2",
   "source_url": "https://omlx.ai/benchmarks/performance/gj006eu2",
   "tg_tokens": null,
   "tps": 39.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-gj006eu2-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai build; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 237.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 64k 237.7 12.4 23.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gj006eu2",
   "source_url": "https://omlx.ai/benchmarks/performance/gj006eu2",
   "tg_tokens": null,
   "tps": 12.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-h0i7l23k-ctx1024",
   "model": "GLM-4.7-Flash",
   "notes": "oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: 8ov7dfs5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 824.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "GLM-4.7-Flash-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 824.5 71.9 30.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h0i7l23k",
   "source_url": "https://omlx.ai/benchmarks/performance/h0i7l23k",
   "tg_tokens": null,
   "tps": 71.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-h0i7l23k-ctx4096",
   "model": "GLM-4.7-Flash",
   "notes": "oMLX community benchmark trace; peak memory 31.3 GB at this context; duplicate traces of the same session: 8ov7dfs5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1937.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "GLM-4.7-Flash-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 1,937 61.3 31.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h0i7l23k",
   "source_url": "https://omlx.ai/benchmarks/performance/h0i7l23k",
   "tg_tokens": null,
   "tps": 61.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-7f1klqfk-ctx1024",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: 06mdde0z, 46wkoeqg; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1309.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,309 72.4 177.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7f1klqfk",
   "source_url": "https://omlx.ai/benchmarks/performance/7f1klqfk",
   "tg_tokens": null,
   "tps": 72.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-7f1klqfk-ctx32768",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 179.2 GB at this context; duplicate traces of the same session: 06mdde0z, 46wkoeqg; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1905.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,905 65.0 179.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7f1klqfk",
   "source_url": "https://omlx.ai/benchmarks/performance/7f1klqfk",
   "tg_tokens": null,
   "tps": 65.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-7f1klqfk-ctx65536",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 179.7 GB at this context; duplicate traces of the same session: 06mdde0z, 46wkoeqg; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1874.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,874 86.0 179.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7f1klqfk",
   "source_url": "https://omlx.ai/benchmarks/performance/7f1klqfk",
   "tg_tokens": null,
   "tps": 86.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-kfixodx3-ctx1024",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: e6w269at, soy921b1; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 407.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 407.1 53.5 177.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kfixodx3",
   "source_url": "https://omlx.ai/benchmarks/performance/kfixodx3",
   "tg_tokens": null,
   "tps": 53.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-kfixodx3-ctx32768",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 179.2 GB at this context; duplicate traces of the same session: e6w269at, soy921b1; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1611.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,611 58.5 179.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kfixodx3",
   "source_url": "https://omlx.ai/benchmarks/performance/kfixodx3",
   "tg_tokens": null,
   "tps": 58.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-kfixodx3-ctx65536",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 179.7 GB at this context; duplicate traces of the same session: e6w269at, soy921b1; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1861.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,861 58.2 179.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kfixodx3",
   "source_url": "https://omlx.ai/benchmarks/performance/kfixodx3",
   "tg_tokens": null,
   "tps": 58.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-vxrcs5rj-ctx1024",
   "model": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: 8vteqt9m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1660.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated-mlx-8bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 1,660 93.0 35.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vxrcs5rj",
   "source_url": "https://omlx.ai/benchmarks/performance/vxrcs5rj",
   "tg_tokens": null,
   "tps": 93.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-vxrcs5rj-ctx4096",
   "model": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated",
   "notes": "oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: 8vteqt9m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2094.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated-mlx-8bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 2,094 94.4 37.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vxrcs5rj",
   "source_url": "https://omlx.ai/benchmarks/performance/vxrcs5rj",
   "tg_tokens": null,
   "tps": 94.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-c8ea90nz-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 65.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1782.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,782 70.7 65.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c8ea90nz",
   "source_url": "https://omlx.ai/benchmarks/performance/c8ea90nz",
   "tg_tokens": null,
   "tps": 70.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-c8ea90nz-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2256.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,256 69.2 66.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c8ea90nz",
   "source_url": "https://omlx.ai/benchmarks/performance/c8ea90nz",
   "tg_tokens": null,
   "tps": 69.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-ltar3byp-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 38.2 GB at this context; duplicate traces of the same session: fb3k8gm4; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1271.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,271 90.6 38.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ltar3byp",
   "source_url": "https://omlx.ai/benchmarks/performance/ltar3byp",
   "tg_tokens": null,
   "tps": 90.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ltar3byp-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 40.1 GB at this context; duplicate traces of the same session: fb3k8gm4; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1431.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 1,431 87.1 40.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ltar3byp",
   "source_url": "https://omlx.ai/benchmarks/performance/ltar3byp",
   "tg_tokens": null,
   "tps": 87.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ltar3byp-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 40.4 GB at this context; duplicate traces of the same session: fb3k8gm4; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1344.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 8k 1,344 72.2 40.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ltar3byp",
   "source_url": "https://omlx.ai/benchmarks/performance/ltar3byp",
   "tg_tokens": null,
   "tps": 72.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-ltar3byp-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 41.1 GB at this context; duplicate traces of the same session: fb3k8gm4; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1169.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 16k 1,169 66.6 41.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ltar3byp",
   "source_url": "https://omlx.ai/benchmarks/performance/ltar3byp",
   "tg_tokens": null,
   "tps": 66.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-e0deuj5n-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: 23lw10ml, vatiz6i1; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1358.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,358 66.3 35.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e0deuj5n",
   "source_url": "https://omlx.ai/benchmarks/performance/e0deuj5n",
   "tg_tokens": null,
   "tps": 66.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-e0deuj5n-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: 23lw10ml, vatiz6i1; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1536.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 1,536 64.3 37.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e0deuj5n",
   "source_url": "https://omlx.ai/benchmarks/performance/e0deuj5n",
   "tg_tokens": null,
   "tps": 64.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-e0deuj5n-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 37.6 GB at this context; duplicate traces of the same session: 23lw10ml, vatiz6i1; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1432.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 8k 1,432 63.5 37.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e0deuj5n",
   "source_url": "https://omlx.ai/benchmarks/performance/e0deuj5n",
   "tg_tokens": null,
   "tps": 63.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-e0deuj5n-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 38.2 GB at this context; duplicate traces of the same session: 23lw10ml, vatiz6i1; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1273.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 16k 1,273 61.1 38.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e0deuj5n",
   "source_url": "https://omlx.ai/benchmarks/performance/e0deuj5n",
   "tg_tokens": null,
   "tps": 61.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-cecdcjmu-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: qjok0j2u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 348.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 348.0 32.7 25.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cecdcjmu",
   "source_url": "https://omlx.ai/benchmarks/performance/cecdcjmu",
   "tg_tokens": null,
   "tps": 32.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-cecdcjmu-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 27.0 GB at this context; duplicate traces of the same session: qjok0j2u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 331.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 331.9 41.7 27.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cecdcjmu",
   "source_url": "https://omlx.ai/benchmarks/performance/cecdcjmu",
   "tg_tokens": null,
   "tps": 41.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fih6ehyw-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 5kkci61m, 8gexczuo, boqm45yo, fn5rsj40, hrit0sgx, lu939rj1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 804.1,
   "provenance": "community",
   "quant": "AWQ-5.0bpw",
   "quote": "Qwen3.8-27B-AWQ-5.0bpw on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 804.1 35.2 20.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fih6ehyw",
   "source_url": "https://omlx.ai/benchmarks/performance/fih6ehyw",
   "tg_tokens": null,
   "tps": 35.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fih6ehyw-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 5kkci61m, 8gexczuo, boqm45yo, fn5rsj40, hrit0sgx, lu939rj1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1248.0,
   "provenance": "community",
   "quant": "AWQ-5.0bpw",
   "quote": "Qwen3.8-27B-AWQ-5.0bpw on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 4k 1,248 69.6 21.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fih6ehyw",
   "source_url": "https://omlx.ai/benchmarks/performance/fih6ehyw",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fih6ehyw-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: 5kkci61m, 8gexczuo, boqm45yo, fn5rsj40, hrit0sgx, lu939rj1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1278.0,
   "provenance": "community",
   "quant": "AWQ-5.0bpw",
   "quote": "Qwen3.8-27B-AWQ-5.0bpw on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 1,278 35.3 22.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fih6ehyw",
   "source_url": "https://omlx.ai/benchmarks/performance/fih6ehyw",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fih6ehyw-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: 5kkci61m, 8gexczuo, boqm45yo, fn5rsj40, hrit0sgx, lu939rj1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1218.0,
   "provenance": "community",
   "quant": "AWQ-5.0bpw",
   "quote": "Qwen3.8-27B-AWQ-5.0bpw on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 1,218 41.2 22.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fih6ehyw",
   "source_url": "https://omlx.ai/benchmarks/performance/fih6ehyw",
   "tg_tokens": null,
   "tps": 41.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fih6ehyw-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: 5kkci61m, 8gexczuo, boqm45yo, fn5rsj40, hrit0sgx, lu939rj1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 891.0,
   "provenance": "community",
   "quant": "AWQ-5.0bpw",
   "quote": "Qwen3.8-27B-AWQ-5.0bpw on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 64k 891.0 35.5 26.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fih6ehyw",
   "source_url": "https://omlx.ai/benchmarks/performance/fih6ehyw",
   "tg_tokens": null,
   "tps": 35.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fih6ehyw-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.3 GB at this context; duplicate traces of the same session: 5kkci61m, 8gexczuo, boqm45yo, fn5rsj40, hrit0sgx, lu939rj1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 582.9,
   "provenance": "community",
   "quant": "AWQ-5.0bpw",
   "quote": "Qwen3.8-27B-AWQ-5.0bpw on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 128k 582.9 30.3 31.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fih6ehyw",
   "source_url": "https://omlx.ai/benchmarks/performance/fih6ehyw",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fih6ehyw-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 38.0 GB at this context; duplicate traces of the same session: 5kkci61m, 8gexczuo, boqm45yo, fn5rsj40, hrit0sgx, lu939rj1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 400.4,
   "provenance": "community",
   "quant": "AWQ-5.0bpw",
   "quote": "Qwen3.8-27B-AWQ-5.0bpw on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 195k 400.4 25.8 38.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fih6ehyw",
   "source_url": "https://omlx.ai/benchmarks/performance/fih6ehyw",
   "tg_tokens": null,
   "tps": 25.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-k3c8pvr4-ctx1024",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "ET-O5.5-G6A-G4.7-DSV4-K3-RLOO MTP DFlash2 build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: y03c6acn; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-ET-O5.5-G6A-G4.7-DSV4-K3-RLOO-MTP-DFlash2-oQ4e on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 125.3 14.1 19.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k3c8pvr4",
   "source_url": "https://omlx.ai/benchmarks/performance/k3c8pvr4",
   "tg_tokens": null,
   "tps": 14.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-k3c8pvr4-ctx4096",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "ET-O5.5-G6A-G4.7-DSV4-K3-RLOO MTP DFlash2 build; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: y03c6acn; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 124.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-ET-O5.5-G6A-G4.7-DSV4-K3-RLOO-MTP-DFlash2-oQ4e on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 124.2 14.0 22.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k3c8pvr4",
   "source_url": "https://omlx.ai/benchmarks/performance/k3c8pvr4",
   "tg_tokens": null,
   "tps": 14.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-cayjov16-ctx1024",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "ET-O5.5-G6A-G4.7-DSV4-K3-RLOO MTP DFlash2 build; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: xlfv9844; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 124.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-ET-O5.5-G6A-G4.7-DSV4-K3-RLOO-MTP-DFlash2-oQ4e on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 124.4 16.0 19.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cayjov16",
   "source_url": "https://omlx.ai/benchmarks/performance/cayjov16",
   "tg_tokens": null,
   "tps": 16.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-cayjov16-ctx4096",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "ET-O5.5-G6A-G4.7-DSV4-K3-RLOO MTP DFlash2 build; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: xlfv9844; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 123.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-ET-O5.5-G6A-G4.7-DSV4-K3-RLOO-MTP-DFlash2-oQ4e on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 123.9 16.2 23.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cayjov16",
   "source_url": "https://omlx.ai/benchmarks/performance/cayjov16",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-2o85jur5-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: l1fa0yv0, sewqksxm; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1254.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,254 77.5 23.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2o85jur5",
   "source_url": "https://omlx.ai/benchmarks/performance/2o85jur5",
   "tg_tokens": null,
   "tps": 77.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-2o85jur5-ctx32768",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: l1fa0yv0, sewqksxm; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1355.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,355 69.2 26.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2o85jur5",
   "source_url": "https://omlx.ai/benchmarks/performance/2o85jur5",
   "tg_tokens": null,
   "tps": 69.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-2o85jur5-ctx65536",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: l1fa0yv0, sewqksxm; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1186.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,186 54.2 28.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2o85jur5",
   "source_url": "https://omlx.ai/benchmarks/performance/2o85jur5",
   "tg_tokens": null,
   "tps": 54.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-t00xa5p2-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: jd8exniv, lbgeigf6, nkqwhg7w, nl6oaz1n, p2k7v6vo, qw3o1n0e, uvmyiffj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 254.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 254.3 64.3 18.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t00xa5p2",
   "source_url": "https://omlx.ai/benchmarks/performance/t00xa5p2",
   "tg_tokens": null,
   "tps": 64.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-t00xa5p2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: jd8exniv, lbgeigf6, nkqwhg7w, nl6oaz1n, p2k7v6vo, qw3o1n0e, uvmyiffj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 269.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 269.4 81.2 21.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t00xa5p2",
   "source_url": "https://omlx.ai/benchmarks/performance/t00xa5p2",
   "tg_tokens": null,
   "tps": 81.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-t00xa5p2-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: jd8exniv, lbgeigf6, nkqwhg7w, nl6oaz1n, p2k7v6vo, qw3o1n0e, uvmyiffj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 264.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 264.6 39.9 22.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t00xa5p2",
   "source_url": "https://omlx.ai/benchmarks/performance/t00xa5p2",
   "tg_tokens": null,
   "tps": 39.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-t00xa5p2-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: jd8exniv, lbgeigf6, nkqwhg7w, nl6oaz1n, p2k7v6vo, qw3o1n0e, uvmyiffj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 254.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 254.4 56.5 23.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t00xa5p2",
   "source_url": "https://omlx.ai/benchmarks/performance/t00xa5p2",
   "tg_tokens": null,
   "tps": 56.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-t00xa5p2-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: jd8exniv, lbgeigf6, nkqwhg7w, nl6oaz1n, p2k7v6vo, qw3o1n0e, uvmyiffj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 235.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 235.7 50.4 25.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t00xa5p2",
   "source_url": "https://omlx.ai/benchmarks/performance/t00xa5p2",
   "tg_tokens": null,
   "tps": 50.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-t00xa5p2-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: jd8exniv, lbgeigf6, nkqwhg7w, nl6oaz1n, p2k7v6vo, qw3o1n0e, uvmyiffj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 205.3 44.9 27.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t00xa5p2",
   "source_url": "https://omlx.ai/benchmarks/performance/t00xa5p2",
   "tg_tokens": null,
   "tps": 44.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-t00xa5p2-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.3 GB at this context; duplicate traces of the same session: jd8exniv, lbgeigf6, nkqwhg7w, nl6oaz1n, p2k7v6vo, qw3o1n0e, uvmyiffj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 164.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 164.0 39.3 32.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t00xa5p2",
   "source_url": "https://omlx.ai/benchmarks/performance/t00xa5p2",
   "tg_tokens": null,
   "tps": 39.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-t00xa5p2-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: jd8exniv, lbgeigf6, nkqwhg7w, nl6oaz1n, p2k7v6vo, qw3o1n0e, uvmyiffj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 135.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 195k 135.3 36.4 37.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t00xa5p2",
   "source_url": "https://omlx.ai/benchmarks/performance/t00xa5p2",
   "tg_tokens": null,
   "tps": 36.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qkrxd4wq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 20v43rc8, kvbxj198, qll24smf, uqa0c5wh, w4fbue8w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 981.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 981.4 46.9 17.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qkrxd4wq",
   "source_url": "https://omlx.ai/benchmarks/performance/qkrxd4wq",
   "tg_tokens": null,
   "tps": 46.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qkrxd4wq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 20v43rc8, kvbxj198, qll24smf, uqa0c5wh, w4fbue8w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1268.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,268 52.1 19.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qkrxd4wq",
   "source_url": "https://omlx.ai/benchmarks/performance/qkrxd4wq",
   "tg_tokens": null,
   "tps": 52.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qkrxd4wq-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 20v43rc8, kvbxj198, qll24smf, uqa0c5wh, w4fbue8w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1173.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,173 65.8 19.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qkrxd4wq",
   "source_url": "https://omlx.ai/benchmarks/performance/qkrxd4wq",
   "tg_tokens": null,
   "tps": 65.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qkrxd4wq-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 20v43rc8, kvbxj198, qll24smf, uqa0c5wh, w4fbue8w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1098.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,098 74.3 19.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qkrxd4wq",
   "source_url": "https://omlx.ai/benchmarks/performance/qkrxd4wq",
   "tg_tokens": null,
   "tps": 74.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qkrxd4wq-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: 20v43rc8, kvbxj198, qll24smf, uqa0c5wh, w4fbue8w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 974.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 974.0 81.3 21.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qkrxd4wq",
   "source_url": "https://omlx.ai/benchmarks/performance/qkrxd4wq",
   "tg_tokens": null,
   "tps": 81.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qkrxd4wq-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 20v43rc8, kvbxj198, qll24smf, uqa0c5wh, w4fbue8w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 808.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 808.3 54.8 23.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qkrxd4wq",
   "source_url": "https://omlx.ai/benchmarks/performance/qkrxd4wq",
   "tg_tokens": null,
   "tps": 54.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-1430gi6s-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 71.6 GB; TTFT 10527 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 778.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k tokens 778.2 PP tok/s 67.6 TG tok/s 10527 TTFT (ms) 71.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1430gi6s",
   "source_url": "https://omlx.ai/benchmarks/performance/1430gi6s",
   "tg_tokens": null,
   "tps": 67.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-oa9hd6nd-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 72.5 GB; TTFT 43669 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 750.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k tokens 750.4 PP tok/s 16.2 TG tok/s 43669 TTFT (ms) 72.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oa9hd6nd",
   "source_url": "https://omlx.ai/benchmarks/performance/oa9hd6nd",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-x1dz3pos-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: 4q45rbi0, bejsu4zv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 832.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 832.2 43.9 24.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x1dz3pos",
   "source_url": "https://omlx.ai/benchmarks/performance/x1dz3pos",
   "tg_tokens": null,
   "tps": 43.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-x1dz3pos-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.6 GB at this context; duplicate traces of the same session: 4q45rbi0, bejsu4zv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 799.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 799.4 41.3 25.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x1dz3pos",
   "source_url": "https://omlx.ai/benchmarks/performance/x1dz3pos",
   "tg_tokens": null,
   "tps": 41.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-x1dz3pos-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ5e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: 4q45rbi0, bejsu4zv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 757.5,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ5e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 757.5 51.4 25.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x1dz3pos",
   "source_url": "https://omlx.ai/benchmarks/performance/x1dz3pos",
   "tg_tokens": null,
   "tps": 51.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-78w6vqnf-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: hxx3s9et; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4187.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,187 122.9 14.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/78w6vqnf",
   "source_url": "https://omlx.ai/benchmarks/performance/78w6vqnf",
   "tg_tokens": null,
   "tps": 122.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-78w6vqnf-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: hxx3s9et; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4181.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 4,181 118.8 14.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/78w6vqnf",
   "source_url": "https://omlx.ai/benchmarks/performance/78w6vqnf",
   "tg_tokens": null,
   "tps": 118.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-r1v590hc-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 15.4 GB; TTFT 8625 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1900.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k tokens 1,900 PP tok/s 65.8 TG tok/s 8625 TTFT (ms) 15.4 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r1v590hc",
   "source_url": "https://omlx.ai/benchmarks/performance/r1v590hc",
   "tg_tokens": null,
   "tps": 65.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-ls675av4-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai build; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 3c43w3a3; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 448.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 448.2 35.4 21.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ls675av4",
   "source_url": "https://omlx.ai/benchmarks/performance/ls675av4",
   "tg_tokens": null,
   "tps": 35.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-ls675av4-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 3c43w3a3; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 390.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 16k 390.6 39.3 21.5 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ls675av4",
   "source_url": "https://omlx.ai/benchmarks/performance/ls675av4",
   "tg_tokens": null,
   "tps": 39.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-ls675av4-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai build; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 3c43w3a3; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 237.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 64k 237.7 12.4 23.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ls675av4",
   "source_url": "https://omlx.ai/benchmarks/performance/ls675av4",
   "tg_tokens": null,
   "tps": 12.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ifrmcx1u-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 8k context; peak memory 13.8 GB; TTFT 8164 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1003.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k tokens 1,003 PP tok/s 98.5 TG tok/s 8164 TTFT (ms) 13.8 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ifrmcx1u",
   "source_url": "https://omlx.ai/benchmarks/performance/ifrmcx1u",
   "tg_tokens": null,
   "tps": 98.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-9nzjvl5z-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 8k context; peak memory 13.8 GB; TTFT 8409 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 974.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k tokens 974.2 PP tok/s 86.1 TG tok/s 8409 TTFT (ms) 13.8 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9nzjvl5z",
   "source_url": "https://omlx.ai/benchmarks/performance/9nzjvl5z",
   "tg_tokens": null,
   "tps": 86.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-m8g8kvlz-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: mnwtw091; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 959.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 959.5 108.5 13.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m8g8kvlz",
   "source_url": "https://omlx.ai/benchmarks/performance/m8g8kvlz",
   "tg_tokens": null,
   "tps": 108.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-m8g8kvlz-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: mnwtw091; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1036.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 1,036 105.2 13.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m8g8kvlz",
   "source_url": "https://omlx.ai/benchmarks/performance/m8g8kvlz",
   "tg_tokens": null,
   "tps": 105.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-m8g8kvlz-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.8 GB at this context; duplicate traces of the same session: mnwtw091; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1008.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 1,008 97.9 13.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m8g8kvlz",
   "source_url": "https://omlx.ai/benchmarks/performance/m8g8kvlz",
   "tg_tokens": null,
   "tps": 97.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-m8g8kvlz-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.9 GB at this context; duplicate traces of the same session: mnwtw091; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 950.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 950.6 87.5 13.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m8g8kvlz",
   "source_url": "https://omlx.ai/benchmarks/performance/m8g8kvlz",
   "tg_tokens": null,
   "tps": 87.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-m8g8kvlz-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 14.2 GB at this context; duplicate traces of the same session: mnwtw091; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 851.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 851.4 69.9 14.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m8g8kvlz",
   "source_url": "https://omlx.ai/benchmarks/performance/m8g8kvlz",
   "tg_tokens": null,
   "tps": 69.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-m8g8kvlz-ctx65536",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 14.9 GB at this context; duplicate traces of the same session: mnwtw091; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 698.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 64k 698.1 50.6 14.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m8g8kvlz",
   "source_url": "https://omlx.ai/benchmarks/performance/m8g8kvlz",
   "tg_tokens": null,
   "tps": 50.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-m8g8kvlz-ctx131072",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: mnwtw091; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 514.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 128k 514.6 32.1 16.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m8g8kvlz",
   "source_url": "https://omlx.ai/benchmarks/performance/m8g8kvlz",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-afzfu9to-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "OptiQ static 5.5bpw build; oMLX community benchmark trace; peak memory 25.6 GB at this context; duplicate traces of the same session: zh4xoptd; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 973.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-static-5.5bpw on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 973.4 65.6 25.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/afzfu9to",
   "source_url": "https://omlx.ai/benchmarks/performance/afzfu9to",
   "tg_tokens": null,
   "tps": 65.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-afzfu9to-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "OptiQ static 5.5bpw build; oMLX community benchmark trace; peak memory 27.3 GB at this context; duplicate traces of the same session: zh4xoptd; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1071.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-static-5.5bpw on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,071 63.5 27.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/afzfu9to",
   "source_url": "https://omlx.ai/benchmarks/performance/afzfu9to",
   "tg_tokens": null,
   "tps": 63.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-vfqqmhib-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "refusal-ablated MTP build; MLX; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: ibzsml2k; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 915.2,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-refusal-ablated-MLX-oQ8-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 915.2 87.2 37.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vfqqmhib",
   "source_url": "https://omlx.ai/benchmarks/performance/vfqqmhib",
   "tg_tokens": null,
   "tps": 87.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-vfqqmhib-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "refusal-ablated MTP build; MLX; oMLX community benchmark trace; peak memory 39.4 GB at this context; duplicate traces of the same session: ibzsml2k; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1026.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-refusal-ablated-MLX-oQ8-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 1,026 63.6 39.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vfqqmhib",
   "source_url": "https://omlx.ai/benchmarks/performance/vfqqmhib",
   "tg_tokens": null,
   "tps": 63.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-z1csuu7z-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "refusal-ablated MTP build; MLX; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: hz2ojmcl; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 875.9,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-refusal-ablated-MLX-oQ8-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 875.9 38.8 37.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z1csuu7z",
   "source_url": "https://omlx.ai/benchmarks/performance/z1csuu7z",
   "tg_tokens": null,
   "tps": 38.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-z1csuu7z-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "refusal-ablated MTP build; MLX; oMLX community benchmark trace; peak memory 39.4 GB at this context; duplicate traces of the same session: hz2ojmcl; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1015.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-refusal-ablated-MLX-oQ8-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 1,015 49.8 39.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z1csuu7z",
   "source_url": "https://omlx.ai/benchmarks/performance/z1csuu7z",
   "tg_tokens": null,
   "tps": 49.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hw1ic7fh-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 65.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1782.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,782 70.7 65.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hw1ic7fh",
   "source_url": "https://omlx.ai/benchmarks/performance/hw1ic7fh",
   "tg_tokens": null,
   "tps": 70.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hw1ic7fh-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2256.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,256 69.2 66.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hw1ic7fh",
   "source_url": "https://omlx.ai/benchmarks/performance/hw1ic7fh",
   "tg_tokens": null,
   "tps": 69.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Pro",
   "id": "omlx-9s8lky01-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: 8wg5f5ra, hd4nlwxu; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 1k 102.3 14.0 17.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9s8lky01",
   "source_url": "https://omlx.ai/benchmarks/performance/9s8lky01",
   "tg_tokens": null,
   "tps": 14.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-9s8lky01-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: 8wg5f5ra, hd4nlwxu; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 4k 103.9 17.9 18.7 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9s8lky01",
   "source_url": "https://omlx.ai/benchmarks/performance/9s8lky01",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-9s8lky01-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: 8wg5f5ra, hd4nlwxu; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 26.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 64k 26.5 1.7 25.4 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9s8lky01",
   "source_url": "https://omlx.ai/benchmarks/performance/9s8lky01",
   "tg_tokens": null,
   "tps": 1.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-hgbp6xiw-ctx1024",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "ET-O5.5-G6A-G4.7-DSV4-K3-RLOO MTP DFlash2 build; oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: 8eyihu4o; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-ET-O5.5-G6A-G4.7-DSV4-K3-RLOO-MTP-DFlash2-oQ4e on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 130.5 16.0 17.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hgbp6xiw",
   "source_url": "https://omlx.ai/benchmarks/performance/hgbp6xiw",
   "tg_tokens": null,
   "tps": 16.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-hgbp6xiw-ctx4096",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "ET-O5.5-G6A-G4.7-DSV4-K3-RLOO MTP DFlash2 build; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 8eyihu4o; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 119.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-ET-O5.5-G6A-G4.7-DSV4-K3-RLOO-MTP-DFlash2-oQ4e on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 119.0 18.6 20.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hgbp6xiw",
   "source_url": "https://omlx.ai/benchmarks/performance/hgbp6xiw",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3q9r6686-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: lw0ysvqx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 935.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 935.4 33.8 16.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3q9r6686",
   "source_url": "https://omlx.ai/benchmarks/performance/3q9r6686",
   "tg_tokens": null,
   "tps": 33.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3q9r6686-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: lw0ysvqx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 970.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 970.3 33.3 17.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3q9r6686",
   "source_url": "https://omlx.ai/benchmarks/performance/3q9r6686",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-xk1p5mql-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 39q6b3fj, 56m5ch0p; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1054.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,054 56.0 23.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xk1p5mql",
   "source_url": "https://omlx.ai/benchmarks/performance/xk1p5mql",
   "tg_tokens": null,
   "tps": 56.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-xk1p5mql-ctx32768",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: 39q6b3fj, 56m5ch0p; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1318.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,318 117.7 26.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xk1p5mql",
   "source_url": "https://omlx.ai/benchmarks/performance/xk1p5mql",
   "tg_tokens": null,
   "tps": 117.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-xk1p5mql-ctx65536",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 28.8 GB at this context; duplicate traces of the same session: 39q6b3fj, 56m5ch0p; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1151.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,151 65.0 28.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xk1p5mql",
   "source_url": "https://omlx.ai/benchmarks/performance/xk1p5mql",
   "tg_tokens": null,
   "tps": 65.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-m70628me-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: asti1x4s; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 189.1,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 189.1 27.8 30.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m70628me",
   "source_url": "https://omlx.ai/benchmarks/performance/m70628me",
   "tg_tokens": null,
   "tps": 27.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-m70628me-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.7 GB at this context; duplicate traces of the same session: asti1x4s; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 273.1,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 273.1 30.8 33.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m70628me",
   "source_url": "https://omlx.ai/benchmarks/performance/m70628me",
   "tg_tokens": null,
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-m70628me-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: asti1x4s; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 274.3,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 274.3 33.0 34.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m70628me",
   "source_url": "https://omlx.ai/benchmarks/performance/m70628me",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-m70628me-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: asti1x4s; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 256.7,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 256.7 33.3 36.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m70628me",
   "source_url": "https://omlx.ai/benchmarks/performance/m70628me",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-uwur8rtu-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 15sxptr7, 2bpqmk2s, 73nq0l2d, c98ij19k, vbn2v83s; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 327.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 327.5 70.0 21.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uwur8rtu",
   "source_url": "https://omlx.ai/benchmarks/performance/uwur8rtu",
   "tg_tokens": null,
   "tps": 70.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-uwur8rtu-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: 15sxptr7, 2bpqmk2s, 73nq0l2d, c98ij19k, vbn2v83s; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 415.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 415.8 70.0 24.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uwur8rtu",
   "source_url": "https://omlx.ai/benchmarks/performance/uwur8rtu",
   "tg_tokens": null,
   "tps": 70.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-uwur8rtu-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.2 GB at this context; duplicate traces of the same session: 15sxptr7, 2bpqmk2s, 73nq0l2d, c98ij19k, vbn2v83s; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 452.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 452.4 69.0 25.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uwur8rtu",
   "source_url": "https://omlx.ai/benchmarks/performance/uwur8rtu",
   "tg_tokens": null,
   "tps": 69.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-uwur8rtu-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: 15sxptr7, 2bpqmk2s, 73nq0l2d, c98ij19k, vbn2v83s; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 430.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 430.1 92.7 26.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uwur8rtu",
   "source_url": "https://omlx.ai/benchmarks/performance/uwur8rtu",
   "tg_tokens": null,
   "tps": 92.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-uwur8rtu-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: 15sxptr7, 2bpqmk2s, 73nq0l2d, c98ij19k, vbn2v83s; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 388.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 388.7 125.1 28.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uwur8rtu",
   "source_url": "https://omlx.ai/benchmarks/performance/uwur8rtu",
   "tg_tokens": null,
   "tps": 125.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-uwur8rtu-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.6 GB at this context; duplicate traces of the same session: 15sxptr7, 2bpqmk2s, 73nq0l2d, c98ij19k, vbn2v83s; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 326.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 326.4 63.5 32.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uwur8rtu",
   "source_url": "https://omlx.ai/benchmarks/performance/uwur8rtu",
   "tg_tokens": null,
   "tps": 63.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ar1f4awt-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 16.9 GB; TTFT 1360 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 752.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 752.9 PP tok/s 54.5 TG tok/s 1360 TTFT (ms) 16.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ar1f4awt",
   "source_url": "https://omlx.ai/benchmarks/performance/ar1f4awt",
   "tg_tokens": null,
   "tps": 54.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-z9fafv82-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 17.6 GB; TTFT 2756 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 371.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k tokens 371.5 PP tok/s 23.4 TG tok/s 2756 TTFT (ms) 17.6 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z9fafv82",
   "source_url": "https://omlx.ai/benchmarks/performance/z9fafv82",
   "tg_tokens": null,
   "tps": 23.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-xtd8jsbc-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: m5cveafh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 468.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 468.2 32.4 17.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xtd8jsbc",
   "source_url": "https://omlx.ai/benchmarks/performance/xtd8jsbc",
   "tg_tokens": null,
   "tps": 32.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-xtd8jsbc-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: m5cveafh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 477.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 477.6 30.5 18.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xtd8jsbc",
   "source_url": "https://omlx.ai/benchmarks/performance/xtd8jsbc",
   "tg_tokens": null,
   "tps": 30.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-2n2t2pwn-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: hmtz38es, i3glyrrz, kiuw40mf, ux0q6qc0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1579.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,579 101.8 17.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2n2t2pwn",
   "source_url": "https://omlx.ai/benchmarks/performance/2n2t2pwn",
   "tg_tokens": null,
   "tps": 101.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-2n2t2pwn-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: hmtz38es, i3glyrrz, kiuw40mf, ux0q6qc0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1900.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 1,900 102.0 18.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2n2t2pwn",
   "source_url": "https://omlx.ai/benchmarks/performance/2n2t2pwn",
   "tg_tokens": null,
   "tps": 102.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-2n2t2pwn-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: hmtz38es, i3glyrrz, kiuw40mf, ux0q6qc0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1889.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 1,889 123.5 19.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2n2t2pwn",
   "source_url": "https://omlx.ai/benchmarks/performance/2n2t2pwn",
   "tg_tokens": null,
   "tps": 123.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-2n2t2pwn-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: hmtz38es, i3glyrrz, kiuw40mf, ux0q6qc0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1849.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 1,849 77.2 19.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2n2t2pwn",
   "source_url": "https://omlx.ai/benchmarks/performance/2n2t2pwn",
   "tg_tokens": null,
   "tps": 77.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-2n2t2pwn-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: hmtz38es, i3glyrrz, kiuw40mf, ux0q6qc0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1728.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 1,728 213.1 20.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2n2t2pwn",
   "source_url": "https://omlx.ai/benchmarks/performance/2n2t2pwn",
   "tg_tokens": null,
   "tps": 213.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-kncttx1l-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 43evg7c3, 593p2p2v, gn20pw4b, vx8vxdd3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1576.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,576 52.0 16.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kncttx1l",
   "source_url": "https://omlx.ai/benchmarks/performance/kncttx1l",
   "tg_tokens": null,
   "tps": 52.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-kncttx1l-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 43evg7c3, 593p2p2v, gn20pw4b, vx8vxdd3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1641.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 1,641 51.1 17.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kncttx1l",
   "source_url": "https://omlx.ai/benchmarks/performance/kncttx1l",
   "tg_tokens": null,
   "tps": 51.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-kncttx1l-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 43evg7c3, 593p2p2v, gn20pw4b, vx8vxdd3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1836.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 1,836 50.5 17.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kncttx1l",
   "source_url": "https://omlx.ai/benchmarks/performance/kncttx1l",
   "tg_tokens": null,
   "tps": 50.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-kncttx1l-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: 43evg7c3, 593p2p2v, gn20pw4b, vx8vxdd3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1783.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 1,783 48.7 18.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kncttx1l",
   "source_url": "https://omlx.ai/benchmarks/performance/kncttx1l",
   "tg_tokens": null,
   "tps": 48.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-kncttx1l-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 43evg7c3, 593p2p2v, gn20pw4b, vx8vxdd3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1678.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 1,678 46.1 19.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kncttx1l",
   "source_url": "https://omlx.ai/benchmarks/performance/kncttx1l",
   "tg_tokens": null,
   "tps": 46.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-0ysullk3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 17.6 GB; TTFT 1011 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1013.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k tokens 1,013 PP tok/s 49.9 TG tok/s 1011 TTFT (ms) 17.6 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ysullk3",
   "source_url": "https://omlx.ai/benchmarks/performance/0ysullk3",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qxt1wfd2-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 69.6 GB at this context; duplicate traces of the same session: 1kj3enza; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2156.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,156 80.5 69.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qxt1wfd2",
   "source_url": "https://omlx.ai/benchmarks/performance/qxt1wfd2",
   "tg_tokens": null,
   "tps": 80.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qxt1wfd2-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 70.2 GB at this context; duplicate traces of the same session: 1kj3enza; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2952.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,952 76.6 70.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qxt1wfd2",
   "source_url": "https://omlx.ai/benchmarks/performance/qxt1wfd2",
   "tg_tokens": null,
   "tps": 76.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-b20i0ujh-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 50.0 GB; TTFT 8675 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 118.0,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 2bit; 1k tokens 118.0 PP tok/s 9.4 TG tok/s 8675 TTFT (ms) 50.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b20i0ujh",
   "source_url": "https://omlx.ai/benchmarks/performance/b20i0ujh",
   "tg_tokens": null,
   "tps": 9.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qnfsytvc-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP requant build; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: 1clq06hj, 4aywig92, 6xofgbzj, tju11b7k, xdxlge96, ylbidquu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1064.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp_requant on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,064 89.3 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qnfsytvc",
   "source_url": "https://omlx.ai/benchmarks/performance/qnfsytvc",
   "tg_tokens": null,
   "tps": 89.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qnfsytvc-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP requant build; oMLX community benchmark trace; peak memory 72.0 GB at this context; duplicate traces of the same session: 1clq06hj, 4aywig92, 6xofgbzj, tju11b7k, xdxlge96, ylbidquu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2014.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp_requant on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,014 79.1 72.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qnfsytvc",
   "source_url": "https://omlx.ai/benchmarks/performance/qnfsytvc",
   "tg_tokens": null,
   "tps": 79.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qnfsytvc-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP requant build; oMLX community benchmark trace; peak memory 73.0 GB at this context; duplicate traces of the same session: 1clq06hj, 4aywig92, 6xofgbzj, tju11b7k, xdxlge96, ylbidquu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2607.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp_requant on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,607 114.6 73.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qnfsytvc",
   "source_url": "https://omlx.ai/benchmarks/performance/qnfsytvc",
   "tg_tokens": null,
   "tps": 114.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qnfsytvc-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP requant build; oMLX community benchmark trace; peak memory 73.6 GB at this context; duplicate traces of the same session: 1clq06hj, 4aywig92, 6xofgbzj, tju11b7k, xdxlge96, ylbidquu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2561.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp_requant on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,561 82.8 73.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qnfsytvc",
   "source_url": "https://omlx.ai/benchmarks/performance/qnfsytvc",
   "tg_tokens": null,
   "tps": 82.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qnfsytvc-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP requant build; oMLX community benchmark trace; peak memory 74.7 GB at this context; duplicate traces of the same session: 1clq06hj, 4aywig92, 6xofgbzj, tju11b7k, xdxlge96, ylbidquu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2445.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp_requant on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,445 65.6 74.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qnfsytvc",
   "source_url": "https://omlx.ai/benchmarks/performance/qnfsytvc",
   "tg_tokens": null,
   "tps": 65.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qnfsytvc-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP requant build; oMLX community benchmark trace; peak memory 77.0 GB at this context; duplicate traces of the same session: 1clq06hj, 4aywig92, 6xofgbzj, tju11b7k, xdxlge96, ylbidquu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2375.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp_requant on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,375 70.1 77.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qnfsytvc",
   "source_url": "https://omlx.ai/benchmarks/performance/qnfsytvc",
   "tg_tokens": null,
   "tps": 70.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qnfsytvc-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP requant build; oMLX community benchmark trace; peak memory 79.8 GB at this context; duplicate traces of the same session: 1clq06hj, 4aywig92, 6xofgbzj, tju11b7k, xdxlge96, ylbidquu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2390.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp_requant on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,390 62.0 79.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qnfsytvc",
   "source_url": "https://omlx.ai/benchmarks/performance/qnfsytvc",
   "tg_tokens": null,
   "tps": 62.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-siygpyqr-ctx1024",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: pnjp92w8; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 81.5,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 2bit; 1k 81.5 10.4 11.8 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/siygpyqr",
   "source_url": "https://omlx.ai/benchmarks/performance/siygpyqr",
   "tg_tokens": null,
   "tps": 10.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Pro",
   "id": "omlx-siygpyqr-ctx4096",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 14.5 GB at this context; duplicate traces of the same session: pnjp92w8; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 84.3,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 2bit; 4k 84.3 9.2 14.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/siygpyqr",
   "source_url": "https://omlx.ai/benchmarks/performance/siygpyqr",
   "tg_tokens": null,
   "tps": 9.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-47eu6cgv-ctx1024",
   "model": "Ternary-Bonsai-27B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 1k context; peak memory 9.1 GB; TTFT 2500 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 409.5,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-27B-mlx-2bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 2bit; 1k tokens 409.5 PP tok/s 32.0 TG tok/s 2500 TTFT (ms) 9.1 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/47eu6cgv",
   "source_url": "https://omlx.ai/benchmarks/performance/47eu6cgv",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-o7rnbf8h-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 22.1 GB; TTFT 2168 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 472.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "gemma-4-26B-A4B-it-oQ6e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 6bit; 1k tokens 472.4 PP tok/s 35.7 TG tok/s 2168 TTFT (ms) 22.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o7rnbf8h",
   "source_url": "https://omlx.ai/benchmarks/performance/o7rnbf8h",
   "tg_tokens": null,
   "tps": 35.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-r0jivd3q-ctx1024",
   "model": "gpt-oss 120B",
   "notes": "MXFP4 experts with Q8 attention; oMLX community benchmark trace; peak memory 59.5 GB at this context; duplicate traces of the same session: jbsp1ocs; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 885.3,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-120b-MXFP4-Q8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 885.3 93.0 59.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r0jivd3q",
   "source_url": "https://omlx.ai/benchmarks/performance/r0jivd3q",
   "tg_tokens": null,
   "tps": 93.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-r0jivd3q-ctx4096",
   "model": "gpt-oss 120B",
   "notes": "MXFP4 experts with Q8 attention; oMLX community benchmark trace; peak memory 59.6 GB at this context; duplicate traces of the same session: jbsp1ocs; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2061.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-120b-MXFP4-Q8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,061 89.4 59.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r0jivd3q",
   "source_url": "https://omlx.ai/benchmarks/performance/r0jivd3q",
   "tg_tokens": null,
   "tps": 89.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-0jr4lxl0-ctx1024",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: nnramdu0, vztq4ped, zud5odlg; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1331.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,331 64.6 178.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0jr4lxl0",
   "source_url": "https://omlx.ai/benchmarks/performance/0jr4lxl0",
   "tg_tokens": null,
   "tps": 64.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-0jr4lxl0-ctx32768",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 180.5 GB at this context; duplicate traces of the same session: nnramdu0, vztq4ped, zud5odlg; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1928.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,928 56.8 180.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0jr4lxl0",
   "source_url": "https://omlx.ai/benchmarks/performance/0jr4lxl0",
   "tg_tokens": null,
   "tps": 56.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-0jr4lxl0-ctx65536",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 180.9 GB at this context; duplicate traces of the same session: nnramdu0, vztq4ped, zud5odlg; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1893.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,893 51.4 180.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0jr4lxl0",
   "source_url": "https://omlx.ai/benchmarks/performance/0jr4lxl0",
   "tg_tokens": null,
   "tps": 51.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-0jr4lxl0-ctx131072",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 181.6 GB at this context; duplicate traces of the same session: nnramdu0, vztq4ped, zud5odlg; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1867.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 1,867 68.7 181.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0jr4lxl0",
   "source_url": "https://omlx.ai/benchmarks/performance/0jr4lxl0",
   "tg_tokens": null,
   "tps": 68.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-cj1397p5-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: aksflov8, c96ssm48, j0br12m2, k2vyhflt; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 959.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 959.5 108.5 13.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cj1397p5",
   "source_url": "https://omlx.ai/benchmarks/performance/cj1397p5",
   "tg_tokens": null,
   "tps": 108.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-cj1397p5-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: aksflov8, c96ssm48, j0br12m2, k2vyhflt; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1036.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 1,036 105.2 13.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cj1397p5",
   "source_url": "https://omlx.ai/benchmarks/performance/cj1397p5",
   "tg_tokens": null,
   "tps": 105.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-cj1397p5-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.8 GB at this context; duplicate traces of the same session: aksflov8, c96ssm48, j0br12m2, k2vyhflt; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1008.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 1,008 97.9 13.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cj1397p5",
   "source_url": "https://omlx.ai/benchmarks/performance/cj1397p5",
   "tg_tokens": null,
   "tps": 97.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-cj1397p5-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.9 GB at this context; duplicate traces of the same session: aksflov8, c96ssm48, j0br12m2, k2vyhflt; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 950.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 950.6 87.5 13.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cj1397p5",
   "source_url": "https://omlx.ai/benchmarks/performance/cj1397p5",
   "tg_tokens": null,
   "tps": 87.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-cj1397p5-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 14.2 GB at this context; duplicate traces of the same session: aksflov8, c96ssm48, j0br12m2, k2vyhflt; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 851.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 851.4 69.9 14.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cj1397p5",
   "source_url": "https://omlx.ai/benchmarks/performance/cj1397p5",
   "tg_tokens": null,
   "tps": 69.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-cj1397p5-ctx65536",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 14.9 GB at this context; duplicate traces of the same session: aksflov8, c96ssm48, j0br12m2, k2vyhflt; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 698.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 64k 698.1 50.6 14.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cj1397p5",
   "source_url": "https://omlx.ai/benchmarks/performance/cj1397p5",
   "tg_tokens": null,
   "tps": 50.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-cj1397p5-ctx131072",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: aksflov8, c96ssm48, j0br12m2, k2vyhflt; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 514.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 128k 514.6 32.1 16.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cj1397p5",
   "source_url": "https://omlx.ai/benchmarks/performance/cj1397p5",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1i7ufhk3-ctx131072",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 128k context; peak memory 27.1 GB; TTFT 27578 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4753.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 128k tokens 4,753 PP tok/s 74.5 TG tok/s 27578 TTFT (ms) 27.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1i7ufhk3",
   "source_url": "https://omlx.ai/benchmarks/performance/1i7ufhk3",
   "tg_tokens": null,
   "tps": 74.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-9m75xz97-ctx1024",
   "model": "Mellum2-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: 0xt1zhg8; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1327.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Mellum2-12B-A2.5B-Thinking-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 1,327 128.5 7.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9m75xz97",
   "source_url": "https://omlx.ai/benchmarks/performance/9m75xz97",
   "tg_tokens": null,
   "tps": 128.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-9m75xz97-ctx4096",
   "model": "Mellum2-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 7.1 GB at this context; duplicate traces of the same session: 0xt1zhg8; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1218.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Mellum2-12B-A2.5B-Thinking-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 1,218 120.9 7.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9m75xz97",
   "source_url": "https://omlx.ai/benchmarks/performance/9m75xz97",
   "tg_tokens": null,
   "tps": 120.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-le20f0eo-ctx1024",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 9.8 GB at this context; duplicate traces of the same session: pdyckm95; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1097.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-6bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 1k 1,097 105.0 9.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/le20f0eo",
   "source_url": "https://omlx.ai/benchmarks/performance/le20f0eo",
   "tg_tokens": null,
   "tps": 105.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-le20f0eo-ctx4096",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 9.9 GB at this context; duplicate traces of the same session: pdyckm95; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 997.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-6bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 4k 997.4 81.6 9.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/le20f0eo",
   "source_url": "https://omlx.ai/benchmarks/performance/le20f0eo",
   "tg_tokens": null,
   "tps": 81.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-7f3lwbbl-ctx1024",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 9.8 GB at this context; duplicate traces of the same session: lz7sehrv; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1296.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-6bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 1k 1,296 105.7 9.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7f3lwbbl",
   "source_url": "https://omlx.ai/benchmarks/performance/7f3lwbbl",
   "tg_tokens": null,
   "tps": 105.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-7f3lwbbl-ctx4096",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 9.9 GB at this context; duplicate traces of the same session: lz7sehrv; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1127.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-6bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 4k 1,127 87.9 9.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7f3lwbbl",
   "source_url": "https://omlx.ai/benchmarks/performance/7f3lwbbl",
   "tg_tokens": null,
   "tps": 87.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-6fk2tq0o-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.6 GB at this context; duplicate traces of the same session: msif6lys; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 981.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 981.0 67.7 35.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6fk2tq0o",
   "source_url": "https://omlx.ai/benchmarks/performance/6fk2tq0o",
   "tg_tokens": null,
   "tps": 67.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-6fk2tq0o-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: msif6lys; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 885.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 885.6 59.7 37.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6fk2tq0o",
   "source_url": "https://omlx.ai/benchmarks/performance/6fk2tq0o",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-cnonxzcz-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.6 GB at this context; duplicate traces of the same session: t3xvo26x; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 939.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 939.5 69.5 35.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cnonxzcz",
   "source_url": "https://omlx.ai/benchmarks/performance/cnonxzcz",
   "tg_tokens": null,
   "tps": 69.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-cnonxzcz-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: t3xvo26x; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 944.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 944.4 58.4 37.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cnonxzcz",
   "source_url": "https://omlx.ai/benchmarks/performance/cnonxzcz",
   "tg_tokens": null,
   "tps": 58.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-02mlvetf-ctx1024",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 5c5ahnv7; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 983.6,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-oQ3.5e on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 3bit; 1k 983.6 96.2 17.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/02mlvetf",
   "source_url": "https://omlx.ai/benchmarks/performance/02mlvetf",
   "tg_tokens": null,
   "tps": 96.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-02mlvetf-ctx4096",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 5c5ahnv7; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 971.6,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-oQ3.5e on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 3bit; 4k 971.6 89.9 19.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/02mlvetf",
   "source_url": "https://omlx.ai/benchmarks/performance/02mlvetf",
   "tg_tokens": null,
   "tps": 89.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-yn12hmn2-ctx1024",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: tkhj357h; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 950.5,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-oQ3.5e on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 3bit; 1k 950.5 84.7 17.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yn12hmn2",
   "source_url": "https://omlx.ai/benchmarks/performance/yn12hmn2",
   "tg_tokens": null,
   "tps": 84.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-yn12hmn2-ctx4096",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: tkhj357h; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 923.7,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-oQ3.5e on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 3bit; 4k 923.7 82.2 19.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yn12hmn2",
   "source_url": "https://omlx.ai/benchmarks/performance/yn12hmn2",
   "tg_tokens": null,
   "tps": 82.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-rj8xk103-ctx1024",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: il0zsui1; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 829.6,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-oQ3.5e on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 3bit; 1k 829.6 94.3 17.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rj8xk103",
   "source_url": "https://omlx.ai/benchmarks/performance/rj8xk103",
   "tg_tokens": null,
   "tps": 94.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-rj8xk103-ctx4096",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: il0zsui1; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1010.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-oQ3.5e on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 3bit; 4k 1,010 82.5 19.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rj8xk103",
   "source_url": "https://omlx.ai/benchmarks/performance/rj8xk103",
   "tg_tokens": null,
   "tps": 82.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-k5q6libm-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2482.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 2,482 199.5 21.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k5q6libm",
   "source_url": "https://omlx.ai/benchmarks/performance/k5q6libm",
   "tg_tokens": null,
   "tps": 199.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-k5q6libm-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3215.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 3,215 157.4 23.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k5q6libm",
   "source_url": "https://omlx.ai/benchmarks/performance/k5q6libm",
   "tg_tokens": null,
   "tps": 157.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8q8smg61-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: tk6mf8ur, vuhuxal1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4646.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 4,646 169.5 22.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8q8smg61",
   "source_url": "https://omlx.ai/benchmarks/performance/8q8smg61",
   "tg_tokens": null,
   "tps": 169.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8q8smg61-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: tk6mf8ur, vuhuxal1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5401.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 5,401 213.5 22.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8q8smg61",
   "source_url": "https://omlx.ai/benchmarks/performance/8q8smg61",
   "tg_tokens": null,
   "tps": 213.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8q8smg61-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: tk6mf8ur, vuhuxal1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4456.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 4,456 225.4 22.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8q8smg61",
   "source_url": "https://omlx.ai/benchmarks/performance/8q8smg61",
   "tg_tokens": null,
   "tps": 225.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wnajbbhl-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: vxrglsrr, wmsqz8m2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1713.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,713 99.1 22.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wnajbbhl",
   "source_url": "https://omlx.ai/benchmarks/performance/wnajbbhl",
   "tg_tokens": null,
   "tps": 99.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wnajbbhl-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: vxrglsrr, wmsqz8m2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5056.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 5,056 81.5 22.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wnajbbhl",
   "source_url": "https://omlx.ai/benchmarks/performance/wnajbbhl",
   "tg_tokens": null,
   "tps": 81.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wnajbbhl-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: vxrglsrr, wmsqz8m2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4204.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 4,204 187.0 22.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wnajbbhl",
   "source_url": "https://omlx.ai/benchmarks/performance/wnajbbhl",
   "tg_tokens": null,
   "tps": 187.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-qn6om0vt-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 53.5 GB at this context; duplicate traces of the same session: jfkzgftx, m4d5ujyv, sejo7x1y, vhxhz8k9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 304.7,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.8-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 1k 304.7 35.4 53.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qn6om0vt",
   "source_url": "https://omlx.ai/benchmarks/performance/qn6om0vt",
   "tg_tokens": null,
   "tps": 35.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qn6om0vt-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 57.0 GB at this context; duplicate traces of the same session: jfkzgftx, m4d5ujyv, sejo7x1y, vhxhz8k9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 329.3,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.8-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 4k 329.3 28.5 57.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qn6om0vt",
   "source_url": "https://omlx.ai/benchmarks/performance/qn6om0vt",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-qn6om0vt-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 57.7 GB at this context; duplicate traces of the same session: jfkzgftx, m4d5ujyv, sejo7x1y, vhxhz8k9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 325.3,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.8-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 8k 325.3 40.5 57.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qn6om0vt",
   "source_url": "https://omlx.ai/benchmarks/performance/qn6om0vt",
   "tg_tokens": null,
   "tps": 40.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-qn6om0vt-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 58.9 GB at this context; duplicate traces of the same session: jfkzgftx, m4d5ujyv, sejo7x1y, vhxhz8k9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 313.7,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.8-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 16k 313.7 41.7 58.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qn6om0vt",
   "source_url": "https://omlx.ai/benchmarks/performance/qn6om0vt",
   "tg_tokens": null,
   "tps": 41.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-qn6om0vt-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 61.1 GB at this context; duplicate traces of the same session: jfkzgftx, m4d5ujyv, sejo7x1y, vhxhz8k9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 291.9,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.8-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 32k 291.9 68.5 61.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qn6om0vt",
   "source_url": "https://omlx.ai/benchmarks/performance/qn6om0vt",
   "tg_tokens": null,
   "tps": 68.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-4ovp82yl-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 53.5 GB at this context; duplicate traces of the same session: qrlq4q1v; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 305.3,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.8-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 1k 305.3 27.5 53.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ovp82yl",
   "source_url": "https://omlx.ai/benchmarks/performance/4ovp82yl",
   "tg_tokens": null,
   "tps": 27.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4ovp82yl-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 57.0 GB at this context; duplicate traces of the same session: qrlq4q1v; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 329.4,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.8-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 4k 329.4 34.1 57.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ovp82yl",
   "source_url": "https://omlx.ai/benchmarks/performance/4ovp82yl",
   "tg_tokens": null,
   "tps": 34.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4ovp82yl-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 57.7 GB at this context; duplicate traces of the same session: qrlq4q1v; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 325.1,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.8-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 8k 325.1 43.3 57.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ovp82yl",
   "source_url": "https://omlx.ai/benchmarks/performance/4ovp82yl",
   "tg_tokens": null,
   "tps": 43.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4ovp82yl-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 58.9 GB at this context; duplicate traces of the same session: qrlq4q1v; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 313.8,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.8-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 16k 313.8 44.7 58.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ovp82yl",
   "source_url": "https://omlx.ai/benchmarks/performance/4ovp82yl",
   "tg_tokens": null,
   "tps": 44.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4ovp82yl-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 61.1 GB at this context; duplicate traces of the same session: qrlq4q1v; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 292.1,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.8-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 32k 292.1 69.6 61.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ovp82yl",
   "source_url": "https://omlx.ai/benchmarks/performance/4ovp82yl",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-alwrgzxh-ctx1024",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "ET-O5.5-G6A-G4.7-DSV4-K3-RLOO MTP DFlash2 build; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: j2si5ff7; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-ET-O5.5-G6A-G4.7-DSV4-K3-RLOO-MTP-DFlash2-oQ4e on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 125.6 15.5 19.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/alwrgzxh",
   "source_url": "https://omlx.ai/benchmarks/performance/alwrgzxh",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-alwrgzxh-ctx4096",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "ET-O5.5-G6A-G4.7-DSV4-K3-RLOO MTP DFlash2 build; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: j2si5ff7; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-ET-O5.5-G6A-G4.7-DSV4-K3-RLOO-MTP-DFlash2-oQ4e on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 125.4 18.1 23.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/alwrgzxh",
   "source_url": "https://omlx.ai/benchmarks/performance/alwrgzxh",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Max",
   "id": "omlx-gnv5we54-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 51.8 GB at this context; duplicate traces of the same session: p83qiso7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 33.4,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.8-27B-bf16 on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 bf16; 1k 33.4 1.5 51.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gnv5we54",
   "source_url": "https://omlx.ai/benchmarks/performance/gnv5we54",
   "tg_tokens": null,
   "tps": 1.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Max",
   "id": "omlx-gnv5we54-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 55.5 GB at this context; duplicate traces of the same session: p83qiso7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 40.4,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.8-27B-bf16 on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 bf16; 4k 40.4 1.5 55.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gnv5we54",
   "source_url": "https://omlx.ai/benchmarks/performance/gnv5we54",
   "tg_tokens": null,
   "tps": 1.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-vzo2medy-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: yfi7x5sp; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 189.1,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 189.1 27.8 30.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vzo2medy",
   "source_url": "https://omlx.ai/benchmarks/performance/vzo2medy",
   "tg_tokens": null,
   "tps": 27.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-vzo2medy-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.7 GB at this context; duplicate traces of the same session: yfi7x5sp; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 273.1,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 273.1 30.8 33.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vzo2medy",
   "source_url": "https://omlx.ai/benchmarks/performance/vzo2medy",
   "tg_tokens": null,
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-vzo2medy-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: yfi7x5sp; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 274.3,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 274.3 33.0 34.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vzo2medy",
   "source_url": "https://omlx.ai/benchmarks/performance/vzo2medy",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-vzo2medy-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: yfi7x5sp; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 256.7,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 256.7 33.3 36.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vzo2medy",
   "source_url": "https://omlx.ai/benchmarks/performance/vzo2medy",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i6jy63zj-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: riu2q884, z5dxgibb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 646.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 646.2 61.3 18.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i6jy63zj",
   "source_url": "https://omlx.ai/benchmarks/performance/i6jy63zj",
   "tg_tokens": null,
   "tps": 61.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i6jy63zj-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: riu2q884, z5dxgibb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 670.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 670.2 78.6 19.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i6jy63zj",
   "source_url": "https://omlx.ai/benchmarks/performance/i6jy63zj",
   "tg_tokens": null,
   "tps": 78.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i6jy63zj-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: riu2q884, z5dxgibb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 633.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 633.8 117.8 20.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i6jy63zj",
   "source_url": "https://omlx.ai/benchmarks/performance/i6jy63zj",
   "tg_tokens": null,
   "tps": 117.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1qrdw5tr-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 8bluyt1e, 9ikpiyjk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1169.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,169 46.5 19.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1qrdw5tr",
   "source_url": "https://omlx.ai/benchmarks/performance/1qrdw5tr",
   "tg_tokens": null,
   "tps": 46.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1qrdw5tr-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 8bluyt1e, 9ikpiyjk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1145.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,145 65.7 19.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1qrdw5tr",
   "source_url": "https://omlx.ai/benchmarks/performance/1qrdw5tr",
   "tg_tokens": null,
   "tps": 65.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1qrdw5tr-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: 8bluyt1e, 9ikpiyjk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 955.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 955.2 111.0 21.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1qrdw5tr",
   "source_url": "https://omlx.ai/benchmarks/performance/1qrdw5tr",
   "tg_tokens": null,
   "tps": 111.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-194qr5ds-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 40.1 GB; TTFT 664494 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 197.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 128k tokens 197.3 PP tok/s 15.9 TG tok/s 664494 TTFT (ms) 40.1 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/194qr5ds",
   "source_url": "https://omlx.ai/benchmarks/performance/194qr5ds",
   "tg_tokens": null,
   "tps": 15.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-owvdnydt-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.7 GB at this context; duplicate traces of the same session: 20lt6flk, 6iuzvjqg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 879.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k 879.6 44.1 28.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/owvdnydt",
   "source_url": "https://omlx.ai/benchmarks/performance/owvdnydt",
   "tg_tokens": null,
   "tps": 44.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-owvdnydt-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: 20lt6flk, 6iuzvjqg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 841.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 841.4 82.2 29.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/owvdnydt",
   "source_url": "https://omlx.ai/benchmarks/performance/owvdnydt",
   "tg_tokens": null,
   "tps": 82.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-owvdnydt-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: 20lt6flk, 6iuzvjqg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 740.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 64k 740.1 47.6 32.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/owvdnydt",
   "source_url": "https://omlx.ai/benchmarks/performance/owvdnydt",
   "tg_tokens": null,
   "tps": 47.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-owvdnydt-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.8 GB at this context; duplicate traces of the same session: 20lt6flk, 6iuzvjqg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 593.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 128k 593.9 43.9 36.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/owvdnydt",
   "source_url": "https://omlx.ai/benchmarks/performance/owvdnydt",
   "tg_tokens": null,
   "tps": 43.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-twz9v6ki-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.2 GB at this context; duplicate traces of the same session: jxwwn0wc, lo129nc8, qhqtgehh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 245.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 245.4 36.2 39.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/twz9v6ki",
   "source_url": "https://omlx.ai/benchmarks/performance/twz9v6ki",
   "tg_tokens": null,
   "tps": 36.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-twz9v6ki-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 42.4 GB at this context; duplicate traces of the same session: jxwwn0wc, lo129nc8, qhqtgehh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 280.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 280.9 33.5 42.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/twz9v6ki",
   "source_url": "https://omlx.ai/benchmarks/performance/twz9v6ki",
   "tg_tokens": null,
   "tps": 33.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-twz9v6ki-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 42.9 GB at this context; duplicate traces of the same session: jxwwn0wc, lo129nc8, qhqtgehh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 293.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 293.9 36.4 42.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/twz9v6ki",
   "source_url": "https://omlx.ai/benchmarks/performance/twz9v6ki",
   "tg_tokens": null,
   "tps": 36.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-twz9v6ki-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.9 GB at this context; duplicate traces of the same session: jxwwn0wc, lo129nc8, qhqtgehh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 290.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 290.2 45.3 43.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/twz9v6ki",
   "source_url": "https://omlx.ai/benchmarks/performance/twz9v6ki",
   "tg_tokens": null,
   "tps": 45.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-fn5jv3l0-ctx1024",
   "model": "AX-Qwen3.8-27B",
   "notes": "AXQ MTP MLX build; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: uhvshnrz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 36.2,
   "provenance": "community",
   "quant": "6bit",
   "quote": "AX-Qwen3.8-27B-MLX-AXQ-6bit-MTP on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 36.2 2.3 20.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fn5jv3l0",
   "source_url": "https://omlx.ai/benchmarks/performance/fn5jv3l0",
   "tg_tokens": null,
   "tps": 2.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-fn5jv3l0-ctx4096",
   "model": "AX-Qwen3.8-27B",
   "notes": "AXQ MTP MLX build; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: uhvshnrz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 38.2,
   "provenance": "community",
   "quant": "6bit",
   "quote": "AX-Qwen3.8-27B-MLX-AXQ-6bit-MTP on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 38.2 4.0 24.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fn5jv3l0",
   "source_url": "https://omlx.ai/benchmarks/performance/fn5jv3l0",
   "tg_tokens": null,
   "tps": 4.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-tp75ojyn-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.9 GB at this context; duplicate traces of the same session: 2n8s0tht, 52jnoogg, iqp579c3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1347.0,
   "provenance": "community",
   "quant": "mixed 4/8bit",
   "quote": "GLM-5.3-Flash-MLX-mixed-4_8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,347 77.6 174.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tp75ojyn",
   "source_url": "https://omlx.ai/benchmarks/performance/tp75ojyn",
   "tg_tokens": null,
   "tps": 77.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-tp75ojyn-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.3 GB at this context; duplicate traces of the same session: 2n8s0tht, 52jnoogg, iqp579c3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1955.0,
   "provenance": "community",
   "quant": "mixed 4/8bit",
   "quote": "GLM-5.3-Flash-MLX-mixed-4_8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,955 102.9 177.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tp75ojyn",
   "source_url": "https://omlx.ai/benchmarks/performance/tp75ojyn",
   "tg_tokens": null,
   "tps": 102.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-tp75ojyn-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.6 GB at this context; duplicate traces of the same session: 2n8s0tht, 52jnoogg, iqp579c3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1922.0,
   "provenance": "community",
   "quant": "mixed 4/8bit",
   "quote": "GLM-5.3-Flash-MLX-mixed-4_8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,922 102.6 177.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tp75ojyn",
   "source_url": "https://omlx.ai/benchmarks/performance/tp75ojyn",
   "tg_tokens": null,
   "tps": 102.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-tp75ojyn-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.4 GB at this context; duplicate traces of the same session: 2n8s0tht, 52jnoogg, iqp579c3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1886.0,
   "provenance": "community",
   "quant": "mixed 4/8bit",
   "quote": "GLM-5.3-Flash-MLX-mixed-4_8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 1,886 111.7 178.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tp75ojyn",
   "source_url": "https://omlx.ai/benchmarks/performance/tp75ojyn",
   "tg_tokens": null,
   "tps": 111.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-jbdnxtxz-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.9 GB at this context; duplicate traces of the same session: j08p3w4t, r5zade3p; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1350.0,
   "provenance": "community",
   "quant": "mixed 4/8bit",
   "quote": "GLM-5.3-Flash-MLX-mixed-4_8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,350 71.4 174.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jbdnxtxz",
   "source_url": "https://omlx.ai/benchmarks/performance/jbdnxtxz",
   "tg_tokens": null,
   "tps": 71.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-jbdnxtxz-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.3 GB at this context; duplicate traces of the same session: j08p3w4t, r5zade3p; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1938.0,
   "provenance": "community",
   "quant": "mixed 4/8bit",
   "quote": "GLM-5.3-Flash-MLX-mixed-4_8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,938 54.8 177.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jbdnxtxz",
   "source_url": "https://omlx.ai/benchmarks/performance/jbdnxtxz",
   "tg_tokens": null,
   "tps": 54.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-jbdnxtxz-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.6 GB at this context; duplicate traces of the same session: j08p3w4t, r5zade3p; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1905.0,
   "provenance": "community",
   "quant": "mixed 4/8bit",
   "quote": "GLM-5.3-Flash-MLX-mixed-4_8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,905 53.1 177.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jbdnxtxz",
   "source_url": "https://omlx.ai/benchmarks/performance/jbdnxtxz",
   "tg_tokens": null,
   "tps": 53.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-jbdnxtxz-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.4 GB at this context; duplicate traces of the same session: j08p3w4t, r5zade3p; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1870.0,
   "provenance": "community",
   "quant": "mixed 4/8bit",
   "quote": "GLM-5.3-Flash-MLX-mixed-4_8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 1,870 51.1 178.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jbdnxtxz",
   "source_url": "https://omlx.ai/benchmarks/performance/jbdnxtxz",
   "tg_tokens": null,
   "tps": 51.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-8vo2yv2r-ctx1024",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: tfwyy52n, yxel71h0; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1329.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,329 65.8 178.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8vo2yv2r",
   "source_url": "https://omlx.ai/benchmarks/performance/8vo2yv2r",
   "tg_tokens": null,
   "tps": 65.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-8vo2yv2r-ctx32768",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 180.5 GB at this context; duplicate traces of the same session: tfwyy52n, yxel71h0; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1938.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,938 102.8 180.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8vo2yv2r",
   "source_url": "https://omlx.ai/benchmarks/performance/8vo2yv2r",
   "tg_tokens": null,
   "tps": 102.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-8vo2yv2r-ctx65536",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 180.9 GB at this context; duplicate traces of the same session: tfwyy52n, yxel71h0; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1908.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,908 81.7 180.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8vo2yv2r",
   "source_url": "https://omlx.ai/benchmarks/performance/8vo2yv2r",
   "tg_tokens": null,
   "tps": 81.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-61jl627n-ctx4096",
   "model": "Qwen3-48B-A4B-Savant-Commander-Distill-12X-Closed-Open-Heretic-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: fpk2vhep; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 576.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3-48B-A4B-Savant-Commander-Distill-12X-Closed-Open-Heretic-Uncensored-oQ4e on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 4k 576.9 46.4 19.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/61jl627n",
   "source_url": "https://omlx.ai/benchmarks/performance/61jl627n",
   "tg_tokens": null,
   "tps": 46.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-61jl627n-ctx16384",
   "model": "Qwen3-48B-A4B-Savant-Commander-Distill-12X-Closed-Open-Heretic-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: fpk2vhep; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 465.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3-48B-A4B-Savant-Commander-Distill-12X-Closed-Open-Heretic-Uncensored-oQ4e on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 16k 465.6 33.6 21.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/61jl627n",
   "source_url": "https://omlx.ai/benchmarks/performance/61jl627n",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-g1hvpe7k-ctx1024",
   "model": "Qwen3-4B-Instruct-2507",
   "notes": "oMLX community benchmark trace; peak memory 2.8 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1497.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-4B-Instruct-2507-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 1,497 47.9 2.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g1hvpe7k",
   "source_url": "https://omlx.ai/benchmarks/performance/g1hvpe7k",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-g1hvpe7k-ctx4096",
   "model": "Qwen3-4B-Instruct-2507",
   "notes": "oMLX community benchmark trace; peak memory 3.2 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1480.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-4B-Instruct-2507-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 1,480 39.9 3.2 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g1hvpe7k",
   "source_url": "https://omlx.ai/benchmarks/performance/g1hvpe7k",
   "tg_tokens": null,
   "tps": 39.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-j5v3x5ah-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2482.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 2,482 199.5 21.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j5v3x5ah",
   "source_url": "https://omlx.ai/benchmarks/performance/j5v3x5ah",
   "tg_tokens": null,
   "tps": 199.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-j5v3x5ah-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3215.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 3,215 157.4 23.0 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j5v3x5ah",
   "source_url": "https://omlx.ai/benchmarks/performance/j5v3x5ah",
   "tg_tokens": null,
   "tps": 157.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-s4cz1aja-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 53.5 GB at this context; duplicate traces of the same session: 1w78nve0, bqgp8xbg; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 305.3,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.8-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 1k 305.3 27.5 53.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s4cz1aja",
   "source_url": "https://omlx.ai/benchmarks/performance/s4cz1aja",
   "tg_tokens": null,
   "tps": 27.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-s4cz1aja-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 57.0 GB at this context; duplicate traces of the same session: 1w78nve0, bqgp8xbg; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 329.4,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.8-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 4k 329.4 34.1 57.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s4cz1aja",
   "source_url": "https://omlx.ai/benchmarks/performance/s4cz1aja",
   "tg_tokens": null,
   "tps": 34.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-s4cz1aja-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 57.7 GB at this context; duplicate traces of the same session: 1w78nve0, bqgp8xbg; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 325.1,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.8-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 8k 325.1 43.3 57.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s4cz1aja",
   "source_url": "https://omlx.ai/benchmarks/performance/s4cz1aja",
   "tg_tokens": null,
   "tps": 43.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-s4cz1aja-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 58.9 GB at this context; duplicate traces of the same session: 1w78nve0, bqgp8xbg; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 313.8,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.8-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 16k 313.8 44.7 58.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s4cz1aja",
   "source_url": "https://omlx.ai/benchmarks/performance/s4cz1aja",
   "tg_tokens": null,
   "tps": 44.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-s4cz1aja-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 61.1 GB at this context; duplicate traces of the same session: 1w78nve0, bqgp8xbg; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 292.1,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.8-27B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 32k 292.1 69.6 61.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s4cz1aja",
   "source_url": "https://omlx.ai/benchmarks/performance/s4cz1aja",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-m5p9hrbh-ctx1024",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "ET-O5.5-G6A-G4.7-DSV4-K3-RLOO MTP DFlash2 build; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 9v8b8we9; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 126.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-ET-O5.5-G6A-G4.7-DSV4-K3-RLOO-MTP-DFlash2-oQ4e on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 126.7 17.8 19.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m5p9hrbh",
   "source_url": "https://omlx.ai/benchmarks/performance/m5p9hrbh",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-m5p9hrbh-ctx4096",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "ET-O5.5-G6A-G4.7-DSV4-K3-RLOO MTP DFlash2 build; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: 9v8b8we9; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-ET-O5.5-G6A-G4.7-DSV4-K3-RLOO-MTP-DFlash2-oQ4e on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 127.5 17.0 23.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m5p9hrbh",
   "source_url": "https://omlx.ai/benchmarks/performance/m5p9hrbh",
   "tg_tokens": null,
   "tps": 17.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-466ikdcm-ctx1024",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "ET-O5.5-G6A-G4.7-DSV4-K3-RLOO MTP DFlash2 build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: uw73x3rr; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 126.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-ET-O5.5-G6A-G4.7-DSV4-K3-RLOO-MTP-DFlash2-oQ4e on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 126.1 14.1 19.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/466ikdcm",
   "source_url": "https://omlx.ai/benchmarks/performance/466ikdcm",
   "tg_tokens": null,
   "tps": 14.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-466ikdcm-ctx4096",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "ET-O5.5-G6A-G4.7-DSV4-K3-RLOO MTP DFlash2 build; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: uw73x3rr; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 123.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-ET-O5.5-G6A-G4.7-DSV4-K3-RLOO-MTP-DFlash2-oQ4e on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 123.7 14.0 22.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/466ikdcm",
   "source_url": "https://omlx.ai/benchmarks/performance/466ikdcm",
   "tg_tokens": null,
   "tps": 14.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-rt2grqdg-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: 0cevvirs, paz0p0jb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 402.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 402.6 25.4 25.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rt2grqdg",
   "source_url": "https://omlx.ai/benchmarks/performance/rt2grqdg",
   "tg_tokens": null,
   "tps": 25.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-rt2grqdg-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 26.9 GB at this context; duplicate traces of the same session: 0cevvirs, paz0p0jb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 471.2,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 471.2 26.9 26.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rt2grqdg",
   "source_url": "https://omlx.ai/benchmarks/performance/rt2grqdg",
   "tg_tokens": null,
   "tps": 26.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-rt2grqdg-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 27.2 GB at this context; duplicate traces of the same session: 0cevvirs, paz0p0jb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 435.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 8k 435.9 28.7 27.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rt2grqdg",
   "source_url": "https://omlx.ai/benchmarks/performance/rt2grqdg",
   "tg_tokens": null,
   "tps": 28.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-8b4lkmc1-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: id5w8259, ikce1eik; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 449.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 449.0 12.8 22.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8b4lkmc1",
   "source_url": "https://omlx.ai/benchmarks/performance/8b4lkmc1",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-8b4lkmc1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: id5w8259, ikce1eik; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 439.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 439.5 12.8 23.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8b4lkmc1",
   "source_url": "https://omlx.ai/benchmarks/performance/8b4lkmc1",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-8b4lkmc1-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: id5w8259, ikce1eik; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 411.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 8k 411.7 12.8 23.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8b4lkmc1",
   "source_url": "https://omlx.ai/benchmarks/performance/8b4lkmc1",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-eix0n69v-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.8 GB at this context; duplicate traces of the same session: 50rt5995, vpnj5r0x; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 432.2,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 432.2 12.8 29.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eix0n69v",
   "source_url": "https://omlx.ai/benchmarks/performance/eix0n69v",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-eix0n69v-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 30.8 GB at this context; duplicate traces of the same session: 50rt5995, vpnj5r0x; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 513.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 513.1 12.7 30.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eix0n69v",
   "source_url": "https://omlx.ai/benchmarks/performance/eix0n69v",
   "tg_tokens": null,
   "tps": 12.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-eix0n69v-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.1 GB at this context; duplicate traces of the same session: 50rt5995, vpnj5r0x; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 504.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 8k 504.6 12.5 31.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eix0n69v",
   "source_url": "https://omlx.ai/benchmarks/performance/eix0n69v",
   "tg_tokens": null,
   "tps": 12.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-cs0dan2q-ctx4096",
   "model": "Qwen3.8-27B-TTURBO-Cold-Fusion-709-L-Uncensored",
   "notes": "mlx 1M context build; oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: woico1yj; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 160.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TTURBO-Cold-Fusion-709-L-Uncensored-mlx-oQ6e-1M on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 6bit; 4k 160.9 9.9 25.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cs0dan2q",
   "source_url": "https://omlx.ai/benchmarks/performance/cs0dan2q",
   "tg_tokens": null,
   "tps": 9.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-cs0dan2q-ctx16384",
   "model": "Qwen3.8-27B-TTURBO-Cold-Fusion-709-L-Uncensored",
   "notes": "mlx 1M context build; oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: woico1yj; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 153.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-TTURBO-Cold-Fusion-709-L-Uncensored-mlx-oQ6e-1M on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 6bit; 16k 153.2 11.2 28.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cs0dan2q",
   "source_url": "https://omlx.ai/benchmarks/performance/cs0dan2q",
   "tg_tokens": null,
   "tps": 11.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hmuyeri3-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 28.8 GB at this context; duplicate traces of the same session: 4xxbhzpx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 786.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ8e-1M on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 786.9 36.6 28.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hmuyeri3",
   "source_url": "https://omlx.ai/benchmarks/performance/hmuyeri3",
   "tg_tokens": null,
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hmuyeri3-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 30.0 GB at this context; duplicate traces of the same session: 4xxbhzpx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 797.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ8e-1M on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 797.0 35.0 30.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hmuyeri3",
   "source_url": "https://omlx.ai/benchmarks/performance/hmuyeri3",
   "tg_tokens": null,
   "tps": 35.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-a0xii0ib-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 12wp29o8; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 124.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 124.7 20.7 18.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a0xii0ib",
   "source_url": "https://omlx.ai/benchmarks/performance/a0xii0ib",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-a0xii0ib-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 12wp29o8; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 115.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 115.3 19.5 22.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a0xii0ib",
   "source_url": "https://omlx.ai/benchmarks/performance/a0xii0ib",
   "tg_tokens": null,
   "tps": 19.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-0oihwuy1-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: o5fpynsq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 338.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 338.6 34.0 22.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0oihwuy1",
   "source_url": "https://omlx.ai/benchmarks/performance/0oihwuy1",
   "tg_tokens": null,
   "tps": 34.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-0oihwuy1-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: o5fpynsq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 328.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 328.2 43.3 23.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0oihwuy1",
   "source_url": "https://omlx.ai/benchmarks/performance/0oihwuy1",
   "tg_tokens": null,
   "tps": 43.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-972c9qr0-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 3165h7qj, mimy7b46, q7a8zmeo, uys5lo91; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 912.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 912.7 45.5 17.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/972c9qr0",
   "source_url": "https://omlx.ai/benchmarks/performance/972c9qr0",
   "tg_tokens": null,
   "tps": 45.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-972c9qr0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 3165h7qj, mimy7b46, q7a8zmeo, uys5lo91; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1081.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 1,081 57.1 19.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/972c9qr0",
   "source_url": "https://omlx.ai/benchmarks/performance/972c9qr0",
   "tg_tokens": null,
   "tps": 57.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-972c9qr0-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 3165h7qj, mimy7b46, q7a8zmeo, uys5lo91; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1132.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 1,132 33.6 19.4 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/972c9qr0",
   "source_url": "https://omlx.ai/benchmarks/performance/972c9qr0",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-972c9qr0-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 3165h7qj, mimy7b46, q7a8zmeo, uys5lo91; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1100.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 1,100 38.7 19.9 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/972c9qr0",
   "source_url": "https://omlx.ai/benchmarks/performance/972c9qr0",
   "tg_tokens": null,
   "tps": 38.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-972c9qr0-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 3165h7qj, mimy7b46, q7a8zmeo, uys5lo91; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1003.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 32k 1,003 29.5 21.1 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/972c9qr0",
   "source_url": "https://omlx.ai/benchmarks/performance/972c9qr0",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-48hh9ob5-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 879.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k 879.6 44.1 28.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/48hh9ob5",
   "source_url": "https://omlx.ai/benchmarks/performance/48hh9ob5",
   "tg_tokens": null,
   "tps": 44.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-48hh9ob5-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 841.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 841.4 82.2 29.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/48hh9ob5",
   "source_url": "https://omlx.ai/benchmarks/performance/48hh9ob5",
   "tg_tokens": null,
   "tps": 82.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-48hh9ob5-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 740.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 64k 740.1 47.6 32.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/48hh9ob5",
   "source_url": "https://omlx.ai/benchmarks/performance/48hh9ob5",
   "tg_tokens": null,
   "tps": 47.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-48hh9ob5-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 593.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 128k 593.9 43.9 36.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/48hh9ob5",
   "source_url": "https://omlx.ai/benchmarks/performance/48hh9ob5",
   "tg_tokens": null,
   "tps": 43.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-o8eh1omv-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 75.1 GB at this context; duplicate traces of the same session: jmjezvhq, y3apaaxb; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 310.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 310.1 24.5 75.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o8eh1omv",
   "source_url": "https://omlx.ai/benchmarks/performance/o8eh1omv",
   "tg_tokens": null,
   "tps": 24.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-o8eh1omv-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 75.2 GB at this context; duplicate traces of the same session: jmjezvhq, y3apaaxb; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 543.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 543.9 24.2 75.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o8eh1omv",
   "source_url": "https://omlx.ai/benchmarks/performance/o8eh1omv",
   "tg_tokens": null,
   "tps": 24.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-o8eh1omv-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 75.7 GB at this context; duplicate traces of the same session: jmjezvhq, y3apaaxb; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 540.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 540.2 33.4 75.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o8eh1omv",
   "source_url": "https://omlx.ai/benchmarks/performance/o8eh1omv",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-l8zt9cr9-ctx1024",
   "model": "Agents-A1",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 9zvfgepc; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1001.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Agents-A1-OptiQ-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 1,001 78.3 21.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l8zt9cr9",
   "source_url": "https://omlx.ai/benchmarks/performance/l8zt9cr9",
   "tg_tokens": null,
   "tps": 78.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-l8zt9cr9-ctx4096",
   "model": "Agents-A1",
   "notes": "oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: 9zvfgepc; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 946.2,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Agents-A1-OptiQ-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 946.2 68.7 23.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l8zt9cr9",
   "source_url": "https://omlx.ai/benchmarks/performance/l8zt9cr9",
   "tg_tokens": null,
   "tps": 68.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qq591cu2-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.0 GB at this context; duplicate traces of the same session: dwcnyyxp, saodjubn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 88.5,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 88.5 12.5 96.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qq591cu2",
   "source_url": "https://omlx.ai/benchmarks/performance/qq591cu2",
   "tg_tokens": null,
   "tps": 12.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qq591cu2-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 97.6 GB at this context; duplicate traces of the same session: dwcnyyxp, saodjubn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 431.1,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 431.1 12.4 97.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qq591cu2",
   "source_url": "https://omlx.ai/benchmarks/performance/qq591cu2",
   "tg_tokens": null,
   "tps": 12.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qq591cu2-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 98.0 GB at this context; duplicate traces of the same session: dwcnyyxp, saodjubn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 446.3,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 446.3 13.8 98.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qq591cu2",
   "source_url": "https://omlx.ai/benchmarks/performance/qq591cu2",
   "tg_tokens": null,
   "tps": 13.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-b9byjazh-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: hqun69g1; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 168.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ4e on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 4k 168.1 25.6 19.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b9byjazh",
   "source_url": "https://omlx.ai/benchmarks/performance/b9byjazh",
   "tg_tokens": null,
   "tps": 25.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-b9byjazh-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: hqun69g1; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 160.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ4e on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 16k 160.9 32.0 20.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b9byjazh",
   "source_url": "https://omlx.ai/benchmarks/performance/b9byjazh",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-wz227ie8-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.9 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1350.0,
   "provenance": "community",
   "quant": "mixed 4/8bit",
   "quote": "GLM-5.3-Flash-MLX-mixed-4_8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,350 71.4 174.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wz227ie8",
   "source_url": "https://omlx.ai/benchmarks/performance/wz227ie8",
   "tg_tokens": null,
   "tps": 71.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-wz227ie8-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.3 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1938.0,
   "provenance": "community",
   "quant": "mixed 4/8bit",
   "quote": "GLM-5.3-Flash-MLX-mixed-4_8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,938 54.8 177.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wz227ie8",
   "source_url": "https://omlx.ai/benchmarks/performance/wz227ie8",
   "tg_tokens": null,
   "tps": 54.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-wz227ie8-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.6 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1905.0,
   "provenance": "community",
   "quant": "mixed 4/8bit",
   "quote": "GLM-5.3-Flash-MLX-mixed-4_8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,905 53.1 177.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wz227ie8",
   "source_url": "https://omlx.ai/benchmarks/performance/wz227ie8",
   "tg_tokens": null,
   "tps": 53.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-wz227ie8-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 178.4 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1870.0,
   "provenance": "community",
   "quant": "mixed 4/8bit",
   "quote": "GLM-5.3-Flash-MLX-mixed-4_8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 1,870 51.1 178.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wz227ie8",
   "source_url": "https://omlx.ai/benchmarks/performance/wz227ie8",
   "tg_tokens": null,
   "tps": 51.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-gwvtcllu-ctx1024",
   "model": "KAT-Coder-V2.5-Dev",
   "notes": "oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: 6ynyesnl; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 993.8,
   "provenance": "community",
   "quant": "OptiQ 4bit",
   "quote": "KAT-Coder-V2.5-Dev-OptiQ-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 993.8 73.7 21.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gwvtcllu",
   "source_url": "https://omlx.ai/benchmarks/performance/gwvtcllu",
   "tg_tokens": null,
   "tps": 73.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-gwvtcllu-ctx4096",
   "model": "KAT-Coder-V2.5-Dev",
   "notes": "oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: 6ynyesnl; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 962.2,
   "provenance": "community",
   "quant": "OptiQ 4bit",
   "quote": "KAT-Coder-V2.5-Dev-OptiQ-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 962.2 68.9 23.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gwvtcllu",
   "source_url": "https://omlx.ai/benchmarks/performance/gwvtcllu",
   "tg_tokens": null,
   "tps": 68.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-xytt8fgr-ctx1024",
   "model": "Ministral-3-3B-Instruct-2512",
   "notes": "oMLX community benchmark trace; peak memory 4.0 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1161.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ministral-3-3B-Instruct-2512-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,161 85.2 4.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xytt8fgr",
   "source_url": "https://omlx.ai/benchmarks/performance/xytt8fgr",
   "tg_tokens": null,
   "tps": 85.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-xytt8fgr-ctx4096",
   "model": "Ministral-3-3B-Instruct-2512",
   "notes": "oMLX community benchmark trace; peak memory 4.5 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 942.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ministral-3-3B-Instruct-2512-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 942.3 61.3 4.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xytt8fgr",
   "source_url": "https://omlx.ai/benchmarks/performance/xytt8fgr",
   "tg_tokens": null,
   "tps": 61.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-9ztw9gma-ctx1024",
   "model": "Ministral-8B-Instruct-2410",
   "notes": "oMLX community benchmark trace; peak memory 8.7 GB at this context; duplicate traces of the same session: uutvco4e; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 708.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ministral-8B-Instruct-2410-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 708.4 41.8 8.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ztw9gma",
   "source_url": "https://omlx.ai/benchmarks/performance/9ztw9gma",
   "tg_tokens": null,
   "tps": 41.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-9ztw9gma-ctx4096",
   "model": "Ministral-8B-Instruct-2410",
   "notes": "oMLX community benchmark trace; peak memory 9.1 GB at this context; duplicate traces of the same session: uutvco4e; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 510.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ministral-8B-Instruct-2410-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 510.7 31.5 9.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ztw9gma",
   "source_url": "https://omlx.ai/benchmarks/performance/9ztw9gma",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-a3i7cnhx-ctx1024",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: 77626fag; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 872.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 872.6 104.2 18.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a3i7cnhx",
   "source_url": "https://omlx.ai/benchmarks/performance/a3i7cnhx",
   "tg_tokens": null,
   "tps": 104.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-a3i7cnhx-ctx4096",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 77626fag; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 786.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 786.4 87.0 19.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a3i7cnhx",
   "source_url": "https://omlx.ai/benchmarks/performance/a3i7cnhx",
   "tg_tokens": null,
   "tps": 87.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-0b7y84sj-ctx1024",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: dv5rs1as; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 874.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 874.1 104.4 18.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0b7y84sj",
   "source_url": "https://omlx.ai/benchmarks/performance/0b7y84sj",
   "tg_tokens": null,
   "tps": 104.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-0b7y84sj-ctx4096",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: dv5rs1as; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 782.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 782.2 86.4 19.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0b7y84sj",
   "source_url": "https://omlx.ai/benchmarks/performance/0b7y84sj",
   "tg_tokens": null,
   "tps": 86.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-7ex8fjku-ctx1024",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: tnc4zqdi; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 866.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 866.7 103.9 18.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ex8fjku",
   "source_url": "https://omlx.ai/benchmarks/performance/7ex8fjku",
   "tg_tokens": null,
   "tps": 103.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-7ex8fjku-ctx4096",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: tnc4zqdi; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 800.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 800.3 87.2 19.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ex8fjku",
   "source_url": "https://omlx.ai/benchmarks/performance/7ex8fjku",
   "tg_tokens": null,
   "tps": 87.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-j2r9if8m-ctx1024",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: n9fb0u4s; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 873.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 873.5 104.0 18.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j2r9if8m",
   "source_url": "https://omlx.ai/benchmarks/performance/j2r9if8m",
   "tg_tokens": null,
   "tps": 104.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-j2r9if8m-ctx4096",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: n9fb0u4s; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 794.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 794.3 87.9 19.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j2r9if8m",
   "source_url": "https://omlx.ai/benchmarks/performance/j2r9if8m",
   "tg_tokens": null,
   "tps": 87.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-wd9qq5ie-ctx1024",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: klc0xn85; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 868.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 868.7 103.8 18.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wd9qq5ie",
   "source_url": "https://omlx.ai/benchmarks/performance/wd9qq5ie",
   "tg_tokens": null,
   "tps": 103.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-wd9qq5ie-ctx4096",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: klc0xn85; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 804.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 804.0 89.4 19.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wd9qq5ie",
   "source_url": "https://omlx.ai/benchmarks/performance/wd9qq5ie",
   "tg_tokens": null,
   "tps": 89.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-c722ttcv-ctx1024",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: 0cmwrsod; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 845.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 845.9 101.1 18.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c722ttcv",
   "source_url": "https://omlx.ai/benchmarks/performance/c722ttcv",
   "tg_tokens": null,
   "tps": 101.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-c722ttcv-ctx4096",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 0cmwrsod; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 802.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 802.0 90.0 19.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c722ttcv",
   "source_url": "https://omlx.ai/benchmarks/performance/c722ttcv",
   "tg_tokens": null,
   "tps": 90.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-2j7hx80t-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 6jxc7a8o; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1671.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 1,671 71.6 19.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2j7hx80t",
   "source_url": "https://omlx.ai/benchmarks/performance/2j7hx80t",
   "tg_tokens": null,
   "tps": 71.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-2j7hx80t-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 6jxc7a8o; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2174.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 2,174 70.4 19.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2j7hx80t",
   "source_url": "https://omlx.ai/benchmarks/performance/2j7hx80t",
   "tg_tokens": null,
   "tps": 70.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-rnotu67r-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 3b6f7zj1; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 668.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 668.3 71.7 19.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rnotu67r",
   "source_url": "https://omlx.ai/benchmarks/performance/rnotu67r",
   "tg_tokens": null,
   "tps": 71.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-rnotu67r-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 3b6f7zj1; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1461.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,461 70.0 19.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rnotu67r",
   "source_url": "https://omlx.ai/benchmarks/performance/rnotu67r",
   "tg_tokens": null,
   "tps": 70.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-2iqpo5vw-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.6 GB at this context; duplicate traces of the same session: 8gdj0wb6; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1008.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,008 72.7 35.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2iqpo5vw",
   "source_url": "https://omlx.ai/benchmarks/performance/2iqpo5vw",
   "tg_tokens": null,
   "tps": 72.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-2iqpo5vw-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: 8gdj0wb6; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 970.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 970.7 63.7 37.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2iqpo5vw",
   "source_url": "https://omlx.ai/benchmarks/performance/2iqpo5vw",
   "tg_tokens": null,
   "tps": 63.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2",
   "id": "omlx-d4d8t8gy-ctx1024",
   "model": "Qwen2.5-Coder-3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 2.2 GB at this context; duplicate traces of the same session: r384ajzm; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 212.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-3B-Instruct on M2 (8c); M2 (8c) \u00b7 8 GB \u00b7 4bit; 1k 212.2 15.8 2.2 GB view",
   "ram_gb": 8,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d4d8t8gy",
   "source_url": "https://omlx.ai/benchmarks/performance/d4d8t8gy",
   "tg_tokens": null,
   "tps": 15.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2",
   "id": "omlx-d4d8t8gy-ctx4096",
   "model": "Qwen2.5-Coder-3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 2.4 GB at this context; duplicate traces of the same session: r384ajzm; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 81.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-3B-Instruct on M2 (8c); M2 (8c) \u00b7 8 GB \u00b7 4bit; 4k 81.7 13.5 2.4 GB current",
   "ram_gb": 8,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d4d8t8gy",
   "source_url": "https://omlx.ai/benchmarks/performance/d4d8t8gy",
   "tg_tokens": null,
   "tps": 13.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-6tz03qav-ctx1024",
   "model": "Qwen3-4B-Instruct-2507",
   "notes": "oMLX community benchmark trace; peak memory 2.8 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1497.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-4B-Instruct-2507-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 1,497 47.9 2.8 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6tz03qav",
   "source_url": "https://omlx.ai/benchmarks/performance/6tz03qav",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-6tz03qav-ctx4096",
   "model": "Qwen3-4B-Instruct-2507",
   "notes": "oMLX community benchmark trace; peak memory 3.2 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1480.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-4B-Instruct-2507-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 1,480 39.9 3.2 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6tz03qav",
   "source_url": "https://omlx.ai/benchmarks/performance/6tz03qav",
   "tg_tokens": null,
   "tps": 39.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9rtr03u8-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 68.3 GB at this context; duplicate traces of the same session: 8ondvp12, t35o7l4e, trin2wz7, xf217msc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2100.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.6-35B-A3B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 1k 2,100 95.6 68.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9rtr03u8",
   "source_url": "https://omlx.ai/benchmarks/performance/9rtr03u8",
   "tg_tokens": null,
   "tps": 95.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9rtr03u8-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 70.0 GB at this context; duplicate traces of the same session: 8ondvp12, t35o7l4e, trin2wz7, xf217msc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2665.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.6-35B-A3B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 4k 2,665 95.0 70.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9rtr03u8",
   "source_url": "https://omlx.ai/benchmarks/performance/9rtr03u8",
   "tg_tokens": null,
   "tps": 95.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9rtr03u8-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 70.2 GB at this context; duplicate traces of the same session: 8ondvp12, t35o7l4e, trin2wz7, xf217msc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2519.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.6-35B-A3B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 8k 2,519 89.1 70.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9rtr03u8",
   "source_url": "https://omlx.ai/benchmarks/performance/9rtr03u8",
   "tg_tokens": null,
   "tps": 89.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9rtr03u8-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: 8ondvp12, t35o7l4e, trin2wz7, xf217msc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2240.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.6-35B-A3B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 16k 2,240 121.2 70.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9rtr03u8",
   "source_url": "https://omlx.ai/benchmarks/performance/9rtr03u8",
   "tg_tokens": null,
   "tps": 121.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9rtr03u8-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 71.6 GB at this context; duplicate traces of the same session: 8ondvp12, t35o7l4e, trin2wz7, xf217msc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1823.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.6-35B-A3B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 32k 1,823 138.3 71.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9rtr03u8",
   "source_url": "https://omlx.ai/benchmarks/performance/9rtr03u8",
   "tg_tokens": null,
   "tps": 138.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9rtr03u8-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: 8ondvp12, t35o7l4e, trin2wz7, xf217msc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1324.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.6-35B-A3B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 64k 1,324 93.4 73.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9rtr03u8",
   "source_url": "https://omlx.ai/benchmarks/performance/9rtr03u8",
   "tg_tokens": null,
   "tps": 93.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9rtr03u8-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 76.6 GB at this context; duplicate traces of the same session: 8ondvp12, t35o7l4e, trin2wz7, xf217msc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 853.8,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.6-35B-A3B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 128k 853.8 79.5 76.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9rtr03u8",
   "source_url": "https://omlx.ai/benchmarks/performance/9rtr03u8",
   "tg_tokens": null,
   "tps": 79.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9rtr03u8-ctx199680",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 79.3 GB at this context; duplicate traces of the same session: 8ondvp12, t35o7l4e, trin2wz7, xf217msc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 623.8,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.6-35B-A3B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 195k 623.8 58.8 79.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9rtr03u8",
   "source_url": "https://omlx.ai/benchmarks/performance/9rtr03u8",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-r2a5anbr-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: z9c2e2e8; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1456.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 1,456 71.9 21.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r2a5anbr",
   "source_url": "https://omlx.ai/benchmarks/performance/r2a5anbr",
   "tg_tokens": null,
   "tps": 71.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-r2a5anbr-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: z9c2e2e8; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2018.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Genesis-Final-dequantized-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 2,018 65.5 21.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r2a5anbr",
   "source_url": "https://omlx.ai/benchmarks/performance/r2a5anbr",
   "tg_tokens": null,
   "tps": 65.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jn3gqm8p-ctx131072",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace, single run at 128k context; peak memory 46.3 GB; TTFT 248241 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 528.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ8e-1M on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k tokens 528.0 PP tok/s 13.5 TG tok/s 248241 TTFT (ms) 46.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jn3gqm8p",
   "source_url": "https://omlx.ai/benchmarks/performance/jn3gqm8p",
   "tg_tokens": null,
   "tps": 13.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-c32cuzhc-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: vwrcdz7w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 220.8 54.9 18.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c32cuzhc",
   "source_url": "https://omlx.ai/benchmarks/performance/c32cuzhc",
   "tg_tokens": null,
   "tps": 54.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-c32cuzhc-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: vwrcdz7w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 193.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 193.5 35.3 22.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c32cuzhc",
   "source_url": "https://omlx.ai/benchmarks/performance/c32cuzhc",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-kt0cddpg-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 4e35dida, 6zp3cj1h, aqoyj9nu, nhow8epp, s9krs2m6, tn0k09kp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 387.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 387.0 27.8 18.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kt0cddpg",
   "source_url": "https://omlx.ai/benchmarks/performance/kt0cddpg",
   "tg_tokens": null,
   "tps": 27.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-kt0cddpg-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 4e35dida, 6zp3cj1h, aqoyj9nu, nhow8epp, s9krs2m6, tn0k09kp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 392.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 392.0 30.0 19.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kt0cddpg",
   "source_url": "https://omlx.ai/benchmarks/performance/kt0cddpg",
   "tg_tokens": null,
   "tps": 30.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-kt0cddpg-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: 4e35dida, 6zp3cj1h, aqoyj9nu, nhow8epp, s9krs2m6, tn0k09kp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 389.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 389.0 34.4 20.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kt0cddpg",
   "source_url": "https://omlx.ai/benchmarks/performance/kt0cddpg",
   "tg_tokens": null,
   "tps": 34.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-kt0cddpg-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 4e35dida, 6zp3cj1h, aqoyj9nu, nhow8epp, s9krs2m6, tn0k09kp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 360.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 360.4 30.2 21.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kt0cddpg",
   "source_url": "https://omlx.ai/benchmarks/performance/kt0cddpg",
   "tg_tokens": null,
   "tps": 30.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-kt0cddpg-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: 4e35dida, 6zp3cj1h, aqoyj9nu, nhow8epp, s9krs2m6, tn0k09kp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 323.3 30.1 24.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kt0cddpg",
   "source_url": "https://omlx.ai/benchmarks/performance/kt0cddpg",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-kt0cddpg-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.8 GB at this context; duplicate traces of the same session: 4e35dida, 6zp3cj1h, aqoyj9nu, nhow8epp, s9krs2m6, tn0k09kp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 266.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 64k 266.9 19.6 29.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kt0cddpg",
   "source_url": "https://omlx.ai/benchmarks/performance/kt0cddpg",
   "tg_tokens": null,
   "tps": 19.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-kt0cddpg-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 40.6 GB at this context; duplicate traces of the same session: 4e35dida, 6zp3cj1h, aqoyj9nu, nhow8epp, s9krs2m6, tn0k09kp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 193.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 128k 193.5 15.5 40.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kt0cddpg",
   "source_url": "https://omlx.ai/benchmarks/performance/kt0cddpg",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-p6s24z4l-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: kbtq5wid; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 217.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 217.1 33.3 33.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p6s24z4l",
   "source_url": "https://omlx.ai/benchmarks/performance/p6s24z4l",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-p6s24z4l-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.7 GB at this context; duplicate traces of the same session: kbtq5wid; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 187.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 187.4 22.3 37.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p6s24z4l",
   "source_url": "https://omlx.ai/benchmarks/performance/p6s24z4l",
   "tg_tokens": null,
   "tps": 22.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-91obp7vj-ctx1024",
   "model": "Gemma 4 E2B",
   "notes": "fp16 weights; oMLX community benchmark trace, single run at 1k context; peak memory 5.2 GB; TTFT 375 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2728.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "gemma-4-E2B-it-oQ8-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k tokens 2,728 PP tok/s 100.9 TG tok/s 375 TTFT (ms) 5.2 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/91obp7vj",
   "source_url": "https://omlx.ai/benchmarks/performance/91obp7vj",
   "tg_tokens": null,
   "tps": 100.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2",
   "id": "omlx-s70yzf8w-ctx1024",
   "model": "granite-4.2-3b",
   "notes": "q4 build; oMLX community benchmark trace; peak memory 2.6 GB at this context; duplicate traces of the same session: hlqvr02o; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 68.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "granite-4.2-3b-q4 on M2 (8c); M2 (8c) \u00b7 8 GB \u00b7 4bit; 1k 68.5 11.8 2.6 GB current",
   "ram_gb": 8,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s70yzf8w",
   "source_url": "https://omlx.ai/benchmarks/performance/s70yzf8w",
   "tg_tokens": null,
   "tps": 11.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2",
   "id": "omlx-s70yzf8w-ctx4096",
   "model": "granite-4.2-3b",
   "notes": "q4 build; oMLX community benchmark trace; peak memory 2.8 GB at this context; duplicate traces of the same session: hlqvr02o; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 64.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "granite-4.2-3b-q4 on M2 (8c); M2 (8c) \u00b7 8 GB \u00b7 4bit; 4k 64.7 9.8 2.8 GB view",
   "ram_gb": 8,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s70yzf8w",
   "source_url": "https://omlx.ai/benchmarks/performance/s70yzf8w",
   "tg_tokens": null,
   "tps": 9.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-x1nckpob-ctx1024",
   "model": "Agents-A1",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: tscmszti; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 997.8,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Agents-A1-OptiQ-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 997.8 80.5 21.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x1nckpob",
   "source_url": "https://omlx.ai/benchmarks/performance/x1nckpob",
   "tg_tokens": null,
   "tps": 80.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-x1nckpob-ctx4096",
   "model": "Agents-A1",
   "notes": "oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: tscmszti; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1004.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Agents-A1-OptiQ-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 1,004 70.4 23.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x1nckpob",
   "source_url": "https://omlx.ai/benchmarks/performance/x1nckpob",
   "tg_tokens": null,
   "tps": 70.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-z2kffns2-ctx1024",
   "model": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated",
   "notes": "oMLX community benchmark trace; peak memory 36.6 GB at this context; duplicate traces of the same session: zez18ngu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1351.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated-mlx-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 1,351 135.5 36.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z2kffns2",
   "source_url": "https://omlx.ai/benchmarks/performance/z2kffns2",
   "tg_tokens": null,
   "tps": 135.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-z2kffns2-ctx4096",
   "model": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated",
   "notes": "oMLX community benchmark trace; peak memory 37.5 GB at this context; duplicate traces of the same session: zez18ngu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4213.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated-mlx-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 4,213 122.9 37.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z2kffns2",
   "source_url": "https://omlx.ai/benchmarks/performance/z2kffns2",
   "tg_tokens": null,
   "tps": 122.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-32bsx4ba-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: i2xwbaur; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3178.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "LFM2.5-24B-A2B-oQ5e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 3,178 160.6 16.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/32bsx4ba",
   "source_url": "https://omlx.ai/benchmarks/performance/32bsx4ba",
   "tg_tokens": null,
   "tps": 160.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-32bsx4ba-ctx131072",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: i2xwbaur; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3950.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "LFM2.5-24B-A2B-oQ5e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 3,950 85.2 19.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/32bsx4ba",
   "source_url": "https://omlx.ai/benchmarks/performance/32bsx4ba",
   "tg_tokens": null,
   "tps": 85.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-tq4cl88a-ctx1024",
   "model": "Ministral-3-3B-Instruct-2512",
   "notes": "oMLX community benchmark trace; peak memory 4.0 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1161.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ministral-3-3B-Instruct-2512-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,161 85.2 4.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tq4cl88a",
   "source_url": "https://omlx.ai/benchmarks/performance/tq4cl88a",
   "tg_tokens": null,
   "tps": 85.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-tq4cl88a-ctx4096",
   "model": "Ministral-3-3B-Instruct-2512",
   "notes": "oMLX community benchmark trace; peak memory 4.5 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 942.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ministral-3-3B-Instruct-2512-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 942.3 61.3 4.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tq4cl88a",
   "source_url": "https://omlx.ai/benchmarks/performance/tq4cl88a",
   "tg_tokens": null,
   "tps": 61.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-g1em1mx0-ctx1024",
   "model": "Ornith-1.0-35B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: ugzp7e83; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 965.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.0-35B-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 965.3 76.0 20.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g1em1mx0",
   "source_url": "https://omlx.ai/benchmarks/performance/g1em1mx0",
   "tg_tokens": null,
   "tps": 76.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-g1em1mx0-ctx4096",
   "model": "Ornith-1.0-35B",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: ugzp7e83; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1043.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.0-35B-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 1,043 77.4 20.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g1em1mx0",
   "source_url": "https://omlx.ai/benchmarks/performance/g1em1mx0",
   "tg_tokens": null,
   "tps": 77.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-w6vuusqa-ctx1024",
   "model": "Ornith-1.0-35B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: flqjbrto; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 773.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.0-35B-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 773.5 79.9 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w6vuusqa",
   "source_url": "https://omlx.ai/benchmarks/performance/w6vuusqa",
   "tg_tokens": null,
   "tps": 79.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-w6vuusqa-ctx4096",
   "model": "Ornith-1.0-35B",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: flqjbrto; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 996.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.0-35B-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 996.2 77.1 20.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w6vuusqa",
   "source_url": "https://omlx.ai/benchmarks/performance/w6vuusqa",
   "tg_tokens": null,
   "tps": 77.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2",
   "id": "omlx-d0mx8teq-ctx1024",
   "model": "Qwen2.5-Coder-3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 2.2 GB at this context; duplicate traces of the same session: yx8xqqlv; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 129.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-3B-Instruct on M2 (8c); M2 (8c) \u00b7 8 GB \u00b7 4bit; 1k 129.7 15.3 2.2 GB view",
   "ram_gb": 8,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d0mx8teq",
   "source_url": "https://omlx.ai/benchmarks/performance/d0mx8teq",
   "tg_tokens": null,
   "tps": 15.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2",
   "id": "omlx-d0mx8teq-ctx4096",
   "model": "Qwen2.5-Coder-3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 2.4 GB at this context; duplicate traces of the same session: yx8xqqlv; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 83.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-3B-Instruct on M2 (8c); M2 (8c) \u00b7 8 GB \u00b7 4bit; 4k 83.6 13.6 2.4 GB current",
   "ram_gb": 8,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d0mx8teq",
   "source_url": "https://omlx.ai/benchmarks/performance/d0mx8teq",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jv972a07-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 68.3 GB at this context; duplicate traces of the same session: 23e1fi86, q23e1lit; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2100.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.6-35B-A3B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 1k 2,100 95.6 68.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jv972a07",
   "source_url": "https://omlx.ai/benchmarks/performance/jv972a07",
   "tg_tokens": null,
   "tps": 95.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jv972a07-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 70.0 GB at this context; duplicate traces of the same session: 23e1fi86, q23e1lit; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2665.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.6-35B-A3B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 4k 2,665 95.0 70.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jv972a07",
   "source_url": "https://omlx.ai/benchmarks/performance/jv972a07",
   "tg_tokens": null,
   "tps": 95.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jv972a07-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 70.2 GB at this context; duplicate traces of the same session: 23e1fi86, q23e1lit; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2519.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.6-35B-A3B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 8k 2,519 89.1 70.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jv972a07",
   "source_url": "https://omlx.ai/benchmarks/performance/jv972a07",
   "tg_tokens": null,
   "tps": 89.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jv972a07-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: 23e1fi86, q23e1lit; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2240.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.6-35B-A3B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 16k 2,240 121.2 70.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jv972a07",
   "source_url": "https://omlx.ai/benchmarks/performance/jv972a07",
   "tg_tokens": null,
   "tps": 121.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jv972a07-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 71.6 GB at this context; duplicate traces of the same session: 23e1fi86, q23e1lit; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1823.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.6-35B-A3B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 32k 1,823 138.3 71.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jv972a07",
   "source_url": "https://omlx.ai/benchmarks/performance/jv972a07",
   "tg_tokens": null,
   "tps": 138.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jv972a07-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: 23e1fi86, q23e1lit; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1324.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.6-35B-A3B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 64k 1,324 93.4 73.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jv972a07",
   "source_url": "https://omlx.ai/benchmarks/performance/jv972a07",
   "tg_tokens": null,
   "tps": 93.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jv972a07-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 76.6 GB at this context; duplicate traces of the same session: 23e1fi86, q23e1lit; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 853.8,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.6-35B-A3B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 128k 853.8 79.5 76.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jv972a07",
   "source_url": "https://omlx.ai/benchmarks/performance/jv972a07",
   "tg_tokens": null,
   "tps": 79.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jv972a07-ctx199680",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 79.3 GB at this context; duplicate traces of the same session: 23e1fi86, q23e1lit; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 623.8,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.6-35B-A3B on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 195k 623.8 58.8 79.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jv972a07",
   "source_url": "https://omlx.ai/benchmarks/performance/jv972a07",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-0x8zj3zt-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: 88069z7c; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1172.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 16k 1,172 88.2 23.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0x8zj3zt",
   "source_url": "https://omlx.ai/benchmarks/performance/0x8zj3zt",
   "tg_tokens": null,
   "tps": 88.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-0x8zj3zt-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: 88069z7c; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 710.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 64k 710.9 50.2 26.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0x8zj3zt",
   "source_url": "https://omlx.ai/benchmarks/performance/0x8zj3zt",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-t9646mnp-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.8 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1092.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 6bit; 16k 1,092 54.3 32.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t9646mnp",
   "source_url": "https://omlx.ai/benchmarks/performance/t9646mnp",
   "tg_tokens": null,
   "tps": 54.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-t9646mnp-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 619.3,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 6bit; 64k 619.3 57.2 36.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t9646mnp",
   "source_url": "https://omlx.ai/benchmarks/performance/t9646mnp",
   "tg_tokens": null,
   "tps": 57.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-di902yx7-ctx4096",
   "model": "Qwen3.6-40B-Fable-Fusion-6-Core-Deckard-Eleanor-Heretic-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.6 GB at this context; duplicate traces of the same session: 8o3cc3ap; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 138.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-40B-Fable-Fusion-6-Core-Deckard-Eleanor-Heretic-Uncensored-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 4k 138.5 20.8 35.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/di902yx7",
   "source_url": "https://omlx.ai/benchmarks/performance/di902yx7",
   "tg_tokens": null,
   "tps": 20.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-di902yx7-ctx16384",
   "model": "Qwen3.6-40B-Fable-Fusion-6-Core-Deckard-Eleanor-Heretic-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.5 GB at this context; duplicate traces of the same session: 8o3cc3ap; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 128.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-40B-Fable-Fusion-6-Core-Deckard-Eleanor-Heretic-Uncensored-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 16k 128.7 27.1 37.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/di902yx7",
   "source_url": "https://omlx.ai/benchmarks/performance/di902yx7",
   "tg_tokens": null,
   "tps": 27.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-6zlv2j11-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.5 GB at this context; duplicate traces of the same session: gbkz2fg5, sb5mkyc7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 440.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 440.5 30.0 33.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6zlv2j11",
   "source_url": "https://omlx.ai/benchmarks/performance/6zlv2j11",
   "tg_tokens": null,
   "tps": 30.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-6zlv2j11-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.5 GB at this context; duplicate traces of the same session: gbkz2fg5, sb5mkyc7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 497.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 497.5 24.4 34.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6zlv2j11",
   "source_url": "https://omlx.ai/benchmarks/performance/6zlv2j11",
   "tg_tokens": null,
   "tps": 24.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-6zlv2j11-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: gbkz2fg5, sb5mkyc7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 485.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 8k 485.8 30.7 34.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6zlv2j11",
   "source_url": "https://omlx.ai/benchmarks/performance/6zlv2j11",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-mf0lvmli-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.5 GB at this context; duplicate traces of the same session: ac3w6pi4, o1est2o5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 440.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 440.1 31.1 31.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mf0lvmli",
   "source_url": "https://omlx.ai/benchmarks/performance/mf0lvmli",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-mf0lvmli-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.6 GB at this context; duplicate traces of the same session: ac3w6pi4, o1est2o5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 503.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 503.9 27.7 32.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mf0lvmli",
   "source_url": "https://omlx.ai/benchmarks/performance/mf0lvmli",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-mf0lvmli-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.8 GB at this context; duplicate traces of the same session: ac3w6pi4, o1est2o5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 471.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 8k 471.4 30.8 32.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mf0lvmli",
   "source_url": "https://omlx.ai/benchmarks/performance/mf0lvmli",
   "tg_tokens": null,
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-li48mqvx-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: 73qqsvfq, x1bfskxt; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 441.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 441.1 27.6 25.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/li48mqvx",
   "source_url": "https://omlx.ai/benchmarks/performance/li48mqvx",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-li48mqvx-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 26.9 GB at this context; duplicate traces of the same session: 73qqsvfq, x1bfskxt; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 479.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 479.7 27.0 26.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/li48mqvx",
   "source_url": "https://omlx.ai/benchmarks/performance/li48mqvx",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-li48mqvx-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 27.2 GB at this context; duplicate traces of the same session: 73qqsvfq, x1bfskxt; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 446.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 8k 446.7 28.4 27.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/li48mqvx",
   "source_url": "https://omlx.ai/benchmarks/performance/li48mqvx",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-hcdo6mg4-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.7,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-27B-nvfp4 on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 130.7 15.5 16.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hcdo6mg4",
   "source_url": "https://omlx.ai/benchmarks/performance/hcdo6mg4",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-hcdo6mg4-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.6,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-27B-nvfp4 on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 131.6 15.2 17.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hcdo6mg4",
   "source_url": "https://omlx.ai/benchmarks/performance/hcdo6mg4",
   "tg_tokens": null,
   "tps": 15.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-l88cl2oi-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 00f215wo, 6wmmyd0n, aip3b9o3, ih5d3rma, l1xwxe3i, okifqoq3, rywxpf2d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 253.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 253.8 103.8 21.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l88cl2oi",
   "source_url": "https://omlx.ai/benchmarks/performance/l88cl2oi",
   "tg_tokens": null,
   "tps": 103.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-l88cl2oi-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: 00f215wo, 6wmmyd0n, aip3b9o3, ih5d3rma, l1xwxe3i, okifqoq3, rywxpf2d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 338.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 338.2 89.7 25.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l88cl2oi",
   "source_url": "https://omlx.ai/benchmarks/performance/l88cl2oi",
   "tg_tokens": null,
   "tps": 89.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-l88cl2oi-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.6 GB at this context; duplicate traces of the same session: 00f215wo, 6wmmyd0n, aip3b9o3, ih5d3rma, l1xwxe3i, okifqoq3, rywxpf2d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 339.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 339.8 41.6 25.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l88cl2oi",
   "source_url": "https://omlx.ai/benchmarks/performance/l88cl2oi",
   "tg_tokens": null,
   "tps": 41.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-l88cl2oi-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: 00f215wo, 6wmmyd0n, aip3b9o3, ih5d3rma, l1xwxe3i, okifqoq3, rywxpf2d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 323.3 57.2 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l88cl2oi",
   "source_url": "https://omlx.ai/benchmarks/performance/l88cl2oi",
   "tg_tokens": null,
   "tps": 57.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-l88cl2oi-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.3 GB at this context; duplicate traces of the same session: 00f215wo, 6wmmyd0n, aip3b9o3, ih5d3rma, l1xwxe3i, okifqoq3, rywxpf2d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 293.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 293.6 84.2 28.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l88cl2oi",
   "source_url": "https://omlx.ai/benchmarks/performance/l88cl2oi",
   "tg_tokens": null,
   "tps": 84.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-l88cl2oi-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.9 GB at this context; duplicate traces of the same session: 00f215wo, 6wmmyd0n, aip3b9o3, ih5d3rma, l1xwxe3i, okifqoq3, rywxpf2d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 231.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 231.4 45.0 30.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l88cl2oi",
   "source_url": "https://omlx.ai/benchmarks/performance/l88cl2oi",
   "tg_tokens": null,
   "tps": 45.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-l88cl2oi-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.3 GB at this context; duplicate traces of the same session: 00f215wo, 6wmmyd0n, aip3b9o3, ih5d3rma, l1xwxe3i, okifqoq3, rywxpf2d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 163.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 163.6 39.2 32.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l88cl2oi",
   "source_url": "https://omlx.ai/benchmarks/performance/l88cl2oi",
   "tg_tokens": null,
   "tps": 39.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-l88cl2oi-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: 00f215wo, 6wmmyd0n, aip3b9o3, ih5d3rma, l1xwxe3i, okifqoq3, rywxpf2d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 135.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 195k 135.0 34.7 37.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l88cl2oi",
   "source_url": "https://omlx.ai/benchmarks/performance/l88cl2oi",
   "tg_tokens": null,
   "tps": 34.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-9x5tyc28-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 105.8 GB at this context; duplicate traces of the same session: 40f46f42, p0h9jg3h; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 740.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 4k 740.8 58.9 105.8 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9x5tyc28",
   "source_url": "https://omlx.ai/benchmarks/performance/9x5tyc28",
   "tg_tokens": null,
   "tps": 58.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-9x5tyc28-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 106.2 GB at this context; duplicate traces of the same session: 40f46f42, p0h9jg3h; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 823.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 16k 823.6 56.9 106.2 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9x5tyc28",
   "source_url": "https://omlx.ai/benchmarks/performance/9x5tyc28",
   "tg_tokens": null,
   "tps": 56.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-9x5tyc28-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 113.6 GB at this context; duplicate traces of the same session: 40f46f42, p0h9jg3h; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 772.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 195k 772.9 52.1 113.6 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9x5tyc28",
   "source_url": "https://omlx.ai/benchmarks/performance/9x5tyc28",
   "tg_tokens": null,
   "tps": 52.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-663jxd4g-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: 23zjth15, iax68mvw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1211.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,211 66.8 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/663jxd4g",
   "source_url": "https://omlx.ai/benchmarks/performance/663jxd4g",
   "tg_tokens": null,
   "tps": 66.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-663jxd4g-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: 23zjth15, iax68mvw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2050.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,050 88.9 71.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/663jxd4g",
   "source_url": "https://omlx.ai/benchmarks/performance/663jxd4g",
   "tg_tokens": null,
   "tps": 88.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-663jxd4g-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.0 GB at this context; duplicate traces of the same session: 23zjth15, iax68mvw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2696.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,696 57.4 72.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/663jxd4g",
   "source_url": "https://omlx.ai/benchmarks/performance/663jxd4g",
   "tg_tokens": null,
   "tps": 57.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ffyfwedy-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: 0rp0r62a, ee2xchie; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 316.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 316.3 72.6 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ffyfwedy",
   "source_url": "https://omlx.ai/benchmarks/performance/ffyfwedy",
   "tg_tokens": null,
   "tps": 72.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ffyfwedy-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: 0rp0r62a, ee2xchie; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 870.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 870.5 61.8 71.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ffyfwedy",
   "source_url": "https://omlx.ai/benchmarks/performance/ffyfwedy",
   "tg_tokens": null,
   "tps": 61.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ffyfwedy-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.0 GB at this context; duplicate traces of the same session: 0rp0r62a, ee2xchie; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1740.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,740 59.3 72.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ffyfwedy",
   "source_url": "https://omlx.ai/benchmarks/performance/ffyfwedy",
   "tg_tokens": null,
   "tps": 59.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-w0ds9gz5-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "WLSRefit MTP build; oMLX community benchmark trace; peak memory 88.0 GB at this context; duplicate traces of the same session: mu7tlzfk, qvw3kqqh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2008.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-WLSRefit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 2,008 57.7 88.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w0ds9gz5",
   "source_url": "https://omlx.ai/benchmarks/performance/w0ds9gz5",
   "tg_tokens": null,
   "tps": 57.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-w0ds9gz5-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "WLSRefit MTP build; oMLX community benchmark trace; peak memory 90.3 GB at this context; duplicate traces of the same session: mu7tlzfk, qvw3kqqh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1981.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-WLSRefit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 1,981 56.1 90.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w0ds9gz5",
   "source_url": "https://omlx.ai/benchmarks/performance/w0ds9gz5",
   "tg_tokens": null,
   "tps": 56.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-w0ds9gz5-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "WLSRefit MTP build; oMLX community benchmark trace; peak memory 94.0 GB at this context; duplicate traces of the same session: mu7tlzfk, qvw3kqqh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1543.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-WLSRefit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 195k 1,543 49.4 94.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w0ds9gz5",
   "source_url": "https://omlx.ai/benchmarks/performance/w0ds9gz5",
   "tg_tokens": null,
   "tps": 49.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-54275il9-ctx1024",
   "model": "Gemma-4-31B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: tddrrua7, vqz8ekhc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 603.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "gemma-4-31B-it-MLX-5bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 5bit; 1k 603.0 21.8 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/54275il9",
   "source_url": "https://omlx.ai/benchmarks/performance/54275il9",
   "tg_tokens": null,
   "tps": 21.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-54275il9-ctx4096",
   "model": "Gemma-4-31B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: tddrrua7, vqz8ekhc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 636.6,
   "provenance": "community",
   "quant": "5bit",
   "quote": "gemma-4-31B-it-MLX-5bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 5bit; 4k 636.6 21.3 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/54275il9",
   "source_url": "https://omlx.ai/benchmarks/performance/54275il9",
   "tg_tokens": null,
   "tps": 21.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-54275il9-ctx65536",
   "model": "Gemma-4-31B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 30.8 GB at this context; duplicate traces of the same session: tddrrua7, vqz8ekhc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 384.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "gemma-4-31B-it-MLX-5bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 5bit; 64k 384.0 12.7 30.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/54275il9",
   "source_url": "https://omlx.ai/benchmarks/performance/54275il9",
   "tg_tokens": null,
   "tps": 12.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-u2fj1qlw-ctx1024",
   "model": "Qwen3-VL-8B-Instruct",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 5.0 GB at this context; duplicate traces of the same session: z71cf9vi; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 467.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3-VL-8B-Instruct-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 467.8 51.7 5.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u2fj1qlw",
   "source_url": "https://omlx.ai/benchmarks/performance/u2fj1qlw",
   "tg_tokens": null,
   "tps": 51.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-u2fj1qlw-ctx4096",
   "model": "Qwen3-VL-8B-Instruct",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 5.5 GB at this context; duplicate traces of the same session: z71cf9vi; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 444.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3-VL-8B-Instruct-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 444.8 47.3 5.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u2fj1qlw",
   "source_url": "https://omlx.ai/benchmarks/performance/u2fj1qlw",
   "tg_tokens": null,
   "tps": 47.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-5yii4t23-ctx4096",
   "model": "AREX-Turbo",
   "notes": "oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: hh9zmtpk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1374.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "AREX-Turbo-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,374 99.2 3.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5yii4t23",
   "source_url": "https://omlx.ai/benchmarks/performance/5yii4t23",
   "tg_tokens": null,
   "tps": 99.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-5yii4t23-ctx16384",
   "model": "AREX-Turbo",
   "notes": "oMLX community benchmark trace; peak memory 4.4 GB at this context; duplicate traces of the same session: hh9zmtpk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1242.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "AREX-Turbo-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 1,242 90.9 4.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5yii4t23",
   "source_url": "https://omlx.ai/benchmarks/performance/5yii4t23",
   "tg_tokens": null,
   "tps": 90.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wbkcd63n-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 94.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 87.4,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 87.4 12.2 94.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wbkcd63n",
   "source_url": "https://omlx.ai/benchmarks/performance/wbkcd63n",
   "tg_tokens": null,
   "tps": 12.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wbkcd63n-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 95.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 431.2,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 431.2 10.7 95.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wbkcd63n",
   "source_url": "https://omlx.ai/benchmarks/performance/wbkcd63n",
   "tg_tokens": null,
   "tps": 10.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wbkcd63n-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 95.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 447.9,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 447.9 13.3 95.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wbkcd63n",
   "source_url": "https://omlx.ai/benchmarks/performance/wbkcd63n",
   "tg_tokens": null,
   "tps": 13.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-snc0iol9-ctx32768",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 94.2 GB; TTFT 81823 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 400.5,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 32k tokens 400.5 PP tok/s 13.5 TG tok/s 81823 TTFT (ms) 94.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/snc0iol9",
   "source_url": "https://omlx.ai/benchmarks/performance/snc0iol9",
   "tg_tokens": null,
   "tps": 13.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-johuuiqg-ctx1024",
   "model": "Jundot:Qwen3.8-Flash-Next",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 72.5 GB at this context; duplicate traces of the same session: 8zpwgsqk, b7ntjiu9, qcp87eca, vglz8mdu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 430.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot:Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 430.4 27.3 72.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/johuuiqg",
   "source_url": "https://omlx.ai/benchmarks/performance/johuuiqg",
   "tg_tokens": null,
   "tps": 27.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-johuuiqg-ctx4096",
   "model": "Jundot:Qwen3.8-Flash-Next",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: 8zpwgsqk, b7ntjiu9, qcp87eca, vglz8mdu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 915.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot:Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 915.2 52.7 73.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/johuuiqg",
   "source_url": "https://omlx.ai/benchmarks/performance/johuuiqg",
   "tg_tokens": null,
   "tps": 52.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-johuuiqg-ctx8192",
   "model": "Jundot:Qwen3.8-Flash-Next",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 76.9 GB at this context; duplicate traces of the same session: 8zpwgsqk, b7ntjiu9, qcp87eca, vglz8mdu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1662.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot:Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,662 50.2 76.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/johuuiqg",
   "source_url": "https://omlx.ai/benchmarks/performance/johuuiqg",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-johuuiqg-ctx32768",
   "model": "Jundot:Qwen3.8-Flash-Next",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 79.8 GB at this context; duplicate traces of the same session: 8zpwgsqk, b7ntjiu9, qcp87eca, vglz8mdu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2050.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot:Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,050 23.5 79.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/johuuiqg",
   "source_url": "https://omlx.ai/benchmarks/performance/johuuiqg",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-johuuiqg-ctx65536",
   "model": "Jundot:Qwen3.8-Flash-Next",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 81.0 GB at this context; duplicate traces of the same session: 8zpwgsqk, b7ntjiu9, qcp87eca, vglz8mdu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1994.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot:Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 1,994 49.3 81.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/johuuiqg",
   "source_url": "https://omlx.ai/benchmarks/performance/johuuiqg",
   "tg_tokens": null,
   "tps": 49.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-w4g6etbz-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: t6amx1d5; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 734.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 734.6 80.7 13.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w4g6etbz",
   "source_url": "https://omlx.ai/benchmarks/performance/w4g6etbz",
   "tg_tokens": null,
   "tps": 80.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-w4g6etbz-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: t6amx1d5; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 786.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 786.6 76.6 13.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w4g6etbz",
   "source_url": "https://omlx.ai/benchmarks/performance/w4g6etbz",
   "tg_tokens": null,
   "tps": 76.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-7tlxyjjk-ctx16384",
   "model": "Muse-Glimmer-30B-heretic",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 16.3 GB; TTFT 38182 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 429.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-heretic-oQ4e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k tokens 429.1 PP tok/s 16.1 TG tok/s 38182 TTFT (ms) 16.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7tlxyjjk",
   "source_url": "https://omlx.ai/benchmarks/performance/7tlxyjjk",
   "tg_tokens": null,
   "tps": 16.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-cgh3cchm-ctx1024",
   "model": "Occamy-1.0",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: geq39apw; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 847.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Occamy-1.0-oQ8-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 1k 847.5 61.5 35.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cgh3cchm",
   "source_url": "https://omlx.ai/benchmarks/performance/cgh3cchm",
   "tg_tokens": null,
   "tps": 61.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-cgh3cchm-ctx32768",
   "model": "Occamy-1.0",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 39.6 GB at this context; duplicate traces of the same session: geq39apw; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 771.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Occamy-1.0-oQ8-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 32k 771.4 52.1 39.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cgh3cchm",
   "source_url": "https://omlx.ai/benchmarks/performance/cgh3cchm",
   "tg_tokens": null,
   "tps": 52.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-7sxtn2s9-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 4249qnbp, 5pay4oze, jlwt1jsb; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 920.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 920.1 100.2 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7sxtn2s9",
   "source_url": "https://omlx.ai/benchmarks/performance/7sxtn2s9",
   "tg_tokens": null,
   "tps": 100.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-7sxtn2s9-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: 4249qnbp, 5pay4oze, jlwt1jsb; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 527.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 64k 527.9 65.8 27.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7sxtn2s9",
   "source_url": "https://omlx.ai/benchmarks/performance/7sxtn2s9",
   "tg_tokens": null,
   "tps": 65.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-7sxtn2s9-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.0 GB at this context; duplicate traces of the same session: 4249qnbp, 5pay4oze, jlwt1jsb; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 342.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 128k 342.1 44.8 29.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7sxtn2s9",
   "source_url": "https://omlx.ai/benchmarks/performance/7sxtn2s9",
   "tg_tokens": null,
   "tps": 44.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-7sxtn2s9-ctx199680",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.8 GB at this context; duplicate traces of the same session: 4249qnbp, 5pay4oze, jlwt1jsb; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 249.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 195k 249.8 31.9 30.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7sxtn2s9",
   "source_url": "https://omlx.ai/benchmarks/performance/7sxtn2s9",
   "tg_tokens": null,
   "tps": 31.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-8gn3nux6-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.8 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1092.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 6bit; 16k 1,092 54.3 32.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8gn3nux6",
   "source_url": "https://omlx.ai/benchmarks/performance/8gn3nux6",
   "tg_tokens": null,
   "tps": 54.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-8gn3nux6-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 619.3,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 6bit; 64k 619.3 57.2 36.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8gn3nux6",
   "source_url": "https://omlx.ai/benchmarks/performance/8gn3nux6",
   "tg_tokens": null,
   "tps": 57.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-y2f2cwqd-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 28.9 GB; TTFT 852453 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 153.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k tokens 153.8 PP tok/s 16.9 TG tok/s 852453 TTFT (ms) 28.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y2f2cwqd",
   "source_url": "https://omlx.ai/benchmarks/performance/y2f2cwqd",
   "tg_tokens": null,
   "tps": 16.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-sne7bh0q-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 46.6 GB; TTFT 611676 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 214.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 128k tokens 214.3 PP tok/s 29.7 TG tok/s 611676 TTFT (ms) 46.6 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sne7bh0q",
   "source_url": "https://omlx.ai/benchmarks/performance/sne7bh0q",
   "tg_tokens": null,
   "tps": 29.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qnp4av3q-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: jcm8bcut; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 329.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 329.3 36.4 33.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qnp4av3q",
   "source_url": "https://omlx.ai/benchmarks/performance/qnp4av3q",
   "tg_tokens": null,
   "tps": 36.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qnp4av3q-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 37.0 GB at this context; duplicate traces of the same session: jcm8bcut; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 341.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 341.3 38.6 37.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qnp4av3q",
   "source_url": "https://omlx.ai/benchmarks/performance/qnp4av3q",
   "tg_tokens": null,
   "tps": 38.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-f6e4ub5a-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: v01gbiyv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 246.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 246.5 16.3 28.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f6e4ub5a",
   "source_url": "https://omlx.ai/benchmarks/performance/f6e4ub5a",
   "tg_tokens": null,
   "tps": 16.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-f6e4ub5a-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.2 GB at this context; duplicate traces of the same session: v01gbiyv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 207.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 207.2 15.0 32.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f6e4ub5a",
   "source_url": "https://omlx.ai/benchmarks/performance/f6e4ub5a",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-4mdhtdm9-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace, single run at 4k context; peak memory 23.1 GB; TTFT 20487 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 199.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k tokens 199.9 PP tok/s 23.9 TG tok/s 20487 TTFT (ms) 23.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4mdhtdm9",
   "source_url": "https://omlx.ai/benchmarks/performance/4mdhtdm9",
   "tg_tokens": null,
   "tps": 23.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-xijdz9pp-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.7,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-27B-nvfp4 on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 130.7 15.5 16.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xijdz9pp",
   "source_url": "https://omlx.ai/benchmarks/performance/xijdz9pp",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-xijdz9pp-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.6,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-27B-nvfp4 on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 131.6 15.2 17.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xijdz9pp",
   "source_url": "https://omlx.ai/benchmarks/performance/xijdz9pp",
   "tg_tokens": null,
   "tps": 15.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-x149v6ne-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 19.8 GB; TTFT 24748 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 662.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k tokens 662.0 PP tok/s 41.0 TG tok/s 24748 TTFT (ms) 19.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x149v6ne",
   "source_url": "https://omlx.ai/benchmarks/performance/x149v6ne",
   "tg_tokens": null,
   "tps": 41.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-gz2s45ib-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 23.4 GB; TTFT 23858 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 686.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k tokens 686.7 PP tok/s 34.9 TG tok/s 23858 TTFT (ms) 23.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gz2s45ib",
   "source_url": "https://omlx.ai/benchmarks/performance/gz2s45ib",
   "tg_tokens": null,
   "tps": 34.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-az9t3uho-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 21.0 GB; TTFT 24354 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 672.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k tokens 672.7 PP tok/s 41.8 TG tok/s 24354 TTFT (ms) 21.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/az9t3uho",
   "source_url": "https://omlx.ai/benchmarks/performance/az9t3uho",
   "tg_tokens": null,
   "tps": 41.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-64b2rwlh-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 23.4 GB; TTFT 24831 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 659.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k tokens 659.8 PP tok/s 32.3 TG tok/s 24831 TTFT (ms) 23.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/64b2rwlh",
   "source_url": "https://omlx.ai/benchmarks/performance/64b2rwlh",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-lwklvpz0-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 23.1 GB; TTFT 130372 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 502.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k tokens 502.7 PP tok/s 21.1 TG tok/s 130372 TTFT (ms) 23.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lwklvpz0",
   "source_url": "https://omlx.ai/benchmarks/performance/lwklvpz0",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-gymljnsk-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.2 GB at this context; duplicate traces of the same session: 1kduydss, 2u9ou0qq, te4e29cl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 249.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 249.6 37.6 39.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gymljnsk",
   "source_url": "https://omlx.ai/benchmarks/performance/gymljnsk",
   "tg_tokens": null,
   "tps": 37.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-gymljnsk-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 42.4 GB at this context; duplicate traces of the same session: 1kduydss, 2u9ou0qq, te4e29cl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 280.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 280.4 31.2 42.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gymljnsk",
   "source_url": "https://omlx.ai/benchmarks/performance/gymljnsk",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-gymljnsk-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 42.9 GB at this context; duplicate traces of the same session: 1kduydss, 2u9ou0qq, te4e29cl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 292.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 292.2 33.7 42.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gymljnsk",
   "source_url": "https://omlx.ai/benchmarks/performance/gymljnsk",
   "tg_tokens": null,
   "tps": 33.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-gymljnsk-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.9 GB at this context; duplicate traces of the same session: 1kduydss, 2u9ou0qq, te4e29cl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 289.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 289.5 44.2 43.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gymljnsk",
   "source_url": "https://omlx.ai/benchmarks/performance/gymljnsk",
   "tg_tokens": null,
   "tps": 44.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-4keyso4n-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.5 GB at this context; duplicate traces of the same session: bce4g8f0, mnnl7gu2, ufyq2oxf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 253.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 253.1 39.2 37.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4keyso4n",
   "source_url": "https://omlx.ai/benchmarks/performance/4keyso4n",
   "tg_tokens": null,
   "tps": 39.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-4keyso4n-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 40.7 GB at this context; duplicate traces of the same session: bce4g8f0, mnnl7gu2, ufyq2oxf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 286.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 286.4 36.6 40.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4keyso4n",
   "source_url": "https://omlx.ai/benchmarks/performance/4keyso4n",
   "tg_tokens": null,
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-4keyso4n-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 41.0 GB at this context; duplicate traces of the same session: bce4g8f0, mnnl7gu2, ufyq2oxf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 301.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 301.0 36.4 41.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4keyso4n",
   "source_url": "https://omlx.ai/benchmarks/performance/4keyso4n",
   "tg_tokens": null,
   "tps": 36.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-4keyso4n-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 41.9 GB at this context; duplicate traces of the same session: bce4g8f0, mnnl7gu2, ufyq2oxf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 297.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 297.7 45.1 41.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4keyso4n",
   "source_url": "https://omlx.ai/benchmarks/performance/4keyso4n",
   "tg_tokens": null,
   "tps": 45.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7cqt5dra-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 36.7 GB; TTFT 33377 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 981.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k tokens 981.7 PP tok/s 47.3 TG tok/s 33377 TTFT (ms) 36.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7cqt5dra",
   "source_url": "https://omlx.ai/benchmarks/performance/7cqt5dra",
   "tg_tokens": null,
   "tps": 47.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-z5fr2bb3-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 105.6 GB; TTFT 3174 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1290.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 1,290 PP tok/s 45.1 TG tok/s 3174 TTFT (ms) 105.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z5fr2bb3",
   "source_url": "https://omlx.ai/benchmarks/performance/z5fr2bb3",
   "tg_tokens": null,
   "tps": 45.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-sdc08y4p-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 101.3 GB; TTFT 2971 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1379.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k tokens 1,379 PP tok/s 61.8 TG tok/s 2971 TTFT (ms) 101.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sdc08y4p",
   "source_url": "https://omlx.ai/benchmarks/performance/sdc08y4p",
   "tg_tokens": null,
   "tps": 61.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-e7iq1w1x-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 30.0 GB; TTFT 34640 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 236.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k tokens 236.5 PP tok/s 6.6 TG tok/s 34640 TTFT (ms) 30.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e7iq1w1x",
   "source_url": "https://omlx.ai/benchmarks/performance/e7iq1w1x",
   "tg_tokens": null,
   "tps": 6.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-52fp2nk9-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: hvr6bumj; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 84.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 84.8 7.6 29.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/52fp2nk9",
   "source_url": "https://omlx.ai/benchmarks/performance/52fp2nk9",
   "tg_tokens": null,
   "tps": 7.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-52fp2nk9-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.4 GB at this context; duplicate traces of the same session: hvr6bumj; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 231.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 231.8 6.4 31.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/52fp2nk9",
   "source_url": "https://omlx.ai/benchmarks/performance/52fp2nk9",
   "tg_tokens": null,
   "tps": 6.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-azuv7322-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 28.0 GB; TTFT 13093 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 78.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k tokens 78.2 PP tok/s 8.9 TG tok/s 13093 TTFT (ms) 28.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/azuv7322",
   "source_url": "https://omlx.ai/benchmarks/performance/azuv7322",
   "tg_tokens": null,
   "tps": 8.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mbtxv87l-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "WLSRefit MTP build; oMLX community benchmark trace; peak memory 88.0 GB at this context; duplicate traces of the same session: 4q38ky2t, oqu5a242; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2112.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-WLSRefit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 2,112 60.2 88.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mbtxv87l",
   "source_url": "https://omlx.ai/benchmarks/performance/mbtxv87l",
   "tg_tokens": null,
   "tps": 60.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mbtxv87l-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "WLSRefit MTP build; oMLX community benchmark trace; peak memory 90.3 GB at this context; duplicate traces of the same session: 4q38ky2t, oqu5a242; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2036.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-WLSRefit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 2,036 57.4 90.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mbtxv87l",
   "source_url": "https://omlx.ai/benchmarks/performance/mbtxv87l",
   "tg_tokens": null,
   "tps": 57.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mbtxv87l-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "WLSRefit MTP build; oMLX community benchmark trace; peak memory 94.0 GB at this context; duplicate traces of the same session: 4q38ky2t, oqu5a242; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1792.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-WLSRefit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 195k 1,792 54.0 94.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mbtxv87l",
   "source_url": "https://omlx.ai/benchmarks/performance/mbtxv87l",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5xu2wv02-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace, single run at 4k context; peak memory 101.4 GB; TTFT 2994 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1368.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k tokens 1,368 PP tok/s 46.4 TG tok/s 2994 TTFT (ms) 101.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5xu2wv02",
   "source_url": "https://omlx.ai/benchmarks/performance/5xu2wv02",
   "tg_tokens": null,
   "tps": 46.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0j5ur3r4-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace, single run at 128k context; peak memory 103.3 GB; TTFT 72332 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1812.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 128k tokens 1,812 PP tok/s 46.2 TG tok/s 72332 TTFT (ms) 103.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0j5ur3r4",
   "source_url": "https://omlx.ai/benchmarks/performance/0j5ur3r4",
   "tg_tokens": null,
   "tps": 46.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-87bo5t0u-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: fq4eqerm, k3tieyaj, xgfsdkrb; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 572.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 572.6 49.3 21.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/87bo5t0u",
   "source_url": "https://omlx.ai/benchmarks/performance/87bo5t0u",
   "tg_tokens": null,
   "tps": 49.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-87bo5t0u-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: fq4eqerm, k3tieyaj, xgfsdkrb; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 799.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 799.7 53.8 21.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/87bo5t0u",
   "source_url": "https://omlx.ai/benchmarks/performance/87bo5t0u",
   "tg_tokens": null,
   "tps": 53.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-87bo5t0u-ctx8192",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: fq4eqerm, k3tieyaj, xgfsdkrb; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 830.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 830.9 58.4 22.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/87bo5t0u",
   "source_url": "https://omlx.ai/benchmarks/performance/87bo5t0u",
   "tg_tokens": null,
   "tps": 58.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-87bo5t0u-ctx16384",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: fq4eqerm, k3tieyaj, xgfsdkrb; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 736.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 736.1 69.7 22.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/87bo5t0u",
   "source_url": "https://omlx.ai/benchmarks/performance/87bo5t0u",
   "tg_tokens": null,
   "tps": 69.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-p9jvxpqa-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.1 GB at this context; duplicate traces of the same session: 3uxvqscd; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1070.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 1,070 74.0 14.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p9jvxpqa",
   "source_url": "https://omlx.ai/benchmarks/performance/p9jvxpqa",
   "tg_tokens": null,
   "tps": 74.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-p9jvxpqa-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: 3uxvqscd; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2481.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,481 81.9 14.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p9jvxpqa",
   "source_url": "https://omlx.ai/benchmarks/performance/p9jvxpqa",
   "tg_tokens": null,
   "tps": 81.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-p9jvxpqa-ctx8192",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: 3uxvqscd; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2324.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 2,324 79.2 14.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p9jvxpqa",
   "source_url": "https://omlx.ai/benchmarks/performance/p9jvxpqa",
   "tg_tokens": null,
   "tps": 79.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zsiay91f-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 94.0 GB at this context; duplicate traces of the same session: kiyp16n8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 87.4,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 87.4 12.2 94.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zsiay91f",
   "source_url": "https://omlx.ai/benchmarks/performance/zsiay91f",
   "tg_tokens": null,
   "tps": 12.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zsiay91f-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 95.7 GB at this context; duplicate traces of the same session: kiyp16n8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 431.2,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 431.2 10.7 95.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zsiay91f",
   "source_url": "https://omlx.ai/benchmarks/performance/zsiay91f",
   "tg_tokens": null,
   "tps": 10.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zsiay91f-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 95.8 GB at this context; duplicate traces of the same session: kiyp16n8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 447.9,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 447.9 13.3 95.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zsiay91f",
   "source_url": "https://omlx.ai/benchmarks/performance/zsiay91f",
   "tg_tokens": null,
   "tps": 13.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lzvbtmyg-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.0 GB at this context; duplicate traces of the same session: p4jn5z4i, r9vn3d7u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 84.2,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 84.2 13.8 96.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lzvbtmyg",
   "source_url": "https://omlx.ai/benchmarks/performance/lzvbtmyg",
   "tg_tokens": null,
   "tps": 13.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lzvbtmyg-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 97.8 GB at this context; duplicate traces of the same session: p4jn5z4i, r9vn3d7u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 426.2,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 426.2 11.6 97.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lzvbtmyg",
   "source_url": "https://omlx.ai/benchmarks/performance/lzvbtmyg",
   "tg_tokens": null,
   "tps": 11.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lzvbtmyg-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 98.0 GB at this context; duplicate traces of the same session: p4jn5z4i, r9vn3d7u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 449.6,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 449.6 13.6 98.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lzvbtmyg",
   "source_url": "https://omlx.ai/benchmarks/performance/lzvbtmyg",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-h6n3r0u1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: yejf6y29; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 229.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ4e on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 4k 229.5 38.1 24.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h6n3r0u1",
   "source_url": "https://omlx.ai/benchmarks/performance/h6n3r0u1",
   "tg_tokens": null,
   "tps": 38.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-h6n3r0u1-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: yejf6y29; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 213.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ4e on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 16k 213.1 43.4 26.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h6n3r0u1",
   "source_url": "https://omlx.ai/benchmarks/performance/h6n3r0u1",
   "tg_tokens": null,
   "tps": 43.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-aortlei3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 6pjxqr1s; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 166.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ4e on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 4k 166.8 24.5 19.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aortlei3",
   "source_url": "https://omlx.ai/benchmarks/performance/aortlei3",
   "tg_tokens": null,
   "tps": 24.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-aortlei3-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: 6pjxqr1s; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 159.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ4e on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 16k 159.4 31.0 20.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aortlei3",
   "source_url": "https://omlx.ai/benchmarks/performance/aortlei3",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-eisb3p3c-ctx1024",
   "model": "Jundot:Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: 659d08wb, egr6yjig, m84cba27; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 573.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot:Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 573.5 43.9 70.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eisb3p3c",
   "source_url": "https://omlx.ai/benchmarks/performance/eisb3p3c",
   "tg_tokens": null,
   "tps": 43.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-eisb3p3c-ctx4096",
   "model": "Jundot:Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: 659d08wb, egr6yjig, m84cba27; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 831.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot:Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 831.8 51.8 71.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eisb3p3c",
   "source_url": "https://omlx.ai/benchmarks/performance/eisb3p3c",
   "tg_tokens": null,
   "tps": 51.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-eisb3p3c-ctx8192",
   "model": "Jundot:Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: 659d08wb, egr6yjig, m84cba27; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1055.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot:Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,055 57.1 71.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eisb3p3c",
   "source_url": "https://omlx.ai/benchmarks/performance/eisb3p3c",
   "tg_tokens": null,
   "tps": 57.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-eisb3p3c-ctx16384",
   "model": "Jundot:Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.0 GB at this context; duplicate traces of the same session: 659d08wb, egr6yjig, m84cba27; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1061.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot:Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 1,061 49.4 72.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eisb3p3c",
   "source_url": "https://omlx.ai/benchmarks/performance/eisb3p3c",
   "tg_tokens": null,
   "tps": 49.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-8pmx7dfq-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Jundot MTP build; oMLX community benchmark trace; peak memory 129.4 GB at this context; duplicate traces of the same session: 0wq5j40u, 3lgfvc8a, h8mm4jqr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 499.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Jundot:Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 499.3 53.2 129.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8pmx7dfq",
   "source_url": "https://omlx.ai/benchmarks/performance/8pmx7dfq",
   "tg_tokens": null,
   "tps": 53.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-8pmx7dfq-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Jundot MTP build; oMLX community benchmark trace; peak memory 130.0 GB at this context; duplicate traces of the same session: 0wq5j40u, 3lgfvc8a, h8mm4jqr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 780.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Jundot:Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 780.7 25.1 130.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8pmx7dfq",
   "source_url": "https://omlx.ai/benchmarks/performance/8pmx7dfq",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-8pmx7dfq-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Jundot MTP build; oMLX community benchmark trace; peak memory 130.4 GB at this context; duplicate traces of the same session: 0wq5j40u, 3lgfvc8a, h8mm4jqr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1022.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Jundot:Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 1,022 57.4 130.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8pmx7dfq",
   "source_url": "https://omlx.ai/benchmarks/performance/8pmx7dfq",
   "tg_tokens": null,
   "tps": 57.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-8pmx7dfq-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "Jundot MTP build; oMLX community benchmark trace; peak memory 130.7 GB at this context; duplicate traces of the same session: 0wq5j40u, 3lgfvc8a, h8mm4jqr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1022.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Jundot:Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 16k 1,022 51.7 130.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8pmx7dfq",
   "source_url": "https://omlx.ai/benchmarks/performance/8pmx7dfq",
   "tg_tokens": null,
   "tps": 51.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-fzc4wqfd-ctx1024",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 396.2,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-oQ5-fp16-mtp on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 5bit; 1k 396.2 39.7 24.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fzc4wqfd",
   "source_url": "https://omlx.ai/benchmarks/performance/fzc4wqfd",
   "tg_tokens": null,
   "tps": 39.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-fzc4wqfd-ctx4096",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 500.9,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-oQ5-fp16-mtp on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 5bit; 4k 500.9 36.9 25.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fzc4wqfd",
   "source_url": "https://omlx.ai/benchmarks/performance/fzc4wqfd",
   "tg_tokens": null,
   "tps": 36.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-ugzz6n3v-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 28.7 GB; TTFT 1216 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 842.4,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k tokens 842.4 PP tok/s 25.4 TG tok/s 1216 TTFT (ms) 28.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ugzz6n3v",
   "source_url": "https://omlx.ai/benchmarks/performance/ugzz6n3v",
   "tg_tokens": null,
   "tps": 25.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-18nyl2pd-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 28.7 GB; TTFT 1232 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 831.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k tokens 831.0 PP tok/s 38.5 TG tok/s 1232 TTFT (ms) 28.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/18nyl2pd",
   "source_url": "https://omlx.ai/benchmarks/performance/18nyl2pd",
   "tg_tokens": null,
   "tps": 38.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-prx9lh7s-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: dz6a62ro, lhzzi0xy, xmnmurxv; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 1k 102.2 17.4 17.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/prx9lh7s",
   "source_url": "https://omlx.ai/benchmarks/performance/prx9lh7s",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-prx9lh7s-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: dz6a62ro, lhzzi0xy, xmnmurxv; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 4k 103.9 18.6 18.7 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/prx9lh7s",
   "source_url": "https://omlx.ai/benchmarks/performance/prx9lh7s",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-prx9lh7s-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: dz6a62ro, lhzzi0xy, xmnmurxv; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 64.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 8k 64.5 5.1 19.7 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/prx9lh7s",
   "source_url": "https://omlx.ai/benchmarks/performance/prx9lh7s",
   "tg_tokens": null,
   "tps": 5.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Pro",
   "id": "omlx-prx9lh7s-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: dz6a62ro, lhzzi0xy, xmnmurxv; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 47.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 16k 47.1 3.5 21.4 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/prx9lh7s",
   "source_url": "https://omlx.ai/benchmarks/performance/prx9lh7s",
   "tg_tokens": null,
   "tps": 3.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-s6pnph5x-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 28.9 GB; TTFT 863752 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 151.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k tokens 151.7 PP tok/s 17.3 TG tok/s 863752 TTFT (ms) 28.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s6pnph5x",
   "source_url": "https://omlx.ai/benchmarks/performance/s6pnph5x",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-6zsyl0uw-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 19.8 GB; TTFT 24484 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 669.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k tokens 669.2 PP tok/s 39.4 TG tok/s 24484 TTFT (ms) 19.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6zsyl0uw",
   "source_url": "https://omlx.ai/benchmarks/performance/6zsyl0uw",
   "tg_tokens": null,
   "tps": 39.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-cqxd9lnd-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.2 GB at this context; duplicate traces of the same session: 3kyob3gd, 56hpmo7f, 59oznfk6, ifjvlu3p, loo6jk33, q9bnnp39; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 243.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 243.8 39.3 33.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cqxd9lnd",
   "source_url": "https://omlx.ai/benchmarks/performance/cqxd9lnd",
   "tg_tokens": null,
   "tps": 39.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-cqxd9lnd-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.4 GB at this context; duplicate traces of the same session: 3kyob3gd, 56hpmo7f, 59oznfk6, ifjvlu3p, loo6jk33, q9bnnp39; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 332.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 332.0 32.5 36.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cqxd9lnd",
   "source_url": "https://omlx.ai/benchmarks/performance/cqxd9lnd",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-cqxd9lnd-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.0 GB at this context; duplicate traces of the same session: 3kyob3gd, 56hpmo7f, 59oznfk6, ifjvlu3p, loo6jk33, q9bnnp39; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 336.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 336.5 43.8 37.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cqxd9lnd",
   "source_url": "https://omlx.ai/benchmarks/performance/cqxd9lnd",
   "tg_tokens": null,
   "tps": 43.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-cqxd9lnd-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.0 GB at this context; duplicate traces of the same session: 3kyob3gd, 56hpmo7f, 59oznfk6, ifjvlu3p, loo6jk33, q9bnnp39; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 320.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 320.5 49.3 38.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cqxd9lnd",
   "source_url": "https://omlx.ai/benchmarks/performance/cqxd9lnd",
   "tg_tokens": null,
   "tps": 49.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-cqxd9lnd-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.7 GB at this context; duplicate traces of the same session: 3kyob3gd, 56hpmo7f, 59oznfk6, ifjvlu3p, loo6jk33, q9bnnp39; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 291.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 291.6 62.6 39.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cqxd9lnd",
   "source_url": "https://omlx.ai/benchmarks/performance/cqxd9lnd",
   "tg_tokens": null,
   "tps": 62.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-cqxd9lnd-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 42.4 GB at this context; duplicate traces of the same session: 3kyob3gd, 56hpmo7f, 59oznfk6, ifjvlu3p, loo6jk33, q9bnnp39; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 246.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k 246.6 35.8 42.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cqxd9lnd",
   "source_url": "https://omlx.ai/benchmarks/performance/cqxd9lnd",
   "tg_tokens": null,
   "tps": 35.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-cqxd9lnd-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 46.9 GB at this context; duplicate traces of the same session: 3kyob3gd, 56hpmo7f, 59oznfk6, ifjvlu3p, loo6jk33, q9bnnp39; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 189.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 128k 189.2 32.2 46.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cqxd9lnd",
   "source_url": "https://omlx.ai/benchmarks/performance/cqxd9lnd",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-63flizik-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 26.6 GB; TTFT 71016 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 115.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 8k tokens 115.4 PP tok/s 6.6 TG tok/s 71016 TTFT (ms) 26.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/63flizik",
   "source_url": "https://omlx.ai/benchmarks/performance/63flizik",
   "tg_tokens": null,
   "tps": 6.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-23ug01jt-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 33.3 GB; TTFT 341470 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 96.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k tokens 96.0 PP tok/s 7.9 TG tok/s 341470 TTFT (ms) 33.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/23ug01jt",
   "source_url": "https://omlx.ai/benchmarks/performance/23ug01jt",
   "tg_tokens": null,
   "tps": 7.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-oprubic0-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 86.4 GB at this context; duplicate traces of the same session: deeg3mmp, qk1b8yxc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 570.6,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 1k 570.6 22.6 86.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oprubic0",
   "source_url": "https://omlx.ai/benchmarks/performance/oprubic0",
   "tg_tokens": null,
   "tps": 22.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-oprubic0-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 87.0 GB at this context; duplicate traces of the same session: deeg3mmp, qk1b8yxc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 804.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 4k 804.1 49.1 87.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oprubic0",
   "source_url": "https://omlx.ai/benchmarks/performance/oprubic0",
   "tg_tokens": null,
   "tps": 49.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-oprubic0-ctx8192",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 87.4 GB at this context; duplicate traces of the same session: deeg3mmp, qk1b8yxc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1031.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 5bit; 8k 1,031 61.8 87.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oprubic0",
   "source_url": "https://omlx.ai/benchmarks/performance/oprubic0",
   "tg_tokens": null,
   "tps": 61.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-5egqi1o5-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 72.4 GB at this context; duplicate traces of the same session: gb20ia0s; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 227.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 227.2 21.2 72.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5egqi1o5",
   "source_url": "https://omlx.ai/benchmarks/performance/5egqi1o5",
   "tg_tokens": null,
   "tps": 21.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-5egqi1o5-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 74.8 GB at this context; duplicate traces of the same session: gb20ia0s; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 287.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 287.6 29.7 74.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5egqi1o5",
   "source_url": "https://omlx.ai/benchmarks/performance/5egqi1o5",
   "tg_tokens": null,
   "tps": 29.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zip12ohl-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: 8nghnqv1, b1zbexv6, gz0b9fmg, hhv02g7r, s2ohhzfj, tbq52mm3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1169.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,169 81.7 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zip12ohl",
   "source_url": "https://omlx.ai/benchmarks/performance/zip12ohl",
   "tg_tokens": null,
   "tps": 81.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zip12ohl-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: 8nghnqv1, b1zbexv6, gz0b9fmg, hhv02g7r, s2ohhzfj, tbq52mm3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2012.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,012 60.2 71.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zip12ohl",
   "source_url": "https://omlx.ai/benchmarks/performance/zip12ohl",
   "tg_tokens": null,
   "tps": 60.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zip12ohl-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.0 GB at this context; duplicate traces of the same session: 8nghnqv1, b1zbexv6, gz0b9fmg, hhv02g7r, s2ohhzfj, tbq52mm3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2632.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,632 74.3 72.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zip12ohl",
   "source_url": "https://omlx.ai/benchmarks/performance/zip12ohl",
   "tg_tokens": null,
   "tps": 74.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zip12ohl-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.0 GB at this context; duplicate traces of the same session: 8nghnqv1, b1zbexv6, gz0b9fmg, hhv02g7r, s2ohhzfj, tbq52mm3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2606.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,606 62.4 73.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zip12ohl",
   "source_url": "https://omlx.ai/benchmarks/performance/zip12ohl",
   "tg_tokens": null,
   "tps": 62.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zip12ohl-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.7 GB at this context; duplicate traces of the same session: 8nghnqv1, b1zbexv6, gz0b9fmg, hhv02g7r, s2ohhzfj, tbq52mm3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2420.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,420 76.7 74.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zip12ohl",
   "source_url": "https://omlx.ai/benchmarks/performance/zip12ohl",
   "tg_tokens": null,
   "tps": 76.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-zip12ohl-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 77.0 GB at this context; duplicate traces of the same session: 8nghnqv1, b1zbexv6, gz0b9fmg, hhv02g7r, s2ohhzfj, tbq52mm3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2417.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,417 71.0 77.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zip12ohl",
   "source_url": "https://omlx.ai/benchmarks/performance/zip12ohl",
   "tg_tokens": null,
   "tps": 71.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zip12ohl-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.8 GB at this context; duplicate traces of the same session: 8nghnqv1, b1zbexv6, gz0b9fmg, hhv02g7r, s2ohhzfj, tbq52mm3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2237.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,237 60.9 79.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zip12ohl",
   "source_url": "https://omlx.ai/benchmarks/performance/zip12ohl",
   "tg_tokens": null,
   "tps": 60.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-70xmjmw9-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "WLSRefit MTP build; oMLX community benchmark trace; peak memory 90.6 GB at this context; duplicate traces of the same session: okpvw9af, qgc038cv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2028.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-WLSRefit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 2,028 65.5 90.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/70xmjmw9",
   "source_url": "https://omlx.ai/benchmarks/performance/70xmjmw9",
   "tg_tokens": null,
   "tps": 65.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-70xmjmw9-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "WLSRefit MTP build; oMLX community benchmark trace; peak memory 94.6 GB at this context; duplicate traces of the same session: okpvw9af, qgc038cv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2010.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-WLSRefit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 2,010 63.4 94.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/70xmjmw9",
   "source_url": "https://omlx.ai/benchmarks/performance/70xmjmw9",
   "tg_tokens": null,
   "tps": 63.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-70xmjmw9-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "WLSRefit MTP build; oMLX community benchmark trace; peak memory 97.0 GB at this context; duplicate traces of the same session: okpvw9af, qgc038cv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1841.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-WLSRefit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 195k 1,841 55.9 97.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/70xmjmw9",
   "source_url": "https://omlx.ai/benchmarks/performance/70xmjmw9",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8heo5yqg-ctx1024",
   "model": "Signal-3.8-27B-Terse-Coder",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: d3sd9f2f, ytn6dco8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 858.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Signal-3.8-27B-Terse-Coder-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 858.7 50.8 17.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8heo5yqg",
   "source_url": "https://omlx.ai/benchmarks/performance/8heo5yqg",
   "tg_tokens": null,
   "tps": 50.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8heo5yqg-ctx4096",
   "model": "Signal-3.8-27B-Terse-Coder",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: d3sd9f2f, ytn6dco8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1256.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Signal-3.8-27B-Terse-Coder-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,256 55.5 19.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8heo5yqg",
   "source_url": "https://omlx.ai/benchmarks/performance/8heo5yqg",
   "tg_tokens": null,
   "tps": 55.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8heo5yqg-ctx8192",
   "model": "Signal-3.8-27B-Terse-Coder",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: d3sd9f2f, ytn6dco8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1347.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Signal-3.8-27B-Terse-Coder-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,347 81.8 19.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8heo5yqg",
   "source_url": "https://omlx.ai/benchmarks/performance/8heo5yqg",
   "tg_tokens": null,
   "tps": 81.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8heo5yqg-ctx16384",
   "model": "Signal-3.8-27B-Terse-Coder",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: d3sd9f2f, ytn6dco8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1288.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Signal-3.8-27B-Terse-Coder-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,288 79.0 19.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8heo5yqg",
   "source_url": "https://omlx.ai/benchmarks/performance/8heo5yqg",
   "tg_tokens": null,
   "tps": 79.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-pzd6rc22-ctx1024",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: 2012wd9i, r2hj284u; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 906.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 906.4 51.7 25.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pzd6rc22",
   "source_url": "https://omlx.ai/benchmarks/performance/pzd6rc22",
   "tg_tokens": null,
   "tps": 51.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-pzd6rc22-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: 2012wd9i, r2hj284u; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2024.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 2,024 51.4 26.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pzd6rc22",
   "source_url": "https://omlx.ai/benchmarks/performance/pzd6rc22",
   "tg_tokens": null,
   "tps": 51.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-pzd6rc22-ctx8192",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: 2012wd9i, r2hj284u; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2227.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 2,227 48.0 26.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pzd6rc22",
   "source_url": "https://omlx.ai/benchmarks/performance/pzd6rc22",
   "tg_tokens": null,
   "tps": 48.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-w3l5o3xi-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1070.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 1,070 74.0 14.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w3l5o3xi",
   "source_url": "https://omlx.ai/benchmarks/performance/w3l5o3xi",
   "tg_tokens": null,
   "tps": 74.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-w3l5o3xi-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2481.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,481 81.9 14.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w3l5o3xi",
   "source_url": "https://omlx.ai/benchmarks/performance/w3l5o3xi",
   "tg_tokens": null,
   "tps": 81.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-w3l5o3xi-ctx8192",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2324.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 2,324 79.2 14.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w3l5o3xi",
   "source_url": "https://omlx.ai/benchmarks/performance/w3l5o3xi",
   "tg_tokens": null,
   "tps": 79.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-3bk1r2vk-ctx1024",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: hhab60i8, m2s84w4w; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1770.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "gemma-4-26b-a4b-it-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 1,770 59.6 19.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3bk1r2vk",
   "source_url": "https://omlx.ai/benchmarks/performance/3bk1r2vk",
   "tg_tokens": null,
   "tps": 59.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-3bk1r2vk-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: hhab60i8, m2s84w4w; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2091.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "gemma-4-26b-a4b-it-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 2,091 61.1 20.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3bk1r2vk",
   "source_url": "https://omlx.ai/benchmarks/performance/3bk1r2vk",
   "tg_tokens": null,
   "tps": 61.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-3bk1r2vk-ctx8192",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: hhab60i8, m2s84w4w; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2182.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "gemma-4-26b-a4b-it-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 8k 2,182 55.7 20.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3bk1r2vk",
   "source_url": "https://omlx.ai/benchmarks/performance/3bk1r2vk",
   "tg_tokens": null,
   "tps": 55.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-djwlv1it-ctx4096",
   "model": "LFM2.5-2.6B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 2.6 GB at this context; duplicate traces of the same session: 6h0orffq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1325.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-2.6B-MLX-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 1,325 128.8 2.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/djwlv1it",
   "source_url": "https://omlx.ai/benchmarks/performance/djwlv1it",
   "tg_tokens": null,
   "tps": 128.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-djwlv1it-ctx16384",
   "model": "LFM2.5-2.6B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 2.8 GB at this context; duplicate traces of the same session: 6h0orffq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1249.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-2.6B-MLX-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 1,249 100.8 2.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/djwlv1it",
   "source_url": "https://omlx.ai/benchmarks/performance/djwlv1it",
   "tg_tokens": null,
   "tps": 100.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-djwlv1it-ctx65536",
   "model": "LFM2.5-2.6B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: 6h0orffq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 922.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-2.6B-MLX-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 922.1 67.8 3.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/djwlv1it",
   "source_url": "https://omlx.ai/benchmarks/performance/djwlv1it",
   "tg_tokens": null,
   "tps": 67.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-zj4tatld-ctx4096",
   "model": "LFM2.5-2.6B",
   "notes": "MLX fp16 build; oMLX community benchmark trace; peak memory 2.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1309.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-2.6B-MLX-oQ4e-fp16 on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 1,309 125.0 2.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zj4tatld",
   "source_url": "https://omlx.ai/benchmarks/performance/zj4tatld",
   "tg_tokens": null,
   "tps": 125.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-zj4tatld-ctx16384",
   "model": "LFM2.5-2.6B",
   "notes": "MLX fp16 build; oMLX community benchmark trace; peak memory 2.8 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1244.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-2.6B-MLX-oQ4e-fp16 on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 1,244 108.6 2.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zj4tatld",
   "source_url": "https://omlx.ai/benchmarks/performance/zj4tatld",
   "tg_tokens": null,
   "tps": 108.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-zj4tatld-ctx65536",
   "model": "LFM2.5-2.6B",
   "notes": "MLX fp16 build; oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 889.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-2.6B-MLX-oQ4e-fp16 on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 889.9 68.3 3.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zj4tatld",
   "source_url": "https://omlx.ai/benchmarks/performance/zj4tatld",
   "tg_tokens": null,
   "tps": 68.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-zikcp3z2-ctx1024",
   "model": "LFM2.5-2.6B",
   "notes": "OptiQ build; oMLX community benchmark trace; peak memory 2.6 GB at this context; duplicate traces of the same session: 465nho5a, 4ith1t23, 96b98hr6; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 672.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-2.6B-OptiQ-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 672.8 48.0 2.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zikcp3z2",
   "source_url": "https://omlx.ai/benchmarks/performance/zikcp3z2",
   "tg_tokens": null,
   "tps": 48.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-zikcp3z2-ctx4096",
   "model": "LFM2.5-2.6B",
   "notes": "OptiQ build; oMLX community benchmark trace; peak memory 2.9 GB at this context; duplicate traces of the same session: 465nho5a, 4ith1t23, 96b98hr6; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 672.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-2.6B-OptiQ-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 672.8 46.6 2.9 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zikcp3z2",
   "source_url": "https://omlx.ai/benchmarks/performance/zikcp3z2",
   "tg_tokens": null,
   "tps": 46.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-zikcp3z2-ctx8192",
   "model": "LFM2.5-2.6B",
   "notes": "OptiQ build; oMLX community benchmark trace; peak memory 3.3 GB at this context; duplicate traces of the same session: 465nho5a, 4ith1t23, 96b98hr6; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 658.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-2.6B-OptiQ-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 8k 658.3 44.9 3.3 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zikcp3z2",
   "source_url": "https://omlx.ai/benchmarks/performance/zikcp3z2",
   "tg_tokens": null,
   "tps": 44.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-zikcp3z2-ctx16384",
   "model": "LFM2.5-2.6B",
   "notes": "OptiQ build; oMLX community benchmark trace; peak memory 4.0 GB at this context; duplicate traces of the same session: 465nho5a, 4ith1t23, 96b98hr6; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 626.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-2.6B-OptiQ-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 16k 626.2 41.1 4.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zikcp3z2",
   "source_url": "https://omlx.ai/benchmarks/performance/zikcp3z2",
   "tg_tokens": null,
   "tps": 41.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ck6u1ub1-ctx1024",
   "model": "Qwen3-30B-A3B-Instruct-2507",
   "notes": "oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: vjhubwsk; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1364.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-30B-A3B-Instruct-2507-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 1,364 114.6 16.6 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ck6u1ub1",
   "source_url": "https://omlx.ai/benchmarks/performance/ck6u1ub1",
   "tg_tokens": null,
   "tps": 114.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ck6u1ub1-ctx4096",
   "model": "Qwen3-30B-A3B-Instruct-2507",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: vjhubwsk; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3199.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-30B-A3B-Instruct-2507-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 3,199 100.3 17.0 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ck6u1ub1",
   "source_url": "https://omlx.ai/benchmarks/performance/ck6u1ub1",
   "tg_tokens": null,
   "tps": 100.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-43grx4e0-ctx1024",
   "model": "Qwen3-4B-Instruct-2507",
   "notes": "DWQ 2510 build; oMLX community benchmark trace; peak memory 2.9 GB at this context; duplicate traces of the same session: d3juu28r; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4948.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-4B-Instruct-2507-4bit-DWQ-2510 on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 4,948 141.5 2.9 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/43grx4e0",
   "source_url": "https://omlx.ai/benchmarks/performance/43grx4e0",
   "tg_tokens": null,
   "tps": 141.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-43grx4e0-ctx4096",
   "model": "Qwen3-4B-Instruct-2507",
   "notes": "DWQ 2510 build; oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: d3juu28r; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4811.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-4B-Instruct-2507-4bit-DWQ-2510 on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 4,811 123.8 3.4 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/43grx4e0",
   "source_url": "https://omlx.ai/benchmarks/performance/43grx4e0",
   "tg_tokens": null,
   "tps": 123.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ed3qtbw9-ctx1024",
   "model": "Qwen3-VL-4B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 2.9 GB at this context; duplicate traces of the same session: 3z1h8yva; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4739.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-VL-4B-Instruct-MLX-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 4,739 141.5 2.9 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ed3qtbw9",
   "source_url": "https://omlx.ai/benchmarks/performance/ed3qtbw9",
   "tg_tokens": null,
   "tps": 141.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ed3qtbw9-ctx4096",
   "model": "Qwen3-VL-4B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: 3z1h8yva; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4218.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-VL-4B-Instruct-MLX-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 4,218 118.2 3.4 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ed3qtbw9",
   "source_url": "https://omlx.ai/benchmarks/performance/ed3qtbw9",
   "tg_tokens": null,
   "tps": 118.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-ccl86th0-ctx1024",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 396.2,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-oQ5-fp16-mtp on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 5bit; 1k 396.2 39.7 24.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ccl86th0",
   "source_url": "https://omlx.ai/benchmarks/performance/ccl86th0",
   "tg_tokens": null,
   "tps": 39.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-ccl86th0-ctx4096",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 500.9,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-oQ5-fp16-mtp on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 5bit; 4k 500.9 36.9 25.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ccl86th0",
   "source_url": "https://omlx.ai/benchmarks/performance/ccl86th0",
   "tg_tokens": null,
   "tps": 36.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-5vevyhzl-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 16.0 GB; TTFT 853 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1200.0,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.6-35B-A3B-oQ3 on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 3bit; 1k tokens 1,200 PP tok/s 58.1 TG tok/s 853 TTFT (ms) 16.0 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5vevyhzl",
   "source_url": "https://omlx.ai/benchmarks/performance/5vevyhzl",
   "tg_tokens": null,
   "tps": 58.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-0k5sfq7x-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: xc4w5es7; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 959.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 959.6 99.7 37.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0k5sfq7x",
   "source_url": "https://omlx.ai/benchmarks/performance/0k5sfq7x",
   "tg_tokens": null,
   "tps": 99.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-0k5sfq7x-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: xc4w5es7; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 900.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 900.5 69.6 39.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0k5sfq7x",
   "source_url": "https://omlx.ai/benchmarks/performance/0k5sfq7x",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-24mh6dfh-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: n4n4hww9; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 891.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 891.0 53.9 37.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/24mh6dfh",
   "source_url": "https://omlx.ai/benchmarks/performance/24mh6dfh",
   "tg_tokens": null,
   "tps": 53.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-24mh6dfh-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: n4n4hww9; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 898.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 898.3 52.2 39.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/24mh6dfh",
   "source_url": "https://omlx.ai/benchmarks/performance/24mh6dfh",
   "tg_tokens": null,
   "tps": 52.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-hp3wwj4a-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: amtdxwor, avcnvsbx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2690.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 2,690 185.0 37.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hp3wwj4a",
   "source_url": "https://omlx.ai/benchmarks/performance/hp3wwj4a",
   "tg_tokens": null,
   "tps": 185.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-hp3wwj4a-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.1 GB at this context; duplicate traces of the same session: amtdxwor, avcnvsbx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3320.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 3,320 162.4 39.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hp3wwj4a",
   "source_url": "https://omlx.ai/benchmarks/performance/hp3wwj4a",
   "tg_tokens": null,
   "tps": 162.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-hp3wwj4a-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: amtdxwor, avcnvsbx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3157.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 3,157 174.1 39.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hp3wwj4a",
   "source_url": "https://omlx.ai/benchmarks/performance/hp3wwj4a",
   "tg_tokens": null,
   "tps": 174.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-8g9k7s4m-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: 6vgfpblk, z5mlgveh; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2627.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 2,627 183.0 37.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8g9k7s4m",
   "source_url": "https://omlx.ai/benchmarks/performance/8g9k7s4m",
   "tg_tokens": null,
   "tps": 183.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-8g9k7s4m-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.1 GB at this context; duplicate traces of the same session: 6vgfpblk, z5mlgveh; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3319.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 3,319 171.1 39.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8g9k7s4m",
   "source_url": "https://omlx.ai/benchmarks/performance/8g9k7s4m",
   "tg_tokens": null,
   "tps": 171.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-8g9k7s4m-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: 6vgfpblk, z5mlgveh; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3138.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 3,138 143.8 39.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8g9k7s4m",
   "source_url": "https://omlx.ai/benchmarks/performance/8g9k7s4m",
   "tg_tokens": null,
   "tps": 143.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-u9glf2hv-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: ddg3wstj; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 729.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 729.3 33.4 16.9 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u9glf2hv",
   "source_url": "https://omlx.ai/benchmarks/performance/u9glf2hv",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-u9glf2hv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: ddg3wstj; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 777.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 777.4 36.9 18.1 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u9glf2hv",
   "source_url": "https://omlx.ai/benchmarks/performance/u9glf2hv",
   "tg_tokens": null,
   "tps": 36.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-tekfrx4d-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 33.8 GB; TTFT 4296 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 238.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k tokens 238.4 PP tok/s 14.7 TG tok/s 4296 TTFT (ms) 33.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tekfrx4d",
   "source_url": "https://omlx.ai/benchmarks/performance/tekfrx4d",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-ykcaojh2-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.7 GB at this context; duplicate traces of the same session: doea8fwl; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 170.7,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 170.7 23.2 30.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ykcaojh2",
   "source_url": "https://omlx.ai/benchmarks/performance/ykcaojh2",
   "tg_tokens": null,
   "tps": 23.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-ykcaojh2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.3 GB at this context; duplicate traces of the same session: doea8fwl; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 198.7,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 198.7 23.4 34.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ykcaojh2",
   "source_url": "https://omlx.ai/benchmarks/performance/ykcaojh2",
   "tg_tokens": null,
   "tps": 23.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-zmtwcso2-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: kmh6cb82; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 167.2,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 167.2 20.1 19.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zmtwcso2",
   "source_url": "https://omlx.ai/benchmarks/performance/zmtwcso2",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-zmtwcso2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: kmh6cb82; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 167.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 167.0 23.5 23.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zmtwcso2",
   "source_url": "https://omlx.ai/benchmarks/performance/zmtwcso2",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-l9jrja52-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: z3nj8wnm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 825.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 825.5 18.6 28.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l9jrja52",
   "source_url": "https://omlx.ai/benchmarks/performance/l9jrja52",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-l9jrja52-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: z3nj8wnm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 812.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 812.8 18.4 29.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l9jrja52",
   "source_url": "https://omlx.ai/benchmarks/performance/l9jrja52",
   "tg_tokens": null,
   "tps": 18.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-0dotswjx-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 105.2 GB at this context; duplicate traces of the same session: eam3tvjp; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 641.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 1k 641.8 31.8 105.2 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0dotswjx",
   "source_url": "https://omlx.ai/benchmarks/performance/0dotswjx",
   "tg_tokens": null,
   "tps": 31.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-0dotswjx-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 105.6 GB at this context; duplicate traces of the same session: eam3tvjp; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 795.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 4k 795.1 27.1 105.6 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0dotswjx",
   "source_url": "https://omlx.ai/benchmarks/performance/0dotswjx",
   "tg_tokens": null,
   "tps": 27.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-1iuvhrca-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 105.6 GB; TTFT 5303 ms; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 772.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 4k tokens 772.5 PP tok/s 31.0 TG tok/s 5303 TTFT (ms) 105.6 Peak mem",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1iuvhrca",
   "source_url": "https://omlx.ai/benchmarks/performance/1iuvhrca",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-e8j1i0m2-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP320",
   "notes": "MTP Vision build; oMLX community benchmark trace, single run at 1k context; peak memory 36.6 GB; TTFT 4619 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 221.7,
   "provenance": "community",
   "quant": "oQ3e-DWQ",
   "quote": "Qwen3.8-Flash-Next-REAP320-oQ3e-DWQ-MTP-Vision-MLX on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 3bit; 1k tokens 221.7 PP tok/s 33.9 TG tok/s 4619 TTFT (ms) 36.6 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e8j1i0m2",
   "source_url": "https://omlx.ai/benchmarks/performance/e8j1i0m2",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jrtr8e40-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace, single run at 4k context; peak memory 86.8 GB; TTFT 3476 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1178.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k tokens 1,178 PP tok/s 41.0 TG tok/s 3476 TTFT (ms) 86.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jrtr8e40",
   "source_url": "https://omlx.ai/benchmarks/performance/jrtr8e40",
   "tg_tokens": null,
   "tps": 41.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-uska368y-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 58i4f7p7; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 335.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 335.0 47.8 70.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uska368y",
   "source_url": "https://omlx.ai/benchmarks/performance/uska368y",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-uska368y-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: 58i4f7p7; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 245.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 245.7 33.9 71.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uska368y",
   "source_url": "https://omlx.ai/benchmarks/performance/uska368y",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0isrh44f-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.2 GB at this context; duplicate traces of the same session: dihpnc3n, p026xuza; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1108.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 1,108 82.8 183.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0isrh44f",
   "source_url": "https://omlx.ai/benchmarks/performance/0isrh44f",
   "tg_tokens": null,
   "tps": 82.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0isrh44f-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.7 GB at this context; duplicate traces of the same session: dihpnc3n, p026xuza; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1372.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 1,372 81.4 183.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0isrh44f",
   "source_url": "https://omlx.ai/benchmarks/performance/0isrh44f",
   "tg_tokens": null,
   "tps": 81.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-0isrh44f-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 184.1 GB at this context; duplicate traces of the same session: dihpnc3n, p026xuza; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1364.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 1,364 68.2 184.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0isrh44f",
   "source_url": "https://omlx.ai/benchmarks/performance/0isrh44f",
   "tg_tokens": null,
   "tps": 68.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4najjr7r-ctx1024",
   "model": "Signal-3.8-27B-Terse-Coder",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 858.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Signal-3.8-27B-Terse-Coder-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 858.7 50.8 17.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4najjr7r",
   "source_url": "https://omlx.ai/benchmarks/performance/4najjr7r",
   "tg_tokens": null,
   "tps": 50.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4najjr7r-ctx4096",
   "model": "Signal-3.8-27B-Terse-Coder",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1256.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Signal-3.8-27B-Terse-Coder-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,256 55.5 19.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4najjr7r",
   "source_url": "https://omlx.ai/benchmarks/performance/4najjr7r",
   "tg_tokens": null,
   "tps": 55.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4najjr7r-ctx8192",
   "model": "Signal-3.8-27B-Terse-Coder",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1347.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Signal-3.8-27B-Terse-Coder-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,347 81.8 19.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4najjr7r",
   "source_url": "https://omlx.ai/benchmarks/performance/4najjr7r",
   "tg_tokens": null,
   "tps": 81.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4najjr7r-ctx16384",
   "model": "Signal-3.8-27B-Terse-Coder",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1288.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Signal-3.8-27B-Terse-Coder-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,288 79.0 19.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4najjr7r",
   "source_url": "https://omlx.ai/benchmarks/performance/4najjr7r",
   "tg_tokens": null,
   "tps": 79.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-578o4agx-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "bf16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: 6e9gqynr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 213.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-bf16-mtp-MLX on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 6bit; 1k 213.1 37.8 26.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/578o4agx",
   "source_url": "https://omlx.ai/benchmarks/performance/578o4agx",
   "tg_tokens": null,
   "tps": 37.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-578o4agx-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "bf16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: 6e9gqynr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-bf16-mtp-MLX on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 6bit; 4k 224.0 29.4 27.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/578o4agx",
   "source_url": "https://omlx.ai/benchmarks/performance/578o4agx",
   "tg_tokens": null,
   "tps": 29.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-j6x8dyxx-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "bf16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: 2v7y7vgk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 212.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-bf16-mtp-MLX on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 6bit; 1k 212.7 36.8 26.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j6x8dyxx",
   "source_url": "https://omlx.ai/benchmarks/performance/j6x8dyxx",
   "tg_tokens": null,
   "tps": 36.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-j6x8dyxx-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "bf16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: 2v7y7vgk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 221.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-bf16-mtp-MLX on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 6bit; 4k 221.9 31.1 27.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j6x8dyxx",
   "source_url": "https://omlx.ai/benchmarks/performance/j6x8dyxx",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-0xiq611d-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "bf16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: 1spcw88f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 215.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-bf16-mtp-MLX on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 6bit; 1k 215.4 32.2 26.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0xiq611d",
   "source_url": "https://omlx.ai/benchmarks/performance/0xiq611d",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-0xiq611d-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ6e",
   "notes": "bf16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: 1spcw88f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 227.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ6e-bf16-mtp-MLX on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 6bit; 4k 227.6 26.3 27.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0xiq611d",
   "source_url": "https://omlx.ai/benchmarks/performance/0xiq611d",
   "tg_tokens": null,
   "tps": 26.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-1bu2oary-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace, single run at 4k context; peak memory 37.5 GB; TTFT 32294 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 126.8,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k tokens 126.8 PP tok/s 17.8 TG tok/s 32294 TTFT (ms) 37.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1bu2oary",
   "source_url": "https://omlx.ai/benchmarks/performance/1bu2oary",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-3o8m4c5s-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace, single run at 4k context; peak memory 31.8 GB; TTFT 32619 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.6,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k tokens 125.6 PP tok/s 6.8 TG tok/s 32619 TTFT (ms) 31.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3o8m4c5s",
   "source_url": "https://omlx.ai/benchmarks/performance/3o8m4c5s",
   "tg_tokens": null,
   "tps": 6.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-tf10ou2l-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: x52i6h25; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 221.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 221.5 38.3 19.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tf10ou2l",
   "source_url": "https://omlx.ai/benchmarks/performance/tf10ou2l",
   "tg_tokens": null,
   "tps": 38.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-tf10ou2l-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "TensorFold build; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: x52i6h25; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 231.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TensorFold--Qwen3.8-27B-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 231.7 39.9 20.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tf10ou2l",
   "source_url": "https://omlx.ai/benchmarks/performance/tf10ou2l",
   "tg_tokens": null,
   "tps": 39.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-g6syxbzt-ctx1024",
   "model": "clef-flash",
   "notes": "generic trace name; oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: jm6iurg8; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2146.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "clef-flash-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 2,146 81.0 5.9 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g6syxbzt",
   "source_url": "https://omlx.ai/benchmarks/performance/g6syxbzt",
   "tg_tokens": null,
   "tps": 81.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-g6syxbzt-ctx4096",
   "model": "clef-flash",
   "notes": "generic trace name; oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: jm6iurg8; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2474.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "clef-flash-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 2,474 80.4 6.3 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g6syxbzt",
   "source_url": "https://omlx.ai/benchmarks/performance/g6syxbzt",
   "tg_tokens": null,
   "tps": 80.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-u5lgijwu-ctx1024",
   "model": "clef-flash",
   "notes": "generic trace name; oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: kyl2w6j4; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2586.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "clef-flash-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 2,586 81.5 5.9 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u5lgijwu",
   "source_url": "https://omlx.ai/benchmarks/performance/u5lgijwu",
   "tg_tokens": null,
   "tps": 81.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-u5lgijwu-ctx4096",
   "model": "clef-flash",
   "notes": "generic trace name; oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: kyl2w6j4; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2837.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "clef-flash-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 2,837 79.6 6.3 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u5lgijwu",
   "source_url": "https://omlx.ai/benchmarks/performance/u5lgijwu",
   "tg_tokens": null,
   "tps": 79.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-95huj06k-ctx16384",
   "model": "Ornith-1.5-35B-A3B-REAP192",
   "notes": "mchen04 MLX build; oMLX community benchmark trace, single run at 16k context; peak memory 15.7 GB; TTFT 53589 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 305.7,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "mchen04--Ornith-1.5-35B-A3B-REAP192-mxfp4-MLX on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k tokens 305.7 PP tok/s 45.0 TG tok/s 53589 TTFT (ms) 15.7 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/95huj06k",
   "source_url": "https://omlx.ai/benchmarks/performance/95huj06k",
   "tg_tokens": null,
   "tps": 45.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-p8x59u01-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 205.5 GB at this context; duplicate traces of the same session: 2b9hvy3h, iv79ca2e, urt9imw4; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 636.1,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 unknown; 1k 636.1 40.6 205.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p8x59u01",
   "source_url": "https://omlx.ai/benchmarks/performance/p8x59u01",
   "tg_tokens": null,
   "tps": 40.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-p8x59u01-ctx65536",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 206.1 GB at this context; duplicate traces of the same session: 2b9hvy3h, iv79ca2e, urt9imw4; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 960.6,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 unknown; 64k 960.6 45.2 206.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p8x59u01",
   "source_url": "https://omlx.ai/benchmarks/performance/p8x59u01",
   "tg_tokens": null,
   "tps": 45.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-p8x59u01-ctx131072",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 206.2 GB at this context; duplicate traces of the same session: 2b9hvy3h, iv79ca2e, urt9imw4; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 734.5,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 unknown; 128k 734.5 42.5 206.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p8x59u01",
   "source_url": "https://omlx.ai/benchmarks/performance/p8x59u01",
   "tg_tokens": null,
   "tps": 42.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-p8x59u01-ctx199680",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 206.2 GB at this context; duplicate traces of the same session: 2b9hvy3h, iv79ca2e, urt9imw4; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 584.9,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 unknown; 195k 584.9 51.7 206.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p8x59u01",
   "source_url": "https://omlx.ai/benchmarks/performance/p8x59u01",
   "tg_tokens": null,
   "tps": 51.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-gx2m22mp-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 205.5 GB at this context; duplicate traces of the same session: 4qiykb1g, zpw2pnfc; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 383.4,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 unknown; 1k 383.4 33.3 205.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gx2m22mp",
   "source_url": "https://omlx.ai/benchmarks/performance/gx2m22mp",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-gx2m22mp-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 205.9 GB at this context; duplicate traces of the same session: 4qiykb1g, zpw2pnfc; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 887.3,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 unknown; 4k 887.3 49.9 205.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gx2m22mp",
   "source_url": "https://omlx.ai/benchmarks/performance/gx2m22mp",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-gx2m22mp-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 206.0 GB at this context; duplicate traces of the same session: 4qiykb1g, zpw2pnfc; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1095.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 unknown; 8k 1,095 47.7 206.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gx2m22mp",
   "source_url": "https://omlx.ai/benchmarks/performance/gx2m22mp",
   "tg_tokens": null,
   "tps": 47.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-bmfd534e-ctx4096",
   "model": "LFM2.5-2.6B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 2.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1325.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-2.6B-MLX-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 1,325 128.8 2.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bmfd534e",
   "source_url": "https://omlx.ai/benchmarks/performance/bmfd534e",
   "tg_tokens": null,
   "tps": 128.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-bmfd534e-ctx16384",
   "model": "LFM2.5-2.6B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 2.8 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1249.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-2.6B-MLX-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 1,249 100.8 2.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bmfd534e",
   "source_url": "https://omlx.ai/benchmarks/performance/bmfd534e",
   "tg_tokens": null,
   "tps": 100.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-bmfd534e-ctx65536",
   "model": "LFM2.5-2.6B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 922.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-2.6B-MLX-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 922.1 67.8 3.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bmfd534e",
   "source_url": "https://omlx.ai/benchmarks/performance/bmfd534e",
   "tg_tokens": null,
   "tps": 67.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-y1px80hp-ctx4096",
   "model": "LFM2.5-2.6B",
   "notes": "MLX fp16 build; oMLX community benchmark trace; peak memory 2.6 GB at this context; duplicate traces of the same session: 3mh14xox; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1309.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-2.6B-MLX-oQ4e-fp16 on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 1,309 125.0 2.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y1px80hp",
   "source_url": "https://omlx.ai/benchmarks/performance/y1px80hp",
   "tg_tokens": null,
   "tps": 125.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-y1px80hp-ctx16384",
   "model": "LFM2.5-2.6B",
   "notes": "MLX fp16 build; oMLX community benchmark trace; peak memory 2.8 GB at this context; duplicate traces of the same session: 3mh14xox; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1244.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-2.6B-MLX-oQ4e-fp16 on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 1,244 108.6 2.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y1px80hp",
   "source_url": "https://omlx.ai/benchmarks/performance/y1px80hp",
   "tg_tokens": null,
   "tps": 108.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-y1px80hp-ctx65536",
   "model": "LFM2.5-2.6B",
   "notes": "MLX fp16 build; oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: 3mh14xox; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 889.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-2.6B-MLX-oQ4e-fp16 on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 889.9 68.3 3.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y1px80hp",
   "source_url": "https://omlx.ai/benchmarks/performance/y1px80hp",
   "tg_tokens": null,
   "tps": 68.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-kvpu7y8k-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: 8is6bmkp, s8qrbcuw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1586.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 1,586 119.6 13.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kvpu7y8k",
   "source_url": "https://omlx.ai/benchmarks/performance/kvpu7y8k",
   "tg_tokens": null,
   "tps": 119.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-kvpu7y8k-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.8 GB at this context; duplicate traces of the same session: 8is6bmkp, s8qrbcuw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1985.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 8k 1,985 114.1 13.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kvpu7y8k",
   "source_url": "https://omlx.ai/benchmarks/performance/kvpu7y8k",
   "tg_tokens": null,
   "tps": 114.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-kvpu7y8k-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 14.2 GB at this context; duplicate traces of the same session: 8is6bmkp, s8qrbcuw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1740.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 32k 1,740 94.7 14.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kvpu7y8k",
   "source_url": "https://omlx.ai/benchmarks/performance/kvpu7y8k",
   "tg_tokens": null,
   "tps": 94.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-g8ezgjxc-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: frj5nut0, jeun3f89; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 173.4,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 173.4 10.3 7.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g8ezgjxc",
   "source_url": "https://omlx.ai/benchmarks/performance/g8ezgjxc",
   "tg_tokens": null,
   "tps": 10.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-g8ezgjxc-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.8 GB at this context; duplicate traces of the same session: frj5nut0, jeun3f89; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 289.4,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 289.4 10.9 7.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g8ezgjxc",
   "source_url": "https://omlx.ai/benchmarks/performance/g8ezgjxc",
   "tg_tokens": null,
   "tps": 10.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-g8ezgjxc-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 10.6 GB at this context; duplicate traces of the same session: frj5nut0, jeun3f89; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 189.6,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 64k 189.6 10.2 10.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g8ezgjxc",
   "source_url": "https://omlx.ai/benchmarks/performance/g8ezgjxc",
   "tg_tokens": null,
   "tps": 10.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-u6tntvi2-ctx1024",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: 0azu84zp; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 100.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-6bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 1k 100.4 11.1 25.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u6tntvi2",
   "source_url": "https://omlx.ai/benchmarks/performance/u6tntvi2",
   "tg_tokens": null,
   "tps": 11.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-u6tntvi2-ctx4096",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 32.8 GB at this context; duplicate traces of the same session: 0azu84zp; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 109.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-6bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 4k 109.8 11.0 32.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u6tntvi2",
   "source_url": "https://omlx.ai/benchmarks/performance/u6tntvi2",
   "tg_tokens": null,
   "tps": 11.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-swqa3l84-ctx1024",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: jk148k1f; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 98.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-6bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 1k 98.5 10.7 25.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/swqa3l84",
   "source_url": "https://omlx.ai/benchmarks/performance/swqa3l84",
   "tg_tokens": null,
   "tps": 10.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-swqa3l84-ctx4096",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 32.8 GB at this context; duplicate traces of the same session: jk148k1f; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-6bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 4k 104.7 10.9 32.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/swqa3l84",
   "source_url": "https://omlx.ai/benchmarks/performance/swqa3l84",
   "tg_tokens": null,
   "tps": 10.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-2rk44p7n-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace, single run at 4k context; peak memory 22.2 GB; TTFT 19675 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 208.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k tokens 208.2 PP tok/s 17.6 TG tok/s 19675 TTFT (ms) 22.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2rk44p7n",
   "source_url": "https://omlx.ai/benchmarks/performance/2rk44p7n",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-r2yoi8up-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace, single run at 4k context; peak memory 23.1 GB; TTFT 20219 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 202.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k tokens 202.6 PP tok/s 34.2 TG tok/s 20219 TTFT (ms) 23.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r2yoi8up",
   "source_url": "https://omlx.ai/benchmarks/performance/r2yoi8up",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-3la92nvi-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace, single run at 4k context; peak memory 23.1 GB; TTFT 20767 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 197.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k tokens 197.2 PP tok/s 25.2 TG tok/s 20767 TTFT (ms) 23.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3la92nvi",
   "source_url": "https://omlx.ai/benchmarks/performance/3la92nvi",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-15901xq5-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: 3yvwwy38, bhz9q473; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 947.6,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-27B-nvfp4 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 947.6 30.8 17.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/15901xq5",
   "source_url": "https://omlx.ai/benchmarks/performance/15901xq5",
   "tg_tokens": null,
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-15901xq5-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: 3yvwwy38, bhz9q473; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1018.0,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-27B-nvfp4 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,018 27.0 18.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/15901xq5",
   "source_url": "https://omlx.ai/benchmarks/performance/15901xq5",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-15901xq5-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 3yvwwy38, bhz9q473; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 788.5,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-27B-nvfp4 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 788.5 29.0 19.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/15901xq5",
   "source_url": "https://omlx.ai/benchmarks/performance/15901xq5",
   "tg_tokens": null,
   "tps": 29.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ls1omwa8-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: ii3cpncp, m7qytdua; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1092.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,092 108.3 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ls1omwa8",
   "source_url": "https://omlx.ai/benchmarks/performance/ls1omwa8",
   "tg_tokens": null,
   "tps": 108.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ls1omwa8-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: ii3cpncp, m7qytdua; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1263.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,263 71.4 19.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ls1omwa8",
   "source_url": "https://omlx.ai/benchmarks/performance/ls1omwa8",
   "tg_tokens": null,
   "tps": 71.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ls1omwa8-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: ii3cpncp, m7qytdua; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 980.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 980.7 50.5 20.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ls1omwa8",
   "source_url": "https://omlx.ai/benchmarks/performance/ls1omwa8",
   "tg_tokens": null,
   "tps": 50.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-j16xatzy-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: urmcmjcd; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 574.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 574.8 39.6 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j16xatzy",
   "source_url": "https://omlx.ai/benchmarks/performance/j16xatzy",
   "tg_tokens": null,
   "tps": 39.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-j16xatzy-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: urmcmjcd; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 627.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 627.5 27.0 22.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j16xatzy",
   "source_url": "https://omlx.ai/benchmarks/performance/j16xatzy",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-suxwcsne-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: k2aqeggp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 568.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 568.4 32.8 21.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/suxwcsne",
   "source_url": "https://omlx.ai/benchmarks/performance/suxwcsne",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-suxwcsne-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: k2aqeggp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 642.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 642.7 31.2 24.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/suxwcsne",
   "source_url": "https://omlx.ai/benchmarks/performance/suxwcsne",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-m47nrwmm-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: vrlju3bc; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 552.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 552.8 37.3 21.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m47nrwmm",
   "source_url": "https://omlx.ai/benchmarks/performance/m47nrwmm",
   "tg_tokens": null,
   "tps": 37.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-m47nrwmm-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: vrlju3bc; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 614.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 614.1 30.9 24.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m47nrwmm",
   "source_url": "https://omlx.ai/benchmarks/performance/m47nrwmm",
   "tg_tokens": null,
   "tps": 30.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-10v9uasm-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: rxmpgcay; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 574.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 574.6 36.0 21.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/10v9uasm",
   "source_url": "https://omlx.ai/benchmarks/performance/10v9uasm",
   "tg_tokens": null,
   "tps": 36.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-10v9uasm-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: rxmpgcay; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 617.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 617.4 26.8 24.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/10v9uasm",
   "source_url": "https://omlx.ai/benchmarks/performance/10v9uasm",
   "tg_tokens": null,
   "tps": 26.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-zcoskrq5-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: wvm8lext; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 576.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 576.1 38.2 17.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zcoskrq5",
   "source_url": "https://omlx.ai/benchmarks/performance/zcoskrq5",
   "tg_tokens": null,
   "tps": 38.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-zcoskrq5-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: wvm8lext; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 598.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 598.6 41.3 20.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zcoskrq5",
   "source_url": "https://omlx.ai/benchmarks/performance/zcoskrq5",
   "tg_tokens": null,
   "tps": 41.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-vlb154um-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.2 GB at this context; duplicate traces of the same session: 02j3yppy, 5gdhuzp1, cb7urh4s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 248.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 248.0 40.6 33.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vlb154um",
   "source_url": "https://omlx.ai/benchmarks/performance/vlb154um",
   "tg_tokens": null,
   "tps": 40.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-vlb154um-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.5 GB at this context; duplicate traces of the same session: 02j3yppy, 5gdhuzp1, cb7urh4s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 332.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 332.1 36.5 36.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vlb154um",
   "source_url": "https://omlx.ai/benchmarks/performance/vlb154um",
   "tg_tokens": null,
   "tps": 36.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-vlb154um-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.0 GB at this context; duplicate traces of the same session: 02j3yppy, 5gdhuzp1, cb7urh4s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 335.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 335.0 45.4 37.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vlb154um",
   "source_url": "https://omlx.ai/benchmarks/performance/vlb154um",
   "tg_tokens": null,
   "tps": 45.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-vlb154um-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.0 GB at this context; duplicate traces of the same session: 02j3yppy, 5gdhuzp1, cb7urh4s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 320.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 320.5 49.5 38.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vlb154um",
   "source_url": "https://omlx.ai/benchmarks/performance/vlb154um",
   "tg_tokens": null,
   "tps": 49.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-vlb154um-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.7 GB at this context; duplicate traces of the same session: 02j3yppy, 5gdhuzp1, cb7urh4s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 290.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 290.9 55.1 39.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vlb154um",
   "source_url": "https://omlx.ai/benchmarks/performance/vlb154um",
   "tg_tokens": null,
   "tps": 55.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-vlb154um-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 42.4 GB at this context; duplicate traces of the same session: 02j3yppy, 5gdhuzp1, cb7urh4s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 246.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k 246.3 31.2 42.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vlb154um",
   "source_url": "https://omlx.ai/benchmarks/performance/vlb154um",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-vlb154um-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 46.9 GB at this context; duplicate traces of the same session: 02j3yppy, 5gdhuzp1, cb7urh4s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 189.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 128k 189.1 29.6 46.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vlb154um",
   "source_url": "https://omlx.ai/benchmarks/performance/vlb154um",
   "tg_tokens": null,
   "tps": 29.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-o0ate1m3-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; build 20261011-c7e8cd38; oMLX community benchmark trace; peak memory 69.3 GB at this context; duplicate traces of the same session: 4418jxwp, 5a83mmfy, 5ckqs4rw, dwv7ym3o, kqd7z90g, wq9qf0lz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 654.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-20261011-c7e8cd38 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 654.0 66.4 69.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o0ate1m3",
   "source_url": "https://omlx.ai/benchmarks/performance/o0ate1m3",
   "tg_tokens": null,
   "tps": 66.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-o0ate1m3-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; build 20261011-c7e8cd38; oMLX community benchmark trace; peak memory 70.1 GB at this context; duplicate traces of the same session: 4418jxwp, 5a83mmfy, 5ckqs4rw, dwv7ym3o, kqd7z90g, wq9qf0lz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1102.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-20261011-c7e8cd38 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 1,102 68.7 70.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o0ate1m3",
   "source_url": "https://omlx.ai/benchmarks/performance/o0ate1m3",
   "tg_tokens": null,
   "tps": 68.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-o0ate1m3-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; build 20261011-c7e8cd38; oMLX community benchmark trace; peak memory 70.3 GB at this context; duplicate traces of the same session: 4418jxwp, 5a83mmfy, 5ckqs4rw, dwv7ym3o, kqd7z90g, wq9qf0lz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1097.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-20261011-c7e8cd38 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 16k 1,097 69.4 70.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o0ate1m3",
   "source_url": "https://omlx.ai/benchmarks/performance/o0ate1m3",
   "tg_tokens": null,
   "tps": 69.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-o0ate1m3-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; build 20261011-c7e8cd38; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 4418jxwp, 5a83mmfy, 5ckqs4rw, dwv7ym3o, kqd7z90g, wq9qf0lz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1091.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-20261011-c7e8cd38 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 1,091 68.3 70.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o0ate1m3",
   "source_url": "https://omlx.ai/benchmarks/performance/o0ate1m3",
   "tg_tokens": null,
   "tps": 68.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-o0ate1m3-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; build 20261011-c7e8cd38; oMLX community benchmark trace; peak memory 71.9 GB at this context; duplicate traces of the same session: 4418jxwp, 5a83mmfy, 5ckqs4rw, dwv7ym3o, kqd7z90g, wq9qf0lz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1077.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-20261011-c7e8cd38 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k 1,077 63.6 71.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o0ate1m3",
   "source_url": "https://omlx.ai/benchmarks/performance/o0ate1m3",
   "tg_tokens": null,
   "tps": 63.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-o0ate1m3-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; build 20261011-c7e8cd38; oMLX community benchmark trace; peak memory 74.5 GB at this context; duplicate traces of the same session: 4418jxwp, 5a83mmfy, 5ckqs4rw, dwv7ym3o, kqd7z90g, wq9qf0lz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1061.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-20261011-c7e8cd38 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 128k 1,061 58.0 74.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o0ate1m3",
   "source_url": "https://omlx.ai/benchmarks/performance/o0ate1m3",
   "tg_tokens": null,
   "tps": 58.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-o0ate1m3-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; build 20261011-c7e8cd38; oMLX community benchmark trace; peak memory 77.8 GB at this context; duplicate traces of the same session: 4418jxwp, 5a83mmfy, 5ckqs4rw, dwv7ym3o, kqd7z90g, wq9qf0lz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1044.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-20261011-c7e8cd38 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 195k 1,044 53.3 77.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o0ate1m3",
   "source_url": "https://omlx.ai/benchmarks/performance/o0ate1m3",
   "tg_tokens": null,
   "tps": 53.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-sw61bsg1-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 133.6 GB; TTFT 15761 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4158.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 64k tokens 4,158 PP tok/s 71.8 TG tok/s 15761 TTFT (ms) 133.6 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sw61bsg1",
   "source_url": "https://omlx.ai/benchmarks/performance/sw61bsg1",
   "tg_tokens": null,
   "tps": 71.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-mafl17x7-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 195k context; peak memory 138.1 GB; TTFT 49792 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4017.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 195k tokens 4,017 PP tok/s 86.1 TG tok/s 49792 TTFT (ms) 138.1 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mafl17x7",
   "source_url": "https://omlx.ai/benchmarks/performance/mafl17x7",
   "tg_tokens": null,
   "tps": 86.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-24a79zrv-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace, single run at 32k context; peak memory 24.6 GB; TTFT 41044 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 798.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k tokens 798.4 PP tok/s 22.4 TG tok/s 41044 TTFT (ms) 24.6 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/24a79zrv",
   "source_url": "https://omlx.ai/benchmarks/performance/24a79zrv",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0eselaho-ctx1024",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 9.1 GB at this context; duplicate traces of the same session: ethy4af4, riy0f17l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1534.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "gemma-4-12B-it-oQ5e on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 5bit; 1k 1,534 51.8 9.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0eselaho",
   "source_url": "https://omlx.ai/benchmarks/performance/0eselaho",
   "tg_tokens": null,
   "tps": 51.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0eselaho-ctx4096",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 9.4 GB at this context; duplicate traces of the same session: ethy4af4, riy0f17l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1823.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "gemma-4-12B-it-oQ5e on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 5bit; 4k 1,823 51.1 9.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0eselaho",
   "source_url": "https://omlx.ai/benchmarks/performance/0eselaho",
   "tg_tokens": null,
   "tps": 51.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0eselaho-ctx65536",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 12.6 GB at this context; duplicate traces of the same session: ethy4af4, riy0f17l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1234.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "gemma-4-12B-it-oQ5e on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 5bit; 64k 1,234 46.2 12.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0eselaho",
   "source_url": "https://omlx.ai/benchmarks/performance/0eselaho",
   "tg_tokens": null,
   "tps": 46.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-ndxk3uld-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: iics2soq; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 664.2,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 4k 664.2 41.2 11.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ndxk3uld",
   "source_url": "https://omlx.ai/benchmarks/performance/ndxk3uld",
   "tg_tokens": null,
   "tps": 41.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-ndxk3uld-ctx16384",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 12.1 GB at this context; duplicate traces of the same session: iics2soq; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 584.5,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 16k 584.5 36.5 12.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ndxk3uld",
   "source_url": "https://omlx.ai/benchmarks/performance/ndxk3uld",
   "tg_tokens": null,
   "tps": 36.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-ndxk3uld-ctx32768",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 12.4 GB at this context; duplicate traces of the same session: iics2soq; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 502.3,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 32k 502.3 30.8 12.4 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ndxk3uld",
   "source_url": "https://omlx.ai/benchmarks/performance/ndxk3uld",
   "tg_tokens": null,
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-ndxk3uld-ctx65536",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 13.2 GB at this context; duplicate traces of the same session: iics2soq; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 390.2,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 64k 390.2 23.1 13.2 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ndxk3uld",
   "source_url": "https://omlx.ai/benchmarks/performance/ndxk3uld",
   "tg_tokens": null,
   "tps": 23.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-g8bhl6y5-ctx16384",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "mlx-community build; oMLX community benchmark trace, single run at 16k context; peak memory 23.3 GB; TTFT 52193 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 313.9,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "mlx-community--MiniCPM-V-4.7-35B-A3B-mxfp4 on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k tokens 313.9 PP tok/s 28.7 TG tok/s 52193 TTFT (ms) 23.3 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g8bhl6y5",
   "source_url": "https://omlx.ai/benchmarks/performance/g8bhl6y5",
   "tg_tokens": null,
   "tps": 28.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-w1y0famo-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "mlx-community; oMLX community benchmark trace, single run at 16k context; peak memory 15.0 GB; TTFT 51981 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 315.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--gemma-4-26b-a4b-it-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k tokens 315.2 PP tok/s 32.9 TG tok/s 51981 TTFT (ms) 15.0 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w1y0famo",
   "source_url": "https://omlx.ai/benchmarks/performance/w1y0famo",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-6lh6kg11-ctx32768",
   "model": "DeepSeek-R1-Distill-Qwen-7B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 6.3 GB; TTFT 1104506 ms; duplicate traces of the same session: none; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 29.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-R1-Distill-Qwen-7B-4bit on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 4bit; 32k tokens 29.7 PP tok/s 2.7 TG tok/s 1104506 TTFT (ms) 6.3 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6lh6kg11",
   "source_url": "https://omlx.ai/benchmarks/performance/6lh6kg11",
   "tg_tokens": null,
   "tps": 2.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-17ow1mfk-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.6 GB at this context; duplicate traces of the same session: qzc571vt, upwtpk3o; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 460.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 460.4 50.7 177.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/17ow1mfk",
   "source_url": "https://omlx.ai/benchmarks/performance/17ow1mfk",
   "tg_tokens": null,
   "tps": 50.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-17ow1mfk-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.8 GB at this context; duplicate traces of the same session: qzc571vt, upwtpk3o; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 568.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 568.0 44.6 178.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/17ow1mfk",
   "source_url": "https://omlx.ai/benchmarks/performance/17ow1mfk",
   "tg_tokens": null,
   "tps": 44.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-17ow1mfk-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.1 GB at this context; duplicate traces of the same session: qzc571vt, upwtpk3o; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 550.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 550.7 47.0 179.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/17ow1mfk",
   "source_url": "https://omlx.ai/benchmarks/performance/17ow1mfk",
   "tg_tokens": null,
   "tps": 47.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-75ptjd57-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 4k context; peak memory 19.2 GB; TTFT 9670 ms; duplicate traces of the same session: none; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 423.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "LFM2.5-24B-A2B-MLX-6bit on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 6bit; 4k tokens 423.6 PP tok/s 40.8 TG tok/s 9670 TTFT (ms) 19.2 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/75ptjd57",
   "source_url": "https://omlx.ai/benchmarks/performance/75ptjd57",
   "tg_tokens": null,
   "tps": 40.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-f1sttb4b-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: bislha7r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2367.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "LFM2.5-24B-A2B-MLX-6bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 2,367 131.0 19.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f1sttb4b",
   "source_url": "https://omlx.ai/benchmarks/performance/f1sttb4b",
   "tg_tokens": null,
   "tps": 131.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-f1sttb4b-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: bislha7r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2220.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "LFM2.5-24B-A2B-MLX-6bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 2,220 118.7 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f1sttb4b",
   "source_url": "https://omlx.ai/benchmarks/performance/f1sttb4b",
   "tg_tokens": null,
   "tps": 118.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-vszsiggo-ctx1024",
   "model": "Muse Glimmer",
   "notes": "oMLX community benchmark trace; peak memory 15.2 GB at this context; duplicate traces of the same session: rmzxlb68; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 767.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-30B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 767.7 25.9 15.2 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vszsiggo",
   "source_url": "https://omlx.ai/benchmarks/performance/vszsiggo",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-vszsiggo-ctx4096",
   "model": "Muse Glimmer",
   "notes": "oMLX community benchmark trace; peak memory 15.5 GB at this context; duplicate traces of the same session: rmzxlb68; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 730.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-30B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 730.1 25.6 15.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vszsiggo",
   "source_url": "https://omlx.ai/benchmarks/performance/vszsiggo",
   "tg_tokens": null,
   "tps": 25.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-mk2bypo6-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 17.3 GB; TTFT 38810 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 4k tokens 105.5 PP tok/s 18.2 TG tok/s 38810 TTFT (ms) 17.3 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mk2bypo6",
   "source_url": "https://omlx.ai/benchmarks/performance/mk2bypo6",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-a57usejd-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: lfk5qh0y; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 136.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 136.0 31.5 23.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a57usejd",
   "source_url": "https://omlx.ai/benchmarks/performance/a57usejd",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-a57usejd-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 27.3 GB at this context; duplicate traces of the same session: lfk5qh0y; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 134.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 134.9 23.6 27.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a57usejd",
   "source_url": "https://omlx.ai/benchmarks/performance/a57usejd",
   "tg_tokens": null,
   "tps": 23.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-t94ozc62-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 3puyszyp, ykmcjset; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 783.1,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-27B-nvfp4 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 783.1 32.9 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t94ozc62",
   "source_url": "https://omlx.ai/benchmarks/performance/t94ozc62",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-t94ozc62-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 3puyszyp, ykmcjset; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 789.4,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-27B-nvfp4 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 789.4 32.9 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t94ozc62",
   "source_url": "https://omlx.ai/benchmarks/performance/t94ozc62",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-t94ozc62-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 3puyszyp, ykmcjset; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 745.7,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-27B-nvfp4 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 745.7 28.1 18.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t94ozc62",
   "source_url": "https://omlx.ai/benchmarks/performance/t94ozc62",
   "tg_tokens": null,
   "tps": 28.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mbslimg2-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: f1mnndpa, wnrues3l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 947.5,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-27B-nvfp4 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 947.5 31.3 17.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mbslimg2",
   "source_url": "https://omlx.ai/benchmarks/performance/mbslimg2",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mbslimg2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: f1mnndpa, wnrues3l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 965.2,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-27B-nvfp4 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 965.2 26.3 18.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mbslimg2",
   "source_url": "https://omlx.ai/benchmarks/performance/mbslimg2",
   "tg_tokens": null,
   "tps": 26.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mbslimg2-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: f1mnndpa, wnrues3l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 748.4,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-27B-nvfp4 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 748.4 27.8 19.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mbslimg2",
   "source_url": "https://omlx.ai/benchmarks/performance/mbslimg2",
   "tg_tokens": null,
   "tps": 27.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-84cnv23f-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: 636qxh1f, koqwmh61, tta4urnn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 428.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 428.6 72.7 21.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/84cnv23f",
   "source_url": "https://omlx.ai/benchmarks/performance/84cnv23f",
   "tg_tokens": null,
   "tps": 72.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-84cnv23f-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: 636qxh1f, koqwmh61, tta4urnn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 571.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 571.3 85.5 25.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/84cnv23f",
   "source_url": "https://omlx.ai/benchmarks/performance/84cnv23f",
   "tg_tokens": null,
   "tps": 85.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-84cnv23f-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: 636qxh1f, koqwmh61, tta4urnn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 569.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 569.6 59.4 25.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/84cnv23f",
   "source_url": "https://omlx.ai/benchmarks/performance/84cnv23f",
   "tg_tokens": null,
   "tps": 59.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-84cnv23f-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.2 GB at this context; duplicate traces of the same session: 636qxh1f, koqwmh61, tta4urnn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 540.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 540.5 102.1 26.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/84cnv23f",
   "source_url": "https://omlx.ai/benchmarks/performance/84cnv23f",
   "tg_tokens": null,
   "tps": 102.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-84cnv23f-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: 636qxh1f, koqwmh61, tta4urnn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 496.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 496.6 164.6 27.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/84cnv23f",
   "source_url": "https://omlx.ai/benchmarks/performance/84cnv23f",
   "tg_tokens": null,
   "tps": 164.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-84cnv23f-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.0 GB at this context; duplicate traces of the same session: 636qxh1f, koqwmh61, tta4urnn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 417.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 417.9 71.3 31.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/84cnv23f",
   "source_url": "https://omlx.ai/benchmarks/performance/84cnv23f",
   "tg_tokens": null,
   "tps": 71.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-58xti0d9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 03z6y125; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 60.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 60.2 14.9 18.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/58xti0d9",
   "source_url": "https://omlx.ai/benchmarks/performance/58xti0d9",
   "tg_tokens": null,
   "tps": 14.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-58xti0d9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 03z6y125; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 97.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 97.8 3.4 22.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/58xti0d9",
   "source_url": "https://omlx.ai/benchmarks/performance/58xti0d9",
   "tg_tokens": null,
   "tps": 3.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-2nj8l00b-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 6g69o65u; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 126.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 126.4 20.4 18.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2nj8l00b",
   "source_url": "https://omlx.ai/benchmarks/performance/2nj8l00b",
   "tg_tokens": null,
   "tps": 20.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-2nj8l00b-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: 6g69o65u; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 131.8 26.6 19.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2nj8l00b",
   "source_url": "https://omlx.ai/benchmarks/performance/2nj8l00b",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-muqog36a-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 2yichs7b; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 580.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 580.3 41.9 17.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/muqog36a",
   "source_url": "https://omlx.ai/benchmarks/performance/muqog36a",
   "tg_tokens": null,
   "tps": 41.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-muqog36a-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 2yichs7b; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 591.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 591.4 46.8 20.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/muqog36a",
   "source_url": "https://omlx.ai/benchmarks/performance/muqog36a",
   "tg_tokens": null,
   "tps": 46.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-wvjjs21w-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: 8a6i6keo; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 580.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 580.0 31.6 18.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wvjjs21w",
   "source_url": "https://omlx.ai/benchmarks/performance/wvjjs21w",
   "tg_tokens": null,
   "tps": 31.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-wvjjs21w-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 8a6i6keo; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 614.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 614.9 23.4 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wvjjs21w",
   "source_url": "https://omlx.ai/benchmarks/performance/wvjjs21w",
   "tg_tokens": null,
   "tps": 23.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-cy8xb5d9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.2 GB at this context; duplicate traces of the same session: rnkhga13; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 571.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 571.2 31.7 18.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cy8xb5d9",
   "source_url": "https://omlx.ai/benchmarks/performance/cy8xb5d9",
   "tg_tokens": null,
   "tps": 31.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-cy8xb5d9-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: rnkhga13; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 630.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 630.7 25.4 21.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cy8xb5d9",
   "source_url": "https://omlx.ai/benchmarks/performance/cy8xb5d9",
   "tg_tokens": null,
   "tps": 25.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-u8h0m1fp-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: jx7n52tt; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 473.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 473.9 36.1 18.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u8h0m1fp",
   "source_url": "https://omlx.ai/benchmarks/performance/u8h0m1fp",
   "tg_tokens": null,
   "tps": 36.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-u8h0m1fp-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: jx7n52tt; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 630.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 630.8 29.1 21.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u8h0m1fp",
   "source_url": "https://omlx.ai/benchmarks/performance/u8h0m1fp",
   "tg_tokens": null,
   "tps": 29.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-a6nr366z-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: sy50sg02; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 563.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 563.4 36.8 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a6nr366z",
   "source_url": "https://omlx.ai/benchmarks/performance/a6nr366z",
   "tg_tokens": null,
   "tps": 36.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-a6nr366z-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: sy50sg02; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 603.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 603.7 31.3 22.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a6nr366z",
   "source_url": "https://omlx.ai/benchmarks/performance/a6nr366z",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-tdq69xab-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.2 GB at this context; duplicate traces of the same session: 6qx6cc9z, ufn4ohal; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 248.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 248.0 40.6 33.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tdq69xab",
   "source_url": "https://omlx.ai/benchmarks/performance/tdq69xab",
   "tg_tokens": null,
   "tps": 40.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-tdq69xab-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.5 GB at this context; duplicate traces of the same session: 6qx6cc9z, ufn4ohal; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 332.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 332.1 36.5 36.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tdq69xab",
   "source_url": "https://omlx.ai/benchmarks/performance/tdq69xab",
   "tg_tokens": null,
   "tps": 36.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-tdq69xab-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.0 GB at this context; duplicate traces of the same session: 6qx6cc9z, ufn4ohal; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 335.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 335.0 45.4 37.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tdq69xab",
   "source_url": "https://omlx.ai/benchmarks/performance/tdq69xab",
   "tg_tokens": null,
   "tps": 45.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-tdq69xab-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.0 GB at this context; duplicate traces of the same session: 6qx6cc9z, ufn4ohal; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 320.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 320.5 49.5 38.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tdq69xab",
   "source_url": "https://omlx.ai/benchmarks/performance/tdq69xab",
   "tg_tokens": null,
   "tps": 49.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-tdq69xab-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.7 GB at this context; duplicate traces of the same session: 6qx6cc9z, ufn4ohal; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 290.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 290.9 55.1 39.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tdq69xab",
   "source_url": "https://omlx.ai/benchmarks/performance/tdq69xab",
   "tg_tokens": null,
   "tps": 55.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-tdq69xab-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 42.4 GB at this context; duplicate traces of the same session: 6qx6cc9z, ufn4ohal; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 246.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k 246.3 31.2 42.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tdq69xab",
   "source_url": "https://omlx.ai/benchmarks/performance/tdq69xab",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-tdq69xab-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 46.9 GB at this context; duplicate traces of the same session: 6qx6cc9z, ufn4ohal; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 189.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 128k 189.1 29.6 46.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tdq69xab",
   "source_url": "https://omlx.ai/benchmarks/performance/tdq69xab",
   "tg_tokens": null,
   "tps": 29.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-l4yz6z1s-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.2 GB at this context; duplicate traces of the same session: 5zr2zb7a, ayg9ius9, d4epk13l, mszvcvm1, zx49bcdx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 248.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 248.5 48.7 33.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l4yz6z1s",
   "source_url": "https://omlx.ai/benchmarks/performance/l4yz6z1s",
   "tg_tokens": null,
   "tps": 48.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-l4yz6z1s-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.4 GB at this context; duplicate traces of the same session: 5zr2zb7a, ayg9ius9, d4epk13l, mszvcvm1, zx49bcdx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 334.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 334.2 40.4 36.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l4yz6z1s",
   "source_url": "https://omlx.ai/benchmarks/performance/l4yz6z1s",
   "tg_tokens": null,
   "tps": 40.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-l4yz6z1s-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.0 GB at this context; duplicate traces of the same session: 5zr2zb7a, ayg9ius9, d4epk13l, mszvcvm1, zx49bcdx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 337.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 337.0 41.5 37.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l4yz6z1s",
   "source_url": "https://omlx.ai/benchmarks/performance/l4yz6z1s",
   "tg_tokens": null,
   "tps": 41.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-l4yz6z1s-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.0 GB at this context; duplicate traces of the same session: 5zr2zb7a, ayg9ius9, d4epk13l, mszvcvm1, zx49bcdx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 321.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 321.1 49.0 38.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l4yz6z1s",
   "source_url": "https://omlx.ai/benchmarks/performance/l4yz6z1s",
   "tg_tokens": null,
   "tps": 49.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-l4yz6z1s-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.7 GB at this context; duplicate traces of the same session: 5zr2zb7a, ayg9ius9, d4epk13l, mszvcvm1, zx49bcdx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 291.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 291.9 60.0 39.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l4yz6z1s",
   "source_url": "https://omlx.ai/benchmarks/performance/l4yz6z1s",
   "tg_tokens": null,
   "tps": 60.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-l4yz6z1s-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 42.4 GB at this context; duplicate traces of the same session: 5zr2zb7a, ayg9ius9, d4epk13l, mszvcvm1, zx49bcdx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 246.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k 246.8 35.1 42.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l4yz6z1s",
   "source_url": "https://omlx.ai/benchmarks/performance/l4yz6z1s",
   "tg_tokens": null,
   "tps": 35.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-f29103f5-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.8 GB at this context; duplicate traces of the same session: 5ce17h1b, ee2jd17h, n39mfmjo, tv6quk20; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1160.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,160 99.5 100.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f29103f5",
   "source_url": "https://omlx.ai/benchmarks/performance/f29103f5",
   "tg_tokens": null,
   "tps": 99.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-f29103f5-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: 5ce17h1b, ee2jd17h, n39mfmjo, tv6quk20; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1402.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 1,402 92.1 101.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f29103f5",
   "source_url": "https://omlx.ai/benchmarks/performance/f29103f5",
   "tg_tokens": null,
   "tps": 92.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-f29103f5-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.8 GB at this context; duplicate traces of the same session: 5ce17h1b, ee2jd17h, n39mfmjo, tv6quk20; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1391.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 1,391 90.9 101.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f29103f5",
   "source_url": "https://omlx.ai/benchmarks/performance/f29103f5",
   "tg_tokens": null,
   "tps": 90.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-f29103f5-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.8 GB at this context; duplicate traces of the same session: 5ce17h1b, ee2jd17h, n39mfmjo, tv6quk20; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1379.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 1,379 85.7 102.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f29103f5",
   "source_url": "https://omlx.ai/benchmarks/performance/f29103f5",
   "tg_tokens": null,
   "tps": 85.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-f29103f5-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 105.5 GB at this context; duplicate traces of the same session: 5ce17h1b, ee2jd17h, n39mfmjo, tv6quk20; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1358.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 1,358 77.4 105.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f29103f5",
   "source_url": "https://omlx.ai/benchmarks/performance/f29103f5",
   "tg_tokens": null,
   "tps": 77.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-a5xufbp4-ctx1024",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: f3blksds, req2trw5; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 263.9,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 2bit; 1k 263.9 31.4 11.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a5xufbp4",
   "source_url": "https://omlx.ai/benchmarks/performance/a5xufbp4",
   "tg_tokens": null,
   "tps": 31.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-a5xufbp4-ctx8192",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: f3blksds, req2trw5; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 262.1,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 2bit; 8k 262.1 28.9 21.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a5xufbp4",
   "source_url": "https://omlx.ai/benchmarks/performance/a5xufbp4",
   "tg_tokens": null,
   "tps": 28.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-a5xufbp4-ctx32768",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: f3blksds, req2trw5; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 227.1,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 2bit; 32k 227.1 21.4 21.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a5xufbp4",
   "source_url": "https://omlx.ai/benchmarks/performance/a5xufbp4",
   "tg_tokens": null,
   "tps": 21.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-zlsx45il-ctx1024",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 3xaj086n, xv3z5045; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 669.1,
   "provenance": "community",
   "quant": "unknown",
   "quote": "gemma-4-12B-it on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 unknown; 1k 669.1 21.4 23.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zlsx45il",
   "source_url": "https://omlx.ai/benchmarks/performance/zlsx45il",
   "tg_tokens": null,
   "tps": 21.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-zlsx45il-ctx4096",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: 3xaj086n, xv3z5045; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1610.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "gemma-4-12B-it on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 unknown; 4k 1,610 21.2 23.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zlsx45il",
   "source_url": "https://omlx.ai/benchmarks/performance/zlsx45il",
   "tg_tokens": null,
   "tps": 21.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-zlsx45il-ctx65536",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: 3xaj086n, xv3z5045; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1354.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "gemma-4-12B-it on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 unknown; 64k 1,354 20.2 26.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zlsx45il",
   "source_url": "https://omlx.ai/benchmarks/performance/zlsx45il",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-1znaydjh-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: aax93cfq; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1482.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 1,482 51.5 14.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1znaydjh",
   "source_url": "https://omlx.ai/benchmarks/performance/1znaydjh",
   "tg_tokens": null,
   "tps": 51.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-1znaydjh-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.5 GB at this context; duplicate traces of the same session: aax93cfq; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1505.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,505 49.5 14.5 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1znaydjh",
   "source_url": "https://omlx.ai/benchmarks/performance/1znaydjh",
   "tg_tokens": null,
   "tps": 49.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-dzn2fjd8-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: 6pa2r7j4; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 664.2,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 4k 664.2 41.2 11.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dzn2fjd8",
   "source_url": "https://omlx.ai/benchmarks/performance/dzn2fjd8",
   "tg_tokens": null,
   "tps": 41.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-dzn2fjd8-ctx16384",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 12.1 GB at this context; duplicate traces of the same session: 6pa2r7j4; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 584.5,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 16k 584.5 36.5 12.1 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dzn2fjd8",
   "source_url": "https://omlx.ai/benchmarks/performance/dzn2fjd8",
   "tg_tokens": null,
   "tps": 36.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-dzn2fjd8-ctx32768",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 12.4 GB at this context; duplicate traces of the same session: 6pa2r7j4; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 502.3,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 32k 502.3 30.8 12.4 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dzn2fjd8",
   "source_url": "https://omlx.ai/benchmarks/performance/dzn2fjd8",
   "tg_tokens": null,
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-dzn2fjd8-ctx65536",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 13.2 GB at this context; duplicate traces of the same session: 6pa2r7j4; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 390.2,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 64k 390.2 23.1 13.2 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dzn2fjd8",
   "source_url": "https://omlx.ai/benchmarks/performance/dzn2fjd8",
   "tg_tokens": null,
   "tps": 23.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-djfzc2c0-ctx1024",
   "model": "CyberTiel-Coder-35B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: 6qeypg5s, dzf4vspd, o0mb3tn8; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1746.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "CyberTiel-Coder-35B-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 1k 1,746 114.2 30.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/djfzc2c0",
   "source_url": "https://omlx.ai/benchmarks/performance/djfzc2c0",
   "tg_tokens": null,
   "tps": 114.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-djfzc2c0-ctx4096",
   "model": "CyberTiel-Coder-35B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: 6qeypg5s, dzf4vspd, o0mb3tn8; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2471.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "CyberTiel-Coder-35B-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 4k 2,471 124.9 32.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/djfzc2c0",
   "source_url": "https://omlx.ai/benchmarks/performance/djfzc2c0",
   "tg_tokens": null,
   "tps": 124.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-djfzc2c0-ctx32768",
   "model": "CyberTiel-Coder-35B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 34.3 GB at this context; duplicate traces of the same session: 6qeypg5s, dzf4vspd, o0mb3tn8; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1667.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "CyberTiel-Coder-35B-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 32k 1,667 104.6 34.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/djfzc2c0",
   "source_url": "https://omlx.ai/benchmarks/performance/djfzc2c0",
   "tg_tokens": null,
   "tps": 104.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-djfzc2c0-ctx65536",
   "model": "CyberTiel-Coder-35B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 35.6 GB at this context; duplicate traces of the same session: 6qeypg5s, dzf4vspd, o0mb3tn8; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1298.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "CyberTiel-Coder-35B-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 64k 1,298 118.6 35.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/djfzc2c0",
   "source_url": "https://omlx.ai/benchmarks/performance/djfzc2c0",
   "tg_tokens": null,
   "tps": 118.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-gv4nfvki-ctx1024",
   "model": "DeepSeek-R1-Distill-Qwen-1.5B-chat",
   "notes": "mlx build; oMLX community benchmark trace, single run at 1k context; peak memory 1.7 GB; TTFT 318 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3222.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "DeepSeek-R1-Distill-Qwen-1.5B-chat-mlx-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 3,222 PP tok/s 336.5 TG tok/s 318 TTFT (ms) 1.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gv4nfvki",
   "source_url": "https://omlx.ai/benchmarks/performance/gv4nfvki",
   "tg_tokens": null,
   "tps": 336.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xk9fdu53-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 93.8 GB at this context; duplicate traces of the same session: htu22mcx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 383.1,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 383.1 13.0 93.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xk9fdu53",
   "source_url": "https://omlx.ai/benchmarks/performance/xk9fdu53",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xk9fdu53-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 93.8 GB at this context; duplicate traces of the same session: htu22mcx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 452.3,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 452.3 13.0 93.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xk9fdu53",
   "source_url": "https://omlx.ai/benchmarks/performance/xk9fdu53",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-nwupfyb3-ctx1024",
   "model": "LFM2-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: 1lvic3jh, pc6scd3i, wl0rhid4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2573.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2-24B-A2B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 2,573 58.6 24.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nwupfyb3",
   "source_url": "https://omlx.ai/benchmarks/performance/nwupfyb3",
   "tg_tokens": null,
   "tps": 58.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-nwupfyb3-ctx4096",
   "model": "LFM2-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: 1lvic3jh, pc6scd3i, wl0rhid4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3113.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2-24B-A2B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 3,113 111.3 25.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nwupfyb3",
   "source_url": "https://omlx.ai/benchmarks/performance/nwupfyb3",
   "tg_tokens": null,
   "tps": 111.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-nwupfyb3-ctx32768",
   "model": "LFM2-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: 1lvic3jh, pc6scd3i, wl0rhid4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2682.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2-24B-A2B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 32k 2,682 100.8 26.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nwupfyb3",
   "source_url": "https://omlx.ai/benchmarks/performance/nwupfyb3",
   "tg_tokens": null,
   "tps": 100.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-nwupfyb3-ctx65536",
   "model": "LFM2-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 26.9 GB at this context; duplicate traces of the same session: 1lvic3jh, pc6scd3i, wl0rhid4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2194.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2-24B-A2B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 64k 2,194 87.8 26.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nwupfyb3",
   "source_url": "https://omlx.ai/benchmarks/performance/nwupfyb3",
   "tg_tokens": null,
   "tps": 87.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-5yf3ujda-ctx1024",
   "model": "NuExtract3",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 5.4 GB at this context; duplicate traces of the same session: nqsmf7sj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1634.0,
   "provenance": "community",
   "quant": "8bits",
   "quote": "NuExtract3-mlx-8bits on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 8bit; 1k 1,634 90.4 5.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5yf3ujda",
   "source_url": "https://omlx.ai/benchmarks/performance/5yf3ujda",
   "tg_tokens": null,
   "tps": 90.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-5yf3ujda-ctx4096",
   "model": "NuExtract3",
   "notes": "mlx build; oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: nqsmf7sj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1643.0,
   "provenance": "community",
   "quant": "8bits",
   "quote": "NuExtract3-mlx-8bits on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 8bit; 4k 1,643 88.1 5.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5yf3ujda",
   "source_url": "https://omlx.ai/benchmarks/performance/5yf3ujda",
   "tg_tokens": null,
   "tps": 88.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-854hzvgz-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.0 GB at this context; duplicate traces of the same session: j8yxah3z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 903.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 8bit; 1k 903.6 52.4 10.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/854hzvgz",
   "source_url": "https://omlx.ai/benchmarks/performance/854hzvgz",
   "tg_tokens": null,
   "tps": 52.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-854hzvgz-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.5 GB at this context; duplicate traces of the same session: j8yxah3z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 936.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 8bit; 4k 936.1 51.4 10.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/854hzvgz",
   "source_url": "https://omlx.ai/benchmarks/performance/854hzvgz",
   "tg_tokens": null,
   "tps": 51.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ozxt186v-ctx4096",
   "model": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: wp21d2ya; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 964.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 964.5 95.4 21.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ozxt186v",
   "source_url": "https://omlx.ai/benchmarks/performance/ozxt186v",
   "tg_tokens": null,
   "tps": 95.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ozxt186v-ctx16384",
   "model": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed",
   "notes": "oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: wp21d2ya; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 767.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 767.6 65.5 22.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ozxt186v",
   "source_url": "https://omlx.ai/benchmarks/performance/ozxt186v",
   "tg_tokens": null,
   "tps": 65.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-2yahu0pi-ctx1024",
   "model": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 9sz0wqku; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 871.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 871.9 91.9 19.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2yahu0pi",
   "source_url": "https://omlx.ai/benchmarks/performance/2yahu0pi",
   "tg_tokens": null,
   "tps": 91.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-2yahu0pi-ctx4096",
   "model": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 9sz0wqku; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1052.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 1,052 88.0 21.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2yahu0pi",
   "source_url": "https://omlx.ai/benchmarks/performance/2yahu0pi",
   "tg_tokens": null,
   "tps": 88.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-ekfdgvoa-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: ln8htkgr; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 136.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 136.3 27.8 24.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ekfdgvoa",
   "source_url": "https://omlx.ai/benchmarks/performance/ekfdgvoa",
   "tg_tokens": null,
   "tps": 27.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ekfdgvoa-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: ln8htkgr; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 139.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 139.3 24.8 27.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ekfdgvoa",
   "source_url": "https://omlx.ai/benchmarks/performance/ekfdgvoa",
   "tg_tokens": null,
   "tps": 24.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-yd28rico-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: atgsnhv6; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 139.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 139.4 17.6 20.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yd28rico",
   "source_url": "https://omlx.ai/benchmarks/performance/yd28rico",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-yd28rico-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: atgsnhv6; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 141.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 141.8 17.4 23.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yd28rico",
   "source_url": "https://omlx.ai/benchmarks/performance/yd28rico",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-4dikpzly-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 188.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 188.0 20.5 30.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4dikpzly",
   "source_url": "https://omlx.ai/benchmarks/performance/4dikpzly",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-4dikpzly-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.7 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 276.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 276.0 20.5 33.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4dikpzly",
   "source_url": "https://omlx.ai/benchmarks/performance/4dikpzly",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-4dikpzly-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 268.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 268.0 22.1 34.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4dikpzly",
   "source_url": "https://omlx.ai/benchmarks/performance/4dikpzly",
   "tg_tokens": null,
   "tps": 22.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-4dikpzly-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 251.9,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 251.9 24.5 36.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4dikpzly",
   "source_url": "https://omlx.ai/benchmarks/performance/4dikpzly",
   "tg_tokens": null,
   "tps": 24.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-33g4okgq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: 1vmgsncg; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 187.2,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 187.2 9.8 32.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/33g4okgq",
   "source_url": "https://omlx.ai/benchmarks/performance/33g4okgq",
   "tg_tokens": null,
   "tps": 9.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-33g4okgq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.4 GB at this context; duplicate traces of the same session: 1vmgsncg; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 281.9,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 281.9 10.6 34.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/33g4okgq",
   "source_url": "https://omlx.ai/benchmarks/performance/33g4okgq",
   "tg_tokens": null,
   "tps": 10.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-33g4okgq-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.4 GB at this context; duplicate traces of the same session: 1vmgsncg; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 274.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 274.0 10.9 35.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/33g4okgq",
   "source_url": "https://omlx.ai/benchmarks/performance/33g4okgq",
   "tg_tokens": null,
   "tps": 10.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-33g4okgq-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.8 GB at this context; duplicate traces of the same session: 1vmgsncg; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 253.9,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 253.9 10.9 36.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/33g4okgq",
   "source_url": "https://omlx.ai/benchmarks/performance/33g4okgq",
   "tg_tokens": null,
   "tps": 10.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-d77sd0if-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 431.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 431.4 35.3 16.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d77sd0if",
   "source_url": "https://omlx.ai/benchmarks/performance/d77sd0if",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-d77sd0if-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 451.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 451.7 34.7 19.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d77sd0if",
   "source_url": "https://omlx.ai/benchmarks/performance/d77sd0if",
   "tg_tokens": null,
   "tps": 34.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-d77sd0if-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 442.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 442.6 31.0 20.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d77sd0if",
   "source_url": "https://omlx.ai/benchmarks/performance/d77sd0if",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-d77sd0if-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 362.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 362.2 23.0 25.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d77sd0if",
   "source_url": "https://omlx.ai/benchmarks/performance/d77sd0if",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ueevaq0d-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: eykrnyhj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 428.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 428.6 72.7 21.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ueevaq0d",
   "source_url": "https://omlx.ai/benchmarks/performance/ueevaq0d",
   "tg_tokens": null,
   "tps": 72.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ueevaq0d-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: eykrnyhj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 571.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 571.3 85.5 25.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ueevaq0d",
   "source_url": "https://omlx.ai/benchmarks/performance/ueevaq0d",
   "tg_tokens": null,
   "tps": 85.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ueevaq0d-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: eykrnyhj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 569.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 569.6 59.4 25.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ueevaq0d",
   "source_url": "https://omlx.ai/benchmarks/performance/ueevaq0d",
   "tg_tokens": null,
   "tps": 59.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ueevaq0d-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.2 GB at this context; duplicate traces of the same session: eykrnyhj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 540.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 540.5 102.1 26.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ueevaq0d",
   "source_url": "https://omlx.ai/benchmarks/performance/ueevaq0d",
   "tg_tokens": null,
   "tps": 102.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ueevaq0d-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: eykrnyhj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 496.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 496.6 164.6 27.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ueevaq0d",
   "source_url": "https://omlx.ai/benchmarks/performance/ueevaq0d",
   "tg_tokens": null,
   "tps": 164.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ueevaq0d-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.0 GB at this context; duplicate traces of the same session: eykrnyhj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 417.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 417.9 71.3 31.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ueevaq0d",
   "source_url": "https://omlx.ai/benchmarks/performance/ueevaq0d",
   "tg_tokens": null,
   "tps": 71.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-kzq47hav-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: 6idwin9c, hskpeggp, jfxi3wy1, kbtdtkw0, yo3rmx89; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 430.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 430.8 84.2 21.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kzq47hav",
   "source_url": "https://omlx.ai/benchmarks/performance/kzq47hav",
   "tg_tokens": null,
   "tps": 84.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-kzq47hav-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: 6idwin9c, hskpeggp, jfxi3wy1, kbtdtkw0, yo3rmx89; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 514.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 514.8 81.4 25.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kzq47hav",
   "source_url": "https://omlx.ai/benchmarks/performance/kzq47hav",
   "tg_tokens": null,
   "tps": 81.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-kzq47hav-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: 6idwin9c, hskpeggp, jfxi3wy1, kbtdtkw0, yo3rmx89; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 570.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 570.3 96.5 25.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kzq47hav",
   "source_url": "https://omlx.ai/benchmarks/performance/kzq47hav",
   "tg_tokens": null,
   "tps": 96.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-kzq47hav-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.2 GB at this context; duplicate traces of the same session: 6idwin9c, hskpeggp, jfxi3wy1, kbtdtkw0, yo3rmx89; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 541.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 541.5 94.0 26.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kzq47hav",
   "source_url": "https://omlx.ai/benchmarks/performance/kzq47hav",
   "tg_tokens": null,
   "tps": 94.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-kzq47hav-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: 6idwin9c, hskpeggp, jfxi3wy1, kbtdtkw0, yo3rmx89; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 491.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 491.4 164.8 27.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kzq47hav",
   "source_url": "https://omlx.ai/benchmarks/performance/kzq47hav",
   "tg_tokens": null,
   "tps": 164.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-kzq47hav-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.9 GB at this context; duplicate traces of the same session: 6idwin9c, hskpeggp, jfxi3wy1, kbtdtkw0, yo3rmx89; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 416.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 416.5 70.6 30.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kzq47hav",
   "source_url": "https://omlx.ai/benchmarks/performance/kzq47hav",
   "tg_tokens": null,
   "tps": 70.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-acl483jb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: r5j8mxym; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 568.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 568.7 41.2 18.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/acl483jb",
   "source_url": "https://omlx.ai/benchmarks/performance/acl483jb",
   "tg_tokens": null,
   "tps": 41.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-acl483jb-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: r5j8mxym; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 619.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 619.5 32.0 22.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/acl483jb",
   "source_url": "https://omlx.ai/benchmarks/performance/acl483jb",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-s1er9dx9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: 66kim0l8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 563.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 563.9 43.7 18.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s1er9dx9",
   "source_url": "https://omlx.ai/benchmarks/performance/s1er9dx9",
   "tg_tokens": null,
   "tps": 43.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-s1er9dx9-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 66kim0l8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 615.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 615.4 30.2 22.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s1er9dx9",
   "source_url": "https://omlx.ai/benchmarks/performance/s1er9dx9",
   "tg_tokens": null,
   "tps": 30.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-9cjpfz96-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: ibesdwvq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 578.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 578.4 43.0 18.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9cjpfz96",
   "source_url": "https://omlx.ai/benchmarks/performance/9cjpfz96",
   "tg_tokens": null,
   "tps": 43.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-9cjpfz96-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.4 GB at this context; duplicate traces of the same session: ibesdwvq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 374.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 128k 374.5 18.2 28.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9cjpfz96",
   "source_url": "https://omlx.ai/benchmarks/performance/9cjpfz96",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-a6zxhfqj-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: bnoimo41, gr2v6wkf; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1079.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,079 100.7 100.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a6zxhfqj",
   "source_url": "https://omlx.ai/benchmarks/performance/a6zxhfqj",
   "tg_tokens": null,
   "tps": 100.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-a6zxhfqj-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: bnoimo41, gr2v6wkf; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1302.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 1,302 83.7 101.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a6zxhfqj",
   "source_url": "https://omlx.ai/benchmarks/performance/a6zxhfqj",
   "tg_tokens": null,
   "tps": 83.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-a6zxhfqj-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.9 GB at this context; duplicate traces of the same session: bnoimo41, gr2v6wkf; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1329.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 1,329 71.5 103.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a6zxhfqj",
   "source_url": "https://omlx.ai/benchmarks/performance/a6zxhfqj",
   "tg_tokens": null,
   "tps": 71.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dmltgb0s-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: flwpo5ob, l2z7wc5y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1044.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,044 71.5 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dmltgb0s",
   "source_url": "https://omlx.ai/benchmarks/performance/dmltgb0s",
   "tg_tokens": null,
   "tps": 71.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dmltgb0s-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: flwpo5ob, l2z7wc5y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2034.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,034 66.1 70.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dmltgb0s",
   "source_url": "https://omlx.ai/benchmarks/performance/dmltgb0s",
   "tg_tokens": null,
   "tps": 66.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dmltgb0s-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 82.9 GB at this context; duplicate traces of the same session: flwpo5ob, l2z7wc5y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2375.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,375 61.8 82.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dmltgb0s",
   "source_url": "https://omlx.ai/benchmarks/performance/dmltgb0s",
   "tg_tokens": null,
   "tps": 61.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ch3jmj96-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.7 GB; TTFT 980 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1045.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,045 PP tok/s 62.2 TG tok/s 980 TTFT (ms) 70.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ch3jmj96",
   "source_url": "https://omlx.ai/benchmarks/performance/ch3jmj96",
   "tg_tokens": null,
   "tps": 62.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vqoopkrh-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 348.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 348.5 53.1 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vqoopkrh",
   "source_url": "https://omlx.ai/benchmarks/performance/vqoopkrh",
   "tg_tokens": null,
   "tps": 53.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vqoopkrh-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1092.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,092 39.4 71.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vqoopkrh",
   "source_url": "https://omlx.ai/benchmarks/performance/vqoopkrh",
   "tg_tokens": null,
   "tps": 39.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-6pw2mtq0-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 195k context; peak memory 191.8 GB; TTFT 48372 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4135.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 195k tokens 4,135 PP tok/s 78.0 TG tok/s 48372 TTFT (ms) 191.8 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6pw2mtq0",
   "source_url": "https://omlx.ai/benchmarks/performance/6pw2mtq0",
   "tg_tokens": null,
   "tps": 78.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-695upq8u-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 191.1 GB; TTFT 30327 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4322.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 128k tokens 4,322 PP tok/s 93.3 TG tok/s 30327 TTFT (ms) 191.1 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/695upq8u",
   "source_url": "https://omlx.ai/benchmarks/performance/695upq8u",
   "tg_tokens": null,
   "tps": 93.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-z8qwvsrz-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 186.9 GB; TTFT 14722 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4452.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 64k tokens 4,452 PP tok/s 94.2 TG tok/s 14722 TTFT (ms) 186.9 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z8qwvsrz",
   "source_url": "https://omlx.ai/benchmarks/performance/z8qwvsrz",
   "tg_tokens": null,
   "tps": 94.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-zudhdfna-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 191.1 GB; TTFT 30688 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4271.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 128k tokens 4,271 PP tok/s 85.0 TG tok/s 30688 TTFT (ms) 191.1 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zudhdfna",
   "source_url": "https://omlx.ai/benchmarks/performance/zudhdfna",
   "tg_tokens": null,
   "tps": 85.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-pzf6y1o2-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 185.9 GB; TTFT 29717 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4411.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 128k tokens 4,411 PP tok/s 65.8 TG tok/s 29717 TTFT (ms) 185.9 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pzf6y1o2",
   "source_url": "https://omlx.ai/benchmarks/performance/pzf6y1o2",
   "tg_tokens": null,
   "tps": 65.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jg4pj6bb-ctx1024",
   "model": "gemma-4-31B-it-uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.1 GB at this context; duplicate traces of the same session: 3j1vzpwo, d3xj953s, wz70c32b; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 250.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-31B-it-uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 250.4 41.1 34.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jg4pj6bb",
   "source_url": "https://omlx.ai/benchmarks/performance/jg4pj6bb",
   "tg_tokens": null,
   "tps": 41.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jg4pj6bb-ctx4096",
   "model": "gemma-4-31B-it-uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.0 GB at this context; duplicate traces of the same session: 3j1vzpwo, d3xj953s, wz70c32b; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 259.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-31B-it-uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 259.2 45.8 35.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jg4pj6bb",
   "source_url": "https://omlx.ai/benchmarks/performance/jg4pj6bb",
   "tg_tokens": null,
   "tps": 45.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jg4pj6bb-ctx8192",
   "model": "gemma-4-31B-it-uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: 3j1vzpwo, d3xj953s, wz70c32b; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 257.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-31B-it-uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 8k 257.2 36.7 35.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jg4pj6bb",
   "source_url": "https://omlx.ai/benchmarks/performance/jg4pj6bb",
   "tg_tokens": null,
   "tps": 36.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jg4pj6bb-ctx16384",
   "model": "gemma-4-31B-it-uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.7 GB at this context; duplicate traces of the same session: 3j1vzpwo, d3xj953s, wz70c32b; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 247.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-31B-it-uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 16k 247.8 36.0 36.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jg4pj6bb",
   "source_url": "https://omlx.ai/benchmarks/performance/jg4pj6bb",
   "tg_tokens": null,
   "tps": 36.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-38jmhn14-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: 8iv6tji9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2905.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.5-9B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 2,905 103.5 5.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/38jmhn14",
   "source_url": "https://omlx.ai/benchmarks/performance/38jmhn14",
   "tg_tokens": null,
   "tps": 103.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-38jmhn14-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "mlx-community build; oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: 8iv6tji9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2896.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.5-9B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 2,896 101.2 6.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/38jmhn14",
   "source_url": "https://omlx.ai/benchmarks/performance/38jmhn14",
   "tg_tokens": null,
   "tps": 101.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-4knbe9wg-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 179.6 GB; TTFT 17304 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1894.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k tokens 1,894 PP tok/s 74.8 TG tok/s 17304 TTFT (ms) 179.6 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4knbe9wg",
   "source_url": "https://omlx.ai/benchmarks/performance/4knbe9wg",
   "tg_tokens": null,
   "tps": 74.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-g95sl88i-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: izoceh7v; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2709.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "LFM2.5-24B-A2B-oQ6e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 2,709 95.2 18.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g95sl88i",
   "source_url": "https://omlx.ai/benchmarks/performance/g95sl88i",
   "tg_tokens": null,
   "tps": 95.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-g95sl88i-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: izoceh7v; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3850.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "LFM2.5-24B-A2B-oQ6e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 3,850 92.0 19.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g95sl88i",
   "source_url": "https://omlx.ai/benchmarks/performance/g95sl88i",
   "tg_tokens": null,
   "tps": 92.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-so16czio-ctx65536",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: ogg3a4t6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2914.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "LFM2.5-24B-A2B-oQ6e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 64k 2,914 60.3 20.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/so16czio",
   "source_url": "https://omlx.ai/benchmarks/performance/so16czio",
   "tg_tokens": null,
   "tps": 60.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-so16czio-ctx131072",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: ogg3a4t6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2091.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "LFM2.5-24B-A2B-oQ6e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 128k 2,091 45.2 21.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/so16czio",
   "source_url": "https://omlx.ai/benchmarks/performance/so16czio",
   "tg_tokens": null,
   "tps": 45.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-m88lyloe-ctx1024",
   "model": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 1i502hfk; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 944.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 944.7 108.3 20.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m88lyloe",
   "source_url": "https://omlx.ai/benchmarks/performance/m88lyloe",
   "tg_tokens": null,
   "tps": 108.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-m88lyloe-ctx4096",
   "model": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 1i502hfk; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 982.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 982.6 93.5 21.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m88lyloe",
   "source_url": "https://omlx.ai/benchmarks/performance/m88lyloe",
   "tg_tokens": null,
   "tps": 93.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Pro",
   "id": "omlx-x0xg2anh-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: hq3m780y, pte5eec1, wxa1c7xo; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 1k 103.9 13.5 19.2 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x0xg2anh",
   "source_url": "https://omlx.ai/benchmarks/performance/x0xg2anh",
   "tg_tokens": null,
   "tps": 13.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Pro",
   "id": "omlx-x0xg2anh-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: hq3m780y, pte5eec1, wxa1c7xo; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 4k 106.5 13.7 20.3 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x0xg2anh",
   "source_url": "https://omlx.ai/benchmarks/performance/x0xg2anh",
   "tg_tokens": null,
   "tps": 13.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-x0xg2anh-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: hq3m780y, pte5eec1, wxa1c7xo; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 40.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 8k 40.7 4.6 20.6 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x0xg2anh",
   "source_url": "https://omlx.ai/benchmarks/performance/x0xg2anh",
   "tg_tokens": null,
   "tps": 4.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-x0xg2anh-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: hq3m780y, pte5eec1, wxa1c7xo; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 39.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 16k 39.3 5.2 21.4 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x0xg2anh",
   "source_url": "https://omlx.ai/benchmarks/performance/x0xg2anh",
   "tg_tokens": null,
   "tps": 5.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-90t6k97l-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: rp83do6q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 973.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 973.0 31.8 17.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/90t6k97l",
   "source_url": "https://omlx.ai/benchmarks/performance/90t6k97l",
   "tg_tokens": null,
   "tps": 31.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-90t6k97l-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: rp83do6q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 972.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 972.0 29.8 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/90t6k97l",
   "source_url": "https://omlx.ai/benchmarks/performance/90t6k97l",
   "tg_tokens": null,
   "tps": 29.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-810dg9ed-ctx32768",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "MTPLX build; oMLX community benchmark trace, single run at 32k context; peak memory 20.6 GB; TTFT 45952 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 713.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Abliterated-MTPLX-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 32k tokens 713.1 PP tok/s 33.1 TG tok/s 45952 TTFT (ms) 20.6 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/810dg9ed",
   "source_url": "https://omlx.ai/benchmarks/performance/810dg9ed",
   "tg_tokens": null,
   "tps": 33.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-w9t3ibfe-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: 3daredln; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 4bit; 8k 205.2 31.1 20.4 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w9t3ibfe",
   "source_url": "https://omlx.ai/benchmarks/performance/w9t3ibfe",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-w9t3ibfe-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: 3daredln; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 148.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 4bit; 64k 148.8 22.8 24.7 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w9t3ibfe",
   "source_url": "https://omlx.ai/benchmarks/performance/w9t3ibfe",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0zhjmcvh-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: gebmvr5a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 941.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 941.3 18.3 29.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0zhjmcvh",
   "source_url": "https://omlx.ai/benchmarks/performance/0zhjmcvh",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0zhjmcvh-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: gebmvr5a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 924.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 16k 924.8 17.7 30.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0zhjmcvh",
   "source_url": "https://omlx.ai/benchmarks/performance/0zhjmcvh",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xk2xoz0z-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: kou375p3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1001.0,
   "provenance": "community",
   "quant": "OptiQ 4bit",
   "quote": "Qwen3.8-27B-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,001 26.9 21.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xk2xoz0z",
   "source_url": "https://omlx.ai/benchmarks/performance/xk2xoz0z",
   "tg_tokens": null,
   "tps": 26.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-xk2xoz0z-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: kou375p3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 968.5,
   "provenance": "community",
   "quant": "OptiQ 4bit",
   "quote": "Qwen3.8-27B-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 968.5 25.7 21.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xk2xoz0z",
   "source_url": "https://omlx.ai/benchmarks/performance/xk2xoz0z",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qcp22mdm-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 14.2 GB at this context; duplicate traces of the same session: g363f7oa, g5yu88il; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 813.3,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-27B-oQ3e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 3bit; 1k 813.3 65.2 14.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qcp22mdm",
   "source_url": "https://omlx.ai/benchmarks/performance/qcp22mdm",
   "tg_tokens": null,
   "tps": 65.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qcp22mdm-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 15.1 GB at this context; duplicate traces of the same session: g363f7oa, g5yu88il; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 878.4,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-27B-oQ3e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 3bit; 4k 878.4 60.5 15.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qcp22mdm",
   "source_url": "https://omlx.ai/benchmarks/performance/qcp22mdm",
   "tg_tokens": null,
   "tps": 60.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qcp22mdm-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 15.4 GB at this context; duplicate traces of the same session: g363f7oa, g5yu88il; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 848.8,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-27B-oQ3e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 3bit; 8k 848.8 62.3 15.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qcp22mdm",
   "source_url": "https://omlx.ai/benchmarks/performance/qcp22mdm",
   "tg_tokens": null,
   "tps": 62.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-2zdo2b99-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: koid4a0x; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 187.2,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 187.2 9.8 32.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2zdo2b99",
   "source_url": "https://omlx.ai/benchmarks/performance/2zdo2b99",
   "tg_tokens": null,
   "tps": 9.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-2zdo2b99-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.4 GB at this context; duplicate traces of the same session: koid4a0x; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 281.9,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 281.9 10.6 34.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2zdo2b99",
   "source_url": "https://omlx.ai/benchmarks/performance/2zdo2b99",
   "tg_tokens": null,
   "tps": 10.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-2zdo2b99-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.4 GB at this context; duplicate traces of the same session: koid4a0x; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 274.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 274.0 10.9 35.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2zdo2b99",
   "source_url": "https://omlx.ai/benchmarks/performance/2zdo2b99",
   "tg_tokens": null,
   "tps": 10.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-2zdo2b99-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.8 GB at this context; duplicate traces of the same session: koid4a0x; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 253.9,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 253.9 10.9 36.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2zdo2b99",
   "source_url": "https://omlx.ai/benchmarks/performance/2zdo2b99",
   "tg_tokens": null,
   "tps": 10.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-o5f5m1io-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: bl1dsrdw, lx2e7214; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 427.8,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 427.8 79.3 22.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o5f5m1io",
   "source_url": "https://omlx.ai/benchmarks/performance/o5f5m1io",
   "tg_tokens": null,
   "tps": 79.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-o5f5m1io-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: bl1dsrdw, lx2e7214; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 444.9,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 444.9 72.1 25.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o5f5m1io",
   "source_url": "https://omlx.ai/benchmarks/performance/o5f5m1io",
   "tg_tokens": null,
   "tps": 72.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-o5f5m1io-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: bl1dsrdw, lx2e7214; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 523.6,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 523.6 94.6 25.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o5f5m1io",
   "source_url": "https://omlx.ai/benchmarks/performance/o5f5m1io",
   "tg_tokens": null,
   "tps": 94.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-o5f5m1io-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: bl1dsrdw, lx2e7214; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 517.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 517.0 78.6 26.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o5f5m1io",
   "source_url": "https://omlx.ai/benchmarks/performance/o5f5m1io",
   "tg_tokens": null,
   "tps": 78.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-o5f5m1io-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.2 GB at this context; duplicate traces of the same session: bl1dsrdw, lx2e7214; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 477.5,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 477.5 155.7 28.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o5f5m1io",
   "source_url": "https://omlx.ai/benchmarks/performance/o5f5m1io",
   "tg_tokens": null,
   "tps": 155.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-o5f5m1io-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.4 GB at this context; duplicate traces of the same session: bl1dsrdw, lx2e7214; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 404.5,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 404.5 71.9 31.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o5f5m1io",
   "source_url": "https://omlx.ai/benchmarks/performance/o5f5m1io",
   "tg_tokens": null,
   "tps": 71.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ikre56nr-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 22274yhk, apjz6bgk; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 431.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 431.4 35.3 16.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ikre56nr",
   "source_url": "https://omlx.ai/benchmarks/performance/ikre56nr",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ikre56nr-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: 22274yhk, apjz6bgk; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 451.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 451.7 34.7 19.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ikre56nr",
   "source_url": "https://omlx.ai/benchmarks/performance/ikre56nr",
   "tg_tokens": null,
   "tps": 34.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ikre56nr-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 22274yhk, apjz6bgk; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 442.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 442.6 31.0 20.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ikre56nr",
   "source_url": "https://omlx.ai/benchmarks/performance/ikre56nr",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ikre56nr-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: 22274yhk, apjz6bgk; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 362.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 362.2 23.0 25.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ikre56nr",
   "source_url": "https://omlx.ai/benchmarks/performance/ikre56nr",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-3u8g9ryx-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: 2b2tt2yr, cq7syb5w, ece0zifs, hkfs64rd, wpfge01p; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 407.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 407.7 40.3 21.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3u8g9ryx",
   "source_url": "https://omlx.ai/benchmarks/performance/3u8g9ryx",
   "tg_tokens": null,
   "tps": 40.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-3u8g9ryx-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: 2b2tt2yr, cq7syb5w, ece0zifs, hkfs64rd, wpfge01p; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 516.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 516.0 64.3 25.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3u8g9ryx",
   "source_url": "https://omlx.ai/benchmarks/performance/3u8g9ryx",
   "tg_tokens": null,
   "tps": 64.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-3u8g9ryx-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: 2b2tt2yr, cq7syb5w, ece0zifs, hkfs64rd, wpfge01p; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 565.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 565.7 97.9 25.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3u8g9ryx",
   "source_url": "https://omlx.ai/benchmarks/performance/3u8g9ryx",
   "tg_tokens": null,
   "tps": 97.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-3u8g9ryx-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.2 GB at this context; duplicate traces of the same session: 2b2tt2yr, cq7syb5w, ece0zifs, hkfs64rd, wpfge01p; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 539.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 539.6 59.7 26.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3u8g9ryx",
   "source_url": "https://omlx.ai/benchmarks/performance/3u8g9ryx",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3u8g9ryx-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: 2b2tt2yr, cq7syb5w, ece0zifs, hkfs64rd, wpfge01p; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 491.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 491.7 163.1 27.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3u8g9ryx",
   "source_url": "https://omlx.ai/benchmarks/performance/3u8g9ryx",
   "tg_tokens": null,
   "tps": 163.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3u8g9ryx-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.9 GB at this context; duplicate traces of the same session: 2b2tt2yr, cq7syb5w, ece0zifs, hkfs64rd, wpfge01p; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 415.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 415.6 70.5 30.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3u8g9ryx",
   "source_url": "https://omlx.ai/benchmarks/performance/3u8g9ryx",
   "tg_tokens": null,
   "tps": 70.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-d3cp2g38-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 4hfmn322, b3w2jdb5, t8nfdwyl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 253.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 253.3 60.6 21.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d3cp2g38",
   "source_url": "https://omlx.ai/benchmarks/performance/d3cp2g38",
   "tg_tokens": null,
   "tps": 60.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-d3cp2g38-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: 4hfmn322, b3w2jdb5, t8nfdwyl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 337.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 337.7 84.6 25.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d3cp2g38",
   "source_url": "https://omlx.ai/benchmarks/performance/d3cp2g38",
   "tg_tokens": null,
   "tps": 84.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-d3cp2g38-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.6 GB at this context; duplicate traces of the same session: 4hfmn322, b3w2jdb5, t8nfdwyl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 337.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 337.8 32.7 25.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d3cp2g38",
   "source_url": "https://omlx.ai/benchmarks/performance/d3cp2g38",
   "tg_tokens": null,
   "tps": 32.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-d3cp2g38-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: 4hfmn322, b3w2jdb5, t8nfdwyl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 322.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 322.7 54.1 26.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d3cp2g38",
   "source_url": "https://omlx.ai/benchmarks/performance/d3cp2g38",
   "tg_tokens": null,
   "tps": 54.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-xrwacrbc-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: hhybc22r; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 580.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 580.1 37.2 18.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xrwacrbc",
   "source_url": "https://omlx.ai/benchmarks/performance/xrwacrbc",
   "tg_tokens": null,
   "tps": 37.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-xrwacrbc-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: hhybc22r; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 635.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 635.2 30.5 22.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xrwacrbc",
   "source_url": "https://omlx.ai/benchmarks/performance/xrwacrbc",
   "tg_tokens": null,
   "tps": 30.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-33jfmnwf-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 62.1 GB at this context; duplicate traces of the same session: 0vpqn2s1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 885.4,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 3bit; 4k 885.4 55.2 62.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/33jfmnwf",
   "source_url": "https://omlx.ai/benchmarks/performance/33jfmnwf",
   "tg_tokens": null,
   "tps": 55.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-33jfmnwf-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 62.5 GB at this context; duplicate traces of the same session: 0vpqn2s1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1054.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 3bit; 16k 1,054 59.5 62.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/33jfmnwf",
   "source_url": "https://omlx.ai/benchmarks/performance/33jfmnwf",
   "tg_tokens": null,
   "tps": 59.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-panip486-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 62.5 GB; TTFT 15651 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1047.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 3bit; 16k tokens 1,047 PP tok/s 15.2 TG tok/s 15651 TTFT (ms) 62.5 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/panip486",
   "source_url": "https://omlx.ai/benchmarks/performance/panip486",
   "tg_tokens": null,
   "tps": 15.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5gd93425-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 348.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 348.5 53.1 70.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5gd93425",
   "source_url": "https://omlx.ai/benchmarks/performance/5gd93425",
   "tg_tokens": null,
   "tps": 53.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5gd93425-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1092.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,092 39.4 71.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5gd93425",
   "source_url": "https://omlx.ai/benchmarks/performance/5gd93425",
   "tg_tokens": null,
   "tps": 39.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-r129q39b-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 185.9 GB; TTFT 29330 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4469.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 128k tokens 4,469 PP tok/s 65.9 TG tok/s 29330 TTFT (ms) 185.9 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r129q39b",
   "source_url": "https://omlx.ai/benchmarks/performance/r129q39b",
   "tg_tokens": null,
   "tps": 65.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-gvx1mqye-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 185.9 GB; TTFT 29368 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4463.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 128k tokens 4,463 PP tok/s 64.9 TG tok/s 29368 TTFT (ms) 185.9 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gvx1mqye",
   "source_url": "https://omlx.ai/benchmarks/performance/gvx1mqye",
   "tg_tokens": null,
   "tps": 64.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-dod4m4lp-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 183.4 GB; TTFT 14282 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4589.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 64k tokens 4,589 PP tok/s 68.1 TG tok/s 14282 TTFT (ms) 183.4 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dod4m4lp",
   "source_url": "https://omlx.ai/benchmarks/performance/dod4m4lp",
   "tg_tokens": null,
   "tps": 68.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-15t3rirv-ctx131072",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 14.8 GB; TTFT 491275 ms; duplicate traces of the same session: none; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 266.8,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 128k tokens 266.8 PP tok/s 17.0 TG tok/s 491275 TTFT (ms) 14.8 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/15t3rirv",
   "source_url": "https://omlx.ai/benchmarks/performance/15t3rirv",
   "tg_tokens": null,
   "tps": 17.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-1u1hq35f-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: rw78j54f; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1322.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 1,322 82.8 27.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1u1hq35f",
   "source_url": "https://omlx.ai/benchmarks/performance/1u1hq35f",
   "tg_tokens": null,
   "tps": 82.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-1u1hq35f-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: rw78j54f; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2197.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 2,197 81.2 28.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1u1hq35f",
   "source_url": "https://omlx.ai/benchmarks/performance/1u1hq35f",
   "tg_tokens": null,
   "tps": 81.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fp3qn2s6-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: u20eha8q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1038.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "qwen3.8-flash-next-oq4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,038 83.0 70.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fp3qn2s6",
   "source_url": "https://omlx.ai/benchmarks/performance/fp3qn2s6",
   "tg_tokens": null,
   "tps": 83.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fp3qn2s6-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: u20eha8q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2044.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "qwen3.8-flash-next-oq4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,044 84.1 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fp3qn2s6",
   "source_url": "https://omlx.ai/benchmarks/performance/fp3qn2s6",
   "tg_tokens": null,
   "tps": 84.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-v3ckad3r-ctx4096",
   "model": "LFM2.5-2.6B",
   "notes": "MLX bf16 matrix-oQ4e build; oMLX community benchmark trace; peak memory 2.6 GB at this context; duplicate traces of the same session: h8hlj5vq, imvd1mkc; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1315.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-2.6B-MLX-oQ4e-bf16-matrixoQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 1,315 124.6 2.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v3ckad3r",
   "source_url": "https://omlx.ai/benchmarks/performance/v3ckad3r",
   "tg_tokens": null,
   "tps": 124.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-v3ckad3r-ctx16384",
   "model": "LFM2.5-2.6B",
   "notes": "MLX bf16 matrix-oQ4e build; oMLX community benchmark trace; peak memory 2.8 GB at this context; duplicate traces of the same session: h8hlj5vq, imvd1mkc; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1247.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-2.6B-MLX-oQ4e-bf16-matrixoQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 1,247 108.6 2.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v3ckad3r",
   "source_url": "https://omlx.ai/benchmarks/performance/v3ckad3r",
   "tg_tokens": null,
   "tps": 108.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-v3ckad3r-ctx65536",
   "model": "LFM2.5-2.6B",
   "notes": "MLX bf16 matrix-oQ4e build; oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: h8hlj5vq, imvd1mkc; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 914.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-2.6B-MLX-oQ4e-bf16-matrixoQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 914.0 67.0 3.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v3ckad3r",
   "source_url": "https://omlx.ai/benchmarks/performance/v3ckad3r",
   "tg_tokens": null,
   "tps": 67.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ykuxq70p-ctx131072",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 128k context; peak memory 27.2 GB; TTFT 28662 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4573.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 128k tokens 4,573 PP tok/s 66.7 TG tok/s 28662 TTFT (ms) 27.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ykuxq70p",
   "source_url": "https://omlx.ai/benchmarks/performance/ykuxq70p",
   "tg_tokens": null,
   "tps": 66.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qt7w3a59-ctx1024",
   "model": "Muse Glimmer",
   "notes": "oMLX community benchmark trace; peak memory 15.2 GB at this context; duplicate traces of the same session: getvrec5, lsjbln05, tii5nqym; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 484.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-30B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 484.1 18.1 15.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qt7w3a59",
   "source_url": "https://omlx.ai/benchmarks/performance/qt7w3a59",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qt7w3a59-ctx4096",
   "model": "Muse Glimmer",
   "notes": "oMLX community benchmark trace; peak memory 15.5 GB at this context; duplicate traces of the same session: getvrec5, lsjbln05, tii5nqym; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 505.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-30B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 505.8 17.9 15.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qt7w3a59",
   "source_url": "https://omlx.ai/benchmarks/performance/qt7w3a59",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qt7w3a59-ctx8192",
   "model": "Muse Glimmer",
   "notes": "oMLX community benchmark trace; peak memory 15.5 GB at this context; duplicate traces of the same session: getvrec5, lsjbln05, tii5nqym; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 495.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-30B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 495.7 17.8 15.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qt7w3a59",
   "source_url": "https://omlx.ai/benchmarks/performance/qt7w3a59",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qt7w3a59-ctx16384",
   "model": "Muse Glimmer",
   "notes": "oMLX community benchmark trace; peak memory 15.6 GB at this context; duplicate traces of the same session: getvrec5, lsjbln05, tii5nqym; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 444.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-30B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 444.2 17.5 15.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qt7w3a59",
   "source_url": "https://omlx.ai/benchmarks/performance/qt7w3a59",
   "tg_tokens": null,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-gmynlj5z-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2386.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 2,386 94.9 3.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gmynlj5z",
   "source_url": "https://omlx.ai/benchmarks/performance/gmynlj5z",
   "tg_tokens": null,
   "tps": 94.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-gmynlj5z-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2449.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,449 90.0 3.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gmynlj5z",
   "source_url": "https://omlx.ai/benchmarks/performance/gmynlj5z",
   "tg_tokens": null,
   "tps": 90.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-gmynlj5z-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.1 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2488.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 2,488 83.7 4.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gmynlj5z",
   "source_url": "https://omlx.ai/benchmarks/performance/gmynlj5z",
   "tg_tokens": null,
   "tps": 83.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-gmynlj5z-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.3 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2358.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 2,358 73.5 4.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gmynlj5z",
   "source_url": "https://omlx.ai/benchmarks/performance/gmynlj5z",
   "tg_tokens": null,
   "tps": 73.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-wc9njur0-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 4shsyh62; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2116.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 2,116 115.3 21.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wc9njur0",
   "source_url": "https://omlx.ai/benchmarks/performance/wc9njur0",
   "tg_tokens": null,
   "tps": 115.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-wc9njur0-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 4shsyh62; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2829.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,829 108.4 22.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wc9njur0",
   "source_url": "https://omlx.ai/benchmarks/performance/wc9njur0",
   "tg_tokens": null,
   "tps": 108.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-wc9njur0-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 4shsyh62; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2766.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 2,766 133.8 22.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wc9njur0",
   "source_url": "https://omlx.ai/benchmarks/performance/wc9njur0",
   "tg_tokens": null,
   "tps": 133.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-wc9njur0-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 4shsyh62; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2546.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 2,546 110.9 22.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wc9njur0",
   "source_url": "https://omlx.ai/benchmarks/performance/wc9njur0",
   "tg_tokens": null,
   "tps": 110.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-uw4dpgih-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: tp273n0j; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 73.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 73.3 7.8 18.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uw4dpgih",
   "source_url": "https://omlx.ai/benchmarks/performance/uw4dpgih",
   "tg_tokens": null,
   "tps": 7.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-uw4dpgih-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: tp273n0j; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 69.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 69.0 7.0 19.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uw4dpgih",
   "source_url": "https://omlx.ai/benchmarks/performance/uw4dpgih",
   "tg_tokens": null,
   "tps": 7.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-b6kz8cz7-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: az79r8bi; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 136.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 136.1 27.7 24.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b6kz8cz7",
   "source_url": "https://omlx.ai/benchmarks/performance/b6kz8cz7",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-b6kz8cz7-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: az79r8bi; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 139.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 139.3 23.7 27.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b6kz8cz7",
   "source_url": "https://omlx.ai/benchmarks/performance/b6kz8cz7",
   "tg_tokens": null,
   "tps": 23.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-kwa7hbpj-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: 8ke0g48i, lyqav9it; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 427.8,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 427.8 79.3 22.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kwa7hbpj",
   "source_url": "https://omlx.ai/benchmarks/performance/kwa7hbpj",
   "tg_tokens": null,
   "tps": 79.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-kwa7hbpj-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: 8ke0g48i, lyqav9it; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 444.9,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 444.9 72.1 25.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kwa7hbpj",
   "source_url": "https://omlx.ai/benchmarks/performance/kwa7hbpj",
   "tg_tokens": null,
   "tps": 72.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-kwa7hbpj-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: 8ke0g48i, lyqav9it; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 523.6,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 523.6 94.6 25.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kwa7hbpj",
   "source_url": "https://omlx.ai/benchmarks/performance/kwa7hbpj",
   "tg_tokens": null,
   "tps": 94.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-kwa7hbpj-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: 8ke0g48i, lyqav9it; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 517.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 517.0 78.6 26.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kwa7hbpj",
   "source_url": "https://omlx.ai/benchmarks/performance/kwa7hbpj",
   "tg_tokens": null,
   "tps": 78.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-kwa7hbpj-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.2 GB at this context; duplicate traces of the same session: 8ke0g48i, lyqav9it; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 477.5,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 477.5 155.7 28.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kwa7hbpj",
   "source_url": "https://omlx.ai/benchmarks/performance/kwa7hbpj",
   "tg_tokens": null,
   "tps": 155.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-kwa7hbpj-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.4 GB at this context; duplicate traces of the same session: 8ke0g48i, lyqav9it; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 404.5,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 404.5 71.9 31.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kwa7hbpj",
   "source_url": "https://omlx.ai/benchmarks/performance/kwa7hbpj",
   "tg_tokens": null,
   "tps": 71.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-q3kc51bk-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: p33e3ie6; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 428.3,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 428.3 77.0 22.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q3kc51bk",
   "source_url": "https://omlx.ai/benchmarks/performance/q3kc51bk",
   "tg_tokens": null,
   "tps": 77.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-q3kc51bk-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: p33e3ie6; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 450.2,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 450.2 84.9 25.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q3kc51bk",
   "source_url": "https://omlx.ai/benchmarks/performance/q3kc51bk",
   "tg_tokens": null,
   "tps": 84.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nxky8b43-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: daewox66; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 426.1,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 426.1 82.7 22.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nxky8b43",
   "source_url": "https://omlx.ai/benchmarks/performance/nxky8b43",
   "tg_tokens": null,
   "tps": 82.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nxky8b43-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: daewox66; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 450.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 450.0 74.4 25.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nxky8b43",
   "source_url": "https://omlx.ai/benchmarks/performance/nxky8b43",
   "tg_tokens": null,
   "tps": 74.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2w4yv2qb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: oshvhh96; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 431.2,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 431.2 77.3 22.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2w4yv2qb",
   "source_url": "https://omlx.ai/benchmarks/performance/2w4yv2qb",
   "tg_tokens": null,
   "tps": 77.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2w4yv2qb-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: oshvhh96; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 463.7,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 463.7 73.4 25.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2w4yv2qb",
   "source_url": "https://omlx.ai/benchmarks/performance/2w4yv2qb",
   "tg_tokens": null,
   "tps": 73.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-sj946xeq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: f56gb9rw; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 413.2,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 413.2 37.8 22.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sj946xeq",
   "source_url": "https://omlx.ai/benchmarks/performance/sj946xeq",
   "tg_tokens": null,
   "tps": 37.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-sj946xeq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: f56gb9rw; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 461.8,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 461.8 65.2 25.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sj946xeq",
   "source_url": "https://omlx.ai/benchmarks/performance/sj946xeq",
   "tg_tokens": null,
   "tps": 65.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-m2ucdnrx-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 0mdmsmy2, 22svf2il, 8fegd5lh, cymc6to2, e5rb77hp, i8vgqx9y, ozfm935h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 253.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 253.5 50.6 21.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m2ucdnrx",
   "source_url": "https://omlx.ai/benchmarks/performance/m2ucdnrx",
   "tg_tokens": null,
   "tps": 50.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-m2ucdnrx-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: 0mdmsmy2, 22svf2il, 8fegd5lh, cymc6to2, e5rb77hp, i8vgqx9y, ozfm935h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 334.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 334.7 76.5 25.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m2ucdnrx",
   "source_url": "https://omlx.ai/benchmarks/performance/m2ucdnrx",
   "tg_tokens": null,
   "tps": 76.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-m2ucdnrx-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.6 GB at this context; duplicate traces of the same session: 0mdmsmy2, 22svf2il, 8fegd5lh, cymc6to2, e5rb77hp, i8vgqx9y, ozfm935h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 338.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 338.4 34.3 25.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m2ucdnrx",
   "source_url": "https://omlx.ai/benchmarks/performance/m2ucdnrx",
   "tg_tokens": null,
   "tps": 34.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-m2ucdnrx-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: 0mdmsmy2, 22svf2il, 8fegd5lh, cymc6to2, e5rb77hp, i8vgqx9y, ozfm935h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 322.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 322.7 52.1 26.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m2ucdnrx",
   "source_url": "https://omlx.ai/benchmarks/performance/m2ucdnrx",
   "tg_tokens": null,
   "tps": 52.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-m2ucdnrx-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.3 GB at this context; duplicate traces of the same session: 0mdmsmy2, 22svf2il, 8fegd5lh, cymc6to2, e5rb77hp, i8vgqx9y, ozfm935h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 292.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 292.9 42.7 28.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m2ucdnrx",
   "source_url": "https://omlx.ai/benchmarks/performance/m2ucdnrx",
   "tg_tokens": null,
   "tps": 42.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-m2ucdnrx-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.0 GB at this context; duplicate traces of the same session: 0mdmsmy2, 22svf2il, 8fegd5lh, cymc6to2, e5rb77hp, i8vgqx9y, ozfm935h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 247.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 247.5 35.5 31.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m2ucdnrx",
   "source_url": "https://omlx.ai/benchmarks/performance/m2ucdnrx",
   "tg_tokens": null,
   "tps": 35.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-m2ucdnrx-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.3 GB at this context; duplicate traces of the same session: 0mdmsmy2, 22svf2il, 8fegd5lh, cymc6to2, e5rb77hp, i8vgqx9y, ozfm935h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 163.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 163.5 37.6 32.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m2ucdnrx",
   "source_url": "https://omlx.ai/benchmarks/performance/m2ucdnrx",
   "tg_tokens": null,
   "tps": 37.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-m2ucdnrx-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: 0mdmsmy2, 22svf2il, 8fegd5lh, cymc6to2, e5rb77hp, i8vgqx9y, ozfm935h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 135.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 195k 135.0 35.3 37.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m2ucdnrx",
   "source_url": "https://omlx.ai/benchmarks/performance/m2ucdnrx",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-hua2htu2-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: nxmt5tu5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 586.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 586.3 38.8 17.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hua2htu2",
   "source_url": "https://omlx.ai/benchmarks/performance/hua2htu2",
   "tg_tokens": null,
   "tps": 38.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-hua2htu2-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.1 GB at this context; duplicate traces of the same session: nxmt5tu5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 364.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 128k 364.3 26.3 29.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hua2htu2",
   "source_url": "https://omlx.ai/benchmarks/performance/hua2htu2",
   "tg_tokens": null,
   "tps": 26.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-y90ck1rq-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: 0f9g6ev0, ke2d2u6a, mhhy8fk1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 314.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 314.8 38.5 33.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y90ck1rq",
   "source_url": "https://omlx.ai/benchmarks/performance/y90ck1rq",
   "tg_tokens": null,
   "tps": 38.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-y90ck1rq-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: 0f9g6ev0, ke2d2u6a, mhhy8fk1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 334.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 334.7 51.0 37.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y90ck1rq",
   "source_url": "https://omlx.ai/benchmarks/performance/y90ck1rq",
   "tg_tokens": null,
   "tps": 51.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-y90ck1rq-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.8 GB at this context; duplicate traces of the same session: 0f9g6ev0, ke2d2u6a, mhhy8fk1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 329.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 8k 329.6 49.1 37.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y90ck1rq",
   "source_url": "https://omlx.ai/benchmarks/performance/y90ck1rq",
   "tg_tokens": null,
   "tps": 49.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-y90ck1rq-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.6 GB at this context; duplicate traces of the same session: 0f9g6ev0, ke2d2u6a, mhhy8fk1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 317.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-uncensored-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 16k 317.4 53.1 38.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y90ck1rq",
   "source_url": "https://omlx.ai/benchmarks/performance/y90ck1rq",
   "tg_tokens": null,
   "tps": 53.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-z9csldb2-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.8 GB at this context; duplicate traces of the same session: 9t5ws9e1, tp64pl0s; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 749.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 749.4 66.8 100.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z9csldb2",
   "source_url": "https://omlx.ai/benchmarks/performance/z9csldb2",
   "tg_tokens": null,
   "tps": 66.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-z9csldb2-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.4 GB at this context; duplicate traces of the same session: 9t5ws9e1, tp64pl0s; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1028.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 1,028 60.1 101.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z9csldb2",
   "source_url": "https://omlx.ai/benchmarks/performance/z9csldb2",
   "tg_tokens": null,
   "tps": 60.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-z9csldb2-ctx8192",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.8 GB at this context; duplicate traces of the same session: 9t5ws9e1, tp64pl0s; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1347.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 8k 1,347 76.1 101.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z9csldb2",
   "source_url": "https://omlx.ai/benchmarks/performance/z9csldb2",
   "tg_tokens": null,
   "tps": 76.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vff8x12e-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 142.6 GB at this context; duplicate traces of the same session: cjfswuxr, mwiwb80e; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1045.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 1,045 92.2 142.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vff8x12e",
   "source_url": "https://omlx.ai/benchmarks/performance/vff8x12e",
   "tg_tokens": null,
   "tps": 92.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vff8x12e-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: cjfswuxr, mwiwb80e; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1383.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 1,383 94.9 143.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vff8x12e",
   "source_url": "https://omlx.ai/benchmarks/performance/vff8x12e",
   "tg_tokens": null,
   "tps": 94.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vff8x12e-ctx8192",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 143.6 GB at this context; duplicate traces of the same session: cjfswuxr, mwiwb80e; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1372.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 8k 1,372 97.0 143.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vff8x12e",
   "source_url": "https://omlx.ai/benchmarks/performance/vff8x12e",
   "tg_tokens": null,
   "tps": 97.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jp2nkgw4-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: i5qlh582, q4dwwtu5, s7gxqwk5, txpb66ry, un1isqw5, yci8t6d8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1309.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,309 73.5 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jp2nkgw4",
   "source_url": "https://omlx.ai/benchmarks/performance/jp2nkgw4",
   "tg_tokens": null,
   "tps": 73.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jp2nkgw4-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.0 GB at this context; duplicate traces of the same session: i5qlh582, q4dwwtu5, s7gxqwk5, txpb66ry, un1isqw5, yci8t6d8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2606.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,606 84.5 72.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jp2nkgw4",
   "source_url": "https://omlx.ai/benchmarks/performance/jp2nkgw4",
   "tg_tokens": null,
   "tps": 84.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jp2nkgw4-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.0 GB at this context; duplicate traces of the same session: i5qlh582, q4dwwtu5, s7gxqwk5, txpb66ry, un1isqw5, yci8t6d8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2588.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,588 72.6 73.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jp2nkgw4",
   "source_url": "https://omlx.ai/benchmarks/performance/jp2nkgw4",
   "tg_tokens": null,
   "tps": 72.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jp2nkgw4-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.6 GB at this context; duplicate traces of the same session: i5qlh582, q4dwwtu5, s7gxqwk5, txpb66ry, un1isqw5, yci8t6d8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2569.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,569 72.6 73.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jp2nkgw4",
   "source_url": "https://omlx.ai/benchmarks/performance/jp2nkgw4",
   "tg_tokens": null,
   "tps": 72.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jp2nkgw4-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.7 GB at this context; duplicate traces of the same session: i5qlh582, q4dwwtu5, s7gxqwk5, txpb66ry, un1isqw5, yci8t6d8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2415.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,415 84.8 74.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jp2nkgw4",
   "source_url": "https://omlx.ai/benchmarks/performance/jp2nkgw4",
   "tg_tokens": null,
   "tps": 84.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jp2nkgw4-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 77.0 GB at this context; duplicate traces of the same session: i5qlh582, q4dwwtu5, s7gxqwk5, txpb66ry, un1isqw5, yci8t6d8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2459.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,459 73.3 77.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jp2nkgw4",
   "source_url": "https://omlx.ai/benchmarks/performance/jp2nkgw4",
   "tg_tokens": null,
   "tps": 73.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jp2nkgw4-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.8 GB at this context; duplicate traces of the same session: i5qlh582, q4dwwtu5, s7gxqwk5, txpb66ry, un1isqw5, yci8t6d8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2359.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,359 61.5 79.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jp2nkgw4",
   "source_url": "https://omlx.ai/benchmarks/performance/jp2nkgw4",
   "tg_tokens": null,
   "tps": 61.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5t6fi05x-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP requant build; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: 41qxu4st, 9r41a06g, idl3wmld, tg751mmx, wkhb0pui, xqnrp16y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1119.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp_requant on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,119 67.3 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5t6fi05x",
   "source_url": "https://omlx.ai/benchmarks/performance/5t6fi05x",
   "tg_tokens": null,
   "tps": 67.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5t6fi05x-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP requant build; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: 41qxu4st, 9r41a06g, idl3wmld, tg751mmx, wkhb0pui, xqnrp16y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2086.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp_requant on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,086 62.8 71.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5t6fi05x",
   "source_url": "https://omlx.ai/benchmarks/performance/5t6fi05x",
   "tg_tokens": null,
   "tps": 62.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5t6fi05x-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP requant build; oMLX community benchmark trace; peak memory 72.0 GB at this context; duplicate traces of the same session: 41qxu4st, 9r41a06g, idl3wmld, tg751mmx, wkhb0pui, xqnrp16y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2614.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp_requant on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,614 84.4 72.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5t6fi05x",
   "source_url": "https://omlx.ai/benchmarks/performance/5t6fi05x",
   "tg_tokens": null,
   "tps": 84.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5t6fi05x-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP requant build; oMLX community benchmark trace; peak memory 73.0 GB at this context; duplicate traces of the same session: 41qxu4st, 9r41a06g, idl3wmld, tg751mmx, wkhb0pui, xqnrp16y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2668.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp_requant on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,668 80.2 73.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5t6fi05x",
   "source_url": "https://omlx.ai/benchmarks/performance/5t6fi05x",
   "tg_tokens": null,
   "tps": 80.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5t6fi05x-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP requant build; oMLX community benchmark trace; peak memory 74.7 GB at this context; duplicate traces of the same session: 41qxu4st, 9r41a06g, idl3wmld, tg751mmx, wkhb0pui, xqnrp16y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2400.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp_requant on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,400 76.6 74.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5t6fi05x",
   "source_url": "https://omlx.ai/benchmarks/performance/5t6fi05x",
   "tg_tokens": null,
   "tps": 76.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5t6fi05x-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP requant build; oMLX community benchmark trace; peak memory 77.0 GB at this context; duplicate traces of the same session: 41qxu4st, 9r41a06g, idl3wmld, tg751mmx, wkhb0pui, xqnrp16y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2300.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp_requant on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,300 67.5 77.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5t6fi05x",
   "source_url": "https://omlx.ai/benchmarks/performance/5t6fi05x",
   "tg_tokens": null,
   "tps": 67.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5t6fi05x-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP requant build; oMLX community benchmark trace; peak memory 79.8 GB at this context; duplicate traces of the same session: 41qxu4st, 9r41a06g, idl3wmld, tg751mmx, wkhb0pui, xqnrp16y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2297.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp_requant on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,297 59.1 79.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5t6fi05x",
   "source_url": "https://omlx.ai/benchmarks/performance/5t6fi05x",
   "tg_tokens": null,
   "tps": 59.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-cr7ylrrz-ctx4096",
   "model": "LFM2.5-2.6B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 4k context; peak memory 2.4 GB; TTFT 3158 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1297.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "LFM2.5-2.6B-MLX-oQ3.5e on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 3bit; 4k tokens 1,297 PP tok/s 140.5 TG tok/s 3158 TTFT (ms) 2.4 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cr7ylrrz",
   "source_url": "https://omlx.ai/benchmarks/performance/cr7ylrrz",
   "tg_tokens": null,
   "tps": 140.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-rcab4w16-ctx4096",
   "model": "LFM2.5-2.6B",
   "notes": "MLX bf16 matrix-oQ4e build; oMLX community benchmark trace, single run at 4k context; peak memory 2.6 GB; TTFT 3187 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1285.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-2.6B-MLX-oQ4e-bf16-matrixoQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k tokens 1,285 PP tok/s 103.3 TG tok/s 3187 TTFT (ms) 2.6 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rcab4w16",
   "source_url": "https://omlx.ai/benchmarks/performance/rcab4w16",
   "tg_tokens": null,
   "tps": 103.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-cjl5qmcp-ctx4096",
   "model": "LFM2.5-2.6B",
   "notes": "MLX fp16 build; oMLX community benchmark trace, single run at 4k context; peak memory 2.6 GB; TTFT 3194 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1283.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-2.6B-MLX-oQ4e-fp16 on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k tokens 1,283 PP tok/s 122.4 TG tok/s 3194 TTFT (ms) 2.6 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cjl5qmcp",
   "source_url": "https://omlx.ai/benchmarks/performance/cjl5qmcp",
   "tg_tokens": null,
   "tps": 122.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-l39b9b96-ctx4096",
   "model": "LFM2.5-2.6B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 4k context; peak memory 2.9 GB; TTFT 3223 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1271.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "LFM2.5-2.6B-MLX-oQ5e on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 5bit; 4k tokens 1,271 PP tok/s 106.5 TG tok/s 3223 TTFT (ms) 2.9 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l39b9b96",
   "source_url": "https://omlx.ai/benchmarks/performance/l39b9b96",
   "tg_tokens": null,
   "tps": 106.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-z1km8wix-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 4k context; peak memory 24.8 GB; TTFT 1340 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3056.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k tokens 3,056 PP tok/s 122.7 TG tok/s 1340 TTFT (ms) 24.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z1km8wix",
   "source_url": "https://omlx.ai/benchmarks/performance/z1km8wix",
   "tg_tokens": null,
   "tps": 122.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-tmc7curk-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 13.6 GB at this context; duplicate traces of the same session: 0dkzbzqq; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 964.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e-fp16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 964.6 89.3 13.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tmc7curk",
   "source_url": "https://omlx.ai/benchmarks/performance/tmc7curk",
   "tg_tokens": null,
   "tps": 89.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-tmc7curk-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 13.9 GB at this context; duplicate traces of the same session: 0dkzbzqq; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1306.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e-fp16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 1,306 86.6 13.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tmc7curk",
   "source_url": "https://omlx.ai/benchmarks/performance/tmc7curk",
   "tg_tokens": null,
   "tps": 86.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-x4gtsvs5-ctx1024",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: kay0y8m9, qjrj0jc4, v2mtf7xy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1663.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "MiniCPM-V-4.7-35B-A3B-mxfp4 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,663 86.5 22.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x4gtsvs5",
   "source_url": "https://omlx.ai/benchmarks/performance/x4gtsvs5",
   "tg_tokens": null,
   "tps": 86.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-x4gtsvs5-ctx4096",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: kay0y8m9, qjrj0jc4, v2mtf7xy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1924.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "MiniCPM-V-4.7-35B-A3B-mxfp4 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,924 85.3 22.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x4gtsvs5",
   "source_url": "https://omlx.ai/benchmarks/performance/x4gtsvs5",
   "tg_tokens": null,
   "tps": 85.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-x4gtsvs5-ctx8192",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: kay0y8m9, qjrj0jc4, v2mtf7xy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1770.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "MiniCPM-V-4.7-35B-A3B-mxfp4 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,770 82.5 22.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x4gtsvs5",
   "source_url": "https://omlx.ai/benchmarks/performance/x4gtsvs5",
   "tg_tokens": null,
   "tps": 82.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-x4gtsvs5-ctx16384",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: kay0y8m9, qjrj0jc4, v2mtf7xy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1533.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "MiniCPM-V-4.7-35B-A3B-mxfp4 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,533 78.5 23.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x4gtsvs5",
   "source_url": "https://omlx.ai/benchmarks/performance/x4gtsvs5",
   "tg_tokens": null,
   "tps": 78.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-a5tqqvlj-ctx1024",
   "model": "Muse-Glimmer-30B-heretic",
   "notes": "oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: bjrl4kym; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 375.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-heretic-oQ4e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 375.9 16.9 15.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a5tqqvlj",
   "source_url": "https://omlx.ai/benchmarks/performance/a5tqqvlj",
   "tg_tokens": null,
   "tps": 16.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-a5tqqvlj-ctx16384",
   "model": "Muse-Glimmer-30B-heretic",
   "notes": "oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: bjrl4kym; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 456.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-heretic-oQ4e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 456.3 15.7 16.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a5tqqvlj",
   "source_url": "https://omlx.ai/benchmarks/performance/a5tqqvlj",
   "tg_tokens": null,
   "tps": 15.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-j2dxparz-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.1 GB at this context; duplicate traces of the same session: mvyfyosy, ytzqvn0e; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 640.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 16k 640.6 53.6 31.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j2dxparz",
   "source_url": "https://omlx.ai/benchmarks/performance/j2dxparz",
   "tg_tokens": null,
   "tps": 53.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-j2dxparz-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.4 GB at this context; duplicate traces of the same session: mvyfyosy, ytzqvn0e; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 520.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 32k 520.2 60.0 31.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j2dxparz",
   "source_url": "https://omlx.ai/benchmarks/performance/j2dxparz",
   "tg_tokens": null,
   "tps": 60.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-j2dxparz-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: mvyfyosy, ytzqvn0e; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 372.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 64k 372.3 43.1 32.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j2dxparz",
   "source_url": "https://omlx.ai/benchmarks/performance/j2dxparz",
   "tg_tokens": null,
   "tps": 43.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-br7213l5-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.0 GB at this context; duplicate traces of the same session: bn5rp0by, gpd3y6h6; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 667.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 16k 667.5 56.2 29.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/br7213l5",
   "source_url": "https://omlx.ai/benchmarks/performance/br7213l5",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-br7213l5-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.3 GB at this context; duplicate traces of the same session: bn5rp0by, gpd3y6h6; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 550.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 32k 550.9 49.4 29.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/br7213l5",
   "source_url": "https://omlx.ai/benchmarks/performance/br7213l5",
   "tg_tokens": null,
   "tps": 49.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-br7213l5-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: bn5rp0by, gpd3y6h6; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 409.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-35B-A3B-oQ6e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 64k 409.3 37.5 29.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/br7213l5",
   "source_url": "https://omlx.ai/benchmarks/performance/br7213l5",
   "tg_tokens": null,
   "tps": 37.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-mxuvv3ys-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: w5gu3zkj, wgt0qnmx; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2386.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 2,386 94.9 3.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mxuvv3ys",
   "source_url": "https://omlx.ai/benchmarks/performance/mxuvv3ys",
   "tg_tokens": null,
   "tps": 94.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-mxuvv3ys-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: w5gu3zkj, wgt0qnmx; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2449.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,449 90.0 3.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mxuvv3ys",
   "source_url": "https://omlx.ai/benchmarks/performance/mxuvv3ys",
   "tg_tokens": null,
   "tps": 90.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-mxuvv3ys-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.1 GB at this context; duplicate traces of the same session: w5gu3zkj, wgt0qnmx; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2488.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 2,488 83.7 4.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mxuvv3ys",
   "source_url": "https://omlx.ai/benchmarks/performance/mxuvv3ys",
   "tg_tokens": null,
   "tps": 83.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-mxuvv3ys-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.3 GB at this context; duplicate traces of the same session: w5gu3zkj, wgt0qnmx; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2358.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 2,358 73.5 4.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mxuvv3ys",
   "source_url": "https://omlx.ai/benchmarks/performance/mxuvv3ys",
   "tg_tokens": null,
   "tps": 73.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-91o1gr7k-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 171ds4hm; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2116.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 2,116 115.3 21.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/91o1gr7k",
   "source_url": "https://omlx.ai/benchmarks/performance/91o1gr7k",
   "tg_tokens": null,
   "tps": 115.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-91o1gr7k-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 171ds4hm; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2829.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,829 108.4 22.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/91o1gr7k",
   "source_url": "https://omlx.ai/benchmarks/performance/91o1gr7k",
   "tg_tokens": null,
   "tps": 108.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-91o1gr7k-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 171ds4hm; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2766.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 2,766 133.8 22.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/91o1gr7k",
   "source_url": "https://omlx.ai/benchmarks/performance/91o1gr7k",
   "tg_tokens": null,
   "tps": 133.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-91o1gr7k-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 171ds4hm; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2546.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 2,546 110.9 22.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/91o1gr7k",
   "source_url": "https://omlx.ai/benchmarks/performance/91o1gr7k",
   "tg_tokens": null,
   "tps": 110.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-lw8r933n-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: qnlyeaat; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 129.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 129.6 15.1 17.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lw8r933n",
   "source_url": "https://omlx.ai/benchmarks/performance/lw8r933n",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-lw8r933n-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: qnlyeaat; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 125.6 14.0 18.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lw8r933n",
   "source_url": "https://omlx.ai/benchmarks/performance/lw8r933n",
   "tg_tokens": null,
   "tps": 14.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ws12i4b3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: aig3iyy2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 768.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 768.5 17.8 28.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ws12i4b3",
   "source_url": "https://omlx.ai/benchmarks/performance/ws12i4b3",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ws12i4b3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: aig3iyy2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 862.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 862.4 17.8 29.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ws12i4b3",
   "source_url": "https://omlx.ai/benchmarks/performance/ws12i4b3",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fhy501co-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: yak9ausb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 932.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 932.6 43.2 16.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fhy501co",
   "source_url": "https://omlx.ai/benchmarks/performance/fhy501co",
   "tg_tokens": null,
   "tps": 43.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fhy501co-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: yak9ausb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 943.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 943.5 44.7 18.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fhy501co",
   "source_url": "https://omlx.ai/benchmarks/performance/fhy501co",
   "tg_tokens": null,
   "tps": 44.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7wm21ln0-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: umlwj66l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 965.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 965.5 45.1 16.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7wm21ln0",
   "source_url": "https://omlx.ai/benchmarks/performance/7wm21ln0",
   "tg_tokens": null,
   "tps": 45.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7wm21ln0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: umlwj66l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 932.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 932.6 41.9 18.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7wm21ln0",
   "source_url": "https://omlx.ai/benchmarks/performance/7wm21ln0",
   "tg_tokens": null,
   "tps": 41.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-b6w22kit-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: k6ypo3tw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 904.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 904.0 49.3 23.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b6w22kit",
   "source_url": "https://omlx.ai/benchmarks/performance/b6w22kit",
   "tg_tokens": null,
   "tps": 49.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-b6w22kit-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: k6ypo3tw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 896.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 896.0 40.8 24.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b6w22kit",
   "source_url": "https://omlx.ai/benchmarks/performance/b6w22kit",
   "tg_tokens": null,
   "tps": 40.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qg1jz1ir-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 727.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 727.7 38.9 23.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qg1jz1ir",
   "source_url": "https://omlx.ai/benchmarks/performance/qg1jz1ir",
   "tg_tokens": null,
   "tps": 38.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qg1jz1ir-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 883.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 883.9 41.0 24.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qg1jz1ir",
   "source_url": "https://omlx.ai/benchmarks/performance/qg1jz1ir",
   "tg_tokens": null,
   "tps": 41.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-yfbe3j00-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: 0gx489sx, xjh2928y; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 188.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 188.0 20.5 30.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yfbe3j00",
   "source_url": "https://omlx.ai/benchmarks/performance/yfbe3j00",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-yfbe3j00-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.7 GB at this context; duplicate traces of the same session: 0gx489sx, xjh2928y; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 276.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 276.0 20.5 33.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yfbe3j00",
   "source_url": "https://omlx.ai/benchmarks/performance/yfbe3j00",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-yfbe3j00-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: 0gx489sx, xjh2928y; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 268.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 268.0 22.1 34.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yfbe3j00",
   "source_url": "https://omlx.ai/benchmarks/performance/yfbe3j00",
   "tg_tokens": null,
   "tps": 22.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-yfbe3j00-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: 0gx489sx, xjh2928y; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 251.9,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 251.9 24.5 36.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yfbe3j00",
   "source_url": "https://omlx.ai/benchmarks/performance/yfbe3j00",
   "tg_tokens": null,
   "tps": 24.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-sgrais35-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 16.6 GB; TTFT 4597 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 222.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k tokens 222.8 PP tok/s 9.2 TG tok/s 4597 TTFT (ms) 16.6 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sgrais35",
   "source_url": "https://omlx.ai/benchmarks/performance/sgrais35",
   "tg_tokens": null,
   "tps": 9.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-8lvu0xbc-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 16.6 GB; TTFT 4447 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 230.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k tokens 230.3 PP tok/s 8.5 TG tok/s 4447 TTFT (ms) 16.6 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8lvu0xbc",
   "source_url": "https://omlx.ai/benchmarks/performance/8lvu0xbc",
   "tg_tokens": null,
   "tps": 8.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-bx8klpo9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: qdrtt1gk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 236.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 236.0 9.3 17.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bx8klpo9",
   "source_url": "https://omlx.ai/benchmarks/performance/bx8klpo9",
   "tg_tokens": null,
   "tps": 9.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-bx8klpo9-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: qdrtt1gk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 211.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 211.4 11.3 19.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bx8klpo9",
   "source_url": "https://omlx.ai/benchmarks/performance/bx8klpo9",
   "tg_tokens": null,
   "tps": 11.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-87mc1tub-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.5 GB at this context; duplicate traces of the same session: w2nbqalu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 225.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 225.6 18.7 17.5 GB",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/87mc1tub",
   "source_url": "https://omlx.ai/benchmarks/performance/87mc1tub",
   "tg_tokens": null,
   "tps": 18.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-w2nbqalu-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 87mc1tub; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 220.9 17.7 17.6 GB",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w2nbqalu",
   "source_url": "https://omlx.ai/benchmarks/performance/w2nbqalu",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-j37zbhob-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: 0tlt9lin; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 195.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 195.7 25.7 34.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j37zbhob",
   "source_url": "https://omlx.ai/benchmarks/performance/j37zbhob",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-j37zbhob-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 37.9 GB at this context; duplicate traces of the same session: 0tlt9lin; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 202.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 202.4 21.6 37.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j37zbhob",
   "source_url": "https://omlx.ai/benchmarks/performance/j37zbhob",
   "tg_tokens": null,
   "tps": 21.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-o4eloq8s-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: 1c3n6zgp, bus8s97c, j3qo5r3h, n72umha2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 603.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 603.6 56.0 71.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o4eloq8s",
   "source_url": "https://omlx.ai/benchmarks/performance/o4eloq8s",
   "tg_tokens": null,
   "tps": 56.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-o4eloq8s-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: 1c3n6zgp, bus8s97c, j3qo5r3h, n72umha2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 748.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 748.9 51.9 71.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o4eloq8s",
   "source_url": "https://omlx.ai/benchmarks/performance/o4eloq8s",
   "tg_tokens": null,
   "tps": 51.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-o4eloq8s-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.9 GB at this context; duplicate traces of the same session: 1c3n6zgp, bus8s97c, j3qo5r3h, n72umha2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 825.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 825.9 54.6 71.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o4eloq8s",
   "source_url": "https://omlx.ai/benchmarks/performance/o4eloq8s",
   "tg_tokens": null,
   "tps": 54.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-o4eloq8s-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.5 GB at this context; duplicate traces of the same session: 1c3n6zgp, bus8s97c, j3qo5r3h, n72umha2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 818.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 818.5 64.5 72.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o4eloq8s",
   "source_url": "https://omlx.ai/benchmarks/performance/o4eloq8s",
   "tg_tokens": null,
   "tps": 64.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-o4eloq8s-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.7 GB at this context; duplicate traces of the same session: 1c3n6zgp, bus8s97c, j3qo5r3h, n72umha2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 811.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 811.5 61.5 73.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o4eloq8s",
   "source_url": "https://omlx.ai/benchmarks/performance/o4eloq8s",
   "tg_tokens": null,
   "tps": 61.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-c85pc4t3-ctx1024",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 5.1 GB at this context; duplicate traces of the same session: yl673gg0; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2227.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-e2b-8bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 8bit; 1k 2,227 38.4 5.1 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c85pc4t3",
   "source_url": "https://omlx.ai/benchmarks/performance/c85pc4t3",
   "tg_tokens": null,
   "tps": 38.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-c85pc4t3-ctx4096",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 5.2 GB at this context; duplicate traces of the same session: yl673gg0; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2345.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-e2b-8bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 8bit; 4k 2,345 38.0 5.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c85pc4t3",
   "source_url": "https://omlx.ai/benchmarks/performance/c85pc4t3",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-xj1oqnec-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.3 GB at this context; duplicate traces of the same session: 3puw76q5, c1m63mpn, gqja7844; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2402.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 2,402 138.6 177.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xj1oqnec",
   "source_url": "https://omlx.ai/benchmarks/performance/xj1oqnec",
   "tg_tokens": null,
   "tps": 138.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-xj1oqnec-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 177.8 GB at this context; duplicate traces of the same session: 3puw76q5, c1m63mpn, gqja7844; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2414.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,414 118.2 177.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xj1oqnec",
   "source_url": "https://omlx.ai/benchmarks/performance/xj1oqnec",
   "tg_tokens": null,
   "tps": 118.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-xj1oqnec-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 179.3 GB at this context; duplicate traces of the same session: 3puw76q5, c1m63mpn, gqja7844; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2390.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,390 94.5 179.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xj1oqnec",
   "source_url": "https://omlx.ai/benchmarks/performance/xj1oqnec",
   "tg_tokens": null,
   "tps": 94.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-xj1oqnec-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 182.3 GB at this context; duplicate traces of the same session: 3puw76q5, c1m63mpn, gqja7844; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2335.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,335 126.4 182.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xj1oqnec",
   "source_url": "https://omlx.ai/benchmarks/performance/xj1oqnec",
   "tg_tokens": null,
   "tps": 126.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-yhw8bf62-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: 2d3nx3dh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1830.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 1,830 140.5 24.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yhw8bf62",
   "source_url": "https://omlx.ai/benchmarks/performance/yhw8bf62",
   "tg_tokens": null,
   "tps": 140.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-yhw8bf62-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: 2d3nx3dh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3283.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 3,283 136.2 24.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yhw8bf62",
   "source_url": "https://omlx.ai/benchmarks/performance/yhw8bf62",
   "tg_tokens": null,
   "tps": 136.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-zb5xxp39-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: 6io0zmlr; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1496.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 1k 1,496 98.0 13.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zb5xxp39",
   "source_url": "https://omlx.ai/benchmarks/performance/zb5xxp39",
   "tg_tokens": null,
   "tps": 98.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-zb5xxp39-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: 6io0zmlr; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1911.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 4k 1,911 96.0 13.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zb5xxp39",
   "source_url": "https://omlx.ai/benchmarks/performance/zb5xxp39",
   "tg_tokens": null,
   "tps": 96.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fhf5hh00-ctx65536",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 22.2 GB; TTFT 13333 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4915.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "LFM2.5-24B-A2B-oQ5e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k tokens 4,915 PP tok/s 78.3 TG tok/s 13333 TTFT (ms) 22.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fhf5hh00",
   "source_url": "https://omlx.ai/benchmarks/performance/fhf5hh00",
   "tg_tokens": null,
   "tps": 78.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-h5oculb7-ctx1024",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: jvgw31kj, pij25xtz, ux5uv1xg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1233.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,233 139.5 18.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h5oculb7",
   "source_url": "https://omlx.ai/benchmarks/performance/h5oculb7",
   "tg_tokens": null,
   "tps": 139.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-h5oculb7-ctx4096",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: jvgw31kj, pij25xtz, ux5uv1xg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 980.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 980.1 122.6 19.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h5oculb7",
   "source_url": "https://omlx.ai/benchmarks/performance/h5oculb7",
   "tg_tokens": null,
   "tps": 122.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-h5oculb7-ctx8192",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: jvgw31kj, pij25xtz, ux5uv1xg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 960.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 960.7 123.3 19.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h5oculb7",
   "source_url": "https://omlx.ai/benchmarks/performance/h5oculb7",
   "tg_tokens": null,
   "tps": 123.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-h5oculb7-ctx16384",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: jvgw31kj, pij25xtz, ux5uv1xg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 941.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 941.8 120.2 19.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h5oculb7",
   "source_url": "https://omlx.ai/benchmarks/performance/h5oculb7",
   "tg_tokens": null,
   "tps": 120.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-ggpam0vj-ctx1024",
   "model": "Qwen3-4B-Instruct-2507",
   "notes": "oMLX community benchmark trace; peak memory 2.8 GB at this context; duplicate traces of the same session: x4644n4d; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1651.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-4B-Instruct-2507-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 1,651 48.2 2.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ggpam0vj",
   "source_url": "https://omlx.ai/benchmarks/performance/ggpam0vj",
   "tg_tokens": null,
   "tps": 48.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-ggpam0vj-ctx4096",
   "model": "Qwen3-4B-Instruct-2507",
   "notes": "oMLX community benchmark trace; peak memory 3.2 GB at this context; duplicate traces of the same session: x4644n4d; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1543.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-4B-Instruct-2507-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 1,543 39.6 3.2 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ggpam0vj",
   "source_url": "https://omlx.ai/benchmarks/performance/ggpam0vj",
   "tg_tokens": null,
   "tps": 39.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-ljx7xdgf-ctx4096",
   "model": "Qwen3 8B",
   "notes": "oMLX community benchmark trace; peak memory 5.5 GB at this context; duplicate traces of the same session: mmjmc89c; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 167.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-8B-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 4k 167.6 31.7 5.5 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ljx7xdgf",
   "source_url": "https://omlx.ai/benchmarks/performance/ljx7xdgf",
   "tg_tokens": null,
   "tps": 31.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-ljx7xdgf-ctx16384",
   "model": "Qwen3 8B",
   "notes": "oMLX community benchmark trace; peak memory 7.1 GB at this context; duplicate traces of the same session: mmjmc89c; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 135.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-8B-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 4bit; 16k 135.7 22.3 7.1 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ljx7xdgf",
   "source_url": "https://omlx.ai/benchmarks/performance/ljx7xdgf",
   "tg_tokens": null,
   "tps": 22.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-fyt5iyx9-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 16.0 GB; TTFT 838 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1222.0,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.6-35B-A3B-oQ3 on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 3bit; 1k tokens 1,222 PP tok/s 58.5 TG tok/s 838 TTFT (ms) 16.0 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fyt5iyx9",
   "source_url": "https://omlx.ai/benchmarks/performance/fyt5iyx9",
   "tg_tokens": null,
   "tps": 58.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-g4hhvz43-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 52.0 GB at this context; duplicate traces of the same session: 3bb7kb03, lrpjfx1p; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 633.6,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.8-27B on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 633.6 10.2 52.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g4hhvz43",
   "source_url": "https://omlx.ai/benchmarks/performance/g4hhvz43",
   "tg_tokens": null,
   "tps": 10.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-g4hhvz43-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 53.3 GB at this context; duplicate traces of the same session: 3bb7kb03, lrpjfx1p; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 964.8,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.8-27B on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 4k 964.8 10.1 53.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g4hhvz43",
   "source_url": "https://omlx.ai/benchmarks/performance/g4hhvz43",
   "tg_tokens": null,
   "tps": 10.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-g4hhvz43-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 61.1 GB at this context; duplicate traces of the same session: 3bb7kb03, lrpjfx1p; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 618.4,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.8-27B on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 128k 618.4 8.7 61.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g4hhvz43",
   "source_url": "https://omlx.ai/benchmarks/performance/g4hhvz43",
   "tg_tokens": null,
   "tps": 8.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-0pdxzy6t-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: iqtig79k; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 238.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 238.9 8.2 15.9 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0pdxzy6t",
   "source_url": "https://omlx.ai/benchmarks/performance/0pdxzy6t",
   "tg_tokens": null,
   "tps": 8.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-0pdxzy6t-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: iqtig79k; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 244.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 244.2 8.0 17.0 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0pdxzy6t",
   "source_url": "https://omlx.ai/benchmarks/performance/0pdxzy6t",
   "tg_tokens": null,
   "tps": 8.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-g1951h8c-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: 97qnhtkm, a4flw7ld, j7y8xc5q, jpftiuw4, oia34jr3, weagglja; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 812.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 812.6 49.6 29.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g1951h8c",
   "source_url": "https://omlx.ai/benchmarks/performance/g1951h8c",
   "tg_tokens": null,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-g1951h8c-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 30.8 GB at this context; duplicate traces of the same session: 97qnhtkm, a4flw7ld, j7y8xc5q, jpftiuw4, oia34jr3, weagglja; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 807.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 807.8 46.1 30.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g1951h8c",
   "source_url": "https://omlx.ai/benchmarks/performance/g1951h8c",
   "tg_tokens": null,
   "tps": 46.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-g1951h8c-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.1 GB at this context; duplicate traces of the same session: 97qnhtkm, a4flw7ld, j7y8xc5q, jpftiuw4, oia34jr3, weagglja; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 728.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 728.0 40.9 31.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g1951h8c",
   "source_url": "https://omlx.ai/benchmarks/performance/g1951h8c",
   "tg_tokens": null,
   "tps": 40.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-g1951h8c-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.6 GB at this context; duplicate traces of the same session: 97qnhtkm, a4flw7ld, j7y8xc5q, jpftiuw4, oia34jr3, weagglja; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 684.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 684.3 53.6 31.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g1951h8c",
   "source_url": "https://omlx.ai/benchmarks/performance/g1951h8c",
   "tg_tokens": null,
   "tps": 53.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-g1951h8c-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.6 GB at this context; duplicate traces of the same session: 97qnhtkm, a4flw7ld, j7y8xc5q, jpftiuw4, oia34jr3, weagglja; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 634.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 634.7 42.7 32.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g1951h8c",
   "source_url": "https://omlx.ai/benchmarks/performance/g1951h8c",
   "tg_tokens": null,
   "tps": 42.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-g1951h8c-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.6 GB at this context; duplicate traces of the same session: 97qnhtkm, a4flw7ld, j7y8xc5q, jpftiuw4, oia34jr3, weagglja; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 582.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 582.4 37.4 34.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g1951h8c",
   "source_url": "https://omlx.ai/benchmarks/performance/g1951h8c",
   "tg_tokens": null,
   "tps": 37.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-g1951h8c-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 38.7 GB at this context; duplicate traces of the same session: 97qnhtkm, a4flw7ld, j7y8xc5q, jpftiuw4, oia34jr3, weagglja; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 469.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 469.7 34.4 38.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g1951h8c",
   "source_url": "https://omlx.ai/benchmarks/performance/g1951h8c",
   "tg_tokens": null,
   "tps": 34.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5wz0dj01-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 727.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 727.7 38.9 23.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5wz0dj01",
   "source_url": "https://omlx.ai/benchmarks/performance/5wz0dj01",
   "tg_tokens": null,
   "tps": 38.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5wz0dj01-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 883.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 883.9 41.0 24.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5wz0dj01",
   "source_url": "https://omlx.ai/benchmarks/performance/5wz0dj01",
   "tg_tokens": null,
   "tps": 41.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-buvm5piu-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 108.5 GB; TTFT 6399 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1280.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k tokens 1,280 PP tok/s 47.6 TG tok/s 6399 TTFT (ms) 108.5 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/buvm5piu",
   "source_url": "https://omlx.ai/benchmarks/performance/buvm5piu",
   "tg_tokens": null,
   "tps": 47.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-p9x49iru-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 107.7 GB; TTFT 1042 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 983.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k tokens 983.0 PP tok/s 70.9 TG tok/s 1042 TTFT (ms) 107.7 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p9x49iru",
   "source_url": "https://omlx.ai/benchmarks/performance/p9x49iru",
   "tg_tokens": null,
   "tps": 70.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-frn9jw8k-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 107.7 GB; TTFT 1301 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 787.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k tokens 787.1 PP tok/s 19.9 TG tok/s 1301 TTFT (ms) 107.7 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/frn9jw8k",
   "source_url": "https://omlx.ai/benchmarks/performance/frn9jw8k",
   "tg_tokens": null,
   "tps": 19.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rqvfl1sy-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 84.7 GB at this context; duplicate traces of the same session: 1iv18i4u, ilpss66i, ohyjfuxk, t5vbcz5h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 988.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 988.7 56.5 84.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rqvfl1sy",
   "source_url": "https://omlx.ai/benchmarks/performance/rqvfl1sy",
   "tg_tokens": null,
   "tps": 56.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rqvfl1sy-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 85.0 GB at this context; duplicate traces of the same session: 1iv18i4u, ilpss66i, ohyjfuxk, t5vbcz5h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1770.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 1,770 56.8 85.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rqvfl1sy",
   "source_url": "https://omlx.ai/benchmarks/performance/rqvfl1sy",
   "tg_tokens": null,
   "tps": 56.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-rqvfl1sy-ctx8192",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 85.9 GB at this context; duplicate traces of the same session: 1iv18i4u, ilpss66i, ohyjfuxk, t5vbcz5h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2086.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 2,086 57.6 85.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rqvfl1sy",
   "source_url": "https://omlx.ai/benchmarks/performance/rqvfl1sy",
   "tg_tokens": null,
   "tps": 57.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rqvfl1sy-ctx65536",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 88.2 GB at this context; duplicate traces of the same session: 1iv18i4u, ilpss66i, ohyjfuxk, t5vbcz5h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1845.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 1,845 57.9 88.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rqvfl1sy",
   "source_url": "https://omlx.ai/benchmarks/performance/rqvfl1sy",
   "tg_tokens": null,
   "tps": 57.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rqvfl1sy-ctx131072",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 90.3 GB at this context; duplicate traces of the same session: 1iv18i4u, ilpss66i, ohyjfuxk, t5vbcz5h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1882.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 1,882 57.5 90.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rqvfl1sy",
   "source_url": "https://omlx.ai/benchmarks/performance/rqvfl1sy",
   "tg_tokens": null,
   "tps": 57.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2wwtzagk-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: da5d458m, sslj94t3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1283.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,283 83.7 70.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2wwtzagk",
   "source_url": "https://omlx.ai/benchmarks/performance/2wwtzagk",
   "tg_tokens": null,
   "tps": 83.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2wwtzagk-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.9 GB at this context; duplicate traces of the same session: da5d458m, sslj94t3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2742.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,742 87.7 72.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2wwtzagk",
   "source_url": "https://omlx.ai/benchmarks/performance/2wwtzagk",
   "tg_tokens": null,
   "tps": 87.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2wwtzagk-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 76.8 GB at this context; duplicate traces of the same session: da5d458m, sslj94t3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2401.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,401 69.9 76.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2wwtzagk",
   "source_url": "https://omlx.ai/benchmarks/performance/2wwtzagk",
   "tg_tokens": null,
   "tps": 69.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-iq22zz9h-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.4 GB at this context; duplicate traces of the same session: g3honfqo, vmgx2uar; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1190.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,190 27.0 71.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iq22zz9h",
   "source_url": "https://omlx.ai/benchmarks/performance/iq22zz9h",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-iq22zz9h-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.9 GB at this context; duplicate traces of the same session: g3honfqo, vmgx2uar; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2080.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,080 74.5 72.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iq22zz9h",
   "source_url": "https://omlx.ai/benchmarks/performance/iq22zz9h",
   "tg_tokens": null,
   "tps": 74.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-iq22zz9h-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 76.8 GB at this context; duplicate traces of the same session: g3honfqo, vmgx2uar; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2274.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,274 60.3 76.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iq22zz9h",
   "source_url": "https://omlx.ai/benchmarks/performance/iq22zz9h",
   "tg_tokens": null,
   "tps": 60.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-p9e5iftz-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: i5dhbuaz, m89jgye4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2014.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,014 77.6 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p9e5iftz",
   "source_url": "https://omlx.ai/benchmarks/performance/p9e5iftz",
   "tg_tokens": null,
   "tps": 77.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-p9e5iftz-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: i5dhbuaz, m89jgye4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2506.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,506 80.5 72.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p9e5iftz",
   "source_url": "https://omlx.ai/benchmarks/performance/p9e5iftz",
   "tg_tokens": null,
   "tps": 80.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-p9e5iftz-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.2 GB at this context; duplicate traces of the same session: i5dhbuaz, m89jgye4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2282.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,282 74.8 78.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p9e5iftz",
   "source_url": "https://omlx.ai/benchmarks/performance/p9e5iftz",
   "tg_tokens": null,
   "tps": 74.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-l61cc5l5-ctx1024",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace, single run at 1k context; peak memory 177.6 GB; TTFT 1999 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 512.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k tokens 512.3 PP tok/s 47.4 TG tok/s 1999 TTFT (ms) 177.6 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l61cc5l5",
   "source_url": "https://omlx.ai/benchmarks/performance/l61cc5l5",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-fyro9ywb-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: uqj39a31; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 189.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 16 GB \u00b7 4bit; 1k 189.4 34.5 5.9 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fyro9ywb",
   "source_url": "https://omlx.ai/benchmarks/performance/fyro9ywb",
   "tg_tokens": null,
   "tps": 34.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-fyro9ywb-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: uqj39a31; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 188.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 16 GB \u00b7 4bit; 4k 188.0 33.9 6.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fyro9ywb",
   "source_url": "https://omlx.ai/benchmarks/performance/fyro9ywb",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-c0okhbj6-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: 2rn0lf93, dwryuzua, xusqahm5, znfmgwk5; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 578.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 1k 578.9 68.4 37.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c0okhbj6",
   "source_url": "https://omlx.ai/benchmarks/performance/c0okhbj6",
   "tg_tokens": null,
   "tps": 68.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-c0okhbj6-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.1 GB at this context; duplicate traces of the same session: 2rn0lf93, dwryuzua, xusqahm5, znfmgwk5; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 645.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 4k 645.5 63.5 39.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c0okhbj6",
   "source_url": "https://omlx.ai/benchmarks/performance/c0okhbj6",
   "tg_tokens": null,
   "tps": 63.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-c0okhbj6-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.4 GB at this context; duplicate traces of the same session: 2rn0lf93, dwryuzua, xusqahm5, znfmgwk5; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 556.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 16k 556.7 55.5 39.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c0okhbj6",
   "source_url": "https://omlx.ai/benchmarks/performance/c0okhbj6",
   "tg_tokens": null,
   "tps": 55.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-c0okhbj6-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.8 GB at this context; duplicate traces of the same session: 2rn0lf93, dwryuzua, xusqahm5, znfmgwk5; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 467.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 32k 467.9 51.6 39.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c0okhbj6",
   "source_url": "https://omlx.ai/benchmarks/performance/c0okhbj6",
   "tg_tokens": null,
   "tps": 51.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-c0okhbj6-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 42.0 GB at this context; duplicate traces of the same session: 2rn0lf93, dwryuzua, xusqahm5, znfmgwk5; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 237.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 128k 237.3 24.7 42.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c0okhbj6",
   "source_url": "https://omlx.ai/benchmarks/performance/c0okhbj6",
   "tg_tokens": null,
   "tps": 24.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-92rjn2pe-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: m5g2yl4j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 262.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 262.0 29.7 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/92rjn2pe",
   "source_url": "https://omlx.ai/benchmarks/performance/92rjn2pe",
   "tg_tokens": null,
   "tps": 29.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-92rjn2pe-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: m5g2yl4j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 274.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 274.3 28.8 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/92rjn2pe",
   "source_url": "https://omlx.ai/benchmarks/performance/92rjn2pe",
   "tg_tokens": null,
   "tps": 28.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-0yo5jzqb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: tlgm9wz2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 235.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 235.5 34.2 18.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0yo5jzqb",
   "source_url": "https://omlx.ai/benchmarks/performance/0yo5jzqb",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-0yo5jzqb-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: tlgm9wz2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 252.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 252.3 51.7 21.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0yo5jzqb",
   "source_url": "https://omlx.ai/benchmarks/performance/0yo5jzqb",
   "tg_tokens": null,
   "tps": 51.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kz1atl8o-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: 0odolcoy, 6wzqfifs, c5mavv5c, qh94of5b, v5n8m2xj, vcixfri2, z95uxaiq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 903.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 903.8 69.8 17.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kz1atl8o",
   "source_url": "https://omlx.ai/benchmarks/performance/kz1atl8o",
   "tg_tokens": null,
   "tps": 69.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kz1atl8o-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 0odolcoy, 6wzqfifs, c5mavv5c, qh94of5b, v5n8m2xj, vcixfri2, z95uxaiq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 991.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 991.7 60.2 18.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kz1atl8o",
   "source_url": "https://omlx.ai/benchmarks/performance/kz1atl8o",
   "tg_tokens": null,
   "tps": 60.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kz1atl8o-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 0odolcoy, 6wzqfifs, c5mavv5c, qh94of5b, v5n8m2xj, vcixfri2, z95uxaiq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 910.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 910.4 65.3 19.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kz1atl8o",
   "source_url": "https://omlx.ai/benchmarks/performance/kz1atl8o",
   "tg_tokens": null,
   "tps": 65.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kz1atl8o-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 0odolcoy, 6wzqfifs, c5mavv5c, qh94of5b, v5n8m2xj, vcixfri2, z95uxaiq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 874.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 874.6 82.7 19.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kz1atl8o",
   "source_url": "https://omlx.ai/benchmarks/performance/kz1atl8o",
   "tg_tokens": null,
   "tps": 82.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kz1atl8o-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 0odolcoy, 6wzqfifs, c5mavv5c, qh94of5b, v5n8m2xj, vcixfri2, z95uxaiq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 811.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 811.3 81.9 20.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kz1atl8o",
   "source_url": "https://omlx.ai/benchmarks/performance/kz1atl8o",
   "tg_tokens": null,
   "tps": 81.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-kz1atl8o-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: 0odolcoy, 6wzqfifs, c5mavv5c, qh94of5b, v5n8m2xj, vcixfri2, z95uxaiq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 715.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 715.0 57.4 23.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kz1atl8o",
   "source_url": "https://omlx.ai/benchmarks/performance/kz1atl8o",
   "tg_tokens": null,
   "tps": 57.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-kz1atl8o-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: 0odolcoy, 6wzqfifs, c5mavv5c, qh94of5b, v5n8m2xj, vcixfri2, z95uxaiq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 572.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 572.4 51.7 27.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kz1atl8o",
   "source_url": "https://omlx.ai/benchmarks/performance/kz1atl8o",
   "tg_tokens": null,
   "tps": 51.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-kz1atl8o-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.1 GB at this context; duplicate traces of the same session: 0odolcoy, 6wzqfifs, c5mavv5c, qh94of5b, v5n8m2xj, vcixfri2, z95uxaiq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 449.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 449.3 34.9 33.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kz1atl8o",
   "source_url": "https://omlx.ai/benchmarks/performance/kz1atl8o",
   "tg_tokens": null,
   "tps": 34.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6682eign-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: kx18eidb, m3wi47u8, tfb2ien8, u0nnwlv9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 692.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 692.8 45.1 17.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6682eign",
   "source_url": "https://omlx.ai/benchmarks/performance/6682eign",
   "tg_tokens": null,
   "tps": 45.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6682eign-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: kx18eidb, m3wi47u8, tfb2ien8, u0nnwlv9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 929.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 929.3 62.4 18.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6682eign",
   "source_url": "https://omlx.ai/benchmarks/performance/6682eign",
   "tg_tokens": null,
   "tps": 62.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6682eign-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: kx18eidb, m3wi47u8, tfb2ien8, u0nnwlv9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 968.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 968.2 58.6 19.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6682eign",
   "source_url": "https://omlx.ai/benchmarks/performance/6682eign",
   "tg_tokens": null,
   "tps": 58.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6682eign-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: kx18eidb, m3wi47u8, tfb2ien8, u0nnwlv9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 950.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 950.3 74.8 19.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6682eign",
   "source_url": "https://omlx.ai/benchmarks/performance/6682eign",
   "tg_tokens": null,
   "tps": 74.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-6682eign-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: kx18eidb, m3wi47u8, tfb2ien8, u0nnwlv9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 594.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 594.3 48.1 27.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6682eign",
   "source_url": "https://omlx.ai/benchmarks/performance/6682eign",
   "tg_tokens": null,
   "tps": 48.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-o1hfrqgl-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 22.9 GB; TTFT 178759 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 366.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k tokens 366.6 PP tok/s 31.0 TG tok/s 178759 TTFT (ms) 22.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o1hfrqgl",
   "source_url": "https://omlx.ai/benchmarks/performance/o1hfrqgl",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-yjpg78pj-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 22.9 GB; TTFT 176917 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 370.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k tokens 370.4 PP tok/s 27.4 TG tok/s 176917 TTFT (ms) 22.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yjpg78pj",
   "source_url": "https://omlx.ai/benchmarks/performance/yjpg78pj",
   "tg_tokens": null,
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-pky187bk-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 20.6 GB; TTFT 77961 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 420.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k tokens 420.3 PP tok/s 47.4 TG tok/s 77961 TTFT (ms) 20.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pky187bk",
   "source_url": "https://omlx.ai/benchmarks/performance/pky187bk",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-srpgz9v2-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: 1iwnl5sg; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 145.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 1k 145.8 18.7 34.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/srpgz9v2",
   "source_url": "https://omlx.ai/benchmarks/performance/srpgz9v2",
   "tg_tokens": null,
   "tps": 18.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-srpgz9v2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 38.7 GB at this context; duplicate traces of the same session: 1iwnl5sg; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 146.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k 146.1 17.4 38.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/srpgz9v2",
   "source_url": "https://omlx.ai/benchmarks/performance/srpgz9v2",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-tkxjq60t-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 107.8 GB at this context; duplicate traces of the same session: l0wdpybf; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 677.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 677.4 53.3 107.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tkxjq60t",
   "source_url": "https://omlx.ai/benchmarks/performance/tkxjq60t",
   "tg_tokens": null,
   "tps": 53.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-tkxjq60t-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 108.3 GB at this context; duplicate traces of the same session: l0wdpybf; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1041.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 1,041 52.2 108.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tkxjq60t",
   "source_url": "https://omlx.ai/benchmarks/performance/tkxjq60t",
   "tg_tokens": null,
   "tps": 52.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-pnybk6lg-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 107.8 GB; TTFT 1209 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 847.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k tokens 847.2 PP tok/s 22.8 TG tok/s 1209 TTFT (ms) 107.8 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pnybk6lg",
   "source_url": "https://omlx.ai/benchmarks/performance/pnybk6lg",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-mk0vaxo7-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 107.7 GB; TTFT 1211 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 845.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k tokens 845.4 PP tok/s 25.5 TG tok/s 1211 TTFT (ms) 107.7 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mk0vaxo7",
   "source_url": "https://omlx.ai/benchmarks/performance/mk0vaxo7",
   "tg_tokens": null,
   "tps": 25.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-avufvrnx-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 108.6 GB; TTFT 1253 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 817.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k tokens 817.6 PP tok/s 22.7 TG tok/s 1253 TTFT (ms) 108.6 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/avufvrnx",
   "source_url": "https://omlx.ai/benchmarks/performance/avufvrnx",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-99x37gj3-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1254.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,254 91.3 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/99x37gj3",
   "source_url": "https://omlx.ai/benchmarks/performance/99x37gj3",
   "tg_tokens": null,
   "tps": 91.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-99x37gj3-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2072.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,072 84.3 71.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/99x37gj3",
   "source_url": "https://omlx.ai/benchmarks/performance/99x37gj3",
   "tg_tokens": null,
   "tps": 84.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-99x37gj3-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2739.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,739 76.4 72.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/99x37gj3",
   "source_url": "https://omlx.ai/benchmarks/performance/99x37gj3",
   "tg_tokens": null,
   "tps": 76.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-99x37gj3-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 76.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2406.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,406 63.3 76.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/99x37gj3",
   "source_url": "https://omlx.ai/benchmarks/performance/99x37gj3",
   "tg_tokens": null,
   "tps": 63.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kjlyy0x4-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP requant build; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 0hgg7ex3, 2hj7m1u4, 5nl1xsjl, gdui4gsq, lge7osc5, z93yb1ir; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1273.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp_requant on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,273 86.3 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kjlyy0x4",
   "source_url": "https://omlx.ai/benchmarks/performance/kjlyy0x4",
   "tg_tokens": null,
   "tps": 86.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-kjlyy0x4-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP requant build; oMLX community benchmark trace; peak memory 72.0 GB at this context; duplicate traces of the same session: 0hgg7ex3, 2hj7m1u4, 5nl1xsjl, gdui4gsq, lge7osc5, z93yb1ir; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2778.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp_requant on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,778 84.4 72.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kjlyy0x4",
   "source_url": "https://omlx.ai/benchmarks/performance/kjlyy0x4",
   "tg_tokens": null,
   "tps": 84.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kjlyy0x4-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP requant build; oMLX community benchmark trace; peak memory 73.0 GB at this context; duplicate traces of the same session: 0hgg7ex3, 2hj7m1u4, 5nl1xsjl, gdui4gsq, lge7osc5, z93yb1ir; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2686.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp_requant on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,686 90.0 73.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kjlyy0x4",
   "source_url": "https://omlx.ai/benchmarks/performance/kjlyy0x4",
   "tg_tokens": null,
   "tps": 90.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kjlyy0x4-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP requant build; oMLX community benchmark trace; peak memory 73.6 GB at this context; duplicate traces of the same session: 0hgg7ex3, 2hj7m1u4, 5nl1xsjl, gdui4gsq, lge7osc5, z93yb1ir; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2583.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp_requant on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,583 75.8 73.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kjlyy0x4",
   "source_url": "https://omlx.ai/benchmarks/performance/kjlyy0x4",
   "tg_tokens": null,
   "tps": 75.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kjlyy0x4-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP requant build; oMLX community benchmark trace; peak memory 74.7 GB at this context; duplicate traces of the same session: 0hgg7ex3, 2hj7m1u4, 5nl1xsjl, gdui4gsq, lge7osc5, z93yb1ir; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2413.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp_requant on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,413 81.0 74.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kjlyy0x4",
   "source_url": "https://omlx.ai/benchmarks/performance/kjlyy0x4",
   "tg_tokens": null,
   "tps": 81.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-kjlyy0x4-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP requant build; oMLX community benchmark trace; peak memory 77.0 GB at this context; duplicate traces of the same session: 0hgg7ex3, 2hj7m1u4, 5nl1xsjl, gdui4gsq, lge7osc5, z93yb1ir; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2402.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp_requant on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,402 75.7 77.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kjlyy0x4",
   "source_url": "https://omlx.ai/benchmarks/performance/kjlyy0x4",
   "tg_tokens": null,
   "tps": 75.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kjlyy0x4-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP requant build; oMLX community benchmark trace; peak memory 79.8 GB at this context; duplicate traces of the same session: 0hgg7ex3, 2hj7m1u4, 5nl1xsjl, gdui4gsq, lge7osc5, z93yb1ir; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2354.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp_requant on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,354 57.4 79.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kjlyy0x4",
   "source_url": "https://omlx.ai/benchmarks/performance/kjlyy0x4",
   "tg_tokens": null,
   "tps": 57.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-i5dalzn6-ctx1024",
   "model": "Swift 1.5 Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 142.6 GB at this context; duplicate traces of the same session: zm1hfjh1; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2199.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 1k 2,199 130.5 142.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i5dalzn6",
   "source_url": "https://omlx.ai/benchmarks/performance/i5dalzn6",
   "tg_tokens": null,
   "tps": 130.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-i5dalzn6-ctx65536",
   "model": "Swift 1.5 Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 147.0 GB at this context; duplicate traces of the same session: zm1hfjh1; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4111.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 64k 4,111 121.7 147.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i5dalzn6",
   "source_url": "https://omlx.ai/benchmarks/performance/i5dalzn6",
   "tg_tokens": null,
   "tps": 121.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9bkt3x2t-ctx1024",
   "model": "CyberTiel-Coder-35B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: oai22vze, rgxj9350, s8oa2dsu, zlvbisrb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2055.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "CyberTiel-Coder-35B-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 1k 2,055 147.0 30.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9bkt3x2t",
   "source_url": "https://omlx.ai/benchmarks/performance/9bkt3x2t",
   "tg_tokens": null,
   "tps": 147.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9bkt3x2t-ctx4096",
   "model": "CyberTiel-Coder-35B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: oai22vze, rgxj9350, s8oa2dsu, zlvbisrb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2516.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "CyberTiel-Coder-35B-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 4k 2,516 151.1 32.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9bkt3x2t",
   "source_url": "https://omlx.ai/benchmarks/performance/9bkt3x2t",
   "tg_tokens": null,
   "tps": 151.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9bkt3x2t-ctx16384",
   "model": "CyberTiel-Coder-35B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 32.5 GB at this context; duplicate traces of the same session: oai22vze, rgxj9350, s8oa2dsu, zlvbisrb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2149.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "CyberTiel-Coder-35B-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 16k 2,149 154.8 32.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9bkt3x2t",
   "source_url": "https://omlx.ai/benchmarks/performance/9bkt3x2t",
   "tg_tokens": null,
   "tps": 154.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9bkt3x2t-ctx32768",
   "model": "CyberTiel-Coder-35B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 33.2 GB at this context; duplicate traces of the same session: oai22vze, rgxj9350, s8oa2dsu, zlvbisrb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1760.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "CyberTiel-Coder-35B-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 32k 1,760 191.7 33.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9bkt3x2t",
   "source_url": "https://omlx.ai/benchmarks/performance/9bkt3x2t",
   "tg_tokens": null,
   "tps": 191.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9bkt3x2t-ctx65536",
   "model": "CyberTiel-Coder-35B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 34.5 GB at this context; duplicate traces of the same session: oai22vze, rgxj9350, s8oa2dsu, zlvbisrb; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1291.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "CyberTiel-Coder-35B-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 64k 1,291 111.2 34.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9bkt3x2t",
   "source_url": "https://omlx.ai/benchmarks/performance/9bkt3x2t",
   "tg_tokens": null,
   "tps": 111.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9e1bqka8-ctx1024",
   "model": "CyberTiel-Coder-35B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1995.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "CyberTiel-Coder-35B-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 1k 1,995 106.6 28.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9e1bqka8",
   "source_url": "https://omlx.ai/benchmarks/performance/9e1bqka8",
   "tg_tokens": null,
   "tps": 106.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9e1bqka8-ctx4096",
   "model": "CyberTiel-Coder-35B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 30.4 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2519.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "CyberTiel-Coder-35B-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 4k 2,519 106.1 30.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9e1bqka8",
   "source_url": "https://omlx.ai/benchmarks/performance/9e1bqka8",
   "tg_tokens": null,
   "tps": 106.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9e1bqka8-ctx16384",
   "model": "CyberTiel-Coder-35B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 30.8 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2240.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "CyberTiel-Coder-35B-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 16k 2,240 112.7 30.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9e1bqka8",
   "source_url": "https://omlx.ai/benchmarks/performance/9e1bqka8",
   "tg_tokens": null,
   "tps": 112.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9e1bqka8-ctx32768",
   "model": "CyberTiel-Coder-35B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 31.4 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1852.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "CyberTiel-Coder-35B-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 32k 1,852 89.3 31.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9e1bqka8",
   "source_url": "https://omlx.ai/benchmarks/performance/9e1bqka8",
   "tg_tokens": null,
   "tps": 89.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9e1bqka8-ctx65536",
   "model": "CyberTiel-Coder-35B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 32.7 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1380.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "CyberTiel-Coder-35B-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 64k 1,380 64.6 32.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9e1bqka8",
   "source_url": "https://omlx.ai/benchmarks/performance/9e1bqka8",
   "tg_tokens": null,
   "tps": 64.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-p928g824-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: fp6aeet4; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 530.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 530.2 49.1 177.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p928g824",
   "source_url": "https://omlx.ai/benchmarks/performance/p928g824",
   "tg_tokens": null,
   "tps": 49.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-p928g824-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: fp6aeet4; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 909.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 909.4 59.4 178.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p928g824",
   "source_url": "https://omlx.ai/benchmarks/performance/p928g824",
   "tg_tokens": null,
   "tps": 59.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5dexaw9c-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Huihui abliterated build; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: bsr2r9kr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1604.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.8-Flash-Next-abliterated-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,604 75.8 71.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5dexaw9c",
   "source_url": "https://omlx.ai/benchmarks/performance/5dexaw9c",
   "tg_tokens": null,
   "tps": 75.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5dexaw9c-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Huihui abliterated build; oMLX community benchmark trace; peak memory 73.3 GB at this context; duplicate traces of the same session: bsr2r9kr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2346.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.8-Flash-Next-abliterated-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,346 69.4 73.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5dexaw9c",
   "source_url": "https://omlx.ai/benchmarks/performance/5dexaw9c",
   "tg_tokens": null,
   "tps": 69.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ukkr7udn-ctx1024",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: zj6g3wd3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 470.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 470.9 17.6 15.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ukkr7udn",
   "source_url": "https://omlx.ai/benchmarks/performance/ukkr7udn",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ukkr7udn-ctx4096",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: zj6g3wd3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 500.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 500.4 17.8 16.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ukkr7udn",
   "source_url": "https://omlx.ai/benchmarks/performance/ukkr7udn",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-1tikgf7y-ctx8192",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace, single run at 8k context; peak memory 16.2 GB; TTFT 16549 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 495.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k tokens 495.0 PP tok/s 17.3 TG tok/s 16549 TTFT (ms) 16.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1tikgf7y",
   "source_url": "https://omlx.ai/benchmarks/performance/1tikgf7y",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-6hshcent-ctx4096",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: r8m4ia0z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1535.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,535 148.1 19.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6hshcent",
   "source_url": "https://omlx.ai/benchmarks/performance/6hshcent",
   "tg_tokens": null,
   "tps": 148.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-6hshcent-ctx16384",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: r8m4ia0z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1486.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,486 136.6 19.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6hshcent",
   "source_url": "https://omlx.ai/benchmarks/performance/6hshcent",
   "tg_tokens": null,
   "tps": 136.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-qgmatb9j-ctx16384",
   "model": "Qwen3.6-35B-A3B-MTPLX-Optimized-Balance",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 31.2 GB; TTFT 14092 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1163.0,
   "provenance": "community",
   "quant": "FP16",
   "quote": "Qwen3.6-35B-A3B-MTPLX-Optimized-Balance-FP16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 6bit; 16k tokens 1,163 PP tok/s 62.1 TG tok/s 14092 TTFT (ms) 31.2 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qgmatb9j",
   "source_url": "https://omlx.ai/benchmarks/performance/qgmatb9j",
   "tg_tokens": null,
   "tps": 62.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-6ux5ups6-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: ikbjnjoy, jco8l6eq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1543.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,543 93.3 21.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ux5ups6",
   "source_url": "https://omlx.ai/benchmarks/performance/6ux5ups6",
   "tg_tokens": null,
   "tps": 93.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-6ux5ups6-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: ikbjnjoy, jco8l6eq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3332.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 3,332 164.6 23.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ux5ups6",
   "source_url": "https://omlx.ai/benchmarks/performance/6ux5ups6",
   "tg_tokens": null,
   "tps": 164.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-6ux5ups6-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: ikbjnjoy, jco8l6eq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3213.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 3,213 101.4 23.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ux5ups6",
   "source_url": "https://omlx.ai/benchmarks/performance/6ux5ups6",
   "tg_tokens": null,
   "tps": 101.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-qeuuv28n-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 6z290f7g, bc6wfver, zxrfdyz9; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 606.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 606.0 62.4 21.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qeuuv28n",
   "source_url": "https://omlx.ai/benchmarks/performance/qeuuv28n",
   "tg_tokens": null,
   "tps": 62.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-qeuuv28n-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 6z290f7g, bc6wfver, zxrfdyz9; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 817.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 817.9 72.5 22.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qeuuv28n",
   "source_url": "https://omlx.ai/benchmarks/performance/qeuuv28n",
   "tg_tokens": null,
   "tps": 72.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-qeuuv28n-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: 6z290f7g, bc6wfver, zxrfdyz9; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 778.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 778.8 99.3 22.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qeuuv28n",
   "source_url": "https://omlx.ai/benchmarks/performance/qeuuv28n",
   "tg_tokens": null,
   "tps": 99.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-qeuuv28n-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: 6z290f7g, bc6wfver, zxrfdyz9; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 696.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 696.7 57.4 22.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qeuuv28n",
   "source_url": "https://omlx.ai/benchmarks/performance/qeuuv28n",
   "tg_tokens": null,
   "tps": 57.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-j15nfvft-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace, single run at 16k context; peak memory 40.2 GB; TTFT 13862 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1182.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.6-35B-A3B-oQ8-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 16k tokens 1,182 PP tok/s 69.2 TG tok/s 13862 TTFT (ms) 40.2 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j15nfvft",
   "source_url": "https://omlx.ai/benchmarks/performance/j15nfvft",
   "tg_tokens": null,
   "tps": 69.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ao04paj4-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTPLX Optimized Quality FP16 build; oMLX community benchmark trace, single run at 16k context; peak memory 45.0 GB; TTFT 72433 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 226.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality-FP16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 16k tokens 226.2 PP tok/s 22.6 TG tok/s 72433 TTFT (ms) 45.0 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ao04paj4",
   "source_url": "https://omlx.ai/benchmarks/performance/ao04paj4",
   "tg_tokens": null,
   "tps": 22.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-gxgtn87i-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "mlx fp16 MTP build; oMLX community benchmark trace, single run at 16k context; peak memory 45.9 GB; TTFT 72847 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.9,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-27B-mlx-oQ8-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 16k tokens 224.9 PP tok/s 15.3 TG tok/s 72847 TTFT (ms) 45.9 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gxgtn87i",
   "source_url": "https://omlx.ai/benchmarks/performance/gxgtn87i",
   "tg_tokens": null,
   "tps": 15.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2v7fn7gm-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 3d61uari, 4mc7y8ls, borfmt8g, sbxkbfil, yq4m2n0a, zel431lt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 791.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 791.2 47.2 17.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2v7fn7gm",
   "source_url": "https://omlx.ai/benchmarks/performance/2v7fn7gm",
   "tg_tokens": null,
   "tps": 47.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2v7fn7gm-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 3d61uari, 4mc7y8ls, borfmt8g, sbxkbfil, yq4m2n0a, zel431lt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1271.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,271 62.3 19.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2v7fn7gm",
   "source_url": "https://omlx.ai/benchmarks/performance/2v7fn7gm",
   "tg_tokens": null,
   "tps": 62.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2v7fn7gm-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 3d61uari, 4mc7y8ls, borfmt8g, sbxkbfil, yq4m2n0a, zel431lt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1312.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,312 57.0 19.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2v7fn7gm",
   "source_url": "https://omlx.ai/benchmarks/performance/2v7fn7gm",
   "tg_tokens": null,
   "tps": 57.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2v7fn7gm-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 3d61uari, 4mc7y8ls, borfmt8g, sbxkbfil, yq4m2n0a, zel431lt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1262.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,262 76.6 19.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2v7fn7gm",
   "source_url": "https://omlx.ai/benchmarks/performance/2v7fn7gm",
   "tg_tokens": null,
   "tps": 76.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2v7fn7gm-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: 3d61uari, 4mc7y8ls, borfmt8g, sbxkbfil, yq4m2n0a, zel431lt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1086.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,086 133.0 21.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2v7fn7gm",
   "source_url": "https://omlx.ai/benchmarks/performance/2v7fn7gm",
   "tg_tokens": null,
   "tps": 133.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2v7fn7gm-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 3d61uari, 4mc7y8ls, borfmt8g, sbxkbfil, yq4m2n0a, zel431lt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 903.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 903.0 55.9 23.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2v7fn7gm",
   "source_url": "https://omlx.ai/benchmarks/performance/2v7fn7gm",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2v7fn7gm-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: 3d61uari, 4mc7y8ls, borfmt8g, sbxkbfil, yq4m2n0a, zel431lt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 598.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 598.0 47.7 28.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2v7fn7gm",
   "source_url": "https://omlx.ai/benchmarks/performance/2v7fn7gm",
   "tg_tokens": null,
   "tps": 47.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-844ccmby-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 7yh1cths; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 933.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 933.4 70.9 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/844ccmby",
   "source_url": "https://omlx.ai/benchmarks/performance/844ccmby",
   "tg_tokens": null,
   "tps": 70.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-844ccmby-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 7yh1cths; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1315.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,315 53.8 19.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/844ccmby",
   "source_url": "https://omlx.ai/benchmarks/performance/844ccmby",
   "tg_tokens": null,
   "tps": 53.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-t8osg78b-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.6 GB at this context; duplicate traces of the same session: cnw51lwj, cv19sxwk, ix3xqfoe, p5l20020; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 155.8,
   "provenance": "community",
   "quant": "oQ6e-fp16",
   "quote": "Qwen3.8-27B-oQ6e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 4k 155.8 13.5 31.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t8osg78b",
   "source_url": "https://omlx.ai/benchmarks/performance/t8osg78b",
   "tg_tokens": null,
   "tps": 13.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-t8osg78b-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.5 GB at this context; duplicate traces of the same session: cnw51lwj, cv19sxwk, ix3xqfoe, p5l20020; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 152.4,
   "provenance": "community",
   "quant": "oQ6e-fp16",
   "quote": "Qwen3.8-27B-oQ6e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 8k 152.4 13.4 32.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t8osg78b",
   "source_url": "https://omlx.ai/benchmarks/performance/t8osg78b",
   "tg_tokens": null,
   "tps": 13.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-t8osg78b-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.5 GB at this context; duplicate traces of the same session: cnw51lwj, cv19sxwk, ix3xqfoe, p5l20020; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 145.3,
   "provenance": "community",
   "quant": "oQ6e-fp16",
   "quote": "Qwen3.8-27B-oQ6e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 16k 145.3 15.6 33.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t8osg78b",
   "source_url": "https://omlx.ai/benchmarks/performance/t8osg78b",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-t8osg78b-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.6 GB at this context; duplicate traces of the same session: cnw51lwj, cv19sxwk, ix3xqfoe, p5l20020; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 133.2,
   "provenance": "community",
   "quant": "oQ6e-fp16",
   "quote": "Qwen3.8-27B-oQ6e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 32k 133.2 15.4 34.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t8osg78b",
   "source_url": "https://omlx.ai/benchmarks/performance/t8osg78b",
   "tg_tokens": null,
   "tps": 15.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-t8osg78b-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.8 GB at this context; duplicate traces of the same session: cnw51lwj, cv19sxwk, ix3xqfoe, p5l20020; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 114.6,
   "provenance": "community",
   "quant": "oQ6e-fp16",
   "quote": "Qwen3.8-27B-oQ6e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 64k 114.6 13.2 36.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t8osg78b",
   "source_url": "https://omlx.ai/benchmarks/performance/t8osg78b",
   "tg_tokens": null,
   "tps": 13.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-ijlb6xye-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: 6zsgfs1s; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 196.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 196.4 24.1 34.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ijlb6xye",
   "source_url": "https://omlx.ai/benchmarks/performance/ijlb6xye",
   "tg_tokens": null,
   "tps": 24.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ijlb6xye-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 38.5 GB at this context; duplicate traces of the same session: 6zsgfs1s; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 201.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 201.3 24.0 38.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ijlb6xye",
   "source_url": "https://omlx.ai/benchmarks/performance/ijlb6xye",
   "tg_tokens": null,
   "tps": 24.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ozucrd05-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: lo9cor73; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1123.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 1k 1,123 57.3 33.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ozucrd05",
   "source_url": "https://omlx.ai/benchmarks/performance/ozucrd05",
   "tg_tokens": null,
   "tps": 57.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ozucrd05-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: lo9cor73; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1359.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 4k 1,359 59.8 34.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ozucrd05",
   "source_url": "https://omlx.ai/benchmarks/performance/ozucrd05",
   "tg_tokens": null,
   "tps": 59.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ygriyv77-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: fcq4dc1h, gmakpv9c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1254.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,254 91.3 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ygriyv77",
   "source_url": "https://omlx.ai/benchmarks/performance/ygriyv77",
   "tg_tokens": null,
   "tps": 91.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ygriyv77-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: fcq4dc1h, gmakpv9c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2072.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,072 84.3 71.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ygriyv77",
   "source_url": "https://omlx.ai/benchmarks/performance/ygriyv77",
   "tg_tokens": null,
   "tps": 84.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ygriyv77-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.9 GB at this context; duplicate traces of the same session: fcq4dc1h, gmakpv9c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2739.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,739 76.4 72.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ygriyv77",
   "source_url": "https://omlx.ai/benchmarks/performance/ygriyv77",
   "tg_tokens": null,
   "tps": 76.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ygriyv77-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 76.8 GB at this context; duplicate traces of the same session: fcq4dc1h, gmakpv9c; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2406.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,406 63.3 76.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ygriyv77",
   "source_url": "https://omlx.ai/benchmarks/performance/ygriyv77",
   "tg_tokens": null,
   "tps": 63.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-81qtm2xr-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: 2berb3rb, api215uw, hwjfdrql, pk9so605, u7i14ofo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 438.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 438.6 22.7 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/81qtm2xr",
   "source_url": "https://omlx.ai/benchmarks/performance/81qtm2xr",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-81qtm2xr-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: 2berb3rb, api215uw, hwjfdrql, pk9so605, u7i14ofo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1324.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,324 35.2 71.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/81qtm2xr",
   "source_url": "https://omlx.ai/benchmarks/performance/81qtm2xr",
   "tg_tokens": null,
   "tps": 35.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-81qtm2xr-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.0 GB at this context; duplicate traces of the same session: 2berb3rb, api215uw, hwjfdrql, pk9so605, u7i14ofo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2614.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,614 51.9 73.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/81qtm2xr",
   "source_url": "https://omlx.ai/benchmarks/performance/81qtm2xr",
   "tg_tokens": null,
   "tps": 51.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-81qtm2xr-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.7 GB at this context; duplicate traces of the same session: 2berb3rb, api215uw, hwjfdrql, pk9so605, u7i14ofo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2483.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,483 76.1 74.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/81qtm2xr",
   "source_url": "https://omlx.ai/benchmarks/performance/81qtm2xr",
   "tg_tokens": null,
   "tps": 76.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-81qtm2xr-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 77.0 GB at this context; duplicate traces of the same session: 2berb3rb, api215uw, hwjfdrql, pk9so605, u7i14ofo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2377.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,377 62.2 77.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/81qtm2xr",
   "source_url": "https://omlx.ai/benchmarks/performance/81qtm2xr",
   "tg_tokens": null,
   "tps": 62.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-81qtm2xr-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.8 GB at this context; duplicate traces of the same session: 2berb3rb, api215uw, hwjfdrql, pk9so605, u7i14ofo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2074.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,074 59.2 79.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/81qtm2xr",
   "source_url": "https://omlx.ai/benchmarks/performance/81qtm2xr",
   "tg_tokens": null,
   "tps": 59.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wn2xankp-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 71.1 GB; TTFT 3235 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1266.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 1,266 PP tok/s 62.1 TG tok/s 3235 TTFT (ms) 71.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wn2xankp",
   "source_url": "https://omlx.ai/benchmarks/performance/wn2xankp",
   "tg_tokens": null,
   "tps": 62.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-8fy1vxn0-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 129.4 GB at this context; duplicate traces of the same session: q0j6ctmj; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 778.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 1k 778.2 68.9 129.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8fy1vxn0",
   "source_url": "https://omlx.ai/benchmarks/performance/8fy1vxn0",
   "tg_tokens": null,
   "tps": 68.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-8fy1vxn0-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 129.7 GB at this context; duplicate traces of the same session: q0j6ctmj; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1826.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 4k 1,826 58.6 129.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8fy1vxn0",
   "source_url": "https://omlx.ai/benchmarks/performance/8fy1vxn0",
   "tg_tokens": null,
   "tps": 58.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-b90mzgcb-ctx1024",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 17.1 GB; TTFT 6058 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 169.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "gemma-4-31b-it-mxfp4 on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 169.0 PP tok/s 6.9 TG tok/s 6058 TTFT (ms) 17.1 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b90mzgcb",
   "source_url": "https://omlx.ai/benchmarks/performance/b90mzgcb",
   "tg_tokens": null,
   "tps": 6.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-21sp2i39-ctx1024",
   "model": "CyberTiel-Coder-35B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: 25hkpn5m, 94a3zg0b, 94s9mpzj; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1995.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "CyberTiel-Coder-35B-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 1k 1,995 106.6 28.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/21sp2i39",
   "source_url": "https://omlx.ai/benchmarks/performance/21sp2i39",
   "tg_tokens": null,
   "tps": 106.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-21sp2i39-ctx4096",
   "model": "CyberTiel-Coder-35B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 30.4 GB at this context; duplicate traces of the same session: 25hkpn5m, 94a3zg0b, 94s9mpzj; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2519.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "CyberTiel-Coder-35B-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 4k 2,519 106.1 30.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/21sp2i39",
   "source_url": "https://omlx.ai/benchmarks/performance/21sp2i39",
   "tg_tokens": null,
   "tps": 106.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-21sp2i39-ctx16384",
   "model": "CyberTiel-Coder-35B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 30.8 GB at this context; duplicate traces of the same session: 25hkpn5m, 94a3zg0b, 94s9mpzj; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2240.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "CyberTiel-Coder-35B-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 16k 2,240 112.7 30.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/21sp2i39",
   "source_url": "https://omlx.ai/benchmarks/performance/21sp2i39",
   "tg_tokens": null,
   "tps": 112.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-21sp2i39-ctx32768",
   "model": "CyberTiel-Coder-35B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 31.4 GB at this context; duplicate traces of the same session: 25hkpn5m, 94a3zg0b, 94s9mpzj; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1852.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "CyberTiel-Coder-35B-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 32k 1,852 89.3 31.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/21sp2i39",
   "source_url": "https://omlx.ai/benchmarks/performance/21sp2i39",
   "tg_tokens": null,
   "tps": 89.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-21sp2i39-ctx65536",
   "model": "CyberTiel-Coder-35B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 32.7 GB at this context; duplicate traces of the same session: 25hkpn5m, 94a3zg0b, 94s9mpzj; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1380.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "CyberTiel-Coder-35B-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 64k 1,380 64.6 32.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/21sp2i39",
   "source_url": "https://omlx.ai/benchmarks/performance/21sp2i39",
   "tg_tokens": null,
   "tps": 64.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-fbetefn9-ctx1024",
   "model": "CyberTiel-Coder-35B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: 33yxe3u2, 8j2o54mz, ldqw3rie, twsswocq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1883.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "CyberTiel-Coder-35B-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 1k 1,883 150.2 30.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fbetefn9",
   "source_url": "https://omlx.ai/benchmarks/performance/fbetefn9",
   "tg_tokens": null,
   "tps": 150.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-fbetefn9-ctx4096",
   "model": "CyberTiel-Coder-35B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: 33yxe3u2, 8j2o54mz, ldqw3rie, twsswocq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2558.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "CyberTiel-Coder-35B-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 4k 2,558 123.2 32.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fbetefn9",
   "source_url": "https://omlx.ai/benchmarks/performance/fbetefn9",
   "tg_tokens": null,
   "tps": 123.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-fbetefn9-ctx16384",
   "model": "CyberTiel-Coder-35B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 32.5 GB at this context; duplicate traces of the same session: 33yxe3u2, 8j2o54mz, ldqw3rie, twsswocq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2163.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "CyberTiel-Coder-35B-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 16k 2,163 201.1 32.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fbetefn9",
   "source_url": "https://omlx.ai/benchmarks/performance/fbetefn9",
   "tg_tokens": null,
   "tps": 201.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-fbetefn9-ctx32768",
   "model": "CyberTiel-Coder-35B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 33.2 GB at this context; duplicate traces of the same session: 33yxe3u2, 8j2o54mz, ldqw3rie, twsswocq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1766.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "CyberTiel-Coder-35B-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 32k 1,766 161.3 33.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fbetefn9",
   "source_url": "https://omlx.ai/benchmarks/performance/fbetefn9",
   "tg_tokens": null,
   "tps": 161.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-fbetefn9-ctx65536",
   "model": "CyberTiel-Coder-35B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 34.5 GB at this context; duplicate traces of the same session: 33yxe3u2, 8j2o54mz, ldqw3rie, twsswocq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1292.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "CyberTiel-Coder-35B-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 64k 1,292 115.9 34.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fbetefn9",
   "source_url": "https://omlx.ai/benchmarks/performance/fbetefn9",
   "tg_tokens": null,
   "tps": 115.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-r018rs7f-ctx1024",
   "model": "CyberTiel-Coder-35B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: 6ckgtrv8, 6i076a32, ad8kqjml, z9uwens2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2019.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "CyberTiel-Coder-35B-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 1k 2,019 124.9 30.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r018rs7f",
   "source_url": "https://omlx.ai/benchmarks/performance/r018rs7f",
   "tg_tokens": null,
   "tps": 124.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-r018rs7f-ctx4096",
   "model": "CyberTiel-Coder-35B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: 6ckgtrv8, 6i076a32, ad8kqjml, z9uwens2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2502.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "CyberTiel-Coder-35B-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 4k 2,502 143.1 32.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r018rs7f",
   "source_url": "https://omlx.ai/benchmarks/performance/r018rs7f",
   "tg_tokens": null,
   "tps": 143.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-r018rs7f-ctx16384",
   "model": "CyberTiel-Coder-35B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 32.5 GB at this context; duplicate traces of the same session: 6ckgtrv8, 6i076a32, ad8kqjml, z9uwens2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2109.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "CyberTiel-Coder-35B-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 16k 2,109 124.0 32.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r018rs7f",
   "source_url": "https://omlx.ai/benchmarks/performance/r018rs7f",
   "tg_tokens": null,
   "tps": 124.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-r018rs7f-ctx32768",
   "model": "CyberTiel-Coder-35B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 33.2 GB at this context; duplicate traces of the same session: 6ckgtrv8, 6i076a32, ad8kqjml, z9uwens2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1747.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "CyberTiel-Coder-35B-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 32k 1,747 110.5 33.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r018rs7f",
   "source_url": "https://omlx.ai/benchmarks/performance/r018rs7f",
   "tg_tokens": null,
   "tps": 110.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-r018rs7f-ctx65536",
   "model": "CyberTiel-Coder-35B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 34.5 GB at this context; duplicate traces of the same session: 6ckgtrv8, 6i076a32, ad8kqjml, z9uwens2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1284.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "CyberTiel-Coder-35B-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 6bit; 64k 1,284 74.0 34.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r018rs7f",
   "source_url": "https://omlx.ai/benchmarks/performance/r018rs7f",
   "tg_tokens": null,
   "tps": 74.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-y4bb5s2s-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: scorc59p; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1148.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,148 65.9 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y4bb5s2s",
   "source_url": "https://omlx.ai/benchmarks/performance/y4bb5s2s",
   "tg_tokens": null,
   "tps": 65.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-y4bb5s2s-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: scorc59p; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1255.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,255 52.6 19.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y4bb5s2s",
   "source_url": "https://omlx.ai/benchmarks/performance/y4bb5s2s",
   "tg_tokens": null,
   "tps": 52.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-80g2ffi3-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 207.1 GB at this context; duplicate traces of the same session: 51bcxgsc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 248.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 248.1 9.6 207.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/80g2ffi3",
   "source_url": "https://omlx.ai/benchmarks/performance/80g2ffi3",
   "tg_tokens": null,
   "tps": 9.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-80g2ffi3-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 213.4 GB at this context; duplicate traces of the same session: 51bcxgsc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 361.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 361.3 7.1 213.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/80g2ffi3",
   "source_url": "https://omlx.ai/benchmarks/performance/80g2ffi3",
   "tg_tokens": null,
   "tps": 7.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-i6p5l2sq-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.4 GB at this context; duplicate traces of the same session: 5j5l5zs0; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1875.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,875 67.9 178.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i6p5l2sq",
   "source_url": "https://omlx.ai/benchmarks/performance/i6p5l2sq",
   "tg_tokens": null,
   "tps": 67.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-i6p5l2sq-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.0 GB at this context; duplicate traces of the same session: 5j5l5zs0; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1902.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,902 139.4 179.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i6p5l2sq",
   "source_url": "https://omlx.ai/benchmarks/performance/i6p5l2sq",
   "tg_tokens": null,
   "tps": 139.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-54hlc2g2-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.4 GB at this context; duplicate traces of the same session: qx7dz9l8; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 731.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 731.6 43.2 178.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/54hlc2g2",
   "source_url": "https://omlx.ai/benchmarks/performance/54hlc2g2",
   "tg_tokens": null,
   "tps": 43.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-54hlc2g2-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.0 GB at this context; duplicate traces of the same session: qx7dz9l8; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1878.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,878 100.9 179.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/54hlc2g2",
   "source_url": "https://omlx.ai/benchmarks/performance/54hlc2g2",
   "tg_tokens": null,
   "tps": 100.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-sc927dqi-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: xaref03l; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1038.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.6-35B-A3B-oQ8-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,038 66.6 37.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sc927dqi",
   "source_url": "https://omlx.ai/benchmarks/performance/sc927dqi",
   "tg_tokens": null,
   "tps": 66.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-sc927dqi-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 39.6 GB at this context; duplicate traces of the same session: xaref03l; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1354.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.6-35B-A3B-oQ8-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 8k 1,354 62.6 39.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sc927dqi",
   "source_url": "https://omlx.ai/benchmarks/performance/sc927dqi",
   "tg_tokens": null,
   "tps": 62.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-h6fjmmd6-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "mlx fp16 MTP build; oMLX community benchmark trace; peak memory 40.2 GB at this context; duplicate traces of the same session: 8tn4h8rq, j3nwn8gz; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 192.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-27B-mlx-oQ8-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 192.0 17.3 40.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h6fjmmd6",
   "source_url": "https://omlx.ai/benchmarks/performance/h6fjmmd6",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-h6fjmmd6-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "mlx fp16 MTP build; oMLX community benchmark trace; peak memory 43.6 GB at this context; duplicate traces of the same session: 8tn4h8rq, j3nwn8gz; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 219.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-27B-mlx-oQ8-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 219.0 20.9 43.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h6fjmmd6",
   "source_url": "https://omlx.ai/benchmarks/performance/h6fjmmd6",
   "tg_tokens": null,
   "tps": 20.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-h6fjmmd6-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "mlx fp16 MTP build; oMLX community benchmark trace; peak memory 44.5 GB at this context; duplicate traces of the same session: 8tn4h8rq, j3nwn8gz; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 222.9,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-27B-mlx-oQ8-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 8k 222.9 12.0 44.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h6fjmmd6",
   "source_url": "https://omlx.ai/benchmarks/performance/h6fjmmd6",
   "tg_tokens": null,
   "tps": 12.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-scqnx4t2-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 26.1 GB; TTFT 135152 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 121.2,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k tokens 121.2 PP tok/s 13.8 TG tok/s 135152 TTFT (ms) 26.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/scqnx4t2",
   "source_url": "https://omlx.ai/benchmarks/performance/scqnx4t2",
   "tg_tokens": null,
   "tps": 13.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-dpk8d0nu-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 297.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 1k 297.0 31.1 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dpk8d0nu",
   "source_url": "https://omlx.ai/benchmarks/performance/dpk8d0nu",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-dpk8d0nu-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 320.3,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 4k 320.3 30.6 19.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dpk8d0nu",
   "source_url": "https://omlx.ai/benchmarks/performance/dpk8d0nu",
   "tg_tokens": null,
   "tps": 30.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zgnm758d-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 3rh7j3eo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 691.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 691.7 47.6 18.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zgnm758d",
   "source_url": "https://omlx.ai/benchmarks/performance/zgnm758d",
   "tg_tokens": null,
   "tps": 47.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-zgnm758d-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 3rh7j3eo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 692.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 692.9 86.7 19.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zgnm758d",
   "source_url": "https://omlx.ai/benchmarks/performance/zgnm758d",
   "tg_tokens": null,
   "tps": 86.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8v4zgfxq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: ofa7esv8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 817.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 817.5 47.8 18.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8v4zgfxq",
   "source_url": "https://omlx.ai/benchmarks/performance/8v4zgfxq",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8v4zgfxq-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: ofa7esv8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 700.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 700.0 58.2 19.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8v4zgfxq",
   "source_url": "https://omlx.ai/benchmarks/performance/8v4zgfxq",
   "tg_tokens": null,
   "tps": 58.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-s2ogq620-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: hlrdo1hq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1169.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,169 56.6 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s2ogq620",
   "source_url": "https://omlx.ai/benchmarks/performance/s2ogq620",
   "tg_tokens": null,
   "tps": 56.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-s2ogq620-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: hlrdo1hq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1313.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,313 56.2 19.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s2ogq620",
   "source_url": "https://omlx.ai/benchmarks/performance/s2ogq620",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-uy6rcjm5-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 38.9 GB at this context; duplicate traces of the same session: 152pptnl, ino5hm20, nc2dmzn6, qle91bvh; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 147.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k 147.0 18.5 38.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uy6rcjm5",
   "source_url": "https://omlx.ai/benchmarks/performance/uy6rcjm5",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-uy6rcjm5-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 39.7 GB at this context; duplicate traces of the same session: 152pptnl, ino5hm20, nc2dmzn6, qle91bvh; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 135.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 8k 135.3 19.5 39.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uy6rcjm5",
   "source_url": "https://omlx.ai/benchmarks/performance/uy6rcjm5",
   "tg_tokens": null,
   "tps": 19.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-uy6rcjm5-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 40.8 GB at this context; duplicate traces of the same session: 152pptnl, ino5hm20, nc2dmzn6, qle91bvh; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 138.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 16k 138.1 21.0 40.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uy6rcjm5",
   "source_url": "https://omlx.ai/benchmarks/performance/uy6rcjm5",
   "tg_tokens": null,
   "tps": 21.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-uy6rcjm5-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 41.8 GB at this context; duplicate traces of the same session: 152pptnl, ino5hm20, nc2dmzn6, qle91bvh; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 32k 127.5 21.4 41.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uy6rcjm5",
   "source_url": "https://omlx.ai/benchmarks/performance/uy6rcjm5",
   "tg_tokens": null,
   "tps": 21.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-uy6rcjm5-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 44.0 GB at this context; duplicate traces of the same session: 152pptnl, ino5hm20, nc2dmzn6, qle91bvh; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 110.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 64k 110.3 15.9 44.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uy6rcjm5",
   "source_url": "https://omlx.ai/benchmarks/performance/uy6rcjm5",
   "tg_tokens": null,
   "tps": 15.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-h5k5x8nf-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: 9yy7e5dj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 419.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 419.7 55.1 34.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h5k5x8nf",
   "source_url": "https://omlx.ai/benchmarks/performance/h5k5x8nf",
   "tg_tokens": null,
   "tps": 55.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-h5k5x8nf-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 38.0 GB at this context; duplicate traces of the same session: 9yy7e5dj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 439.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 439.3 54.5 38.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h5k5x8nf",
   "source_url": "https://omlx.ai/benchmarks/performance/h5k5x8nf",
   "tg_tokens": null,
   "tps": 54.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ba33zk9e-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 418.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 418.9 51.0 33.9 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ba33zk9e",
   "source_url": "https://omlx.ai/benchmarks/performance/ba33zk9e",
   "tg_tokens": null,
   "tps": 51.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ba33zk9e-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 441.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 441.4 43.6 37.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ba33zk9e",
   "source_url": "https://omlx.ai/benchmarks/performance/ba33zk9e",
   "tg_tokens": null,
   "tps": 43.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ba33zk9e-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.6 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 436.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 436.3 66.9 37.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ba33zk9e",
   "source_url": "https://omlx.ai/benchmarks/performance/ba33zk9e",
   "tg_tokens": null,
   "tps": 66.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-1byqpk1m-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: 6dww98ex; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 371.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 371.5 25.8 100.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1byqpk1m",
   "source_url": "https://omlx.ai/benchmarks/performance/1byqpk1m",
   "tg_tokens": null,
   "tps": 25.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-1byqpk1m-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.0 GB at this context; duplicate traces of the same session: 6dww98ex; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1138.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,138 94.3 101.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1byqpk1m",
   "source_url": "https://omlx.ai/benchmarks/performance/1byqpk1m",
   "tg_tokens": null,
   "tps": 94.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-eukm7ljg-ctx1024",
   "model": "Spark-X2.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 8.2 GB at this context; duplicate traces of the same session: h62zit1q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1132.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Spark-X2.5-4B on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 unknown; 1k 1,132 41.2 8.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eukm7ljg",
   "source_url": "https://omlx.ai/benchmarks/performance/eukm7ljg",
   "tg_tokens": null,
   "tps": 41.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-eukm7ljg-ctx4096",
   "model": "Spark-X2.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 8.3 GB at this context; duplicate traces of the same session: h62zit1q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1289.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Spark-X2.5-4B on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 unknown; 4k 1,289 40.7 8.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eukm7ljg",
   "source_url": "https://omlx.ai/benchmarks/performance/eukm7ljg",
   "tg_tokens": null,
   "tps": 40.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-u6nvjqml-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: zcywb06m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4811.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 4,811 170.0 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u6nvjqml",
   "source_url": "https://omlx.ai/benchmarks/performance/u6nvjqml",
   "tg_tokens": null,
   "tps": 170.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-u6nvjqml-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: zcywb06m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6092.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 6,092 176.0 22.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u6nvjqml",
   "source_url": "https://omlx.ai/benchmarks/performance/u6nvjqml",
   "tg_tokens": null,
   "tps": 176.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-evo756gd-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: 9wny1obb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1478.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,478 103.1 14.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/evo756gd",
   "source_url": "https://omlx.ai/benchmarks/performance/evo756gd",
   "tg_tokens": null,
   "tps": 103.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-evo756gd-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 15.4 GB at this context; duplicate traces of the same session: 9wny1obb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1355.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,355 89.9 15.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/evo756gd",
   "source_url": "https://omlx.ai/benchmarks/performance/evo756gd",
   "tg_tokens": null,
   "tps": 89.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-tqah04f8-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: w383vhny; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1320.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,320 103.3 14.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tqah04f8",
   "source_url": "https://omlx.ai/benchmarks/performance/tqah04f8",
   "tg_tokens": null,
   "tps": 103.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-tqah04f8-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 15.4 GB at this context; duplicate traces of the same session: w383vhny; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1353.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,353 89.9 15.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tqah04f8",
   "source_url": "https://omlx.ai/benchmarks/performance/tqah04f8",
   "tg_tokens": null,
   "tps": 89.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-257ocp0v-ctx8192",
   "model": "gemma-4-31B-it-scotoma-2",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 21.8 GB; TTFT 23041 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 355.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-scotoma-2-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k tokens 355.5 PP tok/s 24.7 TG tok/s 23041 TTFT (ms) 21.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/257ocp0v",
   "source_url": "https://omlx.ai/benchmarks/performance/257ocp0v",
   "tg_tokens": null,
   "tps": 24.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-cztafniz-ctx8192",
   "model": "gemma-4-31B-it-scotoma-2",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 21.8 GB; TTFT 22143 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 370.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-scotoma-2-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k tokens 370.0 PP tok/s 21.9 TG tok/s 22143 TTFT (ms) 21.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cztafniz",
   "source_url": "https://omlx.ai/benchmarks/performance/cztafniz",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-wzg4ni9d-ctx1024",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 3.8 GB at this context; duplicate traces of the same session: ckpzcka0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1587.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E2B-it-MLX-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 16 GB \u00b7 4bit; 1k 1,587 57.7 3.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wzg4ni9d",
   "source_url": "https://omlx.ai/benchmarks/performance/wzg4ni9d",
   "tg_tokens": null,
   "tps": 57.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-wzg4ni9d-ctx4096",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 3.8 GB at this context; duplicate traces of the same session: ckpzcka0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1873.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E2B-it-MLX-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 16 GB \u00b7 4bit; 4k 1,873 56.1 3.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wzg4ni9d",
   "source_url": "https://omlx.ai/benchmarks/performance/wzg4ni9d",
   "tg_tokens": null,
   "tps": 56.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2",
   "id": "omlx-cauui9hq-ctx1024",
   "model": "gemma-4-E2B-it-Uncensored-MAX",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 3.0 GB; TTFT 1331 ms; duplicate traces of the same session: none; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 769.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E2B-it-Uncensored-MAX on M2 (8c); M2 (8c) \u00b7 16 GB \u00b7 4bit; 1k tokens 769.2 PP tok/s 50.1 TG tok/s 1331 TTFT (ms) 3.0 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cauui9hq",
   "source_url": "https://omlx.ai/benchmarks/performance/cauui9hq",
   "tg_tokens": null,
   "tps": 50.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-1kl8v8ry-ctx1024",
   "model": "gemma-4-E2B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: xft2fcgg; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 955.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E2B-it-oQ4e-mtp on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 4bit; 1k 955.5 80.3 4.2 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1kl8v8ry",
   "source_url": "https://omlx.ai/benchmarks/performance/1kl8v8ry",
   "tg_tokens": null,
   "tps": 80.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-1kl8v8ry-ctx4096",
   "model": "gemma-4-E2B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: xft2fcgg; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 972.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E2B-it-oQ4e-mtp on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 4bit; 4k 972.0 70.5 4.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1kl8v8ry",
   "source_url": "https://omlx.ai/benchmarks/performance/1kl8v8ry",
   "tg_tokens": null,
   "tps": 70.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-n785v5j1-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: hce44gf3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1338.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 1,338 130.8 13.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n785v5j1",
   "source_url": "https://omlx.ai/benchmarks/performance/n785v5j1",
   "tg_tokens": null,
   "tps": 130.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-n785v5j1-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.5 GB at this context; duplicate traces of the same session: hce44gf3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2925.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 2,925 124.0 13.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n785v5j1",
   "source_url": "https://omlx.ai/benchmarks/performance/n785v5j1",
   "tg_tokens": null,
   "tps": 124.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-h8iz9249-ctx1024",
   "model": "Laguna XS 2.1",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 14.3 GB; TTFT 1482 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 691.1,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Laguna-XS-2.1-3bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 3bit; 1k tokens 691.1 PP tok/s 71.5 TG tok/s 1482 TTFT (ms) 14.3 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h8iz9249",
   "source_url": "https://omlx.ai/benchmarks/performance/h8iz9249",
   "tg_tokens": null,
   "tps": 71.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-003ytqw2-ctx1024",
   "model": "Laguna XS 2.1",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 18.1 GB; TTFT 1011 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1013.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Laguna-XS-2.1-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 1,013 PP tok/s 59.7 TG tok/s 1011 TTFT (ms) 18.1 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/003ytqw2",
   "source_url": "https://omlx.ai/benchmarks/performance/003ytqw2",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-j7152o0b-ctx1024",
   "model": "Muse Glimmer",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 15.2 GB; TTFT 6222 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 164.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-30B-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 164.6 PP tok/s 7.9 TG tok/s 6222 TTFT (ms) 15.2 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j7152o0b",
   "source_url": "https://omlx.ai/benchmarks/performance/j7152o0b",
   "tg_tokens": null,
   "tps": 7.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-j2kk19ue-ctx1024",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 14.4 GB; TTFT 5523 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 185.4,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Muse-Glimmer-30B-mxfp4 on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 185.4 PP tok/s 8.6 TG tok/s 5523 TTFT (ms) 14.4 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j2kk19ue",
   "source_url": "https://omlx.ai/benchmarks/performance/j2kk19ue",
   "tg_tokens": null,
   "tps": 8.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-ezc5wjzk-ctx1024",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 15.2 GB; TTFT 5231 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 195.8,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Muse-Glimmer-30B-nvfp4 on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 195.8 PP tok/s 7.8 TG tok/s 5231 TTFT (ms) 15.2 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ezc5wjzk",
   "source_url": "https://omlx.ai/benchmarks/performance/ezc5wjzk",
   "tg_tokens": null,
   "tps": 7.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-iw4728bi-ctx1024",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 15.9 GB; TTFT 5002 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 204.7 PP tok/s 7.9 TG tok/s 5002 TTFT (ms) 15.9 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iw4728bi",
   "source_url": "https://omlx.ai/benchmarks/performance/iw4728bi",
   "tg_tokens": null,
   "tps": 7.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-cu2onb8r-ctx1024",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: ic04kgp2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 480.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 480.4 17.9 15.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cu2onb8r",
   "source_url": "https://omlx.ai/benchmarks/performance/cu2onb8r",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-cu2onb8r-ctx4096",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: ic04kgp2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 471.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 471.2 17.5 16.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cu2onb8r",
   "source_url": "https://omlx.ai/benchmarks/performance/cu2onb8r",
   "tg_tokens": null,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-goycom11-ctx4096",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: eumi57sj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1615.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,615 146.9 19.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/goycom11",
   "source_url": "https://omlx.ai/benchmarks/performance/goycom11",
   "tg_tokens": null,
   "tps": 146.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-goycom11-ctx16384",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: eumi57sj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1503.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,503 134.1 19.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/goycom11",
   "source_url": "https://omlx.ai/benchmarks/performance/goycom11",
   "tg_tokens": null,
   "tps": 134.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-iobz2gh2-ctx4096",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: x4gc3mwd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1613.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,613 149.4 19.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iobz2gh2",
   "source_url": "https://omlx.ai/benchmarks/performance/iobz2gh2",
   "tg_tokens": null,
   "tps": 149.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-iobz2gh2-ctx16384",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: x4gc3mwd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1500.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,500 135.5 19.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iobz2gh2",
   "source_url": "https://omlx.ai/benchmarks/performance/iobz2gh2",
   "tg_tokens": null,
   "tps": 135.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-hg89jojw-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 1528 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 670.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k tokens 670.2 PP tok/s 73.1 TG tok/s 1528 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hg89jojw",
   "source_url": "https://omlx.ai/benchmarks/performance/hg89jojw",
   "tg_tokens": null,
   "tps": 73.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-msvf7wtj-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 3181 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 321.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k tokens 321.9 PP tok/s 36.8 TG tok/s 3181 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/msvf7wtj",
   "source_url": "https://omlx.ai/benchmarks/performance/msvf7wtj",
   "tg_tokens": null,
   "tps": 36.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-3ra968xu-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1398.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,398 191.7 24.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ra968xu",
   "source_url": "https://omlx.ai/benchmarks/performance/3ra968xu",
   "tg_tokens": null,
   "tps": 191.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-3ra968xu-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1020.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 1,020 102.5 25.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ra968xu",
   "source_url": "https://omlx.ai/benchmarks/performance/3ra968xu",
   "tg_tokens": null,
   "tps": 102.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-4n2p27vc-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 158.9,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 158.9 26.6 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4n2p27vc",
   "source_url": "https://omlx.ai/benchmarks/performance/4n2p27vc",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-4n2p27vc-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 155.3,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 155.3 24.8 24.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4n2p27vc",
   "source_url": "https://omlx.ai/benchmarks/performance/4n2p27vc",
   "tg_tokens": null,
   "tps": 24.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-4n2p27vc-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 148.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 148.0 29.6 25.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4n2p27vc",
   "source_url": "https://omlx.ai/benchmarks/performance/4n2p27vc",
   "tg_tokens": null,
   "tps": 29.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-4n2p27vc-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 135.4,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 32k 135.4 31.3 26.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4n2p27vc",
   "source_url": "https://omlx.ai/benchmarks/performance/4n2p27vc",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-4n2p27vc-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.7 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 115.5,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 64k 115.5 22.5 28.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4n2p27vc",
   "source_url": "https://omlx.ai/benchmarks/performance/4n2p27vc",
   "tg_tokens": null,
   "tps": 22.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-qei5qa9d-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 3esx5om2; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 273.6,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 1k 273.6 25.8 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qei5qa9d",
   "source_url": "https://omlx.ai/benchmarks/performance/qei5qa9d",
   "tg_tokens": null,
   "tps": 25.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-qei5qa9d-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: 3esx5om2; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 306.8,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 4k 306.8 42.4 23.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qei5qa9d",
   "source_url": "https://omlx.ai/benchmarks/performance/qei5qa9d",
   "tg_tokens": null,
   "tps": 42.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-h88hyku9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 297.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 1k 297.0 31.1 16.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h88hyku9",
   "source_url": "https://omlx.ai/benchmarks/performance/h88hyku9",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-h88hyku9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 320.3,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 4k 320.3 30.6 19.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h88hyku9",
   "source_url": "https://omlx.ai/benchmarks/performance/h88hyku9",
   "tg_tokens": null,
   "tps": 30.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-pwd33k2i-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 9g8hmr7n, qle46x7x, r4exr5rq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 123.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 123.2 27.9 18.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pwd33k2i",
   "source_url": "https://omlx.ai/benchmarks/performance/pwd33k2i",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-pwd33k2i-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 9g8hmr7n, qle46x7x, r4exr5rq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 122.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 122.1 25.7 19.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pwd33k2i",
   "source_url": "https://omlx.ai/benchmarks/performance/pwd33k2i",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-pwd33k2i-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: 9g8hmr7n, qle46x7x, r4exr5rq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 111.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 111.2 20.8 20.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pwd33k2i",
   "source_url": "https://omlx.ai/benchmarks/performance/pwd33k2i",
   "tg_tokens": null,
   "tps": 20.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-pwd33k2i-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 9g8hmr7n, qle46x7x, r4exr5rq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 109.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 109.6 32.3 20.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pwd33k2i",
   "source_url": "https://omlx.ai/benchmarks/performance/pwd33k2i",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-1sr67i0f-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: fxfi6ctc, n9yv7j3u, zks78uhz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 123.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 123.5 21.8 18.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1sr67i0f",
   "source_url": "https://omlx.ai/benchmarks/performance/1sr67i0f",
   "tg_tokens": null,
   "tps": 21.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-1sr67i0f-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: fxfi6ctc, n9yv7j3u, zks78uhz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 121.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 121.1 22.2 19.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1sr67i0f",
   "source_url": "https://omlx.ai/benchmarks/performance/1sr67i0f",
   "tg_tokens": null,
   "tps": 22.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-1sr67i0f-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: fxfi6ctc, n9yv7j3u, zks78uhz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 114.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 114.2 28.3 20.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1sr67i0f",
   "source_url": "https://omlx.ai/benchmarks/performance/1sr67i0f",
   "tg_tokens": null,
   "tps": 28.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-1sr67i0f-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: fxfi6ctc, n9yv7j3u, zks78uhz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 109.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 109.4 34.6 20.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1sr67i0f",
   "source_url": "https://omlx.ai/benchmarks/performance/1sr67i0f",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-g3zexovl-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: 2xkm70fh, 3u2lp7om, huz3ssjx, jt2t5bsb, r65gq8xo, y5kin30j, yrkuw5rd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 913.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 913.9 74.3 17.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g3zexovl",
   "source_url": "https://omlx.ai/benchmarks/performance/g3zexovl",
   "tg_tokens": null,
   "tps": 74.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-g3zexovl-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 2xkm70fh, 3u2lp7om, huz3ssjx, jt2t5bsb, r65gq8xo, y5kin30j, yrkuw5rd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1005.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,005 60.9 18.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g3zexovl",
   "source_url": "https://omlx.ai/benchmarks/performance/g3zexovl",
   "tg_tokens": null,
   "tps": 60.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-g3zexovl-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 2xkm70fh, 3u2lp7om, huz3ssjx, jt2t5bsb, r65gq8xo, y5kin30j, yrkuw5rd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 931.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 931.0 67.1 19.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g3zexovl",
   "source_url": "https://omlx.ai/benchmarks/performance/g3zexovl",
   "tg_tokens": null,
   "tps": 67.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-g3zexovl-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 2xkm70fh, 3u2lp7om, huz3ssjx, jt2t5bsb, r65gq8xo, y5kin30j, yrkuw5rd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 896.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 896.5 80.8 19.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g3zexovl",
   "source_url": "https://omlx.ai/benchmarks/performance/g3zexovl",
   "tg_tokens": null,
   "tps": 80.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-g3zexovl-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 2xkm70fh, 3u2lp7om, huz3ssjx, jt2t5bsb, r65gq8xo, y5kin30j, yrkuw5rd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 826.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 826.8 58.4 20.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g3zexovl",
   "source_url": "https://omlx.ai/benchmarks/performance/g3zexovl",
   "tg_tokens": null,
   "tps": 58.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-g3zexovl-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: 2xkm70fh, 3u2lp7om, huz3ssjx, jt2t5bsb, r65gq8xo, y5kin30j, yrkuw5rd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 725.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 725.9 58.0 23.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g3zexovl",
   "source_url": "https://omlx.ai/benchmarks/performance/g3zexovl",
   "tg_tokens": null,
   "tps": 58.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-g3zexovl-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: 2xkm70fh, 3u2lp7om, huz3ssjx, jt2t5bsb, r65gq8xo, y5kin30j, yrkuw5rd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 583.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 583.4 50.4 27.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g3zexovl",
   "source_url": "https://omlx.ai/benchmarks/performance/g3zexovl",
   "tg_tokens": null,
   "tps": 50.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-g3zexovl-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.1 GB at this context; duplicate traces of the same session: 2xkm70fh, 3u2lp7om, huz3ssjx, jt2t5bsb, r65gq8xo, y5kin30j, yrkuw5rd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 480.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 480.8 36.0 33.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g3zexovl",
   "source_url": "https://omlx.ai/benchmarks/performance/g3zexovl",
   "tg_tokens": null,
   "tps": 36.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-on24fph3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: wcbihby3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 879.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 879.0 68.8 17.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/on24fph3",
   "source_url": "https://omlx.ai/benchmarks/performance/on24fph3",
   "tg_tokens": null,
   "tps": 68.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-on24fph3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: wcbihby3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 944.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 944.2 61.7 18.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/on24fph3",
   "source_url": "https://omlx.ai/benchmarks/performance/on24fph3",
   "tg_tokens": null,
   "tps": 61.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-n43f2kjt-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace, single run at 4k context; peak memory 38.9 GB; TTFT 27786 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 147.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k tokens 147.4 PP tok/s 16.7 TG tok/s 27786 TTFT (ms) 38.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n43f2kjt",
   "source_url": "https://omlx.ai/benchmarks/performance/n43f2kjt",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-1qahml37-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 51.1 GB at this context; duplicate traces of the same session: oj08opsj; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 350.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 8bit; 1k 350.6 31.8 51.1 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1qahml37",
   "source_url": "https://omlx.ai/benchmarks/performance/1qahml37",
   "tg_tokens": null,
   "tps": 31.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-1qahml37-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 54.4 GB at this context; duplicate traces of the same session: oj08opsj; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 523.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 8bit; 4k 523.3 34.1 54.4 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1qahml37",
   "source_url": "https://omlx.ai/benchmarks/performance/1qahml37",
   "tg_tokens": null,
   "tps": 34.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3cf5f1ze-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: b0bqsa2z; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 418.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 418.9 51.0 33.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3cf5f1ze",
   "source_url": "https://omlx.ai/benchmarks/performance/3cf5f1ze",
   "tg_tokens": null,
   "tps": 51.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-3cf5f1ze-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: b0bqsa2z; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 441.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 441.4 43.6 37.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3cf5f1ze",
   "source_url": "https://omlx.ai/benchmarks/performance/3cf5f1ze",
   "tg_tokens": null,
   "tps": 43.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3cf5f1ze-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.6 GB at this context; duplicate traces of the same session: b0bqsa2z; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 436.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 436.3 66.9 37.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3cf5f1ze",
   "source_url": "https://omlx.ai/benchmarks/performance/3cf5f1ze",
   "tg_tokens": null,
   "tps": 66.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4y42gvcw-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: r5neemjb, zwo733nu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 168.9,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 2bit; 1k 168.9 17.5 24.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4y42gvcw",
   "source_url": "https://omlx.ai/benchmarks/performance/4y42gvcw",
   "tg_tokens": null,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4y42gvcw-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: r5neemjb, zwo733nu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 550.5,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 2bit; 4k 550.5 19.3 25.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4y42gvcw",
   "source_url": "https://omlx.ai/benchmarks/performance/4y42gvcw",
   "tg_tokens": null,
   "tps": 19.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4y42gvcw-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.0 GB at this context; duplicate traces of the same session: r5neemjb, zwo733nu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 877.1,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 2bit; 128k 877.1 18.1 31.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4y42gvcw",
   "source_url": "https://omlx.ai/benchmarks/performance/4y42gvcw",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-v17p3d4v-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: n8desjnt, xxxqwa21; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 135.1,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 2bit; 1k 135.1 13.7 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v17p3d4v",
   "source_url": "https://omlx.ai/benchmarks/performance/v17p3d4v",
   "tg_tokens": null,
   "tps": 13.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-v17p3d4v-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: n8desjnt, xxxqwa21; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 419.0,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 2bit; 4k 419.0 14.0 21.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v17p3d4v",
   "source_url": "https://omlx.ai/benchmarks/performance/v17p3d4v",
   "tg_tokens": null,
   "tps": 14.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-v17p3d4v-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: n8desjnt, xxxqwa21; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 815.5,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 2bit; 128k 815.5 14.3 26.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v17p3d4v",
   "source_url": "https://omlx.ai/benchmarks/performance/v17p3d4v",
   "tg_tokens": null,
   "tps": 14.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-wd48q3xy-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 73.7 GB; TTFT 279523 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 234.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k tokens 234.5 PP tok/s 19.9 TG tok/s 279523 TTFT (ms) 73.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wd48q3xy",
   "source_url": "https://omlx.ai/benchmarks/performance/wd48q3xy",
   "tg_tokens": null,
   "tps": 19.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-dx7wf7xw-ctx1024",
   "model": "Swift 1.5",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 7y2cb2y1; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 208.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Swift-1.5-4bit-MLX on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 1k 208.5 30.7 18.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dx7wf7xw",
   "source_url": "https://omlx.ai/benchmarks/performance/dx7wf7xw",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-dx7wf7xw-ctx4096",
   "model": "Swift 1.5",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: 7y2cb2y1; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 198.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Swift-1.5-4bit-MLX on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 4k 198.2 30.0 22.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dx7wf7xw",
   "source_url": "https://omlx.ai/benchmarks/performance/dx7wf7xw",
   "tg_tokens": null,
   "tps": 30.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-u1pmqey4-ctx1024",
   "model": "TheCluster--amoral-gemma-3-12B-v2",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 7.8 GB at this context; duplicate traces of the same session: 0qkh6t5h; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 884.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TheCluster--amoral-gemma-3-12B-v2-mlx-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 884.4 65.7 7.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u1pmqey4",
   "source_url": "https://omlx.ai/benchmarks/performance/u1pmqey4",
   "tg_tokens": null,
   "tps": 65.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-u1pmqey4-ctx4096",
   "model": "TheCluster--amoral-gemma-3-12B-v2",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 8.3 GB at this context; duplicate traces of the same session: 0qkh6t5h; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 849.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "TheCluster--amoral-gemma-3-12B-v2-mlx-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 849.3 63.1 8.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u1pmqey4",
   "source_url": "https://omlx.ai/benchmarks/performance/u1pmqey4",
   "tg_tokens": null,
   "tps": 63.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-2yhs0guk-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: u5ci18uv; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 194.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 1k 194.8 19.4 22.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2yhs0guk",
   "source_url": "https://omlx.ai/benchmarks/performance/2yhs0guk",
   "tg_tokens": null,
   "tps": 19.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-2yhs0guk-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap DWQ build; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: u5ci18uv; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 203.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 4k 203.1 23.5 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2yhs0guk",
   "source_url": "https://omlx.ai/benchmarks/performance/2yhs0guk",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-2mhqyuez-ctx4096",
   "model": "K2-Horizon-7B-uno",
   "notes": "oMLX community benchmark trace; peak memory 8.3 GB at this context; duplicate traces of the same session: 10mlq42d, mjy00n5n; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 352.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "k2-horizon-7b-uno on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 4k 352.1 31.4 8.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2mhqyuez",
   "source_url": "https://omlx.ai/benchmarks/performance/2mhqyuez",
   "tg_tokens": null,
   "tps": 31.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-2mhqyuez-ctx16384",
   "model": "K2-Horizon-7B-uno",
   "notes": "oMLX community benchmark trace; peak memory 10.0 GB at this context; duplicate traces of the same session: 10mlq42d, mjy00n5n; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 286.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "k2-horizon-7b-uno on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 16k 286.5 22.6 10.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2mhqyuez",
   "source_url": "https://omlx.ai/benchmarks/performance/2mhqyuez",
   "tg_tokens": null,
   "tps": 22.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-2mhqyuez-ctx131072",
   "model": "K2-Horizon-7B-uno",
   "notes": "oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: 10mlq42d, mjy00n5n; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 99.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "k2-horizon-7b-uno on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 128k 99.6 6.1 26.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2mhqyuez",
   "source_url": "https://omlx.ai/benchmarks/performance/2mhqyuez",
   "tg_tokens": null,
   "tps": 6.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-9uilq7rl-ctx1024",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: w2823lem; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 255.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 255.4 28.2 15.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9uilq7rl",
   "source_url": "https://omlx.ai/benchmarks/performance/9uilq7rl",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-9uilq7rl-ctx4096",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: w2823lem; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 220.1 26.1 16.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9uilq7rl",
   "source_url": "https://omlx.ai/benchmarks/performance/9uilq7rl",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-3xfh92aq-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 0j0rw0l8, 4etuqs0q, bchrgyia, e1ayovj9, gjb4cx9p, iwd0oanu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1809.0,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Ornith-1.5-35B-A3B-MXFP4-MLX on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 1,809 143.5 18.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3xfh92aq",
   "source_url": "https://omlx.ai/benchmarks/performance/3xfh92aq",
   "tg_tokens": null,
   "tps": 143.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-3xfh92aq-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 0j0rw0l8, 4etuqs0q, bchrgyia, e1ayovj9, gjb4cx9p, iwd0oanu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2055.0,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Ornith-1.5-35B-A3B-MXFP4-MLX on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 2,055 140.7 19.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3xfh92aq",
   "source_url": "https://omlx.ai/benchmarks/performance/3xfh92aq",
   "tg_tokens": null,
   "tps": 140.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-3xfh92aq-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 0j0rw0l8, 4etuqs0q, bchrgyia, e1ayovj9, gjb4cx9p, iwd0oanu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1953.0,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Ornith-1.5-35B-A3B-MXFP4-MLX on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 8k 1,953 128.9 19.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3xfh92aq",
   "source_url": "https://omlx.ai/benchmarks/performance/3xfh92aq",
   "tg_tokens": null,
   "tps": 128.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-3xfh92aq-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 0j0rw0l8, 4etuqs0q, bchrgyia, e1ayovj9, gjb4cx9p, iwd0oanu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1747.0,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Ornith-1.5-35B-A3B-MXFP4-MLX on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 1,747 122.8 19.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3xfh92aq",
   "source_url": "https://omlx.ai/benchmarks/performance/3xfh92aq",
   "tg_tokens": null,
   "tps": 122.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-3xfh92aq-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 0j0rw0l8, 4etuqs0q, bchrgyia, e1ayovj9, gjb4cx9p, iwd0oanu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1422.0,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Ornith-1.5-35B-A3B-MXFP4-MLX on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 1,422 115.0 19.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3xfh92aq",
   "source_url": "https://omlx.ai/benchmarks/performance/3xfh92aq",
   "tg_tokens": null,
   "tps": 115.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-3xfh92aq-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 0j0rw0l8, 4etuqs0q, bchrgyia, e1ayovj9, gjb4cx9p, iwd0oanu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1046.0,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Ornith-1.5-35B-A3B-MXFP4-MLX on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 64k 1,046 94.1 20.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3xfh92aq",
   "source_url": "https://omlx.ai/benchmarks/performance/3xfh92aq",
   "tg_tokens": null,
   "tps": 94.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-3xfh92aq-ctx131072",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 0j0rw0l8, 4etuqs0q, bchrgyia, e1ayovj9, gjb4cx9p, iwd0oanu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 614.1,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Ornith-1.5-35B-A3B-MXFP4-MLX on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 128k 614.1 72.7 22.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3xfh92aq",
   "source_url": "https://omlx.ai/benchmarks/performance/3xfh92aq",
   "tg_tokens": null,
   "tps": 72.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ji7f8m18-ctx1024",
   "model": "Qwen3.5-0.8B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 2.3 GB at this context; duplicate traces of the same session: 2y3zdsv1, i2m5s8pt, jgjixpx8, qfp84jh1, s9etho36, v339ma7s, w25vw8t1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4330.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.5-0.8B-oQ8-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 1k 4,330 218.7 2.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ji7f8m18",
   "source_url": "https://omlx.ai/benchmarks/performance/ji7f8m18",
   "tg_tokens": null,
   "tps": 218.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ji7f8m18-ctx4096",
   "model": "Qwen3.5-0.8B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 2.5 GB at this context; duplicate traces of the same session: 2y3zdsv1, i2m5s8pt, jgjixpx8, qfp84jh1, s9etho36, v339ma7s, w25vw8t1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4681.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.5-0.8B-oQ8-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 4k 4,681 242.9 2.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ji7f8m18",
   "source_url": "https://omlx.ai/benchmarks/performance/ji7f8m18",
   "tg_tokens": null,
   "tps": 242.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ji7f8m18-ctx8192",
   "model": "Qwen3.5-0.8B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 2.5 GB at this context; duplicate traces of the same session: 2y3zdsv1, i2m5s8pt, jgjixpx8, qfp84jh1, s9etho36, v339ma7s, w25vw8t1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4332.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.5-0.8B-oQ8-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 8k 4,332 189.6 2.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ji7f8m18",
   "source_url": "https://omlx.ai/benchmarks/performance/ji7f8m18",
   "tg_tokens": null,
   "tps": 189.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ji7f8m18-ctx16384",
   "model": "Qwen3.5-0.8B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 2.8 GB at this context; duplicate traces of the same session: 2y3zdsv1, i2m5s8pt, jgjixpx8, qfp84jh1, s9etho36, v339ma7s, w25vw8t1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3716.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.5-0.8B-oQ8-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 16k 3,716 172.0 2.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ji7f8m18",
   "source_url": "https://omlx.ai/benchmarks/performance/ji7f8m18",
   "tg_tokens": null,
   "tps": 172.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ji7f8m18-ctx32768",
   "model": "Qwen3.5-0.8B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 3.0 GB at this context; duplicate traces of the same session: 2y3zdsv1, i2m5s8pt, jgjixpx8, qfp84jh1, s9etho36, v339ma7s, w25vw8t1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2857.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.5-0.8B-oQ8-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 32k 2,857 165.5 3.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ji7f8m18",
   "source_url": "https://omlx.ai/benchmarks/performance/ji7f8m18",
   "tg_tokens": null,
   "tps": 165.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ji7f8m18-ctx65536",
   "model": "Qwen3.5-0.8B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 3.5 GB at this context; duplicate traces of the same session: 2y3zdsv1, i2m5s8pt, jgjixpx8, qfp84jh1, s9etho36, v339ma7s, w25vw8t1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1960.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.5-0.8B-oQ8-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 64k 1,960 135.7 3.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ji7f8m18",
   "source_url": "https://omlx.ai/benchmarks/performance/ji7f8m18",
   "tg_tokens": null,
   "tps": 135.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ji7f8m18-ctx131072",
   "model": "Qwen3.5-0.8B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.8 GB at this context; duplicate traces of the same session: 2y3zdsv1, i2m5s8pt, jgjixpx8, qfp84jh1, s9etho36, v339ma7s, w25vw8t1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1206.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.5-0.8B-oQ8-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 128k 1,206 97.1 4.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ji7f8m18",
   "source_url": "https://omlx.ai/benchmarks/performance/ji7f8m18",
   "tg_tokens": null,
   "tps": 97.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ji7f8m18-ctx199680",
   "model": "Qwen3.5-0.8B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: 2y3zdsv1, i2m5s8pt, jgjixpx8, qfp84jh1, s9etho36, v339ma7s, w25vw8t1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 857.9,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.5-0.8B-oQ8-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 195k 857.9 73.3 6.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ji7f8m18",
   "source_url": "https://omlx.ai/benchmarks/performance/ji7f8m18",
   "tg_tokens": null,
   "tps": 73.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-a5j2qg6k-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: fyukj1vg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2048.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,048 147.1 23.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a5j2qg6k",
   "source_url": "https://omlx.ai/benchmarks/performance/a5j2qg6k",
   "tg_tokens": null,
   "tps": 147.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-a5j2qg6k-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: fyukj1vg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1716.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,716 154.4 23.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a5j2qg6k",
   "source_url": "https://omlx.ai/benchmarks/performance/a5j2qg6k",
   "tg_tokens": null,
   "tps": 154.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-tgjiisq1-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: qrbyw8ts; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2032.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,032 155.1 23.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tgjiisq1",
   "source_url": "https://omlx.ai/benchmarks/performance/tgjiisq1",
   "tg_tokens": null,
   "tps": 155.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-tgjiisq1-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: qrbyw8ts; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1716.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,716 156.8 23.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tgjiisq1",
   "source_url": "https://omlx.ai/benchmarks/performance/tgjiisq1",
   "tg_tokens": null,
   "tps": 156.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-enxo0p18-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1398.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,398 191.7 24.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/enxo0p18",
   "source_url": "https://omlx.ai/benchmarks/performance/enxo0p18",
   "tg_tokens": null,
   "tps": 191.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-enxo0p18-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1020.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 1,020 102.5 25.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/enxo0p18",
   "source_url": "https://omlx.ai/benchmarks/performance/enxo0p18",
   "tg_tokens": null,
   "tps": 102.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-tc5n1knm-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: 3x5zqyy3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1600.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,600 80.5 23.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tc5n1knm",
   "source_url": "https://omlx.ai/benchmarks/performance/tc5n1knm",
   "tg_tokens": null,
   "tps": 80.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-tc5n1knm-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: 3x5zqyy3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1705.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,705 79.1 23.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tc5n1knm",
   "source_url": "https://omlx.ai/benchmarks/performance/tc5n1knm",
   "tg_tokens": null,
   "tps": 79.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-t14vien5-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 8i3vnqd5, bdt8d4t4, h3l5dlo6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 672.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 672.5 126.8 21.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t14vien5",
   "source_url": "https://omlx.ai/benchmarks/performance/t14vien5",
   "tg_tokens": null,
   "tps": 126.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-t14vien5-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 8i3vnqd5, bdt8d4t4, h3l5dlo6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3305.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 3,305 132.4 22.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t14vien5",
   "source_url": "https://omlx.ai/benchmarks/performance/t14vien5",
   "tg_tokens": null,
   "tps": 132.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-t14vien5-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 8i3vnqd5, bdt8d4t4, h3l5dlo6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6114.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 6,114 117.8 22.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t14vien5",
   "source_url": "https://omlx.ai/benchmarks/performance/t14vien5",
   "tg_tokens": null,
   "tps": 117.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-t14vien5-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 8i3vnqd5, bdt8d4t4, h3l5dlo6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5612.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 5,612 163.3 22.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t14vien5",
   "source_url": "https://omlx.ai/benchmarks/performance/t14vien5",
   "tg_tokens": null,
   "tps": 163.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-y67fbt6a-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: jrzc66rz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1149.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 1,149 87.8 37.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y67fbt6a",
   "source_url": "https://omlx.ai/benchmarks/performance/y67fbt6a",
   "tg_tokens": null,
   "tps": 87.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-y67fbt6a-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.1 GB at this context; duplicate traces of the same session: jrzc66rz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1940.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 1,940 101.1 39.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y67fbt6a",
   "source_url": "https://omlx.ai/benchmarks/performance/y67fbt6a",
   "tg_tokens": null,
   "tps": 101.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-wtcpzzl8-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 8iocfi4f; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 213.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 213.2 45.4 20.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wtcpzzl8",
   "source_url": "https://omlx.ai/benchmarks/performance/wtcpzzl8",
   "tg_tokens": null,
   "tps": 45.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-wtcpzzl8-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: 8iocfi4f; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 266.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 266.8 40.2 24.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wtcpzzl8",
   "source_url": "https://omlx.ai/benchmarks/performance/wtcpzzl8",
   "tg_tokens": null,
   "tps": 40.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-yf2vstq5-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 7e2e5wx0, mhopuypt, ww3128bo; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 158.9,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 158.9 26.6 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yf2vstq5",
   "source_url": "https://omlx.ai/benchmarks/performance/yf2vstq5",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-yf2vstq5-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: 7e2e5wx0, mhopuypt, ww3128bo; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 155.3,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 155.3 24.8 24.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yf2vstq5",
   "source_url": "https://omlx.ai/benchmarks/performance/yf2vstq5",
   "tg_tokens": null,
   "tps": 24.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-yf2vstq5-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: 7e2e5wx0, mhopuypt, ww3128bo; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 148.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 148.0 29.6 25.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yf2vstq5",
   "source_url": "https://omlx.ai/benchmarks/performance/yf2vstq5",
   "tg_tokens": null,
   "tps": 29.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-yf2vstq5-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: 7e2e5wx0, mhopuypt, ww3128bo; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 135.4,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 32k 135.4 31.3 26.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yf2vstq5",
   "source_url": "https://omlx.ai/benchmarks/performance/yf2vstq5",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-yf2vstq5-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.7 GB at this context; duplicate traces of the same session: 7e2e5wx0, mhopuypt, ww3128bo; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 115.5,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 64k 115.5 22.5 28.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yf2vstq5",
   "source_url": "https://omlx.ai/benchmarks/performance/yf2vstq5",
   "tg_tokens": null,
   "tps": 22.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ev6lge1p-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: hf326099, l7h5k0kh, rvdgct8l; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 126.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 126.7 29.8 18.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ev6lge1p",
   "source_url": "https://omlx.ai/benchmarks/performance/ev6lge1p",
   "tg_tokens": null,
   "tps": 29.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ev6lge1p-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: hf326099, l7h5k0kh, rvdgct8l; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 125.6 28.8 19.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ev6lge1p",
   "source_url": "https://omlx.ai/benchmarks/performance/ev6lge1p",
   "tg_tokens": null,
   "tps": 28.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ev6lge1p-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: hf326099, l7h5k0kh, rvdgct8l; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 117.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 117.6 27.8 20.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ev6lge1p",
   "source_url": "https://omlx.ai/benchmarks/performance/ev6lge1p",
   "tg_tokens": null,
   "tps": 27.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ev6lge1p-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: hf326099, l7h5k0kh, rvdgct8l; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 113.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 113.3 34.2 20.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ev6lge1p",
   "source_url": "https://omlx.ai/benchmarks/performance/ev6lge1p",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dmauocsk-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: hqf220j7, kmmycpmt, r7vymliv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 708.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 708.1 42.8 33.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dmauocsk",
   "source_url": "https://omlx.ai/benchmarks/performance/dmauocsk",
   "tg_tokens": null,
   "tps": 42.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dmauocsk-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: hqf220j7, kmmycpmt, r7vymliv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 877.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 877.5 42.1 34.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dmauocsk",
   "source_url": "https://omlx.ai/benchmarks/performance/dmauocsk",
   "tg_tokens": null,
   "tps": 42.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dmauocsk-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: hqf220j7, kmmycpmt, r7vymliv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 851.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 851.9 46.8 35.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dmauocsk",
   "source_url": "https://omlx.ai/benchmarks/performance/dmauocsk",
   "tg_tokens": null,
   "tps": 46.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dmauocsk-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: hqf220j7, kmmycpmt, r7vymliv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 786.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 786.0 54.0 35.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dmauocsk",
   "source_url": "https://omlx.ai/benchmarks/performance/dmauocsk",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tfwvc419-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 33.9 GB; TTFT 1240 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 825.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k tokens 825.9 PP tok/s 40.9 TG tok/s 1240 TTFT (ms) 33.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tfwvc419",
   "source_url": "https://omlx.ai/benchmarks/performance/tfwvc419",
   "tg_tokens": null,
   "tps": 40.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-y627o5h0-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 33.9 GB; TTFT 945 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1083.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k tokens 1,083 PP tok/s 48.5 TG tok/s 945 TTFT (ms) 33.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y627o5h0",
   "source_url": "https://omlx.ai/benchmarks/performance/y627o5h0",
   "tg_tokens": null,
   "tps": 48.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-hsry38nv-ctx16384",
   "model": "Qwen3.8-35B-36.82G-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace, single run at 16k context; peak memory 40.7 GB; TTFT 11434 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1433.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-36.82G-A3B-Distill-oQ8-fp16-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k tokens 1,433 PP tok/s 50.8 TG tok/s 11434 TTFT (ms) 40.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hsry38nv",
   "source_url": "https://omlx.ai/benchmarks/performance/hsry38nv",
   "tg_tokens": null,
   "tps": 50.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-23vfcwmf-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: rveygco4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 694.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 694.9 69.9 71.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/23vfcwmf",
   "source_url": "https://omlx.ai/benchmarks/performance/23vfcwmf",
   "tg_tokens": null,
   "tps": 69.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-23vfcwmf-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 75.9 GB at this context; duplicate traces of the same session: rveygco4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 354.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 354.2 25.8 75.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/23vfcwmf",
   "source_url": "https://omlx.ai/benchmarks/performance/23vfcwmf",
   "tg_tokens": null,
   "tps": 25.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-vm0ud3ra-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: 0a6yrng9, 5yebngg6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 298.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 298.2 26.2 22.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vm0ud3ra",
   "source_url": "https://omlx.ai/benchmarks/performance/vm0ud3ra",
   "tg_tokens": null,
   "tps": 26.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-vm0ud3ra-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: 0a6yrng9, 5yebngg6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 219.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 8k 219.4 25.0 23.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vm0ud3ra",
   "source_url": "https://omlx.ai/benchmarks/performance/vm0ud3ra",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-vm0ud3ra-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: 0a6yrng9, 5yebngg6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 226.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 226.3 36.1 21.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vm0ud3ra",
   "source_url": "https://omlx.ai/benchmarks/performance/vm0ud3ra",
   "tg_tokens": null,
   "tps": 36.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-vm0ud3ra-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 0a6yrng9, 5yebngg6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 171.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 171.9 24.6 23.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vm0ud3ra",
   "source_url": "https://omlx.ai/benchmarks/performance/vm0ud3ra",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-pl8jorr9-ctx1024",
   "model": "gemma-4-E2B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 4.1 GB; TTFT 249 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4109.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E2B-it-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 4,109 PP tok/s 131.2 TG tok/s 249 TTFT (ms) 4.1 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pl8jorr9",
   "source_url": "https://omlx.ai/benchmarks/performance/pl8jorr9",
   "tg_tokens": null,
   "tps": 131.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-kur3bxry-ctx1024",
   "model": "gemma-4-E2B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 3.1 GB; TTFT 245 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4180.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E2B-it-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 4,180 PP tok/s 75.2 TG tok/s 245 TTFT (ms) 3.1 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kur3bxry",
   "source_url": "https://omlx.ai/benchmarks/performance/kur3bxry",
   "tg_tokens": null,
   "tps": 75.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-kvy06tn1-ctx1024",
   "model": "gemma-4-E2B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 4.1 GB; TTFT 248 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4132.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E2B-it-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 4,132 PP tok/s 127.0 TG tok/s 248 TTFT (ms) 4.1 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kvy06tn1",
   "source_url": "https://omlx.ai/benchmarks/performance/kvy06tn1",
   "tg_tokens": null,
   "tps": 127.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-kwqzmssv-ctx1024",
   "model": "gemma-4-E2B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 3.1 GB; TTFT 235 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4354.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E2B-it-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 4,354 PP tok/s 75.6 TG tok/s 235 TTFT (ms) 3.1 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kwqzmssv",
   "source_url": "https://omlx.ai/benchmarks/performance/kwqzmssv",
   "tg_tokens": null,
   "tps": 75.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-n2l17k97-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: 3pw36zas; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1226.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,226 122.3 11.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n2l17k97",
   "source_url": "https://omlx.ai/benchmarks/performance/n2l17k97",
   "tg_tokens": null,
   "tps": 122.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-n2l17k97-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: 3pw36zas; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1652.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,652 118.7 11.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n2l17k97",
   "source_url": "https://omlx.ai/benchmarks/performance/n2l17k97",
   "tg_tokens": null,
   "tps": 118.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-lz8yc1ou-ctx4096",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "prism-ml MLX build; oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: y6eylck8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 144.6,
   "provenance": "community",
   "quant": "2bit",
   "quote": "prism-ml--Ternary-Bonsai-2-27B-mlx-2bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 unknown; 4k 144.6 10.7 14.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lz8yc1ou",
   "source_url": "https://omlx.ai/benchmarks/performance/lz8yc1ou",
   "tg_tokens": null,
   "tps": 10.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-lz8yc1ou-ctx16384",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "prism-ml MLX build; oMLX community benchmark trace; peak memory 15.7 GB at this context; duplicate traces of the same session: y6eylck8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.6,
   "provenance": "community",
   "quant": "2bit",
   "quote": "prism-ml--Ternary-Bonsai-2-27B-mlx-2bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 unknown; 16k 131.6 9.0 15.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lz8yc1ou",
   "source_url": "https://omlx.ai/benchmarks/performance/lz8yc1ou",
   "tg_tokens": null,
   "tps": 9.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-8t7wc8ji-ctx4096",
   "model": "Huihui-Qwen3.8-27B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 19.5 GB; TTFT 31776 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 128.9,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Huihui-Qwen3.8-27B-abliterated-oQ3.5e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 3bit; 4k tokens 128.9 PP tok/s 22.4 TG tok/s 31776 TTFT (ms) 19.5 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8t7wc8ji",
   "source_url": "https://omlx.ai/benchmarks/performance/8t7wc8ji",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-5ebhe3ic-ctx65536",
   "model": "KAT-Coder-V2.5-Dev",
   "notes": "MTP build; oMLX community benchmark trace, single run at 64k context; peak memory 22.6 GB; TTFT 120883 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 542.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "KAT-Coder-V2.5-Dev-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k tokens 542.1 PP tok/s 50.6 TG tok/s 120883 TTFT (ms) 22.6 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ebhe3ic",
   "source_url": "https://omlx.ai/benchmarks/performance/5ebhe3ic",
   "tg_tokens": null,
   "tps": 50.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-rs7gt0aa-ctx4096",
   "model": "KAT-Coder-V2.5-Dev",
   "notes": "MTP build; oMLX community benchmark trace, single run at 4k context; peak memory 22.5 GB; TTFT 4129 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 992.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "KAT-Coder-V2.5-Dev-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k tokens 992.1 PP tok/s 67.8 TG tok/s 4129 TTFT (ms) 22.5 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rs7gt0aa",
   "source_url": "https://omlx.ai/benchmarks/performance/rs7gt0aa",
   "tg_tokens": null,
   "tps": 67.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-sdmu99wl-ctx65536",
   "model": "KAT-Coder-V2.5-Dev",
   "notes": "MTP build; oMLX community benchmark trace, single run at 64k context; peak memory 24.3 GB; TTFT 130508 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 502.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "KAT-Coder-V2.5-Dev-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k tokens 502.2 PP tok/s 45.8 TG tok/s 130508 TTFT (ms) 24.3 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sdmu99wl",
   "source_url": "https://omlx.ai/benchmarks/performance/sdmu99wl",
   "tg_tokens": null,
   "tps": 45.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-4dv5db1d-ctx1024",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: mvzk7bkm; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2367.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 2,367 94.1 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4dv5db1d",
   "source_url": "https://omlx.ai/benchmarks/performance/4dv5db1d",
   "tg_tokens": null,
   "tps": 94.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-4dv5db1d-ctx4096",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: mvzk7bkm; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2655.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 2,655 81.0 17.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4dv5db1d",
   "source_url": "https://omlx.ai/benchmarks/performance/4dv5db1d",
   "tg_tokens": null,
   "tps": 81.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-m9xyl8xh-ctx1024",
   "model": "Qwen3.5-0.8B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 1.9 GB at this context; duplicate traces of the same session: 2e48bot5, 7oigh1o0, bwl67u5m, t9s140ge, vfbgcho6, yd7wcws2, yio8ej9c; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4319.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.5-0.8B-oQ8-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 1k 4,319 207.4 1.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m9xyl8xh",
   "source_url": "https://omlx.ai/benchmarks/performance/m9xyl8xh",
   "tg_tokens": null,
   "tps": 207.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-m9xyl8xh-ctx4096",
   "model": "Qwen3.5-0.8B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 2.1 GB at this context; duplicate traces of the same session: 2e48bot5, 7oigh1o0, bwl67u5m, t9s140ge, vfbgcho6, yd7wcws2, yio8ej9c; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4988.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.5-0.8B-oQ8-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 4k 4,988 210.5 2.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m9xyl8xh",
   "source_url": "https://omlx.ai/benchmarks/performance/m9xyl8xh",
   "tg_tokens": null,
   "tps": 210.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-m9xyl8xh-ctx8192",
   "model": "Qwen3.5-0.8B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 2.2 GB at this context; duplicate traces of the same session: 2e48bot5, 7oigh1o0, bwl67u5m, t9s140ge, vfbgcho6, yd7wcws2, yio8ej9c; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4706.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.5-0.8B-oQ8-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 8k 4,706 204.5 2.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m9xyl8xh",
   "source_url": "https://omlx.ai/benchmarks/performance/m9xyl8xh",
   "tg_tokens": null,
   "tps": 204.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-m9xyl8xh-ctx16384",
   "model": "Qwen3.5-0.8B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 2.3 GB at this context; duplicate traces of the same session: 2e48bot5, 7oigh1o0, bwl67u5m, t9s140ge, vfbgcho6, yd7wcws2, yio8ej9c; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4091.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.5-0.8B-oQ8-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 16k 4,091 181.3 2.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m9xyl8xh",
   "source_url": "https://omlx.ai/benchmarks/performance/m9xyl8xh",
   "tg_tokens": null,
   "tps": 181.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-m9xyl8xh-ctx32768",
   "model": "Qwen3.5-0.8B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 2.5 GB at this context; duplicate traces of the same session: 2e48bot5, 7oigh1o0, bwl67u5m, t9s140ge, vfbgcho6, yd7wcws2, yio8ej9c; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3198.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.5-0.8B-oQ8-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 32k 3,198 157.4 2.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m9xyl8xh",
   "source_url": "https://omlx.ai/benchmarks/performance/m9xyl8xh",
   "tg_tokens": null,
   "tps": 157.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-m9xyl8xh-ctx65536",
   "model": "Qwen3.5-0.8B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 2.9 GB at this context; duplicate traces of the same session: 2e48bot5, 7oigh1o0, bwl67u5m, t9s140ge, vfbgcho6, yd7wcws2, yio8ej9c; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2214.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.5-0.8B-oQ8-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 64k 2,214 121.6 2.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m9xyl8xh",
   "source_url": "https://omlx.ai/benchmarks/performance/m9xyl8xh",
   "tg_tokens": null,
   "tps": 121.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-m9xyl8xh-ctx131072",
   "model": "Qwen3.5-0.8B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.1 GB at this context; duplicate traces of the same session: 2e48bot5, 7oigh1o0, bwl67u5m, t9s140ge, vfbgcho6, yd7wcws2, yio8ej9c; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1396.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.5-0.8B-oQ8-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 128k 1,396 82.9 4.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m9xyl8xh",
   "source_url": "https://omlx.ai/benchmarks/performance/m9xyl8xh",
   "tg_tokens": null,
   "tps": 82.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-m9xyl8xh-ctx199680",
   "model": "Qwen3.5-0.8B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 5.0 GB at this context; duplicate traces of the same session: 2e48bot5, 7oigh1o0, bwl67u5m, t9s140ge, vfbgcho6, yd7wcws2, yio8ej9c; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 938.4,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.5-0.8B-oQ8-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 195k 938.4 62.6 5.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m9xyl8xh",
   "source_url": "https://omlx.ai/benchmarks/performance/m9xyl8xh",
   "tg_tokens": null,
   "tps": 62.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7xm992au-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 1k context; peak memory 37.3 GB; TTFT 1080 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 948.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k tokens 948.6 PP tok/s 109.3 TG tok/s 1080 TTFT (ms) 37.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7xm992au",
   "source_url": "https://omlx.ai/benchmarks/performance/7xm992au",
   "tg_tokens": null,
   "tps": 109.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-zj03tj9z-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: ouqb2rrh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 297.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 297.0 51.6 22.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zj03tj9z",
   "source_url": "https://omlx.ai/benchmarks/performance/zj03tj9z",
   "tg_tokens": null,
   "tps": 51.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-zj03tj9z-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: ouqb2rrh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 303.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 8k 303.1 45.3 23.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zj03tj9z",
   "source_url": "https://omlx.ai/benchmarks/performance/zj03tj9z",
   "tg_tokens": null,
   "tps": 45.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-zj03tj9z-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: ouqb2rrh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 225.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 225.9 34.7 21.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zj03tj9z",
   "source_url": "https://omlx.ai/benchmarks/performance/zj03tj9z",
   "tg_tokens": null,
   "tps": 34.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-zj03tj9z-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: ouqb2rrh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 165.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 165.0 23.3 23.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zj03tj9z",
   "source_url": "https://omlx.ai/benchmarks/performance/zj03tj9z",
   "tg_tokens": null,
   "tps": 23.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-zj03tj9z-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.9 GB at this context; duplicate traces of the same session: ouqb2rrh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 151.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 64k 151.3 16.4 28.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zj03tj9z",
   "source_url": "https://omlx.ai/benchmarks/performance/zj03tj9z",
   "tg_tokens": null,
   "tps": 16.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ge0ktqe0-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 5dtyb1bv, hg82ttaj, rtnd6jf8; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 162.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 1k 162.9 24.5 18.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ge0ktqe0",
   "source_url": "https://omlx.ai/benchmarks/performance/ge0ktqe0",
   "tg_tokens": null,
   "tps": 24.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ge0ktqe0-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 5dtyb1bv, hg82ttaj, rtnd6jf8; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 173.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 4k 173.4 26.4 22.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ge0ktqe0",
   "source_url": "https://omlx.ai/benchmarks/performance/ge0ktqe0",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ge0ktqe0-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 5dtyb1bv, hg82ttaj, rtnd6jf8; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 154.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 8k 154.7 37.1 22.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ge0ktqe0",
   "source_url": "https://omlx.ai/benchmarks/performance/ge0ktqe0",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ge0ktqe0-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: 5dtyb1bv, hg82ttaj, rtnd6jf8; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 146.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 16k 146.8 22.4 24.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ge0ktqe0",
   "source_url": "https://omlx.ai/benchmarks/performance/ge0ktqe0",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-a6zlu3wd-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 0bf3enpb, 2aiam4ng, 3ozv7xlp, 4kl56kdh, fe6v4ic1, idos3pns, mw5mfae6; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 114.4,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 114.4 12.3 21.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a6zlu3wd",
   "source_url": "https://omlx.ai/benchmarks/performance/a6zlu3wd",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-a6zlu3wd-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: 0bf3enpb, 2aiam4ng, 3ozv7xlp, 4kl56kdh, fe6v4ic1, idos3pns, mw5mfae6; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 125.6,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 125.6 9.9 24.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a6zlu3wd",
   "source_url": "https://omlx.ai/benchmarks/performance/a6zlu3wd",
   "tg_tokens": null,
   "tps": 9.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-a6zlu3wd-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: 0bf3enpb, 2aiam4ng, 3ozv7xlp, 4kl56kdh, fe6v4ic1, idos3pns, mw5mfae6; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 120.6,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 120.6 9.9 25.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a6zlu3wd",
   "source_url": "https://omlx.ai/benchmarks/performance/a6zlu3wd",
   "tg_tokens": null,
   "tps": 9.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-a6zlu3wd-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: 0bf3enpb, 2aiam4ng, 3ozv7xlp, 4kl56kdh, fe6v4ic1, idos3pns, mw5mfae6; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 119.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 119.0 14.6 26.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a6zlu3wd",
   "source_url": "https://omlx.ai/benchmarks/performance/a6zlu3wd",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-a6zlu3wd-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.2 GB at this context; duplicate traces of the same session: 0bf3enpb, 2aiam4ng, 3ozv7xlp, 4kl56kdh, fe6v4ic1, idos3pns, mw5mfae6; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 111.2,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 32k 111.2 11.8 27.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a6zlu3wd",
   "source_url": "https://omlx.ai/benchmarks/performance/a6zlu3wd",
   "tg_tokens": null,
   "tps": 11.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-a6zlu3wd-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.2 GB at this context; duplicate traces of the same session: 0bf3enpb, 2aiam4ng, 3ozv7xlp, 4kl56kdh, fe6v4ic1, idos3pns, mw5mfae6; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 99.9,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 64k 99.9 10.1 29.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a6zlu3wd",
   "source_url": "https://omlx.ai/benchmarks/performance/a6zlu3wd",
   "tg_tokens": null,
   "tps": 10.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-a6zlu3wd-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.2 GB at this context; duplicate traces of the same session: 0bf3enpb, 2aiam4ng, 3ozv7xlp, 4kl56kdh, fe6v4ic1, idos3pns, mw5mfae6; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 81.3,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 128k 81.3 8.7 33.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a6zlu3wd",
   "source_url": "https://omlx.ai/benchmarks/performance/a6zlu3wd",
   "tg_tokens": null,
   "tps": 8.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-a6zlu3wd-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: 0bf3enpb, 2aiam4ng, 3ozv7xlp, 4kl56kdh, fe6v4ic1, idos3pns, mw5mfae6; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 68.1,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 195k 68.1 5.7 37.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a6zlu3wd",
   "source_url": "https://omlx.ai/benchmarks/performance/a6zlu3wd",
   "tg_tokens": null,
   "tps": 5.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-sk3s17fj-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.2 GB at this context; duplicate traces of the same session: clfu4ijj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 302.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 302.3 46.4 25.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sk3s17fj",
   "source_url": "https://omlx.ai/benchmarks/performance/sk3s17fj",
   "tg_tokens": null,
   "tps": 46.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-sk3s17fj-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.0 GB at this context; duplicate traces of the same session: clfu4ijj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 85.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 85.5 17.7 35.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sk3s17fj",
   "source_url": "https://omlx.ai/benchmarks/performance/sk3s17fj",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-h74mfrg2-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: a0jcgcup, vpnml22n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 899.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 899.2 63.5 17.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h74mfrg2",
   "source_url": "https://omlx.ai/benchmarks/performance/h74mfrg2",
   "tg_tokens": null,
   "tps": 63.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-h74mfrg2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: a0jcgcup, vpnml22n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 983.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 983.6 67.4 18.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h74mfrg2",
   "source_url": "https://omlx.ai/benchmarks/performance/h74mfrg2",
   "tg_tokens": null,
   "tps": 67.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-h74mfrg2-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: a0jcgcup, vpnml22n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 598.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 598.0 41.3 27.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h74mfrg2",
   "source_url": "https://omlx.ai/benchmarks/performance/h74mfrg2",
   "tg_tokens": null,
   "tps": 41.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-o0imkit7-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.7 GB at this context; duplicate traces of the same session: 6f360j5e, a9zpdf8p, edgai2fm, ew2qnqk0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 378.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 1k 378.5 40.9 39.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o0imkit7",
   "source_url": "https://omlx.ai/benchmarks/performance/o0imkit7",
   "tg_tokens": null,
   "tps": 40.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-o0imkit7-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.0 GB at this context; duplicate traces of the same session: 6f360j5e, a9zpdf8p, edgai2fm, ew2qnqk0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 449.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 4k 449.5 49.6 43.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o0imkit7",
   "source_url": "https://omlx.ai/benchmarks/performance/o0imkit7",
   "tg_tokens": null,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-o0imkit7-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.4 GB at this context; duplicate traces of the same session: 6f360j5e, a9zpdf8p, edgai2fm, ew2qnqk0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 480.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 8k 480.2 53.3 43.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o0imkit7",
   "source_url": "https://omlx.ai/benchmarks/performance/o0imkit7",
   "tg_tokens": null,
   "tps": 53.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-o0imkit7-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 44.2 GB at this context; duplicate traces of the same session: 6f360j5e, a9zpdf8p, edgai2fm, ew2qnqk0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 477.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 16k 477.6 58.4 44.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o0imkit7",
   "source_url": "https://omlx.ai/benchmarks/performance/o0imkit7",
   "tg_tokens": null,
   "tps": 58.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-o0imkit7-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 45.8 GB at this context; duplicate traces of the same session: 6f360j5e, a9zpdf8p, edgai2fm, ew2qnqk0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 445.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 32k 445.3 94.8 45.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o0imkit7",
   "source_url": "https://omlx.ai/benchmarks/performance/o0imkit7",
   "tg_tokens": null,
   "tps": 94.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-o0imkit7-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 49.0 GB at this context; duplicate traces of the same session: 6f360j5e, a9zpdf8p, edgai2fm, ew2qnqk0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 403.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 64k 403.4 53.5 49.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o0imkit7",
   "source_url": "https://omlx.ai/benchmarks/performance/o0imkit7",
   "tg_tokens": null,
   "tps": 53.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-o0imkit7-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 53.5 GB at this context; duplicate traces of the same session: 6f360j5e, a9zpdf8p, edgai2fm, ew2qnqk0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 314.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 128k 314.3 50.6 53.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o0imkit7",
   "source_url": "https://omlx.ai/benchmarks/performance/o0imkit7",
   "tg_tokens": null,
   "tps": 50.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-o0imkit7-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 58.4 GB at this context; duplicate traces of the same session: 6f360j5e, a9zpdf8p, edgai2fm, ew2qnqk0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 254.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 195k 254.3 34.1 58.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o0imkit7",
   "source_url": "https://omlx.ai/benchmarks/performance/o0imkit7",
   "tg_tokens": null,
   "tps": 34.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-l8mfi9bj-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 33.7 GB; TTFT 932 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1098.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k tokens 1,098 PP tok/s 42.2 TG tok/s 932 TTFT (ms) 33.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l8mfi9bj",
   "source_url": "https://omlx.ai/benchmarks/performance/l8mfi9bj",
   "tg_tokens": null,
   "tps": 42.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-10foowox-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: 1mtaqu0l, 222u6zkc, 3yk34juv, 5hhjaw1g, bfbdglvz, gjbt6ptz, k0up5p9k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 812.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 812.3 52.5 33.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/10foowox",
   "source_url": "https://omlx.ai/benchmarks/performance/10foowox",
   "tg_tokens": null,
   "tps": 52.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-10foowox-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: 1mtaqu0l, 222u6zkc, 3yk34juv, 5hhjaw1g, bfbdglvz, gjbt6ptz, k0up5p9k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 917.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 917.7 42.9 34.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/10foowox",
   "source_url": "https://omlx.ai/benchmarks/performance/10foowox",
   "tg_tokens": null,
   "tps": 42.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-10foowox-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: 1mtaqu0l, 222u6zkc, 3yk34juv, 5hhjaw1g, bfbdglvz, gjbt6ptz, k0up5p9k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 852.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 852.1 45.0 35.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/10foowox",
   "source_url": "https://omlx.ai/benchmarks/performance/10foowox",
   "tg_tokens": null,
   "tps": 45.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-10foowox-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: 1mtaqu0l, 222u6zkc, 3yk34juv, 5hhjaw1g, bfbdglvz, gjbt6ptz, k0up5p9k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 802.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 802.7 55.8 35.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/10foowox",
   "source_url": "https://omlx.ai/benchmarks/performance/10foowox",
   "tg_tokens": null,
   "tps": 55.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-10foowox-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: 1mtaqu0l, 222u6zkc, 3yk34juv, 5hhjaw1g, bfbdglvz, gjbt6ptz, k0up5p9k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 750.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 750.1 58.3 36.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/10foowox",
   "source_url": "https://omlx.ai/benchmarks/performance/10foowox",
   "tg_tokens": null,
   "tps": 58.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-10foowox-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: 1mtaqu0l, 222u6zkc, 3yk34juv, 5hhjaw1g, bfbdglvz, gjbt6ptz, k0up5p9k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 666.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 666.8 44.3 39.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/10foowox",
   "source_url": "https://omlx.ai/benchmarks/performance/10foowox",
   "tg_tokens": null,
   "tps": 44.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-10foowox-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 44.0 GB at this context; duplicate traces of the same session: 1mtaqu0l, 222u6zkc, 3yk34juv, 5hhjaw1g, bfbdglvz, gjbt6ptz, k0up5p9k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 545.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 545.1 43.4 44.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/10foowox",
   "source_url": "https://omlx.ai/benchmarks/performance/10foowox",
   "tg_tokens": null,
   "tps": 43.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-10foowox-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 49.2 GB at this context; duplicate traces of the same session: 1mtaqu0l, 222u6zkc, 3yk34juv, 5hhjaw1g, bfbdglvz, gjbt6ptz, k0up5p9k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 455.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 195k 455.5 29.2 49.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/10foowox",
   "source_url": "https://omlx.ai/benchmarks/performance/10foowox",
   "tg_tokens": null,
   "tps": 29.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-8ddbqe5h-ctx4096",
   "model": "Qwen3.8-35B-A3B-DistillAPEX-MiniPlus-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: lc769yc3; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1063.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-35B-A3B-DistillAPEX-MiniPlus-Abliterated on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 1,063 27.9 22.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8ddbqe5h",
   "source_url": "https://omlx.ai/benchmarks/performance/8ddbqe5h",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-8ddbqe5h-ctx16384",
   "model": "Qwen3.8-35B-A3B-DistillAPEX-MiniPlus-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: lc769yc3; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1236.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-35B-A3B-DistillAPEX-MiniPlus-Abliterated on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 1,236 25.9 24.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8ddbqe5h",
   "source_url": "https://omlx.ai/benchmarks/performance/8ddbqe5h",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-u60irsez-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: w5bm793y, wpody1at; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 162.5,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 2bit; 1k 162.5 27.1 37.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u60irsez",
   "source_url": "https://omlx.ai/benchmarks/performance/u60irsez",
   "tg_tokens": null,
   "tps": 27.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-u60irsez-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.4 GB at this context; duplicate traces of the same session: w5bm793y, wpody1at; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 750.3,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 2bit; 4k 750.3 27.3 38.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u60irsez",
   "source_url": "https://omlx.ai/benchmarks/performance/u60irsez",
   "tg_tokens": null,
   "tps": 27.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-u60irsez-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.1 GB at this context; duplicate traces of the same session: w5bm793y, wpody1at; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 738.9,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 2bit; 128k 738.9 16.8 43.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u60irsez",
   "source_url": "https://omlx.ai/benchmarks/performance/u60irsez",
   "tg_tokens": null,
   "tps": 16.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5f36n4u9-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 41.6 GB at this context; duplicate traces of the same session: 7q68lwyc, h1zhgbtl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 158.3,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 2bit; 1k 158.3 32.5 41.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5f36n4u9",
   "source_url": "https://omlx.ai/benchmarks/performance/5f36n4u9",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5f36n4u9-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 42.6 GB at this context; duplicate traces of the same session: 7q68lwyc, h1zhgbtl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 811.4,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 2bit; 4k 811.4 29.2 42.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5f36n4u9",
   "source_url": "https://omlx.ai/benchmarks/performance/5f36n4u9",
   "tg_tokens": null,
   "tps": 29.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5f36n4u9-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 46.3 GB at this context; duplicate traces of the same session: 7q68lwyc, h1zhgbtl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 687.0,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 2bit; 128k 687.0 13.5 46.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5f36n4u9",
   "source_url": "https://omlx.ai/benchmarks/performance/5f36n4u9",
   "tg_tokens": null,
   "tps": 13.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1fnsw5xh-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.5 GB; TTFT 1136 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 901.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 901.8 PP tok/s 55.1 TG tok/s 1136 TTFT (ms) 70.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1fnsw5xh",
   "source_url": "https://omlx.ai/benchmarks/performance/1fnsw5xh",
   "tg_tokens": null,
   "tps": 55.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8d9r9wdw-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.5 GB; TTFT 953 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1075.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,075 PP tok/s 63.6 TG tok/s 953 TTFT (ms) 70.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8d9r9wdw",
   "source_url": "https://omlx.ai/benchmarks/performance/8d9r9wdw",
   "tg_tokens": null,
   "tps": 63.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ept20wiy-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 101.2 GB; TTFT 1842 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2223.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k tokens 2,223 PP tok/s 140.0 TG tok/s 1842 TTFT (ms) 101.2 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ept20wiy",
   "source_url": "https://omlx.ai/benchmarks/performance/ept20wiy",
   "tg_tokens": null,
   "tps": 140.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-637poeya-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 298.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 298.2 26.2 22.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/637poeya",
   "source_url": "https://omlx.ai/benchmarks/performance/637poeya",
   "tg_tokens": null,
   "tps": 26.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-637poeya-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 219.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 8k 219.4 25.0 23.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/637poeya",
   "source_url": "https://omlx.ai/benchmarks/performance/637poeya",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-637poeya-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 226.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 226.3 36.1 21.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/637poeya",
   "source_url": "https://omlx.ai/benchmarks/performance/637poeya",
   "tg_tokens": null,
   "tps": 36.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-637poeya-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 171.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 171.9 24.6 23.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/637poeya",
   "source_url": "https://omlx.ai/benchmarks/performance/637poeya",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-obspu3g2-ctx65536",
   "model": "KAT-Coder-V2.5-Dev",
   "notes": "MTP build; oMLX community benchmark trace, single run at 64k context; peak memory 22.6 GB; TTFT 120825 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 542.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "KAT-Coder-V2.5-Dev-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k tokens 542.4 PP tok/s 51.0 TG tok/s 120825 TTFT (ms) 22.6 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/obspu3g2",
   "source_url": "https://omlx.ai/benchmarks/performance/obspu3g2",
   "tg_tokens": null,
   "tps": 51.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-mbuuvlvq-ctx65536",
   "model": "KAT-Coder-V2.5-Dev",
   "notes": "MTP build; oMLX community benchmark trace, single run at 64k context; peak memory 22.6 GB; TTFT 121252 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 540.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "KAT-Coder-V2.5-Dev-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k tokens 540.5 PP tok/s 51.1 TG tok/s 121252 TTFT (ms) 22.6 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mbuuvlvq",
   "source_url": "https://omlx.ai/benchmarks/performance/mbuuvlvq",
   "tg_tokens": null,
   "tps": 51.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-wzh0c765-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: 2xqt6cg3, hsznepw6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2261.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,261 154.0 13.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wzh0c765",
   "source_url": "https://omlx.ai/benchmarks/performance/wzh0c765",
   "tg_tokens": null,
   "tps": 154.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-wzh0c765-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.6 GB at this context; duplicate traces of the same session: 2xqt6cg3, hsznepw6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2557.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,557 150.6 13.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wzh0c765",
   "source_url": "https://omlx.ai/benchmarks/performance/wzh0c765",
   "tg_tokens": null,
   "tps": 150.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-wzh0c765-ctx131072",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 2xqt6cg3, hsznepw6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 754.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 754.5 16.3 16.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wzh0c765",
   "source_url": "https://omlx.ai/benchmarks/performance/wzh0c765",
   "tg_tokens": null,
   "tps": 16.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4v2lcrr3-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5753.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 5,753 147.8 13.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4v2lcrr3",
   "source_url": "https://omlx.ai/benchmarks/performance/4v2lcrr3",
   "tg_tokens": null,
   "tps": 147.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4v2lcrr3-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6595.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 6,595 159.4 13.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4v2lcrr3",
   "source_url": "https://omlx.ai/benchmarks/performance/4v2lcrr3",
   "tg_tokens": null,
   "tps": 159.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4v2lcrr3-ctx131072",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4873.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 4,873 98.0 17.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4v2lcrr3",
   "source_url": "https://omlx.ai/benchmarks/performance/4v2lcrr3",
   "tg_tokens": null,
   "tps": 98.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-fx23y7r4-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 20.0 GB; TTFT 3660 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1119.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k tokens 1,119 PP tok/s 43.0 TG tok/s 3660 TTFT (ms) 20.0 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fx23y7r4",
   "source_url": "https://omlx.ai/benchmarks/performance/fx23y7r4",
   "tg_tokens": null,
   "tps": 43.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-p5mzh8r9-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 20.0 GB; TTFT 1951 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2100.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k tokens 2,100 PP tok/s 69.9 TG tok/s 1951 TTFT (ms) 20.0 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p5mzh8r9",
   "source_url": "https://omlx.ai/benchmarks/performance/p5mzh8r9",
   "tg_tokens": null,
   "tps": 69.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-w4dcbka2-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 19.9 GB; TTFT 1956 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2094.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k tokens 2,094 PP tok/s 67.6 TG tok/s 1956 TTFT (ms) 19.9 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w4dcbka2",
   "source_url": "https://omlx.ai/benchmarks/performance/w4dcbka2",
   "tg_tokens": null,
   "tps": 67.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-y2dseqxn-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2125.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 2,125 70.0 19.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y2dseqxn",
   "source_url": "https://omlx.ai/benchmarks/performance/y2dseqxn",
   "tg_tokens": null,
   "tps": 70.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-y2dseqxn-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2080.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 8k 2,080 67.3 19.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y2dseqxn",
   "source_url": "https://omlx.ai/benchmarks/performance/y2dseqxn",
   "tg_tokens": null,
   "tps": 67.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-wqxa6pwa-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "REAP-19B build; oMLX community benchmark trace; peak memory 12.8 GB at this context; duplicate traces of the same session: cjxrcv44; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 638.2,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit-REAP-19B on M3 Pro (14c); M3 Pro (14c) \u00b7 36 GB \u00b7 4bit; 1k 638.2 41.6 12.8 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wqxa6pwa",
   "source_url": "https://omlx.ai/benchmarks/performance/wqxa6pwa",
   "tg_tokens": null,
   "tps": 41.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-wqxa6pwa-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "REAP-19B build; oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: cjxrcv44; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 666.3,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit-REAP-19B on M3 Pro (14c); M3 Pro (14c) \u00b7 36 GB \u00b7 4bit; 4k 666.3 40.7 13.4 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wqxa6pwa",
   "source_url": "https://omlx.ai/benchmarks/performance/wqxa6pwa",
   "tg_tokens": null,
   "tps": 40.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-tg06nn2n-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "REAP-19B build; oMLX community benchmark trace; peak memory 12.8 GB at this context; duplicate traces of the same session: j9qj86eq; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 632.7,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit-REAP-19B on M3 Pro (14c); M3 Pro (14c) \u00b7 36 GB \u00b7 4bit; 1k 632.7 41.9 12.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tg06nn2n",
   "source_url": "https://omlx.ai/benchmarks/performance/tg06nn2n",
   "tg_tokens": null,
   "tps": 41.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-tg06nn2n-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "REAP-19B build; oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: j9qj86eq; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 666.9,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit-REAP-19B on M3 Pro (14c); M3 Pro (14c) \u00b7 36 GB \u00b7 4bit; 4k 666.9 40.9 13.4 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tg06nn2n",
   "source_url": "https://omlx.ai/benchmarks/performance/tg06nn2n",
   "tg_tokens": null,
   "tps": 40.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-3qxl6c3c-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: bmvm3hwi, seg32h67; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 297.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 297.0 51.6 22.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3qxl6c3c",
   "source_url": "https://omlx.ai/benchmarks/performance/3qxl6c3c",
   "tg_tokens": null,
   "tps": 51.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-3qxl6c3c-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: bmvm3hwi, seg32h67; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 303.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 8k 303.1 45.3 23.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3qxl6c3c",
   "source_url": "https://omlx.ai/benchmarks/performance/3qxl6c3c",
   "tg_tokens": null,
   "tps": 45.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-3qxl6c3c-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: bmvm3hwi, seg32h67; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 225.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 225.9 34.7 21.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3qxl6c3c",
   "source_url": "https://omlx.ai/benchmarks/performance/3qxl6c3c",
   "tg_tokens": null,
   "tps": 34.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-3qxl6c3c-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: bmvm3hwi, seg32h67; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 165.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 165.0 23.3 23.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3qxl6c3c",
   "source_url": "https://omlx.ai/benchmarks/performance/3qxl6c3c",
   "tg_tokens": null,
   "tps": 23.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-3qxl6c3c-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.9 GB at this context; duplicate traces of the same session: bmvm3hwi, seg32h67; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 151.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 64k 151.3 16.4 28.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3qxl6c3c",
   "source_url": "https://omlx.ai/benchmarks/performance/3qxl6c3c",
   "tg_tokens": null,
   "tps": 16.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-3m46j4qq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 0635a4oy, 1t2fy3o4, 8o2dsmqz, suffvu1d, tmmujr7c, u3h5o7w5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 254.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 254.6 46.5 18.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3m46j4qq",
   "source_url": "https://omlx.ai/benchmarks/performance/3m46j4qq",
   "tg_tokens": null,
   "tps": 46.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-3m46j4qq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 0635a4oy, 1t2fy3o4, 8o2dsmqz, suffvu1d, tmmujr7c, u3h5o7w5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 269.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 269.6 55.7 21.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3m46j4qq",
   "source_url": "https://omlx.ai/benchmarks/performance/3m46j4qq",
   "tg_tokens": null,
   "tps": 55.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-3m46j4qq-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: 0635a4oy, 1t2fy3o4, 8o2dsmqz, suffvu1d, tmmujr7c, u3h5o7w5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 264.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 264.8 64.1 22.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3m46j4qq",
   "source_url": "https://omlx.ai/benchmarks/performance/3m46j4qq",
   "tg_tokens": null,
   "tps": 64.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-3m46j4qq-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 0635a4oy, 1t2fy3o4, 8o2dsmqz, suffvu1d, tmmujr7c, u3h5o7w5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 254.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 254.7 55.1 23.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3m46j4qq",
   "source_url": "https://omlx.ai/benchmarks/performance/3m46j4qq",
   "tg_tokens": null,
   "tps": 55.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-3m46j4qq-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: 0635a4oy, 1t2fy3o4, 8o2dsmqz, suffvu1d, tmmujr7c, u3h5o7w5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 235.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 235.9 112.3 25.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3m46j4qq",
   "source_url": "https://omlx.ai/benchmarks/performance/3m46j4qq",
   "tg_tokens": null,
   "tps": 112.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-3m46j4qq-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: 0635a4oy, 1t2fy3o4, 8o2dsmqz, suffvu1d, tmmujr7c, u3h5o7w5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 205.5 46.3 27.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3m46j4qq",
   "source_url": "https://omlx.ai/benchmarks/performance/3m46j4qq",
   "tg_tokens": null,
   "tps": 46.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-3m46j4qq-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.3 GB at this context; duplicate traces of the same session: 0635a4oy, 1t2fy3o4, 8o2dsmqz, suffvu1d, tmmujr7c, u3h5o7w5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 164.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 164.1 42.7 32.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3m46j4qq",
   "source_url": "https://omlx.ai/benchmarks/performance/3m46j4qq",
   "tg_tokens": null,
   "tps": 42.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1w0qgsnv-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 2uabdbs8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 923.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 923.3 65.9 17.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1w0qgsnv",
   "source_url": "https://omlx.ai/benchmarks/performance/1w0qgsnv",
   "tg_tokens": null,
   "tps": 65.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1w0qgsnv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: 2uabdbs8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 965.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 965.6 60.2 18.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1w0qgsnv",
   "source_url": "https://omlx.ai/benchmarks/performance/1w0qgsnv",
   "tg_tokens": null,
   "tps": 60.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xtjfk38p-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: amfw1qn9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 731.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 731.9 41.7 17.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xtjfk38p",
   "source_url": "https://omlx.ai/benchmarks/performance/xtjfk38p",
   "tg_tokens": null,
   "tps": 41.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xtjfk38p-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: amfw1qn9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 965.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 965.0 58.6 18.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xtjfk38p",
   "source_url": "https://omlx.ai/benchmarks/performance/xtjfk38p",
   "tg_tokens": null,
   "tps": 58.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ka48jiz1-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: izv0ef93; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1066.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,066 77.4 21.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ka48jiz1",
   "source_url": "https://omlx.ai/benchmarks/performance/ka48jiz1",
   "tg_tokens": null,
   "tps": 77.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ka48jiz1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: izv0ef93; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1256.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,256 67.0 22.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ka48jiz1",
   "source_url": "https://omlx.ai/benchmarks/performance/ka48jiz1",
   "tg_tokens": null,
   "tps": 67.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-21frpona-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 705.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 705.6 40.4 21.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/21frpona",
   "source_url": "https://omlx.ai/benchmarks/performance/21frpona",
   "tg_tokens": null,
   "tps": 40.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-21frpona-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1225.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,225 73.3 22.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/21frpona",
   "source_url": "https://omlx.ai/benchmarks/performance/21frpona",
   "tg_tokens": null,
   "tps": 73.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-v8uux6un-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: ey22xe4y; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 472.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 472.6 36.7 18.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v8uux6un",
   "source_url": "https://omlx.ai/benchmarks/performance/v8uux6un",
   "tg_tokens": null,
   "tps": 36.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-v8uux6un-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: ey22xe4y; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 428.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 428.0 43.8 19.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v8uux6un",
   "source_url": "https://omlx.ai/benchmarks/performance/v8uux6un",
   "tg_tokens": null,
   "tps": 43.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-v8uux6un-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: ey22xe4y; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 313.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 313.7 23.3 23.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v8uux6un",
   "source_url": "https://omlx.ai/benchmarks/performance/v8uux6un",
   "tg_tokens": null,
   "tps": 23.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-v8uux6un-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.1 GB at this context; duplicate traces of the same session: ey22xe4y; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 189.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 195k 189.8 15.3 34.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v8uux6un",
   "source_url": "https://omlx.ai/benchmarks/performance/v8uux6un",
   "tg_tokens": null,
   "tps": 15.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2n2v1hk0-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.7 GB at this context; duplicate traces of the same session: 03u4uure, 8huupcat; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 378.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 1k 378.5 40.9 39.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2n2v1hk0",
   "source_url": "https://omlx.ai/benchmarks/performance/2n2v1hk0",
   "tg_tokens": null,
   "tps": 40.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2n2v1hk0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.0 GB at this context; duplicate traces of the same session: 03u4uure, 8huupcat; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 449.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 4k 449.5 49.6 43.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2n2v1hk0",
   "source_url": "https://omlx.ai/benchmarks/performance/2n2v1hk0",
   "tg_tokens": null,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2n2v1hk0-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.4 GB at this context; duplicate traces of the same session: 03u4uure, 8huupcat; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 480.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 8k 480.2 53.3 43.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2n2v1hk0",
   "source_url": "https://omlx.ai/benchmarks/performance/2n2v1hk0",
   "tg_tokens": null,
   "tps": 53.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2n2v1hk0-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 44.2 GB at this context; duplicate traces of the same session: 03u4uure, 8huupcat; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 477.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 16k 477.6 58.4 44.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2n2v1hk0",
   "source_url": "https://omlx.ai/benchmarks/performance/2n2v1hk0",
   "tg_tokens": null,
   "tps": 58.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2n2v1hk0-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 45.8 GB at this context; duplicate traces of the same session: 03u4uure, 8huupcat; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 445.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 32k 445.3 94.8 45.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2n2v1hk0",
   "source_url": "https://omlx.ai/benchmarks/performance/2n2v1hk0",
   "tg_tokens": null,
   "tps": 94.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2n2v1hk0-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 49.0 GB at this context; duplicate traces of the same session: 03u4uure, 8huupcat; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 403.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 64k 403.4 53.5 49.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2n2v1hk0",
   "source_url": "https://omlx.ai/benchmarks/performance/2n2v1hk0",
   "tg_tokens": null,
   "tps": 53.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2n2v1hk0-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 53.5 GB at this context; duplicate traces of the same session: 03u4uure, 8huupcat; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 314.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 128k 314.3 50.6 53.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2n2v1hk0",
   "source_url": "https://omlx.ai/benchmarks/performance/2n2v1hk0",
   "tg_tokens": null,
   "tps": 50.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2n2v1hk0-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 58.4 GB at this context; duplicate traces of the same session: 03u4uure, 8huupcat; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 254.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 195k 254.3 34.1 58.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2n2v1hk0",
   "source_url": "https://omlx.ai/benchmarks/performance/2n2v1hk0",
   "tg_tokens": null,
   "tps": 34.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-69lbdjz3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: tn9v7e3y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 819.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 819.2 52.2 33.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/69lbdjz3",
   "source_url": "https://omlx.ai/benchmarks/performance/69lbdjz3",
   "tg_tokens": null,
   "tps": 52.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-69lbdjz3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: tn9v7e3y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 906.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 906.4 42.5 34.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/69lbdjz3",
   "source_url": "https://omlx.ai/benchmarks/performance/69lbdjz3",
   "tg_tokens": null,
   "tps": 42.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-95slsy1j-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: 348dob8r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 822.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 822.6 48.0 33.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/95slsy1j",
   "source_url": "https://omlx.ai/benchmarks/performance/95slsy1j",
   "tg_tokens": null,
   "tps": 48.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-95slsy1j-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: 348dob8r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 858.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 858.9 43.5 34.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/95slsy1j",
   "source_url": "https://omlx.ai/benchmarks/performance/95slsy1j",
   "tg_tokens": null,
   "tps": 43.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-k91e1f9i-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 69.8 GB at this context; duplicate traces of the same session: 47bruugp, ppa41e57, sk9ayolh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 629.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 629.9 68.6 69.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k91e1f9i",
   "source_url": "https://omlx.ai/benchmarks/performance/k91e1f9i",
   "tg_tokens": null,
   "tps": 68.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-k91e1f9i-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 70.3 GB at this context; duplicate traces of the same session: 47bruugp, ppa41e57, sk9ayolh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 781.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 781.5 59.0 70.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k91e1f9i",
   "source_url": "https://omlx.ai/benchmarks/performance/k91e1f9i",
   "tg_tokens": null,
   "tps": 59.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-k91e1f9i-ctx16384",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: 47bruugp, ppa41e57, sk9ayolh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 810.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 810.7 62.3 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k91e1f9i",
   "source_url": "https://omlx.ai/benchmarks/performance/k91e1f9i",
   "tg_tokens": null,
   "tps": 62.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-k91e1f9i-ctx32768",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: 47bruugp, ppa41e57, sk9ayolh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 796.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 796.9 21.4 71.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k91e1f9i",
   "source_url": "https://omlx.ai/benchmarks/performance/k91e1f9i",
   "tg_tokens": null,
   "tps": 21.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5odhahiu-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: 4w6aw6si, haxfug8n, rkzj4w44, sfqb5iwo; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 414.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 414.1 66.1 70.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5odhahiu",
   "source_url": "https://omlx.ai/benchmarks/performance/5odhahiu",
   "tg_tokens": null,
   "tps": 66.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5odhahiu-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: 4w6aw6si, haxfug8n, rkzj4w44, sfqb5iwo; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1030.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 1,030 66.7 71.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5odhahiu",
   "source_url": "https://omlx.ai/benchmarks/performance/5odhahiu",
   "tg_tokens": null,
   "tps": 66.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5odhahiu-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 4w6aw6si, haxfug8n, rkzj4w44, sfqb5iwo; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1028.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 16k 1,028 68.2 71.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5odhahiu",
   "source_url": "https://omlx.ai/benchmarks/performance/5odhahiu",
   "tg_tokens": null,
   "tps": 68.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5odhahiu-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.6 GB at this context; duplicate traces of the same session: 4w6aw6si, haxfug8n, rkzj4w44, sfqb5iwo; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1015.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 64k 1,015 51.8 74.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5odhahiu",
   "source_url": "https://omlx.ai/benchmarks/performance/5odhahiu",
   "tg_tokens": null,
   "tps": 51.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5odhahiu-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.1 GB at this context; duplicate traces of the same session: 4w6aw6si, haxfug8n, rkzj4w44, sfqb5iwo; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 965.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 128k 965.4 50.3 78.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5odhahiu",
   "source_url": "https://omlx.ai/benchmarks/performance/5odhahiu",
   "tg_tokens": null,
   "tps": 50.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-c5z1gc18-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: 69crwh8k, sygyq9bf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 591.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 591.6 66.9 71.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c5z1gc18",
   "source_url": "https://omlx.ai/benchmarks/performance/c5z1gc18",
   "tg_tokens": null,
   "tps": 66.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-c5z1gc18-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.9 GB at this context; duplicate traces of the same session: 69crwh8k, sygyq9bf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 826.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 826.5 81.3 71.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c5z1gc18",
   "source_url": "https://omlx.ai/benchmarks/performance/c5z1gc18",
   "tg_tokens": null,
   "tps": 81.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-c5z1gc18-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.8 GB at this context; duplicate traces of the same session: 69crwh8k, sygyq9bf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 812.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 812.5 58.1 73.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c5z1gc18",
   "source_url": "https://omlx.ai/benchmarks/performance/c5z1gc18",
   "tg_tokens": null,
   "tps": 58.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cam1ild8-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.7 GB; TTFT 1031 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 993.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 993.2 PP tok/s 68.8 TG tok/s 1031 TTFT (ms) 70.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cam1ild8",
   "source_url": "https://omlx.ai/benchmarks/performance/cam1ild8",
   "tg_tokens": null,
   "tps": 68.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1ai8l3vq-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: xcfwsivn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 374.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 374.0 42.0 70.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ai8l3vq",
   "source_url": "https://omlx.ai/benchmarks/performance/1ai8l3vq",
   "tg_tokens": null,
   "tps": 42.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1ai8l3vq-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: xcfwsivn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1096.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,096 36.2 71.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ai8l3vq",
   "source_url": "https://omlx.ai/benchmarks/performance/1ai8l3vq",
   "tg_tokens": null,
   "tps": 36.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-vddcyxwe-ctx1024",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: rhn570zj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 486.9,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 1k 486.9 44.8 11.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vddcyxwe",
   "source_url": "https://omlx.ai/benchmarks/performance/vddcyxwe",
   "tg_tokens": null,
   "tps": 44.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-vddcyxwe-ctx4096",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 14.2 GB at this context; duplicate traces of the same session: rhn570zj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 634.5,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 4k 634.5 41.9 14.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vddcyxwe",
   "source_url": "https://omlx.ai/benchmarks/performance/vddcyxwe",
   "tg_tokens": null,
   "tps": 41.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-15",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-taofqqx1-ctx1024",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 3.0 GB; TTFT 301 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3400.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e2b-it-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 3,400 PP tok/s 180.6 TG tok/s 301 TTFT (ms) 3.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/taofqqx1",
   "source_url": "https://omlx.ai/benchmarks/performance/taofqqx1",
   "tg_tokens": null,
   "tps": 180.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-nvz5woxe-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: bsyfawuk; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 538.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M3 Pro (14c); M3 Pro (14c) \u00b7 36 GB \u00b7 4bit; 1k 538.3 58.7 19.4 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nvz5woxe",
   "source_url": "https://omlx.ai/benchmarks/performance/nvz5woxe",
   "tg_tokens": null,
   "tps": 58.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-nvz5woxe-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: bsyfawuk; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 630.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-4bit on M3 Pro (14c); M3 Pro (14c) \u00b7 36 GB \u00b7 4bit; 4k 630.4 57.5 20.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nvz5woxe",
   "source_url": "https://omlx.ai/benchmarks/performance/nvz5woxe",
   "tg_tokens": null,
   "tps": 57.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-kt6oxinv-ctx4096",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.3 GB at this context; duplicate traces of the same session: q2j7vadj; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 464.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 464.3 42.3 156.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kt6oxinv",
   "source_url": "https://omlx.ai/benchmarks/performance/kt6oxinv",
   "tg_tokens": null,
   "tps": 42.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-kt6oxinv-ctx8192",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.3 GB at this context; duplicate traces of the same session: q2j7vadj; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 460.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 460.2 44.6 156.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kt6oxinv",
   "source_url": "https://omlx.ai/benchmarks/performance/kt6oxinv",
   "tg_tokens": null,
   "tps": 44.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-kt6oxinv-ctx16384",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.4 GB at this context; duplicate traces of the same session: q2j7vadj; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 451.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 451.6 42.9 156.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kt6oxinv",
   "source_url": "https://omlx.ai/benchmarks/performance/kt6oxinv",
   "tg_tokens": null,
   "tps": 42.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-t1f904q7-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 251.3 GB at this context; duplicate traces of the same session: 6b5any2c, aemihtv5, bhy9bwx9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 430.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 1k 430.6 25.4 251.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t1f904q7",
   "source_url": "https://omlx.ai/benchmarks/performance/t1f904q7",
   "tg_tokens": null,
   "tps": 25.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-t1f904q7-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 252.7 GB at this context; duplicate traces of the same session: 6b5any2c, aemihtv5, bhy9bwx9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 548.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 4k 548.6 32.7 252.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t1f904q7",
   "source_url": "https://omlx.ai/benchmarks/performance/t1f904q7",
   "tg_tokens": null,
   "tps": 32.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-t1f904q7-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 253.5 GB at this context; duplicate traces of the same session: 6b5any2c, aemihtv5, bhy9bwx9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 510.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 16k 510.3 60.6 253.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t1f904q7",
   "source_url": "https://omlx.ai/benchmarks/performance/t1f904q7",
   "tg_tokens": null,
   "tps": 60.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-t1f904q7-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 253.6 GB at this context; duplicate traces of the same session: 6b5any2c, aemihtv5, bhy9bwx9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 524.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 32k 524.1 54.9 253.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t1f904q7",
   "source_url": "https://omlx.ai/benchmarks/performance/t1f904q7",
   "tg_tokens": null,
   "tps": 54.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-8cqbgo8m-ctx65536",
   "model": "KAT-Coder-V2.5-Dev",
   "notes": "MTP build; oMLX community benchmark trace, single run at 64k context; peak memory 24.3 GB; TTFT 130022 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 504.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "KAT-Coder-V2.5-Dev-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k tokens 504.0 PP tok/s 46.8 TG tok/s 130022 TTFT (ms) 24.3 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8cqbgo8m",
   "source_url": "https://omlx.ai/benchmarks/performance/8cqbgo8m",
   "tg_tokens": null,
   "tps": 46.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-jols4hg0-ctx65536",
   "model": "KAT-Coder-V2.5-Dev",
   "notes": "MTP build; oMLX community benchmark trace, single run at 64k context; peak memory 22.6 GB; TTFT 120753 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 542.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "KAT-Coder-V2.5-Dev-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k tokens 542.7 PP tok/s 51.4 TG tok/s 120753 TTFT (ms) 22.6 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jols4hg0",
   "source_url": "https://omlx.ai/benchmarks/performance/jols4hg0",
   "tg_tokens": null,
   "tps": 51.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-xszqmk9v-ctx131072",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 128k context; peak memory 16.0 GB; TTFT 621217 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 211.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 128k tokens 211.0 PP tok/s 9.3 TG tok/s 621217 TTFT (ms) 16.0 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xszqmk9v",
   "source_url": "https://omlx.ai/benchmarks/performance/xszqmk9v",
   "tg_tokens": null,
   "tps": 9.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-h0db7f4w-ctx1024",
   "model": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 48.0 GB at this context; duplicate traces of the same session: pm6hdxhj; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 238.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B-oQ4e-fp16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 238.2 27.9 48.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h0db7f4w",
   "source_url": "https://omlx.ai/benchmarks/performance/h0db7f4w",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-h0db7f4w-ctx4096",
   "model": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 50.2 GB at this context; duplicate traces of the same session: pm6hdxhj; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 248.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B-oQ4e-fp16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 248.1 27.6 50.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h0db7f4w",
   "source_url": "https://omlx.ai/benchmarks/performance/h0db7f4w",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-vjibblxg-ctx1024",
   "model": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 48.0 GB at this context; duplicate traces of the same session: 2cw99dc3, 72kkxh73, doe33koa, k8yzioni, rkib57cz, tcpixgp5; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 241.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B-oQ4e-fp16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 241.7 27.2 48.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vjibblxg",
   "source_url": "https://omlx.ai/benchmarks/performance/vjibblxg",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-vjibblxg-ctx4096",
   "model": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 50.2 GB at this context; duplicate traces of the same session: 2cw99dc3, 72kkxh73, doe33koa, k8yzioni, rkib57cz, tcpixgp5; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 249.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B-oQ4e-fp16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 249.4 27.3 50.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vjibblxg",
   "source_url": "https://omlx.ai/benchmarks/performance/vjibblxg",
   "tg_tokens": null,
   "tps": 27.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-vjibblxg-ctx8192",
   "model": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 50.3 GB at this context; duplicate traces of the same session: 2cw99dc3, 72kkxh73, doe33koa, k8yzioni, rkib57cz, tcpixgp5; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 247.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B-oQ4e-fp16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 8k 247.8 26.2 50.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vjibblxg",
   "source_url": "https://omlx.ai/benchmarks/performance/vjibblxg",
   "tg_tokens": null,
   "tps": 26.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-vjibblxg-ctx16384",
   "model": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 50.3 GB at this context; duplicate traces of the same session: 2cw99dc3, 72kkxh73, doe33koa, k8yzioni, rkib57cz, tcpixgp5; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 243.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B-oQ4e-fp16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k 243.2 25.3 50.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vjibblxg",
   "source_url": "https://omlx.ai/benchmarks/performance/vjibblxg",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-vjibblxg-ctx32768",
   "model": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 50.5 GB at this context; duplicate traces of the same session: 2cw99dc3, 72kkxh73, doe33koa, k8yzioni, rkib57cz, tcpixgp5; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 231.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B-oQ4e-fp16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 32k 231.7 23.9 50.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vjibblxg",
   "source_url": "https://omlx.ai/benchmarks/performance/vjibblxg",
   "tg_tokens": null,
   "tps": 23.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-vjibblxg-ctx65536",
   "model": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 50.7 GB at this context; duplicate traces of the same session: 2cw99dc3, 72kkxh73, doe33koa, k8yzioni, rkib57cz, tcpixgp5; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 214.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B-oQ4e-fp16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 64k 214.1 21.3 50.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vjibblxg",
   "source_url": "https://omlx.ai/benchmarks/performance/vjibblxg",
   "tg_tokens": null,
   "tps": 21.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-vjibblxg-ctx131072",
   "model": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 51.2 GB at this context; duplicate traces of the same session: 2cw99dc3, 72kkxh73, doe33koa, k8yzioni, rkib57cz, tcpixgp5; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 184.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B-oQ4e-fp16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 128k 184.8 17.9 51.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vjibblxg",
   "source_url": "https://omlx.ai/benchmarks/performance/vjibblxg",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-kq0o4ewy-ctx4096",
   "model": "NVIDIA-Nemotron-3.5-Super-VL-120B-A12B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 72.9 GB at this context; duplicate traces of the same session: wsla23wr; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 328.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-MLX-4bit on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 4k 328.5 41.7 72.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kq0o4ewy",
   "source_url": "https://omlx.ai/benchmarks/performance/kq0o4ewy",
   "tg_tokens": null,
   "tps": 41.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-kq0o4ewy-ctx16384",
   "model": "NVIDIA-Nemotron-3.5-Super-VL-120B-A12B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 72.9 GB at this context; duplicate traces of the same session: wsla23wr; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 343.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-MLX-4bit on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 16k 343.2 40.1 72.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kq0o4ewy",
   "source_url": "https://omlx.ai/benchmarks/performance/kq0o4ewy",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-unzup0mb-ctx1024",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: z04mu0at; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 611.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 611.8 28.1 16.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/unzup0mb",
   "source_url": "https://omlx.ai/benchmarks/performance/unzup0mb",
   "tg_tokens": null,
   "tps": 28.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-unzup0mb-ctx4096",
   "model": "Qwen3.5-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: z04mu0at; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 942.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 942.6 32.0 17.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/unzup0mb",
   "source_url": "https://omlx.ai/benchmarks/performance/unzup0mb",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hzsv2x40-ctx1024",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 9dqls3dn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 738.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 738.4 99.3 19.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hzsv2x40",
   "source_url": "https://omlx.ai/benchmarks/performance/hzsv2x40",
   "tg_tokens": null,
   "tps": 99.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hzsv2x40-ctx4096",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 9dqls3dn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2675.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,675 142.5 20.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hzsv2x40",
   "source_url": "https://omlx.ai/benchmarks/performance/hzsv2x40",
   "tg_tokens": null,
   "tps": 142.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-76nskzap-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 24.8 GB; TTFT 33716 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 971.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k tokens 971.9 PP tok/s 133.8 TG tok/s 33716 TTFT (ms) 24.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/76nskzap",
   "source_url": "https://omlx.ai/benchmarks/performance/76nskzap",
   "tg_tokens": null,
   "tps": 133.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-1ri8d927-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 24.5 GB; TTFT 34791 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 941.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k tokens 941.9 PP tok/s 129.8 TG tok/s 34791 TTFT (ms) 24.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ri8d927",
   "source_url": "https://omlx.ai/benchmarks/performance/1ri8d927",
   "tg_tokens": null,
   "tps": 129.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-h8uova9c-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 24.5 GB; TTFT 35031 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 935.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k tokens 935.4 PP tok/s 131.6 TG tok/s 35031 TTFT (ms) 24.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h8uova9c",
   "source_url": "https://omlx.ai/benchmarks/performance/h8uova9c",
   "tg_tokens": null,
   "tps": 131.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-vplb1d7c-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 24.8 GB; TTFT 32122 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1020.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k tokens 1,020 PP tok/s 100.7 TG tok/s 32122 TTFT (ms) 24.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vplb1d7c",
   "source_url": "https://omlx.ai/benchmarks/performance/vplb1d7c",
   "tg_tokens": null,
   "tps": 100.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-qk1b3qko-ctx1024",
   "model": "Qwen3.8-27B-Cold-Fusion-GAIN-V1.1",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: f012m7r5, g88105mc, j9x0yrcb, khl5cjor, kkxpor1v, klii1xoo; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 405.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 405.1 53.8 24.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qk1b3qko",
   "source_url": "https://omlx.ai/benchmarks/performance/qk1b3qko",
   "tg_tokens": null,
   "tps": 53.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-qk1b3qko-ctx4096",
   "model": "Qwen3.8-27B-Cold-Fusion-GAIN-V1.1",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: f012m7r5, g88105mc, j9x0yrcb, khl5cjor, kkxpor1v, klii1xoo; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 467.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 467.9 69.2 27.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qk1b3qko",
   "source_url": "https://omlx.ai/benchmarks/performance/qk1b3qko",
   "tg_tokens": null,
   "tps": 69.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-qk1b3qko-ctx8192",
   "model": "Qwen3.8-27B-Cold-Fusion-GAIN-V1.1",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 28.3 GB at this context; duplicate traces of the same session: f012m7r5, g88105mc, j9x0yrcb, khl5cjor, kkxpor1v, klii1xoo; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 495.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 495.5 57.2 28.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qk1b3qko",
   "source_url": "https://omlx.ai/benchmarks/performance/qk1b3qko",
   "tg_tokens": null,
   "tps": 57.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-qk1b3qko-ctx16384",
   "model": "Qwen3.8-27B-Cold-Fusion-GAIN-V1.1",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 29.1 GB at this context; duplicate traces of the same session: f012m7r5, g88105mc, j9x0yrcb, khl5cjor, kkxpor1v, klii1xoo; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 485.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 485.4 59.6 29.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qk1b3qko",
   "source_url": "https://omlx.ai/benchmarks/performance/qk1b3qko",
   "tg_tokens": null,
   "tps": 59.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-qk1b3qko-ctx32768",
   "model": "Qwen3.8-27B-Cold-Fusion-GAIN-V1.1",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 30.7 GB at this context; duplicate traces of the same session: f012m7r5, g88105mc, j9x0yrcb, khl5cjor, kkxpor1v, klii1xoo; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 442.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 442.2 163.0 30.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qk1b3qko",
   "source_url": "https://omlx.ai/benchmarks/performance/qk1b3qko",
   "tg_tokens": null,
   "tps": 163.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-qk1b3qko-ctx65536",
   "model": "Qwen3.8-27B-Cold-Fusion-GAIN-V1.1",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: f012m7r5, g88105mc, j9x0yrcb, khl5cjor, kkxpor1v, klii1xoo; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 375.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 375.7 78.8 33.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qk1b3qko",
   "source_url": "https://omlx.ai/benchmarks/performance/qk1b3qko",
   "tg_tokens": null,
   "tps": 78.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-qk1b3qko-ctx131072",
   "model": "Qwen3.8-27B-Cold-Fusion-GAIN-V1.1",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 38.4 GB at this context; duplicate traces of the same session: f012m7r5, g88105mc, j9x0yrcb, khl5cjor, kkxpor1v, klii1xoo; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 291.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 291.2 71.4 38.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qk1b3qko",
   "source_url": "https://omlx.ai/benchmarks/performance/qk1b3qko",
   "tg_tokens": null,
   "tps": 71.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-qk1b3qko-ctx199680",
   "model": "Qwen3.8-27B-Cold-Fusion-GAIN-V1.1",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 43.3 GB at this context; duplicate traces of the same session: f012m7r5, g88105mc, j9x0yrcb, khl5cjor, kkxpor1v, klii1xoo; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 235.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 235.5 48.9 43.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qk1b3qko",
   "source_url": "https://omlx.ai/benchmarks/performance/qk1b3qko",
   "tg_tokens": null,
   "tps": 48.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ulb2p9qz-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 3fla4syv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 253.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 253.9 27.1 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ulb2p9qz",
   "source_url": "https://omlx.ai/benchmarks/performance/ulb2p9qz",
   "tg_tokens": null,
   "tps": 27.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ulb2p9qz-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 3fla4syv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 268.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 268.8 28.4 19.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ulb2p9qz",
   "source_url": "https://omlx.ai/benchmarks/performance/ulb2p9qz",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-ofg5n7xc-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 8pjqo6mh; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 212.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 1k 212.9 30.1 20.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ofg5n7xc",
   "source_url": "https://omlx.ai/benchmarks/performance/ofg5n7xc",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-ofg5n7xc-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: 8pjqo6mh; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 265.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 4k 265.4 27.7 24.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ofg5n7xc",
   "source_url": "https://omlx.ai/benchmarks/performance/ofg5n7xc",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "outlier"
   ],
   "hardware": "M3 Max",
   "id": "omlx-8t2dkfsk-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 25.5 GB; TTFT 193991 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 168.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k tokens 168.9 PP tok/s 70.4 TG tok/s 193991 TTFT (ms) 25.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8t2dkfsk",
   "source_url": "https://omlx.ai/benchmarks/performance/8t2dkfsk",
   "tg_tokens": null,
   "tps": 70.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-lo9jc4hv-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: a86qr4o8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 183.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 183.5 26.3 21.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lo9jc4hv",
   "source_url": "https://omlx.ai/benchmarks/performance/lo9jc4hv",
   "tg_tokens": null,
   "tps": 26.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-lo9jc4hv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: a86qr4o8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 250.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 250.1 34.8 24.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lo9jc4hv",
   "source_url": "https://omlx.ai/benchmarks/performance/lo9jc4hv",
   "tg_tokens": null,
   "tps": 34.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "outlier"
   ],
   "hardware": "M3 Max",
   "id": "omlx-bc6mif5e-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 28.2 GB; TTFT 154034 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 212.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k tokens 212.7 PP tok/s 67.8 TG tok/s 154034 TTFT (ms) 28.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bc6mif5e",
   "source_url": "https://omlx.ai/benchmarks/performance/bc6mif5e",
   "tg_tokens": null,
   "tps": 67.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qapf5vb3-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 22.6 GB; TTFT 184376 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 355.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k tokens 355.4 PP tok/s 21.6 TG tok/s 184376 TTFT (ms) 22.6 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qapf5vb3",
   "source_url": "https://omlx.ai/benchmarks/performance/qapf5vb3",
   "tg_tokens": null,
   "tps": 21.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-9edm4nq7-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 19.6 GB; TTFT 36727 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 446.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k tokens 446.1 PP tok/s 40.5 TG tok/s 36727 TTFT (ms) 19.6 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9edm4nq7",
   "source_url": "https://omlx.ai/benchmarks/performance/9edm4nq7",
   "tg_tokens": null,
   "tps": 40.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-9js2yb2f-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 19.6 GB; TTFT 35642 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 459.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k tokens 459.7 PP tok/s 32.1 TG tok/s 35642 TTFT (ms) 19.6 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9js2yb2f",
   "source_url": "https://omlx.ai/benchmarks/performance/9js2yb2f",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-lf3zyf7y-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: 5w3ble1l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 215.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 215.9 35.9 26.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lf3zyf7y",
   "source_url": "https://omlx.ai/benchmarks/performance/lf3zyf7y",
   "tg_tokens": null,
   "tps": 35.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-lf3zyf7y-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.6 GB at this context; duplicate traces of the same session: 5w3ble1l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 197.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 197.2 25.1 30.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lf3zyf7y",
   "source_url": "https://omlx.ai/benchmarks/performance/lf3zyf7y",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2rg58iwl-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.7 GB at this context; duplicate traces of the same session: and0zyf6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 117.7,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 2bit; 1k 117.7 23.0 37.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2rg58iwl",
   "source_url": "https://omlx.ai/benchmarks/performance/2rg58iwl",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2rg58iwl-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 40.0 GB at this context; duplicate traces of the same session: and0zyf6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 311.5,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 2bit; 4k 311.5 26.2 40.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2rg58iwl",
   "source_url": "https://omlx.ai/benchmarks/performance/2rg58iwl",
   "tg_tokens": null,
   "tps": 26.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-b95c19u5-ctx1024",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: mm1q7l5u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 218.6,
   "provenance": "community",
   "quant": "oQ2",
   "quote": "Ternary-Bonsai-2-27B-MLX-oQ2-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 2bit; 1k 218.6 33.7 13.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b95c19u5",
   "source_url": "https://omlx.ai/benchmarks/performance/b95c19u5",
   "tg_tokens": null,
   "tps": 33.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-b95c19u5-ctx4096",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: mm1q7l5u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 191.3,
   "provenance": "community",
   "quant": "oQ2",
   "quote": "Ternary-Bonsai-2-27B-MLX-oQ2-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 2bit; 4k 191.3 23.2 17.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b95c19u5",
   "source_url": "https://omlx.ai/benchmarks/performance/b95c19u5",
   "tg_tokens": null,
   "tps": 23.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-noa23yos-ctx4096",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.3 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 464.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 464.3 42.3 156.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/noa23yos",
   "source_url": "https://omlx.ai/benchmarks/performance/noa23yos",
   "tg_tokens": null,
   "tps": 42.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-noa23yos-ctx8192",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.3 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 460.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 460.2 44.6 156.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/noa23yos",
   "source_url": "https://omlx.ai/benchmarks/performance/noa23yos",
   "tg_tokens": null,
   "tps": 44.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-noa23yos-ctx16384",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.4 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 451.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 451.6 42.9 156.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/noa23yos",
   "source_url": "https://omlx.ai/benchmarks/performance/noa23yos",
   "tg_tokens": null,
   "tps": 42.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-cp2n68wl-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 107.3 GB; TTFT 5964 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 171.7,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "GLM-5.3-Flash-oQ2e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 1k tokens 171.7 PP tok/s 40.7 TG tok/s 5964 TTFT (ms) 107.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cp2n68wl",
   "source_url": "https://omlx.ai/benchmarks/performance/cp2n68wl",
   "tg_tokens": null,
   "tps": 40.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-53zkrz9y-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.0 GB at this context; duplicate traces of the same session: 9x22qkdp, nu0x2sx1, vafc1l0b; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 419.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 419.1 40.3 179.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/53zkrz9y",
   "source_url": "https://omlx.ai/benchmarks/performance/53zkrz9y",
   "tg_tokens": null,
   "tps": 40.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-53zkrz9y-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.1 GB at this context; duplicate traces of the same session: 9x22qkdp, nu0x2sx1, vafc1l0b; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 411.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 411.0 47.1 179.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/53zkrz9y",
   "source_url": "https://omlx.ai/benchmarks/performance/53zkrz9y",
   "tg_tokens": null,
   "tps": 47.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-53zkrz9y-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.3 GB at this context; duplicate traces of the same session: 9x22qkdp, nu0x2sx1, vafc1l0b; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 409.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 409.2 85.6 179.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/53zkrz9y",
   "source_url": "https://omlx.ai/benchmarks/performance/53zkrz9y",
   "tg_tokens": null,
   "tps": 85.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-53zkrz9y-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.4 GB at this context; duplicate traces of the same session: 9x22qkdp, nu0x2sx1, vafc1l0b; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 404.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 404.9 58.9 179.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/53zkrz9y",
   "source_url": "https://omlx.ai/benchmarks/performance/53zkrz9y",
   "tg_tokens": null,
   "tps": 58.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-8l6x1ivz-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.1 GB at this context; duplicate traces of the same session: vl6wcgnq; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 575.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 575.5 55.7 13.1 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8l6x1ivz",
   "source_url": "https://omlx.ai/benchmarks/performance/8l6x1ivz",
   "tg_tokens": null,
   "tps": 55.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-8l6x1ivz-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.5 GB at this context; duplicate traces of the same session: vl6wcgnq; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 628.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 628.0 54.6 13.5 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8l6x1ivz",
   "source_url": "https://omlx.ai/benchmarks/performance/8l6x1ivz",
   "tg_tokens": null,
   "tps": 54.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-hej9o16c-ctx65536",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 14.7 GB at this context; duplicate traces of the same session: 9o4gf545; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 274.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 64k 274.8 23.3 14.7 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hej9o16c",
   "source_url": "https://omlx.ai/benchmarks/performance/hej9o16c",
   "tg_tokens": null,
   "tps": 23.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-hej9o16c-ctx131072",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: 9o4gf545; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 206.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 128k 206.6 15.2 15.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hej9o16c",
   "source_url": "https://omlx.ai/benchmarks/performance/hej9o16c",
   "tg_tokens": null,
   "tps": 15.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-ui1o5y2g-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.1 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 593.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 593.1 59.9 13.1 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ui1o5y2g",
   "source_url": "https://omlx.ai/benchmarks/performance/ui1o5y2g",
   "tg_tokens": null,
   "tps": 59.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-ui1o5y2g-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.5 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 635.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 635.2 57.4 13.5 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ui1o5y2g",
   "source_url": "https://omlx.ai/benchmarks/performance/ui1o5y2g",
   "tg_tokens": null,
   "tps": 57.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-ui1o5y2g-ctx65536",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 14.7 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 429.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 64k 429.5 29.9 14.7 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ui1o5y2g",
   "source_url": "https://omlx.ai/benchmarks/performance/ui1o5y2g",
   "tg_tokens": null,
   "tps": 29.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-zz2f6txi-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: czj2vdde; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 959.7,
   "provenance": "community",
   "quant": "5bit",
   "quote": "LFM2.5-24B-A2B-MLX-5bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 5bit; 1k 959.7 97.7 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zz2f6txi",
   "source_url": "https://omlx.ai/benchmarks/performance/zz2f6txi",
   "tg_tokens": null,
   "tps": 97.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-zz2f6txi-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: czj2vdde; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1017.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "LFM2.5-24B-A2B-MLX-5bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 5bit; 4k 1,017 89.5 16.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zz2f6txi",
   "source_url": "https://omlx.ai/benchmarks/performance/zz2f6txi",
   "tg_tokens": null,
   "tps": 89.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-2yjbywrs-ctx1024",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 768.2,
   "provenance": "community",
   "quant": "5bit",
   "quote": "MiniCPM-V-4.7-35B-A3B-5bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 5bit; 1k 768.2 48.2 26.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2yjbywrs",
   "source_url": "https://omlx.ai/benchmarks/performance/2yjbywrs",
   "tg_tokens": null,
   "tps": 48.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-2yjbywrs-ctx4096",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 841.2,
   "provenance": "community",
   "quant": "5bit",
   "quote": "MiniCPM-V-4.7-35B-A3B-5bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 5bit; 4k 841.2 47.9 27.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2yjbywrs",
   "source_url": "https://omlx.ai/benchmarks/performance/2yjbywrs",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-k8mvpdqd-ctx199680",
   "model": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B",
   "notes": "fp16 build; oMLX community benchmark trace, single run at 195k context; peak memory 51.7 GB; TTFT 1257147 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 159.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B-oQ4e-fp16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 195k tokens 159.1 PP tok/s 15.2 TG tok/s 1257147 TTFT (ms) 51.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k8mvpdqd",
   "source_url": "https://omlx.ai/benchmarks/performance/k8mvpdqd",
   "tg_tokens": null,
   "tps": 15.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-dq6ebkn6-ctx4096",
   "model": "Ornith-1.0-35B",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 2cf4bzlc, b3sftsw7; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2705.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.0-35B-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 2,705 92.4 21.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dq6ebkn6",
   "source_url": "https://omlx.ai/benchmarks/performance/dq6ebkn6",
   "tg_tokens": null,
   "tps": 92.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-dq6ebkn6-ctx8192",
   "model": "Ornith-1.0-35B",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 2cf4bzlc, b3sftsw7; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2546.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.0-35B-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 2,546 88.5 21.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dq6ebkn6",
   "source_url": "https://omlx.ai/benchmarks/performance/dq6ebkn6",
   "tg_tokens": null,
   "tps": 88.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-dq6ebkn6-ctx16384",
   "model": "Ornith-1.0-35B",
   "notes": "oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 2cf4bzlc, b3sftsw7; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2155.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.0-35B-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 2,155 92.0 22.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dq6ebkn6",
   "source_url": "https://omlx.ai/benchmarks/performance/dq6ebkn6",
   "tg_tokens": null,
   "tps": 92.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-27dpaagq-ctx4096",
   "model": "Ornith-1.0-35B",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2729.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.0-35B-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 2,729 91.6 21.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/27dpaagq",
   "source_url": "https://omlx.ai/benchmarks/performance/27dpaagq",
   "tg_tokens": null,
   "tps": 91.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-27dpaagq-ctx8192",
   "model": "Ornith-1.0-35B",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2603.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.0-35B-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 2,603 91.8 21.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/27dpaagq",
   "source_url": "https://omlx.ai/benchmarks/performance/27dpaagq",
   "tg_tokens": null,
   "tps": 91.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-27dpaagq-ctx16384",
   "model": "Ornith-1.0-35B",
   "notes": "oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2333.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.0-35B-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 2,333 87.0 22.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/27dpaagq",
   "source_url": "https://omlx.ai/benchmarks/performance/27dpaagq",
   "tg_tokens": null,
   "tps": 87.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ucdswfpc-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 8ddpb6ok, z3pbx4s0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 771.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 771.6 86.9 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ucdswfpc",
   "source_url": "https://omlx.ai/benchmarks/performance/ucdswfpc",
   "tg_tokens": null,
   "tps": 86.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ucdswfpc-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: 8ddpb6ok, z3pbx4s0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 923.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 923.6 78.1 21.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ucdswfpc",
   "source_url": "https://omlx.ai/benchmarks/performance/ucdswfpc",
   "tg_tokens": null,
   "tps": 78.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ucdswfpc-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: 8ddpb6ok, z3pbx4s0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 594.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 32k 594.2 41.3 22.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ucdswfpc",
   "source_url": "https://omlx.ai/benchmarks/performance/ucdswfpc",
   "tg_tokens": null,
   "tps": 41.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-9k21ybtl-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: prd9oq21; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 772.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 772.4 83.1 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9k21ybtl",
   "source_url": "https://omlx.ai/benchmarks/performance/9k21ybtl",
   "tg_tokens": null,
   "tps": 83.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-9k21ybtl-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: prd9oq21; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 903.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 903.1 78.9 21.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9k21ybtl",
   "source_url": "https://omlx.ai/benchmarks/performance/9k21ybtl",
   "tg_tokens": null,
   "tps": 78.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-98byctii-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 9.6 GB at this context; duplicate traces of the same session: 1e3xep7y, 3szkcp8g, 578xblos, eo2nhqa2, l51ic6h7, m8ghoum1, tpyazfh9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2115.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-9B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 2,115 109.6 9.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/98byctii",
   "source_url": "https://omlx.ai/benchmarks/performance/98byctii",
   "tg_tokens": null,
   "tps": 109.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-98byctii-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 10.2 GB at this context; duplicate traces of the same session: 1e3xep7y, 3szkcp8g, 578xblos, eo2nhqa2, l51ic6h7, m8ghoum1, tpyazfh9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3037.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-9B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 3,037 96.7 10.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/98byctii",
   "source_url": "https://omlx.ai/benchmarks/performance/98byctii",
   "tg_tokens": null,
   "tps": 96.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-98byctii-ctx8192",
   "model": "Ornith-1.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 10.4 GB at this context; duplicate traces of the same session: 1e3xep7y, 3szkcp8g, 578xblos, eo2nhqa2, l51ic6h7, m8ghoum1, tpyazfh9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3119.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-9B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 3,119 115.8 10.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/98byctii",
   "source_url": "https://omlx.ai/benchmarks/performance/98byctii",
   "tg_tokens": null,
   "tps": 115.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-98byctii-ctx16384",
   "model": "Ornith-1.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 10.9 GB at this context; duplicate traces of the same session: 1e3xep7y, 3szkcp8g, 578xblos, eo2nhqa2, l51ic6h7, m8ghoum1, tpyazfh9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2782.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-9B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 2,782 175.5 10.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/98byctii",
   "source_url": "https://omlx.ai/benchmarks/performance/98byctii",
   "tg_tokens": null,
   "tps": 175.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-98byctii-ctx32768",
   "model": "Ornith-1.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 11.9 GB at this context; duplicate traces of the same session: 1e3xep7y, 3szkcp8g, 578xblos, eo2nhqa2, l51ic6h7, m8ghoum1, tpyazfh9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2526.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-9B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 2,526 171.7 11.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/98byctii",
   "source_url": "https://omlx.ai/benchmarks/performance/98byctii",
   "tg_tokens": null,
   "tps": 171.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-98byctii-ctx65536",
   "model": "Ornith-1.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 14.0 GB at this context; duplicate traces of the same session: 1e3xep7y, 3szkcp8g, 578xblos, eo2nhqa2, l51ic6h7, m8ghoum1, tpyazfh9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2031.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-9B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k 2,031 82.6 14.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/98byctii",
   "source_url": "https://omlx.ai/benchmarks/performance/98byctii",
   "tg_tokens": null,
   "tps": 82.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-98byctii-ctx131072",
   "model": "Ornith-1.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.2 GB at this context; duplicate traces of the same session: 1e3xep7y, 3szkcp8g, 578xblos, eo2nhqa2, l51ic6h7, m8ghoum1, tpyazfh9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1628.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-9B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 128k 1,628 66.7 18.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/98byctii",
   "source_url": "https://omlx.ai/benchmarks/performance/98byctii",
   "tg_tokens": null,
   "tps": 66.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-98byctii-ctx199680",
   "model": "Ornith-1.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: 1e3xep7y, 3szkcp8g, 578xblos, eo2nhqa2, l51ic6h7, m8ghoum1, tpyazfh9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1211.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Ornith-1.5-9B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 195k 1,211 43.4 22.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/98byctii",
   "source_url": "https://omlx.ai/benchmarks/performance/98byctii",
   "tg_tokens": null,
   "tps": 43.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-1d7vek2q-ctx4096",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 13.5 GB at this context; duplicate traces of the same session: 8qpk60ji; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 514.4,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-3bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 3bit; 4k 514.4 48.1 13.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1d7vek2q",
   "source_url": "https://omlx.ai/benchmarks/performance/1d7vek2q",
   "tg_tokens": null,
   "tps": 48.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-1d7vek2q-ctx16384",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: 8qpk60ji; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 320.3,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-3bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 3bit; 16k 320.3 31.2 14.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1d7vek2q",
   "source_url": "https://omlx.ai/benchmarks/performance/1d7vek2q",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-23mjn3ee-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 23.7 GB; TTFT 10130 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 808.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 8k tokens 808.7 PP tok/s 72.2 TG tok/s 10130 TTFT (ms) 23.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/23mjn3ee",
   "source_url": "https://omlx.ai/benchmarks/performance/23mjn3ee",
   "tg_tokens": null,
   "tps": 72.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-p6fwd61j-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 21.4 GB; TTFT 1357 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 754.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k tokens 754.8 PP tok/s 90.7 TG tok/s 1357 TTFT (ms) 21.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p6fwd61j",
   "source_url": "https://omlx.ai/benchmarks/performance/p6fwd61j",
   "tg_tokens": null,
   "tps": 90.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-2mqpyq6p-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 21.5 GB; TTFT 1322 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 774.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k tokens 774.9 PP tok/s 86.2 TG tok/s 1322 TTFT (ms) 21.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2mqpyq6p",
   "source_url": "https://omlx.ai/benchmarks/performance/2mqpyq6p",
   "tg_tokens": null,
   "tps": 86.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-dzwndl2e-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 21.4 GB; TTFT 1364 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 750.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k tokens 750.7 PP tok/s 40.4 TG tok/s 1364 TTFT (ms) 21.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dzwndl2e",
   "source_url": "https://omlx.ai/benchmarks/performance/dzwndl2e",
   "tg_tokens": null,
   "tps": 40.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cfa27ekp-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.3 GB at this context; duplicate traces of the same session: hx8pxx6b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 723.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 723.6 18.9 28.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cfa27ekp",
   "source_url": "https://omlx.ai/benchmarks/performance/cfa27ekp",
   "tg_tokens": null,
   "tps": 18.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cfa27ekp-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.8 GB at this context; duplicate traces of the same session: hx8pxx6b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 781.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 781.8 18.6 29.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cfa27ekp",
   "source_url": "https://omlx.ai/benchmarks/performance/cfa27ekp",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-geoqej49-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 22.9 GB; TTFT 876805 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 74.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 64k tokens 74.7 PP tok/s 14.4 TG tok/s 876805 TTFT (ms) 22.9 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/geoqej49",
   "source_url": "https://omlx.ai/benchmarks/performance/geoqej49",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-havpfucv-ctx1024",
   "model": "Qwen3.8-27B-Cold-Fusion-GAIN-V1.1",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 405.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 405.1 53.8 24.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/havpfucv",
   "source_url": "https://omlx.ai/benchmarks/performance/havpfucv",
   "tg_tokens": null,
   "tps": 53.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-havpfucv-ctx4096",
   "model": "Qwen3.8-27B-Cold-Fusion-GAIN-V1.1",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 467.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 467.9 69.2 27.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/havpfucv",
   "source_url": "https://omlx.ai/benchmarks/performance/havpfucv",
   "tg_tokens": null,
   "tps": 69.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-havpfucv-ctx8192",
   "model": "Qwen3.8-27B-Cold-Fusion-GAIN-V1.1",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 28.3 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 495.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 495.5 57.2 28.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/havpfucv",
   "source_url": "https://omlx.ai/benchmarks/performance/havpfucv",
   "tg_tokens": null,
   "tps": 57.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-havpfucv-ctx16384",
   "model": "Qwen3.8-27B-Cold-Fusion-GAIN-V1.1",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 29.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 485.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 485.4 59.6 29.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/havpfucv",
   "source_url": "https://omlx.ai/benchmarks/performance/havpfucv",
   "tg_tokens": null,
   "tps": 59.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-havpfucv-ctx32768",
   "model": "Qwen3.8-27B-Cold-Fusion-GAIN-V1.1",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 30.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 442.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 442.2 163.0 30.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/havpfucv",
   "source_url": "https://omlx.ai/benchmarks/performance/havpfucv",
   "tg_tokens": null,
   "tps": 163.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-havpfucv-ctx65536",
   "model": "Qwen3.8-27B-Cold-Fusion-GAIN-V1.1",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 375.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 375.7 78.8 33.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/havpfucv",
   "source_url": "https://omlx.ai/benchmarks/performance/havpfucv",
   "tg_tokens": null,
   "tps": 78.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-havpfucv-ctx131072",
   "model": "Qwen3.8-27B-Cold-Fusion-GAIN-V1.1",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 38.4 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 291.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 291.2 71.4 38.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/havpfucv",
   "source_url": "https://omlx.ai/benchmarks/performance/havpfucv",
   "tg_tokens": null,
   "tps": 71.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-havpfucv-ctx199680",
   "model": "Qwen3.8-27B-Cold-Fusion-GAIN-V1.1",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 43.3 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 235.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-oQ4e-fp16-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 235.5 48.9 43.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/havpfucv",
   "source_url": "https://omlx.ai/benchmarks/performance/havpfucv",
   "tg_tokens": null,
   "tps": 48.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-xx5e2nmz-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.2 GB at this context; duplicate traces of the same session: 4era0f9b, y4z9y0ir, y9xqkxi4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 258.1,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 258.1 31.1 15.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xx5e2nmz",
   "source_url": "https://omlx.ai/benchmarks/performance/xx5e2nmz",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-xx5e2nmz-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 4era0f9b, y4z9y0ir, y9xqkxi4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 258.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 258.0 29.7 18.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xx5e2nmz",
   "source_url": "https://omlx.ai/benchmarks/performance/xx5e2nmz",
   "tg_tokens": null,
   "tps": 29.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-xx5e2nmz-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: 4era0f9b, y4z9y0ir, y9xqkxi4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 233.9,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 233.9 29.0 18.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xx5e2nmz",
   "source_url": "https://omlx.ai/benchmarks/performance/xx5e2nmz",
   "tg_tokens": null,
   "tps": 29.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-xx5e2nmz-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 4era0f9b, y4z9y0ir, y9xqkxi4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 226.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 226.0 26.6 19.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xx5e2nmz",
   "source_url": "https://omlx.ai/benchmarks/performance/xx5e2nmz",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-e4sot3p2-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 20.2 GB; TTFT 4611 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 222.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 222.1 PP tok/s 23.7 TG tok/s 4611 TTFT (ms) 20.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e4sot3p2",
   "source_url": "https://omlx.ai/benchmarks/performance/e4sot3p2",
   "tg_tokens": null,
   "tps": 23.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-5sa7ytv6-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 18.5 GB; TTFT 4777 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 214.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 214.4 PP tok/s 39.2 TG tok/s 4777 TTFT (ms) 18.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5sa7ytv6",
   "source_url": "https://omlx.ai/benchmarks/performance/5sa7ytv6",
   "tg_tokens": null,
   "tps": 39.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-r2fnjc23-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 18.5 GB; TTFT 4670 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 219.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 219.3 PP tok/s 56.6 TG tok/s 4670 TTFT (ms) 18.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r2fnjc23",
   "source_url": "https://omlx.ai/benchmarks/performance/r2fnjc23",
   "tg_tokens": null,
   "tps": 56.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-mko1fan4-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 20.2 GB; TTFT 4632 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 221.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 221.1 PP tok/s 51.9 TG tok/s 4632 TTFT (ms) 20.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mko1fan4",
   "source_url": "https://omlx.ai/benchmarks/performance/mko1fan4",
   "tg_tokens": null,
   "tps": 51.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-0c8188vf-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 18.5 GB; TTFT 4672 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 219.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 219.2 PP tok/s 53.7 TG tok/s 4672 TTFT (ms) 18.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0c8188vf",
   "source_url": "https://omlx.ai/benchmarks/performance/0c8188vf",
   "tg_tokens": null,
   "tps": 53.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-ewee9ph7-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace, single run at 1k context; peak memory 12.1 GB; TTFT 15777 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 64.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k tokens 64.9 PP tok/s 4.0 TG tok/s 15777 TTFT (ms) 12.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ewee9ph7",
   "source_url": "https://omlx.ai/benchmarks/performance/ewee9ph7",
   "tg_tokens": null,
   "tps": 4.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-fglps72l-ctx1024",
   "model": "ThinkingCap-Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: cfcrm3jo; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 65.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "ThinkingCap-Qwen3.8-27B-oQ5e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 5bit; 1k 65.2 14.3 23.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fglps72l",
   "source_url": "https://omlx.ai/benchmarks/performance/fglps72l",
   "tg_tokens": null,
   "tps": 14.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-fglps72l-ctx4096",
   "model": "ThinkingCap-Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: cfcrm3jo; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 77.3,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "ThinkingCap-Qwen3.8-27B-oQ5e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 5bit; 4k 77.3 12.0 24.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fglps72l",
   "source_url": "https://omlx.ai/benchmarks/performance/fglps72l",
   "tg_tokens": null,
   "tps": 12.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-lguy72sp-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: 5nqla677; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2034.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26b-a4b-it-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 2,034 51.0 25.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lguy72sp",
   "source_url": "https://omlx.ai/benchmarks/performance/lguy72sp",
   "tg_tokens": null,
   "tps": 51.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-lguy72sp-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: 5nqla677; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2136.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26b-a4b-it-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 2,136 49.6 26.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lguy72sp",
   "source_url": "https://omlx.ai/benchmarks/performance/lguy72sp",
   "tg_tokens": null,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-aedr9c8h-ctx1024",
   "model": "Occamy-1.0-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 674.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "occamy-1.0-abliterated-mtp-mlx-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 1k 674.9 58.0 21.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aedr9c8h",
   "source_url": "https://omlx.ai/benchmarks/performance/aedr9c8h",
   "tg_tokens": null,
   "tps": 58.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-aedr9c8h-ctx32768",
   "model": "Occamy-1.0-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 649.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "occamy-1.0-abliterated-mtp-mlx-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 32k 649.1 28.4 25.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aedr9c8h",
   "source_url": "https://omlx.ai/benchmarks/performance/aedr9c8h",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-6h2an05p-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 403.5 GB at this context; duplicate traces of the same session: 6ly38coc, bdzmlf0v, d8y0g22d; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 579.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 579.9 32.0 403.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6h2an05p",
   "source_url": "https://omlx.ai/benchmarks/performance/6h2an05p",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-6h2an05p-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 403.7 GB at this context; duplicate traces of the same session: 6ly38coc, bdzmlf0v, d8y0g22d; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 747.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 747.5 30.0 403.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6h2an05p",
   "source_url": "https://omlx.ai/benchmarks/performance/6h2an05p",
   "tg_tokens": null,
   "tps": 30.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-6h2an05p-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 403.7 GB at this context; duplicate traces of the same session: 6ly38coc, bdzmlf0v, d8y0g22d; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 764.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 16k 764.4 39.1 403.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6h2an05p",
   "source_url": "https://omlx.ai/benchmarks/performance/6h2an05p",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-6h2an05p-ctx32768",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 403.8 GB at this context; duplicate traces of the same session: 6ly38coc, bdzmlf0v, d8y0g22d; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 767.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 32k 767.0 34.6 403.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6h2an05p",
   "source_url": "https://omlx.ai/benchmarks/performance/6h2an05p",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-pyyt8rdt-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 404.9 GB at this context; duplicate traces of the same session: hmpkwoex, k53406jq, yaos04xx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 351.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 351.3 25.1 404.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pyyt8rdt",
   "source_url": "https://omlx.ai/benchmarks/performance/pyyt8rdt",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-pyyt8rdt-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 404.9 GB at this context; duplicate traces of the same session: hmpkwoex, k53406jq, yaos04xx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 702.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 702.5 28.8 404.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pyyt8rdt",
   "source_url": "https://omlx.ai/benchmarks/performance/pyyt8rdt",
   "tg_tokens": null,
   "tps": 28.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-pyyt8rdt-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 405.0 GB at this context; duplicate traces of the same session: hmpkwoex, k53406jq, yaos04xx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 745.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 16k 745.7 38.5 405.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pyyt8rdt",
   "source_url": "https://omlx.ai/benchmarks/performance/pyyt8rdt",
   "tg_tokens": null,
   "tps": 38.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-pyyt8rdt-ctx32768",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 405.0 GB at this context; duplicate traces of the same session: hmpkwoex, k53406jq, yaos04xx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 758.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 32k 758.7 41.8 405.0 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pyyt8rdt",
   "source_url": "https://omlx.ai/benchmarks/performance/pyyt8rdt",
   "tg_tokens": null,
   "tps": 41.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-scui8boi-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.1 GB at this context; duplicate traces of the same session: h2ivi0uq; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 593.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 593.1 59.9 13.1 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/scui8boi",
   "source_url": "https://omlx.ai/benchmarks/performance/scui8boi",
   "tg_tokens": null,
   "tps": 59.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-scui8boi-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.5 GB at this context; duplicate traces of the same session: h2ivi0uq; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 635.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 635.2 57.4 13.5 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/scui8boi",
   "source_url": "https://omlx.ai/benchmarks/performance/scui8boi",
   "tg_tokens": null,
   "tps": 57.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-scui8boi-ctx65536",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 14.7 GB at this context; duplicate traces of the same session: h2ivi0uq; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 429.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 64k 429.5 29.9 14.7 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/scui8boi",
   "source_url": "https://omlx.ai/benchmarks/performance/scui8boi",
   "tg_tokens": null,
   "tps": 29.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-yxf8hyse-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: gi0pg043; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3416.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 3,416 126.0 13.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yxf8hyse",
   "source_url": "https://omlx.ai/benchmarks/performance/yxf8hyse",
   "tg_tokens": null,
   "tps": 126.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-yxf8hyse-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: gi0pg043; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2686.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 2,686 122.4 13.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yxf8hyse",
   "source_url": "https://omlx.ai/benchmarks/performance/yxf8hyse",
   "tg_tokens": null,
   "tps": 122.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-idipi01i-ctx1024",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 768.2,
   "provenance": "community",
   "quant": "5bit",
   "quote": "MiniCPM-V-4.7-35B-A3B-5bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 5bit; 1k 768.2 48.2 26.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/idipi01i",
   "source_url": "https://omlx.ai/benchmarks/performance/idipi01i",
   "tg_tokens": null,
   "tps": 48.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-idipi01i-ctx4096",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 841.2,
   "provenance": "community",
   "quant": "5bit",
   "quote": "MiniCPM-V-4.7-35B-A3B-5bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 5bit; 4k 841.2 47.9 27.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/idipi01i",
   "source_url": "https://omlx.ai/benchmarks/performance/idipi01i",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-v9exv1nr-ctx1024",
   "model": "Muse Glimmer",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: ttrea0q9; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 417.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-q4km-gs128-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 unknown; 1k 417.4 14.8 17.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v9exv1nr",
   "source_url": "https://omlx.ai/benchmarks/performance/v9exv1nr",
   "tg_tokens": null,
   "tps": 14.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-v9exv1nr-ctx4096",
   "model": "Muse Glimmer",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: ttrea0q9; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 424.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-q4km-gs128-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 unknown; 4k 424.4 14.7 18.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v9exv1nr",
   "source_url": "https://omlx.ai/benchmarks/performance/v9exv1nr",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-jxyluw2s-ctx1024",
   "model": "Ornith-1.0-35B",
   "notes": "OptiQ build; oMLX community benchmark trace, single run at 1k context; peak memory 21.9 GB; TTFT 447 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2293.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.0-35B-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k tokens 2,293 PP tok/s 78.3 TG tok/s 447 TTFT (ms) 21.9 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jxyluw2s",
   "source_url": "https://omlx.ai/benchmarks/performance/jxyluw2s",
   "tg_tokens": null,
   "tps": 78.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-n3kk933o-ctx4096",
   "model": "Ornith-1.0-35B",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: zc265017; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2729.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.0-35B-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 2,729 91.6 21.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n3kk933o",
   "source_url": "https://omlx.ai/benchmarks/performance/n3kk933o",
   "tg_tokens": null,
   "tps": 91.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-n3kk933o-ctx8192",
   "model": "Ornith-1.0-35B",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: zc265017; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2603.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.0-35B-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 2,603 91.8 21.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n3kk933o",
   "source_url": "https://omlx.ai/benchmarks/performance/n3kk933o",
   "tg_tokens": null,
   "tps": 91.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-n3kk933o-ctx16384",
   "model": "Ornith-1.0-35B",
   "notes": "oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: zc265017; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2333.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.0-35B-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 2,333 87.0 22.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n3kk933o",
   "source_url": "https://omlx.ai/benchmarks/performance/n3kk933o",
   "tg_tokens": null,
   "tps": 87.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-7ov0g31e-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 6hqgsktw, v9jw7btp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 768.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 768.0 90.8 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ov0g31e",
   "source_url": "https://omlx.ai/benchmarks/performance/7ov0g31e",
   "tg_tokens": null,
   "tps": 90.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-7ov0g31e-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 6hqgsktw, v9jw7btp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 881.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 881.2 86.9 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ov0g31e",
   "source_url": "https://omlx.ai/benchmarks/performance/7ov0g31e",
   "tg_tokens": null,
   "tps": 86.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-7ov0g31e-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: 6hqgsktw, v9jw7btp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 570.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 32k 570.2 62.3 23.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ov0g31e",
   "source_url": "https://omlx.ai/benchmarks/performance/7ov0g31e",
   "tg_tokens": null,
   "tps": 62.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0849b8c5-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 220 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4655.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 4,655 PP tok/s 145.0 TG tok/s 220 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0849b8c5",
   "source_url": "https://omlx.ai/benchmarks/performance/0849b8c5",
   "tg_tokens": null,
   "tps": 145.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-66ewff3x-ctx1024",
   "model": "Ornith-1.5-35B-A3B-REAP192",
   "notes": "oMLX community benchmark trace; peak memory 14.5 GB at this context; duplicate traces of the same session: r4hmu0c8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2723.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Ornith-1.5-35B-A3B-REAP192-mxfp4-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 2,723 109.4 14.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/66ewff3x",
   "source_url": "https://omlx.ai/benchmarks/performance/66ewff3x",
   "tg_tokens": null,
   "tps": 109.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-66ewff3x-ctx4096",
   "model": "Ornith-1.5-35B-A3B-REAP192",
   "notes": "oMLX community benchmark trace; peak memory 15.0 GB at this context; duplicate traces of the same session: r4hmu0c8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3339.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Ornith-1.5-35B-A3B-REAP192-mxfp4-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 3,339 105.2 15.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/66ewff3x",
   "source_url": "https://omlx.ai/benchmarks/performance/66ewff3x",
   "tg_tokens": null,
   "tps": 105.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-e6h93kj1-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: bkayve1u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3137.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 3,137 102.0 5.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e6h93kj1",
   "source_url": "https://omlx.ai/benchmarks/performance/e6h93kj1",
   "tg_tokens": null,
   "tps": 102.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-e6h93kj1-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: bkayve1u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3566.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 3,566 100.5 6.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e6h93kj1",
   "source_url": "https://omlx.ai/benchmarks/performance/e6h93kj1",
   "tg_tokens": null,
   "tps": 100.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-33axuy8d-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 5.9 GB; TTFT 347 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2949.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 2,949 PP tok/s 102.3 TG tok/s 347 TTFT (ms) 5.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/33axuy8d",
   "source_url": "https://omlx.ai/benchmarks/performance/33axuy8d",
   "tg_tokens": null,
   "tps": 102.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-nsxxtu53-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: a2o1w7pp, nj9wrjut; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2637.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 2,637 148.8 23.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nsxxtu53",
   "source_url": "https://omlx.ai/benchmarks/performance/nsxxtu53",
   "tg_tokens": null,
   "tps": 148.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-nsxxtu53-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: a2o1w7pp, nj9wrjut; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2504.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 2,504 182.9 23.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nsxxtu53",
   "source_url": "https://omlx.ai/benchmarks/performance/nsxxtu53",
   "tg_tokens": null,
   "tps": 182.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-nsxxtu53-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: a2o1w7pp, nj9wrjut; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2231.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 2,231 178.1 23.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nsxxtu53",
   "source_url": "https://omlx.ai/benchmarks/performance/nsxxtu53",
   "tg_tokens": null,
   "tps": 178.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-muuqxfoa-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 110.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 110.0 19.5 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/muuqxfoa",
   "source_url": "https://omlx.ai/benchmarks/performance/muuqxfoa",
   "tg_tokens": null,
   "tps": 19.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-muuqxfoa-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 104.1 18.9 19.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/muuqxfoa",
   "source_url": "https://omlx.ai/benchmarks/performance/muuqxfoa",
   "tg_tokens": null,
   "tps": 18.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-t0u56bn8-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: y5xwb4ux; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 328.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 328.1 40.3 33.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t0u56bn8",
   "source_url": "https://omlx.ai/benchmarks/performance/t0u56bn8",
   "tg_tokens": null,
   "tps": 40.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-t0u56bn8-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 37.0 GB at this context; duplicate traces of the same session: y5xwb4ux; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 340.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 340.8 38.6 37.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t0u56bn8",
   "source_url": "https://omlx.ai/benchmarks/performance/t0u56bn8",
   "tg_tokens": null,
   "tps": 38.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-k8d4l39g-ctx1024",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 2afti4zr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 202.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 202.9 42.7 18.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k8d4l39g",
   "source_url": "https://omlx.ai/benchmarks/performance/k8d4l39g",
   "tg_tokens": null,
   "tps": 42.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-k8d4l39g-ctx4096",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: 2afti4zr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 194.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 194.3 36.1 22.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k8d4l39g",
   "source_url": "https://omlx.ai/benchmarks/performance/k8d4l39g",
   "tg_tokens": null,
   "tps": 36.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-na9862p9-ctx1024",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: ah7vpjhp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 221.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 221.0 50.9 18.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/na9862p9",
   "source_url": "https://omlx.ai/benchmarks/performance/na9862p9",
   "tg_tokens": null,
   "tps": 50.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-na9862p9-ctx4096",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: ah7vpjhp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 191.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 191.4 36.4 22.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/na9862p9",
   "source_url": "https://omlx.ai/benchmarks/performance/na9862p9",
   "tg_tokens": null,
   "tps": 36.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-2jrtgdlj-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 26.9 GB; TTFT 350317 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 374.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 128k tokens 374.2 PP tok/s 11.0 TG tok/s 350317 TTFT (ms) 26.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2jrtgdlj",
   "source_url": "https://omlx.ai/benchmarks/performance/2jrtgdlj",
   "tg_tokens": null,
   "tps": 11.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-a4db7ig0-ctx16384",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: 47nfmmfo; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 476.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 476.9 33.4 26.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a4db7ig0",
   "source_url": "https://omlx.ai/benchmarks/performance/a4db7ig0",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-a4db7ig0-ctx65536",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 30.1 GB at this context; duplicate traces of the same session: 47nfmmfo; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 386.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 386.9 24.6 30.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a4db7ig0",
   "source_url": "https://omlx.ai/benchmarks/performance/a4db7ig0",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-4bt29x3d-ctx16384",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: yul7otib; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 461.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 461.3 22.6 26.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4bt29x3d",
   "source_url": "https://omlx.ai/benchmarks/performance/4bt29x3d",
   "tg_tokens": null,
   "tps": 22.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-4bt29x3d-ctx65536",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 29.3 GB at this context; duplicate traces of the same session: yul7otib; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 344.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 344.6 15.6 29.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4bt29x3d",
   "source_url": "https://omlx.ai/benchmarks/performance/4bt29x3d",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-fl9cnh4w-ctx16384",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: 3hnx8kf9; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 462.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 462.4 25.1 26.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fl9cnh4w",
   "source_url": "https://omlx.ai/benchmarks/performance/fl9cnh4w",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-fl9cnh4w-ctx65536",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 29.3 GB at this context; duplicate traces of the same session: 3hnx8kf9; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 372.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 372.7 13.9 29.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fl9cnh4w",
   "source_url": "https://omlx.ai/benchmarks/performance/fl9cnh4w",
   "tg_tokens": null,
   "tps": 13.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-p44le624-ctx16384",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 414.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 414.7 29.3 26.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p44le624",
   "source_url": "https://omlx.ai/benchmarks/performance/p44le624",
   "tg_tokens": null,
   "tps": 29.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-p44le624-ctx65536",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 29.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 340.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 340.6 21.5 29.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p44le624",
   "source_url": "https://omlx.ai/benchmarks/performance/p44le624",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-inb4ko83-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 8416 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 121.7,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k tokens 121.7 PP tok/s 17.6 TG tok/s 8416 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/inb4ko83",
   "source_url": "https://omlx.ai/benchmarks/performance/inb4ko83",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-0o72o68l-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: xyuc2u25; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 199.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 199.2 33.0 18.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0o72o68l",
   "source_url": "https://omlx.ai/benchmarks/performance/0o72o68l",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-0o72o68l-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: xyuc2u25; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 193.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 193.4 38.1 22.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0o72o68l",
   "source_url": "https://omlx.ai/benchmarks/performance/0o72o68l",
   "tg_tokens": null,
   "tps": 38.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-52fw78iu-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: zwgom1zp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 205.1 36.7 18.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/52fw78iu",
   "source_url": "https://omlx.ai/benchmarks/performance/52fw78iu",
   "tg_tokens": null,
   "tps": 36.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-52fw78iu-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: zwgom1zp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 189.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 189.5 30.4 22.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/52fw78iu",
   "source_url": "https://omlx.ai/benchmarks/performance/52fw78iu",
   "tg_tokens": null,
   "tps": 30.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-qplt4mtk-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 195k context; peak memory 34.2 GB; TTFT 2976066 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 67.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 195k tokens 67.2 PP tok/s 13.7 TG tok/s 2976066 TTFT (ms) 34.2 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qplt4mtk",
   "source_url": "https://omlx.ai/benchmarks/performance/qplt4mtk",
   "tg_tokens": null,
   "tps": 13.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-o14mud6d-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 35.9 GB at this context; duplicate traces of the same session: 7t9uojsm; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1513.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 4k 1,513 43.3 35.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o14mud6d",
   "source_url": "https://omlx.ai/benchmarks/performance/o14mud6d",
   "tg_tokens": null,
   "tps": 43.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-o14mud6d-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 36.6 GB at this context; duplicate traces of the same session: 7t9uojsm; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1576.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 16k 1,576 78.1 36.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o14mud6d",
   "source_url": "https://omlx.ai/benchmarks/performance/o14mud6d",
   "tg_tokens": null,
   "tps": 78.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-h8yoxacr-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace, single run at 4k context; peak memory 104.7 GB; TTFT 4251 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 963.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k tokens 963.6 PP tok/s 58.3 TG tok/s 4251 TTFT (ms) 104.7 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h8yoxacr",
   "source_url": "https://omlx.ai/benchmarks/performance/h8yoxacr",
   "tg_tokens": null,
   "tps": 58.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-yvufozpj-ctx1024",
   "model": "Spark-X2.5-4B",
   "notes": "OptiQ build; oMLX community benchmark trace; peak memory 3.7 GB at this context; duplicate traces of the same session: 0mc1ig5p; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2433.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Spark-X2.5-4B-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 2,433 77.3 3.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yvufozpj",
   "source_url": "https://omlx.ai/benchmarks/performance/yvufozpj",
   "tg_tokens": null,
   "tps": 77.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-yvufozpj-ctx4096",
   "model": "Spark-X2.5-4B",
   "notes": "OptiQ build; oMLX community benchmark trace; peak memory 3.8 GB at this context; duplicate traces of the same session: 0mc1ig5p; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2477.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Spark-X2.5-4B-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 2,477 74.1 3.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yvufozpj",
   "source_url": "https://omlx.ai/benchmarks/performance/yvufozpj",
   "tg_tokens": null,
   "tps": 74.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-3xh9dgp1-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: l4lq2xw2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 272.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 272.6 31.2 25.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3xh9dgp1",
   "source_url": "https://omlx.ai/benchmarks/performance/3xh9dgp1",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-3xh9dgp1-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: l4lq2xw2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 233.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 233.2 82.1 27.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3xh9dgp1",
   "source_url": "https://omlx.ai/benchmarks/performance/3xh9dgp1",
   "tg_tokens": null,
   "tps": 82.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-cadspr9v-ctx1024",
   "model": "ThinkingCap-Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 98.8,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "ThinkingCap-Qwen3.8-27B-oQ5e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 5bit; 1k 98.8 19.1 23.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cadspr9v",
   "source_url": "https://omlx.ai/benchmarks/performance/cadspr9v",
   "tg_tokens": null,
   "tps": 19.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-cadspr9v-ctx4096",
   "model": "ThinkingCap-Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 99.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "ThinkingCap-Qwen3.8-27B-oQ5e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 5bit; 4k 99.7 19.0 24.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cadspr9v",
   "source_url": "https://omlx.ai/benchmarks/performance/cadspr9v",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-bm8dcvm0-ctx1024",
   "model": "Occamy-1.0-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 674.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "occamy-1.0-abliterated-mtp-mlx-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 1k 674.9 58.0 21.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bm8dcvm0",
   "source_url": "https://omlx.ai/benchmarks/performance/bm8dcvm0",
   "tg_tokens": null,
   "tps": 58.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-bm8dcvm0-ctx32768",
   "model": "Occamy-1.0-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 649.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "occamy-1.0-abliterated-mtp-mlx-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 32k 649.1 28.4 25.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bm8dcvm0",
   "source_url": "https://omlx.ai/benchmarks/performance/bm8dcvm0",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xmwxnn7c-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.0 GB at this context; duplicate traces of the same session: 0oplgpdi, lag9ymy9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 84.9,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 84.9 12.7 92.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xmwxnn7c",
   "source_url": "https://omlx.ai/benchmarks/performance/xmwxnn7c",
   "tg_tokens": null,
   "tps": 12.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-xmwxnn7c-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 93.8 GB at this context; duplicate traces of the same session: 0oplgpdi, lag9ymy9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 419.8,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 419.8 13.4 93.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xmwxnn7c",
   "source_url": "https://omlx.ai/benchmarks/performance/xmwxnn7c",
   "tg_tokens": null,
   "tps": 13.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xmwxnn7c-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 93.8 GB at this context; duplicate traces of the same session: 0oplgpdi, lag9ymy9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 460.1,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 460.1 14.6 93.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xmwxnn7c",
   "source_url": "https://omlx.ai/benchmarks/performance/xmwxnn7c",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-msmcq8rh-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 323.5 GB at this context; duplicate traces of the same session: i7m8lbpr, lleos7ma, mj97ppmj, tcevevi7, z21ji9ao; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 506.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 506.2 34.9 323.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/msmcq8rh",
   "source_url": "https://omlx.ai/benchmarks/performance/msmcq8rh",
   "tg_tokens": null,
   "tps": 34.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-msmcq8rh-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 324.7 GB at this context; duplicate traces of the same session: i7m8lbpr, lleos7ma, mj97ppmj, tcevevi7, z21ji9ao; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 585.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 585.7 36.9 324.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/msmcq8rh",
   "source_url": "https://omlx.ai/benchmarks/performance/msmcq8rh",
   "tg_tokens": null,
   "tps": 36.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-msmcq8rh-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 325.0 GB at this context; duplicate traces of the same session: i7m8lbpr, lleos7ma, mj97ppmj, tcevevi7, z21ji9ao; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 543.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 543.3 42.0 325.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/msmcq8rh",
   "source_url": "https://omlx.ai/benchmarks/performance/msmcq8rh",
   "tg_tokens": null,
   "tps": 42.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-msmcq8rh-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 325.2 GB at this context; duplicate traces of the same session: i7m8lbpr, lleos7ma, mj97ppmj, tcevevi7, z21ji9ao; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 532.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k 532.9 65.6 325.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/msmcq8rh",
   "source_url": "https://omlx.ai/benchmarks/performance/msmcq8rh",
   "tg_tokens": null,
   "tps": 65.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-msmcq8rh-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 325.4 GB at this context; duplicate traces of the same session: i7m8lbpr, lleos7ma, mj97ppmj, tcevevi7, z21ji9ao; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 525.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 32k 525.8 49.9 325.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/msmcq8rh",
   "source_url": "https://omlx.ai/benchmarks/performance/msmcq8rh",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-msmcq8rh-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 326.1 GB at this context; duplicate traces of the same session: i7m8lbpr, lleos7ma, mj97ppmj, tcevevi7, z21ji9ao; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 518.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 64k 518.4 42.7 326.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/msmcq8rh",
   "source_url": "https://omlx.ai/benchmarks/performance/msmcq8rh",
   "tg_tokens": null,
   "tps": 42.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-msmcq8rh-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 327.4 GB at this context; duplicate traces of the same session: i7m8lbpr, lleos7ma, mj97ppmj, tcevevi7, z21ji9ao; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 507.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 128k 507.6 64.8 327.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/msmcq8rh",
   "source_url": "https://omlx.ai/benchmarks/performance/msmcq8rh",
   "tg_tokens": null,
   "tps": 64.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-msmcq8rh-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 328.7 GB at this context; duplicate traces of the same session: i7m8lbpr, lleos7ma, mj97ppmj, tcevevi7, z21ji9ao; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 497.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 195k 497.8 33.8 328.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/msmcq8rh",
   "source_url": "https://omlx.ai/benchmarks/performance/msmcq8rh",
   "tg_tokens": null,
   "tps": 33.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jkfr2ya2-ctx1024",
   "model": "Huihui-Qwen3.8-27B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 65l4s426, ci7p03bi, dgu7h1td; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 410.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.8-27B-abliterated-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 410.9 35.6 18.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jkfr2ya2",
   "source_url": "https://omlx.ai/benchmarks/performance/jkfr2ya2",
   "tg_tokens": null,
   "tps": 35.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jkfr2ya2-ctx4096",
   "model": "Huihui-Qwen3.8-27B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 65l4s426, ci7p03bi, dgu7h1td; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 461.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.8-27B-abliterated-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 461.1 69.0 21.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jkfr2ya2",
   "source_url": "https://omlx.ai/benchmarks/performance/jkfr2ya2",
   "tg_tokens": null,
   "tps": 69.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jkfr2ya2-ctx8192",
   "model": "Huihui-Qwen3.8-27B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 65l4s426, ci7p03bi, dgu7h1td; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 453.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.8-27B-abliterated-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 453.9 91.7 22.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jkfr2ya2",
   "source_url": "https://omlx.ai/benchmarks/performance/jkfr2ya2",
   "tg_tokens": null,
   "tps": 91.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-jkfr2ya2-ctx16384",
   "model": "Huihui-Qwen3.8-27B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: 65l4s426, ci7p03bi, dgu7h1td; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 437.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.8-27B-abliterated-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 437.3 63.8 23.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jkfr2ya2",
   "source_url": "https://omlx.ai/benchmarks/performance/jkfr2ya2",
   "tg_tokens": null,
   "tps": 63.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ebal8chu-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 32k context; peak memory 22.8 GB; TTFT 13505 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2426.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k tokens 2,426 PP tok/s 90.6 TG tok/s 13505 TTFT (ms) 22.8 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ebal8chu",
   "source_url": "https://omlx.ai/benchmarks/performance/ebal8chu",
   "tg_tokens": null,
   "tps": 90.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-2pm65tq3-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: l0fktwej; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 730.9,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 730.9 72.5 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2pm65tq3",
   "source_url": "https://omlx.ai/benchmarks/performance/2pm65tq3",
   "tg_tokens": null,
   "tps": 72.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-2pm65tq3-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: l0fktwej; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 791.2,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 791.2 69.8 21.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2pm65tq3",
   "source_url": "https://omlx.ai/benchmarks/performance/2pm65tq3",
   "tg_tokens": null,
   "tps": 69.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-7y92l3dc-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.7 GB at this context; duplicate traces of the same session: 81yn43l5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1734.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 1,734 70.0 31.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7y92l3dc",
   "source_url": "https://omlx.ai/benchmarks/performance/7y92l3dc",
   "tg_tokens": null,
   "tps": 70.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-7y92l3dc-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.9 GB at this context; duplicate traces of the same session: 81yn43l5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1343.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 1,343 94.4 32.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7y92l3dc",
   "source_url": "https://omlx.ai/benchmarks/performance/7y92l3dc",
   "tg_tokens": null,
   "tps": 94.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-367ppnuq-ctx1024",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: r4aqoq7r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 192.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 192.2 37.9 18.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/367ppnuq",
   "source_url": "https://omlx.ai/benchmarks/performance/367ppnuq",
   "tg_tokens": null,
   "tps": 37.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-367ppnuq-ctx4096",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: r4aqoq7r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 176.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 176.3 32.1 22.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/367ppnuq",
   "source_url": "https://omlx.ai/benchmarks/performance/367ppnuq",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-t4mrxnis-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: h8c5qi2j, j0fw7hfj, s0g58yam, w20s33po; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 279.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 279.8 45.0 22.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t4mrxnis",
   "source_url": "https://omlx.ai/benchmarks/performance/t4mrxnis",
   "tg_tokens": null,
   "tps": 45.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-t4mrxnis-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: h8c5qi2j, j0fw7hfj, s0g58yam, w20s33po; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 246.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 8k 246.3 43.7 23.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t4mrxnis",
   "source_url": "https://omlx.ai/benchmarks/performance/t4mrxnis",
   "tg_tokens": null,
   "tps": 43.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-t4mrxnis-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: h8c5qi2j, j0fw7hfj, s0g58yam, w20s33po; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 237.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 237.0 41.4 21.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t4mrxnis",
   "source_url": "https://omlx.ai/benchmarks/performance/t4mrxnis",
   "tg_tokens": null,
   "tps": 41.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-t4mrxnis-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: h8c5qi2j, j0fw7hfj, s0g58yam, w20s33po; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 220.5 30.1 23.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t4mrxnis",
   "source_url": "https://omlx.ai/benchmarks/performance/t4mrxnis",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-t4mrxnis-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.9 GB at this context; duplicate traces of the same session: h8c5qi2j, j0fw7hfj, s0g58yam, w20s33po; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 169.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 64k 169.2 15.1 28.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t4mrxnis",
   "source_url": "https://omlx.ai/benchmarks/performance/t4mrxnis",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2bwk6bwm-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: bgevu3zv, kuw27ax9, u4xn4r4f, vh0rre35; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 295.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 295.8 50.8 22.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2bwk6bwm",
   "source_url": "https://omlx.ai/benchmarks/performance/2bwk6bwm",
   "tg_tokens": null,
   "tps": 50.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2bwk6bwm-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: bgevu3zv, kuw27ax9, u4xn4r4f, vh0rre35; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 297.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 8k 297.8 44.5 23.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2bwk6bwm",
   "source_url": "https://omlx.ai/benchmarks/performance/2bwk6bwm",
   "tg_tokens": null,
   "tps": 44.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2bwk6bwm-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: bgevu3zv, kuw27ax9, u4xn4r4f, vh0rre35; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 227.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 227.1 35.7 21.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2bwk6bwm",
   "source_url": "https://omlx.ai/benchmarks/performance/2bwk6bwm",
   "tg_tokens": null,
   "tps": 35.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2bwk6bwm-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: bgevu3zv, kuw27ax9, u4xn4r4f, vh0rre35; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 161.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 161.0 23.6 23.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2bwk6bwm",
   "source_url": "https://omlx.ai/benchmarks/performance/2bwk6bwm",
   "tg_tokens": null,
   "tps": 23.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-2bwk6bwm-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.9 GB at this context; duplicate traces of the same session: bgevu3zv, kuw27ax9, u4xn4r4f, vh0rre35; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 147.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 64k 147.9 17.6 28.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2bwk6bwm",
   "source_url": "https://omlx.ai/benchmarks/performance/2bwk6bwm",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-crx1o04f-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 27.8 GB; TTFT 366609 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 357.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 128k tokens 357.5 PP tok/s 24.4 TG tok/s 366609 TTFT (ms) 27.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/crx1o04f",
   "source_url": "https://omlx.ai/benchmarks/performance/crx1o04f",
   "tg_tokens": null,
   "tps": 24.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-kwgv022n-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 27.8 GB; TTFT 363921 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 360.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 128k tokens 360.2 PP tok/s 27.9 TG tok/s 363921 TTFT (ms) 27.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kwgv022n",
   "source_url": "https://omlx.ai/benchmarks/performance/kwgv022n",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-yx9s3or8-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 4dobcuob; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 470.8,
   "provenance": "community",
   "quant": "OptiQ 4bit",
   "quote": "Qwen3.8-27B-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 470.8 14.2 20.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yx9s3or8",
   "source_url": "https://omlx.ai/benchmarks/performance/yx9s3or8",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-yx9s3or8-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: 4dobcuob; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 482.7,
   "provenance": "community",
   "quant": "OptiQ 4bit",
   "quote": "Qwen3.8-27B-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 482.7 13.9 21.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yx9s3or8",
   "source_url": "https://omlx.ai/benchmarks/performance/yx9s3or8",
   "tg_tokens": null,
   "tps": 13.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-fdunbsmc-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: jqpufryf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 217.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ8e-1M on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 217.2 27.4 33.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fdunbsmc",
   "source_url": "https://omlx.ai/benchmarks/performance/fdunbsmc",
   "tg_tokens": null,
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-fdunbsmc-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 37.7 GB at this context; duplicate traces of the same session: jqpufryf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 185.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ8e-1M on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 185.3 19.3 37.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fdunbsmc",
   "source_url": "https://omlx.ai/benchmarks/performance/fdunbsmc",
   "tg_tokens": null,
   "tps": 19.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-j5p09ptg-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "Jundot build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.2 GB at this context; duplicate traces of the same session: llxz3z9q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 318.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp-Jundot on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 318.5 30.6 25.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j5p09ptg",
   "source_url": "https://omlx.ai/benchmarks/performance/j5p09ptg",
   "tg_tokens": null,
   "tps": 30.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-j5p09ptg-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "Jundot build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: llxz3z9q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 300.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp-Jundot on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 300.0 47.6 26.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j5p09ptg",
   "source_url": "https://omlx.ai/benchmarks/performance/j5p09ptg",
   "tg_tokens": null,
   "tps": 47.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-s95mer5a-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 23.8 GB; TTFT 38504 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.4,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-27B-oQ6 on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 4k tokens 106.4 PP tok/s 10.7 TG tok/s 38504 TTFT (ms) 23.8 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s95mer5a",
   "source_url": "https://omlx.ai/benchmarks/performance/s95mer5a",
   "tg_tokens": null,
   "tps": 10.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-alh9ojhh-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 77.9 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 554.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 554.7 49.7 77.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/alh9ojhh",
   "source_url": "https://omlx.ai/benchmarks/performance/alh9ojhh",
   "tg_tokens": null,
   "tps": 49.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-alh9ojhh-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.4 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 644.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 644.6 36.5 78.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/alh9ojhh",
   "source_url": "https://omlx.ai/benchmarks/performance/alh9ojhh",
   "tg_tokens": null,
   "tps": 36.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-u5mttbtx-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 100.7 GB; TTFT 866 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1182.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-Flash-Next-oQ4-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k tokens 1,182 PP tok/s 103.0 TG tok/s 866 TTFT (ms) 100.7 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u5mttbtx",
   "source_url": "https://omlx.ai/benchmarks/performance/u5mttbtx",
   "tg_tokens": null,
   "tps": 103.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gxhvv9uv-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 100.5 GB; TTFT 977 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1048.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-Flash-Next-oQ4-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k tokens 1,048 PP tok/s 100.8 TG tok/s 977 TTFT (ms) 100.5 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gxhvv9uv",
   "source_url": "https://omlx.ai/benchmarks/performance/gxhvv9uv",
   "tg_tokens": null,
   "tps": 100.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vg5wlckx-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.7 GB at this context; duplicate traces of the same session: 2doql2lm; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1181.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-Flash-Next-oQ4-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,181 100.8 100.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vg5wlckx",
   "source_url": "https://omlx.ai/benchmarks/performance/vg5wlckx",
   "tg_tokens": null,
   "tps": 100.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vg5wlckx-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.5 GB at this context; duplicate traces of the same session: 2doql2lm; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1006.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-Flash-Next-oQ4-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,006 97.0 102.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vg5wlckx",
   "source_url": "https://omlx.ai/benchmarks/performance/vg5wlckx",
   "tg_tokens": null,
   "tps": 97.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-zom8er3o-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: mho3su3o; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 324.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 324.9 30.6 71.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zom8er3o",
   "source_url": "https://omlx.ai/benchmarks/performance/zom8er3o",
   "tg_tokens": null,
   "tps": 30.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-zom8er3o-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.0 GB at this context; duplicate traces of the same session: mho3su3o; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 332.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 64k 332.4 18.0 74.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zom8er3o",
   "source_url": "https://omlx.ai/benchmarks/performance/zom8er3o",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-9od3bbyl-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: m5l9q163; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 313.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 313.7 21.4 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9od3bbyl",
   "source_url": "https://omlx.ai/benchmarks/performance/9od3bbyl",
   "tg_tokens": null,
   "tps": 21.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-9od3bbyl-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: m5l9q163; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 356.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 356.3 39.8 71.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9od3bbyl",
   "source_url": "https://omlx.ai/benchmarks/performance/9od3bbyl",
   "tg_tokens": null,
   "tps": 39.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-v3zpx4ji-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: 1ztbipw2, fjior51w, nz1dqnwh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 358.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 358.1 59.9 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v3zpx4ji",
   "source_url": "https://omlx.ai/benchmarks/performance/v3zpx4ji",
   "tg_tokens": null,
   "tps": 59.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-v3zpx4ji-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: 1ztbipw2, fjior51w, nz1dqnwh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 873.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 873.8 56.3 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v3zpx4ji",
   "source_url": "https://omlx.ai/benchmarks/performance/v3zpx4ji",
   "tg_tokens": null,
   "tps": 56.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-v3zpx4ji-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 1ztbipw2, fjior51w, nz1dqnwh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1714.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,714 62.9 71.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v3zpx4ji",
   "source_url": "https://omlx.ai/benchmarks/performance/v3zpx4ji",
   "tg_tokens": null,
   "tps": 62.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-v3zpx4ji-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: 1ztbipw2, fjior51w, nz1dqnwh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2692.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,692 101.5 72.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v3zpx4ji",
   "source_url": "https://omlx.ai/benchmarks/performance/v3zpx4ji",
   "tg_tokens": null,
   "tps": 101.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ftu93ueh-ctx1024",
   "model": "ThinkingCap-Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: drozusg8; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 95.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "ThinkingCap-Qwen3.8-27B-oQ5e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 5bit; 1k 95.9 21.4 23.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ftu93ueh",
   "source_url": "https://omlx.ai/benchmarks/performance/ftu93ueh",
   "tg_tokens": null,
   "tps": 21.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ftu93ueh-ctx4096",
   "model": "ThinkingCap-Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: drozusg8; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 96.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "ThinkingCap-Qwen3.8-27B-oQ5e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 5bit; 4k 96.9 18.4 24.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ftu93ueh",
   "source_url": "https://omlx.ai/benchmarks/performance/ftu93ueh",
   "tg_tokens": null,
   "tps": 18.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-mw84s3xi-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: mima5w4p; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 787.9,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 787.9 70.4 11.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mw84s3xi",
   "source_url": "https://omlx.ai/benchmarks/performance/mw84s3xi",
   "tg_tokens": null,
   "tps": 70.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-mw84s3xi-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: mima5w4p; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 784.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 784.0 67.8 11.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mw84s3xi",
   "source_url": "https://omlx.ai/benchmarks/performance/mw84s3xi",
   "tg_tokens": null,
   "tps": 67.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-60gqn7d1-ctx65536",
   "model": "gpt-oss 20B",
   "notes": "bf16 MLX build; oMLX community benchmark trace, single run at 64k context; peak memory 16.0 GB; TTFT 281064 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 233.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "gpt-oss-20b-oQ5e-bf16-MLX on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 5bit; 64k tokens 233.2 PP tok/s 16.3 TG tok/s 281064 TTFT (ms) 16.0 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/60gqn7d1",
   "source_url": "https://omlx.ai/benchmarks/performance/60gqn7d1",
   "tg_tokens": null,
   "tps": 16.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-t81hae2h-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 92.0 GB; TTFT 12299 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 83.3,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k tokens 83.3 PP tok/s 12.0 TG tok/s 12299 TTFT (ms) 92.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t81hae2h",
   "source_url": "https://omlx.ai/benchmarks/performance/t81hae2h",
   "tg_tokens": null,
   "tps": 12.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-thgag2ac-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 323.5 GB at this context; duplicate traces of the same session: 534hmjve; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 506.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 506.2 34.9 323.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/thgag2ac",
   "source_url": "https://omlx.ai/benchmarks/performance/thgag2ac",
   "tg_tokens": null,
   "tps": 34.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-thgag2ac-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 324.7 GB at this context; duplicate traces of the same session: 534hmjve; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 585.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 585.7 36.9 324.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/thgag2ac",
   "source_url": "https://omlx.ai/benchmarks/performance/thgag2ac",
   "tg_tokens": null,
   "tps": 36.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-thgag2ac-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 325.0 GB at this context; duplicate traces of the same session: 534hmjve; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 543.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 543.3 42.0 325.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/thgag2ac",
   "source_url": "https://omlx.ai/benchmarks/performance/thgag2ac",
   "tg_tokens": null,
   "tps": 42.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-thgag2ac-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 325.2 GB at this context; duplicate traces of the same session: 534hmjve; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 532.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k 532.9 65.6 325.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/thgag2ac",
   "source_url": "https://omlx.ai/benchmarks/performance/thgag2ac",
   "tg_tokens": null,
   "tps": 65.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-thgag2ac-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 325.4 GB at this context; duplicate traces of the same session: 534hmjve; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 525.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 32k 525.8 49.9 325.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/thgag2ac",
   "source_url": "https://omlx.ai/benchmarks/performance/thgag2ac",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-thgag2ac-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 326.1 GB at this context; duplicate traces of the same session: 534hmjve; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 518.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 64k 518.4 42.7 326.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/thgag2ac",
   "source_url": "https://omlx.ai/benchmarks/performance/thgag2ac",
   "tg_tokens": null,
   "tps": 42.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-thgag2ac-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 327.4 GB at this context; duplicate traces of the same session: 534hmjve; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 507.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 128k 507.6 64.8 327.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/thgag2ac",
   "source_url": "https://omlx.ai/benchmarks/performance/thgag2ac",
   "tg_tokens": null,
   "tps": 64.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-thgag2ac-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 328.7 GB at this context; duplicate traces of the same session: 534hmjve; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 497.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "GLM-5.3-Flash-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 195k 497.8 33.8 328.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/thgag2ac",
   "source_url": "https://omlx.ai/benchmarks/performance/thgag2ac",
   "tg_tokens": null,
   "tps": 33.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-yx6ojo9i-ctx4096",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: r3t2k20f; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 806.9,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "MiniCPM-V-4.7-35B-A3B-mxfp4 on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 806.9 38.4 22.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yx6ojo9i",
   "source_url": "https://omlx.ai/benchmarks/performance/yx6ojo9i",
   "tg_tokens": null,
   "tps": 38.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-yx6ojo9i-ctx16384",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: r3t2k20f; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 700.1,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "MiniCPM-V-4.7-35B-A3B-mxfp4 on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 700.1 35.3 23.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yx6ojo9i",
   "source_url": "https://omlx.ai/benchmarks/performance/yx6ojo9i",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zcks8qyw-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 219 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4667.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 4,667 PP tok/s 143.8 TG tok/s 219 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zcks8qyw",
   "source_url": "https://omlx.ai/benchmarks/performance/zcks8qyw",
   "tg_tokens": null,
   "tps": 143.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-z2cg0iw7-ctx1024",
   "model": "Qwen2.5-Coder-7B-Instruct",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 4.5 GB; TTFT 1069 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 957.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-7B-Instruct-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 957.9 PP tok/s 29.5 TG tok/s 1069 TTFT (ms) 4.5 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z2cg0iw7",
   "source_url": "https://omlx.ai/benchmarks/performance/z2cg0iw7",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-z29no657-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: eebi3n2v, oirebglu; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 432.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 432.9 39.4 3.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z29no657",
   "source_url": "https://omlx.ai/benchmarks/performance/z29no657",
   "tg_tokens": null,
   "tps": 39.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-z29no657-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.7 GB at this context; duplicate traces of the same session: eebi3n2v, oirebglu; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 438.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 438.1 37.6 3.7 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z29no657",
   "source_url": "https://omlx.ai/benchmarks/performance/z29no657",
   "tg_tokens": null,
   "tps": 37.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-z29no657-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.0 GB at this context; duplicate traces of the same session: eebi3n2v, oirebglu; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 427.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 8k 427.3 36.1 4.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z29no657",
   "source_url": "https://omlx.ai/benchmarks/performance/z29no657",
   "tg_tokens": null,
   "tps": 36.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-60vsukuy-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: r76u0q56; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 235.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 235.4 22.0 5.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/60vsukuy",
   "source_url": "https://omlx.ai/benchmarks/performance/60vsukuy",
   "tg_tokens": null,
   "tps": 22.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-60vsukuy-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: r76u0q56; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 242.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 242.1 21.6 6.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/60vsukuy",
   "source_url": "https://omlx.ai/benchmarks/performance/60vsukuy",
   "tg_tokens": null,
   "tps": 21.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-60vsukuy-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: r76u0q56; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 240.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 8k 240.2 21.1 6.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/60vsukuy",
   "source_url": "https://omlx.ai/benchmarks/performance/60vsukuy",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-9zzaj0wf-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 20.3 GB; TTFT 6735 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2433.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k tokens 2,433 PP tok/s 82.8 TG tok/s 6735 TTFT (ms) 20.3 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9zzaj0wf",
   "source_url": "https://omlx.ai/benchmarks/performance/9zzaj0wf",
   "tg_tokens": null,
   "tps": 82.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-kt3yhlm5-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 20.3 GB; TTFT 8182 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2003.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k tokens 2,003 PP tok/s 82.3 TG tok/s 8182 TTFT (ms) 20.3 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kt3yhlm5",
   "source_url": "https://omlx.ai/benchmarks/performance/kt3yhlm5",
   "tg_tokens": null,
   "tps": 82.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-zrwqeg0j-ctx16384",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: mk71awb0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2691.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,691 105.0 20.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zrwqeg0j",
   "source_url": "https://omlx.ai/benchmarks/performance/zrwqeg0j",
   "tg_tokens": null,
   "tps": 105.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-zrwqeg0j-ctx65536",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: mk71awb0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1663.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 1,663 57.1 21.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zrwqeg0j",
   "source_url": "https://omlx.ai/benchmarks/performance/zrwqeg0j",
   "tg_tokens": null,
   "tps": 57.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-sj2di77x-ctx16384",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: e8fruyxu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2643.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,643 88.2 20.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sj2di77x",
   "source_url": "https://omlx.ai/benchmarks/performance/sj2di77x",
   "tg_tokens": null,
   "tps": 88.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-sj2di77x-ctx65536",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: e8fruyxu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1637.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 1,637 60.3 21.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sj2di77x",
   "source_url": "https://omlx.ai/benchmarks/performance/sj2di77x",
   "tg_tokens": null,
   "tps": 60.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-7ym2pfuw-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: xu7vdzee; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 414.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k 414.9 41.7 21.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ym2pfuw",
   "source_url": "https://omlx.ai/benchmarks/performance/7ym2pfuw",
   "tg_tokens": null,
   "tps": 41.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-7ym2pfuw-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: xu7vdzee; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 447.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 447.8 41.9 22.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ym2pfuw",
   "source_url": "https://omlx.ai/benchmarks/performance/7ym2pfuw",
   "tg_tokens": null,
   "tps": 41.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-g8wxsmkn-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 5r9uo7i7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 678.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 678.6 15.4 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g8wxsmkn",
   "source_url": "https://omlx.ai/benchmarks/performance/g8wxsmkn",
   "tg_tokens": null,
   "tps": 15.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-g8wxsmkn-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: 5r9uo7i7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 563.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 563.3 13.7 22.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g8wxsmkn",
   "source_url": "https://omlx.ai/benchmarks/performance/g8wxsmkn",
   "tg_tokens": null,
   "tps": 13.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-b7ap2bvn-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.9 GB at this context; duplicate traces of the same session: 8hipm7gq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 417.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 417.9 11.4 17.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b7ap2bvn",
   "source_url": "https://omlx.ai/benchmarks/performance/b7ap2bvn",
   "tg_tokens": null,
   "tps": 11.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-b7ap2bvn-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 8hipm7gq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 367.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 367.3 11.4 20.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b7ap2bvn",
   "source_url": "https://omlx.ai/benchmarks/performance/b7ap2bvn",
   "tg_tokens": null,
   "tps": 11.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-h8d4uuh0-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.6 GB at this context; duplicate traces of the same session: 7jy3wuei; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 850.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 850.7 16.9 31.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h8d4uuh0",
   "source_url": "https://omlx.ai/benchmarks/performance/h8d4uuh0",
   "tg_tokens": null,
   "tps": 16.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-h8d4uuh0-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: 7jy3wuei; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 598.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 598.6 14.2 37.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h8d4uuh0",
   "source_url": "https://omlx.ai/benchmarks/performance/h8d4uuh0",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-cnqpsqf6-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 7pqogqhw, fah8i704; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 676.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 676.2 36.6 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cnqpsqf6",
   "source_url": "https://omlx.ai/benchmarks/performance/cnqpsqf6",
   "tg_tokens": null,
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-cnqpsqf6-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 7pqogqhw, fah8i704; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 677.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 677.1 58.6 19.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cnqpsqf6",
   "source_url": "https://omlx.ai/benchmarks/performance/cnqpsqf6",
   "tg_tokens": null,
   "tps": 58.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-cnqpsqf6-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: 7pqogqhw, fah8i704; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 504.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 504.8 31.7 23.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cnqpsqf6",
   "source_url": "https://omlx.ai/benchmarks/performance/cnqpsqf6",
   "tg_tokens": null,
   "tps": 31.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-98fohc0u-ctx16384",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: h0d8uwmk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 423.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 423.9 29.7 26.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/98fohc0u",
   "source_url": "https://omlx.ai/benchmarks/performance/98fohc0u",
   "tg_tokens": null,
   "tps": 29.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-98fohc0u-ctx65536",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 30.1 GB at this context; duplicate traces of the same session: h0d8uwmk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 337.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 337.8 19.6 30.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/98fohc0u",
   "source_url": "https://omlx.ai/benchmarks/performance/98fohc0u",
   "tg_tokens": null,
   "tps": 19.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-yb27npcu-ctx16384",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: nwk7bhn6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 444.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 444.3 28.1 26.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yb27npcu",
   "source_url": "https://omlx.ai/benchmarks/performance/yb27npcu",
   "tg_tokens": null,
   "tps": 28.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-yb27npcu-ctx65536",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 30.4 GB at this context; duplicate traces of the same session: nwk7bhn6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 297.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 297.4 20.9 30.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yb27npcu",
   "source_url": "https://omlx.ai/benchmarks/performance/yb27npcu",
   "tg_tokens": null,
   "tps": 20.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ozobp1ss-ctx16384",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: kapduj79; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 386.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 386.2 24.6 26.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ozobp1ss",
   "source_url": "https://omlx.ai/benchmarks/performance/ozobp1ss",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ozobp1ss-ctx65536",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 29.3 GB at this context; duplicate traces of the same session: kapduj79; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 326.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 326.9 22.5 29.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ozobp1ss",
   "source_url": "https://omlx.ai/benchmarks/performance/ozobp1ss",
   "tg_tokens": null,
   "tps": 22.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-dvz2c15k-ctx16384",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: dt8sev0f; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 407.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 407.6 30.5 26.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dvz2c15k",
   "source_url": "https://omlx.ai/benchmarks/performance/dvz2c15k",
   "tg_tokens": null,
   "tps": 30.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-dvz2c15k-ctx65536",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 29.3 GB at this context; duplicate traces of the same session: dt8sev0f; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 331.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 331.7 23.7 29.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dvz2c15k",
   "source_url": "https://omlx.ai/benchmarks/performance/dvz2c15k",
   "tg_tokens": null,
   "tps": 23.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-hjurg9o8-ctx16384",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: ztqiroiz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 427.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 427.5 28.8 26.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hjurg9o8",
   "source_url": "https://omlx.ai/benchmarks/performance/hjurg9o8",
   "tg_tokens": null,
   "tps": 28.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-hjurg9o8-ctx65536",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 29.3 GB at this context; duplicate traces of the same session: ztqiroiz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 305.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 305.7 20.1 29.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hjurg9o8",
   "source_url": "https://omlx.ai/benchmarks/performance/hjurg9o8",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-428z38cc-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: qutc7bta; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1128.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,128 61.0 17.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/428z38cc",
   "source_url": "https://omlx.ai/benchmarks/performance/428z38cc",
   "tg_tokens": null,
   "tps": 61.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-428z38cc-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: qutc7bta; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1364.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,364 59.2 19.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/428z38cc",
   "source_url": "https://omlx.ai/benchmarks/performance/428z38cc",
   "tg_tokens": null,
   "tps": 59.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-8unu2i4w-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: l338k4oq, n84hszkd, yzyk33st; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 312.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 312.3 23.0 21.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8unu2i4w",
   "source_url": "https://omlx.ai/benchmarks/performance/8unu2i4w",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-8unu2i4w-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: l338k4oq, n84hszkd, yzyk33st; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 333.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 333.4 23.4 22.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8unu2i4w",
   "source_url": "https://omlx.ai/benchmarks/performance/8unu2i4w",
   "tg_tokens": null,
   "tps": 23.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-8unu2i4w-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: l338k4oq, n84hszkd, yzyk33st; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 210.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 210.4 31.1 23.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8unu2i4w",
   "source_url": "https://omlx.ai/benchmarks/performance/8unu2i4w",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-8unu2i4w-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: l338k4oq, n84hszkd, yzyk33st; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 248.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 248.6 27.3 23.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8unu2i4w",
   "source_url": "https://omlx.ai/benchmarks/performance/8unu2i4w",
   "tg_tokens": null,
   "tps": 27.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-oaezns2k-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 77.9 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 554.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 554.7 49.7 77.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oaezns2k",
   "source_url": "https://omlx.ai/benchmarks/performance/oaezns2k",
   "tg_tokens": null,
   "tps": 49.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-oaezns2k-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.4 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 644.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 644.6 36.5 78.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oaezns2k",
   "source_url": "https://omlx.ai/benchmarks/performance/oaezns2k",
   "tg_tokens": null,
   "tps": 36.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-ho9kxtcd-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 7ze0gewp, a3cogmu8, gcqz7pth, io4qmtkh, k4zjft3a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 533.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 533.4 46.2 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ho9kxtcd",
   "source_url": "https://omlx.ai/benchmarks/performance/ho9kxtcd",
   "tg_tokens": null,
   "tps": 46.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-ho9kxtcd-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.0 GB at this context; duplicate traces of the same session: 7ze0gewp, a3cogmu8, gcqz7pth, io4qmtkh, k4zjft3a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 674.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 674.6 31.4 72.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ho9kxtcd",
   "source_url": "https://omlx.ai/benchmarks/performance/ho9kxtcd",
   "tg_tokens": null,
   "tps": 31.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-ho9kxtcd-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.5 GB at this context; duplicate traces of the same session: 7ze0gewp, a3cogmu8, gcqz7pth, io4qmtkh, k4zjft3a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 650.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 650.8 21.5 72.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ho9kxtcd",
   "source_url": "https://omlx.ai/benchmarks/performance/ho9kxtcd",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ho9kxtcd-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.7 GB at this context; duplicate traces of the same session: 7ze0gewp, a3cogmu8, gcqz7pth, io4qmtkh, k4zjft3a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 649.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 649.4 44.8 73.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ho9kxtcd",
   "source_url": "https://omlx.ai/benchmarks/performance/ho9kxtcd",
   "tg_tokens": null,
   "tps": 44.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ho9kxtcd-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 76.3 GB at this context; duplicate traces of the same session: 7ze0gewp, a3cogmu8, gcqz7pth, io4qmtkh, k4zjft3a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 611.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 611.0 40.4 76.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ho9kxtcd",
   "source_url": "https://omlx.ai/benchmarks/performance/ho9kxtcd",
   "tg_tokens": null,
   "tps": 40.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-ho9kxtcd-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.9 GB at this context; duplicate traces of the same session: 7ze0gewp, a3cogmu8, gcqz7pth, io4qmtkh, k4zjft3a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 631.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 631.1 35.0 78.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ho9kxtcd",
   "source_url": "https://omlx.ai/benchmarks/performance/ho9kxtcd",
   "tg_tokens": null,
   "tps": 35.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-3h1hm9hw-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 623.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 623.6 50.2 100.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3h1hm9hw",
   "source_url": "https://omlx.ai/benchmarks/performance/3h1hm9hw",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-3h1hm9hw-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 830.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 830.3 45.6 101.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3h1hm9hw",
   "source_url": "https://omlx.ai/benchmarks/performance/3h1hm9hw",
   "tg_tokens": null,
   "tps": 45.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-uuavdfty-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.5 GB at this context; duplicate traces of the same session: l4jdhim9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 363.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 363.8 42.5 70.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uuavdfty",
   "source_url": "https://omlx.ai/benchmarks/performance/uuavdfty",
   "tg_tokens": null,
   "tps": 42.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-uuavdfty-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.9 GB at this context; duplicate traces of the same session: l4jdhim9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2111.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,111 17.1 72.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uuavdfty",
   "source_url": "https://omlx.ai/benchmarks/performance/uuavdfty",
   "tg_tokens": null,
   "tps": 17.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0d5yd867-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.5 GB at this context; duplicate traces of the same session: nm2upyiq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1199.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,199 86.6 70.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0d5yd867",
   "source_url": "https://omlx.ai/benchmarks/performance/0d5yd867",
   "tg_tokens": null,
   "tps": 86.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0d5yd867-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: nm2upyiq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2111.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,111 78.4 71.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0d5yd867",
   "source_url": "https://omlx.ai/benchmarks/performance/0d5yd867",
   "tg_tokens": null,
   "tps": 78.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-88kf9qjj-ctx8192",
   "model": "Spark-X2.5-4B",
   "notes": "OptiQ build; oMLX community benchmark trace, single run at 8k context; peak memory 4.0 GB; TTFT 18097 ms; duplicate traces of the same session: none; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 452.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Spark-X2.5-4B-OptiQ-4bit on M3 Pro (14c); M3 Pro (14c) \u00b7 18 GB \u00b7 4bit; 8k tokens 452.7 PP tok/s 33.0 TG tok/s 18097 TTFT (ms) 4.0 Peak mem",
   "ram_gb": 18,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/88kf9qjj",
   "source_url": "https://omlx.ai/benchmarks/performance/88kf9qjj",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-8avy0hf5-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: qi221ax5; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 104.3 18.8 19.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8avy0hf5",
   "source_url": "https://omlx.ai/benchmarks/performance/8avy0hf5",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-8avy0hf5-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "orcarouter MLX build; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: qi221ax5; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 94.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "orcarouter--Qwen3.8-27B-Uncensored-MLX on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 94.9 17.1 22.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8avy0hf5",
   "source_url": "https://omlx.ai/benchmarks/performance/8avy0hf5",
   "tg_tokens": null,
   "tps": 17.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-lg34wu1e-ctx4096",
   "model": "Mistral-Medium-3.5-128B",
   "notes": "oMLX community benchmark trace; peak memory 70.0 GB at this context; duplicate traces of the same session: r18kfjyn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 381.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Mistral-Medium-3.5-128B-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 381.3 13.7 70.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lg34wu1e",
   "source_url": "https://omlx.ai/benchmarks/performance/lg34wu1e",
   "tg_tokens": null,
   "tps": 13.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-lg34wu1e-ctx16384",
   "model": "Mistral-Medium-3.5-128B",
   "notes": "oMLX community benchmark trace; peak memory 74.1 GB at this context; duplicate traces of the same session: r18kfjyn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 357.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Mistral-Medium-3.5-128B-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 357.7 12.0 74.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lg34wu1e",
   "source_url": "https://omlx.ai/benchmarks/performance/lg34wu1e",
   "tg_tokens": null,
   "tps": 12.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-7gmjizjj-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 68.0 GB at this context; duplicate traces of the same session: 22ht10r8; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1193.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Ornith-1.5-35B-A3B-MLX on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 unknown; 8k 1,193 43.3 68.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7gmjizjj",
   "source_url": "https://omlx.ai/benchmarks/performance/7gmjizjj",
   "tg_tokens": null,
   "tps": 43.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-7gmjizjj-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 68.5 GB at this context; duplicate traces of the same session: 22ht10r8; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1021.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Ornith-1.5-35B-A3B-MLX on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 unknown; 16k 1,021 41.0 68.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7gmjizjj",
   "source_url": "https://omlx.ai/benchmarks/performance/7gmjizjj",
   "tg_tokens": null,
   "tps": 41.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-f301tt1d-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 218 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4690.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 4,690 PP tok/s 147.4 TG tok/s 218 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f301tt1d",
   "source_url": "https://omlx.ai/benchmarks/performance/f301tt1d",
   "tg_tokens": null,
   "tps": 147.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-pkzedvud-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: py40twox; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 815.8,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 815.8 65.1 21.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pkzedvud",
   "source_url": "https://omlx.ai/benchmarks/performance/pkzedvud",
   "tg_tokens": null,
   "tps": 65.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-pkzedvud-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: py40twox; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 926.3,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 926.3 56.7 22.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pkzedvud",
   "source_url": "https://omlx.ai/benchmarks/performance/pkzedvud",
   "tg_tokens": null,
   "tps": 56.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-vty5let8-ctx4096",
   "model": "Qwen2.5-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 6.1 GB at this context; duplicate traces of the same session: wzny1eoo; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 247.4,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "Qwen2.5-7B-Instruct-AWQ on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 247.4 21.1 6.1 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vty5let8",
   "source_url": "https://omlx.ai/benchmarks/performance/vty5let8",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-vty5let8-ctx8192",
   "model": "Qwen2.5-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: wzny1eoo; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 235.4,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "Qwen2.5-7B-Instruct-AWQ on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 8k 235.4 20.2 6.3 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vty5let8",
   "source_url": "https://omlx.ai/benchmarks/performance/vty5let8",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-z53rnwrd-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: bzl2cxxn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4544.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,544 227.6 21.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z53rnwrd",
   "source_url": "https://omlx.ai/benchmarks/performance/z53rnwrd",
   "tg_tokens": null,
   "tps": 227.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-z53rnwrd-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: bzl2cxxn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6346.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,346 191.2 22.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z53rnwrd",
   "source_url": "https://omlx.ai/benchmarks/performance/z53rnwrd",
   "tg_tokens": null,
   "tps": 191.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tck78doy-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: hmntmaxe; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4528.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,528 183.2 21.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tck78doy",
   "source_url": "https://omlx.ai/benchmarks/performance/tck78doy",
   "tg_tokens": null,
   "tps": 183.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-tck78doy-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: hmntmaxe; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6198.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,198 184.4 22.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tck78doy",
   "source_url": "https://omlx.ai/benchmarks/performance/tck78doy",
   "tg_tokens": null,
   "tps": 184.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-frcopq1o-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 1k context; peak memory 20.1 GB; TTFT 215 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4766.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 4,766 PP tok/s 140.6 TG tok/s 215 TTFT (ms) 20.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/frcopq1o",
   "source_url": "https://omlx.ai/benchmarks/performance/frcopq1o",
   "tg_tokens": null,
   "tps": 140.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-1lnsatgc-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 38.3 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 83.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 8bit; 1k 83.4 117.6 38.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1lnsatgc",
   "source_url": "https://omlx.ai/benchmarks/performance/1lnsatgc",
   "tg_tokens": null,
   "tps": 117.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-1lnsatgc-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 40.0 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2283.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 8bit; 4k 2,283 68.0 40.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1lnsatgc",
   "source_url": "https://omlx.ai/benchmarks/performance/1lnsatgc",
   "tg_tokens": null,
   "tps": 68.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-1lnsatgc-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 40.4 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2148.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 8bit; 8k 2,148 105.4 40.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1lnsatgc",
   "source_url": "https://omlx.ai/benchmarks/performance/1lnsatgc",
   "tg_tokens": null,
   "tps": 105.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-1lnsatgc-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 40.9 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1890.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 8bit; 16k 1,890 119.3 40.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1lnsatgc",
   "source_url": "https://omlx.ai/benchmarks/performance/1lnsatgc",
   "tg_tokens": null,
   "tps": 119.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-1lnsatgc-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 42.0 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1511.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 8bit; 32k 1,511 130.3 42.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1lnsatgc",
   "source_url": "https://omlx.ai/benchmarks/performance/1lnsatgc",
   "tg_tokens": null,
   "tps": 130.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-1lnsatgc-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 43.3 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1074.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 8bit; 64k 1,074 92.9 43.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1lnsatgc",
   "source_url": "https://omlx.ai/benchmarks/performance/1lnsatgc",
   "tg_tokens": null,
   "tps": 92.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-1lnsatgc-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 44.8 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 659.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 8bit; 128k 659.7 66.2 44.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1lnsatgc",
   "source_url": "https://omlx.ai/benchmarks/performance/1lnsatgc",
   "tg_tokens": null,
   "tps": 66.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-1lnsatgc-ctx199680",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 46.4 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 474.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 8bit; 195k 474.7 48.7 46.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1lnsatgc",
   "source_url": "https://omlx.ai/benchmarks/performance/1lnsatgc",
   "tg_tokens": null,
   "tps": 48.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-999ydsz8-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 17.8 GB; TTFT 18799 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 871.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k tokens 871.5 PP tok/s 30.7 TG tok/s 18799 TTFT (ms) 17.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/999ydsz8",
   "source_url": "https://omlx.ai/benchmarks/performance/999ydsz8",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qhnw5fya-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.5 GB at this context; duplicate traces of the same session: 78c0lfam; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 493.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 493.1 26.6 17.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qhnw5fya",
   "source_url": "https://omlx.ai/benchmarks/performance/qhnw5fya",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qhnw5fya-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: 78c0lfam; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 485.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 485.1 26.7 18.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qhnw5fya",
   "source_url": "https://omlx.ai/benchmarks/performance/qhnw5fya",
   "tg_tokens": null,
   "tps": 26.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3vpicb3a-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: w4ez9d28; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 927.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 927.3 18.0 29.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3vpicb3a",
   "source_url": "https://omlx.ai/benchmarks/performance/3vpicb3a",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3vpicb3a-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: w4ez9d28; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 738.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 738.0 17.1 30.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3vpicb3a",
   "source_url": "https://omlx.ai/benchmarks/performance/3vpicb3a",
   "tg_tokens": null,
   "tps": 17.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-w7qyhgwy-ctx1024",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "ET-O5.5-G6A-G4.7-DSV4-K3-RLOO MTP DFlash2 build; oMLX community benchmark trace, single run at 1k context; peak memory 16.8 GB; TTFT 3935 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 260.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-ET-O5.5-G6A-G4.7-DSV4-K3-RLOO-MTP-DFlash2-oQ4e on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k tokens 260.2 PP tok/s 27.7 TG tok/s 3935 TTFT (ms) 16.8 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w7qyhgwy",
   "source_url": "https://omlx.ai/benchmarks/performance/w7qyhgwy",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0k9awsnf-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 3na830w8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1130.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,130 63.6 17.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0k9awsnf",
   "source_url": "https://omlx.ai/benchmarks/performance/0k9awsnf",
   "tg_tokens": null,
   "tps": 63.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0k9awsnf-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 3na830w8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1361.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,361 68.5 19.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0k9awsnf",
   "source_url": "https://omlx.ai/benchmarks/performance/0k9awsnf",
   "tg_tokens": null,
   "tps": 68.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hpjqfujr-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 9xth3war; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 816.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 816.1 53.5 17.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hpjqfujr",
   "source_url": "https://omlx.ai/benchmarks/performance/hpjqfujr",
   "tg_tokens": null,
   "tps": 53.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hpjqfujr-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 9xth3war; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1278.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,278 64.7 19.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hpjqfujr",
   "source_url": "https://omlx.ai/benchmarks/performance/hpjqfujr",
   "tg_tokens": null,
   "tps": 64.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-z2ux9xc2-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 40vl5jt5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1123.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,123 54.3 17.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z2ux9xc2",
   "source_url": "https://omlx.ai/benchmarks/performance/z2ux9xc2",
   "tg_tokens": null,
   "tps": 54.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-z2ux9xc2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 40vl5jt5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1349.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,349 65.4 19.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z2ux9xc2",
   "source_url": "https://omlx.ai/benchmarks/performance/z2ux9xc2",
   "tg_tokens": null,
   "tps": 65.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-tw577wpf-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP320",
   "notes": "MTP Vision build; oMLX community benchmark trace, single run at 4k context; peak memory 37.9 GB; TTFT 6139 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 667.2,
   "provenance": "community",
   "quant": "oQ3e-DWQ",
   "quote": "Qwen3.8-Flash-Next-REAP320-oQ3e-DWQ-MTP-Vision-MLX on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 3bit; 4k tokens 667.2 PP tok/s 26.6 TG tok/s 6139 TTFT (ms) 37.9 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tw577wpf",
   "source_url": "https://omlx.ai/benchmarks/performance/tw577wpf",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-13tzc8l0-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: 40jg8fdd, jfbtbkdq, kwelso2p, olmj7k38, ri4au3bq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1068.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,068 53.0 71.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/13tzc8l0",
   "source_url": "https://omlx.ai/benchmarks/performance/13tzc8l0",
   "tg_tokens": null,
   "tps": 53.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-13tzc8l0-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.9 GB at this context; duplicate traces of the same session: 40jg8fdd, jfbtbkdq, kwelso2p, olmj7k38, ri4au3bq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1063.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 1,063 88.3 71.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/13tzc8l0",
   "source_url": "https://omlx.ai/benchmarks/performance/13tzc8l0",
   "tg_tokens": null,
   "tps": 88.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-13tzc8l0-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.5 GB at this context; duplicate traces of the same session: 40jg8fdd, jfbtbkdq, kwelso2p, olmj7k38, ri4au3bq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1056.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 1,056 81.1 72.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/13tzc8l0",
   "source_url": "https://omlx.ai/benchmarks/performance/13tzc8l0",
   "tg_tokens": null,
   "tps": 81.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-13tzc8l0-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.7 GB at this context; duplicate traces of the same session: 40jg8fdd, jfbtbkdq, kwelso2p, olmj7k38, ri4au3bq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1044.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 1,044 73.2 73.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/13tzc8l0",
   "source_url": "https://omlx.ai/benchmarks/performance/13tzc8l0",
   "tg_tokens": null,
   "tps": 73.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-13tzc8l0-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 77.0 GB at this context; duplicate traces of the same session: 40jg8fdd, jfbtbkdq, kwelso2p, olmj7k38, ri4au3bq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1026.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 128k 1,026 57.9 77.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/13tzc8l0",
   "source_url": "https://omlx.ai/benchmarks/performance/13tzc8l0",
   "tg_tokens": null,
   "tps": 57.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-13tzc8l0-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.0 GB at this context; duplicate traces of the same session: 40jg8fdd, jfbtbkdq, kwelso2p, olmj7k38, ri4au3bq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 825.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 195k 825.6 50.9 79.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/13tzc8l0",
   "source_url": "https://omlx.ai/benchmarks/performance/13tzc8l0",
   "tg_tokens": null,
   "tps": 50.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gaxdx18e-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.1 GB at this context; duplicate traces of the same session: 2eworv1e, i3cyt0v6, q33inswa, rtp3iayt, vlxjsipl; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 811.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 811.7 87.4 99.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gaxdx18e",
   "source_url": "https://omlx.ai/benchmarks/performance/gaxdx18e",
   "tg_tokens": null,
   "tps": 87.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gaxdx18e-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.7 GB at this context; duplicate traces of the same session: 2eworv1e, i3cyt0v6, q33inswa, rtp3iayt, vlxjsipl; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1096.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,096 42.9 99.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gaxdx18e",
   "source_url": "https://omlx.ai/benchmarks/performance/gaxdx18e",
   "tg_tokens": null,
   "tps": 42.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gaxdx18e-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.7 GB at this context; duplicate traces of the same session: 2eworv1e, i3cyt0v6, q33inswa, rtp3iayt, vlxjsipl; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1419.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 1,419 79.3 99.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gaxdx18e",
   "source_url": "https://omlx.ai/benchmarks/performance/gaxdx18e",
   "tg_tokens": null,
   "tps": 79.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gaxdx18e-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.0 GB at this context; duplicate traces of the same session: 2eworv1e, i3cyt0v6, q33inswa, rtp3iayt, vlxjsipl; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1411.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 1,411 78.1 100.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gaxdx18e",
   "source_url": "https://omlx.ai/benchmarks/performance/gaxdx18e",
   "tg_tokens": null,
   "tps": 78.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gaxdx18e-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: 2eworv1e, i3cyt0v6, q33inswa, rtp3iayt, vlxjsipl; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1402.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 1,402 26.0 100.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gaxdx18e",
   "source_url": "https://omlx.ai/benchmarks/performance/gaxdx18e",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gaxdx18e-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: 2eworv1e, i3cyt0v6, q33inswa, rtp3iayt, vlxjsipl; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1392.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 1,392 58.5 101.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gaxdx18e",
   "source_url": "https://omlx.ai/benchmarks/performance/gaxdx18e",
   "tg_tokens": null,
   "tps": 58.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-wnetntwi-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: eh10z5hq, i8o1qix8, mvown3am, preeo9r6, zd91zyvr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 543.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 543.5 69.4 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wnetntwi",
   "source_url": "https://omlx.ai/benchmarks/performance/wnetntwi",
   "tg_tokens": null,
   "tps": 69.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-wnetntwi-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: eh10z5hq, i8o1qix8, mvown3am, preeo9r6, zd91zyvr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 651.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 651.1 62.2 71.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wnetntwi",
   "source_url": "https://omlx.ai/benchmarks/performance/wnetntwi",
   "tg_tokens": null,
   "tps": 62.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-wnetntwi-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: eh10z5hq, i8o1qix8, mvown3am, preeo9r6, zd91zyvr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 663.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 663.7 71.6 71.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wnetntwi",
   "source_url": "https://omlx.ai/benchmarks/performance/wnetntwi",
   "tg_tokens": null,
   "tps": 71.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-wnetntwi-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: eh10z5hq, i8o1qix8, mvown3am, preeo9r6, zd91zyvr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 658.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 658.9 70.1 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wnetntwi",
   "source_url": "https://omlx.ai/benchmarks/performance/wnetntwi",
   "tg_tokens": null,
   "tps": 70.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-wnetntwi-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.4 GB at this context; duplicate traces of the same session: eh10z5hq, i8o1qix8, mvown3am, preeo9r6, zd91zyvr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 648.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 648.7 36.6 72.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wnetntwi",
   "source_url": "https://omlx.ai/benchmarks/performance/wnetntwi",
   "tg_tokens": null,
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-wnetntwi-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.3 GB at this context; duplicate traces of the same session: eh10z5hq, i8o1qix8, mvown3am, preeo9r6, zd91zyvr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 649.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 649.0 49.2 74.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wnetntwi",
   "source_url": "https://omlx.ai/benchmarks/performance/wnetntwi",
   "tg_tokens": null,
   "tps": 49.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-oays6i4v-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 78.4 GB; TTFT 48150 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2722.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k tokens 2,722 PP tok/s 63.7 TG tok/s 48150 TTFT (ms) 78.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oays6i4v",
   "source_url": "https://omlx.ai/benchmarks/performance/oays6i4v",
   "tg_tokens": null,
   "tps": 63.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fisrxh3n-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 73.5 GB; TTFT 12851 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2550.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k tokens 2,550 PP tok/s 65.0 TG tok/s 12851 TTFT (ms) 73.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fisrxh3n",
   "source_url": "https://omlx.ai/benchmarks/performance/fisrxh3n",
   "tg_tokens": null,
   "tps": 65.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-3sj9lexf-ctx1024",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace, single run at 1k context; peak memory 11.5 GB; TTFT 22783 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 44.9,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 2bit; 1k tokens 44.9 PP tok/s 8.7 TG tok/s 22783 TTFT (ms) 11.5 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3sj9lexf",
   "source_url": "https://omlx.ai/benchmarks/performance/3sj9lexf",
   "tg_tokens": null,
   "tps": 8.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-d1mipugd-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 099wqtr9, cvq74iw1, h3ic746a, olacg0zw, ylaktj42, yti5vm27; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 912.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 912.5 95.1 21.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d1mipugd",
   "source_url": "https://omlx.ai/benchmarks/performance/d1mipugd",
   "tg_tokens": null,
   "tps": 95.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-d1mipugd-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: 099wqtr9, cvq74iw1, h3ic746a, olacg0zw, ylaktj42, yti5vm27; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1052.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,052 93.6 23.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d1mipugd",
   "source_url": "https://omlx.ai/benchmarks/performance/d1mipugd",
   "tg_tokens": null,
   "tps": 93.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-d1mipugd-ctx8192",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: 099wqtr9, cvq74iw1, h3ic746a, olacg0zw, ylaktj42, yti5vm27; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 989.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 989.0 99.2 23.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d1mipugd",
   "source_url": "https://omlx.ai/benchmarks/performance/d1mipugd",
   "tg_tokens": null,
   "tps": 99.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-d1mipugd-ctx16384",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: 099wqtr9, cvq74iw1, h3ic746a, olacg0zw, ylaktj42, yti5vm27; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 876.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 876.3 106.6 24.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d1mipugd",
   "source_url": "https://omlx.ai/benchmarks/performance/d1mipugd",
   "tg_tokens": null,
   "tps": 106.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-d1mipugd-ctx32768",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: 099wqtr9, cvq74iw1, h3ic746a, olacg0zw, ylaktj42, yti5vm27; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 694.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 694.7 75.3 25.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d1mipugd",
   "source_url": "https://omlx.ai/benchmarks/performance/d1mipugd",
   "tg_tokens": null,
   "tps": 75.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-d1mipugd-ctx65536",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: 099wqtr9, cvq74iw1, h3ic746a, olacg0zw, ylaktj42, yti5vm27; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 502.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 502.8 55.4 26.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d1mipugd",
   "source_url": "https://omlx.ai/benchmarks/performance/d1mipugd",
   "tg_tokens": null,
   "tps": 55.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-d1mipugd-ctx131072",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 28.3 GB at this context; duplicate traces of the same session: 099wqtr9, cvq74iw1, h3ic746a, olacg0zw, ylaktj42, yti5vm27; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 335.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 128k 335.1 47.6 28.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d1mipugd",
   "source_url": "https://omlx.ai/benchmarks/performance/d1mipugd",
   "tg_tokens": null,
   "tps": 47.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-fh3cduvg-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 5.7 GB at this context; duplicate traces of the same session: 3614pr8g; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 602.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 602.2 30.7 5.7 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fh3cduvg",
   "source_url": "https://omlx.ai/benchmarks/performance/fh3cduvg",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-fh3cduvg-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: 3614pr8g; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 645.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 645.5 46.2 5.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fh3cduvg",
   "source_url": "https://omlx.ai/benchmarks/performance/fh3cduvg",
   "tg_tokens": null,
   "tps": 46.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-wfz3xj4y-ctx16384",
   "model": "LFM2.5-1.2B-Instruct",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 1.7 GB at this context; duplicate traces of the same session: xgs9fgbl; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3003.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-1.2B-Instruct-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 3,003 219.1 1.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wfz3xj4y",
   "source_url": "https://omlx.ai/benchmarks/performance/wfz3xj4y",
   "tg_tokens": null,
   "tps": 219.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-wfz3xj4y-ctx32768",
   "model": "LFM2.5-1.2B-Instruct",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 1.8 GB at this context; duplicate traces of the same session: xgs9fgbl; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2627.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-1.2B-Instruct-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 32k 2,627 160.8 1.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wfz3xj4y",
   "source_url": "https://omlx.ai/benchmarks/performance/wfz3xj4y",
   "tg_tokens": null,
   "tps": 160.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-gm4uwy2b-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: 27d69exj; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1151.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 32k 1,151 62.5 25.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gm4uwy2b",
   "source_url": "https://omlx.ai/benchmarks/performance/gm4uwy2b",
   "tg_tokens": null,
   "tps": 62.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-gm4uwy2b-ctx65536",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: 27d69exj; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 961.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 64k 961.0 50.9 26.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gm4uwy2b",
   "source_url": "https://omlx.ai/benchmarks/performance/gm4uwy2b",
   "tg_tokens": null,
   "tps": 50.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-2x10qok4-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: 6ts1bkbh; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1329.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 8k 1,329 77.0 24.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2x10qok4",
   "source_url": "https://omlx.ai/benchmarks/performance/2x10qok4",
   "tg_tokens": null,
   "tps": 77.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-2x10qok4-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: 6ts1bkbh; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1272.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 16k 1,272 71.9 25.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2x10qok4",
   "source_url": "https://omlx.ai/benchmarks/performance/2x10qok4",
   "tg_tokens": null,
   "tps": 71.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-pfoviyfl-ctx1024",
   "model": "Muse-Spark-Open-816B-A42B",
   "notes": "oMLX community benchmark trace; peak memory 341.8 GB at this context; duplicate traces of the same session: 6uu0xznn, c35yyee7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 212.0,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Muse-Spark-Open-816B-A42B-oQ3e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 3bit; 1k 212.0 21.2 341.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pfoviyfl",
   "source_url": "https://omlx.ai/benchmarks/performance/pfoviyfl",
   "tg_tokens": null,
   "tps": 21.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-pfoviyfl-ctx4096",
   "model": "Muse-Spark-Open-816B-A42B",
   "notes": "oMLX community benchmark trace; peak memory 342.8 GB at this context; duplicate traces of the same session: 6uu0xznn, c35yyee7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 227.2,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Muse-Spark-Open-816B-A42B-oQ3e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 3bit; 4k 227.2 20.8 342.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pfoviyfl",
   "source_url": "https://omlx.ai/benchmarks/performance/pfoviyfl",
   "tg_tokens": null,
   "tps": 20.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-pfoviyfl-ctx8192",
   "model": "Muse-Spark-Open-816B-A42B",
   "notes": "oMLX community benchmark trace; peak memory 343.0 GB at this context; duplicate traces of the same session: 6uu0xznn, c35yyee7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.3,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Muse-Spark-Open-816B-A42B-oQ3e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 3bit; 8k 224.3 20.6 343.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pfoviyfl",
   "source_url": "https://omlx.ai/benchmarks/performance/pfoviyfl",
   "tg_tokens": null,
   "tps": 20.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-6xa7u5l5-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 6702p9x0, paz3ukir; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2657.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 2,657 96.6 21.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6xa7u5l5",
   "source_url": "https://omlx.ai/benchmarks/performance/6xa7u5l5",
   "tg_tokens": null,
   "tps": 96.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-6xa7u5l5-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 6702p9x0, paz3ukir; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2604.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 2,604 95.6 21.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6xa7u5l5",
   "source_url": "https://omlx.ai/benchmarks/performance/6xa7u5l5",
   "tg_tokens": null,
   "tps": 95.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-6xa7u5l5-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 6702p9x0, paz3ukir; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2336.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 2,336 95.4 21.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6xa7u5l5",
   "source_url": "https://omlx.ai/benchmarks/performance/6xa7u5l5",
   "tg_tokens": null,
   "tps": 95.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3lbd8vpm-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 6mt1la9y, yfhs2af1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2722.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 2,722 99.0 21.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3lbd8vpm",
   "source_url": "https://omlx.ai/benchmarks/performance/3lbd8vpm",
   "tg_tokens": null,
   "tps": 99.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3lbd8vpm-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 6mt1la9y, yfhs2af1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2604.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 2,604 98.2 21.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3lbd8vpm",
   "source_url": "https://omlx.ai/benchmarks/performance/3lbd8vpm",
   "tg_tokens": null,
   "tps": 98.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3lbd8vpm-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 6mt1la9y, yfhs2af1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2336.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 2,336 95.2 21.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3lbd8vpm",
   "source_url": "https://omlx.ai/benchmarks/performance/3lbd8vpm",
   "tg_tokens": null,
   "tps": 95.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-0zubdwcz-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: jd21pk6z; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 951.5,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 951.5 62.4 24.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0zubdwcz",
   "source_url": "https://omlx.ai/benchmarks/performance/0zubdwcz",
   "tg_tokens": null,
   "tps": 62.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-0zubdwcz-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: jd21pk6z; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 817.2,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 817.2 58.9 24.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0zubdwcz",
   "source_url": "https://omlx.ai/benchmarks/performance/0zubdwcz",
   "tg_tokens": null,
   "tps": 58.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-e92ss6j7-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 8ulpduq0, umfj03b3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2729.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 2,729 94.5 21.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e92ss6j7",
   "source_url": "https://omlx.ai/benchmarks/performance/e92ss6j7",
   "tg_tokens": null,
   "tps": 94.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-e92ss6j7-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 8ulpduq0, umfj03b3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2601.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 2,601 93.5 21.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e92ss6j7",
   "source_url": "https://omlx.ai/benchmarks/performance/e92ss6j7",
   "tg_tokens": null,
   "tps": 93.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-e92ss6j7-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 8ulpduq0, umfj03b3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2332.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 2,332 92.3 22.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e92ss6j7",
   "source_url": "https://omlx.ai/benchmarks/performance/e92ss6j7",
   "tg_tokens": null,
   "tps": 92.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-9g8g3gid-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: m1wbe6gi; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 624.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 1k 624.8 54.0 3.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9g8g3gid",
   "source_url": "https://omlx.ai/benchmarks/performance/9g8g3gid",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-9g8g3gid-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.7 GB at this context; duplicate traces of the same session: m1wbe6gi; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 588.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 4k 588.2 48.6 4.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9g8g3gid",
   "source_url": "https://omlx.ai/benchmarks/performance/9g8g3gid",
   "tg_tokens": null,
   "tps": 48.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ba0f2x7o-ctx1024",
   "model": "Qwen3.6-35B-A3B-Abliterated-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: tdw4bp3h; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 712.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Abliterated-Heretic-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 712.6 96.9 23.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ba0f2x7o",
   "source_url": "https://omlx.ai/benchmarks/performance/ba0f2x7o",
   "tg_tokens": null,
   "tps": 96.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ba0f2x7o-ctx4096",
   "model": "Qwen3.6-35B-A3B-Abliterated-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: tdw4bp3h; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1759.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Abliterated-Heretic-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,759 93.9 24.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ba0f2x7o",
   "source_url": "https://omlx.ai/benchmarks/performance/ba0f2x7o",
   "tg_tokens": null,
   "tps": 93.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-vjmh3oaz-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 6ga08oxk, 9u3810bj; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2731.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 2,731 93.6 21.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vjmh3oaz",
   "source_url": "https://omlx.ai/benchmarks/performance/vjmh3oaz",
   "tg_tokens": null,
   "tps": 93.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-vjmh3oaz-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 6ga08oxk, 9u3810bj; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2604.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 2,604 93.2 21.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vjmh3oaz",
   "source_url": "https://omlx.ai/benchmarks/performance/vjmh3oaz",
   "tg_tokens": null,
   "tps": 93.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-vjmh3oaz-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 6ga08oxk, 9u3810bj; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2334.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 2,334 91.7 22.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vjmh3oaz",
   "source_url": "https://omlx.ai/benchmarks/performance/vjmh3oaz",
   "tg_tokens": null,
   "tps": 91.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-welttbj0-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 38.3 GB at this context; duplicate traces of the same session: 84yqdg6u, e8aqivc2, hkb7y92d, j6gvuxuo, ohtsd1vh, y9ztw8jz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 83.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 8bit; 1k 83.4 117.6 38.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/welttbj0",
   "source_url": "https://omlx.ai/benchmarks/performance/welttbj0",
   "tg_tokens": null,
   "tps": 117.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-welttbj0-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 40.0 GB at this context; duplicate traces of the same session: 84yqdg6u, e8aqivc2, hkb7y92d, j6gvuxuo, ohtsd1vh, y9ztw8jz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2283.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 8bit; 4k 2,283 68.0 40.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/welttbj0",
   "source_url": "https://omlx.ai/benchmarks/performance/welttbj0",
   "tg_tokens": null,
   "tps": 68.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-welttbj0-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 40.4 GB at this context; duplicate traces of the same session: 84yqdg6u, e8aqivc2, hkb7y92d, j6gvuxuo, ohtsd1vh, y9ztw8jz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2148.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 8bit; 8k 2,148 105.4 40.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/welttbj0",
   "source_url": "https://omlx.ai/benchmarks/performance/welttbj0",
   "tg_tokens": null,
   "tps": 105.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-welttbj0-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 40.9 GB at this context; duplicate traces of the same session: 84yqdg6u, e8aqivc2, hkb7y92d, j6gvuxuo, ohtsd1vh, y9ztw8jz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1890.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 8bit; 16k 1,890 119.3 40.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/welttbj0",
   "source_url": "https://omlx.ai/benchmarks/performance/welttbj0",
   "tg_tokens": null,
   "tps": 119.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-welttbj0-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 42.0 GB at this context; duplicate traces of the same session: 84yqdg6u, e8aqivc2, hkb7y92d, j6gvuxuo, ohtsd1vh, y9ztw8jz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1511.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 8bit; 32k 1,511 130.3 42.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/welttbj0",
   "source_url": "https://omlx.ai/benchmarks/performance/welttbj0",
   "tg_tokens": null,
   "tps": 130.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-welttbj0-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 43.3 GB at this context; duplicate traces of the same session: 84yqdg6u, e8aqivc2, hkb7y92d, j6gvuxuo, ohtsd1vh, y9ztw8jz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1074.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 8bit; 64k 1,074 92.9 43.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/welttbj0",
   "source_url": "https://omlx.ai/benchmarks/performance/welttbj0",
   "tg_tokens": null,
   "tps": 92.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-welttbj0-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 44.8 GB at this context; duplicate traces of the same session: 84yqdg6u, e8aqivc2, hkb7y92d, j6gvuxuo, ohtsd1vh, y9ztw8jz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 659.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 8bit; 128k 659.7 66.2 44.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/welttbj0",
   "source_url": "https://omlx.ai/benchmarks/performance/welttbj0",
   "tg_tokens": null,
   "tps": 66.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-welttbj0-ctx199680",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 46.4 GB at this context; duplicate traces of the same session: 84yqdg6u, e8aqivc2, hkb7y92d, j6gvuxuo, ohtsd1vh, y9ztw8jz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 474.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 8bit; 195k 474.7 48.7 46.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/welttbj0",
   "source_url": "https://omlx.ai/benchmarks/performance/welttbj0",
   "tg_tokens": null,
   "tps": 48.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-17jt16po-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 7c3g4h2b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 999.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 999.6 32.9 17.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/17jt16po",
   "source_url": "https://omlx.ai/benchmarks/performance/17jt16po",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-17jt16po-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 7c3g4h2b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 821.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 821.3 29.6 18.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/17jt16po",
   "source_url": "https://omlx.ai/benchmarks/performance/17jt16po",
   "tg_tokens": null,
   "tps": 29.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zau372kr-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.8 GB at this context; duplicate traces of the same session: 0621wvc5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 935.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 935.0 18.3 29.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zau372kr",
   "source_url": "https://omlx.ai/benchmarks/performance/zau372kr",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zau372kr-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.3 GB at this context; duplicate traces of the same session: 0621wvc5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 721.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 32k 721.0 17.2 31.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zau372kr",
   "source_url": "https://omlx.ai/benchmarks/performance/zau372kr",
   "tg_tokens": null,
   "tps": 17.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-a8tbiegb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: h0gl0fsr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 652.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 652.3 42.7 17.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a8tbiegb",
   "source_url": "https://omlx.ai/benchmarks/performance/a8tbiegb",
   "tg_tokens": null,
   "tps": 42.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-a8tbiegb-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: h0gl0fsr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 727.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 727.2 54.1 18.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a8tbiegb",
   "source_url": "https://omlx.ai/benchmarks/performance/a8tbiegb",
   "tg_tokens": null,
   "tps": 54.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9rmunnkh-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: twfqf2js; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 749.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 749.3 39.6 33.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9rmunnkh",
   "source_url": "https://omlx.ai/benchmarks/performance/9rmunnkh",
   "tg_tokens": null,
   "tps": 39.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9rmunnkh-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: twfqf2js; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 904.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 904.5 37.5 34.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9rmunnkh",
   "source_url": "https://omlx.ai/benchmarks/performance/9rmunnkh",
   "tg_tokens": null,
   "tps": 37.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-t3hvcjns-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: hgzho58j, ox6wdbqt, uyhkyzif; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1159.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 1,159 25.4 33.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t3hvcjns",
   "source_url": "https://omlx.ai/benchmarks/performance/t3hvcjns",
   "tg_tokens": null,
   "tps": 25.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-t3hvcjns-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: hgzho58j, ox6wdbqt, uyhkyzif; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1308.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 1,308 28.9 34.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t3hvcjns",
   "source_url": "https://omlx.ai/benchmarks/performance/t3hvcjns",
   "tg_tokens": null,
   "tps": 28.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-t3hvcjns-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: hgzho58j, ox6wdbqt, uyhkyzif; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1228.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 1,228 29.6 35.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t3hvcjns",
   "source_url": "https://omlx.ai/benchmarks/performance/t3hvcjns",
   "tg_tokens": null,
   "tps": 29.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-t3hvcjns-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.7 GB at this context; duplicate traces of the same session: hgzho58j, ox6wdbqt, uyhkyzif; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1020.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 1,020 28.2 35.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t3hvcjns",
   "source_url": "https://omlx.ai/benchmarks/performance/t3hvcjns",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nhsff4nb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: 007i6n4o, 8mr6bv67; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 993.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 993.8 43.5 33.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nhsff4nb",
   "source_url": "https://omlx.ai/benchmarks/performance/nhsff4nb",
   "tg_tokens": null,
   "tps": 43.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nhsff4nb-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: 007i6n4o, 8mr6bv67; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1254.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 1,254 38.5 34.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nhsff4nb",
   "source_url": "https://omlx.ai/benchmarks/performance/nhsff4nb",
   "tg_tokens": null,
   "tps": 38.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nhsff4nb-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: 007i6n4o, 8mr6bv67; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1215.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 1,215 45.7 35.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nhsff4nb",
   "source_url": "https://omlx.ai/benchmarks/performance/nhsff4nb",
   "tg_tokens": null,
   "tps": 45.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nhsff4nb-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.7 GB at this context; duplicate traces of the same session: 007i6n4o, 8mr6bv67; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1039.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 1,039 47.8 35.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nhsff4nb",
   "source_url": "https://omlx.ai/benchmarks/performance/nhsff4nb",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-0md196rj-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.4 GB at this context; duplicate traces of the same session: 0tmalke1, dduuo97w, gcd1zm2r, z8mi5h9i, zr6m7ils; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2752.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 2,752 120.5 100.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0md196rj",
   "source_url": "https://omlx.ai/benchmarks/performance/0md196rj",
   "tg_tokens": null,
   "tps": 120.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-0md196rj-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.8 GB at this context; duplicate traces of the same session: 0tmalke1, dduuo97w, gcd1zm2r, z8mi5h9i, zr6m7ils; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4529.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 4,529 101.4 101.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0md196rj",
   "source_url": "https://omlx.ai/benchmarks/performance/0md196rj",
   "tg_tokens": null,
   "tps": 101.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-0md196rj-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.6 GB at this context; duplicate traces of the same session: 0tmalke1, dduuo97w, gcd1zm2r, z8mi5h9i, zr6m7ils; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4729.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 4,729 131.3 102.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0md196rj",
   "source_url": "https://omlx.ai/benchmarks/performance/0md196rj",
   "tg_tokens": null,
   "tps": 131.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-0md196rj-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.2 GB at this context; duplicate traces of the same session: 0tmalke1, dduuo97w, gcd1zm2r, z8mi5h9i, zr6m7ils; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4750.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 4,750 122.8 103.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0md196rj",
   "source_url": "https://omlx.ai/benchmarks/performance/0md196rj",
   "tg_tokens": null,
   "tps": 122.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-0md196rj-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.4 GB at this context; duplicate traces of the same session: 0tmalke1, dduuo97w, gcd1zm2r, z8mi5h9i, zr6m7ils; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4659.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 4,659 113.8 104.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0md196rj",
   "source_url": "https://omlx.ai/benchmarks/performance/0md196rj",
   "tg_tokens": null,
   "tps": 113.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-0md196rj-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.0 GB at this context; duplicate traces of the same session: 0tmalke1, dduuo97w, gcd1zm2r, z8mi5h9i, zr6m7ils; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4527.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 4,527 109.9 109.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0md196rj",
   "source_url": "https://omlx.ai/benchmarks/performance/0md196rj",
   "tg_tokens": null,
   "tps": 109.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-0md196rj-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.7 GB at this context; duplicate traces of the same session: 0tmalke1, dduuo97w, gcd1zm2r, z8mi5h9i, zr6m7ils; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4389.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 4,389 99.1 110.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0md196rj",
   "source_url": "https://omlx.ai/benchmarks/performance/0md196rj",
   "tg_tokens": null,
   "tps": 99.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qhtmikof-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 3k82ovmp, hvqr17i2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1702.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 1,702 85.7 16.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qhtmikof",
   "source_url": "https://omlx.ai/benchmarks/performance/qhtmikof",
   "tg_tokens": null,
   "tps": 85.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qhtmikof-ctx8192",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: 3k82ovmp, hvqr17i2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1790.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,790 88.5 16.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qhtmikof",
   "source_url": "https://omlx.ai/benchmarks/performance/qhtmikof",
   "tg_tokens": null,
   "tps": 88.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-qhtmikof-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 3k82ovmp, hvqr17i2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1725.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 1,725 117.8 17.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qhtmikof",
   "source_url": "https://omlx.ai/benchmarks/performance/qhtmikof",
   "tg_tokens": null,
   "tps": 117.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-qsg5wlyd-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 4.7 GB at this context; duplicate traces of the same session: o4okeltj; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 877.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-4bit on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 1k 877.5 44.1 4.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qsg5wlyd",
   "source_url": "https://omlx.ai/benchmarks/performance/qsg5wlyd",
   "tg_tokens": null,
   "tps": 44.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-qsg5wlyd-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 4.7 GB at this context; duplicate traces of the same session: o4okeltj; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 931.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-4bit on M3 Max (30c); M3 Max (30c) \u00b7 96 GB \u00b7 4bit; 4k 931.1 43.4 4.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qsg5wlyd",
   "source_url": "https://omlx.ai/benchmarks/performance/qsg5wlyd",
   "tg_tokens": null,
   "tps": 43.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-168s5dve-ctx4096",
   "model": "Huihui-gemma-4-31B-it-v2",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 33.1 GB at this context; duplicate traces of the same session: d6ovealg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 670.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Huihui-gemma-4-31B-it-v2-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 670.5 14.7 33.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/168s5dve",
   "source_url": "https://omlx.ai/benchmarks/performance/168s5dve",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-168s5dve-ctx16384",
   "model": "Huihui-gemma-4-31B-it-v2",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: d6ovealg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 640.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Huihui-gemma-4-31B-it-v2-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 16k 640.6 13.4 34.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/168s5dve",
   "source_url": "https://omlx.ai/benchmarks/performance/168s5dve",
   "tg_tokens": null,
   "tps": 13.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-pbfs2e4k-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 3mkr1vt4, kpb2blmm, sqmvfpb2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3537.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "LFM2.5-24B-A2B-MLX-5bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 5bit; 1k 3,537 136.9 16.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pbfs2e4k",
   "source_url": "https://omlx.ai/benchmarks/performance/pbfs2e4k",
   "tg_tokens": null,
   "tps": 136.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-pbfs2e4k-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: 3mkr1vt4, kpb2blmm, sqmvfpb2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4142.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "LFM2.5-24B-A2B-MLX-5bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 5bit; 4k 4,142 135.1 16.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pbfs2e4k",
   "source_url": "https://omlx.ai/benchmarks/performance/pbfs2e4k",
   "tg_tokens": null,
   "tps": 135.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-pbfs2e4k-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: 3mkr1vt4, kpb2blmm, sqmvfpb2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4056.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "LFM2.5-24B-A2B-MLX-5bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 5bit; 8k 4,056 133.1 16.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pbfs2e4k",
   "source_url": "https://omlx.ai/benchmarks/performance/pbfs2e4k",
   "tg_tokens": null,
   "tps": 133.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-pbfs2e4k-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 3mkr1vt4, kpb2blmm, sqmvfpb2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3844.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "LFM2.5-24B-A2B-MLX-5bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 5bit; 16k 3,844 124.6 16.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pbfs2e4k",
   "source_url": "https://omlx.ai/benchmarks/performance/pbfs2e4k",
   "tg_tokens": null,
   "tps": 124.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-mxukse51-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 7mxou1z2, gnb64z0e; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3543.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "LFM2.5-24B-A2B-MLX-5bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 5bit; 1k 3,543 137.4 16.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mxukse51",
   "source_url": "https://omlx.ai/benchmarks/performance/mxukse51",
   "tg_tokens": null,
   "tps": 137.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-mxukse51-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: 7mxou1z2, gnb64z0e; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4142.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "LFM2.5-24B-A2B-MLX-5bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 5bit; 4k 4,142 135.4 16.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mxukse51",
   "source_url": "https://omlx.ai/benchmarks/performance/mxukse51",
   "tg_tokens": null,
   "tps": 135.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-mxukse51-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: 7mxou1z2, gnb64z0e; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4051.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "LFM2.5-24B-A2B-MLX-5bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 5bit; 8k 4,051 133.2 16.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mxukse51",
   "source_url": "https://omlx.ai/benchmarks/performance/mxukse51",
   "tg_tokens": null,
   "tps": 133.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-pn5e2my7-ctx8192",
   "model": "Mellum2-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 7.3 GB at this context; duplicate traces of the same session: 8tcjsd7h; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1186.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Mellum2-12B-A2.5B-Thinking-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 8k 1,186 116.0 7.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pn5e2my7",
   "source_url": "https://omlx.ai/benchmarks/performance/pn5e2my7",
   "tg_tokens": null,
   "tps": 116.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-pn5e2my7-ctx16384",
   "model": "Mellum2-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 7.3 GB at this context; duplicate traces of the same session: 8tcjsd7h; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1118.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Mellum2-12B-A2.5B-Thinking-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 1,118 108.2 7.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pn5e2my7",
   "source_url": "https://omlx.ai/benchmarks/performance/pn5e2my7",
   "tg_tokens": null,
   "tps": 108.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-xdefgvpm-ctx1024",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: 6hduh3jp; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 826.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 1k 826.9 90.3 18.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xdefgvpm",
   "source_url": "https://omlx.ai/benchmarks/performance/xdefgvpm",
   "tg_tokens": null,
   "tps": 90.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-xdefgvpm-ctx4096",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 6hduh3jp; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1038.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 4k 1,038 87.6 19.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xdefgvpm",
   "source_url": "https://omlx.ai/benchmarks/performance/xdefgvpm",
   "tg_tokens": null,
   "tps": 87.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-bnrgio1b-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: 1rm4wsmd, cd5ne1xa, x2aaqgpa; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2604.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 2,604 89.5 37.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bnrgio1b",
   "source_url": "https://omlx.ai/benchmarks/performance/bnrgio1b",
   "tg_tokens": null,
   "tps": 89.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-bnrgio1b-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: 1rm4wsmd, cd5ne1xa, x2aaqgpa; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2545.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 2,545 88.3 37.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bnrgio1b",
   "source_url": "https://omlx.ai/benchmarks/performance/bnrgio1b",
   "tg_tokens": null,
   "tps": 88.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-bnrgio1b-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.6 GB at this context; duplicate traces of the same session: 1rm4wsmd, cd5ne1xa, x2aaqgpa; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2289.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 16k 2,289 85.6 37.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bnrgio1b",
   "source_url": "https://omlx.ai/benchmarks/performance/bnrgio1b",
   "tg_tokens": null,
   "tps": 85.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-bnrgio1b-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 38.2 GB at this context; duplicate traces of the same session: 1rm4wsmd, cd5ne1xa, x2aaqgpa; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1895.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 32k 1,895 81.9 38.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bnrgio1b",
   "source_url": "https://omlx.ai/benchmarks/performance/bnrgio1b",
   "tg_tokens": null,
   "tps": 81.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-o06449ke-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.6 GB at this context; duplicate traces of the same session: nrn1eer3; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1369.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 1,369 56.8 6.6 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o06449ke",
   "source_url": "https://omlx.ai/benchmarks/performance/o06449ke",
   "tg_tokens": null,
   "tps": 56.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-o06449ke-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: nrn1eer3; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1424.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 1,424 54.8 6.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o06449ke",
   "source_url": "https://omlx.ai/benchmarks/performance/o06449ke",
   "tg_tokens": null,
   "tps": 54.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-acukt487-ctx1024",
   "model": "Qwen3-Next-80B-A3B-Instruct-1M",
   "notes": "qx64n MLX build; oMLX community benchmark trace; peak memory 51.0 GB at this context; duplicate traces of the same session: utgo4bzg; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1585.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Next-80B-A3B-Instruct-1M-qx64n-mlx on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 1k 1,585 58.0 51.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/acukt487",
   "source_url": "https://omlx.ai/benchmarks/performance/acukt487",
   "tg_tokens": null,
   "tps": 58.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-acukt487-ctx4096",
   "model": "Qwen3-Next-80B-A3B-Instruct-1M",
   "notes": "qx64n MLX build; oMLX community benchmark trace; peak memory 51.8 GB at this context; duplicate traces of the same session: utgo4bzg; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2255.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3-Next-80B-A3B-Instruct-1M-qx64n-mlx on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 4k 2,255 69.5 51.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/acukt487",
   "source_url": "https://omlx.ai/benchmarks/performance/acukt487",
   "tg_tokens": null,
   "tps": 69.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-s1efr3be-ctx16384",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: ix8irdy7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2716.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,716 119.6 20.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s1efr3be",
   "source_url": "https://omlx.ai/benchmarks/performance/s1efr3be",
   "tg_tokens": null,
   "tps": 119.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-s1efr3be-ctx65536",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: ix8irdy7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1700.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 1,700 68.8 21.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s1efr3be",
   "source_url": "https://omlx.ai/benchmarks/performance/s1efr3be",
   "tg_tokens": null,
   "tps": 68.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-06axt7co-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.2 GB at this context; duplicate traces of the same session: 5yozj1wk; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3308.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 3,308 159.3 39.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/06axt7co",
   "source_url": "https://omlx.ai/benchmarks/performance/06axt7co",
   "tg_tokens": null,
   "tps": 159.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-06axt7co-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.6 GB at this context; duplicate traces of the same session: 5yozj1wk; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2848.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k 2,848 172.5 39.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/06axt7co",
   "source_url": "https://omlx.ai/benchmarks/performance/06axt7co",
   "tg_tokens": null,
   "tps": 172.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-0go01x8s-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: qmhlgd5w; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 65.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 65.7 14.8 17.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0go01x8s",
   "source_url": "https://omlx.ai/benchmarks/performance/0go01x8s",
   "tg_tokens": null,
   "tps": 14.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-0go01x8s-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: qmhlgd5w; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 64.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k 64.8 12.1 19.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0go01x8s",
   "source_url": "https://omlx.ai/benchmarks/performance/0go01x8s",
   "tg_tokens": null,
   "tps": 12.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-12wbxtss-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: ployhi8i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 261.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 261.5 29.9 16.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/12wbxtss",
   "source_url": "https://omlx.ai/benchmarks/performance/12wbxtss",
   "tg_tokens": null,
   "tps": 29.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-12wbxtss-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: ployhi8i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 257.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 257.8 28.5 19.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/12wbxtss",
   "source_url": "https://omlx.ai/benchmarks/performance/12wbxtss",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ip0xae6i-ctx65536",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2",
   "notes": "fp16 MTP build; oMLX community benchmark trace, single run at 64k context; peak memory 40.1 GB; TTFT 95383 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 687.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 64k tokens 687.1 PP tok/s 37.2 TG tok/s 95383 TTFT (ms) 40.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ip0xae6i",
   "source_url": "https://omlx.ai/benchmarks/performance/ip0xae6i",
   "tg_tokens": null,
   "tps": 37.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-iffnh57h-ctx65536",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2",
   "notes": "fp16 MTP build; oMLX community benchmark trace, single run at 64k context; peak memory 40.1 GB; TTFT 96489 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 679.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 64k tokens 679.2 PP tok/s 33.3 TG tok/s 96489 TTFT (ms) 40.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iffnh57h",
   "source_url": "https://omlx.ai/benchmarks/performance/iffnh57h",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2ve7dycg-ctx65536",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2",
   "notes": "fp16 MTP build; oMLX community benchmark trace, single run at 64k context; peak memory 36.9 GB; TTFT 98443 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 665.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 64k tokens 665.7 PP tok/s 29.8 TG tok/s 98443 TTFT (ms) 36.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ve7dycg",
   "source_url": "https://omlx.ai/benchmarks/performance/2ve7dycg",
   "tg_tokens": null,
   "tps": 29.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-zflrx2nq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: gbt9x6os; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 387.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 387.5 29.6 23.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zflrx2nq",
   "source_url": "https://omlx.ai/benchmarks/performance/zflrx2nq",
   "tg_tokens": null,
   "tps": 29.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-zflrx2nq-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: gbt9x6os; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 337.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 337.9 35.6 24.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zflrx2nq",
   "source_url": "https://omlx.ai/benchmarks/performance/zflrx2nq",
   "tg_tokens": null,
   "tps": 35.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-renh0b3l-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 993.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 993.8 43.5 33.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/renh0b3l",
   "source_url": "https://omlx.ai/benchmarks/performance/renh0b3l",
   "tg_tokens": null,
   "tps": 43.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-renh0b3l-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1254.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 1,254 38.5 34.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/renh0b3l",
   "source_url": "https://omlx.ai/benchmarks/performance/renh0b3l",
   "tg_tokens": null,
   "tps": 38.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-renh0b3l-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1215.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 1,215 45.7 35.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/renh0b3l",
   "source_url": "https://omlx.ai/benchmarks/performance/renh0b3l",
   "tg_tokens": null,
   "tps": 45.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-renh0b3l-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1039.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 1,039 47.8 35.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/renh0b3l",
   "source_url": "https://omlx.ai/benchmarks/performance/renh0b3l",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-q6njmoq2-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: x62dvg6d; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 306.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 1k 306.0 26.2 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q6njmoq2",
   "source_url": "https://omlx.ai/benchmarks/performance/q6njmoq2",
   "tg_tokens": null,
   "tps": 26.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-q6njmoq2-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: x62dvg6d; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 354.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 4k 354.7 24.6 73.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q6njmoq2",
   "source_url": "https://omlx.ai/benchmarks/performance/q6njmoq2",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-ntqry8yq-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: 1rzvki1s; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 432.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 1k 432.3 29.3 70.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ntqry8yq",
   "source_url": "https://omlx.ai/benchmarks/performance/ntqry8yq",
   "tg_tokens": null,
   "tps": 29.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-ntqry8yq-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: 1rzvki1s; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 568.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 4k 568.7 25.7 73.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ntqry8yq",
   "source_url": "https://omlx.ai/benchmarks/performance/ntqry8yq",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-mm453aeo-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: qycx0129; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1104.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,104 83.4 100.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mm453aeo",
   "source_url": "https://omlx.ai/benchmarks/performance/mm453aeo",
   "tg_tokens": null,
   "tps": 83.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-mm453aeo-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: qycx0129; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1378.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,378 84.5 101.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mm453aeo",
   "source_url": "https://omlx.ai/benchmarks/performance/mm453aeo",
   "tg_tokens": null,
   "tps": 84.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-19yr7p02-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 101.3 GB; TTFT 2913 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1406.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k tokens 1,406 PP tok/s 102.3 TG tok/s 2913 TTFT (ms) 101.3 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/19yr7p02",
   "source_url": "https://omlx.ai/benchmarks/performance/19yr7p02",
   "tg_tokens": null,
   "tps": 102.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i5vhbwc0-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 70.1 GB; TTFT 13618 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1203.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k tokens 1,203 PP tok/s 48.6 TG tok/s 13618 TTFT (ms) 70.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i5vhbwc0",
   "source_url": "https://omlx.ai/benchmarks/performance/i5vhbwc0",
   "tg_tokens": null,
   "tps": 48.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-2q1fl5bt-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.4 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2752.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 2,752 120.5 100.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2q1fl5bt",
   "source_url": "https://omlx.ai/benchmarks/performance/2q1fl5bt",
   "tg_tokens": null,
   "tps": 120.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-2q1fl5bt-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.8 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4529.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 4,529 101.4 101.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2q1fl5bt",
   "source_url": "https://omlx.ai/benchmarks/performance/2q1fl5bt",
   "tg_tokens": null,
   "tps": 101.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-2q1fl5bt-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.6 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4729.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 4,729 131.3 102.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2q1fl5bt",
   "source_url": "https://omlx.ai/benchmarks/performance/2q1fl5bt",
   "tg_tokens": null,
   "tps": 131.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-2q1fl5bt-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.2 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4750.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 4,750 122.8 103.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2q1fl5bt",
   "source_url": "https://omlx.ai/benchmarks/performance/2q1fl5bt",
   "tg_tokens": null,
   "tps": 122.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-2q1fl5bt-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.4 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4659.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 4,659 113.8 104.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2q1fl5bt",
   "source_url": "https://omlx.ai/benchmarks/performance/2q1fl5bt",
   "tg_tokens": null,
   "tps": 113.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-2q1fl5bt-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.0 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4527.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 4,527 109.9 109.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2q1fl5bt",
   "source_url": "https://omlx.ai/benchmarks/performance/2q1fl5bt",
   "tg_tokens": null,
   "tps": 109.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-2q1fl5bt-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.7 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4389.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 4,389 99.1 110.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2q1fl5bt",
   "source_url": "https://omlx.ai/benchmarks/performance/2q1fl5bt",
   "tg_tokens": null,
   "tps": 99.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-izz96dc8-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: 2simypz6, 52kll0u6, 62fskjdu, l5ptq60h, ti6z6b7g, x2z3rp02; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2387.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 2,387 136.3 100.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/izz96dc8",
   "source_url": "https://omlx.ai/benchmarks/performance/izz96dc8",
   "tg_tokens": null,
   "tps": 136.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-izz96dc8-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.0 GB at this context; duplicate traces of the same session: 2simypz6, 52kll0u6, 62fskjdu, l5ptq60h, ti6z6b7g, x2z3rp02; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1914.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,914 133.4 101.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/izz96dc8",
   "source_url": "https://omlx.ai/benchmarks/performance/izz96dc8",
   "tg_tokens": null,
   "tps": 133.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-izz96dc8-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.9 GB at this context; duplicate traces of the same session: 2simypz6, 52kll0u6, 62fskjdu, l5ptq60h, ti6z6b7g, x2z3rp02; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4661.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 4,661 148.1 101.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/izz96dc8",
   "source_url": "https://omlx.ai/benchmarks/performance/izz96dc8",
   "tg_tokens": null,
   "tps": 148.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-izz96dc8-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.5 GB at this context; duplicate traces of the same session: 2simypz6, 52kll0u6, 62fskjdu, l5ptq60h, ti6z6b7g, x2z3rp02; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4644.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 4,644 111.9 102.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/izz96dc8",
   "source_url": "https://omlx.ai/benchmarks/performance/izz96dc8",
   "tg_tokens": null,
   "tps": 111.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-izz96dc8-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.2 GB at this context; duplicate traces of the same session: 2simypz6, 52kll0u6, 62fskjdu, l5ptq60h, ti6z6b7g, x2z3rp02; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4755.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 4,755 44.7 103.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/izz96dc8",
   "source_url": "https://omlx.ai/benchmarks/performance/izz96dc8",
   "tg_tokens": null,
   "tps": 44.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-izz96dc8-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.4 GB at this context; duplicate traces of the same session: 2simypz6, 52kll0u6, 62fskjdu, l5ptq60h, ti6z6b7g, x2z3rp02; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4668.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 4,668 98.7 104.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/izz96dc8",
   "source_url": "https://omlx.ai/benchmarks/performance/izz96dc8",
   "tg_tokens": null,
   "tps": 98.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-izz96dc8-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.1 GB at this context; duplicate traces of the same session: 2simypz6, 52kll0u6, 62fskjdu, l5ptq60h, ti6z6b7g, x2z3rp02; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4529.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 4,529 99.1 109.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/izz96dc8",
   "source_url": "https://omlx.ai/benchmarks/performance/izz96dc8",
   "tg_tokens": null,
   "tps": 99.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-oewn0tvz-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: 2408a0ie, de8z2738; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2206.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 2,206 113.6 16.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oewn0tvz",
   "source_url": "https://omlx.ai/benchmarks/performance/oewn0tvz",
   "tg_tokens": null,
   "tps": 113.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-oewn0tvz-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 2408a0ie, de8z2738; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2329.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 2,329 120.3 16.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oewn0tvz",
   "source_url": "https://omlx.ai/benchmarks/performance/oewn0tvz",
   "tg_tokens": null,
   "tps": 120.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-oewn0tvz-ctx8192",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: 2408a0ie, de8z2738; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2272.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 2,272 117.4 16.9 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oewn0tvz",
   "source_url": "https://omlx.ai/benchmarks/performance/oewn0tvz",
   "tg_tokens": null,
   "tps": 117.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-8qlbx8hv-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: 2oq31p9c, k5h6mvy0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2224.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 2,224 142.5 16.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8qlbx8hv",
   "source_url": "https://omlx.ai/benchmarks/performance/8qlbx8hv",
   "tg_tokens": null,
   "tps": 142.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-8qlbx8hv-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 2oq31p9c, k5h6mvy0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2336.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 2,336 95.1 16.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8qlbx8hv",
   "source_url": "https://omlx.ai/benchmarks/performance/8qlbx8hv",
   "tg_tokens": null,
   "tps": 95.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-8qlbx8hv-ctx8192",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: 2oq31p9c, k5h6mvy0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2270.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 2,270 106.7 16.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8qlbx8hv",
   "source_url": "https://omlx.ai/benchmarks/performance/8qlbx8hv",
   "tg_tokens": null,
   "tps": 106.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-3oohbapo-ctx1024",
   "model": "Humanizer-12B",
   "notes": "MLX G64 build; oMLX community benchmark trace; peak memory 9.1 GB at this context; duplicate traces of the same session: p2j77ouk, wp1na0j3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 277.8,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "humanizer-12B-MLX-oQ5-G64 on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 5bit; 1k 277.8 25.6 9.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3oohbapo",
   "source_url": "https://omlx.ai/benchmarks/performance/3oohbapo",
   "tg_tokens": null,
   "tps": 25.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-3oohbapo-ctx4096",
   "model": "Humanizer-12B",
   "notes": "MLX G64 build; oMLX community benchmark trace; peak memory 9.4 GB at this context; duplicate traces of the same session: p2j77ouk, wp1na0j3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 270.2,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "humanizer-12B-MLX-oQ5-G64 on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 5bit; 4k 270.2 25.2 9.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3oohbapo",
   "source_url": "https://omlx.ai/benchmarks/performance/3oohbapo",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-3oohbapo-ctx32768",
   "model": "Humanizer-12B",
   "notes": "MLX G64 build; oMLX community benchmark trace; peak memory 10.7 GB at this context; duplicate traces of the same session: p2j77ouk, wp1na0j3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 229.3,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "humanizer-12B-MLX-oQ5-G64 on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 5bit; 32k 229.3 23.5 10.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3oohbapo",
   "source_url": "https://omlx.ai/benchmarks/performance/3oohbapo",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-roacwu7k-ctx4096",
   "model": "Devstral-Small-2-24B-Instruct-2512",
   "notes": "oMLX community benchmark trace; peak memory 15.1 GB at this context; duplicate traces of the same session: yho9ogk7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1065.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Devstral-Small-2-24B-Instruct-2512-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,065 35.2 15.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/roacwu7k",
   "source_url": "https://omlx.ai/benchmarks/performance/roacwu7k",
   "tg_tokens": null,
   "tps": 35.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-roacwu7k-ctx32768",
   "model": "Devstral-Small-2-24B-Instruct-2512",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: yho9ogk7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 831.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Devstral-Small-2-24B-Instruct-2512-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 831.7 28.1 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/roacwu7k",
   "source_url": "https://omlx.ai/benchmarks/performance/roacwu7k",
   "tg_tokens": null,
   "tps": 28.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1xn3fjcf-ctx4096",
   "model": "Huihui-gemma-4-31B-it-v2",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 33.1 GB at this context; duplicate traces of the same session: 7rs1wwb2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 671.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Huihui-gemma-4-31B-it-v2-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 671.3 14.8 33.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1xn3fjcf",
   "source_url": "https://omlx.ai/benchmarks/performance/1xn3fjcf",
   "tg_tokens": null,
   "tps": 14.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1xn3fjcf-ctx16384",
   "model": "Huihui-gemma-4-31B-it-v2",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: 7rs1wwb2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 641.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Huihui-gemma-4-31B-it-v2-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 16k 641.3 13.5 34.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1xn3fjcf",
   "source_url": "https://omlx.ai/benchmarks/performance/1xn3fjcf",
   "tg_tokens": null,
   "tps": 13.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-jrjh1m87-ctx1024",
   "model": "LFM2.5-1.2B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 1.8 GB at this context; duplicate traces of the same session: io9rskd5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1833.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-1.2B-Thinking on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 8bit; 1k 1,833 195.3 1.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jrjh1m87",
   "source_url": "https://omlx.ai/benchmarks/performance/jrjh1m87",
   "tg_tokens": null,
   "tps": 195.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-jrjh1m87-ctx4096",
   "model": "LFM2.5-1.2B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 2.1 GB at this context; duplicate traces of the same session: io9rskd5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2312.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-1.2B-Thinking on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 8bit; 4k 2,312 178.2 2.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jrjh1m87",
   "source_url": "https://omlx.ai/benchmarks/performance/jrjh1m87",
   "tg_tokens": null,
   "tps": 178.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-1zkab3f3-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 39.7 GB; TTFT 50629 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 647.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 32k tokens 647.2 PP tok/s 52.9 TG tok/s 50629 TTFT (ms) 39.7 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1zkab3f3",
   "source_url": "https://omlx.ai/benchmarks/performance/1zkab3f3",
   "tg_tokens": null,
   "tps": 52.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-g6dxwxg5-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 38.5 GB; TTFT 19685 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 832.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 16k tokens 832.3 PP tok/s 53.7 TG tok/s 19685 TTFT (ms) 38.5 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g6dxwxg5",
   "source_url": "https://omlx.ai/benchmarks/performance/g6dxwxg5",
   "tg_tokens": null,
   "tps": 53.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-g2i0dn00-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 39.7 GB; TTFT 50792 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 645.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 32k tokens 645.1 PP tok/s 53.3 TG tok/s 50792 TTFT (ms) 39.7 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g2i0dn00",
   "source_url": "https://omlx.ai/benchmarks/performance/g2i0dn00",
   "tg_tokens": null,
   "tps": 53.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-099peonl-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: vxz12rk6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1174.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 1,174 109.4 37.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/099peonl",
   "source_url": "https://omlx.ai/benchmarks/performance/099peonl",
   "tg_tokens": null,
   "tps": 109.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-099peonl-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.1 GB at this context; duplicate traces of the same session: vxz12rk6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1693.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 1,693 99.9 39.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/099peonl",
   "source_url": "https://omlx.ai/benchmarks/performance/099peonl",
   "tg_tokens": null,
   "tps": 99.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-x5uyu6ev-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: tjlcwpjb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1537.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 1,537 74.2 35.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x5uyu6ev",
   "source_url": "https://omlx.ai/benchmarks/performance/x5uyu6ev",
   "tg_tokens": null,
   "tps": 74.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-x5uyu6ev-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: tjlcwpjb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1755.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 1,755 75.6 37.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x5uyu6ev",
   "source_url": "https://omlx.ai/benchmarks/performance/x5uyu6ev",
   "tg_tokens": null,
   "tps": 75.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-hws6z4zm-ctx1024",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: h07eww8w; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2352.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,352 95.0 16.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hws6z4zm",
   "source_url": "https://omlx.ai/benchmarks/performance/hws6z4zm",
   "tg_tokens": null,
   "tps": 95.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-hws6z4zm-ctx4096",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: h07eww8w; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2906.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,906 82.7 17.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hws6z4zm",
   "source_url": "https://omlx.ai/benchmarks/performance/hws6z4zm",
   "tg_tokens": null,
   "tps": 82.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-h9fhkaj7-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: tu23hwlo; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 233.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 233.5 21.1 5.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h9fhkaj7",
   "source_url": "https://omlx.ai/benchmarks/performance/h9fhkaj7",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-h9fhkaj7-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: tu23hwlo; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 241.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 241.3 20.4 6.2 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h9fhkaj7",
   "source_url": "https://omlx.ai/benchmarks/performance/h9fhkaj7",
   "tg_tokens": null,
   "tps": 20.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gbnzzdly-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: 39d2tyfd, dt8lh1nt, s7pzcpdq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2001.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 2,001 79.0 29.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gbnzzdly",
   "source_url": "https://omlx.ai/benchmarks/performance/gbnzzdly",
   "tg_tokens": null,
   "tps": 79.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gbnzzdly-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.6 GB at this context; duplicate traces of the same session: 39d2tyfd, dt8lh1nt, s7pzcpdq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3343.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 3,343 152.7 31.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gbnzzdly",
   "source_url": "https://omlx.ai/benchmarks/performance/gbnzzdly",
   "tg_tokens": null,
   "tps": 152.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gbnzzdly-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.8 GB at this context; duplicate traces of the same session: 39d2tyfd, dt8lh1nt, s7pzcpdq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3178.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 8k 3,178 130.9 31.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gbnzzdly",
   "source_url": "https://omlx.ai/benchmarks/performance/gbnzzdly",
   "tg_tokens": null,
   "tps": 130.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gbnzzdly-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 54.9 GB at this context; duplicate traces of the same session: 39d2tyfd, dt8lh1nt, s7pzcpdq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2843.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 16k 2,843 78.2 54.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gbnzzdly",
   "source_url": "https://omlx.ai/benchmarks/performance/gbnzzdly",
   "tg_tokens": null,
   "tps": 78.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9yxzbgpb-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.0 GB at this context; duplicate traces of the same session: 9bk92fqc, hbbe6sqi, k4jacjqr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2693.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 2,693 193.0 30.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9yxzbgpb",
   "source_url": "https://omlx.ai/benchmarks/performance/9yxzbgpb",
   "tg_tokens": null,
   "tps": 193.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9yxzbgpb-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.6 GB at this context; duplicate traces of the same session: 9bk92fqc, hbbe6sqi, k4jacjqr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3346.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 3,346 145.8 31.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9yxzbgpb",
   "source_url": "https://omlx.ai/benchmarks/performance/9yxzbgpb",
   "tg_tokens": null,
   "tps": 145.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9yxzbgpb-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.8 GB at this context; duplicate traces of the same session: 9bk92fqc, hbbe6sqi, k4jacjqr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3178.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 8k 3,178 107.8 31.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9yxzbgpb",
   "source_url": "https://omlx.ai/benchmarks/performance/9yxzbgpb",
   "tg_tokens": null,
   "tps": 107.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9yxzbgpb-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: 9bk92fqc, hbbe6sqi, k4jacjqr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2852.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 16k 2,852 193.3 32.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9yxzbgpb",
   "source_url": "https://omlx.ai/benchmarks/performance/9yxzbgpb",
   "tg_tokens": null,
   "tps": 193.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-unufvrj7-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 22.1 GB; TTFT 133670 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 122.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k tokens 122.6 PP tok/s 17.7 TG tok/s 133670 TTFT (ms) 22.1 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/unufvrj7",
   "source_url": "https://omlx.ai/benchmarks/performance/unufvrj7",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fiv9f1m9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: 47kwrvp0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 936.8,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.8-27B-MLX-5bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 5bit; 4k 936.8 27.0 20.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fiv9f1m9",
   "source_url": "https://omlx.ai/benchmarks/performance/fiv9f1m9",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fiv9f1m9-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 47kwrvp0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 776.2,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.8-27B-MLX-5bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 5bit; 32k 776.2 24.5 21.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fiv9f1m9",
   "source_url": "https://omlx.ai/benchmarks/performance/fiv9f1m9",
   "tg_tokens": null,
   "tps": 24.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-rhnzzp4r-ctx1024",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "oMLX community benchmark trace; peak memory 32.9 GB at this context; duplicate traces of the same session: k0kw3u15; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 155.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 155.0 17.7 32.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rhnzzp4r",
   "source_url": "https://omlx.ai/benchmarks/performance/rhnzzp4r",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-rhnzzp4r-ctx4096",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "oMLX community benchmark trace; peak memory 40.4 GB at this context; duplicate traces of the same session: k0kw3u15; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 158.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 158.2 14.4 40.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rhnzzp4r",
   "source_url": "https://omlx.ai/benchmarks/performance/rhnzzp4r",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-jmb1i1x4-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 5wxnrweb, 9q23z793, ismo5drz, ns5vcc49, v9buqg5t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 253.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 253.1 48.9 21.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jmb1i1x4",
   "source_url": "https://omlx.ai/benchmarks/performance/jmb1i1x4",
   "tg_tokens": null,
   "tps": 48.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-jmb1i1x4-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: 5wxnrweb, 9q23z793, ismo5drz, ns5vcc49, v9buqg5t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 326.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 326.8 83.0 25.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jmb1i1x4",
   "source_url": "https://omlx.ai/benchmarks/performance/jmb1i1x4",
   "tg_tokens": null,
   "tps": 83.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-jmb1i1x4-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: 5wxnrweb, 9q23z793, ismo5drz, ns5vcc49, v9buqg5t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 328.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 328.3 38.4 25.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jmb1i1x4",
   "source_url": "https://omlx.ai/benchmarks/performance/jmb1i1x4",
   "tg_tokens": null,
   "tps": 38.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-jmb1i1x4-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: 5wxnrweb, 9q23z793, ismo5drz, ns5vcc49, v9buqg5t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 309.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 309.9 53.9 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jmb1i1x4",
   "source_url": "https://omlx.ai/benchmarks/performance/jmb1i1x4",
   "tg_tokens": null,
   "tps": 53.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-jmb1i1x4-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.3 GB at this context; duplicate traces of the same session: 5wxnrweb, 9q23z793, ismo5drz, ns5vcc49, v9buqg5t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 280.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 280.6 78.8 28.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jmb1i1x4",
   "source_url": "https://omlx.ai/benchmarks/performance/jmb1i1x4",
   "tg_tokens": null,
   "tps": 78.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-jmb1i1x4-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.8 GB at this context; duplicate traces of the same session: 5wxnrweb, 9q23z793, ismo5drz, ns5vcc49, v9buqg5t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 221.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 221.9 39.9 30.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jmb1i1x4",
   "source_url": "https://omlx.ai/benchmarks/performance/jmb1i1x4",
   "tg_tokens": null,
   "tps": 39.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-cau2tb4l-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: igx8uqjn, ka2nhcwb, krnbpua4, ro4yamaj, y8d7g5hf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 253.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 253.8 69.7 21.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cau2tb4l",
   "source_url": "https://omlx.ai/benchmarks/performance/cau2tb4l",
   "tg_tokens": null,
   "tps": 69.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-cau2tb4l-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: igx8uqjn, ka2nhcwb, krnbpua4, ro4yamaj, y8d7g5hf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 341.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 341.7 79.1 25.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cau2tb4l",
   "source_url": "https://omlx.ai/benchmarks/performance/cau2tb4l",
   "tg_tokens": null,
   "tps": 79.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-cau2tb4l-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.6 GB at this context; duplicate traces of the same session: igx8uqjn, ka2nhcwb, krnbpua4, ro4yamaj, y8d7g5hf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 339.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 339.7 42.1 25.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cau2tb4l",
   "source_url": "https://omlx.ai/benchmarks/performance/cau2tb4l",
   "tg_tokens": null,
   "tps": 42.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-cau2tb4l-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: igx8uqjn, ka2nhcwb, krnbpua4, ro4yamaj, y8d7g5hf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 323.5 62.9 26.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cau2tb4l",
   "source_url": "https://omlx.ai/benchmarks/performance/cau2tb4l",
   "tg_tokens": null,
   "tps": 62.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-cau2tb4l-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.3 GB at this context; duplicate traces of the same session: igx8uqjn, ka2nhcwb, krnbpua4, ro4yamaj, y8d7g5hf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 293.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 293.7 50.1 28.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cau2tb4l",
   "source_url": "https://omlx.ai/benchmarks/performance/cau2tb4l",
   "tg_tokens": null,
   "tps": 50.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-cau2tb4l-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.0 GB at this context; duplicate traces of the same session: igx8uqjn, ka2nhcwb, krnbpua4, ro4yamaj, y8d7g5hf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 245.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 245.8 39.7 31.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cau2tb4l",
   "source_url": "https://omlx.ai/benchmarks/performance/cau2tb4l",
   "tg_tokens": null,
   "tps": 39.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-atw0a8vr-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: 23pmm4q7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 217.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 217.1 33.3 33.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/atw0a8vr",
   "source_url": "https://omlx.ai/benchmarks/performance/atw0a8vr",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-atw0a8vr-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.7 GB at this context; duplicate traces of the same session: 23pmm4q7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 186.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 186.9 21.5 37.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/atw0a8vr",
   "source_url": "https://omlx.ai/benchmarks/performance/atw0a8vr",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-c7w1ewcv-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: xb1y7rne; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 217.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 217.4 32.6 33.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c7w1ewcv",
   "source_url": "https://omlx.ai/benchmarks/performance/c7w1ewcv",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-c7w1ewcv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.7 GB at this context; duplicate traces of the same session: xb1y7rne; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 194.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 194.0 22.7 37.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c7w1ewcv",
   "source_url": "https://omlx.ai/benchmarks/performance/c7w1ewcv",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-da7o3rgw-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 36.9 GB; TTFT 71383 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 918.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k tokens 918.1 PP tok/s 27.6 TG tok/s 71383 TTFT (ms) 36.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/da7o3rgw",
   "source_url": "https://omlx.ai/benchmarks/performance/da7o3rgw",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-8wab4txi-ctx16384",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 40.9 GB; TTFT 67275 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 243.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k tokens 243.5 PP tok/s 24.0 TG tok/s 67275 TTFT (ms) 40.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8wab4txi",
   "source_url": "https://omlx.ai/benchmarks/performance/8wab4txi",
   "tg_tokens": null,
   "tps": 24.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-mqvgjjjx-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: 4m0ftr03; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 429.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 1k 429.0 29.6 70.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mqvgjjjx",
   "source_url": "https://omlx.ai/benchmarks/performance/mqvgjjjx",
   "tg_tokens": null,
   "tps": 29.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-mqvgjjjx-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: 4m0ftr03; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 570.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 4k 570.3 27.2 73.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mqvgjjjx",
   "source_url": "https://omlx.ai/benchmarks/performance/mqvgjjjx",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-pdsvt6q8-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: l2s7mwf2; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 347.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 1k 347.3 26.2 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pdsvt6q8",
   "source_url": "https://omlx.ai/benchmarks/performance/pdsvt6q8",
   "tg_tokens": null,
   "tps": 26.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-pdsvt6q8-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: l2s7mwf2; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 399.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 4k 399.7 25.1 73.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pdsvt6q8",
   "source_url": "https://omlx.ai/benchmarks/performance/pdsvt6q8",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-ek5l5qor-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.3 GB at this context; duplicate traces of the same session: 68xnwa6j; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 296.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 1k 296.6 37.4 69.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ek5l5qor",
   "source_url": "https://omlx.ai/benchmarks/performance/ek5l5qor",
   "tg_tokens": null,
   "tps": 37.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-ek5l5qor-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.8 GB at this context; duplicate traces of the same session: 68xnwa6j; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 335.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 4k 335.0 38.4 69.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ek5l5qor",
   "source_url": "https://omlx.ai/benchmarks/performance/ek5l5qor",
   "tg_tokens": null,
   "tps": 38.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-j6phwmik-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.3 GB at this context; duplicate traces of the same session: lnj3v81s; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 261.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 1k 261.5 36.4 69.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j6phwmik",
   "source_url": "https://omlx.ai/benchmarks/performance/j6phwmik",
   "tg_tokens": null,
   "tps": 36.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-j6phwmik-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.8 GB at this context; duplicate traces of the same session: lnj3v81s; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 4k 323.7 39.8 69.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j6phwmik",
   "source_url": "https://omlx.ai/benchmarks/performance/j6phwmik",
   "tg_tokens": null,
   "tps": 39.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xkqpyv4f-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: 0vrn43kb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1407.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,407 85.7 101.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xkqpyv4f",
   "source_url": "https://omlx.ai/benchmarks/performance/xkqpyv4f",
   "tg_tokens": null,
   "tps": 85.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xkqpyv4f-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.7 GB at this context; duplicate traces of the same session: 0vrn43kb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1395.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 1,395 65.3 101.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xkqpyv4f",
   "source_url": "https://omlx.ai/benchmarks/performance/xkqpyv4f",
   "tg_tokens": null,
   "tps": 65.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-h5yongum-ctx1024",
   "model": "LFM2.5-2.6B",
   "notes": "oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: 9b0fx6s0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4166.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-2.6B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 4,166 86.9 3.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h5yongum",
   "source_url": "https://omlx.ai/benchmarks/performance/h5yongum",
   "tg_tokens": null,
   "tps": 86.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-h5yongum-ctx4096",
   "model": "LFM2.5-2.6B",
   "notes": "oMLX community benchmark trace; peak memory 3.5 GB at this context; duplicate traces of the same session: 9b0fx6s0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4133.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-2.6B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 4,133 82.1 3.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h5yongum",
   "source_url": "https://omlx.ai/benchmarks/performance/h5yongum",
   "tg_tokens": null,
   "tps": 82.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-vqios638-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 38.5 GB; TTFT 19721 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 830.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 16k tokens 830.8 PP tok/s 57.0 TG tok/s 19721 TTFT (ms) 38.5 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vqios638",
   "source_url": "https://omlx.ai/benchmarks/performance/vqios638",
   "tg_tokens": null,
   "tps": 57.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-cdjn0tbz-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.6 GB at this context; duplicate traces of the same session: 1npaoz2p, em7e3xde, ghhgzhh6, ozuoqkod; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2274.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 2,274 92.2 35.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cdjn0tbz",
   "source_url": "https://omlx.ai/benchmarks/performance/cdjn0tbz",
   "tg_tokens": null,
   "tps": 92.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-cdjn0tbz-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: 1npaoz2p, em7e3xde, ghhgzhh6, ozuoqkod; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2718.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 2,718 91.3 37.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cdjn0tbz",
   "source_url": "https://omlx.ai/benchmarks/performance/cdjn0tbz",
   "tg_tokens": null,
   "tps": 91.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-cdjn0tbz-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: 1npaoz2p, em7e3xde, ghhgzhh6, ozuoqkod; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2571.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 2,571 88.9 37.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cdjn0tbz",
   "source_url": "https://omlx.ai/benchmarks/performance/cdjn0tbz",
   "tg_tokens": null,
   "tps": 88.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-cdjn0tbz-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.6 GB at this context; duplicate traces of the same session: 1npaoz2p, em7e3xde, ghhgzhh6, ozuoqkod; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2310.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 16k 2,310 86.2 37.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cdjn0tbz",
   "source_url": "https://omlx.ai/benchmarks/performance/cdjn0tbz",
   "tg_tokens": null,
   "tps": 86.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-cdjn0tbz-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 38.2 GB at this context; duplicate traces of the same session: 1npaoz2p, em7e3xde, ghhgzhh6, ozuoqkod; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1905.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 32k 1,905 82.5 38.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cdjn0tbz",
   "source_url": "https://omlx.ai/benchmarks/performance/cdjn0tbz",
   "tg_tokens": null,
   "tps": 82.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-v5n001i0-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: b26yfyiw; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 822.1,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 822.1 60.9 24.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v5n001i0",
   "source_url": "https://omlx.ai/benchmarks/performance/v5n001i0",
   "tg_tokens": null,
   "tps": 60.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-v5n001i0-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: b26yfyiw; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 638.8,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 32k 638.8 57.6 26.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v5n001i0",
   "source_url": "https://omlx.ai/benchmarks/performance/v5n001i0",
   "tg_tokens": null,
   "tps": 57.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-7q7mw5xk-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: 0htnyi4m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1487.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-fixed-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 1,487 97.0 37.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7q7mw5xk",
   "source_url": "https://omlx.ai/benchmarks/performance/7q7mw5xk",
   "tg_tokens": null,
   "tps": 97.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-7q7mw5xk-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.1 GB at this context; duplicate traces of the same session: 0htnyi4m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1681.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-fixed-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 1,681 99.4 39.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7q7mw5xk",
   "source_url": "https://omlx.ai/benchmarks/performance/7q7mw5xk",
   "tg_tokens": null,
   "tps": 99.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-i21xnj2u-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: 3jeo86am; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1469.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 1,469 108.8 37.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i21xnj2u",
   "source_url": "https://omlx.ai/benchmarks/performance/i21xnj2u",
   "tg_tokens": null,
   "tps": 108.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-i21xnj2u-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.1 GB at this context; duplicate traces of the same session: 3jeo86am; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1668.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 1,668 97.4 39.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i21xnj2u",
   "source_url": "https://omlx.ai/benchmarks/performance/i21xnj2u",
   "tg_tokens": null,
   "tps": 97.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-phl4qq7z-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.4 GB at this context; duplicate traces of the same session: f7qvu378, thrgxf0d, v792bp4t, z46mt3m3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2189.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 2,189 140.1 36.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/phl4qq7z",
   "source_url": "https://omlx.ai/benchmarks/performance/phl4qq7z",
   "tg_tokens": null,
   "tps": 140.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-phl4qq7z-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.1 GB at this context; duplicate traces of the same session: f7qvu378, thrgxf0d, v792bp4t, z46mt3m3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2619.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 2,619 118.3 38.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/phl4qq7z",
   "source_url": "https://omlx.ai/benchmarks/performance/phl4qq7z",
   "tg_tokens": null,
   "tps": 118.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-phl4qq7z-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.3 GB at this context; duplicate traces of the same session: f7qvu378, thrgxf0d, v792bp4t, z46mt3m3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2477.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 2,477 115.5 38.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/phl4qq7z",
   "source_url": "https://omlx.ai/benchmarks/performance/phl4qq7z",
   "tg_tokens": null,
   "tps": 115.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-phl4qq7z-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.6 GB at this context; duplicate traces of the same session: f7qvu378, thrgxf0d, v792bp4t, z46mt3m3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2210.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 16k 2,210 120.7 38.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/phl4qq7z",
   "source_url": "https://omlx.ai/benchmarks/performance/phl4qq7z",
   "tg_tokens": null,
   "tps": 120.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-phl4qq7z-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: f7qvu378, thrgxf0d, v792bp4t, z46mt3m3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1800.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 32k 1,800 91.6 39.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/phl4qq7z",
   "source_url": "https://omlx.ai/benchmarks/performance/phl4qq7z",
   "tg_tokens": null,
   "tps": 91.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-6ak65n1g-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: 0nj7fmps, 1ud4ighk, yjjbg35o; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 234.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 234.4 21.5 5.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ak65n1g",
   "source_url": "https://omlx.ai/benchmarks/performance/6ak65n1g",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-6ak65n1g-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: 0nj7fmps, 1ud4ighk, yjjbg35o; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 241.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 241.5 20.9 6.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ak65n1g",
   "source_url": "https://omlx.ai/benchmarks/performance/6ak65n1g",
   "tg_tokens": null,
   "tps": 20.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-6ak65n1g-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: 0nj7fmps, 1ud4ighk, yjjbg35o; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 239.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 8k 239.5 20.3 6.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ak65n1g",
   "source_url": "https://omlx.ai/benchmarks/performance/6ak65n1g",
   "tg_tokens": null,
   "tps": 20.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-6ak65n1g-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: 0nj7fmps, 1ud4ighk, yjjbg35o; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 230.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 16k 230.4 19.0 6.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ak65n1g",
   "source_url": "https://omlx.ai/benchmarks/performance/6ak65n1g",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-l9vm0jjc-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: gruvwetq, ub9hmq41, xpk8o8i6; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2628.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 2,628 194.2 21.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l9vm0jjc",
   "source_url": "https://omlx.ai/benchmarks/performance/l9vm0jjc",
   "tg_tokens": null,
   "tps": 194.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-l9vm0jjc-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: gruvwetq, ub9hmq41, xpk8o8i6; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3239.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 3,239 162.3 23.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l9vm0jjc",
   "source_url": "https://omlx.ai/benchmarks/performance/l9vm0jjc",
   "tg_tokens": null,
   "tps": 162.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-l9vm0jjc-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: gruvwetq, ub9hmq41, xpk8o8i6; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2700.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 2,700 192.3 23.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l9vm0jjc",
   "source_url": "https://omlx.ai/benchmarks/performance/l9vm0jjc",
   "tg_tokens": null,
   "tps": 192.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-l9vm0jjc-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.3 GB at this context; duplicate traces of the same session: gruvwetq, ub9hmq41, xpk8o8i6; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1094.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 1,094 120.2 27.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l9vm0jjc",
   "source_url": "https://omlx.ai/benchmarks/performance/l9vm0jjc",
   "tg_tokens": null,
   "tps": 120.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-9amfomda-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: c3g57wit; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 423.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 423.2 16.6 16.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9amfomda",
   "source_url": "https://omlx.ai/benchmarks/performance/9amfomda",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-9amfomda-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: c3g57wit; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 478.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 478.9 16.7 17.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9amfomda",
   "source_url": "https://omlx.ai/benchmarks/performance/9amfomda",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-7avr0wca-ctx4096",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 31.1 GB; TTFT 9785 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 418.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Alis-MLX-6bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k tokens 418.6 PP tok/s 39.3 TG tok/s 9785 TTFT (ms) 31.1 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7avr0wca",
   "source_url": "https://omlx.ai/benchmarks/performance/7avr0wca",
   "tg_tokens": null,
   "tps": 39.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-gb0r4z99-ctx1024",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: p2e5b7pt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 221.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 221.0 39.5 18.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gb0r4z99",
   "source_url": "https://omlx.ai/benchmarks/performance/gb0r4z99",
   "tg_tokens": null,
   "tps": 39.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-gb0r4z99-ctx4096",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: p2e5b7pt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 193.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 193.1 28.1 22.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gb0r4z99",
   "source_url": "https://omlx.ai/benchmarks/performance/gb0r4z99",
   "tg_tokens": null,
   "tps": 28.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-swpcji6t-ctx1024",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 7qk23d35; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 221.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 221.2 46.4 18.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/swpcji6t",
   "source_url": "https://omlx.ai/benchmarks/performance/swpcji6t",
   "tg_tokens": null,
   "tps": 46.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-swpcji6t-ctx4096",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: 7qk23d35; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 192.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 192.2 25.4 22.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/swpcji6t",
   "source_url": "https://omlx.ai/benchmarks/performance/swpcji6t",
   "tg_tokens": null,
   "tps": 25.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-2zuxa18p-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 5rol6rhi, zu0bv5p1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 483.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 483.0 16.2 18.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2zuxa18p",
   "source_url": "https://omlx.ai/benchmarks/performance/2zuxa18p",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-2zuxa18p-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 5rol6rhi, zu0bv5p1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 441.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 441.5 14.8 19.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2zuxa18p",
   "source_url": "https://omlx.ai/benchmarks/performance/2zuxa18p",
   "tg_tokens": null,
   "tps": 14.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-2zuxa18p-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 5rol6rhi, zu0bv5p1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 363.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 363.0 13.2 21.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2zuxa18p",
   "source_url": "https://omlx.ai/benchmarks/performance/2zuxa18p",
   "tg_tokens": null,
   "tps": 13.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-wwwkwlgh-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: wpqkgbw8; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 227.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 8bit; 1k 227.8 18.6 28.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wwwkwlgh",
   "source_url": "https://omlx.ai/benchmarks/performance/wwwkwlgh",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-wwwkwlgh-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: wpqkgbw8; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 231.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 8bit; 4k 231.8 18.4 32.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wwwkwlgh",
   "source_url": "https://omlx.ai/benchmarks/performance/wwwkwlgh",
   "tg_tokens": null,
   "tps": 18.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-6fo23jw5-ctx65536",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 42.3 GB at this context; duplicate traces of the same session: ctpdl7l4; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 228.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 64k 228.5 16.1 42.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6fo23jw5",
   "source_url": "https://omlx.ai/benchmarks/performance/6fo23jw5",
   "tg_tokens": null,
   "tps": 16.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-6fo23jw5-ctx131072",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 45.5 GB at this context; duplicate traces of the same session: ctpdl7l4; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 231.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 128k 231.9 20.2 45.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6fo23jw5",
   "source_url": "https://omlx.ai/benchmarks/performance/6fo23jw5",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-h1zup0q2-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 13.6 GB at this context; duplicate traces of the same session: 1xvcqkbv, d30uk0g0, mgdpjkq8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 177.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 177.3 12.8 13.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h1zup0q2",
   "source_url": "https://omlx.ai/benchmarks/performance/h1zup0q2",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-h1zup0q2-ctx32768",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 1xvcqkbv, d30uk0g0, mgdpjkq8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1286.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,286 9.0 16.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h1zup0q2",
   "source_url": "https://omlx.ai/benchmarks/performance/h1zup0q2",
   "tg_tokens": null,
   "tps": 9.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-h1zup0q2-ctx65536",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 17.5 GB at this context; duplicate traces of the same session: 1xvcqkbv, d30uk0g0, mgdpjkq8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1282.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 1,282 13.1 17.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h1zup0q2",
   "source_url": "https://omlx.ai/benchmarks/performance/h1zup0q2",
   "tg_tokens": null,
   "tps": 13.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-h1zup0q2-ctx131072",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 1xvcqkbv, d30uk0g0, mgdpjkq8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1226.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 1,226 12.1 19.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h1zup0q2",
   "source_url": "https://omlx.ai/benchmarks/performance/h1zup0q2",
   "tg_tokens": null,
   "tps": 12.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-s3q5rjt8-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 8.9 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 81.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 81.5 5.1 8.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s3q5rjt8",
   "source_url": "https://omlx.ai/benchmarks/performance/s3q5rjt8",
   "tg_tokens": null,
   "tps": 5.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-s3q5rjt8-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 9.4 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 150.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 150.1 4.7 9.4 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s3q5rjt8",
   "source_url": "https://omlx.ai/benchmarks/performance/s3q5rjt8",
   "tg_tokens": null,
   "tps": 4.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-s3q5rjt8-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 9.4 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 147.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 8k 147.0 5.0 9.4 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s3q5rjt8",
   "source_url": "https://omlx.ai/benchmarks/performance/s3q5rjt8",
   "tg_tokens": null,
   "tps": 5.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qg7re05c-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1204.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,204 86.6 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qg7re05c",
   "source_url": "https://omlx.ai/benchmarks/performance/qg7re05c",
   "tg_tokens": null,
   "tps": 86.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qg7re05c-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2701.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,701 64.2 72.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qg7re05c",
   "source_url": "https://omlx.ai/benchmarks/performance/qg7re05c",
   "tg_tokens": null,
   "tps": 64.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qg7re05c-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2541.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,541 90.9 73.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qg7re05c",
   "source_url": "https://omlx.ai/benchmarks/performance/qg7re05c",
   "tg_tokens": null,
   "tps": 90.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qg7re05c-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2490.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,490 73.4 73.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qg7re05c",
   "source_url": "https://omlx.ai/benchmarks/performance/qg7re05c",
   "tg_tokens": null,
   "tps": 73.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-sbw2rbhf-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.7 GB at this context; duplicate traces of the same session: 7chkomfq, oq0q10fi, rugp2bmd, viy9a2vi; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4685.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 4,685 101.0 102.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sbw2rbhf",
   "source_url": "https://omlx.ai/benchmarks/performance/sbw2rbhf",
   "tg_tokens": null,
   "tps": 101.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-sbw2rbhf-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.3 GB at this context; duplicate traces of the same session: 7chkomfq, oq0q10fi, rugp2bmd, viy9a2vi; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4682.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 4,682 137.8 103.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sbw2rbhf",
   "source_url": "https://omlx.ai/benchmarks/performance/sbw2rbhf",
   "tg_tokens": null,
   "tps": 137.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-sbw2rbhf-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.5 GB at this context; duplicate traces of the same session: 7chkomfq, oq0q10fi, rugp2bmd, viy9a2vi; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4624.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 4,624 113.1 104.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sbw2rbhf",
   "source_url": "https://omlx.ai/benchmarks/performance/sbw2rbhf",
   "tg_tokens": null,
   "tps": 113.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-sbw2rbhf-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 106.8 GB at this context; duplicate traces of the same session: 7chkomfq, oq0q10fi, rugp2bmd, viy9a2vi; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4520.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 4,520 105.4 106.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sbw2rbhf",
   "source_url": "https://omlx.ai/benchmarks/performance/sbw2rbhf",
   "tg_tokens": null,
   "tps": 105.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-sbw2rbhf-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.6 GB at this context; duplicate traces of the same session: 7chkomfq, oq0q10fi, rugp2bmd, viy9a2vi; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4380.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 4,380 103.6 109.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sbw2rbhf",
   "source_url": "https://omlx.ai/benchmarks/performance/sbw2rbhf",
   "tg_tokens": null,
   "tps": 103.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-3jln2pdw-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.1 GB at this context; duplicate traces of the same session: 09eygvjh, iztpje2k, jdaj2foo, ql71rqyq, ssjohfk9; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2810.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 2,810 21.7 102.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3jln2pdw",
   "source_url": "https://omlx.ai/benchmarks/performance/3jln2pdw",
   "tg_tokens": null,
   "tps": 21.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-3jln2pdw-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.7 GB at this context; duplicate traces of the same session: 09eygvjh, iztpje2k, jdaj2foo, ql71rqyq, ssjohfk9; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4494.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 4,494 126.5 102.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3jln2pdw",
   "source_url": "https://omlx.ai/benchmarks/performance/3jln2pdw",
   "tg_tokens": null,
   "tps": 126.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-3jln2pdw-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.4 GB at this context; duplicate traces of the same session: 09eygvjh, iztpje2k, jdaj2foo, ql71rqyq, ssjohfk9; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4699.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 4,699 88.4 103.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3jln2pdw",
   "source_url": "https://omlx.ai/benchmarks/performance/3jln2pdw",
   "tg_tokens": null,
   "tps": 88.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-3jln2pdw-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.5 GB at this context; duplicate traces of the same session: 09eygvjh, iztpje2k, jdaj2foo, ql71rqyq, ssjohfk9; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4630.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 4,630 73.0 104.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3jln2pdw",
   "source_url": "https://omlx.ai/benchmarks/performance/3jln2pdw",
   "tg_tokens": null,
   "tps": 73.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-3jln2pdw-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.5 GB at this context; duplicate traces of the same session: 09eygvjh, iztpje2k, jdaj2foo, ql71rqyq, ssjohfk9; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3368.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 3,368 31.5 109.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3jln2pdw",
   "source_url": "https://omlx.ai/benchmarks/performance/3jln2pdw",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-3jln2pdw-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 112.0 GB at this context; duplicate traces of the same session: 09eygvjh, iztpje2k, jdaj2foo, ql71rqyq, ssjohfk9; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2862.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 2,862 109.2 112.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3jln2pdw",
   "source_url": "https://omlx.ai/benchmarks/performance/3jln2pdw",
   "tg_tokens": null,
   "tps": 109.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-en5syl36-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: 1zngcusm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 722.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 722.9 68.0 71.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/en5syl36",
   "source_url": "https://omlx.ai/benchmarks/performance/en5syl36",
   "tg_tokens": null,
   "tps": 68.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-en5syl36-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift build; oMLX community benchmark trace; peak memory 71.9 GB at this context; duplicate traces of the same session: 1zngcusm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 829.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 829.3 71.4 71.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/en5syl36",
   "source_url": "https://omlx.ai/benchmarks/performance/en5syl36",
   "tg_tokens": null,
   "tps": 71.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-q2ulgfls-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 13.6 GB at this context; duplicate traces of the same session: fq7zfo2t, xvtrw1sv, y482og3o; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1518.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 1,518 111.7 13.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q2ulgfls",
   "source_url": "https://omlx.ai/benchmarks/performance/q2ulgfls",
   "tg_tokens": null,
   "tps": 111.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-q2ulgfls-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 14.1 GB at this context; duplicate traces of the same session: fq7zfo2t, xvtrw1sv, y482og3o; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1713.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 1,713 88.1 14.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q2ulgfls",
   "source_url": "https://omlx.ai/benchmarks/performance/q2ulgfls",
   "tg_tokens": null,
   "tps": 88.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-q2ulgfls-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 15.7 GB at this context; duplicate traces of the same session: fq7zfo2t, xvtrw1sv, y482og3o; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1168.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 1,168 74.6 15.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q2ulgfls",
   "source_url": "https://omlx.ai/benchmarks/performance/q2ulgfls",
   "tg_tokens": null,
   "tps": 74.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-q2ulgfls-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 17.9 GB at this context; duplicate traces of the same session: fq7zfo2t, xvtrw1sv, y482og3o; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1052.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-oQ4e-fp16 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 32k 1,052 66.9 17.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q2ulgfls",
   "source_url": "https://omlx.ai/benchmarks/performance/q2ulgfls",
   "tg_tokens": null,
   "tps": 66.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-sfpliodw-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 26.1 GB; TTFT 51188 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 640.2,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 32k tokens 640.2 PP tok/s 57.2 TG tok/s 51188 TTFT (ms) 26.1 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sfpliodw",
   "source_url": "https://omlx.ai/benchmarks/performance/sfpliodw",
   "tg_tokens": null,
   "tps": 57.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-9lsthwd5-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 24.8 GB; TTFT 19817 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 826.8,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k tokens 826.8 PP tok/s 61.5 TG tok/s 19817 TTFT (ms) 24.8 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9lsthwd5",
   "source_url": "https://omlx.ai/benchmarks/performance/9lsthwd5",
   "tg_tokens": null,
   "tps": 61.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-3598d1y4-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.1 GB at this context; duplicate traces of the same session: ara36n0h, rri5kdw5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1624.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-fixed-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 1,624 100.7 39.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3598d1y4",
   "source_url": "https://omlx.ai/benchmarks/performance/3598d1y4",
   "tg_tokens": null,
   "tps": 100.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-3598d1y4-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.5 GB at this context; duplicate traces of the same session: ara36n0h, rri5kdw5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1378.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-fixed-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 1,378 122.1 39.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3598d1y4",
   "source_url": "https://omlx.ai/benchmarks/performance/3598d1y4",
   "tg_tokens": null,
   "tps": 122.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-3598d1y4-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 40.2 GB at this context; duplicate traces of the same session: ara36n0h, rri5kdw5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1045.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-fixed-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 1,045 127.4 40.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3598d1y4",
   "source_url": "https://omlx.ai/benchmarks/performance/3598d1y4",
   "tg_tokens": null,
   "tps": 127.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-jzrjjn5x-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.1 GB at this context; duplicate traces of the same session: 9eak5ra9, ep0u2xvx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1600.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 1,600 86.8 39.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jzrjjn5x",
   "source_url": "https://omlx.ai/benchmarks/performance/jzrjjn5x",
   "tg_tokens": null,
   "tps": 86.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-jzrjjn5x-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.5 GB at this context; duplicate traces of the same session: 9eak5ra9, ep0u2xvx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1420.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 1,420 114.1 39.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jzrjjn5x",
   "source_url": "https://omlx.ai/benchmarks/performance/jzrjjn5x",
   "tg_tokens": null,
   "tps": 114.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-jzrjjn5x-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 40.2 GB at this context; duplicate traces of the same session: 9eak5ra9, ep0u2xvx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1117.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 1,117 116.6 40.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jzrjjn5x",
   "source_url": "https://omlx.ai/benchmarks/performance/jzrjjn5x",
   "tg_tokens": null,
   "tps": 116.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-mi076qqn-ctx16384",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 12.6 GB; TTFT 35740 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 458.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-9B-MLX-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 16k tokens 458.4 PP tok/s 35.2 TG tok/s 35740 TTFT (ms) 12.6 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mi076qqn",
   "source_url": "https://omlx.ai/benchmarks/performance/mi076qqn",
   "tg_tokens": null,
   "tps": 35.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-v6tkbq9q-ctx32768",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 14.1 GB; TTFT 80209 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 408.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-9B-MLX-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 32k tokens 408.5 PP tok/s 32.9 TG tok/s 80209 TTFT (ms) 14.1 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v6tkbq9q",
   "source_url": "https://omlx.ai/benchmarks/performance/v6tkbq9q",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-953pk62p-ctx16384",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 12.6 GB; TTFT 35850 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 457.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-9B-MLX-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 16k tokens 457.0 PP tok/s 35.2 TG tok/s 35850 TTFT (ms) 12.6 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/953pk62p",
   "source_url": "https://omlx.ai/benchmarks/performance/953pk62p",
   "tg_tokens": null,
   "tps": 35.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-m2k9gd2p-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 20.9 GB; TTFT 1989 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2059.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k tokens 2,059 PP tok/s 91.3 TG tok/s 1989 TTFT (ms) 20.9 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m2k9gd2p",
   "source_url": "https://omlx.ai/benchmarks/performance/m2k9gd2p",
   "tg_tokens": null,
   "tps": 91.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-o3ece0kx-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 21.3 GB; TTFT 1586 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2583.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k tokens 2,583 PP tok/s 96.2 TG tok/s 1586 TTFT (ms) 21.3 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o3ece0kx",
   "source_url": "https://omlx.ai/benchmarks/performance/o3ece0kx",
   "tg_tokens": null,
   "tps": 96.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-sqo0as0n-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 20.9 GB; TTFT 1971 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2078.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k tokens 2,078 PP tok/s 95.7 TG tok/s 1971 TTFT (ms) 20.9 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sqo0as0n",
   "source_url": "https://omlx.ai/benchmarks/performance/sqo0as0n",
   "tg_tokens": null,
   "tps": 95.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-545aswuf-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 11.2 GB at this context; duplicate traces of the same session: lz9kiim0; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 248.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 248.4 22.1 11.2 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/545aswuf",
   "source_url": "https://omlx.ai/benchmarks/performance/545aswuf",
   "tg_tokens": null,
   "tps": 22.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-545aswuf-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: lz9kiim0; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 557.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 557.3 22.4 11.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/545aswuf",
   "source_url": "https://omlx.ai/benchmarks/performance/545aswuf",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-g3mkr0ex-ctx65536",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-DFlash2",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 36.9 GB; TTFT 75971 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 862.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Coder390-EfficientThink-DFlash2-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k tokens 862.6 PP tok/s 34.3 TG tok/s 75971 TTFT (ms) 36.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g3mkr0ex",
   "source_url": "https://omlx.ai/benchmarks/performance/g3mkr0ex",
   "tg_tokens": null,
   "tps": 34.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-62e85tsi-ctx65536",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-DFlash2",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 39.0 GB; TTFT 75955 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 862.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Coder390-EfficientThink-DFlash2-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k tokens 862.8 PP tok/s 38.9 TG tok/s 75955 TTFT (ms) 39.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/62e85tsi",
   "source_url": "https://omlx.ai/benchmarks/performance/62e85tsi",
   "tg_tokens": null,
   "tps": 38.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-hpt4hohl-ctx1024",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 082clk0m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 201.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 201.1 41.1 18.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hpt4hohl",
   "source_url": "https://omlx.ai/benchmarks/performance/hpt4hohl",
   "tg_tokens": null,
   "tps": 41.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-hpt4hohl-ctx4096",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: 082clk0m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 188.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 188.1 34.1 22.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hpt4hohl",
   "source_url": "https://omlx.ai/benchmarks/performance/hpt4hohl",
   "tg_tokens": null,
   "tps": 34.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-1o4efzeu-ctx1024",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: tf79zmoo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 124.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 124.5 15.8 18.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1o4efzeu",
   "source_url": "https://omlx.ai/benchmarks/performance/1o4efzeu",
   "tg_tokens": null,
   "tps": 15.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-1o4efzeu-ctx4096",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: tf79zmoo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 103.1 10.8 22.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1o4efzeu",
   "source_url": "https://omlx.ai/benchmarks/performance/1o4efzeu",
   "tg_tokens": null,
   "tps": 10.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-82v6y5ko-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.2 GB at this context; duplicate traces of the same session: kw2p53w6, t353m4l2, x07hbcfv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 943.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 943.6 41.0 30.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/82v6y5ko",
   "source_url": "https://omlx.ai/benchmarks/performance/82v6y5ko",
   "tg_tokens": null,
   "tps": 41.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-82v6y5ko-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.9 GB at this context; duplicate traces of the same session: kw2p53w6, t353m4l2, x07hbcfv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 924.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 16k 924.5 51.7 30.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/82v6y5ko",
   "source_url": "https://omlx.ai/benchmarks/performance/82v6y5ko",
   "tg_tokens": null,
   "tps": 51.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-82v6y5ko-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.2 GB at this context; duplicate traces of the same session: kw2p53w6, t353m4l2, x07hbcfv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 764.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 64k 764.1 35.0 34.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/82v6y5ko",
   "source_url": "https://omlx.ai/benchmarks/performance/82v6y5ko",
   "tg_tokens": null,
   "tps": 35.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-82v6y5ko-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.0 GB at this context; duplicate traces of the same session: kw2p53w6, t353m4l2, x07hbcfv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 613.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 128k 613.6 33.1 39.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/82v6y5ko",
   "source_url": "https://omlx.ai/benchmarks/performance/82v6y5ko",
   "tg_tokens": null,
   "tps": 33.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-je1sbggs-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 39.7 GB; TTFT 2852 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 359.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k tokens 359.0 PP tok/s 42.7 TG tok/s 2852 TTFT (ms) 39.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/je1sbggs",
   "source_url": "https://omlx.ai/benchmarks/performance/je1sbggs",
   "tg_tokens": null,
   "tps": 42.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-45x1ojib-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP320",
   "notes": "MTP Vision build; oMLX community benchmark trace; peak memory 38.2 GB at this context; duplicate traces of the same session: 04myqsqi, 08fhrbti, lstqgqkh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 622.8,
   "provenance": "community",
   "quant": "oQ3e-DWQ",
   "quote": "Qwen3.8-Flash-Next-REAP320-oQ3e-DWQ-MTP-Vision-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 3bit; 4k 622.8 39.7 38.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/45x1ojib",
   "source_url": "https://omlx.ai/benchmarks/performance/45x1ojib",
   "tg_tokens": null,
   "tps": 39.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-45x1ojib-ctx16384",
   "model": "Qwen3.8-Flash-Next-REAP320",
   "notes": "MTP Vision build; oMLX community benchmark trace; peak memory 38.2 GB at this context; duplicate traces of the same session: 04myqsqi, 08fhrbti, lstqgqkh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2090.0,
   "provenance": "community",
   "quant": "oQ3e-DWQ",
   "quote": "Qwen3.8-Flash-Next-REAP320-oQ3e-DWQ-MTP-Vision-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 3bit; 16k 2,090 61.3 38.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/45x1ojib",
   "source_url": "https://omlx.ai/benchmarks/performance/45x1ojib",
   "tg_tokens": null,
   "tps": 61.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-45x1ojib-ctx65536",
   "model": "Qwen3.8-Flash-Next-REAP320",
   "notes": "MTP Vision build; oMLX community benchmark trace; peak memory 40.9 GB at this context; duplicate traces of the same session: 04myqsqi, 08fhrbti, lstqgqkh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2007.0,
   "provenance": "community",
   "quant": "oQ3e-DWQ",
   "quote": "Qwen3.8-Flash-Next-REAP320-oQ3e-DWQ-MTP-Vision-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 3bit; 64k 2,007 56.7 40.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/45x1ojib",
   "source_url": "https://omlx.ai/benchmarks/performance/45x1ojib",
   "tg_tokens": null,
   "tps": 56.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-45x1ojib-ctx131072",
   "model": "Qwen3.8-Flash-Next-REAP320",
   "notes": "MTP Vision build; oMLX community benchmark trace; peak memory 45.2 GB at this context; duplicate traces of the same session: 04myqsqi, 08fhrbti, lstqgqkh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1865.0,
   "provenance": "community",
   "quant": "oQ3e-DWQ",
   "quote": "Qwen3.8-Flash-Next-REAP320-oQ3e-DWQ-MTP-Vision-MLX on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 3bit; 128k 1,865 59.2 45.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/45x1ojib",
   "source_url": "https://omlx.ai/benchmarks/performance/45x1ojib",
   "tg_tokens": null,
   "tps": 59.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-bvg5b0al-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 69.9 GB; TTFT 15445 ms; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 265.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 4k tokens 265.2 PP tok/s 36.9 TG tok/s 15445 TTFT (ms) 69.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bvg5b0al",
   "source_url": "https://omlx.ai/benchmarks/performance/bvg5b0al",
   "tg_tokens": null,
   "tps": 36.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-2huqw9qj-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 8.9 GB at this context; duplicate traces of the same session: phsms3yn; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 81.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 81.5 5.1 8.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2huqw9qj",
   "source_url": "https://omlx.ai/benchmarks/performance/2huqw9qj",
   "tg_tokens": null,
   "tps": 5.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-2huqw9qj-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 9.4 GB at this context; duplicate traces of the same session: phsms3yn; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 150.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 150.1 4.7 9.4 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2huqw9qj",
   "source_url": "https://omlx.ai/benchmarks/performance/2huqw9qj",
   "tg_tokens": null,
   "tps": 4.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-2huqw9qj-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 9.4 GB at this context; duplicate traces of the same session: phsms3yn; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 147.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 8k 147.0 5.0 9.4 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2huqw9qj",
   "source_url": "https://omlx.ai/benchmarks/performance/2huqw9qj",
   "tg_tokens": null,
   "tps": 5.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-za0fqypl-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: bgzuv0hj, tuamgw2n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1204.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,204 86.6 70.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/za0fqypl",
   "source_url": "https://omlx.ai/benchmarks/performance/za0fqypl",
   "tg_tokens": null,
   "tps": 86.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-za0fqypl-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.0 GB at this context; duplicate traces of the same session: bgzuv0hj, tuamgw2n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2701.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,701 64.2 72.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/za0fqypl",
   "source_url": "https://omlx.ai/benchmarks/performance/za0fqypl",
   "tg_tokens": null,
   "tps": 64.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-za0fqypl-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.0 GB at this context; duplicate traces of the same session: bgzuv0hj, tuamgw2n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2541.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,541 90.9 73.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/za0fqypl",
   "source_url": "https://omlx.ai/benchmarks/performance/za0fqypl",
   "tg_tokens": null,
   "tps": 90.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-za0fqypl-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.6 GB at this context; duplicate traces of the same session: bgzuv0hj, tuamgw2n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2490.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,490 73.4 73.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/za0fqypl",
   "source_url": "https://omlx.ai/benchmarks/performance/za0fqypl",
   "tg_tokens": null,
   "tps": 73.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ulmpj8rl-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 6s0arzwp, 7aja8am3, pe5s2wo7, rrrtowpu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1188.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,188 80.5 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ulmpj8rl",
   "source_url": "https://omlx.ai/benchmarks/performance/ulmpj8rl",
   "tg_tokens": null,
   "tps": 80.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ulmpj8rl-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: 6s0arzwp, 7aja8am3, pe5s2wo7, rrrtowpu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2065.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,065 107.9 71.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ulmpj8rl",
   "source_url": "https://omlx.ai/benchmarks/performance/ulmpj8rl",
   "tg_tokens": null,
   "tps": 107.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ulmpj8rl-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.0 GB at this context; duplicate traces of the same session: 6s0arzwp, 7aja8am3, pe5s2wo7, rrrtowpu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2659.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,659 56.8 72.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ulmpj8rl",
   "source_url": "https://omlx.ai/benchmarks/performance/ulmpj8rl",
   "tg_tokens": null,
   "tps": 56.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ulmpj8rl-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.0 GB at this context; duplicate traces of the same session: 6s0arzwp, 7aja8am3, pe5s2wo7, rrrtowpu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2519.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,519 82.1 73.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ulmpj8rl",
   "source_url": "https://omlx.ai/benchmarks/performance/ulmpj8rl",
   "tg_tokens": null,
   "tps": 82.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ulmpj8rl-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.6 GB at this context; duplicate traces of the same session: 6s0arzwp, 7aja8am3, pe5s2wo7, rrrtowpu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2486.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,486 86.0 73.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ulmpj8rl",
   "source_url": "https://omlx.ai/benchmarks/performance/ulmpj8rl",
   "tg_tokens": null,
   "tps": 86.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-we8a68ph-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: 3rj8hzjd, 3to9wbpv, ahqy2h26, whvpiy5s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 311.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 311.4 37.1 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/we8a68ph",
   "source_url": "https://omlx.ai/benchmarks/performance/we8a68ph",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-we8a68ph-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: 3rj8hzjd, 3to9wbpv, ahqy2h26, whvpiy5s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 936.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 936.7 31.6 71.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/we8a68ph",
   "source_url": "https://omlx.ai/benchmarks/performance/we8a68ph",
   "tg_tokens": null,
   "tps": 31.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-we8a68ph-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.0 GB at this context; duplicate traces of the same session: 3rj8hzjd, 3to9wbpv, ahqy2h26, whvpiy5s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1659.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,659 37.4 72.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/we8a68ph",
   "source_url": "https://omlx.ai/benchmarks/performance/we8a68ph",
   "tg_tokens": null,
   "tps": 37.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-we8a68ph-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.0 GB at this context; duplicate traces of the same session: 3rj8hzjd, 3to9wbpv, ahqy2h26, whvpiy5s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2420.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,420 71.4 73.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/we8a68ph",
   "source_url": "https://omlx.ai/benchmarks/performance/we8a68ph",
   "tg_tokens": null,
   "tps": 71.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-we8a68ph-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.6 GB at this context; duplicate traces of the same session: 3rj8hzjd, 3to9wbpv, ahqy2h26, whvpiy5s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2452.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,452 58.1 73.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/we8a68ph",
   "source_url": "https://omlx.ai/benchmarks/performance/we8a68ph",
   "tg_tokens": null,
   "tps": 58.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zfbwu9v2-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: g9nji3w7, i6ca5v7o, sr2ks6m8, whq4mq1d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 531.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 531.1 68.8 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zfbwu9v2",
   "source_url": "https://omlx.ai/benchmarks/performance/zfbwu9v2",
   "tg_tokens": null,
   "tps": 68.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zfbwu9v2-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: g9nji3w7, i6ca5v7o, sr2ks6m8, whq4mq1d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1121.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,121 55.0 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zfbwu9v2",
   "source_url": "https://omlx.ai/benchmarks/performance/zfbwu9v2",
   "tg_tokens": null,
   "tps": 55.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zfbwu9v2-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: g9nji3w7, i6ca5v7o, sr2ks6m8, whq4mq1d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1633.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,633 62.3 71.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zfbwu9v2",
   "source_url": "https://omlx.ai/benchmarks/performance/zfbwu9v2",
   "tg_tokens": null,
   "tps": 62.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zfbwu9v2-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: g9nji3w7, i6ca5v7o, sr2ks6m8, whq4mq1d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2453.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,453 80.8 72.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zfbwu9v2",
   "source_url": "https://omlx.ai/benchmarks/performance/zfbwu9v2",
   "tg_tokens": null,
   "tps": 80.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zfbwu9v2-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: g9nji3w7, i6ca5v7o, sr2ks6m8, whq4mq1d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2438.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,438 54.6 73.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zfbwu9v2",
   "source_url": "https://omlx.ai/benchmarks/performance/zfbwu9v2",
   "tg_tokens": null,
   "tps": 54.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kxfzjh8f-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: h2z2zy1a, lkilbq20; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 858.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 858.8 47.3 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kxfzjh8f",
   "source_url": "https://omlx.ai/benchmarks/performance/kxfzjh8f",
   "tg_tokens": null,
   "tps": 47.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kxfzjh8f-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: h2z2zy1a, lkilbq20; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1261.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,261 49.5 19.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kxfzjh8f",
   "source_url": "https://omlx.ai/benchmarks/performance/kxfzjh8f",
   "tg_tokens": null,
   "tps": 49.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kxfzjh8f-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: h2z2zy1a, lkilbq20; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1326.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,326 57.4 19.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kxfzjh8f",
   "source_url": "https://omlx.ai/benchmarks/performance/kxfzjh8f",
   "tg_tokens": null,
   "tps": 57.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-npp0194y-ctx4096",
   "model": "Ternary-Bonsai-2-9B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 4k context; peak memory 6.6 GB; TTFT 21147 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 193.7,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-9B-mlx-2bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 2bit; 4k tokens 193.7 PP tok/s 25.1 TG tok/s 21147 TTFT (ms) 6.6 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/npp0194y",
   "source_url": "https://omlx.ai/benchmarks/performance/npp0194y",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-n7vk06hy-ctx32768",
   "model": "Devstral-Small-2-24B-Instruct-2512",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 19.5 GB; TTFT 249209 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Devstral-Small-2-24B-Instruct-2512-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 32k tokens 131.5 PP tok/s 14.1 TG tok/s 249209 TTFT (ms) 19.5 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n7vk06hy",
   "source_url": "https://omlx.ai/benchmarks/performance/n7vk06hy",
   "tg_tokens": null,
   "tps": 14.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-d01s2e61-ctx16384",
   "model": "Devstral-Small-2-24B-Instruct-2512",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 17.0 GB; TTFT 110210 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 148.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Devstral-Small-2-24B-Instruct-2512-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k tokens 148.7 PP tok/s 19.6 TG tok/s 110210 TTFT (ms) 17.0 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d01s2e61",
   "source_url": "https://omlx.ai/benchmarks/performance/d01s2e61",
   "tg_tokens": null,
   "tps": 19.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dxg5fepn-ctx4096",
   "model": "Devstral-Small-2-24B-Instruct-2512",
   "notes": "oMLX community benchmark trace; peak memory 15.1 GB at this context; duplicate traces of the same session: 8gguwd0d, vanmeumn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1094.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Devstral-Small-2-24B-Instruct-2512-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,094 36.2 15.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dxg5fepn",
   "source_url": "https://omlx.ai/benchmarks/performance/dxg5fepn",
   "tg_tokens": null,
   "tps": 36.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dxg5fepn-ctx32768",
   "model": "Devstral-Small-2-24B-Instruct-2512",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 8gguwd0d, vanmeumn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 821.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Devstral-Small-2-24B-Instruct-2512-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 821.9 28.5 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dxg5fepn",
   "source_url": "https://omlx.ai/benchmarks/performance/dxg5fepn",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dxg5fepn-ctx131072",
   "model": "Devstral-Small-2-24B-Instruct-2512",
   "notes": "oMLX community benchmark trace; peak memory 34.0 GB at this context; duplicate traces of the same session: 8gguwd0d, vanmeumn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 440.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Devstral-Small-2-24B-Instruct-2512-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 440.6 13.5 34.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dxg5fepn",
   "source_url": "https://omlx.ai/benchmarks/performance/dxg5fepn",
   "tg_tokens": null,
   "tps": 13.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3",
   "id": "omlx-nfxmxeow-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 4k context; peak memory 13.5 GB; TTFT 8213 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 498.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 4bit; 4k tokens 498.7 PP tok/s 42.7 TG tok/s 8213 TTFT (ms) 13.5 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nfxmxeow",
   "source_url": "https://omlx.ai/benchmarks/performance/nfxmxeow",
   "tg_tokens": null,
   "tps": 42.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-rlktkn8k-ctx4096",
   "model": "LFM2.5-8B-A1B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 5.2 GB; TTFT 5085 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 805.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-8B-A1B-MLX-4bit on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 4bit; 4k tokens 805.5 PP tok/s 68.7 TG tok/s 5085 TTFT (ms) 5.2 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rlktkn8k",
   "source_url": "https://omlx.ai/benchmarks/performance/rlktkn8k",
   "tg_tokens": null,
   "tps": 68.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-cdyo7wqz-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 45.2 GB at this context; duplicate traces of the same session: 6649hbbe; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4544.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "LiquidAI--LFM2.5-24B-A2B-MLX-bf16 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 4,544 86.5 45.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cdyo7wqz",
   "source_url": "https://omlx.ai/benchmarks/performance/cdyo7wqz",
   "tg_tokens": null,
   "tps": 86.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-cdyo7wqz-ctx131072",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 48.1 GB at this context; duplicate traces of the same session: 6649hbbe; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3793.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "LiquidAI--LFM2.5-24B-A2B-MLX-bf16 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 128k 3,793 53.6 48.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cdyo7wqz",
   "source_url": "https://omlx.ai/benchmarks/performance/cdyo7wqz",
   "tg_tokens": null,
   "tps": 53.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-hy6x3wbj-ctx1024",
   "model": "Nail-Qwen3.6-35B-A3B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: km9buiqs, q0oksnlw; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1860.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX-5bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 1,860 65.2 25.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hy6x3wbj",
   "source_url": "https://omlx.ai/benchmarks/performance/hy6x3wbj",
   "tg_tokens": null,
   "tps": 65.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-hy6x3wbj-ctx4096",
   "model": "Nail-Qwen3.6-35B-A3B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: km9buiqs, q0oksnlw; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2752.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX-5bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 2,752 60.2 26.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hy6x3wbj",
   "source_url": "https://omlx.ai/benchmarks/performance/hy6x3wbj",
   "tg_tokens": null,
   "tps": 60.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-hy6x3wbj-ctx8192",
   "model": "Nail-Qwen3.6-35B-A3B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: km9buiqs, q0oksnlw; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2903.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX-5bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 2,903 63.3 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hy6x3wbj",
   "source_url": "https://omlx.ai/benchmarks/performance/hy6x3wbj",
   "tg_tokens": null,
   "tps": 63.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-5rxfiykh-ctx1024",
   "model": "Nail-Qwen3.6-35B-A3B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: 33i2bcc0, hwya1ifb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1783.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX-5bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 1,783 64.2 25.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5rxfiykh",
   "source_url": "https://omlx.ai/benchmarks/performance/5rxfiykh",
   "tg_tokens": null,
   "tps": 64.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-5rxfiykh-ctx4096",
   "model": "Nail-Qwen3.6-35B-A3B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: 33i2bcc0, hwya1ifb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1828.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX-5bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 1,828 61.6 26.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5rxfiykh",
   "source_url": "https://omlx.ai/benchmarks/performance/5rxfiykh",
   "tg_tokens": null,
   "tps": 61.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-5rxfiykh-ctx8192",
   "model": "Nail-Qwen3.6-35B-A3B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: 33i2bcc0, hwya1ifb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2525.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX-5bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 2,525 60.6 26.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5rxfiykh",
   "source_url": "https://omlx.ai/benchmarks/performance/5rxfiykh",
   "tg_tokens": null,
   "tps": 60.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-02y1nmhp-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: 7o7utaoe; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 167.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 167.6 18.1 6.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/02y1nmhp",
   "source_url": "https://omlx.ai/benchmarks/performance/02y1nmhp",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-02y1nmhp-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 8.3 GB at this context; duplicate traces of the same session: 7o7utaoe; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 468.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 468.4 19.5 8.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/02y1nmhp",
   "source_url": "https://omlx.ai/benchmarks/performance/02y1nmhp",
   "tg_tokens": null,
   "tps": 19.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-02y1nmhp-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 10.1 GB at this context; duplicate traces of the same session: 7o7utaoe; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 360.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 32k 360.2 17.0 10.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/02y1nmhp",
   "source_url": "https://omlx.ai/benchmarks/performance/02y1nmhp",
   "tg_tokens": null,
   "tps": 17.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-s57jhg1s-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 35.5 GB; TTFT 1391 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 735.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k tokens 735.9 PP tok/s 108.1 TG tok/s 1391 TTFT (ms) 35.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s57jhg1s",
   "source_url": "https://omlx.ai/benchmarks/performance/s57jhg1s",
   "tg_tokens": null,
   "tps": 108.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-xdv5q0cb-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.7 GB at this context; duplicate traces of the same session: q4uw6a6n; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 655.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 24 GB \u00b7 4bit; 8k 655.5 22.9 7.7 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xdv5q0cb",
   "source_url": "https://omlx.ai/benchmarks/performance/xdv5q0cb",
   "tg_tokens": null,
   "tps": 22.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-xdv5q0cb-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.9 GB at this context; duplicate traces of the same session: q4uw6a6n; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 651.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 24 GB \u00b7 4bit; 16k 651.5 19.8 7.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xdv5q0cb",
   "source_url": "https://omlx.ai/benchmarks/performance/xdv5q0cb",
   "tg_tokens": null,
   "tps": 19.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-tubofwak-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 9.0 GB at this context; duplicate traces of the same session: bpzfkcla, i0yg0wbj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 537.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 24 GB \u00b7 4bit; 4k 537.7 18.0 9.0 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tubofwak",
   "source_url": "https://omlx.ai/benchmarks/performance/tubofwak",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-tubofwak-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 9.1 GB at this context; duplicate traces of the same session: bpzfkcla, i0yg0wbj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 609.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 24 GB \u00b7 4bit; 8k 609.3 17.3 9.1 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tubofwak",
   "source_url": "https://omlx.ai/benchmarks/performance/tubofwak",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-tubofwak-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 9.2 GB at this context; duplicate traces of the same session: bpzfkcla, i0yg0wbj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 635.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 24 GB \u00b7 4bit; 16k 635.3 18.4 9.2 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tubofwak",
   "source_url": "https://omlx.ai/benchmarks/performance/tubofwak",
   "tg_tokens": null,
   "tps": 18.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-h3ud781q-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: 7iva4zq1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 481.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 481.7 24.1 20.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h3ud781q",
   "source_url": "https://omlx.ai/benchmarks/performance/h3ud781q",
   "tg_tokens": null,
   "tps": 24.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-h3ud781q-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: 7iva4zq1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 406.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 406.5 25.4 22.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h3ud781q",
   "source_url": "https://omlx.ai/benchmarks/performance/h3ud781q",
   "tg_tokens": null,
   "tps": 25.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-46dswkh1-ctx4096",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 31.1 GB; TTFT 10546 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 388.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Alis-MLX-6bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k tokens 388.4 PP tok/s 35.6 TG tok/s 10546 TTFT (ms) 31.1 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/46dswkh1",
   "source_url": "https://omlx.ai/benchmarks/performance/46dswkh1",
   "tg_tokens": null,
   "tps": 35.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3jqexy7l-ctx4096",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 31.1 GB; TTFT 10734 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 381.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Alis-MLX-6bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k tokens 381.6 PP tok/s 41.1 TG tok/s 10734 TTFT (ms) 31.1 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3jqexy7l",
   "source_url": "https://omlx.ai/benchmarks/performance/3jqexy7l",
   "tg_tokens": null,
   "tps": 41.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1uc7i1qf-ctx4096",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 31.1 GB; TTFT 10783 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 379.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Alis-MLX-6bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k tokens 379.9 PP tok/s 43.2 TG tok/s 10783 TTFT (ms) 31.1 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1uc7i1qf",
   "source_url": "https://omlx.ai/benchmarks/performance/1uc7i1qf",
   "tg_tokens": null,
   "tps": 43.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-f96eajfq-ctx4096",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 27.5 GB; TTFT 13042 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 314.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Alis-MLX-6bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k tokens 314.1 PP tok/s 40.6 TG tok/s 13042 TTFT (ms) 27.5 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f96eajfq",
   "source_url": "https://omlx.ai/benchmarks/performance/f96eajfq",
   "tg_tokens": null,
   "tps": 40.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-iznlml9h-ctx4096",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 31.1 GB; TTFT 10682 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 383.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Alis-MLX-6bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k tokens 383.4 PP tok/s 41.1 TG tok/s 10682 TTFT (ms) 31.1 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iznlml9h",
   "source_url": "https://omlx.ai/benchmarks/performance/iznlml9h",
   "tg_tokens": null,
   "tps": 41.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-ut83kie4-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: sjy078fv; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 81.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 81.6 14.9 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ut83kie4",
   "source_url": "https://omlx.ai/benchmarks/performance/ut83kie4",
   "tg_tokens": null,
   "tps": 14.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-ut83kie4-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: sjy078fv; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 79.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 79.0 15.0 20.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ut83kie4",
   "source_url": "https://omlx.ai/benchmarks/performance/ut83kie4",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qb43wcxa-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: 3i494msd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 939.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 939.8 32.6 17.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qb43wcxa",
   "source_url": "https://omlx.ai/benchmarks/performance/qb43wcxa",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qb43wcxa-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 3i494msd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 908.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 908.8 30.7 17.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qb43wcxa",
   "source_url": "https://omlx.ai/benchmarks/performance/qb43wcxa",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-l1kvvocb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 8el88w8k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 724.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 724.1 17.8 28.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l1kvvocb",
   "source_url": "https://omlx.ai/benchmarks/performance/l1kvvocb",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-l1kvvocb-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: 8el88w8k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 880.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 880.8 18.7 29.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l1kvvocb",
   "source_url": "https://omlx.ai/benchmarks/performance/l1kvvocb",
   "tg_tokens": null,
   "tps": 18.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-lt8j3471-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "llama.cpp MLX build; oMLX community benchmark trace, single run at 4k context; peak memory 20.0 GB; TTFT 40219 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 101.8,
   "provenance": "community",
   "quant": "Q4_K_M",
   "quote": "Qwen3.8-27B-mlx-Q4_K_M on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k tokens 101.8 PP tok/s 13.5 TG tok/s 40219 TTFT (ms) 20.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lt8j3471",
   "source_url": "https://omlx.ai/benchmarks/performance/lt8j3471",
   "tg_tokens": null,
   "tps": 13.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-t0gqxoie-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 17.8 GB; TTFT 912 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1123.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,123 PP tok/s 58.5 TG tok/s 912 TTFT (ms) 17.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t0gqxoie",
   "source_url": "https://omlx.ai/benchmarks/performance/t0gqxoie",
   "tg_tokens": null,
   "tps": 58.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9jxmq33y-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: 29mot38x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 902.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 902.0 62.2 18.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9jxmq33y",
   "source_url": "https://omlx.ai/benchmarks/performance/9jxmq33y",
   "tg_tokens": null,
   "tps": 62.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9jxmq33y-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 29mot38x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 831.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 831.5 73.5 19.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9jxmq33y",
   "source_url": "https://omlx.ai/benchmarks/performance/9jxmq33y",
   "tg_tokens": null,
   "tps": 73.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-tupyj86x-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 100.3 GB; TTFT 8940 ms; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 458.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 4k tokens 458.2 PP tok/s 36.4 TG tok/s 8940 TTFT (ms) 100.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tupyj86x",
   "source_url": "https://omlx.ai/benchmarks/performance/tupyj86x",
   "tg_tokens": null,
   "tps": 36.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-do1r02kg-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 69.9 GB; TTFT 14659 ms; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 279.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 4k tokens 279.4 PP tok/s 36.8 TG tok/s 14659 TTFT (ms) 69.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/do1r02kg",
   "source_url": "https://omlx.ai/benchmarks/performance/do1r02kg",
   "tg_tokens": null,
   "tps": 36.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-7se6cq31-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: g2ap1h2f; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 252.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 252.3 39.5 70.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7se6cq31",
   "source_url": "https://omlx.ai/benchmarks/performance/7se6cq31",
   "tg_tokens": null,
   "tps": 39.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-7se6cq31-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: g2ap1h2f; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 441.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 441.9 44.0 71.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7se6cq31",
   "source_url": "https://omlx.ai/benchmarks/performance/7se6cq31",
   "tg_tokens": null,
   "tps": 44.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-yxn0qc52-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: m5xpdm5s; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1195.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,195 81.5 100.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yxn0qc52",
   "source_url": "https://omlx.ai/benchmarks/performance/yxn0qc52",
   "tg_tokens": null,
   "tps": 81.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-yxn0qc52-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: m5xpdm5s; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1428.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,428 76.7 101.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yxn0qc52",
   "source_url": "https://omlx.ai/benchmarks/performance/yxn0qc52",
   "tg_tokens": null,
   "tps": 76.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-z8zq8rlz-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 70.3 GB; TTFT 22234 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 736.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k tokens 736.9 PP tok/s 53.6 TG tok/s 22234 TTFT (ms) 70.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z8zq8rlz",
   "source_url": "https://omlx.ai/benchmarks/performance/z8zq8rlz",
   "tg_tokens": null,
   "tps": 53.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hi3f7bbr-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.8 GB; TTFT 1156 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 885.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 885.9 PP tok/s 57.1 TG tok/s 1156 TTFT (ms) 70.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hi3f7bbr",
   "source_url": "https://omlx.ai/benchmarks/performance/hi3f7bbr",
   "tg_tokens": null,
   "tps": 57.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-flu50uvy-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 92.9 GB; TTFT 60196 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2177.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k tokens 2,177 PP tok/s 61.1 TG tok/s 60196 TTFT (ms) 92.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/flu50uvy",
   "source_url": "https://omlx.ai/benchmarks/performance/flu50uvy",
   "tg_tokens": null,
   "tps": 61.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dgzjreia-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 92.9 GB; TTFT 60067 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2182.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k tokens 2,182 PP tok/s 61.5 TG tok/s 60067 TTFT (ms) 92.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dgzjreia",
   "source_url": "https://omlx.ai/benchmarks/performance/dgzjreia",
   "tg_tokens": null,
   "tps": 61.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-dgb46t2o-ctx1024",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: e9p29ukv, ybhvgbv0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 102.6 26.0 20.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dgb46t2o",
   "source_url": "https://omlx.ai/benchmarks/performance/dgb46t2o",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-dgb46t2o-ctx4096",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: e9p29ukv, ybhvgbv0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 149.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 149.7 19.2 23.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dgb46t2o",
   "source_url": "https://omlx.ai/benchmarks/performance/dgb46t2o",
   "tg_tokens": null,
   "tps": 19.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-dgb46t2o-ctx32768",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: e9p29ukv, ybhvgbv0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 119.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 32k 119.6 28.6 27.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dgb46t2o",
   "source_url": "https://omlx.ai/benchmarks/performance/dgb46t2o",
   "tg_tokens": null,
   "tps": 28.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-7mdqwabt-ctx1024",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: aqhtazt0; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 81.9,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 2bit; 1k 81.9 10.7 11.8 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7mdqwabt",
   "source_url": "https://omlx.ai/benchmarks/performance/7mdqwabt",
   "tg_tokens": null,
   "tps": 10.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Pro",
   "id": "omlx-7mdqwabt-ctx4096",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 14.5 GB at this context; duplicate traces of the same session: aqhtazt0; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 83.8,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 2bit; 4k 83.8 9.6 14.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7mdqwabt",
   "source_url": "https://omlx.ai/benchmarks/performance/7mdqwabt",
   "tg_tokens": null,
   "tps": 9.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-l4pktohq-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 288.8 GB at this context; duplicate traces of the same session: 97xprccs; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 519.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 519.1 31.1 288.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l4pktohq",
   "source_url": "https://omlx.ai/benchmarks/performance/l4pktohq",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-l4pktohq-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 289.3 GB at this context; duplicate traces of the same session: 97xprccs; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 679.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 679.8 31.4 289.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l4pktohq",
   "source_url": "https://omlx.ai/benchmarks/performance/l4pktohq",
   "tg_tokens": null,
   "tps": 31.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nh0zsj80-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 288.9 GB at this context; duplicate traces of the same session: ult5a9bq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 369.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 369.4 26.6 288.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nh0zsj80",
   "source_url": "https://omlx.ai/benchmarks/performance/nh0zsj80",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nh0zsj80-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 289.3 GB at this context; duplicate traces of the same session: ult5a9bq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 350.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 350.1 30.2 289.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nh0zsj80",
   "source_url": "https://omlx.ai/benchmarks/performance/nh0zsj80",
   "tg_tokens": null,
   "tps": 30.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-82escbyd-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 403.3 GB at this context; duplicate traces of the same session: wxdv63m6; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 535.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 535.5 28.6 403.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/82escbyd",
   "source_url": "https://omlx.ai/benchmarks/performance/82escbyd",
   "tg_tokens": null,
   "tps": 28.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-82escbyd-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 403.7 GB at this context; duplicate traces of the same session: wxdv63m6; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 719.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 719.6 31.8 403.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/82escbyd",
   "source_url": "https://omlx.ai/benchmarks/performance/82escbyd",
   "tg_tokens": null,
   "tps": 31.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-85tliara-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 420.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ4e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 420.1 30.1 17.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/85tliara",
   "source_url": "https://omlx.ai/benchmarks/performance/85tliara",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-85tliara-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 471.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ4e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 471.2 34.6 18.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/85tliara",
   "source_url": "https://omlx.ai/benchmarks/performance/85tliara",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-85tliara-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 464.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ4e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 464.6 36.1 19.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/85tliara",
   "source_url": "https://omlx.ai/benchmarks/performance/85tliara",
   "tg_tokens": null,
   "tps": 36.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3",
   "id": "omlx-8o76c13w-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 32k context; peak memory 14.1 GB; TTFT 123897 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 264.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 4bit; 32k tokens 264.5 PP tok/s 12.6 TG tok/s 123897 TTFT (ms) 14.1 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8o76c13w",
   "source_url": "https://omlx.ai/benchmarks/performance/8o76c13w",
   "tg_tokens": null,
   "tps": 12.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-g17uxy0l-ctx131072",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 128k context; peak memory 16.3 GB; TTFT 30723 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4266.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "LFM2.5-24B-A2B-MLX-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k tokens 4,266 PP tok/s 75.5 TG tok/s 30723 TTFT (ms) 16.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g17uxy0l",
   "source_url": "https://omlx.ai/benchmarks/performance/g17uxy0l",
   "tg_tokens": null,
   "tps": 75.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-5oypt55j-ctx32768",
   "model": "LFM2.5-8B-A1B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 5.5 GB; TTFT 53068 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 617.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-8B-A1B-MLX-4bit on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 4bit; 32k tokens 617.5 PP tok/s 51.0 TG tok/s 53068 TTFT (ms) 5.5 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5oypt55j",
   "source_url": "https://omlx.ai/benchmarks/performance/5oypt55j",
   "tg_tokens": null,
   "tps": 51.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-22hwu1xz-ctx16384",
   "model": "Mellum2-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 7.4 GB; TTFT 14722 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1113.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Mellum2-12B-A2.5B-Thinking-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k tokens 1,113 PP tok/s 107.8 TG tok/s 14722 TTFT (ms) 7.4 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/22hwu1xz",
   "source_url": "https://omlx.ai/benchmarks/performance/22hwu1xz",
   "tg_tokens": null,
   "tps": 107.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-n5tuilx1-ctx32768",
   "model": "Mellum2-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 7.6 GB; TTFT 33530 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 977.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Mellum2-12B-A2.5B-Thinking-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 32k tokens 977.3 PP tok/s 93.6 TG tok/s 33530 TTFT (ms) 7.6 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n5tuilx1",
   "source_url": "https://omlx.ai/benchmarks/performance/n5tuilx1",
   "tg_tokens": null,
   "tps": 93.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-wc6i68z7-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 167.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 167.6 18.1 6.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wc6i68z7",
   "source_url": "https://omlx.ai/benchmarks/performance/wc6i68z7",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-wc6i68z7-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 8.3 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 468.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 468.4 19.5 8.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wc6i68z7",
   "source_url": "https://omlx.ai/benchmarks/performance/wc6i68z7",
   "tg_tokens": null,
   "tps": 19.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-wc6i68z7-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 10.1 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 360.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 32k 360.2 17.0 10.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wc6i68z7",
   "source_url": "https://omlx.ai/benchmarks/performance/wc6i68z7",
   "tg_tokens": null,
   "tps": 17.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-w0l1ftmo-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "REAP-19B build; oMLX community benchmark trace; peak memory 12.9 GB at this context; duplicate traces of the same session: 7w2hvy7u, midc1yvu, trcs3dfk; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 401.8,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit-REAP-19B on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 401.8 42.5 12.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w0l1ftmo",
   "source_url": "https://omlx.ai/benchmarks/performance/w0l1ftmo",
   "tg_tokens": null,
   "tps": 42.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-w0l1ftmo-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "REAP-19B build; oMLX community benchmark trace; peak memory 13.5 GB at this context; duplicate traces of the same session: 7w2hvy7u, midc1yvu, trcs3dfk; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 414.9,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit-REAP-19B on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 414.9 41.8 13.5 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w0l1ftmo",
   "source_url": "https://omlx.ai/benchmarks/performance/w0l1ftmo",
   "tg_tokens": null,
   "tps": 41.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-w0l1ftmo-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "REAP-19B build; oMLX community benchmark trace; peak memory 13.8 GB at this context; duplicate traces of the same session: 7w2hvy7u, midc1yvu, trcs3dfk; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 406.9,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit-REAP-19B on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 8k 406.9 40.3 13.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w0l1ftmo",
   "source_url": "https://omlx.ai/benchmarks/performance/w0l1ftmo",
   "tg_tokens": null,
   "tps": 40.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-w0l1ftmo-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "REAP-19B build; oMLX community benchmark trace; peak memory 13.9 GB at this context; duplicate traces of the same session: 7w2hvy7u, midc1yvu, trcs3dfk; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 344.1,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit-REAP-19B on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k 344.1 37.7 13.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w0l1ftmo",
   "source_url": "https://omlx.ai/benchmarks/performance/w0l1ftmo",
   "tg_tokens": null,
   "tps": 37.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-e3vtlzqb-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "REAP-19B build; oMLX community benchmark trace; peak memory 12.9 GB at this context; duplicate traces of the same session: g57t6lpr, ltr0nrmk, ofga60jk; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 386.3,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit-REAP-19B on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 386.3 43.2 12.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e3vtlzqb",
   "source_url": "https://omlx.ai/benchmarks/performance/e3vtlzqb",
   "tg_tokens": null,
   "tps": 43.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-e3vtlzqb-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "REAP-19B build; oMLX community benchmark trace; peak memory 13.5 GB at this context; duplicate traces of the same session: g57t6lpr, ltr0nrmk, ofga60jk; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 421.4,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit-REAP-19B on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 421.4 42.0 13.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e3vtlzqb",
   "source_url": "https://omlx.ai/benchmarks/performance/e3vtlzqb",
   "tg_tokens": null,
   "tps": 42.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-e3vtlzqb-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "REAP-19B build; oMLX community benchmark trace; peak memory 13.8 GB at this context; duplicate traces of the same session: g57t6lpr, ltr0nrmk, ofga60jk; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 408.3,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit-REAP-19B on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 8k 408.3 40.5 13.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e3vtlzqb",
   "source_url": "https://omlx.ai/benchmarks/performance/e3vtlzqb",
   "tg_tokens": null,
   "tps": 40.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-e3vtlzqb-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "REAP-19B build; oMLX community benchmark trace; peak memory 13.9 GB at this context; duplicate traces of the same session: g57t6lpr, ltr0nrmk, ofga60jk; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 344.1,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit-REAP-19B on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k 344.1 37.8 13.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e3vtlzqb",
   "source_url": "https://omlx.ai/benchmarks/performance/e3vtlzqb",
   "tg_tokens": null,
   "tps": 37.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-11aojeql-ctx16384",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 8.5 GB; TTFT 38107 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 430.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k tokens 430.0 PP tok/s 55.4 TG tok/s 38107 TTFT (ms) 8.5 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/11aojeql",
   "source_url": "https://omlx.ai/benchmarks/performance/11aojeql",
   "tg_tokens": null,
   "tps": 55.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-lr48dr1n-ctx4096",
   "model": "Qwen3-VL-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: xdxaebc0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3132.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-VL-30B-A3B-Instruct-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,132 88.4 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lr48dr1n",
   "source_url": "https://omlx.ai/benchmarks/performance/lr48dr1n",
   "tg_tokens": null,
   "tps": 88.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-lr48dr1n-ctx16384",
   "model": "Qwen3-VL-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: xdxaebc0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2078.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-VL-30B-A3B-Instruct-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,078 61.2 18.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lr48dr1n",
   "source_url": "https://omlx.ai/benchmarks/performance/lr48dr1n",
   "tg_tokens": null,
   "tps": 61.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-92x1a5oc-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 22.1 GB; TTFT 133259 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 122.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k tokens 122.9 PP tok/s 18.1 TG tok/s 133259 TTFT (ms) 22.1 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/92x1a5oc",
   "source_url": "https://omlx.ai/benchmarks/performance/92x1a5oc",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-0bg1wfw1-ctx4096",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 31.1 GB; TTFT 10539 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 388.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Alis-MLX-6bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k tokens 388.7 PP tok/s 37.1 TG tok/s 10539 TTFT (ms) 31.1 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0bg1wfw1",
   "source_url": "https://omlx.ai/benchmarks/performance/0bg1wfw1",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-b8tush0o-ctx4096",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 31.1 GB; TTFT 10403 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 393.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Alis-MLX-6bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k tokens 393.7 PP tok/s 27.7 TG tok/s 10403 TTFT (ms) 31.1 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b8tush0o",
   "source_url": "https://omlx.ai/benchmarks/performance/b8tush0o",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-7i4dpyk6-ctx4096",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 31.1 GB; TTFT 10332 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 396.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Alis-MLX-6bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k tokens 396.4 PP tok/s 43.0 TG tok/s 10332 TTFT (ms) 31.1 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7i4dpyk6",
   "source_url": "https://omlx.ai/benchmarks/performance/7i4dpyk6",
   "tg_tokens": null,
   "tps": 43.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-v9c2tc39-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: p5xxe1g2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 322.4,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 322.4 36.8 16.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v9c2tc39",
   "source_url": "https://omlx.ai/benchmarks/performance/v9c2tc39",
   "tg_tokens": null,
   "tps": 36.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-v9c2tc39-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: p5xxe1g2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 356.4,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 356.4 36.1 19.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v9c2tc39",
   "source_url": "https://omlx.ai/benchmarks/performance/v9c2tc39",
   "tg_tokens": null,
   "tps": 36.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-uq9o90vd-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 7u45by61; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1543.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,543 52.5 16.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uq9o90vd",
   "source_url": "https://omlx.ai/benchmarks/performance/uq9o90vd",
   "tg_tokens": null,
   "tps": 52.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-uq9o90vd-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 7u45by61; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1812.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 1,812 51.6 17.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uq9o90vd",
   "source_url": "https://omlx.ai/benchmarks/performance/uq9o90vd",
   "tg_tokens": null,
   "tps": 51.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-85ktwvag-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: k9dbmu0v; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1237.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,237 51.9 16.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/85ktwvag",
   "source_url": "https://omlx.ai/benchmarks/performance/85ktwvag",
   "tg_tokens": null,
   "tps": 51.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-85ktwvag-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: k9dbmu0v; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1545.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 1,545 51.2 17.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/85ktwvag",
   "source_url": "https://omlx.ai/benchmarks/performance/85ktwvag",
   "tg_tokens": null,
   "tps": 51.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vmi2y52o-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 17.7 GB; TTFT 1344 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 761.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 761.8 PP tok/s 60.6 TG tok/s 1344 TTFT (ms) 17.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vmi2y52o",
   "source_url": "https://omlx.ai/benchmarks/performance/vmi2y52o",
   "tg_tokens": null,
   "tps": 60.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0ylvbkwe-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6 mixed-quant MTP v2 build; oMLX community benchmark trace; peak memory 77.6 GB at this context; duplicate traces of the same session: 3is2giwd, 6cbq8y2s, ebpto2sp, eihrfxds; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 874.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp-v2 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 874.6 60.7 77.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ylvbkwe",
   "source_url": "https://omlx.ai/benchmarks/performance/0ylvbkwe",
   "tg_tokens": null,
   "tps": 60.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0ylvbkwe-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6 mixed-quant MTP v2 build; oMLX community benchmark trace; peak memory 79.0 GB at this context; duplicate traces of the same session: 3is2giwd, 6cbq8y2s, ebpto2sp, eihrfxds; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2179.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp-v2 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 2,179 69.8 79.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ylvbkwe",
   "source_url": "https://omlx.ai/benchmarks/performance/0ylvbkwe",
   "tg_tokens": null,
   "tps": 69.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0ylvbkwe-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6 mixed-quant MTP v2 build; oMLX community benchmark trace; peak memory 80.1 GB at this context; duplicate traces of the same session: 3is2giwd, 6cbq8y2s, ebpto2sp, eihrfxds; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2400.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp-v2 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 2,400 74.3 80.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ylvbkwe",
   "source_url": "https://omlx.ai/benchmarks/performance/0ylvbkwe",
   "tg_tokens": null,
   "tps": 74.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0ylvbkwe-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6 mixed-quant MTP v2 build; oMLX community benchmark trace; peak memory 80.7 GB at this context; duplicate traces of the same session: 3is2giwd, 6cbq8y2s, ebpto2sp, eihrfxds; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2325.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp-v2 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 2,325 54.8 80.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ylvbkwe",
   "source_url": "https://omlx.ai/benchmarks/performance/0ylvbkwe",
   "tg_tokens": null,
   "tps": 54.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0ylvbkwe-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6 mixed-quant MTP v2 build; oMLX community benchmark trace; peak memory 81.8 GB at this context; duplicate traces of the same session: 3is2giwd, 6cbq8y2s, ebpto2sp, eihrfxds; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2252.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp-v2 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k 2,252 71.9 81.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ylvbkwe",
   "source_url": "https://omlx.ai/benchmarks/performance/0ylvbkwe",
   "tg_tokens": null,
   "tps": 71.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-cfh4ocxu-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.8 GB; TTFT 2376 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 431.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k tokens 431.0 PP tok/s 52.3 TG tok/s 2376 TTFT (ms) 70.8 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cfh4ocxu",
   "source_url": "https://omlx.ai/benchmarks/performance/cfh4ocxu",
   "tg_tokens": null,
   "tps": 52.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-gazv8z54-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 71.8 GB; TTFT 22585 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 725.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k tokens 725.4 PP tok/s 66.2 TG tok/s 22585 TTFT (ms) 71.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gazv8z54",
   "source_url": "https://omlx.ai/benchmarks/performance/gazv8z54",
   "tg_tokens": null,
   "tps": 66.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-3bksansj-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 8.9 GB at this context; duplicate traces of the same session: y5jw26bd; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 86.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 86.2 5.1 8.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3bksansj",
   "source_url": "https://omlx.ai/benchmarks/performance/3bksansj",
   "tg_tokens": null,
   "tps": 5.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-3bksansj-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 9.4 GB at this context; duplicate traces of the same session: y5jw26bd; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 149.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 149.8 4.9 9.4 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3bksansj",
   "source_url": "https://omlx.ai/benchmarks/performance/3bksansj",
   "tg_tokens": null,
   "tps": 4.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-3bksansj-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 9.4 GB at this context; duplicate traces of the same session: y5jw26bd; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 135.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 8k 135.3 4.8 9.4 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3bksansj",
   "source_url": "https://omlx.ai/benchmarks/performance/3bksansj",
   "tg_tokens": null,
   "tps": 4.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-3bksansj-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 10.2 GB at this context; duplicate traces of the same session: y5jw26bd; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 32k 127.1 4.5 10.2 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3bksansj",
   "source_url": "https://omlx.ai/benchmarks/performance/3bksansj",
   "tg_tokens": null,
   "tps": 4.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-3bksansj-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 11.6 GB at this context; duplicate traces of the same session: y5jw26bd; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 122.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 64k 122.6 4.2 11.6 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3bksansj",
   "source_url": "https://omlx.ai/benchmarks/performance/3bksansj",
   "tg_tokens": null,
   "tps": 4.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-3bksansj-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 13.9 GB at this context; duplicate traces of the same session: y5jw26bd; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 113.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 128k 113.2 4.1 13.9 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3bksansj",
   "source_url": "https://omlx.ai/benchmarks/performance/3bksansj",
   "tg_tokens": null,
   "tps": 4.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-3bksansj-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: y5jw26bd; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 195k 106.1 4.1 16.1 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3bksansj",
   "source_url": "https://omlx.ai/benchmarks/performance/3bksansj",
   "tg_tokens": null,
   "tps": 4.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ww9j8u1b-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 1q73rc7t, ku62lrlf, pkuwkq8t, tbzzydoc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1232.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,232 66.5 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ww9j8u1b",
   "source_url": "https://omlx.ai/benchmarks/performance/ww9j8u1b",
   "tg_tokens": null,
   "tps": 66.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ww9j8u1b-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: 1q73rc7t, ku62lrlf, pkuwkq8t, tbzzydoc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2098.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,098 97.8 71.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ww9j8u1b",
   "source_url": "https://omlx.ai/benchmarks/performance/ww9j8u1b",
   "tg_tokens": null,
   "tps": 97.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ww9j8u1b-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.0 GB at this context; duplicate traces of the same session: 1q73rc7t, ku62lrlf, pkuwkq8t, tbzzydoc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2442.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,442 64.9 72.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ww9j8u1b",
   "source_url": "https://omlx.ai/benchmarks/performance/ww9j8u1b",
   "tg_tokens": null,
   "tps": 64.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ww9j8u1b-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.0 GB at this context; duplicate traces of the same session: 1q73rc7t, ku62lrlf, pkuwkq8t, tbzzydoc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2525.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,525 85.7 73.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ww9j8u1b",
   "source_url": "https://omlx.ai/benchmarks/performance/ww9j8u1b",
   "tg_tokens": null,
   "tps": 85.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ww9j8u1b-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.6 GB at this context; duplicate traces of the same session: 1q73rc7t, ku62lrlf, pkuwkq8t, tbzzydoc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2475.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,475 89.3 73.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ww9j8u1b",
   "source_url": "https://omlx.ai/benchmarks/performance/ww9j8u1b",
   "tg_tokens": null,
   "tps": 89.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-f0zsn4uo-ctx4096",
   "model": "Spark-X2.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 8.3 GB at this context; duplicate traces of the same session: 87q1ghr1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1257.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Spark-X2.5-4B on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 unknown; 4k 1,257 39.8 8.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f0zsn4uo",
   "source_url": "https://omlx.ai/benchmarks/performance/f0zsn4uo",
   "tg_tokens": null,
   "tps": 39.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-f0zsn4uo-ctx16384",
   "model": "Spark-X2.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 8.6 GB at this context; duplicate traces of the same session: 87q1ghr1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1153.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Spark-X2.5-4B on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 unknown; 16k 1,153 37.8 8.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f0zsn4uo",
   "source_url": "https://omlx.ai/benchmarks/performance/f0zsn4uo",
   "tg_tokens": null,
   "tps": 37.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-vefvnqu9-ctx4096",
   "model": "Spark-X2.5-4B",
   "notes": "OptiQ build; oMLX community benchmark trace; peak memory 3.8 GB at this context; duplicate traces of the same session: jvfa5456; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1212.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Spark-X2.5-4B-OptiQ-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,212 91.8 3.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vefvnqu9",
   "source_url": "https://omlx.ai/benchmarks/performance/vefvnqu9",
   "tg_tokens": null,
   "tps": 91.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-vefvnqu9-ctx16384",
   "model": "Spark-X2.5-4B",
   "notes": "OptiQ build; oMLX community benchmark trace; peak memory 4.1 GB at this context; duplicate traces of the same session: jvfa5456; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1100.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Spark-X2.5-4B-OptiQ-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 1,100 82.5 4.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vefvnqu9",
   "source_url": "https://omlx.ai/benchmarks/performance/vefvnqu9",
   "tg_tokens": null,
   "tps": 82.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-bz4iwe2j-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 403.3 GB at this context; duplicate traces of the same session: sc6rckqa; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 558.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 558.0 27.8 403.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bz4iwe2j",
   "source_url": "https://omlx.ai/benchmarks/performance/bz4iwe2j",
   "tg_tokens": null,
   "tps": 27.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-bz4iwe2j-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 403.7 GB at this context; duplicate traces of the same session: sc6rckqa; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 721.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 721.3 30.3 403.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bz4iwe2j",
   "source_url": "https://omlx.ai/benchmarks/performance/bz4iwe2j",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ec7u479n-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 6o2mgup0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 420.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ4e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 420.1 30.1 17.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ec7u479n",
   "source_url": "https://omlx.ai/benchmarks/performance/ec7u479n",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ec7u479n-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: 6o2mgup0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 471.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ4e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 471.2 34.6 18.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ec7u479n",
   "source_url": "https://omlx.ai/benchmarks/performance/ec7u479n",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ec7u479n-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "Dirk build; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 6o2mgup0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 464.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Dirk-Qwen3.8-27B-MLX-oQ4e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 464.6 36.1 19.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ec7u479n",
   "source_url": "https://omlx.ai/benchmarks/performance/ec7u479n",
   "tg_tokens": null,
   "tps": 36.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-hpd8khgd-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.6 GB at this context; duplicate traces of the same session: osh0nfut; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 503.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 503.7 51.6 177.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hpd8khgd",
   "source_url": "https://omlx.ai/benchmarks/performance/hpd8khgd",
   "tg_tokens": null,
   "tps": 51.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hpd8khgd-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.8 GB at this context; duplicate traces of the same session: osh0nfut; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 592.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 592.3 43.2 178.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hpd8khgd",
   "source_url": "https://omlx.ai/benchmarks/performance/hpd8khgd",
   "tg_tokens": null,
   "tps": 43.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-0tzcoorx-ctx1024",
   "model": "Nail-Qwen3.6-35B-A3B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: ac7cntkp, m5z42eaj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2004.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX-5bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 2,004 62.5 25.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0tzcoorx",
   "source_url": "https://omlx.ai/benchmarks/performance/0tzcoorx",
   "tg_tokens": null,
   "tps": 62.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-0tzcoorx-ctx4096",
   "model": "Nail-Qwen3.6-35B-A3B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: ac7cntkp, m5z42eaj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2869.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX-5bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 2,869 62.1 26.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0tzcoorx",
   "source_url": "https://omlx.ai/benchmarks/performance/0tzcoorx",
   "tg_tokens": null,
   "tps": 62.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0tzcoorx-ctx8192",
   "model": "Nail-Qwen3.6-35B-A3B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: ac7cntkp, m5z42eaj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2933.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX-5bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 2,933 59.8 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0tzcoorx",
   "source_url": "https://omlx.ai/benchmarks/performance/0tzcoorx",
   "tg_tokens": null,
   "tps": 59.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-puxwkuuq-ctx1024",
   "model": "Nail-Qwen3.6-35B-A3B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: r5wzovaq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2111.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX-5bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 2,111 66.5 25.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/puxwkuuq",
   "source_url": "https://omlx.ai/benchmarks/performance/puxwkuuq",
   "tg_tokens": null,
   "tps": 66.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-puxwkuuq-ctx4096",
   "model": "Nail-Qwen3.6-35B-A3B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: r5wzovaq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2962.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX-5bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 2,962 65.0 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/puxwkuuq",
   "source_url": "https://omlx.ai/benchmarks/performance/puxwkuuq",
   "tg_tokens": null,
   "tps": 65.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-puxwkuuq-ctx8192",
   "model": "Nail-Qwen3.6-35B-A3B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: r5wzovaq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2972.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX-5bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 2,972 70.3 26.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/puxwkuuq",
   "source_url": "https://omlx.ai/benchmarks/performance/puxwkuuq",
   "tg_tokens": null,
   "tps": 70.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-8bb39uxl-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: 6qcu7ed1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1400.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,400 198.4 24.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8bb39uxl",
   "source_url": "https://omlx.ai/benchmarks/performance/8bb39uxl",
   "tg_tokens": null,
   "tps": 198.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-8bb39uxl-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.7 GB at this context; duplicate traces of the same session: 6qcu7ed1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1017.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 1,017 105.6 25.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8bb39uxl",
   "source_url": "https://omlx.ai/benchmarks/performance/8bb39uxl",
   "tg_tokens": null,
   "tps": 105.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gnbdabd7-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: cseppiru, f71b6rai, mujg0bjo, repiq718; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2720.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 1k 2,720 191.5 29.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gnbdabd7",
   "source_url": "https://omlx.ai/benchmarks/performance/gnbdabd7",
   "tg_tokens": null,
   "tps": 191.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gnbdabd7-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.6 GB at this context; duplicate traces of the same session: cseppiru, f71b6rai, mujg0bjo, repiq718; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3327.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 4k 3,327 148.9 31.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gnbdabd7",
   "source_url": "https://omlx.ai/benchmarks/performance/gnbdabd7",
   "tg_tokens": null,
   "tps": 148.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gnbdabd7-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.9 GB at this context; duplicate traces of the same session: cseppiru, f71b6rai, mujg0bjo, repiq718; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3131.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 8k 3,131 129.4 31.9 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gnbdabd7",
   "source_url": "https://omlx.ai/benchmarks/performance/gnbdabd7",
   "tg_tokens": null,
   "tps": 129.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gnbdabd7-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.3 GB at this context; duplicate traces of the same session: cseppiru, f71b6rai, mujg0bjo, repiq718; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2824.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 16k 2,824 114.5 32.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gnbdabd7",
   "source_url": "https://omlx.ai/benchmarks/performance/gnbdabd7",
   "tg_tokens": null,
   "tps": 114.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gnbdabd7-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.1 GB at this context; duplicate traces of the same session: cseppiru, f71b6rai, mujg0bjo, repiq718; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2337.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 32k 2,337 231.6 33.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gnbdabd7",
   "source_url": "https://omlx.ai/benchmarks/performance/gnbdabd7",
   "tg_tokens": null,
   "tps": 231.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ysiuoxem-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: oyxaxd29; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1467.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 1k 1,467 94.3 29.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ysiuoxem",
   "source_url": "https://omlx.ai/benchmarks/performance/ysiuoxem",
   "tg_tokens": null,
   "tps": 94.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ysiuoxem-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.6 GB at this context; duplicate traces of the same session: oyxaxd29; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3235.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 4k 3,235 139.3 31.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ysiuoxem",
   "source_url": "https://omlx.ai/benchmarks/performance/ysiuoxem",
   "tg_tokens": null,
   "tps": 139.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-i0htevt3-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: 4qja5xfi; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2725.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 1k 2,725 184.2 29.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i0htevt3",
   "source_url": "https://omlx.ai/benchmarks/performance/i0htevt3",
   "tg_tokens": null,
   "tps": 184.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-i0htevt3-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.6 GB at this context; duplicate traces of the same session: 4qja5xfi; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3327.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 4k 3,327 137.0 31.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i0htevt3",
   "source_url": "https://omlx.ai/benchmarks/performance/i0htevt3",
   "tg_tokens": null,
   "tps": 137.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-phmxknkr-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: hq3yw4mc; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 298.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 298.8 9.9 15.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/phmxknkr",
   "source_url": "https://omlx.ai/benchmarks/performance/phmxknkr",
   "tg_tokens": null,
   "tps": 9.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-phmxknkr-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: hq3yw4mc; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 314.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 314.1 9.8 17.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/phmxknkr",
   "source_url": "https://omlx.ai/benchmarks/performance/phmxknkr",
   "tg_tokens": null,
   "tps": 9.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-p0okf52q-ctx4096",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 23.4 GB; TTFT 9654 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 424.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Alis-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k tokens 424.3 PP tok/s 49.6 TG tok/s 9654 TTFT (ms) 23.4 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p0okf52q",
   "source_url": "https://omlx.ai/benchmarks/performance/p0okf52q",
   "tg_tokens": null,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-0uvigaaa-ctx4096",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 23.4 GB; TTFT 9897 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 413.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Alis-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k tokens 413.9 PP tok/s 43.5 TG tok/s 9897 TTFT (ms) 23.4 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0uvigaaa",
   "source_url": "https://omlx.ai/benchmarks/performance/0uvigaaa",
   "tg_tokens": null,
   "tps": 43.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-7fxdfgds-ctx4096",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 23.4 GB; TTFT 9789 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 418.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Alis-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k tokens 418.4 PP tok/s 50.3 TG tok/s 9789 TTFT (ms) 23.4 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7fxdfgds",
   "source_url": "https://omlx.ai/benchmarks/performance/7fxdfgds",
   "tg_tokens": null,
   "tps": 50.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-sukt6pnw-ctx4096",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 23.4 GB; TTFT 10154 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 403.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Alis-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k tokens 403.4 PP tok/s 45.1 TG tok/s 10154 TTFT (ms) 23.4 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sukt6pnw",
   "source_url": "https://omlx.ai/benchmarks/performance/sukt6pnw",
   "tg_tokens": null,
   "tps": 45.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zv4t0rqb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: 3pxownxc, vb2p1p01; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 717.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 717.6 59.6 17.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zv4t0rqb",
   "source_url": "https://omlx.ai/benchmarks/performance/zv4t0rqb",
   "tg_tokens": null,
   "tps": 59.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zv4t0rqb-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: 3pxownxc, vb2p1p01; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 747.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 747.1 69.5 19.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zv4t0rqb",
   "source_url": "https://omlx.ai/benchmarks/performance/zv4t0rqb",
   "tg_tokens": null,
   "tps": 69.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zv4t0rqb-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.3 GB at this context; duplicate traces of the same session: 3pxownxc, vb2p1p01; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 662.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 662.2 53.7 27.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zv4t0rqb",
   "source_url": "https://omlx.ai/benchmarks/performance/zv4t0rqb",
   "tg_tokens": null,
   "tps": 53.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-x1libweb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: smnym4gn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 904.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 904.3 42.2 17.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x1libweb",
   "source_url": "https://omlx.ai/benchmarks/performance/x1libweb",
   "tg_tokens": null,
   "tps": 42.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-x1libweb-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: smnym4gn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 962.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 962.8 39.9 18.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x1libweb",
   "source_url": "https://omlx.ai/benchmarks/performance/x1libweb",
   "tg_tokens": null,
   "tps": 39.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-y1zwbvhz-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.5 GB at this context; duplicate traces of the same session: 60qo5n6n, 6wbnogwb, 6y2exp2z, bldodnqb, m28ze1g7, qfw3yptf, xaubtf1g; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 386.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 386.8 33.5 17.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y1zwbvhz",
   "source_url": "https://omlx.ai/benchmarks/performance/y1zwbvhz",
   "tg_tokens": null,
   "tps": 33.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-y1zwbvhz-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 60qo5n6n, 6wbnogwb, 6y2exp2z, bldodnqb, m28ze1g7, qfw3yptf, xaubtf1g; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 403.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 403.1 37.8 18.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y1zwbvhz",
   "source_url": "https://omlx.ai/benchmarks/performance/y1zwbvhz",
   "tg_tokens": null,
   "tps": 37.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-y1zwbvhz-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 60qo5n6n, 6wbnogwb, 6y2exp2z, bldodnqb, m28ze1g7, qfw3yptf, xaubtf1g; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 399.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 399.2 41.2 19.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y1zwbvhz",
   "source_url": "https://omlx.ai/benchmarks/performance/y1zwbvhz",
   "tg_tokens": null,
   "tps": 41.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-y1zwbvhz-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 60qo5n6n, 6wbnogwb, 6y2exp2z, bldodnqb, m28ze1g7, qfw3yptf, xaubtf1g; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 379.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 379.2 41.4 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y1zwbvhz",
   "source_url": "https://omlx.ai/benchmarks/performance/y1zwbvhz",
   "tg_tokens": null,
   "tps": 41.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-y1zwbvhz-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 60qo5n6n, 6wbnogwb, 6y2exp2z, bldodnqb, m28ze1g7, qfw3yptf, xaubtf1g; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 351.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 351.6 38.3 20.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y1zwbvhz",
   "source_url": "https://omlx.ai/benchmarks/performance/y1zwbvhz",
   "tg_tokens": null,
   "tps": 38.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-y1zwbvhz-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: 60qo5n6n, 6wbnogwb, 6y2exp2z, bldodnqb, m28ze1g7, qfw3yptf, xaubtf1g; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 306.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 64k 306.0 29.7 23.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y1zwbvhz",
   "source_url": "https://omlx.ai/benchmarks/performance/y1zwbvhz",
   "tg_tokens": null,
   "tps": 29.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-y1zwbvhz-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: 60qo5n6n, 6wbnogwb, 6y2exp2z, bldodnqb, m28ze1g7, qfw3yptf, xaubtf1g; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 242.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 128k 242.6 24.8 27.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y1zwbvhz",
   "source_url": "https://omlx.ai/benchmarks/performance/y1zwbvhz",
   "tg_tokens": null,
   "tps": 24.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-y1zwbvhz-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.1 GB at this context; duplicate traces of the same session: 60qo5n6n, 6wbnogwb, 6y2exp2z, bldodnqb, m28ze1g7, qfw3yptf, xaubtf1g; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 200.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 195k 200.0 18.3 33.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y1zwbvhz",
   "source_url": "https://omlx.ai/benchmarks/performance/y1zwbvhz",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-g99n6a95-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 39.7 GB; TTFT 1334 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 767.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k tokens 767.9 PP tok/s 42.6 TG tok/s 1334 TTFT (ms) 39.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g99n6a95",
   "source_url": "https://omlx.ai/benchmarks/performance/g99n6a95",
   "tg_tokens": null,
   "tps": 42.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9i7fp60q-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: 6lt0g3cq, kxpryjpv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1156.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,156 90.3 100.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9i7fp60q",
   "source_url": "https://omlx.ai/benchmarks/performance/9i7fp60q",
   "tg_tokens": null,
   "tps": 90.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9i7fp60q-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: 6lt0g3cq, kxpryjpv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1409.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,409 95.6 101.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9i7fp60q",
   "source_url": "https://omlx.ai/benchmarks/performance/9i7fp60q",
   "tg_tokens": null,
   "tps": 95.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9i7fp60q-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: 6lt0g3cq, kxpryjpv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1394.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 1,394 102.2 101.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9i7fp60q",
   "source_url": "https://omlx.ai/benchmarks/performance/9i7fp60q",
   "tg_tokens": null,
   "tps": 102.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-c4gfiq5o-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: vh2vw14r; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1116.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,116 74.8 100.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c4gfiq5o",
   "source_url": "https://omlx.ai/benchmarks/performance/c4gfiq5o",
   "tg_tokens": null,
   "tps": 74.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-c4gfiq5o-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: vh2vw14r; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1406.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,406 58.8 101.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c4gfiq5o",
   "source_url": "https://omlx.ai/benchmarks/performance/c4gfiq5o",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-lbhm0wpm-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 8.9 GB at this context; duplicate traces of the same session: ax7f4c2p, dcd6y8vf, dzdy1e8z, k0fvy6c0; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 86.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 86.2 5.1 8.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lbhm0wpm",
   "source_url": "https://omlx.ai/benchmarks/performance/lbhm0wpm",
   "tg_tokens": null,
   "tps": 5.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-lbhm0wpm-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 9.4 GB at this context; duplicate traces of the same session: ax7f4c2p, dcd6y8vf, dzdy1e8z, k0fvy6c0; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 149.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 149.8 4.9 9.4 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lbhm0wpm",
   "source_url": "https://omlx.ai/benchmarks/performance/lbhm0wpm",
   "tg_tokens": null,
   "tps": 4.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-lbhm0wpm-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 9.4 GB at this context; duplicate traces of the same session: ax7f4c2p, dcd6y8vf, dzdy1e8z, k0fvy6c0; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 135.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 8k 135.3 4.8 9.4 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lbhm0wpm",
   "source_url": "https://omlx.ai/benchmarks/performance/lbhm0wpm",
   "tg_tokens": null,
   "tps": 4.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-lbhm0wpm-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 10.2 GB at this context; duplicate traces of the same session: ax7f4c2p, dcd6y8vf, dzdy1e8z, k0fvy6c0; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 32k 127.1 4.5 10.2 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lbhm0wpm",
   "source_url": "https://omlx.ai/benchmarks/performance/lbhm0wpm",
   "tg_tokens": null,
   "tps": 4.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-lbhm0wpm-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 11.6 GB at this context; duplicate traces of the same session: ax7f4c2p, dcd6y8vf, dzdy1e8z, k0fvy6c0; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 122.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 64k 122.6 4.2 11.6 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lbhm0wpm",
   "source_url": "https://omlx.ai/benchmarks/performance/lbhm0wpm",
   "tg_tokens": null,
   "tps": 4.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-lbhm0wpm-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 13.9 GB at this context; duplicate traces of the same session: ax7f4c2p, dcd6y8vf, dzdy1e8z, k0fvy6c0; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 113.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 128k 113.2 4.1 13.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lbhm0wpm",
   "source_url": "https://omlx.ai/benchmarks/performance/lbhm0wpm",
   "tg_tokens": null,
   "tps": 4.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-lbhm0wpm-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: ax7f4c2p, dcd6y8vf, dzdy1e8z, k0fvy6c0; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 195k 106.1 4.1 16.1 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lbhm0wpm",
   "source_url": "https://omlx.ai/benchmarks/performance/lbhm0wpm",
   "tg_tokens": null,
   "tps": 4.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-chgoz0yh-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 86.5 GB at this context; duplicate traces of the same session: 0bv5am0u, 65nkjb1n, g2d4mbn0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 832.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 832.0 52.8 86.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/chgoz0yh",
   "source_url": "https://omlx.ai/benchmarks/performance/chgoz0yh",
   "tg_tokens": null,
   "tps": 52.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-chgoz0yh-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 87.8 GB at this context; duplicate traces of the same session: 0bv5am0u, 65nkjb1n, g2d4mbn0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2088.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 2,088 65.7 87.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/chgoz0yh",
   "source_url": "https://omlx.ai/benchmarks/performance/chgoz0yh",
   "tg_tokens": null,
   "tps": 65.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-chgoz0yh-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 89.5 GB at this context; duplicate traces of the same session: 0bv5am0u, 65nkjb1n, g2d4mbn0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2086.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 2,086 51.1 89.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/chgoz0yh",
   "source_url": "https://omlx.ai/benchmarks/performance/chgoz0yh",
   "tg_tokens": null,
   "tps": 51.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-chgoz0yh-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 90.7 GB at this context; duplicate traces of the same session: 0bv5am0u, 65nkjb1n, g2d4mbn0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2062.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 2,062 61.4 90.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/chgoz0yh",
   "source_url": "https://omlx.ai/benchmarks/performance/chgoz0yh",
   "tg_tokens": null,
   "tps": 61.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-3ta9qobb-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: ni6f3jg1; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 266.5,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "gemma-4-26B-A4B-it-oQ2.5e-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 16 GB \u00b7 2bit; 1k 266.5 22.1 6.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ta9qobb",
   "source_url": "https://omlx.ai/benchmarks/performance/3ta9qobb",
   "tg_tokens": null,
   "tps": 22.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-3ta9qobb-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: ni6f3jg1; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 280.4,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "gemma-4-26B-A4B-it-oQ2.5e-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 16 GB \u00b7 2bit; 4k 280.4 21.5 6.5 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ta9qobb",
   "source_url": "https://omlx.ai/benchmarks/performance/3ta9qobb",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-64qgp9r0-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: u9t0u2au; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 613.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 16 GB \u00b7 4bit; 1k 613.4 92.7 5.9 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/64qgp9r0",
   "source_url": "https://omlx.ai/benchmarks/performance/64qgp9r0",
   "tg_tokens": null,
   "tps": 92.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-64qgp9r0-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 6.0 GB at this context; duplicate traces of the same session: u9t0u2au; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 627.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M2 Pro (16c); M2 Pro (16c) \u00b7 16 GB \u00b7 4bit; 4k 627.7 42.8 6.0 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/64qgp9r0",
   "source_url": "https://omlx.ai/benchmarks/performance/64qgp9r0",
   "tg_tokens": null,
   "tps": 42.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-l9lr5w5p-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "oMLX community benchmark trace; peak memory 288.9 GB at this context; duplicate traces of the same session: 6nlv8atg; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 269.2,
   "provenance": "community",
   "quant": "unknown",
   "quote": "DeepSeek-V4.1-Flash on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 269.2 26.5 288.9 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l9lr5w5p",
   "source_url": "https://omlx.ai/benchmarks/performance/l9lr5w5p",
   "tg_tokens": null,
   "tps": 26.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-l9lr5w5p-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "oMLX community benchmark trace; peak memory 289.4 GB at this context; duplicate traces of the same session: 6nlv8atg; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 402.9,
   "provenance": "community",
   "quant": "unknown",
   "quote": "DeepSeek-V4.1-Flash on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 402.9 32.2 289.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l9lr5w5p",
   "source_url": "https://omlx.ai/benchmarks/performance/l9lr5w5p",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-3wmmh12j-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 403.3 GB at this context; duplicate traces of the same session: 51aq7bnz; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 558.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 558.7 29.0 403.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3wmmh12j",
   "source_url": "https://omlx.ai/benchmarks/performance/3wmmh12j",
   "tg_tokens": null,
   "tps": 29.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-3wmmh12j-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 403.7 GB at this context; duplicate traces of the same session: 51aq7bnz; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 718.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 718.7 32.1 403.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3wmmh12j",
   "source_url": "https://omlx.ai/benchmarks/performance/3wmmh12j",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-vvd1a6ot-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.6 GB at this context; duplicate traces of the same session: 81k521ks; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 454.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 454.6 34.8 177.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vvd1a6ot",
   "source_url": "https://omlx.ai/benchmarks/performance/vvd1a6ot",
   "tg_tokens": null,
   "tps": 34.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-vvd1a6ot-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.8 GB at this context; duplicate traces of the same session: 81k521ks; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 552.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 552.2 40.9 178.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vvd1a6ot",
   "source_url": "https://omlx.ai/benchmarks/performance/vvd1a6ot",
   "tg_tokens": null,
   "tps": 40.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-4v5aav9e-ctx65536",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 64k context; peak memory 26.0 GB; TTFT 45911 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1428.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 8bit; 64k tokens 1,428 PP tok/s 55.5 TG tok/s 45911 TTFT (ms) 26.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4v5aav9e",
   "source_url": "https://omlx.ai/benchmarks/performance/4v5aav9e",
   "tg_tokens": null,
   "tps": 55.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-klzwm0ku-ctx131072",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 128k context; peak memory 27.2 GB; TTFT 118160 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1109.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 8bit; 128k tokens 1,109 PP tok/s 45.2 TG tok/s 118160 TTFT (ms) 27.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/klzwm0ku",
   "source_url": "https://omlx.ai/benchmarks/performance/klzwm0ku",
   "tg_tokens": null,
   "tps": 45.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-mspeodxm-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: 2cyg8n6c; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1157.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 8bit; 1k 1,157 72.8 24.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mspeodxm",
   "source_url": "https://omlx.ai/benchmarks/performance/mspeodxm",
   "tg_tokens": null,
   "tps": 72.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-mspeodxm-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: 2cyg8n6c; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1841.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 8bit; 4k 1,841 70.0 24.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mspeodxm",
   "source_url": "https://omlx.ai/benchmarks/performance/mspeodxm",
   "tg_tokens": null,
   "tps": 70.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-iryz6z4s-ctx1024",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: ynikmnja; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 606.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-mlx-mxfp4 on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 606.0 64.2 6.5 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iryz6z4s",
   "source_url": "https://omlx.ai/benchmarks/performance/iryz6z4s",
   "tg_tokens": null,
   "tps": 64.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-iryz6z4s-ctx4096",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 6.6 GB at this context; duplicate traces of the same session: ynikmnja; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 557.6,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-mlx-mxfp4 on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 557.6 59.0 6.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iryz6z4s",
   "source_url": "https://omlx.ai/benchmarks/performance/iryz6z4s",
   "tg_tokens": null,
   "tps": 59.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-nhtgdnru-ctx8192",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "MLX serve build; oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: oo0m8jgn; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 340.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM-V-4.7-35B-A3B-MLX-Serve-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 8k 340.5 47.6 20.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nhtgdnru",
   "source_url": "https://omlx.ai/benchmarks/performance/nhtgdnru",
   "tg_tokens": null,
   "tps": 47.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-nhtgdnru-ctx16384",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "MLX serve build; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: oo0m8jgn; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 298.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM-V-4.7-35B-A3B-MLX-Serve-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k 298.0 43.9 20.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nhtgdnru",
   "source_url": "https://omlx.ai/benchmarks/performance/nhtgdnru",
   "tg_tokens": null,
   "tps": 43.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qjyfo0pn-ctx1024",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: dcu0rooh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1465.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3___5-Lightning-30B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,465 175.5 18.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qjyfo0pn",
   "source_url": "https://omlx.ai/benchmarks/performance/qjyfo0pn",
   "tg_tokens": null,
   "tps": 175.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qjyfo0pn-ctx4096",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: dcu0rooh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2458.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3___5-Lightning-30B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,458 170.5 20.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qjyfo0pn",
   "source_url": "https://omlx.ai/benchmarks/performance/qjyfo0pn",
   "tg_tokens": null,
   "tps": 170.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-agt7vsvv-ctx1024",
   "model": "Nail-Qwen3.6-35B-A3B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2111.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX-5bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 2,111 66.5 25.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/agt7vsvv",
   "source_url": "https://omlx.ai/benchmarks/performance/agt7vsvv",
   "tg_tokens": null,
   "tps": 66.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-agt7vsvv-ctx4096",
   "model": "Nail-Qwen3.6-35B-A3B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2962.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX-5bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 2,962 65.0 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/agt7vsvv",
   "source_url": "https://omlx.ai/benchmarks/performance/agt7vsvv",
   "tg_tokens": null,
   "tps": 65.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-agt7vsvv-ctx8192",
   "model": "Nail-Qwen3.6-35B-A3B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2972.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX-5bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 2,972 70.3 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/agt7vsvv",
   "source_url": "https://omlx.ai/benchmarks/performance/agt7vsvv",
   "tg_tokens": null,
   "tps": 70.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-r0hkvp3t-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.7 GB at this context; duplicate traces of the same session: elke8c2m; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 224.9 20.9 5.7 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r0hkvp3t",
   "source_url": "https://omlx.ai/benchmarks/performance/r0hkvp3t",
   "tg_tokens": null,
   "tps": 20.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-r0hkvp3t-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: elke8c2m; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 233.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 233.2 20.1 6.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r0hkvp3t",
   "source_url": "https://omlx.ai/benchmarks/performance/r0hkvp3t",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-ipia16y5-ctx1024",
   "model": "Qwen3.5-2B",
   "notes": "oMLX community benchmark trace; peak memory 1.9 GB at this context; duplicate traces of the same session: 75z31brl; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 931.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-2B-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 931.6 83.5 1.9 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ipia16y5",
   "source_url": "https://omlx.ai/benchmarks/performance/ipia16y5",
   "tg_tokens": null,
   "tps": 83.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-ipia16y5-ctx4096",
   "model": "Qwen3.5-2B",
   "notes": "oMLX community benchmark trace; peak memory 2.1 GB at this context; duplicate traces of the same session: 75z31brl; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 985.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-2B-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 985.1 80.5 2.1 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ipia16y5",
   "source_url": "https://omlx.ai/benchmarks/performance/ipia16y5",
   "tg_tokens": null,
   "tps": 80.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-515i835h-ctx4096",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: p0ajg0dm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1384.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 1,384 84.3 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/515i835h",
   "source_url": "https://omlx.ai/benchmarks/performance/515i835h",
   "tg_tokens": null,
   "tps": 84.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-515i835h-ctx16384",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: p0ajg0dm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2776.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 2,776 85.7 20.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/515i835h",
   "source_url": "https://omlx.ai/benchmarks/performance/515i835h",
   "tg_tokens": null,
   "tps": 85.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-jtj31c7g-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: 6fac0xgw; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 199.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 199.6 20.4 5.8 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jtj31c7g",
   "source_url": "https://omlx.ai/benchmarks/performance/jtj31c7g",
   "tg_tokens": null,
   "tps": 20.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-jtj31c7g-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: 6fac0xgw; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 204.5 20.1 6.2 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jtj31c7g",
   "source_url": "https://omlx.ai/benchmarks/performance/jtj31c7g",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ntop81ya-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 57lhlr6z, 5ila2ztm, 8jznf7mb, e7wrvjks, p50szfdx, szys51tv, vpff4agj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2035.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 2,035 122.0 21.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ntop81ya",
   "source_url": "https://omlx.ai/benchmarks/performance/ntop81ya",
   "tg_tokens": null,
   "tps": 122.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ntop81ya-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 57lhlr6z, 5ila2ztm, 8jznf7mb, e7wrvjks, p50szfdx, szys51tv, vpff4agj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2794.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 2,794 98.0 22.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ntop81ya",
   "source_url": "https://omlx.ai/benchmarks/performance/ntop81ya",
   "tg_tokens": null,
   "tps": 98.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ntop81ya-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 57lhlr6z, 5ila2ztm, 8jznf7mb, e7wrvjks, p50szfdx, szys51tv, vpff4agj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2691.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 2,691 113.9 22.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ntop81ya",
   "source_url": "https://omlx.ai/benchmarks/performance/ntop81ya",
   "tg_tokens": null,
   "tps": 113.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ntop81ya-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 57lhlr6z, 5ila2ztm, 8jznf7mb, e7wrvjks, p50szfdx, szys51tv, vpff4agj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2448.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 2,448 110.0 22.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ntop81ya",
   "source_url": "https://omlx.ai/benchmarks/performance/ntop81ya",
   "tg_tokens": null,
   "tps": 110.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ntop81ya-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 57lhlr6z, 5ila2ztm, 8jznf7mb, e7wrvjks, p50szfdx, szys51tv, vpff4agj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2043.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 2,043 102.1 22.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ntop81ya",
   "source_url": "https://omlx.ai/benchmarks/performance/ntop81ya",
   "tg_tokens": null,
   "tps": 102.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ntop81ya-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 57lhlr6z, 5ila2ztm, 8jznf7mb, e7wrvjks, p50szfdx, szys51tv, vpff4agj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1520.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 64k 1,520 74.5 23.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ntop81ya",
   "source_url": "https://omlx.ai/benchmarks/performance/ntop81ya",
   "tg_tokens": null,
   "tps": 74.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ntop81ya-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: 57lhlr6z, 5ila2ztm, 8jznf7mb, e7wrvjks, p50szfdx, szys51tv, vpff4agj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1005.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 128k 1,005 55.8 24.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ntop81ya",
   "source_url": "https://omlx.ai/benchmarks/performance/ntop81ya",
   "tg_tokens": null,
   "tps": 55.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ntop81ya-ctx199680",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: 57lhlr6z, 5ila2ztm, 8jznf7mb, e7wrvjks, p50szfdx, szys51tv, vpff4agj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 737.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 195k 737.2 35.8 26.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ntop81ya",
   "source_url": "https://omlx.ai/benchmarks/performance/ntop81ya",
   "tg_tokens": null,
   "tps": 35.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-6bg33djk-ctx1024",
   "model": "Qwen3.8-27B-MLX-AD-4.50bpw-DWQ-16.0GB",
   "notes": "DWQ build; oMLX community benchmark trace, single run at 1k context; peak memory 15.8 GB; TTFT 3544 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 288.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-AD-4.50bpw-DWQ-16.0GB on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 1k tokens 288.9 PP tok/s 9.3 TG tok/s 3544 TTFT (ms) 15.8 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6bg33djk",
   "source_url": "https://omlx.ai/benchmarks/performance/6bg33djk",
   "tg_tokens": null,
   "tps": 9.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5lupsll5-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 4cwt6q58; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 326.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 326.7 104.7 18.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5lupsll5",
   "source_url": "https://omlx.ai/benchmarks/performance/5lupsll5",
   "tg_tokens": null,
   "tps": 104.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5lupsll5-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 4cwt6q58; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 347.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 347.7 97.6 21.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5lupsll5",
   "source_url": "https://omlx.ai/benchmarks/performance/5lupsll5",
   "tg_tokens": null,
   "tps": 97.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5lupsll5-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: 4cwt6q58; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 341.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 341.9 86.0 22.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5lupsll5",
   "source_url": "https://omlx.ai/benchmarks/performance/5lupsll5",
   "tg_tokens": null,
   "tps": 86.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5lupsll5-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 4cwt6q58; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 328.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 328.8 47.2 23.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5lupsll5",
   "source_url": "https://omlx.ai/benchmarks/performance/5lupsll5",
   "tg_tokens": null,
   "tps": 47.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ckm0y4v9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 319.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 319.9 41.5 18.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ckm0y4v9",
   "source_url": "https://omlx.ai/benchmarks/performance/ckm0y4v9",
   "tg_tokens": null,
   "tps": 41.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ckm0y4v9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 347.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 347.9 52.4 21.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ckm0y4v9",
   "source_url": "https://omlx.ai/benchmarks/performance/ckm0y4v9",
   "tg_tokens": null,
   "tps": 52.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ckm0y4v9-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 342.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 342.0 86.3 22.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ckm0y4v9",
   "source_url": "https://omlx.ai/benchmarks/performance/ckm0y4v9",
   "tg_tokens": null,
   "tps": 86.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ckm0y4v9-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 328.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 328.6 48.2 23.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ckm0y4v9",
   "source_url": "https://omlx.ai/benchmarks/performance/ckm0y4v9",
   "tg_tokens": null,
   "tps": 48.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-n2q7814r-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: oj8u7oxk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 466.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 466.4 30.9 18.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n2q7814r",
   "source_url": "https://omlx.ai/benchmarks/performance/n2q7814r",
   "tg_tokens": null,
   "tps": 30.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-n2q7814r-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: oj8u7oxk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 415.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 415.8 41.3 19.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n2q7814r",
   "source_url": "https://omlx.ai/benchmarks/performance/n2q7814r",
   "tg_tokens": null,
   "tps": 41.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-n6j7tob8-ctx1024",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: sson4qy0, tydda0z9; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 792.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 792.1 39.9 7.0 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n6j7tob8",
   "source_url": "https://omlx.ai/benchmarks/performance/n6j7tob8",
   "tg_tokens": null,
   "tps": 39.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-n6j7tob8-ctx4096",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace; peak memory 7.3 GB at this context; duplicate traces of the same session: sson4qy0, tydda0z9; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 968.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 968.9 51.0 7.3 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n6j7tob8",
   "source_url": "https://omlx.ai/benchmarks/performance/n6j7tob8",
   "tg_tokens": null,
   "tps": 51.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-n6j7tob8-ctx16384",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace; peak memory 7.7 GB at this context; duplicate traces of the same session: sson4qy0, tydda0z9; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 931.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 16k 931.1 26.8 7.7 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n6j7tob8",
   "source_url": "https://omlx.ai/benchmarks/performance/n6j7tob8",
   "tg_tokens": null,
   "tps": 26.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2rdeyo6q-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 91.3 GB; TTFT 3019 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 339.2,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 1k tokens 339.2 PP tok/s 53.9 TG tok/s 3019 TTFT (ms) 91.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2rdeyo6q",
   "source_url": "https://omlx.ai/benchmarks/performance/2rdeyo6q",
   "tg_tokens": null,
   "tps": 53.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-pstvgwky-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 91.3 GB at this context; duplicate traces of the same session: klb5wk5k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1730.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 1k 1,730 75.1 91.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pstvgwky",
   "source_url": "https://omlx.ai/benchmarks/performance/pstvgwky",
   "tg_tokens": null,
   "tps": 75.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pstvgwky-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 91.7 GB at this context; duplicate traces of the same session: klb5wk5k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2586.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 3bit; 4k 2,586 69.0 91.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pstvgwky",
   "source_url": "https://omlx.ai/benchmarks/performance/pstvgwky",
   "tg_tokens": null,
   "tps": 69.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-zjijne9x-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 99.7 GB; TTFT 8512 ms; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 481.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 4k tokens 481.2 PP tok/s 40.1 TG tok/s 8512 TTFT (ms) 99.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zjijne9x",
   "source_url": "https://omlx.ai/benchmarks/performance/zjijne9x",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-b5dxzeiv-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 69.9 GB; TTFT 22381 ms; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 183.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 4k tokens 183.0 PP tok/s 32.0 TG tok/s 22381 TTFT (ms) 69.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b5dxzeiv",
   "source_url": "https://omlx.ai/benchmarks/performance/b5dxzeiv",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-imybwgxe-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 99.7 GB; TTFT 8507 ms; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 481.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 4k tokens 481.5 PP tok/s 40.1 TG tok/s 8507 TTFT (ms) 99.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/imybwgxe",
   "source_url": "https://omlx.ai/benchmarks/performance/imybwgxe",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5u1ompwi-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: lhusr5u8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1106.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,106 69.2 100.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5u1ompwi",
   "source_url": "https://omlx.ai/benchmarks/performance/5u1ompwi",
   "tg_tokens": null,
   "tps": 69.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5u1ompwi-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: lhusr5u8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1374.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,374 80.7 101.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5u1ompwi",
   "source_url": "https://omlx.ai/benchmarks/performance/5u1ompwi",
   "tg_tokens": null,
   "tps": 80.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-fywxqp0j-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 129.4 GB at this context; duplicate traces of the same session: u8mipl8n; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1422.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 1k 1,422 72.7 129.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fywxqp0j",
   "source_url": "https://omlx.ai/benchmarks/performance/fywxqp0j",
   "tg_tokens": null,
   "tps": 72.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-fywxqp0j-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 129.6 GB at this context; duplicate traces of the same session: u8mipl8n; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2465.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 4k 2,465 61.8 129.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fywxqp0j",
   "source_url": "https://omlx.ai/benchmarks/performance/fywxqp0j",
   "tg_tokens": null,
   "tps": 61.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-e9h0xfkd-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 129.4 GB at this context; duplicate traces of the same session: c3ulkmrc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1241.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 1k 1,241 95.1 129.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e9h0xfkd",
   "source_url": "https://omlx.ai/benchmarks/performance/e9h0xfkd",
   "tg_tokens": null,
   "tps": 95.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-e9h0xfkd-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 129.6 GB at this context; duplicate traces of the same session: c3ulkmrc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2632.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 4k 2,632 81.9 129.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e9h0xfkd",
   "source_url": "https://omlx.ai/benchmarks/performance/e9h0xfkd",
   "tg_tokens": null,
   "tps": 81.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-1jr5mbxz-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 129.4 GB at this context; duplicate traces of the same session: enitvi4b; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1399.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 1k 1,399 34.8 129.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1jr5mbxz",
   "source_url": "https://omlx.ai/benchmarks/performance/1jr5mbxz",
   "tg_tokens": null,
   "tps": 34.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-1jr5mbxz-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 129.6 GB at this context; duplicate traces of the same session: enitvi4b; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2038.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 4k 2,038 44.7 129.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1jr5mbxz",
   "source_url": "https://omlx.ai/benchmarks/performance/1jr5mbxz",
   "tg_tokens": null,
   "tps": 44.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-w6iil4vi-ctx1024",
   "model": "Swift 1.5",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 6wnl6ld4, hqben5oh; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Swift-1.5-4bit-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 104.3 15.2 16.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w6iil4vi",
   "source_url": "https://omlx.ai/benchmarks/performance/w6iil4vi",
   "tg_tokens": null,
   "tps": 15.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-w6iil4vi-ctx4096",
   "model": "Swift 1.5",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 6wnl6ld4, hqben5oh; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 107.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Swift-1.5-4bit-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 107.8 14.5 17.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w6iil4vi",
   "source_url": "https://omlx.ai/benchmarks/performance/w6iil4vi",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-w6iil4vi-ctx8192",
   "model": "Swift 1.5",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 6wnl6ld4, hqben5oh; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Swift-1.5-4bit-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 106.6 14.3 18.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w6iil4vi",
   "source_url": "https://omlx.ai/benchmarks/performance/w6iil4vi",
   "tg_tokens": null,
   "tps": 14.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-tsk7wuj3-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace, single run at 4k context; peak memory 14.6 GB; TTFT 2945 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1391.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 1,391 PP tok/s 96.1 TG tok/s 2945 TTFT (ms) 14.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tsk7wuj3",
   "source_url": "https://omlx.ai/benchmarks/performance/tsk7wuj3",
   "tg_tokens": null,
   "tps": 96.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-2t05bvwb-ctx4096",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "MLX serve build; oMLX community benchmark trace, single run at 4k context; peak memory 20.1 GB; TTFT 11104 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 368.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM-V-4.7-35B-A3B-MLX-Serve-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k tokens 368.9 PP tok/s 49.8 TG tok/s 11104 TTFT (ms) 20.1 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2t05bvwb",
   "source_url": "https://omlx.ai/benchmarks/performance/2t05bvwb",
   "tg_tokens": null,
   "tps": 49.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-atpnduc8-ctx32768",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "MLX serve build; oMLX community benchmark trace, single run at 32k context; peak memory 21.4 GB; TTFT 140357 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 233.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM-V-4.7-35B-A3B-MLX-Serve-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 32k tokens 233.5 PP tok/s 40.6 TG tok/s 140357 TTFT (ms) 21.4 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/atpnduc8",
   "source_url": "https://omlx.ai/benchmarks/performance/atpnduc8",
   "tg_tokens": null,
   "tps": 40.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-s897mdnh-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: mz4zvesm, smx9i7bn; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 547.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 547.6 69.8 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s897mdnh",
   "source_url": "https://omlx.ai/benchmarks/performance/s897mdnh",
   "tg_tokens": null,
   "tps": 69.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-s897mdnh-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: mz4zvesm, smx9i7bn; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 590.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 590.2 68.9 21.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s897mdnh",
   "source_url": "https://omlx.ai/benchmarks/performance/s897mdnh",
   "tg_tokens": null,
   "tps": 68.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-s897mdnh-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: mz4zvesm, smx9i7bn; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 537.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 8k 537.6 62.6 21.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s897mdnh",
   "source_url": "https://omlx.ai/benchmarks/performance/s897mdnh",
   "tg_tokens": null,
   "tps": 62.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-9dpcu5o2-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: 1qzal7o3, eac1np2a; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2514.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 2,514 87.6 20.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9dpcu5o2",
   "source_url": "https://omlx.ai/benchmarks/performance/9dpcu5o2",
   "tg_tokens": null,
   "tps": 87.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-9dpcu5o2-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 1qzal7o3, eac1np2a; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2516.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 2,516 78.6 20.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9dpcu5o2",
   "source_url": "https://omlx.ai/benchmarks/performance/9dpcu5o2",
   "tg_tokens": null,
   "tps": 78.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-9dpcu5o2-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 1qzal7o3, eac1np2a; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1919.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 1,919 62.8 21.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9dpcu5o2",
   "source_url": "https://omlx.ai/benchmarks/performance/9dpcu5o2",
   "tg_tokens": null,
   "tps": 62.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-c09l8s7w-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 4amlco69, vhj1jec3, zflktupw; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 578.5,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 578.5 57.3 21.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c09l8s7w",
   "source_url": "https://omlx.ai/benchmarks/performance/c09l8s7w",
   "tg_tokens": null,
   "tps": 57.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-c09l8s7w-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: 4amlco69, vhj1jec3, zflktupw; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 604.5,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 604.5 54.7 23.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c09l8s7w",
   "source_url": "https://omlx.ai/benchmarks/performance/c09l8s7w",
   "tg_tokens": null,
   "tps": 54.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-c09l8s7w-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: 4amlco69, vhj1jec3, zflktupw; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 554.5,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 8k 554.5 52.4 24.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c09l8s7w",
   "source_url": "https://omlx.ai/benchmarks/performance/c09l8s7w",
   "tg_tokens": null,
   "tps": 52.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-c09l8s7w-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: 4amlco69, vhj1jec3, zflktupw; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 410.3,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Ornith-1.5-35B-A3B-OptiQ-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k 410.3 49.8 25.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c09l8s7w",
   "source_url": "https://omlx.ai/benchmarks/performance/c09l8s7w",
   "tg_tokens": null,
   "tps": 49.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-z6uhqdwr-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: f2t1k0is, igjpqqt9; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 755.8,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 755.8 66.4 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z6uhqdwr",
   "source_url": "https://omlx.ai/benchmarks/performance/z6uhqdwr",
   "tg_tokens": null,
   "tps": 66.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-z6uhqdwr-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: f2t1k0is, igjpqqt9; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 891.4,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 891.4 65.9 22.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z6uhqdwr",
   "source_url": "https://omlx.ai/benchmarks/performance/z6uhqdwr",
   "tg_tokens": null,
   "tps": 65.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-z6uhqdwr-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: f2t1k0is, igjpqqt9; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 839.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 8k 839.0 63.6 22.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z6uhqdwr",
   "source_url": "https://omlx.ai/benchmarks/performance/z6uhqdwr",
   "tg_tokens": null,
   "tps": 63.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-zx446iqm-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: gh0ehikc; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 292.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 292.9 19.0 18.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zx446iqm",
   "source_url": "https://omlx.ai/benchmarks/performance/zx446iqm",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-zx446iqm-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: gh0ehikc; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 318.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 318.8 22.8 18.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zx446iqm",
   "source_url": "https://omlx.ai/benchmarks/performance/zx446iqm",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-6o0kc31z-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 68pycrz2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 426.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 426.9 15.2 17.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6o0kc31z",
   "source_url": "https://omlx.ai/benchmarks/performance/6o0kc31z",
   "tg_tokens": null,
   "tps": 15.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-6o0kc31z-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 68pycrz2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 363.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 363.9 13.9 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6o0kc31z",
   "source_url": "https://omlx.ai/benchmarks/performance/6o0kc31z",
   "tg_tokens": null,
   "tps": 13.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-s1czi5ko-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: ewgxww2j; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 372.3,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.8-27B-MLX-5bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 5bit; 1k 372.3 14.4 19.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s1czi5ko",
   "source_url": "https://omlx.ai/benchmarks/performance/s1czi5ko",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-s1czi5ko-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: ewgxww2j; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 390.9,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.8-27B-MLX-5bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 5bit; 4k 390.9 14.1 20.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s1czi5ko",
   "source_url": "https://omlx.ai/benchmarks/performance/s1czi5ko",
   "tg_tokens": null,
   "tps": 14.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1qwyhy7u-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 34.2 GB at this context; duplicate traces of the same session: 9k3z48x2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 857.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 857.1 35.7 34.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1qwyhy7u",
   "source_url": "https://omlx.ai/benchmarks/performance/1qwyhy7u",
   "tg_tokens": null,
   "tps": 35.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1qwyhy7u-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: 9k3z48x2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 874.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 874.6 28.5 35.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1qwyhy7u",
   "source_url": "https://omlx.ai/benchmarks/performance/1qwyhy7u",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-okw0eckg-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 34.2 GB at this context; duplicate traces of the same session: hfmgfftt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 851.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 851.0 33.7 34.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/okw0eckg",
   "source_url": "https://omlx.ai/benchmarks/performance/okw0eckg",
   "tg_tokens": null,
   "tps": 33.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-okw0eckg-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: hfmgfftt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 827.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 827.9 31.2 35.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/okw0eckg",
   "source_url": "https://omlx.ai/benchmarks/performance/okw0eckg",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-gnmmaire-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: qche3ngs; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 169.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 4k 169.1 29.6 19.7 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gnmmaire",
   "source_url": "https://omlx.ai/benchmarks/performance/gnmmaire",
   "tg_tokens": null,
   "tps": 29.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-gnmmaire-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: qche3ngs; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 154.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 16k 154.2 39.5 20.6 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gnmmaire",
   "source_url": "https://omlx.ai/benchmarks/performance/gnmmaire",
   "tg_tokens": null,
   "tps": 39.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-mu4021wz-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: tl9dtb7e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 220.9 48.4 18.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mu4021wz",
   "source_url": "https://omlx.ai/benchmarks/performance/mu4021wz",
   "tg_tokens": null,
   "tps": 48.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-mu4021wz-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: tl9dtb7e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 199.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 199.0 31.7 22.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mu4021wz",
   "source_url": "https://omlx.ai/benchmarks/performance/mu4021wz",
   "tg_tokens": null,
   "tps": 31.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gqvhz960-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: cpqr6u89, q0ae2ag9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 319.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 319.9 41.5 18.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gqvhz960",
   "source_url": "https://omlx.ai/benchmarks/performance/gqvhz960",
   "tg_tokens": null,
   "tps": 41.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gqvhz960-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: cpqr6u89, q0ae2ag9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 347.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 347.9 52.4 21.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gqvhz960",
   "source_url": "https://omlx.ai/benchmarks/performance/gqvhz960",
   "tg_tokens": null,
   "tps": 52.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gqvhz960-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: cpqr6u89, q0ae2ag9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 342.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 342.0 86.3 22.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gqvhz960",
   "source_url": "https://omlx.ai/benchmarks/performance/gqvhz960",
   "tg_tokens": null,
   "tps": 86.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gqvhz960-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: cpqr6u89, q0ae2ag9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 328.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 328.6 48.2 23.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gqvhz960",
   "source_url": "https://omlx.ai/benchmarks/performance/gqvhz960",
   "tg_tokens": null,
   "tps": 48.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3lwekt4v-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: qz8p6dmo; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 326.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 326.7 104.7 18.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3lwekt4v",
   "source_url": "https://omlx.ai/benchmarks/performance/3lwekt4v",
   "tg_tokens": null,
   "tps": 104.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3lwekt4v-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: qz8p6dmo; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 347.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 347.7 97.6 21.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3lwekt4v",
   "source_url": "https://omlx.ai/benchmarks/performance/3lwekt4v",
   "tg_tokens": null,
   "tps": 97.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3lwekt4v-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: qz8p6dmo; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 341.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 341.9 86.0 22.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3lwekt4v",
   "source_url": "https://omlx.ai/benchmarks/performance/3lwekt4v",
   "tg_tokens": null,
   "tps": 86.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-3lwekt4v-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: qz8p6dmo; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 328.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 328.8 47.2 23.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3lwekt4v",
   "source_url": "https://omlx.ai/benchmarks/performance/3lwekt4v",
   "tg_tokens": null,
   "tps": 47.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-g5fz32p4-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: kefpfhp7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 318.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 318.2 25.3 21.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g5fz32p4",
   "source_url": "https://omlx.ai/benchmarks/performance/g5fz32p4",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-g5fz32p4-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: kefpfhp7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 353.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 353.6 24.3 23.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g5fz32p4",
   "source_url": "https://omlx.ai/benchmarks/performance/g5fz32p4",
   "tg_tokens": null,
   "tps": 24.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-hk3g6xoj-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: kj31ntbm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 676.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 676.4 32.5 19.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hk3g6xoj",
   "source_url": "https://omlx.ai/benchmarks/performance/hk3g6xoj",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-hk3g6xoj-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: kj31ntbm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 692.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 692.5 41.3 20.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hk3g6xoj",
   "source_url": "https://omlx.ai/benchmarks/performance/hk3g6xoj",
   "tg_tokens": null,
   "tps": 41.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-ah7n9x0j-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 13.3 GB at this context; duplicate traces of the same session: xvebhco2; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 70.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 70.7 6.3 13.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ah7n9x0j",
   "source_url": "https://omlx.ai/benchmarks/performance/ah7n9x0j",
   "tg_tokens": null,
   "tps": 6.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ah7n9x0j-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: xvebhco2; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 103.5 4.7 15.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ah7n9x0j",
   "source_url": "https://omlx.ai/benchmarks/performance/ah7n9x0j",
   "tg_tokens": null,
   "tps": 4.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4po10rjf-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: 9hw58jqn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1239.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,239 84.1 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4po10rjf",
   "source_url": "https://omlx.ai/benchmarks/performance/4po10rjf",
   "tg_tokens": null,
   "tps": 84.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4po10rjf-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: 9hw58jqn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2111.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,111 73.9 71.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4po10rjf",
   "source_url": "https://omlx.ai/benchmarks/performance/4po10rjf",
   "tg_tokens": null,
   "tps": 73.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mq52tlf6-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1220.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,220 94.2 70.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mq52tlf6",
   "source_url": "https://omlx.ai/benchmarks/performance/mq52tlf6",
   "tg_tokens": null,
   "tps": 94.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mq52tlf6-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2062.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,062 64.0 71.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mq52tlf6",
   "source_url": "https://omlx.ai/benchmarks/performance/mq52tlf6",
   "tg_tokens": null,
   "tps": 64.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-r4804u41-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 126.5 GB at this context; duplicate traces of the same session: pjv69mba; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 473.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 1k 473.9 62.6 126.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r4804u41",
   "source_url": "https://omlx.ai/benchmarks/performance/r4804u41",
   "tg_tokens": null,
   "tps": 62.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-r4804u41-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 126.6 GB at this context; duplicate traces of the same session: pjv69mba; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1316.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 4k 1,316 50.0 126.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r4804u41",
   "source_url": "https://omlx.ai/benchmarks/performance/r4804u41",
   "tg_tokens": null,
   "tps": 50.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-dhqa3q7o-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 129.4 GB at this context; duplicate traces of the same session: ohyun8ez; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1406.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 1k 1,406 114.3 129.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dhqa3q7o",
   "source_url": "https://omlx.ai/benchmarks/performance/dhqa3q7o",
   "tg_tokens": null,
   "tps": 114.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-dhqa3q7o-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 129.6 GB at this context; duplicate traces of the same session: ohyun8ez; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2593.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 4k 2,593 97.9 129.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dhqa3q7o",
   "source_url": "https://omlx.ai/benchmarks/performance/dhqa3q7o",
   "tg_tokens": null,
   "tps": 97.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-gyt9ie1u-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 129.4 GB at this context; duplicate traces of the same session: maw00il5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1339.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 1k 1,339 119.9 129.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gyt9ie1u",
   "source_url": "https://omlx.ai/benchmarks/performance/gyt9ie1u",
   "tg_tokens": null,
   "tps": 119.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-gyt9ie1u-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 129.6 GB at this context; duplicate traces of the same session: maw00il5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2702.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 4k 2,702 104.4 129.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gyt9ie1u",
   "source_url": "https://omlx.ai/benchmarks/performance/gyt9ie1u",
   "tg_tokens": null,
   "tps": 104.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-w4a8w8oa-ctx1024",
   "model": "Spark-X2.5-4B",
   "notes": "OptiQ build; oMLX community benchmark trace; peak memory 3.6 GB at this context; duplicate traces of the same session: 13x8f2kw; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 391.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Spark-X2.5-4B-OptiQ-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 391.8 31.4 3.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w4a8w8oa",
   "source_url": "https://omlx.ai/benchmarks/performance/w4a8w8oa",
   "tg_tokens": null,
   "tps": 31.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-w4a8w8oa-ctx4096",
   "model": "Spark-X2.5-4B",
   "notes": "OptiQ build; oMLX community benchmark trace; peak memory 3.6 GB at this context; duplicate traces of the same session: 13x8f2kw; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 382.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Spark-X2.5-4B-OptiQ-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 382.5 30.1 3.6 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w4a8w8oa",
   "source_url": "https://omlx.ai/benchmarks/performance/w4a8w8oa",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-qd0xwtls-ctx1024",
   "model": "Spark-X2.5-4B",
   "notes": "OptiQ build; oMLX community benchmark trace; peak memory 3.6 GB at this context; duplicate traces of the same session: vxf9sl66; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 393.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Spark-X2.5-4B-OptiQ-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 393.8 31.3 3.6 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qd0xwtls",
   "source_url": "https://omlx.ai/benchmarks/performance/qd0xwtls",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-qd0xwtls-ctx4096",
   "model": "Spark-X2.5-4B",
   "notes": "OptiQ build; oMLX community benchmark trace; peak memory 3.6 GB at this context; duplicate traces of the same session: vxf9sl66; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 388.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Spark-X2.5-4B-OptiQ-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 388.1 30.1 3.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qd0xwtls",
   "source_url": "https://omlx.ai/benchmarks/performance/qd0xwtls",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-ircjz8sy-ctx1024",
   "model": "Ternary-Bonsai-2-4B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: nzxc2dsp; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 196.9,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-4B-mlx-2bit on M1 (8c); M1 (8c) \u00b7 8 GB \u00b7 2bit; 1k 196.9 27.3 3.9 GB current",
   "ram_gb": 8,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ircjz8sy",
   "source_url": "https://omlx.ai/benchmarks/performance/ircjz8sy",
   "tg_tokens": null,
   "tps": 27.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-ircjz8sy-ctx4096",
   "model": "Ternary-Bonsai-2-4B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 5.2 GB at this context; duplicate traces of the same session: nzxc2dsp; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 194.8,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-4B-mlx-2bit on M1 (8c); M1 (8c) \u00b7 8 GB \u00b7 2bit; 4k 194.8 25.4 5.2 GB view",
   "ram_gb": 8,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ircjz8sy",
   "source_url": "https://omlx.ai/benchmarks/performance/ircjz8sy",
   "tg_tokens": null,
   "tps": 25.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-l2qhb8of-ctx1024",
   "model": "Ternary-Bonsai-2-4B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: 462hxk3z; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 184.8,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-4B-mlx-2bit on M1 (8c); M1 (8c) \u00b7 8 GB \u00b7 2bit; 1k 184.8 27.9 3.9 GB current",
   "ram_gb": 8,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l2qhb8of",
   "source_url": "https://omlx.ai/benchmarks/performance/l2qhb8of",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-l2qhb8of-ctx4096",
   "model": "Ternary-Bonsai-2-4B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 5.2 GB at this context; duplicate traces of the same session: 462hxk3z; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 188.0,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-4B-mlx-2bit on M1 (8c); M1 (8c) \u00b7 8 GB \u00b7 2bit; 4k 188.0 25.0 5.2 GB view",
   "ram_gb": 8,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l2qhb8of",
   "source_url": "https://omlx.ai/benchmarks/performance/l2qhb8of",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-f7h5gv1k-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 4.5 GB at this context; duplicate traces of the same session: vc1yhjj2; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 654.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 654.4 34.9 4.5 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f7h5gv1k",
   "source_url": "https://omlx.ai/benchmarks/performance/f7h5gv1k",
   "tg_tokens": null,
   "tps": 34.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-f7h5gv1k-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 4.6 GB at this context; duplicate traces of the same session: vc1yhjj2; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 651.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 651.3 32.5 4.6 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f7h5gv1k",
   "source_url": "https://omlx.ai/benchmarks/performance/f7h5gv1k",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-klnjgm5b-ctx1024",
   "model": "Qwen2.5-1.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 1.5 GB at this context; duplicate traces of the same session: none; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 859.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen2.5-1.5B-Instruct-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 unknown; 1k 859.7 120.7 1.5 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/klnjgm5b",
   "source_url": "https://omlx.ai/benchmarks/performance/klnjgm5b",
   "tg_tokens": null,
   "tps": 120.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-klnjgm5b-ctx4096",
   "model": "Qwen2.5-1.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 1.6 GB at this context; duplicate traces of the same session: none; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 923.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen2.5-1.5B-Instruct-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 unknown; 4k 923.5 101.1 1.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/klnjgm5b",
   "source_url": "https://omlx.ai/benchmarks/performance/klnjgm5b",
   "tg_tokens": null,
   "tps": 101.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-yfuwuuzo-ctx4096",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.0 GB at this context; duplicate traces of the same session: 23jad3ph, vpdtpa2z; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 453.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 453.2 44.5 179.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yfuwuuzo",
   "source_url": "https://omlx.ai/benchmarks/performance/yfuwuuzo",
   "tg_tokens": null,
   "tps": 44.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-yfuwuuzo-ctx16384",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.7 GB at this context; duplicate traces of the same session: 23jad3ph, vpdtpa2z; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 413.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 413.0 89.6 179.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yfuwuuzo",
   "source_url": "https://omlx.ai/benchmarks/performance/yfuwuuzo",
   "tg_tokens": null,
   "tps": 89.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-yfuwuuzo-ctx65536",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 182.0 GB at this context; duplicate traces of the same session: 23jad3ph, vpdtpa2z; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 401.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 401.3 50.3 182.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yfuwuuzo",
   "source_url": "https://omlx.ai/benchmarks/performance/yfuwuuzo",
   "tg_tokens": null,
   "tps": 50.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-g60x222r-ctx16384",
   "model": "GLM-5.3",
   "notes": "MLX mixed 4/8bit build; oMLX community benchmark trace, single run at 16k context; peak memory 436.1 GB; TTFT 136553 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 120.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-MLX-mixed-4_8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k tokens 120.0 PP tok/s 12.9 TG tok/s 136553 TTFT (ms) 436.1 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g60x222r",
   "source_url": "https://omlx.ai/benchmarks/performance/g60x222r",
   "tg_tokens": null,
   "tps": 12.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-b2zq1f3l-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: 8378qnph, gfh0lkp8, yztbqj4r; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1236.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 1k 1,236 96.7 13.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b2zq1f3l",
   "source_url": "https://omlx.ai/benchmarks/performance/b2zq1f3l",
   "tg_tokens": null,
   "tps": 96.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-b2zq1f3l-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.6 GB at this context; duplicate traces of the same session: 8378qnph, gfh0lkp8, yztbqj4r; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1536.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 4k 1,536 94.5 13.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b2zq1f3l",
   "source_url": "https://omlx.ai/benchmarks/performance/b2zq1f3l",
   "tg_tokens": null,
   "tps": 94.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-b2zq1f3l-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: 8378qnph, gfh0lkp8, yztbqj4r; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1509.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 8k 1,509 93.1 13.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b2zq1f3l",
   "source_url": "https://omlx.ai/benchmarks/performance/b2zq1f3l",
   "tg_tokens": null,
   "tps": 93.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-b2zq1f3l-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.8 GB at this context; duplicate traces of the same session: 8378qnph, gfh0lkp8, yztbqj4r; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1437.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 16k 1,437 90.4 13.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b2zq1f3l",
   "source_url": "https://omlx.ai/benchmarks/performance/b2zq1f3l",
   "tg_tokens": null,
   "tps": 90.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-rs5vmcql-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: atxthizf, dhirztwt; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1368.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 1,368 143.9 3.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rs5vmcql",
   "source_url": "https://omlx.ai/benchmarks/performance/rs5vmcql",
   "tg_tokens": null,
   "tps": 143.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-rs5vmcql-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.8 GB at this context; duplicate traces of the same session: atxthizf, dhirztwt; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1278.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 1,278 137.3 4.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rs5vmcql",
   "source_url": "https://omlx.ai/benchmarks/performance/rs5vmcql",
   "tg_tokens": null,
   "tps": 137.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-rs5vmcql-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 5.1 GB at this context; duplicate traces of the same session: atxthizf, dhirztwt; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1224.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 8k 1,224 132.4 5.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rs5vmcql",
   "source_url": "https://omlx.ai/benchmarks/performance/rs5vmcql",
   "tg_tokens": null,
   "tps": 132.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-rre7nbsd-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: p040jpm0, vx3b1xfq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1358.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 1,358 143.8 3.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rre7nbsd",
   "source_url": "https://omlx.ai/benchmarks/performance/rre7nbsd",
   "tg_tokens": null,
   "tps": 143.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-rre7nbsd-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.8 GB at this context; duplicate traces of the same session: p040jpm0, vx3b1xfq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1285.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 1,285 137.5 4.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rre7nbsd",
   "source_url": "https://omlx.ai/benchmarks/performance/rre7nbsd",
   "tg_tokens": null,
   "tps": 137.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-rre7nbsd-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 5.1 GB at this context; duplicate traces of the same session: p040jpm0, vx3b1xfq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1220.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 8k 1,220 131.4 5.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rre7nbsd",
   "source_url": "https://omlx.ai/benchmarks/performance/rre7nbsd",
   "tg_tokens": null,
   "tps": 131.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-mfxpn3pu-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: f487syic, v9ygemth; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 771.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 771.2 98.9 5.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mfxpn3pu",
   "source_url": "https://omlx.ai/benchmarks/performance/mfxpn3pu",
   "tg_tokens": null,
   "tps": 98.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-mfxpn3pu-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.3 GB at this context; duplicate traces of the same session: f487syic, v9ygemth; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 725.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 725.2 95.5 7.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mfxpn3pu",
   "source_url": "https://omlx.ai/benchmarks/performance/mfxpn3pu",
   "tg_tokens": null,
   "tps": 95.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-mfxpn3pu-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.6 GB at this context; duplicate traces of the same session: f487syic, v9ygemth; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 713.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 8k 713.2 93.2 7.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mfxpn3pu",
   "source_url": "https://omlx.ai/benchmarks/performance/mfxpn3pu",
   "tg_tokens": null,
   "tps": 93.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-txur5247-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: ke5v1gcs, q1ezc3e1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 766.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 766.3 99.1 5.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/txur5247",
   "source_url": "https://omlx.ai/benchmarks/performance/txur5247",
   "tg_tokens": null,
   "tps": 99.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-txur5247-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.3 GB at this context; duplicate traces of the same session: ke5v1gcs, q1ezc3e1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 723.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 723.5 96.0 7.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/txur5247",
   "source_url": "https://omlx.ai/benchmarks/performance/txur5247",
   "tg_tokens": null,
   "tps": 96.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-txur5247-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.6 GB at this context; duplicate traces of the same session: ke5v1gcs, q1ezc3e1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 712.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 8k 712.4 92.9 7.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/txur5247",
   "source_url": "https://omlx.ai/benchmarks/performance/txur5247",
   "tg_tokens": null,
   "tps": 92.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-1643l88k-ctx131072",
   "model": "Qwen3.6-35B-A3B-Fable-5-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace, single run at 128k context; peak memory 37.4 GB; TTFT 267732 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 489.6,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-Fable-5-Distill-oQ6-fp16-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 128k tokens 489.6 PP tok/s 44.3 TG tok/s 267732 TTFT (ms) 37.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1643l88k",
   "source_url": "https://omlx.ai/benchmarks/performance/1643l88k",
   "tg_tokens": null,
   "tps": 44.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-g0df5v38-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: b1e3nobn; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 791.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 791.0 91.3 21.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g0df5v38",
   "source_url": "https://omlx.ai/benchmarks/performance/g0df5v38",
   "tg_tokens": null,
   "tps": 91.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-g0df5v38-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: b1e3nobn; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 840.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 840.0 84.2 22.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g0df5v38",
   "source_url": "https://omlx.ai/benchmarks/performance/g0df5v38",
   "tg_tokens": null,
   "tps": 84.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-a8r7lsyk-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 9qzqiphb; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 672.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 672.0 68.7 21.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a8r7lsyk",
   "source_url": "https://omlx.ai/benchmarks/performance/a8r7lsyk",
   "tg_tokens": null,
   "tps": 68.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-a8r7lsyk-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 9qzqiphb; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 836.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 836.8 78.7 22.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a8r7lsyk",
   "source_url": "https://omlx.ai/benchmarks/performance/a8r7lsyk",
   "tg_tokens": null,
   "tps": 78.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-88voszpo-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 16.0 GB; TTFT 9767 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k tokens 104.8 PP tok/s 15.8 TG tok/s 9767 TTFT (ms) 16.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/88voszpo",
   "source_url": "https://omlx.ai/benchmarks/performance/88voszpo",
   "tg_tokens": null,
   "tps": 15.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pzx94t6e-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.8 GB at this context; duplicate traces of the same session: 73nlc0t6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 958.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 958.3 19.0 29.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pzx94t6e",
   "source_url": "https://omlx.ai/benchmarks/performance/pzx94t6e",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pzx94t6e-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.3 GB at this context; duplicate traces of the same session: 73nlc0t6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 946.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 946.3 17.8 31.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pzx94t6e",
   "source_url": "https://omlx.ai/benchmarks/performance/pzx94t6e",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-w2omc7oa-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 7p0onszr, kp6cxzs5; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1066.0,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-27B-nvfp4 on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k 1,066 53.0 16.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w2omc7oa",
   "source_url": "https://omlx.ai/benchmarks/performance/w2omc7oa",
   "tg_tokens": null,
   "tps": 53.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-w2omc7oa-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 7p0onszr, kp6cxzs5; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1448.0,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-27B-nvfp4 on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 4k 1,448 54.0 17.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w2omc7oa",
   "source_url": "https://omlx.ai/benchmarks/performance/w2omc7oa",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-w2omc7oa-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 7p0onszr, kp6cxzs5; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1229.0,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-27B-nvfp4 on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 64k 1,229 41.8 20.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w2omc7oa",
   "source_url": "https://omlx.ai/benchmarks/performance/w2omc7oa",
   "tg_tokens": null,
   "tps": 41.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-a2ea69d4-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: nsd0z8c7; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 100.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 100.4 24.3 18.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a2ea69d4",
   "source_url": "https://omlx.ai/benchmarks/performance/a2ea69d4",
   "tg_tokens": null,
   "tps": 24.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-a2ea69d4-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: nsd0z8c7; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 104.0 25.3 19.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a2ea69d4",
   "source_url": "https://omlx.ai/benchmarks/performance/a2ea69d4",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-q75p46mz-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: paizafr8; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 104.1 24.4 18.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q75p46mz",
   "source_url": "https://omlx.ai/benchmarks/performance/q75p46mz",
   "tg_tokens": null,
   "tps": 24.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-q75p46mz-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: paizafr8; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 106.5 25.8 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q75p46mz",
   "source_url": "https://omlx.ai/benchmarks/performance/q75p46mz",
   "tg_tokens": null,
   "tps": 25.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-9dkedenu-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: u0zvnlng; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 103.8 25.8 18.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9dkedenu",
   "source_url": "https://omlx.ai/benchmarks/performance/9dkedenu",
   "tg_tokens": null,
   "tps": 25.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-9dkedenu-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: u0zvnlng; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 106.5 28.2 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9dkedenu",
   "source_url": "https://omlx.ai/benchmarks/performance/9dkedenu",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-q9r5hfx2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 5kfykzy1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 127.8 24.3 19.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q9r5hfx2",
   "source_url": "https://omlx.ai/benchmarks/performance/q9r5hfx2",
   "tg_tokens": null,
   "tps": 24.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-q9r5hfx2-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 5kfykzy1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 120.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 120.7 28.1 20.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q9r5hfx2",
   "source_url": "https://omlx.ai/benchmarks/performance/q9r5hfx2",
   "tg_tokens": null,
   "tps": 28.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9mpy3d29-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: ylevzlty; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 306.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 306.3 14.0 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9mpy3d29",
   "source_url": "https://omlx.ai/benchmarks/performance/9mpy3d29",
   "tg_tokens": null,
   "tps": 14.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9mpy3d29-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: ylevzlty; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 352.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 352.1 13.7 20.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9mpy3d29",
   "source_url": "https://omlx.ai/benchmarks/performance/9mpy3d29",
   "tg_tokens": null,
   "tps": 13.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-uw3tud0n-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: c8d6kd33; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 302.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 302.7 13.6 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uw3tud0n",
   "source_url": "https://omlx.ai/benchmarks/performance/uw3tud0n",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-uw3tud0n-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: c8d6kd33; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 346.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 346.9 13.7 20.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uw3tud0n",
   "source_url": "https://omlx.ai/benchmarks/performance/uw3tud0n",
   "tg_tokens": null,
   "tps": 13.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kj3og3v9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: uz9ccj57; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 289.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 289.8 27.3 17.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kj3og3v9",
   "source_url": "https://omlx.ai/benchmarks/performance/kj3og3v9",
   "tg_tokens": null,
   "tps": 27.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kj3og3v9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: uz9ccj57; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 367.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 367.8 22.0 18.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kj3og3v9",
   "source_url": "https://omlx.ai/benchmarks/performance/kj3og3v9",
   "tg_tokens": null,
   "tps": 22.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-1acchwon-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: bw2bepe6, itvaqpx4, j2oya9be; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 912.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 1k 912.4 50.5 33.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1acchwon",
   "source_url": "https://omlx.ai/benchmarks/performance/1acchwon",
   "tg_tokens": null,
   "tps": 50.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-1acchwon-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: bw2bepe6, itvaqpx4, j2oya9be; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1378.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 4k 1,378 72.4 34.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1acchwon",
   "source_url": "https://omlx.ai/benchmarks/performance/1acchwon",
   "tg_tokens": null,
   "tps": 72.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-1acchwon-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.1 GB at this context; duplicate traces of the same session: bw2bepe6, itvaqpx4, j2oya9be; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1111.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 64k 1,111 68.8 39.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1acchwon",
   "source_url": "https://omlx.ai/benchmarks/performance/1acchwon",
   "tg_tokens": null,
   "tps": 68.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-1acchwon-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.8 GB at this context; duplicate traces of the same session: bw2bepe6, itvaqpx4, j2oya9be; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 887.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 128k 887.2 62.9 43.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1acchwon",
   "source_url": "https://omlx.ai/benchmarks/performance/1acchwon",
   "tg_tokens": null,
   "tps": 62.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yl4jziuv-ctx131072",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace, single run at 128k context; peak memory 93.6 GB; TTFT 59238 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2213.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k tokens 2,213 PP tok/s 50.1 TG tok/s 59238 TTFT (ms) 93.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yl4jziuv",
   "source_url": "https://omlx.ai/benchmarks/performance/yl4jziuv",
   "tg_tokens": null,
   "tps": 50.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-a700qiws-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: dbvazq2t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1230.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,230 84.3 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a700qiws",
   "source_url": "https://omlx.ai/benchmarks/performance/a700qiws",
   "tg_tokens": null,
   "tps": 84.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-a700qiws-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: dbvazq2t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2087.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,087 59.9 71.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a700qiws",
   "source_url": "https://omlx.ai/benchmarks/performance/a700qiws",
   "tg_tokens": null,
   "tps": 59.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bammn566-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1220.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,220 94.2 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bammn566",
   "source_url": "https://omlx.ai/benchmarks/performance/bammn566",
   "tg_tokens": null,
   "tps": 94.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bammn566-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2062.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,062 64.0 71.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bammn566",
   "source_url": "https://omlx.ai/benchmarks/performance/bammn566",
   "tg_tokens": null,
   "tps": 64.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-c6lplaa0-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 71.2 GB; TTFT 831 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1232.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,232 PP tok/s 66.2 TG tok/s 831 TTFT (ms) 71.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c6lplaa0",
   "source_url": "https://omlx.ai/benchmarks/performance/c6lplaa0",
   "tg_tokens": null,
   "tps": 66.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fhv8lkp2-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: cja8pptq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 339.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 339.7 28.1 70.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fhv8lkp2",
   "source_url": "https://omlx.ai/benchmarks/performance/fhv8lkp2",
   "tg_tokens": null,
   "tps": 28.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fhv8lkp2-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: cja8pptq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1072.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,072 63.6 71.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fhv8lkp2",
   "source_url": "https://omlx.ai/benchmarks/performance/fhv8lkp2",
   "tg_tokens": null,
   "tps": 63.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-06l986pn-ctx131072",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace, single run at 128k context; peak memory 20.9 GB; TTFT 199057 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 658.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k tokens 658.5 PP tok/s 40.1 TG tok/s 199057 TTFT (ms) 20.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/06l986pn",
   "source_url": "https://omlx.ai/benchmarks/performance/06l986pn",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-i6f7m6w3-ctx1024",
   "model": "Qwen2.5-1.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 1.5 GB at this context; duplicate traces of the same session: none; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 859.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen2.5-1.5B-Instruct-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 unknown; 1k 859.7 120.7 1.5 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i6f7m6w3",
   "source_url": "https://omlx.ai/benchmarks/performance/i6f7m6w3",
   "tg_tokens": null,
   "tps": 120.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-i6f7m6w3-ctx4096",
   "model": "Qwen2.5-1.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 1.6 GB at this context; duplicate traces of the same session: none; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 923.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen2.5-1.5B-Instruct-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 unknown; 4k 923.5 101.1 1.6 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i6f7m6w3",
   "source_url": "https://omlx.ai/benchmarks/performance/i6f7m6w3",
   "tg_tokens": null,
   "tps": 101.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-8jrm04qd-ctx4096",
   "model": "AREX-Turbo",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: ff00asmm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1426.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "AREX-Turbo-8bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 8bit; 4k 1,426 68.8 5.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8jrm04qd",
   "source_url": "https://omlx.ai/benchmarks/performance/8jrm04qd",
   "tg_tokens": null,
   "tps": 68.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-8jrm04qd-ctx16384",
   "model": "AREX-Turbo",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: ff00asmm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1304.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "AREX-Turbo-8bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 8bit; 16k 1,304 63.4 6.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8jrm04qd",
   "source_url": "https://omlx.ai/benchmarks/performance/8jrm04qd",
   "tg_tokens": null,
   "tps": 63.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-xnze9r6y-ctx1024",
   "model": "K2-Horizon-7B",
   "notes": "oMLX community benchmark trace; peak memory 5.5 GB at this context; duplicate traces of the same session: 29y51142, q11bdamv, w837j4zv; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 301.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "K2-Horizon-7B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 1k 301.2 55.5 5.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xnze9r6y",
   "source_url": "https://omlx.ai/benchmarks/performance/xnze9r6y",
   "tg_tokens": null,
   "tps": 55.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-xnze9r6y-ctx4096",
   "model": "K2-Horizon-7B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: 29y51142, q11bdamv, w837j4zv; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 286.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "K2-Horizon-7B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k 286.1 50.0 5.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xnze9r6y",
   "source_url": "https://omlx.ai/benchmarks/performance/xnze9r6y",
   "tg_tokens": null,
   "tps": 50.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-xnze9r6y-ctx8192",
   "model": "K2-Horizon-7B",
   "notes": "oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: 29y51142, q11bdamv, w837j4zv; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 268.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "K2-Horizon-7B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 8k 268.2 43.8 6.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xnze9r6y",
   "source_url": "https://omlx.ai/benchmarks/performance/xnze9r6y",
   "tg_tokens": null,
   "tps": 43.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-xnze9r6y-ctx16384",
   "model": "K2-Horizon-7B",
   "notes": "oMLX community benchmark trace; peak memory 7.5 GB at this context; duplicate traces of the same session: 29y51142, q11bdamv, w837j4zv; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 234.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "K2-Horizon-7B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 16k 234.4 35.2 7.5 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xnze9r6y",
   "source_url": "https://omlx.ai/benchmarks/performance/xnze9r6y",
   "tg_tokens": null,
   "tps": 35.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-dyc93m8r-ctx32768",
   "model": "LFM2-24B-A2B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 14.1 GB; TTFT 49520 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 661.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2-24B-A2B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 32k tokens 661.7 PP tok/s 50.3 TG tok/s 49520 TTFT (ms) 14.1 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dyc93m8r",
   "source_url": "https://omlx.ai/benchmarks/performance/dyc93m8r",
   "tg_tokens": null,
   "tps": 50.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ss5ytznw-ctx32768",
   "model": "LFM2-24B-A2B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 14.1 GB; TTFT 49557 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 661.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2-24B-A2B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 32k tokens 661.2 PP tok/s 49.7 TG tok/s 49557 TTFT (ms) 14.1 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ss5ytznw",
   "source_url": "https://omlx.ai/benchmarks/performance/ss5ytznw",
   "tg_tokens": null,
   "tps": 49.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-cw03vp22-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 32k context; peak memory 14.2 GB; TTFT 52212 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 627.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 32k tokens 627.6 PP tok/s 33.9 TG tok/s 52212 TTFT (ms) 14.2 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cw03vp22",
   "source_url": "https://omlx.ai/benchmarks/performance/cw03vp22",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-q9yilyyp-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1057.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-24B-A2B-oQ8e on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k 1,057 62.3 24.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q9yilyyp",
   "source_url": "https://omlx.ai/benchmarks/performance/q9yilyyp",
   "tg_tokens": null,
   "tps": 62.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-q9yilyyp-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 683.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-24B-A2B-oQ8e on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 16k 683.6 56.8 25.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q9yilyyp",
   "source_url": "https://omlx.ai/benchmarks/performance/q9yilyyp",
   "tg_tokens": null,
   "tps": 56.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-fp4zlduj-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: 2e1xudhx, rrnyy5l1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2159.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-24B-A2B-oQ8e on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 2,159 97.5 24.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fp4zlduj",
   "source_url": "https://omlx.ai/benchmarks/performance/fp4zlduj",
   "tg_tokens": null,
   "tps": 97.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-fp4zlduj-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: 2e1xudhx, rrnyy5l1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2356.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-24B-A2B-oQ8e on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 2,356 92.6 25.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fp4zlduj",
   "source_url": "https://omlx.ai/benchmarks/performance/fp4zlduj",
   "tg_tokens": null,
   "tps": 92.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-fp4zlduj-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 28.8 GB at this context; duplicate traces of the same session: 2e1xudhx, rrnyy5l1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1822.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-24B-A2B-oQ8e on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 1,822 82.1 28.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fp4zlduj",
   "source_url": "https://omlx.ai/benchmarks/performance/fp4zlduj",
   "tg_tokens": null,
   "tps": 82.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-zha2c58n-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.8 GB at this context; duplicate traces of the same session: xdzsldjw; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1346.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.5-9B-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 1,346 23.6 7.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zha2c58n",
   "source_url": "https://omlx.ai/benchmarks/performance/zha2c58n",
   "tg_tokens": null,
   "tps": 23.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-zha2c58n-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.2 GB at this context; duplicate traces of the same session: xdzsldjw; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1446.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.5-9B-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 1,446 31.1 8.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zha2c58n",
   "source_url": "https://omlx.ai/benchmarks/performance/zha2c58n",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xihje9t7-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 20.6 GB; TTFT 6611 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4957.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k tokens 4,957 PP tok/s 121.3 TG tok/s 6611 TTFT (ms) 20.6 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xihje9t7",
   "source_url": "https://omlx.ai/benchmarks/performance/xihje9t7",
   "tg_tokens": null,
   "tps": 121.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-cknr7016-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: d5esg4js; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 850.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 850.5 152.2 19.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cknr7016",
   "source_url": "https://omlx.ai/benchmarks/performance/cknr7016",
   "tg_tokens": null,
   "tps": 152.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-cknr7016-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: d5esg4js; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2347.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 2,347 147.9 20.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cknr7016",
   "source_url": "https://omlx.ai/benchmarks/performance/cknr7016",
   "tg_tokens": null,
   "tps": 147.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-9q925pf3-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: xr2lh8p3; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1424.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 1,424 94.1 24.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9q925pf3",
   "source_url": "https://omlx.ai/benchmarks/performance/9q925pf3",
   "tg_tokens": null,
   "tps": 94.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-9q925pf3-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: xr2lh8p3; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 880.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 880.3 71.5 25.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9q925pf3",
   "source_url": "https://omlx.ai/benchmarks/performance/9q925pf3",
   "tg_tokens": null,
   "tps": 71.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-rp1koaxb-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 8v1opp1s, g7y2cvui, jd4f34v2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 774.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 774.7 102.1 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rp1koaxb",
   "source_url": "https://omlx.ai/benchmarks/performance/rp1koaxb",
   "tg_tokens": null,
   "tps": 102.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-rp1koaxb-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: 8v1opp1s, g7y2cvui, jd4f34v2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 766.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 8k 766.8 76.1 24.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rp1koaxb",
   "source_url": "https://omlx.ai/benchmarks/performance/rp1koaxb",
   "tg_tokens": null,
   "tps": 76.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-rp1koaxb-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 8v1opp1s, g7y2cvui, jd4f34v2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 688.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 16k 688.0 71.0 23.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rp1koaxb",
   "source_url": "https://omlx.ai/benchmarks/performance/rp1koaxb",
   "tg_tokens": null,
   "tps": 71.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-rp1koaxb-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: 8v1opp1s, g7y2cvui, jd4f34v2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 529.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 32k 529.2 85.0 23.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rp1koaxb",
   "source_url": "https://omlx.ai/benchmarks/performance/rp1koaxb",
   "tg_tokens": null,
   "tps": 85.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-rp1koaxb-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: 8v1opp1s, g7y2cvui, jd4f34v2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 403.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 64k 403.0 53.4 24.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rp1koaxb",
   "source_url": "https://omlx.ai/benchmarks/performance/rp1koaxb",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-4vblz1ly-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 11.2 GB at this context; duplicate traces of the same session: 3efgv8dl, dkvpmnp3, f53uci7z, gpb8ieci, q2zo7a6r, rnqnxg98, s91i4avj; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 324.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 324.6 23.1 11.2 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4vblz1ly",
   "source_url": "https://omlx.ai/benchmarks/performance/4vblz1ly",
   "tg_tokens": null,
   "tps": 23.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-4vblz1ly-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 11.9 GB at this context; duplicate traces of the same session: 3efgv8dl, dkvpmnp3, f53uci7z, gpb8ieci, q2zo7a6r, rnqnxg98, s91i4avj; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 607.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 607.1 22.7 11.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4vblz1ly",
   "source_url": "https://omlx.ai/benchmarks/performance/4vblz1ly",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5",
   "id": "omlx-4vblz1ly-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 11.9 GB at this context; duplicate traces of the same session: 3efgv8dl, dkvpmnp3, f53uci7z, gpb8ieci, q2zo7a6r, rnqnxg98, s91i4avj; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 611.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 8k 611.3 22.1 11.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4vblz1ly",
   "source_url": "https://omlx.ai/benchmarks/performance/4vblz1ly",
   "tg_tokens": null,
   "tps": 22.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5",
   "id": "omlx-4vblz1ly-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 12.1 GB at this context; duplicate traces of the same session: 3efgv8dl, dkvpmnp3, f53uci7z, gpb8ieci, q2zo7a6r, rnqnxg98, s91i4avj; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 583.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 16k 583.5 19.6 12.1 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4vblz1ly",
   "source_url": "https://omlx.ai/benchmarks/performance/4vblz1ly",
   "tg_tokens": null,
   "tps": 19.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5",
   "id": "omlx-4vblz1ly-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 12.4 GB at this context; duplicate traces of the same session: 3efgv8dl, dkvpmnp3, f53uci7z, gpb8ieci, q2zo7a6r, rnqnxg98, s91i4avj; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 536.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 32k 536.9 16.8 12.4 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4vblz1ly",
   "source_url": "https://omlx.ai/benchmarks/performance/4vblz1ly",
   "tg_tokens": null,
   "tps": 16.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-4vblz1ly-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 13.0 GB at this context; duplicate traces of the same session: 3efgv8dl, dkvpmnp3, f53uci7z, gpb8ieci, q2zo7a6r, rnqnxg98, s91i4avj; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 447.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 64k 447.1 13.5 13.0 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4vblz1ly",
   "source_url": "https://omlx.ai/benchmarks/performance/4vblz1ly",
   "tg_tokens": null,
   "tps": 13.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-4vblz1ly-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: 3efgv8dl, dkvpmnp3, f53uci7z, gpb8ieci, q2zo7a6r, rnqnxg98, s91i4avj; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 273.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 128k 273.7 5.5 14.3 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4vblz1ly",
   "source_url": "https://omlx.ai/benchmarks/performance/4vblz1ly",
   "tg_tokens": null,
   "tps": 5.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-4vblz1ly-ctx199680",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 15.6 GB at this context; duplicate traces of the same session: 3efgv8dl, dkvpmnp3, f53uci7z, gpb8ieci, q2zo7a6r, rnqnxg98, s91i4avj; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 263.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 195k 263.6 7.0 15.6 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4vblz1ly",
   "source_url": "https://omlx.ai/benchmarks/performance/4vblz1ly",
   "tg_tokens": null,
   "tps": 7.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-itbts558-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 46.5 GB at this context; duplicate traces of the same session: vg7j04y3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 632.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 632.8 14.5 46.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/itbts558",
   "source_url": "https://omlx.ai/benchmarks/performance/itbts558",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-itbts558-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 55.3 GB at this context; duplicate traces of the same session: vg7j04y3; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 520.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 195k 520.2 12.8 55.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/itbts558",
   "source_url": "https://omlx.ai/benchmarks/performance/itbts558",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-p89eo1av-ctx131072",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "MTP text build; oMLX community benchmark trace, single run at 128k context; peak memory 29.7 GB; TTFT 1529658 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 85.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-oQ4e-MTP-Text on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k tokens 85.7 PP tok/s 8.4 TG tok/s 1529658 TTFT (ms) 29.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p89eo1av",
   "source_url": "https://omlx.ai/benchmarks/performance/p89eo1av",
   "tg_tokens": null,
   "tps": 8.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-s0t3uu9h-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 199sea81, eft7gfs7, klsq1610, lr8u5vjb; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1206.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 6bit; 1k 1,206 37.9 22.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s0t3uu9h",
   "source_url": "https://omlx.ai/benchmarks/performance/s0t3uu9h",
   "tg_tokens": null,
   "tps": 37.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-s0t3uu9h-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 199sea81, eft7gfs7, klsq1610, lr8u5vjb; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1370.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 6bit; 4k 1,370 37.8 23.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s0t3uu9h",
   "source_url": "https://omlx.ai/benchmarks/performance/s0t3uu9h",
   "tg_tokens": null,
   "tps": 37.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-s0t3uu9h-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 27.0 GB at this context; duplicate traces of the same session: 199sea81, eft7gfs7, klsq1610, lr8u5vjb; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1142.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 6bit; 64k 1,142 32.1 27.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s0t3uu9h",
   "source_url": "https://omlx.ai/benchmarks/performance/s0t3uu9h",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-s0t3uu9h-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.0 GB at this context; duplicate traces of the same session: 199sea81, eft7gfs7, klsq1610, lr8u5vjb; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 917.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 6bit; 128k 917.7 28.0 31.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s0t3uu9h",
   "source_url": "https://omlx.ai/benchmarks/performance/s0t3uu9h",
   "tg_tokens": null,
   "tps": 28.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-s0t3uu9h-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 35.3 GB at this context; duplicate traces of the same session: 199sea81, eft7gfs7, klsq1610, lr8u5vjb; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 748.2,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 6bit; 195k 748.2 24.9 35.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s0t3uu9h",
   "source_url": "https://omlx.ai/benchmarks/performance/s0t3uu9h",
   "tg_tokens": null,
   "tps": 24.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tvxns3gu-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: 39xpg3cr, 758032r2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 918.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 918.2 73.9 17.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tvxns3gu",
   "source_url": "https://omlx.ai/benchmarks/performance/tvxns3gu",
   "tg_tokens": null,
   "tps": 73.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tvxns3gu-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: 39xpg3cr, 758032r2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 911.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 911.7 82.2 22.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tvxns3gu",
   "source_url": "https://omlx.ai/benchmarks/performance/tvxns3gu",
   "tg_tokens": null,
   "tps": 82.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-tvxns3gu-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 45.6 GB at this context; duplicate traces of the same session: 39xpg3cr, 758032r2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 518.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 195k 518.6 35.4 45.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tvxns3gu",
   "source_url": "https://omlx.ai/benchmarks/performance/tvxns3gu",
   "tg_tokens": null,
   "tps": 35.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-a9g4q0ka-ctx199680",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace, single run at 195k context; peak memory 96.6 GB; TTFT 87443 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2287.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 195k tokens 2,287 PP tok/s 57.5 TG tok/s 87443 TTFT (ms) 96.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a9g4q0ka",
   "source_url": "https://omlx.ai/benchmarks/performance/a9g4q0ka",
   "tg_tokens": null,
   "tps": 57.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jd6jw0ks-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 70fvg9xk, nhagt26w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1994.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,994 100.9 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jd6jw0ks",
   "source_url": "https://omlx.ai/benchmarks/performance/jd6jw0ks",
   "tg_tokens": null,
   "tps": 100.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jd6jw0ks-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: 70fvg9xk, nhagt26w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2412.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,412 78.8 72.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jd6jw0ks",
   "source_url": "https://omlx.ai/benchmarks/performance/jd6jw0ks",
   "tg_tokens": null,
   "tps": 78.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jd6jw0ks-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.7 GB at this context; duplicate traces of the same session: 70fvg9xk, nhagt26w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2446.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,446 67.5 73.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jd6jw0ks",
   "source_url": "https://omlx.ai/benchmarks/performance/jd6jw0ks",
   "tg_tokens": null,
   "tps": 67.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-98cg96wb-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.5 GB at this context; duplicate traces of the same session: 8287w3lz, ibbtm1mv, kekn7bsv, luaz87io, p96b3rmt, xcorgckx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 711.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 711.7 85.7 142.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/98cg96wb",
   "source_url": "https://omlx.ai/benchmarks/performance/98cg96wb",
   "tg_tokens": null,
   "tps": 85.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-98cg96wb-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: 8287w3lz, ibbtm1mv, kekn7bsv, luaz87io, p96b3rmt, xcorgckx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 913.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 913.0 66.1 143.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/98cg96wb",
   "source_url": "https://omlx.ai/benchmarks/performance/98cg96wb",
   "tg_tokens": null,
   "tps": 66.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-98cg96wb-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.4 GB at this context; duplicate traces of the same session: 8287w3lz, ibbtm1mv, kekn7bsv, luaz87io, p96b3rmt, xcorgckx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1333.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 8k 1,333 90.4 143.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/98cg96wb",
   "source_url": "https://omlx.ai/benchmarks/performance/98cg96wb",
   "tg_tokens": null,
   "tps": 90.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-98cg96wb-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.7 GB at this context; duplicate traces of the same session: 8287w3lz, ibbtm1mv, kekn7bsv, luaz87io, p96b3rmt, xcorgckx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1326.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 16k 1,326 86.3 143.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/98cg96wb",
   "source_url": "https://omlx.ai/benchmarks/performance/98cg96wb",
   "tg_tokens": null,
   "tps": 86.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-98cg96wb-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 146.8 GB at this context; duplicate traces of the same session: 8287w3lz, ibbtm1mv, kekn7bsv, luaz87io, p96b3rmt, xcorgckx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1314.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 64k 1,314 65.0 146.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/98cg96wb",
   "source_url": "https://omlx.ai/benchmarks/performance/98cg96wb",
   "tg_tokens": null,
   "tps": 65.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-98cg96wb-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 151.2 GB at this context; duplicate traces of the same session: 8287w3lz, ibbtm1mv, kekn7bsv, luaz87io, p96b3rmt, xcorgckx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1290.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 128k 1,290 62.9 151.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/98cg96wb",
   "source_url": "https://omlx.ai/benchmarks/performance/98cg96wb",
   "tg_tokens": null,
   "tps": 62.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-98cg96wb-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 152.2 GB at this context; duplicate traces of the same session: 8287w3lz, ibbtm1mv, kekn7bsv, luaz87io, p96b3rmt, xcorgckx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1274.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 195k 1,274 51.1 152.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/98cg96wb",
   "source_url": "https://omlx.ai/benchmarks/performance/98cg96wb",
   "tg_tokens": null,
   "tps": 51.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-moq04j4m-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 28.2 GB at this context; duplicate traces of the same session: 9jeoq2ds; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1654.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26b-a4b-it-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 1,654 49.9 28.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/moq04j4m",
   "source_url": "https://omlx.ai/benchmarks/performance/moq04j4m",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-moq04j4m-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: 9jeoq2ds; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1987.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26b-a4b-it-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 16k 1,987 54.2 28.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/moq04j4m",
   "source_url": "https://omlx.ai/benchmarks/performance/moq04j4m",
   "tg_tokens": null,
   "tps": 54.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-a0p6z0cg-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.0 GB at this context; duplicate traces of the same session: h0w7ratk, wfptm4m2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 88.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 88.0 11.5 92.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a0p6z0cg",
   "source_url": "https://omlx.ai/benchmarks/performance/a0p6z0cg",
   "tg_tokens": null,
   "tps": 11.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-a0p6z0cg-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 93.8 GB at this context; duplicate traces of the same session: h0w7ratk, wfptm4m2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 436.6,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 436.6 11.5 93.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a0p6z0cg",
   "source_url": "https://omlx.ai/benchmarks/performance/a0p6z0cg",
   "tg_tokens": null,
   "tps": 11.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-a0p6z0cg-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 94.1 GB at this context; duplicate traces of the same session: h0w7ratk, wfptm4m2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 455.4,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 455.4 13.8 94.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a0p6z0cg",
   "source_url": "https://omlx.ai/benchmarks/performance/a0p6z0cg",
   "tg_tokens": null,
   "tps": 13.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ea2mi5n8-ctx4096",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 175.1 GB at this context; duplicate traces of the same session: 0rzjjm2g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 469.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 469.2 34.0 175.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ea2mi5n8",
   "source_url": "https://omlx.ai/benchmarks/performance/ea2mi5n8",
   "tg_tokens": null,
   "tps": 34.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ea2mi5n8-ctx16384",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 175.7 GB at this context; duplicate traces of the same session: 0rzjjm2g; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 435.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 435.6 34.1 175.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ea2mi5n8",
   "source_url": "https://omlx.ai/benchmarks/performance/ea2mi5n8",
   "tg_tokens": null,
   "tps": 34.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xccyktiz-ctx4096",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 292.5 GB at this context; duplicate traces of the same session: hw8csgqa; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 395.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "GLM-5.3-Flash-Uncensored-MLX-6bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 4k 395.8 5.2 292.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xccyktiz",
   "source_url": "https://omlx.ai/benchmarks/performance/xccyktiz",
   "tg_tokens": null,
   "tps": 5.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xccyktiz-ctx16384",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 293.4 GB at this context; duplicate traces of the same session: hw8csgqa; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 357.2,
   "provenance": "community",
   "quant": "6bit",
   "quote": "GLM-5.3-Flash-Uncensored-MLX-6bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 16k 357.2 5.1 293.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xccyktiz",
   "source_url": "https://omlx.ai/benchmarks/performance/xccyktiz",
   "tg_tokens": null,
   "tps": 5.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ctcr8myl-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 177.6 GB; TTFT 2765 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 370.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k tokens 370.3 PP tok/s 42.1 TG tok/s 2765 TTFT (ms) 177.6 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ctcr8myl",
   "source_url": "https://omlx.ai/benchmarks/performance/ctcr8myl",
   "tg_tokens": null,
   "tps": 42.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-h0k7cap1-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: ose8z5h6, xngrw81e; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 636.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 636.2 67.1 13.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h0k7cap1",
   "source_url": "https://omlx.ai/benchmarks/performance/h0k7cap1",
   "tg_tokens": null,
   "tps": 67.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-h0k7cap1-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: ose8z5h6, xngrw81e; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 568.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 568.9 44.4 13.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h0k7cap1",
   "source_url": "https://omlx.ai/benchmarks/performance/h0k7cap1",
   "tg_tokens": null,
   "tps": 44.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-h0k7cap1-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.9 GB at this context; duplicate traces of the same session: ose8z5h6, xngrw81e; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 530.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 530.5 53.6 13.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h0k7cap1",
   "source_url": "https://omlx.ai/benchmarks/performance/h0k7cap1",
   "tg_tokens": null,
   "tps": 53.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-98x5sgrx-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: l4ta1hn6; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1163.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "LFM2.5-24B-A2B-oQ5e-fp16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 5bit; 4k 1,163 59.7 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/98x5sgrx",
   "source_url": "https://omlx.ai/benchmarks/performance/98x5sgrx",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-98x5sgrx-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 16.8 GB at this context; duplicate traces of the same session: l4ta1hn6; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 682.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "LFM2.5-24B-A2B-oQ5e-fp16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 5bit; 16k 682.2 62.6 16.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/98x5sgrx",
   "source_url": "https://omlx.ai/benchmarks/performance/98x5sgrx",
   "tg_tokens": null,
   "tps": 62.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-k9hwirra-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1057.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-24B-A2B-oQ8e on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k 1,057 62.3 24.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k9hwirra",
   "source_url": "https://omlx.ai/benchmarks/performance/k9hwirra",
   "tg_tokens": null,
   "tps": 62.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-k9hwirra-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 683.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-24B-A2B-oQ8e on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 16k 683.6 56.8 25.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k9hwirra",
   "source_url": "https://omlx.ai/benchmarks/performance/k9hwirra",
   "tg_tokens": null,
   "tps": 56.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-9ct4zc1l-ctx4096",
   "model": "Qwen3 8B",
   "notes": "oMLX community benchmark trace; peak memory 5.4 GB at this context; duplicate traces of the same session: u8lnyezs; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 195.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-8B-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 195.4 19.4 5.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ct4zc1l",
   "source_url": "https://omlx.ai/benchmarks/performance/9ct4zc1l",
   "tg_tokens": null,
   "tps": 19.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-9ct4zc1l-ctx16384",
   "model": "Qwen3 8B",
   "notes": "oMLX community benchmark trace; peak memory 7.1 GB at this context; duplicate traces of the same session: u8lnyezs; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 153.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-8B-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 16k 153.2 12.9 7.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ct4zc1l",
   "source_url": "https://omlx.ai/benchmarks/performance/9ct4zc1l",
   "tg_tokens": null,
   "tps": 12.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-1ncoelsa-ctx1024",
   "model": "Qwen3.6-27B-Claude-Opus-Reasoning-Distill-v2-abliterated",
   "notes": "OptiQ MLX build; oMLX community benchmark trace; peak memory 14.5 GB at this context; duplicate traces of the same session: zcb23say; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 375.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-27B-Claude-Opus-Reasoning-Distill-v2-abliterated-OptiQ-3.7bpw-mlx on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 375.2 18.6 14.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ncoelsa",
   "source_url": "https://omlx.ai/benchmarks/performance/1ncoelsa",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-1ncoelsa-ctx4096",
   "model": "Qwen3.6-27B-Claude-Opus-Reasoning-Distill-v2-abliterated",
   "notes": "OptiQ MLX build; oMLX community benchmark trace; peak memory 15.3 GB at this context; duplicate traces of the same session: zcb23say; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 411.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-27B-Claude-Opus-Reasoning-Distill-v2-abliterated-OptiQ-3.7bpw-mlx on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 411.6 18.5 15.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ncoelsa",
   "source_url": "https://omlx.ai/benchmarks/performance/1ncoelsa",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-omwbj9ml-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 20.6 GB; TTFT 6644 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4932.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k tokens 4,932 PP tok/s 120.8 TG tok/s 6644 TTFT (ms) 20.6 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/omwbj9ml",
   "source_url": "https://omlx.ai/benchmarks/performance/omwbj9ml",
   "tg_tokens": null,
   "tps": 120.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fa6ludqr-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 20.6 GB; TTFT 6583 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4978.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k tokens 4,978 PP tok/s 120.3 TG tok/s 6583 TTFT (ms) 20.6 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fa6ludqr",
   "source_url": "https://omlx.ai/benchmarks/performance/fa6ludqr",
   "tg_tokens": null,
   "tps": 120.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-xpzqdk0h-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 774.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 774.7 102.1 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xpzqdk0h",
   "source_url": "https://omlx.ai/benchmarks/performance/xpzqdk0h",
   "tg_tokens": null,
   "tps": 102.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-xpzqdk0h-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 766.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 8k 766.8 76.1 24.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xpzqdk0h",
   "source_url": "https://omlx.ai/benchmarks/performance/xpzqdk0h",
   "tg_tokens": null,
   "tps": 76.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-xpzqdk0h-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 688.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 16k 688.0 71.0 23.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xpzqdk0h",
   "source_url": "https://omlx.ai/benchmarks/performance/xpzqdk0h",
   "tg_tokens": null,
   "tps": 71.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-xpzqdk0h-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 529.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 32k 529.2 85.0 23.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xpzqdk0h",
   "source_url": "https://omlx.ai/benchmarks/performance/xpzqdk0h",
   "tg_tokens": null,
   "tps": 85.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-xpzqdk0h-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 403.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 64k 403.0 53.4 24.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xpzqdk0h",
   "source_url": "https://omlx.ai/benchmarks/performance/xpzqdk0h",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cqqccxmv-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: p2dbvui9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5497.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 5,497 108.3 21.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cqqccxmv",
   "source_url": "https://omlx.ai/benchmarks/performance/cqqccxmv",
   "tg_tokens": null,
   "tps": 108.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cqqccxmv-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: p2dbvui9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4835.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 4,835 97.1 21.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cqqccxmv",
   "source_url": "https://omlx.ai/benchmarks/performance/cqqccxmv",
   "tg_tokens": null,
   "tps": 97.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-olv523da-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 22.3 GB; TTFT 3168 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5172.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k tokens 5,172 PP tok/s 82.5 TG tok/s 3168 TTFT (ms) 22.3 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/olv523da",
   "source_url": "https://omlx.ai/benchmarks/performance/olv523da",
   "tg_tokens": null,
   "tps": 82.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vf3edn9v-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 20.1 GB; TTFT 208 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4920.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k tokens 4,920 PP tok/s 141.1 TG tok/s 208 TTFT (ms) 20.1 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vf3edn9v",
   "source_url": "https://omlx.ai/benchmarks/performance/vf3edn9v",
   "tg_tokens": null,
   "tps": 141.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-hfdp3nbh-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 31.9 GB; TTFT 38388 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1707.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 64k tokens 1,707 PP tok/s 67.7 TG tok/s 38388 TTFT (ms) 31.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hfdp3nbh",
   "source_url": "https://omlx.ai/benchmarks/performance/hfdp3nbh",
   "tg_tokens": null,
   "tps": 67.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-r3ax15fs-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 31.9 GB; TTFT 40772 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1607.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 64k tokens 1,607 PP tok/s 65.0 TG tok/s 40772 TTFT (ms) 31.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r3ax15fs",
   "source_url": "https://omlx.ai/benchmarks/performance/r3ax15fs",
   "tg_tokens": null,
   "tps": 65.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-hk3xep4w-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 31.9 GB; TTFT 38425 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1706.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 64k tokens 1,706 PP tok/s 69.4 TG tok/s 38425 TTFT (ms) 31.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hk3xep4w",
   "source_url": "https://omlx.ai/benchmarks/performance/hk3xep4w",
   "tg_tokens": null,
   "tps": 69.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-faqmo3f4-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 31.2 GB; TTFT 14860 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2205.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 32k tokens 2,205 PP tok/s 117.4 TG tok/s 14860 TTFT (ms) 31.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/faqmo3f4",
   "source_url": "https://omlx.ai/benchmarks/performance/faqmo3f4",
   "tg_tokens": null,
   "tps": 117.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-jt4icwal-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 31.2 GB; TTFT 14957 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2191.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 32k tokens 2,191 PP tok/s 105.5 TG tok/s 14957 TTFT (ms) 31.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jt4icwal",
   "source_url": "https://omlx.ai/benchmarks/performance/jt4icwal",
   "tg_tokens": null,
   "tps": 105.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3ptsz8cg-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 31.2 GB; TTFT 16203 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2022.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 32k tokens 2,022 PP tok/s 53.2 TG tok/s 16203 TTFT (ms) 31.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ptsz8cg",
   "source_url": "https://omlx.ai/benchmarks/performance/3ptsz8cg",
   "tg_tokens": null,
   "tps": 53.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ou7q9x04-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 31.2 GB; TTFT 14822 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2211.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 32k tokens 2,211 PP tok/s 100.4 TG tok/s 14822 TTFT (ms) 31.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ou7q9x04",
   "source_url": "https://omlx.ai/benchmarks/performance/ou7q9x04",
   "tg_tokens": null,
   "tps": 100.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-f1k43sk6-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 31.2 GB; TTFT 14842 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2208.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 32k tokens 2,208 PP tok/s 113.0 TG tok/s 14842 TTFT (ms) 31.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f1k43sk6",
   "source_url": "https://omlx.ai/benchmarks/performance/f1k43sk6",
   "tg_tokens": null,
   "tps": 113.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-x6spx2tq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: yfemn5g7; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 389.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 389.1 17.7 16.0 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x6spx2tq",
   "source_url": "https://omlx.ai/benchmarks/performance/x6spx2tq",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-x6spx2tq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: yfemn5g7; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 398.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 398.2 16.7 17.0 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x6spx2tq",
   "source_url": "https://omlx.ai/benchmarks/performance/x6spx2tq",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-acklg4wc-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: zze4alyc; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 502.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 502.5 18.0 17.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/acklg4wc",
   "source_url": "https://omlx.ai/benchmarks/performance/acklg4wc",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-acklg4wc-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: zze4alyc; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 608.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 608.5 17.8 18.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/acklg4wc",
   "source_url": "https://omlx.ai/benchmarks/performance/acklg4wc",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-epw310ku-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: dklb1h5h; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 386.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 386.5 18.0 16.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/epw310ku",
   "source_url": "https://omlx.ai/benchmarks/performance/epw310ku",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-epw310ku-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: dklb1h5h; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 384.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 384.3 17.5 17.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/epw310ku",
   "source_url": "https://omlx.ai/benchmarks/performance/epw310ku",
   "tg_tokens": null,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-q7lqukym-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: m5jihmwk; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 299.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 299.0 9.9 15.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q7lqukym",
   "source_url": "https://omlx.ai/benchmarks/performance/q7lqukym",
   "tg_tokens": null,
   "tps": 9.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-q7lqukym-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: m5jihmwk; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 314.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 314.3 9.8 17.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q7lqukym",
   "source_url": "https://omlx.ai/benchmarks/performance/q7lqukym",
   "tg_tokens": null,
   "tps": 9.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zs0cycig-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.4 GB at this context; duplicate traces of the same session: rqqida6i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 937.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 937.3 34.5 33.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zs0cycig",
   "source_url": "https://omlx.ai/benchmarks/performance/zs0cycig",
   "tg_tokens": null,
   "tps": 34.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zs0cycig-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: rqqida6i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 922.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 922.5 54.0 34.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zs0cycig",
   "source_url": "https://omlx.ai/benchmarks/performance/zs0cycig",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-a0m09p0e-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.7 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 165.8,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 165.8 19.1 30.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a0m09p0e",
   "source_url": "https://omlx.ai/benchmarks/performance/a0m09p0e",
   "tg_tokens": null,
   "tps": 19.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-a0m09p0e-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.2 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 185.7,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 185.7 19.8 34.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a0m09p0e",
   "source_url": "https://omlx.ai/benchmarks/performance/a0m09p0e",
   "tg_tokens": null,
   "tps": 19.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-wts14hhf-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: qe1l1sc9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 254.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 254.7 58.9 18.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wts14hhf",
   "source_url": "https://omlx.ai/benchmarks/performance/wts14hhf",
   "tg_tokens": null,
   "tps": 58.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-wts14hhf-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: qe1l1sc9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 266.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 266.2 51.2 21.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wts14hhf",
   "source_url": "https://omlx.ai/benchmarks/performance/wts14hhf",
   "tg_tokens": null,
   "tps": 51.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-kiv2i5xk-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: kyu1x52x; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 389.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 389.7 34.3 17.6 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kiv2i5xk",
   "source_url": "https://omlx.ai/benchmarks/performance/kiv2i5xk",
   "tg_tokens": null,
   "tps": 34.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-kiv2i5xk-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: kyu1x52x; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 385.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 385.9 31.8 18.7 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kiv2i5xk",
   "source_url": "https://omlx.ai/benchmarks/performance/kiv2i5xk",
   "tg_tokens": null,
   "tps": 31.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-fyk4veoc-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.4 GB at this context; duplicate traces of the same session: 8kikg5d1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 228.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 228.6 40.5 30.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fyk4veoc",
   "source_url": "https://omlx.ai/benchmarks/performance/fyk4veoc",
   "tg_tokens": null,
   "tps": 40.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-fyk4veoc-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.6 GB at this context; duplicate traces of the same session: 8kikg5d1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 201.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k 201.5 27.8 31.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fyk4veoc",
   "source_url": "https://omlx.ai/benchmarks/performance/fyk4veoc",
   "tg_tokens": null,
   "tps": 27.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ug2x3zl7-ctx1024",
   "model": "Spark-X2.5-4B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 4.8 GB at this context; duplicate traces of the same session: e0x07khw; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 698.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Spark-X2.5-4B-MLX-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 1k 698.4 49.6 4.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ug2x3zl7",
   "source_url": "https://omlx.ai/benchmarks/performance/ug2x3zl7",
   "tg_tokens": null,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ug2x3zl7-ctx32768",
   "model": "Spark-X2.5-4B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: e0x07khw; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 585.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Spark-X2.5-4B-MLX-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 32k 585.8 38.3 5.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ug2x3zl7",
   "source_url": "https://omlx.ai/benchmarks/performance/ug2x3zl7",
   "tg_tokens": null,
   "tps": 38.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-gqzhdi77-ctx1024",
   "model": "Ternary-Bonsai-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 9.1 GB at this context; duplicate traces of the same session: ahe4ex19; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 422.7,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-27B-mlx-2bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 2bit; 1k 422.7 31.9 9.1 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gqzhdi77",
   "source_url": "https://omlx.ai/benchmarks/performance/gqzhdi77",
   "tg_tokens": null,
   "tps": 31.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-gqzhdi77-ctx4096",
   "model": "Ternary-Bonsai-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 10.1 GB at this context; duplicate traces of the same session: ahe4ex19; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 435.5,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-27B-mlx-2bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 2bit; 4k 435.5 31.1 10.1 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gqzhdi77",
   "source_url": "https://omlx.ai/benchmarks/performance/gqzhdi77",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-kukxd75v-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.1 GB at this context; duplicate traces of the same session: 1k0xdolc; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2061.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 2,061 79.9 14.1 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kukxd75v",
   "source_url": "https://omlx.ai/benchmarks/performance/kukxd75v",
   "tg_tokens": null,
   "tps": 79.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-kukxd75v-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: 1k0xdolc; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1810.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 1,810 71.0 14.4 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kukxd75v",
   "source_url": "https://omlx.ai/benchmarks/performance/kukxd75v",
   "tg_tokens": null,
   "tps": 71.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-cejx1ziv-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 16.5 GB at this context; duplicate traces of the same session: iqzta1yr; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2308.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,308 57.4 16.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cejx1ziv",
   "source_url": "https://omlx.ai/benchmarks/performance/cejx1ziv",
   "tg_tokens": null,
   "tps": 57.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-cejx1ziv-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 16.8 GB at this context; duplicate traces of the same session: iqzta1yr; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2032.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,032 55.3 16.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cejx1ziv",
   "source_url": "https://omlx.ai/benchmarks/performance/cejx1ziv",
   "tg_tokens": null,
   "tps": 55.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-1c8o558n-ctx4096",
   "model": "Occamy-1.0",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: aa5mpe5f; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 830.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "occamy-1.0-oQ4e on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 830.2 74.5 20.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1c8o558n",
   "source_url": "https://omlx.ai/benchmarks/performance/1c8o558n",
   "tg_tokens": null,
   "tps": 74.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-1c8o558n-ctx131072",
   "model": "Occamy-1.0",
   "notes": "oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: aa5mpe5f; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 290.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "occamy-1.0-oQ4e on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 128k 290.3 31.8 24.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1c8o558n",
   "source_url": "https://omlx.ai/benchmarks/performance/1c8o558n",
   "tg_tokens": null,
   "tps": 31.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-f0ewkm1d-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 172.6 GB at this context; duplicate traces of the same session: 325gj6hv, 817y02li, aqtpl8uq, l1pbckbq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 525.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 525.5 65.2 172.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f0ewkm1d",
   "source_url": "https://omlx.ai/benchmarks/performance/f0ewkm1d",
   "tg_tokens": null,
   "tps": 65.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-f0ewkm1d-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 173.7 GB at this context; duplicate traces of the same session: 325gj6hv, 817y02li, aqtpl8uq, l1pbckbq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 579.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 579.2 58.1 173.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f0ewkm1d",
   "source_url": "https://omlx.ai/benchmarks/performance/f0ewkm1d",
   "tg_tokens": null,
   "tps": 58.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-f0ewkm1d-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 174.1 GB at this context; duplicate traces of the same session: 325gj6hv, 817y02li, aqtpl8uq, l1pbckbq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 555.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 555.6 60.3 174.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f0ewkm1d",
   "source_url": "https://omlx.ai/benchmarks/performance/f0ewkm1d",
   "tg_tokens": null,
   "tps": 60.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-f0ewkm1d-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 174.2 GB at this context; duplicate traces of the same session: 325gj6hv, 817y02li, aqtpl8uq, l1pbckbq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 545.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 545.2 104.1 174.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f0ewkm1d",
   "source_url": "https://omlx.ai/benchmarks/performance/f0ewkm1d",
   "tg_tokens": null,
   "tps": 104.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-f0ewkm1d-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 174.3 GB at this context; duplicate traces of the same session: 325gj6hv, 817y02li, aqtpl8uq, l1pbckbq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 537.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 537.8 77.9 174.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f0ewkm1d",
   "source_url": "https://omlx.ai/benchmarks/performance/f0ewkm1d",
   "tg_tokens": null,
   "tps": 77.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-uay4dqyn-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 243.8 GB at this context; duplicate traces of the same session: 5igvqolb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 556.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "GLM-5.3-Flash-MLX-6bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 4k 556.5 35.8 243.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uay4dqyn",
   "source_url": "https://omlx.ai/benchmarks/performance/uay4dqyn",
   "tg_tokens": null,
   "tps": 35.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-uay4dqyn-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 244.3 GB at this context; duplicate traces of the same session: 5igvqolb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 521.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "GLM-5.3-Flash-MLX-6bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 16k 521.0 35.8 244.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uay4dqyn",
   "source_url": "https://omlx.ai/benchmarks/performance/uay4dqyn",
   "tg_tokens": null,
   "tps": 35.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ysdz32dp-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 249.1 GB at this context; duplicate traces of the same session: f000pp9b; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 415.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 1k 415.8 32.1 249.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ysdz32dp",
   "source_url": "https://omlx.ai/benchmarks/performance/ysdz32dp",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ysdz32dp-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 250.2 GB at this context; duplicate traces of the same session: f000pp9b; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 535.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 4k 535.0 37.3 250.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ysdz32dp",
   "source_url": "https://omlx.ai/benchmarks/performance/ysdz32dp",
   "tg_tokens": null,
   "tps": 37.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-gw3tertq-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 65.8 GB at this context; duplicate traces of the same session: 7yuumh8y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1773.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,773 70.1 65.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gw3tertq",
   "source_url": "https://omlx.ai/benchmarks/performance/gw3tertq",
   "tg_tokens": null,
   "tps": 70.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-gw3tertq-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.7 GB at this context; duplicate traces of the same session: 7yuumh8y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2305.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,305 68.6 66.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gw3tertq",
   "source_url": "https://omlx.ai/benchmarks/performance/gw3tertq",
   "tg_tokens": null,
   "tps": 68.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mrsia50p-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 65.8 GB at this context; duplicate traces of the same session: aigmaeli; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1752.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,752 70.0 65.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mrsia50p",
   "source_url": "https://omlx.ai/benchmarks/performance/mrsia50p",
   "tg_tokens": null,
   "tps": 70.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mrsia50p-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.7 GB at this context; duplicate traces of the same session: aigmaeli; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1581.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,581 68.5 66.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mrsia50p",
   "source_url": "https://omlx.ai/benchmarks/performance/mrsia50p",
   "tg_tokens": null,
   "tps": 68.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-d3pxd49f-ctx1024",
   "model": "Qwen3.5-2B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 4.4 GB; TTFT 287 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3567.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.5-2B-bf16 on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 bf16; 1k tokens 3,567 PP tok/s 202.6 TG tok/s 287 TTFT (ms) 4.4 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d3pxd49f",
   "source_url": "https://omlx.ai/benchmarks/performance/d3pxd49f",
   "tg_tokens": null,
   "tps": 202.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ryfsrecc-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: xl4aju3h; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 493.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B-8bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 8bit; 4k 493.4 30.3 12.7 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ryfsrecc",
   "source_url": "https://omlx.ai/benchmarks/performance/ryfsrecc",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ryfsrecc-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 13.1 GB at this context; duplicate traces of the same session: xl4aju3h; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 438.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B-8bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 8bit; 16k 438.8 42.1 13.1 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ryfsrecc",
   "source_url": "https://omlx.ai/benchmarks/performance/ryfsrecc",
   "tg_tokens": null,
   "tps": 42.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-21nzajir-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: cu84goud; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 701.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 701.1 107.6 19.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/21nzajir",
   "source_url": "https://omlx.ai/benchmarks/performance/21nzajir",
   "tg_tokens": null,
   "tps": 107.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-21nzajir-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: cu84goud; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1681.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 1,681 104.1 20.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/21nzajir",
   "source_url": "https://omlx.ai/benchmarks/performance/21nzajir",
   "tg_tokens": null,
   "tps": 104.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7ejnvz40-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 213 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4818.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k tokens 4,818 PP tok/s 146.9 TG tok/s 213 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ejnvz40",
   "source_url": "https://omlx.ai/benchmarks/performance/7ejnvz40",
   "tg_tokens": null,
   "tps": 146.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-o5qn1keg-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 21.0 GB; TTFT 2867 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5716.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k tokens 5,716 PP tok/s 110.4 TG tok/s 2867 TTFT (ms) 21.0 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o5qn1keg",
   "source_url": "https://omlx.ai/benchmarks/performance/o5qn1keg",
   "tg_tokens": null,
   "tps": 110.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ql5bt3s0-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 21.0 GB; TTFT 2888 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5674.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k tokens 5,674 PP tok/s 109.1 TG tok/s 2888 TTFT (ms) 21.0 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ql5bt3s0",
   "source_url": "https://omlx.ai/benchmarks/performance/ql5bt3s0",
   "tg_tokens": null,
   "tps": 109.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7aa5x8vx-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: we74j3p1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4762.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 4,762 124.5 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7aa5x8vx",
   "source_url": "https://omlx.ai/benchmarks/performance/7aa5x8vx",
   "tg_tokens": null,
   "tps": 124.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7aa5x8vx-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: we74j3p1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5189.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 5,189 118.9 20.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7aa5x8vx",
   "source_url": "https://omlx.ai/benchmarks/performance/7aa5x8vx",
   "tg_tokens": null,
   "tps": 118.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2o7yfpcq-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 21.0 GB; TTFT 2893 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5664.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k tokens 5,664 PP tok/s 107.5 TG tok/s 2893 TTFT (ms) 21.0 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2o7yfpcq",
   "source_url": "https://omlx.ai/benchmarks/performance/2o7yfpcq",
   "tg_tokens": null,
   "tps": 107.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-89p7v1pf-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 82eydjah; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3734.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 64k 3,734 79.4 21.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/89p7v1pf",
   "source_url": "https://omlx.ai/benchmarks/performance/89p7v1pf",
   "tg_tokens": null,
   "tps": 79.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-89p7v1pf-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: 82eydjah; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2340.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 128k 2,340 54.9 23.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/89p7v1pf",
   "source_url": "https://omlx.ai/benchmarks/performance/89p7v1pf",
   "tg_tokens": null,
   "tps": 54.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9ur68546-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5698.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 5,698 105.4 21.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ur68546",
   "source_url": "https://omlx.ai/benchmarks/performance/9ur68546",
   "tg_tokens": null,
   "tps": 105.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9ur68546-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4639.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 4,639 97.4 21.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ur68546",
   "source_url": "https://omlx.ai/benchmarks/performance/9ur68546",
   "tg_tokens": null,
   "tps": 97.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i6fy969m-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5681.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 5,681 109.1 21.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i6fy969m",
   "source_url": "https://omlx.ai/benchmarks/performance/i6fy969m",
   "tg_tokens": null,
   "tps": 109.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i6fy969m-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4874.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 4,874 97.3 21.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i6fy969m",
   "source_url": "https://omlx.ai/benchmarks/performance/i6fy969m",
   "tg_tokens": null,
   "tps": 97.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-3gzza82m-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 32.2 GB; TTFT 39799 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1647.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 64k tokens 1,647 PP tok/s 66.0 TG tok/s 39799 TTFT (ms) 32.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3gzza82m",
   "source_url": "https://omlx.ai/benchmarks/performance/3gzza82m",
   "tg_tokens": null,
   "tps": 66.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-y9dpje70-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 32.2 GB; TTFT 40568 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1616.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 64k tokens 1,616 PP tok/s 71.1 TG tok/s 40568 TTFT (ms) 32.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y9dpje70",
   "source_url": "https://omlx.ai/benchmarks/performance/y9dpje70",
   "tg_tokens": null,
   "tps": 71.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-n6pktxxi-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 32.2 GB; TTFT 40606 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1614.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 64k tokens 1,614 PP tok/s 66.0 TG tok/s 40606 TTFT (ms) 32.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n6pktxxi",
   "source_url": "https://omlx.ai/benchmarks/performance/n6pktxxi",
   "tg_tokens": null,
   "tps": 66.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-7cwtklfr-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 30.3 GB; TTFT 1470 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 696.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k tokens 696.5 PP tok/s 59.7 TG tok/s 1470 TTFT (ms) 30.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7cwtklfr",
   "source_url": "https://omlx.ai/benchmarks/performance/7cwtklfr",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ihd7gd3l-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: ppp1ddsj; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 965.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 965.3 98.2 37.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ihd7gd3l",
   "source_url": "https://omlx.ai/benchmarks/performance/ihd7gd3l",
   "tg_tokens": null,
   "tps": 98.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-ihd7gd3l-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: ppp1ddsj; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 893.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 893.7 57.9 39.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ihd7gd3l",
   "source_url": "https://omlx.ai/benchmarks/performance/ihd7gd3l",
   "tg_tokens": null,
   "tps": 57.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-iuicr2xt-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: c0e0etiw; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 952.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 952.7 101.5 37.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iuicr2xt",
   "source_url": "https://omlx.ai/benchmarks/performance/iuicr2xt",
   "tg_tokens": null,
   "tps": 101.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-iuicr2xt-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: c0e0etiw; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 859.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 859.2 68.3 39.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iuicr2xt",
   "source_url": "https://omlx.ai/benchmarks/performance/iuicr2xt",
   "tg_tokens": null,
   "tps": 68.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-36nb6utv-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: vrk7iy4h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 836.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 836.7 19.1 28.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/36nb6utv",
   "source_url": "https://omlx.ai/benchmarks/performance/36nb6utv",
   "tg_tokens": null,
   "tps": 19.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-36nb6utv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: vrk7iy4h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 961.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 961.8 19.0 29.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/36nb6utv",
   "source_url": "https://omlx.ai/benchmarks/performance/36nb6utv",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-v39q3q00-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 42.8 GB at this context; duplicate traces of the same session: mjpsewsh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 604.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 604.6 41.2 42.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v39q3q00",
   "source_url": "https://omlx.ai/benchmarks/performance/v39q3q00",
   "tg_tokens": null,
   "tps": 41.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-v39q3q00-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 57.4 GB at this context; duplicate traces of the same session: mjpsewsh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 495.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 195k 495.7 30.2 57.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v39q3q00",
   "source_url": "https://omlx.ai/benchmarks/performance/v39q3q00",
   "tg_tokens": null,
   "tps": 30.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-vm09lk8y-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: bf7fgpab; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 402.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 402.3 17.9 16.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vm09lk8y",
   "source_url": "https://omlx.ai/benchmarks/performance/vm09lk8y",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-vm09lk8y-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: bf7fgpab; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 419.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 419.1 17.4 17.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vm09lk8y",
   "source_url": "https://omlx.ai/benchmarks/performance/vm09lk8y",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-y2zfjmwx-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.7 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 165.8,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 165.8 19.1 30.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y2zfjmwx",
   "source_url": "https://omlx.ai/benchmarks/performance/y2zfjmwx",
   "tg_tokens": null,
   "tps": 19.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-y2zfjmwx-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.2 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 185.7,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 185.7 19.8 34.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y2zfjmwx",
   "source_url": "https://omlx.ai/benchmarks/performance/y2zfjmwx",
   "tg_tokens": null,
   "tps": 19.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-9173mxfb-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 72.5 GB at this context; duplicate traces of the same session: 513mfrl7; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 300.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 300.5 32.6 72.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9173mxfb",
   "source_url": "https://omlx.ai/benchmarks/performance/9173mxfb",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-9173mxfb-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 74.8 GB at this context; duplicate traces of the same session: 513mfrl7; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 337.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 337.6 31.1 74.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9173mxfb",
   "source_url": "https://omlx.ai/benchmarks/performance/9173mxfb",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lukg0dgn-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.7 GB; TTFT 1424 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 719.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k tokens 719.3 PP tok/s 59.6 TG tok/s 1424 TTFT (ms) 70.7 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lukg0dgn",
   "source_url": "https://omlx.ai/benchmarks/performance/lukg0dgn",
   "tg_tokens": null,
   "tps": 59.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-6du6jf6h-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 69.3 GB; TTFT 1624 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 630.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k tokens 630.4 PP tok/s 56.8 TG tok/s 1624 TTFT (ms) 69.3 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6du6jf6h",
   "source_url": "https://omlx.ai/benchmarks/performance/6du6jf6h",
   "tg_tokens": null,
   "tps": 56.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-puun46gv-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 72.8 GB; TTFT 6796 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2411.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k tokens 2,411 PP tok/s 74.9 TG tok/s 6796 TTFT (ms) 72.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/puun46gv",
   "source_url": "https://omlx.ai/benchmarks/performance/puun46gv",
   "tg_tokens": null,
   "tps": 74.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mm33zj62-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: vak1y3ul; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 336.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 336.9 48.0 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mm33zj62",
   "source_url": "https://omlx.ai/benchmarks/performance/mm33zj62",
   "tg_tokens": null,
   "tps": 48.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mm33zj62-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: vak1y3ul; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 985.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 985.2 53.9 71.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mm33zj62",
   "source_url": "https://omlx.ai/benchmarks/performance/mm33zj62",
   "tg_tokens": null,
   "tps": 53.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ab1u7f32-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace, single run at 16k context; peak memory 143.9 GB; TTFT 11993 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1366.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 16k tokens 1,366 PP tok/s 96.1 TG tok/s 11993 TTFT (ms) 143.9 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ab1u7f32",
   "source_url": "https://omlx.ai/benchmarks/performance/ab1u7f32",
   "tg_tokens": null,
   "tps": 96.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-a1rf95aw-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 195k context; peak memory 137.9 GB; TTFT 51335 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3896.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 195k tokens 3,896 PP tok/s 61.5 TG tok/s 51335 TTFT (ms) 137.9 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a1rf95aw",
   "source_url": "https://omlx.ai/benchmarks/performance/a1rf95aw",
   "tg_tokens": null,
   "tps": 61.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-65mx66zt-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 195k context; peak memory 138.0 GB; TTFT 50032 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3998.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 195k tokens 3,998 PP tok/s 71.5 TG tok/s 50032 TTFT (ms) 138.0 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/65mx66zt",
   "source_url": "https://omlx.ai/benchmarks/performance/65mx66zt",
   "tg_tokens": null,
   "tps": 71.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-r7uouu60-ctx1024",
   "model": "Gemma-4-31B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 6326o0gl, 78qumtmv, m66dpvby, oxq2v7hh; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 899.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31b-it-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k 899.5 42.3 17.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r7uouu60",
   "source_url": "https://omlx.ai/benchmarks/performance/r7uouu60",
   "tg_tokens": null,
   "tps": 42.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-r7uouu60-ctx4096",
   "model": "Gemma-4-31B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 6326o0gl, 78qumtmv, m66dpvby, oxq2v7hh; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 958.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31b-it-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 4k 958.4 39.3 18.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r7uouu60",
   "source_url": "https://omlx.ai/benchmarks/performance/r7uouu60",
   "tg_tokens": null,
   "tps": 39.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-r7uouu60-ctx65536",
   "model": "Gemma-4-31B",
   "notes": "oMLX community benchmark trace; peak memory 27.2 GB at this context; duplicate traces of the same session: 6326o0gl, 78qumtmv, m66dpvby, oxq2v7hh; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 799.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31b-it-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 64k 799.4 17.4 27.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r7uouu60",
   "source_url": "https://omlx.ai/benchmarks/performance/r7uouu60",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-r7uouu60-ctx131072",
   "model": "Gemma-4-31B",
   "notes": "oMLX community benchmark trace; peak memory 36.3 GB at this context; duplicate traces of the same session: 6326o0gl, 78qumtmv, m66dpvby, oxq2v7hh; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 626.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31b-it-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 128k 626.1 10.8 36.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r7uouu60",
   "source_url": "https://omlx.ai/benchmarks/performance/r7uouu60",
   "tg_tokens": null,
   "tps": 10.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-r7uouu60-ctx199680",
   "model": "Gemma-4-31B",
   "notes": "oMLX community benchmark trace; peak memory 37.1 GB at this context; duplicate traces of the same session: 6326o0gl, 78qumtmv, m66dpvby, oxq2v7hh; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 513.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31b-it-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 195k 513.5 7.7 37.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r7uouu60",
   "source_url": "https://omlx.ai/benchmarks/performance/r7uouu60",
   "tg_tokens": null,
   "tps": 7.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-5g583554-ctx1024",
   "model": "gemma-4-E4B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 6.0 GB at this context; duplicate traces of the same session: ys51w3i4; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4928.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "gemma-4-E4B-it-qat-mxfp4 on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k 4,928 126.4 6.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5g583554",
   "source_url": "https://omlx.ai/benchmarks/performance/5g583554",
   "tg_tokens": null,
   "tps": 126.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-5g583554-ctx199680",
   "model": "gemma-4-E4B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 10.6 GB at this context; duplicate traces of the same session: ys51w3i4; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4955.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "gemma-4-E4B-it-qat-mxfp4 on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 195k 4,955 25.6 10.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5g583554",
   "source_url": "https://omlx.ai/benchmarks/performance/5g583554",
   "tg_tokens": null,
   "tps": 25.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-cpsqlm10-ctx4096",
   "model": "AREX-Turbo",
   "notes": "oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: lx4abx28; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1437.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "AREX-Turbo-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,437 110.1 3.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cpsqlm10",
   "source_url": "https://omlx.ai/benchmarks/performance/cpsqlm10",
   "tg_tokens": null,
   "tps": 110.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-cpsqlm10-ctx16384",
   "model": "AREX-Turbo",
   "notes": "oMLX community benchmark trace; peak memory 4.3 GB at this context; duplicate traces of the same session: lx4abx28; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1327.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "AREX-Turbo-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 1,327 101.8 4.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cpsqlm10",
   "source_url": "https://omlx.ai/benchmarks/performance/cpsqlm10",
   "tg_tokens": null,
   "tps": 101.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-pxt80a7d-ctx1024",
   "model": "AREX-Turbo",
   "notes": "oMLX community benchmark trace; peak memory 5.3 GB at this context; duplicate traces of the same session: 0xczanl6, x02ntk9y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 923.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "AREX-Turbo-8bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 8bit; 1k 923.5 60.8 5.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pxt80a7d",
   "source_url": "https://omlx.ai/benchmarks/performance/pxt80a7d",
   "tg_tokens": null,
   "tps": 60.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-pxt80a7d-ctx4096",
   "model": "AREX-Turbo",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: 0xczanl6, x02ntk9y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1012.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "AREX-Turbo-8bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 8bit; 4k 1,012 58.8 5.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pxt80a7d",
   "source_url": "https://omlx.ai/benchmarks/performance/pxt80a7d",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-pxt80a7d-ctx16384",
   "model": "AREX-Turbo",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: 0xczanl6, x02ntk9y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 958.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "AREX-Turbo-8bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 8bit; 16k 958.8 55.1 6.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pxt80a7d",
   "source_url": "https://omlx.ai/benchmarks/performance/pxt80a7d",
   "tg_tokens": null,
   "tps": 55.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-hsciq84t-ctx1024",
   "model": "Agents-A1",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 8.8 GB at this context; duplicate traces of the same session: 7hyxs4x2, 8hg3s1xc, rmt454ws, t5rfoobx, ztlhp6yg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1595.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Agents-A1-4B-mlx on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 unknown; 1k 1,595 52.1 8.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hsciq84t",
   "source_url": "https://omlx.ai/benchmarks/performance/hsciq84t",
   "tg_tokens": null,
   "tps": 52.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-hsciq84t-ctx4096",
   "model": "Agents-A1",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 10.2 GB at this context; duplicate traces of the same session: 7hyxs4x2, 8hg3s1xc, rmt454ws, t5rfoobx, ztlhp6yg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1841.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Agents-A1-4B-mlx on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 unknown; 4k 1,841 51.7 10.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hsciq84t",
   "source_url": "https://omlx.ai/benchmarks/performance/hsciq84t",
   "tg_tokens": null,
   "tps": 51.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-hsciq84t-ctx8192",
   "model": "Agents-A1",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 10.3 GB at this context; duplicate traces of the same session: 7hyxs4x2, 8hg3s1xc, rmt454ws, t5rfoobx, ztlhp6yg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1751.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Agents-A1-4B-mlx on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 unknown; 8k 1,751 50.9 10.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hsciq84t",
   "source_url": "https://omlx.ai/benchmarks/performance/hsciq84t",
   "tg_tokens": null,
   "tps": 50.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-hsciq84t-ctx16384",
   "model": "Agents-A1",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 10.7 GB at this context; duplicate traces of the same session: 7hyxs4x2, 8hg3s1xc, rmt454ws, t5rfoobx, ztlhp6yg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1530.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Agents-A1-4B-mlx on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 unknown; 16k 1,530 49.2 10.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hsciq84t",
   "source_url": "https://omlx.ai/benchmarks/performance/hsciq84t",
   "tg_tokens": null,
   "tps": 49.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-hsciq84t-ctx32768",
   "model": "Agents-A1",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 11.6 GB at this context; duplicate traces of the same session: 7hyxs4x2, 8hg3s1xc, rmt454ws, t5rfoobx, ztlhp6yg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1222.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Agents-A1-4B-mlx on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 unknown; 32k 1,222 45.4 11.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hsciq84t",
   "source_url": "https://omlx.ai/benchmarks/performance/hsciq84t",
   "tg_tokens": null,
   "tps": 45.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-hsciq84t-ctx65536",
   "model": "Agents-A1",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.2 GB at this context; duplicate traces of the same session: 7hyxs4x2, 8hg3s1xc, rmt454ws, t5rfoobx, ztlhp6yg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 946.4,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Agents-A1-4B-mlx on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 unknown; 64k 946.4 39.8 13.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hsciq84t",
   "source_url": "https://omlx.ai/benchmarks/performance/hsciq84t",
   "tg_tokens": null,
   "tps": 39.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-55se1n9i-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 177.6 GB; TTFT 2571 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 398.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k tokens 398.4 PP tok/s 45.8 TG tok/s 2571 TTFT (ms) 177.6 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/55se1n9i",
   "source_url": "https://omlx.ai/benchmarks/performance/55se1n9i",
   "tg_tokens": null,
   "tps": 45.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-i4xfbfve-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Huihui abliterated build; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 0efdm2tm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 355.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.8-Flash-Next-abliterated-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 355.6 45.5 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i4xfbfve",
   "source_url": "https://omlx.ai/benchmarks/performance/i4xfbfve",
   "tg_tokens": null,
   "tps": 45.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-i4xfbfve-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Huihui abliterated build; oMLX community benchmark trace; peak memory 72.5 GB at this context; duplicate traces of the same session: 0efdm2tm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 429.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.8-Flash-Next-abliterated-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 429.5 51.7 72.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i4xfbfve",
   "source_url": "https://omlx.ai/benchmarks/performance/i4xfbfve",
   "tg_tokens": null,
   "tps": 51.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-29hb2p48-ctx4096",
   "model": "Mistral-Small-24B-Instruct-2501",
   "notes": "oMLX community benchmark trace; peak memory 13.9 GB at this context; duplicate traces of the same session: nqo3ma7k; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 878.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Mistral-Small-24B-Instruct-2501-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 878.7 30.0 13.9 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/29hb2p48",
   "source_url": "https://omlx.ai/benchmarks/performance/29hb2p48",
   "tg_tokens": null,
   "tps": 30.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-29hb2p48-ctx16384",
   "model": "Mistral-Small-24B-Instruct-2501",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: nqo3ma7k; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 854.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Mistral-Small-24B-Instruct-2501-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 854.3 26.0 15.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/29hb2p48",
   "source_url": "https://omlx.ai/benchmarks/performance/29hb2p48",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-m1qwggaq-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: p3xqs5du; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5648.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 5,648 106.0 21.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m1qwggaq",
   "source_url": "https://omlx.ai/benchmarks/performance/m1qwggaq",
   "tg_tokens": null,
   "tps": 106.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-m1qwggaq-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: p3xqs5du; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4879.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 4,879 97.5 21.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m1qwggaq",
   "source_url": "https://omlx.ai/benchmarks/performance/m1qwggaq",
   "tg_tokens": null,
   "tps": 97.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-l3noeq7w-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5698.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 5,698 105.4 21.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l3noeq7w",
   "source_url": "https://omlx.ai/benchmarks/performance/l3noeq7w",
   "tg_tokens": null,
   "tps": 105.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-l3noeq7w-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4639.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 4,639 97.4 21.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l3noeq7w",
   "source_url": "https://omlx.ai/benchmarks/performance/l3noeq7w",
   "tg_tokens": null,
   "tps": 97.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-abs68es1-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5681.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 5,681 109.1 21.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/abs68es1",
   "source_url": "https://omlx.ai/benchmarks/performance/abs68es1",
   "tg_tokens": null,
   "tps": 109.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-abs68es1-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4874.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 4,874 97.3 21.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/abs68es1",
   "source_url": "https://omlx.ai/benchmarks/performance/abs68es1",
   "tg_tokens": null,
   "tps": 97.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vilh2sux-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 40.1 GB at this context; duplicate traces of the same session: ikbot9bk, prkvc2kb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2339.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 32k 2,339 68.9 40.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vilh2sux",
   "source_url": "https://omlx.ai/benchmarks/performance/vilh2sux",
   "tg_tokens": null,
   "tps": 68.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vilh2sux-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 41.2 GB at this context; duplicate traces of the same session: ikbot9bk, prkvc2kb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1732.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 64k 1,732 70.0 41.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vilh2sux",
   "source_url": "https://omlx.ai/benchmarks/performance/vilh2sux",
   "tg_tokens": null,
   "tps": 70.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vilh2sux-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 43.6 GB at this context; duplicate traces of the same session: ikbot9bk, prkvc2kb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1128.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 128k 1,128 75.1 43.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vilh2sux",
   "source_url": "https://omlx.ai/benchmarks/performance/vilh2sux",
   "tg_tokens": null,
   "tps": 75.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-887200ed-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: kmph055l; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 421.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 421.7 16.2 16.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/887200ed",
   "source_url": "https://omlx.ai/benchmarks/performance/887200ed",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-887200ed-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: kmph055l; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 447.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 447.6 16.3 18.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/887200ed",
   "source_url": "https://omlx.ai/benchmarks/performance/887200ed",
   "tg_tokens": null,
   "tps": 16.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-xs7yxc08-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: ou1c1ssf; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 409.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 409.3 17.2 16.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xs7yxc08",
   "source_url": "https://omlx.ai/benchmarks/performance/xs7yxc08",
   "tg_tokens": null,
   "tps": 17.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-xs7yxc08-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: ou1c1ssf; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 414.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 414.5 13.5 18.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xs7yxc08",
   "source_url": "https://omlx.ai/benchmarks/performance/xs7yxc08",
   "tg_tokens": null,
   "tps": 13.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-a0ne2onb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: x6xokkhh; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 408.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 408.9 15.0 16.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a0ne2onb",
   "source_url": "https://omlx.ai/benchmarks/performance/a0ne2onb",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-a0ne2onb-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: x6xokkhh; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 403.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 403.8 12.8 18.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a0ne2onb",
   "source_url": "https://omlx.ai/benchmarks/performance/a0ne2onb",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-vmzs93ul-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 18.8 GB; TTFT 37177 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 440.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k tokens 440.7 PP tok/s 15.7 TG tok/s 37177 TTFT (ms) 18.8 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vmzs93ul",
   "source_url": "https://omlx.ai/benchmarks/performance/vmzs93ul",
   "tg_tokens": null,
   "tps": 15.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-955tulsj-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: bl7x1s6g; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 415.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 415.1 16.3 16.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/955tulsj",
   "source_url": "https://omlx.ai/benchmarks/performance/955tulsj",
   "tg_tokens": null,
   "tps": 16.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-955tulsj-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: bl7x1s6g; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 410.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 410.9 14.3 18.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/955tulsj",
   "source_url": "https://omlx.ai/benchmarks/performance/955tulsj",
   "tg_tokens": null,
   "tps": 14.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9or8y035-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 24.7 GB; TTFT 20119 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 814.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 6bit; 16k tokens 814.4 PP tok/s 21.8 TG tok/s 20119 TTFT (ms) 24.7 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9or8y035",
   "source_url": "https://omlx.ai/benchmarks/performance/9or8y035",
   "tg_tokens": null,
   "tps": 21.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-14uma95y-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: hpzucw0o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 852.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 6bit; 1k 852.7 24.5 22.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/14uma95y",
   "source_url": "https://omlx.ai/benchmarks/performance/14uma95y",
   "tg_tokens": null,
   "tps": 24.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-14uma95y-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: hpzucw0o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 910.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 6bit; 4k 910.8 17.3 23.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/14uma95y",
   "source_url": "https://omlx.ai/benchmarks/performance/14uma95y",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-2ebpmcg1-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: 04138k5h, ul7ms2mj; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1144.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 1k 1,144 79.8 33.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ebpmcg1",
   "source_url": "https://omlx.ai/benchmarks/performance/2ebpmcg1",
   "tg_tokens": null,
   "tps": 79.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-2ebpmcg1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: 04138k5h, ul7ms2mj; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1374.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 4k 1,374 85.5 34.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ebpmcg1",
   "source_url": "https://omlx.ai/benchmarks/performance/2ebpmcg1",
   "tg_tokens": null,
   "tps": 85.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-2ebpmcg1-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.7 GB at this context; duplicate traces of the same session: 04138k5h, ul7ms2mj; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1355.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 16k 1,355 79.3 35.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ebpmcg1",
   "source_url": "https://omlx.ai/benchmarks/performance/2ebpmcg1",
   "tg_tokens": null,
   "tps": 79.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-2ebpmcg1-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.6 GB at this context; duplicate traces of the same session: 04138k5h, ul7ms2mj; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 875.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 128k 875.8 58.9 43.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ebpmcg1",
   "source_url": "https://omlx.ai/benchmarks/performance/2ebpmcg1",
   "tg_tokens": null,
   "tps": 58.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-zuzaag0o-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 122.2 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2167.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 5bit; 1k 2,167 135.0 122.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zuzaag0o",
   "source_url": "https://omlx.ai/benchmarks/performance/zuzaag0o",
   "tg_tokens": null,
   "tps": 135.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-zuzaag0o-ctx8192",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 124.1 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4216.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 5bit; 8k 4,216 143.8 124.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zuzaag0o",
   "source_url": "https://omlx.ai/benchmarks/performance/zuzaag0o",
   "tg_tokens": null,
   "tps": 143.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-zhfnlcil-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 195k context; peak memory 137.9 GB; TTFT 48355 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4136.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 195k tokens 4,136 PP tok/s 77.4 TG tok/s 48355 TTFT (ms) 137.9 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zhfnlcil",
   "source_url": "https://omlx.ai/benchmarks/performance/zhfnlcil",
   "tg_tokens": null,
   "tps": 77.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-7vh1x4sc-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 135.5 GB; TTFT 32174 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4074.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 128k tokens 4,074 PP tok/s 94.8 TG tok/s 32174 TTFT (ms) 135.5 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7vh1x4sc",
   "source_url": "https://omlx.ai/benchmarks/performance/7vh1x4sc",
   "tg_tokens": null,
   "tps": 94.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-0xor97x9-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 195k context; peak memory 138.0 GB; TTFT 48545 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4120.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 195k tokens 4,120 PP tok/s 70.2 TG tok/s 48545 TTFT (ms) 138.0 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0xor97x9",
   "source_url": "https://omlx.ai/benchmarks/performance/0xor97x9",
   "tg_tokens": null,
   "tps": 70.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-53k91egg-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 12.5 GB at this context; duplicate traces of the same session: 9qlmjhrm, da1pzi73, mvuon3df; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 433.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 433.3 31.3 12.5 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/53k91egg",
   "source_url": "https://omlx.ai/benchmarks/performance/53k91egg",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-53k91egg-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 12.8 GB at this context; duplicate traces of the same session: 9qlmjhrm, da1pzi73, mvuon3df; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1013.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 1,013 30.9 12.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/53k91egg",
   "source_url": "https://omlx.ai/benchmarks/performance/53k91egg",
   "tg_tokens": null,
   "tps": 30.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-53k91egg-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: 9qlmjhrm, da1pzi73, mvuon3df; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 926.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 16k 926.9 25.6 13.4 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/53k91egg",
   "source_url": "https://omlx.ai/benchmarks/performance/53k91egg",
   "tg_tokens": null,
   "tps": 25.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-53k91egg-ctx32768",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: 9qlmjhrm, da1pzi73, mvuon3df; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 798.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 32k 798.2 23.3 14.3 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/53k91egg",
   "source_url": "https://omlx.ai/benchmarks/performance/53k91egg",
   "tg_tokens": null,
   "tps": 23.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-5bvky56q-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 12.5 GB at this context; duplicate traces of the same session: bnvutqgd, epqm6xjy, m8qb0068; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 476.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 476.1 31.2 12.5 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5bvky56q",
   "source_url": "https://omlx.ai/benchmarks/performance/5bvky56q",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-5bvky56q-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 12.8 GB at this context; duplicate traces of the same session: bnvutqgd, epqm6xjy, m8qb0068; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1019.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 1,019 29.4 12.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5bvky56q",
   "source_url": "https://omlx.ai/benchmarks/performance/5bvky56q",
   "tg_tokens": null,
   "tps": 29.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-5bvky56q-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: bnvutqgd, epqm6xjy, m8qb0068; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 911.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 16k 911.7 26.3 13.4 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5bvky56q",
   "source_url": "https://omlx.ai/benchmarks/performance/5bvky56q",
   "tg_tokens": null,
   "tps": 26.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-5bvky56q-ctx32768",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: bnvutqgd, epqm6xjy, m8qb0068; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 783.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 32k 783.2 22.5 14.3 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5bvky56q",
   "source_url": "https://omlx.ai/benchmarks/performance/5bvky56q",
   "tg_tokens": null,
   "tps": 22.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-9z356qty-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: 0yitxp1b, 2nsit8sh, 8hy4bydd; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 428.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 428.3 31.0 11.8 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9z356qty",
   "source_url": "https://omlx.ai/benchmarks/performance/9z356qty",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-9z356qty-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 12.1 GB at this context; duplicate traces of the same session: 0yitxp1b, 2nsit8sh, 8hy4bydd; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 949.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 949.8 30.6 12.1 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9z356qty",
   "source_url": "https://omlx.ai/benchmarks/performance/9z356qty",
   "tg_tokens": null,
   "tps": 30.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-9z356qty-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 12.8 GB at this context; duplicate traces of the same session: 0yitxp1b, 2nsit8sh, 8hy4bydd; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 880.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 16k 880.8 26.6 12.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9z356qty",
   "source_url": "https://omlx.ai/benchmarks/performance/9z356qty",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-9z356qty-ctx32768",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 13.6 GB at this context; duplicate traces of the same session: 0yitxp1b, 2nsit8sh, 8hy4bydd; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 760.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 32k 760.9 23.1 13.6 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9z356qty",
   "source_url": "https://omlx.ai/benchmarks/performance/9z356qty",
   "tg_tokens": null,
   "tps": 23.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-3a6468xm-ctx4096",
   "model": "Occamy-1.0",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: iis8u2h5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 803.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "occamy-1.0-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 803.4 78.2 21.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3a6468xm",
   "source_url": "https://omlx.ai/benchmarks/performance/3a6468xm",
   "tg_tokens": null,
   "tps": 78.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-3a6468xm-ctx32768",
   "model": "Occamy-1.0",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: iis8u2h5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 554.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "occamy-1.0-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 554.4 59.0 22.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3a6468xm",
   "source_url": "https://omlx.ai/benchmarks/performance/3a6468xm",
   "tg_tokens": null,
   "tps": 59.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-iowcriai-ctx4096",
   "model": "Occamy-1.0",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: hnzfkf99; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 779.5,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "occamy-1.0-oQ5e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 5bit; 4k 779.5 63.2 27.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iowcriai",
   "source_url": "https://omlx.ai/benchmarks/performance/iowcriai",
   "tg_tokens": null,
   "tps": 63.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-iowcriai-ctx32768",
   "model": "Occamy-1.0",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: hnzfkf99; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 541.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "occamy-1.0-oQ5e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 5bit; 32k 541.7 72.1 28.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iowcriai",
   "source_url": "https://omlx.ai/benchmarks/performance/iowcriai",
   "tg_tokens": null,
   "tps": 72.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ahl4uyfs-ctx32768",
   "model": "Occamy-1.0",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 28.6 GB; TTFT 60340 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 543.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "occamy-1.0-oQ5e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 5bit; 32k tokens 543.1 PP tok/s 70.9 TG tok/s 60340 TTFT (ms) 28.6 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ahl4uyfs",
   "source_url": "https://omlx.ai/benchmarks/performance/ahl4uyfs",
   "tg_tokens": null,
   "tps": 70.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-jc5eec5j-ctx4096",
   "model": "Dirk-Qwen3.8-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 36s9zo6z; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 399.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Dirk-Qwen3___8-27B-MLX-oQ4e on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 399.7 16.5 17.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jc5eec5j",
   "source_url": "https://omlx.ai/benchmarks/performance/jc5eec5j",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-jc5eec5j-ctx16384",
   "model": "Dirk-Qwen3.8-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: 36s9zo6z; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 366.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Dirk-Qwen3___8-27B-MLX-oQ4e on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 16k 366.7 16.0 18.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jc5eec5j",
   "source_url": "https://omlx.ai/benchmarks/performance/jc5eec5j",
   "tg_tokens": null,
   "tps": 16.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-vr136ea7-ctx1024",
   "model": "GLM-4.7-Flash",
   "notes": "oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: 3bteae6l, j7c9svim, m7yttmp8; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 674.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-4.7-Flash-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 1k 674.7 49.1 16.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vr136ea7",
   "source_url": "https://omlx.ai/benchmarks/performance/vr136ea7",
   "tg_tokens": null,
   "tps": 49.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-vr136ea7-ctx4096",
   "model": "GLM-4.7-Flash",
   "notes": "oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 3bteae6l, j7c9svim, m7yttmp8; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 742.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-4.7-Flash-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 4k 742.2 45.2 17.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vr136ea7",
   "source_url": "https://omlx.ai/benchmarks/performance/vr136ea7",
   "tg_tokens": null,
   "tps": 45.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-vr136ea7-ctx8192",
   "model": "GLM-4.7-Flash",
   "notes": "oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 3bteae6l, j7c9svim, m7yttmp8; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 634.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-4.7-Flash-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 8k 634.5 39.2 19.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vr136ea7",
   "source_url": "https://omlx.ai/benchmarks/performance/vr136ea7",
   "tg_tokens": null,
   "tps": 39.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-vr136ea7-ctx16384",
   "model": "GLM-4.7-Flash",
   "notes": "oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 3bteae6l, j7c9svim, m7yttmp8; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 472.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-4.7-Flash-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 16k 472.0 32.7 20.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vr136ea7",
   "source_url": "https://omlx.ai/benchmarks/performance/vr136ea7",
   "tg_tokens": null,
   "tps": 32.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-nnwrdzb9-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.1 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2577.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 2,577 85.5 13.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nnwrdzb9",
   "source_url": "https://omlx.ai/benchmarks/performance/nnwrdzb9",
   "tg_tokens": null,
   "tps": 85.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-nnwrdzb9-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.5 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2982.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 2,982 82.3 13.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nnwrdzb9",
   "source_url": "https://omlx.ai/benchmarks/performance/nnwrdzb9",
   "tg_tokens": null,
   "tps": 82.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-32epgta3-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: rq1a993d; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3387.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-6bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 6bit; 1k 3,387 127.4 27.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/32epgta3",
   "source_url": "https://omlx.ai/benchmarks/performance/32epgta3",
   "tg_tokens": null,
   "tps": 127.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-32epgta3-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 30.6 GB at this context; duplicate traces of the same session: rq1a993d; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3752.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-6bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 6bit; 128k 3,752 92.7 30.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/32epgta3",
   "source_url": "https://omlx.ai/benchmarks/performance/32epgta3",
   "tg_tokens": null,
   "tps": 92.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-xehnn9zv-ctx1024",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: 1fjknplp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 577.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 577.5 79.9 12.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xehnn9zv",
   "source_url": "https://omlx.ai/benchmarks/performance/xehnn9zv",
   "tg_tokens": null,
   "tps": 79.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-xehnn9zv-ctx4096",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.3 GB at this context; duplicate traces of the same session: 1fjknplp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1503.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 1,503 77.7 13.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xehnn9zv",
   "source_url": "https://omlx.ai/benchmarks/performance/xehnn9zv",
   "tg_tokens": null,
   "tps": 77.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-s8qe22un-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: cix6j3b1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1506.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 1,506 78.0 37.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s8qe22un",
   "source_url": "https://omlx.ai/benchmarks/performance/s8qe22un",
   "tg_tokens": null,
   "tps": 78.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-s8qe22un-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.1 GB at this context; duplicate traces of the same session: cix6j3b1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3270.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 3,270 154.7 39.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s8qe22un",
   "source_url": "https://omlx.ai/benchmarks/performance/s8qe22un",
   "tg_tokens": null,
   "tps": 154.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-j0a4gwbb-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 18.3 GB; TTFT 18805 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 217.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k tokens 217.8 PP tok/s 14.1 TG tok/s 18805 TTFT (ms) 18.3 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j0a4gwbb",
   "source_url": "https://omlx.ai/benchmarks/performance/j0a4gwbb",
   "tg_tokens": null,
   "tps": 14.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-ja6gwohu-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 18.3 GB; TTFT 18354 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 223.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k tokens 223.2 PP tok/s 15.6 TG tok/s 18354 TTFT (ms) 18.3 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ja6gwohu",
   "source_url": "https://omlx.ai/benchmarks/performance/ja6gwohu",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-42ljj3so-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 18.3 GB; TTFT 18109 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 226.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k tokens 226.2 PP tok/s 18.8 TG tok/s 18109 TTFT (ms) 18.3 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/42ljj3so",
   "source_url": "https://omlx.ai/benchmarks/performance/42ljj3so",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-9710qbmx-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: bjytm1rs; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 405.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 405.6 14.5 16.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9710qbmx",
   "source_url": "https://omlx.ai/benchmarks/performance/9710qbmx",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-9710qbmx-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: bjytm1rs; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 426.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 426.9 15.1 18.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9710qbmx",
   "source_url": "https://omlx.ai/benchmarks/performance/9710qbmx",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-tp7mu4cm-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: pm73tuv3; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 394.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 394.6 9.7 16.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tp7mu4cm",
   "source_url": "https://omlx.ai/benchmarks/performance/tp7mu4cm",
   "tg_tokens": null,
   "tps": 9.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-tp7mu4cm-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: pm73tuv3; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 443.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 443.7 9.2 19.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tp7mu4cm",
   "source_url": "https://omlx.ai/benchmarks/performance/tp7mu4cm",
   "tg_tokens": null,
   "tps": 9.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-3oul8c2s-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: uvkipxdj; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 298.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 298.4 9.8 15.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3oul8c2s",
   "source_url": "https://omlx.ai/benchmarks/performance/3oul8c2s",
   "tg_tokens": null,
   "tps": 9.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-3oul8c2s-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: uvkipxdj; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 288.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 288.6 9.3 17.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3oul8c2s",
   "source_url": "https://omlx.ai/benchmarks/performance/3oul8c2s",
   "tg_tokens": null,
   "tps": 9.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-31czmnvz-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: laapwiky; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 507.9,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-27B-nvfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 507.9 17.6 17.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/31czmnvz",
   "source_url": "https://omlx.ai/benchmarks/performance/31czmnvz",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-31czmnvz-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: laapwiky; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 492.2,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-27B-nvfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 492.2 16.6 17.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/31czmnvz",
   "source_url": "https://omlx.ai/benchmarks/performance/31czmnvz",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-7nrnm7a0-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: v334i1o9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 253.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 253.1 53.5 21.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7nrnm7a0",
   "source_url": "https://omlx.ai/benchmarks/performance/7nrnm7a0",
   "tg_tokens": null,
   "tps": 53.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-7nrnm7a0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: v334i1o9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 324.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 324.4 59.7 25.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7nrnm7a0",
   "source_url": "https://omlx.ai/benchmarks/performance/7nrnm7a0",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-otrqhh9b-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 32.0 GB; TTFT 104817 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 625.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 6bit; 64k tokens 625.2 PP tok/s 36.6 TG tok/s 104817 TTFT (ms) 32.0 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/otrqhh9b",
   "source_url": "https://omlx.ai/benchmarks/performance/otrqhh9b",
   "tg_tokens": null,
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-e4gky8pm-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 29.8 GB; TTFT 45947 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 713.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 6bit; 32k tokens 713.2 PP tok/s 57.0 TG tok/s 45947 TTFT (ms) 29.8 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e4gky8pm",
   "source_url": "https://omlx.ai/benchmarks/performance/e4gky8pm",
   "tg_tokens": null,
   "tps": 57.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wlyddqsb-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 28.7 GB; TTFT 20489 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 799.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 6bit; 16k tokens 799.7 PP tok/s 49.0 TG tok/s 20489 TTFT (ms) 28.7 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wlyddqsb",
   "source_url": "https://omlx.ai/benchmarks/performance/wlyddqsb",
   "tg_tokens": null,
   "tps": 49.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-jwnnz2pc-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: h4brwol0, n7h8a5lq, phi8mg9e, yqq2jlvw; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 753.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 1k 753.6 31.5 26.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jwnnz2pc",
   "source_url": "https://omlx.ai/benchmarks/performance/jwnnz2pc",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-jwnnz2pc-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: h4brwol0, n7h8a5lq, phi8mg9e, yqq2jlvw; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1308.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 4k 1,308 42.3 27.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jwnnz2pc",
   "source_url": "https://omlx.ai/benchmarks/performance/jwnnz2pc",
   "tg_tokens": null,
   "tps": 42.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-jwnnz2pc-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: h4brwol0, n7h8a5lq, phi8mg9e, yqq2jlvw; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1250.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 32k 1,250 45.8 29.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jwnnz2pc",
   "source_url": "https://omlx.ai/benchmarks/performance/jwnnz2pc",
   "tg_tokens": null,
   "tps": 45.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-jwnnz2pc-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: h4brwol0, n7h8a5lq, phi8mg9e, yqq2jlvw; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1099.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 64k 1,099 48.4 32.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jwnnz2pc",
   "source_url": "https://omlx.ai/benchmarks/performance/jwnnz2pc",
   "tg_tokens": null,
   "tps": 48.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-jwnnz2pc-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.8 GB at this context; duplicate traces of the same session: h4brwol0, n7h8a5lq, phi8mg9e, yqq2jlvw; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 873.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 128k 873.2 44.2 36.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jwnnz2pc",
   "source_url": "https://omlx.ai/benchmarks/performance/jwnnz2pc",
   "tg_tokens": null,
   "tps": 44.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-tjjy9uzu-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1144.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 1k 1,144 79.8 33.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tjjy9uzu",
   "source_url": "https://omlx.ai/benchmarks/performance/tjjy9uzu",
   "tg_tokens": null,
   "tps": 79.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-tjjy9uzu-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1374.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 4k 1,374 85.5 34.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tjjy9uzu",
   "source_url": "https://omlx.ai/benchmarks/performance/tjjy9uzu",
   "tg_tokens": null,
   "tps": 85.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-tjjy9uzu-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.7 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1355.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 16k 1,355 79.3 35.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tjjy9uzu",
   "source_url": "https://omlx.ai/benchmarks/performance/tjjy9uzu",
   "tg_tokens": null,
   "tps": 79.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-tjjy9uzu-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.6 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 875.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 128k 875.8 58.9 43.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tjjy9uzu",
   "source_url": "https://omlx.ai/benchmarks/performance/tjjy9uzu",
   "tg_tokens": null,
   "tps": 58.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-r29i5o58-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 122.2 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2167.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 5bit; 1k 2,167 135.0 122.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r29i5o58",
   "source_url": "https://omlx.ai/benchmarks/performance/r29i5o58",
   "tg_tokens": null,
   "tps": 135.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-r29i5o58-ctx8192",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 124.1 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4216.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 5bit; 8k 4,216 143.8 124.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r29i5o58",
   "source_url": "https://omlx.ai/benchmarks/performance/r29i5o58",
   "tg_tokens": null,
   "tps": 143.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-sahpaupz-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 72.4 GB at this context; duplicate traces of the same session: ji0v1p80; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 234.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 234.1 28.3 72.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sahpaupz",
   "source_url": "https://omlx.ai/benchmarks/performance/sahpaupz",
   "tg_tokens": null,
   "tps": 28.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-sahpaupz-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 74.8 GB at this context; duplicate traces of the same session: ji0v1p80; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 303.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 303.9 30.5 74.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sahpaupz",
   "source_url": "https://omlx.ai/benchmarks/performance/sahpaupz",
   "tg_tokens": null,
   "tps": 30.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-0ek1yfi9-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 119.3 GB at this context; duplicate traces of the same session: 3l2kux00, atgoqjl1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1140.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 1k 1,140 96.2 119.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ek1yfi9",
   "source_url": "https://omlx.ai/benchmarks/performance/0ek1yfi9",
   "tg_tokens": null,
   "tps": 96.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0ek1yfi9-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 119.8 GB at this context; duplicate traces of the same session: 3l2kux00, atgoqjl1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1406.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 4k 1,406 112.7 119.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ek1yfi9",
   "source_url": "https://omlx.ai/benchmarks/performance/0ek1yfi9",
   "tg_tokens": null,
   "tps": 112.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-0ek1yfi9-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 122.2 GB at this context; duplicate traces of the same session: 3l2kux00, atgoqjl1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1361.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 64k 1,361 76.6 122.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ek1yfi9",
   "source_url": "https://omlx.ai/benchmarks/performance/0ek1yfi9",
   "tg_tokens": null,
   "tps": 76.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-yzqftuu9-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 119.8 GB; TTFT 2906 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1410.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 4k tokens 1,410 PP tok/s 38.2 TG tok/s 2906 TTFT (ms) 119.8 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yzqftuu9",
   "source_url": "https://omlx.ai/benchmarks/performance/yzqftuu9",
   "tg_tokens": null,
   "tps": 38.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-0ekexhzi-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace, single run at 195k context; peak memory 151.7 GB; TTFT 52978 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3775.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 195k tokens 3,775 PP tok/s 52.0 TG tok/s 52978 TTFT (ms) 151.7 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ekexhzi",
   "source_url": "https://omlx.ai/benchmarks/performance/0ekexhzi",
   "tg_tokens": null,
   "tps": 52.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-oy8kjtye-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 135.4 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4049.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 128k 4,049 88.1 135.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oy8kjtye",
   "source_url": "https://omlx.ai/benchmarks/performance/oy8kjtye",
   "tg_tokens": null,
   "tps": 88.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-oy8kjtye-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 138.2 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4118.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 195k 4,118 76.5 138.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oy8kjtye",
   "source_url": "https://omlx.ai/benchmarks/performance/oy8kjtye",
   "tg_tokens": null,
   "tps": 76.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-00zkqpvc-ctx1024",
   "model": "StartLux-Decision-v1.1-9B-AutoThink",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 10.0 GB at this context; duplicate traces of the same session: 3517ju4t, fms7hnp1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2779.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "StartLux-Decision-v1.1-9B-AutoThink-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 2,779 61.3 10.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/00zkqpvc",
   "source_url": "https://omlx.ai/benchmarks/performance/00zkqpvc",
   "tg_tokens": null,
   "tps": 61.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-00zkqpvc-ctx32768",
   "model": "StartLux-Decision-v1.1-9B-AutoThink",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 12.1 GB at this context; duplicate traces of the same session: 3517ju4t, fms7hnp1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3110.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "StartLux-Decision-v1.1-9B-AutoThink-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 32k 3,110 55.0 12.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/00zkqpvc",
   "source_url": "https://omlx.ai/benchmarks/performance/00zkqpvc",
   "tg_tokens": null,
   "tps": 55.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-00zkqpvc-ctx199680",
   "model": "StartLux-Decision-v1.1-9B-AutoThink",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 3517ju4t, fms7hnp1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1718.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "StartLux-Decision-v1.1-9B-AutoThink-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 195k 1,718 34.2 21.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/00zkqpvc",
   "source_url": "https://omlx.ai/benchmarks/performance/00zkqpvc",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-n16fbthq-ctx1024",
   "model": "Ternary-Bonsai-2-9B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 5.6 GB at this context; duplicate traces of the same session: 4vchtymr, 51jiyg9v; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 172.6,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-9B-mlx-2bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 2bit; 1k 172.6 24.6 5.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n16fbthq",
   "source_url": "https://omlx.ai/benchmarks/performance/n16fbthq",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-n16fbthq-ctx4096",
   "model": "Ternary-Bonsai-2-9B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: 4vchtymr, 51jiyg9v; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 168.5,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-9B-mlx-2bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 2bit; 4k 168.5 23.2 6.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n16fbthq",
   "source_url": "https://omlx.ai/benchmarks/performance/n16fbthq",
   "tg_tokens": null,
   "tps": 23.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-n16fbthq-ctx8192",
   "model": "Ternary-Bonsai-2-9B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 7.3 GB at this context; duplicate traces of the same session: 4vchtymr, 51jiyg9v; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 157.7,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-9B-mlx-2bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 2bit; 8k 157.7 20.7 7.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n16fbthq",
   "source_url": "https://omlx.ai/benchmarks/performance/n16fbthq",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-ahz1h6qh-ctx1024",
   "model": "Ternary-Bonsai-2-9B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 5.6 GB at this context; duplicate traces of the same session: 4gey247t; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 168.5,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-9B-mlx-2bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 2bit; 1k 168.5 24.3 5.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ahz1h6qh",
   "source_url": "https://omlx.ai/benchmarks/performance/ahz1h6qh",
   "tg_tokens": null,
   "tps": 24.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-ahz1h6qh-ctx4096",
   "model": "Ternary-Bonsai-2-9B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: 4gey247t; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 170.7,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-9B-mlx-2bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 2bit; 4k 170.7 23.2 6.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ahz1h6qh",
   "source_url": "https://omlx.ai/benchmarks/performance/ahz1h6qh",
   "tg_tokens": null,
   "tps": 23.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-4f2i978i-ctx1024",
   "model": "gemma-4-12B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 11.3 GB at this context; duplicate traces of the same session: gc84631u, tqz1e20z, utj0owxb; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 882.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-qat-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k 882.3 71.6 11.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4f2i978i",
   "source_url": "https://omlx.ai/benchmarks/performance/4f2i978i",
   "tg_tokens": null,
   "tps": 71.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-4f2i978i-ctx4096",
   "model": "gemma-4-12B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 11.7 GB at this context; duplicate traces of the same session: gc84631u, tqz1e20z, utj0owxb; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2171.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-qat-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 4k 2,171 70.6 11.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4f2i978i",
   "source_url": "https://omlx.ai/benchmarks/performance/4f2i978i",
   "tg_tokens": null,
   "tps": 70.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-4f2i978i-ctx16384",
   "model": "gemma-4-12B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 12.2 GB at this context; duplicate traces of the same session: gc84631u, tqz1e20z, utj0owxb; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2399.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-qat-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 16k 2,399 69.1 12.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4f2i978i",
   "source_url": "https://omlx.ai/benchmarks/performance/4f2i978i",
   "tg_tokens": null,
   "tps": 69.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-4f2i978i-ctx131072",
   "model": "gemma-4-12B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: gc84631u, tqz1e20z, utj0owxb; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1602.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-qat-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 128k 1,602 59.2 18.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4f2i978i",
   "source_url": "https://omlx.ai/benchmarks/performance/4f2i978i",
   "tg_tokens": null,
   "tps": 59.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-d5f6n2ck-ctx4096",
   "model": "Gemma 4 12B",
   "notes": "oMLX community benchmark trace; peak memory 11.7 GB at this context; duplicate traces of the same session: fglaippe; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 956.6,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "gemma-4-12B-nvfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 956.6 23.0 11.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d5f6n2ck",
   "source_url": "https://omlx.ai/benchmarks/performance/d5f6n2ck",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-d5f6n2ck-ctx16384",
   "model": "Gemma 4 12B",
   "notes": "oMLX community benchmark trace; peak memory 12.2 GB at this context; duplicate traces of the same session: fglaippe; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 907.9,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "gemma-4-12B-nvfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 907.9 23.0 12.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d5f6n2ck",
   "source_url": "https://omlx.ai/benchmarks/performance/d5f6n2ck",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-0rrl7273-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 4.7 GB at this context; duplicate traces of the same session: zxae70fl; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4118.0,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "gemma-4-e4b-it-nvfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 4,118 78.2 4.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0rrl7273",
   "source_url": "https://omlx.ai/benchmarks/performance/0rrl7273",
   "tg_tokens": null,
   "tps": 78.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-0rrl7273-ctx16384",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: zxae70fl; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3981.0,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "gemma-4-e4b-it-nvfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 3,981 66.1 4.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0rrl7273",
   "source_url": "https://omlx.ai/benchmarks/performance/0rrl7273",
   "tg_tokens": null,
   "tps": 66.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-c04k87kg-ctx1024",
   "model": "Agents-A1",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 5.3 GB at this context; duplicate traces of the same session: 9ccadrj9, adv7byn0, kt5rokk2, oqa9c11a, rurldhgj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1541.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Agents-A1-4B-mlx-oQ8e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 1,541 90.5 5.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c04k87kg",
   "source_url": "https://omlx.ai/benchmarks/performance/c04k87kg",
   "tg_tokens": null,
   "tps": 90.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-c04k87kg-ctx4096",
   "model": "Agents-A1",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: 9ccadrj9, adv7byn0, kt5rokk2, oqa9c11a, rurldhgj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1685.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Agents-A1-4B-mlx-oQ8e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 1,685 88.9 6.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c04k87kg",
   "source_url": "https://omlx.ai/benchmarks/performance/c04k87kg",
   "tg_tokens": null,
   "tps": 88.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-c04k87kg-ctx8192",
   "model": "Agents-A1",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: 9ccadrj9, adv7byn0, kt5rokk2, oqa9c11a, rurldhgj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1488.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Agents-A1-4B-mlx-oQ8e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 8k 1,488 86.8 6.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c04k87kg",
   "source_url": "https://omlx.ai/benchmarks/performance/c04k87kg",
   "tg_tokens": null,
   "tps": 86.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-c04k87kg-ctx16384",
   "model": "Agents-A1",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 7.1 GB at this context; duplicate traces of the same session: 9ccadrj9, adv7byn0, kt5rokk2, oqa9c11a, rurldhgj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1289.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Agents-A1-4B-mlx-oQ8e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 16k 1,289 81.5 7.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c04k87kg",
   "source_url": "https://omlx.ai/benchmarks/performance/c04k87kg",
   "tg_tokens": null,
   "tps": 81.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-c04k87kg-ctx32768",
   "model": "Agents-A1",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 8.0 GB at this context; duplicate traces of the same session: 9ccadrj9, adv7byn0, kt5rokk2, oqa9c11a, rurldhgj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1079.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Agents-A1-4B-mlx-oQ8e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 32k 1,079 73.2 8.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c04k87kg",
   "source_url": "https://omlx.ai/benchmarks/performance/c04k87kg",
   "tg_tokens": null,
   "tps": 73.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-c04k87kg-ctx65536",
   "model": "Agents-A1",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 9.6 GB at this context; duplicate traces of the same session: 9ccadrj9, adv7byn0, kt5rokk2, oqa9c11a, rurldhgj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 855.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Agents-A1-4B-mlx-oQ8e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 64k 855.2 59.2 9.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c04k87kg",
   "source_url": "https://omlx.ai/benchmarks/performance/c04k87kg",
   "tg_tokens": null,
   "tps": 59.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-y0my9qdr-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.1 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2577.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 2,577 85.5 13.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y0my9qdr",
   "source_url": "https://omlx.ai/benchmarks/performance/y0my9qdr",
   "tg_tokens": null,
   "tps": 85.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-y0my9qdr-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.5 GB at this context; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2982.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 2,982 82.3 13.5 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y0my9qdr",
   "source_url": "https://omlx.ai/benchmarks/performance/y0my9qdr",
   "tg_tokens": null,
   "tps": 82.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-wjk9u1ob-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.1 GB at this context; duplicate traces of the same session: ed8xfkrj; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2577.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 2,577 85.2 13.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wjk9u1ob",
   "source_url": "https://omlx.ai/benchmarks/performance/wjk9u1ob",
   "tg_tokens": null,
   "tps": 85.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-wjk9u1ob-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.5 GB at this context; duplicate traces of the same session: ed8xfkrj; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2991.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 2,991 82.9 13.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wjk9u1ob",
   "source_url": "https://omlx.ai/benchmarks/performance/wjk9u1ob",
   "tg_tokens": null,
   "tps": 82.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-3d4c6suo-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1368.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 1,368 79.7 27.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3d4c6suo",
   "source_url": "https://omlx.ai/benchmarks/performance/3d4c6suo",
   "tg_tokens": null,
   "tps": 79.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-3d4c6suo-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2106.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 2,106 70.4 28.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3d4c6suo",
   "source_url": "https://omlx.ai/benchmarks/performance/3d4c6suo",
   "tg_tokens": null,
   "tps": 70.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-ejfs705k-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "refusal-ablated MTP build; MLX; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: d03rht5s; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 953.6,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-refusal-ablated-MLX-oQ8-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 953.6 78.3 37.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ejfs705k",
   "source_url": "https://omlx.ai/benchmarks/performance/ejfs705k",
   "tg_tokens": null,
   "tps": 78.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ejfs705k-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "refusal-ablated MTP build; MLX; oMLX community benchmark trace; peak memory 39.5 GB at this context; duplicate traces of the same session: d03rht5s; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 883.4,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-refusal-ablated-MLX-oQ8-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 883.4 63.2 39.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ejfs705k",
   "source_url": "https://omlx.ai/benchmarks/performance/ejfs705k",
   "tg_tokens": null,
   "tps": 63.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-vev1cia5-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "refusal-ablated MTP build; MLX; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: c7g65t52; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 946.2,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-refusal-ablated-MLX-oQ8-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 946.2 88.2 37.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vev1cia5",
   "source_url": "https://omlx.ai/benchmarks/performance/vev1cia5",
   "tg_tokens": null,
   "tps": 88.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-vev1cia5-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "refusal-ablated MTP build; MLX; oMLX community benchmark trace; peak memory 39.5 GB at this context; duplicate traces of the same session: c7g65t52; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 889.1,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-refusal-ablated-MLX-oQ8-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 889.1 62.8 39.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vev1cia5",
   "source_url": "https://omlx.ai/benchmarks/performance/vev1cia5",
   "tg_tokens": null,
   "tps": 62.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-6j41ps2n-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "refusal-ablated MTP build; MLX; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: bn49r25b; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 922.8,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-refusal-ablated-MLX-oQ8-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 922.8 77.9 37.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6j41ps2n",
   "source_url": "https://omlx.ai/benchmarks/performance/6j41ps2n",
   "tg_tokens": null,
   "tps": 77.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-6j41ps2n-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "refusal-ablated MTP build; MLX; oMLX community benchmark trace; peak memory 39.5 GB at this context; duplicate traces of the same session: bn49r25b; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 800.3,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-refusal-ablated-MLX-oQ8-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 800.3 58.7 39.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6j41ps2n",
   "source_url": "https://omlx.ai/benchmarks/performance/6j41ps2n",
   "tg_tokens": null,
   "tps": 58.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-fba1cinl-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "refusal-ablated MTP build; MLX; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: vls43mqr; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 930.3,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-refusal-ablated-MLX-oQ8-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 930.3 80.7 37.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fba1cinl",
   "source_url": "https://omlx.ai/benchmarks/performance/fba1cinl",
   "tg_tokens": null,
   "tps": 80.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-fba1cinl-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "refusal-ablated MTP build; MLX; oMLX community benchmark trace; peak memory 39.5 GB at this context; duplicate traces of the same session: vls43mqr; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 832.1,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-refusal-ablated-MLX-oQ8-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 832.1 55.9 39.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fba1cinl",
   "source_url": "https://omlx.ai/benchmarks/performance/fba1cinl",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-rpmkv5t0-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "refusal-ablated MTP build; MLX; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: xhzsbeia; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 910.3,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-refusal-ablated-MLX-oQ8-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 910.3 76.2 37.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rpmkv5t0",
   "source_url": "https://omlx.ai/benchmarks/performance/rpmkv5t0",
   "tg_tokens": null,
   "tps": 76.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-rpmkv5t0-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "refusal-ablated MTP build; MLX; oMLX community benchmark trace; peak memory 39.5 GB at this context; duplicate traces of the same session: xhzsbeia; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 864.9,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-refusal-ablated-MLX-oQ8-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 864.9 61.3 39.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rpmkv5t0",
   "source_url": "https://omlx.ai/benchmarks/performance/rpmkv5t0",
   "tg_tokens": null,
   "tps": 61.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-5268c3au-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "refusal-ablated MTP build; MLX; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: cqdem2xk; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 913.3,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-refusal-ablated-MLX-oQ8-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 913.3 75.1 37.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5268c3au",
   "source_url": "https://omlx.ai/benchmarks/performance/5268c3au",
   "tg_tokens": null,
   "tps": 75.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-5268c3au-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "refusal-ablated MTP build; MLX; oMLX community benchmark trace; peak memory 39.5 GB at this context; duplicate traces of the same session: cqdem2xk; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 864.4,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-refusal-ablated-MLX-oQ8-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 864.4 61.0 39.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5268c3au",
   "source_url": "https://omlx.ai/benchmarks/performance/5268c3au",
   "tg_tokens": null,
   "tps": 61.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-g11acitd-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "refusal-ablated MTP build; MLX; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: lj6cwxr8; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 922.3,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-refusal-ablated-MLX-oQ8-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 922.3 73.1 37.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g11acitd",
   "source_url": "https://omlx.ai/benchmarks/performance/g11acitd",
   "tg_tokens": null,
   "tps": 73.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-g11acitd-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "refusal-ablated MTP build; MLX; oMLX community benchmark trace; peak memory 39.5 GB at this context; duplicate traces of the same session: lj6cwxr8; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 911.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-refusal-ablated-MLX-oQ8-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 911.0 62.7 39.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g11acitd",
   "source_url": "https://omlx.ai/benchmarks/performance/g11acitd",
   "tg_tokens": null,
   "tps": 62.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-w8dgmbc7-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: yrqqbsi8; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 846.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 846.0 102.2 19.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w8dgmbc7",
   "source_url": "https://omlx.ai/benchmarks/performance/w8dgmbc7",
   "tg_tokens": null,
   "tps": 102.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-w8dgmbc7-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: yrqqbsi8; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1018.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 1,018 99.0 21.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w8dgmbc7",
   "source_url": "https://omlx.ai/benchmarks/performance/w8dgmbc7",
   "tg_tokens": null,
   "tps": 99.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-2zklsoew-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: mppcmem7; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 140.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 140.3 21.0 16.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2zklsoew",
   "source_url": "https://omlx.ai/benchmarks/performance/2zklsoew",
   "tg_tokens": null,
   "tps": 21.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-2zklsoew-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: mppcmem7; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 132.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 132.9 20.2 20.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2zklsoew",
   "source_url": "https://omlx.ai/benchmarks/performance/2zklsoew",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0oedpdf1-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 33.8 GB; TTFT 207257 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 632.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k tokens 632.4 PP tok/s 21.4 TG tok/s 207257 TTFT (ms) 33.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0oedpdf1",
   "source_url": "https://omlx.ai/benchmarks/performance/0oedpdf1",
   "tg_tokens": null,
   "tps": 21.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-tzoquznb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: cyc4ketv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 497.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 497.7 24.1 16.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tzoquznb",
   "source_url": "https://omlx.ai/benchmarks/performance/tzoquznb",
   "tg_tokens": null,
   "tps": 24.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-tzoquznb-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.0 GB at this context; duplicate traces of the same session: cyc4ketv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 486.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 486.3 26.1 18.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tzoquznb",
   "source_url": "https://omlx.ai/benchmarks/performance/tzoquznb",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-vqn338pa-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: rnbzgbvv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 367.2,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-27B-NVFP4 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 nvfp4; 1k 367.2 29.6 23.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vqn338pa",
   "source_url": "https://omlx.ai/benchmarks/performance/vqn338pa",
   "tg_tokens": null,
   "tps": 29.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-vqn338pa-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: rnbzgbvv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 424.4,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-27B-NVFP4 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 nvfp4; 4k 424.4 25.4 25.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vqn338pa",
   "source_url": "https://omlx.ai/benchmarks/performance/vqn338pa",
   "tg_tokens": null,
   "tps": 25.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-t2antoc3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: n4uqxa8x; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 139.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 1k 139.4 32.9 18.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t2antoc3",
   "source_url": "https://omlx.ai/benchmarks/performance/t2antoc3",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-t2antoc3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: n4uqxa8x; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k 130.7 28.7 22.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t2antoc3",
   "source_url": "https://omlx.ai/benchmarks/performance/t2antoc3",
   "tg_tokens": null,
   "tps": 28.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-4rgaypu5-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 800s03qb; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 142.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 1k 142.1 19.7 16.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4rgaypu5",
   "source_url": "https://omlx.ai/benchmarks/performance/4rgaypu5",
   "tg_tokens": null,
   "tps": 19.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-4rgaypu5-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 800s03qb; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 132.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k 132.8 19.4 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4rgaypu5",
   "source_url": "https://omlx.ai/benchmarks/performance/4rgaypu5",
   "tg_tokens": null,
   "tps": 19.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-so9tdk6k-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: n9ax1xee; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 441.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 441.0 34.9 21.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/so9tdk6k",
   "source_url": "https://omlx.ai/benchmarks/performance/so9tdk6k",
   "tg_tokens": null,
   "tps": 34.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-so9tdk6k-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: n9ax1xee; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 443.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 443.2 31.0 22.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/so9tdk6k",
   "source_url": "https://omlx.ai/benchmarks/performance/so9tdk6k",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qew7rm1t-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 29.8 GB; TTFT 43911 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 746.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 6bit; 32k tokens 746.2 PP tok/s 62.8 TG tok/s 43911 TTFT (ms) 29.8 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qew7rm1t",
   "source_url": "https://omlx.ai/benchmarks/performance/qew7rm1t",
   "tg_tokens": null,
   "tps": 62.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hwgrx9z5-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 29.8 GB; TTFT 43941 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 745.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 6bit; 32k tokens 745.7 PP tok/s 62.8 TG tok/s 43941 TTFT (ms) 29.8 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hwgrx9z5",
   "source_url": "https://omlx.ai/benchmarks/performance/hwgrx9z5",
   "tg_tokens": null,
   "tps": 62.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-tn8a6x3a-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.5 GB at this context; duplicate traces of the same session: kicp9h66, xrolx04t; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1160.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 1k 1,160 85.0 32.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tn8a6x3a",
   "source_url": "https://omlx.ai/benchmarks/performance/tn8a6x3a",
   "tg_tokens": null,
   "tps": 85.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-tn8a6x3a-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.5 GB at this context; duplicate traces of the same session: kicp9h66, xrolx04t; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1177.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 4k 1,177 79.6 33.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tn8a6x3a",
   "source_url": "https://omlx.ai/benchmarks/performance/tn8a6x3a",
   "tg_tokens": null,
   "tps": 79.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-tn8a6x3a-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.4 GB at this context; duplicate traces of the same session: kicp9h66, xrolx04t; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 970.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 32k 970.7 122.8 35.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tn8a6x3a",
   "source_url": "https://omlx.ai/benchmarks/performance/tn8a6x3a",
   "tg_tokens": null,
   "tps": 122.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-tn8a6x3a-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.7 GB at this context; duplicate traces of the same session: kicp9h66, xrolx04t; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 857.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 64k 857.6 70.7 37.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tn8a6x3a",
   "source_url": "https://omlx.ai/benchmarks/performance/tn8a6x3a",
   "tg_tokens": null,
   "tps": 70.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-b702dyy1-ctx1024",
   "model": "Qwen3.8-Flash-Next-Dynamic",
   "notes": "BF16; PLE; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 75.6 GB at this context; duplicate traces of the same session: 3rukwcdu, byiwa4mf, ieobbmnb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 598.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Dynamic-oQ5e-BF16-PLE-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 598.1 48.9 75.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b702dyy1",
   "source_url": "https://omlx.ai/benchmarks/performance/b702dyy1",
   "tg_tokens": null,
   "tps": 48.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-b702dyy1-ctx65536",
   "model": "Qwen3.8-Flash-Next-Dynamic",
   "notes": "BF16; PLE; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.7 GB at this context; duplicate traces of the same session: 3rukwcdu, byiwa4mf, ieobbmnb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2016.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Dynamic-oQ5e-BF16-PLE-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,016 46.2 78.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b702dyy1",
   "source_url": "https://omlx.ai/benchmarks/performance/b702dyy1",
   "tg_tokens": null,
   "tps": 46.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-b702dyy1-ctx131072",
   "model": "Qwen3.8-Flash-Next-Dynamic",
   "notes": "BF16; PLE; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 81.1 GB at this context; duplicate traces of the same session: 3rukwcdu, byiwa4mf, ieobbmnb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1950.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Dynamic-oQ5e-BF16-PLE-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 1,950 46.4 81.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b702dyy1",
   "source_url": "https://omlx.ai/benchmarks/performance/b702dyy1",
   "tg_tokens": null,
   "tps": 46.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-b702dyy1-ctx199680",
   "model": "Qwen3.8-Flash-Next-Dynamic",
   "notes": "BF16; PLE; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 84.8 GB at this context; duplicate traces of the same session: 3rukwcdu, byiwa4mf, ieobbmnb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1892.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Dynamic-oQ5e-BF16-PLE-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 1,892 44.6 84.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b702dyy1",
   "source_url": "https://omlx.ai/benchmarks/performance/b702dyy1",
   "tg_tokens": null,
   "tps": 44.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bgzkmnbl-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MLX; oMLX community benchmark trace; peak memory 75.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 980.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 980.2 59.5 75.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bgzkmnbl",
   "source_url": "https://omlx.ai/benchmarks/performance/bgzkmnbl",
   "tg_tokens": null,
   "tps": 59.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bgzkmnbl-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MLX; oMLX community benchmark trace; peak memory 78.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2452.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,452 63.3 78.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bgzkmnbl",
   "source_url": "https://omlx.ai/benchmarks/performance/bgzkmnbl",
   "tg_tokens": null,
   "tps": 63.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bgzkmnbl-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MLX; oMLX community benchmark trace; peak memory 81.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2351.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,351 63.4 81.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bgzkmnbl",
   "source_url": "https://omlx.ai/benchmarks/performance/bgzkmnbl",
   "tg_tokens": null,
   "tps": 63.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bgzkmnbl-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MLX; oMLX community benchmark trace; peak memory 84.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2313.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,313 62.2 84.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bgzkmnbl",
   "source_url": "https://omlx.ai/benchmarks/performance/bgzkmnbl",
   "tg_tokens": null,
   "tps": 62.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-l38wqbfd-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 106.6 GB at this context; duplicate traces of the same session: xcz3bcx1; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4005.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 4,005 47.2 106.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l38wqbfd",
   "source_url": "https://omlx.ai/benchmarks/performance/l38wqbfd",
   "tg_tokens": null,
   "tps": 47.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-l38wqbfd-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.3 GB at this context; duplicate traces of the same session: xcz3bcx1; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4242.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 4,242 83.2 109.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l38wqbfd",
   "source_url": "https://omlx.ai/benchmarks/performance/l38wqbfd",
   "tg_tokens": null,
   "tps": 83.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-pbsq2xuu-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 135.4 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4049.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 128k 4,049 88.1 135.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pbsq2xuu",
   "source_url": "https://omlx.ai/benchmarks/performance/pbsq2xuu",
   "tg_tokens": null,
   "tps": 88.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-pbsq2xuu-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 138.2 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4118.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 195k 4,118 76.5 138.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pbsq2xuu",
   "source_url": "https://omlx.ai/benchmarks/performance/pbsq2xuu",
   "tg_tokens": null,
   "tps": 76.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-2uvx297w-ctx1024",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 11.7 GB at this context; duplicate traces of the same session: wlikiirj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 175.7,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 2bit; 1k 175.7 26.1 11.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2uvx297w",
   "source_url": "https://omlx.ai/benchmarks/performance/2uvx297w",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-2uvx297w-ctx8192",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: wlikiirj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 152.0,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 2bit; 8k 152.0 24.7 15.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2uvx297w",
   "source_url": "https://omlx.ai/benchmarks/performance/2uvx297w",
   "tg_tokens": null,
   "tps": 24.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-5gupg4w6-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "mlx-community; oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: 12lnid7q; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 659.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 1k 659.4 59.3 35.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5gupg4w6",
   "source_url": "https://omlx.ai/benchmarks/performance/5gupg4w6",
   "tg_tokens": null,
   "tps": 59.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-5gupg4w6-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "mlx-community; oMLX community benchmark trace; peak memory 37.5 GB at this context; duplicate traces of the same session: 12lnid7q; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 784.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mlx-community--Qwen3.6-35B-A3B-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k 784.9 55.9 37.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5gupg4w6",
   "source_url": "https://omlx.ai/benchmarks/performance/5gupg4w6",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-zfpajnff-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: me7k90d9; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 143.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mlx-community--Qwen3.8-27B-8bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 8bit; 1k 143.2 12.2 28.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zfpajnff",
   "source_url": "https://omlx.ai/benchmarks/performance/zfpajnff",
   "tg_tokens": null,
   "tps": 12.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-zfpajnff-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: me7k90d9; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 143.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mlx-community--Qwen3.8-27B-8bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 8bit; 4k 143.8 11.6 32.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zfpajnff",
   "source_url": "https://omlx.ai/benchmarks/performance/zfpajnff",
   "tg_tokens": null,
   "tps": 11.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-rkoj8yud-ctx65536",
   "model": "GLM-4.5-Air",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 68.5 GB; TTFT 82066 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 798.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-4.5-Air-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 64k tokens 798.6 PP tok/s 24.6 TG tok/s 82066 TTFT (ms) 68.5 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rkoj8yud",
   "source_url": "https://omlx.ai/benchmarks/performance/rkoj8yud",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-72gvuqkz-ctx32768",
   "model": "GLM-4.5-Air",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 62.8 GB; TTFT 26132 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1254.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-4.5-Air-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 32k tokens 1,254 PP tok/s 36.7 TG tok/s 26132 TTFT (ms) 62.8 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/72gvuqkz",
   "source_url": "https://omlx.ai/benchmarks/performance/72gvuqkz",
   "tg_tokens": null,
   "tps": 36.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-6rlacyat-ctx65536",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "MLX serve build; oMLX community benchmark trace, single run at 64k context; peak memory 22.9 GB; TTFT 408766 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 160.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM-V-4.7-35B-A3B-MLX-Serve-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 64k tokens 160.3 PP tok/s 30.9 TG tok/s 408766 TTFT (ms) 22.9 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6rlacyat",
   "source_url": "https://omlx.ai/benchmarks/performance/6rlacyat",
   "tg_tokens": null,
   "tps": 30.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-yh2purtu-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1368.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 1,368 79.7 27.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yh2purtu",
   "source_url": "https://omlx.ai/benchmarks/performance/yh2purtu",
   "tg_tokens": null,
   "tps": 79.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-yh2purtu-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2106.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 2,106 70.4 28.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yh2purtu",
   "source_url": "https://omlx.ai/benchmarks/performance/yh2purtu",
   "tg_tokens": null,
   "tps": 70.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-cuz08hla-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MLX build; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 449.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-MLX-oQ4-FP16 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 449.0 36.2 20.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cuz08hla",
   "source_url": "https://omlx.ai/benchmarks/performance/cuz08hla",
   "tg_tokens": null,
   "tps": 36.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-cuz08hla-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MLX build; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 380.7,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-MLX-oQ4-FP16 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 16k 380.7 31.1 21.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cuz08hla",
   "source_url": "https://omlx.ai/benchmarks/performance/cuz08hla",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4sjvs2kh-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 23.4 GB; TTFT 18267 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3588.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 64k tokens 3,588 PP tok/s 151.7 TG tok/s 18267 TTFT (ms) 23.4 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4sjvs2kh",
   "source_url": "https://omlx.ai/benchmarks/performance/4sjvs2kh",
   "tg_tokens": null,
   "tps": 151.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ngnac083-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 22.7 GB; TTFT 6916 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4738.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k tokens 4,738 PP tok/s 192.6 TG tok/s 6916 TTFT (ms) 22.7 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ngnac083",
   "source_url": "https://omlx.ai/benchmarks/performance/ngnac083",
   "tg_tokens": null,
   "tps": 192.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-o2utx6il-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 22.7 GB; TTFT 6691 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4898.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k tokens 4,898 PP tok/s 193.3 TG tok/s 6691 TTFT (ms) 22.7 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o2utx6il",
   "source_url": "https://omlx.ai/benchmarks/performance/o2utx6il",
   "tg_tokens": null,
   "tps": 193.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-usxxmt70-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 21.3 GB; TTFT 6612 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4956.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k tokens 4,956 PP tok/s 114.7 TG tok/s 6612 TTFT (ms) 21.3 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/usxxmt70",
   "source_url": "https://omlx.ai/benchmarks/performance/usxxmt70",
   "tg_tokens": null,
   "tps": 114.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-0bjv1ac8-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: i0fv6e40; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 968.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 968.1 95.7 37.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0bjv1ac8",
   "source_url": "https://omlx.ai/benchmarks/performance/0bjv1ac8",
   "tg_tokens": null,
   "tps": 95.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-0bjv1ac8-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: i0fv6e40; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 898.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 898.3 69.7 39.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0bjv1ac8",
   "source_url": "https://omlx.ai/benchmarks/performance/0bjv1ac8",
   "tg_tokens": null,
   "tps": 69.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-94zhm24b-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: uz3sebbl; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 953.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 953.3 97.5 37.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/94zhm24b",
   "source_url": "https://omlx.ai/benchmarks/performance/94zhm24b",
   "tg_tokens": null,
   "tps": 97.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-94zhm24b-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: uz3sebbl; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 889.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 889.0 63.4 39.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/94zhm24b",
   "source_url": "https://omlx.ai/benchmarks/performance/94zhm24b",
   "tg_tokens": null,
   "tps": 63.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-ggb6pt0z-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 18.3 GB; TTFT 21895 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 187.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k tokens 187.1 PP tok/s 12.2 TG tok/s 21895 TTFT (ms) 18.3 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ggb6pt0z",
   "source_url": "https://omlx.ai/benchmarks/performance/ggb6pt0z",
   "tg_tokens": null,
   "tps": 12.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kcwj73pb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: hugob5kz, nzgtowlz, salggrxy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 930.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 930.5 33.5 16.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kcwj73pb",
   "source_url": "https://omlx.ai/benchmarks/performance/kcwj73pb",
   "tg_tokens": null,
   "tps": 33.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kcwj73pb-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: hugob5kz, nzgtowlz, salggrxy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 687.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 687.3 25.6 20.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kcwj73pb",
   "source_url": "https://omlx.ai/benchmarks/performance/kcwj73pb",
   "tg_tokens": null,
   "tps": 25.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-kcwj73pb-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: hugob5kz, nzgtowlz, salggrxy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 549.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 549.5 20.8 24.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kcwj73pb",
   "source_url": "https://omlx.ai/benchmarks/performance/kcwj73pb",
   "tg_tokens": null,
   "tps": 20.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-kcwj73pb-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.0 GB at this context; duplicate traces of the same session: hugob5kz, nzgtowlz, salggrxy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 449.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 449.9 17.2 29.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kcwj73pb",
   "source_url": "https://omlx.ai/benchmarks/performance/kcwj73pb",
   "tg_tokens": null,
   "tps": 17.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-t7w6fp7x-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 20.9 GB; TTFT 56508 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1160.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 64k tokens 1,160 PP tok/s 40.1 TG tok/s 56508 TTFT (ms) 20.9 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t7w6fp7x",
   "source_url": "https://omlx.ai/benchmarks/performance/t7w6fp7x",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-p0r5vi7q-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 20.8 GB; TTFT 8436 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 485.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k tokens 485.5 PP tok/s 36.7 TG tok/s 8436 TTFT (ms) 20.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p0r5vi7q",
   "source_url": "https://omlx.ai/benchmarks/performance/p0r5vi7q",
   "tg_tokens": null,
   "tps": 36.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-n0qy0kp5-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: s5jdscya; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 477.7,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-27B-nvfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 477.7 17.1 16.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n0qy0kp5",
   "source_url": "https://omlx.ai/benchmarks/performance/n0qy0kp5",
   "tg_tokens": null,
   "tps": 17.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-n0qy0kp5-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: s5jdscya; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 501.9,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-27B-nvfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 501.9 17.4 17.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n0qy0kp5",
   "source_url": "https://omlx.ai/benchmarks/performance/n0qy0kp5",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-m8b9hq4n-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 18.5 GB; TTFT 7341 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 139.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 1k tokens 139.5 PP tok/s 31.6 TG tok/s 7341 TTFT (ms) 18.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m8b9hq4n",
   "source_url": "https://omlx.ai/benchmarks/performance/m8b9hq4n",
   "tg_tokens": null,
   "tps": 31.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-s5zgd0up-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.5 GB at this context; duplicate traces of the same session: 0xuv9viz; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 380.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 380.8 37.5 17.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s5zgd0up",
   "source_url": "https://omlx.ai/benchmarks/performance/s5zgd0up",
   "tg_tokens": null,
   "tps": 37.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-s5zgd0up-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 0xuv9viz; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 392.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 392.8 40.2 19.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s5zgd0up",
   "source_url": "https://omlx.ai/benchmarks/performance/s5zgd0up",
   "tg_tokens": null,
   "tps": 40.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-xos45or7-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 33b8fqzr; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 466.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 466.6 30.3 17.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xos45or7",
   "source_url": "https://omlx.ai/benchmarks/performance/xos45or7",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-xos45or7-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 33b8fqzr; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 577.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 577.6 44.2 19.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xos45or7",
   "source_url": "https://omlx.ai/benchmarks/performance/xos45or7",
   "tg_tokens": null,
   "tps": 44.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-8nrc1k23-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.5 GB at this context; duplicate traces of the same session: 1luu8eas; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 385.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 385.8 36.9 17.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8nrc1k23",
   "source_url": "https://omlx.ai/benchmarks/performance/8nrc1k23",
   "tg_tokens": null,
   "tps": 36.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-8nrc1k23-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 1luu8eas; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 379.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 379.9 31.6 19.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8nrc1k23",
   "source_url": "https://omlx.ai/benchmarks/performance/8nrc1k23",
   "tg_tokens": null,
   "tps": 31.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-2nwcouow-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.5 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 385.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 385.4 39.8 17.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2nwcouow",
   "source_url": "https://omlx.ai/benchmarks/performance/2nwcouow",
   "tg_tokens": null,
   "tps": 39.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-2nwcouow-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 399.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 399.5 32.2 19.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2nwcouow",
   "source_url": "https://omlx.ai/benchmarks/performance/2nwcouow",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-t2nm0o8u-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: fy5wccth; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 431.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 431.0 23.6 20.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t2nm0o8u",
   "source_url": "https://omlx.ai/benchmarks/performance/t2nm0o8u",
   "tg_tokens": null,
   "tps": 23.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-t2nm0o8u-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: fy5wccth; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 487.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 487.5 32.6 21.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t2nm0o8u",
   "source_url": "https://omlx.ai/benchmarks/performance/t2nm0o8u",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-4l7y357s-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.5 GB at this context; duplicate traces of the same session: qt566tiv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 466.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 466.6 17.4 16.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4l7y357s",
   "source_url": "https://omlx.ai/benchmarks/performance/4l7y357s",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-4l7y357s-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: qt566tiv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 495.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 495.0 17.2 17.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4l7y357s",
   "source_url": "https://omlx.ai/benchmarks/performance/4l7y357s",
   "tg_tokens": null,
   "tps": 17.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-hnry195l-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 4y7jm3x0, xp7krsmk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 582.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 582.6 38.8 17.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hnry195l",
   "source_url": "https://omlx.ai/benchmarks/performance/hnry195l",
   "tg_tokens": null,
   "tps": 38.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-hnry195l-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 4y7jm3x0, xp7krsmk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 685.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 685.2 34.4 19.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hnry195l",
   "source_url": "https://omlx.ai/benchmarks/performance/hnry195l",
   "tg_tokens": null,
   "tps": 34.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-hnry195l-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 4y7jm3x0, xp7krsmk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 682.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 682.0 43.9 19.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hnry195l",
   "source_url": "https://omlx.ai/benchmarks/performance/hnry195l",
   "tg_tokens": null,
   "tps": 43.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-pv3qf2g6-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 17.6 GB; TTFT 1759 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 582.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k tokens 582.2 PP tok/s 32.3 TG tok/s 1759 TTFT (ms) 17.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pv3qf2g6",
   "source_url": "https://omlx.ai/benchmarks/performance/pv3qf2g6",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-yci373lj-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 17.6 GB; TTFT 1767 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 579.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k tokens 579.5 PP tok/s 37.2 TG tok/s 1767 TTFT (ms) 17.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yci373lj",
   "source_url": "https://omlx.ai/benchmarks/performance/yci373lj",
   "tg_tokens": null,
   "tps": 37.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1v05s2t3-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 29.8 GB; TTFT 44673 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 733.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 6bit; 32k tokens 733.5 PP tok/s 34.4 TG tok/s 44673 TTFT (ms) 29.8 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1v05s2t3",
   "source_url": "https://omlx.ai/benchmarks/performance/1v05s2t3",
   "tg_tokens": null,
   "tps": 34.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-rsn81k08-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.5 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1160.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 1k 1,160 85.0 32.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rsn81k08",
   "source_url": "https://omlx.ai/benchmarks/performance/rsn81k08",
   "tg_tokens": null,
   "tps": 85.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-rsn81k08-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.5 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1177.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 4k 1,177 79.6 33.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rsn81k08",
   "source_url": "https://omlx.ai/benchmarks/performance/rsn81k08",
   "tg_tokens": null,
   "tps": 79.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-rsn81k08-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.4 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 970.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 32k 970.7 122.8 35.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rsn81k08",
   "source_url": "https://omlx.ai/benchmarks/performance/rsn81k08",
   "tg_tokens": null,
   "tps": 122.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-rsn81k08-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.7 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 857.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 64k 857.6 70.7 37.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rsn81k08",
   "source_url": "https://omlx.ai/benchmarks/performance/rsn81k08",
   "tg_tokens": null,
   "tps": 70.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ra46clte-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: 7bc8my6j, iofgciun, nqrhqg3m, tdaqd8ju; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1158.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 1k 1,158 85.7 26.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ra46clte",
   "source_url": "https://omlx.ai/benchmarks/performance/ra46clte",
   "tg_tokens": null,
   "tps": 85.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ra46clte-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: 7bc8my6j, iofgciun, nqrhqg3m, tdaqd8ju; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1387.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 4k 1,387 76.2 27.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ra46clte",
   "source_url": "https://omlx.ai/benchmarks/performance/ra46clte",
   "tg_tokens": null,
   "tps": 76.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ra46clte-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: 7bc8my6j, iofgciun, nqrhqg3m, tdaqd8ju; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1244.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 32k 1,244 107.2 29.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ra46clte",
   "source_url": "https://omlx.ai/benchmarks/performance/ra46clte",
   "tg_tokens": null,
   "tps": 107.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ra46clte-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: 7bc8my6j, iofgciun, nqrhqg3m, tdaqd8ju; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1099.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 64k 1,099 74.7 32.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ra46clte",
   "source_url": "https://omlx.ai/benchmarks/performance/ra46clte",
   "tg_tokens": null,
   "tps": 74.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ra46clte-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.8 GB at this context; duplicate traces of the same session: 7bc8my6j, iofgciun, nqrhqg3m, tdaqd8ju; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 871.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 128k 871.5 70.0 36.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ra46clte",
   "source_url": "https://omlx.ai/benchmarks/performance/ra46clte",
   "tg_tokens": null,
   "tps": 70.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dx1w2diq-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MLX; oMLX community benchmark trace; peak memory 75.3 GB at this context; duplicate traces of the same session: fthrgolr, kt80yh9a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 980.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 980.2 59.5 75.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dx1w2diq",
   "source_url": "https://omlx.ai/benchmarks/performance/dx1w2diq",
   "tg_tokens": null,
   "tps": 59.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dx1w2diq-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MLX; oMLX community benchmark trace; peak memory 78.6 GB at this context; duplicate traces of the same session: fthrgolr, kt80yh9a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2452.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,452 63.3 78.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dx1w2diq",
   "source_url": "https://omlx.ai/benchmarks/performance/dx1w2diq",
   "tg_tokens": null,
   "tps": 63.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dx1w2diq-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MLX; oMLX community benchmark trace; peak memory 81.0 GB at this context; duplicate traces of the same session: fthrgolr, kt80yh9a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2351.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,351 63.4 81.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dx1w2diq",
   "source_url": "https://omlx.ai/benchmarks/performance/dx1w2diq",
   "tg_tokens": null,
   "tps": 63.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dx1w2diq-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MLX; oMLX community benchmark trace; peak memory 84.5 GB at this context; duplicate traces of the same session: fthrgolr, kt80yh9a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2313.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,313 62.2 84.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dx1w2diq",
   "source_url": "https://omlx.ai/benchmarks/performance/dx1w2diq",
   "tg_tokens": null,
   "tps": 62.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3oog576d-ctx1024",
   "model": "StartLux-Decision-v1.1-9B-AutoThink",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 10.0 GB at this context; duplicate traces of the same session: gzfac99r, pq1vefpf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2916.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "StartLux-Decision-v1.1-9B-AutoThink-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 2,916 60.9 10.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3oog576d",
   "source_url": "https://omlx.ai/benchmarks/performance/3oog576d",
   "tg_tokens": null,
   "tps": 60.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3oog576d-ctx32768",
   "model": "StartLux-Decision-v1.1-9B-AutoThink",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 12.1 GB at this context; duplicate traces of the same session: gzfac99r, pq1vefpf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3117.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "StartLux-Decision-v1.1-9B-AutoThink-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 32k 3,117 54.6 12.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3oog576d",
   "source_url": "https://omlx.ai/benchmarks/performance/3oog576d",
   "tg_tokens": null,
   "tps": 54.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3oog576d-ctx199680",
   "model": "StartLux-Decision-v1.1-9B-AutoThink",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: gzfac99r, pq1vefpf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1767.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "StartLux-Decision-v1.1-9B-AutoThink-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 195k 1,767 34.9 21.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3oog576d",
   "source_url": "https://omlx.ai/benchmarks/performance/3oog576d",
   "tg_tokens": null,
   "tps": 34.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-b3wca1al-ctx1024",
   "model": "Ternary-Bonsai-2-4B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 4.0 GB at this context; duplicate traces of the same session: ofa0163m, tntz2lpy; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 349.3,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-4B-mlx-2bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 2bit; 1k 349.3 48.1 4.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b3wca1al",
   "source_url": "https://omlx.ai/benchmarks/performance/b3wca1al",
   "tg_tokens": null,
   "tps": 48.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-b3wca1al-ctx4096",
   "model": "Ternary-Bonsai-2-4B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 5.2 GB at this context; duplicate traces of the same session: ofa0163m, tntz2lpy; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 342.0,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-4B-mlx-2bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 2bit; 4k 342.0 34.6 5.2 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b3wca1al",
   "source_url": "https://omlx.ai/benchmarks/performance/b3wca1al",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-b3wca1al-ctx8192",
   "model": "Ternary-Bonsai-2-4B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 6.0 GB at this context; duplicate traces of the same session: ofa0163m, tntz2lpy; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 329.3,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-4B-mlx-2bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 2bit; 8k 329.3 26.3 6.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b3wca1al",
   "source_url": "https://omlx.ai/benchmarks/performance/b3wca1al",
   "tg_tokens": null,
   "tps": 26.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-b3wca1al-ctx16384",
   "model": "Ternary-Bonsai-2-4B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: ofa0163m, tntz2lpy; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 286.0,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-4B-mlx-2bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 2bit; 16k 286.0 18.3 6.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b3wca1al",
   "source_url": "https://omlx.ai/benchmarks/performance/b3wca1al",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-mqm99kba-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace, single run at 1k context; peak memory 32.1 GB; TTFT 7886 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 129.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mlx-community--Qwen3.8-27B-8bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 8bit; 1k tokens 129.9 PP tok/s 19.2 TG tok/s 7886 TTFT (ms) 32.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mqm99kba",
   "source_url": "https://omlx.ai/benchmarks/performance/mqm99kba",
   "tg_tokens": null,
   "tps": 19.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-t8zudt47-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 65.8 GB at this context; duplicate traces of the same session: mwpygb0l; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1189.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Ornith-1.5-35B-A3B-MLX on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 unknown; 1k 1,189 44.6 65.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t8zudt47",
   "source_url": "https://omlx.ai/benchmarks/performance/t8zudt47",
   "tg_tokens": null,
   "tps": 44.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-t8zudt47-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 67.6 GB at this context; duplicate traces of the same session: mwpygb0l; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 958.6,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Ornith-1.5-35B-A3B-MLX on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 unknown; 4k 958.6 38.6 67.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t8zudt47",
   "source_url": "https://omlx.ai/benchmarks/performance/t8zudt47",
   "tg_tokens": null,
   "tps": 38.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-9srxr081-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 65.8 GB at this context; duplicate traces of the same session: i8531efl; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1229.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Ornith-1.5-35B-A3B-MLX on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 unknown; 1k 1,229 45.3 65.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9srxr081",
   "source_url": "https://omlx.ai/benchmarks/performance/9srxr081",
   "tg_tokens": null,
   "tps": 45.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-9srxr081-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 67.6 GB at this context; duplicate traces of the same session: i8531efl; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1234.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Ornith-1.5-35B-A3B-MLX on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 unknown; 4k 1,234 42.0 67.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9srxr081",
   "source_url": "https://omlx.ai/benchmarks/performance/9srxr081",
   "tg_tokens": null,
   "tps": 42.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-asvnxv1r-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "refusal-ablated MTP build; MLX; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: m23jeweq; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 950.1,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-refusal-ablated-MLX-oQ8-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 950.1 77.4 37.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/asvnxv1r",
   "source_url": "https://omlx.ai/benchmarks/performance/asvnxv1r",
   "tg_tokens": null,
   "tps": 77.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-asvnxv1r-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "refusal-ablated MTP build; MLX; oMLX community benchmark trace; peak memory 39.5 GB at this context; duplicate traces of the same session: m23jeweq; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 908.3,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-refusal-ablated-MLX-oQ8-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 908.3 62.1 39.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/asvnxv1r",
   "source_url": "https://omlx.ai/benchmarks/performance/asvnxv1r",
   "tg_tokens": null,
   "tps": 62.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-vodrj5h0-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "refusal-ablated MTP build; MLX; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: rfsv65ka; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 942.1,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-refusal-ablated-MLX-oQ8-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 942.1 86.3 37.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vodrj5h0",
   "source_url": "https://omlx.ai/benchmarks/performance/vodrj5h0",
   "tg_tokens": null,
   "tps": 86.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-vodrj5h0-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "refusal-ablated MTP build; MLX; oMLX community benchmark trace; peak memory 39.5 GB at this context; duplicate traces of the same session: rfsv65ka; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 871.7,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-refusal-ablated-MLX-oQ8-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 871.7 64.0 39.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vodrj5h0",
   "source_url": "https://omlx.ai/benchmarks/performance/vodrj5h0",
   "tg_tokens": null,
   "tps": 64.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-wtdlfwbx-ctx1024",
   "model": "Qwen3.6-27B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: mb3piy88; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 76.2,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-27B-oQ4-fp16-mtp on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 76.2 14.7 19.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wtdlfwbx",
   "source_url": "https://omlx.ai/benchmarks/performance/wtdlfwbx",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-wtdlfwbx-ctx4096",
   "model": "Qwen3.6-27B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: mb3piy88; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 78.6,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-27B-oQ4-fp16-mtp on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 78.6 13.4 20.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wtdlfwbx",
   "source_url": "https://omlx.ai/benchmarks/performance/wtdlfwbx",
   "tg_tokens": null,
   "tps": 13.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-u26q23bl-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MLX build; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 449.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-MLX-oQ4-FP16 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 449.0 36.2 20.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u26q23bl",
   "source_url": "https://omlx.ai/benchmarks/performance/u26q23bl",
   "tg_tokens": null,
   "tps": 36.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-u26q23bl-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 MLX build; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 380.7,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-MLX-oQ4-FP16 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 16k 380.7 31.1 21.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u26q23bl",
   "source_url": "https://omlx.ai/benchmarks/performance/u26q23bl",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-ujv05dt4-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 18.3 GB; TTFT 24362 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 168.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k tokens 168.1 PP tok/s 10.9 TG tok/s 24362 TTFT (ms) 18.3 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ujv05dt4",
   "source_url": "https://omlx.ai/benchmarks/performance/ujv05dt4",
   "tg_tokens": null,
   "tps": 10.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-w27n0bbr-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: z16a2ekr; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 489.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 489.8 17.5 17.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w27n0bbr",
   "source_url": "https://omlx.ai/benchmarks/performance/w27n0bbr",
   "tg_tokens": null,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-w27n0bbr-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: z16a2ekr; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 483.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 483.1 16.4 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w27n0bbr",
   "source_url": "https://omlx.ai/benchmarks/performance/w27n0bbr",
   "tg_tokens": null,
   "tps": 16.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-w51treun-ctx1024",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "MTP MLX build; oMLX community benchmark trace; peak memory 33.1 GB at this context; duplicate traces of the same session: l5qyhzq0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 415.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Abliterated-MTPLX-6bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 1k 415.3 65.1 33.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w51treun",
   "source_url": "https://omlx.ai/benchmarks/performance/w51treun",
   "tg_tokens": null,
   "tps": 65.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-w51treun-ctx4096",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "MTP MLX build; oMLX community benchmark trace; peak memory 36.3 GB at this context; duplicate traces of the same session: l5qyhzq0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 551.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Abliterated-MTPLX-6bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 4k 551.8 62.9 36.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w51treun",
   "source_url": "https://omlx.ai/benchmarks/performance/w51treun",
   "tg_tokens": null,
   "tps": 62.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-k3m0lqh9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 9yza1lwu, tvc80x4h, yv7gji45; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 763.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 763.7 18.8 28.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k3m0lqh9",
   "source_url": "https://omlx.ai/benchmarks/performance/k3m0lqh9",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-k3m0lqh9-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 33.3 GB at this context; duplicate traces of the same session: 9yza1lwu, tvc80x4h, yv7gji45; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 673.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 673.9 16.0 33.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k3m0lqh9",
   "source_url": "https://omlx.ai/benchmarks/performance/k3m0lqh9",
   "tg_tokens": null,
   "tps": 16.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-k3m0lqh9-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: 9yza1lwu, tvc80x4h, yv7gji45; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 542.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 542.2 14.3 37.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k3m0lqh9",
   "source_url": "https://omlx.ai/benchmarks/performance/k3m0lqh9",
   "tg_tokens": null,
   "tps": 14.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-k3m0lqh9-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 41.5 GB at this context; duplicate traces of the same session: 9yza1lwu, tvc80x4h, yv7gji45; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 438.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 195k 438.5 12.4 41.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k3m0lqh9",
   "source_url": "https://omlx.ai/benchmarks/performance/k3m0lqh9",
   "tg_tokens": null,
   "tps": 12.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-zxegoyv8-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: gqravuo1; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 139.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 1k 139.1 32.9 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zxegoyv8",
   "source_url": "https://omlx.ai/benchmarks/performance/zxegoyv8",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-zxegoyv8-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: gqravuo1; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 150.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k 150.4 30.0 25.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zxegoyv8",
   "source_url": "https://omlx.ai/benchmarks/performance/zxegoyv8",
   "tg_tokens": null,
   "tps": 30.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-69ph3hjw-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: kfk9fs9n; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 138.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 1k 138.7 15.1 20.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/69ph3hjw",
   "source_url": "https://omlx.ai/benchmarks/performance/69ph3hjw",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-69ph3hjw-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: kfk9fs9n; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 113.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k 113.1 14.9 21.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/69ph3hjw",
   "source_url": "https://omlx.ai/benchmarks/performance/69ph3hjw",
   "tg_tokens": null,
   "tps": 14.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-yno69p70-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: grqv5p1k; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 424.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 424.1 75.7 18.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yno69p70",
   "source_url": "https://omlx.ai/benchmarks/performance/yno69p70",
   "tg_tokens": null,
   "tps": 75.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-yno69p70-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: grqv5p1k; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 461.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 461.1 70.7 21.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yno69p70",
   "source_url": "https://omlx.ai/benchmarks/performance/yno69p70",
   "tg_tokens": null,
   "tps": 70.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-1i9oprrq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.5 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 385.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 385.4 39.8 17.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1i9oprrq",
   "source_url": "https://omlx.ai/benchmarks/performance/1i9oprrq",
   "tg_tokens": null,
   "tps": 39.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1i9oprrq-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 399.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 399.5 32.2 19.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1i9oprrq",
   "source_url": "https://omlx.ai/benchmarks/performance/1i9oprrq",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-48qulmu3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.5 GB at this context; duplicate traces of the same session: i55v588p; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 380.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 380.8 20.2 17.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/48qulmu3",
   "source_url": "https://omlx.ai/benchmarks/performance/48qulmu3",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-48qulmu3-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: i55v588p; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 345.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 345.8 18.4 20.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/48qulmu3",
   "source_url": "https://omlx.ai/benchmarks/performance/48qulmu3",
   "tg_tokens": null,
   "tps": 18.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-2fvporcn-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.5 GB at this context; duplicate traces of the same session: lnf0xgqd; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 382.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 382.5 22.0 17.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2fvporcn",
   "source_url": "https://omlx.ai/benchmarks/performance/2fvporcn",
   "tg_tokens": null,
   "tps": 22.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-2fvporcn-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: lnf0xgqd; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 336.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 336.1 19.7 20.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2fvporcn",
   "source_url": "https://omlx.ai/benchmarks/performance/2fvporcn",
   "tg_tokens": null,
   "tps": 19.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-rmplovux-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: s2iagmb7; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 90.2,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 2bit; 1k 90.2 8.4 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rmplovux",
   "source_url": "https://omlx.ai/benchmarks/performance/rmplovux",
   "tg_tokens": null,
   "tps": 8.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-rmplovux-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: s2iagmb7; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 191.8,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 2bit; 4k 191.8 9.3 21.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rmplovux",
   "source_url": "https://omlx.ai/benchmarks/performance/rmplovux",
   "tg_tokens": null,
   "tps": 9.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fh8ejk62-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: bmztujp7, fu9gw08m, gbu3hn3d, u5b6givv, u84qqsls; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1026.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,026 60.7 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fh8ejk62",
   "source_url": "https://omlx.ai/benchmarks/performance/fh8ejk62",
   "tg_tokens": null,
   "tps": 60.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fh8ejk62-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: bmztujp7, fu9gw08m, gbu3hn3d, u5b6givv, u84qqsls; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1287.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,287 56.9 71.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fh8ejk62",
   "source_url": "https://omlx.ai/benchmarks/performance/fh8ejk62",
   "tg_tokens": null,
   "tps": 56.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fh8ejk62-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.0 GB at this context; duplicate traces of the same session: bmztujp7, fu9gw08m, gbu3hn3d, u5b6givv, u84qqsls; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1562.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,562 57.1 72.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fh8ejk62",
   "source_url": "https://omlx.ai/benchmarks/performance/fh8ejk62",
   "tg_tokens": null,
   "tps": 57.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fh8ejk62-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.0 GB at this context; duplicate traces of the same session: bmztujp7, fu9gw08m, gbu3hn3d, u5b6givv, u84qqsls; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1637.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,637 61.2 73.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fh8ejk62",
   "source_url": "https://omlx.ai/benchmarks/performance/fh8ejk62",
   "tg_tokens": null,
   "tps": 61.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fh8ejk62-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.6 GB at this context; duplicate traces of the same session: bmztujp7, fu9gw08m, gbu3hn3d, u5b6givv, u84qqsls; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1669.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,669 59.2 73.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fh8ejk62",
   "source_url": "https://omlx.ai/benchmarks/performance/fh8ejk62",
   "tg_tokens": null,
   "tps": 59.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fh8ejk62-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.7 GB at this context; duplicate traces of the same session: bmztujp7, fu9gw08m, gbu3hn3d, u5b6givv, u84qqsls; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1631.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 1,631 50.9 74.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fh8ejk62",
   "source_url": "https://omlx.ai/benchmarks/performance/fh8ejk62",
   "tg_tokens": null,
   "tps": 50.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-o9zp5kng-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: 0xno0y0z, 67siy5n6, 6wpuml4l, e72v14q8, n4eeeuy6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 990.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 990.4 61.7 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9zp5kng",
   "source_url": "https://omlx.ai/benchmarks/performance/o9zp5kng",
   "tg_tokens": null,
   "tps": 61.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-o9zp5kng-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: 0xno0y0z, 67siy5n6, 6wpuml4l, e72v14q8, n4eeeuy6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1321.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,321 58.1 71.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9zp5kng",
   "source_url": "https://omlx.ai/benchmarks/performance/o9zp5kng",
   "tg_tokens": null,
   "tps": 58.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-o9zp5kng-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.0 GB at this context; duplicate traces of the same session: 0xno0y0z, 67siy5n6, 6wpuml4l, e72v14q8, n4eeeuy6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1501.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,501 57.2 72.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9zp5kng",
   "source_url": "https://omlx.ai/benchmarks/performance/o9zp5kng",
   "tg_tokens": null,
   "tps": 57.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-o9zp5kng-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.0 GB at this context; duplicate traces of the same session: 0xno0y0z, 67siy5n6, 6wpuml4l, e72v14q8, n4eeeuy6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1654.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,654 60.3 73.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9zp5kng",
   "source_url": "https://omlx.ai/benchmarks/performance/o9zp5kng",
   "tg_tokens": null,
   "tps": 60.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-o9zp5kng-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.6 GB at this context; duplicate traces of the same session: 0xno0y0z, 67siy5n6, 6wpuml4l, e72v14q8, n4eeeuy6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1649.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,649 59.5 73.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9zp5kng",
   "source_url": "https://omlx.ai/benchmarks/performance/o9zp5kng",
   "tg_tokens": null,
   "tps": 59.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-o9zp5kng-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.7 GB at this context; duplicate traces of the same session: 0xno0y0z, 67siy5n6, 6wpuml4l, e72v14q8, n4eeeuy6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1623.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 1,623 50.0 74.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9zp5kng",
   "source_url": "https://omlx.ai/benchmarks/performance/o9zp5kng",
   "tg_tokens": null,
   "tps": 50.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-io09ztoq-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: 15rt0g47, 6kviq4wo, esi80a03, p5miv0nq, zqgs3l0g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1051.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,051 82.5 71.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/io09ztoq",
   "source_url": "https://omlx.ai/benchmarks/performance/io09ztoq",
   "tg_tokens": null,
   "tps": 82.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-io09ztoq-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: 15rt0g47, 6kviq4wo, esi80a03, p5miv0nq, zqgs3l0g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1299.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,299 64.3 71.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/io09ztoq",
   "source_url": "https://omlx.ai/benchmarks/performance/io09ztoq",
   "tg_tokens": null,
   "tps": 64.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-io09ztoq-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.0 GB at this context; duplicate traces of the same session: 15rt0g47, 6kviq4wo, esi80a03, p5miv0nq, zqgs3l0g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1470.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,470 60.5 72.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/io09ztoq",
   "source_url": "https://omlx.ai/benchmarks/performance/io09ztoq",
   "tg_tokens": null,
   "tps": 60.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-io09ztoq-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.0 GB at this context; duplicate traces of the same session: 15rt0g47, 6kviq4wo, esi80a03, p5miv0nq, zqgs3l0g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1516.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,516 56.4 73.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/io09ztoq",
   "source_url": "https://omlx.ai/benchmarks/performance/io09ztoq",
   "tg_tokens": null,
   "tps": 56.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-io09ztoq-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.6 GB at this context; duplicate traces of the same session: 15rt0g47, 6kviq4wo, esi80a03, p5miv0nq, zqgs3l0g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1490.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,490 55.6 73.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/io09ztoq",
   "source_url": "https://omlx.ai/benchmarks/performance/io09ztoq",
   "tg_tokens": null,
   "tps": 55.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-io09ztoq-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.7 GB at this context; duplicate traces of the same session: 15rt0g47, 6kviq4wo, esi80a03, p5miv0nq, zqgs3l0g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1614.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 1,614 55.2 74.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/io09ztoq",
   "source_url": "https://omlx.ai/benchmarks/performance/io09ztoq",
   "tg_tokens": null,
   "tps": 55.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hrprcyl5-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.5 GB at this context; duplicate traces of the same session: 07oc3keu, 2ggyzvmh, 3x89x43m, 90at92kf, s8gmjr6l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1140.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,140 64.0 70.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hrprcyl5",
   "source_url": "https://omlx.ai/benchmarks/performance/hrprcyl5",
   "tg_tokens": null,
   "tps": 64.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hrprcyl5-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 07oc3keu, 2ggyzvmh, 3x89x43m, 90at92kf, s8gmjr6l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1387.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,387 79.9 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hrprcyl5",
   "source_url": "https://omlx.ai/benchmarks/performance/hrprcyl5",
   "tg_tokens": null,
   "tps": 79.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hrprcyl5-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.0 GB at this context; duplicate traces of the same session: 07oc3keu, 2ggyzvmh, 3x89x43m, 90at92kf, s8gmjr6l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1745.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,745 65.8 72.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hrprcyl5",
   "source_url": "https://omlx.ai/benchmarks/performance/hrprcyl5",
   "tg_tokens": null,
   "tps": 65.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hrprcyl5-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.0 GB at this context; duplicate traces of the same session: 07oc3keu, 2ggyzvmh, 3x89x43m, 90at92kf, s8gmjr6l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1754.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,754 63.5 73.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hrprcyl5",
   "source_url": "https://omlx.ai/benchmarks/performance/hrprcyl5",
   "tg_tokens": null,
   "tps": 63.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hrprcyl5-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.6 GB at this context; duplicate traces of the same session: 07oc3keu, 2ggyzvmh, 3x89x43m, 90at92kf, s8gmjr6l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1625.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,625 60.3 73.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hrprcyl5",
   "source_url": "https://omlx.ai/benchmarks/performance/hrprcyl5",
   "tg_tokens": null,
   "tps": 60.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hrprcyl5-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.7 GB at this context; duplicate traces of the same session: 07oc3keu, 2ggyzvmh, 3x89x43m, 90at92kf, s8gmjr6l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1582.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 1,582 51.7 74.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hrprcyl5",
   "source_url": "https://omlx.ai/benchmarks/performance/hrprcyl5",
   "tg_tokens": null,
   "tps": 51.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ftaf221j-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: xwuk0ux8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1159.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,159 83.0 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ftaf221j",
   "source_url": "https://omlx.ai/benchmarks/performance/ftaf221j",
   "tg_tokens": null,
   "tps": 83.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ftaf221j-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: xwuk0ux8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1514.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,514 76.8 71.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ftaf221j",
   "source_url": "https://omlx.ai/benchmarks/performance/ftaf221j",
   "tg_tokens": null,
   "tps": 76.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ftaf221j-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.0 GB at this context; duplicate traces of the same session: xwuk0ux8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1742.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,742 76.9 72.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ftaf221j",
   "source_url": "https://omlx.ai/benchmarks/performance/ftaf221j",
   "tg_tokens": null,
   "tps": 76.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ftaf221j-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.0 GB at this context; duplicate traces of the same session: xwuk0ux8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1848.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,848 68.1 73.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ftaf221j",
   "source_url": "https://omlx.ai/benchmarks/performance/ftaf221j",
   "tg_tokens": null,
   "tps": 68.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ftaf221j-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.6 GB at this context; duplicate traces of the same session: xwuk0ux8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1856.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,856 64.3 73.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ftaf221j",
   "source_url": "https://omlx.ai/benchmarks/performance/ftaf221j",
   "tg_tokens": null,
   "tps": 64.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ftaf221j-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.7 GB at this context; duplicate traces of the same session: xwuk0ux8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1810.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 1,810 52.2 74.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ftaf221j",
   "source_url": "https://omlx.ai/benchmarks/performance/ftaf221j",
   "tg_tokens": null,
   "tps": 52.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-1ba6ua5e-ctx1024",
   "model": "Ternary-Bonsai-2-4B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 4.0 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 349.3,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-4B-mlx-2bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 2bit; 1k 349.3 48.1 4.0 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ba6ua5e",
   "source_url": "https://omlx.ai/benchmarks/performance/1ba6ua5e",
   "tg_tokens": null,
   "tps": 48.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-1ba6ua5e-ctx4096",
   "model": "Ternary-Bonsai-2-4B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 5.2 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 342.0,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-4B-mlx-2bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 2bit; 4k 342.0 34.6 5.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ba6ua5e",
   "source_url": "https://omlx.ai/benchmarks/performance/1ba6ua5e",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-1ba6ua5e-ctx8192",
   "model": "Ternary-Bonsai-2-4B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 6.0 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 329.3,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-4B-mlx-2bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 2bit; 8k 329.3 26.3 6.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ba6ua5e",
   "source_url": "https://omlx.ai/benchmarks/performance/1ba6ua5e",
   "tg_tokens": null,
   "tps": 26.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-1ba6ua5e-ctx16384",
   "model": "Ternary-Bonsai-2-4B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 286.0,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-4B-mlx-2bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 2bit; 16k 286.0 18.3 6.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ba6ua5e",
   "source_url": "https://omlx.ai/benchmarks/performance/1ba6ua5e",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-6fjy06v3-ctx1024",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 177.5 GB; TTFT 777 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1318.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k tokens 1,318 PP tok/s 67.9 TG tok/s 777 TTFT (ms) 177.5 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6fjy06v3",
   "source_url": "https://omlx.ai/benchmarks/performance/6fjy06v3",
   "tg_tokens": null,
   "tps": 67.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-csy7nh23-ctx1024",
   "model": "Nemotron-3.5-Lightning-30B-A3B-refusal-ablated",
   "notes": "MTP MLX build; oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: nikxifxn; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 869.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Nemotron-3.5-Lightning-30B-A3B-refusal-ablated-MLX-oQ8-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 869.7 75.0 34.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/csy7nh23",
   "source_url": "https://omlx.ai/benchmarks/performance/csy7nh23",
   "tg_tokens": null,
   "tps": 75.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-csy7nh23-ctx4096",
   "model": "Nemotron-3.5-Lightning-30B-A3B-refusal-ablated",
   "notes": "MTP MLX build; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: nikxifxn; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 766.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Nemotron-3.5-Lightning-30B-A3B-refusal-ablated-MLX-oQ8-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 766.3 55.6 35.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/csy7nh23",
   "source_url": "https://omlx.ai/benchmarks/performance/csy7nh23",
   "tg_tokens": null,
   "tps": 55.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-l0yffwd3-ctx1024",
   "model": "Nemotron-3.5-Lightning-30B-A3B-refusal-ablated",
   "notes": "MTP MLX build; oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 861.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Nemotron-3.5-Lightning-30B-A3B-refusal-ablated-MLX-oQ8-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 861.2 79.2 34.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l0yffwd3",
   "source_url": "https://omlx.ai/benchmarks/performance/l0yffwd3",
   "tg_tokens": null,
   "tps": 79.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-l0yffwd3-ctx4096",
   "model": "Nemotron-3.5-Lightning-30B-A3B-refusal-ablated",
   "notes": "MTP MLX build; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 761.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Nemotron-3.5-Lightning-30B-A3B-refusal-ablated-MLX-oQ8-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 761.0 58.3 35.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l0yffwd3",
   "source_url": "https://omlx.ai/benchmarks/performance/l0yffwd3",
   "tg_tokens": null,
   "tps": 58.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-dip8aoc8-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.3 GB at this context; duplicate traces of the same session: uem3ghi9; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 312.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 4bit; 1k 312.5 32.7 3.3 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dip8aoc8",
   "source_url": "https://omlx.ai/benchmarks/performance/dip8aoc8",
   "tg_tokens": null,
   "tps": 32.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-dip8aoc8-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: uem3ghi9; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B on M3 (10c); M3 (10c) \u00b7 24 GB \u00b7 4bit; 4k 323.8 31.5 3.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dip8aoc8",
   "source_url": "https://omlx.ai/benchmarks/performance/dip8aoc8",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-8f2oon8d-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: yprmgm3o; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1529.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 1,529 122.3 19.4 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8f2oon8d",
   "source_url": "https://omlx.ai/benchmarks/performance/8f2oon8d",
   "tg_tokens": null,
   "tps": 122.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-8f2oon8d-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: yprmgm3o; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1669.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 1,669 117.9 20.1 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8f2oon8d",
   "source_url": "https://omlx.ai/benchmarks/performance/8f2oon8d",
   "tg_tokens": null,
   "tps": 117.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-v08xy3ed-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: e086j3yh, e3sm18t7, qotbsgo8, rven91se, xbyo8d1g; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 918.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 918.3 119.6 21.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v08xy3ed",
   "source_url": "https://omlx.ai/benchmarks/performance/v08xy3ed",
   "tg_tokens": null,
   "tps": 119.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-v08xy3ed-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: e086j3yh, e3sm18t7, qotbsgo8, rven91se, xbyo8d1g; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1022.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 1,022 83.8 22.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v08xy3ed",
   "source_url": "https://omlx.ai/benchmarks/performance/v08xy3ed",
   "tg_tokens": null,
   "tps": 83.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-v08xy3ed-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: e086j3yh, e3sm18t7, qotbsgo8, rven91se, xbyo8d1g; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 946.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 946.4 83.3 22.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v08xy3ed",
   "source_url": "https://omlx.ai/benchmarks/performance/v08xy3ed",
   "tg_tokens": null,
   "tps": 83.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-v08xy3ed-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: e086j3yh, e3sm18t7, qotbsgo8, rven91se, xbyo8d1g; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 800.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 800.4 87.8 22.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v08xy3ed",
   "source_url": "https://omlx.ai/benchmarks/performance/v08xy3ed",
   "tg_tokens": null,
   "tps": 87.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-v08xy3ed-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: e086j3yh, e3sm18t7, qotbsgo8, rven91se, xbyo8d1g; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 644.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 644.7 113.7 23.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v08xy3ed",
   "source_url": "https://omlx.ai/benchmarks/performance/v08xy3ed",
   "tg_tokens": null,
   "tps": 113.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-v08xy3ed-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: e086j3yh, e3sm18t7, qotbsgo8, rven91se, xbyo8d1g; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 495.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 495.1 64.0 25.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v08xy3ed",
   "source_url": "https://omlx.ai/benchmarks/performance/v08xy3ed",
   "tg_tokens": null,
   "tps": 64.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-rqisgsve-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 24.8 GB; TTFT 58460 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2242.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 128k tokens 2,242 PP tok/s 110.0 TG tok/s 58460 TTFT (ms) 24.8 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rqisgsve",
   "source_url": "https://omlx.ai/benchmarks/performance/rqisgsve",
   "tg_tokens": null,
   "tps": 110.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zlcu9pvl-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: g1orkj49, jq103bb6, jw4ph4uh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 828.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 828.1 18.7 28.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zlcu9pvl",
   "source_url": "https://omlx.ai/benchmarks/performance/zlcu9pvl",
   "tg_tokens": null,
   "tps": 18.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zlcu9pvl-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.3 GB at this context; duplicate traces of the same session: g1orkj49, jq103bb6, jw4ph4uh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 674.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 674.0 16.0 33.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zlcu9pvl",
   "source_url": "https://omlx.ai/benchmarks/performance/zlcu9pvl",
   "tg_tokens": null,
   "tps": 16.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zlcu9pvl-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: g1orkj49, jq103bb6, jw4ph4uh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 531.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 531.6 14.1 37.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zlcu9pvl",
   "source_url": "https://omlx.ai/benchmarks/performance/zlcu9pvl",
   "tg_tokens": null,
   "tps": 14.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zlcu9pvl-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 41.5 GB at this context; duplicate traces of the same session: g1orkj49, jq103bb6, jw4ph4uh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 431.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 195k 431.3 12.4 41.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zlcu9pvl",
   "source_url": "https://omlx.ai/benchmarks/performance/zlcu9pvl",
   "tg_tokens": null,
   "tps": 12.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-df2uqixb-ctx1024",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "MTP MLX build; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: 8bz00yeq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 395.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Abliterated-MTPLX-6bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 1k 395.6 41.8 26.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/df2uqixb",
   "source_url": "https://omlx.ai/benchmarks/performance/df2uqixb",
   "tg_tokens": null,
   "tps": 41.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-df2uqixb-ctx4096",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "MTP MLX build; oMLX community benchmark trace; peak memory 30.0 GB at this context; duplicate traces of the same session: 8bz00yeq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 454.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Abliterated-MTPLX-6bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 4k 454.7 52.1 30.0 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/df2uqixb",
   "source_url": "https://omlx.ai/benchmarks/performance/df2uqixb",
   "tg_tokens": null,
   "tps": 52.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-w28utyym-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.5 GB at this context; duplicate traces of the same session: 1at1euat; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 380.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 380.8 19.2 17.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w28utyym",
   "source_url": "https://omlx.ai/benchmarks/performance/w28utyym",
   "tg_tokens": null,
   "tps": 19.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-w28utyym-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 1at1euat; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 394.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 394.0 24.4 19.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w28utyym",
   "source_url": "https://omlx.ai/benchmarks/performance/w28utyym",
   "tg_tokens": null,
   "tps": 24.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-5ury32jd-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 17.5 GB; TTFT 2689 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 380.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k tokens 380.8 PP tok/s 36.6 TG tok/s 2689 TTFT (ms) 17.5 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ury32jd",
   "source_url": "https://omlx.ai/benchmarks/performance/5ury32jd",
   "tg_tokens": null,
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-uykke64p-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 17.5 GB; TTFT 2682 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 381.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k tokens 381.8 PP tok/s 22.0 TG tok/s 2682 TTFT (ms) 17.5 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uykke64p",
   "source_url": "https://omlx.ai/benchmarks/performance/uykke64p",
   "tg_tokens": null,
   "tps": 22.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-hzxfbusf-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 17.5 GB; TTFT 2656 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 385.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k tokens 385.6 PP tok/s 21.0 TG tok/s 2656 TTFT (ms) 17.5 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hzxfbusf",
   "source_url": "https://omlx.ai/benchmarks/performance/hzxfbusf",
   "tg_tokens": null,
   "tps": 21.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ncf1dvb0-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 17.5 GB; TTFT 2653 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 386.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k tokens 386.0 PP tok/s 31.1 TG tok/s 2653 TTFT (ms) 17.5 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ncf1dvb0",
   "source_url": "https://omlx.ai/benchmarks/performance/ncf1dvb0",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-kn6ouvh4-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 17.5 GB; TTFT 2640 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 387.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k tokens 387.9 PP tok/s 21.9 TG tok/s 2640 TTFT (ms) 17.5 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kn6ouvh4",
   "source_url": "https://omlx.ai/benchmarks/performance/kn6ouvh4",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-hc3avezn-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 17.5 GB; TTFT 2656 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 385.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k tokens 385.6 PP tok/s 20.1 TG tok/s 2656 TTFT (ms) 17.5 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hc3avezn",
   "source_url": "https://omlx.ai/benchmarks/performance/hc3avezn",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-dwizsj1f-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 50.2 GB at this context; duplicate traces of the same session: 6i1e04xr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 561.2,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 2bit; 1k 561.2 38.3 50.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dwizsj1f",
   "source_url": "https://omlx.ai/benchmarks/performance/dwizsj1f",
   "tg_tokens": null,
   "tps": 38.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-dwizsj1f-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 52.8 GB at this context; duplicate traces of the same session: 6i1e04xr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1000.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 2bit; 32k 1,000 41.5 52.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dwizsj1f",
   "source_url": "https://omlx.ai/benchmarks/performance/dwizsj1f",
   "tg_tokens": null,
   "tps": 41.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2735pb1f-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: e0blklu3, rawym94m, yywghxir; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1159.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,159 83.0 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2735pb1f",
   "source_url": "https://omlx.ai/benchmarks/performance/2735pb1f",
   "tg_tokens": null,
   "tps": 83.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2735pb1f-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: e0blklu3, rawym94m, yywghxir; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1514.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,514 76.8 71.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2735pb1f",
   "source_url": "https://omlx.ai/benchmarks/performance/2735pb1f",
   "tg_tokens": null,
   "tps": 76.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2735pb1f-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.0 GB at this context; duplicate traces of the same session: e0blklu3, rawym94m, yywghxir; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1742.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,742 76.9 72.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2735pb1f",
   "source_url": "https://omlx.ai/benchmarks/performance/2735pb1f",
   "tg_tokens": null,
   "tps": 76.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2735pb1f-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.0 GB at this context; duplicate traces of the same session: e0blklu3, rawym94m, yywghxir; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1848.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,848 68.1 73.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2735pb1f",
   "source_url": "https://omlx.ai/benchmarks/performance/2735pb1f",
   "tg_tokens": null,
   "tps": 68.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2735pb1f-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.6 GB at this context; duplicate traces of the same session: e0blklu3, rawym94m, yywghxir; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1856.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,856 64.3 73.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2735pb1f",
   "source_url": "https://omlx.ai/benchmarks/performance/2735pb1f",
   "tg_tokens": null,
   "tps": 64.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2735pb1f-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.7 GB at this context; duplicate traces of the same session: e0blklu3, rawym94m, yywghxir; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1810.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 1,810 52.2 74.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2735pb1f",
   "source_url": "https://omlx.ai/benchmarks/performance/2735pb1f",
   "tg_tokens": null,
   "tps": 52.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4z00z0jn-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: gbmy9dec, mbjfa74j, rldhy7dx, xcpmechw, xz8oip0x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 709.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 709.6 95.0 71.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4z00z0jn",
   "source_url": "https://omlx.ai/benchmarks/performance/4z00z0jn",
   "tg_tokens": null,
   "tps": 95.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4z00z0jn-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.4 GB at this context; duplicate traces of the same session: gbmy9dec, mbjfa74j, rldhy7dx, xcpmechw, xz8oip0x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2135.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,135 89.1 71.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4z00z0jn",
   "source_url": "https://omlx.ai/benchmarks/performance/4z00z0jn",
   "tg_tokens": null,
   "tps": 89.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4z00z0jn-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.0 GB at this context; duplicate traces of the same session: gbmy9dec, mbjfa74j, rldhy7dx, xcpmechw, xz8oip0x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2270.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,270 71.7 72.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4z00z0jn",
   "source_url": "https://omlx.ai/benchmarks/performance/4z00z0jn",
   "tg_tokens": null,
   "tps": 71.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4z00z0jn-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.0 GB at this context; duplicate traces of the same session: gbmy9dec, mbjfa74j, rldhy7dx, xcpmechw, xz8oip0x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2128.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,128 79.9 73.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4z00z0jn",
   "source_url": "https://omlx.ai/benchmarks/performance/4z00z0jn",
   "tg_tokens": null,
   "tps": 79.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4z00z0jn-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.6 GB at this context; duplicate traces of the same session: gbmy9dec, mbjfa74j, rldhy7dx, xcpmechw, xz8oip0x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2150.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,150 69.7 73.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4z00z0jn",
   "source_url": "https://omlx.ai/benchmarks/performance/4z00z0jn",
   "tg_tokens": null,
   "tps": 69.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4z00z0jn-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.7 GB at this context; duplicate traces of the same session: gbmy9dec, mbjfa74j, rldhy7dx, xcpmechw, xz8oip0x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2038.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,038 61.9 74.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4z00z0jn",
   "source_url": "https://omlx.ai/benchmarks/performance/4z00z0jn",
   "tg_tokens": null,
   "tps": 61.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jash27zs-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: 4htr8yfz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1852.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,852 80.1 71.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jash27zs",
   "source_url": "https://omlx.ai/benchmarks/performance/jash27zs",
   "tg_tokens": null,
   "tps": 80.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jash27zs-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.5 GB at this context; duplicate traces of the same session: 4htr8yfz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2269.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,269 70.6 73.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jash27zs",
   "source_url": "https://omlx.ai/benchmarks/performance/jash27zs",
   "tg_tokens": null,
   "tps": 70.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-my3s343e-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: 50ynzuhu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1716.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,716 74.2 71.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/my3s343e",
   "source_url": "https://omlx.ai/benchmarks/performance/my3s343e",
   "tg_tokens": null,
   "tps": 74.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-my3s343e-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.5 GB at this context; duplicate traces of the same session: 50ynzuhu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2205.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,205 69.6 73.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/my3s343e",
   "source_url": "https://omlx.ai/benchmarks/performance/my3s343e",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-047btonz-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: 31fd562f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1934.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,934 74.3 71.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/047btonz",
   "source_url": "https://omlx.ai/benchmarks/performance/047btonz",
   "tg_tokens": null,
   "tps": 74.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-047btonz-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.5 GB at this context; duplicate traces of the same session: 31fd562f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2273.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,273 78.0 73.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/047btonz",
   "source_url": "https://omlx.ai/benchmarks/performance/047btonz",
   "tg_tokens": null,
   "tps": 78.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1mnbokyp-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: gzxs5vzm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 643.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 643.9 21.7 71.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1mnbokyp",
   "source_url": "https://omlx.ai/benchmarks/performance/1mnbokyp",
   "tg_tokens": null,
   "tps": 21.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1mnbokyp-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.5 GB at this context; duplicate traces of the same session: gzxs5vzm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2001.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,001 61.9 73.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1mnbokyp",
   "source_url": "https://omlx.ai/benchmarks/performance/1mnbokyp",
   "tg_tokens": null,
   "tps": 61.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9f7dtj0x-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.4 GB at this context; duplicate traces of the same session: 12g2bpj0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1650.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,650 58.8 71.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9f7dtj0x",
   "source_url": "https://omlx.ai/benchmarks/performance/9f7dtj0x",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9f7dtj0x-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.5 GB at this context; duplicate traces of the same session: 12g2bpj0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2205.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,205 66.1 73.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9f7dtj0x",
   "source_url": "https://omlx.ai/benchmarks/performance/9f7dtj0x",
   "tg_tokens": null,
   "tps": 66.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-mf1ha1uq-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 135.5 GB at this context; duplicate traces of the same session: jmfncnp8; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4095.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 128k 4,095 92.7 135.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mf1ha1uq",
   "source_url": "https://omlx.ai/benchmarks/performance/mf1ha1uq",
   "tg_tokens": null,
   "tps": 92.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-mf1ha1uq-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 138.2 GB at this context; duplicate traces of the same session: jmfncnp8; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4146.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 195k 4,146 77.3 138.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mf1ha1uq",
   "source_url": "https://omlx.ai/benchmarks/performance/mf1ha1uq",
   "tg_tokens": null,
   "tps": 77.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-gje8wjfs-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: 2gwz6mwc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 900.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-fp16-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 8bit; 1k 900.9 52.5 34.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gje8wjfs",
   "source_url": "https://omlx.ai/benchmarks/performance/gje8wjfs",
   "tg_tokens": null,
   "tps": 52.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-gje8wjfs-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.9 GB at this context; duplicate traces of the same session: 2gwz6mwc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1555.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-fp16-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 8bit; 4k 1,555 60.3 35.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gje8wjfs",
   "source_url": "https://omlx.ai/benchmarks/performance/gje8wjfs",
   "tg_tokens": null,
   "tps": 60.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-u5d6cq81-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: 24iix5vn; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 850.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 850.2 61.6 16.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u5d6cq81",
   "source_url": "https://omlx.ai/benchmarks/performance/u5d6cq81",
   "tg_tokens": null,
   "tps": 61.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-u5d6cq81-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 24iix5vn; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 720.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 720.6 55.9 16.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u5d6cq81",
   "source_url": "https://omlx.ai/benchmarks/performance/u5d6cq81",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-kdn13jop-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: z090rnbm; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 848.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 848.7 61.9 16.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kdn13jop",
   "source_url": "https://omlx.ai/benchmarks/performance/kdn13jop",
   "tg_tokens": null,
   "tps": 61.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-kdn13jop-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: z090rnbm; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 720.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 720.8 56.5 16.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kdn13jop",
   "source_url": "https://omlx.ai/benchmarks/performance/kdn13jop",
   "tg_tokens": null,
   "tps": 56.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-s21jcd4d-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 759.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 759.0 43.0 16.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s21jcd4d",
   "source_url": "https://omlx.ai/benchmarks/performance/s21jcd4d",
   "tg_tokens": null,
   "tps": 43.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-s21jcd4d-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 700.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 700.6 49.6 16.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s21jcd4d",
   "source_url": "https://omlx.ai/benchmarks/performance/s21jcd4d",
   "tg_tokens": null,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-uayn1gun-ctx1024",
   "model": "GLM-4.7-Flash",
   "notes": "oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: 770f9hhq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1823.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-4.7-Flash-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 1,823 75.5 16.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uayn1gun",
   "source_url": "https://omlx.ai/benchmarks/performance/uayn1gun",
   "tg_tokens": null,
   "tps": 75.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-uayn1gun-ctx4096",
   "model": "GLM-4.7-Flash",
   "notes": "oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 770f9hhq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1604.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-4.7-Flash-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 1,604 61.8 17.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uayn1gun",
   "source_url": "https://omlx.ai/benchmarks/performance/uayn1gun",
   "tg_tokens": null,
   "tps": 61.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-o869smfp-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 64.5 GB; TTFT 150322 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 218.0,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "GLM-5.3-Flash-oQ2e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 32k tokens 218.0 PP tok/s 9.9 TG tok/s 150322 TTFT (ms) 64.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o869smfp",
   "source_url": "https://omlx.ai/benchmarks/performance/o869smfp",
   "tg_tokens": null,
   "tps": 9.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bsknfl1m-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 42.3 GB; TTFT 137885 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 237.6,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "GLM-5.3-Flash-oQ2e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 32k tokens 237.6 PP tok/s 5.3 TG tok/s 137885 TTFT (ms) 42.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bsknfl1m",
   "source_url": "https://omlx.ai/benchmarks/performance/bsknfl1m",
   "tg_tokens": null,
   "tps": 5.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-08gdm3vf-ctx1024",
   "model": "K2-Horizon-7B",
   "notes": "oMLX community benchmark trace; peak memory 8.0 GB at this context; duplicate traces of the same session: bigsf4d0, m65vklxh, qzd8a511; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 384.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "K2-Horizon-7B-oQ6e on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 1k 384.8 42.4 8.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/08gdm3vf",
   "source_url": "https://omlx.ai/benchmarks/performance/08gdm3vf",
   "tg_tokens": null,
   "tps": 42.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-08gdm3vf-ctx4096",
   "model": "K2-Horizon-7B",
   "notes": "oMLX community benchmark trace; peak memory 8.3 GB at this context; duplicate traces of the same session: bigsf4d0, m65vklxh, qzd8a511; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 368.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "K2-Horizon-7B-oQ6e on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 4k 368.6 39.0 8.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/08gdm3vf",
   "source_url": "https://omlx.ai/benchmarks/performance/08gdm3vf",
   "tg_tokens": null,
   "tps": 39.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-08gdm3vf-ctx8192",
   "model": "K2-Horizon-7B",
   "notes": "oMLX community benchmark trace; peak memory 8.8 GB at this context; duplicate traces of the same session: bigsf4d0, m65vklxh, qzd8a511; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 346.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "K2-Horizon-7B-oQ6e on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 8k 346.5 36.0 8.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/08gdm3vf",
   "source_url": "https://omlx.ai/benchmarks/performance/08gdm3vf",
   "tg_tokens": null,
   "tps": 36.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-08gdm3vf-ctx16384",
   "model": "K2-Horizon-7B",
   "notes": "oMLX community benchmark trace; peak memory 10.0 GB at this context; duplicate traces of the same session: bigsf4d0, m65vklxh, qzd8a511; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 303.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "K2-Horizon-7B-oQ6e on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 16k 303.7 30.5 10.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/08gdm3vf",
   "source_url": "https://omlx.ai/benchmarks/performance/08gdm3vf",
   "tg_tokens": null,
   "tps": 30.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-b1b7sx4q-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.6 GB at this context; duplicate traces of the same session: eva51t69; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4625.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 4,625 126.4 13.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b1b7sx4q",
   "source_url": "https://omlx.ai/benchmarks/performance/b1b7sx4q",
   "tg_tokens": null,
   "tps": 126.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-b1b7sx4q-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.8 GB at this context; duplicate traces of the same session: eva51t69; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4475.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 4,475 114.3 13.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b1b7sx4q",
   "source_url": "https://omlx.ai/benchmarks/performance/b1b7sx4q",
   "tg_tokens": null,
   "tps": 114.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-k37sn5x0-ctx1024",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "g64; oMLX community benchmark trace; peak memory 9.8 GB at this context; duplicate traces of the same session: 0ydlutln, 5i8tmmm4; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2085.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-mlx-6bit-g64 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 2,085 113.6 9.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k37sn5x0",
   "source_url": "https://omlx.ai/benchmarks/performance/k37sn5x0",
   "tg_tokens": null,
   "tps": 113.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-k37sn5x0-ctx4096",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "g64; oMLX community benchmark trace; peak memory 9.9 GB at this context; duplicate traces of the same session: 0ydlutln, 5i8tmmm4; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2744.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-mlx-6bit-g64 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 2,744 111.3 9.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k37sn5x0",
   "source_url": "https://omlx.ai/benchmarks/performance/k37sn5x0",
   "tg_tokens": null,
   "tps": 111.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-k37sn5x0-ctx16384",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "g64; oMLX community benchmark trace; peak memory 10.1 GB at this context; duplicate traces of the same session: 0ydlutln, 5i8tmmm4; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3450.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-mlx-6bit-g64 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 16k 3,450 102.2 10.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k37sn5x0",
   "source_url": "https://omlx.ai/benchmarks/performance/k37sn5x0",
   "tg_tokens": null,
   "tps": 102.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-6evqk8wj-ctx1024",
   "model": "Nemotron-3.5-Lightning-30B-A3B-refusal-ablated",
   "notes": "MTP MLX build; oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: ozq8ud5k; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 870.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Nemotron-3.5-Lightning-30B-A3B-refusal-ablated-MLX-oQ8-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 870.3 69.1 34.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6evqk8wj",
   "source_url": "https://omlx.ai/benchmarks/performance/6evqk8wj",
   "tg_tokens": null,
   "tps": 69.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-6evqk8wj-ctx4096",
   "model": "Nemotron-3.5-Lightning-30B-A3B-refusal-ablated",
   "notes": "MTP MLX build; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: ozq8ud5k; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 735.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Nemotron-3.5-Lightning-30B-A3B-refusal-ablated-MLX-oQ8-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 735.6 56.0 35.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6evqk8wj",
   "source_url": "https://omlx.ai/benchmarks/performance/6evqk8wj",
   "tg_tokens": null,
   "tps": 56.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-isozh7l0-ctx1024",
   "model": "Nemotron-3.5-Lightning-30B-A3B-refusal-ablated",
   "notes": "MTP MLX build; oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 861.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Nemotron-3.5-Lightning-30B-A3B-refusal-ablated-MLX-oQ8-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 861.2 79.2 34.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/isozh7l0",
   "source_url": "https://omlx.ai/benchmarks/performance/isozh7l0",
   "tg_tokens": null,
   "tps": 79.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-isozh7l0-ctx4096",
   "model": "Nemotron-3.5-Lightning-30B-A3B-refusal-ablated",
   "notes": "MTP MLX build; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 761.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Nemotron-3.5-Lightning-30B-A3B-refusal-ablated-MLX-oQ8-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 761.0 58.3 35.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/isozh7l0",
   "source_url": "https://omlx.ai/benchmarks/performance/isozh7l0",
   "tg_tokens": null,
   "tps": 58.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-x0t6bie9-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 5.4 GB at this context; duplicate traces of the same session: zocwcytw; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2105.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-4B-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 8bit; 1k 2,105 57.2 5.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x0t6bie9",
   "source_url": "https://omlx.ai/benchmarks/performance/x0t6bie9",
   "tg_tokens": null,
   "tps": 57.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-x0t6bie9-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: zocwcytw; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2643.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-4B-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 8bit; 4k 2,643 55.4 5.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x0t6bie9",
   "source_url": "https://omlx.ai/benchmarks/performance/x0t6bie9",
   "tg_tokens": null,
   "tps": 55.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-n95woegb-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: ngfyp6l9, tfzy521l; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2579.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,579 98.7 3.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n95woegb",
   "source_url": "https://omlx.ai/benchmarks/performance/n95woegb",
   "tg_tokens": null,
   "tps": 98.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-n95woegb-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: ngfyp6l9, tfzy521l; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3159.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,159 92.6 3.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n95woegb",
   "source_url": "https://omlx.ai/benchmarks/performance/n95woegb",
   "tg_tokens": null,
   "tps": 92.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-n95woegb-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: ngfyp6l9, tfzy521l; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2919.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,919 84.0 4.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n95woegb",
   "source_url": "https://omlx.ai/benchmarks/performance/n95woegb",
   "tg_tokens": null,
   "tps": 84.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-klv5jp3c-ctx1024",
   "model": "Qwen3.6-27B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: krpia7t1; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 76.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-27B-oQ4-fp16-mtp on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 76.0 13.9 19.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/klv5jp3c",
   "source_url": "https://omlx.ai/benchmarks/performance/klv5jp3c",
   "tg_tokens": null,
   "tps": 13.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-klv5jp3c-ctx4096",
   "model": "Qwen3.6-27B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: krpia7t1; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 78.2,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-27B-oQ4-fp16-mtp on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 78.2 14.9 20.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/klv5jp3c",
   "source_url": "https://omlx.ai/benchmarks/performance/klv5jp3c",
   "tg_tokens": null,
   "tps": 14.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-3fn7oej1-ctx1024",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "ET-O5.5-G6A-G4.7-DSV4-K3-RLOO MTP DFlash2 build; oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: jfriqk2k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 222.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-ET-O5.5-G6A-G4.7-DSV4-K3-RLOO-MTP-DFlash2-oQ4e on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 222.3 37.5 17.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3fn7oej1",
   "source_url": "https://omlx.ai/benchmarks/performance/3fn7oej1",
   "tg_tokens": null,
   "tps": 37.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-3fn7oej1-ctx4096",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "ET-O5.5-G6A-G4.7-DSV4-K3-RLOO MTP DFlash2 build; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: jfriqk2k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 200.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-ET-O5.5-G6A-G4.7-DSV4-K3-RLOO-MTP-DFlash2-oQ4e on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 200.3 28.4 20.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3fn7oej1",
   "source_url": "https://omlx.ai/benchmarks/performance/3fn7oej1",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-fjmjrz3h-ctx32768",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 454.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 454.8 13.0 23.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fjmjrz3h",
   "source_url": "https://omlx.ai/benchmarks/performance/fjmjrz3h",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-fjmjrz3h-ctx65536",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 394.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 394.4 11.7 24.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fjmjrz3h",
   "source_url": "https://omlx.ai/benchmarks/performance/fjmjrz3h",
   "tg_tokens": null,
   "tps": 11.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-fjmjrz3h-ctx131072",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 28.9 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 306.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 128k 306.9 9.9 28.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fjmjrz3h",
   "source_url": "https://omlx.ai/benchmarks/performance/fjmjrz3h",
   "tg_tokens": null,
   "tps": 9.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-w1wkfiri-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 8k context; peak memory 18.6 GB; TTFT 16133 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 507.8,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-27B-nvfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k tokens 507.8 PP tok/s 39.9 TG tok/s 16133 TTFT (ms) 18.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w1wkfiri",
   "source_url": "https://omlx.ai/benchmarks/performance/w1wkfiri",
   "tg_tokens": null,
   "tps": 39.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-8343cmw7-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 8k context; peak memory 18.6 GB; TTFT 16327 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 501.7,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "Qwen3.8-27B-nvfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k tokens 501.7 PP tok/s 27.4 TG tok/s 16327 TTFT (ms) 18.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8343cmw7",
   "source_url": "https://omlx.ai/benchmarks/performance/8343cmw7",
   "tg_tokens": null,
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8vfupf3n-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: gdnuat09, huhwo60t, vsk5v3n4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 914.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 914.6 31.8 16.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8vfupf3n",
   "source_url": "https://omlx.ai/benchmarks/performance/8vfupf3n",
   "tg_tokens": null,
   "tps": 31.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8vfupf3n-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: gdnuat09, huhwo60t, vsk5v3n4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 677.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 677.6 24.7 21.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8vfupf3n",
   "source_url": "https://omlx.ai/benchmarks/performance/8vfupf3n",
   "tg_tokens": null,
   "tps": 24.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8vfupf3n-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: gdnuat09, huhwo60t, vsk5v3n4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 535.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 535.8 19.9 25.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8vfupf3n",
   "source_url": "https://omlx.ai/benchmarks/performance/8vfupf3n",
   "tg_tokens": null,
   "tps": 19.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8vfupf3n-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: gdnuat09, huhwo60t, vsk5v3n4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 439.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 439.6 16.7 29.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8vfupf3n",
   "source_url": "https://omlx.ai/benchmarks/performance/8vfupf3n",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3l0j6t1z-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: fkumhqot; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 499.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 499.8 29.9 17.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3l0j6t1z",
   "source_url": "https://omlx.ai/benchmarks/performance/3l0j6t1z",
   "tg_tokens": null,
   "tps": 29.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3l0j6t1z-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: fkumhqot; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 402.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 64k 402.8 26.5 23.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3l0j6t1z",
   "source_url": "https://omlx.ai/benchmarks/performance/3l0j6t1z",
   "tg_tokens": null,
   "tps": 26.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-mk16m15h-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: uta0g1zj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 377.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 377.5 20.1 22.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mk16m15h",
   "source_url": "https://omlx.ai/benchmarks/performance/mk16m15h",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-mk16m15h-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: uta0g1zj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 395.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 395.0 25.9 23.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mk16m15h",
   "source_url": "https://omlx.ai/benchmarks/performance/mk16m15h",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-dj1eqys3-ctx4096",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 32.8 GB at this context; duplicate traces of the same session: 1ra5ieak, meu8myto; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 948.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ6e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 4k 948.0 69.7 32.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dj1eqys3",
   "source_url": "https://omlx.ai/benchmarks/performance/dj1eqys3",
   "tg_tokens": null,
   "tps": 69.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-dj1eqys3-ctx16384",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 34.1 GB at this context; duplicate traces of the same session: 1ra5ieak, meu8myto; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 834.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ6e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 16k 834.4 80.1 34.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dj1eqys3",
   "source_url": "https://omlx.ai/benchmarks/performance/dj1eqys3",
   "tg_tokens": null,
   "tps": 80.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-dj1eqys3-ctx32768",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 35.1 GB at this context; duplicate traces of the same session: 1ra5ieak, meu8myto; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 673.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ6e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 32k 673.9 70.8 35.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dj1eqys3",
   "source_url": "https://omlx.ai/benchmarks/performance/dj1eqys3",
   "tg_tokens": null,
   "tps": 70.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-uzvps31p-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: 1f6ot2yz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1626.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,626 96.7 100.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uzvps31p",
   "source_url": "https://omlx.ai/benchmarks/performance/uzvps31p",
   "tg_tokens": null,
   "tps": 96.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-uzvps31p-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.0 GB at this context; duplicate traces of the same session: 1f6ot2yz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2648.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,648 72.3 101.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uzvps31p",
   "source_url": "https://omlx.ai/benchmarks/performance/uzvps31p",
   "tg_tokens": null,
   "tps": 72.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pvhy1ajd-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: h1j6ncjg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1554.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,554 75.9 100.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pvhy1ajd",
   "source_url": "https://omlx.ai/benchmarks/performance/pvhy1ajd",
   "tg_tokens": null,
   "tps": 75.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pvhy1ajd-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.0 GB at this context; duplicate traces of the same session: h1j6ncjg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2624.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,624 77.4 101.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pvhy1ajd",
   "source_url": "https://omlx.ai/benchmarks/performance/pvhy1ajd",
   "tg_tokens": null,
   "tps": 77.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-45f4vmqe-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.5 GB at this context; duplicate traces of the same session: b53dd1ln, jbchowjc, k78rnelk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1026.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,026 55.9 70.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/45f4vmqe",
   "source_url": "https://omlx.ai/benchmarks/performance/45f4vmqe",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-45f4vmqe-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: b53dd1ln, jbchowjc, k78rnelk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1339.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,339 63.7 71.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/45f4vmqe",
   "source_url": "https://omlx.ai/benchmarks/performance/45f4vmqe",
   "tg_tokens": null,
   "tps": 63.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-45f4vmqe-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.0 GB at this context; duplicate traces of the same session: b53dd1ln, jbchowjc, k78rnelk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1537.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,537 64.2 72.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/45f4vmqe",
   "source_url": "https://omlx.ai/benchmarks/performance/45f4vmqe",
   "tg_tokens": null,
   "tps": 64.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-45f4vmqe-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.0 GB at this context; duplicate traces of the same session: b53dd1ln, jbchowjc, k78rnelk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1632.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,632 62.0 73.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/45f4vmqe",
   "source_url": "https://omlx.ai/benchmarks/performance/45f4vmqe",
   "tg_tokens": null,
   "tps": 62.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-45f4vmqe-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.6 GB at this context; duplicate traces of the same session: b53dd1ln, jbchowjc, k78rnelk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1660.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,660 57.4 73.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/45f4vmqe",
   "source_url": "https://omlx.ai/benchmarks/performance/45f4vmqe",
   "tg_tokens": null,
   "tps": 57.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-45f4vmqe-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.7 GB at this context; duplicate traces of the same session: b53dd1ln, jbchowjc, k78rnelk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1624.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 1,624 53.9 74.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/45f4vmqe",
   "source_url": "https://omlx.ai/benchmarks/performance/45f4vmqe",
   "tg_tokens": null,
   "tps": 53.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tke2uvml-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.5 GB at this context; duplicate traces of the same session: 30db55dc, 444cktxw, kuo0oqbu, qfvh49tx, ulmux3b4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1028.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,028 60.1 70.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tke2uvml",
   "source_url": "https://omlx.ai/benchmarks/performance/tke2uvml",
   "tg_tokens": null,
   "tps": 60.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tke2uvml-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: 30db55dc, 444cktxw, kuo0oqbu, qfvh49tx, ulmux3b4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1304.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,304 65.1 71.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tke2uvml",
   "source_url": "https://omlx.ai/benchmarks/performance/tke2uvml",
   "tg_tokens": null,
   "tps": 65.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tke2uvml-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.0 GB at this context; duplicate traces of the same session: 30db55dc, 444cktxw, kuo0oqbu, qfvh49tx, ulmux3b4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1529.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,529 67.0 72.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tke2uvml",
   "source_url": "https://omlx.ai/benchmarks/performance/tke2uvml",
   "tg_tokens": null,
   "tps": 67.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tke2uvml-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.0 GB at this context; duplicate traces of the same session: 30db55dc, 444cktxw, kuo0oqbu, qfvh49tx, ulmux3b4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1626.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,626 60.5 73.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tke2uvml",
   "source_url": "https://omlx.ai/benchmarks/performance/tke2uvml",
   "tg_tokens": null,
   "tps": 60.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tke2uvml-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.6 GB at this context; duplicate traces of the same session: 30db55dc, 444cktxw, kuo0oqbu, qfvh49tx, ulmux3b4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1664.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,664 61.2 73.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tke2uvml",
   "source_url": "https://omlx.ai/benchmarks/performance/tke2uvml",
   "tg_tokens": null,
   "tps": 61.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tke2uvml-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.7 GB at this context; duplicate traces of the same session: 30db55dc, 444cktxw, kuo0oqbu, qfvh49tx, ulmux3b4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1656.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 1,656 51.2 74.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tke2uvml",
   "source_url": "https://omlx.ai/benchmarks/performance/tke2uvml",
   "tg_tokens": null,
   "tps": 51.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nctpesw4-ctx1024",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: bxzionca, ld348fxg, whc6x912; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 803.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 803.6 32.1 16.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nctpesw4",
   "source_url": "https://omlx.ai/benchmarks/performance/nctpesw4",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nctpesw4-ctx65536",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: bxzionca, ld348fxg, whc6x912; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 691.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 691.8 24.9 21.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nctpesw4",
   "source_url": "https://omlx.ai/benchmarks/performance/nctpesw4",
   "tg_tokens": null,
   "tps": 24.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nctpesw4-ctx131072",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: bxzionca, ld348fxg, whc6x912; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 540.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 540.4 20.0 25.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nctpesw4",
   "source_url": "https://omlx.ai/benchmarks/performance/nctpesw4",
   "tg_tokens": null,
   "tps": 20.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-nctpesw4-ctx199680",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: bxzionca, ld348fxg, whc6x912; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 439.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 439.1 16.7 29.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nctpesw4",
   "source_url": "https://omlx.ai/benchmarks/performance/nctpesw4",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-jo3uyan8-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 759.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 759.0 43.0 16.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jo3uyan8",
   "source_url": "https://omlx.ai/benchmarks/performance/jo3uyan8",
   "tg_tokens": null,
   "tps": 43.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-jo3uyan8-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 700.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 700.6 49.6 16.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jo3uyan8",
   "source_url": "https://omlx.ai/benchmarks/performance/jo3uyan8",
   "tg_tokens": null,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2k0fhcl8-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: 0c0k2fka, hnupmrj6, o7ipkejf, p9o1iknl, phzr3slb, wg4l7dc3, zifj70f6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2220.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-24B-A2B-oQ8e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 2,220 96.4 24.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2k0fhcl8",
   "source_url": "https://omlx.ai/benchmarks/performance/2k0fhcl8",
   "tg_tokens": null,
   "tps": 96.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2k0fhcl8-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: 0c0k2fka, hnupmrj6, o7ipkejf, p9o1iknl, phzr3slb, wg4l7dc3, zifj70f6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3724.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-24B-A2B-oQ8e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 3,724 90.9 24.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2k0fhcl8",
   "source_url": "https://omlx.ai/benchmarks/performance/2k0fhcl8",
   "tg_tokens": null,
   "tps": 90.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2k0fhcl8-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: 0c0k2fka, hnupmrj6, o7ipkejf, p9o1iknl, phzr3slb, wg4l7dc3, zifj70f6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3189.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-24B-A2B-oQ8e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 3,189 103.9 25.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2k0fhcl8",
   "source_url": "https://omlx.ai/benchmarks/performance/2k0fhcl8",
   "tg_tokens": null,
   "tps": 103.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2k0fhcl8-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: 0c0k2fka, hnupmrj6, o7ipkejf, p9o1iknl, phzr3slb, wg4l7dc3, zifj70f6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3339.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-24B-A2B-oQ8e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 3,339 89.1 26.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2k0fhcl8",
   "source_url": "https://omlx.ai/benchmarks/performance/2k0fhcl8",
   "tg_tokens": null,
   "tps": 89.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2k0fhcl8-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 28.8 GB at this context; duplicate traces of the same session: 0c0k2fka, hnupmrj6, o7ipkejf, p9o1iknl, phzr3slb, wg4l7dc3, zifj70f6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3259.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-24B-A2B-oQ8e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 3,259 99.0 28.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2k0fhcl8",
   "source_url": "https://omlx.ai/benchmarks/performance/2k0fhcl8",
   "tg_tokens": null,
   "tps": 99.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2k0fhcl8-ctx65536",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 33.3 GB at this context; duplicate traces of the same session: 0c0k2fka, hnupmrj6, o7ipkejf, p9o1iknl, phzr3slb, wg4l7dc3, zifj70f6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2999.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-24B-A2B-oQ8e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 2,999 72.5 33.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2k0fhcl8",
   "source_url": "https://omlx.ai/benchmarks/performance/2k0fhcl8",
   "tg_tokens": null,
   "tps": 72.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2k0fhcl8-ctx131072",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 42.9 GB at this context; duplicate traces of the same session: 0c0k2fka, hnupmrj6, o7ipkejf, p9o1iknl, phzr3slb, wg4l7dc3, zifj70f6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2246.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-24B-A2B-oQ8e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 2,246 61.6 42.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2k0fhcl8",
   "source_url": "https://omlx.ai/benchmarks/performance/2k0fhcl8",
   "tg_tokens": null,
   "tps": 61.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2k0fhcl8-ctx199680",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 53.0 GB at this context; duplicate traces of the same session: 0c0k2fka, hnupmrj6, o7ipkejf, p9o1iknl, phzr3slb, wg4l7dc3, zifj70f6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1847.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-24B-A2B-oQ8e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 195k 1,847 46.3 53.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2k0fhcl8",
   "source_url": "https://omlx.ai/benchmarks/performance/2k0fhcl8",
   "tg_tokens": null,
   "tps": 46.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9y7f04kq-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: 77k5u5oa, ngq3ic6j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3571.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "Ornith-1.5-35B-A3B-oQ5-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 3,571 151.5 26.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9y7f04kq",
   "source_url": "https://omlx.ai/benchmarks/performance/9y7f04kq",
   "tg_tokens": null,
   "tps": 151.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9y7f04kq-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: 77k5u5oa, ngq3ic6j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4876.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "Ornith-1.5-35B-A3B-oQ5-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 4,876 146.0 26.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9y7f04kq",
   "source_url": "https://omlx.ai/benchmarks/performance/9y7f04kq",
   "tg_tokens": null,
   "tps": 146.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9y7f04kq-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: 77k5u5oa, ngq3ic6j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4848.0,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "Ornith-1.5-35B-A3B-oQ5-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 4,848 170.6 26.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9y7f04kq",
   "source_url": "https://omlx.ai/benchmarks/performance/9y7f04kq",
   "tg_tokens": null,
   "tps": 170.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ox5guwhg-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.0 GB at this context; duplicate traces of the same session: ld0czu28; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3482.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Ornith-1.5-35B-A3B-oQ6-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 3,482 133.1 30.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ox5guwhg",
   "source_url": "https://omlx.ai/benchmarks/performance/ox5guwhg",
   "tg_tokens": null,
   "tps": 133.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ox5guwhg-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.7 GB at this context; duplicate traces of the same session: ld0czu28; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4801.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Ornith-1.5-35B-A3B-oQ6-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 4,801 142.5 30.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ox5guwhg",
   "source_url": "https://omlx.ai/benchmarks/performance/ox5guwhg",
   "tg_tokens": null,
   "tps": 142.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1ehpy2zo-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.0 GB at this context; duplicate traces of the same session: rng9zdse; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1903.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Ornith-1.5-35B-A3B-oQ6-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 1,903 130.6 30.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ehpy2zo",
   "source_url": "https://omlx.ai/benchmarks/performance/1ehpy2zo",
   "tg_tokens": null,
   "tps": 130.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1ehpy2zo-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.7 GB at this context; duplicate traces of the same session: rng9zdse; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4813.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Ornith-1.5-35B-A3B-oQ6-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 4,813 143.6 30.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ehpy2zo",
   "source_url": "https://omlx.ai/benchmarks/performance/1ehpy2zo",
   "tg_tokens": null,
   "tps": 143.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-hvbuv7sh-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.0 GB at this context; duplicate traces of the same session: 8y0i5nve, tqfd6ajw; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 339.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 16 GB \u00b7 4bit; 4k 339.9 54.2 4.0 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hvbuv7sh",
   "source_url": "https://omlx.ai/benchmarks/performance/hvbuv7sh",
   "tg_tokens": null,
   "tps": 54.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-hvbuv7sh-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: 8y0i5nve, tqfd6ajw; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 318.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 16 GB \u00b7 4bit; 8k 318.8 51.1 4.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hvbuv7sh",
   "source_url": "https://omlx.ai/benchmarks/performance/hvbuv7sh",
   "tg_tokens": null,
   "tps": 51.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-hvbuv7sh-ctx16384",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace; peak memory 4.7 GB at this context; duplicate traces of the same session: 8y0i5nve, tqfd6ajw; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 282.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 16 GB \u00b7 4bit; 16k 282.3 45.5 4.7 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hvbuv7sh",
   "source_url": "https://omlx.ai/benchmarks/performance/hvbuv7sh",
   "tg_tokens": null,
   "tps": 45.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-bsgcgb0n-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: 4re5uma6, ot58unav, pytvxiom; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 194.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 16 GB \u00b7 4bit; 1k 194.1 34.0 5.9 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bsgcgb0n",
   "source_url": "https://omlx.ai/benchmarks/performance/bsgcgb0n",
   "tg_tokens": null,
   "tps": 34.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-bsgcgb0n-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: 4re5uma6, ot58unav, pytvxiom; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 188.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 16 GB \u00b7 4bit; 4k 188.5 32.3 6.4 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bsgcgb0n",
   "source_url": "https://omlx.ai/benchmarks/performance/bsgcgb0n",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-bsgcgb0n-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.6 GB at this context; duplicate traces of the same session: 4re5uma6, ot58unav, pytvxiom; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 180.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 16 GB \u00b7 4bit; 8k 180.0 31.7 6.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bsgcgb0n",
   "source_url": "https://omlx.ai/benchmarks/performance/bsgcgb0n",
   "tg_tokens": null,
   "tps": 31.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-bsgcgb0n-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.1 GB at this context; duplicate traces of the same session: 4re5uma6, ot58unav, pytvxiom; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 167.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 16 GB \u00b7 4bit; 16k 167.4 29.5 7.1 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bsgcgb0n",
   "source_url": "https://omlx.ai/benchmarks/performance/bsgcgb0n",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-l8dk2xg3-ctx4096",
   "model": "Qwen3.6-14B-A3B-FableVibes",
   "notes": "MLX; oMLX community benchmark trace; peak memory 9.1 GB at this context; duplicate traces of the same session: 4p1pjx31; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 543.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-14B-A3B-FableVibes-mlx-q4 on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 543.7 44.6 9.1 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l8dk2xg3",
   "source_url": "https://omlx.ai/benchmarks/performance/l8dk2xg3",
   "tg_tokens": null,
   "tps": 44.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-l8dk2xg3-ctx16384",
   "model": "Qwen3.6-14B-A3B-FableVibes",
   "notes": "MLX; oMLX community benchmark trace; peak memory 9.5 GB at this context; duplicate traces of the same session: 4p1pjx31; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 461.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-14B-A3B-FableVibes-mlx-q4 on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 16k 461.7 36.9 9.5 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l8dk2xg3",
   "source_url": "https://omlx.ai/benchmarks/performance/l8dk2xg3",
   "tg_tokens": null,
   "tps": 36.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-msjlswgy-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1688.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,688 46.3 23.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/msjlswgy",
   "source_url": "https://omlx.ai/benchmarks/performance/msjlswgy",
   "tg_tokens": null,
   "tps": 46.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-msjlswgy-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1482.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,482 132.3 23.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/msjlswgy",
   "source_url": "https://omlx.ai/benchmarks/performance/msjlswgy",
   "tg_tokens": null,
   "tps": 132.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-bdy73r7d-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: 7hg48450, bovgo0lp, db7ox8ke, iy5n8sc4, xk5uvtym; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 918.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 918.0 82.4 21.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bdy73r7d",
   "source_url": "https://omlx.ai/benchmarks/performance/bdy73r7d",
   "tg_tokens": null,
   "tps": 82.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-bdy73r7d-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 7hg48450, bovgo0lp, db7ox8ke, iy5n8sc4, xk5uvtym; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1031.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 1,031 69.1 21.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bdy73r7d",
   "source_url": "https://omlx.ai/benchmarks/performance/bdy73r7d",
   "tg_tokens": null,
   "tps": 69.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-bdy73r7d-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 7hg48450, bovgo0lp, db7ox8ke, iy5n8sc4, xk5uvtym; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 982.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 982.9 66.4 22.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bdy73r7d",
   "source_url": "https://omlx.ai/benchmarks/performance/bdy73r7d",
   "tg_tokens": null,
   "tps": 66.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-bdy73r7d-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: 7hg48450, bovgo0lp, db7ox8ke, iy5n8sc4, xk5uvtym; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 886.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 886.5 85.3 22.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bdy73r7d",
   "source_url": "https://omlx.ai/benchmarks/performance/bdy73r7d",
   "tg_tokens": null,
   "tps": 85.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-bdy73r7d-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: 7hg48450, bovgo0lp, db7ox8ke, iy5n8sc4, xk5uvtym; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 736.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 736.4 62.7 23.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bdy73r7d",
   "source_url": "https://omlx.ai/benchmarks/performance/bdy73r7d",
   "tg_tokens": null,
   "tps": 62.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-bdy73r7d-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: 7hg48450, bovgo0lp, db7ox8ke, iy5n8sc4, xk5uvtym; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 538.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 538.7 55.4 25.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bdy73r7d",
   "source_url": "https://omlx.ai/benchmarks/performance/bdy73r7d",
   "tg_tokens": null,
   "tps": 55.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-a0ib4ksu-ctx32768",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: nece81ac; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 454.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 454.8 13.0 23.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a0ib4ksu",
   "source_url": "https://omlx.ai/benchmarks/performance/a0ib4ksu",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-a0ib4ksu-ctx65536",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: nece81ac; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 394.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 394.4 11.7 24.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a0ib4ksu",
   "source_url": "https://omlx.ai/benchmarks/performance/a0ib4ksu",
   "tg_tokens": null,
   "tps": 11.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-a0ib4ksu-ctx131072",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 28.9 GB at this context; duplicate traces of the same session: nece81ac; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 306.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 128k 306.9 9.9 28.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a0ib4ksu",
   "source_url": "https://omlx.ai/benchmarks/performance/a0ib4ksu",
   "tg_tokens": null,
   "tps": 9.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-grcc711n-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 3jxdmelq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 305.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 305.2 34.5 18.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/grcc711n",
   "source_url": "https://omlx.ai/benchmarks/performance/grcc711n",
   "tg_tokens": null,
   "tps": 34.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-grcc711n-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 3jxdmelq; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 346.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 346.4 65.8 21.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/grcc711n",
   "source_url": "https://omlx.ai/benchmarks/performance/grcc711n",
   "tg_tokens": null,
   "tps": 65.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-6azwfqdu-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: xqvcjnn7; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.4,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "Qwen3.8-27B-oQ2e-fp16-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 2bit; 1k 102.4 27.0 14.4 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6azwfqdu",
   "source_url": "https://omlx.ai/benchmarks/performance/6azwfqdu",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-6azwfqdu-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 15.4 GB at this context; duplicate traces of the same session: xqvcjnn7; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 101.6,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "Qwen3.8-27B-oQ2e-fp16-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 2bit; 4k 101.6 22.4 15.4 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6azwfqdu",
   "source_url": "https://omlx.ai/benchmarks/performance/6azwfqdu",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ydc4k8lg-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: 0uypt472, 2cmjs8sv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 498.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 498.3 25.2 18.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ydc4k8lg",
   "source_url": "https://omlx.ai/benchmarks/performance/ydc4k8lg",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ydc4k8lg-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: 0uypt472, 2cmjs8sv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 438.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 438.3 41.0 19.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ydc4k8lg",
   "source_url": "https://omlx.ai/benchmarks/performance/ydc4k8lg",
   "tg_tokens": null,
   "tps": 41.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ydc4k8lg-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: 0uypt472, 2cmjs8sv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 345.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 345.9 23.4 22.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ydc4k8lg",
   "source_url": "https://omlx.ai/benchmarks/performance/ydc4k8lg",
   "tg_tokens": null,
   "tps": 23.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ggw3qffp-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.5 GB at this context; duplicate traces of the same session: aplv4nzb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1026.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,026 55.9 70.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ggw3qffp",
   "source_url": "https://omlx.ai/benchmarks/performance/ggw3qffp",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ggw3qffp-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: aplv4nzb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1339.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,339 63.7 71.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ggw3qffp",
   "source_url": "https://omlx.ai/benchmarks/performance/ggw3qffp",
   "tg_tokens": null,
   "tps": 63.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ggw3qffp-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.0 GB at this context; duplicate traces of the same session: aplv4nzb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1537.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,537 64.2 72.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ggw3qffp",
   "source_url": "https://omlx.ai/benchmarks/performance/ggw3qffp",
   "tg_tokens": null,
   "tps": 64.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ggw3qffp-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.0 GB at this context; duplicate traces of the same session: aplv4nzb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1632.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,632 62.0 73.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ggw3qffp",
   "source_url": "https://omlx.ai/benchmarks/performance/ggw3qffp",
   "tg_tokens": null,
   "tps": 62.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ggw3qffp-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.6 GB at this context; duplicate traces of the same session: aplv4nzb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1660.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,660 57.4 73.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ggw3qffp",
   "source_url": "https://omlx.ai/benchmarks/performance/ggw3qffp",
   "tg_tokens": null,
   "tps": 57.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ggw3qffp-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.7 GB at this context; duplicate traces of the same session: aplv4nzb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1624.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 1,624 53.9 74.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ggw3qffp",
   "source_url": "https://omlx.ai/benchmarks/performance/ggw3qffp",
   "tg_tokens": null,
   "tps": 53.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-e1x96p86-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 23.0 GB; TTFT 10130 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 101.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k tokens 101.1 PP tok/s 12.1 TG tok/s 10130 TTFT (ms) 23.0 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e1x96p86",
   "source_url": "https://omlx.ai/benchmarks/performance/e1x96p86",
   "tg_tokens": null,
   "tps": 12.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-wc9qgn5f-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.1 GB at this context; duplicate traces of the same session: vd9fsi4l, wdy4ak9l; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1021.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 1,021 90.2 183.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wc9qgn5f",
   "source_url": "https://omlx.ai/benchmarks/performance/wc9qgn5f",
   "tg_tokens": null,
   "tps": 90.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-wc9qgn5f-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.7 GB at this context; duplicate traces of the same session: vd9fsi4l, wdy4ak9l; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1140.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 1,140 82.9 183.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wc9qgn5f",
   "source_url": "https://omlx.ai/benchmarks/performance/wc9qgn5f",
   "tg_tokens": null,
   "tps": 82.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-wc9qgn5f-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 184.1 GB at this context; duplicate traces of the same session: vd9fsi4l, wdy4ak9l; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1367.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 1,367 84.8 184.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wc9qgn5f",
   "source_url": "https://omlx.ai/benchmarks/performance/wc9qgn5f",
   "tg_tokens": null,
   "tps": 84.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-wc9qgn5f-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 184.4 GB at this context; duplicate traces of the same session: vd9fsi4l, wdy4ak9l; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1357.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k 1,357 78.1 184.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wc9qgn5f",
   "source_url": "https://omlx.ai/benchmarks/performance/wc9qgn5f",
   "tg_tokens": null,
   "tps": 78.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-us093gw2-ctx1024",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 7.4 GB at this context; duplicate traces of the same session: zkz4dqgk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1176.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,176 63.8 7.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/us093gw2",
   "source_url": "https://omlx.ai/benchmarks/performance/us093gw2",
   "tg_tokens": null,
   "tps": 63.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-us093gw2-ctx65536",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 10.9 GB at this context; duplicate traces of the same session: zkz4dqgk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1307.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 1,307 56.1 10.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/us093gw2",
   "source_url": "https://omlx.ai/benchmarks/performance/us093gw2",
   "tg_tokens": null,
   "tps": 56.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-us093gw2-ctx131072",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 14.5 GB at this context; duplicate traces of the same session: zkz4dqgk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1023.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 1,023 49.3 14.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/us093gw2",
   "source_url": "https://omlx.ai/benchmarks/performance/us093gw2",
   "tg_tokens": null,
   "tps": 49.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-us093gw2-ctx199680",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 18.2 GB at this context; duplicate traces of the same session: zkz4dqgk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 821.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 821.3 43.5 18.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/us093gw2",
   "source_url": "https://omlx.ai/benchmarks/performance/us093gw2",
   "tg_tokens": null,
   "tps": 43.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-s3n0qqnk-ctx1024",
   "model": "gemma-4-12B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 11.3 GB at this context; duplicate traces of the same session: dic4cboy, ma2za9oo, p6kfyuh6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1704.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-qat-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,704 44.1 11.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s3n0qqnk",
   "source_url": "https://omlx.ai/benchmarks/performance/s3n0qqnk",
   "tg_tokens": null,
   "tps": 44.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-s3n0qqnk-ctx65536",
   "model": "gemma-4-12B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 14.9 GB at this context; duplicate traces of the same session: dic4cboy, ma2za9oo, p6kfyuh6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1278.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-qat-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 1,278 40.5 14.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s3n0qqnk",
   "source_url": "https://omlx.ai/benchmarks/performance/s3n0qqnk",
   "tg_tokens": null,
   "tps": 40.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-s3n0qqnk-ctx131072",
   "model": "gemma-4-12B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: dic4cboy, ma2za9oo, p6kfyuh6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1002.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-qat-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 1,002 36.5 18.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s3n0qqnk",
   "source_url": "https://omlx.ai/benchmarks/performance/s3n0qqnk",
   "tg_tokens": null,
   "tps": 36.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-s3n0qqnk-ctx199680",
   "model": "gemma-4-12B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: dic4cboy, ma2za9oo, p6kfyuh6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 809.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-qat-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 809.9 33.4 22.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s3n0qqnk",
   "source_url": "https://omlx.ai/benchmarks/performance/s3n0qqnk",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-o1qcgfwi-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: 1sr17ely, 8f9ealbg, d37nlmxe; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3577.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 3,577 123.6 14.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o1qcgfwi",
   "source_url": "https://omlx.ai/benchmarks/performance/o1qcgfwi",
   "tg_tokens": null,
   "tps": 123.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-o1qcgfwi-ctx65536",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 1sr17ely, 8f9ealbg, d37nlmxe; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2699.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,699 88.5 17.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o1qcgfwi",
   "source_url": "https://omlx.ai/benchmarks/performance/o1qcgfwi",
   "tg_tokens": null,
   "tps": 88.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-o1qcgfwi-ctx131072",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 1sr17ely, 8f9ealbg, d37nlmxe; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1784.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 1,784 65.0 20.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o1qcgfwi",
   "source_url": "https://omlx.ai/benchmarks/performance/o1qcgfwi",
   "tg_tokens": null,
   "tps": 65.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-o1qcgfwi-ctx199680",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: 1sr17ely, 8f9ealbg, d37nlmxe; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1297.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 1,297 48.3 24.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o1qcgfwi",
   "source_url": "https://omlx.ai/benchmarks/performance/o1qcgfwi",
   "tg_tokens": null,
   "tps": 48.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-f42x1osf-ctx32768",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: chwjtbps, qcl9e3vt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1956.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-oQ8e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 32k 1,956 53.0 27.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f42x1osf",
   "source_url": "https://omlx.ai/benchmarks/performance/f42x1osf",
   "tg_tokens": null,
   "tps": 53.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-f42x1osf-ctx65536",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 29.1 GB at this context; duplicate traces of the same session: chwjtbps, qcl9e3vt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1628.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-oQ8e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 64k 1,628 40.0 29.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f42x1osf",
   "source_url": "https://omlx.ai/benchmarks/performance/f42x1osf",
   "tg_tokens": null,
   "tps": 40.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-f42x1osf-ctx131072",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 32.4 GB at this context; duplicate traces of the same session: chwjtbps, qcl9e3vt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1207.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-oQ8e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 128k 1,207 26.5 32.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f42x1osf",
   "source_url": "https://omlx.ai/benchmarks/performance/f42x1osf",
   "tg_tokens": null,
   "tps": 26.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-gsip5kbw-ctx4096",
   "model": "gemma-4-E2B-it",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 2.8 GB; TTFT 2228 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1839.0,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "gemma-4-E2B-it-oQ3 on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 3bit; 4k tokens 1,839 PP tok/s 66.0 TG tok/s 2228 TTFT (ms) 2.8 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gsip5kbw",
   "source_url": "https://omlx.ai/benchmarks/performance/gsip5kbw",
   "tg_tokens": null,
   "tps": 66.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-r33mxrgl-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Huihui abliterated build; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: y3cvrxvy, zxth0w3o; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2398.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.8-Flash-Next-abliterated-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 4bit; 4k 2,398 97.6 71.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r33mxrgl",
   "source_url": "https://omlx.ai/benchmarks/performance/r33mxrgl",
   "tg_tokens": null,
   "tps": 97.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-r33mxrgl-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Huihui abliterated build; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: y3cvrxvy, zxth0w3o; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4713.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.8-Flash-Next-abliterated-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 4bit; 16k 4,713 89.1 72.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r33mxrgl",
   "source_url": "https://omlx.ai/benchmarks/performance/r33mxrgl",
   "tg_tokens": null,
   "tps": 89.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-r33mxrgl-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Huihui abliterated build; oMLX community benchmark trace; peak memory 76.7 GB at this context; duplicate traces of the same session: y3cvrxvy, zxth0w3o; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2916.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.8-Flash-Next-abliterated-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 4bit; 128k 2,916 97.2 76.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r33mxrgl",
   "source_url": "https://omlx.ai/benchmarks/performance/r33mxrgl",
   "tg_tokens": null,
   "tps": 97.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-c5rj4yx5-ctx32768",
   "model": "K2-Horizon-MoVA-36B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 43.8 GB at this context; duplicate traces of the same session: 3zc5w32b, xqykvdvd; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 791.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "K2-Horizon-MoVA-36B-A4B-oQ8e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 32k 791.2 33.0 43.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c5rj4yx5",
   "source_url": "https://omlx.ai/benchmarks/performance/c5rj4yx5",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-c5rj4yx5-ctx65536",
   "model": "K2-Horizon-MoVA-36B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 49.8 GB at this context; duplicate traces of the same session: 3zc5w32b, xqykvdvd; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 540.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "K2-Horizon-MoVA-36B-A4B-oQ8e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 64k 540.6 24.5 49.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c5rj4yx5",
   "source_url": "https://omlx.ai/benchmarks/performance/c5rj4yx5",
   "tg_tokens": null,
   "tps": 24.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-c5rj4yx5-ctx131072",
   "model": "K2-Horizon-MoVA-36B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 61.8 GB at this context; duplicate traces of the same session: 3zc5w32b, xqykvdvd; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 332.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "K2-Horizon-MoVA-36B-A4B-oQ8e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 128k 332.2 16.1 61.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c5rj4yx5",
   "source_url": "https://omlx.ai/benchmarks/performance/c5rj4yx5",
   "tg_tokens": null,
   "tps": 16.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-vkjwj36v-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: kqfkhafu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4816.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,816 172.0 13.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vkjwj36v",
   "source_url": "https://omlx.ai/benchmarks/performance/vkjwj36v",
   "tg_tokens": null,
   "tps": 172.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vkjwj36v-ctx131072",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: kqfkhafu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4538.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 4,538 84.1 16.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vkjwj36v",
   "source_url": "https://omlx.ai/benchmarks/performance/vkjwj36v",
   "tg_tokens": null,
   "tps": 84.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-3rq6vizv-ctx1024",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: s18aot1j; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 479.6,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-mlx-mxfp4 on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 479.6 61.8 6.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3rq6vizv",
   "source_url": "https://omlx.ai/benchmarks/performance/3rq6vizv",
   "tg_tokens": null,
   "tps": 61.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-3rq6vizv-ctx4096",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: s18aot1j; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 453.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-mlx-mxfp4 on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 453.0 59.8 6.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3rq6vizv",
   "source_url": "https://omlx.ai/benchmarks/performance/3rq6vizv",
   "tg_tokens": null,
   "tps": 59.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-3rq6vizv-ctx16384",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: s18aot1j; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 424.9,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-mlx-mxfp4 on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k 424.9 52.7 6.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3rq6vizv",
   "source_url": "https://omlx.ai/benchmarks/performance/3rq6vizv",
   "tg_tokens": null,
   "tps": 52.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-1zy7ilj6-ctx1024",
   "model": "MiMo-V2.6-Distill-Qwen-9B-Ablitrated",
   "notes": "oMLX community benchmark trace; peak memory 6.0 GB at this context; duplicate traces of the same session: df8huxsg, g26gvnxa, wz0k122h; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 187.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-Ablitrated-oQ4e on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 187.3 30.5 6.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1zy7ilj6",
   "source_url": "https://omlx.ai/benchmarks/performance/1zy7ilj6",
   "tg_tokens": null,
   "tps": 30.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-1zy7ilj6-ctx4096",
   "model": "MiMo-V2.6-Distill-Qwen-9B-Ablitrated",
   "notes": "oMLX community benchmark trace; peak memory 6.6 GB at this context; duplicate traces of the same session: df8huxsg, g26gvnxa, wz0k122h; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 179.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-Ablitrated-oQ4e on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 179.1 26.4 6.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1zy7ilj6",
   "source_url": "https://omlx.ai/benchmarks/performance/1zy7ilj6",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-1zy7ilj6-ctx8192",
   "model": "MiMo-V2.6-Distill-Qwen-9B-Ablitrated",
   "notes": "oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: df8huxsg, g26gvnxa, wz0k122h; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 191.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-Ablitrated-oQ4e on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 8k 191.8 29.1 7.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1zy7ilj6",
   "source_url": "https://omlx.ai/benchmarks/performance/1zy7ilj6",
   "tg_tokens": null,
   "tps": 29.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-1zy7ilj6-ctx16384",
   "model": "MiMo-V2.6-Distill-Qwen-9B-Ablitrated",
   "notes": "oMLX community benchmark trace; peak memory 7.1 GB at this context; duplicate traces of the same session: df8huxsg, g26gvnxa, wz0k122h; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 181.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-Ablitrated-oQ4e on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k 181.8 27.9 7.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1zy7ilj6",
   "source_url": "https://omlx.ai/benchmarks/performance/1zy7ilj6",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fvo18g64-ctx1024",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: wx917wci; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 984.1,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "MiniCPM-V-4.7-35B-A3B-mxfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 984.1 61.6 22.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fvo18g64",
   "source_url": "https://omlx.ai/benchmarks/performance/fvo18g64",
   "tg_tokens": null,
   "tps": 61.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fvo18g64-ctx4096",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: wx917wci; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2303.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "MiniCPM-V-4.7-35B-A3B-mxfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,303 60.7 22.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fvo18g64",
   "source_url": "https://omlx.ai/benchmarks/performance/fvo18g64",
   "tg_tokens": null,
   "tps": 60.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-hrk8f4ah-ctx65536",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 23.9 GB; TTFT 35589 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1842.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "MiniCPM-V-4.7-35B-A3B-mxfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k tokens 1,842 PP tok/s 44.7 TG tok/s 35589 TTFT (ms) 23.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hrk8f4ah",
   "source_url": "https://omlx.ai/benchmarks/performance/hrk8f4ah",
   "tg_tokens": null,
   "tps": 44.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-byt5q16b-ctx16384",
   "model": "Qwen3.6-27B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 19.2 GB; TTFT 214563 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 76.4,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-27B-oQ4-fp16-mtp on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k tokens 76.4 PP tok/s 8.7 TG tok/s 214563 TTFT (ms) 19.2 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/byt5q16b",
   "source_url": "https://omlx.ai/benchmarks/performance/byt5q16b",
   "tg_tokens": null,
   "tps": 8.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-joe127m9-ctx16384",
   "model": "Qwen3.6-27B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 21.7 GB; TTFT 221260 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 74.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-27B-oQ4-fp16-mtp on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k tokens 74.0 PP tok/s 14.6 TG tok/s 221260 TTFT (ms) 21.7 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/joe127m9",
   "source_url": "https://omlx.ai/benchmarks/performance/joe127m9",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-whvmsuxd-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.9 GB at this context; duplicate traces of the same session: tdre32h5; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1986.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 1,986 110.9 35.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/whvmsuxd",
   "source_url": "https://omlx.ai/benchmarks/performance/whvmsuxd",
   "tg_tokens": null,
   "tps": 110.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-whvmsuxd-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 41.9 GB at this context; duplicate traces of the same session: tdre32h5; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 919.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 128k 919.6 54.7 41.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/whvmsuxd",
   "source_url": "https://omlx.ai/benchmarks/performance/whvmsuxd",
   "tg_tokens": null,
   "tps": 54.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-1qlfypf8-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 24.6 GB; TTFT 26928 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1217.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k tokens 1,217 PP tok/s 126.3 TG tok/s 26928 TTFT (ms) 24.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1qlfypf8",
   "source_url": "https://omlx.ai/benchmarks/performance/1qlfypf8",
   "tg_tokens": null,
   "tps": 126.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-1vu2n1jq-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1688.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,688 46.3 23.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1vu2n1jq",
   "source_url": "https://omlx.ai/benchmarks/performance/1vu2n1jq",
   "tg_tokens": null,
   "tps": 46.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-1vu2n1jq-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1482.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,482 132.3 23.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1vu2n1jq",
   "source_url": "https://omlx.ai/benchmarks/performance/1vu2n1jq",
   "tg_tokens": null,
   "tps": 132.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-52e4hsi8-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: zzjhrx23; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 518.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 518.1 16.5 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/52e4hsi8",
   "source_url": "https://omlx.ai/benchmarks/performance/52e4hsi8",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-52e4hsi8-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: zzjhrx23; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 625.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 625.7 16.8 18.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/52e4hsi8",
   "source_url": "https://omlx.ai/benchmarks/performance/52e4hsi8",
   "tg_tokens": null,
   "tps": 16.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-g5edguux-ctx1024",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "MTP text build; oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: gmzky5s1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 222.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-oQ4e-MTP-Text on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 222.0 36.4 17.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g5edguux",
   "source_url": "https://omlx.ai/benchmarks/performance/g5edguux",
   "tg_tokens": null,
   "tps": 36.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-g5edguux-ctx4096",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "MTP text build; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: gmzky5s1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 202.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-oQ4e-MTP-Text on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 202.6 28.5 20.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g5edguux",
   "source_url": "https://omlx.ai/benchmarks/performance/g5edguux",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-i4c38qjr-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTPLX Bare Speed build; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 4c1fvv2p, ltkh0j7e, w0rhdy8q; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 458.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Bare-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 458.9 15.8 19.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i4c38qjr",
   "source_url": "https://omlx.ai/benchmarks/performance/i4c38qjr",
   "tg_tokens": null,
   "tps": 15.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-i4c38qjr-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTPLX Bare Speed build; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 4c1fvv2p, ltkh0j7e, w0rhdy8q; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 402.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Bare-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 402.2 13.9 20.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i4c38qjr",
   "source_url": "https://omlx.ai/benchmarks/performance/i4c38qjr",
   "tg_tokens": null,
   "tps": 13.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-i4c38qjr-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTPLX Bare Speed build; oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: 4c1fvv2p, ltkh0j7e, w0rhdy8q; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 312.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Bare-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 128k 312.5 11.5 24.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i4c38qjr",
   "source_url": "https://omlx.ai/benchmarks/performance/i4c38qjr",
   "tg_tokens": null,
   "tps": 11.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-i4c38qjr-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTPLX Bare Speed build; oMLX community benchmark trace; peak memory 29.0 GB at this context; duplicate traces of the same session: 4c1fvv2p, ltkh0j7e, w0rhdy8q; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 251.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Bare-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 195k 251.8 9.4 29.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i4c38qjr",
   "source_url": "https://omlx.ai/benchmarks/performance/i4c38qjr",
   "tg_tokens": null,
   "tps": 9.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ii7ndw0h-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.2 GB at this context; duplicate traces of the same session: b78417gk; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.1,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 105.1 14.5 16.2 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ii7ndw0h",
   "source_url": "https://omlx.ai/benchmarks/performance/ii7ndw0h",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ii7ndw0h-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: b78417gk; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 99.9,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 99.9 14.2 17.3 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ii7ndw0h",
   "source_url": "https://omlx.ai/benchmarks/performance/ii7ndw0h",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-olbm9f51-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 15.4 GB at this context; duplicate traces of the same session: u81zhfbv; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.3,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-27B-oQ3e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 3bit; 1k 103.3 27.2 15.4 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/olbm9f51",
   "source_url": "https://omlx.ai/benchmarks/performance/olbm9f51",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-olbm9f51-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: u81zhfbv; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 96.8,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-27B-oQ3e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 3bit; 4k 96.8 22.0 16.3 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/olbm9f51",
   "source_url": "https://omlx.ai/benchmarks/performance/olbm9f51",
   "tg_tokens": null,
   "tps": 22.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ldsmg64m-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 6fqmkmg3, dp5lta81, m4t7qoyk, ot7brsys, vzn4ns27; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 127.6 24.7 20.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ldsmg64m",
   "source_url": "https://omlx.ai/benchmarks/performance/ldsmg64m",
   "tg_tokens": null,
   "tps": 24.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ldsmg64m-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 6fqmkmg3, dp5lta81, m4t7qoyk, ot7brsys, vzn4ns27; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 129.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 129.8 22.4 21.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ldsmg64m",
   "source_url": "https://omlx.ai/benchmarks/performance/ldsmg64m",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ldsmg64m-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 6fqmkmg3, dp5lta81, m4t7qoyk, ot7brsys, vzn4ns27; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 127.3 24.1 22.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ldsmg64m",
   "source_url": "https://omlx.ai/benchmarks/performance/ldsmg64m",
   "tg_tokens": null,
   "tps": 24.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ldsmg64m-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 6fqmkmg3, dp5lta81, m4t7qoyk, ot7brsys, vzn4ns27; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 123.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 123.0 27.7 23.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ldsmg64m",
   "source_url": "https://omlx.ai/benchmarks/performance/ldsmg64m",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ldsmg64m-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: 6fqmkmg3, dp5lta81, m4t7qoyk, ot7brsys, vzn4ns27; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 115.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 115.5 21.3 25.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ldsmg64m",
   "source_url": "https://omlx.ai/benchmarks/performance/ldsmg64m",
   "tg_tokens": null,
   "tps": 21.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ldsmg64m-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.9 GB at this context; duplicate traces of the same session: 6fqmkmg3, dp5lta81, m4t7qoyk, ot7brsys, vzn4ns27; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 98.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 98.9 15.2 30.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ldsmg64m",
   "source_url": "https://omlx.ai/benchmarks/performance/ldsmg64m",
   "tg_tokens": null,
   "tps": 15.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bxlc2ckt-ctx8192",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace, single run at 8k context; peak memory 74.6 GB; TTFT 5279 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1552.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k tokens 1,552 PP tok/s 67.0 TG tok/s 5279 TTFT (ms) 74.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bxlc2ckt",
   "source_url": "https://omlx.ai/benchmarks/performance/bxlc2ckt",
   "tg_tokens": null,
   "tps": 67.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-spxcr372-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 32.7 GB; TTFT 7714 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 132.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k tokens 132.8 PP tok/s 17.8 TG tok/s 7714 TTFT (ms) 32.7 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/spxcr372",
   "source_url": "https://omlx.ai/benchmarks/performance/spxcr372",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-93lm7ces-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 32.7 GB; TTFT 8370 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 122.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k tokens 122.3 PP tok/s 14.5 TG tok/s 8370 TTFT (ms) 32.7 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/93lm7ces",
   "source_url": "https://omlx.ai/benchmarks/performance/93lm7ces",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-8ua1wqjw-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1021.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 1,021 90.2 183.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8ua1wqjw",
   "source_url": "https://omlx.ai/benchmarks/performance/8ua1wqjw",
   "tg_tokens": null,
   "tps": 90.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-8ua1wqjw-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1140.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 1,140 82.9 183.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8ua1wqjw",
   "source_url": "https://omlx.ai/benchmarks/performance/8ua1wqjw",
   "tg_tokens": null,
   "tps": 82.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-8ua1wqjw-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 184.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1367.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 1,367 84.8 184.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8ua1wqjw",
   "source_url": "https://omlx.ai/benchmarks/performance/8ua1wqjw",
   "tg_tokens": null,
   "tps": 84.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-8ua1wqjw-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 184.4 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1357.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k 1,357 78.1 184.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8ua1wqjw",
   "source_url": "https://omlx.ai/benchmarks/performance/8ua1wqjw",
   "tg_tokens": null,
   "tps": 78.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-8m25ioke-ctx1024",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: 0ojpocym; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 290.6,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 2bit; 1k 290.6 23.4 11.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8m25ioke",
   "source_url": "https://omlx.ai/benchmarks/performance/8m25ioke",
   "tg_tokens": null,
   "tps": 23.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-8m25ioke-ctx4096",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 14.2 GB at this context; duplicate traces of the same session: 0ojpocym; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 331.8,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 2bit; 4k 331.8 22.7 14.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8m25ioke",
   "source_url": "https://omlx.ai/benchmarks/performance/8m25ioke",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-yqwh6tc7-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 6zz9njn4, d153j9tc; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 839.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 839.5 90.2 20.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yqwh6tc7",
   "source_url": "https://omlx.ai/benchmarks/performance/yqwh6tc7",
   "tg_tokens": null,
   "tps": 90.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-yqwh6tc7-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: 6zz9njn4, d153j9tc; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2363.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,363 87.9 20.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yqwh6tc7",
   "source_url": "https://omlx.ai/benchmarks/performance/yqwh6tc7",
   "tg_tokens": null,
   "tps": 87.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-yqwh6tc7-ctx16384",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 6zz9njn4, d153j9tc; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2798.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,798 79.1 20.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yqwh6tc7",
   "source_url": "https://omlx.ai/benchmarks/performance/yqwh6tc7",
   "tg_tokens": null,
   "tps": 79.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-l15fooex-ctx1024",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 7.4 GB at this context; duplicate traces of the same session: sm3ousxh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1176.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,176 63.8 7.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l15fooex",
   "source_url": "https://omlx.ai/benchmarks/performance/l15fooex",
   "tg_tokens": null,
   "tps": 63.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-l15fooex-ctx65536",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 10.9 GB at this context; duplicate traces of the same session: sm3ousxh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1307.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 1,307 56.1 10.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l15fooex",
   "source_url": "https://omlx.ai/benchmarks/performance/l15fooex",
   "tg_tokens": null,
   "tps": 56.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-l15fooex-ctx131072",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 14.5 GB at this context; duplicate traces of the same session: sm3ousxh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1023.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 1,023 49.3 14.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l15fooex",
   "source_url": "https://omlx.ai/benchmarks/performance/l15fooex",
   "tg_tokens": null,
   "tps": 49.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-l15fooex-ctx199680",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 18.2 GB at this context; duplicate traces of the same session: sm3ousxh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 821.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 821.3 43.5 18.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l15fooex",
   "source_url": "https://omlx.ai/benchmarks/performance/l15fooex",
   "tg_tokens": null,
   "tps": 43.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-yyw9j06z-ctx16384",
   "model": "gemma-4-12b-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 8.5 GB; TTFT 142307 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 115.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-12B-it-oQ4e-mtp on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k tokens 115.1 PP tok/s 19.4 TG tok/s 142307 TTFT (ms) 8.5 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yyw9j06z",
   "source_url": "https://omlx.ai/benchmarks/performance/yyw9j06z",
   "tg_tokens": null,
   "tps": 19.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-unjg7ab0-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 15.4 GB; TTFT 51940 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 315.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k tokens 315.4 PP tok/s 28.8 TG tok/s 51940 TTFT (ms) 15.4 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/unjg7ab0",
   "source_url": "https://omlx.ai/benchmarks/performance/unjg7ab0",
   "tg_tokens": null,
   "tps": 28.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-c7kn7yyq-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 17.2 GB; TTFT 51881 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 315.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k tokens 315.8 PP tok/s 27.6 TG tok/s 51881 TTFT (ms) 17.2 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c7kn7yyq",
   "source_url": "https://omlx.ai/benchmarks/performance/c7kn7yyq",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-c08vz7n8-ctx1024",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: 0wuc0d1z, msdbhho1, z2a9bj8g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1396.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26b-a4b-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 1,396 93.2 25.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c08vz7n8",
   "source_url": "https://omlx.ai/benchmarks/performance/c08vz7n8",
   "tg_tokens": null,
   "tps": 93.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-c08vz7n8-ctx65536",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 29.1 GB at this context; duplicate traces of the same session: 0wuc0d1z, msdbhho1, z2a9bj8g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2618.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26b-a4b-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 2,618 71.0 29.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c08vz7n8",
   "source_url": "https://omlx.ai/benchmarks/performance/c08vz7n8",
   "tg_tokens": null,
   "tps": 71.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-c08vz7n8-ctx131072",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 32.4 GB at this context; duplicate traces of the same session: 0wuc0d1z, msdbhho1, z2a9bj8g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1734.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26b-a4b-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 1,734 56.0 32.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c08vz7n8",
   "source_url": "https://omlx.ai/benchmarks/performance/c08vz7n8",
   "tg_tokens": null,
   "tps": 56.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-c08vz7n8-ctx199680",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 35.9 GB at this context; duplicate traces of the same session: 0wuc0d1z, msdbhho1, z2a9bj8g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1266.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26b-a4b-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 195k 1,266 42.8 35.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c08vz7n8",
   "source_url": "https://omlx.ai/benchmarks/performance/c08vz7n8",
   "tg_tokens": null,
   "tps": 42.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-efm3366c-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 5.9 GB; TTFT 620 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1652.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k tokens 1,652 PP tok/s 30.6 TG tok/s 620 TTFT (ms) 5.9 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/efm3366c",
   "source_url": "https://omlx.ai/benchmarks/performance/efm3366c",
   "tg_tokens": null,
   "tps": 30.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-183phr3x-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 177.4 GB; TTFT 778 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1317.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k tokens 1,317 PP tok/s 72.5 TG tok/s 778 TTFT (ms) 177.4 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/183phr3x",
   "source_url": "https://omlx.ai/benchmarks/performance/183phr3x",
   "tg_tokens": null,
   "tps": 72.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-xcqdfk0s-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 177.4 GB; TTFT 2373 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 431.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k tokens 431.5 PP tok/s 49.8 TG tok/s 2373 TTFT (ms) 177.4 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xcqdfk0s",
   "source_url": "https://omlx.ai/benchmarks/performance/xcqdfk0s",
   "tg_tokens": null,
   "tps": 49.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-8g24h22h-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "DWQ MLX build v2; oMLX community benchmark trace; peak memory 10.6 GB at this context; duplicate traces of the same session: 1qw26mps, 2yiuhv2a; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 465.7,
   "provenance": "community",
   "quant": "3bit",
   "quote": "LFM2.5-24B-A2B-MLX-DWQ-3bit-v2 on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 3bit; 1k 465.7 69.0 10.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8g24h22h",
   "source_url": "https://omlx.ai/benchmarks/performance/8g24h22h",
   "tg_tokens": null,
   "tps": 69.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-8g24h22h-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "DWQ MLX build v2; oMLX community benchmark trace; peak memory 10.9 GB at this context; duplicate traces of the same session: 1qw26mps, 2yiuhv2a; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 526.8,
   "provenance": "community",
   "quant": "3bit",
   "quote": "LFM2.5-24B-A2B-MLX-DWQ-3bit-v2 on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 3bit; 4k 526.8 65.3 10.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8g24h22h",
   "source_url": "https://omlx.ai/benchmarks/performance/8g24h22h",
   "tg_tokens": null,
   "tps": 65.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-8g24h22h-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "DWQ MLX build v2; oMLX community benchmark trace; peak memory 11.1 GB at this context; duplicate traces of the same session: 1qw26mps, 2yiuhv2a; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 490.3,
   "provenance": "community",
   "quant": "3bit",
   "quote": "LFM2.5-24B-A2B-MLX-DWQ-3bit-v2 on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 3bit; 16k 490.3 56.8 11.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8g24h22h",
   "source_url": "https://omlx.ai/benchmarks/performance/8g24h22h",
   "tg_tokens": null,
   "tps": 56.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-a3oe033t-ctx1024",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 479.6,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-mlx-mxfp4 on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 479.6 61.8 6.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a3oe033t",
   "source_url": "https://omlx.ai/benchmarks/performance/a3oe033t",
   "tg_tokens": null,
   "tps": 61.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-a3oe033t-ctx4096",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 453.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-mlx-mxfp4 on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 453.0 59.8 6.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a3oe033t",
   "source_url": "https://omlx.ai/benchmarks/performance/a3oe033t",
   "tg_tokens": null,
   "tps": 59.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-a3oe033t-ctx16384",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 424.9,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-mlx-mxfp4 on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k 424.9 52.7 6.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a3oe033t",
   "source_url": "https://omlx.ai/benchmarks/performance/a3oe033t",
   "tg_tokens": null,
   "tps": 52.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-li77jck1-ctx1024",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: t7svg2xu; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1175.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiniCPM-V-4.7-35B-A3B-oQ4e on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k 1,175 51.8 20.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/li77jck1",
   "source_url": "https://omlx.ai/benchmarks/performance/li77jck1",
   "tg_tokens": null,
   "tps": 51.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-li77jck1-ctx8192",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: t7svg2xu; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1421.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiniCPM-V-4.7-35B-A3B-oQ4e on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 8k 1,421 49.0 21.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/li77jck1",
   "source_url": "https://omlx.ai/benchmarks/performance/li77jck1",
   "tg_tokens": null,
   "tps": 49.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-eks91lis-ctx1024",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: mvotvahy; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 641.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiniCPM-V-4.7-35B-A3B-oQ4e on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k 641.0 45.6 20.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eks91lis",
   "source_url": "https://omlx.ai/benchmarks/performance/eks91lis",
   "tg_tokens": null,
   "tps": 45.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-eks91lis-ctx8192",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: mvotvahy; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1360.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiniCPM-V-4.7-35B-A3B-oQ4e on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 8k 1,360 47.2 21.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eks91lis",
   "source_url": "https://omlx.ai/benchmarks/performance/eks91lis",
   "tg_tokens": null,
   "tps": 47.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-7s4itm7h-ctx1024",
   "model": "Qwen3-4B-Instruct-2507",
   "notes": "oMLX community benchmark trace; peak memory 2.8 GB at this context; duplicate traces of the same session: l0nj2zkc; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 184.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-4B-Instruct-2507-4bit on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 4bit; 1k 184.2 22.9 2.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7s4itm7h",
   "source_url": "https://omlx.ai/benchmarks/performance/7s4itm7h",
   "tg_tokens": null,
   "tps": 22.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-7s4itm7h-ctx4096",
   "model": "Qwen3-4B-Instruct-2507",
   "notes": "oMLX community benchmark trace; peak memory 3.2 GB at this context; duplicate traces of the same session: l0nj2zkc; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 167.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-4B-Instruct-2507-4bit on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 4bit; 4k 167.7 19.5 3.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7s4itm7h",
   "source_url": "https://omlx.ai/benchmarks/performance/7s4itm7h",
   "tg_tokens": null,
   "tps": 19.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-1w4b2iwc-ctx1024",
   "model": "Qwen3-4B-Instruct-2507",
   "notes": "oMLX community benchmark trace; peak memory 2.8 GB at this context; duplicate traces of the same session: gew6h1o9; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 184.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-4B-Instruct-2507-4bit on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 4bit; 1k 184.1 23.0 2.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1w4b2iwc",
   "source_url": "https://omlx.ai/benchmarks/performance/1w4b2iwc",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-1w4b2iwc-ctx4096",
   "model": "Qwen3-4B-Instruct-2507",
   "notes": "oMLX community benchmark trace; peak memory 3.2 GB at this context; duplicate traces of the same session: gew6h1o9; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 167.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-4B-Instruct-2507-4bit on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 4bit; 4k 167.3 19.4 3.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1w4b2iwc",
   "source_url": "https://omlx.ai/benchmarks/performance/1w4b2iwc",
   "tg_tokens": null,
   "tps": 19.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-t04kob1v-ctx32768",
   "model": "Qwen3.5-9b-Sushi-Coder-RL",
   "notes": "MLX build; oMLX community benchmark trace, single run at 32k context; peak memory 7.3 GB; TTFT 1249049 ms; duplicate traces of the same session: none; config=7c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 26.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9b-Sushi-Coder-RL-MLX on M1 (7c); M1 (7c) \u00b7 16 GB \u00b7 4bit; 32k tokens 26.2 PP tok/s 3.1 TG tok/s 1249049 TTFT (ms) 7.3 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t04kob1v",
   "source_url": "https://omlx.ai/benchmarks/performance/t04kob1v",
   "tg_tokens": null,
   "tps": 3.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-7hdmo33l-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: yj9zi7qq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 902.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 902.5 93.8 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7hdmo33l",
   "source_url": "https://omlx.ai/benchmarks/performance/7hdmo33l",
   "tg_tokens": null,
   "tps": 93.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-7hdmo33l-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: yj9zi7qq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1077.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,077 90.6 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7hdmo33l",
   "source_url": "https://omlx.ai/benchmarks/performance/7hdmo33l",
   "tg_tokens": null,
   "tps": 90.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-7shf42mq-ctx1024",
   "model": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: umsiruan; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1159.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k 1,159 65.7 19.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7shf42mq",
   "source_url": "https://omlx.ai/benchmarks/performance/7shf42mq",
   "tg_tokens": null,
   "tps": 65.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-7shf42mq-ctx4096",
   "model": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: umsiruan; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1367.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 1,367 67.9 20.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7shf42mq",
   "source_url": "https://omlx.ai/benchmarks/performance/7shf42mq",
   "tg_tokens": null,
   "tps": 67.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-yzpfqk8r-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: kj3e7pn4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 261.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 261.4 29.3 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yzpfqk8r",
   "source_url": "https://omlx.ai/benchmarks/performance/yzpfqk8r",
   "tg_tokens": null,
   "tps": 29.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-yzpfqk8r-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: kj3e7pn4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 249.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 249.4 26.6 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yzpfqk8r",
   "source_url": "https://omlx.ai/benchmarks/performance/yzpfqk8r",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ahtrjfkr-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 3dphm4q0, nn1x28ww; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 905.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 905.0 32.2 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ahtrjfkr",
   "source_url": "https://omlx.ai/benchmarks/performance/ahtrjfkr",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ahtrjfkr-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 3dphm4q0, nn1x28ww; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 941.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 941.1 31.8 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ahtrjfkr",
   "source_url": "https://omlx.ai/benchmarks/performance/ahtrjfkr",
   "tg_tokens": null,
   "tps": 31.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ahtrjfkr-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 3dphm4q0, nn1x28ww; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 588.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 588.5 22.9 20.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ahtrjfkr",
   "source_url": "https://omlx.ai/benchmarks/performance/ahtrjfkr",
   "tg_tokens": null,
   "tps": 22.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9qoih9t5-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 5lppqzke; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 920.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 920.8 32.6 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9qoih9t5",
   "source_url": "https://omlx.ai/benchmarks/performance/9qoih9t5",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9qoih9t5-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 5lppqzke; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1012.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,012 32.1 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9qoih9t5",
   "source_url": "https://omlx.ai/benchmarks/performance/9qoih9t5",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-g1ph2f6c-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 22.1 GB; TTFT 7961 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 128.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k tokens 128.6 PP tok/s 29.0 TG tok/s 7961 TTFT (ms) 22.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g1ph2f6c",
   "source_url": "https://omlx.ai/benchmarks/performance/g1ph2f6c",
   "tg_tokens": null,
   "tps": 29.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-tzeoka62-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 22.1 GB; TTFT 7978 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 128.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k tokens 128.3 PP tok/s 25.0 TG tok/s 7978 TTFT (ms) 22.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tzeoka62",
   "source_url": "https://omlx.ai/benchmarks/performance/tzeoka62",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-jadzevua-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 23.4 GB; TTFT 66070 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 124.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 8k tokens 124.0 PP tok/s 33.5 TG tok/s 66070 TTFT (ms) 23.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jadzevua",
   "source_url": "https://omlx.ai/benchmarks/performance/jadzevua",
   "tg_tokens": null,
   "tps": 33.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-tgmqrtp0-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: vu2ce8mu; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 244.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k 244.6 17.4 17.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tgmqrtp0",
   "source_url": "https://omlx.ai/benchmarks/performance/tgmqrtp0",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-tgmqrtp0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: vu2ce8mu; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 299.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 299.5 18.2 19.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tgmqrtp0",
   "source_url": "https://omlx.ai/benchmarks/performance/tgmqrtp0",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-fpzmq7a8-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 6pm4s5zj; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 258.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k 258.8 17.4 17.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fpzmq7a8",
   "source_url": "https://omlx.ai/benchmarks/performance/fpzmq7a8",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-fpzmq7a8-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 6pm4s5zj; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 304.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 8k 304.5 20.3 19.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fpzmq7a8",
   "source_url": "https://omlx.ai/benchmarks/performance/fpzmq7a8",
   "tg_tokens": null,
   "tps": 20.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-xs1eq7em-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: wfvghyyk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 578.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 578.7 44.7 17.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xs1eq7em",
   "source_url": "https://omlx.ai/benchmarks/performance/xs1eq7em",
   "tg_tokens": null,
   "tps": 44.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-xs1eq7em-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: wfvghyyk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 672.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 672.0 49.2 19.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xs1eq7em",
   "source_url": "https://omlx.ai/benchmarks/performance/xs1eq7em",
   "tg_tokens": null,
   "tps": 49.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-be1vhohd-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: myh5woiq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 591.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 591.2 35.6 17.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/be1vhohd",
   "source_url": "https://omlx.ai/benchmarks/performance/be1vhohd",
   "tg_tokens": null,
   "tps": 35.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-be1vhohd-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: myh5woiq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 691.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 691.5 40.6 19.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/be1vhohd",
   "source_url": "https://omlx.ai/benchmarks/performance/be1vhohd",
   "tg_tokens": null,
   "tps": 40.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-jte185ae-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 566.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 566.6 40.1 17.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jte185ae",
   "source_url": "https://omlx.ai/benchmarks/performance/jte185ae",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-jte185ae-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 630.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 630.9 46.1 19.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jte185ae",
   "source_url": "https://omlx.ai/benchmarks/performance/jte185ae",
   "tg_tokens": null,
   "tps": 46.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bw4y5kyf-ctx8192",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace, single run at 8k context; peak memory 74.6 GB; TTFT 4505 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1819.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k tokens 1,819 PP tok/s 68.2 TG tok/s 4505 TTFT (ms) 74.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bw4y5kyf",
   "source_url": "https://omlx.ai/benchmarks/performance/bw4y5kyf",
   "tg_tokens": null,
   "tps": 68.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-kncsmy72-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: l6aiq4h7, sfhqdjk1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 53.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 53.1 6.8 21.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kncsmy72",
   "source_url": "https://omlx.ai/benchmarks/performance/kncsmy72",
   "tg_tokens": null,
   "tps": 6.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-kncsmy72-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: l6aiq4h7, sfhqdjk1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 85.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 85.8 8.3 22.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kncsmy72",
   "source_url": "https://omlx.ai/benchmarks/performance/kncsmy72",
   "tg_tokens": null,
   "tps": 8.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-kncsmy72-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: l6aiq4h7, sfhqdjk1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 88.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 88.6 7.2 22.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kncsmy72",
   "source_url": "https://omlx.ai/benchmarks/performance/kncsmy72",
   "tg_tokens": null,
   "tps": 7.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-otviy94f-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 33.1 GB; TTFT 19324 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 212.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k tokens 212.0 PP tok/s 14.0 TG tok/s 19324 TTFT (ms) 33.1 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/otviy94f",
   "source_url": "https://omlx.ai/benchmarks/performance/otviy94f",
   "tg_tokens": null,
   "tps": 14.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-w2jmbvk4-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 31.5 GB; TTFT 18498 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 221.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k tokens 221.4 PP tok/s 14.4 TG tok/s 18498 TTFT (ms) 31.5 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w2jmbvk4",
   "source_url": "https://omlx.ai/benchmarks/performance/w2jmbvk4",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-hhuf3cmc-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 33.0 GB; TTFT 7532 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 136.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k tokens 136.0 PP tok/s 17.8 TG tok/s 7532 TTFT (ms) 33.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hhuf3cmc",
   "source_url": "https://omlx.ai/benchmarks/performance/hhuf3cmc",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-t8z56k5i-ctx1024",
   "model": "Ternary-Bonsai-8B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 2.9 GB at this context; duplicate traces of the same session: ex8hmps1; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 138.7,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-8B-mlx-2bit on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 2bit; 1k 138.7 24.1 2.9 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t8z56k5i",
   "source_url": "https://omlx.ai/benchmarks/performance/t8z56k5i",
   "tg_tokens": null,
   "tps": 24.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-t8z56k5i-ctx4096",
   "model": "Ternary-Bonsai-8B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 3.3 GB at this context; duplicate traces of the same session: ex8hmps1; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 129.9,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-8B-mlx-2bit on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 2bit; 4k 129.9 20.2 3.3 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t8z56k5i",
   "source_url": "https://omlx.ai/benchmarks/performance/t8z56k5i",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-006tv2mb-ctx8192",
   "model": "gemma-4-12b-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: qc45uc27, ty7k7gtp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 519.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-12B-it-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 519.7 34.6 13.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/006tv2mb",
   "source_url": "https://omlx.ai/benchmarks/performance/006tv2mb",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-006tv2mb-ctx32768",
   "model": "gemma-4-12b-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 14.5 GB at this context; duplicate traces of the same session: qc45uc27, ty7k7gtp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 439.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-12B-it-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 439.9 32.2 14.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/006tv2mb",
   "source_url": "https://omlx.ai/benchmarks/performance/006tv2mb",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-006tv2mb-ctx131072",
   "model": "gemma-4-12b-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: qc45uc27, ty7k7gtp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 311.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-12B-it-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 311.6 27.5 19.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/006tv2mb",
   "source_url": "https://omlx.ai/benchmarks/performance/006tv2mb",
   "tg_tokens": null,
   "tps": 27.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-2c3x4t4e-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 12.8 GB; TTFT 19847 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 825.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 16k tokens 825.5 PP tok/s 26.8 TG tok/s 19847 TTFT (ms) 12.8 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2c3x4t4e",
   "source_url": "https://omlx.ai/benchmarks/performance/2c3x4t4e",
   "tg_tokens": null,
   "tps": 26.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-c2th7cpv-ctx8192",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1379.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 1,379 76.6 26.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c2th7cpv",
   "source_url": "https://omlx.ai/benchmarks/performance/c2th7cpv",
   "tg_tokens": null,
   "tps": 76.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-c2th7cpv-ctx32768",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1110.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 1,110 63.2 27.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c2th7cpv",
   "source_url": "https://omlx.ai/benchmarks/performance/c2th7cpv",
   "tg_tokens": null,
   "tps": 63.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-c2th7cpv-ctx131072",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 661.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 661.5 37.4 32.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c2th7cpv",
   "source_url": "https://omlx.ai/benchmarks/performance/c2th7cpv",
   "tg_tokens": null,
   "tps": 37.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-h7z06ci7-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 14.1 GB at this context; duplicate traces of the same session: vryrftfz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1519.0,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "gemma-4-26b-a4b-it-nvfp4 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,519 113.9 14.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h7z06ci7",
   "source_url": "https://omlx.ai/benchmarks/performance/h7z06ci7",
   "tg_tokens": null,
   "tps": 113.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-h7z06ci7-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: vryrftfz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1461.0,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "gemma-4-26b-a4b-it-nvfp4 on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,461 110.6 14.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h7z06ci7",
   "source_url": "https://omlx.ai/benchmarks/performance/h7z06ci7",
   "tg_tokens": null,
   "tps": 110.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-x9pde89t-ctx1024",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: 2962z0nl, wcx846tx, xc5517y8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 640.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 640.4 15.3 32.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x9pde89t",
   "source_url": "https://omlx.ai/benchmarks/performance/x9pde89t",
   "tg_tokens": null,
   "tps": 15.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-x9pde89t-ctx65536",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 41.5 GB at this context; duplicate traces of the same session: 2962z0nl, wcx846tx, xc5517y8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 456.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 456.1 12.4 41.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x9pde89t",
   "source_url": "https://omlx.ai/benchmarks/performance/x9pde89t",
   "tg_tokens": null,
   "tps": 12.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-x9pde89t-ctx131072",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 50.5 GB at this context; duplicate traces of the same session: 2962z0nl, wcx846tx, xc5517y8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 340.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 340.8 10.3 50.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x9pde89t",
   "source_url": "https://omlx.ai/benchmarks/performance/x9pde89t",
   "tg_tokens": null,
   "tps": 10.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-x9pde89t-ctx199680",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 51.3 GB at this context; duplicate traces of the same session: 2962z0nl, wcx846tx, xc5517y8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 266.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 195k 266.3 8.2 51.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x9pde89t",
   "source_url": "https://omlx.ai/benchmarks/performance/x9pde89t",
   "tg_tokens": null,
   "tps": 8.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-a1d4pd11-ctx1024",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 3.0 GB at this context; duplicate traces of the same session: ze4fhes3; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 987.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e2b-it-4bit on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 4bit; 1k 987.6 37.9 3.0 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a1d4pd11",
   "source_url": "https://omlx.ai/benchmarks/performance/a1d4pd11",
   "tg_tokens": null,
   "tps": 37.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1",
   "id": "omlx-a1d4pd11-ctx4096",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 3.0 GB at this context; duplicate traces of the same session: ze4fhes3; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 987.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e2b-it-4bit on M1 (8c); M1 (8c) \u00b7 16 GB \u00b7 4bit; 4k 987.7 35.8 3.0 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a1d4pd11",
   "source_url": "https://omlx.ai/benchmarks/performance/a1d4pd11",
   "tg_tokens": null,
   "tps": 35.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-s2asywmo-ctx1024",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 157.0 GB at this context; duplicate traces of the same session: qonomcqe, ynta0f9c; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 440.7,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "DeepSeek-V4-Flash-0731-MXFP4-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 440.7 41.3 157.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s2asywmo",
   "source_url": "https://omlx.ai/benchmarks/performance/s2asywmo",
   "tg_tokens": null,
   "tps": 41.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-s2asywmo-ctx4096",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 157.4 GB at this context; duplicate traces of the same session: qonomcqe, ynta0f9c; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 478.7,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "DeepSeek-V4-Flash-0731-MXFP4-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 478.7 46.3 157.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s2asywmo",
   "source_url": "https://omlx.ai/benchmarks/performance/s2asywmo",
   "tg_tokens": null,
   "tps": 46.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-s2asywmo-ctx16384",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 157.5 GB at this context; duplicate traces of the same session: qonomcqe, ynta0f9c; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 462.4,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "DeepSeek-V4-Flash-0731-MXFP4-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 462.4 41.8 157.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s2asywmo",
   "source_url": "https://omlx.ai/benchmarks/performance/s2asywmo",
   "tg_tokens": null,
   "tps": 41.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vpb0yrc4-ctx1024",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 156.7 GB at this context; duplicate traces of the same session: o02hvpsy, ty1giecz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 424.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-V2.6-Flash-RL-MLX-4bit-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 424.0 53.1 156.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vpb0yrc4",
   "source_url": "https://omlx.ai/benchmarks/performance/vpb0yrc4",
   "tg_tokens": null,
   "tps": 53.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vpb0yrc4-ctx4096",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 157.8 GB at this context; duplicate traces of the same session: o02hvpsy, ty1giecz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 482.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-V2.6-Flash-RL-MLX-4bit-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 482.2 53.3 157.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vpb0yrc4",
   "source_url": "https://omlx.ai/benchmarks/performance/vpb0yrc4",
   "tg_tokens": null,
   "tps": 53.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vpb0yrc4-ctx16384",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 161.1 GB at this context; duplicate traces of the same session: o02hvpsy, ty1giecz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 447.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-V2.6-Flash-RL-MLX-4bit-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 447.2 53.6 161.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vpb0yrc4",
   "source_url": "https://omlx.ai/benchmarks/performance/vpb0yrc4",
   "tg_tokens": null,
   "tps": 53.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-i0ecubsv-ctx1024",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1179.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "MiniCPM-V-4.7-35B-A3B-mxfp4 on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k 1,179 29.6 22.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i0ecubsv",
   "source_url": "https://omlx.ai/benchmarks/performance/i0ecubsv",
   "tg_tokens": null,
   "tps": 29.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-i0ecubsv-ctx8192",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1460.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "MiniCPM-V-4.7-35B-A3B-mxfp4 on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 8k 1,460 29.1 22.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i0ecubsv",
   "source_url": "https://omlx.ai/benchmarks/performance/i0ecubsv",
   "tg_tokens": null,
   "tps": 29.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-7peongxy-ctx199680",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 195k context; peak memory 26.5 GB; TTFT 213882 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 935.1,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "MiniCPM-V-4.7-35B-A3B-mxfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 195k tokens 935.1 PP tok/s 31.2 TG tok/s 213882 TTFT (ms) 26.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7peongxy",
   "source_url": "https://omlx.ai/benchmarks/performance/7peongxy",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bcu2jd1g-ctx1024",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1883.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ4 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 1,883 132.1 19.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bcu2jd1g",
   "source_url": "https://omlx.ai/benchmarks/performance/bcu2jd1g",
   "tg_tokens": null,
   "tps": 132.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bcu2jd1g-ctx4096",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2073.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ4 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 2,073 128.4 20.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bcu2jd1g",
   "source_url": "https://omlx.ai/benchmarks/performance/bcu2jd1g",
   "tg_tokens": null,
   "tps": 128.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bcu2jd1g-ctx16384",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1939.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ4 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 1,939 120.9 20.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bcu2jd1g",
   "source_url": "https://omlx.ai/benchmarks/performance/bcu2jd1g",
   "tg_tokens": null,
   "tps": 120.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-whxm9xzi-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 12.3 GB at this context; duplicate traces of the same session: 1ief02b5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 363.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B-8bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 8bit; 1k 363.0 36.9 12.3 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/whxm9xzi",
   "source_url": "https://omlx.ai/benchmarks/performance/whxm9xzi",
   "tg_tokens": null,
   "tps": 36.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-whxm9xzi-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 12.6 GB at this context; duplicate traces of the same session: 1ief02b5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 491.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B-8bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 8bit; 4k 491.8 30.7 12.6 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/whxm9xzi",
   "source_url": "https://omlx.ai/benchmarks/performance/whxm9xzi",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-jw6rp31e-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 13.5 GB; TTFT 2812 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 364.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B-8bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 8bit; 1k tokens 364.2 PP tok/s 23.8 TG tok/s 2812 TTFT (ms) 13.5 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jw6rp31e",
   "source_url": "https://omlx.ai/benchmarks/performance/jw6rp31e",
   "tg_tokens": null,
   "tps": 23.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-5v1lz30u-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: ikjsluv1; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 720.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 1k 720.0 58.3 19.4 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5v1lz30u",
   "source_url": "https://omlx.ai/benchmarks/performance/5v1lz30u",
   "tg_tokens": null,
   "tps": 58.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Pro",
   "id": "omlx-5v1lz30u-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: ikjsluv1; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 795.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 4k 795.7 45.9 20.1 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5v1lz30u",
   "source_url": "https://omlx.ai/benchmarks/performance/5v1lz30u",
   "tg_tokens": null,
   "tps": 45.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-t7upgsik-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.2 GB at this context; duplicate traces of the same session: 32dg20l1, mnuyzisc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1788.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 1,788 23.5 39.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t7upgsik",
   "source_url": "https://omlx.ai/benchmarks/performance/t7upgsik",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-t7upgsik-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 40.2 GB at this context; duplicate traces of the same session: 32dg20l1, mnuyzisc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1296.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 1,296 103.7 40.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t7upgsik",
   "source_url": "https://omlx.ai/benchmarks/performance/t7upgsik",
   "tg_tokens": null,
   "tps": 103.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-t7upgsik-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 43.7 GB at this context; duplicate traces of the same session: 32dg20l1, mnuyzisc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 622.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 622.6 74.4 43.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t7upgsik",
   "source_url": "https://omlx.ai/benchmarks/performance/t7upgsik",
   "tg_tokens": null,
   "tps": 74.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-nlyqei5r-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: jlzk580n; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 188.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 188.9 6.5 15.8 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nlyqei5r",
   "source_url": "https://omlx.ai/benchmarks/performance/nlyqei5r",
   "tg_tokens": null,
   "tps": 6.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-nlyqei5r-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: jlzk580n; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 215.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 215.1 7.3 17.2 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nlyqei5r",
   "source_url": "https://omlx.ai/benchmarks/performance/nlyqei5r",
   "tg_tokens": null,
   "tps": 7.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7hrjl6mc-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: uj7chsl8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 905.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 905.7 31.5 17.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7hrjl6mc",
   "source_url": "https://omlx.ai/benchmarks/performance/7hrjl6mc",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7hrjl6mc-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: uj7chsl8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 974.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 974.3 29.7 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7hrjl6mc",
   "source_url": "https://omlx.ai/benchmarks/performance/7hrjl6mc",
   "tg_tokens": null,
   "tps": 29.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-v8e5m6ds-ctx1024",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "ET-O5.5-G6A-G4.7-DSV4-K3-RLOO MTP DFlash2 build; oMLX community benchmark trace; peak memory 16.8 GB at this context; duplicate traces of the same session: 3z67qris; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 439.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-ET-O5.5-G6A-G4.7-DSV4-K3-RLOO-MTP-DFlash2-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 439.6 35.4 16.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v8e5m6ds",
   "source_url": "https://omlx.ai/benchmarks/performance/v8e5m6ds",
   "tg_tokens": null,
   "tps": 35.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-v8e5m6ds-ctx4096",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "ET-O5.5-G6A-G4.7-DSV4-K3-RLOO MTP DFlash2 build; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 3z67qris; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 473.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-ET-O5.5-G6A-G4.7-DSV4-K3-RLOO-MTP-DFlash2-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 473.3 34.8 19.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v8e5m6ds",
   "source_url": "https://omlx.ai/benchmarks/performance/v8e5m6ds",
   "tg_tokens": null,
   "tps": 34.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-a8x2apqm-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: o2r1cyro; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 916.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 916.7 47.5 23.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a8x2apqm",
   "source_url": "https://omlx.ai/benchmarks/performance/a8x2apqm",
   "tg_tokens": null,
   "tps": 47.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-a8x2apqm-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: o2r1cyro; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 981.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 981.5 40.4 24.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a8x2apqm",
   "source_url": "https://omlx.ai/benchmarks/performance/a8x2apqm",
   "tg_tokens": null,
   "tps": 40.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2g2jvrm3-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: uuspc0gz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 829.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 829.0 26.9 20.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2g2jvrm3",
   "source_url": "https://omlx.ai/benchmarks/performance/2g2jvrm3",
   "tg_tokens": null,
   "tps": 26.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2g2jvrm3-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: uuspc0gz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 841.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 841.1 26.5 21.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2g2jvrm3",
   "source_url": "https://omlx.ai/benchmarks/performance/2g2jvrm3",
   "tg_tokens": null,
   "tps": 26.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-51ysy9q4-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 18.5 GB; TTFT 7435 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 137.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 1k tokens 137.7 PP tok/s 33.9 TG tok/s 7435 TTFT (ms) 18.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/51ysy9q4",
   "source_url": "https://omlx.ai/benchmarks/performance/51ysy9q4",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qvg71z8f-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 566.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 566.6 40.1 17.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qvg71z8f",
   "source_url": "https://omlx.ai/benchmarks/performance/qvg71z8f",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qvg71z8f-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 630.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 630.9 46.1 19.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qvg71z8f",
   "source_url": "https://omlx.ai/benchmarks/performance/qvg71z8f",
   "tg_tokens": null,
   "tps": 46.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fuxo1o4r-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 69.6 GB at this context; duplicate traces of the same session: 1kh8yro1, 2duie7mx, pnvg9vfd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2104.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,104 78.9 69.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fuxo1o4r",
   "source_url": "https://omlx.ai/benchmarks/performance/fuxo1o4r",
   "tg_tokens": null,
   "tps": 78.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fuxo1o4r-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 70.2 GB at this context; duplicate traces of the same session: 1kh8yro1, 2duie7mx, pnvg9vfd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2990.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,990 75.0 70.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fuxo1o4r",
   "source_url": "https://omlx.ai/benchmarks/performance/fuxo1o4r",
   "tg_tokens": null,
   "tps": 75.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fuxo1o4r-ctx8192",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: 1kh8yro1, 2duie7mx, pnvg9vfd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3001.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 3,001 74.0 71.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fuxo1o4r",
   "source_url": "https://omlx.ai/benchmarks/performance/fuxo1o4r",
   "tg_tokens": null,
   "tps": 74.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fuxo1o4r-ctx16384",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: 1kh8yro1, 2duie7mx, pnvg9vfd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2764.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,764 72.0 71.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fuxo1o4r",
   "source_url": "https://omlx.ai/benchmarks/performance/fuxo1o4r",
   "tg_tokens": null,
   "tps": 72.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tt851u2j-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 69.6 GB at this context; duplicate traces of the same session: zxamruax; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2119.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,119 80.7 69.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tt851u2j",
   "source_url": "https://omlx.ai/benchmarks/performance/tt851u2j",
   "tg_tokens": null,
   "tps": 80.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tt851u2j-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 70.2 GB at this context; duplicate traces of the same session: zxamruax; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3011.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 3,011 75.6 70.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tt851u2j",
   "source_url": "https://omlx.ai/benchmarks/performance/tt851u2j",
   "tg_tokens": null,
   "tps": 75.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mefiboy8-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 69.6 GB at this context; duplicate traces of the same session: 5f22n132; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2120.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,120 79.3 69.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mefiboy8",
   "source_url": "https://omlx.ai/benchmarks/performance/mefiboy8",
   "tg_tokens": null,
   "tps": 79.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mefiboy8-ctx8192",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: 5f22n132; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3109.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 3,109 74.8 71.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mefiboy8",
   "source_url": "https://omlx.ai/benchmarks/performance/mefiboy8",
   "tg_tokens": null,
   "tps": 74.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ws9bvlhj-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 44.3 GB at this context; duplicate traces of the same session: d1n9dn1i, zi3jp7w5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 95.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 95.6 14.6 44.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ws9bvlhj",
   "source_url": "https://omlx.ai/benchmarks/performance/ws9bvlhj",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ws9bvlhj-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 44.7 GB at this context; duplicate traces of the same session: d1n9dn1i, zi3jp7w5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 356.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 356.7 15.0 44.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ws9bvlhj",
   "source_url": "https://omlx.ai/benchmarks/performance/ws9bvlhj",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ws9bvlhj-ctx8192",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 46.0 GB at this context; duplicate traces of the same session: d1n9dn1i, zi3jp7w5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 295.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 295.0 14.3 46.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ws9bvlhj",
   "source_url": "https://omlx.ai/benchmarks/performance/ws9bvlhj",
   "tg_tokens": null,
   "tps": 14.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-y872z2cr-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.8 GB at this context; duplicate traces of the same session: 8skp60yf; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 89.8,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 2bit; 1k 89.8 11.3 30.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y872z2cr",
   "source_url": "https://omlx.ai/benchmarks/performance/y872z2cr",
   "tg_tokens": null,
   "tps": 11.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-y872z2cr-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.7 GB at this context; duplicate traces of the same session: 8skp60yf; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 191.7,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 2bit; 4k 191.7 10.4 31.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y872z2cr",
   "source_url": "https://omlx.ai/benchmarks/performance/y872z2cr",
   "tg_tokens": null,
   "tps": 10.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lmfxk3z7-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 57.4 GB at this context; duplicate traces of the same session: 45or3yiq, p8ne0cfu; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 790.8,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-Flash-Next-oQ3e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 3bit; 4k 790.8 53.6 57.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lmfxk3z7",
   "source_url": "https://omlx.ai/benchmarks/performance/lmfxk3z7",
   "tg_tokens": null,
   "tps": 53.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-lmfxk3z7-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 57.4 GB at this context; duplicate traces of the same session: 45or3yiq, p8ne0cfu; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1054.0,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-Flash-Next-oQ3e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 3bit; 8k 1,054 78.6 57.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lmfxk3z7",
   "source_url": "https://omlx.ai/benchmarks/performance/lmfxk3z7",
   "tg_tokens": null,
   "tps": 78.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lmfxk3z7-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 58.0 GB at this context; duplicate traces of the same session: 45or3yiq, p8ne0cfu; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1048.0,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-Flash-Next-oQ3e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 3bit; 16k 1,048 73.9 58.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lmfxk3z7",
   "source_url": "https://omlx.ai/benchmarks/performance/lmfxk3z7",
   "tg_tokens": null,
   "tps": 73.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9ztua1pw-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 57.4 GB at this context; duplicate traces of the same session: ca2zxwzx; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1055.0,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-Flash-Next-oQ3e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 3bit; 8k 1,055 38.6 57.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ztua1pw",
   "source_url": "https://omlx.ai/benchmarks/performance/9ztua1pw",
   "tg_tokens": null,
   "tps": 38.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9ztua1pw-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 58.0 GB at this context; duplicate traces of the same session: ca2zxwzx; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1047.0,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-Flash-Next-oQ3e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 3bit; 16k 1,047 21.9 58.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ztua1pw",
   "source_url": "https://omlx.ai/benchmarks/performance/9ztua1pw",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-k29pagua-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 195k context; peak memory 79.3 GB; TTFT 89038 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2246.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k tokens 2,246 PP tok/s 41.7 TG tok/s 89038 TTFT (ms) 79.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k29pagua",
   "source_url": "https://omlx.ai/benchmarks/performance/k29pagua",
   "tg_tokens": null,
   "tps": 41.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-me9e33km-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 35.6 GB; TTFT 33089 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 495.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k tokens 495.1 PP tok/s 7.1 TG tok/s 33089 TTFT (ms) 35.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/me9e33km",
   "source_url": "https://omlx.ai/benchmarks/performance/me9e33km",
   "tg_tokens": null,
   "tps": 7.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-ujb7v8yj-ctx16384",
   "model": "gemma-4-12b-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 9.4 GB; TTFT 142154 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 115.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-12B-it-oQ4e-mtp on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k tokens 115.3 PP tok/s 18.1 TG tok/s 142154 TTFT (ms) 9.4 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ujb7v8yj",
   "source_url": "https://omlx.ai/benchmarks/performance/ujb7v8yj",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-hsi3ub7y-ctx8192",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: qh17i5d4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1379.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 1,379 76.6 26.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hsi3ub7y",
   "source_url": "https://omlx.ai/benchmarks/performance/hsi3ub7y",
   "tg_tokens": null,
   "tps": 76.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-hsi3ub7y-ctx32768",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: qh17i5d4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1110.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 1,110 63.2 27.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hsi3ub7y",
   "source_url": "https://omlx.ai/benchmarks/performance/hsi3ub7y",
   "tg_tokens": null,
   "tps": 63.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-hsi3ub7y-ctx131072",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.4 GB at this context; duplicate traces of the same session: qh17i5d4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 661.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 661.5 37.4 32.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hsi3ub7y",
   "source_url": "https://omlx.ai/benchmarks/performance/hsi3ub7y",
   "tg_tokens": null,
   "tps": 37.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1ek13mxv-ctx1024",
   "model": "Gemma-4-31B",
   "notes": "oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: c7y08082, jnncpyie, z4om5r3y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 639.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31b-it-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 639.7 15.2 32.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ek13mxv",
   "source_url": "https://omlx.ai/benchmarks/performance/1ek13mxv",
   "tg_tokens": null,
   "tps": 15.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1ek13mxv-ctx65536",
   "model": "Gemma-4-31B",
   "notes": "oMLX community benchmark trace; peak memory 41.5 GB at this context; duplicate traces of the same session: c7y08082, jnncpyie, z4om5r3y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 429.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31b-it-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 429.4 12.4 41.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ek13mxv",
   "source_url": "https://omlx.ai/benchmarks/performance/1ek13mxv",
   "tg_tokens": null,
   "tps": 12.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1ek13mxv-ctx131072",
   "model": "Gemma-4-31B",
   "notes": "oMLX community benchmark trace; peak memory 50.5 GB at this context; duplicate traces of the same session: c7y08082, jnncpyie, z4om5r3y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31b-it-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 323.0 10.1 50.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ek13mxv",
   "source_url": "https://omlx.ai/benchmarks/performance/1ek13mxv",
   "tg_tokens": null,
   "tps": 10.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1ek13mxv-ctx199680",
   "model": "Gemma-4-31B",
   "notes": "oMLX community benchmark trace; peak memory 51.3 GB at this context; duplicate traces of the same session: c7y08082, jnncpyie, z4om5r3y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 261.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31b-it-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 195k 261.7 8.1 51.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ek13mxv",
   "source_url": "https://omlx.ai/benchmarks/performance/1ek13mxv",
   "tg_tokens": null,
   "tps": 8.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-zbepa46y-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace, single run at 1k context; peak memory 32.1 GB; TTFT 8767 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 116.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mlx-community--Qwen3.8-27B-8bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 8bit; 1k tokens 116.8 PP tok/s 17.5 TG tok/s 8767 TTFT (ms) 32.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zbepa46y",
   "source_url": "https://omlx.ai/benchmarks/performance/zbepa46y",
   "tg_tokens": null,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-u92upflz-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace, single run at 1k context; peak memory 32.1 GB; TTFT 8129 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 126.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mlx-community--Qwen3.8-27B-8bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 8bit; 1k tokens 126.0 PP tok/s 16.8 TG tok/s 8129 TTFT (ms) 32.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u92upflz",
   "source_url": "https://omlx.ai/benchmarks/performance/u92upflz",
   "tg_tokens": null,
   "tps": 16.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-u9d191bg-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.0 GB at this context; duplicate traces of the same session: 6tl6hih5, bv69clm1, e0kxcc5a; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1895.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,895 60.7 179.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u9d191bg",
   "source_url": "https://omlx.ai/benchmarks/performance/u9d191bg",
   "tg_tokens": null,
   "tps": 60.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-u9d191bg-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.3 GB at this context; duplicate traces of the same session: 6tl6hih5, bv69clm1, e0kxcc5a; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1897.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,897 68.0 179.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u9d191bg",
   "source_url": "https://omlx.ai/benchmarks/performance/u9d191bg",
   "tg_tokens": null,
   "tps": 68.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-u9d191bg-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.7 GB at this context; duplicate traces of the same session: 6tl6hih5, bv69clm1, e0kxcc5a; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1865.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,865 55.3 179.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u9d191bg",
   "source_url": "https://omlx.ai/benchmarks/performance/u9d191bg",
   "tg_tokens": null,
   "tps": 55.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-u9d191bg-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.6 GB at this context; duplicate traces of the same session: 6tl6hih5, bv69clm1, e0kxcc5a; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1827.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 1,827 59.1 180.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u9d191bg",
   "source_url": "https://omlx.ai/benchmarks/performance/u9d191bg",
   "tg_tokens": null,
   "tps": 59.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-nhyycxn2-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.0 GB at this context; duplicate traces of the same session: 07dsm6a6, 684jdc2u, le0763uq; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1636.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,636 89.4 179.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nhyycxn2",
   "source_url": "https://omlx.ai/benchmarks/performance/nhyycxn2",
   "tg_tokens": null,
   "tps": 89.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-nhyycxn2-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.3 GB at this context; duplicate traces of the same session: 07dsm6a6, 684jdc2u, le0763uq; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1911.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,911 77.9 179.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nhyycxn2",
   "source_url": "https://omlx.ai/benchmarks/performance/nhyycxn2",
   "tg_tokens": null,
   "tps": 77.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-nhyycxn2-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.7 GB at this context; duplicate traces of the same session: 07dsm6a6, 684jdc2u, le0763uq; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1882.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,882 74.7 179.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nhyycxn2",
   "source_url": "https://omlx.ai/benchmarks/performance/nhyycxn2",
   "tg_tokens": null,
   "tps": 74.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-nhyycxn2-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.6 GB at this context; duplicate traces of the same session: 07dsm6a6, 684jdc2u, le0763uq; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1847.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 1,847 130.2 180.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nhyycxn2",
   "source_url": "https://omlx.ai/benchmarks/performance/nhyycxn2",
   "tg_tokens": null,
   "tps": 130.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ldcoc45q-ctx1024",
   "model": "Gemma4-26B-MoE",
   "notes": "oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: 0npvadge, 5hd2i9gy, lm0nx18w; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1606.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Gemma4-26B-MoE-Q4 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 1,606 107.7 14.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ldcoc45q",
   "source_url": "https://omlx.ai/benchmarks/performance/ldcoc45q",
   "tg_tokens": null,
   "tps": 107.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ldcoc45q-ctx4096",
   "model": "Gemma4-26B-MoE",
   "notes": "oMLX community benchmark trace; peak memory 14.5 GB at this context; duplicate traces of the same session: 0npvadge, 5hd2i9gy, lm0nx18w; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1898.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Gemma4-26B-MoE-Q4 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 1,898 95.7 14.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ldcoc45q",
   "source_url": "https://omlx.ai/benchmarks/performance/ldcoc45q",
   "tg_tokens": null,
   "tps": 95.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ldcoc45q-ctx8192",
   "model": "Gemma4-26B-MoE",
   "notes": "oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: 0npvadge, 5hd2i9gy, lm0nx18w; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1848.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Gemma4-26B-MoE-Q4 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,848 84.3 14.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ldcoc45q",
   "source_url": "https://omlx.ai/benchmarks/performance/ldcoc45q",
   "tg_tokens": null,
   "tps": 84.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ldcoc45q-ctx16384",
   "model": "Gemma4-26B-MoE",
   "notes": "oMLX community benchmark trace; peak memory 15.1 GB at this context; duplicate traces of the same session: 0npvadge, 5hd2i9gy, lm0nx18w; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1738.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Gemma4-26B-MoE-Q4 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 1,738 73.1 15.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ldcoc45q",
   "source_url": "https://omlx.ai/benchmarks/performance/ldcoc45q",
   "tg_tokens": null,
   "tps": 73.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-5l8lmzzx-ctx1024",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1179.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "MiniCPM-V-4.7-35B-A3B-mxfp4 on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k 1,179 29.6 22.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5l8lmzzx",
   "source_url": "https://omlx.ai/benchmarks/performance/5l8lmzzx",
   "tg_tokens": null,
   "tps": 29.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-5l8lmzzx-ctx8192",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1460.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "MiniCPM-V-4.7-35B-A3B-mxfp4 on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 8k 1,460 29.1 22.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5l8lmzzx",
   "source_url": "https://omlx.ai/benchmarks/performance/5l8lmzzx",
   "tg_tokens": null,
   "tps": 29.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-277yispq-ctx1024",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: 6bvepld4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1883.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ4 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 1,883 132.1 19.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/277yispq",
   "source_url": "https://omlx.ai/benchmarks/performance/277yispq",
   "tg_tokens": null,
   "tps": 132.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-277yispq-ctx4096",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 6bvepld4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2073.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ4 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 2,073 128.4 20.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/277yispq",
   "source_url": "https://omlx.ai/benchmarks/performance/277yispq",
   "tg_tokens": null,
   "tps": 128.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-277yispq-ctx16384",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 6bvepld4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1939.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ4 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 1,939 120.9 20.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/277yispq",
   "source_url": "https://omlx.ai/benchmarks/performance/277yispq",
   "tg_tokens": null,
   "tps": 120.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-t4ksyx7a-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.6 GB at this context; duplicate traces of the same session: ecofxvp5, w4gmw2ld, yk2gpm7k; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2258.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 2,258 92.6 35.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t4ksyx7a",
   "source_url": "https://omlx.ai/benchmarks/performance/t4ksyx7a",
   "tg_tokens": null,
   "tps": 92.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-t4ksyx7a-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: ecofxvp5, w4gmw2ld, yk2gpm7k; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2670.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 2,670 91.6 37.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t4ksyx7a",
   "source_url": "https://omlx.ai/benchmarks/performance/t4ksyx7a",
   "tg_tokens": null,
   "tps": 91.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-t4ksyx7a-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: ecofxvp5, w4gmw2ld, yk2gpm7k; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2571.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 2,571 90.0 37.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t4ksyx7a",
   "source_url": "https://omlx.ai/benchmarks/performance/t4ksyx7a",
   "tg_tokens": null,
   "tps": 90.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-t4ksyx7a-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.6 GB at this context; duplicate traces of the same session: ecofxvp5, w4gmw2ld, yk2gpm7k; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2307.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 16k 2,307 86.1 37.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t4ksyx7a",
   "source_url": "https://omlx.ai/benchmarks/performance/t4ksyx7a",
   "tg_tokens": null,
   "tps": 86.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-505sjrkk-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "refusal-ablated MTP build; MLX; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: tkbv0fjb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 888.7,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-refusal-ablated-MLX-oQ8-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 888.7 62.8 37.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/505sjrkk",
   "source_url": "https://omlx.ai/benchmarks/performance/505sjrkk",
   "tg_tokens": null,
   "tps": 62.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-505sjrkk-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "refusal-ablated MTP build; MLX; oMLX community benchmark trace; peak memory 39.4 GB at this context; duplicate traces of the same session: tkbv0fjb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1023.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-refusal-ablated-MLX-oQ8-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 1,023 63.9 39.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/505sjrkk",
   "source_url": "https://omlx.ai/benchmarks/performance/505sjrkk",
   "tg_tokens": null,
   "tps": 63.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-lqb9d61t-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 4yqbpwjk; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 140.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 140.1 14.2 19.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lqb9d61t",
   "source_url": "https://omlx.ai/benchmarks/performance/lqb9d61t",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-lqb9d61t-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: 4yqbpwjk; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 135.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 135.4 13.8 19.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lqb9d61t",
   "source_url": "https://omlx.ai/benchmarks/performance/lqb9d61t",
   "tg_tokens": null,
   "tps": 13.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-971t35op-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 98.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 98.7 14.3 17.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/971t35op",
   "source_url": "https://omlx.ai/benchmarks/performance/971t35op",
   "tg_tokens": null,
   "tps": 14.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-971t35op-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 101.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 101.5 13.8 17.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/971t35op",
   "source_url": "https://omlx.ai/benchmarks/performance/971t35op",
   "tg_tokens": null,
   "tps": 13.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-fcdbvsii-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: 83o0su1r; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 130.0 37.8 18.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fcdbvsii",
   "source_url": "https://omlx.ai/benchmarks/performance/fcdbvsii",
   "tg_tokens": null,
   "tps": 37.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-fcdbvsii-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: 83o0su1r; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 130.0 21.2 19.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fcdbvsii",
   "source_url": "https://omlx.ai/benchmarks/performance/fcdbvsii",
   "tg_tokens": null,
   "tps": 21.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-6lo464g3-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 22.1 GB; TTFT 7968 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 128.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k tokens 128.5 PP tok/s 30.4 TG tok/s 7968 TTFT (ms) 22.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6lo464g3",
   "source_url": "https://omlx.ai/benchmarks/performance/6lo464g3",
   "tg_tokens": null,
   "tps": 30.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-l1thayfk-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 22.1 GB; TTFT 7967 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 128.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k tokens 128.5 PP tok/s 28.5 TG tok/s 7967 TTFT (ms) 22.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l1thayfk",
   "source_url": "https://omlx.ai/benchmarks/performance/l1thayfk",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-tk5o54z0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: dr8tszrz; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 229.6,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 4k 229.6 18.2 23.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tk5o54z0",
   "source_url": "https://omlx.ai/benchmarks/performance/tk5o54z0",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-tk5o54z0-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: dr8tszrz; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 216.4,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 16k 216.4 25.4 25.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tk5o54z0",
   "source_url": "https://omlx.ai/benchmarks/performance/tk5o54z0",
   "tg_tokens": null,
   "tps": 25.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9azr5ohs-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 9o4cf6mt, o5oqyb0e; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 324.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 324.3 79.6 21.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9azr5ohs",
   "source_url": "https://omlx.ai/benchmarks/performance/9azr5ohs",
   "tg_tokens": null,
   "tps": 79.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9azr5ohs-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: 9o4cf6mt, o5oqyb0e; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 437.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 437.9 68.6 24.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9azr5ohs",
   "source_url": "https://omlx.ai/benchmarks/performance/9azr5ohs",
   "tg_tokens": null,
   "tps": 68.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9azr5ohs-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: 9o4cf6mt, o5oqyb0e; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 425.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 425.3 75.8 25.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9azr5ohs",
   "source_url": "https://omlx.ai/benchmarks/performance/9azr5ohs",
   "tg_tokens": null,
   "tps": 75.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-7xjj2jra-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: xei3nhir; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 324.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 324.6 63.3 21.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7xjj2jra",
   "source_url": "https://omlx.ai/benchmarks/performance/7xjj2jra",
   "tg_tokens": null,
   "tps": 63.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-7xjj2jra-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: xei3nhir; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 439.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 439.6 62.3 24.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7xjj2jra",
   "source_url": "https://omlx.ai/benchmarks/performance/7xjj2jra",
   "tg_tokens": null,
   "tps": 62.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-7xjj2jra-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: xei3nhir; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 423.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 423.5 96.3 25.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7xjj2jra",
   "source_url": "https://omlx.ai/benchmarks/performance/7xjj2jra",
   "tg_tokens": null,
   "tps": 96.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5ujm1l1n-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 323.4 63.0 21.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ujm1l1n",
   "source_url": "https://omlx.ai/benchmarks/performance/5ujm1l1n",
   "tg_tokens": null,
   "tps": 63.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5ujm1l1n-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 427.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 427.9 65.0 24.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ujm1l1n",
   "source_url": "https://omlx.ai/benchmarks/performance/5ujm1l1n",
   "tg_tokens": null,
   "tps": 65.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5ujm1l1n-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 424.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 424.3 93.1 25.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ujm1l1n",
   "source_url": "https://omlx.ai/benchmarks/performance/5ujm1l1n",
   "tg_tokens": null,
   "tps": 93.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-t0yb7mes-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: pg26svga; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 123.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 123.6 20.0 18.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t0yb7mes",
   "source_url": "https://omlx.ai/benchmarks/performance/t0yb7mes",
   "tg_tokens": null,
   "tps": 20.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-t0yb7mes-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: pg26svga; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 130.3 21.5 19.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t0yb7mes",
   "source_url": "https://omlx.ai/benchmarks/performance/t0yb7mes",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3ri2gwfk-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: c41vyrtp, zy33zufd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 894.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 894.7 50.6 17.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ri2gwfk",
   "source_url": "https://omlx.ai/benchmarks/performance/3ri2gwfk",
   "tg_tokens": null,
   "tps": 50.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3ri2gwfk-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: c41vyrtp, zy33zufd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 970.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 970.1 68.1 18.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ri2gwfk",
   "source_url": "https://omlx.ai/benchmarks/performance/3ri2gwfk",
   "tg_tokens": null,
   "tps": 68.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3ri2gwfk-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: c41vyrtp, zy33zufd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 672.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 672.3 45.8 22.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ri2gwfk",
   "source_url": "https://omlx.ai/benchmarks/performance/3ri2gwfk",
   "tg_tokens": null,
   "tps": 45.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4g07c6bo-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 69.6 GB; TTFT 482 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2123.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 2,123 PP tok/s 79.1 TG tok/s 482 TTFT (ms) 69.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4g07c6bo",
   "source_url": "https://omlx.ai/benchmarks/performance/4g07c6bo",
   "tg_tokens": null,
   "tps": 79.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xrej0td5-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 69.6 GB at this context; duplicate traces of the same session: 26mc0oxk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2061.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,061 79.2 69.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xrej0td5",
   "source_url": "https://omlx.ai/benchmarks/performance/xrej0td5",
   "tg_tokens": null,
   "tps": 79.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xrej0td5-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 70.2 GB at this context; duplicate traces of the same session: 26mc0oxk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2940.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,940 75.2 70.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xrej0td5",
   "source_url": "https://omlx.ai/benchmarks/performance/xrej0td5",
   "tg_tokens": null,
   "tps": 75.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wgm9nusj-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 70.2 GB; TTFT 3555 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1152.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 1,152 PP tok/s 59.3 TG tok/s 3555 TTFT (ms) 70.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wgm9nusj",
   "source_url": "https://omlx.ai/benchmarks/performance/wgm9nusj",
   "tg_tokens": null,
   "tps": 59.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-a01n9hta-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 52.4 GB at this context; duplicate traces of the same session: 2objvqz3; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 231.2,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 2bit; 1k 231.2 14.5 52.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a01n9hta",
   "source_url": "https://omlx.ai/benchmarks/performance/a01n9hta",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-a01n9hta-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 52.9 GB at this context; duplicate traces of the same session: 2objvqz3; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 277.2,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 2bit; 4k 277.2 16.4 52.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a01n9hta",
   "source_url": "https://omlx.ai/benchmarks/performance/a01n9hta",
   "tg_tokens": null,
   "tps": 16.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-u6um3axh-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 74.9 GB at this context; duplicate traces of the same session: hkvwukm1; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 286.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 4k 286.6 23.4 74.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u6um3axh",
   "source_url": "https://omlx.ai/benchmarks/performance/u6um3axh",
   "tg_tokens": null,
   "tps": 23.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-u6um3axh-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 75.4 GB at this context; duplicate traces of the same session: hkvwukm1; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 601.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 16k 601.0 27.0 75.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u6um3axh",
   "source_url": "https://omlx.ai/benchmarks/performance/u6um3axh",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-u3rn4932-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: 6rubxig8, rawe8o7m; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 868.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 868.5 92.6 100.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u3rn4932",
   "source_url": "https://omlx.ai/benchmarks/performance/u3rn4932",
   "tg_tokens": null,
   "tps": 92.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-u3rn4932-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: 6rubxig8, rawe8o7m; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1062.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 1,062 73.1 101.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u3rn4932",
   "source_url": "https://omlx.ai/benchmarks/performance/u3rn4932",
   "tg_tokens": null,
   "tps": 73.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-u3rn4932-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: 6rubxig8, rawe8o7m; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1049.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 1,049 82.3 101.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u3rn4932",
   "source_url": "https://omlx.ai/benchmarks/performance/u3rn4932",
   "tg_tokens": null,
   "tps": 82.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-kw8kam8u-ctx1024",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 11.7 GB at this context; duplicate traces of the same session: 9sm999ii; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 100.2,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 2bit; 1k 100.2 16.8 11.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kw8kam8u",
   "source_url": "https://omlx.ai/benchmarks/performance/kw8kam8u",
   "tg_tokens": null,
   "tps": 16.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-kw8kam8u-ctx8192",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: 9sm999ii; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 112.0,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 2bit; 8k 112.0 16.5 15.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kw8kam8u",
   "source_url": "https://omlx.ai/benchmarks/performance/kw8kam8u",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-qtcjjqxc-ctx4096",
   "model": "Occamy-1.0",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: nsvepgqt; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 781.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "occamy-1.0-with-oQ5e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 5bit; 4k 781.4 70.5 26.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qtcjjqxc",
   "source_url": "https://omlx.ai/benchmarks/performance/qtcjjqxc",
   "tg_tokens": null,
   "tps": 70.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-qtcjjqxc-ctx32768",
   "model": "Occamy-1.0",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: nsvepgqt; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 548.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "occamy-1.0-with-oQ5e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 5bit; 32k 548.1 71.4 27.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qtcjjqxc",
   "source_url": "https://omlx.ai/benchmarks/performance/qtcjjqxc",
   "tg_tokens": null,
   "tps": 71.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-m50yfztu-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: 3aitclpi, gnnn1xja, grkeh8un, y6uzflvh; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1147.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,147 50.7 178.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m50yfztu",
   "source_url": "https://omlx.ai/benchmarks/performance/m50yfztu",
   "tg_tokens": null,
   "tps": 50.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-m50yfztu-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.9 GB at this context; duplicate traces of the same session: 3aitclpi, gnnn1xja, grkeh8un, y6uzflvh; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1946.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,946 113.1 179.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m50yfztu",
   "source_url": "https://omlx.ai/benchmarks/performance/m50yfztu",
   "tg_tokens": null,
   "tps": 113.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-m50yfztu-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.2 GB at this context; duplicate traces of the same session: 3aitclpi, gnnn1xja, grkeh8un, y6uzflvh; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1942.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,942 104.1 180.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m50yfztu",
   "source_url": "https://omlx.ai/benchmarks/performance/m50yfztu",
   "tg_tokens": null,
   "tps": 104.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-m50yfztu-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.6 GB at this context; duplicate traces of the same session: 3aitclpi, gnnn1xja, grkeh8un, y6uzflvh; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1911.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,911 82.4 180.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m50yfztu",
   "source_url": "https://omlx.ai/benchmarks/performance/m50yfztu",
   "tg_tokens": null,
   "tps": 82.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-m50yfztu-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.2 GB at this context; duplicate traces of the same session: 3aitclpi, gnnn1xja, grkeh8un, y6uzflvh; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1874.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 1,874 112.9 181.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m50yfztu",
   "source_url": "https://omlx.ai/benchmarks/performance/m50yfztu",
   "tg_tokens": null,
   "tps": 112.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ipq4u9ss-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 214.0 GB at this context; duplicate traces of the same session: kiewhlo5; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1670.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "GLM-5.3-Flash-oQ5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 5bit; 16k 1,670 123.4 214.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ipq4u9ss",
   "source_url": "https://omlx.ai/benchmarks/performance/ipq4u9ss",
   "tg_tokens": null,
   "tps": 123.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ipq4u9ss-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 214.3 GB at this context; duplicate traces of the same session: kiewhlo5; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1681.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "GLM-5.3-Flash-oQ5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 5bit; 32k 1,681 79.8 214.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ipq4u9ss",
   "source_url": "https://omlx.ai/benchmarks/performance/ipq4u9ss",
   "tg_tokens": null,
   "tps": 79.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-gu4l3eew-ctx65536",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 64k context; peak memory 14.8 GB; TTFT 173284 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 378.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LiquidAI--LFM2.5-24B-A2B-MLX-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 64k tokens 378.2 PP tok/s 33.9 TG tok/s 173284 TTFT (ms) 14.8 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gu4l3eew",
   "source_url": "https://omlx.ai/benchmarks/performance/gu4l3eew",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-eh9imhn6-ctx8192",
   "model": "Muse Glimmer",
   "notes": "oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 7wcop0xj, ccy9aivg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 222.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-30B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 222.6 21.0 20.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eh9imhn6",
   "source_url": "https://omlx.ai/benchmarks/performance/eh9imhn6",
   "tg_tokens": null,
   "tps": 21.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-eh9imhn6-ctx32768",
   "model": "Muse Glimmer",
   "notes": "oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: 7wcop0xj, ccy9aivg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 213.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-30B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 213.2 13.7 21.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eh9imhn6",
   "source_url": "https://omlx.ai/benchmarks/performance/eh9imhn6",
   "tg_tokens": null,
   "tps": 13.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-eh9imhn6-ctx131072",
   "model": "Muse Glimmer",
   "notes": "oMLX community benchmark trace; peak memory 25.6 GB at this context; duplicate traces of the same session: 7wcop0xj, ccy9aivg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 182.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-30B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 182.6 9.2 25.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eh9imhn6",
   "source_url": "https://omlx.ai/benchmarks/performance/eh9imhn6",
   "tg_tokens": null,
   "tps": 9.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ln0a254z-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 98.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 98.7 14.3 17.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ln0a254z",
   "source_url": "https://omlx.ai/benchmarks/performance/ln0a254z",
   "tg_tokens": null,
   "tps": 14.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ln0a254z-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 101.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 101.5 13.8 17.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ln0a254z",
   "source_url": "https://omlx.ai/benchmarks/performance/ln0a254z",
   "tg_tokens": null,
   "tps": 13.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ls14nz8n-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: gawn2v2q; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 76.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 76.0 16.4 19.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ls14nz8n",
   "source_url": "https://omlx.ai/benchmarks/performance/ls14nz8n",
   "tg_tokens": null,
   "tps": 16.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ls14nz8n-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: gawn2v2q; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 81.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 81.8 17.6 20.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ls14nz8n",
   "source_url": "https://omlx.ai/benchmarks/performance/ls14nz8n",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ivio2rhz-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.0 GB at this context; duplicate traces of the same session: v31pqw0r; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 100.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 100.2 21.1 18.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ivio2rhz",
   "source_url": "https://omlx.ai/benchmarks/performance/ivio2rhz",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ivio2rhz-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: v31pqw0r; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 99.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 99.6 19.1 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ivio2rhz",
   "source_url": "https://omlx.ai/benchmarks/performance/ivio2rhz",
   "tg_tokens": null,
   "tps": 19.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-dcz1qxw2-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: 7392d38a; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 100.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 100.7 16.5 17.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dcz1qxw2",
   "source_url": "https://omlx.ai/benchmarks/performance/dcz1qxw2",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-dcz1qxw2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: 7392d38a; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 101.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 101.9 17.9 18.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dcz1qxw2",
   "source_url": "https://omlx.ai/benchmarks/performance/dcz1qxw2",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ftvae224-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.0 GB at this context; duplicate traces of the same session: fpvfercq, hxf7sns0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 99.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 99.8 21.3 18.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ftvae224",
   "source_url": "https://omlx.ai/benchmarks/performance/ftvae224",
   "tg_tokens": null,
   "tps": 21.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ftvae224-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: fpvfercq, hxf7sns0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 97.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 97.6 20.1 19.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ftvae224",
   "source_url": "https://omlx.ai/benchmarks/performance/ftvae224",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ftvae224-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: fpvfercq, hxf7sns0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 74.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 74.8 19.8 22.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ftvae224",
   "source_url": "https://omlx.ai/benchmarks/performance/ftvae224",
   "tg_tokens": null,
   "tps": 19.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qfqq5cb9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: 6ek0ja80; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 815.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 815.4 33.9 29.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qfqq5cb9",
   "source_url": "https://omlx.ai/benchmarks/performance/qfqq5cb9",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qfqq5cb9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 30.6 GB at this context; duplicate traces of the same session: 6ek0ja80; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 938.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 938.9 37.3 30.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qfqq5cb9",
   "source_url": "https://omlx.ai/benchmarks/performance/qfqq5cb9",
   "tg_tokens": null,
   "tps": 37.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-87dl9flw-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: ifc3vqp0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 767.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 767.8 19.1 28.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/87dl9flw",
   "source_url": "https://omlx.ai/benchmarks/performance/87dl9flw",
   "tg_tokens": null,
   "tps": 19.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-87dl9flw-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: ifc3vqp0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 905.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 905.3 18.4 29.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/87dl9flw",
   "source_url": "https://omlx.ai/benchmarks/performance/87dl9flw",
   "tg_tokens": null,
   "tps": 18.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-3bq60kno-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: w2jut23f; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 161.9,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 161.9 19.0 20.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3bq60kno",
   "source_url": "https://omlx.ai/benchmarks/performance/3bq60kno",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-3bq60kno-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: w2jut23f; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 150.1,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 150.1 17.7 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3bq60kno",
   "source_url": "https://omlx.ai/benchmarks/performance/3bq60kno",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-hj2jic73-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: nie4aofw; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 108.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 108.4 27.1 18.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hj2jic73",
   "source_url": "https://omlx.ai/benchmarks/performance/hj2jic73",
   "tg_tokens": null,
   "tps": 27.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-hj2jic73-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: nie4aofw; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 101.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 101.9 24.1 22.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hj2jic73",
   "source_url": "https://omlx.ai/benchmarks/performance/hj2jic73",
   "tg_tokens": null,
   "tps": 24.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xwljh67p-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 324.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 324.6 63.3 21.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xwljh67p",
   "source_url": "https://omlx.ai/benchmarks/performance/xwljh67p",
   "tg_tokens": null,
   "tps": 63.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xwljh67p-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 439.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 439.6 62.3 24.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xwljh67p",
   "source_url": "https://omlx.ai/benchmarks/performance/xwljh67p",
   "tg_tokens": null,
   "tps": 62.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xwljh67p-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 423.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 423.5 96.3 25.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xwljh67p",
   "source_url": "https://omlx.ai/benchmarks/performance/xwljh67p",
   "tg_tokens": null,
   "tps": 96.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ab3rmcid-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 76xx4l82; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 323.4 63.0 21.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ab3rmcid",
   "source_url": "https://omlx.ai/benchmarks/performance/ab3rmcid",
   "tg_tokens": null,
   "tps": 63.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ab3rmcid-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: 76xx4l82; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 427.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 427.9 65.0 24.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ab3rmcid",
   "source_url": "https://omlx.ai/benchmarks/performance/ab3rmcid",
   "tg_tokens": null,
   "tps": 65.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ab3rmcid-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: 76xx4l82; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 424.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 424.3 93.1 25.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ab3rmcid",
   "source_url": "https://omlx.ai/benchmarks/performance/ab3rmcid",
   "tg_tokens": null,
   "tps": 93.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qcyd9pos-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: t9azr1tc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 430.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 430.0 43.5 24.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qcyd9pos",
   "source_url": "https://omlx.ai/benchmarks/performance/qcyd9pos",
   "tg_tokens": null,
   "tps": 43.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qcyd9pos-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: t9azr1tc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 425.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 425.2 59.0 25.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qcyd9pos",
   "source_url": "https://omlx.ai/benchmarks/performance/qcyd9pos",
   "tg_tokens": null,
   "tps": 59.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-xkpncmc8-ctx16384",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 38.4 GB at this context; duplicate traces of the same session: t42ofr45; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1401.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 8bit; 16k 1,401 57.5 38.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xkpncmc8",
   "source_url": "https://omlx.ai/benchmarks/performance/xkpncmc8",
   "tg_tokens": null,
   "tps": 57.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-xkpncmc8-ctx65536",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 41.1 GB at this context; duplicate traces of the same session: t42ofr45; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 816.1,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 8bit; 64k 816.1 49.0 41.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xkpncmc8",
   "source_url": "https://omlx.ai/benchmarks/performance/xkpncmc8",
   "tg_tokens": null,
   "tps": 49.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-a38jps3c-ctx4096",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace, single run at 4k context; peak memory 6.3 GB; TTFT 24575 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 166.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k tokens 166.7 PP tok/s 17.4 TG tok/s 24575 TTFT (ms) 6.3 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a38jps3c",
   "source_url": "https://omlx.ai/benchmarks/performance/a38jps3c",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-65vrg16x-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 57.4 GB at this context; duplicate traces of the same session: 73wf587n, lrgz4ye9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 895.1,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-Flash-Next-oQ3e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 3bit; 4k 895.1 81.5 57.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/65vrg16x",
   "source_url": "https://omlx.ai/benchmarks/performance/65vrg16x",
   "tg_tokens": null,
   "tps": 81.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-65vrg16x-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 57.6 GB at this context; duplicate traces of the same session: 73wf587n, lrgz4ye9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1057.0,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-Flash-Next-oQ3e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 3bit; 8k 1,057 63.9 57.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/65vrg16x",
   "source_url": "https://omlx.ai/benchmarks/performance/65vrg16x",
   "tg_tokens": null,
   "tps": 63.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-65vrg16x-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 58.1 GB at this context; duplicate traces of the same session: 73wf587n, lrgz4ye9; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1051.0,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-Flash-Next-oQ3e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 3bit; 16k 1,051 82.5 58.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/65vrg16x",
   "source_url": "https://omlx.ai/benchmarks/performance/65vrg16x",
   "tg_tokens": null,
   "tps": 82.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-cdvqw4gb-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 57.4 GB at this context; duplicate traces of the same session: 5507a48q, kimnemhw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 883.9,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-Flash-Next-oQ3e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 3bit; 4k 883.9 58.3 57.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cdvqw4gb",
   "source_url": "https://omlx.ai/benchmarks/performance/cdvqw4gb",
   "tg_tokens": null,
   "tps": 58.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-cdvqw4gb-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 57.6 GB at this context; duplicate traces of the same session: 5507a48q, kimnemhw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1052.0,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-Flash-Next-oQ3e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 3bit; 8k 1,052 76.1 57.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cdvqw4gb",
   "source_url": "https://omlx.ai/benchmarks/performance/cdvqw4gb",
   "tg_tokens": null,
   "tps": 76.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-cdvqw4gb-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 58.1 GB at this context; duplicate traces of the same session: 5507a48q, kimnemhw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1048.0,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-Flash-Next-oQ3e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 3bit; 16k 1,048 96.1 58.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cdvqw4gb",
   "source_url": "https://omlx.ai/benchmarks/performance/cdvqw4gb",
   "tg_tokens": null,
   "tps": 96.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-vnhcc67o-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: ifjl90t2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1059.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 1,059 69.8 71.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vnhcc67o",
   "source_url": "https://omlx.ai/benchmarks/performance/vnhcc67o",
   "tg_tokens": null,
   "tps": 69.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-vnhcc67o-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.2 GB at this context; duplicate traces of the same session: ifjl90t2; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1058.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 16k 1,058 90.7 72.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vnhcc67o",
   "source_url": "https://omlx.ai/benchmarks/performance/vnhcc67o",
   "tg_tokens": null,
   "tps": 90.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-toc0tn15-ctx8192",
   "model": "Gemma-4-31B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 19fw8acd, epwu9dio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 191.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31b-it-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 191.3 29.7 21.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/toc0tn15",
   "source_url": "https://omlx.ai/benchmarks/performance/toc0tn15",
   "tg_tokens": null,
   "tps": 29.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-toc0tn15-ctx32768",
   "model": "Gemma-4-31B",
   "notes": "oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: 19fw8acd, epwu9dio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 166.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31b-it-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 166.9 22.9 24.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/toc0tn15",
   "source_url": "https://omlx.ai/benchmarks/performance/toc0tn15",
   "tg_tokens": null,
   "tps": 22.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-toc0tn15-ctx131072",
   "model": "Gemma-4-31B",
   "notes": "oMLX community benchmark trace; peak memory 38.5 GB at this context; duplicate traces of the same session: 19fw8acd, epwu9dio; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 117.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31b-it-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 117.3 9.3 38.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/toc0tn15",
   "source_url": "https://omlx.ai/benchmarks/performance/toc0tn15",
   "tg_tokens": null,
   "tps": 9.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ligdv2m1-ctx4096",
   "model": "Occamy-1.0",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: gc5fy940; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 815.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "occamy-1.0-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 815.2 69.9 21.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ligdv2m1",
   "source_url": "https://omlx.ai/benchmarks/performance/ligdv2m1",
   "tg_tokens": null,
   "tps": 69.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ligdv2m1-ctx32768",
   "model": "Occamy-1.0",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: gc5fy940; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 565.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "occamy-1.0-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 565.8 52.7 22.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ligdv2m1",
   "source_url": "https://omlx.ai/benchmarks/performance/ligdv2m1",
   "tg_tokens": null,
   "tps": 52.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-nh335q61-ctx4096",
   "model": "Occamy-1.0",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: ijtz4tp3; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 793.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "occamy-1.0-with-oQ5e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 5bit; 4k 793.2 72.7 26.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nh335q61",
   "source_url": "https://omlx.ai/benchmarks/performance/nh335q61",
   "tg_tokens": null,
   "tps": 72.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-nh335q61-ctx32768",
   "model": "Occamy-1.0",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: ijtz4tp3; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 553.5,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "occamy-1.0-with-oQ5e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 5bit; 32k 553.5 48.5 27.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nh335q61",
   "source_url": "https://omlx.ai/benchmarks/performance/nh335q61",
   "tg_tokens": null,
   "tps": 48.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-6c1beu2g-ctx1024",
   "model": "Occamy-1.0",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 26.1 GB; TTFT 1438 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 711.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "occamy-1.0-with-oQ5e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 5bit; 1k tokens 711.9 PP tok/s 52.2 TG tok/s 1438 TTFT (ms) 26.1 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6c1beu2g",
   "source_url": "https://omlx.ai/benchmarks/performance/6c1beu2g",
   "tg_tokens": null,
   "tps": 52.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-c9m0ovy7-ctx4096",
   "model": "supergemma4-26b-uncensored",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 14.3 GB; TTFT 15926 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 257.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "supergemma4-26b-uncensored-mlx-4bit-v2 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k tokens 257.2 PP tok/s 27.1 TG tok/s 15926 TTFT (ms) 14.3 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c9m0ovy7",
   "source_url": "https://omlx.ai/benchmarks/performance/c9m0ovy7",
   "tg_tokens": null,
   "tps": 27.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-kv0q79cu-ctx16384",
   "model": "Muse Glimmer",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 15.6 GB; TTFT 82768 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 198.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-30B-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k tokens 198.0 PP tok/s 21.0 TG tok/s 82768 TTFT (ms) 15.6 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kv0q79cu",
   "source_url": "https://omlx.ai/benchmarks/performance/kv0q79cu",
   "tg_tokens": null,
   "tps": 21.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-vn2dphlj-ctx1024",
   "model": "Nemotron-3-Super-120B-A12B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 96.7 GB at this context; duplicate traces of the same session: 8bgh35ak; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 363.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Nemotron-3-Super-120B-A12B-MLX-6bit on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 6bit; 1k 363.0 43.6 96.7 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vn2dphlj",
   "source_url": "https://omlx.ai/benchmarks/performance/vn2dphlj",
   "tg_tokens": null,
   "tps": 43.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-vn2dphlj-ctx4096",
   "model": "Nemotron-3-Super-120B-A12B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 99.4 GB at this context; duplicate traces of the same session: 8bgh35ak; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 439.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Nemotron-3-Super-120B-A12B-MLX-6bit on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 6bit; 4k 439.6 43.5 99.4 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vn2dphlj",
   "source_url": "https://omlx.ai/benchmarks/performance/vn2dphlj",
   "tg_tokens": null,
   "tps": 43.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-jraaq47x-ctx1024",
   "model": "Nemotron-3.5-Lightning-30B-A3B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 1k context; peak memory 17.8 GB; TTFT 3926 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 260.8,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Nemotron-3.5-Lightning-30B-A3B-MLX-MXFP4 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 260.8 PP tok/s 37.0 TG tok/s 3926 TTFT (ms) 17.8 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jraaq47x",
   "source_url": "https://omlx.ai/benchmarks/performance/jraaq47x",
   "tg_tokens": null,
   "tps": 37.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-32q104d3-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.3 GB at this context; duplicate traces of the same session: 0xxazf4z; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4832.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 16k 4,832 209.4 38.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/32q104d3",
   "source_url": "https://omlx.ai/benchmarks/performance/32q104d3",
   "tg_tokens": null,
   "tps": 209.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-32q104d3-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: 0xxazf4z; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5331.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 64k 5,331 174.2 39.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/32q104d3",
   "source_url": "https://omlx.ai/benchmarks/performance/32q104d3",
   "tg_tokens": null,
   "tps": 174.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ynn5fsmg-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 9.9 GB; TTFT 2736 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 374.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-9B-MLX-8bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 8bit; 1k tokens 374.3 PP tok/s 27.6 TG tok/s 2736 TTFT (ms) 9.9 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ynn5fsmg",
   "source_url": "https://omlx.ai/benchmarks/performance/ynn5fsmg",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-hn0o1l74-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 12.4 GB; TTFT 2855 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 358.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-9B-MLX-8bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 8bit; 1k tokens 358.6 PP tok/s 53.0 TG tok/s 2855 TTFT (ms) 12.4 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hn0o1l74",
   "source_url": "https://omlx.ai/benchmarks/performance/hn0o1l74",
   "tg_tokens": null,
   "tps": 53.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-srwx9rza-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 12.4 GB; TTFT 2841 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 360.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-9B-MLX-8bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 8bit; 1k tokens 360.4 PP tok/s 48.7 TG tok/s 2841 TTFT (ms) 12.4 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/srwx9rza",
   "source_url": "https://omlx.ai/benchmarks/performance/srwx9rza",
   "tg_tokens": null,
   "tps": 48.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-6xnu9gup-ctx1024",
   "model": "Qwen3.5-0.8B",
   "notes": "MLX MTP build; oMLX community benchmark trace, single run at 1k context; peak memory 2.3 GB; TTFT 68 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 14987.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.5-0.8B-MLX-MTP-5bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k tokens 14,987 PP tok/s 347.3 TG tok/s 68 TTFT (ms) 2.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6xnu9gup",
   "source_url": "https://omlx.ai/benchmarks/performance/6xnu9gup",
   "tg_tokens": null,
   "tps": 347.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-beuc22gg-ctx1024",
   "model": "Qwen3.5-0.8B",
   "notes": "MLX MTP build; oMLX community benchmark trace, single run at 1k context; peak memory 2.3 GB; TTFT 186 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5512.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.5-0.8B-MLX-MTP-5bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k tokens 5,512 PP tok/s 84.6 TG tok/s 186 TTFT (ms) 2.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/beuc22gg",
   "source_url": "https://omlx.ai/benchmarks/performance/beuc22gg",
   "tg_tokens": null,
   "tps": 84.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6h15gm8p-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 20.6 GB; TTFT 7470 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4386.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k tokens 4,386 PP tok/s 107.2 TG tok/s 7470 TTFT (ms) 20.6 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6h15gm8p",
   "source_url": "https://omlx.ai/benchmarks/performance/6h15gm8p",
   "tg_tokens": null,
   "tps": 107.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0651lv6u-ctx32768",
   "model": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2416.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 2,416 91.5 21.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0651lv6u",
   "source_url": "https://omlx.ai/benchmarks/performance/0651lv6u",
   "tg_tokens": null,
   "tps": 91.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0651lv6u-ctx65536",
   "model": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1822.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 1,822 67.7 21.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0651lv6u",
   "source_url": "https://omlx.ai/benchmarks/performance/0651lv6u",
   "tg_tokens": null,
   "tps": 67.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0651lv6u-ctx131072",
   "model": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed",
   "notes": "oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1197.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 128k 1,197 57.5 23.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0651lv6u",
   "source_url": "https://omlx.ai/benchmarks/performance/0651lv6u",
   "tg_tokens": null,
   "tps": 57.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fxkr7bqc-ctx32768",
   "model": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 259t76xz, rju4vduy, se6200xp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2442.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 2,442 146.5 21.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fxkr7bqc",
   "source_url": "https://omlx.ai/benchmarks/performance/fxkr7bqc",
   "tg_tokens": null,
   "tps": 146.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fxkr7bqc-ctx65536",
   "model": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 259t76xz, rju4vduy, se6200xp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1825.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 1,825 85.4 21.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fxkr7bqc",
   "source_url": "https://omlx.ai/benchmarks/performance/fxkr7bqc",
   "tg_tokens": null,
   "tps": 85.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fxkr7bqc-ctx131072",
   "model": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed",
   "notes": "oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 259t76xz, rju4vduy, se6200xp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1197.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 128k 1,197 67.4 23.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fxkr7bqc",
   "source_url": "https://omlx.ai/benchmarks/performance/fxkr7bqc",
   "tg_tokens": null,
   "tps": 67.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fxkr7bqc-ctx199680",
   "model": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed",
   "notes": "oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: 259t76xz, rju4vduy, se6200xp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 856.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 195k 856.1 51.4 25.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fxkr7bqc",
   "source_url": "https://omlx.ai/benchmarks/performance/fxkr7bqc",
   "tg_tokens": null,
   "tps": 51.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-r6yacy37-ctx1024",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 12.8 GB at this context; duplicate traces of the same session: d5rva190, n73bipfi; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 519.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 519.1 82.1 12.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r6yacy37",
   "source_url": "https://omlx.ai/benchmarks/performance/r6yacy37",
   "tg_tokens": null,
   "tps": 82.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-r6yacy37-ctx4096",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: d5rva190, n73bipfi; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1118.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,118 81.6 13.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r6yacy37",
   "source_url": "https://omlx.ai/benchmarks/performance/r6yacy37",
   "tg_tokens": null,
   "tps": 81.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-r6yacy37-ctx8192",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.5 GB at this context; duplicate traces of the same session: d5rva190, n73bipfi; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3378.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 3,378 79.3 13.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r6yacy37",
   "source_url": "https://omlx.ai/benchmarks/performance/r6yacy37",
   "tg_tokens": null,
   "tps": 79.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-88hbtiuj-ctx1024",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.8 GB at this context; duplicate traces of the same session: 40qyd59y, j6t9kp7c; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2772.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,772 93.4 13.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/88hbtiuj",
   "source_url": "https://omlx.ai/benchmarks/performance/88hbtiuj",
   "tg_tokens": null,
   "tps": 93.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-88hbtiuj-ctx4096",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: 40qyd59y, j6t9kp7c; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3438.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,438 94.9 14.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/88hbtiuj",
   "source_url": "https://omlx.ai/benchmarks/performance/88hbtiuj",
   "tg_tokens": null,
   "tps": 94.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-88hbtiuj-ctx8192",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 14.5 GB at this context; duplicate traces of the same session: 40qyd59y, j6t9kp7c; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3314.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 3,314 99.7 14.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/88hbtiuj",
   "source_url": "https://omlx.ai/benchmarks/performance/88hbtiuj",
   "tg_tokens": null,
   "tps": 99.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-08l40nc8-ctx1024",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.8 GB at this context; duplicate traces of the same session: 2s5zpob7, 2xk8cd06; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1756.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 1,756 70.9 13.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/08l40nc8",
   "source_url": "https://omlx.ai/benchmarks/performance/08l40nc8",
   "tg_tokens": null,
   "tps": 70.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-08l40nc8-ctx4096",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: 2s5zpob7, 2xk8cd06; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3393.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,393 100.9 14.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/08l40nc8",
   "source_url": "https://omlx.ai/benchmarks/performance/08l40nc8",
   "tg_tokens": null,
   "tps": 100.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-08l40nc8-ctx8192",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 14.5 GB at this context; duplicate traces of the same session: 2s5zpob7, 2xk8cd06; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3314.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 3,314 71.8 14.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/08l40nc8",
   "source_url": "https://omlx.ai/benchmarks/performance/08l40nc8",
   "tg_tokens": null,
   "tps": 71.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-wxejxdx3-ctx1024",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.8 GB at this context; duplicate traces of the same session: 332phqjj, tybet8vl; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2776.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,776 97.2 13.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wxejxdx3",
   "source_url": "https://omlx.ai/benchmarks/performance/wxejxdx3",
   "tg_tokens": null,
   "tps": 97.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-wxejxdx3-ctx4096",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: 332phqjj, tybet8vl; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3459.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,459 78.5 14.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wxejxdx3",
   "source_url": "https://omlx.ai/benchmarks/performance/wxejxdx3",
   "tg_tokens": null,
   "tps": 78.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-wxejxdx3-ctx8192",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 14.5 GB at this context; duplicate traces of the same session: 332phqjj, tybet8vl; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3351.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 3,351 94.9 14.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wxejxdx3",
   "source_url": "https://omlx.ai/benchmarks/performance/wxejxdx3",
   "tg_tokens": null,
   "tps": 94.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-wbssr9eh-ctx1024",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.8 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2797.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,797 94.8 13.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wbssr9eh",
   "source_url": "https://omlx.ai/benchmarks/performance/wbssr9eh",
   "tg_tokens": null,
   "tps": 94.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-wbssr9eh-ctx4096",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3460.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,460 77.6 14.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wbssr9eh",
   "source_url": "https://omlx.ai/benchmarks/performance/wbssr9eh",
   "tg_tokens": null,
   "tps": 77.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-wbssr9eh-ctx8192",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 14.5 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3352.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 3,352 77.1 14.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wbssr9eh",
   "source_url": "https://omlx.ai/benchmarks/performance/wbssr9eh",
   "tg_tokens": null,
   "tps": 77.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-qmrhd5kf-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: kzd5qb8e; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 842.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 1k 842.9 85.1 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qmrhd5kf",
   "source_url": "https://omlx.ai/benchmarks/performance/qmrhd5kf",
   "tg_tokens": null,
   "tps": 85.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-qmrhd5kf-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: kzd5qb8e; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 942.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k 942.1 85.5 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qmrhd5kf",
   "source_url": "https://omlx.ai/benchmarks/performance/qmrhd5kf",
   "tg_tokens": null,
   "tps": 85.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-mhy5a9lw-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 20.4 GB; TTFT 131234 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 124.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k tokens 124.8 PP tok/s 12.3 TG tok/s 131234 TTFT (ms) 20.4 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mhy5a9lw",
   "source_url": "https://omlx.ai/benchmarks/performance/mhy5a9lw",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-055946v3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: eql07yop; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 93.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 93.3 17.6 19.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/055946v3",
   "source_url": "https://omlx.ai/benchmarks/performance/055946v3",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-055946v3-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: eql07yop; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 90.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 90.2 17.8 20.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/055946v3",
   "source_url": "https://omlx.ai/benchmarks/performance/055946v3",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-za7wdz7y-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: 6vl37029; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 134.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 134.2 46.5 19.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/za7wdz7y",
   "source_url": "https://omlx.ai/benchmarks/performance/za7wdz7y",
   "tg_tokens": null,
   "tps": 46.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-za7wdz7y-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 6vl37029; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 129.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 129.3 17.2 20.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/za7wdz7y",
   "source_url": "https://omlx.ai/benchmarks/performance/za7wdz7y",
   "tg_tokens": null,
   "tps": 17.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-4iquamr4-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: cv4cws6h, expg25vn, ocz4tam3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 323.1 65.2 21.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4iquamr4",
   "source_url": "https://omlx.ai/benchmarks/performance/4iquamr4",
   "tg_tokens": null,
   "tps": 65.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-4iquamr4-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: cv4cws6h, expg25vn, ocz4tam3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 430.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 430.6 62.4 24.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4iquamr4",
   "source_url": "https://omlx.ai/benchmarks/performance/4iquamr4",
   "tg_tokens": null,
   "tps": 62.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-4iquamr4-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: cv4cws6h, expg25vn, ocz4tam3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 427.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 427.1 62.3 25.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4iquamr4",
   "source_url": "https://omlx.ai/benchmarks/performance/4iquamr4",
   "tg_tokens": null,
   "tps": 62.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4iquamr4-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: cv4cws6h, expg25vn, ocz4tam3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 324.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 64k 324.3 58.7 30.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4iquamr4",
   "source_url": "https://omlx.ai/benchmarks/performance/4iquamr4",
   "tg_tokens": null,
   "tps": 58.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-4fy1wiwf-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 2lmiplp4, jdpuvws5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 103.9 26.6 18.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4fy1wiwf",
   "source_url": "https://omlx.ai/benchmarks/performance/4fy1wiwf",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-4fy1wiwf-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: 2lmiplp4, jdpuvws5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 106.3 40.8 19.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4fy1wiwf",
   "source_url": "https://omlx.ai/benchmarks/performance/4fy1wiwf",
   "tg_tokens": null,
   "tps": 40.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-4fy1wiwf-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: 2lmiplp4, jdpuvws5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 105.2 18.1 20.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4fy1wiwf",
   "source_url": "https://omlx.ai/benchmarks/performance/4fy1wiwf",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-f66mavy9-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 45.4 GB at this context; duplicate traces of the same session: n0d1ad14; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 381.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 32k 381.7 65.3 45.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f66mavy9",
   "source_url": "https://omlx.ai/benchmarks/performance/f66mavy9",
   "tg_tokens": null,
   "tps": 65.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-f66mavy9-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 47.3 GB at this context; duplicate traces of the same session: n0d1ad14; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 206.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 128k 206.9 42.3 47.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f66mavy9",
   "source_url": "https://omlx.ai/benchmarks/performance/f66mavy9",
   "tg_tokens": null,
   "tps": 42.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ucfs3llk-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 143.3 GB at this context; duplicate traces of the same session: d6wu28wv, dim0kqlv, dorf6d52, f6nbvead; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2008.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-fp16-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 1k 2,008 95.8 143.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ucfs3llk",
   "source_url": "https://omlx.ai/benchmarks/performance/ucfs3llk",
   "tg_tokens": null,
   "tps": 95.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ucfs3llk-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 149.6 GB at this context; duplicate traces of the same session: d6wu28wv, dim0kqlv, dorf6d52, f6nbvead; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3109.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-fp16-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 8k 3,109 89.8 149.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ucfs3llk",
   "source_url": "https://omlx.ai/benchmarks/performance/ucfs3llk",
   "tg_tokens": null,
   "tps": 89.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ucfs3llk-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 151.9 GB at this context; duplicate traces of the same session: d6wu28wv, dim0kqlv, dorf6d52, f6nbvead; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2921.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-fp16-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 64k 2,921 87.2 151.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ucfs3llk",
   "source_url": "https://omlx.ai/benchmarks/performance/ucfs3llk",
   "tg_tokens": null,
   "tps": 87.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ucfs3llk-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 154.2 GB at this context; duplicate traces of the same session: d6wu28wv, dim0kqlv, dorf6d52, f6nbvead; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2860.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-fp16-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 128k 2,860 88.2 154.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ucfs3llk",
   "source_url": "https://omlx.ai/benchmarks/performance/ucfs3llk",
   "tg_tokens": null,
   "tps": 88.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ucfs3llk-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 157.0 GB at this context; duplicate traces of the same session: d6wu28wv, dim0kqlv, dorf6d52, f6nbvead; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2803.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-fp16-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 195k 2,803 69.7 157.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ucfs3llk",
   "source_url": "https://omlx.ai/benchmarks/performance/ucfs3llk",
   "tg_tokens": null,
   "tps": 69.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-lm9ult3w-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 185.6 GB at this context; duplicate traces of the same session: wy2sezb4; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1867.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 16k 1,867 6.8 185.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lm9ult3w",
   "source_url": "https://omlx.ai/benchmarks/performance/lm9ult3w",
   "tg_tokens": null,
   "tps": 6.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-lm9ult3w-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 187.0 GB at this context; duplicate traces of the same session: wy2sezb4; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4136.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 64k 4,136 58.6 187.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lm9ult3w",
   "source_url": "https://omlx.ai/benchmarks/performance/lm9ult3w",
   "tg_tokens": null,
   "tps": 58.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-cowlhusn-ctx1024",
   "model": "Spark-X2.5-4B",
   "notes": "OptiQ build; oMLX community benchmark trace; peak memory 3.6 GB at this context; duplicate traces of the same session: 5oixey2e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1123.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Spark-X2.5-4B-OptiQ-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 1,123 95.9 3.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cowlhusn",
   "source_url": "https://omlx.ai/benchmarks/performance/cowlhusn",
   "tg_tokens": null,
   "tps": 95.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-cowlhusn-ctx8192",
   "model": "Spark-X2.5-4B",
   "notes": "OptiQ build; oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: 5oixey2e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1177.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Spark-X2.5-4B-OptiQ-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 8k 1,177 90.0 3.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cowlhusn",
   "source_url": "https://omlx.ai/benchmarks/performance/cowlhusn",
   "tg_tokens": null,
   "tps": 90.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-xm9ugyjx-ctx1024",
   "model": "AREX-Turbo",
   "notes": "oMLX community benchmark trace; peak memory 3.8 GB at this context; duplicate traces of the same session: ngp7dv4q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1286.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "AREX-Turbo-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 1,286 109.5 3.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xm9ugyjx",
   "source_url": "https://omlx.ai/benchmarks/performance/xm9ugyjx",
   "tg_tokens": null,
   "tps": 109.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-xm9ugyjx-ctx8192",
   "model": "AREX-Turbo",
   "notes": "oMLX community benchmark trace; peak memory 4.4 GB at this context; duplicate traces of the same session: ngp7dv4q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1416.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "AREX-Turbo-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 8k 1,416 99.8 4.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xm9ugyjx",
   "source_url": "https://omlx.ai/benchmarks/performance/xm9ugyjx",
   "tg_tokens": null,
   "tps": 99.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-fvj3vg2j-ctx1024",
   "model": "AREX-Turbo",
   "notes": "oMLX community benchmark trace; peak memory 5.3 GB at this context; duplicate traces of the same session: ek9043hv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1279.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "AREX-Turbo-8bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 8bit; 1k 1,279 69.9 5.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fvj3vg2j",
   "source_url": "https://omlx.ai/benchmarks/performance/fvj3vg2j",
   "tg_tokens": null,
   "tps": 69.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-fvj3vg2j-ctx8192",
   "model": "AREX-Turbo",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: ek9043hv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1391.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "AREX-Turbo-8bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 8bit; 8k 1,391 67.2 5.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fvj3vg2j",
   "source_url": "https://omlx.ai/benchmarks/performance/fvj3vg2j",
   "tg_tokens": null,
   "tps": 67.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-htcwbs5f-ctx1024",
   "model": "LFM2.5-2.6B",
   "notes": "oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: u48pq4m8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1962.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-2.6B-oQ8e on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 8bit; 1k 1,962 111.0 3.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/htcwbs5f",
   "source_url": "https://omlx.ai/benchmarks/performance/htcwbs5f",
   "tg_tokens": null,
   "tps": 111.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-htcwbs5f-ctx8192",
   "model": "LFM2.5-2.6B",
   "notes": "oMLX community benchmark trace; peak memory 3.6 GB at this context; duplicate traces of the same session: u48pq4m8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2183.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-2.6B-oQ8e on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 8bit; 8k 2,183 100.7 3.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/htcwbs5f",
   "source_url": "https://omlx.ai/benchmarks/performance/htcwbs5f",
   "tg_tokens": null,
   "tps": 100.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qndvl9rp-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.5 GB at this context; duplicate traces of the same session: qdsvphap, yr0qzc0y; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2336.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 2,336 123.4 13.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qndvl9rp",
   "source_url": "https://omlx.ai/benchmarks/performance/qndvl9rp",
   "tg_tokens": null,
   "tps": 123.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qndvl9rp-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 14.1 GB at this context; duplicate traces of the same session: qdsvphap, yr0qzc0y; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4011.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 4,011 99.1 14.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qndvl9rp",
   "source_url": "https://omlx.ai/benchmarks/performance/qndvl9rp",
   "tg_tokens": null,
   "tps": 99.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qndvl9rp-ctx131072",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: qdsvphap, yr0qzc0y; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2463.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 128k 2,463 53.9 16.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qndvl9rp",
   "source_url": "https://omlx.ai/benchmarks/performance/qndvl9rp",
   "tg_tokens": null,
   "tps": 53.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-mlc23fmw-ctx4096",
   "model": "MiMo-V2.6-Distill-Qwen-9B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 8.2 GB; TTFT 28774 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 142.4,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "MiMo-V2.6-Distill-Qwen-9B-OptiQ-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k tokens 142.4 PP tok/s 11.7 TG tok/s 28774 TTFT (ms) 8.2 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mlc23fmw",
   "source_url": "https://omlx.ai/benchmarks/performance/mlc23fmw",
   "tg_tokens": null,
   "tps": 11.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-7x6ohniv-ctx1024",
   "model": "Muse Glimmer",
   "notes": "oMLX community benchmark trace; peak memory 15.2 GB at this context; duplicate traces of the same session: 21bdfzrx, bbkjuntc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 223.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-30B-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 223.2 32.4 15.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7x6ohniv",
   "source_url": "https://omlx.ai/benchmarks/performance/7x6ohniv",
   "tg_tokens": null,
   "tps": 32.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-7x6ohniv-ctx4096",
   "model": "Muse Glimmer",
   "notes": "oMLX community benchmark trace; peak memory 15.5 GB at this context; duplicate traces of the same session: 21bdfzrx, bbkjuntc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 222.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-30B-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 222.8 31.7 15.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7x6ohniv",
   "source_url": "https://omlx.ai/benchmarks/performance/7x6ohniv",
   "tg_tokens": null,
   "tps": 31.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-7x6ohniv-ctx8192",
   "model": "Muse Glimmer",
   "notes": "oMLX community benchmark trace; peak memory 15.5 GB at this context; duplicate traces of the same session: 21bdfzrx, bbkjuntc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 219.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-30B-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 8k 219.1 31.3 15.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7x6ohniv",
   "source_url": "https://omlx.ai/benchmarks/performance/7x6ohniv",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-bgq7sisq-ctx1024",
   "model": "Muse Glimmer",
   "notes": "oMLX community benchmark trace; peak memory 15.2 GB at this context; duplicate traces of the same session: 2a1m47qm; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 221.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-30B-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 1k 221.9 32.4 15.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bgq7sisq",
   "source_url": "https://omlx.ai/benchmarks/performance/bgq7sisq",
   "tg_tokens": null,
   "tps": 32.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-bgq7sisq-ctx4096",
   "model": "Muse Glimmer",
   "notes": "oMLX community benchmark trace; peak memory 15.5 GB at this context; duplicate traces of the same session: 2a1m47qm; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-30B-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 220.3 31.7 15.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bgq7sisq",
   "source_url": "https://omlx.ai/benchmarks/performance/bgq7sisq",
   "tg_tokens": null,
   "tps": 31.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-a8zo8ytb-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: ecz0nalf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 791.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 791.6 62.5 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a8zo8ytb",
   "source_url": "https://omlx.ai/benchmarks/performance/a8zo8ytb",
   "tg_tokens": null,
   "tps": 62.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-a8zo8ytb-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: ecz0nalf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1022.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,022 76.5 23.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a8zo8ytb",
   "source_url": "https://omlx.ai/benchmarks/performance/a8zo8ytb",
   "tg_tokens": null,
   "tps": 76.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-s52tnlyw-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: pk2mc7kz; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 7818.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 16k 7,818 243.6 37.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s52tnlyw",
   "source_url": "https://omlx.ai/benchmarks/performance/s52tnlyw",
   "tg_tokens": null,
   "tps": 243.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-s52tnlyw-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.3 GB at this context; duplicate traces of the same session: pk2mc7kz; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5226.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 64k 5,226 144.2 38.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s52tnlyw",
   "source_url": "https://omlx.ai/benchmarks/performance/s52tnlyw",
   "tg_tokens": null,
   "tps": 144.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1nci2pix-ctx32768",
   "model": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: pj5lmb7i; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2416.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 2,416 91.5 21.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1nci2pix",
   "source_url": "https://omlx.ai/benchmarks/performance/1nci2pix",
   "tg_tokens": null,
   "tps": 91.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1nci2pix-ctx65536",
   "model": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: pj5lmb7i; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1822.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 1,822 67.7 21.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1nci2pix",
   "source_url": "https://omlx.ai/benchmarks/performance/1nci2pix",
   "tg_tokens": null,
   "tps": 67.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1nci2pix-ctx131072",
   "model": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed",
   "notes": "oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: pj5lmb7i; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1197.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 128k 1,197 57.5 23.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1nci2pix",
   "source_url": "https://omlx.ai/benchmarks/performance/1nci2pix",
   "tg_tokens": null,
   "tps": 57.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-w4vzo2ee-ctx1024",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: m0dnfmvw; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3678.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 3,678 108.8 12.7 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w4vzo2ee",
   "source_url": "https://omlx.ai/benchmarks/performance/w4vzo2ee",
   "tg_tokens": null,
   "tps": 108.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-w4vzo2ee-ctx4096",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.3 GB at this context; duplicate traces of the same session: m0dnfmvw; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4485.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 4,485 96.3 13.3 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w4vzo2ee",
   "source_url": "https://omlx.ai/benchmarks/performance/w4vzo2ee",
   "tg_tokens": null,
   "tps": 96.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-nvlrexoh-ctx1024",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.8 GB at this context; duplicate traces of the same session: z4d1w8td; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2797.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,797 94.8 13.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nvlrexoh",
   "source_url": "https://omlx.ai/benchmarks/performance/nvlrexoh",
   "tg_tokens": null,
   "tps": 94.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-nvlrexoh-ctx4096",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: z4d1w8td; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3460.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,460 77.6 14.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nvlrexoh",
   "source_url": "https://omlx.ai/benchmarks/performance/nvlrexoh",
   "tg_tokens": null,
   "tps": 77.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-nvlrexoh-ctx8192",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 14.5 GB at this context; duplicate traces of the same session: z4d1w8td; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3352.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 3,352 77.1 14.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nvlrexoh",
   "source_url": "https://omlx.ai/benchmarks/performance/nvlrexoh",
   "tg_tokens": null,
   "tps": 77.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-g2patxfl-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: q3n7yr2b; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 950.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 950.6 98.1 21.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g2patxfl",
   "source_url": "https://omlx.ai/benchmarks/performance/g2patxfl",
   "tg_tokens": null,
   "tps": 98.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-g2patxfl-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: q3n7yr2b; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1022.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,022 91.3 23.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g2patxfl",
   "source_url": "https://omlx.ai/benchmarks/performance/g2patxfl",
   "tg_tokens": null,
   "tps": 91.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-8weqex3o-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: svnmx22u; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 163.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 1k 163.7 28.4 16.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8weqex3o",
   "source_url": "https://omlx.ai/benchmarks/performance/8weqex3o",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-8weqex3o-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: svnmx22u; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 152.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 4k 152.2 27.7 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8weqex3o",
   "source_url": "https://omlx.ai/benchmarks/performance/8weqex3o",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ibp9la1h-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: njb7afpu, tkc64e6f; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 495.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 495.9 26.7 19.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ibp9la1h",
   "source_url": "https://omlx.ai/benchmarks/performance/ibp9la1h",
   "tg_tokens": null,
   "tps": 26.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ibp9la1h-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: njb7afpu, tkc64e6f; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 506.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 506.0 29.7 21.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ibp9la1h",
   "source_url": "https://omlx.ai/benchmarks/performance/ibp9la1h",
   "tg_tokens": null,
   "tps": 29.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ibp9la1h-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: njb7afpu, tkc64e6f; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 431.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 431.3 31.4 22.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ibp9la1h",
   "source_url": "https://omlx.ai/benchmarks/performance/ibp9la1h",
   "tg_tokens": null,
   "tps": 31.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0tic9xug-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 3a1qc816, 3mwlmclz, 5hu021jw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 818.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 818.6 18.2 28.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0tic9xug",
   "source_url": "https://omlx.ai/benchmarks/performance/0tic9xug",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0tic9xug-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: 3a1qc816, 3mwlmclz, 5hu021jw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 784.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 784.8 18.1 29.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0tic9xug",
   "source_url": "https://omlx.ai/benchmarks/performance/0tic9xug",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0tic9xug-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.8 GB at this context; duplicate traces of the same session: 3a1qc816, 3mwlmclz, 5hu021jw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 718.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 718.5 17.9 29.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0tic9xug",
   "source_url": "https://omlx.ai/benchmarks/performance/0tic9xug",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0tic9xug-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: 3a1qc816, 3mwlmclz, 5hu021jw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 643.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 643.3 17.3 30.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0tic9xug",
   "source_url": "https://omlx.ai/benchmarks/performance/0tic9xug",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0tic9xug-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.3 GB at this context; duplicate traces of the same session: 3a1qc816, 3mwlmclz, 5hu021jw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 611.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 611.6 16.3 31.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0tic9xug",
   "source_url": "https://omlx.ai/benchmarks/performance/0tic9xug",
   "tg_tokens": null,
   "tps": 16.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0tic9xug-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.3 GB at this context; duplicate traces of the same session: 3a1qc816, 3mwlmclz, 5hu021jw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 469.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 469.0 9.7 33.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0tic9xug",
   "source_url": "https://omlx.ai/benchmarks/performance/0tic9xug",
   "tg_tokens": null,
   "tps": 9.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0tic9xug-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: 3a1qc816, 3mwlmclz, 5hu021jw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 345.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 345.9 9.4 37.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0tic9xug",
   "source_url": "https://omlx.ai/benchmarks/performance/0tic9xug",
   "tg_tokens": null,
   "tps": 9.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0tic9xug-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 41.5 GB at this context; duplicate traces of the same session: 3a1qc816, 3mwlmclz, 5hu021jw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 300.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 195k 300.4 8.4 41.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0tic9xug",
   "source_url": "https://omlx.ai/benchmarks/performance/0tic9xug",
   "tg_tokens": null,
   "tps": 8.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-wijn7fiy-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTPLX Bare Speed build; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 4wk58l0t, touibglo, xyjk0hod; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 424.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Bare-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 424.8 67.7 20.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wijn7fiy",
   "source_url": "https://omlx.ai/benchmarks/performance/wijn7fiy",
   "tg_tokens": null,
   "tps": 67.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-wijn7fiy-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTPLX Bare Speed build; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: 4wk58l0t, touibglo, xyjk0hod; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 370.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Bare-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 370.9 28.1 22.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wijn7fiy",
   "source_url": "https://omlx.ai/benchmarks/performance/wijn7fiy",
   "tg_tokens": null,
   "tps": 28.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-wijn7fiy-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTPLX Bare Speed build; oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: 4wk58l0t, touibglo, xyjk0hod; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 288.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Bare-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 128k 288.9 24.9 27.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wijn7fiy",
   "source_url": "https://omlx.ai/benchmarks/performance/wijn7fiy",
   "tg_tokens": null,
   "tps": 24.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-wijn7fiy-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTPLX Bare Speed build; oMLX community benchmark trace; peak memory 34.0 GB at this context; duplicate traces of the same session: 4wk58l0t, touibglo, xyjk0hod; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 231.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Bare-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 195k 231.7 16.7 34.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wijn7fiy",
   "source_url": "https://omlx.ai/benchmarks/performance/wijn7fiy",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-80pmzatm-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: wqwwqe6d; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 131.3 14.8 16.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/80pmzatm",
   "source_url": "https://omlx.ai/benchmarks/performance/80pmzatm",
   "tg_tokens": null,
   "tps": 14.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-80pmzatm-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: wqwwqe6d; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 132.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 132.1 14.4 19.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/80pmzatm",
   "source_url": "https://omlx.ai/benchmarks/performance/80pmzatm",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-k7ntwvec-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 9jiwbuln; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 431.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 431.9 20.2 17.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k7ntwvec",
   "source_url": "https://omlx.ai/benchmarks/performance/k7ntwvec",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-k7ntwvec-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: 9jiwbuln; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 975.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 975.2 62.3 18.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k7ntwvec",
   "source_url": "https://omlx.ai/benchmarks/performance/k7ntwvec",
   "tg_tokens": null,
   "tps": 62.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gzhnd572-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: un5avoji; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 773.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 773.1 49.5 32.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gzhnd572",
   "source_url": "https://omlx.ai/benchmarks/performance/gzhnd572",
   "tg_tokens": null,
   "tps": 49.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-gzhnd572-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.2 GB at this context; duplicate traces of the same session: un5avoji; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 912.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 912.7 68.8 33.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gzhnd572",
   "source_url": "https://omlx.ai/benchmarks/performance/gzhnd572",
   "tg_tokens": null,
   "tps": 68.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vu5tn3dc-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: 7msfhs2f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 385.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 385.0 34.8 32.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vu5tn3dc",
   "source_url": "https://omlx.ai/benchmarks/performance/vu5tn3dc",
   "tg_tokens": null,
   "tps": 34.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-vu5tn3dc-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.2 GB at this context; duplicate traces of the same session: 7msfhs2f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 712.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 712.7 65.3 33.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vu5tn3dc",
   "source_url": "https://omlx.ai/benchmarks/performance/vu5tn3dc",
   "tg_tokens": null,
   "tps": 65.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rc608kfe-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: 8rvarur4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 774.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 774.9 44.1 32.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rc608kfe",
   "source_url": "https://omlx.ai/benchmarks/performance/rc608kfe",
   "tg_tokens": null,
   "tps": 44.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-rc608kfe-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.2 GB at this context; duplicate traces of the same session: 8rvarur4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 917.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 917.8 68.7 33.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rc608kfe",
   "source_url": "https://omlx.ai/benchmarks/performance/rc608kfe",
   "tg_tokens": null,
   "tps": 68.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-x65jcv97-ctx4096",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace, single run at 4k context; peak memory 6.3 GB; TTFT 24596 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 166.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k tokens 166.5 PP tok/s 17.5 TG tok/s 24596 TTFT (ms) 6.3 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x65jcv97",
   "source_url": "https://omlx.ai/benchmarks/performance/x65jcv97",
   "tg_tokens": null,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-clgm84w4-ctx65536",
   "model": "Qwen3.8-9B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 8.2 GB; TTFT 143920 ms; duplicate traces of the same session: none; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 455.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-mlx-4Bit on M5 (8c); M5 (8c) \u00b7 16 GB \u00b7 4bit; 64k tokens 455.4 PP tok/s 17.2 TG tok/s 143920 TTFT (ms) 8.2 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/clgm84w4",
   "source_url": "https://omlx.ai/benchmarks/performance/clgm84w4",
   "tg_tokens": null,
   "tps": 17.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-ggqdd7m2-ctx16384",
   "model": "Qwen3.8-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: 7a149y7m; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 647.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-mlx-4Bit on M5 (8c); M5 (8c) \u00b7 16 GB \u00b7 4bit; 16k 647.0 24.2 6.7 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ggqdd7m2",
   "source_url": "https://omlx.ai/benchmarks/performance/ggqdd7m2",
   "tg_tokens": null,
   "tps": 24.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-ggqdd7m2-ctx32768",
   "model": "Qwen3.8-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.1 GB at this context; duplicate traces of the same session: 7a149y7m; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 582.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-mlx-4Bit on M5 (8c); M5 (8c) \u00b7 16 GB \u00b7 4bit; 32k 582.6 21.7 7.1 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ggqdd7m2",
   "source_url": "https://omlx.ai/benchmarks/performance/ggqdd7m2",
   "tg_tokens": null,
   "tps": 21.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0wsv8188-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.8 GB; TTFT 968 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1058.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,058 PP tok/s 83.4 TG tok/s 968 TTFT (ms) 70.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0wsv8188",
   "source_url": "https://omlx.ai/benchmarks/performance/0wsv8188",
   "tg_tokens": null,
   "tps": 83.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-50244ck6-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 141.7 GB at this context; duplicate traces of the same session: swoq7a62; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3403.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 16k 3,403 80.8 141.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/50244ck6",
   "source_url": "https://omlx.ai/benchmarks/performance/50244ck6",
   "tg_tokens": null,
   "tps": 80.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-50244ck6-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.0 GB at this context; duplicate traces of the same session: swoq7a62; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4220.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 64k 4,220 77.1 143.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/50244ck6",
   "source_url": "https://omlx.ai/benchmarks/performance/50244ck6",
   "tg_tokens": null,
   "tps": 77.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-bgkxczk6-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace, single run at 64k context; peak memory 143.0 GB; TTFT 15530 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4220.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 64k tokens 4,220 PP tok/s 76.3 TG tok/s 15530 TTFT (ms) 143.0 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bgkxczk6",
   "source_url": "https://omlx.ai/benchmarks/performance/bgkxczk6",
   "tg_tokens": null,
   "tps": 76.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-67ubf3dm-ctx1024",
   "model": "Spark-X2.5-4B",
   "notes": "OptiQ build; oMLX community benchmark trace; peak memory 3.6 GB at this context; duplicate traces of the same session: 4b6ps1rq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1166.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Spark-X2.5-4B-OptiQ-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 1,166 90.8 3.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/67ubf3dm",
   "source_url": "https://omlx.ai/benchmarks/performance/67ubf3dm",
   "tg_tokens": null,
   "tps": 90.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-67ubf3dm-ctx8192",
   "model": "Spark-X2.5-4B",
   "notes": "OptiQ build; oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: 4b6ps1rq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1174.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Spark-X2.5-4B-OptiQ-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 8k 1,174 84.9 3.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/67ubf3dm",
   "source_url": "https://omlx.ai/benchmarks/performance/67ubf3dm",
   "tg_tokens": null,
   "tps": 84.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-cg0x8b4t-ctx1024",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: vinhb2w7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 200.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 5bit; 1k 200.0 19.4 22.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cg0x8b4t",
   "source_url": "https://omlx.ai/benchmarks/performance/cg0x8b4t",
   "tg_tokens": null,
   "tps": 19.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-cg0x8b4t-ctx8192",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: vinhb2w7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 203.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 5bit; 8k 203.0 19.0 24.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cg0x8b4t",
   "source_url": "https://omlx.ai/benchmarks/performance/cg0x8b4t",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-81wmk6r2-ctx1024",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 177.6 GB at this context; duplicate traces of the same session: g889wnuw, w6urc3ug; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 529.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 529.4 40.5 177.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/81wmk6r2",
   "source_url": "https://omlx.ai/benchmarks/performance/81wmk6r2",
   "tg_tokens": null,
   "tps": 40.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-81wmk6r2-ctx4096",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 178.8 GB at this context; duplicate traces of the same session: g889wnuw, w6urc3ug; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 605.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 605.6 38.7 178.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/81wmk6r2",
   "source_url": "https://omlx.ai/benchmarks/performance/81wmk6r2",
   "tg_tokens": null,
   "tps": 38.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-81wmk6r2-ctx8192",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 179.1 GB at this context; duplicate traces of the same session: g889wnuw, w6urc3ug; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 558.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 558.4 38.8 179.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/81wmk6r2",
   "source_url": "https://omlx.ai/benchmarks/performance/81wmk6r2",
   "tg_tokens": null,
   "tps": 38.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gpkd4dkd-ctx1024",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 177.6 GB at this context; duplicate traces of the same session: nyxzf3ye, rtdq33cv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 517.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 517.5 48.0 177.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gpkd4dkd",
   "source_url": "https://omlx.ai/benchmarks/performance/gpkd4dkd",
   "tg_tokens": null,
   "tps": 48.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gpkd4dkd-ctx4096",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 178.8 GB at this context; duplicate traces of the same session: nyxzf3ye, rtdq33cv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 590.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 590.2 47.5 178.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gpkd4dkd",
   "source_url": "https://omlx.ai/benchmarks/performance/gpkd4dkd",
   "tg_tokens": null,
   "tps": 47.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gpkd4dkd-ctx8192",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 179.1 GB at this context; duplicate traces of the same session: nyxzf3ye, rtdq33cv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 547.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 547.9 62.2 179.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gpkd4dkd",
   "source_url": "https://omlx.ai/benchmarks/performance/gpkd4dkd",
   "tg_tokens": null,
   "tps": 62.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-vc64tkaj-ctx4096",
   "model": "LFM2.5-24B-A2B-abliterated",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.6 GB at this context; duplicate traces of the same session: t8eqwzs9, wjydn5z1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2202.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-abliterated-4bit-mlx on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,202 123.4 13.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vc64tkaj",
   "source_url": "https://omlx.ai/benchmarks/performance/vc64tkaj",
   "tg_tokens": null,
   "tps": 123.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-vc64tkaj-ctx16384",
   "model": "LFM2.5-24B-A2B-abliterated",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.8 GB at this context; duplicate traces of the same session: t8eqwzs9, wjydn5z1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2056.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-abliterated-4bit-mlx on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,056 112.2 13.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vc64tkaj",
   "source_url": "https://omlx.ai/benchmarks/performance/vc64tkaj",
   "tg_tokens": null,
   "tps": 112.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-vc64tkaj-ctx131072",
   "model": "LFM2.5-24B-A2B-abliterated",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: t8eqwzs9, wjydn5z1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1007.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-abliterated-4bit-mlx on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 1,007 62.2 16.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vc64tkaj",
   "source_url": "https://omlx.ai/benchmarks/performance/vc64tkaj",
   "tg_tokens": null,
   "tps": 62.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-zxs22p6c-ctx1024",
   "model": "Ornith-1.5-35B-A3B-uncensored",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: f09nwjq6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1461.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-uncensored-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 1,461 87.6 19.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zxs22p6c",
   "source_url": "https://omlx.ai/benchmarks/performance/zxs22p6c",
   "tg_tokens": null,
   "tps": 87.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-zxs22p6c-ctx8192",
   "model": "Ornith-1.5-35B-A3B-uncensored",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: f09nwjq6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1577.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-uncensored-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 8k 1,577 75.7 20.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zxs22p6c",
   "source_url": "https://omlx.ai/benchmarks/performance/zxs22p6c",
   "tg_tokens": null,
   "tps": 75.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-r4eotn22-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: maz93qxf; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 718.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 1k 718.8 92.7 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r4eotn22",
   "source_url": "https://omlx.ai/benchmarks/performance/r4eotn22",
   "tg_tokens": null,
   "tps": 92.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-r4eotn22-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: maz93qxf; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 774.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 774.1 89.5 21.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r4eotn22",
   "source_url": "https://omlx.ai/benchmarks/performance/r4eotn22",
   "tg_tokens": null,
   "tps": 89.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-wlpafdtj-ctx16384",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 237ejsa7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2897.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,897 115.7 20.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wlpafdtj",
   "source_url": "https://omlx.ai/benchmarks/performance/wlpafdtj",
   "tg_tokens": null,
   "tps": 115.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-wlpafdtj-ctx65536",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 237ejsa7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1850.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 1,850 74.3 21.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wlpafdtj",
   "source_url": "https://omlx.ai/benchmarks/performance/wlpafdtj",
   "tg_tokens": null,
   "tps": 74.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-yiznw4pg-ctx1024",
   "model": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: xr49uvrv; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1096.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k 1,096 65.0 19.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yiznw4pg",
   "source_url": "https://omlx.ai/benchmarks/performance/yiznw4pg",
   "tg_tokens": null,
   "tps": 65.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-yiznw4pg-ctx4096",
   "model": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: xr49uvrv; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1247.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MTPLX-Optimized-Speed on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 1,247 65.0 20.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yiznw4pg",
   "source_url": "https://omlx.ai/benchmarks/performance/yiznw4pg",
   "tg_tokens": null,
   "tps": 65.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-5etseui3-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "mixed 3/4bit 3.85bpw MLX build; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 8tchfrd7, tz7sqh6q; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2493.0,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.6-35B-A3B-mixed-3-4bit-3.85bpw-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 3bit; 1k 2,493 95.2 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5etseui3",
   "source_url": "https://omlx.ai/benchmarks/performance/5etseui3",
   "tg_tokens": null,
   "tps": 95.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-5etseui3-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "mixed 3/4bit 3.85bpw MLX build; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 8tchfrd7, tz7sqh6q; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2838.0,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.6-35B-A3B-mixed-3-4bit-3.85bpw-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 3bit; 4k 2,838 115.9 18.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5etseui3",
   "source_url": "https://omlx.ai/benchmarks/performance/5etseui3",
   "tg_tokens": null,
   "tps": 115.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-5etseui3-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "mixed 3/4bit 3.85bpw MLX build; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 8tchfrd7, tz7sqh6q; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3044.0,
   "provenance": "community",
   "quant": "3bit",
   "quote": "Qwen3.6-35B-A3B-mixed-3-4bit-3.85bpw-mlx on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 3bit; 8k 3,044 114.0 18.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5etseui3",
   "source_url": "https://omlx.ai/benchmarks/performance/5etseui3",
   "tg_tokens": null,
   "tps": 114.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-rbh0om4z-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 4jttkbq8; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 114.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 1k 114.3 20.2 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rbh0om4z",
   "source_url": "https://omlx.ai/benchmarks/performance/rbh0om4z",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-rbh0om4z-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: 4jttkbq8; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 105.9 19.6 19.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rbh0om4z",
   "source_url": "https://omlx.ai/benchmarks/performance/rbh0om4z",
   "tg_tokens": null,
   "tps": 19.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-y9vp01kj-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 2apbph3l; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 474.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 474.2 31.3 18.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y9vp01kj",
   "source_url": "https://omlx.ai/benchmarks/performance/y9vp01kj",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-y9vp01kj-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 2apbph3l; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 463.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 463.5 24.4 19.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y9vp01kj",
   "source_url": "https://omlx.ai/benchmarks/performance/y9vp01kj",
   "tg_tokens": null,
   "tps": 24.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pysuihd7-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: dq1frnm1, j3b5kl1m, r8lftkyn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 818.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 818.6 18.2 28.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pysuihd7",
   "source_url": "https://omlx.ai/benchmarks/performance/pysuihd7",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pysuihd7-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: dq1frnm1, j3b5kl1m, r8lftkyn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 784.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 784.8 18.1 29.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pysuihd7",
   "source_url": "https://omlx.ai/benchmarks/performance/pysuihd7",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pysuihd7-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.8 GB at this context; duplicate traces of the same session: dq1frnm1, j3b5kl1m, r8lftkyn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 718.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 718.5 17.9 29.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pysuihd7",
   "source_url": "https://omlx.ai/benchmarks/performance/pysuihd7",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pysuihd7-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: dq1frnm1, j3b5kl1m, r8lftkyn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 643.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 643.3 17.3 30.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pysuihd7",
   "source_url": "https://omlx.ai/benchmarks/performance/pysuihd7",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pysuihd7-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.3 GB at this context; duplicate traces of the same session: dq1frnm1, j3b5kl1m, r8lftkyn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 611.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 611.6 16.3 31.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pysuihd7",
   "source_url": "https://omlx.ai/benchmarks/performance/pysuihd7",
   "tg_tokens": null,
   "tps": 16.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pysuihd7-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.3 GB at this context; duplicate traces of the same session: dq1frnm1, j3b5kl1m, r8lftkyn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 469.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 469.0 9.7 33.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pysuihd7",
   "source_url": "https://omlx.ai/benchmarks/performance/pysuihd7",
   "tg_tokens": null,
   "tps": 9.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pysuihd7-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: dq1frnm1, j3b5kl1m, r8lftkyn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 345.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 345.9 9.4 37.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pysuihd7",
   "source_url": "https://omlx.ai/benchmarks/performance/pysuihd7",
   "tg_tokens": null,
   "tps": 9.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pysuihd7-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 41.5 GB at this context; duplicate traces of the same session: dq1frnm1, j3b5kl1m, r8lftkyn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 300.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 195k 300.4 8.4 41.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pysuihd7",
   "source_url": "https://omlx.ai/benchmarks/performance/pysuihd7",
   "tg_tokens": null,
   "tps": 8.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-b8x4nzy4-ctx16384",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: aiqiu7yx; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 477.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 477.4 32.8 26.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b8x4nzy4",
   "source_url": "https://omlx.ai/benchmarks/performance/b8x4nzy4",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-b8x4nzy4-ctx65536",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 29.3 GB at this context; duplicate traces of the same session: aiqiu7yx; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 387.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 387.2 24.4 29.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b8x4nzy4",
   "source_url": "https://omlx.ai/benchmarks/performance/b8x4nzy4",
   "tg_tokens": null,
   "tps": 24.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-m2me3xvq-ctx16384",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: oxqcammw; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3027.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-35B-A3B-Distill-mlx-4Bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 3,027 92.0 20.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m2me3xvq",
   "source_url": "https://omlx.ai/benchmarks/performance/m2me3xvq",
   "tg_tokens": null,
   "tps": 92.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-m2me3xvq-ctx65536",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: oxqcammw; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1964.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-35B-A3B-Distill-mlx-4Bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 1,964 63.8 21.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m2me3xvq",
   "source_url": "https://omlx.ai/benchmarks/performance/m2me3xvq",
   "tg_tokens": null,
   "tps": 63.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-o5xtkytu-ctx1024",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 230.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 230.5 20.2 5.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o5xtkytu",
   "source_url": "https://omlx.ai/benchmarks/performance/o5xtkytu",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-o5xtkytu-ctx4096",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 236.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 236.7 15.7 6.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o5xtkytu",
   "source_url": "https://omlx.ai/benchmarks/performance/o5xtkytu",
   "tg_tokens": null,
   "tps": 15.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ecq1b0ol-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "mixed quantization oMLX build; oMLX community benchmark trace, single run at 4k context; peak memory 74.1 GB; TTFT 6972 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 587.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-Uncensored-Mixed-omlx on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 587.5 PP tok/s 38.3 TG tok/s 6972 TTFT (ms) 74.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ecq1b0ol",
   "source_url": "https://omlx.ai/benchmarks/performance/ecq1b0ol",
   "tg_tokens": null,
   "tps": 38.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-z8jcb66v-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.4 GB at this context; duplicate traces of the same session: 9kse6edh, 9usry3uo, brzcccm3, d85yevkd, dh3bp1sw, q3xr7cdd; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2745.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 2,745 147.0 100.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z8jcb66v",
   "source_url": "https://omlx.ai/benchmarks/performance/z8jcb66v",
   "tg_tokens": null,
   "tps": 147.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-z8jcb66v-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: 9kse6edh, 9usry3uo, brzcccm3, d85yevkd, dh3bp1sw, q3xr7cdd; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5599.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 5,599 133.5 101.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z8jcb66v",
   "source_url": "https://omlx.ai/benchmarks/performance/z8jcb66v",
   "tg_tokens": null,
   "tps": 133.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-z8jcb66v-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.2 GB at this context; duplicate traces of the same session: 9kse6edh, 9usry3uo, brzcccm3, d85yevkd, dh3bp1sw, q3xr7cdd; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6297.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 6,297 135.2 102.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z8jcb66v",
   "source_url": "https://omlx.ai/benchmarks/performance/z8jcb66v",
   "tg_tokens": null,
   "tps": 135.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-z8jcb66v-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.8 GB at this context; duplicate traces of the same session: 9kse6edh, 9usry3uo, brzcccm3, d85yevkd, dh3bp1sw, q3xr7cdd; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6162.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 6,162 107.3 102.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z8jcb66v",
   "source_url": "https://omlx.ai/benchmarks/performance/z8jcb66v",
   "tg_tokens": null,
   "tps": 107.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-z8jcb66v-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.5 GB at this context; duplicate traces of the same session: 9kse6edh, 9usry3uo, brzcccm3, d85yevkd, dh3bp1sw, q3xr7cdd; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6028.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 6,028 100.6 104.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z8jcb66v",
   "source_url": "https://omlx.ai/benchmarks/performance/z8jcb66v",
   "tg_tokens": null,
   "tps": 100.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-z8jcb66v-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 106.8 GB at this context; duplicate traces of the same session: 9kse6edh, 9usry3uo, brzcccm3, d85yevkd, dh3bp1sw, q3xr7cdd; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5893.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,893 104.3 106.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z8jcb66v",
   "source_url": "https://omlx.ai/benchmarks/performance/z8jcb66v",
   "tg_tokens": null,
   "tps": 104.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-z8jcb66v-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.6 GB at this context; duplicate traces of the same session: 9kse6edh, 9usry3uo, brzcccm3, d85yevkd, dh3bp1sw, q3xr7cdd; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5704.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 5,704 100.0 109.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z8jcb66v",
   "source_url": "https://omlx.ai/benchmarks/performance/z8jcb66v",
   "tg_tokens": null,
   "tps": 100.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-smgnjbn9-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: 702guf2i, chjpt2ns, waonzjbd; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3092.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 3,092 152.6 100.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/smgnjbn9",
   "source_url": "https://omlx.ai/benchmarks/performance/smgnjbn9",
   "tg_tokens": null,
   "tps": 152.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-smgnjbn9-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.2 GB at this context; duplicate traces of the same session: 702guf2i, chjpt2ns, waonzjbd; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6298.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 6,298 141.4 102.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/smgnjbn9",
   "source_url": "https://omlx.ai/benchmarks/performance/smgnjbn9",
   "tg_tokens": null,
   "tps": 141.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-smgnjbn9-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.6 GB at this context; duplicate traces of the same session: 702guf2i, chjpt2ns, waonzjbd; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6273.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 6,273 147.6 102.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/smgnjbn9",
   "source_url": "https://omlx.ai/benchmarks/performance/smgnjbn9",
   "tg_tokens": null,
   "tps": 147.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-smgnjbn9-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.2 GB at this context; duplicate traces of the same session: 702guf2i, chjpt2ns, waonzjbd; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6254.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 6,254 131.9 103.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/smgnjbn9",
   "source_url": "https://omlx.ai/benchmarks/performance/smgnjbn9",
   "tg_tokens": null,
   "tps": 131.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-smgnjbn9-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.7 GB at this context; duplicate traces of the same session: 702guf2i, chjpt2ns, waonzjbd; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6107.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 6,107 138.4 104.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/smgnjbn9",
   "source_url": "https://omlx.ai/benchmarks/performance/smgnjbn9",
   "tg_tokens": null,
   "tps": 138.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-smgnjbn9-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.0 GB at this context; duplicate traces of the same session: 702guf2i, chjpt2ns, waonzjbd; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5886.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,886 106.5 109.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/smgnjbn9",
   "source_url": "https://omlx.ai/benchmarks/performance/smgnjbn9",
   "tg_tokens": null,
   "tps": 106.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-smgnjbn9-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.7 GB at this context; duplicate traces of the same session: 702guf2i, chjpt2ns, waonzjbd; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5692.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 5,692 96.6 110.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/smgnjbn9",
   "source_url": "https://omlx.ai/benchmarks/performance/smgnjbn9",
   "tg_tokens": null,
   "tps": 96.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-z3pb4f0t-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.6 GB at this context; duplicate traces of the same session: 9zf5kol0, bmdc82h6, ekoiddi7, gta8rh02, xtb5m297; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2441.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 2,441 151.4 142.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z3pb4f0t",
   "source_url": "https://omlx.ai/benchmarks/performance/z3pb4f0t",
   "tg_tokens": null,
   "tps": 151.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-z3pb4f0t-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: 9zf5kol0, bmdc82h6, ekoiddi7, gta8rh02, xtb5m297; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4470.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 4,470 138.7 143.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z3pb4f0t",
   "source_url": "https://omlx.ai/benchmarks/performance/z3pb4f0t",
   "tg_tokens": null,
   "tps": 138.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-z3pb4f0t-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.1 GB at this context; duplicate traces of the same session: 9zf5kol0, bmdc82h6, ekoiddi7, gta8rh02, xtb5m297; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4962.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 16k 4,962 132.7 145.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z3pb4f0t",
   "source_url": "https://omlx.ai/benchmarks/performance/z3pb4f0t",
   "tg_tokens": null,
   "tps": 132.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-z3pb4f0t-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 146.8 GB at this context; duplicate traces of the same session: 9zf5kol0, bmdc82h6, ekoiddi7, gta8rh02, xtb5m297; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4853.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 64k 4,853 124.7 146.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z3pb4f0t",
   "source_url": "https://omlx.ai/benchmarks/performance/z3pb4f0t",
   "tg_tokens": null,
   "tps": 124.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-z3pb4f0t-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 150.6 GB at this context; duplicate traces of the same session: 9zf5kol0, bmdc82h6, ekoiddi7, gta8rh02, xtb5m297; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4712.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 128k 4,712 100.7 150.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z3pb4f0t",
   "source_url": "https://omlx.ai/benchmarks/performance/z3pb4f0t",
   "tg_tokens": null,
   "tps": 100.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-z3pb4f0t-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 153.3 GB at this context; duplicate traces of the same session: 9zf5kol0, bmdc82h6, ekoiddi7, gta8rh02, xtb5m297; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4588.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 195k 4,588 85.5 153.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z3pb4f0t",
   "source_url": "https://omlx.ai/benchmarks/performance/z3pb4f0t",
   "tg_tokens": null,
   "tps": 85.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-m20fpzmx-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 21.9 GB; TTFT 120358 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 136.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 16k tokens 136.1 PP tok/s 20.0 TG tok/s 120358 TTFT (ms) 21.9 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m20fpzmx",
   "source_url": "https://omlx.ai/benchmarks/performance/m20fpzmx",
   "tg_tokens": null,
   "tps": 20.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-0280kswg-ctx4096",
   "model": "Ternary-Bonsai-2-27B-Uncensored-Heretic-v2",
   "notes": "MLX compact build; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: jsdl6my2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 174.0,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-Uncensored-Heretic-v2-mlx-2bit-compact on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 2bit; 4k 174.0 22.5 18.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0280kswg",
   "source_url": "https://omlx.ai/benchmarks/performance/0280kswg",
   "tg_tokens": null,
   "tps": 22.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-0280kswg-ctx16384",
   "model": "Ternary-Bonsai-2-27B-Uncensored-Heretic-v2",
   "notes": "MLX compact build; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: jsdl6my2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 155.4,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-Uncensored-Heretic-v2-mlx-2bit-compact on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 2bit; 16k 155.4 14.9 19.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0280kswg",
   "source_url": "https://omlx.ai/benchmarks/performance/0280kswg",
   "tg_tokens": null,
   "tps": 14.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1bxyokn5-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "QAT OptiQ build; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: yl0u1dve; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1892.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 1,892 49.9 21.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1bxyokn5",
   "source_url": "https://omlx.ai/benchmarks/performance/1bxyokn5",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-1bxyokn5-ctx65536",
   "model": "gemma4 26B.A4B",
   "notes": "QAT OptiQ build; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: yl0u1dve; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1435.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 1,435 32.9 23.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1bxyokn5",
   "source_url": "https://omlx.ai/benchmarks/performance/1bxyokn5",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-w42nrqat-ctx4096",
   "model": "supergemma4-26b-uncensored",
   "notes": "oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: gnjbr2k0; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 501.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "supergemma4-26b-uncensored-mlx-4bit-v2 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 501.0 51.0 14.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w42nrqat",
   "source_url": "https://omlx.ai/benchmarks/performance/w42nrqat",
   "tg_tokens": null,
   "tps": 51.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-w42nrqat-ctx8192",
   "model": "supergemma4-26b-uncensored",
   "notes": "oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: gnjbr2k0; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 490.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "supergemma4-26b-uncensored-mlx-4bit-v2 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 8k 490.7 47.9 14.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w42nrqat",
   "source_url": "https://omlx.ai/benchmarks/performance/w42nrqat",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-c1j18dv6-ctx4096",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 12.8 GB at this context; duplicate traces of the same session: 3ia1y290; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 752.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-oQ8e on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 4k 752.1 71.8 12.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c1j18dv6",
   "source_url": "https://omlx.ai/benchmarks/performance/c1j18dv6",
   "tg_tokens": null,
   "tps": 71.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-c1j18dv6-ctx8192",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 12.8 GB at this context; duplicate traces of the same session: 3ia1y290; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 722.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-oQ8e on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 8k 722.4 68.6 12.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c1j18dv6",
   "source_url": "https://omlx.ai/benchmarks/performance/c1j18dv6",
   "tg_tokens": null,
   "tps": 68.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-7ntbsnll-ctx1024",
   "model": "Ornith-1.5-35B-A3B-uncensored",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 27ci4jq9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1325.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-uncensored-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 1,325 102.1 19.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ntbsnll",
   "source_url": "https://omlx.ai/benchmarks/performance/7ntbsnll",
   "tg_tokens": null,
   "tps": 102.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-7ntbsnll-ctx8192",
   "model": "Ornith-1.5-35B-A3B-uncensored",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 27ci4jq9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1683.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-uncensored-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 8k 1,683 95.0 20.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ntbsnll",
   "source_url": "https://omlx.ai/benchmarks/performance/7ntbsnll",
   "tg_tokens": null,
   "tps": 95.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-b3imwjf1-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 6.4 GB; TTFT 10896 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 375.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k tokens 375.9 PP tok/s 46.2 TG tok/s 10896 TTFT (ms) 6.4 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b3imwjf1",
   "source_url": "https://omlx.ai/benchmarks/performance/b3imwjf1",
   "tg_tokens": null,
   "tps": 46.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-577vwiw7-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 21.4 GB; TTFT 5276 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 776.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k tokens 776.4 PP tok/s 90.3 TG tok/s 5276 TTFT (ms) 21.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/577vwiw7",
   "source_url": "https://omlx.ai/benchmarks/performance/577vwiw7",
   "tg_tokens": null,
   "tps": 90.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-07hsix7c-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: 8qsd8obu, trq1m6yr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1696.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,696 101.0 22.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/07hsix7c",
   "source_url": "https://omlx.ai/benchmarks/performance/07hsix7c",
   "tg_tokens": null,
   "tps": 101.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-07hsix7c-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: 8qsd8obu, trq1m6yr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1511.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,511 103.6 22.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/07hsix7c",
   "source_url": "https://omlx.ai/benchmarks/performance/07hsix7c",
   "tg_tokens": null,
   "tps": 103.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-07hsix7c-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 8qsd8obu, trq1m6yr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1231.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 1,231 97.3 23.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/07hsix7c",
   "source_url": "https://omlx.ai/benchmarks/performance/07hsix7c",
   "tg_tokens": null,
   "tps": 97.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-36athnad-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 5ckip1r5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1762.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,762 96.3 21.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/36athnad",
   "source_url": "https://omlx.ai/benchmarks/performance/36athnad",
   "tg_tokens": null,
   "tps": 96.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-36athnad-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: 5ckip1r5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1300.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 1,300 83.8 22.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/36athnad",
   "source_url": "https://omlx.ai/benchmarks/performance/36athnad",
   "tg_tokens": null,
   "tps": 83.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-5wtr4o4o-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: mi5iyzp0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1519.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,519 108.2 22.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5wtr4o4o",
   "source_url": "https://omlx.ai/benchmarks/performance/5wtr4o4o",
   "tg_tokens": null,
   "tps": 108.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-5wtr4o4o-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: mi5iyzp0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1211.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 1,211 104.2 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5wtr4o4o",
   "source_url": "https://omlx.ai/benchmarks/performance/5wtr4o4o",
   "tg_tokens": null,
   "tps": 104.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-7exwcf53-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: ihanxn6e, qvv4bckl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1700.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,700 84.6 22.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7exwcf53",
   "source_url": "https://omlx.ai/benchmarks/performance/7exwcf53",
   "tg_tokens": null,
   "tps": 84.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-7exwcf53-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: ihanxn6e, qvv4bckl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1513.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,513 95.2 22.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7exwcf53",
   "source_url": "https://omlx.ai/benchmarks/performance/7exwcf53",
   "tg_tokens": null,
   "tps": 95.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-7exwcf53-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: ihanxn6e, qvv4bckl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1180.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 1,180 96.8 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7exwcf53",
   "source_url": "https://omlx.ai/benchmarks/performance/7exwcf53",
   "tg_tokens": null,
   "tps": 96.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-wrtzfzhk-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 8k context; peak memory 29.9 GB; TTFT 5910 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1386.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 8k tokens 1,386 PP tok/s 31.9 TG tok/s 5910 TTFT (ms) 29.9 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wrtzfzhk",
   "source_url": "https://omlx.ai/benchmarks/performance/wrtzfzhk",
   "tg_tokens": null,
   "tps": 31.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ueka82l0-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 31.4 GB; TTFT 25824 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1269.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 32k tokens 1,269 PP tok/s 29.8 TG tok/s 25824 TTFT (ms) 31.4 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ueka82l0",
   "source_url": "https://omlx.ai/benchmarks/performance/ueka82l0",
   "tg_tokens": null,
   "tps": 29.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-84a348t1-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 245.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 245.7 20.1 22.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/84a348t1",
   "source_url": "https://omlx.ai/benchmarks/performance/84a348t1",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-84a348t1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 227.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 227.6 16.2 25.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/84a348t1",
   "source_url": "https://omlx.ai/benchmarks/performance/84a348t1",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-84a348t1-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 69.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 69.3 6.3 28.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/84a348t1",
   "source_url": "https://omlx.ai/benchmarks/performance/84a348t1",
   "tg_tokens": null,
   "tps": 6.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-g1fwtaex-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 8k context; peak memory 55.4 GB; TTFT 5104 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1605.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.8-27B-bf16 on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 bf16; 8k tokens 1,605 PP tok/s 35.3 TG tok/s 5104 TTFT (ms) 55.4 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g1fwtaex",
   "source_url": "https://omlx.ai/benchmarks/performance/g1fwtaex",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-qkrtdkyn-ctx1024",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 230.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 230.5 20.2 5.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qkrtdkyn",
   "source_url": "https://omlx.ai/benchmarks/performance/qkrtdkyn",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-qkrtdkyn-ctx4096",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 236.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 236.7 15.7 6.2 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qkrtdkyn",
   "source_url": "https://omlx.ai/benchmarks/performance/qkrtdkyn",
   "tg_tokens": null,
   "tps": 15.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-towgrua5-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 75.8 GB; TTFT 232853 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 562.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k tokens 562.9 PP tok/s 34.6 TG tok/s 232853 TTFT (ms) 75.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/towgrua5",
   "source_url": "https://omlx.ai/benchmarks/performance/towgrua5",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-k5rjpftq-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: e5gan2b5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 528.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 528.7 58.5 70.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k5rjpftq",
   "source_url": "https://omlx.ai/benchmarks/performance/k5rjpftq",
   "tg_tokens": null,
   "tps": 58.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-k5rjpftq-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: e5gan2b5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 657.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 657.3 53.3 71.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k5rjpftq",
   "source_url": "https://omlx.ai/benchmarks/performance/k5rjpftq",
   "tg_tokens": null,
   "tps": 53.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-00zf4tru-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 5e26kg2z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 519.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 519.8 64.9 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/00zf4tru",
   "source_url": "https://omlx.ai/benchmarks/performance/00zf4tru",
   "tg_tokens": null,
   "tps": 64.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-00zf4tru-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: 5e26kg2z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 655.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 655.1 49.6 71.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/00zf4tru",
   "source_url": "https://omlx.ai/benchmarks/performance/00zf4tru",
   "tg_tokens": null,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-sd1f3u8t-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: cqc0q1cw, jm2xe0ki; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3092.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 3,092 152.6 100.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sd1f3u8t",
   "source_url": "https://omlx.ai/benchmarks/performance/sd1f3u8t",
   "tg_tokens": null,
   "tps": 152.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-sd1f3u8t-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.2 GB at this context; duplicate traces of the same session: cqc0q1cw, jm2xe0ki; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6298.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 6,298 141.4 102.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sd1f3u8t",
   "source_url": "https://omlx.ai/benchmarks/performance/sd1f3u8t",
   "tg_tokens": null,
   "tps": 141.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-sd1f3u8t-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.6 GB at this context; duplicate traces of the same session: cqc0q1cw, jm2xe0ki; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6273.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 6,273 147.6 102.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sd1f3u8t",
   "source_url": "https://omlx.ai/benchmarks/performance/sd1f3u8t",
   "tg_tokens": null,
   "tps": 147.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-sd1f3u8t-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.2 GB at this context; duplicate traces of the same session: cqc0q1cw, jm2xe0ki; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6254.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 6,254 131.9 103.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sd1f3u8t",
   "source_url": "https://omlx.ai/benchmarks/performance/sd1f3u8t",
   "tg_tokens": null,
   "tps": 131.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-sd1f3u8t-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.7 GB at this context; duplicate traces of the same session: cqc0q1cw, jm2xe0ki; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6107.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 6,107 138.4 104.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sd1f3u8t",
   "source_url": "https://omlx.ai/benchmarks/performance/sd1f3u8t",
   "tg_tokens": null,
   "tps": 138.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-sd1f3u8t-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.0 GB at this context; duplicate traces of the same session: cqc0q1cw, jm2xe0ki; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5886.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,886 106.5 109.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sd1f3u8t",
   "source_url": "https://omlx.ai/benchmarks/performance/sd1f3u8t",
   "tg_tokens": null,
   "tps": 106.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-sd1f3u8t-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.7 GB at this context; duplicate traces of the same session: cqc0q1cw, jm2xe0ki; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5692.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 5,692 96.6 110.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sd1f3u8t",
   "source_url": "https://omlx.ai/benchmarks/performance/sd1f3u8t",
   "tg_tokens": null,
   "tps": 96.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-zjguqoov-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: 3chc5o1u, 7zq8wptr, b98d06m9, u7jyj7xz, wnlquofu; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2748.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 2,748 163.7 100.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zjguqoov",
   "source_url": "https://omlx.ai/benchmarks/performance/zjguqoov",
   "tg_tokens": null,
   "tps": 163.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-zjguqoov-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: 3chc5o1u, 7zq8wptr, b98d06m9, u7jyj7xz, wnlquofu; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5527.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 5,527 137.8 101.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zjguqoov",
   "source_url": "https://omlx.ai/benchmarks/performance/zjguqoov",
   "tg_tokens": null,
   "tps": 137.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-zjguqoov-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.2 GB at this context; duplicate traces of the same session: 3chc5o1u, 7zq8wptr, b98d06m9, u7jyj7xz, wnlquofu; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6289.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 6,289 136.5 102.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zjguqoov",
   "source_url": "https://omlx.ai/benchmarks/performance/zjguqoov",
   "tg_tokens": null,
   "tps": 136.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-zjguqoov-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.4 GB at this context; duplicate traces of the same session: 3chc5o1u, 7zq8wptr, b98d06m9, u7jyj7xz, wnlquofu; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6123.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 6,123 128.6 104.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zjguqoov",
   "source_url": "https://omlx.ai/benchmarks/performance/zjguqoov",
   "tg_tokens": null,
   "tps": 128.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-zjguqoov-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.1 GB at this context; duplicate traces of the same session: 3chc5o1u, 7zq8wptr, b98d06m9, u7jyj7xz, wnlquofu; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5913.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,913 107.4 109.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zjguqoov",
   "source_url": "https://omlx.ai/benchmarks/performance/zjguqoov",
   "tg_tokens": null,
   "tps": 107.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-zjguqoov-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.7 GB at this context; duplicate traces of the same session: 3chc5o1u, 7zq8wptr, b98d06m9, u7jyj7xz, wnlquofu; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5655.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 5,655 89.7 110.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zjguqoov",
   "source_url": "https://omlx.ai/benchmarks/performance/zjguqoov",
   "tg_tokens": null,
   "tps": 89.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-h8uvjrl8-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.4 GB at this context; duplicate traces of the same session: 1ywecdt7, 9dtv846x, ees0if1e, mzynq8za, pdxtj7sg; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2777.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 2,777 142.6 100.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h8uvjrl8",
   "source_url": "https://omlx.ai/benchmarks/performance/h8uvjrl8",
   "tg_tokens": null,
   "tps": 142.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-h8uvjrl8-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.2 GB at this context; duplicate traces of the same session: 1ywecdt7, 9dtv846x, ees0if1e, mzynq8za, pdxtj7sg; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6332.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 6,332 139.0 102.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h8uvjrl8",
   "source_url": "https://omlx.ai/benchmarks/performance/h8uvjrl8",
   "tg_tokens": null,
   "tps": 139.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-h8uvjrl8-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.4 GB at this context; duplicate traces of the same session: 1ywecdt7, 9dtv846x, ees0if1e, mzynq8za, pdxtj7sg; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6167.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 6,167 68.7 103.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h8uvjrl8",
   "source_url": "https://omlx.ai/benchmarks/performance/h8uvjrl8",
   "tg_tokens": null,
   "tps": 68.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-h8uvjrl8-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.5 GB at this context; duplicate traces of the same session: 1ywecdt7, 9dtv846x, ees0if1e, mzynq8za, pdxtj7sg; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6098.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 6,098 138.7 104.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h8uvjrl8",
   "source_url": "https://omlx.ai/benchmarks/performance/h8uvjrl8",
   "tg_tokens": null,
   "tps": 138.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-h8uvjrl8-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 106.8 GB at this context; duplicate traces of the same session: 1ywecdt7, 9dtv846x, ees0if1e, mzynq8za, pdxtj7sg; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5923.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,923 116.9 106.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h8uvjrl8",
   "source_url": "https://omlx.ai/benchmarks/performance/h8uvjrl8",
   "tg_tokens": null,
   "tps": 116.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-h8uvjrl8-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.6 GB at this context; duplicate traces of the same session: 1ywecdt7, 9dtv846x, ees0if1e, mzynq8za, pdxtj7sg; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5719.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 5,719 99.4 109.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h8uvjrl8",
   "source_url": "https://omlx.ai/benchmarks/performance/h8uvjrl8",
   "tg_tokens": null,
   "tps": 99.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-2divfgoj-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: 2y21315v, 388ew41a, i74de9g0, vhk48kxa, z4oc0ktq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2960.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 2,960 154.2 100.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2divfgoj",
   "source_url": "https://omlx.ai/benchmarks/performance/2divfgoj",
   "tg_tokens": null,
   "tps": 154.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-2divfgoj-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: 2y21315v, 388ew41a, i74de9g0, vhk48kxa, z4oc0ktq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5592.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 5,592 105.9 101.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2divfgoj",
   "source_url": "https://omlx.ai/benchmarks/performance/2divfgoj",
   "tg_tokens": null,
   "tps": 105.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-2divfgoj-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.2 GB at this context; duplicate traces of the same session: 2y21315v, 388ew41a, i74de9g0, vhk48kxa, z4oc0ktq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6296.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 6,296 126.2 102.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2divfgoj",
   "source_url": "https://omlx.ai/benchmarks/performance/2divfgoj",
   "tg_tokens": null,
   "tps": 126.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-2divfgoj-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.2 GB at this context; duplicate traces of the same session: 2y21315v, 388ew41a, i74de9g0, vhk48kxa, z4oc0ktq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6265.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 6,265 147.3 103.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2divfgoj",
   "source_url": "https://omlx.ai/benchmarks/performance/2divfgoj",
   "tg_tokens": null,
   "tps": 147.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-2divfgoj-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.4 GB at this context; duplicate traces of the same session: 2y21315v, 388ew41a, i74de9g0, vhk48kxa, z4oc0ktq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6127.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 6,127 135.0 104.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2divfgoj",
   "source_url": "https://omlx.ai/benchmarks/performance/2divfgoj",
   "tg_tokens": null,
   "tps": 135.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-2divfgoj-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.0 GB at this context; duplicate traces of the same session: 2y21315v, 388ew41a, i74de9g0, vhk48kxa, z4oc0ktq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5922.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,922 107.9 109.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2divfgoj",
   "source_url": "https://omlx.ai/benchmarks/performance/2divfgoj",
   "tg_tokens": null,
   "tps": 107.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-2divfgoj-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.7 GB at this context; duplicate traces of the same session: 2y21315v, 388ew41a, i74de9g0, vhk48kxa, z4oc0ktq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5728.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 5,728 89.0 110.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2divfgoj",
   "source_url": "https://omlx.ai/benchmarks/performance/2divfgoj",
   "tg_tokens": null,
   "tps": 89.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ai3vb922-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.6 GB at this context; duplicate traces of the same session: 0m1h86gs, 59mtohd3, i11814z3, ola3ijqj, wgu44cvf; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2463.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 2,463 145.2 142.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ai3vb922",
   "source_url": "https://omlx.ai/benchmarks/performance/ai3vb922",
   "tg_tokens": null,
   "tps": 145.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ai3vb922-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: 0m1h86gs, 59mtohd3, i11814z3, ola3ijqj, wgu44cvf; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4499.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 4,499 132.9 143.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ai3vb922",
   "source_url": "https://omlx.ai/benchmarks/performance/ai3vb922",
   "tg_tokens": null,
   "tps": 132.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ai3vb922-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 144.5 GB at this context; duplicate traces of the same session: 0m1h86gs, 59mtohd3, i11814z3, ola3ijqj, wgu44cvf; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4942.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 8k 4,942 134.4 144.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ai3vb922",
   "source_url": "https://omlx.ai/benchmarks/performance/ai3vb922",
   "tg_tokens": null,
   "tps": 134.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ai3vb922-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 147.0 GB at this context; duplicate traces of the same session: 0m1h86gs, 59mtohd3, i11814z3, ola3ijqj, wgu44cvf; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4826.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 64k 4,826 132.3 147.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ai3vb922",
   "source_url": "https://omlx.ai/benchmarks/performance/ai3vb922",
   "tg_tokens": null,
   "tps": 132.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ai3vb922-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 149.3 GB at this context; duplicate traces of the same session: 0m1h86gs, 59mtohd3, i11814z3, ola3ijqj, wgu44cvf; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4713.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 128k 4,713 105.6 149.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ai3vb922",
   "source_url": "https://omlx.ai/benchmarks/performance/ai3vb922",
   "tg_tokens": null,
   "tps": 105.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ai3vb922-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 152.1 GB at this context; duplicate traces of the same session: 0m1h86gs, 59mtohd3, i11814z3, ola3ijqj, wgu44cvf; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4575.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 195k 4,575 84.6 152.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ai3vb922",
   "source_url": "https://omlx.ai/benchmarks/performance/ai3vb922",
   "tg_tokens": null,
   "tps": 84.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-fkneftgs-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.6 GB at this context; duplicate traces of the same session: 3kqo7zq7, 80kn6tfc, m0d2cgf8, xh20mwsd, zngyfuwj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2440.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 2,440 148.1 142.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fkneftgs",
   "source_url": "https://omlx.ai/benchmarks/performance/fkneftgs",
   "tg_tokens": null,
   "tps": 148.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-fkneftgs-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: 3kqo7zq7, 80kn6tfc, m0d2cgf8, xh20mwsd, zngyfuwj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4476.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 4,476 102.1 143.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fkneftgs",
   "source_url": "https://omlx.ai/benchmarks/performance/fkneftgs",
   "tg_tokens": null,
   "tps": 102.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-fkneftgs-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 144.5 GB at this context; duplicate traces of the same session: 3kqo7zq7, 80kn6tfc, m0d2cgf8, xh20mwsd, zngyfuwj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4928.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 8k 4,928 132.1 144.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fkneftgs",
   "source_url": "https://omlx.ai/benchmarks/performance/fkneftgs",
   "tg_tokens": null,
   "tps": 132.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-fkneftgs-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 146.8 GB at this context; duplicate traces of the same session: 3kqo7zq7, 80kn6tfc, m0d2cgf8, xh20mwsd, zngyfuwj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4848.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 64k 4,848 122.2 146.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fkneftgs",
   "source_url": "https://omlx.ai/benchmarks/performance/fkneftgs",
   "tg_tokens": null,
   "tps": 122.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-fkneftgs-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 150.6 GB at this context; duplicate traces of the same session: 3kqo7zq7, 80kn6tfc, m0d2cgf8, xh20mwsd, zngyfuwj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4707.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 128k 4,707 103.5 150.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fkneftgs",
   "source_url": "https://omlx.ai/benchmarks/performance/fkneftgs",
   "tg_tokens": null,
   "tps": 103.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-fkneftgs-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 153.3 GB at this context; duplicate traces of the same session: 3kqo7zq7, 80kn6tfc, m0d2cgf8, xh20mwsd, zngyfuwj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4584.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 195k 4,584 98.7 153.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fkneftgs",
   "source_url": "https://omlx.ai/benchmarks/performance/fkneftgs",
   "tg_tokens": null,
   "tps": 98.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-bzsyxfv5-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 187.1 GB; TTFT 14784 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4433.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 64k tokens 4,433 PP tok/s 103.6 TG tok/s 14784 TTFT (ms) 187.1 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bzsyxfv5",
   "source_url": "https://omlx.ai/benchmarks/performance/bzsyxfv5",
   "tg_tokens": null,
   "tps": 103.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ur3wdobp-ctx1024",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 177.6 GB at this context; duplicate traces of the same session: f8099sn9, mju7glha; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 515.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 515.6 47.5 177.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ur3wdobp",
   "source_url": "https://omlx.ai/benchmarks/performance/ur3wdobp",
   "tg_tokens": null,
   "tps": 47.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ur3wdobp-ctx4096",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 178.8 GB at this context; duplicate traces of the same session: f8099sn9, mju7glha; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 592.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 592.5 48.2 178.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ur3wdobp",
   "source_url": "https://omlx.ai/benchmarks/performance/ur3wdobp",
   "tg_tokens": null,
   "tps": 48.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ur3wdobp-ctx8192",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 179.1 GB at this context; duplicate traces of the same session: f8099sn9, mju7glha; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 548.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 548.3 57.7 179.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ur3wdobp",
   "source_url": "https://omlx.ai/benchmarks/performance/ur3wdobp",
   "tg_tokens": null,
   "tps": 57.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-dp1a7263-ctx1024",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: cicpxnba, nxbrxcas, t8p49o08; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1334.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-Uncensored-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,334 74.4 177.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dp1a7263",
   "source_url": "https://omlx.ai/benchmarks/performance/dp1a7263",
   "tg_tokens": null,
   "tps": 74.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-dp1a7263-ctx32768",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.2 GB at this context; duplicate traces of the same session: cicpxnba, nxbrxcas, t8p49o08; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1909.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-Uncensored-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,909 93.6 179.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dp1a7263",
   "source_url": "https://omlx.ai/benchmarks/performance/dp1a7263",
   "tg_tokens": null,
   "tps": 93.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-dp1a7263-ctx65536",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.7 GB at this context; duplicate traces of the same session: cicpxnba, nxbrxcas, t8p49o08; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1880.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-Uncensored-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,880 86.7 179.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dp1a7263",
   "source_url": "https://omlx.ai/benchmarks/performance/dp1a7263",
   "tg_tokens": null,
   "tps": 86.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-dp1a7263-ctx131072",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.6 GB at this context; duplicate traces of the same session: cicpxnba, nxbrxcas, t8p49o08; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1845.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-Uncensored-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 1,845 134.8 180.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dp1a7263",
   "source_url": "https://omlx.ai/benchmarks/performance/dp1a7263",
   "tg_tokens": null,
   "tps": 134.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-pjoqltmv-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: 0jr0oj8f, 0kusuabq, 8ghl4vk5, glstm378, jpp5r7ja, m2avet75; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1322.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,322 78.8 177.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pjoqltmv",
   "source_url": "https://omlx.ai/benchmarks/performance/pjoqltmv",
   "tg_tokens": null,
   "tps": 78.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-pjoqltmv-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: 0jr0oj8f, 0kusuabq, 8ghl4vk5, glstm378, jpp5r7ja, m2avet75; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2322.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 2,322 70.6 178.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pjoqltmv",
   "source_url": "https://omlx.ai/benchmarks/performance/pjoqltmv",
   "tg_tokens": null,
   "tps": 70.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-pjoqltmv-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.9 GB at this context; duplicate traces of the same session: 0jr0oj8f, 0kusuabq, 8ghl4vk5, glstm378, jpp5r7ja, m2avet75; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2323.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 2,323 89.5 178.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pjoqltmv",
   "source_url": "https://omlx.ai/benchmarks/performance/pjoqltmv",
   "tg_tokens": null,
   "tps": 89.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-pjoqltmv-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.0 GB at this context; duplicate traces of the same session: 0jr0oj8f, 0kusuabq, 8ghl4vk5, glstm378, jpp5r7ja, m2avet75; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2318.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 2,318 156.0 179.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pjoqltmv",
   "source_url": "https://omlx.ai/benchmarks/performance/pjoqltmv",
   "tg_tokens": null,
   "tps": 156.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-pjoqltmv-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.3 GB at this context; duplicate traces of the same session: 0jr0oj8f, 0kusuabq, 8ghl4vk5, glstm378, jpp5r7ja, m2avet75; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2309.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,309 108.7 179.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pjoqltmv",
   "source_url": "https://omlx.ai/benchmarks/performance/pjoqltmv",
   "tg_tokens": null,
   "tps": 108.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-pjoqltmv-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.7 GB at this context; duplicate traces of the same session: 0jr0oj8f, 0kusuabq, 8ghl4vk5, glstm378, jpp5r7ja, m2avet75; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2274.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,274 85.9 179.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pjoqltmv",
   "source_url": "https://omlx.ai/benchmarks/performance/pjoqltmv",
   "tg_tokens": null,
   "tps": 85.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-pjoqltmv-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.6 GB at this context; duplicate traces of the same session: 0jr0oj8f, 0kusuabq, 8ghl4vk5, glstm378, jpp5r7ja, m2avet75; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2230.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,230 146.2 180.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pjoqltmv",
   "source_url": "https://omlx.ai/benchmarks/performance/pjoqltmv",
   "tg_tokens": null,
   "tps": 146.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-2zgdt05s-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: 0inhdsy7, nqm7t1wm, qvy8ayz6, yplqcijo; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1503.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,503 83.2 177.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2zgdt05s",
   "source_url": "https://omlx.ai/benchmarks/performance/2zgdt05s",
   "tg_tokens": null,
   "tps": 83.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-2zgdt05s-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: 0inhdsy7, nqm7t1wm, qvy8ayz6, yplqcijo; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2324.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 2,324 67.8 178.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2zgdt05s",
   "source_url": "https://omlx.ai/benchmarks/performance/2zgdt05s",
   "tg_tokens": null,
   "tps": 67.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-2zgdt05s-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.9 GB at this context; duplicate traces of the same session: 0inhdsy7, nqm7t1wm, qvy8ayz6, yplqcijo; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2323.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 2,323 69.8 178.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2zgdt05s",
   "source_url": "https://omlx.ai/benchmarks/performance/2zgdt05s",
   "tg_tokens": null,
   "tps": 69.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-2zgdt05s-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.1 GB at this context; duplicate traces of the same session: 0inhdsy7, nqm7t1wm, qvy8ayz6, yplqcijo; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2320.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 2,320 150.1 179.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2zgdt05s",
   "source_url": "https://omlx.ai/benchmarks/performance/2zgdt05s",
   "tg_tokens": null,
   "tps": 150.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-2zgdt05s-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.4 GB at this context; duplicate traces of the same session: 0inhdsy7, nqm7t1wm, qvy8ayz6, yplqcijo; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2308.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,308 101.6 179.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2zgdt05s",
   "source_url": "https://omlx.ai/benchmarks/performance/2zgdt05s",
   "tg_tokens": null,
   "tps": 101.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-2zgdt05s-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.1 GB at this context; duplicate traces of the same session: 0inhdsy7, nqm7t1wm, qvy8ayz6, yplqcijo; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2273.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,273 83.0 180.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2zgdt05s",
   "source_url": "https://omlx.ai/benchmarks/performance/2zgdt05s",
   "tg_tokens": null,
   "tps": 83.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-2zgdt05s-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.3 GB at this context; duplicate traces of the same session: 0inhdsy7, nqm7t1wm, qvy8ayz6, yplqcijo; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2226.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,226 146.5 181.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2zgdt05s",
   "source_url": "https://omlx.ai/benchmarks/performance/2zgdt05s",
   "tg_tokens": null,
   "tps": 146.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-5iwlq0ka-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: 5cfz7sf4, b35siw4t, e05gprfj, erkf1yhr, fxe2hl3p; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1553.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,553 81.2 177.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5iwlq0ka",
   "source_url": "https://omlx.ai/benchmarks/performance/5iwlq0ka",
   "tg_tokens": null,
   "tps": 81.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-5iwlq0ka-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: 5cfz7sf4, b35siw4t, e05gprfj, erkf1yhr, fxe2hl3p; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2311.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 2,311 70.4 178.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5iwlq0ka",
   "source_url": "https://omlx.ai/benchmarks/performance/5iwlq0ka",
   "tg_tokens": null,
   "tps": 70.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-5iwlq0ka-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.0 GB at this context; duplicate traces of the same session: 5cfz7sf4, b35siw4t, e05gprfj, erkf1yhr, fxe2hl3p; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2309.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 2,309 154.4 179.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5iwlq0ka",
   "source_url": "https://omlx.ai/benchmarks/performance/5iwlq0ka",
   "tg_tokens": null,
   "tps": 154.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-5iwlq0ka-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.3 GB at this context; duplicate traces of the same session: 5cfz7sf4, b35siw4t, e05gprfj, erkf1yhr, fxe2hl3p; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2298.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,298 109.4 179.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5iwlq0ka",
   "source_url": "https://omlx.ai/benchmarks/performance/5iwlq0ka",
   "tg_tokens": null,
   "tps": 109.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-5iwlq0ka-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.7 GB at this context; duplicate traces of the same session: 5cfz7sf4, b35siw4t, e05gprfj, erkf1yhr, fxe2hl3p; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2263.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,263 90.4 179.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5iwlq0ka",
   "source_url": "https://omlx.ai/benchmarks/performance/5iwlq0ka",
   "tg_tokens": null,
   "tps": 90.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-5iwlq0ka-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.6 GB at this context; duplicate traces of the same session: 5cfz7sf4, b35siw4t, e05gprfj, erkf1yhr, fxe2hl3p; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2219.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,219 144.9 180.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5iwlq0ka",
   "source_url": "https://omlx.ai/benchmarks/performance/5iwlq0ka",
   "tg_tokens": null,
   "tps": 144.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-6jza9xq8-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: 6qzje0nf, j8p10yd3, o15rj0ve, t11yd96j, t8xwn1dr, yddn4gz2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1557.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,557 76.9 177.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6jza9xq8",
   "source_url": "https://omlx.ai/benchmarks/performance/6jza9xq8",
   "tg_tokens": null,
   "tps": 76.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-6jza9xq8-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: 6qzje0nf, j8p10yd3, o15rj0ve, t11yd96j, t8xwn1dr, yddn4gz2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2317.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 2,317 66.0 178.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6jza9xq8",
   "source_url": "https://omlx.ai/benchmarks/performance/6jza9xq8",
   "tg_tokens": null,
   "tps": 66.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-6jza9xq8-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.8 GB at this context; duplicate traces of the same session: 6qzje0nf, j8p10yd3, o15rj0ve, t11yd96j, t8xwn1dr, yddn4gz2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2316.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 2,316 92.3 178.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6jza9xq8",
   "source_url": "https://omlx.ai/benchmarks/performance/6jza9xq8",
   "tg_tokens": null,
   "tps": 92.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-6jza9xq8-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.0 GB at this context; duplicate traces of the same session: 6qzje0nf, j8p10yd3, o15rj0ve, t11yd96j, t8xwn1dr, yddn4gz2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2313.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 2,313 149.1 179.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6jza9xq8",
   "source_url": "https://omlx.ai/benchmarks/performance/6jza9xq8",
   "tg_tokens": null,
   "tps": 149.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-6jza9xq8-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.4 GB at this context; duplicate traces of the same session: 6qzje0nf, j8p10yd3, o15rj0ve, t11yd96j, t8xwn1dr, yddn4gz2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2305.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,305 103.1 179.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6jza9xq8",
   "source_url": "https://omlx.ai/benchmarks/performance/6jza9xq8",
   "tg_tokens": null,
   "tps": 103.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-6jza9xq8-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.1 GB at this context; duplicate traces of the same session: 6qzje0nf, j8p10yd3, o15rj0ve, t11yd96j, t8xwn1dr, yddn4gz2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2272.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,272 89.6 180.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6jza9xq8",
   "source_url": "https://omlx.ai/benchmarks/performance/6jza9xq8",
   "tg_tokens": null,
   "tps": 89.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-6jza9xq8-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.3 GB at this context; duplicate traces of the same session: 6qzje0nf, j8p10yd3, o15rj0ve, t11yd96j, t8xwn1dr, yddn4gz2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2232.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,232 145.6 181.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6jza9xq8",
   "source_url": "https://omlx.ai/benchmarks/performance/6jza9xq8",
   "tg_tokens": null,
   "tps": 145.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-o1h0axcv-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: m8r1ntgr; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 501.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 501.8 44.3 20.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o1h0axcv",
   "source_url": "https://omlx.ai/benchmarks/performance/o1h0axcv",
   "tg_tokens": null,
   "tps": 44.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-o1h0axcv-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: m8r1ntgr; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 437.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 16k 437.6 36.9 20.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o1h0axcv",
   "source_url": "https://omlx.ai/benchmarks/performance/o1h0axcv",
   "tg_tokens": null,
   "tps": 36.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-yghgb2c3-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: a1dy1912, qsh35g9x, y020ciqu; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 367.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 367.9 47.7 5.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yghgb2c3",
   "source_url": "https://omlx.ai/benchmarks/performance/yghgb2c3",
   "tg_tokens": null,
   "tps": 47.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-yghgb2c3-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: a1dy1912, qsh35g9x, y020ciqu; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 375.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 375.9 46.8 6.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yghgb2c3",
   "source_url": "https://omlx.ai/benchmarks/performance/yghgb2c3",
   "tg_tokens": null,
   "tps": 46.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-yghgb2c3-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: a1dy1912, qsh35g9x, y020ciqu; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 373.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 373.0 45.0 6.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yghgb2c3",
   "source_url": "https://omlx.ai/benchmarks/performance/yghgb2c3",
   "tg_tokens": null,
   "tps": 45.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-yghgb2c3-ctx32768",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.5 GB at this context; duplicate traces of the same session: a1dy1912, qsh35g9x, y020ciqu; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 332.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 332.8 37.2 7.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yghgb2c3",
   "source_url": "https://omlx.ai/benchmarks/performance/yghgb2c3",
   "tg_tokens": null,
   "tps": 37.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-h6kzg0p8-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: atoxfp5o; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 731.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 1k 731.4 97.3 19.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h6kzg0p8",
   "source_url": "https://omlx.ai/benchmarks/performance/h6kzg0p8",
   "tg_tokens": null,
   "tps": 97.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-h6kzg0p8-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: atoxfp5o; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 779.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 4k 779.0 94.7 20.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h6kzg0p8",
   "source_url": "https://omlx.ai/benchmarks/performance/h6kzg0p8",
   "tg_tokens": null,
   "tps": 94.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-cm5c3b2m-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: p91s0z0j; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 694.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 694.0 78.2 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cm5c3b2m",
   "source_url": "https://omlx.ai/benchmarks/performance/cm5c3b2m",
   "tg_tokens": null,
   "tps": 78.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-cm5c3b2m-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: p91s0z0j; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 596.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 596.7 66.4 22.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cm5c3b2m",
   "source_url": "https://omlx.ai/benchmarks/performance/cm5c3b2m",
   "tg_tokens": null,
   "tps": 66.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-mer59u29-ctx1024",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1240.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 1,240 40.1 12.7 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mer59u29",
   "source_url": "https://omlx.ai/benchmarks/performance/mer59u29",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-mer59u29-ctx4096",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.2 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1414.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 1,414 38.7 13.2 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mer59u29",
   "source_url": "https://omlx.ai/benchmarks/performance/mer59u29",
   "tg_tokens": null,
   "tps": 38.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-mer59u29-ctx8192",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.2 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1400.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 8k 1,400 38.4 13.2 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mer59u29",
   "source_url": "https://omlx.ai/benchmarks/performance/mer59u29",
   "tg_tokens": null,
   "tps": 38.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-sl44if4z-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: c7senf5k; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 113.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 1k 113.9 15.7 17.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sl44if4z",
   "source_url": "https://omlx.ai/benchmarks/performance/sl44if4z",
   "tg_tokens": null,
   "tps": 15.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-sl44if4z-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.2 GB at this context; duplicate traces of the same session: c7senf5k; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 109.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 4k 109.6 23.1 18.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sl44if4z",
   "source_url": "https://omlx.ai/benchmarks/performance/sl44if4z",
   "tg_tokens": null,
   "tps": 23.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-difg4ixn-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 8k context; peak memory 29.9 GB; TTFT 5910 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1386.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 8k tokens 1,386 PP tok/s 32.0 TG tok/s 5910 TTFT (ms) 29.9 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/difg4ixn",
   "source_url": "https://omlx.ai/benchmarks/performance/difg4ixn",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-uibcwpbi-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 1wy757g2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 245.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 245.7 20.1 22.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uibcwpbi",
   "source_url": "https://omlx.ai/benchmarks/performance/uibcwpbi",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-uibcwpbi-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: 1wy757g2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 227.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 227.6 16.2 25.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uibcwpbi",
   "source_url": "https://omlx.ai/benchmarks/performance/uibcwpbi",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-uibcwpbi-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 1wy757g2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 69.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 69.3 6.3 28.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uibcwpbi",
   "source_url": "https://omlx.ai/benchmarks/performance/uibcwpbi",
   "tg_tokens": null,
   "tps": 6.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-9t2imvwv-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: e8dxza31; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 491.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 491.0 51.2 70.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9t2imvwv",
   "source_url": "https://omlx.ai/benchmarks/performance/9t2imvwv",
   "tg_tokens": null,
   "tps": 51.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-9t2imvwv-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: e8dxza31; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 644.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 644.4 42.9 71.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9t2imvwv",
   "source_url": "https://omlx.ai/benchmarks/performance/9t2imvwv",
   "tg_tokens": null,
   "tps": 42.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-5s08siki-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.8 GB; TTFT 2934 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 349.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 349.0 PP tok/s 33.0 TG tok/s 2934 TTFT (ms) 70.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5s08siki",
   "source_url": "https://omlx.ai/benchmarks/performance/5s08siki",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ck5uhg3x-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2960.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 2,960 154.2 100.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ck5uhg3x",
   "source_url": "https://omlx.ai/benchmarks/performance/ck5uhg3x",
   "tg_tokens": null,
   "tps": 154.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ck5uhg3x-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5592.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 5,592 105.9 101.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ck5uhg3x",
   "source_url": "https://omlx.ai/benchmarks/performance/ck5uhg3x",
   "tg_tokens": null,
   "tps": 105.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ck5uhg3x-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.2 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6296.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 6,296 126.2 102.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ck5uhg3x",
   "source_url": "https://omlx.ai/benchmarks/performance/ck5uhg3x",
   "tg_tokens": null,
   "tps": 126.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ck5uhg3x-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.2 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6265.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 6,265 147.3 103.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ck5uhg3x",
   "source_url": "https://omlx.ai/benchmarks/performance/ck5uhg3x",
   "tg_tokens": null,
   "tps": 147.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ck5uhg3x-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.4 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6127.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 6,127 135.0 104.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ck5uhg3x",
   "source_url": "https://omlx.ai/benchmarks/performance/ck5uhg3x",
   "tg_tokens": null,
   "tps": 135.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ck5uhg3x-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.0 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5922.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,922 107.9 109.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ck5uhg3x",
   "source_url": "https://omlx.ai/benchmarks/performance/ck5uhg3x",
   "tg_tokens": null,
   "tps": 107.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ck5uhg3x-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5728.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 5,728 89.0 110.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ck5uhg3x",
   "source_url": "https://omlx.ai/benchmarks/performance/ck5uhg3x",
   "tg_tokens": null,
   "tps": 89.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-1o086lbv-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 183.5 GB; TTFT 14432 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4541.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 64k tokens 4,541 PP tok/s 30.2 TG tok/s 14432 TTFT (ms) 183.5 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1o086lbv",
   "source_url": "https://omlx.ai/benchmarks/performance/1o086lbv",
   "tg_tokens": null,
   "tps": 30.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-211rc1rr-ctx199680",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 195k context; peak memory 38.8 GB; TTFT 481038 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 415.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ6e-MTP on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 195k tokens 415.8 PP tok/s 40.2 TG tok/s 481038 TTFT (ms) 38.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/211rc1rr",
   "source_url": "https://omlx.ai/benchmarks/performance/211rc1rr",
   "tg_tokens": null,
   "tps": 40.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ar8cmn8j-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 32.8 GB at this context; duplicate traces of the same session: o638jlnm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1693.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ6e-MTP on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 1,693 112.1 32.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ar8cmn8j",
   "source_url": "https://omlx.ai/benchmarks/performance/ar8cmn8j",
   "tg_tokens": null,
   "tps": 112.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ar8cmn8j-ctx16384",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 33.3 GB at this context; duplicate traces of the same session: o638jlnm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1444.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ6e-MTP on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 1,444 81.2 33.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ar8cmn8j",
   "source_url": "https://omlx.ai/benchmarks/performance/ar8cmn8j",
   "tg_tokens": null,
   "tps": 81.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-dvgr3rej-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 8.1 GB at this context; duplicate traces of the same session: dqilbzzw; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1218.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-E4B-it-MLX-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 24 GB \u00b7 8bit; 1k 1,218 42.9 8.1 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dvgr3rej",
   "source_url": "https://omlx.ai/benchmarks/performance/dvgr3rej",
   "tg_tokens": null,
   "tps": 42.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-dvgr3rej-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 8.1 GB at this context; duplicate traces of the same session: dqilbzzw; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1287.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-E4B-it-MLX-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 24 GB \u00b7 8bit; 4k 1,287 42.0 8.1 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dvgr3rej",
   "source_url": "https://omlx.ai/benchmarks/performance/dvgr3rej",
   "tg_tokens": null,
   "tps": 42.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-mmi22bbk-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: u389iqvk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 504.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 504.4 17.7 17.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mmi22bbk",
   "source_url": "https://omlx.ai/benchmarks/performance/mmi22bbk",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-mmi22bbk-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: u389iqvk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 490.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 490.5 16.8 17.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mmi22bbk",
   "source_url": "https://omlx.ai/benchmarks/performance/mmi22bbk",
   "tg_tokens": null,
   "tps": 16.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4z4ea7lj-ctx1024",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 179.8 GB at this context; duplicate traces of the same session: 5skh11gd, e0f2be0q; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 525.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 525.1 47.6 179.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4z4ea7lj",
   "source_url": "https://omlx.ai/benchmarks/performance/4z4ea7lj",
   "tg_tokens": null,
   "tps": 47.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-4z4ea7lj-ctx4096",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 181.2 GB at this context; duplicate traces of the same session: 5skh11gd, e0f2be0q; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 605.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 605.0 44.3 181.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4z4ea7lj",
   "source_url": "https://omlx.ai/benchmarks/performance/4z4ea7lj",
   "tg_tokens": null,
   "tps": 44.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4z4ea7lj-ctx8192",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 181.5 GB at this context; duplicate traces of the same session: 5skh11gd, e0f2be0q; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 560.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 560.5 60.9 181.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4z4ea7lj",
   "source_url": "https://omlx.ai/benchmarks/performance/4z4ea7lj",
   "tg_tokens": null,
   "tps": 60.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-2ailajyd-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: hp9gm742; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1503.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,503 83.2 177.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ailajyd",
   "source_url": "https://omlx.ai/benchmarks/performance/2ailajyd",
   "tg_tokens": null,
   "tps": 83.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-2ailajyd-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: hp9gm742; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2324.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 2,324 67.8 178.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ailajyd",
   "source_url": "https://omlx.ai/benchmarks/performance/2ailajyd",
   "tg_tokens": null,
   "tps": 67.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-2ailajyd-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.9 GB at this context; duplicate traces of the same session: hp9gm742; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2323.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 2,323 69.8 178.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ailajyd",
   "source_url": "https://omlx.ai/benchmarks/performance/2ailajyd",
   "tg_tokens": null,
   "tps": 69.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-2ailajyd-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.1 GB at this context; duplicate traces of the same session: hp9gm742; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2320.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 2,320 150.1 179.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ailajyd",
   "source_url": "https://omlx.ai/benchmarks/performance/2ailajyd",
   "tg_tokens": null,
   "tps": 150.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-2ailajyd-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.4 GB at this context; duplicate traces of the same session: hp9gm742; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2308.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,308 101.6 179.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ailajyd",
   "source_url": "https://omlx.ai/benchmarks/performance/2ailajyd",
   "tg_tokens": null,
   "tps": 101.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-2ailajyd-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.1 GB at this context; duplicate traces of the same session: hp9gm742; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2273.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,273 83.0 180.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ailajyd",
   "source_url": "https://omlx.ai/benchmarks/performance/2ailajyd",
   "tg_tokens": null,
   "tps": 83.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-2ailajyd-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.3 GB at this context; duplicate traces of the same session: hp9gm742; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2226.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,226 146.5 181.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ailajyd",
   "source_url": "https://omlx.ai/benchmarks/performance/2ailajyd",
   "tg_tokens": null,
   "tps": 146.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-2j0a6rkc-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: 0dkfyt8s, enqf4i12, hxo8vojb, n7g9shr5, rqx82wgb, vszck7kt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1488.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,488 80.1 177.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2j0a6rkc",
   "source_url": "https://omlx.ai/benchmarks/performance/2j0a6rkc",
   "tg_tokens": null,
   "tps": 80.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-2j0a6rkc-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: 0dkfyt8s, enqf4i12, hxo8vojb, n7g9shr5, rqx82wgb, vszck7kt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2318.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 2,318 79.9 178.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2j0a6rkc",
   "source_url": "https://omlx.ai/benchmarks/performance/2j0a6rkc",
   "tg_tokens": null,
   "tps": 79.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-2j0a6rkc-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.9 GB at this context; duplicate traces of the same session: 0dkfyt8s, enqf4i12, hxo8vojb, n7g9shr5, rqx82wgb, vszck7kt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2324.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 2,324 87.6 178.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2j0a6rkc",
   "source_url": "https://omlx.ai/benchmarks/performance/2j0a6rkc",
   "tg_tokens": null,
   "tps": 87.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-2j0a6rkc-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.0 GB at this context; duplicate traces of the same session: 0dkfyt8s, enqf4i12, hxo8vojb, n7g9shr5, rqx82wgb, vszck7kt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2320.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 2,320 156.2 179.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2j0a6rkc",
   "source_url": "https://omlx.ai/benchmarks/performance/2j0a6rkc",
   "tg_tokens": null,
   "tps": 156.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-2j0a6rkc-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.3 GB at this context; duplicate traces of the same session: 0dkfyt8s, enqf4i12, hxo8vojb, n7g9shr5, rqx82wgb, vszck7kt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2308.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,308 115.1 179.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2j0a6rkc",
   "source_url": "https://omlx.ai/benchmarks/performance/2j0a6rkc",
   "tg_tokens": null,
   "tps": 115.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-2j0a6rkc-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.7 GB at this context; duplicate traces of the same session: 0dkfyt8s, enqf4i12, hxo8vojb, n7g9shr5, rqx82wgb, vszck7kt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2271.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,271 83.6 179.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2j0a6rkc",
   "source_url": "https://omlx.ai/benchmarks/performance/2j0a6rkc",
   "tg_tokens": null,
   "tps": 83.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-2j0a6rkc-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.6 GB at this context; duplicate traces of the same session: 0dkfyt8s, enqf4i12, hxo8vojb, n7g9shr5, rqx82wgb, vszck7kt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2230.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,230 145.3 180.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2j0a6rkc",
   "source_url": "https://omlx.ai/benchmarks/performance/2j0a6rkc",
   "tg_tokens": null,
   "tps": 145.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-zk7wgj1o-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: 0emj9d83, co1d3i2a, eo2w98sc, gjpkf22x, ltf5djrl, t9r9iv85; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1548.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,548 81.7 177.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zk7wgj1o",
   "source_url": "https://omlx.ai/benchmarks/performance/zk7wgj1o",
   "tg_tokens": null,
   "tps": 81.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-zk7wgj1o-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: 0emj9d83, co1d3i2a, eo2w98sc, gjpkf22x, ltf5djrl, t9r9iv85; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2323.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 2,323 70.3 178.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zk7wgj1o",
   "source_url": "https://omlx.ai/benchmarks/performance/zk7wgj1o",
   "tg_tokens": null,
   "tps": 70.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-zk7wgj1o-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.9 GB at this context; duplicate traces of the same session: 0emj9d83, co1d3i2a, eo2w98sc, gjpkf22x, ltf5djrl, t9r9iv85; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2320.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 2,320 86.6 178.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zk7wgj1o",
   "source_url": "https://omlx.ai/benchmarks/performance/zk7wgj1o",
   "tg_tokens": null,
   "tps": 86.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-zk7wgj1o-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.0 GB at this context; duplicate traces of the same session: 0emj9d83, co1d3i2a, eo2w98sc, gjpkf22x, ltf5djrl, t9r9iv85; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2307.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 2,307 154.4 179.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zk7wgj1o",
   "source_url": "https://omlx.ai/benchmarks/performance/zk7wgj1o",
   "tg_tokens": null,
   "tps": 154.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-zk7wgj1o-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.3 GB at this context; duplicate traces of the same session: 0emj9d83, co1d3i2a, eo2w98sc, gjpkf22x, ltf5djrl, t9r9iv85; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2306.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,306 100.0 179.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zk7wgj1o",
   "source_url": "https://omlx.ai/benchmarks/performance/zk7wgj1o",
   "tg_tokens": null,
   "tps": 100.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-zk7wgj1o-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.7 GB at this context; duplicate traces of the same session: 0emj9d83, co1d3i2a, eo2w98sc, gjpkf22x, ltf5djrl, t9r9iv85; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2273.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,273 84.1 179.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zk7wgj1o",
   "source_url": "https://omlx.ai/benchmarks/performance/zk7wgj1o",
   "tg_tokens": null,
   "tps": 84.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-zk7wgj1o-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.6 GB at this context; duplicate traces of the same session: 0emj9d83, co1d3i2a, eo2w98sc, gjpkf22x, ltf5djrl, t9r9iv85; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2231.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,231 144.5 180.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zk7wgj1o",
   "source_url": "https://omlx.ai/benchmarks/performance/zk7wgj1o",
   "tg_tokens": null,
   "tps": 144.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-0ctzumus-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 5.0 GB; TTFT 10321 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 396.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-4B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k tokens 396.9 PP tok/s 39.0 TG tok/s 10321 TTFT (ms) 5.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ctzumus",
   "source_url": "https://omlx.ai/benchmarks/performance/0ctzumus",
   "tg_tokens": null,
   "tps": 39.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-tp0f5igb-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 6.4 GB; TTFT 10781 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 379.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k tokens 379.9 PP tok/s 47.3 TG tok/s 10781 TTFT (ms) 6.4 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tp0f5igb",
   "source_url": "https://omlx.ai/benchmarks/performance/tp0f5igb",
   "tg_tokens": null,
   "tps": 47.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-l9joc9iq-ctx1024",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: 313napkt; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1240.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 1k 1,240 40.1 12.7 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l9joc9iq",
   "source_url": "https://omlx.ai/benchmarks/performance/l9joc9iq",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-l9joc9iq-ctx4096",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.2 GB at this context; duplicate traces of the same session: 313napkt; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1414.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 4k 1,414 38.7 13.2 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l9joc9iq",
   "source_url": "https://omlx.ai/benchmarks/performance/l9joc9iq",
   "tg_tokens": null,
   "tps": 38.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-l9joc9iq-ctx8192",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.2 GB at this context; duplicate traces of the same session: 313napkt; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1400.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 8k 1,400 38.4 13.2 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l9joc9iq",
   "source_url": "https://omlx.ai/benchmarks/performance/l9joc9iq",
   "tg_tokens": null,
   "tps": 38.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-49g2a75t-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: geedntk0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 751.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 751.3 36.3 19.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/49g2a75t",
   "source_url": "https://omlx.ai/benchmarks/performance/49g2a75t",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-49g2a75t-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: geedntk0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 988.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 988.0 51.3 20.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/49g2a75t",
   "source_url": "https://omlx.ai/benchmarks/performance/49g2a75t",
   "tg_tokens": null,
   "tps": 51.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-u3egncat-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 43.0 GB at this context; duplicate traces of the same session: xseithd2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 287.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ8e-1M on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 287.0 24.0 43.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u3egncat",
   "source_url": "https://omlx.ai/benchmarks/performance/u3egncat",
   "tg_tokens": null,
   "tps": 24.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-u3egncat-ctx16384",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 44.4 GB at this context; duplicate traces of the same session: xseithd2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 265.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ8e-1M on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 265.5 27.7 44.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u3egncat",
   "source_url": "https://omlx.ai/benchmarks/performance/u3egncat",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-oxkmycfc-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: pd5wvgef; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 258.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 258.2 28.3 18.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oxkmycfc",
   "source_url": "https://omlx.ai/benchmarks/performance/oxkmycfc",
   "tg_tokens": null,
   "tps": 28.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-oxkmycfc-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: pd5wvgef; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 256.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 256.4 25.8 21.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oxkmycfc",
   "source_url": "https://omlx.ai/benchmarks/performance/oxkmycfc",
   "tg_tokens": null,
   "tps": 25.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-k1k1n5bo-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: aknlmi99; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 250.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 250.7 28.9 18.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k1k1n5bo",
   "source_url": "https://omlx.ai/benchmarks/performance/k1k1n5bo",
   "tg_tokens": null,
   "tps": 28.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-k1k1n5bo-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: aknlmi99; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 264.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 264.5 49.8 21.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k1k1n5bo",
   "source_url": "https://omlx.ai/benchmarks/performance/k1k1n5bo",
   "tg_tokens": null,
   "tps": 49.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ipcin5sq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: frbjwx1f, t5h9t7n2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 565.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 565.3 33.6 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ipcin5sq",
   "source_url": "https://omlx.ai/benchmarks/performance/ipcin5sq",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ipcin5sq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: frbjwx1f, t5h9t7n2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 674.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 674.0 35.1 19.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ipcin5sq",
   "source_url": "https://omlx.ai/benchmarks/performance/ipcin5sq",
   "tg_tokens": null,
   "tps": 35.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ipcin5sq-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: frbjwx1f, t5h9t7n2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 697.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 697.3 41.3 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ipcin5sq",
   "source_url": "https://omlx.ai/benchmarks/performance/ipcin5sq",
   "tg_tokens": null,
   "tps": 41.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-0o3zr5ik-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 227.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 6bit; 8k 227.5 25.5 27.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0o3zr5ik",
   "source_url": "https://omlx.ai/benchmarks/performance/0o3zr5ik",
   "tg_tokens": null,
   "tps": 25.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-0o3zr5ik-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 219.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k 219.1 27.7 28.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0o3zr5ik",
   "source_url": "https://omlx.ai/benchmarks/performance/0o3zr5ik",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-uva7afpd-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.8 GB; TTFT 1939 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 528.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 528.2 PP tok/s 59.0 TG tok/s 1939 TTFT (ms) 70.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uva7afpd",
   "source_url": "https://omlx.ai/benchmarks/performance/uva7afpd",
   "tg_tokens": null,
   "tps": 59.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-d846q1nl-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: 2hn3yzjg, 3osbo7fo, bge7fwv0, g0ixr394, gpxvvzzy; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2961.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 2,961 119.7 100.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d846q1nl",
   "source_url": "https://omlx.ai/benchmarks/performance/d846q1nl",
   "tg_tokens": null,
   "tps": 119.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-d846q1nl-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.2 GB at this context; duplicate traces of the same session: 2hn3yzjg, 3osbo7fo, bge7fwv0, g0ixr394, gpxvvzzy; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6317.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 6,317 144.9 102.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d846q1nl",
   "source_url": "https://omlx.ai/benchmarks/performance/d846q1nl",
   "tg_tokens": null,
   "tps": 144.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-d846q1nl-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.2 GB at this context; duplicate traces of the same session: 2hn3yzjg, 3osbo7fo, bge7fwv0, g0ixr394, gpxvvzzy; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6271.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 6,271 119.2 103.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d846q1nl",
   "source_url": "https://omlx.ai/benchmarks/performance/d846q1nl",
   "tg_tokens": null,
   "tps": 119.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-d846q1nl-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.4 GB at this context; duplicate traces of the same session: 2hn3yzjg, 3osbo7fo, bge7fwv0, g0ixr394, gpxvvzzy; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6111.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 6,111 130.8 104.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d846q1nl",
   "source_url": "https://omlx.ai/benchmarks/performance/d846q1nl",
   "tg_tokens": null,
   "tps": 130.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-d846q1nl-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 108.7 GB at this context; duplicate traces of the same session: 2hn3yzjg, 3osbo7fo, bge7fwv0, g0ixr394, gpxvvzzy; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5900.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,900 109.7 108.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d846q1nl",
   "source_url": "https://omlx.ai/benchmarks/performance/d846q1nl",
   "tg_tokens": null,
   "tps": 109.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-d846q1nl-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.7 GB at this context; duplicate traces of the same session: 2hn3yzjg, 3osbo7fo, bge7fwv0, g0ixr394, gpxvvzzy; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5730.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 5,730 97.0 110.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d846q1nl",
   "source_url": "https://omlx.ai/benchmarks/performance/d846q1nl",
   "tg_tokens": null,
   "tps": 97.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-sudt29h3-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace, single run at 64k context; peak memory 143.1 GB; TTFT 15561 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4212.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 64k tokens 4,212 PP tok/s 74.5 TG tok/s 15561 TTFT (ms) 143.1 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sudt29h3",
   "source_url": "https://omlx.ai/benchmarks/performance/sudt29h3",
   "tg_tokens": null,
   "tps": 74.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-c5lah4s0-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace, single run at 64k context; peak memory 147.2 GB; TTFT 15867 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4131.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 64k tokens 4,131 PP tok/s 113.7 TG tok/s 15867 TTFT (ms) 147.2 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c5lah4s0",
   "source_url": "https://omlx.ai/benchmarks/performance/c5lah4s0",
   "tg_tokens": null,
   "tps": 113.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-5vfoizai-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace, single run at 64k context; peak memory 147.2 GB; TTFT 16059 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4081.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 64k tokens 4,081 PP tok/s 94.8 TG tok/s 16059 TTFT (ms) 147.2 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5vfoizai",
   "source_url": "https://omlx.ai/benchmarks/performance/5vfoizai",
   "tg_tokens": null,
   "tps": 94.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-5y4j1mrd-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace, single run at 64k context; peak memory 147.2 GB; TTFT 16082 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4075.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 64k tokens 4,075 PP tok/s 111.6 TG tok/s 16082 TTFT (ms) 147.2 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5y4j1mrd",
   "source_url": "https://omlx.ai/benchmarks/performance/5y4j1mrd",
   "tg_tokens": null,
   "tps": 111.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-jsq3zu4j-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.6 GB at this context; duplicate traces of the same session: 8kpbtu2e, gz6m5ph5, lbf3odtv, qudbqu6h, tk8xdsko, zp2j9rlr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2471.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 2,471 133.5 142.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jsq3zu4j",
   "source_url": "https://omlx.ai/benchmarks/performance/jsq3zu4j",
   "tg_tokens": null,
   "tps": 133.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-jsq3zu4j-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: 8kpbtu2e, gz6m5ph5, lbf3odtv, qudbqu6h, tk8xdsko, zp2j9rlr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4494.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 4,494 136.2 143.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jsq3zu4j",
   "source_url": "https://omlx.ai/benchmarks/performance/jsq3zu4j",
   "tg_tokens": null,
   "tps": 136.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-jsq3zu4j-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 144.5 GB at this context; duplicate traces of the same session: 8kpbtu2e, gz6m5ph5, lbf3odtv, qudbqu6h, tk8xdsko, zp2j9rlr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4946.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 8k 4,946 134.3 144.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jsq3zu4j",
   "source_url": "https://omlx.ai/benchmarks/performance/jsq3zu4j",
   "tg_tokens": null,
   "tps": 134.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-jsq3zu4j-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.9 GB at this context; duplicate traces of the same session: 8kpbtu2e, gz6m5ph5, lbf3odtv, qudbqu6h, tk8xdsko, zp2j9rlr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4874.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 32k 4,874 126.1 145.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jsq3zu4j",
   "source_url": "https://omlx.ai/benchmarks/performance/jsq3zu4j",
   "tg_tokens": null,
   "tps": 126.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-jsq3zu4j-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 147.0 GB at this context; duplicate traces of the same session: 8kpbtu2e, gz6m5ph5, lbf3odtv, qudbqu6h, tk8xdsko, zp2j9rlr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4835.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 64k 4,835 131.1 147.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jsq3zu4j",
   "source_url": "https://omlx.ai/benchmarks/performance/jsq3zu4j",
   "tg_tokens": null,
   "tps": 131.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-jsq3zu4j-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 149.3 GB at this context; duplicate traces of the same session: 8kpbtu2e, gz6m5ph5, lbf3odtv, qudbqu6h, tk8xdsko, zp2j9rlr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4715.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 128k 4,715 103.8 149.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jsq3zu4j",
   "source_url": "https://omlx.ai/benchmarks/performance/jsq3zu4j",
   "tg_tokens": null,
   "tps": 103.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-jsq3zu4j-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 152.1 GB at this context; duplicate traces of the same session: 8kpbtu2e, gz6m5ph5, lbf3odtv, qudbqu6h, tk8xdsko, zp2j9rlr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4586.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 195k 4,586 87.1 152.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jsq3zu4j",
   "source_url": "https://omlx.ai/benchmarks/performance/jsq3zu4j",
   "tg_tokens": null,
   "tps": 87.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-kq52gkv9-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.6 GB at this context; duplicate traces of the same session: dcr6alpu, oesj48rd, yexebxyn, z89djuw9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2477.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 2,477 136.7 142.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kq52gkv9",
   "source_url": "https://omlx.ai/benchmarks/performance/kq52gkv9",
   "tg_tokens": null,
   "tps": 136.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-kq52gkv9-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: dcr6alpu, oesj48rd, yexebxyn, z89djuw9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4490.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 4,490 131.5 143.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kq52gkv9",
   "source_url": "https://omlx.ai/benchmarks/performance/kq52gkv9",
   "tg_tokens": null,
   "tps": 131.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-kq52gkv9-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 146.8 GB at this context; duplicate traces of the same session: dcr6alpu, oesj48rd, yexebxyn, z89djuw9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4856.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 64k 4,856 127.2 146.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kq52gkv9",
   "source_url": "https://omlx.ai/benchmarks/performance/kq52gkv9",
   "tg_tokens": null,
   "tps": 127.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-kq52gkv9-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 150.6 GB at this context; duplicate traces of the same session: dcr6alpu, oesj48rd, yexebxyn, z89djuw9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4709.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 128k 4,709 103.1 150.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kq52gkv9",
   "source_url": "https://omlx.ai/benchmarks/performance/kq52gkv9",
   "tg_tokens": null,
   "tps": 103.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-kq52gkv9-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 153.3 GB at this context; duplicate traces of the same session: dcr6alpu, oesj48rd, yexebxyn, z89djuw9; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4582.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 195k 4,582 87.0 153.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kq52gkv9",
   "source_url": "https://omlx.ai/benchmarks/performance/kq52gkv9",
   "tg_tokens": null,
   "tps": 87.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-xsmz7kxu-ctx1024",
   "model": "Qwen3-30B-A3B-Instruct-2507",
   "notes": "oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: ln5u9bzl; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 313.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 4bit; 1k 313.4 35.5 16.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xsmz7kxu",
   "source_url": "https://omlx.ai/benchmarks/performance/xsmz7kxu",
   "tg_tokens": null,
   "tps": 35.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-xsmz7kxu-ctx4096",
   "model": "Qwen3-30B-A3B-Instruct-2507",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: ln5u9bzl; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 296.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 4bit; 4k 296.0 32.9 17.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xsmz7kxu",
   "source_url": "https://omlx.ai/benchmarks/performance/xsmz7kxu",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ukajburk-ctx1024",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: bewtwdp4, noqtnzud, vgk0mzn8; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1342.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-Uncensored-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,342 68.5 178.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ukajburk",
   "source_url": "https://omlx.ai/benchmarks/performance/ukajburk",
   "tg_tokens": null,
   "tps": 68.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ukajburk-ctx32768",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.1 GB at this context; duplicate traces of the same session: bewtwdp4, noqtnzud, vgk0mzn8; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1924.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-Uncensored-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,924 58.0 180.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ukajburk",
   "source_url": "https://omlx.ai/benchmarks/performance/ukajburk",
   "tg_tokens": null,
   "tps": 58.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ukajburk-ctx65536",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.5 GB at this context; duplicate traces of the same session: bewtwdp4, noqtnzud, vgk0mzn8; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1892.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-Uncensored-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,892 55.2 180.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ukajburk",
   "source_url": "https://omlx.ai/benchmarks/performance/ukajburk",
   "tg_tokens": null,
   "tps": 55.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ukajburk-ctx131072",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.2 GB at this context; duplicate traces of the same session: bewtwdp4, noqtnzud, vgk0mzn8; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1854.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-Uncensored-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 1,854 65.2 181.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ukajburk",
   "source_url": "https://omlx.ai/benchmarks/performance/ukajburk",
   "tg_tokens": null,
   "tps": 65.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-yl2egguu-ctx1024",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: hkt2bl89, hmgn3edi, p8vqfhlr; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1338.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-Uncensored-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,338 83.8 177.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yl2egguu",
   "source_url": "https://omlx.ai/benchmarks/performance/yl2egguu",
   "tg_tokens": null,
   "tps": 83.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-yl2egguu-ctx32768",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.2 GB at this context; duplicate traces of the same session: hkt2bl89, hmgn3edi, p8vqfhlr; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1896.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-Uncensored-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,896 71.0 179.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yl2egguu",
   "source_url": "https://omlx.ai/benchmarks/performance/yl2egguu",
   "tg_tokens": null,
   "tps": 71.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-yl2egguu-ctx65536",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.7 GB at this context; duplicate traces of the same session: hkt2bl89, hmgn3edi, p8vqfhlr; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1862.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-Uncensored-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,862 61.0 179.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yl2egguu",
   "source_url": "https://omlx.ai/benchmarks/performance/yl2egguu",
   "tg_tokens": null,
   "tps": 61.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-yl2egguu-ctx131072",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.6 GB at this context; duplicate traces of the same session: hkt2bl89, hmgn3edi, p8vqfhlr; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1829.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-Uncensored-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 1,829 54.0 180.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yl2egguu",
   "source_url": "https://omlx.ai/benchmarks/performance/yl2egguu",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-cpmv7sqt-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: 1vox3xhi, 5f45dbns, 7g06ebce, foi8cy7f, lx0cibit, wzwybb1i; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1504.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,504 81.5 177.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cpmv7sqt",
   "source_url": "https://omlx.ai/benchmarks/performance/cpmv7sqt",
   "tg_tokens": null,
   "tps": 81.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-cpmv7sqt-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: 1vox3xhi, 5f45dbns, 7g06ebce, foi8cy7f, lx0cibit, wzwybb1i; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2318.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 2,318 71.0 178.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cpmv7sqt",
   "source_url": "https://omlx.ai/benchmarks/performance/cpmv7sqt",
   "tg_tokens": null,
   "tps": 71.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-cpmv7sqt-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.9 GB at this context; duplicate traces of the same session: 1vox3xhi, 5f45dbns, 7g06ebce, foi8cy7f, lx0cibit, wzwybb1i; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2319.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 2,319 78.8 178.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cpmv7sqt",
   "source_url": "https://omlx.ai/benchmarks/performance/cpmv7sqt",
   "tg_tokens": null,
   "tps": 78.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-cpmv7sqt-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.1 GB at this context; duplicate traces of the same session: 1vox3xhi, 5f45dbns, 7g06ebce, foi8cy7f, lx0cibit, wzwybb1i; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2323.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 2,323 150.1 179.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cpmv7sqt",
   "source_url": "https://omlx.ai/benchmarks/performance/cpmv7sqt",
   "tg_tokens": null,
   "tps": 150.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-cpmv7sqt-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.4 GB at this context; duplicate traces of the same session: 1vox3xhi, 5f45dbns, 7g06ebce, foi8cy7f, lx0cibit, wzwybb1i; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2305.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,305 103.1 179.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cpmv7sqt",
   "source_url": "https://omlx.ai/benchmarks/performance/cpmv7sqt",
   "tg_tokens": null,
   "tps": 103.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-cpmv7sqt-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.1 GB at this context; duplicate traces of the same session: 1vox3xhi, 5f45dbns, 7g06ebce, foi8cy7f, lx0cibit, wzwybb1i; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2271.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,271 77.7 180.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cpmv7sqt",
   "source_url": "https://omlx.ai/benchmarks/performance/cpmv7sqt",
   "tg_tokens": null,
   "tps": 77.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-cpmv7sqt-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.3 GB at this context; duplicate traces of the same session: 1vox3xhi, 5f45dbns, 7g06ebce, foi8cy7f, lx0cibit, wzwybb1i; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2228.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,228 146.7 181.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cpmv7sqt",
   "source_url": "https://omlx.ai/benchmarks/performance/cpmv7sqt",
   "tg_tokens": null,
   "tps": 146.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-oc3o1d83-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1478.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,478 74.8 177.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oc3o1d83",
   "source_url": "https://omlx.ai/benchmarks/performance/oc3o1d83",
   "tg_tokens": null,
   "tps": 74.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-oc3o1d83-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2326.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 2,326 75.5 178.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oc3o1d83",
   "source_url": "https://omlx.ai/benchmarks/performance/oc3o1d83",
   "tg_tokens": null,
   "tps": 75.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-oc3o1d83-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.9 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2322.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 2,322 100.8 178.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oc3o1d83",
   "source_url": "https://omlx.ai/benchmarks/performance/oc3o1d83",
   "tg_tokens": null,
   "tps": 100.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-oc3o1d83-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.0 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2319.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 2,319 155.1 179.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oc3o1d83",
   "source_url": "https://omlx.ai/benchmarks/performance/oc3o1d83",
   "tg_tokens": null,
   "tps": 155.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-oc3o1d83-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.3 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2308.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,308 109.9 179.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oc3o1d83",
   "source_url": "https://omlx.ai/benchmarks/performance/oc3o1d83",
   "tg_tokens": null,
   "tps": 109.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-oc3o1d83-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2274.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,274 88.2 179.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oc3o1d83",
   "source_url": "https://omlx.ai/benchmarks/performance/oc3o1d83",
   "tg_tokens": null,
   "tps": 88.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-oc3o1d83-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.6 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2231.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,231 145.1 180.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oc3o1d83",
   "source_url": "https://omlx.ai/benchmarks/performance/oc3o1d83",
   "tg_tokens": null,
   "tps": 145.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-u33p9j27-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: cdpj785u, dl343ye6, ou2gpci7, rnq099kw, w2wqlqcs, x1mb3il4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1536.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,536 79.0 177.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u33p9j27",
   "source_url": "https://omlx.ai/benchmarks/performance/u33p9j27",
   "tg_tokens": null,
   "tps": 79.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-u33p9j27-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: cdpj785u, dl343ye6, ou2gpci7, rnq099kw, w2wqlqcs, x1mb3il4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2321.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 2,321 66.0 178.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u33p9j27",
   "source_url": "https://omlx.ai/benchmarks/performance/u33p9j27",
   "tg_tokens": null,
   "tps": 66.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-u33p9j27-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.8 GB at this context; duplicate traces of the same session: cdpj785u, dl343ye6, ou2gpci7, rnq099kw, w2wqlqcs, x1mb3il4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2321.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 2,321 65.2 178.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u33p9j27",
   "source_url": "https://omlx.ai/benchmarks/performance/u33p9j27",
   "tg_tokens": null,
   "tps": 65.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-u33p9j27-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.1 GB at this context; duplicate traces of the same session: cdpj785u, dl343ye6, ou2gpci7, rnq099kw, w2wqlqcs, x1mb3il4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2313.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 2,313 147.6 179.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u33p9j27",
   "source_url": "https://omlx.ai/benchmarks/performance/u33p9j27",
   "tg_tokens": null,
   "tps": 147.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-u33p9j27-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.4 GB at this context; duplicate traces of the same session: cdpj785u, dl343ye6, ou2gpci7, rnq099kw, w2wqlqcs, x1mb3il4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2304.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,304 102.9 179.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u33p9j27",
   "source_url": "https://omlx.ai/benchmarks/performance/u33p9j27",
   "tg_tokens": null,
   "tps": 102.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-u33p9j27-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.1 GB at this context; duplicate traces of the same session: cdpj785u, dl343ye6, ou2gpci7, rnq099kw, w2wqlqcs, x1mb3il4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2274.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,274 85.4 180.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u33p9j27",
   "source_url": "https://omlx.ai/benchmarks/performance/u33p9j27",
   "tg_tokens": null,
   "tps": 85.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-u33p9j27-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.3 GB at this context; duplicate traces of the same session: cdpj785u, dl343ye6, ou2gpci7, rnq099kw, w2wqlqcs, x1mb3il4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2232.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,232 145.6 181.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u33p9j27",
   "source_url": "https://omlx.ai/benchmarks/performance/u33p9j27",
   "tg_tokens": null,
   "tps": 145.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-gr7dqui0-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: 3hugeydo, 7bd0x5mb, llxjfkjs, ol0p61gb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1545.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,545 78.0 177.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gr7dqui0",
   "source_url": "https://omlx.ai/benchmarks/performance/gr7dqui0",
   "tg_tokens": null,
   "tps": 78.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-gr7dqui0-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: 3hugeydo, 7bd0x5mb, llxjfkjs, ol0p61gb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2328.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 2,328 68.7 178.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gr7dqui0",
   "source_url": "https://omlx.ai/benchmarks/performance/gr7dqui0",
   "tg_tokens": null,
   "tps": 68.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-gr7dqui0-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.0 GB at this context; duplicate traces of the same session: 3hugeydo, 7bd0x5mb, llxjfkjs, ol0p61gb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2323.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 2,323 154.5 179.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gr7dqui0",
   "source_url": "https://omlx.ai/benchmarks/performance/gr7dqui0",
   "tg_tokens": null,
   "tps": 154.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-gr7dqui0-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.3 GB at this context; duplicate traces of the same session: 3hugeydo, 7bd0x5mb, llxjfkjs, ol0p61gb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2308.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,308 109.4 179.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gr7dqui0",
   "source_url": "https://omlx.ai/benchmarks/performance/gr7dqui0",
   "tg_tokens": null,
   "tps": 109.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-gr7dqui0-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.7 GB at this context; duplicate traces of the same session: 3hugeydo, 7bd0x5mb, llxjfkjs, ol0p61gb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2275.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,275 91.0 179.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gr7dqui0",
   "source_url": "https://omlx.ai/benchmarks/performance/gr7dqui0",
   "tg_tokens": null,
   "tps": 91.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-gr7dqui0-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.6 GB at this context; duplicate traces of the same session: 3hugeydo, 7bd0x5mb, llxjfkjs, ol0p61gb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2230.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,230 144.8 180.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gr7dqui0",
   "source_url": "https://omlx.ai/benchmarks/performance/gr7dqui0",
   "tg_tokens": null,
   "tps": 144.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-0jwp5kh6-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.6 GB at this context; duplicate traces of the same session: lr0kjom5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3905.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LiquidAI--LFM2.5-24B-A2B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 3,905 122.4 13.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0jwp5kh6",
   "source_url": "https://omlx.ai/benchmarks/performance/0jwp5kh6",
   "tg_tokens": null,
   "tps": 122.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-0jwp5kh6-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.8 GB at this context; duplicate traces of the same session: lr0kjom5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4299.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LiquidAI--LFM2.5-24B-A2B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 4,299 109.8 13.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0jwp5kh6",
   "source_url": "https://omlx.ai/benchmarks/performance/0jwp5kh6",
   "tg_tokens": null,
   "tps": 109.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-5snl19eu-ctx1024",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: a3r8tpyh, a9sy6f5s, entxk60b, pa7p30bn, zqokxmly; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 130.2 15.4 15.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5snl19eu",
   "source_url": "https://omlx.ai/benchmarks/performance/5snl19eu",
   "tg_tokens": null,
   "tps": 15.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-5snl19eu-ctx4096",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: a3r8tpyh, a9sy6f5s, entxk60b, pa7p30bn, zqokxmly; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 128.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 128.0 15.0 16.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5snl19eu",
   "source_url": "https://omlx.ai/benchmarks/performance/5snl19eu",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-5snl19eu-ctx8192",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 16.2 GB at this context; duplicate traces of the same session: a3r8tpyh, a9sy6f5s, entxk60b, pa7p30bn, zqokxmly; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 121.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 121.0 14.6 16.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5snl19eu",
   "source_url": "https://omlx.ai/benchmarks/performance/5snl19eu",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-5snl19eu-ctx16384",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: a3r8tpyh, a9sy6f5s, entxk60b, pa7p30bn, zqokxmly; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 116.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 116.2 14.3 16.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5snl19eu",
   "source_url": "https://omlx.ai/benchmarks/performance/5snl19eu",
   "tg_tokens": null,
   "tps": 14.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-5snl19eu-ctx32768",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 16.5 GB at this context; duplicate traces of the same session: a3r8tpyh, a9sy6f5s, entxk60b, pa7p30bn, zqokxmly; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 114.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 114.3 13.7 16.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5snl19eu",
   "source_url": "https://omlx.ai/benchmarks/performance/5snl19eu",
   "tg_tokens": null,
   "tps": 13.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-5snl19eu-ctx65536",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: a3r8tpyh, a9sy6f5s, entxk60b, pa7p30bn, zqokxmly; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 105.8 13.0 16.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5snl19eu",
   "source_url": "https://omlx.ai/benchmarks/performance/5snl19eu",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-mik6fb3d-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: 6hqveshw, ialtq1zt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2341.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 2,341 121.8 37.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mik6fb3d",
   "source_url": "https://omlx.ai/benchmarks/performance/mik6fb3d",
   "tg_tokens": null,
   "tps": 121.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-mik6fb3d-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.1 GB at this context; duplicate traces of the same session: 6hqveshw, ialtq1zt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3349.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 3,349 152.2 39.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mik6fb3d",
   "source_url": "https://omlx.ai/benchmarks/performance/mik6fb3d",
   "tg_tokens": null,
   "tps": 152.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-mik6fb3d-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: 6hqveshw, ialtq1zt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3186.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 3,186 127.2 39.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mik6fb3d",
   "source_url": "https://omlx.ai/benchmarks/performance/mik6fb3d",
   "tg_tokens": null,
   "tps": 127.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-pnquog87-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: 0eh02rhg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 371.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 371.7 44.8 6.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pnquog87",
   "source_url": "https://omlx.ai/benchmarks/performance/pnquog87",
   "tg_tokens": null,
   "tps": 44.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-pnquog87-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: 0eh02rhg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 368.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 368.7 43.2 6.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pnquog87",
   "source_url": "https://omlx.ai/benchmarks/performance/pnquog87",
   "tg_tokens": null,
   "tps": 43.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-alxw2i0e-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 6.4 GB; TTFT 10892 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 376.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k tokens 376.1 PP tok/s 45.7 TG tok/s 10892 TTFT (ms) 6.4 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/alxw2i0e",
   "source_url": "https://omlx.ai/benchmarks/performance/alxw2i0e",
   "tg_tokens": null,
   "tps": 45.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-0unyb2q6-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 8k context; peak memory 35.2 GB; TTFT 5953 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1376.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 8k tokens 1,376 PP tok/s 55.8 TG tok/s 5953 TTFT (ms) 35.2 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0unyb2q6",
   "source_url": "https://omlx.ai/benchmarks/performance/0unyb2q6",
   "tg_tokens": null,
   "tps": 55.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-pssmim47-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: jbd55obu, s71fpyks; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 249.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 249.1 19.7 22.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pssmim47",
   "source_url": "https://omlx.ai/benchmarks/performance/pssmim47",
   "tg_tokens": null,
   "tps": 19.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-pssmim47-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: jbd55obu, s71fpyks; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 255.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 255.3 18.8 25.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pssmim47",
   "source_url": "https://omlx.ai/benchmarks/performance/pssmim47",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-pssmim47-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: jbd55obu, s71fpyks; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 82.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 82.9 5.2 28.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pssmim47",
   "source_url": "https://omlx.ai/benchmarks/performance/pssmim47",
   "tg_tokens": null,
   "tps": 5.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-4br03rjv-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 8w8eussn; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 291.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 291.8 24.7 17.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4br03rjv",
   "source_url": "https://omlx.ai/benchmarks/performance/4br03rjv",
   "tg_tokens": null,
   "tps": 24.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-4br03rjv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: 8w8eussn; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 362.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 362.5 21.3 18.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4br03rjv",
   "source_url": "https://omlx.ai/benchmarks/performance/4br03rjv",
   "tg_tokens": null,
   "tps": 21.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-f42qgs13-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: luylayu6; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1511.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 1,511 106.8 19.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f42qgs13",
   "source_url": "https://omlx.ai/benchmarks/performance/f42qgs13",
   "tg_tokens": null,
   "tps": 106.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-f42qgs13-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: luylayu6; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1390.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,390 113.1 20.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f42qgs13",
   "source_url": "https://omlx.ai/benchmarks/performance/f42qgs13",
   "tg_tokens": null,
   "tps": 113.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-emvyptzj-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 227.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 6bit; 8k 227.5 25.5 27.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/emvyptzj",
   "source_url": "https://omlx.ai/benchmarks/performance/emvyptzj",
   "tg_tokens": null,
   "tps": 25.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-emvyptzj-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 219.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k 219.1 27.7 28.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/emvyptzj",
   "source_url": "https://omlx.ai/benchmarks/performance/emvyptzj",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-8vh11f91-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.2 GB at this context; duplicate traces of the same session: 89r34yij, ct0tjhy5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 226.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 6bit; 8k 226.8 25.0 27.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8vh11f91",
   "source_url": "https://omlx.ai/benchmarks/performance/8vh11f91",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-8vh11f91-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: 89r34yij, ct0tjhy5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 219.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k 219.2 28.6 28.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8vh11f91",
   "source_url": "https://omlx.ai/benchmarks/performance/8vh11f91",
   "tg_tokens": null,
   "tps": 28.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-8vh11f91-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.0 GB at this context; duplicate traces of the same session: 89r34yij, ct0tjhy5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 203.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 203.2 25.2 30.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8vh11f91",
   "source_url": "https://omlx.ai/benchmarks/performance/8vh11f91",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ni34io3g-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.2 GB at this context; duplicate traces of the same session: 8k62blri, q43vowgc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 293.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 6bit; 8k 293.1 26.3 33.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ni34io3g",
   "source_url": "https://omlx.ai/benchmarks/performance/ni34io3g",
   "tg_tokens": null,
   "tps": 26.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ni34io3g-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.0 GB at this context; duplicate traces of the same session: 8k62blri, q43vowgc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 279.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k 279.4 35.6 34.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ni34io3g",
   "source_url": "https://omlx.ai/benchmarks/performance/ni34io3g",
   "tg_tokens": null,
   "tps": 35.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-ni34io3g-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.3 GB at this context; duplicate traces of the same session: 8k62blri, q43vowgc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 228.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 228.9 23.4 35.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ni34io3g",
   "source_url": "https://omlx.ai/benchmarks/performance/ni34io3g",
   "tg_tokens": null,
   "tps": 23.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-hhme9o2o-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.9 GB; TTFT 2071 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 494.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 494.5 PP tok/s 63.6 TG tok/s 2071 TTFT (ms) 70.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hhme9o2o",
   "source_url": "https://omlx.ai/benchmarks/performance/hhme9o2o",
   "tg_tokens": null,
   "tps": 63.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-pv035uk6-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace, single run at 64k context; peak memory 147.2 GB; TTFT 15955 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4108.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 64k tokens 4,108 PP tok/s 107.8 TG tok/s 15955 TTFT (ms) 147.2 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pv035uk6",
   "source_url": "https://omlx.ai/benchmarks/performance/pv035uk6",
   "tg_tokens": null,
   "tps": 107.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-dvaqohj8-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace, single run at 64k context; peak memory 147.2 GB; TTFT 16174 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4052.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 64k tokens 4,052 PP tok/s 70.8 TG tok/s 16174 TTFT (ms) 147.2 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dvaqohj8",
   "source_url": "https://omlx.ai/benchmarks/performance/dvaqohj8",
   "tg_tokens": null,
   "tps": 70.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-n9j80abk-ctx1024",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: oln079kh, s58ucar8, wbhd1154; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1331.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-Uncensored-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,331 70.1 178.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n9j80abk",
   "source_url": "https://omlx.ai/benchmarks/performance/n9j80abk",
   "tg_tokens": null,
   "tps": 70.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-n9j80abk-ctx32768",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.1 GB at this context; duplicate traces of the same session: oln079kh, s58ucar8, wbhd1154; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1939.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-Uncensored-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,939 103.0 180.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n9j80abk",
   "source_url": "https://omlx.ai/benchmarks/performance/n9j80abk",
   "tg_tokens": null,
   "tps": 103.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-n9j80abk-ctx65536",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.5 GB at this context; duplicate traces of the same session: oln079kh, s58ucar8, wbhd1154; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1909.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-Uncensored-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,909 81.9 180.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n9j80abk",
   "source_url": "https://omlx.ai/benchmarks/performance/n9j80abk",
   "tg_tokens": null,
   "tps": 81.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-n9j80abk-ctx131072",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.3 GB at this context; duplicate traces of the same session: oln079kh, s58ucar8, wbhd1154; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1873.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-Uncensored-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 1,873 107.8 181.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n9j80abk",
   "source_url": "https://omlx.ai/benchmarks/performance/n9j80abk",
   "tg_tokens": null,
   "tps": 107.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-sez7wzjr-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: 0117s90w, 4kuod9lb, djffgz87, g9n3s2ex, t114nyz5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1478.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,478 74.8 177.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sez7wzjr",
   "source_url": "https://omlx.ai/benchmarks/performance/sez7wzjr",
   "tg_tokens": null,
   "tps": 74.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-sez7wzjr-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: 0117s90w, 4kuod9lb, djffgz87, g9n3s2ex, t114nyz5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2326.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 2,326 75.5 178.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sez7wzjr",
   "source_url": "https://omlx.ai/benchmarks/performance/sez7wzjr",
   "tg_tokens": null,
   "tps": 75.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-sez7wzjr-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.9 GB at this context; duplicate traces of the same session: 0117s90w, 4kuod9lb, djffgz87, g9n3s2ex, t114nyz5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2322.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 2,322 100.8 178.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sez7wzjr",
   "source_url": "https://omlx.ai/benchmarks/performance/sez7wzjr",
   "tg_tokens": null,
   "tps": 100.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-sez7wzjr-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.0 GB at this context; duplicate traces of the same session: 0117s90w, 4kuod9lb, djffgz87, g9n3s2ex, t114nyz5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2319.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 2,319 155.1 179.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sez7wzjr",
   "source_url": "https://omlx.ai/benchmarks/performance/sez7wzjr",
   "tg_tokens": null,
   "tps": 155.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-sez7wzjr-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.3 GB at this context; duplicate traces of the same session: 0117s90w, 4kuod9lb, djffgz87, g9n3s2ex, t114nyz5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2308.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,308 109.9 179.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sez7wzjr",
   "source_url": "https://omlx.ai/benchmarks/performance/sez7wzjr",
   "tg_tokens": null,
   "tps": 109.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-sez7wzjr-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.7 GB at this context; duplicate traces of the same session: 0117s90w, 4kuod9lb, djffgz87, g9n3s2ex, t114nyz5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2274.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,274 88.2 179.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sez7wzjr",
   "source_url": "https://omlx.ai/benchmarks/performance/sez7wzjr",
   "tg_tokens": null,
   "tps": 88.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-sez7wzjr-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.6 GB at this context; duplicate traces of the same session: 0117s90w, 4kuod9lb, djffgz87, g9n3s2ex, t114nyz5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2231.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,231 145.1 180.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sez7wzjr",
   "source_url": "https://omlx.ai/benchmarks/performance/sez7wzjr",
   "tg_tokens": null,
   "tps": 145.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-eggkwb10-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: 8jk8tge3, hs5px8fn, k01vpsnu, maznx7tk, svo5fhmd, x75bkdl5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1484.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,484 75.8 177.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eggkwb10",
   "source_url": "https://omlx.ai/benchmarks/performance/eggkwb10",
   "tg_tokens": null,
   "tps": 75.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-eggkwb10-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: 8jk8tge3, hs5px8fn, k01vpsnu, maznx7tk, svo5fhmd, x75bkdl5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2318.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 2,318 76.0 178.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eggkwb10",
   "source_url": "https://omlx.ai/benchmarks/performance/eggkwb10",
   "tg_tokens": null,
   "tps": 76.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-eggkwb10-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.9 GB at this context; duplicate traces of the same session: 8jk8tge3, hs5px8fn, k01vpsnu, maznx7tk, svo5fhmd, x75bkdl5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2323.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 2,323 83.6 178.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eggkwb10",
   "source_url": "https://omlx.ai/benchmarks/performance/eggkwb10",
   "tg_tokens": null,
   "tps": 83.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-eggkwb10-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.1 GB at this context; duplicate traces of the same session: 8jk8tge3, hs5px8fn, k01vpsnu, maznx7tk, svo5fhmd, x75bkdl5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2318.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 2,318 149.2 179.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eggkwb10",
   "source_url": "https://omlx.ai/benchmarks/performance/eggkwb10",
   "tg_tokens": null,
   "tps": 149.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-eggkwb10-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.4 GB at this context; duplicate traces of the same session: 8jk8tge3, hs5px8fn, k01vpsnu, maznx7tk, svo5fhmd, x75bkdl5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2309.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,309 104.6 179.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eggkwb10",
   "source_url": "https://omlx.ai/benchmarks/performance/eggkwb10",
   "tg_tokens": null,
   "tps": 104.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-eggkwb10-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.1 GB at this context; duplicate traces of the same session: 8jk8tge3, hs5px8fn, k01vpsnu, maznx7tk, svo5fhmd, x75bkdl5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2272.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,272 86.7 180.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eggkwb10",
   "source_url": "https://omlx.ai/benchmarks/performance/eggkwb10",
   "tg_tokens": null,
   "tps": 86.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-eggkwb10-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.3 GB at this context; duplicate traces of the same session: 8jk8tge3, hs5px8fn, k01vpsnu, maznx7tk, svo5fhmd, x75bkdl5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2227.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "GLM-5.3-Flash-oQ4-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,227 147.1 181.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eggkwb10",
   "source_url": "https://omlx.ai/benchmarks/performance/eggkwb10",
   "tg_tokens": null,
   "tps": 147.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-x8kp2eec-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1545.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,545 78.0 177.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x8kp2eec",
   "source_url": "https://omlx.ai/benchmarks/performance/x8kp2eec",
   "tg_tokens": null,
   "tps": 78.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-x8kp2eec-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2328.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 2,328 68.7 178.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x8kp2eec",
   "source_url": "https://omlx.ai/benchmarks/performance/x8kp2eec",
   "tg_tokens": null,
   "tps": 68.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-x8kp2eec-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.0 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2323.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 2,323 154.5 179.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x8kp2eec",
   "source_url": "https://omlx.ai/benchmarks/performance/x8kp2eec",
   "tg_tokens": null,
   "tps": 154.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-x8kp2eec-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.3 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2308.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,308 109.4 179.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x8kp2eec",
   "source_url": "https://omlx.ai/benchmarks/performance/x8kp2eec",
   "tg_tokens": null,
   "tps": 109.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-x8kp2eec-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2275.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,275 91.0 179.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x8kp2eec",
   "source_url": "https://omlx.ai/benchmarks/performance/x8kp2eec",
   "tg_tokens": null,
   "tps": 91.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-x8kp2eec-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.6 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2230.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,230 144.8 180.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x8kp2eec",
   "source_url": "https://omlx.ai/benchmarks/performance/x8kp2eec",
   "tg_tokens": null,
   "tps": 144.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-5iwfkoxc-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: 3quzyh9t, 4mv523kv, jpckif1s, me4bf1rh, p3vltqhw, s7fvduyj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1541.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,541 77.6 177.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5iwfkoxc",
   "source_url": "https://omlx.ai/benchmarks/performance/5iwfkoxc",
   "tg_tokens": null,
   "tps": 77.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-5iwfkoxc-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: 3quzyh9t, 4mv523kv, jpckif1s, me4bf1rh, p3vltqhw, s7fvduyj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2320.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 2,320 66.0 178.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5iwfkoxc",
   "source_url": "https://omlx.ai/benchmarks/performance/5iwfkoxc",
   "tg_tokens": null,
   "tps": 66.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-5iwfkoxc-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.8 GB at this context; duplicate traces of the same session: 3quzyh9t, 4mv523kv, jpckif1s, me4bf1rh, p3vltqhw, s7fvduyj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2320.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 2,320 84.0 178.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5iwfkoxc",
   "source_url": "https://omlx.ai/benchmarks/performance/5iwfkoxc",
   "tg_tokens": null,
   "tps": 84.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-5iwfkoxc-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.0 GB at this context; duplicate traces of the same session: 3quzyh9t, 4mv523kv, jpckif1s, me4bf1rh, p3vltqhw, s7fvduyj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2317.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 2,317 148.3 179.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5iwfkoxc",
   "source_url": "https://omlx.ai/benchmarks/performance/5iwfkoxc",
   "tg_tokens": null,
   "tps": 148.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-5iwfkoxc-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.4 GB at this context; duplicate traces of the same session: 3quzyh9t, 4mv523kv, jpckif1s, me4bf1rh, p3vltqhw, s7fvduyj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2307.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 2,307 103.2 179.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5iwfkoxc",
   "source_url": "https://omlx.ai/benchmarks/performance/5iwfkoxc",
   "tg_tokens": null,
   "tps": 103.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-5iwfkoxc-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.1 GB at this context; duplicate traces of the same session: 3quzyh9t, 4mv523kv, jpckif1s, me4bf1rh, p3vltqhw, s7fvduyj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2275.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 2,275 76.0 180.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5iwfkoxc",
   "source_url": "https://omlx.ai/benchmarks/performance/5iwfkoxc",
   "tg_tokens": null,
   "tps": 76.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-5iwfkoxc-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.3 GB at this context; duplicate traces of the same session: 3quzyh9t, 4mv523kv, jpckif1s, me4bf1rh, p3vltqhw, s7fvduyj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2231.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 2,231 144.4 181.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5iwfkoxc",
   "source_url": "https://omlx.ai/benchmarks/performance/5iwfkoxc",
   "tg_tokens": null,
   "tps": 144.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ezi72tus-ctx4096",
   "model": "Laguna XS 2.1",
   "notes": "oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: wj26prce; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1535.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Laguna-XS-2.1-6bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 4k 1,535 77.8 26.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ezi72tus",
   "source_url": "https://omlx.ai/benchmarks/performance/ezi72tus",
   "tg_tokens": null,
   "tps": 77.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ezi72tus-ctx16384",
   "model": "Laguna XS 2.1",
   "notes": "oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: wj26prce; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1385.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Laguna-XS-2.1-6bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 6bit; 16k 1,385 67.7 26.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ezi72tus",
   "source_url": "https://omlx.ai/benchmarks/performance/ezi72tus",
   "tg_tokens": null,
   "tps": 67.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2e7mirmz-ctx4096",
   "model": "Muse Glimmer",
   "notes": "oMLX community benchmark trace; peak memory 15.5 GB at this context; duplicate traces of the same session: 0ktexi3h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 935.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-30B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 935.1 31.7 15.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2e7mirmz",
   "source_url": "https://omlx.ai/benchmarks/performance/2e7mirmz",
   "tg_tokens": null,
   "tps": 31.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2e7mirmz-ctx16384",
   "model": "Muse Glimmer",
   "notes": "oMLX community benchmark trace; peak memory 15.6 GB at this context; duplicate traces of the same session: 0ktexi3h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 980.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-30B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 980.4 31.0 15.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2e7mirmz",
   "source_url": "https://omlx.ai/benchmarks/performance/2e7mirmz",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-yxxb7khz-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 37.7 GB; TTFT 4892 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6698.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 32k tokens 6,698 PP tok/s 167.8 TG tok/s 4892 TTFT (ms) 37.7 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yxxb7khz",
   "source_url": "https://omlx.ai/benchmarks/performance/yxxb7khz",
   "tg_tokens": null,
   "tps": 167.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-6jolqccr-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 37.7 GB; TTFT 4867 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6733.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 32k tokens 6,733 PP tok/s 168.1 TG tok/s 4867 TTFT (ms) 37.7 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6jolqccr",
   "source_url": "https://omlx.ai/benchmarks/performance/6jolqccr",
   "tg_tokens": null,
   "tps": 168.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-18ui8yzz-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 37.7 GB; TTFT 4890 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6701.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 32k tokens 6,701 PP tok/s 164.7 TG tok/s 4890 TTFT (ms) 37.7 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/18ui8yzz",
   "source_url": "https://omlx.ai/benchmarks/performance/18ui8yzz",
   "tg_tokens": null,
   "tps": 164.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-jce0uc05-ctx16384",
   "model": "Ornith-1.5-9B",
   "notes": "fp16 MTP build; oMLX community benchmark trace, single run at 16k context; peak memory 10.8 GB; TTFT 75906 ms; duplicate traces of the same session: none; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 215.8,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Ornith-1.5-9B-oQ5e-fp16-mtp on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 5bit; 16k tokens 215.8 PP tok/s 23.5 TG tok/s 75906 TTFT (ms) 10.8 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jce0uc05",
   "source_url": "https://omlx.ai/benchmarks/performance/jce0uc05",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-k5mp0kkm-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: sc3yof5i; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1523.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 1,523 56.2 5.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k5mp0kkm",
   "source_url": "https://omlx.ai/benchmarks/performance/k5mp0kkm",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-k5mp0kkm-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: sc3yof5i; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1740.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 1,740 54.7 6.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k5mp0kkm",
   "source_url": "https://omlx.ai/benchmarks/performance/k5mp0kkm",
   "tg_tokens": null,
   "tps": 54.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ok879t1v-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: hmv5tbd2, s7jx7pjf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 590.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 590.8 41.3 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ok879t1v",
   "source_url": "https://omlx.ai/benchmarks/performance/ok879t1v",
   "tg_tokens": null,
   "tps": 41.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ok879t1v-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: hmv5tbd2, s7jx7pjf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 674.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 674.1 48.9 19.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ok879t1v",
   "source_url": "https://omlx.ai/benchmarks/performance/ok879t1v",
   "tg_tokens": null,
   "tps": 48.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ok879t1v-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: hmv5tbd2, s7jx7pjf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 697.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 697.8 24.0 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ok879t1v",
   "source_url": "https://omlx.ai/benchmarks/performance/ok879t1v",
   "tg_tokens": null,
   "tps": 24.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-g62dgt2e-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.4 GB at this context; duplicate traces of the same session: mmrq3d4y; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4926.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 4,926 141.1 102.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g62dgt2e",
   "source_url": "https://omlx.ai/benchmarks/performance/g62dgt2e",
   "tg_tokens": null,
   "tps": 141.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-g62dgt2e-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.3 GB at this context; duplicate traces of the same session: mmrq3d4y; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5198.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 5,198 107.5 103.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g62dgt2e",
   "source_url": "https://omlx.ai/benchmarks/performance/g62dgt2e",
   "tg_tokens": null,
   "tps": 107.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-oqg03k3y-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: 8x8l0vrp, ajeul0mx, fiab0ui5, haa4kgic, xeh5u53h; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2734.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 2,734 156.9 100.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oqg03k3y",
   "source_url": "https://omlx.ai/benchmarks/performance/oqg03k3y",
   "tg_tokens": null,
   "tps": 156.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-oqg03k3y-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: 8x8l0vrp, ajeul0mx, fiab0ui5, haa4kgic, xeh5u53h; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5572.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 5,572 146.0 101.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oqg03k3y",
   "source_url": "https://omlx.ai/benchmarks/performance/oqg03k3y",
   "tg_tokens": null,
   "tps": 146.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-oqg03k3y-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.2 GB at this context; duplicate traces of the same session: 8x8l0vrp, ajeul0mx, fiab0ui5, haa4kgic, xeh5u53h; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6320.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 6,320 158.7 102.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oqg03k3y",
   "source_url": "https://omlx.ai/benchmarks/performance/oqg03k3y",
   "tg_tokens": null,
   "tps": 158.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-oqg03k3y-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.5 GB at this context; duplicate traces of the same session: 8x8l0vrp, ajeul0mx, fiab0ui5, haa4kgic, xeh5u53h; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6103.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 6,103 124.7 104.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oqg03k3y",
   "source_url": "https://omlx.ai/benchmarks/performance/oqg03k3y",
   "tg_tokens": null,
   "tps": 124.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-oqg03k3y-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 106.8 GB at this context; duplicate traces of the same session: 8x8l0vrp, ajeul0mx, fiab0ui5, haa4kgic, xeh5u53h; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5904.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 5,904 104.4 106.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oqg03k3y",
   "source_url": "https://omlx.ai/benchmarks/performance/oqg03k3y",
   "tg_tokens": null,
   "tps": 104.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-oqg03k3y-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.6 GB at this context; duplicate traces of the same session: 8x8l0vrp, ajeul0mx, fiab0ui5, haa4kgic, xeh5u53h; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5717.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 5,717 91.8 109.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oqg03k3y",
   "source_url": "https://omlx.ai/benchmarks/performance/oqg03k3y",
   "tg_tokens": null,
   "tps": 91.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-f4ixspwx-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.6 GB at this context; duplicate traces of the same session: 1a76224e, hwwhrkrx, vfff2tif, w1bn2osr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2488.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 2,488 139.1 142.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f4ixspwx",
   "source_url": "https://omlx.ai/benchmarks/performance/f4ixspwx",
   "tg_tokens": null,
   "tps": 139.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-f4ixspwx-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: 1a76224e, hwwhrkrx, vfff2tif, w1bn2osr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4501.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 4,501 133.4 143.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f4ixspwx",
   "source_url": "https://omlx.ai/benchmarks/performance/f4ixspwx",
   "tg_tokens": null,
   "tps": 133.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-f4ixspwx-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 147.0 GB at this context; duplicate traces of the same session: 1a76224e, hwwhrkrx, vfff2tif, w1bn2osr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4830.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 64k 4,830 126.7 147.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f4ixspwx",
   "source_url": "https://omlx.ai/benchmarks/performance/f4ixspwx",
   "tg_tokens": null,
   "tps": 126.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-f4ixspwx-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 149.3 GB at this context; duplicate traces of the same session: 1a76224e, hwwhrkrx, vfff2tif, w1bn2osr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4712.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 128k 4,712 106.9 149.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f4ixspwx",
   "source_url": "https://omlx.ai/benchmarks/performance/f4ixspwx",
   "tg_tokens": null,
   "tps": 106.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-f4ixspwx-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 152.1 GB at this context; duplicate traces of the same session: 1a76224e, hwwhrkrx, vfff2tif, w1bn2osr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4577.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 195k 4,577 83.8 152.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f4ixspwx",
   "source_url": "https://omlx.ai/benchmarks/performance/f4ixspwx",
   "tg_tokens": null,
   "tps": 83.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-lnbcr33d-ctx1024",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: 6ywfr2q6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2056.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 2,056 50.9 25.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lnbcr33d",
   "source_url": "https://omlx.ai/benchmarks/performance/lnbcr33d",
   "tg_tokens": null,
   "tps": 50.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-lnbcr33d-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: 6ywfr2q6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2160.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 2,160 47.8 26.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lnbcr33d",
   "source_url": "https://omlx.ai/benchmarks/performance/lnbcr33d",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-f61llgr9-ctx1024",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2156.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 2,156 38.9 26.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f61llgr9",
   "source_url": "https://omlx.ai/benchmarks/performance/f61llgr9",
   "tg_tokens": null,
   "tps": 38.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-f61llgr9-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 27.2 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1690.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 1,690 60.0 27.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f61llgr9",
   "source_url": "https://omlx.ai/benchmarks/performance/f61llgr9",
   "tg_tokens": null,
   "tps": 60.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-uylldqhu-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: p7qb21a5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2059.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26b-a4b-it-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 2,059 50.8 25.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uylldqhu",
   "source_url": "https://omlx.ai/benchmarks/performance/uylldqhu",
   "tg_tokens": null,
   "tps": 50.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-uylldqhu-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: p7qb21a5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2183.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26b-a4b-it-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 2,183 49.5 26.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uylldqhu",
   "source_url": "https://omlx.ai/benchmarks/performance/uylldqhu",
   "tg_tokens": null,
   "tps": 49.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-lyg7kzgp-ctx1024",
   "model": "KAT-Coder-V2.5-Dev",
   "notes": "OptiQ REAP-18B build; oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: b6k8x5ri; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 345.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--KAT-Coder-V2.5-Dev-OptiQ-4bit-REAP-18B on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 4bit; 1k 345.1 42.8 12.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lyg7kzgp",
   "source_url": "https://omlx.ai/benchmarks/performance/lyg7kzgp",
   "tg_tokens": null,
   "tps": 42.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-lyg7kzgp-ctx4096",
   "model": "KAT-Coder-V2.5-Dev",
   "notes": "OptiQ REAP-18B build; oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: b6k8x5ri; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 350.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--KAT-Coder-V2.5-Dev-OptiQ-4bit-REAP-18B on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 4bit; 4k 350.9 40.9 13.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lyg7kzgp",
   "source_url": "https://omlx.ai/benchmarks/performance/lyg7kzgp",
   "tg_tokens": null,
   "tps": 40.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-7lnq46qh-ctx1024",
   "model": "Qwen3-30B-A3B-Instruct-2507",
   "notes": "oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: z1p18ip0; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 293.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 4bit; 1k 293.6 43.2 16.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7lnq46qh",
   "source_url": "https://omlx.ai/benchmarks/performance/7lnq46qh",
   "tg_tokens": null,
   "tps": 43.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-7lnq46qh-ctx4096",
   "model": "Qwen3-30B-A3B-Instruct-2507",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: z1p18ip0; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 289.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 4bit; 4k 289.1 35.5 17.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7lnq46qh",
   "source_url": "https://omlx.ai/benchmarks/performance/7lnq46qh",
   "tg_tokens": null,
   "tps": 35.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-d90b9oyq-ctx16384",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.9 GB at this context; duplicate traces of the same session: 49box4ny, ijzdcoid, xs4r4xju, z2hun7pa; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1928.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-Uncensored-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,928 63.0 179.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d90b9oyq",
   "source_url": "https://omlx.ai/benchmarks/performance/d90b9oyq",
   "tg_tokens": null,
   "tps": 63.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-d90b9oyq-ctx32768",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.2 GB at this context; duplicate traces of the same session: 49box4ny, ijzdcoid, xs4r4xju, z2hun7pa; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1943.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-Uncensored-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,943 64.5 180.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d90b9oyq",
   "source_url": "https://omlx.ai/benchmarks/performance/d90b9oyq",
   "tg_tokens": null,
   "tps": 64.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-d90b9oyq-ctx65536",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.5 GB at this context; duplicate traces of the same session: 49box4ny, ijzdcoid, xs4r4xju, z2hun7pa; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1912.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-Uncensored-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,912 104.5 180.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d90b9oyq",
   "source_url": "https://omlx.ai/benchmarks/performance/d90b9oyq",
   "tg_tokens": null,
   "tps": 104.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-d90b9oyq-ctx131072",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.3 GB at this context; duplicate traces of the same session: 49box4ny, ijzdcoid, xs4r4xju, z2hun7pa; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1876.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-Uncensored-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 1,876 79.1 181.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d90b9oyq",
   "source_url": "https://omlx.ai/benchmarks/performance/d90b9oyq",
   "tg_tokens": null,
   "tps": 79.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-d90b9oyq-ctx199680",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 182.0 GB at this context; duplicate traces of the same session: 49box4ny, ijzdcoid, xs4r4xju, z2hun7pa; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1837.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-Uncensored-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 1,837 101.3 182.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d90b9oyq",
   "source_url": "https://omlx.ai/benchmarks/performance/d90b9oyq",
   "tg_tokens": null,
   "tps": 101.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-z1p3lgay-ctx1024",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: e962t7a8, orr5cqvt, v4wlo7yy; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1332.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-Uncensored-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,332 72.2 178.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z1p3lgay",
   "source_url": "https://omlx.ai/benchmarks/performance/z1p3lgay",
   "tg_tokens": null,
   "tps": 72.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-z1p3lgay-ctx32768",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.2 GB at this context; duplicate traces of the same session: e962t7a8, orr5cqvt, v4wlo7yy; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1939.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-Uncensored-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,939 106.0 180.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z1p3lgay",
   "source_url": "https://omlx.ai/benchmarks/performance/z1p3lgay",
   "tg_tokens": null,
   "tps": 106.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-z1p3lgay-ctx65536",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.5 GB at this context; duplicate traces of the same session: e962t7a8, orr5cqvt, v4wlo7yy; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1910.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-Uncensored-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,910 81.1 180.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z1p3lgay",
   "source_url": "https://omlx.ai/benchmarks/performance/z1p3lgay",
   "tg_tokens": null,
   "tps": 81.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-z1p3lgay-ctx131072",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.3 GB at this context; duplicate traces of the same session: e962t7a8, orr5cqvt, v4wlo7yy; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1874.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-Uncensored-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 1,874 112.1 181.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z1p3lgay",
   "source_url": "https://omlx.ai/benchmarks/performance/z1p3lgay",
   "tg_tokens": null,
   "tps": 112.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-l0zmb101-ctx1024",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: 2550k8f5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 761.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 761.1 95.7 18.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l0zmb101",
   "source_url": "https://omlx.ai/benchmarks/performance/l0zmb101",
   "tg_tokens": null,
   "tps": 95.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-l0zmb101-ctx4096",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 2550k8f5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 802.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 802.0 94.2 19.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l0zmb101",
   "source_url": "https://omlx.ai/benchmarks/performance/l0zmb101",
   "tg_tokens": null,
   "tps": 94.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-qzqng1mx-ctx1024",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: yhszfnjd; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 715.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 715.1 96.5 18.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qzqng1mx",
   "source_url": "https://omlx.ai/benchmarks/performance/qzqng1mx",
   "tg_tokens": null,
   "tps": 96.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-qzqng1mx-ctx4096",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: yhszfnjd; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 797.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 797.0 94.2 19.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qzqng1mx",
   "source_url": "https://omlx.ai/benchmarks/performance/qzqng1mx",
   "tg_tokens": null,
   "tps": 94.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-lblujb95-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2716.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 2,716 121.0 37.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lblujb95",
   "source_url": "https://omlx.ai/benchmarks/performance/lblujb95",
   "tg_tokens": null,
   "tps": 121.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lblujb95-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3352.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 3,352 124.0 39.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lblujb95",
   "source_url": "https://omlx.ai/benchmarks/performance/lblujb95",
   "tg_tokens": null,
   "tps": 124.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-lblujb95-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3185.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 3,185 133.3 39.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lblujb95",
   "source_url": "https://omlx.ai/benchmarks/performance/lblujb95",
   "tg_tokens": null,
   "tps": 133.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-0d8t9k5p-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2855.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 2,855 89.8 35.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0d8t9k5p",
   "source_url": "https://omlx.ai/benchmarks/performance/0d8t9k5p",
   "tg_tokens": null,
   "tps": 89.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-0d8t9k5p-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3438.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 3,438 89.3 37.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0d8t9k5p",
   "source_url": "https://omlx.ai/benchmarks/performance/0d8t9k5p",
   "tg_tokens": null,
   "tps": 89.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-0d8t9k5p-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3279.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 3,279 88.0 37.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0d8t9k5p",
   "source_url": "https://omlx.ai/benchmarks/performance/0d8t9k5p",
   "tg_tokens": null,
   "tps": 88.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-yburo8az-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 38.3 GB; TTFT 12768 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5133.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 64k tokens 5,133 PP tok/s 130.7 TG tok/s 12768 TTFT (ms) 38.3 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yburo8az",
   "source_url": "https://omlx.ai/benchmarks/performance/yburo8az",
   "tg_tokens": null,
   "tps": 130.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ri7120ti-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: opeho3js; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 924.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 924.2 93.7 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ri7120ti",
   "source_url": "https://omlx.ai/benchmarks/performance/ri7120ti",
   "tg_tokens": null,
   "tps": 93.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ri7120ti-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: opeho3js; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1068.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,068 90.7 21.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ri7120ti",
   "source_url": "https://omlx.ai/benchmarks/performance/ri7120ti",
   "tg_tokens": null,
   "tps": 90.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-w0cagj3j-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 16.0 GB; TTFT 20758 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 49.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M2 Pro (16c); M2 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k tokens 49.3 PP tok/s 9.4 TG tok/s 20758 TTFT (ms) 16.0 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w0cagj3j",
   "source_url": "https://omlx.ai/benchmarks/performance/w0cagj3j",
   "tg_tokens": null,
   "tps": 9.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-bbdecl07-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 113.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 1k 113.9 11.6 28.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bbdecl07",
   "source_url": "https://omlx.ai/benchmarks/performance/bbdecl07",
   "tg_tokens": null,
   "tps": 11.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-bbdecl07-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.6 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 116.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 4k 116.8 11.4 31.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bbdecl07",
   "source_url": "https://omlx.ai/benchmarks/performance/bbdecl07",
   "tg_tokens": null,
   "tps": 11.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-khuf472i-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: jlrkrvha; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 219.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 219.7 13.0 28.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/khuf472i",
   "source_url": "https://omlx.ai/benchmarks/performance/khuf472i",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-khuf472i-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.7 GB at this context; duplicate traces of the same session: jlrkrvha; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 220.3 12.9 31.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/khuf472i",
   "source_url": "https://omlx.ai/benchmarks/performance/khuf472i",
   "tg_tokens": null,
   "tps": 12.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i8169c3g-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 1fgzs2fn, 4p0tesip, a81ojsv3, c1o2wn3n, f4xdeqm0, sulqzts4, yscokn1s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1090.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,090 54.7 17.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i8169c3g",
   "source_url": "https://omlx.ai/benchmarks/performance/i8169c3g",
   "tg_tokens": null,
   "tps": 54.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i8169c3g-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 1fgzs2fn, 4p0tesip, a81ojsv3, c1o2wn3n, f4xdeqm0, sulqzts4, yscokn1s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1266.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,266 65.0 19.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i8169c3g",
   "source_url": "https://omlx.ai/benchmarks/performance/i8169c3g",
   "tg_tokens": null,
   "tps": 65.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i8169c3g-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 1fgzs2fn, 4p0tesip, a81ojsv3, c1o2wn3n, f4xdeqm0, sulqzts4, yscokn1s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1262.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,262 56.3 19.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i8169c3g",
   "source_url": "https://omlx.ai/benchmarks/performance/i8169c3g",
   "tg_tokens": null,
   "tps": 56.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-i8169c3g-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 1fgzs2fn, 4p0tesip, a81ojsv3, c1o2wn3n, f4xdeqm0, sulqzts4, yscokn1s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1203.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,203 85.9 19.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i8169c3g",
   "source_url": "https://omlx.ai/benchmarks/performance/i8169c3g",
   "tg_tokens": null,
   "tps": 85.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-i8169c3g-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: 1fgzs2fn, 4p0tesip, a81ojsv3, c1o2wn3n, f4xdeqm0, sulqzts4, yscokn1s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1052.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,052 131.4 21.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i8169c3g",
   "source_url": "https://omlx.ai/benchmarks/performance/i8169c3g",
   "tg_tokens": null,
   "tps": 131.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-i8169c3g-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 1fgzs2fn, 4p0tesip, a81ojsv3, c1o2wn3n, f4xdeqm0, sulqzts4, yscokn1s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 886.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 886.0 56.6 23.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i8169c3g",
   "source_url": "https://omlx.ai/benchmarks/performance/i8169c3g",
   "tg_tokens": null,
   "tps": 56.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i8169c3g-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: 1fgzs2fn, 4p0tesip, a81ojsv3, c1o2wn3n, f4xdeqm0, sulqzts4, yscokn1s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 566.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 566.4 43.2 28.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i8169c3g",
   "source_url": "https://omlx.ai/benchmarks/performance/i8169c3g",
   "tg_tokens": null,
   "tps": 43.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i8169c3g-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.2 GB at this context; duplicate traces of the same session: 1fgzs2fn, 4p0tesip, a81ojsv3, c1o2wn3n, f4xdeqm0, sulqzts4, yscokn1s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 413.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 413.7 24.1 33.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i8169c3g",
   "source_url": "https://omlx.ai/benchmarks/performance/i8169c3g",
   "tg_tokens": null,
   "tps": 24.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-irgjtaf3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: mqm190e5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 472.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 472.5 32.5 20.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/irgjtaf3",
   "source_url": "https://omlx.ai/benchmarks/performance/irgjtaf3",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-irgjtaf3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: mqm190e5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 515.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 515.7 33.1 21.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/irgjtaf3",
   "source_url": "https://omlx.ai/benchmarks/performance/irgjtaf3",
   "tg_tokens": null,
   "tps": 33.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zi1hvb8q-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 37.3 GB; TTFT 235305 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 557.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k tokens 557.0 PP tok/s 14.5 TG tok/s 235305 TTFT (ms) 37.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zi1hvb8q",
   "source_url": "https://omlx.ai/benchmarks/performance/zi1hvb8q",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-ac0z5umf-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 72.2 GB at this context; duplicate traces of the same session: 0opyszml, 9sounonn, occxpys2, qem626il, t4iiifwu, ulbtddxr; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 297.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 1k 297.3 22.4 72.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ac0z5umf",
   "source_url": "https://omlx.ai/benchmarks/performance/ac0z5umf",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-ac0z5umf-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 74.8 GB at this context; duplicate traces of the same session: 0opyszml, 9sounonn, occxpys2, qem626il, t4iiifwu, ulbtddxr; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 445.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 4k 445.4 46.8 74.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ac0z5umf",
   "source_url": "https://omlx.ai/benchmarks/performance/ac0z5umf",
   "tg_tokens": null,
   "tps": 46.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-ac0z5umf-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 75.1 GB at this context; duplicate traces of the same session: 0opyszml, 9sounonn, occxpys2, qem626il, t4iiifwu, ulbtddxr; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1032.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 8k 1,032 37.8 75.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ac0z5umf",
   "source_url": "https://omlx.ai/benchmarks/performance/ac0z5umf",
   "tg_tokens": null,
   "tps": 37.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-ac0z5umf-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 75.4 GB at this context; duplicate traces of the same session: 0opyszml, 9sounonn, occxpys2, qem626il, t4iiifwu, ulbtddxr; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1040.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 16k 1,040 40.3 75.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ac0z5umf",
   "source_url": "https://omlx.ai/benchmarks/performance/ac0z5umf",
   "tg_tokens": null,
   "tps": 40.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-ac0z5umf-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 77.1 GB at this context; duplicate traces of the same session: 0opyszml, 9sounonn, occxpys2, qem626il, t4iiifwu, ulbtddxr; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1016.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 64k 1,016 48.8 77.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ac0z5umf",
   "source_url": "https://omlx.ai/benchmarks/performance/ac0z5umf",
   "tg_tokens": null,
   "tps": 48.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-ac0z5umf-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 79.4 GB at this context; duplicate traces of the same session: 0opyszml, 9sounonn, occxpys2, qem626il, t4iiifwu, ulbtddxr; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 992.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 128k 992.3 45.1 79.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ac0z5umf",
   "source_url": "https://omlx.ai/benchmarks/performance/ac0z5umf",
   "tg_tokens": null,
   "tps": 45.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-ac0z5umf-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 82.4 GB at this context; duplicate traces of the same session: 0opyszml, 9sounonn, occxpys2, qem626il, t4iiifwu, ulbtddxr; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 969.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 195k 969.5 37.5 82.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ac0z5umf",
   "source_url": "https://omlx.ai/benchmarks/performance/ac0z5umf",
   "tg_tokens": null,
   "tps": 37.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zgptsv5b-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 7xyc9w5m, v09re0v8, w8nrnba3, zj6kjbcc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2394.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,394 50.2 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zgptsv5b",
   "source_url": "https://omlx.ai/benchmarks/performance/zgptsv5b",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zgptsv5b-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: 7xyc9w5m, v09re0v8, w8nrnba3, zj6kjbcc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2830.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,830 63.0 72.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zgptsv5b",
   "source_url": "https://omlx.ai/benchmarks/performance/zgptsv5b",
   "tg_tokens": null,
   "tps": 63.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zgptsv5b-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: 7xyc9w5m, v09re0v8, w8nrnba3, zj6kjbcc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2708.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,708 38.0 73.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zgptsv5b",
   "source_url": "https://omlx.ai/benchmarks/performance/zgptsv5b",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zgptsv5b-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.9 GB at this context; duplicate traces of the same session: 7xyc9w5m, v09re0v8, w8nrnba3, zj6kjbcc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2598.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,598 69.0 74.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zgptsv5b",
   "source_url": "https://omlx.ai/benchmarks/performance/zgptsv5b",
   "tg_tokens": null,
   "tps": 69.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zgptsv5b-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.2 GB at this context; duplicate traces of the same session: 7xyc9w5m, v09re0v8, w8nrnba3, zj6kjbcc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2524.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,524 60.2 79.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zgptsv5b",
   "source_url": "https://omlx.ai/benchmarks/performance/zgptsv5b",
   "tg_tokens": null,
   "tps": 60.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-qy6lc49l-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace, single run at 64k context; peak memory 147.2 GB; TTFT 16062 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4080.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 64k tokens 4,080 PP tok/s 109.7 TG tok/s 16062 TTFT (ms) 147.2 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qy6lc49l",
   "source_url": "https://omlx.ai/benchmarks/performance/qy6lc49l",
   "tg_tokens": null,
   "tps": 109.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-j22ufm3l-ctx1024",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2156.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 2,156 38.9 26.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j22ufm3l",
   "source_url": "https://omlx.ai/benchmarks/performance/j22ufm3l",
   "tg_tokens": null,
   "tps": 38.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-j22ufm3l-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 27.2 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1690.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 1,690 60.0 27.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j22ufm3l",
   "source_url": "https://omlx.ai/benchmarks/performance/j22ufm3l",
   "tg_tokens": null,
   "tps": 60.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-yrqnh0z6-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: a07u0hx7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 780.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 780.4 68.5 14.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yrqnh0z6",
   "source_url": "https://omlx.ai/benchmarks/performance/yrqnh0z6",
   "tg_tokens": null,
   "tps": 68.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-yrqnh0z6-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: a07u0hx7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 744.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 744.1 66.1 14.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yrqnh0z6",
   "source_url": "https://omlx.ai/benchmarks/performance/yrqnh0z6",
   "tg_tokens": null,
   "tps": 66.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ir21rwm0-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: 49ds2mc9, 633gweg6, 8y55uys9, mndatyuw, nzrdzq9w, ub2rboa9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2918.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,918 131.7 5.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ir21rwm0",
   "source_url": "https://omlx.ai/benchmarks/performance/ir21rwm0",
   "tg_tokens": null,
   "tps": 131.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ir21rwm0-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: 49ds2mc9, 633gweg6, 8y55uys9, mndatyuw, nzrdzq9w, ub2rboa9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6817.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 6,817 150.9 5.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ir21rwm0",
   "source_url": "https://omlx.ai/benchmarks/performance/ir21rwm0",
   "tg_tokens": null,
   "tps": 150.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ir21rwm0-ctx8192",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 6.0 GB at this context; duplicate traces of the same session: 49ds2mc9, 633gweg6, 8y55uys9, mndatyuw, nzrdzq9w, ub2rboa9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6992.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 6,992 124.7 6.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ir21rwm0",
   "source_url": "https://omlx.ai/benchmarks/performance/ir21rwm0",
   "tg_tokens": null,
   "tps": 124.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ir21rwm0-ctx16384",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: 49ds2mc9, 633gweg6, 8y55uys9, mndatyuw, nzrdzq9w, ub2rboa9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6589.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 6,589 135.3 6.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ir21rwm0",
   "source_url": "https://omlx.ai/benchmarks/performance/ir21rwm0",
   "tg_tokens": null,
   "tps": 135.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ir21rwm0-ctx32768",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: 49ds2mc9, 633gweg6, 8y55uys9, mndatyuw, nzrdzq9w, ub2rboa9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6185.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 6,185 94.1 6.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ir21rwm0",
   "source_url": "https://omlx.ai/benchmarks/performance/ir21rwm0",
   "tg_tokens": null,
   "tps": 94.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ir21rwm0-ctx65536",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.7 GB at this context; duplicate traces of the same session: 49ds2mc9, 633gweg6, 8y55uys9, mndatyuw, nzrdzq9w, ub2rboa9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5354.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 5,354 54.8 7.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ir21rwm0",
   "source_url": "https://omlx.ai/benchmarks/performance/ir21rwm0",
   "tg_tokens": null,
   "tps": 54.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ir21rwm0-ctx131072",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 9.9 GB at this context; duplicate traces of the same session: 49ds2mc9, 633gweg6, 8y55uys9, mndatyuw, nzrdzq9w, ub2rboa9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3881.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 3,881 39.1 9.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ir21rwm0",
   "source_url": "https://omlx.ai/benchmarks/performance/ir21rwm0",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-1qr39zwp-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: po1ajnkd; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 805.6,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 805.6 73.5 11.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1qr39zwp",
   "source_url": "https://omlx.ai/benchmarks/performance/1qr39zwp",
   "tg_tokens": null,
   "tps": 73.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-1qr39zwp-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: po1ajnkd; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 808.6,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 808.6 70.3 11.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1qr39zwp",
   "source_url": "https://omlx.ai/benchmarks/performance/1qr39zwp",
   "tg_tokens": null,
   "tps": 70.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-bsn89hfq-ctx1024",
   "model": "Qwen2-VL-2B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 1.5 GB at this context; duplicate traces of the same session: qtskcj5h; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2112.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "mlx-community--Qwen2-VL-2B-Instruct-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 unknown; 1k 2,112 199.4 1.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bsn89hfq",
   "source_url": "https://omlx.ai/benchmarks/performance/bsn89hfq",
   "tg_tokens": null,
   "tps": 199.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-bsn89hfq-ctx4096",
   "model": "Qwen2-VL-2B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 1.6 GB at this context; duplicate traces of the same session: qtskcj5h; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2052.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "mlx-community--Qwen2-VL-2B-Instruct-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 unknown; 4k 2,052 181.2 1.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bsn89hfq",
   "source_url": "https://omlx.ai/benchmarks/performance/bsn89hfq",
   "tg_tokens": null,
   "tps": 181.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-oxtoo3dl-ctx1024",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: xck54jnl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.2,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 5bit; 1k 204.2 18.0 19.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oxtoo3dl",
   "source_url": "https://omlx.ai/benchmarks/performance/oxtoo3dl",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-oxtoo3dl-ctx8192",
   "model": "AREX-2",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: xck54jnl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 185.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "AREX-2-5bit on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 5bit; 8k 185.0 17.3 20.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oxtoo3dl",
   "source_url": "https://omlx.ai/benchmarks/performance/oxtoo3dl",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rh9xok2s-ctx1024",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 177.6 GB at this context; duplicate traces of the same session: 3nvpqlaz, 6jgra9kr, m0tvpm4d, nmvrbhxy, o8shehhy, oaw3tbo7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 514.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 514.7 46.4 177.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rh9xok2s",
   "source_url": "https://omlx.ai/benchmarks/performance/rh9xok2s",
   "tg_tokens": null,
   "tps": 46.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rh9xok2s-ctx4096",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 178.8 GB at this context; duplicate traces of the same session: 3nvpqlaz, 6jgra9kr, m0tvpm4d, nmvrbhxy, o8shehhy, oaw3tbo7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 592.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 592.6 43.8 178.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rh9xok2s",
   "source_url": "https://omlx.ai/benchmarks/performance/rh9xok2s",
   "tg_tokens": null,
   "tps": 43.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rh9xok2s-ctx8192",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 179.1 GB at this context; duplicate traces of the same session: 3nvpqlaz, 6jgra9kr, m0tvpm4d, nmvrbhxy, o8shehhy, oaw3tbo7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 547.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 547.5 58.1 179.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rh9xok2s",
   "source_url": "https://omlx.ai/benchmarks/performance/rh9xok2s",
   "tg_tokens": null,
   "tps": 58.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rh9xok2s-ctx16384",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 179.2 GB at this context; duplicate traces of the same session: 3nvpqlaz, 6jgra9kr, m0tvpm4d, nmvrbhxy, o8shehhy, oaw3tbo7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 537.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 537.1 89.4 179.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rh9xok2s",
   "source_url": "https://omlx.ai/benchmarks/performance/rh9xok2s",
   "tg_tokens": null,
   "tps": 89.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rh9xok2s-ctx32768",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 179.4 GB at this context; duplicate traces of the same session: 3nvpqlaz, 6jgra9kr, m0tvpm4d, nmvrbhxy, o8shehhy, oaw3tbo7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 531.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 531.7 64.7 179.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rh9xok2s",
   "source_url": "https://omlx.ai/benchmarks/performance/rh9xok2s",
   "tg_tokens": null,
   "tps": 64.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rh9xok2s-ctx65536",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 179.8 GB at this context; duplicate traces of the same session: 3nvpqlaz, 6jgra9kr, m0tvpm4d, nmvrbhxy, o8shehhy, oaw3tbo7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 522.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 522.4 48.4 179.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rh9xok2s",
   "source_url": "https://omlx.ai/benchmarks/performance/rh9xok2s",
   "tg_tokens": null,
   "tps": 48.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rh9xok2s-ctx131072",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 180.8 GB at this context; duplicate traces of the same session: 3nvpqlaz, 6jgra9kr, m0tvpm4d, nmvrbhxy, o8shehhy, oaw3tbo7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 493.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 493.3 94.2 180.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rh9xok2s",
   "source_url": "https://omlx.ai/benchmarks/performance/rh9xok2s",
   "tg_tokens": null,
   "tps": 94.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-onwg30qe-ctx1024",
   "model": "LFM2.5-2.6B",
   "notes": "oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: njoygs9j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1998.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-2.6B-oQ8e on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 8bit; 1k 1,998 115.0 3.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/onwg30qe",
   "source_url": "https://omlx.ai/benchmarks/performance/onwg30qe",
   "tg_tokens": null,
   "tps": 115.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-onwg30qe-ctx8192",
   "model": "LFM2.5-2.6B",
   "notes": "oMLX community benchmark trace; peak memory 3.6 GB at this context; duplicate traces of the same session: njoygs9j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2240.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-2.6B-oQ8e on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 8bit; 8k 2,240 108.8 3.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/onwg30qe",
   "source_url": "https://omlx.ai/benchmarks/performance/onwg30qe",
   "tg_tokens": null,
   "tps": 108.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9esfnfxd-ctx1024",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 61.2 GB at this context; duplicate traces of the same session: 9usupinb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2160.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-bf16 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 bf16; 1k 2,160 67.6 61.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9esfnfxd",
   "source_url": "https://omlx.ai/benchmarks/performance/9esfnfxd",
   "tg_tokens": null,
   "tps": 67.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9esfnfxd-ctx4096",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 62.1 GB at this context; duplicate traces of the same session: 9usupinb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2948.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-bf16 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 bf16; 4k 2,948 67.3 62.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9esfnfxd",
   "source_url": "https://omlx.ai/benchmarks/performance/9esfnfxd",
   "tg_tokens": null,
   "tps": 67.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fvv8s4as-ctx1024",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 61.2 GB at this context; duplicate traces of the same session: 1v9b5ml3, rvdfugfj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1994.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-bf16 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 bf16; 1k 1,994 62.9 61.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fvv8s4as",
   "source_url": "https://omlx.ai/benchmarks/performance/fvv8s4as",
   "tg_tokens": null,
   "tps": 62.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fvv8s4as-ctx4096",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 62.1 GB at this context; duplicate traces of the same session: 1v9b5ml3, rvdfugfj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2828.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-bf16 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 bf16; 4k 2,828 65.5 62.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fvv8s4as",
   "source_url": "https://omlx.ai/benchmarks/performance/fvv8s4as",
   "tg_tokens": null,
   "tps": 65.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fvv8s4as-ctx199680",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 63.2 GB at this context; duplicate traces of the same session: 1v9b5ml3, rvdfugfj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1797.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-bf16 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 bf16; 195k 1,797 53.8 63.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fvv8s4as",
   "source_url": "https://omlx.ai/benchmarks/performance/fvv8s4as",
   "tg_tokens": null,
   "tps": 53.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dmmnnot6-ctx1024",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 61.2 GB at this context; duplicate traces of the same session: s4gqytry; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2081.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-bf16 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 bf16; 1k 2,081 67.6 61.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dmmnnot6",
   "source_url": "https://omlx.ai/benchmarks/performance/dmmnnot6",
   "tg_tokens": null,
   "tps": 67.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dmmnnot6-ctx4096",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 62.1 GB at this context; duplicate traces of the same session: s4gqytry; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2899.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-bf16 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 bf16; 4k 2,899 67.2 62.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dmmnnot6",
   "source_url": "https://omlx.ai/benchmarks/performance/dmmnnot6",
   "tg_tokens": null,
   "tps": 67.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-khkcp2h9-ctx1024",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 61.2 GB at this context; duplicate traces of the same session: 3b0c33vo, 7twdg31l, atj91ofp, e366qb0m, qum68gba, vrete7su; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1134.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-bf16 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 bf16; 1k 1,134 67.8 61.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/khkcp2h9",
   "source_url": "https://omlx.ai/benchmarks/performance/khkcp2h9",
   "tg_tokens": null,
   "tps": 67.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-khkcp2h9-ctx4096",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 62.1 GB at this context; duplicate traces of the same session: 3b0c33vo, 7twdg31l, atj91ofp, e366qb0m, qum68gba, vrete7su; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2031.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-bf16 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 bf16; 4k 2,031 67.9 62.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/khkcp2h9",
   "source_url": "https://omlx.ai/benchmarks/performance/khkcp2h9",
   "tg_tokens": null,
   "tps": 67.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-khkcp2h9-ctx8192",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 62.1 GB at this context; duplicate traces of the same session: 3b0c33vo, 7twdg31l, atj91ofp, e366qb0m, qum68gba, vrete7su; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2831.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-bf16 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 bf16; 8k 2,831 67.1 62.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/khkcp2h9",
   "source_url": "https://omlx.ai/benchmarks/performance/khkcp2h9",
   "tg_tokens": null,
   "tps": 67.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-khkcp2h9-ctx16384",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 62.2 GB at this context; duplicate traces of the same session: 3b0c33vo, 7twdg31l, atj91ofp, e366qb0m, qum68gba, vrete7su; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2813.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-bf16 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 bf16; 16k 2,813 67.0 62.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/khkcp2h9",
   "source_url": "https://omlx.ai/benchmarks/performance/khkcp2h9",
   "tg_tokens": null,
   "tps": 67.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-khkcp2h9-ctx32768",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 62.3 GB at this context; duplicate traces of the same session: 3b0c33vo, 7twdg31l, atj91ofp, e366qb0m, qum68gba, vrete7su; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2703.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-bf16 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 bf16; 32k 2,703 66.2 62.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/khkcp2h9",
   "source_url": "https://omlx.ai/benchmarks/performance/khkcp2h9",
   "tg_tokens": null,
   "tps": 66.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-khkcp2h9-ctx65536",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 62.5 GB at this context; duplicate traces of the same session: 3b0c33vo, 7twdg31l, atj91ofp, e366qb0m, qum68gba, vrete7su; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2439.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-bf16 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 bf16; 64k 2,439 63.3 62.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/khkcp2h9",
   "source_url": "https://omlx.ai/benchmarks/performance/khkcp2h9",
   "tg_tokens": null,
   "tps": 63.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-khkcp2h9-ctx131072",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 62.8 GB at this context; duplicate traces of the same session: 3b0c33vo, 7twdg31l, atj91ofp, e366qb0m, qum68gba, vrete7su; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2033.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-bf16 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 bf16; 128k 2,033 58.4 62.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/khkcp2h9",
   "source_url": "https://omlx.ai/benchmarks/performance/khkcp2h9",
   "tg_tokens": null,
   "tps": 58.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-8jz72soo-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: 4eekph0o; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2716.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 2,716 121.0 37.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8jz72soo",
   "source_url": "https://omlx.ai/benchmarks/performance/8jz72soo",
   "tg_tokens": null,
   "tps": 121.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-8jz72soo-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.1 GB at this context; duplicate traces of the same session: 4eekph0o; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3352.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 3,352 124.0 39.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8jz72soo",
   "source_url": "https://omlx.ai/benchmarks/performance/8jz72soo",
   "tg_tokens": null,
   "tps": 124.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-8jz72soo-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: 4eekph0o; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3185.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 3,185 133.3 39.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8jz72soo",
   "source_url": "https://omlx.ai/benchmarks/performance/8jz72soo",
   "tg_tokens": null,
   "tps": 133.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-0n50165w-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: blaffgqt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2855.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 2,855 89.8 35.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0n50165w",
   "source_url": "https://omlx.ai/benchmarks/performance/0n50165w",
   "tg_tokens": null,
   "tps": 89.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-0n50165w-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: blaffgqt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3438.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 3,438 89.3 37.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0n50165w",
   "source_url": "https://omlx.ai/benchmarks/performance/0n50165w",
   "tg_tokens": null,
   "tps": 89.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-0n50165w-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: blaffgqt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3279.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 3,279 88.0 37.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0n50165w",
   "source_url": "https://omlx.ai/benchmarks/performance/0n50165w",
   "tg_tokens": null,
   "tps": 88.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-c0yj7kga-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.5 GB at this context; duplicate traces of the same session: 2l4bn6fx, h7v84es4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2754.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 2,754 127.8 36.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c0yj7kga",
   "source_url": "https://omlx.ai/benchmarks/performance/c0yj7kga",
   "tg_tokens": null,
   "tps": 127.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-c0yj7kga-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.3 GB at this context; duplicate traces of the same session: 2l4bn6fx, h7v84es4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3389.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 3,389 113.2 38.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c0yj7kga",
   "source_url": "https://omlx.ai/benchmarks/performance/c0yj7kga",
   "tg_tokens": null,
   "tps": 113.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-c0yj7kga-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.4 GB at this context; duplicate traces of the same session: 2l4bn6fx, h7v84es4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3257.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 3,257 131.5 38.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c0yj7kga",
   "source_url": "https://omlx.ai/benchmarks/performance/c0yj7kga",
   "tg_tokens": null,
   "tps": 131.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-tze4euhd-ctx16384",
   "model": "Ornith-1.5-9B",
   "notes": "fp16 MTP build; oMLX community benchmark trace, single run at 16k context; peak memory 8.2 GB; TTFT 70388 ms; duplicate traces of the same session: none; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 232.8,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Ornith-1.5-9B-oQ5e-fp16-mtp on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 5bit; 16k tokens 232.8 PP tok/s 23.2 TG tok/s 70388 TTFT (ms) 8.2 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tze4euhd",
   "source_url": "https://omlx.ai/benchmarks/performance/tze4euhd",
   "tg_tokens": null,
   "tps": 23.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-jg8rwk9i-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: k39a5cpx; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 335.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 335.0 45.7 19.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jg8rwk9i",
   "source_url": "https://omlx.ai/benchmarks/performance/jg8rwk9i",
   "tg_tokens": null,
   "tps": 45.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-jg8rwk9i-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: k39a5cpx; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 355.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 355.8 42.6 22.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jg8rwk9i",
   "source_url": "https://omlx.ai/benchmarks/performance/jg8rwk9i",
   "tg_tokens": null,
   "tps": 42.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9nrri848-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: 7wnq1zbs, bg09l8sa, biqy7yuq, mj9d4s44, rlel1168; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 849.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 849.0 32.2 33.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9nrri848",
   "source_url": "https://omlx.ai/benchmarks/performance/9nrri848",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9nrri848-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: 7wnq1zbs, bg09l8sa, biqy7yuq, mj9d4s44, rlel1168; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 794.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 794.5 29.0 34.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9nrri848",
   "source_url": "https://omlx.ai/benchmarks/performance/9nrri848",
   "tg_tokens": null,
   "tps": 29.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9nrri848-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 35.1 GB at this context; duplicate traces of the same session: 7wnq1zbs, bg09l8sa, biqy7yuq, mj9d4s44, rlel1168; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 696.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 696.8 26.1 35.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9nrri848",
   "source_url": "https://omlx.ai/benchmarks/performance/9nrri848",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9nrri848-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 35.6 GB at this context; duplicate traces of the same session: 7wnq1zbs, bg09l8sa, biqy7yuq, mj9d4s44, rlel1168; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 641.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 641.3 35.6 35.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9nrri848",
   "source_url": "https://omlx.ai/benchmarks/performance/9nrri848",
   "tg_tokens": null,
   "tps": 35.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9nrri848-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 36.6 GB at this context; duplicate traces of the same session: 7wnq1zbs, bg09l8sa, biqy7yuq, mj9d4s44, rlel1168; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 619.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 619.0 32.3 36.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9nrri848",
   "source_url": "https://omlx.ai/benchmarks/performance/9nrri848",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9nrri848-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 38.6 GB at this context; duplicate traces of the same session: 7wnq1zbs, bg09l8sa, biqy7yuq, mj9d4s44, rlel1168; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 529.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 529.0 22.7 38.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9nrri848",
   "source_url": "https://omlx.ai/benchmarks/performance/9nrri848",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9nrri848-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 42.6 GB at this context; duplicate traces of the same session: 7wnq1zbs, bg09l8sa, biqy7yuq, mj9d4s44, rlel1168; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 370.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 370.7 20.7 42.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9nrri848",
   "source_url": "https://omlx.ai/benchmarks/performance/9nrri848",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9nrri848-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 46.8 GB at this context; duplicate traces of the same session: 7wnq1zbs, bg09l8sa, biqy7yuq, mj9d4s44, rlel1168; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 279.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 195k 279.5 13.4 46.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9nrri848",
   "source_url": "https://omlx.ai/benchmarks/performance/9nrri848",
   "tg_tokens": null,
   "tps": 13.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-qb8sc63l-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 9opgxnou; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1363.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 1k 1,363 32.4 28.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qb8sc63l",
   "source_url": "https://omlx.ai/benchmarks/performance/qb8sc63l",
   "tg_tokens": null,
   "tps": 32.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-qb8sc63l-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: 9opgxnou; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1712.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 4k 1,712 32.2 29.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qb8sc63l",
   "source_url": "https://omlx.ai/benchmarks/performance/qb8sc63l",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-0j8j29y5-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 113.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 1k 113.9 11.6 28.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0j8j29y5",
   "source_url": "https://omlx.ai/benchmarks/performance/0j8j29y5",
   "tg_tokens": null,
   "tps": 11.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-0j8j29y5-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.6 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 116.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 4k 116.8 11.4 31.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0j8j29y5",
   "source_url": "https://omlx.ai/benchmarks/performance/0j8j29y5",
   "tg_tokens": null,
   "tps": 11.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-a1j5cm10-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 43t1r0sx, 6vy3zut3, od9d7ph8, pd9dpmdk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 674.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 674.5 32.0 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a1j5cm10",
   "source_url": "https://omlx.ai/benchmarks/performance/a1j5cm10",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-a1j5cm10-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 43t1r0sx, 6vy3zut3, od9d7ph8, pd9dpmdk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 611.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 611.4 30.1 22.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a1j5cm10",
   "source_url": "https://omlx.ai/benchmarks/performance/a1j5cm10",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-a1j5cm10-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: 43t1r0sx, 6vy3zut3, od9d7ph8, pd9dpmdk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 517.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 517.8 28.2 24.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a1j5cm10",
   "source_url": "https://omlx.ai/benchmarks/performance/a1j5cm10",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-a1j5cm10-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.9 GB at this context; duplicate traces of the same session: 43t1r0sx, 6vy3zut3, od9d7ph8, pd9dpmdk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 382.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 128k 382.9 24.0 28.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a1j5cm10",
   "source_url": "https://omlx.ai/benchmarks/performance/a1j5cm10",
   "tg_tokens": null,
   "tps": 24.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-a1j5cm10-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.2 GB at this context; duplicate traces of the same session: 43t1r0sx, 6vy3zut3, od9d7ph8, pd9dpmdk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 296.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 195k 296.7 24.7 34.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a1j5cm10",
   "source_url": "https://omlx.ai/benchmarks/performance/a1j5cm10",
   "tg_tokens": null,
   "tps": 24.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ywccktm2-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 42thulqo; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 799.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k 799.2 26.3 17.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ywccktm2",
   "source_url": "https://omlx.ai/benchmarks/performance/ywccktm2",
   "tg_tokens": null,
   "tps": 26.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ywccktm2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: 42thulqo; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 583.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 4k 583.8 36.4 18.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ywccktm2",
   "source_url": "https://omlx.ai/benchmarks/performance/ywccktm2",
   "tg_tokens": null,
   "tps": 36.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wl1k7ysn-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 71.0 GB; TTFT 5314 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 770.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 770.8 PP tok/s 48.7 TG tok/s 5314 TTFT (ms) 71.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wl1k7ysn",
   "source_url": "https://omlx.ai/benchmarks/performance/wl1k7ysn",
   "tg_tokens": null,
   "tps": 48.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-nikicafd-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.8 GB; TTFT 795 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1288.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k tokens 1,288 PP tok/s 125.0 TG tok/s 795 TTFT (ms) 70.8 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nikicafd",
   "source_url": "https://omlx.ai/benchmarks/performance/nikicafd",
   "tg_tokens": null,
   "tps": 125.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-j4m73gck-ctx1024",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: iaxruo06; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 708.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 708.3 25.6 18.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j4m73gck",
   "source_url": "https://omlx.ai/benchmarks/performance/j4m73gck",
   "tg_tokens": null,
   "tps": 25.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-j4m73gck-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: iaxruo06; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 692.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 692.7 24.0 19.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j4m73gck",
   "source_url": "https://omlx.ai/benchmarks/performance/j4m73gck",
   "tg_tokens": null,
   "tps": 24.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-3wasf3lw-ctx4096",
   "model": "gemma-4-E4B-it",
   "notes": "oMLX community benchmark trace; peak memory 8.1 GB at this context; duplicate traces of the same session: 8orci0j2, ypy2v0ve; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1001.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-E4B-it on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 8bit; 4k 1,001 39.6 8.1 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3wasf3lw",
   "source_url": "https://omlx.ai/benchmarks/performance/3wasf3lw",
   "tg_tokens": null,
   "tps": 39.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-3wasf3lw-ctx16384",
   "model": "gemma-4-E4B-it",
   "notes": "oMLX community benchmark trace; peak memory 8.5 GB at this context; duplicate traces of the same session: 8orci0j2, ypy2v0ve; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 973.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-E4B-it on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 8bit; 16k 973.6 36.9 8.5 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3wasf3lw",
   "source_url": "https://omlx.ai/benchmarks/performance/3wasf3lw",
   "tg_tokens": null,
   "tps": 36.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-3wasf3lw-ctx131072",
   "model": "gemma-4-E4B-it",
   "notes": "oMLX community benchmark trace; peak memory 11.3 GB at this context; duplicate traces of the same session: 8orci0j2, ypy2v0ve; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 627.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-E4B-it on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 8bit; 128k 627.0 20.8 11.3 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3wasf3lw",
   "source_url": "https://omlx.ai/benchmarks/performance/3wasf3lw",
   "tg_tokens": null,
   "tps": 20.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-c8ccfgtc-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 11piolfx; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 131.4 15.5 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c8ccfgtc",
   "source_url": "https://omlx.ai/benchmarks/performance/c8ccfgtc",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-c8ccfgtc-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 11piolfx; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 137.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 137.2 14.5 19.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c8ccfgtc",
   "source_url": "https://omlx.ai/benchmarks/performance/c8ccfgtc",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-t8g4hcf3-ctx4096",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.3 GB at this context; duplicate traces of the same session: 6n4yg00i, nzb0ojm4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 590.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 590.1 34.8 156.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t8g4hcf3",
   "source_url": "https://omlx.ai/benchmarks/performance/t8g4hcf3",
   "tg_tokens": null,
   "tps": 34.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-t8g4hcf3-ctx8192",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.3 GB at this context; duplicate traces of the same session: 6n4yg00i, nzb0ojm4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 585.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 585.0 46.4 156.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t8g4hcf3",
   "source_url": "https://omlx.ai/benchmarks/performance/t8g4hcf3",
   "tg_tokens": null,
   "tps": 46.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-t8g4hcf3-ctx16384",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.4 GB at this context; duplicate traces of the same session: 6n4yg00i, nzb0ojm4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 581.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 581.1 41.4 156.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t8g4hcf3",
   "source_url": "https://omlx.ai/benchmarks/performance/t8g4hcf3",
   "tg_tokens": null,
   "tps": 41.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-q8gm2pi0-ctx4096",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.3 GB at this context; duplicate traces of the same session: 0vsfd1gm, 7laki9ij; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 592.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "DeepSeek-V4-Flash-0731-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 592.3 38.2 156.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q8gm2pi0",
   "source_url": "https://omlx.ai/benchmarks/performance/q8gm2pi0",
   "tg_tokens": null,
   "tps": 38.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-q8gm2pi0-ctx8192",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.3 GB at this context; duplicate traces of the same session: 0vsfd1gm, 7laki9ij; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 589.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "DeepSeek-V4-Flash-0731-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 589.8 44.2 156.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q8gm2pi0",
   "source_url": "https://omlx.ai/benchmarks/performance/q8gm2pi0",
   "tg_tokens": null,
   "tps": 44.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-q8gm2pi0-ctx16384",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.4 GB at this context; duplicate traces of the same session: 0vsfd1gm, 7laki9ij; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 581.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "DeepSeek-V4-Flash-0731-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k 581.0 43.4 156.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q8gm2pi0",
   "source_url": "https://omlx.ai/benchmarks/performance/q8gm2pi0",
   "tg_tokens": null,
   "tps": 43.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-kchrd1l4-ctx1024",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 179.8 GB at this context; duplicate traces of the same session: md0zl1yt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 522.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 522.5 44.1 179.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kchrd1l4",
   "source_url": "https://omlx.ai/benchmarks/performance/kchrd1l4",
   "tg_tokens": null,
   "tps": 44.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-kchrd1l4-ctx4096",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 181.2 GB at this context; duplicate traces of the same session: md0zl1yt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 601.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 601.2 52.6 181.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kchrd1l4",
   "source_url": "https://omlx.ai/benchmarks/performance/kchrd1l4",
   "tg_tokens": null,
   "tps": 52.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vxhvdk05-ctx1024",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 177.6 GB at this context; duplicate traces of the same session: fgfgewpg; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 517.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 517.1 50.5 177.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vxhvdk05",
   "source_url": "https://omlx.ai/benchmarks/performance/vxhvdk05",
   "tg_tokens": null,
   "tps": 50.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-vxhvdk05-ctx4096",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 178.8 GB at this context; duplicate traces of the same session: fgfgewpg; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 592.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 592.5 45.8 178.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vxhvdk05",
   "source_url": "https://omlx.ai/benchmarks/performance/vxhvdk05",
   "tg_tokens": null,
   "tps": 45.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-dsc7lm8t-ctx1024",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 177.6 GB at this context; duplicate traces of the same session: 4p8suos8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 517.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 517.2 49.2 177.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dsc7lm8t",
   "source_url": "https://omlx.ai/benchmarks/performance/dsc7lm8t",
   "tg_tokens": null,
   "tps": 49.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-dsc7lm8t-ctx4096",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 178.8 GB at this context; duplicate traces of the same session: 4p8suos8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 592.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 592.7 50.5 178.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dsc7lm8t",
   "source_url": "https://omlx.ai/benchmarks/performance/dsc7lm8t",
   "tg_tokens": null,
   "tps": 50.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-arru9iif-ctx1024",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 177.6 GB at this context; duplicate traces of the same session: zy80muzk; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 528.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 528.0 40.5 177.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/arru9iif",
   "source_url": "https://omlx.ai/benchmarks/performance/arru9iif",
   "tg_tokens": null,
   "tps": 40.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-arru9iif-ctx4096",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 178.8 GB at this context; duplicate traces of the same session: zy80muzk; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 605.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 605.0 38.6 178.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/arru9iif",
   "source_url": "https://omlx.ai/benchmarks/performance/arru9iif",
   "tg_tokens": null,
   "tps": 38.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-p14lcsrp-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.5 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1419.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 1k 1,419 103.4 13.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p14lcsrp",
   "source_url": "https://omlx.ai/benchmarks/performance/p14lcsrp",
   "tg_tokens": null,
   "tps": 103.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-p14lcsrp-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1645.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 4k 1,645 113.2 13.7 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p14lcsrp",
   "source_url": "https://omlx.ai/benchmarks/performance/p14lcsrp",
   "tg_tokens": null,
   "tps": 113.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-p14lcsrp-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.8 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1604.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 8k 1,604 109.8 13.8 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p14lcsrp",
   "source_url": "https://omlx.ai/benchmarks/performance/p14lcsrp",
   "tg_tokens": null,
   "tps": 109.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-p14lcsrp-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.9 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1505.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 16k 1,505 101.5 13.9 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p14lcsrp",
   "source_url": "https://omlx.ai/benchmarks/performance/p14lcsrp",
   "tg_tokens": null,
   "tps": 101.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-p14lcsrp-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 14.2 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1244.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 32k 1,244 88.5 14.2 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p14lcsrp",
   "source_url": "https://omlx.ai/benchmarks/performance/p14lcsrp",
   "tg_tokens": null,
   "tps": 88.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-p14lcsrp-ctx65536",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 14.9 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 977.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 64k 977.0 70.5 14.9 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p14lcsrp",
   "source_url": "https://omlx.ai/benchmarks/performance/p14lcsrp",
   "tg_tokens": null,
   "tps": 70.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ycj27l1o-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 45.2 GB at this context; duplicate traces of the same session: fc4k1m9q; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1545.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "LFM2.5-24B-A2B-MLX-bf16 on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 bf16; 1k 1,545 118.1 45.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ycj27l1o",
   "source_url": "https://omlx.ai/benchmarks/performance/ycj27l1o",
   "tg_tokens": null,
   "tps": 118.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ycj27l1o-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 45.6 GB at this context; duplicate traces of the same session: fc4k1m9q; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4059.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "LFM2.5-24B-A2B-MLX-bf16 on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 bf16; 4k 4,059 116.5 45.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ycj27l1o",
   "source_url": "https://omlx.ai/benchmarks/performance/ycj27l1o",
   "tg_tokens": null,
   "tps": 116.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-j42vlp2m-ctx4096",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 157.8 GB at this context; duplicate traces of the same session: 3oiuyqbr, na0ksh21; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 646.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-V2.6-Flash-RL-MLX-4bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 646.2 43.9 157.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j42vlp2m",
   "source_url": "https://omlx.ai/benchmarks/performance/j42vlp2m",
   "tg_tokens": null,
   "tps": 43.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-j42vlp2m-ctx8192",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 158.9 GB at this context; duplicate traces of the same session: 3oiuyqbr, na0ksh21; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 644.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-V2.6-Flash-RL-MLX-4bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 644.3 49.4 158.9 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j42vlp2m",
   "source_url": "https://omlx.ai/benchmarks/performance/j42vlp2m",
   "tg_tokens": null,
   "tps": 49.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-j42vlp2m-ctx16384",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 161.1 GB at this context; duplicate traces of the same session: 3oiuyqbr, na0ksh21; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 607.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-V2.6-Flash-RL-MLX-4bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 607.0 46.0 161.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j42vlp2m",
   "source_url": "https://omlx.ai/benchmarks/performance/j42vlp2m",
   "tg_tokens": null,
   "tps": 46.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-w4suyh21-ctx1024",
   "model": "Ornith-1.0-35B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 2kans6yu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1195.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.0-35B-oQ4e on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 1,195 102.9 20.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w4suyh21",
   "source_url": "https://omlx.ai/benchmarks/performance/w4suyh21",
   "tg_tokens": null,
   "tps": 102.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-w4suyh21-ctx4096",
   "model": "Ornith-1.0-35B",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: 2kans6yu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1783.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.0-35B-oQ4e on M3 Max (40c); M3 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,783 98.8 20.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w4suyh21",
   "source_url": "https://omlx.ai/benchmarks/performance/w4suyh21",
   "tg_tokens": null,
   "tps": 98.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5h5ot8og-ctx4096",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: 94jzq61z, bopv61mn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3225.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-oQ8-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 3,225 122.8 39.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5h5ot8og",
   "source_url": "https://omlx.ai/benchmarks/performance/5h5ot8og",
   "tg_tokens": null,
   "tps": 122.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5h5ot8og-ctx8192",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.4 GB at this context; duplicate traces of the same session: 94jzq61z, bopv61mn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3150.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-oQ8-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 3,150 139.9 39.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5h5ot8og",
   "source_url": "https://omlx.ai/benchmarks/performance/5h5ot8og",
   "tg_tokens": null,
   "tps": 139.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5h5ot8og-ctx16384",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.7 GB at this context; duplicate traces of the same session: 94jzq61z, bopv61mn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2831.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-oQ8-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k 2,831 177.9 39.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5h5ot8og",
   "source_url": "https://omlx.ai/benchmarks/performance/5h5ot8og",
   "tg_tokens": null,
   "tps": 177.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-6234vpd5-ctx32768",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 9.9 GB; TTFT 84505 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 387.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 32k tokens 387.8 PP tok/s 49.6 TG tok/s 84505 TTFT (ms) 9.9 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6234vpd5",
   "source_url": "https://omlx.ai/benchmarks/performance/6234vpd5",
   "tg_tokens": null,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-bb9p85hz-ctx16384",
   "model": "Ornith-1.5-9B",
   "notes": "fp16 MTP build; oMLX community benchmark trace, single run at 16k context; peak memory 8.2 GB; TTFT 70641 ms; duplicate traces of the same session: none; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 231.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Ornith-1.5-9B-oQ5e-fp16-mtp on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 5bit; 16k tokens 231.9 PP tok/s 23.2 TG tok/s 70641 TTFT (ms) 8.2 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bb9p85hz",
   "source_url": "https://omlx.ai/benchmarks/performance/bb9p85hz",
   "tg_tokens": null,
   "tps": 23.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-goom5xal-ctx16384",
   "model": "Ornith-1.5-9B",
   "notes": "fp16 MTP build; oMLX community benchmark trace, single run at 16k context; peak memory 10.8 GB; TTFT 71252 ms; duplicate traces of the same session: none; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 229.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Ornith-1.5-9B-oQ5e-fp16-mtp on M1 Pro (14c); M1 Pro (14c) \u00b7 16 GB \u00b7 5bit; 16k tokens 229.9 PP tok/s 18.4 TG tok/s 71252 TTFT (ms) 10.8 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/goom5xal",
   "source_url": "https://omlx.ai/benchmarks/performance/goom5xal",
   "tg_tokens": null,
   "tps": 18.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-dxe8apte-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: m9th4aic, ns2x5jh4, zhpd3tka; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 648.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 1k 648.6 93.6 37.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dxe8apte",
   "source_url": "https://omlx.ai/benchmarks/performance/dxe8apte",
   "tg_tokens": null,
   "tps": 93.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-dxe8apte-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 38.0 GB at this context; duplicate traces of the same session: m9th4aic, ns2x5jh4, zhpd3tka; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2903.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 4k 2,903 168.2 38.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dxe8apte",
   "source_url": "https://omlx.ai/benchmarks/performance/dxe8apte",
   "tg_tokens": null,
   "tps": 168.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-dxe8apte-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 40.6 GB at this context; duplicate traces of the same session: m9th4aic, ns2x5jh4, zhpd3tka; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3600.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 128k 3,600 120.0 40.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dxe8apte",
   "source_url": "https://omlx.ai/benchmarks/performance/dxe8apte",
   "tg_tokens": null,
   "tps": 120.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-dxe8apte-ctx199680",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 42.4 GB at this context; duplicate traces of the same session: m9th4aic, ns2x5jh4, zhpd3tka; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2453.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 195k 2,453 59.6 42.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dxe8apte",
   "source_url": "https://omlx.ai/benchmarks/performance/dxe8apte",
   "tg_tokens": null,
   "tps": 59.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tt3uhazp-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: ls4m6k4j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 849.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 849.0 32.2 33.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tt3uhazp",
   "source_url": "https://omlx.ai/benchmarks/performance/tt3uhazp",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tt3uhazp-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: ls4m6k4j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 794.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 794.5 29.0 34.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tt3uhazp",
   "source_url": "https://omlx.ai/benchmarks/performance/tt3uhazp",
   "tg_tokens": null,
   "tps": 29.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tt3uhazp-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 35.1 GB at this context; duplicate traces of the same session: ls4m6k4j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 696.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 696.8 26.1 35.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tt3uhazp",
   "source_url": "https://omlx.ai/benchmarks/performance/tt3uhazp",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tt3uhazp-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 35.6 GB at this context; duplicate traces of the same session: ls4m6k4j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 641.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 641.3 35.6 35.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tt3uhazp",
   "source_url": "https://omlx.ai/benchmarks/performance/tt3uhazp",
   "tg_tokens": null,
   "tps": 35.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tt3uhazp-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 36.6 GB at this context; duplicate traces of the same session: ls4m6k4j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 619.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 619.0 32.3 36.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tt3uhazp",
   "source_url": "https://omlx.ai/benchmarks/performance/tt3uhazp",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tt3uhazp-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 38.6 GB at this context; duplicate traces of the same session: ls4m6k4j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 529.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 529.0 22.7 38.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tt3uhazp",
   "source_url": "https://omlx.ai/benchmarks/performance/tt3uhazp",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tt3uhazp-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 42.6 GB at this context; duplicate traces of the same session: ls4m6k4j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 370.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 370.7 20.7 42.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tt3uhazp",
   "source_url": "https://omlx.ai/benchmarks/performance/tt3uhazp",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tt3uhazp-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 46.8 GB at this context; duplicate traces of the same session: ls4m6k4j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 279.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 195k 279.5 13.4 46.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tt3uhazp",
   "source_url": "https://omlx.ai/benchmarks/performance/tt3uhazp",
   "tg_tokens": null,
   "tps": 13.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-pmhg5fgg-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 72.4 GB at this context; duplicate traces of the same session: 0h8a3eby, 3ytcikbv, duk7twce, r2t8qtue, t9zq5wkc, xo7nial6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 699.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 1k 699.6 35.3 72.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pmhg5fgg",
   "source_url": "https://omlx.ai/benchmarks/performance/pmhg5fgg",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-pmhg5fgg-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 74.8 GB at this context; duplicate traces of the same session: 0h8a3eby, 3ytcikbv, duk7twce, r2t8qtue, t9zq5wkc, xo7nial6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 794.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 4k 794.9 46.0 74.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pmhg5fgg",
   "source_url": "https://omlx.ai/benchmarks/performance/pmhg5fgg",
   "tg_tokens": null,
   "tps": 46.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-pmhg5fgg-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 75.1 GB at this context; duplicate traces of the same session: 0h8a3eby, 3ytcikbv, duk7twce, r2t8qtue, t9zq5wkc, xo7nial6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 851.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 8k 851.5 53.5 75.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pmhg5fgg",
   "source_url": "https://omlx.ai/benchmarks/performance/pmhg5fgg",
   "tg_tokens": null,
   "tps": 53.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-pmhg5fgg-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 75.4 GB at this context; duplicate traces of the same session: 0h8a3eby, 3ytcikbv, duk7twce, r2t8qtue, t9zq5wkc, xo7nial6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 843.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 16k 843.5 40.4 75.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pmhg5fgg",
   "source_url": "https://omlx.ai/benchmarks/performance/pmhg5fgg",
   "tg_tokens": null,
   "tps": 40.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-pmhg5fgg-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 77.1 GB at this context; duplicate traces of the same session: 0h8a3eby, 3ytcikbv, duk7twce, r2t8qtue, t9zq5wkc, xo7nial6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 827.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 64k 827.6 23.4 77.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pmhg5fgg",
   "source_url": "https://omlx.ai/benchmarks/performance/pmhg5fgg",
   "tg_tokens": null,
   "tps": 23.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-pmhg5fgg-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 79.4 GB at this context; duplicate traces of the same session: 0h8a3eby, 3ytcikbv, duk7twce, r2t8qtue, t9zq5wkc, xo7nial6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 802.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 128k 802.3 43.6 79.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pmhg5fgg",
   "source_url": "https://omlx.ai/benchmarks/performance/pmhg5fgg",
   "tg_tokens": null,
   "tps": 43.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-pmhg5fgg-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 82.4 GB at this context; duplicate traces of the same session: 0h8a3eby, 3ytcikbv, duk7twce, r2t8qtue, t9zq5wkc, xo7nial6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 792.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 195k 792.2 29.5 82.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pmhg5fgg",
   "source_url": "https://omlx.ai/benchmarks/performance/pmhg5fgg",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-tmosg15s-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.6 GB at this context; duplicate traces of the same session: 274hptph, 2yiu6m4i, 4ynojwba, me613suz, n0712nv0, tenlrs2y; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 801.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 801.9 85.2 142.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tmosg15s",
   "source_url": "https://omlx.ai/benchmarks/performance/tmosg15s",
   "tg_tokens": null,
   "tps": 85.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-tmosg15s-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 144.5 GB at this context; duplicate traces of the same session: 274hptph, 2yiu6m4i, 4ynojwba, me613suz, n0712nv0, tenlrs2y; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4147.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 8k 4,147 60.4 144.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tmosg15s",
   "source_url": "https://omlx.ai/benchmarks/performance/tmosg15s",
   "tg_tokens": null,
   "tps": 60.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-tmosg15s-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.3 GB at this context; duplicate traces of the same session: 274hptph, 2yiu6m4i, 4ynojwba, me613suz, n0712nv0, tenlrs2y; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4596.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 16k 4,596 113.8 145.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tmosg15s",
   "source_url": "https://omlx.ai/benchmarks/performance/tmosg15s",
   "tg_tokens": null,
   "tps": 113.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-tmosg15s-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.9 GB at this context; duplicate traces of the same session: 274hptph, 2yiu6m4i, 4ynojwba, me613suz, n0712nv0, tenlrs2y; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4817.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 32k 4,817 104.4 145.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tmosg15s",
   "source_url": "https://omlx.ai/benchmarks/performance/tmosg15s",
   "tg_tokens": null,
   "tps": 104.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-tmosg15s-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 147.0 GB at this context; duplicate traces of the same session: 274hptph, 2yiu6m4i, 4ynojwba, me613suz, n0712nv0, tenlrs2y; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4749.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 64k 4,749 88.0 147.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tmosg15s",
   "source_url": "https://omlx.ai/benchmarks/performance/tmosg15s",
   "tg_tokens": null,
   "tps": 88.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-tmosg15s-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 149.3 GB at this context; duplicate traces of the same session: 274hptph, 2yiu6m4i, 4ynojwba, me613suz, n0712nv0, tenlrs2y; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4668.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 128k 4,668 101.5 149.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tmosg15s",
   "source_url": "https://omlx.ai/benchmarks/performance/tmosg15s",
   "tg_tokens": null,
   "tps": 101.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-tmosg15s-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 152.1 GB at this context; duplicate traces of the same session: 274hptph, 2yiu6m4i, 4ynojwba, me613suz, n0712nv0, tenlrs2y; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4532.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 195k 4,532 91.4 152.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tmosg15s",
   "source_url": "https://omlx.ai/benchmarks/performance/tmosg15s",
   "tg_tokens": null,
   "tps": 91.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-j0y31hrd-ctx1024",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: 14cl0s7u, 1mrucgjl, 85ecym2y, i6t0ifmv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2199.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 2,199 67.9 26.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j0y31hrd",
   "source_url": "https://omlx.ai/benchmarks/performance/j0y31hrd",
   "tg_tokens": null,
   "tps": 67.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-j0y31hrd-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 27.2 GB at this context; duplicate traces of the same session: 14cl0s7u, 1mrucgjl, 85ecym2y, i6t0ifmv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2141.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 2,141 55.5 27.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j0y31hrd",
   "source_url": "https://omlx.ai/benchmarks/performance/j0y31hrd",
   "tg_tokens": null,
   "tps": 55.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-j0y31hrd-ctx8192",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: 14cl0s7u, 1mrucgjl, 85ecym2y, i6t0ifmv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1996.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 1,996 35.4 27.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j0y31hrd",
   "source_url": "https://omlx.ai/benchmarks/performance/j0y31hrd",
   "tg_tokens": null,
   "tps": 35.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-j0y31hrd-ctx16384",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 29.0 GB at this context; duplicate traces of the same session: 14cl0s7u, 1mrucgjl, 85ecym2y, i6t0ifmv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1902.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 16k 1,902 43.9 29.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j0y31hrd",
   "source_url": "https://omlx.ai/benchmarks/performance/j0y31hrd",
   "tg_tokens": null,
   "tps": 43.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-j0y31hrd-ctx32768",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 30.9 GB at this context; duplicate traces of the same session: 14cl0s7u, 1mrucgjl, 85ecym2y, i6t0ifmv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1623.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 32k 1,623 42.5 30.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j0y31hrd",
   "source_url": "https://omlx.ai/benchmarks/performance/j0y31hrd",
   "tg_tokens": null,
   "tps": 42.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-013afild-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: iop2ywp1, j7i4ukvl, myh820ky, vva10caq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2054.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26b-a4b-it-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 2,054 49.6 25.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/013afild",
   "source_url": "https://omlx.ai/benchmarks/performance/013afild",
   "tg_tokens": null,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-013afild-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: iop2ywp1, j7i4ukvl, myh820ky, vva10caq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2129.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26b-a4b-it-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 2,129 49.4 26.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/013afild",
   "source_url": "https://omlx.ai/benchmarks/performance/013afild",
   "tg_tokens": null,
   "tps": 49.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-013afild-ctx8192",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: iop2ywp1, j7i4ukvl, myh820ky, vva10caq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2105.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26b-a4b-it-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 2,105 47.5 26.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/013afild",
   "source_url": "https://omlx.ai/benchmarks/performance/013afild",
   "tg_tokens": null,
   "tps": 47.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-013afild-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: iop2ywp1, j7i4ukvl, myh820ky, vva10caq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2028.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26b-a4b-it-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 16k 2,028 43.6 26.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/013afild",
   "source_url": "https://omlx.ai/benchmarks/performance/013afild",
   "tg_tokens": null,
   "tps": 43.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-013afild-ctx32768",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: iop2ywp1, j7i4ukvl, myh820ky, vva10caq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1773.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26b-a4b-it-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 32k 1,773 37.1 27.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/013afild",
   "source_url": "https://omlx.ai/benchmarks/performance/013afild",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-1q9dhn5i-ctx4096",
   "model": "gemma-4-E4B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: tw640qmk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2477.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-qat-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 2,477 89.9 6.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1q9dhn5i",
   "source_url": "https://omlx.ai/benchmarks/performance/1q9dhn5i",
   "tg_tokens": null,
   "tps": 89.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-1q9dhn5i-ctx16384",
   "model": "gemma-4-E4B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: tw640qmk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2337.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-qat-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 2,337 78.4 7.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1q9dhn5i",
   "source_url": "https://omlx.ai/benchmarks/performance/1q9dhn5i",
   "tg_tokens": null,
   "tps": 78.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-hih3amas-ctx16384",
   "model": "Devstral-Small-2-24B-Instruct-2512",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 17.0 GB; TTFT 111390 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 147.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Devstral-Small-2-24B-Instruct-2512-4bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k tokens 147.1 PP tok/s 18.6 TG tok/s 111390 TTFT (ms) 17.0 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hih3amas",
   "source_url": "https://omlx.ai/benchmarks/performance/hih3amas",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-wa19wmgy-ctx1024",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 177.6 GB at this context; duplicate traces of the same session: 5d93e1cz, c1soiu1x, hfen596x; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 517.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 517.5 47.3 177.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wa19wmgy",
   "source_url": "https://omlx.ai/benchmarks/performance/wa19wmgy",
   "tg_tokens": null,
   "tps": 47.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-wa19wmgy-ctx4096",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 178.8 GB at this context; duplicate traces of the same session: 5d93e1cz, c1soiu1x, hfen596x; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 592.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 592.3 48.5 178.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wa19wmgy",
   "source_url": "https://omlx.ai/benchmarks/performance/wa19wmgy",
   "tg_tokens": null,
   "tps": 48.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-wa19wmgy-ctx8192",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 179.1 GB at this context; duplicate traces of the same session: 5d93e1cz, c1soiu1x, hfen596x; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 547.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 547.6 60.3 179.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wa19wmgy",
   "source_url": "https://omlx.ai/benchmarks/performance/wa19wmgy",
   "tg_tokens": null,
   "tps": 60.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-wa19wmgy-ctx16384",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 179.2 GB at this context; duplicate traces of the same session: 5d93e1cz, c1soiu1x, hfen596x; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 537.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 537.8 89.7 179.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wa19wmgy",
   "source_url": "https://omlx.ai/benchmarks/performance/wa19wmgy",
   "tg_tokens": null,
   "tps": 89.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-wa19wmgy-ctx32768",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 179.4 GB at this context; duplicate traces of the same session: 5d93e1cz, c1soiu1x, hfen596x; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 530.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 530.3 67.5 179.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wa19wmgy",
   "source_url": "https://omlx.ai/benchmarks/performance/wa19wmgy",
   "tg_tokens": null,
   "tps": 67.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-wa19wmgy-ctx65536",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 179.8 GB at this context; duplicate traces of the same session: 5d93e1cz, c1soiu1x, hfen596x; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 522.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 522.4 54.2 179.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wa19wmgy",
   "source_url": "https://omlx.ai/benchmarks/performance/wa19wmgy",
   "tg_tokens": null,
   "tps": 54.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-wa19wmgy-ctx131072",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 180.8 GB at this context; duplicate traces of the same session: 5d93e1cz, c1soiu1x, hfen596x; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 505.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 505.9 95.5 180.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wa19wmgy",
   "source_url": "https://omlx.ai/benchmarks/performance/wa19wmgy",
   "tg_tokens": null,
   "tps": 95.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-wa19wmgy-ctx199680",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 181.8 GB at this context; duplicate traces of the same session: 5d93e1cz, c1soiu1x, hfen596x; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 480.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 195k 480.8 42.6 181.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wa19wmgy",
   "source_url": "https://omlx.ai/benchmarks/performance/wa19wmgy",
   "tg_tokens": null,
   "tps": 42.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-9bfh5v8a-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.5 GB at this context; duplicate traces of the same session: 1ym8if3q, 211y3fix, e05ii00s, r9ycf05l; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1419.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 1k 1,419 103.4 13.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9bfh5v8a",
   "source_url": "https://omlx.ai/benchmarks/performance/9bfh5v8a",
   "tg_tokens": null,
   "tps": 103.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-9bfh5v8a-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: 1ym8if3q, 211y3fix, e05ii00s, r9ycf05l; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1645.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 4k 1,645 113.2 13.7 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9bfh5v8a",
   "source_url": "https://omlx.ai/benchmarks/performance/9bfh5v8a",
   "tg_tokens": null,
   "tps": 113.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-9bfh5v8a-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.8 GB at this context; duplicate traces of the same session: 1ym8if3q, 211y3fix, e05ii00s, r9ycf05l; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1604.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 8k 1,604 109.8 13.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9bfh5v8a",
   "source_url": "https://omlx.ai/benchmarks/performance/9bfh5v8a",
   "tg_tokens": null,
   "tps": 109.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-9bfh5v8a-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.9 GB at this context; duplicate traces of the same session: 1ym8if3q, 211y3fix, e05ii00s, r9ycf05l; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1505.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 16k 1,505 101.5 13.9 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9bfh5v8a",
   "source_url": "https://omlx.ai/benchmarks/performance/9bfh5v8a",
   "tg_tokens": null,
   "tps": 101.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-9bfh5v8a-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 14.2 GB at this context; duplicate traces of the same session: 1ym8if3q, 211y3fix, e05ii00s, r9ycf05l; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1244.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 32k 1,244 88.5 14.2 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9bfh5v8a",
   "source_url": "https://omlx.ai/benchmarks/performance/9bfh5v8a",
   "tg_tokens": null,
   "tps": 88.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-9bfh5v8a-ctx65536",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 14.9 GB at this context; duplicate traces of the same session: 1ym8if3q, 211y3fix, e05ii00s, r9ycf05l; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 977.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 64k 977.0 70.5 14.9 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9bfh5v8a",
   "source_url": "https://omlx.ai/benchmarks/performance/9bfh5v8a",
   "tg_tokens": null,
   "tps": 70.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-dg5ji0pw-ctx1024",
   "model": "Nemotron-Cascade-2-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: h02ww9di; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1109.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Nemotron-Cascade-2-30B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 1,109 59.4 18.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dg5ji0pw",
   "source_url": "https://omlx.ai/benchmarks/performance/dg5ji0pw",
   "tg_tokens": null,
   "tps": 59.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-dg5ji0pw-ctx4096",
   "model": "Nemotron-Cascade-2-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: h02ww9di; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1239.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Nemotron-Cascade-2-30B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,239 59.5 19.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dg5ji0pw",
   "source_url": "https://omlx.ai/benchmarks/performance/dg5ji0pw",
   "tg_tokens": null,
   "tps": 59.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wowb1hyg-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 36.4 GB; TTFT 3521 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4653.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k tokens 4,653 PP tok/s 92.8 TG tok/s 3521 TTFT (ms) 36.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wowb1hyg",
   "source_url": "https://omlx.ai/benchmarks/performance/wowb1hyg",
   "tg_tokens": null,
   "tps": 92.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-821rrb59-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 36.4 GB; TTFT 3476 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4713.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k tokens 4,713 PP tok/s 90.9 TG tok/s 3476 TTFT (ms) 36.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/821rrb59",
   "source_url": "https://omlx.ai/benchmarks/performance/821rrb59",
   "tg_tokens": null,
   "tps": 90.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-go8hiqvi-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 36.4 GB; TTFT 3445 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4757.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k tokens 4,757 PP tok/s 92.9 TG tok/s 3445 TTFT (ms) 36.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/go8hiqvi",
   "source_url": "https://omlx.ai/benchmarks/performance/go8hiqvi",
   "tg_tokens": null,
   "tps": 92.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9r0llkgn-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 36.4 GB; TTFT 3695 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4434.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k tokens 4,434 PP tok/s 91.8 TG tok/s 3695 TTFT (ms) 36.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9r0llkgn",
   "source_url": "https://omlx.ai/benchmarks/performance/9r0llkgn",
   "tg_tokens": null,
   "tps": 91.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-a6batd61-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 9r3f6sqd; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1329.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 1,329 49.4 21.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a6batd61",
   "source_url": "https://omlx.ai/benchmarks/performance/a6batd61",
   "tg_tokens": null,
   "tps": 49.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-a6batd61-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 9r3f6sqd; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1860.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,860 42.5 21.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a6batd61",
   "source_url": "https://omlx.ai/benchmarks/performance/a6batd61",
   "tg_tokens": null,
   "tps": 42.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xsq0zefl-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: t6rlok9p; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 334.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 334.5 43.4 19.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xsq0zefl",
   "source_url": "https://omlx.ai/benchmarks/performance/xsq0zefl",
   "tg_tokens": null,
   "tps": 43.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xsq0zefl-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: t6rlok9p; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 354.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 354.2 41.1 22.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xsq0zefl",
   "source_url": "https://omlx.ai/benchmarks/performance/xsq0zefl",
   "tg_tokens": null,
   "tps": 41.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-b254h1hf-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: cox7sgnn; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 334.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 334.3 46.0 19.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b254h1hf",
   "source_url": "https://omlx.ai/benchmarks/performance/b254h1hf",
   "tg_tokens": null,
   "tps": 46.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-b254h1hf-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: cox7sgnn; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 355.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 355.8 41.8 22.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b254h1hf",
   "source_url": "https://omlx.ai/benchmarks/performance/b254h1hf",
   "tg_tokens": null,
   "tps": 41.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-yu0jo7ie-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: am5gr2uh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 486.1,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.8-27B-MLX-5bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 5bit; 4k 486.1 17.6 24.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yu0jo7ie",
   "source_url": "https://omlx.ai/benchmarks/performance/yu0jo7ie",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-yu0jo7ie-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: am5gr2uh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 463.2,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.8-27B-MLX-5bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 5bit; 16k 463.2 18.8 24.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yu0jo7ie",
   "source_url": "https://omlx.ai/benchmarks/performance/yu0jo7ie",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-sjud8hje-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: vsu6jzph; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 441.4,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.8-27B-MLX-5bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 5bit; 4k 441.4 16.2 24.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sjud8hje",
   "source_url": "https://omlx.ai/benchmarks/performance/sjud8hje",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-sjud8hje-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: vsu6jzph; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 455.3,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.8-27B-MLX-5bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 5bit; 16k 455.3 16.7 24.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sjud8hje",
   "source_url": "https://omlx.ai/benchmarks/performance/sjud8hje",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-gthh537n-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 15.4 GB at this context; duplicate traces of the same session: aq6ylqv0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 456.4,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-27B-oQ3e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 24 GB \u00b7 3bit; 1k 456.4 29.1 15.4 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gthh537n",
   "source_url": "https://omlx.ai/benchmarks/performance/gthh537n",
   "tg_tokens": null,
   "tps": 29.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-gthh537n-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.2 GB at this context; duplicate traces of the same session: aq6ylqv0; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 491.4,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-27B-oQ3e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 24 GB \u00b7 3bit; 4k 491.4 34.1 16.2 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gthh537n",
   "source_url": "https://omlx.ai/benchmarks/performance/gthh537n",
   "tg_tokens": null,
   "tps": 34.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-le5ee91i-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 8oh4q82s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 838.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 838.3 67.0 17.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/le5ee91i",
   "source_url": "https://omlx.ai/benchmarks/performance/le5ee91i",
   "tg_tokens": null,
   "tps": 67.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-le5ee91i-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: 8oh4q82s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 955.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 955.7 66.1 18.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/le5ee91i",
   "source_url": "https://omlx.ai/benchmarks/performance/le5ee91i",
   "tg_tokens": null,
   "tps": 66.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-16yqu9pz-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: 1xum0rw6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 923.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 923.7 29.6 18.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/16yqu9pz",
   "source_url": "https://omlx.ai/benchmarks/performance/16yqu9pz",
   "tg_tokens": null,
   "tps": 29.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-16yqu9pz-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 1xum0rw6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 789.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 789.6 56.0 19.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/16yqu9pz",
   "source_url": "https://omlx.ai/benchmarks/performance/16yqu9pz",
   "tg_tokens": null,
   "tps": 56.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-p8mj2w3n-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: oeb253c0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 647.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 647.8 80.7 18.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p8mj2w3n",
   "source_url": "https://omlx.ai/benchmarks/performance/p8mj2w3n",
   "tg_tokens": null,
   "tps": 80.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-p8mj2w3n-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: oeb253c0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 578.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 578.6 51.2 19.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p8mj2w3n",
   "source_url": "https://omlx.ai/benchmarks/performance/p8mj2w3n",
   "tg_tokens": null,
   "tps": 51.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ucgkkb0w-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.0 GB at this context; duplicate traces of the same session: zockd1cn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 650.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 650.6 50.2 35.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ucgkkb0w",
   "source_url": "https://omlx.ai/benchmarks/performance/ucgkkb0w",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ucgkkb0w-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.7 GB at this context; duplicate traces of the same session: zockd1cn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 652.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 652.2 44.8 35.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ucgkkb0w",
   "source_url": "https://omlx.ai/benchmarks/performance/ucgkkb0w",
   "tg_tokens": null,
   "tps": 44.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0lg42lji-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 581.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 581.6 42.2 35.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0lg42lji",
   "source_url": "https://omlx.ai/benchmarks/performance/0lg42lji",
   "tg_tokens": null,
   "tps": 42.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0lg42lji-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 606.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 606.4 40.6 35.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0lg42lji",
   "source_url": "https://omlx.ai/benchmarks/performance/0lg42lji",
   "tg_tokens": null,
   "tps": 40.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4nigbs42-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 75.3 GB at this context; duplicate traces of the same session: 4bf4zbea, gm5a2dwh, s7g6x0bj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 845.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 845.6 38.9 75.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4nigbs42",
   "source_url": "https://omlx.ai/benchmarks/performance/4nigbs42",
   "tg_tokens": null,
   "tps": 38.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4nigbs42-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 75.5 GB at this context; duplicate traces of the same session: 4bf4zbea, gm5a2dwh, s7g6x0bj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1541.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,541 39.4 75.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4nigbs42",
   "source_url": "https://omlx.ai/benchmarks/performance/4nigbs42",
   "tg_tokens": null,
   "tps": 39.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4nigbs42-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 76.0 GB at this context; duplicate traces of the same session: 4bf4zbea, gm5a2dwh, s7g6x0bj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1861.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,861 41.3 76.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4nigbs42",
   "source_url": "https://omlx.ai/benchmarks/performance/4nigbs42",
   "tg_tokens": null,
   "tps": 41.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4nigbs42-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 84.6 GB at this context; duplicate traces of the same session: 4bf4zbea, gm5a2dwh, s7g6x0bj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1944.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 1,944 39.8 84.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4nigbs42",
   "source_url": "https://omlx.ai/benchmarks/performance/4nigbs42",
   "tg_tokens": null,
   "tps": 39.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-29610xqi-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 75.3 GB; TTFT 1132 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 904.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 904.7 PP tok/s 44.1 TG tok/s 1132 TTFT (ms) 75.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/29610xqi",
   "source_url": "https://omlx.ai/benchmarks/performance/29610xqi",
   "tg_tokens": null,
   "tps": 44.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-e4qmk212-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 72.7 GB; TTFT 7403 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2213.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k tokens 2,213 PP tok/s 67.6 TG tok/s 7403 TTFT (ms) 72.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e4qmk212",
   "source_url": "https://omlx.ai/benchmarks/performance/e4qmk212",
   "tg_tokens": null,
   "tps": 67.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-sul1a9ib-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: 6mlzyn33; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2426.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,426 56.2 71.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sul1a9ib",
   "source_url": "https://omlx.ai/benchmarks/performance/sul1a9ib",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-sul1a9ib-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.9 GB at this context; duplicate traces of the same session: 6mlzyn33; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2703.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,703 64.3 72.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sul1a9ib",
   "source_url": "https://omlx.ai/benchmarks/performance/sul1a9ib",
   "tg_tokens": null,
   "tps": 64.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-h0iy4qa1-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 71.1 GB; TTFT 2112 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1939.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 1,939 PP tok/s 25.7 TG tok/s 2112 TTFT (ms) 71.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h0iy4qa1",
   "source_url": "https://omlx.ai/benchmarks/performance/h0iy4qa1",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-k1vj2wrg-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 71.8 GB; TTFT 5954 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1376.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k tokens 1,376 PP tok/s 60.3 TG tok/s 5954 TTFT (ms) 71.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k1vj2wrg",
   "source_url": "https://omlx.ai/benchmarks/performance/k1vj2wrg",
   "tg_tokens": null,
   "tps": 60.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-m0xneq10-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 73.3 GB; TTFT 14547 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2253.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k tokens 2,253 PP tok/s 67.3 TG tok/s 14547 TTFT (ms) 73.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m0xneq10",
   "source_url": "https://omlx.ai/benchmarks/performance/m0xneq10",
   "tg_tokens": null,
   "tps": 67.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nq1fshix-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: u8u75n7h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2259.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,259 58.7 71.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nq1fshix",
   "source_url": "https://omlx.ai/benchmarks/performance/nq1fshix",
   "tg_tokens": null,
   "tps": 58.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nq1fshix-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.9 GB at this context; duplicate traces of the same session: u8u75n7h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2366.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,366 55.5 72.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nq1fshix",
   "source_url": "https://omlx.ai/benchmarks/performance/nq1fshix",
   "tg_tokens": null,
   "tps": 55.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jhrgvl4m-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.7 GB; TTFT 910 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1125.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,125 PP tok/s 68.5 TG tok/s 910 TTFT (ms) 70.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jhrgvl4m",
   "source_url": "https://omlx.ai/benchmarks/performance/jhrgvl4m",
   "tg_tokens": null,
   "tps": 68.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-v99nhkzs-ctx4096",
   "model": "gemma-4-E2B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: zuz1gcwc; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4200.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E2B-it-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 16 GB \u00b7 4bit; 4k 4,200 244.9 4.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v99nhkzs",
   "source_url": "https://omlx.ai/benchmarks/performance/v99nhkzs",
   "tg_tokens": null,
   "tps": 244.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-v99nhkzs-ctx16384",
   "model": "gemma-4-E2B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.3 GB at this context; duplicate traces of the same session: zuz1gcwc; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6511.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E2B-it-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 16 GB \u00b7 4bit; 16k 6,511 143.7 4.3 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v99nhkzs",
   "source_url": "https://omlx.ai/benchmarks/performance/v99nhkzs",
   "tg_tokens": null,
   "tps": 143.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-nycf829q-ctx4096",
   "model": "gemma-4-E2B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: bd3i6dje; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 7866.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E2B-it-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 16 GB \u00b7 4bit; 4k 7,866 224.7 4.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nycf829q",
   "source_url": "https://omlx.ai/benchmarks/performance/nycf829q",
   "tg_tokens": null,
   "tps": 224.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-nycf829q-ctx16384",
   "model": "gemma-4-E2B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.3 GB at this context; duplicate traces of the same session: bd3i6dje; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 7256.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E2B-it-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 16 GB \u00b7 4bit; 16k 7,256 229.0 4.3 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nycf829q",
   "source_url": "https://omlx.ai/benchmarks/performance/nycf829q",
   "tg_tokens": null,
   "tps": 229.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-cq131agb-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: euzz2g7q; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2364.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 24 GB \u00b7 4bit; 1k 2,364 121.5 5.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cq131agb",
   "source_url": "https://omlx.ai/benchmarks/performance/cq131agb",
   "tg_tokens": null,
   "tps": 121.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-cq131agb-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: euzz2g7q; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3393.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 24 GB \u00b7 4bit; 4k 3,393 60.1 5.9 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cq131agb",
   "source_url": "https://omlx.ai/benchmarks/performance/cq131agb",
   "tg_tokens": null,
   "tps": 60.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jknsrnry-ctx1024",
   "model": "medgemma-27b-it",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 54.2 GB; TTFT 1249 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 820.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "medgemma-27b-it-bf16 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 bf16; 1k tokens 820.0 PP tok/s 8.5 TG tok/s 1249 TTFT (ms) 54.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jknsrnry",
   "source_url": "https://omlx.ai/benchmarks/performance/jknsrnry",
   "tg_tokens": null,
   "tps": 8.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-y109od7f-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 5pr0o3dg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 791.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mlx-community--Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 791.0 18.7 28.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y109od7f",
   "source_url": "https://omlx.ai/benchmarks/performance/y109od7f",
   "tg_tokens": null,
   "tps": 18.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-y109od7f-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: 5pr0o3dg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 931.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "mlx-community--Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 931.4 18.5 29.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y109od7f",
   "source_url": "https://omlx.ai/benchmarks/performance/y109od7f",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-btpf6iu7-ctx1024",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 177.6 GB at this context; duplicate traces of the same session: bxeiv837, n97ds73w, pubrzpur; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 517.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 517.5 47.3 177.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/btpf6iu7",
   "source_url": "https://omlx.ai/benchmarks/performance/btpf6iu7",
   "tg_tokens": null,
   "tps": 47.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-btpf6iu7-ctx4096",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 178.8 GB at this context; duplicate traces of the same session: bxeiv837, n97ds73w, pubrzpur; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 592.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 592.3 48.5 178.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/btpf6iu7",
   "source_url": "https://omlx.ai/benchmarks/performance/btpf6iu7",
   "tg_tokens": null,
   "tps": 48.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-btpf6iu7-ctx8192",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 179.1 GB at this context; duplicate traces of the same session: bxeiv837, n97ds73w, pubrzpur; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 547.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 547.6 60.3 179.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/btpf6iu7",
   "source_url": "https://omlx.ai/benchmarks/performance/btpf6iu7",
   "tg_tokens": null,
   "tps": 60.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-btpf6iu7-ctx16384",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 179.2 GB at this context; duplicate traces of the same session: bxeiv837, n97ds73w, pubrzpur; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 537.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 537.8 89.7 179.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/btpf6iu7",
   "source_url": "https://omlx.ai/benchmarks/performance/btpf6iu7",
   "tg_tokens": null,
   "tps": 89.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-btpf6iu7-ctx32768",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 179.4 GB at this context; duplicate traces of the same session: bxeiv837, n97ds73w, pubrzpur; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 530.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 530.3 67.5 179.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/btpf6iu7",
   "source_url": "https://omlx.ai/benchmarks/performance/btpf6iu7",
   "tg_tokens": null,
   "tps": 67.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-btpf6iu7-ctx65536",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 179.8 GB at this context; duplicate traces of the same session: bxeiv837, n97ds73w, pubrzpur; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 522.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 522.4 54.2 179.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/btpf6iu7",
   "source_url": "https://omlx.ai/benchmarks/performance/btpf6iu7",
   "tg_tokens": null,
   "tps": 54.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-btpf6iu7-ctx131072",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 180.8 GB at this context; duplicate traces of the same session: bxeiv837, n97ds73w, pubrzpur; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 505.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 505.9 95.5 180.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/btpf6iu7",
   "source_url": "https://omlx.ai/benchmarks/performance/btpf6iu7",
   "tg_tokens": null,
   "tps": 95.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-btpf6iu7-ctx199680",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "DFlash2 speculative decoding; oMLX community benchmark trace; peak memory 181.8 GB at this context; duplicate traces of the same session: bxeiv837, n97ds73w, pubrzpur; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 480.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 195k 480.8 42.6 181.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/btpf6iu7",
   "source_url": "https://omlx.ai/benchmarks/performance/btpf6iu7",
   "tg_tokens": null,
   "tps": 42.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rmuzlm0n-ctx4096",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 30zi8yqq, sawqp8hb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2404.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 2,404 144.3 19.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rmuzlm0n",
   "source_url": "https://omlx.ai/benchmarks/performance/rmuzlm0n",
   "tg_tokens": null,
   "tps": 144.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rmuzlm0n-ctx8192",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 30zi8yqq, sawqp8hb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2604.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 2,604 143.2 19.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rmuzlm0n",
   "source_url": "https://omlx.ai/benchmarks/performance/rmuzlm0n",
   "tg_tokens": null,
   "tps": 143.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rmuzlm0n-ctx16384",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 30zi8yqq, sawqp8hb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2499.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 2,499 139.4 19.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rmuzlm0n",
   "source_url": "https://omlx.ai/benchmarks/performance/rmuzlm0n",
   "tg_tokens": null,
   "tps": 139.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-z02he08l-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 67.5 GB at this context; duplicate traces of the same session: 1nyf5s4g, ic0zi7bj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3260.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Ornith-1.5-35B-A3B-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 3,260 65.7 67.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z02he08l",
   "source_url": "https://omlx.ai/benchmarks/performance/z02he08l",
   "tg_tokens": null,
   "tps": 65.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-z02he08l-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 67.6 GB at this context; duplicate traces of the same session: 1nyf5s4g, ic0zi7bj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3306.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Ornith-1.5-35B-A3B-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 8k 3,306 65.9 67.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z02he08l",
   "source_url": "https://omlx.ai/benchmarks/performance/z02he08l",
   "tg_tokens": null,
   "tps": 65.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-z02he08l-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 67.8 GB at this context; duplicate traces of the same session: 1nyf5s4g, ic0zi7bj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2979.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Ornith-1.5-35B-A3B-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 16k 2,979 63.6 67.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z02he08l",
   "source_url": "https://omlx.ai/benchmarks/performance/z02he08l",
   "tg_tokens": null,
   "tps": 63.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-tq01wpot-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: 21gl3pgw; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 494.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 494.4 17.6 17.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tq01wpot",
   "source_url": "https://omlx.ai/benchmarks/performance/tq01wpot",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-tq01wpot-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 21gl3pgw; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 476.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 476.3 16.8 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tq01wpot",
   "source_url": "https://omlx.ai/benchmarks/performance/tq01wpot",
   "tg_tokens": null,
   "tps": 16.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-dkkh8sv5-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 30.5 GB at this context; duplicate traces of the same session: 2j7d6zrl, 3fiuji14, 5686l1eu, 8t2swltp, aqscjiu5, kfdgv9ny, nq9vi82l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 814.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 814.5 52.2 30.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dkkh8sv5",
   "source_url": "https://omlx.ai/benchmarks/performance/dkkh8sv5",
   "tg_tokens": null,
   "tps": 52.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dkkh8sv5-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: 2j7d6zrl, 3fiuji14, 5686l1eu, 8t2swltp, aqscjiu5, kfdgv9ny, nq9vi82l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 775.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 775.6 39.1 32.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dkkh8sv5",
   "source_url": "https://omlx.ai/benchmarks/performance/dkkh8sv5",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dkkh8sv5-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.9 GB at this context; duplicate traces of the same session: 2j7d6zrl, 3fiuji14, 5686l1eu, 8t2swltp, aqscjiu5, kfdgv9ny, nq9vi82l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 683.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 683.8 38.2 32.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dkkh8sv5",
   "source_url": "https://omlx.ai/benchmarks/performance/dkkh8sv5",
   "tg_tokens": null,
   "tps": 38.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dkkh8sv5-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.4 GB at this context; duplicate traces of the same session: 2j7d6zrl, 3fiuji14, 5686l1eu, 8t2swltp, aqscjiu5, kfdgv9ny, nq9vi82l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 655.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 655.5 19.6 34.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dkkh8sv5",
   "source_url": "https://omlx.ai/benchmarks/performance/dkkh8sv5",
   "tg_tokens": null,
   "tps": 19.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dkkh8sv5-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 36.8 GB at this context; duplicate traces of the same session: 2j7d6zrl, 3fiuji14, 5686l1eu, 8t2swltp, aqscjiu5, kfdgv9ny, nq9vi82l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 600.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 600.3 7.5 36.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dkkh8sv5",
   "source_url": "https://omlx.ai/benchmarks/performance/dkkh8sv5",
   "tg_tokens": null,
   "tps": 7.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dkkh8sv5-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 41.6 GB at this context; duplicate traces of the same session: 2j7d6zrl, 3fiuji14, 5686l1eu, 8t2swltp, aqscjiu5, kfdgv9ny, nq9vi82l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 406.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 406.1 3.7 41.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dkkh8sv5",
   "source_url": "https://omlx.ai/benchmarks/performance/dkkh8sv5",
   "tg_tokens": null,
   "tps": 3.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dkkh8sv5-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 50.8 GB at this context; duplicate traces of the same session: 2j7d6zrl, 3fiuji14, 5686l1eu, 8t2swltp, aqscjiu5, kfdgv9ny, nq9vi82l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 303.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 303.3 3.1 50.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dkkh8sv5",
   "source_url": "https://omlx.ai/benchmarks/performance/dkkh8sv5",
   "tg_tokens": null,
   "tps": 3.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dkkh8sv5-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 64.6 GB at this context; duplicate traces of the same session: 2j7d6zrl, 3fiuji14, 5686l1eu, 8t2swltp, aqscjiu5, kfdgv9ny, nq9vi82l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 263.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 195k 263.7 1.1 64.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dkkh8sv5",
   "source_url": "https://omlx.ai/benchmarks/performance/dkkh8sv5",
   "tg_tokens": null,
   "tps": 1.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-o955g056-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: bfg8nvz6, v9dsr8ku; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 326.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 326.9 23.0 28.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o955g056",
   "source_url": "https://omlx.ai/benchmarks/performance/o955g056",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-o955g056-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.7 GB at this context; duplicate traces of the same session: bfg8nvz6, v9dsr8ku; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 339.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 339.9 22.7 31.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o955g056",
   "source_url": "https://omlx.ai/benchmarks/performance/o955g056",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-o955g056-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.2 GB at this context; duplicate traces of the same session: bfg8nvz6, v9dsr8ku; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 335.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 8k 335.0 22.1 32.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o955g056",
   "source_url": "https://omlx.ai/benchmarks/performance/o955g056",
   "tg_tokens": null,
   "tps": 22.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4mkjysn5-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 871klm3m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 831.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 831.4 19.0 28.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4mkjysn5",
   "source_url": "https://omlx.ai/benchmarks/performance/4mkjysn5",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4mkjysn5-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: 871klm3m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 961.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 961.3 18.9 29.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4mkjysn5",
   "source_url": "https://omlx.ai/benchmarks/performance/4mkjysn5",
   "tg_tokens": null,
   "tps": 18.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-l3bh3zz6-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 8a5ujzi6, 9gzdsi5y, godm5p6m, t6se95t6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 813.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 813.5 48.1 17.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l3bh3zz6",
   "source_url": "https://omlx.ai/benchmarks/performance/l3bh3zz6",
   "tg_tokens": null,
   "tps": 48.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-l3bh3zz6-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 8a5ujzi6, 9gzdsi5y, godm5p6m, t6se95t6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1300.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,300 60.8 19.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l3bh3zz6",
   "source_url": "https://omlx.ai/benchmarks/performance/l3bh3zz6",
   "tg_tokens": null,
   "tps": 60.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-l3bh3zz6-ctx8192",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 8a5ujzi6, 9gzdsi5y, godm5p6m, t6se95t6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1369.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,369 60.2 19.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l3bh3zz6",
   "source_url": "https://omlx.ai/benchmarks/performance/l3bh3zz6",
   "tg_tokens": null,
   "tps": 60.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-l3bh3zz6-ctx16384",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 8a5ujzi6, 9gzdsi5y, godm5p6m, t6se95t6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1266.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,266 81.6 19.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l3bh3zz6",
   "source_url": "https://omlx.ai/benchmarks/performance/l3bh3zz6",
   "tg_tokens": null,
   "tps": 81.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-l3bh3zz6-ctx32768",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 8a5ujzi6, 9gzdsi5y, godm5p6m, t6se95t6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1128.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,128 146.3 20.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l3bh3zz6",
   "source_url": "https://omlx.ai/benchmarks/performance/l3bh3zz6",
   "tg_tokens": null,
   "tps": 146.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-l3bh3zz6-ctx65536",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 8a5ujzi6, 9gzdsi5y, godm5p6m, t6se95t6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 969.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 969.3 53.0 23.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l3bh3zz6",
   "source_url": "https://omlx.ai/benchmarks/performance/l3bh3zz6",
   "tg_tokens": null,
   "tps": 53.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1fpk58v5-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 581.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 581.6 42.2 35.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1fpk58v5",
   "source_url": "https://omlx.ai/benchmarks/performance/1fpk58v5",
   "tg_tokens": null,
   "tps": 42.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1fpk58v5-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 606.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 606.4 40.6 35.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1fpk58v5",
   "source_url": "https://omlx.ai/benchmarks/performance/1fpk58v5",
   "tg_tokens": null,
   "tps": 40.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yxjh9ka4-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: at00dvr6, d2s6iikz, u6f1jbgc, utzeru1n, x2szi2mc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 809.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 809.2 43.3 33.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yxjh9ka4",
   "source_url": "https://omlx.ai/benchmarks/performance/yxjh9ka4",
   "tg_tokens": null,
   "tps": 43.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yxjh9ka4-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: at00dvr6, d2s6iikz, u6f1jbgc, utzeru1n, x2szi2mc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 706.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 706.5 36.0 34.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yxjh9ka4",
   "source_url": "https://omlx.ai/benchmarks/performance/yxjh9ka4",
   "tg_tokens": null,
   "tps": 36.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yxjh9ka4-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: at00dvr6, d2s6iikz, u6f1jbgc, utzeru1n, x2szi2mc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 671.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 671.3 37.9 35.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yxjh9ka4",
   "source_url": "https://omlx.ai/benchmarks/performance/yxjh9ka4",
   "tg_tokens": null,
   "tps": 37.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yxjh9ka4-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: at00dvr6, d2s6iikz, u6f1jbgc, utzeru1n, x2szi2mc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 640.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 640.9 52.9 35.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yxjh9ka4",
   "source_url": "https://omlx.ai/benchmarks/performance/yxjh9ka4",
   "tg_tokens": null,
   "tps": 52.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yxjh9ka4-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: at00dvr6, d2s6iikz, u6f1jbgc, utzeru1n, x2szi2mc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 586.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 586.3 55.9 36.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yxjh9ka4",
   "source_url": "https://omlx.ai/benchmarks/performance/yxjh9ka4",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yxjh9ka4-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: at00dvr6, d2s6iikz, u6f1jbgc, utzeru1n, x2szi2mc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 458.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 458.3 17.7 39.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yxjh9ka4",
   "source_url": "https://omlx.ai/benchmarks/performance/yxjh9ka4",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yxjh9ka4-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 44.0 GB at this context; duplicate traces of the same session: at00dvr6, d2s6iikz, u6f1jbgc, utzeru1n, x2szi2mc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 305.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 305.1 20.7 44.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yxjh9ka4",
   "source_url": "https://omlx.ai/benchmarks/performance/yxjh9ka4",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yxjh9ka4-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 49.3 GB at this context; duplicate traces of the same session: at00dvr6, d2s6iikz, u6f1jbgc, utzeru1n, x2szi2mc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 263.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 195k 263.5 16.7 49.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yxjh9ka4",
   "source_url": "https://omlx.ai/benchmarks/performance/yxjh9ka4",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-a4t2jjlg-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace, single run at 195k context; peak memory 83.3 GB; TTFT 85395 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2342.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k tokens 2,342 PP tok/s 60.3 TG tok/s 85395 TTFT (ms) 83.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a4t2jjlg",
   "source_url": "https://omlx.ai/benchmarks/performance/a4t2jjlg",
   "tg_tokens": null,
   "tps": 60.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-aqhtjho5-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 75.3 GB; TTFT 837 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1223.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,223 PP tok/s 67.0 TG tok/s 837 TTFT (ms) 75.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aqhtjho5",
   "source_url": "https://omlx.ai/benchmarks/performance/aqhtjho5",
   "tg_tokens": null,
   "tps": 67.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mybon991-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 75.3 GB at this context; duplicate traces of the same session: 0a3uj8u9, 9g6f0445; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 673.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 673.9 58.2 75.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mybon991",
   "source_url": "https://omlx.ai/benchmarks/performance/mybon991",
   "tg_tokens": null,
   "tps": 58.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mybon991-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 76.4 GB at this context; duplicate traces of the same session: 0a3uj8u9, 9g6f0445; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2137.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,137 51.8 76.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mybon991",
   "source_url": "https://omlx.ai/benchmarks/performance/mybon991",
   "tg_tokens": null,
   "tps": 51.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mybon991-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 77.7 GB at this context; duplicate traces of the same session: 0a3uj8u9, 9g6f0445; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2678.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,678 56.7 77.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mybon991",
   "source_url": "https://omlx.ai/benchmarks/performance/mybon991",
   "tg_tokens": null,
   "tps": 56.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1b9rsukg-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.2 GB at this context; duplicate traces of the same session: 6thf2af5, yosiz7hl; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1029.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,029 50.1 101.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1b9rsukg",
   "source_url": "https://omlx.ai/benchmarks/performance/1b9rsukg",
   "tg_tokens": null,
   "tps": 50.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1b9rsukg-ctx8192",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: 6thf2af5, yosiz7hl; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1401.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 1,401 71.4 101.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1b9rsukg",
   "source_url": "https://omlx.ai/benchmarks/performance/1b9rsukg",
   "tg_tokens": null,
   "tps": 71.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1b9rsukg-ctx16384",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.8 GB at this context; duplicate traces of the same session: 6thf2af5, yosiz7hl; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1380.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 1,380 59.1 101.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1b9rsukg",
   "source_url": "https://omlx.ai/benchmarks/performance/1b9rsukg",
   "tg_tokens": null,
   "tps": 59.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ipesr05b-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 69.2 GB; TTFT 947 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1081.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,081 PP tok/s 79.4 TG tok/s 947 TTFT (ms) 69.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ipesr05b",
   "source_url": "https://omlx.ai/benchmarks/performance/ipesr05b",
   "tg_tokens": null,
   "tps": 79.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-jg2eetbg-ctx1024",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: 57v8qmr6, b0hetphw, mncx44tn, qiu9xnks; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2075.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 2,075 50.8 25.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jg2eetbg",
   "source_url": "https://omlx.ai/benchmarks/performance/jg2eetbg",
   "tg_tokens": null,
   "tps": 50.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-jg2eetbg-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: 57v8qmr6, b0hetphw, mncx44tn, qiu9xnks; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2177.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 2,177 49.4 26.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jg2eetbg",
   "source_url": "https://omlx.ai/benchmarks/performance/jg2eetbg",
   "tg_tokens": null,
   "tps": 49.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-jg2eetbg-ctx8192",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: 57v8qmr6, b0hetphw, mncx44tn, qiu9xnks; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2142.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 2,142 47.6 26.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jg2eetbg",
   "source_url": "https://omlx.ai/benchmarks/performance/jg2eetbg",
   "tg_tokens": null,
   "tps": 47.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-jg2eetbg-ctx16384",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: 57v8qmr6, b0hetphw, mncx44tn, qiu9xnks; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2000.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 16k 2,000 42.6 26.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jg2eetbg",
   "source_url": "https://omlx.ai/benchmarks/performance/jg2eetbg",
   "tg_tokens": null,
   "tps": 42.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-jg2eetbg-ctx32768",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: 57v8qmr6, b0hetphw, mncx44tn, qiu9xnks; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1800.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 32k 1,800 36.4 27.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jg2eetbg",
   "source_url": "https://omlx.ai/benchmarks/performance/jg2eetbg",
   "tg_tokens": null,
   "tps": 36.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-et3jspf1-ctx1024",
   "model": "Gemma-4-31B",
   "notes": "oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: 4xtx3jfl, 5v6yib2u, 7lcsdxy0, lcpy5e14; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 586.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31b-it-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 586.5 14.5 32.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/et3jspf1",
   "source_url": "https://omlx.ai/benchmarks/performance/et3jspf1",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-et3jspf1-ctx4096",
   "model": "Gemma-4-31B",
   "notes": "oMLX community benchmark trace; peak memory 33.0 GB at this context; duplicate traces of the same session: 4xtx3jfl, 5v6yib2u, 7lcsdxy0, lcpy5e14; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 626.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31b-it-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 626.1 13.5 33.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/et3jspf1",
   "source_url": "https://omlx.ai/benchmarks/performance/et3jspf1",
   "tg_tokens": null,
   "tps": 13.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-et3jspf1-ctx8192",
   "model": "Gemma-4-31B",
   "notes": "oMLX community benchmark trace; peak memory 33.6 GB at this context; duplicate traces of the same session: 4xtx3jfl, 5v6yib2u, 7lcsdxy0, lcpy5e14; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 635.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31b-it-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 635.0 13.3 33.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/et3jspf1",
   "source_url": "https://omlx.ai/benchmarks/performance/et3jspf1",
   "tg_tokens": null,
   "tps": 13.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-et3jspf1-ctx16384",
   "model": "Gemma-4-31B",
   "notes": "oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: 4xtx3jfl, 5v6yib2u, 7lcsdxy0, lcpy5e14; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 574.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31b-it-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 574.7 12.0 34.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/et3jspf1",
   "source_url": "https://omlx.ai/benchmarks/performance/et3jspf1",
   "tg_tokens": null,
   "tps": 12.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-et3jspf1-ctx32768",
   "model": "Gemma-4-31B",
   "notes": "oMLX community benchmark trace; peak memory 37.0 GB at this context; duplicate traces of the same session: 4xtx3jfl, 5v6yib2u, 7lcsdxy0, lcpy5e14; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 546.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31b-it-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 546.8 10.2 37.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/et3jspf1",
   "source_url": "https://omlx.ai/benchmarks/performance/et3jspf1",
   "tg_tokens": null,
   "tps": 10.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-oo12epr4-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: k1uf7ipe, puands13, sz2l1w1x; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 138.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 138.1 15.2 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oo12epr4",
   "source_url": "https://omlx.ai/benchmarks/performance/oo12epr4",
   "tg_tokens": null,
   "tps": 15.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-oo12epr4-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: k1uf7ipe, puands13, sz2l1w1x; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 120.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 120.8 12.8 22.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oo12epr4",
   "source_url": "https://omlx.ai/benchmarks/performance/oo12epr4",
   "tg_tokens": null,
   "tps": 12.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-oo12epr4-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: k1uf7ipe, puands13, sz2l1w1x; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 106.4 11.5 24.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oo12epr4",
   "source_url": "https://omlx.ai/benchmarks/performance/oo12epr4",
   "tg_tokens": null,
   "tps": 11.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-oo12epr4-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 28.7 GB at this context; duplicate traces of the same session: k1uf7ipe, puands13, sz2l1w1x; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 85.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 128k 85.1 9.1 28.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oo12epr4",
   "source_url": "https://omlx.ai/benchmarks/performance/oo12epr4",
   "tg_tokens": null,
   "tps": 9.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ty187ulu-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: 296i4sth, 32sukqq7, 62y5rknq, 6urndume, 9skau5od, b1420bsq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3486.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 8bit; 1k 3,486 82.6 24.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ty187ulu",
   "source_url": "https://omlx.ai/benchmarks/performance/ty187ulu",
   "tg_tokens": null,
   "tps": 82.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ty187ulu-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: 296i4sth, 32sukqq7, 62y5rknq, 6urndume, 9skau5od, b1420bsq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4301.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 8bit; 4k 4,301 82.4 24.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ty187ulu",
   "source_url": "https://omlx.ai/benchmarks/performance/ty187ulu",
   "tg_tokens": null,
   "tps": 82.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ty187ulu-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: 296i4sth, 32sukqq7, 62y5rknq, 6urndume, 9skau5od, b1420bsq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4284.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 8bit; 8k 4,284 80.6 24.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ty187ulu",
   "source_url": "https://omlx.ai/benchmarks/performance/ty187ulu",
   "tg_tokens": null,
   "tps": 80.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ty187ulu-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: 296i4sth, 32sukqq7, 62y5rknq, 6urndume, 9skau5od, b1420bsq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4158.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 8bit; 16k 4,158 77.2 24.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ty187ulu",
   "source_url": "https://omlx.ai/benchmarks/performance/ty187ulu",
   "tg_tokens": null,
   "tps": 77.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ty187ulu-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 25.2 GB at this context; duplicate traces of the same session: 296i4sth, 32sukqq7, 62y5rknq, 6urndume, 9skau5od, b1420bsq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3724.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 8bit; 32k 3,724 71.6 25.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ty187ulu",
   "source_url": "https://omlx.ai/benchmarks/performance/ty187ulu",
   "tg_tokens": null,
   "tps": 71.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ty187ulu-ctx65536",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: 296i4sth, 32sukqq7, 62y5rknq, 6urndume, 9skau5od, b1420bsq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2944.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 8bit; 64k 2,944 61.7 26.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ty187ulu",
   "source_url": "https://omlx.ai/benchmarks/performance/ty187ulu",
   "tg_tokens": null,
   "tps": 61.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ty187ulu-ctx131072",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: 296i4sth, 32sukqq7, 62y5rknq, 6urndume, 9skau5od, b1420bsq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2173.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 8bit; 128k 2,173 48.2 28.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ty187ulu",
   "source_url": "https://omlx.ai/benchmarks/performance/ty187ulu",
   "tg_tokens": null,
   "tps": 48.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-23l4d34o-ctx1024",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "ET-O5.5-G6A-G4.7-DSV4-K3-RLOO MTP DFlash2 build; oMLX community benchmark trace; peak memory 16.8 GB at this context; duplicate traces of the same session: 9payywrm, hyg0u91h; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 335.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-ET-O5.5-G6A-G4.7-DSV4-K3-RLOO-MTP-DFlash2-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 335.1 37.1 16.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/23l4d34o",
   "source_url": "https://omlx.ai/benchmarks/performance/23l4d34o",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-23l4d34o-ctx4096",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "ET-O5.5-G6A-G4.7-DSV4-K3-RLOO MTP DFlash2 build; oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 9payywrm, hyg0u91h; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 355.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-ET-O5.5-G6A-G4.7-DSV4-K3-RLOO-MTP-DFlash2-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 355.1 28.6 20.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/23l4d34o",
   "source_url": "https://omlx.ai/benchmarks/performance/23l4d34o",
   "tg_tokens": null,
   "tps": 28.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-23l4d34o-ctx8192",
   "model": "Qwen3.8-27B-Coder390",
   "notes": "ET-O5.5-G6A-G4.7-DSV4-K3-RLOO MTP DFlash2 build; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 9payywrm, hyg0u91h; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 349.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Coder390-ET-O5.5-G6A-G4.7-DSV4-K3-RLOO-MTP-DFlash2-oQ4e on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k 349.4 35.2 20.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/23l4d34o",
   "source_url": "https://omlx.ai/benchmarks/performance/23l4d34o",
   "tg_tokens": null,
   "tps": 35.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-h4igdh0i-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 5j42rtjg, is51t0gr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 326.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 326.1 23.0 28.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h4igdh0i",
   "source_url": "https://omlx.ai/benchmarks/performance/h4igdh0i",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-h4igdh0i-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.7 GB at this context; duplicate traces of the same session: 5j42rtjg, is51t0gr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 341.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 341.5 22.7 31.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h4igdh0i",
   "source_url": "https://omlx.ai/benchmarks/performance/h4igdh0i",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-h4igdh0i-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.2 GB at this context; duplicate traces of the same session: 5j42rtjg, is51t0gr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 338.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 8k 338.0 22.3 32.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h4igdh0i",
   "source_url": "https://omlx.ai/benchmarks/performance/h4igdh0i",
   "tg_tokens": null,
   "tps": 22.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-fxozv73h-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: ftpyi8br, y5ys29j4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 327.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 327.5 23.0 28.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fxozv73h",
   "source_url": "https://omlx.ai/benchmarks/performance/fxozv73h",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-fxozv73h-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.7 GB at this context; duplicate traces of the same session: ftpyi8br, y5ys29j4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 340.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 340.9 22.7 31.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fxozv73h",
   "source_url": "https://omlx.ai/benchmarks/performance/fxozv73h",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-fxozv73h-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.2 GB at this context; duplicate traces of the same session: ftpyi8br, y5ys29j4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 336.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 8k 336.0 22.3 32.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fxozv73h",
   "source_url": "https://omlx.ai/benchmarks/performance/fxozv73h",
   "tg_tokens": null,
   "tps": 22.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-6q52qp3g-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 8asu19ao; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 327.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 327.5 23.1 28.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6q52qp3g",
   "source_url": "https://omlx.ai/benchmarks/performance/6q52qp3g",
   "tg_tokens": null,
   "tps": 23.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-6q52qp3g-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.7 GB at this context; duplicate traces of the same session: 8asu19ao; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 340.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 340.1 22.8 31.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6q52qp3g",
   "source_url": "https://omlx.ai/benchmarks/performance/6q52qp3g",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2f2zrocz-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: fwzc39nn, gvvwm82j; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 327.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 327.7 23.0 28.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2f2zrocz",
   "source_url": "https://omlx.ai/benchmarks/performance/2f2zrocz",
   "tg_tokens": null,
   "tps": 23.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2f2zrocz-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.7 GB at this context; duplicate traces of the same session: fwzc39nn, gvvwm82j; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 343.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 343.2 22.8 31.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2f2zrocz",
   "source_url": "https://omlx.ai/benchmarks/performance/2f2zrocz",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2f2zrocz-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.2 GB at this context; duplicate traces of the same session: fwzc39nn, gvvwm82j; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 316.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 8k 316.2 22.2 32.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2f2zrocz",
   "source_url": "https://omlx.ai/benchmarks/performance/2f2zrocz",
   "tg_tokens": null,
   "tps": 22.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4fkukd4t-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 813.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 813.5 48.1 17.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4fkukd4t",
   "source_url": "https://omlx.ai/benchmarks/performance/4fkukd4t",
   "tg_tokens": null,
   "tps": 48.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4fkukd4t-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1300.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,300 60.8 19.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4fkukd4t",
   "source_url": "https://omlx.ai/benchmarks/performance/4fkukd4t",
   "tg_tokens": null,
   "tps": 60.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4fkukd4t-ctx8192",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1369.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,369 60.2 19.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4fkukd4t",
   "source_url": "https://omlx.ai/benchmarks/performance/4fkukd4t",
   "tg_tokens": null,
   "tps": 60.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4fkukd4t-ctx16384",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1266.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,266 81.6 19.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4fkukd4t",
   "source_url": "https://omlx.ai/benchmarks/performance/4fkukd4t",
   "tg_tokens": null,
   "tps": 81.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4fkukd4t-ctx32768",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1128.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 1,128 146.3 20.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4fkukd4t",
   "source_url": "https://omlx.ai/benchmarks/performance/4fkukd4t",
   "tg_tokens": null,
   "tps": 146.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4fkukd4t-ctx65536",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 969.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 969.3 53.0 23.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4fkukd4t",
   "source_url": "https://omlx.ai/benchmarks/performance/4fkukd4t",
   "tg_tokens": null,
   "tps": 53.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8sk3ah0v-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; bf16; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: 9swgtc4k, jydyrx6k, pc6aggr2, sj5z0q3j, uxncucu4, v45ts89h, xmkbn89x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 781.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Uncensored-oQ8e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 781.5 41.7 33.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8sk3ah0v",
   "source_url": "https://omlx.ai/benchmarks/performance/8sk3ah0v",
   "tg_tokens": null,
   "tps": 41.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8sk3ah0v-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; bf16; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: 9swgtc4k, jydyrx6k, pc6aggr2, sj5z0q3j, uxncucu4, v45ts89h, xmkbn89x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 746.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Uncensored-oQ8e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 746.7 38.9 34.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8sk3ah0v",
   "source_url": "https://omlx.ai/benchmarks/performance/8sk3ah0v",
   "tg_tokens": null,
   "tps": 38.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8sk3ah0v-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; bf16; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: 9swgtc4k, jydyrx6k, pc6aggr2, sj5z0q3j, uxncucu4, v45ts89h, xmkbn89x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 679.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Uncensored-oQ8e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 679.2 53.7 35.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8sk3ah0v",
   "source_url": "https://omlx.ai/benchmarks/performance/8sk3ah0v",
   "tg_tokens": null,
   "tps": 53.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8sk3ah0v-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; bf16; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: 9swgtc4k, jydyrx6k, pc6aggr2, sj5z0q3j, uxncucu4, v45ts89h, xmkbn89x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 639.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Uncensored-oQ8e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 639.2 53.0 35.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8sk3ah0v",
   "source_url": "https://omlx.ai/benchmarks/performance/8sk3ah0v",
   "tg_tokens": null,
   "tps": 53.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8sk3ah0v-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; bf16; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: 9swgtc4k, jydyrx6k, pc6aggr2, sj5z0q3j, uxncucu4, v45ts89h, xmkbn89x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 562.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Uncensored-oQ8e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 562.1 53.4 36.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8sk3ah0v",
   "source_url": "https://omlx.ai/benchmarks/performance/8sk3ah0v",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8sk3ah0v-ctx65536",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; bf16; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: 9swgtc4k, jydyrx6k, pc6aggr2, sj5z0q3j, uxncucu4, v45ts89h, xmkbn89x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 405.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Uncensored-oQ8e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 405.0 19.5 39.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8sk3ah0v",
   "source_url": "https://omlx.ai/benchmarks/performance/8sk3ah0v",
   "tg_tokens": null,
   "tps": 19.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8sk3ah0v-ctx131072",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; bf16; oMLX community benchmark trace; peak memory 44.0 GB at this context; duplicate traces of the same session: 9swgtc4k, jydyrx6k, pc6aggr2, sj5z0q3j, uxncucu4, v45ts89h, xmkbn89x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 309.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Uncensored-oQ8e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 309.2 21.6 44.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8sk3ah0v",
   "source_url": "https://omlx.ai/benchmarks/performance/8sk3ah0v",
   "tg_tokens": null,
   "tps": 21.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8sk3ah0v-ctx199680",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; bf16; oMLX community benchmark trace; peak memory 49.3 GB at this context; duplicate traces of the same session: 9swgtc4k, jydyrx6k, pc6aggr2, sj5z0q3j, uxncucu4, v45ts89h, xmkbn89x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 270.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Uncensored-oQ8e-bf16-mtp-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 195k 270.9 15.8 49.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8sk3ah0v",
   "source_url": "https://omlx.ai/benchmarks/performance/8sk3ah0v",
   "tg_tokens": null,
   "tps": 15.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-i0y9j4wy-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 9mvp8mg1; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 426.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 426.5 26.8 17.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i0y9j4wy",
   "source_url": "https://omlx.ai/benchmarks/performance/i0y9j4wy",
   "tg_tokens": null,
   "tps": 26.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-i0y9j4wy-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 9mvp8mg1; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 556.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 556.7 39.1 19.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i0y9j4wy",
   "source_url": "https://omlx.ai/benchmarks/performance/i0y9j4wy",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0iq5r1w8-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: prkttlg2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 809.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 809.2 43.3 33.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0iq5r1w8",
   "source_url": "https://omlx.ai/benchmarks/performance/0iq5r1w8",
   "tg_tokens": null,
   "tps": 43.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0iq5r1w8-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: prkttlg2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 706.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 706.5 36.0 34.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0iq5r1w8",
   "source_url": "https://omlx.ai/benchmarks/performance/0iq5r1w8",
   "tg_tokens": null,
   "tps": 36.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0iq5r1w8-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: prkttlg2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 671.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 671.3 37.9 35.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0iq5r1w8",
   "source_url": "https://omlx.ai/benchmarks/performance/0iq5r1w8",
   "tg_tokens": null,
   "tps": 37.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0iq5r1w8-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: prkttlg2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 640.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 640.9 52.9 35.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0iq5r1w8",
   "source_url": "https://omlx.ai/benchmarks/performance/0iq5r1w8",
   "tg_tokens": null,
   "tps": 52.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0iq5r1w8-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: prkttlg2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 586.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 586.3 55.9 36.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0iq5r1w8",
   "source_url": "https://omlx.ai/benchmarks/performance/0iq5r1w8",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0iq5r1w8-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: prkttlg2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 458.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 458.3 17.7 39.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0iq5r1w8",
   "source_url": "https://omlx.ai/benchmarks/performance/0iq5r1w8",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0iq5r1w8-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 44.0 GB at this context; duplicate traces of the same session: prkttlg2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 305.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 305.1 20.7 44.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0iq5r1w8",
   "source_url": "https://omlx.ai/benchmarks/performance/0iq5r1w8",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0iq5r1w8-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 49.3 GB at this context; duplicate traces of the same session: prkttlg2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 263.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 195k 263.5 16.7 49.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0iq5r1w8",
   "source_url": "https://omlx.ai/benchmarks/performance/0iq5r1w8",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-i464krxc-ctx1024",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 38.2 GB at this context; duplicate traces of the same session: 0dhvk3o5, 80aczidt, b5cz0qfa, ixo0dssi, nafvixta, ww56m6zx; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 853.3,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 8bit; 1k 853.3 40.0 38.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i464krxc",
   "source_url": "https://omlx.ai/benchmarks/performance/i464krxc",
   "tg_tokens": null,
   "tps": 40.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-i464krxc-ctx8192",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 40.6 GB at this context; duplicate traces of the same session: 0dhvk3o5, 80aczidt, b5cz0qfa, ixo0dssi, nafvixta, ww56m6zx; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1502.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 8bit; 8k 1,502 55.7 40.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i464krxc",
   "source_url": "https://omlx.ai/benchmarks/performance/i464krxc",
   "tg_tokens": null,
   "tps": 55.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-i464krxc-ctx16384",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 41.3 GB at this context; duplicate traces of the same session: 0dhvk3o5, 80aczidt, b5cz0qfa, ixo0dssi, nafvixta, ww56m6zx; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1346.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 8bit; 16k 1,346 74.9 41.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i464krxc",
   "source_url": "https://omlx.ai/benchmarks/performance/i464krxc",
   "tg_tokens": null,
   "tps": 74.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-i464krxc-ctx32768",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 42.6 GB at this context; duplicate traces of the same session: 0dhvk3o5, 80aczidt, b5cz0qfa, ixo0dssi, nafvixta, ww56m6zx; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1074.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 8bit; 32k 1,074 75.8 42.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i464krxc",
   "source_url": "https://omlx.ai/benchmarks/performance/i464krxc",
   "tg_tokens": null,
   "tps": 75.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-i464krxc-ctx65536",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 43.5 GB at this context; duplicate traces of the same session: 0dhvk3o5, 80aczidt, b5cz0qfa, ixo0dssi, nafvixta, ww56m6zx; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 774.6,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 8bit; 64k 774.6 47.7 43.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i464krxc",
   "source_url": "https://omlx.ai/benchmarks/performance/i464krxc",
   "tg_tokens": null,
   "tps": 47.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-i464krxc-ctx131072",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 45.0 GB at this context; duplicate traces of the same session: 0dhvk3o5, 80aczidt, b5cz0qfa, ixo0dssi, nafvixta, ww56m6zx; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 499.4,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 8bit; 128k 499.4 54.0 45.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i464krxc",
   "source_url": "https://omlx.ai/benchmarks/performance/i464krxc",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-i464krxc-ctx199680",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 46.8 GB at this context; duplicate traces of the same session: 0dhvk3o5, 80aczidt, b5cz0qfa, ixo0dssi, nafvixta, ww56m6zx; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 364.5,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 8bit; 195k 364.5 36.4 46.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i464krxc",
   "source_url": "https://omlx.ai/benchmarks/performance/i464krxc",
   "tg_tokens": null,
   "tps": 36.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7u6j7g3h-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6 mixed-quant MTP v2 build; oMLX community benchmark trace, single run at 16k context; peak memory 80.2 GB; TTFT 11627 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1409.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp-v2 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k tokens 1,409 PP tok/s 8.5 TG tok/s 11627 TTFT (ms) 80.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7u6j7g3h",
   "source_url": "https://omlx.ai/benchmarks/performance/7u6j7g3h",
   "tg_tokens": null,
   "tps": 8.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n85y44qg-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 71.8 GB; TTFT 5473 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1497.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k tokens 1,497 PP tok/s 63.3 TG tok/s 5473 TTFT (ms) 71.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n85y44qg",
   "source_url": "https://omlx.ai/benchmarks/performance/n85y44qg",
   "tg_tokens": null,
   "tps": 63.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-zqmp2a7e-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mixed MTP build; oMLX community benchmark trace; peak memory 94.1 GB at this context; duplicate traces of the same session: ihxjsruo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 465.4,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "Qwen3.8-Flash-Next-oQ5-mixed-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 465.4 38.6 94.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zqmp2a7e",
   "source_url": "https://omlx.ai/benchmarks/performance/zqmp2a7e",
   "tg_tokens": null,
   "tps": 38.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-zqmp2a7e-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mixed MTP build; oMLX community benchmark trace; peak memory 94.3 GB at this context; duplicate traces of the same session: ihxjsruo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 856.1,
   "provenance": "community",
   "quant": "oQ5",
   "quote": "Qwen3.8-Flash-Next-oQ5-mixed-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 856.1 42.3 94.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zqmp2a7e",
   "source_url": "https://omlx.ai/benchmarks/performance/zqmp2a7e",
   "tg_tokens": null,
   "tps": 42.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-wd48kzbc-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 2zc0vy8k, p002yneu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 464.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 464.1 38.2 22.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wd48kzbc",
   "source_url": "https://omlx.ai/benchmarks/performance/wd48kzbc",
   "tg_tokens": null,
   "tps": 38.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-wd48kzbc-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: 2zc0vy8k, p002yneu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 512.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 512.3 34.0 24.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wd48kzbc",
   "source_url": "https://omlx.ai/benchmarks/performance/wd48kzbc",
   "tg_tokens": null,
   "tps": 34.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-wd48kzbc-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: 2zc0vy8k, p002yneu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 525.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 525.4 37.9 24.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wd48kzbc",
   "source_url": "https://omlx.ai/benchmarks/performance/wd48kzbc",
   "tg_tokens": null,
   "tps": 37.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-pn3t7b4i-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: i8eh6u17, vbbrxrd1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 475.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 475.1 29.8 20.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pn3t7b4i",
   "source_url": "https://omlx.ai/benchmarks/performance/pn3t7b4i",
   "tg_tokens": null,
   "tps": 29.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-pn3t7b4i-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: i8eh6u17, vbbrxrd1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 483.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 483.6 28.9 21.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pn3t7b4i",
   "source_url": "https://omlx.ai/benchmarks/performance/pn3t7b4i",
   "tg_tokens": null,
   "tps": 28.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-pn3t7b4i-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: i8eh6u17, vbbrxrd1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 486.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 486.6 27.9 22.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pn3t7b4i",
   "source_url": "https://omlx.ai/benchmarks/performance/pn3t7b4i",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-27o48x5a-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.5 GB at this context; duplicate traces of the same session: at52bei7, fr7x6zs4; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 461.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 461.8 33.2 17.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/27o48x5a",
   "source_url": "https://omlx.ai/benchmarks/performance/27o48x5a",
   "tg_tokens": null,
   "tps": 33.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-27o48x5a-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: at52bei7, fr7x6zs4; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 475.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 475.4 31.0 18.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/27o48x5a",
   "source_url": "https://omlx.ai/benchmarks/performance/27o48x5a",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-27o48x5a-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: at52bei7, fr7x6zs4; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 472.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 472.0 37.5 19.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/27o48x5a",
   "source_url": "https://omlx.ai/benchmarks/performance/27o48x5a",
   "tg_tokens": null,
   "tps": 37.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-zxg01kpy-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 8.1 GB at this context; duplicate traces of the same session: 1gcn9qy8; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3313.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-E4B-it-MLX-8bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 8bit; 4k 3,313 49.1 8.1 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zxg01kpy",
   "source_url": "https://omlx.ai/benchmarks/performance/zxg01kpy",
   "tg_tokens": null,
   "tps": 49.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-zxg01kpy-ctx16384",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 8.3 GB at this context; duplicate traces of the same session: 1gcn9qy8; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3129.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-E4B-it-MLX-8bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 8bit; 16k 3,129 44.7 8.3 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zxg01kpy",
   "source_url": "https://omlx.ai/benchmarks/performance/zxg01kpy",
   "tg_tokens": null,
   "tps": 44.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-fe8wifjw-ctx1024",
   "model": "gemma-4-E4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 5.7 GB at this context; duplicate traces of the same session: j101tz3m; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 643.7,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "gemma-4-E4B-it-oQ4-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 643.7 54.4 5.7 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fe8wifjw",
   "source_url": "https://omlx.ai/benchmarks/performance/fe8wifjw",
   "tg_tokens": null,
   "tps": 54.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-fe8wifjw-ctx4096",
   "model": "gemma-4-E4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: j101tz3m; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 661.3,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "gemma-4-E4B-it-oQ4-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 661.3 56.9 5.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fe8wifjw",
   "source_url": "https://omlx.ai/benchmarks/performance/fe8wifjw",
   "tg_tokens": null,
   "tps": 56.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-ekn3cdf7-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 5.7 GB at this context; duplicate traces of the same session: pt3geujy; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 650.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 650.2 54.8 5.7 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ekn3cdf7",
   "source_url": "https://omlx.ai/benchmarks/performance/ekn3cdf7",
   "tg_tokens": null,
   "tps": 54.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-ekn3cdf7-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: pt3geujy; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 655.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 655.8 44.9 5.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ekn3cdf7",
   "source_url": "https://omlx.ai/benchmarks/performance/ekn3cdf7",
   "tg_tokens": null,
   "tps": 44.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-6mlbyv43-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.2 GB at this context; duplicate traces of the same session: 7e742lev, eh22pwe0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 96.5,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 96.5 14.3 92.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6mlbyv43",
   "source_url": "https://omlx.ai/benchmarks/performance/6mlbyv43",
   "tg_tokens": null,
   "tps": 14.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-6mlbyv43-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 94.2 GB at this context; duplicate traces of the same session: 7e742lev, eh22pwe0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 456.7,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 456.7 14.7 94.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6mlbyv43",
   "source_url": "https://omlx.ai/benchmarks/performance/6mlbyv43",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6mlbyv43-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 94.6 GB at this context; duplicate traces of the same session: 7e742lev, eh22pwe0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 470.7,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 470.7 14.9 94.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6mlbyv43",
   "source_url": "https://omlx.ai/benchmarks/performance/6mlbyv43",
   "tg_tokens": null,
   "tps": 14.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5h85jbe8-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.0 GB at this context; duplicate traces of the same session: l2e7amgn, vcjaco8z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 83.6,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 83.6 12.0 92.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5h85jbe8",
   "source_url": "https://omlx.ai/benchmarks/performance/5h85jbe8",
   "tg_tokens": null,
   "tps": 12.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5h85jbe8-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.3 GB at this context; duplicate traces of the same session: l2e7amgn, vcjaco8z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 155.3,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 155.3 12.4 92.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5h85jbe8",
   "source_url": "https://omlx.ai/benchmarks/performance/5h85jbe8",
   "tg_tokens": null,
   "tps": 12.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5h85jbe8-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.4 GB at this context; duplicate traces of the same session: l2e7amgn, vcjaco8z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 138.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 138.0 15.1 92.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5h85jbe8",
   "source_url": "https://omlx.ai/benchmarks/performance/5h85jbe8",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-qhbfkavs-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: 2gnx4rej, 4zxmi3cx, alh0az9g, dd3s0nni, dzls1gdp, p2igw565, zlhgf1b2; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1345.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,345 78.9 177.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qhbfkavs",
   "source_url": "https://omlx.ai/benchmarks/performance/qhbfkavs",
   "tg_tokens": null,
   "tps": 78.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-qhbfkavs-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: 2gnx4rej, 4zxmi3cx, alh0az9g, dd3s0nni, dzls1gdp, p2igw565, zlhgf1b2; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1931.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,931 79.1 178.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qhbfkavs",
   "source_url": "https://omlx.ai/benchmarks/performance/qhbfkavs",
   "tg_tokens": null,
   "tps": 79.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-qhbfkavs-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.9 GB at this context; duplicate traces of the same session: 2gnx4rej, 4zxmi3cx, alh0az9g, dd3s0nni, dzls1gdp, p2igw565, zlhgf1b2; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1936.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 1,936 106.3 178.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qhbfkavs",
   "source_url": "https://omlx.ai/benchmarks/performance/qhbfkavs",
   "tg_tokens": null,
   "tps": 106.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-qhbfkavs-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.0 GB at this context; duplicate traces of the same session: 2gnx4rej, 4zxmi3cx, alh0az9g, dd3s0nni, dzls1gdp, p2igw565, zlhgf1b2; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1924.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,924 137.6 179.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qhbfkavs",
   "source_url": "https://omlx.ai/benchmarks/performance/qhbfkavs",
   "tg_tokens": null,
   "tps": 137.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-qhbfkavs-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.3 GB at this context; duplicate traces of the same session: 2gnx4rej, 4zxmi3cx, alh0az9g, dd3s0nni, dzls1gdp, p2igw565, zlhgf1b2; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1919.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,919 109.2 179.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qhbfkavs",
   "source_url": "https://omlx.ai/benchmarks/performance/qhbfkavs",
   "tg_tokens": null,
   "tps": 109.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-qhbfkavs-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.7 GB at this context; duplicate traces of the same session: 2gnx4rej, 4zxmi3cx, alh0az9g, dd3s0nni, dzls1gdp, p2igw565, zlhgf1b2; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1888.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,888 86.4 179.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qhbfkavs",
   "source_url": "https://omlx.ai/benchmarks/performance/qhbfkavs",
   "tg_tokens": null,
   "tps": 86.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-qhbfkavs-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.6 GB at this context; duplicate traces of the same session: 2gnx4rej, 4zxmi3cx, alh0az9g, dd3s0nni, dzls1gdp, p2igw565, zlhgf1b2; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1853.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 1,853 135.9 180.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qhbfkavs",
   "source_url": "https://omlx.ai/benchmarks/performance/qhbfkavs",
   "tg_tokens": null,
   "tps": 135.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-qhbfkavs-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.7 GB at this context; duplicate traces of the same session: 2gnx4rej, 4zxmi3cx, alh0az9g, dd3s0nni, dzls1gdp, p2igw565, zlhgf1b2; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1814.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 1,814 73.4 181.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qhbfkavs",
   "source_url": "https://omlx.ai/benchmarks/performance/qhbfkavs",
   "tg_tokens": null,
   "tps": 73.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-1fmo1zaz-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.4 GB at this context; duplicate traces of the same session: 2k7mo8d1, 5x4w96yk, kevcaun3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1320.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,320 79.4 177.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1fmo1zaz",
   "source_url": "https://omlx.ai/benchmarks/performance/1fmo1zaz",
   "tg_tokens": null,
   "tps": 79.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-1fmo1zaz-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: 2k7mo8d1, 5x4w96yk, kevcaun3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1937.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,937 72.3 178.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1fmo1zaz",
   "source_url": "https://omlx.ai/benchmarks/performance/1fmo1zaz",
   "tg_tokens": null,
   "tps": 72.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-1fmo1zaz-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.9 GB at this context; duplicate traces of the same session: 2k7mo8d1, 5x4w96yk, kevcaun3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1932.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 1,932 87.0 178.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1fmo1zaz",
   "source_url": "https://omlx.ai/benchmarks/performance/1fmo1zaz",
   "tg_tokens": null,
   "tps": 87.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-1fmo1zaz-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.0 GB at this context; duplicate traces of the same session: 2k7mo8d1, 5x4w96yk, kevcaun3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1926.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,926 138.0 179.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1fmo1zaz",
   "source_url": "https://omlx.ai/benchmarks/performance/1fmo1zaz",
   "tg_tokens": null,
   "tps": 138.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-m2zh1b0b-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: 5mjg8pjg; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1444.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 8bit; 1k 1,444 71.1 24.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m2zh1b0b",
   "source_url": "https://omlx.ai/benchmarks/performance/m2zh1b0b",
   "tg_tokens": null,
   "tps": 71.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-m2zh1b0b-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: 5mjg8pjg; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1900.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 8bit; 4k 1,900 69.7 24.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m2zh1b0b",
   "source_url": "https://omlx.ai/benchmarks/performance/m2zh1b0b",
   "tg_tokens": null,
   "tps": 69.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-b3gjzrej-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: j7roq6io; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1468.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 1,468 108.7 19.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b3gjzrej",
   "source_url": "https://omlx.ai/benchmarks/performance/b3gjzrej",
   "tg_tokens": null,
   "tps": 108.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-b3gjzrej-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: j7roq6io; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2196.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 2,196 105.7 20.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b3gjzrej",
   "source_url": "https://omlx.ai/benchmarks/performance/b3gjzrej",
   "tg_tokens": null,
   "tps": 105.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-1dgqjjzm-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.9 GB at this context; duplicate traces of the same session: i9d1839n, osh0n73s; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 451.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 451.4 34.9 7.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1dgqjjzm",
   "source_url": "https://omlx.ai/benchmarks/performance/1dgqjjzm",
   "tg_tokens": null,
   "tps": 34.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-1dgqjjzm-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 9.3 GB at this context; duplicate traces of the same session: i9d1839n, osh0n73s; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 478.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 478.0 33.9 9.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1dgqjjzm",
   "source_url": "https://omlx.ai/benchmarks/performance/1dgqjjzm",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-1dgqjjzm-ctx8192",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 9.5 GB at this context; duplicate traces of the same session: i9d1839n, osh0n73s; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 469.2,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-9B-MLX-6bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 8k 469.2 32.9 9.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1dgqjjzm",
   "source_url": "https://omlx.ai/benchmarks/performance/1dgqjjzm",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-16",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-8davwoyc-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 29.0 GB; TTFT 1670848 ms; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 78.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 128k tokens 78.4 PP tok/s 13.5 TG tok/s 1670848 TTFT (ms) 29.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8davwoyc",
   "source_url": "https://omlx.ai/benchmarks/performance/8davwoyc",
   "tg_tokens": null,
   "tps": 13.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gmtb1qfc-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.6 GB at this context; duplicate traces of the same session: uuu89zg8, ybxlq6vf; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 327.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 327.3 23.1 33.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gmtb1qfc",
   "source_url": "https://omlx.ai/benchmarks/performance/gmtb1qfc",
   "tg_tokens": null,
   "tps": 23.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gmtb1qfc-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 36.7 GB at this context; duplicate traces of the same session: uuu89zg8, ybxlq6vf; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 352.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 352.7 22.7 36.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gmtb1qfc",
   "source_url": "https://omlx.ai/benchmarks/performance/gmtb1qfc",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-16",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gmtb1qfc-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: uuu89zg8, ybxlq6vf; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 446.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 8k 446.0 22.0 37.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gmtb1qfc",
   "source_url": "https://omlx.ai/benchmarks/performance/gmtb1qfc",
   "tg_tokens": null,
   "tps": 22.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-7u57y4kf-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 64.5 GB at this context; duplicate traces of the same session: 0se2oetd, 3wamyxbo, x3yxi59h; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3624.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 3bit; 32k 3,624 102.1 64.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7u57y4kf",
   "source_url": "https://omlx.ai/benchmarks/performance/7u57y4kf",
   "tg_tokens": null,
   "tps": 102.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-7u57y4kf-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 65.3 GB at this context; duplicate traces of the same session: 0se2oetd, 3wamyxbo, x3yxi59h; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4137.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 3bit; 64k 4,137 50.3 65.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7u57y4kf",
   "source_url": "https://omlx.ai/benchmarks/performance/7u57y4kf",
   "tg_tokens": null,
   "tps": 50.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-7u57y4kf-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 67.7 GB at this context; duplicate traces of the same session: 0se2oetd, 3wamyxbo, x3yxi59h; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4035.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 3bit; 128k 4,035 73.7 67.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7u57y4kf",
   "source_url": "https://omlx.ai/benchmarks/performance/7u57y4kf",
   "tg_tokens": null,
   "tps": 73.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-7u57y4kf-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.5 GB at this context; duplicate traces of the same session: 0se2oetd, 3wamyxbo, x3yxi59h; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2712.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 3bit; 195k 2,712 62.2 70.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7u57y4kf",
   "source_url": "https://omlx.ai/benchmarks/performance/7u57y4kf",
   "tg_tokens": null,
   "tps": 62.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-t2khxgi9-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 61.5 GB; TTFT 2484 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 412.2,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 3bit; 1k tokens 412.2 PP tok/s 45.0 TG tok/s 2484 TTFT (ms) 61.5 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t2khxgi9",
   "source_url": "https://omlx.ai/benchmarks/performance/t2khxgi9",
   "tg_tokens": null,
   "tps": 45.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pvh8u16y-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 52.8 GB at this context; duplicate traces of the same session: l09ydqtx, rhpzgwry, tcc87xj4, v1loy93y, w4546pbw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 128.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 128.2 13.4 52.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pvh8u16y",
   "source_url": "https://omlx.ai/benchmarks/performance/pvh8u16y",
   "tg_tokens": null,
   "tps": 13.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pvh8u16y-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 55.0 GB at this context; duplicate traces of the same session: l09ydqtx, rhpzgwry, tcc87xj4, v1loy93y, w4546pbw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 594.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 594.2 12.2 55.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pvh8u16y",
   "source_url": "https://omlx.ai/benchmarks/performance/pvh8u16y",
   "tg_tokens": null,
   "tps": 12.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pvh8u16y-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 56.5 GB at this context; duplicate traces of the same session: l09ydqtx, rhpzgwry, tcc87xj4, v1loy93y, w4546pbw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 575.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 575.5 7.8 56.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pvh8u16y",
   "source_url": "https://omlx.ai/benchmarks/performance/pvh8u16y",
   "tg_tokens": null,
   "tps": 7.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pvh8u16y-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 57.6 GB at this context; duplicate traces of the same session: l09ydqtx, rhpzgwry, tcc87xj4, v1loy93y, w4546pbw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 495.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 495.1 10.3 57.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pvh8u16y",
   "source_url": "https://omlx.ai/benchmarks/performance/pvh8u16y",
   "tg_tokens": null,
   "tps": 10.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pvh8u16y-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 61.0 GB at this context; duplicate traces of the same session: l09ydqtx, rhpzgwry, tcc87xj4, v1loy93y, w4546pbw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 470.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 470.8 9.5 61.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pvh8u16y",
   "source_url": "https://omlx.ai/benchmarks/performance/pvh8u16y",
   "tg_tokens": null,
   "tps": 9.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pvh8u16y-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 64.0 GB at this context; duplicate traces of the same session: l09ydqtx, rhpzgwry, tcc87xj4, v1loy93y, w4546pbw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 483.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 483.3 11.6 64.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pvh8u16y",
   "source_url": "https://omlx.ai/benchmarks/performance/pvh8u16y",
   "tg_tokens": null,
   "tps": 11.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-e3ehb0rl-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 50.6 GB at this context; duplicate traces of the same session: 0ge9maa2, 9dbj21pp, ggwy6b4f, giwth0q1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 409.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 409.1 11.2 50.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e3ehb0rl",
   "source_url": "https://omlx.ai/benchmarks/performance/e3ehb0rl",
   "tg_tokens": null,
   "tps": 11.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-e3ehb0rl-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 52.7 GB at this context; duplicate traces of the same session: 0ge9maa2, 9dbj21pp, ggwy6b4f, giwth0q1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 593.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 593.0 9.0 52.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e3ehb0rl",
   "source_url": "https://omlx.ai/benchmarks/performance/e3ehb0rl",
   "tg_tokens": null,
   "tps": 9.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-e3ehb0rl-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 54.4 GB at this context; duplicate traces of the same session: 0ge9maa2, 9dbj21pp, ggwy6b4f, giwth0q1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 494.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 494.2 9.1 54.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e3ehb0rl",
   "source_url": "https://omlx.ai/benchmarks/performance/e3ehb0rl",
   "tg_tokens": null,
   "tps": 9.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-e3ehb0rl-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 57.8 GB at this context; duplicate traces of the same session: 0ge9maa2, 9dbj21pp, ggwy6b4f, giwth0q1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 499.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 499.6 8.9 57.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e3ehb0rl",
   "source_url": "https://omlx.ai/benchmarks/performance/e3ehb0rl",
   "tg_tokens": null,
   "tps": 8.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-e3ehb0rl-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 60.9 GB at this context; duplicate traces of the same session: 0ge9maa2, 9dbj21pp, ggwy6b4f, giwth0q1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 514.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 514.2 10.4 60.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e3ehb0rl",
   "source_url": "https://omlx.ai/benchmarks/performance/e3ehb0rl",
   "tg_tokens": null,
   "tps": 10.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-7yujoopv-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: wmaishqb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1081.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,081 93.1 70.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7yujoopv",
   "source_url": "https://omlx.ai/benchmarks/performance/7yujoopv",
   "tg_tokens": null,
   "tps": 93.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7yujoopv-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 75.2 GB at this context; duplicate traces of the same session: wmaishqb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2545.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,545 80.7 75.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7yujoopv",
   "source_url": "https://omlx.ai/benchmarks/performance/7yujoopv",
   "tg_tokens": null,
   "tps": 80.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jzleix50-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 39.0 GB; TTFT 11241 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 91.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 91.1 PP tok/s 17.6 TG tok/s 11241 TTFT (ms) 39.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jzleix50",
   "source_url": "https://omlx.ai/benchmarks/performance/jzleix50",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6g6q0cni-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 109.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 109.1 16.5 39.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6g6q0cni",
   "source_url": "https://omlx.ai/benchmarks/performance/6g6q0cni",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6g6q0cni-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 294.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 294.2 17.0 39.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6g6q0cni",
   "source_url": "https://omlx.ai/benchmarks/performance/6g6q0cni",
   "tg_tokens": null,
   "tps": 17.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6g6q0cni-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 40.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 280.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 280.0 14.3 40.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6g6q0cni",
   "source_url": "https://omlx.ai/benchmarks/performance/6g6q0cni",
   "tg_tokens": null,
   "tps": 14.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-mhiifnm1-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace, single run at 4k context; peak memory 143.3 GB; TTFT 4062 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1008.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 6bit; 4k tokens 1,008 PP tok/s 64.7 TG tok/s 4062 TTFT (ms) 143.3 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mhiifnm1",
   "source_url": "https://omlx.ai/benchmarks/performance/mhiifnm1",
   "tg_tokens": null,
   "tps": 64.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-dtttebwa-ctx4096",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace, single run at 4k context; peak memory 13.7 GB; TTFT 27647 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 148.2,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 2bit; 4k tokens 148.2 PP tok/s 9.0 TG tok/s 27647 TTFT (ms) 13.7 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dtttebwa",
   "source_url": "https://omlx.ai/benchmarks/performance/dtttebwa",
   "tg_tokens": null,
   "tps": 9.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-9k9sjxs2-ctx4096",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace, single run at 4k context; peak memory 13.7 GB; TTFT 27495 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 149.0,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 2bit; 4k tokens 149.0 PP tok/s 11.1 TG tok/s 27495 TTFT (ms) 13.7 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9k9sjxs2",
   "source_url": "https://omlx.ai/benchmarks/performance/9k9sjxs2",
   "tg_tokens": null,
   "tps": 11.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-17haynyo-ctx4096",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace, single run at 4k context; peak memory 13.7 GB; TTFT 27477 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 149.1,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 2bit; 4k tokens 149.1 PP tok/s 9.9 TG tok/s 27477 TTFT (ms) 13.7 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/17haynyo",
   "source_url": "https://omlx.ai/benchmarks/performance/17haynyo",
   "tg_tokens": null,
   "tps": 9.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-puv5f9bu-ctx4096",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace, single run at 4k context; peak memory 13.7 GB; TTFT 27461 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 149.2,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 2bit; 4k tokens 149.2 PP tok/s 10.9 TG tok/s 27461 TTFT (ms) 13.7 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/puv5f9bu",
   "source_url": "https://omlx.ai/benchmarks/performance/puv5f9bu",
   "tg_tokens": null,
   "tps": 10.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-77575qfu-ctx4096",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace, single run at 4k context; peak memory 13.7 GB; TTFT 28141 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 145.6,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 2bit; 4k tokens 145.6 PP tok/s 10.1 TG tok/s 28141 TTFT (ms) 13.7 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/77575qfu",
   "source_url": "https://omlx.ai/benchmarks/performance/77575qfu",
   "tg_tokens": null,
   "tps": 10.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-5it3eom5-ctx4096",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace, single run at 4k context; peak memory 13.7 GB; TTFT 26770 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 153.0,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 2bit; 4k tokens 153.0 PP tok/s 11.7 TG tok/s 26770 TTFT (ms) 13.7 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5it3eom5",
   "source_url": "https://omlx.ai/benchmarks/performance/5it3eom5",
   "tg_tokens": null,
   "tps": 11.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-m1b3c0nc-ctx4096",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace, single run at 4k context; peak memory 13.7 GB; TTFT 28337 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 144.5,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 2bit; 4k tokens 144.5 PP tok/s 9.3 TG tok/s 28337 TTFT (ms) 13.7 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m1b3c0nc",
   "source_url": "https://omlx.ai/benchmarks/performance/m1b3c0nc",
   "tg_tokens": null,
   "tps": 9.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-330j3vd3-ctx4096",
   "model": "Ternary-Bonsai-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 10.3 GB at this context; duplicate traces of the same session: mmypfe1d, snrg9phs; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 840.6,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-27B-mlx-2bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 2bit; 4k 840.6 47.7 10.3 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/330j3vd3",
   "source_url": "https://omlx.ai/benchmarks/performance/330j3vd3",
   "tg_tokens": null,
   "tps": 47.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-330j3vd3-ctx16384",
   "model": "Ternary-Bonsai-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 10.8 GB at this context; duplicate traces of the same session: mmypfe1d, snrg9phs; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 720.9,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-27B-mlx-2bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 2bit; 16k 720.9 42.0 10.8 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/330j3vd3",
   "source_url": "https://omlx.ai/benchmarks/performance/330j3vd3",
   "tg_tokens": null,
   "tps": 42.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-330j3vd3-ctx131072",
   "model": "Ternary-Bonsai-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 17.9 GB at this context; duplicate traces of the same session: mmypfe1d, snrg9phs; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 394.6,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-27B-mlx-2bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 2bit; 128k 394.6 17.2 17.9 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/330j3vd3",
   "source_url": "https://omlx.ai/benchmarks/performance/330j3vd3",
   "tg_tokens": null,
   "tps": 17.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-zvdyvdio-ctx4096",
   "model": "DeepSeek-R1-Distill-Qwen-32B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 34.3 GB at this context; duplicate traces of the same session: apif3127; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 708.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "DeepSeek-R1-Distill-Qwen-32B-MLX-8Bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 708.5 15.1 34.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zvdyvdio",
   "source_url": "https://omlx.ai/benchmarks/performance/zvdyvdio",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-zvdyvdio-ctx16384",
   "model": "DeepSeek-R1-Distill-Qwen-32B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 37.1 GB at this context; duplicate traces of the same session: apif3127; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 659.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "DeepSeek-R1-Distill-Qwen-32B-MLX-8Bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 659.1 13.8 37.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zvdyvdio",
   "source_url": "https://omlx.ai/benchmarks/performance/zvdyvdio",
   "tg_tokens": null,
   "tps": 13.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-abk2dq2h-ctx1024",
   "model": "NVIDIA-Nemotron-3.5-Super-VL-120B-A12B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 70.0 GB at this context; duplicate traces of the same session: c0au9881, qsdy9ggj; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 769.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-MLX-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k 769.3 72.0 70.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/abk2dq2h",
   "source_url": "https://omlx.ai/benchmarks/performance/abk2dq2h",
   "tg_tokens": null,
   "tps": 72.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-abk2dq2h-ctx4096",
   "model": "NVIDIA-Nemotron-3.5-Super-VL-120B-A12B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: c0au9881, qsdy9ggj; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1758.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-MLX-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 4k 1,758 71.5 72.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/abk2dq2h",
   "source_url": "https://omlx.ai/benchmarks/performance/abk2dq2h",
   "tg_tokens": null,
   "tps": 71.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-abk2dq2h-ctx131072",
   "model": "NVIDIA-Nemotron-3.5-Super-VL-120B-A12B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 73.6 GB at this context; duplicate traces of the same session: c0au9881, qsdy9ggj; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1574.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-MLX-4bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 128k 1,574 59.5 73.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/abk2dq2h",
   "source_url": "https://omlx.ai/benchmarks/performance/abk2dq2h",
   "tg_tokens": null,
   "tps": 59.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-hiyefy1y-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 1k context; peak memory 17.6 GB; TTFT 692 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1479.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Ornith-1.5-9B-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 unknown; 1k tokens 1,479 PP tok/s 16.8 TG tok/s 692 TTFT (ms) 17.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hiyefy1y",
   "source_url": "https://omlx.ai/benchmarks/performance/hiyefy1y",
   "tg_tokens": null,
   "tps": 16.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-6q9lz6ju-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 23.7 GB; TTFT 67219 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 487.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 32k tokens 487.5 PP tok/s 65.5 TG tok/s 67219 TTFT (ms) 23.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6q9lz6ju",
   "source_url": "https://omlx.ai/benchmarks/performance/6q9lz6ju",
   "tg_tokens": null,
   "tps": 65.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-i07h35jd-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 22.6 GB; TTFT 26272 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 623.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 16k tokens 623.6 PP tok/s 77.0 TG tok/s 26272 TTFT (ms) 22.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i07h35jd",
   "source_url": "https://omlx.ai/benchmarks/performance/i07h35jd",
   "tg_tokens": null,
   "tps": 77.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-nvvyuggf-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: blelkdl1, l5818jl1, psm7l8i2; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1005.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 1k 1,005 103.0 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nvvyuggf",
   "source_url": "https://omlx.ai/benchmarks/performance/nvvyuggf",
   "tg_tokens": null,
   "tps": 103.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-nvvyuggf-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: blelkdl1, l5818jl1, psm7l8i2; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1100.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k 1,100 99.4 21.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nvvyuggf",
   "source_url": "https://omlx.ai/benchmarks/performance/nvvyuggf",
   "tg_tokens": null,
   "tps": 99.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-nvvyuggf-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: blelkdl1, l5818jl1, psm7l8i2; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 961.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 16k 961.3 87.5 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nvvyuggf",
   "source_url": "https://omlx.ai/benchmarks/performance/nvvyuggf",
   "tg_tokens": null,
   "tps": 87.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-nvvyuggf-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: blelkdl1, l5818jl1, psm7l8i2; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 813.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 32k 813.7 75.7 21.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nvvyuggf",
   "source_url": "https://omlx.ai/benchmarks/performance/nvvyuggf",
   "tg_tokens": null,
   "tps": 75.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-m2qk4uis-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: fy40yqhr; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 840.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 1k 840.3 101.6 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m2qk4uis",
   "source_url": "https://omlx.ai/benchmarks/performance/m2qk4uis",
   "tg_tokens": null,
   "tps": 101.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-m2qk4uis-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: fy40yqhr; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1101.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k 1,101 99.1 21.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m2qk4uis",
   "source_url": "https://omlx.ai/benchmarks/performance/m2qk4uis",
   "tg_tokens": null,
   "tps": 99.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-j5ytzjei-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 35.5 GB; TTFT 242 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4228.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k tokens 4,228 PP tok/s 103.0 TG tok/s 242 TTFT (ms) 35.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j5ytzjei",
   "source_url": "https://omlx.ai/benchmarks/performance/j5ytzjei",
   "tg_tokens": null,
   "tps": 103.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8pomd4dg-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 35.5 GB; TTFT 247 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4138.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k tokens 4,138 PP tok/s 100.4 TG tok/s 247 TTFT (ms) 35.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8pomd4dg",
   "source_url": "https://omlx.ai/benchmarks/performance/8pomd4dg",
   "tg_tokens": null,
   "tps": 100.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-75n9ne5c-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: lbd9o0bj, vp6ulatw; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2964.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,964 69.3 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/75n9ne5c",
   "source_url": "https://omlx.ai/benchmarks/performance/75n9ne5c",
   "tg_tokens": null,
   "tps": 69.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-75n9ne5c-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: lbd9o0bj, vp6ulatw; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2566.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 2,566 64.5 21.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/75n9ne5c",
   "source_url": "https://omlx.ai/benchmarks/performance/75n9ne5c",
   "tg_tokens": null,
   "tps": 64.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-75n9ne5c-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: lbd9o0bj, vp6ulatw; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1947.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 1,947 53.5 22.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/75n9ne5c",
   "source_url": "https://omlx.ai/benchmarks/performance/75n9ne5c",
   "tg_tokens": null,
   "tps": 53.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-h7uem6kn-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 0us2qt90, btqama7b, ezz9q7av, fb8182ub; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1918.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 1,918 71.5 20.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h7uem6kn",
   "source_url": "https://omlx.ai/benchmarks/performance/h7uem6kn",
   "tg_tokens": null,
   "tps": 71.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-h7uem6kn-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 0us2qt90, btqama7b, ezz9q7av, fb8182ub; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2191.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 2,191 49.8 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h7uem6kn",
   "source_url": "https://omlx.ai/benchmarks/performance/h7uem6kn",
   "tg_tokens": null,
   "tps": 49.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-h7uem6kn-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 0us2qt90, btqama7b, ezz9q7av, fb8182ub; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2230.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,230 52.2 21.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h7uem6kn",
   "source_url": "https://omlx.ai/benchmarks/performance/h7uem6kn",
   "tg_tokens": null,
   "tps": 52.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-h7uem6kn-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 0us2qt90, btqama7b, ezz9q7av, fb8182ub; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1754.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 1,754 62.8 21.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h7uem6kn",
   "source_url": "https://omlx.ai/benchmarks/performance/h7uem6kn",
   "tg_tokens": null,
   "tps": 62.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-h7uem6kn-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: 0us2qt90, btqama7b, ezz9q7av, fb8182ub; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1400.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 1,400 44.7 22.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h7uem6kn",
   "source_url": "https://omlx.ai/benchmarks/performance/h7uem6kn",
   "tg_tokens": null,
   "tps": 44.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-5u6i23lm-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 6u9fed35, x8lbbw4x, xrfpqc4n; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 144.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 1k 144.7 21.4 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5u6i23lm",
   "source_url": "https://omlx.ai/benchmarks/performance/5u6i23lm",
   "tg_tokens": null,
   "tps": 21.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-5u6i23lm-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 6u9fed35, x8lbbw4x, xrfpqc4n; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 145.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k 145.3 20.9 19.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5u6i23lm",
   "source_url": "https://omlx.ai/benchmarks/performance/5u6i23lm",
   "tg_tokens": null,
   "tps": 20.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-5u6i23lm-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 6u9fed35, x8lbbw4x, xrfpqc4n; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 138.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 16k 138.7 19.2 19.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5u6i23lm",
   "source_url": "https://omlx.ai/benchmarks/performance/5u6i23lm",
   "tg_tokens": null,
   "tps": 19.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-5u6i23lm-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 6u9fed35, x8lbbw4x, xrfpqc4n; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 32k 130.6 18.7 20.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5u6i23lm",
   "source_url": "https://omlx.ai/benchmarks/performance/5u6i23lm",
   "tg_tokens": null,
   "tps": 18.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-adyd2g02-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: tpm4m4nx; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 143.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 1k 143.5 21.7 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/adyd2g02",
   "source_url": "https://omlx.ai/benchmarks/performance/adyd2g02",
   "tg_tokens": null,
   "tps": 21.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-adyd2g02-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: tpm4m4nx; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 145.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k 145.1 20.5 19.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/adyd2g02",
   "source_url": "https://omlx.ai/benchmarks/performance/adyd2g02",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-k7v95s65-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: soien4yp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 371.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 8bit; 1k 371.5 19.4 33.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k7v95s65",
   "source_url": "https://omlx.ai/benchmarks/performance/k7v95s65",
   "tg_tokens": null,
   "tps": 19.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-k7v95s65-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: soien4yp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 389.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 8bit; 4k 389.8 18.1 34.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k7v95s65",
   "source_url": "https://omlx.ai/benchmarks/performance/k7v95s65",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ru513e29-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: f493vbrd; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 345.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 8bit; 1k 345.8 18.3 33.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ru513e29",
   "source_url": "https://omlx.ai/benchmarks/performance/ru513e29",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ru513e29-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: f493vbrd; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 387.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 8bit; 4k 387.4 16.7 34.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ru513e29",
   "source_url": "https://omlx.ai/benchmarks/performance/ru513e29",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-c3s6bf6q-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 43.5 GB; TTFT 247544 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 529.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k tokens 529.5 PP tok/s 40.9 TG tok/s 247544 TTFT (ms) 43.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c3s6bf6q",
   "source_url": "https://omlx.ai/benchmarks/performance/c3s6bf6q",
   "tg_tokens": null,
   "tps": 40.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ckgxdnfq-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 43.5 GB; TTFT 240575 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 544.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k tokens 544.8 PP tok/s 42.6 TG tok/s 240575 TTFT (ms) 43.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ckgxdnfq",
   "source_url": "https://omlx.ai/benchmarks/performance/ckgxdnfq",
   "tg_tokens": null,
   "tps": 42.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-avte0o22-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 43.5 GB; TTFT 253242 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 517.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k tokens 517.6 PP tok/s 43.2 TG tok/s 253242 TTFT (ms) 43.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/avte0o22",
   "source_url": "https://omlx.ai/benchmarks/performance/avte0o22",
   "tg_tokens": null,
   "tps": 43.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-kgr7rj0c-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 43.5 GB; TTFT 240848 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 544.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k tokens 544.2 PP tok/s 41.4 TG tok/s 240848 TTFT (ms) 43.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kgr7rj0c",
   "source_url": "https://omlx.ai/benchmarks/performance/kgr7rj0c",
   "tg_tokens": null,
   "tps": 41.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xd8ho0se-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 40.0 GB; TTFT 3662 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1119.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 1,119 PP tok/s 58.2 TG tok/s 3662 TTFT (ms) 40.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xd8ho0se",
   "source_url": "https://omlx.ai/benchmarks/performance/xd8ho0se",
   "tg_tokens": null,
   "tps": 58.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-7ulfbo6q-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 64.2 GB; TTFT 10627 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3083.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 3bit; 32k tokens 3,083 PP tok/s 53.8 TG tok/s 10627 TTFT (ms) 64.2 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ulfbo6q",
   "source_url": "https://omlx.ai/benchmarks/performance/7ulfbo6q",
   "tg_tokens": null,
   "tps": 53.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-v0hq8kj0-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 52.8 GB at this context; duplicate traces of the same session: k5zl67h9, lrl7k51h, mr8h4dms, sxwa6s6v, tlqdodba; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 131.2 11.5 52.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v0hq8kj0",
   "source_url": "https://omlx.ai/benchmarks/performance/v0hq8kj0",
   "tg_tokens": null,
   "tps": 11.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-v0hq8kj0-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 55.0 GB at this context; duplicate traces of the same session: k5zl67h9, lrl7k51h, mr8h4dms, sxwa6s6v, tlqdodba; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 661.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 661.8 13.8 55.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v0hq8kj0",
   "source_url": "https://omlx.ai/benchmarks/performance/v0hq8kj0",
   "tg_tokens": null,
   "tps": 13.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-v0hq8kj0-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 55.9 GB at this context; duplicate traces of the same session: k5zl67h9, lrl7k51h, mr8h4dms, sxwa6s6v, tlqdodba; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 645.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 645.4 11.6 55.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v0hq8kj0",
   "source_url": "https://omlx.ai/benchmarks/performance/v0hq8kj0",
   "tg_tokens": null,
   "tps": 11.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-v0hq8kj0-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 57.6 GB at this context; duplicate traces of the same session: k5zl67h9, lrl7k51h, mr8h4dms, sxwa6s6v, tlqdodba; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 601.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 601.9 10.8 57.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v0hq8kj0",
   "source_url": "https://omlx.ai/benchmarks/performance/v0hq8kj0",
   "tg_tokens": null,
   "tps": 10.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-v0hq8kj0-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 61.0 GB at this context; duplicate traces of the same session: k5zl67h9, lrl7k51h, mr8h4dms, sxwa6s6v, tlqdodba; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 508.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 508.0 10.9 61.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v0hq8kj0",
   "source_url": "https://omlx.ai/benchmarks/performance/v0hq8kj0",
   "tg_tokens": null,
   "tps": 10.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-v0hq8kj0-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 64.0 GB at this context; duplicate traces of the same session: k5zl67h9, lrl7k51h, mr8h4dms, sxwa6s6v, tlqdodba; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 536.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 536.3 12.3 64.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v0hq8kj0",
   "source_url": "https://omlx.ai/benchmarks/performance/v0hq8kj0",
   "tg_tokens": null,
   "tps": 12.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fzmzp0m9-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 68.7 GB at this context; duplicate traces of the same session: 2o04avbq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 135.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 135.1 21.1 68.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fzmzp0m9",
   "source_url": "https://omlx.ai/benchmarks/performance/fzmzp0m9",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fzmzp0m9-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.6 GB at this context; duplicate traces of the same session: 2o04avbq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 360.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 360.6 20.0 69.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fzmzp0m9",
   "source_url": "https://omlx.ai/benchmarks/performance/fzmzp0m9",
   "tg_tokens": null,
   "tps": 20.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fzmzp0m9-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 2o04avbq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 562.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 562.3 18.0 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fzmzp0m9",
   "source_url": "https://omlx.ai/benchmarks/performance/fzmzp0m9",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fzmzp0m9-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 2o04avbq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 568.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 568.6 14.7 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fzmzp0m9",
   "source_url": "https://omlx.ai/benchmarks/performance/fzmzp0m9",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fzmzp0m9-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.3 GB at this context; duplicate traces of the same session: 2o04avbq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 546.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 546.3 13.8 72.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fzmzp0m9",
   "source_url": "https://omlx.ai/benchmarks/performance/fzmzp0m9",
   "tg_tokens": null,
   "tps": 13.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fzmzp0m9-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.5 GB at this context; duplicate traces of the same session: 2o04avbq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 468.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 468.6 13.9 73.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fzmzp0m9",
   "source_url": "https://omlx.ai/benchmarks/performance/fzmzp0m9",
   "tg_tokens": null,
   "tps": 13.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fzmzp0m9-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 77.1 GB at this context; duplicate traces of the same session: 2o04avbq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 440.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 440.9 17.8 77.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fzmzp0m9",
   "source_url": "https://omlx.ai/benchmarks/performance/fzmzp0m9",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fzmzp0m9-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.9 GB at this context; duplicate traces of the same session: 2o04avbq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 474.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 474.8 15.1 79.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fzmzp0m9",
   "source_url": "https://omlx.ai/benchmarks/performance/fzmzp0m9",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jsitjxqb-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.0 GB at this context; duplicate traces of the same session: 8x6bavzh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 109.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 109.1 16.5 39.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jsitjxqb",
   "source_url": "https://omlx.ai/benchmarks/performance/jsitjxqb",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jsitjxqb-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.4 GB at this context; duplicate traces of the same session: 8x6bavzh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 294.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 294.2 17.0 39.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jsitjxqb",
   "source_url": "https://omlx.ai/benchmarks/performance/jsitjxqb",
   "tg_tokens": null,
   "tps": 17.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jsitjxqb-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 40.0 GB at this context; duplicate traces of the same session: 8x6bavzh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 280.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 280.0 14.3 40.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jsitjxqb",
   "source_url": "https://omlx.ai/benchmarks/performance/jsitjxqb",
   "tg_tokens": null,
   "tps": 14.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-p1p887nz-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: gdaodhip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1026.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3___8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,026 33.1 17.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p1p887nz",
   "source_url": "https://omlx.ai/benchmarks/performance/p1p887nz",
   "tg_tokens": null,
   "tps": 33.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-p1p887nz-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: gdaodhip; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 898.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3___8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 898.1 31.2 17.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p1p887nz",
   "source_url": "https://omlx.ai/benchmarks/performance/p1p887nz",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-sye16mqz-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: kpodzr1f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 992.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3___8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 992.1 32.8 17.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sye16mqz",
   "source_url": "https://omlx.ai/benchmarks/performance/sye16mqz",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-sye16mqz-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: kpodzr1f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 910.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3___8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 910.7 31.1 17.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sye16mqz",
   "source_url": "https://omlx.ai/benchmarks/performance/sye16mqz",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-irg6lmxh-ctx4096",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace, single run at 4k context; peak memory 13.7 GB; TTFT 26219 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 156.2,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 2bit; 4k tokens 156.2 PP tok/s 11.0 TG tok/s 26219 TTFT (ms) 13.7 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/irg6lmxh",
   "source_url": "https://omlx.ai/benchmarks/performance/irg6lmxh",
   "tg_tokens": null,
   "tps": 11.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-cd064fiz-ctx4096",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace, single run at 4k context; peak memory 13.7 GB; TTFT 26915 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 152.2,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 2bit; 4k tokens 152.2 PP tok/s 11.6 TG tok/s 26915 TTFT (ms) 13.7 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cd064fiz",
   "source_url": "https://omlx.ai/benchmarks/performance/cd064fiz",
   "tg_tokens": null,
   "tps": 11.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-p47vn05p-ctx4096",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace, single run at 4k context; peak memory 13.7 GB; TTFT 26899 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 152.3,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 2bit; 4k tokens 152.3 PP tok/s 11.0 TG tok/s 26899 TTFT (ms) 13.7 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p47vn05p",
   "source_url": "https://omlx.ai/benchmarks/performance/p47vn05p",
   "tg_tokens": null,
   "tps": 11.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ajsr3olx-ctx1024",
   "model": "gemma-4-12b-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 8.5 GB; TTFT 1493 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 686.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-12B-it-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 24 GB \u00b7 4bit; 1k tokens 686.0 PP tok/s 26.2 TG tok/s 1493 TTFT (ms) 8.5 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ajsr3olx",
   "source_url": "https://omlx.ai/benchmarks/performance/ajsr3olx",
   "tg_tokens": null,
   "tps": 26.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-j2bndk2b-ctx1024",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 3.6 GB at this context; duplicate traces of the same session: 3bxxmn7p; ",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2295.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E2B-it-MLX-4bit on M1 Max; M1 Max \u00b7 38 GB \u00b7 4bit; 1k 2,295 80.0 3.6 GB current",
   "ram_gb": 38,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j2bndk2b",
   "source_url": "https://omlx.ai/benchmarks/performance/j2bndk2b",
   "tg_tokens": null,
   "tps": 80.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-j2bndk2b-ctx4096",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 3.7 GB at this context; duplicate traces of the same session: 3bxxmn7p; ",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2384.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E2B-it-MLX-4bit on M1 Max; M1 Max \u00b7 38 GB \u00b7 4bit; 4k 2,384 82.5 3.7 GB view",
   "ram_gb": 38,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j2bndk2b",
   "source_url": "https://omlx.ai/benchmarks/performance/j2bndk2b",
   "tg_tokens": null,
   "tps": 82.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1xggsbi4-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 205.7 GB at this context; duplicate traces of the same session: 08z1y66k, 654ghfn3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 279.7,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 1k 279.7 28.7 205.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1xggsbi4",
   "source_url": "https://omlx.ai/benchmarks/performance/1xggsbi4",
   "tg_tokens": null,
   "tps": 28.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-1xggsbi4-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 206.1 GB at this context; duplicate traces of the same session: 08z1y66k, 654ghfn3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 476.1,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 4k 476.1 31.4 206.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1xggsbi4",
   "source_url": "https://omlx.ai/benchmarks/performance/1xggsbi4",
   "tg_tokens": null,
   "tps": 31.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-1xggsbi4-ctx32768",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 206.1 GB at this context; duplicate traces of the same session: 08z1y66k, 654ghfn3; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 504.3,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 32k 504.3 42.0 206.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1xggsbi4",
   "source_url": "https://omlx.ai/benchmarks/performance/1xggsbi4",
   "tg_tokens": null,
   "tps": 42.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-jhh9lper-ctx4096",
   "model": "GLM-5.3-Flash-Abliterated",
   "notes": "MLX; oMLX community benchmark trace, single run at 4k context; peak memory 179.1 GB; TTFT 7130 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 574.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-Abliterated-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k tokens 574.5 PP tok/s 38.8 TG tok/s 7130 TTFT (ms) 179.1 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jhh9lper",
   "source_url": "https://omlx.ai/benchmarks/performance/jhh9lper",
   "tg_tokens": null,
   "tps": 38.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xy2d3acq-ctx4096",
   "model": "GLM-5.3-Flash-Abliterated",
   "notes": "MLX; oMLX community benchmark trace, single run at 4k context; peak memory 179.1 GB; TTFT 8213 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 498.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-Abliterated-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k tokens 498.7 PP tok/s 22.0 TG tok/s 8213 TTFT (ms) 179.1 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xy2d3acq",
   "source_url": "https://omlx.ai/benchmarks/performance/xy2d3acq",
   "tg_tokens": null,
   "tps": 22.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-f23f7ck3-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 98.4 GB at this context; duplicate traces of the same session: 6qa6096f; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 64.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "GLM-5.3-Flash-oQ2.5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 2bit; 1k 64.0 21.8 98.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f23f7ck3",
   "source_url": "https://omlx.ai/benchmarks/performance/f23f7ck3",
   "tg_tokens": null,
   "tps": 21.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-f23f7ck3-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.0 GB at this context; duplicate traces of the same session: 6qa6096f; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 298.5,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "GLM-5.3-Flash-oQ2.5e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 2bit; 8k 298.5 20.6 100.0 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f23f7ck3",
   "source_url": "https://omlx.ai/benchmarks/performance/f23f7ck3",
   "tg_tokens": null,
   "tps": 20.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-mj2drxp5-ctx4096",
   "model": "Qwen3.6-35B-A3B-4bit-MTPLX-Optimized-Speed",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: x4ix4q5y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1604.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit-MTPLX-Optimized-Speed on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,604 103.0 21.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mj2drxp5",
   "source_url": "https://omlx.ai/benchmarks/performance/mj2drxp5",
   "tg_tokens": null,
   "tps": 103.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-mj2drxp5-ctx16384",
   "model": "Qwen3.6-35B-A3B-4bit-MTPLX-Optimized-Speed",
   "notes": "oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: x4ix4q5y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1494.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit-MTPLX-Optimized-Speed on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,494 101.8 22.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mj2drxp5",
   "source_url": "https://omlx.ai/benchmarks/performance/mj2drxp5",
   "tg_tokens": null,
   "tps": 101.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-nneic52k-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: 42ikz1we, peq2oh3g, svl1zz9b; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 372.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 372.3 12.7 16.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nneic52k",
   "source_url": "https://omlx.ai/benchmarks/performance/nneic52k",
   "tg_tokens": null,
   "tps": 12.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-nneic52k-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 42ikz1we, peq2oh3g, svl1zz9b; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 376.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 376.9 12.4 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nneic52k",
   "source_url": "https://omlx.ai/benchmarks/performance/nneic52k",
   "tg_tokens": null,
   "tps": 12.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-nneic52k-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 42ikz1we, peq2oh3g, svl1zz9b; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 368.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 368.6 13.1 17.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nneic52k",
   "source_url": "https://omlx.ai/benchmarks/performance/nneic52k",
   "tg_tokens": null,
   "tps": 13.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-nneic52k-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 42ikz1we, peq2oh3g, svl1zz9b; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 331.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 331.0 12.0 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nneic52k",
   "source_url": "https://omlx.ai/benchmarks/performance/nneic52k",
   "tg_tokens": null,
   "tps": 12.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-nneic52k-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 42ikz1we, peq2oh3g, svl1zz9b; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 318.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 318.8 11.2 18.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nneic52k",
   "source_url": "https://omlx.ai/benchmarks/performance/nneic52k",
   "tg_tokens": null,
   "tps": 11.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-nneic52k-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 42ikz1we, peq2oh3g, svl1zz9b; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 303.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 303.8 11.1 20.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nneic52k",
   "source_url": "https://omlx.ai/benchmarks/performance/nneic52k",
   "tg_tokens": null,
   "tps": 11.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-b0w4xdic-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: pcxz8c4z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 242.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 242.3 28.1 17.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b0w4xdic",
   "source_url": "https://omlx.ai/benchmarks/performance/b0w4xdic",
   "tg_tokens": null,
   "tps": 28.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-b0w4xdic-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.9 GB at this context; duplicate traces of the same session: pcxz8c4z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 237.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 237.0 26.1 17.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b0w4xdic",
   "source_url": "https://omlx.ai/benchmarks/performance/b0w4xdic",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-oka4kbc4-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 2onw2qzk, 3y9l8jqf, 65gp9csl, svf9td7s; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1159.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 1k 1,159 30.5 28.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oka4kbc4",
   "source_url": "https://omlx.ai/benchmarks/performance/oka4kbc4",
   "tg_tokens": null,
   "tps": 30.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-oka4kbc4-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: 2onw2qzk, 3y9l8jqf, 65gp9csl, svf9td7s; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1361.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 4k 1,361 29.8 29.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oka4kbc4",
   "source_url": "https://omlx.ai/benchmarks/performance/oka4kbc4",
   "tg_tokens": null,
   "tps": 29.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-oka4kbc4-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.3 GB at this context; duplicate traces of the same session: 2onw2qzk, 3y9l8jqf, 65gp9csl, svf9td7s; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1128.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 64k 1,128 26.7 33.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oka4kbc4",
   "source_url": "https://omlx.ai/benchmarks/performance/oka4kbc4",
   "tg_tokens": null,
   "tps": 26.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-oka4kbc4-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: 2onw2qzk, 3y9l8jqf, 65gp9csl, svf9td7s; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 911.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 128k 911.7 24.0 37.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oka4kbc4",
   "source_url": "https://omlx.ai/benchmarks/performance/oka4kbc4",
   "tg_tokens": null,
   "tps": 24.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-oka4kbc4-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 41.5 GB at this context; duplicate traces of the same session: 2onw2qzk, 3y9l8jqf, 65gp9csl, svf9td7s; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 753.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 195k 753.2 21.8 41.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oka4kbc4",
   "source_url": "https://omlx.ai/benchmarks/performance/oka4kbc4",
   "tg_tokens": null,
   "tps": 21.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rm4qtjjv-ctx4096",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 40.6 GB; TTFT 11439 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 358.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k tokens 358.1 PP tok/s 16.3 TG tok/s 11439 TTFT (ms) 40.6 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rm4qtjjv",
   "source_url": "https://omlx.ai/benchmarks/performance/rm4qtjjv",
   "tg_tokens": null,
   "tps": 16.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xjlh1d2m-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 42.8 GB; TTFT 240859 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 544.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k tokens 544.2 PP tok/s 3.3 TG tok/s 240859 TTFT (ms) 42.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xjlh1d2m",
   "source_url": "https://omlx.ai/benchmarks/performance/xjlh1d2m",
   "tg_tokens": null,
   "tps": 3.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-xr7wi443-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: ujbdvp3w; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 832.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 8bit; 1k 832.2 51.2 33.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xr7wi443",
   "source_url": "https://omlx.ai/benchmarks/performance/xr7wi443",
   "tg_tokens": null,
   "tps": 51.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-xr7wi443-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: ujbdvp3w; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1522.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 8bit; 4k 1,522 73.4 34.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xr7wi443",
   "source_url": "https://omlx.ai/benchmarks/performance/xr7wi443",
   "tg_tokens": null,
   "tps": 73.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-3zom4q4u-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 192.6 GB; TTFT 3134 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1307.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-Flash-Next-MLX-8bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k tokens 1,307 PP tok/s 56.8 TG tok/s 3134 TTFT (ms) 192.6 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3zom4q4u",
   "source_url": "https://omlx.ai/benchmarks/performance/3zom4q4u",
   "tg_tokens": null,
   "tps": 56.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-y01zdun1-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 192.6 GB; TTFT 3093 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1324.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-Flash-Next-MLX-8bit-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k tokens 1,324 PP tok/s 84.1 TG tok/s 3093 TTFT (ms) 192.6 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y01zdun1",
   "source_url": "https://omlx.ai/benchmarks/performance/y01zdun1",
   "tg_tokens": null,
   "tps": 84.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0ysp3h8c-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 64.0 GB at this context; duplicate traces of the same session: dms382z0, ehgx4uh9, i3s8dfps, rjws2517, xboih0si; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 115.8,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 115.8 11.0 64.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ysp3h8c",
   "source_url": "https://omlx.ai/benchmarks/performance/0ysp3h8c",
   "tg_tokens": null,
   "tps": 11.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0ysp3h8c-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 64.9 GB at this context; duplicate traces of the same session: dms382z0, ehgx4uh9, i3s8dfps, rjws2517, xboih0si; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 336.6,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 336.6 11.9 64.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ysp3h8c",
   "source_url": "https://omlx.ai/benchmarks/performance/0ysp3h8c",
   "tg_tokens": null,
   "tps": 11.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0ysp3h8c-ctx8192",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 66.2 GB at this context; duplicate traces of the same session: dms382z0, ehgx4uh9, i3s8dfps, rjws2517, xboih0si; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 547.8,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 547.8 12.9 66.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ysp3h8c",
   "source_url": "https://omlx.ai/benchmarks/performance/0ysp3h8c",
   "tg_tokens": null,
   "tps": 12.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0ysp3h8c-ctx65536",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 68.8 GB at this context; duplicate traces of the same session: dms382z0, ehgx4uh9, i3s8dfps, rjws2517, xboih0si; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 499.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 499.4 9.1 68.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ysp3h8c",
   "source_url": "https://omlx.ai/benchmarks/performance/0ysp3h8c",
   "tg_tokens": null,
   "tps": 9.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0ysp3h8c-ctx131072",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 72.5 GB at this context; duplicate traces of the same session: dms382z0, ehgx4uh9, i3s8dfps, rjws2517, xboih0si; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 469.3,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 469.3 10.4 72.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ysp3h8c",
   "source_url": "https://omlx.ai/benchmarks/performance/0ysp3h8c",
   "tg_tokens": null,
   "tps": 10.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0ysp3h8c-ctx199680",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 75.2 GB at this context; duplicate traces of the same session: dms382z0, ehgx4uh9, i3s8dfps, rjws2517, xboih0si; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 474.8,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 195k 474.8 10.8 75.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ysp3h8c",
   "source_url": "https://omlx.ai/benchmarks/performance/0ysp3h8c",
   "tg_tokens": null,
   "tps": 10.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gvfnhccq-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.4 GB at this context; duplicate traces of the same session: 1hh79z0t, 5izo6mmh, 5ufoh6q3, ma8j46ap, sw4jk86x, v3ur9n67; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1017.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,017 46.4 100.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gvfnhccq",
   "source_url": "https://omlx.ai/benchmarks/performance/gvfnhccq",
   "tg_tokens": null,
   "tps": 46.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gvfnhccq-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: 1hh79z0t, 5izo6mmh, 5ufoh6q3, ma8j46ap, sw4jk86x, v3ur9n67; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1352.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 1,352 91.5 101.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gvfnhccq",
   "source_url": "https://omlx.ai/benchmarks/performance/gvfnhccq",
   "tg_tokens": null,
   "tps": 91.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gvfnhccq-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: 1hh79z0t, 5izo6mmh, 5ufoh6q3, ma8j46ap, sw4jk86x, v3ur9n67; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1375.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 1,375 99.7 101.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gvfnhccq",
   "source_url": "https://omlx.ai/benchmarks/performance/gvfnhccq",
   "tg_tokens": null,
   "tps": 99.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-gvfnhccq-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.2 GB at this context; duplicate traces of the same session: 1hh79z0t, 5izo6mmh, 5ufoh6q3, ma8j46ap, sw4jk86x, v3ur9n67; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1364.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 1,364 83.8 102.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gvfnhccq",
   "source_url": "https://omlx.ai/benchmarks/performance/gvfnhccq",
   "tg_tokens": null,
   "tps": 83.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gvfnhccq-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.1 GB at this context; duplicate traces of the same session: 1hh79z0t, 5izo6mmh, 5ufoh6q3, ma8j46ap, sw4jk86x, v3ur9n67; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1348.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 64k 1,348 81.1 104.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gvfnhccq",
   "source_url": "https://omlx.ai/benchmarks/performance/gvfnhccq",
   "tg_tokens": null,
   "tps": 81.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gvfnhccq-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 108.6 GB at this context; duplicate traces of the same session: 1hh79z0t, 5izo6mmh, 5ufoh6q3, ma8j46ap, sw4jk86x, v3ur9n67; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1318.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 128k 1,318 66.8 108.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gvfnhccq",
   "source_url": "https://omlx.ai/benchmarks/performance/gvfnhccq",
   "tg_tokens": null,
   "tps": 66.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-gvfnhccq-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.0 GB at this context; duplicate traces of the same session: 1hh79z0t, 5izo6mmh, 5ufoh6q3, ma8j46ap, sw4jk86x, v3ur9n67; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1303.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 195k 1,303 60.5 110.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gvfnhccq",
   "source_url": "https://omlx.ai/benchmarks/performance/gvfnhccq",
   "tg_tokens": null,
   "tps": 60.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-7qe7cel1-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 101.3 GB; TTFT 2982 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1374.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k tokens 1,374 PP tok/s 79.1 TG tok/s 2982 TTFT (ms) 101.3 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7qe7cel1",
   "source_url": "https://omlx.ai/benchmarks/performance/7qe7cel1",
   "tg_tokens": null,
   "tps": 79.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-3438kvtd-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 101.3 GB; TTFT 4947 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 828.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k tokens 828.0 PP tok/s 26.8 TG tok/s 4947 TTFT (ms) 101.3 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3438kvtd",
   "source_url": "https://omlx.ai/benchmarks/performance/3438kvtd",
   "tg_tokens": null,
   "tps": 26.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nf0b7n09-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 68.7 GB at this context; duplicate traces of the same session: 7cjh3j0e, 7oee9kg8, ayvqhsx3, vsjn1n0f, ysfalho8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 135.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 135.1 21.1 68.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nf0b7n09",
   "source_url": "https://omlx.ai/benchmarks/performance/nf0b7n09",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nf0b7n09-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.6 GB at this context; duplicate traces of the same session: 7cjh3j0e, 7oee9kg8, ayvqhsx3, vsjn1n0f, ysfalho8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 360.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 360.6 20.0 69.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nf0b7n09",
   "source_url": "https://omlx.ai/benchmarks/performance/nf0b7n09",
   "tg_tokens": null,
   "tps": 20.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nf0b7n09-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 7cjh3j0e, 7oee9kg8, ayvqhsx3, vsjn1n0f, ysfalho8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 562.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 562.3 18.0 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nf0b7n09",
   "source_url": "https://omlx.ai/benchmarks/performance/nf0b7n09",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nf0b7n09-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 7cjh3j0e, 7oee9kg8, ayvqhsx3, vsjn1n0f, ysfalho8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 568.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 568.6 14.7 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nf0b7n09",
   "source_url": "https://omlx.ai/benchmarks/performance/nf0b7n09",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nf0b7n09-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.3 GB at this context; duplicate traces of the same session: 7cjh3j0e, 7oee9kg8, ayvqhsx3, vsjn1n0f, ysfalho8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 546.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 546.3 13.8 72.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nf0b7n09",
   "source_url": "https://omlx.ai/benchmarks/performance/nf0b7n09",
   "tg_tokens": null,
   "tps": 13.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nf0b7n09-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.5 GB at this context; duplicate traces of the same session: 7cjh3j0e, 7oee9kg8, ayvqhsx3, vsjn1n0f, ysfalho8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 468.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 468.6 13.9 73.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nf0b7n09",
   "source_url": "https://omlx.ai/benchmarks/performance/nf0b7n09",
   "tg_tokens": null,
   "tps": 13.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nf0b7n09-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 77.1 GB at this context; duplicate traces of the same session: 7cjh3j0e, 7oee9kg8, ayvqhsx3, vsjn1n0f, ysfalho8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 440.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 440.9 17.8 77.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nf0b7n09",
   "source_url": "https://omlx.ai/benchmarks/performance/nf0b7n09",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-nf0b7n09-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 79.9 GB at this context; duplicate traces of the same session: 7cjh3j0e, 7oee9kg8, ayvqhsx3, vsjn1n0f, ysfalho8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 474.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 474.8 15.1 79.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nf0b7n09",
   "source_url": "https://omlx.ai/benchmarks/performance/nf0b7n09",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-08twe39k-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: 1i1znmwo, ldxvbakf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1068.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,068 89.7 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/08twe39k",
   "source_url": "https://omlx.ai/benchmarks/performance/08twe39k",
   "tg_tokens": null,
   "tps": 89.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-08twe39k-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 1i1znmwo, ldxvbakf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2128.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,128 72.2 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/08twe39k",
   "source_url": "https://omlx.ai/benchmarks/performance/08twe39k",
   "tg_tokens": null,
   "tps": 72.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-08twe39k-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 75.2 GB at this context; duplicate traces of the same session: 1i1znmwo, ldxvbakf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2553.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,553 68.0 75.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/08twe39k",
   "source_url": "https://omlx.ai/benchmarks/performance/08twe39k",
   "tg_tokens": null,
   "tps": 68.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8fo2if2x-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 78.1 GB; TTFT 56528 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2319.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k tokens 2,319 PP tok/s 61.2 TG tok/s 56528 TTFT (ms) 78.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8fo2if2x",
   "source_url": "https://omlx.ai/benchmarks/performance/8fo2if2x",
   "tg_tokens": null,
   "tps": 61.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-m35pka2x-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.5 GB at this context; duplicate traces of the same session: 2l0decx5, qhexaxrw, xpaz8msc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1026.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 1,026 97.9 142.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m35pka2x",
   "source_url": "https://omlx.ai/benchmarks/performance/m35pka2x",
   "tg_tokens": null,
   "tps": 97.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-m35pka2x-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: 2l0decx5, qhexaxrw, xpaz8msc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1353.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 1,353 89.5 143.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m35pka2x",
   "source_url": "https://omlx.ai/benchmarks/performance/m35pka2x",
   "tg_tokens": null,
   "tps": 89.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-m35pka2x-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.4 GB at this context; duplicate traces of the same session: 2l0decx5, qhexaxrw, xpaz8msc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1353.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 8k 1,353 90.4 143.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m35pka2x",
   "source_url": "https://omlx.ai/benchmarks/performance/m35pka2x",
   "tg_tokens": null,
   "tps": 90.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-m35pka2x-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 146.6 GB at this context; duplicate traces of the same session: 2l0decx5, qhexaxrw, xpaz8msc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1315.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 64k 1,315 78.0 146.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m35pka2x",
   "source_url": "https://omlx.ai/benchmarks/performance/m35pka2x",
   "tg_tokens": null,
   "tps": 78.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-m35pka2x-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 151.2 GB at this context; duplicate traces of the same session: 2l0decx5, qhexaxrw, xpaz8msc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1291.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 128k 1,291 63.5 151.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m35pka2x",
   "source_url": "https://omlx.ai/benchmarks/performance/m35pka2x",
   "tg_tokens": null,
   "tps": 63.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-m35pka2x-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 152.3 GB at this context; duplicate traces of the same session: 2l0decx5, qhexaxrw, xpaz8msc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1276.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 195k 1,276 51.4 152.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m35pka2x",
   "source_url": "https://omlx.ai/benchmarks/performance/m35pka2x",
   "tg_tokens": null,
   "tps": 51.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-fjljh3mx-ctx4096",
   "model": "drowzeys-DeepSeek-V4-Flash-0731-MXFP4-MLX-Abliterated",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 157.7 GB; TTFT 7073 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 579.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "drowzeys-DeepSeek-V4-Flash-0731-MXFP4-MLX-Abliterated on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k tokens 579.1 PP tok/s 35.7 TG tok/s 7073 TTFT (ms) 157.7 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fjljh3mx",
   "source_url": "https://omlx.ai/benchmarks/performance/fjljh3mx",
   "tg_tokens": null,
   "tps": 35.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-hi2nn0q3-ctx4096",
   "model": "drowzeys-DeepSeek-V4-Flash-0731-MXFP4-MLX-Abliterated",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 157.7 GB; TTFT 8646 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 473.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "drowzeys-DeepSeek-V4-Flash-0731-MXFP4-MLX-Abliterated on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k tokens 473.7 PP tok/s 29.1 TG tok/s 8646 TTFT (ms) 157.7 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hi2nn0q3",
   "source_url": "https://omlx.ai/benchmarks/performance/hi2nn0q3",
   "tg_tokens": null,
   "tps": 29.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-x495rsna-ctx65536",
   "model": "gemma-4-E4B-it",
   "notes": "oMLX community benchmark trace; peak memory 9.8 GB at this context; duplicate traces of the same session: u1y9aw5n; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2505.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-E4B-it-8bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 8bit; 64k 2,505 31.5 9.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x495rsna",
   "source_url": "https://omlx.ai/benchmarks/performance/x495rsna",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-x495rsna-ctx131072",
   "model": "gemma-4-E4B-it",
   "notes": "oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: u1y9aw5n; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1862.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-E4B-it-8bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 8bit; 128k 1,862 21.7 11.5 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x495rsna",
   "source_url": "https://omlx.ai/benchmarks/performance/x495rsna",
   "tg_tokens": null,
   "tps": 21.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-aj635g0q-ctx16384",
   "model": "Qwen2.5-7B-Instruct",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 5.5 GB; TTFT 79188 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 206.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen2.5-7B-Instruct-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 unknown; 16k tokens 206.9 PP tok/s 18.7 TG tok/s 79188 TTFT (ms) 5.5 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aj635g0q",
   "source_url": "https://omlx.ai/benchmarks/performance/aj635g0q",
   "tg_tokens": null,
   "tps": 18.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ic2cnw53-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 317.4 GB; TTFT 7116 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 575.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "GLM-5.3-Flash-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k tokens 575.6 PP tok/s 30.3 TG tok/s 7116 TTFT (ms) 317.4 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ic2cnw53",
   "source_url": "https://omlx.ai/benchmarks/performance/ic2cnw53",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xwc5jots-ctx4096",
   "model": "GLM-5.3",
   "notes": "MLX mixed 4/8bit build; oMLX community benchmark trace, single run at 4k context; peak memory 405.6 GB; TTFT 22254 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 184.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-MLX-mixed-4_8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k tokens 184.1 PP tok/s 13.5 TG tok/s 22254 TTFT (ms) 405.6 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xwc5jots",
   "source_url": "https://omlx.ai/benchmarks/performance/xwc5jots",
   "tg_tokens": null,
   "tps": 13.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-p2g4bf10-ctx4096",
   "model": "Kimi-K3",
   "notes": "REAP80 MLX build; oMLX community benchmark trace, single run at 4k context; peak memory 338.7 GB; TTFT 42751 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 95.8,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Kimi-K3-REAP80-MLX-mxfp4-q8 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 mxfp4; 4k tokens 95.8 PP tok/s 6.5 TG tok/s 42751 TTFT (ms) 338.7 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p2g4bf10",
   "source_url": "https://omlx.ai/benchmarks/performance/p2g4bf10",
   "tg_tokens": null,
   "tps": 6.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-b810iglf-ctx4096",
   "model": "MiniMax-M3",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 227.5 GB; TTFT 10875 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 376.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniMax-M3-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k tokens 376.6 PP tok/s 19.7 TG tok/s 10875 TTFT (ms) 227.5 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b810iglf",
   "source_url": "https://omlx.ai/benchmarks/performance/b810iglf",
   "tg_tokens": null,
   "tps": 19.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ghp3dylv-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.8 GB at this context; duplicate traces of the same session: e3hj6jcf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2297.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,297 69.2 66.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ghp3dylv",
   "source_url": "https://omlx.ai/benchmarks/performance/ghp3dylv",
   "tg_tokens": null,
   "tps": 69.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ghp3dylv-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.9 GB at this context; duplicate traces of the same session: e3hj6jcf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2125.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,125 63.2 67.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ghp3dylv",
   "source_url": "https://omlx.ai/benchmarks/performance/ghp3dylv",
   "tg_tokens": null,
   "tps": 63.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-nm1x91h8-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 25.3 GB; TTFT 601069 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 218.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 128k tokens 218.1 PP tok/s 37.3 TG tok/s 601069 TTFT (ms) 25.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nm1x91h8",
   "source_url": "https://omlx.ai/benchmarks/performance/nm1x91h8",
   "tg_tokens": null,
   "tps": 37.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-0ex6ji5v-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: enu89s0z, wvbrv3xs, yolfwb3n; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1008.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 1k 1,008 102.8 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ex6ji5v",
   "source_url": "https://omlx.ai/benchmarks/performance/0ex6ji5v",
   "tg_tokens": null,
   "tps": 102.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-0ex6ji5v-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: enu89s0z, wvbrv3xs, yolfwb3n; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1032.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k 1,032 98.9 21.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ex6ji5v",
   "source_url": "https://omlx.ai/benchmarks/performance/0ex6ji5v",
   "tg_tokens": null,
   "tps": 98.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-0ex6ji5v-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: enu89s0z, wvbrv3xs, yolfwb3n; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 814.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 16k 814.6 86.2 22.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ex6ji5v",
   "source_url": "https://omlx.ai/benchmarks/performance/0ex6ji5v",
   "tg_tokens": null,
   "tps": 86.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-0ex6ji5v-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: enu89s0z, wvbrv3xs, yolfwb3n; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 631.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 32k 631.2 77.2 23.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ex6ji5v",
   "source_url": "https://omlx.ai/benchmarks/performance/0ex6ji5v",
   "tg_tokens": null,
   "tps": 77.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-40ftyyfy-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 22.1 GB; TTFT 20182 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 811.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 16k tokens 811.8 PP tok/s 86.5 TG tok/s 20182 TTFT (ms) 22.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/40ftyyfy",
   "source_url": "https://omlx.ai/benchmarks/performance/40ftyyfy",
   "tg_tokens": null,
   "tps": 86.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-eiws25uy-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 5j43tahg, w8xxcy6b, ww35hzo0, zbwgpkbu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2261.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.6-35B-A3B-5bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 5bit; 1k 2,261 87.3 23.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eiws25uy",
   "source_url": "https://omlx.ai/benchmarks/performance/eiws25uy",
   "tg_tokens": null,
   "tps": 87.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-eiws25uy-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: 5j43tahg, w8xxcy6b, ww35hzo0, zbwgpkbu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2909.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.6-35B-A3B-5bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 5bit; 4k 2,909 85.2 24.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eiws25uy",
   "source_url": "https://omlx.ai/benchmarks/performance/eiws25uy",
   "tg_tokens": null,
   "tps": 85.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-eiws25uy-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: 5j43tahg, w8xxcy6b, ww35hzo0, zbwgpkbu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2880.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.6-35B-A3B-5bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 5bit; 8k 2,880 82.0 24.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eiws25uy",
   "source_url": "https://omlx.ai/benchmarks/performance/eiws25uy",
   "tg_tokens": null,
   "tps": 82.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-eiws25uy-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: 5j43tahg, w8xxcy6b, ww35hzo0, zbwgpkbu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2683.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.6-35B-A3B-5bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 5bit; 16k 2,683 79.8 24.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eiws25uy",
   "source_url": "https://omlx.ai/benchmarks/performance/eiws25uy",
   "tg_tokens": null,
   "tps": 79.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-eiws25uy-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: 5j43tahg, w8xxcy6b, ww35hzo0, zbwgpkbu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2329.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.6-35B-A3B-5bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 5bit; 32k 2,329 71.2 24.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eiws25uy",
   "source_url": "https://omlx.ai/benchmarks/performance/eiws25uy",
   "tg_tokens": null,
   "tps": 71.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-4ihhf8sk-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 6lmyhxw9, jqemmg3u, jy3ty3lq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2247.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.6-35B-A3B-5bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 5bit; 1k 2,247 84.2 23.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ihhf8sk",
   "source_url": "https://omlx.ai/benchmarks/performance/4ihhf8sk",
   "tg_tokens": null,
   "tps": 84.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-4ihhf8sk-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: 6lmyhxw9, jqemmg3u, jy3ty3lq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2912.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.6-35B-A3B-5bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 5bit; 4k 2,912 83.1 24.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ihhf8sk",
   "source_url": "https://omlx.ai/benchmarks/performance/4ihhf8sk",
   "tg_tokens": null,
   "tps": 83.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-4ihhf8sk-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: 6lmyhxw9, jqemmg3u, jy3ty3lq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2877.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.6-35B-A3B-5bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 5bit; 8k 2,877 80.0 24.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ihhf8sk",
   "source_url": "https://omlx.ai/benchmarks/performance/4ihhf8sk",
   "tg_tokens": null,
   "tps": 80.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-4ihhf8sk-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: 6lmyhxw9, jqemmg3u, jy3ty3lq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2684.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.6-35B-A3B-5bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 5bit; 16k 2,684 77.1 24.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ihhf8sk",
   "source_url": "https://omlx.ai/benchmarks/performance/4ihhf8sk",
   "tg_tokens": null,
   "tps": 77.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-famwkwyw-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: kkf5559h; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1626.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.6-35B-A3B-5bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 5bit; 1k 1,626 86.6 23.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/famwkwyw",
   "source_url": "https://omlx.ai/benchmarks/performance/famwkwyw",
   "tg_tokens": null,
   "tps": 86.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-famwkwyw-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: kkf5559h; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2697.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.6-35B-A3B-5bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 5bit; 4k 2,697 84.4 24.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/famwkwyw",
   "source_url": "https://omlx.ai/benchmarks/performance/famwkwyw",
   "tg_tokens": null,
   "tps": 84.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-60gamu0d-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: 3wi4bynd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2359.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 2,359 120.5 27.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/60gamu0d",
   "source_url": "https://omlx.ai/benchmarks/performance/60gamu0d",
   "tg_tokens": null,
   "tps": 120.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-60gamu0d-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: 3wi4bynd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4118.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 4,118 118.8 28.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/60gamu0d",
   "source_url": "https://omlx.ai/benchmarks/performance/60gamu0d",
   "tg_tokens": null,
   "tps": 118.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-bvm39x9p-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: kugb1qzq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 251.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 251.7 26.4 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bvm39x9p",
   "source_url": "https://omlx.ai/benchmarks/performance/bvm39x9p",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-bvm39x9p-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: kugb1qzq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 216.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 216.4 23.1 20.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bvm39x9p",
   "source_url": "https://omlx.ai/benchmarks/performance/bvm39x9p",
   "tg_tokens": null,
   "tps": 23.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-bvm39x9p-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: kugb1qzq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 168.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 168.7 15.5 22.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bvm39x9p",
   "source_url": "https://omlx.ai/benchmarks/performance/bvm39x9p",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-fbmzd6a1-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: e0u6aa71; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 372.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 372.3 12.7 16.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fbmzd6a1",
   "source_url": "https://omlx.ai/benchmarks/performance/fbmzd6a1",
   "tg_tokens": null,
   "tps": 12.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-fbmzd6a1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: e0u6aa71; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 376.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 376.9 12.4 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fbmzd6a1",
   "source_url": "https://omlx.ai/benchmarks/performance/fbmzd6a1",
   "tg_tokens": null,
   "tps": 12.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-fbmzd6a1-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: e0u6aa71; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 368.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 368.6 13.1 17.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fbmzd6a1",
   "source_url": "https://omlx.ai/benchmarks/performance/fbmzd6a1",
   "tg_tokens": null,
   "tps": 13.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-fbmzd6a1-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: e0u6aa71; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 331.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 331.0 12.0 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fbmzd6a1",
   "source_url": "https://omlx.ai/benchmarks/performance/fbmzd6a1",
   "tg_tokens": null,
   "tps": 12.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-fbmzd6a1-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: e0u6aa71; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 318.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 318.8 11.2 18.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fbmzd6a1",
   "source_url": "https://omlx.ai/benchmarks/performance/fbmzd6a1",
   "tg_tokens": null,
   "tps": 11.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-fbmzd6a1-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: e0u6aa71; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 303.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 303.8 11.1 20.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fbmzd6a1",
   "source_url": "https://omlx.ai/benchmarks/performance/fbmzd6a1",
   "tg_tokens": null,
   "tps": 11.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-huyect8d-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: dx4lpod0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 859.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 859.9 24.4 22.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/huyect8d",
   "source_url": "https://omlx.ai/benchmarks/performance/huyect8d",
   "tg_tokens": null,
   "tps": 24.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-huyect8d-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: dx4lpod0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 941.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 941.1 24.1 23.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/huyect8d",
   "source_url": "https://omlx.ai/benchmarks/performance/huyect8d",
   "tg_tokens": null,
   "tps": 24.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-m65ycfh8-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 61.1 GB; TTFT 129104 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1015.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.8-27B-bf16 on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 bf16; 128k tokens 1,015 PP tok/s 13.5 TG tok/s 129104 TTFT (ms) 61.1 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m65ycfh8",
   "source_url": "https://omlx.ai/benchmarks/performance/m65ycfh8",
   "tg_tokens": null,
   "tps": 13.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-o9aelipz-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: hkl3zzzy, x11dwn86; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 592.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 592.9 41.9 17.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9aelipz",
   "source_url": "https://omlx.ai/benchmarks/performance/o9aelipz",
   "tg_tokens": null,
   "tps": 41.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-o9aelipz-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: hkl3zzzy, x11dwn86; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 699.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 699.2 38.3 19.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9aelipz",
   "source_url": "https://omlx.ai/benchmarks/performance/o9aelipz",
   "tg_tokens": null,
   "tps": 38.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-o9aelipz-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: hkl3zzzy, x11dwn86; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 717.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 717.6 39.2 19.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9aelipz",
   "source_url": "https://omlx.ai/benchmarks/performance/o9aelipz",
   "tg_tokens": null,
   "tps": 39.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qs532lf1-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 26ku8a8i; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 593.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 593.5 33.4 17.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qs532lf1",
   "source_url": "https://omlx.ai/benchmarks/performance/qs532lf1",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qs532lf1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 26ku8a8i; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 699.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 699.4 35.6 19.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qs532lf1",
   "source_url": "https://omlx.ai/benchmarks/performance/qs532lf1",
   "tg_tokens": null,
   "tps": 35.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-ivst5o1a-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: ebfh5pn8; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 229.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 8bit; 1k 229.1 18.6 28.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ivst5o1a",
   "source_url": "https://omlx.ai/benchmarks/performance/ivst5o1a",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-ivst5o1a-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: ebfh5pn8; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 233.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 8bit; 4k 233.4 18.3 32.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ivst5o1a",
   "source_url": "https://omlx.ai/benchmarks/performance/ivst5o1a",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-118prtra-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MLX build; oMLX community benchmark trace, single run at 4k context; peak memory 226.7 GB; TTFT 3106 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1319.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-Flash-Next-Uncensored-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k tokens 1,319 PP tok/s 43.3 TG tok/s 3106 TTFT (ms) 226.7 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/118prtra",
   "source_url": "https://omlx.ai/benchmarks/performance/118prtra",
   "tg_tokens": null,
   "tps": 43.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-96fdsadx-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 83.5 GB at this context; duplicate traces of the same session: 7dhw4k46, kktbb9mc, osy2ziac, rcftg4f1, xb5g499b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 118.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 118.1 16.3 83.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/96fdsadx",
   "source_url": "https://omlx.ai/benchmarks/performance/96fdsadx",
   "tg_tokens": null,
   "tps": 16.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-96fdsadx-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 84.3 GB at this context; duplicate traces of the same session: 7dhw4k46, kktbb9mc, osy2ziac, rcftg4f1, xb5g499b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 311.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 311.9 12.2 84.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/96fdsadx",
   "source_url": "https://omlx.ai/benchmarks/performance/96fdsadx",
   "tg_tokens": null,
   "tps": 12.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-96fdsadx-ctx8192",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 85.5 GB at this context; duplicate traces of the same session: 7dhw4k46, kktbb9mc, osy2ziac, rcftg4f1, xb5g499b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 520.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 520.4 14.2 85.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/96fdsadx",
   "source_url": "https://omlx.ai/benchmarks/performance/96fdsadx",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-96fdsadx-ctx65536",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 88.1 GB at this context; duplicate traces of the same session: 7dhw4k46, kktbb9mc, osy2ziac, rcftg4f1, xb5g499b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 478.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 478.9 14.0 88.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/96fdsadx",
   "source_url": "https://omlx.ai/benchmarks/performance/96fdsadx",
   "tg_tokens": null,
   "tps": 14.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-96fdsadx-ctx131072",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 91.7 GB at this context; duplicate traces of the same session: 7dhw4k46, kktbb9mc, osy2ziac, rcftg4f1, xb5g499b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 445.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 445.2 17.7 91.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/96fdsadx",
   "source_url": "https://omlx.ai/benchmarks/performance/96fdsadx",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-96fdsadx-ctx199680",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 94.0 GB at this context; duplicate traces of the same session: 7dhw4k46, kktbb9mc, osy2ziac, rcftg4f1, xb5g499b; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 370.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 195k 370.1 14.2 94.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/96fdsadx",
   "source_url": "https://omlx.ai/benchmarks/performance/96fdsadx",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-zl75ztet-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.5 GB at this context; duplicate traces of the same session: h9bbwg0q; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1026.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 1,026 97.9 142.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zl75ztet",
   "source_url": "https://omlx.ai/benchmarks/performance/zl75ztet",
   "tg_tokens": null,
   "tps": 97.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-zl75ztet-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: h9bbwg0q; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1353.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 1,353 89.5 143.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zl75ztet",
   "source_url": "https://omlx.ai/benchmarks/performance/zl75ztet",
   "tg_tokens": null,
   "tps": 89.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-zl75ztet-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.4 GB at this context; duplicate traces of the same session: h9bbwg0q; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1353.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 8k 1,353 90.4 143.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zl75ztet",
   "source_url": "https://omlx.ai/benchmarks/performance/zl75ztet",
   "tg_tokens": null,
   "tps": 90.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-zl75ztet-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 146.6 GB at this context; duplicate traces of the same session: h9bbwg0q; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1315.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 64k 1,315 78.0 146.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zl75ztet",
   "source_url": "https://omlx.ai/benchmarks/performance/zl75ztet",
   "tg_tokens": null,
   "tps": 78.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-zl75ztet-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 151.2 GB at this context; duplicate traces of the same session: h9bbwg0q; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1291.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 128k 1,291 63.5 151.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zl75ztet",
   "source_url": "https://omlx.ai/benchmarks/performance/zl75ztet",
   "tg_tokens": null,
   "tps": 63.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-zl75ztet-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 152.3 GB at this context; duplicate traces of the same session: h9bbwg0q; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1276.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 195k 1,276 51.4 152.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zl75ztet",
   "source_url": "https://omlx.ai/benchmarks/performance/zl75ztet",
   "tg_tokens": null,
   "tps": 51.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-lezdrptr-ctx4096",
   "model": "Step-3.7-Flash",
   "notes": "MLX build; oMLX community benchmark trace, single run at 4k context; peak memory 196.3 GB; TTFT 6689 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 612.4,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Step-3.7-Flash-oQ8-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k tokens 612.4 PP tok/s 36.8 TG tok/s 6689 TTFT (ms) 196.3 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lezdrptr",
   "source_url": "https://omlx.ai/benchmarks/performance/lezdrptr",
   "tg_tokens": null,
   "tps": 36.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-6hz2oror-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 35.1 GB at this context; duplicate traces of the same session: 0apeboj9, q75h4kyc; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1367.0,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 4k 1,367 82.1 35.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6hz2oror",
   "source_url": "https://omlx.ai/benchmarks/performance/6hz2oror",
   "tg_tokens": null,
   "tps": 82.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-6hz2oror-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 35.7 GB at this context; duplicate traces of the same session: 0apeboj9, q75h4kyc; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1346.0,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 16k 1,346 75.1 35.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6hz2oror",
   "source_url": "https://omlx.ai/benchmarks/performance/6hz2oror",
   "tg_tokens": null,
   "tps": 75.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-6hz2oror-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: 0apeboj9, q75h4kyc; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1263.0,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 32k 1,263 68.4 36.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6hz2oror",
   "source_url": "https://omlx.ai/benchmarks/performance/6hz2oror",
   "tg_tokens": null,
   "tps": 68.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-c49eloob-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: 28gb4ymc, rra9jfio; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1131.0,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 1k 1,131 49.8 33.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c49eloob",
   "source_url": "https://omlx.ai/benchmarks/performance/c49eloob",
   "tg_tokens": null,
   "tps": 49.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-c49eloob-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: 28gb4ymc, rra9jfio; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1358.0,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 4k 1,358 75.0 34.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c49eloob",
   "source_url": "https://omlx.ai/benchmarks/performance/c49eloob",
   "tg_tokens": null,
   "tps": 75.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-c49eloob-ctx131072",
   "model": "Swift-1.5-Qwen3.8-27b",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 43.6 GB at this context; duplicate traces of the same session: 28gb4ymc, rra9jfio; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 881.0,
   "provenance": "community",
   "quant": "oQ8e-bf16",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ8e-bf16-mtp-MLX on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 128k 881.0 70.6 43.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c49eloob",
   "source_url": "https://omlx.ai/benchmarks/performance/c49eloob",
   "tg_tokens": null,
   "tps": 70.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-3db5jpfi-ctx4096",
   "model": "gemma-3-4b-pt",
   "notes": "oMLX community benchmark trace; peak memory 3.5 GB at this context; duplicate traces of the same session: hm3bwko7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1675.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-3-4b-pt-4bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 1,675 144.0 3.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3db5jpfi",
   "source_url": "https://omlx.ai/benchmarks/performance/3db5jpfi",
   "tg_tokens": null,
   "tps": 144.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-3db5jpfi-ctx16384",
   "model": "gemma-3-4b-pt",
   "notes": "oMLX community benchmark trace; peak memory 3.7 GB at this context; duplicate traces of the same session: hm3bwko7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1687.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-3-4b-pt-4bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 1,687 134.2 3.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3db5jpfi",
   "source_url": "https://omlx.ai/benchmarks/performance/3db5jpfi",
   "tg_tokens": null,
   "tps": 134.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-oc5c1azk-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace, single run at 4k context; peak memory 289.6 GB; TTFT 12196 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 335.8,
   "provenance": "community",
   "quant": "oQ3e-2b27-CED",
   "quote": "keys-Mac-oMLX-0.7.0.dev2-DeepSeek-V4.1-Flash-oQ3e-2b27-CED-MTP on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k tokens 335.8 PP tok/s 30.1 TG tok/s 12196 TTFT (ms) 289.6 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oc5c1azk",
   "source_url": "https://omlx.ai/benchmarks/performance/oc5c1azk",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-kp8vc0xr-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "MLX; mlx-community; oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: kxx0nftt; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 240.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 240.1 20.4 6.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kp8vc0xr",
   "source_url": "https://omlx.ai/benchmarks/performance/kp8vc0xr",
   "tg_tokens": null,
   "tps": 20.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-kp8vc0xr-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "MLX; mlx-community; oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: kxx0nftt; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 230.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 16k 230.1 19.2 6.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kp8vc0xr",
   "source_url": "https://omlx.ai/benchmarks/performance/kp8vc0xr",
   "tg_tokens": null,
   "tps": 19.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-n9ls6c9b-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 195k context; peak memory 181.4 GB; TTFT 110796 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1805.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k tokens 1,805 PP tok/s 76.7 TG tok/s 110796 TTFT (ms) 181.4 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n9ls6c9b",
   "source_url": "https://omlx.ai/benchmarks/performance/n9ls6c9b",
   "tg_tokens": null,
   "tps": 76.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-dewuqru6-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.5 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1844.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 1,844 133.3 180.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dewuqru6",
   "source_url": "https://omlx.ai/benchmarks/performance/dewuqru6",
   "tg_tokens": null,
   "tps": 133.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-dewuqru6-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.5 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1799.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 1,799 70.6 181.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dewuqru6",
   "source_url": "https://omlx.ai/benchmarks/performance/dewuqru6",
   "tg_tokens": null,
   "tps": 70.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-4be2zhfq-ctx1024",
   "model": "Huihui-Qwen3.8-27B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: hlda0kzt; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 126.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.8-27B-abliterated-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 126.6 24.4 18.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4be2zhfq",
   "source_url": "https://omlx.ai/benchmarks/performance/4be2zhfq",
   "tg_tokens": null,
   "tps": 24.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-4be2zhfq-ctx4096",
   "model": "Huihui-Qwen3.8-27B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: hlda0kzt; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.8-27B-abliterated-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 131.8 27.3 19.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4be2zhfq",
   "source_url": "https://omlx.ai/benchmarks/performance/4be2zhfq",
   "tg_tokens": null,
   "tps": 27.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-yno5imk4-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "DWQ MLX build v2; oMLX community benchmark trace; peak memory 10.7 GB at this context; duplicate traces of the same session: 7hhnl5ne, qsq1b2dd; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1858.0,
   "provenance": "community",
   "quant": "3bit",
   "quote": "LFM2.5-24B-A2B-MLX-DWQ-3bit-v2 on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 3bit; 4k 1,858 89.1 10.7 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yno5imk4",
   "source_url": "https://omlx.ai/benchmarks/performance/yno5imk4",
   "tg_tokens": null,
   "tps": 89.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-yno5imk4-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "DWQ MLX build v2; oMLX community benchmark trace; peak memory 11.0 GB at this context; duplicate traces of the same session: 7hhnl5ne, qsq1b2dd; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2423.0,
   "provenance": "community",
   "quant": "3bit",
   "quote": "LFM2.5-24B-A2B-MLX-DWQ-3bit-v2 on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 3bit; 16k 2,423 70.0 11.0 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yno5imk4",
   "source_url": "https://omlx.ai/benchmarks/performance/yno5imk4",
   "tg_tokens": null,
   "tps": 70.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-yno5imk4-ctx65536",
   "model": "LFM2.5-24B-A2B",
   "notes": "DWQ MLX build v2; oMLX community benchmark trace; peak memory 12.2 GB at this context; duplicate traces of the same session: 7hhnl5ne, qsq1b2dd; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1902.0,
   "provenance": "community",
   "quant": "3bit",
   "quote": "LFM2.5-24B-A2B-MLX-DWQ-3bit-v2 on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 3bit; 64k 1,902 38.4 12.2 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yno5imk4",
   "source_url": "https://omlx.ai/benchmarks/performance/yno5imk4",
   "tg_tokens": null,
   "tps": 38.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-9p4o6fhr-ctx4096",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 7.4 GB at this context; duplicate traces of the same session: 5vj48u6r; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 345.9,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-oQ4 on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 4bit; 4k 345.9 42.8 7.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9p4o6fhr",
   "source_url": "https://omlx.ai/benchmarks/performance/9p4o6fhr",
   "tg_tokens": null,
   "tps": 42.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-9p4o6fhr-ctx16384",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 7.6 GB at this context; duplicate traces of the same session: 5vj48u6r; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 313.4,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-oQ4 on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 4bit; 16k 313.4 37.9 7.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9p4o6fhr",
   "source_url": "https://omlx.ai/benchmarks/performance/9p4o6fhr",
   "tg_tokens": null,
   "tps": 37.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-fqtqjinf-ctx131072",
   "model": "Ornith-1.0-35B",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 23.1 GB; TTFT 33301 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3936.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.0-35B-oQ4e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 128k tokens 3,936 PP tok/s 93.6 TG tok/s 33301 TTFT (ms) 23.1 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fqtqjinf",
   "source_url": "https://omlx.ai/benchmarks/performance/fqtqjinf",
   "tg_tokens": null,
   "tps": 93.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-56ca7mjw-ctx4096",
   "model": "Qwen2.5-VL-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 8.4 GB at this context; duplicate traces of the same session: 4m4ie388; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 952.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen2.5-VL-7B-Instruct-8bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 8bit; 4k 952.4 57.2 8.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/56ca7mjw",
   "source_url": "https://omlx.ai/benchmarks/performance/56ca7mjw",
   "tg_tokens": null,
   "tps": 57.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-56ca7mjw-ctx16384",
   "model": "Qwen2.5-VL-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 9.0 GB at this context; duplicate traces of the same session: 4m4ie388; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 753.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen2.5-VL-7B-Instruct-8bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 8bit; 16k 753.0 51.8 9.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/56ca7mjw",
   "source_url": "https://omlx.ai/benchmarks/performance/56ca7mjw",
   "tg_tokens": null,
   "tps": 51.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-gx09ehs0-ctx4096",
   "model": "Qwen2.5-VL-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 8.4 GB at this context; duplicate traces of the same session: ue3qj6oe; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 939.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen2.5-VL-7B-Instruct-8bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 8bit; 4k 939.1 57.4 8.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gx09ehs0",
   "source_url": "https://omlx.ai/benchmarks/performance/gx09ehs0",
   "tg_tokens": null,
   "tps": 57.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-gx09ehs0-ctx16384",
   "model": "Qwen2.5-VL-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 9.0 GB at this context; duplicate traces of the same session: ue3qj6oe; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 756.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen2.5-VL-7B-Instruct-8bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 8bit; 16k 756.7 51.7 9.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gx09ehs0",
   "source_url": "https://omlx.ai/benchmarks/performance/gx09ehs0",
   "tg_tokens": null,
   "tps": 51.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-qocph9qx-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.1 GB at this context; duplicate traces of the same session: kge04did; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2286.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 2,286 90.9 43.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qocph9qx",
   "source_url": "https://omlx.ai/benchmarks/performance/qocph9qx",
   "tg_tokens": null,
   "tps": 90.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qocph9qx-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 44.0 GB at this context; duplicate traces of the same session: kge04did; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2919.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,919 87.7 44.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qocph9qx",
   "source_url": "https://omlx.ai/benchmarks/performance/qocph9qx",
   "tg_tokens": null,
   "tps": 87.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-q0lf04ak-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: t6zsjq1p; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1485.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 1,485 55.9 6.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q0lf04ak",
   "source_url": "https://omlx.ai/benchmarks/performance/q0lf04ak",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-q0lf04ak-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: t6zsjq1p; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1475.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 1,475 54.3 6.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q0lf04ak",
   "source_url": "https://omlx.ai/benchmarks/performance/q0lf04ak",
   "tg_tokens": null,
   "tps": 54.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-biy8nzkn-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.7 GB at this context; duplicate traces of the same session: yel92vwj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 702.8,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.5-9B-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 702.8 30.1 7.7 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/biy8nzkn",
   "source_url": "https://omlx.ai/benchmarks/performance/biy8nzkn",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-biy8nzkn-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.2 GB at this context; duplicate traces of the same session: yel92vwj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 832.5,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.5-9B-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 832.5 29.5 8.2 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/biy8nzkn",
   "source_url": "https://omlx.ai/benchmarks/performance/biy8nzkn",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3cntr60z-ctx8192",
   "model": "Qwen3.6-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 22.9 GB; TTFT 6201 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1321.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 8k tokens 1,321 PP tok/s 25.9 TG tok/s 6201 TTFT (ms) 22.9 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3cntr60z",
   "source_url": "https://omlx.ai/benchmarks/performance/3cntr60z",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-88bq64ro-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 24.1 GB; TTFT 187975 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 348.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 64k tokens 348.6 PP tok/s 52.4 TG tok/s 187975 TTFT (ms) 24.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/88bq64ro",
   "source_url": "https://omlx.ai/benchmarks/performance/88bq64ro",
   "tg_tokens": null,
   "tps": 52.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-4n7tufta-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 51gmzm36, gti2ixmi, kh9w406u; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 999.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 1k 999.7 103.8 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4n7tufta",
   "source_url": "https://omlx.ai/benchmarks/performance/4n7tufta",
   "tg_tokens": null,
   "tps": 103.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-4n7tufta-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 51gmzm36, gti2ixmi, kh9w406u; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1103.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k 1,103 99.7 21.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4n7tufta",
   "source_url": "https://omlx.ai/benchmarks/performance/4n7tufta",
   "tg_tokens": null,
   "tps": 99.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-4n7tufta-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 51gmzm36, gti2ixmi, kh9w406u; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 962.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 16k 962.6 87.9 21.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4n7tufta",
   "source_url": "https://omlx.ai/benchmarks/performance/4n7tufta",
   "tg_tokens": null,
   "tps": 87.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-4n7tufta-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: 51gmzm36, gti2ixmi, kh9w406u; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 814.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 32k 814.2 77.0 21.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4n7tufta",
   "source_url": "https://omlx.ai/benchmarks/performance/4n7tufta",
   "tg_tokens": null,
   "tps": 77.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-6t7yvy1e-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: h3mkk026, o16haod5, vf7naev4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2080.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 2,080 90.7 35.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6t7yvy1e",
   "source_url": "https://omlx.ai/benchmarks/performance/6t7yvy1e",
   "tg_tokens": null,
   "tps": 90.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-6t7yvy1e-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: h3mkk026, o16haod5, vf7naev4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2664.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 2,664 89.1 37.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6t7yvy1e",
   "source_url": "https://omlx.ai/benchmarks/performance/6t7yvy1e",
   "tg_tokens": null,
   "tps": 89.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-6t7yvy1e-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.5 GB at this context; duplicate traces of the same session: h3mkk026, o16haod5, vf7naev4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2584.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 8k 2,584 88.6 37.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6t7yvy1e",
   "source_url": "https://omlx.ai/benchmarks/performance/6t7yvy1e",
   "tg_tokens": null,
   "tps": 88.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-6t7yvy1e-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 37.9 GB at this context; duplicate traces of the same session: h3mkk026, o16haod5, vf7naev4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1934.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-MLX-8bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 16k 1,934 81.1 37.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6t7yvy1e",
   "source_url": "https://omlx.ai/benchmarks/performance/6t7yvy1e",
   "tg_tokens": null,
   "tps": 81.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-qcqy1wug-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: r87v7dt1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1011.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,011 80.2 23.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qcqy1wug",
   "source_url": "https://omlx.ai/benchmarks/performance/qcqy1wug",
   "tg_tokens": null,
   "tps": 80.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-qcqy1wug-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: r87v7dt1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 855.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 855.4 94.4 23.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qcqy1wug",
   "source_url": "https://omlx.ai/benchmarks/performance/qcqy1wug",
   "tg_tokens": null,
   "tps": 94.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-leqomy42-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: jn6l5g91, tlc8rxns, u1se9czg; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 144.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 1k 144.2 21.3 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/leqomy42",
   "source_url": "https://omlx.ai/benchmarks/performance/leqomy42",
   "tg_tokens": null,
   "tps": 21.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-leqomy42-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: jn6l5g91, tlc8rxns, u1se9czg; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 132.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k 132.4 20.7 19.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/leqomy42",
   "source_url": "https://omlx.ai/benchmarks/performance/leqomy42",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-leqomy42-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: jn6l5g91, tlc8rxns, u1se9czg; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 121.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 16k 121.4 19.1 21.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/leqomy42",
   "source_url": "https://omlx.ai/benchmarks/performance/leqomy42",
   "tg_tokens": null,
   "tps": 19.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-leqomy42-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: jn6l5g91, tlc8rxns, u1se9czg; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 111.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 32k 111.4 17.5 22.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/leqomy42",
   "source_url": "https://omlx.ai/benchmarks/performance/leqomy42",
   "tg_tokens": null,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-e9r5vkuv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 251.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 251.7 26.4 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e9r5vkuv",
   "source_url": "https://omlx.ai/benchmarks/performance/e9r5vkuv",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-e9r5vkuv-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 216.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 216.4 23.1 20.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e9r5vkuv",
   "source_url": "https://omlx.ai/benchmarks/performance/e9r5vkuv",
   "tg_tokens": null,
   "tps": 23.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-e9r5vkuv-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 168.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 168.7 15.5 22.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e9r5vkuv",
   "source_url": "https://omlx.ai/benchmarks/performance/e9r5vkuv",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-yepc7rlj-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.5 GB at this context; duplicate traces of the same session: fblpywpe; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 704.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 704.3 69.8 17.5 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yepc7rlj",
   "source_url": "https://omlx.ai/benchmarks/performance/yepc7rlj",
   "tg_tokens": null,
   "tps": 69.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-yepc7rlj-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: fblpywpe; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 783.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 783.4 61.5 18.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yepc7rlj",
   "source_url": "https://omlx.ai/benchmarks/performance/yepc7rlj",
   "tg_tokens": null,
   "tps": 61.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-63l4g8po-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 2gr2qcsp, avsk8fpl, rpbdb1fs, rq9tdw44, uuokrn6n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1252.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,252 76.2 19.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/63l4g8po",
   "source_url": "https://omlx.ai/benchmarks/performance/63l4g8po",
   "tg_tokens": null,
   "tps": 76.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-63l4g8po-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 2gr2qcsp, avsk8fpl, rpbdb1fs, rq9tdw44, uuokrn6n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1286.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,286 47.8 19.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/63l4g8po",
   "source_url": "https://omlx.ai/benchmarks/performance/63l4g8po",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-63l4g8po-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 2gr2qcsp, avsk8fpl, rpbdb1fs, rq9tdw44, uuokrn6n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1153.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,153 44.8 19.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/63l4g8po",
   "source_url": "https://omlx.ai/benchmarks/performance/63l4g8po",
   "tg_tokens": null,
   "tps": 44.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-63l4g8po-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: 2gr2qcsp, avsk8fpl, rpbdb1fs, rq9tdw44, uuokrn6n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 953.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 953.3 53.3 21.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/63l4g8po",
   "source_url": "https://omlx.ai/benchmarks/performance/63l4g8po",
   "tg_tokens": null,
   "tps": 53.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-63l4g8po-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 2gr2qcsp, avsk8fpl, rpbdb1fs, rq9tdw44, uuokrn6n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 882.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 882.0 60.3 23.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/63l4g8po",
   "source_url": "https://omlx.ai/benchmarks/performance/63l4g8po",
   "tg_tokens": null,
   "tps": 60.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-63l4g8po-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: 2gr2qcsp, avsk8fpl, rpbdb1fs, rq9tdw44, uuokrn6n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 673.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 673.0 47.2 28.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/63l4g8po",
   "source_url": "https://omlx.ai/benchmarks/performance/63l4g8po",
   "tg_tokens": null,
   "tps": 47.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-hpgc6zn2-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 2gdtbs2h; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 377.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 377.6 34.5 17.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hpgc6zn2",
   "source_url": "https://omlx.ai/benchmarks/performance/hpgc6zn2",
   "tg_tokens": null,
   "tps": 34.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-hpgc6zn2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: 2gdtbs2h; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 392.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 392.6 28.2 18.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hpgc6zn2",
   "source_url": "https://omlx.ai/benchmarks/performance/hpgc6zn2",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-nqvs16yq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: iz6chi7w; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 555.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 555.5 32.7 17.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nqvs16yq",
   "source_url": "https://omlx.ai/benchmarks/performance/nqvs16yq",
   "tg_tokens": null,
   "tps": 32.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-nqvs16yq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: iz6chi7w; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 697.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 697.3 33.5 19.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nqvs16yq",
   "source_url": "https://omlx.ai/benchmarks/performance/nqvs16yq",
   "tg_tokens": null,
   "tps": 33.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-c506gfji-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: js1w6bx2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 249.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 249.7 37.3 33.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c506gfji",
   "source_url": "https://omlx.ai/benchmarks/performance/c506gfji",
   "tg_tokens": null,
   "tps": 37.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-c506gfji-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: js1w6bx2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 256.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 256.8 33.8 37.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c506gfji",
   "source_url": "https://omlx.ai/benchmarks/performance/c506gfji",
   "tg_tokens": null,
   "tps": 33.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-8ywjcjew-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: 1cn1grb8, aclct1n3, uqmk5s3v; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1150.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 1k 1,150 82.2 33.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8ywjcjew",
   "source_url": "https://omlx.ai/benchmarks/performance/8ywjcjew",
   "tg_tokens": null,
   "tps": 82.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-8ywjcjew-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: 1cn1grb8, aclct1n3, uqmk5s3v; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1372.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 4k 1,372 83.3 34.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8ywjcjew",
   "source_url": "https://omlx.ai/benchmarks/performance/8ywjcjew",
   "tg_tokens": null,
   "tps": 83.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-8ywjcjew-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.8 GB at this context; duplicate traces of the same session: 1cn1grb8, aclct1n3, uqmk5s3v; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1262.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 32k 1,262 63.4 36.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8ywjcjew",
   "source_url": "https://omlx.ai/benchmarks/performance/8ywjcjew",
   "tg_tokens": null,
   "tps": 63.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-8ywjcjew-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.7 GB at this context; duplicate traces of the same session: 1cn1grb8, aclct1n3, uqmk5s3v; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 883.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 128k 883.6 65.7 43.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8ywjcjew",
   "source_url": "https://omlx.ai/benchmarks/performance/8ywjcjew",
   "tg_tokens": null,
   "tps": 65.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zb0f0yeh-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 39.0 GB at this context; duplicate traces of the same session: n0nhhx3f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 109.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 109.2 15.3 39.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zb0f0yeh",
   "source_url": "https://omlx.ai/benchmarks/performance/zb0f0yeh",
   "tg_tokens": null,
   "tps": 15.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zb0f0yeh-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 39.4 GB at this context; duplicate traces of the same session: n0nhhx3f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 300.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 300.8 16.2 39.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zb0f0yeh",
   "source_url": "https://omlx.ai/benchmarks/performance/zb0f0yeh",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-h9tizoa5-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Youssofal MTPLX Optimized Speed; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 925.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Youssofal--Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 925.6 50.1 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h9tizoa5",
   "source_url": "https://omlx.ai/benchmarks/performance/h9tizoa5",
   "tg_tokens": null,
   "tps": 50.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-h9tizoa5-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "Youssofal MTPLX Optimized Speed; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1013.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Youssofal--Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,013 81.6 24.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h9tizoa5",
   "source_url": "https://omlx.ai/benchmarks/performance/h9tizoa5",
   "tg_tokens": null,
   "tps": 81.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-qryrx5bb-ctx4096",
   "model": "gemma-3-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 8.4 GB at this context; duplicate traces of the same session: co1hbcos; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 543.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--gemma-3-12b-it-4bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 unknown; 4k 543.8 55.3 8.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qryrx5bb",
   "source_url": "https://omlx.ai/benchmarks/performance/qryrx5bb",
   "tg_tokens": null,
   "tps": 55.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-qryrx5bb-ctx16384",
   "model": "gemma-3-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 9.0 GB at this context; duplicate traces of the same session: co1hbcos; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 485.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--gemma-3-12b-it-4bit on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 unknown; 16k 485.4 48.5 9.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qryrx5bb",
   "source_url": "https://omlx.ai/benchmarks/performance/qryrx5bb",
   "tg_tokens": null,
   "tps": 48.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-c0y8r1hw-ctx4096",
   "model": "GLM-5.3-Flash-Abliterated",
   "notes": "MLX; oMLX community benchmark trace, single run at 4k context; peak memory 180.0 GB; TTFT 7038 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 582.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-Abliterated-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k tokens 582.0 PP tok/s 41.3 TG tok/s 7038 TTFT (ms) 180.0 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c0y8r1hw",
   "source_url": "https://omlx.ai/benchmarks/performance/c0y8r1hw",
   "tg_tokens": null,
   "tps": 41.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-sh5qihpw-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 90.8 GB at this context; duplicate traces of the same session: 0rj4k306, if6gfoa0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 94.9,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "GLM-5.3-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 1k 94.9 12.4 90.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sh5qihpw",
   "source_url": "https://omlx.ai/benchmarks/performance/sh5qihpw",
   "tg_tokens": null,
   "tps": 12.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-sh5qihpw-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 91.5 GB at this context; duplicate traces of the same session: 0rj4k306, if6gfoa0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 296.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "GLM-5.3-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 8k 296.0 12.9 91.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sh5qihpw",
   "source_url": "https://omlx.ai/benchmarks/performance/sh5qihpw",
   "tg_tokens": null,
   "tps": 12.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-sh5qihpw-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 91.9 GB at this context; duplicate traces of the same session: 0rj4k306, if6gfoa0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 279.2,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "GLM-5.3-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 16k 279.2 18.7 91.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sh5qihpw",
   "source_url": "https://omlx.ai/benchmarks/performance/sh5qihpw",
   "tg_tokens": null,
   "tps": 18.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-wacn7ltt-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.5 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1844.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 1,844 133.3 180.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wacn7ltt",
   "source_url": "https://omlx.ai/benchmarks/performance/wacn7ltt",
   "tg_tokens": null,
   "tps": 133.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-wacn7ltt-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.5 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1799.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 1,799 70.6 181.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wacn7ltt",
   "source_url": "https://omlx.ai/benchmarks/performance/wacn7ltt",
   "tg_tokens": null,
   "tps": 70.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-r37eonj0-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.5 GB at this context; duplicate traces of the same session: lpu4fyaz; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1845.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 1,845 135.9 180.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r37eonj0",
   "source_url": "https://omlx.ai/benchmarks/performance/r37eonj0",
   "tg_tokens": null,
   "tps": 135.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-r37eonj0-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.5 GB at this context; duplicate traces of the same session: lpu4fyaz; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1803.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 1,803 71.1 181.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r37eonj0",
   "source_url": "https://omlx.ai/benchmarks/performance/r37eonj0",
   "tg_tokens": null,
   "tps": 71.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-u4634sqj-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.5 GB at this context; duplicate traces of the same session: g95lr12v; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1842.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 1,842 134.1 180.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u4634sqj",
   "source_url": "https://omlx.ai/benchmarks/performance/u4634sqj",
   "tg_tokens": null,
   "tps": 134.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-u4634sqj-ctx199680",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.5 GB at this context; duplicate traces of the same session: g95lr12v; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1806.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 1,806 69.9 181.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u4634sqj",
   "source_url": "https://omlx.ai/benchmarks/performance/u4634sqj",
   "tg_tokens": null,
   "tps": 69.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-jihjs9j9-ctx1024",
   "model": "Nemotron-3-Super-120B-A12B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 68.6 GB at this context; duplicate traces of the same session: h7bqdauk, zhiqx0ol; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 318.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3-Super-120B-A12B-MLX-Q4 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 4bit; 1k 318.1 48.8 68.6 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jihjs9j9",
   "source_url": "https://omlx.ai/benchmarks/performance/jihjs9j9",
   "tg_tokens": null,
   "tps": 48.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-jihjs9j9-ctx4096",
   "model": "Nemotron-3-Super-120B-A12B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: h7bqdauk, zhiqx0ol; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 363.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3-Super-120B-A12B-MLX-Q4 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 4bit; 4k 363.1 48.1 71.5 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jihjs9j9",
   "source_url": "https://omlx.ai/benchmarks/performance/jihjs9j9",
   "tg_tokens": null,
   "tps": 48.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-jihjs9j9-ctx8192",
   "model": "Nemotron-3-Super-120B-A12B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: h7bqdauk, zhiqx0ol; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 369.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3-Super-120B-A12B-MLX-Q4 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 4bit; 8k 369.0 47.5 71.5 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jihjs9j9",
   "source_url": "https://omlx.ai/benchmarks/performance/jihjs9j9",
   "tg_tokens": null,
   "tps": 47.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-qjvf99fp-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: geahqk5c; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1011.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 1k 1,011 82.4 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qjvf99fp",
   "source_url": "https://omlx.ai/benchmarks/performance/qjvf99fp",
   "tg_tokens": null,
   "tps": 82.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-qjvf99fp-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: geahqk5c; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2019.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 4k 2,019 73.4 24.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qjvf99fp",
   "source_url": "https://omlx.ai/benchmarks/performance/qjvf99fp",
   "tg_tokens": null,
   "tps": 73.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-alqiykyz-ctx8192",
   "model": "Qwen3-Coder-Next",
   "notes": "MLX build; oMLX community benchmark trace, single run at 8k context; peak memory 81.1 GB; TTFT 7941 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1032.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next-MLX-8bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 8bit; 8k tokens 1,032 PP tok/s 54.8 TG tok/s 7941 TTFT (ms) 81.1 Peak mem",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/alqiykyz",
   "source_url": "https://omlx.ai/benchmarks/performance/alqiykyz",
   "tg_tokens": null,
   "tps": 54.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-78s1ft3w-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 65.8 GB at this context; duplicate traces of the same session: xn6noafv; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 468.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 4bit; 1k 468.2 52.4 65.8 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/78s1ft3w",
   "source_url": "https://omlx.ai/benchmarks/performance/78s1ft3w",
   "tg_tokens": null,
   "tps": 52.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-78s1ft3w-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 69.5 GB at this context; duplicate traces of the same session: xn6noafv; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 493.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 4bit; 4k 493.8 50.7 69.5 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/78s1ft3w",
   "source_url": "https://omlx.ai/benchmarks/performance/78s1ft3w",
   "tg_tokens": null,
   "tps": 50.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-78s1ft3w-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 70.4 GB at this context; duplicate traces of the same session: xn6noafv; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 466.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 4bit; 8k 466.5 48.5 70.4 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/78s1ft3w",
   "source_url": "https://omlx.ai/benchmarks/performance/78s1ft3w",
   "tg_tokens": null,
   "tps": 48.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-78s1ft3w-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: xn6noafv; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 416.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 4bit; 16k 416.5 46.0 70.9 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/78s1ft3w",
   "source_url": "https://omlx.ai/benchmarks/performance/78s1ft3w",
   "tg_tokens": null,
   "tps": 46.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-nr9jtccx-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: nrs5allp; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1046.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 16 GB \u00b7 4bit; 1k 1,046 28.3 5.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nr9jtccx",
   "source_url": "https://omlx.ai/benchmarks/performance/nr9jtccx",
   "tg_tokens": null,
   "tps": 28.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-nr9jtccx-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: nrs5allp; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1101.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 16 GB \u00b7 4bit; 4k 1,101 28.5 6.2 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nr9jtccx",
   "source_url": "https://omlx.ai/benchmarks/performance/nr9jtccx",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-24fdmi1z-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: ny345ubx; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1076.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k 1,076 98.9 21.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/24fdmi1z",
   "source_url": "https://omlx.ai/benchmarks/performance/24fdmi1z",
   "tg_tokens": null,
   "tps": 98.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-24fdmi1z-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: ny345ubx; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 812.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 32k 812.9 76.0 21.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/24fdmi1z",
   "source_url": "https://omlx.ai/benchmarks/performance/24fdmi1z",
   "tg_tokens": null,
   "tps": 76.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-ldnds3qs-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 8k context; peak memory 21.6 GB; TTFT 5755 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1423.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 4bit; 8k tokens 1,423 PP tok/s 89.8 TG tok/s 5755 TTFT (ms) 21.6 Peak mem",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ldnds3qs",
   "source_url": "https://omlx.ai/benchmarks/performance/ldnds3qs",
   "tg_tokens": null,
   "tps": 89.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-33jawdwe-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 4ctmu4r2, fdzv1i96, j0tpkn1l, koyy0z52; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2225.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.6-35B-A3B-5bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 1k 2,225 96.2 23.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/33jawdwe",
   "source_url": "https://omlx.ai/benchmarks/performance/33jawdwe",
   "tg_tokens": null,
   "tps": 96.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-33jawdwe-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: 4ctmu4r2, fdzv1i96, j0tpkn1l, koyy0z52; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2690.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.6-35B-A3B-5bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 4k 2,690 97.3 25.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/33jawdwe",
   "source_url": "https://omlx.ai/benchmarks/performance/33jawdwe",
   "tg_tokens": null,
   "tps": 97.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-33jawdwe-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: 4ctmu4r2, fdzv1i96, j0tpkn1l, koyy0z52; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2540.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.6-35B-A3B-5bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 8k 2,540 87.4 25.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/33jawdwe",
   "source_url": "https://omlx.ai/benchmarks/performance/33jawdwe",
   "tg_tokens": null,
   "tps": 87.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-33jawdwe-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: 4ctmu4r2, fdzv1i96, j0tpkn1l, koyy0z52; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2270.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.6-35B-A3B-5bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 16k 2,270 84.4 25.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/33jawdwe",
   "source_url": "https://omlx.ai/benchmarks/performance/33jawdwe",
   "tg_tokens": null,
   "tps": 84.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-33jawdwe-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: 4ctmu4r2, fdzv1i96, j0tpkn1l, koyy0z52; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1795.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.6-35B-A3B-5bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 32k 1,795 82.2 26.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/33jawdwe",
   "source_url": "https://omlx.ai/benchmarks/performance/33jawdwe",
   "tg_tokens": null,
   "tps": 82.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-o9tt3ywp-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: 0eihy5fg, 11bfu6yr, cb2nv7o2, f7rslhhm; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2056.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.6-35B-A3B-5bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 1k 2,056 98.3 23.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9tt3ywp",
   "source_url": "https://omlx.ai/benchmarks/performance/o9tt3ywp",
   "tg_tokens": null,
   "tps": 98.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-o9tt3ywp-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: 0eihy5fg, 11bfu6yr, cb2nv7o2, f7rslhhm; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2672.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.6-35B-A3B-5bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 4k 2,672 96.1 25.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9tt3ywp",
   "source_url": "https://omlx.ai/benchmarks/performance/o9tt3ywp",
   "tg_tokens": null,
   "tps": 96.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-o9tt3ywp-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: 0eihy5fg, 11bfu6yr, cb2nv7o2, f7rslhhm; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2550.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.6-35B-A3B-5bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 8k 2,550 90.7 25.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9tt3ywp",
   "source_url": "https://omlx.ai/benchmarks/performance/o9tt3ywp",
   "tg_tokens": null,
   "tps": 90.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-o9tt3ywp-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: 0eihy5fg, 11bfu6yr, cb2nv7o2, f7rslhhm; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2284.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.6-35B-A3B-5bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 16k 2,284 89.3 25.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9tt3ywp",
   "source_url": "https://omlx.ai/benchmarks/performance/o9tt3ywp",
   "tg_tokens": null,
   "tps": 89.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-o9tt3ywp-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: 0eihy5fg, 11bfu6yr, cb2nv7o2, f7rslhhm; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1885.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.6-35B-A3B-5bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 32k 1,885 80.6 26.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9tt3ywp",
   "source_url": "https://omlx.ai/benchmarks/performance/o9tt3ywp",
   "tg_tokens": null,
   "tps": 80.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-656i6nj1-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 0tq7ra10, 6o2zz3nk, iiuydjlo, nodwdgfr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2219.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.6-35B-A3B-5bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 1k 2,219 99.5 23.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/656i6nj1",
   "source_url": "https://omlx.ai/benchmarks/performance/656i6nj1",
   "tg_tokens": null,
   "tps": 99.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-656i6nj1-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: 0tq7ra10, 6o2zz3nk, iiuydjlo, nodwdgfr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2647.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.6-35B-A3B-5bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 4k 2,647 97.6 25.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/656i6nj1",
   "source_url": "https://omlx.ai/benchmarks/performance/656i6nj1",
   "tg_tokens": null,
   "tps": 97.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-656i6nj1-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: 0tq7ra10, 6o2zz3nk, iiuydjlo, nodwdgfr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2548.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.6-35B-A3B-5bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 8k 2,548 95.8 25.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/656i6nj1",
   "source_url": "https://omlx.ai/benchmarks/performance/656i6nj1",
   "tg_tokens": null,
   "tps": 95.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-656i6nj1-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: 0tq7ra10, 6o2zz3nk, iiuydjlo, nodwdgfr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2277.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.6-35B-A3B-5bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 16k 2,277 89.0 25.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/656i6nj1",
   "source_url": "https://omlx.ai/benchmarks/performance/656i6nj1",
   "tg_tokens": null,
   "tps": 89.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-656i6nj1-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: 0tq7ra10, 6o2zz3nk, iiuydjlo, nodwdgfr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1718.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.6-35B-A3B-5bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 32k 1,718 83.8 26.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/656i6nj1",
   "source_url": "https://omlx.ai/benchmarks/performance/656i6nj1",
   "tg_tokens": null,
   "tps": 83.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-wlsvxajq-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: pldkkdpc; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 319.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 128k 319.0 46.1 26.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wlsvxajq",
   "source_url": "https://omlx.ai/benchmarks/performance/wlsvxajq",
   "tg_tokens": null,
   "tps": 46.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-wlsvxajq-ctx199680",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: pldkkdpc; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 228.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 195k 228.6 27.6 28.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wlsvxajq",
   "source_url": "https://omlx.ai/benchmarks/performance/wlsvxajq",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4v6ui8wo-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace, single run at 1k context; peak memory 33.8 GB; TTFT 1044 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 980.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k tokens 980.7 PP tok/s 45.0 TG tok/s 1044 TTFT (ms) 33.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4v6ui8wo",
   "source_url": "https://omlx.ai/benchmarks/performance/4v6ui8wo",
   "tg_tokens": null,
   "tps": 45.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-vzes0l37-ctx1024",
   "model": "Qwen3.8-27B-MindMeld-AREX",
   "notes": "MLX build; oMLX community benchmark trace, single run at 1k context; peak memory 27.7 GB; TTFT 1263 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 810.6,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Qwen3.8-27B-MindMeld-AREX-mxfp8-mlx on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k tokens 810.6 PP tok/s 19.4 TG tok/s 1263 TTFT (ms) 27.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vzes0l37",
   "source_url": "https://omlx.ai/benchmarks/performance/vzes0l37",
   "tg_tokens": null,
   "tps": 19.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-sb85362t-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 26.0 GB; TTFT 1420 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 721.2,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 721.2 PP tok/s 48.4 TG tok/s 1420 TTFT (ms) 26.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sb85362t",
   "source_url": "https://omlx.ai/benchmarks/performance/sb85362t",
   "tg_tokens": null,
   "tps": 48.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-yucppkm9-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 86.4 GB at this context; duplicate traces of the same session: 5y8qizza; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1061.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 1,061 82.8 86.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yucppkm9",
   "source_url": "https://omlx.ai/benchmarks/performance/yucppkm9",
   "tg_tokens": null,
   "tps": 82.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yucppkm9-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 86.6 GB at this context; duplicate traces of the same session: 5y8qizza; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1663.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 1,663 79.2 86.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yucppkm9",
   "source_url": "https://omlx.ai/benchmarks/performance/yucppkm9",
   "tg_tokens": null,
   "tps": 79.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3pxgi9y4-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "WLSRefit MTP build; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: 7h3wxbcv, 7wqwutb8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1049.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-WLSRefit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,049 55.5 70.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3pxgi9y4",
   "source_url": "https://omlx.ai/benchmarks/performance/3pxgi9y4",
   "tg_tokens": null,
   "tps": 55.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3pxgi9y4-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "WLSRefit MTP build; oMLX community benchmark trace; peak memory 76.8 GB at this context; duplicate traces of the same session: 7h3wxbcv, 7wqwutb8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2291.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-WLSRefit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,291 66.7 76.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3pxgi9y4",
   "source_url": "https://omlx.ai/benchmarks/performance/3pxgi9y4",
   "tg_tokens": null,
   "tps": 66.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3pxgi9y4-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "WLSRefit MTP build; oMLX community benchmark trace; peak memory 79.6 GB at this context; duplicate traces of the same session: 7h3wxbcv, 7wqwutb8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2204.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-WLSRefit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,204 61.8 79.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3pxgi9y4",
   "source_url": "https://omlx.ai/benchmarks/performance/3pxgi9y4",
   "tg_tokens": null,
   "tps": 61.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-4yv2g1la-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.6 GB at this context; duplicate traces of the same session: dzarppim; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1122.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 1k 1,122 82.2 142.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4yv2g1la",
   "source_url": "https://omlx.ai/benchmarks/performance/4yv2g1la",
   "tg_tokens": null,
   "tps": 82.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-4yv2g1la-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: dzarppim; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1339.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 4k 1,339 84.6 143.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4yv2g1la",
   "source_url": "https://omlx.ai/benchmarks/performance/4yv2g1la",
   "tg_tokens": null,
   "tps": 84.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-iyddbx68-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "Youssofal MTPLX Optimized Speed; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 925.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Youssofal--Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 925.6 50.1 23.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iyddbx68",
   "source_url": "https://omlx.ai/benchmarks/performance/iyddbx68",
   "tg_tokens": null,
   "tps": 50.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-iyddbx68-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "Youssofal MTPLX Optimized Speed; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1013.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Youssofal--Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,013 81.6 24.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iyddbx68",
   "source_url": "https://omlx.ai/benchmarks/performance/iyddbx68",
   "tg_tokens": null,
   "tps": 81.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-kyx9gpfd-ctx1024",
   "model": "dots.mocr",
   "notes": "oMLX community benchmark trace; peak memory 4.8 GB at this context; duplicate traces of the same session: 0r20h6g7; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4756.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "dots.mocr-8bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 8bit; 1k 4,756 101.8 4.8 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kyx9gpfd",
   "source_url": "https://omlx.ai/benchmarks/performance/kyx9gpfd",
   "tg_tokens": null,
   "tps": 101.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-kyx9gpfd-ctx4096",
   "model": "dots.mocr",
   "notes": "oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: 0r20h6g7; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3456.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "dots.mocr-8bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 8bit; 4k 3,456 99.5 4.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kyx9gpfd",
   "source_url": "https://omlx.ai/benchmarks/performance/kyx9gpfd",
   "tg_tokens": null,
   "tps": 99.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-sh0v58ca-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "leonsarmiento pi MTP; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: 1igghmm3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 208.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "leonsarmiento--Qwen3.8-27B-pi-4bit-mtp-mlx on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 1k 208.0 37.3 18.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sh0v58ca",
   "source_url": "https://omlx.ai/benchmarks/performance/sh0v58ca",
   "tg_tokens": null,
   "tps": 37.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-sh0v58ca-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "leonsarmiento pi MTP; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: 1igghmm3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 197.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "leonsarmiento--Qwen3.8-27B-pi-4bit-mtp-mlx on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 4k 197.0 41.7 22.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sh0v58ca",
   "source_url": "https://omlx.ai/benchmarks/performance/sh0v58ca",
   "tg_tokens": null,
   "tps": 41.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-l8yhgdi4-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: yul61u2p; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 213.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 1k 213.2 26.7 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l8yhgdi4",
   "source_url": "https://omlx.ai/benchmarks/performance/l8yhgdi4",
   "tg_tokens": null,
   "tps": 26.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-l8yhgdi4-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "mlx-community quantized weights; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: yul61u2p; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen3.8-27B-4bit on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 4k 204.2 26.2 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l8yhgdi4",
   "source_url": "https://omlx.ai/benchmarks/performance/l8yhgdi4",
   "tg_tokens": null,
   "tps": 26.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-1n6s5cu7-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai build; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: wi60wk7o; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1120.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-4bit on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 1k 1,120 48.0 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1n6s5cu7",
   "source_url": "https://omlx.ai/benchmarks/performance/1n6s5cu7",
   "tg_tokens": null,
   "tps": 48.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-1n6s5cu7-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai build; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: wi60wk7o; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1398.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-4bit on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 4k 1,398 53.2 21.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1n6s5cu7",
   "source_url": "https://omlx.ai/benchmarks/performance/1n6s5cu7",
   "tg_tokens": null,
   "tps": 53.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-z695iawm-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai build; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: t0e511l8; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1058.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-4bit on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 1k 1,058 57.4 20.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z695iawm",
   "source_url": "https://omlx.ai/benchmarks/performance/z695iawm",
   "tg_tokens": null,
   "tps": 57.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-z695iawm-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "ornith-ai build; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: t0e511l8; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1398.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "ornith-ai--Ornith-1.5-35B-A3B-MLX-4bit on M1 Ultra (64c); M1 Ultra (64c) \u00b7 64 GB \u00b7 4bit; 4k 1,398 73.4 22.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z695iawm",
   "source_url": "https://omlx.ai/benchmarks/performance/z695iawm",
   "tg_tokens": null,
   "tps": 73.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2k8wnff2-ctx1024",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 155.8 GB at this context; duplicate traces of the same session: 5o37l8m8, ez1k6cdc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 422.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 422.5 42.9 155.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2k8wnff2",
   "source_url": "https://omlx.ai/benchmarks/performance/2k8wnff2",
   "tg_tokens": null,
   "tps": 42.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2k8wnff2-ctx4096",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.3 GB at this context; duplicate traces of the same session: 5o37l8m8, ez1k6cdc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 467.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 467.6 39.5 156.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2k8wnff2",
   "source_url": "https://omlx.ai/benchmarks/performance/2k8wnff2",
   "tg_tokens": null,
   "tps": 39.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2k8wnff2-ctx8192",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 156.3 GB at this context; duplicate traces of the same session: 5o37l8m8, ez1k6cdc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 459.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4-Flash-0731-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 459.8 49.3 156.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2k8wnff2",
   "source_url": "https://omlx.ai/benchmarks/performance/2k8wnff2",
   "tg_tokens": null,
   "tps": 49.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-c7z26cji-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 89.7 GB at this context; duplicate traces of the same session: 1qbe2y55, c0hlztu1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 94.7,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "GLM-5.3-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 1k 94.7 13.0 89.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c7z26cji",
   "source_url": "https://omlx.ai/benchmarks/performance/c7z26cji",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-c7z26cji-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 90.3 GB at this context; duplicate traces of the same session: 1qbe2y55, c0hlztu1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 285.2,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "GLM-5.3-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 8k 285.2 12.6 90.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c7z26cji",
   "source_url": "https://omlx.ai/benchmarks/performance/c7z26cji",
   "tg_tokens": null,
   "tps": 12.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-c7z26cji-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 90.9 GB at this context; duplicate traces of the same session: 1qbe2y55, c0hlztu1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 267.8,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "GLM-5.3-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 16k 267.8 18.5 90.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c7z26cji",
   "source_url": "https://omlx.ai/benchmarks/performance/c7z26cji",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-t1ehjkl7-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: none; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 395.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 4bit; 1k 395.2 57.6 13.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t1ehjkl7",
   "source_url": "https://omlx.ai/benchmarks/performance/t1ehjkl7",
   "tg_tokens": null,
   "tps": 57.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-t1ehjkl7-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: none; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 427.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 4bit; 4k 427.3 54.6 13.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t1ehjkl7",
   "source_url": "https://omlx.ai/benchmarks/performance/t1ehjkl7",
   "tg_tokens": null,
   "tps": 54.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-t1ehjkl7-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.8 GB at this context; duplicate traces of the same session: none; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 432.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 4bit; 8k 432.6 51.4 13.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t1ehjkl7",
   "source_url": "https://omlx.ai/benchmarks/performance/t1ehjkl7",
   "tg_tokens": null,
   "tps": 51.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-t1ehjkl7-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.9 GB at this context; duplicate traces of the same session: none; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 413.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 4bit; 16k 413.2 46.3 13.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t1ehjkl7",
   "source_url": "https://omlx.ai/benchmarks/performance/t1ehjkl7",
   "tg_tokens": null,
   "tps": 46.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-l8nlhone-ctx4096",
   "model": "Ministral-3-3B-Reasoning-2512",
   "notes": "oMLX community benchmark trace; peak memory 2.9 GB at this context; duplicate traces of the same session: tqm0eamm; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 305.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ministral-3-3B-Reasoning-2512-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 4bit; 4k 305.8 41.4 2.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l8nlhone",
   "source_url": "https://omlx.ai/benchmarks/performance/l8nlhone",
   "tg_tokens": null,
   "tps": 41.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-l8nlhone-ctx16384",
   "model": "Ministral-3-3B-Reasoning-2512",
   "notes": "oMLX community benchmark trace; peak memory 4.0 GB at this context; duplicate traces of the same session: tqm0eamm; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 231.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ministral-3-3B-Reasoning-2512-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 4bit; 16k 231.4 25.2 4.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l8nlhone",
   "source_url": "https://omlx.ai/benchmarks/performance/l8nlhone",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-lko9yh39-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 38.3 GB; TTFT 12598 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5202.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 64k tokens 5,202 PP tok/s 183.4 TG tok/s 12598 TTFT (ms) 38.3 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lko9yh39",
   "source_url": "https://omlx.ai/benchmarks/performance/lko9yh39",
   "tg_tokens": null,
   "tps": 183.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-p8ij4xly-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 38.3 GB; TTFT 12585 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5208.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 64k tokens 5,208 PP tok/s 140.1 TG tok/s 12585 TTFT (ms) 38.3 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p8ij4xly",
   "source_url": "https://omlx.ai/benchmarks/performance/p8ij4xly",
   "tg_tokens": null,
   "tps": 140.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ldz22g1s-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 38.3 GB; TTFT 12511 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5238.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 64k tokens 5,238 PP tok/s 138.1 TG tok/s 12511 TTFT (ms) 38.3 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ldz22g1s",
   "source_url": "https://omlx.ai/benchmarks/performance/ldz22g1s",
   "tg_tokens": null,
   "tps": 138.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-4xtcyo4i-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 38.3 GB; TTFT 12542 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5226.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 64k tokens 5,226 PP tok/s 145.0 TG tok/s 12542 TTFT (ms) 38.3 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4xtcyo4i",
   "source_url": "https://omlx.ai/benchmarks/performance/4xtcyo4i",
   "tg_tokens": null,
   "tps": 145.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-336ujnjy-ctx1024",
   "model": "Qwen2.5-Coder-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 4.5 GB at this context; duplicate traces of the same session: vtvlnfe5; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1087.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-7B-Instruct-4bit on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 1k 1,087 34.7 4.5 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/336ujnjy",
   "source_url": "https://omlx.ai/benchmarks/performance/336ujnjy",
   "tg_tokens": null,
   "tps": 34.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-336ujnjy-ctx4096",
   "model": "Qwen2.5-Coder-7B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 4.9 GB at this context; duplicate traces of the same session: vtvlnfe5; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1144.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen2.5-Coder-7B-Instruct-4bit on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 4k 1,144 33.7 4.9 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/336ujnjy",
   "source_url": "https://omlx.ai/benchmarks/performance/336ujnjy",
   "tg_tokens": null,
   "tps": 33.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-7b44771y-ctx1024",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: mo4kugsg; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 990.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 1k 990.1 80.9 16.6 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7b44771y",
   "source_url": "https://omlx.ai/benchmarks/performance/7b44771y",
   "tg_tokens": null,
   "tps": 80.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-7b44771y-ctx4096",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: mo4kugsg; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1037.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 4k 1,037 57.5 17.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7b44771y",
   "source_url": "https://omlx.ai/benchmarks/performance/7b44771y",
   "tg_tokens": null,
   "tps": 57.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-s092ufvi-ctx1024",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 65.8 GB at this context; duplicate traces of the same session: 1irtoakk; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 468.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 4bit; 1k 468.2 52.4 65.8 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s092ufvi",
   "source_url": "https://omlx.ai/benchmarks/performance/s092ufvi",
   "tg_tokens": null,
   "tps": 52.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-s092ufvi-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 69.5 GB at this context; duplicate traces of the same session: 1irtoakk; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 493.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 4bit; 4k 493.8 50.7 69.5 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s092ufvi",
   "source_url": "https://omlx.ai/benchmarks/performance/s092ufvi",
   "tg_tokens": null,
   "tps": 50.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-s092ufvi-ctx8192",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 70.4 GB at this context; duplicate traces of the same session: 1irtoakk; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 466.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 4bit; 8k 466.5 48.5 70.4 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s092ufvi",
   "source_url": "https://omlx.ai/benchmarks/performance/s092ufvi",
   "tg_tokens": null,
   "tps": 48.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-s092ufvi-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: 1irtoakk; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 416.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 4bit; 16k 416.5 46.0 70.9 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s092ufvi",
   "source_url": "https://omlx.ai/benchmarks/performance/s092ufvi",
   "tg_tokens": null,
   "tps": 46.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-f2c1bh18-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 69.5 GB; TTFT 8307 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 493.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 4bit; 4k tokens 493.1 PP tok/s 50.5 TG tok/s 8307 TTFT (ms) 69.5 Peak mem",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f2c1bh18",
   "source_url": "https://omlx.ai/benchmarks/performance/f2c1bh18",
   "tg_tokens": null,
   "tps": 50.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-7fuacp8f-ctx1024",
   "model": "Qwen3.6-27B-Architect-Polaris2-Fable-B-F451-Medical",
   "notes": "MLX build; oMLX community benchmark trace, single run at 1k context; peak memory 31.4 GB; TTFT 1570 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 652.2,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Qwen3.6-27B-Architect-Polaris2-Fable-B-F451-Medical-mxfp8-mlx on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k tokens 652.2 PP tok/s 16.3 TG tok/s 1570 TTFT (ms) 31.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7fuacp8f",
   "source_url": "https://omlx.ai/benchmarks/performance/7fuacp8f",
   "tg_tokens": null,
   "tps": 16.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-s53my6rp-ctx1024",
   "model": "Qwen3.6-27B-Architect-Polaris2-Fable-B-F451-Tess-1M",
   "notes": "MLX build; oMLX community benchmark trace, single run at 1k context; peak memory 20.5 GB; TTFT 1208 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 847.7,
   "provenance": "community",
   "quant": "qx64-hi",
   "quote": "Qwen3.6-27B-Architect-Polaris2-Fable-B-F451-Tess-1M-qx64-hi-mlx on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k tokens 847.7 PP tok/s 25.7 TG tok/s 1208 TTFT (ms) 20.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s53my6rp",
   "source_url": "https://omlx.ai/benchmarks/performance/s53my6rp",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-apm2ap6f-ctx1024",
   "model": "Qwen3.6-27B-Architect-Polaris2-Fable-B-F451-Tess-1M",
   "notes": "MLX build; oMLX community benchmark trace, single run at 1k context; peak memory 20.5 GB; TTFT 1327 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 771.9,
   "provenance": "community",
   "quant": "qx64-hi",
   "quote": "Qwen3.6-27B-Architect-Polaris2-Fable-B-F451-Tess-1M-qx64-hi-mlx on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k tokens 771.9 PP tok/s 25.8 TG tok/s 1327 TTFT (ms) 20.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/apm2ap6f",
   "source_url": "https://omlx.ai/benchmarks/performance/apm2ap6f",
   "tg_tokens": null,
   "tps": 25.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wln0u5o6-ctx1024",
   "model": "Qwen3.6-27B-Architect-Polaris2-Fable-B-F451",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 17.6 GB; TTFT 1151 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 890.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-27B-Architect-Polaris2-Fable-B-F451-oQ4-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 890.0 PP tok/s 31.5 TG tok/s 1151 TTFT (ms) 17.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wln0u5o6",
   "source_url": "https://omlx.ai/benchmarks/performance/wln0u5o6",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-zdzrloy3-ctx1024",
   "model": "Qwen3.6-27B-Fable-Fusion-711-BigBubba-717",
   "notes": "MLX build; oMLX community benchmark trace, single run at 1k context; peak memory 27.7 GB; TTFT 1271 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 805.9,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Qwen3.6-27B-Fable-Fusion-711-BigBubba-717-mxfp8-mlx on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k tokens 805.9 PP tok/s 19.4 TG tok/s 1271 TTFT (ms) 27.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zdzrloy3",
   "source_url": "https://omlx.ai/benchmarks/performance/zdzrloy3",
   "tg_tokens": null,
   "tps": 19.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-n2ocylog-ctx1024",
   "model": "Qwen3.6-35B-A3B-Brainwaves-Nex-Uncensored",
   "notes": "mlx build; oMLX community benchmark trace, single run at 1k context; peak memory 29.2 GB; TTFT 395 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2595.0,
   "provenance": "community",
   "quant": "qx64-hi",
   "quote": "Qwen3.6-35B-A3B-Brainwaves-Nex-Uncensored-qx64-hi-mlx on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k tokens 2,595 PP tok/s 115.7 TG tok/s 395 TTFT (ms) 29.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n2ocylog",
   "source_url": "https://omlx.ai/benchmarks/performance/n2ocylog",
   "tg_tokens": null,
   "tps": 115.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-xiq9ce9z-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: wnu7dl7z; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 277.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 277.1 52.0 19.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xiq9ce9z",
   "source_url": "https://omlx.ai/benchmarks/performance/xiq9ce9z",
   "tg_tokens": null,
   "tps": 52.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-xiq9ce9z-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: wnu7dl7z; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 310.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 310.5 49.9 20.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xiq9ce9z",
   "source_url": "https://omlx.ai/benchmarks/performance/xiq9ce9z",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-xiq9ce9z-ctx8192",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: wnu7dl7z; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 295.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 8k 295.4 43.9 20.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xiq9ce9z",
   "source_url": "https://omlx.ai/benchmarks/performance/xiq9ce9z",
   "tg_tokens": null,
   "tps": 43.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hnzvg5sh-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 23.1 GB; TTFT 799 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1282.0,
   "provenance": "community",
   "quant": "VQ-5.4bpw",
   "quote": "Qwen3.6-35B-A3B-VQ-5.4bpw on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 2bit; 1k tokens 1,282 PP tok/s 78.3 TG tok/s 799 TTFT (ms) 23.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hnzvg5sh",
   "source_url": "https://omlx.ai/benchmarks/performance/hnzvg5sh",
   "tg_tokens": null,
   "tps": 78.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1ku8x9su-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 36.6 GB; TTFT 36984 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 886.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k tokens 886.0 PP tok/s 35.2 TG tok/s 36984 TTFT (ms) 36.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ku8x9su",
   "source_url": "https://omlx.ai/benchmarks/performance/1ku8x9su",
   "tg_tokens": null,
   "tps": 35.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9eqd01qz-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 8k context; peak memory 35.1 GB; TTFT 8530 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 960.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k tokens 960.4 PP tok/s 37.8 TG tok/s 8530 TTFT (ms) 35.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9eqd01qz",
   "source_url": "https://omlx.ai/benchmarks/performance/9eqd01qz",
   "tg_tokens": null,
   "tps": 37.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4kp534gt-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 36.6 GB; TTFT 37112 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 883.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k tokens 883.0 PP tok/s 34.2 TG tok/s 37112 TTFT (ms) 36.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4kp534gt",
   "source_url": "https://omlx.ai/benchmarks/performance/4kp534gt",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-v6y62ozp-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace, single run at 4k context; peak memory 23.6 GB; TTFT 40173 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k tokens 102.0 PP tok/s 12.4 TG tok/s 40173 TTFT (ms) 23.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v6y62ozp",
   "source_url": "https://omlx.ai/benchmarks/performance/v6y62ozp",
   "tg_tokens": null,
   "tps": 12.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-r4lfbwuf-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.2 GB at this context; duplicate traces of the same session: a9j6t3cd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 913.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 913.7 44.4 71.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r4lfbwuf",
   "source_url": "https://omlx.ai/benchmarks/performance/r4lfbwuf",
   "tg_tokens": null,
   "tps": 44.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-r4lfbwuf-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.9 GB at this context; duplicate traces of the same session: a9j6t3cd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1241.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,241 58.3 72.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r4lfbwuf",
   "source_url": "https://omlx.ai/benchmarks/performance/r4lfbwuf",
   "tg_tokens": null,
   "tps": 58.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-u10nt52m-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 102.5 GB; TTFT 7573 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2163.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k tokens 2,163 PP tok/s 15.3 TG tok/s 7573 TTFT (ms) 102.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u10nt52m",
   "source_url": "https://omlx.ai/benchmarks/performance/u10nt52m",
   "tg_tokens": null,
   "tps": 15.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ipsg9s8k-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 102.8 GB; TTFT 17580 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 932.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k tokens 932.0 PP tok/s 38.3 TG tok/s 17580 TTFT (ms) 102.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ipsg9s8k",
   "source_url": "https://omlx.ai/benchmarks/performance/ipsg9s8k",
   "tg_tokens": null,
   "tps": 38.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-o3w9qiql-ctx1024",
   "model": "Swift 1.5 Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 142.8 GB at this context; duplicate traces of the same session: 2vjee61p, 9onc3pd7, klmcf16q, yesxutio, zf3c3kgd; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2245.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 1k 2,245 118.7 142.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o3w9qiql",
   "source_url": "https://omlx.ai/benchmarks/performance/o3w9qiql",
   "tg_tokens": null,
   "tps": 118.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-o3w9qiql-ctx16384",
   "model": "Swift 1.5 Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 145.4 GB at this context; duplicate traces of the same session: 2vjee61p, 9onc3pd7, klmcf16q, yesxutio, zf3c3kgd; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4226.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 16k 4,226 105.9 145.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o3w9qiql",
   "source_url": "https://omlx.ai/benchmarks/performance/o3w9qiql",
   "tg_tokens": null,
   "tps": 105.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-o3w9qiql-ctx32768",
   "model": "Swift 1.5 Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 146.0 GB at this context; duplicate traces of the same session: 2vjee61p, 9onc3pd7, klmcf16q, yesxutio, zf3c3kgd; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4180.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 32k 4,180 118.5 146.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o3w9qiql",
   "source_url": "https://omlx.ai/benchmarks/performance/o3w9qiql",
   "tg_tokens": null,
   "tps": 118.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-o3w9qiql-ctx65536",
   "model": "Swift 1.5 Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 147.2 GB at this context; duplicate traces of the same session: 2vjee61p, 9onc3pd7, klmcf16q, yesxutio, zf3c3kgd; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4113.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 64k 4,113 97.4 147.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o3w9qiql",
   "source_url": "https://omlx.ai/benchmarks/performance/o3w9qiql",
   "tg_tokens": null,
   "tps": 97.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-o3w9qiql-ctx131072",
   "model": "Swift 1.5 Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 149.5 GB at this context; duplicate traces of the same session: 2vjee61p, 9onc3pd7, klmcf16q, yesxutio, zf3c3kgd; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4022.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 128k 4,022 84.9 149.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o3w9qiql",
   "source_url": "https://omlx.ai/benchmarks/performance/o3w9qiql",
   "tg_tokens": null,
   "tps": 84.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-o3w9qiql-ctx199680",
   "model": "Swift 1.5 Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 152.3 GB at this context; duplicate traces of the same session: 2vjee61p, 9onc3pd7, klmcf16q, yesxutio, zf3c3kgd; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3911.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 195k 3,911 85.8 152.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o3w9qiql",
   "source_url": "https://omlx.ai/benchmarks/performance/o3w9qiql",
   "tg_tokens": null,
   "tps": 85.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-7lfg59wa-ctx1024",
   "model": "Swift 1.5 Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 142.8 GB at this context; duplicate traces of the same session: 0n10hbf6, 92kzpzt6, iyzroszt, tnwbg1i3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2194.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 1k 2,194 132.2 142.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7lfg59wa",
   "source_url": "https://omlx.ai/benchmarks/performance/7lfg59wa",
   "tg_tokens": null,
   "tps": 132.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-7lfg59wa-ctx16384",
   "model": "Swift 1.5 Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 145.4 GB at this context; duplicate traces of the same session: 0n10hbf6, 92kzpzt6, iyzroszt, tnwbg1i3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4215.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 16k 4,215 111.2 145.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7lfg59wa",
   "source_url": "https://omlx.ai/benchmarks/performance/7lfg59wa",
   "tg_tokens": null,
   "tps": 111.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-7lfg59wa-ctx65536",
   "model": "Swift 1.5 Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 147.2 GB at this context; duplicate traces of the same session: 0n10hbf6, 92kzpzt6, iyzroszt, tnwbg1i3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4107.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 64k 4,107 122.4 147.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7lfg59wa",
   "source_url": "https://omlx.ai/benchmarks/performance/7lfg59wa",
   "tg_tokens": null,
   "tps": 122.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-7lfg59wa-ctx131072",
   "model": "Swift 1.5 Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 149.5 GB at this context; duplicate traces of the same session: 0n10hbf6, 92kzpzt6, iyzroszt, tnwbg1i3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4010.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 128k 4,010 107.8 149.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7lfg59wa",
   "source_url": "https://omlx.ai/benchmarks/performance/7lfg59wa",
   "tg_tokens": null,
   "tps": 107.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-7lfg59wa-ctx199680",
   "model": "Swift 1.5 Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 152.3 GB at this context; duplicate traces of the same session: 0n10hbf6, 92kzpzt6, iyzroszt, tnwbg1i3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3895.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 195k 3,895 88.5 152.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7lfg59wa",
   "source_url": "https://omlx.ai/benchmarks/performance/7lfg59wa",
   "tg_tokens": null,
   "tps": 88.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-luqu77c3-ctx1024",
   "model": "Swift 1.5 Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 142.8 GB at this context; duplicate traces of the same session: 78yhnubm, ayx7x2sx, h33wqh8n, hckyz1gf; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 813.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 1k 813.4 32.6 142.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/luqu77c3",
   "source_url": "https://omlx.ai/benchmarks/performance/luqu77c3",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-luqu77c3-ctx16384",
   "model": "Swift 1.5 Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 145.4 GB at this context; duplicate traces of the same session: 78yhnubm, ayx7x2sx, h33wqh8n, hckyz1gf; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2697.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 16k 2,697 23.6 145.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/luqu77c3",
   "source_url": "https://omlx.ai/benchmarks/performance/luqu77c3",
   "tg_tokens": null,
   "tps": 23.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-luqu77c3-ctx65536",
   "model": "Swift 1.5 Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 147.2 GB at this context; duplicate traces of the same session: 78yhnubm, ayx7x2sx, h33wqh8n, hckyz1gf; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4054.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 64k 4,054 69.2 147.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/luqu77c3",
   "source_url": "https://omlx.ai/benchmarks/performance/luqu77c3",
   "tg_tokens": null,
   "tps": 69.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-luqu77c3-ctx131072",
   "model": "Swift 1.5 Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 149.5 GB at this context; duplicate traces of the same session: 78yhnubm, ayx7x2sx, h33wqh8n, hckyz1gf; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4010.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 128k 4,010 101.2 149.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/luqu77c3",
   "source_url": "https://omlx.ai/benchmarks/performance/luqu77c3",
   "tg_tokens": null,
   "tps": 101.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-luqu77c3-ctx199680",
   "model": "Swift 1.5 Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 152.3 GB at this context; duplicate traces of the same session: 78yhnubm, ayx7x2sx, h33wqh8n, hckyz1gf; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3900.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 195k 3,900 90.8 152.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/luqu77c3",
   "source_url": "https://omlx.ai/benchmarks/performance/luqu77c3",
   "tg_tokens": null,
   "tps": 90.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-sqzb02d6-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: 1degvznu, 49art7op, 93a1renc, lzpble1x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1335.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ6e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 1,335 60.0 28.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sqzb02d6",
   "source_url": "https://omlx.ai/benchmarks/performance/sqzb02d6",
   "tg_tokens": null,
   "tps": 60.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-sqzb02d6-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: 1degvznu, 49art7op, 93a1renc, lzpble1x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1385.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ6e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 1,385 90.6 30.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sqzb02d6",
   "source_url": "https://omlx.ai/benchmarks/performance/sqzb02d6",
   "tg_tokens": null,
   "tps": 90.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-sqzb02d6-ctx8192",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 30.4 GB at this context; duplicate traces of the same session: 1degvznu, 49art7op, 93a1renc, lzpble1x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1734.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ6e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 8k 1,734 60.8 30.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sqzb02d6",
   "source_url": "https://omlx.ai/benchmarks/performance/sqzb02d6",
   "tg_tokens": null,
   "tps": 60.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-sqzb02d6-ctx16384",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 30.7 GB at this context; duplicate traces of the same session: 1degvznu, 49art7op, 93a1renc, lzpble1x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1547.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ6e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k 1,547 118.6 30.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sqzb02d6",
   "source_url": "https://omlx.ai/benchmarks/performance/sqzb02d6",
   "tg_tokens": null,
   "tps": 118.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-sqzb02d6-ctx32768",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 31.4 GB at this context; duplicate traces of the same session: 1degvznu, 49art7op, 93a1renc, lzpble1x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1231.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ6e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 1,231 57.3 31.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sqzb02d6",
   "source_url": "https://omlx.ai/benchmarks/performance/sqzb02d6",
   "tg_tokens": null,
   "tps": 57.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-sqzb02d6-ctx65536",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 32.7 GB at this context; duplicate traces of the same session: 1degvznu, 49art7op, 93a1renc, lzpble1x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 880.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ6e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 64k 880.0 50.6 32.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sqzb02d6",
   "source_url": "https://omlx.ai/benchmarks/performance/sqzb02d6",
   "tg_tokens": null,
   "tps": 50.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-sqzb02d6-ctx131072",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: 1degvznu, 49art7op, 93a1renc, lzpble1x; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 488.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ6e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 128k 488.8 55.2 34.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sqzb02d6",
   "source_url": "https://omlx.ai/benchmarks/performance/sqzb02d6",
   "tg_tokens": null,
   "tps": 55.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ic03gm42-ctx1024",
   "model": "gemma-4-12B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 11.3 GB at this context; duplicate traces of the same session: ld84qb45; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 230.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-qat-4bit on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 1k 230.4 25.3 11.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ic03gm42",
   "source_url": "https://omlx.ai/benchmarks/performance/ic03gm42",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ic03gm42-ctx4096",
   "model": "gemma-4-12B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 11.7 GB at this context; duplicate traces of the same session: ld84qb45; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 238.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-qat-4bit on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 4k 238.2 25.0 11.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ic03gm42",
   "source_url": "https://omlx.ai/benchmarks/performance/ic03gm42",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ff7osnes-ctx1024",
   "model": "GLM-5.3-Flash-Abliterated",
   "notes": "MLX; oMLX community benchmark trace; peak memory 177.7 GB at this context; duplicate traces of the same session: sbddptc1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 522.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-Abliterated-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 522.6 49.9 177.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ff7osnes",
   "source_url": "https://omlx.ai/benchmarks/performance/ff7osnes",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ff7osnes-ctx4096",
   "model": "GLM-5.3-Flash-Abliterated",
   "notes": "MLX; oMLX community benchmark trace; peak memory 178.9 GB at this context; duplicate traces of the same session: sbddptc1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 592.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-Abliterated-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 592.5 46.9 178.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ff7osnes",
   "source_url": "https://omlx.ai/benchmarks/performance/ff7osnes",
   "tg_tokens": null,
   "tps": 46.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-bp3js232-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: 933g49ey, qd5ouazz; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 395.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 4bit; 1k 395.2 57.6 13.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bp3js232",
   "source_url": "https://omlx.ai/benchmarks/performance/bp3js232",
   "tg_tokens": null,
   "tps": 57.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-bp3js232-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: 933g49ey, qd5ouazz; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 427.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 4bit; 4k 427.3 54.6 13.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bp3js232",
   "source_url": "https://omlx.ai/benchmarks/performance/bp3js232",
   "tg_tokens": null,
   "tps": 54.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-bp3js232-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.8 GB at this context; duplicate traces of the same session: 933g49ey, qd5ouazz; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 432.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 4bit; 8k 432.6 51.4 13.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bp3js232",
   "source_url": "https://omlx.ai/benchmarks/performance/bp3js232",
   "tg_tokens": null,
   "tps": 51.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-bp3js232-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.9 GB at this context; duplicate traces of the same session: 933g49ey, qd5ouazz; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 413.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 4bit; 16k 413.2 46.3 13.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bp3js232",
   "source_url": "https://omlx.ai/benchmarks/performance/bp3js232",
   "tg_tokens": null,
   "tps": 46.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-p53zz92u-ctx1024",
   "model": "LFM2.5-8B-A1B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 8.9 GB at this context; duplicate traces of the same session: hu5qg6uy, j6mj7pot, vejykp5e; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 669.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-8B-A1B-MLX-8bit on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 8bit; 1k 669.8 71.2 8.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p53zz92u",
   "source_url": "https://omlx.ai/benchmarks/performance/p53zz92u",
   "tg_tokens": null,
   "tps": 71.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-p53zz92u-ctx8192",
   "model": "LFM2.5-8B-A1B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 9.4 GB at this context; duplicate traces of the same session: hu5qg6uy, j6mj7pot, vejykp5e; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 731.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-8B-A1B-MLX-8bit on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 8bit; 8k 731.7 65.6 9.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p53zz92u",
   "source_url": "https://omlx.ai/benchmarks/performance/p53zz92u",
   "tg_tokens": null,
   "tps": 65.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-p53zz92u-ctx32768",
   "model": "LFM2.5-8B-A1B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 9.7 GB at this context; duplicate traces of the same session: hu5qg6uy, j6mj7pot, vejykp5e; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 637.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-8B-A1B-MLX-8bit on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 8bit; 32k 637.0 52.1 9.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p53zz92u",
   "source_url": "https://omlx.ai/benchmarks/performance/p53zz92u",
   "tg_tokens": null,
   "tps": 52.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-p53zz92u-ctx65536",
   "model": "LFM2.5-8B-A1B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 10.0 GB at this context; duplicate traces of the same session: hu5qg6uy, j6mj7pot, vejykp5e; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 539.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-8B-A1B-MLX-8bit on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 8bit; 64k 539.1 41.0 10.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p53zz92u",
   "source_url": "https://omlx.ai/benchmarks/performance/p53zz92u",
   "tg_tokens": null,
   "tps": 41.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-f0k83n44-ctx4096",
   "model": "LFM2.5-8B-A1B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 9.3 GB at this context; duplicate traces of the same session: xx67z5yc; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 736.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-8B-A1B-MLX-8bit on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 8bit; 4k 736.9 68.5 9.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f0k83n44",
   "source_url": "https://omlx.ai/benchmarks/performance/f0k83n44",
   "tg_tokens": null,
   "tps": 68.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-f0k83n44-ctx16384",
   "model": "LFM2.5-8B-A1B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 9.5 GB at this context; duplicate traces of the same session: xx67z5yc; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 692.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-8B-A1B-MLX-8bit on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 8bit; 16k 692.2 60.7 9.5 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f0k83n44",
   "source_url": "https://omlx.ai/benchmarks/performance/f0k83n44",
   "tg_tokens": null,
   "tps": 60.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-zflme0k1-ctx1024",
   "model": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B",
   "notes": "24GB LiLiCorr MLX build; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 651c4j67, dygbxc49, jdtgopzb, khi7jnrv, sxd1omf8, wveoigv2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 585.6,
   "provenance": "community",
   "quant": "unknown",
   "quote": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-LiLiCorr-MLX on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 unknown; 1k 585.6 60.5 23.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zflme0k1",
   "source_url": "https://omlx.ai/benchmarks/performance/zflme0k1",
   "tg_tokens": null,
   "tps": 60.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-zflme0k1-ctx4096",
   "model": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B",
   "notes": "24GB LiLiCorr MLX build; oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: 651c4j67, dygbxc49, jdtgopzb, khi7jnrv, sxd1omf8, wveoigv2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 597.8,
   "provenance": "community",
   "quant": "unknown",
   "quote": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-LiLiCorr-MLX on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 unknown; 4k 597.8 60.4 24.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zflme0k1",
   "source_url": "https://omlx.ai/benchmarks/performance/zflme0k1",
   "tg_tokens": null,
   "tps": 60.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-zflme0k1-ctx8192",
   "model": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B",
   "notes": "24GB LiLiCorr MLX build; oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: 651c4j67, dygbxc49, jdtgopzb, khi7jnrv, sxd1omf8, wveoigv2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 548.2,
   "provenance": "community",
   "quant": "unknown",
   "quote": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-LiLiCorr-MLX on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 unknown; 8k 548.2 59.4 24.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zflme0k1",
   "source_url": "https://omlx.ai/benchmarks/performance/zflme0k1",
   "tg_tokens": null,
   "tps": 59.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-zflme0k1-ctx16384",
   "model": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B",
   "notes": "24GB LiLiCorr MLX build; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: 651c4j67, dygbxc49, jdtgopzb, khi7jnrv, sxd1omf8, wveoigv2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 521.7,
   "provenance": "community",
   "quant": "unknown",
   "quote": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-LiLiCorr-MLX on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 unknown; 16k 521.7 56.6 24.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zflme0k1",
   "source_url": "https://omlx.ai/benchmarks/performance/zflme0k1",
   "tg_tokens": null,
   "tps": 56.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-zflme0k1-ctx32768",
   "model": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B",
   "notes": "24GB LiLiCorr MLX build; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: 651c4j67, dygbxc49, jdtgopzb, khi7jnrv, sxd1omf8, wveoigv2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 486.4,
   "provenance": "community",
   "quant": "unknown",
   "quote": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-LiLiCorr-MLX on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 unknown; 32k 486.4 54.1 24.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zflme0k1",
   "source_url": "https://omlx.ai/benchmarks/performance/zflme0k1",
   "tg_tokens": null,
   "tps": 54.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-zflme0k1-ctx65536",
   "model": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B",
   "notes": "24GB LiLiCorr MLX build; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: 651c4j67, dygbxc49, jdtgopzb, khi7jnrv, sxd1omf8, wveoigv2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 389.1,
   "provenance": "community",
   "quant": "unknown",
   "quote": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-LiLiCorr-MLX on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 unknown; 64k 389.1 40.7 24.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zflme0k1",
   "source_url": "https://omlx.ai/benchmarks/performance/zflme0k1",
   "tg_tokens": null,
   "tps": 40.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-zflme0k1-ctx131072",
   "model": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B",
   "notes": "24GB LiLiCorr MLX build; oMLX community benchmark trace; peak memory 24.8 GB at this context; duplicate traces of the same session: 651c4j67, dygbxc49, jdtgopzb, khi7jnrv, sxd1omf8, wveoigv2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 313.3,
   "provenance": "community",
   "quant": "unknown",
   "quote": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-LiLiCorr-MLX on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 unknown; 128k 313.3 35.4 24.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zflme0k1",
   "source_url": "https://omlx.ai/benchmarks/performance/zflme0k1",
   "tg_tokens": null,
   "tps": 35.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ynjfdj6q-ctx1024",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: 3i0y7eu1, 8wzi55vx, h5yr0z1y, jc1qgfjf, qm268niz, vk45rkvq, vx3uyyep; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1215.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 1,215 111.8 34.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ynjfdj6q",
   "source_url": "https://omlx.ai/benchmarks/performance/ynjfdj6q",
   "tg_tokens": null,
   "tps": 111.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ynjfdj6q-ctx4096",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: 3i0y7eu1, 8wzi55vx, h5yr0z1y, jc1qgfjf, qm268niz, vk45rkvq, vx3uyyep; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1579.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 1,579 108.0 35.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ynjfdj6q",
   "source_url": "https://omlx.ai/benchmarks/performance/ynjfdj6q",
   "tg_tokens": null,
   "tps": 108.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ynjfdj6q-ctx8192",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.9 GB at this context; duplicate traces of the same session: 3i0y7eu1, 8wzi55vx, h5yr0z1y, jc1qgfjf, qm268niz, vk45rkvq, vx3uyyep; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1464.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 8k 1,464 114.9 35.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ynjfdj6q",
   "source_url": "https://omlx.ai/benchmarks/performance/ynjfdj6q",
   "tg_tokens": null,
   "tps": 114.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ynjfdj6q-ctx16384",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.9 GB at this context; duplicate traces of the same session: 3i0y7eu1, 8wzi55vx, h5yr0z1y, jc1qgfjf, qm268niz, vk45rkvq, vx3uyyep; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1305.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 16k 1,305 92.0 35.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ynjfdj6q",
   "source_url": "https://omlx.ai/benchmarks/performance/ynjfdj6q",
   "tg_tokens": null,
   "tps": 92.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ynjfdj6q-ctx32768",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.0 GB at this context; duplicate traces of the same session: 3i0y7eu1, 8wzi55vx, h5yr0z1y, jc1qgfjf, qm268niz, vk45rkvq, vx3uyyep; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1153.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 32k 1,153 98.5 36.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ynjfdj6q",
   "source_url": "https://omlx.ai/benchmarks/performance/ynjfdj6q",
   "tg_tokens": null,
   "tps": 98.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ynjfdj6q-ctx65536",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.2 GB at this context; duplicate traces of the same session: 3i0y7eu1, 8wzi55vx, h5yr0z1y, jc1qgfjf, qm268niz, vk45rkvq, vx3uyyep; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 969.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 64k 969.2 77.9 36.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ynjfdj6q",
   "source_url": "https://omlx.ai/benchmarks/performance/ynjfdj6q",
   "tg_tokens": null,
   "tps": 77.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ynjfdj6q-ctx131072",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.5 GB at this context; duplicate traces of the same session: 3i0y7eu1, 8wzi55vx, h5yr0z1y, jc1qgfjf, qm268niz, vk45rkvq, vx3uyyep; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 617.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 128k 617.0 63.1 36.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ynjfdj6q",
   "source_url": "https://omlx.ai/benchmarks/performance/ynjfdj6q",
   "tg_tokens": null,
   "tps": 63.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ynjfdj6q-ctx199680",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.0 GB at this context; duplicate traces of the same session: 3i0y7eu1, 8wzi55vx, h5yr0z1y, jc1qgfjf, qm268niz, vk45rkvq, vx3uyyep; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 432.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 195k 432.3 41.5 37.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ynjfdj6q",
   "source_url": "https://omlx.ai/benchmarks/performance/ynjfdj6q",
   "tg_tokens": null,
   "tps": 41.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bw09x2rn-ctx1024",
   "model": "Qwen3.6-35B-A3B-Heretic",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: n3z1o12l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 987.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 987.5 136.0 19.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bw09x2rn",
   "source_url": "https://omlx.ai/benchmarks/performance/bw09x2rn",
   "tg_tokens": null,
   "tps": 136.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bw09x2rn-ctx4096",
   "model": "Qwen3.6-35B-A3B-Heretic",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: n3z1o12l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2903.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Heretic-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,903 121.1 20.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bw09x2rn",
   "source_url": "https://omlx.ai/benchmarks/performance/bw09x2rn",
   "tg_tokens": null,
   "tps": 121.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-a045frv8-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 277.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 277.1 52.0 19.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a045frv8",
   "source_url": "https://omlx.ai/benchmarks/performance/a045frv8",
   "tg_tokens": null,
   "tps": 52.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-a045frv8-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 310.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 310.5 49.9 20.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a045frv8",
   "source_url": "https://omlx.ai/benchmarks/performance/a045frv8",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-a045frv8-ctx8192",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 295.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 8k 295.4 43.9 20.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a045frv8",
   "source_url": "https://omlx.ai/benchmarks/performance/a045frv8",
   "tg_tokens": null,
   "tps": 43.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-ca95nkyw-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: rgnr2qw4, t9bj3pg0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 354.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 354.7 57.5 19.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ca95nkyw",
   "source_url": "https://omlx.ai/benchmarks/performance/ca95nkyw",
   "tg_tokens": null,
   "tps": 57.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-ca95nkyw-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: rgnr2qw4, t9bj3pg0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 394.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 394.1 55.0 20.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ca95nkyw",
   "source_url": "https://omlx.ai/benchmarks/performance/ca95nkyw",
   "tg_tokens": null,
   "tps": 55.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-ca95nkyw-ctx8192",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: rgnr2qw4, t9bj3pg0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 359.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 8k 359.1 51.5 20.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ca95nkyw",
   "source_url": "https://omlx.ai/benchmarks/performance/ca95nkyw",
   "tg_tokens": null,
   "tps": 51.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-2cr3o8s9-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 12.9 GB at this context; duplicate traces of the same session: 0c0h0uu4, 3obw7cn2, 7r966tgh, j6gmwji5, n1gotz8u, riijbz1c, x3vrxgzv; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 184.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k 184.5 26.0 12.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2cr3o8s9",
   "source_url": "https://omlx.ai/benchmarks/performance/2cr3o8s9",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-2cr3o8s9-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: 0c0h0uu4, 3obw7cn2, 7r966tgh, j6gmwji5, n1gotz8u, riijbz1c, x3vrxgzv; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 645.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 645.3 24.2 13.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2cr3o8s9",
   "source_url": "https://omlx.ai/benchmarks/performance/2cr3o8s9",
   "tg_tokens": null,
   "tps": 24.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5",
   "id": "omlx-2cr3o8s9-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: 0c0h0uu4, 3obw7cn2, 7r966tgh, j6gmwji5, n1gotz8u, riijbz1c, x3vrxgzv; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 647.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 8k 647.4 22.7 13.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2cr3o8s9",
   "source_url": "https://omlx.ai/benchmarks/performance/2cr3o8s9",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5",
   "id": "omlx-2cr3o8s9-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 13.9 GB at this context; duplicate traces of the same session: 0c0h0uu4, 3obw7cn2, 7r966tgh, j6gmwji5, n1gotz8u, riijbz1c, x3vrxgzv; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 615.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 16k 615.3 20.3 13.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2cr3o8s9",
   "source_url": "https://omlx.ai/benchmarks/performance/2cr3o8s9",
   "tg_tokens": null,
   "tps": 20.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5",
   "id": "omlx-2cr3o8s9-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 14.2 GB at this context; duplicate traces of the same session: 0c0h0uu4, 3obw7cn2, 7r966tgh, j6gmwji5, n1gotz8u, riijbz1c, x3vrxgzv; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 566.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 32k 566.6 18.4 14.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2cr3o8s9",
   "source_url": "https://omlx.ai/benchmarks/performance/2cr3o8s9",
   "tg_tokens": null,
   "tps": 18.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-2cr3o8s9-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: 0c0h0uu4, 3obw7cn2, 7r966tgh, j6gmwji5, n1gotz8u, riijbz1c, x3vrxgzv; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 473.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 64k 473.3 14.8 14.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2cr3o8s9",
   "source_url": "https://omlx.ai/benchmarks/performance/2cr3o8s9",
   "tg_tokens": null,
   "tps": 14.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-2cr3o8s9-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: 0c0h0uu4, 3obw7cn2, 7r966tgh, j6gmwji5, n1gotz8u, riijbz1c, x3vrxgzv; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 377.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 128k 377.6 10.6 16.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2cr3o8s9",
   "source_url": "https://omlx.ai/benchmarks/performance/2cr3o8s9",
   "tg_tokens": null,
   "tps": 10.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-2cr3o8s9-ctx199680",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: 0c0h0uu4, 3obw7cn2, 7r966tgh, j6gmwji5, n1gotz8u, riijbz1c, x3vrxgzv; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 310.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 195k 310.3 7.7 17.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2cr3o8s9",
   "source_url": "https://omlx.ai/benchmarks/performance/2cr3o8s9",
   "tg_tokens": null,
   "tps": 7.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-al5ukake-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: z8747bj6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 963.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 1k 963.3 60.3 29.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/al5ukake",
   "source_url": "https://omlx.ai/benchmarks/performance/al5ukake",
   "tg_tokens": null,
   "tps": 60.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-al5ukake-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.6 GB at this context; duplicate traces of the same session: z8747bj6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2415.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 4k 2,415 85.8 30.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/al5ukake",
   "source_url": "https://omlx.ai/benchmarks/performance/al5ukake",
   "tg_tokens": null,
   "tps": 85.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-mkmzup9p-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 3b39ztjc, k16pgqb6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 248.3,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 248.3 24.5 19.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mkmzup9p",
   "source_url": "https://omlx.ai/benchmarks/performance/mkmzup9p",
   "tg_tokens": null,
   "tps": 24.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-mkmzup9p-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 3b39ztjc, k16pgqb6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 201.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 201.0 20.5 21.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mkmzup9p",
   "source_url": "https://omlx.ai/benchmarks/performance/mkmzup9p",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-mkmzup9p-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: 3b39ztjc, k16pgqb6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 164.2,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 164.2 13.8 22.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mkmzup9p",
   "source_url": "https://omlx.ai/benchmarks/performance/mkmzup9p",
   "tg_tokens": null,
   "tps": 13.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-v0prj1ws-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 9vo80vn7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 430.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 430.3 29.5 17.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v0prj1ws",
   "source_url": "https://omlx.ai/benchmarks/performance/v0prj1ws",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-v0prj1ws-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: 9vo80vn7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 467.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 467.3 37.1 18.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v0prj1ws",
   "source_url": "https://omlx.ai/benchmarks/performance/v0prj1ws",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-x2wqjzh6-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: h5c7cea9, o11nlxrn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 250.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 250.4 18.4 26.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x2wqjzh6",
   "source_url": "https://omlx.ai/benchmarks/performance/x2wqjzh6",
   "tg_tokens": null,
   "tps": 18.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-x2wqjzh6-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.3 GB at this context; duplicate traces of the same session: h5c7cea9, o11nlxrn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 212.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 212.4 16.6 27.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x2wqjzh6",
   "source_url": "https://omlx.ai/benchmarks/performance/x2wqjzh6",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-x2wqjzh6-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.1 GB at this context; duplicate traces of the same session: h5c7cea9, o11nlxrn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 187.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 187.3 11.4 29.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x2wqjzh6",
   "source_url": "https://omlx.ai/benchmarks/performance/x2wqjzh6",
   "tg_tokens": null,
   "tps": 11.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-50r0ft3g-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.7 GB at this context; duplicate traces of the same session: 2u4gycb2, 9lqedtjb, q16zq0uz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 667.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 667.9 38.6 100.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/50r0ft3g",
   "source_url": "https://omlx.ai/benchmarks/performance/50r0ft3g",
   "tg_tokens": null,
   "tps": 38.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-50r0ft3g-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: 2u4gycb2, 9lqedtjb, q16zq0uz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1079.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,079 81.6 101.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/50r0ft3g",
   "source_url": "https://omlx.ai/benchmarks/performance/50r0ft3g",
   "tg_tokens": null,
   "tps": 81.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-50r0ft3g-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.1 GB at this context; duplicate traces of the same session: 2u4gycb2, 9lqedtjb, q16zq0uz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1064.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 1,064 31.9 102.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/50r0ft3g",
   "source_url": "https://omlx.ai/benchmarks/performance/50r0ft3g",
   "tg_tokens": null,
   "tps": 31.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-50r0ft3g-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.1 GB at this context; duplicate traces of the same session: 2u4gycb2, 9lqedtjb, q16zq0uz; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1062.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 1,062 59.9 103.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/50r0ft3g",
   "source_url": "https://omlx.ai/benchmarks/performance/50r0ft3g",
   "tg_tokens": null,
   "tps": 59.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-u7wi64nf-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.6 GB; TTFT 9752 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 105.0 PP tok/s 18.8 TG tok/s 9752 TTFT (ms) 70.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u7wi64nf",
   "source_url": "https://omlx.ai/benchmarks/performance/u7wi64nf",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-oiczc0pl-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 100.8 GB; TTFT 606 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1689.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,689 PP tok/s 48.0 TG tok/s 606 TTFT (ms) 100.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oiczc0pl",
   "source_url": "https://omlx.ai/benchmarks/performance/oiczc0pl",
   "tg_tokens": null,
   "tps": 48.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-h4vsrjkq-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 958.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 958.0 90.6 21.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h4vsrjkq",
   "source_url": "https://omlx.ai/benchmarks/performance/h4vsrjkq",
   "tg_tokens": null,
   "tps": 90.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-h4vsrjkq-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1029.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 1,029 100.2 22.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h4vsrjkq",
   "source_url": "https://omlx.ai/benchmarks/performance/h4vsrjkq",
   "tg_tokens": null,
   "tps": 100.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-h4vsrjkq-ctx8192",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 972.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 972.5 81.1 22.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h4vsrjkq",
   "source_url": "https://omlx.ai/benchmarks/performance/h4vsrjkq",
   "tg_tokens": null,
   "tps": 81.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-h4vsrjkq-ctx16384",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 865.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 865.1 82.5 22.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h4vsrjkq",
   "source_url": "https://omlx.ai/benchmarks/performance/h4vsrjkq",
   "tg_tokens": null,
   "tps": 82.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-h4vsrjkq-ctx32768",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 699.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 699.8 80.8 23.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h4vsrjkq",
   "source_url": "https://omlx.ai/benchmarks/performance/h4vsrjkq",
   "tg_tokens": null,
   "tps": 80.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-h4vsrjkq-ctx65536",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 513.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 513.9 52.6 24.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h4vsrjkq",
   "source_url": "https://omlx.ai/benchmarks/performance/h4vsrjkq",
   "tg_tokens": null,
   "tps": 52.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-pxydzhg6-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: 53ghjvjt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1335.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ6e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 1,335 60.0 28.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pxydzhg6",
   "source_url": "https://omlx.ai/benchmarks/performance/pxydzhg6",
   "tg_tokens": null,
   "tps": 60.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-pxydzhg6-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: 53ghjvjt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1385.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ6e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 1,385 90.6 30.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pxydzhg6",
   "source_url": "https://omlx.ai/benchmarks/performance/pxydzhg6",
   "tg_tokens": null,
   "tps": 90.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-pxydzhg6-ctx8192",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 30.4 GB at this context; duplicate traces of the same session: 53ghjvjt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1734.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ6e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 8k 1,734 60.8 30.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pxydzhg6",
   "source_url": "https://omlx.ai/benchmarks/performance/pxydzhg6",
   "tg_tokens": null,
   "tps": 60.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-pxydzhg6-ctx16384",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 30.7 GB at this context; duplicate traces of the same session: 53ghjvjt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1547.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ6e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k 1,547 118.6 30.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pxydzhg6",
   "source_url": "https://omlx.ai/benchmarks/performance/pxydzhg6",
   "tg_tokens": null,
   "tps": 118.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-pxydzhg6-ctx32768",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 31.4 GB at this context; duplicate traces of the same session: 53ghjvjt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1231.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ6e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 1,231 57.3 31.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pxydzhg6",
   "source_url": "https://omlx.ai/benchmarks/performance/pxydzhg6",
   "tg_tokens": null,
   "tps": 57.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-pxydzhg6-ctx65536",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 32.7 GB at this context; duplicate traces of the same session: 53ghjvjt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 880.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ6e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 64k 880.0 50.6 32.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pxydzhg6",
   "source_url": "https://omlx.ai/benchmarks/performance/pxydzhg6",
   "tg_tokens": null,
   "tps": 50.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-pxydzhg6-ctx131072",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: 53ghjvjt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 488.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ6e on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 128k 488.8 55.2 34.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pxydzhg6",
   "source_url": "https://omlx.ai/benchmarks/performance/pxydzhg6",
   "tg_tokens": null,
   "tps": 55.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-840e0vx3-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 800.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E4B-it-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k 800.6 46.3 6.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/840e0vx3",
   "source_url": "https://omlx.ai/benchmarks/performance/840e0vx3",
   "tg_tokens": null,
   "tps": 46.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-840e0vx3-ctx16384",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 811.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E4B-it-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 16k 811.4 39.4 6.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/840e0vx3",
   "source_url": "https://omlx.ai/benchmarks/performance/840e0vx3",
   "tg_tokens": null,
   "tps": 39.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-3rww690l-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 175.8 GB at this context; duplicate traces of the same session: 8sqjhev7, gg16fq1n, jt3c7pdo; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 362.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 362.8 30.0 175.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3rww690l",
   "source_url": "https://omlx.ai/benchmarks/performance/3rww690l",
   "tg_tokens": null,
   "tps": 30.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-3rww690l-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.2 GB at this context; duplicate traces of the same session: 8sqjhev7, gg16fq1n, jt3c7pdo; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 438.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 438.6 30.9 177.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3rww690l",
   "source_url": "https://omlx.ai/benchmarks/performance/3rww690l",
   "tg_tokens": null,
   "tps": 30.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3rww690l-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.6 GB at this context; duplicate traces of the same session: 8sqjhev7, gg16fq1n, jt3c7pdo; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 383.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 383.8 63.9 177.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3rww690l",
   "source_url": "https://omlx.ai/benchmarks/performance/3rww690l",
   "tg_tokens": null,
   "tps": 63.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3rww690l-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.0 GB at this context; duplicate traces of the same session: 8sqjhev7, gg16fq1n, jt3c7pdo; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 419.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 419.3 80.6 178.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3rww690l",
   "source_url": "https://omlx.ai/benchmarks/performance/3rww690l",
   "tg_tokens": null,
   "tps": 80.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-wzc0t2j6-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: sple1792; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 201.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M3 (10c); M3 (10c) \u00b7 16 GB \u00b7 4bit; 4k 201.8 18.5 6.3 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wzc0t2j6",
   "source_url": "https://omlx.ai/benchmarks/performance/wzc0t2j6",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3",
   "id": "omlx-wzc0t2j6-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: sple1792; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 199.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M3 (10c); M3 (10c) \u00b7 16 GB \u00b7 4bit; 8k 199.6 17.3 6.4 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wzc0t2j6",
   "source_url": "https://omlx.ai/benchmarks/performance/wzc0t2j6",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-wbh2luux-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: srylx1an; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 544.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k 544.9 67.5 20.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wbh2luux",
   "source_url": "https://omlx.ai/benchmarks/performance/wbh2luux",
   "tg_tokens": null,
   "tps": 67.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-wbh2luux-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: srylx1an; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 449.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 16k 449.6 57.0 20.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wbh2luux",
   "source_url": "https://omlx.ai/benchmarks/performance/wbh2luux",
   "tg_tokens": null,
   "tps": 57.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ncuxz754-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: eh2phj1u; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1361.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 1,361 100.1 24.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ncuxz754",
   "source_url": "https://omlx.ai/benchmarks/performance/ncuxz754",
   "tg_tokens": null,
   "tps": 100.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ncuxz754-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: eh2phj1u; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 883.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k 883.0 79.3 25.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ncuxz754",
   "source_url": "https://omlx.ai/benchmarks/performance/ncuxz754",
   "tg_tokens": null,
   "tps": 79.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-46lmy8ek-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: yvez28o9; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1024.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 1,024 124.7 21.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/46lmy8ek",
   "source_url": "https://omlx.ai/benchmarks/performance/46lmy8ek",
   "tg_tokens": null,
   "tps": 124.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-46lmy8ek-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: yvez28o9; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3162.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 3,162 114.5 22.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/46lmy8ek",
   "source_url": "https://omlx.ai/benchmarks/performance/46lmy8ek",
   "tg_tokens": null,
   "tps": 114.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-dnhx8un1-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: cvz4wk5d; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2037.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 1k 2,037 107.2 29.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dnhx8un1",
   "source_url": "https://omlx.ai/benchmarks/performance/dnhx8un1",
   "tg_tokens": null,
   "tps": 107.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-dnhx8un1-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.6 GB at this context; duplicate traces of the same session: cvz4wk5d; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2763.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 4k 2,763 90.4 30.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dnhx8un1",
   "source_url": "https://omlx.ai/benchmarks/performance/dnhx8un1",
   "tg_tokens": null,
   "tps": 90.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-rdpt24jp-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 16.0 GB; TTFT 9309 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 110.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k tokens 110.0 PP tok/s 19.1 TG tok/s 9309 TTFT (ms) 16.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rdpt24jp",
   "source_url": "https://omlx.ai/benchmarks/performance/rdpt24jp",
   "tg_tokens": null,
   "tps": 19.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-amvo3sje-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 110.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 110.6 13.5 19.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/amvo3sje",
   "source_url": "https://omlx.ai/benchmarks/performance/amvo3sje",
   "tg_tokens": null,
   "tps": 13.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-amvo3sje-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 84.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 32k 84.8 16.6 26.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/amvo3sje",
   "source_url": "https://omlx.ai/benchmarks/performance/amvo3sje",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-iqq3or33-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: fkofqmj9; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 106.3 15.0 20.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iqq3or33",
   "source_url": "https://omlx.ai/benchmarks/performance/iqq3or33",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-iqq3or33-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: fkofqmj9; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 93.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 93.8 21.0 26.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iqq3or33",
   "source_url": "https://omlx.ai/benchmarks/performance/iqq3or33",
   "tg_tokens": null,
   "tps": 21.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xoscnzb6-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace, single run at 1k context; peak memory 73.6 GB; TTFT 943 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1085.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,085 PP tok/s 71.5 TG tok/s 943 TTFT (ms) 73.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xoscnzb6",
   "source_url": "https://omlx.ai/benchmarks/performance/xoscnzb6",
   "tg_tokens": null,
   "tps": 71.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wfd407c4-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 73.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1238.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,238 81.0 73.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wfd407c4",
   "source_url": "https://omlx.ai/benchmarks/performance/wfd407c4",
   "tg_tokens": null,
   "tps": 81.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wfd407c4-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 74.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2147.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,147 69.6 74.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wfd407c4",
   "source_url": "https://omlx.ai/benchmarks/performance/wfd407c4",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-bhymquec-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.7 GB at this context; duplicate traces of the same session: 41gtamau, e40xtgkn, ga1apcg1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 921.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 921.3 80.3 100.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhymquec",
   "source_url": "https://omlx.ai/benchmarks/performance/bhymquec",
   "tg_tokens": null,
   "tps": 80.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-bhymquec-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: 41gtamau, e40xtgkn, ga1apcg1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1091.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 1,091 84.1 101.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhymquec",
   "source_url": "https://omlx.ai/benchmarks/performance/bhymquec",
   "tg_tokens": null,
   "tps": 84.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-bhymquec-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: 41gtamau, e40xtgkn, ga1apcg1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1080.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,080 76.3 101.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhymquec",
   "source_url": "https://omlx.ai/benchmarks/performance/bhymquec",
   "tg_tokens": null,
   "tps": 76.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-bhymquec-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 103.1 GB at this context; duplicate traces of the same session: 41gtamau, e40xtgkn, ga1apcg1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1064.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 1,064 73.2 103.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhymquec",
   "source_url": "https://omlx.ai/benchmarks/performance/bhymquec",
   "tg_tokens": null,
   "tps": 73.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-eq1gh370-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 4gvt58l1, i9qms8ht, yva8hjm4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 603.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 603.1 52.0 70.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eq1gh370",
   "source_url": "https://omlx.ai/benchmarks/performance/eq1gh370",
   "tg_tokens": null,
   "tps": 52.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-eq1gh370-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: 4gvt58l1, i9qms8ht, yva8hjm4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 833.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 833.0 57.8 71.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eq1gh370",
   "source_url": "https://omlx.ai/benchmarks/performance/eq1gh370",
   "tg_tokens": null,
   "tps": 57.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-eq1gh370-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: 4gvt58l1, i9qms8ht, yva8hjm4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1064.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,064 65.5 71.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eq1gh370",
   "source_url": "https://omlx.ai/benchmarks/performance/eq1gh370",
   "tg_tokens": null,
   "tps": 65.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-eq1gh370-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.3 GB at this context; duplicate traces of the same session: 4gvt58l1, i9qms8ht, yva8hjm4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1064.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 1,064 68.1 72.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eq1gh370",
   "source_url": "https://omlx.ai/benchmarks/performance/eq1gh370",
   "tg_tokens": null,
   "tps": 68.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-s768cdkv-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 51.2 GB at this context; duplicate traces of the same session: 4zjab7vv, 9qhtlqo7; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 131.3 27.6 51.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s768cdkv",
   "source_url": "https://omlx.ai/benchmarks/performance/s768cdkv",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-s768cdkv-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 51.9 GB at this context; duplicate traces of the same session: 4zjab7vv, 9qhtlqo7; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 532.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 532.6 26.8 51.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s768cdkv",
   "source_url": "https://omlx.ai/benchmarks/performance/s768cdkv",
   "tg_tokens": null,
   "tps": 26.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-s768cdkv-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 52.9 GB at this context; duplicate traces of the same session: 4zjab7vv, 9qhtlqo7; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 700.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 700.1 23.8 52.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s768cdkv",
   "source_url": "https://omlx.ai/benchmarks/performance/s768cdkv",
   "tg_tokens": null,
   "tps": 23.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-utfkp8e4-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 100.8 GB; TTFT 4537 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 225.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 225.7 PP tok/s 22.2 TG tok/s 4537 TTFT (ms) 100.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/utfkp8e4",
   "source_url": "https://omlx.ai/benchmarks/performance/utfkp8e4",
   "tg_tokens": null,
   "tps": 22.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-z17u5psn-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 76.8 GB; TTFT 202464 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 647.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k tokens 647.4 PP tok/s 22.5 TG tok/s 202464 TTFT (ms) 76.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z17u5psn",
   "source_url": "https://omlx.ai/benchmarks/performance/z17u5psn",
   "tg_tokens": null,
   "tps": 22.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-z2nhr5w8-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 71.1 GB; TTFT 1686 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 607.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 607.2 PP tok/s 20.0 TG tok/s 1686 TTFT (ms) 71.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z2nhr5w8",
   "source_url": "https://omlx.ai/benchmarks/performance/z2nhr5w8",
   "tg_tokens": null,
   "tps": 20.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bazzz9jl-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 101.1 GB; TTFT 2007 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2041.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 2,041 PP tok/s 49.9 TG tok/s 2007 TTFT (ms) 101.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bazzz9jl",
   "source_url": "https://omlx.ai/benchmarks/performance/bazzz9jl",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-dnlaoj9h-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 287.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 287.6 50.5 22.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dnlaoj9h",
   "source_url": "https://omlx.ai/benchmarks/performance/dnlaoj9h",
   "tg_tokens": null,
   "tps": 50.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-dnlaoj9h-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 223.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 8k 223.4 40.9 23.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dnlaoj9h",
   "source_url": "https://omlx.ai/benchmarks/performance/dnlaoj9h",
   "tg_tokens": null,
   "tps": 40.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-dnlaoj9h-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 171.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 171.8 43.3 24.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dnlaoj9h",
   "source_url": "https://omlx.ai/benchmarks/performance/dnlaoj9h",
   "tg_tokens": null,
   "tps": 43.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-dnlaoj9h-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 188.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 188.1 27.3 23.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dnlaoj9h",
   "source_url": "https://omlx.ai/benchmarks/performance/dnlaoj9h",
   "tg_tokens": null,
   "tps": 27.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-nuevu2xq-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: jbtp8tv9, lskmvf54, neuk5481, vgvg8kip; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1513.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 1,513 58.1 19.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nuevu2xq",
   "source_url": "https://omlx.ai/benchmarks/performance/nuevu2xq",
   "tg_tokens": null,
   "tps": 58.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-nuevu2xq-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: jbtp8tv9, lskmvf54, neuk5481, vgvg8kip; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1982.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,982 55.9 20.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nuevu2xq",
   "source_url": "https://omlx.ai/benchmarks/performance/nuevu2xq",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-nuevu2xq-ctx8192",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: jbtp8tv9, lskmvf54, neuk5481, vgvg8kip; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1918.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 8k 1,918 53.2 20.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nuevu2xq",
   "source_url": "https://omlx.ai/benchmarks/performance/nuevu2xq",
   "tg_tokens": null,
   "tps": 53.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-nuevu2xq-ctx16384",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: jbtp8tv9, lskmvf54, neuk5481, vgvg8kip; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1789.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 1,789 47.8 20.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nuevu2xq",
   "source_url": "https://omlx.ai/benchmarks/performance/nuevu2xq",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-nuevu2xq-ctx32768",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: jbtp8tv9, lskmvf54, neuk5481, vgvg8kip; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1508.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 32k 1,508 40.7 21.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nuevu2xq",
   "source_url": "https://omlx.ai/benchmarks/performance/nuevu2xq",
   "tg_tokens": null,
   "tps": 40.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-dlzh3et6-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: w5c05o2d; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1470.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 1,470 56.4 19.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dlzh3et6",
   "source_url": "https://omlx.ai/benchmarks/performance/dlzh3et6",
   "tg_tokens": null,
   "tps": 56.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-dlzh3et6-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: w5c05o2d; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1951.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,951 55.4 20.5 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dlzh3et6",
   "source_url": "https://omlx.ai/benchmarks/performance/dlzh3et6",
   "tg_tokens": null,
   "tps": 55.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-s6wh0ass-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 8k7lvngl, 9u0cq6kt, s21shyxi, xf4vwgc3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 958.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 958.0 90.6 21.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s6wh0ass",
   "source_url": "https://omlx.ai/benchmarks/performance/s6wh0ass",
   "tg_tokens": null,
   "tps": 90.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-s6wh0ass-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 8k7lvngl, 9u0cq6kt, s21shyxi, xf4vwgc3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1029.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 1,029 100.2 22.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s6wh0ass",
   "source_url": "https://omlx.ai/benchmarks/performance/s6wh0ass",
   "tg_tokens": null,
   "tps": 100.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-s6wh0ass-ctx8192",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 8k7lvngl, 9u0cq6kt, s21shyxi, xf4vwgc3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 972.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 972.5 81.1 22.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s6wh0ass",
   "source_url": "https://omlx.ai/benchmarks/performance/s6wh0ass",
   "tg_tokens": null,
   "tps": 81.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-s6wh0ass-ctx16384",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: 8k7lvngl, 9u0cq6kt, s21shyxi, xf4vwgc3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 865.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 865.1 82.5 22.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s6wh0ass",
   "source_url": "https://omlx.ai/benchmarks/performance/s6wh0ass",
   "tg_tokens": null,
   "tps": 82.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-s6wh0ass-ctx32768",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: 8k7lvngl, 9u0cq6kt, s21shyxi, xf4vwgc3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 699.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 699.8 80.8 23.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s6wh0ass",
   "source_url": "https://omlx.ai/benchmarks/performance/s6wh0ass",
   "tg_tokens": null,
   "tps": 80.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-s6wh0ass-ctx65536",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: 8k7lvngl, 9u0cq6kt, s21shyxi, xf4vwgc3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 513.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 513.9 52.6 24.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s6wh0ass",
   "source_url": "https://omlx.ai/benchmarks/performance/s6wh0ass",
   "tg_tokens": null,
   "tps": 52.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-pbkjuodz-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 9evus7nh, dfv4ajrq, epnx65aj, p3c1ueej; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 959.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 959.7 108.6 21.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pbkjuodz",
   "source_url": "https://omlx.ai/benchmarks/performance/pbkjuodz",
   "tg_tokens": null,
   "tps": 108.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-pbkjuodz-ctx8192",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 9evus7nh, dfv4ajrq, epnx65aj, p3c1ueej; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 971.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 971.3 78.7 22.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pbkjuodz",
   "source_url": "https://omlx.ai/benchmarks/performance/pbkjuodz",
   "tg_tokens": null,
   "tps": 78.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-pbkjuodz-ctx16384",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: 9evus7nh, dfv4ajrq, epnx65aj, p3c1ueej; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 865.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 865.4 68.3 22.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pbkjuodz",
   "source_url": "https://omlx.ai/benchmarks/performance/pbkjuodz",
   "tg_tokens": null,
   "tps": 68.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-pbkjuodz-ctx32768",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: 9evus7nh, dfv4ajrq, epnx65aj, p3c1ueej; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 705.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 705.6 75.1 23.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pbkjuodz",
   "source_url": "https://omlx.ai/benchmarks/performance/pbkjuodz",
   "tg_tokens": null,
   "tps": 75.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-pbkjuodz-ctx65536",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: 9evus7nh, dfv4ajrq, epnx65aj, p3c1ueej; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 514.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 514.8 60.0 24.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pbkjuodz",
   "source_url": "https://omlx.ai/benchmarks/performance/pbkjuodz",
   "tg_tokens": null,
   "tps": 60.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-j56ksri1-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: dxpkx33p; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 761.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 761.6 50.7 21.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j56ksri1",
   "source_url": "https://omlx.ai/benchmarks/performance/j56ksri1",
   "tg_tokens": null,
   "tps": 50.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-j56ksri1-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: dxpkx33p; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1029.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 1,029 77.2 22.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j56ksri1",
   "source_url": "https://omlx.ai/benchmarks/performance/j56ksri1",
   "tg_tokens": null,
   "tps": 77.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-drpac0qg-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 800.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E4B-it-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k 800.6 46.3 6.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/drpac0qg",
   "source_url": "https://omlx.ai/benchmarks/performance/drpac0qg",
   "tg_tokens": null,
   "tps": 46.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-drpac0qg-ctx16384",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 811.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E4B-it-MLX-4bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 16k 811.4 39.4 6.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/drpac0qg",
   "source_url": "https://omlx.ai/benchmarks/performance/drpac0qg",
   "tg_tokens": null,
   "tps": 39.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-nvcalkq2-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: iwm7lu1u; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 539.3,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 1k 539.3 69.9 11.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nvcalkq2",
   "source_url": "https://omlx.ai/benchmarks/performance/nvcalkq2",
   "tg_tokens": null,
   "tps": 69.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-nvcalkq2-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: iwm7lu1u; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 627.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 4k 627.0 65.9 11.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nvcalkq2",
   "source_url": "https://omlx.ai/benchmarks/performance/nvcalkq2",
   "tg_tokens": null,
   "tps": 65.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-2x7fnzo4-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: bcmy9ppx; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2355.0,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "mlx-community--Qwen3.6-35B-A3B-nvfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 2,355 102.0 19.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2x7fnzo4",
   "source_url": "https://omlx.ai/benchmarks/performance/2x7fnzo4",
   "tg_tokens": null,
   "tps": 102.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-2x7fnzo4-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: bcmy9ppx; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2985.0,
   "provenance": "community",
   "quant": "nvfp4",
   "quote": "mlx-community--Qwen3.6-35B-A3B-nvfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 2,985 100.7 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2x7fnzo4",
   "source_url": "https://omlx.ai/benchmarks/performance/2x7fnzo4",
   "tg_tokens": null,
   "tps": 100.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-s4sbvj6o-ctx1024",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 157.0 GB at this context; duplicate traces of the same session: 6i8aapkv, aep6bwe8, lomtvlme; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 346.8,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "DeepSeek-V4-Flash-0731-MXFP4-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 346.8 42.1 157.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s4sbvj6o",
   "source_url": "https://omlx.ai/benchmarks/performance/s4sbvj6o",
   "tg_tokens": null,
   "tps": 42.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-s4sbvj6o-ctx4096",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 157.7 GB at this context; duplicate traces of the same session: 6i8aapkv, aep6bwe8, lomtvlme; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 437.0,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "DeepSeek-V4-Flash-0731-MXFP4-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 437.0 40.6 157.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s4sbvj6o",
   "source_url": "https://omlx.ai/benchmarks/performance/s4sbvj6o",
   "tg_tokens": null,
   "tps": 40.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-s4sbvj6o-ctx8192",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 158.1 GB at this context; duplicate traces of the same session: 6i8aapkv, aep6bwe8, lomtvlme; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 459.9,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "DeepSeek-V4-Flash-0731-MXFP4-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 459.9 47.8 158.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s4sbvj6o",
   "source_url": "https://omlx.ai/benchmarks/performance/s4sbvj6o",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-s4sbvj6o-ctx16384",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 159.1 GB at this context; duplicate traces of the same session: 6i8aapkv, aep6bwe8, lomtvlme; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 451.6,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "DeepSeek-V4-Flash-0731-MXFP4-MLX on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 451.6 45.3 159.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s4sbvj6o",
   "source_url": "https://omlx.ai/benchmarks/performance/s4sbvj6o",
   "tg_tokens": null,
   "tps": 45.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-l575eru0-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 175.8 GB at this context; duplicate traces of the same session: gjk41hp1, rocmbv8b, y3v44j3j; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 409.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 409.6 38.1 175.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l575eru0",
   "source_url": "https://omlx.ai/benchmarks/performance/l575eru0",
   "tg_tokens": null,
   "tps": 38.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-l575eru0-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.2 GB at this context; duplicate traces of the same session: gjk41hp1, rocmbv8b, y3v44j3j; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 458.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 458.8 33.2 177.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l575eru0",
   "source_url": "https://omlx.ai/benchmarks/performance/l575eru0",
   "tg_tokens": null,
   "tps": 33.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-l575eru0-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.6 GB at this context; duplicate traces of the same session: gjk41hp1, rocmbv8b, y3v44j3j; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 426.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 426.5 64.2 177.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l575eru0",
   "source_url": "https://omlx.ai/benchmarks/performance/l575eru0",
   "tg_tokens": null,
   "tps": 64.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-l575eru0-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.0 GB at this context; duplicate traces of the same session: gjk41hp1, rocmbv8b, y3v44j3j; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 418.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 418.7 86.3 178.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l575eru0",
   "source_url": "https://omlx.ai/benchmarks/performance/l575eru0",
   "tg_tokens": null,
   "tps": 86.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ty9t73hm-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.4 GB at this context; duplicate traces of the same session: 2vac444i, e7zuahso; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 809.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 809.4 70.3 178.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ty9t73hm",
   "source_url": "https://omlx.ai/benchmarks/performance/ty9t73hm",
   "tg_tokens": null,
   "tps": 70.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ty9t73hm-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.0 GB at this context; duplicate traces of the same session: 2vac444i, e7zuahso; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1874.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,874 105.0 179.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ty9t73hm",
   "source_url": "https://omlx.ai/benchmarks/performance/ty9t73hm",
   "tg_tokens": null,
   "tps": 105.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ty9t73hm-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.7 GB at this context; duplicate traces of the same session: 2vac444i, e7zuahso; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1871.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,871 94.9 179.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ty9t73hm",
   "source_url": "https://omlx.ai/benchmarks/performance/ty9t73hm",
   "tg_tokens": null,
   "tps": 94.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-v9ubkxtt-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.4 GB at this context; duplicate traces of the same session: ftts4a34; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2165.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 2,165 52.5 178.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v9ubkxtt",
   "source_url": "https://omlx.ai/benchmarks/performance/v9ubkxtt",
   "tg_tokens": null,
   "tps": 52.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-v9ubkxtt-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.9 GB at this context; duplicate traces of the same session: ftts4a34; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2248.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 2,248 113.4 178.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v9ubkxtt",
   "source_url": "https://omlx.ai/benchmarks/performance/v9ubkxtt",
   "tg_tokens": null,
   "tps": 113.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-njsdgs2e-ctx1024",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: gx7umq0k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1274.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "MiniCPM-V-4.7-35B-A3B-mxfp4 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,274 100.2 22.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/njsdgs2e",
   "source_url": "https://omlx.ai/benchmarks/performance/njsdgs2e",
   "tg_tokens": null,
   "tps": 100.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-njsdgs2e-ctx4096",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: gx7umq0k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3583.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "MiniCPM-V-4.7-35B-A3B-mxfp4 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 3,583 101.6 22.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/njsdgs2e",
   "source_url": "https://omlx.ai/benchmarks/performance/njsdgs2e",
   "tg_tokens": null,
   "tps": 101.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-93c6f4gr-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 38.3 GB; TTFT 12537 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5227.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 64k tokens 5,227 PP tok/s 155.5 TG tok/s 12537 TTFT (ms) 38.3 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/93c6f4gr",
   "source_url": "https://omlx.ai/benchmarks/performance/93c6f4gr",
   "tg_tokens": null,
   "tps": 155.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-i61wkee7-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.0 GB at this context; duplicate traces of the same session: h2f5ij6t; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 8213.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 4k 8,213 279.0 37.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i61wkee7",
   "source_url": "https://omlx.ai/benchmarks/performance/i61wkee7",
   "tg_tokens": null,
   "tps": 279.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-i61wkee7-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.3 GB at this context; duplicate traces of the same session: h2f5ij6t; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5270.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 64k 5,270 140.9 38.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i61wkee7",
   "source_url": "https://omlx.ai/benchmarks/performance/i61wkee7",
   "tg_tokens": null,
   "tps": 140.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-214sq9t4-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.0 GB at this context; duplicate traces of the same session: izv5u8a8; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 8182.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 4k 8,182 306.3 37.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/214sq9t4",
   "source_url": "https://omlx.ai/benchmarks/performance/214sq9t4",
   "tg_tokens": null,
   "tps": 306.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-214sq9t4-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.3 GB at this context; duplicate traces of the same session: izv5u8a8; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5278.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 64k 5,278 140.5 38.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/214sq9t4",
   "source_url": "https://omlx.ai/benchmarks/performance/214sq9t4",
   "tg_tokens": null,
   "tps": 140.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-cafuurbj-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.0 GB at this context; duplicate traces of the same session: 0um1xj2p, 3nt8u3iu; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 7966.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 4k 7,966 256.9 37.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cafuurbj",
   "source_url": "https://omlx.ai/benchmarks/performance/cafuurbj",
   "tg_tokens": null,
   "tps": 256.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-cafuurbj-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: 0um1xj2p, 3nt8u3iu; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 7804.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 16k 7,804 259.4 37.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cafuurbj",
   "source_url": "https://omlx.ai/benchmarks/performance/cafuurbj",
   "tg_tokens": null,
   "tps": 259.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-cafuurbj-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.3 GB at this context; duplicate traces of the same session: 0um1xj2p, 3nt8u3iu; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5192.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 64k 5,192 163.5 38.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cafuurbj",
   "source_url": "https://omlx.ai/benchmarks/performance/cafuurbj",
   "tg_tokens": null,
   "tps": 163.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ue65cdbd-ctx32768",
   "model": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 10.9 GB at this context; duplicate traces of the same session: i3lg3tma, rc9b4lxs; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 793.9,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-oQ4e-fp16-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 793.9 130.4 10.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ue65cdbd",
   "source_url": "https://omlx.ai/benchmarks/performance/ue65cdbd",
   "tg_tokens": null,
   "tps": 130.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ue65cdbd-ctx65536",
   "model": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 12.8 GB at this context; duplicate traces of the same session: i3lg3tma, rc9b4lxs; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 676.7,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-oQ4e-fp16-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 676.7 98.8 12.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ue65cdbd",
   "source_url": "https://omlx.ai/benchmarks/performance/ue65cdbd",
   "tg_tokens": null,
   "tps": 98.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ue65cdbd-ctx131072",
   "model": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: i3lg3tma, rc9b4lxs; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 520.2,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-oQ4e-fp16-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 520.2 77.8 16.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ue65cdbd",
   "source_url": "https://omlx.ai/benchmarks/performance/ue65cdbd",
   "tg_tokens": null,
   "tps": 77.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-dpd3m9ml-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 110.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 110.6 13.5 19.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dpd3m9ml",
   "source_url": "https://omlx.ai/benchmarks/performance/dpd3m9ml",
   "tg_tokens": null,
   "tps": 13.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-dpd3m9ml-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 84.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 32k 84.8 16.6 26.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dpd3m9ml",
   "source_url": "https://omlx.ai/benchmarks/performance/dpd3m9ml",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-824cw8jz-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 1mt52d7x; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 102.0 19.3 19.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/824cw8jz",
   "source_url": "https://omlx.ai/benchmarks/performance/824cw8jz",
   "tg_tokens": null,
   "tps": 19.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-824cw8jz-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: 1mt52d7x; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 93.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 93.8 17.9 21.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/824cw8jz",
   "source_url": "https://omlx.ai/benchmarks/performance/824cw8jz",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-e1uptlyo-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: ni320q0k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 935.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 935.5 33.7 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e1uptlyo",
   "source_url": "https://omlx.ai/benchmarks/performance/e1uptlyo",
   "tg_tokens": null,
   "tps": 33.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-e1uptlyo-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: ni320q0k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1024.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,024 33.3 17.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e1uptlyo",
   "source_url": "https://omlx.ai/benchmarks/performance/e1uptlyo",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-l3lziymb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 39.0 GB at this context; duplicate traces of the same session: k55fs6ls, u4chi0yv, uv3ubnfj; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 321.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 1k 321.1 56.2 39.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l3lziymb",
   "source_url": "https://omlx.ai/benchmarks/performance/l3lziymb",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-l3lziymb-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 42.3 GB at this context; duplicate traces of the same session: k55fs6ls, u4chi0yv, uv3ubnfj; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 420.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 4k 420.4 46.3 42.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l3lziymb",
   "source_url": "https://omlx.ai/benchmarks/performance/l3lziymb",
   "tg_tokens": null,
   "tps": 46.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-l3lziymb-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 42.9 GB at this context; duplicate traces of the same session: k55fs6ls, u4chi0yv, uv3ubnfj; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 440.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 8k 440.9 62.5 42.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l3lziymb",
   "source_url": "https://omlx.ai/benchmarks/performance/l3lziymb",
   "tg_tokens": null,
   "tps": 62.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-l3lziymb-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 44.1 GB at this context; duplicate traces of the same session: k55fs6ls, u4chi0yv, uv3ubnfj; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 420.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 8bit; 16k 420.9 61.5 44.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l3lziymb",
   "source_url": "https://omlx.ai/benchmarks/performance/l3lziymb",
   "tg_tokens": null,
   "tps": 61.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-h5abr816-ctx1024",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "MLX; oMLX community benchmark trace, single run at 1k context; peak memory 21.0 GB; TTFT 1785 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 573.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 573.8 PP tok/s 38.0 TG tok/s 1785 TTFT (ms) 21.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h5abr816",
   "source_url": "https://omlx.ai/benchmarks/performance/h5abr816",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-sgefqz8c-ctx131072",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 50.3 GB; TTFT 757233 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 173.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k tokens 173.1 PP tok/s 6.3 TG tok/s 757233 TTFT (ms) 50.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sgefqz8c",
   "source_url": "https://omlx.ai/benchmarks/performance/sgefqz8c",
   "tg_tokens": null,
   "tps": 6.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pqlqqj8j-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 0fg1wzfa, 13se8dnr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 675.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 675.3 39.1 17.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pqlqqj8j",
   "source_url": "https://omlx.ai/benchmarks/performance/pqlqqj8j",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pqlqqj8j-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: 0fg1wzfa, 13se8dnr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 826.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 826.5 63.8 18.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pqlqqj8j",
   "source_url": "https://omlx.ai/benchmarks/performance/pqlqqj8j",
   "tg_tokens": null,
   "tps": 63.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-pqlqqj8j-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 0fg1wzfa, 13se8dnr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 807.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 807.4 66.2 19.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pqlqqj8j",
   "source_url": "https://omlx.ai/benchmarks/performance/pqlqqj8j",
   "tg_tokens": null,
   "tps": 66.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-6dr9htic-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: sou6qol5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 127.8 28.7 18.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6dr9htic",
   "source_url": "https://omlx.ai/benchmarks/performance/6dr9htic",
   "tg_tokens": null,
   "tps": 28.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-6dr9htic-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: sou6qol5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 131.8 23.7 19.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6dr9htic",
   "source_url": "https://omlx.ai/benchmarks/performance/6dr9htic",
   "tg_tokens": null,
   "tps": 23.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-x1bl2gjp-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.5 GB at this context; duplicate traces of the same session: nj0ymdrp, yylrod3h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 905.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 905.1 73.6 17.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x1bl2gjp",
   "source_url": "https://omlx.ai/benchmarks/performance/x1bl2gjp",
   "tg_tokens": null,
   "tps": 73.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-x1bl2gjp-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: nj0ymdrp, yylrod3h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1013.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,013 73.0 18.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x1bl2gjp",
   "source_url": "https://omlx.ai/benchmarks/performance/x1bl2gjp",
   "tg_tokens": null,
   "tps": 73.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-x1bl2gjp-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.3 GB at this context; duplicate traces of the same session: nj0ymdrp, yylrod3h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 727.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 727.4 58.1 27.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x1bl2gjp",
   "source_url": "https://omlx.ai/benchmarks/performance/x1bl2gjp",
   "tg_tokens": null,
   "tps": 58.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9154jwx2-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: m4xl6l8t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 825.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k 825.6 32.8 17.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9154jwx2",
   "source_url": "https://omlx.ai/benchmarks/performance/9154jwx2",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9154jwx2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: m4xl6l8t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 913.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 913.8 40.4 18.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9154jwx2",
   "source_url": "https://omlx.ai/benchmarks/performance/9154jwx2",
   "tg_tokens": null,
   "tps": 40.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-it7vhoz9-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 73.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1238.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,238 81.0 73.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/it7vhoz9",
   "source_url": "https://omlx.ai/benchmarks/performance/it7vhoz9",
   "tg_tokens": null,
   "tps": 81.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-it7vhoz9-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 74.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2147.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,147 69.6 74.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/it7vhoz9",
   "source_url": "https://omlx.ai/benchmarks/performance/it7vhoz9",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-llgkipzb-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 704.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 704.6 78.9 70.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/llgkipzb",
   "source_url": "https://omlx.ai/benchmarks/performance/llgkipzb",
   "tg_tokens": null,
   "tps": 78.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-llgkipzb-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 927.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 927.6 73.0 71.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/llgkipzb",
   "source_url": "https://omlx.ai/benchmarks/performance/llgkipzb",
   "tg_tokens": null,
   "tps": 73.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-llgkipzb-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1068.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,068 74.6 71.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/llgkipzb",
   "source_url": "https://omlx.ai/benchmarks/performance/llgkipzb",
   "tg_tokens": null,
   "tps": 74.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-llgkipzb-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.4 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1062.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 1,062 76.1 72.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/llgkipzb",
   "source_url": "https://omlx.ai/benchmarks/performance/llgkipzb",
   "tg_tokens": null,
   "tps": 76.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-9wehnwet-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 8.9 GB; TTFT 11811 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 86.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 86.7 PP tok/s 5.8 TG tok/s 11811 TTFT (ms) 8.9 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9wehnwet",
   "source_url": "https://omlx.ai/benchmarks/performance/9wehnwet",
   "tg_tokens": null,
   "tps": 5.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-rmd0rd5h-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: 6242lk9z, puas1qas; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 287.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 287.6 50.5 22.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rmd0rd5h",
   "source_url": "https://omlx.ai/benchmarks/performance/rmd0rd5h",
   "tg_tokens": null,
   "tps": 50.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-rmd0rd5h-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: 6242lk9z, puas1qas; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 223.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 8k 223.4 40.9 23.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rmd0rd5h",
   "source_url": "https://omlx.ai/benchmarks/performance/rmd0rd5h",
   "tg_tokens": null,
   "tps": 40.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-rmd0rd5h-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: 6242lk9z, puas1qas; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 171.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 171.8 43.3 24.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rmd0rd5h",
   "source_url": "https://omlx.ai/benchmarks/performance/rmd0rd5h",
   "tg_tokens": null,
   "tps": 43.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-rmd0rd5h-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 6242lk9z, puas1qas; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 188.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 188.1 27.3 23.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rmd0rd5h",
   "source_url": "https://omlx.ai/benchmarks/performance/rmd0rd5h",
   "tg_tokens": null,
   "tps": 27.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-c8vt7iu7-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 296.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 296.3 43.8 22.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c8vt7iu7",
   "source_url": "https://omlx.ai/benchmarks/performance/c8vt7iu7",
   "tg_tokens": null,
   "tps": 43.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-c8vt7iu7-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 290.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 8k 290.6 44.4 23.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c8vt7iu7",
   "source_url": "https://omlx.ai/benchmarks/performance/c8vt7iu7",
   "tg_tokens": null,
   "tps": 44.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-c8vt7iu7-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 229.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 229.5 41.3 21.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c8vt7iu7",
   "source_url": "https://omlx.ai/benchmarks/performance/c8vt7iu7",
   "tg_tokens": null,
   "tps": 41.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-c8vt7iu7-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 180.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 180.1 26.0 23.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c8vt7iu7",
   "source_url": "https://omlx.ai/benchmarks/performance/c8vt7iu7",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-273qfdou-ctx32768",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 16.2 GB at this context; duplicate traces of the same session: uxov1hd0, w346j282, xvrmllb2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1901.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 1,901 39.2 16.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/273qfdou",
   "source_url": "https://omlx.ai/benchmarks/performance/273qfdou",
   "tg_tokens": null,
   "tps": 39.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-273qfdou-ctx65536",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: uxov1hd0, w346j282, xvrmllb2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1496.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 1,496 26.0 17.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/273qfdou",
   "source_url": "https://omlx.ai/benchmarks/performance/273qfdou",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-273qfdou-ctx131072",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: uxov1hd0, w346j282, xvrmllb2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1017.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 128k 1,017 15.3 20.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/273qfdou",
   "source_url": "https://omlx.ai/benchmarks/performance/273qfdou",
   "tg_tokens": null,
   "tps": 15.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-273qfdou-ctx199680",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: uxov1hd0, w346j282, xvrmllb2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 759.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 195k 759.4 10.8 24.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/273qfdou",
   "source_url": "https://omlx.ai/benchmarks/performance/273qfdou",
   "tg_tokens": null,
   "tps": 10.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-t3oc9wf1-ctx1024",
   "model": "Ternary-Bonsai-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 9.1 GB at this context; duplicate traces of the same session: 5uv14xyg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 872.7,
   "provenance": "community",
   "quant": "2bit",
   "quote": "prism-ml--Ternary-Bonsai-27B-mlx-2bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 unknown; 1k 872.7 58.5 9.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t3oc9wf1",
   "source_url": "https://omlx.ai/benchmarks/performance/t3oc9wf1",
   "tg_tokens": null,
   "tps": 58.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-t3oc9wf1-ctx4096",
   "model": "Ternary-Bonsai-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 10.1 GB at this context; duplicate traces of the same session: 5uv14xyg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 986.6,
   "provenance": "community",
   "quant": "2bit",
   "quote": "prism-ml--Ternary-Bonsai-27B-mlx-2bit on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 unknown; 4k 986.6 57.8 10.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t3oc9wf1",
   "source_url": "https://omlx.ai/benchmarks/performance/t3oc9wf1",
   "tg_tokens": null,
   "tps": 57.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-pjrtyvx0-ctx1024",
   "model": "GLM-5.3-Flash-Abliterated",
   "notes": "MLX; oMLX community benchmark trace; peak memory 177.7 GB at this context; duplicate traces of the same session: tn9wk8nr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 457.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-Abliterated-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 457.3 38.6 177.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pjrtyvx0",
   "source_url": "https://omlx.ai/benchmarks/performance/pjrtyvx0",
   "tg_tokens": null,
   "tps": 38.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-pjrtyvx0-ctx4096",
   "model": "GLM-5.3-Flash-Abliterated",
   "notes": "MLX; oMLX community benchmark trace; peak memory 178.9 GB at this context; duplicate traces of the same session: tn9wk8nr; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 569.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-Abliterated-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 569.1 35.6 178.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pjrtyvx0",
   "source_url": "https://omlx.ai/benchmarks/performance/pjrtyvx0",
   "tg_tokens": null,
   "tps": 35.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-g40aanr5-ctx1024",
   "model": "GLM-5.3-Flash-Abliterated",
   "notes": "MLX; oMLX community benchmark trace; peak memory 177.6 GB at this context; duplicate traces of the same session: i3n96r3f; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 417.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-Abliterated-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 417.5 33.6 177.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g40aanr5",
   "source_url": "https://omlx.ai/benchmarks/performance/g40aanr5",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-g40aanr5-ctx4096",
   "model": "GLM-5.3-Flash-Abliterated",
   "notes": "MLX; oMLX community benchmark trace; peak memory 178.9 GB at this context; duplicate traces of the same session: i3n96r3f; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 542.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-Abliterated-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 542.8 39.0 178.9 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g40aanr5",
   "source_url": "https://omlx.ai/benchmarks/performance/g40aanr5",
   "tg_tokens": null,
   "tps": 39.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-oo62kzdt-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: 1fks2j9u, j5wfty5p; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2013.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 8k 2,013 120.5 13.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oo62kzdt",
   "source_url": "https://omlx.ai/benchmarks/performance/oo62kzdt",
   "tg_tokens": null,
   "tps": 120.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-oo62kzdt-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.8 GB at this context; duplicate traces of the same session: 1fks2j9u, j5wfty5p; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1950.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 16k 1,950 115.9 13.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oo62kzdt",
   "source_url": "https://omlx.ai/benchmarks/performance/oo62kzdt",
   "tg_tokens": null,
   "tps": 115.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-oo62kzdt-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 14.1 GB at this context; duplicate traces of the same session: 1fks2j9u, j5wfty5p; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1774.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 32k 1,774 106.3 14.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oo62kzdt",
   "source_url": "https://omlx.ai/benchmarks/performance/oo62kzdt",
   "tg_tokens": null,
   "tps": 106.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-qioo5ihx-ctx16384",
   "model": "LFM2.5-8B-A1B",
   "notes": "LiquidAI MLX build; oMLX community benchmark trace, single run at 16k context; peak memory 5.6 GB; TTFT 21080 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 777.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LiquidAI--LFM2.5-8B-A1B-MLX-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k tokens 777.2 PP tok/s 87.8 TG tok/s 21080 TTFT (ms) 5.6 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qioo5ihx",
   "source_url": "https://omlx.ai/benchmarks/performance/qioo5ihx",
   "tg_tokens": null,
   "tps": 87.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1b3wjdgv-ctx4096",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 66.8 GB at this context; duplicate traces of the same session: 0mza65me; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2292.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,292 68.6 66.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1b3wjdgv",
   "source_url": "https://omlx.ai/benchmarks/performance/1b3wjdgv",
   "tg_tokens": null,
   "tps": 68.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1b3wjdgv-ctx16384",
   "model": "Qwen3.5-122B-A10B",
   "notes": "oMLX community benchmark trace; peak memory 67.9 GB at this context; duplicate traces of the same session: 0mza65me; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2134.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-122B-A10B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,134 63.5 67.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1b3wjdgv",
   "source_url": "https://omlx.ai/benchmarks/performance/1b3wjdgv",
   "tg_tokens": null,
   "tps": 63.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-42mvf2jz-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: mvtjl8n5, vmzs8uc5; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1106.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 4k 1,106 72.8 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/42mvf2jz",
   "source_url": "https://omlx.ai/benchmarks/performance/42mvf2jz",
   "tg_tokens": null,
   "tps": 72.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-42mvf2jz-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: mvtjl8n5, vmzs8uc5; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 876.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 16k 876.1 64.8 22.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/42mvf2jz",
   "source_url": "https://omlx.ai/benchmarks/performance/42mvf2jz",
   "tg_tokens": null,
   "tps": 64.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-42mvf2jz-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: mvtjl8n5, vmzs8uc5; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 298.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 128k 298.5 41.8 25.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/42mvf2jz",
   "source_url": "https://omlx.ai/benchmarks/performance/42mvf2jz",
   "tg_tokens": null,
   "tps": 41.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-l7rhuhsv-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 5bhpnasu, sqw6t7uc; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 804.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 1k 804.7 77.5 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l7rhuhsv",
   "source_url": "https://omlx.ai/benchmarks/performance/l7rhuhsv",
   "tg_tokens": null,
   "tps": 77.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-l7rhuhsv-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 5bhpnasu, sqw6t7uc; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1124.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 4k 1,124 77.0 21.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l7rhuhsv",
   "source_url": "https://omlx.ai/benchmarks/performance/l7rhuhsv",
   "tg_tokens": null,
   "tps": 77.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-l7rhuhsv-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 5bhpnasu, sqw6t7uc; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 681.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 32k 681.9 63.5 23.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l7rhuhsv",
   "source_url": "https://omlx.ai/benchmarks/performance/l7rhuhsv",
   "tg_tokens": null,
   "tps": 63.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-9wtb74js-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: 0wk22iao, dm9h8r8s; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 470.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 470.9 30.1 27.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9wtb74js",
   "source_url": "https://omlx.ai/benchmarks/performance/9wtb74js",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-9wtb74js-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.4 GB at this context; duplicate traces of the same session: 0wk22iao, dm9h8r8s; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 469.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 8k 469.1 31.9 28.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9wtb74js",
   "source_url": "https://omlx.ai/benchmarks/performance/9wtb74js",
   "tg_tokens": null,
   "tps": 31.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-9wtb74js-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: 0wk22iao, dm9h8r8s; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 428.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 16k 428.8 35.1 29.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9wtb74js",
   "source_url": "https://omlx.ai/benchmarks/performance/9wtb74js",
   "tg_tokens": null,
   "tps": 35.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-q8jmp3mg-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 53.1 GB; TTFT 47823 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 685.2,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 2bit; 32k tokens 685.2 PP tok/s 30.4 TG tok/s 47823 TTFT (ms) 53.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q8jmp3mg",
   "source_url": "https://omlx.ai/benchmarks/performance/q8jmp3mg",
   "tg_tokens": null,
   "tps": 30.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-29f7n02m-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 49.9 GB; TTFT 35562 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 921.4,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 2bit; 32k tokens 921.4 PP tok/s 31.3 TG tok/s 35562 TTFT (ms) 49.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/29f7n02m",
   "source_url": "https://omlx.ai/benchmarks/performance/29f7n02m",
   "tg_tokens": null,
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-fr7suq5i-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 45.4 GB; TTFT 65883 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 497.4,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 2bit; 32k tokens 497.4 PP tok/s 11.5 TG tok/s 65883 TTFT (ms) 45.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fr7suq5i",
   "source_url": "https://omlx.ai/benchmarks/performance/fr7suq5i",
   "tg_tokens": null,
   "tps": 11.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-ujjqsout-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 72.4 GB at this context; duplicate traces of the same session: qq7vs4a0; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 248.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 248.4 32.2 72.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ujjqsout",
   "source_url": "https://omlx.ai/benchmarks/performance/ujjqsout",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-ujjqsout-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 74.8 GB at this context; duplicate traces of the same session: qq7vs4a0; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 306.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 306.1 38.9 74.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ujjqsout",
   "source_url": "https://omlx.ai/benchmarks/performance/ujjqsout",
   "tg_tokens": null,
   "tps": 38.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-35uyhv0r-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 72.3 GB at this context; duplicate traces of the same session: lsxqmtw0; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 247.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 247.3 27.9 72.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/35uyhv0r",
   "source_url": "https://omlx.ai/benchmarks/performance/35uyhv0r",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-35uyhv0r-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 74.8 GB at this context; duplicate traces of the same session: lsxqmtw0; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 307.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 307.0 26.3 74.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/35uyhv0r",
   "source_url": "https://omlx.ai/benchmarks/performance/35uyhv0r",
   "tg_tokens": null,
   "tps": 26.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-y3xdlh2l-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace, single run at 1k context; peak memory 72.4 GB; TTFT 4418 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 231.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k tokens 231.8 PP tok/s 31.1 TG tok/s 4418 TTFT (ms) 72.4 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y3xdlh2l",
   "source_url": "https://omlx.ai/benchmarks/performance/y3xdlh2l",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-bonhuhz9-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 72.3 GB at this context; duplicate traces of the same session: ohyk199o; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 423.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 423.3 30.6 72.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bonhuhz9",
   "source_url": "https://omlx.ai/benchmarks/performance/bonhuhz9",
   "tg_tokens": null,
   "tps": 30.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-bonhuhz9-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 74.8 GB at this context; duplicate traces of the same session: ohyk199o; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 364.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 364.1 30.9 74.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bonhuhz9",
   "source_url": "https://omlx.ai/benchmarks/performance/bonhuhz9",
   "tg_tokens": null,
   "tps": 30.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-xygm6p33-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace, single run at 1k context; peak memory 72.5 GB; TTFT 4866 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 210.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k tokens 210.5 PP tok/s 30.5 TG tok/s 4866 TTFT (ms) 72.5 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xygm6p33",
   "source_url": "https://omlx.ai/benchmarks/performance/xygm6p33",
   "tg_tokens": null,
   "tps": 30.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-jr0jfa5i-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: alj8m2xx; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 190.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 190.1 24.3 70.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jr0jfa5i",
   "source_url": "https://omlx.ai/benchmarks/performance/jr0jfa5i",
   "tg_tokens": null,
   "tps": 24.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-jr0jfa5i-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: alj8m2xx; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 257.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 257.2 28.4 71.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jr0jfa5i",
   "source_url": "https://omlx.ai/benchmarks/performance/jr0jfa5i",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-vzqfhle6-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.8 GB; TTFT 4656 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 219.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k tokens 219.9 PP tok/s 34.6 TG tok/s 4656 TTFT (ms) 70.8 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vzqfhle6",
   "source_url": "https://omlx.ai/benchmarks/performance/vzqfhle6",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-q8imji60-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: ouzq39p4; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 211.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 211.7 37.3 70.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q8imji60",
   "source_url": "https://omlx.ai/benchmarks/performance/q8imji60",
   "tg_tokens": null,
   "tps": 37.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-q8imji60-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: ouzq39p4; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 267.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 267.0 47.2 71.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q8imji60",
   "source_url": "https://omlx.ai/benchmarks/performance/q8imji60",
   "tg_tokens": null,
   "tps": 47.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-cpebymz3-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: a4rz71pd; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 326.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 326.4 35.5 70.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cpebymz3",
   "source_url": "https://omlx.ai/benchmarks/performance/cpebymz3",
   "tg_tokens": null,
   "tps": 35.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-cpebymz3-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: a4rz71pd; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 331.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 331.5 30.4 71.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cpebymz3",
   "source_url": "https://omlx.ai/benchmarks/performance/cpebymz3",
   "tg_tokens": null,
   "tps": 30.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-s2f21nx3-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.8 GB; TTFT 3949 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 259.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k tokens 259.3 PP tok/s 36.7 TG tok/s 3949 TTFT (ms) 70.8 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s2f21nx3",
   "source_url": "https://omlx.ai/benchmarks/performance/s2f21nx3",
   "tg_tokens": null,
   "tps": 36.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-6vdtvm1y-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: e9dda1rc, r1i9n64n; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 704.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 704.6 78.9 70.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6vdtvm1y",
   "source_url": "https://omlx.ai/benchmarks/performance/6vdtvm1y",
   "tg_tokens": null,
   "tps": 78.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-6vdtvm1y-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: e9dda1rc, r1i9n64n; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 927.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 927.6 73.0 71.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6vdtvm1y",
   "source_url": "https://omlx.ai/benchmarks/performance/6vdtvm1y",
   "tg_tokens": null,
   "tps": 73.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-6vdtvm1y-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: e9dda1rc, r1i9n64n; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1068.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,068 74.6 71.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6vdtvm1y",
   "source_url": "https://omlx.ai/benchmarks/performance/6vdtvm1y",
   "tg_tokens": null,
   "tps": 74.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-6vdtvm1y-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.4 GB at this context; duplicate traces of the same session: e9dda1rc, r1i9n64n; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1062.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 1,062 76.1 72.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6vdtvm1y",
   "source_url": "https://omlx.ai/benchmarks/performance/6vdtvm1y",
   "tg_tokens": null,
   "tps": 76.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-9vjgi46l-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 36.1 GB; TTFT 55831 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 586.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k tokens 586.9 PP tok/s 13.8 TG tok/s 55831 TTFT (ms) 36.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9vjgi46l",
   "source_url": "https://omlx.ai/benchmarks/performance/9vjgi46l",
   "tg_tokens": null,
   "tps": 13.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-v2jle6hq-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 40.3 GB; TTFT 54629 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 599.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k tokens 599.8 PP tok/s 13.8 TG tok/s 54629 TTFT (ms) 40.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v2jle6hq",
   "source_url": "https://omlx.ai/benchmarks/performance/v2jle6hq",
   "tg_tokens": null,
   "tps": 13.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-df9xgzbd-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 34.2 GB; TTFT 54786 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 598.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k tokens 598.1 PP tok/s 15.0 TG tok/s 54786 TTFT (ms) 34.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/df9xgzbd",
   "source_url": "https://omlx.ai/benchmarks/performance/df9xgzbd",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-la5uzuu2-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 34.2 GB; TTFT 63584 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 515.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k tokens 515.3 PP tok/s 13.8 TG tok/s 63584 TTFT (ms) 34.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/la5uzuu2",
   "source_url": "https://omlx.ai/benchmarks/performance/la5uzuu2",
   "tg_tokens": null,
   "tps": 13.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-36hwcm3e-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: 2kf4v1da, q53dnavx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 296.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 296.3 43.8 22.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/36hwcm3e",
   "source_url": "https://omlx.ai/benchmarks/performance/36hwcm3e",
   "tg_tokens": null,
   "tps": 43.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-36hwcm3e-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: 2kf4v1da, q53dnavx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 290.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 8k 290.6 44.4 23.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/36hwcm3e",
   "source_url": "https://omlx.ai/benchmarks/performance/36hwcm3e",
   "tg_tokens": null,
   "tps": 44.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-36hwcm3e-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: 2kf4v1da, q53dnavx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 229.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 229.5 41.3 21.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/36hwcm3e",
   "source_url": "https://omlx.ai/benchmarks/performance/36hwcm3e",
   "tg_tokens": null,
   "tps": 41.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-36hwcm3e-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 2kf4v1da, q53dnavx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 180.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 180.1 26.0 23.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/36hwcm3e",
   "source_url": "https://omlx.ai/benchmarks/performance/36hwcm3e",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-d30ruqkv-ctx1024",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: m0sc8ew9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 487.5,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 1k 487.5 43.9 11.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d30ruqkv",
   "source_url": "https://omlx.ai/benchmarks/performance/d30ruqkv",
   "tg_tokens": null,
   "tps": 43.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-d30ruqkv-ctx4096",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: m0sc8ew9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 589.8,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 4k 589.8 39.9 14.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d30ruqkv",
   "source_url": "https://omlx.ai/benchmarks/performance/d30ruqkv",
   "tg_tokens": null,
   "tps": 39.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-443t9ryp-ctx4096",
   "model": "Underdog-27B-1.1",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 248.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Underdog-27B-1.1-MLX on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 248.5 8.2 17.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/443t9ryp",
   "source_url": "https://omlx.ai/benchmarks/performance/443t9ryp",
   "tg_tokens": null,
   "tps": 8.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-443t9ryp-ctx8192",
   "model": "Underdog-27B-1.1",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 238.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Underdog-27B-1.1-MLX on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 8k 238.2 7.9 17.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/443t9ryp",
   "source_url": "https://omlx.ai/benchmarks/performance/443t9ryp",
   "tg_tokens": null,
   "tps": 7.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-443t9ryp-ctx16384",
   "model": "Underdog-27B-1.1",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 230.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Underdog-27B-1.1-MLX on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 16k 230.9 7.5 17.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/443t9ryp",
   "source_url": "https://omlx.ai/benchmarks/performance/443t9ryp",
   "tg_tokens": null,
   "tps": 7.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-443t9ryp-ctx32768",
   "model": "Underdog-27B-1.1",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Underdog-27B-1.1-MLX on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 32k 220.1 6.6 18.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/443t9ryp",
   "source_url": "https://omlx.ai/benchmarks/performance/443t9ryp",
   "tg_tokens": null,
   "tps": 6.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-840xqixu-ctx32768",
   "model": "gemma-4-12B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 13.3 GB at this context; duplicate traces of the same session: 16bn68n3, 3knvjwjn, 8748u0px; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 825.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-qat-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 825.9 22.9 13.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/840xqixu",
   "source_url": "https://omlx.ai/benchmarks/performance/840xqixu",
   "tg_tokens": null,
   "tps": 22.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-840xqixu-ctx65536",
   "model": "gemma-4-12B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 14.9 GB at this context; duplicate traces of the same session: 16bn68n3, 3knvjwjn, 8748u0px; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 714.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-qat-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 714.0 21.9 14.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/840xqixu",
   "source_url": "https://omlx.ai/benchmarks/performance/840xqixu",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-840xqixu-ctx131072",
   "model": "gemma-4-12B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 16bn68n3, 3knvjwjn, 8748u0px; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 577.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-qat-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 128k 577.4 19.9 18.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/840xqixu",
   "source_url": "https://omlx.ai/benchmarks/performance/840xqixu",
   "tg_tokens": null,
   "tps": 19.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-840xqixu-ctx199680",
   "model": "gemma-4-12B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 16bn68n3, 3knvjwjn, 8748u0px; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 463.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-12B-it-qat-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 195k 463.9 18.5 22.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/840xqixu",
   "source_url": "https://omlx.ai/benchmarks/performance/840xqixu",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-nd93hgrh-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: x8i04439; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 830.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 830.3 75.4 16.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nd93hgrh",
   "source_url": "https://omlx.ai/benchmarks/performance/nd93hgrh",
   "tg_tokens": null,
   "tps": 75.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-nd93hgrh-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: x8i04439; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 723.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 723.8 56.8 16.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nd93hgrh",
   "source_url": "https://omlx.ai/benchmarks/performance/nd93hgrh",
   "tg_tokens": null,
   "tps": 56.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-k7bmpdp4-ctx1024",
   "model": "GLM-5.3-Flash-Abliterated",
   "notes": "MLX; oMLX community benchmark trace; peak memory 177.6 GB at this context; duplicate traces of the same session: s45vqd0b; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 510.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-Abliterated-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 510.7 52.0 177.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k7bmpdp4",
   "source_url": "https://omlx.ai/benchmarks/performance/k7bmpdp4",
   "tg_tokens": null,
   "tps": 52.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-k7bmpdp4-ctx4096",
   "model": "GLM-5.3-Flash-Abliterated",
   "notes": "MLX; oMLX community benchmark trace; peak memory 178.9 GB at this context; duplicate traces of the same session: s45vqd0b; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 582.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-Abliterated-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 582.4 48.0 178.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k7bmpdp4",
   "source_url": "https://omlx.ai/benchmarks/performance/k7bmpdp4",
   "tg_tokens": null,
   "tps": 48.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lesxb8r2-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 89.7 GB at this context; duplicate traces of the same session: gehna044, vqlphlhk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 97.7,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "GLM-5.3-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 1k 97.7 14.5 89.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lesxb8r2",
   "source_url": "https://omlx.ai/benchmarks/performance/lesxb8r2",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lesxb8r2-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 90.3 GB at this context; duplicate traces of the same session: gehna044, vqlphlhk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 286.2,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "GLM-5.3-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 8k 286.2 11.6 90.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lesxb8r2",
   "source_url": "https://omlx.ai/benchmarks/performance/lesxb8r2",
   "tg_tokens": null,
   "tps": 11.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lesxb8r2-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 90.9 GB at this context; duplicate traces of the same session: gehna044, vqlphlhk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 295.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "GLM-5.3-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 16k 295.0 16.9 90.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lesxb8r2",
   "source_url": "https://omlx.ai/benchmarks/performance/lesxb8r2",
   "tg_tokens": null,
   "tps": 16.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8nwp7rbu-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 88.5 GB at this context; duplicate traces of the same session: oom0n3hp, x5t0pawx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 100.5,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "GLM-5.3-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 1k 100.5 13.2 88.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8nwp7rbu",
   "source_url": "https://omlx.ai/benchmarks/performance/8nwp7rbu",
   "tg_tokens": null,
   "tps": 13.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8nwp7rbu-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 88.9 GB at this context; duplicate traces of the same session: oom0n3hp, x5t0pawx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 301.3,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "GLM-5.3-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 8k 301.3 11.1 88.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8nwp7rbu",
   "source_url": "https://omlx.ai/benchmarks/performance/8nwp7rbu",
   "tg_tokens": null,
   "tps": 11.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8nwp7rbu-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 89.6 GB at this context; duplicate traces of the same session: oom0n3hp, x5t0pawx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 297.6,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "GLM-5.3-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 16k 297.6 8.7 89.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8nwp7rbu",
   "source_url": "https://omlx.ai/benchmarks/performance/8nwp7rbu",
   "tg_tokens": null,
   "tps": 8.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-qsjsq34p-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: 8og9vg8q; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1277.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,277 85.9 24.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qsjsq34p",
   "source_url": "https://omlx.ai/benchmarks/performance/qsjsq34p",
   "tg_tokens": null,
   "tps": 85.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-qsjsq34p-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: 8og9vg8q; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1218.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 1,218 75.1 24.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qsjsq34p",
   "source_url": "https://omlx.ai/benchmarks/performance/qsjsq34p",
   "tg_tokens": null,
   "tps": 75.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-8nfeiygo-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: 7gvn597g; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1260.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,260 77.8 24.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8nfeiygo",
   "source_url": "https://omlx.ai/benchmarks/performance/8nfeiygo",
   "tg_tokens": null,
   "tps": 77.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-8nfeiygo-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: 7gvn597g; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1223.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 1,223 72.2 24.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8nfeiygo",
   "source_url": "https://omlx.ai/benchmarks/performance/8nfeiygo",
   "tg_tokens": null,
   "tps": 72.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-dqbhdqrb-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: avdfboze; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1243.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,243 86.7 24.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dqbhdqrb",
   "source_url": "https://omlx.ai/benchmarks/performance/dqbhdqrb",
   "tg_tokens": null,
   "tps": 86.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-dqbhdqrb-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: avdfboze; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1212.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 1,212 76.5 24.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dqbhdqrb",
   "source_url": "https://omlx.ai/benchmarks/performance/dqbhdqrb",
   "tg_tokens": null,
   "tps": 76.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-j31kot8z-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: jko0aq6q; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2019.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "LFM2.5-24B-A2B-oQ6e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 2,019 71.1 18.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j31kot8z",
   "source_url": "https://omlx.ai/benchmarks/performance/j31kot8z",
   "tg_tokens": null,
   "tps": 71.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-j31kot8z-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: jko0aq6q; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2076.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "LFM2.5-24B-A2B-oQ6e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 2,076 87.1 19.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j31kot8z",
   "source_url": "https://omlx.ai/benchmarks/performance/j31kot8z",
   "tg_tokens": null,
   "tps": 87.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-2b3jb24m-ctx131072",
   "model": "LFM2.5-8B-A1B",
   "notes": "LiquidAI MLX build; oMLX community benchmark trace, single run at 128k context; peak memory 6.9 GB; TTFT 283023 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 463.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LiquidAI--LFM2.5-8B-A1B-MLX-4bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 128k tokens 463.1 PP tok/s 32.9 TG tok/s 283023 TTFT (ms) 6.9 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2b3jb24m",
   "source_url": "https://omlx.ai/benchmarks/performance/2b3jb24m",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-eqwxjfhv-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "refusal-ablated MTP build; MLX; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: twpqopeb; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 952.2,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-refusal-ablated-MLX-oQ8-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 952.2 77.8 37.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eqwxjfhv",
   "source_url": "https://omlx.ai/benchmarks/performance/eqwxjfhv",
   "tg_tokens": null,
   "tps": 77.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-eqwxjfhv-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "refusal-ablated MTP build; MLX; oMLX community benchmark trace; peak memory 39.5 GB at this context; duplicate traces of the same session: twpqopeb; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 900.1,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-refusal-ablated-MLX-oQ8-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 900.1 64.1 39.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eqwxjfhv",
   "source_url": "https://omlx.ai/benchmarks/performance/eqwxjfhv",
   "tg_tokens": null,
   "tps": 64.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ey30puer-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "refusal-ablated MTP build; MLX; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: 2e98ybp2; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 946.1,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-refusal-ablated-MLX-oQ8-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 946.1 93.5 37.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ey30puer",
   "source_url": "https://omlx.ai/benchmarks/performance/ey30puer",
   "tg_tokens": null,
   "tps": 93.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ey30puer-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "refusal-ablated MTP build; MLX; oMLX community benchmark trace; peak memory 39.5 GB at this context; duplicate traces of the same session: 2e98ybp2; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 882.3,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-refusal-ablated-MLX-oQ8-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 882.3 62.3 39.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ey30puer",
   "source_url": "https://omlx.ai/benchmarks/performance/ey30puer",
   "tg_tokens": null,
   "tps": 62.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-iktxitxo-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 10.6 GB at this context; duplicate traces of the same session: lsypnwfg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1320.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B-8bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 8bit; 4k 1,320 31.7 10.6 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iktxitxo",
   "source_url": "https://omlx.ai/benchmarks/performance/iktxitxo",
   "tg_tokens": null,
   "tps": 31.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-iktxitxo-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.7 GB at this context; duplicate traces of the same session: lsypnwfg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1251.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B-8bit on M5 Pro (16c); M5 Pro (16c) \u00b7 24 GB \u00b7 8bit; 16k 1,251 30.1 11.7 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iktxitxo",
   "source_url": "https://omlx.ai/benchmarks/performance/iktxitxo",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-lxwdmisx-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 20.0 GB; TTFT 9750 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1681.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 16k tokens 1,681 PP tok/s 60.7 TG tok/s 9750 TTFT (ms) 20.0 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lxwdmisx",
   "source_url": "https://omlx.ai/benchmarks/performance/lxwdmisx",
   "tg_tokens": null,
   "tps": 60.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-70sedq86-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 2p8tei29, gxunkp3t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 826.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 826.3 43.2 20.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/70sedq86",
   "source_url": "https://omlx.ai/benchmarks/performance/70sedq86",
   "tg_tokens": null,
   "tps": 43.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-70sedq86-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 2p8tei29, gxunkp3t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 694.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 694.2 53.0 21.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/70sedq86",
   "source_url": "https://omlx.ai/benchmarks/performance/70sedq86",
   "tg_tokens": null,
   "tps": 53.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-70sedq86-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 2p8tei29, gxunkp3t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 641.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 641.5 58.5 21.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/70sedq86",
   "source_url": "https://omlx.ai/benchmarks/performance/70sedq86",
   "tg_tokens": null,
   "tps": 58.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mgu0gngz-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: iojwvt4x, lxjcc87h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 906.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 906.8 53.2 20.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mgu0gngz",
   "source_url": "https://omlx.ai/benchmarks/performance/mgu0gngz",
   "tg_tokens": null,
   "tps": 53.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mgu0gngz-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: iojwvt4x, lxjcc87h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 704.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 704.4 53.3 21.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mgu0gngz",
   "source_url": "https://omlx.ai/benchmarks/performance/mgu0gngz",
   "tg_tokens": null,
   "tps": 53.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mgu0gngz-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: iojwvt4x, lxjcc87h; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 644.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 644.1 63.4 21.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mgu0gngz",
   "source_url": "https://omlx.ai/benchmarks/performance/mgu0gngz",
   "tg_tokens": null,
   "tps": 63.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-1crubmcc-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: hlboihib; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 130.0 15.4 16.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1crubmcc",
   "source_url": "https://omlx.ai/benchmarks/performance/1crubmcc",
   "tg_tokens": null,
   "tps": 15.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-1crubmcc-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: hlboihib; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 115.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 115.5 14.4 19.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1crubmcc",
   "source_url": "https://omlx.ai/benchmarks/performance/1crubmcc",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-rsx62sr6-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: wp6y2sh2; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 132.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 132.7 15.8 16.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rsx62sr6",
   "source_url": "https://omlx.ai/benchmarks/performance/rsx62sr6",
   "tg_tokens": null,
   "tps": 15.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-rsx62sr6-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: wp6y2sh2; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 120.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 120.6 15.2 19.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rsx62sr6",
   "source_url": "https://omlx.ai/benchmarks/performance/rsx62sr6",
   "tg_tokens": null,
   "tps": 15.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-5cwr58tl-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.7 GB at this context; duplicate traces of the same session: 40a080b6; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 163.1,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 163.1 16.3 30.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5cwr58tl",
   "source_url": "https://omlx.ai/benchmarks/performance/5cwr58tl",
   "tg_tokens": null,
   "tps": 16.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-5cwr58tl-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.3 GB at this context; duplicate traces of the same session: 40a080b6; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 189.8,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 189.8 22.5 34.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5cwr58tl",
   "source_url": "https://omlx.ai/benchmarks/performance/5cwr58tl",
   "tg_tokens": null,
   "tps": 22.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-uc1z9qxp-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 51.5 GB at this context; duplicate traces of the same session: 3nqsgbiv, m48poopl, mhqpsbra; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1920.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 8k 1,920 55.6 51.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uc1z9qxp",
   "source_url": "https://omlx.ai/benchmarks/performance/uc1z9qxp",
   "tg_tokens": null,
   "tps": 55.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-uc1z9qxp-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 54.3 GB at this context; duplicate traces of the same session: 3nqsgbiv, m48poopl, mhqpsbra; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1700.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 64k 1,700 67.9 54.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uc1z9qxp",
   "source_url": "https://omlx.ai/benchmarks/performance/uc1z9qxp",
   "tg_tokens": null,
   "tps": 67.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-uc1z9qxp-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 58.2 GB at this context; duplicate traces of the same session: 3nqsgbiv, m48poopl, mhqpsbra; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1712.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 128k 1,712 69.5 58.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uc1z9qxp",
   "source_url": "https://omlx.ai/benchmarks/performance/uc1z9qxp",
   "tg_tokens": null,
   "tps": 69.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-uc1z9qxp-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 60.5 GB at this context; duplicate traces of the same session: 3nqsgbiv, m48poopl, mhqpsbra; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1452.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 195k 1,452 48.6 60.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uc1z9qxp",
   "source_url": "https://omlx.ai/benchmarks/performance/uc1z9qxp",
   "tg_tokens": null,
   "tps": 48.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-capazsj8-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 72.4 GB at this context; duplicate traces of the same session: ix6asnfu; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 234.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 234.5 27.9 72.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/capazsj8",
   "source_url": "https://omlx.ai/benchmarks/performance/capazsj8",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-capazsj8-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 74.8 GB at this context; duplicate traces of the same session: ix6asnfu; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 293.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 293.0 27.8 74.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/capazsj8",
   "source_url": "https://omlx.ai/benchmarks/performance/capazsj8",
   "tg_tokens": null,
   "tps": 27.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-4rrni2on-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 71.0 GB; TTFT 5306 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 193.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k tokens 193.0 PP tok/s 25.3 TG tok/s 5306 TTFT (ms) 71.0 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4rrni2on",
   "source_url": "https://omlx.ai/benchmarks/performance/4rrni2on",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-q4uq79fa-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 71.4 GB; TTFT 12695 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 322.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k tokens 322.6 PP tok/s 43.2 TG tok/s 12695 TTFT (ms) 71.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q4uq79fa",
   "source_url": "https://omlx.ai/benchmarks/performance/q4uq79fa",
   "tg_tokens": null,
   "tps": 43.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7624ger8-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; Swift 1.5 build; oMLX community benchmark trace; peak memory 89.3 GB at this context; duplicate traces of the same session: 2pbjkdnw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 164.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 164.0 32.0 89.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7624ger8",
   "source_url": "https://omlx.ai/benchmarks/performance/7624ger8",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7624ger8-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; Swift 1.5 build; oMLX community benchmark trace; peak memory 92.0 GB at this context; duplicate traces of the same session: 2pbjkdnw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 250.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 250.8 26.9 92.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7624ger8",
   "source_url": "https://omlx.ai/benchmarks/performance/7624ger8",
   "tg_tokens": null,
   "tps": 26.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-s8kox6tx-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; Swift 1.5 build; oMLX community benchmark trace, single run at 1k context; peak memory 92.5 GB; TTFT 6279 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 163.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 163.1 PP tok/s 34.2 TG tok/s 6279 TTFT (ms) 92.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s8kox6tx",
   "source_url": "https://omlx.ai/benchmarks/performance/s8kox6tx",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-x15kk1a4-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; Swift 1.5 build; oMLX community benchmark trace; peak memory 75.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 594.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 594.8 21.5 75.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x15kk1a4",
   "source_url": "https://omlx.ai/benchmarks/performance/x15kk1a4",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-x15kk1a4-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; Swift 1.5 build; oMLX community benchmark trace; peak memory 78.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 673.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 673.7 18.9 78.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x15kk1a4",
   "source_url": "https://omlx.ai/benchmarks/performance/x15kk1a4",
   "tg_tokens": null,
   "tps": 18.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-b3c7ti1i-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; Swift 1.5 build; oMLX community benchmark trace; peak memory 72.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 578.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 578.0 13.0 72.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b3c7ti1i",
   "source_url": "https://omlx.ai/benchmarks/performance/b3c7ti1i",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-b3c7ti1i-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; Swift 1.5 build; oMLX community benchmark trace; peak memory 76.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 786.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 786.9 13.1 76.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b3c7ti1i",
   "source_url": "https://omlx.ai/benchmarks/performance/b3c7ti1i",
   "tg_tokens": null,
   "tps": 13.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gzc07giw-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; Swift 1.5 build; oMLX community benchmark trace; peak memory 62.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 142.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 142.3 11.9 62.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gzc07giw",
   "source_url": "https://omlx.ai/benchmarks/performance/gzc07giw",
   "tg_tokens": null,
   "tps": 11.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gzc07giw-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; Swift 1.5 build; oMLX community benchmark trace; peak memory 67.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 159.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 159.7 11.8 67.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gzc07giw",
   "source_url": "https://omlx.ai/benchmarks/performance/gzc07giw",
   "tg_tokens": null,
   "tps": 11.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-s5hd2dsj-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; Swift 1.5 build; oMLX community benchmark trace, single run at 1k context; peak memory 74.3 GB; TTFT 1685 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 607.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 607.8 PP tok/s 19.2 TG tok/s 1685 TTFT (ms) 74.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s5hd2dsj",
   "source_url": "https://omlx.ai/benchmarks/performance/s5hd2dsj",
   "tg_tokens": null,
   "tps": 19.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-gah6upvg-ctx1024",
   "model": "Ternary-Bonsai-2-9B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 5.7 GB at this context; duplicate traces of the same session: 3sk6y2tf, i9ux3er2; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 449.1,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-9B-mlx-2bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 2bit; 1k 449.1 64.8 5.7 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gah6upvg",
   "source_url": "https://omlx.ai/benchmarks/performance/gah6upvg",
   "tg_tokens": null,
   "tps": 64.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-gah6upvg-ctx4096",
   "model": "Ternary-Bonsai-2-9B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: 3sk6y2tf, i9ux3er2; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 493.9,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-9B-mlx-2bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 2bit; 4k 493.9 56.0 6.7 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gah6upvg",
   "source_url": "https://omlx.ai/benchmarks/performance/gah6upvg",
   "tg_tokens": null,
   "tps": 56.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-gah6upvg-ctx8192",
   "model": "Ternary-Bonsai-2-9B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 7.4 GB at this context; duplicate traces of the same session: 3sk6y2tf, i9ux3er2; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 439.7,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-9B-mlx-2bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 2bit; 8k 439.7 45.1 7.4 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gah6upvg",
   "source_url": "https://omlx.ai/benchmarks/performance/gah6upvg",
   "tg_tokens": null,
   "tps": 45.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-50kuy5zh-ctx4096",
   "model": "Underdog-27B-1.1",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 0td7n2j7, jcwyr1zd; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 248.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Underdog-27B-1.1-MLX on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 248.5 8.2 17.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/50kuy5zh",
   "source_url": "https://omlx.ai/benchmarks/performance/50kuy5zh",
   "tg_tokens": null,
   "tps": 8.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-50kuy5zh-ctx8192",
   "model": "Underdog-27B-1.1",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 0td7n2j7, jcwyr1zd; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 238.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Underdog-27B-1.1-MLX on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 8k 238.2 7.9 17.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/50kuy5zh",
   "source_url": "https://omlx.ai/benchmarks/performance/50kuy5zh",
   "tg_tokens": null,
   "tps": 7.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-50kuy5zh-ctx16384",
   "model": "Underdog-27B-1.1",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 0td7n2j7, jcwyr1zd; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 230.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Underdog-27B-1.1-MLX on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 16k 230.9 7.5 17.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/50kuy5zh",
   "source_url": "https://omlx.ai/benchmarks/performance/50kuy5zh",
   "tg_tokens": null,
   "tps": 7.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-50kuy5zh-ctx32768",
   "model": "Underdog-27B-1.1",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 0td7n2j7, jcwyr1zd; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Underdog-27B-1.1-MLX on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 32k 220.1 6.6 18.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/50kuy5zh",
   "source_url": "https://omlx.ai/benchmarks/performance/50kuy5zh",
   "tg_tokens": null,
   "tps": 6.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ragpdz2l-ctx1024",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: tg7qb0ll; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 102.8 15.9 20.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ragpdz2l",
   "source_url": "https://omlx.ai/benchmarks/performance/ragpdz2l",
   "tg_tokens": null,
   "tps": 15.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ragpdz2l-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: tg7qb0ll; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 91.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 91.9 17.0 21.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ragpdz2l",
   "source_url": "https://omlx.ai/benchmarks/performance/ragpdz2l",
   "tg_tokens": null,
   "tps": 17.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-b8f3mf0m-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 696.2,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "mlx-community--Ornith-1.5-35B-A3B-OptiQ-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 696.2 57.1 23.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b8f3mf0m",
   "source_url": "https://omlx.ai/benchmarks/performance/b8f3mf0m",
   "tg_tokens": null,
   "tps": 57.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-b8f3mf0m-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 580.1,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "mlx-community--Ornith-1.5-35B-A3B-OptiQ-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 580.1 50.0 23.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b8f3mf0m",
   "source_url": "https://omlx.ai/benchmarks/performance/b8f3mf0m",
   "tg_tokens": null,
   "tps": 50.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bimshlkj-ctx1024",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 102.7 GB at this context; duplicate traces of the same session: xuvqjrhh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 371.5,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4-Flash-0731-oQ2e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 1k 371.5 33.8 102.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bimshlkj",
   "source_url": "https://omlx.ai/benchmarks/performance/bimshlkj",
   "tg_tokens": null,
   "tps": 33.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bimshlkj-ctx4096",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 89.7 GB at this context; duplicate traces of the same session: xuvqjrhh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 712.5,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4-Flash-0731-oQ2e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 4k 712.5 29.4 89.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bimshlkj",
   "source_url": "https://omlx.ai/benchmarks/performance/bimshlkj",
   "tg_tokens": null,
   "tps": 29.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-zpidm3vf-ctx4096",
   "model": "DeepSeek-V4.1-Flash-Abliterated",
   "notes": "MLX build; oMLX community benchmark trace, single run at 4k context; peak memory 313.9 GB; TTFT 10145 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 403.8,
   "provenance": "community",
   "quant": "unknown",
   "quote": "DeepSeek-V4.1-Flash-Abliterated-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k tokens 403.8 PP tok/s 21.5 TG tok/s 10145 TTFT (ms) 313.9 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zpidm3vf",
   "source_url": "https://omlx.ai/benchmarks/performance/zpidm3vf",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ide9yy68-ctx4096",
   "model": "DeepSeek-V4.1-Flash-Abliterated",
   "notes": "MLX build; oMLX community benchmark trace, single run at 4k context; peak memory 313.9 GB; TTFT 5916 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 692.4,
   "provenance": "community",
   "quant": "unknown",
   "quote": "DeepSeek-V4.1-Flash-Abliterated-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k tokens 692.4 PP tok/s 21.3 TG tok/s 5916 TTFT (ms) 313.9 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ide9yy68",
   "source_url": "https://omlx.ai/benchmarks/performance/ide9yy68",
   "tg_tokens": null,
   "tps": 21.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-h4ai9lz9-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 88.6 GB at this context; duplicate traces of the same session: 44bx2y3e, ktz0yrea; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 100.2,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "GLM-5.3-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 1k 100.2 14.5 88.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h4ai9lz9",
   "source_url": "https://omlx.ai/benchmarks/performance/h4ai9lz9",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-h4ai9lz9-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 89.3 GB at this context; duplicate traces of the same session: 44bx2y3e, ktz0yrea; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 297.5,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "GLM-5.3-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 8k 297.5 12.9 89.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h4ai9lz9",
   "source_url": "https://omlx.ai/benchmarks/performance/h4ai9lz9",
   "tg_tokens": null,
   "tps": 12.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-h4ai9lz9-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 89.8 GB at this context; duplicate traces of the same session: 44bx2y3e, ktz0yrea; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 282.9,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "GLM-5.3-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 16k 282.9 18.1 89.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h4ai9lz9",
   "source_url": "https://omlx.ai/benchmarks/performance/h4ai9lz9",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-gph01cjr-ctx16384",
   "model": "Hermes-4.3-36B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: 03m829bd, dj4h0dli; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1044.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Hermes-4.3-36B-mlx-5Bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 5bit; 16k 1,044 27.8 27.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gph01cjr",
   "source_url": "https://omlx.ai/benchmarks/performance/gph01cjr",
   "tg_tokens": null,
   "tps": 27.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-gph01cjr-ctx65536",
   "model": "Hermes-4.3-36B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 39.8 GB at this context; duplicate traces of the same session: 03m829bd, dj4h0dli; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 635.4,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Hermes-4.3-36B-mlx-5Bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 5bit; 64k 635.4 17.9 39.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gph01cjr",
   "source_url": "https://omlx.ai/benchmarks/performance/gph01cjr",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-gph01cjr-ctx131072",
   "model": "Hermes-4.3-36B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 55.8 GB at this context; duplicate traces of the same session: 03m829bd, dj4h0dli; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 415.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Hermes-4.3-36B-mlx-5Bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 5bit; 128k 415.0 12.7 55.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gph01cjr",
   "source_url": "https://omlx.ai/benchmarks/performance/gph01cjr",
   "tg_tokens": null,
   "tps": 12.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-g2k1pbwp-ctx4096",
   "model": "Hermes-4.3-36B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1167.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Hermes-4.3-36B-mlx-5Bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 5bit; 4k 1,167 31.8 25.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g2k1pbwp",
   "source_url": "https://omlx.ai/benchmarks/performance/g2k1pbwp",
   "tg_tokens": null,
   "tps": 31.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-g2k1pbwp-ctx16384",
   "model": "Hermes-4.3-36B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1032.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Hermes-4.3-36B-mlx-5Bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 5bit; 16k 1,032 27.8 27.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g2k1pbwp",
   "source_url": "https://omlx.ai/benchmarks/performance/g2k1pbwp",
   "tg_tokens": null,
   "tps": 27.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-zhcaim17-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.5 GB at this context; duplicate traces of the same session: 2i8tuyfz, c0pfo33o, nwxer39g; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2299.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 2,299 74.2 13.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zhcaim17",
   "source_url": "https://omlx.ai/benchmarks/performance/zhcaim17",
   "tg_tokens": null,
   "tps": 74.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-zhcaim17-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.6 GB at this context; duplicate traces of the same session: 2i8tuyfz, c0pfo33o, nwxer39g; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2335.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 8k 2,335 70.9 13.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zhcaim17",
   "source_url": "https://omlx.ai/benchmarks/performance/zhcaim17",
   "tg_tokens": null,
   "tps": 70.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-zhcaim17-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.8 GB at this context; duplicate traces of the same session: 2i8tuyfz, c0pfo33o, nwxer39g; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2231.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 16k 2,231 60.2 13.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zhcaim17",
   "source_url": "https://omlx.ai/benchmarks/performance/zhcaim17",
   "tg_tokens": null,
   "tps": 60.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-zhcaim17-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 14.1 GB at this context; duplicate traces of the same session: 2i8tuyfz, c0pfo33o, nwxer39g; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2004.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 32k 2,004 56.0 14.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zhcaim17",
   "source_url": "https://omlx.ai/benchmarks/performance/zhcaim17",
   "tg_tokens": null,
   "tps": 56.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-s9ofjjp3-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.7 GB at this context; duplicate traces of the same session: jul20w2w; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 227.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 227.0 20.7 5.7 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s9ofjjp3",
   "source_url": "https://omlx.ai/benchmarks/performance/s9ofjjp3",
   "tg_tokens": null,
   "tps": 20.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-s9ofjjp3-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: jul20w2w; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 235.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 235.8 20.2 6.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s9ofjjp3",
   "source_url": "https://omlx.ai/benchmarks/performance/s9ofjjp3",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1gksn9ml-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 3vkis9a2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 490.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 490.3 28.0 16.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1gksn9ml",
   "source_url": "https://omlx.ai/benchmarks/performance/1gksn9ml",
   "tg_tokens": null,
   "tps": 28.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1gksn9ml-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 3vkis9a2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 650.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 650.5 29.4 17.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1gksn9ml",
   "source_url": "https://omlx.ai/benchmarks/performance/1gksn9ml",
   "tg_tokens": null,
   "tps": 29.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-t9h3r4c3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: k7fjsarl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 912.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 912.1 52.4 20.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t9h3r4c3",
   "source_url": "https://omlx.ai/benchmarks/performance/t9h3r4c3",
   "tg_tokens": null,
   "tps": 52.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-t9h3r4c3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: k7fjsarl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 742.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 742.6 55.4 21.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t9h3r4c3",
   "source_url": "https://omlx.ai/benchmarks/performance/t9h3r4c3",
   "tg_tokens": null,
   "tps": 55.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-o6fudhen-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 128.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 128.7 8.0 32.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o6fudhen",
   "source_url": "https://omlx.ai/benchmarks/performance/o6fudhen",
   "tg_tokens": null,
   "tps": 8.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-o6fudhen-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.7 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 111.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 16k 111.2 8.2 33.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o6fudhen",
   "source_url": "https://omlx.ai/benchmarks/performance/o6fudhen",
   "tg_tokens": null,
   "tps": 8.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-aafnysrk-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: u1udjkj7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 876.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 876.5 32.2 33.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aafnysrk",
   "source_url": "https://omlx.ai/benchmarks/performance/aafnysrk",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-aafnysrk-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: u1udjkj7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 954.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 954.6 30.4 34.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aafnysrk",
   "source_url": "https://omlx.ai/benchmarks/performance/aafnysrk",
   "tg_tokens": null,
   "tps": 30.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-w1cy6s9k-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: h1l4ul0r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 836.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 836.5 19.0 28.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w1cy6s9k",
   "source_url": "https://omlx.ai/benchmarks/performance/w1cy6s9k",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-w1cy6s9k-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: h1l4ul0r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 948.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 948.8 18.9 29.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w1cy6s9k",
   "source_url": "https://omlx.ai/benchmarks/performance/w1cy6s9k",
   "tg_tokens": null,
   "tps": 18.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lxrbbkbq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 033e1qbp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 839.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 839.2 19.0 28.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lxrbbkbq",
   "source_url": "https://omlx.ai/benchmarks/performance/lxrbbkbq",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lxrbbkbq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: 033e1qbp; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 953.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 953.3 18.9 29.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lxrbbkbq",
   "source_url": "https://omlx.ai/benchmarks/performance/lxrbbkbq",
   "tg_tokens": null,
   "tps": 18.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tjr6qou9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: h2n50n90; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 841.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 841.3 19.0 28.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tjr6qou9",
   "source_url": "https://omlx.ai/benchmarks/performance/tjr6qou9",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tjr6qou9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: h2n50n90; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 952.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 952.4 18.2 29.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tjr6qou9",
   "source_url": "https://omlx.ai/benchmarks/performance/tjr6qou9",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hn5e0bhc-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: kahb6tow; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 911.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 911.2 71.8 17.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hn5e0bhc",
   "source_url": "https://omlx.ai/benchmarks/performance/hn5e0bhc",
   "tg_tokens": null,
   "tps": 71.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hn5e0bhc-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: kahb6tow; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 996.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 996.9 62.5 18.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hn5e0bhc",
   "source_url": "https://omlx.ai/benchmarks/performance/hn5e0bhc",
   "tg_tokens": null,
   "tps": 62.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3vs3nrdw-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: mi0z3qt2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 914.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 914.4 75.4 17.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3vs3nrdw",
   "source_url": "https://omlx.ai/benchmarks/performance/3vs3nrdw",
   "tg_tokens": null,
   "tps": 75.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3vs3nrdw-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: mi0z3qt2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 994.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 994.7 74.1 18.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3vs3nrdw",
   "source_url": "https://omlx.ai/benchmarks/performance/3vs3nrdw",
   "tg_tokens": null,
   "tps": 74.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fyxiss2d-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: ikn2gy2e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 789.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 789.5 54.8 23.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fyxiss2d",
   "source_url": "https://omlx.ai/benchmarks/performance/fyxiss2d",
   "tg_tokens": null,
   "tps": 54.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fyxiss2d-ctx4096",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 24.6 GB at this context; duplicate traces of the same session: ikn2gy2e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 970.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 970.1 48.5 24.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fyxiss2d",
   "source_url": "https://omlx.ai/benchmarks/performance/fyxiss2d",
   "tg_tokens": null,
   "tps": 48.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-i0avcj09-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: 3z4z4wgs; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1586.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 4bit; 4k 1,586 53.4 17.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i0avcj09",
   "source_url": "https://omlx.ai/benchmarks/performance/i0avcj09",
   "tg_tokens": null,
   "tps": 53.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-i0avcj09-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 3z4z4wgs; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1804.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 4bit; 16k 1,804 50.4 17.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i0avcj09",
   "source_url": "https://omlx.ai/benchmarks/performance/i0avcj09",
   "tg_tokens": null,
   "tps": 50.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-1ozcd3md-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 25.8 GB; TTFT 54174 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 302.4,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k tokens 302.4 PP tok/s 10.3 TG tok/s 54174 TTFT (ms) 25.8 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ozcd3md",
   "source_url": "https://omlx.ai/benchmarks/performance/1ozcd3md",
   "tg_tokens": null,
   "tps": 10.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bfol8l32-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: npaf1c87; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 764.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 764.4 43.8 17.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bfol8l32",
   "source_url": "https://omlx.ai/benchmarks/performance/bfol8l32",
   "tg_tokens": null,
   "tps": 43.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bfol8l32-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: npaf1c87; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 996.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 996.8 59.0 18.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bfol8l32",
   "source_url": "https://omlx.ai/benchmarks/performance/bfol8l32",
   "tg_tokens": null,
   "tps": 59.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3lzpoce6-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: miip2evy, oi4si5yg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 564.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 564.2 30.9 19.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3lzpoce6",
   "source_url": "https://omlx.ai/benchmarks/performance/3lzpoce6",
   "tg_tokens": null,
   "tps": 30.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3lzpoce6-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: miip2evy, oi4si5yg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 573.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 16k 573.1 27.2 19.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3lzpoce6",
   "source_url": "https://omlx.ai/benchmarks/performance/3lzpoce6",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-3lzpoce6-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: miip2evy, oi4si5yg; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 430.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 64k 430.7 16.0 23.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3lzpoce6",
   "source_url": "https://omlx.ai/benchmarks/performance/3lzpoce6",
   "tg_tokens": null,
   "tps": 16.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-t4wb1b64-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: k4hwijsb, omndzbl9; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 378.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 8bit; 4k 378.3 21.3 34.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t4wb1b64",
   "source_url": "https://omlx.ai/benchmarks/performance/t4wb1b64",
   "tg_tokens": null,
   "tps": 21.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-t4wb1b64-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.7 GB at this context; duplicate traces of the same session: k4hwijsb, omndzbl9; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 358.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 8bit; 16k 358.7 21.5 35.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t4wb1b64",
   "source_url": "https://omlx.ai/benchmarks/performance/t4wb1b64",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-t4wb1b64-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.1 GB at this context; duplicate traces of the same session: k4hwijsb, omndzbl9; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 303.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 8bit; 64k 303.7 12.0 39.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t4wb1b64",
   "source_url": "https://omlx.ai/benchmarks/performance/t4wb1b64",
   "tg_tokens": null,
   "tps": 12.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-3eyhcztt-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.7 GB at this context; duplicate traces of the same session: eydr4awu; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 123.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 123.9 11.0 37.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3eyhcztt",
   "source_url": "https://omlx.ai/benchmarks/performance/3eyhcztt",
   "tg_tokens": null,
   "tps": 11.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-3eyhcztt-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 41.2 GB at this context; duplicate traces of the same session: eydr4awu; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 305.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 4bit; 64k 305.9 9.7 41.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3eyhcztt",
   "source_url": "https://omlx.ai/benchmarks/performance/3eyhcztt",
   "tg_tokens": null,
   "tps": 9.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-108zg2sb-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; Swift 1.5 build; oMLX community benchmark trace; peak memory 62.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 142.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 142.3 11.9 62.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/108zg2sb",
   "source_url": "https://omlx.ai/benchmarks/performance/108zg2sb",
   "tg_tokens": null,
   "tps": 11.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-108zg2sb-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; Swift 1.5 build; oMLX community benchmark trace; peak memory 67.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 159.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 159.7 11.8 67.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/108zg2sb",
   "source_url": "https://omlx.ai/benchmarks/performance/108zg2sb",
   "tg_tokens": null,
   "tps": 11.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ufkuahg6-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; Swift 1.5 build; oMLX community benchmark trace; peak memory 72.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 578.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 578.0 13.0 72.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ufkuahg6",
   "source_url": "https://omlx.ai/benchmarks/performance/ufkuahg6",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ufkuahg6-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; Swift 1.5 build; oMLX community benchmark trace; peak memory 76.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 786.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 786.9 13.1 76.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ufkuahg6",
   "source_url": "https://omlx.ai/benchmarks/performance/ufkuahg6",
   "tg_tokens": null,
   "tps": 13.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-06ul4qfd-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; Swift 1.5 build; oMLX community benchmark trace; peak memory 75.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 594.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 594.8 21.5 75.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/06ul4qfd",
   "source_url": "https://omlx.ai/benchmarks/performance/06ul4qfd",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-06ul4qfd-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; Swift 1.5 build; oMLX community benchmark trace; peak memory 78.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 673.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 673.7 18.9 78.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/06ul4qfd",
   "source_url": "https://omlx.ai/benchmarks/performance/06ul4qfd",
   "tg_tokens": null,
   "tps": 18.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5baq5710-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; Swift 1.5 build; oMLX community benchmark trace; peak memory 91.1 GB at this context; duplicate traces of the same session: 8loqus3s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 158.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 158.0 21.1 91.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5baq5710",
   "source_url": "https://omlx.ai/benchmarks/performance/5baq5710",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5baq5710-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; Swift 1.5 build; oMLX community benchmark trace; peak memory 93.7 GB at this context; duplicate traces of the same session: 8loqus3s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 325.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 325.4 19.2 93.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5baq5710",
   "source_url": "https://omlx.ai/benchmarks/performance/5baq5710",
   "tg_tokens": null,
   "tps": 19.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-lx5822lm-ctx1024",
   "model": "Llama 3.3 70B",
   "notes": "oMLX community benchmark trace; peak memory 37.8 GB at this context; duplicate traces of the same session: awhg6uol; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 114.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Llama-3.3-70B-Instruct-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 1k 114.0 15.3 37.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lx5822lm",
   "source_url": "https://omlx.ai/benchmarks/performance/lx5822lm",
   "tg_tokens": null,
   "tps": 15.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-lx5822lm-ctx4096",
   "model": "Llama 3.3 70B",
   "notes": "oMLX community benchmark trace; peak memory 39.0 GB at this context; duplicate traces of the same session: awhg6uol; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 113.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Llama-3.3-70B-Instruct-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 4k 113.4 14.7 39.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lx5822lm",
   "source_url": "https://omlx.ai/benchmarks/performance/lx5822lm",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-dsb2eczy-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 696.2,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "mlx-community--Ornith-1.5-35B-A3B-OptiQ-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 696.2 57.1 23.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dsb2eczy",
   "source_url": "https://omlx.ai/benchmarks/performance/dsb2eczy",
   "tg_tokens": null,
   "tps": 57.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-dsb2eczy-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 580.1,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "mlx-community--Ornith-1.5-35B-A3B-OptiQ-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k 580.1 50.0 23.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dsb2eczy",
   "source_url": "https://omlx.ai/benchmarks/performance/dsb2eczy",
   "tg_tokens": null,
   "tps": 50.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-7h2c1mkt-ctx4096",
   "model": "Hermes-4.3-36B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1167.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Hermes-4.3-36B-mlx-5Bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 5bit; 4k 1,167 31.8 25.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7h2c1mkt",
   "source_url": "https://omlx.ai/benchmarks/performance/7h2c1mkt",
   "tg_tokens": null,
   "tps": 31.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-7h2c1mkt-ctx16384",
   "model": "Hermes-4.3-36B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1032.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Hermes-4.3-36B-mlx-5Bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 5bit; 16k 1,032 27.8 27.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7h2c1mkt",
   "source_url": "https://omlx.ai/benchmarks/performance/7h2c1mkt",
   "tg_tokens": null,
   "tps": 27.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-q8ig0r7e-ctx1024",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: uatlzlxa; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 571.2,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-35B-A3B-6bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 1k 571.2 56.9 27.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q8ig0r7e",
   "source_url": "https://omlx.ai/benchmarks/performance/q8ig0r7e",
   "tg_tokens": null,
   "tps": 56.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-q8ig0r7e-ctx4096",
   "model": "Qwen3.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: uatlzlxa; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 656.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.5-35B-A3B-6bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 4k 656.5 55.4 29.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q8ig0r7e",
   "source_url": "https://omlx.ai/benchmarks/performance/q8ig0r7e",
   "tg_tokens": null,
   "tps": 55.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-lxhh5cdd-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 9542pxn5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 815.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 815.1 94.3 19.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lxhh5cdd",
   "source_url": "https://omlx.ai/benchmarks/performance/lxhh5cdd",
   "tg_tokens": null,
   "tps": 94.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-lxhh5cdd-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 9542pxn5; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 842.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 842.4 91.7 20.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lxhh5cdd",
   "source_url": "https://omlx.ai/benchmarks/performance/lxhh5cdd",
   "tg_tokens": null,
   "tps": 91.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-7e8bxp2t-ctx1024",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1301.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 1k 1,301 73.8 12.7 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7e8bxp2t",
   "source_url": "https://omlx.ai/benchmarks/performance/7e8bxp2t",
   "tg_tokens": null,
   "tps": 73.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-7e8bxp2t-ctx4096",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1382.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 4k 1,382 72.7 13.4 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7e8bxp2t",
   "source_url": "https://omlx.ai/benchmarks/performance/7e8bxp2t",
   "tg_tokens": null,
   "tps": 72.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-7e8bxp2t-ctx8192",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.5 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1311.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 8k 1,311 70.7 13.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7e8bxp2t",
   "source_url": "https://omlx.ai/benchmarks/performance/7e8bxp2t",
   "tg_tokens": null,
   "tps": 70.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-9f74yow0-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 0owcwut1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 849.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 849.7 108.4 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9f74yow0",
   "source_url": "https://omlx.ai/benchmarks/performance/9f74yow0",
   "tg_tokens": null,
   "tps": 108.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-9f74yow0-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 0owcwut1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2193.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,193 106.2 20.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9f74yow0",
   "source_url": "https://omlx.ai/benchmarks/performance/9f74yow0",
   "tg_tokens": null,
   "tps": 106.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-j6yfr13z-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 9o14vvjv, awlsk7qc, nouyjxza, q31f9pwm, quc3o4cl, z9i1yzw7; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 923.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 923.5 102.5 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j6yfr13z",
   "source_url": "https://omlx.ai/benchmarks/performance/j6yfr13z",
   "tg_tokens": null,
   "tps": 102.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-j6yfr13z-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: 9o14vvjv, awlsk7qc, nouyjxza, q31f9pwm, quc3o4cl, z9i1yzw7; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1107.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 1,107 111.3 24.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j6yfr13z",
   "source_url": "https://omlx.ai/benchmarks/performance/j6yfr13z",
   "tg_tokens": null,
   "tps": 111.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-j6yfr13z-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: 9o14vvjv, awlsk7qc, nouyjxza, q31f9pwm, quc3o4cl, z9i1yzw7; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1034.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 1,034 55.8 24.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j6yfr13z",
   "source_url": "https://omlx.ai/benchmarks/performance/j6yfr13z",
   "tg_tokens": null,
   "tps": 55.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-j6yfr13z-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: 9o14vvjv, awlsk7qc, nouyjxza, q31f9pwm, quc3o4cl, z9i1yzw7; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 907.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 907.5 62.2 25.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j6yfr13z",
   "source_url": "https://omlx.ai/benchmarks/performance/j6yfr13z",
   "tg_tokens": null,
   "tps": 62.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-j6yfr13z-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: 9o14vvjv, awlsk7qc, nouyjxza, q31f9pwm, quc3o4cl, z9i1yzw7; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 729.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 32k 729.1 61.3 26.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j6yfr13z",
   "source_url": "https://omlx.ai/benchmarks/performance/j6yfr13z",
   "tg_tokens": null,
   "tps": 61.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-j6yfr13z-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: 9o14vvjv, awlsk7qc, nouyjxza, q31f9pwm, quc3o4cl, z9i1yzw7; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 528.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 64k 528.1 65.9 27.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j6yfr13z",
   "source_url": "https://omlx.ai/benchmarks/performance/j6yfr13z",
   "tg_tokens": null,
   "tps": 65.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-j6yfr13z-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.1 GB at this context; duplicate traces of the same session: 9o14vvjv, awlsk7qc, nouyjxza, q31f9pwm, quc3o4cl, z9i1yzw7; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 341.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 128k 341.7 43.6 29.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j6yfr13z",
   "source_url": "https://omlx.ai/benchmarks/performance/j6yfr13z",
   "tg_tokens": null,
   "tps": 43.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-iz30kj7m-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 128.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 128.7 8.0 32.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iz30kj7m",
   "source_url": "https://omlx.ai/benchmarks/performance/iz30kj7m",
   "tg_tokens": null,
   "tps": 8.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-iz30kj7m-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.7 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 111.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 16k 111.2 8.2 33.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iz30kj7m",
   "source_url": "https://omlx.ai/benchmarks/performance/iz30kj7m",
   "tg_tokens": null,
   "tps": 8.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jrgarezv-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: qg43x7pc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 837.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 837.5 26.1 33.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jrgarezv",
   "source_url": "https://omlx.ai/benchmarks/performance/jrgarezv",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jrgarezv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: qg43x7pc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 954.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 954.7 29.1 34.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jrgarezv",
   "source_url": "https://omlx.ai/benchmarks/performance/jrgarezv",
   "tg_tokens": null,
   "tps": 29.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-u27dfbcz-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: qmbri9p1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 876.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 876.6 34.6 33.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u27dfbcz",
   "source_url": "https://omlx.ai/benchmarks/performance/u27dfbcz",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-u27dfbcz-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.8 GB at this context; duplicate traces of the same session: qmbri9p1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 953.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 953.5 28.3 34.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u27dfbcz",
   "source_url": "https://omlx.ai/benchmarks/performance/u27dfbcz",
   "tg_tokens": null,
   "tps": 28.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-n05tsnli-ctx16384",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: 6lxp9zmp, qzn5a8se; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1141.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-8bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 8bit; 16k 1,141 25.8 35.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n05tsnli",
   "source_url": "https://omlx.ai/benchmarks/performance/n05tsnli",
   "tg_tokens": null,
   "tps": 25.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-n05tsnli-ctx65536",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "oMLX community benchmark trace; peak memory 41.0 GB at this context; duplicate traces of the same session: 6lxp9zmp, qzn5a8se; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 814.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-8bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 8bit; 64k 814.0 21.9 41.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n05tsnli",
   "source_url": "https://omlx.ai/benchmarks/performance/n05tsnli",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-n05tsnli-ctx131072",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "oMLX community benchmark trace; peak memory 49.0 GB at this context; duplicate traces of the same session: 6lxp9zmp, qzn5a8se; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 577.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-8bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 8bit; 128k 577.5 18.4 49.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n05tsnli",
   "source_url": "https://omlx.ai/benchmarks/performance/n05tsnli",
   "tg_tokens": null,
   "tps": 18.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1g7wz12w-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 75.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 868.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 868.2 42.4 75.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1g7wz12w",
   "source_url": "https://omlx.ai/benchmarks/performance/1g7wz12w",
   "tg_tokens": null,
   "tps": 42.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1g7wz12w-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 75.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1236.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,236 42.7 75.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1g7wz12w",
   "source_url": "https://omlx.ai/benchmarks/performance/1g7wz12w",
   "tg_tokens": null,
   "tps": 42.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-rtipr3vd-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 73.6 GB at this context; duplicate traces of the same session: 8kn4r5do, n8umegx1, vp0bgnp7; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 233.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 1k 233.6 42.5 73.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rtipr3vd",
   "source_url": "https://omlx.ai/benchmarks/performance/rtipr3vd",
   "tg_tokens": null,
   "tps": 42.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-rtipr3vd-ctx8192",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 74.6 GB at this context; duplicate traces of the same session: 8kn4r5do, n8umegx1, vp0bgnp7; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 611.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 8k 611.8 33.0 74.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rtipr3vd",
   "source_url": "https://omlx.ai/benchmarks/performance/rtipr3vd",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-rtipr3vd-ctx16384",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 74.9 GB at this context; duplicate traces of the same session: 8kn4r5do, n8umegx1, vp0bgnp7; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 616.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 16k 616.0 34.5 74.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rtipr3vd",
   "source_url": "https://omlx.ai/benchmarks/performance/rtipr3vd",
   "tg_tokens": null,
   "tps": 34.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-rtipr3vd-ctx32768",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 75.5 GB at this context; duplicate traces of the same session: 8kn4r5do, n8umegx1, vp0bgnp7; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 613.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 32k 613.2 39.2 75.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rtipr3vd",
   "source_url": "https://omlx.ai/benchmarks/performance/rtipr3vd",
   "tg_tokens": null,
   "tps": 39.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-a634targ-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 52.4 GB at this context; duplicate traces of the same session: cmj9x74y; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 154.9,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 2bit; 1k 154.9 20.5 52.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a634targ",
   "source_url": "https://omlx.ai/benchmarks/performance/a634targ",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-a634targ-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 54.2 GB at this context; duplicate traces of the same session: cmj9x74y; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 238.7,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 2bit; 32k 238.7 15.5 54.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a634targ",
   "source_url": "https://omlx.ai/benchmarks/performance/a634targ",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-tnjrqdib-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.8 GB at this context; duplicate traces of the same session: gaqr0q77; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.5,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 2bit; 1k 104.5 14.9 30.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tnjrqdib",
   "source_url": "https://omlx.ai/benchmarks/performance/tnjrqdib",
   "tg_tokens": null,
   "tps": 14.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-tnjrqdib-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.7 GB at this context; duplicate traces of the same session: gaqr0q77; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 259.5,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 2bit; 4k 259.5 14.7 31.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tnjrqdib",
   "source_url": "https://omlx.ai/benchmarks/performance/tnjrqdib",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-zl2pvrek-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.8 GB at this context; duplicate traces of the same session: dod16kwg; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 97.9,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 2bit; 1k 97.9 14.0 30.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zl2pvrek",
   "source_url": "https://omlx.ai/benchmarks/performance/zl2pvrek",
   "tg_tokens": null,
   "tps": 14.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-zl2pvrek-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.7 GB at this context; duplicate traces of the same session: dod16kwg; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 250.4,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 2bit; 4k 250.4 11.0 31.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zl2pvrek",
   "source_url": "https://omlx.ai/benchmarks/performance/zl2pvrek",
   "tg_tokens": null,
   "tps": 11.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-dgb7kx73-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6 mixed-quant MTP v2 build; oMLX community benchmark trace; peak memory 77.5 GB at this context; duplicate traces of the same session: i40kmh9c, ignzk6ki, l4o3lxnb, ra2j38u7; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 279.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp-v2 on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 6bit; 1k 279.2 42.1 77.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dgb7kx73",
   "source_url": "https://omlx.ai/benchmarks/performance/dgb7kx73",
   "tg_tokens": null,
   "tps": 42.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-dgb7kx73-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6 mixed-quant MTP v2 build; oMLX community benchmark trace; peak memory 78.1 GB at this context; duplicate traces of the same session: i40kmh9c, ignzk6ki, l4o3lxnb, ra2j38u7; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 348.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp-v2 on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 6bit; 4k 348.0 62.2 78.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dgb7kx73",
   "source_url": "https://omlx.ai/benchmarks/performance/dgb7kx73",
   "tg_tokens": null,
   "tps": 62.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-dgb7kx73-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6 mixed-quant MTP v2 build; oMLX community benchmark trace; peak memory 78.5 GB at this context; duplicate traces of the same session: i40kmh9c, ignzk6ki, l4o3lxnb, ra2j38u7; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 606.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp-v2 on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 6bit; 8k 606.7 36.9 78.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dgb7kx73",
   "source_url": "https://omlx.ai/benchmarks/performance/dgb7kx73",
   "tg_tokens": null,
   "tps": 36.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-dgb7kx73-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6 mixed-quant MTP v2 build; oMLX community benchmark trace; peak memory 78.8 GB at this context; duplicate traces of the same session: i40kmh9c, ignzk6ki, l4o3lxnb, ra2j38u7; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 607.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp-v2 on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 6bit; 16k 607.6 59.0 78.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dgb7kx73",
   "source_url": "https://omlx.ai/benchmarks/performance/dgb7kx73",
   "tg_tokens": null,
   "tps": 59.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-dgb7kx73-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6 mixed-quant MTP v2 build; oMLX community benchmark trace; peak memory 79.4 GB at this context; duplicate traces of the same session: i40kmh9c, ignzk6ki, l4o3lxnb, ra2j38u7; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 604.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp-v2 on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 6bit; 32k 604.7 38.5 79.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dgb7kx73",
   "source_url": "https://omlx.ai/benchmarks/performance/dgb7kx73",
   "tg_tokens": null,
   "tps": 38.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-aa0bgb5b-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.4 GB at this context; duplicate traces of the same session: aqq90h2o, m1eg9zsx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1367.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,367 71.4 101.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aa0bgb5b",
   "source_url": "https://omlx.ai/benchmarks/performance/aa0bgb5b",
   "tg_tokens": null,
   "tps": 71.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-aa0bgb5b-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: aqq90h2o, m1eg9zsx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1392.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 1,392 96.9 101.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aa0bgb5b",
   "source_url": "https://omlx.ai/benchmarks/performance/aa0bgb5b",
   "tg_tokens": null,
   "tps": 96.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-aa0bgb5b-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: aqq90h2o, m1eg9zsx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1384.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 1,384 101.3 101.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aa0bgb5b",
   "source_url": "https://omlx.ai/benchmarks/performance/aa0bgb5b",
   "tg_tokens": null,
   "tps": 101.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-aa0bgb5b-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.2 GB at this context; duplicate traces of the same session: aqq90h2o, m1eg9zsx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1373.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 1,373 86.0 102.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aa0bgb5b",
   "source_url": "https://omlx.ai/benchmarks/performance/aa0bgb5b",
   "tg_tokens": null,
   "tps": 86.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-aa0bgb5b-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.4 GB at this context; duplicate traces of the same session: aqq90h2o, m1eg9zsx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1350.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 1,350 85.1 104.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aa0bgb5b",
   "source_url": "https://omlx.ai/benchmarks/performance/aa0bgb5b",
   "tg_tokens": null,
   "tps": 85.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-aa0bgb5b-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.1 GB at this context; duplicate traces of the same session: aqq90h2o, m1eg9zsx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1326.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 1,326 65.9 109.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aa0bgb5b",
   "source_url": "https://omlx.ai/benchmarks/performance/aa0bgb5b",
   "tg_tokens": null,
   "tps": 65.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-aa0bgb5b-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.1 GB at this context; duplicate traces of the same session: aqq90h2o, m1eg9zsx; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1309.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 195k 1,309 58.8 110.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aa0bgb5b",
   "source_url": "https://omlx.ai/benchmarks/performance/aa0bgb5b",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-tan5cmv4-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 42.1 GB at this context; duplicate traces of the same session: 9lnv9bmx, c03r2xtl, jdx5sm1s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 95.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 95.5 13.5 42.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tan5cmv4",
   "source_url": "https://omlx.ai/benchmarks/performance/tan5cmv4",
   "tg_tokens": null,
   "tps": 13.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-tan5cmv4-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 42.9 GB at this context; duplicate traces of the same session: 9lnv9bmx, c03r2xtl, jdx5sm1s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 351.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 351.7 16.1 42.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tan5cmv4",
   "source_url": "https://omlx.ai/benchmarks/performance/tan5cmv4",
   "tg_tokens": null,
   "tps": 16.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-tan5cmv4-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 44.1 GB at this context; duplicate traces of the same session: 9lnv9bmx, c03r2xtl, jdx5sm1s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 311.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 311.4 11.9 44.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tan5cmv4",
   "source_url": "https://omlx.ai/benchmarks/performance/tan5cmv4",
   "tg_tokens": null,
   "tps": 11.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-tan5cmv4-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 45.1 GB at this context; duplicate traces of the same session: 9lnv9bmx, c03r2xtl, jdx5sm1s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 336.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 336.4 15.7 45.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tan5cmv4",
   "source_url": "https://omlx.ai/benchmarks/performance/tan5cmv4",
   "tg_tokens": null,
   "tps": 15.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-315pr1ka-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 42.3 GB; TTFT 11172 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 91.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 91.7 PP tok/s 16.0 TG tok/s 11172 TTFT (ms) 42.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/315pr1ka",
   "source_url": "https://omlx.ai/benchmarks/performance/315pr1ka",
   "tg_tokens": null,
   "tps": 16.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dw3fxe65-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; Swift 1.5 build; oMLX community benchmark trace; peak memory 88.0 GB at this context; duplicate traces of the same session: skcvm30g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 159.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 159.7 22.8 88.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dw3fxe65",
   "source_url": "https://omlx.ai/benchmarks/performance/dw3fxe65",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dw3fxe65-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; Swift 1.5 build; oMLX community benchmark trace; peak memory 90.6 GB at this context; duplicate traces of the same session: skcvm30g; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 288.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 288.3 19.6 90.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dw3fxe65",
   "source_url": "https://omlx.ai/benchmarks/performance/dw3fxe65",
   "tg_tokens": null,
   "tps": 19.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6u5jpkdl-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; Swift 1.5 build; oMLX community benchmark trace, single run at 1k context; peak memory 86.2 GB; TTFT 6579 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 155.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 155.6 PP tok/s 25.9 TG tok/s 6579 TTFT (ms) 86.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6u5jpkdl",
   "source_url": "https://omlx.ai/benchmarks/performance/6u5jpkdl",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4nr7czxc-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; Swift 1.5 build; oMLX community benchmark trace; peak memory 92.4 GB at this context; duplicate traces of the same session: 16o8q2lv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 158.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 158.4 31.9 92.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4nr7czxc",
   "source_url": "https://omlx.ai/benchmarks/performance/4nr7czxc",
   "tg_tokens": null,
   "tps": 31.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4nr7czxc-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; Swift 1.5 build; oMLX community benchmark trace; peak memory 95.1 GB at this context; duplicate traces of the same session: 16o8q2lv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 350.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 350.4 27.6 95.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4nr7czxc",
   "source_url": "https://omlx.ai/benchmarks/performance/4nr7czxc",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ayp3t8af-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: 6b9t4o3i; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 632.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 632.8 55.2 16.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ayp3t8af",
   "source_url": "https://omlx.ai/benchmarks/performance/ayp3t8af",
   "tg_tokens": null,
   "tps": 55.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ayp3t8af-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 6b9t4o3i; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 650.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 650.4 53.5 16.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ayp3t8af",
   "source_url": "https://omlx.ai/benchmarks/performance/ayp3t8af",
   "tg_tokens": null,
   "tps": 53.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-t3d35bf3-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.1 GB at this context; duplicate traces of the same session: zjv9xpvo; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 679.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 679.1 57.3 14.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t3d35bf3",
   "source_url": "https://omlx.ai/benchmarks/performance/t3d35bf3",
   "tg_tokens": null,
   "tps": 57.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-t3d35bf3-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: zjv9xpvo; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 654.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 654.4 55.9 14.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t3d35bf3",
   "source_url": "https://omlx.ai/benchmarks/performance/t3d35bf3",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-90125528-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: nuwi8ypb, p49q6idi, zmry5yqi; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 768.2,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "mlx-community--Ornith-1.5-35B-A3B-OptiQ-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 768.2 67.6 21.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/90125528",
   "source_url": "https://omlx.ai/benchmarks/performance/90125528",
   "tg_tokens": null,
   "tps": 67.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-90125528-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: nuwi8ypb, p49q6idi, zmry5yqi; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 836.3,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "mlx-community--Ornith-1.5-35B-A3B-OptiQ-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 836.3 65.3 22.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/90125528",
   "source_url": "https://omlx.ai/benchmarks/performance/90125528",
   "tg_tokens": null,
   "tps": 65.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-90125528-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: nuwi8ypb, p49q6idi, zmry5yqi; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 791.1,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "mlx-community--Ornith-1.5-35B-A3B-OptiQ-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k 791.1 63.0 22.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/90125528",
   "source_url": "https://omlx.ai/benchmarks/performance/90125528",
   "tg_tokens": null,
   "tps": 63.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-90125528-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: nuwi8ypb, p49q6idi, zmry5yqi; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 702.4,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "mlx-community--Ornith-1.5-35B-A3B-OptiQ-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 702.4 56.2 22.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/90125528",
   "source_url": "https://omlx.ai/benchmarks/performance/90125528",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-j9pgoisc-ctx1024",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "OptiQ REAP-19B build; oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: 5tyolexj; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1064.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM-V-4.7-35B-A3B-OptiQ-4bit-REAP-19B on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 1k 1,064 74.2 14.8 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j9pgoisc",
   "source_url": "https://omlx.ai/benchmarks/performance/j9pgoisc",
   "tg_tokens": null,
   "tps": 74.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-j9pgoisc-ctx4096",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "OptiQ REAP-19B build; oMLX community benchmark trace; peak memory 15.3 GB at this context; duplicate traces of the same session: 5tyolexj; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1352.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM-V-4.7-35B-A3B-OptiQ-4bit-REAP-19B on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 4k 1,352 73.1 15.3 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j9pgoisc",
   "source_url": "https://omlx.ai/benchmarks/performance/j9pgoisc",
   "tg_tokens": null,
   "tps": 73.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-7vn1xv7p-ctx1024",
   "model": "Nex-N2.5-mini",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: 9i8lob57; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2143.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Nex-N2.5-mini-MLX-oQ8 on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 8bit; 1k 2,143 70.7 35.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7vn1xv7p",
   "source_url": "https://omlx.ai/benchmarks/performance/7vn1xv7p",
   "tg_tokens": null,
   "tps": 70.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-7vn1xv7p-ctx4096",
   "model": "Nex-N2.5-mini",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: 9i8lob57; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2527.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Nex-N2.5-mini-MLX-oQ8 on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 8bit; 4k 2,527 68.7 36.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7vn1xv7p",
   "source_url": "https://omlx.ai/benchmarks/performance/7vn1xv7p",
   "tg_tokens": null,
   "tps": 68.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-eka49d0e-ctx1024",
   "model": "Qwen3.5-4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 3.5 GB at this context; duplicate traces of the same session: 8jb1647u, m9pk2ajm; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 651.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-4B-oQ4e-mtp on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 1k 651.3 97.8 3.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eka49d0e",
   "source_url": "https://omlx.ai/benchmarks/performance/eka49d0e",
   "tg_tokens": null,
   "tps": 97.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-eka49d0e-ctx4096",
   "model": "Qwen3.5-4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.0 GB at this context; duplicate traces of the same session: 8jb1647u, m9pk2ajm; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 682.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-4B-oQ4e-mtp on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 4k 682.6 94.0 4.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eka49d0e",
   "source_url": "https://omlx.ai/benchmarks/performance/eka49d0e",
   "tg_tokens": null,
   "tps": 94.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-eka49d0e-ctx8192",
   "model": "Qwen3.5-4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: 8jb1647u, m9pk2ajm; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 636.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-4B-oQ4e-mtp on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 8k 636.9 90.1 4.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eka49d0e",
   "source_url": "https://omlx.ai/benchmarks/performance/eka49d0e",
   "tg_tokens": null,
   "tps": 90.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-9qn3zit6-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; MLX build; oMLX community benchmark trace, single run at 8k context; peak memory 31.4 GB; TTFT 11136 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 735.6,
   "provenance": "community",
   "quant": "6bit-XL",
   "quote": "Qwen3.6-35B-A3B-6bit-XL-MTP-mlx on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 8k tokens 735.6 PP tok/s 49.5 TG tok/s 11136 TTFT (ms) 31.4 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9qn3zit6",
   "source_url": "https://omlx.ai/benchmarks/performance/9qn3zit6",
   "tg_tokens": null,
   "tps": 49.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-3bwn53xq-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; MLX build; oMLX community benchmark trace, single run at 8k context; peak memory 31.4 GB; TTFT 11145 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 735.1,
   "provenance": "community",
   "quant": "6bit-XL",
   "quote": "Qwen3.6-35B-A3B-6bit-XL-MTP-mlx on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 8k tokens 735.1 PP tok/s 57.8 TG tok/s 11145 TTFT (ms) 31.4 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3bwn53xq",
   "source_url": "https://omlx.ai/benchmarks/performance/3bwn53xq",
   "tg_tokens": null,
   "tps": 57.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-9e4bi73d-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 8k context; peak memory 29.4 GB; TTFT 10743 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 762.6,
   "provenance": "community",
   "quant": "6bit-XL",
   "quote": "Qwen3.6-35B-A3B-6bit-XL-mlx on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 8k tokens 762.6 PP tok/s 53.1 TG tok/s 10743 TTFT (ms) 29.4 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9e4bi73d",
   "source_url": "https://omlx.ai/benchmarks/performance/9e4bi73d",
   "tg_tokens": null,
   "tps": 53.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-n711mjh7-ctx1024",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: 951cq8vq; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1301.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 1k 1,301 73.8 12.7 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n711mjh7",
   "source_url": "https://omlx.ai/benchmarks/performance/n711mjh7",
   "tg_tokens": null,
   "tps": 73.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-n711mjh7-ctx4096",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: 951cq8vq; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1382.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 4k 1,382 72.7 13.4 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n711mjh7",
   "source_url": "https://omlx.ai/benchmarks/performance/n711mjh7",
   "tg_tokens": null,
   "tps": 72.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-n711mjh7-ctx8192",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.5 GB at this context; duplicate traces of the same session: 951cq8vq; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1311.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 8k 1,311 70.7 13.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n711mjh7",
   "source_url": "https://omlx.ai/benchmarks/performance/n711mjh7",
   "tg_tokens": null,
   "tps": 70.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-70xj0889-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: 0sjh0x0x, 9b1uw2yo, gotzx2vb, jzbyod2s; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 992.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 992.8 68.4 24.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/70xj0889",
   "source_url": "https://omlx.ai/benchmarks/performance/70xj0889",
   "tg_tokens": null,
   "tps": 68.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-70xj0889-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: 0sjh0x0x, 9b1uw2yo, gotzx2vb, jzbyod2s; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 873.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 873.1 63.1 25.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/70xj0889",
   "source_url": "https://omlx.ai/benchmarks/performance/70xj0889",
   "tg_tokens": null,
   "tps": 63.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-70xj0889-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.8 GB at this context; duplicate traces of the same session: 0sjh0x0x, 9b1uw2yo, gotzx2vb, jzbyod2s; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 701.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 32k 701.1 59.6 26.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/70xj0889",
   "source_url": "https://omlx.ai/benchmarks/performance/70xj0889",
   "tg_tokens": null,
   "tps": 59.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-70xj0889-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: 0sjh0x0x, 9b1uw2yo, gotzx2vb, jzbyod2s; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 506.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 64k 506.6 49.6 27.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/70xj0889",
   "source_url": "https://omlx.ai/benchmarks/performance/70xj0889",
   "tg_tokens": null,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-70xj0889-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16 weights; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.1 GB at this context; duplicate traces of the same session: 0sjh0x0x, 9b1uw2yo, gotzx2vb, jzbyod2s; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 327.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 128k 327.6 38.6 29.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/70xj0889",
   "source_url": "https://omlx.ai/benchmarks/performance/70xj0889",
   "tg_tokens": null,
   "tps": 38.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2b06lryr-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 1bvu3hre; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 839.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 839.1 19.1 28.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2b06lryr",
   "source_url": "https://omlx.ai/benchmarks/performance/2b06lryr",
   "tg_tokens": null,
   "tps": 19.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2b06lryr-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: 1bvu3hre; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 631.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 128k 631.8 14.4 37.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2b06lryr",
   "source_url": "https://omlx.ai/benchmarks/performance/2b06lryr",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gby73neg-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: jjqafgrr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 794.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 794.0 19.1 28.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gby73neg",
   "source_url": "https://omlx.ai/benchmarks/performance/gby73neg",
   "tg_tokens": null,
   "tps": 19.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gby73neg-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: jjqafgrr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 953.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 953.5 19.0 29.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gby73neg",
   "source_url": "https://omlx.ai/benchmarks/performance/gby73neg",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-uzrguqaw-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 68000spq, mqt8gqy8, rm3rj4jr, x6hwlh3p; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1186.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-27B-oQ6 on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 1,186 39.8 22.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uzrguqaw",
   "source_url": "https://omlx.ai/benchmarks/performance/uzrguqaw",
   "tg_tokens": null,
   "tps": 39.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-uzrguqaw-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: 68000spq, mqt8gqy8, rm3rj4jr, x6hwlh3p; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1563.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-27B-oQ6 on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 1,563 39.5 23.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uzrguqaw",
   "source_url": "https://omlx.ai/benchmarks/performance/uzrguqaw",
   "tg_tokens": null,
   "tps": 39.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-uzrguqaw-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: 68000spq, mqt8gqy8, rm3rj4jr, x6hwlh3p; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1741.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-27B-oQ6 on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 8k 1,741 39.1 24.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uzrguqaw",
   "source_url": "https://omlx.ai/benchmarks/performance/uzrguqaw",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-uzrguqaw-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: 68000spq, mqt8gqy8, rm3rj4jr, x6hwlh3p; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1695.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-27B-oQ6 on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 16k 1,695 38.0 24.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uzrguqaw",
   "source_url": "https://omlx.ai/benchmarks/performance/uzrguqaw",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-uzrguqaw-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: 68000spq, mqt8gqy8, rm3rj4jr, x6hwlh3p; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1601.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-27B-oQ6 on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 32k 1,601 36.6 25.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uzrguqaw",
   "source_url": "https://omlx.ai/benchmarks/performance/uzrguqaw",
   "tg_tokens": null,
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-uzrguqaw-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: 68000spq, mqt8gqy8, rm3rj4jr, x6hwlh3p; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1428.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-27B-oQ6 on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 64k 1,428 33.6 27.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uzrguqaw",
   "source_url": "https://omlx.ai/benchmarks/performance/uzrguqaw",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-02se8ns2-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: terpds3n; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 196.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 196.6 28.9 34.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/02se8ns2",
   "source_url": "https://omlx.ai/benchmarks/performance/02se8ns2",
   "tg_tokens": null,
   "tps": 28.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-02se8ns2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 38.5 GB at this context; duplicate traces of the same session: terpds3n; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 201.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 201.5 21.6 38.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/02se8ns2",
   "source_url": "https://omlx.ai/benchmarks/performance/02se8ns2",
   "tg_tokens": null,
   "tps": 21.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0ggvto3x-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 75.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 868.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 868.2 42.4 75.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ggvto3x",
   "source_url": "https://omlx.ai/benchmarks/performance/0ggvto3x",
   "tg_tokens": null,
   "tps": 42.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0ggvto3x-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 75.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1236.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,236 42.7 75.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ggvto3x",
   "source_url": "https://omlx.ai/benchmarks/performance/0ggvto3x",
   "tg_tokens": null,
   "tps": 42.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-fpzq025i-ctx1024",
   "model": "Qwen3.8-Flash-Next-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 78.1 GB at this context; duplicate traces of the same session: 6zr36p9q, cpfzfbtu, zza3j4gk; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 321.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-Abliterated-MLX-4bit on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 1k 321.1 29.7 78.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fpzq025i",
   "source_url": "https://omlx.ai/benchmarks/performance/fpzq025i",
   "tg_tokens": null,
   "tps": 29.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-fpzq025i-ctx8192",
   "model": "Qwen3.8-Flash-Next-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 78.9 GB at this context; duplicate traces of the same session: 6zr36p9q, cpfzfbtu, zza3j4gk; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 610.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-Abliterated-MLX-4bit on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 8k 610.2 36.0 78.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fpzq025i",
   "source_url": "https://omlx.ai/benchmarks/performance/fpzq025i",
   "tg_tokens": null,
   "tps": 36.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-fpzq025i-ctx16384",
   "model": "Qwen3.8-Flash-Next-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 79.2 GB at this context; duplicate traces of the same session: 6zr36p9q, cpfzfbtu, zza3j4gk; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 611.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-Abliterated-MLX-4bit on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 16k 611.8 37.5 79.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fpzq025i",
   "source_url": "https://omlx.ai/benchmarks/performance/fpzq025i",
   "tg_tokens": null,
   "tps": 37.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-fpzq025i-ctx32768",
   "model": "Qwen3.8-Flash-Next-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 79.8 GB at this context; duplicate traces of the same session: 6zr36p9q, cpfzfbtu, zza3j4gk; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 609.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-Abliterated-MLX-4bit on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 32k 609.2 44.2 79.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fpzq025i",
   "source_url": "https://omlx.ai/benchmarks/performance/fpzq025i",
   "tg_tokens": null,
   "tps": 44.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-y116akof-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: d27ueavn; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 260.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 260.8 31.0 70.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y116akof",
   "source_url": "https://omlx.ai/benchmarks/performance/y116akof",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-y116akof-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: d27ueavn; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 266.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 266.2 33.9 71.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y116akof",
   "source_url": "https://omlx.ai/benchmarks/performance/y116akof",
   "tg_tokens": null,
   "tps": 33.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-kp3shcsz-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.4 GB at this context; duplicate traces of the same session: 9351romx, 9k75zg12, ltyh2wal; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1367.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,367 71.4 101.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kp3shcsz",
   "source_url": "https://omlx.ai/benchmarks/performance/kp3shcsz",
   "tg_tokens": null,
   "tps": 71.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-kp3shcsz-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.3 GB at this context; duplicate traces of the same session: 9351romx, 9k75zg12, ltyh2wal; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1392.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 1,392 96.9 101.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kp3shcsz",
   "source_url": "https://omlx.ai/benchmarks/performance/kp3shcsz",
   "tg_tokens": null,
   "tps": 96.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-kp3shcsz-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: 9351romx, 9k75zg12, ltyh2wal; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1384.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 1,384 101.3 101.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kp3shcsz",
   "source_url": "https://omlx.ai/benchmarks/performance/kp3shcsz",
   "tg_tokens": null,
   "tps": 101.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-kp3shcsz-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.2 GB at this context; duplicate traces of the same session: 9351romx, 9k75zg12, ltyh2wal; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1373.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 1,373 86.0 102.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kp3shcsz",
   "source_url": "https://omlx.ai/benchmarks/performance/kp3shcsz",
   "tg_tokens": null,
   "tps": 86.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-kp3shcsz-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.4 GB at this context; duplicate traces of the same session: 9351romx, 9k75zg12, ltyh2wal; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1350.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 1,350 85.1 104.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kp3shcsz",
   "source_url": "https://omlx.ai/benchmarks/performance/kp3shcsz",
   "tg_tokens": null,
   "tps": 85.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-kp3shcsz-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 109.1 GB at this context; duplicate traces of the same session: 9351romx, 9k75zg12, ltyh2wal; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1326.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 1,326 65.9 109.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kp3shcsz",
   "source_url": "https://omlx.ai/benchmarks/performance/kp3shcsz",
   "tg_tokens": null,
   "tps": 65.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-kp3shcsz-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.1 GB at this context; duplicate traces of the same session: 9351romx, 9k75zg12, ltyh2wal; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1309.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 195k 1,309 58.8 110.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kp3shcsz",
   "source_url": "https://omlx.ai/benchmarks/performance/kp3shcsz",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-if0g97iz-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.4 GB at this context; duplicate traces of the same session: 7gej0vkf, 7sc0d6oo, ak1ysbhm, hrzhyjop, np4i7l7o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 553.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 553.3 53.0 71.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/if0g97iz",
   "source_url": "https://omlx.ai/benchmarks/performance/if0g97iz",
   "tg_tokens": null,
   "tps": 53.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-if0g97iz-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.5 GB at this context; duplicate traces of the same session: 7gej0vkf, 7sc0d6oo, ak1ysbhm, hrzhyjop, np4i7l7o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 816.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 816.7 61.1 71.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/if0g97iz",
   "source_url": "https://omlx.ai/benchmarks/performance/if0g97iz",
   "tg_tokens": null,
   "tps": 61.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-if0g97iz-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.4 GB at this context; duplicate traces of the same session: 7gej0vkf, 7sc0d6oo, ak1ysbhm, hrzhyjop, np4i7l7o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 813.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 813.5 61.5 72.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/if0g97iz",
   "source_url": "https://omlx.ai/benchmarks/performance/if0g97iz",
   "tg_tokens": null,
   "tps": 61.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-if0g97iz-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.5 GB at this context; duplicate traces of the same session: 7gej0vkf, 7sc0d6oo, ak1ysbhm, hrzhyjop, np4i7l7o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 803.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 803.3 62.2 74.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/if0g97iz",
   "source_url": "https://omlx.ai/benchmarks/performance/if0g97iz",
   "tg_tokens": null,
   "tps": 62.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-if0g97iz-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.6 GB at this context; duplicate traces of the same session: 7gej0vkf, 7sc0d6oo, ak1ysbhm, hrzhyjop, np4i7l7o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 791.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 791.5 53.7 78.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/if0g97iz",
   "source_url": "https://omlx.ai/benchmarks/performance/if0g97iz",
   "tg_tokens": null,
   "tps": 53.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-if0g97iz-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.0 GB at this context; duplicate traces of the same session: 7gej0vkf, 7sc0d6oo, ak1ysbhm, hrzhyjop, np4i7l7o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 781.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 781.9 52.3 80.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/if0g97iz",
   "source_url": "https://omlx.ai/benchmarks/performance/if0g97iz",
   "tg_tokens": null,
   "tps": 52.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1w7o5orp-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 31.2 GB; TTFT 7674 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 133.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 1k tokens 133.4 PP tok/s 16.1 TG tok/s 7674 TTFT (ms) 31.2 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1w7o5orp",
   "source_url": "https://omlx.ai/benchmarks/performance/1w7o5orp",
   "tg_tokens": null,
   "tps": 16.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-p5d3i285-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 31.5 GB; TTFT 18992 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 215.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k tokens 215.7 PP tok/s 14.6 TG tok/s 18992 TTFT (ms) 31.5 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p5d3i285",
   "source_url": "https://omlx.ai/benchmarks/performance/p5d3i285",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-p5yaacj7-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 5x7n089r; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1283.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,283 72.4 17.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p5yaacj7",
   "source_url": "https://omlx.ai/benchmarks/performance/p5yaacj7",
   "tg_tokens": null,
   "tps": 72.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-p5yaacj7-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: 5x7n089r; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1361.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,361 75.7 20.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p5yaacj7",
   "source_url": "https://omlx.ai/benchmarks/performance/p5yaacj7",
   "tg_tokens": null,
   "tps": 75.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-kxkr9dpp-ctx1024",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: nhdwgpwx; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1289.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,289 32.4 17.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kxkr9dpp",
   "source_url": "https://omlx.ai/benchmarks/performance/kxkr9dpp",
   "tg_tokens": null,
   "tps": 32.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-kxkr9dpp-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-Uncensored-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: nhdwgpwx; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1360.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Swift-1.5-Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,360 44.2 20.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kxkr9dpp",
   "source_url": "https://omlx.ai/benchmarks/performance/kxkr9dpp",
   "tg_tokens": null,
   "tps": 44.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-tgwhx95u-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 850.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 850.1 61.6 16.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tgwhx95u",
   "source_url": "https://omlx.ai/benchmarks/performance/tgwhx95u",
   "tg_tokens": null,
   "tps": 61.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-tgwhx95u-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 710.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 710.9 55.6 16.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tgwhx95u",
   "source_url": "https://omlx.ai/benchmarks/performance/tgwhx95u",
   "tg_tokens": null,
   "tps": 55.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-03us9szy-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.4 GB at this context; duplicate traces of the same session: 2vs6exz7, 69zj6bzr, kbjz4q39, stioojmr, tvjb1p7t; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 569.6,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "gpt-oss-20b-OptiQ-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 569.6 68.5 11.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/03us9szy",
   "source_url": "https://omlx.ai/benchmarks/performance/03us9szy",
   "tg_tokens": null,
   "tps": 68.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-03us9szy-ctx8192",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.4 GB at this context; duplicate traces of the same session: 2vs6exz7, 69zj6bzr, kbjz4q39, stioojmr, tvjb1p7t; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 592.8,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "gpt-oss-20b-OptiQ-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 592.8 61.9 11.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/03us9szy",
   "source_url": "https://omlx.ai/benchmarks/performance/03us9szy",
   "tg_tokens": null,
   "tps": 61.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-03us9szy-ctx16384",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.6 GB at this context; duplicate traces of the same session: 2vs6exz7, 69zj6bzr, kbjz4q39, stioojmr, tvjb1p7t; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 554.6,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "gpt-oss-20b-OptiQ-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 554.6 53.2 11.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/03us9szy",
   "source_url": "https://omlx.ai/benchmarks/performance/03us9szy",
   "tg_tokens": null,
   "tps": 53.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-03us9szy-ctx32768",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 12.0 GB at this context; duplicate traces of the same session: 2vs6exz7, 69zj6bzr, kbjz4q39, stioojmr, tvjb1p7t; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 490.2,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "gpt-oss-20b-OptiQ-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 32k 490.2 41.9 12.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/03us9szy",
   "source_url": "https://omlx.ai/benchmarks/performance/03us9szy",
   "tg_tokens": null,
   "tps": 41.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-03us9szy-ctx65536",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: 2vs6exz7, 69zj6bzr, kbjz4q39, stioojmr, tvjb1p7t; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 394.2,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "gpt-oss-20b-OptiQ-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 64k 394.2 29.0 12.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/03us9szy",
   "source_url": "https://omlx.ai/benchmarks/performance/03us9szy",
   "tg_tokens": null,
   "tps": 29.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-03us9szy-ctx131072",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 14.2 GB at this context; duplicate traces of the same session: 2vs6exz7, 69zj6bzr, kbjz4q39, stioojmr, tvjb1p7t; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 283.6,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "gpt-oss-20b-OptiQ-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 128k 283.6 18.0 14.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/03us9szy",
   "source_url": "https://omlx.ai/benchmarks/performance/03us9szy",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-h2k4lubt-ctx1024",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 676c4ruf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1471.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "lmstudio-community--Qwen3-Coder-30B-A3B-Instruct-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,471 111.5 16.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h2k4lubt",
   "source_url": "https://omlx.ai/benchmarks/performance/h2k4lubt",
   "tg_tokens": null,
   "tps": 111.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-h2k4lubt-ctx4096",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 676c4ruf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1600.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "lmstudio-community--Qwen3-Coder-30B-A3B-Instruct-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,600 105.3 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h2k4lubt",
   "source_url": "https://omlx.ai/benchmarks/performance/h2k4lubt",
   "tg_tokens": null,
   "tps": 105.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-a1cl6z4i-ctx1024",
   "model": "Agents-A1-4B",
   "notes": "oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: 1xz2233k; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 682.1,
   "provenance": "community",
   "quant": "oQe4",
   "quote": "Agents-A1-4B-oQe4 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 682.1 90.0 3.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a1cl6z4i",
   "source_url": "https://omlx.ai/benchmarks/performance/a1cl6z4i",
   "tg_tokens": null,
   "tps": 90.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-a1cl6z4i-ctx4096",
   "model": "Agents-A1-4B",
   "notes": "oMLX community benchmark trace; peak memory 5.0 GB at this context; duplicate traces of the same session: 1xz2233k; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 642.7,
   "provenance": "community",
   "quant": "oQe4",
   "quote": "Agents-A1-4B-oQe4 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 642.7 85.6 5.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a1cl6z4i",
   "source_url": "https://omlx.ai/benchmarks/performance/a1cl6z4i",
   "tg_tokens": null,
   "tps": 85.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-8wax9p66-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 40.8 GB at this context; duplicate traces of the same session: fvp6dgjt, nuweacc7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 37.4,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "GLM-5.3-Flash-oQ2e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 2bit; 1k 37.4 5.5 40.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8wax9p66",
   "source_url": "https://omlx.ai/benchmarks/performance/8wax9p66",
   "tg_tokens": null,
   "tps": 5.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-8wax9p66-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 41.6 GB at this context; duplicate traces of the same session: fvp6dgjt, nuweacc7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.5,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "GLM-5.3-Flash-oQ2e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 2bit; 4k 105.5 6.1 41.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8wax9p66",
   "source_url": "https://omlx.ai/benchmarks/performance/8wax9p66",
   "tg_tokens": null,
   "tps": 6.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-8wax9p66-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 42.0 GB at this context; duplicate traces of the same session: fvp6dgjt, nuweacc7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 99.9,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "GLM-5.3-Flash-oQ2e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 2bit; 8k 99.9 5.7 42.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8wax9p66",
   "source_url": "https://omlx.ai/benchmarks/performance/8wax9p66",
   "tg_tokens": null,
   "tps": 5.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-e6kxmi2o-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: mwhc6zv6; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1279.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,279 87.1 24.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e6kxmi2o",
   "source_url": "https://omlx.ai/benchmarks/performance/e6kxmi2o",
   "tg_tokens": null,
   "tps": 87.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-e6kxmi2o-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: mwhc6zv6; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1265.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 1,265 79.5 24.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e6kxmi2o",
   "source_url": "https://omlx.ai/benchmarks/performance/e6kxmi2o",
   "tg_tokens": null,
   "tps": 79.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-o8mzhp2l-ctx1024",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 9.8 GB at this context; duplicate traces of the same session: 1b5zoexu; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 654.2,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-6bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 6bit; 1k 654.2 73.0 9.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o8mzhp2l",
   "source_url": "https://omlx.ai/benchmarks/performance/o8mzhp2l",
   "tg_tokens": null,
   "tps": 73.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-o8mzhp2l-ctx4096",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "oMLX community benchmark trace; peak memory 9.9 GB at this context; duplicate traces of the same session: 1b5zoexu; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 670.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-6bit on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 6bit; 4k 670.8 71.8 9.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o8mzhp2l",
   "source_url": "https://omlx.ai/benchmarks/performance/o8mzhp2l",
   "tg_tokens": null,
   "tps": 71.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-89siy9ql-ctx1024",
   "model": "Nemotron-3.5-Lightning-30B-A3B-refusal-ablated",
   "notes": "MTP MLX build; oMLX community benchmark trace; peak memory 33.5 GB at this context; duplicate traces of the same session: y7fhddqe; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 887.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Nemotron-3.5-Lightning-30B-A3B-refusal-ablated-MLX-oQ8-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 887.8 73.5 33.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/89siy9ql",
   "source_url": "https://omlx.ai/benchmarks/performance/89siy9ql",
   "tg_tokens": null,
   "tps": 73.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-89siy9ql-ctx4096",
   "model": "Nemotron-3.5-Lightning-30B-A3B-refusal-ablated",
   "notes": "MTP MLX build; oMLX community benchmark trace; peak memory 34.5 GB at this context; duplicate traces of the same session: y7fhddqe; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 809.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Nemotron-3.5-Lightning-30B-A3B-refusal-ablated-MLX-oQ8-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 809.8 63.9 34.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/89siy9ql",
   "source_url": "https://omlx.ai/benchmarks/performance/89siy9ql",
   "tg_tokens": null,
   "tps": 63.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-889uakp1-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "refusal-ablated MTP build; MLX; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: 33tpkxyf; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 946.6,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-refusal-ablated-MLX-oQ8-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 946.6 90.4 37.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/889uakp1",
   "source_url": "https://omlx.ai/benchmarks/performance/889uakp1",
   "tg_tokens": null,
   "tps": 90.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-889uakp1-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "refusal-ablated MTP build; MLX; oMLX community benchmark trace; peak memory 39.5 GB at this context; duplicate traces of the same session: 33tpkxyf; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 925.1,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-refusal-ablated-MLX-oQ8-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 925.1 65.4 39.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/889uakp1",
   "source_url": "https://omlx.ai/benchmarks/performance/889uakp1",
   "tg_tokens": null,
   "tps": 65.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-8465yydc-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.4 GB at this context; duplicate traces of the same session: 5db4sosz; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 487.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 1k 487.8 53.7 35.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8465yydc",
   "source_url": "https://omlx.ai/benchmarks/performance/8465yydc",
   "tg_tokens": null,
   "tps": 53.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-8465yydc-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 36.2 GB at this context; duplicate traces of the same session: 5db4sosz; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 593.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k 593.8 51.7 36.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8465yydc",
   "source_url": "https://omlx.ai/benchmarks/performance/8465yydc",
   "tg_tokens": null,
   "tps": 51.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-wc4d42c8-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; MLX build; oMLX community benchmark trace, single run at 8k context; peak memory 31.4 GB; TTFT 11070 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 740.0,
   "provenance": "community",
   "quant": "6bit-XL",
   "quote": "Qwen3.6-35B-A3B-6bit-XL-MTP-mlx on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 8k tokens 740.0 PP tok/s 50.2 TG tok/s 11070 TTFT (ms) 31.4 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wc4d42c8",
   "source_url": "https://omlx.ai/benchmarks/performance/wc4d42c8",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-qwed2jt7-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 8k context; peak memory 29.4 GB; TTFT 10756 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 761.6,
   "provenance": "community",
   "quant": "6bit-XL",
   "quote": "Qwen3.6-35B-A3B-6bit-XL-mlx on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 8k tokens 761.6 PP tok/s 52.8 TG tok/s 10756 TTFT (ms) 29.4 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qwed2jt7",
   "source_url": "https://omlx.ai/benchmarks/performance/qwed2jt7",
   "tg_tokens": null,
   "tps": 52.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-0fieykp5-ctx4096",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.1 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 392.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 392.4 37.4 13.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0fieykp5",
   "source_url": "https://omlx.ai/benchmarks/performance/0fieykp5",
   "tg_tokens": null,
   "tps": 37.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-0fieykp5-ctx16384",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.5 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 314.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k 314.2 29.2 13.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0fieykp5",
   "source_url": "https://omlx.ai/benchmarks/performance/0fieykp5",
   "tg_tokens": null,
   "tps": 29.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-cd5rqbnk-ctx1024",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 14.7 GB at this context; duplicate traces of the same session: 953r6ba6; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1304.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 1k 1,304 58.5 14.7 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cd5rqbnk",
   "source_url": "https://omlx.ai/benchmarks/performance/cd5rqbnk",
   "tg_tokens": null,
   "tps": 58.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-cd5rqbnk-ctx8192",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 15.4 GB at this context; duplicate traces of the same session: 953r6ba6; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1302.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 8k 1,302 74.8 15.4 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cd5rqbnk",
   "source_url": "https://omlx.ai/benchmarks/performance/cd5rqbnk",
   "tg_tokens": null,
   "tps": 74.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-it7gpog5-ctx1024",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: cfhpj16m, ieem9str; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1311.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 1k 1,311 75.3 12.7 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/it7gpog5",
   "source_url": "https://omlx.ai/benchmarks/performance/it7gpog5",
   "tg_tokens": null,
   "tps": 75.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-it7gpog5-ctx4096",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: cfhpj16m, ieem9str; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1380.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 4k 1,380 73.9 13.4 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/it7gpog5",
   "source_url": "https://omlx.ai/benchmarks/performance/it7gpog5",
   "tg_tokens": null,
   "tps": 73.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-it7gpog5-ctx8192",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.5 GB at this context; duplicate traces of the same session: cfhpj16m, ieem9str; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1310.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 8k 1,310 71.5 13.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/it7gpog5",
   "source_url": "https://omlx.ai/benchmarks/performance/it7gpog5",
   "tg_tokens": null,
   "tps": 71.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-vy57b57u-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: zalkixkg; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 967.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 967.9 92.5 37.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vy57b57u",
   "source_url": "https://omlx.ai/benchmarks/performance/vy57b57u",
   "tg_tokens": null,
   "tps": 92.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-vy57b57u-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: zalkixkg; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 926.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 926.7 73.3 39.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vy57b57u",
   "source_url": "https://omlx.ai/benchmarks/performance/vy57b57u",
   "tg_tokens": null,
   "tps": 73.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-xxj5x783-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 1oxv7z33; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 127.7 8.5 28.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xxj5x783",
   "source_url": "https://omlx.ai/benchmarks/performance/xxj5x783",
   "tg_tokens": null,
   "tps": 8.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-xxj5x783-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.5 GB at this context; duplicate traces of the same session: 1oxv7z33; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 121.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 121.5 8.3 32.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xxj5x783",
   "source_url": "https://omlx.ai/benchmarks/performance/xxj5x783",
   "tg_tokens": null,
   "tps": 8.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-otmw2jub-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.7 GB at this context; duplicate traces of the same session: d1uycaq2; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 170.5,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 170.5 22.1 30.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/otmw2jub",
   "source_url": "https://omlx.ai/benchmarks/performance/otmw2jub",
   "tg_tokens": null,
   "tps": 22.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-otmw2jub-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.3 GB at this context; duplicate traces of the same session: d1uycaq2; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 198.2,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 198.2 22.0 34.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/otmw2jub",
   "source_url": "https://omlx.ai/benchmarks/performance/otmw2jub",
   "tg_tokens": null,
   "tps": 22.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-15q641sz-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 12ejxmsy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 771.8,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 771.8 27.2 16.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/15q641sz",
   "source_url": "https://omlx.ai/benchmarks/performance/15q641sz",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-15q641sz-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 12ejxmsy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 799.2,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 799.2 26.9 17.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/15q641sz",
   "source_url": "https://omlx.ai/benchmarks/performance/15q641sz",
   "tg_tokens": null,
   "tps": 26.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-62kj0hv6-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1186.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-27B-oQ6 on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 1,186 39.8 22.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/62kj0hv6",
   "source_url": "https://omlx.ai/benchmarks/performance/62kj0hv6",
   "tg_tokens": null,
   "tps": 39.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-62kj0hv6-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1563.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-27B-oQ6 on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 1,563 39.5 23.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/62kj0hv6",
   "source_url": "https://omlx.ai/benchmarks/performance/62kj0hv6",
   "tg_tokens": null,
   "tps": 39.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-62kj0hv6-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1741.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-27B-oQ6 on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 8k 1,741 39.1 24.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/62kj0hv6",
   "source_url": "https://omlx.ai/benchmarks/performance/62kj0hv6",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-62kj0hv6-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1695.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-27B-oQ6 on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 16k 1,695 38.0 24.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/62kj0hv6",
   "source_url": "https://omlx.ai/benchmarks/performance/62kj0hv6",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-62kj0hv6-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1601.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-27B-oQ6 on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 32k 1,601 36.6 25.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/62kj0hv6",
   "source_url": "https://omlx.ai/benchmarks/performance/62kj0hv6",
   "tg_tokens": null,
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-62kj0hv6-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1428.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-27B-oQ6 on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 64k 1,428 33.6 27.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/62kj0hv6",
   "source_url": "https://omlx.ai/benchmarks/performance/62kj0hv6",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mkjm0axp-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: ppc7ey1a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 289.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 289.5 16.1 33.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mkjm0axp",
   "source_url": "https://omlx.ai/benchmarks/performance/mkjm0axp",
   "tg_tokens": null,
   "tps": 16.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mkjm0axp-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: ppc7ey1a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 356.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 356.7 12.1 34.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mkjm0axp",
   "source_url": "https://omlx.ai/benchmarks/performance/mkjm0axp",
   "tg_tokens": null,
   "tps": 12.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-5jj8d5ku-ctx1024",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: 11otsl9i, rfms0qxj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 378.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 378.9 50.3 5.9 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5jj8d5ku",
   "source_url": "https://omlx.ai/benchmarks/performance/5jj8d5ku",
   "tg_tokens": null,
   "tps": 50.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-5jj8d5ku-ctx4096",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: 11otsl9i, rfms0qxj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 386.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 386.9 48.5 6.3 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5jj8d5ku",
   "source_url": "https://omlx.ai/benchmarks/performance/5jj8d5ku",
   "tg_tokens": null,
   "tps": 48.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-5jj8d5ku-ctx8192",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: 11otsl9i, rfms0qxj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 385.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 8k 385.7 48.4 6.5 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5jj8d5ku",
   "source_url": "https://omlx.ai/benchmarks/performance/5jj8d5ku",
   "tg_tokens": null,
   "tps": 48.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-5jj8d5ku-ctx16384",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: 11otsl9i, rfms0qxj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 371.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 16k 371.7 45.0 6.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5jj8d5ku",
   "source_url": "https://omlx.ai/benchmarks/performance/5jj8d5ku",
   "tg_tokens": null,
   "tps": 45.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-x4z2tpiw-ctx1024",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace, single run at 1k context; peak memory 5.9 GB; TTFT 2706 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 378.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k tokens 378.4 PP tok/s 49.9 TG tok/s 2706 TTFT (ms) 5.9 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x4z2tpiw",
   "source_url": "https://omlx.ai/benchmarks/performance/x4z2tpiw",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-0346ym48-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.6 GB at this context; duplicate traces of the same session: 0dx0makq, 2h29inxb, aft8nj4c, g7cdq5d7, h5l670we, temefm3t; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2511.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 2,511 150.5 142.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0346ym48",
   "source_url": "https://omlx.ai/benchmarks/performance/0346ym48",
   "tg_tokens": null,
   "tps": 150.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-0346ym48-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: 0dx0makq, 2h29inxb, aft8nj4c, g7cdq5d7, h5l670we, temefm3t; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4332.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 4,332 122.2 143.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0346ym48",
   "source_url": "https://omlx.ai/benchmarks/performance/0346ym48",
   "tg_tokens": null,
   "tps": 122.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-0346ym48-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 144.5 GB at this context; duplicate traces of the same session: 0dx0makq, 2h29inxb, aft8nj4c, g7cdq5d7, h5l670we, temefm3t; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4863.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 8k 4,863 120.2 144.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0346ym48",
   "source_url": "https://omlx.ai/benchmarks/performance/0346ym48",
   "tg_tokens": null,
   "tps": 120.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-0346ym48-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.9 GB at this context; duplicate traces of the same session: 0dx0makq, 2h29inxb, aft8nj4c, g7cdq5d7, h5l670we, temefm3t; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4843.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 32k 4,843 75.5 145.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0346ym48",
   "source_url": "https://omlx.ai/benchmarks/performance/0346ym48",
   "tg_tokens": null,
   "tps": 75.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-0346ym48-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 147.0 GB at this context; duplicate traces of the same session: 0dx0makq, 2h29inxb, aft8nj4c, g7cdq5d7, h5l670we, temefm3t; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4786.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 64k 4,786 57.5 147.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0346ym48",
   "source_url": "https://omlx.ai/benchmarks/performance/0346ym48",
   "tg_tokens": null,
   "tps": 57.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-0346ym48-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 149.3 GB at this context; duplicate traces of the same session: 0dx0makq, 2h29inxb, aft8nj4c, g7cdq5d7, h5l670we, temefm3t; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4721.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 128k 4,721 98.6 149.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0346ym48",
   "source_url": "https://omlx.ai/benchmarks/performance/0346ym48",
   "tg_tokens": null,
   "tps": 98.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-0346ym48-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 152.1 GB at this context; duplicate traces of the same session: 0dx0makq, 2h29inxb, aft8nj4c, g7cdq5d7, h5l670we, temefm3t; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4585.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 195k 4,585 80.7 152.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0346ym48",
   "source_url": "https://omlx.ai/benchmarks/performance/0346ym48",
   "tg_tokens": null,
   "tps": 80.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-n4225jwq-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace, single run at 64k context; peak memory 76.6 GB; TTFT 82913 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 790.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k tokens 790.4 PP tok/s 74.5 TG tok/s 82913 TTFT (ms) 76.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n4225jwq",
   "source_url": "https://omlx.ai/benchmarks/performance/n4225jwq",
   "tg_tokens": null,
   "tps": 74.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-gaghvcl1-ctx1024",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: cnolqlkg, doj416iy, hu3cjcvu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2199.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 2,199 47.9 27.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gaghvcl1",
   "source_url": "https://omlx.ai/benchmarks/performance/gaghvcl1",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-gaghvcl1-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: cnolqlkg, doj416iy, hu3cjcvu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2155.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 2,155 40.4 28.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gaghvcl1",
   "source_url": "https://omlx.ai/benchmarks/performance/gaghvcl1",
   "tg_tokens": null,
   "tps": 40.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-gaghvcl1-ctx8192",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 28.2 GB at this context; duplicate traces of the same session: cnolqlkg, doj416iy, hu3cjcvu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2153.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 2,153 31.7 28.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gaghvcl1",
   "source_url": "https://omlx.ai/benchmarks/performance/gaghvcl1",
   "tg_tokens": null,
   "tps": 31.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-gaghvcl1-ctx16384",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: cnolqlkg, doj416iy, hu3cjcvu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2032.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 16k 2,032 36.0 28.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gaghvcl1",
   "source_url": "https://omlx.ai/benchmarks/performance/gaghvcl1",
   "tg_tokens": null,
   "tps": 36.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-v7wle6zs-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 850.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 850.1 61.6 16.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v7wle6zs",
   "source_url": "https://omlx.ai/benchmarks/performance/v7wle6zs",
   "tg_tokens": null,
   "tps": 61.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-v7wle6zs-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 710.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 710.9 55.6 16.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v7wle6zs",
   "source_url": "https://omlx.ai/benchmarks/performance/v7wle6zs",
   "tg_tokens": null,
   "tps": 55.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ths4mi23-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "OptiQ fp16 build; oMLX community benchmark trace; peak memory 11.4 GB at this context; duplicate traces of the same session: 4p98wmrj, 4pr2ikfj, 8wc9mehx, d28zwj4z, tnkgxc40; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 721.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gpt-oss-20b-OptiQ-4bit-fp16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 721.9 71.0 11.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ths4mi23",
   "source_url": "https://omlx.ai/benchmarks/performance/ths4mi23",
   "tg_tokens": null,
   "tps": 71.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ths4mi23-ctx8192",
   "model": "gpt-oss 20B",
   "notes": "OptiQ fp16 build; oMLX community benchmark trace; peak memory 11.4 GB at this context; duplicate traces of the same session: 4p98wmrj, 4pr2ikfj, 8wc9mehx, d28zwj4z, tnkgxc40; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 737.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gpt-oss-20b-OptiQ-4bit-fp16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 737.0 64.4 11.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ths4mi23",
   "source_url": "https://omlx.ai/benchmarks/performance/ths4mi23",
   "tg_tokens": null,
   "tps": 64.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ths4mi23-ctx16384",
   "model": "gpt-oss 20B",
   "notes": "OptiQ fp16 build; oMLX community benchmark trace; peak memory 11.6 GB at this context; duplicate traces of the same session: 4p98wmrj, 4pr2ikfj, 8wc9mehx, d28zwj4z, tnkgxc40; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 693.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gpt-oss-20b-OptiQ-4bit-fp16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 693.7 56.0 11.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ths4mi23",
   "source_url": "https://omlx.ai/benchmarks/performance/ths4mi23",
   "tg_tokens": null,
   "tps": 56.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ths4mi23-ctx32768",
   "model": "gpt-oss 20B",
   "notes": "OptiQ fp16 build; oMLX community benchmark trace; peak memory 12.0 GB at this context; duplicate traces of the same session: 4p98wmrj, 4pr2ikfj, 8wc9mehx, d28zwj4z, tnkgxc40; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 598.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gpt-oss-20b-OptiQ-4bit-fp16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 32k 598.6 45.2 12.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ths4mi23",
   "source_url": "https://omlx.ai/benchmarks/performance/ths4mi23",
   "tg_tokens": null,
   "tps": 45.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ths4mi23-ctx65536",
   "model": "gpt-oss 20B",
   "notes": "OptiQ fp16 build; oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: 4p98wmrj, 4pr2ikfj, 8wc9mehx, d28zwj4z, tnkgxc40; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 469.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gpt-oss-20b-OptiQ-4bit-fp16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 64k 469.9 32.1 12.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ths4mi23",
   "source_url": "https://omlx.ai/benchmarks/performance/ths4mi23",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ths4mi23-ctx131072",
   "model": "gpt-oss 20B",
   "notes": "OptiQ fp16 build; oMLX community benchmark trace; peak memory 14.2 GB at this context; duplicate traces of the same session: 4p98wmrj, 4pr2ikfj, 8wc9mehx, d28zwj4z, tnkgxc40; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 328.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gpt-oss-20b-OptiQ-4bit-fp16 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 128k 328.4 19.7 14.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ths4mi23",
   "source_url": "https://omlx.ai/benchmarks/performance/ths4mi23",
   "tg_tokens": null,
   "tps": 19.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-d2p68hwn-ctx32768",
   "model": "Occamy-1.0",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 21.6 GB; TTFT 54601 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 600.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "occamy-1.0-oQ4e on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k tokens 600.1 PP tok/s 56.4 TG tok/s 54601 TTFT (ms) 21.6 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d2p68hwn",
   "source_url": "https://omlx.ai/benchmarks/performance/d2p68hwn",
   "tg_tokens": null,
   "tps": 56.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-pki1vk5c-ctx131072",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 27.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4338.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-24B-A2B-oQ8e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 4,338 69.9 27.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pki1vk5c",
   "source_url": "https://omlx.ai/benchmarks/performance/pki1vk5c",
   "tg_tokens": null,
   "tps": 69.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-pki1vk5c-ctx199680",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3377.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-24B-A2B-oQ8e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 195k 3,377 55.8 28.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pki1vk5c",
   "source_url": "https://omlx.ai/benchmarks/performance/pki1vk5c",
   "tg_tokens": null,
   "tps": 55.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-n5vpe3wv-ctx1024",
   "model": "Llama 3.3 70B",
   "notes": "oMLX community benchmark trace; peak memory 37.8 GB at this context; duplicate traces of the same session: j0hjem1f; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 114.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Llama-3.3-70B-Instruct-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 1k 114.7 15.2 37.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n5vpe3wv",
   "source_url": "https://omlx.ai/benchmarks/performance/n5vpe3wv",
   "tg_tokens": null,
   "tps": 15.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-n5vpe3wv-ctx4096",
   "model": "Llama 3.3 70B",
   "notes": "oMLX community benchmark trace; peak memory 39.0 GB at this context; duplicate traces of the same session: j0hjem1f; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 113.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Llama-3.3-70B-Instruct-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 4k 113.5 14.6 39.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n5vpe3wv",
   "source_url": "https://omlx.ai/benchmarks/performance/n5vpe3wv",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-d2uklc3b-ctx1024",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "OptiQ REAP-19B build; oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: gx1e0vfu; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 670.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM-V-4.7-35B-A3B-OptiQ-4bit-REAP-19B on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 670.4 45.1 14.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d2uklc3b",
   "source_url": "https://omlx.ai/benchmarks/performance/d2uklc3b",
   "tg_tokens": null,
   "tps": 45.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-d2uklc3b-ctx4096",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "OptiQ REAP-19B build; oMLX community benchmark trace; peak memory 15.2 GB at this context; duplicate traces of the same session: gx1e0vfu; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1277.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM-V-4.7-35B-A3B-OptiQ-4bit-REAP-19B on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,277 43.9 15.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d2uklc3b",
   "source_url": "https://omlx.ai/benchmarks/performance/d2uklc3b",
   "tg_tokens": null,
   "tps": 43.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-pehuc3u6-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 7.3 GB; TTFT 28717 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 570.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k tokens 570.5 PP tok/s 19.8 TG tok/s 28717 TTFT (ms) 7.3 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pehuc3u6",
   "source_url": "https://omlx.ai/benchmarks/performance/pehuc3u6",
   "tg_tokens": null,
   "tps": 19.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-k17dldv8-ctx4096",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.1 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 392.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 392.4 37.4 13.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k17dldv8",
   "source_url": "https://omlx.ai/benchmarks/performance/k17dldv8",
   "tg_tokens": null,
   "tps": 37.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-k17dldv8-ctx16384",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.5 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 314.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k 314.2 29.2 13.5 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k17dldv8",
   "source_url": "https://omlx.ai/benchmarks/performance/k17dldv8",
   "tg_tokens": null,
   "tps": 29.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2kszu6vu-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "MLX build; oMLX community benchmark trace, single run at 1k context; peak memory 27.5 GB; TTFT 304 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3367.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k tokens 3,367 PP tok/s 114.6 TG tok/s 304 TTFT (ms) 27.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2kszu6vu",
   "source_url": "https://omlx.ai/benchmarks/performance/2kszu6vu",
   "tg_tokens": null,
   "tps": 114.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1incj0yi-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored-Heretic",
   "notes": "MLX build; oMLX community benchmark trace, single run at 1k context; peak memory 27.5 GB; TTFT 260 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3932.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k tokens 3,932 PP tok/s 115.3 TG tok/s 260 TTFT (ms) 27.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1incj0yi",
   "source_url": "https://omlx.ai/benchmarks/performance/1incj0yi",
   "tg_tokens": null,
   "tps": 115.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-wdm44jxr-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.2 GB at this context; duplicate traces of the same session: i6r67fog; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 281.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 281.1 60.0 29.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wdm44jxr",
   "source_url": "https://omlx.ai/benchmarks/performance/wdm44jxr",
   "tg_tokens": null,
   "tps": 60.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-wdm44jxr-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.0 GB at this context; duplicate traces of the same session: i6r67fog; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 234.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 234.2 31.6 33.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wdm44jxr",
   "source_url": "https://omlx.ai/benchmarks/performance/wdm44jxr",
   "tg_tokens": null,
   "tps": 31.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-e8gzy9lv-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 27a35j69, f3ck92i2, gcqqjadn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 245.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 245.2 38.4 21.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e8gzy9lv",
   "source_url": "https://omlx.ai/benchmarks/performance/e8gzy9lv",
   "tg_tokens": null,
   "tps": 38.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-e8gzy9lv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: 27a35j69, f3ck92i2, gcqqjadn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 330.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 330.4 52.4 25.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e8gzy9lv",
   "source_url": "https://omlx.ai/benchmarks/performance/e8gzy9lv",
   "tg_tokens": null,
   "tps": 52.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-e8gzy9lv-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: 27a35j69, f3ck92i2, gcqqjadn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 335.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 335.3 33.4 25.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e8gzy9lv",
   "source_url": "https://omlx.ai/benchmarks/performance/e8gzy9lv",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-e8gzy9lv-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.0 GB at this context; duplicate traces of the same session: 27a35j69, f3ck92i2, gcqqjadn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 313.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 313.4 47.5 27.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e8gzy9lv",
   "source_url": "https://omlx.ai/benchmarks/performance/e8gzy9lv",
   "tg_tokens": null,
   "tps": 47.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dg2yhgae-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: 5dg7867l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 995.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 995.7 56.3 18.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dg2yhgae",
   "source_url": "https://omlx.ai/benchmarks/performance/dg2yhgae",
   "tg_tokens": null,
   "tps": 56.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-dg2yhgae-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 5dg7867l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 829.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 829.3 76.2 19.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dg2yhgae",
   "source_url": "https://omlx.ai/benchmarks/performance/dg2yhgae",
   "tg_tokens": null,
   "tps": 76.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-432x3amb-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: vqj3iuki; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 988.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 988.3 58.2 21.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/432x3amb",
   "source_url": "https://omlx.ai/benchmarks/performance/432x3amb",
   "tg_tokens": null,
   "tps": 58.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-432x3amb-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: vqj3iuki; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 849.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 849.6 64.7 22.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/432x3amb",
   "source_url": "https://omlx.ai/benchmarks/performance/432x3amb",
   "tg_tokens": null,
   "tps": 64.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1ushrsde-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: tevcje99; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 942.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 942.2 63.4 18.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ushrsde",
   "source_url": "https://omlx.ai/benchmarks/performance/1ushrsde",
   "tg_tokens": null,
   "tps": 63.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1ushrsde-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: tevcje99; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 825.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 825.0 73.9 19.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ushrsde",
   "source_url": "https://omlx.ai/benchmarks/performance/1ushrsde",
   "tg_tokens": null,
   "tps": 73.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-evw0uqcn-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 3skwcwna; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 829.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 829.6 46.9 21.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/evw0uqcn",
   "source_url": "https://omlx.ai/benchmarks/performance/evw0uqcn",
   "tg_tokens": null,
   "tps": 46.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-evw0uqcn-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: 3skwcwna; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 777.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 777.3 65.1 23.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/evw0uqcn",
   "source_url": "https://omlx.ai/benchmarks/performance/evw0uqcn",
   "tg_tokens": null,
   "tps": 65.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-clymau43-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: 1su69tuf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 952.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 952.0 47.2 18.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/clymau43",
   "source_url": "https://omlx.ai/benchmarks/performance/clymau43",
   "tg_tokens": null,
   "tps": 47.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-clymau43-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 1su69tuf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 871.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 871.1 75.9 19.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/clymau43",
   "source_url": "https://omlx.ai/benchmarks/performance/clymau43",
   "tg_tokens": null,
   "tps": 75.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-emj13d0u-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: ffbn2e15; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1008.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,008 54.4 21.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/emj13d0u",
   "source_url": "https://omlx.ai/benchmarks/performance/emj13d0u",
   "tg_tokens": null,
   "tps": 54.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-emj13d0u-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: ffbn2e15; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 849.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 849.6 62.0 22.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/emj13d0u",
   "source_url": "https://omlx.ai/benchmarks/performance/emj13d0u",
   "tg_tokens": null,
   "tps": 62.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wao38zxb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 2lqe4vho, fzrgo632, h5fys7m7, uf1hcez2, wak1s80o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 693.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 693.0 40.7 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wao38zxb",
   "source_url": "https://omlx.ai/benchmarks/performance/wao38zxb",
   "tg_tokens": null,
   "tps": 40.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wao38zxb-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 2lqe4vho, fzrgo632, h5fys7m7, uf1hcez2, wak1s80o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 525.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 525.4 54.5 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wao38zxb",
   "source_url": "https://omlx.ai/benchmarks/performance/wao38zxb",
   "tg_tokens": null,
   "tps": 54.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wao38zxb-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 2lqe4vho, fzrgo632, h5fys7m7, uf1hcez2, wak1s80o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 580.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 580.7 58.7 22.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wao38zxb",
   "source_url": "https://omlx.ai/benchmarks/performance/wao38zxb",
   "tg_tokens": null,
   "tps": 58.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-wao38zxb-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 2lqe4vho, fzrgo632, h5fys7m7, uf1hcez2, wak1s80o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 570.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 570.7 76.5 23.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wao38zxb",
   "source_url": "https://omlx.ai/benchmarks/performance/wao38zxb",
   "tg_tokens": null,
   "tps": 76.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wao38zxb-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: 2lqe4vho, fzrgo632, h5fys7m7, uf1hcez2, wak1s80o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 547.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 547.1 140.4 24.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wao38zxb",
   "source_url": "https://omlx.ai/benchmarks/performance/wao38zxb",
   "tg_tokens": null,
   "tps": 140.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wao38zxb-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: 2lqe4vho, fzrgo632, h5fys7m7, uf1hcez2, wak1s80o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 497.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 497.4 58.7 26.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wao38zxb",
   "source_url": "https://omlx.ai/benchmarks/performance/wao38zxb",
   "tg_tokens": null,
   "tps": 58.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-3m288bqz-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: pqffxguj; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1184.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 1k 1,184 88.0 17.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3m288bqz",
   "source_url": "https://omlx.ai/benchmarks/performance/3m288bqz",
   "tg_tokens": null,
   "tps": 88.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-3m288bqz-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: pqffxguj; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1469.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 4k 1,469 99.8 18.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3m288bqz",
   "source_url": "https://omlx.ai/benchmarks/performance/3m288bqz",
   "tg_tokens": null,
   "tps": 99.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-xv5gy0kf-ctx1024",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 378.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 378.9 50.3 5.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xv5gy0kf",
   "source_url": "https://omlx.ai/benchmarks/performance/xv5gy0kf",
   "tg_tokens": null,
   "tps": 50.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-xv5gy0kf-ctx4096",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 386.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 386.9 48.5 6.3 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xv5gy0kf",
   "source_url": "https://omlx.ai/benchmarks/performance/xv5gy0kf",
   "tg_tokens": null,
   "tps": 48.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-xv5gy0kf-ctx8192",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 385.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 8k 385.7 48.4 6.5 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xv5gy0kf",
   "source_url": "https://omlx.ai/benchmarks/performance/xv5gy0kf",
   "tg_tokens": null,
   "tps": 48.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-xv5gy0kf-ctx16384",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 371.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 16k 371.7 45.0 6.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xv5gy0kf",
   "source_url": "https://omlx.ai/benchmarks/performance/xv5gy0kf",
   "tg_tokens": null,
   "tps": 45.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-49j1c2b5-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 45.9 GB at this context; duplicate traces of the same session: ke7owyxj, l245vcm8, lrybn8eb; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 78.2,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 2bit; 1k 78.2 14.6 45.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/49j1c2b5",
   "source_url": "https://omlx.ai/benchmarks/performance/49j1c2b5",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-49j1c2b5-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 48.6 GB at this context; duplicate traces of the same session: ke7owyxj, l245vcm8, lrybn8eb; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 161.7,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 2bit; 4k 161.7 21.9 48.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/49j1c2b5",
   "source_url": "https://omlx.ai/benchmarks/performance/49j1c2b5",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-49j1c2b5-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 45.6 GB at this context; duplicate traces of the same session: ke7owyxj, l245vcm8, lrybn8eb; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 160.9,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 2bit; 8k 160.9 7.1 45.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/49j1c2b5",
   "source_url": "https://omlx.ai/benchmarks/performance/49j1c2b5",
   "tg_tokens": null,
   "tps": 7.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-49j1c2b5-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 46.2 GB at this context; duplicate traces of the same session: ke7owyxj, l245vcm8, lrybn8eb; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 182.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 2bit; 16k 182.0 8.1 46.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/49j1c2b5",
   "source_url": "https://omlx.ai/benchmarks/performance/49j1c2b5",
   "tg_tokens": null,
   "tps": 8.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-afqate9b-ctx1024",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: 2njfydr5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 449.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 449.8 34.3 18.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/afqate9b",
   "source_url": "https://omlx.ai/benchmarks/performance/afqate9b",
   "tg_tokens": null,
   "tps": 34.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-afqate9b-ctx4096",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 2njfydr5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 543.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 543.8 39.0 19.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/afqate9b",
   "source_url": "https://omlx.ai/benchmarks/performance/afqate9b",
   "tg_tokens": null,
   "tps": 39.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-wel1rbgk-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: ht7g13lp, wklfnq7a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 455.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 455.7 28.6 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wel1rbgk",
   "source_url": "https://omlx.ai/benchmarks/performance/wel1rbgk",
   "tg_tokens": null,
   "tps": 28.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-wel1rbgk-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: ht7g13lp, wklfnq7a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 623.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 623.1 54.5 71.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wel1rbgk",
   "source_url": "https://omlx.ai/benchmarks/performance/wel1rbgk",
   "tg_tokens": null,
   "tps": 54.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-wel1rbgk-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; Swift 1.5 build; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: ht7g13lp, wklfnq7a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 799.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 799.8 43.4 71.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wel1rbgk",
   "source_url": "https://omlx.ai/benchmarks/performance/wel1rbgk",
   "tg_tokens": null,
   "tps": 43.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-c17hxg82-ctx1024",
   "model": "Ternary-Bonsai-27B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 1k context; peak memory 9.1 GB; TTFT 9960 ms; duplicate traces of the same session: none; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.8,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-27B-mlx-2bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 2bit; 1k tokens 102.8 PP tok/s 17.2 TG tok/s 9960 TTFT (ms) 9.1 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c17hxg82",
   "source_url": "https://omlx.ai/benchmarks/performance/c17hxg82",
   "tg_tokens": null,
   "tps": 17.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-adgtg51i-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: lvg185d1, v5c8vnzi, zict8znq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 700.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26b-a4b-it-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 700.7 45.9 25.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/adgtg51i",
   "source_url": "https://omlx.ai/benchmarks/performance/adgtg51i",
   "tg_tokens": null,
   "tps": 45.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-adgtg51i-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: lvg185d1, v5c8vnzi, zict8znq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 768.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26b-a4b-it-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 768.7 41.4 26.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/adgtg51i",
   "source_url": "https://omlx.ai/benchmarks/performance/adgtg51i",
   "tg_tokens": null,
   "tps": 41.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-adgtg51i-ctx8192",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: lvg185d1, v5c8vnzi, zict8znq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 758.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26b-a4b-it-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 758.7 36.8 26.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/adgtg51i",
   "source_url": "https://omlx.ai/benchmarks/performance/adgtg51i",
   "tg_tokens": null,
   "tps": 36.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-adgtg51i-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: lvg185d1, v5c8vnzi, zict8znq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 718.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26b-a4b-it-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 16k 718.7 30.1 26.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/adgtg51i",
   "source_url": "https://omlx.ai/benchmarks/performance/adgtg51i",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-14hwlgui-ctx1024",
   "model": "gemma-4-31B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: fsfmm40u; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 90.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31B-it-qat-OptiQ-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 90.7 16.5 24.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/14hwlgui",
   "source_url": "https://omlx.ai/benchmarks/performance/14hwlgui",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-14hwlgui-ctx4096",
   "model": "gemma-4-31B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: fsfmm40u; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 89.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31B-it-qat-OptiQ-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 89.8 13.1 25.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/14hwlgui",
   "source_url": "https://omlx.ai/benchmarks/performance/14hwlgui",
   "tg_tokens": null,
   "tps": 13.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-blq4w3gg-ctx1024",
   "model": "gemma-4-31B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: mzezgdba; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 90.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31B-it-qat-OptiQ-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 90.6 11.9 24.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/blq4w3gg",
   "source_url": "https://omlx.ai/benchmarks/performance/blq4w3gg",
   "tg_tokens": null,
   "tps": 11.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-blq4w3gg-ctx4096",
   "model": "gemma-4-31B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: mzezgdba; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 88.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31B-it-qat-OptiQ-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 88.8 10.2 25.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/blq4w3gg",
   "source_url": "https://omlx.ai/benchmarks/performance/blq4w3gg",
   "tg_tokens": null,
   "tps": 10.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-txp3z85r-ctx1024",
   "model": "gemma-4-31B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: zb2tbstl; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 90.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31B-it-qat-OptiQ-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 90.7 12.4 24.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/txp3z85r",
   "source_url": "https://omlx.ai/benchmarks/performance/txp3z85r",
   "tg_tokens": null,
   "tps": 12.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-txp3z85r-ctx4096",
   "model": "gemma-4-31B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: zb2tbstl; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 89.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31B-it-qat-OptiQ-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 89.8 15.6 25.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/txp3z85r",
   "source_url": "https://omlx.ai/benchmarks/performance/txp3z85r",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-u8bk0gmb-ctx1024",
   "model": "gemma-4-31B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: j0pprcj9; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 90.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31B-it-qat-OptiQ-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 90.1 15.0 24.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u8bk0gmb",
   "source_url": "https://omlx.ai/benchmarks/performance/u8bk0gmb",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-u8bk0gmb-ctx4096",
   "model": "gemma-4-31B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: j0pprcj9; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 89.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31B-it-qat-OptiQ-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 89.0 10.0 25.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u8bk0gmb",
   "source_url": "https://omlx.ai/benchmarks/performance/u8bk0gmb",
   "tg_tokens": null,
   "tps": 10.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-6g5gidzm-ctx1024",
   "model": "gemma-4-31B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: y9j5a7s1; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 89.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31B-it-qat-OptiQ-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 89.8 12.9 22.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6g5gidzm",
   "source_url": "https://omlx.ai/benchmarks/performance/6g5gidzm",
   "tg_tokens": null,
   "tps": 12.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-6g5gidzm-ctx4096",
   "model": "gemma-4-31B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: y9j5a7s1; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 89.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31B-it-qat-OptiQ-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 89.7 12.5 23.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6g5gidzm",
   "source_url": "https://omlx.ai/benchmarks/performance/6g5gidzm",
   "tg_tokens": null,
   "tps": 12.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9l11tk8o-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 91.0 GB at this context; duplicate traces of the same session: fzyom898, kvwn8cbo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 116.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 116.0 13.3 91.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9l11tk8o",
   "source_url": "https://omlx.ai/benchmarks/performance/9l11tk8o",
   "tg_tokens": null,
   "tps": 13.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9l11tk8o-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 94.2 GB at this context; duplicate traces of the same session: fzyom898, kvwn8cbo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 526.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 526.0 13.9 94.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9l11tk8o",
   "source_url": "https://omlx.ai/benchmarks/performance/9l11tk8o",
   "tg_tokens": null,
   "tps": 13.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9l11tk8o-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 94.3 GB at this context; duplicate traces of the same session: fzyom898, kvwn8cbo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 613.3,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 613.3 17.6 94.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9l11tk8o",
   "source_url": "https://omlx.ai/benchmarks/performance/9l11tk8o",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-kki11t78-ctx131072",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 27.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4338.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-24B-A2B-oQ8e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 4,338 69.9 27.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kki11t78",
   "source_url": "https://omlx.ai/benchmarks/performance/kki11t78",
   "tg_tokens": null,
   "tps": 69.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-kki11t78-ctx199680",
   "model": "LFM2.5-24B-A2B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3377.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-24B-A2B-oQ8e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 195k 3,377 55.8 28.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kki11t78",
   "source_url": "https://omlx.ai/benchmarks/performance/kki11t78",
   "tg_tokens": null,
   "tps": 55.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-h8gtmaw3-ctx1024",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: 29hviwjc, 93dfho4r, ax2mdtr9, o4542ecg, tugp9nim, x8pxc95c; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2426.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ6 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 2,426 136.1 26.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h8gtmaw3",
   "source_url": "https://omlx.ai/benchmarks/performance/h8gtmaw3",
   "tg_tokens": null,
   "tps": 136.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-h8gtmaw3-ctx4096",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: 29hviwjc, 93dfho4r, ax2mdtr9, o4542ecg, tugp9nim, x8pxc95c; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2644.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ6 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 2,644 135.0 27.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h8gtmaw3",
   "source_url": "https://omlx.ai/benchmarks/performance/h8gtmaw3",
   "tg_tokens": null,
   "tps": 135.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-h8gtmaw3-ctx8192",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: 29hviwjc, 93dfho4r, ax2mdtr9, o4542ecg, tugp9nim, x8pxc95c; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2591.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ6 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 8k 2,591 132.4 27.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h8gtmaw3",
   "source_url": "https://omlx.ai/benchmarks/performance/h8gtmaw3",
   "tg_tokens": null,
   "tps": 132.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-h8gtmaw3-ctx16384",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: 29hviwjc, 93dfho4r, ax2mdtr9, o4542ecg, tugp9nim, x8pxc95c; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2478.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ6 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 16k 2,478 127.9 27.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h8gtmaw3",
   "source_url": "https://omlx.ai/benchmarks/performance/h8gtmaw3",
   "tg_tokens": null,
   "tps": 127.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-h8gtmaw3-ctx32768",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 27.6 GB at this context; duplicate traces of the same session: 29hviwjc, 93dfho4r, ax2mdtr9, o4542ecg, tugp9nim, x8pxc95c; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2271.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ6 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 32k 2,271 122.4 27.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h8gtmaw3",
   "source_url": "https://omlx.ai/benchmarks/performance/h8gtmaw3",
   "tg_tokens": null,
   "tps": 122.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-h8gtmaw3-ctx65536",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: 29hviwjc, 93dfho4r, ax2mdtr9, o4542ecg, tugp9nim, x8pxc95c; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1943.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ6 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 64k 1,943 110.5 27.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h8gtmaw3",
   "source_url": "https://omlx.ai/benchmarks/performance/h8gtmaw3",
   "tg_tokens": null,
   "tps": 110.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-h8gtmaw3-ctx131072",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: 29hviwjc, 93dfho4r, ax2mdtr9, o4542ecg, tugp9nim, x8pxc95c; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1506.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ6 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 128k 1,506 95.8 28.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h8gtmaw3",
   "source_url": "https://omlx.ai/benchmarks/performance/h8gtmaw3",
   "tg_tokens": null,
   "tps": 95.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2rut1aix-ctx1024",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: 2gjksm20; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2001.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ6 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 2,001 134.8 26.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2rut1aix",
   "source_url": "https://omlx.ai/benchmarks/performance/2rut1aix",
   "tg_tokens": null,
   "tps": 134.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2rut1aix-ctx4096",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: 2gjksm20; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2640.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ6 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 2,640 134.0 27.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2rut1aix",
   "source_url": "https://omlx.ai/benchmarks/performance/2rut1aix",
   "tg_tokens": null,
   "tps": 134.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2rut1aix-ctx8192",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: 2gjksm20; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2574.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ6 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 8k 2,574 131.5 27.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2rut1aix",
   "source_url": "https://omlx.ai/benchmarks/performance/2rut1aix",
   "tg_tokens": null,
   "tps": 131.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2rut1aix-ctx16384",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: 2gjksm20; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2469.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ6 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 16k 2,469 128.6 27.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2rut1aix",
   "source_url": "https://omlx.ai/benchmarks/performance/2rut1aix",
   "tg_tokens": null,
   "tps": 128.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-zjilskvq-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 9hr5gvfd, djv50u7a, luiqkuw0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2171.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 2,171 149.7 21.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zjilskvq",
   "source_url": "https://omlx.ai/benchmarks/performance/zjilskvq",
   "tg_tokens": null,
   "tps": 149.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-zjilskvq-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: 9hr5gvfd, djv50u7a, luiqkuw0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2648.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 2,648 143.9 23.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zjilskvq",
   "source_url": "https://omlx.ai/benchmarks/performance/zjilskvq",
   "tg_tokens": null,
   "tps": 143.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-zjilskvq-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 9hr5gvfd, djv50u7a, luiqkuw0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2502.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 2,502 143.3 23.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zjilskvq",
   "source_url": "https://omlx.ai/benchmarks/performance/zjilskvq",
   "tg_tokens": null,
   "tps": 143.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-zjilskvq-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: 9hr5gvfd, djv50u7a, luiqkuw0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2228.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 2,228 187.1 23.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zjilskvq",
   "source_url": "https://omlx.ai/benchmarks/performance/zjilskvq",
   "tg_tokens": null,
   "tps": 187.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2",
   "id": "omlx-3wghp5ap-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: t0ty89pa; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 109.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M2 (8c); M2 (8c) \u00b7 16 GB \u00b7 4bit; 1k 109.9 18.5 5.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3wghp5ap",
   "source_url": "https://omlx.ai/benchmarks/performance/3wghp5ap",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2",
   "id": "omlx-3wghp5ap-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: t0ty89pa; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-4bit on M2 (8c); M2 (8c) \u00b7 16 GB \u00b7 4bit; 4k 105.5 18.3 6.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3wghp5ap",
   "source_url": "https://omlx.ai/benchmarks/performance/3wghp5ap",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hkbqsxwc-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 23.2 GB; TTFT 52429 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2500.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k tokens 2,500 PP tok/s 42.3 TG tok/s 52429 TTFT (ms) 23.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hkbqsxwc",
   "source_url": "https://omlx.ai/benchmarks/performance/hkbqsxwc",
   "tg_tokens": null,
   "tps": 42.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-diyizq2x-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: i7j2yhfs; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 803.9,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 803.9 124.3 20.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/diyizq2x",
   "source_url": "https://omlx.ai/benchmarks/performance/diyizq2x",
   "tg_tokens": null,
   "tps": 124.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-diyizq2x-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: i7j2yhfs; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4161.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 4,161 86.8 21.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/diyizq2x",
   "source_url": "https://omlx.ai/benchmarks/performance/diyizq2x",
   "tg_tokens": null,
   "tps": 86.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9nvd8uj6-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: kw36avbt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 334.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 334.2 34.9 17.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9nvd8uj6",
   "source_url": "https://omlx.ai/benchmarks/performance/9nvd8uj6",
   "tg_tokens": null,
   "tps": 34.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9nvd8uj6-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: kw36avbt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 485.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 485.3 36.0 18.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9nvd8uj6",
   "source_url": "https://omlx.ai/benchmarks/performance/9nvd8uj6",
   "tg_tokens": null,
   "tps": 36.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lpl0umjp-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: dpyw4j27; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 323.3 28.2 32.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lpl0umjp",
   "source_url": "https://omlx.ai/benchmarks/performance/lpl0umjp",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lpl0umjp-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.7 GB at this context; duplicate traces of the same session: dpyw4j27; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 471.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 471.7 22.4 33.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lpl0umjp",
   "source_url": "https://omlx.ai/benchmarks/performance/lpl0umjp",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-o7bzwkut-ctx32768",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "MTPLX build; oMLX community benchmark trace, single run at 32k context; peak memory 22.7 GB; TTFT 46351 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 707.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Abliterated-MTPLX-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 32k tokens 707.0 PP tok/s 33.1 TG tok/s 46351 TTFT (ms) 22.7 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o7bzwkut",
   "source_url": "https://omlx.ai/benchmarks/performance/o7bzwkut",
   "tg_tokens": null,
   "tps": 33.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-c4c0vj2g-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 19.6 GB; TTFT 345243 ms; duplicate traces of the same session: none; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 94.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 32k tokens 94.9 PP tok/s 7.5 TG tok/s 345243 TTFT (ms) 19.6 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c4c0vj2g",
   "source_url": "https://omlx.ai/benchmarks/performance/c4c0vj2g",
   "tg_tokens": null,
   "tps": 7.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9ex9bb9a-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: 183wybgd, 4ymdzwlt, 56hcghek, b59xkyqr, dehba9kv, mrmxta71; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 769.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 769.6 44.1 29.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ex9bb9a",
   "source_url": "https://omlx.ai/benchmarks/performance/9ex9bb9a",
   "tg_tokens": null,
   "tps": 44.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9ex9bb9a-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 30.8 GB at this context; duplicate traces of the same session: 183wybgd, 4ymdzwlt, 56hcghek, b59xkyqr, dehba9kv, mrmxta71; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 927.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 927.6 34.6 30.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ex9bb9a",
   "source_url": "https://omlx.ai/benchmarks/performance/9ex9bb9a",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9ex9bb9a-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.1 GB at this context; duplicate traces of the same session: 183wybgd, 4ymdzwlt, 56hcghek, b59xkyqr, dehba9kv, mrmxta71; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 922.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 8k 922.6 52.4 31.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ex9bb9a",
   "source_url": "https://omlx.ai/benchmarks/performance/9ex9bb9a",
   "tg_tokens": null,
   "tps": 52.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9ex9bb9a-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.6 GB at this context; duplicate traces of the same session: 183wybgd, 4ymdzwlt, 56hcghek, b59xkyqr, dehba9kv, mrmxta71; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 912.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 16k 912.3 59.7 31.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ex9bb9a",
   "source_url": "https://omlx.ai/benchmarks/performance/9ex9bb9a",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9ex9bb9a-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.6 GB at this context; duplicate traces of the same session: 183wybgd, 4ymdzwlt, 56hcghek, b59xkyqr, dehba9kv, mrmxta71; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 868.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 32k 868.7 45.3 32.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ex9bb9a",
   "source_url": "https://omlx.ai/benchmarks/performance/9ex9bb9a",
   "tg_tokens": null,
   "tps": 45.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9ex9bb9a-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.6 GB at this context; duplicate traces of the same session: 183wybgd, 4ymdzwlt, 56hcghek, b59xkyqr, dehba9kv, mrmxta71; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 772.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 64k 772.4 31.2 34.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ex9bb9a",
   "source_url": "https://omlx.ai/benchmarks/performance/9ex9bb9a",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9ex9bb9a-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 38.7 GB at this context; duplicate traces of the same session: 183wybgd, 4ymdzwlt, 56hcghek, b59xkyqr, dehba9kv, mrmxta71; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 624.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 128k 624.9 33.1 38.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ex9bb9a",
   "source_url": "https://omlx.ai/benchmarks/performance/9ex9bb9a",
   "tg_tokens": null,
   "tps": 33.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-6lmhj3o3-ctx1024",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: 4hvztg90, b4adqjql, halaymyn, i9tuc3za, w511sf4h, wcc38q2j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 809.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ8e-1M on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 809.8 41.8 33.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6lmhj3o3",
   "source_url": "https://omlx.ai/benchmarks/performance/6lmhj3o3",
   "tg_tokens": null,
   "tps": 41.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-6lmhj3o3-ctx4096",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: 4hvztg90, b4adqjql, halaymyn, i9tuc3za, w511sf4h, wcc38q2j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 930.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ8e-1M on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 930.1 42.9 34.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6lmhj3o3",
   "source_url": "https://omlx.ai/benchmarks/performance/6lmhj3o3",
   "tg_tokens": null,
   "tps": 42.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-6lmhj3o3-ctx8192",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: 4hvztg90, b4adqjql, halaymyn, i9tuc3za, w511sf4h, wcc38q2j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 933.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ8e-1M on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 8k 933.3 51.4 35.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6lmhj3o3",
   "source_url": "https://omlx.ai/benchmarks/performance/6lmhj3o3",
   "tg_tokens": null,
   "tps": 51.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-6lmhj3o3-ctx16384",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: 4hvztg90, b4adqjql, halaymyn, i9tuc3za, w511sf4h, wcc38q2j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 910.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ8e-1M on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 16k 910.2 58.5 35.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6lmhj3o3",
   "source_url": "https://omlx.ai/benchmarks/performance/6lmhj3o3",
   "tg_tokens": null,
   "tps": 58.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-6lmhj3o3-ctx32768",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 36.9 GB at this context; duplicate traces of the same session: 4hvztg90, b4adqjql, halaymyn, i9tuc3za, w511sf4h, wcc38q2j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 852.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ8e-1M on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 32k 852.6 83.6 36.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6lmhj3o3",
   "source_url": "https://omlx.ai/benchmarks/performance/6lmhj3o3",
   "tg_tokens": null,
   "tps": 83.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-6lmhj3o3-ctx65536",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: 4hvztg90, b4adqjql, halaymyn, i9tuc3za, w511sf4h, wcc38q2j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 751.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ8e-1M on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 64k 751.2 42.8 39.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6lmhj3o3",
   "source_url": "https://omlx.ai/benchmarks/performance/6lmhj3o3",
   "tg_tokens": null,
   "tps": 42.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-6lmhj3o3-ctx131072",
   "model": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored",
   "notes": "NM DAU build, 1M ctx; oMLX community benchmark trace; peak memory 44.0 GB at this context; duplicate traces of the same session: 4hvztg90, b4adqjql, halaymyn, i9tuc3za, w511sf4h, wcc38q2j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 597.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ8e-1M on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 128k 597.0 39.7 44.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6lmhj3o3",
   "source_url": "https://omlx.ai/benchmarks/performance/6lmhj3o3",
   "tg_tokens": null,
   "tps": 39.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i7jdohgs-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: 80w7xduu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 744.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 744.0 35.1 16.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i7jdohgs",
   "source_url": "https://omlx.ai/benchmarks/performance/i7jdohgs",
   "tg_tokens": null,
   "tps": 35.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-i7jdohgs-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.9 GB at this context; duplicate traces of the same session: 80w7xduu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 434.7,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 434.7 31.7 17.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i7jdohgs",
   "source_url": "https://omlx.ai/benchmarks/performance/i7jdohgs",
   "tg_tokens": null,
   "tps": 31.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tllw0132-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 30suoeap; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1002.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,002 50.5 21.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tllw0132",
   "source_url": "https://omlx.ai/benchmarks/performance/tllw0132",
   "tg_tokens": null,
   "tps": 50.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tllw0132-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: 30suoeap; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 853.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 853.0 54.9 22.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tllw0132",
   "source_url": "https://omlx.ai/benchmarks/performance/tllw0132",
   "tg_tokens": null,
   "tps": 54.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-u3y98fwa-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: h8cyn7lh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 919.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 919.0 50.1 21.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u3y98fwa",
   "source_url": "https://omlx.ai/benchmarks/performance/u3y98fwa",
   "tg_tokens": null,
   "tps": 50.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-u3y98fwa-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: h8cyn7lh; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 847.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 847.7 52.0 23.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u3y98fwa",
   "source_url": "https://omlx.ai/benchmarks/performance/u3y98fwa",
   "tg_tokens": null,
   "tps": 52.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mgb0mrcg-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: ahjjtx39; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1002.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,002 47.1 21.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mgb0mrcg",
   "source_url": "https://omlx.ai/benchmarks/performance/mgb0mrcg",
   "tg_tokens": null,
   "tps": 47.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mgb0mrcg-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: ahjjtx39; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 807.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 807.7 59.0 22.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mgb0mrcg",
   "source_url": "https://omlx.ai/benchmarks/performance/mgb0mrcg",
   "tg_tokens": null,
   "tps": 59.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7j5wkh8a-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.6 GB at this context; duplicate traces of the same session: n69b7l88; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 407.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 407.6 10.1 37.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7j5wkh8a",
   "source_url": "https://omlx.ai/benchmarks/performance/7j5wkh8a",
   "tg_tokens": null,
   "tps": 10.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7j5wkh8a-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 41.5 GB at this context; duplicate traces of the same session: n69b7l88; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 306.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 195k 306.7 9.4 41.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7j5wkh8a",
   "source_url": "https://omlx.ai/benchmarks/performance/7j5wkh8a",
   "tg_tokens": null,
   "tps": 9.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-nzvuyvvc-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 46.1 GB; TTFT 108847 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1204.0,
   "provenance": "community",
   "quant": "GSQ-RCO-Q2",
   "quote": "Qwen3.8-Flash-Next-GSQ-RCO-Q2-mlx on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k tokens 1,204 PP tok/s 51.1 TG tok/s 108847 TTFT (ms) 46.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nzvuyvvc",
   "source_url": "https://omlx.ai/benchmarks/performance/nzvuyvvc",
   "tg_tokens": null,
   "tps": 51.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vas2vl2v-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 104.7 GB; TTFT 95606 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1371.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k tokens 1,371 PP tok/s 61.1 TG tok/s 95606 TTFT (ms) 104.7 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vas2vl2v",
   "source_url": "https://omlx.ai/benchmarks/performance/vas2vl2v",
   "tg_tokens": null,
   "tps": 61.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-12w13lba-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 52.8 GB; TTFT 10577 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 96.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 96.8 PP tok/s 11.0 TG tok/s 10577 TTFT (ms) 52.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/12w13lba",
   "source_url": "https://omlx.ai/benchmarks/performance/12w13lba",
   "tg_tokens": null,
   "tps": 11.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5px3mpo0-ctx1024",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: 63wyljpz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 746.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 746.7 50.2 21.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5px3mpo0",
   "source_url": "https://omlx.ai/benchmarks/performance/5px3mpo0",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5px3mpo0-ctx4096",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 63wyljpz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 626.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 626.1 44.9 22.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5px3mpo0",
   "source_url": "https://omlx.ai/benchmarks/performance/5px3mpo0",
   "tg_tokens": null,
   "tps": 44.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-sgn5iezi-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: 7mnqp9z1, teeh31l8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 802.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26b-a4b-it-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 802.2 45.6 25.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sgn5iezi",
   "source_url": "https://omlx.ai/benchmarks/performance/sgn5iezi",
   "tg_tokens": null,
   "tps": 45.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-sgn5iezi-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: 7mnqp9z1, teeh31l8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 784.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26b-a4b-it-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 784.0 41.1 26.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sgn5iezi",
   "source_url": "https://omlx.ai/benchmarks/performance/sgn5iezi",
   "tg_tokens": null,
   "tps": 41.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-sgn5iezi-ctx8192",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: 7mnqp9z1, teeh31l8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 760.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26b-a4b-it-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 760.6 36.6 26.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sgn5iezi",
   "source_url": "https://omlx.ai/benchmarks/performance/sgn5iezi",
   "tg_tokens": null,
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-sgn5iezi-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: 7mnqp9z1, teeh31l8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 712.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26b-a4b-it-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 16k 712.0 30.1 26.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sgn5iezi",
   "source_url": "https://omlx.ai/benchmarks/performance/sgn5iezi",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-w144gxsc-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 187.2 GB at this context; duplicate traces of the same session: djw55gkx, ez6jp0dy; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 62.5,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ3e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 3bit; 1k 62.5 12.9 187.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w144gxsc",
   "source_url": "https://omlx.ai/benchmarks/performance/w144gxsc",
   "tg_tokens": null,
   "tps": 12.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-w144gxsc-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 185.3 GB at this context; duplicate traces of the same session: djw55gkx, ez6jp0dy; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 389.7,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ3e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 3bit; 16k 389.7 17.1 185.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w144gxsc",
   "source_url": "https://omlx.ai/benchmarks/performance/w144gxsc",
   "tg_tokens": null,
   "tps": 17.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-w144gxsc-ctx32768",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 185.3 GB at this context; duplicate traces of the same session: djw55gkx, ez6jp0dy; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 430.6,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ3e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 3bit; 32k 430.6 17.4 185.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w144gxsc",
   "source_url": "https://omlx.ai/benchmarks/performance/w144gxsc",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bz710wd5-ctx1024",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: twsc6eij; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2001.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ6 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 2,001 134.8 26.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bz710wd5",
   "source_url": "https://omlx.ai/benchmarks/performance/bz710wd5",
   "tg_tokens": null,
   "tps": 134.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bz710wd5-ctx4096",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: twsc6eij; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2640.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ6 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 2,640 134.0 27.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bz710wd5",
   "source_url": "https://omlx.ai/benchmarks/performance/bz710wd5",
   "tg_tokens": null,
   "tps": 134.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bz710wd5-ctx8192",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: twsc6eij; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2574.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ6 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 8k 2,574 131.5 27.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bz710wd5",
   "source_url": "https://omlx.ai/benchmarks/performance/bz710wd5",
   "tg_tokens": null,
   "tps": 131.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bz710wd5-ctx16384",
   "model": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: twsc6eij; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2469.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-oQ6 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 16k 2,469 128.6 27.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bz710wd5",
   "source_url": "https://omlx.ai/benchmarks/performance/bz710wd5",
   "tg_tokens": null,
   "tps": 128.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-h50wl9gn-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 22.4 GB; TTFT 26541 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 617.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 16k tokens 617.3 PP tok/s 70.1 TG tok/s 26541 TTFT (ms) 22.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h50wl9gn",
   "source_url": "https://omlx.ai/benchmarks/performance/h50wl9gn",
   "tg_tokens": null,
   "tps": 70.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-pflpwc49-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: nlqllvqi; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 689.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 689.2 58.6 20.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pflpwc49",
   "source_url": "https://omlx.ai/benchmarks/performance/pflpwc49",
   "tg_tokens": null,
   "tps": 58.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-pflpwc49-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: nlqllvqi; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 817.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 817.2 55.9 22.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pflpwc49",
   "source_url": "https://omlx.ai/benchmarks/performance/pflpwc49",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-ety7pnon-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 20.5 GB; TTFT 1514 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 676.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-UD-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k tokens 676.4 PP tok/s 56.9 TG tok/s 1514 TTFT (ms) 20.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ety7pnon",
   "source_url": "https://omlx.ai/benchmarks/performance/ety7pnon",
   "tg_tokens": null,
   "tps": 56.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-iq65unq1-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: zj0qfv9h; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 648.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 648.4 63.1 21.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iq65unq1",
   "source_url": "https://omlx.ai/benchmarks/performance/iq65unq1",
   "tg_tokens": null,
   "tps": 63.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-iq65unq1-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: zj0qfv9h; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 813.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 813.9 66.9 23.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iq65unq1",
   "source_url": "https://omlx.ai/benchmarks/performance/iq65unq1",
   "tg_tokens": null,
   "tps": 66.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-gtknyrpq-ctx32768",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "MTPLX build; oMLX community benchmark trace, single run at 32k context; peak memory 22.7 GB; TTFT 45756 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 716.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Abliterated-MTPLX-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 32k tokens 716.2 PP tok/s 35.2 TG tok/s 45756 TTFT (ms) 22.7 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gtknyrpq",
   "source_url": "https://omlx.ai/benchmarks/performance/gtknyrpq",
   "tg_tokens": null,
   "tps": 35.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lzqqhfzx-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: q5k85q3s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 997.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 997.3 64.4 18.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lzqqhfzx",
   "source_url": "https://omlx.ai/benchmarks/performance/lzqqhfzx",
   "tg_tokens": null,
   "tps": 64.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lzqqhfzx-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: q5k85q3s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 872.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 872.9 77.0 19.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lzqqhfzx",
   "source_url": "https://omlx.ai/benchmarks/performance/lzqqhfzx",
   "tg_tokens": null,
   "tps": 77.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-l0kkoz8d-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: hbgakefr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 987.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 987.0 48.0 21.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l0kkoz8d",
   "source_url": "https://omlx.ai/benchmarks/performance/l0kkoz8d",
   "tg_tokens": null,
   "tps": 48.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-l0kkoz8d-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: hbgakefr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 842.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 842.2 62.6 23.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l0kkoz8d",
   "source_url": "https://omlx.ai/benchmarks/performance/l0kkoz8d",
   "tg_tokens": null,
   "tps": 62.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-24x7hl3m-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: bqj5yiei; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 593.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 593.0 40.6 18.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/24x7hl3m",
   "source_url": "https://omlx.ai/benchmarks/performance/24x7hl3m",
   "tg_tokens": null,
   "tps": 40.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-24x7hl3m-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: bqj5yiei; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 601.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 601.4 64.6 19.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/24x7hl3m",
   "source_url": "https://omlx.ai/benchmarks/performance/24x7hl3m",
   "tg_tokens": null,
   "tps": 64.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bku34344-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: lu63b8c8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 948.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 948.1 45.2 21.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bku34344",
   "source_url": "https://omlx.ai/benchmarks/performance/bku34344",
   "tg_tokens": null,
   "tps": 45.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-bku34344-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: lu63b8c8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 798.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 798.0 57.9 22.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bku34344",
   "source_url": "https://omlx.ai/benchmarks/performance/bku34344",
   "tg_tokens": null,
   "tps": 57.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-f39lvtca-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: 28vywwum; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 853.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 853.2 50.6 18.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f39lvtca",
   "source_url": "https://omlx.ai/benchmarks/performance/f39lvtca",
   "tg_tokens": null,
   "tps": 50.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-f39lvtca-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 28vywwum; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 661.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 661.9 64.3 19.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f39lvtca",
   "source_url": "https://omlx.ai/benchmarks/performance/f39lvtca",
   "tg_tokens": null,
   "tps": 64.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ogdoq5m1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 537w8l1n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 722.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 722.6 42.0 21.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ogdoq5m1",
   "source_url": "https://omlx.ai/benchmarks/performance/ogdoq5m1",
   "tg_tokens": null,
   "tps": 42.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ogdoq5m1-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: 537w8l1n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 699.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 699.3 52.4 23.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ogdoq5m1",
   "source_url": "https://omlx.ai/benchmarks/performance/ogdoq5m1",
   "tg_tokens": null,
   "tps": 52.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-82mz5tim-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: ts1iista; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 907.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 907.9 51.7 21.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/82mz5tim",
   "source_url": "https://omlx.ai/benchmarks/performance/82mz5tim",
   "tg_tokens": null,
   "tps": 51.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-82mz5tim-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: ts1iista; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 706.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 706.5 70.6 22.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/82mz5tim",
   "source_url": "https://omlx.ai/benchmarks/performance/82mz5tim",
   "tg_tokens": null,
   "tps": 70.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-p7rsdbd0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: b4swqawx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 935.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 935.4 49.8 21.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p7rsdbd0",
   "source_url": "https://omlx.ai/benchmarks/performance/p7rsdbd0",
   "tg_tokens": null,
   "tps": 49.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-p7rsdbd0-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: b4swqawx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 846.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 846.7 58.4 22.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p7rsdbd0",
   "source_url": "https://omlx.ai/benchmarks/performance/p7rsdbd0",
   "tg_tokens": null,
   "tps": 58.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-sumy87od-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: 9uu2a6kf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 966.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 966.8 60.2 18.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sumy87od",
   "source_url": "https://omlx.ai/benchmarks/performance/sumy87od",
   "tg_tokens": null,
   "tps": 60.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-sumy87od-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 9uu2a6kf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 819.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 819.1 75.7 19.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sumy87od",
   "source_url": "https://omlx.ai/benchmarks/performance/sumy87od",
   "tg_tokens": null,
   "tps": 75.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-10hx1t7j-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 40.1 GB at this context; duplicate traces of the same session: 8ei6c5m4, m2c7ggeo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 249.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 249.6 41.7 40.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/10hx1t7j",
   "source_url": "https://omlx.ai/benchmarks/performance/10hx1t7j",
   "tg_tokens": null,
   "tps": 41.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-10hx1t7j-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.3 GB at this context; duplicate traces of the same session: 8ei6c5m4, m2c7ggeo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 321.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 321.9 52.9 43.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/10hx1t7j",
   "source_url": "https://omlx.ai/benchmarks/performance/10hx1t7j",
   "tg_tokens": null,
   "tps": 52.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-10hx1t7j-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 44.0 GB at this context; duplicate traces of the same session: 8ei6c5m4, m2c7ggeo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 313.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 313.2 25.1 44.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/10hx1t7j",
   "source_url": "https://omlx.ai/benchmarks/performance/10hx1t7j",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-10hx1t7j-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 45.3 GB at this context; duplicate traces of the same session: 8ei6c5m4, m2c7ggeo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 297.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 297.2 37.8 45.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/10hx1t7j",
   "source_url": "https://omlx.ai/benchmarks/performance/10hx1t7j",
   "tg_tokens": null,
   "tps": 37.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-10hx1t7j-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 47.7 GB at this context; duplicate traces of the same session: 8ei6c5m4, m2c7ggeo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 271.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 271.8 30.0 47.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/10hx1t7j",
   "source_url": "https://omlx.ai/benchmarks/performance/10hx1t7j",
   "tg_tokens": null,
   "tps": 30.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-10hx1t7j-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 51.5 GB at this context; duplicate traces of the same session: 8ei6c5m4, m2c7ggeo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 225.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 225.8 29.1 51.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/10hx1t7j",
   "source_url": "https://omlx.ai/benchmarks/performance/10hx1t7j",
   "tg_tokens": null,
   "tps": 29.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-i4u9m3at-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace, single run at 16k context; peak memory 75.5 GB; TTFT 22502 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 728.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 16k tokens 728.1 PP tok/s 38.8 TG tok/s 22502 TTFT (ms) 75.5 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i4u9m3at",
   "source_url": "https://omlx.ai/benchmarks/performance/i4u9m3at",
   "tg_tokens": null,
   "tps": 38.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-ae42n7kt-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 72.4 GB at this context; duplicate traces of the same session: 6otciasu, ezr4xtd1; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 513.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 1k 513.7 42.1 72.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ae42n7kt",
   "source_url": "https://omlx.ai/benchmarks/performance/ae42n7kt",
   "tg_tokens": null,
   "tps": 42.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-ae42n7kt-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 74.8 GB at this context; duplicate traces of the same session: 6otciasu, ezr4xtd1; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 679.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 4k 679.9 41.3 74.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ae42n7kt",
   "source_url": "https://omlx.ai/benchmarks/performance/ae42n7kt",
   "tg_tokens": null,
   "tps": 41.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-ae42n7kt-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 75.1 GB at this context; duplicate traces of the same session: 6otciasu, ezr4xtd1; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 757.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 8k 757.8 39.2 75.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ae42n7kt",
   "source_url": "https://omlx.ai/benchmarks/performance/ae42n7kt",
   "tg_tokens": null,
   "tps": 39.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-meekn9kc-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 132.5 GB at this context; duplicate traces of the same session: 8dndhmgn, 932v36me, w9f074ju; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1207.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 1k 1,207 91.2 132.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/meekn9kc",
   "source_url": "https://omlx.ai/benchmarks/performance/meekn9kc",
   "tg_tokens": null,
   "tps": 91.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-meekn9kc-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 133.2 GB at this context; duplicate traces of the same session: 8dndhmgn, 932v36me, w9f074ju; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2672.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 4k 2,672 91.0 133.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/meekn9kc",
   "source_url": "https://omlx.ai/benchmarks/performance/meekn9kc",
   "tg_tokens": null,
   "tps": 91.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-meekn9kc-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 137.3 GB at this context; duplicate traces of the same session: 8dndhmgn, 932v36me, w9f074ju; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4979.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 32k 4,979 74.4 137.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/meekn9kc",
   "source_url": "https://omlx.ai/benchmarks/performance/meekn9kc",
   "tg_tokens": null,
   "tps": 74.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-meekn9kc-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 138.3 GB at this context; duplicate traces of the same session: 8dndhmgn, 932v36me, w9f074ju; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5320.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 64k 5,320 102.4 138.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/meekn9kc",
   "source_url": "https://omlx.ai/benchmarks/performance/meekn9kc",
   "tg_tokens": null,
   "tps": 102.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xr0um1gp-ctx1024",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: qa0mc00i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 609.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 609.2 42.7 19.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xr0um1gp",
   "source_url": "https://omlx.ai/benchmarks/performance/xr0um1gp",
   "tg_tokens": null,
   "tps": 42.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xr0um1gp-ctx4096",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: qa0mc00i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 797.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 797.3 41.8 21.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xr0um1gp",
   "source_url": "https://omlx.ai/benchmarks/performance/xr0um1gp",
   "tg_tokens": null,
   "tps": 41.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xn3k31v8-ctx1024",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: lfg2qvvr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 510.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 510.8 45.3 22.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xn3k31v8",
   "source_url": "https://omlx.ai/benchmarks/performance/xn3k31v8",
   "tg_tokens": null,
   "tps": 45.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xn3k31v8-ctx4096",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: lfg2qvvr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 597.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 597.2 39.0 23.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xn3k31v8",
   "source_url": "https://omlx.ai/benchmarks/performance/xn3k31v8",
   "tg_tokens": null,
   "tps": 39.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-f09d2pk2-ctx1024",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: h318ps2y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 535.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 535.6 44.3 22.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f09d2pk2",
   "source_url": "https://omlx.ai/benchmarks/performance/f09d2pk2",
   "tg_tokens": null,
   "tps": 44.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-f09d2pk2-ctx4096",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: h318ps2y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 647.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 647.3 47.4 23.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f09d2pk2",
   "source_url": "https://omlx.ai/benchmarks/performance/f09d2pk2",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-pf7meeuu-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: 6b6g6jvg, izzmsvid, o0u2oly1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1817.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 1,817 107.3 16.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pf7meeuu",
   "source_url": "https://omlx.ai/benchmarks/performance/pf7meeuu",
   "tg_tokens": null,
   "tps": 107.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-pf7meeuu-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: 6b6g6jvg, izzmsvid, o0u2oly1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1898.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 1,898 112.3 16.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pf7meeuu",
   "source_url": "https://omlx.ai/benchmarks/performance/pf7meeuu",
   "tg_tokens": null,
   "tps": 112.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-pf7meeuu-ctx8192",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: 6b6g6jvg, izzmsvid, o0u2oly1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1835.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,835 117.3 16.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pf7meeuu",
   "source_url": "https://omlx.ai/benchmarks/performance/pf7meeuu",
   "tg_tokens": null,
   "tps": 117.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-pf7meeuu-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 6b6g6jvg, izzmsvid, o0u2oly1; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1733.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 1,733 94.9 17.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pf7meeuu",
   "source_url": "https://omlx.ai/benchmarks/performance/pf7meeuu",
   "tg_tokens": null,
   "tps": 94.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-pzelxw5t-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 802.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26b-a4b-it-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 802.2 45.6 25.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pzelxw5t",
   "source_url": "https://omlx.ai/benchmarks/performance/pzelxw5t",
   "tg_tokens": null,
   "tps": 45.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-pzelxw5t-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 784.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26b-a4b-it-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 784.0 41.1 26.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pzelxw5t",
   "source_url": "https://omlx.ai/benchmarks/performance/pzelxw5t",
   "tg_tokens": null,
   "tps": 41.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-pzelxw5t-ctx8192",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 760.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26b-a4b-it-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 760.6 36.6 26.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pzelxw5t",
   "source_url": "https://omlx.ai/benchmarks/performance/pzelxw5t",
   "tg_tokens": null,
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-pzelxw5t-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 712.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26b-a4b-it-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 16k 712.0 30.1 26.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pzelxw5t",
   "source_url": "https://omlx.ai/benchmarks/performance/pzelxw5t",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-dpwm086s-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: 2skbbs79, i1n4qmo9, yi9ivjkh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 801.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26b-a4b-it-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 801.8 45.6 25.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dpwm086s",
   "source_url": "https://omlx.ai/benchmarks/performance/dpwm086s",
   "tg_tokens": null,
   "tps": 45.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-dpwm086s-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: 2skbbs79, i1n4qmo9, yi9ivjkh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 785.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26b-a4b-it-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 785.0 41.2 26.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dpwm086s",
   "source_url": "https://omlx.ai/benchmarks/performance/dpwm086s",
   "tg_tokens": null,
   "tps": 41.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-dpwm086s-ctx8192",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: 2skbbs79, i1n4qmo9, yi9ivjkh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 761.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26b-a4b-it-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 761.2 37.1 26.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dpwm086s",
   "source_url": "https://omlx.ai/benchmarks/performance/dpwm086s",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-dpwm086s-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: 2skbbs79, i1n4qmo9, yi9ivjkh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 717.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26b-a4b-it-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 16k 717.6 30.1 26.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dpwm086s",
   "source_url": "https://omlx.ai/benchmarks/performance/dpwm086s",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-gnv7wm61-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: 0s80n085, 2ihbcc18, a2f0ffcr; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 801.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26b-a4b-it-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 801.7 45.7 25.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gnv7wm61",
   "source_url": "https://omlx.ai/benchmarks/performance/gnv7wm61",
   "tg_tokens": null,
   "tps": 45.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-gnv7wm61-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: 0s80n085, 2ihbcc18, a2f0ffcr; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 783.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26b-a4b-it-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 783.9 41.2 26.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gnv7wm61",
   "source_url": "https://omlx.ai/benchmarks/performance/gnv7wm61",
   "tg_tokens": null,
   "tps": 41.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-gnv7wm61-ctx8192",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: 0s80n085, 2ihbcc18, a2f0ffcr; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 760.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26b-a4b-it-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 760.5 36.6 26.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gnv7wm61",
   "source_url": "https://omlx.ai/benchmarks/performance/gnv7wm61",
   "tg_tokens": null,
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-gnv7wm61-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: 0s80n085, 2ihbcc18, a2f0ffcr; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 717.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26b-a4b-it-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 16k 717.1 30.0 26.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gnv7wm61",
   "source_url": "https://omlx.ai/benchmarks/performance/gnv7wm61",
   "tg_tokens": null,
   "tps": 30.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-r5qbj1r6-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 209.6 GB at this context; duplicate traces of the same session: mft6bplx, qo5dsbsz; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 51.9,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ3e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 3bit; 1k 51.9 20.0 209.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r5qbj1r6",
   "source_url": "https://omlx.ai/benchmarks/performance/r5qbj1r6",
   "tg_tokens": null,
   "tps": 20.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-r5qbj1r6-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 206.9 GB at this context; duplicate traces of the same session: mft6bplx, qo5dsbsz; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 412.3,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ3e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 3bit; 16k 412.3 29.5 206.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r5qbj1r6",
   "source_url": "https://omlx.ai/benchmarks/performance/r5qbj1r6",
   "tg_tokens": null,
   "tps": 29.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-r5qbj1r6-ctx32768",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 206.9 GB at this context; duplicate traces of the same session: mft6bplx, qo5dsbsz; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 517.6,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ3e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 3bit; 32k 517.6 28.2 206.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r5qbj1r6",
   "source_url": "https://omlx.ai/benchmarks/performance/r5qbj1r6",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-accu27uy-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 198.5 GB at this context; duplicate traces of the same session: 2114q6ap, exdcte5v, lzmld39e; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 53.5,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ3e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 3bit; 1k 53.5 18.8 198.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/accu27uy",
   "source_url": "https://omlx.ai/benchmarks/performance/accu27uy",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-accu27uy-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 196.2 GB at this context; duplicate traces of the same session: 2114q6ap, exdcte5v, lzmld39e; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 356.7,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ3e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 3bit; 16k 356.7 29.3 196.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/accu27uy",
   "source_url": "https://omlx.ai/benchmarks/performance/accu27uy",
   "tg_tokens": null,
   "tps": 29.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-accu27uy-ctx32768",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 196.2 GB at this context; duplicate traces of the same session: 2114q6ap, exdcte5v, lzmld39e; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 418.0,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ3e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 3bit; 32k 418.0 27.9 196.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/accu27uy",
   "source_url": "https://omlx.ai/benchmarks/performance/accu27uy",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-accu27uy-ctx65536",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 196.3 GB at this context; duplicate traces of the same session: 2114q6ap, exdcte5v, lzmld39e; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 305.1,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ3e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 3bit; 64k 305.1 27.7 196.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/accu27uy",
   "source_url": "https://omlx.ai/benchmarks/performance/accu27uy",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-1x2tlkkp-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.1 GB at this context; duplicate traces of the same session: xfcabwql, yx5zigak; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 59.7,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ3e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 3bit; 1k 59.7 13.6 180.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1x2tlkkp",
   "source_url": "https://omlx.ai/benchmarks/performance/1x2tlkkp",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-1x2tlkkp-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.0 GB at this context; duplicate traces of the same session: xfcabwql, yx5zigak; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 336.1,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ3e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 3bit; 16k 336.1 15.3 178.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1x2tlkkp",
   "source_url": "https://omlx.ai/benchmarks/performance/1x2tlkkp",
   "tg_tokens": null,
   "tps": 15.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-1x2tlkkp-ctx32768",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.0 GB at this context; duplicate traces of the same session: xfcabwql, yx5zigak; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 358.3,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ3e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 3bit; 32k 358.3 14.0 178.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1x2tlkkp",
   "source_url": "https://omlx.ai/benchmarks/performance/1x2tlkkp",
   "tg_tokens": null,
   "tps": 14.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-wnft55wh-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 187.6 GB at this context; duplicate traces of the same session: pbckw5xv; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 51.4,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ3e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 unknown; 1k 51.4 17.0 187.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wnft55wh",
   "source_url": "https://omlx.ai/benchmarks/performance/wnft55wh",
   "tg_tokens": null,
   "tps": 17.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-wnft55wh-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 185.5 GB at this context; duplicate traces of the same session: pbckw5xv; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 289.7,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ3e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 unknown; 4k 289.7 23.3 185.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wnft55wh",
   "source_url": "https://omlx.ai/benchmarks/performance/wnft55wh",
   "tg_tokens": null,
   "tps": 23.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6upyweos-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 90.8 GB at this context; duplicate traces of the same session: nw5aob7o, pfb5s1st; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 116.9,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 116.9 10.9 90.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6upyweos",
   "source_url": "https://omlx.ai/benchmarks/performance/6upyweos",
   "tg_tokens": null,
   "tps": 10.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-6upyweos-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 94.2 GB at this context; duplicate traces of the same session: nw5aob7o, pfb5s1st; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 531.1,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 531.1 13.3 94.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6upyweos",
   "source_url": "https://omlx.ai/benchmarks/performance/6upyweos",
   "tg_tokens": null,
   "tps": 13.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6upyweos-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 94.3 GB at this context; duplicate traces of the same session: nw5aob7o, pfb5s1st; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 522.4,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 522.4 15.3 94.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6upyweos",
   "source_url": "https://omlx.ai/benchmarks/performance/6upyweos",
   "tg_tokens": null,
   "tps": 15.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-rmj71xzy-ctx4096",
   "model": "Llama 3.3 70B",
   "notes": "oMLX community benchmark trace; peak memory 39.0 GB at this context; duplicate traces of the same session: dc0zrgeu; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 113.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Llama-3.3-70B-Instruct-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 4k 113.5 14.6 39.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rmj71xzy",
   "source_url": "https://omlx.ai/benchmarks/performance/rmj71xzy",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-rmj71xzy-ctx16384",
   "model": "Llama 3.3 70B",
   "notes": "oMLX community benchmark trace; peak memory 42.8 GB at this context; duplicate traces of the same session: dc0zrgeu; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Llama-3.3-70B-Instruct-4bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 128 GB \u00b7 4bit; 16k 102.0 12.2 42.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rmj71xzy",
   "source_url": "https://omlx.ai/benchmarks/performance/rmj71xzy",
   "tg_tokens": null,
   "tps": 12.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-k8bi36v7-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 5gqlyqp6, 6ycoa1kf, 7lqfwm6e, 8zjuoi6s, h5by6715, k0up34k2, ncz0fnv1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2421.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 2,421 64.2 21.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k8bi36v7",
   "source_url": "https://omlx.ai/benchmarks/performance/k8bi36v7",
   "tg_tokens": null,
   "tps": 64.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-k8bi36v7-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 5gqlyqp6, 6ycoa1kf, 7lqfwm6e, 8zjuoi6s, h5by6715, k0up34k2, ncz0fnv1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3086.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 3,086 88.1 22.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k8bi36v7",
   "source_url": "https://omlx.ai/benchmarks/performance/k8bi36v7",
   "tg_tokens": null,
   "tps": 88.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-k8bi36v7-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 5gqlyqp6, 6ycoa1kf, 7lqfwm6e, 8zjuoi6s, h5by6715, k0up34k2, ncz0fnv1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3006.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 3,006 67.9 22.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k8bi36v7",
   "source_url": "https://omlx.ai/benchmarks/performance/k8bi36v7",
   "tg_tokens": null,
   "tps": 67.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-k8bi36v7-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 5gqlyqp6, 6ycoa1kf, 7lqfwm6e, 8zjuoi6s, h5by6715, k0up34k2, ncz0fnv1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2769.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 2,769 97.2 22.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k8bi36v7",
   "source_url": "https://omlx.ai/benchmarks/performance/k8bi36v7",
   "tg_tokens": null,
   "tps": 97.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-k8bi36v7-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 5gqlyqp6, 6ycoa1kf, 7lqfwm6e, 8zjuoi6s, h5by6715, k0up34k2, ncz0fnv1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2349.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 2,349 145.5 22.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k8bi36v7",
   "source_url": "https://omlx.ai/benchmarks/performance/k8bi36v7",
   "tg_tokens": null,
   "tps": 145.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-k8bi36v7-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 5gqlyqp6, 6ycoa1kf, 7lqfwm6e, 8zjuoi6s, h5by6715, k0up34k2, ncz0fnv1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1783.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 1,783 67.3 23.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k8bi36v7",
   "source_url": "https://omlx.ai/benchmarks/performance/k8bi36v7",
   "tg_tokens": null,
   "tps": 67.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-k8bi36v7-ctx131072",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: 5gqlyqp6, 6ycoa1kf, 7lqfwm6e, 8zjuoi6s, h5by6715, k0up34k2, ncz0fnv1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1194.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 128k 1,194 64.7 24.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k8bi36v7",
   "source_url": "https://omlx.ai/benchmarks/performance/k8bi36v7",
   "tg_tokens": null,
   "tps": 64.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-k8bi36v7-ctx199680",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: 5gqlyqp6, 6ycoa1kf, 7lqfwm6e, 8zjuoi6s, h5by6715, k0up34k2, ncz0fnv1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 864.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 195k 864.9 41.7 26.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k8bi36v7",
   "source_url": "https://omlx.ai/benchmarks/performance/k8bi36v7",
   "tg_tokens": null,
   "tps": 41.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-8dhvdhbf-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 24.2 GB; TTFT 193444 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 338.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 64k tokens 338.8 PP tok/s 50.6 TG tok/s 193444 TTFT (ms) 24.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8dhvdhbf",
   "source_url": "https://omlx.ai/benchmarks/performance/8dhvdhbf",
   "tg_tokens": null,
   "tps": 50.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-n08i7b3h-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 23.4 GB; TTFT 67951 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 482.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 32k tokens 482.2 PP tok/s 63.4 TG tok/s 67951 TTFT (ms) 23.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n08i7b3h",
   "source_url": "https://omlx.ai/benchmarks/performance/n08i7b3h",
   "tg_tokens": null,
   "tps": 63.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-04lglquv-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 15.7 GB at this context; duplicate traces of the same session: 2z4f1fir; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 551.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 551.3 30.9 15.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/04lglquv",
   "source_url": "https://omlx.ai/benchmarks/performance/04lglquv",
   "tg_tokens": null,
   "tps": 30.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-04lglquv-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: 2z4f1fir; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 814.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 814.2 30.2 15.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/04lglquv",
   "source_url": "https://omlx.ai/benchmarks/performance/04lglquv",
   "tg_tokens": null,
   "tps": 30.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-k92tx6y1-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: bnr3q302, wdv0yn96; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1247.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 1k 1,247 81.1 21.1 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k92tx6y1",
   "source_url": "https://omlx.ai/benchmarks/performance/k92tx6y1",
   "tg_tokens": null,
   "tps": 81.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-k92tx6y1-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: bnr3q302, wdv0yn96; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1350.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 4k 1,350 80.0 21.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k92tx6y1",
   "source_url": "https://omlx.ai/benchmarks/performance/k92tx6y1",
   "tg_tokens": null,
   "tps": 80.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-k92tx6y1-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: bnr3q302, wdv0yn96; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1282.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 8k 1,282 72.9 21.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k92tx6y1",
   "source_url": "https://omlx.ai/benchmarks/performance/k92tx6y1",
   "tg_tokens": null,
   "tps": 72.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-78rhagr0-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: f0nozt8o, xi4dx3t4; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1163.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 1k 1,163 74.8 20.4 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/78rhagr0",
   "source_url": "https://omlx.ai/benchmarks/performance/78rhagr0",
   "tg_tokens": null,
   "tps": 74.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-78rhagr0-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: f0nozt8o, xi4dx3t4; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1341.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 4k 1,341 92.2 21.1 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/78rhagr0",
   "source_url": "https://omlx.ai/benchmarks/performance/78rhagr0",
   "tg_tokens": null,
   "tps": 92.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-78rhagr0-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: f0nozt8o, xi4dx3t4; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1273.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 8k 1,273 81.5 21.2 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/78rhagr0",
   "source_url": "https://omlx.ai/benchmarks/performance/78rhagr0",
   "tg_tokens": null,
   "tps": 81.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-h1e9svqb-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: c8lbvchw, yi58fi0v; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1252.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 1k 1,252 99.8 19.4 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h1e9svqb",
   "source_url": "https://omlx.ai/benchmarks/performance/h1e9svqb",
   "tg_tokens": null,
   "tps": 99.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-h1e9svqb-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: c8lbvchw, yi58fi0v; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1354.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 4k 1,354 97.1 20.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h1e9svqb",
   "source_url": "https://omlx.ai/benchmarks/performance/h1e9svqb",
   "tg_tokens": null,
   "tps": 97.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-h1e9svqb-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: c8lbvchw, yi58fi0v; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1287.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 8k 1,287 86.5 20.2 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h1e9svqb",
   "source_url": "https://omlx.ai/benchmarks/performance/h1e9svqb",
   "tg_tokens": null,
   "tps": 86.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-dxyx881q-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace, single run at 64k context; peak memory 24.9 GB; TTFT 151758 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 431.8,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 64k tokens 431.8 PP tok/s 31.4 TG tok/s 151758 TTFT (ms) 24.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dxyx881q",
   "source_url": "https://omlx.ai/benchmarks/performance/dxyx881q",
   "tg_tokens": null,
   "tps": 31.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-169val9n-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace, single run at 16k context; peak memory 23.7 GB; TTFT 22169 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 739.1,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k tokens 739.1 PP tok/s 51.1 TG tok/s 22169 TTFT (ms) 23.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/169val9n",
   "source_url": "https://omlx.ai/benchmarks/performance/169val9n",
   "tg_tokens": null,
   "tps": 51.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-6qzbf963-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: h059bxqn; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 146.1,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 1k 146.1 17.4 19.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6qzbf963",
   "source_url": "https://omlx.ai/benchmarks/performance/6qzbf963",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-6qzbf963-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: h059bxqn; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 151.2,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 4k 151.2 26.9 20.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6qzbf963",
   "source_url": "https://omlx.ai/benchmarks/performance/6qzbf963",
   "tg_tokens": null,
   "tps": 26.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-0ef6w3kj-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 40.1 GB at this context; duplicate traces of the same session: 8virlqdf, cpzdmhsw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 249.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 249.6 41.7 40.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ef6w3kj",
   "source_url": "https://omlx.ai/benchmarks/performance/0ef6w3kj",
   "tg_tokens": null,
   "tps": 41.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-0ef6w3kj-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 43.3 GB at this context; duplicate traces of the same session: 8virlqdf, cpzdmhsw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 321.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 321.9 52.9 43.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ef6w3kj",
   "source_url": "https://omlx.ai/benchmarks/performance/0ef6w3kj",
   "tg_tokens": null,
   "tps": 52.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-0ef6w3kj-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 44.0 GB at this context; duplicate traces of the same session: 8virlqdf, cpzdmhsw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 313.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 313.2 25.1 44.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ef6w3kj",
   "source_url": "https://omlx.ai/benchmarks/performance/0ef6w3kj",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-0ef6w3kj-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 45.3 GB at this context; duplicate traces of the same session: 8virlqdf, cpzdmhsw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 297.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 297.2 37.8 45.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ef6w3kj",
   "source_url": "https://omlx.ai/benchmarks/performance/0ef6w3kj",
   "tg_tokens": null,
   "tps": 37.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-0ef6w3kj-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 47.7 GB at this context; duplicate traces of the same session: 8virlqdf, cpzdmhsw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 271.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 271.8 30.0 47.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ef6w3kj",
   "source_url": "https://omlx.ai/benchmarks/performance/0ef6w3kj",
   "tg_tokens": null,
   "tps": 30.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-0ef6w3kj-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 51.5 GB at this context; duplicate traces of the same session: 8virlqdf, cpzdmhsw; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 225.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 225.8 29.1 51.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ef6w3kj",
   "source_url": "https://omlx.ai/benchmarks/performance/0ef6w3kj",
   "tg_tokens": null,
   "tps": 29.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-k0zf2c1s-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 40.5 GB; TTFT 1049 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 976.2,
   "provenance": "community",
   "quant": "GSQ-RCO-Q2",
   "quote": "Qwen3.8-Flash-Next-GSQ-RCO-Q2-mlx on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 976.2 PP tok/s 59.1 TG tok/s 1049 TTFT (ms) 40.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k0zf2c1s",
   "source_url": "https://omlx.ai/benchmarks/performance/k0zf2c1s",
   "tg_tokens": null,
   "tps": 59.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-gye0j8zo-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 37.7 GB; TTFT 10139 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 101.0,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 2bit; 1k tokens 101.0 PP tok/s 18.3 TG tok/s 10139 TTFT (ms) 37.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gye0j8zo",
   "source_url": "https://omlx.ai/benchmarks/performance/gye0j8zo",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-sokdxm8q-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 63.1 GB; TTFT 41070 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 797.9,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 3bit; 32k tokens 797.9 PP tok/s 72.3 TG tok/s 41070 TTFT (ms) 63.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sokdxm8q",
   "source_url": "https://omlx.ai/benchmarks/performance/sokdxm8q",
   "tg_tokens": null,
   "tps": 72.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-ceoj3ykn-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 72.3 GB at this context; duplicate traces of the same session: 5gadm8ha, 5kn46c7d, 87i1lxy6; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 324.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 1k 324.3 28.4 72.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ceoj3ykn",
   "source_url": "https://omlx.ai/benchmarks/performance/ceoj3ykn",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-ceoj3ykn-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 74.8 GB at this context; duplicate traces of the same session: 5gadm8ha, 5kn46c7d, 87i1lxy6; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 400.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 4k 400.7 28.1 74.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ceoj3ykn",
   "source_url": "https://omlx.ai/benchmarks/performance/ceoj3ykn",
   "tg_tokens": null,
   "tps": 28.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-ceoj3ykn-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 75.1 GB at this context; duplicate traces of the same session: 5gadm8ha, 5kn46c7d, 87i1lxy6; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 754.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 8k 754.4 33.2 75.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ceoj3ykn",
   "source_url": "https://omlx.ai/benchmarks/performance/ceoj3ykn",
   "tg_tokens": null,
   "tps": 33.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-ceoj3ykn-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 75.4 GB at this context; duplicate traces of the same session: 5gadm8ha, 5kn46c7d, 87i1lxy6; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 755.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 16k 755.6 37.2 75.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ceoj3ykn",
   "source_url": "https://omlx.ai/benchmarks/performance/ceoj3ykn",
   "tg_tokens": null,
   "tps": 37.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-m2s48xae-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace, single run at 32k context; peak memory 76.2 GB; TTFT 44478 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 736.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 32k tokens 736.7 PP tok/s 45.3 TG tok/s 44478 TTFT (ms) 76.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m2s48xae",
   "source_url": "https://omlx.ai/benchmarks/performance/m2s48xae",
   "tg_tokens": null,
   "tps": 45.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-sv7lt4w7-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 99.1 GB; TTFT 2148 ms; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 476.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 1k tokens 476.8 PP tok/s 65.6 TG tok/s 2148 TTFT (ms) 99.1 Peak mem",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sv7lt4w7",
   "source_url": "https://omlx.ai/benchmarks/performance/sv7lt4w7",
   "tg_tokens": null,
   "tps": 65.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-0p3t0cu3-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: 71m0vlqb; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 347.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 347.7 34.2 14.3 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0p3t0cu3",
   "source_url": "https://omlx.ai/benchmarks/performance/0p3t0cu3",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-0p3t0cu3-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.5 GB at this context; duplicate traces of the same session: 71m0vlqb; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 326.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 326.5 32.9 14.5 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0p3t0cu3",
   "source_url": "https://omlx.ai/benchmarks/performance/0p3t0cu3",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-8thzzmyh-ctx1024",
   "model": "gemma-4-E2B-it",
   "notes": "TurboQuant MLX build; oMLX community benchmark trace; peak memory 5.1 GB at this context; duplicate traces of the same session: i1ek7ero; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2260.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-E2B-it-TurboQuant-MLX-8bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 8bit; 1k 2,260 39.1 5.1 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8thzzmyh",
   "source_url": "https://omlx.ai/benchmarks/performance/8thzzmyh",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-8thzzmyh-ctx4096",
   "model": "gemma-4-E2B-it",
   "notes": "TurboQuant MLX build; oMLX community benchmark trace; peak memory 5.2 GB at this context; duplicate traces of the same session: i1ek7ero; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2352.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-E2B-it-TurboQuant-MLX-8bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 8bit; 4k 2,352 35.7 5.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8thzzmyh",
   "source_url": "https://omlx.ai/benchmarks/performance/8thzzmyh",
   "tg_tokens": null,
   "tps": 35.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-sav24bgy-ctx1024",
   "model": "gemma-4-E2B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 5.1 GB at this context; duplicate traces of the same session: 4f5tv6i5; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2264.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-E2B-it-qat-8bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 8bit; 1k 2,264 38.3 5.1 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sav24bgy",
   "source_url": "https://omlx.ai/benchmarks/performance/sav24bgy",
   "tg_tokens": null,
   "tps": 38.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-sav24bgy-ctx4096",
   "model": "gemma-4-E2B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 5.1 GB at this context; duplicate traces of the same session: 4f5tv6i5; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2343.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-E2B-it-qat-8bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 8bit; 4k 2,343 37.6 5.1 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sav24bgy",
   "source_url": "https://omlx.ai/benchmarks/performance/sav24bgy",
   "tg_tokens": null,
   "tps": 37.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-wdqaei8e-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 187.7 GB at this context; duplicate traces of the same session: 79kjcxp2, aw23xy34; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 52.3,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ3e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 3bit; 1k 52.3 16.1 187.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wdqaei8e",
   "source_url": "https://omlx.ai/benchmarks/performance/wdqaei8e",
   "tg_tokens": null,
   "tps": 16.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-wdqaei8e-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 185.5 GB at this context; duplicate traces of the same session: 79kjcxp2, aw23xy34; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 287.8,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ3e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 3bit; 16k 287.8 30.4 185.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wdqaei8e",
   "source_url": "https://omlx.ai/benchmarks/performance/wdqaei8e",
   "tg_tokens": null,
   "tps": 30.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-wdqaei8e-ctx32768",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 185.5 GB at this context; duplicate traces of the same session: 79kjcxp2, aw23xy34; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 320.9,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ3e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 3bit; 32k 320.9 22.2 185.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wdqaei8e",
   "source_url": "https://omlx.ai/benchmarks/performance/wdqaei8e",
   "tg_tokens": null,
   "tps": 22.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-eweii6es-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 403.4 GB at this context; duplicate traces of the same session: fnk420vm; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 564.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 564.3 27.4 403.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eweii6es",
   "source_url": "https://omlx.ai/benchmarks/performance/eweii6es",
   "tg_tokens": null,
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-eweii6es-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 403.7 GB at this context; duplicate traces of the same session: fnk420vm; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 723.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 723.7 31.6 403.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eweii6es",
   "source_url": "https://omlx.ai/benchmarks/performance/eweii6es",
   "tg_tokens": null,
   "tps": 31.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-k0d9dq6r-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 403.3 GB at this context; duplicate traces of the same session: 9s0qmobf; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 564.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 564.7 29.4 403.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k0d9dq6r",
   "source_url": "https://omlx.ai/benchmarks/performance/k0d9dq6r",
   "tg_tokens": null,
   "tps": 29.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-k0d9dq6r-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 403.8 GB at this context; duplicate traces of the same session: 9s0qmobf; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 713.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 713.4 29.4 403.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k0d9dq6r",
   "source_url": "https://omlx.ai/benchmarks/performance/k0d9dq6r",
   "tg_tokens": null,
   "tps": 29.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-k0d9dq6r-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 403.7 GB at this context; duplicate traces of the same session: 9s0qmobf; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 736.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 16k 736.6 35.9 403.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k0d9dq6r",
   "source_url": "https://omlx.ai/benchmarks/performance/k0d9dq6r",
   "tg_tokens": null,
   "tps": 35.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-k09af1ad-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.7 GB at this context; duplicate traces of the same session: 3bzppuvl, a5xcrdzc, vmzr5dtf; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 232.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 232.7 21.0 5.7 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k09af1ad",
   "source_url": "https://omlx.ai/benchmarks/performance/k09af1ad",
   "tg_tokens": null,
   "tps": 21.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-k09af1ad-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: 3bzppuvl, a5xcrdzc, vmzr5dtf; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 241.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 241.9 20.5 6.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k09af1ad",
   "source_url": "https://omlx.ai/benchmarks/performance/k09af1ad",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-k09af1ad-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: 3bzppuvl, a5xcrdzc, vmzr5dtf; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 240.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 8k 240.0 20.1 6.4 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k09af1ad",
   "source_url": "https://omlx.ai/benchmarks/performance/k09af1ad",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-k09af1ad-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: 3bzppuvl, a5xcrdzc, vmzr5dtf; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 231.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 16k 231.0 18.8 6.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k09af1ad",
   "source_url": "https://omlx.ai/benchmarks/performance/k09af1ad",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M2 Max",
   "id": "omlx-wcd9cqi4-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 21.6 GB; TTFT 28058 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 583.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 16k tokens 583.9 PP tok/s 22.8 TG tok/s 28058 TTFT (ms) 21.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wcd9cqi4",
   "source_url": "https://omlx.ai/benchmarks/performance/wcd9cqi4",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-vommepnr-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 41qle4zw, jo6dxuef, n85jh4kf; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2076.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 2,076 66.0 20.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vommepnr",
   "source_url": "https://omlx.ai/benchmarks/performance/vommepnr",
   "tg_tokens": null,
   "tps": 66.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-vommepnr-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 41qle4zw, jo6dxuef, n85jh4kf; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1924.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 1,924 62.5 20.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vommepnr",
   "source_url": "https://omlx.ai/benchmarks/performance/vommepnr",
   "tg_tokens": null,
   "tps": 62.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-vommepnr-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: 41qle4zw, jo6dxuef, n85jh4kf; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1153.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 64k 1,153 34.6 21.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vommepnr",
   "source_url": "https://omlx.ai/benchmarks/performance/vommepnr",
   "tg_tokens": null,
   "tps": 34.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-vommepnr-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 41qle4zw, jo6dxuef, n85jh4kf; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 728.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 128k 728.9 19.9 22.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vommepnr",
   "source_url": "https://omlx.ai/benchmarks/performance/vommepnr",
   "tg_tokens": null,
   "tps": 19.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-0yqyhaab-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace, single run at 16k context; peak memory 25.3 GB; TTFT 23121 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 708.6,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k tokens 708.6 PP tok/s 30.0 TG tok/s 23121 TTFT (ms) 25.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0yqyhaab",
   "source_url": "https://omlx.ai/benchmarks/performance/0yqyhaab",
   "tg_tokens": null,
   "tps": 30.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-7jj8bzmq-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace, single run at 16k context; peak memory 23.7 GB; TTFT 22077 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 742.1,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k tokens 742.1 PP tok/s 59.9 TG tok/s 22077 TTFT (ms) 23.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7jj8bzmq",
   "source_url": "https://omlx.ai/benchmarks/performance/7jj8bzmq",
   "tg_tokens": null,
   "tps": 59.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-d2e9ux0a-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace, single run at 16k context; peak memory 23.7 GB; TTFT 22109 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 741.1,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-oQ4-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 16k tokens 741.1 PP tok/s 50.3 TG tok/s 22109 TTFT (ms) 23.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d2e9ux0a",
   "source_url": "https://omlx.ai/benchmarks/performance/d2e9ux0a",
   "tg_tokens": null,
   "tps": 50.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-pcybleuv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: l69ajfiu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 99.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 99.5 15.9 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pcybleuv",
   "source_url": "https://omlx.ai/benchmarks/performance/pcybleuv",
   "tg_tokens": null,
   "tps": 15.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-pcybleuv-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: l69ajfiu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 93.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 93.2 20.1 25.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pcybleuv",
   "source_url": "https://omlx.ai/benchmarks/performance/pcybleuv",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-yt0xtp0c-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 5fdctly3, zwl7r0pr; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 127.7 15.1 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yt0xtp0c",
   "source_url": "https://omlx.ai/benchmarks/performance/yt0xtp0c",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-yt0xtp0c-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 5fdctly3, zwl7r0pr; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 122.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 122.6 14.2 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yt0xtp0c",
   "source_url": "https://omlx.ai/benchmarks/performance/yt0xtp0c",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-yt0xtp0c-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 5fdctly3, zwl7r0pr; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 106.4 12.6 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yt0xtp0c",
   "source_url": "https://omlx.ai/benchmarks/performance/yt0xtp0c",
   "tg_tokens": null,
   "tps": 12.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ddoqbe72-ctx1024",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 7xvq7q5f, acgirepl, i4elkivd, ryzrlkgi, s0obb87q, wh08q4bl, wj667i7o; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 429.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 429.1 34.0 17.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ddoqbe72",
   "source_url": "https://omlx.ai/benchmarks/performance/ddoqbe72",
   "tg_tokens": null,
   "tps": 34.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ddoqbe72-ctx4096",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: 7xvq7q5f, acgirepl, i4elkivd, ryzrlkgi, s0obb87q, wh08q4bl, wj667i7o; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 471.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 471.0 36.8 18.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ddoqbe72",
   "source_url": "https://omlx.ai/benchmarks/performance/ddoqbe72",
   "tg_tokens": null,
   "tps": 36.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ddoqbe72-ctx8192",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 7xvq7q5f, acgirepl, i4elkivd, ryzrlkgi, s0obb87q, wh08q4bl, wj667i7o; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 463.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 463.0 33.2 19.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ddoqbe72",
   "source_url": "https://omlx.ai/benchmarks/performance/ddoqbe72",
   "tg_tokens": null,
   "tps": 33.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ddoqbe72-ctx16384",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 7xvq7q5f, acgirepl, i4elkivd, ryzrlkgi, s0obb87q, wh08q4bl, wj667i7o; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 444.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 444.0 36.2 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ddoqbe72",
   "source_url": "https://omlx.ai/benchmarks/performance/ddoqbe72",
   "tg_tokens": null,
   "tps": 36.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "outlier"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ddoqbe72-ctx32768",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 7xvq7q5f, acgirepl, i4elkivd, ryzrlkgi, s0obb87q, wh08q4bl, wj667i7o; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 418.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 418.4 80.8 20.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ddoqbe72",
   "source_url": "https://omlx.ai/benchmarks/performance/ddoqbe72",
   "tg_tokens": null,
   "tps": 80.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ddoqbe72-ctx65536",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: 7xvq7q5f, acgirepl, i4elkivd, ryzrlkgi, s0obb87q, wh08q4bl, wj667i7o; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 378.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 378.2 30.0 23.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ddoqbe72",
   "source_url": "https://omlx.ai/benchmarks/performance/ddoqbe72",
   "tg_tokens": null,
   "tps": 30.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ddoqbe72-ctx131072",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: 7xvq7q5f, acgirepl, i4elkivd, ryzrlkgi, s0obb87q, wh08q4bl, wj667i7o; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 303.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 128k 303.7 25.5 27.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ddoqbe72",
   "source_url": "https://omlx.ai/benchmarks/performance/ddoqbe72",
   "tg_tokens": null,
   "tps": 25.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ddoqbe72-ctx199680",
   "model": "Qwen3.8-27B-Huihui-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 33.1 GB at this context; duplicate traces of the same session: 7xvq7q5f, acgirepl, i4elkivd, ryzrlkgi, s0obb87q, wh08q4bl, wj667i7o; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 244.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Huihui-Abliterated-oQ4e-MTP-MLX on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 195k 244.8 18.2 33.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ddoqbe72",
   "source_url": "https://omlx.ai/benchmarks/performance/ddoqbe72",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hy1spa3l-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: coupnn9v; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 896.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 896.8 51.1 17.6 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hy1spa3l",
   "source_url": "https://omlx.ai/benchmarks/performance/hy1spa3l",
   "tg_tokens": null,
   "tps": 51.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hy1spa3l-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: coupnn9v; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1052.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 1,052 47.4 19.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hy1spa3l",
   "source_url": "https://omlx.ai/benchmarks/performance/hy1spa3l",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wl64x6y0-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: 6jvv68g7, 8qopvvwe, udd2xzbq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 885.3,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 885.3 25.6 16.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wl64x6y0",
   "source_url": "https://omlx.ai/benchmarks/performance/wl64x6y0",
   "tg_tokens": null,
   "tps": 25.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wl64x6y0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 6jvv68g7, 8qopvvwe, udd2xzbq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 741.8,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 741.8 27.7 17.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wl64x6y0",
   "source_url": "https://omlx.ai/benchmarks/performance/wl64x6y0",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wl64x6y0-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 6jvv68g7, 8qopvvwe, udd2xzbq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 683.3,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 683.3 30.4 17.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wl64x6y0",
   "source_url": "https://omlx.ai/benchmarks/performance/wl64x6y0",
   "tg_tokens": null,
   "tps": 30.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wl64x6y0-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: 6jvv68g7, 8qopvvwe, udd2xzbq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 643.8,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 643.8 23.9 18.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wl64x6y0",
   "source_url": "https://omlx.ai/benchmarks/performance/wl64x6y0",
   "tg_tokens": null,
   "tps": 23.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-obsmsmc1-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.3 GB at this context; duplicate traces of the same session: hgg5n7x8, nvrrhtr9, x6f5414i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 793.9,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 793.9 23.9 16.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/obsmsmc1",
   "source_url": "https://omlx.ai/benchmarks/performance/obsmsmc1",
   "tg_tokens": null,
   "tps": 23.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-obsmsmc1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: hgg5n7x8, nvrrhtr9, x6f5414i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 723.7,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 723.7 29.1 17.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/obsmsmc1",
   "source_url": "https://omlx.ai/benchmarks/performance/obsmsmc1",
   "tg_tokens": null,
   "tps": 29.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-obsmsmc1-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: hgg5n7x8, nvrrhtr9, x6f5414i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 723.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 723.0 30.8 17.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/obsmsmc1",
   "source_url": "https://omlx.ai/benchmarks/performance/obsmsmc1",
   "tg_tokens": null,
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-obsmsmc1-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: hgg5n7x8, nvrrhtr9, x6f5414i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 670.6,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.8-27B-mxfp4 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 670.6 25.7 18.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/obsmsmc1",
   "source_url": "https://omlx.ai/benchmarks/performance/obsmsmc1",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-36hvwn7x-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: e9xfx5vi, gd5zx7wn, nzuzfgcc, smeecr4f, t2pt360z; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 473.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 473.2 32.1 17.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/36hvwn7x",
   "source_url": "https://omlx.ai/benchmarks/performance/36hvwn7x",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-36hvwn7x-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: e9xfx5vi, gd5zx7wn, nzuzfgcc, smeecr4f, t2pt360z; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 482.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 482.4 34.7 18.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/36hvwn7x",
   "source_url": "https://omlx.ai/benchmarks/performance/36hvwn7x",
   "tg_tokens": null,
   "tps": 34.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-36hvwn7x-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: e9xfx5vi, gd5zx7wn, nzuzfgcc, smeecr4f, t2pt360z; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 478.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 478.2 40.1 19.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/36hvwn7x",
   "source_url": "https://omlx.ai/benchmarks/performance/36hvwn7x",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-36hvwn7x-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: e9xfx5vi, gd5zx7wn, nzuzfgcc, smeecr4f, t2pt360z; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 458.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 458.3 51.4 19.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/36hvwn7x",
   "source_url": "https://omlx.ai/benchmarks/performance/36hvwn7x",
   "tg_tokens": null,
   "tps": 51.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-36hvwn7x-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: e9xfx5vi, gd5zx7wn, nzuzfgcc, smeecr4f, t2pt360z; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 430.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 430.5 76.3 20.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/36hvwn7x",
   "source_url": "https://omlx.ai/benchmarks/performance/36hvwn7x",
   "tg_tokens": null,
   "tps": 76.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-36hvwn7x-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: e9xfx5vi, gd5zx7wn, nzuzfgcc, smeecr4f, t2pt360z; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 376.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 376.8 30.8 23.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/36hvwn7x",
   "source_url": "https://omlx.ai/benchmarks/performance/36hvwn7x",
   "tg_tokens": null,
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-qb9u694v-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 195k context; peak memory 69.0 GB; TTFT 255843 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 781.7,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 3bit; 195k tokens 781.7 PP tok/s 50.4 TG tok/s 255843 TTFT (ms) 69.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qb9u694v",
   "source_url": "https://omlx.ai/benchmarks/performance/qb9u694v",
   "tg_tokens": null,
   "tps": 50.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-oodxpbas-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.5 GB at this context; duplicate traces of the same session: 3r3jc8f2, alp54fje, b5z3k4c9, no3mad2w, wxq2ce42, yqwvpvyl; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 336.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 1k 336.3 54.0 70.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oodxpbas",
   "source_url": "https://omlx.ai/benchmarks/performance/oodxpbas",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Ultra",
   "id": "omlx-oodxpbas-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: 3r3jc8f2, alp54fje, b5z3k4c9, no3mad2w, wxq2ce42, yqwvpvyl; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 776.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 8k 776.2 60.0 71.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oodxpbas",
   "source_url": "https://omlx.ai/benchmarks/performance/oodxpbas",
   "tg_tokens": null,
   "tps": 60.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-oodxpbas-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.9 GB at this context; duplicate traces of the same session: 3r3jc8f2, alp54fje, b5z3k4c9, no3mad2w, wxq2ce42, yqwvpvyl; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 780.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 16k 780.7 71.6 71.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oodxpbas",
   "source_url": "https://omlx.ai/benchmarks/performance/oodxpbas",
   "tg_tokens": null,
   "tps": 71.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-oodxpbas-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.5 GB at this context; duplicate traces of the same session: 3r3jc8f2, alp54fje, b5z3k4c9, no3mad2w, wxq2ce42, yqwvpvyl; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 778.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 32k 778.2 60.1 72.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oodxpbas",
   "source_url": "https://omlx.ai/benchmarks/performance/oodxpbas",
   "tg_tokens": null,
   "tps": 60.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-oodxpbas-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.7 GB at this context; duplicate traces of the same session: 3r3jc8f2, alp54fje, b5z3k4c9, no3mad2w, wxq2ce42, yqwvpvyl; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 763.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 64k 763.1 55.8 73.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oodxpbas",
   "source_url": "https://omlx.ai/benchmarks/performance/oodxpbas",
   "tg_tokens": null,
   "tps": 55.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-oodxpbas-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 76.1 GB at this context; duplicate traces of the same session: 3r3jc8f2, alp54fje, b5z3k4c9, no3mad2w, wxq2ce42, yqwvpvyl; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 749.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 128k 749.7 51.1 76.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oodxpbas",
   "source_url": "https://omlx.ai/benchmarks/performance/oodxpbas",
   "tg_tokens": null,
   "tps": 51.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-oodxpbas-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.9 GB at this context; duplicate traces of the same session: 3r3jc8f2, alp54fje, b5z3k4c9, no3mad2w, wxq2ce42, yqwvpvyl; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 737.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 128 GB \u00b7 4bit; 195k 737.2 47.9 78.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oodxpbas",
   "source_url": "https://omlx.ai/benchmarks/performance/oodxpbas",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-o80296ic-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 141.1 GB; TTFT 25070 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5228.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 8bit; 128k tokens 5,228 PP tok/s 96.3 TG tok/s 25070 TTFT (ms) 141.1 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o80296ic",
   "source_url": "https://omlx.ai/benchmarks/performance/o80296ic",
   "tg_tokens": null,
   "tps": 96.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-e07e84j8-ctx1024",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 645.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 645.0 15.4 32.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e07e84j8",
   "source_url": "https://omlx.ai/benchmarks/performance/e07e84j8",
   "tg_tokens": null,
   "tps": 15.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-e07e84j8-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 33.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 699.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 699.9 15.0 33.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e07e84j8",
   "source_url": "https://omlx.ai/benchmarks/performance/e07e84j8",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-e07e84j8-ctx8192",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 33.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 687.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 687.4 14.5 33.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e07e84j8",
   "source_url": "https://omlx.ai/benchmarks/performance/e07e84j8",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-e07e84j8-ctx16384",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 662.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 662.8 13.6 34.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e07e84j8",
   "source_url": "https://omlx.ai/benchmarks/performance/e07e84j8",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-2jqyzis6-ctx1024",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 5.1 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2099.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-E2B-it-MLX-8bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 8bit; 1k 2,099 39.0 5.1 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2jqyzis6",
   "source_url": "https://omlx.ai/benchmarks/performance/2jqyzis6",
   "tg_tokens": null,
   "tps": 39.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-2jqyzis6-ctx4096",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 5.1 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2318.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-E2B-it-MLX-8bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 8bit; 4k 2,318 38.5 5.1 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2jqyzis6",
   "source_url": "https://omlx.ai/benchmarks/performance/2jqyzis6",
   "tg_tokens": null,
   "tps": 38.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-iqtctv2o-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 6.0 GB; TTFT 1663 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 615.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E4B-it-MLX-4bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k tokens 615.9 PP tok/s 22.9 TG tok/s 1663 TTFT (ms) 6.0 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iqtctv2o",
   "source_url": "https://omlx.ai/benchmarks/performance/iqtctv2o",
   "tg_tokens": null,
   "tps": 22.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9soidczi-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 403.3 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 564.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 564.7 29.4 403.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9soidczi",
   "source_url": "https://omlx.ai/benchmarks/performance/9soidczi",
   "tg_tokens": null,
   "tps": 29.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9soidczi-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 403.8 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 713.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 713.4 29.4 403.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9soidczi",
   "source_url": "https://omlx.ai/benchmarks/performance/9soidczi",
   "tg_tokens": null,
   "tps": 29.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-9soidczi-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 403.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 736.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 16k 736.6 35.9 403.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9soidczi",
   "source_url": "https://omlx.ai/benchmarks/performance/9soidczi",
   "tg_tokens": null,
   "tps": 35.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-0c658f2a-ctx1024",
   "model": "MiniCPM5-2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 3.1 GB at this context; duplicate traces of the same session: f0bnb1xv; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 756.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "MiniCPM5-2B-MLX-8bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 8bit; 1k 756.9 39.0 3.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0c658f2a",
   "source_url": "https://omlx.ai/benchmarks/performance/0c658f2a",
   "tg_tokens": null,
   "tps": 39.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-0c658f2a-ctx4096",
   "model": "MiniCPM5-2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 3.3 GB at this context; duplicate traces of the same session: f0bnb1xv; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 700.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "MiniCPM5-2B-MLX-8bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 8bit; 4k 700.5 35.5 3.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0c658f2a",
   "source_url": "https://omlx.ai/benchmarks/performance/0c658f2a",
   "tg_tokens": null,
   "tps": 35.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-sj27p2aq-ctx1024",
   "model": "Muse-Spark-Open-816B-A42B",
   "notes": "oMLX community benchmark trace; peak memory 341.7 GB at this context; duplicate traces of the same session: ub9sqdla; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 215.3,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Muse-Spark-Open-816B-A42B-oQ3e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 3bit; 1k 215.3 21.3 341.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sj27p2aq",
   "source_url": "https://omlx.ai/benchmarks/performance/sj27p2aq",
   "tg_tokens": null,
   "tps": 21.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-sj27p2aq-ctx16384",
   "model": "Muse-Spark-Open-816B-A42B",
   "notes": "oMLX community benchmark trace; peak memory 343.5 GB at this context; duplicate traces of the same session: ub9sqdla; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 219.7,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Muse-Spark-Open-816B-A42B-oQ3e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 3bit; 16k 219.7 19.4 343.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sj27p2aq",
   "source_url": "https://omlx.ai/benchmarks/performance/sj27p2aq",
   "tg_tokens": null,
   "tps": 19.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-0pgvdcgn-ctx1024",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "MLX; oMLX community benchmark trace; peak memory 30.8 GB at this context; duplicate traces of the same session: opzelgyu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2489.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 2,489 65.5 30.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0pgvdcgn",
   "source_url": "https://omlx.ai/benchmarks/performance/0pgvdcgn",
   "tg_tokens": null,
   "tps": 65.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-0pgvdcgn-ctx4096",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "MLX; oMLX community benchmark trace; peak memory 31.3 GB at this context; duplicate traces of the same session: opzelgyu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2989.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 2,989 59.4 31.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0pgvdcgn",
   "source_url": "https://omlx.ai/benchmarks/performance/0pgvdcgn",
   "tg_tokens": null,
   "tps": 59.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-c9xutigk-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.7 GB at this context; duplicate traces of the same session: ktanrukv; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 232.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 232.2 20.8 5.7 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c9xutigk",
   "source_url": "https://omlx.ai/benchmarks/performance/c9xutigk",
   "tg_tokens": null,
   "tps": 20.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-c9xutigk-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: ktanrukv; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 241.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 241.1 20.4 6.2 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c9xutigk",
   "source_url": "https://omlx.ai/benchmarks/performance/c9xutigk",
   "tg_tokens": null,
   "tps": 20.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-77wzkgyt-ctx131072",
   "model": "Qwen3.6-35B-A3B-Holo3-Qwopus-Instruct",
   "notes": "MLX build; oMLX community benchmark trace, single run at 128k context; peak memory 32.3 GB; TTFT 55611 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2357.0,
   "provenance": "community",
   "quant": "qx64-hi",
   "quote": "Qwen3.6-35B-A3B-Holo3-Qwopus-Instruct-qx64-hi-mlx on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 128k tokens 2,357 PP tok/s 71.2 TG tok/s 55611 TTFT (ms) 32.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/77wzkgyt",
   "source_url": "https://omlx.ai/benchmarks/performance/77wzkgyt",
   "tg_tokens": null,
   "tps": 71.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ts5jqta1-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: hvk2yl08; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 82.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 82.8 13.3 18.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ts5jqta1",
   "source_url": "https://omlx.ai/benchmarks/performance/ts5jqta1",
   "tg_tokens": null,
   "tps": 13.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ts5jqta1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: hvk2yl08; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 172.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 172.6 14.2 21.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ts5jqta1",
   "source_url": "https://omlx.ai/benchmarks/performance/ts5jqta1",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-owz9kxur-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 0l0slv9m, kl4hipcw, lw2or0hj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 941.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 941.1 33.4 16.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/owz9kxur",
   "source_url": "https://omlx.ai/benchmarks/performance/owz9kxur",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-owz9kxur-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 0l0slv9m, kl4hipcw, lw2or0hj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1017.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,017 32.4 17.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/owz9kxur",
   "source_url": "https://omlx.ai/benchmarks/performance/owz9kxur",
   "tg_tokens": null,
   "tps": 32.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-owz9kxur-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 0l0slv9m, kl4hipcw, lw2or0hj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1016.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,016 32.0 17.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/owz9kxur",
   "source_url": "https://omlx.ai/benchmarks/performance/owz9kxur",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-owz9kxur-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 0l0slv9m, kl4hipcw, lw2or0hj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 986.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 986.3 31.0 17.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/owz9kxur",
   "source_url": "https://omlx.ai/benchmarks/performance/owz9kxur",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7pkwsh50-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 38w3lhwo, bjy4yjrn, vkien0lr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 938.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 938.9 33.5 16.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7pkwsh50",
   "source_url": "https://omlx.ai/benchmarks/performance/7pkwsh50",
   "tg_tokens": null,
   "tps": 33.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7pkwsh50-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 38w3lhwo, bjy4yjrn, vkien0lr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1028.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,028 32.9 17.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7pkwsh50",
   "source_url": "https://omlx.ai/benchmarks/performance/7pkwsh50",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7pkwsh50-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 38w3lhwo, bjy4yjrn, vkien0lr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1024.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,024 32.2 17.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7pkwsh50",
   "source_url": "https://omlx.ai/benchmarks/performance/7pkwsh50",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7pkwsh50-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 38w3lhwo, bjy4yjrn, vkien0lr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 995.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 995.0 31.1 17.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7pkwsh50",
   "source_url": "https://omlx.ai/benchmarks/performance/7pkwsh50",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fng7xx87-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: dq1l6ail, jccd954q, sypywnhu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 836.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 836.8 19.1 28.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fng7xx87",
   "source_url": "https://omlx.ai/benchmarks/performance/fng7xx87",
   "tg_tokens": null,
   "tps": 19.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fng7xx87-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: dq1l6ail, jccd954q, sypywnhu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 966.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 966.0 19.0 29.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fng7xx87",
   "source_url": "https://omlx.ai/benchmarks/performance/fng7xx87",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fng7xx87-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.8 GB at this context; duplicate traces of the same session: dq1l6ail, jccd954q, sypywnhu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 966.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 966.9 18.1 29.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fng7xx87",
   "source_url": "https://omlx.ai/benchmarks/performance/fng7xx87",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fng7xx87-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: dq1l6ail, jccd954q, sypywnhu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 942.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 942.1 17.6 30.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fng7xx87",
   "source_url": "https://omlx.ai/benchmarks/performance/fng7xx87",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-6ls9434x-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 455.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 4k 455.2 26.6 26.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ls9434x",
   "source_url": "https://omlx.ai/benchmarks/performance/6ls9434x",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-6ls9434x-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 436.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 16k 436.8 25.2 27.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ls9434x",
   "source_url": "https://omlx.ai/benchmarks/performance/6ls9434x",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-6ls9434x-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 286.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 128k 286.7 18.2 35.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ls9434x",
   "source_url": "https://omlx.ai/benchmarks/performance/6ls9434x",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-k9oft1a6-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 29.8 GB; TTFT 43219 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 758.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 6bit; 32k tokens 758.2 PP tok/s 75.7 TG tok/s 43219 TTFT (ms) 29.8 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k9oft1a6",
   "source_url": "https://omlx.ai/benchmarks/performance/k9oft1a6",
   "tg_tokens": null,
   "tps": 75.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-82cs0dtp-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 29.8 GB; TTFT 43305 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 756.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 6bit; 32k tokens 756.7 PP tok/s 66.8 TG tok/s 43305 TTFT (ms) 29.8 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/82cs0dtp",
   "source_url": "https://omlx.ai/benchmarks/performance/82cs0dtp",
   "tg_tokens": null,
   "tps": 66.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5ib568mk-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 29.8 GB; TTFT 44867 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 730.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 6bit; 32k tokens 730.3 PP tok/s 79.1 TG tok/s 44867 TTFT (ms) 29.8 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ib568mk",
   "source_url": "https://omlx.ai/benchmarks/performance/5ib568mk",
   "tg_tokens": null,
   "tps": 79.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vgtgx6ia-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: omb7s5z2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 886.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 886.0 44.4 27.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vgtgx6ia",
   "source_url": "https://omlx.ai/benchmarks/performance/vgtgx6ia",
   "tg_tokens": null,
   "tps": 44.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vgtgx6ia-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: omb7s5z2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 761.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k 761.5 57.4 28.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vgtgx6ia",
   "source_url": "https://omlx.ai/benchmarks/performance/vgtgx6ia",
   "tg_tokens": null,
   "tps": 57.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-xh4txjl3-ctx1024",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: pgx7buoy; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2727.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-35B-A3B-Distill-mlx-4Bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,727 109.8 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xh4txjl3",
   "source_url": "https://omlx.ai/benchmarks/performance/xh4txjl3",
   "tg_tokens": null,
   "tps": 109.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-xh4txjl3-ctx4096",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: pgx7buoy; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3483.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-35B-A3B-Distill-mlx-4Bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,483 106.2 20.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xh4txjl3",
   "source_url": "https://omlx.ai/benchmarks/performance/xh4txjl3",
   "tg_tokens": null,
   "tps": 106.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-6wbepx2n-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.7 GB at this context; duplicate traces of the same session: bb8fy3a6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 119.1,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 2bit; 1k 119.1 25.0 37.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6wbepx2n",
   "source_url": "https://omlx.ai/benchmarks/performance/6wbepx2n",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-6wbepx2n-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.8 GB at this context; duplicate traces of the same session: bb8fy3a6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 417.3,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 2bit; 4k 417.3 27.2 38.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6wbepx2n",
   "source_url": "https://omlx.ai/benchmarks/performance/6wbepx2n",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-sf2blgk9-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 40.3 GB; TTFT 29303 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 559.1,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 2bit; 16k tokens 559.1 PP tok/s 14.6 TG tok/s 29303 TTFT (ms) 40.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sf2blgk9",
   "source_url": "https://omlx.ai/benchmarks/performance/sf2blgk9",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-vwk0vrht-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 66.4 GB; TTFT 165446 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 792.2,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-Flash-Next-oQ3.5e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 3bit; 128k tokens 792.2 PP tok/s 54.0 TG tok/s 165446 TTFT (ms) 66.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vwk0vrht",
   "source_url": "https://omlx.ai/benchmarks/performance/vwk0vrht",
   "tg_tokens": null,
   "tps": 54.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-gfphglvt-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 12.0 GB at this context; duplicate traces of the same session: 0b52joun, 9tdpwjxh, benbuosb, fpl7pj9y, q1wwl3zt, s963ky6d; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 90.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k 90.3 6.0 12.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gfphglvt",
   "source_url": "https://omlx.ai/benchmarks/performance/gfphglvt",
   "tg_tokens": null,
   "tps": 6.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-gfphglvt-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 12.5 GB at this context; duplicate traces of the same session: 0b52joun, 9tdpwjxh, benbuosb, fpl7pj9y, q1wwl3zt, s963ky6d; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 151.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 151.4 6.2 12.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gfphglvt",
   "source_url": "https://omlx.ai/benchmarks/performance/gfphglvt",
   "tg_tokens": null,
   "tps": 6.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-gfphglvt-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 12.5 GB at this context; duplicate traces of the same session: 0b52joun, 9tdpwjxh, benbuosb, fpl7pj9y, q1wwl3zt, s963ky6d; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 145.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 8k 145.0 6.4 12.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gfphglvt",
   "source_url": "https://omlx.ai/benchmarks/performance/gfphglvt",
   "tg_tokens": null,
   "tps": 6.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-gfphglvt-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 13.3 GB at this context; duplicate traces of the same session: 0b52joun, 9tdpwjxh, benbuosb, fpl7pj9y, q1wwl3zt, s963ky6d; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 139.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 32k 139.7 5.3 13.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gfphglvt",
   "source_url": "https://omlx.ai/benchmarks/performance/gfphglvt",
   "tg_tokens": null,
   "tps": 5.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-gfphglvt-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: 0b52joun, 9tdpwjxh, benbuosb, fpl7pj9y, q1wwl3zt, s963ky6d; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 129.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 64k 129.2 5.3 14.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gfphglvt",
   "source_url": "https://omlx.ai/benchmarks/performance/gfphglvt",
   "tg_tokens": null,
   "tps": 5.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-gfphglvt-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 0b52joun, 9tdpwjxh, benbuosb, fpl7pj9y, q1wwl3zt, s963ky6d; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 128k 130.6 4.9 17.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gfphglvt",
   "source_url": "https://omlx.ai/benchmarks/performance/gfphglvt",
   "tg_tokens": null,
   "tps": 4.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-gfphglvt-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: 0b52joun, 9tdpwjxh, benbuosb, fpl7pj9y, q1wwl3zt, s963ky6d; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 126.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 195k 126.9 5.2 19.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gfphglvt",
   "source_url": "https://omlx.ai/benchmarks/performance/gfphglvt",
   "tg_tokens": null,
   "tps": 5.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-m2wwxhpr-ctx1024",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: 6ldtzr3l; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 467.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 467.2 37.3 18.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m2wwxhpr",
   "source_url": "https://omlx.ai/benchmarks/performance/m2wwxhpr",
   "tg_tokens": null,
   "tps": 37.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-m2wwxhpr-ctx4096",
   "model": "Swift-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 6ldtzr3l; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 495.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-Qwen3.8-27b-oQ4e-fp16-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 495.3 36.0 19.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m2wwxhpr",
   "source_url": "https://omlx.ai/benchmarks/performance/m2wwxhpr",
   "tg_tokens": null,
   "tps": 36.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-dhot4j17-ctx1024",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 25.8 GB; TTFT 1306 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 784.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k tokens 784.1 PP tok/s 48.5 TG tok/s 1306 TTFT (ms) 25.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dhot4j17",
   "source_url": "https://omlx.ai/benchmarks/performance/dhot4j17",
   "tg_tokens": null,
   "tps": 48.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-tp9m72dt-ctx1024",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: f9cryc2a, fu4fq6ez, yyji3knn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2027.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 2,027 93.2 25.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tp9m72dt",
   "source_url": "https://omlx.ai/benchmarks/performance/tp9m72dt",
   "tg_tokens": null,
   "tps": 93.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-tp9m72dt-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: f9cryc2a, fu4fq6ez, yyji3knn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4072.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 4,072 91.1 26.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tp9m72dt",
   "source_url": "https://omlx.ai/benchmarks/performance/tp9m72dt",
   "tg_tokens": null,
   "tps": 91.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-tp9m72dt-ctx8192",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: f9cryc2a, fu4fq6ez, yyji3knn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4147.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 4,147 87.7 26.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tp9m72dt",
   "source_url": "https://omlx.ai/benchmarks/performance/tp9m72dt",
   "tg_tokens": null,
   "tps": 87.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-tp9m72dt-ctx16384",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: f9cryc2a, fu4fq6ez, yyji3knn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3876.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 3,876 81.7 26.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tp9m72dt",
   "source_url": "https://omlx.ai/benchmarks/performance/tp9m72dt",
   "tg_tokens": null,
   "tps": 81.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-pn909hdp-ctx32768",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 45fje1c3, bxub4h8m, vwujk4gd; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1905.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 1,905 43.2 17.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pn909hdp",
   "source_url": "https://omlx.ai/benchmarks/performance/pn909hdp",
   "tg_tokens": null,
   "tps": 43.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-pn909hdp-ctx65536",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 45fje1c3, bxub4h8m, vwujk4gd; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1496.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 1,496 35.8 19.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pn909hdp",
   "source_url": "https://omlx.ai/benchmarks/performance/pn909hdp",
   "tg_tokens": null,
   "tps": 35.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-pn909hdp-ctx131072",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: 45fje1c3, bxub4h8m, vwujk4gd; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1020.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 128k 1,020 19.0 22.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pn909hdp",
   "source_url": "https://omlx.ai/benchmarks/performance/pn909hdp",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-pn909hdp-ctx199680",
   "model": "gemma-4-26B-A4B-it",
   "notes": "QAT (quantization-aware training); oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: 45fje1c3, bxub4h8m, vwujk4gd; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 760.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 195k 760.6 11.0 26.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pn909hdp",
   "source_url": "https://omlx.ai/benchmarks/performance/pn909hdp",
   "tg_tokens": null,
   "tps": 11.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-te8atfp1-ctx1024",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: greu0vxm, jia6vrmf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 645.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 645.0 15.4 32.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/te8atfp1",
   "source_url": "https://omlx.ai/benchmarks/performance/te8atfp1",
   "tg_tokens": null,
   "tps": 15.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-te8atfp1-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 33.1 GB at this context; duplicate traces of the same session: greu0vxm, jia6vrmf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 699.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 699.9 15.0 33.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/te8atfp1",
   "source_url": "https://omlx.ai/benchmarks/performance/te8atfp1",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-te8atfp1-ctx8192",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 33.6 GB at this context; duplicate traces of the same session: greu0vxm, jia6vrmf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 687.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 687.4 14.5 33.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/te8atfp1",
   "source_url": "https://omlx.ai/benchmarks/performance/te8atfp1",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-te8atfp1-ctx16384",
   "model": "gemma-4-31B-it",
   "notes": "oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: greu0vxm, jia6vrmf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 662.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 662.8 13.6 34.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/te8atfp1",
   "source_url": "https://omlx.ai/benchmarks/performance/te8atfp1",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-3tmwva5m-ctx1024",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 5.1 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2099.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-E2B-it-MLX-8bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 8bit; 1k 2,099 39.0 5.1 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3tmwva5m",
   "source_url": "https://omlx.ai/benchmarks/performance/3tmwva5m",
   "tg_tokens": null,
   "tps": 39.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-3tmwva5m-ctx4096",
   "model": "Gemma 4 E2B",
   "notes": "oMLX community benchmark trace; peak memory 5.1 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2318.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-E2B-it-MLX-8bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 8bit; 4k 2,318 38.5 5.1 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3tmwva5m",
   "source_url": "https://omlx.ai/benchmarks/performance/3tmwva5m",
   "tg_tokens": null,
   "tps": 38.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-px4d3se8-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 403.7 GB at this context; duplicate traces of the same session: cs3p73ej, f2edw8aj, fhhacbpo, kjoco650, kkttkibj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 735.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 8k 735.2 32.5 403.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/px4d3se8",
   "source_url": "https://omlx.ai/benchmarks/performance/px4d3se8",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-px4d3se8-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 403.7 GB at this context; duplicate traces of the same session: cs3p73ej, f2edw8aj, fhhacbpo, kjoco650, kkttkibj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 755.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 16k 755.8 37.2 403.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/px4d3se8",
   "source_url": "https://omlx.ai/benchmarks/performance/px4d3se8",
   "tg_tokens": null,
   "tps": 37.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-px4d3se8-ctx32768",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 403.8 GB at this context; duplicate traces of the same session: cs3p73ej, f2edw8aj, fhhacbpo, kjoco650, kkttkibj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 744.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 32k 744.4 33.2 403.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/px4d3se8",
   "source_url": "https://omlx.ai/benchmarks/performance/px4d3se8",
   "tg_tokens": null,
   "tps": 33.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-px4d3se8-ctx65536",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 403.8 GB at this context; duplicate traces of the same session: cs3p73ej, f2edw8aj, fhhacbpo, kjoco650, kkttkibj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 736.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 64k 736.1 35.4 403.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/px4d3se8",
   "source_url": "https://omlx.ai/benchmarks/performance/px4d3se8",
   "tg_tokens": null,
   "tps": 35.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-px4d3se8-ctx131072",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 403.9 GB at this context; duplicate traces of the same session: cs3p73ej, f2edw8aj, fhhacbpo, kjoco650, kkttkibj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 710.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 128k 710.5 34.5 403.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/px4d3se8",
   "source_url": "https://omlx.ai/benchmarks/performance/px4d3se8",
   "tg_tokens": null,
   "tps": 34.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-px4d3se8-ctx199680",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 403.9 GB at this context; duplicate traces of the same session: cs3p73ej, f2edw8aj, fhhacbpo, kjoco650, kkttkibj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 684.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 195k 684.6 38.1 403.9 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/px4d3se8",
   "source_url": "https://omlx.ai/benchmarks/performance/px4d3se8",
   "tg_tokens": null,
   "tps": 38.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-lig9zoah-ctx4096",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 2.4 GB at this context; duplicate traces of the same session: 38yot9fm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3830.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiniCPM5-2B-oQ4e on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 3,830 136.3 2.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lig9zoah",
   "source_url": "https://omlx.ai/benchmarks/performance/lig9zoah",
   "tg_tokens": null,
   "tps": 136.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-lig9zoah-ctx16384",
   "model": "MiniCPM5-2B",
   "notes": "oMLX community benchmark trace; peak memory 2.6 GB at this context; duplicate traces of the same session: 38yot9fm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3877.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiniCPM5-2B-oQ4e on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 3,877 102.5 2.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lig9zoah",
   "source_url": "https://omlx.ai/benchmarks/performance/lig9zoah",
   "tg_tokens": null,
   "tps": 102.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-hsjkjqq5-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: 1pt7nj5i, 4ax8k1tp, 4zo7v7di, 7vqz1xne, httb6xuw, txl7dab8, vp7b18us; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 833.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k 833.7 26.8 5.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hsjkjqq5",
   "source_url": "https://omlx.ai/benchmarks/performance/hsjkjqq5",
   "tg_tokens": null,
   "tps": 26.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-hsjkjqq5-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: 1pt7nj5i, 4ax8k1tp, 4zo7v7di, 7vqz1xne, httb6xuw, txl7dab8, vp7b18us; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 875.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 875.5 26.1 6.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hsjkjqq5",
   "source_url": "https://omlx.ai/benchmarks/performance/hsjkjqq5",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-hsjkjqq5-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: 1pt7nj5i, 4ax8k1tp, 4zo7v7di, 7vqz1xne, httb6xuw, txl7dab8, vp7b18us; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 873.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 8k 873.9 24.6 6.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hsjkjqq5",
   "source_url": "https://omlx.ai/benchmarks/performance/hsjkjqq5",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-hsjkjqq5-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.6 GB at this context; duplicate traces of the same session: 1pt7nj5i, 4ax8k1tp, 4zo7v7di, 7vqz1xne, httb6xuw, txl7dab8, vp7b18us; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 825.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 16k 825.4 25.1 6.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hsjkjqq5",
   "source_url": "https://omlx.ai/benchmarks/performance/hsjkjqq5",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-hsjkjqq5-ctx32768",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 7.1 GB at this context; duplicate traces of the same session: 1pt7nj5i, 4ax8k1tp, 4zo7v7di, 7vqz1xne, httb6xuw, txl7dab8, vp7b18us; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 750.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 32k 750.3 23.6 7.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hsjkjqq5",
   "source_url": "https://omlx.ai/benchmarks/performance/hsjkjqq5",
   "tg_tokens": null,
   "tps": 23.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-hsjkjqq5-ctx65536",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 8.3 GB at this context; duplicate traces of the same session: 1pt7nj5i, 4ax8k1tp, 4zo7v7di, 7vqz1xne, httb6xuw, txl7dab8, vp7b18us; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 664.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 64k 664.1 21.1 8.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hsjkjqq5",
   "source_url": "https://omlx.ai/benchmarks/performance/hsjkjqq5",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-hsjkjqq5-ctx131072",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 11.9 GB at this context; duplicate traces of the same session: 1pt7nj5i, 4ax8k1tp, 4zo7v7di, 7vqz1xne, httb6xuw, txl7dab8, vp7b18us; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 540.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 128k 540.3 17.6 11.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hsjkjqq5",
   "source_url": "https://omlx.ai/benchmarks/performance/hsjkjqq5",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-hsjkjqq5-ctx199680",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 15.7 GB at this context; duplicate traces of the same session: 1pt7nj5i, 4ax8k1tp, 4zo7v7di, 7vqz1xne, httb6xuw, txl7dab8, vp7b18us; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 447.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 195k 447.6 15.0 15.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hsjkjqq5",
   "source_url": "https://omlx.ai/benchmarks/performance/hsjkjqq5",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-lkmk2dxz-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 22.5 GB; TTFT 25728 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 636.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 16k tokens 636.8 PP tok/s 75.6 TG tok/s 25728 TTFT (ms) 22.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lkmk2dxz",
   "source_url": "https://omlx.ai/benchmarks/performance/lkmk2dxz",
   "tg_tokens": null,
   "tps": 75.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-o20mlc8d-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: p4xjm96f; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 920.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-MLX-6bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 920.1 72.0 27.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o20mlc8d",
   "source_url": "https://omlx.ai/benchmarks/performance/o20mlc8d",
   "tg_tokens": null,
   "tps": 72.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-o20mlc8d-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 29.3 GB at this context; duplicate traces of the same session: p4xjm96f; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1055.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-MLX-6bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 1,055 68.9 29.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o20mlc8d",
   "source_url": "https://omlx.ai/benchmarks/performance/o20mlc8d",
   "tg_tokens": null,
   "tps": 68.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-gjbrrnvu-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: 8forflgr; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 866.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-MLX-6bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 1k 866.6 72.1 27.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gjbrrnvu",
   "source_url": "https://omlx.ai/benchmarks/performance/gjbrrnvu",
   "tg_tokens": null,
   "tps": 72.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-gjbrrnvu-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 29.3 GB at this context; duplicate traces of the same session: 8forflgr; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1044.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.6-35B-A3B-MLX-6bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 6bit; 4k 1,044 70.2 29.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gjbrrnvu",
   "source_url": "https://omlx.ai/benchmarks/performance/gjbrrnvu",
   "tg_tokens": null,
   "tps": 70.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jdx2wrc1-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 22.7 GB; TTFT 9131 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3589.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k tokens 3,589 PP tok/s 63.9 TG tok/s 9131 TTFT (ms) 22.7 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jdx2wrc1",
   "source_url": "https://omlx.ai/benchmarks/performance/jdx2wrc1",
   "tg_tokens": null,
   "tps": 63.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-n2l0176d-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 22.7 GB; TTFT 6960 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4708.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k tokens 4,708 PP tok/s 254.3 TG tok/s 6960 TTFT (ms) 22.7 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n2l0176d",
   "source_url": "https://omlx.ai/benchmarks/performance/n2l0176d",
   "tg_tokens": null,
   "tps": 254.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-1ejiuncp-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 2olmz9da; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 132.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 132.1 15.7 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ejiuncp",
   "source_url": "https://omlx.ai/benchmarks/performance/1ejiuncp",
   "tg_tokens": null,
   "tps": 15.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-1ejiuncp-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 2olmz9da; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 132.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 132.8 15.2 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1ejiuncp",
   "source_url": "https://omlx.ai/benchmarks/performance/1ejiuncp",
   "tg_tokens": null,
   "tps": 15.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "outlier"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-6p20pavo-ctx32768",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: j2hhu953, nutvf4u3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 440.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 440.2 49.3 26.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6p20pavo",
   "source_url": "https://omlx.ai/benchmarks/performance/6p20pavo",
   "tg_tokens": null,
   "tps": 49.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-6p20pavo-ctx65536",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 28.8 GB at this context; duplicate traces of the same session: j2hhu953, nutvf4u3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 382.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 382.2 24.0 28.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6p20pavo",
   "source_url": "https://omlx.ai/benchmarks/performance/6p20pavo",
   "tg_tokens": null,
   "tps": 24.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-6p20pavo-ctx131072",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: j2hhu953, nutvf4u3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 300.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 128k 300.0 22.0 33.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6p20pavo",
   "source_url": "https://omlx.ai/benchmarks/performance/6p20pavo",
   "tg_tokens": null,
   "tps": 22.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-raptzkkj-ctx16384",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "oMLX community benchmark trace; peak memory 37.5 GB at this context; duplicate traces of the same session: bfetbes2, wlxy3cv3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1139.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-8bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 8bit; 16k 1,139 58.2 37.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/raptzkkj",
   "source_url": "https://omlx.ai/benchmarks/performance/raptzkkj",
   "tg_tokens": null,
   "tps": 58.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-raptzkkj-ctx65536",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "oMLX community benchmark trace; peak memory 44.1 GB at this context; duplicate traces of the same session: bfetbes2, wlxy3cv3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 789.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-8bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 8bit; 64k 789.2 29.8 44.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/raptzkkj",
   "source_url": "https://omlx.ai/benchmarks/performance/raptzkkj",
   "tg_tokens": null,
   "tps": 29.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-raptzkkj-ctx131072",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "oMLX community benchmark trace; peak memory 54.6 GB at this context; duplicate traces of the same session: bfetbes2, wlxy3cv3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 554.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-8bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 8bit; 128k 554.3 22.2 54.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/raptzkkj",
   "source_url": "https://omlx.ai/benchmarks/performance/raptzkkj",
   "tg_tokens": null,
   "tps": 22.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-np243xpq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: 0ttdjp5t, yn4q0eml; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 910.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 910.2 64.5 17.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/np243xpq",
   "source_url": "https://omlx.ai/benchmarks/performance/np243xpq",
   "tg_tokens": null,
   "tps": 64.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-np243xpq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 0ttdjp5t, yn4q0eml; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1005.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,005 68.3 18.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/np243xpq",
   "source_url": "https://omlx.ai/benchmarks/performance/np243xpq",
   "tg_tokens": null,
   "tps": 68.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-np243xpq-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 0ttdjp5t, yn4q0eml; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 932.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 932.6 77.0 19.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/np243xpq",
   "source_url": "https://omlx.ai/benchmarks/performance/np243xpq",
   "tg_tokens": null,
   "tps": 77.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-np243xpq-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 0ttdjp5t, yn4q0eml; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 811.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 811.6 82.0 19.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/np243xpq",
   "source_url": "https://omlx.ai/benchmarks/performance/np243xpq",
   "tg_tokens": null,
   "tps": 82.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-np243xpq-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 0ttdjp5t, yn4q0eml; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 764.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 764.0 76.1 20.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/np243xpq",
   "source_url": "https://omlx.ai/benchmarks/performance/np243xpq",
   "tg_tokens": null,
   "tps": 76.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-np243xpq-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: 0ttdjp5t, yn4q0eml; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 715.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 715.9 59.9 23.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/np243xpq",
   "source_url": "https://omlx.ai/benchmarks/performance/np243xpq",
   "tg_tokens": null,
   "tps": 59.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-elp0zkod-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: nnrwtf01; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 455.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 4k 455.2 26.6 26.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/elp0zkod",
   "source_url": "https://omlx.ai/benchmarks/performance/elp0zkod",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-elp0zkod-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.1 GB at this context; duplicate traces of the same session: nnrwtf01; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 436.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 16k 436.8 25.2 27.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/elp0zkod",
   "source_url": "https://omlx.ai/benchmarks/performance/elp0zkod",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-elp0zkod-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: nnrwtf01; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 286.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 128k 286.7 18.2 35.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/elp0zkod",
   "source_url": "https://omlx.ai/benchmarks/performance/elp0zkod",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-vp1i6uwd-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 1uewhgrt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 253.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 253.5 20.3 22.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vp1i6uwd",
   "source_url": "https://omlx.ai/benchmarks/performance/vp1i6uwd",
   "tg_tokens": null,
   "tps": 20.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-vp1i6uwd-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: 1uewhgrt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 271.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 271.1 20.3 25.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vp1i6uwd",
   "source_url": "https://omlx.ai/benchmarks/performance/vp1i6uwd",
   "tg_tokens": null,
   "tps": 20.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2raou1ch-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 29.8 GB; TTFT 44090 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 743.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 48 GB \u00b7 6bit; 32k tokens 743.2 PP tok/s 60.6 TG tok/s 44090 TTFT (ms) 29.8 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2raou1ch",
   "source_url": "https://omlx.ai/benchmarks/performance/2raou1ch",
   "tg_tokens": null,
   "tps": 60.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-8k6ej708-ctx1024",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace, single run at 1k context; peak memory 5.8 GB; TTFT 6895 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 148.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 148.5 PP tok/s 16.4 TG tok/s 6895 TTFT (ms) 5.8 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8k6ej708",
   "source_url": "https://omlx.ai/benchmarks/performance/8k6ej708",
   "tg_tokens": null,
   "tps": 16.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-ovmccztv-ctx1024",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace, single run at 1k context; peak memory 6.6 GB; TTFT 7020 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 145.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 145.9 PP tok/s 16.4 TG tok/s 7020 TTFT (ms) 6.6 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ovmccztv",
   "source_url": "https://omlx.ai/benchmarks/performance/ovmccztv",
   "tg_tokens": null,
   "tps": 16.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-0y0ebapu-ctx1024",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace, single run at 1k context; peak memory 5.8 GB; TTFT 6870 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 149.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 149.1 PP tok/s 16.4 TG tok/s 6870 TTFT (ms) 5.8 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0y0ebapu",
   "source_url": "https://omlx.ai/benchmarks/performance/0y0ebapu",
   "tg_tokens": null,
   "tps": 16.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-76fuogle-ctx1024",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace, single run at 1k context; peak memory 6.6 GB; TTFT 6952 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 147.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 147.3 PP tok/s 16.6 TG tok/s 6952 TTFT (ms) 6.6 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/76fuogle",
   "source_url": "https://omlx.ai/benchmarks/performance/76fuogle",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-ffwafqg5-ctx1024",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MLX build; oMLX community benchmark trace, single run at 1k context; peak memory 7.3 GB; TTFT 7197 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 142.3,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-5bit-MLX on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 5bit; 1k tokens 142.3 PP tok/s 12.1 TG tok/s 7197 TTFT (ms) 7.3 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ffwafqg5",
   "source_url": "https://omlx.ai/benchmarks/performance/ffwafqg5",
   "tg_tokens": null,
   "tps": 12.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-67x1021b-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 75.3 GB at this context; duplicate traces of the same session: 5jevbc8z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 963.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 963.8 61.9 75.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/67x1021b",
   "source_url": "https://omlx.ai/benchmarks/performance/67x1021b",
   "tg_tokens": null,
   "tps": 61.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-67x1021b-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 75.4 GB at this context; duplicate traces of the same session: 5jevbc8z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1208.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,208 49.1 75.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/67x1021b",
   "source_url": "https://omlx.ai/benchmarks/performance/67x1021b",
   "tg_tokens": null,
   "tps": 49.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-y3gtsl75-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 40.9 GB; TTFT 64590 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 507.3,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 2bit; 32k tokens 507.3 PP tok/s 17.6 TG tok/s 64590 TTFT (ms) 40.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y3gtsl75",
   "source_url": "https://omlx.ai/benchmarks/performance/y3gtsl75",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-9r3g4bsx-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 25.3 GB; TTFT 71483 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 458.4,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 2bit; 32k tokens 458.4 PP tok/s 12.2 TG tok/s 71483 TTFT (ms) 25.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9r3g4bsx",
   "source_url": "https://omlx.ai/benchmarks/performance/9r3g4bsx",
   "tg_tokens": null,
   "tps": 12.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-libw2m5n-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 69.3 GB; TTFT 1858 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 551.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 551.2 PP tok/s 53.1 TG tok/s 1858 TTFT (ms) 69.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/libw2m5n",
   "source_url": "https://omlx.ai/benchmarks/performance/libw2m5n",
   "tg_tokens": null,
   "tps": 53.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-kbaeciv7-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: dnt9gill, xn02ikbc, zn69p6qc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 622.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 622.4 63.4 70.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kbaeciv7",
   "source_url": "https://omlx.ai/benchmarks/performance/kbaeciv7",
   "tg_tokens": null,
   "tps": 63.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-kbaeciv7-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: dnt9gill, xn02ikbc, zn69p6qc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 747.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 747.3 74.5 71.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kbaeciv7",
   "source_url": "https://omlx.ai/benchmarks/performance/kbaeciv7",
   "tg_tokens": null,
   "tps": 74.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-kbaeciv7-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.7 GB at this context; duplicate traces of the same session: dnt9gill, xn02ikbc, zn69p6qc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 763.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 763.1 60.1 73.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kbaeciv7",
   "source_url": "https://omlx.ai/benchmarks/performance/kbaeciv7",
   "tg_tokens": null,
   "tps": 60.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-kbaeciv7-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 76.1 GB at this context; duplicate traces of the same session: dnt9gill, xn02ikbc, zn69p6qc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 753.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 753.9 53.9 76.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kbaeciv7",
   "source_url": "https://omlx.ai/benchmarks/performance/kbaeciv7",
   "tg_tokens": null,
   "tps": 53.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-rjcntf92-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 9v0kjo6t, dvdr9oqh, u5s2e5m2, wx51cl5s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 523.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 523.9 74.4 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rjcntf92",
   "source_url": "https://omlx.ai/benchmarks/performance/rjcntf92",
   "tg_tokens": null,
   "tps": 74.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-rjcntf92-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: 9v0kjo6t, dvdr9oqh, u5s2e5m2, wx51cl5s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 692.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 692.3 56.1 71.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rjcntf92",
   "source_url": "https://omlx.ai/benchmarks/performance/rjcntf92",
   "tg_tokens": null,
   "tps": 56.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-rjcntf92-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.5 GB at this context; duplicate traces of the same session: 9v0kjo6t, dvdr9oqh, u5s2e5m2, wx51cl5s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 756.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 756.2 59.1 72.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rjcntf92",
   "source_url": "https://omlx.ai/benchmarks/performance/rjcntf92",
   "tg_tokens": null,
   "tps": 59.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-rjcntf92-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.7 GB at this context; duplicate traces of the same session: 9v0kjo6t, dvdr9oqh, u5s2e5m2, wx51cl5s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 726.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 726.7 53.7 73.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rjcntf92",
   "source_url": "https://omlx.ai/benchmarks/performance/rjcntf92",
   "tg_tokens": null,
   "tps": 53.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-rjcntf92-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 76.1 GB at this context; duplicate traces of the same session: 9v0kjo6t, dvdr9oqh, u5s2e5m2, wx51cl5s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 721.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 721.4 62.2 76.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rjcntf92",
   "source_url": "https://omlx.ai/benchmarks/performance/rjcntf92",
   "tg_tokens": null,
   "tps": 62.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-trenyjox-ctx1024",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: yp70cewg; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 782.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 782.0 48.9 25.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/trenyjox",
   "source_url": "https://omlx.ai/benchmarks/performance/trenyjox",
   "tg_tokens": null,
   "tps": 48.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-trenyjox-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: yp70cewg; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 760.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-it-MLX-8bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 760.9 47.9 26.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/trenyjox",
   "source_url": "https://omlx.ai/benchmarks/performance/trenyjox",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-inrb0fem-ctx65536",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 207.1 GB at this context; duplicate traces of the same session: dmf5lc36; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 612.7,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 unknown; 64k 612.7 47.2 207.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/inrb0fem",
   "source_url": "https://omlx.ai/benchmarks/performance/inrb0fem",
   "tg_tokens": null,
   "tps": 47.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-inrb0fem-ctx131072",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 207.1 GB at this context; duplicate traces of the same session: dmf5lc36; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 512.9,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 unknown; 128k 512.9 43.4 207.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/inrb0fem",
   "source_url": "https://omlx.ai/benchmarks/performance/inrb0fem",
   "tg_tokens": null,
   "tps": 43.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ykm3af2f-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 205.6 GB at this context; duplicate traces of the same session: 9x6vijpr, pxc5s2m1, xcn7lk3y; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 492.2,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 unknown; 1k 492.2 31.0 205.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ykm3af2f",
   "source_url": "https://omlx.ai/benchmarks/performance/ykm3af2f",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ykm3af2f-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 206.1 GB at this context; duplicate traces of the same session: 9x6vijpr, pxc5s2m1, xcn7lk3y; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 637.9,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 unknown; 4k 637.9 38.8 206.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ykm3af2f",
   "source_url": "https://omlx.ai/benchmarks/performance/ykm3af2f",
   "tg_tokens": null,
   "tps": 38.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-ykm3af2f-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 206.9 GB at this context; duplicate traces of the same session: 9x6vijpr, pxc5s2m1, xcn7lk3y; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 671.6,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 unknown; 16k 671.6 46.1 206.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ykm3af2f",
   "source_url": "https://omlx.ai/benchmarks/performance/ykm3af2f",
   "tg_tokens": null,
   "tps": 46.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ykm3af2f-ctx32768",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 207.0 GB at this context; duplicate traces of the same session: 9x6vijpr, pxc5s2m1, xcn7lk3y; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 712.9,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 unknown; 32k 712.9 36.0 207.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ykm3af2f",
   "source_url": "https://omlx.ai/benchmarks/performance/ykm3af2f",
   "tg_tokens": null,
   "tps": 36.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-0g84suc9-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 108.8 GB; TTFT 11899 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 344.2,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "GLM-5.3-Flash-oQ2e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 2bit; 4k tokens 344.2 PP tok/s 38.5 TG tok/s 11899 TTFT (ms) 108.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0g84suc9",
   "source_url": "https://omlx.ai/benchmarks/performance/0g84suc9",
   "tg_tokens": null,
   "tps": 38.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-7ncuy3un-ctx16384",
   "model": "K2-Horizon-7B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 7.7 GB; TTFT 90751 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 180.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "K2-Horizon-7B-oQ4e on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 16k tokens 180.5 PP tok/s 12.1 TG tok/s 90751 TTFT (ms) 7.7 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ncuy3un",
   "source_url": "https://omlx.ai/benchmarks/performance/7ncuy3un",
   "tg_tokens": null,
   "tps": 12.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mb9pp4f6-ctx1024",
   "model": "Kolibri-1",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 77.7 GB at this context; duplicate traces of the same session: uikbscnq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1372.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Kolibri-1-mlx-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 1,372 78.9 77.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mb9pp4f6",
   "source_url": "https://omlx.ai/benchmarks/performance/mb9pp4f6",
   "tg_tokens": null,
   "tps": 78.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mb9pp4f6-ctx4096",
   "model": "Kolibri-1",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 77.8 GB at this context; duplicate traces of the same session: uikbscnq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2136.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Kolibri-1-mlx-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 2,136 73.3 77.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mb9pp4f6",
   "source_url": "https://omlx.ai/benchmarks/performance/mb9pp4f6",
   "tg_tokens": null,
   "tps": 73.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-wsovzfgc-ctx1024",
   "model": "Ornith-1.0-35B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 18496egg, 2st9302o, a1p1acs1, nlg619zu, ztrpzgv3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 983.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.0-35B-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 983.3 79.3 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wsovzfgc",
   "source_url": "https://omlx.ai/benchmarks/performance/wsovzfgc",
   "tg_tokens": null,
   "tps": 79.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-wsovzfgc-ctx4096",
   "model": "Ornith-1.0-35B",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: 18496egg, 2st9302o, a1p1acs1, nlg619zu, ztrpzgv3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1042.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.0-35B-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 1,042 77.5 20.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wsovzfgc",
   "source_url": "https://omlx.ai/benchmarks/performance/wsovzfgc",
   "tg_tokens": null,
   "tps": 77.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-wsovzfgc-ctx8192",
   "model": "Ornith-1.0-35B",
   "notes": "oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 18496egg, 2st9302o, a1p1acs1, nlg619zu, ztrpzgv3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 990.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.0-35B-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 990.5 73.7 20.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wsovzfgc",
   "source_url": "https://omlx.ai/benchmarks/performance/wsovzfgc",
   "tg_tokens": null,
   "tps": 73.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-wsovzfgc-ctx16384",
   "model": "Ornith-1.0-35B",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 18496egg, 2st9302o, a1p1acs1, nlg619zu, ztrpzgv3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 889.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.0-35B-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 889.9 71.1 21.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wsovzfgc",
   "source_url": "https://omlx.ai/benchmarks/performance/wsovzfgc",
   "tg_tokens": null,
   "tps": 71.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-wsovzfgc-ctx32768",
   "model": "Ornith-1.0-35B",
   "notes": "oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 18496egg, 2st9302o, a1p1acs1, nlg619zu, ztrpzgv3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 733.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.0-35B-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 733.5 60.6 21.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wsovzfgc",
   "source_url": "https://omlx.ai/benchmarks/performance/wsovzfgc",
   "tg_tokens": null,
   "tps": 60.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-wsovzfgc-ctx65536",
   "model": "Ornith-1.0-35B",
   "notes": "oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: 18496egg, 2st9302o, a1p1acs1, nlg619zu, ztrpzgv3; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 549.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.0-35B-oQ4e on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 549.3 51.8 22.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wsovzfgc",
   "source_url": "https://omlx.ai/benchmarks/performance/wsovzfgc",
   "tg_tokens": null,
   "tps": 51.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-le9vay6c-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: v8c6pbi8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 190.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 190.0 13.8 21.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/le9vay6c",
   "source_url": "https://omlx.ai/benchmarks/performance/le9vay6c",
   "tg_tokens": null,
   "tps": 13.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-le9vay6c-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: v8c6pbi8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 131.8 12.9 22.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/le9vay6c",
   "source_url": "https://omlx.ai/benchmarks/performance/le9vay6c",
   "tg_tokens": null,
   "tps": 12.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-8n75lwzx-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: yj6mwpcw; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 137.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 1k 137.2 20.5 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8n75lwzx",
   "source_url": "https://omlx.ai/benchmarks/performance/8n75lwzx",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-8n75lwzx-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: yj6mwpcw; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 128.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k 128.7 19.2 19.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8n75lwzx",
   "source_url": "https://omlx.ai/benchmarks/performance/8n75lwzx",
   "tg_tokens": null,
   "tps": 19.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-50096e7w-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: qsyd6gb4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1164.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,164 44.9 20.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/50096e7w",
   "source_url": "https://omlx.ai/benchmarks/performance/50096e7w",
   "tg_tokens": null,
   "tps": 44.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-50096e7w-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: qsyd6gb4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1812.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 1,812 44.8 21.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/50096e7w",
   "source_url": "https://omlx.ai/benchmarks/performance/50096e7w",
   "tg_tokens": null,
   "tps": 44.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-nthjiea0-ctx16384",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: 9v6hweiv, tgtofkb7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1136.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-8bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 8bit; 16k 1,136 53.8 37.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nthjiea0",
   "source_url": "https://omlx.ai/benchmarks/performance/nthjiea0",
   "tg_tokens": null,
   "tps": 53.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-nthjiea0-ctx65536",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "oMLX community benchmark trace; peak memory 43.3 GB at this context; duplicate traces of the same session: 9v6hweiv, tgtofkb7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 790.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-8bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 8bit; 64k 790.7 32.3 43.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nthjiea0",
   "source_url": "https://omlx.ai/benchmarks/performance/nthjiea0",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-nthjiea0-ctx131072",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "oMLX community benchmark trace; peak memory 52.7 GB at this context; duplicate traces of the same session: 9v6hweiv, tgtofkb7; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 550.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-8bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 8bit; 128k 550.1 22.8 52.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nthjiea0",
   "source_url": "https://omlx.ai/benchmarks/performance/nthjiea0",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zmwi7xtk-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: atrf217x, uxng0z6u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 910.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 910.2 64.5 17.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zmwi7xtk",
   "source_url": "https://omlx.ai/benchmarks/performance/zmwi7xtk",
   "tg_tokens": null,
   "tps": 64.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zmwi7xtk-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: atrf217x, uxng0z6u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1005.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,005 68.3 18.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zmwi7xtk",
   "source_url": "https://omlx.ai/benchmarks/performance/zmwi7xtk",
   "tg_tokens": null,
   "tps": 68.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zmwi7xtk-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: atrf217x, uxng0z6u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 932.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 932.6 77.0 19.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zmwi7xtk",
   "source_url": "https://omlx.ai/benchmarks/performance/zmwi7xtk",
   "tg_tokens": null,
   "tps": 77.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zmwi7xtk-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: atrf217x, uxng0z6u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 811.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 811.6 82.0 19.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zmwi7xtk",
   "source_url": "https://omlx.ai/benchmarks/performance/zmwi7xtk",
   "tg_tokens": null,
   "tps": 82.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zmwi7xtk-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: atrf217x, uxng0z6u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 764.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 764.0 76.1 20.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zmwi7xtk",
   "source_url": "https://omlx.ai/benchmarks/performance/zmwi7xtk",
   "tg_tokens": null,
   "tps": 76.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-zmwi7xtk-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: atrf217x, uxng0z6u; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 715.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 715.9 59.9 23.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zmwi7xtk",
   "source_url": "https://omlx.ai/benchmarks/performance/zmwi7xtk",
   "tg_tokens": null,
   "tps": 59.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-q10jcukp-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.2 GB at this context; duplicate traces of the same session: f770yjxt, wsrdgn3n; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1119.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 1k 1,119 75.2 39.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q10jcukp",
   "source_url": "https://omlx.ai/benchmarks/performance/q10jcukp",
   "tg_tokens": null,
   "tps": 75.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-q10jcukp-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 40.2 GB at this context; duplicate traces of the same session: f770yjxt, wsrdgn3n; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1109.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 4k 1,109 66.7 40.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q10jcukp",
   "source_url": "https://omlx.ai/benchmarks/performance/q10jcukp",
   "tg_tokens": null,
   "tps": 66.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-q10jcukp-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 41.0 GB at this context; duplicate traces of the same session: f770yjxt, wsrdgn3n; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1013.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 16k 1,013 96.9 41.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q10jcukp",
   "source_url": "https://omlx.ai/benchmarks/performance/q10jcukp",
   "tg_tokens": null,
   "tps": 96.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-0i10esy8-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: el3br7zr, kvndodgw; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1111.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 1k 1,111 61.0 33.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0i10esy8",
   "source_url": "https://omlx.ai/benchmarks/performance/0i10esy8",
   "tg_tokens": null,
   "tps": 61.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-0i10esy8-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: el3br7zr, kvndodgw; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1316.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 4k 1,316 46.2 34.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0i10esy8",
   "source_url": "https://omlx.ai/benchmarks/performance/0i10esy8",
   "tg_tokens": null,
   "tps": 46.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-0i10esy8-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.7 GB at this context; duplicate traces of the same session: el3br7zr, kvndodgw; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1281.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 16k 1,281 74.9 35.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0i10esy8",
   "source_url": "https://omlx.ai/benchmarks/performance/0i10esy8",
   "tg_tokens": null,
   "tps": 74.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-png7002g-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 33.9 GB; TTFT 915 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1119.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 1k tokens 1,119 PP tok/s 72.5 TG tok/s 915 TTFT (ms) 33.9 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/png7002g",
   "source_url": "https://omlx.ai/benchmarks/performance/png7002g",
   "tg_tokens": null,
   "tps": 72.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-1qp86bm6-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: ofybqr0w, v0c22d7e; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1090.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 1k 1,090 78.3 33.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1qp86bm6",
   "source_url": "https://omlx.ai/benchmarks/performance/1qp86bm6",
   "tg_tokens": null,
   "tps": 78.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-1qp86bm6-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: ofybqr0w, v0c22d7e; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1313.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 4k 1,313 82.3 34.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1qp86bm6",
   "source_url": "https://omlx.ai/benchmarks/performance/1qp86bm6",
   "tg_tokens": null,
   "tps": 82.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-1qp86bm6-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.7 GB at this context; duplicate traces of the same session: ofybqr0w, v0c22d7e; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1282.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 16k 1,282 81.3 35.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1qp86bm6",
   "source_url": "https://omlx.ai/benchmarks/performance/1qp86bm6",
   "tg_tokens": null,
   "tps": 81.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-bek33glo-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: jg1mcmhe, udxi36xf; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1106.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 1k 1,106 87.3 33.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bek33glo",
   "source_url": "https://omlx.ai/benchmarks/performance/bek33glo",
   "tg_tokens": null,
   "tps": 87.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-bek33glo-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: jg1mcmhe, udxi36xf; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1316.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 4k 1,316 72.1 34.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bek33glo",
   "source_url": "https://omlx.ai/benchmarks/performance/bek33glo",
   "tg_tokens": null,
   "tps": 72.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-bek33glo-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 35.7 GB at this context; duplicate traces of the same session: jg1mcmhe, udxi36xf; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1283.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 16k 1,283 95.7 35.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bek33glo",
   "source_url": "https://omlx.ai/benchmarks/performance/bek33glo",
   "tg_tokens": null,
   "tps": 95.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-8yue0jc1-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 03z07hbo, yrarz60e; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1166.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 1k 1,166 32.0 28.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8yue0jc1",
   "source_url": "https://omlx.ai/benchmarks/performance/8yue0jc1",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-8yue0jc1-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: 03z07hbo, yrarz60e; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1376.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 4k 1,376 31.9 29.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8yue0jc1",
   "source_url": "https://omlx.ai/benchmarks/performance/8yue0jc1",
   "tg_tokens": null,
   "tps": 31.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-8yue0jc1-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: 03z07hbo, yrarz60e; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1318.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 16k 1,318 30.7 30.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8yue0jc1",
   "source_url": "https://omlx.ai/benchmarks/performance/8yue0jc1",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-pq24fsgl-ctx1024",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace, single run at 1k context; peak memory 5.8 GB; TTFT 6695 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 152.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 152.9 PP tok/s 17.4 TG tok/s 6695 TTFT (ms) 5.8 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pq24fsgl",
   "source_url": "https://omlx.ai/benchmarks/performance/pq24fsgl",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-c1wuoph8-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.6 GB at this context; duplicate traces of the same session: cnqogggb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 703.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 703.2 43.2 74.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c1wuoph8",
   "source_url": "https://omlx.ai/benchmarks/performance/c1wuoph8",
   "tg_tokens": null,
   "tps": 43.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-c1wuoph8-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 77.8 GB at this context; duplicate traces of the same session: cnqogggb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 692.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 692.6 44.4 77.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c1wuoph8",
   "source_url": "https://omlx.ai/benchmarks/performance/c1wuoph8",
   "tg_tokens": null,
   "tps": 44.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-x277rird-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 69.3 GB; TTFT 1888 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 542.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 542.4 PP tok/s 51.7 TG tok/s 1888 TTFT (ms) 69.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x277rird",
   "source_url": "https://omlx.ai/benchmarks/performance/x277rird",
   "tg_tokens": null,
   "tps": 51.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5lxo17pb-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.2 GB at this context; duplicate traces of the same session: ehbspsyp, ic1hfex5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 951.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 951.4 62.8 69.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5lxo17pb",
   "source_url": "https://omlx.ai/benchmarks/performance/5lxo17pb",
   "tg_tokens": null,
   "tps": 62.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5lxo17pb-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: ehbspsyp, ic1hfex5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2844.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,844 75.8 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5lxo17pb",
   "source_url": "https://omlx.ai/benchmarks/performance/5lxo17pb",
   "tg_tokens": null,
   "tps": 75.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5lxo17pb-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.4 GB at this context; duplicate traces of the same session: ehbspsyp, ic1hfex5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2680.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,680 65.8 71.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5lxo17pb",
   "source_url": "https://omlx.ai/benchmarks/performance/5lxo17pb",
   "tg_tokens": null,
   "tps": 65.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5lxo17pb-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.4 GB at this context; duplicate traces of the same session: ehbspsyp, ic1hfex5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2406.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,406 63.1 72.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5lxo17pb",
   "source_url": "https://omlx.ai/benchmarks/performance/5lxo17pb",
   "tg_tokens": null,
   "tps": 63.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-1404756r-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 100.6 GB; TTFT 1248 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 820.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k tokens 820.3 PP tok/s 140.5 TG tok/s 1248 TTFT (ms) 100.6 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1404756r",
   "source_url": "https://omlx.ai/benchmarks/performance/1404756r",
   "tg_tokens": null,
   "tps": 140.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-4x66se8t-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.6 GB at this context; duplicate traces of the same session: 1hnsrlkl, 3n8vuver, 6denj3gl, 6gzym9m6, fwg2389l; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2493.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 2,493 146.3 142.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4x66se8t",
   "source_url": "https://omlx.ai/benchmarks/performance/4x66se8t",
   "tg_tokens": null,
   "tps": 146.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-4x66se8t-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: 1hnsrlkl, 3n8vuver, 6denj3gl, 6gzym9m6, fwg2389l; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4502.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 4,502 131.1 143.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4x66se8t",
   "source_url": "https://omlx.ai/benchmarks/performance/4x66se8t",
   "tg_tokens": null,
   "tps": 131.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-4x66se8t-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.9 GB at this context; duplicate traces of the same session: 1hnsrlkl, 3n8vuver, 6denj3gl, 6gzym9m6, fwg2389l; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4856.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 32k 4,856 119.5 145.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4x66se8t",
   "source_url": "https://omlx.ai/benchmarks/performance/4x66se8t",
   "tg_tokens": null,
   "tps": 119.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-4x66se8t-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 147.0 GB at this context; duplicate traces of the same session: 1hnsrlkl, 3n8vuver, 6denj3gl, 6gzym9m6, fwg2389l; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4821.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 64k 4,821 118.4 147.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4x66se8t",
   "source_url": "https://omlx.ai/benchmarks/performance/4x66se8t",
   "tg_tokens": null,
   "tps": 118.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-4x66se8t-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 149.3 GB at this context; duplicate traces of the same session: 1hnsrlkl, 3n8vuver, 6denj3gl, 6gzym9m6, fwg2389l; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4717.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 128k 4,717 104.4 149.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4x66se8t",
   "source_url": "https://omlx.ai/benchmarks/performance/4x66se8t",
   "tg_tokens": null,
   "tps": 104.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-4x66se8t-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 152.1 GB at this context; duplicate traces of the same session: 1hnsrlkl, 3n8vuver, 6denj3gl, 6gzym9m6, fwg2389l; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4584.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 195k 4,584 85.0 152.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4x66se8t",
   "source_url": "https://omlx.ai/benchmarks/performance/4x66se8t",
   "tg_tokens": null,
   "tps": 85.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-bpdmqvza-ctx1024",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: 8nzdl3v1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 624.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 624.4 65.8 22.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bpdmqvza",
   "source_url": "https://omlx.ai/benchmarks/performance/bpdmqvza",
   "tg_tokens": null,
   "tps": 65.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-bpdmqvza-ctx4096",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: 8nzdl3v1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2211.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,211 55.7 23.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bpdmqvza",
   "source_url": "https://omlx.ai/benchmarks/performance/bpdmqvza",
   "tg_tokens": null,
   "tps": 55.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qk8ox5qt-ctx4096",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 147.3 GB at this context; duplicate traces of the same session: 4etmd109; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 613.0,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "DeepSeek-V4-Flash-0731-MXFP4-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 613.0 25.4 147.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qk8ox5qt",
   "source_url": "https://omlx.ai/benchmarks/performance/qk8ox5qt",
   "tg_tokens": null,
   "tps": 25.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-qk8ox5qt-ctx16384",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 147.4 GB at this context; duplicate traces of the same session: 4etmd109; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 619.7,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "DeepSeek-V4-Flash-0731-MXFP4-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 619.7 24.7 147.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qk8ox5qt",
   "source_url": "https://omlx.ai/benchmarks/performance/qk8ox5qt",
   "tg_tokens": null,
   "tps": 24.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-tzb5ze5d-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 205.5 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 632.2,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 unknown; 1k 632.2 35.8 205.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tzb5ze5d",
   "source_url": "https://omlx.ai/benchmarks/performance/tzb5ze5d",
   "tg_tokens": null,
   "tps": 35.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-tzb5ze5d-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 206.1 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 934.8,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 unknown; 16k 934.8 47.9 206.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tzb5ze5d",
   "source_url": "https://omlx.ai/benchmarks/performance/tzb5ze5d",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-tzb5ze5d-ctx32768",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 206.0 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1152.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 unknown; 32k 1,152 40.4 206.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tzb5ze5d",
   "source_url": "https://omlx.ai/benchmarks/performance/tzb5ze5d",
   "tg_tokens": null,
   "tps": 40.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-tzb5ze5d-ctx65536",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 206.2 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1085.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 unknown; 64k 1,085 48.5 206.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tzb5ze5d",
   "source_url": "https://omlx.ai/benchmarks/performance/tzb5ze5d",
   "tg_tokens": null,
   "tps": 48.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-tzb5ze5d-ctx131072",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 206.2 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 734.8,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 unknown; 128k 734.8 44.7 206.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tzb5ze5d",
   "source_url": "https://omlx.ai/benchmarks/performance/tzb5ze5d",
   "tg_tokens": null,
   "tps": 44.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-ne023rax-ctx65536",
   "model": "Qwen3.5-2B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 64k context; peak memory 2.8 GB; TTFT 28097 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2333.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-2B-MLX-4bit on M5 (10c); M5 (10c) \u00b7 24 GB \u00b7 4bit; 64k tokens 2,333 PP tok/s 56.7 TG tok/s 28097 TTFT (ms) 2.8 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ne023rax",
   "source_url": "https://omlx.ai/benchmarks/performance/ne023rax",
   "tg_tokens": null,
   "tps": 56.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-ikbgk0ij-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 3321 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 308.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 308.4 PP tok/s 32.7 TG tok/s 3321 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ikbgk0ij",
   "source_url": "https://omlx.ai/benchmarks/performance/ikbgk0ij",
   "tg_tokens": null,
   "tps": 32.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-r9pcpgzu-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: ajblxdsk; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 416.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 416.0 17.4 17.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r9pcpgzu",
   "source_url": "https://omlx.ai/benchmarks/performance/r9pcpgzu",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-r9pcpgzu-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: ajblxdsk; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 395.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 395.5 16.4 17.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r9pcpgzu",
   "source_url": "https://omlx.ai/benchmarks/performance/r9pcpgzu",
   "tg_tokens": null,
   "tps": 16.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-poqpto9g-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 82zzpwk3, 9n6ex66e, gm00i6kw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 327.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.8-27B-MLX-5bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 1k 327.0 31.5 19.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/poqpto9g",
   "source_url": "https://omlx.ai/benchmarks/performance/poqpto9g",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-poqpto9g-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 82zzpwk3, 9n6ex66e, gm00i6kw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 352.6,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.8-27B-MLX-5bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 4k 352.6 31.1 22.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/poqpto9g",
   "source_url": "https://omlx.ai/benchmarks/performance/poqpto9g",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-poqpto9g-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: 82zzpwk3, 9n6ex66e, gm00i6kw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 346.7,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.8-27B-MLX-5bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 8k 346.7 30.2 22.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/poqpto9g",
   "source_url": "https://omlx.ai/benchmarks/performance/poqpto9g",
   "tg_tokens": null,
   "tps": 30.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-poqpto9g-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 82zzpwk3, 9n6ex66e, gm00i6kw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 334.2,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.8-27B-MLX-5bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 16k 334.2 29.0 23.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/poqpto9g",
   "source_url": "https://omlx.ai/benchmarks/performance/poqpto9g",
   "tg_tokens": null,
   "tps": 29.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-poqpto9g-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: 82zzpwk3, 9n6ex66e, gm00i6kw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 310.7,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.8-27B-MLX-5bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 32k 310.7 26.9 25.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/poqpto9g",
   "source_url": "https://omlx.ai/benchmarks/performance/poqpto9g",
   "tg_tokens": null,
   "tps": 26.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-poqpto9g-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: 82zzpwk3, 9n6ex66e, gm00i6kw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 272.4,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.8-27B-MLX-5bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 64k 272.4 23.9 27.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/poqpto9g",
   "source_url": "https://omlx.ai/benchmarks/performance/poqpto9g",
   "tg_tokens": null,
   "tps": 23.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-poqpto9g-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 31.9 GB at this context; duplicate traces of the same session: 82zzpwk3, 9n6ex66e, gm00i6kw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 219.1,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.8-27B-MLX-5bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 128k 219.1 19.6 31.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/poqpto9g",
   "source_url": "https://omlx.ai/benchmarks/performance/poqpto9g",
   "tg_tokens": null,
   "tps": 19.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-poqpto9g-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: 82zzpwk3, 9n6ex66e, gm00i6kw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 181.6,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3.8-27B-MLX-5bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 5bit; 195k 181.6 16.5 36.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/poqpto9g",
   "source_url": "https://omlx.ai/benchmarks/performance/poqpto9g",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3tzi6wfx-ctx1024",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 25.3 GB at this context; duplicate traces of the same session: 1uj9d25v, 20pbhcwb, 6ovx3jjp, ac5mar0o, bq7uy08t, wnq9x91m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 625.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 625.1 19.4 25.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3tzi6wfx",
   "source_url": "https://omlx.ai/benchmarks/performance/3tzi6wfx",
   "tg_tokens": null,
   "tps": 19.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3tzi6wfx-ctx4096",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: 1uj9d25v, 20pbhcwb, 6ovx3jjp, ac5mar0o, bq7uy08t, wnq9x91m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 583.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 583.7 18.5 27.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3tzi6wfx",
   "source_url": "https://omlx.ai/benchmarks/performance/3tzi6wfx",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3tzi6wfx-ctx8192",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: 1uj9d25v, 20pbhcwb, 6ovx3jjp, ac5mar0o, bq7uy08t, wnq9x91m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 507.2,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 507.2 17.6 28.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3tzi6wfx",
   "source_url": "https://omlx.ai/benchmarks/performance/3tzi6wfx",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3tzi6wfx-ctx16384",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 29.0 GB at this context; duplicate traces of the same session: 1uj9d25v, 20pbhcwb, 6ovx3jjp, ac5mar0o, bq7uy08t, wnq9x91m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 412.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 412.5 16.7 29.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3tzi6wfx",
   "source_url": "https://omlx.ai/benchmarks/performance/3tzi6wfx",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3tzi6wfx-ctx32768",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 31.0 GB at this context; duplicate traces of the same session: 1uj9d25v, 20pbhcwb, 6ovx3jjp, ac5mar0o, bq7uy08t, wnq9x91m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 364.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 364.1 15.7 31.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3tzi6wfx",
   "source_url": "https://omlx.ai/benchmarks/performance/3tzi6wfx",
   "tg_tokens": null,
   "tps": 15.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3tzi6wfx-ctx65536",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 35.0 GB at this context; duplicate traces of the same session: 1uj9d25v, 20pbhcwb, 6ovx3jjp, ac5mar0o, bq7uy08t, wnq9x91m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 293.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k 293.7 13.1 35.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3tzi6wfx",
   "source_url": "https://omlx.ai/benchmarks/performance/3tzi6wfx",
   "tg_tokens": null,
   "tps": 13.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3tzi6wfx-ctx131072",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 43.0 GB at this context; duplicate traces of the same session: 1uj9d25v, 20pbhcwb, 6ovx3jjp, ac5mar0o, bq7uy08t, wnq9x91m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 208.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 128k 208.9 10.4 43.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3tzi6wfx",
   "source_url": "https://omlx.ai/benchmarks/performance/3tzi6wfx",
   "tg_tokens": null,
   "tps": 10.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-lgtbg1q7-ctx16384",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "oMLX community benchmark trace; peak memory 42.5 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1137.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-8bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 8bit; 16k 1,137 58.5 42.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lgtbg1q7",
   "source_url": "https://omlx.ai/benchmarks/performance/lgtbg1q7",
   "tg_tokens": null,
   "tps": 58.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-lgtbg1q7-ctx65536",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "oMLX community benchmark trace; peak memory 48.6 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 792.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-8bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 8bit; 64k 792.3 27.6 48.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lgtbg1q7",
   "source_url": "https://omlx.ai/benchmarks/performance/lgtbg1q7",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-lgtbg1q7-ctx131072",
   "model": "Qwen3.8-27B-Uncensored-OrcaRouter",
   "notes": "oMLX community benchmark trace; peak memory 58.1 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 553.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-Uncensored-OrcaRouter-MLX-8bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 96 GB \u00b7 8bit; 128k 553.4 23.3 58.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lgtbg1q7",
   "source_url": "https://omlx.ai/benchmarks/performance/lgtbg1q7",
   "tg_tokens": null,
   "tps": 23.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-orekfgb6-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 667.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 667.3 35.3 32.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/orekfgb6",
   "source_url": "https://omlx.ai/benchmarks/performance/orekfgb6",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-orekfgb6-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 33.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 772.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 772.4 33.3 33.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/orekfgb6",
   "source_url": "https://omlx.ai/benchmarks/performance/orekfgb6",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-bhlsvhyb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 34.6 GB at this context; duplicate traces of the same session: 0by81w08, 2s82c4h3, 5b1pg8qb, fn5m5vzi, gum4beal, s8nirmb0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 325.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 1k 325.9 50.7 34.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhlsvhyb",
   "source_url": "https://omlx.ai/benchmarks/performance/bhlsvhyb",
   "tg_tokens": null,
   "tps": 50.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-bhlsvhyb-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 38.1 GB at this context; duplicate traces of the same session: 0by81w08, 2s82c4h3, 5b1pg8qb, fn5m5vzi, gum4beal, s8nirmb0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 334.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 4k 334.4 46.8 38.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhlsvhyb",
   "source_url": "https://omlx.ai/benchmarks/performance/bhlsvhyb",
   "tg_tokens": null,
   "tps": 46.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-bhlsvhyb-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 38.6 GB at this context; duplicate traces of the same session: 0by81w08, 2s82c4h3, 5b1pg8qb, fn5m5vzi, gum4beal, s8nirmb0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 329.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 8k 329.5 52.3 38.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhlsvhyb",
   "source_url": "https://omlx.ai/benchmarks/performance/bhlsvhyb",
   "tg_tokens": null,
   "tps": 52.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bhlsvhyb-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 39.5 GB at this context; duplicate traces of the same session: 0by81w08, 2s82c4h3, 5b1pg8qb, fn5m5vzi, gum4beal, s8nirmb0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 317.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 16k 317.4 64.7 39.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhlsvhyb",
   "source_url": "https://omlx.ai/benchmarks/performance/bhlsvhyb",
   "tg_tokens": null,
   "tps": 64.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-bhlsvhyb-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 41.1 GB at this context; duplicate traces of the same session: 0by81w08, 2s82c4h3, 5b1pg8qb, fn5m5vzi, gum4beal, s8nirmb0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 293.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 32k 293.7 63.2 41.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhlsvhyb",
   "source_url": "https://omlx.ai/benchmarks/performance/bhlsvhyb",
   "tg_tokens": null,
   "tps": 63.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bhlsvhyb-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 43.8 GB at this context; duplicate traces of the same session: 0by81w08, 2s82c4h3, 5b1pg8qb, fn5m5vzi, gum4beal, s8nirmb0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 255.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 64k 255.3 46.4 43.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhlsvhyb",
   "source_url": "https://omlx.ai/benchmarks/performance/bhlsvhyb",
   "tg_tokens": null,
   "tps": 46.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bhlsvhyb-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 48.0 GB at this context; duplicate traces of the same session: 0by81w08, 2s82c4h3, 5b1pg8qb, fn5m5vzi, gum4beal, s8nirmb0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 203.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 8bit; 128k 203.2 47.0 48.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhlsvhyb",
   "source_url": "https://omlx.ai/benchmarks/performance/bhlsvhyb",
   "tg_tokens": null,
   "tps": 47.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8lzulsbv-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 660.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 660.1 32.9 32.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8lzulsbv",
   "source_url": "https://omlx.ai/benchmarks/performance/8lzulsbv",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8lzulsbv-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 747.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 747.3 35.7 33.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8lzulsbv",
   "source_url": "https://omlx.ai/benchmarks/performance/8lzulsbv",
   "tg_tokens": null,
   "tps": 35.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-l85i51rz-ctx1024",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace, single run at 1k context; peak memory 5.8 GB; TTFT 6870 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 149.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 149.1 PP tok/s 16.3 TG tok/s 6870 TTFT (ms) 5.8 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l85i51rz",
   "source_url": "https://omlx.ai/benchmarks/performance/l85i51rz",
   "tg_tokens": null,
   "tps": 16.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-rmbmjylq-ctx1024",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace, single run at 1k context; peak memory 5.8 GB; TTFT 7580 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 135.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 135.1 PP tok/s 14.0 TG tok/s 7580 TTFT (ms) 5.8 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rmbmjylq",
   "source_url": "https://omlx.ai/benchmarks/performance/rmbmjylq",
   "tg_tokens": null,
   "tps": 14.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-adk5fwh0-ctx1024",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace, single run at 1k context; peak memory 5.8 GB; TTFT 6932 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 147.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 147.7 PP tok/s 16.5 TG tok/s 6932 TTFT (ms) 5.8 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/adk5fwh0",
   "source_url": "https://omlx.ai/benchmarks/performance/adk5fwh0",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ddluekzr-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 45.9 GB; TTFT 226458 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 289.4,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 2bit; 64k tokens 289.4 PP tok/s 19.7 TG tok/s 226458 TTFT (ms) 45.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ddluekzr",
   "source_url": "https://omlx.ai/benchmarks/performance/ddluekzr",
   "tg_tokens": null,
   "tps": 19.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-dv0myc1x-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 53.0 GB at this context; duplicate traces of the same session: s0tafe7n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 349.4,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 2bit; 4k 349.4 36.1 53.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dv0myc1x",
   "source_url": "https://omlx.ai/benchmarks/performance/dv0myc1x",
   "tg_tokens": null,
   "tps": 36.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-dv0myc1x-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 53.6 GB at this context; duplicate traces of the same session: s0tafe7n; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 701.4,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 2bit; 16k 701.4 40.7 53.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dv0myc1x",
   "source_url": "https://omlx.ai/benchmarks/performance/dv0myc1x",
   "tg_tokens": null,
   "tps": 40.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-z1scewe4-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 41.9 GB; TTFT 121579 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 539.0,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 2bit; 64k tokens 539.0 PP tok/s 14.7 TG tok/s 121579 TTFT (ms) 41.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z1scewe4",
   "source_url": "https://omlx.ai/benchmarks/performance/z1scewe4",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-m7i30xm0-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 54.1 GB; TTFT 45899 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 713.9,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 2bit; 32k tokens 713.9 PP tok/s 34.8 TG tok/s 45899 TTFT (ms) 54.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m7i30xm0",
   "source_url": "https://omlx.ai/benchmarks/performance/m7i30xm0",
   "tg_tokens": null,
   "tps": 34.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-5tv7tgu4-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 53.6 GB; TTFT 25163 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 651.1,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 2bit; 16k tokens 651.1 PP tok/s 51.7 TG tok/s 25163 TTFT (ms) 53.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5tv7tgu4",
   "source_url": "https://omlx.ai/benchmarks/performance/5tv7tgu4",
   "tg_tokens": null,
   "tps": 51.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-k0h2t136-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 951.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 951.4 62.8 69.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k0h2t136",
   "source_url": "https://omlx.ai/benchmarks/performance/k0h2t136",
   "tg_tokens": null,
   "tps": 62.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-k0h2t136-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2844.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,844 75.8 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k0h2t136",
   "source_url": "https://omlx.ai/benchmarks/performance/k0h2t136",
   "tg_tokens": null,
   "tps": 75.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-k0h2t136-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2680.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,680 65.8 71.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k0h2t136",
   "source_url": "https://omlx.ai/benchmarks/performance/k0h2t136",
   "tg_tokens": null,
   "tps": 65.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-k0h2t136-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2406.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,406 63.1 72.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k0h2t136",
   "source_url": "https://omlx.ai/benchmarks/performance/k0h2t136",
   "tg_tokens": null,
   "tps": 63.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-l95fi25e-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.7 GB; TTFT 895 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1145.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,145 PP tok/s 85.6 TG tok/s 895 TTFT (ms) 70.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l95fi25e",
   "source_url": "https://omlx.ai/benchmarks/performance/l95fi25e",
   "tg_tokens": null,
   "tps": 85.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-40xedx99-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.2 GB at this context; duplicate traces of the same session: iie5558e, lpgxj6bx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1046.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,046 68.9 69.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/40xedx99",
   "source_url": "https://omlx.ai/benchmarks/performance/40xedx99",
   "tg_tokens": null,
   "tps": 68.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-40xedx99-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.2 GB at this context; duplicate traces of the same session: iie5558e, lpgxj6bx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2417.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,417 70.8 70.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/40xedx99",
   "source_url": "https://omlx.ai/benchmarks/performance/40xedx99",
   "tg_tokens": null,
   "tps": 70.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-40xedx99-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: iie5558e, lpgxj6bx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2817.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,817 74.5 70.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/40xedx99",
   "source_url": "https://omlx.ai/benchmarks/performance/40xedx99",
   "tg_tokens": null,
   "tps": 74.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-40xedx99-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.4 GB at this context; duplicate traces of the same session: iie5558e, lpgxj6bx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2686.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,686 68.8 71.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/40xedx99",
   "source_url": "https://omlx.ai/benchmarks/performance/40xedx99",
   "tg_tokens": null,
   "tps": 68.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-q9tgdx69-ctx131072",
   "model": "gemma-3-4b-it",
   "notes": "DWQ build; oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4192.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-3-4b-it-4bit-DWQ on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 4,192 92.2 5.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q9tgdx69",
   "source_url": "https://omlx.ai/benchmarks/performance/q9tgdx69",
   "tg_tokens": null,
   "tps": 92.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-q9tgdx69-ctx199680",
   "model": "gemma-3-4b-it",
   "notes": "DWQ build; oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3515.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-3-4b-it-4bit-DWQ on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 3,515 73.3 7.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q9tgdx69",
   "source_url": "https://omlx.ai/benchmarks/performance/q9tgdx69",
   "tg_tokens": null,
   "tps": 73.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-5dfdfa31-ctx1024",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "fp16 MLX build; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 25m56q0a, urfn25ym; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 662.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-fp16oqc on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 662.5 35.5 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5dfdfa31",
   "source_url": "https://omlx.ai/benchmarks/performance/5dfdfa31",
   "tg_tokens": null,
   "tps": 35.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-5dfdfa31-ctx4096",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "fp16 MLX build; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: 25m56q0a, urfn25ym; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 878.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-fp16oqc on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k 878.7 38.0 23.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5dfdfa31",
   "source_url": "https://omlx.ai/benchmarks/performance/5dfdfa31",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-5dfdfa31-ctx8192",
   "model": "Cyber-Tiel-Coder-35B-A3B",
   "notes": "fp16 MLX build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 25m56q0a, urfn25ym; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 830.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Cyber-Tiel-Coder-35B-A3B-MLX-oQ4e-fp16oqc on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 8k 830.2 52.6 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5dfdfa31",
   "source_url": "https://omlx.ai/benchmarks/performance/5dfdfa31",
   "tg_tokens": null,
   "tps": 52.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-hfa1vaa4-ctx1024",
   "model": "LFM2.5-8B-A1B",
   "notes": "OptiQ build; oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: 5z6wdse3, h8mnzzsd; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 620.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-8B-A1B-OptiQ-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 4bit; 1k 620.5 77.2 5.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hfa1vaa4",
   "source_url": "https://omlx.ai/benchmarks/performance/hfa1vaa4",
   "tg_tokens": null,
   "tps": 77.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-hfa1vaa4-ctx8192",
   "model": "LFM2.5-8B-A1B",
   "notes": "OptiQ build; oMLX community benchmark trace; peak memory 6.0 GB at this context; duplicate traces of the same session: 5z6wdse3, h8mnzzsd; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 682.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-8B-A1B-OptiQ-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 4bit; 8k 682.7 69.7 6.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hfa1vaa4",
   "source_url": "https://omlx.ai/benchmarks/performance/hfa1vaa4",
   "tg_tokens": null,
   "tps": 69.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-hfa1vaa4-ctx32768",
   "model": "LFM2.5-8B-A1B",
   "notes": "OptiQ build; oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: 5z6wdse3, h8mnzzsd; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 596.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-8B-A1B-OptiQ-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 4bit; 32k 596.7 53.6 6.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hfa1vaa4",
   "source_url": "https://omlx.ai/benchmarks/performance/hfa1vaa4",
   "tg_tokens": null,
   "tps": 53.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-hfa1vaa4-ctx65536",
   "model": "LFM2.5-8B-A1B",
   "notes": "OptiQ build; oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: 5z6wdse3, h8mnzzsd; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 507.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-8B-A1B-OptiQ-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 4bit; 64k 507.7 42.2 6.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hfa1vaa4",
   "source_url": "https://omlx.ai/benchmarks/performance/hfa1vaa4",
   "tg_tokens": null,
   "tps": 42.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-avn4792a-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: 6k5edx00, o6w4bwns, shrz0nbx, si43iui4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1707.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-fixed-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 1,707 126.6 37.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/avn4792a",
   "source_url": "https://omlx.ai/benchmarks/performance/avn4792a",
   "tg_tokens": null,
   "tps": 126.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-avn4792a-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.1 GB at this context; duplicate traces of the same session: 6k5edx00, o6w4bwns, shrz0nbx, si43iui4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2015.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-fixed-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 2,015 110.0 39.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/avn4792a",
   "source_url": "https://omlx.ai/benchmarks/performance/avn4792a",
   "tg_tokens": null,
   "tps": 110.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-avn4792a-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: 6k5edx00, o6w4bwns, shrz0nbx, si43iui4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1905.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-fixed-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 8k 1,905 127.1 39.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/avn4792a",
   "source_url": "https://omlx.ai/benchmarks/performance/avn4792a",
   "tg_tokens": null,
   "tps": 127.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-avn4792a-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.7 GB at this context; duplicate traces of the same session: 6k5edx00, o6w4bwns, shrz0nbx, si43iui4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1699.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-fixed-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 16k 1,699 146.1 39.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/avn4792a",
   "source_url": "https://omlx.ai/benchmarks/performance/avn4792a",
   "tg_tokens": null,
   "tps": 146.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-avn4792a-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 40.4 GB at this context; duplicate traces of the same session: 6k5edx00, o6w4bwns, shrz0nbx, si43iui4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1384.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ornith-1.5-35B-A3B-oQ8e-fixed-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 32k 1,384 128.2 40.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/avn4792a",
   "source_url": "https://omlx.ai/benchmarks/performance/avn4792a",
   "tg_tokens": null,
   "tps": 128.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-21o7cacx-ctx1024",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 78zkpuog; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 111.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 111.2 20.0 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/21o7cacx",
   "source_url": "https://omlx.ai/benchmarks/performance/21o7cacx",
   "tg_tokens": null,
   "tps": 20.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-21o7cacx-ctx4096",
   "model": "Qwen3.5-27B-Claude-4.6-Opus-Distilled",
   "notes": "oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: 78zkpuog; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-27B-Claude-4.6-Opus-Distilled-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 103.3 19.4 19.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/21o7cacx",
   "source_url": "https://omlx.ai/benchmarks/performance/21o7cacx",
   "tg_tokens": null,
   "tps": 19.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-r238xsxt-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: rl4w2lye; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1004.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,004 26.4 21.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r238xsxt",
   "source_url": "https://omlx.ai/benchmarks/performance/r238xsxt",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-r238xsxt-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: rl4w2lye; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 748.3,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 748.3 24.8 21.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r238xsxt",
   "source_url": "https://omlx.ai/benchmarks/performance/r238xsxt",
   "tg_tokens": null,
   "tps": 24.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mdnz41rh-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 667.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 667.3 35.3 32.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mdnz41rh",
   "source_url": "https://omlx.ai/benchmarks/performance/mdnz41rh",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mdnz41rh-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 33.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 772.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 772.4 33.3 33.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mdnz41rh",
   "source_url": "https://omlx.ai/benchmarks/performance/mdnz41rh",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-x3grul5x-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 81d3dljo, 8at2f1l3, gjlov1hw, vr97a6sm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 252.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 252.2 33.3 23.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x3grul5x",
   "source_url": "https://omlx.ai/benchmarks/performance/x3grul5x",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-x3grul5x-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: 81d3dljo, 8at2f1l3, gjlov1hw, vr97a6sm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 233.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 233.7 107.3 25.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x3grul5x",
   "source_url": "https://omlx.ai/benchmarks/performance/x3grul5x",
   "tg_tokens": null,
   "tps": 107.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-x3grul5x-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: 81d3dljo, 8at2f1l3, gjlov1hw, vr97a6sm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 204.0 46.1 27.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x3grul5x",
   "source_url": "https://omlx.ai/benchmarks/performance/x3grul5x",
   "tg_tokens": null,
   "tps": 46.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-x3grul5x-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.3 GB at this context; duplicate traces of the same session: 81d3dljo, 8at2f1l3, gjlov1hw, vr97a6sm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 162.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 162.8 41.7 32.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x3grul5x",
   "source_url": "https://omlx.ai/benchmarks/performance/x3grul5x",
   "tg_tokens": null,
   "tps": 41.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-x3grul5x-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: 81d3dljo, 8at2f1l3, gjlov1hw, vr97a6sm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 134.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 134.2 28.2 37.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x3grul5x",
   "source_url": "https://omlx.ai/benchmarks/performance/x3grul5x",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-zqxtfmau-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: amcvbev2, e5upbna3, t4s0utth, xvn1zmax; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 197.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k 197.5 17.5 17.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zqxtfmau",
   "source_url": "https://omlx.ai/benchmarks/performance/zqxtfmau",
   "tg_tokens": null,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-zqxtfmau-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: amcvbev2, e5upbna3, t4s0utth, xvn1zmax; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 199.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 199.6 17.7 18.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zqxtfmau",
   "source_url": "https://omlx.ai/benchmarks/performance/zqxtfmau",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-zqxtfmau-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: amcvbev2, e5upbna3, t4s0utth, xvn1zmax; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 198.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 8k 198.0 15.2 18.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zqxtfmau",
   "source_url": "https://omlx.ai/benchmarks/performance/zqxtfmau",
   "tg_tokens": null,
   "tps": 15.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-zqxtfmau-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: amcvbev2, e5upbna3, t4s0utth, xvn1zmax; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 191.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 16k 191.0 16.3 19.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zqxtfmau",
   "source_url": "https://omlx.ai/benchmarks/performance/zqxtfmau",
   "tg_tokens": null,
   "tps": 16.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-zqxtfmau-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: amcvbev2, e5upbna3, t4s0utth, xvn1zmax; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 182.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 32k 182.6 16.8 20.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zqxtfmau",
   "source_url": "https://omlx.ai/benchmarks/performance/zqxtfmau",
   "tg_tokens": null,
   "tps": 16.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cr117w42-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: 2vggab53; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 913.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 913.3 58.0 17.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cr117w42",
   "source_url": "https://omlx.ai/benchmarks/performance/cr117w42",
   "tg_tokens": null,
   "tps": 58.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cr117w42-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: 2vggab53; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 911.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 911.6 78.6 22.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cr117w42",
   "source_url": "https://omlx.ai/benchmarks/performance/cr117w42",
   "tg_tokens": null,
   "tps": 78.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ea2mdckp-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 660.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 660.1 32.9 32.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ea2mdckp",
   "source_url": "https://omlx.ai/benchmarks/performance/ea2mdckp",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ea2mdckp-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 747.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 747.3 35.7 33.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ea2mdckp",
   "source_url": "https://omlx.ai/benchmarks/performance/ea2mdckp",
   "tg_tokens": null,
   "tps": 35.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-oohoxwbi-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: 9lmfl0x1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 444.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 444.4 31.2 32.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oohoxwbi",
   "source_url": "https://omlx.ai/benchmarks/performance/oohoxwbi",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-oohoxwbi-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.2 GB at this context; duplicate traces of the same session: 9lmfl0x1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 543.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 543.7 27.1 33.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oohoxwbi",
   "source_url": "https://omlx.ai/benchmarks/performance/oohoxwbi",
   "tg_tokens": null,
   "tps": 27.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-h7ypekps-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 52.9 GB; TTFT 7040 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 581.8,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 2bit; 4k tokens 581.8 PP tok/s 62.0 TG tok/s 7040 TTFT (ms) 52.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h7ypekps",
   "source_url": "https://omlx.ai/benchmarks/performance/h7ypekps",
   "tg_tokens": null,
   "tps": 62.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-2zxc3za4-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 53.3 GB at this context; duplicate traces of the same session: poutvzr0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 280.6,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 2bit; 8k 280.6 57.1 53.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2zxc3za4",
   "source_url": "https://omlx.ai/benchmarks/performance/2zxc3za4",
   "tg_tokens": null,
   "tps": 57.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-2zxc3za4-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 53.6 GB at this context; duplicate traces of the same session: poutvzr0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 673.1,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 2bit; 16k 673.1 60.7 53.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2zxc3za4",
   "source_url": "https://omlx.ai/benchmarks/performance/2zxc3za4",
   "tg_tokens": null,
   "tps": 60.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-jrish4bm-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 55.2 GB; TTFT 130213 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 503.3,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 2bit; 64k tokens 503.3 PP tok/s 37.5 TG tok/s 130213 TTFT (ms) 55.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jrish4bm",
   "source_url": "https://omlx.ai/benchmarks/performance/jrish4bm",
   "tg_tokens": null,
   "tps": 37.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-47huth2a-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 53.6 GB; TTFT 23987 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 683.0,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 2bit; 16k tokens 683.0 PP tok/s 50.6 TG tok/s 23987 TTFT (ms) 53.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/47huth2a",
   "source_url": "https://omlx.ai/benchmarks/performance/47huth2a",
   "tg_tokens": null,
   "tps": 50.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-zo9fhy5e-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 100.7 GB; TTFT 1120 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 914.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k tokens 914.7 PP tok/s 106.1 TG tok/s 1120 TTFT (ms) 100.7 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zo9fhy5e",
   "source_url": "https://omlx.ai/benchmarks/performance/zo9fhy5e",
   "tg_tokens": null,
   "tps": 106.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-id80oeri-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1046.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,046 68.9 69.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/id80oeri",
   "source_url": "https://omlx.ai/benchmarks/performance/id80oeri",
   "tg_tokens": null,
   "tps": 68.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-id80oeri-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2417.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,417 70.8 70.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/id80oeri",
   "source_url": "https://omlx.ai/benchmarks/performance/id80oeri",
   "tg_tokens": null,
   "tps": 70.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-id80oeri-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2817.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,817 74.5 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/id80oeri",
   "source_url": "https://omlx.ai/benchmarks/performance/id80oeri",
   "tg_tokens": null,
   "tps": 74.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-id80oeri-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2686.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,686 68.8 71.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/id80oeri",
   "source_url": "https://omlx.ai/benchmarks/performance/id80oeri",
   "tg_tokens": null,
   "tps": 68.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-cagxw492-ctx131072",
   "model": "gemma-3-4b-it",
   "notes": "DWQ build; oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4192.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-3-4b-it-4bit-DWQ on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 4,192 92.2 5.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cagxw492",
   "source_url": "https://omlx.ai/benchmarks/performance/cagxw492",
   "tg_tokens": null,
   "tps": 92.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-cagxw492-ctx199680",
   "model": "gemma-3-4b-it",
   "notes": "DWQ build; oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3515.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-3-4b-it-4bit-DWQ on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 3,515 73.3 7.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cagxw492",
   "source_url": "https://omlx.ai/benchmarks/performance/cagxw492",
   "tg_tokens": null,
   "tps": 73.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-jmzx7iyl-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: yuev2jhl; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 792.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 792.7 51.5 14.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jmzx7iyl",
   "source_url": "https://omlx.ai/benchmarks/performance/jmzx7iyl",
   "tg_tokens": null,
   "tps": 51.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-jmzx7iyl-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.5 GB at this context; duplicate traces of the same session: yuev2jhl; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1341.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,341 46.3 14.5 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jmzx7iyl",
   "source_url": "https://omlx.ai/benchmarks/performance/jmzx7iyl",
   "tg_tokens": null,
   "tps": 46.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-7zv23vxh-ctx1024",
   "model": "Artemis-31B",
   "notes": "oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 6y9jn262, ucl3rtgn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 722.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Artemis-31B-v1.1-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 722.9 25.8 18.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7zv23vxh",
   "source_url": "https://omlx.ai/benchmarks/performance/7zv23vxh",
   "tg_tokens": null,
   "tps": 25.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-7zv23vxh-ctx4096",
   "model": "Artemis-31B",
   "notes": "oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: 6y9jn262, ucl3rtgn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 715.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Artemis-31B-v1.1-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 715.1 25.0 19.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7zv23vxh",
   "source_url": "https://omlx.ai/benchmarks/performance/7zv23vxh",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-7zv23vxh-ctx16384",
   "model": "Artemis-31B",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 6y9jn262, ucl3rtgn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 583.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Artemis-31B-v1.1-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 583.8 22.1 21.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7zv23vxh",
   "source_url": "https://omlx.ai/benchmarks/performance/7zv23vxh",
   "tg_tokens": null,
   "tps": 22.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-k5vymikb-ctx1024",
   "model": "Bonsai-8B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 2.5 GB at this context; duplicate traces of the same session: 1pxegc25; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 163.1,
   "provenance": "community",
   "quant": "1bit",
   "quote": "Bonsai-8B-mlx-1bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 unknown; 1k 163.1 45.2 2.5 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k5vymikb",
   "source_url": "https://omlx.ai/benchmarks/performance/k5vymikb",
   "tg_tokens": null,
   "tps": 45.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-k5vymikb-ctx4096",
   "model": "Bonsai-8B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 3.0 GB at this context; duplicate traces of the same session: 1pxegc25; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 190.7,
   "provenance": "community",
   "quant": "1bit",
   "quote": "Bonsai-8B-mlx-1bit on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 unknown; 4k 190.7 37.5 3.0 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k5vymikb",
   "source_url": "https://omlx.ai/benchmarks/performance/k5vymikb",
   "tg_tokens": null,
   "tps": 37.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-t6tpijop-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 150.7 GB at this context; duplicate traces of the same session: ibwowbrk, kuq4lhkg; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 251.0,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "GLM-5.3-Flash-oQ3.5e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 3bit; 1k 251.0 35.8 150.7 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t6tpijop",
   "source_url": "https://omlx.ai/benchmarks/performance/t6tpijop",
   "tg_tokens": null,
   "tps": 35.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-t6tpijop-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 151.8 GB at this context; duplicate traces of the same session: ibwowbrk, kuq4lhkg; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 301.1,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "GLM-5.3-Flash-oQ3.5e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 3bit; 4k 301.1 32.3 151.8 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t6tpijop",
   "source_url": "https://omlx.ai/benchmarks/performance/t6tpijop",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-t6tpijop-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 152.2 GB at this context; duplicate traces of the same session: ibwowbrk, kuq4lhkg; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 253.9,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "GLM-5.3-Flash-oQ3.5e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 3bit; 16k 253.9 59.8 152.2 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t6tpijop",
   "source_url": "https://omlx.ai/benchmarks/performance/t6tpijop",
   "tg_tokens": null,
   "tps": 59.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-spiqioj0-ctx1024",
   "model": "LFM2.5-8B-A1B",
   "notes": "OptiQ build; oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: none; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 620.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-8B-A1B-OptiQ-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 4bit; 1k 620.5 77.2 5.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/spiqioj0",
   "source_url": "https://omlx.ai/benchmarks/performance/spiqioj0",
   "tg_tokens": null,
   "tps": 77.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-spiqioj0-ctx8192",
   "model": "LFM2.5-8B-A1B",
   "notes": "OptiQ build; oMLX community benchmark trace; peak memory 6.0 GB at this context; duplicate traces of the same session: none; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 682.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-8B-A1B-OptiQ-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 4bit; 8k 682.7 69.7 6.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/spiqioj0",
   "source_url": "https://omlx.ai/benchmarks/performance/spiqioj0",
   "tg_tokens": null,
   "tps": 69.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-spiqioj0-ctx32768",
   "model": "LFM2.5-8B-A1B",
   "notes": "OptiQ build; oMLX community benchmark trace; peak memory 6.3 GB at this context; duplicate traces of the same session: none; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 596.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-8B-A1B-OptiQ-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 4bit; 32k 596.7 53.6 6.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/spiqioj0",
   "source_url": "https://omlx.ai/benchmarks/performance/spiqioj0",
   "tg_tokens": null,
   "tps": 53.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-spiqioj0-ctx65536",
   "model": "LFM2.5-8B-A1B",
   "notes": "OptiQ build; oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: none; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 507.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-8B-A1B-OptiQ-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 4bit; 64k 507.7 42.2 6.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/spiqioj0",
   "source_url": "https://omlx.ai/benchmarks/performance/spiqioj0",
   "tg_tokens": null,
   "tps": 42.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-cz2mpykw-ctx1024",
   "model": "Ling-mini-2.0",
   "notes": "oMLX community benchmark trace; peak memory 9.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2404.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ling-mini-2.0-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 2,404 202.5 9.0 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cz2mpykw",
   "source_url": "https://omlx.ai/benchmarks/performance/cz2mpykw",
   "tg_tokens": null,
   "tps": 202.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-cz2mpykw-ctx16384",
   "model": "Ling-mini-2.0",
   "notes": "oMLX community benchmark trace; peak memory 9.8 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1645.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ling-mini-2.0-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 16k 1,645 129.8 9.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cz2mpykw",
   "source_url": "https://omlx.ai/benchmarks/performance/cz2mpykw",
   "tg_tokens": null,
   "tps": 129.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-57m331yk-ctx1024",
   "model": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B",
   "notes": "oMLX community benchmark trace; peak memory 54.0 GB at this context; duplicate traces of the same session: 503i8kls; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 495.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B-5bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 495.0 54.1 54.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/57m331yk",
   "source_url": "https://omlx.ai/benchmarks/performance/57m331yk",
   "tg_tokens": null,
   "tps": 54.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-57m331yk-ctx4096",
   "model": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B",
   "notes": "oMLX community benchmark trace; peak memory 57.0 GB at this context; duplicate traces of the same session: 503i8kls; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 552.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B-5bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 552.0 53.6 57.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/57m331yk",
   "source_url": "https://omlx.ai/benchmarks/performance/57m331yk",
   "tg_tokens": null,
   "tps": 53.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-eh9j90qs-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: c7fvn060; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 929.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 929.5 32.7 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eh9j90qs",
   "source_url": "https://omlx.ai/benchmarks/performance/eh9j90qs",
   "tg_tokens": null,
   "tps": 32.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-eh9j90qs-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: c7fvn060; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1001.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,001 32.1 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eh9j90qs",
   "source_url": "https://omlx.ai/benchmarks/performance/eh9j90qs",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-xdmmzewj-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: 3e6v9rox; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 499.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 499.2 17.4 17.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xdmmzewj",
   "source_url": "https://omlx.ai/benchmarks/performance/xdmmzewj",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-xdmmzewj-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 3e6v9rox; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 475.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 475.9 16.3 17.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xdmmzewj",
   "source_url": "https://omlx.ai/benchmarks/performance/xdmmzewj",
   "tg_tokens": null,
   "tps": 16.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-lrwpzdxo-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 1e8wx4jj, p07gl3xp, s47xv8c2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1463.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 1,463 40.2 22.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lrwpzdxo",
   "source_url": "https://omlx.ai/benchmarks/performance/lrwpzdxo",
   "tg_tokens": null,
   "tps": 40.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-lrwpzdxo-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: 1e8wx4jj, p07gl3xp, s47xv8c2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1720.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 1,720 39.9 23.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lrwpzdxo",
   "source_url": "https://omlx.ai/benchmarks/performance/lrwpzdxo",
   "tg_tokens": null,
   "tps": 39.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-lrwpzdxo-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: 1e8wx4jj, p07gl3xp, s47xv8c2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1727.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 8k 1,727 39.5 23.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lrwpzdxo",
   "source_url": "https://omlx.ai/benchmarks/performance/lrwpzdxo",
   "tg_tokens": null,
   "tps": 39.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-lrwpzdxo-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: 1e8wx4jj, p07gl3xp, s47xv8c2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1681.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 16k 1,681 38.4 24.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lrwpzdxo",
   "source_url": "https://omlx.ai/benchmarks/performance/lrwpzdxo",
   "tg_tokens": null,
   "tps": 38.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-drjd0buj-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: ixhcbam0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1071.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 1,071 40.2 22.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/drjd0buj",
   "source_url": "https://omlx.ai/benchmarks/performance/drjd0buj",
   "tg_tokens": null,
   "tps": 40.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-drjd0buj-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: ixhcbam0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1544.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 1,544 39.8 23.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/drjd0buj",
   "source_url": "https://omlx.ai/benchmarks/performance/drjd0buj",
   "tg_tokens": null,
   "tps": 39.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ofmq2pys-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 6cp6rm2r; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 469.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 469.5 36.1 17.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ofmq2pys",
   "source_url": "https://omlx.ai/benchmarks/performance/ofmq2pys",
   "tg_tokens": null,
   "tps": 36.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ofmq2pys-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: 6cp6rm2r; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 478.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 478.9 44.8 18.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ofmq2pys",
   "source_url": "https://omlx.ai/benchmarks/performance/ofmq2pys",
   "tg_tokens": null,
   "tps": 44.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-dvy355pf-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 457.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 457.3 33.6 17.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dvy355pf",
   "source_url": "https://omlx.ai/benchmarks/performance/dvy355pf",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-dvy355pf-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 464.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 464.6 36.2 18.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dvy355pf",
   "source_url": "https://omlx.ai/benchmarks/performance/dvy355pf",
   "tg_tokens": null,
   "tps": 36.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-z0ijk300-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.0 GB at this context; duplicate traces of the same session: ti5iza11; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1158.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 1k 1,158 73.4 29.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z0ijk300",
   "source_url": "https://omlx.ai/benchmarks/performance/z0ijk300",
   "tg_tokens": null,
   "tps": 73.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-z0ijk300-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.0 GB at this context; duplicate traces of the same session: ti5iza11; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1406.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 8bit; 4k 1,406 65.2 30.0 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z0ijk300",
   "source_url": "https://omlx.ai/benchmarks/performance/z0ijk300",
   "tg_tokens": null,
   "tps": 65.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-k1hnjeh8-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 53.7 GB; TTFT 22459 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 729.5,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 2bit; 16k tokens 729.5 PP tok/s 46.1 TG tok/s 22459 TTFT (ms) 53.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k1hnjeh8",
   "source_url": "https://omlx.ai/benchmarks/performance/k1hnjeh8",
   "tg_tokens": null,
   "tps": 46.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-62nqhxl1-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 55.0 GB; TTFT 148971 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 439.9,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 2bit; 64k tokens 439.9 PP tok/s 45.8 TG tok/s 148971 TTFT (ms) 55.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/62nqhxl1",
   "source_url": "https://omlx.ai/benchmarks/performance/62nqhxl1",
   "tg_tokens": null,
   "tps": 45.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-wkhjc803-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 54.1 GB; TTFT 65312 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 501.7,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 2bit; 32k tokens 501.7 PP tok/s 36.1 TG tok/s 65312 TTFT (ms) 54.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wkhjc803",
   "source_url": "https://omlx.ai/benchmarks/performance/wkhjc803",
   "tg_tokens": null,
   "tps": 36.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-l8mich3k-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.7 GB; TTFT 2080 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 492.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 492.3 PP tok/s 46.0 TG tok/s 2080 TTFT (ms) 70.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l8mich3k",
   "source_url": "https://omlx.ai/benchmarks/performance/l8mich3k",
   "tg_tokens": null,
   "tps": 46.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-g6z84bc3-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled, BF16 ngram; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: hwbxxkn8, x3z7zuli, z2olk0l2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 693.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-BF16-ngram on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 693.6 84.2 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g6z84bc3",
   "source_url": "https://omlx.ai/benchmarks/performance/g6z84bc3",
   "tg_tokens": null,
   "tps": 84.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-g6z84bc3-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled, BF16 ngram; oMLX community benchmark trace; peak memory 71.9 GB at this context; duplicate traces of the same session: hwbxxkn8, x3z7zuli, z2olk0l2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 810.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-BF16-ngram on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 810.6 73.3 71.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g6z84bc3",
   "source_url": "https://omlx.ai/benchmarks/performance/g6z84bc3",
   "tg_tokens": null,
   "tps": 73.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-g6z84bc3-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled, BF16 ngram; oMLX community benchmark trace; peak memory 73.7 GB at this context; duplicate traces of the same session: hwbxxkn8, x3z7zuli, z2olk0l2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 773.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-BF16-ngram on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 773.3 71.3 73.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g6z84bc3",
   "source_url": "https://omlx.ai/benchmarks/performance/g6z84bc3",
   "tg_tokens": null,
   "tps": 71.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-g6z84bc3-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled, BF16 ngram; oMLX community benchmark trace; peak memory 76.1 GB at this context; duplicate traces of the same session: hwbxxkn8, x3z7zuli, z2olk0l2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 764.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-BF16-ngram on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 764.7 56.2 76.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g6z84bc3",
   "source_url": "https://omlx.ai/benchmarks/performance/g6z84bc3",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-sj6tf4g2-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.8 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1356.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 16k 1,356 70.5 143.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sj6tf4g2",
   "source_url": "https://omlx.ai/benchmarks/performance/sj6tf4g2",
   "tg_tokens": null,
   "tps": 70.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-sj6tf4g2-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 147.8 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1311.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 128k 1,311 64.8 147.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sj6tf4g2",
   "source_url": "https://omlx.ai/benchmarks/performance/sj6tf4g2",
   "tg_tokens": null,
   "tps": 64.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-qhd31b8h-ctx1024",
   "model": "Ternary-Bonsai-2-9B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 5.7 GB at this context; duplicate traces of the same session: 3ai7amf3; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 320.2,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-9B-mlx-2bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 2bit; 1k 320.2 52.6 5.7 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qhd31b8h",
   "source_url": "https://omlx.ai/benchmarks/performance/qhd31b8h",
   "tg_tokens": null,
   "tps": 52.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-qhd31b8h-ctx16384",
   "model": "Ternary-Bonsai-2-9B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 7.7 GB at this context; duplicate traces of the same session: 3ai7amf3; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 283.4,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-9B-mlx-2bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 2bit; 16k 283.4 27.0 7.7 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qhd31b8h",
   "source_url": "https://omlx.ai/benchmarks/performance/qhd31b8h",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-xonvcz10-ctx1024",
   "model": "Ternary-Bonsai-2-9B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 1k context; peak memory 5.7 GB; TTFT 3357 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 305.0,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-9B-mlx-2bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 2bit; 1k tokens 305.0 PP tok/s 52.7 TG tok/s 3357 TTFT (ms) 5.7 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xonvcz10",
   "source_url": "https://omlx.ai/benchmarks/performance/xonvcz10",
   "tg_tokens": null,
   "tps": 52.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-rgtg8vi5-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.7 GB at this context; duplicate traces of the same session: 2hw0o5uh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2346.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 2,346 74.6 14.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rgtg8vi5",
   "source_url": "https://omlx.ai/benchmarks/performance/rgtg8vi5",
   "tg_tokens": null,
   "tps": 74.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-rgtg8vi5-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 15.2 GB at this context; duplicate traces of the same session: 2hw0o5uh; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2151.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 2,151 62.5 15.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rgtg8vi5",
   "source_url": "https://omlx.ai/benchmarks/performance/rgtg8vi5",
   "tg_tokens": null,
   "tps": 62.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-gdmbz7lb-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 5.7 GB at this context; duplicate traces of the same session: r9jxsm4l; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1089.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M5 (8c); M5 (8c) \u00b7 16 GB \u00b7 4bit; 1k 1,089 41.6 5.7 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gdmbz7lb",
   "source_url": "https://omlx.ai/benchmarks/performance/gdmbz7lb",
   "tg_tokens": null,
   "tps": 41.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-gdmbz7lb-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: r9jxsm4l; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1565.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M5 (8c); M5 (8c) \u00b7 16 GB \u00b7 4bit; 4k 1,565 39.6 5.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gdmbz7lb",
   "source_url": "https://omlx.ai/benchmarks/performance/gdmbz7lb",
   "tg_tokens": null,
   "tps": 39.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-0t7z5rfg-ctx4096",
   "model": "LFM2.5-8B-A1B",
   "notes": "OptiQ build; oMLX community benchmark trace; peak memory 6.0 GB at this context; duplicate traces of the same session: p3qvcpwt; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 685.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-8B-A1B-OptiQ-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 4bit; 4k 685.6 74.6 6.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0t7z5rfg",
   "source_url": "https://omlx.ai/benchmarks/performance/0t7z5rfg",
   "tg_tokens": null,
   "tps": 74.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-0t7z5rfg-ctx16384",
   "model": "LFM2.5-8B-A1B",
   "notes": "OptiQ build; oMLX community benchmark trace; peak memory 6.1 GB at this context; duplicate traces of the same session: p3qvcpwt; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 643.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-8B-A1B-OptiQ-4bit on M1 Pro (14c); M1 Pro (14c) \u00b7 32 GB \u00b7 4bit; 16k 643.1 64.0 6.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0t7z5rfg",
   "source_url": "https://omlx.ai/benchmarks/performance/0t7z5rfg",
   "tg_tokens": null,
   "tps": 64.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-1oabwab4-ctx1024",
   "model": "Ling-3.0-tiny",
   "notes": "bf16; MLX; oMLX community benchmark trace; peak memory 8.7 GB at this context; duplicate traces of the same session: c1111p88; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1782.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ling-3.0-tiny-oQ8e-bf16-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 8bit; 1k 1,782 108.5 8.7 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1oabwab4",
   "source_url": "https://omlx.ai/benchmarks/performance/1oabwab4",
   "tg_tokens": null,
   "tps": 108.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-1oabwab4-ctx16384",
   "model": "Ling-3.0-tiny",
   "notes": "bf16; MLX; oMLX community benchmark trace; peak memory 10.0 GB at this context; duplicate traces of the same session: c1111p88; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1700.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Ling-3.0-tiny-oQ8e-bf16-MLX on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 8bit; 16k 1,700 93.7 10.0 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1oabwab4",
   "source_url": "https://omlx.ai/benchmarks/performance/1oabwab4",
   "tg_tokens": null,
   "tps": 93.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-j2co65r2-ctx1024",
   "model": "Ling-mini-2.0",
   "notes": "oMLX community benchmark trace; peak memory 9.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2404.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ling-mini-2.0-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 2,404 202.5 9.0 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j2co65r2",
   "source_url": "https://omlx.ai/benchmarks/performance/j2co65r2",
   "tg_tokens": null,
   "tps": 202.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-j2co65r2-ctx16384",
   "model": "Ling-mini-2.0",
   "notes": "oMLX community benchmark trace; peak memory 9.8 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1645.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ling-mini-2.0-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 16k 1,645 129.8 9.8 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j2co65r2",
   "source_url": "https://omlx.ai/benchmarks/performance/j2co65r2",
   "tg_tokens": null,
   "tps": 129.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9fn26lxh-ctx1024",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4588.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,588 135.1 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9fn26lxh",
   "source_url": "https://omlx.ai/benchmarks/performance/9fn26lxh",
   "tg_tokens": null,
   "tps": 135.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9fn26lxh-ctx4096",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4588.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 4,588 132.3 20.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9fn26lxh",
   "source_url": "https://omlx.ai/benchmarks/performance/9fn26lxh",
   "tg_tokens": null,
   "tps": 132.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-b2sbtkq9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: vjvz3olf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 897.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 897.1 29.1 17.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b2sbtkq9",
   "source_url": "https://omlx.ai/benchmarks/performance/b2sbtkq9",
   "tg_tokens": null,
   "tps": 29.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-b2sbtkq9-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: vjvz3olf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 776.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 776.2 27.6 17.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b2sbtkq9",
   "source_url": "https://omlx.ai/benchmarks/performance/b2sbtkq9",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-0z88lc4o-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 259.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 259.9 16.5 31.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0z88lc4o",
   "source_url": "https://omlx.ai/benchmarks/performance/0z88lc4o",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-0z88lc4o-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 247.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 16k 247.8 16.0 33.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0z88lc4o",
   "source_url": "https://omlx.ai/benchmarks/performance/0z88lc4o",
   "tg_tokens": null,
   "tps": 16.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-p4t3b01d-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 23.9 GB; TTFT 50711 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 80.8,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k tokens 80.8 PP tok/s 13.5 TG tok/s 50711 TTFT (ms) 23.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p4t3b01d",
   "source_url": "https://omlx.ai/benchmarks/performance/p4t3b01d",
   "tg_tokens": null,
   "tps": 13.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ugt4ghom-ctx1024",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: 750mx8dd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 860.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-abliterated-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 860.4 63.1 21.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ugt4ghom",
   "source_url": "https://omlx.ai/benchmarks/performance/ugt4ghom",
   "tg_tokens": null,
   "tps": 63.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ugt4ghom-ctx4096",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: 750mx8dd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 730.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-abliterated-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 730.5 65.2 22.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ugt4ghom",
   "source_url": "https://omlx.ai/benchmarks/performance/ugt4ghom",
   "tg_tokens": null,
   "tps": 65.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vr3fd8f3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 3d1u0vue, hxrlg12w, kp2wd3mj, mtv7hrv1, nx7fejqa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1101.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,101 68.6 17.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vr3fd8f3",
   "source_url": "https://omlx.ai/benchmarks/performance/vr3fd8f3",
   "tg_tokens": null,
   "tps": 68.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-vr3fd8f3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 3d1u0vue, hxrlg12w, kp2wd3mj, mtv7hrv1, nx7fejqa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1217.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,217 95.1 19.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vr3fd8f3",
   "source_url": "https://omlx.ai/benchmarks/performance/vr3fd8f3",
   "tg_tokens": null,
   "tps": 95.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vr3fd8f3-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 3d1u0vue, hxrlg12w, kp2wd3mj, mtv7hrv1, nx7fejqa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1083.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,083 38.9 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vr3fd8f3",
   "source_url": "https://omlx.ai/benchmarks/performance/vr3fd8f3",
   "tg_tokens": null,
   "tps": 38.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vr3fd8f3-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 3d1u0vue, hxrlg12w, kp2wd3mj, mtv7hrv1, nx7fejqa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1033.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,033 49.3 19.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vr3fd8f3",
   "source_url": "https://omlx.ai/benchmarks/performance/vr3fd8f3",
   "tg_tokens": null,
   "tps": 49.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vr3fd8f3-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: 3d1u0vue, hxrlg12w, kp2wd3mj, mtv7hrv1, nx7fejqa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 863.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 863.7 50.7 21.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vr3fd8f3",
   "source_url": "https://omlx.ai/benchmarks/performance/vr3fd8f3",
   "tg_tokens": null,
   "tps": 50.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vr3fd8f3-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 3d1u0vue, hxrlg12w, kp2wd3mj, mtv7hrv1, nx7fejqa; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 696.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 696.0 27.0 23.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vr3fd8f3",
   "source_url": "https://omlx.ai/benchmarks/performance/vr3fd8f3",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9j22w9tk-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 905.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 905.1 69.6 17.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9j22w9tk",
   "source_url": "https://omlx.ai/benchmarks/performance/9j22w9tk",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9j22w9tk-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 993.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 993.6 65.6 18.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9j22w9tk",
   "source_url": "https://omlx.ai/benchmarks/performance/9j22w9tk",
   "tg_tokens": null,
   "tps": 65.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-9o9tsv60-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 457.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 457.3 33.6 17.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9o9tsv60",
   "source_url": "https://omlx.ai/benchmarks/performance/9o9tsv60",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-9o9tsv60-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 464.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 464.6 36.2 18.7 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9o9tsv60",
   "source_url": "https://omlx.ai/benchmarks/performance/9o9tsv60",
   "tg_tokens": null,
   "tps": 36.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ob86dspj-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "fp16 VLM MTP build; oMLX community benchmark trace; peak memory 32.6 GB at this context; duplicate traces of the same session: psesxr5s; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 118.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-27B-oQ8-fp16-VLM-MTP on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 4k 118.0 10.8 32.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ob86dspj",
   "source_url": "https://omlx.ai/benchmarks/performance/ob86dspj",
   "tg_tokens": null,
   "tps": 10.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-ob86dspj-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "fp16 VLM MTP build; oMLX community benchmark trace; peak memory 34.2 GB at this context; duplicate traces of the same session: psesxr5s; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 111.3,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-27B-oQ8-fp16-VLM-MTP on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 16k 111.3 10.3 34.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ob86dspj",
   "source_url": "https://omlx.ai/benchmarks/performance/ob86dspj",
   "tg_tokens": null,
   "tps": 10.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q6hnyuds-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.2 GB at this context; duplicate traces of the same session: 7yewh50f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 778.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 778.5 38.4 39.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q6hnyuds",
   "source_url": "https://omlx.ai/benchmarks/performance/q6hnyuds",
   "tg_tokens": null,
   "tps": 38.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-q6hnyuds-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 40.2 GB at this context; duplicate traces of the same session: 7yewh50f; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 699.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 699.4 36.6 40.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q6hnyuds",
   "source_url": "https://omlx.ai/benchmarks/performance/q6hnyuds",
   "tg_tokens": null,
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ktlmw25a-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 52.2 GB at this context; duplicate traces of the same session: xryn5axr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 433.6,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 2bit; 1k 433.6 43.6 52.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ktlmw25a",
   "source_url": "https://omlx.ai/benchmarks/performance/ktlmw25a",
   "tg_tokens": null,
   "tps": 43.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-ktlmw25a-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 55.2 GB at this context; duplicate traces of the same session: xryn5axr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 434.0,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 2bit; 64k 434.0 47.4 55.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ktlmw25a",
   "source_url": "https://omlx.ai/benchmarks/performance/ktlmw25a",
   "tg_tokens": null,
   "tps": 47.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-j6smdf4o-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 52.2 GB at this context; duplicate traces of the same session: cb32t2ku; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 392.5,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 2bit; 1k 392.5 48.9 52.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j6smdf4o",
   "source_url": "https://omlx.ai/benchmarks/performance/j6smdf4o",
   "tg_tokens": null,
   "tps": 48.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-j6smdf4o-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 55.2 GB at this context; duplicate traces of the same session: cb32t2ku; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 468.7,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 2bit; 64k 468.7 45.8 55.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j6smdf4o",
   "source_url": "https://omlx.ai/benchmarks/performance/j6smdf4o",
   "tg_tokens": null,
   "tps": 45.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-p8k9lhta-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: 8w7usjjo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 603.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 603.2 61.6 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p8k9lhta",
   "source_url": "https://omlx.ai/benchmarks/performance/p8k9lhta",
   "tg_tokens": null,
   "tps": 61.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-p8k9lhta-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: 8w7usjjo; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 750.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 750.4 52.3 71.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p8k9lhta",
   "source_url": "https://omlx.ai/benchmarks/performance/p8k9lhta",
   "tg_tokens": null,
   "tps": 52.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-j5lz09t9-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled, BF16 ngram; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: lo89izu0, q79juo6t, xl7hp05x, zc5mspnd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 662.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-BF16-ngram on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 662.5 69.2 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j5lz09t9",
   "source_url": "https://omlx.ai/benchmarks/performance/j5lz09t9",
   "tg_tokens": null,
   "tps": 69.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-j5lz09t9-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled, BF16 ngram; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: lo89izu0, q79juo6t, xl7hp05x, zc5mspnd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 790.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-BF16-ngram on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 790.3 58.0 71.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j5lz09t9",
   "source_url": "https://omlx.ai/benchmarks/performance/j5lz09t9",
   "tg_tokens": null,
   "tps": 58.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-j5lz09t9-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled, BF16 ngram; oMLX community benchmark trace; peak memory 72.5 GB at this context; duplicate traces of the same session: lo89izu0, q79juo6t, xl7hp05x, zc5mspnd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 777.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-BF16-ngram on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 777.0 64.9 72.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j5lz09t9",
   "source_url": "https://omlx.ai/benchmarks/performance/j5lz09t9",
   "tg_tokens": null,
   "tps": 64.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-j5lz09t9-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled, BF16 ngram; oMLX community benchmark trace; peak memory 73.7 GB at this context; duplicate traces of the same session: lo89izu0, q79juo6t, xl7hp05x, zc5mspnd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 771.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-BF16-ngram on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 771.8 64.3 73.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j5lz09t9",
   "source_url": "https://omlx.ai/benchmarks/performance/j5lz09t9",
   "tg_tokens": null,
   "tps": 64.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-j5lz09t9-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled, BF16 ngram; oMLX community benchmark trace; peak memory 76.3 GB at this context; duplicate traces of the same session: lo89izu0, q79juo6t, xl7hp05x, zc5mspnd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 764.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-BF16-ngram on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 764.8 54.6 76.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j5lz09t9",
   "source_url": "https://omlx.ai/benchmarks/performance/j5lz09t9",
   "tg_tokens": null,
   "tps": 54.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-oxnpe0wv-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace, single run at 128k context; peak memory 147.8 GB; TTFT 100155 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1309.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 128k tokens 1,309 PP tok/s 33.7 TG tok/s 100155 TTFT (ms) 147.8 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oxnpe0wv",
   "source_url": "https://omlx.ai/benchmarks/performance/oxnpe0wv",
   "tg_tokens": null,
   "tps": 33.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-xfgox2np-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.8 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1356.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 16k 1,356 70.5 143.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xfgox2np",
   "source_url": "https://omlx.ai/benchmarks/performance/xfgox2np",
   "tg_tokens": null,
   "tps": 70.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xfgox2np-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 147.8 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1311.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 128k 1,311 64.8 147.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xfgox2np",
   "source_url": "https://omlx.ai/benchmarks/performance/xfgox2np",
   "tg_tokens": null,
   "tps": 64.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-u1ixhewy-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace, single run at 128k context; peak memory 153.4 GB; TTFT 100726 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1301.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 128k tokens 1,301 PP tok/s 56.6 TG tok/s 100726 TTFT (ms) 153.4 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u1ixhewy",
   "source_url": "https://omlx.ai/benchmarks/performance/u1ixhewy",
   "tg_tokens": null,
   "tps": 56.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ju1sp1ka-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 101.0 GB at this context; duplicate traces of the same session: 4jtofwpy; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1324.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 1,324 86.7 101.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ju1sp1ka",
   "source_url": "https://omlx.ai/benchmarks/performance/ju1sp1ka",
   "tg_tokens": null,
   "tps": 86.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-ju1sp1ka-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 101.4 GB at this context; duplicate traces of the same session: 4jtofwpy; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2461.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 2,461 105.5 101.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ju1sp1ka",
   "source_url": "https://omlx.ai/benchmarks/performance/ju1sp1ka",
   "tg_tokens": null,
   "tps": 105.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-d6ssbdwc-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 101.0 GB at this context; duplicate traces of the same session: 01ho5bdc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1207.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k 1,207 93.6 101.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d6ssbdwc",
   "source_url": "https://omlx.ai/benchmarks/performance/d6ssbdwc",
   "tg_tokens": null,
   "tps": 93.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-d6ssbdwc-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 101.4 GB at this context; duplicate traces of the same session: 01ho5bdc; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2488.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 4k 2,488 87.5 101.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d6ssbdwc",
   "source_url": "https://omlx.ai/benchmarks/performance/d6ssbdwc",
   "tg_tokens": null,
   "tps": 87.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-hi14l5tb-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace, single run at 1k context; peak memory 101.0 GB; TTFT 910 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1125.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 6bit; 1k tokens 1,125 PP tok/s 96.7 TG tok/s 910 TTFT (ms) 101.0 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hi14l5tb",
   "source_url": "https://omlx.ai/benchmarks/performance/hi14l5tb",
   "tg_tokens": null,
   "tps": 96.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wk92usuf-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: q1u33gbd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 693.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "gemma-4-31B-it-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 693.0 39.8 28.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wk92usuf",
   "source_url": "https://omlx.ai/benchmarks/performance/wk92usuf",
   "tg_tokens": null,
   "tps": 39.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wk92usuf-ctx16384",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.1 GB at this context; duplicate traces of the same session: q1u33gbd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 657.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "gemma-4-31B-it-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 657.2 41.7 30.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wk92usuf",
   "source_url": "https://omlx.ai/benchmarks/performance/wk92usuf",
   "tg_tokens": null,
   "tps": 41.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ui7h3uzs-ctx65536",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 64k context; peak memory 14.9 GB; TTFT 20946 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3129.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k tokens 3,129 PP tok/s 55.5 TG tok/s 20946 TTFT (ms) 14.9 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ui7h3uzs",
   "source_url": "https://omlx.ai/benchmarks/performance/ui7h3uzs",
   "tg_tokens": null,
   "tps": 55.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-nfnyn8ax-ctx1024",
   "model": "Ling-3.0-tiny",
   "notes": "oMLX community benchmark trace; peak memory 9.0 GB at this context; duplicate traces of the same session: qhvhvhyp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1976.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ling-3.0-tiny-8bit-XL-mlx on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 8bit; 1k 1,976 107.0 9.0 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nfnyn8ax",
   "source_url": "https://omlx.ai/benchmarks/performance/nfnyn8ax",
   "tg_tokens": null,
   "tps": 107.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-nfnyn8ax-ctx16384",
   "model": "Ling-3.0-tiny",
   "notes": "oMLX community benchmark trace; peak memory 10.9 GB at this context; duplicate traces of the same session: qhvhvhyp; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1810.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ling-3.0-tiny-8bit-XL-mlx on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 8bit; 16k 1,810 75.7 10.9 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nfnyn8ax",
   "source_url": "https://omlx.ai/benchmarks/performance/nfnyn8ax",
   "tg_tokens": null,
   "tps": 75.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-46n7hd02-ctx1024",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4588.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 4,588 135.1 19.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/46n7hd02",
   "source_url": "https://omlx.ai/benchmarks/performance/46n7hd02",
   "tg_tokens": null,
   "tps": 135.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-46n7hd02-ctx4096",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4588.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 4,588 132.3 20.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/46n7hd02",
   "source_url": "https://omlx.ai/benchmarks/performance/46n7hd02",
   "tg_tokens": null,
   "tps": 132.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-kg1mw72o-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 21.9 GB; TTFT 1094 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 936.3,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k tokens 936.3 PP tok/s 69.6 TG tok/s 1094 TTFT (ms) 21.9 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kg1mw72o",
   "source_url": "https://omlx.ai/benchmarks/performance/kg1mw72o",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-uplxnt3q-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 259.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 259.9 16.5 31.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uplxnt3q",
   "source_url": "https://omlx.ai/benchmarks/performance/uplxnt3q",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-uplxnt3q-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 247.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 16k 247.8 16.0 33.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uplxnt3q",
   "source_url": "https://omlx.ai/benchmarks/performance/uplxnt3q",
   "tg_tokens": null,
   "tps": 16.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lscui6xz-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 177j1zx2, 1bx1ezx1, 2mi7lwcp, ee40q62e, s7grb1b4, y0y2u2tz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1132.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,132 66.3 17.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lscui6xz",
   "source_url": "https://omlx.ai/benchmarks/performance/lscui6xz",
   "tg_tokens": null,
   "tps": 66.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lscui6xz-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 177j1zx2, 1bx1ezx1, 2mi7lwcp, ee40q62e, s7grb1b4, y0y2u2tz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1296.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,296 61.9 19.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lscui6xz",
   "source_url": "https://omlx.ai/benchmarks/performance/lscui6xz",
   "tg_tokens": null,
   "tps": 61.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lscui6xz-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 177j1zx2, 1bx1ezx1, 2mi7lwcp, ee40q62e, s7grb1b4, y0y2u2tz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1165.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,165 75.9 19.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lscui6xz",
   "source_url": "https://omlx.ai/benchmarks/performance/lscui6xz",
   "tg_tokens": null,
   "tps": 75.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-lscui6xz-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 177j1zx2, 1bx1ezx1, 2mi7lwcp, ee40q62e, s7grb1b4, y0y2u2tz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1088.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,088 84.3 19.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lscui6xz",
   "source_url": "https://omlx.ai/benchmarks/performance/lscui6xz",
   "tg_tokens": null,
   "tps": 84.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lscui6xz-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: 177j1zx2, 1bx1ezx1, 2mi7lwcp, ee40q62e, s7grb1b4, y0y2u2tz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 936.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 936.1 114.3 21.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lscui6xz",
   "source_url": "https://omlx.ai/benchmarks/performance/lscui6xz",
   "tg_tokens": null,
   "tps": 114.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lscui6xz-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 177j1zx2, 1bx1ezx1, 2mi7lwcp, ee40q62e, s7grb1b4, y0y2u2tz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 751.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 751.0 44.8 23.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lscui6xz",
   "source_url": "https://omlx.ai/benchmarks/performance/lscui6xz",
   "tg_tokens": null,
   "tps": 44.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lscui6xz-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: 177j1zx2, 1bx1ezx1, 2mi7lwcp, ee40q62e, s7grb1b4, y0y2u2tz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 378.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 378.0 28.8 28.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lscui6xz",
   "source_url": "https://omlx.ai/benchmarks/performance/lscui6xz",
   "tg_tokens": null,
   "tps": 28.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9gv245ga-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 6lgvdiwp, gksric47, n9qfnpvw, owdkoy6p, pgsdkdeo, uxmr73jg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 855.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 855.5 44.2 17.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9gv245ga",
   "source_url": "https://omlx.ai/benchmarks/performance/9gv245ga",
   "tg_tokens": null,
   "tps": 44.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9gv245ga-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 6lgvdiwp, gksric47, n9qfnpvw, owdkoy6p, pgsdkdeo, uxmr73jg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1258.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,258 60.4 19.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9gv245ga",
   "source_url": "https://omlx.ai/benchmarks/performance/9gv245ga",
   "tg_tokens": null,
   "tps": 60.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9gv245ga-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 6lgvdiwp, gksric47, n9qfnpvw, owdkoy6p, pgsdkdeo, uxmr73jg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1096.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,096 54.6 19.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9gv245ga",
   "source_url": "https://omlx.ai/benchmarks/performance/9gv245ga",
   "tg_tokens": null,
   "tps": 54.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9gv245ga-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 6lgvdiwp, gksric47, n9qfnpvw, owdkoy6p, pgsdkdeo, uxmr73jg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1033.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,033 82.0 19.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9gv245ga",
   "source_url": "https://omlx.ai/benchmarks/performance/9gv245ga",
   "tg_tokens": null,
   "tps": 82.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-9gv245ga-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: 6lgvdiwp, gksric47, n9qfnpvw, owdkoy6p, pgsdkdeo, uxmr73jg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 902.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 902.3 128.0 21.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9gv245ga",
   "source_url": "https://omlx.ai/benchmarks/performance/9gv245ga",
   "tg_tokens": null,
   "tps": 128.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9gv245ga-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 6lgvdiwp, gksric47, n9qfnpvw, owdkoy6p, pgsdkdeo, uxmr73jg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 720.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 720.1 38.9 23.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9gv245ga",
   "source_url": "https://omlx.ai/benchmarks/performance/9gv245ga",
   "tg_tokens": null,
   "tps": 38.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-9gv245ga-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: 6lgvdiwp, gksric47, n9qfnpvw, owdkoy6p, pgsdkdeo, uxmr73jg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 361.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 361.3 29.0 28.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9gv245ga",
   "source_url": "https://omlx.ai/benchmarks/performance/9gv245ga",
   "tg_tokens": null,
   "tps": 29.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tvlnnvrq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 4nwdu3ov; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 902.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 902.6 56.2 17.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tvlnnvrq",
   "source_url": "https://omlx.ai/benchmarks/performance/tvlnnvrq",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tvlnnvrq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: 4nwdu3ov; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 990.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 990.7 65.8 18.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tvlnnvrq",
   "source_url": "https://omlx.ai/benchmarks/performance/tvlnnvrq",
   "tg_tokens": null,
   "tps": 65.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-k8za7xw3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 905.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 905.1 69.6 17.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k8za7xw3",
   "source_url": "https://omlx.ai/benchmarks/performance/k8za7xw3",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-k8za7xw3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 993.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 993.6 65.6 18.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k8za7xw3",
   "source_url": "https://omlx.ai/benchmarks/performance/k8za7xw3",
   "tg_tokens": null,
   "tps": 65.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-9mn8480f-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 51.1 GB at this context; duplicate traces of the same session: cdjotjr2; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 196.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 196.3 28.3 51.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9mn8480f",
   "source_url": "https://omlx.ai/benchmarks/performance/9mn8480f",
   "tg_tokens": null,
   "tps": 28.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-9mn8480f-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 54.7 GB at this context; duplicate traces of the same session: cdjotjr2; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 285.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 285.3 21.8 54.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9mn8480f",
   "source_url": "https://omlx.ai/benchmarks/performance/9mn8480f",
   "tg_tokens": null,
   "tps": 21.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ml8sly4i-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 1etfdtnd, 9br2tg5i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 385.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 385.4 21.0 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ml8sly4i",
   "source_url": "https://omlx.ai/benchmarks/performance/ml8sly4i",
   "tg_tokens": null,
   "tps": 21.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ml8sly4i-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: 1etfdtnd, 9br2tg5i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 574.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 574.3 45.3 71.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ml8sly4i",
   "source_url": "https://omlx.ai/benchmarks/performance/ml8sly4i",
   "tg_tokens": null,
   "tps": 45.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-ml8sly4i-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: 1etfdtnd, 9br2tg5i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 784.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 784.1 55.2 71.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ml8sly4i",
   "source_url": "https://omlx.ai/benchmarks/performance/ml8sly4i",
   "tg_tokens": null,
   "tps": 55.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-r0v2zxb7-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: 9w94e9ux, a9od1gwm, hf88irnp, wb3o8mct, wdluwkwt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 936.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 936.5 66.6 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r0v2zxb7",
   "source_url": "https://omlx.ai/benchmarks/performance/r0v2zxb7",
   "tg_tokens": null,
   "tps": 66.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-r0v2zxb7-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: 9w94e9ux, a9od1gwm, hf88irnp, wb3o8mct, wdluwkwt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1941.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,941 62.6 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r0v2zxb7",
   "source_url": "https://omlx.ai/benchmarks/performance/r0v2zxb7",
   "tg_tokens": null,
   "tps": 62.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-r0v2zxb7-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 9w94e9ux, a9od1gwm, hf88irnp, wb3o8mct, wdluwkwt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2312.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,312 74.1 71.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r0v2zxb7",
   "source_url": "https://omlx.ai/benchmarks/performance/r0v2zxb7",
   "tg_tokens": null,
   "tps": 74.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-r0v2zxb7-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: 9w94e9ux, a9od1gwm, hf88irnp, wb3o8mct, wdluwkwt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2696.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,696 109.1 72.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r0v2zxb7",
   "source_url": "https://omlx.ai/benchmarks/performance/r0v2zxb7",
   "tg_tokens": null,
   "tps": 109.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-r0v2zxb7-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.4 GB at this context; duplicate traces of the same session: 9w94e9ux, a9od1gwm, hf88irnp, wb3o8mct, wdluwkwt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2558.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,558 80.8 73.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r0v2zxb7",
   "source_url": "https://omlx.ai/benchmarks/performance/r0v2zxb7",
   "tg_tokens": null,
   "tps": 80.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-r0v2zxb7-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.6 GB at this context; duplicate traces of the same session: 9w94e9ux, a9od1gwm, hf88irnp, wb3o8mct, wdluwkwt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2287.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,287 81.1 74.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r0v2zxb7",
   "source_url": "https://omlx.ai/benchmarks/performance/r0v2zxb7",
   "tg_tokens": null,
   "tps": 81.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-o9ogfyhg-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.8 GB at this context; duplicate traces of the same session: p7tjqmk6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 427.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 427.6 14.5 36.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9ogfyhg",
   "source_url": "https://omlx.ai/benchmarks/performance/o9ogfyhg",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Max",
   "id": "omlx-o9ogfyhg-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.7 GB at this context; duplicate traces of the same session: p7tjqmk6; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 859.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 859.7 15.5 38.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o9ogfyhg",
   "source_url": "https://omlx.ai/benchmarks/performance/o9ogfyhg",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-0bsspdja-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 72.8 GB; TTFT 11400 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1437.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 4bit; 16k tokens 1,437 PP tok/s 58.3 TG tok/s 11400 TTFT (ms) 72.8 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0bsspdja",
   "source_url": "https://omlx.ai/benchmarks/performance/0bsspdja",
   "tg_tokens": null,
   "tps": 58.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-16erqs2w-ctx1024",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 11.7 GB at this context; duplicate traces of the same session: o8v09i1h; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.7,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 2bit; 1k 130.7 21.1 11.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/16erqs2w",
   "source_url": "https://omlx.ai/benchmarks/performance/16erqs2w",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-16erqs2w-ctx4096",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: o8v09i1h; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.4,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 2bit; 4k 130.4 20.4 19.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/16erqs2w",
   "source_url": "https://omlx.ai/benchmarks/performance/16erqs2w",
   "tg_tokens": null,
   "tps": 20.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-0qf80ap8-ctx1024",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 11.7 GB at this context; duplicate traces of the same session: olxfngyw; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 126.8,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 2bit; 1k 126.8 21.1 11.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0qf80ap8",
   "source_url": "https://omlx.ai/benchmarks/performance/0qf80ap8",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-0qf80ap8-ctx4096",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: olxfngyw; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.4,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 2bit; 4k 130.4 20.4 19.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0qf80ap8",
   "source_url": "https://omlx.ai/benchmarks/performance/0qf80ap8",
   "tg_tokens": null,
   "tps": 20.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-j9jog0d6-ctx1024",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: 1j9jdygg, 4c359o9x, wc1v37uq; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 155.1,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 2bit; 1k 155.1 12.2 11.5 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j9jog0d6",
   "source_url": "https://omlx.ai/benchmarks/performance/j9jog0d6",
   "tg_tokens": null,
   "tps": 12.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-j9jog0d6-ctx4096",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: 1j9jdygg, 4c359o9x, wc1v37uq; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 158.1,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 2bit; 4k 158.1 11.8 13.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j9jog0d6",
   "source_url": "https://omlx.ai/benchmarks/performance/j9jog0d6",
   "tg_tokens": null,
   "tps": 11.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-j9jog0d6-ctx8192",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 14.2 GB at this context; duplicate traces of the same session: 1j9jdygg, 4c359o9x, wc1v37uq; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 153.1,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 2bit; 8k 153.1 10.8 14.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j9jog0d6",
   "source_url": "https://omlx.ai/benchmarks/performance/j9jog0d6",
   "tg_tokens": null,
   "tps": 10.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-j9jog0d6-ctx16384",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 15.2 GB at this context; duplicate traces of the same session: 1j9jdygg, 4c359o9x, wc1v37uq; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 141.3,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 2bit; 16k 141.3 8.7 15.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j9jog0d6",
   "source_url": "https://omlx.ai/benchmarks/performance/j9jog0d6",
   "tg_tokens": null,
   "tps": 8.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-j9jog0d6-ctx32768",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: 1j9jdygg, 4c359o9x, wc1v37uq; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 126.8,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 2bit; 32k 126.8 6.3 17.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j9jog0d6",
   "source_url": "https://omlx.ai/benchmarks/performance/j9jog0d6",
   "tg_tokens": null,
   "tps": 6.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-j9jog0d6-ctx65536",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 1j9jdygg, 4c359o9x, wc1v37uq; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.3,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 2bit; 64k 103.3 3.7 21.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j9jog0d6",
   "source_url": "https://omlx.ai/benchmarks/performance/j9jog0d6",
   "tg_tokens": null,
   "tps": 3.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-43nuocl6-ctx1024",
   "model": "Muse Glimmer",
   "notes": "oMLX community benchmark trace; peak memory 15.2 GB at this context; duplicate traces of the same session: 4pinpywu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 341.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-30B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 341.1 13.0 15.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/43nuocl6",
   "source_url": "https://omlx.ai/benchmarks/performance/43nuocl6",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-43nuocl6-ctx4096",
   "model": "Muse Glimmer",
   "notes": "oMLX community benchmark trace; peak memory 15.5 GB at this context; duplicate traces of the same session: 4pinpywu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 326.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Muse-Glimmer-30B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 326.4 12.6 15.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/43nuocl6",
   "source_url": "https://omlx.ai/benchmarks/performance/43nuocl6",
   "tg_tokens": null,
   "tps": 12.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-i5agfiqc-ctx1024",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: wng7mzuk; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 978.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 978.0 68.8 16.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i5agfiqc",
   "source_url": "https://omlx.ai/benchmarks/performance/i5agfiqc",
   "tg_tokens": null,
   "tps": 68.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-i5agfiqc-ctx4096",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: wng7mzuk; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1634.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,634 61.6 17.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i5agfiqc",
   "source_url": "https://omlx.ai/benchmarks/performance/i5agfiqc",
   "tg_tokens": null,
   "tps": 61.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-77sx525t-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: jrkpcnjc, tyhqploz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1736.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,736 100.9 21.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/77sx525t",
   "source_url": "https://omlx.ai/benchmarks/performance/77sx525t",
   "tg_tokens": null,
   "tps": 100.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-77sx525t-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: jrkpcnjc, tyhqploz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1966.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 1,966 91.2 23.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/77sx525t",
   "source_url": "https://omlx.ai/benchmarks/performance/77sx525t",
   "tg_tokens": null,
   "tps": 91.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-77sx525t-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: jrkpcnjc, tyhqploz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1462.0,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 1,462 76.9 24.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/77sx525t",
   "source_url": "https://omlx.ai/benchmarks/performance/77sx525t",
   "tg_tokens": null,
   "tps": 76.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-yk45f5rw-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: w2ohly74; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 697.1,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 1k 697.1 80.5 29.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yk45f5rw",
   "source_url": "https://omlx.ai/benchmarks/performance/yk45f5rw",
   "tg_tokens": null,
   "tps": 80.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-yk45f5rw-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.9 GB at this context; duplicate traces of the same session: w2ohly74; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 936.3,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 4k 936.3 63.7 31.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yk45f5rw",
   "source_url": "https://omlx.ai/benchmarks/performance/yk45f5rw",
   "tg_tokens": null,
   "tps": 63.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-yk45f5rw-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.3 GB at this context; duplicate traces of the same session: w2ohly74; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 745.7,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 8k 745.7 62.5 32.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yk45f5rw",
   "source_url": "https://omlx.ai/benchmarks/performance/yk45f5rw",
   "tg_tokens": null,
   "tps": 62.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-kyc3kvkz-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 611.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 1k 611.0 41.8 29.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kyc3kvkz",
   "source_url": "https://omlx.ai/benchmarks/performance/kyc3kvkz",
   "tg_tokens": null,
   "tps": 41.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-kyc3kvkz-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.9 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 979.5,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 4k 979.5 55.8 31.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kyc3kvkz",
   "source_url": "https://omlx.ai/benchmarks/performance/kyc3kvkz",
   "tg_tokens": null,
   "tps": 55.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-kyc3kvkz-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.3 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 782.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 8k 782.0 67.7 32.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kyc3kvkz",
   "source_url": "https://omlx.ai/benchmarks/performance/kyc3kvkz",
   "tg_tokens": null,
   "tps": 67.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-7xxa76mc-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 562gpsx8, d411w1li; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 246.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 246.6 38.4 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7xxa76mc",
   "source_url": "https://omlx.ai/benchmarks/performance/7xxa76mc",
   "tg_tokens": null,
   "tps": 38.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-7xxa76mc-ctx8192",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: 562gpsx8, d411w1li; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 253.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 253.3 27.6 27.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7xxa76mc",
   "source_url": "https://omlx.ai/benchmarks/performance/7xxa76mc",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-7xxa76mc-ctx32768",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 30.2 GB at this context; duplicate traces of the same session: 562gpsx8, d411w1li; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 226.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 226.5 35.8 30.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7xxa76mc",
   "source_url": "https://omlx.ai/benchmarks/performance/7xxa76mc",
   "tg_tokens": null,
   "tps": 35.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-m3z6wumd-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: g0yr7y06; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 184.2,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 184.2 20.1 21.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m3z6wumd",
   "source_url": "https://omlx.ai/benchmarks/performance/m3z6wumd",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-m3z6wumd-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: g0yr7y06; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 154.3,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 154.3 17.5 25.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m3z6wumd",
   "source_url": "https://omlx.ai/benchmarks/performance/m3z6wumd",
   "tg_tokens": null,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-oba76d6o-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: gmwxuhi0, l73tbfv9, r6iw1epj, w9fl4f71; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1115.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,115 66.7 17.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oba76d6o",
   "source_url": "https://omlx.ai/benchmarks/performance/oba76d6o",
   "tg_tokens": null,
   "tps": 66.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-oba76d6o-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: gmwxuhi0, l73tbfv9, r6iw1epj, w9fl4f71; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1287.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,287 59.7 19.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oba76d6o",
   "source_url": "https://omlx.ai/benchmarks/performance/oba76d6o",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-oba76d6o-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: gmwxuhi0, l73tbfv9, r6iw1epj, w9fl4f71; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1156.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,156 75.3 19.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oba76d6o",
   "source_url": "https://omlx.ai/benchmarks/performance/oba76d6o",
   "tg_tokens": null,
   "tps": 75.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-oba76d6o-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: gmwxuhi0, l73tbfv9, r6iw1epj, w9fl4f71; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1063.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,063 84.3 19.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oba76d6o",
   "source_url": "https://omlx.ai/benchmarks/performance/oba76d6o",
   "tg_tokens": null,
   "tps": 84.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-oba76d6o-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: gmwxuhi0, l73tbfv9, r6iw1epj, w9fl4f71; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 925.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 925.7 123.7 21.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oba76d6o",
   "source_url": "https://omlx.ai/benchmarks/performance/oba76d6o",
   "tg_tokens": null,
   "tps": 123.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-oba76d6o-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: gmwxuhi0, l73tbfv9, r6iw1epj, w9fl4f71; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 740.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 740.3 38.0 23.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oba76d6o",
   "source_url": "https://omlx.ai/benchmarks/performance/oba76d6o",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-oba76d6o-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: gmwxuhi0, l73tbfv9, r6iw1epj, w9fl4f71; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 376.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 376.3 27.9 28.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oba76d6o",
   "source_url": "https://omlx.ai/benchmarks/performance/oba76d6o",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-angxtivn-ctx16384",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 38.3 GB at this context; duplicate traces of the same session: jwaat1c1, oogb7zmo; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1801.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 8bit; 16k 1,801 60.6 38.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/angxtivn",
   "source_url": "https://omlx.ai/benchmarks/performance/angxtivn",
   "tg_tokens": null,
   "tps": 60.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-angxtivn-ctx32768",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: jwaat1c1, oogb7zmo; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1476.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 8bit; 32k 1,476 58.1 39.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/angxtivn",
   "source_url": "https://omlx.ai/benchmarks/performance/angxtivn",
   "tg_tokens": null,
   "tps": 58.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-angxtivn-ctx131072",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 41.6 GB at this context; duplicate traces of the same session: jwaat1c1, oogb7zmo; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 708.5,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 8bit; 128k 708.5 41.9 41.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/angxtivn",
   "source_url": "https://omlx.ai/benchmarks/performance/angxtivn",
   "tg_tokens": null,
   "tps": 41.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-qx9srl3d-ctx8192",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 37.7 GB at this context; duplicate traces of the same session: 063hx5f2, 4kb2ka8u, ob23qqf7, v3cq80ei; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1998.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 8bit; 8k 1,998 64.6 37.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qx9srl3d",
   "source_url": "https://omlx.ai/benchmarks/performance/qx9srl3d",
   "tg_tokens": null,
   "tps": 64.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-qx9srl3d-ctx16384",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 38.2 GB at this context; duplicate traces of the same session: 063hx5f2, 4kb2ka8u, ob23qqf7, v3cq80ei; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1800.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 8bit; 16k 1,800 61.7 38.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qx9srl3d",
   "source_url": "https://omlx.ai/benchmarks/performance/qx9srl3d",
   "tg_tokens": null,
   "tps": 61.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-qx9srl3d-ctx32768",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: 063hx5f2, 4kb2ka8u, ob23qqf7, v3cq80ei; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1470.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 8bit; 32k 1,470 56.5 39.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qx9srl3d",
   "source_url": "https://omlx.ai/benchmarks/performance/qx9srl3d",
   "tg_tokens": null,
   "tps": 56.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-qx9srl3d-ctx65536",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 40.3 GB at this context; duplicate traces of the same session: 063hx5f2, 4kb2ka8u, ob23qqf7, v3cq80ei; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1075.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 8bit; 64k 1,075 26.5 40.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qx9srl3d",
   "source_url": "https://omlx.ai/benchmarks/performance/qx9srl3d",
   "tg_tokens": null,
   "tps": 26.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-qx9srl3d-ctx131072",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 41.6 GB at this context; duplicate traces of the same session: 063hx5f2, 4kb2ka8u, ob23qqf7, v3cq80ei; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 702.6,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 8bit; 128k 702.6 43.6 41.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qx9srl3d",
   "source_url": "https://omlx.ai/benchmarks/performance/qx9srl3d",
   "tg_tokens": null,
   "tps": 43.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-vvdr7m2w-ctx1024",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: i2n8546w; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1574.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 8bit; 1k 1,574 65.5 35.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vvdr7m2w",
   "source_url": "https://omlx.ai/benchmarks/performance/vvdr7m2w",
   "tg_tokens": null,
   "tps": 65.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-vvdr7m2w-ctx4096",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: i2n8546w; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2127.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 8bit; 4k 2,127 64.8 37.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vvdr7m2w",
   "source_url": "https://omlx.ai/benchmarks/performance/vvdr7m2w",
   "tg_tokens": null,
   "tps": 64.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zftqtedm-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.3 GB at this context; duplicate traces of the same session: enpsisr2, j02im4t8, sqrjsc0z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2296.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,296 66.3 70.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zftqtedm",
   "source_url": "https://omlx.ai/benchmarks/performance/zftqtedm",
   "tg_tokens": null,
   "tps": 66.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zftqtedm-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: enpsisr2, j02im4t8, sqrjsc0z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2313.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,313 66.2 71.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zftqtedm",
   "source_url": "https://omlx.ai/benchmarks/performance/zftqtedm",
   "tg_tokens": null,
   "tps": 66.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zftqtedm-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.6 GB at this context; duplicate traces of the same session: enpsisr2, j02im4t8, sqrjsc0z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2376.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,376 67.5 71.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zftqtedm",
   "source_url": "https://omlx.ai/benchmarks/performance/zftqtedm",
   "tg_tokens": null,
   "tps": 67.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zftqtedm-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.7 GB at this context; duplicate traces of the same session: enpsisr2, j02im4t8, sqrjsc0z; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2414.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,414 63.5 72.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zftqtedm",
   "source_url": "https://omlx.ai/benchmarks/performance/zftqtedm",
   "tg_tokens": null,
   "tps": 63.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zcjzilef-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.1 GB at this context; duplicate traces of the same session: 00c430rr, 9v5i3ank, irhexf6y, nk3vhpzp, vr3qwnfy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 426.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 426.2 62.9 69.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zcjzilef",
   "source_url": "https://omlx.ai/benchmarks/performance/zcjzilef",
   "tg_tokens": null,
   "tps": 62.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zcjzilef-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.5 GB at this context; duplicate traces of the same session: 00c430rr, 9v5i3ank, irhexf6y, nk3vhpzp, vr3qwnfy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1665.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,665 29.3 69.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zcjzilef",
   "source_url": "https://omlx.ai/benchmarks/performance/zcjzilef",
   "tg_tokens": null,
   "tps": 29.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zcjzilef-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.5 GB at this context; duplicate traces of the same session: 00c430rr, 9v5i3ank, irhexf6y, nk3vhpzp, vr3qwnfy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2361.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,361 66.3 70.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zcjzilef",
   "source_url": "https://omlx.ai/benchmarks/performance/zcjzilef",
   "tg_tokens": null,
   "tps": 66.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zcjzilef-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: 00c430rr, 9v5i3ank, irhexf6y, nk3vhpzp, vr3qwnfy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2436.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,436 71.8 71.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zcjzilef",
   "source_url": "https://omlx.ai/benchmarks/performance/zcjzilef",
   "tg_tokens": null,
   "tps": 71.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zcjzilef-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.7 GB at this context; duplicate traces of the same session: 00c430rr, 9v5i3ank, irhexf6y, nk3vhpzp, vr3qwnfy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2363.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,363 59.0 72.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zcjzilef",
   "source_url": "https://omlx.ai/benchmarks/performance/zcjzilef",
   "tg_tokens": null,
   "tps": 59.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zcjzilef-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 74.6 GB at this context; duplicate traces of the same session: 00c430rr, 9v5i3ank, irhexf6y, nk3vhpzp, vr3qwnfy; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2274.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,274 60.4 74.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zcjzilef",
   "source_url": "https://omlx.ai/benchmarks/performance/zcjzilef",
   "tg_tokens": null,
   "tps": 60.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-whc1xhim-ctx1024",
   "model": "Ternary-Bonsai-2-27B-Abliterated-v2",
   "notes": "PQ2_0; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 11.8 GB; TTFT 18639 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 54.9,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Ternary-Bonsai-2-27B-Abliterated-v2-PQ2_0-MTP-MLX on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 unknown; 1k tokens 54.9 PP tok/s 7.6 TG tok/s 18639 TTFT (ms) 11.8 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/whc1xhim",
   "source_url": "https://omlx.ai/benchmarks/performance/whc1xhim",
   "tg_tokens": null,
   "tps": 7.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-s0wl448e-ctx1024",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 11.5 GB at this context; duplicate traces of the same session: evzvl2wx; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 155.1,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 2bit; 1k 155.1 12.2 11.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s0wl448e",
   "source_url": "https://omlx.ai/benchmarks/performance/s0wl448e",
   "tg_tokens": null,
   "tps": 12.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-s0wl448e-ctx4096",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: evzvl2wx; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 158.1,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 2bit; 4k 158.1 11.8 13.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s0wl448e",
   "source_url": "https://omlx.ai/benchmarks/performance/s0wl448e",
   "tg_tokens": null,
   "tps": 11.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-s0wl448e-ctx8192",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 14.2 GB at this context; duplicate traces of the same session: evzvl2wx; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 153.1,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 2bit; 8k 153.1 10.8 14.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s0wl448e",
   "source_url": "https://omlx.ai/benchmarks/performance/s0wl448e",
   "tg_tokens": null,
   "tps": 10.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-s0wl448e-ctx16384",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 15.2 GB at this context; duplicate traces of the same session: evzvl2wx; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 141.3,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 2bit; 16k 141.3 8.7 15.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s0wl448e",
   "source_url": "https://omlx.ai/benchmarks/performance/s0wl448e",
   "tg_tokens": null,
   "tps": 8.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-s0wl448e-ctx32768",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: evzvl2wx; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 126.8,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 2bit; 32k 126.8 6.3 17.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s0wl448e",
   "source_url": "https://omlx.ai/benchmarks/performance/s0wl448e",
   "tg_tokens": null,
   "tps": 6.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-s0wl448e-ctx65536",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: evzvl2wx; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.3,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 2bit; 64k 103.3 3.7 21.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s0wl448e",
   "source_url": "https://omlx.ai/benchmarks/performance/s0wl448e",
   "tg_tokens": null,
   "tps": 3.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-xf0f263b-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 5.7 GB at this context; duplicate traces of the same session: x2f1zsnh; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2385.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 2,385 199.7 5.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xf0f263b",
   "source_url": "https://omlx.ai/benchmarks/performance/xf0f263b",
   "tg_tokens": null,
   "tps": 199.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-xf0f263b-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: x2f1zsnh; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2683.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 2,683 72.9 5.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xf0f263b",
   "source_url": "https://omlx.ai/benchmarks/performance/xf0f263b",
   "tg_tokens": null,
   "tps": 72.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ezkmumpa-ctx1024",
   "model": "Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored",
   "notes": "MTP speculative decoding enabled; DavidAU build; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: eknwid8o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 799.5,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "DavidAU_Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored-oQ8-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 799.5 45.5 33.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ezkmumpa",
   "source_url": "https://omlx.ai/benchmarks/performance/ezkmumpa",
   "tg_tokens": null,
   "tps": 45.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ezkmumpa-ctx4096",
   "model": "Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored",
   "notes": "MTP speculative decoding enabled; DavidAU build; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: eknwid8o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 888.5,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "DavidAU_Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored-oQ8-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 888.5 44.6 34.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ezkmumpa",
   "source_url": "https://omlx.ai/benchmarks/performance/ezkmumpa",
   "tg_tokens": null,
   "tps": 44.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-xiaidzlt-ctx131072",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: 0iiuztu5; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1102.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 128k 1,102 49.0 16.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xiaidzlt",
   "source_url": "https://omlx.ai/benchmarks/performance/xiaidzlt",
   "tg_tokens": null,
   "tps": 49.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-xiaidzlt-ctx199680",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: 0iiuztu5; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 888.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 4bit; 195k 888.1 37.8 17.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xiaidzlt",
   "source_url": "https://omlx.ai/benchmarks/performance/xiaidzlt",
   "tg_tokens": null,
   "tps": 37.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-up1ygpja-ctx1024",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "MLX; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: mleazosn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2390.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 2,390 55.2 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/up1ygpja",
   "source_url": "https://omlx.ai/benchmarks/performance/up1ygpja",
   "tg_tokens": null,
   "tps": 55.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-up1ygpja-ctx4096",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "MLX; oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: mleazosn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1434.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,434 52.7 17.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/up1ygpja",
   "source_url": "https://omlx.ai/benchmarks/performance/up1ygpja",
   "tg_tokens": null,
   "tps": 52.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xo6l7uys-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MLX Serve build; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: ut12ztym; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 761.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-Serve-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 761.9 149.7 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xo6l7uys",
   "source_url": "https://omlx.ai/benchmarks/performance/xo6l7uys",
   "tg_tokens": null,
   "tps": 149.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-xo6l7uys-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MLX Serve build; oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: ut12ztym; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2274.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-Serve-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,274 149.1 20.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xo6l7uys",
   "source_url": "https://omlx.ai/benchmarks/performance/xo6l7uys",
   "tg_tokens": null,
   "tps": 149.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-jlebgtf8-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 697.1,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 1k 697.1 80.5 29.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jlebgtf8",
   "source_url": "https://omlx.ai/benchmarks/performance/jlebgtf8",
   "tg_tokens": null,
   "tps": 80.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-jlebgtf8-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.9 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 936.3,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 4k 936.3 63.7 31.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jlebgtf8",
   "source_url": "https://omlx.ai/benchmarks/performance/jlebgtf8",
   "tg_tokens": null,
   "tps": 63.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-jlebgtf8-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.3 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 745.7,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 8k 745.7 62.5 32.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jlebgtf8",
   "source_url": "https://omlx.ai/benchmarks/performance/jlebgtf8",
   "tg_tokens": null,
   "tps": 62.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-j1a30t5q-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.9 GB at this context; duplicate traces of the same session: z5vvibvi; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 611.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 1k 611.0 41.8 29.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j1a30t5q",
   "source_url": "https://omlx.ai/benchmarks/performance/j1a30t5q",
   "tg_tokens": null,
   "tps": 41.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-j1a30t5q-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 31.9 GB at this context; duplicate traces of the same session: z5vvibvi; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 979.5,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 4k 979.5 55.8 31.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j1a30t5q",
   "source_url": "https://omlx.ai/benchmarks/performance/j1a30t5q",
   "tg_tokens": null,
   "tps": 55.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-j1a30t5q-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "fp16; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.3 GB at this context; duplicate traces of the same session: z5vvibvi; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 782.0,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.6-35B-A3B-oQ6-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 8k 782.0 67.7 32.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j1a30t5q",
   "source_url": "https://omlx.ai/benchmarks/performance/j1a30t5q",
   "tg_tokens": null,
   "tps": 67.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-n7k5v793-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: ynhmias6; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 393.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 393.0 18.1 16.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n7k5v793",
   "source_url": "https://omlx.ai/benchmarks/performance/n7k5v793",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-n7k5v793-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: ynhmias6; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 407.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 407.8 17.9 17.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n7k5v793",
   "source_url": "https://omlx.ai/benchmarks/performance/n7k5v793",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ida0tul8-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 3zeeipy6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 409.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 409.3 18.0 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ida0tul8",
   "source_url": "https://omlx.ai/benchmarks/performance/ida0tul8",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ida0tul8-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 3zeeipy6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 434.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 434.1 17.9 17.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ida0tul8",
   "source_url": "https://omlx.ai/benchmarks/performance/ida0tul8",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qej7n8ln-ctx4096",
   "model": "Qwen3.8-27B-JANG_4D",
   "notes": "oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: o3tp1big; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 275.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-JANG_4D on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 275.5 17.6 23.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qej7n8ln",
   "source_url": "https://omlx.ai/benchmarks/performance/qej7n8ln",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qej7n8ln-ctx16384",
   "model": "Qwen3.8-27B-JANG_4D",
   "notes": "oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: o3tp1big; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 266.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-JANG_4D on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 266.0 20.6 24.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qej7n8ln",
   "source_url": "https://omlx.ai/benchmarks/performance/qej7n8ln",
   "tg_tokens": null,
   "tps": 20.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-3nc3nptb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 8ha232qq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 256.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 256.2 16.7 28.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3nc3nptb",
   "source_url": "https://omlx.ai/benchmarks/performance/3nc3nptb",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-3nc3nptb-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 31.7 GB at this context; duplicate traces of the same session: 8ha232qq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 255.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 255.0 16.5 31.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3nc3nptb",
   "source_url": "https://omlx.ai/benchmarks/performance/3nc3nptb",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-llehz8s6-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: 0syj0epb, iqyx5tjo; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 194.1,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 194.1 19.5 21.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/llehz8s6",
   "source_url": "https://omlx.ai/benchmarks/performance/llehz8s6",
   "tg_tokens": null,
   "tps": 19.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-llehz8s6-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.9 GB at this context; duplicate traces of the same session: 0syj0epb, iqyx5tjo; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 161.6,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 161.6 17.4 25.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/llehz8s6",
   "source_url": "https://omlx.ai/benchmarks/performance/llehz8s6",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-llehz8s6-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.0 GB at this context; duplicate traces of the same session: 0syj0epb, iqyx5tjo; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 152.4,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 152.4 20.8 27.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/llehz8s6",
   "source_url": "https://omlx.ai/benchmarks/performance/llehz8s6",
   "tg_tokens": null,
   "tps": 20.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0q9u8mwp-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 26rjiz6l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1115.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,115 66.7 17.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0q9u8mwp",
   "source_url": "https://omlx.ai/benchmarks/performance/0q9u8mwp",
   "tg_tokens": null,
   "tps": 66.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0q9u8mwp-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 26rjiz6l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1287.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,287 59.7 19.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0q9u8mwp",
   "source_url": "https://omlx.ai/benchmarks/performance/0q9u8mwp",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0q9u8mwp-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: 26rjiz6l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1156.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,156 75.3 19.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0q9u8mwp",
   "source_url": "https://omlx.ai/benchmarks/performance/0q9u8mwp",
   "tg_tokens": null,
   "tps": 75.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0q9u8mwp-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 26rjiz6l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1063.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 1,063 84.3 19.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0q9u8mwp",
   "source_url": "https://omlx.ai/benchmarks/performance/0q9u8mwp",
   "tg_tokens": null,
   "tps": 84.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0q9u8mwp-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: 26rjiz6l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 925.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 925.7 123.7 21.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0q9u8mwp",
   "source_url": "https://omlx.ai/benchmarks/performance/0q9u8mwp",
   "tg_tokens": null,
   "tps": 123.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0q9u8mwp-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 26rjiz6l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 740.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 740.3 38.0 23.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0q9u8mwp",
   "source_url": "https://omlx.ai/benchmarks/performance/0q9u8mwp",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-0q9u8mwp-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: 26rjiz6l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 376.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 376.3 27.9 28.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0q9u8mwp",
   "source_url": "https://omlx.ai/benchmarks/performance/0q9u8mwp",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yt1mcu4g-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: 7ki0e8s7, 95gpyul1, ckqjd1or, ujjomjc6, v6dgghwi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 901.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 901.9 65.4 17.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yt1mcu4g",
   "source_url": "https://omlx.ai/benchmarks/performance/yt1mcu4g",
   "tg_tokens": null,
   "tps": 65.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yt1mcu4g-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 7ki0e8s7, 95gpyul1, ckqjd1or, ujjomjc6, v6dgghwi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1005.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,005 68.9 18.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yt1mcu4g",
   "source_url": "https://omlx.ai/benchmarks/performance/yt1mcu4g",
   "tg_tokens": null,
   "tps": 68.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yt1mcu4g-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 7ki0e8s7, 95gpyul1, ckqjd1or, ujjomjc6, v6dgghwi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 942.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 942.8 75.3 19.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yt1mcu4g",
   "source_url": "https://omlx.ai/benchmarks/performance/yt1mcu4g",
   "tg_tokens": null,
   "tps": 75.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yt1mcu4g-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 7ki0e8s7, 95gpyul1, ckqjd1or, ujjomjc6, v6dgghwi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 845.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 845.6 77.5 19.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yt1mcu4g",
   "source_url": "https://omlx.ai/benchmarks/performance/yt1mcu4g",
   "tg_tokens": null,
   "tps": 77.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yt1mcu4g-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 7ki0e8s7, 95gpyul1, ckqjd1or, ujjomjc6, v6dgghwi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 760.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 760.2 76.9 20.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yt1mcu4g",
   "source_url": "https://omlx.ai/benchmarks/performance/yt1mcu4g",
   "tg_tokens": null,
   "tps": 76.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-yt1mcu4g-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: 7ki0e8s7, 95gpyul1, ckqjd1or, ujjomjc6, v6dgghwi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 712.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 712.0 59.6 23.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yt1mcu4g",
   "source_url": "https://omlx.ai/benchmarks/performance/yt1mcu4g",
   "tg_tokens": null,
   "tps": 59.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-1e4paorb-ctx1024",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: 94or7tvw; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 340.7,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 2bit; 1k 340.7 23.7 11.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1e4paorb",
   "source_url": "https://omlx.ai/benchmarks/performance/1e4paorb",
   "tg_tokens": null,
   "tps": 23.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-1e4paorb-ctx4096",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 14.2 GB at this context; duplicate traces of the same session: 94or7tvw; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 344.9,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 2bit; 4k 344.9 22.7 14.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1e4paorb",
   "source_url": "https://omlx.ai/benchmarks/performance/1e4paorb",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-m278vt4d-ctx1024",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 286.0,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 2bit; 1k 286.0 23.6 11.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m278vt4d",
   "source_url": "https://omlx.ai/benchmarks/performance/m278vt4d",
   "tg_tokens": null,
   "tps": 23.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-m278vt4d-ctx4096",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 14.2 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 330.8,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 2bit; 4k 330.8 22.6 14.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m278vt4d",
   "source_url": "https://omlx.ai/benchmarks/performance/m278vt4d",
   "tg_tokens": null,
   "tps": 22.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-7h3608f2-ctx1024",
   "model": "Ternary-Bonsai-2-2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 3.4 GB at this context; duplicate traces of the same session: 570fsk1s; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3016.0,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-2B-mlx-2bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 2bit; 1k 3,016 239.9 3.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7h3608f2",
   "source_url": "https://omlx.ai/benchmarks/performance/7h3608f2",
   "tg_tokens": null,
   "tps": 239.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-7h3608f2-ctx4096",
   "model": "Ternary-Bonsai-2-2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 3.7 GB at this context; duplicate traces of the same session: 570fsk1s; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4534.0,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-2B-mlx-2bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 2bit; 4k 4,534 222.2 3.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7h3608f2",
   "source_url": "https://omlx.ai/benchmarks/performance/7h3608f2",
   "tg_tokens": null,
   "tps": 222.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-7dylpw6w-ctx1024",
   "model": "Ternary-Bonsai-2-4B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 4.4 GB at this context; duplicate traces of the same session: iplo5c8m; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1978.0,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-4B-mlx-2bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 2bit; 1k 1,978 123.6 4.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7dylpw6w",
   "source_url": "https://omlx.ai/benchmarks/performance/7dylpw6w",
   "tg_tokens": null,
   "tps": 123.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-7dylpw6w-ctx4096",
   "model": "Ternary-Bonsai-2-4B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 5.2 GB at this context; duplicate traces of the same session: iplo5c8m; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1974.0,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-4B-mlx-2bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 2bit; 4k 1,974 112.7 5.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7dylpw6w",
   "source_url": "https://omlx.ai/benchmarks/performance/7dylpw6w",
   "tg_tokens": null,
   "tps": 112.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-4ul29knj-ctx1024",
   "model": "Ternary-Bonsai-2-9B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 5.7 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 967.8,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-9B-mlx-2bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 2bit; 1k 967.8 74.9 5.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ul29knj",
   "source_url": "https://omlx.ai/benchmarks/performance/4ul29knj",
   "tg_tokens": null,
   "tps": 74.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-4ul29knj-ctx4096",
   "model": "Ternary-Bonsai-2-9B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1090.0,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-9B-mlx-2bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 2bit; 4k 1,090 70.8 6.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ul29knj",
   "source_url": "https://omlx.ai/benchmarks/performance/4ul29knj",
   "tg_tokens": null,
   "tps": 70.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-vjosq72e-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: 3h57amvf, kytiflt2; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 591.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 591.5 52.6 14.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vjosq72e",
   "source_url": "https://omlx.ai/benchmarks/performance/vjosq72e",
   "tg_tokens": null,
   "tps": 52.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-vjosq72e-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: 3h57amvf, kytiflt2; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 585.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 585.0 49.9 14.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vjosq72e",
   "source_url": "https://omlx.ai/benchmarks/performance/vjosq72e",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-vjosq72e-ctx8192",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 15.0 GB at this context; duplicate traces of the same session: 3h57amvf, kytiflt2; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 535.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 535.4 47.0 15.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vjosq72e",
   "source_url": "https://omlx.ai/benchmarks/performance/vjosq72e",
   "tg_tokens": null,
   "tps": 47.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-8a4zp5a6-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 5.7 GB at this context; duplicate traces of the same session: 2w6yn8jw; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2386.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 1k 2,386 167.3 5.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8a4zp5a6",
   "source_url": "https://omlx.ai/benchmarks/performance/8a4zp5a6",
   "tg_tokens": null,
   "tps": 167.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-8a4zp5a6-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: 2w6yn8jw; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2697.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 2,697 136.9 5.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8a4zp5a6",
   "source_url": "https://omlx.ai/benchmarks/performance/8a4zp5a6",
   "tg_tokens": null,
   "tps": 136.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-5grlwabw-ctx1024",
   "model": "gemma-4-E4B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: vo2l89c7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1275.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-qat-OptiQ-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 1,275 56.7 6.8 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5grlwabw",
   "source_url": "https://omlx.ai/benchmarks/performance/5grlwabw",
   "tg_tokens": null,
   "tps": 56.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-5grlwabw-ctx4096",
   "model": "gemma-4-E4B-it-qat",
   "notes": "OptiQ MTP build; oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: vo2l89c7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1302.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-e4b-it-qat-OptiQ-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 1,302 55.0 6.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5grlwabw",
   "source_url": "https://omlx.ai/benchmarks/performance/5grlwabw",
   "tg_tokens": null,
   "tps": 55.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-rbkmp686-ctx4096",
   "model": "ALoDLM-8B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 17.0 GB; TTFT 2332 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1756.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "ALoDLM-8B on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 unknown; 4k tokens 1,756 PP tok/s 18.0 TG tok/s 2332 TTFT (ms) 17.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rbkmp686",
   "source_url": "https://omlx.ai/benchmarks/performance/rbkmp686",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-dgaafqsy-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: v3jt2615; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 827.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 827.9 78.0 13.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dgaafqsy",
   "source_url": "https://omlx.ai/benchmarks/performance/dgaafqsy",
   "tg_tokens": null,
   "tps": 78.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-dgaafqsy-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.6 GB at this context; duplicate traces of the same session: v3jt2615; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1032.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 1,032 82.6 13.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dgaafqsy",
   "source_url": "https://omlx.ai/benchmarks/performance/dgaafqsy",
   "tg_tokens": null,
   "tps": 82.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ayprqwft-ctx1024",
   "model": "Muse Glimmer",
   "notes": "oMLX community benchmark trace; peak memory 28.2 GB at this context; duplicate traces of the same session: nmgq51lz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 403.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Muse-Glimmer-30B-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 403.5 9.7 28.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ayprqwft",
   "source_url": "https://omlx.ai/benchmarks/performance/ayprqwft",
   "tg_tokens": null,
   "tps": 9.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ayprqwft-ctx4096",
   "model": "Muse Glimmer",
   "notes": "oMLX community benchmark trace; peak memory 28.4 GB at this context; duplicate traces of the same session: nmgq51lz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 423.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Muse-Glimmer-30B-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 423.0 9.4 28.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ayprqwft",
   "source_url": "https://omlx.ai/benchmarks/performance/ayprqwft",
   "tg_tokens": null,
   "tps": 9.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-h4x4k4lu-ctx131072",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 MLX build; oMLX community benchmark trace, single run at 128k context; peak memory 28.2 GB; TTFT 142152 ms; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 922.1,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Ornith-1.5-35B-A3B-MLX-oQ4-fp16 on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 128k tokens 922.1 PP tok/s 63.4 TG tok/s 142152 TTFT (ms) 28.2 Peak mem",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h4x4k4lu",
   "source_url": "https://omlx.ai/benchmarks/performance/h4x4k4lu",
   "tg_tokens": null,
   "tps": 63.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-hp60lwd0-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 MLX build; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: i1wtvyfr; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2167.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Ornith-1.5-35B-A3B-MLX-oQ4-fp16 on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 1k 2,167 98.1 20.2 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hp60lwd0",
   "source_url": "https://omlx.ai/benchmarks/performance/hp60lwd0",
   "tg_tokens": null,
   "tps": 98.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-hp60lwd0-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 MLX build; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: i1wtvyfr; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2819.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Ornith-1.5-35B-A3B-MLX-oQ4-fp16 on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 4k 2,819 95.7 22.1 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hp60lwd0",
   "source_url": "https://omlx.ai/benchmarks/performance/hp60lwd0",
   "tg_tokens": null,
   "tps": 95.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-r02q9vxu-ctx16384",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: ij5ni9tz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2888.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,888 128.6 20.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r02q9vxu",
   "source_url": "https://omlx.ai/benchmarks/performance/r02q9vxu",
   "tg_tokens": null,
   "tps": 128.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-r02q9vxu-ctx65536",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: ij5ni9tz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1863.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 1,863 74.3 21.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r02q9vxu",
   "source_url": "https://omlx.ai/benchmarks/performance/r02q9vxu",
   "tg_tokens": null,
   "tps": 74.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ub52gen3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 0ba69ne0, 0yaw2ybc, dgy7capg, evvyxjep; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 760.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 760.5 25.8 16.0 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ub52gen3",
   "source_url": "https://omlx.ai/benchmarks/performance/ub52gen3",
   "tg_tokens": null,
   "tps": 25.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ub52gen3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 0ba69ne0, 0yaw2ybc, dgy7capg, evvyxjep; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 814.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 814.6 25.6 17.0 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ub52gen3",
   "source_url": "https://omlx.ai/benchmarks/performance/ub52gen3",
   "tg_tokens": null,
   "tps": 25.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ub52gen3-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 0ba69ne0, 0yaw2ybc, dgy7capg, evvyxjep; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 812.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 8k 812.8 25.1 17.3 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ub52gen3",
   "source_url": "https://omlx.ai/benchmarks/performance/ub52gen3",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ub52gen3-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 0ba69ne0, 0yaw2ybc, dgy7capg, evvyxjep; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 790.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 790.4 24.1 17.8 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ub52gen3",
   "source_url": "https://omlx.ai/benchmarks/performance/ub52gen3",
   "tg_tokens": null,
   "tps": 24.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ub52gen3-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 0ba69ne0, 0yaw2ybc, dgy7capg, evvyxjep; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 740.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 32k 740.5 23.3 18.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ub52gen3",
   "source_url": "https://omlx.ai/benchmarks/performance/ub52gen3",
   "tg_tokens": null,
   "tps": 23.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-3ebrh1i2-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.4 GB at this context; duplicate traces of the same session: i3ero1cr; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 106.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 4k 106.1 10.9 32.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ebrh1i2",
   "source_url": "https://omlx.ai/benchmarks/performance/3ebrh1i2",
   "tg_tokens": null,
   "tps": 10.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-3ebrh1i2-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 34.2 GB at this context; duplicate traces of the same session: i3ero1cr; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 98.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 8bit; 16k 98.2 10.9 34.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ebrh1i2",
   "source_url": "https://omlx.ai/benchmarks/performance/3ebrh1i2",
   "tg_tokens": null,
   "tps": 10.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-beccm28a-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: 8osjnzm5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 897.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 897.5 17.7 29.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/beccm28a",
   "source_url": "https://omlx.ai/benchmarks/performance/beccm28a",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-beccm28a-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: 8osjnzm5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 774.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 774.7 17.1 30.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/beccm28a",
   "source_url": "https://omlx.ai/benchmarks/performance/beccm28a",
   "tg_tokens": null,
   "tps": 17.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-qqs35wlc-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 522dqbeo; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 131.3 15.3 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qqs35wlc",
   "source_url": "https://omlx.ai/benchmarks/performance/qqs35wlc",
   "tg_tokens": null,
   "tps": 15.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-qqs35wlc-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 522dqbeo; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 124.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 124.8 14.4 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qqs35wlc",
   "source_url": "https://omlx.ai/benchmarks/performance/qqs35wlc",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-nkw2ca2z-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 38.7 GB at this context; duplicate traces of the same session: z57m3ns2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 240.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 240.1 32.5 38.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nkw2ca2z",
   "source_url": "https://omlx.ai/benchmarks/performance/nkw2ca2z",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-nkw2ca2z-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized quality build; oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: z57m3ns2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 239.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Quality on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 239.6 33.4 37.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nkw2ca2z",
   "source_url": "https://omlx.ai/benchmarks/performance/nkw2ca2z",
   "tg_tokens": null,
   "tps": 33.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-oi9v6qnn-ctx16384",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: 9zljj0dc; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 476.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 476.9 32.8 26.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oi9v6qnn",
   "source_url": "https://omlx.ai/benchmarks/performance/oi9v6qnn",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-oi9v6qnn-ctx65536",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 29.3 GB at this context; duplicate traces of the same session: 9zljj0dc; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 385.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 385.5 24.1 29.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oi9v6qnn",
   "source_url": "https://omlx.ai/benchmarks/performance/oi9v6qnn",
   "tg_tokens": null,
   "tps": 24.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-rt47yfbm-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: vhk102sd; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 153.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 153.3 19.2 16.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rt47yfbm",
   "source_url": "https://omlx.ai/benchmarks/performance/rt47yfbm",
   "tg_tokens": null,
   "tps": 19.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M1 Max",
   "id": "omlx-rt47yfbm-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 32.3 GB at this context; duplicate traces of the same session: vhk102sd; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 163.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 163.9 6.4 32.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rt47yfbm",
   "source_url": "https://omlx.ai/benchmarks/performance/rt47yfbm",
   "tg_tokens": null,
   "tps": 6.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-3xudx6rs-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: u296oop3; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 151.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 151.0 25.2 19.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3xudx6rs",
   "source_url": "https://omlx.ai/benchmarks/performance/3xudx6rs",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-3xudx6rs-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: u296oop3; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 161.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 161.2 24.9 23.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3xudx6rs",
   "source_url": "https://omlx.ai/benchmarks/performance/3xudx6rs",
   "tg_tokens": null,
   "tps": 24.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-kve0vaw3-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: kzgy5jgt; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 147.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 147.3 18.7 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kve0vaw3",
   "source_url": "https://omlx.ai/benchmarks/performance/kve0vaw3",
   "tg_tokens": null,
   "tps": 18.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-kve0vaw3-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: kzgy5jgt; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 161.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 161.4 22.8 23.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kve0vaw3",
   "source_url": "https://omlx.ai/benchmarks/performance/kve0vaw3",
   "tg_tokens": null,
   "tps": 22.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-3zlc79ew-ctx16384",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: 6celel5s; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3026.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-35B-A3B-Distill-mlx-4Bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 3,026 92.1 20.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3zlc79ew",
   "source_url": "https://omlx.ai/benchmarks/performance/3zlc79ew",
   "tg_tokens": null,
   "tps": 92.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-3zlc79ew-ctx65536",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 6celel5s; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1971.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-35B-A3B-Distill-mlx-4Bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 1,971 63.9 21.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3zlc79ew",
   "source_url": "https://omlx.ai/benchmarks/performance/3zlc79ew",
   "tg_tokens": null,
   "tps": 63.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-vflsqrro-ctx199680",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "fp16 MTP build; oMLX community benchmark trace, single run at 195k context; peak memory 43.0 GB; TTFT 383713 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 521.2,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Qwen3.8-35B-A3B-Distill-oQ8-fp16-mtp on M1 Ultra (64c); M1 Ultra (64c) \u00b7 128 GB \u00b7 8bit; 195k tokens 521.2 PP tok/s 37.2 TG tok/s 383713 TTFT (ms) 43.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vflsqrro",
   "source_url": "https://omlx.ai/benchmarks/performance/vflsqrro",
   "tg_tokens": null,
   "tps": 37.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-drowv8sb-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.2 GB at this context; duplicate traces of the same session: 5gqutopa, aiu3144z, ha47zzty, kavjxsyx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1240.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,240 72.0 69.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/drowv8sb",
   "source_url": "https://omlx.ai/benchmarks/performance/drowv8sb",
   "tg_tokens": null,
   "tps": 72.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-drowv8sb-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.3 GB at this context; duplicate traces of the same session: 5gqutopa, aiu3144z, ha47zzty, kavjxsyx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2270.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,270 68.0 70.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/drowv8sb",
   "source_url": "https://omlx.ai/benchmarks/performance/drowv8sb",
   "tg_tokens": null,
   "tps": 68.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-drowv8sb-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: 5gqutopa, aiu3144z, ha47zzty, kavjxsyx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2205.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,205 66.1 71.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/drowv8sb",
   "source_url": "https://omlx.ai/benchmarks/performance/drowv8sb",
   "tg_tokens": null,
   "tps": 66.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-drowv8sb-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.6 GB at this context; duplicate traces of the same session: 5gqutopa, aiu3144z, ha47zzty, kavjxsyx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2172.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,172 68.5 71.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/drowv8sb",
   "source_url": "https://omlx.ai/benchmarks/performance/drowv8sb",
   "tg_tokens": null,
   "tps": 68.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-drowv8sb-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.7 GB at this context; duplicate traces of the same session: 5gqutopa, aiu3144z, ha47zzty, kavjxsyx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2047.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,047 61.7 72.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/drowv8sb",
   "source_url": "https://omlx.ai/benchmarks/performance/drowv8sb",
   "tg_tokens": null,
   "tps": 61.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-zzsk1tzr-ctx1024",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 286.0,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 2bit; 1k 286.0 23.6 11.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zzsk1tzr",
   "source_url": "https://omlx.ai/benchmarks/performance/zzsk1tzr",
   "tg_tokens": null,
   "tps": 23.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-zzsk1tzr-ctx4096",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 14.2 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 330.8,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 2bit; 4k 330.8 22.6 14.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zzsk1tzr",
   "source_url": "https://omlx.ai/benchmarks/performance/zzsk1tzr",
   "tg_tokens": null,
   "tps": 22.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-o516gfcb-ctx1024",
   "model": "Ternary-Bonsai-2-9B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 5.7 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 967.8,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-9B-mlx-2bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 2bit; 1k 967.8 74.9 5.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o516gfcb",
   "source_url": "https://omlx.ai/benchmarks/performance/o516gfcb",
   "tg_tokens": null,
   "tps": 74.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-o516gfcb-ctx4096",
   "model": "Ternary-Bonsai-2-9B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1090.0,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-9B-mlx-2bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 2bit; 4k 1,090 70.8 6.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o516gfcb",
   "source_url": "https://omlx.ai/benchmarks/performance/o516gfcb",
   "tg_tokens": null,
   "tps": 70.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-tkp0k9ra-ctx16384",
   "model": "gemma4 26B.A4B",
   "notes": "QAT OptiQ build; oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: iko59mvk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2081.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,081 49.6 21.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tkp0k9ra",
   "source_url": "https://omlx.ai/benchmarks/performance/tkp0k9ra",
   "tg_tokens": null,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-tkp0k9ra-ctx65536",
   "model": "gemma4 26B.A4B",
   "notes": "QAT OptiQ build; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: iko59mvk; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1438.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-qat-OptiQ-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 1,438 32.8 23.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tkp0k9ra",
   "source_url": "https://omlx.ai/benchmarks/performance/tkp0k9ra",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-14denid1-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 6.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1251.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E4B-it-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 1,251 47.5 6.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/14denid1",
   "source_url": "https://omlx.ai/benchmarks/performance/14denid1",
   "tg_tokens": null,
   "tps": 47.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-14denid1-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1292.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E4B-it-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,292 45.4 6.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/14denid1",
   "source_url": "https://omlx.ai/benchmarks/performance/14denid1",
   "tg_tokens": null,
   "tps": 45.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-mn6vs93m-ctx16384",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "mlx-community build; oMLX community benchmark trace, single run at 16k context; peak memory 20.1 GB; TTFT 63502 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 258.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "mlx-community--MiniCPM-V-4.7-35B-A3B-mxfp4 on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k tokens 258.0 PP tok/s 18.2 TG tok/s 63502 TTFT (ms) 20.1 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mn6vs93m",
   "source_url": "https://omlx.ai/benchmarks/performance/mn6vs93m",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-16cqrty9-ctx4096",
   "model": "Huihui-MiniCPM-V-4.6-abliterated",
   "notes": "MLX build; oMLX community benchmark trace, single run at 4k context; peak memory 3.5 GB; TTFT 2252 ms; duplicate traces of the same session: none; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1819.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Huihui-MiniCPM-V-4.6-abliterated-mlx-bf16 on M4 (8c); M4 (8c) \u00b7 16 GB \u00b7 bf16; 4k tokens 1,819 PP tok/s 57.7 TG tok/s 2252 TTFT (ms) 3.5 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/16cqrty9",
   "source_url": "https://omlx.ai/benchmarks/performance/16cqrty9",
   "tg_tokens": null,
   "tps": 57.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-5akqortz-ctx1024",
   "model": "Hy-Memory-30B-A3B",
   "notes": "MLX Mixed4-6 build; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 0zx3zdra, 30154e3x, cfcfjj0t, luny408e, n9omx1ey, r5ql05wb, ujayxrek; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1340.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Hy-Memory-30B-A3B-MLX-Mixed4-6-omlx on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 1,340 108.6 17.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5akqortz",
   "source_url": "https://omlx.ai/benchmarks/performance/5akqortz",
   "tg_tokens": null,
   "tps": 108.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-5akqortz-ctx4096",
   "model": "Hy-Memory-30B-A3B",
   "notes": "MLX Mixed4-6 build; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 0zx3zdra, 30154e3x, cfcfjj0t, luny408e, n9omx1ey, r5ql05wb, ujayxrek; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1466.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Hy-Memory-30B-A3B-MLX-Mixed4-6-omlx on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 unknown; 4k 1,466 98.0 18.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5akqortz",
   "source_url": "https://omlx.ai/benchmarks/performance/5akqortz",
   "tg_tokens": null,
   "tps": 98.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-5akqortz-ctx8192",
   "model": "Hy-Memory-30B-A3B",
   "notes": "MLX Mixed4-6 build; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: 0zx3zdra, 30154e3x, cfcfjj0t, luny408e, n9omx1ey, r5ql05wb, ujayxrek; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1266.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Hy-Memory-30B-A3B-MLX-Mixed4-6-omlx on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 1,266 82.6 18.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5akqortz",
   "source_url": "https://omlx.ai/benchmarks/performance/5akqortz",
   "tg_tokens": null,
   "tps": 82.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-5akqortz-ctx16384",
   "model": "Hy-Memory-30B-A3B",
   "notes": "MLX Mixed4-6 build; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 0zx3zdra, 30154e3x, cfcfjj0t, luny408e, n9omx1ey, r5ql05wb, ujayxrek; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 975.8,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Hy-Memory-30B-A3B-MLX-Mixed4-6-omlx on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 975.8 72.0 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5akqortz",
   "source_url": "https://omlx.ai/benchmarks/performance/5akqortz",
   "tg_tokens": null,
   "tps": 72.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-5akqortz-ctx32768",
   "model": "Hy-Memory-30B-A3B",
   "notes": "MLX Mixed4-6 build; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 0zx3zdra, 30154e3x, cfcfjj0t, luny408e, n9omx1ey, r5ql05wb, ujayxrek; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 663.6,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Hy-Memory-30B-A3B-MLX-Mixed4-6-omlx on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 unknown; 32k 663.6 54.4 20.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5akqortz",
   "source_url": "https://omlx.ai/benchmarks/performance/5akqortz",
   "tg_tokens": null,
   "tps": 54.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-5akqortz-ctx65536",
   "model": "Hy-Memory-30B-A3B",
   "notes": "MLX Mixed4-6 build; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: 0zx3zdra, 30154e3x, cfcfjj0t, luny408e, n9omx1ey, r5ql05wb, ujayxrek; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 404.2,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Hy-Memory-30B-A3B-MLX-Mixed4-6-omlx on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 unknown; 64k 404.2 37.0 23.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5akqortz",
   "source_url": "https://omlx.ai/benchmarks/performance/5akqortz",
   "tg_tokens": null,
   "tps": 37.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-5akqortz-ctx131072",
   "model": "Hy-Memory-30B-A3B",
   "notes": "MLX Mixed4-6 build; oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: 0zx3zdra, 30154e3x, cfcfjj0t, luny408e, n9omx1ey, r5ql05wb, ujayxrek; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 226.7,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Hy-Memory-30B-A3B-MLX-Mixed4-6-omlx on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 unknown; 128k 226.7 24.2 29.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5akqortz",
   "source_url": "https://omlx.ai/benchmarks/performance/5akqortz",
   "tg_tokens": null,
   "tps": 24.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-5akqortz-ctx199680",
   "model": "Hy-Memory-30B-A3B",
   "notes": "MLX Mixed4-6 build; oMLX community benchmark trace; peak memory 36.0 GB at this context; duplicate traces of the same session: 0zx3zdra, 30154e3x, cfcfjj0t, luny408e, n9omx1ey, r5ql05wb, ujayxrek; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 155.1,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Hy-Memory-30B-A3B-MLX-Mixed4-6-omlx on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 unknown; 195k 155.1 17.8 36.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5akqortz",
   "source_url": "https://omlx.ai/benchmarks/performance/5akqortz",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9uoaqeg8-ctx4096",
   "model": "Jundot--GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 174.9 GB at this context; duplicate traces of the same session: lzkw3fl5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 592.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 592.6 38.6 174.9 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9uoaqeg8",
   "source_url": "https://omlx.ai/benchmarks/performance/9uoaqeg8",
   "tg_tokens": null,
   "tps": 38.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9uoaqeg8-ctx16384",
   "model": "Jundot--GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 175.5 GB at this context; duplicate traces of the same session: lzkw3fl5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 525.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--GLM-5.3-Flash-oQ4e on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 525.0 38.3 175.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9uoaqeg8",
   "source_url": "https://omlx.ai/benchmarks/performance/9uoaqeg8",
   "tg_tokens": null,
   "tps": 38.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-pkdxhklq-ctx1024",
   "model": "Ling-mini-2.0",
   "notes": "DWQ build; oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: j6jbvr9z; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2266.0,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ling-mini-2.0-2bit-DWQ on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 2bit; 1k 2,266 225.0 6.2 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pkdxhklq",
   "source_url": "https://omlx.ai/benchmarks/performance/pkdxhklq",
   "tg_tokens": null,
   "tps": 225.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-pkdxhklq-ctx16384",
   "model": "Ling-mini-2.0",
   "notes": "DWQ build; oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: j6jbvr9z; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1651.0,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ling-mini-2.0-2bit-DWQ on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 2bit; 16k 1,651 140.1 7.0 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pkdxhklq",
   "source_url": "https://omlx.ai/benchmarks/performance/pkdxhklq",
   "tg_tokens": null,
   "tps": 140.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-b1qooxmc-ctx1024",
   "model": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B",
   "notes": "oMLX community benchmark trace; peak memory 54.0 GB at this context; duplicate traces of the same session: wbm0oigc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 517.2,
   "provenance": "community",
   "quant": "5bit",
   "quote": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B-5bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 517.2 54.3 54.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b1qooxmc",
   "source_url": "https://omlx.ai/benchmarks/performance/b1qooxmc",
   "tg_tokens": null,
   "tps": 54.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-b1qooxmc-ctx4096",
   "model": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B",
   "notes": "oMLX community benchmark trace; peak memory 57.0 GB at this context; duplicate traces of the same session: wbm0oigc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 552.1,
   "provenance": "community",
   "quant": "5bit",
   "quote": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B-5bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 552.1 54.1 57.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b1qooxmc",
   "source_url": "https://omlx.ai/benchmarks/performance/b1qooxmc",
   "tg_tokens": null,
   "tps": 54.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-g8zxc3qw-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 8.0 GB at this context; duplicate traces of the same session: strwzcu9; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 363.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-9B-oQ4e-fp16-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 363.3 44.6 8.0 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g8zxc3qw",
   "source_url": "https://omlx.ai/benchmarks/performance/g8zxc3qw",
   "tg_tokens": null,
   "tps": 44.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-g8zxc3qw-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 8.5 GB at this context; duplicate traces of the same session: strwzcu9; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 377.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-9B-oQ4e-fp16-mtp on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 4k 377.4 57.9 8.5 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g8zxc3qw",
   "source_url": "https://omlx.ai/benchmarks/performance/g8zxc3qw",
   "tg_tokens": null,
   "tps": 57.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-grtoubo9-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: he3p6xrz, ibp13seh; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1945.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 1,945 108.6 24.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/grtoubo9",
   "source_url": "https://omlx.ai/benchmarks/performance/grtoubo9",
   "tg_tokens": null,
   "tps": 108.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-grtoubo9-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: he3p6xrz, ibp13seh; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2518.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 2,518 110.9 28.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/grtoubo9",
   "source_url": "https://omlx.ai/benchmarks/performance/grtoubo9",
   "tg_tokens": null,
   "tps": 110.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-grtoubo9-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 30.1 GB at this context; duplicate traces of the same session: he3p6xrz, ibp13seh; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1761.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 1,761 97.0 30.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/grtoubo9",
   "source_url": "https://omlx.ai/benchmarks/performance/grtoubo9",
   "tg_tokens": null,
   "tps": 97.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-xsz8uw01-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 2u9op1zc, yw13b47i; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1053.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 1,053 89.9 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xsz8uw01",
   "source_url": "https://omlx.ai/benchmarks/performance/xsz8uw01",
   "tg_tokens": null,
   "tps": 89.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-xsz8uw01-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: 2u9op1zc, yw13b47i; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 899.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 899.1 77.7 20.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xsz8uw01",
   "source_url": "https://omlx.ai/benchmarks/performance/xsz8uw01",
   "tg_tokens": null,
   "tps": 77.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-xsz8uw01-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 2u9op1zc, yw13b47i; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 733.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 733.9 67.1 20.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xsz8uw01",
   "source_url": "https://omlx.ai/benchmarks/performance/xsz8uw01",
   "tg_tokens": null,
   "tps": 67.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1pfb2zu8-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: o7qi9skh; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 729.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 729.0 26.0 16.0 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1pfb2zu8",
   "source_url": "https://omlx.ai/benchmarks/performance/1pfb2zu8",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1pfb2zu8-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: o7qi9skh; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 801.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 801.0 25.6 17.0 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1pfb2zu8",
   "source_url": "https://omlx.ai/benchmarks/performance/1pfb2zu8",
   "tg_tokens": null,
   "tps": 25.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-k2rr3thy-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: g0pvdpth; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 465.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 465.6 18.2 16.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k2rr3thy",
   "source_url": "https://omlx.ai/benchmarks/performance/k2rr3thy",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-k2rr3thy-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: g0pvdpth; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 503.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 503.3 17.9 17.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k2rr3thy",
   "source_url": "https://omlx.ai/benchmarks/performance/k2rr3thy",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-t7heexau-ctx1024",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 45.6 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 316.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Alis-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 316.4 48.4 45.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t7heexau",
   "source_url": "https://omlx.ai/benchmarks/performance/t7heexau",
   "tg_tokens": null,
   "tps": 48.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-t7heexau-ctx4096",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 49.1 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 415.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Alis-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 415.9 41.9 49.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t7heexau",
   "source_url": "https://omlx.ai/benchmarks/performance/t7heexau",
   "tg_tokens": null,
   "tps": 41.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-t7heexau-ctx32768",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 52.2 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 354.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Alis-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 354.6 42.0 52.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t7heexau",
   "source_url": "https://omlx.ai/benchmarks/performance/t7heexau",
   "tg_tokens": null,
   "tps": 42.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-8yjlmb2a-ctx1024",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: mq1nd6ey; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 381.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Alis-MLX-6bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 381.9 42.7 27.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8yjlmb2a",
   "source_url": "https://omlx.ai/benchmarks/performance/8yjlmb2a",
   "tg_tokens": null,
   "tps": 42.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-8yjlmb2a-ctx4096",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 30.9 GB at this context; duplicate traces of the same session: mq1nd6ey; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 410.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Alis-MLX-6bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 410.7 41.4 30.9 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8yjlmb2a",
   "source_url": "https://omlx.ai/benchmarks/performance/8yjlmb2a",
   "tg_tokens": null,
   "tps": 41.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-8yjlmb2a-ctx32768",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 34.1 GB at this context; duplicate traces of the same session: mq1nd6ey; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 348.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Alis-MLX-6bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 348.9 36.6 34.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8yjlmb2a",
   "source_url": "https://omlx.ai/benchmarks/performance/8yjlmb2a",
   "tg_tokens": null,
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-kqs79kb9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "fp16 MLX build, MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: 68iyc85d; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 146.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-MLX-oQ6e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 6bit; 1k 146.6 17.9 29.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kqs79kb9",
   "source_url": "https://omlx.ai/benchmarks/performance/kqs79kb9",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-kqs79kb9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "fp16 MLX build, MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.8 GB at this context; duplicate traces of the same session: 68iyc85d; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 149.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-MLX-oQ6e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 6bit; 4k 149.7 17.0 32.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kqs79kb9",
   "source_url": "https://omlx.ai/benchmarks/performance/kqs79kb9",
   "tg_tokens": null,
   "tps": 17.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-tt5xryi6-ctx1024",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: icxgqrda; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 253.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 253.7 67.9 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tt5xryi6",
   "source_url": "https://omlx.ai/benchmarks/performance/tt5xryi6",
   "tg_tokens": null,
   "tps": 67.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-tt5xryi6-ctx8192",
   "model": "Qwen3.8-27B-MTPLX",
   "notes": "Optimized-Speed build; oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: icxgqrda; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 253.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 253.0 34.4 27.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tt5xryi6",
   "source_url": "https://omlx.ai/benchmarks/performance/tt5xryi6",
   "tg_tokens": null,
   "tps": 34.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-mwqu6dwe-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized speed FP16 build; oMLX community benchmark trace, single run at 1k context; peak memory 19.9 GB; TTFT 8579 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 119.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed-FP16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k tokens 119.4 PP tok/s 15.9 TG tok/s 8579 TTFT (ms) 19.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mwqu6dwe",
   "source_url": "https://omlx.ai/benchmarks/performance/mwqu6dwe",
   "tg_tokens": null,
   "tps": 15.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-8rtns3or-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: ves5xwwm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 259.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 259.0 28.0 16.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8rtns3or",
   "source_url": "https://omlx.ai/benchmarks/performance/8rtns3or",
   "tg_tokens": null,
   "tps": 28.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-8rtns3or-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: ves5xwwm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 268.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 268.6 26.6 20.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8rtns3or",
   "source_url": "https://omlx.ai/benchmarks/performance/8rtns3or",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-hnfpe5jc-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 69.3 GB; TTFT 2294 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 446.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 446.3 PP tok/s 52.7 TG tok/s 2294 TTFT (ms) 69.3 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hnfpe5jc",
   "source_url": "https://omlx.ai/benchmarks/performance/hnfpe5jc",
   "tg_tokens": null,
   "tps": 52.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vtj9kcop-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.6 GB; TTFT 1064 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 962.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 962.4 PP tok/s 63.0 TG tok/s 1064 TTFT (ms) 70.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vtj9kcop",
   "source_url": "https://omlx.ai/benchmarks/performance/vtj9kcop",
   "tg_tokens": null,
   "tps": 63.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-xb63in4x-ctx1024",
   "model": "Spark-X2.5-4B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: nayx6kon; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 293.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Spark-X2.5-4B-oQ6e-fp16 on M4 (8c); M4 (8c) \u00b7 16 GB \u00b7 6bit; 1k 293.1 26.1 3.9 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xb63in4x",
   "source_url": "https://omlx.ai/benchmarks/performance/xb63in4x",
   "tg_tokens": null,
   "tps": 26.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-xb63in4x-ctx4096",
   "model": "Spark-X2.5-4B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 4.0 GB at this context; duplicate traces of the same session: nayx6kon; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 286.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Spark-X2.5-4B-oQ6e-fp16 on M4 (8c); M4 (8c) \u00b7 16 GB \u00b7 6bit; 4k 286.1 25.0 4.0 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xb63in4x",
   "source_url": "https://omlx.ai/benchmarks/performance/xb63in4x",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-sya9ox8c-ctx65536",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 16.7 GB; TTFT 122604 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 534.5,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "gemma-4-26b-a4b-it-mxfp4 on M4 Pro (20c); M4 Pro (20c) \u00b7 24 GB \u00b7 4bit; 64k tokens 534.5 PP tok/s 31.5 TG tok/s 122604 TTFT (ms) 16.7 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sya9ox8c",
   "source_url": "https://omlx.ai/benchmarks/performance/sya9ox8c",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-octj35sr-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 14.2 GB; TTFT 25152 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 651.4,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "gemma-4-26b-a4b-it-mxfp4 on M4 Pro (20c); M4 Pro (20c) \u00b7 24 GB \u00b7 4bit; 16k tokens 651.4 PP tok/s 52.5 TG tok/s 25152 TTFT (ms) 14.2 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/octj35sr",
   "source_url": "https://omlx.ai/benchmarks/performance/octj35sr",
   "tg_tokens": null,
   "tps": 52.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-tkq3zocn-ctx32768",
   "model": "gemma-4-26B-A4B-it",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 15.0 GB; TTFT 55066 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 595.1,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "gemma-4-26b-a4b-it-mxfp4 on M4 Pro (20c); M4 Pro (20c) \u00b7 24 GB \u00b7 4bit; 32k tokens 595.1 PP tok/s 42.8 TG tok/s 55066 TTFT (ms) 15.0 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tkq3zocn",
   "source_url": "https://omlx.ai/benchmarks/performance/tkq3zocn",
   "tg_tokens": null,
   "tps": 42.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-vvmzclft-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 6.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1251.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E4B-it-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 1,251 47.5 6.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vvmzclft",
   "source_url": "https://omlx.ai/benchmarks/performance/vvmzclft",
   "tg_tokens": null,
   "tps": 47.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-vvmzclft-ctx4096",
   "model": "Gemma 4 E4B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1292.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E4B-it-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,292 45.4 6.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vvmzclft",
   "source_url": "https://omlx.ai/benchmarks/performance/vvmzclft",
   "tg_tokens": null,
   "tps": 45.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-wsgz5vjy-ctx1024",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: z0gl0xa7; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1037.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Mellum2-12B-A2.5B-Instruct-mxfp4 on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 1k 1,037 116.5 6.7 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wsgz5vjy",
   "source_url": "https://omlx.ai/benchmarks/performance/wsgz5vjy",
   "tg_tokens": null,
   "tps": 116.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-wsgz5vjy-ctx16384",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 6.9 GB at this context; duplicate traces of the same session: z0gl0xa7; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 860.6,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Mellum2-12B-A2.5B-Instruct-mxfp4 on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 4bit; 16k 860.6 105.0 6.9 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wsgz5vjy",
   "source_url": "https://omlx.ai/benchmarks/performance/wsgz5vjy",
   "tg_tokens": null,
   "tps": 105.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-jreww5ok-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.2 GB at this context; duplicate traces of the same session: b30zftwx; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 707.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-6bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 1k 707.4 47.9 28.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jreww5ok",
   "source_url": "https://omlx.ai/benchmarks/performance/jreww5ok",
   "tg_tokens": null,
   "tps": 47.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-jreww5ok-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 29.2 GB at this context; duplicate traces of the same session: b30zftwx; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 702.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-6bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 4k 702.5 34.1 29.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jreww5ok",
   "source_url": "https://omlx.ai/benchmarks/performance/jreww5ok",
   "tg_tokens": null,
   "tps": 34.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-c4us32aa-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: zh5q00be; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2166.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16 on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 1k 2,166 97.8 20.1 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c4us32aa",
   "source_url": "https://omlx.ai/benchmarks/performance/c4us32aa",
   "tg_tokens": null,
   "tps": 97.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-c4us32aa-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: zh5q00be; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2840.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16 on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 4k 2,840 39.8 21.9 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c4us32aa",
   "source_url": "https://omlx.ai/benchmarks/performance/c4us32aa",
   "tg_tokens": null,
   "tps": 39.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-b8fqk6fc-ctx131072",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "fp16 build; oMLX community benchmark trace, single run at 128k context; peak memory 28.0 GB; TTFT 142098 ms; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 922.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16 on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 128k tokens 922.4 PP tok/s 62.8 TG tok/s 142098 TTFT (ms) 28.0 Peak mem",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b8fqk6fc",
   "source_url": "https://omlx.ai/benchmarks/performance/b8fqk6fc",
   "tg_tokens": null,
   "tps": 62.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-adw9mscs-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: x0j2fr3w, yjveq7ev; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 830.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 1k 830.5 62.6 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/adw9mscs",
   "source_url": "https://omlx.ai/benchmarks/performance/adw9mscs",
   "tg_tokens": null,
   "tps": 62.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-adw9mscs-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: x0j2fr3w, yjveq7ev; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 943.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k 943.4 69.2 21.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/adw9mscs",
   "source_url": "https://omlx.ai/benchmarks/performance/adw9mscs",
   "tg_tokens": null,
   "tps": 69.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-adw9mscs-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: x0j2fr3w, yjveq7ev; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 860.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 8k 860.0 62.7 21.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/adw9mscs",
   "source_url": "https://omlx.ai/benchmarks/performance/adw9mscs",
   "tg_tokens": null,
   "tps": 62.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wwdoeooh-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 53.7 GB at this context; duplicate traces of the same session: 3rr0gao9, 7rjtz2jl, p13quycz, t7xfhvv0, uptmem36; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 855.8,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.8-27B on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 855.8 29.7 53.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wwdoeooh",
   "source_url": "https://omlx.ai/benchmarks/performance/wwdoeooh",
   "tg_tokens": null,
   "tps": 29.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wwdoeooh-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 55.2 GB at this context; duplicate traces of the same session: 3rr0gao9, 7rjtz2jl, p13quycz, t7xfhvv0, uptmem36; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1037.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.8-27B on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 4k 1,037 36.3 55.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wwdoeooh",
   "source_url": "https://omlx.ai/benchmarks/performance/wwdoeooh",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wwdoeooh-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 55.7 GB at this context; duplicate traces of the same session: 3rr0gao9, 7rjtz2jl, p13quycz, t7xfhvv0, uptmem36; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1015.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.8-27B on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 1,015 22.5 55.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wwdoeooh",
   "source_url": "https://omlx.ai/benchmarks/performance/wwdoeooh",
   "tg_tokens": null,
   "tps": 22.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wwdoeooh-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 56.9 GB at this context; duplicate traces of the same session: 3rr0gao9, 7rjtz2jl, p13quycz, t7xfhvv0, uptmem36; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 955.9,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.8-27B on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 955.9 30.2 56.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wwdoeooh",
   "source_url": "https://omlx.ai/benchmarks/performance/wwdoeooh",
   "tg_tokens": null,
   "tps": 30.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wwdoeooh-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 59.1 GB at this context; duplicate traces of the same session: 3rr0gao9, 7rjtz2jl, p13quycz, t7xfhvv0, uptmem36; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 883.6,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.8-27B on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 32k 883.6 35.0 59.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wwdoeooh",
   "source_url": "https://omlx.ai/benchmarks/performance/wwdoeooh",
   "tg_tokens": null,
   "tps": 35.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wwdoeooh-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 63.7 GB at this context; duplicate traces of the same session: 3rr0gao9, 7rjtz2jl, p13quycz, t7xfhvv0, uptmem36; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 768.2,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen3.8-27B on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 64k 768.2 25.6 63.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wwdoeooh",
   "source_url": "https://omlx.ai/benchmarks/performance/wwdoeooh",
   "tg_tokens": null,
   "tps": 25.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-anb5feof-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: b5dh31km; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 54.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 64k 54.4 10.9 24.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/anb5feof",
   "source_url": "https://omlx.ai/benchmarks/performance/anb5feof",
   "tg_tokens": null,
   "tps": 10.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-anb5feof-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.7 GB at this context; duplicate traces of the same session: b5dh31km; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 40.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 128k 40.9 8.3 28.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/anb5feof",
   "source_url": "https://omlx.ai/benchmarks/performance/anb5feof",
   "tg_tokens": null,
   "tps": 8.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-0u6kxpd0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: ycipr0z1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 155.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 155.4 15.1 22.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0u6kxpd0",
   "source_url": "https://omlx.ai/benchmarks/performance/0u6kxpd0",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-0u6kxpd0-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: ycipr0z1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 155.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 155.8 14.1 23.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0u6kxpd0",
   "source_url": "https://omlx.ai/benchmarks/performance/0u6kxpd0",
   "tg_tokens": null,
   "tps": 14.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-zaptxauu-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: duaji8wq; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 783.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 783.9 36.6 16.9 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zaptxauu",
   "source_url": "https://omlx.ai/benchmarks/performance/zaptxauu",
   "tg_tokens": null,
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-zaptxauu-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: duaji8wq; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 790.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 790.0 36.6 18.1 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zaptxauu",
   "source_url": "https://omlx.ai/benchmarks/performance/zaptxauu",
   "tg_tokens": null,
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rdlv4jgf-ctx1024",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 45.6 GB at this context; duplicate traces of the same session: gxyy7dfc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 316.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Alis-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 316.4 48.4 45.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rdlv4jgf",
   "source_url": "https://omlx.ai/benchmarks/performance/rdlv4jgf",
   "tg_tokens": null,
   "tps": 48.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rdlv4jgf-ctx4096",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 49.1 GB at this context; duplicate traces of the same session: gxyy7dfc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 415.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Alis-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 415.9 41.9 49.1 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rdlv4jgf",
   "source_url": "https://omlx.ai/benchmarks/performance/rdlv4jgf",
   "tg_tokens": null,
   "tps": 41.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-rdlv4jgf-ctx32768",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 52.2 GB at this context; duplicate traces of the same session: gxyy7dfc; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 354.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Alis-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 354.6 42.0 52.2 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rdlv4jgf",
   "source_url": "https://omlx.ai/benchmarks/performance/rdlv4jgf",
   "tg_tokens": null,
   "tps": 42.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2qkth3f1-ctx1024",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 381.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Alis-MLX-6bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 381.9 42.7 27.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2qkth3f1",
   "source_url": "https://omlx.ai/benchmarks/performance/2qkth3f1",
   "tg_tokens": null,
   "tps": 42.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2qkth3f1-ctx4096",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 30.9 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 410.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Alis-MLX-6bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 410.7 41.4 30.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2qkth3f1",
   "source_url": "https://omlx.ai/benchmarks/performance/2qkth3f1",
   "tg_tokens": null,
   "tps": 41.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-2qkth3f1-ctx32768",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 34.1 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 348.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Alis-MLX-6bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 348.9 36.6 34.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2qkth3f1",
   "source_url": "https://omlx.ai/benchmarks/performance/2qkth3f1",
   "tg_tokens": null,
   "tps": 36.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-tp9g0snl-ctx1024",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 27.4 GB at this context; duplicate traces of the same session: 3bbuv81j, jk554ihf; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 377.9,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Alis-MLX-6bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 377.9 39.2 27.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tp9g0snl",
   "source_url": "https://omlx.ai/benchmarks/performance/tp9g0snl",
   "tg_tokens": null,
   "tps": 39.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-tp9g0snl-ctx4096",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 30.9 GB at this context; duplicate traces of the same session: 3bbuv81j, jk554ihf; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 409.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Alis-MLX-6bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 409.3 37.6 30.9 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tp9g0snl",
   "source_url": "https://omlx.ai/benchmarks/performance/tp9g0snl",
   "tg_tokens": null,
   "tps": 37.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-tp9g0snl-ctx32768",
   "model": "Qwen3.8-27B-Alis",
   "notes": "oMLX community benchmark trace; peak memory 34.1 GB at this context; duplicate traces of the same session: 3bbuv81j, jk554ihf; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 350.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Alis-MLX-6bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 32k 350.5 35.0 34.1 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tp9g0snl",
   "source_url": "https://omlx.ai/benchmarks/performance/tp9g0snl",
   "tg_tokens": null,
   "tps": 35.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-rt3evl78-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "fp16 MLX build, MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: bxpez9bg; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 145.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-MLX-oQ6e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 6bit; 1k 145.0 13.6 27.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rt3evl78",
   "source_url": "https://omlx.ai/benchmarks/performance/rt3evl78",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-rt3evl78-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "fp16 MLX build, MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.7 GB at this context; duplicate traces of the same session: bxpez9bg; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 150.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-MLX-oQ6e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 6bit; 4k 150.0 16.3 30.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rt3evl78",
   "source_url": "https://omlx.ai/benchmarks/performance/rt3evl78",
   "tg_tokens": null,
   "tps": 16.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-nq0rn3g7-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MPTLS optimized speed FP16 build; oMLX community benchmark trace, single run at 1k context; peak memory 23.4 GB; TTFT 8781 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 116.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MTPLX-Optimized-Speed-FP16 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k tokens 116.6 PP tok/s 15.2 TG tok/s 8781 TTFT (ms) 23.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nq0rn3g7",
   "source_url": "https://omlx.ai/benchmarks/performance/nq0rn3g7",
   "tg_tokens": null,
   "tps": 15.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-rakanwxu-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 17.6 GB; TTFT 8934 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 114.6,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k tokens 114.6 PP tok/s 16.8 TG tok/s 8934 TTFT (ms) 17.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rakanwxu",
   "source_url": "https://omlx.ai/benchmarks/performance/rakanwxu",
   "tg_tokens": null,
   "tps": 16.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-amfs66wp-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.9 GB at this context; duplicate traces of the same session: quf9s1c1; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 774.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 774.8 38.2 17.9 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/amfs66wp",
   "source_url": "https://omlx.ai/benchmarks/performance/amfs66wp",
   "tg_tokens": null,
   "tps": 38.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-amfs66wp-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: quf9s1c1; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 808.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 808.2 34.7 18.9 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/amfs66wp",
   "source_url": "https://omlx.ai/benchmarks/performance/amfs66wp",
   "tg_tokens": null,
   "tps": 34.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-f1zh23va-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: oz8ot9tp; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 638.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 638.9 39.3 17.7 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f1zh23va",
   "source_url": "https://omlx.ai/benchmarks/performance/f1zh23va",
   "tg_tokens": null,
   "tps": 39.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-f1zh23va-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: oz8ot9tp; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 759.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 759.0 49.9 18.8 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f1zh23va",
   "source_url": "https://omlx.ai/benchmarks/performance/f1zh23va",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-w36hlsjt-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: b0omd1og, dd3cuut3, fg2ynzev, g4n1sesw, h31wt9pb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 840.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 840.9 53.8 23.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w36hlsjt",
   "source_url": "https://omlx.ai/benchmarks/performance/w36hlsjt",
   "tg_tokens": null,
   "tps": 53.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-w36hlsjt-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: b0omd1og, dd3cuut3, fg2ynzev, g4n1sesw, h31wt9pb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 920.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 920.1 76.1 24.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w36hlsjt",
   "source_url": "https://omlx.ai/benchmarks/performance/w36hlsjt",
   "tg_tokens": null,
   "tps": 76.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-w36hlsjt-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: b0omd1og, dd3cuut3, fg2ynzev, g4n1sesw, h31wt9pb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 837.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 837.5 35.3 24.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w36hlsjt",
   "source_url": "https://omlx.ai/benchmarks/performance/w36hlsjt",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-w36hlsjt-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: b0omd1og, dd3cuut3, fg2ynzev, g4n1sesw, h31wt9pb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 807.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 807.8 48.0 26.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w36hlsjt",
   "source_url": "https://omlx.ai/benchmarks/performance/w36hlsjt",
   "tg_tokens": null,
   "tps": 48.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-w36hlsjt-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.3 GB at this context; duplicate traces of the same session: b0omd1og, dd3cuut3, fg2ynzev, g4n1sesw, h31wt9pb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 763.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 763.0 39.4 28.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w36hlsjt",
   "source_url": "https://omlx.ai/benchmarks/performance/w36hlsjt",
   "tg_tokens": null,
   "tps": 39.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-w36hlsjt-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 33.0 GB at this context; duplicate traces of the same session: b0omd1og, dd3cuut3, fg2ynzev, g4n1sesw, h31wt9pb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 675.7,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k 675.7 32.9 33.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w36hlsjt",
   "source_url": "https://omlx.ai/benchmarks/performance/w36hlsjt",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1t4anrgp-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.0 GB at this context; duplicate traces of the same session: jrqqu1rv, zg6o1vp9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 411.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 411.6 80.8 99.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1t4anrgp",
   "source_url": "https://omlx.ai/benchmarks/performance/1t4anrgp",
   "tg_tokens": null,
   "tps": 80.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1t4anrgp-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.4 GB at this context; duplicate traces of the same session: jrqqu1rv, zg6o1vp9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2032.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,032 50.7 100.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1t4anrgp",
   "source_url": "https://omlx.ai/benchmarks/performance/1t4anrgp",
   "tg_tokens": null,
   "tps": 50.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1t4anrgp-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: jrqqu1rv, zg6o1vp9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2736.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,736 73.4 100.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1t4anrgp",
   "source_url": "https://omlx.ai/benchmarks/performance/1t4anrgp",
   "tg_tokens": null,
   "tps": 73.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-4dj2ut6n-ctx1024",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 9.8 GB at this context; duplicate traces of the same session: v0qnwna0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1002.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2-12B-A2.5B-Instruct-6bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 6bit; 1k 1,002 87.5 9.8 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4dj2ut6n",
   "source_url": "https://omlx.ai/benchmarks/performance/4dj2ut6n",
   "tg_tokens": null,
   "tps": 87.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-4dj2ut6n-ctx16384",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 10.1 GB at this context; duplicate traces of the same session: v0qnwna0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 820.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2-12B-A2.5B-Instruct-6bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 6bit; 16k 820.0 80.6 10.1 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4dj2ut6n",
   "source_url": "https://omlx.ai/benchmarks/performance/4dj2ut6n",
   "tg_tokens": null,
   "tps": 80.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-nfyht767-ctx4096",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: xh5okvdr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1643.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,643 161.3 19.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nfyht767",
   "source_url": "https://omlx.ai/benchmarks/performance/nfyht767",
   "tg_tokens": null,
   "tps": 161.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-nfyht767-ctx16384",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: xh5okvdr; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3826.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 3,826 152.2 19.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nfyht767",
   "source_url": "https://omlx.ai/benchmarks/performance/nfyht767",
   "tg_tokens": null,
   "tps": 152.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-bm5j0zs8-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: wnimyxsf; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 312.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 32k 312.6 50.2 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bm5j0zs8",
   "source_url": "https://omlx.ai/benchmarks/performance/bm5j0zs8",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-bm5j0zs8-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: wnimyxsf; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 298.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 64k 298.8 43.7 24.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bm5j0zs8",
   "source_url": "https://omlx.ai/benchmarks/performance/bm5j0zs8",
   "tg_tokens": null,
   "tps": 43.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-lhlhon77-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.3 GB at this context; duplicate traces of the same session: x23yal4t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 247.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 247.1 28.6 19.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lhlhon77",
   "source_url": "https://omlx.ai/benchmarks/performance/lhlhon77",
   "tg_tokens": null,
   "tps": 28.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-lhlhon77-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: x23yal4t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 332.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 332.7 27.6 22.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lhlhon77",
   "source_url": "https://omlx.ai/benchmarks/performance/lhlhon77",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-z4k6dch9-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 906.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 906.1 32.5 16.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z4k6dch9",
   "source_url": "https://omlx.ai/benchmarks/performance/z4k6dch9",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-z4k6dch9-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 843.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 843.1 28.8 17.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z4k6dch9",
   "source_url": "https://omlx.ai/benchmarks/performance/z4k6dch9",
   "tg_tokens": null,
   "tps": 28.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-3difu0vn-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 19.2 GB; TTFT 8912 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 114.9,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k tokens 114.9 PP tok/s 16.6 TG tok/s 8912 TTFT (ms) 19.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3difu0vn",
   "source_url": "https://omlx.ai/benchmarks/performance/3difu0vn",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-1t4gukk4-ctx16384",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 41.5 GB; TTFT 10876 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1506.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k tokens 1,506 PP tok/s 44.8 TG tok/s 10876 TTFT (ms) 41.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1t4gukk4",
   "source_url": "https://omlx.ai/benchmarks/performance/1t4gukk4",
   "tg_tokens": null,
   "tps": 44.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-51i6j3vu-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 100.5 GB; TTFT 1468 ms; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 697.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 1k tokens 697.4 PP tok/s 76.6 TG tok/s 1468 TTFT (ms) 100.5 Peak mem",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/51i6j3vu",
   "source_url": "https://omlx.ai/benchmarks/performance/51i6j3vu",
   "tg_tokens": null,
   "tps": 76.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fmwp0ala-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.0 GB at this context; duplicate traces of the same session: 00g2pdks, 1k1y4wtq, f5f35c4p, kcybky8v, pe98fd1b, sp6n43hb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1920.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,920 80.8 99.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fmwp0ala",
   "source_url": "https://omlx.ai/benchmarks/performance/fmwp0ala",
   "tg_tokens": null,
   "tps": 80.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fmwp0ala-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.5 GB at this context; duplicate traces of the same session: 00g2pdks, 1k1y4wtq, f5f35c4p, kcybky8v, pe98fd1b, sp6n43hb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2788.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,788 76.8 99.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fmwp0ala",
   "source_url": "https://omlx.ai/benchmarks/performance/fmwp0ala",
   "tg_tokens": null,
   "tps": 76.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fmwp0ala-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.4 GB at this context; duplicate traces of the same session: 00g2pdks, 1k1y4wtq, f5f35c4p, kcybky8v, pe98fd1b, sp6n43hb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2971.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,971 76.7 100.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fmwp0ala",
   "source_url": "https://omlx.ai/benchmarks/performance/fmwp0ala",
   "tg_tokens": null,
   "tps": 76.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fmwp0ala-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: 00g2pdks, 1k1y4wtq, f5f35c4p, kcybky8v, pe98fd1b, sp6n43hb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2824.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,824 70.8 100.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fmwp0ala",
   "source_url": "https://omlx.ai/benchmarks/performance/fmwp0ala",
   "tg_tokens": null,
   "tps": 70.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fmwp0ala-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.2 GB at this context; duplicate traces of the same session: 00g2pdks, 1k1y4wtq, f5f35c4p, kcybky8v, pe98fd1b, sp6n43hb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2713.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,713 66.8 101.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fmwp0ala",
   "source_url": "https://omlx.ai/benchmarks/performance/fmwp0ala",
   "tg_tokens": null,
   "tps": 66.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fmwp0ala-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.3 GB at this context; duplicate traces of the same session: 00g2pdks, 1k1y4wtq, f5f35c4p, kcybky8v, pe98fd1b, sp6n43hb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2597.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,597 64.4 102.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fmwp0ala",
   "source_url": "https://omlx.ai/benchmarks/performance/fmwp0ala",
   "tg_tokens": null,
   "tps": 64.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fmwp0ala-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.7 GB at this context; duplicate traces of the same session: 00g2pdks, 1k1y4wtq, f5f35c4p, kcybky8v, pe98fd1b, sp6n43hb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2499.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,499 62.2 104.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fmwp0ala",
   "source_url": "https://omlx.ai/benchmarks/performance/fmwp0ala",
   "tg_tokens": null,
   "tps": 62.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-g977ure9-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.0 GB at this context; duplicate traces of the same session: 1awxgno0, c3uts92l, cqlx3cxg, fl59srqb, ve91u000; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1931.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,931 78.9 99.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g977ure9",
   "source_url": "https://omlx.ai/benchmarks/performance/g977ure9",
   "tg_tokens": null,
   "tps": 78.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-g977ure9-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.4 GB at this context; duplicate traces of the same session: 1awxgno0, c3uts92l, cqlx3cxg, fl59srqb, ve91u000; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2956.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,956 74.4 100.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g977ure9",
   "source_url": "https://omlx.ai/benchmarks/performance/g977ure9",
   "tg_tokens": null,
   "tps": 74.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-g977ure9-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: 1awxgno0, c3uts92l, cqlx3cxg, fl59srqb, ve91u000; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2773.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 2,773 71.0 100.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g977ure9",
   "source_url": "https://omlx.ai/benchmarks/performance/g977ure9",
   "tg_tokens": null,
   "tps": 71.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-g977ure9-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.2 GB at this context; duplicate traces of the same session: 1awxgno0, c3uts92l, cqlx3cxg, fl59srqb, ve91u000; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2728.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,728 68.5 101.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g977ure9",
   "source_url": "https://omlx.ai/benchmarks/performance/g977ure9",
   "tg_tokens": null,
   "tps": 68.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-g977ure9-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 102.3 GB at this context; duplicate traces of the same session: 1awxgno0, c3uts92l, cqlx3cxg, fl59srqb, ve91u000; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2605.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,605 64.7 102.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g977ure9",
   "source_url": "https://omlx.ai/benchmarks/performance/g977ure9",
   "tg_tokens": null,
   "tps": 64.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-g977ure9-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 104.7 GB at this context; duplicate traces of the same session: 1awxgno0, c3uts92l, cqlx3cxg, fl59srqb, ve91u000; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2551.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,551 65.8 104.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g977ure9",
   "source_url": "https://omlx.ai/benchmarks/performance/g977ure9",
   "tg_tokens": null,
   "tps": 65.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-7s1qvs8v-ctx1024",
   "model": "ThinkingCap-Qwen3.8-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 30.0 GB at this context; duplicate traces of the same session: 96608fm1, dh7n26h5, o7ktmk6h, seo3oyuq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 822.0,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "ThinkingCap-Qwen3.8-27B-mlx-mxfp8 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 822.0 18.6 30.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7s1qvs8v",
   "source_url": "https://omlx.ai/benchmarks/performance/7s1qvs8v",
   "tg_tokens": null,
   "tps": 18.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-7s1qvs8v-ctx4096",
   "model": "ThinkingCap-Qwen3.8-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 31.1 GB at this context; duplicate traces of the same session: 96608fm1, dh7n26h5, o7ktmk6h, seo3oyuq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 945.9,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "ThinkingCap-Qwen3.8-27B-mlx-mxfp8 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 945.9 18.5 31.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7s1qvs8v",
   "source_url": "https://omlx.ai/benchmarks/performance/7s1qvs8v",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-7s1qvs8v-ctx8192",
   "model": "ThinkingCap-Qwen3.8-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 31.3 GB at this context; duplicate traces of the same session: 96608fm1, dh7n26h5, o7ktmk6h, seo3oyuq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 946.6,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "ThinkingCap-Qwen3.8-27B-mlx-mxfp8 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 8k 946.6 18.2 31.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7s1qvs8v",
   "source_url": "https://omlx.ai/benchmarks/performance/7s1qvs8v",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-7s1qvs8v-ctx16384",
   "model": "ThinkingCap-Qwen3.8-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 31.8 GB at this context; duplicate traces of the same session: 96608fm1, dh7n26h5, o7ktmk6h, seo3oyuq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 924.2,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "ThinkingCap-Qwen3.8-27B-mlx-mxfp8 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 16k 924.2 17.8 31.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7s1qvs8v",
   "source_url": "https://omlx.ai/benchmarks/performance/7s1qvs8v",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-7s1qvs8v-ctx32768",
   "model": "ThinkingCap-Qwen3.8-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 32.8 GB at this context; duplicate traces of the same session: 96608fm1, dh7n26h5, o7ktmk6h, seo3oyuq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 867.5,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "ThinkingCap-Qwen3.8-27B-mlx-mxfp8 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 32k 867.5 17.2 32.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7s1qvs8v",
   "source_url": "https://omlx.ai/benchmarks/performance/7s1qvs8v",
   "tg_tokens": null,
   "tps": 17.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-7xjeakhx-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1053.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,053 88.7 23.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7xjeakhx",
   "source_url": "https://omlx.ai/benchmarks/performance/7xjeakhx",
   "tg_tokens": null,
   "tps": 88.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-7xjeakhx-ctx8192",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 989.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 989.0 103.5 23.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7xjeakhx",
   "source_url": "https://omlx.ai/benchmarks/performance/7xjeakhx",
   "tg_tokens": null,
   "tps": 103.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-7xjeakhx-ctx16384",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 876.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 876.1 106.6 24.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7xjeakhx",
   "source_url": "https://omlx.ai/benchmarks/performance/7xjeakhx",
   "tg_tokens": null,
   "tps": 106.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-7xjeakhx-ctx32768",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 691.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 691.9 71.5 25.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7xjeakhx",
   "source_url": "https://omlx.ai/benchmarks/performance/7xjeakhx",
   "tg_tokens": null,
   "tps": 71.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-tvapvi5p-ctx4096",
   "model": "Gemma 4 12B",
   "notes": "oMLX community benchmark trace; peak memory 13.2 GB at this context; duplicate traces of the same session: hwzhk31q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1224.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-12B-it-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 1,224 37.6 13.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tvapvi5p",
   "source_url": "https://omlx.ai/benchmarks/performance/tvapvi5p",
   "tg_tokens": null,
   "tps": 37.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-tvapvi5p-ctx16384",
   "model": "Gemma 4 12B",
   "notes": "oMLX community benchmark trace; peak memory 13.8 GB at this context; duplicate traces of the same session: hwzhk31q; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1610.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-12B-it-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 1,610 37.4 13.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tvapvi5p",
   "source_url": "https://omlx.ai/benchmarks/performance/tvapvi5p",
   "tg_tokens": null,
   "tps": 37.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qdmdez9w-ctx1024",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 2ywla861, 9k4rmc9i, di3jq1g3, glw6elye, zclk68qo; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2575.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,575 106.5 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qdmdez9w",
   "source_url": "https://omlx.ai/benchmarks/performance/qdmdez9w",
   "tg_tokens": null,
   "tps": 106.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qdmdez9w-ctx4096",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 2ywla861, 9k4rmc9i, di3jq1g3, glw6elye, zclk68qo; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3292.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,292 98.0 22.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qdmdez9w",
   "source_url": "https://omlx.ai/benchmarks/performance/qdmdez9w",
   "tg_tokens": null,
   "tps": 98.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qdmdez9w-ctx8192",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 2ywla861, 9k4rmc9i, di3jq1g3, glw6elye, zclk68qo; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3167.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 3,167 93.9 22.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qdmdez9w",
   "source_url": "https://omlx.ai/benchmarks/performance/qdmdez9w",
   "tg_tokens": null,
   "tps": 93.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qdmdez9w-ctx16384",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 2ywla861, 9k4rmc9i, di3jq1g3, glw6elye, zclk68qo; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2896.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,896 87.4 22.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qdmdez9w",
   "source_url": "https://omlx.ai/benchmarks/performance/qdmdez9w",
   "tg_tokens": null,
   "tps": 87.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qdmdez9w-ctx32768",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 2ywla861, 9k4rmc9i, di3jq1g3, glw6elye, zclk68qo; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2460.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 2,460 72.0 22.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qdmdez9w",
   "source_url": "https://omlx.ai/benchmarks/performance/qdmdez9w",
   "tg_tokens": null,
   "tps": 72.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qdmdez9w-ctx65536",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 2ywla861, 9k4rmc9i, di3jq1g3, glw6elye, zclk68qo; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1743.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 1,743 59.7 23.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qdmdez9w",
   "source_url": "https://omlx.ai/benchmarks/performance/qdmdez9w",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-gevt2up9-ctx1024",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: f4ndx711; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2094.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,094 85.9 21.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gevt2up9",
   "source_url": "https://omlx.ai/benchmarks/performance/gevt2up9",
   "tg_tokens": null,
   "tps": 85.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-gevt2up9-ctx4096",
   "model": "Huihui-Qwen3.6-35B-A3B-abliterated",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: f4ndx711; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2040.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Huihui-Qwen3.6-35B-A3B-abliterated-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,040 114.2 22.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gevt2up9",
   "source_url": "https://omlx.ai/benchmarks/performance/gevt2up9",
   "tg_tokens": null,
   "tps": 114.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-c97h1vef-ctx1024",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 12.6 GB at this context; duplicate traces of the same session: swh0wgty; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1070.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Mellum2-12B-A2.5B-Instruct-8bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 8bit; 1k 1,070 79.2 12.6 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c97h1vef",
   "source_url": "https://omlx.ai/benchmarks/performance/c97h1vef",
   "tg_tokens": null,
   "tps": 79.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-c97h1vef-ctx16384",
   "model": "Mellum2-12B-A2.5B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 12.9 GB at this context; duplicate traces of the same session: swh0wgty; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 831.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Mellum2-12B-A2.5B-Instruct-8bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 8bit; 16k 831.8 70.7 12.9 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c97h1vef",
   "source_url": "https://omlx.ai/benchmarks/performance/c97h1vef",
   "tg_tokens": null,
   "tps": 70.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-hxx782zz-ctx1024",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 156.7 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 344.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-V2.6-Flash-RL-MLX-4bit-MTP on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 1k 344.0 50.2 156.7 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hxx782zz",
   "source_url": "https://omlx.ai/benchmarks/performance/hxx782zz",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-hxx782zz-ctx4096",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 157.8 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 375.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-V2.6-Flash-RL-MLX-4bit-MTP on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 4k 375.5 48.5 157.8 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hxx782zz",
   "source_url": "https://omlx.ai/benchmarks/performance/hxx782zz",
   "tg_tokens": null,
   "tps": 48.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-h3i95jt4-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 2vx4ajt3, 5qaripr9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1410.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,410 148.7 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h3i95jt4",
   "source_url": "https://omlx.ai/benchmarks/performance/h3i95jt4",
   "tg_tokens": null,
   "tps": 148.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-h3i95jt4-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 2vx4ajt3, 5qaripr9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4891.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 4,891 121.4 20.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h3i95jt4",
   "source_url": "https://omlx.ai/benchmarks/performance/h3i95jt4",
   "tg_tokens": null,
   "tps": 121.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-h3i95jt4-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 2vx4ajt3, 5qaripr9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3817.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 3,817 101.0 21.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h3i95jt4",
   "source_url": "https://omlx.ai/benchmarks/performance/h3i95jt4",
   "tg_tokens": null,
   "tps": 101.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qlhw60nq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 906.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 906.1 32.5 16.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qlhw60nq",
   "source_url": "https://omlx.ai/benchmarks/performance/qlhw60nq",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-qlhw60nq-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 843.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 843.1 28.8 17.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qlhw60nq",
   "source_url": "https://omlx.ai/benchmarks/performance/qlhw60nq",
   "tg_tokens": null,
   "tps": 28.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-jvq8g5pq-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: s1lcaqnk; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 208.4,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 208.4 18.4 20.0 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jvq8g5pq",
   "source_url": "https://omlx.ai/benchmarks/performance/jvq8g5pq",
   "tg_tokens": null,
   "tps": 18.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-jvq8g5pq-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: s1lcaqnk; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 197.4,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 197.4 17.8 21.0 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jvq8g5pq",
   "source_url": "https://omlx.ai/benchmarks/performance/jvq8g5pq",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-cwrx0w1d-ctx1024",
   "model": "Qwen3.8-3.6-27B-blend",
   "notes": "blend of Qwen3.8 and Qwen3.6 27B; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: d5pkslgb; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 112.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-3.6-27B-blend-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 112.0 14.8 19.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cwrx0w1d",
   "source_url": "https://omlx.ai/benchmarks/performance/cwrx0w1d",
   "tg_tokens": null,
   "tps": 14.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-cwrx0w1d-ctx4096",
   "model": "Qwen3.8-3.6-27B-blend",
   "notes": "blend of Qwen3.8 and Qwen3.6 27B; oMLX community benchmark trace; peak memory 23.0 GB at this context; duplicate traces of the same session: d5pkslgb; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-3.6-27B-blend-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 104.8 13.9 23.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cwrx0w1d",
   "source_url": "https://omlx.ai/benchmarks/performance/cwrx0w1d",
   "tg_tokens": null,
   "tps": 13.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-2g3rbn16-ctx1024",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: die0xc59; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 204.0 20.9 5.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2g3rbn16",
   "source_url": "https://omlx.ai/benchmarks/performance/2g3rbn16",
   "tg_tokens": null,
   "tps": 20.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-2g3rbn16-ctx4096",
   "model": "Qwen3.8-9B-heretic-uncensored",
   "notes": "MTPLX; oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: die0xc59; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 211.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-heretic-uncensored-4bit-MTPLX on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 211.0 20.5 6.2 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2g3rbn16",
   "source_url": "https://omlx.ai/benchmarks/performance/2g3rbn16",
   "tg_tokens": null,
   "tps": 20.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0wq5vfic-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 52.4 GB at this context; duplicate traces of the same session: 9xntgzof; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 309.5,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 2bit; 1k 309.5 53.7 52.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0wq5vfic",
   "source_url": "https://omlx.ai/benchmarks/performance/0wq5vfic",
   "tg_tokens": null,
   "tps": 53.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0wq5vfic-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 52.8 GB at this context; duplicate traces of the same session: 9xntgzof; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 739.3,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Qwen3.8-Flash-Next-oQ2.7e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 2bit; 4k 739.3 53.7 52.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0wq5vfic",
   "source_url": "https://omlx.ai/benchmarks/performance/0wq5vfic",
   "tg_tokens": null,
   "tps": 53.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-xpc23kf0-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 77.5 GB at this context; duplicate traces of the same session: 0kv648if, 9rxtcv7w, e1vek9t3, wdqepbj4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 619.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 619.5 66.8 77.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xpc23kf0",
   "source_url": "https://omlx.ai/benchmarks/performance/xpc23kf0",
   "tg_tokens": null,
   "tps": 66.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-xpc23kf0-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.1 GB at this context; duplicate traces of the same session: 0kv648if, 9rxtcv7w, e1vek9t3, wdqepbj4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 765.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 765.6 63.2 78.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xpc23kf0",
   "source_url": "https://omlx.ai/benchmarks/performance/xpc23kf0",
   "tg_tokens": null,
   "tps": 63.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-xpc23kf0-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.8 GB at this context; duplicate traces of the same session: 0kv648if, 9rxtcv7w, e1vek9t3, wdqepbj4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 820.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 820.1 71.6 78.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xpc23kf0",
   "source_url": "https://omlx.ai/benchmarks/performance/xpc23kf0",
   "tg_tokens": null,
   "tps": 71.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-xpc23kf0-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.5 GB at this context; duplicate traces of the same session: 0kv648if, 9rxtcv7w, e1vek9t3, wdqepbj4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 807.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k 807.0 63.6 80.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xpc23kf0",
   "source_url": "https://omlx.ai/benchmarks/performance/xpc23kf0",
   "tg_tokens": null,
   "tps": 63.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-xpc23kf0-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 83.0 GB at this context; duplicate traces of the same session: 0kv648if, 9rxtcv7w, e1vek9t3, wdqepbj4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 795.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 128k 795.1 52.1 83.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xpc23kf0",
   "source_url": "https://omlx.ai/benchmarks/performance/xpc23kf0",
   "tg_tokens": null,
   "tps": 52.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-mde8lkst-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 77.5 GB at this context; duplicate traces of the same session: 25pqjpdj, hfni9vc3, u8uluw1i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 618.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 618.6 71.3 77.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mde8lkst",
   "source_url": "https://omlx.ai/benchmarks/performance/mde8lkst",
   "tg_tokens": null,
   "tps": 71.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-mde8lkst-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.5 GB at this context; duplicate traces of the same session: 25pqjpdj, hfni9vc3, u8uluw1i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 825.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 825.8 71.7 78.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mde8lkst",
   "source_url": "https://omlx.ai/benchmarks/performance/mde8lkst",
   "tg_tokens": null,
   "tps": 71.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-mde8lkst-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.5 GB at this context; duplicate traces of the same session: 25pqjpdj, hfni9vc3, u8uluw1i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 807.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k 807.0 65.0 80.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mde8lkst",
   "source_url": "https://omlx.ai/benchmarks/performance/mde8lkst",
   "tg_tokens": null,
   "tps": 65.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-mde8lkst-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 83.0 GB at this context; duplicate traces of the same session: 25pqjpdj, hfni9vc3, u8uluw1i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 795.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 128k 795.1 55.1 83.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mde8lkst",
   "source_url": "https://omlx.ai/benchmarks/performance/mde8lkst",
   "tg_tokens": null,
   "tps": 55.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-zi76n5dw-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: b64y4230, et0wg5yq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1053.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,053 88.7 23.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zi76n5dw",
   "source_url": "https://omlx.ai/benchmarks/performance/zi76n5dw",
   "tg_tokens": null,
   "tps": 88.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-zi76n5dw-ctx8192",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: b64y4230, et0wg5yq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 989.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 989.0 103.5 23.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zi76n5dw",
   "source_url": "https://omlx.ai/benchmarks/performance/zi76n5dw",
   "tg_tokens": null,
   "tps": 103.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-zi76n5dw-ctx16384",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: b64y4230, et0wg5yq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 876.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 876.1 106.6 24.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zi76n5dw",
   "source_url": "https://omlx.ai/benchmarks/performance/zi76n5dw",
   "tg_tokens": null,
   "tps": 106.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-zi76n5dw-ctx32768",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: b64y4230, et0wg5yq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 691.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 691.9 71.5 25.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zi76n5dw",
   "source_url": "https://omlx.ai/benchmarks/performance/zi76n5dw",
   "tg_tokens": null,
   "tps": 71.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-df6789us-ctx1024",
   "model": "gemma-4-12b-it",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.6 GB at this context; duplicate traces of the same session: nfexxr6i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1721.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-12B-it-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 1,721 72.9 13.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/df6789us",
   "source_url": "https://omlx.ai/benchmarks/performance/df6789us",
   "tg_tokens": null,
   "tps": 72.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-df6789us-ctx4096",
   "model": "gemma-4-12b-it",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: nfexxr6i; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1787.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-12B-it-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 1,787 33.1 13.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/df6789us",
   "source_url": "https://omlx.ai/benchmarks/performance/df6789us",
   "tg_tokens": null,
   "tps": 33.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-v9apa7po-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: ycry85u0; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1079.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "gemma-4-26B-A4B-it-oQ6e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 6bit; 1k 1,079 92.2 22.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v9apa7po",
   "source_url": "https://omlx.ai/benchmarks/performance/v9apa7po",
   "tg_tokens": null,
   "tps": 92.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-v9apa7po-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: ycry85u0; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1183.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "gemma-4-26B-A4B-it-oQ6e-mtp on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 6bit; 4k 1,183 103.3 22.4 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v9apa7po",
   "source_url": "https://omlx.ai/benchmarks/performance/v9apa7po",
   "tg_tokens": null,
   "tps": 103.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-716wsski-ctx1024",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.5 GB at this context; duplicate traces of the same session: 8zuy9u1u; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 105.5 19.9 20.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/716wsski",
   "source_url": "https://omlx.ai/benchmarks/performance/716wsski",
   "tg_tokens": null,
   "tps": 19.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-716wsski-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 8zuy9u1u; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 104.1 20.2 21.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/716wsski",
   "source_url": "https://omlx.ai/benchmarks/performance/716wsski",
   "tg_tokens": null,
   "tps": 20.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-tjvg15lr-ctx1024",
   "model": "DeepSeek-R1-Distill-Qwen-14B",
   "notes": "oMLX community benchmark trace; peak memory 8.5 GB at this context; duplicate traces of the same session: s4ymnqu0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 197.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-R1-Distill-Qwen-14B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 197.5 27.8 8.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tjvg15lr",
   "source_url": "https://omlx.ai/benchmarks/performance/tjvg15lr",
   "tg_tokens": null,
   "tps": 27.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-tjvg15lr-ctx4096",
   "model": "DeepSeek-R1-Distill-Qwen-14B",
   "notes": "oMLX community benchmark trace; peak memory 9.1 GB at this context; duplicate traces of the same session: s4ymnqu0; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 181.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-R1-Distill-Qwen-14B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 181.0 25.4 9.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tjvg15lr",
   "source_url": "https://omlx.ai/benchmarks/performance/tjvg15lr",
   "tg_tokens": null,
   "tps": 25.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-kpxgbztt-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 205.7 GB at this context; duplicate traces of the same session: 6i7xnugq, pf9ln7bx, vtaovxum; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 213.7,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 1k 213.7 27.6 205.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kpxgbztt",
   "source_url": "https://omlx.ai/benchmarks/performance/kpxgbztt",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-kpxgbztt-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 206.2 GB at this context; duplicate traces of the same session: 6i7xnugq, pf9ln7bx, vtaovxum; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 296.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 4k 296.0 30.1 206.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kpxgbztt",
   "source_url": "https://omlx.ai/benchmarks/performance/kpxgbztt",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-kpxgbztt-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 206.6 GB at this context; duplicate traces of the same session: 6i7xnugq, pf9ln7bx, vtaovxum; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 312.9,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 8k 312.9 29.9 206.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kpxgbztt",
   "source_url": "https://omlx.ai/benchmarks/performance/kpxgbztt",
   "tg_tokens": null,
   "tps": 29.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-kpxgbztt-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 207.2 GB at this context; duplicate traces of the same session: 6i7xnugq, pf9ln7bx, vtaovxum; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 316.4,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 16k 316.4 40.0 207.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kpxgbztt",
   "source_url": "https://omlx.ai/benchmarks/performance/kpxgbztt",
   "tg_tokens": null,
   "tps": 40.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-mo460lgp-ctx1024",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "oMLX community benchmark trace; peak memory 208.2 GB at this context; duplicate traces of the same session: 769y8kyn, n9bk7sta; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 929.6,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "GLM-5.3-Flash-Uncensored-oQ5e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 5bit; 1k 929.6 61.5 208.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mo460lgp",
   "source_url": "https://omlx.ai/benchmarks/performance/mo460lgp",
   "tg_tokens": null,
   "tps": 61.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-mo460lgp-ctx16384",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "oMLX community benchmark trace; peak memory 210.1 GB at this context; duplicate traces of the same session: 769y8kyn, n9bk7sta; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1712.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "GLM-5.3-Flash-Uncensored-oQ5e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 5bit; 16k 1,712 101.8 210.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mo460lgp",
   "source_url": "https://omlx.ai/benchmarks/performance/mo460lgp",
   "tg_tokens": null,
   "tps": 101.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-mo460lgp-ctx32768",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "oMLX community benchmark trace; peak memory 210.3 GB at this context; duplicate traces of the same session: 769y8kyn, n9bk7sta; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1708.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "GLM-5.3-Flash-Uncensored-oQ5e on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 5bit; 32k 1,708 91.7 210.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mo460lgp",
   "source_url": "https://omlx.ai/benchmarks/performance/mo460lgp",
   "tg_tokens": null,
   "tps": 91.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-58trei4j-ctx1024",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "g64; oMLX community benchmark trace; peak memory 9.8 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1037.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-mlx-6bit-g64 on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 6bit; 1k 1,037 91.9 9.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/58trei4j",
   "source_url": "https://omlx.ai/benchmarks/performance/58trei4j",
   "tg_tokens": null,
   "tps": 91.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-58trei4j-ctx16384",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "g64; oMLX community benchmark trace; peak memory 10.1 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 837.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-mlx-6bit-g64 on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 6bit; 16k 837.8 80.9 10.1 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/58trei4j",
   "source_url": "https://omlx.ai/benchmarks/performance/58trei4j",
   "tg_tokens": null,
   "tps": 80.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-ut0yfutj-ctx1024",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 156.7 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 344.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-V2.6-Flash-RL-MLX-4bit-MTP on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 1k 344.0 50.2 156.7 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ut0yfutj",
   "source_url": "https://omlx.ai/benchmarks/performance/ut0yfutj",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-ut0yfutj-ctx4096",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 157.8 GB at this context; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 375.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-V2.6-Flash-RL-MLX-4bit-MTP on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 4k 375.5 48.5 157.8 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ut0yfutj",
   "source_url": "https://omlx.ai/benchmarks/performance/ut0yfutj",
   "tg_tokens": null,
   "tps": 48.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-2r31ph0a-ctx1024",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: oltt3dep; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 205.0 22.5 15.9 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2r31ph0a",
   "source_url": "https://omlx.ai/benchmarks/performance/2r31ph0a",
   "tg_tokens": null,
   "tps": 22.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-2r31ph0a-ctx4096",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: oltt3dep; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 184.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 184.8 21.5 16.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2r31ph0a",
   "source_url": "https://omlx.ai/benchmarks/performance/2r31ph0a",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-o7leeauk-ctx1024",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: 60iyghf6; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 204.6 22.4 15.9 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o7leeauk",
   "source_url": "https://omlx.ai/benchmarks/performance/o7leeauk",
   "tg_tokens": null,
   "tps": 22.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-o7leeauk-ctx4096",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: 60iyghf6; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 180.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 180.2 21.1 16.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o7leeauk",
   "source_url": "https://omlx.ai/benchmarks/performance/o7leeauk",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-a0dli1gs-ctx1024",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 15.9 GB at this context; duplicate traces of the same session: 7rcwof1b; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 207.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 207.6 22.3 15.9 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a0dli1gs",
   "source_url": "https://omlx.ai/benchmarks/performance/a0dli1gs",
   "tg_tokens": null,
   "tps": 22.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-a0dli1gs-ctx4096",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: 7rcwof1b; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 185.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 185.0 21.4 16.1 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a0dli1gs",
   "source_url": "https://omlx.ai/benchmarks/performance/a0dli1gs",
   "tg_tokens": null,
   "tps": 21.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ygqzyc71-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: w0cco678, zzzazc69; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 730.2,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-6bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 1k 730.2 67.4 27.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ygqzyc71",
   "source_url": "https://omlx.ai/benchmarks/performance/ygqzyc71",
   "tg_tokens": null,
   "tps": 67.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ygqzyc71-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: w0cco678, zzzazc69; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 811.4,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-6bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 4k 811.4 65.2 28.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ygqzyc71",
   "source_url": "https://omlx.ai/benchmarks/performance/ygqzyc71",
   "tg_tokens": null,
   "tps": 65.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ygqzyc71-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 29.0 GB at this context; duplicate traces of the same session: w0cco678, zzzazc69; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 577.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-6bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 32k 577.8 49.7 29.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ygqzyc71",
   "source_url": "https://omlx.ai/benchmarks/performance/ygqzyc71",
   "tg_tokens": null,
   "tps": 49.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-vkppvq0m-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: vnbm2xth, xbf08sso; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 945.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 945.4 93.2 19.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vkppvq0m",
   "source_url": "https://omlx.ai/benchmarks/performance/vkppvq0m",
   "tg_tokens": null,
   "tps": 93.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-vkppvq0m-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: vnbm2xth, xbf08sso; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1034.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 1,034 89.7 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vkppvq0m",
   "source_url": "https://omlx.ai/benchmarks/performance/vkppvq0m",
   "tg_tokens": null,
   "tps": 89.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-vkppvq0m-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: vnbm2xth, xbf08sso; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 351.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 128k 351.1 37.0 23.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vkppvq0m",
   "source_url": "https://omlx.ai/benchmarks/performance/vkppvq0m",
   "tg_tokens": null,
   "tps": 37.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-jdafx7pk-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: jy9mtron; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 952.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 952.9 94.3 19.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jdafx7pk",
   "source_url": "https://omlx.ai/benchmarks/performance/jdafx7pk",
   "tg_tokens": null,
   "tps": 94.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-jdafx7pk-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: jy9mtron; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1038.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 1,038 90.7 20.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jdafx7pk",
   "source_url": "https://omlx.ai/benchmarks/performance/jdafx7pk",
   "tg_tokens": null,
   "tps": 90.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-ymfvpfho-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 17.2 GB; TTFT 23670 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 173.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 4k tokens 173.0 PP tok/s 16.8 TG tok/s 23670 TTFT (ms) 17.2 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ymfvpfho",
   "source_url": "https://omlx.ai/benchmarks/performance/ymfvpfho",
   "tg_tokens": null,
   "tps": 16.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rdzz0mdl-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: ax68yvgc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 909.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 909.5 31.6 16.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rdzz0mdl",
   "source_url": "https://omlx.ai/benchmarks/performance/rdzz0mdl",
   "tg_tokens": null,
   "tps": 31.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rdzz0mdl-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: ax68yvgc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 672.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 672.8 21.3 18.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rdzz0mdl",
   "source_url": "https://omlx.ai/benchmarks/performance/rdzz0mdl",
   "tg_tokens": null,
   "tps": 21.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-1f87qrwv-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: gqzgu68i; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 208.5,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 208.5 17.6 20.0 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1f87qrwv",
   "source_url": "https://omlx.ai/benchmarks/performance/1f87qrwv",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-1f87qrwv-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: gqzgu68i; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 196.4,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M4 Max (32c); M4 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 196.4 17.7 21.0 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1f87qrwv",
   "source_url": "https://omlx.ai/benchmarks/performance/1f87qrwv",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3c4nrt7f-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.6 GB; TTFT 1019 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1005.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,005 PP tok/s 53.6 TG tok/s 1019 TTFT (ms) 70.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3c4nrt7f",
   "source_url": "https://omlx.ai/benchmarks/performance/3c4nrt7f",
   "tg_tokens": null,
   "tps": 53.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pcclxiwn-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 69.2 GB; TTFT 894 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1145.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,145 PP tok/s 70.7 TG tok/s 894 TTFT (ms) 69.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pcclxiwn",
   "source_url": "https://omlx.ai/benchmarks/performance/pcclxiwn",
   "tg_tokens": null,
   "tps": 70.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2flmglaz-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: lsiz9l4p; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1197.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,197 69.0 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2flmglaz",
   "source_url": "https://omlx.ai/benchmarks/performance/2flmglaz",
   "tg_tokens": null,
   "tps": 69.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-2flmglaz-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.3 GB at this context; duplicate traces of the same session: lsiz9l4p; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2050.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,050 80.2 73.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2flmglaz",
   "source_url": "https://omlx.ai/benchmarks/performance/2flmglaz",
   "tg_tokens": null,
   "tps": 80.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hl4u6xl0-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.2 GB at this context; duplicate traces of the same session: 6u78mnew; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1137.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,137 71.4 69.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hl4u6xl0",
   "source_url": "https://omlx.ai/benchmarks/performance/hl4u6xl0",
   "tg_tokens": null,
   "tps": 71.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hl4u6xl0-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.4 GB at this context; duplicate traces of the same session: 6u78mnew; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2143.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,143 64.7 71.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hl4u6xl0",
   "source_url": "https://omlx.ai/benchmarks/performance/hl4u6xl0",
   "tg_tokens": null,
   "tps": 64.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jp0kdyi9-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 72.9 GB; TTFT 9522 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1721.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k tokens 1,721 PP tok/s 80.1 TG tok/s 9522 TTFT (ms) 72.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jp0kdyi9",
   "source_url": "https://omlx.ai/benchmarks/performance/jp0kdyi9",
   "tg_tokens": null,
   "tps": 80.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-itztjp86-ctx65536",
   "model": "ThinkingCap-Qwen3.8-27B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 64k context; peak memory 35.0 GB; TTFT 87125 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 752.2,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "ThinkingCap-Qwen3.8-27B-mlx-mxfp8 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 64k tokens 752.2 PP tok/s 15.9 TG tok/s 87125 TTFT (ms) 35.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/itztjp86",
   "source_url": "https://omlx.ai/benchmarks/performance/itztjp86",
   "tg_tokens": null,
   "tps": 15.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-o0p0odp2-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 616.7,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 616.7 69.7 11.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o0p0odp2",
   "source_url": "https://omlx.ai/benchmarks/performance/o0p0odp2",
   "tg_tokens": null,
   "tps": 69.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-o0p0odp2-ctx16384",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 12.1 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 594.5,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 16k 594.5 62.5 12.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o0p0odp2",
   "source_url": "https://omlx.ai/benchmarks/performance/o0p0odp2",
   "tg_tokens": null,
   "tps": 62.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-qiy3q9g1-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 205.9 GB at this context; duplicate traces of the same session: rmewszv0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1037.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 unknown; 16k 1,037 48.8 205.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qiy3q9g1",
   "source_url": "https://omlx.ai/benchmarks/performance/qiy3q9g1",
   "tg_tokens": null,
   "tps": 48.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-qiy3q9g1-ctx32768",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 206.0 GB at this context; duplicate traces of the same session: rmewszv0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1349.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 unknown; 32k 1,349 46.0 206.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qiy3q9g1",
   "source_url": "https://omlx.ai/benchmarks/performance/qiy3q9g1",
   "tg_tokens": null,
   "tps": 46.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-rj3pabjb-ctx65536",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 206.0 GB at this context; duplicate traces of the same session: zqlc8x20; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1024.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 unknown; 64k 1,024 49.9 206.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rj3pabjb",
   "source_url": "https://omlx.ai/benchmarks/performance/rj3pabjb",
   "tg_tokens": null,
   "tps": 49.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-rj3pabjb-ctx131072",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 206.2 GB at this context; duplicate traces of the same session: zqlc8x20; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 723.2,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 unknown; 128k 723.2 48.7 206.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rj3pabjb",
   "source_url": "https://omlx.ai/benchmarks/performance/rj3pabjb",
   "tg_tokens": null,
   "tps": 48.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vvexhwyg-ctx1024",
   "model": "GLM-5.3",
   "notes": "oMLX community benchmark trace; peak memory 390.9 GB at this context; duplicate traces of the same session: 3s9spl74, 6ip8k7br, 6p05yhm6, 7hu792da, adutch7l, crk97iwb, mgze4frd; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 219.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 219.8 19.1 390.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vvexhwyg",
   "source_url": "https://omlx.ai/benchmarks/performance/vvexhwyg",
   "tg_tokens": null,
   "tps": 19.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vvexhwyg-ctx4096",
   "model": "GLM-5.3",
   "notes": "oMLX community benchmark trace; peak memory 392.6 GB at this context; duplicate traces of the same session: 3s9spl74, 6ip8k7br, 6p05yhm6, 7hu792da, adutch7l, crk97iwb, mgze4frd; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 226.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 226.5 18.0 392.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vvexhwyg",
   "source_url": "https://omlx.ai/benchmarks/performance/vvexhwyg",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vvexhwyg-ctx8192",
   "model": "GLM-5.3",
   "notes": "oMLX community benchmark trace; peak memory 394.0 GB at this context; duplicate traces of the same session: 3s9spl74, 6ip8k7br, 6p05yhm6, 7hu792da, adutch7l, crk97iwb, mgze4frd; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 202.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 8k 202.6 17.8 394.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vvexhwyg",
   "source_url": "https://omlx.ai/benchmarks/performance/vvexhwyg",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vvexhwyg-ctx16384",
   "model": "GLM-5.3",
   "notes": "oMLX community benchmark trace; peak memory 394.7 GB at this context; duplicate traces of the same session: 3s9spl74, 6ip8k7br, 6p05yhm6, 7hu792da, adutch7l, crk97iwb, mgze4frd; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 187.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 16k 187.5 17.6 394.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vvexhwyg",
   "source_url": "https://omlx.ai/benchmarks/performance/vvexhwyg",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vvexhwyg-ctx32768",
   "model": "GLM-5.3",
   "notes": "oMLX community benchmark trace; peak memory 396.2 GB at this context; duplicate traces of the same session: 3s9spl74, 6ip8k7br, 6p05yhm6, 7hu792da, adutch7l, crk97iwb, mgze4frd; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 181.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 32k 181.1 17.2 396.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vvexhwyg",
   "source_url": "https://omlx.ai/benchmarks/performance/vvexhwyg",
   "tg_tokens": null,
   "tps": 17.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vvexhwyg-ctx65536",
   "model": "GLM-5.3",
   "notes": "oMLX community benchmark trace; peak memory 399.1 GB at this context; duplicate traces of the same session: 3s9spl74, 6ip8k7br, 6p05yhm6, 7hu792da, adutch7l, crk97iwb, mgze4frd; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 174.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 174.7 16.7 399.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vvexhwyg",
   "source_url": "https://omlx.ai/benchmarks/performance/vvexhwyg",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vvexhwyg-ctx131072",
   "model": "GLM-5.3",
   "notes": "oMLX community benchmark trace; peak memory 407.6 GB at this context; duplicate traces of the same session: 3s9spl74, 6ip8k7br, 6p05yhm6, 7hu792da, adutch7l, crk97iwb, mgze4frd; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 166.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 166.1 15.0 407.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vvexhwyg",
   "source_url": "https://omlx.ai/benchmarks/performance/vvexhwyg",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-vvexhwyg-ctx199680",
   "model": "GLM-5.3",
   "notes": "oMLX community benchmark trace; peak memory 413.5 GB at this context; duplicate traces of the same session: 3s9spl74, 6ip8k7br, 6p05yhm6, 7hu792da, adutch7l, crk97iwb, mgze4frd; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 158.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 195k 158.7 14.4 413.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vvexhwyg",
   "source_url": "https://omlx.ai/benchmarks/performance/vvexhwyg",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-9bnz788b-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.7 GB at this context; duplicate traces of the same session: 802yqvtl, bsmgupjc, fnbqcsxo, gjuykim1, x089975e, yjggzij7; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1201.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 1,201 58.8 177.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9bnz788b",
   "source_url": "https://omlx.ai/benchmarks/performance/9bnz788b",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-9bnz788b-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.7 GB at this context; duplicate traces of the same session: 802yqvtl, bsmgupjc, fnbqcsxo, gjuykim1, x089975e, yjggzij7; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1861.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 1,861 58.5 179.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9bnz788b",
   "source_url": "https://omlx.ai/benchmarks/performance/9bnz788b",
   "tg_tokens": null,
   "tps": 58.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-9bnz788b-ctx8192",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.2 GB at this context; duplicate traces of the same session: 802yqvtl, bsmgupjc, fnbqcsxo, gjuykim1, x089975e, yjggzij7; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1889.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 1,889 71.0 180.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9bnz788b",
   "source_url": "https://omlx.ai/benchmarks/performance/9bnz788b",
   "tg_tokens": null,
   "tps": 71.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-9bnz788b-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.3 GB at this context; duplicate traces of the same session: 802yqvtl, bsmgupjc, fnbqcsxo, gjuykim1, x089975e, yjggzij7; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1892.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,892 81.8 180.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9bnz788b",
   "source_url": "https://omlx.ai/benchmarks/performance/9bnz788b",
   "tg_tokens": null,
   "tps": 81.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-9bnz788b-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.7 GB at this context; duplicate traces of the same session: 802yqvtl, bsmgupjc, fnbqcsxo, gjuykim1, x089975e, yjggzij7; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1882.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,882 80.4 180.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9bnz788b",
   "source_url": "https://omlx.ai/benchmarks/performance/9bnz788b",
   "tg_tokens": null,
   "tps": 80.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-9bnz788b-ctx65536",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 181.5 GB at this context; duplicate traces of the same session: 802yqvtl, bsmgupjc, fnbqcsxo, gjuykim1, x089975e, yjggzij7; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1856.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,856 80.0 181.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9bnz788b",
   "source_url": "https://omlx.ai/benchmarks/performance/9bnz788b",
   "tg_tokens": null,
   "tps": 80.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-9bnz788b-ctx131072",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 183.4 GB at this context; duplicate traces of the same session: 802yqvtl, bsmgupjc, fnbqcsxo, gjuykim1, x089975e, yjggzij7; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1824.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 1,824 77.7 183.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9bnz788b",
   "source_url": "https://omlx.ai/benchmarks/performance/9bnz788b",
   "tg_tokens": null,
   "tps": 77.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-f0klm1o1-ctx1024",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "g64; oMLX community benchmark trace; peak memory 9.8 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1037.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-mlx-6bit-g64 on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 6bit; 1k 1,037 91.9 9.8 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f0klm1o1",
   "source_url": "https://omlx.ai/benchmarks/performance/f0klm1o1",
   "tg_tokens": null,
   "tps": 91.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-f0klm1o1-ctx16384",
   "model": "Mellum2.1-12B-A2.5B-Thinking",
   "notes": "g64; oMLX community benchmark trace; peak memory 10.1 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 837.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Mellum2.1-12B-A2.5B-Thinking-mlx-6bit-g64 on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 6bit; 16k 837.8 80.9 10.1 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f0klm1o1",
   "source_url": "https://omlx.ai/benchmarks/performance/f0klm1o1",
   "tg_tokens": null,
   "tps": 80.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-dgf42sw3-ctx1024",
   "model": "Phi-4-reasoning-plus",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 8.6 GB at this context; duplicate traces of the same session: knxj2q2k; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 193.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Phi-4-reasoning-plus-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 193.8 27.5 8.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dgf42sw3",
   "source_url": "https://omlx.ai/benchmarks/performance/dgf42sw3",
   "tg_tokens": null,
   "tps": 27.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-dgf42sw3-ctx4096",
   "model": "Phi-4-reasoning-plus",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 9.9 GB at this context; duplicate traces of the same session: knxj2q2k; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 186.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Phi-4-reasoning-plus-MLX-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 186.7 25.8 9.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dgf42sw3",
   "source_url": "https://omlx.ai/benchmarks/performance/dgf42sw3",
   "tg_tokens": null,
   "tps": 25.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-n2nukihw-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: pqgvwllf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 941.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 941.8 94.0 19.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n2nukihw",
   "source_url": "https://omlx.ai/benchmarks/performance/n2nukihw",
   "tg_tokens": null,
   "tps": 94.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-n2nukihw-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: pqgvwllf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1035.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 1,035 90.5 20.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n2nukihw",
   "source_url": "https://omlx.ai/benchmarks/performance/n2nukihw",
   "tg_tokens": null,
   "tps": 90.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-pjn84iqd-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: lnz2fu12, nyehnwm5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 6945.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 6,945 179.8 21.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pjn84iqd",
   "source_url": "https://omlx.ai/benchmarks/performance/pjn84iqd",
   "tg_tokens": null,
   "tps": 179.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-pjn84iqd-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: lnz2fu12, nyehnwm5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5236.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 5,236 109.4 22.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pjn84iqd",
   "source_url": "https://omlx.ai/benchmarks/performance/pjn84iqd",
   "tg_tokens": null,
   "tps": 109.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-pjn84iqd-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: lnz2fu12, nyehnwm5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 7403.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 7,403 313.7 22.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pjn84iqd",
   "source_url": "https://omlx.ai/benchmarks/performance/pjn84iqd",
   "tg_tokens": null,
   "tps": 313.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-ujed5039-ctx1024",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 31.7 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 190.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 unknown; 1k 190.9 25.1 31.7 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ujed5039",
   "source_url": "https://omlx.ai/benchmarks/performance/ujed5039",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-ujed5039-ctx4096",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 35.4 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 197.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 unknown; 4k 197.2 26.6 35.4 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ujed5039",
   "source_url": "https://omlx.ai/benchmarks/performance/ujed5039",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-ujed5039-ctx8192",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 36.2 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 192.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 unknown; 8k 192.2 21.5 36.2 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ujed5039",
   "source_url": "https://omlx.ai/benchmarks/performance/ujed5039",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-ujed5039-ctx16384",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 37.7 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 182.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 unknown; 16k 182.0 35.0 37.7 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ujed5039",
   "source_url": "https://omlx.ai/benchmarks/performance/ujed5039",
   "tg_tokens": null,
   "tps": 35.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-b88gkuvp-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 16.6 GB; TTFT 641 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1597.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k tokens 1,597 PP tok/s 52.1 TG tok/s 641 TTFT (ms) 16.6 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b88gkuvp",
   "source_url": "https://omlx.ai/benchmarks/performance/b88gkuvp",
   "tg_tokens": null,
   "tps": 52.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-quxiragy-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: aabx1e8v, cgbwiryo; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1450.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 1,450 87.1 17.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/quxiragy",
   "source_url": "https://omlx.ai/benchmarks/performance/quxiragy",
   "tg_tokens": null,
   "tps": 87.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-quxiragy-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: aabx1e8v, cgbwiryo; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1856.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 1,856 116.3 19.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/quxiragy",
   "source_url": "https://omlx.ai/benchmarks/performance/quxiragy",
   "tg_tokens": null,
   "tps": 116.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-quxiragy-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: aabx1e8v, cgbwiryo; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1741.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 1,741 129.4 20.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/quxiragy",
   "source_url": "https://omlx.ai/benchmarks/performance/quxiragy",
   "tg_tokens": null,
   "tps": 129.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-etk7ywlj-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 72.9 GB; TTFT 9538 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1718.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k tokens 1,718 PP tok/s 56.7 TG tok/s 9538 TTFT (ms) 72.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/etk7ywlj",
   "source_url": "https://omlx.ai/benchmarks/performance/etk7ywlj",
   "tg_tokens": null,
   "tps": 56.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5ah65gq9-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 72.9 GB; TTFT 9869 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1660.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k tokens 1,660 PP tok/s 56.8 TG tok/s 9869 TTFT (ms) 72.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ah65gq9",
   "source_url": "https://omlx.ai/benchmarks/performance/5ah65gq9",
   "tg_tokens": null,
   "tps": 56.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-4bhp5q8d-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.0 GB at this context; duplicate traces of the same session: 8bsxyo7i, yysmj0dv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3231.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 3,231 124.5 99.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4bhp5q8d",
   "source_url": "https://omlx.ai/benchmarks/performance/4bhp5q8d",
   "tg_tokens": null,
   "tps": 124.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-4bhp5q8d-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.3 GB at this context; duplicate traces of the same session: 8bsxyo7i, yysmj0dv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4241.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 4,241 113.9 99.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4bhp5q8d",
   "source_url": "https://omlx.ai/benchmarks/performance/4bhp5q8d",
   "tg_tokens": null,
   "tps": 113.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-4bhp5q8d-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.2 GB at this context; duplicate traces of the same session: 8bsxyo7i, yysmj0dv; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4201.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 4,201 108.8 100.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4bhp5q8d",
   "source_url": "https://omlx.ai/benchmarks/performance/4bhp5q8d",
   "tg_tokens": null,
   "tps": 108.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-fli4peua-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.4 GB at this context; duplicate traces of the same session: s0u7jia8, z6rpykn3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 403.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 403.3 153.8 100.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fli4peua",
   "source_url": "https://omlx.ai/benchmarks/performance/fli4peua",
   "tg_tokens": null,
   "tps": 153.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-fli4peua-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.8 GB at this context; duplicate traces of the same session: s0u7jia8, z6rpykn3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1826.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 1,826 113.1 100.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fli4peua",
   "source_url": "https://omlx.ai/benchmarks/performance/fli4peua",
   "tg_tokens": null,
   "tps": 113.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-fli4peua-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.8 GB at this context; duplicate traces of the same session: s0u7jia8, z6rpykn3; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4014.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 4,014 24.1 101.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fli4peua",
   "source_url": "https://omlx.ai/benchmarks/performance/fli4peua",
   "tg_tokens": null,
   "tps": 24.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-o96guxpe-ctx1024",
   "model": "Ternary-Bonsai-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 9.1 GB at this context; duplicate traces of the same session: 1cbmq655; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 479.3,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-27B-mlx-2bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 2bit; 1k 479.3 34.0 9.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o96guxpe",
   "source_url": "https://omlx.ai/benchmarks/performance/o96guxpe",
   "tg_tokens": null,
   "tps": 34.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-o96guxpe-ctx4096",
   "model": "Ternary-Bonsai-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 10.1 GB at this context; duplicate traces of the same session: 1cbmq655; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 528.9,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-27B-mlx-2bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 2bit; 4k 528.9 33.2 10.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o96guxpe",
   "source_url": "https://omlx.ai/benchmarks/performance/o96guxpe",
   "tg_tokens": null,
   "tps": 33.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-mgby0pm9-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 616.7,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 4k 616.7 69.7 11.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mgby0pm9",
   "source_url": "https://omlx.ai/benchmarks/performance/mgby0pm9",
   "tg_tokens": null,
   "tps": 69.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-mgby0pm9-ctx16384",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 12.1 GB at this context; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 594.5,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 4bit; 16k 594.5 62.5 12.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mgby0pm9",
   "source_url": "https://omlx.ai/benchmarks/performance/mgby0pm9",
   "tg_tokens": null,
   "tps": 62.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-mafrx6w3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: h61hrsxj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "p4ik--Qwen3.8-27B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 130.0 15.4 16.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mafrx6w3",
   "source_url": "https://omlx.ai/benchmarks/performance/mafrx6w3",
   "tg_tokens": null,
   "tps": 15.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-mafrx6w3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: h61hrsxj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 132.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "p4ik--Qwen3.8-27B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 132.2 15.1 17.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mafrx6w3",
   "source_url": "https://omlx.ai/benchmarks/performance/mafrx6w3",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-d8ootnvr-ctx1024",
   "model": "BigBang-v1",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 30.0 GB at this context; duplicate traces of the same session: 5dir1euf; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 630.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "BigBang-v1-mtp-oQ6e on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 1k 630.3 49.7 30.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d8ootnvr",
   "source_url": "https://omlx.ai/benchmarks/performance/d8ootnvr",
   "tg_tokens": null,
   "tps": 49.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-d8ootnvr-ctx4096",
   "model": "BigBang-v1",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: 5dir1euf; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 739.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "BigBang-v1-mtp-oQ6e on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 4k 739.9 49.6 32.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d8ootnvr",
   "source_url": "https://omlx.ai/benchmarks/performance/d8ootnvr",
   "tg_tokens": null,
   "tps": 49.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-xy3t3y1i-ctx4096",
   "model": "DeepSeek-R1-Distill-Qwen-14B",
   "notes": "oMLX community benchmark trace; peak memory 9.6 GB at this context; duplicate traces of the same session: q7jh3msn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 929.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-R1-Distill-Qwen-14B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 929.6 30.5 9.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xy3t3y1i",
   "source_url": "https://omlx.ai/benchmarks/performance/xy3t3y1i",
   "tg_tokens": null,
   "tps": 30.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-xy3t3y1i-ctx16384",
   "model": "DeepSeek-R1-Distill-Qwen-14B",
   "notes": "oMLX community benchmark trace; peak memory 11.4 GB at this context; duplicate traces of the same session: q7jh3msn; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 740.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-R1-Distill-Qwen-14B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 740.4 23.8 11.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xy3t3y1i",
   "source_url": "https://omlx.ai/benchmarks/performance/xy3t3y1i",
   "tg_tokens": null,
   "tps": 23.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-h2iavxj4-ctx4096",
   "model": "GLM-4.7-Flash",
   "notes": "oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: 35hd4emo; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1494.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-4.7-Flash-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,494 60.9 17.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h2iavxj4",
   "source_url": "https://omlx.ai/benchmarks/performance/h2iavxj4",
   "tg_tokens": null,
   "tps": 60.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-h2iavxj4-ctx16384",
   "model": "GLM-4.7-Flash",
   "notes": "oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 35hd4emo; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 753.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-4.7-Flash-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 753.5 48.8 20.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h2iavxj4",
   "source_url": "https://omlx.ai/benchmarks/performance/h2iavxj4",
   "tg_tokens": null,
   "tps": 48.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-9zdrtilm-ctx1024",
   "model": "LFM2.5-8B-A1B",
   "notes": "oMLX community benchmark trace; peak memory 5.3 GB at this context; duplicate traces of the same session: dffzp3s3; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1897.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-8B-A1B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 1k 1,897 173.0 5.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9zdrtilm",
   "source_url": "https://omlx.ai/benchmarks/performance/9zdrtilm",
   "tg_tokens": null,
   "tps": 173.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-9zdrtilm-ctx4096",
   "model": "LFM2.5-8B-A1B",
   "notes": "oMLX community benchmark trace; peak memory 5.5 GB at this context; duplicate traces of the same session: dffzp3s3; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2088.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-8B-A1B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k 2,088 163.3 5.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9zdrtilm",
   "source_url": "https://omlx.ai/benchmarks/performance/9zdrtilm",
   "tg_tokens": null,
   "tps": 163.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-0vv2taqg-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: ckp4kj59; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 137.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 137.1 15.0 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0vv2taqg",
   "source_url": "https://omlx.ai/benchmarks/performance/0vv2taqg",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-0vv2taqg-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: ckp4kj59; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 123.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 123.5 14.2 21.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0vv2taqg",
   "source_url": "https://omlx.ai/benchmarks/performance/0vv2taqg",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-j7wl0sfw-ctx1024",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 31.7 GB at this context; duplicate traces of the same session: 3zxu4938, tim1210k; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 190.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 unknown; 1k 190.9 25.1 31.7 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j7wl0sfw",
   "source_url": "https://omlx.ai/benchmarks/performance/j7wl0sfw",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-j7wl0sfw-ctx4096",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 35.4 GB at this context; duplicate traces of the same session: 3zxu4938, tim1210k; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 197.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 unknown; 4k 197.2 26.6 35.4 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j7wl0sfw",
   "source_url": "https://omlx.ai/benchmarks/performance/j7wl0sfw",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-j7wl0sfw-ctx8192",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 36.2 GB at this context; duplicate traces of the same session: 3zxu4938, tim1210k; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 192.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 unknown; 8k 192.2 21.5 36.2 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j7wl0sfw",
   "source_url": "https://omlx.ai/benchmarks/performance/j7wl0sfw",
   "tg_tokens": null,
   "tps": 21.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-j7wl0sfw-ctx16384",
   "model": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2",
   "notes": "fp16 MTP build; oMLX community benchmark trace; peak memory 37.7 GB at this context; duplicate traces of the same session: 3zxu4938, tim1210k; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 182.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2 on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 unknown; 16k 182.0 35.0 37.7 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j7wl0sfw",
   "source_url": "https://omlx.ai/benchmarks/performance/j7wl0sfw",
   "tg_tokens": null,
   "tps": 35.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5urde0q2-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.1 GB at this context; duplicate traces of the same session: 6datn11j, r38g8xtg; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 899.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 899.2 87.7 99.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5urde0q2",
   "source_url": "https://omlx.ai/benchmarks/performance/5urde0q2",
   "tg_tokens": null,
   "tps": 87.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5urde0q2-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.7 GB at this context; duplicate traces of the same session: 6datn11j, r38g8xtg; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1098.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 4k 1,098 78.6 99.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5urde0q2",
   "source_url": "https://omlx.ai/benchmarks/performance/5urde0q2",
   "tg_tokens": null,
   "tps": 78.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5urde0q2-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.7 GB at this context; duplicate traces of the same session: 6datn11j, r38g8xtg; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1091.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,091 80.0 99.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5urde0q2",
   "source_url": "https://omlx.ai/benchmarks/performance/5urde0q2",
   "tg_tokens": null,
   "tps": 80.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-sd5dklvv-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 72.9 GB; TTFT 9705 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1688.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k tokens 1,688 PP tok/s 77.5 TG tok/s 9705 TTFT (ms) 72.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sd5dklvv",
   "source_url": "https://omlx.ai/benchmarks/performance/sd5dklvv",
   "tg_tokens": null,
   "tps": 77.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-lah3wqqv-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 72.9 GB; TTFT 11495 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1425.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k tokens 1,425 PP tok/s 56.1 TG tok/s 11495 TTFT (ms) 72.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lah3wqqv",
   "source_url": "https://omlx.ai/benchmarks/performance/lah3wqqv",
   "tg_tokens": null,
   "tps": 56.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-zx06bmll-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 72.9 GB; TTFT 10369 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1580.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k tokens 1,580 PP tok/s 73.7 TG tok/s 10369 TTFT (ms) 72.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zx06bmll",
   "source_url": "https://omlx.ai/benchmarks/performance/zx06bmll",
   "tg_tokens": null,
   "tps": 73.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fe7bp1av-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 72.9 GB; TTFT 6970 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2351.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k tokens 2,351 PP tok/s 85.2 TG tok/s 6970 TTFT (ms) 72.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fe7bp1av",
   "source_url": "https://omlx.ai/benchmarks/performance/fe7bp1av",
   "tg_tokens": null,
   "tps": 85.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-offbm93p-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 72.9 GB; TTFT 7225 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2268.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k tokens 2,268 PP tok/s 73.8 TG tok/s 7225 TTFT (ms) 72.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/offbm93p",
   "source_url": "https://omlx.ai/benchmarks/performance/offbm93p",
   "tg_tokens": null,
   "tps": 73.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jdh4zeu7-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 72.9 GB; TTFT 9460 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1732.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k tokens 1,732 PP tok/s 64.9 TG tok/s 9460 TTFT (ms) 72.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jdh4zeu7",
   "source_url": "https://omlx.ai/benchmarks/performance/jdh4zeu7",
   "tg_tokens": null,
   "tps": 64.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-kf8gvjak-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.8 GB at this context; duplicate traces of the same session: 0fy8r8rd; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4078.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 4,078 104.7 100.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kf8gvjak",
   "source_url": "https://omlx.ai/benchmarks/performance/kf8gvjak",
   "tg_tokens": null,
   "tps": 104.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-kf8gvjak-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.8 GB at this context; duplicate traces of the same session: 0fy8r8rd; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4093.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 4,093 121.1 101.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kf8gvjak",
   "source_url": "https://omlx.ai/benchmarks/performance/kf8gvjak",
   "tg_tokens": null,
   "tps": 121.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-36unnvez-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3048.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 3,048 160.7 100.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/36unnvez",
   "source_url": "https://omlx.ai/benchmarks/performance/36unnvez",
   "tg_tokens": null,
   "tps": 160.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-36unnvez-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.8 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4086.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 4,086 135.6 100.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/36unnvez",
   "source_url": "https://omlx.ai/benchmarks/performance/36unnvez",
   "tg_tokens": null,
   "tps": 135.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-a1a49ve2-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.4 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4237.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 4,237 112.6 99.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a1a49ve2",
   "source_url": "https://omlx.ai/benchmarks/performance/a1a49ve2",
   "tg_tokens": null,
   "tps": 112.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-a1a49ve2-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.2 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4264.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 4,264 109.5 100.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a1a49ve2",
   "source_url": "https://omlx.ai/benchmarks/performance/a1a49ve2",
   "tg_tokens": null,
   "tps": 109.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-03gdejee-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oct11 build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.1 GB at this context; duplicate traces of the same session: 5kp1xus5, w18us3lf; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 897.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-oct11 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k 897.2 87.0 99.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/03gdejee",
   "source_url": "https://omlx.ai/benchmarks/performance/03gdejee",
   "tg_tokens": null,
   "tps": 87.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-03gdejee-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oct11 build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.7 GB at this context; duplicate traces of the same session: 5kp1xus5, w18us3lf; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1090.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-oct11 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 8k 1,090 75.5 99.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/03gdejee",
   "source_url": "https://omlx.ai/benchmarks/performance/03gdejee",
   "tg_tokens": null,
   "tps": 75.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-03gdejee-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oct11 build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.0 GB at this context; duplicate traces of the same session: 5kp1xus5, w18us3lf; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1078.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-oct11 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 1,078 77.3 100.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/03gdejee",
   "source_url": "https://omlx.ai/benchmarks/performance/03gdejee",
   "tg_tokens": null,
   "tps": 77.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-s73ft3pf-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oct11 build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.0 GB at this context; duplicate traces of the same session: he16md5g; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1082.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-oct11 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 16k 1,082 79.2 100.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s73ft3pf",
   "source_url": "https://omlx.ai/benchmarks/performance/s73ft3pf",
   "tg_tokens": null,
   "tps": 79.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-s73ft3pf-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oct11 build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: he16md5g; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1065.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-oct11 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 32k 1,065 34.8 100.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s73ft3pf",
   "source_url": "https://omlx.ai/benchmarks/performance/s73ft3pf",
   "tg_tokens": null,
   "tps": 34.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-7055jknh-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oct11 build; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 99.1 GB; TTFT 1548 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 661.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-oct11 on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 4bit; 1k tokens 661.5 PP tok/s 88.1 TG tok/s 1548 TTFT (ms) 99.1 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7055jknh",
   "source_url": "https://omlx.ai/benchmarks/performance/7055jknh",
   "tg_tokens": null,
   "tps": 88.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-f4woj0pk-ctx1024",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 177.7 GB at this context; duplicate traces of the same session: 1xa9mh8c, 9fr78ojj, 9iw8hr2q, hdfpimqf, pxc28yvo, ts411vd8, z2p0e1i5; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 393.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "glm-5.3-flash-uncensored-oq4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 1k 393.4 48.8 177.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f4woj0pk",
   "source_url": "https://omlx.ai/benchmarks/performance/f4woj0pk",
   "tg_tokens": null,
   "tps": 48.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-f4woj0pk-ctx4096",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: 1xa9mh8c, 9fr78ojj, 9iw8hr2q, hdfpimqf, pxc28yvo, ts411vd8, z2p0e1i5; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 999.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "glm-5.3-flash-uncensored-oq4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 4k 999.0 59.3 178.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f4woj0pk",
   "source_url": "https://omlx.ai/benchmarks/performance/f4woj0pk",
   "tg_tokens": null,
   "tps": 59.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-f4woj0pk-ctx8192",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.2 GB at this context; duplicate traces of the same session: 1xa9mh8c, 9fr78ojj, 9iw8hr2q, hdfpimqf, pxc28yvo, ts411vd8, z2p0e1i5; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1646.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "glm-5.3-flash-uncensored-oq4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 8k 1,646 58.7 178.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f4woj0pk",
   "source_url": "https://omlx.ai/benchmarks/performance/f4woj0pk",
   "tg_tokens": null,
   "tps": 58.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-f4woj0pk-ctx16384",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.4 GB at this context; duplicate traces of the same session: 1xa9mh8c, 9fr78ojj, 9iw8hr2q, hdfpimqf, pxc28yvo, ts411vd8, z2p0e1i5; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1639.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "glm-5.3-flash-uncensored-oq4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 16k 1,639 131.5 178.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f4woj0pk",
   "source_url": "https://omlx.ai/benchmarks/performance/f4woj0pk",
   "tg_tokens": null,
   "tps": 131.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-f4woj0pk-ctx32768",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.8 GB at this context; duplicate traces of the same session: 1xa9mh8c, 9fr78ojj, 9iw8hr2q, hdfpimqf, pxc28yvo, ts411vd8, z2p0e1i5; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1627.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "glm-5.3-flash-uncensored-oq4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 32k 1,627 84.0 178.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f4woj0pk",
   "source_url": "https://omlx.ai/benchmarks/performance/f4woj0pk",
   "tg_tokens": null,
   "tps": 84.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-f4woj0pk-ctx65536",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.3 GB at this context; duplicate traces of the same session: 1xa9mh8c, 9fr78ojj, 9iw8hr2q, hdfpimqf, pxc28yvo, ts411vd8, z2p0e1i5; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1598.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "glm-5.3-flash-uncensored-oq4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 64k 1,598 71.4 179.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f4woj0pk",
   "source_url": "https://omlx.ai/benchmarks/performance/f4woj0pk",
   "tg_tokens": null,
   "tps": 71.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-f4woj0pk-ctx131072",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 180.4 GB at this context; duplicate traces of the same session: 1xa9mh8c, 9fr78ojj, 9iw8hr2q, hdfpimqf, pxc28yvo, ts411vd8, z2p0e1i5; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1547.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "glm-5.3-flash-uncensored-oq4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 128k 1,547 134.3 180.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f4woj0pk",
   "source_url": "https://omlx.ai/benchmarks/performance/f4woj0pk",
   "tg_tokens": null,
   "tps": 134.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-f4woj0pk-ctx199680",
   "model": "GLM-5.3-Flash-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 182.1 GB at this context; duplicate traces of the same session: 1xa9mh8c, 9fr78ojj, 9iw8hr2q, hdfpimqf, pxc28yvo, ts411vd8, z2p0e1i5; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1475.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "glm-5.3-flash-uncensored-oq4e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 4bit; 195k 1,475 60.8 182.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f4woj0pk",
   "source_url": "https://omlx.ai/benchmarks/performance/f4woj0pk",
   "tg_tokens": null,
   "tps": 60.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-ayoccga2-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "mlx-community; oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: iga554si, l6jmln4e; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 339.9,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "mlx-community--gpt-oss-20b-MXFP4-Q8 on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 339.9 34.2 11.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ayoccga2",
   "source_url": "https://omlx.ai/benchmarks/performance/ayoccga2",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-ayoccga2-ctx8192",
   "model": "gpt-oss 20B",
   "notes": "mlx-community; oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: iga554si, l6jmln4e; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 324.9,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "mlx-community--gpt-oss-20b-MXFP4-Q8 on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 8k 324.9 32.6 11.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ayoccga2",
   "source_url": "https://omlx.ai/benchmarks/performance/ayoccga2",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-ayoccga2-ctx16384",
   "model": "gpt-oss 20B",
   "notes": "mlx-community; oMLX community benchmark trace; peak memory 12.0 GB at this context; duplicate traces of the same session: iga554si, l6jmln4e; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 265.3,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "mlx-community--gpt-oss-20b-MXFP4-Q8 on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 16k 265.3 27.8 12.0 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ayoccga2",
   "source_url": "https://omlx.ai/benchmarks/performance/ayoccga2",
   "tg_tokens": null,
   "tps": 27.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-4vm8lcbs-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "mlx-community; oMLX community benchmark trace; peak memory 11.7 GB at this context; duplicate traces of the same session: 9411c9l1; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 373.6,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "mlx-community--gpt-oss-20b-MXFP4-Q8 on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 1k 373.6 34.9 11.7 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4vm8lcbs",
   "source_url": "https://omlx.ai/benchmarks/performance/4vm8lcbs",
   "tg_tokens": null,
   "tps": 34.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-4vm8lcbs-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "mlx-community; oMLX community benchmark trace; peak memory 11.7 GB at this context; duplicate traces of the same session: 9411c9l1; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 349.6,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "mlx-community--gpt-oss-20b-MXFP4-Q8 on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 349.6 34.2 11.7 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4vm8lcbs",
   "source_url": "https://omlx.ai/benchmarks/performance/4vm8lcbs",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-957ggcpt-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 168.5 GB at this context; duplicate traces of the same session: 2ukwtscm; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 564.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 564.5 76.7 168.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/957ggcpt",
   "source_url": "https://omlx.ai/benchmarks/performance/957ggcpt",
   "tg_tokens": null,
   "tps": 76.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-957ggcpt-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 169.3 GB at this context; duplicate traces of the same session: 2ukwtscm; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2000.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 2,000 74.9 169.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/957ggcpt",
   "source_url": "https://omlx.ai/benchmarks/performance/957ggcpt",
   "tg_tokens": null,
   "tps": 74.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-by57xbdw-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 171.9 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2033.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 2,033 34.5 171.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/by57xbdw",
   "source_url": "https://omlx.ai/benchmarks/performance/by57xbdw",
   "tg_tokens": null,
   "tps": 34.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-by57xbdw-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 172.6 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2308.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 2,308 132.8 172.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/by57xbdw",
   "source_url": "https://omlx.ai/benchmarks/performance/by57xbdw",
   "tg_tokens": null,
   "tps": 132.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-l2irk6jk-ctx1024",
   "model": "Occamy-1.0-6bit-XL-mlx",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 663.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Occamy-1.0-6bit-XL-mlx on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 1k 663.8 30.7 29.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l2irk6jk",
   "source_url": "https://omlx.ai/benchmarks/performance/l2irk6jk",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-l2irk6jk-ctx4096",
   "model": "Occamy-1.0-6bit-XL-mlx",
   "notes": null,
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 781.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Occamy-1.0-6bit-XL-mlx on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 6bit; 4k 781.5 32.8 30.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l2irk6jk",
   "source_url": "https://omlx.ai/benchmarks/performance/l2irk6jk",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-yavbq32g-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.1 GB at this context; duplicate traces of the same session: ww5mqsre; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3081.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 3,081 93.2 43.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yavbq32g",
   "source_url": "https://omlx.ai/benchmarks/performance/yavbq32g",
   "tg_tokens": null,
   "tps": 93.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yavbq32g-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.9 GB at this context; duplicate traces of the same session: ww5mqsre; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4241.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 4,241 87.8 43.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yavbq32g",
   "source_url": "https://omlx.ai/benchmarks/performance/yavbq32g",
   "tg_tokens": null,
   "tps": 87.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-7eesc520-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: r4tf7thb; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 439.9,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k 439.9 29.0 21.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7eesc520",
   "source_url": "https://omlx.ai/benchmarks/performance/7eesc520",
   "tg_tokens": null,
   "tps": 29.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-7eesc520-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: r4tf7thb; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 221.9,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 221.9 21.6 22.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7eesc520",
   "source_url": "https://omlx.ai/benchmarks/performance/7eesc520",
   "tg_tokens": null,
   "tps": 21.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-hgclqwab-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 1weulczv; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 101.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 101.3 19.2 19.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hgclqwab",
   "source_url": "https://omlx.ai/benchmarks/performance/hgclqwab",
   "tg_tokens": null,
   "tps": 19.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-hgclqwab-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: 1weulczv; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 93.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 93.2 17.8 21.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hgclqwab",
   "source_url": "https://omlx.ai/benchmarks/performance/hgclqwab",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-6ira140s-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 5y7yuqms, a06lhr0h, c61wmp2z, ssj8yf0v, xwu18ufv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 471.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 471.0 17.7 16.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ira140s",
   "source_url": "https://omlx.ai/benchmarks/performance/6ira140s",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-6ira140s-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 5y7yuqms, a06lhr0h, c61wmp2z, ssj8yf0v, xwu18ufv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 496.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 496.4 17.5 17.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ira140s",
   "source_url": "https://omlx.ai/benchmarks/performance/6ira140s",
   "tg_tokens": null,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-6ira140s-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: 5y7yuqms, a06lhr0h, c61wmp2z, ssj8yf0v, xwu18ufv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 486.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 8k 486.9 17.2 17.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ira140s",
   "source_url": "https://omlx.ai/benchmarks/performance/6ira140s",
   "tg_tokens": null,
   "tps": 17.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-6ira140s-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: 5y7yuqms, a06lhr0h, c61wmp2z, ssj8yf0v, xwu18ufv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 472.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 472.9 16.5 17.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ira140s",
   "source_url": "https://omlx.ai/benchmarks/performance/6ira140s",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-6ira140s-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: 5y7yuqms, a06lhr0h, c61wmp2z, ssj8yf0v, xwu18ufv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 465.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k 465.0 15.8 18.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ira140s",
   "source_url": "https://omlx.ai/benchmarks/performance/6ira140s",
   "tg_tokens": null,
   "tps": 15.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-6ira140s-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 5y7yuqms, a06lhr0h, c61wmp2z, ssj8yf0v, xwu18ufv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 411.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 64k 411.4 13.8 20.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6ira140s",
   "source_url": "https://omlx.ai/benchmarks/performance/6ira140s",
   "tg_tokens": null,
   "tps": 13.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-x3zwco5y-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: 4rngiszz; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 441.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 441.8 9.6 18.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x3zwco5y",
   "source_url": "https://omlx.ai/benchmarks/performance/x3zwco5y",
   "tg_tokens": null,
   "tps": 9.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-x3zwco5y-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 4rngiszz; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 448.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 448.2 9.3 19.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x3zwco5y",
   "source_url": "https://omlx.ai/benchmarks/performance/x3zwco5y",
   "tg_tokens": null,
   "tps": 9.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-v35aujn8-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: 2kiqvgct, m28jnwj7, zccfp9ct; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 195.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 195.6 45.5 22.8 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v35aujn8",
   "source_url": "https://omlx.ai/benchmarks/performance/v35aujn8",
   "tg_tokens": null,
   "tps": 45.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-v35aujn8-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: 2kiqvgct, m28jnwj7, zccfp9ct; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 258.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 8k 258.8 45.1 23.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v35aujn8",
   "source_url": "https://omlx.ai/benchmarks/performance/v35aujn8",
   "tg_tokens": null,
   "tps": 45.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-v35aujn8-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: 2kiqvgct, m28jnwj7, zccfp9ct; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 220.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 220.8 38.0 21.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v35aujn8",
   "source_url": "https://omlx.ai/benchmarks/performance/v35aujn8",
   "tg_tokens": null,
   "tps": 38.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-v35aujn8-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 2kiqvgct, m28jnwj7, zccfp9ct; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 168.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 168.0 26.4 23.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v35aujn8",
   "source_url": "https://omlx.ai/benchmarks/performance/v35aujn8",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1dpxrr7c-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 1czzpy2s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 930.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 930.7 33.8 16.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1dpxrr7c",
   "source_url": "https://omlx.ai/benchmarks/performance/1dpxrr7c",
   "tg_tokens": null,
   "tps": 33.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-1dpxrr7c-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 1czzpy2s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1024.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-Uncensored-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,024 33.1 17.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1dpxrr7c",
   "source_url": "https://omlx.ai/benchmarks/performance/1dpxrr7c",
   "tg_tokens": null,
   "tps": 33.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-2abvvjr0-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: fw3me0c6; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 109.3,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 1k 109.3 14.9 19.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2abvvjr0",
   "source_url": "https://omlx.ai/benchmarks/performance/2abvvjr0",
   "tg_tokens": null,
   "tps": 14.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-2abvvjr0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: fw3me0c6; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 116.7,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 32 GB \u00b7 4bit; 4k 116.7 23.7 20.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2abvvjr0",
   "source_url": "https://omlx.ai/benchmarks/performance/2abvvjr0",
   "tg_tokens": null,
   "tps": 23.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-y8jf8kqv-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3048.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k 3,048 160.7 100.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y8jf8kqv",
   "source_url": "https://omlx.ai/benchmarks/performance/y8jf8kqv",
   "tg_tokens": null,
   "tps": 160.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-y8jf8kqv-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.8 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4086.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 4,086 135.6 100.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y8jf8kqv",
   "source_url": "https://omlx.ai/benchmarks/performance/y8jf8kqv",
   "tg_tokens": null,
   "tps": 135.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-f6k9q114-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 99.0 GB; TTFT 318 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3222.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 1k tokens 3,222 PP tok/s 124.1 TG tok/s 318 TTFT (ms) 99.0 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f6k9q114",
   "source_url": "https://omlx.ai/benchmarks/performance/f6k9q114",
   "tg_tokens": null,
   "tps": 124.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-bs7xptgy-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.4 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4237.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 8k 4,237 112.6 99.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bs7xptgy",
   "source_url": "https://omlx.ai/benchmarks/performance/bs7xptgy",
   "tg_tokens": null,
   "tps": 112.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-17",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-bs7xptgy-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.2 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4264.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 32k 4,264 109.5 100.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bs7xptgy",
   "source_url": "https://omlx.ai/benchmarks/performance/bs7xptgy",
   "tg_tokens": null,
   "tps": 109.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-1xydwls7-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: fk33eu7s, m6p0po65, w8ffadqn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 290.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 290.1 46.4 22.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1xydwls7",
   "source_url": "https://omlx.ai/benchmarks/performance/1xydwls7",
   "tg_tokens": null,
   "tps": 46.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-1xydwls7-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: fk33eu7s, m6p0po65, w8ffadqn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 185.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 8k 185.2 46.3 23.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1xydwls7",
   "source_url": "https://omlx.ai/benchmarks/performance/1xydwls7",
   "tg_tokens": null,
   "tps": 46.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-1xydwls7-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: fk33eu7s, m6p0po65, w8ffadqn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 208.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 208.8 30.7 21.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1xydwls7",
   "source_url": "https://omlx.ai/benchmarks/performance/1xydwls7",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-1xydwls7-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: fk33eu7s, m6p0po65, w8ffadqn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 162.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 162.0 27.7 23.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1xydwls7",
   "source_url": "https://omlx.ai/benchmarks/performance/1xydwls7",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-hbjcrfre-ctx4096",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: k0xo26pv, ot3uzttt, pnwy2qqu, ub0gap5e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 285.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 4k 285.8 45.0 22.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hbjcrfre",
   "source_url": "https://omlx.ai/benchmarks/performance/hbjcrfre",
   "tg_tokens": null,
   "tps": 45.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-hbjcrfre-ctx8192",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.5 GB at this context; duplicate traces of the same session: k0xo26pv, ot3uzttt, pnwy2qqu, ub0gap5e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 227.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 8k 227.0 46.7 23.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hbjcrfre",
   "source_url": "https://omlx.ai/benchmarks/performance/hbjcrfre",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-hbjcrfre-ctx16384",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: k0xo26pv, ot3uzttt, pnwy2qqu, ub0gap5e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 232.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 16k 232.2 43.0 21.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hbjcrfre",
   "source_url": "https://omlx.ai/benchmarks/performance/hbjcrfre",
   "tg_tokens": null,
   "tps": 43.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-hbjcrfre-ctx32768",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: k0xo26pv, ot3uzttt, pnwy2qqu, ub0gap5e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 211.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 32k 211.5 30.6 23.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hbjcrfre",
   "source_url": "https://omlx.ai/benchmarks/performance/hbjcrfre",
   "tg_tokens": null,
   "tps": 30.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-hbjcrfre-ctx65536",
   "model": "Swift-1.5-Qwen3.8-27b-oQ4e",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.9 GB at this context; duplicate traces of the same session: k0xo26pv, ot3uzttt, pnwy2qqu, ub0gap5e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 161.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift-1.5-Qwen3.8-27b-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 48 GB \u00b7 4bit; 64k 161.7 19.4 28.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hbjcrfre",
   "source_url": "https://omlx.ai/benchmarks/performance/hbjcrfre",
   "tg_tokens": null,
   "tps": 19.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-hlgxqu01-ctx1024",
   "model": "Ternary-Bonsai-2-2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 3.5 GB at this context; duplicate traces of the same session: cim0ulgd; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2420.0,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-2B-mlx-2bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 2bit; 1k 2,420 215.1 3.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hlgxqu01",
   "source_url": "https://omlx.ai/benchmarks/performance/hlgxqu01",
   "tg_tokens": null,
   "tps": 215.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-hlgxqu01-ctx32768",
   "model": "Ternary-Bonsai-2-2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 5.2 GB at this context; duplicate traces of the same session: cim0ulgd; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4534.0,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-2B-mlx-2bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 2bit; 32k 4,534 187.8 5.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hlgxqu01",
   "source_url": "https://omlx.ai/benchmarks/performance/hlgxqu01",
   "tg_tokens": null,
   "tps": 187.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-vlezz98t-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "OptiQ build; oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: p1eqjzlj; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 391.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-OptiQ-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k 391.2 22.3 17.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vlezz98t",
   "source_url": "https://omlx.ai/benchmarks/performance/vlezz98t",
   "tg_tokens": null,
   "tps": 22.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-vlezz98t-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "OptiQ build; oMLX community benchmark trace; peak memory 17.5 GB at this context; duplicate traces of the same session: p1eqjzlj; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 365.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-OptiQ-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 365.2 20.9 17.5 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vlezz98t",
   "source_url": "https://omlx.ai/benchmarks/performance/vlezz98t",
   "tg_tokens": null,
   "tps": 20.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-it1lacha-ctx1024",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: c9p95j4y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1255.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,255 132.6 11.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/it1lacha",
   "source_url": "https://omlx.ai/benchmarks/performance/it1lacha",
   "tg_tokens": null,
   "tps": 132.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-it1lacha-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: c9p95j4y; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4520.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 4,520 123.9 11.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/it1lacha",
   "source_url": "https://omlx.ai/benchmarks/performance/it1lacha",
   "tg_tokens": null,
   "tps": 123.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xon879lq-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 93.7 GB at this context; duplicate traces of the same session: w9lqjzwn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.6,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 104.6 11.2 93.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xon879lq",
   "source_url": "https://omlx.ai/benchmarks/performance/xon879lq",
   "tg_tokens": null,
   "tps": 11.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-xon879lq-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 97.1 GB at this context; duplicate traces of the same session: w9lqjzwn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 464.2,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 464.2 14.6 97.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xon879lq",
   "source_url": "https://omlx.ai/benchmarks/performance/xon879lq",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xon879lq-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 97.2 GB at this context; duplicate traces of the same session: w9lqjzwn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 547.6,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 547.6 15.6 97.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xon879lq",
   "source_url": "https://omlx.ai/benchmarks/performance/xon879lq",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ymp9vv21-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 288.7 GB at this context; duplicate traces of the same session: eo5fnsty; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 342.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 342.4 28.9 288.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ymp9vv21",
   "source_url": "https://omlx.ai/benchmarks/performance/ymp9vv21",
   "tg_tokens": null,
   "tps": 28.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ymp9vv21-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 290.1 GB at this context; duplicate traces of the same session: eo5fnsty; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 399.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 16k 399.1 37.3 290.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ymp9vv21",
   "source_url": "https://omlx.ai/benchmarks/performance/ymp9vv21",
   "tg_tokens": null,
   "tps": 37.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-cm2va9ss-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 403.3 GB at this context; duplicate traces of the same session: a5ofvvr8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 249.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 249.9 26.8 403.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cm2va9ss",
   "source_url": "https://omlx.ai/benchmarks/performance/cm2va9ss",
   "tg_tokens": null,
   "tps": 26.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-cm2va9ss-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 404.6 GB at this context; duplicate traces of the same session: a5ofvvr8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 400.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 16k 400.9 37.0 404.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cm2va9ss",
   "source_url": "https://omlx.ai/benchmarks/performance/cm2va9ss",
   "tg_tokens": null,
   "tps": 37.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nzslx2j5-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 403.4 GB at this context; duplicate traces of the same session: 4buknjee, jp8j4pe6, ny8y83vz, q2jpw9q6, sbqsnccl, vt1fyxuo, xy3coe17; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 373.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 373.4 27.9 403.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nzslx2j5",
   "source_url": "https://omlx.ai/benchmarks/performance/nzslx2j5",
   "tg_tokens": null,
   "tps": 27.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nzslx2j5-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 403.9 GB at this context; duplicate traces of the same session: 4buknjee, jp8j4pe6, ny8y83vz, q2jpw9q6, sbqsnccl, vt1fyxuo, xy3coe17; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 427.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 427.4 33.3 403.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nzslx2j5",
   "source_url": "https://omlx.ai/benchmarks/performance/nzslx2j5",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-nzslx2j5-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 404.2 GB at this context; duplicate traces of the same session: 4buknjee, jp8j4pe6, ny8y83vz, q2jpw9q6, sbqsnccl, vt1fyxuo, xy3coe17; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 430.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 8k 430.7 36.1 404.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nzslx2j5",
   "source_url": "https://omlx.ai/benchmarks/performance/nzslx2j5",
   "tg_tokens": null,
   "tps": 36.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-nzslx2j5-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 404.7 GB at this context; duplicate traces of the same session: 4buknjee, jp8j4pe6, ny8y83vz, q2jpw9q6, sbqsnccl, vt1fyxuo, xy3coe17; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 426.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 16k 426.0 39.4 404.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nzslx2j5",
   "source_url": "https://omlx.ai/benchmarks/performance/nzslx2j5",
   "tg_tokens": null,
   "tps": 39.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nzslx2j5-ctx32768",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 404.7 GB at this context; duplicate traces of the same session: 4buknjee, jp8j4pe6, ny8y83vz, q2jpw9q6, sbqsnccl, vt1fyxuo, xy3coe17; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 418.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 32k 418.4 34.7 404.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nzslx2j5",
   "source_url": "https://omlx.ai/benchmarks/performance/nzslx2j5",
   "tg_tokens": null,
   "tps": 34.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nzslx2j5-ctx65536",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 404.8 GB at this context; duplicate traces of the same session: 4buknjee, jp8j4pe6, ny8y83vz, q2jpw9q6, sbqsnccl, vt1fyxuo, xy3coe17; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 406.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 64k 406.5 34.9 404.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nzslx2j5",
   "source_url": "https://omlx.ai/benchmarks/performance/nzslx2j5",
   "tg_tokens": null,
   "tps": 34.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nzslx2j5-ctx131072",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 404.8 GB at this context; duplicate traces of the same session: 4buknjee, jp8j4pe6, ny8y83vz, q2jpw9q6, sbqsnccl, vt1fyxuo, xy3coe17; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 392.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 128k 392.2 31.8 404.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nzslx2j5",
   "source_url": "https://omlx.ai/benchmarks/performance/nzslx2j5",
   "tg_tokens": null,
   "tps": 31.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-nzslx2j5-ctx199680",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 404.9 GB at this context; duplicate traces of the same session: 4buknjee, jp8j4pe6, ny8y83vz, q2jpw9q6, sbqsnccl, vt1fyxuo, xy3coe17; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 379.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 195k 379.1 39.6 404.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nzslx2j5",
   "source_url": "https://omlx.ai/benchmarks/performance/nzslx2j5",
   "tg_tokens": null,
   "tps": 39.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-pklegtzd-ctx4096",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 171.9 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2033.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 4k 2,033 34.5 171.9 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pklegtzd",
   "source_url": "https://omlx.ai/benchmarks/performance/pklegtzd",
   "tg_tokens": null,
   "tps": 34.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-pklegtzd-ctx16384",
   "model": "GLM-5.3-Flash",
   "notes": "oMLX community benchmark trace; peak memory 172.6 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2308.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "GLM-5.3-Flash-MLX-4bit on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 4bit; 16k 2,308 132.8 172.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pklegtzd",
   "source_url": "https://omlx.ai/benchmarks/performance/pklegtzd",
   "tg_tokens": null,
   "tps": 132.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-lr0zbg15-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: 0rtghc4p, rulnu2gv; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 754.3,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-6bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 1k 754.3 66.8 27.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lr0zbg15",
   "source_url": "https://omlx.ai/benchmarks/performance/lr0zbg15",
   "tg_tokens": null,
   "tps": 66.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-lr0zbg15-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: 0rtghc4p, rulnu2gv; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 803.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-6bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 4k 803.8 65.7 28.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lr0zbg15",
   "source_url": "https://omlx.ai/benchmarks/performance/lr0zbg15",
   "tg_tokens": null,
   "tps": 65.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-lr0zbg15-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 29.0 GB at this context; duplicate traces of the same session: 0rtghc4p, rulnu2gv; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 574.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-6bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 32k 574.8 49.0 29.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lr0zbg15",
   "source_url": "https://omlx.ai/benchmarks/performance/lr0zbg15",
   "tg_tokens": null,
   "tps": 49.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-n0z1jtp5-ctx1024",
   "model": "Ornith-1.5-35B-A3B-uncensored",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: k2h9ed6m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1117.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-uncensored-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 1,117 108.1 35.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n0z1jtp5",
   "source_url": "https://omlx.ai/benchmarks/performance/n0z1jtp5",
   "tg_tokens": null,
   "tps": 108.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-n0z1jtp5-ctx4096",
   "model": "Ornith-1.5-35B-A3B-uncensored",
   "notes": "oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: k2h9ed6m; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3284.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-uncensored-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 3,284 106.5 36.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n0z1jtp5",
   "source_url": "https://omlx.ai/benchmarks/performance/n0z1jtp5",
   "tg_tokens": null,
   "tps": 106.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-xvdlv0ix-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: 6gzycwrv, 79bofwuz, tnh6d35r, uu017uq1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2301.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 2,301 108.5 37.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xvdlv0ix",
   "source_url": "https://omlx.ai/benchmarks/performance/xvdlv0ix",
   "tg_tokens": null,
   "tps": 108.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-xvdlv0ix-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 38.0 GB at this context; duplicate traces of the same session: 6gzycwrv, 79bofwuz, tnh6d35r, uu017uq1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2953.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 2,953 88.7 38.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xvdlv0ix",
   "source_url": "https://omlx.ai/benchmarks/performance/xvdlv0ix",
   "tg_tokens": null,
   "tps": 88.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-xvdlv0ix-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 38.1 GB at this context; duplicate traces of the same session: 6gzycwrv, 79bofwuz, tnh6d35r, uu017uq1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2870.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 8k 2,870 109.2 38.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xvdlv0ix",
   "source_url": "https://omlx.ai/benchmarks/performance/xvdlv0ix",
   "tg_tokens": null,
   "tps": 109.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-xvdlv0ix-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 38.3 GB at this context; duplicate traces of the same session: 6gzycwrv, 79bofwuz, tnh6d35r, uu017uq1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2658.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 16k 2,658 111.4 38.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xvdlv0ix",
   "source_url": "https://omlx.ai/benchmarks/performance/xvdlv0ix",
   "tg_tokens": null,
   "tps": 111.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-xvdlv0ix-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 38.6 GB at this context; duplicate traces of the same session: 6gzycwrv, 79bofwuz, tnh6d35r, uu017uq1; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2211.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 32k 2,211 130.8 38.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xvdlv0ix",
   "source_url": "https://omlx.ai/benchmarks/performance/xvdlv0ix",
   "tg_tokens": null,
   "tps": 130.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-mwbzsag5-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2241.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 2,241 104.6 37.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mwbzsag5",
   "source_url": "https://omlx.ai/benchmarks/performance/mwbzsag5",
   "tg_tokens": null,
   "tps": 104.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-mwbzsag5-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 38.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1720.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 1,720 94.5 38.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mwbzsag5",
   "source_url": "https://omlx.ai/benchmarks/performance/mwbzsag5",
   "tg_tokens": null,
   "tps": 94.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-6sig9yyy-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 102.8 14.0 23.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6sig9yyy",
   "source_url": "https://omlx.ai/benchmarks/performance/6sig9yyy",
   "tg_tokens": null,
   "tps": 14.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-6sig9yyy-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 93.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 93.8 14.2 25.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6sig9yyy",
   "source_url": "https://omlx.ai/benchmarks/performance/6sig9yyy",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-k76qm128-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.4 GB at this context; duplicate traces of the same session: a7lhvvex; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 256.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k 256.0 16.7 29.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k76qm128",
   "source_url": "https://omlx.ai/benchmarks/performance/k76qm128",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-k76qm128-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.5 GB at this context; duplicate traces of the same session: a7lhvvex; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 261.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 4k 261.9 16.6 32.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k76qm128",
   "source_url": "https://omlx.ai/benchmarks/performance/k76qm128",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-705tw3q9-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: b0s0ag67; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 101.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 101.8 14.7 23.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/705tw3q9",
   "source_url": "https://omlx.ai/benchmarks/performance/705tw3q9",
   "tg_tokens": null,
   "tps": 14.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-705tw3q9-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: b0s0ag67; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 93.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 93.4 14.1 25.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/705tw3q9",
   "source_url": "https://omlx.ai/benchmarks/performance/705tw3q9",
   "tg_tokens": null,
   "tps": 14.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-1aphuyuo-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 23.9 GB; TTFT 50983 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 80.3,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 4k tokens 80.3 PP tok/s 13.9 TG tok/s 50983 TTFT (ms) 23.9 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1aphuyuo",
   "source_url": "https://omlx.ai/benchmarks/performance/1aphuyuo",
   "tg_tokens": null,
   "tps": 13.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-npjxwy5n-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: hox9eiwk, s5x0vn6z, wuzypmf7, xiqt0ajn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 905.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 905.6 58.5 17.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/npjxwy5n",
   "source_url": "https://omlx.ai/benchmarks/performance/npjxwy5n",
   "tg_tokens": null,
   "tps": 58.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-npjxwy5n-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: hox9eiwk, s5x0vn6z, wuzypmf7, xiqt0ajn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1003.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,003 64.0 18.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/npjxwy5n",
   "source_url": "https://omlx.ai/benchmarks/performance/npjxwy5n",
   "tg_tokens": null,
   "tps": 64.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-npjxwy5n-ctx8192",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: hox9eiwk, s5x0vn6z, wuzypmf7, xiqt0ajn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 927.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 927.9 69.8 19.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/npjxwy5n",
   "source_url": "https://omlx.ai/benchmarks/performance/npjxwy5n",
   "tg_tokens": null,
   "tps": 69.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-npjxwy5n-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: hox9eiwk, s5x0vn6z, wuzypmf7, xiqt0ajn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 814.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 814.4 83.9 19.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/npjxwy5n",
   "source_url": "https://omlx.ai/benchmarks/performance/npjxwy5n",
   "tg_tokens": null,
   "tps": 83.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-npjxwy5n-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: hox9eiwk, s5x0vn6z, wuzypmf7, xiqt0ajn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 752.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 752.5 137.6 20.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/npjxwy5n",
   "source_url": "https://omlx.ai/benchmarks/performance/npjxwy5n",
   "tg_tokens": null,
   "tps": 137.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-vg446mya-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.2 GB at this context; duplicate traces of the same session: kyqabal3; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 100.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 100.9 13.0 23.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vg446mya",
   "source_url": "https://omlx.ai/benchmarks/performance/vg446mya",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-vg446mya-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: kyqabal3; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 92.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 92.9 12.9 25.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vg446mya",
   "source_url": "https://omlx.ai/benchmarks/performance/vg446mya",
   "tg_tokens": null,
   "tps": 12.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-0tc0dguw-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 72.1 GB at this context; duplicate traces of the same session: f8m19kqe; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 505.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 505.6 56.8 72.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0tc0dguw",
   "source_url": "https://omlx.ai/benchmarks/performance/0tc0dguw",
   "tg_tokens": null,
   "tps": 56.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-0tc0dguw-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 72.8 GB at this context; duplicate traces of the same session: f8m19kqe; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 663.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 663.2 54.6 72.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0tc0dguw",
   "source_url": "https://omlx.ai/benchmarks/performance/0tc0dguw",
   "tg_tokens": null,
   "tps": 54.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ezxf9pvw-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: fl49w9hp; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4223.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 4,223 104.5 11.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ezxf9pvw",
   "source_url": "https://omlx.ai/benchmarks/performance/ezxf9pvw",
   "tg_tokens": null,
   "tps": 104.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ezxf9pvw-ctx16384",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 12.1 GB at this context; duplicate traces of the same session: fl49w9hp; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4029.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 4,029 90.5 12.1 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ezxf9pvw",
   "source_url": "https://omlx.ai/benchmarks/performance/ezxf9pvw",
   "tg_tokens": null,
   "tps": 90.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2zt8cul5-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: bmoxcfph; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4204.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 4,204 104.6 11.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2zt8cul5",
   "source_url": "https://omlx.ai/benchmarks/performance/2zt8cul5",
   "tg_tokens": null,
   "tps": 104.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2zt8cul5-ctx16384",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 12.1 GB at this context; duplicate traces of the same session: bmoxcfph; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4048.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 4,048 90.5 12.1 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2zt8cul5",
   "source_url": "https://omlx.ai/benchmarks/performance/2zt8cul5",
   "tg_tokens": null,
   "tps": 90.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-psqt12qp-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "1017 build; oMLX community benchmark trace; peak memory 288.8 GB at this context; duplicate traces of the same session: tddrjg65; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 435.4,
   "provenance": "community",
   "quant": "unknown",
   "quote": "DeepSeek-V4.1-Flash-1017 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 435.4 31.0 288.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/psqt12qp",
   "source_url": "https://omlx.ai/benchmarks/performance/psqt12qp",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-psqt12qp-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "1017 build; oMLX community benchmark trace; peak memory 292.3 GB at this context; duplicate traces of the same session: tddrjg65; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 645.8,
   "provenance": "community",
   "quant": "unknown",
   "quote": "DeepSeek-V4.1-Flash-1017 on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 8k 645.8 32.1 292.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/psqt12qp",
   "source_url": "https://omlx.ai/benchmarks/performance/psqt12qp",
   "tg_tokens": null,
   "tps": 32.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-rp6ncep2-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.8 GB at this context; duplicate traces of the same session: en0951m6, l19o4rl0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 110.6,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 110.6 11.6 92.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rp6ncep2",
   "source_url": "https://omlx.ai/benchmarks/performance/rp6ncep2",
   "tg_tokens": null,
   "tps": 11.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-rp6ncep2-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.4 GB at this context; duplicate traces of the same session: en0951m6, l19o4rl0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 510.7,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 510.7 14.5 96.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rp6ncep2",
   "source_url": "https://omlx.ai/benchmarks/performance/rp6ncep2",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-rp6ncep2-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.3 GB at this context; duplicate traces of the same session: en0951m6, l19o4rl0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 589.4,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 589.4 15.9 96.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rp6ncep2",
   "source_url": "https://omlx.ai/benchmarks/performance/rp6ncep2",
   "tg_tokens": null,
   "tps": 15.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-m29c9vw1-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 93.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 104.6,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 104.6 11.2 93.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m29c9vw1",
   "source_url": "https://omlx.ai/benchmarks/performance/m29c9vw1",
   "tg_tokens": null,
   "tps": 11.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-m29c9vw1-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 97.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 464.2,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 464.2 14.6 97.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m29c9vw1",
   "source_url": "https://omlx.ai/benchmarks/performance/m29c9vw1",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-m29c9vw1-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 97.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 547.6,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 547.6 15.6 97.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m29c9vw1",
   "source_url": "https://omlx.ai/benchmarks/performance/m29c9vw1",
   "tg_tokens": null,
   "tps": 15.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-vp1l62ba-ctx4096",
   "model": "Huihui-GLM-4.7-Flash-abliterated",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 1twfp7n6; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 650.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Huihui-GLM-4.7-Flash-abliterated-mlx-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k 650.9 45.2 17.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vp1l62ba",
   "source_url": "https://omlx.ai/benchmarks/performance/vp1l62ba",
   "tg_tokens": null,
   "tps": 45.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-vp1l62ba-ctx16384",
   "model": "Huihui-GLM-4.7-Flash-abliterated",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 22.0 GB at this context; duplicate traces of the same session: 1twfp7n6; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 394.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Huihui-GLM-4.7-Flash-abliterated-mlx-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 16k 394.9 34.1 22.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vp1l62ba",
   "source_url": "https://omlx.ai/benchmarks/performance/vp1l62ba",
   "tg_tokens": null,
   "tps": 34.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-2shlydy3-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.1 GB at this context; duplicate traces of the same session: fyuu0zpw; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 618.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k 618.4 60.8 13.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2shlydy3",
   "source_url": "https://omlx.ai/benchmarks/performance/2shlydy3",
   "tg_tokens": null,
   "tps": 60.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-2shlydy3-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 13.5 GB at this context; duplicate traces of the same session: fyuu0zpw; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 655.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 655.7 58.8 13.5 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2shlydy3",
   "source_url": "https://omlx.ai/benchmarks/performance/2shlydy3",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-m141jjcj-ctx1024",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "OptiQ REAP-19B build; oMLX community benchmark trace; peak memory 14.7 GB at this context; duplicate traces of the same session: widbxmu2; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 507.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM-V-4.7-35B-A3B-OptiQ-4bit-REAP-19B on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k 507.9 33.3 14.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m141jjcj",
   "source_url": "https://omlx.ai/benchmarks/performance/m141jjcj",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-m141jjcj-ctx4096",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "OptiQ REAP-19B build; oMLX community benchmark trace; peak memory 15.3 GB at this context; duplicate traces of the same session: widbxmu2; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 540.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM-V-4.7-35B-A3B-OptiQ-4bit-REAP-19B on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 540.1 32.7 15.3 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m141jjcj",
   "source_url": "https://omlx.ai/benchmarks/performance/m141jjcj",
   "tg_tokens": null,
   "tps": 32.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-15bmldpv-ctx1024",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "OptiQ REAP-19B build; oMLX community benchmark trace; peak memory 14.7 GB at this context; duplicate traces of the same session: r10fdx4i; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 490.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM-V-4.7-35B-A3B-OptiQ-4bit-REAP-19B on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k 490.0 33.2 14.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/15bmldpv",
   "source_url": "https://omlx.ai/benchmarks/performance/15bmldpv",
   "tg_tokens": null,
   "tps": 33.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-15bmldpv-ctx4096",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "OptiQ REAP-19B build; oMLX community benchmark trace; peak memory 15.3 GB at this context; duplicate traces of the same session: r10fdx4i; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 540.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM-V-4.7-35B-A3B-OptiQ-4bit-REAP-19B on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 540.3 32.7 15.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/15bmldpv",
   "source_url": "https://omlx.ai/benchmarks/performance/15bmldpv",
   "tg_tokens": null,
   "tps": 32.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-pw3u30m6-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: yizcxvp3; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 487.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k 487.0 49.4 19.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pw3u30m6",
   "source_url": "https://omlx.ai/benchmarks/performance/pw3u30m6",
   "tg_tokens": null,
   "tps": 49.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-pw3u30m6-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: yizcxvp3; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 530.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 530.2 48.1 20.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pw3u30m6",
   "source_url": "https://omlx.ai/benchmarks/performance/pw3u30m6",
   "tg_tokens": null,
   "tps": 48.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-9mbjzrnq-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: oaejxxm6; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 477.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k 477.1 48.9 19.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9mbjzrnq",
   "source_url": "https://omlx.ai/benchmarks/performance/9mbjzrnq",
   "tg_tokens": null,
   "tps": 48.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-9mbjzrnq-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: oaejxxm6; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 527.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 527.9 47.7 20.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9mbjzrnq",
   "source_url": "https://omlx.ai/benchmarks/performance/9mbjzrnq",
   "tg_tokens": null,
   "tps": 47.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-pkvgb6fi-ctx1024",
   "model": "Qwen3.6-28B-REAP",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 16.6 GB; TTFT 2071 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 494.5,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.6-28B-REAP-oQ4 on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k tokens 494.5 PP tok/s 40.3 TG tok/s 2071 TTFT (ms) 16.6 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pkvgb6fi",
   "source_url": "https://omlx.ai/benchmarks/performance/pkvgb6fi",
   "tg_tokens": null,
   "tps": 40.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-b2cqccbn-ctx1024",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 12.7 GB at this context; duplicate traces of the same session: seuoytpu; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 507.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k 507.8 34.5 12.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b2cqccbn",
   "source_url": "https://omlx.ai/benchmarks/performance/b2cqccbn",
   "tg_tokens": null,
   "tps": 34.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-b2cqccbn-ctx4096",
   "model": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B",
   "notes": "oMLX community benchmark trace; peak memory 13.3 GB at this context; duplicate traces of the same session: seuoytpu; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 540.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-OptiQ-4bit-REAP-19B on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 540.8 33.8 13.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b2cqccbn",
   "source_url": "https://omlx.ai/benchmarks/performance/b2cqccbn",
   "tg_tokens": null,
   "tps": 33.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1gvcz4eq-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: pso5wndx; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2891.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,891 77.3 22.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1gvcz4eq",
   "source_url": "https://omlx.ai/benchmarks/performance/1gvcz4eq",
   "tg_tokens": null,
   "tps": 77.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-1gvcz4eq-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: pso5wndx; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2409.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 2,409 168.3 22.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1gvcz4eq",
   "source_url": "https://omlx.ai/benchmarks/performance/1gvcz4eq",
   "tg_tokens": null,
   "tps": 168.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-duzufb23-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2241.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 2,241 104.6 37.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/duzufb23",
   "source_url": "https://omlx.ai/benchmarks/performance/duzufb23",
   "tg_tokens": null,
   "tps": 104.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-duzufb23-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 38.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1720.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 1,720 94.5 38.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/duzufb23",
   "source_url": "https://omlx.ai/benchmarks/performance/duzufb23",
   "tg_tokens": null,
   "tps": 94.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-ujyhqcxm-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.7 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 102.8 14.0 23.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ujyhqcxm",
   "source_url": "https://omlx.ai/benchmarks/performance/ujyhqcxm",
   "tg_tokens": null,
   "tps": 14.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-ujyhqcxm-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 93.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 93.8 14.2 25.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ujyhqcxm",
   "source_url": "https://omlx.ai/benchmarks/performance/ujyhqcxm",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-g1uq9ap9-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 330.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 128k 330.9 11.3 24.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g1uq9ap9",
   "source_url": "https://omlx.ai/benchmarks/performance/g1uq9ap9",
   "tg_tokens": null,
   "tps": 11.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-g1uq9ap9-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 273.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 195k 273.6 9.4 29.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g1uq9ap9",
   "source_url": "https://omlx.ai/benchmarks/performance/g1uq9ap9",
   "tg_tokens": null,
   "tps": 9.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-we1z4ceh-ctx1024",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "MLX MTP build; oMLX community benchmark trace; peak memory 53.5 GB at this context; duplicate traces of the same session: p3blpfg8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 132.9,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.8-27B-Abliterated-MLX-bf16-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 bf16; 1k 132.9 10.3 53.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/we1z4ceh",
   "source_url": "https://omlx.ai/benchmarks/performance/we1z4ceh",
   "tg_tokens": null,
   "tps": 10.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-we1z4ceh-ctx4096",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "MLX MTP build; oMLX community benchmark trace; peak memory 56.2 GB at this context; duplicate traces of the same session: p3blpfg8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 137.8,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.8-27B-Abliterated-MLX-bf16-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 bf16; 4k 137.8 11.1 56.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/we1z4ceh",
   "source_url": "https://omlx.ai/benchmarks/performance/we1z4ceh",
   "tg_tokens": null,
   "tps": 11.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-uu3ldvfo-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: qcjz0chc; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 103.1 18.1 20.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uu3ldvfo",
   "source_url": "https://omlx.ai/benchmarks/performance/uu3ldvfo",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-uu3ldvfo-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: qcjz0chc; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 93.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 93.4 17.1 22.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uu3ldvfo",
   "source_url": "https://omlx.ai/benchmarks/performance/uu3ldvfo",
   "tg_tokens": null,
   "tps": 17.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-q9xc3abg-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.6 GB at this context; duplicate traces of the same session: cn9rf4yt; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 99.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 99.2 23.5 22.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q9xc3abg",
   "source_url": "https://omlx.ai/benchmarks/performance/q9xc3abg",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-q9xc3abg-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: cn9rf4yt; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 90.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 90.8 28.5 24.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q9xc3abg",
   "source_url": "https://omlx.ai/benchmarks/performance/q9xc3abg",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-d8hkcjrb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: 8iv8v4f2, g7am21su, hhv71i9k, o5o9u90r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 784.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 784.2 60.5 26.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d8hkcjrb",
   "source_url": "https://omlx.ai/benchmarks/performance/d8hkcjrb",
   "tg_tokens": null,
   "tps": 60.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-d8hkcjrb-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.8 GB at this context; duplicate traces of the same session: 8iv8v4f2, g7am21su, hhv71i9k, o5o9u90r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 897.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 897.2 52.1 27.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d8hkcjrb",
   "source_url": "https://omlx.ai/benchmarks/performance/d8hkcjrb",
   "tg_tokens": null,
   "tps": 52.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-d8hkcjrb-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: 8iv8v4f2, g7am21su, hhv71i9k, o5o9u90r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 760.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 760.1 52.1 28.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d8hkcjrb",
   "source_url": "https://omlx.ai/benchmarks/performance/d8hkcjrb",
   "tg_tokens": null,
   "tps": 52.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-d8hkcjrb-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.6 GB at this context; duplicate traces of the same session: 8iv8v4f2, g7am21su, hhv71i9k, o5o9u90r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 679.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 679.5 57.4 28.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d8hkcjrb",
   "source_url": "https://omlx.ai/benchmarks/performance/d8hkcjrb",
   "tg_tokens": null,
   "tps": 57.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-d8hkcjrb-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 29.7 GB at this context; duplicate traces of the same session: 8iv8v4f2, g7am21su, hhv71i9k, o5o9u90r; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 668.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 32k 668.9 73.7 29.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d8hkcjrb",
   "source_url": "https://omlx.ai/benchmarks/performance/d8hkcjrb",
   "tg_tokens": null,
   "tps": 73.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-4oq9gyux-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 86.4 GB at this context; duplicate traces of the same session: 1gwrodos, 2y52a4mf, 60ecpk6j, b4uuf0ne, f9ybayl9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 402.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 402.0 45.9 86.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4oq9gyux",
   "source_url": "https://omlx.ai/benchmarks/performance/4oq9gyux",
   "tg_tokens": null,
   "tps": 45.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-4oq9gyux-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 87.0 GB at this context; duplicate traces of the same session: 1gwrodos, 2y52a4mf, 60ecpk6j, b4uuf0ne, f9ybayl9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 540.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 540.2 48.1 87.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4oq9gyux",
   "source_url": "https://omlx.ai/benchmarks/performance/4oq9gyux",
   "tg_tokens": null,
   "tps": 48.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-4oq9gyux-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 87.4 GB at this context; duplicate traces of the same session: 1gwrodos, 2y52a4mf, 60ecpk6j, b4uuf0ne, f9ybayl9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 603.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 603.9 41.2 87.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4oq9gyux",
   "source_url": "https://omlx.ai/benchmarks/performance/4oq9gyux",
   "tg_tokens": null,
   "tps": 41.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-4oq9gyux-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 88.3 GB at this context; duplicate traces of the same session: 1gwrodos, 2y52a4mf, 60ecpk6j, b4uuf0ne, f9ybayl9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 575.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 575.2 43.9 88.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4oq9gyux",
   "source_url": "https://omlx.ai/benchmarks/performance/4oq9gyux",
   "tg_tokens": null,
   "tps": 43.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-4oq9gyux-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 89.4 GB at this context; duplicate traces of the same session: 1gwrodos, 2y52a4mf, 60ecpk6j, b4uuf0ne, f9ybayl9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 657.3,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 657.3 44.7 89.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4oq9gyux",
   "source_url": "https://omlx.ai/benchmarks/performance/4oq9gyux",
   "tg_tokens": null,
   "tps": 44.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-4oq9gyux-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 91.8 GB at this context; duplicate traces of the same session: 1gwrodos, 2y52a4mf, 60ecpk6j, b4uuf0ne, f9ybayl9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 660.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 660.0 41.3 91.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4oq9gyux",
   "source_url": "https://omlx.ai/benchmarks/performance/4oq9gyux",
   "tg_tokens": null,
   "tps": 41.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-m8f2e6k1-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.0 GB at this context; duplicate traces of the same session: bofegjaa; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 409.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k 409.8 38.4 14.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m8f2e6k1",
   "source_url": "https://omlx.ai/benchmarks/performance/m8f2e6k1",
   "tg_tokens": null,
   "tps": 38.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-m8f2e6k1-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: bofegjaa; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 393.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26b-a4b-it-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 4k 393.8 31.9 14.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m8f2e6k1",
   "source_url": "https://omlx.ai/benchmarks/performance/m8f2e6k1",
   "tg_tokens": null,
   "tps": 31.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-x5gtvha4-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 90.8 GB at this context; duplicate traces of the same session: d1gxxneh, d66ntx6l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 109.3,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 109.3 10.7 90.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x5gtvha4",
   "source_url": "https://omlx.ai/benchmarks/performance/x5gtvha4",
   "tg_tokens": null,
   "tps": 10.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-x5gtvha4-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 94.5 GB at this context; duplicate traces of the same session: d1gxxneh, d66ntx6l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 508.6,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 508.6 13.4 94.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x5gtvha4",
   "source_url": "https://omlx.ai/benchmarks/performance/x5gtvha4",
   "tg_tokens": null,
   "tps": 13.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-x5gtvha4-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 94.6 GB at this context; duplicate traces of the same session: d1gxxneh, d66ntx6l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 593.7,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 593.7 16.2 94.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x5gtvha4",
   "source_url": "https://omlx.ai/benchmarks/performance/x5gtvha4",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-i4ehmwue-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 290.5 GB at this context; duplicate traces of the same session: 3wrb9f95, 5s9pohgj, qaozbz85, vhroq0gj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 494.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 494.8 27.2 290.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i4ehmwue",
   "source_url": "https://omlx.ai/benchmarks/performance/i4ehmwue",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-i4ehmwue-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 290.5 GB at this context; duplicate traces of the same session: 3wrb9f95, 5s9pohgj, qaozbz85, vhroq0gj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 650.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 650.4 29.6 290.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i4ehmwue",
   "source_url": "https://omlx.ai/benchmarks/performance/i4ehmwue",
   "tg_tokens": null,
   "tps": 29.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-i4ehmwue-ctx32768",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 290.6 GB at this context; duplicate traces of the same session: 3wrb9f95, 5s9pohgj, qaozbz85, vhroq0gj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 686.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 32k 686.7 34.0 290.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i4ehmwue",
   "source_url": "https://omlx.ai/benchmarks/performance/i4ehmwue",
   "tg_tokens": null,
   "tps": 34.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-i4ehmwue-ctx131072",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 290.7 GB at this context; duplicate traces of the same session: 3wrb9f95, 5s9pohgj, qaozbz85, vhroq0gj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 684.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 128k 684.6 32.9 290.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i4ehmwue",
   "source_url": "https://omlx.ai/benchmarks/performance/i4ehmwue",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-i4ehmwue-ctx199680",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 290.8 GB at this context; duplicate traces of the same session: 3wrb9f95, 5s9pohgj, qaozbz85, vhroq0gj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 669.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 195k 669.9 40.3 290.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i4ehmwue",
   "source_url": "https://omlx.ai/benchmarks/performance/i4ehmwue",
   "tg_tokens": null,
   "tps": 40.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-olhpnsgh-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 207.3 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 375.3,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 1k 375.3 18.0 207.3 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/olhpnsgh",
   "source_url": "https://omlx.ai/benchmarks/performance/olhpnsgh",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-olhpnsgh-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 207.4 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 500.3,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 4k 500.3 20.9 207.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/olhpnsgh",
   "source_url": "https://omlx.ai/benchmarks/performance/olhpnsgh",
   "tg_tokens": null,
   "tps": 20.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-olhpnsgh-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 207.6 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 515.3,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 16k 515.3 20.6 207.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/olhpnsgh",
   "source_url": "https://omlx.ai/benchmarks/performance/olhpnsgh",
   "tg_tokens": null,
   "tps": 20.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-g11gcakl-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 207.3 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 392.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 1k 392.0 26.4 207.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g11gcakl",
   "source_url": "https://omlx.ai/benchmarks/performance/g11gcakl",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-g11gcakl-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 207.4 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 507.5,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 4k 507.5 31.8 207.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g11gcakl",
   "source_url": "https://omlx.ai/benchmarks/performance/g11gcakl",
   "tg_tokens": null,
   "tps": 31.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-g11gcakl-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 207.6 GB at this context; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 524.9,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 16k 524.9 39.4 207.6 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g11gcakl",
   "source_url": "https://omlx.ai/benchmarks/performance/g11gcakl",
   "tg_tokens": null,
   "tps": 39.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-nz3mu455-ctx4096",
   "model": "Huihui-GLM-4.7-Flash-abliterated",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: uz03g9f4; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 646.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Huihui-GLM-4.7-Flash-abliterated-mlx-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 4k 646.6 50.7 17.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nz3mu455",
   "source_url": "https://omlx.ai/benchmarks/performance/nz3mu455",
   "tg_tokens": null,
   "tps": 50.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-nz3mu455-ctx16384",
   "model": "Huihui-GLM-4.7-Flash-abliterated",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: uz03g9f4; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 395.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Huihui-GLM-4.7-Flash-abliterated-mlx-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 16k 395.3 34.5 20.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nz3mu455",
   "source_url": "https://omlx.ai/benchmarks/performance/nz3mu455",
   "tg_tokens": null,
   "tps": 34.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-zxcuobba-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX DWQ v2 build; oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: 08vwkghy, 6jok4fwt, czgywm52, jn2s431k, xtt5cr7g, yyknjg00; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 869.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-DWQ-4bit-v2 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 869.1 86.1 13.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zxcuobba",
   "source_url": "https://omlx.ai/benchmarks/performance/zxcuobba",
   "tg_tokens": null,
   "tps": 86.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-zxcuobba-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX DWQ v2 build; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: 08vwkghy, 6jok4fwt, czgywm52, jn2s431k, xtt5cr7g, yyknjg00; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1013.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-DWQ-4bit-v2 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 1,013 84.2 13.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zxcuobba",
   "source_url": "https://omlx.ai/benchmarks/performance/zxcuobba",
   "tg_tokens": null,
   "tps": 84.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-zxcuobba-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX DWQ v2 build; oMLX community benchmark trace; peak memory 13.8 GB at this context; duplicate traces of the same session: 08vwkghy, 6jok4fwt, czgywm52, jn2s431k, xtt5cr7g, yyknjg00; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 992.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-DWQ-4bit-v2 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 992.6 81.3 13.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zxcuobba",
   "source_url": "https://omlx.ai/benchmarks/performance/zxcuobba",
   "tg_tokens": null,
   "tps": 81.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-zxcuobba-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX DWQ v2 build; oMLX community benchmark trace; peak memory 13.9 GB at this context; duplicate traces of the same session: 08vwkghy, 6jok4fwt, czgywm52, jn2s431k, xtt5cr7g, yyknjg00; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 946.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-DWQ-4bit-v2 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 946.6 75.5 13.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zxcuobba",
   "source_url": "https://omlx.ai/benchmarks/performance/zxcuobba",
   "tg_tokens": null,
   "tps": 75.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-zxcuobba-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX DWQ v2 build; oMLX community benchmark trace; peak memory 14.2 GB at this context; duplicate traces of the same session: 08vwkghy, 6jok4fwt, czgywm52, jn2s431k, xtt5cr7g, yyknjg00; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 866.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-DWQ-4bit-v2 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 32k 866.1 63.4 14.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zxcuobba",
   "source_url": "https://omlx.ai/benchmarks/performance/zxcuobba",
   "tg_tokens": null,
   "tps": 63.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-zxcuobba-ctx65536",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX DWQ v2 build; oMLX community benchmark trace; peak memory 14.9 GB at this context; duplicate traces of the same session: 08vwkghy, 6jok4fwt, czgywm52, jn2s431k, xtt5cr7g, yyknjg00; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 736.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-DWQ-4bit-v2 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 64k 736.3 47.7 14.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zxcuobba",
   "source_url": "https://omlx.ai/benchmarks/performance/zxcuobba",
   "tg_tokens": null,
   "tps": 47.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-zxcuobba-ctx131072",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX DWQ v2 build; oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: 08vwkghy, 6jok4fwt, czgywm52, jn2s431k, xtt5cr7g, yyknjg00; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 563.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-DWQ-4bit-v2 on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 128k 563.8 36.8 16.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zxcuobba",
   "source_url": "https://omlx.ai/benchmarks/performance/zxcuobba",
   "tg_tokens": null,
   "tps": 36.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-vb1vrkgn-ctx1024",
   "model": "Ornith-1.5-35B-A3B-uncensored",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 5392y5xh, uz0vgzby; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1012.0,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Ornith-1.5-35B-A3B-uncensored-MLX-MXFP4 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,012 142.9 18.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vb1vrkgn",
   "source_url": "https://omlx.ai/benchmarks/performance/vb1vrkgn",
   "tg_tokens": null,
   "tps": 142.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-vb1vrkgn-ctx4096",
   "model": "Ornith-1.5-35B-A3B-uncensored",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 5392y5xh, uz0vgzby; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2950.0,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Ornith-1.5-35B-A3B-uncensored-MLX-MXFP4 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,950 140.1 19.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vb1vrkgn",
   "source_url": "https://omlx.ai/benchmarks/performance/vb1vrkgn",
   "tg_tokens": null,
   "tps": 140.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-vb1vrkgn-ctx32768",
   "model": "Ornith-1.5-35B-A3B-uncensored",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 5392y5xh, uz0vgzby; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4891.0,
   "provenance": "community",
   "quant": "MXFP4",
   "quote": "Ornith-1.5-35B-A3B-uncensored-MLX-MXFP4 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 4,891 120.0 19.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vb1vrkgn",
   "source_url": "https://omlx.ai/benchmarks/performance/vb1vrkgn",
   "tg_tokens": null,
   "tps": 120.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-2wkw4iwv-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: q8j9jjur; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2775.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,775 110.2 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2wkw4iwv",
   "source_url": "https://omlx.ai/benchmarks/performance/2wkw4iwv",
   "tg_tokens": null,
   "tps": 110.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-2wkw4iwv-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: q8j9jjur; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3474.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,474 107.7 20.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2wkw4iwv",
   "source_url": "https://omlx.ai/benchmarks/performance/2wkw4iwv",
   "tg_tokens": null,
   "tps": 107.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-2wkw4iwv-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: q8j9jjur; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3331.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 3,331 103.5 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2wkw4iwv",
   "source_url": "https://omlx.ai/benchmarks/performance/2wkw4iwv",
   "tg_tokens": null,
   "tps": 103.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-2wkw4iwv-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: q8j9jjur; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3069.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 3,069 97.5 20.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2wkw4iwv",
   "source_url": "https://omlx.ai/benchmarks/performance/2wkw4iwv",
   "tg_tokens": null,
   "tps": 97.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-2wkw4iwv-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: q8j9jjur; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2585.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 2,585 85.2 20.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2wkw4iwv",
   "source_url": "https://omlx.ai/benchmarks/performance/2wkw4iwv",
   "tg_tokens": null,
   "tps": 85.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-2wkw4iwv-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: q8j9jjur; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1903.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 1,903 64.5 21.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2wkw4iwv",
   "source_url": "https://omlx.ai/benchmarks/performance/2wkw4iwv",
   "tg_tokens": null,
   "tps": 64.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-lko8qkiy-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 21.4 GB; TTFT 597 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1716.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,716 PP tok/s 173.1 TG tok/s 597 TTFT (ms) 21.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lko8qkiy",
   "source_url": "https://omlx.ai/benchmarks/performance/lko8qkiy",
   "tg_tokens": null,
   "tps": 173.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-rx6f16br-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 28.9 GB; TTFT 11240 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1458.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k tokens 1,458 PP tok/s 27.2 TG tok/s 11240 TTFT (ms) 28.9 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rx6f16br",
   "source_url": "https://omlx.ai/benchmarks/performance/rx6f16br",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-crue30b3-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: tatp7biu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 101.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 101.2 18.2 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/crue30b3",
   "source_url": "https://omlx.ai/benchmarks/performance/crue30b3",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-crue30b3-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: tatp7biu; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 93.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 93.3 24.8 25.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/crue30b3",
   "source_url": "https://omlx.ai/benchmarks/performance/crue30b3",
   "tg_tokens": null,
   "tps": 24.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0ifh0935-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 24.9 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 330.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 128k 330.9 11.3 24.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ifh0935",
   "source_url": "https://omlx.ai/benchmarks/performance/0ifh0935",
   "tg_tokens": null,
   "tps": 11.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-0ifh0935-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 273.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 195k 273.6 9.4 29.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ifh0935",
   "source_url": "https://omlx.ai/benchmarks/performance/0ifh0935",
   "tg_tokens": null,
   "tps": 9.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-3rxq52k9-ctx1024",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 50.8 GB at this context; duplicate traces of the same session: 8uu2yc2n; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 117.9,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.8-27B-Abliteratedbf16-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 bf16; 1k 117.9 4.8 50.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3rxq52k9",
   "source_url": "https://omlx.ai/benchmarks/performance/3rxq52k9",
   "tg_tokens": null,
   "tps": 4.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-3rxq52k9-ctx4096",
   "model": "Qwen3.8-27B-Abliterated",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 51.9 GB at this context; duplicate traces of the same session: 8uu2yc2n; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 138.7,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.8-27B-Abliteratedbf16-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 bf16; 4k 138.7 4.8 51.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3rxq52k9",
   "source_url": "https://omlx.ai/benchmarks/performance/3rxq52k9",
   "tg_tokens": null,
   "tps": 4.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-xtq8evol-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 7ka12w68; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 100.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 100.9 17.9 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xtq8evol",
   "source_url": "https://omlx.ai/benchmarks/performance/xtq8evol",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-xtq8evol-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: 7ka12w68; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 93.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 93.3 24.9 25.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xtq8evol",
   "source_url": "https://omlx.ai/benchmarks/performance/xtq8evol",
   "tg_tokens": null,
   "tps": 24.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-pk3bf3p8-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: 59su7lxr; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 102.5 19.2 19.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pk3bf3p8",
   "source_url": "https://omlx.ai/benchmarks/performance/pk3bf3p8",
   "tg_tokens": null,
   "tps": 19.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-pk3bf3p8-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.7 GB at this context; duplicate traces of the same session: 59su7lxr; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 93.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 93.2 17.8 21.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pk3bf3p8",
   "source_url": "https://omlx.ai/benchmarks/performance/pk3bf3p8",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-uqm9idqz-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 9o0mrk1y; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 130.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 130.6 14.3 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uqm9idqz",
   "source_url": "https://omlx.ai/benchmarks/performance/uqm9idqz",
   "tg_tokens": null,
   "tps": 14.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-uqm9idqz-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 9o0mrk1y; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 129.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 129.2 14.1 19.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uqm9idqz",
   "source_url": "https://omlx.ai/benchmarks/performance/uqm9idqz",
   "tg_tokens": null,
   "tps": 14.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-j62waq2j-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: gt1o85z8; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 101.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 101.0 18.4 20.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j62waq2j",
   "source_url": "https://omlx.ai/benchmarks/performance/j62waq2j",
   "tg_tokens": null,
   "tps": 18.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-j62waq2j-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: gt1o85z8; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 92.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 92.9 15.2 22.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j62waq2j",
   "source_url": "https://omlx.ai/benchmarks/performance/j62waq2j",
   "tg_tokens": null,
   "tps": 15.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-dor8ymtr-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: 15aa0jex, bcxzbx01, f3m4l5td, nus6rnal, t3ene283; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 468.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 468.6 27.4 70.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dor8ymtr",
   "source_url": "https://omlx.ai/benchmarks/performance/dor8ymtr",
   "tg_tokens": null,
   "tps": 27.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-dor8ymtr-ctx8192",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.7 GB at this context; duplicate traces of the same session: 15aa0jex, bcxzbx01, f3m4l5td, nus6rnal, t3ene283; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 711.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 711.3 44.6 71.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dor8ymtr",
   "source_url": "https://omlx.ai/benchmarks/performance/dor8ymtr",
   "tg_tokens": null,
   "tps": 44.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-dor8ymtr-ctx16384",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.0 GB at this context; duplicate traces of the same session: 15aa0jex, bcxzbx01, f3m4l5td, nus6rnal, t3ene283; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 690.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 690.0 27.6 72.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dor8ymtr",
   "source_url": "https://omlx.ai/benchmarks/performance/dor8ymtr",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-dor8ymtr-ctx32768",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 72.5 GB at this context; duplicate traces of the same session: 15aa0jex, bcxzbx01, f3m4l5td, nus6rnal, t3ene283; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 677.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 677.3 44.3 72.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dor8ymtr",
   "source_url": "https://omlx.ai/benchmarks/performance/dor8ymtr",
   "tg_tokens": null,
   "tps": 44.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-dor8ymtr-ctx65536",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.7 GB at this context; duplicate traces of the same session: 15aa0jex, bcxzbx01, f3m4l5td, nus6rnal, t3ene283; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 674.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 674.7 52.3 73.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dor8ymtr",
   "source_url": "https://omlx.ai/benchmarks/performance/dor8ymtr",
   "tg_tokens": null,
   "tps": 52.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-dor8ymtr-ctx131072",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 76.9 GB at this context; duplicate traces of the same session: 15aa0jex, bcxzbx01, f3m4l5td, nus6rnal, t3ene283; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 671.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-mtp on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 671.4 41.9 76.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dor8ymtr",
   "source_url": "https://omlx.ai/benchmarks/performance/dor8ymtr",
   "tg_tokens": null,
   "tps": 41.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-ef78hit2-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 290.5 GB at this context; duplicate traces of the same session: kvz6s5d5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 504.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 504.5 26.4 290.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ef78hit2",
   "source_url": "https://omlx.ai/benchmarks/performance/ef78hit2",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-ef78hit2-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 290.5 GB at this context; duplicate traces of the same session: kvz6s5d5; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 696.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 16k 696.4 37.7 290.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ef78hit2",
   "source_url": "https://omlx.ai/benchmarks/performance/ef78hit2",
   "tg_tokens": null,
   "tps": 37.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-8siqgalq-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 290.5 GB at this context; duplicate traces of the same session: 2ztkpa3o; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 320.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 320.5 28.0 290.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8siqgalq",
   "source_url": "https://omlx.ai/benchmarks/performance/8siqgalq",
   "tg_tokens": null,
   "tps": 28.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-8siqgalq-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 290.5 GB at this context; duplicate traces of the same session: 2ztkpa3o; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 386.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 16k 386.5 41.2 290.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8siqgalq",
   "source_url": "https://omlx.ai/benchmarks/performance/8siqgalq",
   "tg_tokens": null,
   "tps": 41.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-kaev8lpq-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 403.2 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 549.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 549.2 27.3 403.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kaev8lpq",
   "source_url": "https://omlx.ai/benchmarks/performance/kaev8lpq",
   "tg_tokens": null,
   "tps": 27.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-kaev8lpq-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 403.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 712.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 712.3 32.0 403.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kaev8lpq",
   "source_url": "https://omlx.ai/benchmarks/performance/kaev8lpq",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-kt3sbxaw-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 207.3 GB at this context; duplicate traces of the same session: g0xdctjw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 375.3,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 1k 375.3 18.0 207.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kt3sbxaw",
   "source_url": "https://omlx.ai/benchmarks/performance/kt3sbxaw",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-kt3sbxaw-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 207.4 GB at this context; duplicate traces of the same session: g0xdctjw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 500.3,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 4k 500.3 20.9 207.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kt3sbxaw",
   "source_url": "https://omlx.ai/benchmarks/performance/kt3sbxaw",
   "tg_tokens": null,
   "tps": 20.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-kt3sbxaw-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 207.6 GB at this context; duplicate traces of the same session: g0xdctjw; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 515.3,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 16k 515.3 20.6 207.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kt3sbxaw",
   "source_url": "https://omlx.ai/benchmarks/performance/kt3sbxaw",
   "tg_tokens": null,
   "tps": 20.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4r22hhrc-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 207.3 GB at this context; duplicate traces of the same session: dso7zp3t; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 392.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 1k 392.0 26.4 207.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4r22hhrc",
   "source_url": "https://omlx.ai/benchmarks/performance/4r22hhrc",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-4r22hhrc-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 207.4 GB at this context; duplicate traces of the same session: dso7zp3t; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 507.5,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 4k 507.5 31.8 207.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4r22hhrc",
   "source_url": "https://omlx.ai/benchmarks/performance/4r22hhrc",
   "tg_tokens": null,
   "tps": 31.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4r22hhrc-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 207.6 GB at this context; duplicate traces of the same session: dso7zp3t; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 524.9,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-oQ2.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 16k 524.9 39.4 207.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4r22hhrc",
   "source_url": "https://omlx.ai/benchmarks/performance/4r22hhrc",
   "tg_tokens": null,
   "tps": 39.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mvyuc5rq-ctx1024",
   "model": "Jundot:Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 41.6 GB at this context; duplicate traces of the same session: 0g4hfhy1, 4z2wpv1f, 6jrkr297; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 172.7,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Jundot--Qwen3.8-Flash-Next-oQ2.7e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 2bit; 1k 172.7 34.5 41.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mvyuc5rq",
   "source_url": "https://omlx.ai/benchmarks/performance/mvyuc5rq",
   "tg_tokens": null,
   "tps": 34.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mvyuc5rq-ctx4096",
   "model": "Jundot:Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 42.6 GB at this context; duplicate traces of the same session: 0g4hfhy1, 4z2wpv1f, 6jrkr297; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 829.9,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Jundot--Qwen3.8-Flash-Next-oQ2.7e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 2bit; 4k 829.9 35.7 42.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mvyuc5rq",
   "source_url": "https://omlx.ai/benchmarks/performance/mvyuc5rq",
   "tg_tokens": null,
   "tps": 35.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mvyuc5rq-ctx8192",
   "model": "Jundot:Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 44.5 GB at this context; duplicate traces of the same session: 0g4hfhy1, 4z2wpv1f, 6jrkr297; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 907.1,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Jundot--Qwen3.8-Flash-Next-oQ2.7e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 2bit; 8k 907.1 34.7 44.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mvyuc5rq",
   "source_url": "https://omlx.ai/benchmarks/performance/mvyuc5rq",
   "tg_tokens": null,
   "tps": 34.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mvyuc5rq-ctx16384",
   "model": "Jundot:Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 44.2 GB at this context; duplicate traces of the same session: 0g4hfhy1, 4z2wpv1f, 6jrkr297; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1019.0,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Jundot--Qwen3.8-Flash-Next-oQ2.7e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 2bit; 16k 1,019 46.7 44.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mvyuc5rq",
   "source_url": "https://omlx.ai/benchmarks/performance/mvyuc5rq",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mvyuc5rq-ctx65536",
   "model": "Jundot:Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 45.8 GB at this context; duplicate traces of the same session: 0g4hfhy1, 4z2wpv1f, 6jrkr297; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1021.0,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Jundot--Qwen3.8-Flash-Next-oQ2.7e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 2bit; 64k 1,021 28.6 45.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mvyuc5rq",
   "source_url": "https://omlx.ai/benchmarks/performance/mvyuc5rq",
   "tg_tokens": null,
   "tps": 28.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jg6l6c5z-ctx1024",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 22.1 GB; TTFT 249 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4112.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "MiniCPM-V-4.7-35B-A3B-mxfp4 on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 4,112 PP tok/s 99.4 TG tok/s 249 TTFT (ms) 22.1 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jg6l6c5z",
   "source_url": "https://omlx.ai/benchmarks/performance/jg6l6c5z",
   "tg_tokens": null,
   "tps": 99.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-x0wth64x-ctx1024",
   "model": "Qwen3.5-0.8B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 2.5 GB; TTFT 99 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 10357.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Qwen--Qwen3.5-0.8B-MTP on M5 Ultra (64c); M5 Ultra (64c) \u00b7 96 GB \u00b7 unknown; 1k tokens 10,357 PP tok/s 71.3 TG tok/s 99 TTFT (ms) 2.5 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x0wth64x",
   "source_url": "https://omlx.ai/benchmarks/performance/x0wth64x",
   "tg_tokens": null,
   "tps": 71.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-v1gn7du8-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 6p12ctzg, jxxjy4bg, y23o5okx; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2775.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,775 110.2 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v1gn7du8",
   "source_url": "https://omlx.ai/benchmarks/performance/v1gn7du8",
   "tg_tokens": null,
   "tps": 110.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-v1gn7du8-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 6p12ctzg, jxxjy4bg, y23o5okx; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3474.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,474 107.7 20.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v1gn7du8",
   "source_url": "https://omlx.ai/benchmarks/performance/v1gn7du8",
   "tg_tokens": null,
   "tps": 107.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-v1gn7du8-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 6p12ctzg, jxxjy4bg, y23o5okx; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3331.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 3,331 103.5 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v1gn7du8",
   "source_url": "https://omlx.ai/benchmarks/performance/v1gn7du8",
   "tg_tokens": null,
   "tps": 103.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-v1gn7du8-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: 6p12ctzg, jxxjy4bg, y23o5okx; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3069.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 3,069 97.5 20.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v1gn7du8",
   "source_url": "https://omlx.ai/benchmarks/performance/v1gn7du8",
   "tg_tokens": null,
   "tps": 97.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-v1gn7du8-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 6p12ctzg, jxxjy4bg, y23o5okx; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2585.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 2,585 85.2 20.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v1gn7du8",
   "source_url": "https://omlx.ai/benchmarks/performance/v1gn7du8",
   "tg_tokens": null,
   "tps": 85.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-v1gn7du8-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 6p12ctzg, jxxjy4bg, y23o5okx; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1903.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 1,903 64.5 21.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v1gn7du8",
   "source_url": "https://omlx.ai/benchmarks/performance/v1gn7du8",
   "tg_tokens": null,
   "tps": 64.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ljqwzzfz-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 2dawwbt8, 69o5xdf1, flhq8nbx, jdual8og; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2778.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,778 110.9 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ljqwzzfz",
   "source_url": "https://omlx.ai/benchmarks/performance/ljqwzzfz",
   "tg_tokens": null,
   "tps": 110.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ljqwzzfz-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 2dawwbt8, 69o5xdf1, flhq8nbx, jdual8og; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3439.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,439 108.5 20.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ljqwzzfz",
   "source_url": "https://omlx.ai/benchmarks/performance/ljqwzzfz",
   "tg_tokens": null,
   "tps": 108.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ljqwzzfz-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 2dawwbt8, 69o5xdf1, flhq8nbx, jdual8og; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3346.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 3,346 103.9 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ljqwzzfz",
   "source_url": "https://omlx.ai/benchmarks/performance/ljqwzzfz",
   "tg_tokens": null,
   "tps": 103.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ljqwzzfz-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: 2dawwbt8, 69o5xdf1, flhq8nbx, jdual8og; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3066.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 3,066 97.8 20.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ljqwzzfz",
   "source_url": "https://omlx.ai/benchmarks/performance/ljqwzzfz",
   "tg_tokens": null,
   "tps": 97.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ljqwzzfz-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 2dawwbt8, 69o5xdf1, flhq8nbx, jdual8og; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2607.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 2,607 85.9 20.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ljqwzzfz",
   "source_url": "https://omlx.ai/benchmarks/performance/ljqwzzfz",
   "tg_tokens": null,
   "tps": 85.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-9hpmmeei-ctx199680",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 195k context; peak memory 23.8 GB; TTFT 210394 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 950.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 195k tokens 950.6 PP tok/s 39.0 TG tok/s 210394 TTFT (ms) 23.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9hpmmeei",
   "source_url": "https://omlx.ai/benchmarks/performance/9hpmmeei",
   "tg_tokens": null,
   "tps": 39.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-mj3dj6nq-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: agemipe4, clbksj8f, o7760z4l, rk4ksi6j; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 233.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 1k 233.7 33.5 14.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mj3dj6nq",
   "source_url": "https://omlx.ai/benchmarks/performance/mj3dj6nq",
   "tg_tokens": null,
   "tps": 33.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-mj3dj6nq-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 15.5 GB at this context; duplicate traces of the same session: agemipe4, clbksj8f, o7760z4l, rk4ksi6j; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 228.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 4k 228.7 30.6 15.5 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mj3dj6nq",
   "source_url": "https://omlx.ai/benchmarks/performance/mj3dj6nq",
   "tg_tokens": null,
   "tps": 30.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-mj3dj6nq-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 15.6 GB at this context; duplicate traces of the same session: agemipe4, clbksj8f, o7760z4l, rk4ksi6j; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 202.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 8k 202.8 29.7 15.6 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mj3dj6nq",
   "source_url": "https://omlx.ai/benchmarks/performance/mj3dj6nq",
   "tg_tokens": null,
   "tps": 29.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-mj3dj6nq-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: agemipe4, clbksj8f, o7760z4l, rk4ksi6j; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 186.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 16k 186.9 26.4 15.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mj3dj6nq",
   "source_url": "https://omlx.ai/benchmarks/performance/mj3dj6nq",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-mj3dj6nq-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 16.1 GB at this context; duplicate traces of the same session: agemipe4, clbksj8f, o7760z4l, rk4ksi6j; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 176.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 32k 176.6 24.5 16.1 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mj3dj6nq",
   "source_url": "https://omlx.ai/benchmarks/performance/mj3dj6nq",
   "tg_tokens": null,
   "tps": 24.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-dm83i9q8-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 19.2 GB; TTFT 656 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1562.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 1k tokens 1,562 PP tok/s 69.3 TG tok/s 656 TTFT (ms) 19.2 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dm83i9q8",
   "source_url": "https://omlx.ai/benchmarks/performance/dm83i9q8",
   "tg_tokens": null,
   "tps": 69.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-0zshv561-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MLX fp16 MTP build; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: hee9zwla; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 123.9,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-Uncensored-MLX-oQ5e-fp16-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 5bit; 1k 123.9 25.2 24.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0zshv561",
   "source_url": "https://omlx.ai/benchmarks/performance/0zshv561",
   "tg_tokens": null,
   "tps": 25.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-0zshv561-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MLX fp16 MTP build; oMLX community benchmark trace; peak memory 27.3 GB at this context; duplicate traces of the same session: hee9zwla; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 124.4,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-27B-Uncensored-MLX-oQ5e-fp16-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 5bit; 4k 124.4 21.7 27.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0zshv561",
   "source_url": "https://omlx.ai/benchmarks/performance/0zshv561",
   "tg_tokens": null,
   "tps": 21.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-608f6ofz-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: pnc3o1qu; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 319.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 4k 319.6 40.0 23.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/608f6ofz",
   "source_url": "https://omlx.ai/benchmarks/performance/608f6ofz",
   "tg_tokens": null,
   "tps": 40.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-608f6ofz-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 26.4 GB at this context; duplicate traces of the same session: pnc3o1qu; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 277.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 32k 277.2 61.2 26.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/608f6ofz",
   "source_url": "https://omlx.ai/benchmarks/performance/608f6ofz",
   "tg_tokens": null,
   "tps": 61.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-zv2zanpt-ctx65536",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace, single run at 64k context; peak memory 28.5 GB; TTFT 279926 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 234.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 64k tokens 234.1 PP tok/s 43.5 TG tok/s 279926 TTFT (ms) 28.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zv2zanpt",
   "source_url": "https://omlx.ai/benchmarks/performance/zv2zanpt",
   "tg_tokens": null,
   "tps": 43.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-7il5cnvx-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: jigp4l0b; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 470.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 470.1 38.2 17.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7il5cnvx",
   "source_url": "https://omlx.ai/benchmarks/performance/7il5cnvx",
   "tg_tokens": null,
   "tps": 38.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-7il5cnvx-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: jigp4l0b; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 472.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 472.3 34.3 18.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7il5cnvx",
   "source_url": "https://omlx.ai/benchmarks/performance/7il5cnvx",
   "tg_tokens": null,
   "tps": 34.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-53tpr9hu-ctx4096",
   "model": "Qwen3.8-35B-A3B-Distill-Heretic",
   "notes": "MTP speculative decoding enabled; fp16 build; oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: y94z2akm; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1356.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-35B-A3B-Distill-Heretic-oQ4-fp16-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,356 59.2 22.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/53tpr9hu",
   "source_url": "https://omlx.ai/benchmarks/performance/53tpr9hu",
   "tg_tokens": null,
   "tps": 59.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-53tpr9hu-ctx16384",
   "model": "Qwen3.8-35B-A3B-Distill-Heretic",
   "notes": "MTP speculative decoding enabled; fp16 build; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: y94z2akm; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1847.0,
   "provenance": "community",
   "quant": "oQ4",
   "quote": "Qwen3.8-35B-A3B-Distill-Heretic-oQ4-fp16-mtp on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 1,847 43.4 23.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/53tpr9hu",
   "source_url": "https://omlx.ai/benchmarks/performance/53tpr9hu",
   "tg_tokens": null,
   "tps": 43.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-a4oi4svd-ctx4096",
   "model": "Qwen3.8-35B-A3B-Distill",
   "notes": "APEX MiniPlus Abliterated MLX build; oMLX community benchmark trace, single run at 4k context; peak memory 13.2 GB; TTFT 10440 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 392.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-35B-A3B-Distill-MLX-APEX-MiniPlus-Abliterated on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k tokens 392.3 PP tok/s 14.5 TG tok/s 10440 TTFT (ms) 13.2 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a4oi4svd",
   "source_url": "https://omlx.ai/benchmarks/performance/a4oi4svd",
   "tg_tokens": null,
   "tps": 14.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-oigj3gp0-ctx4096",
   "model": "gemma-4-12b-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.8 GB at this context; duplicate traces of the same session: apgbkm44; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 135.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-12B-it-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 135.2 13.2 7.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oigj3gp0",
   "source_url": "https://omlx.ai/benchmarks/performance/oigj3gp0",
   "tg_tokens": null,
   "tps": 13.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-oigj3gp0-ctx16384",
   "model": "gemma-4-12b-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 8.3 GB at this context; duplicate traces of the same session: apgbkm44; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 122.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-12B-it-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 16k 122.1 12.7 8.3 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oigj3gp0",
   "source_url": "https://omlx.ai/benchmarks/performance/oigj3gp0",
   "tg_tokens": null,
   "tps": 12.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-9tjadopc-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.6 GB at this context; duplicate traces of the same session: smopz3g2; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 388.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 388.9 25.3 16.6 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9tjadopc",
   "source_url": "https://omlx.ai/benchmarks/performance/9tjadopc",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-9tjadopc-ctx16384",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: smopz3g2; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 348.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-26B-A4B-it-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 16k 348.9 22.7 17.2 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9tjadopc",
   "source_url": "https://omlx.ai/benchmarks/performance/9tjadopc",
   "tg_tokens": null,
   "tps": 22.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2hnbt41l-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.5 GB at this context; duplicate traces of the same session: vxra8o9e, ypqy2sq4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 112.5,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 112.5 13.4 96.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2hnbt41l",
   "source_url": "https://omlx.ai/benchmarks/performance/2hnbt41l",
   "tg_tokens": null,
   "tps": 13.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2hnbt41l-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.5 GB at this context; duplicate traces of the same session: vxra8o9e, ypqy2sq4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 538.9,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 538.9 13.5 99.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2hnbt41l",
   "source_url": "https://omlx.ai/benchmarks/performance/2hnbt41l",
   "tg_tokens": null,
   "tps": 13.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-2hnbt41l-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.7 GB at this context; duplicate traces of the same session: vxra8o9e, ypqy2sq4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 612.2,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 612.2 15.9 99.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2hnbt41l",
   "source_url": "https://omlx.ai/benchmarks/performance/2hnbt41l",
   "tg_tokens": null,
   "tps": 15.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-j9lc2gfa-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 403.2 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 549.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 549.2 27.3 403.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j9lc2gfa",
   "source_url": "https://omlx.ai/benchmarks/performance/j9lc2gfa",
   "tg_tokens": null,
   "tps": 27.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-j9lc2gfa-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 403.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 712.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 712.3 32.0 403.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j9lc2gfa",
   "source_url": "https://omlx.ai/benchmarks/performance/j9lc2gfa",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-if08wss2-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 288.7 GB at this context; duplicate traces of the same session: 6g85ztad; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 502.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 502.5 30.7 288.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/if08wss2",
   "source_url": "https://omlx.ai/benchmarks/performance/if08wss2",
   "tg_tokens": null,
   "tps": 30.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-if08wss2-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 289.3 GB at this context; duplicate traces of the same session: 6g85ztad; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 683.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 16k 683.0 37.8 289.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/if08wss2",
   "source_url": "https://omlx.ai/benchmarks/performance/if08wss2",
   "tg_tokens": null,
   "tps": 37.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-lnwki81s-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 288.7 GB at this context; duplicate traces of the same session: tief1n2c; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 448.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 448.8 27.5 288.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lnwki81s",
   "source_url": "https://omlx.ai/benchmarks/performance/lnwki81s",
   "tg_tokens": null,
   "tps": 27.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-lnwki81s-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 289.3 GB at this context; duplicate traces of the same session: tief1n2c; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 575.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 16k 575.1 38.7 289.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lnwki81s",
   "source_url": "https://omlx.ai/benchmarks/performance/lnwki81s",
   "tg_tokens": null,
   "tps": 38.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-hxy35oko-ctx16384",
   "model": "Huihui-Qwen3.8-27B-abliterated",
   "notes": "MLX build; oMLX community benchmark trace, single run at 16k context; peak memory 24.0 GB; TTFT 37538 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 436.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Huihui-Qwen3.8-27B-abliterated-MLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 6bit; 16k tokens 436.5 PP tok/s 11.4 TG tok/s 37538 TTFT (ms) 24.0 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hxy35oko",
   "source_url": "https://omlx.ai/benchmarks/performance/hxy35oko",
   "tg_tokens": null,
   "tps": 11.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fy03bhdc-ctx1024",
   "model": "Jundot:Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 41.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 172.7,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Jundot--Qwen3.8-Flash-Next-oQ2.7e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 2bit; 1k 172.7 34.5 41.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fy03bhdc",
   "source_url": "https://omlx.ai/benchmarks/performance/fy03bhdc",
   "tg_tokens": null,
   "tps": 34.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fy03bhdc-ctx4096",
   "model": "Jundot:Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 42.6 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 829.9,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Jundot--Qwen3.8-Flash-Next-oQ2.7e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 2bit; 4k 829.9 35.7 42.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fy03bhdc",
   "source_url": "https://omlx.ai/benchmarks/performance/fy03bhdc",
   "tg_tokens": null,
   "tps": 35.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fy03bhdc-ctx8192",
   "model": "Jundot:Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 44.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 907.1,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Jundot--Qwen3.8-Flash-Next-oQ2.7e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 2bit; 8k 907.1 34.7 44.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fy03bhdc",
   "source_url": "https://omlx.ai/benchmarks/performance/fy03bhdc",
   "tg_tokens": null,
   "tps": 34.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fy03bhdc-ctx16384",
   "model": "Jundot:Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 44.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1019.0,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Jundot--Qwen3.8-Flash-Next-oQ2.7e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 2bit; 16k 1,019 46.7 44.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fy03bhdc",
   "source_url": "https://omlx.ai/benchmarks/performance/fy03bhdc",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fy03bhdc-ctx65536",
   "model": "Jundot:Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 45.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1021.0,
   "provenance": "community",
   "quant": "oQ2.7e",
   "quote": "Jundot--Qwen3.8-Flash-Next-oQ2.7e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 2bit; 64k 1,021 28.6 45.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fy03bhdc",
   "source_url": "https://omlx.ai/benchmarks/performance/fy03bhdc",
   "tg_tokens": null,
   "tps": 28.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-sphoseaa-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX Serve build; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: b5xjcfwi, cyj6d0hr, einwm71w, nvr5rakf; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 887.6,
   "provenance": "community",
   "quant": "Mixed-4-8",
   "quote": "LFM2.5-24B-A2B-MLX-Serve-mixed-4-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 887.6 81.7 13.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sphoseaa",
   "source_url": "https://omlx.ai/benchmarks/performance/sphoseaa",
   "tg_tokens": null,
   "tps": 81.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-sphoseaa-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX Serve build; oMLX community benchmark trace; peak memory 14.0 GB at this context; duplicate traces of the same session: b5xjcfwi, cyj6d0hr, einwm71w, nvr5rakf; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1024.0,
   "provenance": "community",
   "quant": "Mixed-4-8",
   "quote": "LFM2.5-24B-A2B-MLX-Serve-mixed-4-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 1,024 77.6 14.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sphoseaa",
   "source_url": "https://omlx.ai/benchmarks/performance/sphoseaa",
   "tg_tokens": null,
   "tps": 77.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-sphoseaa-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX Serve build; oMLX community benchmark trace; peak memory 14.1 GB at this context; duplicate traces of the same session: b5xjcfwi, cyj6d0hr, einwm71w, nvr5rakf; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1006.0,
   "provenance": "community",
   "quant": "Mixed-4-8",
   "quote": "LFM2.5-24B-A2B-MLX-Serve-mixed-4-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 1,006 76.4 14.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sphoseaa",
   "source_url": "https://omlx.ai/benchmarks/performance/sphoseaa",
   "tg_tokens": null,
   "tps": 76.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-sphoseaa-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX Serve build; oMLX community benchmark trace; peak memory 14.2 GB at this context; duplicate traces of the same session: b5xjcfwi, cyj6d0hr, einwm71w, nvr5rakf; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 960.5,
   "provenance": "community",
   "quant": "Mixed-4-8",
   "quote": "LFM2.5-24B-A2B-MLX-Serve-mixed-4-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 960.5 70.5 14.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sphoseaa",
   "source_url": "https://omlx.ai/benchmarks/performance/sphoseaa",
   "tg_tokens": null,
   "tps": 70.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-sphoseaa-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX Serve build; oMLX community benchmark trace; peak memory 14.5 GB at this context; duplicate traces of the same session: b5xjcfwi, cyj6d0hr, einwm71w, nvr5rakf; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 873.7,
   "provenance": "community",
   "quant": "Mixed-4-8",
   "quote": "LFM2.5-24B-A2B-MLX-Serve-mixed-4-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 32k 873.7 57.6 14.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sphoseaa",
   "source_url": "https://omlx.ai/benchmarks/performance/sphoseaa",
   "tg_tokens": null,
   "tps": 57.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-sphoseaa-ctx65536",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX Serve build; oMLX community benchmark trace; peak memory 15.2 GB at this context; duplicate traces of the same session: b5xjcfwi, cyj6d0hr, einwm71w, nvr5rakf; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 739.9,
   "provenance": "community",
   "quant": "Mixed-4-8",
   "quote": "LFM2.5-24B-A2B-MLX-Serve-mixed-4-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 64k 739.9 45.5 15.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sphoseaa",
   "source_url": "https://omlx.ai/benchmarks/performance/sphoseaa",
   "tg_tokens": null,
   "tps": 45.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-sphoseaa-ctx131072",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX Serve build; oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: b5xjcfwi, cyj6d0hr, einwm71w, nvr5rakf; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 565.9,
   "provenance": "community",
   "quant": "Mixed-4-8",
   "quote": "LFM2.5-24B-A2B-MLX-Serve-mixed-4-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 128k 565.9 32.0 16.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sphoseaa",
   "source_url": "https://omlx.ai/benchmarks/performance/sphoseaa",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-83vxjpfl-ctx65536",
   "model": "LFM2.5-24B-A2B-Uncensored-Heretic",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 14.8 GB at this context; duplicate traces of the same session: fwnrn0yu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1200.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-Uncensored-Heretic-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 1,200 76.3 14.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/83vxjpfl",
   "source_url": "https://omlx.ai/benchmarks/performance/83vxjpfl",
   "tg_tokens": null,
   "tps": 76.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-83vxjpfl-ctx131072",
   "model": "LFM2.5-24B-A2B-Uncensored-Heretic",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: fwnrn0yu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 485.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-Uncensored-Heretic-MLX-4bit on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 485.8 17.3 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/83vxjpfl",
   "source_url": "https://omlx.ai/benchmarks/performance/83vxjpfl",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-aaaq9l56-ctx32768",
   "model": "Qwen3-0.6B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 4.5 GB; TTFT 7080 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4629.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-0.6B-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 32k tokens 4,629 PP tok/s 118.6 TG tok/s 7080 TTFT (ms) 4.5 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aaaq9l56",
   "source_url": "https://omlx.ai/benchmarks/performance/aaaq9l56",
   "tg_tokens": null,
   "tps": 118.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-oqvagky9-ctx199680",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 195k context; peak memory 23.8 GB; TTFT 209926 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 952.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 195k tokens 952.7 PP tok/s 40.4 TG tok/s 209926 TTFT (ms) 23.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oqvagky9",
   "source_url": "https://omlx.ai/benchmarks/performance/oqvagky9",
   "tg_tokens": null,
   "tps": 40.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-i2pa8bsz-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 797y29mx, d4fk6dvw, deyde62m, ne6x4880, t1c76afh, y83tkivv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2758.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,758 109.1 19.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i2pa8bsz",
   "source_url": "https://omlx.ai/benchmarks/performance/i2pa8bsz",
   "tg_tokens": null,
   "tps": 109.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-i2pa8bsz-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 797y29mx, d4fk6dvw, deyde62m, ne6x4880, t1c76afh, y83tkivv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3483.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,483 108.1 20.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i2pa8bsz",
   "source_url": "https://omlx.ai/benchmarks/performance/i2pa8bsz",
   "tg_tokens": null,
   "tps": 108.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-i2pa8bsz-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 797y29mx, d4fk6dvw, deyde62m, ne6x4880, t1c76afh, y83tkivv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3336.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 3,336 103.5 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i2pa8bsz",
   "source_url": "https://omlx.ai/benchmarks/performance/i2pa8bsz",
   "tg_tokens": null,
   "tps": 103.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-i2pa8bsz-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: 797y29mx, d4fk6dvw, deyde62m, ne6x4880, t1c76afh, y83tkivv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3060.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 3,060 97.1 20.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i2pa8bsz",
   "source_url": "https://omlx.ai/benchmarks/performance/i2pa8bsz",
   "tg_tokens": null,
   "tps": 97.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-i2pa8bsz-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 797y29mx, d4fk6dvw, deyde62m, ne6x4880, t1c76afh, y83tkivv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2598.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 32k 2,598 85.4 20.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i2pa8bsz",
   "source_url": "https://omlx.ai/benchmarks/performance/i2pa8bsz",
   "tg_tokens": null,
   "tps": 85.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-i2pa8bsz-ctx65536",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: 797y29mx, d4fk6dvw, deyde62m, ne6x4880, t1c76afh, y83tkivv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1934.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k 1,934 66.8 21.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i2pa8bsz",
   "source_url": "https://omlx.ai/benchmarks/performance/i2pa8bsz",
   "tg_tokens": null,
   "tps": 66.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-i2pa8bsz-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: 797y29mx, d4fk6dvw, deyde62m, ne6x4880, t1c76afh, y83tkivv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1282.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 128k 1,282 50.8 22.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i2pa8bsz",
   "source_url": "https://omlx.ai/benchmarks/performance/i2pa8bsz",
   "tg_tokens": null,
   "tps": 50.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-nk3ug5hf-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 23.6 GB; TTFT 55811 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 587.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 32k tokens 587.1 PP tok/s 56.7 TG tok/s 55811 TTFT (ms) 23.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nk3ug5hf",
   "source_url": "https://omlx.ai/benchmarks/performance/nk3ug5hf",
   "tg_tokens": null,
   "tps": 56.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-q3lv3o2j-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 23.6 GB; TTFT 55773 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 587.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 32k tokens 587.5 PP tok/s 66.3 TG tok/s 55773 TTFT (ms) 23.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q3lv3o2j",
   "source_url": "https://omlx.ai/benchmarks/performance/q3lv3o2j",
   "tg_tokens": null,
   "tps": 66.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-ny21ufcj-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.8 GB at this context; duplicate traces of the same session: x1zth8pt; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 517.6,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.6-35B-A3B-oQ3e-mtp on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 3bit; 4k 517.6 41.6 16.8 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ny21ufcj",
   "source_url": "https://omlx.ai/benchmarks/performance/ny21ufcj",
   "tg_tokens": null,
   "tps": 41.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-ny21ufcj-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: x1zth8pt; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 420.3,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.6-35B-A3B-oQ3e-mtp on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 3bit; 16k 420.3 36.9 17.1 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ny21ufcj",
   "source_url": "https://omlx.ai/benchmarks/performance/ny21ufcj",
   "tg_tokens": null,
   "tps": 36.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-5s39j0tb-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: 62zaxfgs; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 492.3,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.6-35B-A3B-oQ3e-mtp on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 3bit; 4k 492.3 48.0 18.3 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5s39j0tb",
   "source_url": "https://omlx.ai/benchmarks/performance/5s39j0tb",
   "tg_tokens": null,
   "tps": 48.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-5s39j0tb-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: 62zaxfgs; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 399.2,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.6-35B-A3B-oQ3e-mtp on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 3bit; 16k 399.2 52.6 18.7 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5s39j0tb",
   "source_url": "https://omlx.ai/benchmarks/performance/5s39j0tb",
   "tg_tokens": null,
   "tps": 52.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-5hrmaunw-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: h6qux69q; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 132.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 132.1 14.6 17.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5hrmaunw",
   "source_url": "https://omlx.ai/benchmarks/performance/5hrmaunw",
   "tg_tokens": null,
   "tps": 14.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-5hrmaunw-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.0 GB at this context; duplicate traces of the same session: h6qux69q; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 124.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 124.1 14.1 18.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5hrmaunw",
   "source_url": "https://omlx.ai/benchmarks/performance/5hrmaunw",
   "tg_tokens": null,
   "tps": 14.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-9ypj5yap-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 25.7 GB at this context; duplicate traces of the same session: hsy7638t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 226.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 226.1 13.9 25.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ypj5yap",
   "source_url": "https://omlx.ai/benchmarks/performance/9ypj5yap",
   "tg_tokens": null,
   "tps": 13.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-9ypj5yap-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 26.9 GB at this context; duplicate traces of the same session: hsy7638t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 209.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-MLX-6bit on M3 Max (40c); M3 Max (40c) \u00b7 64 GB \u00b7 6bit; 16k 209.0 14.3 26.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9ypj5yap",
   "source_url": "https://omlx.ai/benchmarks/performance/9ypj5yap",
   "tg_tokens": null,
   "tps": 14.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-iprxfpzq-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 51.8 GB at this context; duplicate traces of the same session: 0dwl3ehi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 226.1,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.8-27B-bf16 on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 bf16; 1k 226.1 6.8 51.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iprxfpzq",
   "source_url": "https://omlx.ai/benchmarks/performance/iprxfpzq",
   "tg_tokens": null,
   "tps": 6.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-iprxfpzq-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 55.2 GB at this context; duplicate traces of the same session: 0dwl3ehi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 211.7,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.8-27B-bf16 on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 bf16; 4k 211.7 6.5 55.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iprxfpzq",
   "source_url": "https://omlx.ai/benchmarks/performance/iprxfpzq",
   "tg_tokens": null,
   "tps": 6.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-22zmjl9v-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 55.5 GB at this context; duplicate traces of the same session: cbxdt31v; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 200.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.8-27B-bf16 on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 bf16; 1k 200.0 15.4 55.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/22zmjl9v",
   "source_url": "https://omlx.ai/benchmarks/performance/22zmjl9v",
   "tg_tokens": null,
   "tps": 15.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-22zmjl9v-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 57.2 GB at this context; duplicate traces of the same session: cbxdt31v; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 213.6,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Qwen3.8-27B-bf16 on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 bf16; 4k 213.6 15.0 57.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/22zmjl9v",
   "source_url": "https://omlx.ai/benchmarks/performance/22zmjl9v",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-5wv5pzci-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 40.0 GB at this context; duplicate traces of the same session: uba31w30, v4vklwyl; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 227.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 227.6 30.8 40.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5wv5pzci",
   "source_url": "https://omlx.ai/benchmarks/performance/5wv5pzci",
   "tg_tokens": null,
   "tps": 30.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-5wv5pzci-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 40.5 GB at this context; duplicate traces of the same session: uba31w30, v4vklwyl; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 245.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 245.3 31.1 40.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5wv5pzci",
   "source_url": "https://omlx.ai/benchmarks/performance/5wv5pzci",
   "tg_tokens": null,
   "tps": 31.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-5wv5pzci-ctx8192",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 40.6 GB at this context; duplicate traces of the same session: uba31w30, v4vklwyl; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 336.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 336.2 31.2 40.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5wv5pzci",
   "source_url": "https://omlx.ai/benchmarks/performance/5wv5pzci",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-4tofr9tl-ctx4096",
   "model": "gemma-4-12b-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 8.6 GB at this context; duplicate traces of the same session: 0xac1cx3; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 137.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-12B-it-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 137.1 12.5 8.6 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4tofr9tl",
   "source_url": "https://omlx.ai/benchmarks/performance/4tofr9tl",
   "tg_tokens": null,
   "tps": 12.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-4tofr9tl-ctx16384",
   "model": "gemma-4-12b-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 9.2 GB at this context; duplicate traces of the same session: 0xac1cx3; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 123.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-12B-it-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 16k 123.2 15.8 9.2 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4tofr9tl",
   "source_url": "https://omlx.ai/benchmarks/performance/4tofr9tl",
   "tg_tokens": null,
   "tps": 15.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-3p2ghym3-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 97.3 GB at this context; duplicate traces of the same session: 11l05wjp, 43d2cutu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 108.3,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 108.3 12.2 97.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3p2ghym3",
   "source_url": "https://omlx.ai/benchmarks/performance/3p2ghym3",
   "tg_tokens": null,
   "tps": 12.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3p2ghym3-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: 11l05wjp, 43d2cutu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 500.1,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 500.1 13.9 100.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3p2ghym3",
   "source_url": "https://omlx.ai/benchmarks/performance/3p2ghym3",
   "tg_tokens": null,
   "tps": 13.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-3p2ghym3-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.9 GB at this context; duplicate traces of the same session: 11l05wjp, 43d2cutu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 570.5,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 570.5 16.6 100.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3p2ghym3",
   "source_url": "https://omlx.ai/benchmarks/performance/3p2ghym3",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jmut20i9-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 98.5 GB at this context; duplicate traces of the same session: hmegcy9u, ias1u7h7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 107.5,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 107.5 10.7 98.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jmut20i9",
   "source_url": "https://omlx.ai/benchmarks/performance/jmut20i9",
   "tg_tokens": null,
   "tps": 10.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-jmut20i9-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.5 GB at this context; duplicate traces of the same session: hmegcy9u, ias1u7h7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 532.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 532.0 13.8 101.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jmut20i9",
   "source_url": "https://omlx.ai/benchmarks/performance/jmut20i9",
   "tg_tokens": null,
   "tps": 13.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jmut20i9-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.6 GB at this context; duplicate traces of the same session: hmegcy9u, ias1u7h7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 604.7,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 604.7 15.8 101.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jmut20i9",
   "source_url": "https://omlx.ai/benchmarks/performance/jmut20i9",
   "tg_tokens": null,
   "tps": 15.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-m2p8v3d6-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 288.8 GB at this context; duplicate traces of the same session: 21w45gdw, 43h4n4bs, 8tbkbwx8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 496.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 496.5 30.1 288.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m2p8v3d6",
   "source_url": "https://omlx.ai/benchmarks/performance/m2p8v3d6",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-m2p8v3d6-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 289.3 GB at this context; duplicate traces of the same session: 21w45gdw, 43h4n4bs, 8tbkbwx8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 693.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 693.3 31.6 289.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m2p8v3d6",
   "source_url": "https://omlx.ai/benchmarks/performance/m2p8v3d6",
   "tg_tokens": null,
   "tps": 31.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-m2p8v3d6-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 289.3 GB at this context; duplicate traces of the same session: 21w45gdw, 43h4n4bs, 8tbkbwx8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 730.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 8k 730.7 29.4 289.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m2p8v3d6",
   "source_url": "https://omlx.ai/benchmarks/performance/m2p8v3d6",
   "tg_tokens": null,
   "tps": 29.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-m2p8v3d6-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 289.3 GB at this context; duplicate traces of the same session: 21w45gdw, 43h4n4bs, 8tbkbwx8; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 748.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 16k 748.8 37.1 289.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m2p8v3d6",
   "source_url": "https://omlx.ai/benchmarks/performance/m2p8v3d6",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-vy6s6otk-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX Serve build; oMLX community benchmark trace; peak memory 13.7 GB at this context; duplicate traces of the same session: 578exclt; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 887.6,
   "provenance": "community",
   "quant": "Mixed-4-8",
   "quote": "LFM2.5-24B-A2B-MLX-Serve-mixed-4-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 887.6 81.7 13.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vy6s6otk",
   "source_url": "https://omlx.ai/benchmarks/performance/vy6s6otk",
   "tg_tokens": null,
   "tps": 81.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-vy6s6otk-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX Serve build; oMLX community benchmark trace; peak memory 14.0 GB at this context; duplicate traces of the same session: 578exclt; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1024.0,
   "provenance": "community",
   "quant": "Mixed-4-8",
   "quote": "LFM2.5-24B-A2B-MLX-Serve-mixed-4-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 1,024 77.6 14.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vy6s6otk",
   "source_url": "https://omlx.ai/benchmarks/performance/vy6s6otk",
   "tg_tokens": null,
   "tps": 77.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-vy6s6otk-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX Serve build; oMLX community benchmark trace; peak memory 14.1 GB at this context; duplicate traces of the same session: 578exclt; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1006.0,
   "provenance": "community",
   "quant": "Mixed-4-8",
   "quote": "LFM2.5-24B-A2B-MLX-Serve-mixed-4-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 8k 1,006 76.4 14.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vy6s6otk",
   "source_url": "https://omlx.ai/benchmarks/performance/vy6s6otk",
   "tg_tokens": null,
   "tps": 76.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-vy6s6otk-ctx16384",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX Serve build; oMLX community benchmark trace; peak memory 14.2 GB at this context; duplicate traces of the same session: 578exclt; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 960.5,
   "provenance": "community",
   "quant": "Mixed-4-8",
   "quote": "LFM2.5-24B-A2B-MLX-Serve-mixed-4-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 960.5 70.5 14.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vy6s6otk",
   "source_url": "https://omlx.ai/benchmarks/performance/vy6s6otk",
   "tg_tokens": null,
   "tps": 70.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-vy6s6otk-ctx32768",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX Serve build; oMLX community benchmark trace; peak memory 14.5 GB at this context; duplicate traces of the same session: 578exclt; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 873.7,
   "provenance": "community",
   "quant": "Mixed-4-8",
   "quote": "LFM2.5-24B-A2B-MLX-Serve-mixed-4-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 32k 873.7 57.6 14.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vy6s6otk",
   "source_url": "https://omlx.ai/benchmarks/performance/vy6s6otk",
   "tg_tokens": null,
   "tps": 57.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-vy6s6otk-ctx65536",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX Serve build; oMLX community benchmark trace; peak memory 15.2 GB at this context; duplicate traces of the same session: 578exclt; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 739.9,
   "provenance": "community",
   "quant": "Mixed-4-8",
   "quote": "LFM2.5-24B-A2B-MLX-Serve-mixed-4-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 64k 739.9 45.5 15.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vy6s6otk",
   "source_url": "https://omlx.ai/benchmarks/performance/vy6s6otk",
   "tg_tokens": null,
   "tps": 45.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-vy6s6otk-ctx131072",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX Serve build; oMLX community benchmark trace; peak memory 16.4 GB at this context; duplicate traces of the same session: 578exclt; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 565.9,
   "provenance": "community",
   "quant": "Mixed-4-8",
   "quote": "LFM2.5-24B-A2B-MLX-Serve-mixed-4-8bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 128k 565.9 32.0 16.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vy6s6otk",
   "source_url": "https://omlx.ai/benchmarks/performance/vy6s6otk",
   "tg_tokens": null,
   "tps": 32.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-13ocxo7j-ctx4096",
   "model": "Nail-Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: luxr8gvx; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 990.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 990.6 63.9 21.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/13ocxo7j",
   "source_url": "https://omlx.ai/benchmarks/performance/13ocxo7j",
   "tg_tokens": null,
   "tps": 63.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-13ocxo7j-ctx16384",
   "model": "Nail-Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: luxr8gvx; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 872.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Nail-Qwen3.6-35B-A3B-MLX on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 872.6 60.3 21.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/13ocxo7j",
   "source_url": "https://omlx.ai/benchmarks/performance/13ocxo7j",
   "tg_tokens": null,
   "tps": 60.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-hfgs0fmm-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 0xoju66m; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1042.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 1,042 84.0 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hfgs0fmm",
   "source_url": "https://omlx.ai/benchmarks/performance/hfgs0fmm",
   "tg_tokens": null,
   "tps": 84.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-hfgs0fmm-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.4 GB at this context; duplicate traces of the same session: 0xoju66m; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 886.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 886.7 77.6 20.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hfgs0fmm",
   "source_url": "https://omlx.ai/benchmarks/performance/hfgs0fmm",
   "tg_tokens": null,
   "tps": 77.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-b03p28ju-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.1 GB at this context; duplicate traces of the same session: 2b8m8l3m, din90iwc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1060.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,060 99.9 43.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b03p28ju",
   "source_url": "https://omlx.ai/benchmarks/performance/b03p28ju",
   "tg_tokens": null,
   "tps": 99.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-b03p28ju-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 43.9 GB at this context; duplicate traces of the same session: 2b8m8l3m, din90iwc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2615.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,615 99.5 43.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b03p28ju",
   "source_url": "https://omlx.ai/benchmarks/performance/b03p28ju",
   "tg_tokens": null,
   "tps": 99.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-b03p28ju-ctx32768",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 44.5 GB at this context; duplicate traces of the same session: 2b8m8l3m, din90iwc; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3443.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3-Coder-Next-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 3,443 86.9 44.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b03p28ju",
   "source_url": "https://omlx.ai/benchmarks/performance/b03p28ju",
   "tg_tokens": null,
   "tps": 86.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-qrjjbsn5-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: ejsu4dg9; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 990.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 1k 990.5 30.4 5.8 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qrjjbsn5",
   "source_url": "https://omlx.ai/benchmarks/performance/qrjjbsn5",
   "tg_tokens": null,
   "tps": 30.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-qrjjbsn5-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: ejsu4dg9; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1104.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 4k 1,104 29.9 6.2 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qrjjbsn5",
   "source_url": "https://omlx.ai/benchmarks/performance/qrjjbsn5",
   "tg_tokens": null,
   "tps": 29.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-mhnttl69-ctx1024",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: 2gr3798y; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 102.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-27B-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 1k 102.0 8.3 16.0 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mhnttl69",
   "source_url": "https://omlx.ai/benchmarks/performance/mhnttl69",
   "tg_tokens": null,
   "tps": 8.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-mhnttl69-ctx4096",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: 2gr3798y; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 103.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-27B-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 4k 103.5 8.6 17.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mhnttl69",
   "source_url": "https://omlx.ai/benchmarks/performance/mhnttl69",
   "tg_tokens": null,
   "tps": 8.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-e6uupznc-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 34c36ru0; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 678.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 1k 678.6 58.4 19.4 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e6uupznc",
   "source_url": "https://omlx.ai/benchmarks/performance/e6uupznc",
   "tg_tokens": null,
   "tps": 58.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-e6uupznc-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 34c36ru0; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 773.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 4k 773.7 57.4 20.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e6uupznc",
   "source_url": "https://omlx.ai/benchmarks/performance/e6uupznc",
   "tg_tokens": null,
   "tps": 57.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-1wlz3v4s-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 22.3 GB; TTFT 20410 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 802.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M2 Max (38c); M2 Max (38c) \u00b7 64 GB \u00b7 4bit; 16k tokens 802.8 PP tok/s 71.5 TG tok/s 20410 TTFT (ms) 22.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1wlz3v4s",
   "source_url": "https://omlx.ai/benchmarks/performance/1wlz3v4s",
   "tg_tokens": null,
   "tps": 71.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-gumtyz72-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: 64epghn0; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 494.0,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.6-35B-A3B-oQ3e-mtp on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 3bit; 4k 494.0 32.2 18.3 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gumtyz72",
   "source_url": "https://omlx.ai/benchmarks/performance/gumtyz72",
   "tg_tokens": null,
   "tps": 32.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-gumtyz72-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: 64epghn0; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 421.7,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.6-35B-A3B-oQ3e-mtp on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 3bit; 16k 421.7 47.1 18.7 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gumtyz72",
   "source_url": "https://omlx.ai/benchmarks/performance/gumtyz72",
   "tg_tokens": null,
   "tps": 47.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-aimrego7-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: vojkr6t5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 970.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 970.9 51.7 22.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aimrego7",
   "source_url": "https://omlx.ai/benchmarks/performance/aimrego7",
   "tg_tokens": null,
   "tps": 51.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-aimrego7-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: vojkr6t5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 847.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 847.3 53.9 22.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aimrego7",
   "source_url": "https://omlx.ai/benchmarks/performance/aimrego7",
   "tg_tokens": null,
   "tps": 53.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-14exbqgb-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 33.2 GB; TTFT 66550 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1970.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 128k tokens 1,970 PP tok/s 69.3 TG tok/s 66550 TTFT (ms) 33.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/14exbqgb",
   "source_url": "https://omlx.ai/benchmarks/performance/14exbqgb",
   "tg_tokens": null,
   "tps": 69.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-lehbks9z-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.1 GB at this context; duplicate traces of the same session: xnq78jab; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 507.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 507.1 17.5 17.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lehbks9z",
   "source_url": "https://omlx.ai/benchmarks/performance/lehbks9z",
   "tg_tokens": null,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-lehbks9z-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: xnq78jab; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 468.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 468.1 16.6 18.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lehbks9z",
   "source_url": "https://omlx.ai/benchmarks/performance/lehbks9z",
   "tg_tokens": null,
   "tps": 16.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-ehs2y1ak-ctx131072",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace, single run at 128k context; peak memory 32.7 GB; TTFT 712669 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 183.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-fp16-mtp on M2 Ultra (60c); M2 Ultra (60c) \u00b7 64 GB \u00b7 4bit; 128k tokens 183.9 PP tok/s 37.9 TG tok/s 712669 TTFT (ms) 32.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ehs2y1ak",
   "source_url": "https://omlx.ai/benchmarks/performance/ehs2y1ak",
   "tg_tokens": null,
   "tps": 37.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-j99c0wqg-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: t3wiw0r0; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 66.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 4k 66.0 6.4 17.8 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j99c0wqg",
   "source_url": "https://omlx.ai/benchmarks/performance/j99c0wqg",
   "tg_tokens": null,
   "tps": 6.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-j99c0wqg-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: t3wiw0r0; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 60.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 16k 60.4 5.7 18.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j99c0wqg",
   "source_url": "https://omlx.ai/benchmarks/performance/j99c0wqg",
   "tg_tokens": null,
   "tps": 5.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-y9ru9geo-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 20.9 GB; TTFT 274695 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 59.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 16k tokens 59.6 PP tok/s 13.0 TG tok/s 274695 TTFT (ms) 20.9 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y9ru9geo",
   "source_url": "https://omlx.ai/benchmarks/performance/y9ru9geo",
   "tg_tokens": null,
   "tps": 13.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-rtck1ls0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.7 GB at this context; duplicate traces of the same session: 5ifh2v5c; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 124.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 124.9 17.3 19.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rtck1ls0",
   "source_url": "https://omlx.ai/benchmarks/performance/rtck1ls0",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-rtck1ls0-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.6 GB at this context; duplicate traces of the same session: 5ifh2v5c; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 119.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 119.4 25.5 20.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rtck1ls0",
   "source_url": "https://omlx.ai/benchmarks/performance/rtck1ls0",
   "tg_tokens": null,
   "tps": 25.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-rdbp0236-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build 2615fc0; oMLX community benchmark trace, single run at 128k context; peak memory 78.4 GB; TTFT 255965 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 512.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-2615fc0 on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k tokens 512.1 PP tok/s 48.5 TG tok/s 255965 TTFT (ms) 78.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rdbp0236",
   "source_url": "https://omlx.ai/benchmarks/performance/rdbp0236",
   "tg_tokens": null,
   "tps": 48.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-j5omwp95-ctx1024",
   "model": "gemma-4-12b-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 13.6 GB at this context; duplicate traces of the same session: du7d0652, z4hwd3y2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 277.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-12B-it-oQ8e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 1k 277.2 40.9 13.6 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j5omwp95",
   "source_url": "https://omlx.ai/benchmarks/performance/j5omwp95",
   "tg_tokens": null,
   "tps": 40.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-j5omwp95-ctx4096",
   "model": "gemma-4-12b-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 14.1 GB at this context; duplicate traces of the same session: du7d0652, z4hwd3y2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 278.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-12B-it-oQ8e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 4k 278.7 28.3 14.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j5omwp95",
   "source_url": "https://omlx.ai/benchmarks/performance/j5omwp95",
   "tg_tokens": null,
   "tps": 28.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-j5omwp95-ctx131072",
   "model": "gemma-4-12b-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: du7d0652, z4hwd3y2; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 174.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-12B-it-oQ8e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 128k 174.9 13.7 19.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j5omwp95",
   "source_url": "https://omlx.ai/benchmarks/performance/j5omwp95",
   "tg_tokens": null,
   "tps": 13.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1u3mrk5l-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: 9z56s5dt; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4287.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 4,287 104.9 11.8 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1u3mrk5l",
   "source_url": "https://omlx.ai/benchmarks/performance/1u3mrk5l",
   "tg_tokens": null,
   "tps": 104.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1u3mrk5l-ctx16384",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 12.1 GB at this context; duplicate traces of the same session: 9z56s5dt; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4019.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 16k 4,019 90.2 12.1 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1u3mrk5l",
   "source_url": "https://omlx.ai/benchmarks/performance/1u3mrk5l",
   "tg_tokens": null,
   "tps": 90.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-ec273yj1-ctx4096",
   "model": "4-bit",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 17.2 GB; TTFT 13298 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 308.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "4-bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k tokens 308.0 PP tok/s 9.7 TG tok/s 13298 TTFT (ms) 17.2 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ec273yj1",
   "source_url": "https://omlx.ai/benchmarks/performance/ec273yj1",
   "tg_tokens": null,
   "tps": 9.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-vxop8m5g-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 156.3 GB at this context; duplicate traces of the same session: zjabode6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 25.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 1k 25.5 5.1 156.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vxop8m5g",
   "source_url": "https://omlx.ai/benchmarks/performance/vxop8m5g",
   "tg_tokens": null,
   "tps": 5.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-vxop8m5g-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 156.8 GB at this context; duplicate traces of the same session: zjabode6; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 59.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 4k 59.1 5.6 156.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vxop8m5g",
   "source_url": "https://omlx.ai/benchmarks/performance/vxop8m5g",
   "tg_tokens": null,
   "tps": 5.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-yju954d0-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 403.1 GB at this context; duplicate traces of the same session: 77tup651, qixnitpn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 557.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 557.6 33.0 403.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yju954d0",
   "source_url": "https://omlx.ai/benchmarks/performance/yju954d0",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-yju954d0-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 406.8 GB at this context; duplicate traces of the same session: 77tup651, qixnitpn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 863.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 8k 863.7 32.8 406.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yju954d0",
   "source_url": "https://omlx.ai/benchmarks/performance/yju954d0",
   "tg_tokens": null,
   "tps": 32.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-yju954d0-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 406.8 GB at this context; duplicate traces of the same session: 77tup651, qixnitpn; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 886.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 16k 886.3 36.2 406.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yju954d0",
   "source_url": "https://omlx.ai/benchmarks/performance/yju954d0",
   "tg_tokens": null,
   "tps": 36.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5lk5anvp-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 288.8 GB at this context; duplicate traces of the same session: 0kvbdfr5, jp5fplk2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 474.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 474.0 26.6 288.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5lk5anvp",
   "source_url": "https://omlx.ai/benchmarks/performance/5lk5anvp",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5lk5anvp-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 292.1 GB at this context; duplicate traces of the same session: 0kvbdfr5, jp5fplk2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 438.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 8k 438.0 30.1 292.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5lk5anvp",
   "source_url": "https://omlx.ai/benchmarks/performance/5lk5anvp",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5lk5anvp-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 292.2 GB at this context; duplicate traces of the same session: 0kvbdfr5, jp5fplk2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 595.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 16k 595.9 38.9 292.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5lk5anvp",
   "source_url": "https://omlx.ai/benchmarks/performance/5lk5anvp",
   "tg_tokens": null,
   "tps": 38.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-mwgyw6dg-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 288.9 GB at this context; duplicate traces of the same session: 1axefnfp, t5wipska, uehl0gbq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 491.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 491.0 29.0 288.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mwgyw6dg",
   "source_url": "https://omlx.ai/benchmarks/performance/mwgyw6dg",
   "tg_tokens": null,
   "tps": 29.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-mwgyw6dg-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 289.3 GB at this context; duplicate traces of the same session: 1axefnfp, t5wipska, uehl0gbq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 652.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 652.9 32.3 289.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mwgyw6dg",
   "source_url": "https://omlx.ai/benchmarks/performance/mwgyw6dg",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-mwgyw6dg-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 289.3 GB at this context; duplicate traces of the same session: 1axefnfp, t5wipska, uehl0gbq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 705.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 8k 705.6 33.2 289.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mwgyw6dg",
   "source_url": "https://omlx.ai/benchmarks/performance/mwgyw6dg",
   "tg_tokens": null,
   "tps": 33.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-mwgyw6dg-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 289.3 GB at this context; duplicate traces of the same session: 1axefnfp, t5wipska, uehl0gbq; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 734.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 16k 734.0 37.5 289.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mwgyw6dg",
   "source_url": "https://omlx.ai/benchmarks/performance/mwgyw6dg",
   "tg_tokens": null,
   "tps": 37.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-5jtw80sd-ctx32768",
   "model": "MiMo-V2.6-Flash-RL",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 165.4 GB; TTFT 100590 ms; duplicate traces of the same session: none; config=76c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 325.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiMo-V2.6-Flash-RL-MLX-4bit-MTP on M2 Ultra (76c); M2 Ultra (76c) \u00b7 192 GB \u00b7 4bit; 32k tokens 325.8 PP tok/s 40.7 TG tok/s 100590 TTFT (ms) 165.4 Peak mem",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5jtw80sd",
   "source_url": "https://omlx.ai/benchmarks/performance/5jtw80sd",
   "tg_tokens": null,
   "tps": 40.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-pskl1x93-ctx1024",
   "model": "MiMo-V2.6-Flash-RL-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 158.0 GB at this context; duplicate traces of the same session: 1o64mxzo, 8p3hwm7u, jev3svtb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 571.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "MiMo-V2.6-Flash-RL-UNCENSORED-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 571.5 43.2 158.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pskl1x93",
   "source_url": "https://omlx.ai/benchmarks/performance/pskl1x93",
   "tg_tokens": null,
   "tps": 43.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-pskl1x93-ctx4096",
   "model": "MiMo-V2.6-Flash-RL-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 159.3 GB at this context; duplicate traces of the same session: 1o64mxzo, 8p3hwm7u, jev3svtb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 661.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "MiMo-V2.6-Flash-RL-UNCENSORED-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 661.2 50.5 159.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pskl1x93",
   "source_url": "https://omlx.ai/benchmarks/performance/pskl1x93",
   "tg_tokens": null,
   "tps": 50.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-pskl1x93-ctx8192",
   "model": "MiMo-V2.6-Flash-RL-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 160.4 GB at this context; duplicate traces of the same session: 1o64mxzo, 8p3hwm7u, jev3svtb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 650.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "MiMo-V2.6-Flash-RL-UNCENSORED-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 650.6 52.9 160.4 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pskl1x93",
   "source_url": "https://omlx.ai/benchmarks/performance/pskl1x93",
   "tg_tokens": null,
   "tps": 52.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-pskl1x93-ctx16384",
   "model": "MiMo-V2.6-Flash-RL-UNCENSORED",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 162.6 GB at this context; duplicate traces of the same session: 1o64mxzo, 8p3hwm7u, jev3svtb; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 612.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "MiMo-V2.6-Flash-RL-UNCENSORED-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 16k 612.7 43.9 162.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pskl1x93",
   "source_url": "https://omlx.ai/benchmarks/performance/pskl1x93",
   "tg_tokens": null,
   "tps": 43.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9k2hpg2u-ctx1024",
   "model": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B",
   "notes": "24GB LiLiCorr MLX build; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 949.9,
   "provenance": "community",
   "quant": "unknown",
   "quote": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-LiLiCorr-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 949.9 75.4 23.9 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9k2hpg2u",
   "source_url": "https://omlx.ai/benchmarks/performance/9k2hpg2u",
   "tg_tokens": null,
   "tps": 75.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-9k2hpg2u-ctx4096",
   "model": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B",
   "notes": "24GB LiLiCorr MLX build; oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1032.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-LiLiCorr-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 1,032 73.7 24.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9k2hpg2u",
   "source_url": "https://omlx.ai/benchmarks/performance/9k2hpg2u",
   "tg_tokens": null,
   "tps": 73.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-zrjgydi0-ctx65536",
   "model": "Qwen3.5-35B-A3B-Uncensored-HauhauCS-Aggressive",
   "notes": "MLX build; oMLX community benchmark trace, single run at 64k context; peak memory 20.3 GB; TTFT 35694 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1836.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "Qwen3.5-35B-A3B-Uncensored-HauhauCS-Aggressive-MLX-mxfp4 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k tokens 1,836 PP tok/s 61.0 TG tok/s 35694 TTFT (ms) 20.3 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zrjgydi0",
   "source_url": "https://omlx.ai/benchmarks/performance/zrjgydi0",
   "tg_tokens": null,
   "tps": 61.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-51k3itib-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 9u5paffr; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 720.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 1k 720.7 73.5 19.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/51k3itib",
   "source_url": "https://omlx.ai/benchmarks/performance/51k3itib",
   "tg_tokens": null,
   "tps": 73.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-51k3itib-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 9u5paffr; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 749.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 4k 749.9 69.3 20.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/51k3itib",
   "source_url": "https://omlx.ai/benchmarks/performance/51k3itib",
   "tg_tokens": null,
   "tps": 69.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-zcfvzj0k-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 7nxetqkq; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 713.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 1k 713.5 73.3 19.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zcfvzj0k",
   "source_url": "https://omlx.ai/benchmarks/performance/zcfvzj0k",
   "tg_tokens": null,
   "tps": 73.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-zcfvzj0k-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: 7nxetqkq; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 745.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 4k 745.5 72.5 20.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zcfvzj0k",
   "source_url": "https://omlx.ai/benchmarks/performance/zcfvzj0k",
   "tg_tokens": null,
   "tps": 72.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-ek1s7dd0-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: bao59efv; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 250.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 1k 250.2 40.1 17.3 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ek1s7dd0",
   "source_url": "https://omlx.ai/benchmarks/performance/ek1s7dd0",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-ek1s7dd0-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 18.0 GB at this context; duplicate traces of the same session: bao59efv; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 265.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 4k 265.8 38.2 18.0 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ek1s7dd0",
   "source_url": "https://omlx.ai/benchmarks/performance/ek1s7dd0",
   "tg_tokens": null,
   "tps": 38.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-6af4vhkg-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 17.3 GB; TTFT 4016 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 255.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 24 GB \u00b7 4bit; 1k tokens 255.0 PP tok/s 40.0 TG tok/s 4016 TTFT (ms) 17.3 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6af4vhkg",
   "source_url": "https://omlx.ai/benchmarks/performance/6af4vhkg",
   "tg_tokens": null,
   "tps": 40.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-tyjwtlzo-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.2 GB at this context; duplicate traces of the same session: mwp5k9qe; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 406.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 406.7 16.2 18.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tyjwtlzo",
   "source_url": "https://omlx.ai/benchmarks/performance/tyjwtlzo",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-tyjwtlzo-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: mwp5k9qe; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 436.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 436.0 15.5 18.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tyjwtlzo",
   "source_url": "https://omlx.ai/benchmarks/performance/tyjwtlzo",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-evpiopfu-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 18.8 GB; TTFT 37520 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 436.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k tokens 436.7 PP tok/s 15.0 TG tok/s 37520 TTFT (ms) 18.8 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/evpiopfu",
   "source_url": "https://omlx.ai/benchmarks/performance/evpiopfu",
   "tg_tokens": null,
   "tps": 15.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-p3fo0a6a-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 18.2 GB; TTFT 10026 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 408.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k tokens 408.6 PP tok/s 14.4 TG tok/s 10026 TTFT (ms) 18.2 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p3fo0a6a",
   "source_url": "https://omlx.ai/benchmarks/performance/p3fo0a6a",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-iti94d8s-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 128k context; peak memory 43.5 GB; TTFT 219119 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 598.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 128k tokens 598.2 PP tok/s 31.7 TG tok/s 219119 TTFT (ms) 43.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iti94d8s",
   "source_url": "https://omlx.ai/benchmarks/performance/iti94d8s",
   "tg_tokens": null,
   "tps": 31.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-yzsc1rj3-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 32.3 GB; TTFT 37514 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 873.5,
   "provenance": "community",
   "quant": "GSQ-RCO-Q2",
   "quote": "Qwen3.8-Flash-Next-GSQ-RCO-Q2-mlx on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k tokens 873.5 PP tok/s 27.1 TG tok/s 37514 TTFT (ms) 32.3 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yzsc1rj3",
   "source_url": "https://omlx.ai/benchmarks/performance/yzsc1rj3",
   "tg_tokens": null,
   "tps": 27.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-lpuuim7h-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace, single run at 8k context; peak memory 30.8 GB; TTFT 11721 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 698.9,
   "provenance": "community",
   "quant": "GSQ-RCO-Q2",
   "quote": "Qwen3.8-Flash-Next-GSQ-RCO-Q2-mlx on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 8k tokens 698.9 PP tok/s 32.3 TG tok/s 11721 TTFT (ms) 30.8 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lpuuim7h",
   "source_url": "https://omlx.ai/benchmarks/performance/lpuuim7h",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-z2nrdlku-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 32.2 GB; TTFT 38191 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 858.0,
   "provenance": "community",
   "quant": "GSQ-RCO-Q2",
   "quote": "Qwen3.8-Flash-Next-GSQ-RCO-Q2-mlx on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 32k tokens 858.0 PP tok/s 24.8 TG tok/s 38191 TTFT (ms) 32.2 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z2nrdlku",
   "source_url": "https://omlx.ai/benchmarks/performance/z2nrdlku",
   "tg_tokens": null,
   "tps": 24.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-tfeg5da3-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 28.9 GB; TTFT 1901 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 538.8,
   "provenance": "community",
   "quant": "GSQ-RCO-Q2",
   "quote": "Qwen3.8-Flash-Next-GSQ-RCO-Q2-mlx on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k tokens 538.8 PP tok/s 26.2 TG tok/s 1901 TTFT (ms) 28.9 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tfeg5da3",
   "source_url": "https://omlx.ai/benchmarks/performance/tfeg5da3",
   "tg_tokens": null,
   "tps": 26.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-38tyv8x9-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 77.0 GB at this context; duplicate traces of the same session: 47r962gt, ur8tn4n4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 895.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 895.8 57.3 77.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/38tyv8x9",
   "source_url": "https://omlx.ai/benchmarks/performance/38tyv8x9",
   "tg_tokens": null,
   "tps": 57.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-38tyv8x9-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.1 GB at this context; duplicate traces of the same session: 47r962gt, ur8tn4n4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1705.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 1,705 60.7 78.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/38tyv8x9",
   "source_url": "https://omlx.ai/benchmarks/performance/38tyv8x9",
   "tg_tokens": null,
   "tps": 60.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-38tyv8x9-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 86.2 GB at this context; duplicate traces of the same session: 47r962gt, ur8tn4n4; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2237.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k 2,237 60.4 86.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/38tyv8x9",
   "source_url": "https://omlx.ai/benchmarks/performance/38tyv8x9",
   "tg_tokens": null,
   "tps": 60.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-5ri4kp3u-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6; MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 86.2 GB; TTFT 27292 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2401.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k tokens 2,401 PP tok/s 63.3 TG tok/s 27292 TTFT (ms) 86.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ri4kp3u",
   "source_url": "https://omlx.ai/benchmarks/performance/5ri4kp3u",
   "tg_tokens": null,
   "tps": 63.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8bn0vy9q-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.1 GB at this context; duplicate traces of the same session: 5ydgs8sm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1934.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 1,934 56.4 78.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8bn0vy9q",
   "source_url": "https://omlx.ai/benchmarks/performance/8bn0vy9q",
   "tg_tokens": null,
   "tps": 56.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8bn0vy9q-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 86.2 GB at this context; duplicate traces of the same session: 5ydgs8sm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2257.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k 2,257 58.4 86.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8bn0vy9q",
   "source_url": "https://omlx.ai/benchmarks/performance/8bn0vy9q",
   "tg_tokens": null,
   "tps": 58.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-w7x026iu-ctx1024",
   "model": "gemma-4-E4B-it-ultra-uncensored-heretic",
   "notes": "oMLX community benchmark trace; peak memory 5.5 GB at this context; duplicate traces of the same session: p852j08a; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 651.2,
   "provenance": "community",
   "quant": "mixed 4+6bit",
   "quote": "gemma-4-E4B-it-ultra-uncensored-heretic-MLX-4bit-mixed_4_6 on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 651.2 31.4 5.5 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w7x026iu",
   "source_url": "https://omlx.ai/benchmarks/performance/w7x026iu",
   "tg_tokens": null,
   "tps": 31.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-w7x026iu-ctx4096",
   "model": "gemma-4-E4B-it-ultra-uncensored-heretic",
   "notes": "oMLX community benchmark trace; peak memory 5.5 GB at this context; duplicate traces of the same session: p852j08a; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 660.2,
   "provenance": "community",
   "quant": "mixed 4+6bit",
   "quote": "gemma-4-E4B-it-ultra-uncensored-heretic-MLX-4bit-mixed_4_6 on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 660.2 30.5 5.5 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w7x026iu",
   "source_url": "https://omlx.ai/benchmarks/performance/w7x026iu",
   "tg_tokens": null,
   "tps": 30.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-bhuic7ky-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2831.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,831 81.4 11.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhuic7ky",
   "source_url": "https://omlx.ai/benchmarks/performance/bhuic7ky",
   "tg_tokens": null,
   "tps": 81.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-bhuic7ky-ctx16384",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 12.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2565.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,565 68.7 12.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhuic7ky",
   "source_url": "https://omlx.ai/benchmarks/performance/bhuic7ky",
   "tg_tokens": null,
   "tps": 68.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vzmmny1z-ctx1024",
   "model": "Agnes-3.0-Flash",
   "notes": "MLX MTP build; oMLX community benchmark trace; peak memory 39.1 GB at this context; duplicate traces of the same session: z6fr1j0a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 329.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Agnes-3.0-Flash-MLX-8bit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 329.0 12.6 39.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vzmmny1z",
   "source_url": "https://omlx.ai/benchmarks/performance/vzmmny1z",
   "tg_tokens": null,
   "tps": 12.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-vzmmny1z-ctx4096",
   "model": "Agnes-3.0-Flash",
   "notes": "MLX MTP build; oMLX community benchmark trace; peak memory 41.6 GB at this context; duplicate traces of the same session: z6fr1j0a; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 456.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Agnes-3.0-Flash-MLX-8bit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 456.9 12.9 41.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vzmmny1z",
   "source_url": "https://omlx.ai/benchmarks/performance/vzmmny1z",
   "tg_tokens": null,
   "tps": 12.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-iufmum3f-ctx16384",
   "model": "DeepSeek-Coder-V2-Lite-Instruct",
   "notes": "MLX; oMLX community benchmark trace, single run at 16k context; peak memory 15.1 GB; TTFT 41037 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 399.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-Coder-V2-Lite-Instruct-4bit-mlx on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 4bit; 16k tokens 399.2 PP tok/s 16.7 TG tok/s 41037 TTFT (ms) 15.1 Peak mem",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iufmum3f",
   "source_url": "https://omlx.ai/benchmarks/performance/iufmum3f",
   "tg_tokens": null,
   "tps": 16.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-l5n9ud7k-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 207.3 GB at this context; duplicate traces of the same session: 8849oqlb, h9wroi5k, qs66tli4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 381.6,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 1k 381.6 25.5 207.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l5n9ud7k",
   "source_url": "https://omlx.ai/benchmarks/performance/l5n9ud7k",
   "tg_tokens": null,
   "tps": 25.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-l5n9ud7k-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 207.4 GB at this context; duplicate traces of the same session: 8849oqlb, h9wroi5k, qs66tli4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 496.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 4k 496.0 30.6 207.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l5n9ud7k",
   "source_url": "https://omlx.ai/benchmarks/performance/l5n9ud7k",
   "tg_tokens": null,
   "tps": 30.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-l5n9ud7k-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 207.6 GB at this context; duplicate traces of the same session: 8849oqlb, h9wroi5k, qs66tli4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 514.1,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 16k 514.1 33.3 207.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l5n9ud7k",
   "source_url": "https://omlx.ai/benchmarks/performance/l5n9ud7k",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-l5n9ud7k-ctx32768",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 207.8 GB at this context; duplicate traces of the same session: 8849oqlb, h9wroi5k, qs66tli4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 522.5,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 32k 522.5 34.1 207.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l5n9ud7k",
   "source_url": "https://omlx.ai/benchmarks/performance/l5n9ud7k",
   "tg_tokens": null,
   "tps": 34.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-7dklx18d-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 205.5 GB at this context; duplicate traces of the same session: 4kjyh7ue, bd516rrh, xzgmq81z; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 620.7,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 unknown; 1k 620.7 43.7 205.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7dklx18d",
   "source_url": "https://omlx.ai/benchmarks/performance/7dklx18d",
   "tg_tokens": null,
   "tps": 43.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-7dklx18d-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 205.9 GB at this context; duplicate traces of the same session: 4kjyh7ue, bd516rrh, xzgmq81z; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1089.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 unknown; 8k 1,089 44.9 205.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7dklx18d",
   "source_url": "https://omlx.ai/benchmarks/performance/7dklx18d",
   "tg_tokens": null,
   "tps": 44.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-7dklx18d-ctx32768",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 206.0 GB at this context; duplicate traces of the same session: 4kjyh7ue, bd516rrh, xzgmq81z; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1208.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 unknown; 32k 1,208 54.2 206.0 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7dklx18d",
   "source_url": "https://omlx.ai/benchmarks/performance/7dklx18d",
   "tg_tokens": null,
   "tps": 54.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-7dklx18d-ctx131072",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 206.2 GB at this context; duplicate traces of the same session: 4kjyh7ue, bd516rrh, xzgmq81z; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 771.9,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 unknown; 128k 771.9 50.0 206.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7dklx18d",
   "source_url": "https://omlx.ai/benchmarks/performance/7dklx18d",
   "tg_tokens": null,
   "tps": 50.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-lqmnj3bb-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 406.8 GB; TTFT 9365 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 874.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 8k tokens 874.7 PP tok/s 32.6 TG tok/s 9365 TTFT (ms) 406.8 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lqmnj3bb",
   "source_url": "https://omlx.ai/benchmarks/performance/lqmnj3bb",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0widvdwf-ctx131072",
   "model": "muse-glimmer 30B",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 17.7 GB; TTFT 162035 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 808.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Muse-Glimmer-30B-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k tokens 808.9 PP tok/s 27.0 TG tok/s 162035 TTFT (ms) 17.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0widvdwf",
   "source_url": "https://omlx.ai/benchmarks/performance/0widvdwf",
   "tg_tokens": null,
   "tps": 27.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bwanmkp8-ctx1024",
   "model": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B",
   "notes": "24GB LiLiCorr MLX build; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 949.9,
   "provenance": "community",
   "quant": "unknown",
   "quote": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-LiLiCorr-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 949.9 75.4 23.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bwanmkp8",
   "source_url": "https://omlx.ai/benchmarks/performance/bwanmkp8",
   "tg_tokens": null,
   "tps": 75.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-bwanmkp8-ctx4096",
   "model": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B",
   "notes": "24GB LiLiCorr MLX build; oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1032.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-LiLiCorr-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 1,032 73.7 24.0 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bwanmkp8",
   "source_url": "https://omlx.ai/benchmarks/performance/bwanmkp8",
   "tg_tokens": null,
   "tps": 73.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4rvtd5p1-ctx1024",
   "model": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B",
   "notes": "24GB LiLiCorr MLX build; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: 34w81zue, hd1emolq, x8qbibm2, zn7ms5v0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 946.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-LiLiCorr-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 946.0 74.5 23.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4rvtd5p1",
   "source_url": "https://omlx.ai/benchmarks/performance/4rvtd5p1",
   "tg_tokens": null,
   "tps": 74.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4rvtd5p1-ctx4096",
   "model": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B",
   "notes": "24GB LiLiCorr MLX build; oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: 34w81zue, hd1emolq, x8qbibm2, zn7ms5v0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1078.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-LiLiCorr-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 1,078 72.8 24.0 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4rvtd5p1",
   "source_url": "https://omlx.ai/benchmarks/performance/4rvtd5p1",
   "tg_tokens": null,
   "tps": 72.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4rvtd5p1-ctx8192",
   "model": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B",
   "notes": "24GB LiLiCorr MLX build; oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: 34w81zue, hd1emolq, x8qbibm2, zn7ms5v0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1070.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-LiLiCorr-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 8k 1,070 72.6 24.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4rvtd5p1",
   "source_url": "https://omlx.ai/benchmarks/performance/4rvtd5p1",
   "tg_tokens": null,
   "tps": 72.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4rvtd5p1-ctx16384",
   "model": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B",
   "notes": "24GB LiLiCorr MLX build; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: 34w81zue, hd1emolq, x8qbibm2, zn7ms5v0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1045.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-LiLiCorr-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 16k 1,045 71.2 24.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4rvtd5p1",
   "source_url": "https://omlx.ai/benchmarks/performance/4rvtd5p1",
   "tg_tokens": null,
   "tps": 71.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-4rvtd5p1-ctx32768",
   "model": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B",
   "notes": "24GB LiLiCorr MLX build; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: 34w81zue, hd1emolq, x8qbibm2, zn7ms5v0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 963.4,
   "provenance": "community",
   "quant": "unknown",
   "quote": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-LiLiCorr-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 32k 963.4 31.9 24.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4rvtd5p1",
   "source_url": "https://omlx.ai/benchmarks/performance/4rvtd5p1",
   "tg_tokens": null,
   "tps": 31.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-zn48nvg9-ctx1024",
   "model": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B",
   "notes": "24GB LiLiCorr MLX build; oMLX community benchmark trace; peak memory 23.8 GB at this context; duplicate traces of the same session: 6d6w9dbm, d533w53u, k3uunuxg, m5ow1kv6; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 802.4,
   "provenance": "community",
   "quant": "unknown",
   "quote": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-LiLiCorr-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 802.4 73.3 23.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zn48nvg9",
   "source_url": "https://omlx.ai/benchmarks/performance/zn48nvg9",
   "tg_tokens": null,
   "tps": 73.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-zn48nvg9-ctx4096",
   "model": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B",
   "notes": "24GB LiLiCorr MLX build; oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: 6d6w9dbm, d533w53u, k3uunuxg, m5ow1kv6; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1044.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-LiLiCorr-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 1,044 71.0 24.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zn48nvg9",
   "source_url": "https://omlx.ai/benchmarks/performance/zn48nvg9",
   "tg_tokens": null,
   "tps": 71.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-zn48nvg9-ctx8192",
   "model": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B",
   "notes": "24GB LiLiCorr MLX build; oMLX community benchmark trace; peak memory 24.0 GB at this context; duplicate traces of the same session: 6d6w9dbm, d533w53u, k3uunuxg, m5ow1kv6; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1053.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-LiLiCorr-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 8k 1,053 71.8 24.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zn48nvg9",
   "source_url": "https://omlx.ai/benchmarks/performance/zn48nvg9",
   "tg_tokens": null,
   "tps": 71.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-zn48nvg9-ctx16384",
   "model": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B",
   "notes": "24GB LiLiCorr MLX build; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: 6d6w9dbm, d533w53u, k3uunuxg, m5ow1kv6; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1045.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-LiLiCorr-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 16k 1,045 68.9 24.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zn48nvg9",
   "source_url": "https://omlx.ai/benchmarks/performance/zn48nvg9",
   "tg_tokens": null,
   "tps": 68.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-zn48nvg9-ctx32768",
   "model": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B",
   "notes": "24GB LiLiCorr MLX build; oMLX community benchmark trace; peak memory 24.2 GB at this context; duplicate traces of the same session: 6d6w9dbm, d533w53u, k3uunuxg, m5ow1kv6; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 990.7,
   "provenance": "community",
   "quant": "unknown",
   "quote": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-LiLiCorr-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 32k 990.7 69.6 24.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zn48nvg9",
   "source_url": "https://omlx.ai/benchmarks/performance/zn48nvg9",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mij0ofqp-ctx131072",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 128k context; peak memory 40.6 GB; TTFT 57875 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2265.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 8bit; 128k tokens 2,265 PP tok/s 50.1 TG tok/s 57875 TTFT (ms) 40.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mij0ofqp",
   "source_url": "https://omlx.ai/benchmarks/performance/mij0ofqp",
   "tg_tokens": null,
   "tps": 50.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-zju6o4ho-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 4k context; peak memory 29.8 GB; TTFT 35226 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 116.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k tokens 116.3 PP tok/s 11.8 TG tok/s 35226 TTFT (ms) 29.8 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zju6o4ho",
   "source_url": "https://omlx.ai/benchmarks/performance/zju6o4ho",
   "tg_tokens": null,
   "tps": 11.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-4ss3tuwu-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 5dgvzsbf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 204.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 204.1 12.2 28.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ss3tuwu",
   "source_url": "https://omlx.ai/benchmarks/performance/4ss3tuwu",
   "tg_tokens": null,
   "tps": 12.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-4ss3tuwu-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.7 GB at this context; duplicate traces of the same session: 5dgvzsbf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 176.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 176.6 11.8 31.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4ss3tuwu",
   "source_url": "https://omlx.ai/benchmarks/performance/4ss3tuwu",
   "tg_tokens": null,
   "tps": 11.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-rbz025iu-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: 3hy4nzbm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 197.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 197.6 28.8 32.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rbz025iu",
   "source_url": "https://omlx.ai/benchmarks/performance/rbz025iu",
   "tg_tokens": null,
   "tps": 28.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-rbz025iu-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.7 GB at this context; duplicate traces of the same session: 3hy4nzbm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 192.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 192.7 21.3 33.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rbz025iu",
   "source_url": "https://omlx.ai/benchmarks/performance/rbz025iu",
   "tg_tokens": null,
   "tps": 21.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-pfmsjrez-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 32.1 GB at this context; duplicate traces of the same session: 0tzgksl5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 198.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 198.5 29.1 32.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pfmsjrez",
   "source_url": "https://omlx.ai/benchmarks/performance/pfmsjrez",
   "tg_tokens": null,
   "tps": 29.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-pfmsjrez-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 33.7 GB at this context; duplicate traces of the same session: 0tzgksl5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 189.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 189.3 21.9 33.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pfmsjrez",
   "source_url": "https://omlx.ai/benchmarks/performance/pfmsjrez",
   "tg_tokens": null,
   "tps": 21.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-fldpbzgp-ctx199680",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 195k context; peak memory 49.8 GB; TTFT 1978659 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 101.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M3 Max (40c); M3 Max (40c) \u00b7 128 GB \u00b7 8bit; 195k tokens 101.1 PP tok/s 0.6 TG tok/s 1978659 TTFT (ms) 49.8 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fldpbzgp",
   "source_url": "https://omlx.ai/benchmarks/performance/fldpbzgp",
   "tg_tokens": null,
   "tps": 0.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-7pml2ith-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 55.2 GB; TTFT 127506 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 128.5,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-Flash-Next-oQ3e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 3bit; 16k tokens 128.5 PP tok/s 19.2 TG tok/s 127506 TTFT (ms) 55.2 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7pml2ith",
   "source_url": "https://omlx.ai/benchmarks/performance/7pml2ith",
   "tg_tokens": null,
   "tps": 19.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-t4midpo3-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 51.5 GB at this context; duplicate traces of the same session: ym6mm8cu; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 75.9,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-Flash-Next-oQ3e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 3bit; 1k 75.9 20.6 51.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t4midpo3",
   "source_url": "https://omlx.ai/benchmarks/performance/t4midpo3",
   "tg_tokens": null,
   "tps": 20.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-t4midpo3-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 52.0 GB at this context; duplicate traces of the same session: ym6mm8cu; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 231.2,
   "provenance": "community",
   "quant": "oQ3e",
   "quote": "Qwen3.8-Flash-Next-oQ3e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 64 GB \u00b7 3bit; 4k 231.2 26.4 52.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t4midpo3",
   "source_url": "https://omlx.ai/benchmarks/performance/t4midpo3",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-czyqz2u0-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 77.0 GB at this context; duplicate traces of the same session: cg4ahonv, uqv240mp, xfu7ofr7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 958.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 958.5 49.1 77.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/czyqz2u0",
   "source_url": "https://omlx.ai/benchmarks/performance/czyqz2u0",
   "tg_tokens": null,
   "tps": 49.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-czyqz2u0-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 78.1 GB at this context; duplicate traces of the same session: cg4ahonv, uqv240mp, xfu7ofr7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1987.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 1,987 68.2 78.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/czyqz2u0",
   "source_url": "https://omlx.ai/benchmarks/performance/czyqz2u0",
   "tg_tokens": null,
   "tps": 68.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-czyqz2u0-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 84.5 GB at this context; duplicate traces of the same session: cg4ahonv, uqv240mp, xfu7ofr7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2585.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 16k 2,585 74.1 84.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/czyqz2u0",
   "source_url": "https://omlx.ai/benchmarks/performance/czyqz2u0",
   "tg_tokens": null,
   "tps": 74.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-czyqz2u0-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "mix6; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 86.2 GB at this context; duplicate traces of the same session: cg4ahonv, uqv240mp, xfu7ofr7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2403.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mix6-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k 2,403 63.4 86.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/czyqz2u0",
   "source_url": "https://omlx.ai/benchmarks/performance/czyqz2u0",
   "tg_tokens": null,
   "tps": 63.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-2unps8t5-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap MTP; oMLX community benchmark trace; peak memory 17.9 GB at this context; duplicate traces of the same session: pzsg20q7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 509.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "ThinkingCap-Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 509.6 17.2 17.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2unps8t5",
   "source_url": "https://omlx.ai/benchmarks/performance/2unps8t5",
   "tg_tokens": null,
   "tps": 17.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-2unps8t5-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "ThinkingCap MTP; oMLX community benchmark trace; peak memory 18.3 GB at this context; duplicate traces of the same session: pzsg20q7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 435.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "ThinkingCap-Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 435.5 16.3 18.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2unps8t5",
   "source_url": "https://omlx.ai/benchmarks/performance/2unps8t5",
   "tg_tokens": null,
   "tps": 16.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-4136xo1v-ctx1024",
   "model": "gemma-4-12b-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 13.6 GB at this context; duplicate traces of the same session: n9nx6k77; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 280.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-12B-it-oQ8e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 1k 280.9 37.2 13.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4136xo1v",
   "source_url": "https://omlx.ai/benchmarks/performance/4136xo1v",
   "tg_tokens": null,
   "tps": 37.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-4136xo1v-ctx4096",
   "model": "gemma-4-12b-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 14.1 GB at this context; duplicate traces of the same session: n9nx6k77; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 279.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-12B-it-oQ8e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 4k 279.4 56.9 14.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4136xo1v",
   "source_url": "https://omlx.ai/benchmarks/performance/4136xo1v",
   "tg_tokens": null,
   "tps": 56.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-4136xo1v-ctx8192",
   "model": "gemma-4-12b-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: n9nx6k77; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 274.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-12B-it-oQ8e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 8k 274.2 32.3 14.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4136xo1v",
   "source_url": "https://omlx.ai/benchmarks/performance/4136xo1v",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-4136xo1v-ctx16384",
   "model": "gemma-4-12b-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 14.7 GB at this context; duplicate traces of the same session: n9nx6k77; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 261.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-12B-it-oQ8e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 16k 261.9 46.4 14.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4136xo1v",
   "source_url": "https://omlx.ai/benchmarks/performance/4136xo1v",
   "tg_tokens": null,
   "tps": 46.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-4136xo1v-ctx32768",
   "model": "gemma-4-12b-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 15.4 GB at this context; duplicate traces of the same session: n9nx6k77; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 246.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-12B-it-oQ8e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 32k 246.6 16.9 15.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4136xo1v",
   "source_url": "https://omlx.ai/benchmarks/performance/4136xo1v",
   "tg_tokens": null,
   "tps": 16.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-4136xo1v-ctx65536",
   "model": "gemma-4-12b-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: n9nx6k77; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 217.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-12B-it-oQ8e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 64k 217.4 16.2 16.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4136xo1v",
   "source_url": "https://omlx.ai/benchmarks/performance/4136xo1v",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-4136xo1v-ctx131072",
   "model": "gemma-4-12b-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: n9nx6k77; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 175.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-12B-it-oQ8e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 128k 175.1 13.8 20.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4136xo1v",
   "source_url": "https://omlx.ai/benchmarks/performance/4136xo1v",
   "tg_tokens": null,
   "tps": 13.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-b2fv148v-ctx4096",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2831.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 2,831 81.4 11.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b2fv148v",
   "source_url": "https://omlx.ai/benchmarks/performance/b2fv148v",
   "tg_tokens": null,
   "tps": 81.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-b2fv148v-ctx16384",
   "model": "gpt-oss 20B",
   "notes": "oMLX community benchmark trace; peak memory 12.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2565.0,
   "provenance": "community",
   "quant": "MXFP4-Q8",
   "quote": "gpt-oss-20b-MXFP4-Q8 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 2,565 68.7 12.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b2fv148v",
   "source_url": "https://omlx.ai/benchmarks/performance/b2fv148v",
   "tg_tokens": null,
   "tps": 68.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-c2b03rqp-ctx1024",
   "model": "Agnes-3.0-Flash",
   "notes": "MLX MTP build; oMLX community benchmark trace; peak memory 38.7 GB at this context; duplicate traces of the same session: yk7mrpr7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 356.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Agnes-3.0-Flash-MLX-8bit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 356.6 14.9 38.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c2b03rqp",
   "source_url": "https://omlx.ai/benchmarks/performance/c2b03rqp",
   "tg_tokens": null,
   "tps": 14.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-c2b03rqp-ctx4096",
   "model": "Agnes-3.0-Flash",
   "notes": "MLX MTP build; oMLX community benchmark trace; peak memory 41.3 GB at this context; duplicate traces of the same session: yk7mrpr7; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 631.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Agnes-3.0-Flash-MLX-8bit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 631.9 15.8 41.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c2b03rqp",
   "source_url": "https://omlx.ai/benchmarks/performance/c2b03rqp",
   "tg_tokens": null,
   "tps": 15.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-fj4h8n7v-ctx65536",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 208.2 GB at this context; duplicate traces of the same session: t4hpyfc4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 497.9,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 64k 497.9 35.6 208.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fj4h8n7v",
   "source_url": "https://omlx.ai/benchmarks/performance/fj4h8n7v",
   "tg_tokens": null,
   "tps": 35.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-fj4h8n7v-ctx131072",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 209.1 GB at this context; duplicate traces of the same session: t4hpyfc4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 444.5,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 128k 444.5 40.5 209.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fj4h8n7v",
   "source_url": "https://omlx.ai/benchmarks/performance/fj4h8n7v",
   "tg_tokens": null,
   "tps": 40.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-k2vt93a2-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 207.4 GB at this context; duplicate traces of the same session: yx11qgck; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 477.3,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 4k 477.3 28.4 207.4 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k2vt93a2",
   "source_url": "https://omlx.ai/benchmarks/performance/k2vt93a2",
   "tg_tokens": null,
   "tps": 28.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-k2vt93a2-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 207.6 GB at this context; duplicate traces of the same session: yx11qgck; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 522.4,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 16k 522.4 37.8 207.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/k2vt93a2",
   "source_url": "https://omlx.ai/benchmarks/performance/k2vt93a2",
   "tg_tokens": null,
   "tps": 37.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0ixbe4gn-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 207.4 GB; TTFT 8869 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 461.8,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 4k tokens 461.8 PP tok/s 43.8 TG tok/s 8869 TTFT (ms) 207.4 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ixbe4gn",
   "source_url": "https://omlx.ai/benchmarks/performance/0ixbe4gn",
   "tg_tokens": null,
   "tps": 43.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-0puod68z-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 156.4 GB at this context; duplicate traces of the same session: pqnuv5gp; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 38.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 1k 38.5 7.3 156.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0puod68z",
   "source_url": "https://omlx.ai/benchmarks/performance/0puod68z",
   "tg_tokens": null,
   "tps": 7.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-0puod68z-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 156.8 GB at this context; duplicate traces of the same session: pqnuv5gp; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 83.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 4k 83.5 10.2 156.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0puod68z",
   "source_url": "https://omlx.ai/benchmarks/performance/0puod68z",
   "tg_tokens": null,
   "tps": 10.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ct3bkq0c-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: xhpawnaj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 384.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 384.5 30.1 18.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ct3bkq0c",
   "source_url": "https://omlx.ai/benchmarks/performance/ct3bkq0c",
   "tg_tokens": null,
   "tps": 30.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ct3bkq0c-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "Jundot repack; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: xhpawnaj; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 377.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Jundot--Qwen3.8-27B-oQ4e-mtp on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 16k 377.4 36.1 19.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ct3bkq0c",
   "source_url": "https://omlx.ai/benchmarks/performance/ct3bkq0c",
   "tg_tokens": null,
   "tps": 36.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-27eavj0z-ctx16384",
   "model": "MiniCPM-V-4.7-1B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 4.0 GB; TTFT 5074 ms; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3229.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM-V-4.7-1B-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 16k tokens 3,229 PP tok/s 125.7 TG tok/s 5074 TTFT (ms) 4.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/27eavj0z",
   "source_url": "https://omlx.ai/benchmarks/performance/27eavj0z",
   "tg_tokens": null,
   "tps": 125.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-slpl5hbf-ctx4096",
   "model": "MiniCPM-V-4.7-1B",
   "notes": "oMLX community benchmark trace; peak memory 3.9 GB at this context; duplicate traces of the same session: i6emw9d1; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3485.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "MiniCPM-V-4.7-1B-8bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 8bit; 4k 3,485 116.1 3.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/slpl5hbf",
   "source_url": "https://omlx.ai/benchmarks/performance/slpl5hbf",
   "tg_tokens": null,
   "tps": 116.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-slpl5hbf-ctx16384",
   "model": "MiniCPM-V-4.7-1B",
   "notes": "oMLX community benchmark trace; peak memory 4.1 GB at this context; duplicate traces of the same session: i6emw9d1; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3132.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "MiniCPM-V-4.7-1B-8bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 8bit; 16k 3,132 113.6 4.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/slpl5hbf",
   "source_url": "https://omlx.ai/benchmarks/performance/slpl5hbf",
   "tg_tokens": null,
   "tps": 113.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-0v2vkvth-ctx4096",
   "model": "MiniCPM-V-4.7-1B",
   "notes": "oMLX community benchmark trace; peak memory 4.2 GB at this context; duplicate traces of the same session: mmcluy8l; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3802.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "MiniCPM-V-4.7-1B-bf16 on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 bf16; 4k 3,802 62.2 4.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0v2vkvth",
   "source_url": "https://omlx.ai/benchmarks/performance/0v2vkvth",
   "tg_tokens": null,
   "tps": 62.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-0v2vkvth-ctx16384",
   "model": "MiniCPM-V-4.7-1B",
   "notes": "oMLX community benchmark trace; peak memory 4.4 GB at this context; duplicate traces of the same session: mmcluy8l; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3119.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "MiniCPM-V-4.7-1B-bf16 on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 bf16; 16k 3,119 110.3 4.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0v2vkvth",
   "source_url": "https://omlx.ai/benchmarks/performance/0v2vkvth",
   "tg_tokens": null,
   "tps": 110.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-bhpdoeqd-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: 1uqhmmme, ux9q6r18; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1485.0,
   "provenance": "community",
   "quant": "MXFP8",
   "quote": "Qwen3.6-35B-A3B-MXFP8-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 unknown; 1k 1,485 58.8 37.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhpdoeqd",
   "source_url": "https://omlx.ai/benchmarks/performance/bhpdoeqd",
   "tg_tokens": null,
   "tps": 58.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-bhpdoeqd-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.9 GB at this context; duplicate traces of the same session: 1uqhmmme, ux9q6r18; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1573.0,
   "provenance": "community",
   "quant": "MXFP8",
   "quote": "Qwen3.6-35B-A3B-MXFP8-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 unknown; 8k 1,573 53.8 38.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhpdoeqd",
   "source_url": "https://omlx.ai/benchmarks/performance/bhpdoeqd",
   "tg_tokens": null,
   "tps": 53.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-bhpdoeqd-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: 1uqhmmme, ux9q6r18; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1383.0,
   "provenance": "community",
   "quant": "MXFP8",
   "quote": "Qwen3.6-35B-A3B-MXFP8-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 unknown; 16k 1,383 50.0 39.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhpdoeqd",
   "source_url": "https://omlx.ai/benchmarks/performance/bhpdoeqd",
   "tg_tokens": null,
   "tps": 50.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ba15zil5-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: w3v1vo7x; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 468.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 468.0 17.8 16.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ba15zil5",
   "source_url": "https://omlx.ai/benchmarks/performance/ba15zil5",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ba15zil5-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: w3v1vo7x; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 498.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 498.0 17.6 17.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ba15zil5",
   "source_url": "https://omlx.ai/benchmarks/performance/ba15zil5",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fhk9hh1q-ctx1024",
   "model": "Qwen3.8-27B-JANG_2D",
   "notes": "oMLX community benchmark trace; peak memory 14.9 GB at this context; duplicate traces of the same session: 9rfb77d5, hndze4am, rkjzdg0e; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 325.3,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Qwen3.8-27B-JANG_2D on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 2bit; 1k 325.3 23.8 14.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fhk9hh1q",
   "source_url": "https://omlx.ai/benchmarks/performance/fhk9hh1q",
   "tg_tokens": null,
   "tps": 23.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fhk9hh1q-ctx4096",
   "model": "Qwen3.8-27B-JANG_2D",
   "notes": "oMLX community benchmark trace; peak memory 16.7 GB at this context; duplicate traces of the same session: 9rfb77d5, hndze4am, rkjzdg0e; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 341.2,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Qwen3.8-27B-JANG_2D on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 2bit; 4k 341.2 29.2 16.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fhk9hh1q",
   "source_url": "https://omlx.ai/benchmarks/performance/fhk9hh1q",
   "tg_tokens": null,
   "tps": 29.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fhk9hh1q-ctx8192",
   "model": "Qwen3.8-27B-JANG_2D",
   "notes": "oMLX community benchmark trace; peak memory 17.2 GB at this context; duplicate traces of the same session: 9rfb77d5, hndze4am, rkjzdg0e; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 303.1,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Qwen3.8-27B-JANG_2D on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 2bit; 8k 303.1 21.3 17.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fhk9hh1q",
   "source_url": "https://omlx.ai/benchmarks/performance/fhk9hh1q",
   "tg_tokens": null,
   "tps": 21.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-fhk9hh1q-ctx16384",
   "model": "Qwen3.8-27B-JANG_2D",
   "notes": "oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: 9rfb77d5, hndze4am, rkjzdg0e; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 283.8,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Qwen3.8-27B-JANG_2D on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 2bit; 16k 283.8 25.8 18.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fhk9hh1q",
   "source_url": "https://omlx.ai/benchmarks/performance/fhk9hh1q",
   "tg_tokens": null,
   "tps": 25.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-797c879e-ctx1024",
   "model": "Qwen3.8-27B-JANG_4D",
   "notes": "oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: iswaralu, ka4w8mpx, la0heta8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 269.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-JANG_4D on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 269.2 19.8 20.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/797c879e",
   "source_url": "https://omlx.ai/benchmarks/performance/797c879e",
   "tg_tokens": null,
   "tps": 19.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-797c879e-ctx4096",
   "model": "Qwen3.8-27B-JANG_4D",
   "notes": "oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: iswaralu, ka4w8mpx, la0heta8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 302.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-JANG_4D on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 302.2 25.7 22.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/797c879e",
   "source_url": "https://omlx.ai/benchmarks/performance/797c879e",
   "tg_tokens": null,
   "tps": 25.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-797c879e-ctx8192",
   "model": "Qwen3.8-27B-JANG_4D",
   "notes": "oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: iswaralu, ka4w8mpx, la0heta8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 303.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-JANG_4D on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 8k 303.4 13.4 23.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/797c879e",
   "source_url": "https://omlx.ai/benchmarks/performance/797c879e",
   "tg_tokens": null,
   "tps": 13.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-797c879e-ctx16384",
   "model": "Qwen3.8-27B-JANG_4D",
   "notes": "oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: iswaralu, ka4w8mpx, la0heta8; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 275.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-JANG_4D on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 275.8 17.5 24.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/797c879e",
   "source_url": "https://omlx.ai/benchmarks/performance/797c879e",
   "tg_tokens": null,
   "tps": 17.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ijdy2hn5-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 27.6 GB; TTFT 8854 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 115.7,
   "provenance": "community",
   "quant": "mxfp8",
   "quote": "Qwen3.8-27B-mxfp8 on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k tokens 115.7 PP tok/s 12.2 TG tok/s 8854 TTFT (ms) 27.6 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ijdy2hn5",
   "source_url": "https://omlx.ai/benchmarks/performance/ijdy2hn5",
   "tg_tokens": null,
   "tps": 12.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-7rhrfr0k-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 22.5 GB; TTFT 15910 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 257.5,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k tokens 257.5 PP tok/s 54.7 TG tok/s 15910 TTFT (ms) 22.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7rhrfr0k",
   "source_url": "https://omlx.ai/benchmarks/performance/7rhrfr0k",
   "tg_tokens": null,
   "tps": 54.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-uy0zs1q0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.7 GB at this context; duplicate traces of the same session: lihdsdqn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 505.6,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-27B-oQ6-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 505.6 26.7 27.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uy0zs1q0",
   "source_url": "https://omlx.ai/benchmarks/performance/uy0zs1q0",
   "tg_tokens": null,
   "tps": 26.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-uy0zs1q0-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: lihdsdqn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 505.7,
   "provenance": "community",
   "quant": "oQ6",
   "quote": "Qwen3.8-27B-oQ6-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 8k 505.7 25.3 28.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uy0zs1q0",
   "source_url": "https://omlx.ai/benchmarks/performance/uy0zs1q0",
   "tg_tokens": null,
   "tps": 25.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-talwwoox-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build 20261011; oMLX community benchmark trace; peak memory 70.4 GB at this context; duplicate traces of the same session: 5aub91fp, yl1l5qe5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2042.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-20261011 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,042 67.5 70.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/talwwoox",
   "source_url": "https://omlx.ai/benchmarks/performance/talwwoox",
   "tg_tokens": null,
   "tps": 67.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-talwwoox-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build 20261011; oMLX community benchmark trace; peak memory 72.7 GB at this context; duplicate traces of the same session: 5aub91fp, yl1l5qe5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2891.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-20261011 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,891 66.3 72.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/talwwoox",
   "source_url": "https://omlx.ai/benchmarks/performance/talwwoox",
   "tg_tokens": null,
   "tps": 66.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-talwwoox-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build 20261011; oMLX community benchmark trace; peak memory 74.6 GB at this context; duplicate traces of the same session: 5aub91fp, yl1l5qe5; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2615.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp-20261011 on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,615 67.6 74.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/talwwoox",
   "source_url": "https://omlx.ai/benchmarks/performance/talwwoox",
   "tg_tokens": null,
   "tps": 67.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-7ghq6rv3-ctx1024",
   "model": "gemma-4-12b-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 13.6 GB at this context; duplicate traces of the same session: 375mc6rt, 9xh1olnl, my8xt6dj, xx47g3oa; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 280.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-12B-it-oQ8e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 1k 280.9 37.2 13.6 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ghq6rv3",
   "source_url": "https://omlx.ai/benchmarks/performance/7ghq6rv3",
   "tg_tokens": null,
   "tps": 37.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-7ghq6rv3-ctx4096",
   "model": "gemma-4-12b-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 14.1 GB at this context; duplicate traces of the same session: 375mc6rt, 9xh1olnl, my8xt6dj, xx47g3oa; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 279.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-12B-it-oQ8e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 4k 279.4 56.9 14.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ghq6rv3",
   "source_url": "https://omlx.ai/benchmarks/performance/7ghq6rv3",
   "tg_tokens": null,
   "tps": 56.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-7ghq6rv3-ctx8192",
   "model": "gemma-4-12b-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: 375mc6rt, 9xh1olnl, my8xt6dj, xx47g3oa; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 274.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-12B-it-oQ8e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 8k 274.2 32.3 14.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ghq6rv3",
   "source_url": "https://omlx.ai/benchmarks/performance/7ghq6rv3",
   "tg_tokens": null,
   "tps": 32.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-7ghq6rv3-ctx16384",
   "model": "gemma-4-12b-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 14.7 GB at this context; duplicate traces of the same session: 375mc6rt, 9xh1olnl, my8xt6dj, xx47g3oa; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 261.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-12B-it-oQ8e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 16k 261.9 46.4 14.7 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ghq6rv3",
   "source_url": "https://omlx.ai/benchmarks/performance/7ghq6rv3",
   "tg_tokens": null,
   "tps": 46.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-7ghq6rv3-ctx32768",
   "model": "gemma-4-12b-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 15.4 GB at this context; duplicate traces of the same session: 375mc6rt, 9xh1olnl, my8xt6dj, xx47g3oa; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 246.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-12B-it-oQ8e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 32k 246.6 16.9 15.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ghq6rv3",
   "source_url": "https://omlx.ai/benchmarks/performance/7ghq6rv3",
   "tg_tokens": null,
   "tps": 16.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-7ghq6rv3-ctx65536",
   "model": "gemma-4-12b-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: 375mc6rt, 9xh1olnl, my8xt6dj, xx47g3oa; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 217.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-12B-it-oQ8e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 64k 217.4 16.2 16.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ghq6rv3",
   "source_url": "https://omlx.ai/benchmarks/performance/7ghq6rv3",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-7ghq6rv3-ctx131072",
   "model": "gemma-4-12b-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.0 GB at this context; duplicate traces of the same session: 375mc6rt, 9xh1olnl, my8xt6dj, xx47g3oa; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 175.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-12B-it-oQ8e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 8bit; 128k 175.1 13.8 20.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ghq6rv3",
   "source_url": "https://omlx.ai/benchmarks/performance/7ghq6rv3",
   "tg_tokens": null,
   "tps": 13.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ldchh951-ctx1024",
   "model": "Codestral-22B-v0.1",
   "notes": "oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 1jpa6rud, 7vi3qkpk, mekywl0k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1006.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Codestral-22B-v0.1-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 1,006 21.1 22.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ldchh951",
   "source_url": "https://omlx.ai/benchmarks/performance/ldchh951",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ldchh951-ctx4096",
   "model": "Codestral-22B-v0.1",
   "notes": "oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 1jpa6rud, 7vi3qkpk, mekywl0k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 937.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Codestral-22B-v0.1-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 937.4 20.4 23.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ldchh951",
   "source_url": "https://omlx.ai/benchmarks/performance/ldchh951",
   "tg_tokens": null,
   "tps": 20.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ldchh951-ctx8192",
   "model": "Codestral-22B-v0.1",
   "notes": "oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: 1jpa6rud, 7vi3qkpk, mekywl0k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 757.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Codestral-22B-v0.1-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 757.2 19.6 24.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ldchh951",
   "source_url": "https://omlx.ai/benchmarks/performance/ldchh951",
   "tg_tokens": null,
   "tps": 19.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ldchh951-ctx16384",
   "model": "Codestral-22B-v0.1",
   "notes": "oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: 1jpa6rud, 7vi3qkpk, mekywl0k; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 605.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Codestral-22B-v0.1-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 605.9 18.4 26.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ldchh951",
   "source_url": "https://omlx.ai/benchmarks/performance/ldchh951",
   "tg_tokens": null,
   "tps": 18.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-h2f3d625-ctx4096",
   "model": "DeepSeek-V4-Pro-Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: 18f9tz1w; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 469.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-V4-Pro-Qwen3.5-9B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 24 GB \u00b7 4bit; 4k 469.0 47.8 6.4 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h2f3d625",
   "source_url": "https://omlx.ai/benchmarks/performance/h2f3d625",
   "tg_tokens": null,
   "tps": 47.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-h2f3d625-ctx16384",
   "model": "DeepSeek-V4-Pro-Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.8 GB at this context; duplicate traces of the same session: 18f9tz1w; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 449.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-V4-Pro-Qwen3.5-9B-4bit on M4 Pro (20c); M4 Pro (20c) \u00b7 24 GB \u00b7 4bit; 16k 449.3 43.7 6.8 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/h2f3d625",
   "source_url": "https://omlx.ai/benchmarks/performance/h2f3d625",
   "tg_tokens": null,
   "tps": 43.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-uir24zyb-ctx1024",
   "model": "MinerU2.5-2509-1.2B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 1.4 GB; TTFT 244 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4201.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "MinerU2.5-2509-1.2B on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 unknown; 1k tokens 4,201 PP tok/s 87.0 TG tok/s 244 TTFT (ms) 1.4 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uir24zyb",
   "source_url": "https://omlx.ai/benchmarks/performance/uir24zyb",
   "tg_tokens": null,
   "tps": 87.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-750nodjm-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: de2dydb4, wyorss15; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1527.0,
   "provenance": "community",
   "quant": "MXFP8",
   "quote": "Qwen3.6-35B-A3B-MXFP8-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 unknown; 1k 1,527 60.1 37.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/750nodjm",
   "source_url": "https://omlx.ai/benchmarks/performance/750nodjm",
   "tg_tokens": null,
   "tps": 60.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-750nodjm-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.9 GB at this context; duplicate traces of the same session: de2dydb4, wyorss15; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1667.0,
   "provenance": "community",
   "quant": "MXFP8",
   "quote": "Qwen3.6-35B-A3B-MXFP8-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 unknown; 8k 1,667 55.5 38.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/750nodjm",
   "source_url": "https://omlx.ai/benchmarks/performance/750nodjm",
   "tg_tokens": null,
   "tps": 55.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-750nodjm-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: de2dydb4, wyorss15; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1444.0,
   "provenance": "community",
   "quant": "MXFP8",
   "quote": "Qwen3.6-35B-A3B-MXFP8-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 unknown; 16k 1,444 50.3 39.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/750nodjm",
   "source_url": "https://omlx.ai/benchmarks/performance/750nodjm",
   "tg_tokens": null,
   "tps": 50.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-xy1z7blp-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 1k context; peak memory 37.4 GB; TTFT 756 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1354.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 8bit; 1k tokens 1,354 PP tok/s 108.6 TG tok/s 756 TTFT (ms) 37.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xy1z7blp",
   "source_url": "https://omlx.ai/benchmarks/performance/xy1z7blp",
   "tg_tokens": null,
   "tps": 108.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-28bghty8-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.4 GB at this context; duplicate traces of the same session: w8x0uozr; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 63.9,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-27B-oQ3.5e-mtp on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 3bit; 4k 63.9 12.7 17.4 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/28bghty8",
   "source_url": "https://omlx.ai/benchmarks/performance/28bghty8",
   "tg_tokens": null,
   "tps": 12.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-28bghty8-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.6 GB at this context; duplicate traces of the same session: w8x0uozr; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 57.1,
   "provenance": "community",
   "quant": "oQ3.5e",
   "quote": "Qwen3.8-27B-oQ3.5e-mtp on M4 (10c); M4 (10c) \u00b7 24 GB \u00b7 3bit; 16k 57.1 14.2 18.6 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/28bghty8",
   "source_url": "https://omlx.ai/benchmarks/performance/28bghty8",
   "tg_tokens": null,
   "tps": 14.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5mrbwdmg-ctx1024",
   "model": "Qwen3.8-Flash-Next-heretic-2",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: vihb21g8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 935.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-heretic-2-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 935.5 73.9 70.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5mrbwdmg",
   "source_url": "https://omlx.ai/benchmarks/performance/5mrbwdmg",
   "tg_tokens": null,
   "tps": 73.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5mrbwdmg-ctx4096",
   "model": "Qwen3.8-Flash-Next-heretic-2",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.2 GB at this context; duplicate traces of the same session: vihb21g8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2128.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-heretic-2-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,128 71.2 71.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5mrbwdmg",
   "source_url": "https://omlx.ai/benchmarks/performance/5mrbwdmg",
   "tg_tokens": null,
   "tps": 71.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-z55sksqg-ctx1024",
   "model": "Qwen3.8-Flash-Next-heretic-2",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: iyiq0x24; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 981.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-heretic-2-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 981.4 59.5 70.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z55sksqg",
   "source_url": "https://omlx.ai/benchmarks/performance/z55sksqg",
   "tg_tokens": null,
   "tps": 59.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-z55sksqg-ctx4096",
   "model": "Qwen3.8-Flash-Next-heretic-2",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.1 GB at this context; duplicate traces of the same session: iyiq0x24; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1924.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-heretic-2-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,924 64.1 71.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z55sksqg",
   "source_url": "https://omlx.ai/benchmarks/performance/z55sksqg",
   "tg_tokens": null,
   "tps": 64.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-fg2o1wdn-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "WLSRefit MTP build; oMLX community benchmark trace; peak memory 100.4 GB at this context; duplicate traces of the same session: 3r8kk0eg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1931.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-WLSRefit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,931 95.3 100.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fg2o1wdn",
   "source_url": "https://omlx.ai/benchmarks/performance/fg2o1wdn",
   "tg_tokens": null,
   "tps": 95.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-fg2o1wdn-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "WLSRefit MTP build; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: 3r8kk0eg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3037.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-WLSRefit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 3,037 83.1 101.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fg2o1wdn",
   "source_url": "https://omlx.ai/benchmarks/performance/fg2o1wdn",
   "tg_tokens": null,
   "tps": 83.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-c3aamv9a-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "WLSRefit MTP build; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: uptqsef0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1955.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-WLSRefit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,955 93.9 100.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c3aamv9a",
   "source_url": "https://omlx.ai/benchmarks/performance/c3aamv9a",
   "tg_tokens": null,
   "tps": 93.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-c3aamv9a-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "WLSRefit MTP build; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: uptqsef0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3017.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-WLSRefit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 3,017 96.4 101.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c3aamv9a",
   "source_url": "https://omlx.ai/benchmarks/performance/c3aamv9a",
   "tg_tokens": null,
   "tps": 96.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-08lqej8r-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "WLSRefit MTP build; oMLX community benchmark trace, single run at 1k context; peak memory 100.6 GB; TTFT 523 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1959.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-WLSRefit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,959 PP tok/s 99.3 TG tok/s 523 TTFT (ms) 100.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/08lqej8r",
   "source_url": "https://omlx.ai/benchmarks/performance/08lqej8r",
   "tg_tokens": null,
   "tps": 99.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-uxfwlflm-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "WLSRefit MTP build; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: q7xmuzui; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 890.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-WLSRefit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 890.5 84.1 100.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uxfwlflm",
   "source_url": "https://omlx.ai/benchmarks/performance/uxfwlflm",
   "tg_tokens": null,
   "tps": 84.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-uxfwlflm-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "WLSRefit MTP build; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: q7xmuzui; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2144.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-WLSRefit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 2,144 90.8 101.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uxfwlflm",
   "source_url": "https://omlx.ai/benchmarks/performance/uxfwlflm",
   "tg_tokens": null,
   "tps": 90.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wovlucoe-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "WLSRefit MTP build; oMLX community benchmark trace, single run at 1k context; peak memory 100.6 GB; TTFT 514 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1992.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-WLSRefit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 1,992 PP tok/s 98.5 TG tok/s 514 TTFT (ms) 100.6 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wovlucoe",
   "source_url": "https://omlx.ai/benchmarks/performance/wovlucoe",
   "tg_tokens": null,
   "tps": 98.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-adpt1ugr-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.6 GB at this context; duplicate traces of the same session: 1u1lm6a8, 4s8gbebu, imnmlgzr, j7gtcwok, rnsjncbg, tarriimd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1223.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,223 71.9 70.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/adpt1ugr",
   "source_url": "https://omlx.ai/benchmarks/performance/adpt1ugr",
   "tg_tokens": null,
   "tps": 71.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-adpt1ugr-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: 1u1lm6a8, 4s8gbebu, imnmlgzr, j7gtcwok, rnsjncbg, tarriimd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1982.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,982 65.4 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/adpt1ugr",
   "source_url": "https://omlx.ai/benchmarks/performance/adpt1ugr",
   "tg_tokens": null,
   "tps": 65.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-adpt1ugr-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.8 GB at this context; duplicate traces of the same session: 1u1lm6a8, 4s8gbebu, imnmlgzr, j7gtcwok, rnsjncbg, tarriimd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2430.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 2,430 85.2 71.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/adpt1ugr",
   "source_url": "https://omlx.ai/benchmarks/performance/adpt1ugr",
   "tg_tokens": null,
   "tps": 85.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-adpt1ugr-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 73.7 GB at this context; duplicate traces of the same session: 1u1lm6a8, 4s8gbebu, imnmlgzr, j7gtcwok, rnsjncbg, tarriimd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2587.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 2,587 81.7 73.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/adpt1ugr",
   "source_url": "https://omlx.ai/benchmarks/performance/adpt1ugr",
   "tg_tokens": null,
   "tps": 81.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-adpt1ugr-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 75.3 GB at this context; duplicate traces of the same session: 1u1lm6a8, 4s8gbebu, imnmlgzr, j7gtcwok, rnsjncbg, tarriimd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2504.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,504 80.7 75.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/adpt1ugr",
   "source_url": "https://omlx.ai/benchmarks/performance/adpt1ugr",
   "tg_tokens": null,
   "tps": 80.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-adpt1ugr-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 80.4 GB at this context; duplicate traces of the same session: 1u1lm6a8, 4s8gbebu, imnmlgzr, j7gtcwok, rnsjncbg, tarriimd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2415.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 2,415 72.6 80.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/adpt1ugr",
   "source_url": "https://omlx.ai/benchmarks/performance/adpt1ugr",
   "tg_tokens": null,
   "tps": 72.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-adpt1ugr-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 83.4 GB at this context; duplicate traces of the same session: 1u1lm6a8, 4s8gbebu, imnmlgzr, j7gtcwok, rnsjncbg, tarriimd; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2311.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 195k 2,311 62.8 83.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/adpt1ugr",
   "source_url": "https://omlx.ai/benchmarks/performance/adpt1ugr",
   "tg_tokens": null,
   "tps": 62.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ywbtyqzl-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "WLSRefit MTP build; oMLX community benchmark trace; peak memory 86.5 GB at this context; duplicate traces of the same session: m44z74t2, tq80pdqm, w69c4541; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1137.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-WLSRefit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 1,137 86.7 86.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ywbtyqzl",
   "source_url": "https://omlx.ai/benchmarks/performance/ywbtyqzl",
   "tg_tokens": null,
   "tps": 86.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ywbtyqzl-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "WLSRefit MTP build; oMLX community benchmark trace; peak memory 86.8 GB at this context; duplicate traces of the same session: m44z74t2, tq80pdqm, w69c4541; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1877.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-WLSRefit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 1,877 64.5 86.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ywbtyqzl",
   "source_url": "https://omlx.ai/benchmarks/performance/ywbtyqzl",
   "tg_tokens": null,
   "tps": 64.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ywbtyqzl-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "WLSRefit MTP build; oMLX community benchmark trace; peak memory 89.7 GB at this context; duplicate traces of the same session: m44z74t2, tq80pdqm, w69c4541; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2409.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-WLSRefit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 2,409 74.7 89.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ywbtyqzl",
   "source_url": "https://omlx.ai/benchmarks/performance/ywbtyqzl",
   "tg_tokens": null,
   "tps": 74.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ywbtyqzl-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "WLSRefit MTP build; oMLX community benchmark trace; peak memory 90.9 GB at this context; duplicate traces of the same session: m44z74t2, tq80pdqm, w69c4541; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2121.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-WLSRefit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 2,121 71.2 90.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ywbtyqzl",
   "source_url": "https://omlx.ai/benchmarks/performance/ywbtyqzl",
   "tg_tokens": null,
   "tps": 71.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ywbtyqzl-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "WLSRefit MTP build; oMLX community benchmark trace; peak memory 93.2 GB at this context; duplicate traces of the same session: m44z74t2, tq80pdqm, w69c4541; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1921.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-WLSRefit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 1,921 65.9 93.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ywbtyqzl",
   "source_url": "https://omlx.ai/benchmarks/performance/ywbtyqzl",
   "tg_tokens": null,
   "tps": 65.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-2jp8v0gp-ctx1024",
   "model": "deepseek-v41-lsq",
   "notes": "oMLX community benchmark trace; peak memory 219.9 GB at this context; duplicate traces of the same session: ebiq8wh0, ifq58s03, qluo75b0, rm6om9nx; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 518.3,
   "provenance": "community",
   "quant": "unknown",
   "quote": "deepseek-v41-lsq on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 unknown; 1k 518.3 62.3 219.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2jp8v0gp",
   "source_url": "https://omlx.ai/benchmarks/performance/2jp8v0gp",
   "tg_tokens": null,
   "tps": 62.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-2jp8v0gp-ctx4096",
   "model": "deepseek-v41-lsq",
   "notes": "oMLX community benchmark trace; peak memory 221.1 GB at this context; duplicate traces of the same session: ebiq8wh0, ifq58s03, qluo75b0, rm6om9nx; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 914.9,
   "provenance": "community",
   "quant": "unknown",
   "quote": "deepseek-v41-lsq on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 unknown; 4k 914.9 77.6 221.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2jp8v0gp",
   "source_url": "https://omlx.ai/benchmarks/performance/2jp8v0gp",
   "tg_tokens": null,
   "tps": 77.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-2jp8v0gp-ctx8192",
   "model": "deepseek-v41-lsq",
   "notes": "oMLX community benchmark trace; peak memory 223.4 GB at this context; duplicate traces of the same session: ebiq8wh0, ifq58s03, qluo75b0, rm6om9nx; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1313.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "deepseek-v41-lsq on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 unknown; 8k 1,313 80.6 223.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2jp8v0gp",
   "source_url": "https://omlx.ai/benchmarks/performance/2jp8v0gp",
   "tg_tokens": null,
   "tps": 80.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-2jp8v0gp-ctx16384",
   "model": "deepseek-v41-lsq",
   "notes": "oMLX community benchmark trace; peak memory 223.7 GB at this context; duplicate traces of the same session: ebiq8wh0, ifq58s03, qluo75b0, rm6om9nx; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2085.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "deepseek-v41-lsq on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 unknown; 16k 2,085 85.9 223.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2jp8v0gp",
   "source_url": "https://omlx.ai/benchmarks/performance/2jp8v0gp",
   "tg_tokens": null,
   "tps": 85.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-2jp8v0gp-ctx32768",
   "model": "deepseek-v41-lsq",
   "notes": "oMLX community benchmark trace; peak memory 223.8 GB at this context; duplicate traces of the same session: ebiq8wh0, ifq58s03, qluo75b0, rm6om9nx; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2106.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "deepseek-v41-lsq on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 unknown; 32k 2,106 85.4 223.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2jp8v0gp",
   "source_url": "https://omlx.ai/benchmarks/performance/2jp8v0gp",
   "tg_tokens": null,
   "tps": 85.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-1fwm0536-ctx1024",
   "model": "deepseek-v41-lsq",
   "notes": "oMLX community benchmark trace; peak memory 219.6 GB at this context; duplicate traces of the same session: 0du8ub0a, 4x4sb9tw, w5c1yigt; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 885.9,
   "provenance": "community",
   "quant": "unknown",
   "quote": "deepseek-v41-lsq on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 unknown; 1k 885.9 65.8 219.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1fwm0536",
   "source_url": "https://omlx.ai/benchmarks/performance/1fwm0536",
   "tg_tokens": null,
   "tps": 65.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-1fwm0536-ctx4096",
   "model": "deepseek-v41-lsq",
   "notes": "oMLX community benchmark trace; peak memory 221.1 GB at this context; duplicate traces of the same session: 0du8ub0a, 4x4sb9tw, w5c1yigt; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 767.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "deepseek-v41-lsq on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 unknown; 4k 767.0 74.5 221.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1fwm0536",
   "source_url": "https://omlx.ai/benchmarks/performance/1fwm0536",
   "tg_tokens": null,
   "tps": 74.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-1fwm0536-ctx8192",
   "model": "deepseek-v41-lsq",
   "notes": "oMLX community benchmark trace; peak memory 223.4 GB at this context; duplicate traces of the same session: 0du8ub0a, 4x4sb9tw, w5c1yigt; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1201.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "deepseek-v41-lsq on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 unknown; 8k 1,201 79.8 223.4 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1fwm0536",
   "source_url": "https://omlx.ai/benchmarks/performance/1fwm0536",
   "tg_tokens": null,
   "tps": 79.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-1fwm0536-ctx16384",
   "model": "deepseek-v41-lsq",
   "notes": "oMLX community benchmark trace; peak memory 223.7 GB at this context; duplicate traces of the same session: 0du8ub0a, 4x4sb9tw, w5c1yigt; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2072.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "deepseek-v41-lsq on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 unknown; 16k 2,072 92.5 223.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1fwm0536",
   "source_url": "https://omlx.ai/benchmarks/performance/1fwm0536",
   "tg_tokens": null,
   "tps": 92.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-epvffjww-ctx1024",
   "model": "Codestral-22B-v0.1",
   "notes": "oMLX community benchmark trace; peak memory 22.7 GB at this context; duplicate traces of the same session: 6my2un0g, ebjv3nyu, kanz5dc0, u9xuoiyg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1000.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Codestral-22B-v0.1-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 1,000 21.1 22.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/epvffjww",
   "source_url": "https://omlx.ai/benchmarks/performance/epvffjww",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-epvffjww-ctx4096",
   "model": "Codestral-22B-v0.1",
   "notes": "oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 6my2un0g, ebjv3nyu, kanz5dc0, u9xuoiyg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 802.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Codestral-22B-v0.1-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 802.7 20.4 23.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/epvffjww",
   "source_url": "https://omlx.ai/benchmarks/performance/epvffjww",
   "tg_tokens": null,
   "tps": 20.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-epvffjww-ctx8192",
   "model": "Codestral-22B-v0.1",
   "notes": "oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: 6my2un0g, ebjv3nyu, kanz5dc0, u9xuoiyg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 723.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Codestral-22B-v0.1-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 723.3 19.5 24.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/epvffjww",
   "source_url": "https://omlx.ai/benchmarks/performance/epvffjww",
   "tg_tokens": null,
   "tps": 19.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-epvffjww-ctx16384",
   "model": "Codestral-22B-v0.1",
   "notes": "oMLX community benchmark trace; peak memory 26.0 GB at this context; duplicate traces of the same session: 6my2un0g, ebjv3nyu, kanz5dc0, u9xuoiyg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 611.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Codestral-22B-v0.1-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 611.4 15.7 26.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/epvffjww",
   "source_url": "https://omlx.ai/benchmarks/performance/epvffjww",
   "tg_tokens": null,
   "tps": 15.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-epvffjww-ctx32768",
   "model": "Codestral-22B-v0.1",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: 6my2un0g, ebjv3nyu, kanz5dc0, u9xuoiyg; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 393.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Codestral-22B-v0.1-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 393.6 9.4 29.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/epvffjww",
   "source_url": "https://omlx.ai/benchmarks/performance/epvffjww",
   "tg_tokens": null,
   "tps": 9.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-r23tc5oh-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 224.9 GB at this context; duplicate traces of the same session: l0tdlasy, m3fcs4ds, wzgqqirr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 37.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 1k 37.7 10.5 224.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r23tc5oh",
   "source_url": "https://omlx.ai/benchmarks/performance/r23tc5oh",
   "tg_tokens": null,
   "tps": 10.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-r23tc5oh-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 223.2 GB at this context; duplicate traces of the same session: l0tdlasy, m3fcs4ds, wzgqqirr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 147.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 4k 147.5 18.3 223.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r23tc5oh",
   "source_url": "https://omlx.ai/benchmarks/performance/r23tc5oh",
   "tg_tokens": null,
   "tps": 18.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-r23tc5oh-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 223.2 GB at this context; duplicate traces of the same session: l0tdlasy, m3fcs4ds, wzgqqirr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 169.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 8k 169.3 12.7 223.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r23tc5oh",
   "source_url": "https://omlx.ai/benchmarks/performance/r23tc5oh",
   "tg_tokens": null,
   "tps": 12.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-r23tc5oh-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 223.2 GB at this context; duplicate traces of the same session: l0tdlasy, m3fcs4ds, wzgqqirr; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 145.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 256 GB \u00b7 unknown; 16k 145.8 14.3 223.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r23tc5oh",
   "source_url": "https://omlx.ai/benchmarks/performance/r23tc5oh",
   "tg_tokens": null,
   "tps": 14.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-ccnx20xy-ctx16384",
   "model": "MiniCPM-V-4.7-1B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 4.1 GB; TTFT 4156 ms; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3943.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "MiniCPM-V-4.7-1B-8bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 8bit; 16k tokens 3,943 PP tok/s 207.6 TG tok/s 4156 TTFT (ms) 4.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ccnx20xy",
   "source_url": "https://omlx.ai/benchmarks/performance/ccnx20xy",
   "tg_tokens": null,
   "tps": 207.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3h4g63iw-ctx1024",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 35.6 GB at this context; duplicate traces of the same session: b26x2gs0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2781.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 2,781 88.4 35.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3h4g63iw",
   "source_url": "https://omlx.ai/benchmarks/performance/3h4g63iw",
   "tg_tokens": null,
   "tps": 88.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-3h4g63iw-ctx4096",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: b26x2gs0; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3362.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 3,362 87.6 37.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3h4g63iw",
   "source_url": "https://omlx.ai/benchmarks/performance/3h4g63iw",
   "tg_tokens": null,
   "tps": 87.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-j3i5fria-ctx1024",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 35.6 GB at this context; duplicate traces of the same session: xxdygfta; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2755.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 2,755 87.8 35.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j3i5fria",
   "source_url": "https://omlx.ai/benchmarks/performance/j3i5fria",
   "tg_tokens": null,
   "tps": 87.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-j3i5fria-ctx4096",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: xxdygfta; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3341.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 3,341 87.5 37.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j3i5fria",
   "source_url": "https://omlx.ai/benchmarks/performance/j3i5fria",
   "tg_tokens": null,
   "tps": 87.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-vem0r53h-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: 1ywuj8v3, n23wyjvk; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 752.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-6bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 1k 752.1 65.6 27.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vem0r53h",
   "source_url": "https://omlx.ai/benchmarks/performance/vem0r53h",
   "tg_tokens": null,
   "tps": 65.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-vem0r53h-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: 1ywuj8v3, n23wyjvk; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 812.1,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-6bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 4k 812.1 65.1 28.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vem0r53h",
   "source_url": "https://omlx.ai/benchmarks/performance/vem0r53h",
   "tg_tokens": null,
   "tps": 65.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-vem0r53h-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 29.0 GB at this context; duplicate traces of the same session: 1ywuj8v3, n23wyjvk; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 576.6,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-6bit on M4 Pro (16c); M4 Pro (16c) \u00b7 48 GB \u00b7 6bit; 32k 576.6 49.3 29.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vem0r53h",
   "source_url": "https://omlx.ai/benchmarks/performance/vem0r53h",
   "tg_tokens": null,
   "tps": 49.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-oday4ujb-ctx1024",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 30.8 GB at this context; duplicate traces of the same session: 5huyd91y, g2zav980, l2afd9v1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4754.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 4,754 93.7 30.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oday4ujb",
   "source_url": "https://omlx.ai/benchmarks/performance/oday4ujb",
   "tg_tokens": null,
   "tps": 93.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-oday4ujb-ctx8192",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 31.6 GB at this context; duplicate traces of the same session: 5huyd91y, g2zav980, l2afd9v1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4107.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 4,107 82.2 31.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oday4ujb",
   "source_url": "https://omlx.ai/benchmarks/performance/oday4ujb",
   "tg_tokens": null,
   "tps": 82.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-oday4ujb-ctx16384",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 32.4 GB at this context; duplicate traces of the same session: 5huyd91y, g2zav980, l2afd9v1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2918.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 2,918 73.3 32.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oday4ujb",
   "source_url": "https://omlx.ai/benchmarks/performance/oday4ujb",
   "tg_tokens": null,
   "tps": 73.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-oday4ujb-ctx32768",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 33.8 GB at this context; duplicate traces of the same session: 5huyd91y, g2zav980, l2afd9v1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1737.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 1,737 55.6 33.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oday4ujb",
   "source_url": "https://omlx.ai/benchmarks/performance/oday4ujb",
   "tg_tokens": null,
   "tps": 55.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-z36v4xxk-ctx1024",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 30.8 GB at this context; duplicate traces of the same session: abssc54y, jby5m9j1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4766.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 4,766 96.1 30.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z36v4xxk",
   "source_url": "https://omlx.ai/benchmarks/performance/z36v4xxk",
   "tg_tokens": null,
   "tps": 96.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-z36v4xxk-ctx8192",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 31.6 GB at this context; duplicate traces of the same session: abssc54y, jby5m9j1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4844.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 4,844 82.6 31.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z36v4xxk",
   "source_url": "https://omlx.ai/benchmarks/performance/z36v4xxk",
   "tg_tokens": null,
   "tps": 82.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-z36v4xxk-ctx16384",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 32.4 GB at this context; duplicate traces of the same session: abssc54y, jby5m9j1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3114.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 3,114 73.5 32.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z36v4xxk",
   "source_url": "https://omlx.ai/benchmarks/performance/z36v4xxk",
   "tg_tokens": null,
   "tps": 73.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-j83q8lg3-ctx1024",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 30.8 GB at this context; duplicate traces of the same session: d9mw0r31, yqo14y6d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4793.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 4,793 96.5 30.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j83q8lg3",
   "source_url": "https://omlx.ai/benchmarks/performance/j83q8lg3",
   "tg_tokens": null,
   "tps": 96.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-j83q8lg3-ctx8192",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 31.6 GB at this context; duplicate traces of the same session: d9mw0r31, yqo14y6d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4651.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 4,651 82.8 31.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j83q8lg3",
   "source_url": "https://omlx.ai/benchmarks/performance/j83q8lg3",
   "tg_tokens": null,
   "tps": 82.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-j83q8lg3-ctx16384",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 32.4 GB at this context; duplicate traces of the same session: d9mw0r31, yqo14y6d; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2993.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 2,993 73.3 32.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/j83q8lg3",
   "source_url": "https://omlx.ai/benchmarks/performance/j83q8lg3",
   "tg_tokens": null,
   "tps": 73.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wwe299oq-ctx1024",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 30.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3849.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 3,849 96.2 30.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wwe299oq",
   "source_url": "https://omlx.ai/benchmarks/performance/wwe299oq",
   "tg_tokens": null,
   "tps": 96.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-wwe299oq-ctx4096",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 31.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4232.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 4,232 90.0 31.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wwe299oq",
   "source_url": "https://omlx.ai/benchmarks/performance/wwe299oq",
   "tg_tokens": null,
   "tps": 90.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4j9h6h73-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 52.4 GB at this context; duplicate traces of the same session: 6geed40n, 6wjwhy08, pbrxmpjj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1826.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3-Coder-Next-5bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 1,826 91.0 52.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4j9h6h73",
   "source_url": "https://omlx.ai/benchmarks/performance/4j9h6h73",
   "tg_tokens": null,
   "tps": 91.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4j9h6h73-ctx4096",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 53.2 GB at this context; duplicate traces of the same session: 6geed40n, 6wjwhy08, pbrxmpjj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3204.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3-Coder-Next-5bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 3,204 91.5 53.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4j9h6h73",
   "source_url": "https://omlx.ai/benchmarks/performance/4j9h6h73",
   "tg_tokens": null,
   "tps": 91.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4j9h6h73-ctx16384",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 53.5 GB at this context; duplicate traces of the same session: 6geed40n, 6wjwhy08, pbrxmpjj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2682.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3-Coder-Next-5bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 16k 2,682 86.8 53.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4j9h6h73",
   "source_url": "https://omlx.ai/benchmarks/performance/4j9h6h73",
   "tg_tokens": null,
   "tps": 86.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-4j9h6h73-ctx32768",
   "model": "Qwen3-Coder-Next",
   "notes": "oMLX community benchmark trace; peak memory 53.9 GB at this context; duplicate traces of the same session: 6geed40n, 6wjwhy08, pbrxmpjj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2263.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "Qwen3-Coder-Next-5bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 2,263 79.7 53.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4j9h6h73",
   "source_url": "https://omlx.ai/benchmarks/performance/4j9h6h73",
   "tg_tokens": null,
   "tps": 79.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-stt0mjoc-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: 09d0q0yw, 5sxrn12k; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 169.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 1k 169.0 23.6 17.0 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/stt0mjoc",
   "source_url": "https://omlx.ai/benchmarks/performance/stt0mjoc",
   "tg_tokens": null,
   "tps": 23.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-stt0mjoc-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: 09d0q0yw, 5sxrn12k; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 165.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 4k 165.4 35.0 18.1 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/stt0mjoc",
   "source_url": "https://omlx.ai/benchmarks/performance/stt0mjoc",
   "tg_tokens": null,
   "tps": 35.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-stt0mjoc-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: 09d0q0yw, 5sxrn12k; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 152.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 8k 152.3 30.9 18.5 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/stt0mjoc",
   "source_url": "https://omlx.ai/benchmarks/performance/stt0mjoc",
   "tg_tokens": null,
   "tps": 30.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-7cr9nw67-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "WLSRefit MTP build; oMLX community benchmark trace; peak memory 86.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1137.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-WLSRefit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 1,137 86.7 86.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7cr9nw67",
   "source_url": "https://omlx.ai/benchmarks/performance/7cr9nw67",
   "tg_tokens": null,
   "tps": 86.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7cr9nw67-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "WLSRefit MTP build; oMLX community benchmark trace; peak memory 86.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1877.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-WLSRefit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 1,877 64.5 86.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7cr9nw67",
   "source_url": "https://omlx.ai/benchmarks/performance/7cr9nw67",
   "tg_tokens": null,
   "tps": 64.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-7cr9nw67-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "WLSRefit MTP build; oMLX community benchmark trace; peak memory 89.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2409.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-WLSRefit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 2,409 74.7 89.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7cr9nw67",
   "source_url": "https://omlx.ai/benchmarks/performance/7cr9nw67",
   "tg_tokens": null,
   "tps": 74.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7cr9nw67-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "WLSRefit MTP build; oMLX community benchmark trace; peak memory 90.9 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2121.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-WLSRefit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 2,121 71.2 90.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7cr9nw67",
   "source_url": "https://omlx.ai/benchmarks/performance/7cr9nw67",
   "tg_tokens": null,
   "tps": 71.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7cr9nw67-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "WLSRefit MTP build; oMLX community benchmark trace; peak memory 93.2 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1921.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-WLSRefit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 1,921 65.9 93.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7cr9nw67",
   "source_url": "https://omlx.ai/benchmarks/performance/7cr9nw67",
   "tg_tokens": null,
   "tps": 65.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-onhzmt1o-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "WLSRefit MTP build; oMLX community benchmark trace; peak memory 86.5 GB at this context; duplicate traces of the same session: 23lu4p9p, 57q67vzl, 5kilwq6j, xbcg94kl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 908.8,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-WLSRefit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 1k 908.8 63.6 86.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/onhzmt1o",
   "source_url": "https://omlx.ai/benchmarks/performance/onhzmt1o",
   "tg_tokens": null,
   "tps": 63.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-onhzmt1o-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "WLSRefit MTP build; oMLX community benchmark trace; peak memory 86.8 GB at this context; duplicate traces of the same session: 23lu4p9p, 57q67vzl, 5kilwq6j, xbcg94kl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1745.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-WLSRefit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 1,745 66.0 86.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/onhzmt1o",
   "source_url": "https://omlx.ai/benchmarks/performance/onhzmt1o",
   "tg_tokens": null,
   "tps": 66.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-onhzmt1o-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "WLSRefit MTP build; oMLX community benchmark trace; peak memory 89.7 GB at this context; duplicate traces of the same session: 23lu4p9p, 57q67vzl, 5kilwq6j, xbcg94kl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2368.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-WLSRefit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 32k 2,368 74.0 89.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/onhzmt1o",
   "source_url": "https://omlx.ai/benchmarks/performance/onhzmt1o",
   "tg_tokens": null,
   "tps": 74.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-onhzmt1o-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "WLSRefit MTP build; oMLX community benchmark trace; peak memory 90.9 GB at this context; duplicate traces of the same session: 23lu4p9p, 57q67vzl, 5kilwq6j, xbcg94kl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2144.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-WLSRefit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 64k 2,144 73.9 90.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/onhzmt1o",
   "source_url": "https://omlx.ai/benchmarks/performance/onhzmt1o",
   "tg_tokens": null,
   "tps": 73.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-onhzmt1o-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "WLSRefit MTP build; oMLX community benchmark trace; peak memory 93.2 GB at this context; duplicate traces of the same session: 23lu4p9p, 57q67vzl, 5kilwq6j, xbcg94kl; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2003.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-oQ5e-WLSRefit-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 128k 2,003 75.2 93.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/onhzmt1o",
   "source_url": "https://omlx.ai/benchmarks/performance/onhzmt1o",
   "tg_tokens": null,
   "tps": 75.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-nmxmxg50-ctx1024",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.3 GB at this context; duplicate traces of the same session: hh90j0t1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2381.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 2,381 103.3 14.3 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nmxmxg50",
   "source_url": "https://omlx.ai/benchmarks/performance/nmxmxg50",
   "tg_tokens": null,
   "tps": 103.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-nmxmxg50-ctx4096",
   "model": "gemma4 26B.A4B",
   "notes": "QAT build; oMLX community benchmark trace; peak memory 14.6 GB at this context; duplicate traces of the same session: hh90j0t1; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2412.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-26B-A4B-it-QAT-MLX-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 2,412 96.9 14.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nmxmxg50",
   "source_url": "https://omlx.ai/benchmarks/performance/nmxmxg50",
   "tg_tokens": null,
   "tps": 96.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-8loyhpe1-ctx1024",
   "model": "gemma-4-26B-A4B-it-qat-JANG_4M",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: jztvy9a7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 799.5,
   "provenance": "community",
   "quant": "unknown",
   "quote": "gemma-4-26B-A4B-it-qat-JANG_4M on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 unknown; 1k 799.5 54.4 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8loyhpe1",
   "source_url": "https://omlx.ai/benchmarks/performance/8loyhpe1",
   "tg_tokens": null,
   "tps": 54.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-8loyhpe1-ctx4096",
   "model": "gemma-4-26B-A4B-it-qat-JANG_4M",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: jztvy9a7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2175.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "gemma-4-26B-A4B-it-qat-JANG_4M on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 unknown; 4k 2,175 53.3 17.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8loyhpe1",
   "source_url": "https://omlx.ai/benchmarks/performance/8loyhpe1",
   "tg_tokens": null,
   "tps": 53.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-8loyhpe1-ctx8192",
   "model": "gemma-4-26B-A4B-it-qat-JANG_4M",
   "notes": "oMLX community benchmark trace; peak memory 17.5 GB at this context; duplicate traces of the same session: jztvy9a7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2250.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "gemma-4-26B-A4B-it-qat-JANG_4M on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 unknown; 8k 2,250 52.1 17.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8loyhpe1",
   "source_url": "https://omlx.ai/benchmarks/performance/8loyhpe1",
   "tg_tokens": null,
   "tps": 52.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-8loyhpe1-ctx16384",
   "model": "gemma-4-26B-A4B-it-qat-JANG_4M",
   "notes": "oMLX community benchmark trace; peak memory 17.9 GB at this context; duplicate traces of the same session: jztvy9a7; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2006.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "gemma-4-26B-A4B-it-qat-JANG_4M on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 unknown; 16k 2,006 49.4 17.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8loyhpe1",
   "source_url": "https://omlx.ai/benchmarks/performance/8loyhpe1",
   "tg_tokens": null,
   "tps": 49.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-7e37512q-ctx1024",
   "model": "DeepSeek-Coder-V2-Lite-Instruct",
   "notes": "MLX; oMLX community benchmark trace, single run at 1k context; peak memory 9.0 GB; TTFT 1741 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 588.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-Coder-V2-Lite-Instruct-4bit-mlx on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k tokens 588.2 PP tok/s 54.5 TG tok/s 1741 TTFT (ms) 9.0 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7e37512q",
   "source_url": "https://omlx.ai/benchmarks/performance/7e37512q",
   "tg_tokens": null,
   "tps": 54.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-68392r8z-ctx4096",
   "model": "MiniCPM-V-4.7-1B",
   "notes": "oMLX community benchmark trace; peak memory 3.8 GB at this context; duplicate traces of the same session: bnr7ikh9; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4997.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM-V-4.7-1B-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 4k 4,997 239.2 3.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/68392r8z",
   "source_url": "https://omlx.ai/benchmarks/performance/68392r8z",
   "tg_tokens": null,
   "tps": 239.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-68392r8z-ctx16384",
   "model": "MiniCPM-V-4.7-1B",
   "notes": "oMLX community benchmark trace; peak memory 4.0 GB at this context; duplicate traces of the same session: bnr7ikh9; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3830.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM-V-4.7-1B-4bit on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 16k 3,830 219.7 4.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/68392r8z",
   "source_url": "https://omlx.ai/benchmarks/performance/68392r8z",
   "tg_tokens": null,
   "tps": 219.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-ekebk3zk-ctx16384",
   "model": "MiniCPM-V-4.7-1B",
   "notes": "oMLX community benchmark trace, single run at 16k context; peak memory 4.4 GB; TTFT 4025 ms; duplicate traces of the same session: none; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4070.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "MiniCPM-V-4.7-1B-bf16 on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 bf16; 16k tokens 4,070 PP tok/s 188.7 TG tok/s 4025 TTFT (ms) 4.4 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ekebk3zk",
   "source_url": "https://omlx.ai/benchmarks/performance/ekebk3zk",
   "tg_tokens": null,
   "tps": 188.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-e4hjnsva-ctx1024",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 66.7 GB at this context; duplicate traces of the same session: f430ofb3, u5z3xqnz; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 720.6,
   "provenance": "community",
   "quant": "bf16",
   "quote": "MiniCPM-V-4.7-35B-A3B-bf16 on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 bf16; 1k 720.6 52.5 66.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e4hjnsva",
   "source_url": "https://omlx.ai/benchmarks/performance/e4hjnsva",
   "tg_tokens": null,
   "tps": 52.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-e4hjnsva-ctx4096",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 67.4 GB at this context; duplicate traces of the same session: f430ofb3, u5z3xqnz; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1190.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "MiniCPM-V-4.7-35B-A3B-bf16 on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 bf16; 4k 1,190 51.6 67.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e4hjnsva",
   "source_url": "https://omlx.ai/benchmarks/performance/e4hjnsva",
   "tg_tokens": null,
   "tps": 51.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-e4hjnsva-ctx16384",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 67.7 GB at this context; duplicate traces of the same session: f430ofb3, u5z3xqnz; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1044.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "MiniCPM-V-4.7-35B-A3B-bf16 on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 bf16; 16k 1,044 48.1 67.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e4hjnsva",
   "source_url": "https://omlx.ai/benchmarks/performance/e4hjnsva",
   "tg_tokens": null,
   "tps": 48.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-0x43yubx-ctx1024",
   "model": "Ornith-1.5-35B-A3B-uncensored",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: oncdwqx6, v2kh32uu; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 473.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-uncensored-MLX-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k 473.6 47.7 19.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0x43yubx",
   "source_url": "https://omlx.ai/benchmarks/performance/0x43yubx",
   "tg_tokens": null,
   "tps": 47.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-0x43yubx-ctx8192",
   "model": "Ornith-1.5-35B-A3B-uncensored",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: oncdwqx6, v2kh32uu; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 492.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-uncensored-MLX-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 8k 492.5 44.2 20.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0x43yubx",
   "source_url": "https://omlx.ai/benchmarks/performance/0x43yubx",
   "tg_tokens": null,
   "tps": 44.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-0x43yubx-ctx32768",
   "model": "Ornith-1.5-35B-A3B-uncensored",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 21.0 GB at this context; duplicate traces of the same session: oncdwqx6, v2kh32uu; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 363.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-35B-A3B-uncensored-MLX-4bit on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 32k 363.8 33.3 21.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0x43yubx",
   "source_url": "https://omlx.ai/benchmarks/performance/0x43yubx",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bg0nory2-ctx1024",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 30.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3849.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 3,849 96.2 30.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bg0nory2",
   "source_url": "https://omlx.ai/benchmarks/performance/bg0nory2",
   "tg_tokens": null,
   "tps": 96.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-bg0nory2-ctx4096",
   "model": "Qwen3-Coder-30B-A3B-Instruct",
   "notes": "oMLX community benchmark trace; peak memory 31.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4232.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-30B-A3B-Instruct-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 4,232 90.0 31.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bg0nory2",
   "source_url": "https://omlx.ai/benchmarks/performance/bg0nory2",
   "tg_tokens": null,
   "tps": 90.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-ebemdo1t-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 10.3 GB at this context; duplicate traces of the same session: yjotgw5a; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 416.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B-MLX-8bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 8bit; 4k 416.0 12.4 10.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ebemdo1t",
   "source_url": "https://omlx.ai/benchmarks/performance/ebemdo1t",
   "tg_tokens": null,
   "tps": 12.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5",
   "id": "omlx-ebemdo1t-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 10.6 GB at this context; duplicate traces of the same session: yjotgw5a; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 441.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.5-9B-MLX-8bit on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 8bit; 16k 441.8 11.7 10.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ebemdo1t",
   "source_url": "https://omlx.ai/benchmarks/performance/ebemdo1t",
   "tg_tokens": null,
   "tps": 11.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-abifgeb7-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: 03idbbj6, hs8p46xx; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 223.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-9B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k 223.0 35.0 7.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/abifgeb7",
   "source_url": "https://omlx.ai/benchmarks/performance/abifgeb7",
   "tg_tokens": null,
   "tps": 35.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-abifgeb7-ctx8192",
   "model": "Qwen3.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.7 GB at this context; duplicate traces of the same session: 03idbbj6, hs8p46xx; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 227.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-9B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 8k 227.2 37.7 7.7 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/abifgeb7",
   "source_url": "https://omlx.ai/benchmarks/performance/abifgeb7",
   "tg_tokens": null,
   "tps": 37.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-abifgeb7-ctx32768",
   "model": "Qwen3.5-9B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 9.2 GB at this context; duplicate traces of the same session: 03idbbj6, hs8p46xx; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 200.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.5-9B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 32k 200.4 33.3 9.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/abifgeb7",
   "source_url": "https://omlx.ai/benchmarks/performance/abifgeb7",
   "tg_tokens": null,
   "tps": 33.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-rakigg33-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 36.5 GB at this context; duplicate traces of the same session: 5wjniat1, cr812iau; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1536.0,
   "provenance": "community",
   "quant": "MXFP8",
   "quote": "Qwen3.6-35B-A3B-MXFP8-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 unknown; 1k 1,536 60.1 36.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rakigg33",
   "source_url": "https://omlx.ai/benchmarks/performance/rakigg33",
   "tg_tokens": null,
   "tps": 60.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-rakigg33-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.1 GB at this context; duplicate traces of the same session: 5wjniat1, cr812iau; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1654.0,
   "provenance": "community",
   "quant": "MXFP8",
   "quote": "Qwen3.6-35B-A3B-MXFP8-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 unknown; 8k 1,654 55.5 38.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rakigg33",
   "source_url": "https://omlx.ai/benchmarks/performance/rakigg33",
   "tg_tokens": null,
   "tps": 55.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-rakigg33-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 38.5 GB at this context; duplicate traces of the same session: 5wjniat1, cr812iau; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1440.0,
   "provenance": "community",
   "quant": "MXFP8",
   "quote": "Qwen3.6-35B-A3B-MXFP8-MTP on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 unknown; 16k 1,440 51.7 38.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rakigg33",
   "source_url": "https://omlx.ai/benchmarks/performance/rakigg33",
   "tg_tokens": null,
   "tps": 51.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-dn3mm4x4-ctx1024",
   "model": "Qwen3.8-27B-aeon-ultimate-uncensored",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 15.8 GB at this context; duplicate traces of the same session: qcf43f0q; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 139.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-4bit-MLX on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k 139.7 6.8 15.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dn3mm4x4",
   "source_url": "https://omlx.ai/benchmarks/performance/dn3mm4x4",
   "tg_tokens": null,
   "tps": 6.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-dn3mm4x4-ctx4096",
   "model": "Qwen3.8-27B-aeon-ultimate-uncensored",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: qcf43f0q; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 140.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-4bit-MLX on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 140.1 6.6 17.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dn3mm4x4",
   "source_url": "https://omlx.ai/benchmarks/performance/dn3mm4x4",
   "tg_tokens": null,
   "tps": 6.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-vqj1nt26-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: nxy4xn5a; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 403.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 403.5 18.2 16.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vqj1nt26",
   "source_url": "https://omlx.ai/benchmarks/performance/vqj1nt26",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-vqj1nt26-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: nxy4xn5a; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 419.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 419.1 17.9 17.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vqj1nt26",
   "source_url": "https://omlx.ai/benchmarks/performance/vqj1nt26",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-vc8ittus-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: 61rlexh9; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 414.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 414.1 28.5 16.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vc8ittus",
   "source_url": "https://omlx.ai/benchmarks/performance/vc8ittus",
   "tg_tokens": null,
   "tps": 28.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-vc8ittus-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: 61rlexh9; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 403.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 403.0 26.0 18.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vc8ittus",
   "source_url": "https://omlx.ai/benchmarks/performance/vc8ittus",
   "tg_tokens": null,
   "tps": 26.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-rtuv1ax0-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 385.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 385.4 17.9 16.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rtuv1ax0",
   "source_url": "https://omlx.ai/benchmarks/performance/rtuv1ax0",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-rtuv1ax0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 418.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 418.3 17.8 17.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rtuv1ax0",
   "source_url": "https://omlx.ai/benchmarks/performance/rtuv1ax0",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-8ma03c02-ctx65536",
   "model": "Qwen3.8-27B-Uncensored-HauhauCS-Aggressive",
   "notes": "MTPLX build; oMLX community benchmark trace, single run at 64k context; peak memory 28.5 GB; TTFT 182719 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 358.7,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTPLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 64k tokens 358.7 PP tok/s 21.1 TG tok/s 182719 TTFT (ms) 28.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8ma03c02",
   "source_url": "https://omlx.ai/benchmarks/performance/8ma03c02",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-jm6qmo0q-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 31.7 GB; TTFT 265615 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 493.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 128k tokens 493.5 PP tok/s 15.1 TG tok/s 265615 TTFT (ms) 31.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jm6qmo0q",
   "source_url": "https://omlx.ai/benchmarks/performance/jm6qmo0q",
   "tg_tokens": null,
   "tps": 15.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-ae1qew88-ctx1024",
   "model": "Qwen3.8-9B-Distill-Uncensored-Heretic",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-9B-Distill-uncensored-heretic-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k 224.0 32.6 7.0 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ae1qew88",
   "source_url": "https://omlx.ai/benchmarks/performance/ae1qew88",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-ae1qew88-ctx8192",
   "model": "Qwen3.8-9B-Distill-Uncensored-Heretic",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.7 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 229.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-9B-Distill-uncensored-heretic-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 8k 229.3 40.8 7.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ae1qew88",
   "source_url": "https://omlx.ai/benchmarks/performance/ae1qew88",
   "tg_tokens": null,
   "tps": 40.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-ae1qew88-ctx32768",
   "model": "Qwen3.8-9B-Distill-Uncensored-Heretic",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 9.2 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 201.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-9B-Distill-uncensored-heretic-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 32k 201.3 36.3 9.2 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ae1qew88",
   "source_url": "https://omlx.ai/benchmarks/performance/ae1qew88",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-a38akdz4-ctx1024",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 13.4 GB at this context; duplicate traces of the same session: dru1hfaq, laoxzlms, ojvmgfix; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 610.9,
   "provenance": "community",
   "quant": "MXFP8",
   "quote": "gemma-4-12B-it-MXFP8 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 unknown; 1k 610.9 19.6 13.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a38akdz4",
   "source_url": "https://omlx.ai/benchmarks/performance/a38akdz4",
   "tg_tokens": null,
   "tps": 19.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-a38akdz4-ctx4096",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 13.8 GB at this context; duplicate traces of the same session: dru1hfaq, laoxzlms, ojvmgfix; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 833.3,
   "provenance": "community",
   "quant": "MXFP8",
   "quote": "gemma-4-12B-it-MXFP8 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 unknown; 4k 833.3 19.8 13.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a38akdz4",
   "source_url": "https://omlx.ai/benchmarks/performance/a38akdz4",
   "tg_tokens": null,
   "tps": 19.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-a38akdz4-ctx8192",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 14.0 GB at this context; duplicate traces of the same session: dru1hfaq, laoxzlms, ojvmgfix; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 940.7,
   "provenance": "community",
   "quant": "MXFP8",
   "quote": "gemma-4-12B-it-MXFP8 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 unknown; 8k 940.7 19.8 14.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a38akdz4",
   "source_url": "https://omlx.ai/benchmarks/performance/a38akdz4",
   "tg_tokens": null,
   "tps": 19.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-a38akdz4-ctx16384",
   "model": "gemma-4-12b-it",
   "notes": "oMLX community benchmark trace; peak memory 14.4 GB at this context; duplicate traces of the same session: dru1hfaq, laoxzlms, ojvmgfix; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 827.4,
   "provenance": "community",
   "quant": "MXFP8",
   "quote": "gemma-4-12B-it-MXFP8 on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 unknown; 16k 827.4 19.6 14.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/a38akdz4",
   "source_url": "https://omlx.ai/benchmarks/performance/a38akdz4",
   "tg_tokens": null,
   "tps": 19.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-iot3ujj3-ctx1024",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: 8t506gw6, st03o4dt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3800.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 3,800 91.6 25.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iot3ujj3",
   "source_url": "https://omlx.ai/benchmarks/performance/iot3ujj3",
   "tg_tokens": null,
   "tps": 91.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-iot3ujj3-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: 8t506gw6, st03o4dt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4050.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 4,050 88.8 26.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iot3ujj3",
   "source_url": "https://omlx.ai/benchmarks/performance/iot3ujj3",
   "tg_tokens": null,
   "tps": 88.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-iot3ujj3-ctx8192",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: 8t506gw6, st03o4dt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3940.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 3,940 86.5 26.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iot3ujj3",
   "source_url": "https://omlx.ai/benchmarks/performance/iot3ujj3",
   "tg_tokens": null,
   "tps": 86.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-iot3ujj3-ctx16384",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: 8t506gw6, st03o4dt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3397.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 3,397 80.1 26.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iot3ujj3",
   "source_url": "https://omlx.ai/benchmarks/performance/iot3ujj3",
   "tg_tokens": null,
   "tps": 80.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-50jun0g4-ctx1024",
   "model": "gemma-4-26B-A4B-it-qat-JANG_4M",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: ewrfhb8v; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 799.5,
   "provenance": "community",
   "quant": "unknown",
   "quote": "gemma-4-26B-A4B-it-qat-JANG_4M on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 unknown; 1k 799.5 54.4 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/50jun0g4",
   "source_url": "https://omlx.ai/benchmarks/performance/50jun0g4",
   "tg_tokens": null,
   "tps": 54.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-50jun0g4-ctx4096",
   "model": "gemma-4-26B-A4B-it-qat-JANG_4M",
   "notes": "oMLX community benchmark trace; peak memory 17.3 GB at this context; duplicate traces of the same session: ewrfhb8v; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2175.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "gemma-4-26B-A4B-it-qat-JANG_4M on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 unknown; 4k 2,175 53.3 17.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/50jun0g4",
   "source_url": "https://omlx.ai/benchmarks/performance/50jun0g4",
   "tg_tokens": null,
   "tps": 53.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-50jun0g4-ctx8192",
   "model": "gemma-4-26B-A4B-it-qat-JANG_4M",
   "notes": "oMLX community benchmark trace; peak memory 17.5 GB at this context; duplicate traces of the same session: ewrfhb8v; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2250.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "gemma-4-26B-A4B-it-qat-JANG_4M on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 unknown; 8k 2,250 52.1 17.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/50jun0g4",
   "source_url": "https://omlx.ai/benchmarks/performance/50jun0g4",
   "tg_tokens": null,
   "tps": 52.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-50jun0g4-ctx16384",
   "model": "gemma-4-26B-A4B-it-qat-JANG_4M",
   "notes": "oMLX community benchmark trace; peak memory 17.9 GB at this context; duplicate traces of the same session: ewrfhb8v; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2006.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "gemma-4-26B-A4B-it-qat-JANG_4M on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 unknown; 16k 2,006 49.4 17.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/50jun0g4",
   "source_url": "https://omlx.ai/benchmarks/performance/50jun0g4",
   "tg_tokens": null,
   "tps": 49.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-iu74o1as-ctx1024",
   "model": "gemma-4-31B-it",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 32.0 GB at this context; duplicate traces of the same session: 853sxcbi, dnf7nejz, n7agmmhe; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 637.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-mlx-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 637.4 14.4 32.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iu74o1as",
   "source_url": "https://omlx.ai/benchmarks/performance/iu74o1as",
   "tg_tokens": null,
   "tps": 14.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-iu74o1as-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 33.1 GB at this context; duplicate traces of the same session: 853sxcbi, dnf7nejz, n7agmmhe; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 531.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-mlx-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 531.2 14.1 33.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iu74o1as",
   "source_url": "https://omlx.ai/benchmarks/performance/iu74o1as",
   "tg_tokens": null,
   "tps": 14.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-iu74o1as-ctx8192",
   "model": "gemma-4-31B-it",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 33.6 GB at this context; duplicate traces of the same session: 853sxcbi, dnf7nejz, n7agmmhe; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 471.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-mlx-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 471.0 13.6 33.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iu74o1as",
   "source_url": "https://omlx.ai/benchmarks/performance/iu74o1as",
   "tg_tokens": null,
   "tps": 13.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-iu74o1as-ctx16384",
   "model": "gemma-4-31B-it",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: 853sxcbi, dnf7nejz, n7agmmhe; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 435.8,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-31B-it-mlx-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 435.8 13.1 34.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/iu74o1as",
   "source_url": "https://omlx.ai/benchmarks/performance/iu74o1as",
   "tg_tokens": null,
   "tps": 13.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0lu4vp3o-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: 1ktk67vk, 25vj2q9k, 34rd3mg2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4006.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 4,006 104.3 35.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0lu4vp3o",
   "source_url": "https://omlx.ai/benchmarks/performance/0lu4vp3o",
   "tg_tokens": null,
   "tps": 104.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0lu4vp3o-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 36.1 GB at this context; duplicate traces of the same session: 1ktk67vk, 25vj2q9k, 34rd3mg2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3470.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 3,470 102.4 36.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0lu4vp3o",
   "source_url": "https://omlx.ai/benchmarks/performance/0lu4vp3o",
   "tg_tokens": null,
   "tps": 102.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0lu4vp3o-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 36.2 GB at this context; duplicate traces of the same session: 1ktk67vk, 25vj2q9k, 34rd3mg2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3821.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 3,821 102.4 36.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0lu4vp3o",
   "source_url": "https://omlx.ai/benchmarks/performance/0lu4vp3o",
   "tg_tokens": null,
   "tps": 102.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-0lu4vp3o-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 36.4 GB at this context; duplicate traces of the same session: 1ktk67vk, 25vj2q9k, 34rd3mg2; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3799.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 3,799 98.0 36.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0lu4vp3o",
   "source_url": "https://omlx.ai/benchmarks/performance/0lu4vp3o",
   "tg_tokens": null,
   "tps": 98.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-zlv9a9hq-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2901.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Ornith-1.5-9B-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 2,901 30.5 17.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zlv9a9hq",
   "source_url": "https://omlx.ai/benchmarks/performance/zlv9a9hq",
   "tg_tokens": null,
   "tps": 30.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-zlv9a9hq-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1966.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Ornith-1.5-9B-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 4k 1,966 30.3 18.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zlv9a9hq",
   "source_url": "https://omlx.ai/benchmarks/performance/zlv9a9hq",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-zlv9a9hq-ctx8192",
   "model": "Ornith-1.5-9B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2428.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Ornith-1.5-9B-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 2,428 30.3 18.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zlv9a9hq",
   "source_url": "https://omlx.ai/benchmarks/performance/zlv9a9hq",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-zlv9a9hq-ctx16384",
   "model": "Ornith-1.5-9B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2460.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Ornith-1.5-9B-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 2,460 29.7 18.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zlv9a9hq",
   "source_url": "https://omlx.ai/benchmarks/performance/zlv9a9hq",
   "tg_tokens": null,
   "tps": 29.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-aavw7a8z-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: ets7zei2, w3iatsyu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4249.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 4,249 103.0 35.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aavw7a8z",
   "source_url": "https://omlx.ai/benchmarks/performance/aavw7a8z",
   "tg_tokens": null,
   "tps": 103.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-aavw7a8z-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 36.2 GB at this context; duplicate traces of the same session: ets7zei2, w3iatsyu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 5834.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 5,834 102.8 36.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aavw7a8z",
   "source_url": "https://omlx.ai/benchmarks/performance/aavw7a8z",
   "tg_tokens": null,
   "tps": 102.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-aavw7a8z-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 36.4 GB at this context; duplicate traces of the same session: ets7zei2, w3iatsyu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4757.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-35B-A3B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 4,757 88.6 36.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aavw7a8z",
   "source_url": "https://omlx.ai/benchmarks/performance/aavw7a8z",
   "tg_tokens": null,
   "tps": 88.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-drujiofq-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: 1iee3k1d; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2286.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 2,286 134.6 37.4 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/drujiofq",
   "source_url": "https://omlx.ai/benchmarks/performance/drujiofq",
   "tg_tokens": null,
   "tps": 134.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-drujiofq-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.1 GB at this context; duplicate traces of the same session: 1iee3k1d; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2975.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 2,975 147.5 39.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/drujiofq",
   "source_url": "https://omlx.ai/benchmarks/performance/drujiofq",
   "tg_tokens": null,
   "tps": 147.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gbmuy045-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 4sm7j3hl, d8rgoa4u, odapxe4l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 504.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 504.9 17.8 28.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gbmuy045",
   "source_url": "https://omlx.ai/benchmarks/performance/gbmuy045",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gbmuy045-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: 4sm7j3hl, d8rgoa4u, odapxe4l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 586.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 586.6 17.7 29.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gbmuy045",
   "source_url": "https://omlx.ai/benchmarks/performance/gbmuy045",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gbmuy045-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.8 GB at this context; duplicate traces of the same session: 4sm7j3hl, d8rgoa4u, odapxe4l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 600.9,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 600.9 16.9 29.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gbmuy045",
   "source_url": "https://omlx.ai/benchmarks/performance/gbmuy045",
   "tg_tokens": null,
   "tps": 16.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-gbmuy045-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 30.3 GB at this context; duplicate traces of the same session: 4sm7j3hl, d8rgoa4u, odapxe4l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 563.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 563.1 16.1 30.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gbmuy045",
   "source_url": "https://omlx.ai/benchmarks/performance/gbmuy045",
   "tg_tokens": null,
   "tps": 16.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-xa7i7oia-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: xmfigsg2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 413.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 413.1 25.6 16.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xa7i7oia",
   "source_url": "https://omlx.ai/benchmarks/performance/xa7i7oia",
   "tg_tokens": null,
   "tps": 25.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-xa7i7oia-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: xmfigsg2; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 394.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 394.7 26.9 18.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xa7i7oia",
   "source_url": "https://omlx.ai/benchmarks/performance/xa7i7oia",
   "tg_tokens": null,
   "tps": 26.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qs6h8c7u-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 385.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 385.4 17.9 16.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qs6h8c7u",
   "source_url": "https://omlx.ai/benchmarks/performance/qs6h8c7u",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-qs6h8c7u-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 418.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 418.3 17.8 17.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qs6h8c7u",
   "source_url": "https://omlx.ai/benchmarks/performance/qs6h8c7u",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-rgx6cm12-ctx65536",
   "model": "Qwen3.8-27B-Uncensored-HauhauCS-Aggressive",
   "notes": "MTPLX build; oMLX community benchmark trace, single run at 64k context; peak memory 28.5 GB; TTFT 183653 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 356.8,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTPLX-6bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 6bit; 64k tokens 356.8 PP tok/s 20.1 TG tok/s 183653 TTFT (ms) 28.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rgx6cm12",
   "source_url": "https://omlx.ai/benchmarks/performance/rgx6cm12",
   "tg_tokens": null,
   "tps": 20.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-jvm6p6qg-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: h3coewhi, q88oh1yl; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 64.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k 64.1 12.9 18.4 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jvm6p6qg",
   "source_url": "https://omlx.ai/benchmarks/performance/jvm6p6qg",
   "tg_tokens": null,
   "tps": 12.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-jvm6p6qg-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: h3coewhi, q88oh1yl; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 64.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 8k 64.1 16.0 20.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jvm6p6qg",
   "source_url": "https://omlx.ai/benchmarks/performance/jvm6p6qg",
   "tg_tokens": null,
   "tps": 16.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-jvm6p6qg-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.8 GB at this context; duplicate traces of the same session: h3coewhi, q88oh1yl; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 58.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 32k 58.1 17.8 22.8 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jvm6p6qg",
   "source_url": "https://omlx.ai/benchmarks/performance/jvm6p6qg",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-xm43n10j-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: 9lm4t1yv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 127.4 23.5 20.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xm43n10j",
   "source_url": "https://omlx.ai/benchmarks/performance/xm43n10j",
   "tg_tokens": null,
   "tps": 23.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-xm43n10j-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.4 GB at this context; duplicate traces of the same session: 9lm4t1yv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 189.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 189.0 30.2 24.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xm43n10j",
   "source_url": "https://omlx.ai/benchmarks/performance/xm43n10j",
   "tg_tokens": null,
   "tps": 30.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-ryb13t05-ctx1024",
   "model": "Qwen3.8-9B-Distill-Uncensored-Heretic",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.0 GB at this context; duplicate traces of the same session: d9uzlhuo; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-9B-Distill-uncensored-heretic-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k 224.0 32.6 7.0 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ryb13t05",
   "source_url": "https://omlx.ai/benchmarks/performance/ryb13t05",
   "tg_tokens": null,
   "tps": 32.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-ryb13t05-ctx8192",
   "model": "Qwen3.8-9B-Distill-Uncensored-Heretic",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 7.7 GB at this context; duplicate traces of the same session: d9uzlhuo; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 229.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-9B-Distill-uncensored-heretic-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 8k 229.3 40.8 7.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ryb13t05",
   "source_url": "https://omlx.ai/benchmarks/performance/ryb13t05",
   "tg_tokens": null,
   "tps": 40.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-ryb13t05-ctx32768",
   "model": "Qwen3.8-9B-Distill-Uncensored-Heretic",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 9.2 GB at this context; duplicate traces of the same session: d9uzlhuo; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 201.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-9B-Distill-uncensored-heretic-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 32k 201.3 36.3 9.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ryb13t05",
   "source_url": "https://omlx.ai/benchmarks/performance/ryb13t05",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-x63n3e9q-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "fp16 MTP build; oMLX community benchmark trace, single run at 64k context; peak memory 53.7 GB; TTFT 250665 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 261.4,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "Qwen3.8-Flash-Next-oQ2.5e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 2bit; 64k tokens 261.4 PP tok/s 29.1 TG tok/s 250665 TTFT (ms) 53.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x63n3e9q",
   "source_url": "https://omlx.ai/benchmarks/performance/x63n3e9q",
   "tg_tokens": null,
   "tps": 29.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-ucx78nv4-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 70.9 GB at this context; duplicate traces of the same session: cqrmxf2b, ud1u9tqj; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 94.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 1k 94.5 24.7 70.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ucx78nv4",
   "source_url": "https://omlx.ai/benchmarks/performance/ucx78nv4",
   "tg_tokens": null,
   "tps": 24.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-ucx78nv4-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 73.6 GB at this context; duplicate traces of the same session: cqrmxf2b, ud1u9tqj; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 624.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 8k 624.1 25.0 73.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ucx78nv4",
   "source_url": "https://omlx.ai/benchmarks/performance/ucx78nv4",
   "tg_tokens": null,
   "tps": 25.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-ucx78nv4-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 73.8 GB at this context; duplicate traces of the same session: cqrmxf2b, ud1u9tqj; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 624.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 16k 624.3 26.4 73.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ucx78nv4",
   "source_url": "https://omlx.ai/benchmarks/performance/ucx78nv4",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-x16xbtzs-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.6 GB at this context; duplicate traces of the same session: nhyzwdzy; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1192.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,192 101.5 100.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x16xbtzs",
   "source_url": "https://omlx.ai/benchmarks/performance/x16xbtzs",
   "tg_tokens": null,
   "tps": 101.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-x16xbtzs-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: nhyzwdzy; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1436.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,436 94.8 101.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x16xbtzs",
   "source_url": "https://omlx.ai/benchmarks/performance/x16xbtzs",
   "tg_tokens": null,
   "tps": 94.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-8hq2rwuv-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 100.5 GB at this context; duplicate traces of the same session: skhymg1h; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1182.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,182 94.8 100.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8hq2rwuv",
   "source_url": "https://omlx.ai/benchmarks/performance/8hq2rwuv",
   "tg_tokens": null,
   "tps": 94.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-8hq2rwuv-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 101.1 GB at this context; duplicate traces of the same session: skhymg1h; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1437.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,437 94.8 101.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8hq2rwuv",
   "source_url": "https://omlx.ai/benchmarks/performance/8hq2rwuv",
   "tg_tokens": null,
   "tps": 94.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-ik1b66fs-ctx4096",
   "model": "Qwen3.6-35B-A3B-MTPLX-Optimized-Balance",
   "notes": "Youssofal build; oMLX community benchmark trace; peak memory 29.2 GB at this context; duplicate traces of the same session: lyqo04z6, p3fh3scw; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1668.0,
   "provenance": "community",
   "quant": "FP16",
   "quote": "Youssofal--Qwen3.6-35B-A3B-MTPLX-Optimized-Balance-FP16 on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 6bit; 4k 1,668 58.9 29.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ik1b66fs",
   "source_url": "https://omlx.ai/benchmarks/performance/ik1b66fs",
   "tg_tokens": null,
   "tps": 58.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-ik1b66fs-ctx8192",
   "model": "Qwen3.6-35B-A3B-MTPLX-Optimized-Balance",
   "notes": "Youssofal build; oMLX community benchmark trace; peak memory 29.6 GB at this context; duplicate traces of the same session: lyqo04z6, p3fh3scw; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1577.0,
   "provenance": "community",
   "quant": "FP16",
   "quote": "Youssofal--Qwen3.6-35B-A3B-MTPLX-Optimized-Balance-FP16 on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 6bit; 8k 1,577 57.7 29.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ik1b66fs",
   "source_url": "https://omlx.ai/benchmarks/performance/ik1b66fs",
   "tg_tokens": null,
   "tps": 57.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-ik1b66fs-ctx16384",
   "model": "Qwen3.6-35B-A3B-MTPLX-Optimized-Balance",
   "notes": "Youssofal build; oMLX community benchmark trace; peak memory 30.1 GB at this context; duplicate traces of the same session: lyqo04z6, p3fh3scw; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1401.0,
   "provenance": "community",
   "quant": "FP16",
   "quote": "Youssofal--Qwen3.6-35B-A3B-MTPLX-Optimized-Balance-FP16 on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 6bit; 16k 1,401 56.4 30.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ik1b66fs",
   "source_url": "https://omlx.ai/benchmarks/performance/ik1b66fs",
   "tg_tokens": null,
   "tps": 56.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8dwi93jf-ctx1024",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 25.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3800.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 3,800 91.6 25.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8dwi93jf",
   "source_url": "https://omlx.ai/benchmarks/performance/8dwi93jf",
   "tg_tokens": null,
   "tps": 91.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8dwi93jf-ctx4096",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4050.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 4,050 88.8 26.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8dwi93jf",
   "source_url": "https://omlx.ai/benchmarks/performance/8dwi93jf",
   "tg_tokens": null,
   "tps": 88.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8dwi93jf-ctx8192",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.3 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3940.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 3,940 86.5 26.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8dwi93jf",
   "source_url": "https://omlx.ai/benchmarks/performance/8dwi93jf",
   "tg_tokens": null,
   "tps": 86.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-8dwi93jf-ctx16384",
   "model": "Gemma 4 26B-A4B",
   "notes": "oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3397.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "gemma-4-26B-A4B-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 3,397 80.1 26.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8dwi93jf",
   "source_url": "https://omlx.ai/benchmarks/performance/8dwi93jf",
   "tg_tokens": null,
   "tps": 80.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-b1s4rw16-ctx1024",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 5.7 GB at this context; duplicate traces of the same session: r0kzb7kj, ri6tgi0z; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 649.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 1k 649.4 60.8 5.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b1s4rw16",
   "source_url": "https://omlx.ai/benchmarks/performance/b1s4rw16",
   "tg_tokens": null,
   "tps": 60.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-b1s4rw16-ctx8192",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 5.8 GB at this context; duplicate traces of the same session: r0kzb7kj, ri6tgi0z; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 649.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 8k 649.3 46.0 5.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b1s4rw16",
   "source_url": "https://omlx.ai/benchmarks/performance/b1s4rw16",
   "tg_tokens": null,
   "tps": 46.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-b1s4rw16-ctx32768",
   "model": "Gemma 4 E4B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: r0kzb7kj, ri6tgi0z; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 576.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-E4B-it-oQ4e-mtp on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 32k 576.9 24.1 6.4 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/b1s4rw16",
   "source_url": "https://omlx.ai/benchmarks/performance/b1s4rw16",
   "tg_tokens": null,
   "tps": 24.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-vcae31xm-ctx16384",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "mlx-community build; oMLX community benchmark trace, single run at 16k context; peak memory 20.1 GB; TTFT 63908 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 256.4,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "mlx-community--MiniCPM-V-4.7-35B-A3B-mxfp4 on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k tokens 256.4 PP tok/s 17.6 TG tok/s 63908 TTFT (ms) 20.1 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vcae31xm",
   "source_url": "https://omlx.ai/benchmarks/performance/vcae31xm",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-g3zr28fm-ctx16384",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "mlx-community build; oMLX community benchmark trace, single run at 16k context; peak memory 18.4 GB; TTFT 66040 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 248.1,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "mlx-community--MiniCPM-V-4.7-35B-A3B-mxfp4 on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k tokens 248.1 PP tok/s 17.0 TG tok/s 66040 TTFT (ms) 18.4 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g3zr28fm",
   "source_url": "https://omlx.ai/benchmarks/performance/g3zr28fm",
   "tg_tokens": null,
   "tps": 17.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-za3mf795-ctx1024",
   "model": "Qwen2.5-Coder-32B-Instruct",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 17.9 GB at this context; duplicate traces of the same session: 938kqflm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 772.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen2.5-Coder-32B-Instruct-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 772.3 27.8 17.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/za3mf795",
   "source_url": "https://omlx.ai/benchmarks/performance/za3mf795",
   "tg_tokens": null,
   "tps": 27.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-za3mf795-ctx4096",
   "model": "Qwen2.5-Coder-32B-Instruct",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 19.0 GB at this context; duplicate traces of the same session: 938kqflm; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 798.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "mlx-community--Qwen2.5-Coder-32B-Instruct-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 798.3 26.3 19.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/za3mf795",
   "source_url": "https://omlx.ai/benchmarks/performance/za3mf795",
   "tg_tokens": null,
   "tps": 26.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-wb2q00m2-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "1017 build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.8 GB at this context; duplicate traces of the same session: gej4egqb, uzllsiyd, z3vyoyzh, zomrlzrt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 776.9,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 unknown; 4k 776.9 41.4 178.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wb2q00m2",
   "source_url": "https://omlx.ai/benchmarks/performance/wb2q00m2",
   "tg_tokens": null,
   "tps": 41.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-wb2q00m2-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "1017 build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.9 GB at this context; duplicate traces of the same session: gej4egqb, uzllsiyd, z3vyoyzh, zomrlzrt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 895.3,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 unknown; 8k 895.3 38.2 178.9 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wb2q00m2",
   "source_url": "https://omlx.ai/benchmarks/performance/wb2q00m2",
   "tg_tokens": null,
   "tps": 38.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-wb2q00m2-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "1017 build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.6 GB at this context; duplicate traces of the same session: gej4egqb, uzllsiyd, z3vyoyzh, zomrlzrt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 878.1,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 unknown; 16k 878.1 53.3 179.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wb2q00m2",
   "source_url": "https://omlx.ai/benchmarks/performance/wb2q00m2",
   "tg_tokens": null,
   "tps": 53.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-wb2q00m2-ctx32768",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "1017 build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.7 GB at this context; duplicate traces of the same session: gej4egqb, uzllsiyd, z3vyoyzh, zomrlzrt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 880.1,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 unknown; 32k 880.1 53.1 179.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wb2q00m2",
   "source_url": "https://omlx.ai/benchmarks/performance/wb2q00m2",
   "tg_tokens": null,
   "tps": 53.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-wb2q00m2-ctx65536",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "1017 build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 179.7 GB at this context; duplicate traces of the same session: gej4egqb, uzllsiyd, z3vyoyzh, zomrlzrt; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 857.6,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 unknown; 64k 857.6 48.6 179.7 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wb2q00m2",
   "source_url": "https://omlx.ai/benchmarks/performance/wb2q00m2",
   "tg_tokens": null,
   "tps": 48.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-573eoqvm-ctx1024",
   "model": "NVIDIA-Nemotron-3.5-Super-VL-120B-A12B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 69.8 GB at this context; duplicate traces of the same session: 9q4gbjp1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 400.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Duckets--NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 400.7 50.8 69.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/573eoqvm",
   "source_url": "https://omlx.ai/benchmarks/performance/573eoqvm",
   "tg_tokens": null,
   "tps": 50.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-573eoqvm-ctx4096",
   "model": "NVIDIA-Nemotron-3.5-Super-VL-120B-A12B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 72.9 GB at this context; duplicate traces of the same session: 9q4gbjp1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 426.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Duckets--NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-MLX-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 426.8 50.2 72.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/573eoqvm",
   "source_url": "https://omlx.ai/benchmarks/performance/573eoqvm",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-e0irpnm2-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: 1u3vqpcc; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1668.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-24B-A2B-oQ8e-fp16 on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 8bit; 1k 1,668 79.2 24.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e0irpnm2",
   "source_url": "https://omlx.ai/benchmarks/performance/e0irpnm2",
   "tg_tokens": null,
   "tps": 79.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-e0irpnm2-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "fp16 build; oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: 1u3vqpcc; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2069.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "LFM2.5-24B-A2B-oQ8e-fp16 on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 8bit; 4k 2,069 78.1 24.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e0irpnm2",
   "source_url": "https://omlx.ai/benchmarks/performance/e0irpnm2",
   "tg_tokens": null,
   "tps": 78.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mgb4ivsf-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: b9ep2cly, ct25znn8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2901.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Ornith-1.5-9B-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 2,901 30.5 17.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mgb4ivsf",
   "source_url": "https://omlx.ai/benchmarks/performance/mgb4ivsf",
   "tg_tokens": null,
   "tps": 30.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mgb4ivsf-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 18.4 GB at this context; duplicate traces of the same session: b9ep2cly, ct25znn8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1966.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Ornith-1.5-9B-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 4k 1,966 30.3 18.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mgb4ivsf",
   "source_url": "https://omlx.ai/benchmarks/performance/mgb4ivsf",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mgb4ivsf-ctx8192",
   "model": "Ornith-1.5-9B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: b9ep2cly, ct25znn8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2428.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Ornith-1.5-9B-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 2,428 30.3 18.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mgb4ivsf",
   "source_url": "https://omlx.ai/benchmarks/performance/mgb4ivsf",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-mgb4ivsf-ctx16384",
   "model": "Ornith-1.5-9B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: b9ep2cly, ct25znn8; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2460.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Ornith-1.5-9B-MLX on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 2,460 29.7 18.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mgb4ivsf",
   "source_url": "https://omlx.ai/benchmarks/performance/mgb4ivsf",
   "tg_tokens": null,
   "tps": 29.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-d5e94mjd-ctx1024",
   "model": "Qwen3-Coder-Next",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 80.2 GB at this context; duplicate traces of the same session: f0kbck0f, irzixos4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 959.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next-MLX-8bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 8bit; 1k 959.6 58.6 80.2 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d5e94mjd",
   "source_url": "https://omlx.ai/benchmarks/performance/d5e94mjd",
   "tg_tokens": null,
   "tps": 58.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-d5e94mjd-ctx8192",
   "model": "Qwen3-Coder-Next",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 81.1 GB at this context; duplicate traces of the same session: f0kbck0f, irzixos4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1049.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next-MLX-8bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 8bit; 8k 1,049 55.9 81.1 GB view",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d5e94mjd",
   "source_url": "https://omlx.ai/benchmarks/performance/d5e94mjd",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Ultra",
   "id": "omlx-d5e94mjd-ctx16384",
   "model": "Qwen3-Coder-Next",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 81.3 GB at this context; duplicate traces of the same session: f0kbck0f, irzixos4; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 920.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3-Coder-Next-MLX-8bit on M2 Ultra (60c); M2 Ultra (60c) \u00b7 192 GB \u00b7 8bit; 16k 920.1 53.5 81.3 GB current",
   "ram_gb": 192,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d5e94mjd",
   "source_url": "https://omlx.ai/benchmarks/performance/d5e94mjd",
   "tg_tokens": null,
   "tps": 53.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-jkd7nrk0-ctx65536",
   "model": "Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive",
   "notes": "OptiQ MLX build; oMLX community benchmark trace, single run at 64k context; peak memory 21.8 GB; TTFT 257029 ms; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 255.0,
   "provenance": "community",
   "quant": "6bpw",
   "quote": "Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-OptiQ-6bpw-MLX on M4 (10c); M4 (10c) \u00b7 32 GB \u00b7 4bit; 64k tokens 255.0 PP tok/s 12.0 TG tok/s 257029 TTFT (ms) 21.8 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jkd7nrk0",
   "source_url": "https://omlx.ai/benchmarks/performance/jkd7nrk0",
   "tg_tokens": null,
   "tps": 12.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-vmqt1kgs-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1306.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k 1,306 65.1 21.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vmqt1kgs",
   "source_url": "https://omlx.ai/benchmarks/performance/vmqt1kgs",
   "tg_tokens": null,
   "tps": 65.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-vmqt1kgs-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1514.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 1,514 55.7 21.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vmqt1kgs",
   "source_url": "https://omlx.ai/benchmarks/performance/vmqt1kgs",
   "tg_tokens": null,
   "tps": 55.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-vmqt1kgs-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1514.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 8k 1,514 71.7 21.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vmqt1kgs",
   "source_url": "https://omlx.ai/benchmarks/performance/vmqt1kgs",
   "tg_tokens": null,
   "tps": 71.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-vmqt1kgs-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1389.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 16k 1,389 70.8 22.1 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vmqt1kgs",
   "source_url": "https://omlx.ai/benchmarks/performance/vmqt1kgs",
   "tg_tokens": null,
   "tps": 70.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-vmqt1kgs-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: none; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1165.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 32k 1,165 75.4 22.5 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vmqt1kgs",
   "source_url": "https://omlx.ai/benchmarks/performance/vmqt1kgs",
   "tg_tokens": null,
   "tps": 75.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-nsag5bej-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 29.9 GB; TTFT 1139 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 898.9,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.6-35B-A3B-oQ6e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 6bit; 1k tokens 898.9 PP tok/s 95.4 TG tok/s 1139 TTFT (ms) 29.9 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nsag5bej",
   "source_url": "https://omlx.ai/benchmarks/performance/nsag5bej",
   "tg_tokens": null,
   "tps": 95.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-z4mlwouh-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 23.6 GB at this context; duplicate traces of the same session: 6f4yrxol; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 98.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 98.5 18.1 23.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z4mlwouh",
   "source_url": "https://omlx.ai/benchmarks/performance/z4mlwouh",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-z4mlwouh-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: 6f4yrxol; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 87.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 16k 87.8 22.6 25.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z4mlwouh",
   "source_url": "https://omlx.ai/benchmarks/performance/z4mlwouh",
   "tg_tokens": null,
   "tps": 22.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ztooa92l-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: qz5pm1oy; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 549.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 549.9 26.7 17.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ztooa92l",
   "source_url": "https://omlx.ai/benchmarks/performance/ztooa92l",
   "tg_tokens": null,
   "tps": 26.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ztooa92l-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 18.1 GB at this context; duplicate traces of the same session: qz5pm1oy; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 607.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 607.2 26.3 18.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ztooa92l",
   "source_url": "https://omlx.ai/benchmarks/performance/ztooa92l",
   "tg_tokens": null,
   "tps": 26.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-rkkb22v3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: 7kzas6x4; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 224.4,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 1k 224.4 31.2 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rkkb22v3",
   "source_url": "https://omlx.ai/benchmarks/performance/rkkb22v3",
   "tg_tokens": null,
   "tps": 31.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-rkkb22v3-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.9 GB at this context; duplicate traces of the same session: 7kzas6x4; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 235.2,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 4bit; 8k 235.2 43.2 23.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rkkb22v3",
   "source_url": "https://omlx.ai/benchmarks/performance/rkkb22v3",
   "tg_tokens": null,
   "tps": 43.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-3sbwvxf3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 34.7 GB at this context; duplicate traces of the same session: 0usg16pc; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 219.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 8bit; 1k 219.0 18.5 34.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3sbwvxf3",
   "source_url": "https://omlx.ai/benchmarks/performance/3sbwvxf3",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-3sbwvxf3-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: 0usg16pc; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 226.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 128 GB \u00b7 8bit; 8k 226.0 21.8 39.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3sbwvxf3",
   "source_url": "https://omlx.ai/benchmarks/performance/3sbwvxf3",
   "tg_tokens": null,
   "tps": 21.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-dng1gole-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 105.2 GB at this context; duplicate traces of the same session: blux1oxj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1200.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,200 69.6 105.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dng1gole",
   "source_url": "https://omlx.ai/benchmarks/performance/dng1gole",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-dng1gole-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 105.7 GB at this context; duplicate traces of the same session: blux1oxj; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1466.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-4bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,466 69.2 105.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dng1gole",
   "source_url": "https://omlx.ai/benchmarks/performance/dng1gole",
   "tg_tokens": null,
   "tps": 69.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cwzmd9vm-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 83.0 GB at this context; duplicate traces of the same session: 2auinnfx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 361.2,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 4k 361.2 50.2 83.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cwzmd9vm",
   "source_url": "https://omlx.ai/benchmarks/performance/cwzmd9vm",
   "tg_tokens": null,
   "tps": 50.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-cwzmd9vm-ctx8192",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 83.7 GB at this context; duplicate traces of the same session: 2auinnfx; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 586.7,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 5bit; 8k 586.7 30.0 83.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cwzmd9vm",
   "source_url": "https://omlx.ai/benchmarks/performance/cwzmd9vm",
   "tg_tokens": null,
   "tps": 30.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M2 Max",
   "id": "omlx-7658w239-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 38.8 GB; TTFT 16035 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 63.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k tokens 63.9 PP tok/s 9.6 TG tok/s 16035 TTFT (ms) 38.8 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7658w239",
   "source_url": "https://omlx.ai/benchmarks/performance/7658w239",
   "tg_tokens": null,
   "tps": 9.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M2 Max",
   "id": "omlx-tc9p5orf-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 40.2 GB; TTFT 124960 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 131.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 16k tokens 131.1 PP tok/s 7.5 TG tok/s 124960 TTFT (ms) 40.2 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tc9p5orf",
   "source_url": "https://omlx.ai/benchmarks/performance/tc9p5orf",
   "tg_tokens": null,
   "tps": 7.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-r30a8ki7-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.2 GB at this context; duplicate traces of the same session: kciie4ga; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 512.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 512.0 50.9 69.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r30a8ki7",
   "source_url": "https://omlx.ai/benchmarks/performance/r30a8ki7",
   "tg_tokens": null,
   "tps": 50.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-r30a8ki7-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.0 GB at this context; duplicate traces of the same session: kciie4ga; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 561.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 561.8 50.0 71.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r30a8ki7",
   "source_url": "https://omlx.ai/benchmarks/performance/r30a8ki7",
   "tg_tokens": null,
   "tps": 50.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-5bwhjs3x-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "1017 build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.5 GB at this context; duplicate traces of the same session: noa78157, qufeid9h; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1102.0,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 unknown; 4k 1,102 43.9 178.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5bwhjs3x",
   "source_url": "https://omlx.ai/benchmarks/performance/5bwhjs3x",
   "tg_tokens": null,
   "tps": 43.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-5bwhjs3x-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "1017 build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.5 GB at this context; duplicate traces of the same session: noa78157, qufeid9h; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1379.0,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 unknown; 8k 1,379 39.1 178.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5bwhjs3x",
   "source_url": "https://omlx.ai/benchmarks/performance/5bwhjs3x",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-5bwhjs3x-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "1017 build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.5 GB at this context; duplicate traces of the same session: noa78157, qufeid9h; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1384.0,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 unknown; 16k 1,384 51.7 178.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5bwhjs3x",
   "source_url": "https://omlx.ai/benchmarks/performance/5bwhjs3x",
   "tg_tokens": null,
   "tps": 51.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-5bwhjs3x-ctx32768",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "1017 build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.5 GB at this context; duplicate traces of the same session: noa78157, qufeid9h; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1429.0,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 unknown; 32k 1,429 49.8 178.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5bwhjs3x",
   "source_url": "https://omlx.ai/benchmarks/performance/5bwhjs3x",
   "tg_tokens": null,
   "tps": 49.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-5bwhjs3x-ctx65536",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "1017 build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.6 GB at this context; duplicate traces of the same session: noa78157, qufeid9h; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1407.0,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 unknown; 64k 1,407 48.5 178.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5bwhjs3x",
   "source_url": "https://omlx.ai/benchmarks/performance/5bwhjs3x",
   "tg_tokens": null,
   "tps": 48.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2hd6kpbn-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 249.0 GB at this context; duplicate traces of the same session: 91m1pa0r; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 487.4,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 1k 487.4 42.5 249.0 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2hd6kpbn",
   "source_url": "https://omlx.ai/benchmarks/performance/2hd6kpbn",
   "tg_tokens": null,
   "tps": 42.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-2hd6kpbn-ctx32768",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 252.5 GB at this context; duplicate traces of the same session: 91m1pa0r; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 507.2,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 32k 507.2 50.4 252.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2hd6kpbn",
   "source_url": "https://omlx.ai/benchmarks/performance/2hd6kpbn",
   "tg_tokens": null,
   "tps": 50.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-i51pfrxe-ctx1024",
   "model": "GLM-5.3-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 249.0 GB; TTFT 2122 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 482.6,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "GLM-5.3-Flash-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 1k tokens 482.6 PP tok/s 44.4 TG tok/s 2122 TTFT (ms) 249.0 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i51pfrxe",
   "source_url": "https://omlx.ai/benchmarks/performance/i51pfrxe",
   "tg_tokens": null,
   "tps": 44.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-nrod6dji-ctx1024",
   "model": "MiniCPM-V-4.7-1B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 3.5 GB; TTFT 351 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2914.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM-V-4.7-1B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 2,914 PP tok/s 344.3 TG tok/s 351 TTFT (ms) 3.5 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nrod6dji",
   "source_url": "https://omlx.ai/benchmarks/performance/nrod6dji",
   "tg_tokens": null,
   "tps": 344.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xbf5f9ri-ctx1024",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 35.6 GB at this context; duplicate traces of the same session: 98aohztd, a62w3j1q; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2739.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 2,739 86.8 35.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xbf5f9ri",
   "source_url": "https://omlx.ai/benchmarks/performance/xbf5f9ri",
   "tg_tokens": null,
   "tps": 86.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xbf5f9ri-ctx4096",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: 98aohztd, a62w3j1q; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3400.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 3,400 86.9 37.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xbf5f9ri",
   "source_url": "https://omlx.ai/benchmarks/performance/xbf5f9ri",
   "tg_tokens": null,
   "tps": 86.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-xbf5f9ri-ctx8192",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: 98aohztd, a62w3j1q; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3258.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 8k 3,258 84.2 37.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xbf5f9ri",
   "source_url": "https://omlx.ai/benchmarks/performance/xbf5f9ri",
   "tg_tokens": null,
   "tps": 84.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-n0pg8ztf-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: ehh3ssyp, gle0jpvs, lw2k46tz, ov714hbk; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1334.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 1k 1,334 56.4 22.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n0pg8ztf",
   "source_url": "https://omlx.ai/benchmarks/performance/n0pg8ztf",
   "tg_tokens": null,
   "tps": 56.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-n0pg8ztf-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: ehh3ssyp, gle0jpvs, lw2k46tz, ov714hbk; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1679.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 4k 1,679 84.1 24.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n0pg8ztf",
   "source_url": "https://omlx.ai/benchmarks/performance/n0pg8ztf",
   "tg_tokens": null,
   "tps": 84.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-n0pg8ztf-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: ehh3ssyp, gle0jpvs, lw2k46tz, ov714hbk; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1534.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 8k 1,534 93.3 24.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n0pg8ztf",
   "source_url": "https://omlx.ai/benchmarks/performance/n0pg8ztf",
   "tg_tokens": null,
   "tps": 93.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-n0pg8ztf-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.2 GB at this context; duplicate traces of the same session: ehh3ssyp, gle0jpvs, lw2k46tz, ov714hbk; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1376.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 16k 1,376 92.6 25.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n0pg8ztf",
   "source_url": "https://omlx.ai/benchmarks/performance/n0pg8ztf",
   "tg_tokens": null,
   "tps": 92.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-n0pg8ztf-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: ehh3ssyp, gle0jpvs, lw2k46tz, ov714hbk; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1100.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 32k 1,100 121.9 26.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n0pg8ztf",
   "source_url": "https://omlx.ai/benchmarks/performance/n0pg8ztf",
   "tg_tokens": null,
   "tps": 121.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-lv3as6mo-ctx131072",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 10.1 GB; TTFT 203645 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 643.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 128k tokens 643.6 PP tok/s 14.8 TG tok/s 203645 TTFT (ms) 10.1 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lv3as6mo",
   "source_url": "https://omlx.ai/benchmarks/performance/lv3as6mo",
   "tg_tokens": null,
   "tps": 14.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-dex0cd9f-ctx65536",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 8.1 GB; TTFT 77166 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 849.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 64k tokens 849.3 PP tok/s 20.8 TG tok/s 77166 TTFT (ms) 8.1 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dex0cd9f",
   "source_url": "https://omlx.ai/benchmarks/performance/dex0cd9f",
   "tg_tokens": null,
   "tps": 20.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M6",
   "id": "omlx-p75h4ok5-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: rbp4g1jk; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1129.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 4k 1,129 31.7 6.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p75h4ok5",
   "source_url": "https://omlx.ai/benchmarks/performance/p75h4ok5",
   "tg_tokens": null,
   "tps": 31.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-p75h4ok5-ctx16384",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.6 GB at this context; duplicate traces of the same session: rbp4g1jk; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1095.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 16k 1,095 29.2 6.6 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p75h4ok5",
   "source_url": "https://omlx.ai/benchmarks/performance/p75h4ok5",
   "tg_tokens": null,
   "tps": 29.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-on51aoqg-ctx65536",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: vua50etu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 200.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-27B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 64k 200.2 13.9 37.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/on51aoqg",
   "source_url": "https://omlx.ai/benchmarks/performance/on51aoqg",
   "tg_tokens": null,
   "tps": 13.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-on51aoqg-ctx131072",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace; peak memory 41.2 GB at this context; duplicate traces of the same session: vua50etu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 142.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-27B-MLX-8bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 8bit; 128k 142.0 11.5 41.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/on51aoqg",
   "source_url": "https://omlx.ai/benchmarks/performance/on51aoqg",
   "tg_tokens": null,
   "tps": 11.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-7j3j6tls-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: xqst3mmp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2944.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 2,944 100.6 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7j3j6tls",
   "source_url": "https://omlx.ai/benchmarks/performance/7j3j6tls",
   "tg_tokens": null,
   "tps": 100.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-7j3j6tls-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: xqst3mmp; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2949.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 16k 2,949 90.4 20.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7j3j6tls",
   "source_url": "https://omlx.ai/benchmarks/performance/7j3j6tls",
   "tg_tokens": null,
   "tps": 90.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-8cdpm7jg-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 6mxvfsz4, 8qtbxtaa, ydqiuu61; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1306.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 1k 1,306 65.1 21.3 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8cdpm7jg",
   "source_url": "https://omlx.ai/benchmarks/performance/8cdpm7jg",
   "tg_tokens": null,
   "tps": 65.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-8cdpm7jg-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 6mxvfsz4, 8qtbxtaa, ydqiuu61; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1514.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 4k 1,514 55.7 21.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8cdpm7jg",
   "source_url": "https://omlx.ai/benchmarks/performance/8cdpm7jg",
   "tg_tokens": null,
   "tps": 55.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-8cdpm7jg-ctx8192",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: 6mxvfsz4, 8qtbxtaa, ydqiuu61; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1514.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 8k 1,514 71.7 21.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8cdpm7jg",
   "source_url": "https://omlx.ai/benchmarks/performance/8cdpm7jg",
   "tg_tokens": null,
   "tps": 71.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-8cdpm7jg-ctx16384",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 6mxvfsz4, 8qtbxtaa, ydqiuu61; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1389.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 16k 1,389 70.8 22.1 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8cdpm7jg",
   "source_url": "https://omlx.ai/benchmarks/performance/8cdpm7jg",
   "tg_tokens": null,
   "tps": 70.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-8cdpm7jg-ctx32768",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: 6mxvfsz4, 8qtbxtaa, ydqiuu61; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1165.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 4bit; 32k 1,165 75.4 22.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8cdpm7jg",
   "source_url": "https://omlx.ai/benchmarks/performance/8cdpm7jg",
   "tg_tokens": null,
   "tps": 75.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-4v1919mk-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 19.4 GB; TTFT 4045 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 253.1,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 253.1 PP tok/s 62.8 TG tok/s 4045 TTFT (ms) 19.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4v1919mk",
   "source_url": "https://omlx.ai/benchmarks/performance/4v1919mk",
   "tg_tokens": null,
   "tps": 62.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-99306jcx-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 27.6 GB; TTFT 4150 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 246.7,
   "provenance": "community",
   "quant": "oQ6e-fp16",
   "quote": "Qwen3.8-27B-oQ6e-fp16-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k tokens 246.7 PP tok/s 49.4 TG tok/s 4150 TTFT (ms) 27.6 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/99306jcx",
   "source_url": "https://omlx.ai/benchmarks/performance/99306jcx",
   "tg_tokens": null,
   "tps": 49.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-wptwmdvv-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 122.2 GB at this context; duplicate traces of the same session: fmtryrih, jsj3aw2h, lexcci0g, xryoy63i, zjgeb8ec; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 670.1,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 5bit; 1k 670.1 35.3 122.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wptwmdvv",
   "source_url": "https://omlx.ai/benchmarks/performance/wptwmdvv",
   "tg_tokens": null,
   "tps": 35.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-wptwmdvv-ctx4096",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 122.5 GB at this context; duplicate traces of the same session: fmtryrih, jsj3aw2h, lexcci0g, xryoy63i, zjgeb8ec; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1769.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 5bit; 4k 1,769 83.6 122.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wptwmdvv",
   "source_url": "https://omlx.ai/benchmarks/performance/wptwmdvv",
   "tg_tokens": null,
   "tps": 83.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-wptwmdvv-ctx8192",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 122.5 GB at this context; duplicate traces of the same session: fmtryrih, jsj3aw2h, lexcci0g, xryoy63i, zjgeb8ec; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2987.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 5bit; 8k 2,987 136.5 122.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wptwmdvv",
   "source_url": "https://omlx.ai/benchmarks/performance/wptwmdvv",
   "tg_tokens": null,
   "tps": 136.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-wptwmdvv-ctx65536",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 125.3 GB at this context; duplicate traces of the same session: fmtryrih, jsj3aw2h, lexcci0g, xryoy63i, zjgeb8ec; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2859.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 5bit; 64k 2,859 67.4 125.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wptwmdvv",
   "source_url": "https://omlx.ai/benchmarks/performance/wptwmdvv",
   "tg_tokens": null,
   "tps": 67.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-wptwmdvv-ctx131072",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 128.0 GB at this context; duplicate traces of the same session: fmtryrih, jsj3aw2h, lexcci0g, xryoy63i, zjgeb8ec; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2727.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 5bit; 128k 2,727 84.1 128.0 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wptwmdvv",
   "source_url": "https://omlx.ai/benchmarks/performance/wptwmdvv",
   "tg_tokens": null,
   "tps": 84.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-wptwmdvv-ctx199680",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 131.3 GB at this context; duplicate traces of the same session: fmtryrih, jsj3aw2h, lexcci0g, xryoy63i, zjgeb8ec; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2557.0,
   "provenance": "community",
   "quant": "oQ5e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 5bit; 195k 2,557 83.0 131.3 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wptwmdvv",
   "source_url": "https://omlx.ai/benchmarks/performance/wptwmdvv",
   "tg_tokens": null,
   "tps": 83.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ynxugfog-ctx1024",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: g1pvvkbu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 242.7,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 2bit; 1k 242.7 17.4 11.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ynxugfog",
   "source_url": "https://omlx.ai/benchmarks/performance/ynxugfog",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ynxugfog-ctx4096",
   "model": "Ternary-Bonsai-2-27B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 14.2 GB at this context; duplicate traces of the same session: g1pvvkbu; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 250.1,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-27B-mlx-2bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 2bit; 4k 250.1 16.5 14.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ynxugfog",
   "source_url": "https://omlx.ai/benchmarks/performance/ynxugfog",
   "tg_tokens": null,
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1dszdj6y-ctx1024",
   "model": "Ternary-Bonsai-2-9B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 5.7 GB at this context; duplicate traces of the same session: ycd19c02; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 871.2,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-9B-mlx-2bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 2bit; 1k 871.2 55.0 5.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1dszdj6y",
   "source_url": "https://omlx.ai/benchmarks/performance/1dszdj6y",
   "tg_tokens": null,
   "tps": 55.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1dszdj6y-ctx4096",
   "model": "Ternary-Bonsai-2-9B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: ycd19c02; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 853.9,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Ternary-Bonsai-2-9B-mlx-2bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 2bit; 4k 853.9 50.5 6.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1dszdj6y",
   "source_url": "https://omlx.ai/benchmarks/performance/1dszdj6y",
   "tg_tokens": null,
   "tps": 50.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ogz0fgeb-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 93.0 GB at this context; duplicate traces of the same session: a1z71r28; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 108.7,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 108.7 11.2 93.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ogz0fgeb",
   "source_url": "https://omlx.ai/benchmarks/performance/ogz0fgeb",
   "tg_tokens": null,
   "tps": 11.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ogz0fgeb-ctx131072",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 97.7 GB at this context; duplicate traces of the same session: a1z71r28; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 520.5,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 128k 520.5 16.2 97.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ogz0fgeb",
   "source_url": "https://omlx.ai/benchmarks/performance/ogz0fgeb",
   "tg_tokens": null,
   "tps": 16.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-hioq2v8n-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 92.8 GB at this context; duplicate traces of the same session: 6q2o6px6, mk73bmlf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 110.6,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 110.6 10.8 92.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hioq2v8n",
   "source_url": "https://omlx.ai/benchmarks/performance/hioq2v8n",
   "tg_tokens": null,
   "tps": 10.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hioq2v8n-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.2 GB at this context; duplicate traces of the same session: 6q2o6px6, mk73bmlf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 514.1,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 8k 514.1 13.4 96.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hioq2v8n",
   "source_url": "https://omlx.ai/benchmarks/performance/hioq2v8n",
   "tg_tokens": null,
   "tps": 13.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-hioq2v8n-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.3 GB at this context; duplicate traces of the same session: 6q2o6px6, mk73bmlf; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 579.1,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 16k 579.1 16.4 96.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hioq2v8n",
   "source_url": "https://omlx.ai/benchmarks/performance/hioq2v8n",
   "tg_tokens": null,
   "tps": 16.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-3wrd5ydp-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "1017 build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.5 GB at this context; duplicate traces of the same session: 0gea47g4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1102.0,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 unknown; 4k 1,102 43.9 178.5 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3wrd5ydp",
   "source_url": "https://omlx.ai/benchmarks/performance/3wrd5ydp",
   "tg_tokens": null,
   "tps": 43.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-3wrd5ydp-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "1017 build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.5 GB at this context; duplicate traces of the same session: 0gea47g4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1379.0,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 unknown; 8k 1,379 39.1 178.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3wrd5ydp",
   "source_url": "https://omlx.ai/benchmarks/performance/3wrd5ydp",
   "tg_tokens": null,
   "tps": 39.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-3wrd5ydp-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "1017 build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.5 GB at this context; duplicate traces of the same session: 0gea47g4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1384.0,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 unknown; 16k 1,384 51.7 178.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3wrd5ydp",
   "source_url": "https://omlx.ai/benchmarks/performance/3wrd5ydp",
   "tg_tokens": null,
   "tps": 51.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-3wrd5ydp-ctx32768",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "1017 build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.5 GB at this context; duplicate traces of the same session: 0gea47g4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1429.0,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 unknown; 32k 1,429 49.8 178.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3wrd5ydp",
   "source_url": "https://omlx.ai/benchmarks/performance/3wrd5ydp",
   "tg_tokens": null,
   "tps": 49.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-3wrd5ydp-ctx65536",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "1017 build; MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 178.6 GB at this context; duplicate traces of the same session: 0gea47g4; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1407.0,
   "provenance": "community",
   "quant": "oQ2e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2e-mtp on M5 Ultra (80c); M5 Ultra (80c) \u00b7 256 GB \u00b7 unknown; 64k 1,407 48.5 178.6 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3wrd5ydp",
   "source_url": "https://omlx.ai/benchmarks/performance/3wrd5ydp",
   "tg_tokens": null,
   "tps": 48.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5ivir1h7-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 403.3 GB at this context; duplicate traces of the same session: 3uk8r8cb, 4jytcjzf, hgjifuhc, iobr8rqs, rjfg54fn, sr5goclk, wal5851w; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 529.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 529.8 26.8 403.3 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ivir1h7",
   "source_url": "https://omlx.ai/benchmarks/performance/5ivir1h7",
   "tg_tokens": null,
   "tps": 26.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5ivir1h7-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 403.7 GB at this context; duplicate traces of the same session: 3uk8r8cb, 4jytcjzf, hgjifuhc, iobr8rqs, rjfg54fn, sr5goclk, wal5851w; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 680.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 680.1 34.9 403.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ivir1h7",
   "source_url": "https://omlx.ai/benchmarks/performance/5ivir1h7",
   "tg_tokens": null,
   "tps": 34.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5ivir1h7-ctx8192",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 403.7 GB at this context; duplicate traces of the same session: 3uk8r8cb, 4jytcjzf, hgjifuhc, iobr8rqs, rjfg54fn, sr5goclk, wal5851w; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 719.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 8k 719.1 25.9 403.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ivir1h7",
   "source_url": "https://omlx.ai/benchmarks/performance/5ivir1h7",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5ivir1h7-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 403.7 GB at this context; duplicate traces of the same session: 3uk8r8cb, 4jytcjzf, hgjifuhc, iobr8rqs, rjfg54fn, sr5goclk, wal5851w; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 715.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 16k 715.0 33.1 403.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ivir1h7",
   "source_url": "https://omlx.ai/benchmarks/performance/5ivir1h7",
   "tg_tokens": null,
   "tps": 33.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5ivir1h7-ctx32768",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 403.8 GB at this context; duplicate traces of the same session: 3uk8r8cb, 4jytcjzf, hgjifuhc, iobr8rqs, rjfg54fn, sr5goclk, wal5851w; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 714.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 32k 714.0 38.7 403.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ivir1h7",
   "source_url": "https://omlx.ai/benchmarks/performance/5ivir1h7",
   "tg_tokens": null,
   "tps": 38.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-5ivir1h7-ctx65536",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 403.8 GB at this context; duplicate traces of the same session: 3uk8r8cb, 4jytcjzf, hgjifuhc, iobr8rqs, rjfg54fn, sr5goclk, wal5851w; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 709.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 64k 709.9 35.7 403.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ivir1h7",
   "source_url": "https://omlx.ai/benchmarks/performance/5ivir1h7",
   "tg_tokens": null,
   "tps": 35.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5ivir1h7-ctx131072",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 403.9 GB at this context; duplicate traces of the same session: 3uk8r8cb, 4jytcjzf, hgjifuhc, iobr8rqs, rjfg54fn, sr5goclk, wal5851w; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 690.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 128k 690.2 36.9 403.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ivir1h7",
   "source_url": "https://omlx.ai/benchmarks/performance/5ivir1h7",
   "tg_tokens": null,
   "tps": 36.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-5ivir1h7-ctx199680",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 403.9 GB at this context; duplicate traces of the same session: 3uk8r8cb, 4jytcjzf, hgjifuhc, iobr8rqs, rjfg54fn, sr5goclk, wal5851w; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 670.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 195k 670.6 37.6 403.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ivir1h7",
   "source_url": "https://omlx.ai/benchmarks/performance/5ivir1h7",
   "tg_tokens": null,
   "tps": 37.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-t5k0c4dn-ctx1024",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 23.0 GB; TTFT 789 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1298.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "MiniCPM-V-4.7-35B-A3B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k tokens 1,298 PP tok/s 100.5 TG tok/s 789 TTFT (ms) 23.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/t5k0c4dn",
   "source_url": "https://omlx.ai/benchmarks/performance/t5k0c4dn",
   "tg_tokens": null,
   "tps": 100.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-d1o1prb9-ctx1024",
   "model": "Ornith-1.0-35B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: c3izk29l; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 508.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.0-35B-oQ4e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 508.4 67.8 20.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d1o1prb9",
   "source_url": "https://omlx.ai/benchmarks/performance/d1o1prb9",
   "tg_tokens": null,
   "tps": 67.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-d1o1prb9-ctx4096",
   "model": "Ornith-1.0-35B",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: c3izk29l; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1456.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.0-35B-oQ4e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 1,456 60.8 20.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d1o1prb9",
   "source_url": "https://omlx.ai/benchmarks/performance/d1o1prb9",
   "tg_tokens": null,
   "tps": 60.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-0l38ezh3-ctx4096",
   "model": "Qwen3.6-35B-A3B-TurboQuant",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3398.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-TurboQuant-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,398 103.2 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0l38ezh3",
   "source_url": "https://omlx.ai/benchmarks/performance/0l38ezh3",
   "tg_tokens": null,
   "tps": 103.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-0l38ezh3-ctx16384",
   "model": "Qwen3.6-35B-A3B-TurboQuant",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3062.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-TurboQuant-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 3,062 93.4 20.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0l38ezh3",
   "source_url": "https://omlx.ai/benchmarks/performance/0l38ezh3",
   "tg_tokens": null,
   "tps": 93.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-e66n3icy-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: o5c440mz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 784.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 784.2 48.5 37.4 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e66n3icy",
   "source_url": "https://omlx.ai/benchmarks/performance/e66n3icy",
   "tg_tokens": null,
   "tps": 48.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-e66n3icy-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 38.0 GB at this context; duplicate traces of the same session: o5c440mz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1688.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 1,688 57.6 38.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/e66n3icy",
   "source_url": "https://omlx.ai/benchmarks/performance/e66n3icy",
   "tg_tokens": null,
   "tps": 57.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-rcz0mw7c-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 64k context; peak memory 20.8 GB; TTFT 273835 ms; duplicate traces of the same session: none; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 239.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M6 (12c); M6 (12c) \u00b7 32 GB \u00b7 4bit; 64k tokens 239.3 PP tok/s 7.0 TG tok/s 273835 TTFT (ms) 20.8 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rcz0mw7c",
   "source_url": "https://omlx.ai/benchmarks/performance/rcz0mw7c",
   "tg_tokens": null,
   "tps": 7.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-exqtheg0-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: 1j7tt17c; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 324.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 324.3 7.4 28.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/exqtheg0",
   "source_url": "https://omlx.ai/benchmarks/performance/exqtheg0",
   "tg_tokens": null,
   "tps": 7.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-exqtheg0-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: 1j7tt17c; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 299.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 299.0 7.5 29.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/exqtheg0",
   "source_url": "https://omlx.ai/benchmarks/performance/exqtheg0",
   "tg_tokens": null,
   "tps": 7.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-vesqsidy-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 390.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 390.7 18.2 16.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vesqsidy",
   "source_url": "https://omlx.ai/benchmarks/performance/vesqsidy",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-vesqsidy-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 415.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 415.2 17.7 17.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vesqsidy",
   "source_url": "https://omlx.ai/benchmarks/performance/vesqsidy",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7y4xs0nx-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.7 GB at this context; duplicate traces of the same session: ovperb6l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 892.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 892.6 69.1 17.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7y4xs0nx",
   "source_url": "https://omlx.ai/benchmarks/performance/7y4xs0nx",
   "tg_tokens": null,
   "tps": 69.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-7y4xs0nx-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: ovperb6l; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 859.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 859.7 76.9 19.6 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7y4xs0nx",
   "source_url": "https://omlx.ai/benchmarks/performance/7y4xs0nx",
   "tg_tokens": null,
   "tps": 76.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-etk4chqp-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: fqhmbvhv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 460.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 460.9 32.7 17.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/etk4chqp",
   "source_url": "https://omlx.ai/benchmarks/performance/etk4chqp",
   "tg_tokens": null,
   "tps": 32.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-etk4chqp-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: fqhmbvhv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 477.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 477.9 35.8 18.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/etk4chqp",
   "source_url": "https://omlx.ai/benchmarks/performance/etk4chqp",
   "tg_tokens": null,
   "tps": 35.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-zyqrsokf-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 38.9 GB at this context; duplicate traces of the same session: 9zyunn1n; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 152.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 4k 152.1 18.4 38.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zyqrsokf",
   "source_url": "https://omlx.ai/benchmarks/performance/zyqrsokf",
   "tg_tokens": null,
   "tps": 18.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-zyqrsokf-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP build, fp16 attention; oMLX community benchmark trace; peak memory 40.7 GB at this context; duplicate traces of the same session: 9zyunn1n; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 147.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-oQ8e-fp16-mtp on M2 Max (30c); M2 Max (30c) \u00b7 64 GB \u00b7 8bit; 16k 147.4 17.4 40.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zyqrsokf",
   "source_url": "https://omlx.ai/benchmarks/performance/zyqrsokf",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-stb27sr9-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.8 GB at this context; duplicate traces of the same session: 08zuk5fp; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2062.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 1k 2,062 120.5 142.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/stb27sr9",
   "source_url": "https://omlx.ai/benchmarks/performance/stb27sr9",
   "tg_tokens": null,
   "tps": 120.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-stb27sr9-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: 08zuk5fp; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3572.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 4k 3,572 122.8 143.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/stb27sr9",
   "source_url": "https://omlx.ai/benchmarks/performance/stb27sr9",
   "tg_tokens": null,
   "tps": 122.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-w5g5okk2-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.6 GB at this context; duplicate traces of the same session: jd5k4czq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 794.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 794.1 45.7 27.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w5g5okk2",
   "source_url": "https://omlx.ai/benchmarks/performance/w5g5okk2",
   "tg_tokens": null,
   "tps": 45.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-w5g5okk2-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: jd5k4czq; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1467.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 1,467 34.2 28.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w5g5okk2",
   "source_url": "https://omlx.ai/benchmarks/performance/w5g5okk2",
   "tg_tokens": null,
   "tps": 34.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-3ps06845-ctx4096",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 4k context; peak memory 21.3 GB; TTFT 45889 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 89.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 4k tokens 89.3 PP tok/s 18.5 TG tok/s 45889 TTFT (ms) 21.3 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ps06845",
   "source_url": "https://omlx.ai/benchmarks/performance/3ps06845",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-sdsiidt8-ctx1024",
   "model": "gemma-4-E4B-it",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 6.6 GB at this context; duplicate traces of the same session: wkq5pj1a; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2272.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "gemma-4-E4B-it-MLX-5bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 5bit; 1k 2,272 40.7 6.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sdsiidt8",
   "source_url": "https://omlx.ai/benchmarks/performance/sdsiidt8",
   "tg_tokens": null,
   "tps": 40.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-sdsiidt8-ctx4096",
   "model": "gemma-4-E4B-it",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: wkq5pj1a; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2922.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "gemma-4-E4B-it-MLX-5bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 5bit; 4k 2,922 39.6 6.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sdsiidt8",
   "source_url": "https://omlx.ai/benchmarks/performance/sdsiidt8",
   "tg_tokens": null,
   "tps": 39.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-i3ej0qo2-ctx4096",
   "model": "gemma-4-E4B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 6.1 GB at this context; duplicate traces of the same session: z0vb8w8h; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2275.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E4B-it-qat-4bit on M6 (12c); M6 (12c) \u00b7 16 GB \u00b7 4bit; 4k 2,275 28.2 6.1 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i3ej0qo2",
   "source_url": "https://omlx.ai/benchmarks/performance/i3ej0qo2",
   "tg_tokens": null,
   "tps": 28.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M6",
   "id": "omlx-i3ej0qo2-ctx16384",
   "model": "gemma-4-E4B-it-qat",
   "notes": "oMLX community benchmark trace; peak memory 6.4 GB at this context; duplicate traces of the same session: z0vb8w8h; config=12c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2352.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-E4B-it-qat-4bit on M6 (12c); M6 (12c) \u00b7 16 GB \u00b7 4bit; 16k 2,352 26.2 6.4 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i3ej0qo2",
   "source_url": "https://omlx.ai/benchmarks/performance/i3ej0qo2",
   "tg_tokens": null,
   "tps": 26.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tmjxhlxd-ctx1024",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 89.0 GB at this context; duplicate traces of the same session: 6rxzipbt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 381.7,
   "provenance": "community",
   "quant": "2.4bit mixed",
   "quote": "DeepSeek-V4-Flash-0731-2.4bit-mixed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 1k 381.7 17.0 89.0 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tmjxhlxd",
   "source_url": "https://omlx.ai/benchmarks/performance/tmjxhlxd",
   "tg_tokens": null,
   "tps": 17.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-tmjxhlxd-ctx4096",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 89.8 GB at this context; duplicate traces of the same session: 6rxzipbt; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 417.4,
   "provenance": "community",
   "quant": "2.4bit mixed",
   "quote": "DeepSeek-V4-Flash-0731-2.4bit-mixed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 4k 417.4 25.1 89.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tmjxhlxd",
   "source_url": "https://omlx.ai/benchmarks/performance/tmjxhlxd",
   "tg_tokens": null,
   "tps": 25.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-et0mt19b-ctx32768",
   "model": "Huihui-MiniCPM-V-4.6-abliterated",
   "notes": "MLX build; oMLX community benchmark trace, single run at 32k context; peak memory 4.0 GB; TTFT 25709 ms; duplicate traces of the same session: none; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1275.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Huihui-MiniCPM-V-4.6-abliterated-mlx-bf16 on M4 (8c); M4 (8c) \u00b7 16 GB \u00b7 bf16; 32k tokens 1,275 PP tok/s 37.4 TG tok/s 25709 TTFT (ms) 4.0 Peak mem",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/et0mt19b",
   "source_url": "https://omlx.ai/benchmarks/performance/et0mt19b",
   "tg_tokens": null,
   "tps": 37.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-gkhbns97-ctx4096",
   "model": "Huihui-MiniCPM-V-4.6-abliterated",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 3.6 GB at this context; duplicate traces of the same session: yx2qiask; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2033.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Huihui-MiniCPM-V-4.6-abliterated-mlx-bf16 on M4 (8c); M4 (8c) \u00b7 16 GB \u00b7 bf16; 4k 2,033 56.3 3.6 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gkhbns97",
   "source_url": "https://omlx.ai/benchmarks/performance/gkhbns97",
   "tg_tokens": null,
   "tps": 56.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-gkhbns97-ctx16384",
   "model": "Huihui-MiniCPM-V-4.6-abliterated",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 3.8 GB at this context; duplicate traces of the same session: yx2qiask; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1631.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Huihui-MiniCPM-V-4.6-abliterated-mlx-bf16 on M4 (8c); M4 (8c) \u00b7 16 GB \u00b7 bf16; 16k 1,631 46.0 3.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/gkhbns97",
   "source_url": "https://omlx.ai/benchmarks/performance/gkhbns97",
   "tg_tokens": null,
   "tps": 46.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-g46ccqw2-ctx4096",
   "model": "Huihui-MiniCPM-V-4.6-abliterated",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 3.5 GB at this context; duplicate traces of the same session: ldn155mi; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2036.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Huihui-MiniCPM-V-4.6-abliterated-mlx-bf16 on M4 (8c); M4 (8c) \u00b7 16 GB \u00b7 bf16; 4k 2,036 55.9 3.5 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g46ccqw2",
   "source_url": "https://omlx.ai/benchmarks/performance/g46ccqw2",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-g46ccqw2-ctx16384",
   "model": "Huihui-MiniCPM-V-4.6-abliterated",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 3.8 GB at this context; duplicate traces of the same session: ldn155mi; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1627.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Huihui-MiniCPM-V-4.6-abliterated-mlx-bf16 on M4 (8c); M4 (8c) \u00b7 16 GB \u00b7 bf16; 16k 1,627 46.0 3.8 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g46ccqw2",
   "source_url": "https://omlx.ai/benchmarks/performance/g46ccqw2",
   "tg_tokens": null,
   "tps": 46.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-5fvpkowt-ctx4096",
   "model": "Huihui-MiniCPM-V-4.6-abliterated",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 3.5 GB at this context; duplicate traces of the same session: yfoo85e3; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2033.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Huihui-MiniCPM-V-4.6-abliterated-mlx-bf16 on M4 (8c); M4 (8c) \u00b7 16 GB \u00b7 bf16; 4k 2,033 56.1 3.5 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5fvpkowt",
   "source_url": "https://omlx.ai/benchmarks/performance/5fvpkowt",
   "tg_tokens": null,
   "tps": 56.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4",
   "id": "omlx-5fvpkowt-ctx16384",
   "model": "Huihui-MiniCPM-V-4.6-abliterated",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 3.8 GB at this context; duplicate traces of the same session: yfoo85e3; config=8c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1612.0,
   "provenance": "community",
   "quant": "bf16",
   "quote": "Huihui-MiniCPM-V-4.6-abliterated-mlx-bf16 on M4 (8c); M4 (8c) \u00b7 16 GB \u00b7 bf16; 16k 1,612 45.4 3.8 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5fvpkowt",
   "source_url": "https://omlx.ai/benchmarks/performance/5fvpkowt",
   "tg_tokens": null,
   "tps": 45.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-2ob3dqei-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: nzezo57y, s4a52rgs, v2fqzgye; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1653.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 4k 1,653 76.7 24.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ob3dqei",
   "source_url": "https://omlx.ai/benchmarks/performance/2ob3dqei",
   "tg_tokens": null,
   "tps": 76.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-2ob3dqei-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: nzezo57y, s4a52rgs, v2fqzgye; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1577.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 8k 1,577 82.1 24.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ob3dqei",
   "source_url": "https://omlx.ai/benchmarks/performance/2ob3dqei",
   "tg_tokens": null,
   "tps": 82.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-2ob3dqei-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.2 GB at this context; duplicate traces of the same session: nzezo57y, s4a52rgs, v2fqzgye; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1382.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 16k 1,382 96.4 25.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ob3dqei",
   "source_url": "https://omlx.ai/benchmarks/performance/2ob3dqei",
   "tg_tokens": null,
   "tps": 96.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-2ob3dqei-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: nzezo57y, s4a52rgs, v2fqzgye; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1102.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 32k 1,102 121.7 26.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ob3dqei",
   "source_url": "https://omlx.ai/benchmarks/performance/2ob3dqei",
   "tg_tokens": null,
   "tps": 121.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-2ob3dqei-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.3 GB at this context; duplicate traces of the same session: nzezo57y, s4a52rgs, v2fqzgye; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 786.1,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 64k 786.1 65.2 27.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ob3dqei",
   "source_url": "https://omlx.ai/benchmarks/performance/2ob3dqei",
   "tg_tokens": null,
   "tps": 65.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-2ob3dqei-ctx131072",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.8 GB at this context; duplicate traces of the same session: nzezo57y, s4a52rgs, v2fqzgye; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 503.9,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 128k 503.9 42.7 28.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2ob3dqei",
   "source_url": "https://omlx.ai/benchmarks/performance/2ob3dqei",
   "tg_tokens": null,
   "tps": 42.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-mpn0y99i-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 345k62tj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 901.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 901.9 111.4 21.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mpn0y99i",
   "source_url": "https://omlx.ai/benchmarks/performance/mpn0y99i",
   "tg_tokens": null,
   "tps": 111.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-mpn0y99i-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: 345k62tj; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 952.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 952.5 90.7 22.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mpn0y99i",
   "source_url": "https://omlx.ai/benchmarks/performance/mpn0y99i",
   "tg_tokens": null,
   "tps": 90.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-y4po72ml-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: do43hcm6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 906.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 906.8 95.7 21.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y4po72ml",
   "source_url": "https://omlx.ai/benchmarks/performance/y4po72ml",
   "tg_tokens": null,
   "tps": 95.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-y4po72ml-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: do43hcm6; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 951.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 951.7 91.7 22.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/y4po72ml",
   "source_url": "https://omlx.ai/benchmarks/performance/y4po72ml",
   "tg_tokens": null,
   "tps": 91.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-1al2gh0w-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: none; config=19c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 249.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M2 Pro (19c); M2 Pro (19c) \u00b7 32 GB \u00b7 4bit; 1k 249.3 35.4 5.9 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1al2gh0w",
   "source_url": "https://omlx.ai/benchmarks/performance/1al2gh0w",
   "tg_tokens": null,
   "tps": 35.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-1al2gh0w-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: none; config=19c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 253.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M2 Pro (19c); M2 Pro (19c) \u00b7 32 GB \u00b7 4bit; 4k 253.1 34.1 6.5 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1al2gh0w",
   "source_url": "https://omlx.ai/benchmarks/performance/1al2gh0w",
   "tg_tokens": null,
   "tps": 34.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-z1llxo80-ctx4096",
   "model": "Qwen3.6-35B-A3B-TurboQuant",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3398.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-TurboQuant-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,398 103.2 20.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z1llxo80",
   "source_url": "https://omlx.ai/benchmarks/performance/z1llxo80",
   "tg_tokens": null,
   "tps": 103.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-z1llxo80-ctx16384",
   "model": "Qwen3.6-35B-A3B-TurboQuant",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 20.3 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3062.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-TurboQuant-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 16k 3,062 93.4 20.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/z1llxo80",
   "source_url": "https://omlx.ai/benchmarks/performance/z1llxo80",
   "tg_tokens": null,
   "tps": 93.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-awj9j2j7-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: mah5dooj; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 756.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 1k 756.3 26.3 16.0 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/awj9j2j7",
   "source_url": "https://omlx.ai/benchmarks/performance/awj9j2j7",
   "tg_tokens": null,
   "tps": 26.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-awj9j2j7-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: mah5dooj; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 830.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (32c); M5 Max (32c) \u00b7 36 GB \u00b7 4bit; 4k 830.0 25.9 17.0 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/awj9j2j7",
   "source_url": "https://omlx.ai/benchmarks/performance/awj9j2j7",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-w8o97baw-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 390.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 1k 390.7 18.2 16.0 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w8o97baw",
   "source_url": "https://omlx.ai/benchmarks/performance/w8o97baw",
   "tg_tokens": null,
   "tps": 18.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-w8o97baw-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 415.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Pro (16c); M5 Pro (16c) \u00b7 48 GB \u00b7 4bit; 4k 415.2 17.7 17.0 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w8o97baw",
   "source_url": "https://omlx.ai/benchmarks/performance/w8o97baw",
   "tg_tokens": null,
   "tps": 17.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-nbskrqxb-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 0495t6we, dbstkv9j, lkuo3f8t, rww4nv2w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 254.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 254.0 37.1 21.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nbskrqxb",
   "source_url": "https://omlx.ai/benchmarks/performance/nbskrqxb",
   "tg_tokens": null,
   "tps": 37.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-nbskrqxb-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.0 GB at this context; duplicate traces of the same session: 0495t6we, dbstkv9j, lkuo3f8t, rww4nv2w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 340.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 340.2 36.3 25.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nbskrqxb",
   "source_url": "https://omlx.ai/benchmarks/performance/nbskrqxb",
   "tg_tokens": null,
   "tps": 36.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-nbskrqxb-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.5 GB at this context; duplicate traces of the same session: 0495t6we, dbstkv9j, lkuo3f8t, rww4nv2w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 340.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 340.3 37.2 25.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nbskrqxb",
   "source_url": "https://omlx.ai/benchmarks/performance/nbskrqxb",
   "tg_tokens": null,
   "tps": 37.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-nbskrqxb-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: 0495t6we, dbstkv9j, lkuo3f8t, rww4nv2w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 324.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 324.2 37.8 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nbskrqxb",
   "source_url": "https://omlx.ai/benchmarks/performance/nbskrqxb",
   "tg_tokens": null,
   "tps": 37.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-nbskrqxb-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.3 GB at this context; duplicate traces of the same session: 0495t6we, dbstkv9j, lkuo3f8t, rww4nv2w; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 294.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 294.2 41.2 28.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nbskrqxb",
   "source_url": "https://omlx.ai/benchmarks/performance/nbskrqxb",
   "tg_tokens": null,
   "tps": 41.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-nelcgcx8-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: 4ihnlq6h, r42xbhr9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 254.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 254.1 60.6 21.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nelcgcx8",
   "source_url": "https://omlx.ai/benchmarks/performance/nelcgcx8",
   "tg_tokens": null,
   "tps": 60.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-nelcgcx8-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: 4ihnlq6h, r42xbhr9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 338.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 338.5 78.5 25.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nelcgcx8",
   "source_url": "https://omlx.ai/benchmarks/performance/nelcgcx8",
   "tg_tokens": null,
   "tps": 78.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-nelcgcx8-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.6 GB at this context; duplicate traces of the same session: 4ihnlq6h, r42xbhr9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 340.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 340.6 41.2 25.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nelcgcx8",
   "source_url": "https://omlx.ai/benchmarks/performance/nelcgcx8",
   "tg_tokens": null,
   "tps": 41.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-nelcgcx8-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: 4ihnlq6h, r42xbhr9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 323.7 56.7 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nelcgcx8",
   "source_url": "https://omlx.ai/benchmarks/performance/nelcgcx8",
   "tg_tokens": null,
   "tps": 56.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-nelcgcx8-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.3 GB at this context; duplicate traces of the same session: 4ihnlq6h, r42xbhr9; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 289.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 289.4 47.7 28.3 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nelcgcx8",
   "source_url": "https://omlx.ai/benchmarks/performance/nelcgcx8",
   "tg_tokens": null,
   "tps": 47.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-av8k2gjf-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 16.9 GB at this context; duplicate traces of the same session: 1ye3v5vb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 835.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 835.0 52.0 16.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/av8k2gjf",
   "source_url": "https://omlx.ai/benchmarks/performance/av8k2gjf",
   "tg_tokens": null,
   "tps": 52.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-av8k2gjf-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: 1ye3v5vb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 810.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 810.3 55.5 19.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/av8k2gjf",
   "source_url": "https://omlx.ai/benchmarks/performance/av8k2gjf",
   "tg_tokens": null,
   "tps": 55.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-fnm1pcln-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: u8i29zoy; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 472.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 472.0 22.2 17.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fnm1pcln",
   "source_url": "https://omlx.ai/benchmarks/performance/fnm1pcln",
   "tg_tokens": null,
   "tps": 22.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-fnm1pcln-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: u8i29zoy; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 481.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 481.9 24.3 18.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fnm1pcln",
   "source_url": "https://omlx.ai/benchmarks/performance/fnm1pcln",
   "tg_tokens": null,
   "tps": 24.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Pro",
   "id": "omlx-q4k8hdqx-ctx1024",
   "model": "Qwen3.8-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.6 GB at this context; duplicate traces of the same session: x2n375id; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 158.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-mlx-4Bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 1k 158.1 26.4 6.6 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q4k8hdqx",
   "source_url": "https://omlx.ai/benchmarks/performance/q4k8hdqx",
   "tg_tokens": null,
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-q4k8hdqx-ctx4096",
   "model": "Qwen3.8-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: x2n375id; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 198.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-9B-mlx-4Bit on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 4k 198.2 25.9 6.7 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q4k8hdqx",
   "source_url": "https://omlx.ai/benchmarks/performance/q4k8hdqx",
   "tg_tokens": null,
   "tps": 25.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M4 Max",
   "id": "omlx-mbrs6z5z-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 16k context; peak memory 44.7 GB; TTFT 45750 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 358.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k tokens 358.1 PP tok/s 9.2 TG tok/s 45750 TTFT (ms) 44.7 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mbrs6z5z",
   "source_url": "https://omlx.ai/benchmarks/performance/mbrs6z5z",
   "tg_tokens": null,
   "tps": 9.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-we947119-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; Swift 1.5 build; oMLX community benchmark trace; peak memory 92.4 GB at this context; duplicate traces of the same session: c5zbjzcn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 158.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 158.6 31.0 92.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/we947119",
   "source_url": "https://omlx.ai/benchmarks/performance/we947119",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-we947119-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; Swift 1.5 build; oMLX community benchmark trace; peak memory 95.1 GB at this context; duplicate traces of the same session: c5zbjzcn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 303.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Swift1.5-Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 303.8 28.0 95.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/we947119",
   "source_url": "https://omlx.ai/benchmarks/performance/we947119",
   "tg_tokens": null,
   "tps": 28.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-jworcvrb-ctx32768",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 32k context; peak memory 22.8 GB; TTFT 61151 ms; duplicate traces of the same session: none; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 535.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 32k tokens 535.9 PP tok/s 48.8 TG tok/s 61151 TTFT (ms) 22.8 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jworcvrb",
   "source_url": "https://omlx.ai/benchmarks/performance/jworcvrb",
   "tg_tokens": null,
   "tps": 48.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-3jln48ue-ctx65536",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 64k context; peak memory 23.9 GB; TTFT 173556 ms; duplicate traces of the same session: none; config=18c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 377.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M3 Pro (18c); M3 Pro (18c) \u00b7 36 GB \u00b7 4bit; 64k tokens 377.6 PP tok/s 30.3 TG tok/s 173556 TTFT (ms) 23.9 Peak mem",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3jln48ue",
   "source_url": "https://omlx.ai/benchmarks/performance/3jln48ue",
   "tg_tokens": null,
   "tps": 30.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-slutg0p2-ctx1024",
   "model": "gemma-4-31B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 18.1 GB; TTFT 11503 ms; duplicate traces of the same session: none; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 89.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "gemma-4-31B-it-oQ4e-mtp on M2 Max (30c); M2 Max (30c) \u00b7 32 GB \u00b7 4bit; 1k tokens 89.0 PP tok/s 15.5 TG tok/s 11503 TTFT (ms) 18.1 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/slutg0p2",
   "source_url": "https://omlx.ai/benchmarks/performance/slutg0p2",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-s3m73pkb-ctx1024",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 89.0 GB at this context; duplicate traces of the same session: i7yrrg1t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 348.7,
   "provenance": "community",
   "quant": "2.4bit mixed",
   "quote": "DeepSeek-V4-Flash-0731-2.4bit-mixed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 1k 348.7 41.9 89.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s3m73pkb",
   "source_url": "https://omlx.ai/benchmarks/performance/s3m73pkb",
   "tg_tokens": null,
   "tps": 41.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-s3m73pkb-ctx4096",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 89.9 GB at this context; duplicate traces of the same session: i7yrrg1t; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 532.9,
   "provenance": "community",
   "quant": "2.4bit mixed",
   "quote": "DeepSeek-V4-Flash-0731-2.4bit-mixed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 4k 532.9 31.5 89.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s3m73pkb",
   "source_url": "https://omlx.ai/benchmarks/performance/s3m73pkb",
   "tg_tokens": null,
   "tps": 31.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-6qo43czy-ctx1024",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 93.0 GB at this context; duplicate traces of the same session: r8vxuvwk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 109.3,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 1k 109.3 11.1 93.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6qo43czy",
   "source_url": "https://omlx.ai/benchmarks/performance/6qo43czy",
   "tg_tokens": null,
   "tps": 11.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-6qo43czy-ctx32768",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 96.7 GB at this context; duplicate traces of the same session: r8vxuvwk; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 525.0,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 unknown; 32k 525.0 14.1 96.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6qo43czy",
   "source_url": "https://omlx.ai/benchmarks/performance/6qo43czy",
   "tg_tokens": null,
   "tps": 14.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-uioi0dfc-ctx1024",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 35.6 GB at this context; duplicate traces of the same session: ieqm4dxa; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2818.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 1k 2,818 90.0 35.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uioi0dfc",
   "source_url": "https://omlx.ai/benchmarks/performance/uioi0dfc",
   "tg_tokens": null,
   "tps": 90.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-uioi0dfc-ctx4096",
   "model": "Ornith-1.5-35B-A3B-Abliterated",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 37.2 GB at this context; duplicate traces of the same session: ieqm4dxa; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3352.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Ornith-1.5-35B-A3B-Abliterated-MLX-8bit on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 4k 3,352 89.2 37.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uioi0dfc",
   "source_url": "https://omlx.ai/benchmarks/performance/uioi0dfc",
   "tg_tokens": null,
   "tps": 89.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-leqkaajr-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 65.8 GB at this context; duplicate traces of the same session: 89ukvi2p; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2036.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Ornith-1.5-35B-A3B-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 2,036 68.8 65.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/leqkaajr",
   "source_url": "https://omlx.ai/benchmarks/performance/leqkaajr",
   "tg_tokens": null,
   "tps": 68.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-leqkaajr-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 67.5 GB at this context; duplicate traces of the same session: 89ukvi2p; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3384.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Ornith-1.5-35B-A3B-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 3,384 68.2 67.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/leqkaajr",
   "source_url": "https://omlx.ai/benchmarks/performance/leqkaajr",
   "tg_tokens": null,
   "tps": 68.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-cbh3ur6r-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 65.8 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2580.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Ornith-1.5-35B-A3B-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 2,580 69.4 65.8 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cbh3ur6r",
   "source_url": "https://omlx.ai/benchmarks/performance/cbh3ur6r",
   "tg_tokens": null,
   "tps": 69.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-cbh3ur6r-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 67.5 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3411.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Ornith-1.5-35B-A3B-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 3,411 68.2 67.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cbh3ur6r",
   "source_url": "https://omlx.ai/benchmarks/performance/cbh3ur6r",
   "tg_tokens": null,
   "tps": 68.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-n9uq6srr-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 27.5 GB at this context; duplicate traces of the same session: 4na8ydf2, bfd8l3i9, remiog2c, zosoei7j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2380.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 1k 2,380 108.8 27.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n9uq6srr",
   "source_url": "https://omlx.ai/benchmarks/performance/n9uq6srr",
   "tg_tokens": null,
   "tps": 108.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-n9uq6srr-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.1 GB at this context; duplicate traces of the same session: 4na8ydf2, bfd8l3i9, remiog2c, zosoei7j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4194.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 4k 4,194 106.2 28.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n9uq6srr",
   "source_url": "https://omlx.ai/benchmarks/performance/n9uq6srr",
   "tg_tokens": null,
   "tps": 106.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-n9uq6srr-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 28.7 GB at this context; duplicate traces of the same session: 4na8ydf2, bfd8l3i9, remiog2c, zosoei7j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4424.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 32k 4,424 99.7 28.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n9uq6srr",
   "source_url": "https://omlx.ai/benchmarks/performance/n9uq6srr",
   "tg_tokens": null,
   "tps": 99.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-n9uq6srr-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 29.3 GB at this context; duplicate traces of the same session: 4na8ydf2, bfd8l3i9, remiog2c, zosoei7j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3469.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 64k 3,469 82.5 29.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n9uq6srr",
   "source_url": "https://omlx.ai/benchmarks/performance/n9uq6srr",
   "tg_tokens": null,
   "tps": 82.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-n9uq6srr-ctx131072",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 30.6 GB at this context; duplicate traces of the same session: 4na8ydf2, bfd8l3i9, remiog2c, zosoei7j; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2410.0,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Ornith-1.5-35B-A3B-MLX-6bit on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 6bit; 128k 2,410 71.2 30.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n9uq6srr",
   "source_url": "https://omlx.ai/benchmarks/performance/n9uq6srr",
   "tg_tokens": null,
   "tps": 71.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-zyhyy01k-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.1 GB at this context; duplicate traces of the same session: ifr28r3o; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1653.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 4k 1,653 76.7 24.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zyhyy01k",
   "source_url": "https://omlx.ai/benchmarks/performance/zyhyy01k",
   "tg_tokens": null,
   "tps": 76.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Ultra",
   "id": "omlx-zyhyy01k-ctx8192",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: ifr28r3o; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1577.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 8k 1,577 82.1 24.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zyhyy01k",
   "source_url": "https://omlx.ai/benchmarks/performance/zyhyy01k",
   "tg_tokens": null,
   "tps": 82.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-zyhyy01k-ctx16384",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.2 GB at this context; duplicate traces of the same session: ifr28r3o; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1382.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 16k 1,382 96.4 25.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zyhyy01k",
   "source_url": "https://omlx.ai/benchmarks/performance/zyhyy01k",
   "tg_tokens": null,
   "tps": 96.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-zyhyy01k-ctx32768",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.5 GB at this context; duplicate traces of the same session: ifr28r3o; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1102.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 32k 1,102 121.7 26.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zyhyy01k",
   "source_url": "https://omlx.ai/benchmarks/performance/zyhyy01k",
   "tg_tokens": null,
   "tps": 121.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-zyhyy01k-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.3 GB at this context; duplicate traces of the same session: ifr28r3o; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 786.1,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 64k 786.1 65.2 27.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zyhyy01k",
   "source_url": "https://omlx.ai/benchmarks/performance/zyhyy01k",
   "tg_tokens": null,
   "tps": 65.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Ultra",
   "id": "omlx-zyhyy01k-ctx131072",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.8 GB at this context; duplicate traces of the same session: ifr28r3o; config=48c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 503.9,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-fp16-mtp on M1 Ultra (48c); M1 Ultra (48c) \u00b7 64 GB \u00b7 4bit; 128k 503.9 42.7 28.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zyhyy01k",
   "source_url": "https://omlx.ai/benchmarks/performance/zyhyy01k",
   "tg_tokens": null,
   "tps": 42.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-sii0mmwp-ctx1024",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.9 GB at this context; duplicate traces of the same session: none; config=19c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 249.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M2 Pro (19c); M2 Pro (19c) \u00b7 32 GB \u00b7 4bit; 1k 249.3 35.4 5.9 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sii0mmwp",
   "source_url": "https://omlx.ai/benchmarks/performance/sii0mmwp",
   "tg_tokens": null,
   "tps": 35.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Pro",
   "id": "omlx-sii0mmwp-ctx4096",
   "model": "Qwen3.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.5 GB at this context; duplicate traces of the same session: none; config=19c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 253.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.5-9B-MLX-4bit on M2 Pro (19c); M2 Pro (19c) \u00b7 32 GB \u00b7 4bit; 4k 253.1 34.1 6.5 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/sii0mmwp",
   "source_url": "https://omlx.ai/benchmarks/performance/sii0mmwp",
   "tg_tokens": null,
   "tps": 34.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-knx1oq9e-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1568.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,568 141.6 19.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/knx1oq9e",
   "source_url": "https://omlx.ai/benchmarks/performance/knx1oq9e",
   "tg_tokens": null,
   "tps": 141.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-knx1oq9e-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1995.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,995 138.1 21.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/knx1oq9e",
   "source_url": "https://omlx.ai/benchmarks/performance/knx1oq9e",
   "tg_tokens": null,
   "tps": 138.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-9837k61p-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 16.0 GB; TTFT 10506 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 97.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 1k tokens 97.5 PP tok/s 17.3 TG tok/s 10506 TTFT (ms) 16.0 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9837k61p",
   "source_url": "https://omlx.ai/benchmarks/performance/9837k61p",
   "tg_tokens": null,
   "tps": 17.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Max",
   "id": "omlx-es7iejuh-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: u9g6ndxq; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 150.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 1k 150.9 19.7 18.5 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/es7iejuh",
   "source_url": "https://omlx.ai/benchmarks/performance/es7iejuh",
   "tg_tokens": null,
   "tps": 19.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Max",
   "id": "omlx-es7iejuh-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.5 GB at this context; duplicate traces of the same session: u9g6ndxq; config=30c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 127.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Max (30c); M3 Max (30c) \u00b7 36 GB \u00b7 4bit; 4k 127.7 27.7 19.5 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/es7iejuh",
   "source_url": "https://omlx.ai/benchmarks/performance/es7iejuh",
   "tg_tokens": null,
   "tps": 27.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-v2yo0znw-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.8 GB at this context; duplicate traces of the same session: k4onfq54; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 254.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 254.1 60.6 21.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v2yo0znw",
   "source_url": "https://omlx.ai/benchmarks/performance/v2yo0znw",
   "tg_tokens": null,
   "tps": 60.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-v2yo0znw-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.1 GB at this context; duplicate traces of the same session: k4onfq54; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 338.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 338.5 78.5 25.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v2yo0znw",
   "source_url": "https://omlx.ai/benchmarks/performance/v2yo0znw",
   "tg_tokens": null,
   "tps": 78.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-v2yo0znw-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 25.6 GB at this context; duplicate traces of the same session: k4onfq54; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 340.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 340.6 41.2 25.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v2yo0znw",
   "source_url": "https://omlx.ai/benchmarks/performance/v2yo0znw",
   "tg_tokens": null,
   "tps": 41.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-v2yo0znw-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.6 GB at this context; duplicate traces of the same session: k4onfq54; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 323.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 323.7 56.7 26.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v2yo0znw",
   "source_url": "https://omlx.ai/benchmarks/performance/v2yo0znw",
   "tg_tokens": null,
   "tps": 56.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-v2yo0znw-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.3 GB at this context; duplicate traces of the same session: k4onfq54; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 289.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 289.4 47.7 28.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/v2yo0znw",
   "source_url": "https://omlx.ai/benchmarks/performance/v2yo0znw",
   "tg_tokens": null,
   "tps": 47.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0l49wu0u-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 409.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 1k 409.1 55.8 26.7 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0l49wu0u",
   "source_url": "https://omlx.ai/benchmarks/performance/0l49wu0u",
   "tg_tokens": null,
   "tps": 55.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0l49wu0u-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.6 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 396.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 32k 396.5 66.3 32.6 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0l49wu0u",
   "source_url": "https://omlx.ai/benchmarks/performance/0l49wu0u",
   "tg_tokens": null,
   "tps": 66.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-oafdowtp-ctx65536",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 43.2 GB at this context; duplicate traces of the same session: lbw5risz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2122.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 64k 2,122 55.2 43.2 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oafdowtp",
   "source_url": "https://omlx.ai/benchmarks/performance/oafdowtp",
   "tg_tokens": null,
   "tps": 55.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-oafdowtp-ctx131072",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 45.5 GB at this context; duplicate traces of the same session: lbw5risz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1956.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k 1,956 59.1 45.5 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oafdowtp",
   "source_url": "https://omlx.ai/benchmarks/performance/oafdowtp",
   "tg_tokens": null,
   "tps": 59.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mbbc0pcu-ctx1024",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 39.8 GB at this context; duplicate traces of the same session: ty1r5wp1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 960.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 960.6 43.8 39.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mbbc0pcu",
   "source_url": "https://omlx.ai/benchmarks/performance/mbbc0pcu",
   "tg_tokens": null,
   "tps": 43.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-mbbc0pcu-ctx4096",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace; peak memory 40.0 GB at this context; duplicate traces of the same session: ty1r5wp1; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1601.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,601 46.7 40.0 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/mbbc0pcu",
   "source_url": "https://omlx.ai/benchmarks/performance/mbbc0pcu",
   "tg_tokens": null,
   "tps": 46.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-hbuaono9-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: q34z4gk5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 892.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 892.1 99.9 21.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hbuaono9",
   "source_url": "https://omlx.ai/benchmarks/performance/hbuaono9",
   "tg_tokens": null,
   "tps": 99.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-hbuaono9-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: q34z4gk5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 900.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 900.6 100.6 22.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/hbuaono9",
   "source_url": "https://omlx.ai/benchmarks/performance/hbuaono9",
   "tg_tokens": null,
   "tps": 100.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ekj6v7tg-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 9x2sh3fv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 905.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 905.2 88.8 21.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ekj6v7tg",
   "source_url": "https://omlx.ai/benchmarks/performance/ekj6v7tg",
   "tg_tokens": null,
   "tps": 88.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ekj6v7tg-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 9x2sh3fv; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 954.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 954.1 88.9 22.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ekj6v7tg",
   "source_url": "https://omlx.ai/benchmarks/performance/ekj6v7tg",
   "tg_tokens": null,
   "tps": 88.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-aex977o3-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: m14tweei; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 904.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 904.9 91.6 21.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aex977o3",
   "source_url": "https://omlx.ai/benchmarks/performance/aex977o3",
   "tg_tokens": null,
   "tps": 91.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-aex977o3-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: m14tweei; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 918.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 918.2 104.9 22.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/aex977o3",
   "source_url": "https://omlx.ai/benchmarks/performance/aex977o3",
   "tg_tokens": null,
   "tps": 104.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-eylvk30a-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 06os2t5y; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 901.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 901.3 98.9 21.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eylvk30a",
   "source_url": "https://omlx.ai/benchmarks/performance/eylvk30a",
   "tg_tokens": null,
   "tps": 98.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-eylvk30a-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 06os2t5y; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 916.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 916.0 95.1 22.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/eylvk30a",
   "source_url": "https://omlx.ai/benchmarks/performance/eylvk30a",
   "tg_tokens": null,
   "tps": 95.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-wmo7xalq-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 9fo6j503; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 905.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 905.7 108.9 21.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wmo7xalq",
   "source_url": "https://omlx.ai/benchmarks/performance/wmo7xalq",
   "tg_tokens": null,
   "tps": 108.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-wmo7xalq-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 9fo6j503; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 952.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 952.4 96.0 22.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wmo7xalq",
   "source_url": "https://omlx.ai/benchmarks/performance/wmo7xalq",
   "tg_tokens": null,
   "tps": 96.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-2n0kn4l3-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: tmfqfyyz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 903.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 903.4 94.9 21.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2n0kn4l3",
   "source_url": "https://omlx.ai/benchmarks/performance/2n0kn4l3",
   "tg_tokens": null,
   "tps": 94.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-2n0kn4l3-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: tmfqfyyz; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 954.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 954.4 92.3 22.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2n0kn4l3",
   "source_url": "https://omlx.ai/benchmarks/performance/2n0kn4l3",
   "tg_tokens": null,
   "tps": 92.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-nv42y8vk-ctx32768",
   "model": "DeepSeek-R1-0528-Qwen3-8B",
   "notes": "oMLX community benchmark trace, single run at 32k context; peak memory 9.3 GB; TTFT 56921 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 575.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "DeepSeek-R1-0528-Qwen3-8B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 48 GB \u00b7 4bit; 32k tokens 575.7 PP tok/s 15.3 TG tok/s 56921 TTFT (ms) 9.3 Peak mem",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nv42y8vk",
   "source_url": "https://omlx.ai/benchmarks/performance/nv42y8vk",
   "tg_tokens": null,
   "tps": 15.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-yiwc8qgn-ctx1024",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 87.8 GB at this context; duplicate traces of the same session: thfnsk3s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 642.4,
   "provenance": "community",
   "quant": "2.4bit mixed",
   "quote": "DeepSeek-V4-Flash-0731-2.4bit-mixed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 1k 642.4 33.2 87.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yiwc8qgn",
   "source_url": "https://omlx.ai/benchmarks/performance/yiwc8qgn",
   "tg_tokens": null,
   "tps": 33.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-yiwc8qgn-ctx4096",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 88.7 GB at this context; duplicate traces of the same session: thfnsk3s; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 678.2,
   "provenance": "community",
   "quant": "2.4bit mixed",
   "quote": "DeepSeek-V4-Flash-0731-2.4bit-mixed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 4k 678.2 44.4 88.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/yiwc8qgn",
   "source_url": "https://omlx.ai/benchmarks/performance/yiwc8qgn",
   "tg_tokens": null,
   "tps": 44.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-w7en5yib-ctx1024",
   "model": "Inkling-Small",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 85.8 GB at this context; duplicate traces of the same session: c7cnjsyi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 889.7,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Inkling-Small-mlx-2bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 1k 889.7 52.3 85.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w7en5yib",
   "source_url": "https://omlx.ai/benchmarks/performance/w7en5yib",
   "tg_tokens": null,
   "tps": 52.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-w7en5yib-ctx4096",
   "model": "Inkling-Small",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 92.8 GB at this context; duplicate traces of the same session: c7cnjsyi; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1248.0,
   "provenance": "community",
   "quant": "2bit",
   "quote": "Inkling-Small-mlx-2bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 4k 1,248 52.5 92.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/w7en5yib",
   "source_url": "https://omlx.ai/benchmarks/performance/w7en5yib",
   "tg_tokens": null,
   "tps": 52.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-niy142gj-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1280.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,280 86.7 24.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/niy142gj",
   "source_url": "https://omlx.ai/benchmarks/performance/niy142gj",
   "tg_tokens": null,
   "tps": 86.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-niy142gj-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1188.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 1,188 75.3 24.7 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/niy142gj",
   "source_url": "https://omlx.ai/benchmarks/performance/niy142gj",
   "tg_tokens": null,
   "tps": 75.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1f8z3g59-ctx1024",
   "model": "Laguna-S-2.1",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 67.8 GB at this context; duplicate traces of the same session: 5mp1yusu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 780.6,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Laguna-S-2.1-NVFP4-mlx on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 780.6 39.2 67.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1f8z3g59",
   "source_url": "https://omlx.ai/benchmarks/performance/1f8z3g59",
   "tg_tokens": null,
   "tps": 39.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-1f8z3g59-ctx4096",
   "model": "Laguna-S-2.1",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 67.9 GB at this context; duplicate traces of the same session: 5mp1yusu; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1355.0,
   "provenance": "community",
   "quant": "NVFP4",
   "quote": "Laguna-S-2.1-NVFP4-mlx on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,355 44.5 67.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1f8z3g59",
   "source_url": "https://omlx.ai/benchmarks/performance/1f8z3g59",
   "tg_tokens": null,
   "tps": 44.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-s3ry0d2e-ctx1024",
   "model": "Laguna-S-2.1",
   "notes": "oMLX community benchmark trace; peak memory 60.4 GB at this context; duplicate traces of the same session: xuribbwn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 680.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Laguna-S-2.1-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 680.8 49.8 60.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s3ry0d2e",
   "source_url": "https://omlx.ai/benchmarks/performance/s3ry0d2e",
   "tg_tokens": null,
   "tps": 49.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-s3ry0d2e-ctx4096",
   "model": "Laguna-S-2.1",
   "notes": "oMLX community benchmark trace; peak memory 60.5 GB at this context; duplicate traces of the same session: xuribbwn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1496.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Laguna-S-2.1-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,496 63.4 60.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/s3ry0d2e",
   "source_url": "https://omlx.ai/benchmarks/performance/s3ry0d2e",
   "tg_tokens": null,
   "tps": 63.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-8wj04xiw-ctx1024",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 18.9 GB at this context; duplicate traces of the same session: fxfe3x55; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 675.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 675.7 94.7 18.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8wj04xiw",
   "source_url": "https://omlx.ai/benchmarks/performance/8wj04xiw",
   "tg_tokens": null,
   "tps": 94.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-8wj04xiw-ctx4096",
   "model": "NVIDIA Nemotron 3.5 Lightning",
   "notes": "oMLX community benchmark trace; peak memory 19.8 GB at this context; duplicate traces of the same session: fxfe3x55; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 700.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 700.8 93.2 19.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8wj04xiw",
   "source_url": "https://omlx.ai/benchmarks/performance/8wj04xiw",
   "tg_tokens": null,
   "tps": 93.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-omcwwx59-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 65.8 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2580.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Ornith-1.5-35B-A3B-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 1k 2,580 69.4 65.8 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/omcwwx59",
   "source_url": "https://omlx.ai/benchmarks/performance/omcwwx59",
   "tg_tokens": null,
   "tps": 69.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-omcwwx59-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MLX; oMLX community benchmark trace; peak memory 67.5 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3411.0,
   "provenance": "community",
   "quant": "unknown",
   "quote": "Ornith-1.5-35B-A3B-MLX on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 3,411 68.2 67.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/omcwwx59",
   "source_url": "https://omlx.ai/benchmarks/performance/omcwwx59",
   "tg_tokens": null,
   "tps": 68.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-x5kbtvtu-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "refusal-ablated MTP build; MLX; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 950.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-refusal-ablated-MLX-oQ8-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 950.0 83.4 37.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x5kbtvtu",
   "source_url": "https://omlx.ai/benchmarks/performance/x5kbtvtu",
   "tg_tokens": null,
   "tps": 83.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-x5kbtvtu-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "refusal-ablated MTP build; MLX; oMLX community benchmark trace; peak memory 39.5 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 879.8,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-refusal-ablated-MLX-oQ8-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 879.8 59.4 39.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/x5kbtvtu",
   "source_url": "https://omlx.ai/benchmarks/performance/x5kbtvtu",
   "tg_tokens": null,
   "tps": 59.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-jyfyq96u-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: vwyw3gbn; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 765.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 1k 765.4 93.9 19.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jyfyq96u",
   "source_url": "https://omlx.ai/benchmarks/performance/jyfyq96u",
   "tg_tokens": null,
   "tps": 93.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-jyfyq96u-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: vwyw3gbn; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 898.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-4bit on M4 Pro (16c); M4 Pro (16c) \u00b7 64 GB \u00b7 4bit; 4k 898.0 91.4 21.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jyfyq96u",
   "source_url": "https://omlx.ai/benchmarks/performance/jyfyq96u",
   "tg_tokens": null,
   "tps": 91.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-lkv66wih-ctx1024",
   "model": "Qwen3.6-35B-A3B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 19.4 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1568.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 1,568 141.6 19.4 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lkv66wih",
   "source_url": "https://omlx.ai/benchmarks/performance/lkv66wih",
   "tg_tokens": null,
   "tps": 141.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-lkv66wih-ctx4096",
   "model": "Qwen3.6-35B-A3B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 21.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1995.0,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-35B-A3B-Uncensored-4bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 1,995 138.1 21.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lkv66wih",
   "source_url": "https://omlx.ai/benchmarks/performance/lkv66wih",
   "tg_tokens": null,
   "tps": 138.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-3ifvs22b-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 19.2 GB at this context; duplicate traces of the same session: ekxz1rav; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 246.2,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 246.2 32.5 19.2 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ifvs22b",
   "source_url": "https://omlx.ai/benchmarks/performance/3ifvs22b",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-3ifvs22b-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: ekxz1rav; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 231.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M4 Max (40c); M4 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 231.7 32.5 22.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3ifvs22b",
   "source_url": "https://omlx.ai/benchmarks/performance/3ifvs22b",
   "tg_tokens": null,
   "tps": 32.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-cfk3k0mr-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: y00phc65; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 254.2,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Uncensored-6bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k 254.2 20.9 22.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cfk3k0mr",
   "source_url": "https://omlx.ai/benchmarks/performance/cfk3k0mr",
   "tg_tokens": null,
   "tps": 20.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-cfk3k0mr-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 25.4 GB at this context; duplicate traces of the same session: y00phc65; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 269.5,
   "provenance": "community",
   "quant": "6bit",
   "quote": "Qwen3.8-27B-Uncensored-6bit on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 6bit; 4k 269.5 20.6 25.4 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/cfk3k0mr",
   "source_url": "https://omlx.ai/benchmarks/performance/cfk3k0mr",
   "tg_tokens": null,
   "tps": 20.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-lb7ca9pq-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: 3sm1v5wk; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 81.9,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M3 Pro (14c); M3 Pro (14c) \u00b7 36 GB \u00b7 4bit; 4k 81.9 7.1 21.3 GB current",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lb7ca9pq",
   "source_url": "https://omlx.ai/benchmarks/performance/lb7ca9pq",
   "tg_tokens": null,
   "tps": 7.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Pro",
   "id": "omlx-lb7ca9pq-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 3sm1v5wk; config=14c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 77.9,
   "provenance": "community",
   "quant": "OptiQ-4bit",
   "quote": "Qwen3.8-27B-Uncensored-OptiQ-4bit on M3 Pro (14c); M3 Pro (14c) \u00b7 36 GB \u00b7 4bit; 16k 77.9 6.8 22.3 GB view",
   "ram_gb": 36,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lb7ca9pq",
   "source_url": "https://omlx.ai/benchmarks/performance/lb7ca9pq",
   "tg_tokens": null,
   "tps": 6.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ytaiw2a4-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: f6t7lkvq, iky6n84e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1330.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,330 60.4 19.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ytaiw2a4",
   "source_url": "https://omlx.ai/benchmarks/performance/ytaiw2a4",
   "tg_tokens": null,
   "tps": 60.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ytaiw2a4-ctx16384",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.9 GB at this context; duplicate traces of the same session: f6t7lkvq, iky6n84e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1358.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 1,358 75.7 19.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ytaiw2a4",
   "source_url": "https://omlx.ai/benchmarks/performance/ytaiw2a4",
   "tg_tokens": null,
   "tps": 75.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-ytaiw2a4-ctx65536",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.3 GB at this context; duplicate traces of the same session: f6t7lkvq, iky6n84e; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1019.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-Uncensored-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 1,019 61.7 23.3 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ytaiw2a4",
   "source_url": "https://omlx.ai/benchmarks/performance/ytaiw2a4",
   "tg_tokens": null,
   "tps": 61.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-at8wc1y4-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.2 GB at this context; duplicate traces of the same session: k8d7bmw0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 322.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 322.7 41.0 21.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/at8wc1y4",
   "source_url": "https://omlx.ai/benchmarks/performance/at8wc1y4",
   "tg_tokens": null,
   "tps": 41.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-at8wc1y4-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 24.5 GB at this context; duplicate traces of the same session: k8d7bmw0; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 428.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 428.4 62.3 24.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/at8wc1y4",
   "source_url": "https://omlx.ai/benchmarks/performance/at8wc1y4",
   "tg_tokens": null,
   "tps": 62.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-oic915qj-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: zx7fncjv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 878.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 878.5 40.3 17.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oic915qj",
   "source_url": "https://omlx.ai/benchmarks/performance/oic915qj",
   "tg_tokens": null,
   "tps": 40.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-oic915qj-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: zx7fncjv; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 917.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 917.9 56.1 18.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oic915qj",
   "source_url": "https://omlx.ai/benchmarks/performance/oic915qj",
   "tg_tokens": null,
   "tps": 56.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0ur0zkty-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 26.7 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 409.1,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 1k 409.1 55.8 26.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ur0zkty",
   "source_url": "https://omlx.ai/benchmarks/performance/0ur0zkty",
   "tg_tokens": null,
   "tps": 55.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-0ur0zkty-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 32.6 GB at this context; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 396.5,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-oQ6e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 6bit; 32k 396.5 66.3 32.6 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/0ur0zkty",
   "source_url": "https://omlx.ai/benchmarks/performance/0ur0zkty",
   "tg_tokens": null,
   "tps": 66.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-7cnrdtm5-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 70.1 GB at this context; duplicate traces of the same session: 8w8983xq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 813.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 813.7 59.0 70.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7cnrdtm5",
   "source_url": "https://omlx.ai/benchmarks/performance/7cnrdtm5",
   "tg_tokens": null,
   "tps": 59.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Max",
   "id": "omlx-7cnrdtm5-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: 8w8983xq; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 762.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 32k 762.1 60.4 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7cnrdtm5",
   "source_url": "https://omlx.ai/benchmarks/performance/7cnrdtm5",
   "tg_tokens": null,
   "tps": 60.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-rftpzc53-ctx8192",
   "model": "Qwen3.8-Flash-Next-REAP-288",
   "notes": "oMLX community benchmark trace, single run at 8k context; peak memory 40.6 GB; TTFT 8951 ms; duplicate traces of the same session: none; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 915.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-Flash-Next-REAP-288-MLX-4bit on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 8k tokens 915.3 PP tok/s 53.7 TG tok/s 8951 TTFT (ms) 40.6 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rftpzc53",
   "source_url": "https://omlx.ai/benchmarks/performance/rftpzc53",
   "tg_tokens": null,
   "tps": 53.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-ppyvqim4-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.8 GB; TTFT 3091 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 331.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k tokens 331.3 PP tok/s 35.5 TG tok/s 3091 TTFT (ms) 70.8 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ppyvqim4",
   "source_url": "https://omlx.ai/benchmarks/performance/ppyvqim4",
   "tg_tokens": null,
   "tps": 35.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-fll909kj-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.7 GB; TTFT 5030 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 203.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k tokens 203.6 PP tok/s 39.0 TG tok/s 5030 TTFT (ms) 70.7 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fll909kj",
   "source_url": "https://omlx.ai/benchmarks/performance/fll909kj",
   "tg_tokens": null,
   "tps": 39.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-kf3zc7b8-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.3 GB at this context; duplicate traces of the same session: b6suf82h; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 509.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 1k 509.4 58.7 69.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kf3zc7b8",
   "source_url": "https://omlx.ai/benchmarks/performance/kf3zc7b8",
   "tg_tokens": null,
   "tps": 58.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-kf3zc7b8-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 69.8 GB at this context; duplicate traces of the same session: b6suf82h; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 802.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 4bit; 4k 802.9 56.2 69.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kf3zc7b8",
   "source_url": "https://omlx.ai/benchmarks/performance/kf3zc7b8",
   "tg_tokens": null,
   "tps": 56.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-nebb0raa-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 191.2 GB at this context; duplicate traces of the same session: rrb0jq2c, w1mklga2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1303.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 64k 1,303 66.7 191.2 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nebb0raa",
   "source_url": "https://omlx.ai/benchmarks/performance/nebb0raa",
   "tg_tokens": null,
   "tps": 66.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nebb0raa-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 192.9 GB at this context; duplicate traces of the same session: rrb0jq2c, w1mklga2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1282.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 128k 1,282 54.3 192.9 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nebb0raa",
   "source_url": "https://omlx.ai/benchmarks/performance/nebb0raa",
   "tg_tokens": null,
   "tps": 54.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-nebb0raa-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 196.5 GB at this context; duplicate traces of the same session: rrb0jq2c, w1mklga2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1262.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 8bit; 195k 1,262 58.3 196.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nebb0raa",
   "source_url": "https://omlx.ai/benchmarks/performance/nebb0raa",
   "tg_tokens": null,
   "tps": 58.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-063cy4ki-ctx16384",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "mlx-community build; oMLX community benchmark trace, single run at 16k context; peak memory 20.1 GB; TTFT 64168 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 255.3,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "mlx-community--MiniCPM-V-4.7-35B-A3B-mxfp4 on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k tokens 255.3 PP tok/s 17.8 TG tok/s 64168 TTFT (ms) 20.1 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/063cy4ki",
   "source_url": "https://omlx.ai/benchmarks/performance/063cy4ki",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-2d6aezv2-ctx32768",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 206.0 GB; TTFT 34622 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 946.4,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 unknown; 32k tokens 946.4 PP tok/s 44.8 TG tok/s 34622 TTFT (ms) 206.0 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2d6aezv2",
   "source_url": "https://omlx.ai/benchmarks/performance/2d6aezv2",
   "tg_tokens": null,
   "tps": 44.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-obvtt00j-ctx32768",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 32k context; peak memory 206.0 GB; TTFT 33363 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 982.2,
   "provenance": "community",
   "quant": "oQ2.5e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 unknown; 32k tokens 982.2 PP tok/s 53.1 TG tok/s 33363 TTFT (ms) 206.0 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/obvtt00j",
   "source_url": "https://omlx.ai/benchmarks/performance/obvtt00j",
   "tg_tokens": null,
   "tps": 53.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-2j92ignq-ctx4096",
   "model": "Gemma-4-26B-A4B-Heretic",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 20.2 GB at this context; duplicate traces of the same session: 53zgbm9b; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 880.2,
   "provenance": "community",
   "quant": "Q6",
   "quote": "Gemma-4-26B-A4B-Heretic-MLX-q6 on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 6bit; 4k 880.2 31.6 20.2 GB current",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2j92ignq",
   "source_url": "https://omlx.ai/benchmarks/performance/2j92ignq",
   "tg_tokens": null,
   "tps": 31.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5",
   "id": "omlx-2j92ignq-ctx16384",
   "model": "Gemma-4-26B-A4B-Heretic",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 20.8 GB at this context; duplicate traces of the same session: 53zgbm9b; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 841.3,
   "provenance": "community",
   "quant": "Q6",
   "quote": "Gemma-4-26B-A4B-Heretic-MLX-q6 on M5 (10c); M5 (10c) \u00b7 32 GB \u00b7 6bit; 16k 841.3 27.2 20.8 GB view",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2j92ignq",
   "source_url": "https://omlx.ai/benchmarks/performance/2j92ignq",
   "tg_tokens": null,
   "tps": 27.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-n2jem4jx-ctx1024",
   "model": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated",
   "notes": "oMLX community benchmark trace; peak memory 35.6 GB at this context; duplicate traces of the same session: axwol7k8; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1014.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated-mlx-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,014 69.3 35.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n2jem4jx",
   "source_url": "https://omlx.ai/benchmarks/performance/n2jem4jx",
   "tg_tokens": null,
   "tps": 69.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-n2jem4jx-ctx4096",
   "model": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated",
   "notes": "oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: axwol7k8; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 922.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated-mlx-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 922.0 60.8 37.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n2jem4jx",
   "source_url": "https://omlx.ai/benchmarks/performance/n2jem4jx",
   "tg_tokens": null,
   "tps": 60.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-lzktj2lx-ctx1024",
   "model": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated",
   "notes": "oMLX community benchmark trace; peak memory 35.6 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1013.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated-mlx-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,013 72.0 35.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lzktj2lx",
   "source_url": "https://omlx.ai/benchmarks/performance/lzktj2lx",
   "tg_tokens": null,
   "tps": 72.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-lzktj2lx-ctx4096",
   "model": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated",
   "notes": "oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 952.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated-mlx-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 952.0 62.9 37.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lzktj2lx",
   "source_url": "https://omlx.ai/benchmarks/performance/lzktj2lx",
   "tg_tokens": null,
   "tps": 62.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-g8a5a3da-ctx1024",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 24.3 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1280.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,280 86.7 24.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g8a5a3da",
   "source_url": "https://omlx.ai/benchmarks/performance/g8a5a3da",
   "tg_tokens": null,
   "tps": 86.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-g8a5a3da-ctx4096",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 24.7 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1188.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "LFM2.5-24B-A2B-MLX-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 1,188 75.3 24.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g8a5a3da",
   "source_url": "https://omlx.ai/benchmarks/performance/g8a5a3da",
   "tg_tokens": null,
   "tps": 75.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-41wdu09m-ctx1024",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: bepiw1g0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1098.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiniCPM-V-4.7-35B-A3B-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 1,098 141.8 20.9 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/41wdu09m",
   "source_url": "https://omlx.ai/benchmarks/performance/41wdu09m",
   "tg_tokens": null,
   "tps": 141.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-41wdu09m-ctx4096",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 21.6 GB at this context; duplicate traces of the same session: bepiw1g0; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3948.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "MiniCPM-V-4.7-35B-A3B-oQ4e on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 3,948 137.6 21.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/41wdu09m",
   "source_url": "https://omlx.ai/benchmarks/performance/41wdu09m",
   "tg_tokens": null,
   "tps": 137.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-3znqcr99-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "refusal-ablated MTP build; MLX; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 950.0,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-refusal-ablated-MLX-oQ8-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 950.0 83.4 37.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3znqcr99",
   "source_url": "https://omlx.ai/benchmarks/performance/3znqcr99",
   "tg_tokens": null,
   "tps": 83.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-3znqcr99-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "refusal-ablated MTP build; MLX; oMLX community benchmark trace; peak memory 39.5 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 879.8,
   "provenance": "community",
   "quant": "oQ8",
   "quote": "Ornith-1.5-35B-A3B-refusal-ablated-MLX-oQ8-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 879.8 59.4 39.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3znqcr99",
   "source_url": "https://omlx.ai/benchmarks/performance/3znqcr99",
   "tg_tokens": null,
   "tps": 59.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-72rnkbzx-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: wakqj8wt; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 948.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 948.2 110.8 21.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/72rnkbzx",
   "source_url": "https://omlx.ai/benchmarks/performance/72rnkbzx",
   "tg_tokens": null,
   "tps": 110.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-72rnkbzx-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.4 GB at this context; duplicate traces of the same session: wakqj8wt; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 948.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.6-35B-A3B-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 948.1 83.6 23.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/72rnkbzx",
   "source_url": "https://omlx.ai/benchmarks/performance/72rnkbzx",
   "tg_tokens": null,
   "tps": 83.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-htanugkh-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: wvvn4gbc; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 952.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 952.3 102.4 37.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/htanugkh",
   "source_url": "https://omlx.ai/benchmarks/performance/htanugkh",
   "tg_tokens": null,
   "tps": 102.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-htanugkh-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: wvvn4gbc; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 867.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 867.1 62.8 39.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/htanugkh",
   "source_url": "https://omlx.ai/benchmarks/performance/htanugkh",
   "tg_tokens": null,
   "tps": 62.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-i1g77rxg-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: 6cbztkpr; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 880.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 880.1 104.1 37.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i1g77rxg",
   "source_url": "https://omlx.ai/benchmarks/performance/i1g77rxg",
   "tg_tokens": null,
   "tps": 104.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-i1g77rxg-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: 6cbztkpr; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 899.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 899.0 64.9 39.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/i1g77rxg",
   "source_url": "https://omlx.ai/benchmarks/performance/i1g77rxg",
   "tg_tokens": null,
   "tps": 64.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-whfezfk0-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 961.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 961.4 83.2 37.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/whfezfk0",
   "source_url": "https://omlx.ai/benchmarks/performance/whfezfk0",
   "tg_tokens": null,
   "tps": 83.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-whfezfk0-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 912.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 912.7 68.6 39.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/whfezfk0",
   "source_url": "https://omlx.ai/benchmarks/performance/whfezfk0",
   "tg_tokens": null,
   "tps": 68.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pj7auy83-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 128k context; peak memory 25.1 GB; TTFT 203116 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 645.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 128k tokens 645.3 PP tok/s 21.6 TG tok/s 203116 TTFT (ms) 25.1 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pj7auy83",
   "source_url": "https://omlx.ai/benchmarks/performance/pj7auy83",
   "tg_tokens": null,
   "tps": 21.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M1 Max",
   "id": "omlx-ge9bze6n-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 8k context; peak memory 20.9 GB; TTFT 149060 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 55.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 8k tokens 55.0 PP tok/s 2.1 TG tok/s 149060 TTFT (ms) 20.9 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ge9bze6n",
   "source_url": "https://omlx.ai/benchmarks/performance/ge9bze6n",
   "tg_tokens": null,
   "tps": 2.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-r1lyoa1u-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 18.9 GB; TTFT 12605 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 81.2,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 1k tokens 81.2 PP tok/s 14.1 TG tok/s 12605 TTFT (ms) 18.9 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r1lyoa1u",
   "source_url": "https://omlx.ai/benchmarks/performance/r1lyoa1u",
   "tg_tokens": null,
   "tps": 14.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction",
    "outlier"
   ],
   "hardware": "M1 Max",
   "id": "omlx-d094n1f3-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 16.6 GB; TTFT 48764 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 21.0,
   "provenance": "community",
   "quant": "oQ4e-fp16",
   "quote": "Qwen3.8-27B-oQ4e-fp16-mtp on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 1k tokens 21.0 PP tok/s 2.4 TG tok/s 48764 TTFT (ms) 16.6 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/d094n1f3",
   "source_url": "https://omlx.ai/benchmarks/performance/d094n1f3",
   "tg_tokens": null,
   "tps": 2.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-p1ag24wn-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: snzdm9xf; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 105.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 1k 105.4 16.3 18.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p1ag24wn",
   "source_url": "https://omlx.ai/benchmarks/performance/p1ag24wn",
   "tg_tokens": null,
   "tps": 16.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-p1ag24wn-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 22.4 GB at this context; duplicate traces of the same session: snzdm9xf; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 101.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M1 Max (32c); M1 Max (32c) \u00b7 64 GB \u00b7 4bit; 4k 101.5 22.2 22.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/p1ag24wn",
   "source_url": "https://omlx.ai/benchmarks/performance/p1ag24wn",
   "tg_tokens": null,
   "tps": 22.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-4q1kz33e-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace, single run at 1k context; peak memory 72.4 GB; TTFT 3402 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 301.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k tokens 301.0 PP tok/s 26.6 TG tok/s 3402 TTFT (ms) 72.4 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4q1kz33e",
   "source_url": "https://omlx.ai/benchmarks/performance/4q1kz33e",
   "tg_tokens": null,
   "tps": 26.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-kmuyvsxr-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 72.5 GB at this context; duplicate traces of the same session: q9487x6d; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 247.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 247.3 31.0 72.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kmuyvsxr",
   "source_url": "https://omlx.ai/benchmarks/performance/kmuyvsxr",
   "tg_tokens": null,
   "tps": 31.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-kmuyvsxr-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 74.8 GB at this context; duplicate traces of the same session: q9487x6d; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 308.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 308.8 29.0 74.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kmuyvsxr",
   "source_url": "https://omlx.ai/benchmarks/performance/kmuyvsxr",
   "tg_tokens": null,
   "tps": 29.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-c7gpxlur-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 72.5 GB at this context; duplicate traces of the same session: 8t11ej8p; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 423.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 423.1 32.9 72.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c7gpxlur",
   "source_url": "https://omlx.ai/benchmarks/performance/c7gpxlur",
   "tg_tokens": null,
   "tps": 32.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-c7gpxlur-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace; peak memory 74.8 GB at this context; duplicate traces of the same session: 8t11ej8p; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 417.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 417.4 25.5 74.8 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/c7gpxlur",
   "source_url": "https://omlx.ai/benchmarks/performance/c7gpxlur",
   "tg_tokens": null,
   "tps": 25.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-dmy9kd4r-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; fp16; oMLX community benchmark trace, single run at 4k context; peak memory 74.8 GB; TTFT 13097 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 312.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-fp16-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k tokens 312.8 PP tok/s 27.6 TG tok/s 13097 TTFT (ms) 74.8 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dmy9kd4r",
   "source_url": "https://omlx.ai/benchmarks/performance/dmy9kd4r",
   "tg_tokens": null,
   "tps": 27.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-g8m810yc-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.8 GB at this context; duplicate traces of the same session: rpg4d7gp; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 205.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k 205.9 37.0 70.8 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g8m810yc",
   "source_url": "https://omlx.ai/benchmarks/performance/g8m810yc",
   "tg_tokens": null,
   "tps": 37.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-g8m810yc-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: rpg4d7gp; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 264.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 4k 264.8 29.6 71.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g8m810yc",
   "source_url": "https://omlx.ai/benchmarks/performance/g8m810yc",
   "tg_tokens": null,
   "tps": 29.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-2dki8x81-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 70.8 GB; TTFT 4749 ms; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 215.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 4bit; 1k tokens 215.6 PP tok/s 33.0 TG tok/s 4749 TTFT (ms) 70.8 Peak mem",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2dki8x81",
   "source_url": "https://omlx.ai/benchmarks/performance/2dki8x81",
   "tg_tokens": null,
   "tps": 33.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-u5q86imq-ctx65536",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 110.2 GB at this context; duplicate traces of the same session: 38b5jxzh, k2yekqtw; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1344.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 64k 1,344 73.1 110.2 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u5q86imq",
   "source_url": "https://omlx.ai/benchmarks/performance/u5q86imq",
   "tg_tokens": null,
   "tps": 73.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-u5q86imq-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 112.7 GB at this context; duplicate traces of the same session: 38b5jxzh, k2yekqtw; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1322.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 128k 1,322 54.3 112.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u5q86imq",
   "source_url": "https://omlx.ai/benchmarks/performance/u5q86imq",
   "tg_tokens": null,
   "tps": 54.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-u5q86imq-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 115.1 GB at this context; duplicate traces of the same session: 38b5jxzh, k2yekqtw; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1301.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 195k 1,301 61.9 115.1 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/u5q86imq",
   "source_url": "https://omlx.ai/benchmarks/performance/u5q86imq",
   "tg_tokens": null,
   "tps": 61.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-fpgwo98c-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.8 GB at this context; duplicate traces of the same session: 4uuml9em, 7r04heog, 88ipvg1k; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2074.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 1k 2,074 73.4 142.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fpgwo98c",
   "source_url": "https://omlx.ai/benchmarks/performance/fpgwo98c",
   "tg_tokens": null,
   "tps": 73.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-fpgwo98c-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: 4uuml9em, 7r04heog, 88ipvg1k; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3562.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 4k 3,562 124.2 143.2 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fpgwo98c",
   "source_url": "https://omlx.ai/benchmarks/performance/fpgwo98c",
   "tg_tokens": null,
   "tps": 124.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-fpgwo98c-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 144.5 GB at this context; duplicate traces of the same session: 4uuml9em, 7r04heog, 88ipvg1k; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3992.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 8k 3,992 117.4 144.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fpgwo98c",
   "source_url": "https://omlx.ai/benchmarks/performance/fpgwo98c",
   "tg_tokens": null,
   "tps": 117.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-fpgwo98c-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 145.1 GB at this context; duplicate traces of the same session: 4uuml9em, 7r04heog, 88ipvg1k; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4024.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 16k 4,024 106.7 145.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fpgwo98c",
   "source_url": "https://omlx.ai/benchmarks/performance/fpgwo98c",
   "tg_tokens": null,
   "tps": 106.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-1vbyomv9-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 142.8 GB at this context; duplicate traces of the same session: 9kfy8f1m, xj1uctzg; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2111.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 1k 2,111 133.7 142.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1vbyomv9",
   "source_url": "https://omlx.ai/benchmarks/performance/1vbyomv9",
   "tg_tokens": null,
   "tps": 133.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-1vbyomv9-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 143.2 GB at this context; duplicate traces of the same session: 9kfy8f1m, xj1uctzg; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3571.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 4k 3,571 114.3 143.2 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1vbyomv9",
   "source_url": "https://omlx.ai/benchmarks/performance/1vbyomv9",
   "tg_tokens": null,
   "tps": 114.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Ultra",
   "id": "omlx-1vbyomv9-ctx8192",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 144.5 GB at this context; duplicate traces of the same session: 9kfy8f1m, xj1uctzg; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3990.0,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-Flash-Next-oQ6e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 6bit; 8k 3,990 148.5 144.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1vbyomv9",
   "source_url": "https://omlx.ai/benchmarks/performance/1vbyomv9",
   "tg_tokens": null,
   "tps": 148.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-g97wzqxl-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 132.1 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4450.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 128k 4,450 73.3 132.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g97wzqxl",
   "source_url": "https://omlx.ai/benchmarks/performance/g97wzqxl",
   "tg_tokens": null,
   "tps": 73.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-g97wzqxl-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 134.8 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4476.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 195k 4,476 70.3 134.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/g97wzqxl",
   "source_url": "https://omlx.ai/benchmarks/performance/g97wzqxl",
   "tg_tokens": null,
   "tps": 70.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-6asin4rl-ctx16384",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "mlx-community build; oMLX community benchmark trace, single run at 16k context; peak memory 20.1 GB; TTFT 63601 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 257.6,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "mlx-community--MiniCPM-V-4.7-35B-A3B-mxfp4 on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k tokens 257.6 PP tok/s 18.0 TG tok/s 63601 TTFT (ms) 20.1 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6asin4rl",
   "source_url": "https://omlx.ai/benchmarks/performance/6asin4rl",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-jhqj28e1-ctx4096",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 292.0 GB at this context; duplicate traces of the same session: et7bmra2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 171.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 4k 171.8 19.0 292.0 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jhqj28e1",
   "source_url": "https://omlx.ai/benchmarks/performance/jhqj28e1",
   "tg_tokens": null,
   "tps": 19.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-jhqj28e1-ctx16384",
   "model": "DeepSeek-V4.1-Flash",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 292.5 GB at this context; duplicate traces of the same session: et7bmra2; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 175.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "DeepSeek-V4.1-Flash-1017-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 unknown; 16k 175.9 19.1 292.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jhqj28e1",
   "source_url": "https://omlx.ai/benchmarks/performance/jhqj28e1",
   "tg_tokens": null,
   "tps": 19.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-m9s44hyi-ctx1024",
   "model": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated",
   "notes": "oMLX community benchmark trace; peak memory 35.6 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1013.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated-mlx-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,013 72.0 35.6 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m9s44hyi",
   "source_url": "https://omlx.ai/benchmarks/performance/m9s44hyi",
   "tg_tokens": null,
   "tps": 72.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-m9s44hyi-ctx4096",
   "model": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated",
   "notes": "oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 952.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated-mlx-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 952.0 62.9 37.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/m9s44hyi",
   "source_url": "https://omlx.ai/benchmarks/performance/m9s44hyi",
   "tg_tokens": null,
   "tps": 62.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-kheqyxmg-ctx1024",
   "model": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated",
   "notes": "oMLX community benchmark trace; peak memory 35.6 GB at this context; duplicate traces of the same session: tohap1lo; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1008.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated-mlx-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,008 71.2 35.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kheqyxmg",
   "source_url": "https://omlx.ai/benchmarks/performance/kheqyxmg",
   "tg_tokens": null,
   "tps": 71.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-kheqyxmg-ctx4096",
   "model": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated",
   "notes": "oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: tohap1lo; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 955.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated-mlx-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 955.2 63.0 37.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kheqyxmg",
   "source_url": "https://omlx.ai/benchmarks/performance/kheqyxmg",
   "tg_tokens": null,
   "tps": 63.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-6rzfbqrl-ctx1024",
   "model": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated",
   "notes": "oMLX community benchmark trace; peak memory 35.6 GB at this context; duplicate traces of the same session: b9pm4gxl; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1013.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated-mlx-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,013 72.5 35.6 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6rzfbqrl",
   "source_url": "https://omlx.ai/benchmarks/performance/6rzfbqrl",
   "tg_tokens": null,
   "tps": 72.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-6rzfbqrl-ctx4096",
   "model": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated",
   "notes": "oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: b9pm4gxl; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1013.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated-mlx-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 1,013 64.9 37.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/6rzfbqrl",
   "source_url": "https://omlx.ai/benchmarks/performance/6rzfbqrl",
   "tg_tokens": null,
   "tps": 64.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-3tbsoip7-ctx1024",
   "model": "K2-Horizon-7B",
   "notes": "oMLX community benchmark trace; peak memory 9.7 GB at this context; duplicate traces of the same session: c3376u6r; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 367.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "K2-Horizon-7B-8bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 8bit; 1k 367.6 25.5 9.7 GB current",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3tbsoip7",
   "source_url": "https://omlx.ai/benchmarks/performance/3tbsoip7",
   "tg_tokens": null,
   "tps": 25.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-3tbsoip7-ctx16384",
   "model": "K2-Horizon-7B",
   "notes": "oMLX community benchmark trace; peak memory 11.8 GB at this context; duplicate traces of the same session: c3376u6r; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 278.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "K2-Horizon-7B-8bit on M4 Pro (16c); M4 Pro (16c) \u00b7 24 GB \u00b7 8bit; 16k 278.5 20.9 11.8 GB view",
   "ram_gb": 24,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3tbsoip7",
   "source_url": "https://omlx.ai/benchmarks/performance/3tbsoip7",
   "tg_tokens": null,
   "tps": 20.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Max",
   "id": "omlx-bn1d22sw-ctx8192",
   "model": "LFM2.5-24B-A2B",
   "notes": "MLX build; oMLX community benchmark trace, single run at 8k context; peak memory 13.7 GB; TTFT 48524 ms; duplicate traces of the same session: none; config=32c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 168.8,
   "provenance": "community",
   "quant": "4bit",
   "quote": "LFM2.5-24B-A2B-MLX-4bit on M1 Max (32c); M1 Max (32c) \u00b7 32 GB \u00b7 4bit; 8k tokens 168.8 PP tok/s 18.8 TG tok/s 48524 TTFT (ms) 13.7 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bn1d22sw",
   "source_url": "https://omlx.ai/benchmarks/performance/bn1d22sw",
   "tg_tokens": null,
   "tps": 18.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-78i23o7n-ctx8192",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace, single run at 8k context; peak memory 39.0 GB; TTFT 14038 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 583.6,
   "provenance": "community",
   "quant": "8bit",
   "quote": "MiniCPM-V-4.7-35B-A3B-8bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 8k tokens 583.6 PP tok/s 42.8 TG tok/s 14038 TTFT (ms) 39.0 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/78i23o7n",
   "source_url": "https://omlx.ai/benchmarks/performance/78i23o7n",
   "tg_tokens": null,
   "tps": 42.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-zx3vpok3-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: g7kwi9ln; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 904.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 904.2 100.0 21.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zx3vpok3",
   "source_url": "https://omlx.ai/benchmarks/performance/zx3vpok3",
   "tg_tokens": null,
   "tps": 100.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-zx3vpok3-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: g7kwi9ln; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 952.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 952.6 98.1 22.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zx3vpok3",
   "source_url": "https://omlx.ai/benchmarks/performance/zx3vpok3",
   "tg_tokens": null,
   "tps": 98.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-9a11fwim-ctx1024",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 5.7 GB at this context; duplicate traces of the same session: 92dlhpw4; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 227.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 1k 227.7 19.1 5.7 GB view",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9a11fwim",
   "source_url": "https://omlx.ai/benchmarks/performance/9a11fwim",
   "tg_tokens": null,
   "tps": 19.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4",
   "id": "omlx-9a11fwim-ctx4096",
   "model": "Ornith-1.5-9B",
   "notes": "oMLX community benchmark trace; peak memory 6.2 GB at this context; duplicate traces of the same session: 92dlhpw4; config=10c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 235.4,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Ornith-1.5-9B-MLX-4bit on M4 (10c); M4 (10c) \u00b7 16 GB \u00b7 4bit; 4k 235.4 18.5 6.2 GB current",
   "ram_gb": 16,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/9a11fwim",
   "source_url": "https://omlx.ai/benchmarks/performance/9a11fwim",
   "tg_tokens": null,
   "tps": 18.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-quypf5ne-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 961.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 961.4 83.2 37.4 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/quypf5ne",
   "source_url": "https://omlx.ai/benchmarks/performance/quypf5ne",
   "tg_tokens": null,
   "tps": 83.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-quypf5ne-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: none; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 912.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 912.7 68.6 39.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/quypf5ne",
   "source_url": "https://omlx.ai/benchmarks/performance/quypf5ne",
   "tg_tokens": null,
   "tps": 68.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-157yq1fh-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: mxwpn0kw; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 962.9,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 962.9 91.8 37.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/157yq1fh",
   "source_url": "https://omlx.ai/benchmarks/performance/157yq1fh",
   "tg_tokens": null,
   "tps": 91.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-157yq1fh-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: mxwpn0kw; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 946.7,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 946.7 75.6 39.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/157yq1fh",
   "source_url": "https://omlx.ai/benchmarks/performance/157yq1fh",
   "tg_tokens": null,
   "tps": 75.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ebmjtf4h-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace, single run at 1k context; peak memory 16.0 GB; TTFT 1095 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 935.6,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-MLX-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k tokens 935.6 PP tok/s 33.6 TG tok/s 1095 TTFT (ms) 16.0 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ebmjtf4h",
   "source_url": "https://omlx.ai/benchmarks/performance/ebmjtf4h",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-5ga11zl1-ctx1024",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 26.7 GB; TTFT 1225 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 836.3,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-Uncensored-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 1k tokens 836.3 PP tok/s 58.4 TG tok/s 1225 TTFT (ms) 26.7 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5ga11zl1",
   "source_url": "https://omlx.ai/benchmarks/performance/5ga11zl1",
   "tg_tokens": null,
   "tps": 58.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-446z0apu-ctx65536",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 64k context; peak memory 32.2 GB; TTFT 88227 ms; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 742.8,
   "provenance": "community",
   "quant": "oQ6e",
   "quote": "Qwen3.8-27B-Uncensored-oQ6e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 6bit; 64k tokens 742.8 PP tok/s 48.6 TG tok/s 88227 TTFT (ms) 32.2 Peak mem",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/446z0apu",
   "source_url": "https://omlx.ai/benchmarks/performance/446z0apu",
   "tg_tokens": null,
   "tps": 48.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-oo5re95w-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: 7r55wqpb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 915.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 915.9 60.4 17.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oo5re95w",
   "source_url": "https://omlx.ai/benchmarks/performance/oo5re95w",
   "tg_tokens": null,
   "tps": 60.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-oo5re95w-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: 7r55wqpb; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1000.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 1,000 63.6 18.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/oo5re95w",
   "source_url": "https://omlx.ai/benchmarks/performance/oo5re95w",
   "tg_tokens": null,
   "tps": 63.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8v6n0kul-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 17.6 GB at this context; duplicate traces of the same session: kn8xg1tj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 912.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 1k 912.3 69.2 17.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8v6n0kul",
   "source_url": "https://omlx.ai/benchmarks/performance/8v6n0kul",
   "tg_tokens": null,
   "tps": 69.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-8v6n0kul-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.7 GB at this context; duplicate traces of the same session: kn8xg1tj; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 995.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 4k 995.7 65.9 18.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8v6n0kul",
   "source_url": "https://omlx.ai/benchmarks/performance/8v6n0kul",
   "tg_tokens": null,
   "tps": 65.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-kyrzm66l-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 132.1 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4450.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 128k 4,450 73.3 132.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kyrzm66l",
   "source_url": "https://omlx.ai/benchmarks/performance/kyrzm66l",
   "tg_tokens": null,
   "tps": 73.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-kyrzm66l-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 134.8 GB at this context; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4476.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 195k 4,476 70.3 134.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kyrzm66l",
   "source_url": "https://omlx.ai/benchmarks/performance/kyrzm66l",
   "tg_tokens": null,
   "tps": 70.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-kosmo8ps-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 132.1 GB at this context; duplicate traces of the same session: 5kcj2og3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4234.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 128k 4,234 70.8 132.1 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kosmo8ps",
   "source_url": "https://omlx.ai/benchmarks/performance/kosmo8ps",
   "tg_tokens": null,
   "tps": 70.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-kosmo8ps-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 134.8 GB at this context; duplicate traces of the same session: 5kcj2og3; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4284.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 195k 4,284 65.3 134.8 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kosmo8ps",
   "source_url": "https://omlx.ai/benchmarks/performance/kosmo8ps",
   "tg_tokens": null,
   "tps": 65.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-4v935pjl-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 132.1 GB at this context; duplicate traces of the same session: vxvvckva; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4421.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 128k 4,421 71.5 132.1 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4v935pjl",
   "source_url": "https://omlx.ai/benchmarks/performance/4v935pjl",
   "tg_tokens": null,
   "tps": 71.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-4v935pjl-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 134.8 GB at this context; duplicate traces of the same session: vxvvckva; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4466.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 195k 4,466 69.6 134.8 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4v935pjl",
   "source_url": "https://omlx.ai/benchmarks/performance/4v935pjl",
   "tg_tokens": null,
   "tps": 69.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ig3kvnst-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: zsu1m006; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 904.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 904.9 90.8 21.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ig3kvnst",
   "source_url": "https://omlx.ai/benchmarks/performance/ig3kvnst",
   "tg_tokens": null,
   "tps": 90.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-ig3kvnst-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: zsu1m006; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 951.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 951.6 95.3 22.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ig3kvnst",
   "source_url": "https://omlx.ai/benchmarks/performance/ig3kvnst",
   "tg_tokens": null,
   "tps": 95.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-8kewwuig-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: s9rfme03; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 903.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 903.2 100.1 21.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8kewwuig",
   "source_url": "https://omlx.ai/benchmarks/performance/8kewwuig",
   "tg_tokens": null,
   "tps": 100.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-8kewwuig-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: s9rfme03; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 953.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 953.9 101.1 22.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8kewwuig",
   "source_url": "https://omlx.ai/benchmarks/performance/8kewwuig",
   "tg_tokens": null,
   "tps": 101.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-kdbveo8h-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 5e5xvras; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 880.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 880.6 98.9 21.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kdbveo8h",
   "source_url": "https://omlx.ai/benchmarks/performance/kdbveo8h",
   "tg_tokens": null,
   "tps": 98.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-kdbveo8h-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 5e5xvras; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 878.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 878.2 91.2 22.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/kdbveo8h",
   "source_url": "https://omlx.ai/benchmarks/performance/kdbveo8h",
   "tg_tokens": null,
   "tps": 91.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ffcb7uqs-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: m0vss1mc; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 902.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 902.2 88.7 21.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ffcb7uqs",
   "source_url": "https://omlx.ai/benchmarks/performance/ffcb7uqs",
   "tg_tokens": null,
   "tps": 88.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ffcb7uqs-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: m0vss1mc; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 913.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 913.2 92.5 22.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ffcb7uqs",
   "source_url": "https://omlx.ai/benchmarks/performance/ffcb7uqs",
   "tg_tokens": null,
   "tps": 92.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-uoroc5gc-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: whyd9zia; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 881.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 881.8 94.9 21.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uoroc5gc",
   "source_url": "https://omlx.ai/benchmarks/performance/uoroc5gc",
   "tg_tokens": null,
   "tps": 94.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-uoroc5gc-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: whyd9zia; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 879.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 879.7 96.2 22.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uoroc5gc",
   "source_url": "https://omlx.ai/benchmarks/performance/uoroc5gc",
   "tg_tokens": null,
   "tps": 96.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-uib37msm-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 7wdbvcbm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 904.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 904.7 100.5 21.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uib37msm",
   "source_url": "https://omlx.ai/benchmarks/performance/uib37msm",
   "tg_tokens": null,
   "tps": 100.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-uib37msm-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 7wdbvcbm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 933.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 933.8 91.6 22.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/uib37msm",
   "source_url": "https://omlx.ai/benchmarks/performance/uib37msm",
   "tg_tokens": null,
   "tps": 91.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-55wb23on-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 9yt20m4w; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 889.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 889.5 95.4 21.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/55wb23on",
   "source_url": "https://omlx.ai/benchmarks/performance/55wb23on",
   "tg_tokens": null,
   "tps": 95.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-55wb23on-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 9yt20m4w; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 893.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 893.7 83.8 22.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/55wb23on",
   "source_url": "https://omlx.ai/benchmarks/performance/55wb23on",
   "tg_tokens": null,
   "tps": 83.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-o8up743j-ctx16384",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "mlx-community build; oMLX community benchmark trace, single run at 16k context; peak memory 20.1 GB; TTFT 63481 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 258.1,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "mlx-community--MiniCPM-V-4.7-35B-A3B-mxfp4 on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k tokens 258.1 PP tok/s 17.8 TG tok/s 63481 TTFT (ms) 20.1 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o8up743j",
   "source_url": "https://omlx.ai/benchmarks/performance/o8up743j",
   "tg_tokens": null,
   "tps": 17.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4lt76rsf-ctx1024",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 65.7 GB at this context; duplicate traces of the same session: tu17uydz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 166.3,
   "provenance": "community",
   "quant": "2.4bit mixed",
   "quote": "DeepSeek-V4-Flash-0731-2.4bit-mixed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 1k 166.3 24.0 65.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4lt76rsf",
   "source_url": "https://omlx.ai/benchmarks/performance/4lt76rsf",
   "tg_tokens": null,
   "tps": 24.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-4lt76rsf-ctx4096",
   "model": "DeepSeek-V4-Flash-0731",
   "notes": "oMLX community benchmark trace; peak memory 66.5 GB at this context; duplicate traces of the same session: tu17uydz; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 449.7,
   "provenance": "community",
   "quant": "2.4bit mixed",
   "quote": "DeepSeek-V4-Flash-0731-2.4bit-mixed on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 2bit; 4k 449.7 21.1 66.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4lt76rsf",
   "source_url": "https://omlx.ai/benchmarks/performance/4lt76rsf",
   "tg_tokens": null,
   "tps": 21.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-jtn5s0ss-ctx1024",
   "model": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: g9zcsq54; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1016.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated-mlx-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,016 72.1 35.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jtn5s0ss",
   "source_url": "https://omlx.ai/benchmarks/performance/jtn5s0ss",
   "tg_tokens": null,
   "tps": 72.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-jtn5s0ss-ctx4096",
   "model": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated",
   "notes": "oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: g9zcsq54; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 970.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated-mlx-8bit on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 970.2 63.6 37.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/jtn5s0ss",
   "source_url": "https://omlx.ai/benchmarks/performance/jtn5s0ss",
   "tg_tokens": null,
   "tps": 63.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-3fgpdqyt-ctx1024",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 38.2 GB at this context; duplicate traces of the same session: vizjie9v; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 636.1,
   "provenance": "community",
   "quant": "8bit",
   "quote": "MiniCPM-V-4.7-35B-A3B-8bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 1k 636.1 45.6 38.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3fgpdqyt",
   "source_url": "https://omlx.ai/benchmarks/performance/3fgpdqyt",
   "tg_tokens": null,
   "tps": 45.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-3fgpdqyt-ctx131072",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 43.1 GB at this context; duplicate traces of the same session: vizjie9v; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 155.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "MiniCPM-V-4.7-35B-A3B-8bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 128k 155.2 24.6 43.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/3fgpdqyt",
   "source_url": "https://omlx.ai/benchmarks/performance/3fgpdqyt",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-heqry9kf-ctx1024",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 602.3,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "MiniCPM-V-4.7-35B-A3B-mxfp4 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 602.3 48.2 22.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/heqry9kf",
   "source_url": "https://omlx.ai/benchmarks/performance/heqry9kf",
   "tg_tokens": null,
   "tps": 48.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-heqry9kf-ctx131072",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 27.0 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 152.9,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "MiniCPM-V-4.7-35B-A3B-mxfp4 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 128k 152.9 24.6 27.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/heqry9kf",
   "source_url": "https://omlx.ai/benchmarks/performance/heqry9kf",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-rvjloqnm-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 7alxl6iy; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 902.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 902.5 95.7 21.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rvjloqnm",
   "source_url": "https://omlx.ai/benchmarks/performance/rvjloqnm",
   "tg_tokens": null,
   "tps": 95.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-rvjloqnm-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: 7alxl6iy; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 953.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 953.0 87.3 22.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/rvjloqnm",
   "source_url": "https://omlx.ai/benchmarks/performance/rvjloqnm",
   "tg_tokens": null,
   "tps": 87.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-8sociooa-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 7lysjfrt; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 904.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 904.2 109.5 21.4 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8sociooa",
   "source_url": "https://omlx.ai/benchmarks/performance/8sociooa",
   "tg_tokens": null,
   "tps": 109.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-8sociooa-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: 7lysjfrt; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 952.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 952.7 96.1 22.2 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/8sociooa",
   "source_url": "https://omlx.ai/benchmarks/performance/8sociooa",
   "tg_tokens": null,
   "tps": 96.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-2uiq27nj-ctx1024",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.4 GB at this context; duplicate traces of the same session: 7ll6gdyb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 901.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 901.8 97.4 21.4 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2uiq27nj",
   "source_url": "https://omlx.ai/benchmarks/performance/2uiq27nj",
   "tg_tokens": null,
   "tps": 97.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-2uiq27nj-ctx4096",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.2 GB at this context; duplicate traces of the same session: 7ll6gdyb; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 938.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 938.1 97.1 22.2 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/2uiq27nj",
   "source_url": "https://omlx.ai/benchmarks/performance/2uiq27nj",
   "tg_tokens": null,
   "tps": 97.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-onf6a2gf-ctx65536",
   "model": "Ornith-1.5-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace, single run at 64k context; peak memory 23.4 GB; TTFT 38583 ms; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1699.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.5-35B-A3B-oQ4e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 64k tokens 1,699 PP tok/s 45.1 TG tok/s 38583 TTFT (ms) 23.4 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/onf6a2gf",
   "source_url": "https://omlx.ai/benchmarks/performance/onf6a2gf",
   "tg_tokens": null,
   "tps": 45.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ukmh84mi-ctx1024",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 189.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 189.7 10.7 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ukmh84mi",
   "source_url": "https://omlx.ai/benchmarks/performance/ukmh84mi",
   "tg_tokens": null,
   "tps": 10.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-ukmh84mi-ctx4096",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 240.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 240.3 10.7 17.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ukmh84mi",
   "source_url": "https://omlx.ai/benchmarks/performance/ukmh84mi",
   "tg_tokens": null,
   "tps": 10.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-7zg2wvnb-ctx1024",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 172.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-27B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 172.2 6.1 28.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7zg2wvnb",
   "source_url": "https://omlx.ai/benchmarks/performance/7zg2wvnb",
   "tg_tokens": null,
   "tps": 6.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-7zg2wvnb-ctx4096",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 212.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-27B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 212.0 6.1 29.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7zg2wvnb",
   "source_url": "https://omlx.ai/benchmarks/performance/7zg2wvnb",
   "tg_tokens": null,
   "tps": 6.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-7g4lcxwx-ctx1024",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: yj7r9umk; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1011.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-oQ8e on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,011 70.2 35.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7g4lcxwx",
   "source_url": "https://omlx.ai/benchmarks/performance/7g4lcxwx",
   "tg_tokens": null,
   "tps": 70.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-7g4lcxwx-ctx4096",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: yj7r9umk; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 971.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-oQ8e on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 971.6 62.5 37.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7g4lcxwx",
   "source_url": "https://omlx.ai/benchmarks/performance/7g4lcxwx",
   "tg_tokens": null,
   "tps": 62.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-vnratgu5-ctx1024",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: g3ffpwed; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1011.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-oQ8e on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,011 70.1 35.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vnratgu5",
   "source_url": "https://omlx.ai/benchmarks/performance/vnratgu5",
   "tg_tokens": null,
   "tps": 70.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-vnratgu5-ctx4096",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: g3ffpwed; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 986.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-oQ8e on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 986.6 63.8 37.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vnratgu5",
   "source_url": "https://omlx.ai/benchmarks/performance/vnratgu5",
   "tg_tokens": null,
   "tps": 63.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-n5fhn71v-ctx1024",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "oMLX community benchmark trace; peak memory 35.5 GB at this context; duplicate traces of the same session: ijnbpuct; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1016.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-oQ8e on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 1,016 70.8 35.5 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n5fhn71v",
   "source_url": "https://omlx.ai/benchmarks/performance/n5fhn71v",
   "tg_tokens": null,
   "tps": 70.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-n5fhn71v-ctx4096",
   "model": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled",
   "notes": "oMLX community benchmark trace; peak memory 37.3 GB at this context; duplicate traces of the same session: ijnbpuct; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1028.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-oQ8e on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 1,028 63.3 37.3 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n5fhn71v",
   "source_url": "https://omlx.ai/benchmarks/performance/n5fhn71v",
   "tg_tokens": null,
   "tps": 63.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M2 Max",
   "id": "omlx-bhbe3spr-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: tm7cch75; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 962.4,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 1k 962.4 97.6 37.4 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhbe3spr",
   "source_url": "https://omlx.ai/benchmarks/performance/bhbe3spr",
   "tg_tokens": null,
   "tps": 97.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M2 Max",
   "id": "omlx-bhbe3spr-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 39.3 GB at this context; duplicate traces of the same session: tm7cch75; config=38c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 938.3,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M2 Max (38c); M2 Max (38c) \u00b7 96 GB \u00b7 8bit; 4k 938.3 70.0 39.3 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/bhbe3spr",
   "source_url": "https://omlx.ai/benchmarks/performance/bhbe3spr",
   "tg_tokens": null,
   "tps": 70.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-xmomeydp-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 35.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 867.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 867.8 61.7 35.1 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xmomeydp",
   "source_url": "https://omlx.ai/benchmarks/performance/xmomeydp",
   "tg_tokens": null,
   "tps": 61.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xmomeydp-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 36.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 614.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 614.5 37.6 36.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xmomeydp",
   "source_url": "https://omlx.ai/benchmarks/performance/xmomeydp",
   "tg_tokens": null,
   "tps": 37.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pidb56af-ctx8192",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.1 GB at this context; duplicate traces of the same session: 35u532ba, ipjyg2r4, reckb573, rzzcxbel; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 917.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 8k 917.1 40.1 19.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pidb56af",
   "source_url": "https://omlx.ai/benchmarks/performance/pidb56af",
   "tg_tokens": null,
   "tps": 40.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-pidb56af-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 19.6 GB at this context; duplicate traces of the same session: 35u532ba, ipjyg2r4, reckb573, rzzcxbel; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 964.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 16k 964.9 59.7 19.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pidb56af",
   "source_url": "https://omlx.ai/benchmarks/performance/pidb56af",
   "tg_tokens": null,
   "tps": 59.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-pidb56af-ctx32768",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: 35u532ba, ipjyg2r4, reckb573, rzzcxbel; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 900.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 32k 900.4 130.9 20.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pidb56af",
   "source_url": "https://omlx.ai/benchmarks/performance/pidb56af",
   "tg_tokens": null,
   "tps": 130.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 65536,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-pidb56af-ctx65536",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 23.1 GB at this context; duplicate traces of the same session: 35u532ba, ipjyg2r4, reckb573, rzzcxbel; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 788.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 64k 788.6 55.9 23.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pidb56af",
   "source_url": "https://omlx.ai/benchmarks/performance/pidb56af",
   "tg_tokens": null,
   "tps": 55.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-pidb56af-ctx131072",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.9 GB at this context; duplicate traces of the same session: 35u532ba, ipjyg2r4, reckb573, rzzcxbel; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 625.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-oQ4e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 64 GB \u00b7 4bit; 128k 625.5 47.5 27.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/pidb56af",
   "source_url": "https://omlx.ai/benchmarks/performance/pidb56af",
   "tg_tokens": null,
   "tps": 47.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-l005b6eh-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 70.7 GB at this context; duplicate traces of the same session: lunv6s9o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 467.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 467.2 54.1 70.7 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l005b6eh",
   "source_url": "https://omlx.ai/benchmarks/performance/l005b6eh",
   "tg_tokens": null,
   "tps": 54.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Max",
   "id": "omlx-l005b6eh-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 71.3 GB at this context; duplicate traces of the same session: lunv6s9o; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 612.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M4 Max (40c); M4 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 612.9 51.8 71.3 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l005b6eh",
   "source_url": "https://omlx.ai/benchmarks/performance/l005b6eh",
   "tg_tokens": null,
   "tps": 51.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-izs7ldi2-ctx131072",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 131.7 GB at this context; duplicate traces of the same session: 2db8qsgu; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4251.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 128k 4,251 48.7 131.7 GB current",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/izs7ldi2",
   "source_url": "https://omlx.ai/benchmarks/performance/izs7ldi2",
   "tg_tokens": null,
   "tps": 48.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-izs7ldi2-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 134.5 GB at this context; duplicate traces of the same session: 2db8qsgu; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4286.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 195k 4,286 62.9 134.5 GB view",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/izs7ldi2",
   "source_url": "https://omlx.ai/benchmarks/performance/izs7ldi2",
   "tg_tokens": null,
   "tps": 62.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 199680,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Ultra",
   "id": "omlx-psgakviy-ctx199680",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 195k context; peak memory 134.8 GB; TTFT 45926 ms; duplicate traces of the same session: none; config=64c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 4355.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-Flash-Next-oQ8e-mtp on M5 Ultra (64c); M5 Ultra (64c) \u00b7 256 GB \u00b7 8bit; 195k tokens 4,355 PP tok/s 71.0 TG tok/s 45926 TTFT (ms) 134.8 Peak mem",
   "ram_gb": 256,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/psgakviy",
   "source_url": "https://omlx.ai/benchmarks/performance/psgakviy",
   "tg_tokens": null,
   "tps": 71.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-l2gibjc9-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: hc1dkxzt; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 904.6,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 904.6 97.0 21.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l2gibjc9",
   "source_url": "https://omlx.ai/benchmarks/performance/l2gibjc9",
   "tg_tokens": null,
   "tps": 97.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-l2gibjc9-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: hc1dkxzt; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 925.9,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 925.9 92.5 22.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/l2gibjc9",
   "source_url": "https://omlx.ai/benchmarks/performance/l2gibjc9",
   "tg_tokens": null,
   "tps": 92.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-lh9fjgc7-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: zb6bwca5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 899.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 899.7 97.1 21.5 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lh9fjgc7",
   "source_url": "https://omlx.ai/benchmarks/performance/lh9fjgc7",
   "tg_tokens": null,
   "tps": 97.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-lh9fjgc7-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: zb6bwca5; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 913.2,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 913.2 101.1 22.3 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/lh9fjgc7",
   "source_url": "https://omlx.ai/benchmarks/performance/lh9fjgc7",
   "tg_tokens": null,
   "tps": 101.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-o3wb659v-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 21.5 GB at this context; duplicate traces of the same session: 4jo6eesd; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 892.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 892.7 95.9 21.5 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o3wb659v",
   "source_url": "https://omlx.ai/benchmarks/performance/o3wb659v",
   "tg_tokens": null,
   "tps": 95.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M4 Pro",
   "id": "omlx-o3wb659v-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 22.3 GB at this context; duplicate traces of the same session: 4jo6eesd; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 902.1,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 902.1 99.5 22.3 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/o3wb659v",
   "source_url": "https://omlx.ai/benchmarks/performance/o3wb659v",
   "tg_tokens": null,
   "tps": 99.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ig907vs5-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: af7xtavt; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 930.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 930.7 78.0 20.1 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ig907vs5",
   "source_url": "https://omlx.ai/benchmarks/performance/ig907vs5",
   "tg_tokens": null,
   "tps": 78.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-ig907vs5-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: af7xtavt; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 977.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 977.8 74.1 20.9 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ig907vs5",
   "source_url": "https://omlx.ai/benchmarks/performance/ig907vs5",
   "tg_tokens": null,
   "tps": 74.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-qz2e5ar7-ctx1024",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: vn0stolf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 937.3,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 1k 937.3 80.0 20.1 GB current",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qz2e5ar7",
   "source_url": "https://omlx.ai/benchmarks/performance/qz2e5ar7",
   "tg_tokens": null,
   "tps": 80.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M4 Pro",
   "id": "omlx-qz2e5ar7-ctx4096",
   "model": "Tiel-Coder-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 20.9 GB at this context; duplicate traces of the same session: vn0stolf; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 971.7,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Tiel-Coder-35B-A3B-MLX-oQ4e-MTP on M4 Pro (20c); M4 Pro (20c) \u00b7 48 GB \u00b7 4bit; 4k 971.7 77.2 20.9 GB view",
   "ram_gb": 48,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qz2e5ar7",
   "source_url": "https://omlx.ai/benchmarks/performance/qz2e5ar7",
   "tg_tokens": null,
   "tps": 77.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-wnc4axs9-ctx1024",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 38.2 GB at this context; duplicate traces of the same session: y2ehq4gf; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 634.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "MiniCPM-V-4.7-35B-A3B-8bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 1k 634.2 45.6 38.2 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wnc4axs9",
   "source_url": "https://omlx.ai/benchmarks/performance/wnc4axs9",
   "tg_tokens": null,
   "tps": 45.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-wnc4axs9-ctx8192",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 39.0 GB at this context; duplicate traces of the same session: y2ehq4gf; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 600.5,
   "provenance": "community",
   "quant": "8bit",
   "quote": "MiniCPM-V-4.7-35B-A3B-8bit on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 8k 600.5 42.8 39.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/wnc4axs9",
   "source_url": "https://omlx.ai/benchmarks/performance/wnc4axs9",
   "tg_tokens": null,
   "tps": 42.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-xjqcdi6p-ctx1024",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: 5ndc7l22; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 579.9,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "MiniCPM-V-4.7-35B-A3B-mxfp4 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 579.9 48.1 22.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xjqcdi6p",
   "source_url": "https://omlx.ai/benchmarks/performance/xjqcdi6p",
   "tg_tokens": null,
   "tps": 48.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-xjqcdi6p-ctx8192",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.9 GB at this context; duplicate traces of the same session: 5ndc7l22; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 567.9,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "MiniCPM-V-4.7-35B-A3B-mxfp4 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 8k 567.9 45.0 22.9 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xjqcdi6p",
   "source_url": "https://omlx.ai/benchmarks/performance/xjqcdi6p",
   "tg_tokens": null,
   "tps": 45.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-tn6wj6ni-ctx1024",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 22.1 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 602.3,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "MiniCPM-V-4.7-35B-A3B-mxfp4 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 1k 602.3 48.2 22.1 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tn6wj6ni",
   "source_url": "https://omlx.ai/benchmarks/performance/tn6wj6ni",
   "tg_tokens": null,
   "tps": 48.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 131072,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-tn6wj6ni-ctx131072",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "oMLX community benchmark trace; peak memory 27.0 GB at this context; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 152.9,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "MiniCPM-V-4.7-35B-A3B-mxfp4 on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 4bit; 128k 152.9 24.6 27.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tn6wj6ni",
   "source_url": "https://omlx.ai/benchmarks/performance/tn6wj6ni",
   "tg_tokens": null,
   "tps": 24.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M1 Max",
   "id": "omlx-vb83pabf-ctx1024",
   "model": "Nex-N2.5-mini",
   "notes": "fp16 MTP build; oMLX community benchmark trace, single run at 1k context; peak memory 37.8 GB; TTFT 1459 ms; duplicate traces of the same session: none; config=24c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 701.6,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Nex-N2.5-mini-oQ8e-fp16-mtp on M1 Max (24c); M1 Max (24c) \u00b7 64 GB \u00b7 8bit; 1k tokens 701.6 PP tok/s 66.4 TG tok/s 1459 TTFT (ms) 37.8 Peak mem",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vb83pabf",
   "source_url": "https://omlx.ai/benchmarks/performance/vb83pabf",
   "tg_tokens": null,
   "tps": 66.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qvud7dqj-ctx1024",
   "model": "Ornith-1.0-35B",
   "notes": "oMLX community benchmark trace; peak memory 20.1 GB at this context; duplicate traces of the same session: zrd19y38; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2715.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.0-35B-oQ4e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 2,715 92.3 20.1 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qvud7dqj",
   "source_url": "https://omlx.ai/benchmarks/performance/qvud7dqj",
   "tg_tokens": null,
   "tps": 92.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-qvud7dqj-ctx4096",
   "model": "Ornith-1.0-35B",
   "notes": "oMLX community benchmark trace; peak memory 20.7 GB at this context; duplicate traces of the same session: zrd19y38; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3396.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Ornith-1.0-35B-oQ4e on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 3,396 90.8 20.7 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/qvud7dqj",
   "source_url": "https://omlx.ai/benchmarks/performance/qvud7dqj",
   "tg_tokens": null,
   "tps": 90.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-xtzaib7m-ctx1024",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 189.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 189.7 10.7 16.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xtzaib7m",
   "source_url": "https://omlx.ai/benchmarks/performance/xtzaib7m",
   "tg_tokens": null,
   "tps": 10.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-xtzaib7m-ctx4096",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 240.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.6-27B-MLX-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 240.3 10.7 17.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xtzaib7m",
   "source_url": "https://omlx.ai/benchmarks/performance/xtzaib7m",
   "tg_tokens": null,
   "tps": 10.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-vt2bp42y-ctx1024",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 172.2,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-27B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 172.2 6.1 28.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vt2bp42y",
   "source_url": "https://omlx.ai/benchmarks/performance/vt2bp42y",
   "tg_tokens": null,
   "tps": 6.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-vt2bp42y-ctx4096",
   "model": "Qwen3.6-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 212.0,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.6-27B-MLX-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 212.0 6.1 29.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/vt2bp42y",
   "source_url": "https://omlx.ai/benchmarks/performance/vt2bp42y",
   "tg_tokens": null,
   "tps": 6.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-f3vmos1t-ctx1024",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 37.4 GB at this context; duplicate traces of the same session: 9f7511gm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2261.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 2,261 106.9 37.4 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f3vmos1t",
   "source_url": "https://omlx.ai/benchmarks/performance/f3vmos1t",
   "tg_tokens": null,
   "tps": 106.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-f3vmos1t-ctx4096",
   "model": "Qwen3.6-35B-A3B",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 38.0 GB at this context; duplicate traces of the same session: 9f7511gm; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3081.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.6-35B-A3B-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 3,081 88.7 38.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f3vmos1t",
   "source_url": "https://omlx.ai/benchmarks/performance/f3vmos1t",
   "tg_tokens": null,
   "tps": 88.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n42fyydt-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 21.3 GB at this context; duplicate traces of the same session: x2gow1ai; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 886.9,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 4k 886.9 42.4 21.3 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n42fyydt",
   "source_url": "https://omlx.ai/benchmarks/performance/n42fyydt",
   "tg_tokens": null,
   "tps": 42.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-n42fyydt-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 22.5 GB at this context; duplicate traces of the same session: x2gow1ai; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 584.1,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 16k 584.1 54.4 22.5 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n42fyydt",
   "source_url": "https://omlx.ai/benchmarks/performance/n42fyydt",
   "tg_tokens": null,
   "tps": 54.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-tjlwxb0c-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 16.0 GB at this context; duplicate traces of the same session: s83q6qfw; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 491.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 491.3 18.1 16.0 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tjlwxb0c",
   "source_url": "https://omlx.ai/benchmarks/performance/tjlwxb0c",
   "tg_tokens": null,
   "tps": 18.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-tjlwxb0c-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 17.0 GB at this context; duplicate traces of the same session: s83q6qfw; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 518.7,
   "provenance": "community",
   "quant": "4bit",
   "quote": "Qwen3.8-27B-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 518.7 17.6 17.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/tjlwxb0c",
   "source_url": "https://omlx.ai/benchmarks/performance/tjlwxb0c",
   "tg_tokens": null,
   "tps": 17.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1v3p3l1f-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 28.5 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 432.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 432.4 10.0 28.5 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1v3p3l1f",
   "source_url": "https://omlx.ai/benchmarks/performance/1v3p3l1f",
   "tg_tokens": null,
   "tps": 10.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Pro",
   "id": "omlx-1v3p3l1f-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.5 GB at this context; duplicate traces of the same session: none; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 449.3,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-8bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 449.3 10.0 29.5 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/1v3p3l1f",
   "source_url": "https://omlx.ai/benchmarks/performance/1v3p3l1f",
   "tg_tokens": null,
   "tps": 10.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xoiaot0z-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 29.8 GB at this context; duplicate traces of the same session: ezv6dtxn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 843.4,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 843.4 17.4 29.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xoiaot0z",
   "source_url": "https://omlx.ai/benchmarks/performance/xoiaot0z",
   "tg_tokens": null,
   "tps": 17.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-xoiaot0z-ctx16384",
   "model": "Qwen3.8-27B",
   "notes": "oMLX community benchmark trace; peak memory 31.7 GB at this context; duplicate traces of the same session: ezv6dtxn; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 652.7,
   "provenance": "community",
   "quant": "8bit",
   "quote": "Qwen3.8-27B-MLX-8bit on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 652.7 15.5 31.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/xoiaot0z",
   "source_url": "https://omlx.ai/benchmarks/performance/xoiaot0z",
   "tg_tokens": null,
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-q7jk0w3t-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: iarc66td; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 426.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 426.8 86.7 18.5 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q7jk0w3t",
   "source_url": "https://omlx.ai/benchmarks/performance/q7jk0w3t",
   "tg_tokens": null,
   "tps": 86.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-q7jk0w3t-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: iarc66td; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 461.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 461.5 85.5 21.9 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/q7jk0w3t",
   "source_url": "https://omlx.ai/benchmarks/performance/q7jk0w3t",
   "tg_tokens": null,
   "tps": 85.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-4qh45pga-ctx1024",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 18.5 GB at this context; duplicate traces of the same session: gyzcn81d; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 425.8,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 425.8 86.6 18.5 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4qh45pga",
   "source_url": "https://omlx.ai/benchmarks/performance/4qh45pga",
   "tg_tokens": null,
   "tps": 86.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-4qh45pga-ctx4096",
   "model": "Qwen3.8-27B",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 21.9 GB at this context; duplicate traces of the same session: gyzcn81d; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 461.5,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-27B-MLX-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 461.5 83.5 21.9 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/4qh45pga",
   "source_url": "https://omlx.ai/benchmarks/performance/4qh45pga",
   "tg_tokens": null,
   "tps": 83.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-r0x8f4s9-ctx1024",
   "model": "Qwen3.8-27B-TTURBO-Cold-Fusion-709-L-Uncensored",
   "notes": "mlx 1M context build; oMLX community benchmark trace; peak memory 33.9 GB at this context; duplicate traces of the same session: gu1ea1g9, h9w3mio6, lyidqa57; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 807.1,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TTURBO-Cold-Fusion-709-L-Uncensored-mlx-oQ8e-1M on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 1k 807.1 39.8 33.9 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r0x8f4s9",
   "source_url": "https://omlx.ai/benchmarks/performance/r0x8f4s9",
   "tg_tokens": null,
   "tps": 39.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-r0x8f4s9-ctx4096",
   "model": "Qwen3.8-27B-TTURBO-Cold-Fusion-709-L-Uncensored",
   "notes": "mlx 1M context build; oMLX community benchmark trace; peak memory 34.9 GB at this context; duplicate traces of the same session: gu1ea1g9, h9w3mio6, lyidqa57; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 920.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TTURBO-Cold-Fusion-709-L-Uncensored-mlx-oQ8e-1M on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 920.2 49.5 34.9 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r0x8f4s9",
   "source_url": "https://omlx.ai/benchmarks/performance/r0x8f4s9",
   "tg_tokens": null,
   "tps": 49.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-r0x8f4s9-ctx8192",
   "model": "Qwen3.8-27B-TTURBO-Cold-Fusion-709-L-Uncensored",
   "notes": "mlx 1M context build; oMLX community benchmark trace; peak memory 35.2 GB at this context; duplicate traces of the same session: gu1ea1g9, h9w3mio6, lyidqa57; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 856.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TTURBO-Cold-Fusion-709-L-Uncensored-mlx-oQ8e-1M on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 8k 856.0 54.5 35.2 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r0x8f4s9",
   "source_url": "https://omlx.ai/benchmarks/performance/r0x8f4s9",
   "tg_tokens": null,
   "tps": 54.5,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-r0x8f4s9-ctx16384",
   "model": "Qwen3.8-27B-TTURBO-Cold-Fusion-709-L-Uncensored",
   "notes": "mlx 1M context build; oMLX community benchmark trace; peak memory 35.8 GB at this context; duplicate traces of the same session: gu1ea1g9, h9w3mio6, lyidqa57; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 803.2,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-TTURBO-Cold-Fusion-709-L-Uncensored-mlx-oQ8e-1M on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 16k 803.2 52.2 35.8 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/r0x8f4s9",
   "source_url": "https://omlx.ai/benchmarks/performance/r0x8f4s9",
   "tg_tokens": null,
   "tps": 52.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Max",
   "id": "omlx-f729t5ry-ctx4096",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 35.1 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 867.8,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 4k 867.8 61.7 35.1 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f729t5ry",
   "source_url": "https://omlx.ai/benchmarks/performance/f729t5ry",
   "tg_tokens": null,
   "tps": 61.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-f729t5ry-ctx32768",
   "model": "Qwen3.8-27B-Uncensored",
   "notes": "MTP build; oMLX community benchmark trace; peak memory 36.8 GB at this context; duplicate traces of the same session: none; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 614.5,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "Qwen3.8-27B-Uncensored-oQ8e-mtp on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 8bit; 32k 614.5 37.6 36.8 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/f729t5ry",
   "source_url": "https://omlx.ai/benchmarks/performance/f729t5ry",
   "tg_tokens": null,
   "tps": 37.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-fefkdlqt-ctx16384",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 63.0 GB at this context; duplicate traces of the same session: 7ho3phvv; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 903.5,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ3-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 3bit; 16k 903.5 41.2 63.0 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fefkdlqt",
   "source_url": "https://omlx.ai/benchmarks/performance/fefkdlqt",
   "tg_tokens": null,
   "tps": 41.2,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 32768,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-fefkdlqt-ctx32768",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 63.5 GB at this context; duplicate traces of the same session: 7ho3phvv; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1006.0,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ3-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 3bit; 32k 1,006 33.6 63.5 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/fefkdlqt",
   "source_url": "https://omlx.ai/benchmarks/performance/fefkdlqt",
   "tg_tokens": null,
   "tps": 33.6,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-zyqsfjky-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 62.2 GB at this context; duplicate traces of the same session: bsgf38yu; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 424.5,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ3-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 3bit; 1k 424.5 43.1 62.2 GB current",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zyqsfjky",
   "source_url": "https://omlx.ai/benchmarks/performance/zyqsfjky",
   "tg_tokens": null,
   "tps": 43.1,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M3 Ultra",
   "id": "omlx-zyqsfjky-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 62.7 GB at this context; duplicate traces of the same session: bsgf38yu; config=60c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 615.1,
   "provenance": "community",
   "quant": "oQ3",
   "quote": "Qwen3.8-Flash-Next-MLX-oQ3-MTP on M3 Ultra (60c); M3 Ultra (60c) \u00b7 96 GB \u00b7 3bit; 4k 615.1 36.8 62.7 GB view",
   "ram_gb": 96,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zyqsfjky",
   "source_url": "https://omlx.ai/benchmarks/performance/zyqsfjky",
   "tg_tokens": null,
   "tps": 36.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ix93wl7x-ctx1024",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 73.6 GB at this context; duplicate traces of the same session: wisu0mec; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 998.4,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 1k 998.4 62.8 73.6 GB view",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ix93wl7x",
   "source_url": "https://omlx.ai/benchmarks/performance/ix93wl7x",
   "tg_tokens": null,
   "tps": 62.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 8192,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M5 Max",
   "id": "omlx-ix93wl7x-ctx8192",
   "model": "Qwen3.8-Flash-Next-Uncensored",
   "notes": "MTP 100K build; oMLX community benchmark trace; peak memory 74.7 GB at this context; duplicate traces of the same session: wisu0mec; config=40c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1809.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP on M5 Max (40c); M5 Max (40c) \u00b7 128 GB \u00b7 4bit; 8k 1,809 58.0 74.7 GB current",
   "ram_gb": 128,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ix93wl7x",
   "source_url": "https://omlx.ai/benchmarks/performance/ix93wl7x",
   "tg_tokens": null,
   "tps": 58.0,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-nl04vm2a-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace, single run at 1k context; peak memory 99.1 GB; TTFT 893 ms; duplicate traces of the same session: none; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1147.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k tokens 1,147 PP tok/s 86.7 TG tok/s 893 TTFT (ms) 99.1 Peak mem",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/nl04vm2a",
   "source_url": "https://omlx.ai/benchmarks/performance/nl04vm2a",
   "tg_tokens": null,
   "tps": 86.7,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-n80337qe-ctx1024",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.1 GB at this context; duplicate traces of the same session: 9x273aye; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1149.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 1k 1,149 87.8 99.1 GB current",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n80337qe",
   "source_url": "https://omlx.ai/benchmarks/performance/n80337qe",
   "tg_tokens": null,
   "tps": 87.8,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M3 Ultra",
   "id": "omlx-n80337qe-ctx4096",
   "model": "Qwen3.8-Flash-Next",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 99.7 GB at this context; duplicate traces of the same session: 9x273aye; config=80c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 1431.0,
   "provenance": "community",
   "quant": "oQ4e",
   "quote": "Qwen3.8-Flash-Next-oQ4e-mtp on M3 Ultra (80c); M3 Ultra (80c) \u00b7 512 GB \u00b7 4bit; 4k 1,431 80.3 99.7 GB view",
   "ram_gb": 512,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/n80337qe",
   "source_url": "https://omlx.ai/benchmarks/performance/n80337qe",
   "tg_tokens": null,
   "tps": 80.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-zrmxgfke-ctx1024",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 27.6 GB at this context; duplicate traces of the same session: w27dczdx; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2069.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 1k 2,069 75.9 27.6 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zrmxgfke",
   "source_url": "https://omlx.ai/benchmarks/performance/zrmxgfke",
   "tg_tokens": null,
   "tps": 75.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-zrmxgfke-ctx4096",
   "model": "gemma-4-26B-A4B-it",
   "notes": "MTP speculative decoding enabled; oMLX community benchmark trace; peak memory 28.0 GB at this context; duplicate traces of the same session: w27dczdx; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 2260.0,
   "provenance": "community",
   "quant": "oQ8e",
   "quote": "gemma-4-26B-A4B-it-oQ8e-mtp on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 8bit; 4k 2,260 52.9 28.0 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/zrmxgfke",
   "source_url": "https://omlx.ai/benchmarks/performance/zrmxgfke",
   "tg_tokens": null,
   "tps": 52.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ythwb3zo-ctx1024",
   "model": "Gemma-4-31B",
   "notes": "oMLX community benchmark trace; peak memory 17.8 GB at this context; duplicate traces of the same session: x629v023; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 380.5,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 1k 380.5 14.3 17.8 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ythwb3zo",
   "source_url": "https://omlx.ai/benchmarks/performance/ythwb3zo",
   "tg_tokens": null,
   "tps": 14.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-ythwb3zo-ctx4096",
   "model": "Gemma-4-31B",
   "notes": "oMLX community benchmark trace; peak memory 18.8 GB at this context; duplicate traces of the same session: x629v023; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 389.3,
   "provenance": "community",
   "quant": "4bit",
   "quote": "gemma-4-31b-it-4bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 4bit; 4k 389.3 13.4 18.8 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/ythwb3zo",
   "source_url": "https://omlx.ai/benchmarks/performance/ythwb3zo",
   "tg_tokens": null,
   "tps": 13.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 1024,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-dj9r78zc-ctx1024",
   "model": "gemma-4-E4B-it",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 6.6 GB at this context; duplicate traces of the same session: igjqewxi; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3325.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "gemma-4-E4B-it-MLX-5bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 5bit; 1k 3,325 53.3 6.6 GB current",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dj9r78zc",
   "source_url": "https://omlx.ai/benchmarks/performance/dj9r78zc",
   "tg_tokens": null,
   "tps": 53.3,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 4096,
   "date": "2026-10-18",
   "flags": [],
   "hardware": "M5 Pro",
   "id": "omlx-dj9r78zc-ctx4096",
   "model": "gemma-4-E4B-it",
   "notes": "MLX build; oMLX community benchmark trace; peak memory 6.7 GB at this context; duplicate traces of the same session: igjqewxi; config=20c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 3848.0,
   "provenance": "community",
   "quant": "5bit",
   "quote": "gemma-4-E4B-it-MLX-5bit on M5 Pro (20c); M5 Pro (20c) \u00b7 64 GB \u00b7 5bit; 4k 3,848 52.4 6.7 GB view",
   "ram_gb": 64,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/dj9r78zc",
   "source_url": "https://omlx.ai/benchmarks/performance/dj9r78zc",
   "tg_tokens": null,
   "tps": 52.4,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-7ot552wp-ctx16384",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "mlx-community build; oMLX community benchmark trace, single run at 16k context; peak memory 20.1 GB; TTFT 64240 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 255.0,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "mlx-community--MiniCPM-V-4.7-35B-A3B-mxfp4 on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k tokens 255.0 PP tok/s 17.9 TG tok/s 64240 TTFT (ms) 20.1 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/7ot552wp",
   "source_url": "https://omlx.ai/benchmarks/performance/7ot552wp",
   "tg_tokens": null,
   "tps": 17.9,
   "ttft_s": null
  },
  {
   "backend": "oMLX",
   "batch": 1,
   "ctx": 16384,
   "date": "2026-10-18",
   "flags": [
    "contradiction"
   ],
   "hardware": "M1 Pro",
   "id": "omlx-5mxqlvmd-ctx16384",
   "model": "MiniCPM-V-4.7-35B-A3B",
   "notes": "mlx-community build; oMLX community benchmark trace, single run at 16k context; peak memory 20.1 GB; TTFT 63311 ms; duplicate traces of the same session: none; config=16c",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": 258.8,
   "provenance": "community",
   "quant": "mxfp4",
   "quote": "mlx-community--MiniCPM-V-4.7-35B-A3B-mxfp4 on M1 Pro (16c); M1 Pro (16c) \u00b7 32 GB \u00b7 4bit; 16k tokens 258.8 PP tok/s 18.0 TG tok/s 63311 TTFT (ms) 20.1 Peak mem",
   "ram_gb": 32,
   "retrieved": "2026-10-07",
   "source_name": "oMLX community benchmark trace: https://omlx.ai/benchmarks/performance/5mxqlvmd",
   "source_url": "https://omlx.ai/benchmarks/performance/5mxqlvmd",
   "tg_tokens": null,
   "tps": 18.0,
   "ttft_s": null
  },
  {
   "backend": "SGLang",
   "batch": 5,
   "ctx": 8192,
   "date": "2025-04",
   "flags": [],
   "hardware": "A10",
   "id": "qsl17221-a10-qwen257b-awq-sglang-c5",
   "model": "Qwen2.5-7B",
   "notes": "Single A10 24GB, SGLang 0.4.5 with --context-length 8192, 20 identical prompts at 5 concurrency, max_tokens 256; this is the stable repeat run (the cold first run is 313.00 t/s in the project's OneA10 sample files; the README documents the SGLang warm-up effect); aggregate server generation rate",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "Single, dual and quad NVIDIA A10 GPU (24GB) configurations; Model: Qwen2.5 7B-AWQ (quantized); Single-GPU Results; SGLang Stress Test Results - 2025-04-25 21:27:57; Concurrent Requests 5; Tokens Per Second 333.00",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "qiulang/vllm-sglang-perf README (via vLLM discussion #17221)",
   "source_url": "https://github.com/qiulang/vllm-sglang-perf",
   "tg_tokens": 256,
   "tps": "333.00",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 5,
   "ctx": 8192,
   "date": "2025-04",
   "flags": [],
   "hardware": "A10",
   "id": "qsl17221-a10-qwen257b-awq-vllm-c5",
   "model": "Qwen2.5-7B",
   "notes": "Single A10 24GB, vLLM v0.8.4 with --max-model-len 8192, 20 identical prompts at 5 concurrency, max_tokens 256; the README notes vLLM needs no warm-up; aggregate server generation rate",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "Single, dual and quad NVIDIA A10 GPU (24GB) configurations; Model: Qwen2.5 7B-AWQ (quantized); Single-GPU Results; vLLM Stress Test Results - 2025-04-25 22:08:34; Concurrent Requests 5; Tokens Per Second 259.53",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "qiulang/vllm-sglang-perf README (via vLLM discussion #17221)",
   "source_url": "https://github.com/qiulang/vllm-sglang-perf",
   "tg_tokens": 256,
   "tps": "259.53",
   "ttft_s": null
  },
  {
   "backend": "SGLang",
   "batch": 30,
   "ctx": 8192,
   "date": "2025-04",
   "flags": [],
   "hardware": "A10",
   "id": "qsl17221-a10-qwen257b-awq-sglang-c30",
   "model": "Qwen2.5-7B",
   "notes": "Single A10 24GB, SGLang 0.4.5 with --context-length 8192; aggregate token generation at 30 concurrency over 300 requests from the project's single-GPU comparison table (same row: vLLM 961.05 tok/s)",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "Single, dual and quad NVIDIA A10 GPU (24GB) configurations; Model: Qwen2.5 7B-AWQ (quantized); Single-GPU Comparison; 30 Concurrent Requests (300 total); Token Generation 1544.94 tok/s 961.05 tok/s",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "qiulang/vllm-sglang-perf README (via vLLM discussion #17221)",
   "source_url": "https://github.com/qiulang/vllm-sglang-perf",
   "tg_tokens": 256,
   "tps": "1544.94",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 30,
   "ctx": 8192,
   "date": "2025-04",
   "flags": [],
   "hardware": "A10",
   "id": "qsl17221-a10-qwen257b-awq-vllm-c30",
   "model": "Qwen2.5-7B",
   "notes": "Single A10 24GB, vLLM v0.8.4 with --max-model-len 8192; aggregate token generation at 30 concurrency over 300 requests from the project's single-GPU comparison table (same row: SGLang 1544.94 tok/s)",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "Single, dual and quad NVIDIA A10 GPU (24GB) configurations; Model: Qwen2.5 7B-AWQ (quantized); Single-GPU Comparison; 30 Concurrent Requests (300 total); Token Generation 1544.94 tok/s 961.05 tok/s",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "qiulang/vllm-sglang-perf README (via vLLM discussion #17221)",
   "source_url": "https://github.com/qiulang/vllm-sglang-perf",
   "tg_tokens": 256,
   "tps": "961.05",
   "ttft_s": null
  },
  {
   "backend": "SGLang",
   "batch": 30,
   "ctx": 8192,
   "date": "2025-04",
   "flags": [],
   "hardware": "2 A10",
   "id": "qsl17221-2a10-qwen257b-awq-sglang-dp-c30-r1",
   "model": "Qwen2.5-7B",
   "notes": "2 A10 with SGLang 0.4.5 --dp (data parallelism, a full model copy per GPU): the author documents heavy warm-up variance on this path; run 1 of 2 published DP runs (the repeat is 511.92 tok/s); 300 requests at 30 concurrency, max_tokens 256; config=dp",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "SGLang with Data Parallelism; SGLang Stress Test Results - 2025-04-27 14:00:37; Concurrent Requests 30; Tokens Per Second 853.94; Model: Qwen2.5 7B-AWQ (quantized); For comparison, vLLM's results on 2 A10 GPUs:",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "qiulang/vllm-sglang-perf README (via vLLM discussion #17221)",
   "source_url": "https://github.com/qiulang/vllm-sglang-perf",
   "tg_tokens": 256,
   "tps": "853.94",
   "ttft_s": null
  },
  {
   "backend": "SGLang",
   "batch": 30,
   "ctx": 8192,
   "date": "2025-04",
   "flags": [],
   "hardware": "2 A10",
   "id": "qsl17221-2a10-qwen257b-awq-sglang-dp-c30-r2",
   "model": "Qwen2.5-7B",
   "notes": "2 A10 with SGLang 0.4.5 --dp (data parallelism, a full model copy per GPU): the author documents heavy warm-up variance on this path; run 2 of 2 published DP runs (the first was 853.94 tok/s); 300 requests at 30 concurrency, max_tokens 256; config=dp",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "SGLang with Data Parallelism; SGLang Stress Test Results - 2025-04-27 14:10:34; Concurrent Requests 30; Tokens Per Second 511.92; Model: Qwen2.5 7B-AWQ (quantized); For comparison, vLLM's results on 2 A10 GPUs:",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "qiulang/vllm-sglang-perf README (via vLLM discussion #17221)",
   "source_url": "https://github.com/qiulang/vllm-sglang-perf",
   "tg_tokens": 256,
   "tps": "511.92",
   "ttft_s": null
  },
  {
   "backend": "SGLang",
   "batch": 30,
   "ctx": 8192,
   "date": "2025-04",
   "flags": [],
   "hardware": "2 A10",
   "id": "qsl17221-2a10-qwen257b-awq-sglang-tp-c30-r1",
   "model": "Qwen2.5-7B",
   "notes": "2 A10 with SGLang 0.4.5 --tp (tensor parallelism): after the author switched from --dp on advice from SGLang issue #5808, consistency returned to single-GPU levels; run 1 of 2 published TP runs (the repeat is 1151.33 tok/s); 300 requests at 30 concurrency, max_tokens 256; config=tp",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "SGLang with Tensor Parallelism; SGLang Stress Test Results - 2025-04-28 23:17:05; Concurrent Requests 30; Tokens Per Second 1157.76; Model: Qwen2.5 7B-AWQ (quantized); For comparison, vLLM's results on 2 A10 GPUs:",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "qiulang/vllm-sglang-perf README (via vLLM discussion #17221)",
   "source_url": "https://github.com/qiulang/vllm-sglang-perf",
   "tg_tokens": 256,
   "tps": "1157.76",
   "ttft_s": null
  },
  {
   "backend": "SGLang",
   "batch": 30,
   "ctx": 8192,
   "date": "2025-04",
   "flags": [],
   "hardware": "2 A10",
   "id": "qsl17221-2a10-qwen257b-awq-sglang-tp-c30-r2",
   "model": "Qwen2.5-7B",
   "notes": "2 A10 with SGLang 0.4.5 --tp (tensor parallelism); run 2 of 2 published TP runs (the first was 1157.76 tok/s); 300 requests at 30 concurrency, max_tokens 256; config=tp",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "SGLang with Tensor Parallelism; SGLang Stress Test Results - 2025-04-28 23:17:38; Concurrent Requests 30; Tokens Per Second 1151.33; Model: Qwen2.5 7B-AWQ (quantized); For comparison, vLLM's results on 2 A10 GPUs:",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "qiulang/vllm-sglang-perf README (via vLLM discussion #17221)",
   "source_url": "https://github.com/qiulang/vllm-sglang-perf",
   "tg_tokens": 256,
   "tps": "1151.33",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 30,
   "ctx": 8192,
   "date": "2025-04",
   "flags": [],
   "hardware": "2 A10",
   "id": "qsl17221-2a10-qwen257b-awq-vllm-c30-r1",
   "model": "Qwen2.5-7B",
   "notes": "2 A10 with vLLM v0.8.4 --tensor-parallel-size 2, 300 requests at 30 concurrency, max_tokens 256; run 1 of 2 published runs (the repeat is 1025.81 tok/s); config=tp",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "For comparison, vLLM's results on 2 A10 GPUs: vLLM Stress Test Results - 2025-04-27 14:26:27; Concurrent Requests 30; Tokens Per Second 1074.01; Model: Qwen2.5 7B-AWQ (quantized)",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "qiulang/vllm-sglang-perf README (via vLLM discussion #17221)",
   "source_url": "https://github.com/qiulang/vllm-sglang-perf",
   "tg_tokens": 256,
   "tps": "1074.01",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 30,
   "ctx": 8192,
   "date": "2025-04",
   "flags": [],
   "hardware": "2 A10",
   "id": "qsl17221-2a10-qwen257b-awq-vllm-c30-r2",
   "model": "Qwen2.5-7B",
   "notes": "2 A10 with vLLM v0.8.4 --tensor-parallel-size 2, 300 requests at 30 concurrency, max_tokens 256; run 2 of 2 published runs (the first was 1074.01 tok/s); config=tp",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "Another vLLM run showing similar consistency on 2 A10; vLLM Stress Test Results - 2025-04-27 14:25:12; Concurrent Requests 30; Tokens Per Second 1025.81; Model: Qwen2.5 7B-AWQ (quantized)",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "qiulang/vllm-sglang-perf README (via vLLM discussion #17221)",
   "source_url": "https://github.com/qiulang/vllm-sglang-perf",
   "tg_tokens": 256,
   "tps": "1025.81",
   "ttft_s": null
  },
  {
   "backend": "SGLang",
   "batch": 30,
   "ctx": 8192,
   "date": "2025-04",
   "flags": [],
   "hardware": "4 A10",
   "id": "qsl17221-4a10-qwen7bawq-sglang-c30-r1",
   "model": "Qwen7B-awq",
   "notes": "4 A10, SGLang 0.4.5 tensor-parallel 4 (nvidia-smi shows all four A10 loaded at 22 GB each in the 4.27-28 section); the results file names the model only by its local path /home/vllm/llm/Qwen7B-awq, which is the project's Qwen2.5-7B-AWQ download (per the README); 300 requests at 30 concurrency, max_tokens 256; run 1 of 2 (the repeat is 2123.63 tok/s); config=tp4",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "4A10.md; Model /home/vllm/llm/Qwen7B-awq; SGLang Stress Test Results - 2025-04-28 20:22:04; Concurrent Requests 30; Tokens Per Second 2086.65; vLLM Performance Comparison Between 2 A10 and 4 A10 GPUs",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "qiulang/vllm-sglang-perf 4A10.md (via vLLM discussion #17221)",
   "source_url": "https://raw.githubusercontent.com/qiulang/vllm-sglang-perf/main/4A10.md",
   "tg_tokens": 256,
   "tps": "2086.65",
   "ttft_s": null
  },
  {
   "backend": "SGLang",
   "batch": 30,
   "ctx": 8192,
   "date": "2025-04",
   "flags": [],
   "hardware": "4 A10",
   "id": "qsl17221-4a10-qwen7bawq-sglang-c30-r2",
   "model": "Qwen7B-awq",
   "notes": "4 A10, SGLang 0.4.5 tensor-parallel 4; model recorded by its local path /home/vllm/llm/Qwen7B-awq (the project's Qwen2.5-7B-AWQ download, per the README); 300 requests at 30 concurrency, max_tokens 256; run 2 of 2 (the first was 2086.65 tok/s); config=tp4",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "4A10.md; Model /home/vllm/llm/Qwen7B-awq; SGLang Stress Test Results - 2025-04-28 20:23:05; Concurrent Requests 30; Tokens Per Second 2123.63; vLLM Performance Comparison Between 2 A10 and 4 A10 GPUs",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "qiulang/vllm-sglang-perf 4A10.md (via vLLM discussion #17221)",
   "source_url": "https://raw.githubusercontent.com/qiulang/vllm-sglang-perf/main/4A10.md",
   "tg_tokens": 256,
   "tps": "2123.63",
   "ttft_s": null
  },
  {
   "backend": "SGLang",
   "batch": 50,
   "ctx": 8192,
   "date": "2025-04",
   "flags": [],
   "hardware": "4 A10",
   "id": "qsl17221-4a10-qwen7bawq-sglang-c50-r1",
   "model": "Qwen7B-awq",
   "notes": "4 A10, SGLang 0.4.5 tensor-parallel 4; model recorded by its local path /home/vllm/llm/Qwen7B-awq (the project's Qwen2.5-7B-AWQ download, per the README); 500 requests at 50 concurrency, max_tokens 256; run 1 of 2 (the repeat is 2526.98 tok/s); config=tp4",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "4A10.md; Model /home/vllm/llm/Qwen7B-awq; SGLang Stress Test Results - 2025-04-28 20:23:47; Concurrent Requests 50; Tokens Per Second 2560.58; vLLM Performance Comparison Between 2 A10 and 4 A10 GPUs",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "qiulang/vllm-sglang-perf 4A10.md (via vLLM discussion #17221)",
   "source_url": "https://raw.githubusercontent.com/qiulang/vllm-sglang-perf/main/4A10.md",
   "tg_tokens": 256,
   "tps": "2560.58",
   "ttft_s": null
  },
  {
   "backend": "SGLang",
   "batch": 50,
   "ctx": 8192,
   "date": "2025-04",
   "flags": [],
   "hardware": "4 A10",
   "id": "qsl17221-4a10-qwen7bawq-sglang-c50-r2",
   "model": "Qwen7B-awq",
   "notes": "4 A10, SGLang 0.4.5 tensor-parallel 4; model recorded by its local path /home/vllm/llm/Qwen7B-awq (the project's Qwen2.5-7B-AWQ download, per the README); 500 requests at 50 concurrency, max_tokens 256; run 2 of 2 (the first was 2560.58 tok/s); config=tp4",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "4A10.md; Model /home/vllm/llm/Qwen7B-awq; SGLang Stress Test Results - 2025-04-28 20:24:47; Concurrent Requests 50; Tokens Per Second 2526.98; vLLM Performance Comparison Between 2 A10 and 4 A10 GPUs",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "qiulang/vllm-sglang-perf 4A10.md (via vLLM discussion #17221)",
   "source_url": "https://raw.githubusercontent.com/qiulang/vllm-sglang-perf/main/4A10.md",
   "tg_tokens": 256,
   "tps": "2526.98",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 30,
   "ctx": 8192,
   "date": "2025-04",
   "flags": [],
   "hardware": "4 A10",
   "id": "qsl17221-4a10-qwen7bawq-vllm-c30-r1",
   "model": "Qwen7B-awq",
   "notes": "4 A10, vLLM --tensor-parallel-size 4, 300 requests at 30 concurrency, max_tokens 256; run 1 of 2 (the repeat is 1309.40 tok/s); config=tp4",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "vllm serve /home/vllm/llm/Qwen7B-awq --max-model-len 8192 --tensor-parallel-size 4; vLLM Stress Test Results - 2025-04-29 14:17:40; Concurrent Requests 30; Tokens Per Second 1318.94; vLLM Performance Comparison Between 2 A10 and 4 A10 GPUs",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "qiulang/vllm-sglang-perf 4A10.md (via vLLM discussion #17221)",
   "source_url": "https://raw.githubusercontent.com/qiulang/vllm-sglang-perf/main/4A10.md",
   "tg_tokens": 256,
   "tps": "1318.94",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 30,
   "ctx": 8192,
   "date": "2025-04",
   "flags": [],
   "hardware": "4 A10",
   "id": "qsl17221-4a10-qwen7bawq-vllm-c30-r2",
   "model": "Qwen7B-awq",
   "notes": "4 A10, vLLM --tensor-parallel-size 4, 300 requests at 30 concurrency, max_tokens 256; run 2 of 2 (the first was 1318.94 tok/s); config=tp4",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "vllm serve /home/vllm/llm/Qwen7B-awq --max-model-len 8192 --tensor-parallel-size 4; vLLM Stress Test Results - 2025-04-29 14:18:24; Concurrent Requests 30; Tokens Per Second 1309.40; vLLM Performance Comparison Between 2 A10 and 4 A10 GPUs",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "qiulang/vllm-sglang-perf 4A10.md (via vLLM discussion #17221)",
   "source_url": "https://raw.githubusercontent.com/qiulang/vllm-sglang-perf/main/4A10.md",
   "tg_tokens": 256,
   "tps": "1309.40",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 50,
   "ctx": 8192,
   "date": "2025-04",
   "flags": [],
   "hardware": "4 A10",
   "id": "qsl17221-4a10-qwen7bawq-vllm-c50-r1",
   "model": "Qwen7B-awq",
   "notes": "4 A10, vLLM --tensor-parallel-size 4, 500 requests at 50 concurrency, max_tokens 256; run 1 of 2 (the repeat is 1534.41 tok/s); config=tp4",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "vllm serve /home/vllm/llm/Qwen7B-awq --max-model-len 8192 --tensor-parallel-size 4; vLLM Stress Test Results - 2025-04-29 14:19:28; Concurrent Requests 50; Tokens Per Second 1566.10; vLLM Performance Comparison Between 2 A10 and 4 A10 GPUs",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "qiulang/vllm-sglang-perf 4A10.md (via vLLM discussion #17221)",
   "source_url": "https://raw.githubusercontent.com/qiulang/vllm-sglang-perf/main/4A10.md",
   "tg_tokens": 256,
   "tps": "1566.10",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 50,
   "ctx": 8192,
   "date": "2025-04",
   "flags": [],
   "hardware": "4 A10",
   "id": "qsl17221-4a10-qwen7bawq-vllm-c50-r2",
   "model": "Qwen7B-awq",
   "notes": "4 A10, vLLM --tensor-parallel-size 4, 500 requests at 50 concurrency, max_tokens 256; run 2 of 2 (the first was 1566.10 tok/s); config=tp4",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "vllm serve /home/vllm/llm/Qwen7B-awq --max-model-len 8192 --tensor-parallel-size 4; vLLM Stress Test Results - 2025-04-29 14:20:16; Concurrent Requests 50; Tokens Per Second 1534.41; vLLM Performance Comparison Between 2 A10 and 4 A10 GPUs",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "qiulang/vllm-sglang-perf 4A10.md (via vLLM discussion #17221)",
   "source_url": "https://raw.githubusercontent.com/qiulang/vllm-sglang-perf/main/4A10.md",
   "tg_tokens": 256,
   "tps": "1534.41",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 100,
   "ctx": 8192,
   "date": "2025-04",
   "flags": [],
   "hardware": "4 A10",
   "id": "qsl17221-4a10-qwen7bawq-vllm-c100-r1",
   "model": "Qwen7B-awq",
   "notes": "4 A10, vLLM --tensor-parallel-size 4, 1000 requests at 100 concurrency, max_tokens 256; run 1 of 2 (the repeat is 1887.62 tok/s); the project's scaling analysis reports only 2-10% gain from 2 to 4 GPUs at high concurrency; config=tp4",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "vllm serve /home/vllm/llm/Qwen7B-awq --max-model-len 8192 --tensor-parallel-size 4; vLLM Stress Test Results - 2025-04-29 14:34:25; Concurrent Requests 100; Tokens Per Second 1914.26; vLLM Performance Comparison Between 2 A10 and 4 A10 GPUs",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "qiulang/vllm-sglang-perf 4A10.md (via vLLM discussion #17221)",
   "source_url": "https://raw.githubusercontent.com/qiulang/vllm-sglang-perf/main/4A10.md",
   "tg_tokens": 256,
   "tps": "1914.26",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 100,
   "ctx": 8192,
   "date": "2025-04",
   "flags": [],
   "hardware": "4 A10",
   "id": "qsl17221-4a10-qwen7bawq-vllm-c100-r2",
   "model": "Qwen7B-awq",
   "notes": "4 A10, vLLM --tensor-parallel-size 4, 1000 requests at 100 concurrency, max_tokens 256; run 2 of 2 (the first was 1914.26 tok/s); config=tp4",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "vllm serve /home/vllm/llm/Qwen7B-awq --max-model-len 8192 --tensor-parallel-size 4; vLLM Stress Test Results - 2025-04-29 14:35:37; Concurrent Requests 100; Tokens Per Second 1887.62; vLLM Performance Comparison Between 2 A10 and 4 A10 GPUs",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "qiulang/vllm-sglang-perf 4A10.md (via vLLM discussion #17221)",
   "source_url": "https://raw.githubusercontent.com/qiulang/vllm-sglang-perf/main/4A10.md",
   "tg_tokens": 256,
   "tps": "1887.62",
   "ttft_s": null
  },
  {
   "backend": "SGLang",
   "batch": 30,
   "ctx": 8192,
   "date": "2025-04",
   "flags": [],
   "hardware": "4 A10",
   "id": "qsl17221-4a10-qwen7bawq-sglang-c30-r3",
   "model": "Qwen7B-awq",
   "notes": "4 A10, SGLang 0.4.5 tensor-parallel 4, 300 requests at 30 concurrency, max_tokens 256; the 4.29 warm-up retest of the 4.27-28 4-GPU section (its c30 runs were 2086.65 and 2123.63 tok/s); config=tp4",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "4A10.md; Model /home/vllm/llm/Qwen7B-awq; SGLang Stress Test Results - 2025-04-29 14:05:46; Concurrent Requests 30; Tokens Per Second 2119.20; vLLM Performance Comparison Between 2 A10 and 4 A10 GPUs",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "qiulang/vllm-sglang-perf 4A10.md (via vLLM discussion #17221)",
   "source_url": "https://raw.githubusercontent.com/qiulang/vllm-sglang-perf/main/4A10.md",
   "tg_tokens": 256,
   "tps": "2119.20",
   "ttft_s": null
  },
  {
   "backend": "SGLang",
   "batch": 50,
   "ctx": 8192,
   "date": "2025-04",
   "flags": [],
   "hardware": "4 A10",
   "id": "qsl17221-4a10-qwen7bawq-sglang-c50-r3",
   "model": "Qwen7B-awq",
   "notes": "4 A10, SGLang 0.4.5 tensor-parallel 4, 500 requests at 50 concurrency, max_tokens 256; the 4.29 warm-up retest of the 4.27-28 4-GPU section (its c50 runs were 2560.58 and 2526.98 tok/s); config=tp4",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "4A10.md; Model /home/vllm/llm/Qwen7B-awq; SGLang Stress Test Results - 2025-04-29 14:06:42; Concurrent Requests 50; Tokens Per Second 2540.23; vLLM Performance Comparison Between 2 A10 and 4 A10 GPUs",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "qiulang/vllm-sglang-perf 4A10.md (via vLLM discussion #17221)",
   "source_url": "https://raw.githubusercontent.com/qiulang/vllm-sglang-perf/main/4A10.md",
   "tg_tokens": 256,
   "tps": "2540.23",
   "ttft_s": null
  },
  {
   "backend": "SGLang",
   "batch": 5,
   "ctx": 8192,
   "date": "2025-04",
   "flags": [],
   "hardware": "2 A10",
   "id": "qsl17221-2a10-qwen7bawq-sglang-c5",
   "model": "Qwen7B-awq",
   "notes": "2 of the 4 A10, SGLang 0.4.5 --tp 2 (the 4.29 comparison sweep, context-length 8192); model recorded by its local path /home/vllm/llm/Qwen7B-awq (the project's Qwen2.5-7B-AWQ download, per the README); 20 requests at 5 concurrency, max_tokens 256; config=tp2",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "4A10.md; Model /home/vllm/llm/Qwen7B-awq; SGLang Stress Test Results - 2025-04-29 14:08:06; Concurrent Requests 5; Tokens Per Second 422.97; vLLM Performance Comparison Between 2 A10 and 4 A10 GPUs",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "qiulang/vllm-sglang-perf 4A10.md (via vLLM discussion #17221)",
   "source_url": "https://raw.githubusercontent.com/qiulang/vllm-sglang-perf/main/4A10.md",
   "tg_tokens": 256,
   "tps": "422.97",
   "ttft_s": null
  },
  {
   "backend": "SGLang",
   "batch": 30,
   "ctx": 8192,
   "date": "2025-04",
   "flags": [],
   "hardware": "2 A10",
   "id": "qsl17221-2a10-qwen7bawq-sglang-c30-r1",
   "model": "Qwen7B-awq",
   "notes": "2 of the 4 A10, SGLang 0.4.5 --tp 2 (the 4.29 comparison sweep, context-length 8192); model recorded by its local path /home/vllm/llm/Qwen7B-awq (the project's Qwen2.5-7B-AWQ download, per the README); 300 requests at 30 concurrency, max_tokens 256; run 1 of 2 (the repeat is 1115.89 tok/s); config=tp2",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "4A10.md; Model /home/vllm/llm/Qwen7B-awq; SGLang Stress Test Results - 2025-04-29 14:09:01; Concurrent Requests 30; Tokens Per Second 1117.99; vLLM Performance Comparison Between 2 A10 and 4 A10 GPUs",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "qiulang/vllm-sglang-perf 4A10.md (via vLLM discussion #17221)",
   "source_url": "https://raw.githubusercontent.com/qiulang/vllm-sglang-perf/main/4A10.md",
   "tg_tokens": 256,
   "tps": "1117.99",
   "ttft_s": null
  },
  {
   "backend": "SGLang",
   "batch": 50,
   "ctx": 8192,
   "date": "2025-04",
   "flags": [],
   "hardware": "2 A10",
   "id": "qsl17221-2a10-qwen7bawq-sglang-c50-r1",
   "model": "Qwen7B-awq",
   "notes": "2 of the 4 A10, SGLang 0.4.5 --tp 2 (the 4.29 comparison sweep, context-length 8192); model recorded by its local path /home/vllm/llm/Qwen7B-awq (the project's Qwen2.5-7B-AWQ download, per the README); 500 requests at 50 concurrency, max_tokens 256; run 1 of 2 (the repeat is 1811.35 tok/s); config=tp2",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "4A10.md; Model /home/vllm/llm/Qwen7B-awq; SGLang Stress Test Results - 2025-04-29 14:10:00; Concurrent Requests 50; Tokens Per Second 1822.21; vLLM Performance Comparison Between 2 A10 and 4 A10 GPUs",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "qiulang/vllm-sglang-perf 4A10.md (via vLLM discussion #17221)",
   "source_url": "https://raw.githubusercontent.com/qiulang/vllm-sglang-perf/main/4A10.md",
   "tg_tokens": 256,
   "tps": "1822.21",
   "ttft_s": null
  },
  {
   "backend": "SGLang",
   "batch": 50,
   "ctx": 8192,
   "date": "2025-04",
   "flags": [],
   "hardware": "2 A10",
   "id": "qsl17221-2a10-qwen7bawq-sglang-c50-r2",
   "model": "Qwen7B-awq",
   "notes": "2 of the 4 A10, SGLang 0.4.5 --tp 2 (the 4.29 comparison sweep, context-length 8192); model recorded by its local path /home/vllm/llm/Qwen7B-awq (the project's Qwen2.5-7B-AWQ download, per the README); 500 requests at 50 concurrency, max_tokens 256; run 2 of 2 (the first was 1822.21 tok/s); config=tp2",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "4A10.md; Model /home/vllm/llm/Qwen7B-awq; SGLang Stress Test Results - 2025-04-29 14:11:38; Concurrent Requests 50; Tokens Per Second 1811.35; vLLM Performance Comparison Between 2 A10 and 4 A10 GPUs",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "qiulang/vllm-sglang-perf 4A10.md (via vLLM discussion #17221)",
   "source_url": "https://raw.githubusercontent.com/qiulang/vllm-sglang-perf/main/4A10.md",
   "tg_tokens": 256,
   "tps": "1811.35",
   "ttft_s": null
  },
  {
   "backend": "SGLang",
   "batch": 30,
   "ctx": 8192,
   "date": "2025-04",
   "flags": [],
   "hardware": "2 A10",
   "id": "qsl17221-2a10-qwen7bawq-sglang-c30-r2",
   "model": "Qwen7B-awq",
   "notes": "2 of the 4 A10, SGLang 0.4.5 --tp 2 (the 4.29 comparison sweep, context-length 8192); model recorded by its local path /home/vllm/llm/Qwen7B-awq (the project's Qwen2.5-7B-AWQ download, per the README); 300 requests at 30 concurrency, max_tokens 256; run 2 of 2 (the first was 1117.99 tok/s); config=tp2",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "4A10.md; Model /home/vllm/llm/Qwen7B-awq; SGLang Stress Test Results - 2025-04-29 14:12:25; Concurrent Requests 30; Tokens Per Second 1115.89; vLLM Performance Comparison Between 2 A10 and 4 A10 GPUs",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "qiulang/vllm-sglang-perf 4A10.md (via vLLM discussion #17221)",
   "source_url": "https://raw.githubusercontent.com/qiulang/vllm-sglang-perf/main/4A10.md",
   "tg_tokens": 256,
   "tps": "1115.89",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 30,
   "ctx": 8192,
   "date": "2025-04",
   "flags": [],
   "hardware": "2 A10",
   "id": "qsl17221-2a10-qwen7bawq-vllm-c30-r1",
   "model": "Qwen7B-awq",
   "notes": "2 of the 4 A10, vLLM --tensor-parallel-size 2 (the 4.29 comparison sweep); 300 requests at 30 concurrency, max_tokens 256; run 1 of 2 (the repeat is 1217.92 tok/s); config=tp2",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "vllm serve /home/vllm/llm/Qwen7B-awq --max-model-len 8192 --tensor-parallel-size 2; vLLM Stress Test Results - 2025-04-29 14:25:50; Concurrent Requests 30; Tokens Per Second 1131.43; vLLM Performance Comparison Between 2 A10 and 4 A10 GPUs",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "qiulang/vllm-sglang-perf 4A10.md (via vLLM discussion #17221)",
   "source_url": "https://raw.githubusercontent.com/qiulang/vllm-sglang-perf/main/4A10.md",
   "tg_tokens": 256,
   "tps": "1131.43",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 30,
   "ctx": 8192,
   "date": "2025-04",
   "flags": [],
   "hardware": "2 A10",
   "id": "qsl17221-2a10-qwen7bawq-vllm-c30-r2",
   "model": "Qwen7B-awq",
   "notes": "2 of the 4 A10, vLLM --tensor-parallel-size 2 (the 4.29 comparison sweep); 300 requests at 30 concurrency, max_tokens 256; run 2 of 2 (the first was 1131.43 tok/s); config=tp2",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "vllm serve /home/vllm/llm/Qwen7B-awq --max-model-len 8192 --tensor-parallel-size 2; vLLM Stress Test Results - 2025-04-29 14:26:58; Concurrent Requests 30; Tokens Per Second 1217.92; vLLM Performance Comparison Between 2 A10 and 4 A10 GPUs",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "qiulang/vllm-sglang-perf 4A10.md (via vLLM discussion #17221)",
   "source_url": "https://raw.githubusercontent.com/qiulang/vllm-sglang-perf/main/4A10.md",
   "tg_tokens": 256,
   "tps": "1217.92",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 50,
   "ctx": 8192,
   "date": "2025-04",
   "flags": [],
   "hardware": "2 A10",
   "id": "qsl17221-2a10-qwen7bawq-vllm-c50-r1",
   "model": "Qwen7B-awq",
   "notes": "2 of the 4 A10, vLLM --tensor-parallel-size 2 (the 4.29 comparison sweep); 500 requests at 50 concurrency, max_tokens 256; run 1 of 2 (the repeat is 1519.86 tok/s); config=tp2",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "vllm serve /home/vllm/llm/Qwen7B-awq --max-model-len 8192 --tensor-parallel-size 2; vLLM Stress Test Results - 2025-04-29 14:27:46; Concurrent Requests 50; Tokens Per Second 1534.72; vLLM Performance Comparison Between 2 A10 and 4 A10 GPUs",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "qiulang/vllm-sglang-perf 4A10.md (via vLLM discussion #17221)",
   "source_url": "https://raw.githubusercontent.com/qiulang/vllm-sglang-perf/main/4A10.md",
   "tg_tokens": 256,
   "tps": "1534.72",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 50,
   "ctx": 8192,
   "date": "2025-04",
   "flags": [],
   "hardware": "2 A10",
   "id": "qsl17221-2a10-qwen7bawq-vllm-c50-r2",
   "model": "Qwen7B-awq",
   "notes": "2 of the 4 A10, vLLM --tensor-parallel-size 2 (the 4.29 comparison sweep); 500 requests at 50 concurrency, max_tokens 256; run 2 of 2 (the first was 1534.72 tok/s); config=tp2",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "vllm serve /home/vllm/llm/Qwen7B-awq --max-model-len 8192 --tensor-parallel-size 2; vLLM Stress Test Results - 2025-04-29 14:28:54; Concurrent Requests 50; Tokens Per Second 1519.86; vLLM Performance Comparison Between 2 A10 and 4 A10 GPUs",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "qiulang/vllm-sglang-perf 4A10.md (via vLLM discussion #17221)",
   "source_url": "https://raw.githubusercontent.com/qiulang/vllm-sglang-perf/main/4A10.md",
   "tg_tokens": 256,
   "tps": "1519.86",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 100,
   "ctx": 8192,
   "date": "2025-04",
   "flags": [],
   "hardware": "2 A10",
   "id": "qsl17221-2a10-qwen7bawq-vllm-c100-r1",
   "model": "Qwen7B-awq",
   "notes": "2 of the 4 A10, vLLM --tensor-parallel-size 2 (the 4.29 comparison sweep); 1000 requests at 100 concurrency, max_tokens 256; run 1 of 2 (the repeat is 1816.33 tok/s); config=tp2",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "vllm serve /home/vllm/llm/Qwen7B-awq --max-model-len 8192 --tensor-parallel-size 2; vLLM Stress Test Results - 2025-04-29 14:30:29; Concurrent Requests 100; Tokens Per Second 1757.90; vLLM Performance Comparison Between 2 A10 and 4 A10 GPUs",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "qiulang/vllm-sglang-perf 4A10.md (via vLLM discussion #17221)",
   "source_url": "https://raw.githubusercontent.com/qiulang/vllm-sglang-perf/main/4A10.md",
   "tg_tokens": 256,
   "tps": "1757.90",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 100,
   "ctx": 8192,
   "date": "2025-04",
   "flags": [],
   "hardware": "2 A10",
   "id": "qsl17221-2a10-qwen7bawq-vllm-c100-r2",
   "model": "Qwen7B-awq",
   "notes": "2 of the 4 A10, vLLM --tensor-parallel-size 2 (the 4.29 comparison sweep); 1000 requests at 100 concurrency, max_tokens 256; run 2 of 2 (the first was 1757.90 tok/s); config=tp2",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "vllm serve /home/vllm/llm/Qwen7B-awq --max-model-len 8192 --tensor-parallel-size 2; vLLM Stress Test Results - 2025-04-29 14:31:44; Concurrent Requests 100; Tokens Per Second 1816.33; vLLM Performance Comparison Between 2 A10 and 4 A10 GPUs",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "qiulang/vllm-sglang-perf 4A10.md (via vLLM discussion #17221)",
   "source_url": "https://raw.githubusercontent.com/qiulang/vllm-sglang-perf/main/4A10.md",
   "tg_tokens": 256,
   "tps": "1816.33",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 10,
   "ctx": null,
   "date": "2025-05",
   "flags": [],
   "hardware": "A10",
   "id": "qsl17221-a10-qwen34b-vllm-c10-think-mt3096-r1",
   "model": "Qwen3 4B",
   "notes": "Qwen3-4B, vLLM, OpenAI-compatible API at 10 concurrency, 100 requests, max_tokens 3096, thinking mode on; the project's appendix ran the Qwen3-4B pass on the same A10 setup as the original tests; run 1 of 2 (the repeat is 395.12 tok/s); config=think",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "Model /home/vllm/llm/Qwen/Qwen3-4B; Max Tokens 3096; Concurrent Requests 10; Tokens Per Second 377.98; 4A10.md\",\"contentType\":\"file\"",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "qiulang/vllm-sglang-perf qwen3-final.md (via vLLM discussion #17221)",
   "source_url": "https://raw.githubusercontent.com/qiulang/vllm-sglang-perf/main/qwen3-final.md",
   "tg_tokens": 3096,
   "tps": "377.98",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 10,
   "ctx": null,
   "date": "2025-05",
   "flags": [],
   "hardware": "A10",
   "id": "qsl17221-a10-qwen34b-vllm-c10-think-mt3096-r2",
   "model": "Qwen3 4B",
   "notes": "Qwen3-4B, vLLM, OpenAI-compatible API at 10 concurrency, 100 requests, max_tokens 3096, thinking mode on; run 2 of 2 (the first was 377.98 tok/s); config=think",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "Model /home/vllm/llm/Qwen/Qwen3-4B; Max Tokens 3096; Concurrent Requests 10; Tokens Per Second 395.12; 4A10.md\",\"contentType\":\"file\"",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "qiulang/vllm-sglang-perf qwen3-final.md (via vLLM discussion #17221)",
   "source_url": "https://raw.githubusercontent.com/qiulang/vllm-sglang-perf/main/qwen3-final.md",
   "tg_tokens": 3096,
   "tps": "395.12",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 10,
   "ctx": null,
   "date": "2025-05",
   "flags": [],
   "hardware": "A10",
   "id": "qsl17221-a10-qwen34b-vllm-c10-notthink-mt3096-r1",
   "model": "Qwen3 4B",
   "notes": "Qwen3-4B, vLLM, OpenAI-compatible API at 10 concurrency, 100 requests, max_tokens 3096, thinking mode off; run 1 of 2 (the repeat is 249.99 tok/s); config=notthink",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "Model /home/vllm/llm/Qwen/Qwen3-4B; Max Tokens 3096; Concurrent Requests 10; Tokens Per Second 278.73; 4A10.md\",\"contentType\":\"file\"",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "qiulang/vllm-sglang-perf qwen3-final.md (via vLLM discussion #17221)",
   "source_url": "https://raw.githubusercontent.com/qiulang/vllm-sglang-perf/main/qwen3-final.md",
   "tg_tokens": 3096,
   "tps": "278.73",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 10,
   "ctx": null,
   "date": "2025-05",
   "flags": [],
   "hardware": "A10",
   "id": "qsl17221-a10-qwen34b-vllm-c10-notthink-mt3096-r2",
   "model": "Qwen3 4B",
   "notes": "Qwen3-4B, vLLM, OpenAI-compatible API at 10 concurrency, 100 requests, max_tokens 3096, thinking mode off; run 2 of 2 (the first was 278.73 tok/s); config=notthink",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "Model /home/vllm/llm/Qwen/Qwen3-4B; Max Tokens 3096; Concurrent Requests 10; Tokens Per Second 249.99; 4A10.md\",\"contentType\":\"file\"",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "qiulang/vllm-sglang-perf qwen3-final.md (via vLLM discussion #17221)",
   "source_url": "https://raw.githubusercontent.com/qiulang/vllm-sglang-perf/main/qwen3-final.md",
   "tg_tokens": 3096,
   "tps": "249.99",
   "ttft_s": null
  },
  {
   "backend": "SGLang",
   "batch": 10,
   "ctx": null,
   "date": "2025-05",
   "flags": [],
   "hardware": "A10",
   "id": "qsl17221-a10-qwen34b-sglang-c10-notthink-mt3096-r1",
   "model": "Qwen3 4B",
   "notes": "Qwen3-4B, SGLang, OpenAI-compatible API at 10 concurrency, 100 requests, max_tokens 3096, thinking mode off; run 1 of 2 (the repeat is 239.82 tok/s); config=notthink",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "Model /home/vllm/llm/Qwen/Qwen3-4B; Max Tokens 3096; Concurrent Requests 10; Tokens Per Second 238.21; 4A10.md\",\"contentType\":\"file\"",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "qiulang/vllm-sglang-perf qwen3-final.md (via vLLM discussion #17221)",
   "source_url": "https://raw.githubusercontent.com/qiulang/vllm-sglang-perf/main/qwen3-final.md",
   "tg_tokens": 3096,
   "tps": "238.21",
   "ttft_s": null
  },
  {
   "backend": "SGLang",
   "batch": 10,
   "ctx": null,
   "date": "2025-05",
   "flags": [],
   "hardware": "A10",
   "id": "qsl17221-a10-qwen34b-sglang-c10-notthink-mt3096-r2",
   "model": "Qwen3 4B",
   "notes": "Qwen3-4B, SGLang, OpenAI-compatible API at 10 concurrency, 100 requests, max_tokens 3096, thinking mode off; run 2 of 2 (the first was 238.21 tok/s); config=notthink",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "Model /home/vllm/llm/Qwen/Qwen3-4B; Max Tokens 3096; Concurrent Requests 10; Tokens Per Second 239.82; 4A10.md\",\"contentType\":\"file\"",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "qiulang/vllm-sglang-perf qwen3-final.md (via vLLM discussion #17221)",
   "source_url": "https://raw.githubusercontent.com/qiulang/vllm-sglang-perf/main/qwen3-final.md",
   "tg_tokens": 3096,
   "tps": "239.82",
   "ttft_s": null
  },
  {
   "backend": "SGLang",
   "batch": 10,
   "ctx": null,
   "date": "2025-05",
   "flags": [],
   "hardware": "A10",
   "id": "qsl17221-a10-qwen34b-sglang-c10-think-mt3096-r1",
   "model": "Qwen3 4B",
   "notes": "Qwen3-4B, SGLang, OpenAI-compatible API at 10 concurrency, 100 requests, max_tokens 3096, thinking mode on; run 1 of 2 (the repeat is 369.19 tok/s); the README appendix notes thinking mode hit the 256/3096 token cap early, shortening responses; config=think",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "Model /home/vllm/llm/Qwen/Qwen3-4B; Max Tokens 3096; Concurrent Requests 10; Tokens Per Second 337.17; 4A10.md\",\"contentType\":\"file\"",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "qiulang/vllm-sglang-perf qwen3-final.md (via vLLM discussion #17221)",
   "source_url": "https://raw.githubusercontent.com/qiulang/vllm-sglang-perf/main/qwen3-final.md",
   "tg_tokens": 3096,
   "tps": "337.17",
   "ttft_s": null
  },
  {
   "backend": "SGLang",
   "batch": 10,
   "ctx": null,
   "date": "2025-05",
   "flags": [],
   "hardware": "A10",
   "id": "qsl17221-a10-qwen34b-sglang-c10-think-mt3096-r2",
   "model": "Qwen3 4B",
   "notes": "Qwen3-4B, SGLang, OpenAI-compatible API at 10 concurrency, 100 requests, max_tokens 3096, thinking mode on; run 2 of 2 (the first was 337.17 tok/s); config=think",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "Model /home/vllm/llm/Qwen/Qwen3-4B; Max Tokens 3096; Concurrent Requests 10; Tokens Per Second 369.19; 4A10.md\",\"contentType\":\"file\"",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "qiulang/vllm-sglang-perf qwen3-final.md (via vLLM discussion #17221)",
   "source_url": "https://raw.githubusercontent.com/qiulang/vllm-sglang-perf/main/qwen3-final.md",
   "tg_tokens": 3096,
   "tps": "369.19",
   "ttft_s": null
  },
  {
   "backend": "SGLang",
   "batch": 10,
   "ctx": null,
   "date": "2025-05",
   "flags": [],
   "hardware": "A10",
   "id": "qsl17221-a10-qwen34b-sglang-c10-notthink-mt512-r1",
   "model": "Qwen3 4B",
   "notes": "Qwen3-4B, SGLang, OpenAI-compatible API at 10 concurrency, 100 requests, max_tokens 512, thinking mode off; run 1 of 2 (the repeat is 413.54 tok/s); config=notthink",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "max_token 512 with no think mode; Model /home/vllm/llm/Qwen/Qwen3-4B; Max Tokens 512; Concurrent Requests 10; Tokens Per Second 401.26; 4A10.md\",\"contentType\":\"file\"",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "qiulang/vllm-sglang-perf qwen3-final.md (via vLLM discussion #17221)",
   "source_url": "https://raw.githubusercontent.com/qiulang/vllm-sglang-perf/main/qwen3-final.md",
   "tg_tokens": 512,
   "tps": "401.26",
   "ttft_s": null
  },
  {
   "backend": "SGLang",
   "batch": 10,
   "ctx": null,
   "date": "2025-05",
   "flags": [],
   "hardware": "A10",
   "id": "qsl17221-a10-qwen34b-sglang-c10-notthink-mt512-r2",
   "model": "Qwen3 4B",
   "notes": "Qwen3-4B, SGLang, OpenAI-compatible API at 10 concurrency, 100 requests, max_tokens 512, thinking mode off; run 2 of 2 (the first was 401.26 tok/s); config=notthink",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "max_token 512 with no think mode; Model /home/vllm/llm/Qwen/Qwen3-4B; Max Tokens 512; Concurrent Requests 10; Tokens Per Second 413.54; 4A10.md\",\"contentType\":\"file\"",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "qiulang/vllm-sglang-perf qwen3-final.md (via vLLM discussion #17221)",
   "source_url": "https://raw.githubusercontent.com/qiulang/vllm-sglang-perf/main/qwen3-final.md",
   "tg_tokens": 512,
   "tps": "413.54",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 10,
   "ctx": null,
   "date": "2025-05",
   "flags": [],
   "hardware": "A10",
   "id": "qsl17221-a10-qwen34b-vllm-c10-notthink-mt512-r1",
   "model": "Qwen3 4B",
   "notes": "Qwen3-4B, vLLM, OpenAI-compatible API at 10 concurrency, 100 requests, max_tokens 512, thinking mode off; run 1 of 2 (the repeat is 433.77 tok/s); config=notthink",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "max_token 512 with no think mode; Model /home/vllm/llm/Qwen/Qwen3-4B; Max Tokens 512; Concurrent Requests 10; Tokens Per Second 469.62; 4A10.md\",\"contentType\":\"file\"",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "qiulang/vllm-sglang-perf qwen3-final.md (via vLLM discussion #17221)",
   "source_url": "https://raw.githubusercontent.com/qiulang/vllm-sglang-perf/main/qwen3-final.md",
   "tg_tokens": 512,
   "tps": "469.62",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 10,
   "ctx": null,
   "date": "2025-05",
   "flags": [],
   "hardware": "A10",
   "id": "qsl17221-a10-qwen34b-vllm-c10-notthink-mt512-r2",
   "model": "Qwen3 4B",
   "notes": "Qwen3-4B, vLLM, OpenAI-compatible API at 10 concurrency, 100 requests, max_tokens 512, thinking mode off; run 2 of 2 (the first was 469.62 tok/s); config=notthink",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "max_token 512 with no think mode; Model /home/vllm/llm/Qwen/Qwen3-4B; Max Tokens 512; Concurrent Requests 10; Tokens Per Second 433.77; 4A10.md\",\"contentType\":\"file\"",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "qiulang/vllm-sglang-perf qwen3-final.md (via vLLM discussion #17221)",
   "source_url": "https://raw.githubusercontent.com/qiulang/vllm-sglang-perf/main/qwen3-final.md",
   "tg_tokens": 512,
   "tps": "433.77",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 10,
   "ctx": null,
   "date": "2025-05",
   "flags": [],
   "hardware": "A10",
   "id": "qsl17221-a10-qwen34b-vllm-native-mt512",
   "model": "Qwen3 4B",
   "notes": "Qwen3-4B, vLLM native /v1/completions API, max_tokens 512; the project notes vLLM generates exactly the requested token count on the native API; the SGLang native API ignores max_tokens entirely (its runs are capped at 128 tokens, so the two are not directly comparable at this max_tokens)",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "native api; Model /home/vllm/llm/Qwen/Qwen3-4B; Max Tokens 512; Tokens Per Second 656.60; 4A10.md\",\"contentType\":\"file\"",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "qiulang/vllm-sglang-perf qwen3-final.md (via vLLM discussion #17221)",
   "source_url": "https://raw.githubusercontent.com/qiulang/vllm-sglang-perf/main/qwen3-final.md",
   "tg_tokens": 512,
   "tps": "656.60",
   "ttft_s": null
  },
  {
   "backend": "vLLM",
   "batch": 10,
   "ctx": null,
   "date": "2025-05",
   "flags": [],
   "hardware": "A10",
   "id": "qsl17221-a10-qwen34b-vllm-native-mt1024",
   "model": "Qwen3 4B",
   "notes": "Qwen3-4B, vLLM native /v1/completions API, max_tokens 1024",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "native api; Model /home/vllm/llm/Qwen/Qwen3-4B; Max Tokens 1024; Tokens Per Second 607.98; 4A10.md\",\"contentType\":\"file\"",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "qiulang/vllm-sglang-perf qwen3-final.md (via vLLM discussion #17221)",
   "source_url": "https://raw.githubusercontent.com/qiulang/vllm-sglang-perf/main/qwen3-final.md",
   "tg_tokens": 1024,
   "tps": "607.98",
   "ttft_s": null
  },
  {
   "backend": "SGLang",
   "batch": 10,
   "ctx": null,
   "date": "2025-05",
   "flags": [],
   "hardware": "A10",
   "id": "qsl17221-a10-qwen34b-sglang-native-mt512",
   "model": "Qwen3 4B",
   "notes": "Qwen3-4B, SGLang native /generate API, max_tokens 512 as requested; SGLang's native API ignores max_tokens (responses end at 128 tokens with finish_reason length 128), so the measured span is a 128-token completion",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "native api; Model /home/vllm/llm/Qwen/Qwen3-4B; Max Tokens 512; Tokens Per Second 424.58; 4A10.md\",\"contentType\":\"file\"",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "qiulang/vllm-sglang-perf qwen3-final.md (via vLLM discussion #17221)",
   "source_url": "https://raw.githubusercontent.com/qiulang/vllm-sglang-perf/main/qwen3-final.md",
   "tg_tokens": 512,
   "tps": "424.58",
   "ttft_s": null
  },
  {
   "backend": "SGLang",
   "batch": 10,
   "ctx": null,
   "date": "2025-05",
   "flags": [],
   "hardware": "A10",
   "id": "qsl17221-a10-qwen34b-sglang-native-mt1024",
   "model": "Qwen3 4B",
   "notes": "Qwen3-4B, SGLang native /generate API, max_tokens 1024 as requested; SGLang's native API ignores max_tokens (responses end at 128 tokens with finish_reason length 128), so the measured span is a 128-token completion",
   "params": null,
   "power_w": null,
   "pp_tokens": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "AWQ",
   "quote": "native api; Model /home/vllm/llm/Qwen/Qwen3-4B; Max Tokens 1024; Tokens Per Second 423.26; 4A10.md\",\"contentType\":\"file\"",
   "ram_gb": null,
   "retrieved": "2026-09-27",
   "scope": null,
   "source_name": "qiulang/vllm-sglang-perf qwen3-final.md (via vLLM discussion #17221)",
   "source_url": "https://raw.githubusercontent.com/qiulang/vllm-sglang-perf/main/qwen3-final.md",
   "tg_tokens": 1024,
   "tps": "423.26",
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "M1 Max (64 GB)",
   "id": "siliconscore-m1-max-64-gb-glm-4-7-flash-q4-k-xl-ctx4096",
   "model": "GLM-4.7-Flash",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 99.4,
   "provenance": "sourced",
   "quant": "Q4_K_XL",
   "quote": "M1 Max (64 GB),GLM-4.7-Flash,Q4_K_XL,ctx 4096,36.8 tok/s",
   "ram_gb": 17,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1rdx2c7/ran_3_popular_30b_moe_models_on_my_apple_silicon/",
   "tps": 36.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "M1 Max (64 GB)",
   "id": "siliconscore-m1-max-64-gb-nemotron-3-nano-30b-a3b-q4-k-xl-ctx4096",
   "model": "Nemotron 3 Nano 30B-A3B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 136.9,
   "provenance": "sourced",
   "quant": "Q4_K_XL",
   "quote": "M1 Max (64 GB),Nemotron-3-Nano-30B-A3B,Q4_K_XL,ctx 4096,43.7 tok/s",
   "ram_gb": 22,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1rdx2c7/ran_3_popular_30b_moe_models_on_my_apple_silicon/",
   "tps": 43.7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "M1 Max (64 GB)",
   "id": "siliconscore-m1-max-64-gb-qwen3-coder-30b-a3b-iq4-xs-ctx4096",
   "model": "Qwen3-Coder-30B-A3B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 132.1,
   "provenance": "sourced",
   "quant": "IQ4_XS",
   "quote": "M1 Max (64 GB),Qwen3-Coder-30B-A3B,IQ4_XS,ctx 4096,58.5 tok/s",
   "ram_gb": 16.1,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1rdx2c7/ran_3_popular_30b_moe_models_on_my_apple_silicon/",
   "tps": 58.5,
   "ttft_s": null
  },
  {
   "backend": "Ollama",
   "batch": null,
   "ctx": 128000,
   "date": null,
   "flags": [],
   "hardware": "M4 Max (40-core GPU, 64 GB)",
   "id": "siliconscore-m4-max-40-core-gpu-64-gb-qwen-3-32b-q4-k-medium-ctx128000",
   "model": "Qwen 3 32B",
   "notes": "source label: Silicon Score Lab; verified=True",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "community",
   "quant": "Q4_K - Medium",
   "quote": "M4 Max (40-core GPU, 64 GB),Qwen 3 32B,Q4_K - Medium,ctx 128000,22 tok/s",
   "ram_gb": 20,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://siliconscore.com/benchmarks.json",
   "tps": 22,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": 2048,
   "date": null,
   "flags": [],
   "hardware": "M4 Max (40-core GPU, 64 GB)",
   "id": "siliconscore-m4-max-40-core-gpu-64-gb-qwen-3-4b-q4-ctx2048",
   "model": "Qwen 3 4B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 2976.68,
   "provenance": "sourced",
   "quant": "Q4",
   "quote": "M4 Max (40-core GPU, 64 GB),Qwen 3 4B,Q4,ctx 2048,148.08 tok/s",
   "ram_gb": 2.54,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://gist.github.com/awni/c1790e4c3a39be6e8f1c4afd42423d2d",
   "tps": 148.08,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": 2048,
   "date": null,
   "flags": [],
   "hardware": "M4 Max (40-core GPU, 64 GB)",
   "id": "siliconscore-m4-max-40-core-gpu-64-gb-qwen-3-4b-q4-g32-ctx2048",
   "model": "Qwen 3 4B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 2838.42,
   "provenance": "sourced",
   "quant": "Q4_G32",
   "quote": "M4 Max (40-core GPU, 64 GB),Qwen 3 4B,Q4_G32,ctx 2048,149.07 tok/s",
   "ram_gb": 2.78,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://gist.github.com/awni/c1790e4c3a39be6e8f1c4afd42423d2d",
   "tps": 149.07,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": 2048,
   "date": null,
   "flags": [],
   "hardware": "M4 Max (40-core GPU, 64 GB)",
   "id": "siliconscore-m4-max-40-core-gpu-64-gb-qwen-3-4b-q5-ctx2048",
   "model": "Qwen 3 4B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 2736.15,
   "provenance": "sourced",
   "quant": "Q5",
   "quote": "M4 Max (40-core GPU, 64 GB),Qwen 3 4B,Q5,ctx 2048,143.2 tok/s",
   "ram_gb": 3.26,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://gist.github.com/awni/c1790e4c3a39be6e8f1c4afd42423d2d",
   "tps": 143.2,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": 2048,
   "date": null,
   "flags": [],
   "hardware": "M4 Max (40-core GPU, 64 GB)",
   "id": "siliconscore-m4-max-40-core-gpu-64-gb-qwen-3-4b-q5-g32-ctx2048",
   "model": "Qwen 3 4B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 2754.53,
   "provenance": "sourced",
   "quant": "Q5_G32",
   "quote": "M4 Max (40-core GPU, 64 GB),Qwen 3 4B,Q5_G32,ctx 2048,142.98 tok/s",
   "ram_gb": 3.5,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://gist.github.com/awni/c1790e4c3a39be6e8f1c4afd42423d2d",
   "tps": 142.98,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": 2048,
   "date": null,
   "flags": [],
   "hardware": "M4 Max (40-core GPU, 64 GB)",
   "id": "siliconscore-m4-max-40-core-gpu-64-gb-qwen-3-4b-q6-ctx2048",
   "model": "Qwen 3 4B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 2735.69,
   "provenance": "sourced",
   "quant": "Q6",
   "quote": "M4 Max (40-core GPU, 64 GB),Qwen 3 4B,Q6,ctx 2048,136.58 tok/s",
   "ram_gb": 3.98,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://gist.github.com/awni/c1790e4c3a39be6e8f1c4afd42423d2d",
   "tps": 136.58,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": 2048,
   "date": null,
   "flags": [],
   "hardware": "M4 Max (40-core GPU, 64 GB)",
   "id": "siliconscore-m4-max-40-core-gpu-64-gb-qwen-3-4b-q8-ctx2048",
   "model": "Qwen 3 4B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 1780.63,
   "provenance": "sourced",
   "quant": "Q8",
   "quote": "M4 Max (40-core GPU, 64 GB),Qwen 3 4B,Q8,ctx 2048,111.55 tok/s",
   "ram_gb": 5.06,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://gist.github.com/awni/c1790e4c3a39be6e8f1c4afd42423d2d",
   "tps": 111.55,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": 2048,
   "date": null,
   "flags": [],
   "hardware": "M4 Max (40-core GPU, 64 GB)",
   "id": "siliconscore-m4-max-40-core-gpu-64-gb-qwen-3-30b-a3b-q4-ctx2048",
   "model": "Qwen 3 30B-A3B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 822.59,
   "provenance": "sourced",
   "quant": "Q4",
   "quote": "M4 Max (40-core GPU, 64 GB),Qwen 3 30B-A3B,Q4,ctx 2048,92.09 tok/s",
   "ram_gb": 16.12,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://gist.github.com/awni/c1790e4c3a39be6e8f1c4afd42423d2d",
   "tps": 92.09,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": 2048,
   "date": null,
   "flags": [],
   "hardware": "M4 Max (40-core GPU, 64 GB)",
   "id": "siliconscore-m4-max-40-core-gpu-64-gb-qwen-3-30b-a3b-q5-ctx2048",
   "model": "Qwen 3 30B-A3B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 819.83,
   "provenance": "sourced",
   "quant": "Q5",
   "quote": "M4 Max (40-core GPU, 64 GB),Qwen 3 30B-A3B,Q5,ctx 2048,84.89 tok/s",
   "ram_gb": 18.09,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://gist.github.com/awni/c1790e4c3a39be6e8f1c4afd42423d2d",
   "tps": 84.89,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": 2048,
   "date": null,
   "flags": [],
   "hardware": "M4 Max (40-core GPU, 64 GB)",
   "id": "siliconscore-m4-max-40-core-gpu-64-gb-qwen-3-30b-a3b-q6-ctx2048",
   "model": "Qwen 3 30B-A3B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 817.63,
   "provenance": "sourced",
   "quant": "Q6",
   "quote": "M4 Max (40-core GPU, 64 GB),Qwen 3 30B-A3B,Q6,ctx 2048,76.74 tok/s",
   "ram_gb": 21.87,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://gist.github.com/awni/c1790e4c3a39be6e8f1c4afd42423d2d",
   "tps": 76.74,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": 2048,
   "date": null,
   "flags": [],
   "hardware": "M4 Max (40-core GPU, 64 GB)",
   "id": "siliconscore-m4-max-40-core-gpu-64-gb-qwen-3-30b-a3b-q8-ctx2048",
   "model": "Qwen 3 30B-A3B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 772.59,
   "provenance": "sourced",
   "quant": "Q8",
   "quote": "M4 Max (40-core GPU, 64 GB),Qwen 3 30B-A3B,Q8,ctx 2048,52.58 tok/s",
   "ram_gb": 29.78,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://gist.github.com/awni/c1790e4c3a39be6e8f1c4afd42423d2d",
   "tps": 52.58,
   "ttft_s": null
  },
  {
   "backend": "LM Studio",
   "batch": null,
   "ctx": 10000,
   "date": null,
   "flags": [],
   "hardware": "M4 Max (128 GB)",
   "id": "siliconscore-m4-max-128-gb-qwen-3-0-6b-q8-0-ctx10000",
   "model": "qwen-3-0-6b",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q8_0",
   "quote": "M4 Max (128 GB),qwen-3-0-6b,Q8_0,ctx 10000,184.45 tok/s; ttft 0.04s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://gist.github.com/estsauver/a70c929398479f3166f3d69bcededac3",
   "tps": 184.45,
   "ttft_s": 0.04
  },
  {
   "backend": "LM Studio",
   "batch": null,
   "ctx": 10000,
   "date": null,
   "flags": [],
   "hardware": "M4 Max (128 GB)",
   "id": "siliconscore-m4-max-128-gb-qwen-3-8b-q4-k-medium-ctx10000",
   "model": "Qwen 3 8B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Max (128 GB),Qwen 3 8B,Q4_K - Medium,ctx 10000,63.15 tok/s; ttft 0.14s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://gist.github.com/estsauver/a70c929398479f3166f3d69bcededac3",
   "tps": 63.15,
   "ttft_s": 0.14
  },
  {
   "backend": "LM Studio",
   "batch": null,
   "ctx": 10000,
   "date": null,
   "flags": [],
   "hardware": "M4 Max (128 GB)",
   "id": "siliconscore-m4-max-128-gb-qwen-3-30b-a3b-q4-k-medium-ctx10000",
   "model": "Qwen 3 30B-A3B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Max (128 GB),Qwen 3 30B-A3B,Q4_K - Medium,ctx 10000,70.15 tok/s; ttft 0.62s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://gist.github.com/estsauver/a70c929398479f3166f3d69bcededac3",
   "tps": 70.15,
   "ttft_s": 0.62
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Pro (48 GB)",
   "id": "siliconscore-m4-pro-48-gb-qwen-3-30b-a3b-8bit-ctxnone",
   "model": "Qwen 3 30B-A3B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "8bit",
   "quote": "M4 Pro (48 GB),Qwen 3 30B-A3B,8bit,ctx None,55 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1m6tf9v/m4_pro_owners_i_want_your_biased_hottakes/",
   "tps": 55,
   "ttft_s": null
  },
  {
   "backend": "LM Studio",
   "batch": null,
   "ctx": 131072,
   "date": null,
   "flags": [],
   "hardware": "M4 Max (128 GB)",
   "id": "siliconscore-m4-max-128-gb-gemma-3-27b-q8-0-ctx131072",
   "model": "Gemma 3 27B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q8_0",
   "quote": "M4 Max (128 GB),Gemma 3 27B,Q8_0,ctx 131072,14.49 tok/s; ttft 1.71s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://gist.github.com/estsauver/a70c929398479f3166f3d69bcededac3",
   "tps": 14.49,
   "ttft_s": 1.71
  },
  {
   "backend": "LM Studio",
   "batch": null,
   "ctx": 10000,
   "date": null,
   "flags": [],
   "hardware": "M4 Max (128 GB)",
   "id": "siliconscore-m4-max-128-gb-gemma-3-27b-q8-0-ctx10000",
   "model": "Gemma 3 27B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q8_0",
   "quote": "M4 Max (128 GB),Gemma 3 27B,Q8_0,ctx 10000,12.99 tok/s; ttft 0.79s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://gist.github.com/estsauver/a70c929398479f3166f3d69bcededac3",
   "tps": 12.99,
   "ttft_s": 0.79
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 512,
   "date": null,
   "flags": [],
   "hardware": "M4 (32 GB)",
   "id": "siliconscore-m4-32-gb-gemma-3-27b-q4-0-ctx512",
   "model": "Gemma 3 27B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 47.51,
   "provenance": "sourced",
   "quant": "Q4_0",
   "quote": "M4 (32 GB),Gemma 3 27B,Q4_0,ctx 512,5.72 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://gist.github.com/zachrattner/52a6b56d70ed024b18c992ef14b89656",
   "tps": 5.72,
   "ttft_s": null
  },
  {
   "backend": "LM Studio",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "M4 Max (128 GB)",
   "id": "siliconscore-m4-max-128-gb-gemma-3-4b-q4-0-ctx4096",
   "model": "Gemma 3 4B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_0",
   "quote": "M4 Max (128 GB),Gemma 3 4B,Q4_0,ctx 4096,100.54 tok/s; ttft 0.14s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://gist.github.com/estsauver/a70c929398479f3166f3d69bcededac3",
   "tps": 100.54,
   "ttft_s": 0.14
  },
  {
   "backend": "LM Studio",
   "batch": null,
   "ctx": 10000,
   "date": null,
   "flags": [],
   "hardware": "M4 Max (128 GB)",
   "id": "siliconscore-m4-max-128-gb-qwen-2-5-7b-instruct-q8-0-ctx10000",
   "model": "qwen-2-5-7b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q8_0",
   "quote": "M4 Max (128 GB),qwen-2-5-7b-instruct,Q8_0,ctx 10000,49.67 tok/s; ttft 0.26s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://gist.github.com/estsauver/a70c929398479f3166f3d69bcededac3",
   "tps": 49.67,
   "ttft_s": 0.26
  },
  {
   "backend": "LM Studio",
   "batch": null,
   "ctx": 10000,
   "date": null,
   "flags": [],
   "hardware": "M4 Max (128 GB)",
   "id": "siliconscore-m4-max-128-gb-qwen-3-235b-a22b-q4-k-medium-ctx10000",
   "model": "Qwen 3 235B-A22B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Max (128 GB),Qwen 3 235B-A22B,Q4_K - Medium,ctx 10000,8.08 tok/s; ttft 1.87s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://gist.github.com/estsauver/a70c929398479f3166f3d69bcededac3",
   "tps": 8.08,
   "ttft_s": 1.87
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": 1400,
   "date": null,
   "flags": [],
   "hardware": "M4 Max (128 GB)",
   "id": "siliconscore-m4-max-128-gb-qwen-3-235b-a22b-3bit-ctx1400",
   "model": "Qwen 3 235B-A22B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "3bit",
   "quote": "M4 Max (128 GB),Qwen 3 235B-A22B,3bit,ctx 1400,30 tok/s; ttft 14s",
   "ram_gb": 100,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1kfmoyx/qwen3_235b_pairs_extremely_well_with_a_macbook/",
   "tps": 30,
   "ttft_s": 14
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M2 Ultra (192 GB)",
   "id": "siliconscore-m2-ultra-192-gb-qwen-3-235b-a22b-q4-ctxnone",
   "model": "Qwen 3 235B-A22B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4",
   "quote": "M2 Ultra (192 GB),Qwen 3 235B-A22B,Q4,ctx None,26.4 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1kr9d9b/how_are_you_running_qwen3235b_locally/",
   "tps": 26.4,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 Ultra (256 GB)",
   "id": "siliconscore-m3-ultra-256-gb-qwen-3-235b-a22b-q4-ctxnone",
   "model": "Qwen 3 235B-A22B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4",
   "quote": "M3 Ultra (256 GB),Qwen 3 235B-A22B,Q4,ctx None,27 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1kr9d9b/how_are_you_running_qwen3235b_locally/",
   "tps": 27,
   "ttft_s": null
  },
  {
   "backend": "LM Studio",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 Ultra (512 GB)",
   "id": "siliconscore-m3-ultra-512-gb-qwen-3-235b-a22b-4bit-ctxnone",
   "model": "Qwen 3 235B-A22B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "4bit",
   "quote": "M3 Ultra (512 GB),Qwen 3 235B-A22B,4bit,ctx None,27.36 tok/s; ttft 1.73s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1lumsd2/mac_studio_512gb_online/",
   "tps": 27.36,
   "ttft_s": 1.73
  },
  {
   "backend": "LM Studio",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Max (128 GB)",
   "id": "siliconscore-m4-max-128-gb-llama-3-3-70b-4bit-ctxnone",
   "model": "Llama 3.3 70B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "4bit",
   "quote": "M4 Max (128 GB),Llama 3.3 70B,4bit,ctx None,11.8 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1i7b3r1/i_did_a_quick_test_of_macbook_m4_max_128_gb/",
   "tps": 11.8,
   "ttft_s": null
  },
  {
   "backend": "LM Studio",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Max (128 GB)",
   "id": "siliconscore-m4-max-128-gb-llama-3-3-70b-8bit-ctxnone",
   "model": "Llama 3.3 70B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "8bit",
   "quote": "M4 Max (128 GB),Llama 3.3 70B,8bit,ctx None,6.5 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1i7b3r1/i_did_a_quick_test_of_macbook_m4_max_128_gb/",
   "tps": 6.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 35550,
   "date": null,
   "flags": [],
   "hardware": "M4 Max (128 GB)",
   "id": "siliconscore-m4-max-128-gb-llama-3-3-70b-q8-0-ctx35550",
   "model": "Llama 3.3 70B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 58,
   "provenance": "sourced",
   "quant": "Q8_0",
   "quote": "M4 Max (128 GB),Llama 3.3 70B,Q8_0,ctx 35550,4.5 tok/s; ttft 600s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1hax4ue/llama_32_3b_and_llama_33_70b_models_on_a_mac_mini/",
   "tps": 4.5,
   "ttft_s": 600
  },
  {
   "backend": "LM Studio",
   "batch": null,
   "ctx": 7800,
   "date": null,
   "flags": [],
   "hardware": "M3 Ultra (512 GB)",
   "id": "siliconscore-m3-ultra-512-gb-llama-3-3-70b-4bit-ctx7800",
   "model": "Llama 3.3 70B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 150,
   "provenance": "sourced",
   "quant": "4bit",
   "quote": "M3 Ultra (512 GB),Llama 3.3 70B,4bit,ctx 7800,15.5 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1jdvk7c/any_m3_ultra_test_requests_for_mlx_models_in_lm/",
   "tps": 15.5,
   "ttft_s": null
  },
  {
   "backend": "LM Studio",
   "batch": null,
   "ctx": 7800,
   "date": null,
   "flags": [],
   "hardware": "M3 Ultra (512 GB)",
   "id": "siliconscore-m3-ultra-512-gb-llama-3-3-70b-8bit-ctx7800",
   "model": "Llama 3.3 70B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 150,
   "provenance": "sourced",
   "quant": "8bit",
   "quote": "M3 Ultra (512 GB),Llama 3.3 70B,8bit,ctx 7800,8.5 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1jdvk7c/any_m3_ultra_test_requests_for_mlx_models_in_lm/",
   "tps": 8.5,
   "ttft_s": null
  },
  {
   "backend": "LM Studio",
   "batch": null,
   "ctx": 39686,
   "date": null,
   "flags": [],
   "hardware": "M3 Ultra (512 GB)",
   "id": "siliconscore-m3-ultra-512-gb-llama-3-3-70b-4bit-ctx39686",
   "model": "Llama 3.3 70B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 103,
   "provenance": "sourced",
   "quant": "4bit",
   "quote": "M3 Ultra (512 GB),Llama 3.3 70B,4bit,ctx 39686,9.57 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1jdvk7c/any_m3_ultra_test_requests_for_mlx_models_in_lm/",
   "tps": 9.57,
   "ttft_s": null
  },
  {
   "backend": "LM Studio",
   "batch": null,
   "ctx": 39686,
   "date": null,
   "flags": [],
   "hardware": "M3 Ultra (512 GB)",
   "id": "siliconscore-m3-ultra-512-gb-llama-3-3-70b-8bit-ctx39686",
   "model": "Llama 3.3 70B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 101,
   "provenance": "sourced",
   "quant": "8bit",
   "quote": "M3 Ultra (512 GB),Llama 3.3 70B,8bit,ctx 39686,6.53 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1jdvk7c/any_m3_ultra_test_requests_for_mlx_models_in_lm/",
   "tps": 6.53,
   "ttft_s": null
  },
  {
   "backend": "LM Studio",
   "batch": null,
   "ctx": 3991,
   "date": null,
   "flags": [],
   "hardware": "M1 Ultra (64 GB)",
   "id": "siliconscore-m1-ultra-64-gb-llama-3-3-70b-4bit-ctx3991",
   "model": "Llama 3.3 70B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "4bit",
   "quote": "M1 Ultra (64 GB),Llama 3.3 70B,4bit,ctx 3991,12.55 tok/s; ttft 49.8s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1he2v2n/speed_test_llama3370b_on_2xrtx3090_vs_m3max_64gb/",
   "tps": 12.55,
   "ttft_s": 49.8
  },
  {
   "backend": "LM Studio",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M2 Max (38-core GPU, 64 GB)",
   "id": "siliconscore-m2-max-38-core-gpu-64-gb-llama-3-3-70b-4bit-ctxnone",
   "model": "Llama 3.3 70B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "4bit",
   "quote": "M2 Max (38-core GPU, 64 GB),Llama 3.3 70B,4bit,ctx None,8.8 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1hvwwsq/exolab_nvidias_digits_outperforms_apples_m4_chips/",
   "tps": 8.8,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 258,
   "date": null,
   "flags": [],
   "hardware": "M3 Max (GPU count not published, 64 GB)",
   "id": "siliconscore-m3-max-gpu-count-not-published-64-gb-llama-3-3-70b-q4-k-medium-ctx258",
   "model": "Llama 3.3 70B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 67.86,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M3 Max (GPU count not published, 64 GB),Llama 3.3 70B,Q4_K - Medium,ctx 258,8.15 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1he2v2n/speed_test_llama3370b_on_2xrtx3090_vs_m3max_64gb/",
   "tps": 8.15,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 8013,
   "date": null,
   "flags": [],
   "hardware": "M3 Max (GPU count not published, 64 GB)",
   "id": "siliconscore-m3-max-gpu-count-not-published-64-gb-llama-3-3-70b-q4-k-medium-ctx8013",
   "model": "Llama 3.3 70B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 65.17,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M3 Max (GPU count not published, 64 GB),Llama 3.3 70B,Q4_K - Medium,ctx 8013,7.48 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1he2v2n/speed_test_llama3370b_on_2xrtx3090_vs_m3max_64gb/",
   "tps": 7.48,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16001,
   "date": null,
   "flags": [],
   "hardware": "M3 Max (GPU count not published, 64 GB)",
   "id": "siliconscore-m3-max-gpu-count-not-published-64-gb-llama-3-3-70b-q4-k-medium-ctx16001",
   "model": "Llama 3.3 70B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 59.5,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M3 Max (GPU count not published, 64 GB),Llama 3.3 70B,Q4_K - Medium,ctx 16001,7 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1he2v2n/speed_test_llama3370b_on_2xrtx3090_vs_m3max_64gb/",
   "tps": 7,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 32170,
   "date": null,
   "flags": [],
   "hardware": "M3 Max (GPU count not published, 64 GB)",
   "id": "siliconscore-m3-max-gpu-count-not-published-64-gb-llama-3-3-70b-q4-k-medium-ctx32170",
   "model": "Llama 3.3 70B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 50.32,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M3 Max (GPU count not published, 64 GB),Llama 3.3 70B,Q4_K - Medium,ctx 32170,6.13 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1he2v2n/speed_test_llama3370b_on_2xrtx3090_vs_m3max_64gb/",
   "tps": 6.13,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 264,
   "date": null,
   "flags": [],
   "hardware": "M3 Max (GPU count not published, 64 GB)",
   "id": "siliconscore-m3-max-gpu-count-not-published-64-gb-qwen-3-32b-q8-0-ctx264",
   "model": "Qwen 3 32B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 153.63,
   "provenance": "sourced",
   "quant": "Q8_0",
   "quote": "M3 Max (GPU count not published, 64 GB),Qwen 3 32B,Q8_0,ctx 264,10.41 tok/s; ttft 1.72s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1kk0ghi/speed_comparison_with_qwen332bq8_0_ollama/",
   "tps": 10.41,
   "ttft_s": 1.72
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 450,
   "date": null,
   "flags": [],
   "hardware": "M3 Max (GPU count not published, 64 GB)",
   "id": "siliconscore-m3-max-gpu-count-not-published-64-gb-qwen-3-32b-q8-0-ctx450",
   "model": "Qwen 3 32B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 171.37,
   "provenance": "sourced",
   "quant": "Q8_0",
   "quote": "M3 Max (GPU count not published, 64 GB),Qwen 3 32B,Q8_0,ctx 450,10.28 tok/s; ttft 2.63s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1kk0ghi/speed_comparison_with_qwen332bq8_0_ollama/",
   "tps": 10.28,
   "ttft_s": 2.63
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 723,
   "date": null,
   "flags": [],
   "hardware": "M3 Max (GPU count not published, 64 GB)",
   "id": "siliconscore-m3-max-gpu-count-not-published-64-gb-qwen-3-32b-q8-0-ctx723",
   "model": "Qwen 3 32B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 164.83,
   "provenance": "sourced",
   "quant": "Q8_0",
   "quote": "M3 Max (GPU count not published, 64 GB),Qwen 3 32B,Q8_0,ctx 723,10.29 tok/s; ttft 4.39s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1kk0ghi/speed_comparison_with_qwen332bq8_0_ollama/",
   "tps": 10.29,
   "ttft_s": 4.39
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 1219,
   "date": null,
   "flags": [],
   "hardware": "M3 Max (GPU count not published, 64 GB)",
   "id": "siliconscore-m3-max-gpu-count-not-published-64-gb-qwen-3-32b-q8-0-ctx1219",
   "model": "Qwen 3 32B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 169.15,
   "provenance": "sourced",
   "quant": "Q8_0",
   "quote": "M3 Max (GPU count not published, 64 GB),Qwen 3 32B,Q8_0,ctx 1219,10.19 tok/s; ttft 7.21s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1kk0ghi/speed_comparison_with_qwen332bq8_0_ollama/",
   "tps": 10.19,
   "ttft_s": 7.21
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 1858,
   "date": null,
   "flags": [],
   "hardware": "M3 Max (GPU count not published, 64 GB)",
   "id": "siliconscore-m3-max-gpu-count-not-published-64-gb-qwen-3-32b-q8-0-ctx1858",
   "model": "Qwen 3 32B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 166.81,
   "provenance": "sourced",
   "quant": "Q8_0",
   "quote": "M3 Max (GPU count not published, 64 GB),Qwen 3 32B,Q8_0,ctx 1858,10.09 tok/s; ttft 11.14s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1kk0ghi/speed_comparison_with_qwen332bq8_0_ollama/",
   "tps": 10.09,
   "ttft_s": 11.14
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 2979,
   "date": null,
   "flags": [],
   "hardware": "M3 Max (GPU count not published, 64 GB)",
   "id": "siliconscore-m3-max-gpu-count-not-published-64-gb-qwen-3-32b-q8-0-ctx2979",
   "model": "Qwen 3 32B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 162.22,
   "provenance": "sourced",
   "quant": "Q8_0",
   "quote": "M3 Max (GPU count not published, 64 GB),Qwen 3 32B,Q8_0,ctx 2979,9.89 tok/s; ttft 18.36s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1kk0ghi/speed_comparison_with_qwen332bq8_0_ollama/",
   "tps": 9.89,
   "ttft_s": 18.36
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 4669,
   "date": null,
   "flags": [],
   "hardware": "M3 Max (GPU count not published, 64 GB)",
   "id": "siliconscore-m3-max-gpu-count-not-published-64-gb-qwen-3-32b-q8-0-ctx4669",
   "model": "Qwen 3 32B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 154.16,
   "provenance": "sourced",
   "quant": "Q8_0",
   "quote": "M3 Max (GPU count not published, 64 GB),Qwen 3 32B,Q8_0,ctx 4669,9.67 tok/s; ttft 30.29s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1kk0ghi/speed_comparison_with_qwen332bq8_0_ollama/",
   "tps": 9.67,
   "ttft_s": 30.29
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 7948,
   "date": null,
   "flags": [],
   "hardware": "M3 Max (GPU count not published, 64 GB)",
   "id": "siliconscore-m3-max-gpu-count-not-published-64-gb-qwen-3-32b-q8-0-ctx7948",
   "model": "Qwen 3 32B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 140.11,
   "provenance": "sourced",
   "quant": "Q8_0",
   "quote": "M3 Max (GPU count not published, 64 GB),Qwen 3 32B,Q8_0,ctx 7948,9.2 tok/s; ttft 56.73s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1kk0ghi/speed_comparison_with_qwen332bq8_0_ollama/",
   "tps": 9.2,
   "ttft_s": 56.73
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 12416,
   "date": null,
   "flags": [],
   "hardware": "M3 Max (GPU count not published, 64 GB)",
   "id": "siliconscore-m3-max-gpu-count-not-published-64-gb-qwen-3-32b-q8-0-ctx12416",
   "model": "Qwen 3 32B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 127.96,
   "provenance": "sourced",
   "quant": "Q8_0",
   "quote": "M3 Max (GPU count not published, 64 GB),Qwen 3 32B,Q8_0,ctx 12416,8.6 tok/s; ttft 97.03s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1kk0ghi/speed_comparison_with_qwen332bq8_0_ollama/",
   "tps": 8.6,
   "ttft_s": 97.03
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 20172,
   "date": null,
   "flags": [],
   "hardware": "M3 Max (GPU count not published, 64 GB)",
   "id": "siliconscore-m3-max-gpu-count-not-published-64-gb-qwen-3-32b-q8-0-ctx20172",
   "model": "Qwen 3 32B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 111.18,
   "provenance": "sourced",
   "quant": "Q8_0",
   "quote": "M3 Max (GPU count not published, 64 GB),Qwen 3 32B,Q8_0,ctx 20172,7.58 tok/s; ttft 181.44s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1kk0ghi/speed_comparison_with_qwen332bq8_0_ollama/",
   "tps": 7.58,
   "ttft_s": 181.44
  },
  {
   "backend": "Ollama",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Pro (64 GB)",
   "id": "siliconscore-m4-pro-64-gb-llama-3-3-70b-q4-k-medium-ctxnone",
   "model": "Llama 3.3 70B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Pro (64 GB),Llama 3.3 70B,Q4_K - Medium,ctx None,5 tok/s; ttft 3.75s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1hax4ue/llama_32_3b_and_llama_33_70b_models_on_a_mac_mini/",
   "tps": 5,
   "ttft_s": 3.75
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 512,
   "date": null,
   "flags": [],
   "hardware": "M1 Pro (16-core GPU)",
   "id": "siliconscore-m1-pro-16-core-gpu-llama-2-7b-q4-0-ctx512",
   "model": "Llama 2 7B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 266.25,
   "provenance": "sourced",
   "quant": "Q4_0",
   "quote": "M1 Pro (16-core GPU),Llama 2 7B,Q4_0,ctx 512,36.41 tok/s",
   "ram_gb": 3.56,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tps": 36.41,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 512,
   "date": null,
   "flags": [],
   "hardware": "M2 Ultra (76-core GPU, 192 GB)",
   "id": "siliconscore-m2-ultra-76-core-gpu-192-gb-llama-2-7b-q4-0-ctx512",
   "model": "Llama 2 7B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 1238.48,
   "provenance": "sourced",
   "quant": "Q4_0",
   "quote": "M2 Ultra (76-core GPU, 192 GB),Llama 2 7B,Q4_0,ctx 512,94.27 tok/s",
   "ram_gb": 3.56,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tps": 94.27,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 512,
   "date": null,
   "flags": [],
   "hardware": "M3 Pro (18-core GPU)",
   "id": "siliconscore-m3-pro-18-core-gpu-llama-2-7b-q4-0-ctx512",
   "model": "Llama 2 7B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 341.67,
   "provenance": "sourced",
   "quant": "Q4_0",
   "quote": "M3 Pro (18-core GPU),Llama 2 7B,Q4_0,ctx 512,30.74 tok/s",
   "ram_gb": 3.56,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tps": 30.74,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 512,
   "date": null,
   "flags": [],
   "hardware": "M3 Max (40-core GPU, 48 GB)",
   "id": "siliconscore-m3-max-40-core-gpu-48-gb-llama-2-7b-q4-0-ctx512",
   "model": "Llama 2 7B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 690.99,
   "provenance": "sourced",
   "quant": "Q4_0",
   "quote": "M3 Max (40-core GPU, 48 GB),Llama 2 7B,Q4_0,ctx 512,65.85 tok/s",
   "ram_gb": 3.56,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tps": 65.85,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 512,
   "date": null,
   "flags": [],
   "hardware": "M4 (10-core GPU, 16 GB)",
   "id": "siliconscore-m4-10-core-gpu-16-gb-llama-2-7b-q4-0-ctx512",
   "model": "Llama 2 7B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 221.29,
   "provenance": "sourced",
   "quant": "Q4_0",
   "quote": "M4 (10-core GPU, 16 GB),Llama 2 7B,Q4_0,ctx 512,24.11 tok/s",
   "ram_gb": 3.56,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "tps": 24.11,
   "ttft_s": null
  },
  {
   "backend": "LM Studio",
   "batch": null,
   "ctx": 10000,
   "date": null,
   "flags": [],
   "hardware": "M4 Max (128 GB)",
   "id": "siliconscore-m4-max-128-gb-qwen-3-32b-q4-k-medium-ctx10000",
   "model": "Qwen 3 32B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Max (128 GB),Qwen 3 32B,Q4_K - Medium,ctx 10000,11.72 tok/s; ttft 0.81s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://gist.github.com/estsauver/a70c929398479f3166f3d69bcededac3",
   "tps": 11.72,
   "ttft_s": 0.81
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 (7-core GPU, 8 GB)",
   "id": "siliconscore-m1-7-core-gpu-8-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 108.71905633333331,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M1 (7-core GPU, 8 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,13.385236666666666 tok/s; ttft 11.513906754444445s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/1281",
   "tps": 13.385236666666666,
   "ttft_s": 11.513906754444445
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 (7-core GPU, 8 GB)",
   "id": "siliconscore-m1-7-core-gpu-8-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 532.9676827037036,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M1 (7-core GPU, 8 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,38.51773918518518 tok/s; ttft 2.4294918625925925s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/1281",
   "tps": 38.51773918518518,
   "ttft_s": 2.4294918625925925
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 (7-core GPU, 16 GB)",
   "id": "siliconscore-m1-7-core-gpu-16-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 83.93792277777777,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M1 (7-core GPU, 16 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,9.387483333333332 tok/s; ttft 14.82420303238889s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/223",
   "tps": 9.387483333333332,
   "ttft_s": 14.82420303238889
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 (7-core GPU, 16 GB)",
   "id": "siliconscore-m1-7-core-gpu-16-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 530.8859519722223,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M1 (7-core GPU, 16 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,37.872518722222225 tok/s; ttft 2.4528570579444446s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/223",
   "tps": 37.872518722222225,
   "ttft_s": 2.4528570579444446
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 (7-core GPU, 16 GB)",
   "id": "siliconscore-m1-7-core-gpu-16-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 40.612479388888886,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M1 (7-core GPU, 16 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,4.806673055555556 tok/s; ttft 30.980614497666668s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/223",
   "tps": 4.806673055555556,
   "ttft_s": 30.980614497666668
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 (8-core GPU, 8 GB)",
   "id": "siliconscore-m1-8-core-gpu-8-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 133.8709728888889,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M1 (8-core GPU, 8 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,14.552260666666667 tok/s; ttft 9.648145671333332s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/17",
   "tps": 14.552260666666667,
   "ttft_s": 9.648145671333332
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 (8-core GPU, 8 GB)",
   "id": "siliconscore-m1-8-core-gpu-8-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 607.4231519333333,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M1 (8-core GPU, 8 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,40.41441422222222 tok/s; ttft 2.1677008980666663s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/17",
   "tps": 40.41441422222222,
   "ttft_s": 2.1677008980666663
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 (8-core GPU, 16 GB)",
   "id": "siliconscore-m1-8-core-gpu-16-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 585.7259774074074,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M1 (8-core GPU, 16 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,40.18323162962963 tok/s; ttft 2.24642097837037s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/432",
   "tps": 40.18323162962963,
   "ttft_s": 2.24642097837037
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 (8-core GPU, 16 GB)",
   "id": "siliconscore-m1-8-core-gpu-16-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 53.26808444444444,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M1 (8-core GPU, 16 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,5.429473666666667 tok/s; ttft 24.55238074088889s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/432",
   "tps": 5.429473666666667,
   "ttft_s": 24.55238074088889
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 Max (24-core GPU, 32 GB)",
   "id": "siliconscore-m1-max-24-core-gpu-32-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 1582.032366,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M1 Max (24-core GPU, 32 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,93.859194 tok/s; ttft 0.7243009050000001s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/1943",
   "tps": 93.859194,
   "ttft_s": 0.7243009050000001
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 Max (24-core GPU, 32 GB)",
   "id": "siliconscore-m1-max-24-core-gpu-32-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 155.93443555555555,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M1 Max (24-core GPU, 32 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,17.355148333333332 tok/s; ttft 7.862980407555557s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/1943",
   "tps": 17.355148333333332,
   "ttft_s": 7.862980407555557
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 Max (24-core GPU, 64 GB)",
   "id": "siliconscore-m1-max-24-core-gpu-64-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 306.18835666666666,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M1 Max (24-core GPU, 64 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,32.11692033333334 tok/s; ttft 3.9939120186666677s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/688",
   "tps": 32.11692033333334,
   "ttft_s": 3.9939120186666677
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 Max (24-core GPU, 64 GB)",
   "id": "siliconscore-m1-max-24-core-gpu-64-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 1669.4850333333334,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M1 Max (24-core GPU, 64 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,105.66463296296295 tok/s; ttft 0.6893040093333332s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/688",
   "tps": 105.66463296296295,
   "ttft_s": 0.6893040093333332
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 Max (24-core GPU, 64 GB)",
   "id": "siliconscore-m1-max-24-core-gpu-64-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 140.19277200000002,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M1 Max (24-core GPU, 64 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,15.135508527777777 tok/s; ttft 8.830140840277778s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/688",
   "tps": 15.135508527777777,
   "ttft_s": 8.830140840277778
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 Max (32-core GPU, 32 GB)",
   "id": "siliconscore-m1-max-32-core-gpu-32-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 359.8733261944444,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M1 Max (32-core GPU, 32 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,35.40364952777778 tok/s; ttft 3.4528548829722223s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/597",
   "tps": 35.40364952777778,
   "ttft_s": 3.4528548829722223
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 Max (32-core GPU, 32 GB)",
   "id": "siliconscore-m1-max-32-core-gpu-32-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 2025.1921614074072,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M1 Max (32-core GPU, 32 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,125.77495338888887 tok/s; ttft 0.5596487252407406s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/597",
   "tps": 125.77495338888887,
   "ttft_s": 0.5596487252407406
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 Max (32-core GPU, 32 GB)",
   "id": "siliconscore-m1-max-32-core-gpu-32-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 195.8191797407407,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M1 Max (32-core GPU, 32 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,20.07296611111111 tok/s; ttft 6.301562639796296s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/597",
   "tps": 20.07296611111111,
   "ttft_s": 6.301562639796296
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 Max (32-core GPU, 64 GB)",
   "id": "siliconscore-m1-max-32-core-gpu-64-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 380.7280225277778,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M1 Max (32-core GPU, 64 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,37.75919747222222 tok/s; ttft 3.12193795138889s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/364",
   "tps": 37.75919747222222,
   "ttft_s": 3.12193795138889
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 Max (32-core GPU, 64 GB)",
   "id": "siliconscore-m1-max-32-core-gpu-64-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 1972.074889361111,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M1 Max (32-core GPU, 64 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,120.67175636111112 tok/s; ttft 0.5782512951666667s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/364",
   "tps": 120.67175636111112,
   "ttft_s": 0.5782512951666667
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 Max (32-core GPU, 64 GB)",
   "id": "siliconscore-m1-max-32-core-gpu-64-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 185.93018677777775,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M1 Max (32-core GPU, 64 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,19.01644427777778 tok/s; ttft 6.8818776490777775s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/364",
   "tps": 19.01644427777778,
   "ttft_s": 6.8818776490777775
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 Pro (14-core GPU, 16 GB)",
   "id": "siliconscore-m1-pro-14-core-gpu-16-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 176.96636692592594,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M1 Pro (14-core GPU, 16 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,20.14017855555555 tok/s; ttft 6.991519057185186s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/412",
   "tps": 20.14017855555555,
   "ttft_s": 6.991519057185186
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 Pro (14-core GPU, 16 GB)",
   "id": "siliconscore-m1-pro-14-core-gpu-16-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 1040.067598027778,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M1 Pro (14-core GPU, 16 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,71.81051661111111 tok/s; ttft 1.1483909050555556s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/412",
   "tps": 71.81051661111111,
   "ttft_s": 1.1483909050555556
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 Pro (14-core GPU, 16 GB)",
   "id": "siliconscore-m1-pro-14-core-gpu-16-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 92.4920569777778,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M1 Pro (14-core GPU, 16 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,10.8301478 tok/s; ttft 13.4879233686s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/412",
   "tps": 10.8301478,
   "ttft_s": 13.4879233686
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 Pro (14-core GPU, 32 GB)",
   "id": "siliconscore-m1-pro-14-core-gpu-32-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 173.08191844444443,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M1 Pro (14-core GPU, 32 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,19.971878777777775 tok/s; ttft 7.230578513999999s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/418",
   "tps": 19.971878777777775,
   "ttft_s": 7.230578513999999
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 Pro (14-core GPU, 32 GB)",
   "id": "siliconscore-m1-pro-14-core-gpu-32-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 1063.8479185555557,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M1 Pro (14-core GPU, 32 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,71.01946466666666 tok/s; ttft 1.1336526296666667s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/418",
   "tps": 71.01946466666666,
   "ttft_s": 1.1336526296666667
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 Pro (14-core GPU, 32 GB)",
   "id": "siliconscore-m1-pro-14-core-gpu-32-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 88.86977233333333,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M1 Pro (14-core GPU, 32 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,10.399665111111112 tok/s; ttft 14.166348648222224s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/1602",
   "tps": 10.399665111111112,
   "ttft_s": 14.166348648222224
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 Pro (16-core GPU, 16 GB)",
   "id": "siliconscore-m1-pro-16-core-gpu-16-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 204.5403022222222,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M1 Pro (16-core GPU, 16 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,21.88205522222222 tok/s; ttft 6.089909109518518s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/929",
   "tps": 21.88205522222222,
   "ttft_s": 6.089909109518518
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 Pro (16-core GPU, 16 GB)",
   "id": "siliconscore-m1-pro-16-core-gpu-16-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 1158.147661425926,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M1 Pro (16-core GPU, 16 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,78.15808337037036 tok/s; ttft 1.0274150409074074s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/929",
   "tps": 78.15808337037036,
   "ttft_s": 1.0274150409074074
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 Pro (16-core GPU, 16 GB)",
   "id": "siliconscore-m1-pro-16-core-gpu-16-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 106.6591231111111,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M1 Pro (16-core GPU, 16 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,11.90798088888889 tok/s; ttft 11.83519856488889s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/929",
   "tps": 11.90798088888889,
   "ttft_s": 11.83519856488889
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 Pro (16-core GPU, 32 GB)",
   "id": "siliconscore-m1-pro-16-core-gpu-32-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 200.28393007407408,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M1 Pro (16-core GPU, 32 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,21.698664740740742 tok/s; ttft 6.2272433256296305s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/4",
   "tps": 21.698664740740742,
   "ttft_s": 6.2272433256296305
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 Pro (16-core GPU, 32 GB)",
   "id": "siliconscore-m1-pro-16-core-gpu-32-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 1165.9555193555557,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M1 Pro (16-core GPU, 32 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,77.23935995555556 tok/s; ttft 1.0365038463555556s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/4",
   "tps": 77.23935995555556,
   "ttft_s": 1.0365038463555556
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 Pro (16-core GPU, 32 GB)",
   "id": "siliconscore-m1-pro-16-core-gpu-32-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 104.4587133888889,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M1 Pro (16-core GPU, 32 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,11.619242305555556 tok/s; ttft 12.083205306861112s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/4",
   "tps": 11.619242305555556,
   "ttft_s": 12.083205306861112
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 Ultra (48-core GPU, 128 GB)",
   "id": "siliconscore-m1-ultra-48-core-gpu-128-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 534.3503048148149,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M1 Ultra (48-core GPU, 128 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,48.89361162962964 tok/s; ttft 2.1556520740370373s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/306",
   "tps": 48.89361162962964,
   "ttft_s": 2.1556520740370373
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 Ultra (48-core GPU, 128 GB)",
   "id": "siliconscore-m1-ultra-48-core-gpu-128-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 2582.382230222222,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M1 Ultra (48-core GPU, 128 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,137.97023270370372 tok/s; ttft 0.4251138565185186s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/306",
   "tps": 137.97023270370372,
   "ttft_s": 0.4251138565185186
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 Ultra (48-core GPU, 128 GB)",
   "id": "siliconscore-m1-ultra-48-core-gpu-128-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 289.8969081111111,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M1 Ultra (48-core GPU, 128 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,27.773420277777777 tok/s; ttft 4.044066250166667s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/306",
   "tps": 27.773420277777777,
   "ttft_s": 4.044066250166667
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 Ultra (64-core GPU, 128 GB)",
   "id": "siliconscore-m1-ultra-64-core-gpu-128-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 667.5689454444446,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M1 Ultra (64-core GPU, 128 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,54.30790400000001 tok/s; ttft 1.6929770138888889s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/1087",
   "tps": 54.30790400000001,
   "ttft_s": 1.6929770138888889
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 Ultra (64-core GPU, 128 GB)",
   "id": "siliconscore-m1-ultra-64-core-gpu-128-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 2977.086063277778,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M1 Ultra (64-core GPU, 128 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,151.0678697777778 tok/s; ttft 0.3675812592777778s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/1087",
   "tps": 151.0678697777778,
   "ttft_s": 0.3675812592777778
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 Ultra (64-core GPU, 128 GB)",
   "id": "siliconscore-m1-ultra-64-core-gpu-128-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 371.7215576111111,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M1 Ultra (64-core GPU, 128 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,32.417089222222224 tok/s; ttft 3.126813851888889s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/1087",
   "tps": 32.417089222222224,
   "ttft_s": 3.126813851888889
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 Ultra (GPU count not published, 128 GB)",
   "id": "siliconscore-m1-ultra-gpu-count-not-published-128-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 71.55644355555556,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M1 Ultra (GPU count not published, 128 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,15.24847544444444 tok/s; ttft 17.47675613911111s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/307",
   "tps": 15.24847544444444,
   "ttft_s": 17.47675613911111
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 Ultra (GPU count not published, 128 GB)",
   "id": "siliconscore-m1-ultra-gpu-count-not-published-128-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 283.7757922222222,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M1 Ultra (GPU count not published, 128 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,57.07094177777779 tok/s; ttft 4.787078634222223s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/307",
   "tps": 57.07094177777779,
   "ttft_s": 4.787078634222223
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 Ultra (GPU count not published, 128 GB)",
   "id": "siliconscore-m1-ultra-gpu-count-not-published-128-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 38.83732133333333,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M1 Ultra (GPU count not published, 128 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,7.9883163333333345 tok/s; ttft 32.935429014111115s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/307",
   "tps": 7.9883163333333345,
   "ttft_s": 32.935429014111115
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M2 (8-core GPU, 8 GB)",
   "id": "siliconscore-m2-8-core-gpu-8-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 148.82072661111113,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M2 (8-core GPU, 8 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,18.342969166666663 tok/s; ttft 8.390452365666668s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/245",
   "tps": 18.342969166666663,
   "ttft_s": 8.390452365666668
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M2 (8-core GPU, 8 GB)",
   "id": "siliconscore-m2-8-core-gpu-8-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 523.2109411111111,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M2 (8-core GPU, 8 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,34.48367238888889 tok/s; ttft 2.7225200694166665s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/245",
   "tps": 34.48367238888889,
   "ttft_s": 2.7225200694166665
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M2 (8-core GPU, 16 GB)",
   "id": "siliconscore-m2-8-core-gpu-16-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 113.95904577777776,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M2 (8-core GPU, 16 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,12.906848888888891 tok/s; ttft 10.946365963333333s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/1939",
   "tps": 12.906848888888891,
   "ttft_s": 10.946365963333333
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M2 (8-core GPU, 16 GB)",
   "id": "siliconscore-m2-8-core-gpu-16-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 523.1656983888889,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M2 (8-core GPU, 16 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,35.320437777777784 tok/s; ttft 3.046989798555555s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/1939",
   "tps": 35.320437777777784,
   "ttft_s": 3.046989798555555
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M2 (8-core GPU, 16 GB)",
   "id": "siliconscore-m2-8-core-gpu-16-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 59.9820798888889,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M2 (8-core GPU, 16 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,6.960648666666668 tok/s; ttft 21.432082513888886s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/1939",
   "tps": 6.960648666666668,
   "ttft_s": 21.432082513888886
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M2 (10-core GPU, 8 GB)",
   "id": "siliconscore-m2-10-core-gpu-8-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 804.8249559999999,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M2 (10-core GPU, 8 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,56.46111377777777 tok/s; ttft 1.6694184073333336s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/847",
   "tps": 56.46111377777777,
   "ttft_s": 1.6694184073333336
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M2 (10-core GPU, 16 GB)",
   "id": "siliconscore-m2-10-core-gpu-16-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 140.73035655555557,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M2 (10-core GPU, 16 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,14.68716522222222 tok/s; ttft 9.125759541666666s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/499",
   "tps": 14.68716522222222,
   "ttft_s": 9.125759541666666
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M2 (10-core GPU, 16 GB)",
   "id": "siliconscore-m2-10-core-gpu-16-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 801.8366849333333,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M2 (10-core GPU, 16 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,55.565495822222225 tok/s; ttft 1.589305650066667s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/499",
   "tps": 55.565495822222225,
   "ttft_s": 1.589305650066667
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M2 (10-core GPU, 16 GB)",
   "id": "siliconscore-m2-10-core-gpu-16-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 74.75765844444444,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M2 (10-core GPU, 16 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,8.076025111111111 tok/s; ttft 17.296274291666677s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/499",
   "tps": 8.076025111111111,
   "ttft_s": 17.296274291666677
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M2 (10-core GPU, 24 GB)",
   "id": "siliconscore-m2-10-core-gpu-24-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 141.73761555555558,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M2 (10-core GPU, 24 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,14.668569888888888 tok/s; ttft 9.023242666666668s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/202",
   "tps": 14.668569888888888,
   "ttft_s": 9.023242666666668
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M2 (10-core GPU, 24 GB)",
   "id": "siliconscore-m2-10-core-gpu-24-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 813.7094306666667,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M2 (10-core GPU, 24 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,54.7565285 tok/s; ttft 1.5797905915s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/202",
   "tps": 54.7565285,
   "ttft_s": 1.5797905915
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M2 (10-core GPU, 24 GB)",
   "id": "siliconscore-m2-10-core-gpu-24-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 70.25649544444445,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M2 (10-core GPU, 24 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,7.345236666666668 tok/s; ttft 18.027677185222224s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/202",
   "tps": 7.345236666666668,
   "ttft_s": 18.027677185222224
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M2 Max (30-core GPU, 32 GB)",
   "id": "siliconscore-m2-max-30-core-gpu-32-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 325.001513,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M2 Max (30-core GPU, 32 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,31.185154777777775 tok/s; ttft 3.7657567592222216s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/1079",
   "tps": 31.185154777777775,
   "ttft_s": 3.7657567592222216
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M2 Max (30-core GPU, 32 GB)",
   "id": "siliconscore-m2-max-30-core-gpu-32-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 2031.665092888889,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M2 Max (30-core GPU, 32 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,127.58507908333334 tok/s; ttft 0.5575502119166668s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/1079",
   "tps": 127.58507908333334,
   "ttft_s": 0.5575502119166668
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M2 Max (30-core GPU, 64 GB)",
   "id": "siliconscore-m2-max-30-core-gpu-64-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 149.0023927777778,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M2 Max (30-core GPU, 64 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,14.487775777777776 tok/s; ttft 7.678328407666665s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/1473",
   "tps": 14.487775777777776,
   "ttft_s": 7.678328407666665
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M2 Max (38-core GPU, 32 GB)",
   "id": "siliconscore-m2-max-38-core-gpu-32-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 473.660562,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M2 Max (38-core GPU, 32 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,44.66334788888888 tok/s; ttft 2.4878861063333333s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/1291",
   "tps": 44.66334788888888,
   "ttft_s": 2.4878861063333333
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M2 Max (38-core GPU, 32 GB)",
   "id": "siliconscore-m2-max-38-core-gpu-32-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 2551.365566666667,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M2 Max (38-core GPU, 32 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,152.97702255555558 tok/s; ttft 0.44196865744444436s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/1291",
   "tps": 152.97702255555558,
   "ttft_s": 0.44196865744444436
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M2 Max (38-core GPU, 32 GB)",
   "id": "siliconscore-m2-max-38-core-gpu-32-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 225.5740881111111,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M2 Max (38-core GPU, 32 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,20.586609111111116 tok/s; ttft 4.983570564777778s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/1291",
   "tps": 20.586609111111116,
   "ttft_s": 4.983570564777778
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M2 Max (38-core GPU, 64 GB)",
   "id": "siliconscore-m2-max-38-core-gpu-64-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 222.7697787777778,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M2 Max (38-core GPU, 64 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,21.958976814814815 tok/s; ttft 5.511166165148148s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/1577",
   "tps": 21.958976814814815,
   "ttft_s": 5.511166165148148
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M2 Max (38-core GPU, 96 GB)",
   "id": "siliconscore-m2-max-38-core-gpu-96-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 484.057812,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M2 Max (38-core GPU, 96 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,46.41150744444444 tok/s; ttft 2.4422629954444446s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/515",
   "tps": 46.41150744444444,
   "ttft_s": 2.4422629954444446
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M2 Max (38-core GPU, 96 GB)",
   "id": "siliconscore-m2-max-38-core-gpu-96-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 252.470057,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M2 Max (38-core GPU, 96 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,25.230792444444447 tok/s; ttft 4.798504768333334s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/515",
   "tps": 25.230792444444447,
   "ttft_s": 4.798504768333334
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M2 Pro (16-core GPU, 16 GB)",
   "id": "siliconscore-m2-pro-16-core-gpu-16-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 224.56031674074077,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M2 Pro (16-core GPU, 16 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,24.331960740740744 tok/s; ttft 5.574148605037037s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/1220",
   "tps": 24.331960740740744,
   "ttft_s": 5.574148605037037
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M2 Pro (16-core GPU, 16 GB)",
   "id": "siliconscore-m2-pro-16-core-gpu-16-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 1328.2018266666669,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M2 Pro (16-core GPU, 16 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,91.14859188888887 tok/s; ttft 0.9017461759444444s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/1220",
   "tps": 91.14859188888887,
   "ttft_s": 0.9017461759444444
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M2 Pro (16-core GPU, 16 GB)",
   "id": "siliconscore-m2-pro-16-core-gpu-16-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 118.980526,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M2 Pro (16-core GPU, 16 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,13.350706388888891 tok/s; ttft 10.621650037111111s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/1220",
   "tps": 13.350706388888891,
   "ttft_s": 10.621650037111111
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M2 Pro (16-core GPU, 32 GB)",
   "id": "siliconscore-m2-pro-16-core-gpu-32-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 208.68323722222223,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M2 Pro (16-core GPU, 32 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,23.809723111111115 tok/s; ttft 6.06785997211111s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/1298",
   "tps": 23.809723111111115,
   "ttft_s": 6.06785997211111
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M2 Pro (16-core GPU, 32 GB)",
   "id": "siliconscore-m2-pro-16-core-gpu-32-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 1281.403522111111,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M2 Pro (16-core GPU, 32 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,91.45281944444444 tok/s; ttft 0.9141989862222222s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/1298",
   "tps": 91.45281944444444,
   "ttft_s": 0.9141989862222222
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M2 Pro (19-core GPU, 16 GB)",
   "id": "siliconscore-m2-pro-19-core-gpu-16-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 1457.8697797777777,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M2 Pro (19-core GPU, 16 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,99.52299688888888 tok/s; ttft 0.8091630742222221s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/1862",
   "tps": 99.52299688888888,
   "ttft_s": 0.8091630742222221
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M2 Pro (19-core GPU, 32 GB)",
   "id": "siliconscore-m2-pro-19-core-gpu-32-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 261.761128,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M2 Pro (19-core GPU, 32 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,26.311742333333328 tok/s; ttft 4.701914784666666s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/470",
   "tps": 26.311742333333328,
   "ttft_s": 4.701914784666666
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M2 Pro (19-core GPU, 32 GB)",
   "id": "siliconscore-m2-pro-19-core-gpu-32-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 1487.5642871333334,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M2 Pro (19-core GPU, 32 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,100.2557703111111 tok/s; ttft 0.8063303454888889s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/470",
   "tps": 100.2557703111111,
   "ttft_s": 0.8063303454888889
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M2 Pro (19-core GPU, 32 GB)",
   "id": "siliconscore-m2-pro-19-core-gpu-32-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 137.32948566666667,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M2 Pro (19-core GPU, 32 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,14.116209888888887 tok/s; ttft 9.101328361111113s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/470",
   "tps": 14.116209888888887,
   "ttft_s": 9.101328361111113
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M2 Ultra (60-core GPU, 64 GB)",
   "id": "siliconscore-m2-ultra-60-core-gpu-64-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 703.2592348888888,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M2 Ultra (60-core GPU, 64 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,59.480982 tok/s; ttft 1.6398754306666665s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/2309",
   "tps": 59.480982,
   "ttft_s": 1.6398754306666665
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M2 Ultra (60-core GPU, 64 GB)",
   "id": "siliconscore-m2-ultra-60-core-gpu-64-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 3290.1508422222223,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M2 Ultra (60-core GPU, 64 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,174.08661300000003 tok/s; ttft 0.3351846761111111s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/2309",
   "tps": 174.08661300000003,
   "ttft_s": 0.3351846761111111
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M2 Ultra (60-core GPU, 64 GB)",
   "id": "siliconscore-m2-ultra-60-core-gpu-64-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 381.12729266666673,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M2 Ultra (60-core GPU, 64 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,34.21236922222223 tok/s; ttft 3.0647734165555556s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/2309",
   "tps": 34.21236922222223,
   "ttft_s": 3.0647734165555556
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M2 Ultra (60-core GPU, 128 GB)",
   "id": "siliconscore-m2-ultra-60-core-gpu-128-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 3295.522693222222,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M2 Ultra (60-core GPU, 128 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,176.3882528888889 tok/s; ttft 0.33357317588888885s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/617",
   "tps": 176.3882528888889,
   "ttft_s": 0.33357317588888885
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M2 Ultra (60-core GPU, 192 GB)",
   "id": "siliconscore-m2-ultra-60-core-gpu-192-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 3272.209784333334,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M2 Ultra (60-core GPU, 192 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,169.76082655555555 tok/s; ttft 0.3388111018888889s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/273",
   "tps": 169.76082655555555,
   "ttft_s": 0.3388111018888889
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M2 Ultra (76-core GPU, 128 GB)",
   "id": "siliconscore-m2-ultra-76-core-gpu-128-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 470.5342577777778,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M2 Ultra (76-core GPU, 128 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,36.626677 tok/s; ttft 2.484028763888889s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/1222",
   "tps": 36.626677,
   "ttft_s": 2.484028763888889
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M2 Ultra (GPU count not published, 128 GB)",
   "id": "siliconscore-m2-ultra-gpu-count-not-published-128-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 659.8306766666667,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M2 Ultra (GPU count not published, 128 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,120.4092842222222 tok/s; ttft 1.9676880275555553s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/621",
   "tps": 120.4092842222222,
   "ttft_s": 1.9676880275555553
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 (10-core GPU, 16 GB)",
   "id": "siliconscore-m3-10-core-gpu-16-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 138.71456447222224,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M3 (10-core GPU, 16 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,13.499802861111112 tok/s; ttft 9.024639849583332s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/414",
   "tps": 13.499802861111112,
   "ttft_s": 9.024639849583332
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 (10-core GPU, 16 GB)",
   "id": "siliconscore-m3-10-core-gpu-16-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 931.5106880888889,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M3 (10-core GPU, 16 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,67.18826562222223 tok/s; ttft 1.4093825610888888s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/414",
   "tps": 67.18826562222223,
   "ttft_s": 1.4093825610888888
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 (10-core GPU, 24 GB)",
   "id": "siliconscore-m3-10-core-gpu-24-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 109.8980422777778,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M3 (10-core GPU, 24 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,10.22738411111111 tok/s; ttft 11.670486025444445s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/562",
   "tps": 10.22738411111111,
   "ttft_s": 11.670486025444445
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 (10-core GPU, 24 GB)",
   "id": "siliconscore-m3-10-core-gpu-24-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 916.5074562888888,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M3 (10-core GPU, 24 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,64.71307648888889 tok/s; ttft 1.4258484739777777s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/562",
   "tps": 64.71307648888889,
   "ttft_s": 1.4258484739777777
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 (10-core GPU, 24 GB)",
   "id": "siliconscore-m3-10-core-gpu-24-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 65.80348074074074,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M3 (10-core GPU, 24 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,6.133619851851852 tok/s; ttft 19.91686526244444s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/562",
   "tps": 6.133619851851852,
   "ttft_s": 19.91686526244444
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 (GPU count not published, 16 GB)",
   "id": "siliconscore-m3-gpu-count-not-published-16-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 33.58097844444444,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M3 (GPU count not published, 16 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,11.83115477777778 tok/s; ttft 40.49901329166667s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/1139",
   "tps": 11.83115477777778,
   "ttft_s": 40.49901329166667
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 (GPU count not published, 16 GB)",
   "id": "siliconscore-m3-gpu-count-not-published-16-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 229.7350115555556,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M3 (GPU count not published, 16 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,61.584239 tok/s; ttft 6.018140875s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/1139",
   "tps": 61.584239,
   "ttft_s": 6.018140875
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 Max (30-core GPU, 36 GB)",
   "id": "siliconscore-m3-max-30-core-gpu-36-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 443.4985221666666,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M3 Max (30-core GPU, 36 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,37.47979072222222 tok/s; ttft 2.725718886444444s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/805",
   "tps": 37.47979072222222,
   "ttft_s": 2.725718886444444
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 Max (30-core GPU, 36 GB)",
   "id": "siliconscore-m3-max-30-core-gpu-36-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 2553.2959846666668,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M3 Max (30-core GPU, 36 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,132.9925948888889 tok/s; ttft 0.4669405787777778s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/805",
   "tps": 132.9925948888889,
   "ttft_s": 0.4669405787777778
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 Max (30-core GPU, 36 GB)",
   "id": "siliconscore-m3-max-30-core-gpu-36-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 226.180183,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M3 Max (30-core GPU, 36 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,19.768525888888888 tok/s; ttft 5.407046111222224s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/805",
   "tps": 19.768525888888888,
   "ttft_s": 5.407046111222224
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 Max (30-core GPU, 96 GB)",
   "id": "siliconscore-m3-max-30-core-gpu-96-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 456.8050157777777,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M3 Max (30-core GPU, 96 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,37.74030144444444 tok/s; ttft 2.6877091111111118s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/789",
   "tps": 37.74030144444444,
   "ttft_s": 2.6877091111111118
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 Max (30-core GPU, 96 GB)",
   "id": "siliconscore-m3-max-30-core-gpu-96-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 2519.9553065555556,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M3 Max (30-core GPU, 96 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,132.87379466666664 tok/s; ttft 0.4758228193333332s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/789",
   "tps": 132.87379466666664,
   "ttft_s": 0.4758228193333332
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 Max (30-core GPU, 96 GB)",
   "id": "siliconscore-m3-max-30-core-gpu-96-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 238.76133155555556,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M3 Max (30-core GPU, 96 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,20.766045 tok/s; ttft 5.162620243055555s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/789",
   "tps": 20.766045,
   "ttft_s": 5.162620243055555
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 Max (40-core GPU, 48 GB)",
   "id": "siliconscore-m3-max-40-core-gpu-48-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 3399.1241412222225,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M3 Max (40-core GPU, 48 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,148.95889516666665 tok/s; ttft 0.34880203711111113s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/1114",
   "tps": 148.95889516666665,
   "ttft_s": 0.34880203711111113
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 Max (40-core GPU, 64 GB)",
   "id": "siliconscore-m3-max-40-core-gpu-64-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 377.04858177777777,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M3 Max (40-core GPU, 64 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,25.448832000000003 tok/s; ttft 3.3699854838888887s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/1469",
   "tps": 25.448832000000003,
   "ttft_s": 3.3699854838888887
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 Max (40-core GPU, 64 GB)",
   "id": "siliconscore-m3-max-40-core-gpu-64-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 2521.617585666667,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M3 Max (40-core GPU, 64 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,106.97443855555557 tok/s; ttft 0.520643525388889s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/1469",
   "tps": 106.97443855555557,
   "ttft_s": 0.520643525388889
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 Max (40-core GPU, 64 GB)",
   "id": "siliconscore-m3-max-40-core-gpu-64-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 200.82679227777783,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M3 Max (40-core GPU, 64 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,13.801572 tok/s; ttft 6.657162884277778s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/1469",
   "tps": 13.801572,
   "ttft_s": 6.657162884277778
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 Max (40-core GPU, 128 GB)",
   "id": "siliconscore-m3-max-40-core-gpu-128-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 587.0791163333333,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M3 Max (40-core GPU, 128 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,45.792675277777775 tok/s; ttft 2.0508701621111114s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/750",
   "tps": 45.792675277777775,
   "ttft_s": 2.0508701621111114
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 Max (40-core GPU, 128 GB)",
   "id": "siliconscore-m3-max-40-core-gpu-128-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 3291.9091724444447,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M3 Max (40-core GPU, 128 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,146.33974311111112 tok/s; ttft 0.3584107959999999s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/750",
   "tps": 146.33974311111112,
   "ttft_s": 0.3584107959999999
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 Max (40-core GPU, 128 GB)",
   "id": "siliconscore-m3-max-40-core-gpu-128-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 302.37069566666673,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M3 Max (40-core GPU, 128 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,25.48905755555555 tok/s; ttft 4.070913819333334s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/750",
   "tps": 25.48905755555555,
   "ttft_s": 4.070913819333334
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 Pro (14-core GPU, 18 GB)",
   "id": "siliconscore-m3-pro-14-core-gpu-18-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 199.51528166666668,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M3 Pro (14-core GPU, 18 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,19.10761944444444 tok/s; ttft 6.427559875000002s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/263",
   "tps": 19.10761944444444,
   "ttft_s": 6.427559875000002
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 Pro (14-core GPU, 18 GB)",
   "id": "siliconscore-m3-pro-14-core-gpu-18-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 1343.9570948888888,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M3 Pro (14-core GPU, 18 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,88.10604988888889 tok/s; ttft 0.9594814676666664s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/263",
   "tps": 88.10604988888889,
   "ttft_s": 0.9594814676666664
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 Pro (14-core GPU, 18 GB)",
   "id": "siliconscore-m3-pro-14-core-gpu-18-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 117.03684152777775,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M3 Pro (14-core GPU, 18 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,11.886003166666669 tok/s; ttft 10.918680665472221s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/263",
   "tps": 11.886003166666669,
   "ttft_s": 10.918680665472221
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 Pro (14-core GPU, 36 GB)",
   "id": "siliconscore-m3-pro-14-core-gpu-36-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 223.43918544444443,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M3 Pro (14-core GPU, 36 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,21.455100444444447 tok/s; ttft 5.651565199111111s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/862",
   "tps": 21.455100444444447,
   "ttft_s": 5.651565199111111
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 Pro (14-core GPU, 36 GB)",
   "id": "siliconscore-m3-pro-14-core-gpu-36-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 1327.8959733333334,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M3 Pro (14-core GPU, 36 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,88.23620972222221 tok/s; ttft 0.9533793772777779s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/862",
   "tps": 88.23620972222221,
   "ttft_s": 0.9533793772777779
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 Pro (14-core GPU, 36 GB)",
   "id": "siliconscore-m3-pro-14-core-gpu-36-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 119.77962366666664,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M3 Pro (14-core GPU, 36 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,12.10437772222222 tok/s; ttft 10.648140152777778s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/862",
   "tps": 12.10437772222222,
   "ttft_s": 10.648140152777778
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 Pro (18-core GPU, 18 GB)",
   "id": "siliconscore-m3-pro-18-core-gpu-18-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 279.21399711111115,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M3 Pro (18-core GPU, 18 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,20.80849266666667 tok/s; ttft 4.529956342555556s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/401",
   "tps": 20.80849266666667,
   "ttft_s": 4.529956342555556
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 Pro (18-core GPU, 18 GB)",
   "id": "siliconscore-m3-pro-18-core-gpu-18-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 1573.6998116666664,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M3 Pro (18-core GPU, 18 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,85.63227766666667 tok/s; ttft 0.8258373006666666s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/401",
   "tps": 85.63227766666667,
   "ttft_s": 0.8258373006666666
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 Pro (18-core GPU, 18 GB)",
   "id": "siliconscore-m3-pro-18-core-gpu-18-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 144.77590885185188,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M3 Pro (18-core GPU, 18 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,11.63579448148148 tok/s; ttft 8.883480046185184s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/401",
   "tps": 11.63579448148148,
   "ttft_s": 8.883480046185184
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 Pro (18-core GPU, 36 GB)",
   "id": "siliconscore-m3-pro-18-core-gpu-36-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 283.8059724444445,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M3 Pro (18-core GPU, 36 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,22.06217488888889 tok/s; ttft 4.460495907444445s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/428",
   "tps": 22.06217488888889,
   "ttft_s": 4.460495907444445
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 Pro (18-core GPU, 36 GB)",
   "id": "siliconscore-m3-pro-18-core-gpu-36-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 1586.168225333333,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M3 Pro (18-core GPU, 36 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,89.75982788888888 tok/s; ttft 0.8094999791666667s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/428",
   "tps": 89.75982788888888,
   "ttft_s": 0.8094999791666667
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 Pro (18-core GPU, 36 GB)",
   "id": "siliconscore-m3-pro-18-core-gpu-36-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 147.43616580555556,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M3 Pro (18-core GPU, 36 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,11.988618277777778 tok/s; ttft 8.729495730361112s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/428",
   "tps": 11.988618277777778,
   "ttft_s": 8.729495730361112
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 Ultra (60-core GPU, 96 GB)",
   "id": "siliconscore-m3-ultra-60-core-gpu-96-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 444.62243322222224,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M3 Ultra (60-core GPU, 96 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,34.43578599999999 tok/s; ttft 2.672898578888889s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/479",
   "tps": 34.43578599999999,
   "ttft_s": 2.672898578888889
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 Ultra (80-core GPU, 256 GB)",
   "id": "siliconscore-m3-ultra-80-core-gpu-256-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 1062.184084222222,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M3 Ultra (80-core GPU, 256 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,63.304543444444434 tok/s; ttft 1.1019065137777775s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/1359",
   "tps": 63.304543444444434,
   "ttft_s": 1.1019065137777775
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 Ultra (80-core GPU, 256 GB)",
   "id": "siliconscore-m3-ultra-80-core-gpu-256-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 4999.496534111112,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M3 Ultra (80-core GPU, 256 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,177.900228 tok/s; ttft 0.22653121744444446s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/1359",
   "tps": 177.900228,
   "ttft_s": 0.22653121744444446
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 Ultra (80-core GPU, 256 GB)",
   "id": "siliconscore-m3-ultra-80-core-gpu-256-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 568.2534414444444,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M3 Ultra (80-core GPU, 256 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,36.711573666666666 tok/s; ttft 2.077086245444445s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/1359",
   "tps": 36.711573666666666,
   "ttft_s": 2.077086245444445
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 Ultra (80-core GPU, 512 GB)",
   "id": "siliconscore-m3-ultra-80-core-gpu-512-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 1109.4627288333336,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M3 Ultra (80-core GPU, 512 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,62.67566477777777 tok/s; ttft 1.060675240722222s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/404",
   "tps": 62.67566477777777,
   "ttft_s": 1.060675240722222
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 Ultra (80-core GPU, 512 GB)",
   "id": "siliconscore-m3-ultra-80-core-gpu-512-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 5600.99501711111,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M3 Ultra (80-core GPU, 512 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,178.83792144444442 tok/s; ttft 0.20634095138888886s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/404",
   "tps": 178.83792144444442,
   "ttft_s": 0.20634095138888886
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 Ultra (80-core GPU, 512 GB)",
   "id": "siliconscore-m3-ultra-80-core-gpu-512-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 577.4241827222222,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M3 Ultra (80-core GPU, 512 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,35.80064133333334 tok/s; ttft 2.0507156840555556s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/404",
   "tps": 35.80064133333334,
   "ttft_s": 2.0507156840555556
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 (8-core GPU, 16 GB)",
   "id": "siliconscore-m4-8-core-gpu-16-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 134.02995044444447,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 (8-core GPU, 16 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,15.26954622222222 tok/s; ttft 9.18359174066667s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/1394",
   "tps": 15.26954622222222,
   "ttft_s": 9.18359174066667
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 (8-core GPU, 16 GB)",
   "id": "siliconscore-m4-8-core-gpu-16-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 897.0363834222222,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 (8-core GPU, 16 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,65.86037953333333 tok/s; ttft 1.4411192250444445s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/1394",
   "tps": 65.86037953333333,
   "ttft_s": 1.4411192250444445
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 (8-core GPU, 16 GB)",
   "id": "siliconscore-m4-8-core-gpu-16-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 61.06916633333333,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 (8-core GPU, 16 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,7.234274277777778 tok/s; ttft 20.47596530788889s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/1394",
   "tps": 7.234274277777778,
   "ttft_s": 20.47596530788889
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 (10-core GPU, 16 GB)",
   "id": "siliconscore-m4-10-core-gpu-16-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 166.82581797222224,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 (10-core GPU, 16 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,15.983290222222225 tok/s; ttft 7.980498432833332s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/32",
   "tps": 15.983290222222225,
   "ttft_s": 7.980498432833332
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 (10-core GPU, 16 GB)",
   "id": "siliconscore-m4-10-core-gpu-16-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 1091.1117940111112,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 (10-core GPU, 16 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,76.16243562222222 tok/s; ttft 1.1983527194666668s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/32",
   "tps": 76.16243562222222,
   "ttft_s": 1.1983527194666668
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 (10-core GPU, 16 GB)",
   "id": "siliconscore-m4-10-core-gpu-16-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 83.05415,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 (10-core GPU, 16 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,8.726884422222223 tok/s; ttft 15.662719702777778s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/32",
   "tps": 8.726884422222223,
   "ttft_s": 15.662719702777778
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 (10-core GPU, 24 GB)",
   "id": "siliconscore-m4-10-core-gpu-24-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 148.96748985185187,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 (10-core GPU, 24 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,15.935378370370373 tok/s; ttft 8.236942598851853s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/1085",
   "tps": 15.935378370370373,
   "ttft_s": 8.236942598851853
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 (10-core GPU, 24 GB)",
   "id": "siliconscore-m4-10-core-gpu-24-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 1035.9960293703705,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 (10-core GPU, 24 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,75.4330094074074 tok/s; ttft 1.2527001605185184s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/1085",
   "tps": 75.4330094074074,
   "ttft_s": 1.2527001605185184
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 (10-core GPU, 24 GB)",
   "id": "siliconscore-m4-10-core-gpu-24-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 93.30326355555556,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 (10-core GPU, 24 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,9.238281555555554 tok/s; ttft 13.714502268555552s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/1085",
   "tps": 9.238281555555554,
   "ttft_s": 13.714502268555552
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 (10-core GPU, 32 GB)",
   "id": "siliconscore-m4-10-core-gpu-32-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 166.14551366666666,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 (10-core GPU, 32 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,16.809439277777777 tok/s; ttft 7.554811377277778s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/546",
   "tps": 16.809439277777777,
   "ttft_s": 7.554811377277778
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 (10-core GPU, 32 GB)",
   "id": "siliconscore-m4-10-core-gpu-32-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 1069.8806100277775,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 (10-core GPU, 32 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,75.56749613888888 tok/s; ttft 1.2130988726666667s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/546",
   "tps": 75.56749613888888,
   "ttft_s": 1.2130988726666667
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 (10-core GPU, 32 GB)",
   "id": "siliconscore-m4-10-core-gpu-32-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 79.33710781481481,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 (10-core GPU, 32 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,8.58771625925926 tok/s; ttft 15.81842497840741s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/546",
   "tps": 8.58771625925926,
   "ttft_s": 15.81842497840741
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 (GPU count not published, 16 GB)",
   "id": "siliconscore-m4-gpu-count-not-published-16-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 238.95714649999996,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 (GPU count not published, 16 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,67.95033205555555 tok/s; ttft 5.743805437444445s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/35",
   "tps": 67.95033205555555,
   "ttft_s": 5.743805437444445
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Max (32-core GPU, 36 GB)",
   "id": "siliconscore-m4-max-32-core-gpu-36-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 570.2711772,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Max (32-core GPU, 36 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,48.071329866666666 tok/s; ttft 2.1545828000444445s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/162",
   "tps": 48.071329866666666,
   "ttft_s": 2.1545828000444445
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Max (32-core GPU, 36 GB)",
   "id": "siliconscore-m4-max-32-core-gpu-36-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 3268.880837361111,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Max (32-core GPU, 36 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,166.45011455555556 tok/s; ttft 0.366560225611111s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/162",
   "tps": 166.45011455555556,
   "ttft_s": 0.366560225611111
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Max (32-core GPU, 36 GB)",
   "id": "siliconscore-m4-max-32-core-gpu-36-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 273.40494424444444,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Max (32-core GPU, 36 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,24.60766388888889 tok/s; ttft 4.441582821466667s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/162",
   "tps": 24.60766388888889,
   "ttft_s": 4.441582821466667
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Max (40-core GPU, 48 GB)",
   "id": "siliconscore-m4-max-40-core-gpu-48-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 663.3646092592592,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Max (40-core GPU, 48 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,55.0887822962963 tok/s; ttft 1.7937055694814816s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/1188",
   "tps": 55.0887822962963,
   "ttft_s": 1.7937055694814816
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Max (40-core GPU, 48 GB)",
   "id": "siliconscore-m4-max-40-core-gpu-48-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 3750.568106722222,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Max (40-core GPU, 48 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,178.9852262222222 tok/s; ttft 0.30928531144444443s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/1188",
   "tps": 178.9852262222222,
   "ttft_s": 0.30928531144444443
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Max (40-core GPU, 48 GB)",
   "id": "siliconscore-m4-max-40-core-gpu-48-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 347.03543255555553,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Max (40-core GPU, 48 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,30.13459711111111 tok/s; ttft 3.4913595508333337s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/1188",
   "tps": 30.13459711111111,
   "ttft_s": 3.4913595508333337
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Max (40-core GPU, 64 GB)",
   "id": "siliconscore-m4-max-40-core-gpu-64-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 557.1200590555555,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Max (40-core GPU, 64 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,47.095953944444446 tok/s; ttft 2.1445155647222225s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/776",
   "tps": 47.095953944444446,
   "ttft_s": 2.1445155647222225
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Max (40-core GPU, 64 GB)",
   "id": "siliconscore-m4-max-40-core-gpu-64-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 3857.499544666666,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Max (40-core GPU, 64 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,180.24040433333337 tok/s; ttft 0.30436414811111107s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/776",
   "tps": 180.24040433333337,
   "ttft_s": 0.30436414811111107
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Max (40-core GPU, 64 GB)",
   "id": "siliconscore-m4-max-40-core-gpu-64-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 286.74860906666663,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Max (40-core GPU, 64 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,25.874467866666663 tok/s; ttft 4.498645011088889s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/776",
   "tps": 25.874467866666663,
   "ttft_s": 4.498645011088889
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Max (40-core GPU, 128 GB)",
   "id": "siliconscore-m4-max-40-core-gpu-128-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 617.642594037037,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Max (40-core GPU, 128 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,51.63987292592592 tok/s; ttft 1.9420189397222225s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/6",
   "tps": 51.63987292592592,
   "ttft_s": 1.9420189397222225
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Max (40-core GPU, 128 GB)",
   "id": "siliconscore-m4-max-40-core-gpu-128-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 3833.843851847222,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Max (40-core GPU, 128 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,182.5577164513889 tok/s; ttft 0.30469161659027777s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/6",
   "tps": 182.5577164513889,
   "ttft_s": 0.30469161659027777
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Max (40-core GPU, 128 GB)",
   "id": "siliconscore-m4-max-40-core-gpu-128-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 326.6194374666667,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Max (40-core GPU, 128 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,28.74764654814815 tok/s; ttft 3.7867302675777776s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/6",
   "tps": 28.74764654814815,
   "ttft_s": 3.7867302675777776
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Max (GPU count not published, 128 GB)",
   "id": "siliconscore-m4-max-gpu-count-not-published-128-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 679.7101832222222,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Max (GPU count not published, 128 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,156.3000006666667 tok/s; ttft 1.9661574258888888s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/8",
   "tps": 156.3000006666667,
   "ttft_s": 1.9661574258888888
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Pro (16-core GPU, 24 GB)",
   "id": "siliconscore-m4-pro-16-core-gpu-24-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 298.0409531777778,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Pro (16-core GPU, 24 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,30.508745533333336 tok/s; ttft 4.201011161111111s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/380",
   "tps": 30.508745533333336,
   "ttft_s": 4.201011161111111
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Pro (16-core GPU, 24 GB)",
   "id": "siliconscore-m4-pro-16-core-gpu-24-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 1823.817922962963,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Pro (16-core GPU, 24 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,110.94556159259258 tok/s; ttft 0.6715746882962962s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/380",
   "tps": 110.94556159259258,
   "ttft_s": 0.6715746882962962
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Pro (16-core GPU, 24 GB)",
   "id": "siliconscore-m4-pro-16-core-gpu-24-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 144.2954926666667,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Pro (16-core GPU, 24 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,15.160405666666666 tok/s; ttft 8.903966308944444s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/380",
   "tps": 15.160405666666666,
   "ttft_s": 8.903966308944444
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Pro (16-core GPU, 48 GB)",
   "id": "siliconscore-m4-pro-16-core-gpu-48-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 302.390073,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Pro (16-core GPU, 48 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,30.176443111111105 tok/s; ttft 4.143549694333333s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/550",
   "tps": 30.176443111111105,
   "ttft_s": 4.143549694333333
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Pro (16-core GPU, 48 GB)",
   "id": "siliconscore-m4-pro-16-core-gpu-48-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 1754.577648111111,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Pro (16-core GPU, 48 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,111.00253472222222 tok/s; ttft 0.6816085969999999s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/550",
   "tps": 111.00253472222222,
   "ttft_s": 0.6816085969999999
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Pro (16-core GPU, 48 GB)",
   "id": "siliconscore-m4-pro-16-core-gpu-48-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 161.09253933333332,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Pro (16-core GPU, 48 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,16.762545222222222 tok/s; ttft 7.843727675888888s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/550",
   "tps": 16.762545222222222,
   "ttft_s": 7.843727675888888
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Pro (16-core GPU, 64 GB)",
   "id": "siliconscore-m4-pro-16-core-gpu-64-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 1858.9418227777776,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Pro (16-core GPU, 64 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,111.93509922222222 tok/s; ttft 0.6703511943333332s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/2277",
   "tps": 111.93509922222222,
   "ttft_s": 0.6703511943333332
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Pro (16-core GPU, 64 GB)",
   "id": "siliconscore-m4-pro-16-core-gpu-64-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 151.01082800000003,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Pro (16-core GPU, 64 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,16.132661555555554 tok/s; ttft 8.460523194444445s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/2277",
   "tps": 16.132661555555554,
   "ttft_s": 8.460523194444445
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Pro (20-core GPU, 24 GB)",
   "id": "siliconscore-m4-pro-20-core-gpu-24-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 359.5577164,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Pro (20-core GPU, 24 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,32.533236711111115 tok/s; ttft 3.419087840755556s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/561",
   "tps": 32.533236711111115,
   "ttft_s": 3.419087840755556
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Pro (20-core GPU, 24 GB)",
   "id": "siliconscore-m4-pro-20-core-gpu-24-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 2128.7541195555555,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Pro (20-core GPU, 24 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,119.24874207407407 tok/s; ttft 0.5725979937407406s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/561",
   "tps": 119.24874207407407,
   "ttft_s": 0.5725979937407406
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Pro (20-core GPU, 24 GB)",
   "id": "siliconscore-m4-pro-20-core-gpu-24-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 190.44284311111113,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Pro (20-core GPU, 24 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,17.982970555555557 tok/s; ttft 6.6294910853968245s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/561",
   "tps": 17.982970555555557,
   "ttft_s": 6.6294910853968245
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Pro (20-core GPU, 48 GB)",
   "id": "siliconscore-m4-pro-20-core-gpu-48-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 361.8889191604938,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Pro (20-core GPU, 48 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,32.67357938271605 tok/s; ttft 3.435018682604938s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/777",
   "tps": 32.67357938271605,
   "ttft_s": 3.435018682604938
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Pro (20-core GPU, 48 GB)",
   "id": "siliconscore-m4-pro-20-core-gpu-48-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 2134.135712188034,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Pro (20-core GPU, 48 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,118.8963297948718 tok/s; ttft 0.5720378254615385s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/777",
   "tps": 118.8963297948718,
   "ttft_s": 0.5720378254615385
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Pro (20-core GPU, 48 GB)",
   "id": "siliconscore-m4-pro-20-core-gpu-48-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 189.7956117692308,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Pro (20-core GPU, 48 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,17.981375188034185 tok/s; ttft 6.62966483617094s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/777",
   "tps": 17.981375188034185,
   "ttft_s": 6.62966483617094
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Pro (20-core GPU, 64 GB)",
   "id": "siliconscore-m4-pro-20-core-gpu-64-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 349.4138904444444,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Pro (20-core GPU, 64 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,32.90144462962963 tok/s; ttft 3.563006918222223s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/369",
   "tps": 32.90144462962963,
   "ttft_s": 3.563006918222223
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Pro (20-core GPU, 64 GB)",
   "id": "siliconscore-m4-pro-20-core-gpu-64-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 2145.125925888889,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Pro (20-core GPU, 64 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,118.5574338888889 tok/s; ttft 0.5718198889722222s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/369",
   "tps": 118.5574338888889,
   "ttft_s": 0.5718198889722222
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Pro (20-core GPU, 64 GB)",
   "id": "siliconscore-m4-pro-20-core-gpu-64-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 182.97843991666667,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Pro (20-core GPU, 64 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,18.008562833333333 tok/s; ttft 6.9151872395s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/369",
   "tps": 18.008562833333333,
   "ttft_s": 6.9151872395
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M5 (10-core GPU, 16 GB)",
   "id": "siliconscore-m5-10-core-gpu-16-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 1244.9073107777776,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M5 (10-core GPU, 16 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,98.08477433333331 tok/s; ttft 0.995430389s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/1716",
   "tps": 98.08477433333331,
   "ttft_s": 0.995430389
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M5 (10-core GPU, 32 GB)",
   "id": "siliconscore-m5-10-core-gpu-32-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 210.70696922222223,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M5 (10-core GPU, 32 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,22.29346603703704 tok/s; ttft 5.997633993703704s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/2912",
   "tps": 22.29346603703704,
   "ttft_s": 5.997633993703704
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M5 (10-core GPU, 32 GB)",
   "id": "siliconscore-m5-10-core-gpu-32-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 1271.5481898333337,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M5 (10-core GPU, 32 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,98.37004172222224 tok/s; ttft 0.989687108888889s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/2912",
   "tps": 98.37004172222224,
   "ttft_s": 0.989687108888889
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M5 (10-core GPU, 32 GB)",
   "id": "siliconscore-m5-10-core-gpu-32-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 110.4192515,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M5 (10-core GPU, 32 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,11.536774722222225 tok/s; ttft 11.637239780333331s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/2912",
   "tps": 11.536774722222225,
   "ttft_s": 11.637239780333331
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M5 Max (32-core GPU, 36 GB)",
   "id": "siliconscore-m5-max-32-core-gpu-36-gb-llama-3-1-8b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-1-8b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 630.2577913333333,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M5 Max (32-core GPU, 36 GB),llama-3-1-8b-instruct,Q4_K - Medium,ctx None,61.634733222222216 tok/s; ttft 1.914077976777778s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/3105",
   "tps": 61.634733222222216,
   "ttft_s": 1.914077976777778
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M5 Max (32-core GPU, 36 GB)",
   "id": "siliconscore-m5-max-32-core-gpu-36-gb-llama-3-2-1b-instruct-q4-k-medium-ctxnone",
   "model": "llama-3-2-1b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 3620.6608126666665,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M5 Max (32-core GPU, 36 GB),llama-3-2-1b-instruct,Q4_K - Medium,ctx None,228.99372777777776 tok/s; ttft 0.32193552311111107s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/3105",
   "tps": 228.99372777777776,
   "ttft_s": 0.32193552311111107
  },
  {
   "backend": "llamafile",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M5 Max (32-core GPU, 36 GB)",
   "id": "siliconscore-m5-max-32-core-gpu-36-gb-qwen-2-5-14b-instruct-q4-k-medium-ctxnone",
   "model": "qwen-2-5-14b-instruct",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 343.244466,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M5 Max (32-core GPU, 36 GB),qwen-2-5-14b-instruct,Q4_K - Medium,ctx None,34.28566722222222 tok/s; ttft 3.5801585140000003s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.localscore.ai/accelerator/3105",
   "tps": 34.28566722222222,
   "ttft_s": 3.5801585140000003
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": 8000,
   "date": null,
   "flags": [],
   "hardware": "M5 Max (48 GB)",
   "id": "siliconscore-m5-max-48-gb-qwen3-5-35b-a3b-4bit-ctx8000",
   "model": "Qwen3.5-35B-A3B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 3235,
   "provenance": "sourced",
   "quant": "4bit",
   "quote": "M5 Max (48 GB),Qwen3.5-35B-A3B,4bit,ctx 8000,128 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1s4bggo/benchmarked_qwen35_35b_moe_27b_dense_122b_moe/",
   "tps": 128,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": 8000,
   "date": null,
   "flags": [],
   "hardware": "M1 Max (64 GB)",
   "id": "siliconscore-m1-max-64-gb-qwen3-5-35b-a3b-4bit-ctx8000",
   "model": "Qwen3.5-35B-A3B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 431,
   "provenance": "sourced",
   "quant": "4bit",
   "quote": "M1 Max (64 GB),Qwen3.5-35B-A3B,4bit,ctx 8000,57.6 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1s4bggo/benchmarked_qwen35_35b_moe_27b_dense_122b_moe/",
   "tps": 57.6,
   "ttft_s": null
  },
  {
   "backend": "LM Studio (MLX)",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 Max (64 GB)",
   "id": "siliconscore-m1-max-64-gb-qwen3-5-35b-a3b-4bit-ctxnone",
   "model": "Qwen3.5-35B-A3B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "4bit",
   "quote": "M1 Max (64 GB),Qwen3.5-35B-A3B,4bit,ctx None,57 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://famstack.dev/guides/mlx-vs-gguf-apple-silicon/",
   "tps": 57,
   "ttft_s": null
  },
  {
   "backend": "LM Studio (llama.cpp)",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 Max (64 GB)",
   "id": "siliconscore-m1-max-64-gb-qwen3-5-35b-a3b-q4-k-medium-ctxnone",
   "model": "Qwen3.5-35B-A3B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M1 Max (64 GB),Qwen3.5-35B-A3B,Q4_K - Medium,ctx None,29 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://famstack.dev/guides/mlx-vs-gguf-apple-silicon/",
   "tps": 29,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": 8000,
   "date": null,
   "flags": [],
   "hardware": "M5 Max (48 GB)",
   "id": "siliconscore-m5-max-48-gb-qwen3-5-27b-4bit-ctx8000",
   "model": "Qwen3.5-27B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 779,
   "provenance": "sourced",
   "quant": "4bit",
   "quote": "M5 Max (48 GB),Qwen3.5-27B,4bit,ctx 8000,31.3 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1s4bggo/benchmarked_qwen35_35b_moe_27b_dense_122b_moe/",
   "tps": 31.3,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": 8000,
   "date": null,
   "flags": [],
   "hardware": "M1 Max (64 GB)",
   "id": "siliconscore-m1-max-64-gb-qwen3-5-27b-4bit-ctx8000",
   "model": "Qwen3.5-27B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 67,
   "provenance": "sourced",
   "quant": "4bit",
   "quote": "M1 Max (64 GB),Qwen3.5-27B,4bit,ctx 8000,15 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1s4bggo/benchmarked_qwen35_35b_moe_27b_dense_122b_moe/",
   "tps": 15,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 Max (64 GB)",
   "id": "siliconscore-m1-max-64-gb-qwen3-5-27b-q8-0-ctxnone",
   "model": "Qwen3.5-27B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q8_0",
   "quote": "M1 Max (64 GB),Qwen3.5-27B,Q8_0,ctx None,10.5 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1s0xmq0/question_llamacpp_performance_on_m1_max_qwen_27b/",
   "tps": 10.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 Max (64 GB)",
   "id": "siliconscore-m1-max-64-gb-qwen3-5-27b-q6-k-ctxnone",
   "model": "Qwen3.5-27B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q6_K",
   "quote": "M1 Max (64 GB),Qwen3.5-27B,Q6_K,ctx None,12 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1s0xmq0/question_llamacpp_performance_on_m1_max_qwen_27b/",
   "tps": 12,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 Max (64 GB)",
   "id": "siliconscore-m1-max-64-gb-qwen3-5-27b-q4-k-medium-ctxnone",
   "model": "Qwen3.5-27B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M1 Max (64 GB),Qwen3.5-27B,Q4_K - Medium,ctx None,11.5 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1s0xmq0/question_llamacpp_performance_on_m1_max_qwen_27b/",
   "tps": 11.5,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 8000,
   "date": null,
   "flags": [],
   "hardware": "M5 Max (48 GB)",
   "id": "siliconscore-m5-max-48-gb-qwen3-5-35b-a3b-q4-k-medium-ctx8000",
   "model": "Qwen3.5-35B-A3B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 783,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M5 Max (48 GB),Qwen3.5-35B-A3B,Q4_K - Medium,ctx 8000,89.4 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1s4bggo/benchmarked_qwen35_35b_moe_27b_dense_122b_moe/",
   "tps": 89.4,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 8000,
   "date": null,
   "flags": [],
   "hardware": "M5 Max (48 GB)",
   "id": "siliconscore-m5-max-48-gb-qwen3-5-27b-q4-k-medium-ctx8000",
   "model": "Qwen3.5-27B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 171,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M5 Max (48 GB),Qwen3.5-27B,Q4_K - Medium,ctx 8000,23.7 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1s4bggo/benchmarked_qwen35_35b_moe_27b_dense_122b_moe/",
   "tps": 23.7,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 Ultra (256 GB)",
   "id": "siliconscore-m3-ultra-256-gb-qwen3-5-122b-a10b-8bit-ctxnone",
   "model": "Qwen3.5-122B-A10B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "8bit",
   "quote": "M3 Ultra (256 GB),Qwen3.5-122B-A10B,8bit,ctx None,43 tok/s",
   "ram_gb": 129.8,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1rkcvqa/benchmarked_11_mlx_models_on_m3_ultra_heres_which/",
   "tps": 43,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 Ultra (256 GB)",
   "id": "siliconscore-m3-ultra-256-gb-qwen3-5-122b-a10b-mxfp4-ctxnone",
   "model": "Qwen3.5-122B-A10B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "MXFP4",
   "quote": "M3 Ultra (256 GB),Qwen3.5-122B-A10B,MXFP4,ctx None,57 tok/s",
   "ram_gb": 65,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1rkcvqa/benchmarked_11_mlx_models_on_m3_ultra_heres_which/",
   "tps": 57,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 Ultra (256 GB)",
   "id": "siliconscore-m3-ultra-256-gb-qwen3-coder-next-6bit-ctxnone",
   "model": "Qwen3-Coder-Next",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "6bit",
   "quote": "M3 Ultra (256 GB),Qwen3-Coder-Next,6bit,ctx None,66 tok/s",
   "ram_gb": 64.8,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1rkcvqa/benchmarked_11_mlx_models_on_m3_ultra_heres_which/",
   "tps": 66,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 Ultra (256 GB)",
   "id": "siliconscore-m3-ultra-256-gb-qwen3-coder-next-4bit-ctxnone",
   "model": "Qwen3-Coder-Next",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "4bit",
   "quote": "M3 Ultra (256 GB),Qwen3-Coder-Next,4bit,ctx None,74 tok/s",
   "ram_gb": 44.9,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1rkcvqa/benchmarked_11_mlx_models_on_m3_ultra_heres_which/",
   "tps": 74,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 Ultra (256 GB)",
   "id": "siliconscore-m3-ultra-256-gb-glm-4-5-air-4bit-ctxnone",
   "model": "GLM-4.5-Air",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "4bit",
   "quote": "M3 Ultra (256 GB),GLM-4.5-Air,4bit,ctx None,54 tok/s",
   "ram_gb": 60.3,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1rkcvqa/benchmarked_11_mlx_models_on_m3_ultra_heres_which/",
   "tps": 54,
   "ttft_s": null
  },
  {
   "backend": "LM Studio",
   "batch": null,
   "ctx": 10758,
   "date": null,
   "flags": [],
   "hardware": "M4 Max (128 GB)",
   "id": "siliconscore-m4-max-128-gb-glm-4-5-air-4bit-ctx10758",
   "model": "GLM-4.5-Air",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 183.25,
   "provenance": "sourced",
   "quant": "4bit",
   "quote": "M4 Max (128 GB),GLM-4.5-Air,4bit,ctx 10758,27.75 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1mt3epi/m4_max_generation_speed_vs_context_size/",
   "tps": 27.75,
   "ttft_s": null
  },
  {
   "backend": "LM Studio",
   "batch": null,
   "ctx": 21002,
   "date": null,
   "flags": [],
   "hardware": "M4 Max (128 GB)",
   "id": "siliconscore-m4-max-128-gb-glm-4-5-air-4bit-ctx21002",
   "model": "GLM-4.5-Air",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 120.82,
   "provenance": "sourced",
   "quant": "4bit",
   "quote": "M4 Max (128 GB),GLM-4.5-Air,4bit,ctx 21002,17.99 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1mt3epi/m4_max_generation_speed_vs_context_size/",
   "tps": 17.99,
   "ttft_s": null
  },
  {
   "backend": "LM Studio",
   "batch": null,
   "ctx": 31182,
   "date": null,
   "flags": [],
   "hardware": "M4 Max (128 GB)",
   "id": "siliconscore-m4-max-128-gb-glm-4-5-air-4bit-ctx31182",
   "model": "GLM-4.5-Air",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 75.5,
   "provenance": "sourced",
   "quant": "4bit",
   "quote": "M4 Max (128 GB),GLM-4.5-Air,4bit,ctx 31182,15.07 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1mt3epi/m4_max_generation_speed_vs_context_size/",
   "tps": 15.07,
   "ttft_s": null
  },
  {
   "backend": "LM Studio",
   "batch": null,
   "ctx": 62037,
   "date": null,
   "flags": [],
   "hardware": "M4 Max (128 GB)",
   "id": "siliconscore-m4-max-128-gb-glm-4-5-air-4bit-ctx62037",
   "model": "GLM-4.5-Air",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 41.46,
   "provenance": "sourced",
   "quant": "4bit",
   "quote": "M4 Max (128 GB),GLM-4.5-Air,4bit,ctx 62037,9.58 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1mt3epi/m4_max_generation_speed_vs_context_size/",
   "tps": 9.58,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 Ultra (256 GB)",
   "id": "siliconscore-m3-ultra-256-gb-devstral-small-2-24b-4bit-ctxnone",
   "model": "Devstral Small 2 24B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "4bit",
   "quote": "M3 Ultra (256 GB),Devstral Small 2 24B,4bit,ctx None,47 tok/s",
   "ram_gb": 13.4,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1rkcvqa/benchmarked_11_mlx_models_on_m3_ultra_heres_which/",
   "tps": 47,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 Ultra (256 GB)",
   "id": "siliconscore-m3-ultra-256-gb-glm-4-7-flash-8bit-ctxnone",
   "model": "GLM-4.7-Flash",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "8bit",
   "quote": "M3 Ultra (256 GB),GLM-4.7-Flash,8bit,ctx None,58 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1rkcvqa/benchmarked_11_mlx_models_on_m3_ultra_heres_which/",
   "tps": 58,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 Ultra (256 GB)",
   "id": "siliconscore-m3-ultra-256-gb-qwen-3-0-6b-4bit-ctxnone",
   "model": "qwen-3-0-6b",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "4bit",
   "quote": "M3 Ultra (256 GB),Qwen 3 0.6B,4bit,ctx None,370 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1rkcvqa/benchmarked_11_mlx_models_on_m3_ultra_heres_which/",
   "tps": 370,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 Ultra (256 GB)",
   "id": "siliconscore-m3-ultra-256-gb-qwen3-5-9b-4bit-ctxnone",
   "model": "Qwen3.5-9B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "4bit",
   "quote": "M3 Ultra (256 GB),Qwen3.5-9B,4bit,ctx None,106 tok/s",
   "ram_gb": 5.1,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1rkcvqa/benchmarked_11_mlx_models_on_m3_ultra_heres_which/",
   "tps": 106,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 Ultra (256 GB)",
   "id": "siliconscore-m3-ultra-256-gb-qwen3-5-27b-4bit-ctxnone",
   "model": "Qwen3.5-27B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "4bit",
   "quote": "M3 Ultra (256 GB),Qwen3.5-27B,4bit,ctx None,38 tok/s",
   "ram_gb": 15.3,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1rkcvqa/benchmarked_11_mlx_models_on_m3_ultra_heres_which/",
   "tps": 38,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 Ultra (256 GB)",
   "id": "siliconscore-m3-ultra-256-gb-qwen3-5-35b-a3b-8bit-ctxnone",
   "model": "Qwen3.5-35B-A3B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "8bit",
   "quote": "M3 Ultra (256 GB),Qwen3.5-35B-A3B,8bit,ctx None,80 tok/s",
   "ram_gb": 36.9,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1rkcvqa/benchmarked_11_mlx_models_on_m3_ultra_heres_which/",
   "tps": 80,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 Ultra (256 GB)",
   "id": "siliconscore-m3-ultra-256-gb-qwen3-5-35b-a3b-4bit-ctxnone",
   "model": "Qwen3.5-35B-A3B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "4bit",
   "quote": "M3 Ultra (256 GB),Qwen3.5-35B-A3B,4bit,ctx None,95 tok/s",
   "ram_gb": 19.6,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1rkcvqa/benchmarked_11_mlx_models_on_m3_ultra_heres_which/",
   "tps": 95,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 (16 GB)",
   "id": "siliconscore-m4-16-gb-qwen3-5-9b-q4-0-ctxnone",
   "model": "Qwen3.5-9B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_0",
   "quote": "M4 (16 GB),Qwen3.5-9B,Q4_0,ctx None,4.0648 tok/s; ttft 0.5057s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://huggingface.co/Manojb/macmini-16gb-bench-gguf-mlx/raw/main/SUMMARY.md",
   "tps": 4.0648,
   "ttft_s": 0.5057
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 (16 GB)",
   "id": "siliconscore-m4-16-gb-qwen3-5-9b-q4-k-small-ctxnone",
   "model": "Qwen3.5-9B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K - Small",
   "quote": "M4 (16 GB),Qwen3.5-9B,Q4_K - Small,ctx None,3.142 tok/s; ttft 0.8445s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://huggingface.co/Manojb/macmini-16gb-bench-gguf-mlx/raw/main/SUMMARY.md",
   "tps": 3.142,
   "ttft_s": 0.8445
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 (16 GB)",
   "id": "siliconscore-m4-16-gb-qwen3-5-9b-q6-k-ctxnone",
   "model": "Qwen3.5-9B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q6_K",
   "quote": "M4 (16 GB),Qwen3.5-9B,Q6_K,ctx None,2.1849 tok/s; ttft 1.5031s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://huggingface.co/Manojb/macmini-16gb-bench-gguf-mlx/raw/main/SUMMARY.md",
   "tps": 2.1849,
   "ttft_s": 1.5031
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 (16 GB)",
   "id": "siliconscore-m4-16-gb-devstral-small-2-24b-q4-0-ctxnone",
   "model": "Devstral Small 2 24B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_0",
   "quote": "M4 (16 GB),Devstral Small 2 24B,Q4_0,ctx None,3.3641 tok/s; ttft 2.6874s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://huggingface.co/Manojb/macmini-16gb-bench-gguf-mlx/raw/main/SUMMARY.md",
   "tps": 3.3641,
   "ttft_s": 2.6874
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 (16 GB)",
   "id": "siliconscore-m4-16-gb-devstral-small-2-24b-q4-1-ctxnone",
   "model": "Devstral Small 2 24B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_1",
   "quote": "M4 (16 GB),Devstral Small 2 24B,Q4_1,ctx None,0.0542 tok/s; ttft 17.48s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://huggingface.co/Manojb/macmini-16gb-bench-gguf-mlx/raw/main/SUMMARY.md",
   "tps": 0.0542,
   "ttft_s": 17.48
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 (16 GB)",
   "id": "siliconscore-m4-16-gb-devstral-small-2-24b-q4-k-medium-ctxnone",
   "model": "Devstral Small 2 24B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 (16 GB),Devstral Small 2 24B,Q4_K - Medium,ctx None,0.0198 tok/s; ttft 65.32s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://huggingface.co/Manojb/macmini-16gb-bench-gguf-mlx/raw/main/SUMMARY.md",
   "tps": 0.0198,
   "ttft_s": 65.32
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 (16 GB)",
   "id": "siliconscore-m4-16-gb-qwen3-5-35b-a3b-q4-k-medium-ctxnone",
   "model": "Qwen3.5-35B-A3B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 (16 GB),Qwen3.5-35B-A3B,Q4_K - Medium,ctx None,1.2956 tok/s; ttft 13.0595s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://huggingface.co/Manojb/macmini-16gb-bench-gguf-mlx/raw/main/SUMMARY.md",
   "tps": 1.2956,
   "ttft_s": 13.0595
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 (16 GB)",
   "id": "siliconscore-m4-16-gb-qwen3-5-27b-q4-k-medium-ctxnone",
   "model": "Qwen3.5-27B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 (16 GB),Qwen3.5-27B,Q4_K - Medium,ctx None,0.0084 tok/s; ttft 66.9331s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://huggingface.co/Manojb/macmini-16gb-bench-gguf-mlx/raw/main/SUMMARY.md",
   "tps": 0.0084,
   "ttft_s": 66.9331
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 Ultra (64 GB)",
   "id": "siliconscore-m1-ultra-64-gb-devstral-small-2-24b-4bit-ctxnone",
   "model": "Devstral Small 2 24B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "4bit",
   "quote": "M1 Ultra (64 GB),Devstral Small 2 24B,4bit,ctx None,29.68 tok/s; ttft 6.63s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1plbjqg/devstral_small_2_on_macos/",
   "tps": 29.68,
   "ttft_s": 6.63
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 Ultra (64 GB)",
   "id": "siliconscore-m1-ultra-64-gb-devstral-small-2-24b-8bit-ctxnone",
   "model": "Devstral Small 2 24B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "8bit",
   "quote": "M1 Ultra (64 GB),Devstral Small 2 24B,8bit,ctx None,22.32 tok/s; ttft 7.57s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1plbjqg/devstral_small_2_on_macos/",
   "tps": 22.32,
   "ttft_s": 7.57
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 Ultra (64 GB)",
   "id": "siliconscore-m1-ultra-64-gb-devstral-small-2-24b-q4-k-medium-ctxnone",
   "model": "Devstral Small 2 24B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M1 Ultra (64 GB),Devstral Small 2 24B,Q4_K - Medium,ctx None,25.3 tok/s; ttft 5.89s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1plbjqg/devstral_small_2_on_macos/",
   "tps": 25.3,
   "ttft_s": 5.89
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 Ultra (64 GB)",
   "id": "siliconscore-m1-ultra-64-gb-devstral-small-2-24b-q8-ctxnone",
   "model": "Devstral Small 2 24B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q8",
   "quote": "M1 Ultra (64 GB),Devstral Small 2 24B,Q8,ctx None,23.37 tok/s; ttft 5.66s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1plbjqg/devstral_small_2_on_macos/",
   "tps": 23.37,
   "ttft_s": 5.66
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M1 Pro (16 GB)",
   "id": "siliconscore-m1-pro-16-gb-qwen3-5-9b-4bit-ctxnone",
   "model": "Qwen3.5-9B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "4bit",
   "quote": "M1 Pro (16 GB),Qwen3.5-9B,4bit,ctx None,30 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1rj3ay3/qwen359b_4bit_quant_acting_weird/",
   "tps": 30,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Pro (48 GB)",
   "id": "siliconscore-m4-pro-48-gb-qwen3-5-27b-8bit-ctxnone",
   "model": "Qwen3.5-27B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "8bit",
   "quote": "M4 Pro (48 GB),Qwen3.5-27B,8bit,ctx None,8.5 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1roetd9/qwen_35_27b_macbook_m4_pro_48gb/",
   "tps": 8.5,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M2 Max (38-core GPU, 96 GB)",
   "id": "siliconscore-m2-max-38-core-gpu-96-gb-qwen3-5-27b-8bit-ctxnone",
   "model": "Qwen3.5-27B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "8bit",
   "quote": "M2 Max (38-core GPU, 96 GB),Qwen3.5-27B,8bit,ctx None,12 tok/s",
   "ram_gb": 30,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1rkkeqh/qwen35_models_ultra_slow_for_anyone_else_compared/",
   "tps": 12,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M2 Ultra (GPU count not published, 128 GB)",
   "id": "siliconscore-m2-ultra-gpu-count-not-published-128-gb-qwen3-5-27b-8bit-ctxnone",
   "model": "Qwen3.5-27B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "8bit",
   "quote": "M2 Ultra (GPU count not published, 128 GB),Qwen3.5-27B,8bit,ctx None,20.6 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://github.com/ml-explore/mlx-lm/pull/990",
   "tps": 20.6,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M5 Max (128 GB)",
   "id": "siliconscore-m5-max-128-gb-qwen3-5-27b-4bit-ctxnone",
   "model": "Qwen3.5-27B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "4bit",
   "quote": "M5 Max (128 GB),Qwen3.5-27B,4bit,ctx None,31.6 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1rzkw4x/m5_max_128g_performance_tests_i_just_got_my_new/",
   "tps": 31.6,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 8192,
   "date": null,
   "flags": [],
   "hardware": "M5 Max (128 GB)",
   "id": "siliconscore-m5-max-128-gb-qwen3-5-27b-q6-k-ctx8192",
   "model": "Qwen3.5-27B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q6_K",
   "quote": "M5 Max (128 GB),Qwen3.5-27B,Q6_K,ctx 8192,16.5 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1rzkw4x/m5_max_128g_performance_tests_i_just_got_my_new/",
   "tps": 16.5,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": 4106,
   "date": null,
   "flags": [],
   "hardware": "M5 Max (128 GB)",
   "id": "siliconscore-m5-max-128-gb-qwen3-5-122b-a10b-4bit-ctx4106",
   "model": "Qwen3.5-122B-A10B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 881.466,
   "provenance": "sourced",
   "quant": "4bit",
   "quote": "M5 Max (128 GB),Qwen3.5-122B-A10B,4bit,ctx 4106,65.853 tok/s",
   "ram_gb": 71.91,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1rqnpvj/m5_max_just_arrived_benchmarks_incoming/",
   "tps": 65.853,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": 16394,
   "date": null,
   "flags": [],
   "hardware": "M5 Max (128 GB)",
   "id": "siliconscore-m5-max-128-gb-qwen3-5-122b-a10b-4bit-ctx16394",
   "model": "Qwen3.5-122B-A10B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 1239.734,
   "provenance": "sourced",
   "quant": "4bit",
   "quote": "M5 Max (128 GB),Qwen3.5-122B-A10B,4bit,ctx 16394,60.639 tok/s",
   "ram_gb": 73.803,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1rqnpvj/m5_max_just_arrived_benchmarks_incoming/",
   "tps": 60.639,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": 32778,
   "date": null,
   "flags": [],
   "hardware": "M5 Max (128 GB)",
   "id": "siliconscore-m5-max-128-gb-qwen3-5-122b-a10b-4bit-ctx32778",
   "model": "Qwen3.5-122B-A10B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 1067.824,
   "provenance": "sourced",
   "quant": "4bit",
   "quote": "M5 Max (128 GB),Qwen3.5-122B-A10B,4bit,ctx 32778,54.923 tok/s",
   "ram_gb": 76.397,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1rqnpvj/m5_max_just_arrived_benchmarks_incoming/",
   "tps": 54.923,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": 4105,
   "date": null,
   "flags": [],
   "hardware": "M5 Max (128 GB)",
   "id": "siliconscore-m5-max-128-gb-qwen3-coder-next-8bit-ctx4105",
   "model": "Qwen3-Coder-Next",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 754.927,
   "provenance": "sourced",
   "quant": "8bit",
   "quote": "M5 Max (128 GB),Qwen3-Coder-Next,8bit,ctx 4105,79.296 tok/s",
   "ram_gb": 87.068,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1rqnpvj/m5_max_just_arrived_benchmarks_incoming/",
   "tps": 79.296,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": 16393,
   "date": null,
   "flags": [],
   "hardware": "M5 Max (128 GB)",
   "id": "siliconscore-m5-max-128-gb-qwen3-coder-next-8bit-ctx16393",
   "model": "Qwen3-Coder-Next",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 1802.144,
   "provenance": "sourced",
   "quant": "8bit",
   "quote": "M5 Max (128 GB),Qwen3-Coder-Next,8bit,ctx 16393,74.293 tok/s",
   "ram_gb": 88.176,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1rqnpvj/m5_max_just_arrived_benchmarks_incoming/",
   "tps": 74.293,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": 32777,
   "date": null,
   "flags": [],
   "hardware": "M5 Max (128 GB)",
   "id": "siliconscore-m5-max-128-gb-qwen3-coder-next-8bit-ctx32777",
   "model": "Qwen3-Coder-Next",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 1887.158,
   "provenance": "sourced",
   "quant": "8bit",
   "quote": "M5 Max (128 GB),Qwen3-Coder-Next,8bit,ctx 32777,68.624 tok/s",
   "ram_gb": 89.652,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1rqnpvj/m5_max_just_arrived_benchmarks_incoming/",
   "tps": 68.624,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": 65545,
   "date": null,
   "flags": [],
   "hardware": "M5 Max (128 GB)",
   "id": "siliconscore-m5-max-128-gb-qwen3-coder-next-8bit-ctx65545",
   "model": "Qwen3-Coder-Next",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 1432.73,
   "provenance": "sourced",
   "quant": "8bit",
   "quote": "M5 Max (128 GB),Qwen3-Coder-Next,8bit,ctx 65545,48.212 tok/s",
   "ram_gb": 92.605,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1rqnpvj/m5_max_just_arrived_benchmarks_incoming/",
   "tps": 48.212,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M5 Pro (64 GB)",
   "id": "siliconscore-m5-pro-64-gb-qwen3-5-9b-q4-k-medium-ctxnone",
   "model": "Qwen3.5-9B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M5 Pro (64 GB),Qwen3.5-9B,Q4_K - Medium,ctx None,25 tok/s; ttft 0.765s",
   "ram_gb": 13.8,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.sharpai.org/benchmark/",
   "tps": 25,
   "ttft_s": 0.765
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M5 Pro (64 GB)",
   "id": "siliconscore-m5-pro-64-gb-qwen3-5-27b-q4-k-medium-ctxnone",
   "model": "Qwen3.5-27B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M5 Pro (64 GB),Qwen3.5-27B,Q4_K - Medium,ctx None,10 tok/s; ttft 2.156s",
   "ram_gb": 24.9,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.sharpai.org/benchmark/",
   "tps": 10,
   "ttft_s": 2.156
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M5 Pro (64 GB)",
   "id": "siliconscore-m5-pro-64-gb-qwen3-5-35b-a3b-q4-k-l-ctxnone",
   "model": "Qwen3.5-35B-A3B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K_L",
   "quote": "M5 Pro (64 GB),Qwen3.5-35B-A3B,Q4_K_L,ctx None,41.9 tok/s; ttft 0.435s",
   "ram_gb": 27.2,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.sharpai.org/benchmark/",
   "tps": 41.9,
   "ttft_s": 0.435
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M5 Pro (64 GB)",
   "id": "siliconscore-m5-pro-64-gb-qwen3-5-122b-a10b-iq1-m-ctxnone",
   "model": "Qwen3.5-122B-A10B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "IQ1_M",
   "quote": "M5 Pro (64 GB),Qwen3.5-122B-A10B,IQ1_M,ctx None,18 tok/s; ttft 1.627s",
   "ram_gb": 40.8,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.sharpai.org/benchmark/",
   "tps": 18,
   "ttft_s": 1.627
  },
  {
   "backend": "flash-moe",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M5 Max (128 GB)",
   "id": "siliconscore-m5-max-128-gb-qwen3-5-397b-a17b-4bit-ctxnone",
   "model": "Qwen3.5-397B-A17B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "4bit",
   "quote": "M5 Max (128 GB),Qwen3.5-397B-A17B,4bit,ctx None,12.99 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1s30vs2/took_the_48gb_flashmoe_benchmark_and_ran_it_on/",
   "tps": 12.99,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 Ultra (512 GB)",
   "id": "siliconscore-m3-ultra-512-gb-qwen3-5-397b-a17b-q4-1bit-ctxnone",
   "model": "Qwen3.5-397B-A17B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "q4.1bit",
   "quote": "M3 Ultra (512 GB),Qwen3.5-397B-A17B,q4.1bit,ctx None,40.2 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://siliconscore.com/benchmarks.json",
   "tps": 40.2,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": 1024,
   "date": null,
   "flags": [],
   "hardware": "M3 Ultra (512 GB)",
   "id": "siliconscore-m3-ultra-512-gb-glm-5-4bit-ctx1024",
   "model": "GLM-5",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 187,
   "provenance": "sourced",
   "quant": "4bit",
   "quote": "M3 Ultra (512 GB),GLM-5,4bit,ctx 1024,16.7 tok/s; ttft 5.477s",
   "ram_gb": 391.82,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1rdkze3/m3_ultra_512gb_realworld_performance_of/",
   "tps": 16.7,
   "ttft_s": 5.477
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": 4096,
   "date": null,
   "flags": [],
   "hardware": "M3 Ultra (512 GB)",
   "id": "siliconscore-m3-ultra-512-gb-glm-5-4bit-ctx4096",
   "model": "GLM-5",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 180.1,
   "provenance": "sourced",
   "quant": "4bit",
   "quote": "M3 Ultra (512 GB),GLM-5,4bit,ctx 4096,13.7 tok/s; ttft 22.745s",
   "ram_gb": 394.07,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1rdkze3/m3_ultra_512gb_realworld_performance_of/",
   "tps": 13.7,
   "ttft_s": 22.745
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": 8192,
   "date": null,
   "flags": [],
   "hardware": "M3 Ultra (512 GB)",
   "id": "siliconscore-m3-ultra-512-gb-glm-5-4bit-ctx8192",
   "model": "GLM-5",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 154.1,
   "provenance": "sourced",
   "quant": "4bit",
   "quote": "M3 Ultra (512 GB),GLM-5,4bit,ctx 8192,13.2 tok/s; ttft 53.169s",
   "ram_gb": 396.69,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1rdkze3/m3_ultra_512gb_realworld_performance_of/",
   "tps": 13.2,
   "ttft_s": 53.169
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "M3 Ultra (512 GB)",
   "id": "siliconscore-m3-ultra-512-gb-glm-5-4bit-ctx16384",
   "model": "GLM-5",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 117.4,
   "provenance": "sourced",
   "quant": "4bit",
   "quote": "M3 Ultra (512 GB),GLM-5,4bit,ctx 16384,12 tok/s; ttft 139.545s",
   "ram_gb": 402.72,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1rdkze3/m3_ultra_512gb_realworld_performance_of/",
   "tps": 12,
   "ttft_s": 139.545
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": 32768,
   "date": null,
   "flags": [],
   "hardware": "M3 Ultra (512 GB)",
   "id": "siliconscore-m3-ultra-512-gb-glm-5-4bit-ctx32768",
   "model": "GLM-5",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 77.7,
   "provenance": "sourced",
   "quant": "4bit",
   "quote": "M3 Ultra (512 GB),GLM-5,4bit,ctx 32768,10.7 tok/s; ttft 421.955s",
   "ram_gb": 415.41,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1rdkze3/m3_ultra_512gb_realworld_performance_of/",
   "tps": 10.7,
   "ttft_s": 421.955
  },
  {
   "backend": "LM Studio",
   "batch": null,
   "ctx": 131072,
   "date": null,
   "flags": [],
   "hardware": "M4 Pro (48 GB)",
   "id": "siliconscore-m4-pro-48-gb-devstral-small-1-1-6bit-ctx131072",
   "model": "Devstral Small 1.1",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "6bit",
   "quote": "M4 Pro (48 GB),Devstral Small 1.1,6bit,ctx 131072,12.88 tok/s; ttft 5.91s",
   "ram_gb": 18.51,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1nkfvrl/local_llm_coding_stack_24gb_minimum_ideal_36gb/",
   "tps": 12.88,
   "ttft_s": 5.91
  },
  {
   "backend": "LM Studio",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Max (128 GB)",
   "id": "siliconscore-m4-max-128-gb-devstral-small-1-1-4bit-ctxnone",
   "model": "Devstral Small 1.1",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "4bit",
   "quote": "M4 Max (128 GB),Devstral Small 1.1,4bit,ctx None,33 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1m1t19r/any_experiences_running_llms_on_a_macbook/",
   "tps": 33,
   "ttft_s": null
  },
  {
   "backend": "LM Studio",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M2 (24 GB)",
   "id": "siliconscore-m2-24-gb-devstral-small-1-1-4bit-ctxnone",
   "model": "Devstral Small 1.1",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "4bit",
   "quote": "M2 (24 GB),Devstral Small 1.1,4bit,ctx None,6 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1m1t19r/any_experiences_running_llms_on_a_macbook/",
   "tps": 6,
   "ttft_s": null
  },
  {
   "backend": "LM Studio",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M3 Ultra (512 GB)",
   "id": "siliconscore-m3-ultra-512-gb-devstral-small-1-1-4bit-ctxnone",
   "model": "Devstral Small 1.1",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "4bit",
   "quote": "M3 Ultra (512 GB),Devstral Small 1.1,4bit,ctx None,43 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1pry2v7/people_using_devstral_2_123b_how_has_it_been/",
   "tps": 43,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 8192,
   "date": null,
   "flags": [],
   "hardware": "M5 Max (128 GB)",
   "id": "siliconscore-m5-max-128-gb-gemma-3-27b-q6-k-ctx8192",
   "model": "Gemma 3 27B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 391,
   "provenance": "sourced",
   "quant": "Q6_K",
   "quote": "M5 Max (128 GB),Gemma 3 27B,Q6_K,ctx 8192,20 tok/s",
   "ram_gb": null,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1s0czc4/round_2_followup_m5_max_128g_performance_tests_i/",
   "tps": 20,
   "ttft_s": null
  },
  {
   "backend": "LM Studio",
   "batch": null,
   "ctx": 128000,
   "date": null,
   "flags": [],
   "hardware": "M3 Ultra (512 GB)",
   "id": "siliconscore-m3-ultra-512-gb-gemma-3-27b-bf16-ctx128000",
   "model": "Gemma 3 27B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "bf16",
   "quote": "M3 Ultra (512 GB),Gemma 3 27B,bf16,ctx 128000,11.19 tok/s; ttft 1.72s",
   "ram_gb": 52.57,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1kfi8xh/benchmark_quickanddirty_test_of_5_models_on_a_mac/",
   "tps": 11.19,
   "ttft_s": 1.72
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 2048,
   "date": null,
   "flags": [],
   "hardware": "M4 Max",
   "id": "siliconscore-m4-max-qwen3-5-27b-q4-k-ctx2048",
   "model": "Qwen3.5-27B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 222.23,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "M4 Max,Qwen3.5-27B,Q4_K,ctx 2048,16.69 tok/s",
   "ram_gb": 16.4,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1rlyukb/m4_max_llamacpp_benchmarks_of_qwen35_35b_and_27b/",
   "tps": 16.69,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 8192,
   "date": null,
   "flags": [],
   "hardware": "M4 Max",
   "id": "siliconscore-m4-max-qwen3-5-27b-q4-k-ctx8192",
   "model": "Qwen3.5-27B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 209.3,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "M4 Max,Qwen3.5-27B,Q4_K,ctx 8192,16.14 tok/s",
   "ram_gb": 16.4,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1rlyukb/m4_max_llamacpp_benchmarks_of_qwen35_35b_and_27b/",
   "tps": 16.14,
   "ttft_s": null
  },
  {
   "backend": "llama.cpp",
   "batch": null,
   "ctx": 16384,
   "date": null,
   "flags": [],
   "hardware": "M4 Max",
   "id": "siliconscore-m4-max-qwen3-5-27b-q4-k-ctx16384",
   "model": "Qwen3.5-27B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 195.44,
   "provenance": "sourced",
   "quant": "Q4_K",
   "quote": "M4 Max,Qwen3.5-27B,Q4_K,ctx 16384,15.75 tok/s",
   "ram_gb": 16.4,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://www.reddit.com/r/LocalLLaMA/comments/1rlyukb/m4_max_llamacpp_benchmarks_of_qwen35_35b_and_27b/",
   "tps": 15.75,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M5 Max (128 GB)",
   "id": "siliconscore-m5-max-128-gb-gemma-4-26b-a4b-q4-k-medium-ctxnone",
   "model": "Gemma 4 26B-A4B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M5 Max (128 GB),Gemma 4 26B-A4B,Q4_K - Medium,ctx None,50 tok/s; ttft 0.5s",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://llmcheck.net/data/benchmarks.csv",
   "tps": 50,
   "ttft_s": 0.5
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Max (48 GB)",
   "id": "siliconscore-m4-max-48-gb-gemma-4-26b-a4b-q4-k-medium-ctxnone",
   "model": "Gemma 4 26B-A4B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Max (48 GB),Gemma 4 26B-A4B,Q4_K - Medium,ctx None,40 tok/s; ttft 0.7s",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://llmcheck.net/data/benchmarks.csv",
   "tps": 40,
   "ttft_s": 0.7
  },
  {
   "backend": "Ollama",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Pro (24 GB)",
   "id": "siliconscore-m4-pro-24-gb-gemma-4-26b-a4b-q4-k-medium-ctxnone",
   "model": "Gemma 4 26B-A4B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Pro (24 GB),Gemma 4 26B-A4B,Q4_K - Medium,ctx None,28 tok/s; ttft 1s",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://llmcheck.net/data/benchmarks.csv",
   "tps": 28,
   "ttft_s": 1
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M5 Max (128 GB)",
   "id": "siliconscore-m5-max-128-gb-gemma-4-31b-q4-k-medium-ctxnone",
   "model": "Gemma-4-31B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M5 Max (128 GB),Gemma 4 31B,Q4_K - Medium,ctx None,26 tok/s; ttft 0.7s",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://llmcheck.net/data/benchmarks.csv",
   "tps": 26,
   "ttft_s": 0.7
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Max (48 GB)",
   "id": "siliconscore-m4-max-48-gb-gemma-4-31b-q4-k-medium-ctxnone",
   "model": "Gemma-4-31B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Max (48 GB),Gemma 4 31B,Q4_K - Medium,ctx None,24 tok/s; ttft 1.1s",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://llmcheck.net/data/benchmarks.csv",
   "tps": 24,
   "ttft_s": 1.1
  },
  {
   "backend": "Ollama",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Pro (24 GB)",
   "id": "siliconscore-m4-pro-24-gb-gemma-4-31b-q4-k-medium-ctxnone",
   "model": "Gemma-4-31B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Pro (24 GB),Gemma 4 31B,Q4_K - Medium,ctx None,12 tok/s; ttft 1.4s",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://llmcheck.net/data/benchmarks.csv",
   "tps": 12,
   "ttft_s": 1.4
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M5 Max (128 GB)",
   "id": "siliconscore-m5-max-128-gb-qwen3-6-35b-a3b-q4-k-medium-ctxnone",
   "model": "Qwen3.6-35B-A3B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M5 Max (128 GB),Qwen3.6-35B-A3B,Q4_K - Medium,ctx None,55 tok/s; ttft 0.6s",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://llmcheck.net/data/benchmarks.csv",
   "tps": 55,
   "ttft_s": 0.6
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Max (48 GB)",
   "id": "siliconscore-m4-max-48-gb-qwen3-6-35b-a3b-q4-k-medium-ctxnone",
   "model": "Qwen3.6-35B-A3B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Max (48 GB),Qwen3.6-35B-A3B,Q4_K - Medium,ctx None,42 tok/s; ttft 0.9s",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://llmcheck.net/data/benchmarks.csv",
   "tps": 42,
   "ttft_s": 0.9
  },
  {
   "backend": "Ollama",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Pro (24 GB)",
   "id": "siliconscore-m4-pro-24-gb-qwen3-6-35b-a3b-q4-k-medium-ctxnone",
   "model": "Qwen3.6-35B-A3B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Pro (24 GB),Qwen3.6-35B-A3B,Q4_K - Medium,ctx None,32 tok/s; ttft 1.2s",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://llmcheck.net/data/benchmarks.csv",
   "tps": 32,
   "ttft_s": 1.2
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Ultra (192 GB)",
   "id": "siliconscore-m4-ultra-192-gb-mistral-small-4-119b-q4-k-medium-ctxnone",
   "model": "Mistral Small 4 119B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Ultra (192 GB),Mistral Small 4 119B,Q4_K - Medium,ctx None,45 tok/s; ttft 0.7s",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://llmcheck.net/data/benchmarks.json",
   "tps": 45,
   "ttft_s": 0.7
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Ultra (192 GB)",
   "id": "siliconscore-m4-ultra-192-gb-llama-3-3-70b-q4-k-medium-ctxnone",
   "model": "Llama 3.3 70B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Ultra (192 GB),Llama 3.3 70B,Q4_K - Medium,ctx None,21 tok/s; ttft 2s",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://llmcheck.net/data/benchmarks.json",
   "tps": 21,
   "ttft_s": 2
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Ultra (192 GB)",
   "id": "siliconscore-m4-ultra-192-gb-deepseek-r1-distill-llama-70b-q4-k-medium-ctxnone",
   "model": "DeepSeek R1 70B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Ultra (192 GB),DeepSeek R1 70B,Q4_K - Medium,ctx None,21 tok/s; ttft 2.2s",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://llmcheck.net/data/benchmarks.json",
   "tps": 21,
   "ttft_s": 2.2
  },
  {
   "backend": "Ollama",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Ultra (192 GB)",
   "id": "siliconscore-m4-ultra-192-gb-qwen-2-5-72b-q4-k-medium-ctxnone",
   "model": "Qwen 2.5 72B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Ultra (192 GB),Qwen 2.5 72B,Q4_K - Medium,ctx None,21 tok/s; ttft 2.5s",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://llmcheck.net/data/benchmarks.json",
   "tps": 21,
   "ttft_s": 2.5
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Ultra (192 GB)",
   "id": "siliconscore-m4-ultra-192-gb-qwen-3-235b-a22b-q4-k-medium-ctxnone",
   "model": "Qwen 3 235B-A22B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Ultra (192 GB),Qwen 3 235B-A22B,Q4_K - Medium,ctx None,22 tok/s; ttft 1.5s",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://llmcheck.net/data/benchmarks.json",
   "tps": 22,
   "ttft_s": 1.5
  },
  {
   "backend": "Ollama",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Max (128 GB)",
   "id": "siliconscore-m4-max-128-gb-qwen3-6-27b-q4-k-medium-ctxnone",
   "model": "Qwen3.6-27B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": 114.49,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Max (128 GB),Qwen3.6-27B,Q4_K - Medium,ctx None,16.56 tok/s",
   "ram_gb": 29,
   "retrieved": "2026-10-03",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://huggingface.co/batiai/Qwen3.6-27B-GGUF",
   "tps": 16.56,
   "ttft_s": null
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Ultra (192 GB)",
   "id": "siliconscore-m4-ultra-192-gb-llama-4-scout-17b-16e-q4-k-medium-ctxnone",
   "model": "Llama 4 Scout 17B-16E",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Ultra (192 GB),Llama 4 Scout 17B-16E,Q4_K - Medium,ctx None,30 tok/s; ttft 1.3s",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://llmcheck.net/data/benchmarks.json",
   "tps": 30,
   "ttft_s": 1.3
  },
  {
   "backend": "MLX",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Ultra (192 GB)",
   "id": "siliconscore-m4-ultra-192-gb-gpt-oss-120b-q4-k-medium-ctxnone",
   "model": "gpt-oss 120B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Ultra (192 GB),gpt-oss 120B,Q4_K - Medium,ctx None,13 tok/s; ttft 4.2s",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://llmcheck.net/data/benchmarks.json",
   "tps": 13,
   "ttft_s": 4.2
  },
  {
   "backend": "Ollama",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M5 Max (64 GB)",
   "id": "siliconscore-m5-max-64-gb-gemma-4-31b-q4-k-medium-ctxnone",
   "model": "Gemma-4-31B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M5 Max (64 GB),Gemma 4 31B,Q4_K - Medium,ctx None,26 tok/s; ttft 0.9s",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://llmcheck.net/data/benchmarks.json",
   "tps": 26,
   "ttft_s": 0.9
  },
  {
   "backend": "Ollama",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M5 Max (64 GB)",
   "id": "siliconscore-m5-max-64-gb-qwen3-6-35b-a3b-q4-k-medium-ctxnone",
   "model": "Qwen3.6-35B-A3B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M5 Max (64 GB),Qwen3.6-35B-A3B,Q4_K - Medium,ctx None,48 tok/s; ttft 0.8s",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://llmcheck.net/data/benchmarks.json",
   "tps": 48,
   "ttft_s": 0.8
  },
  {
   "backend": "Ollama",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Max (48 GB)",
   "id": "siliconscore-m4-max-48-gb-qwen-3-30b-a3b-q4-k-medium-ctxnone",
   "model": "Qwen 3 30B-A3B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Max (48 GB),Qwen 3 30B-A3B,Q4_K - Medium,ctx None,42 tok/s; ttft 0.9s",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://llmcheck.net/data/benchmarks.json",
   "tps": 42,
   "ttft_s": 0.9
  },
  {
   "backend": "Ollama",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M5 Max (128 GB)",
   "id": "siliconscore-m5-max-128-gb-qwen3-5-35b-a3b-q4-k-medium-ctxnone",
   "model": "Qwen3.5-35B-A3B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M5 Max (128 GB),Qwen3.5-35B-A3B,Q4_K - Medium,ctx None,48 tok/s; ttft 0.9s",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://llmcheck.net/data/benchmarks.json",
   "tps": 48,
   "ttft_s": 0.9
  },
  {
   "backend": "Ollama",
   "batch": null,
   "ctx": null,
   "date": null,
   "flags": [],
   "hardware": "M4 Ultra (192 GB)",
   "id": "siliconscore-m4-ultra-192-gb-qwen-3-32b-q4-k-medium-ctxnone",
   "model": "Qwen 3 32B",
   "notes": "source label: reference run; verified=False",
   "params": null,
   "power_w": null,
   "pp_tps": null,
   "provenance": "sourced",
   "quant": "Q4_K - Medium",
   "quote": "M4 Ultra (192 GB),Qwen 3 32B,Q4_K - Medium,ctx None,45 tok/s; ttft 1s",
   "ram_gb": null,
   "retrieved": "2026-10-06",
   "source_name": "Silicon Score benchmark audit",
   "source_url": "https://llmcheck.net/data/benchmarks.json",
   "tps": 45,
   "ttft_s": 1
  }
 ],
 "retrieved": "2026-10-22"
}